diff --git a/.gitattributes b/.gitattributes index 71e6abe8f61ca70dc7cf9b606e68fae741b03e87..beb0f62facd0b348d2e9b8e86ddee5a3aeb76a93 100644 --- a/.gitattributes +++ b/.gitattributes @@ -1108,3 +1108,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2024/In[[:space:]]Defense[[:space:]]of[[:space:]]Lazy[[:space:]]Visual[[:space:]]Grounding[[:space:]]for[[:space:]]Open-Vocabulary[[:space:]]Semantic[[:space:]]Segmentation/cdf9d4c4-1527-4261-89bb-f885a518daa3_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Inf-DiT_[[:space:]]Upsampling[[:space:]]any-resolution[[:space:]]image[[:space:]]with[[:space:]]memory-efficient[[:space:]]diffusion[[:space:]]transformer./f3002cff-588f-4f87-ae99-26bf23b35c9e_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/InfMAE_[[:space:]]A[[:space:]]Foundation[[:space:]]Model[[:space:]]in[[:space:]]The[[:space:]]Infrared[[:space:]]Modality/27243eca-1887-4c8e-8b92-89b601b015c1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Infinite-ID_[[:space:]]Identity-preserved[[:space:]]Personalization[[:space:]]via[[:space:]]ID-semantics[[:space:]]Decoupling[[:space:]]Paradigm/6b2c0220-7e0d-410c-b7de-7b3ff87a9f85_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/InfoNorm_[[:space:]]Mutual[[:space:]]Information[[:space:]]Shaping[[:space:]]of[[:space:]]Normals[[:space:]]for[[:space:]]Sparse-View[[:space:]]Reconstruction/8f9c186f-c37a-4b28-8f4e-c701ceacc591_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Information[[:space:]]Bottleneck[[:space:]]Based[[:space:]]Data[[:space:]]Correction[[:space:]]in[[:space:]]Continual[[:space:]]Learning/6c4e72fe-4072-432b-96c2-9a488daff608_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/InsMapper_[[:space:]]Exploring[[:space:]]Inner-instance[[:space:]]Information[[:space:]]for[[:space:]]Vectorized[[:space:]]HD[[:space:]]Mapping/703739e6-709f-487f-9515-53b438aabcdc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Insect[[:space:]]Identification[[:space:]]in[[:space:]]the[[:space:]]Wild_[[:space:]]The[[:space:]]AMI[[:space:]]Dataset/deef9d64-c259-4177-ab17-679ea66a461f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/InstaStyle_[[:space:]]Inversion[[:space:]]Noise[[:space:]]of[[:space:]]a[[:space:]]Stylized[[:space:]]Image[[:space:]]is[[:space:]]Secretly[[:space:]]a[[:space:]]Style[[:space:]]Adviser/938fb53e-3bec-4f8a-bc7e-f5b36dd7b576_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Instance-dependent[[:space:]]Noisy-label[[:space:]]Learning[[:space:]]with[[:space:]]Graphical[[:space:]]Model[[:space:]]Based[[:space:]]Noise-rate[[:space:]]Estimation/8698146c-3791-42ec-912e-34881bbacb25_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Instant[[:space:]]3D[[:space:]]Human[[:space:]]Avatar[[:space:]]Generation[[:space:]]using[[:space:]]Image[[:space:]]Diffusion[[:space:]]Models/848a10c5-4ff1-45f7-8010-a0858cf611ea_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Instant[[:space:]]Uncertainty[[:space:]]Calibration[[:space:]]of[[:space:]]NeRFs[[:space:]]Using[[:space:]]a[[:space:]]Meta-Calibrator/6c9ab90c-9fb8-405a-a3f6-eaa68a4cea3f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/InstructGIE_[[:space:]]Towards[[:space:]]Generalizable[[:space:]]Image[[:space:]]Editing/d8d7c0e8-e8d9-4aa1-9b02-c4948b5c2541_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/InstructIR_[[:space:]]High-Quality[[:space:]]Image[[:space:]]Restoration[[:space:]]Following[[:space:]]Human[[:space:]]Instructions/86f2cef3-3642-43ee-930e-b6ddf921d714_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Instruction[[:space:]]Tuning-free[[:space:]]Visual[[:space:]]Token[[:space:]]Complement[[:space:]]for[[:space:]]Multimodal[[:space:]]LLMs/95b8eb87-0839-45ae-a393-d76dd91f7ccf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Integer-Valued[[:space:]]Training[[:space:]]and[[:space:]]Spike-driven[[:space:]]Inference[[:space:]]Spiking[[:space:]]Neural[[:space:]]Network[[:space:]]for[[:space:]]High-performance[[:space:]]and[[:space:]]Energy-efficient[[:space:]]Object[[:space:]]Detection/b44a5b42-9423-4159-817b-8bb17aae60da_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Integrating[[:space:]]Markov[[:space:]]Blanket[[:space:]]Discovery[[:space:]]into[[:space:]]Causal[[:space:]]Representation[[:space:]]Learning[[:space:]]for[[:space:]]Domain[[:space:]]Generalization/bb833fed-6970-4db4-b3da-646e63e5493d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Integration[[:space:]]of[[:space:]]Global[[:space:]]and[[:space:]]Local[[:space:]]Representations[[:space:]]for[[:space:]]Fine-grained[[:space:]]Cross-modal[[:space:]]Alignment/35329370-3365-4c03-a82d-856d1e6d798d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Inter-Class[[:space:]]Topology[[:space:]]Alignment[[:space:]]for[[:space:]]Efficient[[:space:]]Black-Box[[:space:]]Substitute[[:space:]]Attacks/5bfa27b2-acad-4e1e-b905-e3f3b1d14c81_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/InterFusion_[[:space:]]Text-Driven[[:space:]]Generation[[:space:]]of[[:space:]]3D[[:space:]]Human-Object[[:space:]]Interaction/d1b8e0cb-6c9e-44c9-80dd-158238285207_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Interaction-centric[[:space:]]Spatio-Temporal[[:space:]]Context[[:space:]]Reasoning[[:space:]]for[[:space:]]Multi-Person[[:space:]]Video[[:space:]]HOI[[:space:]]Recognition/0a79d390-9d09-4042-af06-6b62e335816f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Interactive[[:space:]]3D[[:space:]]Object[[:space:]]Detection[[:space:]]with[[:space:]]Prompts/20a1b205-7505-40a6-b499-29fffaad19e8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Interleaving[[:space:]]One-Class[[:space:]]and[[:space:]]Weakly-Supervised[[:space:]]Models[[:space:]]with[[:space:]]Adaptive[[:space:]]Thresholding[[:space:]]for[[:space:]]Unsupervised[[:space:]]Video[[:space:]]Anomaly[[:space:]]Detection/0ff966a8-e3c5-432a-8beb-e808413549e5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/InternVideo2_[[:space:]]Scaling[[:space:]]Foundation[[:space:]]Models[[:space:]]for[[:space:]]Multimodal[[:space:]]Video[[:space:]]Understanding/3f170c5b-f8dd-444c-a833-330298501085_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Interpretability-Guided[[:space:]]Test-Time[[:space:]]Adversarial[[:space:]]Defense/d8bed3ff-2ae8-4bfb-abf2-7d81aee96721_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Intrinsic[[:space:]]Single-Image[[:space:]]HDR[[:space:]]Reconstruction/e55f56f5-9c00-4140-aab7-15323b235acd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/IntrinsicAnything_[[:space:]]Learning[[:space:]]Diffusion[[:space:]]Priors[[:space:]]for[[:space:]]Inverse[[:space:]]Rendering[[:space:]]Under[[:space:]]Unknown[[:space:]]Illumination/b4b4420f-76ed-4ddf-baa7-e217742f08de_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Introducing[[:space:]]Routing[[:space:]]Functions[[:space:]]to[[:space:]]Vision-Language[[:space:]]Parameter-Efficient[[:space:]]Fine-Tuning[[:space:]]with[[:space:]]Low-Rank[[:space:]]Bottlenecks/61df44b9-0f88-4bc8-892b-514f717f8da8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Invertible[[:space:]]Neural[[:space:]]Warp[[:space:]]for[[:space:]]NeRF/9222ab52-5e8c-4d58-9015-e57bdb13f7a5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Investigating[[:space:]]Style[[:space:]]Similarity[[:space:]]in[[:space:]]Diffusion[[:space:]]Models/e6cae220-dbdc-4ecf-90b6-28f1684a0527_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Is[[:space:]]Retain[[:space:]]Set[[:space:]]All[[:space:]]You[[:space:]]Need[[:space:]]in[[:space:]]Machine[[:space:]]Unlearning_[[:space:]]Restoring[[:space:]]Performance[[:space:]]of[[:space:]]Unlearned[[:space:]]Models[[:space:]]with[[:space:]]Out-Of-Distribution[[:space:]]Images/ef16a8f0-48f4-4778-ba31-599cb1e6436c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Is[[:space:]]user[[:space:]]feedback[[:space:]]always[[:space:]]informative_[[:space:]]Retrieval[[:space:]]Latent[[:space:]]Defending[[:space:]]for[[:space:]]Semi-Supervised[[:space:]]Domain[[:space:]]Adaptation[[:space:]]without[[:space:]]Source[[:space:]]Data/0589fa53-73c5-413a-aac9-f37670a1de05_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Isomorphic[[:space:]]Pruning[[:space:]]for[[:space:]]Vision[[:space:]]Models/bf292aec-a3cb-46ef-92f7-27df67a69e35_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ItTakesTwo_[[:space:]]Leveraging[[:space:]]Peer[[:space:]]Representations[[:space:]]for[[:space:]]Semi-supervised[[:space:]]LiDAR[[:space:]]Semantic[[:space:]]Segmentation/69e10020-97d7-4fa9-96f3-57b936f4e871_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Iterative[[:space:]]Ensemble[[:space:]]Training[[:space:]]with[[:space:]]Anti-Gradient[[:space:]]Control[[:space:]]for[[:space:]]Mitigating[[:space:]]Memorization[[:space:]]in[[:space:]]Diffusion[[:space:]]Models/d774ff0f-2daf-4eeb-815c-c6da2eb3ab17_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/JDT3D_[[:space:]]Addressing[[:space:]]the[[:space:]]Gaps[[:space:]]in[[:space:]]LiDAR-Based[[:space:]]Tracking-by-Attention/b2b98b8d-db63-4b59-ae29-0da5121d4534_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Joint[[:space:]]RGB-Spectral[[:space:]]Decomposition[[:space:]]Model[[:space:]]Guided[[:space:]]Image[[:space:]]Enhancement[[:space:]]in[[:space:]]Mobile[[:space:]]Photography/2ac99c58-8fa1-4a8d-b568-c7870f80ab53_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/JointDreamer_[[:space:]]Ensuring[[:space:]]Geometry[[:space:]]Consistency[[:space:]]and[[:space:]]Text[[:space:]]Congruence[[:space:]]in[[:space:]]Text-to-3D[[:space:]]Generation[[:space:]]via[[:space:]]Joint[[:space:]]Score[[:space:]]Distillation/4667ae70-c38c-469d-a697-9d92865631f2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Just[[:space:]]a[[:space:]]Hint_[[:space:]]Point-Supervised[[:space:]]Camouflaged[[:space:]]Object[[:space:]]Detection/6f40032b-910c-409c-99b0-064b71681182_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/KDProR_[[:space:]]A[[:space:]]Knowledge-Decoupling[[:space:]]Probabilistic[[:space:]]Framework[[:space:]]for[[:space:]]Video-Text[[:space:]]Retrieval/8778d22d-a6f9-4fed-bc5f-fddacba7cf0b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/KFD-NeRF_[[:space:]]Rethinking[[:space:]]Dynamic[[:space:]]NeRF[[:space:]]with[[:space:]]Kalman[[:space:]]Filter/3bae14dd-a29b-4634-bf7b-d258a05d1f55_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/KMTalk_[[:space:]]Speech-Driven[[:space:]]3D[[:space:]]Facial[[:space:]]Animation[[:space:]]with[[:space:]]Key[[:space:]]Motion[[:space:]]Embedding/8f222025-36f0-4f3b-ab78-77e262291c3e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Kalman-Inspired[[:space:]]Feature[[:space:]]Propagation[[:space:]]for[[:space:]]Video[[:space:]]Face[[:space:]]Super-Resolution/86d2672e-c2f0-471a-a7a6-101117941b3f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Kernel[[:space:]]Diffusion_[[:space:]]An[[:space:]]Alternate[[:space:]]Approach[[:space:]]to[[:space:]]Blind[[:space:]]Deconvolution/0f568881-c0ff-47c0-b18b-592b7cbf5a93_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Keypoint[[:space:]]Promptable[[:space:]]Re-Identification/21228ec1-ee24-4ba9-acb3-211d122a99e2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/KeypointDETR_[[:space:]]An[[:space:]]End-to-End[[:space:]]3D[[:space:]]Keypoint[[:space:]]Detector/d05c910b-e5c7-4df0-b517-a9b20e7a9b64_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Kinetic[[:space:]]Typography[[:space:]]Diffusion[[:space:]]Model/d7a51e44-9061-4ac8-9c76-37fd7685fe8a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Knowledge[[:space:]]Transfer[[:space:]]with[[:space:]]Simulated[[:space:]]Inter-Image[[:space:]]Erasing[[:space:]]for[[:space:]]Weakly[[:space:]]Supervised[[:space:]]Semantic[[:space:]]Segmentation/41d0edfc-3cb8-4f88-8e86-7449093dd76b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Knowledge-enhanced[[:space:]]Visual-Language[[:space:]]Pretraining[[:space:]]for[[:space:]]Computational[[:space:]]Pathology/a04fca37-34eb-47aa-8e9d-2ee68fcc1f47_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/L-DiffER_[[:space:]]Single[[:space:]]Image[[:space:]]Reflection[[:space:]]Removal[[:space:]]with[[:space:]]Language-based[[:space:]]Diffusion[[:space:]]Model/84990c91-4cd8-4a1b-83d4-8e97d608b20a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LAPT_[[:space:]]Label-driven[[:space:]]Automated[[:space:]]Prompt[[:space:]]Tuning[[:space:]]for[[:space:]]OOD[[:space:]]Detection[[:space:]]with[[:space:]]Vision-Language[[:space:]]Models/64854d06-c2d0-416b-8dda-0ced42b5873e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LASS3D_[[:space:]]Language-Assisted[[:space:]]Semi-Supervised[[:space:]]3D[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]with[[:space:]]Progressive[[:space:]]Unreliable[[:space:]]Data[[:space:]]Exploitation/c8ab9b1f-da9d-4178-ac58-4d667a9330b3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LATTE3D_[[:space:]]Large-scale[[:space:]]Amortized[[:space:]]Text-To-Enhanced3D[[:space:]]Synthesis/a662796f-9e3f-4f91-946c-d99cfdaedd9c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LCM-Lookahead[[:space:]]for[[:space:]]Encoder-based[[:space:]]Text-to-Image[[:space:]]Personalization/bbfbe553-d1d9-40ca-841b-fd60b8f4aefe_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LEGO_[[:space:]]Learning[[:space:]]EGOcentric[[:space:]]Action[[:space:]]Frame[[:space:]]Generation[[:space:]]via[[:space:]]Visual[[:space:]]Instruction[[:space:]]Tuning/13a5a5c4-078d-4d85-be9d-7966d6d6437e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LEIA_[[:space:]]Latent[[:space:]]View-invariant[[:space:]]Embeddings[[:space:]]for[[:space:]]Implicit[[:space:]]3D[[:space:]]Articulation/0e9336bc-d9b2-471b-94f5-34342577a604_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LEROjD_[[:space:]]Lidar[[:space:]]Extended[[:space:]]Radar-Only[[:space:]]Object[[:space:]]Detection/e0f1cd56-0b49-499c-b5fc-daef469190ce_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LG-Gaze_[[:space:]]Learning[[:space:]]Geometry-aware[[:space:]]Continuous[[:space:]]Prompts[[:space:]]for[[:space:]]Language-Guided[[:space:]]Gaze[[:space:]]Estimation/82982acf-4d4b-4f4d-a9b4-ee45c99a1bec_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LGM_[[:space:]]Large[[:space:]]Multi-View[[:space:]]Gaussian[[:space:]]Model[[:space:]]for[[:space:]]High-Resolution[[:space:]]3D[[:space:]]Content[[:space:]]Creation/269a8871-4800-4336-9c13-fbe2b1d2c82b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LHRS-Bot_[[:space:]]Empowering[[:space:]]Remote[[:space:]]Sensing[[:space:]]with[[:space:]]VGI-Enhanced[[:space:]]Large[[:space:]]Multimodal[[:space:]]Language[[:space:]]Model/2d33d60a-efea-43ed-9423-c305740d4cd0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LISO_[[:space:]]Lidar-only[[:space:]]Self-Supervised[[:space:]]3D[[:space:]]Object[[:space:]]Detection/d6e4374b-831f-4119-9bbf-2ac849c4dff4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LITA_[[:space:]]Language[[:space:]]Instructed[[:space:]]Temporal-Localization[[:space:]]Assistant/94963324-666f-4292-a883-31e4dd7adc61_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LLM[[:space:]]as[[:space:]]Copilot[[:space:]]for[[:space:]]Coarse-grained[[:space:]]Vision-and-Language[[:space:]]Navigation/5a4eb16e-327f-48aa-9238-afeb5ed269ac_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LLM[[:space:]]as[[:space:]]Dataset[[:space:]]Analyst_[[:space:]]Subpopulation[[:space:]]Structure[[:space:]]Discovery[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Model/cfcf5c50-f6b7-4af0-a369-8bdd05a30c95_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LLMCO4MR_[[:space:]]LLMs-aided[[:space:]]Neural[[:space:]]Combinatorial[[:space:]]Optimization[[:space:]]for[[:space:]]Ancient[[:space:]]Manuscript[[:space:]]Restoration[[:space:]]from[[:space:]]Fragments[[:space:]]with[[:space:]]Case[[:space:]]Studies[[:space:]]on[[:space:]]Dunhuang/33e2dc3a-2be0-4119-88f6-223087048685_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LLMGA_[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]based[[:space:]]Generation[[:space:]]Assistant/d91671aa-dc99-4c79-b4c9-9d98440c12f4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LLaMA-VID_[[:space:]]An[[:space:]]Image[[:space:]]is[[:space:]]Worth[[:space:]]2[[:space:]]Tokens[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/022c7268-b560-44a0-867d-3ec58c86bb1a_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2024/Infinite-ID_ Identity-preserved Personalization via ID-semantics Decoupling Paradigm/6b2c0220-7e0d-410c-b7de-7b3ff87a9f85_content_list.json b/2024/Infinite-ID_ Identity-preserved Personalization via ID-semantics Decoupling Paradigm/6b2c0220-7e0d-410c-b7de-7b3ff87a9f85_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5d226c249e6d6832d5d3c35164d88ec810f25d29 --- /dev/null +++ b/2024/Infinite-ID_ Identity-preserved Personalization via ID-semantics Decoupling Paradigm/6b2c0220-7e0d-410c-b7de-7b3ff87a9f85_content_list.json @@ -0,0 +1,1849 @@ +[ + { + "type": "text", + "text": "Infinite-ID: Identity-preserved Personalization via ID-semantics Decoupling Paradigm", + "text_level": 1, + "bbox": [ + 217, + 140, + 785, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yi Wu $^{1\\star}$ , Ziqiang Li $^{1,2\\star}$ , Heliang Zheng $^{1}$ , Chaoyue Wang $^{3\\star\\star}$ , and Bin Li $^{1\\star\\star}$", + "bbox": [ + 228, + 210, + 771, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Science and Technology of China, China", + "bbox": [ + 313, + 238, + 689, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Nanjing University of Information Science and Technology, China", + "bbox": [ + 274, + 253, + 725, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 The University of Sydney, Australia", + "bbox": [ + 372, + 267, + 630, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wuyi2021@mail.ustc.edu.cn, iceli@mail.ustc.edu.cn,", + "bbox": [ + 312, + 282, + 689, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhenghl@mail.ustc.edu.cn,", + "bbox": [ + 403, + 297, + 596, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chaoyue.wang@outlook.com, binli@ustc.edu.cn", + "bbox": [ + 338, + 311, + 663, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Drawing on recent advancements in diffusion models for text-to-image generation, identity-preserved personalization has made significant progress in accurately capturing specific identities with just a single reference image. However, existing methods primarily integrate reference images within the text embedding space, leading to a complex entanglement of image and text information, which poses challenges for preserving both identity fidelity and semantic consistency. To tackle this challenge, we propose Infinite-ID, an ID-semantics decoupling paradigm for identity-preserved personalization. Specifically, we introduce identity-enhanced training, incorporating an additional image cross-attention module to capture sufficient ID information while deactivating the original text cross-attention module of the diffusion model. This ensures that the image stream faithfully represents the identity provided by the reference image while mitigating interference from textual input. Additionally, we introduce a feature interaction mechanism that combines a mixed attention module with an AdaIN-mean operation to seamlessly merge the two streams. This mechanism not only enhances the fidelity of identity and semantic consistency but also enables convenient control over the styles of the generated images. Extensive experimental results on both raw photo generation and style image generation demonstrate the superior performance of our proposed method.", + "bbox": [ + 259, + 359, + 743, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Personalized Text-to-image Generation, Stable Diffusion, Identity-preserved Personalization", + "bbox": [ + 259, + 664, + 740, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 739, + 375, + 757 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human photo synthesis [19,38] has experienced notable advancements, particularly with the introduction of large text-to-image diffusion models such as Stable", + "bbox": [ + 212, + 772, + 787, + 803 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* First two authors contributed equally to this work.", + "bbox": [ + 215, + 810, + 573, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*\\* Corresponding Authors.", + "bbox": [ + 215, + 825, + 393, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4e341a16035d1d8bddf81feb9456875b5cb78d771d719f6b70dffc7434288e56.jpg", + "image_caption": [ + "Fig. 1: With just a single reference image, our Infinite-ID framework excels in synthesizing high-quality images while maintaining superior identity fidelity and text semantic consistency in various styles." + ], + "image_footnote": [], + "bbox": [ + 228, + 142, + 772, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion (SD) [27], Imagen [30], and DALL-E 3 [3]. Benefit from personalized text-to-image generation, recent researches focus on Identity-preserved personalization. This specialized area aims to produce highly customized photos that faithfully reflect a specific identity in novel scenes, actions, and styles, drawing inspiration from one or more reference images. This task has garnered considerable attention, leading to the development of numerous applications, including personalized AI portraits and virtual try-on scenarios. In the context of Identity-preserved personalization, the emphasis is placed on maintaining the invariance of human facial identity (ID), requiring a heightened level of detail and fidelity compared to more general styles or objects.", + "bbox": [ + 212, + 415, + 787, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent tuning-free methods exhibit promise for large-scale deployment, yet they face a notable challenge in balancing the trade-off between the fidelity of identity representation (ID fidelity) and the consistency of semantic understanding conveyed by the text prompt. This challenge arises due to the inherent entanglement of image and text information. Typically, tuning-free methods extract ID information from reference images and integrate it into the semantic information in two distinct ways. The first type, exemplified by PhotoMaker [17], incorporates text information with ID details in the text embedding space of the text encoder. While this merging approach aids in achieving semantic consistency, it compresses image features into the text embedding space, thereby weakening the ID information of the image and compromising identity fidelity. The second type, demonstrated by IP-Adapter [40], directly injects ID information into the U-Net of the diffusion model through an additional trainable cross-attention module. Although this approach aims to enhance the strength of ID information for improved fidelity, it tends to favor the image branch during training, consequently weakening the text branch and compromising semantic consistency. In summary, existing methods entangle image and text information, resulting in a significant trade-off between ID fidelity and semantic consistency.", + "bbox": [ + 212, + 568, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Yi Wu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the entanglement between image and text information, we propose Infinite-ID, an innovative approach to personalized text-to-image generation. Our method tackles the trade-off between maintaining high fidelity of identity and ensuring semantic consistency of the text prompt by implementing the ID-semantics decoupling paradigm. Specifically, we adopt identity-enhanced training that introduces an additional image cross-attention module to capture sufficient ID information and deactivate the original text cross-attention module to avoid text interference during training stage. Accordingly, our method can faithfully capture identity information from reference image, significantly improving the ID fidelity. Additionally, we employ a novel feature interaction mechanism that leverages a mixed attention module and an AdaIN-mean operation to effectively merge text information and identity information. Notably, our feature interaction mechanism not only preserves both identity and semantic details effectively but also enables convenient control over the styles of the generated images (as depicted in Fig. 1). Our contributions are summarized as:", + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel ID-semantics decoupling paradigm to resolve the entanglement between image and text information, acquiring a remarkable balance between ID fidelity and semantic consistency in Identity-preserved personalization.", + "- We propose a novel feature interaction mechanism incorporating a mixed attention module and an AdaIN-mean operation to effectively merge ID information and text information and also conveniently control the styles of the generated image in diffusion models.", + "- Experimental results demonstrate the excellent performance of our proposed method as compared with current state-of-the-art methods on both raw photo generation and style image generation." + ], + "bbox": [ + 225, + 382, + 784, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 571, + 395, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Text-to-image Diffusion Models", + "text_level": 1, + "bbox": [ + 215, + 603, + 522, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-image diffusion models, such as those explored in [3, 13, 27, 30, 43, 45], have garnered significant attention due to their impressive image generation capabilities. Current research endeavors aim to further enhance these models along multiple fronts, including the utilization of high-quality and large-scale datasets [18, 20, 32, 33], refinements to foundational architectures [3, 27, 30], and advancements in controllability [12, 29, 45]. Present iterations of text-to-image diffusion models typically follow a two-step process: first, encoding the text prompt using pre-trained text encoders such as CLIP [24] or T5 [26], and then utilizing the resulting text embedding as a condition for generating corresponding images through the diffusion process. Notably, the widely adopted Stable Diffusion model [27] distinguishes itself by executing the diffusion process in latent space instead of the original pixel space, leading to significant reductions in computation and time costs. An important extension to this framework is the Stable Diffusion XL (SDXL) [23], which enhances performance by scaling", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Infinite-ID", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "up the U-Net architecture and introducing an additional text encoder. Thus, our proposed method builds upon the SDXL. However, our method can also be extended to other text-to-image diffusion models.", + "bbox": [ + 212, + 146, + 784, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Identity-preserved personalization", + "text_level": 1, + "bbox": [ + 215, + 224, + 542, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Identity-preserved personalization aims to generate highly customized photos that accurately reflect a specific identity across various scenes, actions, and styles, drawing inspiration from one or more reference images. Initially, tuning-based methods, exemplified by DreamBooth [29], Textual Inversion [8] and Celeb-Basis [41], employ images of the same identity (ID) to fine-tune the model. While these methods yield results with high fidelity in preserving facial identity (ID), a significant drawback emerges: the customization of each ID necessitates a time investment of 10-30 minutes [15], consuming substantial computing resources and time. This limitation poses a significant obstacle to large-scale deployment in commercial applications. Consequently, recent advancements in tuning-free methods [6,7,9,16,17,21,35,36,39,40] have been introduced to streamline the generation process. These methods specifically leverage the construction of a vast amount of domain-specific data and the training of an encoder or hyper-network to represent input ID images as embeddings or LoRA weights within the model. Post-training, users need only input an image of the ID for customization, enabling personalized generation within seconds during the inference phase. These tuning-free methods typically contain two distinct manners.", + "bbox": [ + 212, + 261, + 787, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On one hand, methods [1, 17, 22, 39] incorporate text information alongside identity details within the text embedding space of the text encoder. For example, PhotoMaker [17] extracts identity embeddings from single or multiple reference images and merges them with corresponding class embeddings (e.g., \"man\" and \"woman\") in the text embedding space of the text encoder. While this stacking operation aids in achieving semantic consistency, it compresses image features into the text embedding space, leading to compromised identity fidelity. On the other hand, some studies [6, 36, 40] directly integrate identity information into the U-Net of the diffusion model. IP-Adapter [40] distinguishes itself by incorporating a additional cross-attention layer for each existing cross-attention layer within the original UNet model. This approach merges identity information with semantic details directly within the U-net but leads to distortion of the semantic space of the U-Net model. Consequently, this compromises the semantic consistency of the text prompt.", + "bbox": [ + 212, + 520, + 787, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, existing methods entangle the identity and text information, leading to a significant trade-off between ID fidelity and semantic consistency. To mitigate these limitations, we propose an identity-enhanced training to capture ID and text information separately. Moreover, we design an effective feature interaction mechanism leveraging a mixed attention module and an AdaIN-mean operation to preserve both identity and semantic details while also enabling convenient control over the styles of generated images.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Yi Wu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Attention Control in Diffusion model", + "text_level": 1, + "bbox": [ + 215, + 146, + 568, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous studies have investigated various attention control techniques within diffusion models. Hertz et al. [10] employed a shared attention mechanism, concatenating and applying an AdaIN module on the key and value between reference and synthesis images within the self-attention layer to ensure style-consistent image generation using a reference style. Cao et al. [5] utilized a mutual self-attention approach to achieve consistent image generation and non-rigid image editing, wherein the key and value of the synthesis image were replaced with those of the reference image within the self-attention layers of the diffusion model. Similarly, Shi et al. [34] proposed a method termed reference attention, enabling consistent multi-view generation of target objects by concatenating the key and value features between the condition signal and the synthesis image in the self-attention layers. Wang et al. [39] and Avrahami et al. [2] exploited attention maps within the cross-attention layers to guide the optimization process towards disentangling learned concepts in personalized generation tasks.", + "bbox": [ + 212, + 171, + 787, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 405, + 330, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 436, + 372, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stable Diffusion XL. Our method builds upon Stable Diffusion XL [23], comprising three core components: a Variational AutoEncoder (VAE) denoted as $\\xi(\\cdot)$ , a conditional U-Net [28] represented by $\\epsilon_{\\theta}(\\cdot)$ , and two pre-trained text encoders [25] denoted as $\\Theta_1(\\cdot)$ and $\\Theta_2(\\cdot)$ . Specifically, given a training image $x_0$ and its corresponding text prompt $T$ , the VAE encoder $\\xi$ transforms $x_0$ from its original space $R^{H\\times W\\times 3}$ to a compressed latent representation $z_0 = \\xi(x_0)$ , where $z_0 \\in R^{h\\times w\\times c}$ and $c$ denotes the latent dimension. Subsequently, the diffusion process operates within this compressed latent space to conserve computational resources and memory. Once the two text encoders process the text prompt $T$ into a text embedding $c = \\mathrm{Concat}(\\Theta_1(T), \\Theta_2(T))$ , the conditional U-Net $\\epsilon_{\\theta}$ predicts the noise $\\epsilon$ based on the current timestep $t$ , the $t$ -th latent representation $z_t$ , and the text embedding $c$ . The training objective is formulated as follows:", + "bbox": [ + 212, + 460, + 787, + 643 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {d i f f u s i o n}} = E _ {z _ {t}, t, c, \\epsilon \\in N (0, I)} [ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, c) | | _ {2} ^ {2} ]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 657, + 785, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attention mechanism in diffusion models. The fundamental unit of the stable diffusion model comprises a resblock, a self-attention layer, and a cross-attention layer. The attention mechanism is represented as follows:", + "bbox": [ + 212, + 680, + 787, + 726 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t n} (Q, K, V) = \\operatorname {S o f t m a x} (\\frac {Q K ^ {T}}{\\sqrt {d}}) V, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 736, + 785, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Q$ denotes the query feature projected from the spatial features generated by the preceding resblock, $K$ and $V$ represent the key and value features projected from the same spatial features as the query feature (in self-attention) or the text embedding extracted from the text prompt (in cross-attention).", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Infinite-ID", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bc8b81633e4218c09ab22f6c46b8e81ccab461952e9aab4409f34a539fb94514.jpg", + "image_caption": [ + "Fig. 2: Framework of ID-semantics Decoupling Paradigm. In the training phase, we adopt Face embeddings extractor to extract rich identity information and identity-enhanced training for faithfully representing the identity provided by the reference image while mitigating interference from textual input. In the inference stage, a mixed attention module is introduced to replace the original self-attention mechanism within the denoising U-Net model, facilitating the fusion of both identity and text information." + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 772, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Methodology", + "text_level": 1, + "bbox": [ + 215, + 523, + 370, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overview. In this section, we introduce our ID-semantics decoupling paradigm, as illustrated in Fig. 2, which effectively addresses the severe trade-off between high-fidelity identity and semantic consistency within identity-preserved personalization. Subsequently, we present our mixed attention mechanism, depicted in Fig. 3, designed to seamlessly integrate ID information and semantic information within the diffusion model during the inference stage. Additionally, we utilize an adaptive mean normalization (AdaIN-mean) operation to precisely align the style of the synthesized image with the desired style prompts.", + "bbox": [ + 212, + 551, + 787, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ID-semantics Decoupling Paradigm. To faithfully capture high-fidelity identity, we implement a novel identity-enhanced strategy during the training stage, as depicted in Fig. 2. Diverging from conventional methods [6,17,40] that utilize text-image pairs for training, we opt to exclude the text prompt input and deactivate cross-attention modules for text embeddings within the U-Net model. Instead, we establish a training pair consisting of an ID image, where the face is aligned to extract identity information, and denoising image ( $x_{t}$ in Fig. 2). Both the denoising image used for training and the ID image belong to the same individual, but they vary in factors such as viewpoints and facial expressions. This approach fosters a more comprehensive learning process [17]. We adopt Face embeddings extractor to accurately capture and leverage identity information from", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Yi Wu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/77312140525ed59cbd83583c6bab7affe65f8d203ba21ec5e32ac9b6e5dbcb74.jpg", + "image_caption": [ + "(a) Mixed Attention", + "Fig. 3: Mixed attention mechanism. On the left side, we employ mixed attention to fuse identity and text information. This involves concatenating their respective key and value features and subsequently applying mixed attention, where identity features are updated based on the concatenated key and value features. On the right side, for style merging, we introduce an additional AdaIN-mean operation (as depicted in Eq. 8) to the concatenated key and value features." + ], + "image_footnote": [], + "bbox": [ + 233, + 146, + 493, + 405 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0b4fd8e1db8bf40fe9c7f7a5c537cbda2c1af8d258d98ef0d3b1ffded6a26f95.jpg", + "image_caption": [ + "(b) Mixed Attention with AdaIN-mean" + ], + "image_footnote": [], + "bbox": [ + 509, + 146, + 769, + 405 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the input ID image. Additionally, to seamlessly integrate identity information into the denoising U-Net model, we introduce an extra trainable cross-attention mechanism ( $K_{id}'$ and $V_{id}'$ in Fig. 2) for image embeddings.", + "bbox": [ + 212, + 547, + 784, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Throughout the training phase, we exclusively optimize the parameters associated with the face mapper, CLIP mapper, and the image cross-attention module, while keeping the parameters of the pre-trained diffusion model fixed. The optimization loss closely resembles the original diffusion loss formulation (as delineated in Eq. 1), with the sole distinction being the shift from a text condition to an identity condition as the conditional input.", + "bbox": [ + 212, + 594, + 784, + 683 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {d i f f u s i o n}} = E _ {z _ {t}, t, c _ {i d}, \\epsilon \\in N (0, I)} [ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, c _ {i d}) | | _ {2} ^ {2} ], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 691, + 784, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the $c_{id}$ is the identity embeddings of the input ID image.", + "bbox": [ + 212, + 719, + 669, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During the inference phase, we align the desired face within the ID image to afford identity information. Following the approach adopted during the training phase, we utilize a face recognition backbone and a CLIP image encoder to extract identity features from the aligned face image. Leveraging the trained face mapper, clip mapper, and image cross-attention mechanisms, these identity features are seamlessly integrated into the denoising U-Net model. Subsequently, we compute the key and value features for self-attention in the original stable", + "bbox": [ + 212, + 734, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Infinite-ID", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "diffusion model, considering only the text prompt input. These text key and value features are instrumental in the mixed attention process (illustrated in Fig. 3), facilitating the fusion of text and identity information. Moreover, to further augment the text information during the denoising process, we incorporate the original text cross-attention mechanism, integrating the resulting text hidden states with the output image hidden states obtained from the image cross-attention module.", + "bbox": [ + 212, + 146, + 787, + 251 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Face Embeddings Extractor. We adopt a multifaceted approach by incorporating pre-trained models to extract facial features. Firstly, following the methodologies of prior research, we utilize a pre-trained CLIP image encoder as one of our facial feature extractors. Specifically, we leverage the local embeddings, comprising the last hidden states obtained from the CLIP image encoder, forming a sequence of embeddings with a length of N (N=257 in our implementation). Subsequently, we employ a CLIP mapper to project these image embeddings (with a dimensionality of 1664) to the same dimension as the text features in the pre-trained diffusion model. As elucidated in [40], the features extracted by the CLIP image encoder are instrumental in capturing the structural information pertinent to the identity face within identity-preserved personalization tasks. Additionally, we leverage the backbone of a face recognition model as another facial feature extractor. As highlighted in [40], features extracted by the face recognition backbone are adept at capturing the characteristics associated with human facial features within identity-preserved personalization tasks. More specifically, we utilize the global image embedding derived from the extracted features and subsequently employ a face mapper to align the dimensionality (512 dimensions) of the extracted global image embedding with the dimensionality of the text features in the pre-trained diffusion model.", + "bbox": [ + 212, + 252, + 787, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, the identity embeddings $c_{id}$ corresponding to the ID image $x$ can be expressed as:", + "bbox": [ + 212, + 541, + 789, + 571 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nc _ {i d} = \\operatorname {C o n c a t} \\left(\\mathrm {M} _ {\\text {c l i p}} \\left(\\mathrm {E} _ {\\text {c l i p}} (\\mathrm {F A} (x))\\right), \\mathrm {M} _ {\\text {f a c e}} \\left(\\mathrm {E} _ {\\text {f a c e}} (\\mathrm {F A} (x))\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 587, + 785, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathrm{Concat}(\\cdot, \\cdot)$ , $\\mathrm{M}_{\\mathrm{clip}}(\\cdot)$ , $\\mathrm{E}_{\\mathrm{clip}}(\\cdot)$ , $\\mathrm{M}_{\\mathrm{face}}(\\cdot)$ , $\\mathrm{E}_{\\mathrm{face}}(\\cdot)$ , and $\\mathrm{FA}(\\cdot)$ denote the concatenation function, CLIP mapper, CLIP image encoder, face mapper, face recognition backbone, and face alignment module [44], respectively.", + "bbox": [ + 212, + 627, + 787, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mixed Attention Mechanism. As explored in previous studies [5, 14, 37], the features in the self attention layers play a crucial role in consistency image generation (across-frame in text-to-video works), which indicates these features provide a refined and detailed semantics information. In our study, we extract features from the self-attention layers of the original text-to-image diffusion model to capture rich semantic information, represented as $K_{t}$ and $V_{t}$ . We enhance the self-attention mechanism by incorporating it into a mixed attention framework, depicted in Fig. 3 (a). This fusion enables the integration of semantic features $(K_{t}$ and $V_{t}$ ) with the identity-based features $(K_{id}$ and $V_{id}$ ), thereby encapsulating identity information. Through this integration, the mixed attention mechanism seamlessly merges semantic details into the generated features", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Yi Wu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "across different resolutions. The formulation of the mixed attention mechanism is as follows:", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t n} _ {\\operatorname {m i x}} (Q, K, V) \\triangleq \\operatorname {A t t n} (Q, \\hat {K}, \\hat {V})\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 183, + 611, + 199 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l l} \\text {w . r . t} & \\hat {K} = \\underset {i d} {\\operatorname {C o n c a t}} \\left(K _ {i d}, K _ {t}\\right), \\quad \\hat {V} = \\underset {t} {\\operatorname {C o n c a t}} \\left(V _ {i d}, V _ {t}\\right), \\end{array} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 204, + 784, + 229 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nK _ {i d} = W _ {k} ^ {i d} Z _ {i d}, \\quad K _ {t} = W _ {k} ^ {t} Z _ {t},\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 223, + 575, + 242 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nV _ {i d} = W _ {v} ^ {i d} Z _ {i d}, \\quad V _ {t} = W _ {v} ^ {t} Z _ {t},\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 246, + 566, + 262 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $Z_{id}$ and $Z_{t}$ are the corresponding spatial features of generated features and semantic features, respectively. The parameters $W_{k}^{id}$ , $W_{k}^{t}$ , $W_{v}^{id}$ , and $W_{v}^{t}$ correspond to the weights of the corresponding fully connected layers.", + "bbox": [ + 212, + 270, + 784, + 315 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Cross-attention Merging. To further refine semantic control, we incorporate text features into the identity feature within the cross-attention layers using the following formulation:", + "bbox": [ + 212, + 316, + 782, + 359 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t n} _ {\\text {c r o s s}} (Q, K, V) \\triangleq \\operatorname {A t t n} \\left(Q, K _ {i d} ^ {\\prime}, V _ {i d} ^ {\\prime}\\right) + \\operatorname {A t t n} \\left(Q, K _ {t} ^ {\\prime}, V _ {t} ^ {\\prime}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 297, + 369, + 784, + 386 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $K_{id}^{\\prime} = \\hat{W}_{k}^{id}c_{id}$ , $V_{id}^{\\prime} = \\hat{W}_{v}^{id}c_{id}$ , $K_{t}^{\\prime} = \\hat{W}_{k}^{t}c_{t}$ , and $V_{t}^{\\prime} = \\hat{W}_{v}^{t}c_{t}$ . $c_{id}$ and $c_{t}$ are the identity embedding and text embedding, respectively. $\\hat{W}_{k}^{id}$ , $\\hat{W}_{v}^{id}$ , $\\hat{W}_{k}^{t}$ , and $\\hat{W}_{v}^{t}$ correspond to the weights of the trainable fully connected layers within cross-attention module.", + "bbox": [ + 212, + 393, + 784, + 455 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Style Information Merging. Inspired by [10], we propose an adaptive mean normalization (AdaIN-mean) operation to further align the style of the synthesis image with the style prompts. Concretely, we align the key and value features projected from identity features in both mixed attention and cross-attention with the key and value features projected from text features, formulated as follows:", + "bbox": [ + 212, + 458, + 784, + 532 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l l l} K _ {i d} = \\operatorname {A d a I N - m} \\left(K _ {i d}, K _ {t}\\right), & V _ {i d} = \\operatorname {A d a I N - m} \\left(V _ {i d}, V _ {t}\\right), & \\text {F o r M i x e d A t t e n t i o n} \\\\ K _ {i d} ^ {\\prime} = \\operatorname {A d a I N - m} \\left(K _ {i d} ^ {\\prime}, K _ {t} ^ {\\prime}\\right), & V _ {i d} ^ {\\prime} = \\operatorname {A d a I N - m} \\left(V _ {i d} ^ {\\prime}, V _ {t} ^ {\\prime}\\right), & \\text {F o r C r o s s A t t e n t i o n} \\end{array} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 541, + 784, + 590 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where the AdaIN-mean operation (AdaIN-m(\\cdot)) is defined as:", + "bbox": [ + 214, + 590, + 656, + 606 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A d a I N} - \\mathrm {m} (x, y) = x - \\mu (x) + \\mu (y), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 616, + 784, + 631 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mu (x)\\in R^{d_k}$ is the mean of key and value features across different pixels. The mixed attention with AdaIN-mean has been illustrated in Fig. 3 (b).", + "bbox": [ + 212, + 638, + 782, + 669 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 690, + 375, + 707 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "After outlining the experimental setup in Sec. 4.1, we conduct a comparative analysis of raw photo generation and style image generation in Sec. 4.2. Ablation studies, highlighting the significance of various components, are presented in Sec. 4.3. Additionally, experiments involving multiple input ID images including identity mixing and identity linear interpolation are detailed in Sec. A.7 of the supplementary materials. For further insights, Sec. A.8 and Sec. A.9 of the supplementary materials provide additional qualitative results for raw photo generation and style photo generation, respectively.", + "bbox": [ + 212, + 719, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Infinite-ID", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ce66ee3a3d30418f47ba6bf5dc8d7764ec132a6e50227e46137dfc4dbb4ba01d.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparison on raw photo generation. The results demonstrate that our Infinite-ID consistently maintains identity fidelity and achieves high-quality semantic consistency with just a single input image." + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 781, + 460 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 544, + 426, + 560 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details. Our experiments leverage a pre-trained Stable Diffusion XL (SDXL) model [23]. For image encoding, we utilize the OpenCLIP ViT-H/14 [42] and the backbone of ArcFace [11]. The SDXL model consists of 70 cross-attention layers, to each of which we append an additional image cross-attention module. Training is conducted on 16 A100 GPUs for 1 million steps, with a batch size of 4 per GPU. We employ the AdamW optimizer with a fixed learning rate of 1e-4 and weight decay set to 0.01. During inference, we employ the DDIM Sampler with 30 steps and guidance scale is set to 5.0. Training data are sourced from multiple datasets, including the LAION-2B dataset [33], the LAION-Face dataset [46], and few high-quality images collected from the Websites4. We curate a dataset where each individual is represented by multiple photographs.", + "bbox": [ + 212, + 571, + 787, + 752 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation. We assess the efficacy of our approach in preserving both identity fidelity and semantic consistency. Specifically, we measure identity fidelity, utilizing metrics such as $M_{\\mathrm{FaceNet}}$ (measured by FaceNet [31]) and CLIP-I [8]. Identity fidelity is evaluated based on the similarity of detected faces between the", + "bbox": [ + 212, + 753, + 787, + 816 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Yi Wu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "4 https://www.pexels.com, https://pixabay.com.", + "bbox": [ + 217, + 824, + 545, + 839 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/01d499f8c7db4e1c3c67dccca8ae2f10d8c2b4af6eeadea587c0d1c24f4ba24e.jpg", + "image_caption": [ + "Fig. 5: Qualitative comparison on style image generation. The results demonstrate that our method maintains strong identity fidelity, high-quality semantic consistency, and precise stylization using only a single reference image." + ], + "image_footnote": [], + "bbox": [ + 223, + 147, + 781, + 453 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "reference image and generated images. Semantic consistency is quantified using CLIP text-image consistency (CLIP-T [24]), which compares the text prompt with the corresponding generated images. More details of evaluation are shown in Sec. A.1 and Sec. A.2 of the supplementary materials.", + "bbox": [ + 212, + 537, + 787, + 599 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Comparison to Previous Methods", + "text_level": 1, + "bbox": [ + 214, + 619, + 539, + 635 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Raw Photo Generation. We benchmark our Infinite-ID against established identity-preserving personalization approaches. The qualitative outcomes are illustrated in Fig. 4, while quantitative results are provided in Table 1. Notably, all methods are tuning-free, necessitating no test-time adjustments. FastComposer [39] exhibits challenges in maintaining identity fidelity, often presenting undesired artifacts in synthesized images. While IP-Adapter [40] and IP-Adapter-Face [40] demonstrates relatively fewer artifacts, its semantic consistency falls short. This phenomenon arises from the direct fusion of identity information with semantic details within the U-net model, leading to a compromise in semantic consistency. In contrast, PhotoMaker [17] exhibits commendable semantic consistency but falls short in preserving identity fidelity. Leveraging our ID-semantics decoupling paradigm, our method excels in preserving identity fidelity. Furthermore, our mixed attention mechanism effectively integrate the semantic", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Infinite-ID", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/2b75eef9a25fc951129ca904fc62075ac9b77aa18a1485e1c02e9c17b8452f5a.jpg", + "table_caption": [ + "Table 1: Quantitative comparison. The evaluation metrics encompass CLIP-T, CLIP-I, and $M_{\\text{FaceNet}}$ . Our approach outperforms other methods in terms of identity fidelity while simultaneously achieving satisfactory semantic consistency. The best result is shown in bold, and the second best is underlined. Additionally, the quantitative comparison of ablation studies have been shown in the gray part." + ], + "table_footnote": [], + "table_body": "
CLIP-T ↑CLIP-I ↑M_FaceNet ↑
FastComposer [39]0.2920.8870.556
IP-Adapter [40]0.2740.9050.474
IP-Adapter-Face [40]0.3130.9190.513
PhotoMaker [17]0.3430.8140.502
Ours0.3400.9130.689
Ours (w/o identity-enhanced training)0.3290.8910.593
Ours (w/o Mixed Attention)0.3310.9050.700
Ours (Mixed Attention ⇒ Mutual Attention)0.3160.8080.398
", + "bbox": [ + 240, + 226, + 758, + 359 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "information into the denoising process, positioning our method favorably against existing techniques. More comparison results are shown in Sec. A.4, Sec. A.5 and Sec. A.6 of the supplementary materials.", + "bbox": [ + 212, + 388, + 782, + 434 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Style Image Generation. We demonstrate the results of the stylization results in Fig. 5, which compare our method with state-of-the-art tuning-free identity-preserving personalization methods. The style of the synthesis images include anime style, comic book style, and line art style. According to the stylization results, the IP-Adapter [40] and IP-Adapter-Face fails to depict the desired style in the prompt and the style of generation results always obey the tone in the reference image. The training pipeline of the IP-Adapter and IP-Adapter-Face entangles the text embedding and image embedding which leads to the distortion of the text embeddings space. FastComposer [39] also fails in stylization generation and shows undesired artifacts. PhotoMaker [17] achieves a better semantic consistency and stylization, but the identity fidelity is still unsatisfactory. In contrast, our method achieves high identity fidelity, appealing semantic consistency, and precise stylization meantime.", + "bbox": [ + 212, + 435, + 787, + 632 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 657, + 388, + 674 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we begin by conducting ablation studies to assess the influence of our identity-enhanced training and mixed attention mechanism. Furthermore, we conduct style image generation to evaluate the effectiveness of our AdaIN-mean operation in regulating the style of the generated images. More ablation studies are demonstrated in Sec. A.3 of the supplementary materials.", + "bbox": [ + 212, + 688, + 785, + 763 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation of identity-enhanced training. In Fig. 6 and Table 1, we compare our method with \"Ours (w/o identity-enhanced training)\", which is implemented using an identity-semantics entangled training strategy. This strategy utilizes text-image pairs during training and activates cross-attention modules for text embeddings within the original U-Net model. It is noteworthy that both methods", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Yi Wu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2337fa1f5f28299abc98a7e44b4eb7e3ab6aca5208d2f2f8925f440af4c5d3ad.jpg", + "image_caption": [], + "image_footnote": [ + "a woman wearing purple sweater, facing and smiling to the camera, in a forest" + ], + "bbox": [ + 292, + 143, + 364, + 200 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8bcc8b60ad7cba6d6f482b5c449c104eb748bcb263cf27be832b956c32954ca2.jpg", + "image_caption": [], + "image_footnote": [ + "a photo of a man wearing a colorful shirt, in the living room" + ], + "bbox": [ + 292, + 200, + 364, + 253 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ab5fdf84a765b7d1e7e633e134b54e7ef4a12f2ec0abe906dd12b8c51b8bc05e.jpg", + "image_caption": [ + "ID image" + ], + "image_footnote": [], + "bbox": [ + 292, + 253, + 354, + 309 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/10a0c5816810dbe015c626d4041327b6f61c4f147cc8463e6962a9f478caf51c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 143, + 496, + 199 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/df1a192a31f1f6bad742e8cdef1e7990cb0235d6f9f194f3ec1024d796f1fdb3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 143, + 637, + 199 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c9720f8d102dd20dd49713e63b974affa08146157b85f3616678df783acddf2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 199, + 637, + 253 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/939905821752ac9b1156d38a0bb51381a3134c39e960af462d2ae55427e69975.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [ + "a woman wearing a red hat, smiling and facing to the camera, sunset, on the beach", + "Ours (w/o identity-enhanced training)" + ], + "bbox": [ + 424, + 253, + 635, + 309 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3b164df0b3cddb42c19fe4e4042690a18573786ea330f15b130117cb7d17a85d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 143, + 707, + 199 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f0297b2b1e480601e5270170046b85ebda0b05559abe54dc644e74df6863dc73.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 199, + 707, + 253 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/49a3475032c54d73935d3f51366f60732f52fbc8699a4b055164cc5ed54c7923.jpg", + "image_caption": [ + "Ours (M-A $\\Rightarrow$ MU-A)" + ], + "image_footnote": [], + "bbox": [ + 637, + 253, + 707, + 309 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8fbf6616ec2483404a4b7d2399eee2847d481cebb3030e2bba8de7834fe9a83c.jpg", + "image_caption": [ + "Fig. 6: Ablation study of our identity-enhanced training and mixed attention (M-A). It is evident that identity-enhanced training significantly improves the identity fidelity, and mixed-attention mechanism enhances semantic consistency compared to mutual attention (MU-A) approach [5]." + ], + "image_footnote": [ + "awearing a blue clothing, facing and smiling to the snow. A snowy, snowy, in an anime style" + ], + "bbox": [ + 310, + 417, + 395, + 484 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9a1c3b99492fc94af69f9a0838f250c115911615070ac6db8a20e09586ce23e7.jpg", + "image_caption": [ + "ID Image", + "Fig. 7: Ablation study of our AdaIN-mean operation. The results show that AdaIN-mean play a crucial role in style image generation. Compare to AdaIN [10] module, our AdaIN-mean helps to achieve a higher identity fidelity." + ], + "image_footnote": [ + "a woman wearing a purple dress and smiling to the camera, on a beach, sunset, comic book style" + ], + "bbox": [ + 310, + 484, + 395, + 547 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8d1c57c74125779832a2b402dbf91d09755bd61eca819c7b94c846fadd7592a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 417, + 607, + 483 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e46e5ccadcd2615e9b1fe5dadd204d460988396d27a7e752dc92d9cebc2e94e5.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 442, + 483, + 607, + 547 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0bcf974314745e99ea2c32b5ec3637174b3d274b3dd1ec56afad195a9dfe6e6f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 417, + 692, + 483 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/22a87f3a44b42531fb9abd58dadcbe61d4125c02c416795ca24e670f5f0ef109.jpg", + "image_caption": [ + "Ours (AdaIN-mean $\\Rightarrow$ AdaIN)" + ], + "image_footnote": [], + "bbox": [ + 609, + 483, + 691, + 547 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "share the same inference processing. The qualitative comparison demonstrates that our identity-enhanced training notably enhances identity fidelity.", + "bbox": [ + 212, + 642, + 785, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation of mixed attention mechanism. To assess the effectiveness of our proposed mixed attention (M-A) mechanism, we compare our method with \"Ours (w/o M-A)\" and \"Ours (M-A $\\Rightarrow$ MU-A)\" in Fig. 6 and Table 1. Mutual self-attention (MU-A) [5] converts the existing self-attention into 'cross-attention' for consistent image editing, where the crossing operation happens in the self-attention of two related diffusion processes. The results show that our mixed attention mechanism demonstrates superior ability to improve semantic consistency while maintaining identity fidelity.", + "bbox": [ + 212, + 672, + 787, + 792 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation of AdaIN-mean operation. To assess the effectiveness of our proposed AdaIN-mean operation, we compare our Infinite-ID model with variations, namely Ours (w/o AdaIN-mean) and Ours (AdaIN-mean $\\Rightarrow$ AdaIN), as depicted", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Infinite-ID", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "in Fig. 7. The results reveal that: i) Ours (w/o AdaIN-mean) exhibits superior ID fidelity compared to Infinite-ID but fails to achieve style consistency with the text prompt; ii) Both our AdaIN-mean and AdaIN modules successfully achieve style consistency with the text prompt, yet AdaIN-mean maintains better ID fidelity than AdaIN. In conclusion, our proposed AdaIN-mean operation facilitates precise stylization while concurrently preserving ID fidelity.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion and limitations.", + "text_level": 1, + "bbox": [ + 215, + 261, + 516, + 277 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce Infinite-ID, an innovative identity-preserved personalization method designed to meet the requirement of identity (ID) fidelity and semantic consistency of the text prompt, all achievable with just one reference image and completed within seconds. Infinite-ID comprises three key components: the identity-enhanced training, mixed attention mechanism, and adaptive mean normalization (AdaIN-mean). Through extensive experimentation, our results illustrate that Infinite-ID outperforms baseline methods, delivering strong ID fidelity, superior generation quality, and precise semantic consistency in both raw photo generation and style image generation tasks. However, it's important to note that our method lacks multi-object personalization capability. Moreover, artifacts may occur when the human face occupies only a small portion of the entire image, attributed to limitations inherent in the original diffusion model. Furthermore, the image encoders employed in the identity-preserved personalization remain to be further explored.", + "bbox": [ + 212, + 295, + 787, + 507 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Copyright of Training Data and Ethical Considerations.", + "text_level": 1, + "bbox": [ + 215, + 531, + 784, + 549 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Copyright of Training Data. We construct our training dataset from the open-source public image dataset, including LAION-2B dataset, the LAION-Face dataset and few high-quality images collected from the Website5 which supports the CC0 (Creative Commons Zero) 1.0 Universal license6.", + "bbox": [ + 212, + 564, + 787, + 625 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ethical Considerations. We recognize the ethical considerations about improper use of such high-quality human photo synthesis techniques. Therefore, we emphasize the significance of establishing and complying to ethical guidelines [4] and using this technology responsibly. We hope that our work will drive further discussions and research on the security and ethical use of the human photo generation in computer vision. Moreover, we will employ the watermark and safety checker (NSFW classifiers) to attempt to avoid inappropriate content generation.", + "bbox": [ + 212, + 625, + 787, + 744 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Yi Wu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "page_footnote", + "text": "$^{5}$ https://www.pexels.com, https://pixabay.com.", + "bbox": [ + 217, + 753, + 545, + 770 + ], + "page_idx": 13 + }, + { + "type": "page_footnote", + "text": "6 The person who associated a work with this deed has dedicated the work to the public domain by waiving all of his or her rights to the work worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law. You can copy, modify, distribute and perform the work, even for commercial purposes, all without asking permission.", + "bbox": [ + 220, + 771, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research was funded by the National Natural Science Foundation of China (U19B2044), supported by the Fundamental Research Funds for the Central Universities (WK2100000045).", + "bbox": [ + 215, + 172, + 785, + 219 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 239, + 321, + 255 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achlioptas, P., Benetatos, A., Fostiropoulos, I., Skourtis, D.: Stellar: Systematic evaluation of human-centric personalized text-to-image methods. arXiv preprint arXiv:2312.06116 (2023)", + "2. Avrahami, O., Aberman, K., Fried, O., Cohen-Or, D., Lischinski, D.: Break-a-scene: Extracting multiple concepts from a single image. arXiv preprint arXiv:2305.16311 (2023)", + "3. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf 2, 3 (2023)", + "4. Biden, J.R.: Executive order on the safe, secure, and trustworthy development and use of artificial intelligence (2023)", + "5. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. arXiv preprint arXiv:2304.08465 (2023)", + "6. Chen, L., Zhao, M., Liu, Y., Ding, M., Song, Y., Wang, S., Wang, X., Yang, H., Liu, J., Du, K., et al.: Photovverse: Tuning-free image customization with text-to-image diffusion models. arXiv preprint arXiv:2309.05793 (2023)", + "7. Chen, X., Huang, L., Liu, Y., Shen, Y., Zhao, D., Zhao, H.: Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481 (2023)", + "8. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618 (2022)", + "9. Gal, R., Arar, M., Atzmon, Y., Bermano, A.H., Chechik, G., Cohen-Or, D.: Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG) 42(4), 1-13 (2023)", + "0. Hertz, A., Voynov, A., Fruchter, S., Cohen-Or, D.: Style aligned image generation via shared attention. arXiv preprint arXiv:2312.02133 (2023)", + "1. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: 2017 IEEE International Conference on Computer Vision (ICCV) (Oct 2017). https://doi.org/10.1109/iccv.2017.167, http://dx.doi.org/10.1109/iccv.2017.167", + "2. Inoue, N., Kikuchi, K., Simo-Serra, E., Otani, M., Yamaguchi, K.: Layoutdm: Discrete diffusion model for controllable layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10167-10176 (2023)", + "3. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)", + "4. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439 (2023)" + ], + "bbox": [ + 225, + 266, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Infinite-ID", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1931-1941 (2023)", + "16. Li, D., Li, J., Hoi, S.C.: Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720 (2023)", + "17. Li, Z., Cao, M., Wang, X., Qi, Z., Cheng, M.M., Shan, Y.: Photomaker: Customizing realistic human photos via stacked id embedding. arXiv preprint arXiv:2312.04461 (2023)", + "18. Li, Z., Usman, M., Tao, R., Xia, P., Wang, C., Chen, H., Li, B.: A systematic survey of regularization and normalization in gans. ACM Computing Surveys 55(11), 1-37 (2023)", + "19. Li, Z., Wang, C., Zheng, H., Zhang, J., Li, B.: Fakeclr: Exploring contrastive learning for solving latent discontinuity in data-efficient gans. In: European Conference on Computer Vision. pp. 598-615. Springer (2022)", + "20. Li, Z., Xia, P., Tao, R., Niu, H., Li, B.: A new perspective on stabilizing gans training: Direct adversarial training. IEEE Transactions on Emerging Topics in Computational Intelligence 7(1), 178-189 (2022)", + "21. Ma, J., Liang, J., Chen, C., Lu, H.: Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. arXiv preprint arXiv:2307.11410 (2023)", + "22. Peng, X., Zhu, J., Jiang, B., Tai, Y., Luo, D., Zhang, J., Lin, W., Jin, T., Wang, C., Ji, R.: Portrait booth: A versatile portrait model for fast identity-preserved personalization. arXiv preprint arXiv:2312.06354 (2023)", + "23. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)", + "24. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "25. Radford, A., Kim, J., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Amanda, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. Cornell University - arXiv, Cornell University - arXiv (Feb 2021)", + "26. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)", + "27. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "28. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)", + "29. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Yi Wu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)", + "31. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015)", + "32. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)", + "33. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)", + "34. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110 (2023)", + "35. Wang, Q., Bai, X., Wang, H., Qin, Z., Chen, A.: Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519 (2024)", + "36. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848 (2023)", + "37. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7623-7633 (2023)", + "38. Wu, Y., Li, Z., Wang, C., Zheng, H., Zhao, S., Li, B., Tao, D.: Domain remodulation for few-shot generative domain adaptation. Advances in Neural Information Processing Systems 36 (2024)", + "39. Xiao, G., Yin, T., Freeman, W.T., Durand, F., Han, S.: Fastcomposer: Tuning-free multi-subject image generation with localized attention. arXiv preprint arXiv:2305.10431 (2023)", + "40. Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721 (2023)", + "41. Yuan, G., Cun, X., Zhang, Y., Li, M., Qi, C., Wang, X., Shan, Y., Zheng, H.: Inserting anybody in diffusion models via celeb basis. In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 72958-72982. Curran Associates, Inc. (2023), https://proceedings.neurips.cc/paper_files/paper/2023/file/e6d37cc5723e810b793c834bc6647cf-Paper-Conference.pdf", + "42. Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djolonga, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., et al.: A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867 (2019)", + "43. Zhang, C., Zhang, C., Zhang, M., Kweon, I.S.: Text-to-image diffusion model in generative ai: A survey. arXiv preprint arXiv:2303.07909 (2023)", + "44. Zhang, K., Zhang, Z., Li, Z., Qiao, Y.: Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters 23(10), 1499-1503 (2016)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Infinite-ID", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "45. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023)", + "46. Zheng, Y., Yang, H., Zhang, T., Bao, J., Chen, D., Huang, Y., Yuan, L., Chen, D., Zeng, M., Wen, F.: General facial representation learning in a visual-linguistic manner. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18697-18709 (2022)" + ], + "bbox": [ + 217, + 146, + 785, + 244 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Yi Wu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Infinite-ID_ Identity-preserved Personalization via ID-semantics Decoupling Paradigm/6b2c0220-7e0d-410c-b7de-7b3ff87a9f85_model.json b/2024/Infinite-ID_ Identity-preserved Personalization via ID-semantics Decoupling Paradigm/6b2c0220-7e0d-410c-b7de-7b3ff87a9f85_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c66e489f8633eeef2f148726fbad1923514c642d --- /dev/null +++ b/2024/Infinite-ID_ Identity-preserved Personalization via ID-semantics Decoupling Paradigm/6b2c0220-7e0d-410c-b7de-7b3ff87a9f85_model.json @@ -0,0 +1,2513 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.787, + 0.187 + ], + "angle": 0, + "content": "Infinite-ID: Identity-preserved Personalization via ID-semantics Decoupling Paradigm" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.212, + 0.772, + 0.229 + ], + "angle": 0, + "content": "Yi Wu\\(^{1\\star}\\), Ziqiang Li\\(^{1,2\\star}\\), Heliang Zheng\\(^{1}\\), Chaoyue Wang\\(^{3\\star\\star}\\), and Bin Li\\(^{1\\star\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.314, + 0.239, + 0.69, + 0.254 + ], + "angle": 0, + "content": "1 University of Science and Technology of China, China" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.254, + 0.726, + 0.269 + ], + "angle": 0, + "content": "\\(^{2}\\) Nanjing University of Information Science and Technology, China" + }, + { + "type": "text", + "bbox": [ + 0.374, + 0.268, + 0.631, + 0.282 + ], + "angle": 0, + "content": "3 The University of Sydney, Australia" + }, + { + "type": "text", + "bbox": [ + 0.313, + 0.284, + 0.691, + 0.296 + ], + "angle": 0, + "content": "wuyi2021@mail.ustc.edu.cn, iceli@mail.ustc.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.405, + 0.298, + 0.598, + 0.31 + ], + "angle": 0, + "content": "zhenghl@mail.ustc.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.339, + 0.312, + 0.665, + 0.324 + ], + "angle": 0, + "content": "chaoyue.wang@outlook.com, binli@ustc.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.361, + 0.744, + 0.653 + ], + "angle": 0, + "content": "Abstract. Drawing on recent advancements in diffusion models for text-to-image generation, identity-preserved personalization has made significant progress in accurately capturing specific identities with just a single reference image. However, existing methods primarily integrate reference images within the text embedding space, leading to a complex entanglement of image and text information, which poses challenges for preserving both identity fidelity and semantic consistency. To tackle this challenge, we propose Infinite-ID, an ID-semantics decoupling paradigm for identity-preserved personalization. Specifically, we introduce identity-enhanced training, incorporating an additional image cross-attention module to capture sufficient ID information while deactivating the original text cross-attention module of the diffusion model. This ensures that the image stream faithfully represents the identity provided by the reference image while mitigating interference from textual input. Additionally, we introduce a feature interaction mechanism that combines a mixed attention module with an AdaIN-mean operation to seamlessly merge the two streams. This mechanism not only enhances the fidelity of identity and semantic consistency but also enables convenient control over the styles of the generated images. Extensive experimental results on both raw photo generation and style image generation demonstrate the superior performance of our proposed method." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.665, + 0.741, + 0.694 + ], + "angle": 0, + "content": "Keywords: Personalized Text-to-image Generation, Stable Diffusion, Identity-preserved Personalization" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.741, + 0.377, + 0.758 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.773, + 0.788, + 0.804 + ], + "angle": 0, + "content": "Human photo synthesis [19,38] has experienced notable advancements, particularly with the introduction of large text-to-image diffusion models such as Stable" + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.811, + 0.574, + 0.826 + ], + "angle": 0, + "content": "* First two authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.826, + 0.395, + 0.84 + ], + "angle": 0, + "content": "\\*\\* Corresponding Authors." + }, + { + "type": "list", + "bbox": [ + 0.216, + 0.811, + 0.574, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Yi Wu et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.143, + 0.773, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Fig. 1: With just a single reference image, our Infinite-ID framework excels in synthesizing high-quality images while maintaining superior identity fidelity and text semantic consistency in various styles." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.788, + 0.569 + ], + "angle": 0, + "content": "Diffusion (SD) [27], Imagen [30], and DALL-E 3 [3]. Benefit from personalized text-to-image generation, recent researches focus on Identity-preserved personalization. This specialized area aims to produce highly customized photos that faithfully reflect a specific identity in novel scenes, actions, and styles, drawing inspiration from one or more reference images. This task has garnered considerable attention, leading to the development of numerous applications, including personalized AI portraits and virtual try-on scenarios. In the context of Identity-preserved personalization, the emphasis is placed on maintaining the invariance of human facial identity (ID), requiring a heightened level of detail and fidelity compared to more general styles or objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Recent tuning-free methods exhibit promise for large-scale deployment, yet they face a notable challenge in balancing the trade-off between the fidelity of identity representation (ID fidelity) and the consistency of semantic understanding conveyed by the text prompt. This challenge arises due to the inherent entanglement of image and text information. Typically, tuning-free methods extract ID information from reference images and integrate it into the semantic information in two distinct ways. The first type, exemplified by PhotoMaker [17], incorporates text information with ID details in the text embedding space of the text encoder. While this merging approach aids in achieving semantic consistency, it compresses image features into the text embedding space, thereby weakening the ID information of the image and compromising identity fidelity. The second type, demonstrated by IP-Adapter [40], directly injects ID information into the U-Net of the diffusion model through an additional trainable cross-attention module. Although this approach aims to enhance the strength of ID information for improved fidelity, it tends to favor the image branch during training, consequently weakening the text branch and compromising semantic consistency. In summary, existing methods entangle image and text information, resulting in a significant trade-off between ID fidelity and semantic consistency." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Infinite-ID" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": "To address the entanglement between image and text information, we propose Infinite-ID, an innovative approach to personalized text-to-image generation. Our method tackles the trade-off between maintaining high fidelity of identity and ensuring semantic consistency of the text prompt by implementing the ID-semantics decoupling paradigm. Specifically, we adopt identity-enhanced training that introduces an additional image cross-attention module to capture sufficient ID information and deactivate the original text cross-attention module to avoid text interference during training stage. Accordingly, our method can faithfully capture identity information from reference image, significantly improving the ID fidelity. Additionally, we employ a novel feature interaction mechanism that leverages a mixed attention module and an AdaIN-mean operation to effectively merge text information and identity information. Notably, our feature interaction mechanism not only preserves both identity and semantic details effectively but also enables convenient control over the styles of the generated images (as depicted in Fig. 1). Our contributions are summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.383, + 0.785, + 0.44 + ], + "angle": 0, + "content": "- We propose a novel ID-semantics decoupling paradigm to resolve the entanglement between image and text information, acquiring a remarkable balance between ID fidelity and semantic consistency in Identity-preserved personalization." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.444, + 0.785, + 0.5 + ], + "angle": 0, + "content": "- We propose a novel feature interaction mechanism incorporating a mixed attention module and an AdaIN-mean operation to effectively merge ID information and text information and also conveniently control the styles of the generated image in diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.505, + 0.785, + 0.547 + ], + "angle": 0, + "content": "- Experimental results demonstrate the excellent performance of our proposed method as compared with current state-of-the-art methods on both raw photo generation and style image generation." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.383, + 0.785, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.572, + 0.396, + 0.587 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.604, + 0.523, + 0.619 + ], + "angle": 0, + "content": "2.1 Text-to-image Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Text-to-image diffusion models, such as those explored in [3, 13, 27, 30, 43, 45], have garnered significant attention due to their impressive image generation capabilities. Current research endeavors aim to further enhance these models along multiple fronts, including the utilization of high-quality and large-scale datasets [18, 20, 32, 33], refinements to foundational architectures [3, 27, 30], and advancements in controllability [12, 29, 45]. Present iterations of text-to-image diffusion models typically follow a two-step process: first, encoding the text prompt using pre-trained text encoders such as CLIP [24] or T5 [26], and then utilizing the resulting text embedding as a condition for generating corresponding images through the diffusion process. Notably, the widely adopted Stable Diffusion model [27] distinguishes itself by executing the diffusion process in latent space instead of the original pixel space, leading to significant reductions in computation and time costs. An important extension to this framework is the Stable Diffusion XL (SDXL) [23], which enhances performance by scaling" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Yi Wu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.193 + ], + "angle": 0, + "content": "up the U-Net architecture and introducing an additional text encoder. Thus, our proposed method builds upon the SDXL. However, our method can also be extended to other text-to-image diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.226, + 0.544, + 0.241 + ], + "angle": 0, + "content": "2.2 Identity-preserved personalization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.262, + 0.788, + 0.519 + ], + "angle": 0, + "content": "Identity-preserved personalization aims to generate highly customized photos that accurately reflect a specific identity across various scenes, actions, and styles, drawing inspiration from one or more reference images. Initially, tuning-based methods, exemplified by DreamBooth [29], Textual Inversion [8] and Celeb-Basis [41], employ images of the same identity (ID) to fine-tune the model. While these methods yield results with high fidelity in preserving facial identity (ID), a significant drawback emerges: the customization of each ID necessitates a time investment of 10-30 minutes [15], consuming substantial computing resources and time. This limitation poses a significant obstacle to large-scale deployment in commercial applications. Consequently, recent advancements in tuning-free methods [6,7,9,16,17,21,35,36,39,40] have been introduced to streamline the generation process. These methods specifically leverage the construction of a vast amount of domain-specific data and the training of an encoder or hyper-network to represent input ID images as embeddings or LoRA weights within the model. Post-training, users need only input an image of the ID for customization, enabling personalized generation within seconds during the inference phase. These tuning-free methods typically contain two distinct manners." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.734 + ], + "angle": 0, + "content": "On one hand, methods [1, 17, 22, 39] incorporate text information alongside identity details within the text embedding space of the text encoder. For example, PhotoMaker [17] extracts identity embeddings from single or multiple reference images and merges them with corresponding class embeddings (e.g., \"man\" and \"woman\") in the text embedding space of the text encoder. While this stacking operation aids in achieving semantic consistency, it compresses image features into the text embedding space, leading to compromised identity fidelity. On the other hand, some studies [6, 36, 40] directly integrate identity information into the U-Net of the diffusion model. IP-Adapter [40] distinguishes itself by incorporating a additional cross-attention layer for each existing cross-attention layer within the original UNet model. This approach merges identity information with semantic details directly within the U-net but leads to distortion of the semantic space of the U-Net model. Consequently, this compromises the semantic consistency of the text prompt." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In summary, existing methods entangle the identity and text information, leading to a significant trade-off between ID fidelity and semantic consistency. To mitigate these limitations, we propose an identity-enhanced training to capture ID and text information separately. Moreover, we design an effective feature interaction mechanism leveraging a mixed attention module and an AdaIN-mean operation to preserve both identity and semantic details while also enabling convenient control over the styles of generated images." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.659, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Infinite-ID" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.569, + 0.162 + ], + "angle": 0, + "content": "2.3 Attention Control in Diffusion model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Previous studies have investigated various attention control techniques within diffusion models. Hertz et al. [10] employed a shared attention mechanism, concatenating and applying an AdaIN module on the key and value between reference and synthesis images within the self-attention layer to ensure style-consistent image generation using a reference style. Cao et al. [5] utilized a mutual self-attention approach to achieve consistent image generation and non-rigid image editing, wherein the key and value of the synthesis image were replaced with those of the reference image within the self-attention layers of the diffusion model. Similarly, Shi et al. [34] proposed a method termed reference attention, enabling consistent multi-view generation of target objects by concatenating the key and value features between the condition signal and the synthesis image in the self-attention layers. Wang et al. [39] and Avrahami et al. [2] exploited attention maps within the cross-attention layers to guide the optimization process towards disentangling learned concepts in personalized generation tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.406, + 0.331, + 0.421 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.438, + 0.373, + 0.453 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.789, + 0.645 + ], + "angle": 0, + "content": "Stable Diffusion XL. Our method builds upon Stable Diffusion XL [23], comprising three core components: a Variational AutoEncoder (VAE) denoted as \\(\\xi(\\cdot)\\), a conditional U-Net [28] represented by \\(\\epsilon_{\\theta}(\\cdot)\\), and two pre-trained text encoders [25] denoted as \\(\\Theta_1(\\cdot)\\) and \\(\\Theta_2(\\cdot)\\). Specifically, given a training image \\(x_0\\) and its corresponding text prompt \\(T\\), the VAE encoder \\(\\xi\\) transforms \\(x_0\\) from its original space \\(R^{H\\times W\\times 3}\\) to a compressed latent representation \\(z_0 = \\xi(x_0)\\), where \\(z_0 \\in R^{h\\times w\\times c}\\) and \\(c\\) denotes the latent dimension. Subsequently, the diffusion process operates within this compressed latent space to conserve computational resources and memory. Once the two text encoders process the text prompt \\(T\\) into a text embedding \\(c = \\mathrm{Concat}(\\Theta_1(T), \\Theta_2(T))\\), the conditional U-Net \\(\\epsilon_{\\theta}\\) predicts the noise \\(\\epsilon\\) based on the current timestep \\(t\\), the \\(t\\)-th latent representation \\(z_t\\), and the text embedding \\(c\\). The training objective is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.658, + 0.787, + 0.677 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {d i f f u s i o n}} = E _ {z _ {t}, t, c, \\epsilon \\in N (0, I)} [ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, c) | | _ {2} ^ {2} ]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.788, + 0.727 + ], + "angle": 0, + "content": "Attention mechanism in diffusion models. The fundamental unit of the stable diffusion model comprises a resblock, a self-attention layer, and a cross-attention layer. The attention mechanism is represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.737, + 0.787, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t n} (Q, K, V) = \\operatorname {S o f t m a x} (\\frac {Q K ^ {T}}{\\sqrt {d}}) V, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "where \\( Q \\) denotes the query feature projected from the spatial features generated by the preceding resblock, \\( K \\) and \\( V \\) represent the key and value features projected from the same spatial features as the query feature (in self-attention) or the text embedding extracted from the text prompt (in cross-attention)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Yi Wu et al." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.144, + 0.773, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.408, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Fig. 2: Framework of ID-semantics Decoupling Paradigm. In the training phase, we adopt Face embeddings extractor to extract rich identity information and identity-enhanced training for faithfully representing the identity provided by the reference image while mitigating interference from textual input. In the inference stage, a mixed attention module is introduced to replace the original self-attention mechanism within the denoising U-Net model, facilitating the fusion of both identity and text information." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.524, + 0.371, + 0.54 + ], + "angle": 0, + "content": "3.2 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Overview. In this section, we introduce our ID-semantics decoupling paradigm, as illustrated in Fig. 2, which effectively addresses the severe trade-off between high-fidelity identity and semantic consistency within identity-preserved personalization. Subsequently, we present our mixed attention mechanism, depicted in Fig. 3, designed to seamlessly integrate ID information and semantic information within the diffusion model during the inference stage. Additionally, we utilize an adaptive mean normalization (AdaIN-mean) operation to precisely align the style of the synthesized image with the desired style prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.789, + 0.842 + ], + "angle": 0, + "content": "ID-semantics Decoupling Paradigm. To faithfully capture high-fidelity identity, we implement a novel identity-enhanced strategy during the training stage, as depicted in Fig. 2. Diverging from conventional methods [6,17,40] that utilize text-image pairs for training, we opt to exclude the text prompt input and deactivate cross-attention modules for text embeddings within the U-Net model. Instead, we establish a training pair consisting of an ID image, where the face is aligned to extract identity information, and denoising image (\\( x_{t} \\) in Fig. 2). Both the denoising image used for training and the ID image belong to the same individual, but they vary in factors such as viewpoints and facial expressions. This approach fosters a more comprehensive learning process [17]. We adopt Face embeddings extractor to accurately capture and leverage identity information from" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Infinite-ID" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.147, + 0.495, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.409, + 0.429, + 0.422 + ], + "angle": 0, + "content": "(a) Mixed Attention" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.147, + 0.77, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.408, + 0.762, + 0.422 + ], + "angle": 0, + "content": "(b) Mixed Attention with AdaIN-mean" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.436, + 0.785, + 0.519 + ], + "angle": 0, + "content": "Fig. 3: Mixed attention mechanism. On the left side, we employ mixed attention to fuse identity and text information. This involves concatenating their respective key and value features and subsequently applying mixed attention, where identity features are updated based on the concatenated key and value features. On the right side, for style merging, we introduce an additional AdaIN-mean operation (as depicted in Eq. 8) to the concatenated key and value features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.785, + 0.594 + ], + "angle": 0, + "content": "the input ID image. Additionally, to seamlessly integrate identity information into the denoising U-Net model, we introduce an extra trainable cross-attention mechanism (\\( K_{id}' \\) and \\( V_{id}' \\) in Fig. 2) for image embeddings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.785, + 0.684 + ], + "angle": 0, + "content": "Throughout the training phase, we exclusively optimize the parameters associated with the face mapper, CLIP mapper, and the image cross-attention module, while keeping the parameters of the pre-trained diffusion model fixed. The optimization loss closely resembles the original diffusion loss formulation (as delineated in Eq. 1), with the sole distinction being the shift from a text condition to an identity condition as the conditional input." + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.693, + 0.785, + 0.711 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {d i f f u s i o n}} = E _ {z _ {t}, t, c _ {i d}, \\epsilon \\in N (0, I)} [ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, c _ {i d}) | | _ {2} ^ {2} ], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.671, + 0.735 + ], + "angle": 0, + "content": "where the \\( c_{id} \\) is the identity embeddings of the input ID image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.785, + 0.84 + ], + "angle": 0, + "content": "During the inference phase, we align the desired face within the ID image to afford identity information. Following the approach adopted during the training phase, we utilize a face recognition backbone and a CLIP image encoder to extract identity features from the aligned face image. Leveraging the trained face mapper, clip mapper, and image cross-attention mechanisms, these identity features are seamlessly integrated into the denoising U-Net model. Subsequently, we compute the key and value features for self-attention in the original stable" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Yi Wu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.252 + ], + "angle": 0, + "content": "diffusion model, considering only the text prompt input. These text key and value features are instrumental in the mixed attention process (illustrated in Fig. 3), facilitating the fusion of text and identity information. Moreover, to further augment the text information during the denoising process, we incorporate the original text cross-attention mechanism, integrating the resulting text hidden states with the output image hidden states obtained from the image cross-attention module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.789, + 0.541 + ], + "angle": 0, + "content": "Face Embeddings Extractor. We adopt a multifaceted approach by incorporating pre-trained models to extract facial features. Firstly, following the methodologies of prior research, we utilize a pre-trained CLIP image encoder as one of our facial feature extractors. Specifically, we leverage the local embeddings, comprising the last hidden states obtained from the CLIP image encoder, forming a sequence of embeddings with a length of N (N=257 in our implementation). Subsequently, we employ a CLIP mapper to project these image embeddings (with a dimensionality of 1664) to the same dimension as the text features in the pre-trained diffusion model. As elucidated in [40], the features extracted by the CLIP image encoder are instrumental in capturing the structural information pertinent to the identity face within identity-preserved personalization tasks. Additionally, we leverage the backbone of a face recognition model as another facial feature extractor. As highlighted in [40], features extracted by the face recognition backbone are adept at capturing the characteristics associated with human facial features within identity-preserved personalization tasks. More specifically, we utilize the global image embedding derived from the extracted features and subsequently employ a face mapper to align the dimensionality (512 dimensions) of the extracted global image embedding with the dimensionality of the text features in the pre-trained diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.79, + 0.573 + ], + "angle": 0, + "content": "In summary, the identity embeddings \\( c_{id} \\) corresponding to the ID image \\( x \\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.297, + 0.588, + 0.787, + 0.62 + ], + "angle": 0, + "content": "\\[\nc _ {i d} = \\operatorname {C o n c a t} \\left(\\mathrm {M} _ {\\text {c l i p}} \\left(\\mathrm {E} _ {\\text {c l i p}} (\\mathrm {F A} (x))\\right), \\mathrm {M} _ {\\text {f a c e}} \\left(\\mathrm {E} _ {\\text {f a c e}} (\\mathrm {F A} (x))\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.673 + ], + "angle": 0, + "content": "where \\(\\mathrm{Concat}(\\cdot, \\cdot)\\), \\(\\mathrm{M}_{\\mathrm{clip}}(\\cdot)\\), \\(\\mathrm{E}_{\\mathrm{clip}}(\\cdot)\\), \\(\\mathrm{M}_{\\mathrm{face}}(\\cdot)\\), \\(\\mathrm{E}_{\\mathrm{face}}(\\cdot)\\), and \\(\\mathrm{FA}(\\cdot)\\) denote the concatenation function, CLIP mapper, CLIP image encoder, face mapper, face recognition backbone, and face alignment module [44], respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Mixed Attention Mechanism. As explored in previous studies [5, 14, 37], the features in the self attention layers play a crucial role in consistency image generation (across-frame in text-to-video works), which indicates these features provide a refined and detailed semantics information. In our study, we extract features from the self-attention layers of the original text-to-image diffusion model to capture rich semantic information, represented as \\( K_{t} \\) and \\( V_{t} \\). We enhance the self-attention mechanism by incorporating it into a mixed attention framework, depicted in Fig. 3 (a). This fusion enables the integration of semantic features \\( (K_{t} \\) and \\( V_{t} \\)) with the identity-based features \\( (K_{id} \\) and \\( V_{id} \\)), thereby encapsulating identity information. Through this integration, the mixed attention mechanism seamlessly merges semantic details into the generated features" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.659, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Infinite-ID" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "across different resolutions. The formulation of the mixed attention mechanism is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.184, + 0.612, + 0.2 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t n} _ {\\operatorname {m i x}} (Q, K, V) \\triangleq \\operatorname {A t t n} (Q, \\hat {K}, \\hat {V})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.205, + 0.785, + 0.23 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l l} \\text {w . r . t} & \\hat {K} = \\underset {i d} {\\operatorname {C o n c a t}} \\left(K _ {i d}, K _ {t}\\right), \\quad \\hat {V} = \\underset {t} {\\operatorname {C o n c a t}} \\left(V _ {i d}, V _ {t}\\right), \\end{array} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.366, + 0.224, + 0.576, + 0.243 + ], + "angle": 0, + "content": "\\[\nK _ {i d} = W _ {k} ^ {i d} Z _ {i d}, \\quad K _ {t} = W _ {k} ^ {t} Z _ {t},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.366, + 0.247, + 0.568, + 0.263 + ], + "angle": 0, + "content": "\\[\nV _ {i d} = W _ {v} ^ {i d} Z _ {i d}, \\quad V _ {t} = W _ {v} ^ {t} Z _ {t},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.271, + 0.785, + 0.316 + ], + "angle": 0, + "content": "where \\( Z_{id} \\) and \\( Z_{t} \\) are the corresponding spatial features of generated features and semantic features, respectively. The parameters \\( W_{k}^{id} \\), \\( W_{k}^{t} \\), \\( W_{v}^{id} \\), and \\( W_{v}^{t} \\) correspond to the weights of the corresponding fully connected layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.317, + 0.784, + 0.361 + ], + "angle": 0, + "content": "Cross-attention Merging. To further refine semantic control, we incorporate text features into the identity feature within the cross-attention layers using the following formulation:" + }, + { + "type": "equation", + "bbox": [ + 0.299, + 0.37, + 0.785, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t n} _ {\\text {c r o s s}} (Q, K, V) \\triangleq \\operatorname {A t t n} \\left(Q, K _ {i d} ^ {\\prime}, V _ {i d} ^ {\\prime}\\right) + \\operatorname {A t t n} \\left(Q, K _ {t} ^ {\\prime}, V _ {t} ^ {\\prime}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.785, + 0.457 + ], + "angle": 0, + "content": "where \\( K_{id}^{\\prime} = \\hat{W}_{k}^{id}c_{id} \\), \\( V_{id}^{\\prime} = \\hat{W}_{v}^{id}c_{id} \\), \\( K_{t}^{\\prime} = \\hat{W}_{k}^{t}c_{t} \\), and \\( V_{t}^{\\prime} = \\hat{W}_{v}^{t}c_{t} \\). \\( c_{id} \\) and \\( c_{t} \\) are the identity embedding and text embedding, respectively. \\( \\hat{W}_{k}^{id} \\), \\( \\hat{W}_{v}^{id} \\), \\( \\hat{W}_{k}^{t} \\), and \\( \\hat{W}_{v}^{t} \\) correspond to the weights of the trainable fully connected layers within cross-attention module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.785, + 0.534 + ], + "angle": 0, + "content": "Style Information Merging. Inspired by [10], we propose an adaptive mean normalization (AdaIN-mean) operation to further align the style of the synthesis image with the style prompts. Concretely, we align the key and value features projected from identity features in both mixed attention and cross-attention with the key and value features projected from text features, formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.542, + 0.785, + 0.592 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l l l} K _ {i d} = \\operatorname {A d a I N - m} \\left(K _ {i d}, K _ {t}\\right), & V _ {i d} = \\operatorname {A d a I N - m} \\left(V _ {i d}, V _ {t}\\right), & \\text {F o r M i x e d A t t e n t i o n} \\\\ K _ {i d} ^ {\\prime} = \\operatorname {A d a I N - m} \\left(K _ {i d} ^ {\\prime}, K _ {t} ^ {\\prime}\\right), & V _ {i d} ^ {\\prime} = \\operatorname {A d a I N - m} \\left(V _ {i d} ^ {\\prime}, V _ {t} ^ {\\prime}\\right), & \\text {F o r C r o s s A t t e n t i o n} \\end{array} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.592, + 0.658, + 0.607 + ], + "angle": 0, + "content": "where the AdaIN-mean operation (AdaIN-m(\\cdot)) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.617, + 0.785, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A d a I N} - \\mathrm {m} (x, y) = x - \\mu (x) + \\mu (y), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.784, + 0.67 + ], + "angle": 0, + "content": "where \\(\\mu (x)\\in R^{d_k}\\) is the mean of key and value features across different pixels. The mixed attention with AdaIN-mean has been illustrated in Fig. 3 (b)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.691, + 0.376, + 0.708 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.841 + ], + "angle": 0, + "content": "After outlining the experimental setup in Sec. 4.1, we conduct a comparative analysis of raw photo generation and style image generation in Sec. 4.2. Ablation studies, highlighting the significance of various components, are presented in Sec. 4.3. Additionally, experiments involving multiple input ID images including identity mixing and identity linear interpolation are detailed in Sec. A.7 of the supplementary materials. For further insights, Sec. A.8 and Sec. A.9 of the supplementary materials provide additional qualitative results for raw photo generation and style photo generation, respectively." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Yi Wu et al." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.145, + 0.782, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.515 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparison on raw photo generation. The results demonstrate that our Infinite-ID consistently maintains identity fidelity and achieves high-quality semantic consistency with just a single input image." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.545, + 0.427, + 0.561 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.573, + 0.788, + 0.753 + ], + "angle": 0, + "content": "Implementation Details. Our experiments leverage a pre-trained Stable Diffusion XL (SDXL) model [23]. For image encoding, we utilize the OpenCLIP ViT-H/14 [42] and the backbone of ArcFace [11]. The SDXL model consists of 70 cross-attention layers, to each of which we append an additional image cross-attention module. Training is conducted on 16 A100 GPUs for 1 million steps, with a batch size of 4 per GPU. We employ the AdamW optimizer with a fixed learning rate of 1e-4 and weight decay set to 0.01. During inference, we employ the DDIM Sampler with 30 steps and guidance scale is set to 5.0. Training data are sourced from multiple datasets, including the LAION-2B dataset [33], the LAION-Face dataset [46], and few high-quality images collected from the Websites4. We curate a dataset where each individual is represented by multiple photographs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.755, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Evaluation. We assess the efficacy of our approach in preserving both identity fidelity and semantic consistency. Specifically, we measure identity fidelity, utilizing metrics such as \\( M_{\\mathrm{FaceNet}} \\) (measured by FaceNet [31]) and CLIP-I [8]. Identity fidelity is evaluated based on the similarity of detected faces between the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.546, + 0.84 + ], + "angle": 0, + "content": "4 https://www.pexels.com, https://pixabay.com." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Infinite-ID" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.148, + 0.782, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.467, + 0.785, + 0.509 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparison on style image generation. The results demonstrate that our method maintains strong identity fidelity, high-quality semantic consistency, and precise stylization using only a single reference image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.6 + ], + "angle": 0, + "content": "reference image and generated images. Semantic consistency is quantified using CLIP text-image consistency (CLIP-T [24]), which compares the text prompt with the corresponding generated images. More details of evaluation are shown in Sec. A.1 and Sec. A.2 of the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.62, + 0.54, + 0.636 + ], + "angle": 0, + "content": "4.2 Comparison to Previous Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Raw Photo Generation. We benchmark our Infinite-ID against established identity-preserving personalization approaches. The qualitative outcomes are illustrated in Fig. 4, while quantitative results are provided in Table 1. Notably, all methods are tuning-free, necessitating no test-time adjustments. FastComposer [39] exhibits challenges in maintaining identity fidelity, often presenting undesired artifacts in synthesized images. While IP-Adapter [40] and IP-Adapter-Face [40] demonstrates relatively fewer artifacts, its semantic consistency falls short. This phenomenon arises from the direct fusion of identity information with semantic details within the U-net model, leading to a compromise in semantic consistency. In contrast, PhotoMaker [17] exhibits commendable semantic consistency but falls short in preserving identity fidelity. Leveraging our ID-semantics decoupling paradigm, our method excels in preserving identity fidelity. Furthermore, our mixed attention mechanism effectively integrate the semantic" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Yi Wu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison. The evaluation metrics encompass CLIP-T, CLIP-I, and \\( M_{\\text{FaceNet}} \\). Our approach outperforms other methods in terms of identity fidelity while simultaneously achieving satisfactory semantic consistency. The best result is shown in bold, and the second best is underlined. Additionally, the quantitative comparison of ablation studies have been shown in the gray part." + }, + { + "type": "table", + "bbox": [ + 0.241, + 0.227, + 0.759, + 0.361 + ], + "angle": 0, + "content": "
CLIP-T ↑CLIP-I ↑M_FaceNet ↑
FastComposer [39]0.2920.8870.556
IP-Adapter [40]0.2740.9050.474
IP-Adapter-Face [40]0.3130.9190.513
PhotoMaker [17]0.3430.8140.502
Ours0.3400.9130.689
Ours (w/o identity-enhanced training)0.3290.8910.593
Ours (w/o Mixed Attention)0.3310.9050.700
Ours (Mixed Attention ⇒ Mutual Attention)0.3160.8080.398
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.784, + 0.435 + ], + "angle": 0, + "content": "information into the denoising process, positioning our method favorably against existing techniques. More comparison results are shown in Sec. A.4, Sec. A.5 and Sec. A.6 of the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.788, + 0.633 + ], + "angle": 0, + "content": "Style Image Generation. We demonstrate the results of the stylization results in Fig. 5, which compare our method with state-of-the-art tuning-free identity-preserving personalization methods. The style of the synthesis images include anime style, comic book style, and line art style. According to the stylization results, the IP-Adapter [40] and IP-Adapter-Face fails to depict the desired style in the prompt and the style of generation results always obey the tone in the reference image. The training pipeline of the IP-Adapter and IP-Adapter-Face entangles the text embedding and image embedding which leads to the distortion of the text embeddings space. FastComposer [39] also fails in stylization generation and shows undesired artifacts. PhotoMaker [17] achieves a better semantic consistency and stylization, but the identity fidelity is still unsatisfactory. In contrast, our method achieves high identity fidelity, appealing semantic consistency, and precise stylization meantime." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.659, + 0.39, + 0.675 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.764 + ], + "angle": 0, + "content": "In this section, we begin by conducting ablation studies to assess the influence of our identity-enhanced training and mixed attention mechanism. Furthermore, we conduct style image generation to evaluate the effectiveness of our AdaIN-mean operation in regulating the style of the generated images. More ablation studies are demonstrated in Sec. A.3 of the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Ablation of identity-enhanced training. In Fig. 6 and Table 1, we compare our method with \"Ours (w/o identity-enhanced training)\", which is implemented using an identity-semantics entangled training strategy. This strategy utilizes text-image pairs during training and activates cross-attention modules for text embeddings within the original U-Net model. It is noteworthy that both methods" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Infinite-ID" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.144, + 0.365, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.366, + 0.151, + 0.424, + 0.197 + ], + "angle": 0, + "content": "a woman wearing purple sweater, facing and smiling to the camera, in a forest" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.201, + 0.365, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.366, + 0.208, + 0.422, + 0.246 + ], + "angle": 0, + "content": "a photo of a man wearing a colorful shirt, in the living room" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.255, + 0.356, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.366, + 0.258, + 0.422, + 0.31 + ], + "angle": 0, + "content": "a woman wearing a red hat, smiling and facing to the camera, sunset, on the beach" + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.312, + 0.347, + 0.32 + ], + "angle": 0, + "content": "ID image" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.144, + 0.498, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.144, + 0.638, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.2, + 0.638, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.255, + 0.637, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.449, + 0.311, + 0.467, + 0.318 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_footnote", + "bbox": [ + 0.493, + 0.312, + 0.572, + 0.335 + ], + "angle": 0, + "content": "Ours (w/o identity-enhanced training)" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.144, + 0.708, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.2, + 0.708, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.255, + 0.708, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.312, + 0.7, + 0.326 + ], + "angle": 0, + "content": "Ours (M-A \\(\\Rightarrow\\) MU-A)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.345, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Fig. 6: Ablation study of our identity-enhanced training and mixed attention (M-A). It is evident that identity-enhanced training significantly improves the identity fidelity, and mixed-attention mechanism enhances semantic consistency compared to mutual attention (MU-A) approach [5]." + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.419, + 0.396, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.396, + 0.434, + 0.444, + 0.47 + ], + "angle": 0, + "content": "awearing a blue clothing, facing and smiling to the snow. A snowy, snowy, in an anime style" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.485, + 0.396, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.548, + 0.361, + 0.553 + ], + "angle": 0, + "content": "ID Image" + }, + { + "type": "image_footnote", + "bbox": [ + 0.397, + 0.498, + 0.444, + 0.535 + ], + "angle": 0, + "content": "a woman wearing a purple dress and smiling to the camera, on a beach, sunset, comic book style" + }, + { + "type": "image", + "bbox": [ + 0.444, + 0.419, + 0.609, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.444, + 0.484, + 0.609, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.479, + 0.548, + 0.492, + 0.554 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.419, + 0.693, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.484, + 0.692, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.548, + 0.685, + 0.554 + ], + "angle": 0, + "content": "Ours (AdaIN-mean \\(\\Rightarrow\\) AdaIN)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.568, + 0.788, + 0.611 + ], + "angle": 0, + "content": "Fig. 7: Ablation study of our AdaIN-mean operation. The results show that AdaIN-mean play a crucial role in style image generation. Compare to AdaIN [10] module, our AdaIN-mean helps to achieve a higher identity fidelity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.787, + 0.673 + ], + "angle": 0, + "content": "share the same inference processing. The qualitative comparison demonstrates that our identity-enhanced training notably enhances identity fidelity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Ablation of mixed attention mechanism. To assess the effectiveness of our proposed mixed attention (M-A) mechanism, we compare our method with \"Ours (w/o M-A)\" and \"Ours (M-A \\(\\Rightarrow\\) MU-A)\" in Fig. 6 and Table 1. Mutual self-attention (MU-A) [5] converts the existing self-attention into 'cross-attention' for consistent image editing, where the crossing operation happens in the self-attention of two related diffusion processes. The results show that our mixed attention mechanism demonstrates superior ability to improve semantic consistency while maintaining identity fidelity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Ablation of AdaIN-mean operation. To assess the effectiveness of our proposed AdaIN-mean operation, we compare our Infinite-ID model with variations, namely Ours (w/o AdaIN-mean) and Ours (AdaIN-mean \\(\\Rightarrow\\) AdaIN), as depicted" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Yi Wu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "in Fig. 7. The results reveal that: i) Ours (w/o AdaIN-mean) exhibits superior ID fidelity compared to Infinite-ID but fails to achieve style consistency with the text prompt; ii) Both our AdaIN-mean and AdaIN modules successfully achieve style consistency with the text prompt, yet AdaIN-mean maintains better ID fidelity than AdaIN. In conclusion, our proposed AdaIN-mean operation facilitates precise stylization while concurrently preserving ID fidelity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.262, + 0.517, + 0.279 + ], + "angle": 0, + "content": "5 Conclusion and limitations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.296, + 0.789, + 0.508 + ], + "angle": 0, + "content": "In this paper, we introduce Infinite-ID, an innovative identity-preserved personalization method designed to meet the requirement of identity (ID) fidelity and semantic consistency of the text prompt, all achievable with just one reference image and completed within seconds. Infinite-ID comprises three key components: the identity-enhanced training, mixed attention mechanism, and adaptive mean normalization (AdaIN-mean). Through extensive experimentation, our results illustrate that Infinite-ID outperforms baseline methods, delivering strong ID fidelity, superior generation quality, and precise semantic consistency in both raw photo generation and style image generation tasks. However, it's important to note that our method lacks multi-object personalization capability. Moreover, artifacts may occur when the human face occupies only a small portion of the entire image, attributed to limitations inherent in the original diffusion model. Furthermore, the image encoders employed in the identity-preserved personalization remain to be further explored." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.532, + 0.785, + 0.55 + ], + "angle": 0, + "content": "6 Copyright of Training Data and Ethical Considerations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.626 + ], + "angle": 0, + "content": "Copyright of Training Data. We construct our training dataset from the open-source public image dataset, including LAION-2B dataset, the LAION-Face dataset and few high-quality images collected from the Website5 which supports the CC0 (Creative Commons Zero) 1.0 Universal license6." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.788, + 0.746 + ], + "angle": 0, + "content": "Ethical Considerations. We recognize the ethical considerations about improper use of such high-quality human photo synthesis techniques. Therefore, we emphasize the significance of establishing and complying to ethical guidelines [4] and using this technology responsibly. We hope that our work will drive further discussions and research on the security and ethical use of the human photo generation in computer vision. Moreover, we will employ the watermark and safety checker (NSFW classifiers) to attempt to avoid inappropriate content generation." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.755, + 0.546, + 0.771 + ], + "angle": 0, + "content": "\\(^{5}\\) https://www.pexels.com, https://pixabay.com." + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.772, + 0.787, + 0.84 + ], + "angle": 0, + "content": "6 The person who associated a work with this deed has dedicated the work to the public domain by waiving all of his or her rights to the work worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law. You can copy, modify, distribute and perform the work, even for commercial purposes, all without asking permission." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.755, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Infinite-ID" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.174, + 0.787, + 0.22 + ], + "angle": 0, + "content": "This research was funded by the National Natural Science Foundation of China (U19B2044), supported by the Fundamental Research Funds for the Central Universities (WK2100000045)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.24, + 0.323, + 0.256 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.786, + 0.309 + ], + "angle": 0, + "content": "1. Achlioptas, P., Benetatos, A., Fostiropoulos, I., Skourtis, D.: Stellar: Systematic evaluation of human-centric personalized text-to-image methods. arXiv preprint arXiv:2312.06116 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.786, + 0.35 + ], + "angle": 0, + "content": "2. Avrahami, O., Aberman, K., Fried, O., Cohen-Or, D., Lischinski, D.: Break-a-scene: Extracting multiple concepts from a single image. arXiv preprint arXiv:2305.16311 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.35, + 0.786, + 0.392 + ], + "angle": 0, + "content": "3. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf 2, 3 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.786, + 0.418 + ], + "angle": 0, + "content": "4. Biden, J.R.: Executive order on the safe, secure, and trustworthy development and use of artificial intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.418, + 0.786, + 0.459 + ], + "angle": 0, + "content": "5. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. arXiv preprint arXiv:2304.08465 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.459, + 0.786, + 0.5 + ], + "angle": 0, + "content": "6. Chen, L., Zhao, M., Liu, Y., Ding, M., Song, Y., Wang, S., Wang, X., Yang, H., Liu, J., Du, K., et al.: Photovverse: Tuning-free image customization with text-to-image diffusion models. arXiv preprint arXiv:2309.05793 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.5, + 0.786, + 0.527 + ], + "angle": 0, + "content": "7. Chen, X., Huang, L., Liu, Y., Shen, Y., Zhao, D., Zhao, H.: Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.527, + 0.786, + 0.568 + ], + "angle": 0, + "content": "8. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.568, + 0.786, + 0.609 + ], + "angle": 0, + "content": "9. Gal, R., Arar, M., Atzmon, Y., Bermano, A.H., Chechik, G., Cohen-Or, D.: Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG) 42(4), 1-13 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.609, + 0.786, + 0.636 + ], + "angle": 0, + "content": "0. Hertz, A., Voynov, A., Fruchter, S., Cohen-Or, D.: Style aligned image generation via shared attention. arXiv preprint arXiv:2312.02133 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.636, + 0.786, + 0.69 + ], + "angle": 0, + "content": "1. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: 2017 IEEE International Conference on Computer Vision (ICCV) (Oct 2017). https://doi.org/10.1109/iccv.2017.167, http://dx.doi.org/10.1109/iccv.2017.167" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.69, + 0.786, + 0.745 + ], + "angle": 0, + "content": "2. Inoue, N., Kikuchi, K., Simo-Serra, E., Otani, M., Yamaguchi, K.: Layoutdm: Discrete diffusion model for controllable layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10167-10176 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.786, + 0.8 + ], + "angle": 0, + "content": "3. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.786, + 0.841 + ], + "angle": 0, + "content": "4. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.267, + 0.786, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Yi Wu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "15. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1931-1941 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.191, + 0.787, + 0.233 + ], + "angle": 0, + "content": "16. Li, D., Li, J., Hoi, S.C.: Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.785, + 0.274 + ], + "angle": 0, + "content": "17. Li, Z., Cao, M., Wang, X., Qi, Z., Cheng, M.M., Shan, Y.: Photomaker: Customizing realistic human photos via stacked id embedding. arXiv preprint arXiv:2312.04461 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.785, + 0.317 + ], + "angle": 0, + "content": "18. Li, Z., Usman, M., Tao, R., Xia, P., Wang, C., Chen, H., Li, B.: A systematic survey of regularization and normalization in gans. ACM Computing Surveys 55(11), 1-37 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.785, + 0.359 + ], + "angle": 0, + "content": "19. Li, Z., Wang, C., Zheng, H., Zhang, J., Li, B.: Fakeclr: Exploring contrastive learning for solving latent discontinuity in data-efficient gans. In: European Conference on Computer Vision. pp. 598-615. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.785, + 0.402 + ], + "angle": 0, + "content": "20. Li, Z., Xia, P., Tao, R., Niu, H., Li, B.: A new perspective on stabilizing gans training: Direct adversarial training. IEEE Transactions on Emerging Topics in Computational Intelligence 7(1), 178-189 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.404, + 0.785, + 0.444 + ], + "angle": 0, + "content": "21. Ma, J., Liang, J., Chen, C., Lu, H.: Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. arXiv preprint arXiv:2307.11410 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.785, + 0.488 + ], + "angle": 0, + "content": "22. Peng, X., Zhu, J., Jiang, B., Tai, Y., Luo, D., Zhang, J., Lin, W., Jin, T., Wang, C., Ji, R.: Portrait booth: A versatile portrait model for fast identity-preserved personalization. arXiv preprint arXiv:2312.06354 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.489, + 0.785, + 0.53 + ], + "angle": 0, + "content": "23. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.531, + 0.785, + 0.585 + ], + "angle": 0, + "content": "24. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.785, + 0.642 + ], + "angle": 0, + "content": "25. Radford, A., Kim, J., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Amanda, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. Cornell University - arXiv, Cornell University - arXiv (Feb 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.785, + 0.685 + ], + "angle": 0, + "content": "26. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.687, + 0.785, + 0.728 + ], + "angle": 0, + "content": "27. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.785, + 0.784 + ], + "angle": 0, + "content": "28. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "29. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Infinite-ID" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "30. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "31. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.787, + 0.3 + ], + "angle": 0, + "content": "32. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "33. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.343, + 0.787, + 0.384 + ], + "angle": 0, + "content": "34. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.384, + 0.787, + 0.412 + ], + "angle": 0, + "content": "35. Wang, Q., Bai, X., Wang, H., Qin, Z., Chen, A.: Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.412, + 0.787, + 0.452 + ], + "angle": 0, + "content": "36. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.453, + 0.787, + 0.508 + ], + "angle": 0, + "content": "37. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7623-7633 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "38. Wu, Y., Li, Z., Wang, C., Zheng, H., Zhao, S., Li, B., Tao, D.: Domain remodulation for few-shot generative domain adaptation. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.55, + 0.787, + 0.591 + ], + "angle": 0, + "content": "39. Xiao, G., Yin, T., Freeman, W.T., Durand, F., Han, S.: Fastcomposer: Tuning-free multi-subject image generation with localized attention. arXiv preprint arXiv:2305.10431 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.592, + 0.787, + 0.632 + ], + "angle": 0, + "content": "40. Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.633, + 0.787, + 0.715 + ], + "angle": 0, + "content": "41. Yuan, G., Cun, X., Zhang, Y., Li, M., Qi, C., Wang, X., Shan, Y., Zheng, H.: Inserting anybody in diffusion models via celeb basis. In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 72958-72982. Curran Associates, Inc. (2023), https://proceedings.neurips.cc/paper_files/paper/2023/file/e6d37cc5723e810b793c834bc6647cf-Paper-Conference.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.716, + 0.787, + 0.77 + ], + "angle": 0, + "content": "42. Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djolonga, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., et al.: A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.771, + 0.787, + 0.799 + ], + "angle": 0, + "content": "43. Zhang, C., Zhang, C., Zhang, M., Kweon, I.S.: Text-to-image diffusion model in generative ai: A survey. arXiv preprint arXiv:2303.07909 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "44. Zhang, K., Zhang, Z., Li, Z., Qiao, Y.: Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters 23(10), 1499-1503 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Yi Wu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "45. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "46. Zheng, Y., Yang, H., Zhang, T., Bao, J., Chen, D., Huang, Y., Yuan, L., Chen, D., Zeng, M., Wen, F.: General facial representation learning in a visual-linguistic manner. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18697-18709 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.245 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Infinite-ID_ Identity-preserved Personalization via ID-semantics Decoupling Paradigm/6b2c0220-7e0d-410c-b7de-7b3ff87a9f85_origin.pdf b/2024/Infinite-ID_ Identity-preserved Personalization via ID-semantics Decoupling Paradigm/6b2c0220-7e0d-410c-b7de-7b3ff87a9f85_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cd7df6f4d6057d3dfce6859afd8963602a8c19f5 --- /dev/null +++ b/2024/Infinite-ID_ Identity-preserved Personalization via ID-semantics Decoupling Paradigm/6b2c0220-7e0d-410c-b7de-7b3ff87a9f85_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70f145505341d578e93d48489957bc77f32ad7b37f40b20c3804c86db383d4d0 +size 2271117 diff --git a/2024/Infinite-ID_ Identity-preserved Personalization via ID-semantics Decoupling Paradigm/full.md b/2024/Infinite-ID_ Identity-preserved Personalization via ID-semantics Decoupling Paradigm/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4e0ab9d13e40a946fd37c5f069f66c282376cca0 --- /dev/null +++ b/2024/Infinite-ID_ Identity-preserved Personalization via ID-semantics Decoupling Paradigm/full.md @@ -0,0 +1,314 @@ +# Infinite-ID: Identity-preserved Personalization via ID-semantics Decoupling Paradigm + +Yi Wu $^{1\star}$ , Ziqiang Li $^{1,2\star}$ , Heliang Zheng $^{1}$ , Chaoyue Wang $^{3\star\star}$ , and Bin Li $^{1\star\star}$ + +1 University of Science and Technology of China, China + +$^{2}$ Nanjing University of Information Science and Technology, China + +3 The University of Sydney, Australia + +wuyi2021@mail.ustc.edu.cn, iceli@mail.ustc.edu.cn, + +zhenghl@mail.ustc.edu.cn, + +chaoyue.wang@outlook.com, binli@ustc.edu.cn + +Abstract. Drawing on recent advancements in diffusion models for text-to-image generation, identity-preserved personalization has made significant progress in accurately capturing specific identities with just a single reference image. However, existing methods primarily integrate reference images within the text embedding space, leading to a complex entanglement of image and text information, which poses challenges for preserving both identity fidelity and semantic consistency. To tackle this challenge, we propose Infinite-ID, an ID-semantics decoupling paradigm for identity-preserved personalization. Specifically, we introduce identity-enhanced training, incorporating an additional image cross-attention module to capture sufficient ID information while deactivating the original text cross-attention module of the diffusion model. This ensures that the image stream faithfully represents the identity provided by the reference image while mitigating interference from textual input. Additionally, we introduce a feature interaction mechanism that combines a mixed attention module with an AdaIN-mean operation to seamlessly merge the two streams. This mechanism not only enhances the fidelity of identity and semantic consistency but also enables convenient control over the styles of the generated images. Extensive experimental results on both raw photo generation and style image generation demonstrate the superior performance of our proposed method. + +Keywords: Personalized Text-to-image Generation, Stable Diffusion, Identity-preserved Personalization + +# 1 Introduction + +Human photo synthesis [19,38] has experienced notable advancements, particularly with the introduction of large text-to-image diffusion models such as Stable + +![](images/4e341a16035d1d8bddf81feb9456875b5cb78d771d719f6b70dffc7434288e56.jpg) +Fig. 1: With just a single reference image, our Infinite-ID framework excels in synthesizing high-quality images while maintaining superior identity fidelity and text semantic consistency in various styles. + +Diffusion (SD) [27], Imagen [30], and DALL-E 3 [3]. Benefit from personalized text-to-image generation, recent researches focus on Identity-preserved personalization. This specialized area aims to produce highly customized photos that faithfully reflect a specific identity in novel scenes, actions, and styles, drawing inspiration from one or more reference images. This task has garnered considerable attention, leading to the development of numerous applications, including personalized AI portraits and virtual try-on scenarios. In the context of Identity-preserved personalization, the emphasis is placed on maintaining the invariance of human facial identity (ID), requiring a heightened level of detail and fidelity compared to more general styles or objects. + +Recent tuning-free methods exhibit promise for large-scale deployment, yet they face a notable challenge in balancing the trade-off between the fidelity of identity representation (ID fidelity) and the consistency of semantic understanding conveyed by the text prompt. This challenge arises due to the inherent entanglement of image and text information. Typically, tuning-free methods extract ID information from reference images and integrate it into the semantic information in two distinct ways. The first type, exemplified by PhotoMaker [17], incorporates text information with ID details in the text embedding space of the text encoder. While this merging approach aids in achieving semantic consistency, it compresses image features into the text embedding space, thereby weakening the ID information of the image and compromising identity fidelity. The second type, demonstrated by IP-Adapter [40], directly injects ID information into the U-Net of the diffusion model through an additional trainable cross-attention module. Although this approach aims to enhance the strength of ID information for improved fidelity, it tends to favor the image branch during training, consequently weakening the text branch and compromising semantic consistency. In summary, existing methods entangle image and text information, resulting in a significant trade-off between ID fidelity and semantic consistency. + +To address the entanglement between image and text information, we propose Infinite-ID, an innovative approach to personalized text-to-image generation. Our method tackles the trade-off between maintaining high fidelity of identity and ensuring semantic consistency of the text prompt by implementing the ID-semantics decoupling paradigm. Specifically, we adopt identity-enhanced training that introduces an additional image cross-attention module to capture sufficient ID information and deactivate the original text cross-attention module to avoid text interference during training stage. Accordingly, our method can faithfully capture identity information from reference image, significantly improving the ID fidelity. Additionally, we employ a novel feature interaction mechanism that leverages a mixed attention module and an AdaIN-mean operation to effectively merge text information and identity information. Notably, our feature interaction mechanism not only preserves both identity and semantic details effectively but also enables convenient control over the styles of the generated images (as depicted in Fig. 1). Our contributions are summarized as: + +- We propose a novel ID-semantics decoupling paradigm to resolve the entanglement between image and text information, acquiring a remarkable balance between ID fidelity and semantic consistency in Identity-preserved personalization. +- We propose a novel feature interaction mechanism incorporating a mixed attention module and an AdaIN-mean operation to effectively merge ID information and text information and also conveniently control the styles of the generated image in diffusion models. +- Experimental results demonstrate the excellent performance of our proposed method as compared with current state-of-the-art methods on both raw photo generation and style image generation. + +# 2 Related Works + +# 2.1 Text-to-image Diffusion Models + +Text-to-image diffusion models, such as those explored in [3, 13, 27, 30, 43, 45], have garnered significant attention due to their impressive image generation capabilities. Current research endeavors aim to further enhance these models along multiple fronts, including the utilization of high-quality and large-scale datasets [18, 20, 32, 33], refinements to foundational architectures [3, 27, 30], and advancements in controllability [12, 29, 45]. Present iterations of text-to-image diffusion models typically follow a two-step process: first, encoding the text prompt using pre-trained text encoders such as CLIP [24] or T5 [26], and then utilizing the resulting text embedding as a condition for generating corresponding images through the diffusion process. Notably, the widely adopted Stable Diffusion model [27] distinguishes itself by executing the diffusion process in latent space instead of the original pixel space, leading to significant reductions in computation and time costs. An important extension to this framework is the Stable Diffusion XL (SDXL) [23], which enhances performance by scaling + +up the U-Net architecture and introducing an additional text encoder. Thus, our proposed method builds upon the SDXL. However, our method can also be extended to other text-to-image diffusion models. + +# 2.2 Identity-preserved personalization + +Identity-preserved personalization aims to generate highly customized photos that accurately reflect a specific identity across various scenes, actions, and styles, drawing inspiration from one or more reference images. Initially, tuning-based methods, exemplified by DreamBooth [29], Textual Inversion [8] and Celeb-Basis [41], employ images of the same identity (ID) to fine-tune the model. While these methods yield results with high fidelity in preserving facial identity (ID), a significant drawback emerges: the customization of each ID necessitates a time investment of 10-30 minutes [15], consuming substantial computing resources and time. This limitation poses a significant obstacle to large-scale deployment in commercial applications. Consequently, recent advancements in tuning-free methods [6,7,9,16,17,21,35,36,39,40] have been introduced to streamline the generation process. These methods specifically leverage the construction of a vast amount of domain-specific data and the training of an encoder or hyper-network to represent input ID images as embeddings or LoRA weights within the model. Post-training, users need only input an image of the ID for customization, enabling personalized generation within seconds during the inference phase. These tuning-free methods typically contain two distinct manners. + +On one hand, methods [1, 17, 22, 39] incorporate text information alongside identity details within the text embedding space of the text encoder. For example, PhotoMaker [17] extracts identity embeddings from single or multiple reference images and merges them with corresponding class embeddings (e.g., "man" and "woman") in the text embedding space of the text encoder. While this stacking operation aids in achieving semantic consistency, it compresses image features into the text embedding space, leading to compromised identity fidelity. On the other hand, some studies [6, 36, 40] directly integrate identity information into the U-Net of the diffusion model. IP-Adapter [40] distinguishes itself by incorporating a additional cross-attention layer for each existing cross-attention layer within the original UNet model. This approach merges identity information with semantic details directly within the U-net but leads to distortion of the semantic space of the U-Net model. Consequently, this compromises the semantic consistency of the text prompt. + +In summary, existing methods entangle the identity and text information, leading to a significant trade-off between ID fidelity and semantic consistency. To mitigate these limitations, we propose an identity-enhanced training to capture ID and text information separately. Moreover, we design an effective feature interaction mechanism leveraging a mixed attention module and an AdaIN-mean operation to preserve both identity and semantic details while also enabling convenient control over the styles of generated images. + +# 2.3 Attention Control in Diffusion model + +Previous studies have investigated various attention control techniques within diffusion models. Hertz et al. [10] employed a shared attention mechanism, concatenating and applying an AdaIN module on the key and value between reference and synthesis images within the self-attention layer to ensure style-consistent image generation using a reference style. Cao et al. [5] utilized a mutual self-attention approach to achieve consistent image generation and non-rigid image editing, wherein the key and value of the synthesis image were replaced with those of the reference image within the self-attention layers of the diffusion model. Similarly, Shi et al. [34] proposed a method termed reference attention, enabling consistent multi-view generation of target objects by concatenating the key and value features between the condition signal and the synthesis image in the self-attention layers. Wang et al. [39] and Avrahami et al. [2] exploited attention maps within the cross-attention layers to guide the optimization process towards disentangling learned concepts in personalized generation tasks. + +# 3 Method + +# 3.1 Preliminaries + +Stable Diffusion XL. Our method builds upon Stable Diffusion XL [23], comprising three core components: a Variational AutoEncoder (VAE) denoted as $\xi(\cdot)$ , a conditional U-Net [28] represented by $\epsilon_{\theta}(\cdot)$ , and two pre-trained text encoders [25] denoted as $\Theta_1(\cdot)$ and $\Theta_2(\cdot)$ . Specifically, given a training image $x_0$ and its corresponding text prompt $T$ , the VAE encoder $\xi$ transforms $x_0$ from its original space $R^{H\times W\times 3}$ to a compressed latent representation $z_0 = \xi(x_0)$ , where $z_0 \in R^{h\times w\times c}$ and $c$ denotes the latent dimension. Subsequently, the diffusion process operates within this compressed latent space to conserve computational resources and memory. Once the two text encoders process the text prompt $T$ into a text embedding $c = \mathrm{Concat}(\Theta_1(T), \Theta_2(T))$ , the conditional U-Net $\epsilon_{\theta}$ predicts the noise $\epsilon$ based on the current timestep $t$ , the $t$ -th latent representation $z_t$ , and the text embedding $c$ . The training objective is formulated as follows: + +$$ +L _ {\text {d i f f u s i o n}} = E _ {z _ {t}, t, c, \epsilon \in N (0, I)} [ | | \epsilon - \epsilon_ {\theta} (z _ {t}, t, c) | | _ {2} ^ {2} ]. \tag {1} +$$ + +Attention mechanism in diffusion models. The fundamental unit of the stable diffusion model comprises a resblock, a self-attention layer, and a cross-attention layer. The attention mechanism is represented as follows: + +$$ +\operatorname {A t t n} (Q, K, V) = \operatorname {S o f t m a x} (\frac {Q K ^ {T}}{\sqrt {d}}) V, \tag {2} +$$ + +where $Q$ denotes the query feature projected from the spatial features generated by the preceding resblock, $K$ and $V$ represent the key and value features projected from the same spatial features as the query feature (in self-attention) or the text embedding extracted from the text prompt (in cross-attention). + +![](images/bc8b81633e4218c09ab22f6c46b8e81ccab461952e9aab4409f34a539fb94514.jpg) +Fig. 2: Framework of ID-semantics Decoupling Paradigm. In the training phase, we adopt Face embeddings extractor to extract rich identity information and identity-enhanced training for faithfully representing the identity provided by the reference image while mitigating interference from textual input. In the inference stage, a mixed attention module is introduced to replace the original self-attention mechanism within the denoising U-Net model, facilitating the fusion of both identity and text information. + +# 3.2 Methodology + +Overview. In this section, we introduce our ID-semantics decoupling paradigm, as illustrated in Fig. 2, which effectively addresses the severe trade-off between high-fidelity identity and semantic consistency within identity-preserved personalization. Subsequently, we present our mixed attention mechanism, depicted in Fig. 3, designed to seamlessly integrate ID information and semantic information within the diffusion model during the inference stage. Additionally, we utilize an adaptive mean normalization (AdaIN-mean) operation to precisely align the style of the synthesized image with the desired style prompts. + +ID-semantics Decoupling Paradigm. To faithfully capture high-fidelity identity, we implement a novel identity-enhanced strategy during the training stage, as depicted in Fig. 2. Diverging from conventional methods [6,17,40] that utilize text-image pairs for training, we opt to exclude the text prompt input and deactivate cross-attention modules for text embeddings within the U-Net model. Instead, we establish a training pair consisting of an ID image, where the face is aligned to extract identity information, and denoising image ( $x_{t}$ in Fig. 2). Both the denoising image used for training and the ID image belong to the same individual, but they vary in factors such as viewpoints and facial expressions. This approach fosters a more comprehensive learning process [17]. We adopt Face embeddings extractor to accurately capture and leverage identity information from + +![](images/77312140525ed59cbd83583c6bab7affe65f8d203ba21ec5e32ac9b6e5dbcb74.jpg) +(a) Mixed Attention +Fig. 3: Mixed attention mechanism. On the left side, we employ mixed attention to fuse identity and text information. This involves concatenating their respective key and value features and subsequently applying mixed attention, where identity features are updated based on the concatenated key and value features. On the right side, for style merging, we introduce an additional AdaIN-mean operation (as depicted in Eq. 8) to the concatenated key and value features. + +![](images/0b4fd8e1db8bf40fe9c7f7a5c537cbda2c1af8d258d98ef0d3b1ffded6a26f95.jpg) +(b) Mixed Attention with AdaIN-mean + +the input ID image. Additionally, to seamlessly integrate identity information into the denoising U-Net model, we introduce an extra trainable cross-attention mechanism ( $K_{id}'$ and $V_{id}'$ in Fig. 2) for image embeddings. + +Throughout the training phase, we exclusively optimize the parameters associated with the face mapper, CLIP mapper, and the image cross-attention module, while keeping the parameters of the pre-trained diffusion model fixed. The optimization loss closely resembles the original diffusion loss formulation (as delineated in Eq. 1), with the sole distinction being the shift from a text condition to an identity condition as the conditional input. + +$$ +L _ {\text {d i f f u s i o n}} = E _ {z _ {t}, t, c _ {i d}, \epsilon \in N (0, I)} [ | | \epsilon - \epsilon_ {\theta} (z _ {t}, t, c _ {i d}) | | _ {2} ^ {2} ], \tag {3} +$$ + +where the $c_{id}$ is the identity embeddings of the input ID image. + +During the inference phase, we align the desired face within the ID image to afford identity information. Following the approach adopted during the training phase, we utilize a face recognition backbone and a CLIP image encoder to extract identity features from the aligned face image. Leveraging the trained face mapper, clip mapper, and image cross-attention mechanisms, these identity features are seamlessly integrated into the denoising U-Net model. Subsequently, we compute the key and value features for self-attention in the original stable + +diffusion model, considering only the text prompt input. These text key and value features are instrumental in the mixed attention process (illustrated in Fig. 3), facilitating the fusion of text and identity information. Moreover, to further augment the text information during the denoising process, we incorporate the original text cross-attention mechanism, integrating the resulting text hidden states with the output image hidden states obtained from the image cross-attention module. + +Face Embeddings Extractor. We adopt a multifaceted approach by incorporating pre-trained models to extract facial features. Firstly, following the methodologies of prior research, we utilize a pre-trained CLIP image encoder as one of our facial feature extractors. Specifically, we leverage the local embeddings, comprising the last hidden states obtained from the CLIP image encoder, forming a sequence of embeddings with a length of N (N=257 in our implementation). Subsequently, we employ a CLIP mapper to project these image embeddings (with a dimensionality of 1664) to the same dimension as the text features in the pre-trained diffusion model. As elucidated in [40], the features extracted by the CLIP image encoder are instrumental in capturing the structural information pertinent to the identity face within identity-preserved personalization tasks. Additionally, we leverage the backbone of a face recognition model as another facial feature extractor. As highlighted in [40], features extracted by the face recognition backbone are adept at capturing the characteristics associated with human facial features within identity-preserved personalization tasks. More specifically, we utilize the global image embedding derived from the extracted features and subsequently employ a face mapper to align the dimensionality (512 dimensions) of the extracted global image embedding with the dimensionality of the text features in the pre-trained diffusion model. + +In summary, the identity embeddings $c_{id}$ corresponding to the ID image $x$ can be expressed as: + +$$ +c _ {i d} = \operatorname {C o n c a t} \left(\mathrm {M} _ {\text {c l i p}} \left(\mathrm {E} _ {\text {c l i p}} (\mathrm {F A} (x))\right), \mathrm {M} _ {\text {f a c e}} \left(\mathrm {E} _ {\text {f a c e}} (\mathrm {F A} (x))\right)\right), \tag {4} +$$ + +where $\mathrm{Concat}(\cdot, \cdot)$ , $\mathrm{M}_{\mathrm{clip}}(\cdot)$ , $\mathrm{E}_{\mathrm{clip}}(\cdot)$ , $\mathrm{M}_{\mathrm{face}}(\cdot)$ , $\mathrm{E}_{\mathrm{face}}(\cdot)$ , and $\mathrm{FA}(\cdot)$ denote the concatenation function, CLIP mapper, CLIP image encoder, face mapper, face recognition backbone, and face alignment module [44], respectively. + +Mixed Attention Mechanism. As explored in previous studies [5, 14, 37], the features in the self attention layers play a crucial role in consistency image generation (across-frame in text-to-video works), which indicates these features provide a refined and detailed semantics information. In our study, we extract features from the self-attention layers of the original text-to-image diffusion model to capture rich semantic information, represented as $K_{t}$ and $V_{t}$ . We enhance the self-attention mechanism by incorporating it into a mixed attention framework, depicted in Fig. 3 (a). This fusion enables the integration of semantic features $(K_{t}$ and $V_{t}$ ) with the identity-based features $(K_{id}$ and $V_{id}$ ), thereby encapsulating identity information. Through this integration, the mixed attention mechanism seamlessly merges semantic details into the generated features + +across different resolutions. The formulation of the mixed attention mechanism is as follows: + +$$ +\operatorname {A t t n} _ {\operatorname {m i x}} (Q, K, V) \triangleq \operatorname {A t t n} (Q, \hat {K}, \hat {V}) +$$ + +$$ +\begin{array}{l l} \text {w . r . t} & \hat {K} = \underset {i d} {\operatorname {C o n c a t}} \left(K _ {i d}, K _ {t}\right), \quad \hat {V} = \underset {t} {\operatorname {C o n c a t}} \left(V _ {i d}, V _ {t}\right), \end{array} \tag {5} +$$ + +$$ +K _ {i d} = W _ {k} ^ {i d} Z _ {i d}, \quad K _ {t} = W _ {k} ^ {t} Z _ {t}, +$$ + +$$ +V _ {i d} = W _ {v} ^ {i d} Z _ {i d}, \quad V _ {t} = W _ {v} ^ {t} Z _ {t}, +$$ + +where $Z_{id}$ and $Z_{t}$ are the corresponding spatial features of generated features and semantic features, respectively. The parameters $W_{k}^{id}$ , $W_{k}^{t}$ , $W_{v}^{id}$ , and $W_{v}^{t}$ correspond to the weights of the corresponding fully connected layers. + +Cross-attention Merging. To further refine semantic control, we incorporate text features into the identity feature within the cross-attention layers using the following formulation: + +$$ +\operatorname {A t t n} _ {\text {c r o s s}} (Q, K, V) \triangleq \operatorname {A t t n} \left(Q, K _ {i d} ^ {\prime}, V _ {i d} ^ {\prime}\right) + \operatorname {A t t n} \left(Q, K _ {t} ^ {\prime}, V _ {t} ^ {\prime}\right), \tag {6} +$$ + +where $K_{id}^{\prime} = \hat{W}_{k}^{id}c_{id}$ , $V_{id}^{\prime} = \hat{W}_{v}^{id}c_{id}$ , $K_{t}^{\prime} = \hat{W}_{k}^{t}c_{t}$ , and $V_{t}^{\prime} = \hat{W}_{v}^{t}c_{t}$ . $c_{id}$ and $c_{t}$ are the identity embedding and text embedding, respectively. $\hat{W}_{k}^{id}$ , $\hat{W}_{v}^{id}$ , $\hat{W}_{k}^{t}$ , and $\hat{W}_{v}^{t}$ correspond to the weights of the trainable fully connected layers within cross-attention module. + +Style Information Merging. Inspired by [10], we propose an adaptive mean normalization (AdaIN-mean) operation to further align the style of the synthesis image with the style prompts. Concretely, we align the key and value features projected from identity features in both mixed attention and cross-attention with the key and value features projected from text features, formulated as follows: + +$$ +\begin{array}{l l l} K _ {i d} = \operatorname {A d a I N - m} \left(K _ {i d}, K _ {t}\right), & V _ {i d} = \operatorname {A d a I N - m} \left(V _ {i d}, V _ {t}\right), & \text {F o r M i x e d A t t e n t i o n} \\ K _ {i d} ^ {\prime} = \operatorname {A d a I N - m} \left(K _ {i d} ^ {\prime}, K _ {t} ^ {\prime}\right), & V _ {i d} ^ {\prime} = \operatorname {A d a I N - m} \left(V _ {i d} ^ {\prime}, V _ {t} ^ {\prime}\right), & \text {F o r C r o s s A t t e n t i o n} \end{array} \tag {7} +$$ + +where the AdaIN-mean operation (AdaIN-m(\cdot)) is defined as: + +$$ +\operatorname {A d a I N} - \mathrm {m} (x, y) = x - \mu (x) + \mu (y), \tag {8} +$$ + +where $\mu (x)\in R^{d_k}$ is the mean of key and value features across different pixels. The mixed attention with AdaIN-mean has been illustrated in Fig. 3 (b). + +# 4 Experiments + +After outlining the experimental setup in Sec. 4.1, we conduct a comparative analysis of raw photo generation and style image generation in Sec. 4.2. Ablation studies, highlighting the significance of various components, are presented in Sec. 4.3. Additionally, experiments involving multiple input ID images including identity mixing and identity linear interpolation are detailed in Sec. A.7 of the supplementary materials. For further insights, Sec. A.8 and Sec. A.9 of the supplementary materials provide additional qualitative results for raw photo generation and style photo generation, respectively. + +![](images/ce66ee3a3d30418f47ba6bf5dc8d7764ec132a6e50227e46137dfc4dbb4ba01d.jpg) +Fig. 4: Qualitative comparison on raw photo generation. The results demonstrate that our Infinite-ID consistently maintains identity fidelity and achieves high-quality semantic consistency with just a single input image. + +# 4.1 Experimental Setup + +Implementation Details. Our experiments leverage a pre-trained Stable Diffusion XL (SDXL) model [23]. For image encoding, we utilize the OpenCLIP ViT-H/14 [42] and the backbone of ArcFace [11]. The SDXL model consists of 70 cross-attention layers, to each of which we append an additional image cross-attention module. Training is conducted on 16 A100 GPUs for 1 million steps, with a batch size of 4 per GPU. We employ the AdamW optimizer with a fixed learning rate of 1e-4 and weight decay set to 0.01. During inference, we employ the DDIM Sampler with 30 steps and guidance scale is set to 5.0. Training data are sourced from multiple datasets, including the LAION-2B dataset [33], the LAION-Face dataset [46], and few high-quality images collected from the Websites4. We curate a dataset where each individual is represented by multiple photographs. + +Evaluation. We assess the efficacy of our approach in preserving both identity fidelity and semantic consistency. Specifically, we measure identity fidelity, utilizing metrics such as $M_{\mathrm{FaceNet}}$ (measured by FaceNet [31]) and CLIP-I [8]. Identity fidelity is evaluated based on the similarity of detected faces between the + +![](images/01d499f8c7db4e1c3c67dccca8ae2f10d8c2b4af6eeadea587c0d1c24f4ba24e.jpg) +Fig. 5: Qualitative comparison on style image generation. The results demonstrate that our method maintains strong identity fidelity, high-quality semantic consistency, and precise stylization using only a single reference image. + +reference image and generated images. Semantic consistency is quantified using CLIP text-image consistency (CLIP-T [24]), which compares the text prompt with the corresponding generated images. More details of evaluation are shown in Sec. A.1 and Sec. A.2 of the supplementary materials. + +# 4.2 Comparison to Previous Methods + +Raw Photo Generation. We benchmark our Infinite-ID against established identity-preserving personalization approaches. The qualitative outcomes are illustrated in Fig. 4, while quantitative results are provided in Table 1. Notably, all methods are tuning-free, necessitating no test-time adjustments. FastComposer [39] exhibits challenges in maintaining identity fidelity, often presenting undesired artifacts in synthesized images. While IP-Adapter [40] and IP-Adapter-Face [40] demonstrates relatively fewer artifacts, its semantic consistency falls short. This phenomenon arises from the direct fusion of identity information with semantic details within the U-net model, leading to a compromise in semantic consistency. In contrast, PhotoMaker [17] exhibits commendable semantic consistency but falls short in preserving identity fidelity. Leveraging our ID-semantics decoupling paradigm, our method excels in preserving identity fidelity. Furthermore, our mixed attention mechanism effectively integrate the semantic + +Table 1: Quantitative comparison. The evaluation metrics encompass CLIP-T, CLIP-I, and $M_{\text{FaceNet}}$ . Our approach outperforms other methods in terms of identity fidelity while simultaneously achieving satisfactory semantic consistency. The best result is shown in bold, and the second best is underlined. Additionally, the quantitative comparison of ablation studies have been shown in the gray part. + +
CLIP-T ↑CLIP-I ↑M_FaceNet ↑
FastComposer [39]0.2920.8870.556
IP-Adapter [40]0.2740.9050.474
IP-Adapter-Face [40]0.3130.9190.513
PhotoMaker [17]0.3430.8140.502
Ours0.3400.9130.689
Ours (w/o identity-enhanced training)0.3290.8910.593
Ours (w/o Mixed Attention)0.3310.9050.700
Ours (Mixed Attention ⇒ Mutual Attention)0.3160.8080.398
+ +information into the denoising process, positioning our method favorably against existing techniques. More comparison results are shown in Sec. A.4, Sec. A.5 and Sec. A.6 of the supplementary materials. + +Style Image Generation. We demonstrate the results of the stylization results in Fig. 5, which compare our method with state-of-the-art tuning-free identity-preserving personalization methods. The style of the synthesis images include anime style, comic book style, and line art style. According to the stylization results, the IP-Adapter [40] and IP-Adapter-Face fails to depict the desired style in the prompt and the style of generation results always obey the tone in the reference image. The training pipeline of the IP-Adapter and IP-Adapter-Face entangles the text embedding and image embedding which leads to the distortion of the text embeddings space. FastComposer [39] also fails in stylization generation and shows undesired artifacts. PhotoMaker [17] achieves a better semantic consistency and stylization, but the identity fidelity is still unsatisfactory. In contrast, our method achieves high identity fidelity, appealing semantic consistency, and precise stylization meantime. + +# 4.3 Ablation Study + +In this section, we begin by conducting ablation studies to assess the influence of our identity-enhanced training and mixed attention mechanism. Furthermore, we conduct style image generation to evaluate the effectiveness of our AdaIN-mean operation in regulating the style of the generated images. More ablation studies are demonstrated in Sec. A.3 of the supplementary materials. + +Ablation of identity-enhanced training. In Fig. 6 and Table 1, we compare our method with "Ours (w/o identity-enhanced training)", which is implemented using an identity-semantics entangled training strategy. This strategy utilizes text-image pairs during training and activates cross-attention modules for text embeddings within the original U-Net model. It is noteworthy that both methods + +![](images/2337fa1f5f28299abc98a7e44b4eb7e3ab6aca5208d2f2f8925f440af4c5d3ad.jpg) +a woman wearing purple sweater, facing and smiling to the camera, in a forest + +![](images/8bcc8b60ad7cba6d6f482b5c449c104eb748bcb263cf27be832b956c32954ca2.jpg) +a photo of a man wearing a colorful shirt, in the living room + +![](images/ab5fdf84a765b7d1e7e633e134b54e7ef4a12f2ec0abe906dd12b8c51b8bc05e.jpg) +ID image + +![](images/10a0c5816810dbe015c626d4041327b6f61c4f147cc8463e6962a9f478caf51c.jpg) + +![](images/df1a192a31f1f6bad742e8cdef1e7990cb0235d6f9f194f3ec1024d796f1fdb3.jpg) + +![](images/c9720f8d102dd20dd49713e63b974affa08146157b85f3616678df783acddf2b.jpg) + +Ours +![](images/939905821752ac9b1156d38a0bb51381a3134c39e960af462d2ae55427e69975.jpg) +a woman wearing a red hat, smiling and facing to the camera, sunset, on the beach +Ours (w/o identity-enhanced training) + +![](images/3b164df0b3cddb42c19fe4e4042690a18573786ea330f15b130117cb7d17a85d.jpg) + +![](images/f0297b2b1e480601e5270170046b85ebda0b05559abe54dc644e74df6863dc73.jpg) + +![](images/49a3475032c54d73935d3f51366f60732f52fbc8699a4b055164cc5ed54c7923.jpg) +Ours (M-A $\Rightarrow$ MU-A) + +Fig. 6: Ablation study of our identity-enhanced training and mixed attention (M-A). It is evident that identity-enhanced training significantly improves the identity fidelity, and mixed-attention mechanism enhances semantic consistency compared to mutual attention (MU-A) approach [5]. +![](images/8fbf6616ec2483404a4b7d2399eee2847d481cebb3030e2bba8de7834fe9a83c.jpg) +awearing a blue clothing, facing and smiling to the snow. A snowy, snowy, in an anime style + +ID Image +Fig. 7: Ablation study of our AdaIN-mean operation. The results show that AdaIN-mean play a crucial role in style image generation. Compare to AdaIN [10] module, our AdaIN-mean helps to achieve a higher identity fidelity. +![](images/9a1c3b99492fc94af69f9a0838f250c115911615070ac6db8a20e09586ce23e7.jpg) +a woman wearing a purple dress and smiling to the camera, on a beach, sunset, comic book style + +![](images/8d1c57c74125779832a2b402dbf91d09755bd61eca819c7b94c846fadd7592a2.jpg) + +![](images/e46e5ccadcd2615e9b1fe5dadd204d460988396d27a7e752dc92d9cebc2e94e5.jpg) +Ours + +![](images/0bcf974314745e99ea2c32b5ec3637174b3d274b3dd1ec56afad195a9dfe6e6f.jpg) + +![](images/22a87f3a44b42531fb9abd58dadcbe61d4125c02c416795ca24e670f5f0ef109.jpg) +Ours (AdaIN-mean $\Rightarrow$ AdaIN) + +share the same inference processing. The qualitative comparison demonstrates that our identity-enhanced training notably enhances identity fidelity. + +Ablation of mixed attention mechanism. To assess the effectiveness of our proposed mixed attention (M-A) mechanism, we compare our method with "Ours (w/o M-A)" and "Ours (M-A $\Rightarrow$ MU-A)" in Fig. 6 and Table 1. Mutual self-attention (MU-A) [5] converts the existing self-attention into 'cross-attention' for consistent image editing, where the crossing operation happens in the self-attention of two related diffusion processes. The results show that our mixed attention mechanism demonstrates superior ability to improve semantic consistency while maintaining identity fidelity. + +Ablation of AdaIN-mean operation. To assess the effectiveness of our proposed AdaIN-mean operation, we compare our Infinite-ID model with variations, namely Ours (w/o AdaIN-mean) and Ours (AdaIN-mean $\Rightarrow$ AdaIN), as depicted + +in Fig. 7. The results reveal that: i) Ours (w/o AdaIN-mean) exhibits superior ID fidelity compared to Infinite-ID but fails to achieve style consistency with the text prompt; ii) Both our AdaIN-mean and AdaIN modules successfully achieve style consistency with the text prompt, yet AdaIN-mean maintains better ID fidelity than AdaIN. In conclusion, our proposed AdaIN-mean operation facilitates precise stylization while concurrently preserving ID fidelity. + +# 5 Conclusion and limitations. + +In this paper, we introduce Infinite-ID, an innovative identity-preserved personalization method designed to meet the requirement of identity (ID) fidelity and semantic consistency of the text prompt, all achievable with just one reference image and completed within seconds. Infinite-ID comprises three key components: the identity-enhanced training, mixed attention mechanism, and adaptive mean normalization (AdaIN-mean). Through extensive experimentation, our results illustrate that Infinite-ID outperforms baseline methods, delivering strong ID fidelity, superior generation quality, and precise semantic consistency in both raw photo generation and style image generation tasks. However, it's important to note that our method lacks multi-object personalization capability. Moreover, artifacts may occur when the human face occupies only a small portion of the entire image, attributed to limitations inherent in the original diffusion model. Furthermore, the image encoders employed in the identity-preserved personalization remain to be further explored. + +# 6 Copyright of Training Data and Ethical Considerations. + +Copyright of Training Data. We construct our training dataset from the open-source public image dataset, including LAION-2B dataset, the LAION-Face dataset and few high-quality images collected from the Website5 which supports the CC0 (Creative Commons Zero) 1.0 Universal license6. + +Ethical Considerations. We recognize the ethical considerations about improper use of such high-quality human photo synthesis techniques. Therefore, we emphasize the significance of establishing and complying to ethical guidelines [4] and using this technology responsibly. We hope that our work will drive further discussions and research on the security and ethical use of the human photo generation in computer vision. Moreover, we will employ the watermark and safety checker (NSFW classifiers) to attempt to avoid inappropriate content generation. + +# Acknowledgements + +This research was funded by the National Natural Science Foundation of China (U19B2044), supported by the Fundamental Research Funds for the Central Universities (WK2100000045). + +# References + +1. Achlioptas, P., Benetatos, A., Fostiropoulos, I., Skourtis, D.: Stellar: Systematic evaluation of human-centric personalized text-to-image methods. arXiv preprint arXiv:2312.06116 (2023) +2. Avrahami, O., Aberman, K., Fried, O., Cohen-Or, D., Lischinski, D.: Break-a-scene: Extracting multiple concepts from a single image. arXiv preprint arXiv:2305.16311 (2023) +3. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf 2, 3 (2023) +4. Biden, J.R.: Executive order on the safe, secure, and trustworthy development and use of artificial intelligence (2023) +5. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. arXiv preprint arXiv:2304.08465 (2023) +6. Chen, L., Zhao, M., Liu, Y., Ding, M., Song, Y., Wang, S., Wang, X., Yang, H., Liu, J., Du, K., et al.: Photovverse: Tuning-free image customization with text-to-image diffusion models. arXiv preprint arXiv:2309.05793 (2023) +7. Chen, X., Huang, L., Liu, Y., Shen, Y., Zhao, D., Zhao, H.: Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481 (2023) +8. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618 (2022) +9. Gal, R., Arar, M., Atzmon, Y., Bermano, A.H., Chechik, G., Cohen-Or, D.: Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG) 42(4), 1-13 (2023) +0. Hertz, A., Voynov, A., Fruchter, S., Cohen-Or, D.: Style aligned image generation via shared attention. arXiv preprint arXiv:2312.02133 (2023) +1. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: 2017 IEEE International Conference on Computer Vision (ICCV) (Oct 2017). https://doi.org/10.1109/iccv.2017.167, http://dx.doi.org/10.1109/iccv.2017.167 +2. Inoue, N., Kikuchi, K., Simo-Serra, E., Otani, M., Yamaguchi, K.: Layoutdm: Discrete diffusion model for controllable layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10167-10176 (2023) +3. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023) +4. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439 (2023) + +15. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1931-1941 (2023) +16. Li, D., Li, J., Hoi, S.C.: Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720 (2023) +17. Li, Z., Cao, M., Wang, X., Qi, Z., Cheng, M.M., Shan, Y.: Photomaker: Customizing realistic human photos via stacked id embedding. arXiv preprint arXiv:2312.04461 (2023) +18. Li, Z., Usman, M., Tao, R., Xia, P., Wang, C., Chen, H., Li, B.: A systematic survey of regularization and normalization in gans. ACM Computing Surveys 55(11), 1-37 (2023) +19. Li, Z., Wang, C., Zheng, H., Zhang, J., Li, B.: Fakeclr: Exploring contrastive learning for solving latent discontinuity in data-efficient gans. In: European Conference on Computer Vision. pp. 598-615. Springer (2022) +20. Li, Z., Xia, P., Tao, R., Niu, H., Li, B.: A new perspective on stabilizing gans training: Direct adversarial training. IEEE Transactions on Emerging Topics in Computational Intelligence 7(1), 178-189 (2022) +21. Ma, J., Liang, J., Chen, C., Lu, H.: Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. arXiv preprint arXiv:2307.11410 (2023) +22. Peng, X., Zhu, J., Jiang, B., Tai, Y., Luo, D., Zhang, J., Lin, W., Jin, T., Wang, C., Ji, R.: Portrait booth: A versatile portrait model for fast identity-preserved personalization. arXiv preprint arXiv:2312.06354 (2023) +23. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023) +24. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +25. Radford, A., Kim, J., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Amanda, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. Cornell University - arXiv, Cornell University - arXiv (Feb 2021) +26. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) +27. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +28. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) +29. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023) + +30. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) +31. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015) +32. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) +33. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021) +34. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110 (2023) +35. Wang, Q., Bai, X., Wang, H., Qin, Z., Chen, A.: Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519 (2024) +36. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848 (2023) +37. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7623-7633 (2023) +38. Wu, Y., Li, Z., Wang, C., Zheng, H., Zhao, S., Li, B., Tao, D.: Domain remodulation for few-shot generative domain adaptation. Advances in Neural Information Processing Systems 36 (2024) +39. Xiao, G., Yin, T., Freeman, W.T., Durand, F., Han, S.: Fastcomposer: Tuning-free multi-subject image generation with localized attention. arXiv preprint arXiv:2305.10431 (2023) +40. Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721 (2023) +41. Yuan, G., Cun, X., Zhang, Y., Li, M., Qi, C., Wang, X., Shan, Y., Zheng, H.: Inserting anybody in diffusion models via celeb basis. In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 72958-72982. Curran Associates, Inc. (2023), https://proceedings.neurips.cc/paper_files/paper/2023/file/e6d37cc5723e810b793c834bc6647cf-Paper-Conference.pdf +42. Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djolonga, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., et al.: A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867 (2019) +43. Zhang, C., Zhang, C., Zhang, M., Kweon, I.S.: Text-to-image diffusion model in generative ai: A survey. arXiv preprint arXiv:2303.07909 (2023) +44. Zhang, K., Zhang, Z., Li, Z., Qiao, Y.: Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters 23(10), 1499-1503 (2016) + +45. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023) +46. Zheng, Y., Yang, H., Zhang, T., Bao, J., Chen, D., Huang, Y., Yuan, L., Chen, D., Zeng, M., Wen, F.: General facial representation learning in a visual-linguistic manner. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18697-18709 (2022) \ No newline at end of file diff --git a/2024/Infinite-ID_ Identity-preserved Personalization via ID-semantics Decoupling Paradigm/images.zip b/2024/Infinite-ID_ Identity-preserved Personalization via ID-semantics Decoupling Paradigm/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c54f45b644ea31bb1d67e69d3e2cc604a6168bbd --- /dev/null +++ b/2024/Infinite-ID_ Identity-preserved Personalization via ID-semantics Decoupling Paradigm/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d4d38278f076deb22f48050872b94a35af367f4159cefd0de811e2eb5bd9c997 +size 733755 diff --git a/2024/Infinite-ID_ Identity-preserved Personalization via ID-semantics Decoupling Paradigm/layout.json b/2024/Infinite-ID_ Identity-preserved Personalization via ID-semantics Decoupling Paradigm/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a19ab05edafc04ec6d2202754631fb87621c0e4f --- /dev/null +++ b/2024/Infinite-ID_ Identity-preserved Personalization via ID-semantics Decoupling Paradigm/layout.json @@ -0,0 +1,9146 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "type": "text", + "content": "Infinite-ID: Identity-preserved Personalization via ID-semantics Decoupling Paradigm" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "spans": [ + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "text", + "content": "Yi Wu" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "text", + "content": ", Ziqiang Li" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "inline_equation", + "content": "^{1,2\\star}" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "text", + "content": ", Heliang Zheng" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "text", + "content": ", Chaoyue Wang" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "inline_equation", + "content": "^{3\\star\\star}" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "text", + "content": ", and Bin Li" + }, + { + "bbox": [ + 140, + 167, + 472, + 181 + ], + "type": "inline_equation", + "content": "^{1\\star\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 192, + 189, + 422, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 189, + 422, + 201 + ], + "spans": [ + { + "bbox": [ + 192, + 189, + 422, + 201 + ], + "type": "text", + "content": "1 University of Science and Technology of China, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 168, + 201, + 444, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 201, + 444, + 213 + ], + "spans": [ + { + "bbox": [ + 168, + 201, + 444, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 168, + 201, + 444, + 213 + ], + "type": "text", + "content": " Nanjing University of Information Science and Technology, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 228, + 212, + 386, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 212, + 386, + 223 + ], + "spans": [ + { + "bbox": [ + 228, + 212, + 386, + 223 + ], + "type": "text", + "content": "3 The University of Sydney, Australia" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 191, + 224, + 422, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 191, + 224, + 422, + 234 + ], + "spans": [ + { + "bbox": [ + 191, + 224, + 422, + 234 + ], + "type": "text", + "content": "wuyi2021@mail.ustc.edu.cn, iceli@mail.ustc.edu.cn," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 247, + 236, + 365, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 236, + 365, + 245 + ], + "spans": [ + { + "bbox": [ + 247, + 236, + 365, + 245 + ], + "type": "text", + "content": "zhenghl@mail.ustc.edu.cn," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 207, + 247, + 406, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 247, + 406, + 256 + ], + "spans": [ + { + "bbox": [ + 207, + 247, + 406, + 256 + ], + "type": "text", + "content": "chaoyue.wang@outlook.com, binli@ustc.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 285, + 455, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 285, + 455, + 517 + ], + "spans": [ + { + "bbox": [ + 159, + 285, + 455, + 517 + ], + "type": "text", + "content": "Abstract. Drawing on recent advancements in diffusion models for text-to-image generation, identity-preserved personalization has made significant progress in accurately capturing specific identities with just a single reference image. However, existing methods primarily integrate reference images within the text embedding space, leading to a complex entanglement of image and text information, which poses challenges for preserving both identity fidelity and semantic consistency. To tackle this challenge, we propose Infinite-ID, an ID-semantics decoupling paradigm for identity-preserved personalization. Specifically, we introduce identity-enhanced training, incorporating an additional image cross-attention module to capture sufficient ID information while deactivating the original text cross-attention module of the diffusion model. This ensures that the image stream faithfully represents the identity provided by the reference image while mitigating interference from textual input. Additionally, we introduce a feature interaction mechanism that combines a mixed attention module with an AdaIN-mean operation to seamlessly merge the two streams. This mechanism not only enhances the fidelity of identity and semantic consistency but also enables convenient control over the styles of the generated images. Extensive experimental results on both raw photo generation and style image generation demonstrate the superior performance of our proposed method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 159, + 526, + 453, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 526, + 453, + 549 + ], + "spans": [ + { + "bbox": [ + 159, + 526, + 453, + 549 + ], + "type": "text", + "content": "Keywords: Personalized Text-to-image Generation, Stable Diffusion, Identity-preserved Personalization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 586, + 230, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 230, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 230, + 600 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 612, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 612, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 612, + 482, + 636 + ], + "type": "text", + "content": "Human photo synthesis [19,38] has experienced notable advancements, particularly with the introduction of large text-to-image diffusion models such as Stable" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 642, + 351, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 642, + 351, + 654 + ], + "spans": [ + { + "bbox": [ + 132, + 642, + 351, + 654 + ], + "type": "text", + "content": "* First two authors contributed equally to this work." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 654, + 241, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 654, + 241, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 654, + 241, + 665 + ], + "type": "text", + "content": "\\*\\* Corresponding Authors." + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 113, + 473, + 264 + ], + "blocks": [ + { + "bbox": [ + 140, + 113, + 473, + 264 + ], + "lines": [ + { + "bbox": [ + 140, + 113, + 473, + 264 + ], + "spans": [ + { + "bbox": [ + 140, + 113, + 473, + 264 + ], + "type": "image", + "image_path": "4e341a16035d1d8bddf81feb9456875b5cb78d771d719f6b70dffc7434288e56.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 272, + 482, + 307 + ], + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 307 + ], + "type": "text", + "content": "Fig. 1: With just a single reference image, our Infinite-ID framework excels in synthesizing high-quality images while maintaining superior identity fidelity and text semantic consistency in various styles." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 329, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 450 + ], + "type": "text", + "content": "Diffusion (SD) [27], Imagen [30], and DALL-E 3 [3]. Benefit from personalized text-to-image generation, recent researches focus on Identity-preserved personalization. This specialized area aims to produce highly customized photos that faithfully reflect a specific identity in novel scenes, actions, and styles, drawing inspiration from one or more reference images. This task has garnered considerable attention, leading to the development of numerous applications, including personalized AI portraits and virtual try-on scenarios. In the context of Identity-preserved personalization, the emphasis is placed on maintaining the invariance of human facial identity (ID), requiring a heightened level of detail and fidelity compared to more general styles or objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": "Recent tuning-free methods exhibit promise for large-scale deployment, yet they face a notable challenge in balancing the trade-off between the fidelity of identity representation (ID fidelity) and the consistency of semantic understanding conveyed by the text prompt. This challenge arises due to the inherent entanglement of image and text information. Typically, tuning-free methods extract ID information from reference images and integrate it into the semantic information in two distinct ways. The first type, exemplified by PhotoMaker [17], incorporates text information with ID details in the text embedding space of the text encoder. While this merging approach aids in achieving semantic consistency, it compresses image features into the text embedding space, thereby weakening the ID information of the image and compromising identity fidelity. The second type, demonstrated by IP-Adapter [40], directly injects ID information into the U-Net of the diffusion model through an additional trainable cross-attention module. Although this approach aims to enhance the strength of ID information for improved fidelity, it tends to favor the image branch during training, consequently weakening the text branch and compromising semantic consistency. In summary, existing methods entangle image and text information, resulting in a significant trade-off between ID fidelity and semantic consistency." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Yi Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "To address the entanglement between image and text information, we propose Infinite-ID, an innovative approach to personalized text-to-image generation. Our method tackles the trade-off between maintaining high fidelity of identity and ensuring semantic consistency of the text prompt by implementing the ID-semantics decoupling paradigm. Specifically, we adopt identity-enhanced training that introduces an additional image cross-attention module to capture sufficient ID information and deactivate the original text cross-attention module to avoid text interference during training stage. Accordingly, our method can faithfully capture identity information from reference image, significantly improving the ID fidelity. Additionally, we employ a novel feature interaction mechanism that leverages a mixed attention module and an AdaIN-mean operation to effectively merge text information and identity information. Notably, our feature interaction mechanism not only preserves both identity and semantic details effectively but also enables convenient control over the styles of the generated images (as depicted in Fig. 1). Our contributions are summarized as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 303, + 480, + 433 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 303, + 480, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 303, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 303, + 480, + 348 + ], + "type": "text", + "content": "- We propose a novel ID-semantics decoupling paradigm to resolve the entanglement between image and text information, acquiring a remarkable balance between ID fidelity and semantic consistency in Identity-preserved personalization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 351, + 480, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 351, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 351, + 480, + 396 + ], + "type": "text", + "content": "- We propose a novel feature interaction mechanism incorporating a mixed attention module and an AdaIN-mean operation to effectively merge ID information and text information and also conveniently control the styles of the generated image in diffusion models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 399, + 480, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 480, + 433 + ], + "type": "text", + "content": "- Experimental results demonstrate the excellent performance of our proposed method as compared with current state-of-the-art methods on both raw photo generation and style image generation." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 453, + 242, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 242, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 242, + 464 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 478, + 320, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 478, + 320, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 478, + 320, + 490 + ], + "type": "text", + "content": "2.1 Text-to-image Diffusion Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Text-to-image diffusion models, such as those explored in [3, 13, 27, 30, 43, 45], have garnered significant attention due to their impressive image generation capabilities. Current research endeavors aim to further enhance these models along multiple fronts, including the utilization of high-quality and large-scale datasets [18, 20, 32, 33], refinements to foundational architectures [3, 27, 30], and advancements in controllability [12, 29, 45]. Present iterations of text-to-image diffusion models typically follow a two-step process: first, encoding the text prompt using pre-trained text encoders such as CLIP [24] or T5 [26], and then utilizing the resulting text embedding as a condition for generating corresponding images through the diffusion process. Notably, the widely adopted Stable Diffusion model [27] distinguishes itself by executing the diffusion process in latent space instead of the original pixel space, leading to significant reductions in computation and time costs. An important extension to this framework is the Stable Diffusion XL (SDXL) [23], which enhances performance by scaling" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "Infinite-ID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": "up the U-Net architecture and introducing an additional text encoder. Thus, our proposed method builds upon the SDXL. However, our method can also be extended to other text-to-image diffusion models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 178, + 332, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 178, + 332, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 332, + 190 + ], + "type": "text", + "content": "2.2 Identity-preserved personalization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 207, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 207, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 482, + 411 + ], + "type": "text", + "content": "Identity-preserved personalization aims to generate highly customized photos that accurately reflect a specific identity across various scenes, actions, and styles, drawing inspiration from one or more reference images. Initially, tuning-based methods, exemplified by DreamBooth [29], Textual Inversion [8] and Celeb-Basis [41], employ images of the same identity (ID) to fine-tune the model. While these methods yield results with high fidelity in preserving facial identity (ID), a significant drawback emerges: the customization of each ID necessitates a time investment of 10-30 minutes [15], consuming substantial computing resources and time. This limitation poses a significant obstacle to large-scale deployment in commercial applications. Consequently, recent advancements in tuning-free methods [6,7,9,16,17,21,35,36,39,40] have been introduced to streamline the generation process. These methods specifically leverage the construction of a vast amount of domain-specific data and the training of an encoder or hyper-network to represent input ID images as embeddings or LoRA weights within the model. Post-training, users need only input an image of the ID for customization, enabling personalized generation within seconds during the inference phase. These tuning-free methods typically contain two distinct manners." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 412, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 581 + ], + "type": "text", + "content": "On one hand, methods [1, 17, 22, 39] incorporate text information alongside identity details within the text embedding space of the text encoder. For example, PhotoMaker [17] extracts identity embeddings from single or multiple reference images and merges them with corresponding class embeddings (e.g., \"man\" and \"woman\") in the text embedding space of the text encoder. While this stacking operation aids in achieving semantic consistency, it compresses image features into the text embedding space, leading to compromised identity fidelity. On the other hand, some studies [6, 36, 40] directly integrate identity information into the U-Net of the diffusion model. IP-Adapter [40] distinguishes itself by incorporating a additional cross-attention layer for each existing cross-attention layer within the original UNet model. This approach merges identity information with semantic details directly within the U-net but leads to distortion of the semantic space of the U-Net model. Consequently, this compromises the semantic consistency of the text prompt." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "In summary, existing methods entangle the identity and text information, leading to a significant trade-off between ID fidelity and semantic consistency. To mitigate these limitations, we propose an identity-enhanced training to capture ID and text information separately. Moreover, we design an effective feature interaction mechanism leveraging a mixed attention module and an AdaIN-mean operation to preserve both identity and semantic details while also enabling convenient control over the styles of generated images." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Yi Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 348, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 348, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 348, + 128 + ], + "type": "text", + "content": "2.3 Attention Control in Diffusion model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 304 + ], + "type": "text", + "content": "Previous studies have investigated various attention control techniques within diffusion models. Hertz et al. [10] employed a shared attention mechanism, concatenating and applying an AdaIN module on the key and value between reference and synthesis images within the self-attention layer to ensure style-consistent image generation using a reference style. Cao et al. [5] utilized a mutual self-attention approach to achieve consistent image generation and non-rigid image editing, wherein the key and value of the synthesis image were replaced with those of the reference image within the self-attention layers of the diffusion model. Similarly, Shi et al. [34] proposed a method termed reference attention, enabling consistent multi-view generation of target objects by concatenating the key and value features between the condition signal and the synthesis image in the self-attention layers. Wang et al. [39] and Avrahami et al. [2] exploited attention maps within the cross-attention layers to guide the optimization process towards disentangling learned concepts in personalized generation tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 321, + 202, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 202, + 333 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 202, + 333 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 346, + 228, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 346, + 228, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 346, + 228, + 358 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": "Stable Diffusion XL. Our method builds upon Stable Diffusion XL [23], comprising three core components: a Variational AutoEncoder (VAE) denoted as " + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\xi(\\cdot)" + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": ", a conditional U-Net [28] represented by " + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": ", and two pre-trained text encoders [25] denoted as " + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\Theta_1(\\cdot)" + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\Theta_2(\\cdot)" + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": ". Specifically, given a training image " + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": " and its corresponding text prompt " + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": ", the VAE encoder " + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": " transforms " + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": " from its original space " + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "inline_equation", + "content": "R^{H\\times W\\times 3}" + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": " to a compressed latent representation " + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "inline_equation", + "content": "z_0 = \\xi(x_0)" + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "inline_equation", + "content": "z_0 \\in R^{h\\times w\\times c}" + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": " denotes the latent dimension. Subsequently, the diffusion process operates within this compressed latent space to conserve computational resources and memory. Once the two text encoders process the text prompt " + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": " into a text embedding " + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "inline_equation", + "content": "c = \\mathrm{Concat}(\\Theta_1(T), \\Theta_2(T))" + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": ", the conditional U-Net " + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": " predicts the noise " + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": " based on the current timestep " + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": "-th latent representation " + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": ", and the text embedding " + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 365, + 482, + 510 + ], + "type": "text", + "content": ". The training objective is formulated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 211, + 521, + 481, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 521, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 211, + 521, + 481, + 536 + ], + "type": "interline_equation", + "content": "L _ {\\text {d i f f u s i o n}} = E _ {z _ {t}, t, c, \\epsilon \\in N (0, I)} [ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, c) | | _ {2} ^ {2} ]. \\tag {1}", + "image_path": "e833903b4e9a09d173df724538dbee7c10c480b6171a0c08ba32e56d855c3e76.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 539, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 482, + 575 + ], + "type": "text", + "content": "Attention mechanism in diffusion models. The fundamental unit of the stable diffusion model comprises a resblock, a self-attention layer, and a cross-attention layer. The attention mechanism is represented as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 230, + 583, + 481, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 583, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 230, + 583, + 481, + 611 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t n} (Q, K, V) = \\operatorname {S o f t m a x} (\\frac {Q K ^ {T}}{\\sqrt {d}}) V, \\tag {2}", + "image_path": "25954590b58c0228f007d5efb372766807d6869ef80df252c35f4a16bcadbcd6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " denotes the query feature projected from the spatial features generated by the preceding resblock, " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " represent the key and value features projected from the same spatial features as the query feature (in self-attention) or the text embedding extracted from the text prompt (in cross-attention)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "Infinite-ID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 473, + 315 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 473, + 315 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 473, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 473, + 315 + ], + "type": "image", + "image_path": "bc8b81633e4218c09ab22f6c46b8e81ccab461952e9aab4409f34a539fb94514.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 323, + 482, + 390 + ], + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 390 + ], + "type": "text", + "content": "Fig. 2: Framework of ID-semantics Decoupling Paradigm. In the training phase, we adopt Face embeddings extractor to extract rich identity information and identity-enhanced training for faithfully representing the identity provided by the reference image while mitigating interference from textual input. In the inference stage, a mixed attention module is introduced to replace the original self-attention mechanism within the denoising U-Net model, facilitating the fusion of both identity and text information." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 415, + 227, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 415, + 227, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 227, + 427 + ], + "type": "text", + "content": "3.2 Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "type": "text", + "content": "Overview. In this section, we introduce our ID-semantics decoupling paradigm, as illustrated in Fig. 2, which effectively addresses the severe trade-off between high-fidelity identity and semantic consistency within identity-preserved personalization. Subsequently, we present our mixed attention mechanism, depicted in Fig. 3, designed to seamlessly integrate ID information and semantic information within the diffusion model during the inference stage. Additionally, we utilize an adaptive mean normalization (AdaIN-mean) operation to precisely align the style of the synthesized image with the desired style prompts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "ID-semantics Decoupling Paradigm. To faithfully capture high-fidelity identity, we implement a novel identity-enhanced strategy during the training stage, as depicted in Fig. 2. Diverging from conventional methods [6,17,40] that utilize text-image pairs for training, we opt to exclude the text prompt input and deactivate cross-attention modules for text embeddings within the U-Net model. Instead, we establish a training pair consisting of an ID image, where the face is aligned to extract identity information, and denoising image (" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " in Fig. 2). Both the denoising image used for training and the ID image belong to the same individual, but they vary in factors such as viewpoints and facial expressions. This approach fosters a more comprehensive learning process [17]. We adopt Face embeddings extractor to accurately capture and leverage identity information from" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Yi Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 116, + 302, + 321 + ], + "blocks": [ + { + "bbox": [ + 143, + 116, + 302, + 321 + ], + "lines": [ + { + "bbox": [ + 143, + 116, + 302, + 321 + ], + "spans": [ + { + "bbox": [ + 143, + 116, + 302, + 321 + ], + "type": "image", + "image_path": "77312140525ed59cbd83583c6bab7affe65f8d203ba21ec5e32ac9b6e5dbcb74.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 323, + 262, + 334 + ], + "lines": [ + { + "bbox": [ + 182, + 323, + 262, + 334 + ], + "spans": [ + { + "bbox": [ + 182, + 323, + 262, + 334 + ], + "type": "text", + "content": "(a) Mixed Attention" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 345, + 480, + 411 + ], + "lines": [ + { + "bbox": [ + 130, + 345, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 480, + 411 + ], + "type": "text", + "content": "Fig. 3: Mixed attention mechanism. On the left side, we employ mixed attention to fuse identity and text information. This involves concatenating their respective key and value features and subsequently applying mixed attention, where identity features are updated based on the concatenated key and value features. On the right side, for style merging, we introduce an additional AdaIN-mean operation (as depicted in Eq. 8) to the concatenated key and value features." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 116, + 471, + 321 + ], + "blocks": [ + { + "bbox": [ + 312, + 116, + 471, + 321 + ], + "lines": [ + { + "bbox": [ + 312, + 116, + 471, + 321 + ], + "spans": [ + { + "bbox": [ + 312, + 116, + 471, + 321 + ], + "type": "image", + "image_path": "0b4fd8e1db8bf40fe9c7f7a5c537cbda2c1af8d258d98ef0d3b1ffded6a26f95.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 323, + 466, + 334 + ], + "lines": [ + { + "bbox": [ + 313, + 323, + 466, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 466, + 334 + ], + "type": "text", + "content": "(b) Mixed Attention with AdaIN-mean" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 434, + 480, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 480, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 480, + 470 + ], + "type": "text", + "content": "the input ID image. Additionally, to seamlessly integrate identity information into the denoising U-Net model, we introduce an extra trainable cross-attention mechanism (" + }, + { + "bbox": [ + 130, + 434, + 480, + 470 + ], + "type": "inline_equation", + "content": "K_{id}'" + }, + { + "bbox": [ + 130, + 434, + 480, + 470 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 434, + 480, + 470 + ], + "type": "inline_equation", + "content": "V_{id}'" + }, + { + "bbox": [ + 130, + 434, + 480, + 470 + ], + "type": "text", + "content": " in Fig. 2) for image embeddings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 471, + 480, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 480, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 480, + 541 + ], + "type": "text", + "content": "Throughout the training phase, we exclusively optimize the parameters associated with the face mapper, CLIP mapper, and the image cross-attention module, while keeping the parameters of the pre-trained diffusion model fixed. The optimization loss closely resembles the original diffusion loss formulation (as delineated in Eq. 1), with the sole distinction being the shift from a text condition to an identity condition as the conditional input." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 200, + 548, + 480, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 548, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 200, + 548, + 480, + 563 + ], + "type": "interline_equation", + "content": "L _ {\\text {d i f f u s i o n}} = E _ {z _ {t}, t, c _ {i d}, \\epsilon \\in N (0, I)} [ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, c _ {i d}) | | _ {2} ^ {2} ], \\tag {3}", + "image_path": "223d99678925886d067be7c79498453251a9a23055709a0c0f3ef7534fc34b80.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 410, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 410, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 410, + 582 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 130, + 570, + 410, + 582 + ], + "type": "inline_equation", + "content": "c_{id}" + }, + { + "bbox": [ + 130, + 570, + 410, + 582 + ], + "type": "text", + "content": " is the identity embeddings of the input ID image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "type": "text", + "content": "During the inference phase, we align the desired face within the ID image to afford identity information. Following the approach adopted during the training phase, we utilize a face recognition backbone and a CLIP image encoder to extract identity features from the aligned face image. Leveraging the trained face mapper, clip mapper, and image cross-attention mechanisms, these identity features are seamlessly integrated into the denoising U-Net model. Subsequently, we compute the key and value features for self-attention in the original stable" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "Infinite-ID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": "diffusion model, considering only the text prompt input. These text key and value features are instrumental in the mixed attention process (illustrated in Fig. 3), facilitating the fusion of text and identity information. Moreover, to further augment the text information during the denoising process, we incorporate the original text cross-attention mechanism, integrating the resulting text hidden states with the output image hidden states obtained from the image cross-attention module." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 428 + ], + "type": "text", + "content": "Face Embeddings Extractor. We adopt a multifaceted approach by incorporating pre-trained models to extract facial features. Firstly, following the methodologies of prior research, we utilize a pre-trained CLIP image encoder as one of our facial feature extractors. Specifically, we leverage the local embeddings, comprising the last hidden states obtained from the CLIP image encoder, forming a sequence of embeddings with a length of N (N=257 in our implementation). Subsequently, we employ a CLIP mapper to project these image embeddings (with a dimensionality of 1664) to the same dimension as the text features in the pre-trained diffusion model. As elucidated in [40], the features extracted by the CLIP image encoder are instrumental in capturing the structural information pertinent to the identity face within identity-preserved personalization tasks. Additionally, we leverage the backbone of a face recognition model as another facial feature extractor. As highlighted in [40], features extracted by the face recognition backbone are adept at capturing the characteristics associated with human facial features within identity-preserved personalization tasks. More specifically, we utilize the global image embedding derived from the extracted features and subsequently employ a face mapper to align the dimensionality (512 dimensions) of the extracted global image embedding with the dimensionality of the text features in the pre-trained diffusion model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 429, + 483, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 483, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 483, + 453 + ], + "type": "text", + "content": "In summary, the identity embeddings " + }, + { + "bbox": [ + 130, + 429, + 483, + 453 + ], + "type": "inline_equation", + "content": "c_{id}" + }, + { + "bbox": [ + 130, + 429, + 483, + 453 + ], + "type": "text", + "content": " corresponding to the ID image " + }, + { + "bbox": [ + 130, + 429, + 483, + 453 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 429, + 483, + 453 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 181, + 465, + 481, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 465, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 181, + 465, + 481, + 491 + ], + "type": "interline_equation", + "content": "c _ {i d} = \\operatorname {C o n c a t} \\left(\\mathrm {M} _ {\\text {c l i p}} \\left(\\mathrm {E} _ {\\text {c l i p}} (\\mathrm {F A} (x))\\right), \\mathrm {M} _ {\\text {f a c e}} \\left(\\mathrm {E} _ {\\text {f a c e}} (\\mathrm {F A} (x))\\right)\\right), \\tag {4}", + "image_path": "ea060218347202021953490b198f1bf9b7de44db9c1d4487ca89ac2bcc0a682f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 497, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 497, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{Concat}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 497, + 482, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 497, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{M}_{\\mathrm{clip}}(\\cdot)" + }, + { + "bbox": [ + 130, + 497, + 482, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 497, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{\\mathrm{clip}}(\\cdot)" + }, + { + "bbox": [ + 130, + 497, + 482, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 497, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{M}_{\\mathrm{face}}(\\cdot)" + }, + { + "bbox": [ + 130, + 497, + 482, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 497, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{\\mathrm{face}}(\\cdot)" + }, + { + "bbox": [ + 130, + 497, + 482, + 533 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 497, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{FA}(\\cdot)" + }, + { + "bbox": [ + 130, + 497, + 482, + 533 + ], + "type": "text", + "content": " denote the concatenation function, CLIP mapper, CLIP image encoder, face mapper, face recognition backbone, and face alignment module [44], respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Mixed Attention Mechanism. As explored in previous studies [5, 14, 37], the features in the self attention layers play a crucial role in consistency image generation (across-frame in text-to-video works), which indicates these features provide a refined and detailed semantics information. In our study, we extract features from the self-attention layers of the original text-to-image diffusion model to capture rich semantic information, represented as " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "K_{t}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ". We enhance the self-attention mechanism by incorporating it into a mixed attention framework, depicted in Fig. 3 (a). This fusion enables the integration of semantic features " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "(K_{t}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ") with the identity-based features " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "(K_{id}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "V_{id}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "), thereby encapsulating identity information. Through this integration, the mixed attention mechanism seamlessly merges semantic details into the generated features" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Yi Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "across different resolutions. The formulation of the mixed attention mechanism is as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 223, + 145, + 374, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 145, + 374, + 158 + ], + "spans": [ + { + "bbox": [ + 223, + 145, + 374, + 158 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t n} _ {\\operatorname {m i x}} (Q, K, V) \\triangleq \\operatorname {A t t n} (Q, \\hat {K}, \\hat {V})", + "image_path": "f926cca7f5227e90b1b620f08ced2af17c2984f2505eee8af3840d71c0b04de7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 192, + 162, + 480, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 162, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 192, + 162, + 480, + 182 + ], + "type": "interline_equation", + "content": "\\begin{array}{l l} \\text {w . r . t} & \\hat {K} = \\underset {i d} {\\operatorname {C o n c a t}} \\left(K _ {i d}, K _ {t}\\right), \\quad \\hat {V} = \\underset {t} {\\operatorname {C o n c a t}} \\left(V _ {i d}, V _ {t}\\right), \\end{array} \\tag {5}", + "image_path": "91e968af68faf5d4d21e0cf1ccf36b41a0acef88292cffc9418c393968b80668.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 223, + 177, + 352, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 177, + 352, + 192 + ], + "spans": [ + { + "bbox": [ + 223, + 177, + 352, + 192 + ], + "type": "interline_equation", + "content": "K _ {i d} = W _ {k} ^ {i d} Z _ {i d}, \\quad K _ {t} = W _ {k} ^ {t} Z _ {t},", + "image_path": "7165f1e66679379772329ad199384369b01cca18385af4db37f7072d4764b9bd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 223, + 195, + 347, + 208 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 195, + 347, + 208 + ], + "spans": [ + { + "bbox": [ + 223, + 195, + 347, + 208 + ], + "type": "interline_equation", + "content": "V _ {i d} = W _ {v} ^ {i d} Z _ {i d}, \\quad V _ {t} = W _ {v} ^ {t} Z _ {t},", + "image_path": "d962354e84f28c2099252fcae8103a9b6e518a5ec28344be229acd40226bdaa0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 214, + 480, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 480, + 250 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 480, + 250 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 214, + 480, + 250 + ], + "type": "inline_equation", + "content": "Z_{id}" + }, + { + "bbox": [ + 130, + 214, + 480, + 250 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 214, + 480, + 250 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 130, + 214, + 480, + 250 + ], + "type": "text", + "content": " are the corresponding spatial features of generated features and semantic features, respectively. The parameters " + }, + { + "bbox": [ + 130, + 214, + 480, + 250 + ], + "type": "inline_equation", + "content": "W_{k}^{id}" + }, + { + "bbox": [ + 130, + 214, + 480, + 250 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 214, + 480, + 250 + ], + "type": "inline_equation", + "content": "W_{k}^{t}" + }, + { + "bbox": [ + 130, + 214, + 480, + 250 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 214, + 480, + 250 + ], + "type": "inline_equation", + "content": "W_{v}^{id}" + }, + { + "bbox": [ + 130, + 214, + 480, + 250 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 214, + 480, + 250 + ], + "type": "inline_equation", + "content": "W_{v}^{t}" + }, + { + "bbox": [ + 130, + 214, + 480, + 250 + ], + "type": "text", + "content": " correspond to the weights of the corresponding fully connected layers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 251, + 479, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 479, + 285 + ], + "type": "text", + "content": "Cross-attention Merging. To further refine semantic control, we incorporate text features into the identity feature within the cross-attention layers using the following formulation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 182, + 293, + 480, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 293, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 182, + 293, + 480, + 306 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t n} _ {\\text {c r o s s}} (Q, K, V) \\triangleq \\operatorname {A t t n} \\left(Q, K _ {i d} ^ {\\prime}, V _ {i d} ^ {\\prime}\\right) + \\operatorname {A t t n} \\left(Q, K _ {t} ^ {\\prime}, V _ {t} ^ {\\prime}\\right), \\tag {6}", + "image_path": "dd3988031534cb63e8a305422e8d9e4e301916265acdcb2e96518158b843d705.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "inline_equation", + "content": "K_{id}^{\\prime} = \\hat{W}_{k}^{id}c_{id}" + }, + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "inline_equation", + "content": "V_{id}^{\\prime} = \\hat{W}_{v}^{id}c_{id}" + }, + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "inline_equation", + "content": "K_{t}^{\\prime} = \\hat{W}_{k}^{t}c_{t}" + }, + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "inline_equation", + "content": "V_{t}^{\\prime} = \\hat{W}_{v}^{t}c_{t}" + }, + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "inline_equation", + "content": "c_{id}" + }, + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "text", + "content": " are the identity embedding and text embedding, respectively. " + }, + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "inline_equation", + "content": "\\hat{W}_{k}^{id}" + }, + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "inline_equation", + "content": "\\hat{W}_{v}^{id}" + }, + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "inline_equation", + "content": "\\hat{W}_{k}^{t}" + }, + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "inline_equation", + "content": "\\hat{W}_{v}^{t}" + }, + { + "bbox": [ + 130, + 312, + 480, + 361 + ], + "type": "text", + "content": " correspond to the weights of the trainable fully connected layers within cross-attention module." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 363, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 480, + 422 + ], + "type": "text", + "content": "Style Information Merging. Inspired by [10], we propose an adaptive mean normalization (AdaIN-mean) operation to further align the style of the synthesis image with the style prompts. Concretely, we align the key and value features projected from identity features in both mixed attention and cross-attention with the key and value features projected from text features, formulated as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 429, + 480, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 480, + 468 + ], + "type": "interline_equation", + "content": "\\begin{array}{l l l} K _ {i d} = \\operatorname {A d a I N - m} \\left(K _ {i d}, K _ {t}\\right), & V _ {i d} = \\operatorname {A d a I N - m} \\left(V _ {i d}, V _ {t}\\right), & \\text {F o r M i x e d A t t e n t i o n} \\\\ K _ {i d} ^ {\\prime} = \\operatorname {A d a I N - m} \\left(K _ {i d} ^ {\\prime}, K _ {t} ^ {\\prime}\\right), & V _ {i d} ^ {\\prime} = \\operatorname {A d a I N - m} \\left(V _ {i d} ^ {\\prime}, V _ {t} ^ {\\prime}\\right), & \\text {F o r C r o s s A t t e n t i o n} \\end{array} \\tag {7}", + "image_path": "a556cce1a37b7e87d69d285ab91c3d0cc15e675e6c958baceb12d08c6e070db7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 468, + 402, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 468, + 402, + 480 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 402, + 480 + ], + "type": "text", + "content": "where the AdaIN-mean operation (AdaIN-m(\\cdot)) is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 233, + 488, + 480, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 488, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 233, + 488, + 480, + 500 + ], + "type": "interline_equation", + "content": "\\operatorname {A d a I N} - \\mathrm {m} (x, y) = x - \\mu (x) + \\mu (y), \\tag {8}", + "image_path": "d78574dae030b66908034521950cc91d45bbc176e4fb8ebbd86bf572bea2e1b5.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 506, + 479, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 479, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 479, + 530 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 506, + 479, + 530 + ], + "type": "inline_equation", + "content": "\\mu (x)\\in R^{d_k}" + }, + { + "bbox": [ + 130, + 506, + 479, + 530 + ], + "type": "text", + "content": " is the mean of key and value features across different pixels. The mixed attention with AdaIN-mean has been illustrated in Fig. 3 (b)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 547, + 230, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 547, + 230, + 560 + ], + "spans": [ + { + "bbox": [ + 131, + 547, + 230, + 560 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": "After outlining the experimental setup in Sec. 4.1, we conduct a comparative analysis of raw photo generation and style image generation in Sec. 4.2. Ablation studies, highlighting the significance of various components, are presented in Sec. 4.3. Additionally, experiments involving multiple input ID images including identity mixing and identity linear interpolation are detailed in Sec. A.7 of the supplementary materials. For further insights, Sec. A.8 and Sec. A.9 of the supplementary materials provide additional qualitative results for raw photo generation and style photo generation, respectively." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "Infinite-ID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 478, + 365 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 478, + 365 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 478, + 365 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 478, + 365 + ], + "type": "image", + "image_path": "ce66ee3a3d30418f47ba6bf5dc8d7764ec132a6e50227e46137dfc4dbb4ba01d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 374, + 482, + 407 + ], + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 407 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparison on raw photo generation. The results demonstrate that our Infinite-ID consistently maintains identity fidelity and achieves high-quality semantic consistency with just a single input image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 431, + 261, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 261, + 444 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 261, + 444 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 453, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 596 + ], + "type": "text", + "content": "Implementation Details. Our experiments leverage a pre-trained Stable Diffusion XL (SDXL) model [23]. For image encoding, we utilize the OpenCLIP ViT-H/14 [42] and the backbone of ArcFace [11]. The SDXL model consists of 70 cross-attention layers, to each of which we append an additional image cross-attention module. Training is conducted on 16 A100 GPUs for 1 million steps, with a batch size of 4 per GPU. We employ the AdamW optimizer with a fixed learning rate of 1e-4 and weight decay set to 0.01. During inference, we employ the DDIM Sampler with 30 steps and guidance scale is set to 5.0. Training data are sourced from multiple datasets, including the LAION-2B dataset [33], the LAION-Face dataset [46], and few high-quality images collected from the Websites4. We curate a dataset where each individual is represented by multiple photographs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "type": "text", + "content": "Evaluation. We assess the efficacy of our approach in preserving both identity fidelity and semantic consistency. Specifically, we measure identity fidelity, utilizing metrics such as " + }, + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{FaceNet}}" + }, + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "type": "text", + "content": " (measured by FaceNet [31]) and CLIP-I [8]. Identity fidelity is evaluated based on the similarity of detected faces between the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Yi Wu et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 334, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 334, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 334, + 665 + ], + "type": "text", + "content": "4 https://www.pexels.com, https://pixabay.com." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 117, + 478, + 359 + ], + "blocks": [ + { + "bbox": [ + 137, + 117, + 478, + 359 + ], + "lines": [ + { + "bbox": [ + 137, + 117, + 478, + 359 + ], + "spans": [ + { + "bbox": [ + 137, + 117, + 478, + 359 + ], + "type": "image", + "image_path": "01d499f8c7db4e1c3c67dccca8ae2f10d8c2b4af6eeadea587c0d1c24f4ba24e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 369, + 480, + 403 + ], + "lines": [ + { + "bbox": [ + 130, + 369, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 480, + 403 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparison on style image generation. The results demonstrate that our method maintains strong identity fidelity, high-quality semantic consistency, and precise stylization using only a single reference image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 426, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 475 + ], + "type": "text", + "content": "reference image and generated images. Semantic consistency is quantified using CLIP text-image consistency (CLIP-T [24]), which compares the text prompt with the corresponding generated images. More details of evaluation are shown in Sec. A.1 and Sec. A.2 of the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 491, + 330, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 330, + 503 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 330, + 503 + ], + "type": "text", + "content": "4.2 Comparison to Previous Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Raw Photo Generation. We benchmark our Infinite-ID against established identity-preserving personalization approaches. The qualitative outcomes are illustrated in Fig. 4, while quantitative results are provided in Table 1. Notably, all methods are tuning-free, necessitating no test-time adjustments. FastComposer [39] exhibits challenges in maintaining identity fidelity, often presenting undesired artifacts in synthesized images. While IP-Adapter [40] and IP-Adapter-Face [40] demonstrates relatively fewer artifacts, its semantic consistency falls short. This phenomenon arises from the direct fusion of identity information with semantic details within the U-net model, leading to a compromise in semantic consistency. In contrast, PhotoMaker [17] exhibits commendable semantic consistency but falls short in preserving identity fidelity. Leveraging our ID-semantics decoupling paradigm, our method excels in preserving identity fidelity. Furthermore, our mixed attention mechanism effectively integrate the semantic" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "Infinite-ID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 147, + 179, + 464, + 285 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 1: Quantitative comparison. The evaluation metrics encompass CLIP-T, CLIP-I, and " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "M_{\\text{FaceNet}}" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": ". Our approach outperforms other methods in terms of identity fidelity while simultaneously achieving satisfactory semantic consistency. The best result is shown in bold, and the second best is underlined. Additionally, the quantitative comparison of ablation studies have been shown in the gray part." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 147, + 179, + 464, + 285 + ], + "lines": [ + { + "bbox": [ + 147, + 179, + 464, + 285 + ], + "spans": [ + { + "bbox": [ + 147, + 179, + 464, + 285 + ], + "type": "table", + "html": "
CLIP-T ↑CLIP-I ↑M_FaceNet ↑
FastComposer [39]0.2920.8870.556
IP-Adapter [40]0.2740.9050.474
IP-Adapter-Face [40]0.3130.9190.513
PhotoMaker [17]0.3430.8140.502
Ours0.3400.9130.689
Ours (w/o identity-enhanced training)0.3290.8910.593
Ours (w/o Mixed Attention)0.3310.9050.700
Ours (Mixed Attention ⇒ Mutual Attention)0.3160.8080.398
", + "image_path": "2b75eef9a25fc951129ca904fc62075ac9b77aa18a1485e1c02e9c17b8452f5a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 479, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 479, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 479, + 344 + ], + "type": "text", + "content": "information into the denoising process, positioning our method favorably against existing techniques. More comparison results are shown in Sec. A.4, Sec. A.5 and Sec. A.6 of the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 345, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 482, + 501 + ], + "type": "text", + "content": "Style Image Generation. We demonstrate the results of the stylization results in Fig. 5, which compare our method with state-of-the-art tuning-free identity-preserving personalization methods. The style of the synthesis images include anime style, comic book style, and line art style. According to the stylization results, the IP-Adapter [40] and IP-Adapter-Face fails to depict the desired style in the prompt and the style of generation results always obey the tone in the reference image. The training pipeline of the IP-Adapter and IP-Adapter-Face entangles the text embedding and image embedding which leads to the distortion of the text embeddings space. FastComposer [39] also fails in stylization generation and shows undesired artifacts. PhotoMaker [17] achieves a better semantic consistency and stylization, but the identity fidelity is still unsatisfactory. In contrast, our method achieves high identity fidelity, appealing semantic consistency, and precise stylization meantime." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 521, + 238, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 238, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 238, + 534 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "text", + "content": "In this section, we begin by conducting ablation studies to assess the influence of our identity-enhanced training and mixed attention mechanism. Furthermore, we conduct style image generation to evaluate the effectiveness of our AdaIN-mean operation in regulating the style of the generated images. More ablation studies are demonstrated in Sec. A.3 of the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Ablation of identity-enhanced training. In Fig. 6 and Table 1, we compare our method with \"Ours (w/o identity-enhanced training)\", which is implemented using an identity-semantics entangled training strategy. This strategy utilizes text-image pairs during training and activates cross-attention modules for text embeddings within the original U-Net model. It is noteworthy that both methods" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Yi Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 179, + 114, + 223, + 159 + ], + "blocks": [ + { + "bbox": [ + 179, + 114, + 223, + 159 + ], + "lines": [ + { + "bbox": [ + 179, + 114, + 223, + 159 + ], + "spans": [ + { + "bbox": [ + 179, + 114, + 223, + 159 + ], + "type": "image", + "image_path": "2337fa1f5f28299abc98a7e44b4eb7e3ab6aca5208d2f2f8925f440af4c5d3ad.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 119, + 259, + 156 + ], + "lines": [ + { + "bbox": [ + 223, + 119, + 259, + 156 + ], + "spans": [ + { + "bbox": [ + 223, + 119, + 259, + 156 + ], + "type": "text", + "content": "a woman wearing purple sweater, facing and smiling to the camera, in a forest" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 179, + 159, + 223, + 201 + ], + "blocks": [ + { + "bbox": [ + 179, + 159, + 223, + 201 + ], + "lines": [ + { + "bbox": [ + 179, + 159, + 223, + 201 + ], + "spans": [ + { + "bbox": [ + 179, + 159, + 223, + 201 + ], + "type": "image", + "image_path": "8bcc8b60ad7cba6d6f482b5c449c104eb748bcb263cf27be832b956c32954ca2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 164, + 258, + 194 + ], + "lines": [ + { + "bbox": [ + 223, + 164, + 258, + 194 + ], + "spans": [ + { + "bbox": [ + 223, + 164, + 258, + 194 + ], + "type": "text", + "content": "a photo of a man wearing a colorful shirt, in the living room" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 179, + 201, + 217, + 245 + ], + "blocks": [ + { + "bbox": [ + 179, + 201, + 217, + 245 + ], + "lines": [ + { + "bbox": [ + 179, + 201, + 217, + 245 + ], + "spans": [ + { + "bbox": [ + 179, + 201, + 217, + 245 + ], + "type": "image", + "image_path": "ab5fdf84a765b7d1e7e633e134b54e7ef4a12f2ec0abe906dd12b8c51b8bc05e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 247, + 212, + 253 + ], + "lines": [ + { + "bbox": [ + 190, + 247, + 212, + 253 + ], + "spans": [ + { + "bbox": [ + 190, + 247, + 212, + 253 + ], + "type": "text", + "content": "ID image" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 260, + 114, + 304, + 158 + ], + "blocks": [ + { + "bbox": [ + 260, + 114, + 304, + 158 + ], + "lines": [ + { + "bbox": [ + 260, + 114, + 304, + 158 + ], + "spans": [ + { + "bbox": [ + 260, + 114, + 304, + 158 + ], + "type": "image", + "image_path": "10a0c5816810dbe015c626d4041327b6f61c4f147cc8463e6962a9f478caf51c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 304, + 114, + 390, + 158 + ], + "blocks": [ + { + "bbox": [ + 304, + 114, + 390, + 158 + ], + "lines": [ + { + "bbox": [ + 304, + 114, + 390, + 158 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 390, + 158 + ], + "type": "image", + "image_path": "df1a192a31f1f6bad742e8cdef1e7990cb0235d6f9f194f3ec1024d796f1fdb3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 260, + 158, + 390, + 201 + ], + "blocks": [ + { + "bbox": [ + 260, + 158, + 390, + 201 + ], + "lines": [ + { + "bbox": [ + 260, + 158, + 390, + 201 + ], + "spans": [ + { + "bbox": [ + 260, + 158, + 390, + 201 + ], + "type": "image", + "image_path": "c9720f8d102dd20dd49713e63b974affa08146157b85f3616678df783acddf2b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 260, + 201, + 389, + 245 + ], + "blocks": [ + { + "bbox": [ + 223, + 204, + 258, + 245 + ], + "lines": [ + { + "bbox": [ + 223, + 204, + 258, + 245 + ], + "spans": [ + { + "bbox": [ + 223, + 204, + 258, + 245 + ], + "type": "text", + "content": "a woman wearing a red hat, smiling and facing to the camera, sunset, on the beach" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 260, + 201, + 389, + 245 + ], + "lines": [ + { + "bbox": [ + 260, + 201, + 389, + 245 + ], + "spans": [ + { + "bbox": [ + 260, + 201, + 389, + 245 + ], + "type": "image", + "image_path": "939905821752ac9b1156d38a0bb51381a3134c39e960af462d2ae55427e69975.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 246, + 285, + 251 + ], + "lines": [ + { + "bbox": [ + 274, + 246, + 285, + 251 + ], + "spans": [ + { + "bbox": [ + 274, + 246, + 285, + 251 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 301, + 247, + 350, + 265 + ], + "lines": [ + { + "bbox": [ + 301, + 247, + 350, + 265 + ], + "spans": [ + { + "bbox": [ + 301, + 247, + 350, + 265 + ], + "type": "text", + "content": "Ours (w/o identity-enhanced training)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 390, + 114, + 433, + 158 + ], + "blocks": [ + { + "bbox": [ + 390, + 114, + 433, + 158 + ], + "lines": [ + { + "bbox": [ + 390, + 114, + 433, + 158 + ], + "spans": [ + { + "bbox": [ + 390, + 114, + 433, + 158 + ], + "type": "image", + "image_path": "3b164df0b3cddb42c19fe4e4042690a18573786ea330f15b130117cb7d17a85d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 390, + 158, + 433, + 201 + ], + "blocks": [ + { + "bbox": [ + 390, + 158, + 433, + 201 + ], + "lines": [ + { + "bbox": [ + 390, + 158, + 433, + 201 + ], + "spans": [ + { + "bbox": [ + 390, + 158, + 433, + 201 + ], + "type": "image", + "image_path": "f0297b2b1e480601e5270170046b85ebda0b05559abe54dc644e74df6863dc73.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 390, + 201, + 433, + 245 + ], + "blocks": [ + { + "bbox": [ + 390, + 201, + 433, + 245 + ], + "lines": [ + { + "bbox": [ + 390, + 201, + 433, + 245 + ], + "spans": [ + { + "bbox": [ + 390, + 201, + 433, + 245 + ], + "type": "image", + "image_path": "49a3475032c54d73935d3f51366f60732f52fbc8699a4b055164cc5ed54c7923.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 247, + 428, + 258 + ], + "lines": [ + { + "bbox": [ + 394, + 247, + 428, + 258 + ], + "spans": [ + { + "bbox": [ + 394, + 247, + 428, + 258 + ], + "type": "text", + "content": "Ours (M-A " + }, + { + "bbox": [ + 394, + 247, + 428, + 258 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 394, + 247, + 428, + 258 + ], + "type": "text", + "content": " MU-A)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 190, + 331, + 242, + 384 + ], + "blocks": [ + { + "bbox": [ + 131, + 273, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 131, + 273, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 273, + 482, + 319 + ], + "type": "text", + "content": "Fig. 6: Ablation study of our identity-enhanced training and mixed attention (M-A). It is evident that identity-enhanced training significantly improves the identity fidelity, and mixed-attention mechanism enhances semantic consistency compared to mutual attention (MU-A) approach [5]." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 190, + 331, + 242, + 384 + ], + "lines": [ + { + "bbox": [ + 190, + 331, + 242, + 384 + ], + "spans": [ + { + "bbox": [ + 190, + 331, + 242, + 384 + ], + "type": "image", + "image_path": "8fbf6616ec2483404a4b7d2399eee2847d481cebb3030e2bba8de7834fe9a83c.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 343, + 271, + 372 + ], + "lines": [ + { + "bbox": [ + 242, + 343, + 271, + 372 + ], + "spans": [ + { + "bbox": [ + 242, + 343, + 271, + 372 + ], + "type": "text", + "content": "awearing a blue clothing, facing and smiling to the snow. A snowy, snowy, in an anime style" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 190, + 384, + 242, + 434 + ], + "blocks": [ + { + "bbox": [ + 190, + 384, + 242, + 434 + ], + "lines": [ + { + "bbox": [ + 190, + 384, + 242, + 434 + ], + "spans": [ + { + "bbox": [ + 190, + 384, + 242, + 434 + ], + "type": "image", + "image_path": "9a1c3b99492fc94af69f9a0838f250c115911615070ac6db8a20e09586ce23e7.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 434, + 220, + 437 + ], + "lines": [ + { + "bbox": [ + 206, + 434, + 220, + 437 + ], + "spans": [ + { + "bbox": [ + 206, + 434, + 220, + 437 + ], + "type": "text", + "content": "ID Image" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 242, + 394, + 271, + 423 + ], + "lines": [ + { + "bbox": [ + 242, + 394, + 271, + 423 + ], + "spans": [ + { + "bbox": [ + 242, + 394, + 271, + 423 + ], + "type": "text", + "content": "a woman wearing a purple dress and smiling to the camera, on a beach, sunset, comic book style" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 131, + 449, + 482, + 483 + ], + "lines": [ + { + "bbox": [ + 131, + 449, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 482, + 483 + ], + "type": "text", + "content": "Fig. 7: Ablation study of our AdaIN-mean operation. The results show that AdaIN-mean play a crucial role in style image generation. Compare to AdaIN [10] module, our AdaIN-mean helps to achieve a higher identity fidelity." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 271, + 331, + 372, + 383 + ], + "blocks": [ + { + "bbox": [ + 271, + 331, + 372, + 383 + ], + "lines": [ + { + "bbox": [ + 271, + 331, + 372, + 383 + ], + "spans": [ + { + "bbox": [ + 271, + 331, + 372, + 383 + ], + "type": "image", + "image_path": "8d1c57c74125779832a2b402dbf91d09755bd61eca819c7b94c846fadd7592a2.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 271, + 383, + 372, + 434 + ], + "blocks": [ + { + "bbox": [ + 271, + 383, + 372, + 434 + ], + "lines": [ + { + "bbox": [ + 271, + 383, + 372, + 434 + ], + "spans": [ + { + "bbox": [ + 271, + 383, + 372, + 434 + ], + "type": "image", + "image_path": "e46e5ccadcd2615e9b1fe5dadd204d460988396d27a7e752dc92d9cebc2e94e5.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 293, + 434, + 301, + 438 + ], + "lines": [ + { + "bbox": [ + 293, + 434, + 301, + 438 + ], + "spans": [ + { + "bbox": [ + 293, + 434, + 301, + 438 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 373, + 331, + 424, + 383 + ], + "blocks": [ + { + "bbox": [ + 373, + 331, + 424, + 383 + ], + "lines": [ + { + "bbox": [ + 373, + 331, + 424, + 383 + ], + "spans": [ + { + "bbox": [ + 373, + 331, + 424, + 383 + ], + "type": "image", + "image_path": "0bcf974314745e99ea2c32b5ec3637174b3d274b3dd1ec56afad195a9dfe6e6f.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 373, + 383, + 423, + 434 + ], + "blocks": [ + { + "bbox": [ + 373, + 383, + 423, + 434 + ], + "lines": [ + { + "bbox": [ + 373, + 383, + 423, + 434 + ], + "spans": [ + { + "bbox": [ + 373, + 383, + 423, + 434 + ], + "type": "image", + "image_path": "22a87f3a44b42531fb9abd58dadcbe61d4125c02c416795ca24e670f5f0ef109.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 434, + 419, + 438 + ], + "lines": [ + { + "bbox": [ + 375, + 434, + 419, + 438 + ], + "spans": [ + { + "bbox": [ + 375, + 434, + 419, + 438 + ], + "type": "text", + "content": "Ours (AdaIN-mean " + }, + { + "bbox": [ + 375, + 434, + 419, + 438 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 375, + 434, + 419, + 438 + ], + "type": "text", + "content": " AdaIN)" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 130, + 509, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 481, + 533 + ], + "type": "text", + "content": "share the same inference processing. The qualitative comparison demonstrates that our identity-enhanced training notably enhances identity fidelity." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "content": "Ablation of mixed attention mechanism. To assess the effectiveness of our proposed mixed attention (M-A) mechanism, we compare our method with \"Ours (w/o M-A)\" and \"Ours (M-A " + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "content": " MU-A)\" in Fig. 6 and Table 1. Mutual self-attention (MU-A) [5] converts the existing self-attention into 'cross-attention' for consistent image editing, where the crossing operation happens in the self-attention of two related diffusion processes. The results show that our mixed attention mechanism demonstrates superior ability to improve semantic consistency while maintaining identity fidelity." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Ablation of AdaIN-mean operation. To assess the effectiveness of our proposed AdaIN-mean operation, we compare our Infinite-ID model with variations, namely Ours (w/o AdaIN-mean) and Ours (AdaIN-mean " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " AdaIN), as depicted" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "Infinite-ID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "in Fig. 7. The results reveal that: i) Ours (w/o AdaIN-mean) exhibits superior ID fidelity compared to Infinite-ID but fails to achieve style consistency with the text prompt; ii) Both our AdaIN-mean and AdaIN modules successfully achieve style consistency with the text prompt, yet AdaIN-mean maintains better ID fidelity than AdaIN. In conclusion, our proposed AdaIN-mean operation facilitates precise stylization while concurrently preserving ID fidelity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 207, + 316, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 316, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 316, + 220 + ], + "type": "text", + "content": "5 Conclusion and limitations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 234, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 234, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 482, + 402 + ], + "type": "text", + "content": "In this paper, we introduce Infinite-ID, an innovative identity-preserved personalization method designed to meet the requirement of identity (ID) fidelity and semantic consistency of the text prompt, all achievable with just one reference image and completed within seconds. Infinite-ID comprises three key components: the identity-enhanced training, mixed attention mechanism, and adaptive mean normalization (AdaIN-mean). Through extensive experimentation, our results illustrate that Infinite-ID outperforms baseline methods, delivering strong ID fidelity, superior generation quality, and precise semantic consistency in both raw photo generation and style image generation tasks. However, it's important to note that our method lacks multi-object personalization capability. Moreover, artifacts may occur when the human face occupies only a small portion of the entire image, attributed to limitations inherent in the original diffusion model. Furthermore, the image encoders employed in the identity-preserved personalization remain to be further explored." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 421, + 480, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 421, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 421, + 480, + 435 + ], + "type": "text", + "content": "6 Copyright of Training Data and Ethical Considerations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 447, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 495 + ], + "type": "text", + "content": "Copyright of Training Data. We construct our training dataset from the open-source public image dataset, including LAION-2B dataset, the LAION-Face dataset and few high-quality images collected from the Website5 which supports the CC0 (Creative Commons Zero) 1.0 Universal license6." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 495, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 590 + ], + "type": "text", + "content": "Ethical Considerations. We recognize the ethical considerations about improper use of such high-quality human photo synthesis techniques. Therefore, we emphasize the significance of establishing and complying to ethical guidelines [4] and using this technology responsibly. We hope that our work will drive further discussions and research on the security and ethical use of the human photo generation in computer vision. Moreover, we will employ the watermark and safety checker (NSFW classifiers) to attempt to avoid inappropriate content generation." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Yi Wu et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 597, + 334, + 610 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 597, + 334, + 610 + ], + "spans": [ + { + "bbox": [ + 133, + 597, + 334, + 610 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 133, + 597, + 334, + 610 + ], + "type": "text", + "content": " https://www.pexels.com, https://pixabay.com." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 611, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 611, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 611, + 481, + 665 + ], + "type": "text", + "content": "6 The person who associated a work with this deed has dedicated the work to the public domain by waiving all of his or her rights to the work worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law. You can copy, modify, distribute and perform the work, even for commercial purposes, all without asking permission." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 481, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 481, + 174 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 481, + 174 + ], + "type": "text", + "content": "This research was funded by the National Natural Science Foundation of China (U19B2044), supported by the Fundamental Research Funds for the Central Universities (WK2100000045)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 211, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 211, + 481, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 481, + 244 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 481, + 244 + ], + "type": "text", + "content": "1. Achlioptas, P., Benetatos, A., Fostiropoulos, I., Skourtis, D.: Stellar: Systematic evaluation of human-centric personalized text-to-image methods. arXiv preprint arXiv:2312.06116 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "type": "text", + "content": "2. Avrahami, O., Aberman, K., Fried, O., Cohen-Or, D., Lischinski, D.: Break-a-scene: Extracting multiple concepts from a single image. arXiv preprint arXiv:2305.16311 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 277, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 481, + 310 + ], + "type": "text", + "content": "3. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf 2, 3 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 310, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 481, + 331 + ], + "type": "text", + "content": "4. Biden, J.R.: Executive order on the safe, secure, and trustworthy development and use of artificial intelligence (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 331, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 481, + 363 + ], + "type": "text", + "content": "5. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. arXiv preprint arXiv:2304.08465 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 363, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 363, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 363, + 481, + 396 + ], + "type": "text", + "content": "6. Chen, L., Zhao, M., Liu, Y., Ding, M., Song, Y., Wang, S., Wang, X., Yang, H., Liu, J., Du, K., et al.: Photovverse: Tuning-free image customization with text-to-image diffusion models. arXiv preprint arXiv:2309.05793 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 396, + 481, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 481, + 417 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 481, + 417 + ], + "type": "text", + "content": "7. Chen, X., Huang, L., Liu, Y., Shen, Y., Zhao, D., Zhao, H.: Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 417, + 481, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 417, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 417, + 481, + 449 + ], + "type": "text", + "content": "8. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 449, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 481, + 482 + ], + "type": "text", + "content": "9. Gal, R., Arar, M., Atzmon, Y., Bermano, A.H., Chechik, G., Cohen-Or, D.: Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG) 42(4), 1-13 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 482, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 482, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 482, + 481, + 503 + ], + "type": "text", + "content": "0. Hertz, A., Voynov, A., Fruchter, S., Cohen-Or, D.: Style aligned image generation via shared attention. arXiv preprint arXiv:2312.02133 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 503, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 481, + 546 + ], + "type": "text", + "content": "1. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: 2017 IEEE International Conference on Computer Vision (ICCV) (Oct 2017). https://doi.org/10.1109/iccv.2017.167, http://dx.doi.org/10.1109/iccv.2017.167" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 546, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 481, + 590 + ], + "type": "text", + "content": "2. Inoue, N., Kikuchi, K., Simo-Serra, E., Otani, M., Yamaguchi, K.: Layoutdm: Discrete diffusion model for controllable layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10167-10176 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 590, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 481, + 633 + ], + "type": "text", + "content": "3. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 666 + ], + "type": "text", + "content": "4. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "Infinite-ID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "15. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1931-1941 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 151, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 151, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 133, + 151, + 481, + 184 + ], + "type": "text", + "content": "16. Li, D., Li, J., Hoi, S.C.: Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "type": "text", + "content": "17. Li, Z., Cao, M., Wang, X., Qi, Z., Cheng, M.M., Shan, Y.: Photomaker: Customizing realistic human photos via stacked id embedding. arXiv preprint arXiv:2312.04461 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "type": "text", + "content": "18. Li, Z., Usman, M., Tao, R., Xia, P., Wang, C., Chen, H., Li, B.: A systematic survey of regularization and normalization in gans. ACM Computing Surveys 55(11), 1-37 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 480, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 480, + 284 + ], + "type": "text", + "content": "19. Li, Z., Wang, C., Zheng, H., Zhang, J., Li, B.: Fakeclr: Exploring contrastive learning for solving latent discontinuity in data-efficient gans. In: European Conference on Computer Vision. pp. 598-615. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 480, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 318 + ], + "type": "text", + "content": "20. Li, Z., Xia, P., Tao, R., Niu, H., Li, B.: A new perspective on stabilizing gans training: Direct adversarial training. IEEE Transactions on Emerging Topics in Computational Intelligence 7(1), 178-189 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 319, + 480, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 480, + 351 + ], + "type": "text", + "content": "21. Ma, J., Liang, J., Chen, C., Lu, H.: Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. arXiv preprint arXiv:2307.11410 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 353, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 480, + 386 + ], + "type": "text", + "content": "22. Peng, X., Zhu, J., Jiang, B., Tai, Y., Luo, D., Zhang, J., Lin, W., Jin, T., Wang, C., Ji, R.: Portrait booth: A versatile portrait model for fast identity-preserved personalization. arXiv preprint arXiv:2312.06354 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 387, + 480, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 480, + 419 + ], + "type": "text", + "content": "23. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 420, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 480, + 463 + ], + "type": "text", + "content": "24. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 464, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 508 + ], + "type": "text", + "content": "25. Radford, A., Kim, J., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Amanda, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. Cornell University - arXiv, Cornell University - arXiv (Feb 2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 509, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 480, + 542 + ], + "type": "text", + "content": "26. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 544, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 576 + ], + "type": "text", + "content": "27. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 577, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 480, + 620 + ], + "type": "text", + "content": "28. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "type": "text", + "content": "29. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Yi Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "30. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "text", + "content": "31. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "type": "text", + "content": "32. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "text", + "content": "33. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "type": "text", + "content": "34. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 326 + ], + "type": "text", + "content": "35. Wang, Q., Bai, X., Wang, H., Qin, Z., Chen, A.: Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "text", + "content": "36. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "text", + "content": "37. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7623-7633 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "38. Wu, Y., Li, Z., Wang, C., Zheng, H., Zhao, S., Li, B., Tao, D.: Domain remodulation for few-shot generative domain adaptation. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "39. Xiao, G., Yin, T., Freeman, W.T., Durand, F., Han, S.: Fastcomposer: Tuning-free multi-subject image generation with localized attention. arXiv preprint arXiv:2305.10431 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "text", + "content": "40. Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 501, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 566 + ], + "type": "text", + "content": "41. Yuan, G., Cun, X., Zhang, Y., Li, M., Qi, C., Wang, X., Shan, Y., Zheng, H.: Inserting anybody in diffusion models via celeb basis. In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 72958-72982. Curran Associates, Inc. (2023), https://proceedings.neurips.cc/paper_files/paper/2023/file/e6d37cc5723e810b793c834bc6647cf-Paper-Conference.pdf" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 567, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 609 + ], + "type": "text", + "content": "42. Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djolonga, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., et al.: A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "text", + "content": "43. Zhang, C., Zhang, C., Zhang, M., Kweon, I.S.: Text-to-image diffusion model in generative ai: A survey. arXiv preprint arXiv:2303.07909 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "44. Zhang, K., Zhang, Z., Li, Z., Qiao, Y.: Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters 23(10), 1499-1503 (2016)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "Infinite-ID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 194 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "45. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "46. Zheng, Y., Yang, H., Zhang, T., Bao, J., Chen, D., Huang, Y., Yuan, L., Chen, D., Zeng, M., Wen, F.: General facial representation learning in a visual-linguistic manner. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18697-18709 (2022)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Yi Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/InfoNorm_ Mutual Information Shaping of Normals for Sparse-View Reconstruction/8f9c186f-c37a-4b28-8f4e-c701ceacc591_content_list.json b/2024/InfoNorm_ Mutual Information Shaping of Normals for Sparse-View Reconstruction/8f9c186f-c37a-4b28-8f4e-c701ceacc591_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ff7354ed7e893e2ca3a2514c5ce8ea3ea553be44 --- /dev/null +++ b/2024/InfoNorm_ Mutual Information Shaping of Normals for Sparse-View Reconstruction/8f9c186f-c37a-4b28-8f4e-c701ceacc591_content_list.json @@ -0,0 +1,1880 @@ +[ + { + "type": "text", + "text": "InfoNorm: Mutual Information Shaping of Normals for Sparse-View Reconstruction", + "text_level": 1, + "bbox": [ + 259, + 140, + 743, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xulong Wang $^{1,2*}$ Siyan Dong $^{3*}$ Youyi Zheng $^{1\\dagger}$ Yanchao Yang $^{3,4\\dagger}$", + "bbox": [ + 241, + 210, + 756, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 State Key Lab of CAD&CG, Zhejiang University 2 Chohotech co. Ltd.", + "bbox": [ + 261, + 238, + 738, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Institute of Data Science, The University of Hong Kong", + "bbox": [ + 307, + 253, + 692, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Department of Electrical and Electronic Engineering, The University of Hong Kong", + "bbox": [ + 215, + 267, + 784, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. 3D surface reconstruction from multi-view images is essential for scene understanding and interaction. However, complex indoor scenes pose challenges such as ambiguity due to limited observations. Recent implicit surface representations, such as Neural Radiance Fields (NeRFs) and signed distance functions (SDFs), employ various geometric priors to resolve the lack of observed information. Nevertheless, their performance heavily depends on the quality of the pre-trained geometry estimation models. To ease such dependence, we propose regularizing the geometric modeling by explicitly encouraging the mutual information among surface normals of highly correlated scene points. In this way, the geometry learning process is modulated by the second-order correlations from noisy (first-order) geometric priors, thus eliminating the bias due to poor generalization. Additionally, we introduce a simple yet effective scheme that utilizes semantic and geometric features to identify correlated points, enhancing their mutual information accordingly. The proposed technique can serve as a plugin for SDF-based neural surface representations. Our experiments demonstrate the effectiveness of the proposed in improving the surface reconstruction quality of major states of the arts. Our code is available at: https://github.com/Muliphein/InfoNorm.", + "bbox": [ + 261, + 318, + 740, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D from multi-view images $\\cdot$ surface reconstruction", + "bbox": [ + 261, + 595, + 689, + 608 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 635, + 375, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D surface reconstruction from multi-view images is an essential task in the computer vision and graphics community, with practical applications including content creation for virtual reality and robot-scene interaction. However, indoor scenes present challenges due to the large and complex scenes captured from sparse viewpoints. These challenges encompass issues such as occlusion and ambiguity arising from limited observations. As a result, traditional Multi-View Stereo (MVS) methods [28,36], which require substantial overlap among images, may not produce satisfactory results in these scenarios.", + "bbox": [ + 212, + 666, + 787, + 787 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contributions (wangxulong@zju.edu.cn, siyan3d@hku.hk).", + "bbox": [ + 217, + 796, + 658, + 811 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 220, + 811, + 392, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "2 This work was done during the author's internship at Chohotech Co. Ltd..", + "bbox": [ + 220, + 825, + 720, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9b8401e3fa530d78df7d772cbc467ce6c418352d13234aeaf0b6384a2139842b.jpg", + "image_caption": [ + "VoISDF" + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 401, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/806ffe2576270f50b515b669174c1942b2de766ad6891de348a0e98e77682aa9.jpg", + "image_caption": [ + "VoISDF+ (Ours)" + ], + "image_footnote": [], + "bbox": [ + 408, + 146, + 591, + 229 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4fa19faf7ef2ee3e5792b72952735bf2af6b7b5b9c5badae8ff9e895d4e9b334.jpg", + "image_caption": [ + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 598, + 146, + 781, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5a2d9d718867f7ac2f1ce0a97aca24769f738ae92246bee6311fadca4d578343.jpg", + "image_caption": [ + "GeoNeuS", + "Fig. 1: 3D scene reconstruction from sparse views on Replica [26] (first row) and ScanNet++ [38] (second row). By enforcing the mutual information between the normals of highly correlated scene points, the proposed method can effectively enhance the reconstruction quality of the baselines (VolSDF [37] and GeoNeuS [8])." + ], + "image_footnote": [], + "bbox": [ + 222, + 243, + 403, + 306 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ca123373e0b175854094b84eb9138c7da3f4b7698a126c30ac38be6287709a8c.jpg", + "image_caption": [ + "GeoNeuS $^{+}$ (Ours)" + ], + "image_footnote": [], + "bbox": [ + 411, + 243, + 589, + 306 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/163c73317edb184b56866c497313fa562e0cfff80ed358069df6bcaa7a5f696a.jpg", + "image_caption": [ + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 596, + 244, + 779, + 306 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, implicit scene representations, e.g., Neural Radiance Fields (NeRFs) [21], have been proposed to encode 3D scenes into a set of neural network parameters, with subsequent works enhancing the representational capacity for geometry to enable surface extraction. The key idea involves incorporating signed distance functions (SDFs) [23] into the learned radiance field. Although NeRFs with SDF produce high-quality reconstructions of simple scenes with sufficient images, they still underperform in large and complex scenes captured from sparse viewpoints. More recent approaches, such as NeuRIS [29] and MonoSDF [40], propose using monocular geometric priors to guide the learning of the radiance field. However, they can be affected by inevitable errors and noise in the monocular geometry estimation modules.", + "bbox": [ + 212, + 417, + 784, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We explore geometric regularization in a second-order perspective. We argue that both the geometry itself and the correlation of geometry among scene regions are crucial for surface reconstruction, especially when dealing with varying indoor scenes under sparse views with noisy geometry estimates. The correlation is expressed as mutual information between scene entities under random perturbations of the network weights. More explicitly, we propose to enforce the mutual information between the normals of two points in the scene that are considered geometrically correlated, where the mutual information can be efficiently computed by measuring the cosine similarity between the gradients of the normals with respect to the perturbed weights. Further, we leverage a combination of pretrained semantic and geometric features [4,7,12] to identify the correlated scene regions that endorse high mutual information. This results in an easy-to-use geometric shaping1 technique that can be applied to any SDF-based neural radiance fields. To thoroughly evaluate its effectiveness, we test the proposed with", + "bbox": [ + 212, + 585, + 785, + 796 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "X. Wang & S. Dong et al.", + "bbox": [ + 271, + 114, + 444, + 128 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1 We use the term \"shaping\" to refer to the process of optimizing network parameters to enforce consistencies in the scene geometry via aligning the gradients.", + "bbox": [ + 217, + 810, + 784, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "multiple state-of-the-art neural surface reconstruction models. The experimental results confirm the effectiveness of the proposed shaping technique, showing improved surface reconstruction quality (e.g., Fig. 1) on challenging scenes with limited views. To summarize:", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose to leverage second-order geometric correlation in the form of mutual information between normals to regularize the surface reconstruction of SDF-based NeRF representations.", + "- We develop a pipeline that efficiently enforces the mutual information of the normals of points, deemed geometrically correlated with pre-trained multimodal features, to improve the reconstruction quality of the scene.", + "- We verify the effectiveness of the proposed second-order regularization technique across a broad spectrum of baselines and demonstrate its usefulness as an easy-to-use plugin in improving 3D surface modeling." + ], + "bbox": [ + 225, + 218, + 785, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 378, + 387, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Reconstruction with neural radiance fields (NeRFs). NeRF [21] is originally designed for novel view synthesis, with its variants serving as scene representations [15, 32, 33, 35, 39] for tasks related to editing [2, 31, 44], semantic segmentation [5, 14, 16, 34], 3D shape generation [20, 27], and so on. We mainly focus on neural surface representations that enable the extraction of high-quality 3D meshes from the trained NeRFs.", + "bbox": [ + 212, + 414, + 787, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early works such as NeuS [30] and VolSDF [37] combine signed distance functions (SDFs) with volume rendering to extract high-quality surfaces. Recently, SparseNeuS [17] introduces a cascaded geometry reasoning framework that generalizes well to novel scenes. GeoNeuS [8] uses surface points from structure-from-motion and neighbor-view patches to reconstruct surfaces, while NeuDA [3] employs a deformable anchor to adaptively encode geometric details. NeAT [19] broadens the scope of reconstruction to arbitrary topology, no longer confining to watertight surfaces. Moreover, Neuralangelo [13] uses a hash grid for position encoding and implements a coarse-to-fine optimization with numerical gradients. Despite their promising results, most of these methods need sufficient training views, and their performance declines when dealing with large and complex scenes under sparse viewpoints.", + "bbox": [ + 212, + 505, + 787, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Incorporating priors into NeRFs. One can augment scene representations by introducing various priors to NeRFs. VDN-NeRF [43] normalizes the spatial feature to align with monocular features. LERF [11] introduces a new branch whose outputs align with DINO [4] and CLIP [24] features. Semantic-NeRF [42] encodes semantic information for segmentation tasks, while JacobiNeRF [34] shapes the color gradients to encode the mutual information in terms of semantic similarity from DINO [4] features for label propagation.", + "bbox": [ + 212, + 703, + 787, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For surface reconstruction, GeoNeuS [8] incorporates patch similarity from neighboring images and sparse surface points derived from structure-from-motion", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "InfoNorm", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[25], while I²-SDF [44] combines the rendering equation with VolSDF [37] for better novel views and geometry. To cope with complex indoor scenes, NeuRIS [29] incorporates the normal priors as additional geometric supervision. Further, MonoSDF [40] uses both depth and normal to improve the surface reconstruction quality. However, these methods highly depend on the quality of the supplemented geometric information. In contrast, our method leverages the second-order information in geometric prior, rather than directly using the prior as a first-order supervision signal, which makes it more robust to errors and noise. There are also methods utilizing strong assumptions derived for human-made scenes. For instance, ManhattanSDF [10] works under the assumption that all captured scenes adhere to the Manhattan-world concept, thus, predicting semantic segmentation to identify walls and floors. While these methods primarily focus on planar scenes, the proposed technique can work for general scenes under noisy correlation information.", + "bbox": [ + 212, + 145, + 787, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 383, + 330, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We start with a discussion on the preliminaries in Sec. 3.1 to provide the background on how we can learn from posed images an implicit scene representation (NeRF), improve 3D scene reconstruction via the combination of NeRF and SDF representations, and employ mutual information as a constraint in the NeRF training process. Following this, in Sec. 3.2, we delve into how mutual information can be used as constraints on NeRF's implied surfaces, so that we can learn better geometries for higher-quality 3D reconstruction of the scene. Finally, we elaborate on how we can employ off-the-shelf semantic and geometry features to assist the encoding of mutual information through the proposed geometric shaping framework.", + "bbox": [ + 212, + 417, + 787, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 595, + 372, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Neural radiance fields. NeRF [21] learns an implicit scene representation from a set of posed images. For a specific pixel in an image, NeRF samples a list of 3D points, $\\{x|x_{t} = \\mathbf{o} + t\\mathbf{v},t\\in [t_{near},t_{far}]\\}$ , from the camera center $\\mathbf{o}$ along the viewing direction $\\mathbf{v}$ within a bounded space. Specifically, it employs a Multilayer Perceptron (MLP) to encode the volume density and color of each sampled point. We denote all the parameters of NeRF as $\\theta$ , the density function as $\\sigma$ , and the color function as $c$ . The pixel's rendered color $\\hat{\\mathbf{C}}$ from a NeRF is calculated by the discrete volume rendering:", + "bbox": [ + 212, + 622, + 787, + 744 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {C}} (\\mathbf {o}, \\mathbf {v}; \\theta) = \\sum_ {i = 1} ^ {N} \\omega (\\mathbf {o}, \\mathbf {v}, t _ {i}; \\theta) c \\left(x _ {t _ {i}}, \\mathbf {v}; \\theta\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 758, + 784, + 806 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} \\omega (\\mathbf {o}, \\mathbf {v}, t _ {i}; \\theta) = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}) \\alpha_ {i}.\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 800, + 635, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "X. Wang & S. Dong et al.", + "bbox": [ + 271, + 114, + 442, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\alpha_{i} = 1 - \\exp (-\\sigma_{i}\\delta_{i})$ , where $\\sigma_{i}$ is the abbreviation of $\\sigma (x_{t_i};\\theta)$ and $\\delta_{i} = t_{i + 1} - t_{i}$ . A photometric loss is utilized to minimize the discrepancy between the rendered color and its ground truth:", + "bbox": [ + 214, + 146, + 787, + 191 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {C} = \\left\\| \\mathbf {C} - \\hat {\\mathbf {C}} \\right\\| _ {1}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 441, + 204, + 784, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "so that the parameters $\\theta$ can be optimized to fit the scene. Despite the ability to handle photorealistic novel view synthesis, NeRFs suffer from noisy and unrealistic surface extraction when it comes to 3D reconstruction, as discussed in the literature [30, 37].", + "bbox": [ + 212, + 233, + 782, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NeRFs with signed distance functions. To produce high-quality surfaces, it is beneficial to use signed distance functions (SDFs) [6, 23] to represent geometry. Consequently, the surface can be reconstructed by extracting the zero-level set [18] of the SDF values. The combination of NeRF and SDF can be achieved by deriving the NeRF's densities from the SDF, or defining Laplace's cumulative distribution function. Accordingly, $\\alpha_{i}$ in Eq. 1 can be represented by a function $\\varPhi$ which converts SDF values to volume densities as", + "bbox": [ + 214, + 310, + 782, + 414 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} = \\Phi (x, f (x; \\theta), i). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 430, + 784, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generally, NeRFs with an SDF achieve improved surface reconstruction quality, where the key to maintaining the characteristics of SDF is an eikonal loss [9]:", + "bbox": [ + 214, + 457, + 782, + 488 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {E} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\left| \\left| \\nabla f \\left(x _ {i}\\right) \\right| \\right| _ {2} - 1\\right) ^ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 501, + 784, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mutual information in NeRFs. Mutual information (MI) quantifies the statistical dependence between two random variables, providing an estimate of how much information they share. It has been introduced for semantic segmentation and motion modeling tasks [34, 41]. In this paper, We study the mutual information under perturbations of the NeRF parameters, and focus more on the geometry properties. Let $I(p_{i})$ and $I(p_{j})$ denote two pixels derived from the NeRF rendering process $\\mathbb{F}$ . And we assume that a subset of the NeRF weights, denoted by $\\theta^{D}$ , is perturbed by a random noise vector $\\mathbf{n} \\in \\mathbb{R}^{D}$ sampled from a uniform distribution on the sphere $\\mathbb{S}^{D - 1}$ by a small step $\\gamma \\ll 1.0$ . The random variables corresponding to the perturbed pixels can be written through a Taylor expansion:", + "bbox": [ + 214, + 556, + 787, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} (p _ {i}) = I (p _ {i}) + \\gamma \\mathbf {n} \\frac {\\partial \\mathbb {F} _ {i}}{\\partial \\theta^ {D}},\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 734, + 588, + 762 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} (p _ {j}) = I (p _ {j}) + \\gamma \\mathbf {n} \\frac {\\partial \\mathbb {F} _ {j}}{\\partial \\theta^ {D}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 765, + 588, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can now characterize the mutual information between $\\hat{I}(p_i)$ and $\\hat{I}(p_j)$ under the joint probability distribution $\\mathbb{P}(\\hat{I}(p_i), \\hat{I}(p_j))$ . As derived in JacobiNeRF, the", + "bbox": [ + 214, + 806, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "InfoNorm", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mutual information can be computed as:", + "bbox": [ + 215, + 146, + 511, + 161 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbb {I} (\\hat {I} (p _ {i}), \\hat {I} (p _ {j})) = \\mathbb {H} (\\hat {I} (p _ {i})) - \\mathbb {H} (\\hat {I} (p _ {i}) | \\hat {I} (p _ {j})) \\\\ = \\log (\\frac {1}{\\sqrt {1 - \\cos^ {2} \\beta}}) + \\mathrm {c o n s t .}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 171, + 656, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\beta$ denotes the angle between $\\partial \\mathbb{F}_i / \\partial \\theta^D$ and $\\partial \\mathbb{F}_j / \\partial \\theta^D$ . It is observed that:", + "bbox": [ + 215, + 233, + 779, + 251 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {I} (\\hat {I} (p _ {i}), \\hat {I} (p _ {j})) \\propto | | \\cos \\beta | |. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 261, + 785, + 279 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, to enhance the MI between two random pixels induced by NeRF perturbations, one can enforce the absolute value of the cosine similarity. That is, if two pixels $p_i$ , $p_j$ come with high correlation (e.g., from the same object in the scene), their gradients regarding the perturbed parameters should be in the same or opposite direction.", + "bbox": [ + 214, + 286, + 785, + 362 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Mutual Information Shaping of Geometry in SDF-based NeRFs", + "text_level": 1, + "bbox": [ + 215, + 382, + 785, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While the previous approach focuses on enforcing mutual information for image synthesis, we explore a mutual information shaping of the geometries for quality surface reconstruction in SDF-based NeRFs.", + "bbox": [ + 214, + 406, + 785, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, we represent a scene via a NeRF backbone with an SDF head and a color head, as shown in Fig. 2. Note that the method to be proposed is capable of being applied to any network similar to NeRF and SDF representations. Given that the two heads partially share parameters, we denote all the parameters as $\\theta$ , and use $f(x;\\theta)$ and $c(x,\\mathbf{v};\\theta)$ to represent the SDF and color branches, respectively. To achieve mutual information shaping of the geometry, we focus on manipulating the parameters of $f(x;\\theta)$ , while keeping $c(x,\\mathbf{v};\\theta)$ unchanged. Below, we first explain why and how we compute mutual information with surface normals derived from SDF, i.e., $f(x;\\theta)$ , as a geometry-aware constraint. Next, we present a simple yet effective strategy that extracts regions with high correlation, which need to be encoded into the weights using the mutual information shaping technique. Finally, we elaborate on our loss function and training process to reconstruct the scene from posed images.", + "bbox": [ + 214, + 452, + 787, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mutual information among surface normals. We explicitly encode mutual information into the function $f(x; \\theta)$ . However, instead of computing mutual information between the density values, we proceed by manipulating the surface normal vectors that can be derived from $f$ . Specifically, for a 3D point $x$ on a surface of the scene, its normal vector $N(x; \\theta)$ can be calculated as the partial derivative of the 3D coordinate:", + "bbox": [ + 214, + 660, + 787, + 751 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nN (x; \\theta) = \\frac {\\partial f (x ; \\theta)}{\\partial x}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 758, + 785, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "There are several reasons to manipulate surface normal vectors but not SDF values: 1) Shaping the SDF values with respect to a subset $\\theta^D$ of the parameters may not effectively impose the constraints on all parameters. However, the", + "bbox": [ + 215, + 794, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "X. Wang & S. Dong et al.", + "bbox": [ + 271, + 114, + 444, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6f17f63afda6c3119d5acae9d04874923ea056d465e48b6553b8bf4685feb5f3.jpg", + "image_caption": [ + "Fig. 2: An overview of the pipeline, where we apply mutual information shaping on the geometric branch to enforce consistencies that help enhance the surface reconstruction. Specifically, the model consists of the NeRF backbone, an SDF head, and a color head. The predicted density and color are supervised by the classic eikonal loss $L_{E}$ and the photometric reconstruction loss $L_{C}$ , respectively. We encode geometry-aware mutual information into a subset of the parameters $\\theta^{D}$ of $f(x)$ to constrain the density field learning for better surface quality, which is achieved by the proposed mutual information loss $L_{M}$ computed on top of the estimated surface normal." + ], + "image_footnote": [], + "bbox": [ + 222, + 145, + 781, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "calculation of the surface normal includes the entire set of parameters in $f$ , thus a more global regularity. 2) Moreover, shaping $f$ with the mutual information between its values may induce side effects as density could be distorted to satisfy the constraints. However, shaping the normal $\\partial f / \\partial x$ leave more space for the SDF values to cope with the scene reconstruction task, thus minimizing potential negative effects.", + "bbox": [ + 212, + 420, + 782, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "According to Eq. 5, we need to calculate the partial gradients $\\partial N / \\partial \\theta^D$ to compute the mutual information. When it comes to a vector rather than a scalar, the gradients consist of three components corresponding to the $\\mathbf{x}$ , $\\mathbf{y}$ , and $\\mathbf{z}$ axes:", + "bbox": [ + 214, + 511, + 784, + 556 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial N (x ; \\theta)}{\\partial \\theta^ {D}} = \\frac {\\partial N _ {\\mathbf {x}} (x ; \\theta)}{\\partial \\theta^ {D}}, \\frac {\\partial N _ {\\mathbf {y}} (x ; \\theta)}{\\partial \\theta^ {D}}, \\frac {\\partial N _ {\\mathbf {z}} (x ; \\theta)}{\\partial \\theta^ {D}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 566, + 784, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For ease of computation, we concatenate the gradients into a vector, denoted as $\\partial N$ for simplicity:", + "bbox": [ + 214, + 601, + 782, + 630 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\partial N = \\operatorname {C o n c a t} \\left(\\frac {\\partial N _ {\\mathbf {x}}}{\\partial \\theta^ {D}}, \\frac {\\partial N _ {\\mathbf {y}}}{\\partial \\theta^ {D}}, \\frac {\\partial N _ {\\mathbf {z}}}{\\partial \\theta^ {D}}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 628, + 784, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since a 2D pixel is accumulated from 3D, we aggregate the normal-based information along a viewing ray as in Eq. 1 as:", + "bbox": [ + 214, + 662, + 782, + 693 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {N} \\left(p _ {i}\\right) = \\sum_ {x \\in \\text {r a y}} \\omega \\left(t _ {i}; \\theta\\right) \\partial N. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 702, + 784, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we obtain that the mutual information between two normals is:", + "bbox": [ + 214, + 743, + 732, + 758 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {I} (\\hat {N} (p _ {i}), \\hat {N} (p _ {j})) \\propto | | \\frac {\\mathbb {N} (p _ {i}) \\cdot \\mathbb {N} (p _ {j})}{| \\mathbb {N} (p _ {i}) | | \\mathbb {N} (p _ {j}) |} | |. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 767, + 784, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Please refer to our supplementary material for a detailed derivation. From Eq. 10, we observe that the mutual information in surface normal vectors correlates", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "InfoNorm", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a1037c099eeadeb4e12569c2335a9d39eec82a6e6ce806b409303ba67e34c5c1.jpg", + "image_caption": [ + "Anchor pixel" + ], + "image_footnote": [], + "bbox": [ + 220, + 145, + 356, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fdc6e8d50ef3f3f196ffc7ad2e20073c5eefb613a549457884295d170dcec5a5.jpg", + "image_caption": [ + "Similar semantic features", + "Fig. 3: An example of positive samples from different features. Given an anchor pixel (marked by a blue circle in the first image) on the wall-like cabinet, semantic features like DINO often correlate both the cabinet and the ceiling (as shown by the red area in the second image). Meanwhile, geometric features such as (noisy) monocular normals can not distinguish between parallel planes that are not connected (the third image). By combining semantic and geometric features, we can obtain positive samples of the anchor pixel with better geometric consistency, thus, high mutual correlation." + ], + "image_footnote": [], + "bbox": [ + 361, + 145, + 498, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0713b265af24b7ebbaeed91bc4e2fbefb2b1432c4c0011904bb28b4c17d96bbc.jpg", + "image_caption": [ + "Similar geometric features" + ], + "image_footnote": [], + "bbox": [ + 504, + 145, + 638, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c5e4fb51e95f325dfde32be6d991727a9691233107efa75288aab21b7ee2d042.jpg", + "image_caption": [ + "Combined (Ours)" + ], + "image_footnote": [], + "bbox": [ + 643, + 145, + 782, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "positively with the cosine similarity of their gradients to the network parameters. This conclusion is similar to that of Eq. 5. As a result, we can now identify the pixels that are of high correlations, and increase the aggregated mutual information during the NeRF training process.", + "bbox": [ + 212, + 375, + 787, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Extracting correlated scene regions from image pixels. To extract scene regions that need to be shaped to have high mutual information, we present a simple yet effective approach with off-the-shelf visual features. Given the input images, we extract pixel-wise semantic and geometric features to identify correlated scene regions. Following previous research [34], we apply DINO [4] as the semantic feature extractor to obtain a high-level understanding of the scene regions. Meanwhile, as we shape mutual information on top of the surface normal, we employ a monocular normal estimator [7] as the geometric feature extractor for a detailed low-level perception of the 3D points in the scene.", + "bbox": [ + 212, + 446, + 787, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DINO features can group different parts of an object, like the seat and back of a chair, which share similar semantics. However, the surface normal vectors from these parts have limited correlation, thus should not be constrained. Conversely, using only geometric features can wrongly suggest high correlation between unrelated areas, like the seat and the floor. As these areas aren't closely connected in 3D space, there is no reason to have better surface reconstruction when enhancing their mutual information. Similar phenomena on wall-like structures can be observed in Fig. 3, where we show that by combining semantic and geometric features together we can better identify scene regions that are spatially connected and have similar normal directions.", + "bbox": [ + 212, + 583, + 787, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "More explicitly, we combine semantic and geometric features with an intersection operator. It is simple yet effective. The process of identifying correlated scene regions (in normal) is implemented as positive pixel pair selection from input images. Given a randomly selected pixel from an image, we first extract similar pixels from neighboring images using semantic and geometric features, respectively. This extraction is implemented by applying two thresholds, $\\beta_{S}$ and $\\beta_{G}$ , to filter the cosine similarity for semantic and geometric features. We then", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "X. Wang & S. Dong et al.", + "bbox": [ + 271, + 114, + 444, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "feed the intersection of the top similar pixels from each feature type into NeRF's training process as positive samples, allowing us to shape them with high mutual information. In order to leverage the contrastive learning framework to encourage mutual information, we randomly select pixels not involved in the aforementioned process to serve as negative samples (independent pixels that are not correlated to the anchor pixel).", + "bbox": [ + 212, + 146, + 782, + 238 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Loss functions for training. The aforementioned contrastive learning process is performed with a tailored InfoNCE loss [22]:", + "bbox": [ + 214, + 248, + 782, + 280 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {M} = - \\log \\frac {\\sum \\exp (| | \\cos (\\partial \\mathbb {N} _ {i} , \\partial \\mathbb {N} _ {i +}) | |)}{\\sum_ {j} \\exp (| | \\cos (\\partial \\mathbb {N} _ {i} , \\partial \\mathbb {N} _ {j}) | |)}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 286, + 785, + 321 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $i$ , $i+$ , and $j$ denote the anchor, positive, and negative pixels, respectively. This loss function effectively enhances mutual information (according to Eq. 10) among correlated pixels, while maintaining low mutual information among independent pixels. Finally, to train the NeRF backbone and retain the characteristics of the SDF, we adopt the basic photometric loss (Eq. 2) and the eikonal loss (Eq. 4). The latter two are derived from single images, while the former can be calculated from either single images or a set of neighboring images. The full loss function is defined as the weighted sum of three loss terms:", + "bbox": [ + 214, + 327, + 784, + 446 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {C} + \\lambda_ {E} L _ {E} + \\lambda_ {M} L _ {M}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 455, + 784, + 470 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "With the above, all network parameters $\\theta$ can be trained end-to-end.", + "bbox": [ + 214, + 479, + 712, + 494 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 513, + 375, + 532 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we evaluate the effectiveness of our method through experiments conducted on multiple datasets with various baselines. First, we introduce the two datasets used to conduct the experiments in Sec. 4.1. Next, we detail the implementation of the proposed method in Sec. 4.2, where we adapt a spectrum of state-of-the-art models for validating the effectiveness of the geometric shaping. Then, we report the comparisons between the baselines and our models in Sec. 4.3. Lastly, we conduct a series of ablation studies in Sec. 4.4, to verify the necessity of each component in our method.", + "bbox": [ + 212, + 542, + 784, + 662 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 214, + 681, + 334, + 696 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate the proposed method using two public datasets from the literature: ScanNet++ [38] and the Replica [26] dataset. For the ScanNet++ dataset, we choose 8 representative scenes, while for the Replica dataset, we select 4 representative scenes. The scenes from both datasets represent a variety of indoor settings, including offices, bedrooms, and so on. As our focus is on scene reconstruction from sparse views, we uniformly sample 50-85 images from the official (and Replica from Semantic-NeRF [42]) scanning sequences for each scene to use as input in our evaluation. For detailed statistics of each scene, please refer to our supplementary material.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "InfoNorm", + "bbox": [ + 663, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation metrics. Following previous works [13, 30, 40], we evaluate the quality of surface reconstruction using two metrics: Chamfer distance and F-score. Chamfer distance is a measure that evaluates the difference between reconstructed and ground-truth points sets. F-score is a statistical metric that integrates precision and recall, providing a more comprehensive understanding of the reconstruction quality.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 258, + 457, + 273 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As the proposed geometric shaping can be easily integrated into any network similar to NeRF with SDF representations, we select several state-of-the-art models as our baselines and apply the proposed to them for demonstrating the effectiveness of our approach. We first choose NeuS [30] and VolSDF [37], the two representative pioneers for NeRFs with SDFs. We also apply our method to the following works: GeoNeuS [8], I²-SDF [44], NeuRIS [29], MonoSDF [40], and Neuralangelo [13]. In addition to the basic loss functions for NeRF and SDF, GeoNeuS incorporates patch similarity from neighboring views and surface points derived from structure-from-motion. I²-SDF integrates the rendering equation with VolSDF. We disregard its normal and depth loss terms. NeuRIS utilizes monocular normal estimation as a supervision on the geometry, and discards the low-quality normal information that lacks sufficient confidence through multi-view patching. MonoSDF employs both depth and normal estimation, for which we set the decay of the normal and depth at 30k iterations to prevent the method from degeneration. Neuralangelo utilizes a multi-resolution hash grid and numerical gradient. We set the hash encoding dictionary size to 20, and the encoding dimension to 4. For all the base models mentioned above, we enhance them by incorporating our proposed mutual information loss during training. The resulting methods are respectively labeled as $\\mathrm{NeuS^{+}}$ , VolSDF+, GeoNeuS+, and so on.", + "bbox": [ + 212, + 281, + 787, + 583 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Comparisons", + "text_level": 1, + "bbox": [ + 215, + 604, + 370, + 619 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results on the ScanNet++ dataset. The quantitative comparisons are reported in Tab. 1. We also visualize the surface reconstruction results in Fig. 4 for qualitative comparisons. We can observe that by applying the proposed mutual information shaping on the geometry, the surface reconstruction quality can be improved across all baselines, as measured by mean Chamfer distance, F-score, and the visuals.", + "bbox": [ + 212, + 628, + 787, + 718 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "NeuS and VolSDF tend to produce rough surfaces and excessively smoothed boundaries. However, with our geometry-aware mutual information, the surfaces become smoother, while the boundaries are more crispy. Compared to GeoNeuS, a method that uses patch-similarity among neighboring views and requires sufficient textures, our method provides significant improvements. Applying the proposed technique to $\\mathrm{I}^2$ -SDF can also enhance its reconstruction quality.", + "bbox": [ + 212, + 719, + 787, + 809 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As for more recent works, we find that NeuRIS is somewhat sensitive to the quality of normal estimation. It can either be affected by noise in the normal", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "X. Wang & S. Dong et al.", + "bbox": [ + 271, + 114, + 442, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/6f95acba95b601c8f059e1f4c6c843acaf0fecec1f6f13a14a91ffef6e2753f6.jpg", + "table_caption": [ + "Table 1: Quantitative results on the ScanNet++ dataset. Each cell includes the original baseline number in black, followed by the improvement using our mutual information. Positive and negative improvements are marked in green and orange, respectively. The last column shows the average improvement over all scenes, validating the effectiveness of the proposed shaping for better geometric reconstruction (despite a small potion of noise on a few scenes)." + ], + "table_footnote": [], + "table_body": "
0a7c0a186ee27b6456a09460a08de0abMean
NeuS [30]+0.08-0.040.05-0.030.03+0.010.06-0.040.01-0.000.24-0.140.04-0.030.02-0.000.07-0.03
VolSDF [37]+0.90-0.830.31-0.290.04+0.020.10-0.030.03-0.010.14+0.030.39-0.380.04-0.000.24-0.19
GeoNeuS [8]+0.03-0.021.03-1.020.28-0.270.05-0.030.02-0.010.22-0.200.01-0.010.26-0.250.24-0.23
I2-SDF [44]+0.27-0.240.22-0.030.04-0.010.47-0.450.03-0.020.13+0.050.03+0.020.30-0.070.19-0.09
NeuRIS [29]+0.01-0.000.01-0.000.03+0.020.07-0.040.04-0.030.41-0.210.03-0.010.03-0.000.08-0.03
MonoSDF [40]+0.01-0.000.01+0.020.02-0.010.02-0.000.01-0.000.05-0.020.01-0.000.01-0.000.02-0.00
Neuralangelo [13]+0.17-0.010.84-0.660.52-0.330.05-0.010.04+0.024.76-2.850.33+0.050.12+0.030.85-0.48
NeuS [30]+0.49+0.060.48+0.100.67+0.000.58+0.160.81-0.010.58-0.020.67+0.170.75+0.090.63+0.07
VolSDF [37]+0.45+0.350.47+0.270.67+0.000.55+0.050.71+0.060.52+0.180.48+0.350.70+0.020.57+0.16
GeoNeuS [8]+0.69+0.200.47+0.440.52+0.340.70+0.060.87+0.030.68+0.230.75+0.140.37+0.510.63+0.24
I2-SDF [44]+0.33+0.520.28+0.000.58+0.150.41+0.430.75+0.100.65+0.070.65+0.070.40+0.250.51+0.20
NeuRIS [29]+0.72+0.030.68+0.010.65+0.010.69+0.060.85+0.010.24+0.170.64+0.130.69+0.090.65+0.06
MonoSDF [40]+0.92-0.010.88-0.010.81+0.040.79+0.060.89+0.000.88+0.050.91+0.000.89-0.010.87+0.01
Neuralangelo [13]+0.42+0.050.14+0.300.35+0.120.63+0.050.71-0.010.17+0.220.29+0.060.51-0.010.40+0.10
", + "bbox": [ + 220, + 239, + 782, + 406 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/1ab590405cf2035dc2662145551ba9a56b42fa584ba3629c28d927463d808a7c.jpg", + "table_caption": [ + "Table 2: Quantitative results on the Replica dataset." + ], + "table_footnote": [], + "table_body": "
office0office1room0room1Mean
NeuS [30]+0.005+0.0010.023-0.0120.068-0.0380.130-0.0870.057-0.035
VolSDF [37]+0.004+0.0060.016-0.0040.080-0.0710.034-0.0250.033-0.023
GeoNeuS [8]+0.023-0.0210.014-0.0090.040-0.0350.002-0.0000.020-0.017
I2-SDF [44]+0.032-0.0090.056-0.033--0.044-0.021
NeuRIS [29]+0.014-0.0063.975-3.9630.025-0.0170.015+0.0021.007-0.996
MonoSDF [40]+0.003-0.0000.005-0.0010.004-0.0000.004-0.0010.004-0.001
Neuralangelo [13]+0.024-0.0180.030-0.0190.146-0.0120.082-0.0470.070-0.024
NeuS [30]+0.92+0.000.80+0.010.78-0.030.72+0.080.80+0.02
VolSDF [37]+0.94-0.040.78+0.010.78+0.160.82+0.070.83+0.05
GeoNeuS [8]+0.89+0.080.91+0.040.90+0.090.98+0.010.920+0.05
I2-SDF [44]+0.72+0.090.56+0.20--0.64+0.14
NeuRIS [29]+0.79+0.060.01+0.680.89+0.050.77+0.030.61+0.20
MonoSDF [40]+0.96+0.000.90+0.020.98+0.000.95+0.020.95+0.01
Neuralangelo [13]+0.89+0.070.49+0.360.65+0.150.71+0.180.69+0.19
", + "bbox": [ + 290, + 452, + 712, + 625 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "or completely disregard the normal information due to low confidence. Particularly, when images are sparse, the regions covered by fewer viewpoints tend to lose supervision. Our mutual information scheme does not depend on the absolute values of the surface normal, but on the similarity among these values. This makes it more robust to noise and less likely to discard low-quality but useful normal information. Consequently, our method enhances the reconstruction quality of NeuRIS. MonoSDF achieves the best overall performance among the baselines because it uses both depth and normal estimations as geometry supervision. However, our method still offers improvements. Notably, it produces smoother planes and clearer boundaries. This indicates that not only the supervision of geometric cues is important, but also the correlation between the geometric entities matters.", + "bbox": [ + 212, + 657, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "InfoNorm", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/85fcbe70ebb8b419ecd0f5736e2a72107774160c5cf350c90b8e8223928071d9.jpg", + "image_caption": [ + "Fig. 4: Visual results on the ScanNet++ dataset. Each row shows a comparison with a different baseline, from top to bottom, NeuS $[30]^+$ , VolSDF $[37]^+$ , GeoNeuS $[8]^+$ , I²-SDF $[44]^+$ , NeuRIS $[29]^+$ , MonoSDF $[40]^+$ , and Neuralangelo $[13]^+$ . Red boxes are overlaid to help the contrast." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 781, + 728 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "X. Wang & S. Dong et al.", + "bbox": [ + 271, + 114, + 442, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c3b66e13e0a8712a4ccb6d1eeba48ff3a6730ef1a34898ee902321c2a9046255.jpg", + "image_caption": [ + "Reference image" + ], + "image_footnote": [], + "bbox": [ + 220, + 145, + 352, + 223 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/12b7055978d076432f27e37e27bcf5105d4b8a6caddf0c26c3ce3c4e5a6b914d.jpg", + "image_caption": [ + "NeuS", + "Fig. 5: Reconstruction results after 10K iterations in the training process with $(\\mathrm{NeuS}^{+})$ and without (NeuS) the proposed geometric shaping technique." + ], + "image_footnote": [], + "bbox": [ + 364, + 145, + 496, + 223 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1934989b75e764148aeab0fdca00ee343919dd3f6fea03c82d22737ae5d6ba6b.jpg", + "image_caption": [ + "Neu $^{+}$ (Ours)" + ], + "image_footnote": [], + "bbox": [ + 506, + 145, + 638, + 223 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3d686b691ed9650ea1c7a36938a87db1535ca26e1281f93b6ada1fecee39a744.jpg", + "image_caption": [ + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 650, + 145, + 781, + 223 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Last but not least, the experiments on Neuralangelo demonstrate the potential of our method in working with hash grid encoding and numerical gradient. This highlights our method's versatility and ease of use, which can be broadly applied in various reconstruction architectures.", + "bbox": [ + 212, + 306, + 782, + 367 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results on the Replica dataset. In Tab. 2 we report the numerical results on the Replica dataset, with respect to NeuS [30], VolSDF [37], GeoNeuS [8], I²-SDF [44], NeuRIS [29], MonoSDF [40], and Neuralangelo [13]. As observed, the proposed geometric shaping technique still provides positive improvements when incorporated into these baselines, aligning with the conclusion drawn from the ScanNet++ data. Qualitative examples can also be found in Fig. 1.", + "bbox": [ + 212, + 382, + 784, + 473 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Analyses", + "text_level": 1, + "bbox": [ + 215, + 496, + 334, + 512 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation studies. We select three representative scenes from the ScanNet++ dataset and conduct ablation studies to evaluate the effectiveness of our method designs. We conduct ablation studies using two top-performing baselines: NeuRIS and MonoSDF.", + "bbox": [ + 212, + 522, + 429, + 657 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The results are reported in Tab. 3. Instead of using the surface normal formula to inject mutual information, we first perform a basic mutual information shaping separately in the color and SDF values. These two variations are labeled as base $^{+}$ (color)", + "bbox": [ + 212, + 659, + 429, + 794 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "and base $^{+}$ (SDF), respectively. Interestingly, we find that, even though color is not directly related to geometry, by shaping it, we can still obtain improvements on the geometry in some cases. Moreover, shaping directly on the SDF values", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/2199c25153cdaabb203df08322dabc11494f48b939ed42552af69e28972dbd91.jpg", + "table_caption": [ + "Table 3: Ablation studies on three scenes from ScanNet++. The best and second-best results are marked in bold and underlined." + ], + "table_footnote": [], + "table_body": "
Chamfer (m)↓F-score ↑
6ee27b649460Mean6ee27b649460Mean
NeuRIS [30]0.0290.0700.4050.1680.650.690.240.53
NeuRIS+ (color)0.0230.0380.3150.1250.720.680.340.58
NeuRIS+ (SDF)0.1270.0530.3600.1800.470.610.300.46
NeuRIS+ w/o normal0.0360.0250.3380.1330.660.750.320.58
NeuRIS+ w/o DINO0.0610.0330.3810.1580.630.720.320.56
NeuRIS+ (Full)0.0440.0330.1980.0920.660.760.410.61
MonoSDF [40]0.0200.0160.0460.0280.810.790.880.83
MonoSDF+ (color)0.0240.0190.0330.0250.790.780.870.81
MonoSDF+ (SDF)0.5880.0293.4431.3530.380.650.210.42
MonoSDF+ w/o normal0.1280.0230.0660.0720.680.730.800.74
MonoSDF+ w/o DINO0.0640.0160.0220.0340.780.830.930.85
MonoSDF+ (Full)0.0140.0200.0230.0190.850.850.930.87
", + "bbox": [ + 444, + 609, + 782, + 758 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "InfoNorm", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "severely degrade the reconstruction quality, which evidences that the shaping on SDF values could incur a competition with the goal to faithfully learn the geometry within the neural radiance field. We further examine the effectiveness of combining semantic and geometric features to identify geometrically highly correlated regions. As demonstrated in Fig. 3, using only one type of the features can lead to incorrect correlations. Tab. 3 also shows that eliminating any one of the features can decrease the performance, confirming the effectiveness of our positive-negative sample selection mechanism leveraging the multimodal features.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Time and memory consumption. Our method provides benefits for 3D reconstruction with limited additional effort. Specifically, it only increases the optimization time and memory usage during the training process, without affecting the model architecture and final storage. The time and memory consumption are reported in Tab. 4. It is observed that the overhead on the training time highly depends on the architecture used in each baseline. For example, the train", + "bbox": [ + 212, + 282, + 486, + 477 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "ing time increase on NeuRIS is marginal while more significant on VolSDF. These studies provide comprehensive information on the trade-offs between training time and performance gain. However, we do find that baselines with the proposed shaping technique can achieve a better performance (e.g., early geometries from NeuS and $\\mathrm{NeuS^{+}}$ shown in Fig. 5) at the same training steps due to the effective regularization by mutual information.", + "bbox": [ + 212, + 479, + 787, + 570 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/c427bf7809e6fe8cede9ddd17ce7837b9eceb758239fe1b10de076dcb76eabca.jpg", + "table_caption": [ + "Table 4: Time and memory consumption." + ], + "table_footnote": [], + "table_body": "
Training Time (h)VRAM (GB)
NeuS [30]+3.00+0.377.89+1.68
VolSDF [37]+3.18+1.9616.36+4.68
GeoNeuS [8]+3.26+0.165.72+3.56
I2-SDF [44]+3.47+2.8410.08+8.09
NeuRIS [29]+3.70+0.2011.73+4.12
MonoSDF [40]+2.65+1.3212.83+9.42
Neuralangelo [13]+3.12+2.0616.44+1.00
", + "bbox": [ + 503, + 340, + 782, + 444 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 592, + 359, + 608 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We explore the usage of the second-order geometric correlations as a regularization and introduce a mutual information shaping technique of the normals for better surface reconstruction quality. With a simple yet effective correlated point selection mechanism that does not depend on accurate geometry information, we can perform a contrastive learning that maximizes the detected correlations. In return, we can observe a boost in the geometric quality of the reconstructed surfaces in state-of-the-art neural scene representations.", + "bbox": [ + 212, + 625, + 787, + 729 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Extensive evaluations on various baselines and datasets show the effectiveness of our method, which improves 3D reconstruction quality without resorting to a more sophisticated surface modeling. The overhead caused by the shaping of the normals varies with different neural architectures of the scene representation, e.g., some are marginal, and some are significant, which we deem as a limitation of our current pipeline. We believe that an acceleration in computation can be achieved using a coarse-to-fine strategy, which we leave as a future work.", + "bbox": [ + 212, + 729, + 787, + 835 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "X. Wang & S. Dong et al.", + "bbox": [ + 271, + 114, + 442, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments. We thank the anonymous reviewers for their valuable feedback. This work is supported by the Early Career Scheme of the Research Grants Council (grant # 27207224), the HKU-100 Award, and in part by NSF China (No. 62172363). Siyan Dong would also like to thank the support from HKU Musketeers Foundation Institute of Data Science for the Postdoctoral Research Fellowship.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 262, + 325, + 277 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bae, G., Budvytis, I., Cipolla, R.: Estimating and exploiting the aleatoric uncertainty in surface normal estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13137-13146 (2021)", + "2. Bao, C., Zhang, Y., Yang, B., Fan, T., Yang, Z., Bao, H., Zhang, G., Cui, Z.: Sine: Semantic-driven image-based nerf editing with prior-guided editing field. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20919-20929 (2023)", + "3. Cai, B., Huang, J., Jia, R., Lv, C., Fu, H.: Neuda: Neural deformable anchor for high-fidelity implicit surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8476-8485 (2023)", + "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)", + "5. Chen, Y., Wu, Q., Zheng, C., Cham, T.J., Cai, J.: Sem2nerf: Converting single-view semantic masks to neural radiance fields. In: European Conference on Computer Vision. pp. 730-748. Springer (2022)", + "6. Curless, B., Levoy, M.: A volumetric method for building complex models from range images. In: Proceedings of the 23rd annual conference on Computer graphics and interactive techniques. pp. 303-312 (1996)", + "7. Do, T., Vuong, K., Roumeliotis, S.I., Park, H.S.: Surface normal estimation of tilted images via spatial rectifier. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 265-280. Springer (2020)", + "8. Fu, Q., Xu, Q., Ong, Y.S., Tao, W.: Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems 35, 3403-3416 (2022)", + "9. Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099 (2020)", + "10. Guo, H., Peng, S., Lin, H., Wang, Q., Zhang, G., Bao, H., Zhou, X.: Neural 3d scene reconstruction with the manhattan-world assumption. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5511-5520 (2022)", + "11. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: Lerf: Language embedded radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19729-19739 (2023)", + "12. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023)" + ], + "bbox": [ + 218, + 295, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "InfoNorm", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023)", + "14. Liu, F., Zhang, C., Zheng, Y., Duan, Y.: Semantic ray: Learning a generalizable semantic field with cross-reprojection attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17386-17396 (2023)", + "15. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13-23 (2023)", + "16. Liu, Z., Milano, F., Frey, J., Siegwart, R., Blum, H., Cadena, C.: Unsupervised continual semantic adaptation through neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3031-3040 (2023)", + "17. Long, X., Lin, C., Wang, P., Komura, T., Wang, W.: Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In: European Conference on Computer Vision. pp. 210-227. Springer (2022)", + "18. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. In: Seminal graphics: pioneering efforts that shaped the field, pp. 347-353 (1998)", + "19. Meng, X., Chen, W., Yang, B.: Neat: Learning neural implicit surfaces with arbitrary topologies from multi-view images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 248-258 (2023)", + "20. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12663-12673 (2023)", + "21. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "22. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)", + "23. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019)", + "24. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "25. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016)", + "26. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., Clarkson, A., Yan, M., Budge, B., Yan, Y., Pan, X., Yon, J., Zou, Y., Leon, K., Carter, N., Briales, J., Gillingham, T., Mueggler, E., Pesqueira, L., Savva, M., Batra, D., Strasdat, H.M., Nardi, R.D., Goesele, M., Lovegrove, S., Newcombe, R.: The replica dataset: A digital replica of indoor spaces (2019)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "X. Wang & S. Dong et al.", + "bbox": [ + 271, + 114, + 442, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "27. Tertikas, K., Despoina, P., Pan, B., Park, J.J., Uy, M.A., Emiris, I., Avrithis, Y., Guibas, L.: Partnerf: Generating part-aware editable 3d shapes without 3d supervision. arXiv preprint arXiv:2303.09554 (2023)", + "28. Wang, F., Galliani, S., Vogel, C., Speciale, P., Pollefeys, M.: Patchmatchnet: Learned multi-view patchmatch stereo. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14194-14203 (2021)", + "29. Wang, J., Wang, P., Long, X., Theobalt, C., Komura, T., Liu, L., Wang, W.: Neuris: Neural reconstruction of indoor scenes using normal priors. In: European Conference on Computer Vision. pp. 139-155. Springer (2022)", + "30. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689 (2021)", + "31. Weder, S., Garcia-Hernando, G., Monszpart, A., Pollefeys, M., Brostow, G.J., Firman, M., Vicente, S.: Removing objects from neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16528-16538 (2023)", + "32. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In: European conference on computer vision. pp. 106-122. Springer (2022)", + "33. Xu, L., Xiangli, Y., Peng, S., Pan, X., Zhao, N., Theobalt, C., Dai, B., Lin, D.: Grid-guided neural radiance fields for large urban scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8296-8306 (2023)", + "34. Xu, X., Yang, Y., Mo, K., Pan, B., Yi, L., Guibas, L.: Jacobinerf: Nerf shaping with mutual information gradients. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16498-16507 (2023)", + "35. Yan, Z., Li, C., Lee, G.H.: Nerf-ds: Neural radiance fields for dynamic specular objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8285-8295 (2023)", + "36. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multi-view stereo. In: Proceedings of the European conference on computer vision (ECCV). pp. 767-783 (2018)", + "37. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021)", + "38. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: Proceedings of the International Conference on Computer Vision (ICCV) (2023)", + "39. Yu, H., Julin, J., Milacski, Z.A., Niinuma, K., Jeni, L.A.: Dylin: Making light field networks dynamic. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12397-12406 (2023)", + "40. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems 35, 25018-25032 (2022)", + "41. Zhang, Y., Yang, G., Guibas, L., Yang, Y.: Infogaussian: Structure-aware dynamic gaussians through lightweight information shaping. arXiv preprint arXiv:2406.05897 (2024)", + "42. Zhi, S., Laidlow, T., Leutenegger, S., Davison, A.J.: In-place scene labelling and understanding with implicit scene representation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15838-15847 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "InfoNorm", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Zhu, B., Yang, Y., Wang, X., Zheng, Y., Guibas, L.: Vdn-nerf: Resolving shape- radiance ambiguity via view-dependence normalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 35-45 (2023)", + "44. Zhu, J., Huo, Y., Ye, Q., Luan, F., Li, J., Xi, D., Wang, L., Tang, R., Hua, W., Bao, H., Wang, R.: I $^2$ -sdf: Intrinsic indoor scene reconstruction and editing via raytracing in neural sdfs. In: CVPR (2023)" + ], + "bbox": [ + 217, + 146, + 785, + 244 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "X. Wang & S. Dong et al.", + "bbox": [ + 271, + 114, + 442, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/InfoNorm_ Mutual Information Shaping of Normals for Sparse-View Reconstruction/8f9c186f-c37a-4b28-8f4e-c701ceacc591_model.json b/2024/InfoNorm_ Mutual Information Shaping of Normals for Sparse-View Reconstruction/8f9c186f-c37a-4b28-8f4e-c701ceacc591_model.json new file mode 100644 index 0000000000000000000000000000000000000000..98eec23e35f9c9932f870fd4cdcfca78f108b688 --- /dev/null +++ b/2024/InfoNorm_ Mutual Information Shaping of Normals for Sparse-View Reconstruction/8f9c186f-c37a-4b28-8f4e-c701ceacc591_model.json @@ -0,0 +1,2524 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.26, + 0.141, + 0.745, + 0.187 + ], + "angle": 0, + "content": "InfoNorm: Mutual Information Shaping of Normals for Sparse-View Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.212, + 0.758, + 0.229 + ], + "angle": 0, + "content": "Xulong Wang\\(^{1,2*}\\) Siyan Dong\\(^{3*}\\) Youyi Zheng\\(^{1\\dagger}\\) Yanchao Yang\\(^{3,4\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.239, + 0.74, + 0.253 + ], + "angle": 0, + "content": "1 State Key Lab of CAD&CG, Zhejiang University 2 Chohotech co. Ltd." + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.254, + 0.694, + 0.268 + ], + "angle": 0, + "content": "3 Institute of Data Science, The University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.268, + 0.785, + 0.282 + ], + "angle": 0, + "content": "\\(^{4}\\) Department of Electrical and Electronic Engineering, The University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.319, + 0.741, + 0.582 + ], + "angle": 0, + "content": "Abstract. 3D surface reconstruction from multi-view images is essential for scene understanding and interaction. However, complex indoor scenes pose challenges such as ambiguity due to limited observations. Recent implicit surface representations, such as Neural Radiance Fields (NeRFs) and signed distance functions (SDFs), employ various geometric priors to resolve the lack of observed information. Nevertheless, their performance heavily depends on the quality of the pre-trained geometry estimation models. To ease such dependence, we propose regularizing the geometric modeling by explicitly encouraging the mutual information among surface normals of highly correlated scene points. In this way, the geometry learning process is modulated by the second-order correlations from noisy (first-order) geometric priors, thus eliminating the bias due to poor generalization. Additionally, we introduce a simple yet effective scheme that utilizes semantic and geometric features to identify correlated points, enhancing their mutual information accordingly. The proposed technique can serve as a plugin for SDF-based neural surface representations. Our experiments demonstrate the effectiveness of the proposed in improving the surface reconstruction quality of major states of the arts. Our code is available at: https://github.com/Muliphein/InfoNorm." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.596, + 0.691, + 0.609 + ], + "angle": 0, + "content": "Keywords: 3D from multi-view images \\(\\cdot\\) surface reconstruction" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.636, + 0.377, + 0.652 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.788, + 0.789 + ], + "angle": 0, + "content": "3D surface reconstruction from multi-view images is an essential task in the computer vision and graphics community, with practical applications including content creation for virtual reality and robot-scene interaction. However, indoor scenes present challenges due to the large and complex scenes captured from sparse viewpoints. These challenges encompass issues such as occlusion and ambiguity arising from limited observations. As a result, traditional Multi-View Stereo (MVS) methods [28,36], which require substantial overlap among images, may not produce satisfactory results in these scenarios." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.797, + 0.66, + 0.812 + ], + "angle": 0, + "content": "* Equal contributions (wangxulong@zju.edu.cn, siyan3d@hku.hk)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.812, + 0.393, + 0.826 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.826, + 0.722, + 0.84 + ], + "angle": 0, + "content": "2 This work was done during the author's internship at Chohotech Co. Ltd.." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.797, + 0.722, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.446, + 0.129 + ], + "angle": 0, + "content": "X. Wang & S. Dong et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.402, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.231, + 0.326, + 0.24 + ], + "angle": 0, + "content": "VoISDF" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.147, + 0.593, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.231, + 0.535, + 0.242 + ], + "angle": 0, + "content": "VoISDF+ (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.147, + 0.782, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.231, + 0.72, + 0.241 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.244, + 0.405, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.287, + 0.308, + 0.332, + 0.317 + ], + "angle": 0, + "content": "GeoNeuS" + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.244, + 0.591, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.461, + 0.308, + 0.542, + 0.318 + ], + "angle": 0, + "content": "GeoNeuS\\(^{+}\\) (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.245, + 0.78, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.308, + 0.72, + 0.317 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.329, + 0.787, + 0.386 + ], + "angle": 0, + "content": "Fig. 1: 3D scene reconstruction from sparse views on Replica [26] (first row) and ScanNet++ [38] (second row). By enforcing the mutual information between the normals of highly correlated scene points, the proposed method can effectively enhance the reconstruction quality of the baselines (VolSDF [37] and GeoNeuS [8])." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.785, + 0.584 + ], + "angle": 0, + "content": "Recently, implicit scene representations, e.g., Neural Radiance Fields (NeRFs) [21], have been proposed to encode 3D scenes into a set of neural network parameters, with subsequent works enhancing the representational capacity for geometry to enable surface extraction. The key idea involves incorporating signed distance functions (SDFs) [23] into the learned radiance field. Although NeRFs with SDF produce high-quality reconstructions of simple scenes with sufficient images, they still underperform in large and complex scenes captured from sparse viewpoints. More recent approaches, such as NeuRIS [29] and MonoSDF [40], propose using monocular geometric priors to guide the learning of the radiance field. However, they can be affected by inevitable errors and noise in the monocular geometry estimation modules." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.586, + 0.787, + 0.797 + ], + "angle": 0, + "content": "We explore geometric regularization in a second-order perspective. We argue that both the geometry itself and the correlation of geometry among scene regions are crucial for surface reconstruction, especially when dealing with varying indoor scenes under sparse views with noisy geometry estimates. The correlation is expressed as mutual information between scene entities under random perturbations of the network weights. More explicitly, we propose to enforce the mutual information between the normals of two points in the scene that are considered geometrically correlated, where the mutual information can be efficiently computed by measuring the cosine similarity between the gradients of the normals with respect to the perturbed weights. Further, we leverage a combination of pretrained semantic and geometric features [4,7,12] to identify the correlated scene regions that endorse high mutual information. This results in an easy-to-use geometric shaping1 technique that can be applied to any SDF-based neural radiance fields. To thoroughly evaluate its effectiveness, we test the proposed with" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.785, + 0.84 + ], + "angle": 0, + "content": "1 We use the term \"shaping\" to refer to the process of optimizing network parameters to enforce consistencies in the scene geometry via aligning the gradients." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InfoNorm" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "multiple state-of-the-art neural surface reconstruction models. The experimental results confirm the effectiveness of the proposed shaping technique, showing improved surface reconstruction quality (e.g., Fig. 1) on challenging scenes with limited views. To summarize:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.219, + 0.787, + 0.264 + ], + "angle": 0, + "content": "- We propose to leverage second-order geometric correlation in the form of mutual information between normals to regularize the surface reconstruction of SDF-based NeRF representations." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.264, + 0.787, + 0.31 + ], + "angle": 0, + "content": "- We develop a pipeline that efficiently enforces the mutual information of the normals of points, deemed geometrically correlated with pre-trained multimodal features, to improve the reconstruction quality of the scene." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.311, + 0.787, + 0.356 + ], + "angle": 0, + "content": "- We verify the effectiveness of the proposed second-order regularization technique across a broad spectrum of baselines and demonstrate its usefulness as an easy-to-use plugin in improving 3D surface modeling." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.219, + 0.787, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.38, + 0.388, + 0.397 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.788, + 0.505 + ], + "angle": 0, + "content": "3D Reconstruction with neural radiance fields (NeRFs). NeRF [21] is originally designed for novel view synthesis, with its variants serving as scene representations [15, 32, 33, 35, 39] for tasks related to editing [2, 31, 44], semantic segmentation [5, 14, 16, 34], 3D shape generation [20, 27], and so on. We mainly focus on neural surface representations that enable the extraction of high-quality 3D meshes from the trained NeRFs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.789, + 0.688 + ], + "angle": 0, + "content": "Early works such as NeuS [30] and VolSDF [37] combine signed distance functions (SDFs) with volume rendering to extract high-quality surfaces. Recently, SparseNeuS [17] introduces a cascaded geometry reasoning framework that generalizes well to novel scenes. GeoNeuS [8] uses surface points from structure-from-motion and neighbor-view patches to reconstruct surfaces, while NeuDA [3] employs a deformable anchor to adaptively encode geometric details. NeAT [19] broadens the scope of reconstruction to arbitrary topology, no longer confining to watertight surfaces. Moreover, Neuralangelo [13] uses a hash grid for position encoding and implements a coarse-to-fine optimization with numerical gradients. Despite their promising results, most of these methods need sufficient training views, and their performance declines when dealing with large and complex scenes under sparse viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Incorporating priors into NeRFs. One can augment scene representations by introducing various priors to NeRFs. VDN-NeRF [43] normalizes the spatial feature to align with monocular features. LERF [11] introduces a new branch whose outputs align with DINO [4] and CLIP [24] features. Semantic-NeRF [42] encodes semantic information for segmentation tasks, while JacobiNeRF [34] shapes the color gradients to encode the mutual information in terms of semantic similarity from DINO [4] features for label propagation." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "For surface reconstruction, GeoNeuS [8] incorporates patch similarity from neighboring images and sparse surface points derived from structure-from-motion" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.444, + 0.128 + ], + "angle": 0, + "content": "X. Wang & S. Dong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.359 + ], + "angle": 0, + "content": "[25], while I²-SDF [44] combines the rendering equation with VolSDF [37] for better novel views and geometry. To cope with complex indoor scenes, NeuRIS [29] incorporates the normal priors as additional geometric supervision. Further, MonoSDF [40] uses both depth and normal to improve the surface reconstruction quality. However, these methods highly depend on the quality of the supplemented geometric information. In contrast, our method leverages the second-order information in geometric prior, rather than directly using the prior as a first-order supervision signal, which makes it more robust to errors and noise. There are also methods utilizing strong assumptions derived for human-made scenes. For instance, ManhattanSDF [10] works under the assumption that all captured scenes adhere to the Manhattan-world concept, thus, predicting semantic segmentation to identify walls and floors. While these methods primarily focus on planar scenes, the proposed technique can work for general scenes under noisy correlation information." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.385, + 0.331, + 0.4 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.789, + 0.571 + ], + "angle": 0, + "content": "We start with a discussion on the preliminaries in Sec. 3.1 to provide the background on how we can learn from posed images an implicit scene representation (NeRF), improve 3D scene reconstruction via the combination of NeRF and SDF representations, and employ mutual information as a constraint in the NeRF training process. Following this, in Sec. 3.2, we delve into how mutual information can be used as constraints on NeRF's implied surfaces, so that we can learn better geometries for higher-quality 3D reconstruction of the scene. Finally, we elaborate on how we can employ off-the-shelf semantic and geometry features to assist the encoding of mutual information through the proposed geometric shaping framework." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.596, + 0.373, + 0.61 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.623, + 0.788, + 0.746 + ], + "angle": 0, + "content": "Neural radiance fields. NeRF [21] learns an implicit scene representation from a set of posed images. For a specific pixel in an image, NeRF samples a list of 3D points, \\(\\{x|x_{t} = \\mathbf{o} + t\\mathbf{v},t\\in [t_{near},t_{far}]\\}\\), from the camera center \\(\\mathbf{o}\\) along the viewing direction \\(\\mathbf{v}\\) within a bounded space. Specifically, it employs a Multilayer Perceptron (MLP) to encode the volume density and color of each sampled point. We denote all the parameters of NeRF as \\(\\theta\\), the density function as \\(\\sigma\\), and the color function as \\(c\\). The pixel's rendered color \\(\\hat{\\mathbf{C}}\\) from a NeRF is calculated by the discrete volume rendering:" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.759, + 0.785, + 0.808 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {C}} (\\mathbf {o}, \\mathbf {v}; \\theta) = \\sum_ {i = 1} ^ {N} \\omega (\\mathbf {o}, \\mathbf {v}, t _ {i}; \\theta) c \\left(x _ {t _ {i}}, \\mathbf {v}; \\theta\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.801, + 0.637, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} \\omega (\\mathbf {o}, \\mathbf {v}, t _ {i}; \\theta) = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}) \\alpha_ {i}.\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InfoNorm" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "Here, \\(\\alpha_{i} = 1 - \\exp (-\\sigma_{i}\\delta_{i})\\), where \\(\\sigma_{i}\\) is the abbreviation of \\(\\sigma (x_{t_i};\\theta)\\) and \\(\\delta_{i} = t_{i + 1} - t_{i}\\). A photometric loss is utilized to minimize the discrepancy between the rendered color and its ground truth:" + }, + { + "type": "equation", + "bbox": [ + 0.442, + 0.205, + 0.785, + 0.222 + ], + "angle": 0, + "content": "\\[\nL _ {C} = \\left\\| \\mathbf {C} - \\hat {\\mathbf {C}} \\right\\| _ {1}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.234, + 0.784, + 0.295 + ], + "angle": 0, + "content": "so that the parameters \\(\\theta\\) can be optimized to fit the scene. Despite the ability to handle photorealistic novel view synthesis, NeRFs suffer from noisy and unrealistic surface extraction when it comes to 3D reconstruction, as discussed in the literature [30, 37]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.311, + 0.784, + 0.415 + ], + "angle": 0, + "content": "NeRFs with signed distance functions. To produce high-quality surfaces, it is beneficial to use signed distance functions (SDFs) [6, 23] to represent geometry. Consequently, the surface can be reconstructed by extracting the zero-level set [18] of the SDF values. The combination of NeRF and SDF can be achieved by deriving the NeRF's densities from the SDF, or defining Laplace's cumulative distribution function. Accordingly, \\(\\alpha_{i}\\) in Eq. 1 can be represented by a function \\(\\varPhi\\) which converts SDF values to volume densities as" + }, + { + "type": "equation", + "bbox": [ + 0.431, + 0.431, + 0.785, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} = \\Phi (x, f (x; \\theta), i). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.458, + 0.784, + 0.489 + ], + "angle": 0, + "content": "Generally, NeRFs with an SDF achieve improved surface reconstruction quality, where the key to maintaining the characteristics of SDF is an eikonal loss [9]:" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.502, + 0.785, + 0.542 + ], + "angle": 0, + "content": "\\[\nL _ {E} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\left| \\left| \\nabla f \\left(x _ {i}\\right) \\right| \\right| _ {2} - 1\\right) ^ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.557, + 0.788, + 0.725 + ], + "angle": 0, + "content": "Mutual information in NeRFs. Mutual information (MI) quantifies the statistical dependence between two random variables, providing an estimate of how much information they share. It has been introduced for semantic segmentation and motion modeling tasks [34, 41]. In this paper, We study the mutual information under perturbations of the NeRF parameters, and focus more on the geometry properties. Let \\( I(p_{i}) \\) and \\( I(p_{j}) \\) denote two pixels derived from the NeRF rendering process \\( \\mathbb{F} \\). And we assume that a subset of the NeRF weights, denoted by \\( \\theta^{D} \\), is perturbed by a random noise vector \\( \\mathbf{n} \\in \\mathbb{R}^{D} \\) sampled from a uniform distribution on the sphere \\( \\mathbb{S}^{D - 1} \\) by a small step \\( \\gamma \\ll 1.0 \\). The random variables corresponding to the perturbed pixels can be written through a Taylor expansion:" + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.735, + 0.589, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\hat {I} (p _ {i}) = I (p _ {i}) + \\gamma \\mathbf {n} \\frac {\\partial \\mathbb {F} _ {i}}{\\partial \\theta^ {D}},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.766, + 0.589, + 0.795 + ], + "angle": 0, + "content": "\\[\n\\hat {I} (p _ {j}) = I (p _ {j}) + \\gamma \\mathbf {n} \\frac {\\partial \\mathbb {F} _ {j}}{\\partial \\theta^ {D}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.807, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We can now characterize the mutual information between \\(\\hat{I}(p_i)\\) and \\(\\hat{I}(p_j)\\) under the joint probability distribution \\(\\mathbb{P}(\\hat{I}(p_i), \\hat{I}(p_j))\\). As derived in JacobiNeRF, the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.446, + 0.129 + ], + "angle": 0, + "content": "X. Wang & S. Dong et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.512, + 0.162 + ], + "angle": 0, + "content": "mutual information can be computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.172, + 0.657, + 0.226 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbb {I} (\\hat {I} (p _ {i}), \\hat {I} (p _ {j})) = \\mathbb {H} (\\hat {I} (p _ {i})) - \\mathbb {H} (\\hat {I} (p _ {i}) | \\hat {I} (p _ {j})) \\\\ = \\log (\\frac {1}{\\sqrt {1 - \\cos^ {2} \\beta}}) + \\mathrm {c o n s t .}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.234, + 0.78, + 0.252 + ], + "angle": 0, + "content": "Here, \\(\\beta\\) denotes the angle between \\(\\partial \\mathbb{F}_i / \\partial \\theta^D\\) and \\(\\partial \\mathbb{F}_j / \\partial \\theta^D\\). It is observed that:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.262, + 0.786, + 0.28 + ], + "angle": 0, + "content": "\\[\n\\mathbb {I} (\\hat {I} (p _ {i}), \\hat {I} (p _ {j})) \\propto | | \\cos \\beta | |. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.287, + 0.787, + 0.363 + ], + "angle": 0, + "content": "Therefore, to enhance the MI between two random pixels induced by NeRF perturbations, one can enforce the absolute value of the cosine similarity. That is, if two pixels \\( p_i \\), \\( p_j \\) come with high correlation (e.g., from the same object in the scene), their gradients regarding the perturbed parameters should be in the same or opposite direction." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.383, + 0.786, + 0.399 + ], + "angle": 0, + "content": "3.2 Mutual Information Shaping of Geometry in SDF-based NeRFs" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.407, + 0.786, + 0.451 + ], + "angle": 0, + "content": "While the previous approach focuses on enforcing mutual information for image synthesis, we explore a mutual information shaping of the geometries for quality surface reconstruction in SDF-based NeRFs." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.453, + 0.788, + 0.649 + ], + "angle": 0, + "content": "Specifically, we represent a scene via a NeRF backbone with an SDF head and a color head, as shown in Fig. 2. Note that the method to be proposed is capable of being applied to any network similar to NeRF and SDF representations. Given that the two heads partially share parameters, we denote all the parameters as \\(\\theta\\), and use \\(f(x;\\theta)\\) and \\(c(x,\\mathbf{v};\\theta)\\) to represent the SDF and color branches, respectively. To achieve mutual information shaping of the geometry, we focus on manipulating the parameters of \\(f(x;\\theta)\\), while keeping \\(c(x,\\mathbf{v};\\theta)\\) unchanged. Below, we first explain why and how we compute mutual information with surface normals derived from SDF, i.e., \\(f(x;\\theta)\\), as a geometry-aware constraint. Next, we present a simple yet effective strategy that extracts regions with high correlation, which need to be encoded into the weights using the mutual information shaping technique. Finally, we elaborate on our loss function and training process to reconstruct the scene from posed images." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.661, + 0.788, + 0.752 + ], + "angle": 0, + "content": "Mutual information among surface normals. We explicitly encode mutual information into the function \\( f(x; \\theta) \\). However, instead of computing mutual information between the density values, we proceed by manipulating the surface normal vectors that can be derived from \\( f \\). Specifically, for a 3D point \\( x \\) on a surface of the scene, its normal vector \\( N(x; \\theta) \\) can be calculated as the partial derivative of the 3D coordinate:" + }, + { + "type": "equation", + "bbox": [ + 0.433, + 0.759, + 0.786, + 0.789 + ], + "angle": 0, + "content": "\\[\nN (x; \\theta) = \\frac {\\partial f (x ; \\theta)}{\\partial x}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "There are several reasons to manipulate surface normal vectors but not SDF values: 1) Shaping the SDF values with respect to a subset \\(\\theta^D\\) of the parameters may not effectively impose the constraints on all parameters. However, the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InfoNorm" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.146, + 0.782, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.281, + 0.788, + 0.392 + ], + "angle": 0, + "content": "Fig. 2: An overview of the pipeline, where we apply mutual information shaping on the geometric branch to enforce consistencies that help enhance the surface reconstruction. Specifically, the model consists of the NeRF backbone, an SDF head, and a color head. The predicted density and color are supervised by the classic eikonal loss \\( L_{E} \\) and the photometric reconstruction loss \\( L_{C} \\), respectively. We encode geometry-aware mutual information into a subset of the parameters \\( \\theta^{D} \\) of \\( f(x) \\) to constrain the density field learning for better surface quality, which is achieved by the proposed mutual information loss \\( L_{M} \\) computed on top of the estimated surface normal." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.784, + 0.51 + ], + "angle": 0, + "content": "calculation of the surface normal includes the entire set of parameters in \\( f \\), thus a more global regularity. 2) Moreover, shaping \\( f \\) with the mutual information between its values may induce side effects as density could be distorted to satisfy the constraints. However, shaping the normal \\( \\partial f / \\partial x \\) leave more space for the SDF values to cope with the scene reconstruction task, thus minimizing potential negative effects." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.512, + 0.785, + 0.558 + ], + "angle": 0, + "content": "According to Eq. 5, we need to calculate the partial gradients \\(\\partial N / \\partial \\theta^D\\) to compute the mutual information. When it comes to a vector rather than a scalar, the gradients consist of three components corresponding to the \\(\\mathbf{x}\\), \\(\\mathbf{y}\\), and \\(\\mathbf{z}\\) axes:" + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.567, + 0.785, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial N (x ; \\theta)}{\\partial \\theta^ {D}} = \\frac {\\partial N _ {\\mathbf {x}} (x ; \\theta)}{\\partial \\theta^ {D}}, \\frac {\\partial N _ {\\mathbf {y}} (x ; \\theta)}{\\partial \\theta^ {D}}, \\frac {\\partial N _ {\\mathbf {z}} (x ; \\theta)}{\\partial \\theta^ {D}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.602, + 0.784, + 0.631 + ], + "angle": 0, + "content": "For ease of computation, we concatenate the gradients into a vector, denoted as \\(\\partial N\\) for simplicity:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.63, + 0.785, + 0.659 + ], + "angle": 0, + "content": "\\[\n\\partial N = \\operatorname {C o n c a t} \\left(\\frac {\\partial N _ {\\mathbf {x}}}{\\partial \\theta^ {D}}, \\frac {\\partial N _ {\\mathbf {y}}}{\\partial \\theta^ {D}}, \\frac {\\partial N _ {\\mathbf {z}}}{\\partial \\theta^ {D}}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.663, + 0.784, + 0.694 + ], + "angle": 0, + "content": "Since a 2D pixel is accumulated from 3D, we aggregate the normal-based information along a viewing ray as in Eq. 1 as:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.703, + 0.785, + 0.735 + ], + "angle": 0, + "content": "\\[\n\\mathbb {N} \\left(p _ {i}\\right) = \\sum_ {x \\in \\text {r a y}} \\omega \\left(t _ {i}; \\theta\\right) \\partial N. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.744, + 0.733, + 0.759 + ], + "angle": 0, + "content": "Finally, we obtain that the mutual information between two normals is:" + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.768, + 0.785, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\mathbb {I} (\\hat {N} (p _ {i}), \\hat {N} (p _ {j})) \\propto | | \\frac {\\mathbb {N} (p _ {i}) \\cdot \\mathbb {N} (p _ {j})}{| \\mathbb {N} (p _ {i}) | | \\mathbb {N} (p _ {j}) |} | |. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Please refer to our supplementary material for a detailed derivation. From Eq. 10, we observe that the mutual information in surface normal vectors correlates" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.446, + 0.129 + ], + "angle": 0, + "content": "X. Wang & S. Dong et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.146, + 0.357, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.227, + 0.319, + 0.237 + ], + "angle": 0, + "content": "Anchor pixel" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.146, + 0.499, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.227, + 0.49, + 0.237 + ], + "angle": 0, + "content": "Similar semantic features" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.146, + 0.64, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.227, + 0.635, + 0.237 + ], + "angle": 0, + "content": "Similar geometric features" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.146, + 0.783, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.227, + 0.756, + 0.237 + ], + "angle": 0, + "content": "Combined (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.25, + 0.788, + 0.348 + ], + "angle": 0, + "content": "Fig. 3: An example of positive samples from different features. Given an anchor pixel (marked by a blue circle in the first image) on the wall-like cabinet, semantic features like DINO often correlate both the cabinet and the ceiling (as shown by the red area in the second image). Meanwhile, geometric features such as (noisy) monocular normals can not distinguish between parallel planes that are not connected (the third image). By combining semantic and geometric features, we can obtain positive samples of the anchor pixel with better geometric consistency, thus, high mutual correlation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.437 + ], + "angle": 0, + "content": "positively with the cosine similarity of their gradients to the network parameters. This conclusion is similar to that of Eq. 5. As a result, we can now identify the pixels that are of high correlations, and increase the aggregated mutual information during the NeRF training process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.583 + ], + "angle": 0, + "content": "Extracting correlated scene regions from image pixels. To extract scene regions that need to be shaped to have high mutual information, we present a simple yet effective approach with off-the-shelf visual features. Given the input images, we extract pixel-wise semantic and geometric features to identify correlated scene regions. Following previous research [34], we apply DINO [4] as the semantic feature extractor to obtain a high-level understanding of the scene regions. Meanwhile, as we shape mutual information on top of the surface normal, we employ a monocular normal estimator [7] as the geometric feature extractor for a detailed low-level perception of the 3D points in the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.734 + ], + "angle": 0, + "content": "DINO features can group different parts of an object, like the seat and back of a chair, which share similar semantics. However, the surface normal vectors from these parts have limited correlation, thus should not be constrained. Conversely, using only geometric features can wrongly suggest high correlation between unrelated areas, like the seat and the floor. As these areas aren't closely connected in 3D space, there is no reason to have better surface reconstruction when enhancing their mutual information. Similar phenomena on wall-like structures can be observed in Fig. 3, where we show that by combining semantic and geometric features together we can better identify scene regions that are spatially connected and have similar normal directions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "More explicitly, we combine semantic and geometric features with an intersection operator. It is simple yet effective. The process of identifying correlated scene regions (in normal) is implemented as positive pixel pair selection from input images. Given a randomly selected pixel from an image, we first extract similar pixels from neighboring images using semantic and geometric features, respectively. This extraction is implemented by applying two thresholds, \\(\\beta_{S}\\) and \\(\\beta_{G}\\), to filter the cosine similarity for semantic and geometric features. We then" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "InfoNorm" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.239 + ], + "angle": 0, + "content": "feed the intersection of the top similar pixels from each feature type into NeRF's training process as positive samples, allowing us to shape them with high mutual information. In order to leverage the contrastive learning framework to encourage mutual information, we randomly select pixels not involved in the aforementioned process to serve as negative samples (independent pixels that are not correlated to the anchor pixel)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.249, + 0.784, + 0.281 + ], + "angle": 0, + "content": "Loss functions for training. The aforementioned contrastive learning process is performed with a tailored InfoNCE loss [22]:" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.287, + 0.786, + 0.323 + ], + "angle": 0, + "content": "\\[\nL _ {M} = - \\log \\frac {\\sum \\exp (| | \\cos (\\partial \\mathbb {N} _ {i} , \\partial \\mathbb {N} _ {i +}) | |)}{\\sum_ {j} \\exp (| | \\cos (\\partial \\mathbb {N} _ {i} , \\partial \\mathbb {N} _ {j}) | |)}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.328, + 0.785, + 0.448 + ], + "angle": 0, + "content": "where \\(i\\), \\(i+\\), and \\(j\\) denote the anchor, positive, and negative pixels, respectively. This loss function effectively enhances mutual information (according to Eq. 10) among correlated pixels, while maintaining low mutual information among independent pixels. Finally, to train the NeRF backbone and retain the characteristics of the SDF, we adopt the basic photometric loss (Eq. 2) and the eikonal loss (Eq. 4). The latter two are derived from single images, while the former can be calculated from either single images or a set of neighboring images. The full loss function is defined as the weighted sum of three loss terms:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.457, + 0.785, + 0.472 + ], + "angle": 0, + "content": "\\[\nL = L _ {C} + \\lambda_ {E} L _ {E} + \\lambda_ {M} L _ {M}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.48, + 0.713, + 0.495 + ], + "angle": 0, + "content": "With the above, all network parameters \\(\\theta\\) can be trained end-to-end." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.515, + 0.376, + 0.533 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.785, + 0.664 + ], + "angle": 0, + "content": "In this section, we evaluate the effectiveness of our method through experiments conducted on multiple datasets with various baselines. First, we introduce the two datasets used to conduct the experiments in Sec. 4.1. Next, we detail the implementation of the proposed method in Sec. 4.2, where we adapt a spectrum of state-of-the-art models for validating the effectiveness of the geometric shaping. Then, we report the comparisons between the baselines and our models in Sec. 4.3. Lastly, we conduct a series of ablation studies in Sec. 4.4, to verify the necessity of each component in our method." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.683, + 0.336, + 0.697 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We evaluate the proposed method using two public datasets from the literature: ScanNet++ [38] and the Replica [26] dataset. For the ScanNet++ dataset, we choose 8 representative scenes, while for the Replica dataset, we select 4 representative scenes. The scenes from both datasets represent a variety of indoor settings, including offices, bedrooms, and so on. As our focus is on scene reconstruction from sparse views, we uniformly sample 50-85 images from the official (and Replica from Semantic-NeRF [42]) scanning sequences for each scene to use as input in our evaluation. For detailed statistics of each scene, please refer to our supplementary material." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.444, + 0.129 + ], + "angle": 0, + "content": "X. Wang & S. Dong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Evaluation metrics. Following previous works [13, 30, 40], we evaluate the quality of surface reconstruction using two metrics: Chamfer distance and F-score. Chamfer distance is a measure that evaluates the difference between reconstructed and ground-truth points sets. F-score is a statistical metric that integrates precision and recall, providing a more comprehensive understanding of the reconstruction quality." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.259, + 0.458, + 0.274 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.282, + 0.789, + 0.584 + ], + "angle": 0, + "content": "As the proposed geometric shaping can be easily integrated into any network similar to NeRF with SDF representations, we select several state-of-the-art models as our baselines and apply the proposed to them for demonstrating the effectiveness of our approach. We first choose NeuS [30] and VolSDF [37], the two representative pioneers for NeRFs with SDFs. We also apply our method to the following works: GeoNeuS [8], I²-SDF [44], NeuRIS [29], MonoSDF [40], and Neuralangelo [13]. In addition to the basic loss functions for NeRF and SDF, GeoNeuS incorporates patch similarity from neighboring views and surface points derived from structure-from-motion. I²-SDF integrates the rendering equation with VolSDF. We disregard its normal and depth loss terms. NeuRIS utilizes monocular normal estimation as a supervision on the geometry, and discards the low-quality normal information that lacks sufficient confidence through multi-view patching. MonoSDF employs both depth and normal estimation, for which we set the decay of the normal and depth at 30k iterations to prevent the method from degeneration. Neuralangelo utilizes a multi-resolution hash grid and numerical gradient. We set the hash encoding dictionary size to 20, and the encoding dimension to 4. For all the base models mentioned above, we enhance them by incorporating our proposed mutual information loss during training. The resulting methods are respectively labeled as \\(\\mathrm{NeuS^{+}}\\), VolSDF+, GeoNeuS+, and so on." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.606, + 0.371, + 0.621 + ], + "angle": 0, + "content": "4.3 Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Results on the ScanNet++ dataset. The quantitative comparisons are reported in Tab. 1. We also visualize the surface reconstruction results in Fig. 4 for qualitative comparisons. We can observe that by applying the proposed mutual information shaping on the geometry, the surface reconstruction quality can be improved across all baselines, as measured by mean Chamfer distance, F-score, and the visuals." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.81 + ], + "angle": 0, + "content": "NeuS and VolSDF tend to produce rough surfaces and excessively smoothed boundaries. However, with our geometry-aware mutual information, the surfaces become smoother, while the boundaries are more crispy. Compared to GeoNeuS, a method that uses patch-similarity among neighboring views and requires sufficient textures, our method provides significant improvements. Applying the proposed technique to \\(\\mathrm{I}^2\\)-SDF can also enhance its reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "As for more recent works, we find that NeuRIS is somewhat sensitive to the quality of normal estimation. It can either be affected by noise in the normal" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InfoNorm" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 1: Quantitative results on the ScanNet++ dataset. Each cell includes the original baseline number in black, followed by the improvement using our mutual information. Positive and negative improvements are marked in green and orange, respectively. The last column shows the average improvement over all scenes, validating the effectiveness of the proposed shaping for better geometric reconstruction (despite a small potion of noise on a few scenes)." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.241, + 0.784, + 0.407 + ], + "angle": 0, + "content": "
0a7c0a186ee27b6456a09460a08de0abMean
NeuS [30]+0.08-0.040.05-0.030.03+0.010.06-0.040.01-0.000.24-0.140.04-0.030.02-0.000.07-0.03
VolSDF [37]+0.90-0.830.31-0.290.04+0.020.10-0.030.03-0.010.14+0.030.39-0.380.04-0.000.24-0.19
GeoNeuS [8]+0.03-0.021.03-1.020.28-0.270.05-0.030.02-0.010.22-0.200.01-0.010.26-0.250.24-0.23
I2-SDF [44]+0.27-0.240.22-0.030.04-0.010.47-0.450.03-0.020.13+0.050.03+0.020.30-0.070.19-0.09
NeuRIS [29]+0.01-0.000.01-0.000.03+0.020.07-0.040.04-0.030.41-0.210.03-0.010.03-0.000.08-0.03
MonoSDF [40]+0.01-0.000.01+0.020.02-0.010.02-0.000.01-0.000.05-0.020.01-0.000.01-0.000.02-0.00
Neuralangelo [13]+0.17-0.010.84-0.660.52-0.330.05-0.010.04+0.024.76-2.850.33+0.050.12+0.030.85-0.48
NeuS [30]+0.49+0.060.48+0.100.67+0.000.58+0.160.81-0.010.58-0.020.67+0.170.75+0.090.63+0.07
VolSDF [37]+0.45+0.350.47+0.270.67+0.000.55+0.050.71+0.060.52+0.180.48+0.350.70+0.020.57+0.16
GeoNeuS [8]+0.69+0.200.47+0.440.52+0.340.70+0.060.87+0.030.68+0.230.75+0.140.37+0.510.63+0.24
I2-SDF [44]+0.33+0.520.28+0.000.58+0.150.41+0.430.75+0.100.65+0.070.65+0.070.40+0.250.51+0.20
NeuRIS [29]+0.72+0.030.68+0.010.65+0.010.69+0.060.85+0.010.24+0.170.64+0.130.69+0.090.65+0.06
MonoSDF [40]+0.92-0.010.88-0.010.81+0.040.79+0.060.89+0.000.88+0.050.91+0.000.89-0.010.87+0.01
Neuralangelo [13]+0.42+0.050.14+0.300.35+0.120.63+0.050.71-0.010.17+0.220.29+0.060.51-0.010.40+0.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.321, + 0.427, + 0.68, + 0.441 + ], + "angle": 0, + "content": "Table 2: Quantitative results on the Replica dataset." + }, + { + "type": "table", + "bbox": [ + 0.292, + 0.453, + 0.713, + 0.625 + ], + "angle": 0, + "content": "
office0office1room0room1Mean
NeuS [30]+0.005+0.0010.023-0.0120.068-0.0380.130-0.0870.057-0.035
VolSDF [37]+0.004+0.0060.016-0.0040.080-0.0710.034-0.0250.033-0.023
GeoNeuS [8]+0.023-0.0210.014-0.0090.040-0.0350.002-0.0000.020-0.017
I2-SDF [44]+0.032-0.0090.056-0.033--0.044-0.021
NeuRIS [29]+0.014-0.0063.975-3.9630.025-0.0170.015+0.0021.007-0.996
MonoSDF [40]+0.003-0.0000.005-0.0010.004-0.0000.004-0.0010.004-0.001
Neuralangelo [13]+0.024-0.0180.030-0.0190.146-0.0120.082-0.0470.070-0.024
NeuS [30]+0.92+0.000.80+0.010.78-0.030.72+0.080.80+0.02
VolSDF [37]+0.94-0.040.78+0.010.78+0.160.82+0.070.83+0.05
GeoNeuS [8]+0.89+0.080.91+0.040.90+0.090.98+0.010.920+0.05
I2-SDF [44]+0.72+0.090.56+0.20--0.64+0.14
NeuRIS [29]+0.79+0.060.01+0.680.89+0.050.77+0.030.61+0.20
MonoSDF [40]+0.96+0.000.90+0.020.98+0.000.95+0.020.95+0.01
Neuralangelo [13]+0.89+0.070.49+0.360.65+0.150.71+0.180.69+0.19
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.84 + ], + "angle": 0, + "content": "or completely disregard the normal information due to low confidence. Particularly, when images are sparse, the regions covered by fewer viewpoints tend to lose supervision. Our mutual information scheme does not depend on the absolute values of the surface normal, but on the similarity among these values. This makes it more robust to noise and less likely to discard low-quality but useful normal information. Consequently, our method enhances the reconstruction quality of NeuRIS. MonoSDF achieves the best overall performance among the baselines because it uses both depth and normal estimations as geometry supervision. However, our method still offers improvements. Notably, it produces smoother planes and clearer boundaries. This indicates that not only the supervision of geometric cues is important, but also the correlation between the geometric entities matters." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.444, + 0.129 + ], + "angle": 0, + "content": "X. Wang & S. Dong et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.145, + 0.782, + 0.729 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.739, + 0.788, + 0.796 + ], + "angle": 0, + "content": "Fig. 4: Visual results on the ScanNet++ dataset. Each row shows a comparison with a different baseline, from top to bottom, NeuS \\([30]^+\\), VolSDF \\([37]^+\\), GeoNeuS \\([8]^+\\), I²-SDF \\([44]^+\\), NeuRIS \\([29]^+\\), MonoSDF \\([40]^+\\), and Neuralangelo \\([13]^+\\). Red boxes are overlaid to help the contrast." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InfoNorm" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.146, + 0.354, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.247, + 0.226, + 0.328, + 0.236 + ], + "angle": 0, + "content": "Reference image" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.146, + 0.497, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.417, + 0.226, + 0.444, + 0.235 + ], + "angle": 0, + "content": "NeuS" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.146, + 0.64, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.226, + 0.606, + 0.236 + ], + "angle": 0, + "content": "Neu\\(^{+}\\) (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.146, + 0.782, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.226, + 0.748, + 0.236 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.248, + 0.785, + 0.277 + ], + "angle": 0, + "content": "Fig. 5: Reconstruction results after 10K iterations in the training process with \\((\\mathrm{NeuS}^{+})\\) and without (NeuS) the proposed geometric shaping technique." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.784, + 0.368 + ], + "angle": 0, + "content": "Last but not least, the experiments on Neuralangelo demonstrate the potential of our method in working with hash grid encoding and numerical gradient. This highlights our method's versatility and ease of use, which can be broadly applied in various reconstruction architectures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.785, + 0.474 + ], + "angle": 0, + "content": "Results on the Replica dataset. In Tab. 2 we report the numerical results on the Replica dataset, with respect to NeuS [30], VolSDF [37], GeoNeuS [8], I²-SDF [44], NeuRIS [29], MonoSDF [40], and Neuralangelo [13]. As observed, the proposed geometric shaping technique still provides positive improvements when incorporated into these baselines, aligning with the conclusion drawn from the ScanNet++ data. Qualitative examples can also be found in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.497, + 0.336, + 0.513 + ], + "angle": 0, + "content": "4.4 Analyses" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.431, + 0.658 + ], + "angle": 0, + "content": "Ablation studies. We select three representative scenes from the ScanNet++ dataset and conduct ablation studies to evaluate the effectiveness of our method designs. We conduct ablation studies using two top-performing baselines: NeuRIS and MonoSDF." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.431, + 0.795 + ], + "angle": 0, + "content": "The results are reported in Tab. 3. Instead of using the surface normal formula to inject mutual information, we first perform a basic mutual information shaping separately in the color and SDF values. These two variations are labeled as base\\(^{+}\\) (color)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "and base\\(^{+}\\) (SDF), respectively. Interestingly, we find that, even though color is not directly related to geometry, by shaping it, we can still obtain improvements on the geometry in some cases. Moreover, shaping directly on the SDF values" + }, + { + "type": "table_caption", + "bbox": [ + 0.442, + 0.567, + 0.785, + 0.608 + ], + "angle": 0, + "content": "Table 3: Ablation studies on three scenes from ScanNet++. The best and second-best results are marked in bold and underlined." + }, + { + "type": "table", + "bbox": [ + 0.446, + 0.61, + 0.784, + 0.76 + ], + "angle": 0, + "content": "
Chamfer (m)↓F-score ↑
6ee27b649460Mean6ee27b649460Mean
NeuRIS [30]0.0290.0700.4050.1680.650.690.240.53
NeuRIS+ (color)0.0230.0380.3150.1250.720.680.340.58
NeuRIS+ (SDF)0.1270.0530.3600.1800.470.610.300.46
NeuRIS+ w/o normal0.0360.0250.3380.1330.660.750.320.58
NeuRIS+ w/o DINO0.0610.0330.3810.1580.630.720.320.56
NeuRIS+ (Full)0.0440.0330.1980.0920.660.760.410.61
MonoSDF [40]0.0200.0160.0460.0280.810.790.880.83
MonoSDF+ (color)0.0240.0190.0330.0250.790.780.870.81
MonoSDF+ (SDF)0.5880.0293.4431.3530.380.650.210.42
MonoSDF+ w/o normal0.1280.0230.0660.0720.680.730.800.74
MonoSDF+ w/o DINO0.0640.0160.0220.0340.780.830.930.85
MonoSDF+ (Full)0.0140.0200.0230.0190.850.850.930.87
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.444, + 0.129 + ], + "angle": 0, + "content": "X. Wang & S. Dong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "severely degrade the reconstruction quality, which evidences that the shaping on SDF values could incur a competition with the goal to faithfully learn the geometry within the neural radiance field. We further examine the effectiveness of combining semantic and geometric features to identify geometrically highly correlated regions. As demonstrated in Fig. 3, using only one type of the features can lead to incorrect correlations. Tab. 3 also shows that eliminating any one of the features can decrease the performance, confirming the effectiveness of our positive-negative sample selection mechanism leveraging the multimodal features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.488, + 0.478 + ], + "angle": 0, + "content": "Time and memory consumption. Our method provides benefits for 3D reconstruction with limited additional effort. Specifically, it only increases the optimization time and memory usage during the training process, without affecting the model architecture and final storage. The time and memory consumption are reported in Tab. 4. It is observed that the overhead on the training time highly depends on the architecture used in each baseline. For example, the train" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.571 + ], + "angle": 0, + "content": "ing time increase on NeuRIS is marginal while more significant on VolSDF. These studies provide comprehensive information on the trade-offs between training time and performance gain. However, we do find that baselines with the proposed shaping technique can achieve a better performance (e.g., early geometries from NeuS and \\(\\mathrm{NeuS^{+}}\\) shown in Fig. 5) at the same training steps due to the effective regularization by mutual information." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.328, + 0.784, + 0.341 + ], + "angle": 0, + "content": "Table 4: Time and memory consumption." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.342, + 0.784, + 0.445 + ], + "angle": 0, + "content": "
Training Time (h)VRAM (GB)
NeuS [30]+3.00+0.377.89+1.68
VolSDF [37]+3.18+1.9616.36+4.68
GeoNeuS [8]+3.26+0.165.72+3.56
I2-SDF [44]+3.47+2.8410.08+8.09
NeuRIS [29]+3.70+0.2011.73+4.12
MonoSDF [40]+2.65+1.3212.83+9.42
Neuralangelo [13]+3.12+2.0616.44+1.00
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.593, + 0.36, + 0.609 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.788, + 0.731 + ], + "angle": 0, + "content": "We explore the usage of the second-order geometric correlations as a regularization and introduce a mutual information shaping technique of the normals for better surface reconstruction quality. With a simple yet effective correlated point selection mechanism that does not depend on accurate geometry information, we can perform a contrastive learning that maximizes the detected correlations. In return, we can observe a boost in the geometric quality of the reconstructed surfaces in state-of-the-art neural scene representations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.731, + 0.788, + 0.837 + ], + "angle": 0, + "content": "Extensive evaluations on various baselines and datasets show the effectiveness of our method, which improves 3D reconstruction quality without resorting to a more sophisticated surface modeling. The overhead caused by the shaping of the normals varies with different neural architectures of the scene representation, e.g., some are marginal, and some are significant, which we deem as a limitation of our current pipeline. We believe that an acceleration in computation can be achieved using a coarse-to-fine strategy, which we leave as a future work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InfoNorm" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Acknowledgments. We thank the anonymous reviewers for their valuable feedback. This work is supported by the Early Career Scheme of the Research Grants Council (grant # 27207224), the HKU-100 Award, and in part by NSF China (No. 62172363). Siyan Dong would also like to thank the support from HKU Musketeers Foundation Institute of Data Science for the Postdoctoral Research Fellowship." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.263, + 0.326, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.338 + ], + "angle": 0, + "content": "1. Bae, G., Budvytis, I., Cipolla, R.: Estimating and exploiting the aleatoric uncertainty in surface normal estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13137-13146 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.339, + 0.785, + 0.393 + ], + "angle": 0, + "content": "2. Bao, C., Zhang, Y., Yang, B., Fan, T., Yang, Z., Bao, H., Zhang, G., Cui, Z.: Sine: Semantic-driven image-based nerf editing with prior-guided editing field. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20919-20929 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.395, + 0.785, + 0.435 + ], + "angle": 0, + "content": "3. Cai, B., Huang, J., Jia, R., Lv, C., Fu, H.: Neuda: Neural deformable anchor for high-fidelity implicit surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8476-8485 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.437, + 0.785, + 0.478 + ], + "angle": 0, + "content": "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.479, + 0.785, + 0.519 + ], + "angle": 0, + "content": "5. Chen, Y., Wu, Q., Zheng, C., Cham, T.J., Cai, J.: Sem2nerf: Converting single-view semantic masks to neural radiance fields. In: European Conference on Computer Vision. pp. 730-748. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.521, + 0.785, + 0.561 + ], + "angle": 0, + "content": "6. Curless, B., Levoy, M.: A volumetric method for building complex models from range images. In: Proceedings of the 23rd annual conference on Computer graphics and interactive techniques. pp. 303-312 (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.563, + 0.785, + 0.616 + ], + "angle": 0, + "content": "7. Do, T., Vuong, K., Roumeliotis, S.I., Park, H.S.: Surface normal estimation of tilted images via spatial rectifier. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 265-280. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.618, + 0.785, + 0.657 + ], + "angle": 0, + "content": "8. Fu, Q., Xu, Q., Ong, Y.S., Tao, W.: Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems 35, 3403-3416 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.785, + 0.687 + ], + "angle": 0, + "content": "9. Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.688, + 0.785, + 0.741 + ], + "angle": 0, + "content": "10. Guo, H., Peng, S., Lin, H., Wang, Q., Zhang, G., Bao, H., Zhou, X.: Neural 3d scene reconstruction with the manhattan-world assumption. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5511-5520 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.743, + 0.785, + 0.784 + ], + "angle": 0, + "content": "11. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: Lerf: Language embedded radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19729-19739 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "12. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.296, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.444, + 0.128 + ], + "angle": 0, + "content": "X. Wang & S. Dong et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "13. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.787, + 0.247 + ], + "angle": 0, + "content": "14. Liu, F., Zhang, C., Zheng, Y., Duan, Y.: Semantic ray: Learning a generalizable semantic field with cross-reprojection attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17386-17396 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.248, + 0.787, + 0.302 + ], + "angle": 0, + "content": "15. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13-23 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.787, + 0.359 + ], + "angle": 0, + "content": "16. Liu, Z., Milano, F., Frey, J., Siegwart, R., Blum, H., Cadena, C.: Unsupervised continual semantic adaptation through neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3031-3040 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.787, + 0.402 + ], + "angle": 0, + "content": "17. Long, X., Lin, C., Wang, P., Komura, T., Wang, W.: Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In: European Conference on Computer Vision. pp. 210-227. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.787, + 0.445 + ], + "angle": 0, + "content": "18. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. In: Seminal graphics: pioneering efforts that shaped the field, pp. 347-353 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.787, + 0.488 + ], + "angle": 0, + "content": "19. Meng, X., Chen, W., Yang, B.: Neat: Learning neural implicit surfaces with arbitrary topologies from multi-view images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 248-258 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.489, + 0.787, + 0.543 + ], + "angle": 0, + "content": "20. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12663-12673 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.545, + 0.787, + 0.585 + ], + "angle": 0, + "content": "21. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.787, + 0.614 + ], + "angle": 0, + "content": "22. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.616, + 0.787, + 0.671 + ], + "angle": 0, + "content": "23. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.673, + 0.787, + 0.728 + ], + "angle": 0, + "content": "24. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.787, + 0.756 + ], + "angle": 0, + "content": "25. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.841 + ], + "angle": 0, + "content": "26. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., Clarkson, A., Yan, M., Budge, B., Yan, Y., Pan, X., Yon, J., Zou, Y., Leon, K., Carter, N., Briales, J., Gillingham, T., Mueggler, E., Pesqueira, L., Savva, M., Batra, D., Strasdat, H.M., Nardi, R.D., Goesele, M., Lovegrove, S., Newcombe, R.: The replica dataset: A digital replica of indoor spaces (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InfoNorm" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "27. Tertikas, K., Despoina, P., Pan, B., Park, J.J., Uy, M.A., Emiris, I., Avrithis, Y., Guibas, L.: Partnerf: Generating part-aware editable 3d shapes without 3d supervision. arXiv preprint arXiv:2303.09554 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.233 + ], + "angle": 0, + "content": "28. Wang, F., Galliani, S., Vogel, C., Speciale, P., Pollefeys, M.: Patchmatchnet: Learned multi-view patchmatch stereo. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14194-14203 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.234, + 0.788, + 0.275 + ], + "angle": 0, + "content": "29. Wang, J., Wang, P., Long, X., Theobalt, C., Komura, T., Liu, L., Wang, W.: Neuris: Neural reconstruction of indoor scenes using normal priors. In: European Conference on Computer Vision. pp. 139-155. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.276, + 0.788, + 0.317 + ], + "angle": 0, + "content": "30. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.318, + 0.788, + 0.373 + ], + "angle": 0, + "content": "31. Weder, S., Garcia-Hernando, G., Monszpart, A., Pollefeys, M., Brostow, G.J., Firman, M., Vicente, S.: Removing objects from neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16528-16538 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.416 + ], + "angle": 0, + "content": "32. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In: European conference on computer vision. pp. 106-122. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.471 + ], + "angle": 0, + "content": "33. Xu, L., Xiangli, Y., Peng, S., Pan, X., Zhao, N., Theobalt, C., Dai, B., Lin, D.: Grid-guided neural radiance fields for large urban scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8296-8306 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.514 + ], + "angle": 0, + "content": "34. Xu, X., Yang, Y., Mo, K., Pan, B., Yi, L., Guibas, L.: Jacobinerf: Nerf shaping with mutual information gradients. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16498-16507 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.557 + ], + "angle": 0, + "content": "35. Yan, Z., Li, C., Lee, G.H.: Nerf-ds: Neural radiance fields for dynamic specular objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8285-8295 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.558, + 0.788, + 0.6 + ], + "angle": 0, + "content": "36. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multi-view stereo. In: Proceedings of the European conference on computer vision (ECCV). pp. 767-783 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.601, + 0.788, + 0.628 + ], + "angle": 0, + "content": "37. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.67 + ], + "angle": 0, + "content": "38. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: Proceedings of the International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.713 + ], + "angle": 0, + "content": "39. Yu, H., Julin, J., Milacski, Z.A., Niinuma, K., Jeni, L.A.: Dylin: Making light field networks dynamic. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12397-12406 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.714, + 0.788, + 0.755 + ], + "angle": 0, + "content": "40. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems 35, 25018-25032 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.757, + 0.788, + 0.798 + ], + "angle": 0, + "content": "41. Zhang, Y., Yang, G., Guibas, L., Yang, Y.: Infogaussian: Structure-aware dynamic gaussians through lightweight information shaping. arXiv preprint arXiv:2406.05897 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.841 + ], + "angle": 0, + "content": "42. Zhi, S., Laidlow, T., Leutenegger, S., Davison, A.J.: In-place scene labelling and understanding with implicit scene representation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15838-15847 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.444, + 0.129 + ], + "angle": 0, + "content": "X. Wang & S. Dong et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "43. Zhu, B., Yang, Y., Wang, X., Zheng, Y., Guibas, L.: Vdn-nerf: Resolving shape- radiance ambiguity via view-dependence normalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 35-45 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "44. Zhu, J., Huo, Y., Ye, Q., Luan, F., Li, J., Xi, D., Wang, L., Tang, R., Hua, W., Bao, H., Wang, R.: I\\(^2\\)-sdf: Intrinsic indoor scene reconstruction and editing via raytracing in neural sdfs. In: CVPR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.245 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/InfoNorm_ Mutual Information Shaping of Normals for Sparse-View Reconstruction/8f9c186f-c37a-4b28-8f4e-c701ceacc591_origin.pdf b/2024/InfoNorm_ Mutual Information Shaping of Normals for Sparse-View Reconstruction/8f9c186f-c37a-4b28-8f4e-c701ceacc591_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..acc20d264fc4d65c7275f231d77eefbb0221b9a3 --- /dev/null +++ b/2024/InfoNorm_ Mutual Information Shaping of Normals for Sparse-View Reconstruction/8f9c186f-c37a-4b28-8f4e-c701ceacc591_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c82ef6b7ec49359cb30acd6d0a165c43788f0332d7c99cda187acd92dad30b7 +size 5988635 diff --git a/2024/InfoNorm_ Mutual Information Shaping of Normals for Sparse-View Reconstruction/full.md b/2024/InfoNorm_ Mutual Information Shaping of Normals for Sparse-View Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..18ab1fdf7de7f1b30d80b9c24f32e8cf303991aa --- /dev/null +++ b/2024/InfoNorm_ Mutual Information Shaping of Normals for Sparse-View Reconstruction/full.md @@ -0,0 +1,334 @@ +# InfoNorm: Mutual Information Shaping of Normals for Sparse-View Reconstruction + +Xulong Wang $^{1,2*}$ Siyan Dong $^{3*}$ Youyi Zheng $^{1\dagger}$ Yanchao Yang $^{3,4\dagger}$ + +1 State Key Lab of CAD&CG, Zhejiang University 2 Chohotech co. Ltd. + +3 Institute of Data Science, The University of Hong Kong + +$^{4}$ Department of Electrical and Electronic Engineering, The University of Hong Kong + +Abstract. 3D surface reconstruction from multi-view images is essential for scene understanding and interaction. However, complex indoor scenes pose challenges such as ambiguity due to limited observations. Recent implicit surface representations, such as Neural Radiance Fields (NeRFs) and signed distance functions (SDFs), employ various geometric priors to resolve the lack of observed information. Nevertheless, their performance heavily depends on the quality of the pre-trained geometry estimation models. To ease such dependence, we propose regularizing the geometric modeling by explicitly encouraging the mutual information among surface normals of highly correlated scene points. In this way, the geometry learning process is modulated by the second-order correlations from noisy (first-order) geometric priors, thus eliminating the bias due to poor generalization. Additionally, we introduce a simple yet effective scheme that utilizes semantic and geometric features to identify correlated points, enhancing their mutual information accordingly. The proposed technique can serve as a plugin for SDF-based neural surface representations. Our experiments demonstrate the effectiveness of the proposed in improving the surface reconstruction quality of major states of the arts. Our code is available at: https://github.com/Muliphein/InfoNorm. + +Keywords: 3D from multi-view images $\cdot$ surface reconstruction + +# 1 Introduction + +3D surface reconstruction from multi-view images is an essential task in the computer vision and graphics community, with practical applications including content creation for virtual reality and robot-scene interaction. However, indoor scenes present challenges due to the large and complex scenes captured from sparse viewpoints. These challenges encompass issues such as occlusion and ambiguity arising from limited observations. As a result, traditional Multi-View Stereo (MVS) methods [28,36], which require substantial overlap among images, may not produce satisfactory results in these scenarios. + +![](images/9b8401e3fa530d78df7d772cbc467ce6c418352d13234aeaf0b6384a2139842b.jpg) +VoISDF + +![](images/806ffe2576270f50b515b669174c1942b2de766ad6891de348a0e98e77682aa9.jpg) +VoISDF+ (Ours) + +![](images/4fa19faf7ef2ee3e5792b72952735bf2af6b7b5b9c5badae8ff9e895d4e9b334.jpg) +Ground truth + +![](images/5a2d9d718867f7ac2f1ce0a97aca24769f738ae92246bee6311fadca4d578343.jpg) +GeoNeuS +Fig. 1: 3D scene reconstruction from sparse views on Replica [26] (first row) and ScanNet++ [38] (second row). By enforcing the mutual information between the normals of highly correlated scene points, the proposed method can effectively enhance the reconstruction quality of the baselines (VolSDF [37] and GeoNeuS [8]). + +![](images/ca123373e0b175854094b84eb9138c7da3f4b7698a126c30ac38be6287709a8c.jpg) +GeoNeuS $^{+}$ (Ours) + +![](images/163c73317edb184b56866c497313fa562e0cfff80ed358069df6bcaa7a5f696a.jpg) +Ground truth + +Recently, implicit scene representations, e.g., Neural Radiance Fields (NeRFs) [21], have been proposed to encode 3D scenes into a set of neural network parameters, with subsequent works enhancing the representational capacity for geometry to enable surface extraction. The key idea involves incorporating signed distance functions (SDFs) [23] into the learned radiance field. Although NeRFs with SDF produce high-quality reconstructions of simple scenes with sufficient images, they still underperform in large and complex scenes captured from sparse viewpoints. More recent approaches, such as NeuRIS [29] and MonoSDF [40], propose using monocular geometric priors to guide the learning of the radiance field. However, they can be affected by inevitable errors and noise in the monocular geometry estimation modules. + +We explore geometric regularization in a second-order perspective. We argue that both the geometry itself and the correlation of geometry among scene regions are crucial for surface reconstruction, especially when dealing with varying indoor scenes under sparse views with noisy geometry estimates. The correlation is expressed as mutual information between scene entities under random perturbations of the network weights. More explicitly, we propose to enforce the mutual information between the normals of two points in the scene that are considered geometrically correlated, where the mutual information can be efficiently computed by measuring the cosine similarity between the gradients of the normals with respect to the perturbed weights. Further, we leverage a combination of pretrained semantic and geometric features [4,7,12] to identify the correlated scene regions that endorse high mutual information. This results in an easy-to-use geometric shaping1 technique that can be applied to any SDF-based neural radiance fields. To thoroughly evaluate its effectiveness, we test the proposed with + +multiple state-of-the-art neural surface reconstruction models. The experimental results confirm the effectiveness of the proposed shaping technique, showing improved surface reconstruction quality (e.g., Fig. 1) on challenging scenes with limited views. To summarize: + +- We propose to leverage second-order geometric correlation in the form of mutual information between normals to regularize the surface reconstruction of SDF-based NeRF representations. +- We develop a pipeline that efficiently enforces the mutual information of the normals of points, deemed geometrically correlated with pre-trained multimodal features, to improve the reconstruction quality of the scene. +- We verify the effectiveness of the proposed second-order regularization technique across a broad spectrum of baselines and demonstrate its usefulness as an easy-to-use plugin in improving 3D surface modeling. + +# 2 Related Work + +3D Reconstruction with neural radiance fields (NeRFs). NeRF [21] is originally designed for novel view synthesis, with its variants serving as scene representations [15, 32, 33, 35, 39] for tasks related to editing [2, 31, 44], semantic segmentation [5, 14, 16, 34], 3D shape generation [20, 27], and so on. We mainly focus on neural surface representations that enable the extraction of high-quality 3D meshes from the trained NeRFs. + +Early works such as NeuS [30] and VolSDF [37] combine signed distance functions (SDFs) with volume rendering to extract high-quality surfaces. Recently, SparseNeuS [17] introduces a cascaded geometry reasoning framework that generalizes well to novel scenes. GeoNeuS [8] uses surface points from structure-from-motion and neighbor-view patches to reconstruct surfaces, while NeuDA [3] employs a deformable anchor to adaptively encode geometric details. NeAT [19] broadens the scope of reconstruction to arbitrary topology, no longer confining to watertight surfaces. Moreover, Neuralangelo [13] uses a hash grid for position encoding and implements a coarse-to-fine optimization with numerical gradients. Despite their promising results, most of these methods need sufficient training views, and their performance declines when dealing with large and complex scenes under sparse viewpoints. + +Incorporating priors into NeRFs. One can augment scene representations by introducing various priors to NeRFs. VDN-NeRF [43] normalizes the spatial feature to align with monocular features. LERF [11] introduces a new branch whose outputs align with DINO [4] and CLIP [24] features. Semantic-NeRF [42] encodes semantic information for segmentation tasks, while JacobiNeRF [34] shapes the color gradients to encode the mutual information in terms of semantic similarity from DINO [4] features for label propagation. + +For surface reconstruction, GeoNeuS [8] incorporates patch similarity from neighboring images and sparse surface points derived from structure-from-motion + +[25], while I²-SDF [44] combines the rendering equation with VolSDF [37] for better novel views and geometry. To cope with complex indoor scenes, NeuRIS [29] incorporates the normal priors as additional geometric supervision. Further, MonoSDF [40] uses both depth and normal to improve the surface reconstruction quality. However, these methods highly depend on the quality of the supplemented geometric information. In contrast, our method leverages the second-order information in geometric prior, rather than directly using the prior as a first-order supervision signal, which makes it more robust to errors and noise. There are also methods utilizing strong assumptions derived for human-made scenes. For instance, ManhattanSDF [10] works under the assumption that all captured scenes adhere to the Manhattan-world concept, thus, predicting semantic segmentation to identify walls and floors. While these methods primarily focus on planar scenes, the proposed technique can work for general scenes under noisy correlation information. + +# 3 Method + +We start with a discussion on the preliminaries in Sec. 3.1 to provide the background on how we can learn from posed images an implicit scene representation (NeRF), improve 3D scene reconstruction via the combination of NeRF and SDF representations, and employ mutual information as a constraint in the NeRF training process. Following this, in Sec. 3.2, we delve into how mutual information can be used as constraints on NeRF's implied surfaces, so that we can learn better geometries for higher-quality 3D reconstruction of the scene. Finally, we elaborate on how we can employ off-the-shelf semantic and geometry features to assist the encoding of mutual information through the proposed geometric shaping framework. + +# 3.1 Preliminaries + +Neural radiance fields. NeRF [21] learns an implicit scene representation from a set of posed images. For a specific pixel in an image, NeRF samples a list of 3D points, $\{x|x_{t} = \mathbf{o} + t\mathbf{v},t\in [t_{near},t_{far}]\}$ , from the camera center $\mathbf{o}$ along the viewing direction $\mathbf{v}$ within a bounded space. Specifically, it employs a Multilayer Perceptron (MLP) to encode the volume density and color of each sampled point. We denote all the parameters of NeRF as $\theta$ , the density function as $\sigma$ , and the color function as $c$ . The pixel's rendered color $\hat{\mathbf{C}}$ from a NeRF is calculated by the discrete volume rendering: + +$$ +\hat {\mathbf {C}} (\mathbf {o}, \mathbf {v}; \theta) = \sum_ {i = 1} ^ {N} \omega (\mathbf {o}, \mathbf {v}, t _ {i}; \theta) c \left(x _ {t _ {i}}, \mathbf {v}; \theta\right), \tag {1} +$$ + +$$ +\text {w h e r e} \omega (\mathbf {o}, \mathbf {v}, t _ {i}; \theta) = \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}) \alpha_ {i}. +$$ + +Here, $\alpha_{i} = 1 - \exp (-\sigma_{i}\delta_{i})$ , where $\sigma_{i}$ is the abbreviation of $\sigma (x_{t_i};\theta)$ and $\delta_{i} = t_{i + 1} - t_{i}$ . A photometric loss is utilized to minimize the discrepancy between the rendered color and its ground truth: + +$$ +L _ {C} = \left\| \mathbf {C} - \hat {\mathbf {C}} \right\| _ {1}, \tag {2} +$$ + +so that the parameters $\theta$ can be optimized to fit the scene. Despite the ability to handle photorealistic novel view synthesis, NeRFs suffer from noisy and unrealistic surface extraction when it comes to 3D reconstruction, as discussed in the literature [30, 37]. + +NeRFs with signed distance functions. To produce high-quality surfaces, it is beneficial to use signed distance functions (SDFs) [6, 23] to represent geometry. Consequently, the surface can be reconstructed by extracting the zero-level set [18] of the SDF values. The combination of NeRF and SDF can be achieved by deriving the NeRF's densities from the SDF, or defining Laplace's cumulative distribution function. Accordingly, $\alpha_{i}$ in Eq. 1 can be represented by a function $\varPhi$ which converts SDF values to volume densities as + +$$ +\alpha_ {i} = \Phi (x, f (x; \theta), i). \tag {3} +$$ + +Generally, NeRFs with an SDF achieve improved surface reconstruction quality, where the key to maintaining the characteristics of SDF is an eikonal loss [9]: + +$$ +L _ {E} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left(\left| \left| \nabla f \left(x _ {i}\right) \right| \right| _ {2} - 1\right) ^ {2}. \tag {4} +$$ + +Mutual information in NeRFs. Mutual information (MI) quantifies the statistical dependence between two random variables, providing an estimate of how much information they share. It has been introduced for semantic segmentation and motion modeling tasks [34, 41]. In this paper, We study the mutual information under perturbations of the NeRF parameters, and focus more on the geometry properties. Let $I(p_{i})$ and $I(p_{j})$ denote two pixels derived from the NeRF rendering process $\mathbb{F}$ . And we assume that a subset of the NeRF weights, denoted by $\theta^{D}$ , is perturbed by a random noise vector $\mathbf{n} \in \mathbb{R}^{D}$ sampled from a uniform distribution on the sphere $\mathbb{S}^{D - 1}$ by a small step $\gamma \ll 1.0$ . The random variables corresponding to the perturbed pixels can be written through a Taylor expansion: + +$$ +\hat {I} (p _ {i}) = I (p _ {i}) + \gamma \mathbf {n} \frac {\partial \mathbb {F} _ {i}}{\partial \theta^ {D}}, +$$ + +$$ +\hat {I} (p _ {j}) = I (p _ {j}) + \gamma \mathbf {n} \frac {\partial \mathbb {F} _ {j}}{\partial \theta^ {D}}. +$$ + +We can now characterize the mutual information between $\hat{I}(p_i)$ and $\hat{I}(p_j)$ under the joint probability distribution $\mathbb{P}(\hat{I}(p_i), \hat{I}(p_j))$ . As derived in JacobiNeRF, the + +mutual information can be computed as: + +$$ +\begin{array}{l} \mathbb {I} (\hat {I} (p _ {i}), \hat {I} (p _ {j})) = \mathbb {H} (\hat {I} (p _ {i})) - \mathbb {H} (\hat {I} (p _ {i}) | \hat {I} (p _ {j})) \\ = \log (\frac {1}{\sqrt {1 - \cos^ {2} \beta}}) + \mathrm {c o n s t .}. \\ \end{array} +$$ + +Here, $\beta$ denotes the angle between $\partial \mathbb{F}_i / \partial \theta^D$ and $\partial \mathbb{F}_j / \partial \theta^D$ . It is observed that: + +$$ +\mathbb {I} (\hat {I} (p _ {i}), \hat {I} (p _ {j})) \propto | | \cos \beta | |. \tag {5} +$$ + +Therefore, to enhance the MI between two random pixels induced by NeRF perturbations, one can enforce the absolute value of the cosine similarity. That is, if two pixels $p_i$ , $p_j$ come with high correlation (e.g., from the same object in the scene), their gradients regarding the perturbed parameters should be in the same or opposite direction. + +# 3.2 Mutual Information Shaping of Geometry in SDF-based NeRFs + +While the previous approach focuses on enforcing mutual information for image synthesis, we explore a mutual information shaping of the geometries for quality surface reconstruction in SDF-based NeRFs. + +Specifically, we represent a scene via a NeRF backbone with an SDF head and a color head, as shown in Fig. 2. Note that the method to be proposed is capable of being applied to any network similar to NeRF and SDF representations. Given that the two heads partially share parameters, we denote all the parameters as $\theta$ , and use $f(x;\theta)$ and $c(x,\mathbf{v};\theta)$ to represent the SDF and color branches, respectively. To achieve mutual information shaping of the geometry, we focus on manipulating the parameters of $f(x;\theta)$ , while keeping $c(x,\mathbf{v};\theta)$ unchanged. Below, we first explain why and how we compute mutual information with surface normals derived from SDF, i.e., $f(x;\theta)$ , as a geometry-aware constraint. Next, we present a simple yet effective strategy that extracts regions with high correlation, which need to be encoded into the weights using the mutual information shaping technique. Finally, we elaborate on our loss function and training process to reconstruct the scene from posed images. + +Mutual information among surface normals. We explicitly encode mutual information into the function $f(x; \theta)$ . However, instead of computing mutual information between the density values, we proceed by manipulating the surface normal vectors that can be derived from $f$ . Specifically, for a 3D point $x$ on a surface of the scene, its normal vector $N(x; \theta)$ can be calculated as the partial derivative of the 3D coordinate: + +$$ +N (x; \theta) = \frac {\partial f (x ; \theta)}{\partial x}. \tag {6} +$$ + +There are several reasons to manipulate surface normal vectors but not SDF values: 1) Shaping the SDF values with respect to a subset $\theta^D$ of the parameters may not effectively impose the constraints on all parameters. However, the + +![](images/6f17f63afda6c3119d5acae9d04874923ea056d465e48b6553b8bf4685feb5f3.jpg) +Fig. 2: An overview of the pipeline, where we apply mutual information shaping on the geometric branch to enforce consistencies that help enhance the surface reconstruction. Specifically, the model consists of the NeRF backbone, an SDF head, and a color head. The predicted density and color are supervised by the classic eikonal loss $L_{E}$ and the photometric reconstruction loss $L_{C}$ , respectively. We encode geometry-aware mutual information into a subset of the parameters $\theta^{D}$ of $f(x)$ to constrain the density field learning for better surface quality, which is achieved by the proposed mutual information loss $L_{M}$ computed on top of the estimated surface normal. + +calculation of the surface normal includes the entire set of parameters in $f$ , thus a more global regularity. 2) Moreover, shaping $f$ with the mutual information between its values may induce side effects as density could be distorted to satisfy the constraints. However, shaping the normal $\partial f / \partial x$ leave more space for the SDF values to cope with the scene reconstruction task, thus minimizing potential negative effects. + +According to Eq. 5, we need to calculate the partial gradients $\partial N / \partial \theta^D$ to compute the mutual information. When it comes to a vector rather than a scalar, the gradients consist of three components corresponding to the $\mathbf{x}$ , $\mathbf{y}$ , and $\mathbf{z}$ axes: + +$$ +\frac {\partial N (x ; \theta)}{\partial \theta^ {D}} = \frac {\partial N _ {\mathbf {x}} (x ; \theta)}{\partial \theta^ {D}}, \frac {\partial N _ {\mathbf {y}} (x ; \theta)}{\partial \theta^ {D}}, \frac {\partial N _ {\mathbf {z}} (x ; \theta)}{\partial \theta^ {D}}. \tag {7} +$$ + +For ease of computation, we concatenate the gradients into a vector, denoted as $\partial N$ for simplicity: + +$$ +\partial N = \operatorname {C o n c a t} \left(\frac {\partial N _ {\mathbf {x}}}{\partial \theta^ {D}}, \frac {\partial N _ {\mathbf {y}}}{\partial \theta^ {D}}, \frac {\partial N _ {\mathbf {z}}}{\partial \theta^ {D}}\right). \tag {8} +$$ + +Since a 2D pixel is accumulated from 3D, we aggregate the normal-based information along a viewing ray as in Eq. 1 as: + +$$ +\mathbb {N} \left(p _ {i}\right) = \sum_ {x \in \text {r a y}} \omega \left(t _ {i}; \theta\right) \partial N. \tag {9} +$$ + +Finally, we obtain that the mutual information between two normals is: + +$$ +\mathbb {I} (\hat {N} (p _ {i}), \hat {N} (p _ {j})) \propto | | \frac {\mathbb {N} (p _ {i}) \cdot \mathbb {N} (p _ {j})}{| \mathbb {N} (p _ {i}) | | \mathbb {N} (p _ {j}) |} | |. \tag {10} +$$ + +Please refer to our supplementary material for a detailed derivation. From Eq. 10, we observe that the mutual information in surface normal vectors correlates + +![](images/a1037c099eeadeb4e12569c2335a9d39eec82a6e6ce806b409303ba67e34c5c1.jpg) +Anchor pixel + +![](images/fdc6e8d50ef3f3f196ffc7ad2e20073c5eefb613a549457884295d170dcec5a5.jpg) +Similar semantic features +Fig. 3: An example of positive samples from different features. Given an anchor pixel (marked by a blue circle in the first image) on the wall-like cabinet, semantic features like DINO often correlate both the cabinet and the ceiling (as shown by the red area in the second image). Meanwhile, geometric features such as (noisy) monocular normals can not distinguish between parallel planes that are not connected (the third image). By combining semantic and geometric features, we can obtain positive samples of the anchor pixel with better geometric consistency, thus, high mutual correlation. + +![](images/0713b265af24b7ebbaeed91bc4e2fbefb2b1432c4c0011904bb28b4c17d96bbc.jpg) +Similar geometric features + +![](images/c5e4fb51e95f325dfde32be6d991727a9691233107efa75288aab21b7ee2d042.jpg) +Combined (Ours) + +positively with the cosine similarity of their gradients to the network parameters. This conclusion is similar to that of Eq. 5. As a result, we can now identify the pixels that are of high correlations, and increase the aggregated mutual information during the NeRF training process. + +Extracting correlated scene regions from image pixels. To extract scene regions that need to be shaped to have high mutual information, we present a simple yet effective approach with off-the-shelf visual features. Given the input images, we extract pixel-wise semantic and geometric features to identify correlated scene regions. Following previous research [34], we apply DINO [4] as the semantic feature extractor to obtain a high-level understanding of the scene regions. Meanwhile, as we shape mutual information on top of the surface normal, we employ a monocular normal estimator [7] as the geometric feature extractor for a detailed low-level perception of the 3D points in the scene. + +DINO features can group different parts of an object, like the seat and back of a chair, which share similar semantics. However, the surface normal vectors from these parts have limited correlation, thus should not be constrained. Conversely, using only geometric features can wrongly suggest high correlation between unrelated areas, like the seat and the floor. As these areas aren't closely connected in 3D space, there is no reason to have better surface reconstruction when enhancing their mutual information. Similar phenomena on wall-like structures can be observed in Fig. 3, where we show that by combining semantic and geometric features together we can better identify scene regions that are spatially connected and have similar normal directions. + +More explicitly, we combine semantic and geometric features with an intersection operator. It is simple yet effective. The process of identifying correlated scene regions (in normal) is implemented as positive pixel pair selection from input images. Given a randomly selected pixel from an image, we first extract similar pixels from neighboring images using semantic and geometric features, respectively. This extraction is implemented by applying two thresholds, $\beta_{S}$ and $\beta_{G}$ , to filter the cosine similarity for semantic and geometric features. We then + +feed the intersection of the top similar pixels from each feature type into NeRF's training process as positive samples, allowing us to shape them with high mutual information. In order to leverage the contrastive learning framework to encourage mutual information, we randomly select pixels not involved in the aforementioned process to serve as negative samples (independent pixels that are not correlated to the anchor pixel). + +Loss functions for training. The aforementioned contrastive learning process is performed with a tailored InfoNCE loss [22]: + +$$ +L _ {M} = - \log \frac {\sum \exp (| | \cos (\partial \mathbb {N} _ {i} , \partial \mathbb {N} _ {i +}) | |)}{\sum_ {j} \exp (| | \cos (\partial \mathbb {N} _ {i} , \partial \mathbb {N} _ {j}) | |)}, \tag {11} +$$ + +where $i$ , $i+$ , and $j$ denote the anchor, positive, and negative pixels, respectively. This loss function effectively enhances mutual information (according to Eq. 10) among correlated pixels, while maintaining low mutual information among independent pixels. Finally, to train the NeRF backbone and retain the characteristics of the SDF, we adopt the basic photometric loss (Eq. 2) and the eikonal loss (Eq. 4). The latter two are derived from single images, while the former can be calculated from either single images or a set of neighboring images. The full loss function is defined as the weighted sum of three loss terms: + +$$ +L = L _ {C} + \lambda_ {E} L _ {E} + \lambda_ {M} L _ {M}. \tag {12} +$$ + +With the above, all network parameters $\theta$ can be trained end-to-end. + +# 4 Experiments + +In this section, we evaluate the effectiveness of our method through experiments conducted on multiple datasets with various baselines. First, we introduce the two datasets used to conduct the experiments in Sec. 4.1. Next, we detail the implementation of the proposed method in Sec. 4.2, where we adapt a spectrum of state-of-the-art models for validating the effectiveness of the geometric shaping. Then, we report the comparisons between the baselines and our models in Sec. 4.3. Lastly, we conduct a series of ablation studies in Sec. 4.4, to verify the necessity of each component in our method. + +# 4.1 Datasets + +We evaluate the proposed method using two public datasets from the literature: ScanNet++ [38] and the Replica [26] dataset. For the ScanNet++ dataset, we choose 8 representative scenes, while for the Replica dataset, we select 4 representative scenes. The scenes from both datasets represent a variety of indoor settings, including offices, bedrooms, and so on. As our focus is on scene reconstruction from sparse views, we uniformly sample 50-85 images from the official (and Replica from Semantic-NeRF [42]) scanning sequences for each scene to use as input in our evaluation. For detailed statistics of each scene, please refer to our supplementary material. + +Evaluation metrics. Following previous works [13, 30, 40], we evaluate the quality of surface reconstruction using two metrics: Chamfer distance and F-score. Chamfer distance is a measure that evaluates the difference between reconstructed and ground-truth points sets. F-score is a statistical metric that integrates precision and recall, providing a more comprehensive understanding of the reconstruction quality. + +# 4.2 Implementation Details + +As the proposed geometric shaping can be easily integrated into any network similar to NeRF with SDF representations, we select several state-of-the-art models as our baselines and apply the proposed to them for demonstrating the effectiveness of our approach. We first choose NeuS [30] and VolSDF [37], the two representative pioneers for NeRFs with SDFs. We also apply our method to the following works: GeoNeuS [8], I²-SDF [44], NeuRIS [29], MonoSDF [40], and Neuralangelo [13]. In addition to the basic loss functions for NeRF and SDF, GeoNeuS incorporates patch similarity from neighboring views and surface points derived from structure-from-motion. I²-SDF integrates the rendering equation with VolSDF. We disregard its normal and depth loss terms. NeuRIS utilizes monocular normal estimation as a supervision on the geometry, and discards the low-quality normal information that lacks sufficient confidence through multi-view patching. MonoSDF employs both depth and normal estimation, for which we set the decay of the normal and depth at 30k iterations to prevent the method from degeneration. Neuralangelo utilizes a multi-resolution hash grid and numerical gradient. We set the hash encoding dictionary size to 20, and the encoding dimension to 4. For all the base models mentioned above, we enhance them by incorporating our proposed mutual information loss during training. The resulting methods are respectively labeled as $\mathrm{NeuS^{+}}$ , VolSDF+, GeoNeuS+, and so on. + +# 4.3 Comparisons + +Results on the ScanNet++ dataset. The quantitative comparisons are reported in Tab. 1. We also visualize the surface reconstruction results in Fig. 4 for qualitative comparisons. We can observe that by applying the proposed mutual information shaping on the geometry, the surface reconstruction quality can be improved across all baselines, as measured by mean Chamfer distance, F-score, and the visuals. + +NeuS and VolSDF tend to produce rough surfaces and excessively smoothed boundaries. However, with our geometry-aware mutual information, the surfaces become smoother, while the boundaries are more crispy. Compared to GeoNeuS, a method that uses patch-similarity among neighboring views and requires sufficient textures, our method provides significant improvements. Applying the proposed technique to $\mathrm{I}^2$ -SDF can also enhance its reconstruction quality. + +As for more recent works, we find that NeuRIS is somewhat sensitive to the quality of normal estimation. It can either be affected by noise in the normal + +Table 1: Quantitative results on the ScanNet++ dataset. Each cell includes the original baseline number in black, followed by the improvement using our mutual information. Positive and negative improvements are marked in green and orange, respectively. The last column shows the average improvement over all scenes, validating the effectiveness of the proposed shaping for better geometric reconstruction (despite a small potion of noise on a few scenes). + +
0a7c0a186ee27b6456a09460a08de0abMean
NeuS [30]+0.08-0.040.05-0.030.03+0.010.06-0.040.01-0.000.24-0.140.04-0.030.02-0.000.07-0.03
VolSDF [37]+0.90-0.830.31-0.290.04+0.020.10-0.030.03-0.010.14+0.030.39-0.380.04-0.000.24-0.19
GeoNeuS [8]+0.03-0.021.03-1.020.28-0.270.05-0.030.02-0.010.22-0.200.01-0.010.26-0.250.24-0.23
I2-SDF [44]+0.27-0.240.22-0.030.04-0.010.47-0.450.03-0.020.13+0.050.03+0.020.30-0.070.19-0.09
NeuRIS [29]+0.01-0.000.01-0.000.03+0.020.07-0.040.04-0.030.41-0.210.03-0.010.03-0.000.08-0.03
MonoSDF [40]+0.01-0.000.01+0.020.02-0.010.02-0.000.01-0.000.05-0.020.01-0.000.01-0.000.02-0.00
Neuralangelo [13]+0.17-0.010.84-0.660.52-0.330.05-0.010.04+0.024.76-2.850.33+0.050.12+0.030.85-0.48
NeuS [30]+0.49+0.060.48+0.100.67+0.000.58+0.160.81-0.010.58-0.020.67+0.170.75+0.090.63+0.07
VolSDF [37]+0.45+0.350.47+0.270.67+0.000.55+0.050.71+0.060.52+0.180.48+0.350.70+0.020.57+0.16
GeoNeuS [8]+0.69+0.200.47+0.440.52+0.340.70+0.060.87+0.030.68+0.230.75+0.140.37+0.510.63+0.24
I2-SDF [44]+0.33+0.520.28+0.000.58+0.150.41+0.430.75+0.100.65+0.070.65+0.070.40+0.250.51+0.20
NeuRIS [29]+0.72+0.030.68+0.010.65+0.010.69+0.060.85+0.010.24+0.170.64+0.130.69+0.090.65+0.06
MonoSDF [40]+0.92-0.010.88-0.010.81+0.040.79+0.060.89+0.000.88+0.050.91+0.000.89-0.010.87+0.01
Neuralangelo [13]+0.42+0.050.14+0.300.35+0.120.63+0.050.71-0.010.17+0.220.29+0.060.51-0.010.40+0.10
+ +Table 2: Quantitative results on the Replica dataset. + +
office0office1room0room1Mean
NeuS [30]+0.005+0.0010.023-0.0120.068-0.0380.130-0.0870.057-0.035
VolSDF [37]+0.004+0.0060.016-0.0040.080-0.0710.034-0.0250.033-0.023
GeoNeuS [8]+0.023-0.0210.014-0.0090.040-0.0350.002-0.0000.020-0.017
I2-SDF [44]+0.032-0.0090.056-0.033--0.044-0.021
NeuRIS [29]+0.014-0.0063.975-3.9630.025-0.0170.015+0.0021.007-0.996
MonoSDF [40]+0.003-0.0000.005-0.0010.004-0.0000.004-0.0010.004-0.001
Neuralangelo [13]+0.024-0.0180.030-0.0190.146-0.0120.082-0.0470.070-0.024
NeuS [30]+0.92+0.000.80+0.010.78-0.030.72+0.080.80+0.02
VolSDF [37]+0.94-0.040.78+0.010.78+0.160.82+0.070.83+0.05
GeoNeuS [8]+0.89+0.080.91+0.040.90+0.090.98+0.010.920+0.05
I2-SDF [44]+0.72+0.090.56+0.20--0.64+0.14
NeuRIS [29]+0.79+0.060.01+0.680.89+0.050.77+0.030.61+0.20
MonoSDF [40]+0.96+0.000.90+0.020.98+0.000.95+0.020.95+0.01
Neuralangelo [13]+0.89+0.070.49+0.360.65+0.150.71+0.180.69+0.19
+ +or completely disregard the normal information due to low confidence. Particularly, when images are sparse, the regions covered by fewer viewpoints tend to lose supervision. Our mutual information scheme does not depend on the absolute values of the surface normal, but on the similarity among these values. This makes it more robust to noise and less likely to discard low-quality but useful normal information. Consequently, our method enhances the reconstruction quality of NeuRIS. MonoSDF achieves the best overall performance among the baselines because it uses both depth and normal estimations as geometry supervision. However, our method still offers improvements. Notably, it produces smoother planes and clearer boundaries. This indicates that not only the supervision of geometric cues is important, but also the correlation between the geometric entities matters. + +![](images/85fcbe70ebb8b419ecd0f5736e2a72107774160c5cf350c90b8e8223928071d9.jpg) +Fig. 4: Visual results on the ScanNet++ dataset. Each row shows a comparison with a different baseline, from top to bottom, NeuS $[30]^+$ , VolSDF $[37]^+$ , GeoNeuS $[8]^+$ , I²-SDF $[44]^+$ , NeuRIS $[29]^+$ , MonoSDF $[40]^+$ , and Neuralangelo $[13]^+$ . Red boxes are overlaid to help the contrast. + +![](images/c3b66e13e0a8712a4ccb6d1eeba48ff3a6730ef1a34898ee902321c2a9046255.jpg) +Reference image + +![](images/12b7055978d076432f27e37e27bcf5105d4b8a6caddf0c26c3ce3c4e5a6b914d.jpg) +NeuS +Fig. 5: Reconstruction results after 10K iterations in the training process with $(\mathrm{NeuS}^{+})$ and without (NeuS) the proposed geometric shaping technique. + +![](images/1934989b75e764148aeab0fdca00ee343919dd3f6fea03c82d22737ae5d6ba6b.jpg) +Neu $^{+}$ (Ours) + +![](images/3d686b691ed9650ea1c7a36938a87db1535ca26e1281f93b6ada1fecee39a744.jpg) +Ground truth + +Last but not least, the experiments on Neuralangelo demonstrate the potential of our method in working with hash grid encoding and numerical gradient. This highlights our method's versatility and ease of use, which can be broadly applied in various reconstruction architectures. + +Results on the Replica dataset. In Tab. 2 we report the numerical results on the Replica dataset, with respect to NeuS [30], VolSDF [37], GeoNeuS [8], I²-SDF [44], NeuRIS [29], MonoSDF [40], and Neuralangelo [13]. As observed, the proposed geometric shaping technique still provides positive improvements when incorporated into these baselines, aligning with the conclusion drawn from the ScanNet++ data. Qualitative examples can also be found in Fig. 1. + +# 4.4 Analyses + +Ablation studies. We select three representative scenes from the ScanNet++ dataset and conduct ablation studies to evaluate the effectiveness of our method designs. We conduct ablation studies using two top-performing baselines: NeuRIS and MonoSDF. + +The results are reported in Tab. 3. Instead of using the surface normal formula to inject mutual information, we first perform a basic mutual information shaping separately in the color and SDF values. These two variations are labeled as base $^{+}$ (color) + +and base $^{+}$ (SDF), respectively. Interestingly, we find that, even though color is not directly related to geometry, by shaping it, we can still obtain improvements on the geometry in some cases. Moreover, shaping directly on the SDF values + +Table 3: Ablation studies on three scenes from ScanNet++. The best and second-best results are marked in bold and underlined. + +
Chamfer (m)↓F-score ↑
6ee27b649460Mean6ee27b649460Mean
NeuRIS [30]0.0290.0700.4050.1680.650.690.240.53
NeuRIS+ (color)0.0230.0380.3150.1250.720.680.340.58
NeuRIS+ (SDF)0.1270.0530.3600.1800.470.610.300.46
NeuRIS+ w/o normal0.0360.0250.3380.1330.660.750.320.58
NeuRIS+ w/o DINO0.0610.0330.3810.1580.630.720.320.56
NeuRIS+ (Full)0.0440.0330.1980.0920.660.760.410.61
MonoSDF [40]0.0200.0160.0460.0280.810.790.880.83
MonoSDF+ (color)0.0240.0190.0330.0250.790.780.870.81
MonoSDF+ (SDF)0.5880.0293.4431.3530.380.650.210.42
MonoSDF+ w/o normal0.1280.0230.0660.0720.680.730.800.74
MonoSDF+ w/o DINO0.0640.0160.0220.0340.780.830.930.85
MonoSDF+ (Full)0.0140.0200.0230.0190.850.850.930.87
+ +severely degrade the reconstruction quality, which evidences that the shaping on SDF values could incur a competition with the goal to faithfully learn the geometry within the neural radiance field. We further examine the effectiveness of combining semantic and geometric features to identify geometrically highly correlated regions. As demonstrated in Fig. 3, using only one type of the features can lead to incorrect correlations. Tab. 3 also shows that eliminating any one of the features can decrease the performance, confirming the effectiveness of our positive-negative sample selection mechanism leveraging the multimodal features. + +Time and memory consumption. Our method provides benefits for 3D reconstruction with limited additional effort. Specifically, it only increases the optimization time and memory usage during the training process, without affecting the model architecture and final storage. The time and memory consumption are reported in Tab. 4. It is observed that the overhead on the training time highly depends on the architecture used in each baseline. For example, the train + +ing time increase on NeuRIS is marginal while more significant on VolSDF. These studies provide comprehensive information on the trade-offs between training time and performance gain. However, we do find that baselines with the proposed shaping technique can achieve a better performance (e.g., early geometries from NeuS and $\mathrm{NeuS^{+}}$ shown in Fig. 5) at the same training steps due to the effective regularization by mutual information. + +Table 4: Time and memory consumption. + +
Training Time (h)VRAM (GB)
NeuS [30]+3.00+0.377.89+1.68
VolSDF [37]+3.18+1.9616.36+4.68
GeoNeuS [8]+3.26+0.165.72+3.56
I2-SDF [44]+3.47+2.8410.08+8.09
NeuRIS [29]+3.70+0.2011.73+4.12
MonoSDF [40]+2.65+1.3212.83+9.42
Neuralangelo [13]+3.12+2.0616.44+1.00
+ +# 5 Conclusion + +We explore the usage of the second-order geometric correlations as a regularization and introduce a mutual information shaping technique of the normals for better surface reconstruction quality. With a simple yet effective correlated point selection mechanism that does not depend on accurate geometry information, we can perform a contrastive learning that maximizes the detected correlations. In return, we can observe a boost in the geometric quality of the reconstructed surfaces in state-of-the-art neural scene representations. + +Extensive evaluations on various baselines and datasets show the effectiveness of our method, which improves 3D reconstruction quality without resorting to a more sophisticated surface modeling. The overhead caused by the shaping of the normals varies with different neural architectures of the scene representation, e.g., some are marginal, and some are significant, which we deem as a limitation of our current pipeline. We believe that an acceleration in computation can be achieved using a coarse-to-fine strategy, which we leave as a future work. + +Acknowledgments. We thank the anonymous reviewers for their valuable feedback. This work is supported by the Early Career Scheme of the Research Grants Council (grant # 27207224), the HKU-100 Award, and in part by NSF China (No. 62172363). Siyan Dong would also like to thank the support from HKU Musketeers Foundation Institute of Data Science for the Postdoctoral Research Fellowship. + +# References + +1. Bae, G., Budvytis, I., Cipolla, R.: Estimating and exploiting the aleatoric uncertainty in surface normal estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13137-13146 (2021) +2. Bao, C., Zhang, Y., Yang, B., Fan, T., Yang, Z., Bao, H., Zhang, G., Cui, Z.: Sine: Semantic-driven image-based nerf editing with prior-guided editing field. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20919-20929 (2023) +3. Cai, B., Huang, J., Jia, R., Lv, C., Fu, H.: Neuda: Neural deformable anchor for high-fidelity implicit surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8476-8485 (2023) +4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) +5. Chen, Y., Wu, Q., Zheng, C., Cham, T.J., Cai, J.: Sem2nerf: Converting single-view semantic masks to neural radiance fields. In: European Conference on Computer Vision. pp. 730-748. Springer (2022) +6. Curless, B., Levoy, M.: A volumetric method for building complex models from range images. In: Proceedings of the 23rd annual conference on Computer graphics and interactive techniques. pp. 303-312 (1996) +7. Do, T., Vuong, K., Roumeliotis, S.I., Park, H.S.: Surface normal estimation of tilted images via spatial rectifier. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 265-280. Springer (2020) +8. Fu, Q., Xu, Q., Ong, Y.S., Tao, W.: Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems 35, 3403-3416 (2022) +9. Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099 (2020) +10. Guo, H., Peng, S., Lin, H., Wang, Q., Zhang, G., Bao, H., Zhou, X.: Neural 3d scene reconstruction with the manhattan-world assumption. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5511-5520 (2022) +11. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: Lerf: Language embedded radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19729-19739 (2023) +12. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023) + +13. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023) +14. Liu, F., Zhang, C., Zheng, Y., Duan, Y.: Semantic ray: Learning a generalizable semantic field with cross-reprojection attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17386-17396 (2023) +15. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13-23 (2023) +16. Liu, Z., Milano, F., Frey, J., Siegwart, R., Blum, H., Cadena, C.: Unsupervised continual semantic adaptation through neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3031-3040 (2023) +17. Long, X., Lin, C., Wang, P., Komura, T., Wang, W.: Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In: European Conference on Computer Vision. pp. 210-227. Springer (2022) +18. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. In: Seminal graphics: pioneering efforts that shaped the field, pp. 347-353 (1998) +19. Meng, X., Chen, W., Yang, B.: Neat: Learning neural implicit surfaces with arbitrary topologies from multi-view images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 248-258 (2023) +20. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12663-12673 (2023) +21. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +22. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) +23. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019) +24. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +25. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016) +26. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., Clarkson, A., Yan, M., Budge, B., Yan, Y., Pan, X., Yon, J., Zou, Y., Leon, K., Carter, N., Briales, J., Gillingham, T., Mueggler, E., Pesqueira, L., Savva, M., Batra, D., Strasdat, H.M., Nardi, R.D., Goesele, M., Lovegrove, S., Newcombe, R.: The replica dataset: A digital replica of indoor spaces (2019) + +27. Tertikas, K., Despoina, P., Pan, B., Park, J.J., Uy, M.A., Emiris, I., Avrithis, Y., Guibas, L.: Partnerf: Generating part-aware editable 3d shapes without 3d supervision. arXiv preprint arXiv:2303.09554 (2023) +28. Wang, F., Galliani, S., Vogel, C., Speciale, P., Pollefeys, M.: Patchmatchnet: Learned multi-view patchmatch stereo. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14194-14203 (2021) +29. Wang, J., Wang, P., Long, X., Theobalt, C., Komura, T., Liu, L., Wang, W.: Neuris: Neural reconstruction of indoor scenes using normal priors. In: European Conference on Computer Vision. pp. 139-155. Springer (2022) +30. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689 (2021) +31. Weder, S., Garcia-Hernando, G., Monszpart, A., Pollefeys, M., Brostow, G.J., Firman, M., Vicente, S.: Removing objects from neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16528-16538 (2023) +32. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In: European conference on computer vision. pp. 106-122. Springer (2022) +33. Xu, L., Xiangli, Y., Peng, S., Pan, X., Zhao, N., Theobalt, C., Dai, B., Lin, D.: Grid-guided neural radiance fields for large urban scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8296-8306 (2023) +34. Xu, X., Yang, Y., Mo, K., Pan, B., Yi, L., Guibas, L.: Jacobinerf: Nerf shaping with mutual information gradients. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16498-16507 (2023) +35. Yan, Z., Li, C., Lee, G.H.: Nerf-ds: Neural radiance fields for dynamic specular objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8285-8295 (2023) +36. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multi-view stereo. In: Proceedings of the European conference on computer vision (ECCV). pp. 767-783 (2018) +37. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021) +38. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: Proceedings of the International Conference on Computer Vision (ICCV) (2023) +39. Yu, H., Julin, J., Milacski, Z.A., Niinuma, K., Jeni, L.A.: Dylin: Making light field networks dynamic. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12397-12406 (2023) +40. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems 35, 25018-25032 (2022) +41. Zhang, Y., Yang, G., Guibas, L., Yang, Y.: Infogaussian: Structure-aware dynamic gaussians through lightweight information shaping. arXiv preprint arXiv:2406.05897 (2024) +42. Zhi, S., Laidlow, T., Leutenegger, S., Davison, A.J.: In-place scene labelling and understanding with implicit scene representation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15838-15847 (2021) + +43. Zhu, B., Yang, Y., Wang, X., Zheng, Y., Guibas, L.: Vdn-nerf: Resolving shape- radiance ambiguity via view-dependence normalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 35-45 (2023) +44. Zhu, J., Huo, Y., Ye, Q., Luan, F., Li, J., Xi, D., Wang, L., Tang, R., Hua, W., Bao, H., Wang, R.: I $^2$ -sdf: Intrinsic indoor scene reconstruction and editing via raytracing in neural sdfs. In: CVPR (2023) \ No newline at end of file diff --git a/2024/InfoNorm_ Mutual Information Shaping of Normals for Sparse-View Reconstruction/images.zip b/2024/InfoNorm_ Mutual Information Shaping of Normals for Sparse-View Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..672849c8df01404bd2dbb9ea63355c0079f1ca86 --- /dev/null +++ b/2024/InfoNorm_ Mutual Information Shaping of Normals for Sparse-View Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:532e3f0f10e9376f7413a796f055a26a18dbee05da3c554a48480961a9d92d02 +size 586536 diff --git a/2024/InfoNorm_ Mutual Information Shaping of Normals for Sparse-View Reconstruction/layout.json b/2024/InfoNorm_ Mutual Information Shaping of Normals for Sparse-View Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f1659e97615aae61f06d07012ab467e85531e70b --- /dev/null +++ b/2024/InfoNorm_ Mutual Information Shaping of Normals for Sparse-View Reconstruction/layout.json @@ -0,0 +1,9394 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "spans": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "type": "text", + "content": "InfoNorm: Mutual Information Shaping of Normals for Sparse-View Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 148, + 167, + 463, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 167, + 463, + 181 + ], + "spans": [ + { + "bbox": [ + 148, + 167, + 463, + 181 + ], + "type": "text", + "content": "Xulong Wang" + }, + { + "bbox": [ + 148, + 167, + 463, + 181 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 148, + 167, + 463, + 181 + ], + "type": "text", + "content": " Siyan Dong" + }, + { + "bbox": [ + 148, + 167, + 463, + 181 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 148, + 167, + 463, + 181 + ], + "type": "text", + "content": " Youyi Zheng" + }, + { + "bbox": [ + 148, + 167, + 463, + 181 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 148, + 167, + 463, + 181 + ], + "type": "text", + "content": " Yanchao Yang" + }, + { + "bbox": [ + 148, + 167, + 463, + 181 + ], + "type": "inline_equation", + "content": "^{3,4\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 160, + 189, + 452, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 189, + 452, + 200 + ], + "spans": [ + { + "bbox": [ + 160, + 189, + 452, + 200 + ], + "type": "text", + "content": "1 State Key Lab of CAD&CG, Zhejiang University 2 Chohotech co. Ltd." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 188, + 201, + 424, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 201, + 424, + 212 + ], + "spans": [ + { + "bbox": [ + 188, + 201, + 424, + 212 + ], + "type": "text", + "content": "3 Institute of Data Science, The University of Hong Kong" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 212, + 480, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 212, + 480, + 223 + ], + "spans": [ + { + "bbox": [ + 132, + 212, + 480, + 223 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 132, + 212, + 480, + 223 + ], + "type": "text", + "content": " Department of Electrical and Electronic Engineering, The University of Hong Kong" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 252, + 453, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 252, + 453, + 460 + ], + "spans": [ + { + "bbox": [ + 160, + 252, + 453, + 460 + ], + "type": "text", + "content": "Abstract. 3D surface reconstruction from multi-view images is essential for scene understanding and interaction. However, complex indoor scenes pose challenges such as ambiguity due to limited observations. Recent implicit surface representations, such as Neural Radiance Fields (NeRFs) and signed distance functions (SDFs), employ various geometric priors to resolve the lack of observed information. Nevertheless, their performance heavily depends on the quality of the pre-trained geometry estimation models. To ease such dependence, we propose regularizing the geometric modeling by explicitly encouraging the mutual information among surface normals of highly correlated scene points. In this way, the geometry learning process is modulated by the second-order correlations from noisy (first-order) geometric priors, thus eliminating the bias due to poor generalization. Additionally, we introduce a simple yet effective scheme that utilizes semantic and geometric features to identify correlated points, enhancing their mutual information accordingly. The proposed technique can serve as a plugin for SDF-based neural surface representations. Our experiments demonstrate the effectiveness of the proposed in improving the surface reconstruction quality of major states of the arts. Our code is available at: https://github.com/Muliphein/InfoNorm." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 472, + 422, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 472, + 422, + 482 + ], + "spans": [ + { + "bbox": [ + 160, + 472, + 422, + 482 + ], + "type": "text", + "content": "Keywords: 3D from multi-view images " + }, + { + "bbox": [ + 160, + 472, + 422, + 482 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 472, + 422, + 482 + ], + "type": "text", + "content": " surface reconstruction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 503, + 230, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 230, + 516 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 230, + 516 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "content": "3D surface reconstruction from multi-view images is an essential task in the computer vision and graphics community, with practical applications including content creation for virtual reality and robot-scene interaction. However, indoor scenes present challenges due to the large and complex scenes captured from sparse viewpoints. These challenges encompass issues such as occlusion and ambiguity arising from limited observations. As a result, traditional Multi-View Stereo (MVS) methods [28,36], which require substantial overlap among images, may not produce satisfactory results in these scenarios." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 631, + 403, + 643 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 631, + 403, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 631, + 403, + 643 + ], + "type": "text", + "content": "* Equal contributions (wangxulong@zju.edu.cn, siyan3d@hku.hk)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 135, + 643, + 240, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 643, + 240, + 654 + ], + "spans": [ + { + "bbox": [ + 135, + 643, + 240, + 654 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 135, + 654, + 441, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 654, + 441, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 654, + 441, + 665 + ], + "type": "text", + "content": "2 This work was done during the author's internship at Chohotech Co. Ltd.." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 246, + 178 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 246, + 178 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 246, + 178 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 246, + 178 + ], + "type": "image", + "image_path": "9b8401e3fa530d78df7d772cbc467ce6c418352d13234aeaf0b6384a2139842b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 182, + 199, + 190 + ], + "lines": [ + { + "bbox": [ + 178, + 182, + 199, + 190 + ], + "spans": [ + { + "bbox": [ + 178, + 182, + 199, + 190 + ], + "type": "text", + "content": "VoISDF" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 116, + 362, + 182 + ], + "blocks": [ + { + "bbox": [ + 250, + 116, + 362, + 182 + ], + "lines": [ + { + "bbox": [ + 250, + 116, + 362, + 182 + ], + "spans": [ + { + "bbox": [ + 250, + 116, + 362, + 182 + ], + "type": "image", + "image_path": "806ffe2576270f50b515b669174c1942b2de766ad6891de348a0e98e77682aa9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 182, + 327, + 191 + ], + "lines": [ + { + "bbox": [ + 284, + 182, + 327, + 191 + ], + "spans": [ + { + "bbox": [ + 284, + 182, + 327, + 191 + ], + "type": "text", + "content": "VoISDF+ (Ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 366, + 116, + 478, + 178 + ], + "blocks": [ + { + "bbox": [ + 366, + 116, + 478, + 178 + ], + "lines": [ + { + "bbox": [ + 366, + 116, + 478, + 178 + ], + "spans": [ + { + "bbox": [ + 366, + 116, + 478, + 178 + ], + "type": "image", + "image_path": "4fa19faf7ef2ee3e5792b72952735bf2af6b7b5b9c5badae8ff9e895d4e9b334.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 182, + 440, + 190 + ], + "lines": [ + { + "bbox": [ + 402, + 182, + 440, + 190 + ], + "spans": [ + { + "bbox": [ + 402, + 182, + 440, + 190 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 136, + 193, + 247, + 243 + ], + "blocks": [ + { + "bbox": [ + 136, + 193, + 247, + 243 + ], + "lines": [ + { + "bbox": [ + 136, + 193, + 247, + 243 + ], + "spans": [ + { + "bbox": [ + 136, + 193, + 247, + 243 + ], + "type": "image", + "image_path": "5a2d9d718867f7ac2f1ce0a97aca24769f738ae92246bee6311fadca4d578343.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 243, + 203, + 251 + ], + "lines": [ + { + "bbox": [ + 175, + 243, + 203, + 251 + ], + "spans": [ + { + "bbox": [ + 175, + 243, + 203, + 251 + ], + "type": "text", + "content": "GeoNeuS" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 260, + 481, + 305 + ], + "lines": [ + { + "bbox": [ + 130, + 260, + 481, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 481, + 305 + ], + "type": "text", + "content": "Fig. 1: 3D scene reconstruction from sparse views on Replica [26] (first row) and ScanNet++ [38] (second row). By enforcing the mutual information between the normals of highly correlated scene points, the proposed method can effectively enhance the reconstruction quality of the baselines (VolSDF [37] and GeoNeuS [8])." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 252, + 193, + 361, + 243 + ], + "blocks": [ + { + "bbox": [ + 252, + 193, + 361, + 243 + ], + "lines": [ + { + "bbox": [ + 252, + 193, + 361, + 243 + ], + "spans": [ + { + "bbox": [ + 252, + 193, + 361, + 243 + ], + "type": "image", + "image_path": "ca123373e0b175854094b84eb9138c7da3f4b7698a126c30ac38be6287709a8c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 243, + 331, + 251 + ], + "lines": [ + { + "bbox": [ + 282, + 243, + 331, + 251 + ], + "spans": [ + { + "bbox": [ + 282, + 243, + 331, + 251 + ], + "type": "text", + "content": "GeoNeuS" + }, + { + "bbox": [ + 282, + 243, + 331, + 251 + ], + "type": "inline_equation", + "content": "^{+}" + }, + { + "bbox": [ + 282, + 243, + 331, + 251 + ], + "type": "text", + "content": " (Ours)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 365, + 194, + 477, + 243 + ], + "blocks": [ + { + "bbox": [ + 365, + 194, + 477, + 243 + ], + "lines": [ + { + "bbox": [ + 365, + 194, + 477, + 243 + ], + "spans": [ + { + "bbox": [ + 365, + 194, + 477, + 243 + ], + "type": "image", + "image_path": "163c73317edb184b56866c497313fa562e0cfff80ed358069df6bcaa7a5f696a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 243, + 440, + 251 + ], + "lines": [ + { + "bbox": [ + 402, + 243, + 440, + 251 + ], + "spans": [ + { + "bbox": [ + 402, + 243, + 440, + 251 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 331, + 480, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 480, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 480, + 462 + ], + "type": "text", + "content": "Recently, implicit scene representations, e.g., Neural Radiance Fields (NeRFs) [21], have been proposed to encode 3D scenes into a set of neural network parameters, with subsequent works enhancing the representational capacity for geometry to enable surface extraction. The key idea involves incorporating signed distance functions (SDFs) [23] into the learned radiance field. Although NeRFs with SDF produce high-quality reconstructions of simple scenes with sufficient images, they still underperform in large and complex scenes captured from sparse viewpoints. More recent approaches, such as NeuRIS [29] and MonoSDF [40], propose using monocular geometric priors to guide the learning of the radiance field. However, they can be affected by inevitable errors and noise in the monocular geometry estimation modules." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 464, + 481, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 481, + 631 + ], + "type": "text", + "content": "We explore geometric regularization in a second-order perspective. We argue that both the geometry itself and the correlation of geometry among scene regions are crucial for surface reconstruction, especially when dealing with varying indoor scenes under sparse views with noisy geometry estimates. The correlation is expressed as mutual information between scene entities under random perturbations of the network weights. More explicitly, we propose to enforce the mutual information between the normals of two points in the scene that are considered geometrically correlated, where the mutual information can be efficiently computed by measuring the cosine similarity between the gradients of the normals with respect to the perturbed weights. Further, we leverage a combination of pretrained semantic and geometric features [4,7,12] to identify the correlated scene regions that endorse high mutual information. This results in an easy-to-use geometric shaping1 technique that can be applied to any SDF-based neural radiance fields. To thoroughly evaluate its effectiveness, we test the proposed with" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 272, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 272, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 272, + 102 + ], + "type": "text", + "content": "X. Wang & S. Dong et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 480, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 480, + 665 + ], + "type": "text", + "content": "1 We use the term \"shaping\" to refer to the process of optimizing network parameters to enforce consistencies in the scene geometry via aligning the gradients." + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "multiple state-of-the-art neural surface reconstruction models. The experimental results confirm the effectiveness of the proposed shaping technique, showing improved surface reconstruction quality (e.g., Fig. 1) on challenging scenes with limited views. To summarize:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 173, + 481, + 281 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 173, + 481, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 481, + 209 + ], + "type": "text", + "content": "- We propose to leverage second-order geometric correlation in the form of mutual information between normals to regularize the surface reconstruction of SDF-based NeRF representations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 209, + 481, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 481, + 245 + ], + "type": "text", + "content": "- We develop a pipeline that efficiently enforces the mutual information of the normals of points, deemed geometrically correlated with pre-trained multimodal features, to improve the reconstruction quality of the scene." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 246, + 481, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 281 + ], + "type": "text", + "content": "- We verify the effectiveness of the proposed second-order regularization technique across a broad spectrum of baselines and demonstrate its usefulness as an easy-to-use plugin in improving 3D surface modeling." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 300, + 237, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 237, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 237, + 314 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 328, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 482, + 399 + ], + "type": "text", + "content": "3D Reconstruction with neural radiance fields (NeRFs). NeRF [21] is originally designed for novel view synthesis, with its variants serving as scene representations [15, 32, 33, 35, 39] for tasks related to editing [2, 31, 44], semantic segmentation [5, 14, 16, 34], 3D shape generation [20, 27], and so on. We mainly focus on neural surface representations that enable the extraction of high-quality 3D meshes from the trained NeRFs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 400, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 544 + ], + "type": "text", + "content": "Early works such as NeuS [30] and VolSDF [37] combine signed distance functions (SDFs) with volume rendering to extract high-quality surfaces. Recently, SparseNeuS [17] introduces a cascaded geometry reasoning framework that generalizes well to novel scenes. GeoNeuS [8] uses surface points from structure-from-motion and neighbor-view patches to reconstruct surfaces, while NeuDA [3] employs a deformable anchor to adaptively encode geometric details. NeAT [19] broadens the scope of reconstruction to arbitrary topology, no longer confining to watertight surfaces. Moreover, Neuralangelo [13] uses a hash grid for position encoding and implements a coarse-to-fine optimization with numerical gradients. Despite their promising results, most of these methods need sufficient training views, and their performance declines when dealing with large and complex scenes under sparse viewpoints." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": "Incorporating priors into NeRFs. One can augment scene representations by introducing various priors to NeRFs. VDN-NeRF [43] normalizes the spatial feature to align with monocular features. LERF [11] introduces a new branch whose outputs align with DINO [4] and CLIP [24] features. Semantic-NeRF [42] encodes semantic information for segmentation tasks, while JacobiNeRF [34] shapes the color gradients to encode the mutual information in terms of semantic similarity from DINO [4] features for label propagation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "For surface reconstruction, GeoNeuS [8] incorporates patch similarity from neighboring images and sparse surface points derived from structure-from-motion" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "InfoNorm" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "text", + "content": "[25], while I²-SDF [44] combines the rendering equation with VolSDF [37] for better novel views and geometry. To cope with complex indoor scenes, NeuRIS [29] incorporates the normal priors as additional geometric supervision. Further, MonoSDF [40] uses both depth and normal to improve the surface reconstruction quality. However, these methods highly depend on the quality of the supplemented geometric information. In contrast, our method leverages the second-order information in geometric prior, rather than directly using the prior as a first-order supervision signal, which makes it more robust to errors and noise. There are also methods utilizing strong assumptions derived for human-made scenes. For instance, ManhattanSDF [10] works under the assumption that all captured scenes adhere to the Manhattan-world concept, thus, predicting semantic segmentation to identify walls and floors. While these methods primarily focus on planar scenes, the proposed technique can work for general scenes under noisy correlation information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 304, + 202, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 202, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 202, + 316 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "content": "We start with a discussion on the preliminaries in Sec. 3.1 to provide the background on how we can learn from posed images an implicit scene representation (NeRF), improve 3D scene reconstruction via the combination of NeRF and SDF representations, and employ mutual information as a constraint in the NeRF training process. Following this, in Sec. 3.2, we delve into how mutual information can be used as constraints on NeRF's implied surfaces, so that we can learn better geometries for higher-quality 3D reconstruction of the scene. Finally, we elaborate on how we can employ off-the-shelf semantic and geometry features to assist the encoding of mutual information through the proposed geometric shaping framework." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 472, + 228, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 228, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 228, + 483 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 493, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 482, + 590 + ], + "type": "text", + "content": "Neural radiance fields. NeRF [21] learns an implicit scene representation from a set of posed images. For a specific pixel in an image, NeRF samples a list of 3D points, " + }, + { + "bbox": [ + 130, + 493, + 482, + 590 + ], + "type": "inline_equation", + "content": "\\{x|x_{t} = \\mathbf{o} + t\\mathbf{v},t\\in [t_{near},t_{far}]\\}" + }, + { + "bbox": [ + 130, + 493, + 482, + 590 + ], + "type": "text", + "content": ", from the camera center " + }, + { + "bbox": [ + 130, + 493, + 482, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 130, + 493, + 482, + 590 + ], + "type": "text", + "content": " along the viewing direction " + }, + { + "bbox": [ + 130, + 493, + 482, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 130, + 493, + 482, + 590 + ], + "type": "text", + "content": " within a bounded space. Specifically, it employs a Multilayer Perceptron (MLP) to encode the volume density and color of each sampled point. We denote all the parameters of NeRF as " + }, + { + "bbox": [ + 130, + 493, + 482, + 590 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 493, + 482, + 590 + ], + "type": "text", + "content": ", the density function as " + }, + { + "bbox": [ + 130, + 493, + 482, + 590 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 493, + 482, + 590 + ], + "type": "text", + "content": ", and the color function as " + }, + { + "bbox": [ + 130, + 493, + 482, + 590 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 493, + 482, + 590 + ], + "type": "text", + "content": ". The pixel's rendered color " + }, + { + "bbox": [ + 130, + 493, + 482, + 590 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{C}}" + }, + { + "bbox": [ + 130, + 493, + 482, + 590 + ], + "type": "text", + "content": " from a NeRF is calculated by the discrete volume rendering:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 222, + 601, + 480, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 601, + 480, + 639 + ], + "spans": [ + { + "bbox": [ + 222, + 601, + 480, + 639 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {C}} (\\mathbf {o}, \\mathbf {v}; \\theta) = \\sum_ {i = 1} ^ {N} \\omega (\\mathbf {o}, \\mathbf {v}, t _ {i}; \\theta) c \\left(x _ {t _ {i}}, \\mathbf {v}; \\theta\\right), \\tag {1}", + "image_path": "1ea951629e72e5a3424b8a7d1b425cd4fe1affa1572996905c975d82e3df0bad.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 233, + 634, + 389, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 634, + 389, + 667 + ], + "spans": [ + { + "bbox": [ + 233, + 634, + 389, + 667 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} \\omega (\\mathbf {o}, \\mathbf {v}, t _ {i}; \\theta) = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}) \\alpha_ {i}.", + "image_path": "657c3c31108bc97879394b0a6fc2bb63ce752b60361214db40f05ded8f0d9c63.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 271, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 271, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 271, + 101 + ], + "type": "text", + "content": "X. Wang & S. Dong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\alpha_{i} = 1 - \\exp (-\\sigma_{i}\\delta_{i})" + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "content": " is the abbreviation of " + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\sigma (x_{t_i};\\theta)" + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\delta_{i} = t_{i + 1} - t_{i}" + }, + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "content": ". A photometric loss is utilized to minimize the discrepancy between the rendered color and its ground truth:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 270, + 162, + 480, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 162, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 270, + 162, + 480, + 175 + ], + "type": "interline_equation", + "content": "L _ {C} = \\left\\| \\mathbf {C} - \\hat {\\mathbf {C}} \\right\\| _ {1}, \\tag {2}", + "image_path": "697073eed0a760cc5322e474c49c026222996e461f1586414d78b8223ac1e4b7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 185, + 479, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 185, + 479, + 233 + ], + "spans": [ + { + "bbox": [ + 130, + 185, + 479, + 233 + ], + "type": "text", + "content": "so that the parameters " + }, + { + "bbox": [ + 130, + 185, + 479, + 233 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 185, + 479, + 233 + ], + "type": "text", + "content": " can be optimized to fit the scene. Despite the ability to handle photorealistic novel view synthesis, NeRFs suffer from noisy and unrealistic surface extraction when it comes to 3D reconstruction, as discussed in the literature [30, 37]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 246, + 479, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 246, + 479, + 328 + ], + "spans": [ + { + "bbox": [ + 131, + 246, + 479, + 328 + ], + "type": "text", + "content": "NeRFs with signed distance functions. To produce high-quality surfaces, it is beneficial to use signed distance functions (SDFs) [6, 23] to represent geometry. Consequently, the surface can be reconstructed by extracting the zero-level set [18] of the SDF values. The combination of NeRF and SDF can be achieved by deriving the NeRF's densities from the SDF, or defining Laplace's cumulative distribution function. Accordingly, " + }, + { + "bbox": [ + 131, + 246, + 479, + 328 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 131, + 246, + 479, + 328 + ], + "type": "text", + "content": " in Eq. 1 can be represented by a function " + }, + { + "bbox": [ + 131, + 246, + 479, + 328 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 131, + 246, + 479, + 328 + ], + "type": "text", + "content": " which converts SDF values to volume densities as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 263, + 341, + 480, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 341, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 263, + 341, + 480, + 353 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} = \\Phi (x, f (x; \\theta), i). \\tag {3}", + "image_path": "c641b3e38db4648e47063ddc12a51cce84fff77e99c9bc53375801473d6dc098.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 362, + 479, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 362, + 479, + 387 + ], + "spans": [ + { + "bbox": [ + 131, + 362, + 479, + 387 + ], + "type": "text", + "content": "Generally, NeRFs with an SDF achieve improved surface reconstruction quality, where the key to maintaining the characteristics of SDF is an eikonal loss [9]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 240, + 397, + 480, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 397, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 240, + 397, + 480, + 429 + ], + "type": "interline_equation", + "content": "L _ {E} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\left| \\left| \\nabla f \\left(x _ {i}\\right) \\right| \\right| _ {2} - 1\\right) ^ {2}. \\tag {4}", + "image_path": "dabc536d5bb9109686bae8eba5a463adeee21c6fe8ca96e954c4b75356ceb238.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 441, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 441, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 441, + 482, + 574 + ], + "type": "text", + "content": "Mutual information in NeRFs. Mutual information (MI) quantifies the statistical dependence between two random variables, providing an estimate of how much information they share. It has been introduced for semantic segmentation and motion modeling tasks [34, 41]. In this paper, We study the mutual information under perturbations of the NeRF parameters, and focus more on the geometry properties. Let " + }, + { + "bbox": [ + 131, + 441, + 482, + 574 + ], + "type": "inline_equation", + "content": "I(p_{i})" + }, + { + "bbox": [ + 131, + 441, + 482, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 441, + 482, + 574 + ], + "type": "inline_equation", + "content": "I(p_{j})" + }, + { + "bbox": [ + 131, + 441, + 482, + 574 + ], + "type": "text", + "content": " denote two pixels derived from the NeRF rendering process " + }, + { + "bbox": [ + 131, + 441, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\mathbb{F}" + }, + { + "bbox": [ + 131, + 441, + 482, + 574 + ], + "type": "text", + "content": ". And we assume that a subset of the NeRF weights, denoted by " + }, + { + "bbox": [ + 131, + 441, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\theta^{D}" + }, + { + "bbox": [ + 131, + 441, + 482, + 574 + ], + "type": "text", + "content": ", is perturbed by a random noise vector " + }, + { + "bbox": [ + 131, + 441, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{n} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 131, + 441, + 482, + 574 + ], + "type": "text", + "content": " sampled from a uniform distribution on the sphere " + }, + { + "bbox": [ + 131, + 441, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^{D - 1}" + }, + { + "bbox": [ + 131, + 441, + 482, + 574 + ], + "type": "text", + "content": " by a small step " + }, + { + "bbox": [ + 131, + 441, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\gamma \\ll 1.0" + }, + { + "bbox": [ + 131, + 441, + 482, + 574 + ], + "type": "text", + "content": ". The random variables corresponding to the perturbed pixels can be written through a Taylor expansion:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 255, + 582, + 360, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 582, + 360, + 604 + ], + "spans": [ + { + "bbox": [ + 255, + 582, + 360, + 604 + ], + "type": "interline_equation", + "content": "\\hat {I} (p _ {i}) = I (p _ {i}) + \\gamma \\mathbf {n} \\frac {\\partial \\mathbb {F} _ {i}}{\\partial \\theta^ {D}},", + "image_path": "cc37ed55efc61d1be7213b125e89ef00727e69d7ac41dcd1cb438fef66fbb445.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 254, + 606, + 360, + 629 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 606, + 360, + 629 + ], + "spans": [ + { + "bbox": [ + 254, + 606, + 360, + 629 + ], + "type": "interline_equation", + "content": "\\hat {I} (p _ {j}) = I (p _ {j}) + \\gamma \\mathbf {n} \\frac {\\partial \\mathbb {F} _ {j}}{\\partial \\theta^ {D}}.", + "image_path": "1869b2415c62d0d92a8c27aa6a15ea2e126f7798db4fc231114924933239342a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 639, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 639, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 639, + 481, + 666 + ], + "type": "text", + "content": "We can now characterize the mutual information between " + }, + { + "bbox": [ + 131, + 639, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\hat{I}(p_i)" + }, + { + "bbox": [ + 131, + 639, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 639, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\hat{I}(p_j)" + }, + { + "bbox": [ + 131, + 639, + 481, + 666 + ], + "type": "text", + "content": " under the joint probability distribution " + }, + { + "bbox": [ + 131, + 639, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(\\hat{I}(p_i), \\hat{I}(p_j))" + }, + { + "bbox": [ + 131, + 639, + 481, + 666 + ], + "type": "text", + "content": ". As derived in JacobiNeRF, the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "InfoNorm" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 313, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 313, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 313, + 128 + ], + "type": "text", + "content": "mutual information can be computed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 212, + 136, + 402, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 136, + 402, + 178 + ], + "spans": [ + { + "bbox": [ + 212, + 136, + 402, + 178 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbb {I} (\\hat {I} (p _ {i}), \\hat {I} (p _ {j})) = \\mathbb {H} (\\hat {I} (p _ {i})) - \\mathbb {H} (\\hat {I} (p _ {i}) | \\hat {I} (p _ {j})) \\\\ = \\log (\\frac {1}{\\sqrt {1 - \\cos^ {2} \\beta}}) + \\mathrm {c o n s t .}. \\\\ \\end{array}", + "image_path": "3bd7053ceafbde92165211c213f109fc8c1a45a18393f8a5b610edc20c12ffbc.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 477, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 477, + 199 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 477, + 199 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 132, + 185, + 477, + 199 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 132, + 185, + 477, + 199 + ], + "type": "text", + "content": " denotes the angle between " + }, + { + "bbox": [ + 132, + 185, + 477, + 199 + ], + "type": "inline_equation", + "content": "\\partial \\mathbb{F}_i / \\partial \\theta^D" + }, + { + "bbox": [ + 132, + 185, + 477, + 199 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 185, + 477, + 199 + ], + "type": "inline_equation", + "content": "\\partial \\mathbb{F}_j / \\partial \\theta^D" + }, + { + "bbox": [ + 132, + 185, + 477, + 199 + ], + "type": "text", + "content": ". It is observed that:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 252, + 207, + 481, + 221 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 207, + 481, + 221 + ], + "spans": [ + { + "bbox": [ + 252, + 207, + 481, + 221 + ], + "type": "interline_equation", + "content": "\\mathbb {I} (\\hat {I} (p _ {i}), \\hat {I} (p _ {j})) \\propto | | \\cos \\beta | |. \\tag {5}", + "image_path": "176872d45630eb76ec7bc67dde8d4e9f46e22ba0585ddfdbd3f897ad842be99a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 227, + 481, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 227, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 131, + 227, + 481, + 287 + ], + "type": "text", + "content": "Therefore, to enhance the MI between two random pixels induced by NeRF perturbations, one can enforce the absolute value of the cosine similarity. That is, if two pixels " + }, + { + "bbox": [ + 131, + 227, + 481, + 287 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 131, + 227, + 481, + 287 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 227, + 481, + 287 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 131, + 227, + 481, + 287 + ], + "type": "text", + "content": " come with high correlation (e.g., from the same object in the scene), their gradients regarding the perturbed parameters should be in the same or opposite direction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 303, + 481, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 481, + 316 + ], + "type": "text", + "content": "3.2 Mutual Information Shaping of Geometry in SDF-based NeRFs" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 322, + 481, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 322, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 131, + 322, + 481, + 357 + ], + "type": "text", + "content": "While the previous approach focuses on enforcing mutual information for image synthesis, we explore a mutual information shaping of the geometries for quality surface reconstruction in SDF-based NeRFs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 358, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 358, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 358, + 482, + 514 + ], + "type": "text", + "content": "Specifically, we represent a scene via a NeRF backbone with an SDF head and a color head, as shown in Fig. 2. Note that the method to be proposed is capable of being applied to any network similar to NeRF and SDF representations. Given that the two heads partially share parameters, we denote all the parameters as " + }, + { + "bbox": [ + 131, + 358, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 358, + 482, + 514 + ], + "type": "text", + "content": ", and use " + }, + { + "bbox": [ + 131, + 358, + 482, + 514 + ], + "type": "inline_equation", + "content": "f(x;\\theta)" + }, + { + "bbox": [ + 131, + 358, + 482, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 358, + 482, + 514 + ], + "type": "inline_equation", + "content": "c(x,\\mathbf{v};\\theta)" + }, + { + "bbox": [ + 131, + 358, + 482, + 514 + ], + "type": "text", + "content": " to represent the SDF and color branches, respectively. To achieve mutual information shaping of the geometry, we focus on manipulating the parameters of " + }, + { + "bbox": [ + 131, + 358, + 482, + 514 + ], + "type": "inline_equation", + "content": "f(x;\\theta)" + }, + { + "bbox": [ + 131, + 358, + 482, + 514 + ], + "type": "text", + "content": ", while keeping " + }, + { + "bbox": [ + 131, + 358, + 482, + 514 + ], + "type": "inline_equation", + "content": "c(x,\\mathbf{v};\\theta)" + }, + { + "bbox": [ + 131, + 358, + 482, + 514 + ], + "type": "text", + "content": " unchanged. Below, we first explain why and how we compute mutual information with surface normals derived from SDF, i.e., " + }, + { + "bbox": [ + 131, + 358, + 482, + 514 + ], + "type": "inline_equation", + "content": "f(x;\\theta)" + }, + { + "bbox": [ + 131, + 358, + 482, + 514 + ], + "type": "text", + "content": ", as a geometry-aware constraint. Next, we present a simple yet effective strategy that extracts regions with high correlation, which need to be encoded into the weights using the mutual information shaping technique. Finally, we elaborate on our loss function and training process to reconstruct the scene from posed images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 523, + 482, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 523, + 482, + 595 + ], + "spans": [ + { + "bbox": [ + 131, + 523, + 482, + 595 + ], + "type": "text", + "content": "Mutual information among surface normals. We explicitly encode mutual information into the function " + }, + { + "bbox": [ + 131, + 523, + 482, + 595 + ], + "type": "inline_equation", + "content": "f(x; \\theta)" + }, + { + "bbox": [ + 131, + 523, + 482, + 595 + ], + "type": "text", + "content": ". However, instead of computing mutual information between the density values, we proceed by manipulating the surface normal vectors that can be derived from " + }, + { + "bbox": [ + 131, + 523, + 482, + 595 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 131, + 523, + 482, + 595 + ], + "type": "text", + "content": ". Specifically, for a 3D point " + }, + { + "bbox": [ + 131, + 523, + 482, + 595 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 523, + 482, + 595 + ], + "type": "text", + "content": " on a surface of the scene, its normal vector " + }, + { + "bbox": [ + 131, + 523, + 482, + 595 + ], + "type": "inline_equation", + "content": "N(x; \\theta)" + }, + { + "bbox": [ + 131, + 523, + 482, + 595 + ], + "type": "text", + "content": " can be calculated as the partial derivative of the 3D coordinate:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 264, + 601, + 481, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 601, + 481, + 624 + ], + "spans": [ + { + "bbox": [ + 264, + 601, + 481, + 624 + ], + "type": "interline_equation", + "content": "N (x; \\theta) = \\frac {\\partial f (x ; \\theta)}{\\partial x}. \\tag {6}", + "image_path": "a989b3759ae2db480737a1486711dbd21b60335436fd6fdb7e981cd65d1d136c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 481, + 666 + ], + "type": "text", + "content": "There are several reasons to manipulate surface normal vectors but not SDF values: 1) Shaping the SDF values with respect to a subset " + }, + { + "bbox": [ + 132, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\theta^D" + }, + { + "bbox": [ + 132, + 629, + 481, + 666 + ], + "type": "text", + "content": " of the parameters may not effectively impose the constraints on all parameters. However, the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 272, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 272, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 272, + 102 + ], + "type": "text", + "content": "X. Wang & S. Dong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 115, + 478, + 213 + ], + "blocks": [ + { + "bbox": [ + 136, + 115, + 478, + 213 + ], + "lines": [ + { + "bbox": [ + 136, + 115, + 478, + 213 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 478, + 213 + ], + "type": "image", + "image_path": "6f17f63afda6c3119d5acae9d04874923ea056d465e48b6553b8bf4685feb5f3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 222, + 482, + 310 + ], + "lines": [ + { + "bbox": [ + 131, + 222, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 131, + 222, + 482, + 310 + ], + "type": "text", + "content": "Fig. 2: An overview of the pipeline, where we apply mutual information shaping on the geometric branch to enforce consistencies that help enhance the surface reconstruction. Specifically, the model consists of the NeRF backbone, an SDF head, and a color head. The predicted density and color are supervised by the classic eikonal loss " + }, + { + "bbox": [ + 131, + 222, + 482, + 310 + ], + "type": "inline_equation", + "content": "L_{E}" + }, + { + "bbox": [ + 131, + 222, + 482, + 310 + ], + "type": "text", + "content": " and the photometric reconstruction loss " + }, + { + "bbox": [ + 131, + 222, + 482, + 310 + ], + "type": "inline_equation", + "content": "L_{C}" + }, + { + "bbox": [ + 131, + 222, + 482, + 310 + ], + "type": "text", + "content": ", respectively. We encode geometry-aware mutual information into a subset of the parameters " + }, + { + "bbox": [ + 131, + 222, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\theta^{D}" + }, + { + "bbox": [ + 131, + 222, + 482, + 310 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 131, + 222, + 482, + 310 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 131, + 222, + 482, + 310 + ], + "type": "text", + "content": " to constrain the density field learning for better surface quality, which is achieved by the proposed mutual information loss " + }, + { + "bbox": [ + 131, + 222, + 482, + 310 + ], + "type": "inline_equation", + "content": "L_{M}" + }, + { + "bbox": [ + 131, + 222, + 482, + 310 + ], + "type": "text", + "content": " computed on top of the estimated surface normal." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "text", + "content": "calculation of the surface normal includes the entire set of parameters in " + }, + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "text", + "content": ", thus a more global regularity. 2) Moreover, shaping " + }, + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "text", + "content": " with the mutual information between its values may induce side effects as density could be distorted to satisfy the constraints. However, shaping the normal " + }, + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "inline_equation", + "content": "\\partial f / \\partial x" + }, + { + "bbox": [ + 130, + 333, + 479, + 403 + ], + "type": "text", + "content": " leave more space for the SDF values to cope with the scene reconstruction task, thus minimizing potential negative effects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 405, + 480, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 405, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 131, + 405, + 480, + 441 + ], + "type": "text", + "content": "According to Eq. 5, we need to calculate the partial gradients " + }, + { + "bbox": [ + 131, + 405, + 480, + 441 + ], + "type": "inline_equation", + "content": "\\partial N / \\partial \\theta^D" + }, + { + "bbox": [ + 131, + 405, + 480, + 441 + ], + "type": "text", + "content": " to compute the mutual information. When it comes to a vector rather than a scalar, the gradients consist of three components corresponding to the " + }, + { + "bbox": [ + 131, + 405, + 480, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 131, + 405, + 480, + 441 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 405, + 480, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 131, + 405, + 480, + 441 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 405, + 480, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 131, + 405, + 480, + 441 + ], + "type": "text", + "content": " axes:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 206, + 449, + 480, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 449, + 480, + 473 + ], + "spans": [ + { + "bbox": [ + 206, + 449, + 480, + 473 + ], + "type": "interline_equation", + "content": "\\frac {\\partial N (x ; \\theta)}{\\partial \\theta^ {D}} = \\frac {\\partial N _ {\\mathbf {x}} (x ; \\theta)}{\\partial \\theta^ {D}}, \\frac {\\partial N _ {\\mathbf {y}} (x ; \\theta)}{\\partial \\theta^ {D}}, \\frac {\\partial N _ {\\mathbf {z}} (x ; \\theta)}{\\partial \\theta^ {D}}. \\tag {7}", + "image_path": "c657f92d15b8c9ea64f95cd857ad7b879499875b252449198b68e1ef95638b1f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 476, + 479, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 476, + 479, + 499 + ], + "spans": [ + { + "bbox": [ + 131, + 476, + 479, + 499 + ], + "type": "text", + "content": "For ease of computation, we concatenate the gradients into a vector, denoted as " + }, + { + "bbox": [ + 131, + 476, + 479, + 499 + ], + "type": "inline_equation", + "content": "\\partial N" + }, + { + "bbox": [ + 131, + 476, + 479, + 499 + ], + "type": "text", + "content": " for simplicity:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 233, + 498, + 480, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 498, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 233, + 498, + 480, + 521 + ], + "type": "interline_equation", + "content": "\\partial N = \\operatorname {C o n c a t} \\left(\\frac {\\partial N _ {\\mathbf {x}}}{\\partial \\theta^ {D}}, \\frac {\\partial N _ {\\mathbf {y}}}{\\partial \\theta^ {D}}, \\frac {\\partial N _ {\\mathbf {z}}}{\\partial \\theta^ {D}}\\right). \\tag {8}", + "image_path": "68971032f805a5328f9b897aa40336badb0f23a5f0c41ab666c64d0893e0eae3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 525, + 479, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 525, + 479, + 549 + ], + "spans": [ + { + "bbox": [ + 131, + 525, + 479, + 549 + ], + "type": "text", + "content": "Since a 2D pixel is accumulated from 3D, we aggregate the normal-based information along a viewing ray as in Eq. 1 as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 250, + 556, + 480, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 556, + 480, + 582 + ], + "spans": [ + { + "bbox": [ + 250, + 556, + 480, + 582 + ], + "type": "interline_equation", + "content": "\\mathbb {N} \\left(p _ {i}\\right) = \\sum_ {x \\in \\text {r a y}} \\omega \\left(t _ {i}; \\theta\\right) \\partial N. \\tag {9}", + "image_path": "67c0e02ce385267d3a1d98073fd628bdc30dda8df8c4cd5278059a3101f16079.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 589, + 448, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 589, + 448, + 601 + ], + "spans": [ + { + "bbox": [ + 131, + 589, + 448, + 601 + ], + "type": "text", + "content": "Finally, we obtain that the mutual information between two normals is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 226, + 608, + 480, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 608, + 480, + 635 + ], + "spans": [ + { + "bbox": [ + 226, + 608, + 480, + 635 + ], + "type": "interline_equation", + "content": "\\mathbb {I} (\\hat {N} (p _ {i}), \\hat {N} (p _ {j})) \\propto | | \\frac {\\mathbb {N} (p _ {i}) \\cdot \\mathbb {N} (p _ {j})}{| \\mathbb {N} (p _ {i}) | | \\mathbb {N} (p _ {j}) |} | |. \\tag {10}", + "image_path": "396161b59d1ed756cba0e508873eb36c2d2383bcd0e33576a8d7eea87655f76e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "Please refer to our supplementary material for a detailed derivation. From Eq. 10, we observe that the mutual information in surface normal vectors correlates" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "InfoNorm" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 115, + 218, + 178 + ], + "blocks": [ + { + "bbox": [ + 135, + 115, + 218, + 178 + ], + "lines": [ + { + "bbox": [ + 135, + 115, + 218, + 178 + ], + "spans": [ + { + "bbox": [ + 135, + 115, + 218, + 178 + ], + "type": "image", + "image_path": "a1037c099eeadeb4e12569c2335a9d39eec82a6e6ce806b409303ba67e34c5c1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 179, + 195, + 187 + ], + "lines": [ + { + "bbox": [ + 157, + 179, + 195, + 187 + ], + "spans": [ + { + "bbox": [ + 157, + 179, + 195, + 187 + ], + "type": "text", + "content": "Anchor pixel" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 221, + 115, + 305, + 178 + ], + "blocks": [ + { + "bbox": [ + 221, + 115, + 305, + 178 + ], + "lines": [ + { + "bbox": [ + 221, + 115, + 305, + 178 + ], + "spans": [ + { + "bbox": [ + 221, + 115, + 305, + 178 + ], + "type": "image", + "image_path": "fdc6e8d50ef3f3f196ffc7ad2e20073c5eefb613a549457884295d170dcec5a5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 179, + 299, + 187 + ], + "lines": [ + { + "bbox": [ + 227, + 179, + 299, + 187 + ], + "spans": [ + { + "bbox": [ + 227, + 179, + 299, + 187 + ], + "type": "text", + "content": "Similar semantic features" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 198, + 482, + 275 + ], + "lines": [ + { + "bbox": [ + 130, + 198, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 482, + 275 + ], + "type": "text", + "content": "Fig. 3: An example of positive samples from different features. Given an anchor pixel (marked by a blue circle in the first image) on the wall-like cabinet, semantic features like DINO often correlate both the cabinet and the ceiling (as shown by the red area in the second image). Meanwhile, geometric features such as (noisy) monocular normals can not distinguish between parallel planes that are not connected (the third image). By combining semantic and geometric features, we can obtain positive samples of the anchor pixel with better geometric consistency, thus, high mutual correlation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 115, + 391, + 178 + ], + "blocks": [ + { + "bbox": [ + 309, + 115, + 391, + 178 + ], + "lines": [ + { + "bbox": [ + 309, + 115, + 391, + 178 + ], + "spans": [ + { + "bbox": [ + 309, + 115, + 391, + 178 + ], + "type": "image", + "image_path": "0713b265af24b7ebbaeed91bc4e2fbefb2b1432c4c0011904bb28b4c17d96bbc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 179, + 388, + 187 + ], + "lines": [ + { + "bbox": [ + 312, + 179, + 388, + 187 + ], + "spans": [ + { + "bbox": [ + 312, + 179, + 388, + 187 + ], + "type": "text", + "content": "Similar geometric features" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 394, + 115, + 479, + 178 + ], + "blocks": [ + { + "bbox": [ + 394, + 115, + 479, + 178 + ], + "lines": [ + { + "bbox": [ + 394, + 115, + 479, + 178 + ], + "spans": [ + { + "bbox": [ + 394, + 115, + 479, + 178 + ], + "type": "image", + "image_path": "c5e4fb51e95f325dfde32be6d991727a9691233107efa75288aab21b7ee2d042.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 179, + 462, + 187 + ], + "lines": [ + { + "bbox": [ + 411, + 179, + 462, + 187 + ], + "spans": [ + { + "bbox": [ + 411, + 179, + 462, + 187 + ], + "type": "text", + "content": "Combined (Ours)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 297, + 482, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 346 + ], + "type": "text", + "content": "positively with the cosine similarity of their gradients to the network parameters. This conclusion is similar to that of Eq. 5. As a result, we can now identify the pixels that are of high correlations, and increase the aggregated mutual information during the NeRF training process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 354, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 461 + ], + "type": "text", + "content": "Extracting correlated scene regions from image pixels. To extract scene regions that need to be shaped to have high mutual information, we present a simple yet effective approach with off-the-shelf visual features. Given the input images, we extract pixel-wise semantic and geometric features to identify correlated scene regions. Following previous research [34], we apply DINO [4] as the semantic feature extractor to obtain a high-level understanding of the scene regions. Meanwhile, as we shape mutual information on top of the surface normal, we employ a monocular normal estimator [7] as the geometric feature extractor for a detailed low-level perception of the 3D points in the scene." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 581 + ], + "type": "text", + "content": "DINO features can group different parts of an object, like the seat and back of a chair, which share similar semantics. However, the surface normal vectors from these parts have limited correlation, thus should not be constrained. Conversely, using only geometric features can wrongly suggest high correlation between unrelated areas, like the seat and the floor. As these areas aren't closely connected in 3D space, there is no reason to have better surface reconstruction when enhancing their mutual information. Similar phenomena on wall-like structures can be observed in Fig. 3, where we show that by combining semantic and geometric features together we can better identify scene regions that are spatially connected and have similar normal directions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "More explicitly, we combine semantic and geometric features with an intersection operator. It is simple yet effective. The process of identifying correlated scene regions (in normal) is implemented as positive pixel pair selection from input images. Given a randomly selected pixel from an image, we first extract similar pixels from neighboring images using semantic and geometric features, respectively. This extraction is implemented by applying two thresholds, " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta_{S}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta_{G}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", to filter the cosine similarity for semantic and geometric features. We then" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 272, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 272, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 272, + 102 + ], + "type": "text", + "content": "X. Wang & S. Dong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 189 + ], + "type": "text", + "content": "feed the intersection of the top similar pixels from each feature type into NeRF's training process as positive samples, allowing us to shape them with high mutual information. In order to leverage the contrastive learning framework to encourage mutual information, we randomly select pixels not involved in the aforementioned process to serve as negative samples (independent pixels that are not correlated to the anchor pixel)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 197, + 479, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 197, + 479, + 222 + ], + "spans": [ + { + "bbox": [ + 131, + 197, + 479, + 222 + ], + "type": "text", + "content": "Loss functions for training. The aforementioned contrastive learning process is performed with a tailored InfoNCE loss [22]:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 222, + 227, + 481, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 227, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 222, + 227, + 481, + 255 + ], + "type": "interline_equation", + "content": "L _ {M} = - \\log \\frac {\\sum \\exp (| | \\cos (\\partial \\mathbb {N} _ {i} , \\partial \\mathbb {N} _ {i +}) | |)}{\\sum_ {j} \\exp (| | \\cos (\\partial \\mathbb {N} _ {i} , \\partial \\mathbb {N} _ {j}) | |)}, \\tag {11}", + "image_path": "7b8631b9e44835ba14fcf0a9b93cb7a8e9bdfaf1e3bfc3e1e0531b88e9a7e763.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 259, + 480, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 259, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 131, + 259, + 480, + 354 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 259, + 480, + 354 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 259, + 480, + 354 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 259, + 480, + 354 + ], + "type": "inline_equation", + "content": "i+" + }, + { + "bbox": [ + 131, + 259, + 480, + 354 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 259, + 480, + 354 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 131, + 259, + 480, + 354 + ], + "type": "text", + "content": " denote the anchor, positive, and negative pixels, respectively. This loss function effectively enhances mutual information (according to Eq. 10) among correlated pixels, while maintaining low mutual information among independent pixels. Finally, to train the NeRF backbone and retain the characteristics of the SDF, we adopt the basic photometric loss (Eq. 2) and the eikonal loss (Eq. 4). The latter two are derived from single images, while the former can be calculated from either single images or a set of neighboring images. The full loss function is defined as the weighted sum of three loss terms:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 249, + 361, + 480, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 361, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 249, + 361, + 480, + 373 + ], + "type": "interline_equation", + "content": "L = L _ {C} + \\lambda_ {E} L _ {E} + \\lambda_ {M} L _ {M}. \\tag {12}", + "image_path": "2a88d576e3a734f9ce41b33dd278b8ffddb9ec902b49a8ae87573fb73f804fca.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 380, + 436, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 380, + 436, + 392 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 436, + 392 + ], + "type": "text", + "content": "With the above, all network parameters " + }, + { + "bbox": [ + 131, + 380, + 436, + 392 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 380, + 436, + 392 + ], + "type": "text", + "content": " can be trained end-to-end." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 407, + 230, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 407, + 230, + 422 + ], + "spans": [ + { + "bbox": [ + 131, + 407, + 230, + 422 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 430, + 480, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 480, + 525 + ], + "type": "text", + "content": "In this section, we evaluate the effectiveness of our method through experiments conducted on multiple datasets with various baselines. First, we introduce the two datasets used to conduct the experiments in Sec. 4.1. Next, we detail the implementation of the proposed method in Sec. 4.2, where we adapt a spectrum of state-of-the-art models for validating the effectiveness of the geometric shaping. Then, we report the comparisons between the baselines and our models in Sec. 4.3. Lastly, we conduct a series of ablation studies in Sec. 4.4, to verify the necessity of each component in our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 540, + 205, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 540, + 205, + 552 + ], + "spans": [ + { + "bbox": [ + 131, + 540, + 205, + 552 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "We evaluate the proposed method using two public datasets from the literature: ScanNet++ [38] and the Replica [26] dataset. For the ScanNet++ dataset, we choose 8 representative scenes, while for the Replica dataset, we select 4 representative scenes. The scenes from both datasets represent a variety of indoor settings, including offices, bedrooms, and so on. As our focus is on scene reconstruction from sparse views, we uniformly sample 50-85 images from the official (and Replica from Semantic-NeRF [42]) scanning sequences for each scene to use as input in our evaluation. For detailed statistics of each scene, please refer to our supplementary material." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 101 + ], + "type": "text", + "content": "InfoNorm" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Evaluation metrics. Following previous works [13, 30, 40], we evaluate the quality of surface reconstruction using two metrics: Chamfer distance and F-score. Chamfer distance is a measure that evaluates the difference between reconstructed and ground-truth points sets. F-score is a statistical metric that integrates precision and recall, providing a more comprehensive understanding of the reconstruction quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 205, + 280, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 280, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 280, + 217 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 223, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 482, + 462 + ], + "type": "text", + "content": "As the proposed geometric shaping can be easily integrated into any network similar to NeRF with SDF representations, we select several state-of-the-art models as our baselines and apply the proposed to them for demonstrating the effectiveness of our approach. We first choose NeuS [30] and VolSDF [37], the two representative pioneers for NeRFs with SDFs. We also apply our method to the following works: GeoNeuS [8], I²-SDF [44], NeuRIS [29], MonoSDF [40], and Neuralangelo [13]. In addition to the basic loss functions for NeRF and SDF, GeoNeuS incorporates patch similarity from neighboring views and surface points derived from structure-from-motion. I²-SDF integrates the rendering equation with VolSDF. We disregard its normal and depth loss terms. NeuRIS utilizes monocular normal estimation as a supervision on the geometry, and discards the low-quality normal information that lacks sufficient confidence through multi-view patching. MonoSDF employs both depth and normal estimation, for which we set the decay of the normal and depth at 30k iterations to prevent the method from degeneration. Neuralangelo utilizes a multi-resolution hash grid and numerical gradient. We set the hash encoding dictionary size to 20, and the encoding dimension to 4. For all the base models mentioned above, we enhance them by incorporating our proposed mutual information loss during training. The resulting methods are respectively labeled as " + }, + { + "bbox": [ + 130, + 223, + 482, + 462 + ], + "type": "inline_equation", + "content": "\\mathrm{NeuS^{+}}" + }, + { + "bbox": [ + 130, + 223, + 482, + 462 + ], + "type": "text", + "content": ", VolSDF+, GeoNeuS+, and so on." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 479, + 227, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 227, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 227, + 491 + ], + "type": "text", + "content": "4.3 Comparisons" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 569 + ], + "type": "text", + "content": "Results on the ScanNet++ dataset. The quantitative comparisons are reported in Tab. 1. We also visualize the surface reconstruction results in Fig. 4 for qualitative comparisons. We can observe that by applying the proposed mutual information shaping on the geometry, the surface reconstruction quality can be improved across all baselines, as measured by mean Chamfer distance, F-score, and the visuals." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "text", + "content": "NeuS and VolSDF tend to produce rough surfaces and excessively smoothed boundaries. However, with our geometry-aware mutual information, the surfaces become smoother, while the boundaries are more crispy. Compared to GeoNeuS, a method that uses patch-similarity among neighboring views and requires sufficient textures, our method provides significant improvements. Applying the proposed technique to " + }, + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathrm{I}^2" + }, + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "text", + "content": "-SDF can also enhance its reconstruction quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "As for more recent works, we find that NeuRIS is somewhat sensitive to the quality of normal estimation. It can either be affected by noise in the normal" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "text", + "content": "X. Wang & S. Dong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 190, + 479, + 322 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 1: Quantitative results on the ScanNet++ dataset. Each cell includes the original baseline number in black, followed by the improvement using our mutual information. Positive and negative improvements are marked in green and orange, respectively. The last column shows the average improvement over all scenes, validating the effectiveness of the proposed shaping for better geometric reconstruction (despite a small potion of noise on a few scenes)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 190, + 479, + 322 + ], + "lines": [ + { + "bbox": [ + 135, + 190, + 479, + 322 + ], + "spans": [ + { + "bbox": [ + 135, + 190, + 479, + 322 + ], + "type": "table", + "html": "
0a7c0a186ee27b6456a09460a08de0abMean
NeuS [30]+0.08-0.040.05-0.030.03+0.010.06-0.040.01-0.000.24-0.140.04-0.030.02-0.000.07-0.03
VolSDF [37]+0.90-0.830.31-0.290.04+0.020.10-0.030.03-0.010.14+0.030.39-0.380.04-0.000.24-0.19
GeoNeuS [8]+0.03-0.021.03-1.020.28-0.270.05-0.030.02-0.010.22-0.200.01-0.010.26-0.250.24-0.23
I2-SDF [44]+0.27-0.240.22-0.030.04-0.010.47-0.450.03-0.020.13+0.050.03+0.020.30-0.070.19-0.09
NeuRIS [29]+0.01-0.000.01-0.000.03+0.020.07-0.040.04-0.030.41-0.210.03-0.010.03-0.000.08-0.03
MonoSDF [40]+0.01-0.000.01+0.020.02-0.010.02-0.000.01-0.000.05-0.020.01-0.000.01-0.000.02-0.00
Neuralangelo [13]+0.17-0.010.84-0.660.52-0.330.05-0.010.04+0.024.76-2.850.33+0.050.12+0.030.85-0.48
NeuS [30]+0.49+0.060.48+0.100.67+0.000.58+0.160.81-0.010.58-0.020.67+0.170.75+0.090.63+0.07
VolSDF [37]+0.45+0.350.47+0.270.67+0.000.55+0.050.71+0.060.52+0.180.48+0.350.70+0.020.57+0.16
GeoNeuS [8]+0.69+0.200.47+0.440.52+0.340.70+0.060.87+0.030.68+0.230.75+0.140.37+0.510.63+0.24
I2-SDF [44]+0.33+0.520.28+0.000.58+0.150.41+0.430.75+0.100.65+0.070.65+0.070.40+0.250.51+0.20
NeuRIS [29]+0.72+0.030.68+0.010.65+0.010.69+0.060.85+0.010.24+0.170.64+0.130.69+0.090.65+0.06
MonoSDF [40]+0.92-0.010.88-0.010.81+0.040.79+0.060.89+0.000.88+0.050.91+0.000.89-0.010.87+0.01
Neuralangelo [13]+0.42+0.050.14+0.300.35+0.120.63+0.050.71-0.010.17+0.220.29+0.060.51-0.010.40+0.10
", + "image_path": "6f95acba95b601c8f059e1f4c6c843acaf0fecec1f6f13a14a91ffef6e2753f6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 178, + 358, + 436, + 495 + ], + "blocks": [ + { + "bbox": [ + 196, + 338, + 416, + 349 + ], + "lines": [ + { + "bbox": [ + 196, + 338, + 416, + 349 + ], + "spans": [ + { + "bbox": [ + 196, + 338, + 416, + 349 + ], + "type": "text", + "content": "Table 2: Quantitative results on the Replica dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 178, + 358, + 436, + 495 + ], + "lines": [ + { + "bbox": [ + 178, + 358, + 436, + 495 + ], + "spans": [ + { + "bbox": [ + 178, + 358, + 436, + 495 + ], + "type": "table", + "html": "
office0office1room0room1Mean
NeuS [30]+0.005+0.0010.023-0.0120.068-0.0380.130-0.0870.057-0.035
VolSDF [37]+0.004+0.0060.016-0.0040.080-0.0710.034-0.0250.033-0.023
GeoNeuS [8]+0.023-0.0210.014-0.0090.040-0.0350.002-0.0000.020-0.017
I2-SDF [44]+0.032-0.0090.056-0.033--0.044-0.021
NeuRIS [29]+0.014-0.0063.975-3.9630.025-0.0170.015+0.0021.007-0.996
MonoSDF [40]+0.003-0.0000.005-0.0010.004-0.0000.004-0.0010.004-0.001
Neuralangelo [13]+0.024-0.0180.030-0.0190.146-0.0120.082-0.0470.070-0.024
NeuS [30]+0.92+0.000.80+0.010.78-0.030.72+0.080.80+0.02
VolSDF [37]+0.94-0.040.78+0.010.78+0.160.82+0.070.83+0.05
GeoNeuS [8]+0.89+0.080.91+0.040.90+0.090.98+0.010.920+0.05
I2-SDF [44]+0.72+0.090.56+0.20--0.64+0.14
NeuRIS [29]+0.79+0.060.01+0.680.89+0.050.77+0.030.61+0.20
MonoSDF [40]+0.96+0.000.90+0.020.98+0.000.95+0.020.95+0.01
Neuralangelo [13]+0.89+0.070.49+0.360.65+0.150.71+0.180.69+0.19
", + "image_path": "1ab590405cf2035dc2662145551ba9a56b42fa584ba3629c28d927463d808a7c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": "or completely disregard the normal information due to low confidence. Particularly, when images are sparse, the regions covered by fewer viewpoints tend to lose supervision. Our mutual information scheme does not depend on the absolute values of the surface normal, but on the similarity among these values. This makes it more robust to noise and less likely to discard low-quality but useful normal information. Consequently, our method enhances the reconstruction quality of NeuRIS. MonoSDF achieves the best overall performance among the baselines because it uses both depth and normal estimations as geometry supervision. However, our method still offers improvements. Notably, it produces smoother planes and clearer boundaries. This indicates that not only the supervision of geometric cues is important, but also the correlation between the geometric entities matters." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "InfoNorm" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 478, + 577 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 478, + 577 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 478, + 577 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 478, + 577 + ], + "type": "image", + "image_path": "85fcbe70ebb8b419ecd0f5736e2a72107774160c5cf350c90b8e8223928071d9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 585, + 482, + 630 + ], + "lines": [ + { + "bbox": [ + 130, + 585, + 482, + 630 + ], + "spans": [ + { + "bbox": [ + 130, + 585, + 482, + 630 + ], + "type": "text", + "content": "Fig. 4: Visual results on the ScanNet++ dataset. Each row shows a comparison with a different baseline, from top to bottom, NeuS " + }, + { + "bbox": [ + 130, + 585, + 482, + 630 + ], + "type": "inline_equation", + "content": "[30]^+" + }, + { + "bbox": [ + 130, + 585, + 482, + 630 + ], + "type": "text", + "content": ", VolSDF " + }, + { + "bbox": [ + 130, + 585, + 482, + 630 + ], + "type": "inline_equation", + "content": "[37]^+" + }, + { + "bbox": [ + 130, + 585, + 482, + 630 + ], + "type": "text", + "content": ", GeoNeuS " + }, + { + "bbox": [ + 130, + 585, + 482, + 630 + ], + "type": "inline_equation", + "content": "[8]^+" + }, + { + "bbox": [ + 130, + 585, + 482, + 630 + ], + "type": "text", + "content": ", I²-SDF " + }, + { + "bbox": [ + 130, + 585, + 482, + 630 + ], + "type": "inline_equation", + "content": "[44]^+" + }, + { + "bbox": [ + 130, + 585, + 482, + 630 + ], + "type": "text", + "content": ", NeuRIS " + }, + { + "bbox": [ + 130, + 585, + 482, + 630 + ], + "type": "inline_equation", + "content": "[29]^+" + }, + { + "bbox": [ + 130, + 585, + 482, + 630 + ], + "type": "text", + "content": ", MonoSDF " + }, + { + "bbox": [ + 130, + 585, + 482, + 630 + ], + "type": "inline_equation", + "content": "[40]^+" + }, + { + "bbox": [ + 130, + 585, + 482, + 630 + ], + "type": "text", + "content": ", and Neuralangelo " + }, + { + "bbox": [ + 130, + 585, + 482, + 630 + ], + "type": "inline_equation", + "content": "[13]^+" + }, + { + "bbox": [ + 130, + 585, + 482, + 630 + ], + "type": "text", + "content": ". Red boxes are overlaid to help the contrast." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "text", + "content": "X. Wang & S. Dong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 115, + 216, + 177 + ], + "blocks": [ + { + "bbox": [ + 135, + 115, + 216, + 177 + ], + "lines": [ + { + "bbox": [ + 135, + 115, + 216, + 177 + ], + "spans": [ + { + "bbox": [ + 135, + 115, + 216, + 177 + ], + "type": "image", + "image_path": "c3b66e13e0a8712a4ccb6d1eeba48ff3a6730ef1a34898ee902321c2a9046255.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 178, + 200, + 186 + ], + "lines": [ + { + "bbox": [ + 151, + 178, + 200, + 186 + ], + "spans": [ + { + "bbox": [ + 151, + 178, + 200, + 186 + ], + "type": "text", + "content": "Reference image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 115, + 304, + 177 + ], + "blocks": [ + { + "bbox": [ + 223, + 115, + 304, + 177 + ], + "lines": [ + { + "bbox": [ + 223, + 115, + 304, + 177 + ], + "spans": [ + { + "bbox": [ + 223, + 115, + 304, + 177 + ], + "type": "image", + "image_path": "12b7055978d076432f27e37e27bcf5105d4b8a6caddf0c26c3ce3c4e5a6b914d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 178, + 271, + 186 + ], + "lines": [ + { + "bbox": [ + 255, + 178, + 271, + 186 + ], + "spans": [ + { + "bbox": [ + 255, + 178, + 271, + 186 + ], + "type": "text", + "content": "NeuS" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 196, + 480, + 219 + ], + "lines": [ + { + "bbox": [ + 131, + 196, + 480, + 219 + ], + "spans": [ + { + "bbox": [ + 131, + 196, + 480, + 219 + ], + "type": "text", + "content": "Fig. 5: Reconstruction results after 10K iterations in the training process with " + }, + { + "bbox": [ + 131, + 196, + 480, + 219 + ], + "type": "inline_equation", + "content": "(\\mathrm{NeuS}^{+})" + }, + { + "bbox": [ + 131, + 196, + 480, + 219 + ], + "type": "text", + "content": " and without (NeuS) the proposed geometric shaping technique." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 115, + 391, + 177 + ], + "blocks": [ + { + "bbox": [ + 310, + 115, + 391, + 177 + ], + "lines": [ + { + "bbox": [ + 310, + 115, + 391, + 177 + ], + "spans": [ + { + "bbox": [ + 310, + 115, + 391, + 177 + ], + "type": "image", + "image_path": "1934989b75e764148aeab0fdca00ee343919dd3f6fea03c82d22737ae5d6ba6b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 178, + 370, + 186 + ], + "lines": [ + { + "bbox": [ + 331, + 178, + 370, + 186 + ], + "spans": [ + { + "bbox": [ + 331, + 178, + 370, + 186 + ], + "type": "text", + "content": "Neu" + }, + { + "bbox": [ + 331, + 178, + 370, + 186 + ], + "type": "inline_equation", + "content": "^{+}" + }, + { + "bbox": [ + 331, + 178, + 370, + 186 + ], + "type": "text", + "content": " (Ours)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 398, + 115, + 478, + 177 + ], + "blocks": [ + { + "bbox": [ + 398, + 115, + 478, + 177 + ], + "lines": [ + { + "bbox": [ + 398, + 115, + 478, + 177 + ], + "spans": [ + { + "bbox": [ + 398, + 115, + 478, + 177 + ], + "type": "image", + "image_path": "3d686b691ed9650ea1c7a36938a87db1535ca26e1281f93b6ada1fecee39a744.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 178, + 457, + 186 + ], + "lines": [ + { + "bbox": [ + 419, + 178, + 457, + 186 + ], + "spans": [ + { + "bbox": [ + 419, + 178, + 457, + 186 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 243, + 479, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 479, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 479, + 291 + ], + "type": "text", + "content": "Last but not least, the experiments on Neuralangelo demonstrate the potential of our method in working with hash grid encoding and numerical gradient. This highlights our method's versatility and ease of use, which can be broadly applied in various reconstruction architectures." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 303, + 480, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 480, + 375 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 480, + 375 + ], + "type": "text", + "content": "Results on the Replica dataset. In Tab. 2 we report the numerical results on the Replica dataset, with respect to NeuS [30], VolSDF [37], GeoNeuS [8], I²-SDF [44], NeuRIS [29], MonoSDF [40], and Neuralangelo [13]. As observed, the proposed geometric shaping technique still provides positive improvements when incorporated into these baselines, aligning with the conclusion drawn from the ScanNet++ data. Qualitative examples can also be found in Fig. 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 393, + 205, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 393, + 205, + 406 + ], + "spans": [ + { + "bbox": [ + 132, + 393, + 205, + 406 + ], + "type": "text", + "content": "4.4 Analyses" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 414, + 263, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 263, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 263, + 521 + ], + "type": "text", + "content": "Ablation studies. We select three representative scenes from the ScanNet++ dataset and conduct ablation studies to evaluate the effectiveness of our method designs. We conduct ablation studies using two top-performing baselines: NeuRIS and MonoSDF." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 522, + 263, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 263, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 263, + 629 + ], + "type": "text", + "content": "The results are reported in Tab. 3. Instead of using the surface normal formula to inject mutual information, we first perform a basic mutual information shaping separately in the color and SDF values. These two variations are labeled as base" + }, + { + "bbox": [ + 130, + 522, + 263, + 629 + ], + "type": "inline_equation", + "content": "^{+}" + }, + { + "bbox": [ + 130, + 522, + 263, + 629 + ], + "type": "text", + "content": " (color)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "and base" + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "^{+}" + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": " (SDF), respectively. Interestingly, we find that, even though color is not directly related to geometry, by shaping it, we can still obtain improvements on the geometry in some cases. Moreover, shaping directly on the SDF values" + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 272, + 483, + 479, + 601 + ], + "blocks": [ + { + "bbox": [ + 270, + 449, + 480, + 481 + ], + "lines": [ + { + "bbox": [ + 270, + 449, + 480, + 481 + ], + "spans": [ + { + "bbox": [ + 270, + 449, + 480, + 481 + ], + "type": "text", + "content": "Table 3: Ablation studies on three scenes from ScanNet++. The best and second-best results are marked in bold and underlined." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 272, + 483, + 479, + 601 + ], + "lines": [ + { + "bbox": [ + 272, + 483, + 479, + 601 + ], + "spans": [ + { + "bbox": [ + 272, + 483, + 479, + 601 + ], + "type": "table", + "html": "
Chamfer (m)↓F-score ↑
6ee27b649460Mean6ee27b649460Mean
NeuRIS [30]0.0290.0700.4050.1680.650.690.240.53
NeuRIS+ (color)0.0230.0380.3150.1250.720.680.340.58
NeuRIS+ (SDF)0.1270.0530.3600.1800.470.610.300.46
NeuRIS+ w/o normal0.0360.0250.3380.1330.660.750.320.58
NeuRIS+ w/o DINO0.0610.0330.3810.1580.630.720.320.56
NeuRIS+ (Full)0.0440.0330.1980.0920.660.760.410.61
MonoSDF [40]0.0200.0160.0460.0280.810.790.880.83
MonoSDF+ (color)0.0240.0190.0330.0250.790.780.870.81
MonoSDF+ (SDF)0.5880.0293.4431.3530.380.650.210.42
MonoSDF+ w/o normal0.1280.0230.0660.0720.680.730.800.74
MonoSDF+ w/o DINO0.0640.0160.0220.0340.780.830.930.85
MonoSDF+ (Full)0.0140.0200.0230.0190.850.850.930.87
", + "image_path": "2199c25153cdaabb203df08322dabc11494f48b939ed42552af69e28972dbd91.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "InfoNorm" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "severely degrade the reconstruction quality, which evidences that the shaping on SDF values could incur a competition with the goal to faithfully learn the geometry within the neural radiance field. We further examine the effectiveness of combining semantic and geometric features to identify geometrically highly correlated regions. As demonstrated in Fig. 3, using only one type of the features can lead to incorrect correlations. Tab. 3 also shows that eliminating any one of the features can decrease the performance, confirming the effectiveness of our positive-negative sample selection mechanism leveraging the multimodal features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 298, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 298, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 298, + 378 + ], + "type": "text", + "content": "Time and memory consumption. Our method provides benefits for 3D reconstruction with limited additional effort. Specifically, it only increases the optimization time and memory usage during the training process, without affecting the model architecture and final storage. The time and memory consumption are reported in Tab. 4. It is observed that the overhead on the training time highly depends on the architecture used in each baseline. For example, the train" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 380, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 452 + ], + "type": "text", + "content": "ing time increase on NeuRIS is marginal while more significant on VolSDF. These studies provide comprehensive information on the trade-offs between training time and performance gain. However, we do find that baselines with the proposed shaping technique can achieve a better performance (e.g., early geometries from NeuS and " + }, + { + "bbox": [ + 130, + 380, + 482, + 452 + ], + "type": "inline_equation", + "content": "\\mathrm{NeuS^{+}}" + }, + { + "bbox": [ + 130, + 380, + 482, + 452 + ], + "type": "text", + "content": " shown in Fig. 5) at the same training steps due to the effective regularization by mutual information." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 308, + 270, + 479, + 352 + ], + "blocks": [ + { + "bbox": [ + 306, + 259, + 479, + 270 + ], + "lines": [ + { + "bbox": [ + 306, + 259, + 479, + 270 + ], + "spans": [ + { + "bbox": [ + 306, + 259, + 479, + 270 + ], + "type": "text", + "content": "Table 4: Time and memory consumption." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 270, + 479, + 352 + ], + "lines": [ + { + "bbox": [ + 308, + 270, + 479, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 270, + 479, + 352 + ], + "type": "table", + "html": "
Training Time (h)VRAM (GB)
NeuS [30]+3.00+0.377.89+1.68
VolSDF [37]+3.18+1.9616.36+4.68
GeoNeuS [8]+3.26+0.165.72+3.56
I2-SDF [44]+3.47+2.8410.08+8.09
NeuRIS [29]+3.70+0.2011.73+4.12
MonoSDF [40]+2.65+1.3212.83+9.42
Neuralangelo [13]+3.12+2.0616.44+1.00
", + "image_path": "c427bf7809e6fe8cede9ddd17ce7837b9eceb758239fe1b10de076dcb76eabca.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 469, + 220, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 469, + 220, + 482 + ], + "spans": [ + { + "bbox": [ + 131, + 469, + 220, + 482 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 495, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 578 + ], + "type": "text", + "content": "We explore the usage of the second-order geometric correlations as a regularization and introduce a mutual information shaping technique of the normals for better surface reconstruction quality. With a simple yet effective correlated point selection mechanism that does not depend on accurate geometry information, we can perform a contrastive learning that maximizes the detected correlations. In return, we can observe a boost in the geometric quality of the reconstructed surfaces in state-of-the-art neural scene representations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 578, + 482, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 482, + 662 + ], + "type": "text", + "content": "Extensive evaluations on various baselines and datasets show the effectiveness of our method, which improves 3D reconstruction quality without resorting to a more sophisticated surface modeling. The overhead caused by the shaping of the normals varies with different neural architectures of the scene representation, e.g., some are marginal, and some are significant, which we deem as a limitation of our current pipeline. We believe that an acceleration in computation can be achieved using a coarse-to-fine strategy, which we leave as a future work." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "text", + "content": "X. Wang & S. Dong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Acknowledgments. We thank the anonymous reviewers for their valuable feedback. This work is supported by the Early Career Scheme of the Research Grants Council (grant # 27207224), the HKU-100 Award, and in part by NSF China (No. 62172363). Siyan Dong would also like to thank the support from HKU Musketeers Foundation Institute of Data Science for the Postdoctoral Research Fellowship." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 208, + 199, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 208, + 199, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 208, + 199, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 234, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 234, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 480, + 267 + ], + "type": "text", + "content": "1. Bae, G., Budvytis, I., Cipolla, R.: Estimating and exploiting the aleatoric uncertainty in surface normal estimation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13137-13146 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 268, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 268, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 268, + 480, + 311 + ], + "type": "text", + "content": "2. Bao, C., Zhang, Y., Yang, B., Fan, T., Yang, Z., Bao, H., Zhang, G., Cui, Z.: Sine: Semantic-driven image-based nerf editing with prior-guided editing field. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20919-20929 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 312, + 480, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 480, + 344 + ], + "type": "text", + "content": "3. Cai, B., Huang, J., Jia, R., Lv, C., Fu, H.: Neuda: Neural deformable anchor for high-fidelity implicit surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8476-8485 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "type": "text", + "content": "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 379, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 480, + 411 + ], + "type": "text", + "content": "5. Chen, Y., Wu, Q., Zheng, C., Cham, T.J., Cai, J.: Sem2nerf: Converting single-view semantic masks to neural radiance fields. In: European Conference on Computer Vision. pp. 730-748. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 412, + 480, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 480, + 444 + ], + "type": "text", + "content": "6. Curless, B., Levoy, M.: A volumetric method for building complex models from range images. In: Proceedings of the 23rd annual conference on Computer graphics and interactive techniques. pp. 303-312 (1996)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 445, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 480, + 487 + ], + "type": "text", + "content": "7. Do, T., Vuong, K., Roumeliotis, S.I., Park, H.S.: Surface normal estimation of tilted images via spatial rectifier. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 265-280. Springer (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 489, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 489, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 489, + 480, + 520 + ], + "type": "text", + "content": "8. Fu, Q., Xu, Q., Ong, Y.S., Tao, W.: Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems 35, 3403-3416 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 522, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 480, + 544 + ], + "type": "text", + "content": "9. Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 544, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 544, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 134, + 544, + 480, + 586 + ], + "type": "text", + "content": "10. Guo, H., Peng, S., Lin, H., Wang, Q., Zhang, G., Bao, H., Zhou, X.: Neural 3d scene reconstruction with the manhattan-world assumption. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5511-5520 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 588, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 588, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 134, + 588, + 480, + 620 + ], + "type": "text", + "content": "11. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: Lerf: Language embedded radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19729-19739 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "type": "text", + "content": "12. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "InfoNorm" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "13. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "type": "text", + "content": "14. Liu, F., Zhang, C., Zheng, Y., Duan, Y.: Semantic ray: Learning a generalizable semantic field with cross-reprojection attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17386-17396 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 196, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 481, + 239 + ], + "type": "text", + "content": "15. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13-23 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 240, + 481, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 284 + ], + "type": "text", + "content": "16. Liu, Z., Milano, F., Frey, J., Siegwart, R., Blum, H., Cadena, C.: Unsupervised continual semantic adaptation through neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3031-3040 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "type": "text", + "content": "17. Long, X., Lin, C., Wang, P., Komura, T., Wang, W.: Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In: European Conference on Computer Vision. pp. 210-227. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "text", + "content": "18. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. In: Seminal graphics: pioneering efforts that shaped the field, pp. 347-353 (1998)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 353, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 481, + 386 + ], + "type": "text", + "content": "19. Meng, X., Chen, W., Yang, B.: Neat: Learning neural implicit surfaces with arbitrary topologies from multi-view images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 248-258 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 387, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 481, + 430 + ], + "type": "text", + "content": "20. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12663-12673 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 431, + 481, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 481, + 463 + ], + "type": "text", + "content": "21. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 464, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 486 + ], + "type": "text", + "content": "22. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 487, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 531 + ], + "type": "text", + "content": "23. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 533, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 481, + 576 + ], + "type": "text", + "content": "24. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 577, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 598 + ], + "type": "text", + "content": "25. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 599, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 666 + ], + "type": "text", + "content": "26. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., Clarkson, A., Yan, M., Budge, B., Yan, Y., Pan, X., Yon, J., Zou, Y., Leon, K., Carter, N., Briales, J., Gillingham, T., Mueggler, E., Pesqueira, L., Savva, M., Batra, D., Strasdat, H.M., Nardi, R.D., Goesele, M., Lovegrove, S., Newcombe, R.: The replica dataset: A digital replica of indoor spaces (2019)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 271, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 271, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 271, + 101 + ], + "type": "text", + "content": "X. Wang & S. Dong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "27. Tertikas, K., Despoina, P., Pan, B., Park, J.J., Uy, M.A., Emiris, I., Avrithis, Y., Guibas, L.: Partnerf: Generating part-aware editable 3d shapes without 3d supervision. arXiv preprint arXiv:2303.09554 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 184 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 184 + ], + "type": "text", + "content": "28. Wang, F., Galliani, S., Vogel, C., Speciale, P., Pollefeys, M.: Patchmatchnet: Learned multi-view patchmatch stereo. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14194-14203 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 185, + 482, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 185, + 482, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 185, + 482, + 217 + ], + "type": "text", + "content": "29. Wang, J., Wang, P., Long, X., Theobalt, C., Komura, T., Liu, L., Wang, W.: Neuris: Neural reconstruction of indoor scenes using normal priors. In: European Conference on Computer Vision. pp. 139-155. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 218, + 482, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 218, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 482, + 251 + ], + "type": "text", + "content": "30. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 251, + 482, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 295 + ], + "type": "text", + "content": "31. Weder, S., Garcia-Hernando, G., Monszpart, A., Pollefeys, M., Brostow, G.J., Firman, M., Vicente, S.: Removing objects from neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16528-16538 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 296, + 482, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 329 + ], + "type": "text", + "content": "32. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In: European conference on computer vision. pp. 106-122. Springer (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 330, + 482, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 373 + ], + "type": "text", + "content": "33. Xu, L., Xiangli, Y., Peng, S., Pan, X., Zhao, N., Theobalt, C., Dai, B., Lin, D.: Grid-guided neural radiance fields for large urban scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8296-8306 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 374, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 407 + ], + "type": "text", + "content": "34. Xu, X., Yang, Y., Mo, K., Pan, B., Yi, L., Guibas, L.: Jacobinerf: Nerf shaping with mutual information gradients. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16498-16507 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 407, + 482, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 441 + ], + "type": "text", + "content": "35. Yan, Z., Li, C., Lee, G.H.: Nerf-ds: Neural radiance fields for dynamic specular objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8285-8295 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 441, + 482, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 475 + ], + "type": "text", + "content": "36. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multi-view stereo. In: Proceedings of the European conference on computer vision (ECCV). pp. 767-783 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 475, + 482, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 497 + ], + "type": "text", + "content": "37. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 498, + 482, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 530 + ], + "type": "text", + "content": "38. Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: Proceedings of the International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 531, + 482, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 564 + ], + "type": "text", + "content": "39. Yu, H., Julin, J., Milacski, Z.A., Niinuma, K., Jeni, L.A.: Dylin: Making light field networks dynamic. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12397-12406 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 565, + 482, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 565, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 565, + 482, + 597 + ], + "type": "text", + "content": "40. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems 35, 25018-25032 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 599, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 482, + 632 + ], + "type": "text", + "content": "41. Zhang, Y., Yang, G., Guibas, L., Yang, Y.: Infogaussian: Structure-aware dynamic gaussians through lightweight information shaping. arXiv preprint arXiv:2406.05897 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "42. Zhi, S., Laidlow, T., Leutenegger, S., Davison, A.J.: In-place scene labelling and understanding with implicit scene representation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15838-15847 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "InfoNorm" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 194 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "43. Zhu, B., Yang, Y., Wang, X., Zheng, Y., Guibas, L.: Vdn-nerf: Resolving shape- radiance ambiguity via view-dependence normalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 35-45 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "text", + "content": "44. Zhu, J., Huo, Y., Ye, Q., Luan, F., Li, J., Xi, D., Wang, L., Tang, R., Hua, W., Bao, H., Wang, R.: I" + }, + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "text", + "content": "-sdf: Intrinsic indoor scene reconstruction and editing via raytracing in neural sdfs. In: CVPR (2023)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "text", + "content": "X. Wang & S. Dong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Information Bottleneck Based Data Correction in Continual Learning/6c4e72fe-4072-432b-96c2-9a488daff608_content_list.json b/2024/Information Bottleneck Based Data Correction in Continual Learning/6c4e72fe-4072-432b-96c2-9a488daff608_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3f20132bd00a4c08d0757425390d162ffcf008f3 --- /dev/null +++ b/2024/Information Bottleneck Based Data Correction in Continual Learning/6c4e72fe-4072-432b-96c2-9a488daff608_content_list.json @@ -0,0 +1,1756 @@ +[ + { + "type": "text", + "text": "Information Bottleneck Based Data Correction in Continual Learning", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shuai Chen $^{1,2}$ , Mingyi Zhang $^{1}$ , Junge Zhang $^{1*}$ , and Kaiqi Huang $^{1,2,3*}$", + "bbox": [ + 225, + 210, + 777, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ CRISE, Institute of Automation, Chinese Academy of Sciences \n $^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences \n $^{3}$ CAS Center for Excellence in Brain Science and Intelligence Technology", + "bbox": [ + 243, + 239, + 756, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Continual Learning (CL) requires model to retain previously learned knowledge while learning new tasks. Recently, experience replay-based methods have made significant progress in addressing this challenge. These methods primarily select data from old tasks and store them in a buffer. When learning new task, they train the model using both the current and buffered data. However, the limited number of old data can lead to the model being influenced by new tasks. The repeated replaying of buffer data and the gradual discarding of old task data (un-sampled data) also result in a biased estimation of the model towards the old tasks, causing overfitting issues. All these factors can affect the CL performance. Therefore, we propose a data correction algorithm based on the Information Bottleneck (IBCL) to enhance the performance of the replay-based CL system. This algorithm comprises two components: the Information Bottleneck Task Agnostic Constraints (IBTA), which encourages the buffer data to learn task-relevant features related to the old tasks, thereby reducing the impact of new tasks. The Information Bottleneck Unsampled Data Surrogate (IBDS), which models the information of the unsampled data in the old tasks to alleviate data bias. Our method can be flexibly combined with most existing experience replay methods. We have verified the effectiveness of our method through a series of experiments, demonstrating its potential for improving the performance of CL algorithms.", + "bbox": [ + 261, + 311, + 743, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Continual learning $\\cdot$ Incremental learning $\\cdot$ Experience replay", + "bbox": [ + 259, + 630, + 740, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 681, + 375, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modern deep learning algorithms have achieved remarkable performance in a diverse range of tasks [48-50]. However, one of their significant limitations is the inability to retain knowledge from previously learned tasks when receiving training on new tasks, known as catastrophic forgetting [18, 25]. Addressing this issue has been a subject of extensive research in the field of continual learning (CL) [12, 36, 52, 53]. Among the proposed methods, the experience replay approaches have received significant attention [11, 13, 31, 54]. These methods aim to", + "bbox": [ + 212, + 710, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 218, + 825, + 385, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8085fa4ba7d2799a4275bbe20286b80cd1bdbd100f9f00759e6f7bcdafeea710.jpg", + "image_caption": [ + "Fig. 1: During the process of CL, when the old task is sampled, there is a consequent discarding of some unsampled data. This leads to bias in the model's estimation of the distribution of the old task." + ], + "image_footnote": [], + "bbox": [ + 232, + 146, + 767, + 261 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "overcome catastrophic forgetting by selectively storing a subset of data from old tasks in the buffer and then combining buffer data with new task data for joint training. Various replay methods have been investigated from different perspectives to address the challenge of sample selection and balancing the contributions of new and old samples during the replay process [6,14,37].", + "bbox": [ + 212, + 353, + 782, + 430 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the progress made in reducing catastrophic forgetting through replay methods, certain limitations still exist. A major issue is that the memory of the buffer is too small, and the model cannot store enough data to restore previously learned knowledge [4, 21, 35]. This will give rise to two significant problems. Firstly, the scarcity of old data will result in an imbalance between the number of new and old tasks, leading to a disruption of knowledge from the old task during the learning process of the new task. Secondly, in the CL process, each time the old task data is sampled, some unsampled data will be discarded, introducing bias in the model's estimation of the old task distribution. Moreover, the repeated replaying of buffer data can contribute to overfitting of this specific data. All of these factors can significantly impact the CL performance. Currently, there are some methods to address the above issues, such as selecting more representative samples [20], increasing the diversity of old samples through data expansion [3], and improving the generalization ability of the model [5].", + "bbox": [ + 212, + 433, + 787, + 645 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This work aims to tackle the aforementioned challenges from a different perspective. Regarding the first issue, we propose encouraging the data in the buffer to acquire task agnostic features related to the old task, thereby mitigating the impact of the new task. For the second issue, we focus on an important yet under-investigated question: How can unsampled data effectively contribute to learning new tasks? In Figure 1, we illustrate the cause of model overfitting, which arises because replay-based methods often utilize only a subset of data from old task, leading to biased results. When training on new tasks, we design the objective function as follows:", + "bbox": [ + 212, + 647, + 787, + 784 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {n e w}} = \\frac {1}{N _ {\\text {b u f f e r}} + N _ {\\text {n e w}}} \\sum_ {i = 1} ^ {N _ {n}} \\ell_ {i} (\\hat {Y}, Y). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 801, + 785, + 844 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Chen et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\ell_i(\\hat{Y}, Y)$ represents the loss function that takes the true outcome $Y$ and predicts outcome $\\hat{Y}$ as inputs. $N_{buffer}$ and $N_{new}$ denote the numbers of buffer and new task data, respectively. $N_n$ is the sum of these two quantities. However, this loss function does not provide an unbiased estimate of the loss function for both new and old tasks, as the model gradually loses some data from past tasks during training. The unbiased loss function for new and old tasks should be formulated as follows:", + "bbox": [ + 212, + 145, + 782, + 252 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} [ \\mathcal {L} _ {\\text {n e w}} ] \\neq \\mathcal {L} := \\frac {1}{N _ {\\text {o l d}} + N _ {\\text {n e w}}} \\sum_ {i = 1} ^ {N} \\ell_ {i} (\\hat {Y}, Y). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 266, + 785, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $N_{old}$ and $N_{new}$ represent the numbers of old task and new task data, respectively, and $N$ signifies the sum of these two quantities. To address the bias arising from the discarding of unsampled data from old task, our method aims to estimate the information of the unsampled data using specific strategies. However, unsampled data cannot directly participate in the training of new tasks, we indirectly obtain a surrogate for the information about unsampled data through the relationship between sampled and unsampled data. This information related to the unsampled data serves as a regularization term to constrain the training process of the model, thereby achieving bias correction for the dataset.", + "bbox": [ + 212, + 320, + 784, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on the intuition presented above, we propose an Information Bottleneck based data correction algorithm (IBCL). The information bottleneck method is a technique in information theory used to find the best compromise between accuracy and complexity [41]. It is regarded as the theoretical foundation of deep learning and has been applied in many fields, such as robust or invariant representation learning, causal reasoning, and so on [17]. This theory suggests that the model gradually eliminates irrelevant noise information during the learning process, retaining the most relevant feature information to the task. Borrowing the ability of this method, our algorithm mainly consists of two parts: Information Bottleneck Task Agnostic Constraints (IBTA): During the update process of new tasks, we introduce a constraint on the learning process of the buffer data. This constraint encourages the model to learn task agnostic features that are relevant to the old tasks, effectively preventing interference from the new tasks. Information Bottleneck Unsampled Data Surrogate (IBDS): This module is designed to model the information of the unsampled data. As the unsampled data cannot be directly trained, our method decouples the features of the unsampled data from the sampled data after the completion of learning old tasks. This decoupling allows model to leverage the relationship between the two types of data to estimate the surrogate influence of unsampled data. Consequently, this achieves an indirect implementation of the impact of unsampled data on learning new tasks.", + "bbox": [ + 212, + 458, + 785, + 775 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall, our work makes the following contributions:", + "bbox": [ + 238, + 777, + 617, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We propose a data correction framework based on information bottleneck to mitigate model bias when the number of old task samples is small.", + "bbox": [ + 225, + 809, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "IB Based Data Correction in CL", + "bbox": [ + 511, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- To mitigate the problem of the influence of new tasks on old sample features during the learning process, we propose IBTA to encourage the model to learn task agnostic features.", + "bbox": [ + 223, + 146, + 782, + 190 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- To counter the issue of overfitting caused by the repeated replay of buffer data and gradual discarding of unsampled data, we propose IBDS to incorporate the information from unsampled data to assist in the training of the new model.", + "bbox": [ + 225, + 191, + 784, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Our method can be combined with most replay based methods to improve their performance.", + "bbox": [ + 223, + 252, + 782, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 306, + 387, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Experience Replay Continual Learning", + "text_level": 1, + "bbox": [ + 215, + 340, + 581, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the context of CL, there are three main approaches to retaining knowledge of past tasks: using a fixed-sized buffer for replaying past samples (replay-based approach) [2, 3, 7, 8, 43], regulating model parameter changes through learning (regularization-based approach) [16, 40, 42], or dynamically expanding model architecture as needed (expansion-based approach) [15, 28, 47]. Despite its simplicity, the replay-based approach has shown promising results in CL settings. Experience Replay (ER) [19] is a widely-used technique, which maintains a fixed-sized buffer to store and replay selected data, helping to mitigate the forgetting of previously learned knowledge. Incremental Classifier and Representation Learning (iCaRL) [37] uses the herding strategy to select a few representative samples. Rainbow Memory [3] proposes a diversity sampling strategy based on per-sample classification uncertainty. However, these methods primarily concentrate on how to choose which old data to store in the buffer, while neglecting the unsampled data. We argue that the unsampled data can also be beneficial in mitigating deviation and should be utilized. To this end, we propose a data correction method that employs the information bottleneck theory to model the information of the unsampled data in the old task, thereby enhancing the performance of several state-of-the-art methods.", + "bbox": [ + 212, + 366, + 787, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Information Bottleneck", + "text_level": 1, + "bbox": [ + 215, + 662, + 454, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Information bottleneck (IB) is a classical concept in information theory, which is used to explain the behavior of deep learning [41]. Furthermore, IB has been extensively utilized in diverse applications, including but not limited to, representation learning decoupling [23, 33], recommendation systems [29, 45], and enhancing model robustness [17]. Recently, there have been some other works that apply information theory to CL, such as [27, 46], but there are fewer works based on the IB theory. The fundamental construct of IB is based on mutual information. In a previous work [20], the maximization of mutual information is employed to learn adequate feature expression. In contrast, our proposed method concentrates on modeling the unsampled data in the previous task.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Chen et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ff358228f58fb0907ffd9d15fd833630364b4c1bd2fd93c8aa1a9041b1ec48ae.jpg", + "image_caption": [ + "Fig. 2: During the updating of new tasks, we constrain the learning process of buffer data, encouraging the model to learn task agnostic features related to the old tasks, thus avoiding the interference of new tasks." + ], + "image_footnote": [], + "bbox": [ + 222, + 148, + 782, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 349, + 330, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method consists of two key processes: the learning process of new tasks and the processing process after learning old tasks. We introduce IBTA and IBDS separately.", + "bbox": [ + 212, + 378, + 782, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem Formulation", + "text_level": 1, + "bbox": [ + 215, + 444, + 437, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In CL scenarios, tasks are introduced one after another and handled separately. We consider a total of $T$ tasks. Each task $D_{t} = \\{x_{i}^{t},y_{i}^{t}\\}_{i = 1}^{n_{t}}$ comprises $n_t$ input-target pairs $(x_i^t,y_i^t)$ , which are assumed to be independent and identically distributed (i.i.d.). The aim for each task is to train a model $f_{\\theta}:X\\to Y$ with parameters $\\theta$ , capable of predicting the class label $y$ for a new image $x$ . In CL algorithms that utilize experience replay, the methods typically store some data from previous task data $x_{old}$ in a buffer. This stored data is referred to as sampled data $x_{m}$ , with its features denoted as $z_{m}$ . The data not selected for storage is called unsampled data $x_{u}$ , and its features are represented as $z_{u}$ .", + "bbox": [ + 212, + 467, + 787, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Information Bottleneck Task Agnostic Constraints", + "text_level": 1, + "bbox": [ + 215, + 623, + 676, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the learning of new tasks, the model focuses on the new task due to the limited data available in the buffer. In order to mitigate the impact of new tasks, our method encourages the buffer data to focus on learning the task-specific features, thus minimizing their influence. In order to compress the information in the dataset into the features of the sampled data as much as possible, the mutual information between the features $z_{m}$ of the sampled data and the buffer dataset $x_{m}$ should be minimized; In order to make the feature be able to distinguish well, the feature $z_{m}$ of the sample needs to predict the label $y_{old}$ as accurately as possible. The weighting parameters $\\alpha$ is used to balance the contributions of different components in the overall objective function.", + "bbox": [ + 212, + 646, + 787, + 797 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I B T A} := \\min \\underbrace {\\alpha I \\left(z _ {m} ; x _ {\\mathrm {m}}\\right)} _ {(1)} - \\underbrace {I \\left(z _ {m} ; y _ {\\text {o l d}}\\right)} _ {(2)}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 805, + 785, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "IB Based Data Correction in CL", + "bbox": [ + 511, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the given framework, the expression (1) represents a compression component that characterizes the mutual information between the buffer dataset $x_{m}$ and the sampled data features $z_{m}$ . This component aims to capture the relevance of $z_{m}$ to the old dataset $x_{m}$ . The expression (2) corresponds to an accuracy component that measures the effectiveness of $z_{m}$ in representing the underlying distribution of the dataset. It aims to ensure that the sampled features $z_{m}$ preserve sufficient information about the original data. Next, we will proceed to optimize the compression term $I(z_{m};x_{m})$ . This term can be quantified using the Kullback-Leibler (KL) divergence as follows:", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} I (z _ {m}; x _ {\\mathrm {m}}) = \\mathbb {E} _ {x} \\left[ \\mathrm {K L} (p (z _ {m} | x _ {\\mathrm {m}}) \\parallel p (z _ {m})) \\right] \\\\ = \\mathbb {E} _ {x} \\left[ \\int p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) \\log p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) d z - \\int p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) \\log p \\left(z _ {m}\\right) d z \\right]. \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 292, + 784, + 343 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Nonetheless, due to the computational complexity involved in calculating $p(z_{m}) = \\int p(z_{m}|x_{m})p(x_{m})dx$ , we resort to the utilization of a variational approximation approach. Instead of using $p(z_{m})$ , we adopt a variational distribution denoted as $q(z_{m})$ . Leveraging Gibbs' inequality, we recognize the non-negativity of the Kullback-Leibler (KL) divergence. Thus, we can establish an upper bound for Equation (4),", + "bbox": [ + 214, + 349, + 789, + 441 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} - \\int p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) \\log p \\left(z _ {m}\\right) d z \\leq - \\int p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) \\log q \\left(z _ {m}\\right) d z \\tag {5} \\\\ \\Rightarrow \\operatorname {K L} \\left(p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) \\| p \\left(z _ {m}\\right)\\right) \\leq \\operatorname {K L} \\left(p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) \\| q \\left(z _ {m}\\right)\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 450, + 784, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similar to prior research [1, 29, 34, 44, 45], we can make the assumption that the posterior distribution $p(z_{m} \\mid x_{m})$ follows a Gaussian distribution and set it $\\mathcal{N}(e(x_m),\\mathrm{diag}(\\sigma (x_m)))$ . In this context, $e(x_{m})$ represents the encoded embedding of the variables $x_{m}$ , and $\\mathrm{diag}(\\sigma (x_m))$ indicates the variance in a diagonal matrix form. By employing the reparameterization trick, the embedding $z_{m}$ can be generated through the formula $z_{m} = e(x_{m}) + \\epsilon \\odot \\sigma (x_{m})$ , where $\\epsilon \\sim \\mathcal{N}(0,I)$ . Notably, if we set $\\sigma (x_{m})$ as an all-zero matrix, $z_{m}$ will effectively become a deterministic embedding. Conversely, the prior distribution $q(z_{m})$ is assumed to adhere to a standard Gaussian variational distribution, namely $q(z_{m}) = \\mathcal{N}(0,I)$ . Ultimately, the previously mentioned upper bound can be rephrased as follows:", + "bbox": [ + 212, + 507, + 787, + 657 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {K L} \\left(p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) \\| q \\left(z _ {m}\\right)\\right) = \\| e \\left(x _ {\\mathrm {m}}\\right) \\| _ {2} ^ {2} + \\sum_ {d} \\left(\\sigma_ {d} - \\frac {1}{2} \\log \\sigma_ {d} - 1\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 279, + 666, + 784, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\sigma_d$ corresponds to an element within $\\operatorname{diag}\\{\\sigma(x_m)\\}$ . This signifies that for a deterministic embedding $z_m$ , we can enhance the optimization of this upper bound by directly implementing $\\ell_2$ -norm regularization on the embedding vector $z_m$ . This process is equivalent to refining the compression term $I(z_m; x_m)$ . When considering the mutual information $I(z_m; y_{old})$ within Equation (3), we can express it as $I(z_m; y_{old}) = H(y_{old}) - H(y_{old} \\mid z_m)$ . As $H(y_{old})$ is a positive constant that can be disregarded, the following inequality holds:", + "bbox": [ + 212, + 709, + 787, + 816 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI \\left(z _ {m}; y _ {\\text {o l d}}\\right) \\geq - H \\left(y _ {\\text {o l d}} \\mid z _ {m}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 825, + 784, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Chen et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/65ac90514e9c443546e403e7d5158f0ce4fbb1558b966208a78394fa8412f3ea.jpg", + "image_caption": [ + "Fig. 3: This module is used to model the information of unsampled data. Because unsampled data cannot be trained directly, our method decouples the features of unsampled data and sampled data after the completion of the old task learning, so that it is convenient to estimate the surrogate loss of unsampled data in the new task by using the relationship between them, and indirectly realize the influence of unsampled data on the new task learning." + ], + "image_footnote": [], + "bbox": [ + 222, + 148, + 779, + 266 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This term can be seen as a cross entropy loss function. Therefore, we have implemented the constraint process for learning buffer data. However, the loss function obtained in this way is still biased, as there is missing unsampled data, which means that the contribution to the old task label $y_{old}$ should be $I(z; y) = I(z_m, z_u; y)$ , where $z_u$ is unsampled data feature. During the learning process of the new task, this item cannot be directly calculated. We will calculate it indirectly in the following section.", + "bbox": [ + 212, + 398, + 787, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Information Bottleneck Unsampled Data Surrogate", + "text_level": 1, + "bbox": [ + 214, + 534, + 684, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To estimate the mutual information $I(z; y) = I(z_{m}, z_{u}; y)$ during the learning process of new task, it becomes essential to obtain $z_{u}$ . Given that $x_{u}$ denotes unsampled data, we cannot directly calculate $z_{u}$ from $x_{u}$ when learning new task. This necessitates an indirect estimation of $z_{u}$ from old task model. As shown in Fig.3, during the learning stage of old task, both sampled $x_{m}$ and unsampled $x_{u}$ data are involved in the training process. After training completion, the features corresponding to these two parts $z_{m}$ and $z_{u}$ are tightly coupled. Consequently, a specific operation is needed after training ends to decouple these two features. If we can decouple the features of sampled and unsampled data by optimizing $I(z_{m}; z_{u})$ , we can get $I(z; y) = I(z_{m}; y) + I(z_{u}; y)$ . In this case, the estimate of $I(z_{u}; y)$ can be used as a surrogate for the unsampled data information during learning new task. The feature $z_{m}$ of the sampled data and the feature $z_{u}$ of the unsampled data must maintain as much independence from one another as possible to obtain better discrimination. Drawing on principles from information theory, we are inspired to derive a minimization objective function based on the aforementioned analysis:", + "bbox": [ + 212, + 566, + 787, + 809 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I B D S} := \\min I \\left(z _ {m}; z _ {u}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 824, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "IB Based Data Correction in CL", + "bbox": [ + 511, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "represents a decoupling component, capturing the level of interdependence between $z_{u}$ and $z_{m}$ . It aims to minimize the dependence of $z_{u}$ on $z_{m}$ , ensuring that $z_{u}$ provides additional information beyond what is already captured by $z_{m}$ .", + "bbox": [ + 215, + 146, + 784, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inspired by previous research work [1, 29, 34, 44, 45], utilizing the chain rule of mutual information, we derive the subsequent equation pertaining to the decoupling component:", + "bbox": [ + 212, + 191, + 785, + 238 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nI \\left(z _ {m}; z _ {u}\\right) = I \\left(z _ {m}; y _ {o l d}\\right) - I \\left(z _ {m}; y _ {o l d} \\mid z _ {u}\\right) + I \\left(z _ {m}; z _ {u} \\mid y _ {o l d}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 294, + 250, + 785, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proceed to analyze the expression $I(z_{m};z_{u}\\mid y_{old})$ in greater detail. Given that the distribution of $z_{m}$ exclusively hinges on the variables $x_{m}$ , and the latter is influenced by $y_{old}$ , we establish $H(z_{m}\\mid y_{old},z_{u}) = H(z_{m}\\mid y_{old})$ , where $H(\\cdot \\mid \\cdot)$ signifies conditional entropy. By combining the attributes of mutual information, we arrive at the following relationship:", + "bbox": [ + 214, + 277, + 785, + 353 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} I (z _ {m}; z _ {u} \\mid y _ {\\mathrm {o l d}}) = H (z _ {m} \\mid y _ {\\mathrm {o l d}}) - H (z _ {m} \\mid y _ {\\mathrm {o l d}}, z _ {u}) \\\\ = H \\left(z _ {m} \\mid y _ {\\text {o l d}}\\right) - H \\left(z _ {m} \\mid y _ {\\text {o l d}}\\right) = 0. \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 366, + 784, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By substituting Eq.(9), we have,", + "bbox": [ + 215, + 412, + 452, + 428 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nI \\left(z _ {m}; z _ {u}\\right) = I \\left(z _ {m}; y _ {\\text {o l d}}\\right) - I \\left(z _ {m}; y _ {\\text {o l d}} \\mid z _ {u}\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 440, + 784, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since the term $I(z_{m};y_{old} \\mid z_{u})$ in Eq.(11) is still difficult to be calculated, we simplify it further by using the form of conditional entropy:", + "bbox": [ + 212, + 465, + 785, + 498 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} I \\left(z _ {m}; y _ {\\text {o l d}} \\mid z _ {u}\\right) = H \\left(z _ {m} \\mid z _ {u}\\right) + H \\left(y _ {\\text {o l d}} \\mid z _ {u}\\right) - I \\left(z _ {m}; y _ {\\text {o l d}} \\mid z _ {u}\\right) \\\\ = H \\left(z _ {m} \\mid z _ {u}\\right) - H \\left(z _ {m} \\mid y _ {\\text {o l d}}, z _ {u}\\right) \\tag {12} \\\\ = H \\left(y _ {\\text {o l d}} \\mid z _ {u}\\right) - H \\left(y _ {\\text {o l d}} \\mid z _ {m}, z _ {u}\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 287, + 510, + 784, + 563 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nI \\left(z _ {m}; z _ {u}\\right) = I \\left(z _ {m}; y _ {o l d}\\right) - H \\left(y _ {o l d} \\mid z _ {u}\\right) + H \\left(y _ {o l d} \\mid z _ {m}, z _ {u}\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 302, + 580, + 784, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, combining Eq.(7) and Eq.(13), we can rewrite the objective function $\\mathcal{L}_{IBDS}$ in Eq.(8) as follows,", + "bbox": [ + 215, + 603, + 784, + 635 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I B D S} = - H \\left(y _ {o l d} \\mid z _ {m}\\right) - H \\left(y _ {o l d} \\mid z _ {u}\\right) + H \\left(y _ {o l d} \\mid z _ {m}, z _ {u}\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 297, + 647, + 784, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By optimizing $\\mathcal{L}_{IBDS}$ , we get the decoupled expression $z_{u}$ . Now we can compute $I(z_{u};y)$ , which acts as a surrogate loss to compensate for missing weights due to unsampled data.", + "bbox": [ + 215, + 672, + 785, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During the training process of neural network, the model will encounter the problem of semantic drift [22,24,51], which means that the feature distribution of sampled data will change with the learning of new tasks. This concept highlights a significant challenge: using the fixed features of unsampled data as constraints in the learning of new task becomes problematic. As the model trains on new tasks, the estimation of unsampled data features gradually deviates from their true values, which not only affects the effectiveness of bias correction but also, to some extent, impacts the overall learning process. Just as the \"no free lunch\"", + "bbox": [ + 212, + 719, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Chen et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "theorem advocated in machine learning. Once samples are lost, it becomes difficult to perfectly estimate the distribution of unsampled data, the best one can do is to minimize the model's bias when addressing old tasks. To mitigate this impact, it's essential to consider the characteristics of neural networks in CL [9]. Early in the learning of new tasks, gradients undergo significant changes, and the model frequently encounters considerable biases. A viable strategy involves imposing stricter constraints on the new task learning process during the initial stages of model training. This approach instructs the model in reducing biases when learning new tasks. As the training advances, these constraints are gradually relaxed, permitting the model to better adjust to the learning requirements of new tasks during the slower update phases. To address the processing of unsampled data features, a dynamic decay strategy is employed to balance the learning conflicts between old and new tasks, thereby optimizing the model's overall performance. This method allows the model to discover a more balanced learning point between preserving old knowledge and acquiring new tasks, enhancing its learning efficiency and effectiveness. Specifically, inspired by [30], we propose a dynamic cosine annealing factor $\\eta$ for constraint terms when updating unsampled data.", + "bbox": [ + 212, + 146, + 787, + 417 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\eta = \\left(\\frac {1}{2} (1 + \\cos \\left(\\frac {e _ {\\text {c u r}}}{e _ {\\text {m a x}}}\\right) \\pi)\\right) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 417, + 785, + 450 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $e_{cur}$ and $e_{max}$ represent current and maximum number of epochs.", + "bbox": [ + 214, + 455, + 736, + 470 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Algorithm", + "text_level": 1, + "bbox": [ + 214, + 493, + 346, + 508 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "By combining above two parts, we can get the total loss function as:", + "bbox": [ + 214, + 518, + 709, + 535 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {C I B} = \\alpha I (z _ {m}; x _ {\\mathrm {m}}) - I (z; y _ {\\mathrm {o l d}}) \\\\ = \\alpha I \\left(z _ {m}; x _ {\\mathrm {m}}\\right) - \\beta I \\left(z _ {u}; y\\right) - \\gamma I \\left(z _ {m}; y\\right) \\tag {16} \\\\ \\leq \\alpha \\| e \\left(x _ {\\mathrm {m}}\\right) \\| _ {2} ^ {2} + \\beta \\cdot \\eta H \\left(y _ {\\text {o l d}} \\mid z _ {u}\\right) + \\gamma H \\left(y _ {\\text {o l d}} \\mid z _ {m}\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 547, + 784, + 602 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The weighting parameters $\\alpha$ , $\\beta$ , $\\gamma$ are used to balance the contributions of different components in the overall objective function.", + "bbox": [ + 212, + 612, + 785, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The final objective function comprises three terms: term (1) represents a regularization constraint on the input data, term (2) represents surrogate loss to reflects the influence of the unsampled data; term (3) denotes the cross-entropy between $z_{m}$ and $y_{old}$ , this can be seen as the cross-entropy loss function. The aim is to make the learned feature expression fully predict the final label.", + "bbox": [ + 212, + 643, + 787, + 718 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Discussion This approach comprises two primary modules: the IBTA module facilitates the learning of new tasks, whereas the IBDS module is utilized post-learning of old tasks to derive features from unsampled data. These features are then integrated as constraints within the training of new tasks, thereby shaping the overall optimization loss function. Our method can be combined with most replay-based strategies. Taking the naive ER method as an example, it randomly selects some samples from the old task and stores them in a buffer, which are trained together with the samples from the new task. Upon finalizing", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "IB Based Data Correction in CL", + "bbox": [ + 511, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/48ef1e08468ed897951a3379377e38bca32a21e4cea047a7fdbf52ad2f8d4678.jpg", + "table_caption": [ + "Table 1: Comparison of the average accuracy (the higher the better) of four split datasets with different buffer sizes." + ], + "table_footnote": [], + "table_body": "
MethodSplit CIFAR-10Split CIFAR-100Split ImageNet-100Split miniImageNet
Joint92.3873.2980.2353.55
FT19.679.298.689.52
Buffer size10020050020050020002005002000100020005000
ER36.3944.7957.7414.3519.6636.7613.6318.3734.258.3716.4924.17
+IBCL45.4351.9163.1223.9828.3242.0122.7228.9343.1114.7922.7227.69
ER-ACE53.9063.4170.5326.2836.4848.4124.2337.1249.5517.9522.6027.92
+IBCL63.8570.9774.8232.1240.9451.8930.7241.8152.6224.7727.6331.02
DER++57.6564.8872.7025.1137.1352.0826.5043.6558.0518.0223.4430.43
+IBCL66.4172.5276.6133.8244.2154.8932.6346.5759.9227.7231.4135.79
X-DER59.2965.1968.1035.3444.6254.4433.2146.7255.2325.2426.3829.91
+IBCL67.9173.8274.1443.8648.7255.9143.2549.6256.5929.9030.2132.66
", + "bbox": [ + 222, + 184, + 787, + 359 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "the learning of old task, the sampled data is processed through the old model to yield $z_{m}$ , and both sampled and unsampled data are processed to extract coupled features $z$ . Subsequently, by optimizing IBDS, features $z_{u}$ for each class of unsampled data are obtained. During the training of the new task, IBTA and $z_{u}$ are combined to optimize the total loss function for learning the new task. Detailed pseudocode of this procedure is provided in the appendix.", + "bbox": [ + 212, + 387, + 784, + 479 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 501, + 375, + 518 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we will present the datasets and evaluation metrics used in the experiment, and describe the implementation details. Then we describe the experimental results of our method on multiple standard benchmarks. Finally, we conduct an ablation study to illustrate the effectiveness of each module of our method.", + "bbox": [ + 212, + 532, + 784, + 608 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Dataset", + "text_level": 1, + "bbox": [ + 215, + 631, + 328, + 645 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We validate our results on four distinct datasets: CIFAR-10, CIFAR-100, miniImageNet, and ImageNet-100. CIFAR-10 and CIFAR-100 each consist of 50,000 training samples and 10,000 testing samples, with 10 and 100 classes respectively. miniImageNet and ImageNet-100 are subsets of ImageNet, both containing 100 classes. miniImageNet includes 500 images per category for training, while ImageNet-100 includes 1300 images per category for training.", + "bbox": [ + 212, + 656, + 784, + 748 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Details", + "text_level": 1, + "bbox": [ + 215, + 768, + 320, + 782 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To better compare with other methods, we follow the settings of [46] and [5]. For the CIFAR-10/100 and ImageNet-100 datasets, we adopt ResNet-18, and for the miniImageNet dataset, we adopt EfficientNet-B2. For CIFAR-10/100, we", + "bbox": [ + 212, + 794, + 784, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Chen et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/039be141bd4bba34a17cb65083d22e77e756b69b1f0cd790cafd3ca0ed886846.jpg", + "table_caption": [ + "Table 2: Comparison of our method with other regularization methods on three datasets, we use two typical methods DER++ and ER-ACE as benchmarks for comparison." + ], + "table_footnote": [], + "table_body": "
MethodSplit CIFAR-10Split CIFAR-100Split miniImageNet
Buffer size1002005002005002000100020005000
ER-ACE53.9063.4170.5326.2836.4848.4117.9522.6027.92
+ sSGD56.2664.7371.4528.0739.5949.7018.1122.4324.12
+ oEwC52.3661.0968.7024.9335.0645.5919.0424.3229.46
+ oLAP52.7663.1970.3226.4236.5847.6618.3423.1928.77
+ OCM57.1864.6570.8628.1837.7449.0320.3224.3228.57
+ LiDER56.0865.3271.7527.9438.4350.3219.6924.1330.00
+ DualHSIC60.5268.0873.7829.0838.9450.5522.3325.4130.12
+ Our63.8570.9774.8232.1240.9451.8924.7727.6331.02
DER++57.6564.8872.7025.1137.1352.0818.0223.4430.43
+ sSGD55.8164.4472.0524.7638.4850.7416.3119.2924.24
+ oEwC55.7863.0271.6424.5135.2251.5318.8724.5331.91
+ oLAP54.8662.5471.3823.2634.4850.8018.9125.0232.78
+ OCM59.2565.8173.5327.4638.9452.2520.9324.7531.16
+ LiDER58.4366.0273.3927.3239.2553.2721.5828.3335.04
+ DualHSIC64.9870.2875.9431.4641.8653.5324.7829.3734.98
+ Our66.4172.5276.6133.8244.2154.8927.7231.4135.79
", + "bbox": [ + 228, + 198, + 776, + 478 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "adopt a training epoch of 50 and set the batch size to 32 and 64, respectively. For MiniImageNet and ImageNet-100, we adopt training epochs of 80 and 120, with batch size set to 128 and 64, respectively. In order to better compare the experimental results, we adopt the experimental results reported in [46] and [5], or the hyperparametric design in the original article.", + "bbox": [ + 212, + 508, + 787, + 585 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Evaluation Metrics", + "text_level": 1, + "bbox": [ + 215, + 614, + 419, + 630 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We mainly evaluate on class incremental learning scenario. In this setup, each sequential task introduces a distinct set of classes that do not overlap with those from previous tasks. The challenge lies in training a classifier that can effectively classify all classes encountered up to the current stage, without the need for explicit task ids. We evaluate the final performance of the method by average accuracy.", + "bbox": [ + 212, + 648, + 787, + 741 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nA C C = \\frac {1}{T} \\sum_ {i = 1} ^ {T} R _ {T, i}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 758, + 785, + 800 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $R_{T,i}$ denotes the test accuracy on task $i$ after the model has finished task $T$ .", + "bbox": [ + 214, + 809, + 785, + 838 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "IB Based Data Correction in CL", + "bbox": [ + 511, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/2d38f47d97a1ae860d0b7e74d5b321446d60c15efb8ed5508636400a0615466f.jpg", + "table_caption": [ + "Table 3: Abalation study of the average accuracy (the higher the better) of split ImageNet-100 and split CIFAR-100 with different buffer sizes." + ], + "table_footnote": [], + "table_body": "
MethodSplit CIFAR-100Split ImageNet-100
Buffer size20050020002005002000
ER-ACE26.2836.4848.4124.2337.1249.55
+ IBTA27.3637.4949.1225.6337.9250.12
+ IBDS30.4439.5250.9129.8340.7151.85
+IBCL32.1240.9451.8930.7241.8152.62
DER++25.1137.1352.0826.5043.6558.05
+IBTA27.8838.7453.9627.3945.7259.33
+ IBDS31.5342.5954.3329.5846.1559.70
+IBCL33.8244.2154.8932.6346.5759.92
", + "bbox": [ + 218, + 184, + 784, + 353 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Baseline Models", + "text_level": 1, + "bbox": [ + 215, + 377, + 397, + 391 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conduct experiments on multiple state-of-the-art experience replay CL models: ER [10], DER++ [8], X-DER [6], and ER-ACE [9]. In addition to the original loss function based on replay, our method can be considered as a new regularization method. Therefore, we also compare our method with existing regularization-based techniques, including sSGD [32], oEWC [39], oLAP [38], and more recent SOTA methods, OCM [20], LiDER [5] and DualHSIC [46].", + "bbox": [ + 212, + 398, + 785, + 491 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Results", + "text_level": 1, + "bbox": [ + 215, + 507, + 323, + 520 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The comparison results are reported in Table 1. For the method of listing several baselines, our proposed method has improved the model's performance in various datasets and buffer sizes. In addition, we have the following main observations: 1) There has been a significant improvement in basic methods such as ER and DER++, as these two methods adopt more basic methods for sampling. Our method can reduce the bias caused by the sampling process to a certain extent. 2) Our methods have also shown some effectiveness on X-DER and ER-ACE, which take additional measures to alleviate sample imbalance and improve sample utilization, resulting in improved model performance. Our method further enhances the model's capabilities on top of this.", + "bbox": [ + 212, + 527, + 785, + 679 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To further demonstrate the effectiveness of our method, we compare our method with other regularization methods, and it can be found that we achieve the best performance, recent methods such as OCM, LiDER and DualHSIC achieve significant progress compared to baseline when the number of buffer is small, and our method further improves the performance when the buffer data is small by considering modeling for the unsampled data.", + "bbox": [ + 212, + 680, + 785, + 771 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 787, + 388, + 803 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In order to understand the effect of each of the items in the loss function on the performance of the algorithm, an ablation study has been carried out and the", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Chen et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e17b5e35974228cf5ec5776a2af0dda49e6c658d92ad831dc2a12a503c651496.jpg", + "image_caption": [ + "Fig. 4: This figure shows the CKA matrix based on the ER algorithm, FT and incorporating our algorithm. It can be observed from it that due to our modeling of unsampled data, the model shows more information about the old task." + ], + "image_footnote": [], + "bbox": [ + 222, + 148, + 419, + 313 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6c4cc5c7089bf47269767d347ffd7406f8ca2b259b93520de11a4d1a7be87f2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 148, + 599, + 313 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/305984db984788dc1f90a3b3b41fd0744b949396a47aee300cf03cb7baf59ca2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 148, + 779, + 313 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "results are shown in Table 3. It is observed that in most of the cases, increasing any of the components has an impact on the performance of the algorithm. However, increasing the IBDS gives a large performance improvement, which indicates that it plays a vital role in the performance of the algorithm. This component represents the loss of substitution of unsampled data, which is necessary for training on new tasks using unsampled data features. Therefore, if this aspect is not well learned, it may negatively affect the subsequent performance of the algorithm. IBTA also has an impact on the performance of the algorithm, the involves regularization constraints on the input features in order to make the model learn enough robust features.", + "bbox": [ + 212, + 404, + 787, + 555 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Characteristic Analysis", + "text_level": 1, + "bbox": [ + 215, + 585, + 477, + 603 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We analyze various characteristics of the model to gain insight into the factors contributing to the performance enhancements achieved by our approach.", + "bbox": [ + 212, + 625, + 785, + 657 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.1 CKA Matrix Analysis", + "text_level": 1, + "bbox": [ + 214, + 686, + 444, + 700 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We use Central Kernel Alignment (CKA) [26] in Fig.4, which is an effective tool to measure the similarity of network representations and evaluate the relationships between these backbones. We send the same batch of instances to the ER (2000 buffer size) algorithm model and fine-tuning based model respectively to obtain corresponding feature maps, thus intuitively demonstrating the similarity of algorithm feature maps. Our algorithm can be used in conjunction with the experience replay method to model unsampled data. When the sampled data is 0, i.e. fine-tuning, our method can also get some benefits.", + "bbox": [ + 212, + 718, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "IB Based Data Correction in CL", + "bbox": [ + 511, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f776ee0a368bf83264a3019df57278fd23a062b9eca311ea385184ae50334934.jpg", + "image_caption": [ + "Initialization" + ], + "image_footnote": [], + "bbox": [ + 253, + 148, + 416, + 250 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9deb011b3e4c1988a3809b8798e4c7015966b38b5352829e6520b0359a26812f.jpg", + "image_caption": [ + "Epoch $= 20$", + "Fig. 5: This image uses TSNE to display the changes in the characteristics of sampled and unsampled data with the number of training epochs of the model. We have shown one class of situation, with a sample data size of 200 and an unsampled data size of 1100." + ], + "image_footnote": [], + "bbox": [ + 421, + 148, + 584, + 250 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e3e5fdfec78084223105586c8257c1a5fb7c74e63c28ad3d3144c282a36632b5.jpg", + "image_caption": [ + "Epoch $= 50$" + ], + "image_footnote": [], + "bbox": [ + 589, + 148, + 751, + 250 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.2 Feature Decoupling Analysis", + "text_level": 1, + "bbox": [ + 214, + 369, + 498, + 386 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A significant advance of our method is the decoupling of sampled and unsampled data features. We demonstrate this in Fig.5, where we use TSNE technology to visualize the changes in these two features with the increase of network training epochs. We test our method, the ER algorithm, on 200 sampled data and 1100 unsampled data from the ImageNet-100 dataset. As the network training epochs increase, we can observe that the two sets of features begin to separate from their initial coupling, which indicates that our method has achieved decoupling of the two data.", + "bbox": [ + 212, + 396, + 787, + 516 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 540, + 359, + 556 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In conclusion, our proposed method is designed to tackle the issue of data bias in continual learning through the application of experience replay. Drawing inspiration from information bottleneck theory, we steer the incorporation of unsampled data information, thereby bolstering the model's capacity to preserve past knowledge while accommodating novel tasks. Our conducted experiments offer empirical validation of the advantages conferred by our approach. While we have presented a method for computing unsampled data, it's important to acknowledge the existence of several alternative strategies for estimating these. This opens up exciting possibilities for continued exploration and innovation in this field.", + "bbox": [ + 212, + 571, + 787, + 723 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 747, + 403, + 763 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported by National Science and Technology Major Project, Grant No.2022ZD0116403, the Youth Innovation Promotion Association CAS and Strategic Priority Research Program of Chinese Academy of Sciences XDA27010300.", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Chen et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alemi, A.A., Fischer, I., Dillon, J.V., Murphy, K.: Deep variational information bottleneck. arXiv (2016)", + "2. Aljundi, R., Lin, M., Goujaud, B., Bengio, Y.: Gradient based sample selection for online continual learning. In: Adv. Neural Inform. Process. Syst. (2019)", + "3. Bang, J., Kim, H., Yoo, Y., Ha, J.W., Choi, J.: Rainbow memory: Continual learning with a memory of diverse samples. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)", + "4. Biesialska, M., Biesialska, K., Costa-Jussa, M.R.: Continual lifelong learning in natural language processing: A survey. arXiv (2020)", + "5. Bonicelli, L., Boschini, M., Porrello, A., Spampinato, C., Calderara, S.: On the effectiveness of lipschitz-driven rehearsal in continual learning. In: Adv. Neural Inform. Process. Syst. (2022)", + "6. Boschini, M., Bonicelli, L., Buzzega, P., Porrello, A., Calderara, S.: Class incremental continual learning into the extended der-verse. IEEE Trans. Pattern Anal. Mach. Intell. (2022)", + "7. Boschini, M., Buzzega, P., Bonicelli, L., Porrello, A., Calderara, S.: Continual semisupervised learning through contrastive interpolation consistency. Pattern Recognition Letters 162, 9-14 (2022)", + "8. Buzzega, P., Boschini, M., Porrello, A., Abati, D., Calderara, S.: Dark experience for general continual learning: a strong, simple baseline. In: Adv. Neural Inform. Process. Syst. (2020)", + "9. Caccia, L., Aljundi, R., Asadi, N., Tuytelaars, T., Pineau, J., Belilovsky, E.: New insights on reducing abrupt representation change in online continual learning. In: Int. Conf. Learn. Represent. (2022)", + "0. Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H.S., Ranzato, M.: Continual learning with tiny episodic memories. Arxiv (2019)", + "1. Chen, S., Zhang, M., Zhang, J., Huang, K.: Exemplar-based continual learning via contrastive learning. IEEE Transactions on Artificial Intelligence 5(7), 3313-3324 (2024)", + "2. Cossu, A., Carta, A., Lomonaco, V., Bacciu, D.: Continual learning for recurrent neural networks: an empirical evaluation. Neural Networks (2021)", + "3. Delange, M., Aljundi, R., Masana, M., Parisot, S., Jia, X., Leonardis, A., Slabaugh, G., Tuytelaars, T.: A continual learning survey: Defying forgetting in classification tasks. IEEE Trans. Pattern Anal. Mach. Intell. (2021)", + "4. Douillard, A., Cord, M., Ollion, C., Robert, T., Valle, E.: PODNet: Pooled Outputs Distillation for Small-Tasks Incremental Learning. In: Eur. Conf. Comput. Vis. (2020)", + "5. Douillard, A., Ram, A., Couairon, G., Cord, M.: Dytox: Transformers for continual learning with dynamic token expansion. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)", + "6. Ebrahimi, S., Elhoseiny, M., Darrell, T., Rohrbach, M.: Uncertainty-guided continual learning with bayesian neural networks. In: Int. Conf. Learn. Represent. (2020)", + "7. Federici, M., Dutta, A., Forre, P., Kushman, N., Akata, Z.: Learning robust representations via multi-view information bottleneck. In: Int. Conf. Learn. Represent. (2020)", + "8. French, R.M.: Catastrophic forgetting in connectionist networks. Trends in cognitive sciences (1999)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "IB Based Data Correction in CL", + "bbox": [ + 511, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Goodfellow, I.J., Mirza, M., Xiao, D., Courville, A., Bengio, Y.: An empirical investigation of catastrophic forgetting in gradient-based neural networks. arXiv (2015)", + "20. Guo, Y., Liu, B., Zhao, D.: Online continual learning through mutual information maximization. In: Int. Conf. Mach. Learn. (2022)", + "21. Hadsell, R., Rao, D., Rusu, A.A., Pascanu, R.: Embracing change: Continual learning in deep neural networks. Trends in cognitive sciences (2020)", + "22. Iscen, A., Zhang, J., Lazebnik, S., Schmid, C.: Memory-efficient incremental learning through feature adaptation. In: Eur. Conf. Comput. Vis. (2020)", + "23. Jaiswal, A., Brekelmans, R., Moyer, D., Steeg, G.V., AbdAlmageed, W., Natarajan, P.: Discovery and separation of features for invariant representation learning. arXiv (2019)", + "24. Jie, S., Deng, Z.H., Li, Z.: Alleviating representational shift for continual fin-tuning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)", + "25. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proc. of the national academy of sciences (2017)", + "26. Kornblith, S., Norouzi, M., Lee, H., Hinton, G.E.: Similarity of neural network representations revisited. In: Int. Conf. Mach. Learn. (2019)", + "27. Li, D., Wang, T., Chen, J., Ren, Q., Kawaguchi, K., Zeng, Z.: Towards continual learning desiderata via hsic-bottleneck orthogonalization and equiangular embedding. In: AAAI (2024)", + "28. Li, X., Zhou, Y., Wu, T., Socher, R., Xiong, C.: Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In: Int. Conf. Mach. Learn. (2019)", + "29. Liu, D., Cheng, P., Zhu, H., Dong, Z., He, X., Pan, W., Ming, Z.: Mitigating confounding bias in recommendation via information bottleneck. In: Proceedings of the 15th ACM Conference on Recommender Systems (2021)", + "30. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv (2016)", + "31. Masana, M., Liu, X., Twardowski, B., Menta, M., Bagdanov, A.D., van de Weijer, J.: Class-incremental learning: Survey and performance evaluation on image classification. IEEE Trans. Pattern Anal. Mach. Intell. (2023)", + "32. Mirzadeh, S.I., Farajtabar, M., Pascanu, R., Ghasemzadeh, H.: Understanding the role of training regimes in continual learning. Arxiv (2020)", + "33. Moyer, D., Gao, S., Brekelmans, R., Steeg, G.V., Galstyan, A.: Evading the adversary in invariant representation. CoRR (2018)", + "34. Pan, Z., Niu, L., Zhang, J., Zhang, L.: Disentangled information bottleneck. In: AAAI (2021)", + "35. Parisi, G.I., Kemker, R., Part, J.L., Kanan, C., Wermter, S.: Continual lifelong learning with neural networks: A review. Neural Networks (2019)", + "36. Qu, H., Rahmani, H., Xu, L., Williams, B., Liu, J.: Recent advances of continual learning in computer vision: An overview. arXiv (2021)", + "37. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017)", + "38. Ritter, H., Botev, A., Barber, D.: Online structured laplace approximations for overcoming catastrophic forgetting. In: Adv. Neural Inform. Process. Syst. (2018)", + "39. Schuster, M., Paliwal, K.K.: Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing 45(11), 2673-2681 (1997)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Chen et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "40. Swaroop, S., Nguyen, C.V., Bui, T.D., Turner, R.E.: Improving and understanding variational continual learning. ArXiv (2019)", + "41. Tishby, N., Pereira, F.C., Bialek, W.: The information bottleneck method (2000)", + "42. Titsias, M.K., Schwarz, J., Matthews, A.G.d.G., Pascanu, R., Teh, Y.W.: Functional regularisation for continual learning with gaussian processes. In: Int. Conf. Learn. Represent. (2019)", + "43. Von Oswald, J., Zhao, D., Kobayashi, S., Schug, S., Caccia, M., Zucchet, N., Sacramento, J.: Learning where to learn: Gradient sparsity in meta and continual learning. In: Adv. Neural Inform. Process. Syst. (2021)", + "44. Wang, Y., Rudner, T.G., Wilson, A.G.: Visual explanations of image-text representations via multi-modal information bottleneck attribution. In: Adv. Neural Inform. Process. Syst. (2024)", + "45. Wang, Z., Chen, X., Wen, R., Huang, S.L., Kuruoglu, E.E., Zheng, Y.: Information theoretic counterfactual learning from missing-not-at-random feedback. In: Adv. Neural Inform. Process. Syst. (2020)", + "46. Wang, Z., Zhan, Z., Gong, Y., Shao, Y., Ioannidis, S., Wang, Y., Dy, J.: Dualhsic: Hsic-bottleneck and alignment for continual learning. In: Int. Conf. Mach. Learn. (2023)", + "47. Wortsman, M., Ramanujan, V., Liu, R., Kembhavi, A., Rastegari, M., Yosinski, J., Farhadi, A.: Supermasks in superposition. In: Adv. Neural Inform. Process. Syst. (2020)", + "48. Xu, P., Zhang, J., Huang, K.: Exploration via joint policy diversity for sparse-reward multi-agent tasks. In: IJCAI (2023)", + "49. Xu, P., Zhang, J., Huang, K.: Population-based diverse exploration for sparse-reward multi-agent tasks. In: IJCAI (2024)", + "50. Xu, P., Zhang, J., Yin, Q., Yu, C., Yang, Y., Huang, K.: Subspace-aware exploration for sparse-reward multi-agent tasks. In: AAAI (2023)", + "51. Yu, L., Twardowski, B., Liu, X., Herranz, L., Wang, K., Cheng, Y., Jui, S., Weijer, J.v.d.: Semantic drift compensation for class-incremental learning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)", + "52. Zhang, X., Yi, J., Tao, J., Wang, C., Zhang, C.Y.: Do you remember? overcoming catastrophic forgetting for fake audio detection. In: Int. Conf. Mach. Learn. (2023)", + "53. Zhang, X., Yi, J., Wang, C., Zhang, C.Y., Zeng, S., Tao, J.: What to remember: Self-adaptive continual learning for audio deepfake detection. In: AAAI (2024)", + "54. Zhou, D.W., Wang, Q.W., Qi, Z.H., Ye, H.J., Zhan, D.C., Liu, Z.: Deep class-incremental learning: A survey. arXiv (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 646 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "IB Based Data Correction in CL", + "bbox": [ + 511, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Information Bottleneck Based Data Correction in Continual Learning/6c4e72fe-4072-432b-96c2-9a488daff608_model.json b/2024/Information Bottleneck Based Data Correction in Continual Learning/6c4e72fe-4072-432b-96c2-9a488daff608_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f0117499b1a1bc65727485a2e51f4eabffc0dd09 --- /dev/null +++ b/2024/Information Bottleneck Based Data Correction in Continual Learning/6c4e72fe-4072-432b-96c2-9a488daff608_model.json @@ -0,0 +1,2357 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.219, + 0.141, + 0.784, + 0.187 + ], + "angle": 0, + "content": "Information Bottleneck Based Data Correction in Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.212, + 0.778, + 0.229 + ], + "angle": 0, + "content": "Shuai Chen\\(^{1,2}\\), Mingyi Zhang\\(^{1}\\), Junge Zhang\\(^{1*}\\), and Kaiqi Huang\\(^{1,2,3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.24, + 0.758, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\) CRISE, Institute of Automation, Chinese Academy of Sciences \n\\(^{2}\\) School of Artificial Intelligence, University of Chinese Academy of Sciences \n\\(^{3}\\) CAS Center for Excellence in Brain Science and Intelligence Technology" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.313, + 0.744, + 0.619 + ], + "angle": 0, + "content": "Abstract. Continual Learning (CL) requires model to retain previously learned knowledge while learning new tasks. Recently, experience replay-based methods have made significant progress in addressing this challenge. These methods primarily select data from old tasks and store them in a buffer. When learning new task, they train the model using both the current and buffered data. However, the limited number of old data can lead to the model being influenced by new tasks. The repeated replaying of buffer data and the gradual discarding of old task data (un-sampled data) also result in a biased estimation of the model towards the old tasks, causing overfitting issues. All these factors can affect the CL performance. Therefore, we propose a data correction algorithm based on the Information Bottleneck (IBCL) to enhance the performance of the replay-based CL system. This algorithm comprises two components: the Information Bottleneck Task Agnostic Constraints (IBTA), which encourages the buffer data to learn task-relevant features related to the old tasks, thereby reducing the impact of new tasks. The Information Bottleneck Unsampled Data Surrogate (IBDS), which models the information of the unsampled data in the old tasks to alleviate data bias. Our method can be flexibly combined with most existing experience replay methods. We have verified the effectiveness of our method through a series of experiments, demonstrating its potential for improving the performance of CL algorithms." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.631, + 0.741, + 0.659 + ], + "angle": 0, + "content": "Keywords: Continual learning \\(\\cdot\\) Incremental learning \\(\\cdot\\) Experience replay" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.682, + 0.377, + 0.699 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Modern deep learning algorithms have achieved remarkable performance in a diverse range of tasks [48-50]. However, one of their significant limitations is the inability to retain knowledge from previously learned tasks when receiving training on new tasks, known as catastrophic forgetting [18, 25]. Addressing this issue has been a subject of extensive research in the field of continual learning (CL) [12, 36, 52, 53]. Among the proposed methods, the experience replay approaches have received significant attention [11, 13, 31, 54]. These methods aim to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.386, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "S. Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.147, + 0.768, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.275, + 0.788, + 0.318 + ], + "angle": 0, + "content": "Fig. 1: During the process of CL, when the old task is sampled, there is a consequent discarding of some unsampled data. This leads to bias in the model's estimation of the distribution of the old task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.784, + 0.431 + ], + "angle": 0, + "content": "overcome catastrophic forgetting by selectively storing a subset of data from old tasks in the buffer and then combining buffer data with new task data for joint training. Various replay methods have been investigated from different perspectives to address the challenge of sample selection and balancing the contributions of new and old samples during the replay process [6,14,37]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.646 + ], + "angle": 0, + "content": "Despite the progress made in reducing catastrophic forgetting through replay methods, certain limitations still exist. A major issue is that the memory of the buffer is too small, and the model cannot store enough data to restore previously learned knowledge [4, 21, 35]. This will give rise to two significant problems. Firstly, the scarcity of old data will result in an imbalance between the number of new and old tasks, leading to a disruption of knowledge from the old task during the learning process of the new task. Secondly, in the CL process, each time the old task data is sampled, some unsampled data will be discarded, introducing bias in the model's estimation of the old task distribution. Moreover, the repeated replaying of buffer data can contribute to overfitting of this specific data. All of these factors can significantly impact the CL performance. Currently, there are some methods to address the above issues, such as selecting more representative samples [20], increasing the diversity of old samples through data expansion [3], and improving the generalization ability of the model [5]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.648, + 0.788, + 0.785 + ], + "angle": 0, + "content": "This work aims to tackle the aforementioned challenges from a different perspective. Regarding the first issue, we propose encouraging the data in the buffer to acquire task agnostic features related to the old task, thereby mitigating the impact of the new task. For the second issue, we focus on an important yet under-investigated question: How can unsampled data effectively contribute to learning new tasks? In Figure 1, we illustrate the cause of model overfitting, which arises because replay-based methods often utilize only a subset of data from old task, leading to biased results. When training on new tasks, we design the objective function as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.802, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {n e w}} = \\frac {1}{N _ {\\text {b u f f e r}} + N _ {\\text {n e w}}} \\sum_ {i = 1} ^ {N _ {n}} \\ell_ {i} (\\hat {Y}, Y). \\tag {1}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.513, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "IB Based Data Correction in CL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.253 + ], + "angle": 0, + "content": "where \\(\\ell_i(\\hat{Y}, Y)\\) represents the loss function that takes the true outcome \\(Y\\) and predicts outcome \\(\\hat{Y}\\) as inputs. \\(N_{buffer}\\) and \\(N_{new}\\) denote the numbers of buffer and new task data, respectively. \\(N_n\\) is the sum of these two quantities. However, this loss function does not provide an unbiased estimate of the loss function for both new and old tasks, as the model gradually loses some data from past tasks during training. The unbiased loss function for new and old tasks should be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.267, + 0.786, + 0.306 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} [ \\mathcal {L} _ {\\text {n e w}} ] \\neq \\mathcal {L} := \\frac {1}{N _ {\\text {o l d}} + N _ {\\text {n e w}}} \\sum_ {i = 1} ^ {N} \\ell_ {i} (\\hat {Y}, Y). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.321, + 0.785, + 0.457 + ], + "angle": 0, + "content": "where \\( N_{old} \\) and \\( N_{new} \\) represent the numbers of old task and new task data, respectively, and \\( N \\) signifies the sum of these two quantities. To address the bias arising from the discarding of unsampled data from old task, our method aims to estimate the information of the unsampled data using specific strategies. However, unsampled data cannot directly participate in the training of new tasks, we indirectly obtain a surrogate for the information about unsampled data through the relationship between sampled and unsampled data. This information related to the unsampled data serves as a regularization term to constrain the training process of the model, thereby achieving bias correction for the dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.786, + 0.776 + ], + "angle": 0, + "content": "Based on the intuition presented above, we propose an Information Bottleneck based data correction algorithm (IBCL). The information bottleneck method is a technique in information theory used to find the best compromise between accuracy and complexity [41]. It is regarded as the theoretical foundation of deep learning and has been applied in many fields, such as robust or invariant representation learning, causal reasoning, and so on [17]. This theory suggests that the model gradually eliminates irrelevant noise information during the learning process, retaining the most relevant feature information to the task. Borrowing the ability of this method, our algorithm mainly consists of two parts: Information Bottleneck Task Agnostic Constraints (IBTA): During the update process of new tasks, we introduce a constraint on the learning process of the buffer data. This constraint encourages the model to learn task agnostic features that are relevant to the old tasks, effectively preventing interference from the new tasks. Information Bottleneck Unsampled Data Surrogate (IBDS): This module is designed to model the information of the unsampled data. As the unsampled data cannot be directly trained, our method decouples the features of the unsampled data from the sampled data after the completion of learning old tasks. This decoupling allows model to leverage the relationship between the two types of data to estimate the surrogate influence of unsampled data. Consequently, this achieves an indirect implementation of the impact of unsampled data on learning new tasks." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.778, + 0.618, + 0.792 + ], + "angle": 0, + "content": "Overall, our work makes the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "- We propose a data correction framework based on information bottleneck to mitigate model bias when the number of old task samples is small." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.128 + ], + "angle": 0, + "content": "S. Chen et al." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "- To mitigate the problem of the influence of new tasks on old sample features during the learning process, we propose IBTA to encourage the model to learn task agnostic features." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.193, + 0.785, + 0.25 + ], + "angle": 0, + "content": "- To counter the issue of overfitting caused by the repeated replay of buffer data and gradual discarding of unsampled data, we propose IBDS to incorporate the information from unsampled data to assist in the training of the new model." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.253, + 0.784, + 0.282 + ], + "angle": 0, + "content": "- Our method can be combined with most replay based methods to improve their performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.308, + 0.388, + 0.323 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.341, + 0.582, + 0.357 + ], + "angle": 0, + "content": "2.1 Experience Replay Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.64 + ], + "angle": 0, + "content": "In the context of CL, there are three main approaches to retaining knowledge of past tasks: using a fixed-sized buffer for replaying past samples (replay-based approach) [2, 3, 7, 8, 43], regulating model parameter changes through learning (regularization-based approach) [16, 40, 42], or dynamically expanding model architecture as needed (expansion-based approach) [15, 28, 47]. Despite its simplicity, the replay-based approach has shown promising results in CL settings. Experience Replay (ER) [19] is a widely-used technique, which maintains a fixed-sized buffer to store and replay selected data, helping to mitigate the forgetting of previously learned knowledge. Incremental Classifier and Representation Learning (iCaRL) [37] uses the herding strategy to select a few representative samples. Rainbow Memory [3] proposes a diversity sampling strategy based on per-sample classification uncertainty. However, these methods primarily concentrate on how to choose which old data to store in the buffer, while neglecting the unsampled data. We argue that the unsampled data can also be beneficial in mitigating deviation and should be utilized. To this end, we propose a data correction method that employs the information bottleneck theory to model the information of the unsampled data in the old task, thereby enhancing the performance of several state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.663, + 0.455, + 0.677 + ], + "angle": 0, + "content": "2.2 Information Bottleneck" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Information bottleneck (IB) is a classical concept in information theory, which is used to explain the behavior of deep learning [41]. Furthermore, IB has been extensively utilized in diverse applications, including but not limited to, representation learning decoupling [23, 33], recommendation systems [29, 45], and enhancing model robustness [17]. Recently, there have been some other works that apply information theory to CL, such as [27, 46], but there are fewer works based on the IB theory. The fundamental construct of IB is based on mutual information. In a previous work [20], the maximization of mutual information is employed to learn adequate feature expression. In contrast, our proposed method concentrates on modeling the unsampled data in the previous task." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.513, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "IB Based Data Correction in CL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.149, + 0.784, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.282, + 0.788, + 0.325 + ], + "angle": 0, + "content": "Fig. 2: During the updating of new tasks, we constrain the learning process of buffer data, encouraging the model to learn task agnostic features related to the old tasks, thus avoiding the interference of new tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.351, + 0.331, + 0.367 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.784, + 0.426 + ], + "angle": 0, + "content": "Our method consists of two key processes: the learning process of new tasks and the processing process after learning old tasks. We introduce IBTA and IBDS separately." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.445, + 0.439, + 0.459 + ], + "angle": 0, + "content": "3.1 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.468, + 0.788, + 0.605 + ], + "angle": 0, + "content": "In CL scenarios, tasks are introduced one after another and handled separately. We consider a total of \\( T \\) tasks. Each task \\( D_{t} = \\{x_{i}^{t},y_{i}^{t}\\}_{i = 1}^{n_{t}} \\) comprises \\( n_t \\) input-target pairs \\( (x_i^t,y_i^t) \\), which are assumed to be independent and identically distributed (i.i.d.). The aim for each task is to train a model \\( f_{\\theta}:X\\to Y \\) with parameters \\( \\theta \\), capable of predicting the class label \\( y \\) for a new image \\( x \\). In CL algorithms that utilize experience replay, the methods typically store some data from previous task data \\( x_{old} \\) in a buffer. This stored data is referred to as sampled data \\( x_{m} \\), with its features denoted as \\( z_{m} \\). The data not selected for storage is called unsampled data \\( x_{u} \\), and its features are represented as \\( z_{u} \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.624, + 0.677, + 0.64 + ], + "angle": 0, + "content": "3.2 Information Bottleneck Task Agnostic Constraints" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.647, + 0.788, + 0.798 + ], + "angle": 0, + "content": "During the learning of new tasks, the model focuses on the new task due to the limited data available in the buffer. In order to mitigate the impact of new tasks, our method encourages the buffer data to focus on learning the task-specific features, thus minimizing their influence. In order to compress the information in the dataset into the features of the sampled data as much as possible, the mutual information between the features \\( z_{m} \\) of the sampled data and the buffer dataset \\( x_{m} \\) should be minimized; In order to make the feature be able to distinguish well, the feature \\( z_{m} \\) of the sample needs to predict the label \\( y_{old} \\) as accurately as possible. The weighting parameters \\( \\alpha \\) is used to balance the contributions of different components in the overall objective function." + }, + { + "type": "equation", + "bbox": [ + 0.356, + 0.806, + 0.786, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I B T A} := \\min \\underbrace {\\alpha I \\left(z _ {m} ; x _ {\\mathrm {m}}\\right)} _ {(1)} - \\underbrace {I \\left(z _ {m} ; y _ {\\text {o l d}}\\right)} _ {(2)}. \\tag {3}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.128 + ], + "angle": 0, + "content": "S. Chen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "In the given framework, the expression (1) represents a compression component that characterizes the mutual information between the buffer dataset \\( x_{m} \\) and the sampled data features \\( z_{m} \\). This component aims to capture the relevance of \\( z_{m} \\) to the old dataset \\( x_{m} \\). The expression (2) corresponds to an accuracy component that measures the effectiveness of \\( z_{m} \\) in representing the underlying distribution of the dataset. It aims to ensure that the sampled features \\( z_{m} \\) preserve sufficient information about the original data. Next, we will proceed to optimize the compression term \\( I(z_{m};x_{m}) \\). This term can be quantified using the Kullback-Leibler (KL) divergence as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.293, + 0.785, + 0.344 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} I (z _ {m}; x _ {\\mathrm {m}}) = \\mathbb {E} _ {x} \\left[ \\mathrm {K L} (p (z _ {m} | x _ {\\mathrm {m}}) \\parallel p (z _ {m})) \\right] \\\\ = \\mathbb {E} _ {x} \\left[ \\int p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) \\log p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) d z - \\int p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) \\log p \\left(z _ {m}\\right) d z \\right]. \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.351, + 0.79, + 0.442 + ], + "angle": 0, + "content": "Nonetheless, due to the computational complexity involved in calculating \\( p(z_{m}) = \\int p(z_{m}|x_{m})p(x_{m})dx \\), we resort to the utilization of a variational approximation approach. Instead of using \\( p(z_{m}) \\), we adopt a variational distribution denoted as \\( q(z_{m}) \\). Leveraging Gibbs' inequality, we recognize the non-negativity of the Kullback-Leibler (KL) divergence. Thus, we can establish an upper bound for Equation (4)," + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.451, + 0.785, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} - \\int p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) \\log p \\left(z _ {m}\\right) d z \\leq - \\int p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) \\log q \\left(z _ {m}\\right) d z \\tag {5} \\\\ \\Rightarrow \\operatorname {K L} \\left(p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) \\| p \\left(z _ {m}\\right)\\right) \\leq \\operatorname {K L} \\left(p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) \\| q \\left(z _ {m}\\right)\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.659 + ], + "angle": 0, + "content": "Similar to prior research [1, 29, 34, 44, 45], we can make the assumption that the posterior distribution \\( p(z_{m} \\mid x_{m}) \\) follows a Gaussian distribution and set it \\( \\mathcal{N}(e(x_m),\\mathrm{diag}(\\sigma (x_m))) \\). In this context, \\( e(x_{m}) \\) represents the encoded embedding of the variables \\( x_{m} \\), and \\( \\mathrm{diag}(\\sigma (x_m)) \\) indicates the variance in a diagonal matrix form. By employing the reparameterization trick, the embedding \\( z_{m} \\) can be generated through the formula \\( z_{m} = e(x_{m}) + \\epsilon \\odot \\sigma (x_{m}) \\), where \\( \\epsilon \\sim \\mathcal{N}(0,I) \\). Notably, if we set \\( \\sigma (x_{m}) \\) as an all-zero matrix, \\( z_{m} \\) will effectively become a deterministic embedding. Conversely, the prior distribution \\( q(z_{m}) \\) is assumed to adhere to a standard Gaussian variational distribution, namely \\( q(z_{m}) = \\mathcal{N}(0,I) \\). Ultimately, the previously mentioned upper bound can be rephrased as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.28, + 0.667, + 0.785, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\operatorname {K L} \\left(p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) \\| q \\left(z _ {m}\\right)\\right) = \\| e \\left(x _ {\\mathrm {m}}\\right) \\| _ {2} ^ {2} + \\sum_ {d} \\left(\\sigma_ {d} - \\frac {1}{2} \\log \\sigma_ {d} - 1\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.71, + 0.788, + 0.817 + ], + "angle": 0, + "content": "where \\(\\sigma_d\\) corresponds to an element within \\(\\operatorname{diag}\\{\\sigma(x_m)\\}\\). This signifies that for a deterministic embedding \\(z_m\\), we can enhance the optimization of this upper bound by directly implementing \\(\\ell_2\\)-norm regularization on the embedding vector \\(z_m\\). This process is equivalent to refining the compression term \\(I(z_m; x_m)\\). When considering the mutual information \\(I(z_m; y_{old})\\) within Equation (3), we can express it as \\(I(z_m; y_{old}) = H(y_{old}) - H(y_{old} \\mid z_m)\\). As \\(H(y_{old})\\) is a positive constant that can be disregarded, the following inequality holds:" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.826, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\nI \\left(z _ {m}; y _ {\\text {o l d}}\\right) \\geq - H \\left(y _ {\\text {o l d}} \\mid z _ {m}\\right). \\tag {7}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.513, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "IB Based Data Correction in CL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.149, + 0.78, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.282, + 0.788, + 0.366 + ], + "angle": 0, + "content": "Fig. 3: This module is used to model the information of unsampled data. Because unsampled data cannot be trained directly, our method decouples the features of unsampled data and sampled data after the completion of the old task learning, so that it is convenient to estimate the surrogate loss of unsampled data in the new task by using the relationship between them, and indirectly realize the influence of unsampled data on the new task learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.788, + 0.506 + ], + "angle": 0, + "content": "This term can be seen as a cross entropy loss function. Therefore, we have implemented the constraint process for learning buffer data. However, the loss function obtained in this way is still biased, as there is missing unsampled data, which means that the contribution to the old task label \\( y_{old} \\) should be \\( I(z; y) = I(z_m, z_u; y) \\), where \\( z_u \\) is unsampled data feature. During the learning process of the new task, this item cannot be directly calculated. We will calculate it indirectly in the following section." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.535, + 0.685, + 0.551 + ], + "angle": 0, + "content": "3.3 Information Bottleneck Unsampled Data Surrogate" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.789, + 0.81 + ], + "angle": 0, + "content": "To estimate the mutual information \\( I(z; y) = I(z_{m}, z_{u}; y) \\) during the learning process of new task, it becomes essential to obtain \\( z_{u} \\). Given that \\( x_{u} \\) denotes unsampled data, we cannot directly calculate \\( z_{u} \\) from \\( x_{u} \\) when learning new task. This necessitates an indirect estimation of \\( z_{u} \\) from old task model. As shown in Fig.3, during the learning stage of old task, both sampled \\( x_{m} \\) and unsampled \\( x_{u} \\) data are involved in the training process. After training completion, the features corresponding to these two parts \\( z_{m} \\) and \\( z_{u} \\) are tightly coupled. Consequently, a specific operation is needed after training ends to decouple these two features. If we can decouple the features of sampled and unsampled data by optimizing \\( I(z_{m}; z_{u}) \\), we can get \\( I(z; y) = I(z_{m}; y) + I(z_{u}; y) \\). In this case, the estimate of \\( I(z_{u}; y) \\) can be used as a surrogate for the unsampled data information during learning new task. The feature \\( z_{m} \\) of the sampled data and the feature \\( z_{u} \\) of the unsampled data must maintain as much independence from one another as possible to obtain better discrimination. Drawing on principles from information theory, we are inspired to derive a minimization objective function based on the aforementioned analysis:" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.825, + 0.786, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I B D S} := \\min I \\left(z _ {m}; z _ {u}\\right). \\tag {8}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.128 + ], + "angle": 0, + "content": "S. Chen et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.785, + 0.192 + ], + "angle": 0, + "content": "represents a decoupling component, capturing the level of interdependence between \\( z_{u} \\) and \\( z_{m} \\). It aims to minimize the dependence of \\( z_{u} \\) on \\( z_{m} \\), ensuring that \\( z_{u} \\) provides additional information beyond what is already captured by \\( z_{m} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.786, + 0.239 + ], + "angle": 0, + "content": "Inspired by previous research work [1, 29, 34, 44, 45], utilizing the chain rule of mutual information, we derive the subsequent equation pertaining to the decoupling component:" + }, + { + "type": "equation", + "bbox": [ + 0.295, + 0.251, + 0.786, + 0.268 + ], + "angle": 0, + "content": "\\[\nI \\left(z _ {m}; z _ {u}\\right) = I \\left(z _ {m}; y _ {o l d}\\right) - I \\left(z _ {m}; y _ {o l d} \\mid z _ {u}\\right) + I \\left(z _ {m}; z _ {u} \\mid y _ {o l d}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.278, + 0.787, + 0.354 + ], + "angle": 0, + "content": "We proceed to analyze the expression \\( I(z_{m};z_{u}\\mid y_{old}) \\) in greater detail. Given that the distribution of \\( z_{m} \\) exclusively hinges on the variables \\( x_{m} \\), and the latter is influenced by \\( y_{old} \\), we establish \\( H(z_{m}\\mid y_{old},z_{u}) = H(z_{m}\\mid y_{old}) \\), where \\( H(\\cdot \\mid \\cdot) \\) signifies conditional entropy. By combining the attributes of mutual information, we arrive at the following relationship:" + }, + { + "type": "equation", + "bbox": [ + 0.325, + 0.367, + 0.785, + 0.401 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} I (z _ {m}; z _ {u} \\mid y _ {\\mathrm {o l d}}) = H (z _ {m} \\mid y _ {\\mathrm {o l d}}) - H (z _ {m} \\mid y _ {\\mathrm {o l d}}, z _ {u}) \\\\ = H \\left(z _ {m} \\mid y _ {\\text {o l d}}\\right) - H \\left(z _ {m} \\mid y _ {\\text {o l d}}\\right) = 0. \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.413, + 0.453, + 0.429 + ], + "angle": 0, + "content": "By substituting Eq.(9), we have," + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.441, + 0.785, + 0.457 + ], + "angle": 0, + "content": "\\[\nI \\left(z _ {m}; z _ {u}\\right) = I \\left(z _ {m}; y _ {\\text {o l d}}\\right) - I \\left(z _ {m}; y _ {\\text {o l d}} \\mid z _ {u}\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.786, + 0.499 + ], + "angle": 0, + "content": "Since the term \\( I(z_{m};y_{old} \\mid z_{u}) \\) in Eq.(11) is still difficult to be calculated, we simplify it further by using the form of conditional entropy:" + }, + { + "type": "equation", + "bbox": [ + 0.288, + 0.511, + 0.785, + 0.564 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} I \\left(z _ {m}; y _ {\\text {o l d}} \\mid z _ {u}\\right) = H \\left(z _ {m} \\mid z _ {u}\\right) + H \\left(y _ {\\text {o l d}} \\mid z _ {u}\\right) - I \\left(z _ {m}; y _ {\\text {o l d}} \\mid z _ {u}\\right) \\\\ = H \\left(z _ {m} \\mid z _ {u}\\right) - H \\left(z _ {m} \\mid y _ {\\text {o l d}}, z _ {u}\\right) \\tag {12} \\\\ = H \\left(y _ {\\text {o l d}} \\mid z _ {u}\\right) - H \\left(y _ {\\text {o l d}} \\mid z _ {m}, z _ {u}\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.303, + 0.582, + 0.785, + 0.598 + ], + "angle": 0, + "content": "\\[\nI \\left(z _ {m}; z _ {u}\\right) = I \\left(z _ {m}; y _ {o l d}\\right) - H \\left(y _ {o l d} \\mid z _ {u}\\right) + H \\left(y _ {o l d} \\mid z _ {m}, z _ {u}\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.604, + 0.785, + 0.636 + ], + "angle": 0, + "content": "Finally, combining Eq.(7) and Eq.(13), we can rewrite the objective function \\(\\mathcal{L}_{IBDS}\\) in Eq.(8) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.299, + 0.648, + 0.785, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I B D S} = - H \\left(y _ {o l d} \\mid z _ {m}\\right) - H \\left(y _ {o l d} \\mid z _ {u}\\right) + H \\left(y _ {o l d} \\mid z _ {m}, z _ {u}\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.674, + 0.786, + 0.719 + ], + "angle": 0, + "content": "By optimizing \\(\\mathcal{L}_{IBDS}\\), we get the decoupled expression \\(z_{u}\\). Now we can compute \\(I(z_{u};y)\\), which acts as a surrogate loss to compensate for missing weights due to unsampled data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.841 + ], + "angle": 0, + "content": "During the training process of neural network, the model will encounter the problem of semantic drift [22,24,51], which means that the feature distribution of sampled data will change with the learning of new tasks. This concept highlights a significant challenge: using the fixed features of unsampled data as constraints in the learning of new task becomes problematic. As the model trains on new tasks, the estimation of unsampled data features gradually deviates from their true values, which not only affects the effectiveness of bias correction but also, to some extent, impacts the overall learning process. Just as the \"no free lunch\"" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.513, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "IB Based Data Correction in CL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.418 + ], + "angle": 0, + "content": "theorem advocated in machine learning. Once samples are lost, it becomes difficult to perfectly estimate the distribution of unsampled data, the best one can do is to minimize the model's bias when addressing old tasks. To mitigate this impact, it's essential to consider the characteristics of neural networks in CL [9]. Early in the learning of new tasks, gradients undergo significant changes, and the model frequently encounters considerable biases. A viable strategy involves imposing stricter constraints on the new task learning process during the initial stages of model training. This approach instructs the model in reducing biases when learning new tasks. As the training advances, these constraints are gradually relaxed, permitting the model to better adjust to the learning requirements of new tasks during the slower update phases. To address the processing of unsampled data features, a dynamic decay strategy is employed to balance the learning conflicts between old and new tasks, thereby optimizing the model's overall performance. This method allows the model to discover a more balanced learning point between preserving old knowledge and acquiring new tasks, enhancing its learning efficiency and effectiveness. Specifically, inspired by [30], we propose a dynamic cosine annealing factor \\(\\eta\\) for constraint terms when updating unsampled data." + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.419, + 0.786, + 0.451 + ], + "angle": 0, + "content": "\\[\n\\eta = \\left(\\frac {1}{2} (1 + \\cos \\left(\\frac {e _ {\\text {c u r}}}{e _ {\\text {m a x}}}\\right) \\pi)\\right) \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.456, + 0.738, + 0.472 + ], + "angle": 0, + "content": "where \\( e_{cur} \\) and \\( e_{max} \\) represent current and maximum number of epochs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.494, + 0.348, + 0.509 + ], + "angle": 0, + "content": "3.4 Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.52, + 0.71, + 0.536 + ], + "angle": 0, + "content": "By combining above two parts, we can get the total loss function as:" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.548, + 0.785, + 0.603 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {C I B} = \\alpha I (z _ {m}; x _ {\\mathrm {m}}) - I (z; y _ {\\mathrm {o l d}}) \\\\ = \\alpha I \\left(z _ {m}; x _ {\\mathrm {m}}\\right) - \\beta I \\left(z _ {u}; y\\right) - \\gamma I \\left(z _ {m}; y\\right) \\tag {16} \\\\ \\leq \\alpha \\| e \\left(x _ {\\mathrm {m}}\\right) \\| _ {2} ^ {2} + \\beta \\cdot \\eta H \\left(y _ {\\text {o l d}} \\mid z _ {u}\\right) + \\gamma H \\left(y _ {\\text {o l d}} \\mid z _ {m}\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.786, + 0.644 + ], + "angle": 0, + "content": "The weighting parameters \\(\\alpha\\), \\(\\beta\\), \\(\\gamma\\) are used to balance the contributions of different components in the overall objective function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.719 + ], + "angle": 0, + "content": "The final objective function comprises three terms: term (1) represents a regularization constraint on the input data, term (2) represents surrogate loss to reflects the influence of the unsampled data; term (3) denotes the cross-entropy between \\( z_{m} \\) and \\( y_{old} \\), this can be seen as the cross-entropy loss function. The aim is to make the learned feature expression fully predict the final label." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Discussion This approach comprises two primary modules: the IBTA module facilitates the learning of new tasks, whereas the IBDS module is utilized post-learning of old tasks to derive features from unsampled data. These features are then integrated as constraints within the training of new tasks, thereby shaping the overall optimization loss function. Our method can be combined with most replay-based strategies. Taking the naive ER method as an example, it randomly selects some samples from the old task and stores them in a buffer, which are trained together with the samples from the new task. Upon finalizing" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "S. Chen et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 1: Comparison of the average accuracy (the higher the better) of four split datasets with different buffer sizes." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.185, + 0.789, + 0.361 + ], + "angle": 0, + "content": "
MethodSplit CIFAR-10Split CIFAR-100Split ImageNet-100Split miniImageNet
Joint92.3873.2980.2353.55
FT19.679.298.689.52
Buffer size10020050020050020002005002000100020005000
ER36.3944.7957.7414.3519.6636.7613.6318.3734.258.3716.4924.17
+IBCL45.4351.9163.1223.9828.3242.0122.7228.9343.1114.7922.7227.69
ER-ACE53.9063.4170.5326.2836.4848.4124.2337.1249.5517.9522.6027.92
+IBCL63.8570.9774.8232.1240.9451.8930.7241.8152.6224.7727.6331.02
DER++57.6564.8872.7025.1137.1352.0826.5043.6558.0518.0223.4430.43
+IBCL66.4172.5276.6133.8244.2154.8932.6346.5759.9227.7231.4135.79
X-DER59.2965.1968.1035.3444.6254.4433.2146.7255.2325.2426.3829.91
+IBCL67.9173.8274.1443.8648.7255.9143.2549.6256.5929.9030.2132.66
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.785, + 0.48 + ], + "angle": 0, + "content": "the learning of old task, the sampled data is processed through the old model to yield \\( z_{m} \\), and both sampled and unsampled data are processed to extract coupled features \\( z \\). Subsequently, by optimizing IBDS, features \\( z_{u} \\) for each class of unsampled data are obtained. During the training of the new task, IBTA and \\( z_{u} \\) are combined to optimize the total loss function for learning the new task. Detailed pseudocode of this procedure is provided in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.502, + 0.377, + 0.519 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.785, + 0.609 + ], + "angle": 0, + "content": "In this section, we will present the datasets and evaluation metrics used in the experiment, and describe the implementation details. Then we describe the experimental results of our method on multiple standard benchmarks. Finally, we conduct an ablation study to illustrate the effectiveness of each module of our method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.632, + 0.329, + 0.646 + ], + "angle": 0, + "content": "4.1 Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.785, + 0.749 + ], + "angle": 0, + "content": "We validate our results on four distinct datasets: CIFAR-10, CIFAR-100, miniImageNet, and ImageNet-100. CIFAR-10 and CIFAR-100 each consist of 50,000 training samples and 10,000 testing samples, with 10 and 100 classes respectively. miniImageNet and ImageNet-100 are subsets of ImageNet, both containing 100 classes. miniImageNet includes 500 images per category for training, while ImageNet-100 includes 1300 images per category for training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.321, + 0.784 + ], + "angle": 0, + "content": "4.2 Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.841 + ], + "angle": 0, + "content": "To better compare with other methods, we follow the settings of [46] and [5]. For the CIFAR-10/100 and ImageNet-100 datasets, we adopt ResNet-18, and for the miniImageNet dataset, we adopt EfficientNet-B2. For CIFAR-10/100, we" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.513, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "IB Based Data Correction in CL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Comparison of our method with other regularization methods on three datasets, we use two typical methods DER++ and ER-ACE as benchmarks for comparison." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.199, + 0.777, + 0.479 + ], + "angle": 0, + "content": "
MethodSplit CIFAR-10Split CIFAR-100Split miniImageNet
Buffer size1002005002005002000100020005000
ER-ACE53.9063.4170.5326.2836.4848.4117.9522.6027.92
+ sSGD56.2664.7371.4528.0739.5949.7018.1122.4324.12
+ oEwC52.3661.0968.7024.9335.0645.5919.0424.3229.46
+ oLAP52.7663.1970.3226.4236.5847.6618.3423.1928.77
+ OCM57.1864.6570.8628.1837.7449.0320.3224.3228.57
+ LiDER56.0865.3271.7527.9438.4350.3219.6924.1330.00
+ DualHSIC60.5268.0873.7829.0838.9450.5522.3325.4130.12
+ Our63.8570.9774.8232.1240.9451.8924.7727.6331.02
DER++57.6564.8872.7025.1137.1352.0818.0223.4430.43
+ sSGD55.8164.4472.0524.7638.4850.7416.3119.2924.24
+ oEwC55.7863.0271.6424.5135.2251.5318.8724.5331.91
+ oLAP54.8662.5471.3823.2634.4850.8018.9125.0232.78
+ OCM59.2565.8173.5327.4638.9452.2520.9324.7531.16
+ LiDER58.4366.0273.3927.3239.2553.2721.5828.3335.04
+ DualHSIC64.9870.2875.9431.4641.8653.5324.7829.3734.98
+ Our66.4172.5276.6133.8244.2154.8927.7231.4135.79
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.587 + ], + "angle": 0, + "content": "adopt a training epoch of 50 and set the batch size to 32 and 64, respectively. For MiniImageNet and ImageNet-100, we adopt training epochs of 80 and 120, with batch size set to 128 and 64, respectively. In order to better compare the experimental results, we adopt the experimental results reported in [46] and [5], or the hyperparametric design in the original article." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.616, + 0.421, + 0.631 + ], + "angle": 0, + "content": "4.3 Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.742 + ], + "angle": 0, + "content": "We mainly evaluate on class incremental learning scenario. In this setup, each sequential task introduces a distinct set of classes that do not overlap with those from previous tasks. The challenge lies in training a classifier that can effectively classify all classes encountered up to the current stage, without the need for explicit task ids. We evaluate the final performance of the method by average accuracy." + }, + { + "type": "equation", + "bbox": [ + 0.427, + 0.759, + 0.787, + 0.801 + ], + "angle": 0, + "content": "\\[\nA C C = \\frac {1}{T} \\sum_ {i = 1} ^ {T} R _ {T, i}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.839 + ], + "angle": 0, + "content": "where \\( R_{T,i} \\) denotes the test accuracy on task \\( i \\) after the model has finished task \\( T \\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.128 + ], + "angle": 0, + "content": "S. Chen et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.174 + ], + "angle": 0, + "content": "Table 3: Abalation study of the average accuracy (the higher the better) of split ImageNet-100 and split CIFAR-100 with different buffer sizes." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.185, + 0.785, + 0.354 + ], + "angle": 0, + "content": "
MethodSplit CIFAR-100Split ImageNet-100
Buffer size20050020002005002000
ER-ACE26.2836.4848.4124.2337.1249.55
+ IBTA27.3637.4949.1225.6337.9250.12
+ IBDS30.4439.5250.9129.8340.7151.85
+IBCL32.1240.9451.8930.7241.8152.62
DER++25.1137.1352.0826.5043.6558.05
+IBTA27.8838.7453.9627.3945.7259.33
+ IBDS31.5342.5954.3329.5846.1559.70
+IBCL33.8244.2154.8932.6346.5759.92
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.378, + 0.398, + 0.392 + ], + "angle": 0, + "content": "4.4 Baseline Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.787, + 0.492 + ], + "angle": 0, + "content": "We conduct experiments on multiple state-of-the-art experience replay CL models: ER [10], DER++ [8], X-DER [6], and ER-ACE [9]. In addition to the original loss function based on replay, our method can be considered as a new regularization method. Therefore, we also compare our method with existing regularization-based techniques, including sSGD [32], oEWC [39], oLAP [38], and more recent SOTA methods, OCM [20], LiDER [5] and DualHSIC [46]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.508, + 0.325, + 0.521 + ], + "angle": 0, + "content": "4.5 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.787, + 0.68 + ], + "angle": 0, + "content": "The comparison results are reported in Table 1. For the method of listing several baselines, our proposed method has improved the model's performance in various datasets and buffer sizes. In addition, we have the following main observations: 1) There has been a significant improvement in basic methods such as ER and DER++, as these two methods adopt more basic methods for sampling. Our method can reduce the bias caused by the sampling process to a certain extent. 2) Our methods have also shown some effectiveness on X-DER and ER-ACE, which take additional measures to alleviate sample imbalance and improve sample utilization, resulting in improved model performance. Our method further enhances the model's capabilities on top of this." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.787, + 0.772 + ], + "angle": 0, + "content": "To further demonstrate the effectiveness of our method, we compare our method with other regularization methods, and it can be found that we achieve the best performance, recent methods such as OCM, LiDER and DualHSIC achieve significant progress compared to baseline when the number of buffer is small, and our method further improves the performance when the buffer data is small by considering modeling for the unsampled data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.789, + 0.39, + 0.804 + ], + "angle": 0, + "content": "4.6 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In order to understand the effect of each of the items in the loss function on the performance of the algorithm, an ablation study has been carried out and the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.513, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "IB Based Data Correction in CL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.149, + 0.42, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.149, + 0.601, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.149, + 0.78, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.372 + ], + "angle": 0, + "content": "Fig. 4: This figure shows the CKA matrix based on the ER algorithm, FT and incorporating our algorithm. It can be observed from it that due to our modeling of unsampled data, the model shows more information about the old task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.789, + 0.556 + ], + "angle": 0, + "content": "results are shown in Table 3. It is observed that in most of the cases, increasing any of the components has an impact on the performance of the algorithm. However, increasing the IBDS gives a large performance improvement, which indicates that it plays a vital role in the performance of the algorithm. This component represents the loss of substitution of unsampled data, which is necessary for training on new tasks using unsampled data features. Therefore, if this aspect is not well learned, it may negatively affect the subsequent performance of the algorithm. IBTA also has an impact on the performance of the algorithm, the involves regularization constraints on the input features in order to make the model learn enough robust features." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.587, + 0.478, + 0.604 + ], + "angle": 0, + "content": "5 Characteristic Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.787, + 0.658 + ], + "angle": 0, + "content": "We analyze various characteristics of the model to gain insight into the factors contributing to the performance enhancements achieved by our approach." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.687, + 0.445, + 0.702 + ], + "angle": 0, + "content": "5.1 CKA Matrix Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We use Central Kernel Alignment (CKA) [26] in Fig.4, which is an effective tool to measure the similarity of network representations and evaluate the relationships between these backbones. We send the same batch of instances to the ER (2000 buffer size) algorithm model and fine-tuning based model respectively to obtain corresponding feature maps, thus intuitively demonstrating the similarity of algorithm feature maps. Our algorithm can be used in conjunction with the experience replay method to model unsampled data. When the sampled data is 0, i.e. fine-tuning, our method can also get some benefits." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.127 + ], + "angle": 0, + "content": "S. Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.149, + 0.418, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.259, + 0.35, + 0.269 + ], + "angle": 0, + "content": "Initialization" + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.149, + 0.586, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.475, + 0.259, + 0.531, + 0.269 + ], + "angle": 0, + "content": "Epoch \\(= 20\\)" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.149, + 0.753, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.259, + 0.691, + 0.269 + ], + "angle": 0, + "content": "Epoch \\(= 50\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.285, + 0.789, + 0.34 + ], + "angle": 0, + "content": "Fig. 5: This image uses TSNE to display the changes in the characteristics of sampled and unsampled data with the number of training epochs of the model. We have shown one class of situation, with a sample data size of 200 and an unsampled data size of 1100." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.371, + 0.499, + 0.387 + ], + "angle": 0, + "content": "5.2 Feature Decoupling Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.789, + 0.517 + ], + "angle": 0, + "content": "A significant advance of our method is the decoupling of sampled and unsampled data features. We demonstrate this in Fig.5, where we use TSNE technology to visualize the changes in these two features with the increase of network training epochs. We test our method, the ER algorithm, on 200 sampled data and 1100 unsampled data from the ImageNet-100 dataset. As the network training epochs increase, we can observe that the two sets of features begin to separate from their initial coupling, which indicates that our method has achieved decoupling of the two data." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.541, + 0.36, + 0.557 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.573, + 0.789, + 0.724 + ], + "angle": 0, + "content": "In conclusion, our proposed method is designed to tackle the issue of data bias in continual learning through the application of experience replay. Drawing inspiration from information bottleneck theory, we steer the incorporation of unsampled data information, thereby bolstering the model's capacity to preserve past knowledge while accommodating novel tasks. Our conducted experiments offer empirical validation of the advantages conferred by our approach. While we have presented a method for computing unsampled data, it's important to acknowledge the existence of several alternative strategies for estimating these. This opens up exciting possibilities for continued exploration and innovation in this field." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.748, + 0.405, + 0.765 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "This work was supported by National Science and Technology Major Project, Grant No.2022ZD0116403, the Youth Innovation Promotion Association CAS and Strategic Priority Research Program of Chinese Academy of Sciences XDA27010300." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.513, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "IB Based Data Correction in CL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Alemi, A.A., Fischer, I., Dillon, J.V., Murphy, K.: Deep variational information bottleneck. arXiv (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.232 + ], + "angle": 0, + "content": "2. Aljundi, R., Lin, M., Goujaud, B., Bengio, Y.: Gradient based sample selection for online continual learning. In: Adv. Neural Inform. Process. Syst. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.785, + 0.272 + ], + "angle": 0, + "content": "3. Bang, J., Kim, H., Yoo, Y., Ha, J.W., Choi, J.: Rainbow memory: Continual learning with a memory of diverse samples. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "4. Biesialska, M., Biesialska, K., Costa-Jussa, M.R.: Continual lifelong learning in natural language processing: A survey. arXiv (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.301, + 0.785, + 0.342 + ], + "angle": 0, + "content": "5. Bonicelli, L., Boschini, M., Porrello, A., Spampinato, C., Calderara, S.: On the effectiveness of lipschitz-driven rehearsal in continual learning. In: Adv. Neural Inform. Process. Syst. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "6. Boschini, M., Bonicelli, L., Buzzega, P., Porrello, A., Calderara, S.: Class incremental continual learning into the extended der-verse. IEEE Trans. Pattern Anal. Mach. Intell. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "7. Boschini, M., Buzzega, P., Bonicelli, L., Porrello, A., Calderara, S.: Continual semisupervised learning through contrastive interpolation consistency. Pattern Recognition Letters 162, 9-14 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.466 + ], + "angle": 0, + "content": "8. Buzzega, P., Boschini, M., Porrello, A., Abati, D., Calderara, S.: Dark experience for general continual learning: a strong, simple baseline. In: Adv. Neural Inform. Process. Syst. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.785, + 0.508 + ], + "angle": 0, + "content": "9. Caccia, L., Aljundi, R., Asadi, N., Tuytelaars, T., Pineau, J., Belilovsky, E.: New insights on reducing abrupt representation change in online continual learning. In: Int. Conf. Learn. Represent. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "0. Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H.S., Ranzato, M.: Continual learning with tiny episodic memories. Arxiv (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.785, + 0.577 + ], + "angle": 0, + "content": "1. Chen, S., Zhang, M., Zhang, J., Huang, K.: Exemplar-based continual learning via contrastive learning. IEEE Transactions on Artificial Intelligence 5(7), 3313-3324 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "2. Cossu, A., Carta, A., Lomonaco, V., Bacciu, D.: Continual learning for recurrent neural networks: an empirical evaluation. Neural Networks (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "3. Delange, M., Aljundi, R., Masana, M., Parisot, S., Jia, X., Leonardis, A., Slabaugh, G., Tuytelaars, T.: A continual learning survey: Defying forgetting in classification tasks. IEEE Trans. Pattern Anal. Mach. Intell. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.687 + ], + "angle": 0, + "content": "4. Douillard, A., Cord, M., Ollion, C., Robert, T., Valle, E.: PODNet: Pooled Outputs Distillation for Small-Tasks Incremental Learning. In: Eur. Conf. Comput. Vis. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.785, + 0.729 + ], + "angle": 0, + "content": "5. Douillard, A., Ram, A., Couairon, G., Cord, M.: Dytox: Transformers for continual learning with dynamic token expansion. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.785, + 0.771 + ], + "angle": 0, + "content": "6. Ebrahimi, S., Elhoseiny, M., Darrell, T., Rohrbach, M.: Uncertainty-guided continual learning with bayesian neural networks. In: Int. Conf. Learn. Represent. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "7. Federici, M., Dutta, A., Forre, P., Kushman, N., Akata, Z.: Learning robust representations via multi-view information bottleneck. In: Int. Conf. Learn. Represent. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "8. French, R.M.: Catastrophic forgetting in connectionist networks. Trends in cognitive sciences (1999)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "S. Chen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "19. Goodfellow, I.J., Mirza, M., Xiao, D., Courville, A., Bengio, Y.: An empirical investigation of catastrophic forgetting in gradient-based neural networks. arXiv (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.217 + ], + "angle": 0, + "content": "20. Guo, Y., Liu, B., Zhao, D.: Online continual learning through mutual information maximization. In: Int. Conf. Mach. Learn. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "21. Hadsell, R., Rao, D., Rusu, A.A., Pascanu, R.: Embracing change: Continual learning in deep neural networks. Trends in cognitive sciences (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.272 + ], + "angle": 0, + "content": "22. Iscen, A., Zhang, J., Lazebnik, S., Schmid, C.: Memory-efficient incremental learning through feature adaptation. In: Eur. Conf. Comput. Vis. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.313 + ], + "angle": 0, + "content": "23. Jaiswal, A., Brekelmans, R., Moyer, D., Steeg, G.V., AbdAlmageed, W., Natarajan, P.: Discovery and separation of features for invariant representation learning. arXiv (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "24. Jie, S., Deng, Z.H., Li, Z.: Alleviating representational shift for continual fin-tuning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.396 + ], + "angle": 0, + "content": "25. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proc. of the national academy of sciences (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "26. Kornblith, S., Norouzi, M., Lee, H., Hinton, G.E.: Similarity of neural network representations revisited. In: Int. Conf. Mach. Learn. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "27. Li, D., Wang, T., Chen, J., Ren, Q., Kawaguchi, K., Zeng, Z.: Towards continual learning desiderata via hsic-bottleneck orthogonalization and equiangular embedding. In: AAAI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "28. Li, X., Zhou, Y., Wu, T., Socher, R., Xiong, C.: Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In: Int. Conf. Mach. Learn. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.549 + ], + "angle": 0, + "content": "29. Liu, D., Cheng, P., Zhu, H., Dong, Z., He, X., Pan, W., Ming, Z.: Mitigating confounding bias in recommendation via information bottleneck. In: Proceedings of the 15th ACM Conference on Recommender Systems (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "30. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "31. Masana, M., Liu, X., Twardowski, B., Menta, M., Bagdanov, A.D., van de Weijer, J.: Class-incremental learning: Survey and performance evaluation on image classification. IEEE Trans. Pattern Anal. Mach. Intell. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.646 + ], + "angle": 0, + "content": "32. Mirzadeh, S.I., Farajtabar, M., Pascanu, R., Ghasemzadeh, H.: Understanding the role of training regimes in continual learning. Arxiv (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.673 + ], + "angle": 0, + "content": "33. Moyer, D., Gao, S., Brekelmans, R., Steeg, G.V., Galstyan, A.: Evading the adversary in invariant representation. CoRR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.785, + 0.702 + ], + "angle": 0, + "content": "34. Pan, Z., Niu, L., Zhang, J., Zhang, L.: Disentangled information bottleneck. In: AAAI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.729 + ], + "angle": 0, + "content": "35. Parisi, G.I., Kemker, R., Part, J.L., Kanan, C., Wermter, S.: Continual lifelong learning with neural networks: A review. Neural Networks (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.757 + ], + "angle": 0, + "content": "36. Qu, H., Rahmani, H., Xu, L., Williams, B., Liu, J.: Recent advances of continual learning in computer vision: An overview. arXiv (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "37. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "38. Ritter, H., Botev, A., Barber, D.: Online structured laplace approximations for overcoming catastrophic forgetting. In: Adv. Neural Inform. Process. Syst. (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "39. Schuster, M., Paliwal, K.K.: Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing 45(11), 2673-2681 (1997)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.513, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "IB Based Data Correction in CL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "40. Swaroop, S., Nguyen, C.V., Bui, T.D., Turner, R.E.: Improving and understanding variational continual learning. ArXiv (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.78, + 0.189 + ], + "angle": 0, + "content": "41. Tishby, N., Pereira, F.C., Bialek, W.: The information bottleneck method (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "42. Titsias, M.K., Schwarz, J., Matthews, A.G.d.G., Pascanu, R., Teh, Y.W.: Functional regularisation for continual learning with gaussian processes. In: Int. Conf. Learn. Represent. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.273 + ], + "angle": 0, + "content": "43. Von Oswald, J., Zhao, D., Kobayashi, S., Schug, S., Caccia, M., Zucchet, N., Sacramento, J.: Learning where to learn: Gradient sparsity in meta and continual learning. In: Adv. Neural Inform. Process. Syst. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "44. Wang, Y., Rudner, T.G., Wilson, A.G.: Visual explanations of image-text representations via multi-modal information bottleneck attribution. In: Adv. Neural Inform. Process. Syst. (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "45. Wang, Z., Chen, X., Wen, R., Huang, S.L., Kuruoglu, E.E., Zheng, Y.: Information theoretic counterfactual learning from missing-not-at-random feedback. In: Adv. Neural Inform. Process. Syst. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.396 + ], + "angle": 0, + "content": "46. Wang, Z., Zhan, Z., Gong, Y., Shao, Y., Ioannidis, S., Wang, Y., Dy, J.: Dualhsic: Hsic-bottleneck and alignment for continual learning. In: Int. Conf. Mach. Learn. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "47. Wortsman, M., Ramanujan, V., Liu, R., Kembhavi, A., Rastegari, M., Yosinski, J., Farhadi, A.: Supermasks in superposition. In: Adv. Neural Inform. Process. Syst. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "48. Xu, P., Zhang, J., Huang, K.: Exploration via joint policy diversity for sparse-reward multi-agent tasks. In: IJCAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "49. Xu, P., Zhang, J., Huang, K.: Population-based diverse exploration for sparse-reward multi-agent tasks. In: IJCAI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "50. Xu, P., Zhang, J., Yin, Q., Yu, C., Yang, Y., Huang, K.: Subspace-aware exploration for sparse-reward multi-agent tasks. In: AAAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "51. Yu, L., Twardowski, B., Liu, X., Herranz, L., Wang, K., Cheng, Y., Jui, S., Weijer, J.v.d.: Semantic drift compensation for class-incremental learning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "52. Zhang, X., Yi, J., Tao, J., Wang, C., Zhang, C.Y.: Do you remember? overcoming catastrophic forgetting for fake audio detection. In: Int. Conf. Mach. Learn. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "53. Zhang, X., Yi, J., Wang, C., Zhang, C.Y., Zeng, S., Tao, J.: What to remember: Self-adaptive continual learning for audio deepfake detection. In: AAAI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.647 + ], + "angle": 0, + "content": "54. Zhou, D.W., Wang, Q.W., Qi, Z.H., Ye, H.J., Zhan, D.C., Liu, Z.: Deep class-incremental learning: A survey. arXiv (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.647 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Information Bottleneck Based Data Correction in Continual Learning/6c4e72fe-4072-432b-96c2-9a488daff608_origin.pdf b/2024/Information Bottleneck Based Data Correction in Continual Learning/6c4e72fe-4072-432b-96c2-9a488daff608_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2a9f954a37d95340b1684ef48125d0a587f98ae5 --- /dev/null +++ b/2024/Information Bottleneck Based Data Correction in Continual Learning/6c4e72fe-4072-432b-96c2-9a488daff608_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0cb1bde804b103fc029025796a5e990452dc5d2a59c696462d1c0358918a132 +size 966796 diff --git a/2024/Information Bottleneck Based Data Correction in Continual Learning/full.md b/2024/Information Bottleneck Based Data Correction in Continual Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..12e6217f8953bf92b8086770396d1e1595fff92b --- /dev/null +++ b/2024/Information Bottleneck Based Data Correction in Continual Learning/full.md @@ -0,0 +1,326 @@ +# Information Bottleneck Based Data Correction in Continual Learning + +Shuai Chen $^{1,2}$ , Mingyi Zhang $^{1}$ , Junge Zhang $^{1*}$ , and Kaiqi Huang $^{1,2,3*}$ + +$^{1}$ CRISE, Institute of Automation, Chinese Academy of Sciences + $^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences + $^{3}$ CAS Center for Excellence in Brain Science and Intelligence Technology + +Abstract. Continual Learning (CL) requires model to retain previously learned knowledge while learning new tasks. Recently, experience replay-based methods have made significant progress in addressing this challenge. These methods primarily select data from old tasks and store them in a buffer. When learning new task, they train the model using both the current and buffered data. However, the limited number of old data can lead to the model being influenced by new tasks. The repeated replaying of buffer data and the gradual discarding of old task data (un-sampled data) also result in a biased estimation of the model towards the old tasks, causing overfitting issues. All these factors can affect the CL performance. Therefore, we propose a data correction algorithm based on the Information Bottleneck (IBCL) to enhance the performance of the replay-based CL system. This algorithm comprises two components: the Information Bottleneck Task Agnostic Constraints (IBTA), which encourages the buffer data to learn task-relevant features related to the old tasks, thereby reducing the impact of new tasks. The Information Bottleneck Unsampled Data Surrogate (IBDS), which models the information of the unsampled data in the old tasks to alleviate data bias. Our method can be flexibly combined with most existing experience replay methods. We have verified the effectiveness of our method through a series of experiments, demonstrating its potential for improving the performance of CL algorithms. + +Keywords: Continual learning $\cdot$ Incremental learning $\cdot$ Experience replay + +# 1 Introduction + +Modern deep learning algorithms have achieved remarkable performance in a diverse range of tasks [48-50]. However, one of their significant limitations is the inability to retain knowledge from previously learned tasks when receiving training on new tasks, known as catastrophic forgetting [18, 25]. Addressing this issue has been a subject of extensive research in the field of continual learning (CL) [12, 36, 52, 53]. Among the proposed methods, the experience replay approaches have received significant attention [11, 13, 31, 54]. These methods aim to + +![](images/8085fa4ba7d2799a4275bbe20286b80cd1bdbd100f9f00759e6f7bcdafeea710.jpg) +Fig. 1: During the process of CL, when the old task is sampled, there is a consequent discarding of some unsampled data. This leads to bias in the model's estimation of the distribution of the old task. + +overcome catastrophic forgetting by selectively storing a subset of data from old tasks in the buffer and then combining buffer data with new task data for joint training. Various replay methods have been investigated from different perspectives to address the challenge of sample selection and balancing the contributions of new and old samples during the replay process [6,14,37]. + +Despite the progress made in reducing catastrophic forgetting through replay methods, certain limitations still exist. A major issue is that the memory of the buffer is too small, and the model cannot store enough data to restore previously learned knowledge [4, 21, 35]. This will give rise to two significant problems. Firstly, the scarcity of old data will result in an imbalance between the number of new and old tasks, leading to a disruption of knowledge from the old task during the learning process of the new task. Secondly, in the CL process, each time the old task data is sampled, some unsampled data will be discarded, introducing bias in the model's estimation of the old task distribution. Moreover, the repeated replaying of buffer data can contribute to overfitting of this specific data. All of these factors can significantly impact the CL performance. Currently, there are some methods to address the above issues, such as selecting more representative samples [20], increasing the diversity of old samples through data expansion [3], and improving the generalization ability of the model [5]. + +This work aims to tackle the aforementioned challenges from a different perspective. Regarding the first issue, we propose encouraging the data in the buffer to acquire task agnostic features related to the old task, thereby mitigating the impact of the new task. For the second issue, we focus on an important yet under-investigated question: How can unsampled data effectively contribute to learning new tasks? In Figure 1, we illustrate the cause of model overfitting, which arises because replay-based methods often utilize only a subset of data from old task, leading to biased results. When training on new tasks, we design the objective function as follows: + +$$ +\mathcal {L} _ {\text {n e w}} = \frac {1}{N _ {\text {b u f f e r}} + N _ {\text {n e w}}} \sum_ {i = 1} ^ {N _ {n}} \ell_ {i} (\hat {Y}, Y). \tag {1} +$$ + +where $\ell_i(\hat{Y}, Y)$ represents the loss function that takes the true outcome $Y$ and predicts outcome $\hat{Y}$ as inputs. $N_{buffer}$ and $N_{new}$ denote the numbers of buffer and new task data, respectively. $N_n$ is the sum of these two quantities. However, this loss function does not provide an unbiased estimate of the loss function for both new and old tasks, as the model gradually loses some data from past tasks during training. The unbiased loss function for new and old tasks should be formulated as follows: + +$$ +\mathbb {E} [ \mathcal {L} _ {\text {n e w}} ] \neq \mathcal {L} := \frac {1}{N _ {\text {o l d}} + N _ {\text {n e w}}} \sum_ {i = 1} ^ {N} \ell_ {i} (\hat {Y}, Y). \tag {2} +$$ + +where $N_{old}$ and $N_{new}$ represent the numbers of old task and new task data, respectively, and $N$ signifies the sum of these two quantities. To address the bias arising from the discarding of unsampled data from old task, our method aims to estimate the information of the unsampled data using specific strategies. However, unsampled data cannot directly participate in the training of new tasks, we indirectly obtain a surrogate for the information about unsampled data through the relationship between sampled and unsampled data. This information related to the unsampled data serves as a regularization term to constrain the training process of the model, thereby achieving bias correction for the dataset. + +Based on the intuition presented above, we propose an Information Bottleneck based data correction algorithm (IBCL). The information bottleneck method is a technique in information theory used to find the best compromise between accuracy and complexity [41]. It is regarded as the theoretical foundation of deep learning and has been applied in many fields, such as robust or invariant representation learning, causal reasoning, and so on [17]. This theory suggests that the model gradually eliminates irrelevant noise information during the learning process, retaining the most relevant feature information to the task. Borrowing the ability of this method, our algorithm mainly consists of two parts: Information Bottleneck Task Agnostic Constraints (IBTA): During the update process of new tasks, we introduce a constraint on the learning process of the buffer data. This constraint encourages the model to learn task agnostic features that are relevant to the old tasks, effectively preventing interference from the new tasks. Information Bottleneck Unsampled Data Surrogate (IBDS): This module is designed to model the information of the unsampled data. As the unsampled data cannot be directly trained, our method decouples the features of the unsampled data from the sampled data after the completion of learning old tasks. This decoupling allows model to leverage the relationship between the two types of data to estimate the surrogate influence of unsampled data. Consequently, this achieves an indirect implementation of the impact of unsampled data on learning new tasks. + +Overall, our work makes the following contributions: + +- We propose a data correction framework based on information bottleneck to mitigate model bias when the number of old task samples is small. + +- To mitigate the problem of the influence of new tasks on old sample features during the learning process, we propose IBTA to encourage the model to learn task agnostic features. + +- To counter the issue of overfitting caused by the repeated replay of buffer data and gradual discarding of unsampled data, we propose IBDS to incorporate the information from unsampled data to assist in the training of the new model. + +- Our method can be combined with most replay based methods to improve their performance. + +# 2 Related Work + +# 2.1 Experience Replay Continual Learning + +In the context of CL, there are three main approaches to retaining knowledge of past tasks: using a fixed-sized buffer for replaying past samples (replay-based approach) [2, 3, 7, 8, 43], regulating model parameter changes through learning (regularization-based approach) [16, 40, 42], or dynamically expanding model architecture as needed (expansion-based approach) [15, 28, 47]. Despite its simplicity, the replay-based approach has shown promising results in CL settings. Experience Replay (ER) [19] is a widely-used technique, which maintains a fixed-sized buffer to store and replay selected data, helping to mitigate the forgetting of previously learned knowledge. Incremental Classifier and Representation Learning (iCaRL) [37] uses the herding strategy to select a few representative samples. Rainbow Memory [3] proposes a diversity sampling strategy based on per-sample classification uncertainty. However, these methods primarily concentrate on how to choose which old data to store in the buffer, while neglecting the unsampled data. We argue that the unsampled data can also be beneficial in mitigating deviation and should be utilized. To this end, we propose a data correction method that employs the information bottleneck theory to model the information of the unsampled data in the old task, thereby enhancing the performance of several state-of-the-art methods. + +# 2.2 Information Bottleneck + +Information bottleneck (IB) is a classical concept in information theory, which is used to explain the behavior of deep learning [41]. Furthermore, IB has been extensively utilized in diverse applications, including but not limited to, representation learning decoupling [23, 33], recommendation systems [29, 45], and enhancing model robustness [17]. Recently, there have been some other works that apply information theory to CL, such as [27, 46], but there are fewer works based on the IB theory. The fundamental construct of IB is based on mutual information. In a previous work [20], the maximization of mutual information is employed to learn adequate feature expression. In contrast, our proposed method concentrates on modeling the unsampled data in the previous task. + +![](images/ff358228f58fb0907ffd9d15fd833630364b4c1bd2fd93c8aa1a9041b1ec48ae.jpg) +Fig. 2: During the updating of new tasks, we constrain the learning process of buffer data, encouraging the model to learn task agnostic features related to the old tasks, thus avoiding the interference of new tasks. + +# 3 Method + +Our method consists of two key processes: the learning process of new tasks and the processing process after learning old tasks. We introduce IBTA and IBDS separately. + +# 3.1 Problem Formulation + +In CL scenarios, tasks are introduced one after another and handled separately. We consider a total of $T$ tasks. Each task $D_{t} = \{x_{i}^{t},y_{i}^{t}\}_{i = 1}^{n_{t}}$ comprises $n_t$ input-target pairs $(x_i^t,y_i^t)$ , which are assumed to be independent and identically distributed (i.i.d.). The aim for each task is to train a model $f_{\theta}:X\to Y$ with parameters $\theta$ , capable of predicting the class label $y$ for a new image $x$ . In CL algorithms that utilize experience replay, the methods typically store some data from previous task data $x_{old}$ in a buffer. This stored data is referred to as sampled data $x_{m}$ , with its features denoted as $z_{m}$ . The data not selected for storage is called unsampled data $x_{u}$ , and its features are represented as $z_{u}$ . + +# 3.2 Information Bottleneck Task Agnostic Constraints + +During the learning of new tasks, the model focuses on the new task due to the limited data available in the buffer. In order to mitigate the impact of new tasks, our method encourages the buffer data to focus on learning the task-specific features, thus minimizing their influence. In order to compress the information in the dataset into the features of the sampled data as much as possible, the mutual information between the features $z_{m}$ of the sampled data and the buffer dataset $x_{m}$ should be minimized; In order to make the feature be able to distinguish well, the feature $z_{m}$ of the sample needs to predict the label $y_{old}$ as accurately as possible. The weighting parameters $\alpha$ is used to balance the contributions of different components in the overall objective function. + +$$ +\mathcal {L} _ {I B T A} := \min \underbrace {\alpha I \left(z _ {m} ; x _ {\mathrm {m}}\right)} _ {(1)} - \underbrace {I \left(z _ {m} ; y _ {\text {o l d}}\right)} _ {(2)}. \tag {3} +$$ + +In the given framework, the expression (1) represents a compression component that characterizes the mutual information between the buffer dataset $x_{m}$ and the sampled data features $z_{m}$ . This component aims to capture the relevance of $z_{m}$ to the old dataset $x_{m}$ . The expression (2) corresponds to an accuracy component that measures the effectiveness of $z_{m}$ in representing the underlying distribution of the dataset. It aims to ensure that the sampled features $z_{m}$ preserve sufficient information about the original data. Next, we will proceed to optimize the compression term $I(z_{m};x_{m})$ . This term can be quantified using the Kullback-Leibler (KL) divergence as follows: + +$$ +\begin{array}{l} I (z _ {m}; x _ {\mathrm {m}}) = \mathbb {E} _ {x} \left[ \mathrm {K L} (p (z _ {m} | x _ {\mathrm {m}}) \parallel p (z _ {m})) \right] \\ = \mathbb {E} _ {x} \left[ \int p \left(z _ {m} \mid x _ {\mathrm {m}}\right) \log p \left(z _ {m} \mid x _ {\mathrm {m}}\right) d z - \int p \left(z _ {m} \mid x _ {\mathrm {m}}\right) \log p \left(z _ {m}\right) d z \right]. \tag {4} \\ \end{array} +$$ + +Nonetheless, due to the computational complexity involved in calculating $p(z_{m}) = \int p(z_{m}|x_{m})p(x_{m})dx$ , we resort to the utilization of a variational approximation approach. Instead of using $p(z_{m})$ , we adopt a variational distribution denoted as $q(z_{m})$ . Leveraging Gibbs' inequality, we recognize the non-negativity of the Kullback-Leibler (KL) divergence. Thus, we can establish an upper bound for Equation (4), + +$$ +\begin{array}{l} - \int p \left(z _ {m} \mid x _ {\mathrm {m}}\right) \log p \left(z _ {m}\right) d z \leq - \int p \left(z _ {m} \mid x _ {\mathrm {m}}\right) \log q \left(z _ {m}\right) d z \tag {5} \\ \Rightarrow \operatorname {K L} \left(p \left(z _ {m} \mid x _ {\mathrm {m}}\right) \| p \left(z _ {m}\right)\right) \leq \operatorname {K L} \left(p \left(z _ {m} \mid x _ {\mathrm {m}}\right) \| q \left(z _ {m}\right)\right). \\ \end{array} +$$ + +Similar to prior research [1, 29, 34, 44, 45], we can make the assumption that the posterior distribution $p(z_{m} \mid x_{m})$ follows a Gaussian distribution and set it $\mathcal{N}(e(x_m),\mathrm{diag}(\sigma (x_m)))$ . In this context, $e(x_{m})$ represents the encoded embedding of the variables $x_{m}$ , and $\mathrm{diag}(\sigma (x_m))$ indicates the variance in a diagonal matrix form. By employing the reparameterization trick, the embedding $z_{m}$ can be generated through the formula $z_{m} = e(x_{m}) + \epsilon \odot \sigma (x_{m})$ , where $\epsilon \sim \mathcal{N}(0,I)$ . Notably, if we set $\sigma (x_{m})$ as an all-zero matrix, $z_{m}$ will effectively become a deterministic embedding. Conversely, the prior distribution $q(z_{m})$ is assumed to adhere to a standard Gaussian variational distribution, namely $q(z_{m}) = \mathcal{N}(0,I)$ . Ultimately, the previously mentioned upper bound can be rephrased as follows: + +$$ +\operatorname {K L} \left(p \left(z _ {m} \mid x _ {\mathrm {m}}\right) \| q \left(z _ {m}\right)\right) = \| e \left(x _ {\mathrm {m}}\right) \| _ {2} ^ {2} + \sum_ {d} \left(\sigma_ {d} - \frac {1}{2} \log \sigma_ {d} - 1\right). \tag {6} +$$ + +where $\sigma_d$ corresponds to an element within $\operatorname{diag}\{\sigma(x_m)\}$ . This signifies that for a deterministic embedding $z_m$ , we can enhance the optimization of this upper bound by directly implementing $\ell_2$ -norm regularization on the embedding vector $z_m$ . This process is equivalent to refining the compression term $I(z_m; x_m)$ . When considering the mutual information $I(z_m; y_{old})$ within Equation (3), we can express it as $I(z_m; y_{old}) = H(y_{old}) - H(y_{old} \mid z_m)$ . As $H(y_{old})$ is a positive constant that can be disregarded, the following inequality holds: + +$$ +I \left(z _ {m}; y _ {\text {o l d}}\right) \geq - H \left(y _ {\text {o l d}} \mid z _ {m}\right). \tag {7} +$$ + +![](images/65ac90514e9c443546e403e7d5158f0ce4fbb1558b966208a78394fa8412f3ea.jpg) +Fig. 3: This module is used to model the information of unsampled data. Because unsampled data cannot be trained directly, our method decouples the features of unsampled data and sampled data after the completion of the old task learning, so that it is convenient to estimate the surrogate loss of unsampled data in the new task by using the relationship between them, and indirectly realize the influence of unsampled data on the new task learning. + +This term can be seen as a cross entropy loss function. Therefore, we have implemented the constraint process for learning buffer data. However, the loss function obtained in this way is still biased, as there is missing unsampled data, which means that the contribution to the old task label $y_{old}$ should be $I(z; y) = I(z_m, z_u; y)$ , where $z_u$ is unsampled data feature. During the learning process of the new task, this item cannot be directly calculated. We will calculate it indirectly in the following section. + +# 3.3 Information Bottleneck Unsampled Data Surrogate + +To estimate the mutual information $I(z; y) = I(z_{m}, z_{u}; y)$ during the learning process of new task, it becomes essential to obtain $z_{u}$ . Given that $x_{u}$ denotes unsampled data, we cannot directly calculate $z_{u}$ from $x_{u}$ when learning new task. This necessitates an indirect estimation of $z_{u}$ from old task model. As shown in Fig.3, during the learning stage of old task, both sampled $x_{m}$ and unsampled $x_{u}$ data are involved in the training process. After training completion, the features corresponding to these two parts $z_{m}$ and $z_{u}$ are tightly coupled. Consequently, a specific operation is needed after training ends to decouple these two features. If we can decouple the features of sampled and unsampled data by optimizing $I(z_{m}; z_{u})$ , we can get $I(z; y) = I(z_{m}; y) + I(z_{u}; y)$ . In this case, the estimate of $I(z_{u}; y)$ can be used as a surrogate for the unsampled data information during learning new task. The feature $z_{m}$ of the sampled data and the feature $z_{u}$ of the unsampled data must maintain as much independence from one another as possible to obtain better discrimination. Drawing on principles from information theory, we are inspired to derive a minimization objective function based on the aforementioned analysis: + +$$ +\mathcal {L} _ {I B D S} := \min I \left(z _ {m}; z _ {u}\right). \tag {8} +$$ + +represents a decoupling component, capturing the level of interdependence between $z_{u}$ and $z_{m}$ . It aims to minimize the dependence of $z_{u}$ on $z_{m}$ , ensuring that $z_{u}$ provides additional information beyond what is already captured by $z_{m}$ . + +Inspired by previous research work [1, 29, 34, 44, 45], utilizing the chain rule of mutual information, we derive the subsequent equation pertaining to the decoupling component: + +$$ +I \left(z _ {m}; z _ {u}\right) = I \left(z _ {m}; y _ {o l d}\right) - I \left(z _ {m}; y _ {o l d} \mid z _ {u}\right) + I \left(z _ {m}; z _ {u} \mid y _ {o l d}\right). \tag {9} +$$ + +We proceed to analyze the expression $I(z_{m};z_{u}\mid y_{old})$ in greater detail. Given that the distribution of $z_{m}$ exclusively hinges on the variables $x_{m}$ , and the latter is influenced by $y_{old}$ , we establish $H(z_{m}\mid y_{old},z_{u}) = H(z_{m}\mid y_{old})$ , where $H(\cdot \mid \cdot)$ signifies conditional entropy. By combining the attributes of mutual information, we arrive at the following relationship: + +$$ +\begin{array}{l} I (z _ {m}; z _ {u} \mid y _ {\mathrm {o l d}}) = H (z _ {m} \mid y _ {\mathrm {o l d}}) - H (z _ {m} \mid y _ {\mathrm {o l d}}, z _ {u}) \\ = H \left(z _ {m} \mid y _ {\text {o l d}}\right) - H \left(z _ {m} \mid y _ {\text {o l d}}\right) = 0. \tag {10} \\ \end{array} +$$ + +By substituting Eq.(9), we have, + +$$ +I \left(z _ {m}; z _ {u}\right) = I \left(z _ {m}; y _ {\text {o l d}}\right) - I \left(z _ {m}; y _ {\text {o l d}} \mid z _ {u}\right) \tag {11} +$$ + +Since the term $I(z_{m};y_{old} \mid z_{u})$ in Eq.(11) is still difficult to be calculated, we simplify it further by using the form of conditional entropy: + +$$ +\begin{array}{l} I \left(z _ {m}; y _ {\text {o l d}} \mid z _ {u}\right) = H \left(z _ {m} \mid z _ {u}\right) + H \left(y _ {\text {o l d}} \mid z _ {u}\right) - I \left(z _ {m}; y _ {\text {o l d}} \mid z _ {u}\right) \\ = H \left(z _ {m} \mid z _ {u}\right) - H \left(z _ {m} \mid y _ {\text {o l d}}, z _ {u}\right) \tag {12} \\ = H \left(y _ {\text {o l d}} \mid z _ {u}\right) - H \left(y _ {\text {o l d}} \mid z _ {m}, z _ {u}\right) \\ \end{array} +$$ + +$$ +I \left(z _ {m}; z _ {u}\right) = I \left(z _ {m}; y _ {o l d}\right) - H \left(y _ {o l d} \mid z _ {u}\right) + H \left(y _ {o l d} \mid z _ {m}, z _ {u}\right). \tag {13} +$$ + +Finally, combining Eq.(7) and Eq.(13), we can rewrite the objective function $\mathcal{L}_{IBDS}$ in Eq.(8) as follows, + +$$ +\mathcal {L} _ {I B D S} = - H \left(y _ {o l d} \mid z _ {m}\right) - H \left(y _ {o l d} \mid z _ {u}\right) + H \left(y _ {o l d} \mid z _ {m}, z _ {u}\right). \tag {14} +$$ + +By optimizing $\mathcal{L}_{IBDS}$ , we get the decoupled expression $z_{u}$ . Now we can compute $I(z_{u};y)$ , which acts as a surrogate loss to compensate for missing weights due to unsampled data. + +During the training process of neural network, the model will encounter the problem of semantic drift [22,24,51], which means that the feature distribution of sampled data will change with the learning of new tasks. This concept highlights a significant challenge: using the fixed features of unsampled data as constraints in the learning of new task becomes problematic. As the model trains on new tasks, the estimation of unsampled data features gradually deviates from their true values, which not only affects the effectiveness of bias correction but also, to some extent, impacts the overall learning process. Just as the "no free lunch" + +theorem advocated in machine learning. Once samples are lost, it becomes difficult to perfectly estimate the distribution of unsampled data, the best one can do is to minimize the model's bias when addressing old tasks. To mitigate this impact, it's essential to consider the characteristics of neural networks in CL [9]. Early in the learning of new tasks, gradients undergo significant changes, and the model frequently encounters considerable biases. A viable strategy involves imposing stricter constraints on the new task learning process during the initial stages of model training. This approach instructs the model in reducing biases when learning new tasks. As the training advances, these constraints are gradually relaxed, permitting the model to better adjust to the learning requirements of new tasks during the slower update phases. To address the processing of unsampled data features, a dynamic decay strategy is employed to balance the learning conflicts between old and new tasks, thereby optimizing the model's overall performance. This method allows the model to discover a more balanced learning point between preserving old knowledge and acquiring new tasks, enhancing its learning efficiency and effectiveness. Specifically, inspired by [30], we propose a dynamic cosine annealing factor $\eta$ for constraint terms when updating unsampled data. + +$$ +\eta = \left(\frac {1}{2} (1 + \cos \left(\frac {e _ {\text {c u r}}}{e _ {\text {m a x}}}\right) \pi)\right) \tag {15} +$$ + +where $e_{cur}$ and $e_{max}$ represent current and maximum number of epochs. + +# 3.4 Algorithm + +By combining above two parts, we can get the total loss function as: + +$$ +\begin{array}{l} \mathcal {L} _ {C I B} = \alpha I (z _ {m}; x _ {\mathrm {m}}) - I (z; y _ {\mathrm {o l d}}) \\ = \alpha I \left(z _ {m}; x _ {\mathrm {m}}\right) - \beta I \left(z _ {u}; y\right) - \gamma I \left(z _ {m}; y\right) \tag {16} \\ \leq \alpha \| e \left(x _ {\mathrm {m}}\right) \| _ {2} ^ {2} + \beta \cdot \eta H \left(y _ {\text {o l d}} \mid z _ {u}\right) + \gamma H \left(y _ {\text {o l d}} \mid z _ {m}\right). \\ \end{array} +$$ + +The weighting parameters $\alpha$ , $\beta$ , $\gamma$ are used to balance the contributions of different components in the overall objective function. + +The final objective function comprises three terms: term (1) represents a regularization constraint on the input data, term (2) represents surrogate loss to reflects the influence of the unsampled data; term (3) denotes the cross-entropy between $z_{m}$ and $y_{old}$ , this can be seen as the cross-entropy loss function. The aim is to make the learned feature expression fully predict the final label. + +Discussion This approach comprises two primary modules: the IBTA module facilitates the learning of new tasks, whereas the IBDS module is utilized post-learning of old tasks to derive features from unsampled data. These features are then integrated as constraints within the training of new tasks, thereby shaping the overall optimization loss function. Our method can be combined with most replay-based strategies. Taking the naive ER method as an example, it randomly selects some samples from the old task and stores them in a buffer, which are trained together with the samples from the new task. Upon finalizing + +Table 1: Comparison of the average accuracy (the higher the better) of four split datasets with different buffer sizes. + +
MethodSplit CIFAR-10Split CIFAR-100Split ImageNet-100Split miniImageNet
Joint92.3873.2980.2353.55
FT19.679.298.689.52
Buffer size10020050020050020002005002000100020005000
ER36.3944.7957.7414.3519.6636.7613.6318.3734.258.3716.4924.17
+IBCL45.4351.9163.1223.9828.3242.0122.7228.9343.1114.7922.7227.69
ER-ACE53.9063.4170.5326.2836.4848.4124.2337.1249.5517.9522.6027.92
+IBCL63.8570.9774.8232.1240.9451.8930.7241.8152.6224.7727.6331.02
DER++57.6564.8872.7025.1137.1352.0826.5043.6558.0518.0223.4430.43
+IBCL66.4172.5276.6133.8244.2154.8932.6346.5759.9227.7231.4135.79
X-DER59.2965.1968.1035.3444.6254.4433.2146.7255.2325.2426.3829.91
+IBCL67.9173.8274.1443.8648.7255.9143.2549.6256.5929.9030.2132.66
+ +the learning of old task, the sampled data is processed through the old model to yield $z_{m}$ , and both sampled and unsampled data are processed to extract coupled features $z$ . Subsequently, by optimizing IBDS, features $z_{u}$ for each class of unsampled data are obtained. During the training of the new task, IBTA and $z_{u}$ are combined to optimize the total loss function for learning the new task. Detailed pseudocode of this procedure is provided in the appendix. + +# 4 Experiments + +In this section, we will present the datasets and evaluation metrics used in the experiment, and describe the implementation details. Then we describe the experimental results of our method on multiple standard benchmarks. Finally, we conduct an ablation study to illustrate the effectiveness of each module of our method. + +# 4.1 Dataset + +We validate our results on four distinct datasets: CIFAR-10, CIFAR-100, miniImageNet, and ImageNet-100. CIFAR-10 and CIFAR-100 each consist of 50,000 training samples and 10,000 testing samples, with 10 and 100 classes respectively. miniImageNet and ImageNet-100 are subsets of ImageNet, both containing 100 classes. miniImageNet includes 500 images per category for training, while ImageNet-100 includes 1300 images per category for training. + +# 4.2 Details + +To better compare with other methods, we follow the settings of [46] and [5]. For the CIFAR-10/100 and ImageNet-100 datasets, we adopt ResNet-18, and for the miniImageNet dataset, we adopt EfficientNet-B2. For CIFAR-10/100, we + +Table 2: Comparison of our method with other regularization methods on three datasets, we use two typical methods DER++ and ER-ACE as benchmarks for comparison. + +
MethodSplit CIFAR-10Split CIFAR-100Split miniImageNet
Buffer size1002005002005002000100020005000
ER-ACE53.9063.4170.5326.2836.4848.4117.9522.6027.92
+ sSGD56.2664.7371.4528.0739.5949.7018.1122.4324.12
+ oEwC52.3661.0968.7024.9335.0645.5919.0424.3229.46
+ oLAP52.7663.1970.3226.4236.5847.6618.3423.1928.77
+ OCM57.1864.6570.8628.1837.7449.0320.3224.3228.57
+ LiDER56.0865.3271.7527.9438.4350.3219.6924.1330.00
+ DualHSIC60.5268.0873.7829.0838.9450.5522.3325.4130.12
+ Our63.8570.9774.8232.1240.9451.8924.7727.6331.02
DER++57.6564.8872.7025.1137.1352.0818.0223.4430.43
+ sSGD55.8164.4472.0524.7638.4850.7416.3119.2924.24
+ oEwC55.7863.0271.6424.5135.2251.5318.8724.5331.91
+ oLAP54.8662.5471.3823.2634.4850.8018.9125.0232.78
+ OCM59.2565.8173.5327.4638.9452.2520.9324.7531.16
+ LiDER58.4366.0273.3927.3239.2553.2721.5828.3335.04
+ DualHSIC64.9870.2875.9431.4641.8653.5324.7829.3734.98
+ Our66.4172.5276.6133.8244.2154.8927.7231.4135.79
+ +adopt a training epoch of 50 and set the batch size to 32 and 64, respectively. For MiniImageNet and ImageNet-100, we adopt training epochs of 80 and 120, with batch size set to 128 and 64, respectively. In order to better compare the experimental results, we adopt the experimental results reported in [46] and [5], or the hyperparametric design in the original article. + +# 4.3 Evaluation Metrics + +We mainly evaluate on class incremental learning scenario. In this setup, each sequential task introduces a distinct set of classes that do not overlap with those from previous tasks. The challenge lies in training a classifier that can effectively classify all classes encountered up to the current stage, without the need for explicit task ids. We evaluate the final performance of the method by average accuracy. + +$$ +A C C = \frac {1}{T} \sum_ {i = 1} ^ {T} R _ {T, i}, \tag {17} +$$ + +where $R_{T,i}$ denotes the test accuracy on task $i$ after the model has finished task $T$ . + +Table 3: Abalation study of the average accuracy (the higher the better) of split ImageNet-100 and split CIFAR-100 with different buffer sizes. + +
MethodSplit CIFAR-100Split ImageNet-100
Buffer size20050020002005002000
ER-ACE26.2836.4848.4124.2337.1249.55
+ IBTA27.3637.4949.1225.6337.9250.12
+ IBDS30.4439.5250.9129.8340.7151.85
+IBCL32.1240.9451.8930.7241.8152.62
DER++25.1137.1352.0826.5043.6558.05
+IBTA27.8838.7453.9627.3945.7259.33
+ IBDS31.5342.5954.3329.5846.1559.70
+IBCL33.8244.2154.8932.6346.5759.92
+ +# 4.4 Baseline Models + +We conduct experiments on multiple state-of-the-art experience replay CL models: ER [10], DER++ [8], X-DER [6], and ER-ACE [9]. In addition to the original loss function based on replay, our method can be considered as a new regularization method. Therefore, we also compare our method with existing regularization-based techniques, including sSGD [32], oEWC [39], oLAP [38], and more recent SOTA methods, OCM [20], LiDER [5] and DualHSIC [46]. + +# 4.5 Results + +The comparison results are reported in Table 1. For the method of listing several baselines, our proposed method has improved the model's performance in various datasets and buffer sizes. In addition, we have the following main observations: 1) There has been a significant improvement in basic methods such as ER and DER++, as these two methods adopt more basic methods for sampling. Our method can reduce the bias caused by the sampling process to a certain extent. 2) Our methods have also shown some effectiveness on X-DER and ER-ACE, which take additional measures to alleviate sample imbalance and improve sample utilization, resulting in improved model performance. Our method further enhances the model's capabilities on top of this. + +To further demonstrate the effectiveness of our method, we compare our method with other regularization methods, and it can be found that we achieve the best performance, recent methods such as OCM, LiDER and DualHSIC achieve significant progress compared to baseline when the number of buffer is small, and our method further improves the performance when the buffer data is small by considering modeling for the unsampled data. + +# 4.6 Ablation Study + +In order to understand the effect of each of the items in the loss function on the performance of the algorithm, an ablation study has been carried out and the + +![](images/e17b5e35974228cf5ec5776a2af0dda49e6c658d92ad831dc2a12a503c651496.jpg) +Fig. 4: This figure shows the CKA matrix based on the ER algorithm, FT and incorporating our algorithm. It can be observed from it that due to our modeling of unsampled data, the model shows more information about the old task. + +![](images/6c4cc5c7089bf47269767d347ffd7406f8ca2b259b93520de11a4d1a7be87f2c.jpg) + +![](images/305984db984788dc1f90a3b3b41fd0744b949396a47aee300cf03cb7baf59ca2.jpg) + +results are shown in Table 3. It is observed that in most of the cases, increasing any of the components has an impact on the performance of the algorithm. However, increasing the IBDS gives a large performance improvement, which indicates that it plays a vital role in the performance of the algorithm. This component represents the loss of substitution of unsampled data, which is necessary for training on new tasks using unsampled data features. Therefore, if this aspect is not well learned, it may negatively affect the subsequent performance of the algorithm. IBTA also has an impact on the performance of the algorithm, the involves regularization constraints on the input features in order to make the model learn enough robust features. + +# 5 Characteristic Analysis + +We analyze various characteristics of the model to gain insight into the factors contributing to the performance enhancements achieved by our approach. + +# 5.1 CKA Matrix Analysis + +We use Central Kernel Alignment (CKA) [26] in Fig.4, which is an effective tool to measure the similarity of network representations and evaluate the relationships between these backbones. We send the same batch of instances to the ER (2000 buffer size) algorithm model and fine-tuning based model respectively to obtain corresponding feature maps, thus intuitively demonstrating the similarity of algorithm feature maps. Our algorithm can be used in conjunction with the experience replay method to model unsampled data. When the sampled data is 0, i.e. fine-tuning, our method can also get some benefits. + +![](images/f776ee0a368bf83264a3019df57278fd23a062b9eca311ea385184ae50334934.jpg) +Initialization + +![](images/9deb011b3e4c1988a3809b8798e4c7015966b38b5352829e6520b0359a26812f.jpg) +Epoch $= 20$ +Fig. 5: This image uses TSNE to display the changes in the characteristics of sampled and unsampled data with the number of training epochs of the model. We have shown one class of situation, with a sample data size of 200 and an unsampled data size of 1100. + +![](images/e3e5fdfec78084223105586c8257c1a5fb7c74e63c28ad3d3144c282a36632b5.jpg) +Epoch $= 50$ + +# 5.2 Feature Decoupling Analysis + +A significant advance of our method is the decoupling of sampled and unsampled data features. We demonstrate this in Fig.5, where we use TSNE technology to visualize the changes in these two features with the increase of network training epochs. We test our method, the ER algorithm, on 200 sampled data and 1100 unsampled data from the ImageNet-100 dataset. As the network training epochs increase, we can observe that the two sets of features begin to separate from their initial coupling, which indicates that our method has achieved decoupling of the two data. + +# 6 Conclusion + +In conclusion, our proposed method is designed to tackle the issue of data bias in continual learning through the application of experience replay. Drawing inspiration from information bottleneck theory, we steer the incorporation of unsampled data information, thereby bolstering the model's capacity to preserve past knowledge while accommodating novel tasks. Our conducted experiments offer empirical validation of the advantages conferred by our approach. While we have presented a method for computing unsampled data, it's important to acknowledge the existence of several alternative strategies for estimating these. This opens up exciting possibilities for continued exploration and innovation in this field. + +# Acknowledgements + +This work was supported by National Science and Technology Major Project, Grant No.2022ZD0116403, the Youth Innovation Promotion Association CAS and Strategic Priority Research Program of Chinese Academy of Sciences XDA27010300. + +# References + +1. Alemi, A.A., Fischer, I., Dillon, J.V., Murphy, K.: Deep variational information bottleneck. arXiv (2016) +2. Aljundi, R., Lin, M., Goujaud, B., Bengio, Y.: Gradient based sample selection for online continual learning. In: Adv. Neural Inform. Process. Syst. (2019) +3. Bang, J., Kim, H., Yoo, Y., Ha, J.W., Choi, J.: Rainbow memory: Continual learning with a memory of diverse samples. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) +4. Biesialska, M., Biesialska, K., Costa-Jussa, M.R.: Continual lifelong learning in natural language processing: A survey. arXiv (2020) +5. Bonicelli, L., Boschini, M., Porrello, A., Spampinato, C., Calderara, S.: On the effectiveness of lipschitz-driven rehearsal in continual learning. In: Adv. Neural Inform. Process. Syst. (2022) +6. Boschini, M., Bonicelli, L., Buzzega, P., Porrello, A., Calderara, S.: Class incremental continual learning into the extended der-verse. IEEE Trans. Pattern Anal. Mach. Intell. (2022) +7. Boschini, M., Buzzega, P., Bonicelli, L., Porrello, A., Calderara, S.: Continual semisupervised learning through contrastive interpolation consistency. Pattern Recognition Letters 162, 9-14 (2022) +8. Buzzega, P., Boschini, M., Porrello, A., Abati, D., Calderara, S.: Dark experience for general continual learning: a strong, simple baseline. In: Adv. Neural Inform. Process. Syst. (2020) +9. Caccia, L., Aljundi, R., Asadi, N., Tuytelaars, T., Pineau, J., Belilovsky, E.: New insights on reducing abrupt representation change in online continual learning. In: Int. Conf. Learn. Represent. (2022) +0. Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H.S., Ranzato, M.: Continual learning with tiny episodic memories. Arxiv (2019) +1. Chen, S., Zhang, M., Zhang, J., Huang, K.: Exemplar-based continual learning via contrastive learning. IEEE Transactions on Artificial Intelligence 5(7), 3313-3324 (2024) +2. Cossu, A., Carta, A., Lomonaco, V., Bacciu, D.: Continual learning for recurrent neural networks: an empirical evaluation. Neural Networks (2021) +3. Delange, M., Aljundi, R., Masana, M., Parisot, S., Jia, X., Leonardis, A., Slabaugh, G., Tuytelaars, T.: A continual learning survey: Defying forgetting in classification tasks. IEEE Trans. Pattern Anal. Mach. Intell. (2021) +4. Douillard, A., Cord, M., Ollion, C., Robert, T., Valle, E.: PODNet: Pooled Outputs Distillation for Small-Tasks Incremental Learning. In: Eur. Conf. Comput. Vis. (2020) +5. Douillard, A., Ram, A., Couairon, G., Cord, M.: Dytox: Transformers for continual learning with dynamic token expansion. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) +6. Ebrahimi, S., Elhoseiny, M., Darrell, T., Rohrbach, M.: Uncertainty-guided continual learning with bayesian neural networks. In: Int. Conf. Learn. Represent. (2020) +7. Federici, M., Dutta, A., Forre, P., Kushman, N., Akata, Z.: Learning robust representations via multi-view information bottleneck. In: Int. Conf. Learn. Represent. (2020) +8. French, R.M.: Catastrophic forgetting in connectionist networks. Trends in cognitive sciences (1999) + +19. Goodfellow, I.J., Mirza, M., Xiao, D., Courville, A., Bengio, Y.: An empirical investigation of catastrophic forgetting in gradient-based neural networks. arXiv (2015) +20. Guo, Y., Liu, B., Zhao, D.: Online continual learning through mutual information maximization. In: Int. Conf. Mach. Learn. (2022) +21. Hadsell, R., Rao, D., Rusu, A.A., Pascanu, R.: Embracing change: Continual learning in deep neural networks. Trends in cognitive sciences (2020) +22. Iscen, A., Zhang, J., Lazebnik, S., Schmid, C.: Memory-efficient incremental learning through feature adaptation. In: Eur. Conf. Comput. Vis. (2020) +23. Jaiswal, A., Brekelmans, R., Moyer, D., Steeg, G.V., AbdAlmageed, W., Natarajan, P.: Discovery and separation of features for invariant representation learning. arXiv (2019) +24. Jie, S., Deng, Z.H., Li, Z.: Alleviating representational shift for continual fin-tuning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) +25. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proc. of the national academy of sciences (2017) +26. Kornblith, S., Norouzi, M., Lee, H., Hinton, G.E.: Similarity of neural network representations revisited. In: Int. Conf. Mach. Learn. (2019) +27. Li, D., Wang, T., Chen, J., Ren, Q., Kawaguchi, K., Zeng, Z.: Towards continual learning desiderata via hsic-bottleneck orthogonalization and equiangular embedding. In: AAAI (2024) +28. Li, X., Zhou, Y., Wu, T., Socher, R., Xiong, C.: Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In: Int. Conf. Mach. Learn. (2019) +29. Liu, D., Cheng, P., Zhu, H., Dong, Z., He, X., Pan, W., Ming, Z.: Mitigating confounding bias in recommendation via information bottleneck. In: Proceedings of the 15th ACM Conference on Recommender Systems (2021) +30. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv (2016) +31. Masana, M., Liu, X., Twardowski, B., Menta, M., Bagdanov, A.D., van de Weijer, J.: Class-incremental learning: Survey and performance evaluation on image classification. IEEE Trans. Pattern Anal. Mach. Intell. (2023) +32. Mirzadeh, S.I., Farajtabar, M., Pascanu, R., Ghasemzadeh, H.: Understanding the role of training regimes in continual learning. Arxiv (2020) +33. Moyer, D., Gao, S., Brekelmans, R., Steeg, G.V., Galstyan, A.: Evading the adversary in invariant representation. CoRR (2018) +34. Pan, Z., Niu, L., Zhang, J., Zhang, L.: Disentangled information bottleneck. In: AAAI (2021) +35. Parisi, G.I., Kemker, R., Part, J.L., Kanan, C., Wermter, S.: Continual lifelong learning with neural networks: A review. Neural Networks (2019) +36. Qu, H., Rahmani, H., Xu, L., Williams, B., Liu, J.: Recent advances of continual learning in computer vision: An overview. arXiv (2021) +37. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017) +38. Ritter, H., Botev, A., Barber, D.: Online structured laplace approximations for overcoming catastrophic forgetting. In: Adv. Neural Inform. Process. Syst. (2018) +39. Schuster, M., Paliwal, K.K.: Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing 45(11), 2673-2681 (1997) + +40. Swaroop, S., Nguyen, C.V., Bui, T.D., Turner, R.E.: Improving and understanding variational continual learning. ArXiv (2019) +41. Tishby, N., Pereira, F.C., Bialek, W.: The information bottleneck method (2000) +42. Titsias, M.K., Schwarz, J., Matthews, A.G.d.G., Pascanu, R., Teh, Y.W.: Functional regularisation for continual learning with gaussian processes. In: Int. Conf. Learn. Represent. (2019) +43. Von Oswald, J., Zhao, D., Kobayashi, S., Schug, S., Caccia, M., Zucchet, N., Sacramento, J.: Learning where to learn: Gradient sparsity in meta and continual learning. In: Adv. Neural Inform. Process. Syst. (2021) +44. Wang, Y., Rudner, T.G., Wilson, A.G.: Visual explanations of image-text representations via multi-modal information bottleneck attribution. In: Adv. Neural Inform. Process. Syst. (2024) +45. Wang, Z., Chen, X., Wen, R., Huang, S.L., Kuruoglu, E.E., Zheng, Y.: Information theoretic counterfactual learning from missing-not-at-random feedback. In: Adv. Neural Inform. Process. Syst. (2020) +46. Wang, Z., Zhan, Z., Gong, Y., Shao, Y., Ioannidis, S., Wang, Y., Dy, J.: Dualhsic: Hsic-bottleneck and alignment for continual learning. In: Int. Conf. Mach. Learn. (2023) +47. Wortsman, M., Ramanujan, V., Liu, R., Kembhavi, A., Rastegari, M., Yosinski, J., Farhadi, A.: Supermasks in superposition. In: Adv. Neural Inform. Process. Syst. (2020) +48. Xu, P., Zhang, J., Huang, K.: Exploration via joint policy diversity for sparse-reward multi-agent tasks. In: IJCAI (2023) +49. Xu, P., Zhang, J., Huang, K.: Population-based diverse exploration for sparse-reward multi-agent tasks. In: IJCAI (2024) +50. Xu, P., Zhang, J., Yin, Q., Yu, C., Yang, Y., Huang, K.: Subspace-aware exploration for sparse-reward multi-agent tasks. In: AAAI (2023) +51. Yu, L., Twardowski, B., Liu, X., Herranz, L., Wang, K., Cheng, Y., Jui, S., Weijer, J.v.d.: Semantic drift compensation for class-incremental learning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020) +52. Zhang, X., Yi, J., Tao, J., Wang, C., Zhang, C.Y.: Do you remember? overcoming catastrophic forgetting for fake audio detection. In: Int. Conf. Mach. Learn. (2023) +53. Zhang, X., Yi, J., Wang, C., Zhang, C.Y., Zeng, S., Tao, J.: What to remember: Self-adaptive continual learning for audio deepfake detection. In: AAAI (2024) +54. Zhou, D.W., Wang, Q.W., Qi, Z.H., Ye, H.J., Zhan, D.C., Liu, Z.: Deep class-incremental learning: A survey. arXiv (2023) \ No newline at end of file diff --git a/2024/Information Bottleneck Based Data Correction in Continual Learning/images.zip b/2024/Information Bottleneck Based Data Correction in Continual Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..255241f9dd6d42288566d5418f4f12acf34f16fa --- /dev/null +++ b/2024/Information Bottleneck Based Data Correction in Continual Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c81b2372ea4cda0aa5c3148e210625f35ab512920aefc6d5ff0e8db79ccf8af +size 543842 diff --git a/2024/Information Bottleneck Based Data Correction in Continual Learning/layout.json b/2024/Information Bottleneck Based Data Correction in Continual Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..51fb5e6b2dc9521870891aef61e03d6872331141 --- /dev/null +++ b/2024/Information Bottleneck Based Data Correction in Continual Learning/layout.json @@ -0,0 +1,9603 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "text", + "content": "Information Bottleneck Based Data Correction in Continual Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 167, + 476, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 167, + 476, + 181 + ], + "spans": [ + { + "bbox": [ + 138, + 167, + 476, + 181 + ], + "type": "text", + "content": "Shuai Chen" + }, + { + "bbox": [ + 138, + 167, + 476, + 181 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 138, + 167, + 476, + 181 + ], + "type": "text", + "content": ", Mingyi Zhang" + }, + { + "bbox": [ + 138, + 167, + 476, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 167, + 476, + 181 + ], + "type": "text", + "content": ", Junge Zhang" + }, + { + "bbox": [ + 138, + 167, + 476, + 181 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 138, + 167, + 476, + 181 + ], + "type": "text", + "content": ", and Kaiqi Huang" + }, + { + "bbox": [ + 138, + 167, + 476, + 181 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 149, + 190, + 463, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 190, + 463, + 223 + ], + "spans": [ + { + "bbox": [ + 149, + 190, + 463, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 190, + 463, + 223 + ], + "type": "text", + "content": " CRISE, Institute of Automation, Chinese Academy of Sciences \n" + }, + { + "bbox": [ + 149, + 190, + 463, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 190, + 463, + 223 + ], + "type": "text", + "content": " School of Artificial Intelligence, University of Chinese Academy of Sciences \n" + }, + { + "bbox": [ + 149, + 190, + 463, + 223 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 149, + 190, + 463, + 223 + ], + "type": "text", + "content": " CAS Center for Excellence in Brain Science and Intelligence Technology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 247, + 455, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 247, + 455, + 490 + ], + "spans": [ + { + "bbox": [ + 160, + 247, + 455, + 490 + ], + "type": "text", + "content": "Abstract. Continual Learning (CL) requires model to retain previously learned knowledge while learning new tasks. Recently, experience replay-based methods have made significant progress in addressing this challenge. These methods primarily select data from old tasks and store them in a buffer. When learning new task, they train the model using both the current and buffered data. However, the limited number of old data can lead to the model being influenced by new tasks. The repeated replaying of buffer data and the gradual discarding of old task data (un-sampled data) also result in a biased estimation of the model towards the old tasks, causing overfitting issues. All these factors can affect the CL performance. Therefore, we propose a data correction algorithm based on the Information Bottleneck (IBCL) to enhance the performance of the replay-based CL system. This algorithm comprises two components: the Information Bottleneck Task Agnostic Constraints (IBTA), which encourages the buffer data to learn task-relevant features related to the old tasks, thereby reducing the impact of new tasks. The Information Bottleneck Unsampled Data Surrogate (IBDS), which models the information of the unsampled data in the old tasks to alleviate data bias. Our method can be flexibly combined with most existing experience replay methods. We have verified the effectiveness of our method through a series of experiments, demonstrating its potential for improving the performance of CL algorithms." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 499, + 453, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 499, + 453, + 521 + ], + "spans": [ + { + "bbox": [ + 159, + 499, + 453, + 521 + ], + "type": "text", + "content": "Keywords: Continual learning " + }, + { + "bbox": [ + 159, + 499, + 453, + 521 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 499, + 453, + 521 + ], + "type": "text", + "content": " Incremental learning " + }, + { + "bbox": [ + 159, + 499, + 453, + 521 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 499, + 453, + 521 + ], + "type": "text", + "content": " Experience replay" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 540, + 230, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 540, + 230, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 540, + 230, + 553 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 563, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 648 + ], + "type": "text", + "content": "Modern deep learning algorithms have achieved remarkable performance in a diverse range of tasks [48-50]. However, one of their significant limitations is the inability to retain knowledge from previously learned tasks when receiving training on new tasks, known as catastrophic forgetting [18, 25]. Addressing this issue has been a subject of extensive research in the field of continual learning (CL) [12, 36, 52, 53]. Among the proposed methods, the experience replay approaches have received significant attention [11, 13, 31, 54]. These methods aim to" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 654, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 236, + 665 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 116, + 470, + 207 + ], + "blocks": [ + { + "bbox": [ + 142, + 116, + 470, + 207 + ], + "lines": [ + { + "bbox": [ + 142, + 116, + 470, + 207 + ], + "spans": [ + { + "bbox": [ + 142, + 116, + 470, + 207 + ], + "type": "image", + "image_path": "8085fa4ba7d2799a4275bbe20286b80cd1bdbd100f9f00759e6f7bcdafeea710.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 217, + 482, + 251 + ], + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 251 + ], + "type": "text", + "content": "Fig. 1: During the process of CL, when the old task is sampled, there is a consequent discarding of some unsampled data. This leads to bias in the model's estimation of the distribution of the old task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 280, + 479, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 479, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 479, + 341 + ], + "type": "text", + "content": "overcome catastrophic forgetting by selectively storing a subset of data from old tasks in the buffer and then combining buffer data with new task data for joint training. Various replay methods have been investigated from different perspectives to address the challenge of sample selection and balancing the contributions of new and old samples during the replay process [6,14,37]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 343, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 511 + ], + "type": "text", + "content": "Despite the progress made in reducing catastrophic forgetting through replay methods, certain limitations still exist. A major issue is that the memory of the buffer is too small, and the model cannot store enough data to restore previously learned knowledge [4, 21, 35]. This will give rise to two significant problems. Firstly, the scarcity of old data will result in an imbalance between the number of new and old tasks, leading to a disruption of knowledge from the old task during the learning process of the new task. Secondly, in the CL process, each time the old task data is sampled, some unsampled data will be discarded, introducing bias in the model's estimation of the old task distribution. Moreover, the repeated replaying of buffer data can contribute to overfitting of this specific data. All of these factors can significantly impact the CL performance. Currently, there are some methods to address the above issues, such as selecting more representative samples [20], increasing the diversity of old samples through data expansion [3], and improving the generalization ability of the model [5]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 513, + 482, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 513, + 482, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 513, + 482, + 621 + ], + "type": "text", + "content": "This work aims to tackle the aforementioned challenges from a different perspective. Regarding the first issue, we propose encouraging the data in the buffer to acquire task agnostic features related to the old task, thereby mitigating the impact of the new task. For the second issue, we focus on an important yet under-investigated question: How can unsampled data effectively contribute to learning new tasks? In Figure 1, we illustrate the cause of model overfitting, which arises because replay-based methods often utilize only a subset of data from old task, leading to biased results. When training on new tasks, we design the objective function as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 223, + 635, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 635, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 223, + 635, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {n e w}} = \\frac {1}{N _ {\\text {b u f f e r}} + N _ {\\text {n e w}}} \\sum_ {i = 1} ^ {N _ {n}} \\ell_ {i} (\\hat {Y}, Y). \\tag {1}", + "image_path": "e726442cc5fb3792aed77f67388134f338646da068d1b3e55afa4a8cbf9b768a.jpg" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "S. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "inline_equation", + "content": "\\ell_i(\\hat{Y}, Y)" + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "text", + "content": " represents the loss function that takes the true outcome " + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "text", + "content": " and predicts outcome " + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "text", + "content": " as inputs. " + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "inline_equation", + "content": "N_{buffer}" + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "inline_equation", + "content": "N_{new}" + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "text", + "content": " denote the numbers of buffer and new task data, respectively. " + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "inline_equation", + "content": "N_n" + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "text", + "content": " is the sum of these two quantities. However, this loss function does not provide an unbiased estimate of the loss function for both new and old tasks, as the model gradually loses some data from past tasks during training. The unbiased loss function for new and old tasks should be formulated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 215, + 211, + 481, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 211, + 481, + 242 + ], + "spans": [ + { + "bbox": [ + 215, + 211, + 481, + 242 + ], + "type": "interline_equation", + "content": "\\mathbb {E} [ \\mathcal {L} _ {\\text {n e w}} ] \\neq \\mathcal {L} := \\frac {1}{N _ {\\text {o l d}} + N _ {\\text {n e w}}} \\sum_ {i = 1} ^ {N} \\ell_ {i} (\\hat {Y}, Y). \\tag {2}", + "image_path": "15ec0b33b2c122aef4a4b3d8bde6307594369d267776766ea0cfa4d8809fbd74.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 254, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 254, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 480, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 254, + 480, + 361 + ], + "type": "inline_equation", + "content": "N_{old}" + }, + { + "bbox": [ + 130, + 254, + 480, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 254, + 480, + 361 + ], + "type": "inline_equation", + "content": "N_{new}" + }, + { + "bbox": [ + 130, + 254, + 480, + 361 + ], + "type": "text", + "content": " represent the numbers of old task and new task data, respectively, and " + }, + { + "bbox": [ + 130, + 254, + 480, + 361 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 254, + 480, + 361 + ], + "type": "text", + "content": " signifies the sum of these two quantities. To address the bias arising from the discarding of unsampled data from old task, our method aims to estimate the information of the unsampled data using specific strategies. However, unsampled data cannot directly participate in the training of new tasks, we indirectly obtain a surrogate for the information about unsampled data through the relationship between sampled and unsampled data. This information related to the unsampled data serves as a regularization term to constrain the training process of the model, thereby achieving bias correction for the dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 363, + 481, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 481, + 614 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 481, + 614 + ], + "type": "text", + "content": "Based on the intuition presented above, we propose an Information Bottleneck based data correction algorithm (IBCL). The information bottleneck method is a technique in information theory used to find the best compromise between accuracy and complexity [41]. It is regarded as the theoretical foundation of deep learning and has been applied in many fields, such as robust or invariant representation learning, causal reasoning, and so on [17]. This theory suggests that the model gradually eliminates irrelevant noise information during the learning process, retaining the most relevant feature information to the task. Borrowing the ability of this method, our algorithm mainly consists of two parts: Information Bottleneck Task Agnostic Constraints (IBTA): During the update process of new tasks, we introduce a constraint on the learning process of the buffer data. This constraint encourages the model to learn task agnostic features that are relevant to the old tasks, effectively preventing interference from the new tasks. Information Bottleneck Unsampled Data Surrogate (IBDS): This module is designed to model the information of the unsampled data. As the unsampled data cannot be directly trained, our method decouples the features of the unsampled data from the sampled data after the completion of learning old tasks. This decoupling allows model to leverage the relationship between the two types of data to estimate the surrogate influence of unsampled data. Consequently, this achieves an indirect implementation of the impact of unsampled data on learning new tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 616, + 378, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 616, + 378, + 627 + ], + "spans": [ + { + "bbox": [ + 146, + 616, + 378, + 627 + ], + "type": "text", + "content": "Overall, our work makes the following contributions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 641, + 480, + 665 + ], + "type": "text", + "content": "- We propose a data correction framework based on information bottleneck to mitigate model bias when the number of old task samples is small." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 447, + 101 + ], + "type": "text", + "content": "IB Based Data Correction in CL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 479, + 151 + ], + "type": "text", + "content": "- To mitigate the problem of the influence of new tasks on old sample features during the learning process, we propose IBTA to encourage the model to learn task agnostic features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 152, + 480, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 152, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 138, + 152, + 480, + 198 + ], + "type": "text", + "content": "- To counter the issue of overfitting caused by the repeated replay of buffer data and gradual discarding of unsampled data, we propose IBDS to incorporate the information from unsampled data to assist in the training of the new model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 200, + 479, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 200, + 479, + 223 + ], + "spans": [ + { + "bbox": [ + 137, + 200, + 479, + 223 + ], + "type": "text", + "content": "- Our method can be combined with most replay based methods to improve their performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 243, + 237, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 243, + 237, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 237, + 255 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 270, + 356, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 356, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 356, + 282 + ], + "type": "text", + "content": "2.1 Experience Replay Continual Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 290, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 506 + ], + "type": "text", + "content": "In the context of CL, there are three main approaches to retaining knowledge of past tasks: using a fixed-sized buffer for replaying past samples (replay-based approach) [2, 3, 7, 8, 43], regulating model parameter changes through learning (regularization-based approach) [16, 40, 42], or dynamically expanding model architecture as needed (expansion-based approach) [15, 28, 47]. Despite its simplicity, the replay-based approach has shown promising results in CL settings. Experience Replay (ER) [19] is a widely-used technique, which maintains a fixed-sized buffer to store and replay selected data, helping to mitigate the forgetting of previously learned knowledge. Incremental Classifier and Representation Learning (iCaRL) [37] uses the herding strategy to select a few representative samples. Rainbow Memory [3] proposes a diversity sampling strategy based on per-sample classification uncertainty. However, these methods primarily concentrate on how to choose which old data to store in the buffer, while neglecting the unsampled data. We argue that the unsampled data can also be beneficial in mitigating deviation and should be utilized. To this end, we propose a data correction method that employs the information bottleneck theory to model the information of the unsampled data in the old task, thereby enhancing the performance of several state-of-the-art methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 525, + 278, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 278, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 278, + 536 + ], + "type": "text", + "content": "2.2 Information Bottleneck" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Information bottleneck (IB) is a classical concept in information theory, which is used to explain the behavior of deep learning [41]. Furthermore, IB has been extensively utilized in diverse applications, including but not limited to, representation learning decoupling [23, 33], recommendation systems [29, 45], and enhancing model robustness [17]. Recently, there have been some other works that apply information theory to CL, such as [27, 46], but there are fewer works based on the IB theory. The fundamental construct of IB is based on mutual information. In a previous work [20], the maximization of mutual information is employed to learn adequate feature expression. In contrast, our proposed method concentrates on modeling the unsampled data in the previous task." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "S. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 118, + 479, + 213 + ], + "blocks": [ + { + "bbox": [ + 136, + 118, + 479, + 213 + ], + "lines": [ + { + "bbox": [ + 136, + 118, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 136, + 118, + 479, + 213 + ], + "type": "image", + "image_path": "ff358228f58fb0907ffd9d15fd833630364b4c1bd2fd93c8aa1a9041b1ec48ae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 223, + 482, + 257 + ], + "lines": [ + { + "bbox": [ + 130, + 223, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 482, + 257 + ], + "type": "text", + "content": "Fig. 2: During the updating of new tasks, we constrain the learning process of buffer data, encouraging the model to learn task agnostic features related to the old tasks, thus avoiding the interference of new tasks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 277, + 202, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 202, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 202, + 290 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 479, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 479, + 337 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 479, + 337 + ], + "type": "text", + "content": "Our method consists of two key processes: the learning process of new tasks and the processing process after learning old tasks. We introduce IBTA and IBDS separately." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 352, + 268, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 268, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 268, + 363 + ], + "type": "text", + "content": "3.1 Problem Formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "text", + "content": "In CL scenarios, tasks are introduced one after another and handled separately. We consider a total of " + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "text", + "content": " tasks. Each task " + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "inline_equation", + "content": "D_{t} = \\{x_{i}^{t},y_{i}^{t}\\}_{i = 1}^{n_{t}}" + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "text", + "content": " comprises " + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "inline_equation", + "content": "n_t" + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "text", + "content": " input-target pairs " + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "inline_equation", + "content": "(x_i^t,y_i^t)" + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "text", + "content": ", which are assumed to be independent and identically distributed (i.i.d.). The aim for each task is to train a model " + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "inline_equation", + "content": "f_{\\theta}:X\\to Y" + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "text", + "content": ", capable of predicting the class label " + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "text", + "content": " for a new image " + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "text", + "content": ". In CL algorithms that utilize experience replay, the methods typically store some data from previous task data " + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "inline_equation", + "content": "x_{old}" + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "text", + "content": " in a buffer. This stored data is referred to as sampled data " + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "text", + "content": ", with its features denoted as " + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "inline_equation", + "content": "z_{m}" + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "text", + "content": ". The data not selected for storage is called unsampled data " + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "inline_equation", + "content": "x_{u}" + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "text", + "content": ", and its features are represented as " + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "inline_equation", + "content": "z_{u}" + }, + { + "bbox": [ + 130, + 370, + 482, + 479 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 494, + 414, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 414, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 414, + 506 + ], + "type": "text", + "content": "3.2 Information Bottleneck Task Agnostic Constraints" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 512, + 482, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 482, + 632 + ], + "type": "text", + "content": "During the learning of new tasks, the model focuses on the new task due to the limited data available in the buffer. In order to mitigate the impact of new tasks, our method encourages the buffer data to focus on learning the task-specific features, thus minimizing their influence. In order to compress the information in the dataset into the features of the sampled data as much as possible, the mutual information between the features " + }, + { + "bbox": [ + 130, + 512, + 482, + 632 + ], + "type": "inline_equation", + "content": "z_{m}" + }, + { + "bbox": [ + 130, + 512, + 482, + 632 + ], + "type": "text", + "content": " of the sampled data and the buffer dataset " + }, + { + "bbox": [ + 130, + 512, + 482, + 632 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 130, + 512, + 482, + 632 + ], + "type": "text", + "content": " should be minimized; In order to make the feature be able to distinguish well, the feature " + }, + { + "bbox": [ + 130, + 512, + 482, + 632 + ], + "type": "inline_equation", + "content": "z_{m}" + }, + { + "bbox": [ + 130, + 512, + 482, + 632 + ], + "type": "text", + "content": " of the sample needs to predict the label " + }, + { + "bbox": [ + 130, + 512, + 482, + 632 + ], + "type": "inline_equation", + "content": "y_{old}" + }, + { + "bbox": [ + 130, + 512, + 482, + 632 + ], + "type": "text", + "content": " as accurately as possible. The weighting parameters " + }, + { + "bbox": [ + 130, + 512, + 482, + 632 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 512, + 482, + 632 + ], + "type": "text", + "content": " is used to balance the contributions of different components in the overall objective function." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 217, + 638, + 481, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 638, + 481, + 668 + ], + "spans": [ + { + "bbox": [ + 217, + 638, + 481, + 668 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I B T A} := \\min \\underbrace {\\alpha I \\left(z _ {m} ; x _ {\\mathrm {m}}\\right)} _ {(1)} - \\underbrace {I \\left(z _ {m} ; y _ {\\text {o l d}}\\right)} _ {(2)}. \\tag {3}", + "image_path": "872f394a27210def720add431dc092f94a60832a5157a24c157dfbecd11380c5.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 447, + 100 + ], + "type": "text", + "content": "IB Based Data Correction in CL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "In the given framework, the expression (1) represents a compression component that characterizes the mutual information between the buffer dataset " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " and the sampled data features " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "z_{m}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ". This component aims to capture the relevance of " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "z_{m}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " to the old dataset " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ". The expression (2) corresponds to an accuracy component that measures the effectiveness of " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "z_{m}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " in representing the underlying distribution of the dataset. It aims to ensure that the sampled features " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "z_{m}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " preserve sufficient information about the original data. Next, we will proceed to optimize the compression term " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "I(z_{m};x_{m})" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ". This term can be quantified using the Kullback-Leibler (KL) divergence as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 232, + 480, + 272 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 232, + 480, + 272 + ], + "spans": [ + { + "bbox": [ + 140, + 232, + 480, + 272 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} I (z _ {m}; x _ {\\mathrm {m}}) = \\mathbb {E} _ {x} \\left[ \\mathrm {K L} (p (z _ {m} | x _ {\\mathrm {m}}) \\parallel p (z _ {m})) \\right] \\\\ = \\mathbb {E} _ {x} \\left[ \\int p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) \\log p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) d z - \\int p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) \\log p \\left(z _ {m}\\right) d z \\right]. \\tag {4} \\\\ \\end{array}", + "image_path": "194dd122fd7408733ad4bcace633ea74cdaa8c2b8ba8eda5672505a7313dd223.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 277, + 483, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 277, + 483, + 350 + ], + "spans": [ + { + "bbox": [ + 131, + 277, + 483, + 350 + ], + "type": "text", + "content": "Nonetheless, due to the computational complexity involved in calculating " + }, + { + "bbox": [ + 131, + 277, + 483, + 350 + ], + "type": "inline_equation", + "content": "p(z_{m}) = \\int p(z_{m}|x_{m})p(x_{m})dx" + }, + { + "bbox": [ + 131, + 277, + 483, + 350 + ], + "type": "text", + "content": ", we resort to the utilization of a variational approximation approach. Instead of using " + }, + { + "bbox": [ + 131, + 277, + 483, + 350 + ], + "type": "inline_equation", + "content": "p(z_{m})" + }, + { + "bbox": [ + 131, + 277, + 483, + 350 + ], + "type": "text", + "content": ", we adopt a variational distribution denoted as " + }, + { + "bbox": [ + 131, + 277, + 483, + 350 + ], + "type": "inline_equation", + "content": "q(z_{m})" + }, + { + "bbox": [ + 131, + 277, + 483, + 350 + ], + "type": "text", + "content": ". Leveraging Gibbs' inequality, we recognize the non-negativity of the Kullback-Leibler (KL) divergence. Thus, we can establish an upper bound for Equation (4)," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 188, + 357, + 480, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 357, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 188, + 357, + 480, + 396 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} - \\int p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) \\log p \\left(z _ {m}\\right) d z \\leq - \\int p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) \\log q \\left(z _ {m}\\right) d z \\tag {5} \\\\ \\Rightarrow \\operatorname {K L} \\left(p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) \\| p \\left(z _ {m}\\right)\\right) \\leq \\operatorname {K L} \\left(p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) \\| q \\left(z _ {m}\\right)\\right). \\\\ \\end{array}", + "image_path": "47066fcd25eeb48d43caaa8973b40b65526b654cbdd5c147667fab2934267a2b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "content": "Similar to prior research [1, 29, 34, 44, 45], we can make the assumption that the posterior distribution " + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "inline_equation", + "content": "p(z_{m} \\mid x_{m})" + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "content": " follows a Gaussian distribution and set it " + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(e(x_m),\\mathrm{diag}(\\sigma (x_m)))" + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "content": ". In this context, " + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "inline_equation", + "content": "e(x_{m})" + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "content": " represents the encoded embedding of the variables " + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathrm{diag}(\\sigma (x_m))" + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "content": " indicates the variance in a diagonal matrix form. By employing the reparameterization trick, the embedding " + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "inline_equation", + "content": "z_{m}" + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "content": " can be generated through the formula " + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "inline_equation", + "content": "z_{m} = e(x_{m}) + \\epsilon \\odot \\sigma (x_{m})" + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "content": ". Notably, if we set " + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\sigma (x_{m})" + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "content": " as an all-zero matrix, " + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "inline_equation", + "content": "z_{m}" + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "content": " will effectively become a deterministic embedding. Conversely, the prior distribution " + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "inline_equation", + "content": "q(z_{m})" + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "content": " is assumed to adhere to a standard Gaussian variational distribution, namely " + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "inline_equation", + "content": "q(z_{m}) = \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "content": ". Ultimately, the previously mentioned upper bound can be rephrased as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 171, + 528, + 480, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 528, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 171, + 528, + 480, + 556 + ], + "type": "interline_equation", + "content": "\\operatorname {K L} \\left(p \\left(z _ {m} \\mid x _ {\\mathrm {m}}\\right) \\| q \\left(z _ {m}\\right)\\right) = \\| e \\left(x _ {\\mathrm {m}}\\right) \\| _ {2} ^ {2} + \\sum_ {d} \\left(\\sigma_ {d} - \\frac {1}{2} \\log \\sigma_ {d} - 1\\right). \\tag {6}", + "image_path": "cf2ca92726cdf59feae4c9a05194c8525ef059f89f9726a10c5ca9bfce9b1976.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\sigma_d" + }, + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "text", + "content": " corresponds to an element within " + }, + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\operatorname{diag}\\{\\sigma(x_m)\\}" + }, + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "text", + "content": ". This signifies that for a deterministic embedding " + }, + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "inline_equation", + "content": "z_m" + }, + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "text", + "content": ", we can enhance the optimization of this upper bound by directly implementing " + }, + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "text", + "content": "-norm regularization on the embedding vector " + }, + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "inline_equation", + "content": "z_m" + }, + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "text", + "content": ". This process is equivalent to refining the compression term " + }, + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "inline_equation", + "content": "I(z_m; x_m)" + }, + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "text", + "content": ". When considering the mutual information " + }, + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "inline_equation", + "content": "I(z_m; y_{old})" + }, + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "text", + "content": " within Equation (3), we can express it as " + }, + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "inline_equation", + "content": "I(z_m; y_{old}) = H(y_{old}) - H(y_{old} \\mid z_m)" + }, + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "text", + "content": ". As " + }, + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "inline_equation", + "content": "H(y_{old})" + }, + { + "bbox": [ + 130, + 562, + 482, + 647 + ], + "type": "text", + "content": " is a positive constant that can be disregarded, the following inequality holds:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 246, + 654, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 654, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 246, + 654, + 480, + 666 + ], + "type": "interline_equation", + "content": "I \\left(z _ {m}; y _ {\\text {o l d}}\\right) \\geq - H \\left(y _ {\\text {o l d}} \\mid z _ {m}\\right). \\tag {7}", + "image_path": "20b5e69a1f4cdf2936ca4c9496030e2d51ec7ff9b8deb06217110a3bf3aea3cc.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "S. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 118, + 477, + 211 + ], + "blocks": [ + { + "bbox": [ + 136, + 118, + 477, + 211 + ], + "lines": [ + { + "bbox": [ + 136, + 118, + 477, + 211 + ], + "spans": [ + { + "bbox": [ + 136, + 118, + 477, + 211 + ], + "type": "image", + "image_path": "65ac90514e9c443546e403e7d5158f0ce4fbb1558b966208a78394fa8412f3ea.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 223, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 130, + 223, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 482, + 289 + ], + "type": "text", + "content": "Fig. 3: This module is used to model the information of unsampled data. Because unsampled data cannot be trained directly, our method decouples the features of unsampled data and sampled data after the completion of the old task learning, so that it is convenient to estimate the surrogate loss of unsampled data in the new task by using the relationship between them, and indirectly realize the influence of unsampled data on the new task learning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 316, + 482, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 400 + ], + "type": "text", + "content": "This term can be seen as a cross entropy loss function. Therefore, we have implemented the constraint process for learning buffer data. However, the loss function obtained in this way is still biased, as there is missing unsampled data, which means that the contribution to the old task label " + }, + { + "bbox": [ + 130, + 316, + 482, + 400 + ], + "type": "inline_equation", + "content": "y_{old}" + }, + { + "bbox": [ + 130, + 316, + 482, + 400 + ], + "type": "text", + "content": " should be " + }, + { + "bbox": [ + 130, + 316, + 482, + 400 + ], + "type": "inline_equation", + "content": "I(z; y) = I(z_m, z_u; y)" + }, + { + "bbox": [ + 130, + 316, + 482, + 400 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 316, + 482, + 400 + ], + "type": "inline_equation", + "content": "z_u" + }, + { + "bbox": [ + 130, + 316, + 482, + 400 + ], + "type": "text", + "content": " is unsampled data feature. During the learning process of the new task, this item cannot be directly calculated. We will calculate it indirectly in the following section." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 423, + 419, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 423, + 419, + 436 + ], + "spans": [ + { + "bbox": [ + 131, + 423, + 419, + 436 + ], + "type": "text", + "content": "3.3 Information Bottleneck Unsampled Data Surrogate" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "text", + "content": "To estimate the mutual information " + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "inline_equation", + "content": "I(z; y) = I(z_{m}, z_{u}; y)" + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "text", + "content": " during the learning process of new task, it becomes essential to obtain " + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "inline_equation", + "content": "z_{u}" + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "text", + "content": ". Given that " + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "inline_equation", + "content": "x_{u}" + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "text", + "content": " denotes unsampled data, we cannot directly calculate " + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "inline_equation", + "content": "z_{u}" + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "inline_equation", + "content": "x_{u}" + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "text", + "content": " when learning new task. This necessitates an indirect estimation of " + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "inline_equation", + "content": "z_{u}" + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "text", + "content": " from old task model. As shown in Fig.3, during the learning stage of old task, both sampled " + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "text", + "content": " and unsampled " + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "inline_equation", + "content": "x_{u}" + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "text", + "content": " data are involved in the training process. After training completion, the features corresponding to these two parts " + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "inline_equation", + "content": "z_{m}" + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "inline_equation", + "content": "z_{u}" + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "text", + "content": " are tightly coupled. Consequently, a specific operation is needed after training ends to decouple these two features. If we can decouple the features of sampled and unsampled data by optimizing " + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "inline_equation", + "content": "I(z_{m}; z_{u})" + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "text", + "content": ", we can get " + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "inline_equation", + "content": "I(z; y) = I(z_{m}; y) + I(z_{u}; y)" + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "text", + "content": ". In this case, the estimate of " + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "inline_equation", + "content": "I(z_{u}; y)" + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "text", + "content": " can be used as a surrogate for the unsampled data information during learning new task. The feature " + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "inline_equation", + "content": "z_{m}" + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "text", + "content": " of the sampled data and the feature " + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "inline_equation", + "content": "z_{u}" + }, + { + "bbox": [ + 130, + 449, + 482, + 641 + ], + "type": "text", + "content": " of the unsampled data must maintain as much independence from one another as possible to obtain better discrimination. Drawing on principles from information theory, we are inspired to derive a minimization objective function based on the aforementioned analysis:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 251, + 653, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 653, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 251, + 653, + 481, + 666 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I B D S} := \\min I \\left(z _ {m}; z _ {u}\\right). \\tag {8}", + "image_path": "f64a60d79b60aef97ee3a6a742ff982393c6ff7b668fefbd24910e4aabecedd4.jpg" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 447, + 100 + ], + "type": "text", + "content": "IB Based Data Correction in CL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "content": "represents a decoupling component, capturing the level of interdependence between " + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "z_{u}" + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "z_{m}" + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "content": ". It aims to minimize the dependence of " + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "z_{u}" + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "z_{m}" + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "content": ", ensuring that " + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "z_{u}" + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "content": " provides additional information beyond what is already captured by " + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "z_{m}" + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 481, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 481, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 481, + 189 + ], + "type": "text", + "content": "Inspired by previous research work [1, 29, 34, 44, 45], utilizing the chain rule of mutual information, we derive the subsequent equation pertaining to the decoupling component:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 180, + 198, + 481, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 198, + 481, + 212 + ], + "spans": [ + { + "bbox": [ + 180, + 198, + 481, + 212 + ], + "type": "interline_equation", + "content": "I \\left(z _ {m}; z _ {u}\\right) = I \\left(z _ {m}; y _ {o l d}\\right) - I \\left(z _ {m}; y _ {o l d} \\mid z _ {u}\\right) + I \\left(z _ {m}; z _ {u} \\mid y _ {o l d}\\right). \\tag {9}", + "image_path": "a09d11f433cc52f39a5e54b8ee8c989b814ec6bf6e685cc1c8a39dece91d1094.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 220, + 481, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 220, + 481, + 280 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 481, + 280 + ], + "type": "text", + "content": "We proceed to analyze the expression " + }, + { + "bbox": [ + 131, + 220, + 481, + 280 + ], + "type": "inline_equation", + "content": "I(z_{m};z_{u}\\mid y_{old})" + }, + { + "bbox": [ + 131, + 220, + 481, + 280 + ], + "type": "text", + "content": " in greater detail. Given that the distribution of " + }, + { + "bbox": [ + 131, + 220, + 481, + 280 + ], + "type": "inline_equation", + "content": "z_{m}" + }, + { + "bbox": [ + 131, + 220, + 481, + 280 + ], + "type": "text", + "content": " exclusively hinges on the variables " + }, + { + "bbox": [ + 131, + 220, + 481, + 280 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 131, + 220, + 481, + 280 + ], + "type": "text", + "content": ", and the latter is influenced by " + }, + { + "bbox": [ + 131, + 220, + 481, + 280 + ], + "type": "inline_equation", + "content": "y_{old}" + }, + { + "bbox": [ + 131, + 220, + 481, + 280 + ], + "type": "text", + "content": ", we establish " + }, + { + "bbox": [ + 131, + 220, + 481, + 280 + ], + "type": "inline_equation", + "content": "H(z_{m}\\mid y_{old},z_{u}) = H(z_{m}\\mid y_{old})" + }, + { + "bbox": [ + 131, + 220, + 481, + 280 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 220, + 481, + 280 + ], + "type": "inline_equation", + "content": "H(\\cdot \\mid \\cdot)" + }, + { + "bbox": [ + 131, + 220, + 481, + 280 + ], + "type": "text", + "content": " signifies conditional entropy. By combining the attributes of mutual information, we arrive at the following relationship:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 198, + 290, + 480, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 290, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 198, + 290, + 480, + 317 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} I (z _ {m}; z _ {u} \\mid y _ {\\mathrm {o l d}}) = H (z _ {m} \\mid y _ {\\mathrm {o l d}}) - H (z _ {m} \\mid y _ {\\mathrm {o l d}}, z _ {u}) \\\\ = H \\left(z _ {m} \\mid y _ {\\text {o l d}}\\right) - H \\left(z _ {m} \\mid y _ {\\text {o l d}}\\right) = 0. \\tag {10} \\\\ \\end{array}", + "image_path": "ec452ba5d42c89e02683f2a4429aa53bba86ce27726a68feca4fade0fafbd686.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 327, + 277, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 277, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 277, + 339 + ], + "type": "text", + "content": "By substituting Eq.(9), we have," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 220, + 349, + 480, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 349, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 220, + 349, + 480, + 361 + ], + "type": "interline_equation", + "content": "I \\left(z _ {m}; z _ {u}\\right) = I \\left(z _ {m}; y _ {\\text {o l d}}\\right) - I \\left(z _ {m}; y _ {\\text {o l d}} \\mid z _ {u}\\right) \\tag {11}", + "image_path": "9c4811b34c09ec41ef1a78712a7e57fa2c17c95e222e67de9f5898edc744545b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 369, + 481, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 481, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 481, + 395 + ], + "type": "text", + "content": "Since the term " + }, + { + "bbox": [ + 130, + 369, + 481, + 395 + ], + "type": "inline_equation", + "content": "I(z_{m};y_{old} \\mid z_{u})" + }, + { + "bbox": [ + 130, + 369, + 481, + 395 + ], + "type": "text", + "content": " in Eq.(11) is still difficult to be calculated, we simplify it further by using the form of conditional entropy:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 176, + 404, + 480, + 446 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 404, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 176, + 404, + 480, + 446 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} I \\left(z _ {m}; y _ {\\text {o l d}} \\mid z _ {u}\\right) = H \\left(z _ {m} \\mid z _ {u}\\right) + H \\left(y _ {\\text {o l d}} \\mid z _ {u}\\right) - I \\left(z _ {m}; y _ {\\text {o l d}} \\mid z _ {u}\\right) \\\\ = H \\left(z _ {m} \\mid z _ {u}\\right) - H \\left(z _ {m} \\mid y _ {\\text {o l d}}, z _ {u}\\right) \\tag {12} \\\\ = H \\left(y _ {\\text {o l d}} \\mid z _ {u}\\right) - H \\left(y _ {\\text {o l d}} \\mid z _ {m}, z _ {u}\\right) \\\\ \\end{array}", + "image_path": "49ba126861786ea7b607817ed4f010c0c631dca64f544e9cacb0cac5b809d541.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 185, + 460, + 480, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 460, + 480, + 473 + ], + "spans": [ + { + "bbox": [ + 185, + 460, + 480, + 473 + ], + "type": "interline_equation", + "content": "I \\left(z _ {m}; z _ {u}\\right) = I \\left(z _ {m}; y _ {o l d}\\right) - H \\left(y _ {o l d} \\mid z _ {u}\\right) + H \\left(y _ {o l d} \\mid z _ {m}, z _ {u}\\right). \\tag {13}", + "image_path": "e717509b69d2b23ac3e011f6c3a041dcdd7198c82615b276e239d23e27800ec6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 478, + 480, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 478, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 478, + 480, + 503 + ], + "type": "text", + "content": "Finally, combining Eq.(7) and Eq.(13), we can rewrite the objective function " + }, + { + "bbox": [ + 132, + 478, + 480, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{IBDS}" + }, + { + "bbox": [ + 132, + 478, + 480, + 503 + ], + "type": "text", + "content": " in Eq.(8) as follows," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 182, + 513, + 480, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 513, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 182, + 513, + 480, + 525 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I B D S} = - H \\left(y _ {o l d} \\mid z _ {m}\\right) - H \\left(y _ {o l d} \\mid z _ {u}\\right) + H \\left(y _ {o l d} \\mid z _ {m}, z _ {u}\\right). \\tag {14}", + "image_path": "a3798e3bb8bb9faf0765d092f719ed90025849f21551276e826af8b287957251.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 533, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 481, + 569 + ], + "type": "text", + "content": "By optimizing " + }, + { + "bbox": [ + 132, + 533, + 481, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{IBDS}" + }, + { + "bbox": [ + 132, + 533, + 481, + 569 + ], + "type": "text", + "content": ", we get the decoupled expression " + }, + { + "bbox": [ + 132, + 533, + 481, + 569 + ], + "type": "inline_equation", + "content": "z_{u}" + }, + { + "bbox": [ + 132, + 533, + 481, + 569 + ], + "type": "text", + "content": ". Now we can compute " + }, + { + "bbox": [ + 132, + 533, + 481, + 569 + ], + "type": "inline_equation", + "content": "I(z_{u};y)" + }, + { + "bbox": [ + 132, + 533, + 481, + 569 + ], + "type": "text", + "content": ", which acts as a surrogate loss to compensate for missing weights due to unsampled data." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": "During the training process of neural network, the model will encounter the problem of semantic drift [22,24,51], which means that the feature distribution of sampled data will change with the learning of new tasks. This concept highlights a significant challenge: using the fixed features of unsampled data as constraints in the learning of new task becomes problematic. As the model trains on new tasks, the estimation of unsampled data features gradually deviates from their true values, which not only affects the effectiveness of bias correction but also, to some extent, impacts the overall learning process. Just as the \"no free lunch\"" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "S. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "type": "text", + "content": "theorem advocated in machine learning. Once samples are lost, it becomes difficult to perfectly estimate the distribution of unsampled data, the best one can do is to minimize the model's bias when addressing old tasks. To mitigate this impact, it's essential to consider the characteristics of neural networks in CL [9]. Early in the learning of new tasks, gradients undergo significant changes, and the model frequently encounters considerable biases. A viable strategy involves imposing stricter constraints on the new task learning process during the initial stages of model training. This approach instructs the model in reducing biases when learning new tasks. As the training advances, these constraints are gradually relaxed, permitting the model to better adjust to the learning requirements of new tasks during the slower update phases. To address the processing of unsampled data features, a dynamic decay strategy is employed to balance the learning conflicts between old and new tasks, thereby optimizing the model's overall performance. This method allows the model to discover a more balanced learning point between preserving old knowledge and acquiring new tasks, enhancing its learning efficiency and effectiveness. Specifically, inspired by [30], we propose a dynamic cosine annealing factor " + }, + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 116, + 482, + 331 + ], + "type": "text", + "content": " for constraint terms when updating unsampled data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 249, + 331, + 481, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 331, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 249, + 331, + 481, + 357 + ], + "type": "interline_equation", + "content": "\\eta = \\left(\\frac {1}{2} (1 + \\cos \\left(\\frac {e _ {\\text {c u r}}}{e _ {\\text {m a x}}}\\right) \\pi)\\right) \\tag {15}", + "image_path": "16b85f9aa694120f5b45dab99307e7cd04d00550d375fd7d9f658f2c846775ba.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 361, + 451, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 361, + 451, + 373 + ], + "spans": [ + { + "bbox": [ + 131, + 361, + 451, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 361, + 451, + 373 + ], + "type": "inline_equation", + "content": "e_{cur}" + }, + { + "bbox": [ + 131, + 361, + 451, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 361, + 451, + 373 + ], + "type": "inline_equation", + "content": "e_{max}" + }, + { + "bbox": [ + 131, + 361, + 451, + 373 + ], + "type": "text", + "content": " represent current and maximum number of epochs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 391, + 212, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 391, + 212, + 403 + ], + "spans": [ + { + "bbox": [ + 131, + 391, + 212, + 403 + ], + "type": "text", + "content": "3.4 Algorithm" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 411, + 434, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 411, + 434, + 424 + ], + "spans": [ + { + "bbox": [ + 131, + 411, + 434, + 424 + ], + "type": "text", + "content": "By combining above two parts, we can get the total loss function as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 188, + 434, + 480, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 434, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 188, + 434, + 480, + 477 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {C I B} = \\alpha I (z _ {m}; x _ {\\mathrm {m}}) - I (z; y _ {\\mathrm {o l d}}) \\\\ = \\alpha I \\left(z _ {m}; x _ {\\mathrm {m}}\\right) - \\beta I \\left(z _ {u}; y\\right) - \\gamma I \\left(z _ {m}; y\\right) \\tag {16} \\\\ \\leq \\alpha \\| e \\left(x _ {\\mathrm {m}}\\right) \\| _ {2} ^ {2} + \\beta \\cdot \\eta H \\left(y _ {\\text {o l d}} \\mid z _ {u}\\right) + \\gamma H \\left(y _ {\\text {o l d}} \\mid z _ {m}\\right). \\\\ \\end{array}", + "image_path": "1db5b817423742ce09cec59a8d9f30be47c63990e5de065464128f02573fdb04.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 485, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 481, + 510 + ], + "type": "text", + "content": "The weighting parameters " + }, + { + "bbox": [ + 130, + 485, + 481, + 510 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 485, + 481, + 510 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 485, + 481, + 510 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 485, + 481, + 510 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 485, + 481, + 510 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 485, + 481, + 510 + ], + "type": "text", + "content": " are used to balance the contributions of different components in the overall objective function." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "text", + "content": "The final objective function comprises three terms: term (1) represents a regularization constraint on the input data, term (2) represents surrogate loss to reflects the influence of the unsampled data; term (3) denotes the cross-entropy between " + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "inline_equation", + "content": "z_{m}" + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "inline_equation", + "content": "y_{old}" + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "text", + "content": ", this can be seen as the cross-entropy loss function. The aim is to make the learned feature expression fully predict the final label." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Discussion This approach comprises two primary modules: the IBTA module facilitates the learning of new tasks, whereas the IBDS module is utilized post-learning of old tasks to derive features from unsampled data. These features are then integrated as constraints within the training of new tasks, thereby shaping the overall optimization loss function. Our method can be combined with most replay-based strategies. Taking the naive ER method as an example, it randomly selects some samples from the old task and stores them in a buffer, which are trained together with the samples from the new task. Upon finalizing" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 447, + 101 + ], + "type": "text", + "content": "IB Based Data Correction in CL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 146, + 482, + 285 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 1: Comparison of the average accuracy (the higher the better) of four split datasets with different buffer sizes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 146, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 136, + 146, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 136, + 146, + 482, + 285 + ], + "type": "table", + "html": "
MethodSplit CIFAR-10Split CIFAR-100Split ImageNet-100Split miniImageNet
Joint92.3873.2980.2353.55
FT19.679.298.689.52
Buffer size10020050020050020002005002000100020005000
ER36.3944.7957.7414.3519.6636.7613.6318.3734.258.3716.4924.17
+IBCL45.4351.9163.1223.9828.3242.0122.7228.9343.1114.7922.7227.69
ER-ACE53.9063.4170.5326.2836.4848.4124.2337.1249.5517.9522.6027.92
+IBCL63.8570.9774.8232.1240.9451.8930.7241.8152.6224.7727.6331.02
DER++57.6564.8872.7025.1137.1352.0826.5043.6558.0518.0223.4430.43
+IBCL66.4172.5276.6133.8244.2154.8932.6346.5759.9227.7231.4135.79
X-DER59.2965.1968.1035.3444.6254.4433.2146.7255.2325.2426.3829.91
+IBCL67.9173.8274.1443.8648.7255.9143.2549.6256.5929.9030.2132.66
", + "image_path": "48ef1e08468ed897951a3379377e38bca32a21e4cea047a7fdbf52ad2f8d4678.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 307, + 480, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 480, + 380 + ], + "type": "text", + "content": "the learning of old task, the sampled data is processed through the old model to yield " + }, + { + "bbox": [ + 130, + 307, + 480, + 380 + ], + "type": "inline_equation", + "content": "z_{m}" + }, + { + "bbox": [ + 130, + 307, + 480, + 380 + ], + "type": "text", + "content": ", and both sampled and unsampled data are processed to extract coupled features " + }, + { + "bbox": [ + 130, + 307, + 480, + 380 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 307, + 480, + 380 + ], + "type": "text", + "content": ". Subsequently, by optimizing IBDS, features " + }, + { + "bbox": [ + 130, + 307, + 480, + 380 + ], + "type": "inline_equation", + "content": "z_{u}" + }, + { + "bbox": [ + 130, + 307, + 480, + 380 + ], + "type": "text", + "content": " for each class of unsampled data are obtained. During the training of the new task, IBTA and " + }, + { + "bbox": [ + 130, + 307, + 480, + 380 + ], + "type": "inline_equation", + "content": "z_{u}" + }, + { + "bbox": [ + 130, + 307, + 480, + 380 + ], + "type": "text", + "content": " are combined to optimize the total loss function for learning the new task. Detailed pseudocode of this procedure is provided in the appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 397, + 230, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 230, + 411 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 230, + 411 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 422, + 480, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 480, + 482 + ], + "type": "text", + "content": "In this section, we will present the datasets and evaluation metrics used in the experiment, and describe the implementation details. Then we describe the experimental results of our method on multiple standard benchmarks. Finally, we conduct an ablation study to illustrate the effectiveness of each module of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 500, + 201, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 201, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 201, + 511 + ], + "type": "text", + "content": "4.1 Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 520, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 480, + 593 + ], + "type": "text", + "content": "We validate our results on four distinct datasets: CIFAR-10, CIFAR-100, miniImageNet, and ImageNet-100. CIFAR-10 and CIFAR-100 each consist of 50,000 training samples and 10,000 testing samples, with 10 and 100 classes respectively. miniImageNet and ImageNet-100 are subsets of ImageNet, both containing 100 classes. miniImageNet includes 500 images per category for training, while ImageNet-100 includes 1300 images per category for training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 609, + 196, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 196, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 196, + 620 + ], + "type": "text", + "content": "4.2 Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 666 + ], + "type": "text", + "content": "To better compare with other methods, we follow the settings of [46] and [5]. For the CIFAR-10/100 and ImageNet-100 datasets, we adopt ResNet-18, and for the miniImageNet dataset, we adopt EfficientNet-B2. For CIFAR-10/100, we" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "S. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 157, + 475, + 379 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Comparison of our method with other regularization methods on three datasets, we use two typical methods DER++ and ER-ACE as benchmarks for comparison." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 157, + 475, + 379 + ], + "lines": [ + { + "bbox": [ + 140, + 157, + 475, + 379 + ], + "spans": [ + { + "bbox": [ + 140, + 157, + 475, + 379 + ], + "type": "table", + "html": "
MethodSplit CIFAR-10Split CIFAR-100Split miniImageNet
Buffer size1002005002005002000100020005000
ER-ACE53.9063.4170.5326.2836.4848.4117.9522.6027.92
+ sSGD56.2664.7371.4528.0739.5949.7018.1122.4324.12
+ oEwC52.3661.0968.7024.9335.0645.5919.0424.3229.46
+ oLAP52.7663.1970.3226.4236.5847.6618.3423.1928.77
+ OCM57.1864.6570.8628.1837.7449.0320.3224.3228.57
+ LiDER56.0865.3271.7527.9438.4350.3219.6924.1330.00
+ DualHSIC60.5268.0873.7829.0838.9450.5522.3325.4130.12
+ Our63.8570.9774.8232.1240.9451.8924.7727.6331.02
DER++57.6564.8872.7025.1137.1352.0818.0223.4430.43
+ sSGD55.8164.4472.0524.7638.4850.7416.3119.2924.24
+ oEwC55.7863.0271.6424.5135.2251.5318.8724.5331.91
+ oLAP54.8662.5471.3823.2634.4850.8018.9125.0232.78
+ OCM59.2565.8173.5327.4638.9452.2520.9324.7531.16
+ LiDER58.4366.0273.3927.3239.2553.2721.5828.3335.04
+ DualHSIC64.9870.2875.9431.4641.8653.5324.7829.3734.98
+ Our66.4172.5276.6133.8244.2154.8927.7231.4135.79
", + "image_path": "039be141bd4bba34a17cb65083d22e77e756b69b1f0cd790cafd3ca0ed886846.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 403, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 464 + ], + "type": "text", + "content": "adopt a training epoch of 50 and set the batch size to 32 and 64, respectively. For MiniImageNet and ImageNet-100, we adopt training epochs of 80 and 120, with batch size set to 128 and 64, respectively. In order to better compare the experimental results, we adopt the experimental results reported in [46] and [5], or the hyperparametric design in the original article." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 487, + 257, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 257, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 257, + 499 + ], + "type": "text", + "content": "4.3 Evaluation Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 514, + 482, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 587 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 587 + ], + "type": "text", + "content": "We mainly evaluate on class incremental learning scenario. In this setup, each sequential task introduces a distinct set of classes that do not overlap with those from previous tasks. The challenge lies in training a classifier that can effectively classify all classes encountered up to the current stage, without the need for explicit task ids. We evaluate the final performance of the method by average accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 261, + 601, + 481, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 601, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 261, + 601, + 481, + 634 + ], + "type": "interline_equation", + "content": "A C C = \\frac {1}{T} \\sum_ {i = 1} ^ {T} R _ {T, i}, \\tag {17}", + "image_path": "216ae06cf6121c464f24fd34e43549f327b2d4fa39e37847d954bc607e084274.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 641, + 481, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 664 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 641, + 481, + 664 + ], + "type": "inline_equation", + "content": "R_{T,i}" + }, + { + "bbox": [ + 131, + 641, + 481, + 664 + ], + "type": "text", + "content": " denotes the test accuracy on task " + }, + { + "bbox": [ + 131, + 641, + 481, + 664 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 641, + 481, + 664 + ], + "type": "text", + "content": " after the model has finished task " + }, + { + "bbox": [ + 131, + 641, + 481, + 664 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 641, + 481, + 664 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 447, + 101 + ], + "type": "text", + "content": "IB Based Data Correction in CL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 146, + 480, + 280 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 3: Abalation study of the average accuracy (the higher the better) of split ImageNet-100 and split CIFAR-100 with different buffer sizes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 146, + 480, + 280 + ], + "lines": [ + { + "bbox": [ + 134, + 146, + 480, + 280 + ], + "spans": [ + { + "bbox": [ + 134, + 146, + 480, + 280 + ], + "type": "table", + "html": "
MethodSplit CIFAR-100Split ImageNet-100
Buffer size20050020002005002000
ER-ACE26.2836.4848.4124.2337.1249.55
+ IBTA27.3637.4949.1225.6337.9250.12
+ IBDS30.4439.5250.9129.8340.7151.85
+IBCL32.1240.9451.8930.7241.8152.62
DER++25.1137.1352.0826.5043.6558.05
+IBTA27.8838.7453.9627.3945.7259.33
+ IBDS31.5342.5954.3329.5846.1559.70
+IBCL33.8244.2154.8932.6346.5759.92
", + "image_path": "2d38f47d97a1ae860d0b7e74d5b321446d60c15efb8ed5508636400a0615466f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 299, + 243, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 299, + 243, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 299, + 243, + 310 + ], + "type": "text", + "content": "4.4 Baseline Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 316, + 481, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 481, + 389 + ], + "type": "text", + "content": "We conduct experiments on multiple state-of-the-art experience replay CL models: ER [10], DER++ [8], X-DER [6], and ER-ACE [9]. In addition to the original loss function based on replay, our method can be considered as a new regularization method. Therefore, we also compare our method with existing regularization-based techniques, including sSGD [32], oEWC [39], oLAP [38], and more recent SOTA methods, OCM [20], LiDER [5] and DualHSIC [46]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 402, + 198, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 198, + 412 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 198, + 412 + ], + "type": "text", + "content": "4.5 Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 418, + 481, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 481, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 481, + 538 + ], + "type": "text", + "content": "The comparison results are reported in Table 1. For the method of listing several baselines, our proposed method has improved the model's performance in various datasets and buffer sizes. In addition, we have the following main observations: 1) There has been a significant improvement in basic methods such as ER and DER++, as these two methods adopt more basic methods for sampling. Our method can reduce the bias caused by the sampling process to a certain extent. 2) Our methods have also shown some effectiveness on X-DER and ER-ACE, which take additional measures to alleviate sample imbalance and improve sample utilization, resulting in improved model performance. Our method further enhances the model's capabilities on top of this." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 539, + 481, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 481, + 611 + ], + "type": "text", + "content": "To further demonstrate the effectiveness of our method, we compare our method with other regularization methods, and it can be found that we achieve the best performance, recent methods such as OCM, LiDER and DualHSIC achieve significant progress compared to baseline when the number of buffer is small, and our method further improves the performance when the buffer data is small by considering modeling for the unsampled data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 624, + 238, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 238, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 238, + 636 + ], + "type": "text", + "content": "4.6 Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "In order to understand the effect of each of the items in the loss function on the performance of the algorithm, an ablation study has been carried out and the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "S. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 118, + 257, + 248 + ], + "blocks": [ + { + "bbox": [ + 136, + 118, + 257, + 248 + ], + "lines": [ + { + "bbox": [ + 136, + 118, + 257, + 248 + ], + "spans": [ + { + "bbox": [ + 136, + 118, + 257, + 248 + ], + "type": "image", + "image_path": "e17b5e35974228cf5ec5776a2af0dda49e6c658d92ad831dc2a12a503c651496.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 261, + 482, + 294 + ], + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 294 + ], + "type": "text", + "content": "Fig. 4: This figure shows the CKA matrix based on the ER algorithm, FT and incorporating our algorithm. It can be observed from it that due to our modeling of unsampled data, the model shows more information about the old task." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 260, + 118, + 367, + 248 + ], + "blocks": [ + { + "bbox": [ + 260, + 118, + 367, + 248 + ], + "lines": [ + { + "bbox": [ + 260, + 118, + 367, + 248 + ], + "spans": [ + { + "bbox": [ + 260, + 118, + 367, + 248 + ], + "type": "image", + "image_path": "6c4cc5c7089bf47269767d347ffd7406f8ca2b259b93520de11a4d1a7be87f2c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 369, + 118, + 477, + 248 + ], + "blocks": [ + { + "bbox": [ + 369, + 118, + 477, + 248 + ], + "lines": [ + { + "bbox": [ + 369, + 118, + 477, + 248 + ], + "spans": [ + { + "bbox": [ + 369, + 118, + 477, + 248 + ], + "type": "image", + "image_path": "305984db984788dc1f90a3b3b41fd0744b949396a47aee300cf03cb7baf59ca2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 320, + 482, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 440 + ], + "type": "text", + "content": "results are shown in Table 3. It is observed that in most of the cases, increasing any of the components has an impact on the performance of the algorithm. However, increasing the IBDS gives a large performance improvement, which indicates that it plays a vital role in the performance of the algorithm. This component represents the loss of substitution of unsampled data, which is necessary for training on new tasks using unsampled data features. Therefore, if this aspect is not well learned, it may negatively affect the subsequent performance of the algorithm. IBTA also has an impact on the performance of the algorithm, the involves regularization constraints on the input features in order to make the model learn enough robust features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 464, + 292, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 292, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 292, + 478 + ], + "type": "text", + "content": "5 Characteristic Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 495, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 481, + 521 + ], + "type": "text", + "content": "We analyze various characteristics of the model to gain insight into the factors contributing to the performance enhancements achieved by our approach." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 544, + 272, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 544, + 272, + 555 + ], + "spans": [ + { + "bbox": [ + 131, + 544, + 272, + 555 + ], + "type": "text", + "content": "5.1 CKA Matrix Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "content": "We use Central Kernel Alignment (CKA) [26] in Fig.4, which is an effective tool to measure the similarity of network representations and evaluate the relationships between these backbones. We send the same batch of instances to the ER (2000 buffer size) algorithm model and fine-tuning based model respectively to obtain corresponding feature maps, thus intuitively demonstrating the similarity of algorithm feature maps. Our algorithm can be used in conjunction with the experience replay method to model unsampled data. When the sampled data is 0, i.e. fine-tuning, our method can also get some benefits." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 447, + 100 + ], + "type": "text", + "content": "IB Based Data Correction in CL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 155, + 118, + 255, + 198 + ], + "blocks": [ + { + "bbox": [ + 155, + 118, + 255, + 198 + ], + "lines": [ + { + "bbox": [ + 155, + 118, + 255, + 198 + ], + "spans": [ + { + "bbox": [ + 155, + 118, + 255, + 198 + ], + "type": "image", + "image_path": "f776ee0a368bf83264a3019df57278fd23a062b9eca311ea385184ae50334934.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 205, + 214, + 213 + ], + "lines": [ + { + "bbox": [ + 176, + 205, + 214, + 213 + ], + "spans": [ + { + "bbox": [ + 176, + 205, + 214, + 213 + ], + "type": "text", + "content": "Initialization" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 258, + 118, + 358, + 198 + ], + "blocks": [ + { + "bbox": [ + 258, + 118, + 358, + 198 + ], + "lines": [ + { + "bbox": [ + 258, + 118, + 358, + 198 + ], + "spans": [ + { + "bbox": [ + 258, + 118, + 358, + 198 + ], + "type": "image", + "image_path": "9deb011b3e4c1988a3809b8798e4c7015966b38b5352829e6520b0359a26812f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 205, + 324, + 213 + ], + "lines": [ + { + "bbox": [ + 290, + 205, + 324, + 213 + ], + "spans": [ + { + "bbox": [ + 290, + 205, + 324, + 213 + ], + "type": "text", + "content": "Epoch " + }, + { + "bbox": [ + 290, + 205, + 324, + 213 + ], + "type": "inline_equation", + "content": "= 20" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 225, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 269 + ], + "type": "text", + "content": "Fig. 5: This image uses TSNE to display the changes in the characteristics of sampled and unsampled data with the number of training epochs of the model. We have shown one class of situation, with a sample data size of 200 and an unsampled data size of 1100." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 361, + 118, + 460, + 198 + ], + "blocks": [ + { + "bbox": [ + 361, + 118, + 460, + 198 + ], + "lines": [ + { + "bbox": [ + 361, + 118, + 460, + 198 + ], + "spans": [ + { + "bbox": [ + 361, + 118, + 460, + 198 + ], + "type": "image", + "image_path": "e3e5fdfec78084223105586c8257c1a5fb7c74e63c28ad3d3144c282a36632b5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 205, + 422, + 213 + ], + "lines": [ + { + "bbox": [ + 388, + 205, + 422, + 213 + ], + "spans": [ + { + "bbox": [ + 388, + 205, + 422, + 213 + ], + "type": "text", + "content": "Epoch " + }, + { + "bbox": [ + 388, + 205, + 422, + 213 + ], + "type": "inline_equation", + "content": "= 50" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 293, + 305, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 293, + 305, + 306 + ], + "spans": [ + { + "bbox": [ + 131, + 293, + 305, + 306 + ], + "type": "text", + "content": "5.2 Feature Decoupling Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 314, + 482, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 409 + ], + "type": "text", + "content": "A significant advance of our method is the decoupling of sampled and unsampled data features. We demonstrate this in Fig.5, where we use TSNE technology to visualize the changes in these two features with the increase of network training epochs. We test our method, the ER algorithm, on 200 sampled data and 1100 unsampled data from the ImageNet-100 dataset. As the network training epochs increase, we can observe that the two sets of features begin to separate from their initial coupling, which indicates that our method has achieved decoupling of the two data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 428, + 220, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 428, + 220, + 441 + ], + "spans": [ + { + "bbox": [ + 131, + 428, + 220, + 441 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 453, + 482, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 573 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 573 + ], + "type": "text", + "content": "In conclusion, our proposed method is designed to tackle the issue of data bias in continual learning through the application of experience replay. Drawing inspiration from information bottleneck theory, we steer the incorporation of unsampled data information, thereby bolstering the model's capacity to preserve past knowledge while accommodating novel tasks. Our conducted experiments offer empirical validation of the advantages conferred by our approach. While we have presented a method for computing unsampled data, it's important to acknowledge the existence of several alternative strategies for estimating these. This opens up exciting possibilities for continued exploration and innovation in this field." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 592, + 247, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 592, + 247, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 592, + 247, + 605 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "This work was supported by National Science and Technology Major Project, Grant No.2022ZD0116403, the Youth Innovation Promotion Association CAS and Strategic Priority Research Program of Chinese Academy of Sciences XDA27010300." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "S. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. Alemi, A.A., Fischer, I., Dillon, J.V., Murphy, K.: Deep variational information bottleneck. arXiv (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "text", + "content": "2. Aljundi, R., Lin, M., Goujaud, B., Bengio, Y.: Gradient based sample selection for online continual learning. In: Adv. Neural Inform. Process. Syst. (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 184, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 480, + 215 + ], + "type": "text", + "content": "3. Bang, J., Kim, H., Yoo, Y., Ha, J.W., Choi, J.: Rainbow memory: Continual learning with a memory of diverse samples. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 216, + 480, + 237 + ], + "type": "text", + "content": "4. Biesialska, M., Biesialska, K., Costa-Jussa, M.R.: Continual lifelong learning in natural language processing: A survey. arXiv (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "type": "text", + "content": "5. Bonicelli, L., Boschini, M., Porrello, A., Spampinato, C., Calderara, S.: On the effectiveness of lipschitz-driven rehearsal in continual learning. In: Adv. Neural Inform. Process. Syst. (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 303 + ], + "type": "text", + "content": "6. Boschini, M., Bonicelli, L., Buzzega, P., Porrello, A., Calderara, S.: Class incremental continual learning into the extended der-verse. IEEE Trans. Pattern Anal. Mach. Intell. (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 335 + ], + "type": "text", + "content": "7. Boschini, M., Buzzega, P., Bonicelli, L., Porrello, A., Calderara, S.: Continual semisupervised learning through contrastive interpolation consistency. Pattern Recognition Letters 162, 9-14 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 337, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 480, + 369 + ], + "type": "text", + "content": "8. Buzzega, P., Boschini, M., Porrello, A., Abati, D., Calderara, S.: Dark experience for general continual learning: a strong, simple baseline. In: Adv. Neural Inform. Process. Syst. (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 370, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 480, + 402 + ], + "type": "text", + "content": "9. Caccia, L., Aljundi, R., Asadi, N., Tuytelaars, T., Pineau, J., Belilovsky, E.: New insights on reducing abrupt representation change in online continual learning. In: Int. Conf. Learn. Represent. (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "type": "text", + "content": "0. Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H.S., Ranzato, M.: Continual learning with tiny episodic memories. Arxiv (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 425, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 456 + ], + "type": "text", + "content": "1. Chen, S., Zhang, M., Zhang, J., Huang, K.: Exemplar-based continual learning via contrastive learning. IEEE Transactions on Artificial Intelligence 5(7), 3313-3324 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 480, + 479 + ], + "type": "text", + "content": "2. Cossu, A., Carta, A., Lomonaco, V., Bacciu, D.: Continual learning for recurrent neural networks: an empirical evaluation. Neural Networks (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "type": "text", + "content": "3. Delange, M., Aljundi, R., Masana, M., Parisot, S., Jia, X., Leonardis, A., Slabaugh, G., Tuytelaars, T.: A continual learning survey: Defying forgetting in classification tasks. IEEE Trans. Pattern Anal. Mach. Intell. (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "text", + "content": "4. Douillard, A., Cord, M., Ollion, C., Robert, T., Valle, E.: PODNet: Pooled Outputs Distillation for Small-Tasks Incremental Learning. In: Eur. Conf. Comput. Vis. (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 544, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 577 + ], + "type": "text", + "content": "5. Douillard, A., Ram, A., Couairon, G., Cord, M.: Dytox: Transformers for continual learning with dynamic token expansion. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 610 + ], + "type": "text", + "content": "6. Ebrahimi, S., Elhoseiny, M., Darrell, T., Rohrbach, M.: Uncertainty-guided continual learning with bayesian neural networks. In: Int. Conf. Learn. Represent. (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "text", + "content": "7. Federici, M., Dutta, A., Forre, P., Kushman, N., Akata, Z.: Learning robust representations via multi-view information bottleneck. In: Int. Conf. Learn. Represent. (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "8. French, R.M.: Catastrophic forgetting in connectionist networks. Trends in cognitive sciences (1999)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 447, + 100 + ], + "type": "text", + "content": "IB Based Data Correction in CL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "19. Goodfellow, I.J., Mirza, M., Xiao, D., Courville, A., Bengio, Y.: An empirical investigation of catastrophic forgetting in gradient-based neural networks. arXiv (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "text", + "content": "20. Guo, Y., Liu, B., Zhao, D.: Online continual learning through mutual information maximization. In: Int. Conf. Mach. Learn. (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "21. Hadsell, R., Rao, D., Rusu, A.A., Pascanu, R.: Embracing change: Continual learning in deep neural networks. Trends in cognitive sciences (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "text", + "content": "22. Iscen, A., Zhang, J., Lazebnik, S., Schmid, C.: Memory-efficient incremental learning through feature adaptation. In: Eur. Conf. Comput. Vis. (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "text", + "content": "23. Jaiswal, A., Brekelmans, R., Moyer, D., Steeg, G.V., AbdAlmageed, W., Natarajan, P.: Discovery and separation of features for invariant representation learning. arXiv (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "24. Jie, S., Deng, Z.H., Li, Z.: Alleviating representational shift for continual fin-tuning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 313 + ], + "type": "text", + "content": "25. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proc. of the national academy of sciences (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "26. Kornblith, S., Norouzi, M., Lee, H., Hinton, G.E.: Similarity of neural network representations revisited. In: Int. Conf. Mach. Learn. (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "27. Li, D., Wang, T., Chen, J., Ren, Q., Kawaguchi, K., Zeng, Z.: Towards continual learning desiderata via hsic-bottleneck orthogonalization and equiangular embedding. In: AAAI (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "28. Li, X., Zhou, Y., Wu, T., Socher, R., Xiong, C.: Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting. In: Int. Conf. Mach. Learn. (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "type": "text", + "content": "29. Liu, D., Cheng, P., Zhu, H., Dong, Z., He, X., Pan, W., Ming, Z.: Mitigating confounding bias in recommendation via information bottleneck. In: Proceedings of the 15th ACM Conference on Recommender Systems (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "30. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "text", + "content": "31. Masana, M., Liu, X., Twardowski, B., Menta, M., Bagdanov, A.D., van de Weijer, J.: Class-incremental learning: Survey and performance evaluation on image classification. IEEE Trans. Pattern Anal. Mach. Intell. (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "type": "text", + "content": "32. Mirzadeh, S.I., Farajtabar, M., Pascanu, R., Ghasemzadeh, H.: Understanding the role of training regimes in continual learning. Arxiv (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "33. Moyer, D., Gao, S., Brekelmans, R., Steeg, G.V., Galstyan, A.: Evading the adversary in invariant representation. CoRR (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 533, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 555 + ], + "type": "text", + "content": "34. Pan, Z., Niu, L., Zhang, J., Zhang, L.: Disentangled information bottleneck. In: AAAI (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 555, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 577 + ], + "type": "text", + "content": "35. Parisi, G.I., Kemker, R., Part, J.L., Kanan, C., Wermter, S.: Continual lifelong learning with neural networks: A review. Neural Networks (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "text", + "content": "36. Qu, H., Rahmani, H., Xu, L., Williams, B., Liu, J.: Recent advances of continual learning in computer vision: An overview. arXiv (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "text", + "content": "37. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "38. Ritter, H., Botev, A., Barber, D.: Online structured laplace approximations for overcoming catastrophic forgetting. In: Adv. Neural Inform. Process. Syst. (2018)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "39. Schuster, M., Paliwal, K.K.: Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing 45(11), 2673-2681 (1997)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "S. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 512 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "40. Swaroop, S., Nguyen, C.V., Bui, T.D., Turner, R.E.: Improving and understanding variational continual learning. ArXiv (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 477, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 477, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 477, + 149 + ], + "type": "text", + "content": "41. Tishby, N., Pereira, F.C., Bialek, W.: The information bottleneck method (2000)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "text", + "content": "42. Titsias, M.K., Schwarz, J., Matthews, A.G.d.G., Pascanu, R., Teh, Y.W.: Functional regularisation for continual learning with gaussian processes. In: Int. Conf. Learn. Represent. (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 183, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 216 + ], + "type": "text", + "content": "43. Von Oswald, J., Zhao, D., Kobayashi, S., Schug, S., Caccia, M., Zucchet, N., Sacramento, J.: Learning where to learn: Gradient sparsity in meta and continual learning. In: Adv. Neural Inform. Process. Syst. (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "44. Wang, Y., Rudner, T.G., Wilson, A.G.: Visual explanations of image-text representations via multi-modal information bottleneck attribution. In: Adv. Neural Inform. Process. Syst. (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "45. Wang, Z., Chen, X., Wen, R., Huang, S.L., Kuruoglu, E.E., Zheng, Y.: Information theoretic counterfactual learning from missing-not-at-random feedback. In: Adv. Neural Inform. Process. Syst. (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 313 + ], + "type": "text", + "content": "46. Wang, Z., Zhan, Z., Gong, Y., Shao, Y., Ioannidis, S., Wang, Y., Dy, J.: Dualhsic: Hsic-bottleneck and alignment for continual learning. In: Int. Conf. Mach. Learn. (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "47. Wortsman, M., Ramanujan, V., Liu, R., Kembhavi, A., Rastegari, M., Yosinski, J., Farhadi, A.: Supermasks in superposition. In: Adv. Neural Inform. Process. Syst. (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "48. Xu, P., Zhang, J., Huang, K.: Exploration via joint policy diversity for sparse-reward multi-agent tasks. In: IJCAI (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "49. Xu, P., Zhang, J., Huang, K.: Population-based diverse exploration for sparse-reward multi-agent tasks. In: IJCAI (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "50. Xu, P., Zhang, J., Yin, Q., Yu, C., Yang, Y., Huang, K.: Subspace-aware exploration for sparse-reward multi-agent tasks. In: AAAI (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "51. Yu, L., Twardowski, B., Liu, X., Herranz, L., Wang, K., Cheng, Y., Jui, S., Weijer, J.v.d.: Semantic drift compensation for class-incremental learning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "text", + "content": "52. Zhang, X., Yi, J., Tao, J., Wang, C., Zhang, C.Y.: Do you remember? overcoming catastrophic forgetting for fake audio detection. In: Int. Conf. Mach. Learn. (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "53. Zhang, X., Yi, J., Wang, C., Zhang, C.Y., Zeng, S., Tao, J.: What to remember: Self-adaptive continual learning for audio deepfake detection. In: AAAI (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 491, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 512 + ], + "type": "text", + "content": "54. Zhou, D.W., Wang, Q.W., Qi, Z.H., Ye, H.J., Zhan, D.C., Liu, Z.: Deep class-incremental learning: A survey. arXiv (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 447, + 100 + ], + "type": "text", + "content": "IB Based Data Correction in CL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/InsMapper_ Exploring Inner-instance Information for Vectorized HD Mapping/703739e6-709f-487f-9515-53b438aabcdc_content_list.json b/2024/InsMapper_ Exploring Inner-instance Information for Vectorized HD Mapping/703739e6-709f-487f-9515-53b438aabcdc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..13b1fd2539a2359871fcf64ff3d7cdcdee35ea61 --- /dev/null +++ b/2024/InsMapper_ Exploring Inner-instance Information for Vectorized HD Mapping/703739e6-709f-487f-9515-53b438aabcdc_content_list.json @@ -0,0 +1,1648 @@ +[ + { + "type": "text", + "text": "InsMapper: Exploring Inner-instance Information for Vectorized HD Mapping", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhenhua Xu Kwan-Yee K. Wong Hengshuang Zhao\\*", + "bbox": [ + 277, + 212, + 725, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The University of Hong Kong", + "bbox": [ + 401, + 239, + 602, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project webpage: https://tonyxuqq.github.io/InsMapper", + "bbox": [ + 276, + 279, + 725, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Vectorized high-definition (HD) maps contain detailed information about surrounding road elements, which are crucial for various downstream tasks in modern autonomous vehicles, such as motion planning and vehicle control. Recent works attempt to directly detect the vectorized HD map as a point set prediction task, achieving notable detection performance improvements. However, these methods usually overlook and fail to analyze the important inner-instance correlations between predicted points, impeding further advancements. To address this issue, we investigate the utilization of inner-instance information for vectorized high-definition mapping through transformers, and propose a powerful system named InsMapper, which effectively harnesses inner-instance information with three exquisite designs, including hybrid query generation, inner-instance query fusion, and inner-instance feature aggregation. The first two modules can better initialize queries for line detection, while the last one refines predicted line instances. InsMapper is highly adaptable and can be seamlessly modified to align with the most recent HD map detection frameworks. Extensive experimental evaluations are conducted on the challenging NuScenes and Argovverse 2 datasets, where InsMapper surpasses the previous state-of-the-art method, demonstrating its effectiveness and generality.", + "bbox": [ + 261, + 318, + 743, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Autonomous Driving $\\cdot$ High-definition Map", + "bbox": [ + 261, + 609, + 640, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 648, + 375, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vectorized high-definition maps (HD maps) play a critical role in today's autonomous vehicles [6, 23, 27], as they contain detailed information about the road, including the position of road elements (e.g., road boundaries, lane splits, pedestrian crossings, and lane centerlines), connectivity, and topology of the road. Without the assistance of HD maps for perceiving and understanding road elements, unexpected vehicle behaviors may be encountered, such as incorrect path planning results or even vehicle collisions.", + "bbox": [ + 212, + 680, + 787, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Typically, HD maps are created by offline human annotation, which is labor-intensive, inefficient, and expensive. Although there are works proposing to make", + "bbox": [ + 212, + 786, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 217, + 824, + 385, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "such an offline process automatic [38, 40, 41], it is not possible to recreate and update the HD map frequently when the road has been modified, such as when a new road is built or an existing road is removed. To address this issue, several recent studies propose to detect local HD maps using an online method [7, 20, 24, 25, 27, 32]. This paper aims to further investigate online HD map detection techniques based on vehicle-mounted sensors (i.e., cameras and LiDARs).", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early online HD map detection works consider road element mapping as a semantic segmentation task in bird's-eye view (BEV) [22,30,31], in which road elements are predicted in raster format (i.e., pixel-level semantic segmentation mask). However, rasterized road maps cannot be effectively utilized by downstream tasks of autonomous vehicles, such as motion planning and vehicle control [23,26,27]. Moreover, it is challenging to distinguish instances from the rasterized map, especially when some road elements overlap with each other or when a complicated topology is encountered (e.g., road split or road intersections).", + "bbox": [ + 212, + 251, + 787, + 372 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To alleviate these problems, HDMap-", + "bbox": [ + 212, + 372, + 480, + 387 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Net [20] proposes hand-crafted post-processing algorithms to better obtain road element instances for vectorized HD maps. However, HDMapNet still heavily relies on rasterized results, restricting it from handling complicated urban scenes. Recently, some works have resorted to predicting vectorized HD maps directly [3,24,27,35]. A two-stage hierarchical set prediction method is proposed in VectorMapNet [27]. After predicting key points of road elements in the HD map, VectorMapNet sequentially generates intermediate points. Although VectorMapNet is considered to be the first online vectorized HD map detection work, the sequential operation degrades its efficiency and model performance. MapTR [24] further proposes to use DETR [5, 46] for vectorized HD map detection as a point set prediction problem, and the output point sets are then", + "bbox": [ + 212, + 387, + 460, + 779 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7eb90e49242fce19ab9f25d5d0a95a1d9ee8ba290014cdd3d56266bab7472bd0.jpg", + "image_caption": [ + "Fig.1: Comparison of vectorized HD map detection methods. All solutions are evaluated on the NuScenes validation set. The x-axis represents the mAP results and the y-axis displays topology level correctness [13]. InsMapper is flexible and adaptable, so it can be seamlessly modified to align with multiple existing frameworks. InsMapper is built on MapTR [24]. $\\dagger$ means InsMapper based on more recent framework PivotNet [7], while $\\ddagger$ indicates InsMapper based on previous SOTA, MapTR-V2 [25]." + ], + "image_footnote": [], + "bbox": [ + 470, + 402, + 785, + 585 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "grouped into road element instances. Some recent works further improve the performance of MapTR [7, 25, 32]. Among them, MapTR-V2 achieves state-of-the-art performance, but it overlooks and fails to utilize the essential inner-instance correlations between points to further boost the performance.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Xu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we focus on investigating the utilization of inner-instance information for vectorized high-definition mapping via transformer, and propose a new framework named InsMapper, which can utilize inner-instance point information for improved online HD map detection via three exquisite designs. First, a hybrid query generation module is introduced to better initialize queries for detection. Then, an inner-instance query fusion module is added before the transformer decoder, which fuses inner-instance object queries to refine the initialized line instances. Finally, a transformer-based inner-instance aggregation module is incorporated to further refine the predicted lines. Extensive experimental evaluations are conducted on the popular and challenging NuScenes [2] and Argoverse 2 [37] datasets. InsMapper is based on MapTR [24], and it can be seamlessly modified based on the latest frameworks with more enhanced performance, such as PivotNet [7] and MapTR-V2 [25]. No matter what the base framework is, InsMapper harvests state-of-the-art performance, surpassing the previous highest solutions. The intuitive performance comparisons are illustrated in Figure 1, and our main contributions can be summarized as follows:", + "bbox": [ + 212, + 146, + 787, + 388 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We analyze the limitations of existing HD map detection algorithms and propose a new framework named InsMapper, which can effectively utilize the easily overlooked inner-instance point correlation information for accurate and generalizable online HD map detection.", + "- We incorporate three useful modules to leverage inner-instance information, including hybrid query generation, inner-instance query fusion, and inner-instance feature aggregation. The first two modules better initialize queries for detection and the last one refines detected line instances.", + "- We conduct experiments on the challenging NuScenes and Argoverse 2 datasets. InsMapper outperforms all baselines by a large margin. The achieved state-of-the-art results demonstrate its effectiveness and generality." + ], + "bbox": [ + 225, + 402, + 800, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 594, + 387, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vector map detection. Vector maps use vectors to represent road elements, such as road networks, road curbs, and lane lines. A high-definition map (HD map) is a type of vector map, but a vector map may not qualify as an HD map if it lacks sufficient resolution or fails to provide lane-level information, among other factors. Online HD map detection can benefit from insights gained from a broader scope of vector map detection works, including road-network detection [1,14,39, 41], road-curb detection in aerial images [42, 43], and road lane line detection [15, 16], etc. DagMapper [16] aims to detect vectorized lane lines from pre-built point cloud maps using iterations. However, DagMapper only handles simple topology changes of lane lines and struggles to handle complex road intersections. RNGDet++ [41] applies DETR along with an imitation learning algorithm [34] for road-network detection, achieving state-of-the-art performance on multiple datasets. Although these methods achieve satisfactory graph detection accuracy, they suffer from poor efficiency due to their iterative algorithms. Given the", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "InsMapper", + "bbox": [ + 656, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "stringent real-time requirements of autonomous vehicles, they are not suitable for online HD map detection.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "HD map detection. HD map detection was initially a subtask of BEV detection [21, 22, 28, 30, 31]. Recently, given the importance of HD maps, several works have directly focused on HD map detection [4, 20, 24, 27, 29, 33, 35, 38, 40]. However, most of these works either involve offline HD map creation [12, 38, 40] or only detect one specific road element [3, 4, 33]. HDMapNet [20] is considered to be the first work specifically designed for multiple road element detection. However, HDMapNet outputs rasterized results, requiring complicated handcrafted post-processing to obtain vectorized HD maps. To address this issue, VectorMapNet [27] is believed to be the first work detecting vectorized HD maps in real-time. However, it consists of two stages, and its efficiency is significantly impacted by sequential operations. In contrast, MapTR [24] uses deformable DETR [46] to design an end-to-end vectorized HD map detection model, which greatly simplifies the pipeline and delivers better detection performance. Some recent works further improve the performance of MapTR [7, 25, 32]. Among them, MapTR-V2 achieves state-of-the-art performance on vectorized HD map detection, but it still overlooks and does not investigate the inner-instance correlations between points, limiting further improvements.", + "bbox": [ + 212, + 176, + 785, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Detection by Transformer. DETR [5] is the first end-to-end transformer-based object detection framework. Compared with previous CNN-based methods [8, 10], DETR removes the need for anchor proposals and non-maximum suppression (NMS), making it simpler and more effective. To address the issue of slow convergence, subsequent works propose accelerating DETR training through deformable attention [46], denoising [18], and dynamic query generation [19, 36, 45]. Intuitively, most of these DETR refinements could be adapted to the HD map detection task since our proposed InsMapper relies on DETR. However, unlike typical object detection tasks in which objects are assumed to be independent and identically distributed (i.i.d.), detected points in HD maps have strong correlations, especially among inner-instance points. This inherent difference makes some refined DETR methods unsuitable to our task and additional investigations are required.", + "bbox": [ + 212, + 433, + 787, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Point Correlation", + "text_level": 1, + "bbox": [ + 215, + 651, + 423, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Vector Map Decomposition and Sampling", + "text_level": 1, + "bbox": [ + 215, + 680, + 606, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $G$ denote the original vector map label of the scene, which consists of vertices $V$ and edges $E$ . The vector map contains multiple classes of road elements, including pedestrian crossings, road dividers, road boundaries, and lane centerlines. Among them, the first three classes of road elements are simple polylines or polygons without intersection points. While lane centerlines have more complicated topology, such as lane split, lane merge, and lane intersection. To unify all vector elements, the vector map is decomposed into simpler shapes (i.e., polylines and polygons) without intersections. Any vertices in the vector map with degrees larger than two (i.e., intersection vertices) are removed from $G$ , but", + "bbox": [ + 212, + 704, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Xu et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "incident edges are kept. In this way, a set of simple polylines and polygons without intersections is obtained as $G^{*} = \\{l_{i}\\}_{i=1}^{N^{*}}$ , where $G^{*}$ is an undirected graph. Each shape $l_{i}$ is defined as an instance, and $N^{*}$ denotes the overall number of instances in a vector map.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following MapTR and MapTR-V2, to enhance the parallelization capacity of the model, each instance is evenly re-sampled with fixed-length points as $l_{i} = (v_{1},\\dots,v_{j},\\dots,v_{n_{p}})$ . $l_{i}$ is ordered from $v_{1}$ to $v_{n_p}$ , with $n_p$ being the number of sampled points for each instance. For polygon instances, $v_{1}$ is equal to $v_{n_p}$ . Please refer to the supplementary document for the visualization of the pre-processing.", + "bbox": [ + 212, + 207, + 784, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Inner-instance Correlation", + "text_level": 1, + "bbox": [ + 214, + 306, + 478, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike conventional object detection tasks, where objects can be approximated as independent and identically distributed (i.i.d.), strong correlations exist between points within the same line instance in the vectorized HD map detection task. Please refer to the supplementary document for the visualization of the point correlation.", + "bbox": [ + 212, + 332, + 782, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inner-instance correlation is crucial for point coordinates prediction. Points within the same instance can collaborate by sharing inner-instance information, leading to smoother and more accurate predictions. Without this collaboration, points of the same instance may produce independent predictions, leading to zig-zag or incorrect instances. The inner-instance correlation can serve as additional constraints to facilitate the query initialization and refinement of predicted lines. In previous methods, the correlation of points is not correctly analyzed and leveraged, limiting further improvement. In subsequent sections, we introduce InsMapper to more effectively utilize point correlations and improve vectorized map detection performance.", + "bbox": [ + 212, + 409, + 784, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Methodology", + "text_level": 1, + "bbox": [ + 214, + 584, + 380, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Overall Framework", + "text_level": 1, + "bbox": [ + 214, + 617, + 419, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Building on MapTR [24], our proposed InsMapper is an end-to-end trainable framework for online vectorized HD map detection, as illustrated in Figure 2.", + "bbox": [ + 212, + 642, + 782, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "InsMapper is adaptive and can be seamlessly modified based on the latest frameworks with more enhanced performance, such as PivotNet [7] and MapTR-V2 [25]. Following BEVformer [22], InsMapper first projects perspective view camera images into the bird's-eye-view (BEV). After obtaining the BEV features, InsMapper uses deformable attention layers [46] to process input object queries. Each object query predicts one point, including the object class and the regression point position. To better leverage inner-instance information and further enhance the final detection performance, we introduce three designs in InsMapper: hybrid query generation, which replaces the hierarchical query in MapTR, generating line queries with better diversity; inner-instance query fusion, which fuses generated queries within the same instance based on", + "bbox": [ + 212, + 674, + 784, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "InsMapper", + "bbox": [ + 656, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/153f5c94707310daf2c3f6a5c59eb2405a1a7953aa5e3a05b673c0fd415fe27b.jpg", + "image_caption": [ + "Fig. 2: Overall framework. InsMapper is an end-to-end transformer model with an encoder-decoder structure. The transformer encoder projects perspective-view camera images into a bird's-eye view (BEV). Subsequently, the transformer decoder detects vector instances by predicting point sets. To enhance the utilization of inner-instance information, we introduce the following three components: a hybrid query generation scheme (orange module), an inner-instance query fusion module (yellow module), and an inner-instance feature aggregation module (blue module). The first two modules better initialize queries for detection and the final one refines detected line instances." + ], + "image_footnote": [], + "bbox": [ + 222, + 145, + 787, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "inner-instance features to further enhance the quality of initialized line queries; and inner-instance feature aggregation, which adds a new inner-instance self-attention module to decoder layers to further aggregate inner-instance features for line refinement.", + "bbox": [ + 212, + 547, + 787, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Different from past works, which treat queries as independently distributed, InsMapper utilizes inner-instance information to fuse object queries and aggregate intermediate features in deformable transformer decoder layers. Otherwise, adjacent vertices may produce independent position regression results, leading to zigzag or even incorrect instances. With the assistance of inner-instance information, the model can better initialize queries for detection and refine the positions of detected points within an instance, significantly enhancing the final performance, as demonstrated in the experiments section.", + "bbox": [ + 212, + 607, + 787, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Query Generation", + "text_level": 1, + "bbox": [ + 215, + 752, + 413, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In contrast to traditional object detection tasks, where queries are independent and identically distributed (i.i.d.), the points to be detected in our task exhibit significant correlations. Various methods can be used to generate object queries as input for the transformer decoder, such as the basic scheme, hierar", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Xu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2b3026b0a16709faa46790e0f959aea1b377cf19f46caa56809a437cfd9980f5.jpg", + "image_caption": [ + "Fig.3: Query generation schemes. For concise visualization, the number of instances $N_{I}$ is 2, and the number of points per instance $n_p$ is 3. The hybrid scheme can initialize lines with both good diversity and quality." + ], + "image_footnote": [], + "bbox": [ + 217, + 174, + 485, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d97ac786feed7392a30d050bc18f5bde601ec124e187d05874aa7cb4d95f00c7.jpg", + "image_caption": [ + "(a) Basic" + ], + "image_footnote": [], + "bbox": [ + 506, + 150, + 591, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6f3776c816987d234da648677839b80c2689b13c85b31dc2f64d3b1f91713c44.jpg", + "image_caption": [ + "(b) Hierarchical" + ], + "image_footnote": [], + "bbox": [ + 602, + 148, + 691, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a87f30e0e3d703d02d5300a33ffccd98957e6509fe176317d56eef30638cf4dd.jpg", + "image_caption": [ + "(c) Hybrid", + "Fig. 4: Visualization of initialized lines with different query generation schemes. The quality and diversity of initialized lines are crucial for the final performance. Each color represents one initialized line." + ], + "image_footnote": [], + "bbox": [ + 707, + 148, + 787, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "chical scheme, and hybrid scheme. All query generation schemes are illustrated in Figure 3. Initialized queries are organized into line instances. The quality and diversity of the initialized lines are both crucial.", + "bbox": [ + 212, + 406, + 784, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Basic query generation. This method is leveraged in the original DETR [5] for query generation. It assumes that points are i.i.d. and generates queries randomly, without utilizing inner-instance information. Let $N_{I}$ denote the number of predicted instances, which is obviously larger than $N^{*}$ . Since each instance contains $n_p$ points, the basic object queries consist of $N_{I} \\cdot n_{p}$ randomly generated i.i.d. queries, resulting in irregular line shapes. Due to the lack of constraints on inner-instance points, the initialized line instances are noisy and have poor quality, which leads to degraded performance. MapTR with the basic query scheme is visualized in Figure 4a.", + "bbox": [ + 212, + 452, + 787, + 587 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hierarchical query generation. To address the aforementioned issue of the basic scheme, a hierarchical query generation scheme is proposed in MapTR [24] and MapTR-V2 [25]. Let $Q_{ins} = \\{q_i^I\\}_{i=1}^{N_I}$ denote instance queries and $Q_{pts} = \\{q_j^P\\}_{j=1}^{n_p}$ denote point queries. The object queries for HD map detection are then the pairwise addition of $Q_{ins}$ and $Q_{pts}$ :", + "bbox": [ + 214, + 588, + 787, + 665 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nQ = \\left\\{q _ {i, j} = q _ {i} ^ {I} + q _ {j} ^ {P} \\mid q _ {i} ^ {I} \\in Q _ {\\text {i n s}}, q _ {j} ^ {P} \\in Q _ {\\text {p t s}} \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 672, + 785, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $|Q| = |Q_{ins}| \\cdot |Q_{pts}| = N_I \\cdot n_p$ . For each line instance $i$ , the instance query $q_i^I$ denotes its uniqueness. By using $q_i^I$ as a bridge, object queries within the same instance can better collaborate with each other for point prediction. But all line instances share the same point queries so all initialized lines tend to have quite similar shapes, degrading initialization diversity, which harms the final detection performance. The initialized lines generated by MapTR with the hierarchical query scheme are visualized in Figure 4b.", + "bbox": [ + 212, + 703, + 787, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hybrid query generation. The basic scheme does not take into account any information exchange between queries, whereas the hierarchical one generates", + "bbox": [ + 214, + 809, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "InsMapper", + "bbox": [ + 656, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "repetitive line shapes. To address these issues, this paper presents a hybrid query generation method that mitigates the drawbacks of the aforementioned schemes while maintaining appropriate inner-instance information exchange.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Let the instance queries be $Q_{ins} = \\{q_i^I\\}_{i=1}^{N_I}$ , and the point queries be $Q_{pts} = \\{q_j^P\\}_{j=1}^{N_P} = \\{q_j^P\\}_{j=1}^{N_I \\cdot n_p}$ . The point queries are divided into $N_I$ instance groups, and a point query $q_j^P$ is assigned to the $\\lceil \\frac{j}{n_p} \\rceil$ -th instance. Consequently, the final object query is the sum of a point query $p_j^P$ and its assigned instance query $p_k^I$ , where $k = \\lceil \\frac{j}{n_p} \\rceil$ . The object query set can be expressed as:", + "bbox": [ + 212, + 191, + 784, + 282 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nQ = \\left\\{q _ {k} ^ {I} + q _ {j} ^ {P} \\mid q _ {j} ^ {P} \\in Q _ {p t s} \\right\\} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 294, + 785, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $|Q| = |Q_{pts}| = N_P = N_I \\cdot n_p$ . In contrast to the hierarchical scheme, each point query in the hybrid scheme is utilized only once to generate the object query, thereby preventing repetitive line shapes. Simultaneously, point queries belonging to the same instance are summed with a shared instance query, establishing the inner-instance connection. The hybrid query generation method can be considered as a combination of the basic and hierarchical schemes, effectively mitigating their respective drawbacks. InsMapper with the hybrid query generation scheme is visualized in Figure 4c.", + "bbox": [ + 212, + 316, + 784, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Inner-instance Query Fusion", + "text_level": 1, + "bbox": [ + 215, + 462, + 496, + 478 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The input object queries for the transformer decoder are generated from instance queries and point queries. Although the generated queries can leverage some inner-instance information, this information exchange is indirect and inflexible. The instance query is distributed equally among all inner-instance points, while a more precise point-to-point information exchange cannot be realized. As a result, a query fusion module is introduced to further utilize inner-instance information and refine initialized lines.", + "bbox": [ + 212, + 488, + 784, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Let $Q_{i} = \\{q_{i,j}\\}_{j=1}^{n_{p}}$ represent the set of object queries belonging to the $i$ -th instance, and $q_{i,j}$ denote the $j$ -th point of the $i$ -th instance. $q_{i,j}$ is correlated with all other queries in $Q_{i}$ . To better fuse inner-instance object queries, each query is updated by a weighted sum of all queries in $Q_{i}$ as:", + "bbox": [ + 212, + 594, + 784, + 656 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nq _ {i, j} = f \\left(q _ {i, j}, Q _ {i}\\right) = \\sum_ {k = 1} ^ {n _ {p}} w _ {i, j, k} \\phi \\left(q _ {i, k}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 667, + 785, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $w_{i,j,k}$ demonstrates weights for query fusion. $f(\\cdot)$ is the fusion function and $\\phi(\\cdot)$ is the kernel in case nonlinear transformation is needed. $f(\\cdot)$ could be realized by handcraft weights, fully connected layers, or self-attention. In this work, self-attention is applied as $f(\\cdot)$ .", + "bbox": [ + 212, + 719, + 782, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conventional object detection tasks, object queries are assumed to be independent and identically distributed (i.i.d.), making query fusion unnecessary. However, in the task of HD map detection, the query fusion module effectively aligns the update of queries and enables each point to pay more attention to", + "bbox": [ + 212, + 780, + 784, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Xu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/69c49678df6c7cb8472dd3c690821fde218551d962898247891c2624b3127a37.jpg", + "image_caption": [ + "(a) No fusion" + ], + "image_footnote": [], + "bbox": [ + 225, + 167, + 330, + 325 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2e3e38d7b3ab1a38ad9f0535c2ec7191e14301154bec10b5aeda39e1526c2965.jpg", + "image_caption": [ + "(b) With fusion" + ], + "image_footnote": [], + "bbox": [ + 352, + 170, + 450, + 321 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4d4a9fcfdb1c5a3058a2349065c04ff2c1c0f86f45216eb70bddfbebe100d8c9.jpg", + "image_caption": [ + "Fig.5: Visualization of the effect of the inner-instance query fusion module. (a) No query fusion. (b) With query fusion. The query fusion module can better refine the initialized lines based on inner-instance information. In this example, the proposed query fusion module significantly enhances the diversity of initialized line instances.", + "Fig.6: Decoder of InsMapper. An inner-instance multi-head self-attention module is incorporated into decoder layers. In this module, the attention between points belonging to different instances is blocked (grey grids). Only inner-instance attention is allowed (colored grids). Colored grids are randomly blocked with $\\epsilon$ probability (set to grey) for robustness. The proposed module can further utilize inner-instance information and refine the predicted lines." + ], + "image_footnote": [], + "bbox": [ + 480, + 147, + 789, + 345 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "neighboring points. It can greatly refine the initialized lines for better query initialization, leading to enhanced performance. Without this module, queries within the same instance cannot be aware of each other, rendering them \"blind\". Poor line initialization tends to have degraded detection scores. The effect of the query fusion module is visualized in Figure 5.", + "bbox": [ + 212, + 529, + 787, + 606 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 Inner-instance Feature Aggregation", + "text_level": 1, + "bbox": [ + 215, + 630, + 553, + 645 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In addition to object query manipulation, InsMapper performs inner-instance feature aggregation for predicted line refinement in the transformer decoder layers by incorporating an additional inner-instance self-attention module. Inspired by [9], inner-instance attentions are randomly blocked with probability $\\epsilon$ for robustness. The decoder layer of InsMapper is depicted in Figure 6.", + "bbox": [ + 212, + 657, + 785, + 734 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The inner-instance self-attention module resembles the original self-attention module but features a specially designed attention mask. As illustrated in Figure 6, the attention mask of inner-instance points is set to zero (colored grids), indicating that attention between points within the same instance is allowed. Conversely, for points belonging to different instances (i.e., inter-instance points), the corresponding attention mask values are set to one, blocking attention between them. This method encourages the model to focus more on inner-instance", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "InsMapper", + "bbox": [ + 656, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "information, resulting in more consistent predictions. To further enhance the robustness of this module, the inner-instance attention (colored grids) has an $e$ probability of being blocked.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "An alternative method involves placing the proposed inner-instance attention module before the cross-attention module. However, the self-attention module before cross-attention should not block inter-instance information. Otherwise, some instances may produce duplicate predictions since they cannot \"clearly see\" other instances, which is analyzed in the original DETR [5]. Consequently, implementing inner-instance self attention before the cross-attention module leads to degraded final performance. Thus the proposed inner-instance feature aggregation should be placed after the cross-attention layer. More experiments about the decoder structure are provided in the appendix.", + "bbox": [ + 212, + 191, + 784, + 328 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 345, + 375, + 363 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 214, + 373, + 428, + 388 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. Extensive experiments are conducted on the popular and challenging NuScenes dataset [2] and the Argoverse 2 dataset [37]. These two datasets comprise hundreds of data sequences captured in various autonomous driving scenarios, encompassing diverse weather conditions and time periods. The perception range for the X-axis and Y-axis in the BEV is set to $[-15m, 15m]$ and $[-30m, 30m]$ , respectively. The detected vectorized HD map should encompass four types of road elements simultaneously: road boundary, lane split, pedestrian crossing, and lane centerline. Among these, the lane centerline is a crucial element for ensuring effective vehicle planning and control. Detecting lane centerlines poses greater difficulty compared to the other three elements.", + "bbox": [ + 212, + 393, + 784, + 545 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation details. We perform all the experiments on a machine equipped with 8 RTX-3090 GPUs. During the training phase, all GPUs are utilized, whereas only a single GPU is employed for inference. For our proposed InsMapper, we adopt multiple settings akin to the previous SOTA method, MapTR. All ablation studies are conducted for 24 epochs, with ResNet50 [11] as the backbone. For a fair comparison, except for the proposed modules, we keep the experiment settings exactly the same, such as batch size, network depth, the number of input queries, etc.", + "bbox": [ + 212, + 545, + 792, + 665 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation metrics. Models are assessed using two types of metrics. The first metric, average precision (AP), gauges the instance-level detection performance, employing Chamfer distance for matching predictions with ground truth labels. To ensure a fair comparison, we follow previous works to calculate multiple $\\mathrm{AP}_{\\tau}$ values with $\\tau \\in \\{0.5, 1.0, 1.5\\}$ , and report the average AP. To measure the topology correctness, the metric TOPO that is widely used in past works [12-14, 41] is reported. For all metrics, a larger value indicates better performance.", + "bbox": [ + 212, + 666, + 785, + 772 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Performance Comparison", + "text_level": 1, + "bbox": [ + 214, + 789, + 472, + 804 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we compare InsMapper with previous state-of-the-art methods using the aforementioned evaluation metrics. The quantitative comparison results", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Xu et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/a30545cd2c236ff090daae4d43839ad4307e0255bd42faec8e4409d65afc3c0b.jpg", + "table_caption": [ + "Table 1: Quantitative results of comparison experiments on the NuScenes validation set. Colored numbers show differences between InsMapper and the SOTA baseline under the same experiment setting. \"C\" and \"L\" denote camera and LiDAR, respectively. \"-\" denotes that the result is not available. \"V2-99\" and \"Sec\" correspond to VoVNetV2-99 [17] and SECOND [44]. $\\dagger$ represents InsMapper based on PivotNet [7], $\\ddagger$ indicates InsMapper based on MapTR-V2 [25]." + ], + "table_footnote": [], + "table_body": "
MethodsEpochsBackboneModality\\(AP_{ped}\\)\\(AP_{div}\\)\\(AP_{bound}\\)\\(AP_{center}\\)mAPTOPO
HDMapNet30Effi-B0C4.4123.7358.1737.2930.9029.79
STSU110R50C---31.2131.2132.11
VectorMapNet130R50C28.6639.7433.0634.9334.1039.90
MapTR24R50C37.9246.2550.0737.4742.9346.77
MapTR24V2-99C44.5756.2560.5846.3051.9255.16
MapTR24R50&SecC&L49.4758.9866.7244.8955.0156.77
MapTR110R50C49.1659.1258.9347.2653.6259.16
BeMapNet24R50C43.5954.1151.9038.6047.0549.90
PivotNet24R50C42.7353.5552.0142.4747.6951.04
MapTR-V224R50C48.2054.7556.6645.8651.3751.15
MapTR-V2110R50C58.6865.7267.1256.1661.9263.26
InsMapper24R50C44.3653.3652.7742.3548.3151.58
InsMapper24V2-99C51.1663.7164.4751.4057.6860.00
InsMapper24R50&SecC&L56.0063.4271.6152.8560.9762.51
InsMapper110R50C55.4263.8765.8054.2059.4066.19
InsMapper†24R50C46.4554.9154.1645.2050.1853.57
InsMapper‡24R50C49.0757.9859.8447.9953.7253.00
InsMapper†110R50C62.0967.6068.1558.4164.0664.72
", + "bbox": [ + 215, + 239, + 785, + 458 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ff2567bdaeeaa76a643da0e4ac71e86aea9f272153d40971038885f6761e9c3c.jpg", + "table_caption": [ + "Table 2: Quantitative results of comparison experiments on the Argoverse 2 validation set. $\\ddagger$ indicates InsMapper based on MapTR-V2 [25]." + ], + "table_footnote": [], + "table_body": "
MethodsEpochsBackboneModalityAPpedAPdivAPboundAPcentermAPTOPO
MapTR6R50C52.8863.6861.1859.7359.3775.79
MapTR-V26R50C57.1667.9665.2563.2063.3977.94
InsMapper6R50C55.6166.6062.5862.6761.8777.58
InsMapper†6R50C58.9171.3566.9064.7065.4679.26
", + "bbox": [ + 215, + 512, + 785, + 580 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/c9cf344cd196342bce229d65f7d6e60b5be68877b7c5f8f400aff0974a95c1ab.jpg", + "table_caption": [ + "Table 3: Quantitative results of comparison experiments on the NuScenes validation set without centerline. $\\ddagger$ indicates InsMapper based on MapTR-V2 [25]." + ], + "table_footnote": [], + "table_body": "
MethodsEpochsBackboneModalityAPpedAPdivAPboundmAP
MapTR24R50C46.0451.5853.0850.23
MapTR-V224R50C59.8062.4062.4061.50
InsMapper24R50C48.4454.6856.9253.35
InsMapper‡24R50C61.5465.0964.6263.75
", + "bbox": [ + 215, + 636, + 785, + 705 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "are reported in Table 1. InsMapper is flexible and adaptive, making it seamless to modify it based on existing transformer frameworks. InsMapper is built on MapTR for experiments, we also report the results of InsMapper with different base frameworks, e.g. InsMapper† is based on PivotNet [7] and InsMapper‡ relies on the previous SOTA method MapTR-V2 [25]. Compared to the previous methods, InsMapper improves the AP of all road elements by around 2. The topological metric also sees an improvement of around 1.5. The inference FPS of", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "InsMapper", + "bbox": [ + 656, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/e4dd68f52ddfa6eba12b27b688195bd1fd82907b204c2b159d06cf9d783dc2dc.jpg", + "table_caption": [ + "Table 4: Quantitative results of comparison experiments on the NuScenes validation set. Centerlines are directed graphs in this table. $\\ddagger$ indicates InsMapper based on MapTR-V2 [25]." + ], + "table_footnote": [], + "table_body": "
MethodsEpochsBackboneModality\\(AP_{ped}\\)\\(AP_{div}\\)\\(AP_{bound}\\)\\(AP_{center}\\)mAPTOPO
MapTR24R50C37.3946.6150.6937.2242.9846.98
MapTR-V224R50C49.5355.4658.4853.8254.3254.90
InsMapper24R50C43.2453.7753.8943.2248.5352.09
\\(InsMapper^{\\dagger}\\)24R50C51.6257.8860.9056.5256.7356.01
", + "bbox": [ + 215, + 198, + 787, + 268 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "MapTR and InsMapper are 8.5 and 7.5, respectively. The evaluation outcomes on the Argoverse 2 dataset [37] can be found in Table 2. InsMapper outperforms past works with notable enhancements. Table 3 presents the results without centerlines. InsMapper with different bases attains an improvement of more than $2\\mathrm{mAP}$ . In the aforementioned experiments, all road elements are treated as undirected graphs. Table 4 shows the results when centerlines are represented as directed graphs. Based on the results, InsMapper still presents superior results than past works for directed HD map detection. Furthermore, we illustrate the qualitative visualizations in Figure 7. Owing to the effective employment of inner-instance information, InsMapper generates smoother and more precise HD maps compared to previous works.", + "bbox": [ + 212, + 297, + 787, + 462 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Consequently, InsMapper demonstrates superior performance compared to prior approaches across multiple datasets, exhibiting outstanding detection outcomes. Concurrently, InsMapper maintains a comparable inference speed to the previous state-of-the-art method. These exceptional experimental results strongly demonstrate the effectiveness and generality of the proposed designs.", + "bbox": [ + 212, + 464, + 787, + 541 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 566, + 401, + 580 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Key component designs. We conduct several ablation study experiments to confirm the necessity of the proposed modules. Initially, we incorporate these modules into the previous state-of-the-art baseline model incrementally, and the results are presented in Table 5. It is evident that all three modules contribute to performance improvements, thereby validating their necessity. To further investigate the impact of each module's design, we perform comprehensive ablation studies in the following paragraphs.", + "bbox": [ + 212, + 595, + 787, + 702 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Hybrid query generation. Compared to basic or hierarchical query generation schemes, the proposed hybrid query generation scheme generates queries with better quality and diversity, due to effectively utilizing inner-instance correlation. The evaluation results of the different query generation schemes are presented in Table 6. These results reveal that the hybrid query generation scheme outperforms its counterparts, thereby demonstrating the soundness of this design.", + "bbox": [ + 212, + 703, + 787, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Inner-instance query fusion. Query fusion is proposed to leverage the correlation between inner-instance queries to enhance prediction performance. Various methods can be employed for fusing queries, including mean fusing (i.e.,", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Xu et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9e81c143736942d0d6a4bb29a6b7d0ab832ae314d6a85d767761e8c8af2a39bd.jpg", + "image_caption": [ + "Fig. 7: Qualitative visualization. The predicted map contains four classes, i.e., road boundaries (green), lane splits (red), pedestrian crossing (blue), and lane centerlines (pink). InsMapper presents better detection results compared with the past work." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 571, + 558 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/503858f18474dde558d326dc82abed08358b72a575badea6db98b137a2268064.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 143, + 782, + 558 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "each query is summed with the mean of all queries of the same instance), fusion by the feed-forward network, and fusion by self-attention. The evaluation results are presented in Table 7. The outcomes of the query fusion methods vary significantly, with self-attention-based query fusion achieving the best results. Consequently, self-attention query fusion is incorporated into InsMapper.", + "bbox": [ + 212, + 642, + 787, + 717 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Inner-instance feature aggregation. In InsMapper, the inner-instance self-attention module is incorporated into the decoder layers following the cross-attention module. We evaluate InsMapper under four different conditions: without the inner-instance self-attention module (no attention), replace the inner-instance self-attention module with a vanilla self-attention module (vanilla attention), without the random blocking (no random blocking, i.e., $\\epsilon = 0$ ), and with the module placed before the cross-attention (change position). The results are shown in Table 8. The outcomes reveal that removing or altering the proposed", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "InsMapper", + "bbox": [ + 656, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e76427544833258e5817c6b9646dc499e0eff9fb340dd6cda978a54ee2d6b268.jpg", + "table_caption": [ + "Table 5: Ablation studies on key component designs. \"QG\", \"QF\" and \"FA\" represent hybrid query generation, inner-instance query fusion, and inner-instance feature aggregation, respectively." + ], + "table_footnote": [], + "table_body": "
QGQFFAmAPTOPO
42.9346.77
✓✓45.42(↑2.49)49.19(↑2.42)
✓✓45.45(↑2.52)48.54(↑1.77)
✓✓46.93(↑4.00)50.16(↑3.39)
✓✓✓48.31(↑5.38)51.58(↑4.81)
", + "bbox": [ + 218, + 226, + 488, + 325 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/3d6ac60e8cc309f99b6c5c8c3caaa76cf6819ecd289a029b00aa3a6401118bff.jpg", + "table_caption": [ + "Table 6: Ablation studies on inner-instance query generation." + ], + "table_footnote": [], + "table_body": "
Query SchememAPTOPO
InsMapper-basic46.6249.39
InsMapper-hierarchical46.9350.16
InsMapper-hybrid48.3151.58
", + "bbox": [ + 217, + 364, + 491, + 438 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/60467ad1dcabcde060cf7384529273568e4ffcd11675ff327319086a3e8335c4.jpg", + "table_caption": [ + "Table 7: Ablation studies on inner-instance query fusion." + ], + "table_footnote": [], + "table_body": "
Query FusionmAPTOPO
No fusion45.4548.54
Mean43.3446.00
Feed-Forward47.8050.93
Self-attention48.3151.58
", + "bbox": [ + 509, + 185, + 785, + 271 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/44be73fd04a7f1432fe73293e2e9318887cb5e9b550adfa2643f738220674af7.jpg", + "table_caption": [ + "Table 8: Ablation studies on inner-instance feature aggregation. Multiple designs of the incorporated inner-instance self-attention layers are examined." + ], + "table_footnote": [], + "table_body": "
MethodmAPTOPO
No attention45.4249.19
Vanilla attention46.2349.51
No random blocking47.6751.49
Change position45.2448.16
InsMapper48.3151.58
", + "bbox": [ + 513, + 338, + 782, + 436 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "module can impair model performance. This observation confirms the effectiveness of the proposed module design.", + "bbox": [ + 212, + 472, + 782, + 503 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 544, + 359, + 561 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduced InsMapper, an end-to-end transformer-based model for on-the-fly vectorized HD map detection. InsMapper surpasses previous works by effectively leveraging inner-instance information to improve detection outcomes. Three exquisite designs have been proposed to utilize inner-instance information, including hybrid query generation, inner-instance query fusion, and inner-instance feature aggregation. The first two modules improve the quality of initialized line instances for detection, while the last module effectively refines the detected lines. The superiority of InsMapper is well exhibited through experiments on the challenging NuScenes and Argoverse 2 datasets. The evaluation results demonstrate the effectiveness and generality of InsMapper, making it a promising solution for the vectorized HD map detection task. In the future, InsMapper will be further enhanced for temporal-consistent vector map detection, making it more suitable for real-world applications.", + "bbox": [ + 212, + 594, + 787, + 792 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement. This work is supported by the National Natural Science Foundation of China (No. 62201484), HKU Startup Fund, and HKU Seed Fund for Basic Research.", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Xu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bastani, F., He, S., Abbar, S., Alizadeh, M., Balakrishnan, H., Chawla, S., Madden, S., DeWitt, D.: Roadtracer: Automatic extraction of road networks from aerial images. In: CVPR (2018) 3", + "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020) 3, 10", + "3. Can, Y.B., Liniger, A., Paudel, D.P., Van Gool, L.: Structured bird's-eye-view traffic scene understanding from onboard images. In: ICCV (2021) 2, 4", + "4. Can, Y.B., Liniger, A., Paudel, D.P., Van Gool, L.: Topology preserving local road network estimation from single onboard camera image. In: CVPR (2022) 4", + "5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020) 2, 4, 7, 10", + "6. Da, F., Zhang, Y.: Path-aware graph attention for hd maps in motion prediction. In: ICRA (2022) 1", + "7. Ding, W., Qiao, L., Qiu, X., Zhang, C.: Pivotnet: Vectorized pivot learning for end-to-end hd map construction. In: ICCV (2023) 2, 3, 4, 5, 11", + "8. Girshick, R.: Fast r-cnn. In: ICCV (2015) 4", + "9. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022) 9", + "0. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. In: ICCV (2017) 4", + "1. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) 10", + "2. He, S., Balakrishnan, H.: Lane-level street map extraction from aerial imagery. In: WACV (2022) 4, 10", + "3. He, S., Bastani, F., Abbar, S., Alizadeh, M., Balakrishnan, H., Chawla, S., Madden, S.: Roadrunner: improving the precision of road network inference fromgps trajectories. In: SIGSPATIAL (2018) 2, 10", + "4. He, S., Bastani, F., Jagwani, S., Alizadeh, M., Balakrishnan, H., Chawla, S., Elshrif, M.M., Madden, S., Sadeghi, M.A.: Sat2graph: road graph extraction through graph-tensor encoding. In: ECCV (2020) 3, 10", + "5. Homayounfar, N., Ma, W.C., Kowshika Lakshmikanth, S., Urtasun, R.: Hierarchical recurrent attention networks for structured online maps. In: CVPR (2018) 3", + "6. Homayounfar, N., Ma, W.C., Liang, J., Wu, X., Fan, J., Urtasun, R.: Dagmapper: Learning to map by discovering lane topology. In: ICCV (2019) 3", + "7. Lee, Y., Park, J.: Centermask: Real-time anchor-free instance segmentation. In: CVPR (2020) 11", + "8. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: CVPR (2022) 4", + "9. Li, F., Zhang, H., Xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y.: Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In: CVPR (2023) 4", + "20. Li, Q., Wang, Y., Wang, Y., Zhao, H.: Hdmapnet: A local semantic map learning and evaluation framework. In: ICRA (2022) 2, 4", + "21. Li, S., Yang, K., Shi, H., Zhang, J., Lin, J., Teng, Z., Li, Z.: Bi-mapper: Holistic bev semantic mapping for autonomous driving. RAL (2023) 4", + "22. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: ECCV (2022) 2, 4, 5" + ], + "bbox": [ + 225, + 172, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "InsMapper", + "bbox": [ + 656, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "23. Liang, M., Yang, B., Hu, R., Chen, Y., Liao, R., Feng, S., Urtasun, R.: Learning lane graph representations for motion forecasting. In: ECCV (2020) 1, 2", + "24. Liao, B., Chen, S., Wang, X., Cheng, T., Zhang, Q., Liu, W., Huang, C.: Maptr: Structured modeling and learning for online vectorized hd map construction. In: ICLR (2023) 2, 3, 4, 5, 7", + "25. Liao, B., Chen, S., Zhang, Y., Jiang, B., Zhang, Q., Liu, W., Huang, C., Wang, X.: Maptrv2: An end-to-end framework for online vectorized hd map construction. arXiv:2308.05736 (2023) 2, 3, 4, 5, 7, 11, 12", + "26. Liu, T., hai Liao, Q., Gan, L., Ma, F., Cheng, J., Xie, X., Wang, Z., Chen, Y., Zhu, Y., Zhang, S., et al.: The role of the hercules autonomous vehicle during the COVID-19 pandemic: An autonomous logistic vehicle for contactless goods transportation. IEEE Robotics & Automation Magazine (2021) 2", + "27. Liu, Y., Wang, Y., Wang, Y., Zhao, H.: Vectormapnet: End-to-end vectorized hd map learning. In: ICML (2023) 1, 2, 4", + "28. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: ICRA (2023) 4", + "29. Mi, L., Zhao, H., Nash, C., Jin, X., Gao, J., Sun, C., Schmid, C., Shavit, N., Chai, Y., Anguelov, D.: Hdmapgen: A hierarchical graph generative model of high definition maps. In: CVPR (2021) 4", + "30. Ng, M.H., Radia, K., Chen, J., Wang, D., Gog, I., Gonzalez, J.E.: Bev-seg: Bird's eye view semantic segmentation using geometry and semantic point cloud. In: CVPRW (2020) 2, 4", + "31. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: ECCV (2020) 2, 4", + "32. Qiao, L., Ding, W., Qiu, X., Zhang, C.: End-to-end vectorized hd-map construction with piecewiseBezier curve. In: CVPR (2023) 2, 4", + "33. Qiao, Z., Yu, Z., Yin, H., Shen, S.: Online monocular lane mapping using catmullrom spline. In: IROS (2023) 4", + "34. Ross, S., Gordon, G., Bagnell, D.: A reduction of imitation learning and structured prediction to no-regret online learning. In: AISTATS (2011) 3", + "35. Shin, J., Rameau, F., Jeong, H., Kum, D.: Instagram: Instance-level graph modeling for vectorized hd map learning. In: CVPRW (2023) 2, 4", + "36. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: AAAI (2022) 4", + "37. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. In: NeurIPS (2021) 3, 10, 12", + "38. Xie, Z., Pang, Z., Wang, Y.: Mv-map: Offboard hd-map generation with multi-view consistency. In: ICCV (2023) 2, 4", + "39. Xu, Z., Liu, Y., Gan, L., Sun, Y., Liu, M., Wang, L.: Ringdet: Road network graph detection by transformer in aerial images. TGRS (2022) 3", + "40. Xu, Z., Liu, Y., Sun, Y., Liu, M., Wang, L.: Centerlinedet: Road lane centerline graph detection with vehicle-mounted sensors by transformer for high-definition map creation. In: ICRA (2023) 2, 4", + "41. Xu, Z., Liu, Y., Sun, Y., Liu, M., Wang, L.: Ringdet++: Road network graph detection by transformer with instance segmentation and multi-scale features enhancement. RAL (2023) 2, 3, 10", + "42. Xu, Z., Sun, Y., Liu, M.: imputation learning-based detection of road curbs using aerial images for autonomous driving. RAL (2021) 3" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Xu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Xu, Z., Sun, Y., Liu, M.: Topo-boundary: A benchmark dataset on topological road-boundary detection using aerial images for autonomous driving. RAL (2021) 3", + "44. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors (2018) 11", + "45. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In: ICLR (2023) 4", + "46. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: ICLR (2021) 2, 4, 5" + ], + "bbox": [ + 212, + 146, + 787, + 286 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "InsMapper", + "bbox": [ + 656, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/InsMapper_ Exploring Inner-instance Information for Vectorized HD Mapping/703739e6-709f-487f-9515-53b438aabcdc_model.json b/2024/InsMapper_ Exploring Inner-instance Information for Vectorized HD Mapping/703739e6-709f-487f-9515-53b438aabcdc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..242a537ba198b39444f758b32b574fa1b0b6dfb4 --- /dev/null +++ b/2024/InsMapper_ Exploring Inner-instance Information for Vectorized HD Mapping/703739e6-709f-487f-9515-53b438aabcdc_model.json @@ -0,0 +1,2280 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.22, + 0.141, + 0.783, + 0.187 + ], + "angle": 0, + "content": "InsMapper: Exploring Inner-instance Information for Vectorized HD Mapping" + }, + { + "type": "text", + "bbox": [ + 0.279, + 0.213, + 0.726, + 0.228 + ], + "angle": 0, + "content": "Zhenhua Xu Kwan-Yee K. Wong Hengshuang Zhao\\*" + }, + { + "type": "text", + "bbox": [ + 0.402, + 0.24, + 0.603, + 0.255 + ], + "angle": 0, + "content": "The University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.28, + 0.726, + 0.296 + ], + "angle": 0, + "content": "Project webpage: https://tonyxuqq.github.io/InsMapper" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.319, + 0.744, + 0.597 + ], + "angle": 0, + "content": "Abstract. Vectorized high-definition (HD) maps contain detailed information about surrounding road elements, which are crucial for various downstream tasks in modern autonomous vehicles, such as motion planning and vehicle control. Recent works attempt to directly detect the vectorized HD map as a point set prediction task, achieving notable detection performance improvements. However, these methods usually overlook and fail to analyze the important inner-instance correlations between predicted points, impeding further advancements. To address this issue, we investigate the utilization of inner-instance information for vectorized high-definition mapping through transformers, and propose a powerful system named InsMapper, which effectively harnesses inner-instance information with three exquisite designs, including hybrid query generation, inner-instance query fusion, and inner-instance feature aggregation. The first two modules can better initialize queries for line detection, while the last one refines predicted line instances. InsMapper is highly adaptable and can be seamlessly modified to align with the most recent HD map detection frameworks. Extensive experimental evaluations are conducted on the challenging NuScenes and Argovverse 2 datasets, where InsMapper surpasses the previous state-of-the-art method, demonstrating its effectiveness and generality." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.61, + 0.641, + 0.624 + ], + "angle": 0, + "content": "Keywords: Autonomous Driving \\(\\cdot\\) High-definition Map" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.65, + 0.377, + 0.667 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.788, + 0.786 + ], + "angle": 0, + "content": "Vectorized high-definition maps (HD maps) play a critical role in today's autonomous vehicles [6, 23, 27], as they contain detailed information about the road, including the position of road elements (e.g., road boundaries, lane splits, pedestrian crossings, and lane centerlines), connectivity, and topology of the road. Without the assistance of HD maps for perceiving and understanding road elements, unexpected vehicle behaviors may be encountered, such as incorrect path planning results or even vehicle collisions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.787, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Typically, HD maps are created by offline human annotation, which is labor-intensive, inefficient, and expensive. Although there are works proposing to make" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.386, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Z. Xu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "such an offline process automatic [38, 40, 41], it is not possible to recreate and update the HD map frequently when the road has been modified, such as when a new road is built or an existing road is removed. To address this issue, several recent studies propose to detect local HD maps using an online method [7, 20, 24, 25, 27, 32]. This paper aims to further investigate online HD map detection techniques based on vehicle-mounted sensors (i.e., cameras and LiDARs)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.252, + 0.788, + 0.373 + ], + "angle": 0, + "content": "Early online HD map detection works consider road element mapping as a semantic segmentation task in bird's-eye view (BEV) [22,30,31], in which road elements are predicted in raster format (i.e., pixel-level semantic segmentation mask). However, rasterized road maps cannot be effectively utilized by downstream tasks of autonomous vehicles, such as motion planning and vehicle control [23,26,27]. Moreover, it is challenging to distinguish instances from the rasterized map, especially when some road elements overlap with each other or when a complicated topology is encountered (e.g., road split or road intersections)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.482, + 0.388 + ], + "angle": 0, + "content": "To alleviate these problems, HDMap-" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.461, + 0.78 + ], + "angle": 0, + "content": "Net [20] proposes hand-crafted post-processing algorithms to better obtain road element instances for vectorized HD maps. However, HDMapNet still heavily relies on rasterized results, restricting it from handling complicated urban scenes. Recently, some works have resorted to predicting vectorized HD maps directly [3,24,27,35]. A two-stage hierarchical set prediction method is proposed in VectorMapNet [27]. After predicting key points of road elements in the HD map, VectorMapNet sequentially generates intermediate points. Although VectorMapNet is considered to be the first online vectorized HD map detection work, the sequential operation degrades its efficiency and model performance. MapTR [24] further proposes to use DETR [5, 46] for vectorized HD map detection as a point set prediction problem, and the output point sets are then" + }, + { + "type": "image", + "bbox": [ + 0.472, + 0.404, + 0.787, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.599, + 0.789, + 0.753 + ], + "angle": 0, + "content": "Fig.1: Comparison of vectorized HD map detection methods. All solutions are evaluated on the NuScenes validation set. The x-axis represents the mAP results and the y-axis displays topology level correctness [13]. InsMapper is flexible and adaptable, so it can be seamlessly modified to align with multiple existing frameworks. InsMapper is built on MapTR [24]. \\(\\dagger\\) means InsMapper based on more recent framework PivotNet [7], while \\(\\ddagger\\) indicates InsMapper based on previous SOTA, MapTR-V2 [25]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "grouped into road element instances. Some recent works further improve the performance of MapTR [7, 25, 32]. Among them, MapTR-V2 achieves state-of-the-art performance, but it overlooks and fails to utilize the essential inner-instance correlations between points to further boost the performance." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "InsMapper" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.389 + ], + "angle": 0, + "content": "In this work, we focus on investigating the utilization of inner-instance information for vectorized high-definition mapping via transformer, and propose a new framework named InsMapper, which can utilize inner-instance point information for improved online HD map detection via three exquisite designs. First, a hybrid query generation module is introduced to better initialize queries for detection. Then, an inner-instance query fusion module is added before the transformer decoder, which fuses inner-instance object queries to refine the initialized line instances. Finally, a transformer-based inner-instance aggregation module is incorporated to further refine the predicted lines. Extensive experimental evaluations are conducted on the popular and challenging NuScenes [2] and Argoverse 2 [37] datasets. InsMapper is based on MapTR [24], and it can be seamlessly modified based on the latest frameworks with more enhanced performance, such as PivotNet [7] and MapTR-V2 [25]. No matter what the base framework is, InsMapper harvests state-of-the-art performance, surpassing the previous highest solutions. The intuitive performance comparisons are illustrated in Figure 1, and our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.403, + 0.787, + 0.463 + ], + "angle": 0, + "content": "- We analyze the limitations of existing HD map detection algorithms and propose a new framework named InsMapper, which can effectively utilize the easily overlooked inner-instance point correlation information for accurate and generalizable online HD map detection." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.464, + 0.787, + 0.523 + ], + "angle": 0, + "content": "- We incorporate three useful modules to leverage inner-instance information, including hybrid query generation, inner-instance query fusion, and inner-instance feature aggregation. The first two modules better initialize queries for detection and the last one refines detected line instances." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.525, + 0.802, + 0.57 + ], + "angle": 0, + "content": "- We conduct experiments on the challenging NuScenes and Argoverse 2 datasets. InsMapper outperforms all baselines by a large margin. The achieved state-of-the-art results demonstrate its effectiveness and generality." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.403, + 0.802, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.595, + 0.388, + 0.611 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Vector map detection. Vector maps use vectors to represent road elements, such as road networks, road curbs, and lane lines. A high-definition map (HD map) is a type of vector map, but a vector map may not qualify as an HD map if it lacks sufficient resolution or fails to provide lane-level information, among other factors. Online HD map detection can benefit from insights gained from a broader scope of vector map detection works, including road-network detection [1,14,39, 41], road-curb detection in aerial images [42, 43], and road lane line detection [15, 16], etc. DagMapper [16] aims to detect vectorized lane lines from pre-built point cloud maps using iterations. However, DagMapper only handles simple topology changes of lane lines and struggles to handle complex road intersections. RNGDet++ [41] applies DETR along with an imitation learning algorithm [34] for road-network detection, achieving state-of-the-art performance on multiple datasets. Although these methods achieve satisfactory graph detection accuracy, they suffer from poor efficiency due to their iterative algorithms. Given the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "Z. Xu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "stringent real-time requirements of autonomous vehicles, they are not suitable for online HD map detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.433 + ], + "angle": 0, + "content": "HD map detection. HD map detection was initially a subtask of BEV detection [21, 22, 28, 30, 31]. Recently, given the importance of HD maps, several works have directly focused on HD map detection [4, 20, 24, 27, 29, 33, 35, 38, 40]. However, most of these works either involve offline HD map creation [12, 38, 40] or only detect one specific road element [3, 4, 33]. HDMapNet [20] is considered to be the first work specifically designed for multiple road element detection. However, HDMapNet outputs rasterized results, requiring complicated handcrafted post-processing to obtain vectorized HD maps. To address this issue, VectorMapNet [27] is believed to be the first work detecting vectorized HD maps in real-time. However, it consists of two stages, and its efficiency is significantly impacted by sequential operations. In contrast, MapTR [24] uses deformable DETR [46] to design an end-to-end vectorized HD map detection model, which greatly simplifies the pipeline and delivers better detection performance. Some recent works further improve the performance of MapTR [7, 25, 32]. Among them, MapTR-V2 achieves state-of-the-art performance on vectorized HD map detection, but it still overlooks and does not investigate the inner-instance correlations between points, limiting further improvements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Detection by Transformer. DETR [5] is the first end-to-end transformer-based object detection framework. Compared with previous CNN-based methods [8, 10], DETR removes the need for anchor proposals and non-maximum suppression (NMS), making it simpler and more effective. To address the issue of slow convergence, subsequent works propose accelerating DETR training through deformable attention [46], denoising [18], and dynamic query generation [19, 36, 45]. Intuitively, most of these DETR refinements could be adapted to the HD map detection task since our proposed InsMapper relies on DETR. However, unlike typical object detection tasks in which objects are assumed to be independent and identically distributed (i.i.d.), detected points in HD maps have strong correlations, especially among inner-instance points. This inherent difference makes some refined DETR methods unsuitable to our task and additional investigations are required." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.652, + 0.424, + 0.667 + ], + "angle": 0, + "content": "3 Point Correlation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.681, + 0.607, + 0.697 + ], + "angle": 0, + "content": "3.1 Vector Map Decomposition and Sampling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Let \\( G \\) denote the original vector map label of the scene, which consists of vertices \\( V \\) and edges \\( E \\). The vector map contains multiple classes of road elements, including pedestrian crossings, road dividers, road boundaries, and lane centerlines. Among them, the first three classes of road elements are simple polylines or polygons without intersection points. While lane centerlines have more complicated topology, such as lane split, lane merge, and lane intersection. To unify all vector elements, the vector map is decomposed into simpler shapes (i.e., polylines and polygons) without intersections. Any vertices in the vector map with degrees larger than two (i.e., intersection vertices) are removed from \\( G \\), but" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "InsMapper" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "incident edges are kept. In this way, a set of simple polylines and polygons without intersections is obtained as \\( G^{*} = \\{l_{i}\\}_{i=1}^{N^{*}} \\), where \\( G^{*} \\) is an undirected graph. Each shape \\( l_{i} \\) is defined as an instance, and \\( N^{*} \\) denotes the overall number of instances in a vector map." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.284 + ], + "angle": 0, + "content": "Following MapTR and MapTR-V2, to enhance the parallelization capacity of the model, each instance is evenly re-sampled with fixed-length points as \\( l_{i} = (v_{1},\\dots,v_{j},\\dots,v_{n_{p}}) \\). \\( l_{i} \\) is ordered from \\( v_{1} \\) to \\( v_{n_p} \\), with \\( n_p \\) being the number of sampled points for each instance. For polygon instances, \\( v_{1} \\) is equal to \\( v_{n_p} \\). Please refer to the supplementary document for the visualization of the pre-processing." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.307, + 0.48, + 0.321 + ], + "angle": 0, + "content": "3.2 Inner-instance Correlation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.784, + 0.408 + ], + "angle": 0, + "content": "Unlike conventional object detection tasks, where objects can be approximated as independent and identically distributed (i.i.d.), strong correlations exist between points within the same line instance in the vectorized HD map detection task. Please refer to the supplementary document for the visualization of the point correlation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.785, + 0.56 + ], + "angle": 0, + "content": "Inner-instance correlation is crucial for point coordinates prediction. Points within the same instance can collaborate by sharing inner-instance information, leading to smoother and more accurate predictions. Without this collaboration, points of the same instance may produce independent predictions, leading to zig-zag or incorrect instances. The inner-instance correlation can serve as additional constraints to facilitate the query initialization and refinement of predicted lines. In previous methods, the correlation of points is not correctly analyzed and leveraged, limiting further improvement. In subsequent sections, we introduce InsMapper to more effectively utilize point correlations and improve vectorized map detection performance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.585, + 0.381, + 0.603 + ], + "angle": 0, + "content": "4 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.618, + 0.421, + 0.632 + ], + "angle": 0, + "content": "4.1 Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.784, + 0.673 + ], + "angle": 0, + "content": "Building on MapTR [24], our proposed InsMapper is an end-to-end trainable framework for online vectorized HD map detection, as illustrated in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.785, + 0.84 + ], + "angle": 0, + "content": "InsMapper is adaptive and can be seamlessly modified based on the latest frameworks with more enhanced performance, such as PivotNet [7] and MapTR-V2 [25]. Following BEVformer [22], InsMapper first projects perspective view camera images into the bird's-eye-view (BEV). After obtaining the BEV features, InsMapper uses deformable attention layers [46] to process input object queries. Each object query predicts one point, including the object class and the regression point position. To better leverage inner-instance information and further enhance the final detection performance, we introduce three designs in InsMapper: hybrid query generation, which replaces the hierarchical query in MapTR, generating line queries with better diversity; inner-instance query fusion, which fuses generated queries within the same instance based on" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Z. Xu et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.146, + 0.788, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.406, + 0.788, + 0.518 + ], + "angle": 0, + "content": "Fig. 2: Overall framework. InsMapper is an end-to-end transformer model with an encoder-decoder structure. The transformer encoder projects perspective-view camera images into a bird's-eye view (BEV). Subsequently, the transformer decoder detects vector instances by predicting point sets. To enhance the utilization of inner-instance information, we introduce the following three components: a hybrid query generation scheme (orange module), an inner-instance query fusion module (yellow module), and an inner-instance feature aggregation module (blue module). The first two modules better initialize queries for detection and the final one refines detected line instances." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.607 + ], + "angle": 0, + "content": "inner-instance features to further enhance the quality of initialized line queries; and inner-instance feature aggregation, which adds a new inner-instance self-attention module to decoder layers to further aggregate inner-instance features for line refinement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.788, + 0.73 + ], + "angle": 0, + "content": "Different from past works, which treat queries as independently distributed, InsMapper utilizes inner-instance information to fuse object queries and aggregate intermediate features in deformable transformer decoder layers. Otherwise, adjacent vertices may produce independent position regression results, leading to zigzag or even incorrect instances. With the assistance of inner-instance information, the model can better initialize queries for detection and refine the positions of detected points within an instance, significantly enhancing the final performance, as demonstrated in the experiments section." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.753, + 0.414, + 0.769 + ], + "angle": 0, + "content": "4.2 Query Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In contrast to traditional object detection tasks, where queries are independent and identically distributed (i.i.d.), the points to be detected in our task exhibit significant correlations. Various methods can be used to generate object queries as input for the transformer decoder, such as the basic scheme, hierar" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "InsMapper" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.175, + 0.486, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.293, + 0.487, + 0.378 + ], + "angle": 0, + "content": "Fig.3: Query generation schemes. For concise visualization, the number of instances \\(N_{I}\\) is 2, and the number of points per instance \\(n_p\\) is 3. The hybrid scheme can initialize lines with both good diversity and quality." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.151, + 0.593, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.284, + 0.579, + 0.296 + ], + "angle": 0, + "content": "(a) Basic" + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.149, + 0.692, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.603, + 0.283, + 0.697, + 0.296 + ], + "angle": 0, + "content": "(b) Hierarchical" + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.15, + 0.788, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.284, + 0.781, + 0.296 + ], + "angle": 0, + "content": "(c) Hybrid" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.307, + 0.8, + 0.377 + ], + "angle": 0, + "content": "Fig. 4: Visualization of initialized lines with different query generation schemes. The quality and diversity of initialized lines are crucial for the final performance. Each color represents one initialized line." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.785, + 0.451 + ], + "angle": 0, + "content": "chical scheme, and hybrid scheme. All query generation schemes are illustrated in Figure 3. Initialized queries are organized into line instances. The quality and diversity of the initialized lines are both crucial." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.588 + ], + "angle": 0, + "content": "Basic query generation. This method is leveraged in the original DETR [5] for query generation. It assumes that points are i.i.d. and generates queries randomly, without utilizing inner-instance information. Let \\( N_{I} \\) denote the number of predicted instances, which is obviously larger than \\( N^{*} \\). Since each instance contains \\( n_p \\) points, the basic object queries consist of \\( N_{I} \\cdot n_{p} \\) randomly generated i.i.d. queries, resulting in irregular line shapes. Due to the lack of constraints on inner-instance points, the initialized line instances are noisy and have poor quality, which leads to degraded performance. MapTR with the basic query scheme is visualized in Figure 4a." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.589, + 0.788, + 0.666 + ], + "angle": 0, + "content": "Hierarchical query generation. To address the aforementioned issue of the basic scheme, a hierarchical query generation scheme is proposed in MapTR [24] and MapTR-V2 [25]. Let \\( Q_{ins} = \\{q_i^I\\}_{i=1}^{N_I} \\) denote instance queries and \\( Q_{pts} = \\{q_j^P\\}_{j=1}^{n_p} \\) denote point queries. The object queries for HD map detection are then the pairwise addition of \\( Q_{ins} \\) and \\( Q_{pts} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.346, + 0.674, + 0.787, + 0.695 + ], + "angle": 0, + "content": "\\[\nQ = \\left\\{q _ {i, j} = q _ {i} ^ {I} + q _ {j} ^ {P} \\mid q _ {i} ^ {I} \\in Q _ {\\text {i n s}}, q _ {j} ^ {P} \\in Q _ {\\text {p t s}} \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.81 + ], + "angle": 0, + "content": "where \\( |Q| = |Q_{ins}| \\cdot |Q_{pts}| = N_I \\cdot n_p \\). For each line instance \\( i \\), the instance query \\( q_i^I \\) denotes its uniqueness. By using \\( q_i^I \\) as a bridge, object queries within the same instance can better collaborate with each other for point prediction. But all line instances share the same point queries so all initialized lines tend to have quite similar shapes, degrading initialization diversity, which harms the final detection performance. The initialized lines generated by MapTR with the hierarchical query scheme are visualized in Figure 4b." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Hybrid query generation. The basic scheme does not take into account any information exchange between queries, whereas the hierarchical one generates" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Z. Xu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "repetitive line shapes. To address these issues, this paper presents a hybrid query generation method that mitigates the drawbacks of the aforementioned schemes while maintaining appropriate inner-instance information exchange." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.785, + 0.284 + ], + "angle": 0, + "content": "Let the instance queries be \\( Q_{ins} = \\{q_i^I\\}_{i=1}^{N_I} \\), and the point queries be \\( Q_{pts} = \\{q_j^P\\}_{j=1}^{N_P} = \\{q_j^P\\}_{j=1}^{N_I \\cdot n_p} \\). The point queries are divided into \\( N_I \\) instance groups, and a point query \\( q_j^P \\) is assigned to the \\( \\lceil \\frac{j}{n_p} \\rceil \\)-th instance. Consequently, the final object query is the sum of a point query \\( p_j^P \\) and its assigned instance query \\( p_k^I \\), where \\( k = \\lceil \\frac{j}{n_p} \\rceil \\). The object query set can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.295, + 0.786, + 0.314 + ], + "angle": 0, + "content": "\\[\nQ = \\left\\{q _ {k} ^ {I} + q _ {j} ^ {P} \\mid q _ {j} ^ {P} \\in Q _ {p t s} \\right\\} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.318, + 0.785, + 0.44 + ], + "angle": 0, + "content": "where \\( |Q| = |Q_{pts}| = N_P = N_I \\cdot n_p \\). In contrast to the hierarchical scheme, each point query in the hybrid scheme is utilized only once to generate the object query, thereby preventing repetitive line shapes. Simultaneously, point queries belonging to the same instance are summed with a shared instance query, establishing the inner-instance connection. The hybrid query generation method can be considered as a combination of the basic and hierarchical schemes, effectively mitigating their respective drawbacks. InsMapper with the hybrid query generation scheme is visualized in Figure 4c." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.463, + 0.498, + 0.479 + ], + "angle": 0, + "content": "4.3 Inner-instance Query Fusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.785, + 0.594 + ], + "angle": 0, + "content": "The input object queries for the transformer decoder are generated from instance queries and point queries. Although the generated queries can leverage some inner-instance information, this information exchange is indirect and inflexible. The instance query is distributed equally among all inner-instance points, while a more precise point-to-point information exchange cannot be realized. As a result, a query fusion module is introduced to further utilize inner-instance information and refine initialized lines." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.785, + 0.657 + ], + "angle": 0, + "content": "Let \\( Q_{i} = \\{q_{i,j}\\}_{j=1}^{n_{p}} \\) represent the set of object queries belonging to the \\( i \\)-th instance, and \\( q_{i,j} \\) denote the \\( j \\)-th point of the \\( i \\)-th instance. \\( q_{i,j} \\) is correlated with all other queries in \\( Q_{i} \\). To better fuse inner-instance object queries, each query is updated by a weighted sum of all queries in \\( Q_{i} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.669, + 0.786, + 0.709 + ], + "angle": 0, + "content": "\\[\nq _ {i, j} = f \\left(q _ {i, j}, Q _ {i}\\right) = \\sum_ {k = 1} ^ {n _ {p}} w _ {i, j, k} \\phi \\left(q _ {i, k}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.78 + ], + "angle": 0, + "content": "where \\( w_{i,j,k} \\) demonstrates weights for query fusion. \\( f(\\cdot) \\) is the fusion function and \\( \\phi(\\cdot) \\) is the kernel in case nonlinear transformation is needed. \\( f(\\cdot) \\) could be realized by handcraft weights, fully connected layers, or self-attention. In this work, self-attention is applied as \\( f(\\cdot) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.781, + 0.785, + 0.841 + ], + "angle": 0, + "content": "In conventional object detection tasks, object queries are assumed to be independent and identically distributed (i.i.d.), making query fusion unnecessary. However, in the task of HD map detection, the query fusion module effectively aligns the update of queries and enables each point to pay more attention to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "InsMapper" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.169, + 0.331, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.337, + 0.319, + 0.349 + ], + "angle": 0, + "content": "(a) No fusion" + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.171, + 0.452, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.357, + 0.337, + 0.45, + 0.349 + ], + "angle": 0, + "content": "(b) With fusion" + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.148, + 0.79, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.36, + 0.465, + 0.498 + ], + "angle": 0, + "content": "Fig.5: Visualization of the effect of the inner-instance query fusion module. (a) No query fusion. (b) With query fusion. The query fusion module can better refine the initialized lines based on inner-instance information. In this example, the proposed query fusion module significantly enhances the diversity of initialized line instances." + }, + { + "type": "image_caption", + "bbox": [ + 0.475, + 0.36, + 0.794, + 0.499 + ], + "angle": 0, + "content": "Fig.6: Decoder of InsMapper. An inner-instance multi-head self-attention module is incorporated into decoder layers. In this module, the attention between points belonging to different instances is blocked (grey grids). Only inner-instance attention is allowed (colored grids). Colored grids are randomly blocked with \\(\\epsilon\\) probability (set to grey) for robustness. The proposed module can further utilize inner-instance information and refine the predicted lines." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.788, + 0.607 + ], + "angle": 0, + "content": "neighboring points. It can greatly refine the initialized lines for better query initialization, leading to enhanced performance. Without this module, queries within the same instance cannot be aware of each other, rendering them \"blind\". Poor line initialization tends to have degraded detection scores. The effect of the query fusion module is visualized in Figure 5." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.631, + 0.555, + 0.646 + ], + "angle": 0, + "content": "4.4 Inner-instance Feature Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.735 + ], + "angle": 0, + "content": "In addition to object query manipulation, InsMapper performs inner-instance feature aggregation for predicted line refinement in the transformer decoder layers by incorporating an additional inner-instance self-attention module. Inspired by [9], inner-instance attentions are randomly blocked with probability \\(\\epsilon\\) for robustness. The decoder layer of InsMapper is depicted in Figure 6." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The inner-instance self-attention module resembles the original self-attention module but features a specially designed attention mask. As illustrated in Figure 6, the attention mask of inner-instance points is set to zero (colored grids), indicating that attention between points within the same instance is allowed. Conversely, for points belonging to different instances (i.e., inter-instance points), the corresponding attention mask values are set to one, blocking attention between them. This method encourages the model to focus more on inner-instance" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "Z. Xu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "information, resulting in more consistent predictions. To further enhance the robustness of this module, the inner-instance attention (colored grids) has an \\( e \\) probability of being blocked." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.329 + ], + "angle": 0, + "content": "An alternative method involves placing the proposed inner-instance attention module before the cross-attention module. However, the self-attention module before cross-attention should not block inter-instance information. Otherwise, some instances may produce duplicate predictions since they cannot \"clearly see\" other instances, which is analyzed in the original DETR [5]. Consequently, implementing inner-instance self attention before the cross-attention module leads to degraded final performance. Thus the proposed inner-instance feature aggregation should be placed after the cross-attention layer. More experiments about the decoder structure are provided in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.347, + 0.377, + 0.364 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.374, + 0.429, + 0.39 + ], + "angle": 0, + "content": "5.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.785, + 0.546 + ], + "angle": 0, + "content": "Datasets. Extensive experiments are conducted on the popular and challenging NuScenes dataset [2] and the Argoverse 2 dataset [37]. These two datasets comprise hundreds of data sequences captured in various autonomous driving scenarios, encompassing diverse weather conditions and time periods. The perception range for the X-axis and Y-axis in the BEV is set to \\([-15m, 15m]\\) and \\([-30m, 30m]\\), respectively. The detected vectorized HD map should encompass four types of road elements simultaneously: road boundary, lane split, pedestrian crossing, and lane centerline. Among these, the lane centerline is a crucial element for ensuring effective vehicle planning and control. Detecting lane centerlines poses greater difficulty compared to the other three elements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.794, + 0.666 + ], + "angle": 0, + "content": "Implementation details. We perform all the experiments on a machine equipped with 8 RTX-3090 GPUs. During the training phase, all GPUs are utilized, whereas only a single GPU is employed for inference. For our proposed InsMapper, we adopt multiple settings akin to the previous SOTA method, MapTR. All ablation studies are conducted for 24 epochs, with ResNet50 [11] as the backbone. For a fair comparison, except for the proposed modules, we keep the experiment settings exactly the same, such as batch size, network depth, the number of input queries, etc." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.787, + 0.773 + ], + "angle": 0, + "content": "Evaluation metrics. Models are assessed using two types of metrics. The first metric, average precision (AP), gauges the instance-level detection performance, employing Chamfer distance for matching predictions with ground truth labels. To ensure a fair comparison, we follow previous works to calculate multiple \\(\\mathrm{AP}_{\\tau}\\) values with \\(\\tau \\in \\{0.5, 1.0, 1.5\\}\\), and report the average AP. To measure the topology correctness, the metric TOPO that is widely used in past works [12-14, 41] is reported. For all metrics, a larger value indicates better performance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.79, + 0.473, + 0.805 + ], + "angle": 0, + "content": "5.2 Performance Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "In this section, we compare InsMapper with previous state-of-the-art methods using the aforementioned evaluation metrics. The quantitative comparison results" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "InsMapper" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.229 + ], + "angle": 0, + "content": "Table 1: Quantitative results of comparison experiments on the NuScenes validation set. Colored numbers show differences between InsMapper and the SOTA baseline under the same experiment setting. \"C\" and \"L\" denote camera and LiDAR, respectively. \"-\" denotes that the result is not available. \"V2-99\" and \"Sec\" correspond to VoVNetV2-99 [17] and SECOND [44]. \\(\\dagger\\) represents InsMapper based on PivotNet [7], \\(\\ddagger\\) indicates InsMapper based on MapTR-V2 [25]." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.24, + 0.787, + 0.459 + ], + "angle": 0, + "content": "
MethodsEpochsBackboneModality\\(AP_{ped}\\)\\(AP_{div}\\)\\(AP_{bound}\\)\\(AP_{center}\\)mAPTOPO
HDMapNet30Effi-B0C4.4123.7358.1737.2930.9029.79
STSU110R50C---31.2131.2132.11
VectorMapNet130R50C28.6639.7433.0634.9334.1039.90
MapTR24R50C37.9246.2550.0737.4742.9346.77
MapTR24V2-99C44.5756.2560.5846.3051.9255.16
MapTR24R50&SecC&L49.4758.9866.7244.8955.0156.77
MapTR110R50C49.1659.1258.9347.2653.6259.16
BeMapNet24R50C43.5954.1151.9038.6047.0549.90
PivotNet24R50C42.7353.5552.0142.4747.6951.04
MapTR-V224R50C48.2054.7556.6645.8651.3751.15
MapTR-V2110R50C58.6865.7267.1256.1661.9263.26
InsMapper24R50C44.3653.3652.7742.3548.3151.58
InsMapper24V2-99C51.1663.7164.4751.4057.6860.00
InsMapper24R50&SecC&L56.0063.4271.6152.8560.9762.51
InsMapper110R50C55.4263.8765.8054.2059.4066.19
InsMapper†24R50C46.4554.9154.1645.2050.1853.57
InsMapper‡24R50C49.0757.9859.8447.9953.7253.00
InsMapper†110R50C62.0967.6068.1558.4164.0664.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.473, + 0.785, + 0.501 + ], + "angle": 0, + "content": "Table 2: Quantitative results of comparison experiments on the Argoverse 2 validation set. \\(\\ddagger\\) indicates InsMapper based on MapTR-V2 [25]." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.513, + 0.787, + 0.582 + ], + "angle": 0, + "content": "
MethodsEpochsBackboneModalityAPpedAPdivAPboundAPcentermAPTOPO
MapTR6R50C52.8863.6861.1859.7359.3775.79
MapTR-V26R50C57.1667.9665.2563.2063.3977.94
InsMapper6R50C55.6166.6062.5862.6761.8777.58
InsMapper†6R50C58.9171.3566.9064.7065.4679.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.596, + 0.785, + 0.625 + ], + "angle": 0, + "content": "Table 3: Quantitative results of comparison experiments on the NuScenes validation set without centerline. \\(\\ddagger\\) indicates InsMapper based on MapTR-V2 [25]." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.637, + 0.787, + 0.706 + ], + "angle": 0, + "content": "
MethodsEpochsBackboneModalityAPpedAPdivAPboundmAP
MapTR24R50C46.0451.5853.0850.23
MapTR-V224R50C59.8062.4062.4061.50
InsMapper24R50C48.4454.6856.9253.35
InsMapper‡24R50C61.5465.0964.6263.75
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "are reported in Table 1. InsMapper is flexible and adaptive, making it seamless to modify it based on existing transformer frameworks. InsMapper is built on MapTR for experiments, we also report the results of InsMapper with different base frameworks, e.g. InsMapper† is based on PivotNet [7] and InsMapper‡ relies on the previous SOTA method MapTR-V2 [25]. Compared to the previous methods, InsMapper improves the AP of all road elements by around 2. The topological metric also sees an improvement of around 1.5. The inference FPS of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "Z. Xu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 4: Quantitative results of comparison experiments on the NuScenes validation set. Centerlines are directed graphs in this table. \\(\\ddagger\\) indicates InsMapper based on MapTR-V2 [25]." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.199, + 0.789, + 0.269 + ], + "angle": 0, + "content": "
MethodsEpochsBackboneModality\\(AP_{ped}\\)\\(AP_{div}\\)\\(AP_{bound}\\)\\(AP_{center}\\)mAPTOPO
MapTR24R50C37.3946.6150.6937.2242.9846.98
MapTR-V224R50C49.5355.4658.4853.8254.3254.90
InsMapper24R50C43.2453.7753.8943.2248.5352.09
\\(InsMapper^{\\dagger}\\)24R50C51.6257.8860.9056.5256.7356.01
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.789, + 0.463 + ], + "angle": 0, + "content": "MapTR and InsMapper are 8.5 and 7.5, respectively. The evaluation outcomes on the Argoverse 2 dataset [37] can be found in Table 2. InsMapper outperforms past works with notable enhancements. Table 3 presents the results without centerlines. InsMapper with different bases attains an improvement of more than \\(2\\mathrm{mAP}\\). In the aforementioned experiments, all road elements are treated as undirected graphs. Table 4 shows the results when centerlines are represented as directed graphs. Based on the results, InsMapper still presents superior results than past works for directed HD map detection. Furthermore, we illustrate the qualitative visualizations in Figure 7. Owing to the effective employment of inner-instance information, InsMapper generates smoother and more precise HD maps compared to previous works." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.789, + 0.542 + ], + "angle": 0, + "content": "Consequently, InsMapper demonstrates superior performance compared to prior approaches across multiple datasets, exhibiting outstanding detection outcomes. Concurrently, InsMapper maintains a comparable inference speed to the previous state-of-the-art method. These exceptional experimental results strongly demonstrate the effectiveness and generality of the proposed designs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.568, + 0.403, + 0.582 + ], + "angle": 0, + "content": "5.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.789, + 0.703 + ], + "angle": 0, + "content": "Key component designs. We conduct several ablation study experiments to confirm the necessity of the proposed modules. Initially, we incorporate these modules into the previous state-of-the-art baseline model incrementally, and the results are presented in Table 5. It is evident that all three modules contribute to performance improvements, thereby validating their necessity. To further investigate the impact of each module's design, we perform comprehensive ablation studies in the following paragraphs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.794 + ], + "angle": 0, + "content": "Hybrid query generation. Compared to basic or hierarchical query generation schemes, the proposed hybrid query generation scheme generates queries with better quality and diversity, due to effectively utilizing inner-instance correlation. The evaluation results of the different query generation schemes are presented in Table 6. These results reveal that the hybrid query generation scheme outperforms its counterparts, thereby demonstrating the soundness of this design." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Inner-instance query fusion. Query fusion is proposed to leverage the correlation between inner-instance queries to enhance prediction performance. Various methods can be employed for fusing queries, including mean fusing (i.e.," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "InsMapper" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.573, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.144, + 0.784, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.568, + 0.788, + 0.611 + ], + "angle": 0, + "content": "Fig. 7: Qualitative visualization. The predicted map contains four classes, i.e., road boundaries (green), lane splits (red), pedestrian crossing (blue), and lane centerlines (pink). InsMapper presents better detection results compared with the past work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.718 + ], + "angle": 0, + "content": "each query is summed with the mean of all queries of the same instance), fusion by the feed-forward network, and fusion by self-attention. The evaluation results are presented in Table 7. The outcomes of the query fusion methods vary significantly, with self-attention-based query fusion achieving the best results. Consequently, self-attention query fusion is incorporated into InsMapper." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Inner-instance feature aggregation. In InsMapper, the inner-instance self-attention module is incorporated into the decoder layers following the cross-attention module. We evaluate InsMapper under four different conditions: without the inner-instance self-attention module (no attention), replace the inner-instance self-attention module with a vanilla self-attention module (vanilla attention), without the random blocking (no random blocking, i.e., \\(\\epsilon = 0\\)), and with the module placed before the cross-attention (change position). The results are shown in Table 8. The outcomes reveal that removing or altering the proposed" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Z. Xu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.158, + 0.492, + 0.226 + ], + "angle": 0, + "content": "Table 5: Ablation studies on key component designs. \"QG\", \"QF\" and \"FA\" represent hybrid query generation, inner-instance query fusion, and inner-instance feature aggregation, respectively." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.227, + 0.49, + 0.327 + ], + "angle": 0, + "content": "
QGQFFAmAPTOPO
42.9346.77
✓✓45.42(↑2.49)49.19(↑2.42)
✓✓45.45(↑2.52)48.54(↑1.77)
✓✓46.93(↑4.00)50.16(↑3.39)
✓✓✓48.31(↑5.38)51.58(↑4.81)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.338, + 0.492, + 0.364 + ], + "angle": 0, + "content": "Table 6: Ablation studies on inner-instance query generation." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.366, + 0.493, + 0.439 + ], + "angle": 0, + "content": "
Query SchememAPTOPO
InsMapper-basic46.6249.39
InsMapper-hierarchical46.9350.16
InsMapper-hybrid48.3151.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.158, + 0.786, + 0.185 + ], + "angle": 0, + "content": "Table 7: Ablation studies on inner-instance query fusion." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.186, + 0.787, + 0.272 + ], + "angle": 0, + "content": "
Query FusionmAPTOPO
No fusion45.4548.54
Mean43.3446.00
Feed-Forward47.8050.93
Self-attention48.3151.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.282, + 0.786, + 0.337 + ], + "angle": 0, + "content": "Table 8: Ablation studies on inner-instance feature aggregation. Multiple designs of the incorporated inner-instance self-attention layers are examined." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.339, + 0.784, + 0.438 + ], + "angle": 0, + "content": "
MethodmAPTOPO
No attention45.4249.19
Vanilla attention46.2349.51
No random blocking47.6751.49
Change position45.2448.16
InsMapper48.3151.58
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.784, + 0.504 + ], + "angle": 0, + "content": "module can impair model performance. This observation confirms the effectiveness of the proposed module design." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.545, + 0.36, + 0.563 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.793 + ], + "angle": 0, + "content": "In this paper, we introduced InsMapper, an end-to-end transformer-based model for on-the-fly vectorized HD map detection. InsMapper surpasses previous works by effectively leveraging inner-instance information to improve detection outcomes. Three exquisite designs have been proposed to utilize inner-instance information, including hybrid query generation, inner-instance query fusion, and inner-instance feature aggregation. The first two modules improve the quality of initialized line instances for detection, while the last module effectively refines the detected lines. The superiority of InsMapper is well exhibited through experiments on the challenging NuScenes and Argoverse 2 datasets. The evaluation results demonstrate the effectiveness and generality of InsMapper, making it a promising solution for the vectorized HD map detection task. In the future, InsMapper will be further enhanced for temporal-consistent vector map detection, making it more suitable for real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Acknowledgement. This work is supported by the National Natural Science Foundation of China (No. 62201484), HKU Startup Fund, and HKU Seed Fund for Basic Research." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "InsMapper" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.174, + 0.786, + 0.215 + ], + "angle": 0, + "content": "1. Bastani, F., He, S., Abbar, S., Alizadeh, M., Balakrishnan, H., Chawla, S., Madden, S., DeWitt, D.: Roadtracer: Automatic extraction of road networks from aerial images. In: CVPR (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.216, + 0.786, + 0.257 + ], + "angle": 0, + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020) 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.257, + 0.786, + 0.284 + ], + "angle": 0, + "content": "3. Can, Y.B., Liniger, A., Paudel, D.P., Van Gool, L.: Structured bird's-eye-view traffic scene understanding from onboard images. In: ICCV (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.284, + 0.786, + 0.311 + ], + "angle": 0, + "content": "4. Can, Y.B., Liniger, A., Paudel, D.P., Van Gool, L.: Topology preserving local road network estimation from single onboard camera image. In: CVPR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.786, + 0.338 + ], + "angle": 0, + "content": "5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020) 2, 4, 7, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.786, + 0.365 + ], + "angle": 0, + "content": "6. Da, F., Zhang, Y.: Path-aware graph attention for hd maps in motion prediction. In: ICRA (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.786, + 0.392 + ], + "angle": 0, + "content": "7. Ding, W., Qiao, L., Qiu, X., Zhang, C.: Pivotnet: Vectorized pivot learning for end-to-end hd map construction. In: ICCV (2023) 2, 3, 4, 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.535, + 0.405 + ], + "angle": 0, + "content": "8. Girshick, R.: Fast r-cnn. In: ICCV (2015) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.786, + 0.433 + ], + "angle": 0, + "content": "9. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.786, + 0.446 + ], + "angle": 0, + "content": "0. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. In: ICCV (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.786, + 0.472 + ], + "angle": 0, + "content": "1. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.472, + 0.786, + 0.499 + ], + "angle": 0, + "content": "2. He, S., Balakrishnan, H.: Lane-level street map extraction from aerial imagery. In: WACV (2022) 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.5, + 0.786, + 0.541 + ], + "angle": 0, + "content": "3. He, S., Bastani, F., Abbar, S., Alizadeh, M., Balakrishnan, H., Chawla, S., Madden, S.: Roadrunner: improving the precision of road network inference fromgps trajectories. In: SIGSPATIAL (2018) 2, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.541, + 0.786, + 0.582 + ], + "angle": 0, + "content": "4. He, S., Bastani, F., Jagwani, S., Alizadeh, M., Balakrishnan, H., Chawla, S., Elshrif, M.M., Madden, S., Sadeghi, M.A.: Sat2graph: road graph extraction through graph-tensor encoding. In: ECCV (2020) 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.786, + 0.621 + ], + "angle": 0, + "content": "5. Homayounfar, N., Ma, W.C., Kowshika Lakshmikanth, S., Urtasun, R.: Hierarchical recurrent attention networks for structured online maps. In: CVPR (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.622, + 0.786, + 0.65 + ], + "angle": 0, + "content": "6. Homayounfar, N., Ma, W.C., Liang, J., Wu, X., Fan, J., Urtasun, R.: Dagmapper: Learning to map by discovering lane topology. In: ICCV (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.786, + 0.676 + ], + "angle": 0, + "content": "7. Lee, Y., Park, J.: Centermask: Real-time anchor-free instance segmentation. In: CVPR (2020) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.786, + 0.704 + ], + "angle": 0, + "content": "8. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: CVPR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.704, + 0.786, + 0.745 + ], + "angle": 0, + "content": "9. Li, F., Zhang, H., Xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y.: Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In: CVPR (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.786, + 0.772 + ], + "angle": 0, + "content": "20. Li, Q., Wang, Y., Wang, Y., Zhao, H.: Hdmapnet: A local semantic map learning and evaluation framework. In: ICRA (2022) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.786, + 0.799 + ], + "angle": 0, + "content": "21. Li, S., Yang, K., Shi, H., Zhang, J., Lin, J., Teng, Z., Li, Z.: Bi-mapper: Holistic bev semantic mapping for autonomous driving. RAL (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "22. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: ECCV (2022) 2, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.174, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "Z. Xu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "23. Liang, M., Yang, B., Hu, R., Chen, Y., Liao, R., Feng, S., Urtasun, R.: Learning lane graph representations for motion forecasting. In: ECCV (2020) 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "24. Liao, B., Chen, S., Wang, X., Cheng, T., Zhang, Q., Liu, W., Huang, C.: Maptr: Structured modeling and learning for online vectorized hd map construction. In: ICLR (2023) 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.258 + ], + "angle": 0, + "content": "25. Liao, B., Chen, S., Zhang, Y., Jiang, B., Zhang, Q., Liu, W., Huang, C., Wang, X.: Maptrv2: An end-to-end framework for online vectorized hd map construction. arXiv:2308.05736 (2023) 2, 3, 4, 5, 7, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.313 + ], + "angle": 0, + "content": "26. Liu, T., hai Liao, Q., Gan, L., Ma, F., Cheng, J., Xie, X., Wang, Z., Chen, Y., Zhu, Y., Zhang, S., et al.: The role of the hercules autonomous vehicle during the COVID-19 pandemic: An autonomous logistic vehicle for contactless goods transportation. IEEE Robotics & Automation Magazine (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "27. Liu, Y., Wang, Y., Wang, Y., Zhao, H.: Vectormapnet: End-to-end vectorized hd map learning. In: ICML (2023) 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "28. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: ICRA (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "29. Mi, L., Zhao, H., Nash, C., Jin, X., Gao, J., Sun, C., Schmid, C., Shavit, N., Chai, Y., Anguelov, D.: Hdmapgen: A hierarchical graph generative model of high definition maps. In: CVPR (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.465 + ], + "angle": 0, + "content": "30. Ng, M.H., Radia, K., Chen, J., Wang, D., Gog, I., Gonzalez, J.E.: Bev-seg: Bird's eye view semantic segmentation using geometry and semantic point cloud. In: CVPRW (2020) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "31. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: ECCV (2020) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "32. Qiao, L., Ding, W., Qiu, X., Zhang, C.: End-to-end vectorized hd-map construction with piecewiseBezier curve. In: CVPR (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.549 + ], + "angle": 0, + "content": "33. Qiao, Z., Yu, Z., Yin, H., Shen, S.: Online monocular lane mapping using catmullrom spline. In: IROS (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "34. Ross, S., Gordon, G., Bagnell, D.: A reduction of imitation learning and structured prediction to no-regret online learning. In: AISTATS (2011) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "35. Shin, J., Rameau, F., Jeong, H., Kum, D.: Instagram: Instance-level graph modeling for vectorized hd map learning. In: CVPRW (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "36. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: AAAI (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.785, + 0.674 + ], + "angle": 0, + "content": "37. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. In: NeurIPS (2021) 3, 10, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "38. Xie, Z., Pang, Z., Wang, Y.: Mv-map: Offboard hd-map generation with multi-view consistency. In: ICCV (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.785, + 0.73 + ], + "angle": 0, + "content": "39. Xu, Z., Liu, Y., Gan, L., Sun, Y., Liu, M., Wang, L.: Ringdet: Road network graph detection by transformer in aerial images. TGRS (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.771 + ], + "angle": 0, + "content": "40. Xu, Z., Liu, Y., Sun, Y., Liu, M., Wang, L.: Centerlinedet: Road lane centerline graph detection with vehicle-mounted sensors by transformer for high-definition map creation. In: ICRA (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "41. Xu, Z., Liu, Y., Sun, Y., Liu, M., Wang, L.: Ringdet++: Road network graph detection by transformer with instance segmentation and multi-scale features enhancement. RAL (2023) 2, 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "42. Xu, Z., Sun, Y., Liu, M.: imputation learning-based detection of road curbs using aerial images for autonomous driving. RAL (2021) 3" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "InsMapper" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.188 + ], + "angle": 0, + "content": "43. Xu, Z., Sun, Y., Liu, M.: Topo-boundary: A benchmark dataset on topological road-boundary detection using aerial images for autonomous driving. RAL (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.218 + ], + "angle": 0, + "content": "44. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors (2018) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.218, + 0.788, + 0.259 + ], + "angle": 0, + "content": "45. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In: ICLR (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.215, + 0.259, + 0.788, + 0.287 + ], + "angle": 0, + "content": "46. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: ICLR (2021) 2, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.287 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/InsMapper_ Exploring Inner-instance Information for Vectorized HD Mapping/703739e6-709f-487f-9515-53b438aabcdc_origin.pdf b/2024/InsMapper_ Exploring Inner-instance Information for Vectorized HD Mapping/703739e6-709f-487f-9515-53b438aabcdc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..576f03f44e51db12dd1d33de7566b8e961f58291 --- /dev/null +++ b/2024/InsMapper_ Exploring Inner-instance Information for Vectorized HD Mapping/703739e6-709f-487f-9515-53b438aabcdc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30a6e6d5f581c8a152caa30b1778b97ffa27e619ad6213d6a3c82e06c7c85644 +size 3016423 diff --git a/2024/InsMapper_ Exploring Inner-instance Information for Vectorized HD Mapping/full.md b/2024/InsMapper_ Exploring Inner-instance Information for Vectorized HD Mapping/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eb3166ab23a8c734f20c7bcf6b2fdab39650a2ef --- /dev/null +++ b/2024/InsMapper_ Exploring Inner-instance Information for Vectorized HD Mapping/full.md @@ -0,0 +1,282 @@ +# InsMapper: Exploring Inner-instance Information for Vectorized HD Mapping + +Zhenhua Xu Kwan-Yee K. Wong Hengshuang Zhao\* + +The University of Hong Kong + +Project webpage: https://tonyxuqq.github.io/InsMapper + +Abstract. Vectorized high-definition (HD) maps contain detailed information about surrounding road elements, which are crucial for various downstream tasks in modern autonomous vehicles, such as motion planning and vehicle control. Recent works attempt to directly detect the vectorized HD map as a point set prediction task, achieving notable detection performance improvements. However, these methods usually overlook and fail to analyze the important inner-instance correlations between predicted points, impeding further advancements. To address this issue, we investigate the utilization of inner-instance information for vectorized high-definition mapping through transformers, and propose a powerful system named InsMapper, which effectively harnesses inner-instance information with three exquisite designs, including hybrid query generation, inner-instance query fusion, and inner-instance feature aggregation. The first two modules can better initialize queries for line detection, while the last one refines predicted line instances. InsMapper is highly adaptable and can be seamlessly modified to align with the most recent HD map detection frameworks. Extensive experimental evaluations are conducted on the challenging NuScenes and Argovverse 2 datasets, where InsMapper surpasses the previous state-of-the-art method, demonstrating its effectiveness and generality. + +Keywords: Autonomous Driving $\cdot$ High-definition Map + +# 1 Introduction + +Vectorized high-definition maps (HD maps) play a critical role in today's autonomous vehicles [6, 23, 27], as they contain detailed information about the road, including the position of road elements (e.g., road boundaries, lane splits, pedestrian crossings, and lane centerlines), connectivity, and topology of the road. Without the assistance of HD maps for perceiving and understanding road elements, unexpected vehicle behaviors may be encountered, such as incorrect path planning results or even vehicle collisions. + +Typically, HD maps are created by offline human annotation, which is labor-intensive, inefficient, and expensive. Although there are works proposing to make + +such an offline process automatic [38, 40, 41], it is not possible to recreate and update the HD map frequently when the road has been modified, such as when a new road is built or an existing road is removed. To address this issue, several recent studies propose to detect local HD maps using an online method [7, 20, 24, 25, 27, 32]. This paper aims to further investigate online HD map detection techniques based on vehicle-mounted sensors (i.e., cameras and LiDARs). + +Early online HD map detection works consider road element mapping as a semantic segmentation task in bird's-eye view (BEV) [22,30,31], in which road elements are predicted in raster format (i.e., pixel-level semantic segmentation mask). However, rasterized road maps cannot be effectively utilized by downstream tasks of autonomous vehicles, such as motion planning and vehicle control [23,26,27]. Moreover, it is challenging to distinguish instances from the rasterized map, especially when some road elements overlap with each other or when a complicated topology is encountered (e.g., road split or road intersections). + +To alleviate these problems, HDMap- + +Net [20] proposes hand-crafted post-processing algorithms to better obtain road element instances for vectorized HD maps. However, HDMapNet still heavily relies on rasterized results, restricting it from handling complicated urban scenes. Recently, some works have resorted to predicting vectorized HD maps directly [3,24,27,35]. A two-stage hierarchical set prediction method is proposed in VectorMapNet [27]. After predicting key points of road elements in the HD map, VectorMapNet sequentially generates intermediate points. Although VectorMapNet is considered to be the first online vectorized HD map detection work, the sequential operation degrades its efficiency and model performance. MapTR [24] further proposes to use DETR [5, 46] for vectorized HD map detection as a point set prediction problem, and the output point sets are then + +![](images/7eb90e49242fce19ab9f25d5d0a95a1d9ee8ba290014cdd3d56266bab7472bd0.jpg) +Fig.1: Comparison of vectorized HD map detection methods. All solutions are evaluated on the NuScenes validation set. The x-axis represents the mAP results and the y-axis displays topology level correctness [13]. InsMapper is flexible and adaptable, so it can be seamlessly modified to align with multiple existing frameworks. InsMapper is built on MapTR [24]. $\dagger$ means InsMapper based on more recent framework PivotNet [7], while $\ddagger$ indicates InsMapper based on previous SOTA, MapTR-V2 [25]. + +grouped into road element instances. Some recent works further improve the performance of MapTR [7, 25, 32]. Among them, MapTR-V2 achieves state-of-the-art performance, but it overlooks and fails to utilize the essential inner-instance correlations between points to further boost the performance. + +In this work, we focus on investigating the utilization of inner-instance information for vectorized high-definition mapping via transformer, and propose a new framework named InsMapper, which can utilize inner-instance point information for improved online HD map detection via three exquisite designs. First, a hybrid query generation module is introduced to better initialize queries for detection. Then, an inner-instance query fusion module is added before the transformer decoder, which fuses inner-instance object queries to refine the initialized line instances. Finally, a transformer-based inner-instance aggregation module is incorporated to further refine the predicted lines. Extensive experimental evaluations are conducted on the popular and challenging NuScenes [2] and Argoverse 2 [37] datasets. InsMapper is based on MapTR [24], and it can be seamlessly modified based on the latest frameworks with more enhanced performance, such as PivotNet [7] and MapTR-V2 [25]. No matter what the base framework is, InsMapper harvests state-of-the-art performance, surpassing the previous highest solutions. The intuitive performance comparisons are illustrated in Figure 1, and our main contributions can be summarized as follows: + +- We analyze the limitations of existing HD map detection algorithms and propose a new framework named InsMapper, which can effectively utilize the easily overlooked inner-instance point correlation information for accurate and generalizable online HD map detection. +- We incorporate three useful modules to leverage inner-instance information, including hybrid query generation, inner-instance query fusion, and inner-instance feature aggregation. The first two modules better initialize queries for detection and the last one refines detected line instances. +- We conduct experiments on the challenging NuScenes and Argoverse 2 datasets. InsMapper outperforms all baselines by a large margin. The achieved state-of-the-art results demonstrate its effectiveness and generality. + +# 2 Related Work + +Vector map detection. Vector maps use vectors to represent road elements, such as road networks, road curbs, and lane lines. A high-definition map (HD map) is a type of vector map, but a vector map may not qualify as an HD map if it lacks sufficient resolution or fails to provide lane-level information, among other factors. Online HD map detection can benefit from insights gained from a broader scope of vector map detection works, including road-network detection [1,14,39, 41], road-curb detection in aerial images [42, 43], and road lane line detection [15, 16], etc. DagMapper [16] aims to detect vectorized lane lines from pre-built point cloud maps using iterations. However, DagMapper only handles simple topology changes of lane lines and struggles to handle complex road intersections. RNGDet++ [41] applies DETR along with an imitation learning algorithm [34] for road-network detection, achieving state-of-the-art performance on multiple datasets. Although these methods achieve satisfactory graph detection accuracy, they suffer from poor efficiency due to their iterative algorithms. Given the + +stringent real-time requirements of autonomous vehicles, they are not suitable for online HD map detection. + +HD map detection. HD map detection was initially a subtask of BEV detection [21, 22, 28, 30, 31]. Recently, given the importance of HD maps, several works have directly focused on HD map detection [4, 20, 24, 27, 29, 33, 35, 38, 40]. However, most of these works either involve offline HD map creation [12, 38, 40] or only detect one specific road element [3, 4, 33]. HDMapNet [20] is considered to be the first work specifically designed for multiple road element detection. However, HDMapNet outputs rasterized results, requiring complicated handcrafted post-processing to obtain vectorized HD maps. To address this issue, VectorMapNet [27] is believed to be the first work detecting vectorized HD maps in real-time. However, it consists of two stages, and its efficiency is significantly impacted by sequential operations. In contrast, MapTR [24] uses deformable DETR [46] to design an end-to-end vectorized HD map detection model, which greatly simplifies the pipeline and delivers better detection performance. Some recent works further improve the performance of MapTR [7, 25, 32]. Among them, MapTR-V2 achieves state-of-the-art performance on vectorized HD map detection, but it still overlooks and does not investigate the inner-instance correlations between points, limiting further improvements. + +Detection by Transformer. DETR [5] is the first end-to-end transformer-based object detection framework. Compared with previous CNN-based methods [8, 10], DETR removes the need for anchor proposals and non-maximum suppression (NMS), making it simpler and more effective. To address the issue of slow convergence, subsequent works propose accelerating DETR training through deformable attention [46], denoising [18], and dynamic query generation [19, 36, 45]. Intuitively, most of these DETR refinements could be adapted to the HD map detection task since our proposed InsMapper relies on DETR. However, unlike typical object detection tasks in which objects are assumed to be independent and identically distributed (i.i.d.), detected points in HD maps have strong correlations, especially among inner-instance points. This inherent difference makes some refined DETR methods unsuitable to our task and additional investigations are required. + +# 3 Point Correlation + +# 3.1 Vector Map Decomposition and Sampling + +Let $G$ denote the original vector map label of the scene, which consists of vertices $V$ and edges $E$ . The vector map contains multiple classes of road elements, including pedestrian crossings, road dividers, road boundaries, and lane centerlines. Among them, the first three classes of road elements are simple polylines or polygons without intersection points. While lane centerlines have more complicated topology, such as lane split, lane merge, and lane intersection. To unify all vector elements, the vector map is decomposed into simpler shapes (i.e., polylines and polygons) without intersections. Any vertices in the vector map with degrees larger than two (i.e., intersection vertices) are removed from $G$ , but + +incident edges are kept. In this way, a set of simple polylines and polygons without intersections is obtained as $G^{*} = \{l_{i}\}_{i=1}^{N^{*}}$ , where $G^{*}$ is an undirected graph. Each shape $l_{i}$ is defined as an instance, and $N^{*}$ denotes the overall number of instances in a vector map. + +Following MapTR and MapTR-V2, to enhance the parallelization capacity of the model, each instance is evenly re-sampled with fixed-length points as $l_{i} = (v_{1},\dots,v_{j},\dots,v_{n_{p}})$ . $l_{i}$ is ordered from $v_{1}$ to $v_{n_p}$ , with $n_p$ being the number of sampled points for each instance. For polygon instances, $v_{1}$ is equal to $v_{n_p}$ . Please refer to the supplementary document for the visualization of the pre-processing. + +# 3.2 Inner-instance Correlation + +Unlike conventional object detection tasks, where objects can be approximated as independent and identically distributed (i.i.d.), strong correlations exist between points within the same line instance in the vectorized HD map detection task. Please refer to the supplementary document for the visualization of the point correlation. + +Inner-instance correlation is crucial for point coordinates prediction. Points within the same instance can collaborate by sharing inner-instance information, leading to smoother and more accurate predictions. Without this collaboration, points of the same instance may produce independent predictions, leading to zig-zag or incorrect instances. The inner-instance correlation can serve as additional constraints to facilitate the query initialization and refinement of predicted lines. In previous methods, the correlation of points is not correctly analyzed and leveraged, limiting further improvement. In subsequent sections, we introduce InsMapper to more effectively utilize point correlations and improve vectorized map detection performance. + +# 4 Methodology + +# 4.1 Overall Framework + +Building on MapTR [24], our proposed InsMapper is an end-to-end trainable framework for online vectorized HD map detection, as illustrated in Figure 2. + +InsMapper is adaptive and can be seamlessly modified based on the latest frameworks with more enhanced performance, such as PivotNet [7] and MapTR-V2 [25]. Following BEVformer [22], InsMapper first projects perspective view camera images into the bird's-eye-view (BEV). After obtaining the BEV features, InsMapper uses deformable attention layers [46] to process input object queries. Each object query predicts one point, including the object class and the regression point position. To better leverage inner-instance information and further enhance the final detection performance, we introduce three designs in InsMapper: hybrid query generation, which replaces the hierarchical query in MapTR, generating line queries with better diversity; inner-instance query fusion, which fuses generated queries within the same instance based on + +![](images/153f5c94707310daf2c3f6a5c59eb2405a1a7953aa5e3a05b673c0fd415fe27b.jpg) +Fig. 2: Overall framework. InsMapper is an end-to-end transformer model with an encoder-decoder structure. The transformer encoder projects perspective-view camera images into a bird's-eye view (BEV). Subsequently, the transformer decoder detects vector instances by predicting point sets. To enhance the utilization of inner-instance information, we introduce the following three components: a hybrid query generation scheme (orange module), an inner-instance query fusion module (yellow module), and an inner-instance feature aggregation module (blue module). The first two modules better initialize queries for detection and the final one refines detected line instances. + +inner-instance features to further enhance the quality of initialized line queries; and inner-instance feature aggregation, which adds a new inner-instance self-attention module to decoder layers to further aggregate inner-instance features for line refinement. + +Different from past works, which treat queries as independently distributed, InsMapper utilizes inner-instance information to fuse object queries and aggregate intermediate features in deformable transformer decoder layers. Otherwise, adjacent vertices may produce independent position regression results, leading to zigzag or even incorrect instances. With the assistance of inner-instance information, the model can better initialize queries for detection and refine the positions of detected points within an instance, significantly enhancing the final performance, as demonstrated in the experiments section. + +# 4.2 Query Generation + +In contrast to traditional object detection tasks, where queries are independent and identically distributed (i.i.d.), the points to be detected in our task exhibit significant correlations. Various methods can be used to generate object queries as input for the transformer decoder, such as the basic scheme, hierar + +![](images/2b3026b0a16709faa46790e0f959aea1b377cf19f46caa56809a437cfd9980f5.jpg) +Fig.3: Query generation schemes. For concise visualization, the number of instances $N_{I}$ is 2, and the number of points per instance $n_p$ is 3. The hybrid scheme can initialize lines with both good diversity and quality. + +![](images/d97ac786feed7392a30d050bc18f5bde601ec124e187d05874aa7cb4d95f00c7.jpg) +(a) Basic + +![](images/6f3776c816987d234da648677839b80c2689b13c85b31dc2f64d3b1f91713c44.jpg) +(b) Hierarchical + +![](images/a87f30e0e3d703d02d5300a33ffccd98957e6509fe176317d56eef30638cf4dd.jpg) +(c) Hybrid +Fig. 4: Visualization of initialized lines with different query generation schemes. The quality and diversity of initialized lines are crucial for the final performance. Each color represents one initialized line. + +chical scheme, and hybrid scheme. All query generation schemes are illustrated in Figure 3. Initialized queries are organized into line instances. The quality and diversity of the initialized lines are both crucial. + +Basic query generation. This method is leveraged in the original DETR [5] for query generation. It assumes that points are i.i.d. and generates queries randomly, without utilizing inner-instance information. Let $N_{I}$ denote the number of predicted instances, which is obviously larger than $N^{*}$ . Since each instance contains $n_p$ points, the basic object queries consist of $N_{I} \cdot n_{p}$ randomly generated i.i.d. queries, resulting in irregular line shapes. Due to the lack of constraints on inner-instance points, the initialized line instances are noisy and have poor quality, which leads to degraded performance. MapTR with the basic query scheme is visualized in Figure 4a. + +Hierarchical query generation. To address the aforementioned issue of the basic scheme, a hierarchical query generation scheme is proposed in MapTR [24] and MapTR-V2 [25]. Let $Q_{ins} = \{q_i^I\}_{i=1}^{N_I}$ denote instance queries and $Q_{pts} = \{q_j^P\}_{j=1}^{n_p}$ denote point queries. The object queries for HD map detection are then the pairwise addition of $Q_{ins}$ and $Q_{pts}$ : + +$$ +Q = \left\{q _ {i, j} = q _ {i} ^ {I} + q _ {j} ^ {P} \mid q _ {i} ^ {I} \in Q _ {\text {i n s}}, q _ {j} ^ {P} \in Q _ {\text {p t s}} \right\}, \tag {1} +$$ + +where $|Q| = |Q_{ins}| \cdot |Q_{pts}| = N_I \cdot n_p$ . For each line instance $i$ , the instance query $q_i^I$ denotes its uniqueness. By using $q_i^I$ as a bridge, object queries within the same instance can better collaborate with each other for point prediction. But all line instances share the same point queries so all initialized lines tend to have quite similar shapes, degrading initialization diversity, which harms the final detection performance. The initialized lines generated by MapTR with the hierarchical query scheme are visualized in Figure 4b. + +Hybrid query generation. The basic scheme does not take into account any information exchange between queries, whereas the hierarchical one generates + +repetitive line shapes. To address these issues, this paper presents a hybrid query generation method that mitigates the drawbacks of the aforementioned schemes while maintaining appropriate inner-instance information exchange. + +Let the instance queries be $Q_{ins} = \{q_i^I\}_{i=1}^{N_I}$ , and the point queries be $Q_{pts} = \{q_j^P\}_{j=1}^{N_P} = \{q_j^P\}_{j=1}^{N_I \cdot n_p}$ . The point queries are divided into $N_I$ instance groups, and a point query $q_j^P$ is assigned to the $\lceil \frac{j}{n_p} \rceil$ -th instance. Consequently, the final object query is the sum of a point query $p_j^P$ and its assigned instance query $p_k^I$ , where $k = \lceil \frac{j}{n_p} \rceil$ . The object query set can be expressed as: + +$$ +Q = \left\{q _ {k} ^ {I} + q _ {j} ^ {P} \mid q _ {j} ^ {P} \in Q _ {p t s} \right\} \tag {2} +$$ + +where $|Q| = |Q_{pts}| = N_P = N_I \cdot n_p$ . In contrast to the hierarchical scheme, each point query in the hybrid scheme is utilized only once to generate the object query, thereby preventing repetitive line shapes. Simultaneously, point queries belonging to the same instance are summed with a shared instance query, establishing the inner-instance connection. The hybrid query generation method can be considered as a combination of the basic and hierarchical schemes, effectively mitigating their respective drawbacks. InsMapper with the hybrid query generation scheme is visualized in Figure 4c. + +# 4.3 Inner-instance Query Fusion + +The input object queries for the transformer decoder are generated from instance queries and point queries. Although the generated queries can leverage some inner-instance information, this information exchange is indirect and inflexible. The instance query is distributed equally among all inner-instance points, while a more precise point-to-point information exchange cannot be realized. As a result, a query fusion module is introduced to further utilize inner-instance information and refine initialized lines. + +Let $Q_{i} = \{q_{i,j}\}_{j=1}^{n_{p}}$ represent the set of object queries belonging to the $i$ -th instance, and $q_{i,j}$ denote the $j$ -th point of the $i$ -th instance. $q_{i,j}$ is correlated with all other queries in $Q_{i}$ . To better fuse inner-instance object queries, each query is updated by a weighted sum of all queries in $Q_{i}$ as: + +$$ +q _ {i, j} = f \left(q _ {i, j}, Q _ {i}\right) = \sum_ {k = 1} ^ {n _ {p}} w _ {i, j, k} \phi \left(q _ {i, k}\right) \tag {3} +$$ + +where $w_{i,j,k}$ demonstrates weights for query fusion. $f(\cdot)$ is the fusion function and $\phi(\cdot)$ is the kernel in case nonlinear transformation is needed. $f(\cdot)$ could be realized by handcraft weights, fully connected layers, or self-attention. In this work, self-attention is applied as $f(\cdot)$ . + +In conventional object detection tasks, object queries are assumed to be independent and identically distributed (i.i.d.), making query fusion unnecessary. However, in the task of HD map detection, the query fusion module effectively aligns the update of queries and enables each point to pay more attention to + +![](images/69c49678df6c7cb8472dd3c690821fde218551d962898247891c2624b3127a37.jpg) +(a) No fusion + +![](images/2e3e38d7b3ab1a38ad9f0535c2ec7191e14301154bec10b5aeda39e1526c2965.jpg) +(b) With fusion + +![](images/4d4a9fcfdb1c5a3058a2349065c04ff2c1c0f86f45216eb70bddfbebe100d8c9.jpg) +Fig.5: Visualization of the effect of the inner-instance query fusion module. (a) No query fusion. (b) With query fusion. The query fusion module can better refine the initialized lines based on inner-instance information. In this example, the proposed query fusion module significantly enhances the diversity of initialized line instances. +Fig.6: Decoder of InsMapper. An inner-instance multi-head self-attention module is incorporated into decoder layers. In this module, the attention between points belonging to different instances is blocked (grey grids). Only inner-instance attention is allowed (colored grids). Colored grids are randomly blocked with $\epsilon$ probability (set to grey) for robustness. The proposed module can further utilize inner-instance information and refine the predicted lines. + +neighboring points. It can greatly refine the initialized lines for better query initialization, leading to enhanced performance. Without this module, queries within the same instance cannot be aware of each other, rendering them "blind". Poor line initialization tends to have degraded detection scores. The effect of the query fusion module is visualized in Figure 5. + +# 4.4 Inner-instance Feature Aggregation + +In addition to object query manipulation, InsMapper performs inner-instance feature aggregation for predicted line refinement in the transformer decoder layers by incorporating an additional inner-instance self-attention module. Inspired by [9], inner-instance attentions are randomly blocked with probability $\epsilon$ for robustness. The decoder layer of InsMapper is depicted in Figure 6. + +The inner-instance self-attention module resembles the original self-attention module but features a specially designed attention mask. As illustrated in Figure 6, the attention mask of inner-instance points is set to zero (colored grids), indicating that attention between points within the same instance is allowed. Conversely, for points belonging to different instances (i.e., inter-instance points), the corresponding attention mask values are set to one, blocking attention between them. This method encourages the model to focus more on inner-instance + +information, resulting in more consistent predictions. To further enhance the robustness of this module, the inner-instance attention (colored grids) has an $e$ probability of being blocked. + +An alternative method involves placing the proposed inner-instance attention module before the cross-attention module. However, the self-attention module before cross-attention should not block inter-instance information. Otherwise, some instances may produce duplicate predictions since they cannot "clearly see" other instances, which is analyzed in the original DETR [5]. Consequently, implementing inner-instance self attention before the cross-attention module leads to degraded final performance. Thus the proposed inner-instance feature aggregation should be placed after the cross-attention layer. More experiments about the decoder structure are provided in the appendix. + +# 5 Experiments + +# 5.1 Experimental Setup + +Datasets. Extensive experiments are conducted on the popular and challenging NuScenes dataset [2] and the Argoverse 2 dataset [37]. These two datasets comprise hundreds of data sequences captured in various autonomous driving scenarios, encompassing diverse weather conditions and time periods. The perception range for the X-axis and Y-axis in the BEV is set to $[-15m, 15m]$ and $[-30m, 30m]$ , respectively. The detected vectorized HD map should encompass four types of road elements simultaneously: road boundary, lane split, pedestrian crossing, and lane centerline. Among these, the lane centerline is a crucial element for ensuring effective vehicle planning and control. Detecting lane centerlines poses greater difficulty compared to the other three elements. + +Implementation details. We perform all the experiments on a machine equipped with 8 RTX-3090 GPUs. During the training phase, all GPUs are utilized, whereas only a single GPU is employed for inference. For our proposed InsMapper, we adopt multiple settings akin to the previous SOTA method, MapTR. All ablation studies are conducted for 24 epochs, with ResNet50 [11] as the backbone. For a fair comparison, except for the proposed modules, we keep the experiment settings exactly the same, such as batch size, network depth, the number of input queries, etc. + +Evaluation metrics. Models are assessed using two types of metrics. The first metric, average precision (AP), gauges the instance-level detection performance, employing Chamfer distance for matching predictions with ground truth labels. To ensure a fair comparison, we follow previous works to calculate multiple $\mathrm{AP}_{\tau}$ values with $\tau \in \{0.5, 1.0, 1.5\}$ , and report the average AP. To measure the topology correctness, the metric TOPO that is widely used in past works [12-14, 41] is reported. For all metrics, a larger value indicates better performance. + +# 5.2 Performance Comparison + +In this section, we compare InsMapper with previous state-of-the-art methods using the aforementioned evaluation metrics. The quantitative comparison results + +Table 1: Quantitative results of comparison experiments on the NuScenes validation set. Colored numbers show differences between InsMapper and the SOTA baseline under the same experiment setting. "C" and "L" denote camera and LiDAR, respectively. "-" denotes that the result is not available. "V2-99" and "Sec" correspond to VoVNetV2-99 [17] and SECOND [44]. $\dagger$ represents InsMapper based on PivotNet [7], $\ddagger$ indicates InsMapper based on MapTR-V2 [25]. + +
MethodsEpochsBackboneModality\(AP_{ped}\)\(AP_{div}\)\(AP_{bound}\)\(AP_{center}\)mAPTOPO
HDMapNet30Effi-B0C4.4123.7358.1737.2930.9029.79
STSU110R50C---31.2131.2132.11
VectorMapNet130R50C28.6639.7433.0634.9334.1039.90
MapTR24R50C37.9246.2550.0737.4742.9346.77
MapTR24V2-99C44.5756.2560.5846.3051.9255.16
MapTR24R50&SecC&L49.4758.9866.7244.8955.0156.77
MapTR110R50C49.1659.1258.9347.2653.6259.16
BeMapNet24R50C43.5954.1151.9038.6047.0549.90
PivotNet24R50C42.7353.5552.0142.4747.6951.04
MapTR-V224R50C48.2054.7556.6645.8651.3751.15
MapTR-V2110R50C58.6865.7267.1256.1661.9263.26
InsMapper24R50C44.3653.3652.7742.3548.3151.58
InsMapper24V2-99C51.1663.7164.4751.4057.6860.00
InsMapper24R50&SecC&L56.0063.4271.6152.8560.9762.51
InsMapper110R50C55.4263.8765.8054.2059.4066.19
InsMapper†24R50C46.4554.9154.1645.2050.1853.57
InsMapper‡24R50C49.0757.9859.8447.9953.7253.00
InsMapper†110R50C62.0967.6068.1558.4164.0664.72
+ +Table 2: Quantitative results of comparison experiments on the Argoverse 2 validation set. $\ddagger$ indicates InsMapper based on MapTR-V2 [25]. + +
MethodsEpochsBackboneModalityAPpedAPdivAPboundAPcentermAPTOPO
MapTR6R50C52.8863.6861.1859.7359.3775.79
MapTR-V26R50C57.1667.9665.2563.2063.3977.94
InsMapper6R50C55.6166.6062.5862.6761.8777.58
InsMapper†6R50C58.9171.3566.9064.7065.4679.26
+ +Table 3: Quantitative results of comparison experiments on the NuScenes validation set without centerline. $\ddagger$ indicates InsMapper based on MapTR-V2 [25]. + +
MethodsEpochsBackboneModalityAPpedAPdivAPboundmAP
MapTR24R50C46.0451.5853.0850.23
MapTR-V224R50C59.8062.4062.4061.50
InsMapper24R50C48.4454.6856.9253.35
InsMapper‡24R50C61.5465.0964.6263.75
+ +are reported in Table 1. InsMapper is flexible and adaptive, making it seamless to modify it based on existing transformer frameworks. InsMapper is built on MapTR for experiments, we also report the results of InsMapper with different base frameworks, e.g. InsMapper† is based on PivotNet [7] and InsMapper‡ relies on the previous SOTA method MapTR-V2 [25]. Compared to the previous methods, InsMapper improves the AP of all road elements by around 2. The topological metric also sees an improvement of around 1.5. The inference FPS of + +Table 4: Quantitative results of comparison experiments on the NuScenes validation set. Centerlines are directed graphs in this table. $\ddagger$ indicates InsMapper based on MapTR-V2 [25]. + +
MethodsEpochsBackboneModality\(AP_{ped}\)\(AP_{div}\)\(AP_{bound}\)\(AP_{center}\)mAPTOPO
MapTR24R50C37.3946.6150.6937.2242.9846.98
MapTR-V224R50C49.5355.4658.4853.8254.3254.90
InsMapper24R50C43.2453.7753.8943.2248.5352.09
\(InsMapper^{\dagger}\)24R50C51.6257.8860.9056.5256.7356.01
+ +MapTR and InsMapper are 8.5 and 7.5, respectively. The evaluation outcomes on the Argoverse 2 dataset [37] can be found in Table 2. InsMapper outperforms past works with notable enhancements. Table 3 presents the results without centerlines. InsMapper with different bases attains an improvement of more than $2\mathrm{mAP}$ . In the aforementioned experiments, all road elements are treated as undirected graphs. Table 4 shows the results when centerlines are represented as directed graphs. Based on the results, InsMapper still presents superior results than past works for directed HD map detection. Furthermore, we illustrate the qualitative visualizations in Figure 7. Owing to the effective employment of inner-instance information, InsMapper generates smoother and more precise HD maps compared to previous works. + +Consequently, InsMapper demonstrates superior performance compared to prior approaches across multiple datasets, exhibiting outstanding detection outcomes. Concurrently, InsMapper maintains a comparable inference speed to the previous state-of-the-art method. These exceptional experimental results strongly demonstrate the effectiveness and generality of the proposed designs. + +# 5.3 Ablation Studies + +Key component designs. We conduct several ablation study experiments to confirm the necessity of the proposed modules. Initially, we incorporate these modules into the previous state-of-the-art baseline model incrementally, and the results are presented in Table 5. It is evident that all three modules contribute to performance improvements, thereby validating their necessity. To further investigate the impact of each module's design, we perform comprehensive ablation studies in the following paragraphs. + +Hybrid query generation. Compared to basic or hierarchical query generation schemes, the proposed hybrid query generation scheme generates queries with better quality and diversity, due to effectively utilizing inner-instance correlation. The evaluation results of the different query generation schemes are presented in Table 6. These results reveal that the hybrid query generation scheme outperforms its counterparts, thereby demonstrating the soundness of this design. + +Inner-instance query fusion. Query fusion is proposed to leverage the correlation between inner-instance queries to enhance prediction performance. Various methods can be employed for fusing queries, including mean fusing (i.e., + +![](images/9e81c143736942d0d6a4bb29a6b7d0ab832ae314d6a85d767761e8c8af2a39bd.jpg) +Fig. 7: Qualitative visualization. The predicted map contains four classes, i.e., road boundaries (green), lane splits (red), pedestrian crossing (blue), and lane centerlines (pink). InsMapper presents better detection results compared with the past work. + +![](images/503858f18474dde558d326dc82abed08358b72a575badea6db98b137a2268064.jpg) + +each query is summed with the mean of all queries of the same instance), fusion by the feed-forward network, and fusion by self-attention. The evaluation results are presented in Table 7. The outcomes of the query fusion methods vary significantly, with self-attention-based query fusion achieving the best results. Consequently, self-attention query fusion is incorporated into InsMapper. + +Inner-instance feature aggregation. In InsMapper, the inner-instance self-attention module is incorporated into the decoder layers following the cross-attention module. We evaluate InsMapper under four different conditions: without the inner-instance self-attention module (no attention), replace the inner-instance self-attention module with a vanilla self-attention module (vanilla attention), without the random blocking (no random blocking, i.e., $\epsilon = 0$ ), and with the module placed before the cross-attention (change position). The results are shown in Table 8. The outcomes reveal that removing or altering the proposed + +Table 5: Ablation studies on key component designs. "QG", "QF" and "FA" represent hybrid query generation, inner-instance query fusion, and inner-instance feature aggregation, respectively. + +
QGQFFAmAPTOPO
42.9346.77
✓✓45.42(↑2.49)49.19(↑2.42)
✓✓45.45(↑2.52)48.54(↑1.77)
✓✓46.93(↑4.00)50.16(↑3.39)
✓✓✓48.31(↑5.38)51.58(↑4.81)
+ +Table 6: Ablation studies on inner-instance query generation. + +
Query SchememAPTOPO
InsMapper-basic46.6249.39
InsMapper-hierarchical46.9350.16
InsMapper-hybrid48.3151.58
+ +Table 7: Ablation studies on inner-instance query fusion. + +
Query FusionmAPTOPO
No fusion45.4548.54
Mean43.3446.00
Feed-Forward47.8050.93
Self-attention48.3151.58
+ +Table 8: Ablation studies on inner-instance feature aggregation. Multiple designs of the incorporated inner-instance self-attention layers are examined. + +
MethodmAPTOPO
No attention45.4249.19
Vanilla attention46.2349.51
No random blocking47.6751.49
Change position45.2448.16
InsMapper48.3151.58
+ +module can impair model performance. This observation confirms the effectiveness of the proposed module design. + +# 6 Conclusion + +In this paper, we introduced InsMapper, an end-to-end transformer-based model for on-the-fly vectorized HD map detection. InsMapper surpasses previous works by effectively leveraging inner-instance information to improve detection outcomes. Three exquisite designs have been proposed to utilize inner-instance information, including hybrid query generation, inner-instance query fusion, and inner-instance feature aggregation. The first two modules improve the quality of initialized line instances for detection, while the last module effectively refines the detected lines. The superiority of InsMapper is well exhibited through experiments on the challenging NuScenes and Argoverse 2 datasets. The evaluation results demonstrate the effectiveness and generality of InsMapper, making it a promising solution for the vectorized HD map detection task. In the future, InsMapper will be further enhanced for temporal-consistent vector map detection, making it more suitable for real-world applications. + +Acknowledgement. This work is supported by the National Natural Science Foundation of China (No. 62201484), HKU Startup Fund, and HKU Seed Fund for Basic Research. + +# References + +1. Bastani, F., He, S., Abbar, S., Alizadeh, M., Balakrishnan, H., Chawla, S., Madden, S., DeWitt, D.: Roadtracer: Automatic extraction of road networks from aerial images. In: CVPR (2018) 3 +2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020) 3, 10 +3. Can, Y.B., Liniger, A., Paudel, D.P., Van Gool, L.: Structured bird's-eye-view traffic scene understanding from onboard images. In: ICCV (2021) 2, 4 +4. Can, Y.B., Liniger, A., Paudel, D.P., Van Gool, L.: Topology preserving local road network estimation from single onboard camera image. In: CVPR (2022) 4 +5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020) 2, 4, 7, 10 +6. Da, F., Zhang, Y.: Path-aware graph attention for hd maps in motion prediction. In: ICRA (2022) 1 +7. Ding, W., Qiao, L., Qiu, X., Zhang, C.: Pivotnet: Vectorized pivot learning for end-to-end hd map construction. In: ICCV (2023) 2, 3, 4, 5, 11 +8. Girshick, R.: Fast r-cnn. In: ICCV (2015) 4 +9. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022) 9 +0. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. In: ICCV (2017) 4 +1. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) 10 +2. He, S., Balakrishnan, H.: Lane-level street map extraction from aerial imagery. In: WACV (2022) 4, 10 +3. He, S., Bastani, F., Abbar, S., Alizadeh, M., Balakrishnan, H., Chawla, S., Madden, S.: Roadrunner: improving the precision of road network inference fromgps trajectories. In: SIGSPATIAL (2018) 2, 10 +4. He, S., Bastani, F., Jagwani, S., Alizadeh, M., Balakrishnan, H., Chawla, S., Elshrif, M.M., Madden, S., Sadeghi, M.A.: Sat2graph: road graph extraction through graph-tensor encoding. In: ECCV (2020) 3, 10 +5. Homayounfar, N., Ma, W.C., Kowshika Lakshmikanth, S., Urtasun, R.: Hierarchical recurrent attention networks for structured online maps. In: CVPR (2018) 3 +6. Homayounfar, N., Ma, W.C., Liang, J., Wu, X., Fan, J., Urtasun, R.: Dagmapper: Learning to map by discovering lane topology. In: ICCV (2019) 3 +7. Lee, Y., Park, J.: Centermask: Real-time anchor-free instance segmentation. In: CVPR (2020) 11 +8. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: CVPR (2022) 4 +9. Li, F., Zhang, H., Xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y.: Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In: CVPR (2023) 4 +20. Li, Q., Wang, Y., Wang, Y., Zhao, H.: Hdmapnet: A local semantic map learning and evaluation framework. In: ICRA (2022) 2, 4 +21. Li, S., Yang, K., Shi, H., Zhang, J., Lin, J., Teng, Z., Li, Z.: Bi-mapper: Holistic bev semantic mapping for autonomous driving. RAL (2023) 4 +22. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: ECCV (2022) 2, 4, 5 + +23. Liang, M., Yang, B., Hu, R., Chen, Y., Liao, R., Feng, S., Urtasun, R.: Learning lane graph representations for motion forecasting. In: ECCV (2020) 1, 2 +24. Liao, B., Chen, S., Wang, X., Cheng, T., Zhang, Q., Liu, W., Huang, C.: Maptr: Structured modeling and learning for online vectorized hd map construction. In: ICLR (2023) 2, 3, 4, 5, 7 +25. Liao, B., Chen, S., Zhang, Y., Jiang, B., Zhang, Q., Liu, W., Huang, C., Wang, X.: Maptrv2: An end-to-end framework for online vectorized hd map construction. arXiv:2308.05736 (2023) 2, 3, 4, 5, 7, 11, 12 +26. Liu, T., hai Liao, Q., Gan, L., Ma, F., Cheng, J., Xie, X., Wang, Z., Chen, Y., Zhu, Y., Zhang, S., et al.: The role of the hercules autonomous vehicle during the COVID-19 pandemic: An autonomous logistic vehicle for contactless goods transportation. IEEE Robotics & Automation Magazine (2021) 2 +27. Liu, Y., Wang, Y., Wang, Y., Zhao, H.: Vectormapnet: End-to-end vectorized hd map learning. In: ICML (2023) 1, 2, 4 +28. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: ICRA (2023) 4 +29. Mi, L., Zhao, H., Nash, C., Jin, X., Gao, J., Sun, C., Schmid, C., Shavit, N., Chai, Y., Anguelov, D.: Hdmapgen: A hierarchical graph generative model of high definition maps. In: CVPR (2021) 4 +30. Ng, M.H., Radia, K., Chen, J., Wang, D., Gog, I., Gonzalez, J.E.: Bev-seg: Bird's eye view semantic segmentation using geometry and semantic point cloud. In: CVPRW (2020) 2, 4 +31. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: ECCV (2020) 2, 4 +32. Qiao, L., Ding, W., Qiu, X., Zhang, C.: End-to-end vectorized hd-map construction with piecewiseBezier curve. In: CVPR (2023) 2, 4 +33. Qiao, Z., Yu, Z., Yin, H., Shen, S.: Online monocular lane mapping using catmullrom spline. In: IROS (2023) 4 +34. Ross, S., Gordon, G., Bagnell, D.: A reduction of imitation learning and structured prediction to no-regret online learning. In: AISTATS (2011) 3 +35. Shin, J., Rameau, F., Jeong, H., Kum, D.: Instagram: Instance-level graph modeling for vectorized hd map learning. In: CVPRW (2023) 2, 4 +36. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: AAAI (2022) 4 +37. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. In: NeurIPS (2021) 3, 10, 12 +38. Xie, Z., Pang, Z., Wang, Y.: Mv-map: Offboard hd-map generation with multi-view consistency. In: ICCV (2023) 2, 4 +39. Xu, Z., Liu, Y., Gan, L., Sun, Y., Liu, M., Wang, L.: Ringdet: Road network graph detection by transformer in aerial images. TGRS (2022) 3 +40. Xu, Z., Liu, Y., Sun, Y., Liu, M., Wang, L.: Centerlinedet: Road lane centerline graph detection with vehicle-mounted sensors by transformer for high-definition map creation. In: ICRA (2023) 2, 4 +41. Xu, Z., Liu, Y., Sun, Y., Liu, M., Wang, L.: Ringdet++: Road network graph detection by transformer with instance segmentation and multi-scale features enhancement. RAL (2023) 2, 3, 10 +42. Xu, Z., Sun, Y., Liu, M.: imputation learning-based detection of road curbs using aerial images for autonomous driving. RAL (2021) 3 + +43. Xu, Z., Sun, Y., Liu, M.: Topo-boundary: A benchmark dataset on topological road-boundary detection using aerial images for autonomous driving. RAL (2021) 3 +44. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors (2018) 11 +45. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In: ICLR (2023) 4 +46. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: ICLR (2021) 2, 4, 5 \ No newline at end of file diff --git a/2024/InsMapper_ Exploring Inner-instance Information for Vectorized HD Mapping/images.zip b/2024/InsMapper_ Exploring Inner-instance Information for Vectorized HD Mapping/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c0fc37d788be6499a1c2c86ffcd205dfe59a9206 --- /dev/null +++ b/2024/InsMapper_ Exploring Inner-instance Information for Vectorized HD Mapping/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b5146f0d32ec22048ac2ac977592fc828295369481d234aef3bb5319e58cf19 +size 614333 diff --git a/2024/InsMapper_ Exploring Inner-instance Information for Vectorized HD Mapping/layout.json b/2024/InsMapper_ Exploring Inner-instance Information for Vectorized HD Mapping/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b7f36a7e2835f2b8d0106b9597bf8379fa884c71 --- /dev/null +++ b/2024/InsMapper_ Exploring Inner-instance Information for Vectorized HD Mapping/layout.json @@ -0,0 +1,8448 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "text", + "content": "InsMapper: Exploring Inner-instance Information for Vectorized HD Mapping" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 170, + 168, + 444, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 168, + 444, + 180 + ], + "spans": [ + { + "bbox": [ + 170, + 168, + 444, + 180 + ], + "type": "text", + "content": "Zhenhua Xu Kwan-Yee K. Wong Hengshuang Zhao\\*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 246, + 190, + 369, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 190, + 369, + 201 + ], + "spans": [ + { + "bbox": [ + 246, + 190, + 369, + 201 + ], + "type": "text", + "content": "The University of Hong Kong" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 169, + 221, + 444, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 221, + 444, + 234 + ], + "spans": [ + { + "bbox": [ + 169, + 221, + 444, + 234 + ], + "type": "text", + "content": "Project webpage: https://tonyxuqq.github.io/InsMapper" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 252, + 455, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 252, + 455, + 472 + ], + "spans": [ + { + "bbox": [ + 160, + 252, + 455, + 472 + ], + "type": "text", + "content": "Abstract. Vectorized high-definition (HD) maps contain detailed information about surrounding road elements, which are crucial for various downstream tasks in modern autonomous vehicles, such as motion planning and vehicle control. Recent works attempt to directly detect the vectorized HD map as a point set prediction task, achieving notable detection performance improvements. However, these methods usually overlook and fail to analyze the important inner-instance correlations between predicted points, impeding further advancements. To address this issue, we investigate the utilization of inner-instance information for vectorized high-definition mapping through transformers, and propose a powerful system named InsMapper, which effectively harnesses inner-instance information with three exquisite designs, including hybrid query generation, inner-instance query fusion, and inner-instance feature aggregation. The first two modules can better initialize queries for line detection, while the last one refines predicted line instances. InsMapper is highly adaptable and can be seamlessly modified to align with the most recent HD map detection frameworks. Extensive experimental evaluations are conducted on the challenging NuScenes and Argovverse 2 datasets, where InsMapper surpasses the previous state-of-the-art method, demonstrating its effectiveness and generality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 483, + 392, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 483, + 392, + 494 + ], + "spans": [ + { + "bbox": [ + 160, + 483, + 392, + 494 + ], + "type": "text", + "content": "Keywords: Autonomous Driving " + }, + { + "bbox": [ + 160, + 483, + 392, + 494 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 483, + 392, + 494 + ], + "type": "text", + "content": " High-definition Map" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 539, + 482, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 482, + 622 + ], + "type": "text", + "content": "Vectorized high-definition maps (HD maps) play a critical role in today's autonomous vehicles [6, 23, 27], as they contain detailed information about the road, including the position of road elements (e.g., road boundaries, lane splits, pedestrian crossings, and lane centerlines), connectivity, and topology of the road. Without the assistance of HD maps for perceiving and understanding road elements, unexpected vehicle behaviors may be encountered, such as incorrect path planning results or even vehicle collisions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 623, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 623, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 623, + 482, + 647 + ], + "type": "text", + "content": "Typically, HD maps are created by offline human annotation, which is labor-intensive, inefficient, and expensive. Although there are works proposing to make" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "such an offline process automatic [38, 40, 41], it is not possible to recreate and update the HD map frequently when the road has been modified, such as when a new road is built or an existing road is removed. To address this issue, several recent studies propose to detect local HD maps using an online method [7, 20, 24, 25, 27, 32]. This paper aims to further investigate online HD map detection techniques based on vehicle-mounted sensors (i.e., cameras and LiDARs)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 199, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 199, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 199, + 482, + 295 + ], + "type": "text", + "content": "Early online HD map detection works consider road element mapping as a semantic segmentation task in bird's-eye view (BEV) [22,30,31], in which road elements are predicted in raster format (i.e., pixel-level semantic segmentation mask). However, rasterized road maps cannot be effectively utilized by downstream tasks of autonomous vehicles, such as motion planning and vehicle control [23,26,27]. Moreover, it is challenging to distinguish instances from the rasterized map, especially when some road elements overlap with each other or when a complicated topology is encountered (e.g., road split or road intersections)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 295, + 294, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 294, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 294, + 307 + ], + "type": "text", + "content": "To alleviate these problems, HDMap-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 307, + 282, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 282, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 282, + 617 + ], + "type": "text", + "content": "Net [20] proposes hand-crafted post-processing algorithms to better obtain road element instances for vectorized HD maps. However, HDMapNet still heavily relies on rasterized results, restricting it from handling complicated urban scenes. Recently, some works have resorted to predicting vectorized HD maps directly [3,24,27,35]. A two-stage hierarchical set prediction method is proposed in VectorMapNet [27]. After predicting key points of road elements in the HD map, VectorMapNet sequentially generates intermediate points. Although VectorMapNet is considered to be the first online vectorized HD map detection work, the sequential operation degrades its efficiency and model performance. MapTR [24] further proposes to use DETR [5, 46] for vectorized HD map detection as a point set prediction problem, and the output point sets are then" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 288, + 319, + 481, + 464 + ], + "blocks": [ + { + "bbox": [ + 288, + 319, + 481, + 464 + ], + "lines": [ + { + "bbox": [ + 288, + 319, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 288, + 319, + 481, + 464 + ], + "type": "image", + "image_path": "7eb90e49242fce19ab9f25d5d0a95a1d9ee8ba290014cdd3d56266bab7472bd0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 286, + 474, + 482, + 596 + ], + "lines": [ + { + "bbox": [ + 286, + 474, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 286, + 474, + 482, + 596 + ], + "type": "text", + "content": "Fig.1: Comparison of vectorized HD map detection methods. All solutions are evaluated on the NuScenes validation set. The x-axis represents the mAP results and the y-axis displays topology level correctness [13]. InsMapper is flexible and adaptable, so it can be seamlessly modified to align with multiple existing frameworks. InsMapper is built on MapTR [24]. " + }, + { + "bbox": [ + 286, + 474, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 286, + 474, + 482, + 596 + ], + "type": "text", + "content": " means InsMapper based on more recent framework PivotNet [7], while " + }, + { + "bbox": [ + 286, + 474, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 286, + 474, + 482, + 596 + ], + "type": "text", + "content": " indicates InsMapper based on previous SOTA, MapTR-V2 [25]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "grouped into road element instances. Some recent works further improve the performance of MapTR [7, 25, 32]. Among them, MapTR-V2 achieves state-of-the-art performance, but it overlooks and fails to utilize the essential inner-instance correlations between points to further boost the performance." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Z. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": "In this work, we focus on investigating the utilization of inner-instance information for vectorized high-definition mapping via transformer, and propose a new framework named InsMapper, which can utilize inner-instance point information for improved online HD map detection via three exquisite designs. First, a hybrid query generation module is introduced to better initialize queries for detection. Then, an inner-instance query fusion module is added before the transformer decoder, which fuses inner-instance object queries to refine the initialized line instances. Finally, a transformer-based inner-instance aggregation module is incorporated to further refine the predicted lines. Extensive experimental evaluations are conducted on the popular and challenging NuScenes [2] and Argoverse 2 [37] datasets. InsMapper is based on MapTR [24], and it can be seamlessly modified based on the latest frameworks with more enhanced performance, such as PivotNet [7] and MapTR-V2 [25]. No matter what the base framework is, InsMapper harvests state-of-the-art performance, surpassing the previous highest solutions. The intuitive performance comparisons are illustrated in Figure 1, and our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 319, + 490, + 451 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 319, + 481, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 319, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 138, + 319, + 481, + 366 + ], + "type": "text", + "content": "- We analyze the limitations of existing HD map detection algorithms and propose a new framework named InsMapper, which can effectively utilize the easily overlooked inner-instance point correlation information for accurate and generalizable online HD map detection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 367, + 481, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 367, + 481, + 414 + ], + "spans": [ + { + "bbox": [ + 138, + 367, + 481, + 414 + ], + "type": "text", + "content": "- We incorporate three useful modules to leverage inner-instance information, including hybrid query generation, inner-instance query fusion, and inner-instance feature aggregation. The first two modules better initialize queries for detection and the last one refines detected line instances." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 415, + 490, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 490, + 451 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 490, + 451 + ], + "type": "text", + "content": "- We conduct experiments on the challenging NuScenes and Argoverse 2 datasets. InsMapper outperforms all baselines by a large margin. The achieved state-of-the-art results demonstrate its effectiveness and generality." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 471, + 237, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 237, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 237, + 483 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Vector map detection. Vector maps use vectors to represent road elements, such as road networks, road curbs, and lane lines. A high-definition map (HD map) is a type of vector map, but a vector map may not qualify as an HD map if it lacks sufficient resolution or fails to provide lane-level information, among other factors. Online HD map detection can benefit from insights gained from a broader scope of vector map detection works, including road-network detection [1,14,39, 41], road-curb detection in aerial images [42, 43], and road lane line detection [15, 16], etc. DagMapper [16] aims to detect vectorized lane lines from pre-built point cloud maps using iterations. However, DagMapper only handles simple topology changes of lane lines and struggles to handle complex road intersections. RNGDet++ [41] applies DETR along with an imitation learning algorithm [34] for road-network detection, achieving state-of-the-art performance on multiple datasets. Although these methods achieve satisfactory graph detection accuracy, they suffer from poor efficiency due to their iterative algorithms. Given the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "text", + "content": "InsMapper" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "stringent real-time requirements of autonomous vehicles, they are not suitable for online HD map detection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 342 + ], + "type": "text", + "content": "HD map detection. HD map detection was initially a subtask of BEV detection [21, 22, 28, 30, 31]. Recently, given the importance of HD maps, several works have directly focused on HD map detection [4, 20, 24, 27, 29, 33, 35, 38, 40]. However, most of these works either involve offline HD map creation [12, 38, 40] or only detect one specific road element [3, 4, 33]. HDMapNet [20] is considered to be the first work specifically designed for multiple road element detection. However, HDMapNet outputs rasterized results, requiring complicated handcrafted post-processing to obtain vectorized HD maps. To address this issue, VectorMapNet [27] is believed to be the first work detecting vectorized HD maps in real-time. However, it consists of two stages, and its efficiency is significantly impacted by sequential operations. In contrast, MapTR [24] uses deformable DETR [46] to design an end-to-end vectorized HD map detection model, which greatly simplifies the pipeline and delivers better detection performance. Some recent works further improve the performance of MapTR [7, 25, 32]. Among them, MapTR-V2 achieves state-of-the-art performance on vectorized HD map detection, but it still overlooks and does not investigate the inner-instance correlations between points, limiting further improvements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 343, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 499 + ], + "type": "text", + "content": "Detection by Transformer. DETR [5] is the first end-to-end transformer-based object detection framework. Compared with previous CNN-based methods [8, 10], DETR removes the need for anchor proposals and non-maximum suppression (NMS), making it simpler and more effective. To address the issue of slow convergence, subsequent works propose accelerating DETR training through deformable attention [46], denoising [18], and dynamic query generation [19, 36, 45]. Intuitively, most of these DETR refinements could be adapted to the HD map detection task since our proposed InsMapper relies on DETR. However, unlike typical object detection tasks in which objects are assumed to be independent and identically distributed (i.i.d.), detected points in HD maps have strong correlations, especially among inner-instance points. This inherent difference makes some refined DETR methods unsuitable to our task and additional investigations are required." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 516, + 259, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 516, + 259, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 516, + 259, + 528 + ], + "type": "text", + "content": "3 Point Correlation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 539, + 371, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 539, + 371, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 539, + 371, + 552 + ], + "type": "text", + "content": "3.1 Vector Map Decomposition and Sampling" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": " denote the original vector map label of the scene, which consists of vertices " + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": " and edges " + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": ". The vector map contains multiple classes of road elements, including pedestrian crossings, road dividers, road boundaries, and lane centerlines. Among them, the first three classes of road elements are simple polylines or polygons without intersection points. While lane centerlines have more complicated topology, such as lane split, lane merge, and lane intersection. To unify all vector elements, the vector map is decomposed into simpler shapes (i.e., polylines and polygons) without intersections. Any vertices in the vector map with degrees larger than two (i.e., intersection vertices) are removed from " + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": ", but" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "Z. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "incident edges are kept. In this way, a set of simple polylines and polygons without intersections is obtained as " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "G^{*} = \\{l_{i}\\}_{i=1}^{N^{*}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "G^{*}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " is an undirected graph. Each shape " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "l_{i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " is defined as an instance, and " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "N^{*}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " denotes the overall number of instances in a vector map." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "text", + "content": "Following MapTR and MapTR-V2, to enhance the parallelization capacity of the model, each instance is evenly re-sampled with fixed-length points as " + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "inline_equation", + "content": "l_{i} = (v_{1},\\dots,v_{j},\\dots,v_{n_{p}})" + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "inline_equation", + "content": "l_{i}" + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "text", + "content": " is ordered from " + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "inline_equation", + "content": "v_{1}" + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "inline_equation", + "content": "v_{n_p}" + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "inline_equation", + "content": "n_p" + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "text", + "content": " being the number of sampled points for each instance. For polygon instances, " + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "inline_equation", + "content": "v_{1}" + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "text", + "content": " is equal to " + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "inline_equation", + "content": "v_{n_p}" + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "text", + "content": ". Please refer to the supplementary document for the visualization of the pre-processing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 243, + 293, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 243, + 293, + 254 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 293, + 254 + ], + "type": "text", + "content": "3.2 Inner-instance Correlation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 263, + 479, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 479, + 323 + ], + "type": "text", + "content": "Unlike conventional object detection tasks, where objects can be approximated as independent and identically distributed (i.i.d.), strong correlations exist between points within the same line instance in the vectorized HD map detection task. Please refer to the supplementary document for the visualization of the point correlation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 324, + 480, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 480, + 443 + ], + "type": "text", + "content": "Inner-instance correlation is crucial for point coordinates prediction. Points within the same instance can collaborate by sharing inner-instance information, leading to smoother and more accurate predictions. Without this collaboration, points of the same instance may produce independent predictions, leading to zig-zag or incorrect instances. The inner-instance correlation can serve as additional constraints to facilitate the query initialization and refinement of predicted lines. In previous methods, the correlation of points is not correctly analyzed and leveraged, limiting further improvement. In subsequent sections, we introduce InsMapper to more effectively utilize point correlations and improve vectorized map detection performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 463, + 233, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 463, + 233, + 477 + ], + "spans": [ + { + "bbox": [ + 131, + 463, + 233, + 477 + ], + "type": "text", + "content": "4 Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 489, + 257, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 489, + 257, + 500 + ], + "spans": [ + { + "bbox": [ + 131, + 489, + 257, + 500 + ], + "type": "text", + "content": "4.1 Overall Framework" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 509, + 479, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 479, + 533 + ], + "type": "text", + "content": "Building on MapTR [24], our proposed InsMapper is an end-to-end trainable framework for online vectorized HD map detection, as illustrated in Figure 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 534, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 480, + 665 + ], + "type": "text", + "content": "InsMapper is adaptive and can be seamlessly modified based on the latest frameworks with more enhanced performance, such as PivotNet [7] and MapTR-V2 [25]. Following BEVformer [22], InsMapper first projects perspective view camera images into the bird's-eye-view (BEV). After obtaining the BEV features, InsMapper uses deformable attention layers [46] to process input object queries. Each object query predicts one point, including the object class and the regression point position. To better leverage inner-instance information and further enhance the final detection performance, we introduce three designs in InsMapper: hybrid query generation, which replaces the hierarchical query in MapTR, generating line queries with better diversity; inner-instance query fusion, which fuses generated queries within the same instance based on" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "text", + "content": "InsMapper" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 115, + 482, + 311 + ], + "blocks": [ + { + "bbox": [ + 136, + 115, + 482, + 311 + ], + "lines": [ + { + "bbox": [ + 136, + 115, + 482, + 311 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 482, + 311 + ], + "type": "image", + "image_path": "153f5c94707310daf2c3f6a5c59eb2405a1a7953aa5e3a05b673c0fd415fe27b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 321, + 482, + 410 + ], + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 410 + ], + "type": "text", + "content": "Fig. 2: Overall framework. InsMapper is an end-to-end transformer model with an encoder-decoder structure. The transformer encoder projects perspective-view camera images into a bird's-eye view (BEV). Subsequently, the transformer decoder detects vector instances by predicting point sets. To enhance the utilization of inner-instance information, we introduce the following three components: a hybrid query generation scheme (orange module), an inner-instance query fusion module (yellow module), and an inner-instance feature aggregation module (blue module). The first two modules better initialize queries for detection and the final one refines detected line instances." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 434, + 482, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 480 + ], + "type": "text", + "content": "inner-instance features to further enhance the quality of initialized line queries; and inner-instance feature aggregation, which adds a new inner-instance self-attention module to decoder layers to further aggregate inner-instance features for line refinement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 481, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 482, + 578 + ], + "type": "text", + "content": "Different from past works, which treat queries as independently distributed, InsMapper utilizes inner-instance information to fuse object queries and aggregate intermediate features in deformable transformer decoder layers. Otherwise, adjacent vertices may produce independent position regression results, leading to zigzag or even incorrect instances. With the assistance of inner-instance information, the model can better initialize queries for detection and refine the positions of detected points within an instance, significantly enhancing the final performance, as demonstrated in the experiments section." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 596, + 253, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 596, + 253, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 596, + 253, + 609 + ], + "type": "text", + "content": "4.2 Query Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In contrast to traditional object detection tasks, where queries are independent and identically distributed (i.i.d.), the points to be detected in our task exhibit significant correlations. Various methods can be used to generate object queries as input for the transformer decoder, such as the basic scheme, hierar" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Z. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 138, + 297, + 223 + ], + "blocks": [ + { + "bbox": [ + 133, + 138, + 297, + 223 + ], + "lines": [ + { + "bbox": [ + 133, + 138, + 297, + 223 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 297, + 223 + ], + "type": "image", + "image_path": "2b3026b0a16709faa46790e0f959aea1b377cf19f46caa56809a437cfd9980f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 232, + 298, + 299 + ], + "lines": [ + { + "bbox": [ + 131, + 232, + 298, + 299 + ], + "spans": [ + { + "bbox": [ + 131, + 232, + 298, + 299 + ], + "type": "text", + "content": "Fig.3: Query generation schemes. For concise visualization, the number of instances " + }, + { + "bbox": [ + 131, + 232, + 298, + 299 + ], + "type": "inline_equation", + "content": "N_{I}" + }, + { + "bbox": [ + 131, + 232, + 298, + 299 + ], + "type": "text", + "content": " is 2, and the number of points per instance " + }, + { + "bbox": [ + 131, + 232, + 298, + 299 + ], + "type": "inline_equation", + "content": "n_p" + }, + { + "bbox": [ + 131, + 232, + 298, + 299 + ], + "type": "text", + "content": " is 3. The hybrid scheme can initialize lines with both good diversity and quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 119, + 362, + 219 + ], + "blocks": [ + { + "bbox": [ + 310, + 119, + 362, + 219 + ], + "lines": [ + { + "bbox": [ + 310, + 119, + 362, + 219 + ], + "spans": [ + { + "bbox": [ + 310, + 119, + 362, + 219 + ], + "type": "image", + "image_path": "d97ac786feed7392a30d050bc18f5bde601ec124e187d05874aa7cb4d95f00c7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 224, + 354, + 234 + ], + "lines": [ + { + "bbox": [ + 319, + 224, + 354, + 234 + ], + "spans": [ + { + "bbox": [ + 319, + 224, + 354, + 234 + ], + "type": "text", + "content": "(a) Basic" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 369, + 118, + 423, + 219 + ], + "blocks": [ + { + "bbox": [ + 369, + 118, + 423, + 219 + ], + "lines": [ + { + "bbox": [ + 369, + 118, + 423, + 219 + ], + "spans": [ + { + "bbox": [ + 369, + 118, + 423, + 219 + ], + "type": "image", + "image_path": "6f3776c816987d234da648677839b80c2689b13c85b31dc2f64d3b1f91713c44.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 224, + 426, + 234 + ], + "lines": [ + { + "bbox": [ + 369, + 224, + 426, + 234 + ], + "spans": [ + { + "bbox": [ + 369, + 224, + 426, + 234 + ], + "type": "text", + "content": "(b) Hierarchical" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 433, + 118, + 482, + 220 + ], + "blocks": [ + { + "bbox": [ + 433, + 118, + 482, + 220 + ], + "lines": [ + { + "bbox": [ + 433, + 118, + 482, + 220 + ], + "spans": [ + { + "bbox": [ + 433, + 118, + 482, + 220 + ], + "type": "image", + "image_path": "a87f30e0e3d703d02d5300a33ffccd98957e6509fe176317d56eef30638cf4dd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 224, + 477, + 234 + ], + "lines": [ + { + "bbox": [ + 438, + 224, + 477, + 234 + ], + "spans": [ + { + "bbox": [ + 438, + 224, + 477, + 234 + ], + "type": "text", + "content": "(c) Hybrid" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 243, + 489, + 298 + ], + "lines": [ + { + "bbox": [ + 305, + 243, + 489, + 298 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 489, + 298 + ], + "type": "text", + "content": "Fig. 4: Visualization of initialized lines with different query generation schemes. The quality and diversity of initialized lines are crucial for the final performance. Each color represents one initialized line." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 322, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 480, + 357 + ], + "type": "text", + "content": "chical scheme, and hybrid scheme. All query generation schemes are illustrated in Figure 3. Initialized queries are organized into line instances. The quality and diversity of the initialized lines are both crucial." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 358, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 465 + ], + "type": "text", + "content": "Basic query generation. This method is leveraged in the original DETR [5] for query generation. It assumes that points are i.i.d. and generates queries randomly, without utilizing inner-instance information. Let " + }, + { + "bbox": [ + 130, + 358, + 482, + 465 + ], + "type": "inline_equation", + "content": "N_{I}" + }, + { + "bbox": [ + 130, + 358, + 482, + 465 + ], + "type": "text", + "content": " denote the number of predicted instances, which is obviously larger than " + }, + { + "bbox": [ + 130, + 358, + 482, + 465 + ], + "type": "inline_equation", + "content": "N^{*}" + }, + { + "bbox": [ + 130, + 358, + 482, + 465 + ], + "type": "text", + "content": ". Since each instance contains " + }, + { + "bbox": [ + 130, + 358, + 482, + 465 + ], + "type": "inline_equation", + "content": "n_p" + }, + { + "bbox": [ + 130, + 358, + 482, + 465 + ], + "type": "text", + "content": " points, the basic object queries consist of " + }, + { + "bbox": [ + 130, + 358, + 482, + 465 + ], + "type": "inline_equation", + "content": "N_{I} \\cdot n_{p}" + }, + { + "bbox": [ + 130, + 358, + 482, + 465 + ], + "type": "text", + "content": " randomly generated i.i.d. queries, resulting in irregular line shapes. Due to the lack of constraints on inner-instance points, the initialized line instances are noisy and have poor quality, which leads to degraded performance. MapTR with the basic query scheme is visualized in Figure 4a." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 466, + 482, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 466, + 482, + 527 + ], + "spans": [ + { + "bbox": [ + 131, + 466, + 482, + 527 + ], + "type": "text", + "content": "Hierarchical query generation. To address the aforementioned issue of the basic scheme, a hierarchical query generation scheme is proposed in MapTR [24] and MapTR-V2 [25]. Let " + }, + { + "bbox": [ + 131, + 466, + 482, + 527 + ], + "type": "inline_equation", + "content": "Q_{ins} = \\{q_i^I\\}_{i=1}^{N_I}" + }, + { + "bbox": [ + 131, + 466, + 482, + 527 + ], + "type": "text", + "content": " denote instance queries and " + }, + { + "bbox": [ + 131, + 466, + 482, + 527 + ], + "type": "inline_equation", + "content": "Q_{pts} = \\{q_j^P\\}_{j=1}^{n_p}" + }, + { + "bbox": [ + 131, + 466, + 482, + 527 + ], + "type": "text", + "content": " denote point queries. The object queries for HD map detection are then the pairwise addition of " + }, + { + "bbox": [ + 131, + 466, + 482, + 527 + ], + "type": "inline_equation", + "content": "Q_{ins}" + }, + { + "bbox": [ + 131, + 466, + 482, + 527 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 466, + 482, + 527 + ], + "type": "inline_equation", + "content": "Q_{pts}" + }, + { + "bbox": [ + 131, + 466, + 482, + 527 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 211, + 533, + 481, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 533, + 481, + 550 + ], + "spans": [ + { + "bbox": [ + 211, + 533, + 481, + 550 + ], + "type": "interline_equation", + "content": "Q = \\left\\{q _ {i, j} = q _ {i} ^ {I} + q _ {j} ^ {P} \\mid q _ {i} ^ {I} \\in Q _ {\\text {i n s}}, q _ {j} ^ {P} \\in Q _ {\\text {p t s}} \\right\\}, \\tag {1}", + "image_path": "091e493ec51e87c50750b1866fa5fb06a2bce551fa5cc3ae89c0404a3bbc1035.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "|Q| = |Q_{ins}| \\cdot |Q_{pts}| = N_I \\cdot n_p" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": ". For each line instance " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": ", the instance query " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "q_i^I" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": " denotes its uniqueness. By using " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "q_i^I" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": " as a bridge, object queries within the same instance can better collaborate with each other for point prediction. But all line instances share the same point queries so all initialized lines tend to have quite similar shapes, degrading initialization diversity, which harms the final detection performance. The initialized lines generated by MapTR with the hierarchical query scheme are visualized in Figure 4b." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "content": "Hybrid query generation. The basic scheme does not take into account any information exchange between queries, whereas the hierarchical one generates" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "text", + "content": "InsMapper" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "repetitive line shapes. To address these issues, this paper presents a hybrid query generation method that mitigates the drawbacks of the aforementioned schemes while maintaining appropriate inner-instance information exchange." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": "Let the instance queries be " + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "inline_equation", + "content": "Q_{ins} = \\{q_i^I\\}_{i=1}^{N_I}" + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": ", and the point queries be " + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "inline_equation", + "content": "Q_{pts} = \\{q_j^P\\}_{j=1}^{N_P} = \\{q_j^P\\}_{j=1}^{N_I \\cdot n_p}" + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": ". The point queries are divided into " + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "inline_equation", + "content": "N_I" + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": " instance groups, and a point query " + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "inline_equation", + "content": "q_j^P" + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": " is assigned to the " + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "inline_equation", + "content": "\\lceil \\frac{j}{n_p} \\rceil" + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": "-th instance. Consequently, the final object query is the sum of a point query " + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "inline_equation", + "content": "p_j^P" + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": " and its assigned instance query " + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "inline_equation", + "content": "p_k^I" + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "inline_equation", + "content": "k = \\lceil \\frac{j}{n_p} \\rceil" + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": ". The object query set can be expressed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 252, + 233, + 481, + 248 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 233, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 252, + 233, + 481, + 248 + ], + "type": "interline_equation", + "content": "Q = \\left\\{q _ {k} ^ {I} + q _ {j} ^ {P} \\mid q _ {j} ^ {P} \\in Q _ {p t s} \\right\\} \\tag {2}", + "image_path": "808b376cf9c1ebbf2d2b3e5fc45bcbaa7c3c1f54408af255acfc7cf7ea294e65.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 251, + 480, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 480, + 348 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 251, + 480, + 348 + ], + "type": "inline_equation", + "content": "|Q| = |Q_{pts}| = N_P = N_I \\cdot n_p" + }, + { + "bbox": [ + 130, + 251, + 480, + 348 + ], + "type": "text", + "content": ". In contrast to the hierarchical scheme, each point query in the hybrid scheme is utilized only once to generate the object query, thereby preventing repetitive line shapes. Simultaneously, point queries belonging to the same instance are summed with a shared instance query, establishing the inner-instance connection. The hybrid query generation method can be considered as a combination of the basic and hierarchical schemes, effectively mitigating their respective drawbacks. InsMapper with the hybrid query generation scheme is visualized in Figure 4c." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 366, + 304, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 366, + 304, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 366, + 304, + 379 + ], + "type": "text", + "content": "4.3 Inner-instance Query Fusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 387, + 480, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 480, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 480, + 470 + ], + "type": "text", + "content": "The input object queries for the transformer decoder are generated from instance queries and point queries. Although the generated queries can leverage some inner-instance information, this information exchange is indirect and inflexible. The instance query is distributed equally among all inner-instance points, while a more precise point-to-point information exchange cannot be realized. As a result, a query fusion module is introduced to further utilize inner-instance information and refine initialized lines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 471, + 480, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 480, + 520 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 471, + 480, + 520 + ], + "type": "inline_equation", + "content": "Q_{i} = \\{q_{i,j}\\}_{j=1}^{n_{p}}" + }, + { + "bbox": [ + 130, + 471, + 480, + 520 + ], + "type": "text", + "content": " represent the set of object queries belonging to the " + }, + { + "bbox": [ + 130, + 471, + 480, + 520 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 471, + 480, + 520 + ], + "type": "text", + "content": "-th instance, and " + }, + { + "bbox": [ + 130, + 471, + 480, + 520 + ], + "type": "inline_equation", + "content": "q_{i,j}" + }, + { + "bbox": [ + 130, + 471, + 480, + 520 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 130, + 471, + 480, + 520 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 471, + 480, + 520 + ], + "type": "text", + "content": "-th point of the " + }, + { + "bbox": [ + 130, + 471, + 480, + 520 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 471, + 480, + 520 + ], + "type": "text", + "content": "-th instance. " + }, + { + "bbox": [ + 130, + 471, + 480, + 520 + ], + "type": "inline_equation", + "content": "q_{i,j}" + }, + { + "bbox": [ + 130, + 471, + 480, + 520 + ], + "type": "text", + "content": " is correlated with all other queries in " + }, + { + "bbox": [ + 130, + 471, + 480, + 520 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 130, + 471, + 480, + 520 + ], + "type": "text", + "content": ". To better fuse inner-instance object queries, each query is updated by a weighted sum of all queries in " + }, + { + "bbox": [ + 130, + 471, + 480, + 520 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 130, + 471, + 480, + 520 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 231, + 529, + 481, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 529, + 481, + 561 + ], + "spans": [ + { + "bbox": [ + 231, + 529, + 481, + 561 + ], + "type": "interline_equation", + "content": "q _ {i, j} = f \\left(q _ {i, j}, Q _ {i}\\right) = \\sum_ {k = 1} ^ {n _ {p}} w _ {i, j, k} \\phi \\left(q _ {i, k}\\right) \\tag {3}", + "image_path": "ddccfe38e52b79889b5d59d630999b3c9e8a1a4b1f3df9e842916e7b638b6359.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 479, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 570, + 479, + 617 + ], + "type": "inline_equation", + "content": "w_{i,j,k}" + }, + { + "bbox": [ + 130, + 570, + 479, + 617 + ], + "type": "text", + "content": " demonstrates weights for query fusion. " + }, + { + "bbox": [ + 130, + 570, + 479, + 617 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 570, + 479, + 617 + ], + "type": "text", + "content": " is the fusion function and " + }, + { + "bbox": [ + 130, + 570, + 479, + 617 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 130, + 570, + 479, + 617 + ], + "type": "text", + "content": " is the kernel in case nonlinear transformation is needed. " + }, + { + "bbox": [ + 130, + 570, + 479, + 617 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 570, + 479, + 617 + ], + "type": "text", + "content": " could be realized by handcraft weights, fully connected layers, or self-attention. In this work, self-attention is applied as " + }, + { + "bbox": [ + 130, + 570, + 479, + 617 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 570, + 479, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 618, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 618, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 618, + 480, + 666 + ], + "type": "text", + "content": "In conventional object detection tasks, object queries are assumed to be independent and identically distributed (i.i.d.), making query fusion unnecessary. However, in the task of HD map detection, the query fusion module effectively aligns the update of queries and enables each point to pay more attention to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Z. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 133, + 202, + 258 + ], + "blocks": [ + { + "bbox": [ + 138, + 133, + 202, + 258 + ], + "lines": [ + { + "bbox": [ + 138, + 133, + 202, + 258 + ], + "spans": [ + { + "bbox": [ + 138, + 133, + 202, + 258 + ], + "type": "image", + "image_path": "69c49678df6c7cb8472dd3c690821fde218551d962898247891c2624b3127a37.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 266, + 195, + 276 + ], + "lines": [ + { + "bbox": [ + 146, + 266, + 195, + 276 + ], + "spans": [ + { + "bbox": [ + 146, + 266, + 195, + 276 + ], + "type": "text", + "content": "(a) No fusion" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 216, + 135, + 276, + 255 + ], + "blocks": [ + { + "bbox": [ + 216, + 135, + 276, + 255 + ], + "lines": [ + { + "bbox": [ + 216, + 135, + 276, + 255 + ], + "spans": [ + { + "bbox": [ + 216, + 135, + 276, + 255 + ], + "type": "image", + "image_path": "2e3e38d7b3ab1a38ad9f0535c2ec7191e14301154bec10b5aeda39e1526c2965.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 218, + 266, + 275, + 276 + ], + "lines": [ + { + "bbox": [ + 218, + 266, + 275, + 276 + ], + "spans": [ + { + "bbox": [ + 218, + 266, + 275, + 276 + ], + "type": "text", + "content": "(b) With fusion" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 294, + 117, + 483, + 274 + ], + "blocks": [ + { + "bbox": [ + 294, + 117, + 483, + 274 + ], + "lines": [ + { + "bbox": [ + 294, + 117, + 483, + 274 + ], + "spans": [ + { + "bbox": [ + 294, + 117, + 483, + 274 + ], + "type": "image", + "image_path": "4d4a9fcfdb1c5a3058a2349065c04ff2c1c0f86f45216eb70bddfbebe100d8c9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 285, + 284, + 394 + ], + "lines": [ + { + "bbox": [ + 130, + 285, + 284, + 394 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 284, + 394 + ], + "type": "text", + "content": "Fig.5: Visualization of the effect of the inner-instance query fusion module. (a) No query fusion. (b) With query fusion. The query fusion module can better refine the initialized lines based on inner-instance information. In this example, the proposed query fusion module significantly enhances the diversity of initialized line instances." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 290, + 285, + 485, + 395 + ], + "lines": [ + { + "bbox": [ + 290, + 285, + 485, + 395 + ], + "spans": [ + { + "bbox": [ + 290, + 285, + 485, + 395 + ], + "type": "text", + "content": "Fig.6: Decoder of InsMapper. An inner-instance multi-head self-attention module is incorporated into decoder layers. In this module, the attention between points belonging to different instances is blocked (grey grids). Only inner-instance attention is allowed (colored grids). Colored grids are randomly blocked with " + }, + { + "bbox": [ + 290, + 285, + 485, + 395 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 290, + 285, + 485, + 395 + ], + "type": "text", + "content": " probability (set to grey) for robustness. The proposed module can further utilize inner-instance information and refine the predicted lines." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 419, + 482, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 482, + 480 + ], + "type": "text", + "content": "neighboring points. It can greatly refine the initialized lines for better query initialization, leading to enhanced performance. Without this module, queries within the same instance cannot be aware of each other, rendering them \"blind\". Poor line initialization tends to have degraded detection scores. The effect of the query fusion module is visualized in Figure 5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 499, + 339, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 499, + 339, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 499, + 339, + 511 + ], + "type": "text", + "content": "4.4 Inner-instance Feature Aggregation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 521, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 582 + ], + "type": "text", + "content": "In addition to object query manipulation, InsMapper performs inner-instance feature aggregation for predicted line refinement in the transformer decoder layers by incorporating an additional inner-instance self-attention module. Inspired by [9], inner-instance attentions are randomly blocked with probability " + }, + { + "bbox": [ + 130, + 521, + 481, + 582 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 521, + 481, + 582 + ], + "type": "text", + "content": " for robustness. The decoder layer of InsMapper is depicted in Figure 6." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "The inner-instance self-attention module resembles the original self-attention module but features a specially designed attention mask. As illustrated in Figure 6, the attention mask of inner-instance points is set to zero (colored grids), indicating that attention between points within the same instance is allowed. Conversely, for points belonging to different instances (i.e., inter-instance points), the corresponding attention mask values are set to one, blocking attention between them. This method encourages the model to focus more on inner-instance" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "text", + "content": "InsMapper" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "information, resulting in more consistent predictions. To further enhance the robustness of this module, the inner-instance attention (colored grids) has an " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " probability of being blocked." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 260 + ], + "type": "text", + "content": "An alternative method involves placing the proposed inner-instance attention module before the cross-attention module. However, the self-attention module before cross-attention should not block inter-instance information. Otherwise, some instances may produce duplicate predictions since they cannot \"clearly see\" other instances, which is analyzed in the original DETR [5]. Consequently, implementing inner-instance self attention before the cross-attention module leads to degraded final performance. Thus the proposed inner-instance feature aggregation should be placed after the cross-attention layer. More experiments about the decoder structure are provided in the appendix." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 274, + 230, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 274, + 230, + 288 + ], + "spans": [ + { + "bbox": [ + 131, + 274, + 230, + 288 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 296, + 262, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 296, + 262, + 308 + ], + "spans": [ + { + "bbox": [ + 131, + 296, + 262, + 308 + ], + "type": "text", + "content": "5.1 Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 312, + 480, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 480, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 480, + 432 + ], + "type": "text", + "content": "Datasets. Extensive experiments are conducted on the popular and challenging NuScenes dataset [2] and the Argoverse 2 dataset [37]. These two datasets comprise hundreds of data sequences captured in various autonomous driving scenarios, encompassing diverse weather conditions and time periods. The perception range for the X-axis and Y-axis in the BEV is set to " + }, + { + "bbox": [ + 130, + 312, + 480, + 432 + ], + "type": "inline_equation", + "content": "[-15m, 15m]" + }, + { + "bbox": [ + 130, + 312, + 480, + 432 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 312, + 480, + 432 + ], + "type": "inline_equation", + "content": "[-30m, 30m]" + }, + { + "bbox": [ + 130, + 312, + 480, + 432 + ], + "type": "text", + "content": ", respectively. The detected vectorized HD map should encompass four types of road elements simultaneously: road boundary, lane split, pedestrian crossing, and lane centerline. Among these, the lane centerline is a crucial element for ensuring effective vehicle planning and control. Detecting lane centerlines poses greater difficulty compared to the other three elements." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 432, + 485, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 485, + 527 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 485, + 527 + ], + "type": "text", + "content": "Implementation details. We perform all the experiments on a machine equipped with 8 RTX-3090 GPUs. During the training phase, all GPUs are utilized, whereas only a single GPU is employed for inference. For our proposed InsMapper, we adopt multiple settings akin to the previous SOTA method, MapTR. All ablation studies are conducted for 24 epochs, with ResNet50 [11] as the backbone. For a fair comparison, except for the proposed modules, we keep the experiment settings exactly the same, such as batch size, network depth, the number of input queries, etc." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 528, + 481, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 481, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 481, + 612 + ], + "type": "text", + "content": "Evaluation metrics. Models are assessed using two types of metrics. The first metric, average precision (AP), gauges the instance-level detection performance, employing Chamfer distance for matching predictions with ground truth labels. To ensure a fair comparison, we follow previous works to calculate multiple " + }, + { + "bbox": [ + 130, + 528, + 481, + 612 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\tau}" + }, + { + "bbox": [ + 130, + 528, + 481, + 612 + ], + "type": "text", + "content": " values with " + }, + { + "bbox": [ + 130, + 528, + 481, + 612 + ], + "type": "inline_equation", + "content": "\\tau \\in \\{0.5, 1.0, 1.5\\}" + }, + { + "bbox": [ + 130, + 528, + 481, + 612 + ], + "type": "text", + "content": ", and report the average AP. To measure the topology correctness, the metric TOPO that is widely used in past works [12-14, 41] is reported. For all metrics, a larger value indicates better performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 625, + 289, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 625, + 289, + 637 + ], + "spans": [ + { + "bbox": [ + 131, + 625, + 289, + 637 + ], + "type": "text", + "content": "5.2 Performance Comparison" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "In this section, we compare InsMapper with previous state-of-the-art methods using the aforementioned evaluation metrics. The quantitative comparison results" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "Z. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 190, + 481, + 363 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "text", + "content": "Table 1: Quantitative results of comparison experiments on the NuScenes validation set. Colored numbers show differences between InsMapper and the SOTA baseline under the same experiment setting. \"C\" and \"L\" denote camera and LiDAR, respectively. \"-\" denotes that the result is not available. \"V2-99\" and \"Sec\" correspond to VoVNetV2-99 [17] and SECOND [44]. " + }, + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "text", + "content": " represents InsMapper based on PivotNet [7], " + }, + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "text", + "content": " indicates InsMapper based on MapTR-V2 [25]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 190, + 481, + 363 + ], + "lines": [ + { + "bbox": [ + 132, + 190, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 190, + 481, + 363 + ], + "type": "table", + "html": "
MethodsEpochsBackboneModality\\(AP_{ped}\\)\\(AP_{div}\\)\\(AP_{bound}\\)\\(AP_{center}\\)mAPTOPO
HDMapNet30Effi-B0C4.4123.7358.1737.2930.9029.79
STSU110R50C---31.2131.2132.11
VectorMapNet130R50C28.6639.7433.0634.9334.1039.90
MapTR24R50C37.9246.2550.0737.4742.9346.77
MapTR24V2-99C44.5756.2560.5846.3051.9255.16
MapTR24R50&SecC&L49.4758.9866.7244.8955.0156.77
MapTR110R50C49.1659.1258.9347.2653.6259.16
BeMapNet24R50C43.5954.1151.9038.6047.0549.90
PivotNet24R50C42.7353.5552.0142.4747.6951.04
MapTR-V224R50C48.2054.7556.6645.8651.3751.15
MapTR-V2110R50C58.6865.7267.1256.1661.9263.26
InsMapper24R50C44.3653.3652.7742.3548.3151.58
InsMapper24V2-99C51.1663.7164.4751.4057.6860.00
InsMapper24R50&SecC&L56.0063.4271.6152.8560.9762.51
InsMapper110R50C55.4263.8765.8054.2059.4066.19
InsMapper†24R50C46.4554.9154.1645.2050.1853.57
InsMapper‡24R50C49.0757.9859.8447.9953.7253.00
InsMapper†110R50C62.0967.6068.1558.4164.0664.72
", + "image_path": "a30545cd2c236ff090daae4d43839ad4307e0255bd42faec8e4409d65afc3c0b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 132, + 406, + 481, + 460 + ], + "blocks": [ + { + "bbox": [ + 130, + 374, + 480, + 396 + ], + "lines": [ + { + "bbox": [ + 130, + 374, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 480, + 396 + ], + "type": "text", + "content": "Table 2: Quantitative results of comparison experiments on the Argoverse 2 validation set. " + }, + { + "bbox": [ + 130, + 374, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 130, + 374, + 480, + 396 + ], + "type": "text", + "content": " indicates InsMapper based on MapTR-V2 [25]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 406, + 481, + 460 + ], + "lines": [ + { + "bbox": [ + 132, + 406, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 406, + 481, + 460 + ], + "type": "table", + "html": "
MethodsEpochsBackboneModalityAPpedAPdivAPboundAPcentermAPTOPO
MapTR6R50C52.8863.6861.1859.7359.3775.79
MapTR-V26R50C57.1667.9665.2563.2063.3977.94
InsMapper6R50C55.6166.6062.5862.6761.8777.58
InsMapper†6R50C58.9171.3566.9064.7065.4679.26
", + "image_path": "ff2567bdaeeaa76a643da0e4ac71e86aea9f272153d40971038885f6761e9c3c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 132, + 504, + 481, + 559 + ], + "blocks": [ + { + "bbox": [ + 130, + 472, + 480, + 495 + ], + "lines": [ + { + "bbox": [ + 130, + 472, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 480, + 495 + ], + "type": "text", + "content": "Table 3: Quantitative results of comparison experiments on the NuScenes validation set without centerline. " + }, + { + "bbox": [ + 130, + 472, + 480, + 495 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 130, + 472, + 480, + 495 + ], + "type": "text", + "content": " indicates InsMapper based on MapTR-V2 [25]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 504, + 481, + 559 + ], + "lines": [ + { + "bbox": [ + 132, + 504, + 481, + 559 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 481, + 559 + ], + "type": "table", + "html": "
MethodsEpochsBackboneModalityAPpedAPdivAPboundmAP
MapTR24R50C46.0451.5853.0850.23
MapTR-V224R50C59.8062.4062.4061.50
InsMapper24R50C48.4454.6856.9253.35
InsMapper‡24R50C61.5465.0964.6263.75
", + "image_path": "c9cf344cd196342bce229d65f7d6e60b5be68877b7c5f8f400aff0974a95c1ab.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "are reported in Table 1. InsMapper is flexible and adaptive, making it seamless to modify it based on existing transformer frameworks. InsMapper is built on MapTR for experiments, we also report the results of InsMapper with different base frameworks, e.g. InsMapper† is based on PivotNet [7] and InsMapper‡ relies on the previous SOTA method MapTR-V2 [25]. Compared to the previous methods, InsMapper improves the AP of all road elements by around 2. The topological metric also sees an improvement of around 1.5. The inference FPS of" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "text", + "content": "InsMapper" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 157, + 482, + 213 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 4: Quantitative results of comparison experiments on the NuScenes validation set. Centerlines are directed graphs in this table. " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " indicates InsMapper based on MapTR-V2 [25]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 157, + 482, + 213 + ], + "lines": [ + { + "bbox": [ + 132, + 157, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 482, + 213 + ], + "type": "table", + "html": "
MethodsEpochsBackboneModality\\(AP_{ped}\\)\\(AP_{div}\\)\\(AP_{bound}\\)\\(AP_{center}\\)mAPTOPO
MapTR24R50C37.3946.6150.6937.2242.9846.98
MapTR-V224R50C49.5355.4658.4853.8254.3254.90
InsMapper24R50C43.2453.7753.8943.2248.5352.09
\\(InsMapper^{\\dagger}\\)24R50C51.6257.8860.9056.5256.7356.01
", + "image_path": "e4dd68f52ddfa6eba12b27b688195bd1fd82907b204c2b159d06cf9d783dc2dc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 482, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 366 + ], + "type": "text", + "content": "MapTR and InsMapper are 8.5 and 7.5, respectively. The evaluation outcomes on the Argoverse 2 dataset [37] can be found in Table 2. InsMapper outperforms past works with notable enhancements. Table 3 presents the results without centerlines. InsMapper with different bases attains an improvement of more than " + }, + { + "bbox": [ + 130, + 236, + 482, + 366 + ], + "type": "inline_equation", + "content": "2\\mathrm{mAP}" + }, + { + "bbox": [ + 130, + 236, + 482, + 366 + ], + "type": "text", + "content": ". In the aforementioned experiments, all road elements are treated as undirected graphs. Table 4 shows the results when centerlines are represented as directed graphs. Based on the results, InsMapper still presents superior results than past works for directed HD map detection. Furthermore, we illustrate the qualitative visualizations in Figure 7. Owing to the effective employment of inner-instance information, InsMapper generates smoother and more precise HD maps compared to previous works." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 368, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 429 + ], + "type": "text", + "content": "Consequently, InsMapper demonstrates superior performance compared to prior approaches across multiple datasets, exhibiting outstanding detection outcomes. Concurrently, InsMapper maintains a comparable inference speed to the previous state-of-the-art method. These exceptional experimental results strongly demonstrate the effectiveness and generality of the proposed designs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 449, + 246, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 246, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 246, + 460 + ], + "type": "text", + "content": "5.3 Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 472, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 556 + ], + "type": "text", + "content": "Key component designs. We conduct several ablation study experiments to confirm the necessity of the proposed modules. Initially, we incorporate these modules into the previous state-of-the-art baseline model incrementally, and the results are presented in Table 5. It is evident that all three modules contribute to performance improvements, thereby validating their necessity. To further investigate the impact of each module's design, we perform comprehensive ablation studies in the following paragraphs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 628 + ], + "type": "text", + "content": "Hybrid query generation. Compared to basic or hierarchical query generation schemes, the proposed hybrid query generation scheme generates queries with better quality and diversity, due to effectively utilizing inner-instance correlation. The evaluation results of the different query generation schemes are presented in Table 6. These results reveal that the hybrid query generation scheme outperforms its counterparts, thereby demonstrating the soundness of this design." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Inner-instance query fusion. Query fusion is proposed to leverage the correlation between inner-instance queries to enhance prediction performance. Various methods can be employed for fusing queries, including mean fusing (i.e.," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "Z. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 350, + 442 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 350, + 442 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 350, + 442 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 350, + 442 + ], + "type": "image", + "image_path": "9e81c143736942d0d6a4bb29a6b7d0ab832ae314d6a85d767761e8c8af2a39bd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 449, + 482, + 483 + ], + "lines": [ + { + "bbox": [ + 132, + 449, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 482, + 483 + ], + "type": "text", + "content": "Fig. 7: Qualitative visualization. The predicted map contains four classes, i.e., road boundaries (green), lane splits (red), pedestrian crossing (blue), and lane centerlines (pink). InsMapper presents better detection results compared with the past work." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 351, + 114, + 479, + 442 + ], + "blocks": [ + { + "bbox": [ + 351, + 114, + 479, + 442 + ], + "lines": [ + { + "bbox": [ + 351, + 114, + 479, + 442 + ], + "spans": [ + { + "bbox": [ + 351, + 114, + 479, + 442 + ], + "type": "image", + "image_path": "503858f18474dde558d326dc82abed08358b72a575badea6db98b137a2268064.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 568 + ], + "type": "text", + "content": "each query is summed with the mean of all queries of the same instance), fusion by the feed-forward network, and fusion by self-attention. The evaluation results are presented in Table 7. The outcomes of the query fusion methods vary significantly, with self-attention-based query fusion achieving the best results. Consequently, self-attention query fusion is incorporated into InsMapper." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Inner-instance feature aggregation. In InsMapper, the inner-instance self-attention module is incorporated into the decoder layers following the cross-attention module. We evaluate InsMapper under four different conditions: without the inner-instance self-attention module (no attention), replace the inner-instance self-attention module with a vanilla self-attention module (vanilla attention), without the random blocking (no random blocking, i.e., " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon = 0" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "), and with the module placed before the cross-attention (change position). The results are shown in Table 8. The outcomes reveal that removing or altering the proposed" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "text", + "content": "InsMapper" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 179, + 299, + 258 + ], + "blocks": [ + { + "bbox": [ + 132, + 125, + 301, + 178 + ], + "lines": [ + { + "bbox": [ + 132, + 125, + 301, + 178 + ], + "spans": [ + { + "bbox": [ + 132, + 125, + 301, + 178 + ], + "type": "text", + "content": "Table 5: Ablation studies on key component designs. \"QG\", \"QF\" and \"FA\" represent hybrid query generation, inner-instance query fusion, and inner-instance feature aggregation, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 179, + 299, + 258 + ], + "lines": [ + { + "bbox": [ + 134, + 179, + 299, + 258 + ], + "spans": [ + { + "bbox": [ + 134, + 179, + 299, + 258 + ], + "type": "table", + "html": "
QGQFFAmAPTOPO
42.9346.77
✓✓45.42(↑2.49)49.19(↑2.42)
✓✓45.45(↑2.52)48.54(↑1.77)
✓✓46.93(↑4.00)50.16(↑3.39)
✓✓✓48.31(↑5.38)51.58(↑4.81)
", + "image_path": "e76427544833258e5817c6b9646dc499e0eff9fb340dd6cda978a54ee2d6b268.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 289, + 301, + 347 + ], + "blocks": [ + { + "bbox": [ + 132, + 267, + 301, + 288 + ], + "lines": [ + { + "bbox": [ + 132, + 267, + 301, + 288 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 301, + 288 + ], + "type": "text", + "content": "Table 6: Ablation studies on inner-instance query generation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 289, + 301, + 347 + ], + "lines": [ + { + "bbox": [ + 133, + 289, + 301, + 347 + ], + "spans": [ + { + "bbox": [ + 133, + 289, + 301, + 347 + ], + "type": "table", + "html": "
Query SchememAPTOPO
InsMapper-basic46.6249.39
InsMapper-hierarchical46.9350.16
InsMapper-hybrid48.3151.58
", + "image_path": "3d6ac60e8cc309f99b6c5c8c3caaa76cf6819ecd289a029b00aa3a6401118bff.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 312, + 147, + 481, + 215 + ], + "blocks": [ + { + "bbox": [ + 311, + 125, + 481, + 146 + ], + "lines": [ + { + "bbox": [ + 311, + 125, + 481, + 146 + ], + "spans": [ + { + "bbox": [ + 311, + 125, + 481, + 146 + ], + "type": "text", + "content": "Table 7: Ablation studies on inner-instance query fusion." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 147, + 481, + 215 + ], + "lines": [ + { + "bbox": [ + 312, + 147, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 312, + 147, + 481, + 215 + ], + "type": "table", + "html": "
Query FusionmAPTOPO
No fusion45.4548.54
Mean43.3446.00
Feed-Forward47.8050.93
Self-attention48.3151.58
", + "image_path": "60467ad1dcabcde060cf7384529273568e4ffcd11675ff327319086a3e8335c4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 314, + 268, + 479, + 346 + ], + "blocks": [ + { + "bbox": [ + 311, + 223, + 481, + 266 + ], + "lines": [ + { + "bbox": [ + 311, + 223, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 311, + 223, + 481, + 266 + ], + "type": "text", + "content": "Table 8: Ablation studies on inner-instance feature aggregation. Multiple designs of the incorporated inner-instance self-attention layers are examined." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 268, + 479, + 346 + ], + "lines": [ + { + "bbox": [ + 314, + 268, + 479, + 346 + ], + "spans": [ + { + "bbox": [ + 314, + 268, + 479, + 346 + ], + "type": "table", + "html": "
MethodmAPTOPO
No attention45.4249.19
Vanilla attention46.2349.51
No random blocking47.6751.49
Change position45.2448.16
InsMapper48.3151.58
", + "image_path": "44be73fd04a7f1432fe73293e2e9318887cb5e9b550adfa2643f738220674af7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 374, + 479, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 479, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 479, + 399 + ], + "type": "text", + "content": "module can impair model performance. This observation confirms the effectiveness of the proposed module design." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 431, + 220, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 220, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 220, + 445 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 628 + ], + "type": "text", + "content": "In this paper, we introduced InsMapper, an end-to-end transformer-based model for on-the-fly vectorized HD map detection. InsMapper surpasses previous works by effectively leveraging inner-instance information to improve detection outcomes. Three exquisite designs have been proposed to utilize inner-instance information, including hybrid query generation, inner-instance query fusion, and inner-instance feature aggregation. The first two modules improve the quality of initialized line instances for detection, while the last module effectively refines the detected lines. The superiority of InsMapper is well exhibited through experiments on the challenging NuScenes and Argoverse 2 datasets. The evaluation results demonstrate the effectiveness and generality of InsMapper, making it a promising solution for the vectorized HD map detection task. In the future, InsMapper will be further enhanced for temporal-consistent vector map detection, making it more suitable for real-world applications." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Acknowledgement. This work is supported by the National Natural Science Foundation of China (No. 62201484), HKU Startup Fund, and HKU Seed Fund for Basic Research." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Z. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 138, + 137, + 481, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 481, + 170 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 481, + 170 + ], + "type": "text", + "content": "1. Bastani, F., He, S., Abbar, S., Alizadeh, M., Balakrishnan, H., Chawla, S., Madden, S., DeWitt, D.: Roadtracer: Automatic extraction of road networks from aerial images. In: CVPR (2018) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 171, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 171, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 138, + 171, + 481, + 203 + ], + "type": "text", + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020) 3, 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 203, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 203, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 203, + 481, + 224 + ], + "type": "text", + "content": "3. Can, Y.B., Liniger, A., Paudel, D.P., Van Gool, L.: Structured bird's-eye-view traffic scene understanding from onboard images. In: ICCV (2021) 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 224, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 481, + 246 + ], + "type": "text", + "content": "4. Can, Y.B., Liniger, A., Paudel, D.P., Van Gool, L.: Topology preserving local road network estimation from single onboard camera image. In: CVPR (2022) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 246, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 267 + ], + "type": "text", + "content": "5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020) 2, 4, 7, 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 267, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 481, + 289 + ], + "type": "text", + "content": "6. Da, F., Zhang, Y.: Path-aware graph attention for hd maps in motion prediction. In: ICRA (2022) 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 289, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 481, + 310 + ], + "type": "text", + "content": "7. Ding, W., Qiao, L., Qiu, X., Zhang, C.: Pivotnet: Vectorized pivot learning for end-to-end hd map construction. In: ICCV (2023) 2, 3, 4, 5, 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 310, + 327, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 327, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 327, + 320 + ], + "type": "text", + "content": "8. Girshick, R.: Fast r-cnn. In: ICCV (2015) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 320, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 481, + 342 + ], + "type": "text", + "content": "9. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022) 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 342, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 481, + 353 + ], + "type": "text", + "content": "0. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. In: ICCV (2017) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 353, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 481, + 373 + ], + "type": "text", + "content": "1. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 373, + 481, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 373, + 481, + 395 + ], + "spans": [ + { + "bbox": [ + 138, + 373, + 481, + 395 + ], + "type": "text", + "content": "2. He, S., Balakrishnan, H.: Lane-level street map extraction from aerial imagery. In: WACV (2022) 4, 10" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 396, + 481, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 481, + 428 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 481, + 428 + ], + "type": "text", + "content": "3. He, S., Bastani, F., Abbar, S., Alizadeh, M., Balakrishnan, H., Chawla, S., Madden, S.: Roadrunner: improving the precision of road network inference fromgps trajectories. In: SIGSPATIAL (2018) 2, 10" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 428, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 428, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 428, + 481, + 460 + ], + "type": "text", + "content": "4. He, S., Bastani, F., Jagwani, S., Alizadeh, M., Balakrishnan, H., Chawla, S., Elshrif, M.M., Madden, S., Sadeghi, M.A.: Sat2graph: road graph extraction through graph-tensor encoding. In: ECCV (2020) 3, 10" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 460, + 481, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 481, + 491 + ], + "type": "text", + "content": "5. Homayounfar, N., Ma, W.C., Kowshika Lakshmikanth, S., Urtasun, R.: Hierarchical recurrent attention networks for structured online maps. In: CVPR (2018) 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 492, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 492, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 492, + 481, + 514 + ], + "type": "text", + "content": "6. Homayounfar, N., Ma, W.C., Liang, J., Wu, X., Fan, J., Urtasun, R.: Dagmapper: Learning to map by discovering lane topology. In: ICCV (2019) 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 514, + 481, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 481, + 535 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 481, + 535 + ], + "type": "text", + "content": "7. Lee, Y., Park, J.: Centermask: Real-time anchor-free instance segmentation. In: CVPR (2020) 11" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 536, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 481, + 557 + ], + "type": "text", + "content": "8. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: CVPR (2022) 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 557, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 481, + 590 + ], + "type": "text", + "content": "9. Li, F., Zhang, H., Xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y.: Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In: CVPR (2023) 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 590, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 481, + 611 + ], + "type": "text", + "content": "20. Li, Q., Wang, Y., Wang, Y., Zhao, H.: Hdmapnet: A local semantic map learning and evaluation framework. In: ICRA (2022) 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "text", + "content": "21. Li, S., Yang, K., Shi, H., Zhang, J., Lin, J., Teng, Z., Li, Z.: Bi-mapper: Holistic bev semantic mapping for autonomous driving. RAL (2023) 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "22. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: ECCV (2022) 2, 4, 5" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "text", + "content": "InsMapper" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "23. Liang, M., Yang, B., Hu, R., Chen, Y., Liao, R., Feng, S., Urtasun, R.: Learning lane graph representations for motion forecasting. In: ECCV (2020) 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "text", + "content": "24. Liao, B., Chen, S., Wang, X., Cheng, T., Zhang, Q., Liu, W., Huang, C.: Maptr: Structured modeling and learning for online vectorized hd map construction. In: ICLR (2023) 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "type": "text", + "content": "25. Liao, B., Chen, S., Zhang, Y., Jiang, B., Zhang, Q., Liu, W., Huang, C., Wang, X.: Maptrv2: An end-to-end framework for online vectorized hd map construction. arXiv:2308.05736 (2023) 2, 3, 4, 5, 7, 11, 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 247 + ], + "type": "text", + "content": "26. Liu, T., hai Liao, Q., Gan, L., Ma, F., Cheng, J., Xie, X., Wang, Z., Chen, Y., Zhu, Y., Zhang, S., et al.: The role of the hercules autonomous vehicle during the COVID-19 pandemic: An autonomous logistic vehicle for contactless goods transportation. IEEE Robotics & Automation Magazine (2021) 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "27. Liu, Y., Wang, Y., Wang, Y., Zhao, H.: Vectormapnet: End-to-end vectorized hd map learning. In: ICML (2023) 1, 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "text", + "content": "28. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: ICRA (2023) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "29. Mi, L., Zhao, H., Nash, C., Jin, X., Gao, J., Sun, C., Schmid, C., Shavit, N., Chai, Y., Anguelov, D.: Hdmapgen: A hierarchical graph generative model of high definition maps. In: CVPR (2021) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 480, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 368 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 368 + ], + "type": "text", + "content": "30. Ng, M.H., Radia, K., Chen, J., Wang, D., Gog, I., Gonzalez, J.E.: Bev-seg: Bird's eye view semantic segmentation using geometry and semantic point cloud. In: CVPRW (2020) 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "31. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: ECCV (2020) 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "32. Qiao, L., Ding, W., Qiu, X., Zhang, C.: End-to-end vectorized hd-map construction with piecewiseBezier curve. In: CVPR (2023) 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 414, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 434 + ], + "type": "text", + "content": "33. Qiao, Z., Yu, Z., Yin, H., Shen, S.: Online monocular lane mapping using catmullrom spline. In: IROS (2023) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "34. Ross, S., Gordon, G., Bagnell, D.: A reduction of imitation learning and structured prediction to no-regret online learning. In: AISTATS (2011) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "text", + "content": "35. Shin, J., Rameau, F., Jeong, H., Kum, D.: Instagram: Instance-level graph modeling for vectorized hd map learning. In: CVPRW (2023) 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "36. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: AAAI (2022) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "text", + "content": "37. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. In: NeurIPS (2021) 3, 10, 12" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "text", + "content": "38. Xie, Z., Pang, Z., Wang, Y.: Mv-map: Offboard hd-map generation with multi-view consistency. In: ICCV (2023) 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 556, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 480, + 578 + ], + "type": "text", + "content": "39. Xu, Z., Liu, Y., Gan, L., Sun, Y., Liu, M., Wang, L.: Ringdet: Road network graph detection by transformer in aerial images. TGRS (2022) 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "text", + "content": "40. Xu, Z., Liu, Y., Sun, Y., Liu, M., Wang, L.: Centerlinedet: Road lane centerline graph detection with vehicle-mounted sensors by transformer for high-definition map creation. In: ICRA (2023) 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "41. Xu, Z., Liu, Y., Sun, Y., Liu, M., Wang, L.: Ringdet++: Road network graph detection by transformer with instance segmentation and multi-scale features enhancement. RAL (2023) 2, 3, 10" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "42. Xu, Z., Sun, Y., Liu, M.: imputation learning-based detection of road curbs using aerial images for autonomous driving. RAL (2021) 3" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "Z. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 227 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 148 + ], + "type": "text", + "content": "43. Xu, Z., Sun, Y., Liu, M.: Topo-boundary: A benchmark dataset on topological road-boundary detection using aerial images for autonomous driving. RAL (2021) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 172 + ], + "type": "text", + "content": "44. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors (2018) 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 172, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 482, + 205 + ], + "type": "text", + "content": "45. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In: ICLR (2023) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 205, + 482, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 205, + 482, + 227 + ], + "spans": [ + { + "bbox": [ + 131, + 205, + 482, + 227 + ], + "type": "text", + "content": "46. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: ICLR (2021) 2, 4, 5" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 102 + ], + "type": "text", + "content": "InsMapper" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Insect Identification in the Wild_ The AMI Dataset/deef9d64-c259-4177-ab17-679ea66a461f_content_list.json b/2024/Insect Identification in the Wild_ The AMI Dataset/deef9d64-c259-4177-ab17-679ea66a461f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3a44fecbcac9218fcd7bda2b72f920caf2a95794 --- /dev/null +++ b/2024/Insect Identification in the Wild_ The AMI Dataset/deef9d64-c259-4177-ab17-679ea66a461f_content_list.json @@ -0,0 +1,1707 @@ +[ + { + "type": "text", + "text": "Insect Identification in the Wild: The AMI Dataset", + "text_level": 1, + "bbox": [ + 313, + 140, + 689, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aditya Jain1*, Fagner Cunha2*, Michael James Bunsen1*, Juan Sebastián Canas1, Léonard Pasi1, Nathan Pinoy3, Flemming Helsing3, JoAnne Russo4, Marc Botham5, Michael Sabourin6, Jonathan Fréchette12, Alexandre Anctil7, Yacksecari Lopez8, Eduardo Navarro8, Filonila Perez Pimentel8, Ana Cecilia Zamora8, José Alejandro Ramirez Silva8, Jonathan Gagnon12, Tom August5, Kim Bjerge3, Alba Gomez Segura5, Marc Bélisle9, Yves Basset8, Kent P. McFarland4, David Roy5, Toke Thomas Hoye3, Maxim Larrivée10, and David Rolnick1,11", + "bbox": [ + 218, + 210, + 782, + 334 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Mila - Quebec AI Institute", + "bbox": [ + 405, + 345, + 599, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Federal University of Amazonas", + "bbox": [ + 388, + 359, + 614, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Aarhus University", + "bbox": [ + 433, + 372, + 568, + 387 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Vermont Center for Ecostudies", + "bbox": [ + 392, + 387, + 611, + 398 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "UK Centre for Ecology & Hydrology", + "bbox": [ + 372, + 400, + 630, + 415 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "McGuire Center", + "bbox": [ + 439, + 415, + 563, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "7 Centre de données sur le patrimoine naturel du Québec", + "bbox": [ + 310, + 428, + 692, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Smithsonian Tropical Research Institute", + "bbox": [ + 359, + 441, + 642, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "9 Université de Sherbrooke", + "bbox": [ + 410, + 455, + 591, + 468 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "10 Montreal Insectarium", + "bbox": [ + 419, + 469, + 583, + 483 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "11 McGill University", + "bbox": [ + 431, + 484, + 571, + 497 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{12}$ Independent Researcher", + "bbox": [ + 411, + 497, + 591, + 511 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Insects represent half of all global biodiversity, yet many of the world's insects are disappearing, with severe implications for ecosystems and agriculture. Despite this crisis, data on insect diversity and abundance remain woefully inadequate, due to the scarcity of human experts and the lack of scalable tools for monitoring. Ecologists have started to adopt camera traps to record and study insects, and have proposed computer vision algorithms as an answer for scalable data processing. However, insect monitoring in the wild poses unique challenges that have not yet been addressed within computer vision, including the combination of long-tailed data, extremely similar classes, and significant distribution shifts. We provide the first large-scale machine learning benchmarks for fine-grained insect recognition, designed to match real-world tasks faced by ecologists. Our contributions include a curated dataset of images from citizen science platforms and museums, and an expert-annotated dataset drawn from automated camera traps across multiple continents, designed to test out-of-distribution generalization under field conditions. We train and evaluate a variety of baseline algorithms and introduce a combination of data augmentation techniques that enhance generalization across geographies and hardware setups. The dataset is made publicly available1.", + "bbox": [ + 259, + 529, + 740, + 806 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Denotes equal contribution. Correspondence to: moth-ai@mila.quebec", + "bbox": [ + 218, + 811, + 702, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ https://github.com/RolnickLab/ami-dataset", + "bbox": [ + 220, + 825, + 552, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 374, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Computer vision algorithms for species identification have become important tools in studying and protecting the Earth's wildlife. Biodiversity applications have also been beneficial to computer vision, serving as a fruitful source of new benchmarks and algorithms [11,23,68,75]. However, research in computer vision for biodiversity has often focused on large animals, not reflecting the full set of challenges faced in ecology and conservation.", + "bbox": [ + 212, + 183, + 787, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are over a million known species of insects on Earth, representing nearly one half of all known organisms (by comparison, there are only 10,000 species of birds, and even fewer mammals) [51,62]. Insects play essential roles in nearly all terrestrial ecosystems, as well as providing ecosystem services such as pollination that are critical for humanity [60], yet a looming \"insect apocalyptic\" threatens both their abundance and diversity [72]. Despite the crisis, data on insect populations is shockingly sparse [27], as a result of less resources, a plethora of species, and the high level of expertise typically required to study insects. Amid this data gap, computer vision tools for automated monitoring have the potential to radically reshape the study and conservation of insects. In particular, recent breakthroughs in camera traps for insects have made it possible to automatically photograph large numbers of individuals in the wild [38,41,58], opening up the the potential for image-processing algorithms to parse this raw data.", + "bbox": [ + 212, + 275, + 787, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Species recognition from camera traps is a well-studied problem in the computer vision community [55], with common challenges including poor lighting, occlusion of individuals, camouflage, and blur [14]. However, working with insects presents distinct challenges from those faced in traditional camera traps, which are often designed for large animals. For example, while traditional camera trap images may only rarely contain a target species, almost every image from an insect camera trap includes insects – often hundreds of them. Where traditional camera traps may have a short list of species of interest, insect monitoring requires considering thousands of species at any given location, with many of them being extremely similar, and most with very little labeled data.", + "bbox": [ + 212, + 488, + 787, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While recent authors have proposed the use of computer vision tools for detecting and identifying insects in camera trap data [8, 17, 20, 32, 43, 63], such studies have involved small datasets with only a few species represented. There has to date been no work assessing the performance of algorithms in identifying large number of insect species, or addressing the challenges of generalization across in-the-wild deployments. Our work here offers the following contributions:", + "bbox": [ + 212, + 638, + 787, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We introduce the AMI (Automated Monitoring of Insects) dataset, consisting of two parts: 1) AMI-GBIF, a dataset of $\\sim 2.5\\mathrm{M}$ human-captured images curated from citizen science platforms and museum collections, 2) AMI-Traps, an expert-annotated dataset of 2,893 insect camera trap images (representing 52,948 labeled insects) collected from a global network of automated camera traps, designed to test in-the-wild performance (Fig. 1).", + "bbox": [ + 225, + 748, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "A. Jain et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/00003212e95be6e5dd272b91109bf82f0a6193888b5755b4b1ef396d96a576bc.jpg", + "image_caption": [ + "Fig. 1: Illustration of the difference between insect species identification from human-captured images and in the wild. On left are shown images of three moth species from our AMI-GBIF dataset, curated from citizen science data and museum collections. On right are shown images of the same species (along with other insects) within a photograph taken by an automated camera trap in the wild from our AMI-Traps dataset. One of the challenges of the AMI dataset is generalizing from AMI-GBIF to AMI-Traps without additional labeled data." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 787, + 276 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We frame the problem of practical insect monitoring in the field as a set of computer vision tasks, entailing very fine-grained classification from long-tailed data, with the additional challenge of generalization across multiple kinds of imagery.", + "- We test a variety of strong baseline models for the proposed tasks, including a combination of data augmentation approaches that significantly improve performance under field conditions." + ], + "bbox": [ + 225, + 411, + 785, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 551, + 387, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Biodiversity Image Datasets. Advancements in computer vision have been closely intertwined with the creation of curated datasets such as ImageNet [26, 59], Pascal VOC [30], Open Images [44], and MS-COCO [46]. In recent years, the computer vision and machine learning communities have increasingly directed their focus towards addressing pressing human threats, including biodiversity loss [68] and climate change [57]. These domains not only hold significant potential for societal impact but also offer fertile ground for exploring intriguing machine learning questions such as domain generalization [77], long-tailed distributions [71], and fine-grained classification [74]. Such challenges have found especially fruitful testbeds in the classification of species, where datasets such as iNaturalist [70] have proven useful for both the computer vision and ecology communities. While some challenges have considered plant identification [15, 45, 53, 56], the majority have focused on animals, including datasets focused on birds [16, 69, 73], mammals [9, 14, 64], and fish [22, 25, 39, 40]. Such datasets have often leveraged concurrent advancements in sensor technology, including in satellite imagery and drones. In particular, camera traps have been widely adopted by biologists as an in situ and cost-effective means of wildlife", + "bbox": [ + 212, + 582, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Insect Identification in the Wild", + "bbox": [ + 514, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c8422f237016fb6a7084b360cb3732bc8a7d1c35b2e6e6375849c53ca34a5014.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 147, + 295, + 183 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c8d7dee1236803f8056af0ee0b3845fd33dd3d277603922e0f9d64c8b7e2393e.jpg", + "image_caption": [ + "(a) Incomplete" + ], + "image_footnote": [], + "bbox": [ + 227, + 200, + 290, + 241 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ff4a585ec332620364206eefd14b9081a9fe36e330ff7b5f7ca3df710832ddc0.jpg", + "image_caption": [ + "(b) Overlapping", + "(c) Motion" + ], + "image_footnote": [], + "bbox": [ + 225, + 255, + 292, + 292 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d9d2b1c95c7f02a1d8fca63c3d025a0f755d16b8d30d32f621e88939ab03b5e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 302, + 143, + 496, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2107424f9c29b4c8e3cc810131011b20ef05a0ea9a3822c40be081770f90c764.jpg", + "image_caption": [ + "Fig. 2: Examples of challenges in the AMI-Traps dataset. (a) Incomplete: In some occasions, the body of an insect is incomplete due to its position in the automatically taken image. (b) Overlapping: Sometimes insects occlude each other. (c) Motion: Moving insects can be blurred and hard to classify. (d) Lighting: The lighting and camera exposure of the images is variable and can lead to poor contrast that masks some or all detail. (e) Resolution: Some insects are very small, leading to low-resolution images. (f) Perspective: While insects are generally positioned at a fixed distance, sometimes insects appear in the air or perched on the camera itself. Additionally, unexpected objects such as spider webs, dirt, and vegetation can be captured by the cameras and may be confused with objects of interest." + ], + "image_footnote": [], + "bbox": [ + 302, + 222, + 434, + 306 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d25edc3113fff52e6dd677f6c5d7786b8f7aa6715e2d2171963f3f77c8b59014.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 143, + 699, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/afdce32fb78a285bbb6342f87374635c6e7ed62ef5730bc0ab098be0a2df700b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 222, + 699, + 306 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4b7b837aeb3abd3c022276d253d1d6aa0761af1fe5cfe8d7f98afd06b4869868.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 148, + 774, + 185 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bde0083e1d7e69c0259918e0346dc0eff44e7c7ad1adc146b61a2226a13e5f33.jpg", + "image_caption": [ + "(d) Lighting", + "(e) Resolution" + ], + "image_footnote": [], + "bbox": [ + 714, + 200, + 777, + 241 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/65f593ddb7b468337474b2dd8a2f838564d7358c5ba6be8132ef0e21091b0f05.jpg", + "image_caption": [ + "(f) Perspective" + ], + "image_footnote": [], + "bbox": [ + 712, + 255, + 777, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "monitoring [55], and have in turn presented fascinating challenges for the computer vision community [14], including in classification [12, 29], object detection [13,31], out-of-distribution generalization [34,42,50], and active learning [54]. Such algorithms have to date, however, been designed mostly for large animals such as mammals.", + "bbox": [ + 212, + 489, + 784, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Insect Image Datasets. As algorithms and datasets for species recognition have proliferated, insects have traditionally been underrepresented. Within ImageNet1k [59], only 27 classes depicted insects, as compared to 57 for birds and 93 classes for wild mammals [49]. Some popular fine-grained classification datasets such as NABirds and CUB have focused exclusively on birds [69, 73], while camera trap datasets such as iWildCam have been focused on mammals [12]. The iNat2017 dataset [70] included 1,021 classes of insects out of 5,089 total classes in the dataset. By largely focusing on other taxonomic groups, prior works have neglected the challenges posed by insect data, such as extremely similar classes (reflecting far greater diversity than any other group of organisms), small size, hyperabundance, etc. In recent years, however, some works have introduced insect-focused benchmark datasets. The newly released BIOSCAN-1M dataset includes images and DNA sequences, with a focus on fly identification [33]. Despite the very large size of the dataset (1.1 million datapoints), the number of object classes involved is relatively small, with the focus being to distinguish between 40 taxonomic families of flies, reflecting the challenging nature of insect identification. Other recent datasets for computer vision include one focused on 8 classes of pollinators across Europe [61], and the Indian Butterflies dataset,", + "bbox": [ + 212, + 568, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "A. Jain et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "encompassing 30k images of butterflies from India across 315 species [67]. A variety of publications have used computer vision algorithms for detection of pest insects in specific agricultural applications [10, 52, 66].", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Insect Camera Traps. The use of camera traps for studying insect biodiversity is relatively recent. Geissmann et al. [32] introduced the Sticky Pi device, which uses sticky cards to capture (and kill) insects, which are then automatically photographed. Bjerge et al. [21] introduced an automated camera trap using UV light to attract and photograph moths and other insects (without killing them), which was built on by subsequent works [18, 20]. A similar approach focused on diurnal pollinators via a camera trap installed above flowers [17, 19]. These works also propose deep learning methods for processing the resulting images, considering both object detection and classification of a small number of insect categories. The lack of fine-grained classification is largely due to the time required to annotate images from the traps, making it difficult to obtain a large enough labeled dataset from trap data alone.", + "bbox": [ + 212, + 191, + 787, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Benchmark Design", + "text_level": 1, + "bbox": [ + 215, + 398, + 434, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we propose a benchmark dataset for insect species identification that addresses the challenges faced in real-world use cases. These include both (1) challenges that distinguish insect identification from other kinds of species identification, notably the massive diversity of often-similar species, and (2) the particular challenges associated with deploying these algorithms \"in the wild\".", + "bbox": [ + 212, + 431, + 782, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the first family of challenges, we introduce a dataset for fine-grained insect classification that (to our knowledge) is by far the largest yet released, with $\\sim 2.5\\mathrm{M}$ images across over 5,000 classes. We concentrate on moths, which are readily drawn to UV camera traps and can often be visually identified to species (unlike some other insects) [21].", + "bbox": [ + 212, + 507, + 782, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The second family of challenges is more complex - involving the gap between field conditions and the high-quality images typically used for training species identification algorithms. Figure 2 illustrates some of the challenges that computer vision algorithms face on data from automated insect camera traps, in contrast to images taken by humans. Insects may occlude each other or be partially \"out of frame\". Motion blur and lighting issues may occur, along with objects in unexpected positions. The resolution of images may be very low, owing to the small size of insects at a stereotyped distance from the camera. Also, insect camera traps typically use webcams and PCB camera modules due to their cost and software integration capabilities, and the resulting images are often both structurally different and overall lower quality than those taken by human photographers.", + "bbox": [ + 212, + 583, + 787, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As described in the preceding section, previous authors working with insect camera traps have addressed the problem of challenging camera trap data by training algorithms on annotated data from the particular camera trap in question. However, manual annotation is extremely time-intensive, since it must be performed by expert entomologists, of whom there are very few in any given", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Insect Identification in the Wild", + "bbox": [ + 514, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "region. Insect species vary greatly across regions, meaning that fine-grained annotations in one location are generally not useful in another. Furthermore, the hardware used in insect camera traps also varies greatly (with considerable differences in lighting and background, as well as image sharpness, contrast, noise reduction settings, and compression artifacts, which are exacerbated on small scale crops), meaning that even two datasets from the same region may not be comparable.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this paper, we remove the problem of annotation by solving a more challenging computer vision problem. Namely, we propose to train algorithms on abundant labeled images of insects taken by human photographers, then use these algorithms on out-of-distribution camera trap images. Our benchmark dataset is designed to test this functionality, by including both a large training set of curated photographs taken by humans and an out-of-distribution test set drawn from a geographically diverse set of insect camera traps.", + "bbox": [ + 212, + 252, + 787, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The AMI dataset is therefore designed to test fine-grained classification of moth species both in in-distribution and out-of-distribution contexts. It also includes the task of binary classification of insects as moth or non-moth, which is a prerequisite for camera traps since moths are not the only insects to appear. Note that we do not consider the task of object detection, which in applications is a precursor to insect identification. This is because object detection against a relatively plain background is a simpler task, where good results have already been obtained in the insect camera trap literature [18,20,21,32].", + "bbox": [ + 212, + 358, + 787, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 The AMI Dataset", + "text_level": 1, + "bbox": [ + 215, + 502, + 426, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we describe the construction of the datasets that we release in this work: AMI-Traps and AMI-GBIF. For AMI-Traps, we collected insect camera trap data by field deployments in three regions: NE-America (Northeast of North America), which includes data from Quebec in Canada and Vermont in the USA; W-Europe (Western Europe), which includes data from the UK and Denmark; and C-America (Central America), comprising data from Panama. For AMI-GBIF, we compiled public data for the same three regions using the Global Biodiversity Information Facility (GBIF) [7]. The GBIF platform provides access to biodiversity data aggregated from multiple sources, such as museums, research institutions, and citizen science platforms such as iNaturalist [4], Observation.org [5], and BugGuide [1]. For brevity, we sometimes refer to AMI-Traps and AMI-GBIF as AMI-T and AMI-G, respectively. Figure 3 shows examples of images from the two datasets and Tab. 1 compares the number of images and coverage of classes at different taxonomic levels.", + "bbox": [ + 212, + 534, + 787, + 746 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 AMI-Traps Dataset", + "text_level": 1, + "bbox": [ + 215, + 768, + 426, + 784 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data Collection. The source images in the AMI-Traps dataset were collected from a variety of different insect camera traps deployed across 22 monitoring stations in the three regions previously described. The cameras were configured", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "A. Jain et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6d3d09004541e7bbf9f29d11bdc159376009b17c5cc75bd7e740c4a592d92ee7.jpg", + "image_caption": [ + "Fig. 3: Sample images from the AMI Dataset. The AMI Dataset is composed of (1) AMI-GBIF, curated from a number of sources with imagery from citizen science and museum collections, and (2) AMI-Traps, drawn from automated camera traps for insects across five countries in three regions. The number of individual insects annotated per sub-dataset is denoted in parentheses." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 787, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to be motion-sensitive, as well as taking photos on a fixed interval of 10 minutes; we selected a subset of images for annotation so as to avoid highly similar images. In total, 2,893 trap images were sampled from all cameras and made available on the Label Studio annotation platform (see Appendix B for more details).", + "bbox": [ + 212, + 532, + 784, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Annotation Process. Images were annotated by entomologists with expertise in the species of the particular region. Labeling was conducted according a three-step process. 1) We designed a custom object detection model to suggest boxes around objects in the image (refer to Appendix B for details on this model), to save annotators time. Annotators were asked to correct errors in the proposed boxes as well as drawing any missing boxes. Insects appearing smaller than 1 centimeter were ignored. 2) Annotators were asked to label image crops according to one of three coarse categories: moth, non-moth, or unidentifiable. Some insects could not be labeled due to motion blur, occlusion, or other factors. 3) For insects labeled as moths, annotators were asked to select the narrowest taxonomic group that could be determined from the image - identifying moths to species where possible, if not species then genus, and if not genus then family.", + "bbox": [ + 212, + 595, + 785, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data Statistics. 2,893 source images from the traps were annotated (see Tab. B.2 in the supplementary for further details) which includes Vermont (900), Denmark (892), Quebec (533), UK (446), and Panama (122). A total of 52,948 insects were labeled within these images, including 37,105 $(70.08\\%)$ non-moths,", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Insect Identification in the Wild", + "bbox": [ + 514, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "14,105 (26.64%) moths, and 1,738 (3.28%) unidentifiable. Out of the 14,105 moth crops, 10,854 (76.95%), 7,065 (50.09%) and 5,374 (38.10%) crops were labeled at the family, genus, and species level respectively.", + "bbox": [ + 212, + 146, + 787, + 193 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 AMI-GBIF Dataset", + "text_level": 1, + "bbox": [ + 215, + 212, + 426, + 227 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AMI-GBIF has two sub-datasets: binary and fine-grained classification. The fine-grained classification dataset contains images for moth species in each of the regions covered by AMI-Traps. The binary classification dataset contains images of moths aggregated across all regions and non-moths from anywhere on earth.", + "bbox": [ + 212, + 234, + 785, + 296 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset Construction. First, in collaboration with regional moth specialists, we created \"checklists\" of species that could occur in each of the three regions of interest. It is worth noting that some species are included in more than one checklist, though most are unique to a single region. Next, we compared the checklists against the GBIF taxonomic backbone [2], merged synonym names (since some species have multiple accepted names), and manually investigated doubtful or unmatched names. Following that, we used the GBIF occurrence search tool to download metadata for all observations from the order Lepidoptera with images, which returned 15.1 million observations [3]. We then fetched images from observations in which the species is included in our regional checklists. The number of training images per species were limited to 1,000, though for many species it was considerably lower due to the long-tailed nature of the data. All images were resized to have the smallest dimension of 450px and packaged using the WebDataset format [6].", + "bbox": [ + 212, + 296, + 787, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quality Control. After downloading the images, we applied a set of cleaning steps to ensure the quality of the dataset. First, we removed duplicate URLs, i.e., images linked to more than one observation. This may occur when an image contains multiple individual insects, or when a placeholder is used for an observation (for example, Fig. A.1b in Appendix A shows a picture used as a placeholder for more than 100,000 observations). Next, we excluded images from specific datasets (refer to Tab. A.1 in Appendix A)) that we found to have a high percentage of images unsuitable for training our models, for instance, images of body parts or text only. Another issue we identified is that some observations contained thumbnail images that are too small to provide sufficient information to train our models; we removed all images with a width or height less than 64 pixels. Finally, as our dataset targets only adult insects, we removed images containing non-adult individuals such as eggs and caterpillars. For observations where the life stage was reported by the observer, approximately $60\\%$ of the cases, we removed all occurrences tagged as non-adult. For the remaining images with a missing life stage tag, we designed a classification model to predict the life stage (refer to Appendix A for details about this model) and removed all images classified as non-adult.", + "bbox": [ + 212, + 507, + 787, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset Split. Each GBIF observation may contain more than one image, but each image is considered an independent instance of the species. However, as images belonging to the same observation may be very similar, we keep images from the same observation in the same data split. To do so, we split the dataset", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "A. Jain et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "using the observation ID as reference, with the proportion of $70\\%$ , $10\\%$ , and $20\\%$ for the train, validation and test sets respectively. We apply a stratified split, i.e., we guarantee that this proportion is also maintained at the species level. We exclude species with fewer than 5 images, so the minimum number of images for a species in the training set is 4, and in the test set is one. Due to the data split based on observation ID, the validation set might lack images for some species. As some species may belong to multiple checklists, we keep the train/val/test splits consistent across the regions to prevent data leakage when training a unified model for all regions.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Binary Classification. Since moths are not the only insects that appear on the camera trap screen, for our proposed benchmark task of binary classification, we curate a dataset comprising 350,000 images for each of the moth and non-moth categories. The moth class consists of adult moth images randomly sampled from the AMI-GBIF dataset. For the non-moth category, we leveraged expert knowledge of the taxonomic groups likely to appear on the screen, fetching images of adult individuals from GBIF of the following groups: Diptera (85,474), Hemiptera (75,053), Odonata (68,770), Coleoptera (45,954), Araneae (21,321), Orthoptera (21,261), Formicidae (12,206), Ichneumonidae (6,581), Trichoptera (5,316), Neuroptera (4,880), Opiliones (1,880), Ephemeroptera (1,304) — this includes mostly insects, with some Arachnids. The dataset is split into train, validation, and test sets with proportions of $70\\%$ , $10\\%$ , and $20\\%$ , respectively. Specifically for the moth category, the splits are kept consistent with the splits for the fine-grained species classification task.", + "bbox": [ + 212, + 282, + 787, + 493 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Statistics. The final dataset contains $\\sim 2.5\\mathrm{M}$ images from 5,364 species, 1,734 genera and 77 families of moths, along with 350,000 additional images of non-moth groups. The binary dataset contains 700,000 images, composed of 350,000 moth images and the 350,000 non-moth images. The fine-grained dataset contains 2,564,392 moth images, and Tab. 1 shows the number of images and species for each region. Though species are selected based on their presence in regions of interest, we include all GBIF observations of these species from anywhere on Earth (see Fig. A.2 in supplementary for the global distribution of images). As in other biodiversity datasets, the number of images per species in AMI-GBIF follows a long-tailed distribution (see Fig. C.3 in supplementary). Specifically, only 3,099 species have more than 100 images and are considered under the \"many-shot\" bucket, 1,358 species contains between 20 and 100 images and are considered \"medium-shot\", and there are 907 species with less than 20 training images, considered as \"few-shot\" categories.", + "bbox": [ + 212, + 494, + 787, + 705 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Methods", + "text_level": 1, + "bbox": [ + 215, + 729, + 338, + 747 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we describe the methods used to assess the difficulty of the proposed benchmark tasks. We evaluate a variety of high-performing image classification architectures and use extensive data augmentation to address the challenges of domain shift between the two datasets. Motivated by the particular structure of the AMI-Traps dataset, we add a mixed-resolution (MixRes) data", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Insect Identification in the Wild", + "bbox": [ + 516, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/cf4692da1b0ca236aaa7751bf606018707568c1a42204c55a51125fd9061c300.jpg", + "table_caption": [ + "Table 1: Comparison between the AMI-GBIF and AMI-Traps dataset for the fine-grained classification task. Images for the traps dataset denote the individual insect crops and not the raw trap images. The last three rows show the unique classes for each of the three taxonomic levels for the two datasets for different regions. It is also important to note that regions have species in common." + ], + "table_footnote": [], + "table_body": "
All RegionsNE-AmericaW-EuropeC-America
AMI-GAMI-TAMI-GAMI-TAMI-GAMI-TAMI-GAMI-T
Images2,564,39214,1051,179,9439,0661,579,3334,06699,405973
Families7743683465262711
Genera1,73434488720495719238512
Species5,3645162,4972772,6032446366
", + "bbox": [ + 241, + 226, + 761, + 314 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "augmentation technique. As we describe in the following section, our approaches perform quite well, with MixRes leading to an especially significant performance improvement.", + "bbox": [ + 212, + 349, + 782, + 393 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Models. We work with five computer vision models widely used in identification of wildlife: ResNet50 [35], MobileNetV3-Large [37], ConvNeXt-T [47], ConvNeXt-B [47], and ViT-B/16 [28]. We select ResNet50 because it is the most popular backbone used as a baseline for almost any computer vision task. Considering the possibility that the trained models may need to run on hardware with limited resources, such as edge devices and personal computers, for in-field predictions, we choose MobileNetV3Large as a lightweight model alternative. Furthermore, considering that vision transformers have surpassed convolutional neural networks and become the state-of-the-art models for image classification, we select the ViT-B/16 model to provide a baseline from this family of architectures. Finally, for representing modern convolutional neural networks, we choose the recently introduced ConvNeXt architecture, which has proven to be competitive with transformers using only standard ConvNet modules. From this family of models, we select two versions: the base version (ConvNeXt-B) and, considering lightweight usage scenarios, we also include the tiny version (ConvNeXt-T). Unless otherwise stated, we use a default input resolution of 128x128 for all models.", + "bbox": [ + 212, + 398, + 787, + 654 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Data Augmentation Approaches. One characteristic of the in-the-wild images is the large variety of crops due to the diversity of insect sizes. A considerable number of them have small resolutions, below $64 \\times 64$ . At this scale, some important visual features that the model may have learned from high-resolution GBIF images might not be identifiable, leading to inaccurate predictions. This issue can be further amplified when these low-resolution crops are scaled up to the input model size. To compel the model to learn features for low-resolution inputs and avoid bias towards visual features only distinguishable in high-resolution images from other domains, we use a data augmentation technique we refer to as Mixed Resolution (MixRes) that simulates this low-resolution input during training (see Fig. D.4 in supplementary). This approach consists of randomly resizing the training input image to $25\\%$ or $50\\%$ of the model input size with a", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "A. Jain et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "probability $\\rho$ for each case. In our experiments $\\rho$ is set to 0.25, maintaining the original input size with probability 0.5. We employ MixRes in all experiments unless otherwise stated.", + "bbox": [ + 218, + 146, + 784, + 191 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In addition to MixRes, we also use a set of standard data augmentation operations: random resize crop with scale of 0.3 to 1.0 and an aspect ratio between $3/4$ and $4/3$ , random horizontal flip with $50\\%$ probability, and RandAugment [24] with $N = 2$ operations and magnitude $M = 9$ . Finally, the image is normalized using the mean and standard deviation of ImageNet and scaled to the input network size.", + "bbox": [ + 218, + 196, + 784, + 287 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details. We train all models with the AMI-GBIF training sets using the ImageNet-1K pretrained weights from the timm library [76] (for ConvNext and ViT we used ImageNet-22K weights). Each model is trained with a batch size of 128 and the AdamW optimizer [48] for 30 epochs. The initial learning rate is set to 0.001, with a weight decay of 1e-5. The learning rate is linearly warmed up from 0 to the initial value for 2 epochs and then decayed using cosine scheduling [36]. We also apply label smoothing [65] with a value of 0.1.", + "bbox": [ + 218, + 294, + 784, + 412 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Inference Settings. At test time, the insect crops from AMI-Traps are padded with black pixels to make them square, preventing the distortion of the shape of the insect before prediction. Next, the images are normalized with the mean and standard deviation of ImageNet and, finally, resized to the model input size. No padding is done for the test images from AMI-GBIF.", + "bbox": [ + 218, + 420, + 784, + 494 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Higher Taxonomic Rank Prediction. We consider the accuracy of our baseline models both with respect to the exact insect species predicted and with respect to higher level taxonomic groups such as genus and family. Indeed some images in our dataset cannot be labeled at the species level at all due to the quality or perspective of the image, and our annotators instead provided labels at the genus or family level. In our experiments, to make predictions at a higher taxonomic level, we simply sum together species-level predictions to get genus results, and similarly, the genus predictions are grouped together to get family predictions.", + "bbox": [ + 218, + 501, + 784, + 637 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation Metrics. During the binary classification step, it is important to achieve high precision by filtering out as many non-moths as possible to prevent the propagation of errors in the pipeline. At the same time, it is also important to assess recall because if too many moths are discarded at this step, it may impact the utility of the data for ecological studies. Therefore, for the binary classification task, along with accuracy, we also consider precision, recall, and F1-score as evaluation metrics. For the fine-grained classification task, we evaluate the model in terms of micro- and macro-averaged accuracy at the species, genus and family levels. Micro-averaging aggregates all class samples to calculate the accuracy while the macro-average is the average of per-class accuracy. Unless otherwise stated, the reported accuracy is micro-averaged. We conduct each experiment for three independent runs and report the mean and standard deviation.", + "bbox": [ + 218, + 643, + 784, + 838 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Insect Identification in the Wild", + "bbox": [ + 516, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6 Experiments and Results", + "text_level": 1, + "bbox": [ + 215, + 143, + 495, + 162 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Binary Classification Results. This task involves distinguishing moths from non-moths. Here we train a ResNet50 model on the AMI-GBIF dataset and evaluate performance both on the AMI-GBIF held-out test set and on the entirety of the AMI-Traps dataset. On AMI-GBIF (in-distribution), the model achieves a test accuracy of $97.77\\% \\pm 0.02\\%$ . Table 2 shows the results for the AMI-Traps dataset in more detail. We observe that the model has a high recall but a lower precision, which means it is over-predicting the moth class. Also, the bigger the image crop, the better the accuracy of the insect classification; many non-moths where the model errs are extremely small and are therefore highly blurred in the trap images. Table E.3 in the supplementary illustrates the improvement in model performance when crops are square padded (our default at inference time) versus an ablation where padding is not performed.", + "bbox": [ + 212, + 178, + 787, + 359 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/8650a1fd72c60892336133d2d9c26a31e49babb0317771eb2b06cb36c8805ad4.jpg", + "table_caption": [ + "Table 2: Binary classification model evaluation results on the AMI-Traps dataset after training on AMI-GBIF, showing that recall is high throughout, while precision drops on very small insects, which are predominantly non-moths. Crops $>100\\mathrm{px}$ or $>150\\mathrm{px}$ represent that atleast one side of the box is greater than the threshold." + ], + "table_footnote": [], + "table_body": "
MothsNon-mothsTotalAccuracyPrecisionRecallF1-score
All crops14,10537,10551,12086.48±1.0068.31±1.7095.03±0.3979.48±1.27
Crops > 100px11,63016,53428,16488.39±0.9779.46±1.4396.99±0.2087.35±0.94
Crops > 150px6,4465,96112,40792.95±0.2289.83±0.4397.48±0.1293.70±0.30
", + "bbox": [ + 217, + 454, + 785, + 508 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Fine-grained Classification Results. The fine-grained classification task aims to identify moths to the species level (and secondarily to genus or family level). In our experiments, we train models on AMI-GBIF data for each of the three regions of the AMI dataset, evaluating both on AMI-GBIF (in-distribution) and AMI-Traps (out-of-distribution) data (see Tab. 3). Not all moth species present in a region appear regularly at camera traps; hence, the list of species in AMI-Traps is a subset of species in AMI-GBIF for a given area (see Tab. 1). Therefore, we also evaluate the models on an in-distribution subset $(\\mathrm{Test}\\ddagger)$ from AMI-GBIF containing only the species present in AMI-Traps. Meanwhile, for AMI-Traps, there are 16 species not present in AMI-GBIF that were removed during our evaluation procedure (these represent unexpected species not included in the regional checklists used to define AMI-GBIF). Overall, the best model at the species level is ConvNeXt-B for both AMI-GBIF and AMI-traps. For C-America, annotations are mostly provided at the genus or family level, due to the challenging nature of visual identification of moths in this biodiversity-rich region. Furthermore, while the genus and family accuracies are similar between N-America and W-Europe, they are much lower for C-America. This can be attributed to the fact that there are 10 times fewer images in the training set for this region. Note that for genus and family evaluation, there are some cases where the accuracy is lower than the species level. This occurs because those", + "bbox": [ + 212, + 537, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "A. Jain et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/6bdaeca35aa9620bae8bfbffb510c666742eadef579b4842e70fc4724076e6d2.jpg", + "table_caption": [ + "Table 3: Accuracy results for the fine-grained classification task. Overall, ConvNeXt-B achieves the best results for both AMI-GBIF and AMI-traps, with ViT-B/16 achieving slightly better results for some cases at the genus and family levels. Test‡ represents a subset of the AMI-GBIF test set, containing only the species present in AMI-Traps. Species-level results on AMI-Traps for the C-America region are not available since most annotations were provided only at higher taxonomic levels." + ], + "table_footnote": [], + "table_body": "
RegionModelAMI-GBIFAMI-Traps
SpeciesSpeciesGenusFamily
Micro TestMicro Test‡Micro Top1Macro Top1Micro Top1Macro Top1Micro Top1Macro Top1
NE-AmericaResNet5087.56±0.1289.39±0.1471.86±1.0264.96±2.0080.07±0.4974.66±0.3590.44±0.4169.30±4.15
MBNetV3L83.73±0.1785.80±0.2967.41±0.7261.21±1.2074.83±1.3071.20±2.4789.31±0.6264.48±1.87
ConvNeXt-T89.81±0.0591.21±0.0674.89±0.0568.00±0.7683.26±1.2277.26±0.8590.61±0.5374.60±0.64
ConvNeXt-B90.45±0.0391.66±0.0376.75±1.5268.56±0.6285.46±0.8378.28±0.1491.07±0.2976.87±1.10
ViT-B/1686.39±0.0587.94±0.1077.43±0.7968.02±0.3485.57±0.9878.97±0.7291.45±0.3473.66±1.49
W-EuropeResNet5086.45±0.0890.25±0.0379.39±0.3972.84±0.2178.95±0.2673.22±0.6089.13±0.4275.24±1.85
MBNetV3L82.20±0.0586.47±0.1376.30±1.2970.35±0.9376.35±1.1972.88±1.2888.33±0.4772.21±0.38
ConvNeXt-T88.82±0.0392.05±0.0582.44±0.8974.59±0.6281.86±0.7974.15±0.5990.08±0.5974.58±2.93
ConvNeXt-B89.35±0.0192.48±0.0582.87±0.3976.41±0.5082.45±0.4475.98±1.0591.27±0.6079.39±2.70
ViT-B/1685.45±0.0389.22±0.1182.44±0.3675.25±0.9883.01±0.2675.96±0.3591.68±0.1775.81±0.79
C-AmericaResNet5087.46±0.2395.92±0.68--47.44±1.1166.59±0.6069.18±0.2442.40±0.31
MBNetV3L83.68±0.2492.89±0.27--41.03±2.2259.32±1.1868.34±1.6941.42±2.98
ConvNeXt-T87.65±0.1195.46±0.46--55.77±3.3370.68±1.5869.46±0.4245.35±0.67
ConvNeXt-B88.62±0.0996.22±0.26--51.92±3.8568.63±1.9472.24±1.4744.31±1.32
ViT-B/1679.80±0.2688.36±1.02--47.44±4.4464.70±4.0769.87±1.5142.00±3.60
", + "bbox": [ + 215, + 239, + 785, + 422 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "higher taxonomic levels include crops that have labels only at higher levels. In the supplementary material, we include additional results for AMI-GBIF long-tail accuracies (Tab. E.4), a pre-training ablation study (Tab. E.5), and a study analyzing classification accuracy as a function of prediction confidence (Fig. E.5).", + "bbox": [ + 212, + 452, + 782, + 513 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Data augmentation approaches. To assess the effectiveness of the mixed resolution data augmentation technique, we conduct an ablation study using ResNet50 for the NE-America region. As the effects of MixRes could already be covered by other data augmentation operations used in our preprocessing, such as random resize crop or one of the operations of RandAugment, we analyze multiple combinations of these operations. Additionally, we evaluate a higher input resolution, using the original 224 input size of ResNet50, to determine whether this leads to better predictions. Table 4 shows the results. While applying MixRes results in a slight decrease in accuracy within the in-distribution AMI-GBIF test set, for the in-the-wild AMI-Traps test set there is a substantial improvement, ranging between $4\\%$ and $7\\%$ when compared to models not trained with MixRes. Another important observation is that, although RandAugment and random resize together contribute to an overall improvement of around $16\\%$ , MixRes remains additionally beneficial. Finally, despite an increase of more than $4\\%$ in accuracy within the original domain when using a higher resolution, there is a marginal decrease in the accuracy of the in-the-wild results.", + "bbox": [ + 212, + 513, + 787, + 753 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 777, + 359, + 792 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Many ecological applications require species-level image classification to meet their goals. Despite advances in computer vision, there remain challenges in", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Insect Identification in the Wild", + "bbox": [ + 516, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/f6c860a8441da95dff8cbccbac4a26ab243dad7232acb9d117ff1fd7b9021d02.jpg", + "table_caption": [ + "Table 4: Ablation study on data augmentation techniques using a ResNet50 for the NE-America region evaluated on micro-averaged top-1 species accuracy. \"RandAug\" denotes the combination of the RandAugment and random resize crop data augmentation techniques. AMI-GBIF† test set contains only the species present in AMI-Traps." + ], + "table_footnote": [], + "table_body": "
ResolutionRandAugMixResAMI-GBIF‡AMI-Traps
128××89.18±0.0851.46±2.73
128✓×89.71±0.1667.88±1.27
128×✓87.41±0.0658.41±0.30
128✓✓89.39±0.1471.86±1.02
224✓✓92.97±0.1768.41±1.06
", + "bbox": [ + 313, + 224, + 689, + 311 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "implementing fine-grained classification that is effective in real-world scenarios. Here we outline the challenges of insect identification in the wild and propose a benchmark dataset to tackle them. We present AMI-GBIF, a dataset with 2.5 million labeled photographs of 5,000 moth species. We also introduce AMI-Traps, a dataset of images captured by automated insect monitoring systems across five countries, labeled by regional experts. We evaluate multiple classification models both on the AMI-GBIF dataset and in out-of-distribution generalization to AMI-Traps, with ConvNeXt-B generally performing the best. We also evaluate a range of data augmentation techniques, finding that a combination of simple approaches, notably mixed-resolution augmentation, considerably improves generalization under field conditions.", + "bbox": [ + 212, + 378, + 787, + 544 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Several limitations of this work should be noted. First, while annotating AMI-Traps already occupied a considerable time for expert entomologists, further expanding it to a greater number of species and geographic regions would be helpful in further stress-testing field deployments of fine-grained classification algorithms. Second, while we focus on camera traps for nocturnal insects, it would be useful also to consider diurnal insect camera traps, for example those aimed at pollinators, which pose some additional challenges, such as multiple perspectives and a diversity of backgrounds. We hope furthermore that future works will consider more sophisticated domain adaptation approaches building on the baselines we present here.", + "bbox": [ + 212, + 556, + 787, + 708 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work represents a step towards computer vision algorithms for species recognition that are effective in the ultra-challenging regime of insect identification and work effectively in field conditions. The approach we propose should make it possible in particular to generalize algorithms for insect camera traps rapidly and cost-effectively across different hardware setups and regions of interest. We hope that the datasets, challenges and strategies presented here will propel further research in computer vision for insects and help accelerate initiatives to understand and protect our planet's biodiversity.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "A. Jain et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We thank Annie-Shan Morin, Anna Viklund, and Jamie Alison for their helpful input and support to the project.", + "bbox": [ + 212, + 176, + 782, + 207 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research was generously funded by Mila Quebec AI Institute and Ministère de l'Économie, de l'Innovation et de l'Energie. The work was additionally supported in part by the Canada CIFAR AI Chairs program (NSERC 585136), the Global Climate Center on AI and Biodiversity Change (ABC) (NSF OISE2330423), the Montreal Insectarium, the Coordenação de Aperfeçoamento de Pessoal de Nível Superior - Brasil (CAPES-PROEX) - Finance Code 001, the Amazonas State Research Support Foundation - FAPEAM - through the POSGRAD project 2024/2025, the AMBER project funded by the abrdn Charitable Foundation, the Aage V. Jensen Naturfond grant no. N2024-0013, the European Union's Horizon Europe Research and Innovation programme under grant agreement no. 101060639 (MAMBO), and a Discovery Grant by NSERC. The computing resources were provided by Mila Quebec AI Institute, the Digital Research Alliance of Canada, and NVIDIA Corporation.", + "bbox": [ + 212, + 207, + 785, + 402 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "A.J., F.C., and M.J.B. led dataset creation and experiments. J.S.C. and L.P. contributed to machine learning pipeline development. N.P., F.H, J.R., M.B., M.S., J.F., A.A., Y.L., E.N., F.P.P., A.C.Z., J.A.R.S., and J.G. performed insect identification for the AMI-Traps dataset. T.A, K.B., A.G.S., M.B., Y.B., K.P.M, D. Roy, T.T.H., M.L., and D. Rolnick guided and advised the project.", + "bbox": [ + 212, + 402, + 785, + 479 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 502, + 323, + 517 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. BugGuide.Net: Identification, Images, & Information For Insects, Spiders & Their Kin For the United States & Canada, https://bugguide.net/ accessed Feb 12 2024", + "2. GBIF Backbone Taxonomy, https://doi.org/10.15468/39omei accessed Nov 13 2023", + "3. GBIF Occurrence Download, https://doi.org/10.15468/dl.4ua6wg accessed Oct 19 2023", + "4. iNaturalist, www.inaturalist.org, accessed Feb 9 2024", + "5. Observation.org, https://observation.org, accessed Feb 12 2024", + "6. WebDataset, https://github.com/webdataset/webdataset, accessed on April 2 2024", + "7. What is GBIF?, https://www.gbif.org/what-is-gbif accessed Sep 28 2023", + "8. Alison, J., Alexander, J.M., Diaz Zeugin, N., Dupont, Y.L., Iseli, E., Mann, H.M., Hoye, T.T.: Moths complement bumblebee pollination of red clover: a case for day-and-night insect surveillance. Biology Letters 18(7), 20220187 (2022)", + "9. Anton, V., Hartley, S., Geldenhuis, A., Wittmer, H.U.: Monitoring the mammalian fauna of urban areas using remote cameras and citizen science. Journal of Urban Ecology 4(1), juy002 (2018)", + "10. Badgujar, C.M., Armstrong, P.R., Gerken, A.R., Pordesimo, L.O., Campbell, J.F.: Real-time stored product insect detection and identification using deep learning: System integration and extensibility to mobile platforms. Journal of Stored Products Research 104, 102196 (2023)" + ], + "bbox": [ + 217, + 534, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Insect Identification in the Wild", + "bbox": [ + 514, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "11. Beery, S.: Scaling biodiversity monitoring for the data age. XRDS: Crossroads, The ACM Magazine for Students 27(4), 14-18 (2021)", + "12. Beery, S., Agarwal, A., Cole, E., Birodkar, V.: The iwildcam 2021 competition dataset. arXiv preprint arXiv:2105.03494 (2021)", + "13. Beery, S., Morris, D., Yang, S.: Efficient pipeline for camera trap image review. arXiv preprint arXiv:1907.06772 (2019)", + "14. Beery, S., Van Horn, G., Perona, P.: Recognition in terra incognita. In: Proceedings of the European conference on computer vision (ECCV). pp. 456-473 (2018)", + "15. Beery, S., Wu, G., Edwards, T., Pavetic, F., Majewski, B., Mukherjee, S., Chan, S., Morgan, J., Rathod, V., Huang, J.: The auto arborist dataset: A large-scale benchmark for multiview urban forest monitoring under domain shift. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21294-21307 (2022)", + "16. Berg, T., Liu, J., Woo Lee, S., Alexander, M.L., Jacobs, D.W., Belhumeur, P.N.: Birdsnap: Large-scale fine-grained visual categorization of birds. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2011-2018 (2014)", + "17. Bjerge, K., Alison, J., Dyrmann, M., Frigaard, C.E., Mann, H.M., Hoye, T.T.: Accurate detection and identification of insects from camera trap images with deep learning. PLOS Sustainability and Transformation 2(3), e0000051 (2023)", + "18. Bjerge, K., Frigaard, C.E., Karstoft, H.: Object detection of small insects in timelapse camera recordings. Sensors 23(16), 7242 (2023)", + "19. Bjerge, K., Geissmann, Q., Alison, J., Mann, H.M., Hoye, T.T., Dyrmann, M., Karstoft, H.: Hierarchical classification of insects with multitask learning and anomaly detection. Ecological Informatics 77, 102278 (2023)", + "20. Bjerge, K., Mann, H.M., Hoye, T.T.: Real-time insect tracking and monitoring with computer vision and deep learning. Remote Sensing in Ecology and Conservation 8(3), 315-327 (2022)", + "21. Bjerge, K., Nielsen, J.B., Sepstrup, M.V., Helsing-Nielsen, F., Hoye, T.T.: An automated light trap to monitor moths (lepidoptera) using computer vision-based tracking and deep learning. Sensors 21(2), 343 (2021)", + "22. Boom, B.J., He, J., Palazzo, S., Huang, P.X., Bayan, C., Chou, H.M., Lin, F.P., Spampinato, C., Fisher, R.B.: A research tool for long-term and continuous analysis of fish assemblage in coral-reefs using underwater camera footage. Ecological Informatics 23, 83-97 (2014)", + "23. Borowiec, M.L., Dikow, R.B., Frandsen, P.B., McKeeken, A., Valentini, G., White, A.E.: Deep learning as a tool for ecology and evolution. Methods in Ecology and Evolution 13(8), 1640-1660 (2022)", + "24. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 702-703 (2020)", + "25. Cutter, G., Stierhoff, K., Zeng, J.: Automated detection of rockfish in unconstrained underwater videos using haar cascades and a new image dataset: Labeled fishes in the wild. In: 2015 IEEE Winter Applications and Computer Vision Workshops. pp. 57-62. IEEE (2015)", + "26. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 248-255. IEEE (2009)" + ], + "bbox": [ + 217, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "A. Jain et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "27. Didham, R.K., Basset, Y., Collins, C.M., Leather, S.R., Littlewood, N.A., Menz, M.H., Müller, J., Packer, L., Saunders, M.E., Schonrogge, K., et al.: Interpreting insect declines: seven challenges and a way forward. Insect Conservation and Diversity 13(2), 103-114 (2020)", + "28. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net (2021), https://openreview.net/forum?id=YicbFdNTTy", + "29. Dubey, A., Ramanathan, V., Pentland, A., Mahajan, D.: Adaptive methods for real-world domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14340-14349 (2021)", + "30. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision 88, 303-338 (2010)", + "31. Fennell, M., Beirne, C., Burton, A.C.: Use of object detection in camera trap image identification: Assessing a method to rapidly and accurately classify human and animal detections for research and application in recreation ecology. Global Ecology and Conservation 35, e02104 (2022)", + "32. Geissmann, Q., Abram, P.K., Wu, D., Haney, C.H., Carrillo, J.: Sticky pi is a high-frequency smart trap that enables the study of insect circadian activity under natural conditions. PLoS Biology 20(7), e3001689 (2022)", + "33. Gharaee, Z., Gong, Z., Pellegrino, N., Zarubiieva, I., Haurum, J.B., Lowe, S.C., McKeown, J.T., Ho, C.C., McLeod, J., Wei, Y.Y.C., et al.: A step towards worldwide biodiversity assessment: The bioscan-1m insect dataset. arXiv preprint arXiv:2307.10455 (2023)", + "34. Gulrajani, I., Lopez-Paz, D.: In search of lost domain generalization. arXiv preprint arXiv:2007.01434 (2020)", + "35. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "36. He, T., Zhang, Z., Zhang, H., Zhang, Z., Xie, J., Li, M.: Bag of tricks for image classification with convolutional neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 558-567 (2019)", + "37. Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al.: Searching for mobilenetv3. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1314-1324 (2019)", + "38. Hoye, T.T., Arje, J., Bjerge, K., Hansen, O.L., Iosifidis, A., Leese, F., Mann, H.M., Meissner, K., Melvad, C., Raitoharju, J.: Deep learning and computer vision will transform entomology. Proceedings of the National Academy of Sciences 118(2), e2002545117 (2021)", + "39. Katija, K., Orenstein, E., Schlining, B., Lundsten, L., Barnard, K., Sainz, G., Boulais, O., Cromwell, M., Butler, E., Woodward, B., et al.: Fathomnet: A global image database for enabling artificial intelligence in the ocean. Scientific reports 12(1), 15914 (2022)", + "40. Kay, J., Kulits, P., Stathatos, S., Deng, S., Young, E., Beery, S., Van Horn, G., Perona, P.: The caltech fish counting dataset: A benchmark for multiple-object tracking and counting. In: European Conference on Computer Vision. pp. 290-311. Springer (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Insect Identification in the Wild", + "bbox": [ + 514, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "41. van Klink, R., August, T., Bas, Y., Bodesheim, P., Bonn, A., Fossoy, F., Høye, T.T., Jongejans, E., Menz, M.H., Miraldo, A., et al.: Emerging technologies revolutionise insect ecology and monitoring. Trends in ecology & evolution (2022)", + "42. Koh, P.W., Sagawa, S., Marklund, H., Xie, S.M., Zhang, M., Balsubramani, A., Hu, W., Yasunaga, M., Phillips, R.L., Gao, I., Lee, T., David, E., Stavness, I., Guo, W., Earnshaw, B.A., Haque, I.S., Beery, S., Leskovec, J., Kundaje, A., Pierson, E., Levine, S., Finn, C., Liang, P.: WILDS: A benchmark of in-the-wild distribution shifts. In: International Conference on Machine Learning (ICML) (2021)", + "43. Korsch, D., Bodesheim, P., Denzler, J.: Deep learning pipeline for automated visual moth monitoring: insect localization and species classification. arXiv preprint arXiv:2307.15427 (2023)", + "44. Krasin, I., Duerig, T., Alldrin, N., Ferrari, V., Abu-El-Haija, S., Kuznetsova, A., Rom, H., Uijlings, J., Popov, S., Veit, A., et al.: Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://github.com/openimages2(3), 18 (2017)", + "45. Kumar, N., Belhumeur, P.N., Biswas, A., Jacobs, D.W., Kress, W.J., Lopez, I.C., Soares, J.V.: Leafsnap: A computer vision system for automatic plant species identification. In: Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part II 12. pp. 502-516. Springer (2012)", + "46. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "47. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022)", + "48. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019)", + "49. Luccioni, A.S., Rolnick, D.: Bugs in the data: How imagenet misrepresents biodiversity. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 14382-14390 (2023)", + "50. Ming, Y., Yin, H., Li, Y.: On the impact of spurious correlation for out-of-distribution detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 10051-10059 (2022)", + "51. Mora, C., Tittensor, D.P., Adl, S., Simpson, A.G., Worm, B.: How many species are there on earth and in the ocean? PLoS biology 9(8), e1001127 (2011)", + "52. Nawoya, S., Ssemakula, F., Akol, R., Geissmann, Q., Karstoft, H., Bjerge, K., Mwikirize, C., Katumba, A., Gebreyesus, G.: Computer vision and deep learning in insects for food and feed production: A review. Computers and Electronics in Agriculture 216, 108503 (2024)", + "53. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian conference on computer vision, graphics & image processing. pp. 722-729. IEEE (2008)", + "54. Norouzzadeh, M.S., Morris, D., Beery, S., Joshi, N., Jojic, N., Clune, J.: A deep active learning system for species identification and counting in camera trap images. Methods in ecology and evolution 12(1), 150-161 (2021)", + "55. Oliver, R.Y., Iannarilli, F., Ahumada, J., Fegraus, E., Flores, N., Kays, R., Birch, T., Ranipeta, A., Rogan, M.S., Sica, Y.V., et al.: Camera trapping expands the view into global biodiversity and its change. Philosophical Transactions of the Royal Society B 378(1881), 20220232 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "A. Jain et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "56. Ouaknine, A., Kattenborn, T., Laliberte, E., Rolnick, D.: Openforest: A data catalogue for machine learning in forest monitoring. arXiv preprint arXiv:2311.00277 (2023)", + "57. Rolnick, D., Donti, P.L., Kaack, L.H., Kochanski, K., Lacoste, A., Sankaran, K., Ross, A.S., Milojevic-Dupont, N., Jaques, N., Waldman-Brown, A., et al.: Tackling climate change with machine learning. ACM Computing Surveys (CSUR) 55(2), 1-96 (2022)", + "58. Roy, D., et al.: Towards a standardised framework for ai-assisted, image-based monitoring of nocturnal insects. Philosophical Transactions of the Royal Society B (2024). https://doi.org/10.1098/rstb.2023.0108, (Forthcoming)", + "59. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) 115(3), 211-252 (2015)", + "60. Seibold, S., Gossner, M.M., Simons, N.K., Bluthgen, N., Müller, J., Ambarh, D., Ammer, C., Bauhus, J., Fischer, M., Habel, J.C., et al.: Arthropod decline in grasslands and forests is associated with landscape-level drivers. Nature 574(7780), 671-674 (2019)", + "61. Stark, T., Stefan, V., Wurm, M., Spanier, R., Taubenbock, H., Knight, T.M.: Yolo object detection models can locate and classify broad groups of flower-visiting arthropods in images. Scientific Reports 13(1), 16364 (2023)", + "62. Stork, N.E., et al.: How many species of insects and other terrestrial arthropods are there on earth. Annual review of entomology 63(1), 31-45 (2018)", + "63. Suto, J.: Codling moth monitoring with camera-equipped automated traps: A review. Agriculture 12(10), 1721 (2022)", + "64. Swanson, A., Kosmala, M., Lintott, C., Simpson, R., Smith, A., Packer, C.: Snapshot serengeti, high-frequency annotated camera trap images of 40 mammalian species in an african savanna. Scientific data 2(1), 1-14 (2015)", + "65. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818-2826 (2016)", + "66. Teixeira, A.C., Ribeiro, J., Morais, R., Sousa, J.J., Cunha, A.: A systematic review on automatic insect detection using deep learning. Agriculture 13(3), 713 (2023)", + "67. Theivaprakasham, H.: Identification of indian butterflies using deep convolutional neural network. Journal of Asia-Pacific Entomology 24(1), 329-340 (2021)", + "68. Tuia, D., Kellenberger, B., Beery, S., Costelloe, B.R., Zuffi, S., Risse, B., Mathis, A., Mathis, M.W., van Langevelde, F., Burghardt, T., et al.: Perspectives in machine learning for wildlife conservation. Nature communications 13(1), 792 (2022)", + "69. Van Horn, G., Branson, S., Farrell, R., Haber, S., Barry, J., Ipeirotis, P., Perona, P., Belongie, S.: Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 595-604 (2015)", + "70. Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist species classification and detection dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8769-8778 (2018)", + "71. Van Horn, G., Perona, P.: The devil is in the tails: Fine-grained classification in the wild. arXiv preprint arXiv:1709.01450 (2017)", + "72. Wagner, D.L.: Insect declines in the anthropocene. Annual review of entomology 65, 457-480 (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Insect Identification in the Wild", + "bbox": [ + 514, + 114, + 730, + 127 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "73. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset. Tech. rep. (2011)", + "74. Wei, X.S., Song, Y.Z., Mac Aodha, O., Wu, J., Peng, Y., Tang, J., Yang, J., Belongie, S.: Fine-grained image analysis with deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence 44(12), 8927-8948 (2021)", + "75. Weinstein, B.G.: A computer vision for animal ecology. Journal of Animal Ecology 87(3), 533-545 (2018)", + "76. Wightman, R.: Pytorch image models. https://github.com/rwrightman/pytorch-image-models (2019)", + "77. Zhou, K., Liu, Z., Qiao, Y., Xiang, T., Loy, C.C.: Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 299 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "A. Jain et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/Insect Identification in the Wild_ The AMI Dataset/deef9d64-c259-4177-ab17-679ea66a461f_model.json b/2024/Insect Identification in the Wild_ The AMI Dataset/deef9d64-c259-4177-ab17-679ea66a461f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..40d859044458d46f4285b69cc30488c8137e4aba --- /dev/null +++ b/2024/Insect Identification in the Wild_ The AMI Dataset/deef9d64-c259-4177-ab17-679ea66a461f_model.json @@ -0,0 +1,2605 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.315, + 0.141, + 0.69, + 0.185 + ], + "angle": 0, + "content": "Insect Identification in the Wild: The AMI Dataset" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.212, + 0.784, + 0.335 + ], + "angle": 0, + "content": "Aditya Jain1*, Fagner Cunha2*, Michael James Bunsen1*, Juan Sebastián Canas1, Léonard Pasi1, Nathan Pinoy3, Flemming Helsing3, JoAnne Russo4, Marc Botham5, Michael Sabourin6, Jonathan Fréchette12, Alexandre Anctil7, Yacksecari Lopez8, Eduardo Navarro8, Filonila Perez Pimentel8, Ana Cecilia Zamora8, José Alejandro Ramirez Silva8, Jonathan Gagnon12, Tom August5, Kim Bjerge3, Alba Gomez Segura5, Marc Bélisle9, Yves Basset8, Kent P. McFarland4, David Roy5, Toke Thomas Hoye3, Maxim Larrivée10, and David Rolnick1,11" + }, + { + "type": "text", + "bbox": [ + 0.406, + 0.346, + 0.6, + 0.36 + ], + "angle": 0, + "content": "1 Mila - Quebec AI Institute" + }, + { + "type": "text", + "bbox": [ + 0.389, + 0.36, + 0.615, + 0.373 + ], + "angle": 0, + "content": "2 Federal University of Amazonas" + }, + { + "type": "text", + "bbox": [ + 0.434, + 0.373, + 0.57, + 0.388 + ], + "angle": 0, + "content": "3 Aarhus University" + }, + { + "type": "text", + "bbox": [ + 0.393, + 0.388, + 0.612, + 0.4 + ], + "angle": 0, + "content": "4 Vermont Center for Ecostudies" + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.401, + 0.631, + 0.416 + ], + "angle": 0, + "content": "UK Centre for Ecology & Hydrology" + }, + { + "type": "text", + "bbox": [ + 0.44, + 0.416, + 0.564, + 0.428 + ], + "angle": 0, + "content": "McGuire Center" + }, + { + "type": "text", + "bbox": [ + 0.311, + 0.429, + 0.693, + 0.443 + ], + "angle": 0, + "content": "7 Centre de données sur le patrimoine naturel du Québec" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.443, + 0.643, + 0.457 + ], + "angle": 0, + "content": "Smithsonian Tropical Research Institute" + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.457, + 0.593, + 0.469 + ], + "angle": 0, + "content": "9 Université de Sherbrooke" + }, + { + "type": "text", + "bbox": [ + 0.42, + 0.47, + 0.584, + 0.484 + ], + "angle": 0, + "content": "10 Montreal Insectarium" + }, + { + "type": "text", + "bbox": [ + 0.432, + 0.485, + 0.572, + 0.498 + ], + "angle": 0, + "content": "11 McGill University" + }, + { + "type": "text", + "bbox": [ + 0.413, + 0.498, + 0.592, + 0.512 + ], + "angle": 0, + "content": "\\(^{12}\\) Independent Researcher" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.53, + 0.741, + 0.807 + ], + "angle": 0, + "content": "Abstract. Insects represent half of all global biodiversity, yet many of the world's insects are disappearing, with severe implications for ecosystems and agriculture. Despite this crisis, data on insect diversity and abundance remain woefully inadequate, due to the scarcity of human experts and the lack of scalable tools for monitoring. Ecologists have started to adopt camera traps to record and study insects, and have proposed computer vision algorithms as an answer for scalable data processing. However, insect monitoring in the wild poses unique challenges that have not yet been addressed within computer vision, including the combination of long-tailed data, extremely similar classes, and significant distribution shifts. We provide the first large-scale machine learning benchmarks for fine-grained insect recognition, designed to match real-world tasks faced by ecologists. Our contributions include a curated dataset of images from citizen science platforms and museums, and an expert-annotated dataset drawn from automated camera traps across multiple continents, designed to test out-of-distribution generalization under field conditions. We train and evaluate a variety of baseline algorithms and introduce a combination of data augmentation techniques that enhance generalization across geographies and hardware setups. The dataset is made publicly available1." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.812, + 0.703, + 0.826 + ], + "angle": 0, + "content": "* Denotes equal contribution. Correspondence to: moth-ai@mila.quebec" + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.826, + 0.553, + 0.84 + ], + "angle": 0, + "content": "\\(^{1}\\) https://github.com/RolnickLab/ami-dataset" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.812, + 0.703, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "A. Jain et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.375, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.184, + 0.788, + 0.273 + ], + "angle": 0, + "content": "Computer vision algorithms for species identification have become important tools in studying and protecting the Earth's wildlife. Biodiversity applications have also been beneficial to computer vision, serving as a fruitful source of new benchmarks and algorithms [11,23,68,75]. However, research in computer vision for biodiversity has often focused on large animals, not reflecting the full set of challenges faced in ecology and conservation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.276, + 0.788, + 0.486 + ], + "angle": 0, + "content": "There are over a million known species of insects on Earth, representing nearly one half of all known organisms (by comparison, there are only 10,000 species of birds, and even fewer mammals) [51,62]. Insects play essential roles in nearly all terrestrial ecosystems, as well as providing ecosystem services such as pollination that are critical for humanity [60], yet a looming \"insect apocalyptic\" threatens both their abundance and diversity [72]. Despite the crisis, data on insect populations is shockingly sparse [27], as a result of less resources, a plethora of species, and the high level of expertise typically required to study insects. Amid this data gap, computer vision tools for automated monitoring have the potential to radically reshape the study and conservation of insects. In particular, recent breakthroughs in camera traps for insects have made it possible to automatically photograph large numbers of individuals in the wild [38,41,58], opening up the the potential for image-processing algorithms to parse this raw data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.639 + ], + "angle": 0, + "content": "Species recognition from camera traps is a well-studied problem in the computer vision community [55], with common challenges including poor lighting, occlusion of individuals, camouflage, and blur [14]. However, working with insects presents distinct challenges from those faced in traditional camera traps, which are often designed for large animals. For example, while traditional camera trap images may only rarely contain a target species, almost every image from an insect camera trap includes insects – often hundreds of them. Where traditional camera traps may have a short list of species of interest, insect monitoring requires considering thousands of species at any given location, with many of them being extremely similar, and most with very little labeled data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.788, + 0.732 + ], + "angle": 0, + "content": "While recent authors have proposed the use of computer vision tools for detecting and identifying insects in camera trap data [8, 17, 20, 32, 43, 63], such studies have involved small datasets with only a few species represented. There has to date been no work assessing the performance of algorithms in identifying large number of insect species, or addressing the challenges of generalization across in-the-wild deployments. Our work here offers the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.749, + 0.789, + 0.842 + ], + "angle": 0, + "content": "- We introduce the AMI (Automated Monitoring of Insects) dataset, consisting of two parts: 1) AMI-GBIF, a dataset of \\(\\sim 2.5\\mathrm{M}\\) human-captured images curated from citizen science platforms and museum collections, 2) AMI-Traps, an expert-annotated dataset of 2,893 insect camera trap images (representing 52,948 labeled insects) collected from a global network of automated camera traps, designed to test in-the-wild performance (Fig. 1)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.516, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Insect Identification in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.143, + 0.788, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.285, + 0.788, + 0.382 + ], + "angle": 0, + "content": "Fig. 1: Illustration of the difference between insect species identification from human-captured images and in the wild. On left are shown images of three moth species from our AMI-GBIF dataset, curated from citizen science data and museum collections. On right are shown images of the same species (along with other insects) within a photograph taken by an automated camera trap in the wild from our AMI-Traps dataset. One of the challenges of the AMI dataset is generalizing from AMI-GBIF to AMI-Traps without additional labeled data." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.412, + 0.787, + 0.472 + ], + "angle": 0, + "content": "- We frame the problem of practical insect monitoring in the field as a set of computer vision tasks, entailing very fine-grained classification from long-tailed data, with the additional challenge of generalization across multiple kinds of imagery." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.473, + 0.787, + 0.517 + ], + "angle": 0, + "content": "- We test a variety of strong baseline models for the proposed tasks, including a combination of data augmentation approaches that significantly improve performance under field conditions." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.412, + 0.787, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.552, + 0.388, + 0.568 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Biodiversity Image Datasets. Advancements in computer vision have been closely intertwined with the creation of curated datasets such as ImageNet [26, 59], Pascal VOC [30], Open Images [44], and MS-COCO [46]. In recent years, the computer vision and machine learning communities have increasingly directed their focus towards addressing pressing human threats, including biodiversity loss [68] and climate change [57]. These domains not only hold significant potential for societal impact but also offer fertile ground for exploring intriguing machine learning questions such as domain generalization [77], long-tailed distributions [71], and fine-grained classification [74]. Such challenges have found especially fruitful testbeds in the classification of species, where datasets such as iNaturalist [70] have proven useful for both the computer vision and ecology communities. While some challenges have considered plant identification [15, 45, 53, 56], the majority have focused on animals, including datasets focused on birds [16, 69, 73], mammals [9, 14, 64], and fish [22, 25, 39, 40]. Such datasets have often leveraged concurrent advancements in sensor technology, including in satellite imagery and drones. In particular, camera traps have been widely adopted by biologists as an in situ and cost-effective means of wildlife" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "A. Jain et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.148, + 0.296, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.189, + 0.293, + 0.2 + ], + "angle": 0, + "content": "(a) Incomplete" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.201, + 0.291, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.243, + 0.297, + 0.254 + ], + "angle": 0, + "content": "(b) Overlapping" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.256, + 0.293, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.295, + 0.284, + 0.305 + ], + "angle": 0, + "content": "(c) Motion" + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.144, + 0.498, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.223, + 0.436, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.144, + 0.7, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.223, + 0.7, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.149, + 0.776, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.718, + 0.189, + 0.774, + 0.2 + ], + "angle": 0, + "content": "(d) Lighting" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.202, + 0.778, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.243, + 0.779, + 0.253 + ], + "angle": 0, + "content": "(e) Resolution" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.256, + 0.779, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.295, + 0.78, + 0.305 + ], + "angle": 0, + "content": "(f) Perspective" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.319, + 0.788, + 0.457 + ], + "angle": 0, + "content": "Fig. 2: Examples of challenges in the AMI-Traps dataset. (a) Incomplete: In some occasions, the body of an insect is incomplete due to its position in the automatically taken image. (b) Overlapping: Sometimes insects occlude each other. (c) Motion: Moving insects can be blurred and hard to classify. (d) Lighting: The lighting and camera exposure of the images is variable and can lead to poor contrast that masks some or all detail. (e) Resolution: Some insects are very small, leading to low-resolution images. (f) Perspective: While insects are generally positioned at a fixed distance, sometimes insects appear in the air or perched on the camera itself. Additionally, unexpected objects such as spider webs, dirt, and vegetation can be captured by the cameras and may be confused with objects of interest." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.785, + 0.566 + ], + "angle": 0, + "content": "monitoring [55], and have in turn presented fascinating challenges for the computer vision community [14], including in classification [12, 29], object detection [13,31], out-of-distribution generalization [34,42,50], and active learning [54]. Such algorithms have to date, however, been designed mostly for large animals such as mammals." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Insect Image Datasets. As algorithms and datasets for species recognition have proliferated, insects have traditionally been underrepresented. Within ImageNet1k [59], only 27 classes depicted insects, as compared to 57 for birds and 93 classes for wild mammals [49]. Some popular fine-grained classification datasets such as NABirds and CUB have focused exclusively on birds [69, 73], while camera trap datasets such as iWildCam have been focused on mammals [12]. The iNat2017 dataset [70] included 1,021 classes of insects out of 5,089 total classes in the dataset. By largely focusing on other taxonomic groups, prior works have neglected the challenges posed by insect data, such as extremely similar classes (reflecting far greater diversity than any other group of organisms), small size, hyperabundance, etc. In recent years, however, some works have introduced insect-focused benchmark datasets. The newly released BIOSCAN-1M dataset includes images and DNA sequences, with a focus on fly identification [33]. Despite the very large size of the dataset (1.1 million datapoints), the number of object classes involved is relatively small, with the focus being to distinguish between 40 taxonomic families of flies, reflecting the challenging nature of insect identification. Other recent datasets for computer vision include one focused on 8 classes of pollinators across Europe [61], and the Indian Butterflies dataset," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.516, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Insect Identification in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "encompassing 30k images of butterflies from India across 315 species [67]. A variety of publications have used computer vision algorithms for detection of pest insects in specific agricultural applications [10, 52, 66]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Insect Camera Traps. The use of camera traps for studying insect biodiversity is relatively recent. Geissmann et al. [32] introduced the Sticky Pi device, which uses sticky cards to capture (and kill) insects, which are then automatically photographed. Bjerge et al. [21] introduced an automated camera trap using UV light to attract and photograph moths and other insects (without killing them), which was built on by subsequent works [18, 20]. A similar approach focused on diurnal pollinators via a camera trap installed above flowers [17, 19]. These works also propose deep learning methods for processing the resulting images, considering both object detection and classification of a small number of insect categories. The lack of fine-grained classification is largely due to the time required to annotate images from the traps, making it difficult to obtain a large enough labeled dataset from trap data alone." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.399, + 0.436, + 0.417 + ], + "angle": 0, + "content": "3 Benchmark Design" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.784, + 0.507 + ], + "angle": 0, + "content": "In this paper, we propose a benchmark dataset for insect species identification that addresses the challenges faced in real-world use cases. These include both (1) challenges that distinguish insect identification from other kinds of species identification, notably the massive diversity of often-similar species, and (2) the particular challenges associated with deploying these algorithms \"in the wild\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.784, + 0.583 + ], + "angle": 0, + "content": "To address the first family of challenges, we introduce a dataset for fine-grained insect classification that (to our knowledge) is by far the largest yet released, with \\(\\sim 2.5\\mathrm{M}\\) images across over 5,000 classes. We concentrate on moths, which are readily drawn to UV camera traps and can often be visually identified to species (unlike some other insects) [21]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.765 + ], + "angle": 0, + "content": "The second family of challenges is more complex - involving the gap between field conditions and the high-quality images typically used for training species identification algorithms. Figure 2 illustrates some of the challenges that computer vision algorithms face on data from automated insect camera traps, in contrast to images taken by humans. Insects may occlude each other or be partially \"out of frame\". Motion blur and lighting issues may occur, along with objects in unexpected positions. The resolution of images may be very low, owing to the small size of insects at a stereotyped distance from the camera. Also, insect camera traps typically use webcams and PCB camera modules due to their cost and software integration capabilities, and the resulting images are often both structurally different and overall lower quality than those taken by human photographers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.842 + ], + "angle": 0, + "content": "As described in the preceding section, previous authors working with insect camera traps have addressed the problem of challenging camera trap data by training algorithms on annotated data from the particular camera trap in question. However, manual annotation is extremely time-intensive, since it must be performed by expert entomologists, of whom there are very few in any given" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "A. Jain et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "region. Insect species vary greatly across regions, meaning that fine-grained annotations in one location are generally not useful in another. Furthermore, the hardware used in insect camera traps also varies greatly (with considerable differences in lighting and background, as well as image sharpness, contrast, noise reduction settings, and compression artifacts, which are exacerbated on small scale crops), meaning that even two datasets from the same region may not be comparable." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.358 + ], + "angle": 0, + "content": "In this paper, we remove the problem of annotation by solving a more challenging computer vision problem. Namely, we propose to train algorithms on abundant labeled images of insects taken by human photographers, then use these algorithms on out-of-distribution camera trap images. Our benchmark dataset is designed to test this functionality, by including both a large training set of curated photographs taken by humans and an out-of-distribution test set drawn from a geographically diverse set of insect camera traps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.48 + ], + "angle": 0, + "content": "The AMI dataset is therefore designed to test fine-grained classification of moth species both in in-distribution and out-of-distribution contexts. It also includes the task of binary classification of insects as moth or non-moth, which is a prerequisite for camera traps since moths are not the only insects to appear. Note that we do not consider the task of object detection, which in applications is a precursor to insect identification. This is because object detection against a relatively plain background is a simpler task, where good results have already been obtained in the insect camera trap literature [18,20,21,32]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.503, + 0.427, + 0.518 + ], + "angle": 0, + "content": "4 The AMI Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.747 + ], + "angle": 0, + "content": "In this section, we describe the construction of the datasets that we release in this work: AMI-Traps and AMI-GBIF. For AMI-Traps, we collected insect camera trap data by field deployments in three regions: NE-America (Northeast of North America), which includes data from Quebec in Canada and Vermont in the USA; W-Europe (Western Europe), which includes data from the UK and Denmark; and C-America (Central America), comprising data from Panama. For AMI-GBIF, we compiled public data for the same three regions using the Global Biodiversity Information Facility (GBIF) [7]. The GBIF platform provides access to biodiversity data aggregated from multiple sources, such as museums, research institutions, and citizen science platforms such as iNaturalist [4], Observation.org [5], and BugGuide [1]. For brevity, we sometimes refer to AMI-Traps and AMI-GBIF as AMI-T and AMI-G, respectively. Figure 3 shows examples of images from the two datasets and Tab. 1 compares the number of images and coverage of classes at different taxonomic levels." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.769, + 0.427, + 0.785 + ], + "angle": 0, + "content": "4.1 AMI-Traps Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Data Collection. The source images in the AMI-Traps dataset were collected from a variety of different insect camera traps deployed across 22 monitoring stations in the three regions previously described. The cameras were configured" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.516, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Insect Identification in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.146, + 0.788, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.43, + 0.788, + 0.501 + ], + "angle": 0, + "content": "Fig. 3: Sample images from the AMI Dataset. The AMI Dataset is composed of (1) AMI-GBIF, curated from a number of sources with imagery from citizen science and museum collections, and (2) AMI-Traps, drawn from automated camera traps for insects across five countries in three regions. The number of individual insects annotated per sub-dataset is denoted in parentheses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.785, + 0.595 + ], + "angle": 0, + "content": "to be motion-sensitive, as well as taking photos on a fixed interval of 10 minutes; we selected a subset of images for annotation so as to avoid highly similar images. In total, 2,893 trap images were sampled from all cameras and made available on the Label Studio annotation platform (see Appendix B for more details)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.787, + 0.779 + ], + "angle": 0, + "content": "Annotation Process. Images were annotated by entomologists with expertise in the species of the particular region. Labeling was conducted according a three-step process. 1) We designed a custom object detection model to suggest boxes around objects in the image (refer to Appendix B for details on this model), to save annotators time. Annotators were asked to correct errors in the proposed boxes as well as drawing any missing boxes. Insects appearing smaller than 1 centimeter were ignored. 2) Annotators were asked to label image crops according to one of three coarse categories: moth, non-moth, or unidentifiable. Some insects could not be labeled due to motion blur, occlusion, or other factors. 3) For insects labeled as moths, annotators were asked to select the narrowest taxonomic group that could be determined from the image - identifying moths to species where possible, if not species then genus, and if not genus then family." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Data Statistics. 2,893 source images from the traps were annotated (see Tab. B.2 in the supplementary for further details) which includes Vermont (900), Denmark (892), Quebec (533), UK (446), and Panama (122). A total of 52,948 insects were labeled within these images, including 37,105 \\((70.08\\%)\\) non-moths," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "A. Jain et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.194 + ], + "angle": 0, + "content": "14,105 (26.64%) moths, and 1,738 (3.28%) unidentifiable. Out of the 14,105 moth crops, 10,854 (76.95%), 7,065 (50.09%) and 5,374 (38.10%) crops were labeled at the family, genus, and species level respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.213, + 0.427, + 0.228 + ], + "angle": 0, + "content": "4.2 AMI-GBIF Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.236, + 0.786, + 0.297 + ], + "angle": 0, + "content": "AMI-GBIF has two sub-datasets: binary and fine-grained classification. The fine-grained classification dataset contains images for moth species in each of the regions covered by AMI-Traps. The binary classification dataset contains images of moths aggregated across all regions and non-moths from anywhere on earth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.788, + 0.508 + ], + "angle": 0, + "content": "Dataset Construction. First, in collaboration with regional moth specialists, we created \"checklists\" of species that could occur in each of the three regions of interest. It is worth noting that some species are included in more than one checklist, though most are unique to a single region. Next, we compared the checklists against the GBIF taxonomic backbone [2], merged synonym names (since some species have multiple accepted names), and manually investigated doubtful or unmatched names. Following that, we used the GBIF occurrence search tool to download metadata for all observations from the order Lepidoptera with images, which returned 15.1 million observations [3]. We then fetched images from observations in which the species is included in our regional checklists. The number of training images per species were limited to 1,000, though for many species it was considerably lower due to the long-tailed nature of the data. All images were resized to have the smallest dimension of 450px and packaged using the WebDataset format [6]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Quality Control. After downloading the images, we applied a set of cleaning steps to ensure the quality of the dataset. First, we removed duplicate URLs, i.e., images linked to more than one observation. This may occur when an image contains multiple individual insects, or when a placeholder is used for an observation (for example, Fig. A.1b in Appendix A shows a picture used as a placeholder for more than 100,000 observations). Next, we excluded images from specific datasets (refer to Tab. A.1 in Appendix A)) that we found to have a high percentage of images unsuitable for training our models, for instance, images of body parts or text only. Another issue we identified is that some observations contained thumbnail images that are too small to provide sufficient information to train our models; we removed all images with a width or height less than 64 pixels. Finally, as our dataset targets only adult insects, we removed images containing non-adult individuals such as eggs and caterpillars. For observations where the life stage was reported by the observer, approximately \\(60\\%\\) of the cases, we removed all occurrences tagged as non-adult. For the remaining images with a missing life stage tag, we designed a classification model to predict the life stage (refer to Appendix A for details about this model) and removed all images classified as non-adult." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Dataset Split. Each GBIF observation may contain more than one image, but each image is considered an independent instance of the species. However, as images belonging to the same observation may be very similar, we keep images from the same observation in the same data split. To do so, we split the dataset" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.517, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Insect Identification in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "using the observation ID as reference, with the proportion of \\(70\\%\\), \\(10\\%\\), and \\(20\\%\\) for the train, validation and test sets respectively. We apply a stratified split, i.e., we guarantee that this proportion is also maintained at the species level. We exclude species with fewer than 5 images, so the minimum number of images for a species in the training set is 4, and in the test set is one. Due to the data split based on observation ID, the validation set might lack images for some species. As some species may belong to multiple checklists, we keep the train/val/test splits consistent across the regions to prevent data leakage when training a unified model for all regions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.494 + ], + "angle": 0, + "content": "Binary Classification. Since moths are not the only insects that appear on the camera trap screen, for our proposed benchmark task of binary classification, we curate a dataset comprising 350,000 images for each of the moth and non-moth categories. The moth class consists of adult moth images randomly sampled from the AMI-GBIF dataset. For the non-moth category, we leveraged expert knowledge of the taxonomic groups likely to appear on the screen, fetching images of adult individuals from GBIF of the following groups: Diptera (85,474), Hemiptera (75,053), Odonata (68,770), Coleoptera (45,954), Araneae (21,321), Orthoptera (21,261), Formicidae (12,206), Ichneumonidae (6,581), Trichoptera (5,316), Neuroptera (4,880), Opiliones (1,880), Ephemeroptera (1,304) — this includes mostly insects, with some Arachnids. The dataset is split into train, validation, and test sets with proportions of \\(70\\%\\), \\(10\\%\\), and \\(20\\%\\), respectively. Specifically for the moth category, the splits are kept consistent with the splits for the fine-grained species classification task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.706 + ], + "angle": 0, + "content": "Statistics. The final dataset contains \\(\\sim 2.5\\mathrm{M}\\) images from 5,364 species, 1,734 genera and 77 families of moths, along with 350,000 additional images of non-moth groups. The binary dataset contains 700,000 images, composed of 350,000 moth images and the 350,000 non-moth images. The fine-grained dataset contains 2,564,392 moth images, and Tab. 1 shows the number of images and species for each region. Though species are selected based on their presence in regions of interest, we include all GBIF observations of these species from anywhere on Earth (see Fig. A.2 in supplementary for the global distribution of images). As in other biodiversity datasets, the number of images per species in AMI-GBIF follows a long-tailed distribution (see Fig. C.3 in supplementary). Specifically, only 3,099 species have more than 100 images and are considered under the \"many-shot\" bucket, 1,358 species contains between 20 and 100 images and are considered \"medium-shot\", and there are 907 species with less than 20 training images, considered as \"few-shot\" categories." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.731, + 0.339, + 0.748 + ], + "angle": 0, + "content": "5 Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this section, we describe the methods used to assess the difficulty of the proposed benchmark tasks. We evaluate a variety of high-performing image classification architectures and use extensive data augmentation to address the challenges of domain shift between the two datasets. Motivated by the particular structure of the AMI-Traps dataset, we add a mixed-resolution (MixRes) data" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "A. Jain et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 1: Comparison between the AMI-GBIF and AMI-Traps dataset for the fine-grained classification task. Images for the traps dataset denote the individual insect crops and not the raw trap images. The last three rows show the unique classes for each of the three taxonomic levels for the two datasets for different regions. It is also important to note that regions have species in common." + }, + { + "type": "table", + "bbox": [ + 0.243, + 0.227, + 0.763, + 0.315 + ], + "angle": 0, + "content": "
All RegionsNE-AmericaW-EuropeC-America
AMI-GAMI-TAMI-GAMI-TAMI-GAMI-TAMI-GAMI-T
Images2,564,39214,1051,179,9439,0661,579,3334,06699,405973
Families7743683465262711
Genera1,73434488720495719238512
Species5,3645162,4972772,6032446366
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.784, + 0.395 + ], + "angle": 0, + "content": "augmentation technique. As we describe in the following section, our approaches perform quite well, with MixRes leading to an especially significant performance improvement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.789, + 0.655 + ], + "angle": 0, + "content": "Models. We work with five computer vision models widely used in identification of wildlife: ResNet50 [35], MobileNetV3-Large [37], ConvNeXt-T [47], ConvNeXt-B [47], and ViT-B/16 [28]. We select ResNet50 because it is the most popular backbone used as a baseline for almost any computer vision task. Considering the possibility that the trained models may need to run on hardware with limited resources, such as edge devices and personal computers, for in-field predictions, we choose MobileNetV3Large as a lightweight model alternative. Furthermore, considering that vision transformers have surpassed convolutional neural networks and become the state-of-the-art models for image classification, we select the ViT-B/16 model to provide a baseline from this family of architectures. Finally, for representing modern convolutional neural networks, we choose the recently introduced ConvNeXt architecture, which has proven to be competitive with transformers using only standard ConvNet modules. From this family of models, we select two versions: the base version (ConvNeXt-B) and, considering lightweight usage scenarios, we also include the tiny version (ConvNeXt-T). Unless otherwise stated, we use a default input resolution of 128x128 for all models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Data Augmentation Approaches. One characteristic of the in-the-wild images is the large variety of crops due to the diversity of insect sizes. A considerable number of them have small resolutions, below \\(64 \\times 64\\). At this scale, some important visual features that the model may have learned from high-resolution GBIF images might not be identifiable, leading to inaccurate predictions. This issue can be further amplified when these low-resolution crops are scaled up to the input model size. To compel the model to learn features for low-resolution inputs and avoid bias towards visual features only distinguishable in high-resolution images from other domains, we use a data augmentation technique we refer to as Mixed Resolution (MixRes) that simulates this low-resolution input during training (see Fig. D.4 in supplementary). This approach consists of randomly resizing the training input image to \\(25\\%\\) or \\(50\\%\\) of the model input size with a" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.517, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Insect Identification in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.147, + 0.785, + 0.192 + ], + "angle": 0, + "content": "probability \\(\\rho\\) for each case. In our experiments \\(\\rho\\) is set to 0.25, maintaining the original input size with probability 0.5. We employ MixRes in all experiments unless otherwise stated." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.198, + 0.785, + 0.288 + ], + "angle": 0, + "content": "In addition to MixRes, we also use a set of standard data augmentation operations: random resize crop with scale of 0.3 to 1.0 and an aspect ratio between \\(3/4\\) and \\(4/3\\), random horizontal flip with \\(50\\%\\) probability, and RandAugment [24] with \\(N = 2\\) operations and magnitude \\(M = 9\\). Finally, the image is normalized using the mean and standard deviation of ImageNet and scaled to the input network size." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.295, + 0.785, + 0.414 + ], + "angle": 0, + "content": "Implementation Details. We train all models with the AMI-GBIF training sets using the ImageNet-1K pretrained weights from the timm library [76] (for ConvNext and ViT we used ImageNet-22K weights). Each model is trained with a batch size of 128 and the AdamW optimizer [48] for 30 epochs. The initial learning rate is set to 0.001, with a weight decay of 1e-5. The learning rate is linearly warmed up from 0 to the initial value for 2 epochs and then decayed using cosine scheduling [36]. We also apply label smoothing [65] with a value of 0.1." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.421, + 0.785, + 0.496 + ], + "angle": 0, + "content": "Inference Settings. At test time, the insect crops from AMI-Traps are padded with black pixels to make them square, preventing the distortion of the shape of the insect before prediction. Next, the images are normalized with the mean and standard deviation of ImageNet and, finally, resized to the model input size. No padding is done for the test images from AMI-GBIF." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.502, + 0.785, + 0.638 + ], + "angle": 0, + "content": "Higher Taxonomic Rank Prediction. We consider the accuracy of our baseline models both with respect to the exact insect species predicted and with respect to higher level taxonomic groups such as genus and family. Indeed some images in our dataset cannot be labeled at the species level at all due to the quality or perspective of the image, and our annotators instead provided labels at the genus or family level. In our experiments, to make predictions at a higher taxonomic level, we simply sum together species-level predictions to get genus results, and similarly, the genus predictions are grouped together to get family predictions." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.645, + 0.785, + 0.839 + ], + "angle": 0, + "content": "Evaluation Metrics. During the binary classification step, it is important to achieve high precision by filtering out as many non-moths as possible to prevent the propagation of errors in the pipeline. At the same time, it is also important to assess recall because if too many moths are discarded at this step, it may impact the utility of the data for ecological studies. Therefore, for the binary classification task, along with accuracy, we also consider precision, recall, and F1-score as evaluation metrics. For the fine-grained classification task, we evaluate the model in terms of micro- and macro-averaged accuracy at the species, genus and family levels. Micro-averaging aggregates all class samples to calculate the accuracy while the macro-average is the average of per-class accuracy. Unless otherwise stated, the reported accuracy is micro-averaged. We conduct each experiment for three independent runs and report the mean and standard deviation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "A. Jain et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.496, + 0.163 + ], + "angle": 0, + "content": "6 Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.36 + ], + "angle": 0, + "content": "Binary Classification Results. This task involves distinguishing moths from non-moths. Here we train a ResNet50 model on the AMI-GBIF dataset and evaluate performance both on the AMI-GBIF held-out test set and on the entirety of the AMI-Traps dataset. On AMI-GBIF (in-distribution), the model achieves a test accuracy of \\(97.77\\% \\pm 0.02\\%\\). Table 2 shows the results for the AMI-Traps dataset in more detail. We observe that the model has a high recall but a lower precision, which means it is over-predicting the moth class. Also, the bigger the image crop, the better the accuracy of the insect classification; many non-moths where the model errs are extremely small and are therefore highly blurred in the trap images. Table E.3 in the supplementary illustrates the improvement in model performance when crops are square padded (our default at inference time) versus an ablation where padding is not performed." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.387, + 0.788, + 0.443 + ], + "angle": 0, + "content": "Table 2: Binary classification model evaluation results on the AMI-Traps dataset after training on AMI-GBIF, showing that recall is high throughout, while precision drops on very small insects, which are predominantly non-moths. Crops \\(>100\\mathrm{px}\\) or \\(>150\\mathrm{px}\\) represent that atleast one side of the box is greater than the threshold." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.455, + 0.787, + 0.51 + ], + "angle": 0, + "content": "
MothsNon-mothsTotalAccuracyPrecisionRecallF1-score
All crops14,10537,10551,12086.48±1.0068.31±1.7095.03±0.3979.48±1.27
Crops > 100px11,63016,53428,16488.39±0.9779.46±1.4396.99±0.2087.35±0.94
Crops > 150px6,4465,96112,40792.95±0.2289.83±0.4397.48±0.1293.70±0.30
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Fine-grained Classification Results. The fine-grained classification task aims to identify moths to the species level (and secondarily to genus or family level). In our experiments, we train models on AMI-GBIF data for each of the three regions of the AMI dataset, evaluating both on AMI-GBIF (in-distribution) and AMI-Traps (out-of-distribution) data (see Tab. 3). Not all moth species present in a region appear regularly at camera traps; hence, the list of species in AMI-Traps is a subset of species in AMI-GBIF for a given area (see Tab. 1). Therefore, we also evaluate the models on an in-distribution subset \\((\\mathrm{Test}\\ddagger)\\) from AMI-GBIF containing only the species present in AMI-Traps. Meanwhile, for AMI-Traps, there are 16 species not present in AMI-GBIF that were removed during our evaluation procedure (these represent unexpected species not included in the regional checklists used to define AMI-GBIF). Overall, the best model at the species level is ConvNeXt-B for both AMI-GBIF and AMI-traps. For C-America, annotations are mostly provided at the genus or family level, due to the challenging nature of visual identification of moths in this biodiversity-rich region. Furthermore, while the genus and family accuracies are similar between N-America and W-Europe, they are much lower for C-America. This can be attributed to the fact that there are 10 times fewer images in the training set for this region. Note that for genus and family evaluation, there are some cases where the accuracy is lower than the species level. This occurs because those" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.517, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Insect Identification in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 3: Accuracy results for the fine-grained classification task. Overall, ConvNeXt-B achieves the best results for both AMI-GBIF and AMI-traps, with ViT-B/16 achieving slightly better results for some cases at the genus and family levels. Test‡ represents a subset of the AMI-GBIF test set, containing only the species present in AMI-Traps. Species-level results on AMI-Traps for the C-America region are not available since most annotations were provided only at higher taxonomic levels." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.241, + 0.787, + 0.424 + ], + "angle": 0, + "content": "
RegionModelAMI-GBIFAMI-Traps
SpeciesSpeciesGenusFamily
Micro TestMicro Test‡Micro Top1Macro Top1Micro Top1Macro Top1Micro Top1Macro Top1
NE-AmericaResNet5087.56±0.1289.39±0.1471.86±1.0264.96±2.0080.07±0.4974.66±0.3590.44±0.4169.30±4.15
MBNetV3L83.73±0.1785.80±0.2967.41±0.7261.21±1.2074.83±1.3071.20±2.4789.31±0.6264.48±1.87
ConvNeXt-T89.81±0.0591.21±0.0674.89±0.0568.00±0.7683.26±1.2277.26±0.8590.61±0.5374.60±0.64
ConvNeXt-B90.45±0.0391.66±0.0376.75±1.5268.56±0.6285.46±0.8378.28±0.1491.07±0.2976.87±1.10
ViT-B/1686.39±0.0587.94±0.1077.43±0.7968.02±0.3485.57±0.9878.97±0.7291.45±0.3473.66±1.49
W-EuropeResNet5086.45±0.0890.25±0.0379.39±0.3972.84±0.2178.95±0.2673.22±0.6089.13±0.4275.24±1.85
MBNetV3L82.20±0.0586.47±0.1376.30±1.2970.35±0.9376.35±1.1972.88±1.2888.33±0.4772.21±0.38
ConvNeXt-T88.82±0.0392.05±0.0582.44±0.8974.59±0.6281.86±0.7974.15±0.5990.08±0.5974.58±2.93
ConvNeXt-B89.35±0.0192.48±0.0582.87±0.3976.41±0.5082.45±0.4475.98±1.0591.27±0.6079.39±2.70
ViT-B/1685.45±0.0389.22±0.1182.44±0.3675.25±0.9883.01±0.2675.96±0.3591.68±0.1775.81±0.79
C-AmericaResNet5087.46±0.2395.92±0.68--47.44±1.1166.59±0.6069.18±0.2442.40±0.31
MBNetV3L83.68±0.2492.89±0.27--41.03±2.2259.32±1.1868.34±1.6941.42±2.98
ConvNeXt-T87.65±0.1195.46±0.46--55.77±3.3370.68±1.5869.46±0.4245.35±0.67
ConvNeXt-B88.62±0.0996.22±0.26--51.92±3.8568.63±1.9472.24±1.4744.31±1.32
ViT-B/1679.80±0.2688.36±1.02--47.44±4.4464.70±4.0769.87±1.5142.00±3.60
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.784, + 0.514 + ], + "angle": 0, + "content": "higher taxonomic levels include crops that have labels only at higher levels. In the supplementary material, we include additional results for AMI-GBIF long-tail accuracies (Tab. E.4), a pre-training ablation study (Tab. E.5), and a study analyzing classification accuracy as a function of prediction confidence (Fig. E.5)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.755 + ], + "angle": 0, + "content": "Data augmentation approaches. To assess the effectiveness of the mixed resolution data augmentation technique, we conduct an ablation study using ResNet50 for the NE-America region. As the effects of MixRes could already be covered by other data augmentation operations used in our preprocessing, such as random resize crop or one of the operations of RandAugment, we analyze multiple combinations of these operations. Additionally, we evaluate a higher input resolution, using the original 224 input size of ResNet50, to determine whether this leads to better predictions. Table 4 shows the results. While applying MixRes results in a slight decrease in accuracy within the in-distribution AMI-GBIF test set, for the in-the-wild AMI-Traps test set there is a substantial improvement, ranging between \\(4\\%\\) and \\(7\\%\\) when compared to models not trained with MixRes. Another important observation is that, although RandAugment and random resize together contribute to an overall improvement of around \\(16\\%\\), MixRes remains additionally beneficial. Finally, despite an increase of more than \\(4\\%\\) in accuracy within the original domain when using a higher resolution, there is a marginal decrease in the accuracy of the in-the-wild results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.778, + 0.36, + 0.794 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Many ecological applications require species-level image classification to meet their goals. Despite advances in computer vision, there remain challenges in" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "A. Jain et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 4: Ablation study on data augmentation techniques using a ResNet50 for the NE-America region evaluated on micro-averaged top-1 species accuracy. \"RandAug\" denotes the combination of the RandAugment and random resize crop data augmentation techniques. AMI-GBIF† test set contains only the species present in AMI-Traps." + }, + { + "type": "table", + "bbox": [ + 0.315, + 0.226, + 0.691, + 0.313 + ], + "angle": 0, + "content": "
ResolutionRandAugMixResAMI-GBIF‡AMI-Traps
128××89.18±0.0851.46±2.73
128✓×89.71±0.1667.88±1.27
128×✓87.41±0.0658.41±0.30
128✓✓89.39±0.1471.86±1.02
224✓✓92.97±0.1768.41±1.06
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.545 + ], + "angle": 0, + "content": "implementing fine-grained classification that is effective in real-world scenarios. Here we outline the challenges of insect identification in the wild and propose a benchmark dataset to tackle them. We present AMI-GBIF, a dataset with 2.5 million labeled photographs of 5,000 moth species. We also introduce AMI-Traps, a dataset of images captured by automated insect monitoring systems across five countries, labeled by regional experts. We evaluate multiple classification models both on the AMI-GBIF dataset and in out-of-distribution generalization to AMI-Traps, with ConvNeXt-B generally performing the best. We also evaluate a range of data augmentation techniques, finding that a combination of simple approaches, notably mixed-resolution augmentation, considerably improves generalization under field conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.788, + 0.709 + ], + "angle": 0, + "content": "Several limitations of this work should be noted. First, while annotating AMI-Traps already occupied a considerable time for expert entomologists, further expanding it to a greater number of species and geographic regions would be helpful in further stress-testing field deployments of fine-grained classification algorithms. Second, while we focus on camera traps for nocturnal insects, it would be useful also to consider diurnal insect camera traps, for example those aimed at pollinators, which pose some additional challenges, such as multiple perspectives and a diversity of backgrounds. We hope furthermore that future works will consider more sophisticated domain adaptation approaches building on the baselines we present here." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This work represents a step towards computer vision algorithms for species recognition that are effective in the ultra-challenging regime of insect identification and work effectively in field conditions. The approach we propose should make it possible in particular to generalize algorithms for insect camera traps rapidly and cost-effectively across different hardware setups and regions of interest. We hope that the datasets, challenges and strategies presented here will propel further research in computer vision for insects and help accelerate initiatives to understand and protect our planet's biodiversity." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.516, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Insect Identification in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.784, + 0.208 + ], + "angle": 0, + "content": "We thank Annie-Shan Morin, Anna Viklund, and Jamie Alison for their helpful input and support to the project." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.787, + 0.403 + ], + "angle": 0, + "content": "This research was generously funded by Mila Quebec AI Institute and Ministère de l'Économie, de l'Innovation et de l'Energie. The work was additionally supported in part by the Canada CIFAR AI Chairs program (NSERC 585136), the Global Climate Center on AI and Biodiversity Change (ABC) (NSF OISE2330423), the Montreal Insectarium, the Coordenação de Aperfeçoamento de Pessoal de Nível Superior - Brasil (CAPES-PROEX) - Finance Code 001, the Amazonas State Research Support Foundation - FAPEAM - through the POSGRAD project 2024/2025, the AMBER project funded by the abrdn Charitable Foundation, the Aage V. Jensen Naturfond grant no. N2024-0013, the European Union's Horizon Europe Research and Innovation programme under grant agreement no. 101060639 (MAMBO), and a Discovery Grant by NSERC. The computing resources were provided by Mila Quebec AI Institute, the Digital Research Alliance of Canada, and NVIDIA Corporation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.787, + 0.48 + ], + "angle": 0, + "content": "A.J., F.C., and M.J.B. led dataset creation and experiments. J.S.C. and L.P. contributed to machine learning pipeline development. N.P., F.H, J.R., M.B., M.S., J.F., A.A., Y.L., E.N., F.P.P., A.C.Z., J.A.R.S., and J.G. performed insect identification for the AMI-Traps dataset. T.A, K.B., A.G.S., M.B., Y.B., K.P.M, D. Roy, T.T.H., M.L., and D. Rolnick guided and advised the project." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.503, + 0.325, + 0.518 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.787, + 0.576 + ], + "angle": 0, + "content": "1. BugGuide.Net: Identification, Images, & Information For Insects, Spiders & Their Kin For the United States & Canada, https://bugguide.net/ accessed Feb 12 2024" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.787, + 0.603 + ], + "angle": 0, + "content": "2. GBIF Backbone Taxonomy, https://doi.org/10.15468/39omei accessed Nov 13 2023" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.605, + 0.787, + 0.631 + ], + "angle": 0, + "content": "3. GBIF Occurrence Download, https://doi.org/10.15468/dl.4ua6wg accessed Oct 19 2023" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.614, + 0.646 + ], + "angle": 0, + "content": "4. iNaturalist, www.inaturalist.org, accessed Feb 9 2024" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.687, + 0.66 + ], + "angle": 0, + "content": "5. Observation.org, https://observation.org, accessed Feb 12 2024" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.787, + 0.686 + ], + "angle": 0, + "content": "6. WebDataset, https://github.com/webdataset/webdataset, accessed on April 2 2024" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.759, + 0.702 + ], + "angle": 0, + "content": "7. What is GBIF?, https://www.gbif.org/what-is-gbif accessed Sep 28 2023" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.787, + 0.743 + ], + "angle": 0, + "content": "8. Alison, J., Alexander, J.M., Diaz Zeugin, N., Dupont, Y.L., Iseli, E., Mann, H.M., Hoye, T.T.: Moths complement bumblebee pollination of red clover: a case for day-and-night insect surveillance. Biology Letters 18(7), 20220187 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.787, + 0.785 + ], + "angle": 0, + "content": "9. Anton, V., Hartley, S., Geldenhuis, A., Wittmer, H.U.: Monitoring the mammalian fauna of urban areas using remote cameras and citizen science. Journal of Urban Ecology 4(1), juy002 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "10. Badgujar, C.M., Armstrong, P.R., Gerken, A.R., Pordesimo, L.O., Campbell, J.F.: Real-time stored product insect detection and identification using deep learning: System integration and extensibility to mobile platforms. Journal of Stored Products Research 104, 102196 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.535, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "A. Jain et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "11. Beery, S.: Scaling biodiversity monitoring for the data age. XRDS: Crossroads, The ACM Magazine for Students 27(4), 14-18 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "12. Beery, S., Agarwal, A., Cole, E., Birodkar, V.: The iwildcam 2021 competition dataset. arXiv preprint arXiv:2105.03494 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.206, + 0.785, + 0.233 + ], + "angle": 0, + "content": "13. Beery, S., Morris, D., Yang, S.: Efficient pipeline for camera trap image review. arXiv preprint arXiv:1907.06772 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.235, + 0.785, + 0.261 + ], + "angle": 0, + "content": "14. Beery, S., Van Horn, G., Perona, P.: Recognition in terra incognita. In: Proceedings of the European conference on computer vision (ECCV). pp. 456-473 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.263, + 0.785, + 0.331 + ], + "angle": 0, + "content": "15. Beery, S., Wu, G., Edwards, T., Pavetic, F., Majewski, B., Mukherjee, S., Chan, S., Morgan, J., Rathod, V., Huang, J.: The auto arborist dataset: A large-scale benchmark for multiview urban forest monitoring under domain shift. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21294-21307 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.333, + 0.785, + 0.387 + ], + "angle": 0, + "content": "16. Berg, T., Liu, J., Woo Lee, S., Alexander, M.L., Jacobs, D.W., Belhumeur, P.N.: Birdsnap: Large-scale fine-grained visual categorization of birds. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2011-2018 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.389, + 0.785, + 0.431 + ], + "angle": 0, + "content": "17. Bjerge, K., Alison, J., Dyrmann, M., Frigaard, C.E., Mann, H.M., Hoye, T.T.: Accurate detection and identification of insects from camera trap images with deep learning. PLOS Sustainability and Transformation 2(3), e0000051 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.433, + 0.785, + 0.459 + ], + "angle": 0, + "content": "18. Bjerge, K., Frigaard, C.E., Karstoft, H.: Object detection of small insects in timelapse camera recordings. Sensors 23(16), 7242 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.461, + 0.785, + 0.5 + ], + "angle": 0, + "content": "19. Bjerge, K., Geissmann, Q., Alison, J., Mann, H.M., Hoye, T.T., Dyrmann, M., Karstoft, H.: Hierarchical classification of insects with multitask learning and anomaly detection. Ecological Informatics 77, 102278 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.503, + 0.785, + 0.543 + ], + "angle": 0, + "content": "20. Bjerge, K., Mann, H.M., Hoye, T.T.: Real-time insect tracking and monitoring with computer vision and deep learning. Remote Sensing in Ecology and Conservation 8(3), 315-327 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.546, + 0.785, + 0.586 + ], + "angle": 0, + "content": "21. Bjerge, K., Nielsen, J.B., Sepstrup, M.V., Helsing-Nielsen, F., Hoye, T.T.: An automated light trap to monitor moths (lepidoptera) using computer vision-based tracking and deep learning. Sensors 21(2), 343 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.588, + 0.785, + 0.642 + ], + "angle": 0, + "content": "22. Boom, B.J., He, J., Palazzo, S., Huang, P.X., Bayan, C., Chou, H.M., Lin, F.P., Spampinato, C., Fisher, R.B.: A research tool for long-term and continuous analysis of fish assemblage in coral-reefs using underwater camera footage. Ecological Informatics 23, 83-97 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.645, + 0.785, + 0.685 + ], + "angle": 0, + "content": "23. Borowiec, M.L., Dikow, R.B., Frandsen, P.B., McKeeken, A., Valentini, G., White, A.E.: Deep learning as a tool for ecology and evolution. Methods in Ecology and Evolution 13(8), 1640-1660 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.687, + 0.785, + 0.741 + ], + "angle": 0, + "content": "24. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 702-703 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.743, + 0.785, + 0.797 + ], + "angle": 0, + "content": "25. Cutter, G., Stierhoff, K., Zeng, J.: Automated detection of rockfish in unconstrained underwater videos using haar cascades and a new image dataset: Labeled fishes in the wild. In: 2015 IEEE Winter Applications and Computer Vision Workshops. pp. 57-62. IEEE (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "26. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.516, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Insect Identification in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "27. Didham, R.K., Basset, Y., Collins, C.M., Leather, S.R., Littlewood, N.A., Menz, M.H., Müller, J., Packer, L., Saunders, M.E., Schonrogge, K., et al.: Interpreting insect declines: seven challenges and a way forward. Insect Conservation and Diversity 13(2), 103-114 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.788, + 0.285 + ], + "angle": 0, + "content": "28. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net (2021), https://openreview.net/forum?id=YicbFdNTTy" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.287, + 0.788, + 0.327 + ], + "angle": 0, + "content": "29. Dubey, A., Ramanathan, V., Pentland, A., Mahajan, D.: Adaptive methods for real-world domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14340-14349 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.328, + 0.787, + 0.369 + ], + "angle": 0, + "content": "30. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision 88, 303-338 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.37, + 0.787, + 0.424 + ], + "angle": 0, + "content": "31. Fennell, M., Beirne, C., Burton, A.C.: Use of object detection in camera trap image identification: Assessing a method to rapidly and accurately classify human and animal detections for research and application in recreation ecology. Global Ecology and Conservation 35, e02104 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.425, + 0.787, + 0.466 + ], + "angle": 0, + "content": "32. Geissmann, Q., Abram, P.K., Wu, D., Haney, C.H., Carrillo, J.: Sticky pi is a high-frequency smart trap that enables the study of insect circadian activity under natural conditions. PLoS Biology 20(7), e3001689 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.467, + 0.787, + 0.522 + ], + "angle": 0, + "content": "33. Gharaee, Z., Gong, Z., Pellegrino, N., Zarubiieva, I., Haurum, J.B., Lowe, S.C., McKeown, J.T., Ho, C.C., McLeod, J., Wei, Y.Y.C., et al.: A step towards worldwide biodiversity assessment: The bioscan-1m insect dataset. arXiv preprint arXiv:2307.10455 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.523, + 0.787, + 0.549 + ], + "angle": 0, + "content": "34. Gulrajani, I., Lopez-Paz, D.: In search of lost domain generalization. arXiv preprint arXiv:2007.01434 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.55, + 0.787, + 0.591 + ], + "angle": 0, + "content": "35. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.591, + 0.787, + 0.633 + ], + "angle": 0, + "content": "36. He, T., Zhang, Z., Zhang, H., Zhang, Z., Xie, J., Li, M.: Bag of tricks for image classification with convolutional neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 558-567 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.633, + 0.787, + 0.674 + ], + "angle": 0, + "content": "37. Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al.: Searching for mobilenetv3. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1314-1324 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.675, + 0.787, + 0.729 + ], + "angle": 0, + "content": "38. Hoye, T.T., Arje, J., Bjerge, K., Hansen, O.L., Iosifidis, A., Leese, F., Mann, H.M., Meissner, K., Melvad, C., Raitoharju, J.: Deep learning and computer vision will transform entomology. Proceedings of the National Academy of Sciences 118(2), e2002545117 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.729, + 0.787, + 0.785 + ], + "angle": 0, + "content": "39. Katija, K., Orenstein, E., Schlining, B., Lundsten, L., Barnard, K., Sainz, G., Boulais, O., Cromwell, M., Butler, E., Woodward, B., et al.: Fathomnet: A global image database for enabling artificial intelligence in the ocean. Scientific reports 12(1), 15914 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "40. Kay, J., Kulits, P., Stathatos, S., Deng, S., Young, E., Beery, S., Van Horn, G., Perona, P.: The caltech fish counting dataset: A benchmark for multiple-object tracking and counting. In: European Conference on Computer Vision. pp. 290-311. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "A. Jain et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "41. van Klink, R., August, T., Bas, Y., Bodesheim, P., Bonn, A., Fossoy, F., Høye, T.T., Jongejans, E., Menz, M.H., Miraldo, A., et al.: Emerging technologies revolutionise insect ecology and monitoring. Trends in ecology & evolution (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.258 + ], + "angle": 0, + "content": "42. Koh, P.W., Sagawa, S., Marklund, H., Xie, S.M., Zhang, M., Balsubramani, A., Hu, W., Yasunaga, M., Phillips, R.L., Gao, I., Lee, T., David, E., Stavness, I., Guo, W., Earnshaw, B.A., Haque, I.S., Beery, S., Leskovec, J., Kundaje, A., Pierson, E., Levine, S., Finn, C., Liang, P.: WILDS: A benchmark of in-the-wild distribution shifts. In: International Conference on Machine Learning (ICML) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.258, + 0.785, + 0.298 + ], + "angle": 0, + "content": "43. Korsch, D., Bodesheim, P., Denzler, J.: Deep learning pipeline for automated visual moth monitoring: insect localization and species classification. arXiv preprint arXiv:2307.15427 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.352 + ], + "angle": 0, + "content": "44. Krasin, I., Duerig, T., Alldrin, N., Ferrari, V., Abu-El-Haija, S., Kuznetsova, A., Rom, H., Uijlings, J., Popov, S., Veit, A., et al.: Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://github.com/openimages2(3), 18 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.421 + ], + "angle": 0, + "content": "45. Kumar, N., Belhumeur, P.N., Biswas, A., Jacobs, D.W., Kress, W.J., Lopez, I.C., Soares, J.V.: Leafsnap: A computer vision system for automatic plant species identification. In: Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part II 12. pp. 502-516. Springer (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.421, + 0.785, + 0.476 + ], + "angle": 0, + "content": "46. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.476, + 0.785, + 0.516 + ], + "angle": 0, + "content": "47. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.785, + 0.543 + ], + "angle": 0, + "content": "48. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.543, + 0.785, + 0.583 + ], + "angle": 0, + "content": "49. Luccioni, A.S., Rolnick, D.: Bugs in the data: How imagenet misrepresents biodiversity. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 14382-14390 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.623 + ], + "angle": 0, + "content": "50. Ming, Y., Yin, H., Li, Y.: On the impact of spurious correlation for out-of-distribution detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 10051-10059 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.785, + 0.651 + ], + "angle": 0, + "content": "51. Mora, C., Tittensor, D.P., Adl, S., Simpson, A.G., Worm, B.: How many species are there on earth and in the ocean? PLoS biology 9(8), e1001127 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.785, + 0.704 + ], + "angle": 0, + "content": "52. Nawoya, S., Ssemakula, F., Akol, R., Geissmann, Q., Karstoft, H., Bjerge, K., Mwikirize, C., Katumba, A., Gebreyesus, G.: Computer vision and deep learning in insects for food and feed production: A review. Computers and Electronics in Agriculture 216, 108503 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.745 + ], + "angle": 0, + "content": "53. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian conference on computer vision, graphics & image processing. pp. 722-729. IEEE (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.786 + ], + "angle": 0, + "content": "54. Norouzzadeh, M.S., Morris, D., Beery, S., Joshi, N., Jojic, N., Clune, J.: A deep active learning system for species identification and counting in camera trap images. Methods in ecology and evolution 12(1), 150-161 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "55. Oliver, R.Y., Iannarilli, F., Ahumada, J., Fegraus, E., Flores, N., Kays, R., Birch, T., Ranipeta, A., Rogan, M.S., Sica, Y.V., et al.: Camera trapping expands the view into global biodiversity and its change. Philosophical Transactions of the Royal Society B 378(1881), 20220232 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.516, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Insect Identification in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "56. Ouaknine, A., Kattenborn, T., Laliberte, E., Rolnick, D.: Openforest: A data catalogue for machine learning in forest monitoring. arXiv preprint arXiv:2311.00277 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.245 + ], + "angle": 0, + "content": "57. Rolnick, D., Donti, P.L., Kaack, L.H., Kochanski, K., Lacoste, A., Sankaran, K., Ross, A.S., Milojevic-Dupont, N., Jaques, N., Waldman-Brown, A., et al.: Tackling climate change with machine learning. ACM Computing Surveys (CSUR) 55(2), 1-96 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.788, + 0.286 + ], + "angle": 0, + "content": "58. Roy, D., et al.: Towards a standardised framework for ai-assisted, image-based monitoring of nocturnal insects. Philosophical Transactions of the Royal Society B (2024). https://doi.org/10.1098/rstb.2023.0108, (Forthcoming)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.288, + 0.788, + 0.342 + ], + "angle": 0, + "content": "59. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) 115(3), 211-252 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.343, + 0.788, + 0.397 + ], + "angle": 0, + "content": "60. Seibold, S., Gossner, M.M., Simons, N.K., Bluthgen, N., Müller, J., Ambarh, D., Ammer, C., Bauhus, J., Fischer, M., Habel, J.C., et al.: Arthropod decline in grasslands and forests is associated with landscape-level drivers. Nature 574(7780), 671-674 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.398, + 0.788, + 0.439 + ], + "angle": 0, + "content": "61. Stark, T., Stefan, V., Wurm, M., Spanier, R., Taubenbock, H., Knight, T.M.: Yolo object detection models can locate and classify broad groups of flower-visiting arthropods in images. Scientific Reports 13(1), 16364 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.44, + 0.788, + 0.467 + ], + "angle": 0, + "content": "62. Stork, N.E., et al.: How many species of insects and other terrestrial arthropods are there on earth. Annual review of entomology 63(1), 31-45 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.468, + 0.788, + 0.494 + ], + "angle": 0, + "content": "63. Suto, J.: Codling moth monitoring with camera-equipped automated traps: A review. Agriculture 12(10), 1721 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.496, + 0.788, + 0.536 + ], + "angle": 0, + "content": "64. Swanson, A., Kosmala, M., Lintott, C., Simpson, R., Smith, A., Packer, C.: Snapshot serengeti, high-frequency annotated camera trap images of 40 mammalian species in an african savanna. Scientific data 2(1), 1-14 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.537, + 0.788, + 0.577 + ], + "angle": 0, + "content": "65. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818-2826 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.578, + 0.788, + 0.605 + ], + "angle": 0, + "content": "66. Teixeira, A.C., Ribeiro, J., Morais, R., Sousa, J.J., Cunha, A.: A systematic review on automatic insect detection using deep learning. Agriculture 13(3), 713 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.606, + 0.788, + 0.633 + ], + "angle": 0, + "content": "67. Theivaprakasham, H.: Identification of indian butterflies using deep convolutional neural network. Journal of Asia-Pacific Entomology 24(1), 329-340 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.634, + 0.788, + 0.674 + ], + "angle": 0, + "content": "68. Tuia, D., Kellenberger, B., Beery, S., Costelloe, B.R., Zuffi, S., Risse, B., Mathis, A., Mathis, M.W., van Langevelde, F., Burghardt, T., et al.: Perspectives in machine learning for wildlife conservation. Nature communications 13(1), 792 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.675, + 0.788, + 0.73 + ], + "angle": 0, + "content": "69. Van Horn, G., Branson, S., Farrell, R., Haber, S., Barry, J., Ipeirotis, P., Perona, P., Belongie, S.: Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 595-604 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.731, + 0.788, + 0.785 + ], + "angle": 0, + "content": "70. Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist species classification and detection dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8769-8778 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.786, + 0.788, + 0.813 + ], + "angle": 0, + "content": "71. Van Horn, G., Perona, P.: The devil is in the tails: Fine-grained classification in the wild. arXiv preprint arXiv:1709.01450 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.814, + 0.788, + 0.841 + ], + "angle": 0, + "content": "72. Wagner, D.L.: Insect declines in the anthropocene. Annual review of entomology 65, 457-480 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.361, + 0.127 + ], + "angle": 0, + "content": "A. Jain et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "73. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset. Tech. rep. (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.217 + ], + "angle": 0, + "content": "74. Wei, X.S., Song, Y.Z., Mac Aodha, O., Wu, J., Peng, Y., Tang, J., Yang, J., Belongie, S.: Fine-grained image analysis with deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence 44(12), 8927-8948 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "75. Weinstein, B.G.: A computer vision for animal ecology. Journal of Animal Ecology 87(3), 533-545 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.272 + ], + "angle": 0, + "content": "76. Wightman, R.: Pytorch image models. https://github.com/rwrightman/pytorch-image-models (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "77. Zhou, K., Liu, Z., Qiao, Y., Xiang, T., Loy, C.C.: Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.3 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Insect Identification in the Wild_ The AMI Dataset/deef9d64-c259-4177-ab17-679ea66a461f_origin.pdf b/2024/Insect Identification in the Wild_ The AMI Dataset/deef9d64-c259-4177-ab17-679ea66a461f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e0e0e4d26235a24d105151e72ca05d95ea019a9b --- /dev/null +++ b/2024/Insect Identification in the Wild_ The AMI Dataset/deef9d64-c259-4177-ab17-679ea66a461f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:696665fb5c32235dec752e15b2c967daf1b82ca7aaf7483ace01a51153af9839 +size 4032230 diff --git a/2024/Insect Identification in the Wild_ The AMI Dataset/full.md b/2024/Insect Identification in the Wild_ The AMI Dataset/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d0bdfb9b430e5fc168a9efccf9f86f392761b71c --- /dev/null +++ b/2024/Insect Identification in the Wild_ The AMI Dataset/full.md @@ -0,0 +1,288 @@ +# Insect Identification in the Wild: The AMI Dataset + +Aditya Jain1*, Fagner Cunha2*, Michael James Bunsen1*, Juan Sebastián Canas1, Léonard Pasi1, Nathan Pinoy3, Flemming Helsing3, JoAnne Russo4, Marc Botham5, Michael Sabourin6, Jonathan Fréchette12, Alexandre Anctil7, Yacksecari Lopez8, Eduardo Navarro8, Filonila Perez Pimentel8, Ana Cecilia Zamora8, José Alejandro Ramirez Silva8, Jonathan Gagnon12, Tom August5, Kim Bjerge3, Alba Gomez Segura5, Marc Bélisle9, Yves Basset8, Kent P. McFarland4, David Roy5, Toke Thomas Hoye3, Maxim Larrivée10, and David Rolnick1,11 + +1 Mila - Quebec AI Institute + +2 Federal University of Amazonas + +3 Aarhus University + +4 Vermont Center for Ecostudies + +UK Centre for Ecology & Hydrology + +McGuire Center + +7 Centre de données sur le patrimoine naturel du Québec + +Smithsonian Tropical Research Institute + +9 Université de Sherbrooke + +10 Montreal Insectarium + +11 McGill University + +$^{12}$ Independent Researcher + +Abstract. Insects represent half of all global biodiversity, yet many of the world's insects are disappearing, with severe implications for ecosystems and agriculture. Despite this crisis, data on insect diversity and abundance remain woefully inadequate, due to the scarcity of human experts and the lack of scalable tools for monitoring. Ecologists have started to adopt camera traps to record and study insects, and have proposed computer vision algorithms as an answer for scalable data processing. However, insect monitoring in the wild poses unique challenges that have not yet been addressed within computer vision, including the combination of long-tailed data, extremely similar classes, and significant distribution shifts. We provide the first large-scale machine learning benchmarks for fine-grained insect recognition, designed to match real-world tasks faced by ecologists. Our contributions include a curated dataset of images from citizen science platforms and museums, and an expert-annotated dataset drawn from automated camera traps across multiple continents, designed to test out-of-distribution generalization under field conditions. We train and evaluate a variety of baseline algorithms and introduce a combination of data augmentation techniques that enhance generalization across geographies and hardware setups. The dataset is made publicly available1. + +# 1 Introduction + +Computer vision algorithms for species identification have become important tools in studying and protecting the Earth's wildlife. Biodiversity applications have also been beneficial to computer vision, serving as a fruitful source of new benchmarks and algorithms [11,23,68,75]. However, research in computer vision for biodiversity has often focused on large animals, not reflecting the full set of challenges faced in ecology and conservation. + +There are over a million known species of insects on Earth, representing nearly one half of all known organisms (by comparison, there are only 10,000 species of birds, and even fewer mammals) [51,62]. Insects play essential roles in nearly all terrestrial ecosystems, as well as providing ecosystem services such as pollination that are critical for humanity [60], yet a looming "insect apocalyptic" threatens both their abundance and diversity [72]. Despite the crisis, data on insect populations is shockingly sparse [27], as a result of less resources, a plethora of species, and the high level of expertise typically required to study insects. Amid this data gap, computer vision tools for automated monitoring have the potential to radically reshape the study and conservation of insects. In particular, recent breakthroughs in camera traps for insects have made it possible to automatically photograph large numbers of individuals in the wild [38,41,58], opening up the the potential for image-processing algorithms to parse this raw data. + +Species recognition from camera traps is a well-studied problem in the computer vision community [55], with common challenges including poor lighting, occlusion of individuals, camouflage, and blur [14]. However, working with insects presents distinct challenges from those faced in traditional camera traps, which are often designed for large animals. For example, while traditional camera trap images may only rarely contain a target species, almost every image from an insect camera trap includes insects – often hundreds of them. Where traditional camera traps may have a short list of species of interest, insect monitoring requires considering thousands of species at any given location, with many of them being extremely similar, and most with very little labeled data. + +While recent authors have proposed the use of computer vision tools for detecting and identifying insects in camera trap data [8, 17, 20, 32, 43, 63], such studies have involved small datasets with only a few species represented. There has to date been no work assessing the performance of algorithms in identifying large number of insect species, or addressing the challenges of generalization across in-the-wild deployments. Our work here offers the following contributions: + +- We introduce the AMI (Automated Monitoring of Insects) dataset, consisting of two parts: 1) AMI-GBIF, a dataset of $\sim 2.5\mathrm{M}$ human-captured images curated from citizen science platforms and museum collections, 2) AMI-Traps, an expert-annotated dataset of 2,893 insect camera trap images (representing 52,948 labeled insects) collected from a global network of automated camera traps, designed to test in-the-wild performance (Fig. 1). + +![](images/00003212e95be6e5dd272b91109bf82f0a6193888b5755b4b1ef396d96a576bc.jpg) +Fig. 1: Illustration of the difference between insect species identification from human-captured images and in the wild. On left are shown images of three moth species from our AMI-GBIF dataset, curated from citizen science data and museum collections. On right are shown images of the same species (along with other insects) within a photograph taken by an automated camera trap in the wild from our AMI-Traps dataset. One of the challenges of the AMI dataset is generalizing from AMI-GBIF to AMI-Traps without additional labeled data. + +- We frame the problem of practical insect monitoring in the field as a set of computer vision tasks, entailing very fine-grained classification from long-tailed data, with the additional challenge of generalization across multiple kinds of imagery. +- We test a variety of strong baseline models for the proposed tasks, including a combination of data augmentation approaches that significantly improve performance under field conditions. + +# 2 Related Work + +Biodiversity Image Datasets. Advancements in computer vision have been closely intertwined with the creation of curated datasets such as ImageNet [26, 59], Pascal VOC [30], Open Images [44], and MS-COCO [46]. In recent years, the computer vision and machine learning communities have increasingly directed their focus towards addressing pressing human threats, including biodiversity loss [68] and climate change [57]. These domains not only hold significant potential for societal impact but also offer fertile ground for exploring intriguing machine learning questions such as domain generalization [77], long-tailed distributions [71], and fine-grained classification [74]. Such challenges have found especially fruitful testbeds in the classification of species, where datasets such as iNaturalist [70] have proven useful for both the computer vision and ecology communities. While some challenges have considered plant identification [15, 45, 53, 56], the majority have focused on animals, including datasets focused on birds [16, 69, 73], mammals [9, 14, 64], and fish [22, 25, 39, 40]. Such datasets have often leveraged concurrent advancements in sensor technology, including in satellite imagery and drones. In particular, camera traps have been widely adopted by biologists as an in situ and cost-effective means of wildlife + +![](images/c8422f237016fb6a7084b360cb3732bc8a7d1c35b2e6e6375849c53ca34a5014.jpg) + +![](images/c8d7dee1236803f8056af0ee0b3845fd33dd3d277603922e0f9d64c8b7e2393e.jpg) +(a) Incomplete + +![](images/ff4a585ec332620364206eefd14b9081a9fe36e330ff7b5f7ca3df710832ddc0.jpg) +(b) Overlapping +(c) Motion + +![](images/d9d2b1c95c7f02a1d8fca63c3d025a0f755d16b8d30d32f621e88939ab03b5e2.jpg) + +![](images/2107424f9c29b4c8e3cc810131011b20ef05a0ea9a3822c40be081770f90c764.jpg) +Fig. 2: Examples of challenges in the AMI-Traps dataset. (a) Incomplete: In some occasions, the body of an insect is incomplete due to its position in the automatically taken image. (b) Overlapping: Sometimes insects occlude each other. (c) Motion: Moving insects can be blurred and hard to classify. (d) Lighting: The lighting and camera exposure of the images is variable and can lead to poor contrast that masks some or all detail. (e) Resolution: Some insects are very small, leading to low-resolution images. (f) Perspective: While insects are generally positioned at a fixed distance, sometimes insects appear in the air or perched on the camera itself. Additionally, unexpected objects such as spider webs, dirt, and vegetation can be captured by the cameras and may be confused with objects of interest. + +![](images/d25edc3113fff52e6dd677f6c5d7786b8f7aa6715e2d2171963f3f77c8b59014.jpg) + +![](images/afdce32fb78a285bbb6342f87374635c6e7ed62ef5730bc0ab098be0a2df700b.jpg) + +![](images/4b7b837aeb3abd3c022276d253d1d6aa0761af1fe5cfe8d7f98afd06b4869868.jpg) + +![](images/bde0083e1d7e69c0259918e0346dc0eff44e7c7ad1adc146b61a2226a13e5f33.jpg) +(d) Lighting +(e) Resolution + +![](images/65f593ddb7b468337474b2dd8a2f838564d7358c5ba6be8132ef0e21091b0f05.jpg) +(f) Perspective + +monitoring [55], and have in turn presented fascinating challenges for the computer vision community [14], including in classification [12, 29], object detection [13,31], out-of-distribution generalization [34,42,50], and active learning [54]. Such algorithms have to date, however, been designed mostly for large animals such as mammals. + +Insect Image Datasets. As algorithms and datasets for species recognition have proliferated, insects have traditionally been underrepresented. Within ImageNet1k [59], only 27 classes depicted insects, as compared to 57 for birds and 93 classes for wild mammals [49]. Some popular fine-grained classification datasets such as NABirds and CUB have focused exclusively on birds [69, 73], while camera trap datasets such as iWildCam have been focused on mammals [12]. The iNat2017 dataset [70] included 1,021 classes of insects out of 5,089 total classes in the dataset. By largely focusing on other taxonomic groups, prior works have neglected the challenges posed by insect data, such as extremely similar classes (reflecting far greater diversity than any other group of organisms), small size, hyperabundance, etc. In recent years, however, some works have introduced insect-focused benchmark datasets. The newly released BIOSCAN-1M dataset includes images and DNA sequences, with a focus on fly identification [33]. Despite the very large size of the dataset (1.1 million datapoints), the number of object classes involved is relatively small, with the focus being to distinguish between 40 taxonomic families of flies, reflecting the challenging nature of insect identification. Other recent datasets for computer vision include one focused on 8 classes of pollinators across Europe [61], and the Indian Butterflies dataset, + +encompassing 30k images of butterflies from India across 315 species [67]. A variety of publications have used computer vision algorithms for detection of pest insects in specific agricultural applications [10, 52, 66]. + +Insect Camera Traps. The use of camera traps for studying insect biodiversity is relatively recent. Geissmann et al. [32] introduced the Sticky Pi device, which uses sticky cards to capture (and kill) insects, which are then automatically photographed. Bjerge et al. [21] introduced an automated camera trap using UV light to attract and photograph moths and other insects (without killing them), which was built on by subsequent works [18, 20]. A similar approach focused on diurnal pollinators via a camera trap installed above flowers [17, 19]. These works also propose deep learning methods for processing the resulting images, considering both object detection and classification of a small number of insect categories. The lack of fine-grained classification is largely due to the time required to annotate images from the traps, making it difficult to obtain a large enough labeled dataset from trap data alone. + +# 3 Benchmark Design + +In this paper, we propose a benchmark dataset for insect species identification that addresses the challenges faced in real-world use cases. These include both (1) challenges that distinguish insect identification from other kinds of species identification, notably the massive diversity of often-similar species, and (2) the particular challenges associated with deploying these algorithms "in the wild". + +To address the first family of challenges, we introduce a dataset for fine-grained insect classification that (to our knowledge) is by far the largest yet released, with $\sim 2.5\mathrm{M}$ images across over 5,000 classes. We concentrate on moths, which are readily drawn to UV camera traps and can often be visually identified to species (unlike some other insects) [21]. + +The second family of challenges is more complex - involving the gap between field conditions and the high-quality images typically used for training species identification algorithms. Figure 2 illustrates some of the challenges that computer vision algorithms face on data from automated insect camera traps, in contrast to images taken by humans. Insects may occlude each other or be partially "out of frame". Motion blur and lighting issues may occur, along with objects in unexpected positions. The resolution of images may be very low, owing to the small size of insects at a stereotyped distance from the camera. Also, insect camera traps typically use webcams and PCB camera modules due to their cost and software integration capabilities, and the resulting images are often both structurally different and overall lower quality than those taken by human photographers. + +As described in the preceding section, previous authors working with insect camera traps have addressed the problem of challenging camera trap data by training algorithms on annotated data from the particular camera trap in question. However, manual annotation is extremely time-intensive, since it must be performed by expert entomologists, of whom there are very few in any given + +region. Insect species vary greatly across regions, meaning that fine-grained annotations in one location are generally not useful in another. Furthermore, the hardware used in insect camera traps also varies greatly (with considerable differences in lighting and background, as well as image sharpness, contrast, noise reduction settings, and compression artifacts, which are exacerbated on small scale crops), meaning that even two datasets from the same region may not be comparable. + +In this paper, we remove the problem of annotation by solving a more challenging computer vision problem. Namely, we propose to train algorithms on abundant labeled images of insects taken by human photographers, then use these algorithms on out-of-distribution camera trap images. Our benchmark dataset is designed to test this functionality, by including both a large training set of curated photographs taken by humans and an out-of-distribution test set drawn from a geographically diverse set of insect camera traps. + +The AMI dataset is therefore designed to test fine-grained classification of moth species both in in-distribution and out-of-distribution contexts. It also includes the task of binary classification of insects as moth or non-moth, which is a prerequisite for camera traps since moths are not the only insects to appear. Note that we do not consider the task of object detection, which in applications is a precursor to insect identification. This is because object detection against a relatively plain background is a simpler task, where good results have already been obtained in the insect camera trap literature [18,20,21,32]. + +# 4 The AMI Dataset + +In this section, we describe the construction of the datasets that we release in this work: AMI-Traps and AMI-GBIF. For AMI-Traps, we collected insect camera trap data by field deployments in three regions: NE-America (Northeast of North America), which includes data from Quebec in Canada and Vermont in the USA; W-Europe (Western Europe), which includes data from the UK and Denmark; and C-America (Central America), comprising data from Panama. For AMI-GBIF, we compiled public data for the same three regions using the Global Biodiversity Information Facility (GBIF) [7]. The GBIF platform provides access to biodiversity data aggregated from multiple sources, such as museums, research institutions, and citizen science platforms such as iNaturalist [4], Observation.org [5], and BugGuide [1]. For brevity, we sometimes refer to AMI-Traps and AMI-GBIF as AMI-T and AMI-G, respectively. Figure 3 shows examples of images from the two datasets and Tab. 1 compares the number of images and coverage of classes at different taxonomic levels. + +# 4.1 AMI-Traps Dataset + +Data Collection. The source images in the AMI-Traps dataset were collected from a variety of different insect camera traps deployed across 22 monitoring stations in the three regions previously described. The cameras were configured + +![](images/6d3d09004541e7bbf9f29d11bdc159376009b17c5cc75bd7e740c4a592d92ee7.jpg) +Fig. 3: Sample images from the AMI Dataset. The AMI Dataset is composed of (1) AMI-GBIF, curated from a number of sources with imagery from citizen science and museum collections, and (2) AMI-Traps, drawn from automated camera traps for insects across five countries in three regions. The number of individual insects annotated per sub-dataset is denoted in parentheses. + +to be motion-sensitive, as well as taking photos on a fixed interval of 10 minutes; we selected a subset of images for annotation so as to avoid highly similar images. In total, 2,893 trap images were sampled from all cameras and made available on the Label Studio annotation platform (see Appendix B for more details). + +Annotation Process. Images were annotated by entomologists with expertise in the species of the particular region. Labeling was conducted according a three-step process. 1) We designed a custom object detection model to suggest boxes around objects in the image (refer to Appendix B for details on this model), to save annotators time. Annotators were asked to correct errors in the proposed boxes as well as drawing any missing boxes. Insects appearing smaller than 1 centimeter were ignored. 2) Annotators were asked to label image crops according to one of three coarse categories: moth, non-moth, or unidentifiable. Some insects could not be labeled due to motion blur, occlusion, or other factors. 3) For insects labeled as moths, annotators were asked to select the narrowest taxonomic group that could be determined from the image - identifying moths to species where possible, if not species then genus, and if not genus then family. + +Data Statistics. 2,893 source images from the traps were annotated (see Tab. B.2 in the supplementary for further details) which includes Vermont (900), Denmark (892), Quebec (533), UK (446), and Panama (122). A total of 52,948 insects were labeled within these images, including 37,105 $(70.08\%)$ non-moths, + +14,105 (26.64%) moths, and 1,738 (3.28%) unidentifiable. Out of the 14,105 moth crops, 10,854 (76.95%), 7,065 (50.09%) and 5,374 (38.10%) crops were labeled at the family, genus, and species level respectively. + +# 4.2 AMI-GBIF Dataset + +AMI-GBIF has two sub-datasets: binary and fine-grained classification. The fine-grained classification dataset contains images for moth species in each of the regions covered by AMI-Traps. The binary classification dataset contains images of moths aggregated across all regions and non-moths from anywhere on earth. + +Dataset Construction. First, in collaboration with regional moth specialists, we created "checklists" of species that could occur in each of the three regions of interest. It is worth noting that some species are included in more than one checklist, though most are unique to a single region. Next, we compared the checklists against the GBIF taxonomic backbone [2], merged synonym names (since some species have multiple accepted names), and manually investigated doubtful or unmatched names. Following that, we used the GBIF occurrence search tool to download metadata for all observations from the order Lepidoptera with images, which returned 15.1 million observations [3]. We then fetched images from observations in which the species is included in our regional checklists. The number of training images per species were limited to 1,000, though for many species it was considerably lower due to the long-tailed nature of the data. All images were resized to have the smallest dimension of 450px and packaged using the WebDataset format [6]. + +Quality Control. After downloading the images, we applied a set of cleaning steps to ensure the quality of the dataset. First, we removed duplicate URLs, i.e., images linked to more than one observation. This may occur when an image contains multiple individual insects, or when a placeholder is used for an observation (for example, Fig. A.1b in Appendix A shows a picture used as a placeholder for more than 100,000 observations). Next, we excluded images from specific datasets (refer to Tab. A.1 in Appendix A)) that we found to have a high percentage of images unsuitable for training our models, for instance, images of body parts or text only. Another issue we identified is that some observations contained thumbnail images that are too small to provide sufficient information to train our models; we removed all images with a width or height less than 64 pixels. Finally, as our dataset targets only adult insects, we removed images containing non-adult individuals such as eggs and caterpillars. For observations where the life stage was reported by the observer, approximately $60\%$ of the cases, we removed all occurrences tagged as non-adult. For the remaining images with a missing life stage tag, we designed a classification model to predict the life stage (refer to Appendix A for details about this model) and removed all images classified as non-adult. + +Dataset Split. Each GBIF observation may contain more than one image, but each image is considered an independent instance of the species. However, as images belonging to the same observation may be very similar, we keep images from the same observation in the same data split. To do so, we split the dataset + +using the observation ID as reference, with the proportion of $70\%$ , $10\%$ , and $20\%$ for the train, validation and test sets respectively. We apply a stratified split, i.e., we guarantee that this proportion is also maintained at the species level. We exclude species with fewer than 5 images, so the minimum number of images for a species in the training set is 4, and in the test set is one. Due to the data split based on observation ID, the validation set might lack images for some species. As some species may belong to multiple checklists, we keep the train/val/test splits consistent across the regions to prevent data leakage when training a unified model for all regions. + +Binary Classification. Since moths are not the only insects that appear on the camera trap screen, for our proposed benchmark task of binary classification, we curate a dataset comprising 350,000 images for each of the moth and non-moth categories. The moth class consists of adult moth images randomly sampled from the AMI-GBIF dataset. For the non-moth category, we leveraged expert knowledge of the taxonomic groups likely to appear on the screen, fetching images of adult individuals from GBIF of the following groups: Diptera (85,474), Hemiptera (75,053), Odonata (68,770), Coleoptera (45,954), Araneae (21,321), Orthoptera (21,261), Formicidae (12,206), Ichneumonidae (6,581), Trichoptera (5,316), Neuroptera (4,880), Opiliones (1,880), Ephemeroptera (1,304) — this includes mostly insects, with some Arachnids. The dataset is split into train, validation, and test sets with proportions of $70\%$ , $10\%$ , and $20\%$ , respectively. Specifically for the moth category, the splits are kept consistent with the splits for the fine-grained species classification task. + +Statistics. The final dataset contains $\sim 2.5\mathrm{M}$ images from 5,364 species, 1,734 genera and 77 families of moths, along with 350,000 additional images of non-moth groups. The binary dataset contains 700,000 images, composed of 350,000 moth images and the 350,000 non-moth images. The fine-grained dataset contains 2,564,392 moth images, and Tab. 1 shows the number of images and species for each region. Though species are selected based on their presence in regions of interest, we include all GBIF observations of these species from anywhere on Earth (see Fig. A.2 in supplementary for the global distribution of images). As in other biodiversity datasets, the number of images per species in AMI-GBIF follows a long-tailed distribution (see Fig. C.3 in supplementary). Specifically, only 3,099 species have more than 100 images and are considered under the "many-shot" bucket, 1,358 species contains between 20 and 100 images and are considered "medium-shot", and there are 907 species with less than 20 training images, considered as "few-shot" categories. + +# 5 Methods + +In this section, we describe the methods used to assess the difficulty of the proposed benchmark tasks. We evaluate a variety of high-performing image classification architectures and use extensive data augmentation to address the challenges of domain shift between the two datasets. Motivated by the particular structure of the AMI-Traps dataset, we add a mixed-resolution (MixRes) data + +Table 1: Comparison between the AMI-GBIF and AMI-Traps dataset for the fine-grained classification task. Images for the traps dataset denote the individual insect crops and not the raw trap images. The last three rows show the unique classes for each of the three taxonomic levels for the two datasets for different regions. It is also important to note that regions have species in common. + +
All RegionsNE-AmericaW-EuropeC-America
AMI-GAMI-TAMI-GAMI-TAMI-GAMI-TAMI-GAMI-T
Images2,564,39214,1051,179,9439,0661,579,3334,06699,405973
Families7743683465262711
Genera1,73434488720495719238512
Species5,3645162,4972772,6032446366
+ +augmentation technique. As we describe in the following section, our approaches perform quite well, with MixRes leading to an especially significant performance improvement. + +Models. We work with five computer vision models widely used in identification of wildlife: ResNet50 [35], MobileNetV3-Large [37], ConvNeXt-T [47], ConvNeXt-B [47], and ViT-B/16 [28]. We select ResNet50 because it is the most popular backbone used as a baseline for almost any computer vision task. Considering the possibility that the trained models may need to run on hardware with limited resources, such as edge devices and personal computers, for in-field predictions, we choose MobileNetV3Large as a lightweight model alternative. Furthermore, considering that vision transformers have surpassed convolutional neural networks and become the state-of-the-art models for image classification, we select the ViT-B/16 model to provide a baseline from this family of architectures. Finally, for representing modern convolutional neural networks, we choose the recently introduced ConvNeXt architecture, which has proven to be competitive with transformers using only standard ConvNet modules. From this family of models, we select two versions: the base version (ConvNeXt-B) and, considering lightweight usage scenarios, we also include the tiny version (ConvNeXt-T). Unless otherwise stated, we use a default input resolution of 128x128 for all models. + +Data Augmentation Approaches. One characteristic of the in-the-wild images is the large variety of crops due to the diversity of insect sizes. A considerable number of them have small resolutions, below $64 \times 64$ . At this scale, some important visual features that the model may have learned from high-resolution GBIF images might not be identifiable, leading to inaccurate predictions. This issue can be further amplified when these low-resolution crops are scaled up to the input model size. To compel the model to learn features for low-resolution inputs and avoid bias towards visual features only distinguishable in high-resolution images from other domains, we use a data augmentation technique we refer to as Mixed Resolution (MixRes) that simulates this low-resolution input during training (see Fig. D.4 in supplementary). This approach consists of randomly resizing the training input image to $25\%$ or $50\%$ of the model input size with a + +probability $\rho$ for each case. In our experiments $\rho$ is set to 0.25, maintaining the original input size with probability 0.5. We employ MixRes in all experiments unless otherwise stated. + +In addition to MixRes, we also use a set of standard data augmentation operations: random resize crop with scale of 0.3 to 1.0 and an aspect ratio between $3/4$ and $4/3$ , random horizontal flip with $50\%$ probability, and RandAugment [24] with $N = 2$ operations and magnitude $M = 9$ . Finally, the image is normalized using the mean and standard deviation of ImageNet and scaled to the input network size. + +Implementation Details. We train all models with the AMI-GBIF training sets using the ImageNet-1K pretrained weights from the timm library [76] (for ConvNext and ViT we used ImageNet-22K weights). Each model is trained with a batch size of 128 and the AdamW optimizer [48] for 30 epochs. The initial learning rate is set to 0.001, with a weight decay of 1e-5. The learning rate is linearly warmed up from 0 to the initial value for 2 epochs and then decayed using cosine scheduling [36]. We also apply label smoothing [65] with a value of 0.1. + +Inference Settings. At test time, the insect crops from AMI-Traps are padded with black pixels to make them square, preventing the distortion of the shape of the insect before prediction. Next, the images are normalized with the mean and standard deviation of ImageNet and, finally, resized to the model input size. No padding is done for the test images from AMI-GBIF. + +Higher Taxonomic Rank Prediction. We consider the accuracy of our baseline models both with respect to the exact insect species predicted and with respect to higher level taxonomic groups such as genus and family. Indeed some images in our dataset cannot be labeled at the species level at all due to the quality or perspective of the image, and our annotators instead provided labels at the genus or family level. In our experiments, to make predictions at a higher taxonomic level, we simply sum together species-level predictions to get genus results, and similarly, the genus predictions are grouped together to get family predictions. + +Evaluation Metrics. During the binary classification step, it is important to achieve high precision by filtering out as many non-moths as possible to prevent the propagation of errors in the pipeline. At the same time, it is also important to assess recall because if too many moths are discarded at this step, it may impact the utility of the data for ecological studies. Therefore, for the binary classification task, along with accuracy, we also consider precision, recall, and F1-score as evaluation metrics. For the fine-grained classification task, we evaluate the model in terms of micro- and macro-averaged accuracy at the species, genus and family levels. Micro-averaging aggregates all class samples to calculate the accuracy while the macro-average is the average of per-class accuracy. Unless otherwise stated, the reported accuracy is micro-averaged. We conduct each experiment for three independent runs and report the mean and standard deviation. + +# 6 Experiments and Results + +Binary Classification Results. This task involves distinguishing moths from non-moths. Here we train a ResNet50 model on the AMI-GBIF dataset and evaluate performance both on the AMI-GBIF held-out test set and on the entirety of the AMI-Traps dataset. On AMI-GBIF (in-distribution), the model achieves a test accuracy of $97.77\% \pm 0.02\%$ . Table 2 shows the results for the AMI-Traps dataset in more detail. We observe that the model has a high recall but a lower precision, which means it is over-predicting the moth class. Also, the bigger the image crop, the better the accuracy of the insect classification; many non-moths where the model errs are extremely small and are therefore highly blurred in the trap images. Table E.3 in the supplementary illustrates the improvement in model performance when crops are square padded (our default at inference time) versus an ablation where padding is not performed. + +Table 2: Binary classification model evaluation results on the AMI-Traps dataset after training on AMI-GBIF, showing that recall is high throughout, while precision drops on very small insects, which are predominantly non-moths. Crops $>100\mathrm{px}$ or $>150\mathrm{px}$ represent that atleast one side of the box is greater than the threshold. + +
MothsNon-mothsTotalAccuracyPrecisionRecallF1-score
All crops14,10537,10551,12086.48±1.0068.31±1.7095.03±0.3979.48±1.27
Crops > 100px11,63016,53428,16488.39±0.9779.46±1.4396.99±0.2087.35±0.94
Crops > 150px6,4465,96112,40792.95±0.2289.83±0.4397.48±0.1293.70±0.30
+ +Fine-grained Classification Results. The fine-grained classification task aims to identify moths to the species level (and secondarily to genus or family level). In our experiments, we train models on AMI-GBIF data for each of the three regions of the AMI dataset, evaluating both on AMI-GBIF (in-distribution) and AMI-Traps (out-of-distribution) data (see Tab. 3). Not all moth species present in a region appear regularly at camera traps; hence, the list of species in AMI-Traps is a subset of species in AMI-GBIF for a given area (see Tab. 1). Therefore, we also evaluate the models on an in-distribution subset $(\mathrm{Test}\ddagger)$ from AMI-GBIF containing only the species present in AMI-Traps. Meanwhile, for AMI-Traps, there are 16 species not present in AMI-GBIF that were removed during our evaluation procedure (these represent unexpected species not included in the regional checklists used to define AMI-GBIF). Overall, the best model at the species level is ConvNeXt-B for both AMI-GBIF and AMI-traps. For C-America, annotations are mostly provided at the genus or family level, due to the challenging nature of visual identification of moths in this biodiversity-rich region. Furthermore, while the genus and family accuracies are similar between N-America and W-Europe, they are much lower for C-America. This can be attributed to the fact that there are 10 times fewer images in the training set for this region. Note that for genus and family evaluation, there are some cases where the accuracy is lower than the species level. This occurs because those + +Table 3: Accuracy results for the fine-grained classification task. Overall, ConvNeXt-B achieves the best results for both AMI-GBIF and AMI-traps, with ViT-B/16 achieving slightly better results for some cases at the genus and family levels. Test‡ represents a subset of the AMI-GBIF test set, containing only the species present in AMI-Traps. Species-level results on AMI-Traps for the C-America region are not available since most annotations were provided only at higher taxonomic levels. + +
RegionModelAMI-GBIFAMI-Traps
SpeciesSpeciesGenusFamily
Micro TestMicro Test‡Micro Top1Macro Top1Micro Top1Macro Top1Micro Top1Macro Top1
NE-AmericaResNet5087.56±0.1289.39±0.1471.86±1.0264.96±2.0080.07±0.4974.66±0.3590.44±0.4169.30±4.15
MBNetV3L83.73±0.1785.80±0.2967.41±0.7261.21±1.2074.83±1.3071.20±2.4789.31±0.6264.48±1.87
ConvNeXt-T89.81±0.0591.21±0.0674.89±0.0568.00±0.7683.26±1.2277.26±0.8590.61±0.5374.60±0.64
ConvNeXt-B90.45±0.0391.66±0.0376.75±1.5268.56±0.6285.46±0.8378.28±0.1491.07±0.2976.87±1.10
ViT-B/1686.39±0.0587.94±0.1077.43±0.7968.02±0.3485.57±0.9878.97±0.7291.45±0.3473.66±1.49
W-EuropeResNet5086.45±0.0890.25±0.0379.39±0.3972.84±0.2178.95±0.2673.22±0.6089.13±0.4275.24±1.85
MBNetV3L82.20±0.0586.47±0.1376.30±1.2970.35±0.9376.35±1.1972.88±1.2888.33±0.4772.21±0.38
ConvNeXt-T88.82±0.0392.05±0.0582.44±0.8974.59±0.6281.86±0.7974.15±0.5990.08±0.5974.58±2.93
ConvNeXt-B89.35±0.0192.48±0.0582.87±0.3976.41±0.5082.45±0.4475.98±1.0591.27±0.6079.39±2.70
ViT-B/1685.45±0.0389.22±0.1182.44±0.3675.25±0.9883.01±0.2675.96±0.3591.68±0.1775.81±0.79
C-AmericaResNet5087.46±0.2395.92±0.68--47.44±1.1166.59±0.6069.18±0.2442.40±0.31
MBNetV3L83.68±0.2492.89±0.27--41.03±2.2259.32±1.1868.34±1.6941.42±2.98
ConvNeXt-T87.65±0.1195.46±0.46--55.77±3.3370.68±1.5869.46±0.4245.35±0.67
ConvNeXt-B88.62±0.0996.22±0.26--51.92±3.8568.63±1.9472.24±1.4744.31±1.32
ViT-B/1679.80±0.2688.36±1.02--47.44±4.4464.70±4.0769.87±1.5142.00±3.60
+ +higher taxonomic levels include crops that have labels only at higher levels. In the supplementary material, we include additional results for AMI-GBIF long-tail accuracies (Tab. E.4), a pre-training ablation study (Tab. E.5), and a study analyzing classification accuracy as a function of prediction confidence (Fig. E.5). + +Data augmentation approaches. To assess the effectiveness of the mixed resolution data augmentation technique, we conduct an ablation study using ResNet50 for the NE-America region. As the effects of MixRes could already be covered by other data augmentation operations used in our preprocessing, such as random resize crop or one of the operations of RandAugment, we analyze multiple combinations of these operations. Additionally, we evaluate a higher input resolution, using the original 224 input size of ResNet50, to determine whether this leads to better predictions. Table 4 shows the results. While applying MixRes results in a slight decrease in accuracy within the in-distribution AMI-GBIF test set, for the in-the-wild AMI-Traps test set there is a substantial improvement, ranging between $4\%$ and $7\%$ when compared to models not trained with MixRes. Another important observation is that, although RandAugment and random resize together contribute to an overall improvement of around $16\%$ , MixRes remains additionally beneficial. Finally, despite an increase of more than $4\%$ in accuracy within the original domain when using a higher resolution, there is a marginal decrease in the accuracy of the in-the-wild results. + +# 7 Conclusion + +Many ecological applications require species-level image classification to meet their goals. Despite advances in computer vision, there remain challenges in + +Table 4: Ablation study on data augmentation techniques using a ResNet50 for the NE-America region evaluated on micro-averaged top-1 species accuracy. "RandAug" denotes the combination of the RandAugment and random resize crop data augmentation techniques. AMI-GBIF† test set contains only the species present in AMI-Traps. + +
ResolutionRandAugMixResAMI-GBIF‡AMI-Traps
128××89.18±0.0851.46±2.73
128✓×89.71±0.1667.88±1.27
128×✓87.41±0.0658.41±0.30
128✓✓89.39±0.1471.86±1.02
224✓✓92.97±0.1768.41±1.06
+ +implementing fine-grained classification that is effective in real-world scenarios. Here we outline the challenges of insect identification in the wild and propose a benchmark dataset to tackle them. We present AMI-GBIF, a dataset with 2.5 million labeled photographs of 5,000 moth species. We also introduce AMI-Traps, a dataset of images captured by automated insect monitoring systems across five countries, labeled by regional experts. We evaluate multiple classification models both on the AMI-GBIF dataset and in out-of-distribution generalization to AMI-Traps, with ConvNeXt-B generally performing the best. We also evaluate a range of data augmentation techniques, finding that a combination of simple approaches, notably mixed-resolution augmentation, considerably improves generalization under field conditions. + +Several limitations of this work should be noted. First, while annotating AMI-Traps already occupied a considerable time for expert entomologists, further expanding it to a greater number of species and geographic regions would be helpful in further stress-testing field deployments of fine-grained classification algorithms. Second, while we focus on camera traps for nocturnal insects, it would be useful also to consider diurnal insect camera traps, for example those aimed at pollinators, which pose some additional challenges, such as multiple perspectives and a diversity of backgrounds. We hope furthermore that future works will consider more sophisticated domain adaptation approaches building on the baselines we present here. + +This work represents a step towards computer vision algorithms for species recognition that are effective in the ultra-challenging regime of insect identification and work effectively in field conditions. The approach we propose should make it possible in particular to generalize algorithms for insect camera traps rapidly and cost-effectively across different hardware setups and regions of interest. We hope that the datasets, challenges and strategies presented here will propel further research in computer vision for insects and help accelerate initiatives to understand and protect our planet's biodiversity. + +# Acknowledgements + +We thank Annie-Shan Morin, Anna Viklund, and Jamie Alison for their helpful input and support to the project. + +This research was generously funded by Mila Quebec AI Institute and Ministère de l'Économie, de l'Innovation et de l'Energie. The work was additionally supported in part by the Canada CIFAR AI Chairs program (NSERC 585136), the Global Climate Center on AI and Biodiversity Change (ABC) (NSF OISE2330423), the Montreal Insectarium, the Coordenação de Aperfeçoamento de Pessoal de Nível Superior - Brasil (CAPES-PROEX) - Finance Code 001, the Amazonas State Research Support Foundation - FAPEAM - through the POSGRAD project 2024/2025, the AMBER project funded by the abrdn Charitable Foundation, the Aage V. Jensen Naturfond grant no. N2024-0013, the European Union's Horizon Europe Research and Innovation programme under grant agreement no. 101060639 (MAMBO), and a Discovery Grant by NSERC. The computing resources were provided by Mila Quebec AI Institute, the Digital Research Alliance of Canada, and NVIDIA Corporation. + +A.J., F.C., and M.J.B. led dataset creation and experiments. J.S.C. and L.P. contributed to machine learning pipeline development. N.P., F.H, J.R., M.B., M.S., J.F., A.A., Y.L., E.N., F.P.P., A.C.Z., J.A.R.S., and J.G. performed insect identification for the AMI-Traps dataset. T.A, K.B., A.G.S., M.B., Y.B., K.P.M, D. Roy, T.T.H., M.L., and D. Rolnick guided and advised the project. + +# References + +1. BugGuide.Net: Identification, Images, & Information For Insects, Spiders & Their Kin For the United States & Canada, https://bugguide.net/ accessed Feb 12 2024 +2. GBIF Backbone Taxonomy, https://doi.org/10.15468/39omei accessed Nov 13 2023 +3. GBIF Occurrence Download, https://doi.org/10.15468/dl.4ua6wg accessed Oct 19 2023 +4. iNaturalist, www.inaturalist.org, accessed Feb 9 2024 +5. Observation.org, https://observation.org, accessed Feb 12 2024 +6. WebDataset, https://github.com/webdataset/webdataset, accessed on April 2 2024 +7. What is GBIF?, https://www.gbif.org/what-is-gbif accessed Sep 28 2023 +8. Alison, J., Alexander, J.M., Diaz Zeugin, N., Dupont, Y.L., Iseli, E., Mann, H.M., Hoye, T.T.: Moths complement bumblebee pollination of red clover: a case for day-and-night insect surveillance. Biology Letters 18(7), 20220187 (2022) +9. Anton, V., Hartley, S., Geldenhuis, A., Wittmer, H.U.: Monitoring the mammalian fauna of urban areas using remote cameras and citizen science. Journal of Urban Ecology 4(1), juy002 (2018) +10. Badgujar, C.M., Armstrong, P.R., Gerken, A.R., Pordesimo, L.O., Campbell, J.F.: Real-time stored product insect detection and identification using deep learning: System integration and extensibility to mobile platforms. Journal of Stored Products Research 104, 102196 (2023) + +11. Beery, S.: Scaling biodiversity monitoring for the data age. XRDS: Crossroads, The ACM Magazine for Students 27(4), 14-18 (2021) +12. Beery, S., Agarwal, A., Cole, E., Birodkar, V.: The iwildcam 2021 competition dataset. arXiv preprint arXiv:2105.03494 (2021) +13. Beery, S., Morris, D., Yang, S.: Efficient pipeline for camera trap image review. arXiv preprint arXiv:1907.06772 (2019) +14. Beery, S., Van Horn, G., Perona, P.: Recognition in terra incognita. In: Proceedings of the European conference on computer vision (ECCV). pp. 456-473 (2018) +15. Beery, S., Wu, G., Edwards, T., Pavetic, F., Majewski, B., Mukherjee, S., Chan, S., Morgan, J., Rathod, V., Huang, J.: The auto arborist dataset: A large-scale benchmark for multiview urban forest monitoring under domain shift. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21294-21307 (2022) +16. Berg, T., Liu, J., Woo Lee, S., Alexander, M.L., Jacobs, D.W., Belhumeur, P.N.: Birdsnap: Large-scale fine-grained visual categorization of birds. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2011-2018 (2014) +17. Bjerge, K., Alison, J., Dyrmann, M., Frigaard, C.E., Mann, H.M., Hoye, T.T.: Accurate detection and identification of insects from camera trap images with deep learning. PLOS Sustainability and Transformation 2(3), e0000051 (2023) +18. Bjerge, K., Frigaard, C.E., Karstoft, H.: Object detection of small insects in timelapse camera recordings. Sensors 23(16), 7242 (2023) +19. Bjerge, K., Geissmann, Q., Alison, J., Mann, H.M., Hoye, T.T., Dyrmann, M., Karstoft, H.: Hierarchical classification of insects with multitask learning and anomaly detection. Ecological Informatics 77, 102278 (2023) +20. Bjerge, K., Mann, H.M., Hoye, T.T.: Real-time insect tracking and monitoring with computer vision and deep learning. Remote Sensing in Ecology and Conservation 8(3), 315-327 (2022) +21. Bjerge, K., Nielsen, J.B., Sepstrup, M.V., Helsing-Nielsen, F., Hoye, T.T.: An automated light trap to monitor moths (lepidoptera) using computer vision-based tracking and deep learning. Sensors 21(2), 343 (2021) +22. Boom, B.J., He, J., Palazzo, S., Huang, P.X., Bayan, C., Chou, H.M., Lin, F.P., Spampinato, C., Fisher, R.B.: A research tool for long-term and continuous analysis of fish assemblage in coral-reefs using underwater camera footage. Ecological Informatics 23, 83-97 (2014) +23. Borowiec, M.L., Dikow, R.B., Frandsen, P.B., McKeeken, A., Valentini, G., White, A.E.: Deep learning as a tool for ecology and evolution. Methods in Ecology and Evolution 13(8), 1640-1660 (2022) +24. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 702-703 (2020) +25. Cutter, G., Stierhoff, K., Zeng, J.: Automated detection of rockfish in unconstrained underwater videos using haar cascades and a new image dataset: Labeled fishes in the wild. In: 2015 IEEE Winter Applications and Computer Vision Workshops. pp. 57-62. IEEE (2015) +26. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 248-255. IEEE (2009) + +27. Didham, R.K., Basset, Y., Collins, C.M., Leather, S.R., Littlewood, N.A., Menz, M.H., Müller, J., Packer, L., Saunders, M.E., Schonrogge, K., et al.: Interpreting insect declines: seven challenges and a way forward. Insect Conservation and Diversity 13(2), 103-114 (2020) +28. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net (2021), https://openreview.net/forum?id=YicbFdNTTy +29. Dubey, A., Ramanathan, V., Pentland, A., Mahajan, D.: Adaptive methods for real-world domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14340-14349 (2021) +30. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision 88, 303-338 (2010) +31. Fennell, M., Beirne, C., Burton, A.C.: Use of object detection in camera trap image identification: Assessing a method to rapidly and accurately classify human and animal detections for research and application in recreation ecology. Global Ecology and Conservation 35, e02104 (2022) +32. Geissmann, Q., Abram, P.K., Wu, D., Haney, C.H., Carrillo, J.: Sticky pi is a high-frequency smart trap that enables the study of insect circadian activity under natural conditions. PLoS Biology 20(7), e3001689 (2022) +33. Gharaee, Z., Gong, Z., Pellegrino, N., Zarubiieva, I., Haurum, J.B., Lowe, S.C., McKeown, J.T., Ho, C.C., McLeod, J., Wei, Y.Y.C., et al.: A step towards worldwide biodiversity assessment: The bioscan-1m insect dataset. arXiv preprint arXiv:2307.10455 (2023) +34. Gulrajani, I., Lopez-Paz, D.: In search of lost domain generalization. arXiv preprint arXiv:2007.01434 (2020) +35. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +36. He, T., Zhang, Z., Zhang, H., Zhang, Z., Xie, J., Li, M.: Bag of tricks for image classification with convolutional neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 558-567 (2019) +37. Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al.: Searching for mobilenetv3. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1314-1324 (2019) +38. Hoye, T.T., Arje, J., Bjerge, K., Hansen, O.L., Iosifidis, A., Leese, F., Mann, H.M., Meissner, K., Melvad, C., Raitoharju, J.: Deep learning and computer vision will transform entomology. Proceedings of the National Academy of Sciences 118(2), e2002545117 (2021) +39. Katija, K., Orenstein, E., Schlining, B., Lundsten, L., Barnard, K., Sainz, G., Boulais, O., Cromwell, M., Butler, E., Woodward, B., et al.: Fathomnet: A global image database for enabling artificial intelligence in the ocean. Scientific reports 12(1), 15914 (2022) +40. Kay, J., Kulits, P., Stathatos, S., Deng, S., Young, E., Beery, S., Van Horn, G., Perona, P.: The caltech fish counting dataset: A benchmark for multiple-object tracking and counting. In: European Conference on Computer Vision. pp. 290-311. Springer (2022) + +41. van Klink, R., August, T., Bas, Y., Bodesheim, P., Bonn, A., Fossoy, F., Høye, T.T., Jongejans, E., Menz, M.H., Miraldo, A., et al.: Emerging technologies revolutionise insect ecology and monitoring. Trends in ecology & evolution (2022) +42. Koh, P.W., Sagawa, S., Marklund, H., Xie, S.M., Zhang, M., Balsubramani, A., Hu, W., Yasunaga, M., Phillips, R.L., Gao, I., Lee, T., David, E., Stavness, I., Guo, W., Earnshaw, B.A., Haque, I.S., Beery, S., Leskovec, J., Kundaje, A., Pierson, E., Levine, S., Finn, C., Liang, P.: WILDS: A benchmark of in-the-wild distribution shifts. In: International Conference on Machine Learning (ICML) (2021) +43. Korsch, D., Bodesheim, P., Denzler, J.: Deep learning pipeline for automated visual moth monitoring: insect localization and species classification. arXiv preprint arXiv:2307.15427 (2023) +44. Krasin, I., Duerig, T., Alldrin, N., Ferrari, V., Abu-El-Haija, S., Kuznetsova, A., Rom, H., Uijlings, J., Popov, S., Veit, A., et al.: Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://github.com/openimages2(3), 18 (2017) +45. Kumar, N., Belhumeur, P.N., Biswas, A., Jacobs, D.W., Kress, W.J., Lopez, I.C., Soares, J.V.: Leafsnap: A computer vision system for automatic plant species identification. In: Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part II 12. pp. 502-516. Springer (2012) +46. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +47. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022) +48. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019) +49. Luccioni, A.S., Rolnick, D.: Bugs in the data: How imagenet misrepresents biodiversity. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 14382-14390 (2023) +50. Ming, Y., Yin, H., Li, Y.: On the impact of spurious correlation for out-of-distribution detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 10051-10059 (2022) +51. Mora, C., Tittensor, D.P., Adl, S., Simpson, A.G., Worm, B.: How many species are there on earth and in the ocean? PLoS biology 9(8), e1001127 (2011) +52. Nawoya, S., Ssemakula, F., Akol, R., Geissmann, Q., Karstoft, H., Bjerge, K., Mwikirize, C., Katumba, A., Gebreyesus, G.: Computer vision and deep learning in insects for food and feed production: A review. Computers and Electronics in Agriculture 216, 108503 (2024) +53. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian conference on computer vision, graphics & image processing. pp. 722-729. IEEE (2008) +54. Norouzzadeh, M.S., Morris, D., Beery, S., Joshi, N., Jojic, N., Clune, J.: A deep active learning system for species identification and counting in camera trap images. Methods in ecology and evolution 12(1), 150-161 (2021) +55. Oliver, R.Y., Iannarilli, F., Ahumada, J., Fegraus, E., Flores, N., Kays, R., Birch, T., Ranipeta, A., Rogan, M.S., Sica, Y.V., et al.: Camera trapping expands the view into global biodiversity and its change. Philosophical Transactions of the Royal Society B 378(1881), 20220232 (2023) + +56. Ouaknine, A., Kattenborn, T., Laliberte, E., Rolnick, D.: Openforest: A data catalogue for machine learning in forest monitoring. arXiv preprint arXiv:2311.00277 (2023) +57. Rolnick, D., Donti, P.L., Kaack, L.H., Kochanski, K., Lacoste, A., Sankaran, K., Ross, A.S., Milojevic-Dupont, N., Jaques, N., Waldman-Brown, A., et al.: Tackling climate change with machine learning. ACM Computing Surveys (CSUR) 55(2), 1-96 (2022) +58. Roy, D., et al.: Towards a standardised framework for ai-assisted, image-based monitoring of nocturnal insects. Philosophical Transactions of the Royal Society B (2024). https://doi.org/10.1098/rstb.2023.0108, (Forthcoming) +59. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) 115(3), 211-252 (2015) +60. Seibold, S., Gossner, M.M., Simons, N.K., Bluthgen, N., Müller, J., Ambarh, D., Ammer, C., Bauhus, J., Fischer, M., Habel, J.C., et al.: Arthropod decline in grasslands and forests is associated with landscape-level drivers. Nature 574(7780), 671-674 (2019) +61. Stark, T., Stefan, V., Wurm, M., Spanier, R., Taubenbock, H., Knight, T.M.: Yolo object detection models can locate and classify broad groups of flower-visiting arthropods in images. Scientific Reports 13(1), 16364 (2023) +62. Stork, N.E., et al.: How many species of insects and other terrestrial arthropods are there on earth. Annual review of entomology 63(1), 31-45 (2018) +63. Suto, J.: Codling moth monitoring with camera-equipped automated traps: A review. Agriculture 12(10), 1721 (2022) +64. Swanson, A., Kosmala, M., Lintott, C., Simpson, R., Smith, A., Packer, C.: Snapshot serengeti, high-frequency annotated camera trap images of 40 mammalian species in an african savanna. Scientific data 2(1), 1-14 (2015) +65. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818-2826 (2016) +66. Teixeira, A.C., Ribeiro, J., Morais, R., Sousa, J.J., Cunha, A.: A systematic review on automatic insect detection using deep learning. Agriculture 13(3), 713 (2023) +67. Theivaprakasham, H.: Identification of indian butterflies using deep convolutional neural network. Journal of Asia-Pacific Entomology 24(1), 329-340 (2021) +68. Tuia, D., Kellenberger, B., Beery, S., Costelloe, B.R., Zuffi, S., Risse, B., Mathis, A., Mathis, M.W., van Langevelde, F., Burghardt, T., et al.: Perspectives in machine learning for wildlife conservation. Nature communications 13(1), 792 (2022) +69. Van Horn, G., Branson, S., Farrell, R., Haber, S., Barry, J., Ipeirotis, P., Perona, P., Belongie, S.: Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 595-604 (2015) +70. Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist species classification and detection dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8769-8778 (2018) +71. Van Horn, G., Perona, P.: The devil is in the tails: Fine-grained classification in the wild. arXiv preprint arXiv:1709.01450 (2017) +72. Wagner, D.L.: Insect declines in the anthropocene. Annual review of entomology 65, 457-480 (2020) + +73. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset. Tech. rep. (2011) +74. Wei, X.S., Song, Y.Z., Mac Aodha, O., Wu, J., Peng, Y., Tang, J., Yang, J., Belongie, S.: Fine-grained image analysis with deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence 44(12), 8927-8948 (2021) +75. Weinstein, B.G.: A computer vision for animal ecology. Journal of Animal Ecology 87(3), 533-545 (2018) +76. Wightman, R.: Pytorch image models. https://github.com/rwrightman/pytorch-image-models (2019) +77. Zhou, K., Liu, Z., Qiao, Y., Xiang, T., Loy, C.C.: Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022) \ No newline at end of file diff --git a/2024/Insect Identification in the Wild_ The AMI Dataset/images.zip b/2024/Insect Identification in the Wild_ The AMI Dataset/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..91e8d224c537f24d07b4c4e28095ea0cea8f9e35 --- /dev/null +++ b/2024/Insect Identification in the Wild_ The AMI Dataset/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f933fd395a1e90ee7dd7469944730d536dae3791477714d482d7f0e98f4a903e +size 406038 diff --git a/2024/Insect Identification in the Wild_ The AMI Dataset/layout.json b/2024/Insect Identification in the Wild_ The AMI Dataset/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..82f16fea4c3cf6aff6c0759c6df6bf69fdc2d1ec --- /dev/null +++ b/2024/Insect Identification in the Wild_ The AMI Dataset/layout.json @@ -0,0 +1,8545 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 192, + 111, + 422, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 111, + 422, + 146 + ], + "spans": [ + { + "bbox": [ + 192, + 111, + 422, + 146 + ], + "type": "text", + "content": "Insect Identification in the Wild: The AMI Dataset" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 134, + 167, + 479, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 167, + 479, + 265 + ], + "spans": [ + { + "bbox": [ + 134, + 167, + 479, + 265 + ], + "type": "text", + "content": "Aditya Jain1*, Fagner Cunha2*, Michael James Bunsen1*, Juan Sebastián Canas1, Léonard Pasi1, Nathan Pinoy3, Flemming Helsing3, JoAnne Russo4, Marc Botham5, Michael Sabourin6, Jonathan Fréchette12, Alexandre Anctil7, Yacksecari Lopez8, Eduardo Navarro8, Filonila Perez Pimentel8, Ana Cecilia Zamora8, José Alejandro Ramirez Silva8, Jonathan Gagnon12, Tom August5, Kim Bjerge3, Alba Gomez Segura5, Marc Bélisle9, Yves Basset8, Kent P. McFarland4, David Roy5, Toke Thomas Hoye3, Maxim Larrivée10, and David Rolnick1,11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 248, + 274, + 367, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 274, + 367, + 285 + ], + "spans": [ + { + "bbox": [ + 248, + 274, + 367, + 285 + ], + "type": "text", + "content": "1 Mila - Quebec AI Institute" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 238, + 285, + 376, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 285, + 376, + 295 + ], + "spans": [ + { + "bbox": [ + 238, + 285, + 376, + 295 + ], + "type": "text", + "content": "2 Federal University of Amazonas" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 265, + 295, + 348, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 295, + 348, + 307 + ], + "spans": [ + { + "bbox": [ + 265, + 295, + 348, + 307 + ], + "type": "text", + "content": "3 Aarhus University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 240, + 307, + 374, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 307, + 374, + 316 + ], + "spans": [ + { + "bbox": [ + 240, + 307, + 374, + 316 + ], + "type": "text", + "content": "4 Vermont Center for Ecostudies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 228, + 317, + 386, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 317, + 386, + 329 + ], + "spans": [ + { + "bbox": [ + 228, + 317, + 386, + 329 + ], + "type": "text", + "content": "UK Centre for Ecology & Hydrology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 269, + 329, + 345, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 329, + 345, + 338 + ], + "spans": [ + { + "bbox": [ + 269, + 329, + 345, + 338 + ], + "type": "text", + "content": "McGuire Center" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 190, + 339, + 424, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 339, + 424, + 350 + ], + "spans": [ + { + "bbox": [ + 190, + 339, + 424, + 350 + ], + "type": "text", + "content": "7 Centre de données sur le patrimoine naturel du Québec" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 220, + 350, + 393, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 350, + 393, + 361 + ], + "spans": [ + { + "bbox": [ + 220, + 350, + 393, + 361 + ], + "type": "text", + "content": "Smithsonian Tropical Research Institute" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 251, + 361, + 362, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 361, + 362, + 371 + ], + "spans": [ + { + "bbox": [ + 251, + 361, + 362, + 371 + ], + "type": "text", + "content": "9 Université de Sherbrooke" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 257, + 372, + 357, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 372, + 357, + 383 + ], + "spans": [ + { + "bbox": [ + 257, + 372, + 357, + 383 + ], + "type": "text", + "content": "10 Montreal Insectarium" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 264, + 384, + 350, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 384, + 350, + 394 + ], + "spans": [ + { + "bbox": [ + 264, + 384, + 350, + 394 + ], + "type": "text", + "content": "11 McGill University" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 252, + 394, + 362, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 394, + 362, + 405 + ], + "spans": [ + { + "bbox": [ + 252, + 394, + 362, + 405 + ], + "type": "inline_equation", + "content": "^{12}" + }, + { + "bbox": [ + 252, + 394, + 362, + 405 + ], + "type": "text", + "content": " Independent Researcher" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 159, + 419, + 453, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 419, + 453, + 639 + ], + "spans": [ + { + "bbox": [ + 159, + 419, + 453, + 639 + ], + "type": "text", + "content": "Abstract. Insects represent half of all global biodiversity, yet many of the world's insects are disappearing, with severe implications for ecosystems and agriculture. Despite this crisis, data on insect diversity and abundance remain woefully inadequate, due to the scarcity of human experts and the lack of scalable tools for monitoring. Ecologists have started to adopt camera traps to record and study insects, and have proposed computer vision algorithms as an answer for scalable data processing. However, insect monitoring in the wild poses unique challenges that have not yet been addressed within computer vision, including the combination of long-tailed data, extremely similar classes, and significant distribution shifts. We provide the first large-scale machine learning benchmarks for fine-grained insect recognition, designed to match real-world tasks faced by ecologists. Our contributions include a curated dataset of images from citizen science platforms and museums, and an expert-annotated dataset drawn from automated camera traps across multiple continents, designed to test out-of-distribution generalization under field conditions. We train and evaluate a variety of baseline algorithms and introduce a combination of data augmentation techniques that enhance generalization across geographies and hardware setups. The dataset is made publicly available1." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 643, + 430, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 430, + 654 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 430, + 654 + ], + "type": "text", + "content": "* Denotes equal contribution. Correspondence to: moth-ai@mila.quebec" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 135, + 654, + 338, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 654, + 338, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 654, + 338, + 665 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 654, + 338, + 665 + ], + "type": "text", + "content": " https://github.com/RolnickLab/ami-dataset" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 145, + 482, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 145, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 145, + 482, + 216 + ], + "type": "text", + "content": "Computer vision algorithms for species identification have become important tools in studying and protecting the Earth's wildlife. Biodiversity applications have also been beneficial to computer vision, serving as a fruitful source of new benchmarks and algorithms [11,23,68,75]. However, research in computer vision for biodiversity has often focused on large animals, not reflecting the full set of challenges faced in ecology and conservation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 218, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 218, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 482, + 384 + ], + "type": "text", + "content": "There are over a million known species of insects on Earth, representing nearly one half of all known organisms (by comparison, there are only 10,000 species of birds, and even fewer mammals) [51,62]. Insects play essential roles in nearly all terrestrial ecosystems, as well as providing ecosystem services such as pollination that are critical for humanity [60], yet a looming \"insect apocalyptic\" threatens both their abundance and diversity [72]. Despite the crisis, data on insect populations is shockingly sparse [27], as a result of less resources, a plethora of species, and the high level of expertise typically required to study insects. Amid this data gap, computer vision tools for automated monitoring have the potential to radically reshape the study and conservation of insects. In particular, recent breakthroughs in camera traps for insects have made it possible to automatically photograph large numbers of individuals in the wild [38,41,58], opening up the the potential for image-processing algorithms to parse this raw data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 387, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 506 + ], + "type": "text", + "content": "Species recognition from camera traps is a well-studied problem in the computer vision community [55], with common challenges including poor lighting, occlusion of individuals, camouflage, and blur [14]. However, working with insects presents distinct challenges from those faced in traditional camera traps, which are often designed for large animals. For example, while traditional camera trap images may only rarely contain a target species, almost every image from an insect camera trap includes insects – often hundreds of them. Where traditional camera traps may have a short list of species of interest, insect monitoring requires considering thousands of species at any given location, with many of them being extremely similar, and most with very little labeled data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 506, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 579 + ], + "type": "text", + "content": "While recent authors have proposed the use of computer vision tools for detecting and identifying insects in camera trap data [8, 17, 20, 32, 43, 63], such studies have involved small datasets with only a few species represented. There has to date been no work assessing the performance of algorithms in identifying large number of insect species, or addressing the challenges of generalization across in-the-wild deployments. Our work here offers the following contributions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "text", + "content": "- We introduce the AMI (Automated Monitoring of Insects) dataset, consisting of two parts: 1) AMI-GBIF, a dataset of " + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sim 2.5\\mathrm{M}" + }, + { + "bbox": [ + 138, + 593, + 482, + 666 + ], + "type": "text", + "content": " human-captured images curated from citizen science platforms and museum collections, 2) AMI-Traps, an expert-annotated dataset of 2,893 insect camera trap images (representing 52,948 labeled insects) collected from a global network of automated camera traps, designed to test in-the-wild performance (Fig. 1)." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "A. Jain et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 482, + 219 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 482, + 219 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 482, + 219 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 482, + 219 + ], + "type": "image", + "image_path": "00003212e95be6e5dd272b91109bf82f0a6193888b5755b4b1ef396d96a576bc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 225, + 482, + 302 + ], + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 302 + ], + "type": "text", + "content": "Fig. 1: Illustration of the difference between insect species identification from human-captured images and in the wild. On left are shown images of three moth species from our AMI-GBIF dataset, curated from citizen science data and museum collections. On right are shown images of the same species (along with other insects) within a photograph taken by an automated camera trap in the wild from our AMI-Traps dataset. One of the challenges of the AMI dataset is generalizing from AMI-GBIF to AMI-Traps without additional labeled data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 326, + 481, + 409 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 326, + 481, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 481, + 373 + ], + "type": "text", + "content": "- We frame the problem of practical insect monitoring in the field as a set of computer vision tasks, entailing very fine-grained classification from long-tailed data, with the additional challenge of generalization across multiple kinds of imagery." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 374, + 481, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 481, + 409 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 481, + 409 + ], + "type": "text", + "content": "- We test a variety of strong baseline models for the proposed tasks, including a combination of data augmentation approaches that significantly improve performance under field conditions." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 437, + 237, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 237, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 237, + 449 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "content": "Biodiversity Image Datasets. Advancements in computer vision have been closely intertwined with the creation of curated datasets such as ImageNet [26, 59], Pascal VOC [30], Open Images [44], and MS-COCO [46]. In recent years, the computer vision and machine learning communities have increasingly directed their focus towards addressing pressing human threats, including biodiversity loss [68] and climate change [57]. These domains not only hold significant potential for societal impact but also offer fertile ground for exploring intriguing machine learning questions such as domain generalization [77], long-tailed distributions [71], and fine-grained classification [74]. Such challenges have found especially fruitful testbeds in the classification of species, where datasets such as iNaturalist [70] have proven useful for both the computer vision and ecology communities. While some challenges have considered plant identification [15, 45, 53, 56], the majority have focused on animals, including datasets focused on birds [16, 69, 73], mammals [9, 14, 64], and fish [22, 25, 39, 40]. Such datasets have often leveraged concurrent advancements in sensor technology, including in satellite imagery and drones. In particular, camera traps have been widely adopted by biologists as an in situ and cost-effective means of wildlife" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 315, + 91, + 447, + 100 + ], + "type": "text", + "content": "Insect Identification in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 117, + 181, + 145 + ], + "blocks": [ + { + "bbox": [ + 138, + 117, + 181, + 145 + ], + "lines": [ + { + "bbox": [ + 138, + 117, + 181, + 145 + ], + "spans": [ + { + "bbox": [ + 138, + 117, + 181, + 145 + ], + "type": "image", + "image_path": "c8422f237016fb6a7084b360cb3732bc8a7d1c35b2e6e6375849c53ca34a5014.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 139, + 159, + 178, + 191 + ], + "blocks": [ + { + "bbox": [ + 138, + 149, + 179, + 158 + ], + "lines": [ + { + "bbox": [ + 138, + 149, + 179, + 158 + ], + "spans": [ + { + "bbox": [ + 138, + 149, + 179, + 158 + ], + "type": "text", + "content": "(a) Incomplete" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 139, + 159, + 178, + 191 + ], + "lines": [ + { + "bbox": [ + 139, + 159, + 178, + 191 + ], + "spans": [ + { + "bbox": [ + 139, + 159, + 178, + 191 + ], + "type": "image", + "image_path": "c8d7dee1236803f8056af0ee0b3845fd33dd3d277603922e0f9d64c8b7e2393e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 138, + 202, + 179, + 232 + ], + "blocks": [ + { + "bbox": [ + 138, + 192, + 181, + 201 + ], + "lines": [ + { + "bbox": [ + 138, + 192, + 181, + 201 + ], + "spans": [ + { + "bbox": [ + 138, + 192, + 181, + 201 + ], + "type": "text", + "content": "(b) Overlapping" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 202, + 179, + 232 + ], + "lines": [ + { + "bbox": [ + 138, + 202, + 179, + 232 + ], + "spans": [ + { + "bbox": [ + 138, + 202, + 179, + 232 + ], + "type": "image", + "image_path": "ff4a585ec332620364206eefd14b9081a9fe36e330ff7b5f7ca3df710832ddc0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 233, + 173, + 241 + ], + "lines": [ + { + "bbox": [ + 143, + 233, + 173, + 241 + ], + "spans": [ + { + "bbox": [ + 143, + 233, + 173, + 241 + ], + "type": "text", + "content": "(c) Motion" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 185, + 114, + 304, + 175 + ], + "blocks": [ + { + "bbox": [ + 185, + 114, + 304, + 175 + ], + "lines": [ + { + "bbox": [ + 185, + 114, + 304, + 175 + ], + "spans": [ + { + "bbox": [ + 185, + 114, + 304, + 175 + ], + "type": "image", + "image_path": "d9d2b1c95c7f02a1d8fca63c3d025a0f755d16b8d30d32f621e88939ab03b5e2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 185, + 176, + 266, + 243 + ], + "blocks": [ + { + "bbox": [ + 185, + 176, + 266, + 243 + ], + "lines": [ + { + "bbox": [ + 185, + 176, + 266, + 243 + ], + "spans": [ + { + "bbox": [ + 185, + 176, + 266, + 243 + ], + "type": "image", + "image_path": "2107424f9c29b4c8e3cc810131011b20ef05a0ea9a3822c40be081770f90c764.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 252, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 132, + 252, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 482, + 361 + ], + "type": "text", + "content": "Fig. 2: Examples of challenges in the AMI-Traps dataset. (a) Incomplete: In some occasions, the body of an insect is incomplete due to its position in the automatically taken image. (b) Overlapping: Sometimes insects occlude each other. (c) Motion: Moving insects can be blurred and hard to classify. (d) Lighting: The lighting and camera exposure of the images is variable and can lead to poor contrast that masks some or all detail. (e) Resolution: Some insects are very small, leading to low-resolution images. (f) Perspective: While insects are generally positioned at a fixed distance, sometimes insects appear in the air or perched on the camera itself. Additionally, unexpected objects such as spider webs, dirt, and vegetation can be captured by the cameras and may be confused with objects of interest." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 428, + 175 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 428, + 175 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 428, + 175 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 428, + 175 + ], + "type": "image", + "image_path": "d25edc3113fff52e6dd677f6c5d7786b8f7aa6715e2d2171963f3f77c8b59014.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 266, + 176, + 428, + 243 + ], + "blocks": [ + { + "bbox": [ + 266, + 176, + 428, + 243 + ], + "lines": [ + { + "bbox": [ + 266, + 176, + 428, + 243 + ], + "spans": [ + { + "bbox": [ + 266, + 176, + 428, + 243 + ], + "type": "image", + "image_path": "afdce32fb78a285bbb6342f87374635c6e7ed62ef5730bc0ab098be0a2df700b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 438, + 118, + 474, + 147 + ], + "blocks": [ + { + "bbox": [ + 438, + 118, + 474, + 147 + ], + "lines": [ + { + "bbox": [ + 438, + 118, + 474, + 147 + ], + "spans": [ + { + "bbox": [ + 438, + 118, + 474, + 147 + ], + "type": "image", + "image_path": "4b7b837aeb3abd3c022276d253d1d6aa0761af1fe5cfe8d7f98afd06b4869868.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 437, + 159, + 476, + 191 + ], + "blocks": [ + { + "bbox": [ + 439, + 149, + 473, + 158 + ], + "lines": [ + { + "bbox": [ + 439, + 149, + 473, + 158 + ], + "spans": [ + { + "bbox": [ + 439, + 149, + 473, + 158 + ], + "type": "text", + "content": "(d) Lighting" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 437, + 159, + 476, + 191 + ], + "lines": [ + { + "bbox": [ + 437, + 159, + 476, + 191 + ], + "spans": [ + { + "bbox": [ + 437, + 159, + 476, + 191 + ], + "type": "image", + "image_path": "bde0083e1d7e69c0259918e0346dc0eff44e7c7ad1adc146b61a2226a13e5f33.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 192, + 476, + 200 + ], + "lines": [ + { + "bbox": [ + 436, + 192, + 476, + 200 + ], + "spans": [ + { + "bbox": [ + 436, + 192, + 476, + 200 + ], + "type": "text", + "content": "(e) Resolution" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 436, + 202, + 476, + 232 + ], + "blocks": [ + { + "bbox": [ + 436, + 202, + 476, + 232 + ], + "lines": [ + { + "bbox": [ + 436, + 202, + 476, + 232 + ], + "spans": [ + { + "bbox": [ + 436, + 202, + 476, + 232 + ], + "type": "image", + "image_path": "65f593ddb7b468337474b2dd8a2f838564d7358c5ba6be8132ef0e21091b0f05.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 233, + 477, + 241 + ], + "lines": [ + { + "bbox": [ + 436, + 233, + 477, + 241 + ], + "spans": [ + { + "bbox": [ + 436, + 233, + 477, + 241 + ], + "type": "text", + "content": "(f) Perspective" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 388, + 480, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 480, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 480, + 448 + ], + "type": "text", + "content": "monitoring [55], and have in turn presented fascinating challenges for the computer vision community [14], including in classification [12, 29], object detection [13,31], out-of-distribution generalization [34,42,50], and active learning [54]. Such algorithms have to date, however, been designed mostly for large animals such as mammals." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": "Insect Image Datasets. As algorithms and datasets for species recognition have proliferated, insects have traditionally been underrepresented. Within ImageNet1k [59], only 27 classes depicted insects, as compared to 57 for birds and 93 classes for wild mammals [49]. Some popular fine-grained classification datasets such as NABirds and CUB have focused exclusively on birds [69, 73], while camera trap datasets such as iWildCam have been focused on mammals [12]. The iNat2017 dataset [70] included 1,021 classes of insects out of 5,089 total classes in the dataset. By largely focusing on other taxonomic groups, prior works have neglected the challenges posed by insect data, such as extremely similar classes (reflecting far greater diversity than any other group of organisms), small size, hyperabundance, etc. In recent years, however, some works have introduced insect-focused benchmark datasets. The newly released BIOSCAN-1M dataset includes images and DNA sequences, with a focus on fly identification [33]. Despite the very large size of the dataset (1.1 million datapoints), the number of object classes involved is relatively small, with the focus being to distinguish between 40 taxonomic families of flies, reflecting the challenging nature of insect identification. Other recent datasets for computer vision include one focused on 8 classes of pollinators across Europe [61], and the Indian Butterflies dataset," + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "A. Jain et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "encompassing 30k images of butterflies from India across 315 species [67]. A variety of publications have used computer vision algorithms for detection of pest insects in specific agricultural applications [10, 52, 66]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 297 + ], + "type": "text", + "content": "Insect Camera Traps. The use of camera traps for studying insect biodiversity is relatively recent. Geissmann et al. [32] introduced the Sticky Pi device, which uses sticky cards to capture (and kill) insects, which are then automatically photographed. Bjerge et al. [21] introduced an automated camera trap using UV light to attract and photograph moths and other insects (without killing them), which was built on by subsequent works [18, 20]. A similar approach focused on diurnal pollinators via a camera trap installed above flowers [17, 19]. These works also propose deep learning methods for processing the resulting images, considering both object detection and classification of a small number of insect categories. The lack of fine-grained classification is largely due to the time required to annotate images from the traps, making it difficult to obtain a large enough labeled dataset from trap data alone." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 316, + 266, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 266, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 266, + 330 + ], + "type": "text", + "content": "3 Benchmark Design" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 479, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 479, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 479, + 401 + ], + "type": "text", + "content": "In this paper, we propose a benchmark dataset for insect species identification that addresses the challenges faced in real-world use cases. These include both (1) challenges that distinguish insect identification from other kinds of species identification, notably the massive diversity of often-similar species, and (2) the particular challenges associated with deploying these algorithms \"in the wild\"." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 479, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 479, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 479, + 461 + ], + "type": "text", + "content": "To address the first family of challenges, we introduce a dataset for fine-grained insect classification that (to our knowledge) is by far the largest yet released, with " + }, + { + "bbox": [ + 130, + 402, + 479, + 461 + ], + "type": "inline_equation", + "content": "\\sim 2.5\\mathrm{M}" + }, + { + "bbox": [ + 130, + 402, + 479, + 461 + ], + "type": "text", + "content": " images across over 5,000 classes. We concentrate on moths, which are readily drawn to UV camera traps and can often be visually identified to species (unlike some other insects) [21]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "type": "text", + "content": "The second family of challenges is more complex - involving the gap between field conditions and the high-quality images typically used for training species identification algorithms. Figure 2 illustrates some of the challenges that computer vision algorithms face on data from automated insect camera traps, in contrast to images taken by humans. Insects may occlude each other or be partially \"out of frame\". Motion blur and lighting issues may occur, along with objects in unexpected positions. The resolution of images may be very low, owing to the small size of insects at a stereotyped distance from the camera. Also, insect camera traps typically use webcams and PCB camera modules due to their cost and software integration capabilities, and the resulting images are often both structurally different and overall lower quality than those taken by human photographers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "As described in the preceding section, previous authors working with insect camera traps have addressed the problem of challenging camera trap data by training algorithms on annotated data from the particular camera trap in question. However, manual annotation is extremely time-intensive, since it must be performed by expert entomologists, of whom there are very few in any given" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 315, + 91, + 447, + 101 + ], + "type": "text", + "content": "Insect Identification in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "region. Insect species vary greatly across regions, meaning that fine-grained annotations in one location are generally not useful in another. Furthermore, the hardware used in insect camera traps also varies greatly (with considerable differences in lighting and background, as well as image sharpness, contrast, noise reduction settings, and compression artifacts, which are exacerbated on small scale crops), meaning that even two datasets from the same region may not be comparable." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 283 + ], + "type": "text", + "content": "In this paper, we remove the problem of annotation by solving a more challenging computer vision problem. Namely, we propose to train algorithms on abundant labeled images of insects taken by human photographers, then use these algorithms on out-of-distribution camera trap images. Our benchmark dataset is designed to test this functionality, by including both a large training set of curated photographs taken by humans and an out-of-distribution test set drawn from a geographically diverse set of insect camera traps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 284, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 380 + ], + "type": "text", + "content": "The AMI dataset is therefore designed to test fine-grained classification of moth species both in in-distribution and out-of-distribution contexts. It also includes the task of binary classification of insects as moth or non-moth, which is a prerequisite for camera traps since moths are not the only insects to appear. Note that we do not consider the task of object detection, which in applications is a precursor to insect identification. This is because object detection against a relatively plain background is a simpler task, where good results have already been obtained in the insect camera trap literature [18,20,21,32]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 398, + 261, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 261, + 410 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 261, + 410 + ], + "type": "text", + "content": "4 The AMI Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 423, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 591 + ], + "type": "text", + "content": "In this section, we describe the construction of the datasets that we release in this work: AMI-Traps and AMI-GBIF. For AMI-Traps, we collected insect camera trap data by field deployments in three regions: NE-America (Northeast of North America), which includes data from Quebec in Canada and Vermont in the USA; W-Europe (Western Europe), which includes data from the UK and Denmark; and C-America (Central America), comprising data from Panama. For AMI-GBIF, we compiled public data for the same three regions using the Global Biodiversity Information Facility (GBIF) [7]. The GBIF platform provides access to biodiversity data aggregated from multiple sources, such as museums, research institutions, and citizen science platforms such as iNaturalist [4], Observation.org [5], and BugGuide [1]. For brevity, we sometimes refer to AMI-Traps and AMI-GBIF as AMI-T and AMI-G, respectively. Figure 3 shows examples of images from the two datasets and Tab. 1 compares the number of images and coverage of classes at different taxonomic levels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 609, + 261, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 261, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 261, + 621 + ], + "type": "text", + "content": "4.1 AMI-Traps Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Data Collection. The source images in the AMI-Traps dataset were collected from a variety of different insect camera traps deployed across 22 monitoring stations in the three regions previously described. The cameras were configured" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "A. Jain et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 482, + 331 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 482, + 331 + ], + "type": "image", + "image_path": "6d3d09004541e7bbf9f29d11bdc159376009b17c5cc75bd7e740c4a592d92ee7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 340, + 482, + 396 + ], + "lines": [ + { + "bbox": [ + 132, + 340, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 340, + 482, + 396 + ], + "type": "text", + "content": "Fig. 3: Sample images from the AMI Dataset. The AMI Dataset is composed of (1) AMI-GBIF, curated from a number of sources with imagery from citizen science and museum collections, and (2) AMI-Traps, drawn from automated camera traps for insects across five countries in three regions. The number of individual insects annotated per sub-dataset is denoted in parentheses." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 422, + 480, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 480, + 471 + ], + "type": "text", + "content": "to be motion-sensitive, as well as taking photos on a fixed interval of 10 minutes; we selected a subset of images for annotation so as to avoid highly similar images. In total, 2,893 trap images were sampled from all cameras and made available on the Label Studio annotation platform (see Appendix B for more details)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "text", + "content": "Annotation Process. Images were annotated by entomologists with expertise in the species of the particular region. Labeling was conducted according a three-step process. 1) We designed a custom object detection model to suggest boxes around objects in the image (refer to Appendix B for details on this model), to save annotators time. Annotators were asked to correct errors in the proposed boxes as well as drawing any missing boxes. Insects appearing smaller than 1 centimeter were ignored. 2) Annotators were asked to label image crops according to one of three coarse categories: moth, non-moth, or unidentifiable. Some insects could not be labeled due to motion blur, occlusion, or other factors. 3) For insects labeled as moths, annotators were asked to select the narrowest taxonomic group that could be determined from the image - identifying moths to species where possible, if not species then genus, and if not genus then family." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Data Statistics. 2,893 source images from the traps were annotated (see Tab. B.2 in the supplementary for further details) which includes Vermont (900), Denmark (892), Quebec (533), UK (446), and Panama (122). A total of 52,948 insects were labeled within these images, including 37,105 " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "(70.08\\%)" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " non-moths," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 315, + 91, + 447, + 101 + ], + "type": "text", + "content": "Insect Identification in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": "14,105 (26.64%) moths, and 1,738 (3.28%) unidentifiable. Out of the 14,105 moth crops, 10,854 (76.95%), 7,065 (50.09%) and 5,374 (38.10%) crops were labeled at the family, genus, and species level respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 168, + 261, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 168, + 261, + 180 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 261, + 180 + ], + "type": "text", + "content": "4.2 AMI-GBIF Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 186, + 481, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 186, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 481, + 235 + ], + "type": "text", + "content": "AMI-GBIF has two sub-datasets: binary and fine-grained classification. The fine-grained classification dataset contains images for moth species in each of the regions covered by AMI-Traps. The binary classification dataset contains images of moths aggregated across all regions and non-moths from anywhere on earth." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 235, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 482, + 402 + ], + "type": "text", + "content": "Dataset Construction. First, in collaboration with regional moth specialists, we created \"checklists\" of species that could occur in each of the three regions of interest. It is worth noting that some species are included in more than one checklist, though most are unique to a single region. Next, we compared the checklists against the GBIF taxonomic backbone [2], merged synonym names (since some species have multiple accepted names), and manually investigated doubtful or unmatched names. Following that, we used the GBIF occurrence search tool to download metadata for all observations from the order Lepidoptera with images, which returned 15.1 million observations [3]. We then fetched images from observations in which the species is included in our regional checklists. The number of training images per species were limited to 1,000, though for many species it was considerably lower due to the long-tailed nature of the data. All images were resized to have the smallest dimension of 450px and packaged using the WebDataset format [6]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 617 + ], + "type": "text", + "content": "Quality Control. After downloading the images, we applied a set of cleaning steps to ensure the quality of the dataset. First, we removed duplicate URLs, i.e., images linked to more than one observation. This may occur when an image contains multiple individual insects, or when a placeholder is used for an observation (for example, Fig. A.1b in Appendix A shows a picture used as a placeholder for more than 100,000 observations). Next, we excluded images from specific datasets (refer to Tab. A.1 in Appendix A)) that we found to have a high percentage of images unsuitable for training our models, for instance, images of body parts or text only. Another issue we identified is that some observations contained thumbnail images that are too small to provide sufficient information to train our models; we removed all images with a width or height less than 64 pixels. Finally, as our dataset targets only adult insects, we removed images containing non-adult individuals such as eggs and caterpillars. For observations where the life stage was reported by the observer, approximately " + }, + { + "bbox": [ + 130, + 402, + 482, + 617 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 402, + 482, + 617 + ], + "type": "text", + "content": " of the cases, we removed all occurrences tagged as non-adult. For the remaining images with a missing life stage tag, we designed a classification model to predict the life stage (refer to Appendix A for details about this model) and removed all images classified as non-adult." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Dataset Split. Each GBIF observation may contain more than one image, but each image is considered an independent instance of the species. However, as images belonging to the same observation may be very similar, we keep images from the same observation in the same data split. To do so, we split the dataset" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "A. Jain et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "using the observation ID as reference, with the proportion of " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " for the train, validation and test sets respectively. We apply a stratified split, i.e., we guarantee that this proportion is also maintained at the species level. We exclude species with fewer than 5 images, so the minimum number of images for a species in the training set is 4, and in the test set is one. Due to the data split based on observation ID, the validation set might lack images for some species. As some species may belong to multiple checklists, we keep the train/val/test splits consistent across the regions to prevent data leakage when training a unified model for all regions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 391 + ], + "type": "text", + "content": "Binary Classification. Since moths are not the only insects that appear on the camera trap screen, for our proposed benchmark task of binary classification, we curate a dataset comprising 350,000 images for each of the moth and non-moth categories. The moth class consists of adult moth images randomly sampled from the AMI-GBIF dataset. For the non-moth category, we leveraged expert knowledge of the taxonomic groups likely to appear on the screen, fetching images of adult individuals from GBIF of the following groups: Diptera (85,474), Hemiptera (75,053), Odonata (68,770), Coleoptera (45,954), Araneae (21,321), Orthoptera (21,261), Formicidae (12,206), Ichneumonidae (6,581), Trichoptera (5,316), Neuroptera (4,880), Opiliones (1,880), Ephemeroptera (1,304) — this includes mostly insects, with some Arachnids. The dataset is split into train, validation, and test sets with proportions of " + }, + { + "bbox": [ + 130, + 224, + 482, + 391 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 224, + 482, + 391 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 224, + 482, + 391 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 224, + 482, + 391 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 224, + 482, + 391 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 224, + 482, + 391 + ], + "type": "text", + "content": ", respectively. Specifically for the moth category, the splits are kept consistent with the splits for the fine-grained species classification task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 392, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 559 + ], + "type": "text", + "content": "Statistics. The final dataset contains " + }, + { + "bbox": [ + 130, + 392, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\sim 2.5\\mathrm{M}" + }, + { + "bbox": [ + 130, + 392, + 482, + 559 + ], + "type": "text", + "content": " images from 5,364 species, 1,734 genera and 77 families of moths, along with 350,000 additional images of non-moth groups. The binary dataset contains 700,000 images, composed of 350,000 moth images and the 350,000 non-moth images. The fine-grained dataset contains 2,564,392 moth images, and Tab. 1 shows the number of images and species for each region. Though species are selected based on their presence in regions of interest, we include all GBIF observations of these species from anywhere on Earth (see Fig. A.2 in supplementary for the global distribution of images). As in other biodiversity datasets, the number of images per species in AMI-GBIF follows a long-tailed distribution (see Fig. C.3 in supplementary). Specifically, only 3,099 species have more than 100 images and are considered under the \"many-shot\" bucket, 1,358 species contains between 20 and 100 images and are considered \"medium-shot\", and there are 907 species with less than 20 training images, considered as \"few-shot\" categories." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 578, + 207, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 207, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 207, + 592 + ], + "type": "text", + "content": "5 Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "In this section, we describe the methods used to assess the difficulty of the proposed benchmark tasks. We evaluate a variety of high-performing image classification architectures and use extensive data augmentation to address the challenges of domain shift between the two datasets. Motivated by the particular structure of the AMI-Traps dataset, we add a mixed-resolution (MixRes) data" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 316, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 316, + 91, + 447, + 101 + ], + "type": "text", + "content": "Insect Identification in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 179, + 466, + 249 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 1: Comparison between the AMI-GBIF and AMI-Traps dataset for the fine-grained classification task. Images for the traps dataset denote the individual insect crops and not the raw trap images. The last three rows show the unique classes for each of the three taxonomic levels for the two datasets for different regions. It is also important to note that regions have species in common." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 179, + 466, + 249 + ], + "lines": [ + { + "bbox": [ + 148, + 179, + 466, + 249 + ], + "spans": [ + { + "bbox": [ + 148, + 179, + 466, + 249 + ], + "type": "table", + "html": "
All RegionsNE-AmericaW-EuropeC-America
AMI-GAMI-TAMI-GAMI-TAMI-GAMI-TAMI-GAMI-T
Images2,564,39214,1051,179,9439,0661,579,3334,06699,405973
Families7743683465262711
Genera1,73434488720495719238512
Species5,3645162,4972772,6032446366
", + "image_path": "cf4692da1b0ca236aaa7751bf606018707568c1a42204c55a51125fd9061c300.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 277, + 479, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 479, + 312 + ], + "type": "text", + "content": "augmentation technique. As we describe in the following section, our approaches perform quite well, with MixRes leading to an especially significant performance improvement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 316, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 518 + ], + "type": "text", + "content": "Models. We work with five computer vision models widely used in identification of wildlife: ResNet50 [35], MobileNetV3-Large [37], ConvNeXt-T [47], ConvNeXt-B [47], and ViT-B/16 [28]. We select ResNet50 because it is the most popular backbone used as a baseline for almost any computer vision task. Considering the possibility that the trained models may need to run on hardware with limited resources, such as edge devices and personal computers, for in-field predictions, we choose MobileNetV3Large as a lightweight model alternative. Furthermore, considering that vision transformers have surpassed convolutional neural networks and become the state-of-the-art models for image classification, we select the ViT-B/16 model to provide a baseline from this family of architectures. Finally, for representing modern convolutional neural networks, we choose the recently introduced ConvNeXt architecture, which has proven to be competitive with transformers using only standard ConvNet modules. From this family of models, we select two versions: the base version (ConvNeXt-B) and, considering lightweight usage scenarios, we also include the tiny version (ConvNeXt-T). Unless otherwise stated, we use a default input resolution of 128x128 for all models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Data Augmentation Approaches. One characteristic of the in-the-wild images is the large variety of crops due to the diversity of insect sizes. A considerable number of them have small resolutions, below " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ". At this scale, some important visual features that the model may have learned from high-resolution GBIF images might not be identifiable, leading to inaccurate predictions. This issue can be further amplified when these low-resolution crops are scaled up to the input model size. To compel the model to learn features for low-resolution inputs and avoid bias towards visual features only distinguishable in high-resolution images from other domains, we use a data augmentation technique we refer to as Mixed Resolution (MixRes) that simulates this low-resolution input during training (see Fig. D.4 in supplementary). This approach consists of randomly resizing the training input image to " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " of the model input size with a" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "A. Jain et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 480, + 152 + ], + "type": "text", + "content": "probability " + }, + { + "bbox": [ + 134, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 134, + 116, + 480, + 152 + ], + "type": "text", + "content": " for each case. In our experiments " + }, + { + "bbox": [ + 134, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 134, + 116, + 480, + 152 + ], + "type": "text", + "content": " is set to 0.25, maintaining the original input size with probability 0.5. We employ MixRes in all experiments unless otherwise stated." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 156, + 480, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 156, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 134, + 156, + 480, + 228 + ], + "type": "text", + "content": "In addition to MixRes, we also use a set of standard data augmentation operations: random resize crop with scale of 0.3 to 1.0 and an aspect ratio between " + }, + { + "bbox": [ + 134, + 156, + 480, + 228 + ], + "type": "inline_equation", + "content": "3/4" + }, + { + "bbox": [ + 134, + 156, + 480, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 134, + 156, + 480, + 228 + ], + "type": "inline_equation", + "content": "4/3" + }, + { + "bbox": [ + 134, + 156, + 480, + 228 + ], + "type": "text", + "content": ", random horizontal flip with " + }, + { + "bbox": [ + 134, + 156, + 480, + 228 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 134, + 156, + 480, + 228 + ], + "type": "text", + "content": " probability, and RandAugment [24] with " + }, + { + "bbox": [ + 134, + 156, + 480, + 228 + ], + "type": "inline_equation", + "content": "N = 2" + }, + { + "bbox": [ + 134, + 156, + 480, + 228 + ], + "type": "text", + "content": " operations and magnitude " + }, + { + "bbox": [ + 134, + 156, + 480, + 228 + ], + "type": "inline_equation", + "content": "M = 9" + }, + { + "bbox": [ + 134, + 156, + 480, + 228 + ], + "type": "text", + "content": ". Finally, the image is normalized using the mean and standard deviation of ImageNet and scaled to the input network size." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 233, + 480, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 233, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 134, + 233, + 480, + 327 + ], + "type": "text", + "content": "Implementation Details. We train all models with the AMI-GBIF training sets using the ImageNet-1K pretrained weights from the timm library [76] (for ConvNext and ViT we used ImageNet-22K weights). Each model is trained with a batch size of 128 and the AdamW optimizer [48] for 30 epochs. The initial learning rate is set to 0.001, with a weight decay of 1e-5. The learning rate is linearly warmed up from 0 to the initial value for 2 epochs and then decayed using cosine scheduling [36]. We also apply label smoothing [65] with a value of 0.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 333, + 480, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 333, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 134, + 333, + 480, + 392 + ], + "type": "text", + "content": "Inference Settings. At test time, the insect crops from AMI-Traps are padded with black pixels to make them square, preventing the distortion of the shape of the insect before prediction. Next, the images are normalized with the mean and standard deviation of ImageNet and, finally, resized to the model input size. No padding is done for the test images from AMI-GBIF." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 397, + 480, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 397, + 480, + 505 + ], + "spans": [ + { + "bbox": [ + 134, + 397, + 480, + 505 + ], + "type": "text", + "content": "Higher Taxonomic Rank Prediction. We consider the accuracy of our baseline models both with respect to the exact insect species predicted and with respect to higher level taxonomic groups such as genus and family. Indeed some images in our dataset cannot be labeled at the species level at all due to the quality or perspective of the image, and our annotators instead provided labels at the genus or family level. In our experiments, to make predictions at a higher taxonomic level, we simply sum together species-level predictions to get genus results, and similarly, the genus predictions are grouped together to get family predictions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 510, + 480, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 510, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 134, + 510, + 480, + 664 + ], + "type": "text", + "content": "Evaluation Metrics. During the binary classification step, it is important to achieve high precision by filtering out as many non-moths as possible to prevent the propagation of errors in the pipeline. At the same time, it is also important to assess recall because if too many moths are discarded at this step, it may impact the utility of the data for ecological studies. Therefore, for the binary classification task, along with accuracy, we also consider precision, recall, and F1-score as evaluation metrics. For the fine-grained classification task, we evaluate the model in terms of micro- and macro-averaged accuracy at the species, genus and family levels. Micro-averaging aggregates all class samples to calculate the accuracy while the macro-average is the average of per-class accuracy. Unless otherwise stated, the reported accuracy is micro-averaged. We conduct each experiment for three independent runs and report the mean and standard deviation." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 316, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 316, + 91, + 447, + 101 + ], + "type": "text", + "content": "Insect Identification in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 303, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 303, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 303, + 129 + ], + "type": "text", + "content": "6 Experiments and Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 285 + ], + "type": "text", + "content": "Binary Classification Results. This task involves distinguishing moths from non-moths. Here we train a ResNet50 model on the AMI-GBIF dataset and evaluate performance both on the AMI-GBIF held-out test set and on the entirety of the AMI-Traps dataset. On AMI-GBIF (in-distribution), the model achieves a test accuracy of " + }, + { + "bbox": [ + 130, + 141, + 482, + 285 + ], + "type": "inline_equation", + "content": "97.77\\% \\pm 0.02\\%" + }, + { + "bbox": [ + 130, + 141, + 482, + 285 + ], + "type": "text", + "content": ". Table 2 shows the results for the AMI-Traps dataset in more detail. We observe that the model has a high recall but a lower precision, which means it is over-predicting the moth class. Also, the bigger the image crop, the better the accuracy of the insect classification; many non-moths where the model errs are extremely small and are therefore highly blurred in the trap images. Table E.3 in the supplementary illustrates the improvement in model performance when crops are square padded (our default at inference time) versus an ablation where padding is not performed." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 360, + 481, + 403 + ], + "blocks": [ + { + "bbox": [ + 130, + 306, + 482, + 350 + ], + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 350 + ], + "type": "text", + "content": "Table 2: Binary classification model evaluation results on the AMI-Traps dataset after training on AMI-GBIF, showing that recall is high throughout, while precision drops on very small insects, which are predominantly non-moths. Crops " + }, + { + "bbox": [ + 130, + 306, + 482, + 350 + ], + "type": "inline_equation", + "content": ">100\\mathrm{px}" + }, + { + "bbox": [ + 130, + 306, + 482, + 350 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 306, + 482, + 350 + ], + "type": "inline_equation", + "content": ">150\\mathrm{px}" + }, + { + "bbox": [ + 130, + 306, + 482, + 350 + ], + "type": "text", + "content": " represent that atleast one side of the box is greater than the threshold." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 360, + 481, + 403 + ], + "lines": [ + { + "bbox": [ + 133, + 360, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 133, + 360, + 481, + 403 + ], + "type": "table", + "html": "
MothsNon-mothsTotalAccuracyPrecisionRecallF1-score
All crops14,10537,10551,12086.48±1.0068.31±1.7095.03±0.3979.48±1.27
Crops > 100px11,63016,53428,16488.39±0.9779.46±1.4396.99±0.2087.35±0.94
Crops > 150px6,4465,96112,40792.95±0.2289.83±0.4397.48±0.1293.70±0.30
", + "image_path": "8650a1fd72c60892336133d2d9c26a31e49babb0317771eb2b06cb36c8805ad4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": "Fine-grained Classification Results. The fine-grained classification task aims to identify moths to the species level (and secondarily to genus or family level). In our experiments, we train models on AMI-GBIF data for each of the three regions of the AMI dataset, evaluating both on AMI-GBIF (in-distribution) and AMI-Traps (out-of-distribution) data (see Tab. 3). Not all moth species present in a region appear regularly at camera traps; hence, the list of species in AMI-Traps is a subset of species in AMI-GBIF for a given area (see Tab. 1). Therefore, we also evaluate the models on an in-distribution subset " + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathrm{Test}\\ddagger)" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": " from AMI-GBIF containing only the species present in AMI-Traps. Meanwhile, for AMI-Traps, there are 16 species not present in AMI-GBIF that were removed during our evaluation procedure (these represent unexpected species not included in the regional checklists used to define AMI-GBIF). Overall, the best model at the species level is ConvNeXt-B for both AMI-GBIF and AMI-traps. For C-America, annotations are mostly provided at the genus or family level, due to the challenging nature of visual identification of moths in this biodiversity-rich region. Furthermore, while the genus and family accuracies are similar between N-America and W-Europe, they are much lower for C-America. This can be attributed to the fact that there are 10 times fewer images in the training set for this region. Note that for genus and family evaluation, there are some cases where the accuracy is lower than the species level. This occurs because those" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "A. Jain et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 190, + 481, + 335 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 3: Accuracy results for the fine-grained classification task. Overall, ConvNeXt-B achieves the best results for both AMI-GBIF and AMI-traps, with ViT-B/16 achieving slightly better results for some cases at the genus and family levels. Test‡ represents a subset of the AMI-GBIF test set, containing only the species present in AMI-Traps. Species-level results on AMI-Traps for the C-America region are not available since most annotations were provided only at higher taxonomic levels." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 190, + 481, + 335 + ], + "lines": [ + { + "bbox": [ + 132, + 190, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 190, + 481, + 335 + ], + "type": "table", + "html": "
RegionModelAMI-GBIFAMI-Traps
SpeciesSpeciesGenusFamily
Micro TestMicro Test‡Micro Top1Macro Top1Micro Top1Macro Top1Micro Top1Macro Top1
NE-AmericaResNet5087.56±0.1289.39±0.1471.86±1.0264.96±2.0080.07±0.4974.66±0.3590.44±0.4169.30±4.15
MBNetV3L83.73±0.1785.80±0.2967.41±0.7261.21±1.2074.83±1.3071.20±2.4789.31±0.6264.48±1.87
ConvNeXt-T89.81±0.0591.21±0.0674.89±0.0568.00±0.7683.26±1.2277.26±0.8590.61±0.5374.60±0.64
ConvNeXt-B90.45±0.0391.66±0.0376.75±1.5268.56±0.6285.46±0.8378.28±0.1491.07±0.2976.87±1.10
ViT-B/1686.39±0.0587.94±0.1077.43±0.7968.02±0.3485.57±0.9878.97±0.7291.45±0.3473.66±1.49
W-EuropeResNet5086.45±0.0890.25±0.0379.39±0.3972.84±0.2178.95±0.2673.22±0.6089.13±0.4275.24±1.85
MBNetV3L82.20±0.0586.47±0.1376.30±1.2970.35±0.9376.35±1.1972.88±1.2888.33±0.4772.21±0.38
ConvNeXt-T88.82±0.0392.05±0.0582.44±0.8974.59±0.6281.86±0.7974.15±0.5990.08±0.5974.58±2.93
ConvNeXt-B89.35±0.0192.48±0.0582.87±0.3976.41±0.5082.45±0.4475.98±1.0591.27±0.6079.39±2.70
ViT-B/1685.45±0.0389.22±0.1182.44±0.3675.25±0.9883.01±0.2675.96±0.3591.68±0.1775.81±0.79
C-AmericaResNet5087.46±0.2395.92±0.68--47.44±1.1166.59±0.6069.18±0.2442.40±0.31
MBNetV3L83.68±0.2492.89±0.27--41.03±2.2259.32±1.1868.34±1.6941.42±2.98
ConvNeXt-T87.65±0.1195.46±0.46--55.77±3.3370.68±1.5869.46±0.4245.35±0.67
ConvNeXt-B88.62±0.0996.22±0.26--51.92±3.8568.63±1.9472.24±1.4744.31±1.32
ViT-B/1679.80±0.2688.36±1.02--47.44±4.4464.70±4.0769.87±1.5142.00±3.60
", + "image_path": "6bdaeca35aa9620bae8bfbffb510c666742eadef579b4842e70fc4724076e6d2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 358, + 479, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 479, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 479, + 407 + ], + "type": "text", + "content": "higher taxonomic levels include crops that have labels only at higher levels. In the supplementary material, we include additional results for AMI-GBIF long-tail accuracies (Tab. E.4), a pre-training ablation study (Tab. E.5), and a study analyzing classification accuracy as a function of prediction confidence (Fig. E.5)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 407, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 597 + ], + "type": "text", + "content": "Data augmentation approaches. To assess the effectiveness of the mixed resolution data augmentation technique, we conduct an ablation study using ResNet50 for the NE-America region. As the effects of MixRes could already be covered by other data augmentation operations used in our preprocessing, such as random resize crop or one of the operations of RandAugment, we analyze multiple combinations of these operations. Additionally, we evaluate a higher input resolution, using the original 224 input size of ResNet50, to determine whether this leads to better predictions. Table 4 shows the results. While applying MixRes results in a slight decrease in accuracy within the in-distribution AMI-GBIF test set, for the in-the-wild AMI-Traps test set there is a substantial improvement, ranging between " + }, + { + "bbox": [ + 130, + 407, + 482, + 597 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 130, + 407, + 482, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 407, + 482, + 597 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 130, + 407, + 482, + 597 + ], + "type": "text", + "content": " when compared to models not trained with MixRes. Another important observation is that, although RandAugment and random resize together contribute to an overall improvement of around " + }, + { + "bbox": [ + 130, + 407, + 482, + 597 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 130, + 407, + 482, + 597 + ], + "type": "text", + "content": ", MixRes remains additionally beneficial. Finally, despite an increase of more than " + }, + { + "bbox": [ + 130, + 407, + 482, + 597 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 130, + 407, + 482, + 597 + ], + "type": "text", + "content": " in accuracy within the original domain when using a higher resolution, there is a marginal decrease in the accuracy of the in-the-wild results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 616, + 220, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 220, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 220, + 628 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Many ecological applications require species-level image classification to meet their goals. Despite advances in computer vision, there remain challenges in" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 316, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 316, + 91, + 447, + 101 + ], + "type": "text", + "content": "Insect Identification in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 192, + 178, + 422, + 247 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 4: Ablation study on data augmentation techniques using a ResNet50 for the NE-America region evaluated on micro-averaged top-1 species accuracy. \"RandAug\" denotes the combination of the RandAugment and random resize crop data augmentation techniques. AMI-GBIF† test set contains only the species present in AMI-Traps." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 192, + 178, + 422, + 247 + ], + "lines": [ + { + "bbox": [ + 192, + 178, + 422, + 247 + ], + "spans": [ + { + "bbox": [ + 192, + 178, + 422, + 247 + ], + "type": "table", + "html": "
ResolutionRandAugMixResAMI-GBIF‡AMI-Traps
128××89.18±0.0851.46±2.73
128✓×89.71±0.1667.88±1.27
128×✓87.41±0.0658.41±0.30
128✓✓89.39±0.1471.86±1.02
224✓✓92.97±0.1768.41±1.06
", + "image_path": "f6c860a8441da95dff8cbccbac4a26ab243dad7232acb9d117ff1fd7b9021d02.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 300, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 431 + ], + "type": "text", + "content": "implementing fine-grained classification that is effective in real-world scenarios. Here we outline the challenges of insect identification in the wild and propose a benchmark dataset to tackle them. We present AMI-GBIF, a dataset with 2.5 million labeled photographs of 5,000 moth species. We also introduce AMI-Traps, a dataset of images captured by automated insect monitoring systems across five countries, labeled by regional experts. We evaluate multiple classification models both on the AMI-GBIF dataset and in out-of-distribution generalization to AMI-Traps, with ConvNeXt-B generally performing the best. We also evaluate a range of data augmentation techniques, finding that a combination of simple approaches, notably mixed-resolution augmentation, considerably improves generalization under field conditions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 441, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 561 + ], + "type": "text", + "content": "Several limitations of this work should be noted. First, while annotating AMI-Traps already occupied a considerable time for expert entomologists, further expanding it to a greater number of species and geographic regions would be helpful in further stress-testing field deployments of fine-grained classification algorithms. Second, while we focus on camera traps for nocturnal insects, it would be useful also to consider diurnal insect camera traps, for example those aimed at pollinators, which pose some additional challenges, such as multiple perspectives and a diversity of backgrounds. We hope furthermore that future works will consider more sophisticated domain adaptation approaches building on the baselines we present here." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "This work represents a step towards computer vision algorithms for species recognition that are effective in the ultra-challenging regime of insect identification and work effectively in field conditions. The approach we propose should make it possible in particular to generalize algorithms for insect camera traps rapidly and cost-effectively across different hardware setups and regions of interest. We hope that the datasets, challenges and strategies presented here will propel further research in computer vision for insects and help accelerate initiatives to understand and protect our planet's biodiversity." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "A. Jain et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 479, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 479, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 479, + 164 + ], + "type": "text", + "content": "We thank Annie-Shan Morin, Anna Viklund, and Jamie Alison for their helpful input and support to the project." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 164, + 481, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 481, + 319 + ], + "type": "text", + "content": "This research was generously funded by Mila Quebec AI Institute and Ministère de l'Économie, de l'Innovation et de l'Energie. The work was additionally supported in part by the Canada CIFAR AI Chairs program (NSERC 585136), the Global Climate Center on AI and Biodiversity Change (ABC) (NSF OISE2330423), the Montreal Insectarium, the Coordenação de Aperfeçoamento de Pessoal de Nível Superior - Brasil (CAPES-PROEX) - Finance Code 001, the Amazonas State Research Support Foundation - FAPEAM - through the POSGRAD project 2024/2025, the AMBER project funded by the abrdn Charitable Foundation, the Aage V. Jensen Naturfond grant no. N2024-0013, the European Union's Horizon Europe Research and Innovation programme under grant agreement no. 101060639 (MAMBO), and a Discovery Grant by NSERC. The computing resources were provided by Mila Quebec AI Institute, the Digital Research Alliance of Canada, and NVIDIA Corporation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 481, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 380 + ], + "type": "text", + "content": "A.J., F.C., and M.J.B. led dataset creation and experiments. J.S.C. and L.P. contributed to machine learning pipeline development. N.P., F.H, J.R., M.B., M.S., J.F., A.A., Y.L., E.N., F.P.P., A.C.Z., J.A.R.S., and J.G. performed insect identification for the AMI-Traps dataset. T.A, K.B., A.G.S., M.B., Y.B., K.P.M, D. Roy, T.T.H., M.L., and D. Rolnick guided and advised the project." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 398, + 198, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 198, + 410 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 198, + 410 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 423, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 423, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 481, + 456 + ], + "type": "text", + "content": "1. BugGuide.Net: Identification, Images, & Information For Insects, Spiders & Their Kin For the United States & Canada, https://bugguide.net/ accessed Feb 12 2024" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 456, + 481, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 481, + 477 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 481, + 477 + ], + "type": "text", + "content": "2. GBIF Backbone Taxonomy, https://doi.org/10.15468/39omei accessed Nov 13 2023" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 479, + 481, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 481, + 499 + ], + "type": "text", + "content": "3. GBIF Occurrence Download, https://doi.org/10.15468/dl.4ua6wg accessed Oct 19 2023" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 501, + 375, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 375, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 375, + 511 + ], + "type": "text", + "content": "4. iNaturalist, www.inaturalist.org, accessed Feb 9 2024" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 512, + 420, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 420, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 420, + 522 + ], + "type": "text", + "content": "5. Observation.org, https://observation.org, accessed Feb 12 2024" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 523, + 481, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 481, + 543 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 481, + 543 + ], + "type": "text", + "content": "6. WebDataset, https://github.com/webdataset/webdataset, accessed on April 2 2024" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 544, + 464, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 464, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 464, + 555 + ], + "type": "text", + "content": "7. What is GBIF?, https://www.gbif.org/what-is-gbif accessed Sep 28 2023" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 556, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 481, + 588 + ], + "type": "text", + "content": "8. Alison, J., Alexander, J.M., Diaz Zeugin, N., Dupont, Y.L., Iseli, E., Mann, H.M., Hoye, T.T.: Moths complement bumblebee pollination of red clover: a case for day-and-night insect surveillance. Biology Letters 18(7), 20220187 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 481, + 621 + ], + "type": "text", + "content": "9. Anton, V., Hartley, S., Geldenhuis, A., Wittmer, H.U.: Monitoring the mammalian fauna of urban areas using remote cameras and citizen science. Journal of Urban Ecology 4(1), juy002 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 621, + 481, + 665 + ], + "type": "text", + "content": "10. Badgujar, C.M., Armstrong, P.R., Gerken, A.R., Pordesimo, L.O., Campbell, J.F.: Real-time stored product insect detection and identification using deep learning: System integration and extensibility to mobile platforms. Journal of Stored Products Research 104, 102196 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 315, + 91, + 447, + 101 + ], + "type": "text", + "content": "Insect Identification in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "11. Beery, S.: Scaling biodiversity monitoring for the data age. XRDS: Crossroads, The ACM Magazine for Students 27(4), 14-18 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "type": "text", + "content": "12. Beery, S., Agarwal, A., Cole, E., Birodkar, V.: The iwildcam 2021 competition dataset. arXiv preprint arXiv:2105.03494 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 163, + 480, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 163, + 480, + 184 + ], + "spans": [ + { + "bbox": [ + 133, + 163, + 480, + 184 + ], + "type": "text", + "content": "13. Beery, S., Morris, D., Yang, S.: Efficient pipeline for camera trap image review. arXiv preprint arXiv:1907.06772 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 186, + 480, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 186, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 133, + 186, + 480, + 206 + ], + "type": "text", + "content": "14. Beery, S., Van Horn, G., Perona, P.: Recognition in terra incognita. In: Proceedings of the European conference on computer vision (ECCV). pp. 456-473 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 208, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 208, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 133, + 208, + 480, + 262 + ], + "type": "text", + "content": "15. Beery, S., Wu, G., Edwards, T., Pavetic, F., Majewski, B., Mukherjee, S., Chan, S., Morgan, J., Rathod, V., Huang, J.: The auto arborist dataset: A large-scale benchmark for multiview urban forest monitoring under domain shift. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21294-21307 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 263, + 480, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 263, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 133, + 263, + 480, + 306 + ], + "type": "text", + "content": "16. Berg, T., Liu, J., Woo Lee, S., Alexander, M.L., Jacobs, D.W., Belhumeur, P.N.: Birdsnap: Large-scale fine-grained visual categorization of birds. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2011-2018 (2014)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 308, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 308, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 133, + 308, + 480, + 341 + ], + "type": "text", + "content": "17. Bjerge, K., Alison, J., Dyrmann, M., Frigaard, C.E., Mann, H.M., Hoye, T.T.: Accurate detection and identification of insects from camera trap images with deep learning. PLOS Sustainability and Transformation 2(3), e0000051 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 342, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 342, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 133, + 342, + 480, + 363 + ], + "type": "text", + "content": "18. Bjerge, K., Frigaard, C.E., Karstoft, H.: Object detection of small insects in timelapse camera recordings. Sensors 23(16), 7242 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 365, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 365, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 133, + 365, + 480, + 396 + ], + "type": "text", + "content": "19. Bjerge, K., Geissmann, Q., Alison, J., Mann, H.M., Hoye, T.T., Dyrmann, M., Karstoft, H.: Hierarchical classification of insects with multitask learning and anomaly detection. Ecological Informatics 77, 102278 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 398, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 398, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 133, + 398, + 480, + 430 + ], + "type": "text", + "content": "20. Bjerge, K., Mann, H.M., Hoye, T.T.: Real-time insect tracking and monitoring with computer vision and deep learning. Remote Sensing in Ecology and Conservation 8(3), 315-327 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 432, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 432, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 133, + 432, + 480, + 464 + ], + "type": "text", + "content": "21. Bjerge, K., Nielsen, J.B., Sepstrup, M.V., Helsing-Nielsen, F., Hoye, T.T.: An automated light trap to monitor moths (lepidoptera) using computer vision-based tracking and deep learning. Sensors 21(2), 343 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 465, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 465, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 133, + 465, + 480, + 508 + ], + "type": "text", + "content": "22. Boom, B.J., He, J., Palazzo, S., Huang, P.X., Bayan, C., Chou, H.M., Lin, F.P., Spampinato, C., Fisher, R.B.: A research tool for long-term and continuous analysis of fish assemblage in coral-reefs using underwater camera footage. Ecological Informatics 23, 83-97 (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 510, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 510, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 133, + 510, + 480, + 542 + ], + "type": "text", + "content": "23. Borowiec, M.L., Dikow, R.B., Frandsen, P.B., McKeeken, A., Valentini, G., White, A.E.: Deep learning as a tool for ecology and evolution. Methods in Ecology and Evolution 13(8), 1640-1660 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 544, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 544, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 133, + 544, + 480, + 586 + ], + "type": "text", + "content": "24. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 702-703 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 588, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 588, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 133, + 588, + 480, + 631 + ], + "type": "text", + "content": "25. Cutter, G., Stierhoff, K., Zeng, J.: Automated detection of rockfish in unconstrained underwater videos using haar cascades and a new image dataset: Labeled fishes in the wild. In: 2015 IEEE Winter Applications and Computer Vision Workshops. pp. 57-62. IEEE (2015)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "text", + "content": "26. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "A. Jain et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "27. Didham, R.K., Basset, Y., Collins, C.M., Leather, S.R., Littlewood, N.A., Menz, M.H., Müller, J., Packer, L., Saunders, M.E., Schonrogge, K., et al.: Interpreting insect declines: seven challenges and a way forward. Insect Conservation and Diversity 13(2), 103-114 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 482, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 482, + 225 + ], + "type": "text", + "content": "28. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net (2021), https://openreview.net/forum?id=YicbFdNTTy" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 227, + 482, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 482, + 258 + ], + "type": "text", + "content": "29. Dubey, A., Ramanathan, V., Pentland, A., Mahajan, D.: Adaptive methods for real-world domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14340-14349 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 259, + 481, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 292 + ], + "type": "text", + "content": "30. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision 88, 303-338 (2010)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "text", + "content": "31. Fennell, M., Beirne, C., Burton, A.C.: Use of object detection in camera trap image identification: Assessing a method to rapidly and accurately classify human and animal detections for research and application in recreation ecology. Global Ecology and Conservation 35, e02104 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "32. Geissmann, Q., Abram, P.K., Wu, D., Haney, C.H., Carrillo, J.: Sticky pi is a high-frequency smart trap that enables the study of insect circadian activity under natural conditions. PLoS Biology 20(7), e3001689 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 369, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 413 + ], + "type": "text", + "content": "33. Gharaee, Z., Gong, Z., Pellegrino, N., Zarubiieva, I., Haurum, J.B., Lowe, S.C., McKeown, J.T., Ho, C.C., McLeod, J., Wei, Y.Y.C., et al.: A step towards worldwide biodiversity assessment: The bioscan-1m insect dataset. arXiv preprint arXiv:2307.10455 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 414, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 434 + ], + "type": "text", + "content": "34. Gulrajani, I., Lopez-Paz, D.: In search of lost domain generalization. arXiv preprint arXiv:2007.01434 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "35. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "36. He, T., Zhang, Z., Zhang, H., Zhang, Z., Xie, J., Li, M.: Bag of tricks for image classification with convolutional neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 558-567 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "text", + "content": "37. Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al.: Searching for mobilenetv3. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1314-1324 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 534, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 577 + ], + "type": "text", + "content": "38. Hoye, T.T., Arje, J., Bjerge, K., Hansen, O.L., Iosifidis, A., Leese, F., Mann, H.M., Meissner, K., Melvad, C., Raitoharju, J.: Deep learning and computer vision will transform entomology. Proceedings of the National Academy of Sciences 118(2), e2002545117 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 577, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 621 + ], + "type": "text", + "content": "39. Katija, K., Orenstein, E., Schlining, B., Lundsten, L., Barnard, K., Sainz, G., Boulais, O., Cromwell, M., Butler, E., Woodward, B., et al.: Fathomnet: A global image database for enabling artificial intelligence in the ocean. Scientific reports 12(1), 15914 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "40. Kay, J., Kulits, P., Stathatos, S., Deng, S., Young, E., Beery, S., Van Horn, G., Perona, P.: The caltech fish counting dataset: A benchmark for multiple-object tracking and counting. In: European Conference on Computer Vision. pp. 290-311. Springer (2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 315, + 91, + 447, + 101 + ], + "type": "text", + "content": "Insect Identification in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "41. van Klink, R., August, T., Bas, Y., Bodesheim, P., Bonn, A., Fossoy, F., Høye, T.T., Jongejans, E., Menz, M.H., Miraldo, A., et al.: Emerging technologies revolutionise insect ecology and monitoring. Trends in ecology & evolution (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 204 + ], + "type": "text", + "content": "42. Koh, P.W., Sagawa, S., Marklund, H., Xie, S.M., Zhang, M., Balsubramani, A., Hu, W., Yasunaga, M., Phillips, R.L., Gao, I., Lee, T., David, E., Stavness, I., Guo, W., Earnshaw, B.A., Haque, I.S., Beery, S., Leskovec, J., Kundaje, A., Pierson, E., Levine, S., Finn, C., Liang, P.: WILDS: A benchmark of in-the-wild distribution shifts. In: International Conference on Machine Learning (ICML) (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 204, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 204, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 204, + 480, + 236 + ], + "type": "text", + "content": "43. Korsch, D., Bodesheim, P., Denzler, J.: Deep learning pipeline for automated visual moth monitoring: insect localization and species classification. arXiv preprint arXiv:2307.15427 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 236, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 278 + ], + "type": "text", + "content": "44. Krasin, I., Duerig, T., Alldrin, N., Ferrari, V., Abu-El-Haija, S., Kuznetsova, A., Rom, H., Uijlings, J., Popov, S., Veit, A., et al.: Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://github.com/openimages2(3), 18 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 278, + 480, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 333 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 333 + ], + "type": "text", + "content": "45. Kumar, N., Belhumeur, P.N., Biswas, A., Jacobs, D.W., Kress, W.J., Lopez, I.C., Soares, J.V.: Leafsnap: A computer vision system for automatic plant species identification. In: Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part II 12. pp. 502-516. Springer (2012)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 333, + 480, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 480, + 376 + ], + "type": "text", + "content": "46. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 376, + 480, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 480, + 408 + ], + "type": "text", + "content": "47. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 408, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 480, + 430 + ], + "type": "text", + "content": "48. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 430, + 480, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 480, + 461 + ], + "type": "text", + "content": "49. Luccioni, A.S., Rolnick, D.: Bugs in the data: How imagenet misrepresents biodiversity. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 14382-14390 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 461, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 493 + ], + "type": "text", + "content": "50. Ming, Y., Yin, H., Li, Y.: On the impact of spurious correlation for out-of-distribution detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 10051-10059 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 493, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 480, + 515 + ], + "type": "text", + "content": "51. Mora, C., Tittensor, D.P., Adl, S., Simpson, A.G., Worm, B.: How many species are there on earth and in the ocean? PLoS biology 9(8), e1001127 (2011)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 515, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 480, + 557 + ], + "type": "text", + "content": "52. Nawoya, S., Ssemakula, F., Akol, R., Geissmann, Q., Karstoft, H., Bjerge, K., Mwikirize, C., Katumba, A., Gebreyesus, G.: Computer vision and deep learning in insects for food and feed production: A review. Computers and Electronics in Agriculture 216, 108503 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "type": "text", + "content": "53. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: 2008 Sixth Indian conference on computer vision, graphics & image processing. pp. 722-729. IEEE (2008)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "type": "text", + "content": "54. Norouzzadeh, M.S., Morris, D., Beery, S., Joshi, N., Jojic, N., Clune, J.: A deep active learning system for species identification and counting in camera trap images. Methods in ecology and evolution 12(1), 150-161 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "text", + "content": "55. Oliver, R.Y., Iannarilli, F., Ahumada, J., Fegraus, E., Flores, N., Kays, R., Birch, T., Ranipeta, A., Rogan, M.S., Sica, Y.V., et al.: Camera trapping expands the view into global biodiversity and its change. Philosophical Transactions of the Royal Society B 378(1881), 20220232 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "A. Jain et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "56. Ouaknine, A., Kattenborn, T., Laliberte, E., Rolnick, D.: Openforest: A data catalogue for machine learning in forest monitoring. arXiv preprint arXiv:2311.00277 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "text", + "content": "57. Rolnick, D., Donti, P.L., Kaack, L.H., Kochanski, K., Lacoste, A., Sankaran, K., Ross, A.S., Milojevic-Dupont, N., Jaques, N., Waldman-Brown, A., et al.: Tackling climate change with machine learning. ACM Computing Surveys (CSUR) 55(2), 1-96 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "type": "text", + "content": "58. Roy, D., et al.: Towards a standardised framework for ai-assisted, image-based monitoring of nocturnal insects. Philosophical Transactions of the Royal Society B (2024). https://doi.org/10.1098/rstb.2023.0108, (Forthcoming)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 482, + 270 + ], + "type": "text", + "content": "59. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) 115(3), 211-252 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 482, + 314 + ], + "type": "text", + "content": "60. Seibold, S., Gossner, M.M., Simons, N.K., Bluthgen, N., Müller, J., Ambarh, D., Ammer, C., Bauhus, J., Fischer, M., Habel, J.C., et al.: Arthropod decline in grasslands and forests is associated with landscape-level drivers. Nature 574(7780), 671-674 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 315, + 482, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 482, + 347 + ], + "type": "text", + "content": "61. Stark, T., Stefan, V., Wurm, M., Spanier, R., Taubenbock, H., Knight, T.M.: Yolo object detection models can locate and classify broad groups of flower-visiting arthropods in images. Scientific Reports 13(1), 16364 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 348, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 482, + 369 + ], + "type": "text", + "content": "62. Stork, N.E., et al.: How many species of insects and other terrestrial arthropods are there on earth. Annual review of entomology 63(1), 31-45 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 370, + 482, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 482, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 482, + 391 + ], + "type": "text", + "content": "63. Suto, J.: Codling moth monitoring with camera-equipped automated traps: A review. Agriculture 12(10), 1721 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 482, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 482, + 424 + ], + "type": "text", + "content": "64. Swanson, A., Kosmala, M., Lintott, C., Simpson, R., Smith, A., Packer, C.: Snapshot serengeti, high-frequency annotated camera trap images of 40 mammalian species in an african savanna. Scientific data 2(1), 1-14 (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 425, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 482, + 456 + ], + "type": "text", + "content": "65. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818-2826 (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 457, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 482, + 479 + ], + "type": "text", + "content": "66. Teixeira, A.C., Ribeiro, J., Morais, R., Sousa, J.J., Cunha, A.: A systematic review on automatic insect detection using deep learning. Agriculture 13(3), 713 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 482, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 482, + 501 + ], + "type": "text", + "content": "67. Theivaprakasham, H.: Identification of indian butterflies using deep convolutional neural network. Journal of Asia-Pacific Entomology 24(1), 329-340 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 502, + 482, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 482, + 533 + ], + "type": "text", + "content": "68. Tuia, D., Kellenberger, B., Beery, S., Costelloe, B.R., Zuffi, S., Risse, B., Mathis, A., Mathis, M.W., van Langevelde, F., Burghardt, T., et al.: Perspectives in machine learning for wildlife conservation. Nature communications 13(1), 792 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 482, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 482, + 578 + ], + "type": "text", + "content": "69. Van Horn, G., Branson, S., Farrell, R., Haber, S., Barry, J., Ipeirotis, P., Perona, P., Belongie, S.: Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 595-604 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 482, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 482, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 482, + 621 + ], + "type": "text", + "content": "70. Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist species classification and detection dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8769-8778 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 622, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 482, + 643 + ], + "type": "text", + "content": "71. Van Horn, G., Perona, P.: The devil is in the tails: Fine-grained classification in the wild. arXiv preprint arXiv:1709.01450 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 644, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 482, + 666 + ], + "type": "text", + "content": "72. Wagner, D.L.: Insect declines in the anthropocene. Annual review of entomology 65, 457-480 (2020)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 315, + 91, + 447, + 101 + ], + "type": "text", + "content": "Insect Identification in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 237 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "73. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset. Tech. rep. (2011)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "text", + "content": "74. Wei, X.S., Song, Y.Z., Mac Aodha, O., Wu, J., Peng, Y., Tang, J., Yang, J., Belongie, S.: Fine-grained image analysis with deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence 44(12), 8927-8948 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "75. Weinstein, B.G.: A computer vision for animal ecology. Journal of Animal Ecology 87(3), 533-545 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "text", + "content": "76. Wightman, R.: Pytorch image models. https://github.com/rwrightman/pytorch-image-models (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "77. Zhou, K., Liu, Z., Qiao, Y., Xiang, T., Loy, C.C.: Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "A. Jain et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/InstaStyle_ Inversion Noise of a Stylized Image is Secretly a Style Adviser/938fb53e-3bec-4f8a-bc7e-f5b36dd7b576_content_list.json b/2024/InstaStyle_ Inversion Noise of a Stylized Image is Secretly a Style Adviser/938fb53e-3bec-4f8a-bc7e-f5b36dd7b576_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..56a1c486924514c5ce5f0767a1fdb13e506314dd --- /dev/null +++ b/2024/InstaStyle_ Inversion Noise of a Stylized Image is Secretly a Style Adviser/938fb53e-3bec-4f8a-bc7e-f5b36dd7b576_content_list.json @@ -0,0 +1,2926 @@ +[ + { + "type": "text", + "text": "INSTASTYLE: Inversion Noise of a Stylized Image is Secretly a Style Adviser", + "text_level": 1, + "bbox": [ + 223, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xing Cui $^{1}$ Zekun Li $^{2}$ Peipei Li $^{1\\star}$ Huaibo Huang $^{3}$ Xuannan Liu $^{1}$ Zhaofeng He $^{1}$", + "bbox": [ + 238, + 210, + 763, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Beijing University of Posts and Telecommunications", + "bbox": [ + 320, + 253, + 681, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 University of California, Santa Barbara", + "bbox": [ + 362, + 268, + 640, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ MAIS & NLPR, Institute of Automation, Chinese Academy of Sciences", + "bbox": [ + 256, + 282, + 745, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{cuixing, lipeipei, liuxuannan, zhaofenghe}@bupt.edu.cn", + "bbox": [ + 287, + 297, + 712, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zekunli@cs.ucsb.edu huaibo.huang@cripac.ia.ac.cn", + "bbox": [ + 305, + 311, + 696, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://cuixing100876.github.io/instastyle.github.io/", + "bbox": [ + 295, + 325, + 705, + 338 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1e63eab505223bc53f6724f12ef348ed063d982a0d5eed9da5777533207015f9.jpg", + "image_caption": [ + "(a) Stylized image generation" + ], + "image_footnote": [], + "bbox": [ + 243, + 364, + 313, + 476 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/169ca768357d38194f2bb25e7bf82d15cb71c2e183bd28dbc4ac52b090739872.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 315, + 364, + 383, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/69cb3e298b998b86a6927958978f4559c255977f7d7a973f04cf3751e19c297d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 392, + 364, + 449, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/84c629f341a3fd09e2ee3f6cd287e8f9f227a9693baefa5afcb66f43fc32c2af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 454, + 364, + 521, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b473543400ab34673d2adfbf995d8edb1c31fb93a1762a0942448bb5d951a3dc.jpg", + "image_caption": [ + "(b) Combination of two styles" + ], + "image_footnote": [], + "bbox": [ + 535, + 364, + 616, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1c4d6b7cda07a19095eb7c0f005b26060b39492853b3c4481b232b0aab1effce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 364, + 684, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/97529d597569d6ed5e4480cfb297ad0f2af082ea4440564d09051db94f66be3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 364, + 756, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9330de21c23d5472d0ff22763d378e135a93505a0d1b824f021dedc8d5609c02.jpg", + "image_caption": [ + "Fig.1: Visualization of INSTASTYLE. (a) Our method excels at capturing style details and distinguishing between similar styles. (b) The first and third columns show images styled with reference to style 1 and style 2, respectively. The middle column shows images in a combined style. (c) Our method supports adjusting the degree of two styles during combination, dynamically ranging from one style to another." + ], + "image_footnote": [], + "bbox": [ + 246, + 489, + 313, + 549 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/907bae5c8e0dc894846f9e6f17ff37fd2c55f8afe2aa130360b62c3694a522a6.jpg", + "image_caption": [ + "(c) Smooth style combination between two styles" + ], + "image_footnote": [], + "bbox": [ + 326, + 489, + 395, + 549 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/55ba7284fe8c72a89735b12aa9a149385830693aae246f036db71da7804d6512.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 489, + 465, + 549 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/267eaa01b35a34ab943f9ada63666182b2a808ac49138da21b5861ffa48e9bce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 489, + 535, + 549 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c07f653d646e4259d232a88b8f4c396b95f600fd4f1138d064b6859e612e65a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 489, + 604, + 549 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8f12bad59ba600af0c00a833e3d7ab06fdbc0c2280cf101b86b5e6ab0fce7cc1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 606, + 489, + 674, + 549 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4bcbdf770e23e41f3ada368e2fd9598647f1c8198b5ab2243a224d827bbd285c.jpg", + "image_caption": [ + "Style 2" + ], + "image_footnote": [], + "bbox": [ + 689, + 500, + 756, + 549 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Stylized text-to-image generation focuses on creating images from textual descriptions while adhering to a style specified by reference images. However, subtle style variations within different reference images can hinder the model from accurately learning the target style. In this paper, we propose INSTASTYLE, a novel approach that excels in generating high-fidelity stylized images with only a single reference image. Our approach is based on the finding that the inversion noise from a stylized reference image inherently carries the style signal, as evidenced by their non-zero signal-to-noise ratio. We employ DDIM inversion to extract this noise from the reference image and leverage a", + "bbox": [ + 259, + 676, + 740, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 218, + 825, + 383, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "diffusion model to generate new stylized images from the \"style\" noise. Additionally, the inherent ambiguity and bias of textual prompts impede the precise conveying of style during image inversion. To address this, we devise prompt refinement, which learns a style token assisted by human feedback. Qualitative and quantitative experimental results demonstrate that INSTASTYLE achieves superior performance compared to current benchmarks. Furthermore, our approach also showcases its capability in the creative task of style combination with mixed inversion noise.", + "bbox": [ + 259, + 147, + 743, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: Stylized image generation $\\cdot$ Inversion noise $\\cdot$ Signal-to-noise ratio $\\cdot$ Prompt refinement", + "bbox": [ + 259, + 272, + 743, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 332, + 375, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, the advent of DDPMs [16, 51] has ushered in a new era of image generation. Existing text-to-image generation methods [5, 43, 45, 46] can generate images in coarse-grained styles by incorporating style descriptions in the textual prompt. However, the textual prompt is ambiguous which makes it hard to express style precisely [11]. Recently, personalized generation [10,48,63] has been introduced to generate novel concepts, which can be utilized for stylized generation by viewing the reference style as a new concept. These approaches typically necessitate mul", + "bbox": [ + 212, + 369, + 488, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tiple images as references. However, the subtle style variations in these multiple images present a challenge for the model to accurately learn and replicate the intended style. As shown in the first column of Fig. 1 (a), although the reference image Houses at Auvers (top) and Thatched Cottages by a Hill (bottom) are all Van Gogh's artworks, the differences in terms of color and texture [2] pose a significant challenge. Some recent approaches [3,53,69] are designed to more accurately describe the target style during generation. However, they may ignore the fine-grained style information in the reference images.", + "bbox": [ + 212, + 612, + 787, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose INSTASTYLE (Inversion Noise of a Stylized Image is Secretly a Style Adviser) based on the observation that the inversion noise from a stylized reference image inherently contains the style signal. Specifically, our approach only requires a single reference image, which is transformed into a noise via DDIM inversion. The inversion noise, which preserves the style signal, is then used to generate stylized images by utilizing diffusion models. To shed light on this phenomenon, we demonstrate that the inversion noise maintains", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c6e36f47eb5ca9567d1b717d049362c833b3659b33beacdc9ab7cb5b91e60c4c.jpg", + "image_caption": [ + "Fig. 2: Motivation. Sampling from the inversion noise of a reference image can generate stylized images. However, the optimal style token varies for each case among human-written style tokens. Our learnable style token, i.e., \"\", shows greater universality across various scenarios." + ], + "image_footnote": [], + "bbox": [ + 501, + 375, + 787, + 481 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "X. Cui et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a non-zero signal-to-noise ratio in Sec. 3.2, indicating that it retains essential information (including the style signal) from the reference image.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Nevertheless, directly using human-written prompts to describe styles for inverting the reference image often encounters challenges due to the inherent ambiguity of natural language. For instance, specific style descriptors like \"ink painting\" don't always work effectively with various objects. As shown in Fig. 2, the style \"ink painting\" might suit an object like a \"chair\" but not as well with a \"boy\". This inconsistency can arise because \"ink painting\" is typically associated with landscapes and might not yield optimal results for human subjects. Conversely, when using vague descriptors like \"specific style\", they may fail to provide enough information, leading to unpredictable generation quality (see the second row in Fig. 2). To address this, we propose Prompt Refinement to learn a style token with the assistance of human feedback. As illustrated in the last row of Fig. 2, the style information is boosted by using the learned style token in the noising process to obtain a better \"style\" noise. Therefore, our prompt refinement stands apart from previous methods [3, 53, 69], where style tokens are utilized in the denoising process to generate target images. During prompt refinement, we first collect images via human feedback. Then we optimize the embedding of the style token and the key and value projection matrices in the cross-attention layers of the diffusion model.", + "bbox": [ + 212, + 176, + 787, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 1 (a), our approach can effectively retain the fine-grained style in the reference image as well as generate new objects in high fidelity. Furthermore, our method supports the creative generation of style combination (Fig. 1 (b) and (c)) and allows adjusting the degree of two styles dynamically (Fig. 1 (c)). To sum up, our contributions are as follows:", + "bbox": [ + 212, + 449, + 787, + 525 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We find that the inversion noise of a reference image via DDIM inversion potentially retains style information and evidence this observation by the non-zero signal-to-noise ratio of the inversion noise.", + "- We propose INSTASTYLE based on our observation, which leverages the style signal in inversion noise. Additionally, to better represent the style during inversion, a prompt refinement scheme is designed to learn the style token. Human feedback is also introduced to enhance model alignment with human preferences during prompt refinement.", + "- Qualitative and quantitative experimental results show the superiority of our INSTASTYLE in generating high-quality stylized images. Moreover, it exhibits promising potential in creative scenarios like style combinations." + ], + "bbox": [ + 225, + 532, + 784, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 722, + 387, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Text-to-image Synthesis", + "text_level": 1, + "bbox": [ + 215, + 753, + 460, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image synthesis is an essential subject in computer vision [7,23,41,56]. Previous works are based on variational auto-encoder [29, 33] or generative adversarial networks [5,14,32,35,57]. With the advance of pretrained vison-language models [44,61] and diffusion models [34,54,73], text-to-image generation [4,8,43,45,", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "INSTASTYLE", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[60] has been widely studied and shown remarkable generalization ability. Recently, Lin et al. [36] have pointed out that the noise of the last step during noisering in the training process has a non-zero signal-to-noise ratio (SNR), i.e., there exists a signal leak. This results in a misalignment between the training and inference process. Therefore, some works [36, 50] propose to train diffusion models by enforcing the SNR to zero to avoid the signal leak. On the contrary, our approach leverages the leaked signal, which potentially includes the style details, from the reference image for the stylized generation.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As these methods ignore the concepts that do not appear in the training set [10], some works [10,48,49] study personalized text-to-image generation which aims to adapt text-to-image models to new concepts given several reference images. For example, Textual Inversion [10] introduces and optimizes a word vector for each new concept. Subsequent works [1,63,74] further enhance the flexibility and adaptiveness of the learning strategy. Some methods [20, 48, 49] propose to finetune the original diffusion model, showing more satisfactory results. However, maintaining fine-grained style while simultaneously generating new objects remains a challenge for current personalized generation methods [52].", + "bbox": [ + 212, + 268, + 787, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Stylized Image Generation", + "text_level": 1, + "bbox": [ + 214, + 430, + 480, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stylized image generation is a new paradigm for image generation which aims to generate content in a specific style given a few reference images. Although it is similar to the neural style transfer task [12, 13, 22, 27, 37, 59] which generates stylized images as well, they are fundamentally different. Style transfer solves an image translation task, focusing on translating a content image to target style [15, 21, 24, 28, 62, 64, 67, 68, 70]. For example, some works explore the global and local information to preserve the content of the source image [6, 18, 25, 26, 55, 65, 71, 75]. On the contrary, stylized image generation is geared towards generating a new image in specific style conditioned on a text. For example, ZDAIS [53] views stylized image generation as a domain adaptive task, where each style belongs to a domain. It learns disentangled prompts to adapt the model to new domains. Some approaches [9, 52] fine-tune the model to capture style properties. Our work differs from theirs by proposing a framework based on our observation that inversion noise contains style information.", + "bbox": [ + 212, + 459, + 787, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 698, + 330, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 3, our proposed method involves two main stages. In the first stage, we employ DDIM inversion to transform the reference image into noise. Notably, the inversion noise exhibits a non-zero signal-to-noise ratio, suggesting the presence of style signals from the reference image. Subsequently, we generate $M$ stylized images from the inversion noise conditioned on the given textual prompts. Due to the inherent ambiguity of the textual prompts that describe the style, it is challenging to precisely convey the desired style. Addressing this,", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "X. Cui et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eb3fe407cfbedfbac602427d450029f373f0a543a3e1a5fda5646613d6d87ca0.jpg", + "image_caption": [ + "Fig. 3: The training process of INSTASTYLE. (a) The first stage is initial stylized image generation. The reference image is inverted to noise conditioned on a prompt via DDIM Inversion. Then the inversion noise is utilized to generate initial stylized images. (b) The second stage is prompt refinement which leverages the selected high-quality initial stylized images to learn a style token." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 787, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the second stage involves the incorporation of human feedback to select $N$ high-quality generated images from the first stage. The selected images are then used to learn a style token via prompt refinement.", + "bbox": [ + 212, + 388, + 787, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 457, + 372, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We begin by reviewing the fundamental diffusion model. Subsequently, we introduce Stable Diffusion [46], a key component of our framework. Finally, we discuss both DDIM and DDIM inversion. In our method, DDIM serves the purpose of denoising a noise to generate new images. Concurrently, DDIM inversion is utilized to transform the reference image into a corresponding \"style\" noise.", + "bbox": [ + 212, + 482, + 787, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffusion models. Diffusion models [16, 51] contain a forward process and a backward process. The forward process adds noise to the data according to a predetermined, non-learned variance schedule $\\beta_{1},\\ldots,\\beta_{T}$ :", + "bbox": [ + 212, + 556, + 787, + 603 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(z _ {t} \\mid z _ {t - 1}\\right) := \\mathcal {N} \\left(z _ {t}; \\sqrt {1 - \\beta_ {t}} z _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 613, + 785, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An important property of the forward process is that we can obtain the latent variable $z_{t}$ directly based on $z_{0}$ :", + "bbox": [ + 212, + 641, + 785, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} = \\sqrt {\\alpha_ {t}} z _ {0} + \\sqrt {1 - \\alpha_ {t}} \\varepsilon , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 681, + 785, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varepsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ , $\\alpha_{t}\\coloneqq \\prod_{i = 1}^{t}(1 - \\beta_{i})$ . Diffusion models restore the information by learning the backward process:", + "bbox": [ + 212, + 709, + 785, + 741 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(z _ {t - 1} \\mid z _ {t}\\right) := \\mathcal {N} \\left(z _ {t - 1}; \\mu_ {\\theta} \\left(z _ {t}, t\\right), \\Sigma_ {\\theta} \\left(z _ {t}, t\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 752, + 785, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stable Diffusion. In Stable Diffusion, the diffusion process happens in latent space [46]. It utilizes a pretrained autoencoder which consists of an encoder and a decoder. The encoder $\\mathcal{E}(\\cdot)$ maps an image $x$ into a latent code $z$ and the decoder $\\mathcal{D}(\\cdot)$ transforms the latent code back into an image. The denoising model $\\varepsilon_{\\theta}(\\cdot)$", + "bbox": [ + 212, + 779, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "INSTASTYLE", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is an Unet [47]. During training, artificial noise is added to the sampled data $z_0$ according to timestamp $t$ based on Eq. (2), resulting in a noised sample $z_t$ . Denoising model starts with $z_t$ and is trained to predict the artificial noise:", + "bbox": [ + 215, + 146, + 782, + 191 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} E _ {z _ {0}, \\varepsilon \\sim N (0, I), t \\sim \\operatorname {U n i f o r m} (1, T)} \\| \\varepsilon - \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C}\\right) \\| _ {2} ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 202, + 785, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{C} = \\psi (\\mathcal{P})$ is the embedding of text condition $\\mathcal{P}$", + "bbox": [ + 215, + 234, + 607, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DDIM. In inference time, given a noise vector $z_{T}$ , the noise is gradually removed by sequentially predicting the added noise for $T$ steps. DDIM [54] is one of the denoising approaches with a deterministic process:", + "bbox": [ + 215, + 250, + 782, + 295 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t - 1} = \\sqrt {\\frac {\\alpha_ {t - 1}}{\\alpha_ {t}}} z _ {t} + \\left(\\sqrt {\\frac {1}{\\alpha_ {t - 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\cdot \\tilde {\\varepsilon} _ {\\theta}, \\qquad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 306, + 785, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tilde{\\varepsilon}_{\\theta}$ is the estimated noise.", + "bbox": [ + 215, + 344, + 446, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DDIM inversion [42, 54] transforms an image to a noise conditioned on a prompt. The diffusion process is performed in the reverse direction, i.e., $z_0 \\to z_T$ :", + "bbox": [ + 215, + 359, + 782, + 390 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t + 1} = \\sqrt {\\frac {\\alpha_ {t + 1}}{\\alpha_ {t}}} z _ {t} + \\left(\\sqrt {\\frac {1}{\\alpha_ {t + 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\cdot \\tilde {\\varepsilon} _ {\\theta}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 401, + 785, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Initial Stylized Image Generation", + "text_level": 1, + "bbox": [ + 215, + 449, + 537, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 3 (a), the first stage of our INSTASTYLE involves obtaining the inversion noise via DDIM inversion and sampling images via DDIM.", + "bbox": [ + 215, + 473, + 782, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In DDIM inversion, the added noise for each step is calculated conditioned on a prompt and gradually incorporated into the reference image following Eq. (6). Specifically, we set the learnable style token as a human-written style description in this stage. We demonstrate that the inversion noise from a stylized reference image inherently carries the style signal from the perspective of the signal-to-noise ratio. As the estimated noise $\\varepsilon_{\\theta}(z_t,t,\\mathcal{C})$ is trained to approximate the artificial noise $\\varepsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ , we assume that $\\varepsilon_{\\theta}(z_t,t,\\mathcal{C}) \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . According to Eq. (6), $z_{t + 1}$ can be approximated in a closed form:", + "bbox": [ + 215, + 503, + 785, + 625 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t + 1} = \\sqrt {\\frac {\\alpha_ {t + 1}}{\\alpha_ {0}}} z _ {0} + \\sqrt {\\sum_ {i = 0} ^ {t} \\frac {\\alpha_ {t + 1}}{\\alpha_ {i + 1}} \\left(\\sqrt {\\frac {1}{\\alpha_ {i + 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {i}} - 1}\\right) ^ {2}} \\cdot \\bar {\\varepsilon} _ {0}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 281, + 637, + 785, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\bar{\\varepsilon}_0\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$", + "bbox": [ + 215, + 678, + 359, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Signal-to-Noise ratio. Signal-to-noise ratio (SNR) is introduced to measure the ratio of signals from the original image preserved in the noise [36,46]. The SNR of the inversion noise can be calculated as:", + "bbox": [ + 215, + 694, + 782, + 738 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S N R} (t) := \\frac {1}{\\sum_ {i = 0} ^ {t} \\frac {\\alpha_ {0}}{\\alpha_ {i + 1}} \\left(\\sqrt {\\frac {1}{\\alpha_ {i + 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {i}} - 1}\\right) ^ {2}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 750, + 785, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present detailed derivations in Sec. B in Supplementary Material. In Stable Diffusion [46], the predetermined variance schedule $\\beta_{t} = (\\sqrt{0.00085}\\cdot (1 - j) + \\sqrt{0.012}\\cdot j)^{2}$ , where $j = \\frac{t - 1}{T - 1}$ . At the terminal timestep $T = 1000$ , the", + "bbox": [ + 215, + 794, + 782, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "X. Cui et al.", + "bbox": [ + 271, + 114, + 356, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$SNR(T) = 0.015144$ , i.e., $z_{T}$ has a none-zero signal-to-noise ratio. This nonzero signal-to-noise ratio suggests that the noise obtained via DDIM inversion still retains information from the reference image and deviates from white noise. Besides, we conduct qualitative experiments and observe that style information can consistently be preserved at the terminal timestep (refer to Sec. C.3 in Supplementary Material). Therefore, we can generate target stylized images by sampling from the inversion noise based on Eq. (5).", + "bbox": [ + 212, + 146, + 782, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Prompt Refinement", + "text_level": 1, + "bbox": [ + 214, + 273, + 426, + 289 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As the natural language prompts may not precisely convey the style (Fig. 2), we propose Prompt Refinement to learn the style tokens as shown in Fig. 3 (b). We utilize the generated stylized images in the first stage to constitute the training data for prompt refinement learning. While these generated images may not always be precise, we observe that most of them successfully retain the style information alongside appropriate content. We manually select images that distinctly embody the reference style and the target object to build the dataset.", + "bbox": [ + 212, + 299, + 784, + 404 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, we introduce a new token, i.e., “” to represent the style descriptor and learn its embedding. In practice, we initialize the new token with the embeddings of the textual style descriptor of the reference image. The text condition is input to the model via two projection matrices in the cross-attention block of diffusion model $\\varepsilon_{\\theta}$ , i.e., $W^{k} \\in \\mathbb{R}^{d \\times d'}$ and $W^{v} \\in \\mathbb{R}^{d \\times d'}$ . Therefore, we also fine-tune these two projection matrices. The text feature $\\mathbf{c} \\in \\mathbb{R}^{s \\times d}$ is projected to key $K = \\mathbf{c}W^{k}$ and value $V = \\mathbf{c}W^{v}$ . The query matrix $W^{q} \\in \\mathbb{R}^{l \\times d'}$ projects the latent image feature $\\mathbf{f} \\in \\mathbb{R}^{(h \\times w) \\times l}$ into query feature $Q = \\mathbf{f}W^{q}$ . Then the cross-attention [58] is calculated as:", + "bbox": [ + 212, + 405, + 785, + 541 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d ^ {\\prime}}}\\right) V. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 553, + 784, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We utilize the LoRA [20] for model fine-tuning. Specifically, for a projection matrix $W \\in \\mathbb{R}^{d \\times d'}$ (i.e., $W = W^k$ or $W = W^v$ ) to be fine-tuned, we update a low-rank residual rather than directly fine-tuning $W$ . Formally, we denote the fine-tuned projection matrix as $W' = W + BA$ , where $B \\in \\mathbb{R}^{d \\times r}$ , $A \\in \\mathbb{R}^{r \\times d'}$ , and the rank $r \\ll \\min(d, d')$ . During training, only $A$ and $B$ are learnable. For $y = Wx$ , the modified forward pass is $y = Wx + BAx$ . Putting the two stages together, our full algorithm is shown in Algorithm 1.", + "bbox": [ + 214, + 595, + 782, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Inference", + "text_level": 1, + "bbox": [ + 214, + 724, + 339, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Stylized image generation. For stylized image generation, we first learn the style token \"\" to describe the style in the reference image. During inference, the learned style token is integrated into the prompt that inverts image and the prompt that generates the target image. Besides, the diffusion model $\\varepsilon_{\\theta}$ is utilized as the backbone of DDIM inversion and DDIM. Specifically, the reference image is first inverted to a noise $z_{T}$ via DDIM inversion conditioned", + "bbox": [ + 212, + 750, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "INSTASTYLE", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Input: A source prompt embedding $\\mathcal{C} = \\psi(\\mathcal{P})$ and a reference image $\\mathcal{I}$ . Initial style token embedding $v$ is included in the prompt embedding $\\mathcal{C}$ .", + "2 Output: Optimized style token embedding $v^{*}$ and diffusion model $\\varepsilon_{\\theta}^{*}$ ." + ], + "bbox": [ + 225, + 167, + 759, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "// Initial Stylized Image Generation", + "bbox": [ + 240, + 218, + 519, + 231 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "3 Compute the inversion noise $z_{T}$ using DDIM inversion over $\\mathcal{I}$ conditioned on $\\mathcal{C}$ based on Eq. (6);", + "4 Generate $M$ images conditioned on a predefined prompt set based on Eq. (5); // Prompt Refinement", + "5 Select $N$ images from $M$ images with human feedback as the training dataset $q(\\mathbf{x}_0, c)$ ;" + ], + "bbox": [ + 225, + 232, + 763, + 315 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6 repeat", + "bbox": [ + 225, + 316, + 292, + 328 + ], + "page_idx": 7 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1: INSTASTYLE" + ], + "code_body": "7 $\\mathbf{x}_0,c\\sim q(\\mathbf{x}_0,c)$ \n8 $z_{0} = \\mathcal{E}(\\mathbf{x}_{0})$ \n9 $t\\sim \\mathrm{Uniform}(\\{1,\\dots ,T\\})$ \n10 $\\varepsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ \n11 $z_{t} = \\sqrt{\\alpha_{t}} z_{0} + \\sqrt{1 - \\alpha_{t}}\\varepsilon$ \n12 Take gradient step on $\\nabla_{\\theta ,v}\\| \\varepsilon -\\varepsilon_{\\theta}(z_t,t,c)\\| _2^2;$", + "guess_lang": "txt", + "bbox": [ + 225, + 329, + 565, + 416 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "13 until converged;", + "14 Return $v^{*}$ , $\\varepsilon_{\\theta}^{*}$" + ], + "bbox": [ + 218, + 417, + 352, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "on a prompt containing the learned style token. Then we sample from the inversion noise conditioned on a target prompt which contains the target content and the learned style token via DDIM. Specifically, the estimated noise is calculated based on the classifier-free guidance [17]:", + "bbox": [ + 212, + 479, + 787, + 541 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\varepsilon} _ {\\theta} \\left(z _ {t}, t, \\mathcal {C}, \\varnothing\\right) = \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\varnothing\\right) + w \\cdot \\left(\\varepsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C}\\right) - \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\varnothing\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 261, + 549, + 785, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\varnothing = \\psi(\")$ is the embedding of a null text. $\\varepsilon_{\\theta}(z_t, t, \\mathcal{C})$ represents the conditional predictions. $w$ is the guidance scale parameter.", + "bbox": [ + 212, + 577, + 785, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Combination of two styles. To combine two styles, we learn style tokens for each style, i.e., “” and “”. Specifically, we use the selected images of both styles to constitute the training set. Then we jointly optimize the embeddings of the style tokens and fine-tune the key projection and value projection in the cross-attention block. During inference, we individually transform two reference images into their corresponding inversion noise $z_{t1} \\in \\mathbb{R}^{H \\times W \\times C}$ and $z_{t2} \\in \\mathbb{R}^{H \\times W \\times C}$ via DDIM inversion conditioned on a prompt containing their learned style token. As the style can be described by both the “style” noise and style token, we combine the styles from these two perspectives. For the “style” noise, the combined inversion noise $z_{t}$ is obtained based on a masking strategy:", + "bbox": [ + 212, + 607, + 787, + 758 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} = (\\mathbf {1} - \\mathbf {M}) \\odot z _ {t 1} + \\mathbf {M} \\odot z _ {t 2}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 768, + 785, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\odot$ is element-wise multiplication and $\\mathbf{1}$ is a binary mask filled with ones. $\\mathbf{M} \\in \\{0, 1\\}^{H \\times W}$ denotes a random binary mask indicating where to drop out and fill in from two inversion noises. The noise mix ratio between two inversion", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "X. Cui et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "noises is $\\alpha$ , representing the percent of $\\mathbf{M}$ that is set to 1 (the rest is set to 0). Finally, we utilize a composed guidance mechanism [38] to estimate the noise:", + "bbox": [ + 214, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {\\varepsilon} _ {\\theta} \\left(z _ {t}, t, \\mathcal {C}, \\varnothing\\right) = \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\varnothing\\right) \\\\ + w \\cdot (1 - \\beta) \\cdot \\left(\\varepsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C} _ {1}\\right) - \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\varnothing\\right)\\right) \\tag {12} \\\\ + w \\cdot \\beta \\cdot \\left(\\varepsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C} _ {2}\\right) - \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\varnothing\\right)\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 303, + 185, + 785, + 243 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\beta$ is a prompt mix ratio. $\\mathcal{C}_1$ and $\\mathcal{C}_2$ are the embeddings of the target prompts for the two styles, respectively. Denote the target object as $<\\mathrm{obj}>$ , they can be formulated as $\\mathcal{C}_1 = \\psi(\"A < \\mathrm{obj}>$ in the style of $<\\mathrm{style}1>$ ) and $\\mathcal{C}_2 = \\psi(\"A < \\mathrm{obj}>$ in the style of $<\\mathrm{style}2>$ ).", + "bbox": [ + 214, + 251, + 782, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 332, + 364, + 349 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Setting", + "text_level": 1, + "bbox": [ + 215, + 362, + 439, + 378 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We collect 60 images as the reference style dataset. The complete information on these images is listed in Tab. 2 in the Supplementary Material. Our method is implemented using PyTorch and executed on a single NVIDIA GeForce RTX 3090 GPU. The training procedure consists of 500 iterations, employing the Adam optimizer with a learning rate of $1 \\times 10^{-5}$ . During sampling, we set the guidance scale $w = 2.5$ . The number of generated images $M$ in the first stage is 15 whose details are shown in Sec. C.1 in the Supplementary Material. The number of selected images $N$ in the second stage is set to 5.", + "bbox": [ + 214, + 383, + 787, + 506 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c80cd58cfac2a6c96bc4d5b69f6c86d483da87d9c4e1d3427d3057bcb40f2567.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 530, + 285, + 580 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/62bb4229f8af2972e988e022ba0d231ef3ad458bc58eb5fedee8e8358b7441be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 582, + 282, + 630 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/bd86b46287db8e57f5d00d11245aa8a8a3df65f210ac82d2bf536f74c6df529a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 632, + 282, + 679 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f76010f551154aba616b502ebb7632c7ed8a8dc21e847c25fe407ec728079064.jpg", + "image_caption": [ + "Reference" + ], + "image_footnote": [], + "bbox": [ + 217, + 680, + 282, + 729 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/7da6a41e381c596b65db7c7f3decaf9953cc2a7dd0161b4941397773a727d28a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 531, + 393, + 580 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ad889a0748d7d4b91bc4e81feab077c629a9fff69edfcd9d34d3c23ae55d6273.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 582, + 393, + 630 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ec16abb08643cb1c6a287e592936c53702e6390106467c49c48455a9c122504f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 632, + 393, + 680 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a4f917a9ca3a35ce6227ed953403d9a6a49b24239d416c7cadfd7b3ae105b922.jpg", + "image_caption": [ + "(a)InstaStyle(Ours)" + ], + "image_footnote": [], + "bbox": [ + 300, + 680, + 395, + 729 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/820c13bc6a331fdd56672ae44865209d89456ea7e66cc8269458d290bf5066b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 531, + 491, + 580 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f4ea39848124db46872acc74bb26bef845260cfadd46797d37f9938aa2eb0ba4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 582, + 491, + 630 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f30cfcc739e5c8d127bd62253d7bf6fd9854cea2b82a4bef23f836be27277f2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 631, + 491, + 680 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/55ceaa9fe7cd17f7c80fc91b8bafb11df3a9528725f082866ec54ca2184e4fcc.jpg", + "image_caption": [ + "(b) StyleDrop" + ], + "image_footnote": [], + "bbox": [ + 398, + 680, + 491, + 729 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a44d63fc11adf0886ed91375a8e74f7fae3953358ce775aa66cefe1ac9d7695e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 531, + 586, + 580 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4674897d31eeb69fbacd88345437a172ba213fbd74048934bf294f05392867fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 582, + 586, + 630 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0107c43360cc799d70b54fe3f4e7e4e8b60c9f736ad4ba37ad6fb88f2ea1bc34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 631, + 586, + 680 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/cedd09f2c56d58e9316e4c472b9610fbbd96ec4b0292eeb9a36619ac9f8bb852.jpg", + "image_caption": [ + "(c) DreamBooth", + "Fig. 4: Qualitative comparison of stylized image generation on various styles. Objects for synthesis are Bicycle, Sunflowers and Chair. Our method excels at capturing fine-grained style information, such as color, textures, and brushstrokes." + ], + "image_footnote": [], + "bbox": [ + 495, + 680, + 586, + 729 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/002147886c24b0ab81221df2673671661e13f714c9273a98d253f706318af2ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 531, + 684, + 580 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/653565d4528b7c0528cf9491e47b6acdccc83b98af0dc302b7a61fafab4ca3fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 582, + 684, + 630 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6983507207ecd7191875b40ee1de1704c34080d70e3b1344588e9ce3ba303710.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 631, + 684, + 680 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/83b87eac91516065c091178a75e7d91cba991330dfc8b756dcd910b4ccddb590.jpg", + "image_caption": [ + "(d) Custom Diffusion" + ], + "image_footnote": [], + "bbox": [ + 591, + 680, + 684, + 729 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/40a69f8153fe9d2dbf1fbe2c7145211952e9bf6c2a33c76174eb41e0829690d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 531, + 782, + 580 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/88421700aef09d80582a937676a89610ad3f6f07859c37df56059aa7ebbfea5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 582, + 782, + 630 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/79e448d2d647579d093015077ce84c9df2da00dec33a313e6c9974613d4136ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 631, + 782, + 680 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d5b43c83b2a50bb2f1d961f2fb79cdcb39838166256d1b4b070d8709d0b2ca67.jpg", + "image_caption": [ + "(e) Textual Inversion" + ], + "image_footnote": [], + "bbox": [ + 689, + 680, + 782, + 729 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "INSTASTYLE", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/93a947d27a8f62ea37793f74e43126f818fb863980e76daf4c312e7d6151b1a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 287, + 196 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e723e52182e240f95fbbe1b7a875a9b19283d0e8e74851d803ff4d831ec3d572.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 199, + 285, + 248 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/49116d4c68b22b28a783b598b4d337418063adbb1b2519c0a2bf9d78da5f3f88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 250, + 285, + 301 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/6c1a1cdaf0f6f75b8d8923bfa18fd5b89f2d51c76d7d7dc52e08972b08ca6aa6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 303, + 285, + 352 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a8d3e58b66bc7a859d63d2473cbdf9c7094fcc83fa576d996983b959fa348831.jpg", + "image_caption": [ + "Reference" + ], + "image_footnote": [], + "bbox": [ + 217, + 353, + 285, + 406 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/6308999419633e125169ded48e8b7eff36ed76541bc56a75a94058289b96d5f0.jpg", + "image_caption": [ + "Elephant" + ], + "image_footnote": [], + "bbox": [ + 303, + 145, + 370, + 185 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/edc3f3da5f4238def5eac7a6ed55fdab14f7825f302830fd0521a3e20a4fd899.jpg", + "image_caption": [ + "Castle" + ], + "image_footnote": [], + "bbox": [ + 303, + 199, + 370, + 239 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/dd8c65a113b26c9d1a7ab6c21cd6573ca34f853337ef4a1d7a5000375691bd2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 250, + 370, + 297 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/de7afc33523f9555c9b8ce501cbfdd3109d3fbc90863b3439ab6346959c63a42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 302, + 301, + 370, + 352 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d7c07f514c16acc211d896020f09eae1047f00ab72e06c84dd7ccecb71e2a852.jpg", + "image_caption": [ + "InstaStyle (Ours) AsePA-Net" + ], + "image_footnote": [], + "bbox": [ + 303, + 353, + 370, + 405 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/1f685ed8d931e04fcb720f28c5e00c1b0b51feb446af776d043f685b254cc413.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 145, + 441, + 196 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/330ec054cd93cdd9ed17ca7eb53fd48bddd2eef517b06bc473b1e77534e8c038.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 198, + 439, + 250 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/ef77012026b7637a45a696c5a90362734f2fbf4bd85a431bdbf604570b051001.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 250, + 439, + 300 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/206cfea526225d6930e12f88a1c78d4df1041164ce2bd67524b1ba578a9e2a94.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 301, + 439, + 352 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d10c581de5238462787cc8f7fff64353771ac95d9828003356f85a3eb8091a98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 353, + 439, + 405 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "AsePA-Net", + "bbox": [ + 379, + 407, + 431, + 416 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/73463e740a620daeeb15fcfe9c403a856ba19fd6a67bb97c6da8d3d443e15c3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 145, + 508, + 196 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/95bfccb1ea87ba82468c96b26daec57cc902ad2b653de2bcfeef06fd6fa3ecd4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 198, + 506, + 250 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/35d60c178be4da672a7f060d94fa71e8479354998dd926f364096d7bbe1c0838.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 250, + 506, + 300 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/cbfcc14075106e7f94a26eb32d48e078074cd840d83d012f91adf7a93a083600.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 301, + 506, + 352 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d7180e1caddb0d3a86177ba9e3e5d782cb58390549923bd61eaf4ac902533446.jpg", + "image_caption": [ + "QuantArt" + ], + "image_footnote": [], + "bbox": [ + 442, + 353, + 506, + 405 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/05598e7175a982d5487822da50fa43e462e996653c5f4523c0726f50f9d32d62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 145, + 576, + 196 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/0c9d46d1c3d0bcf4267c7bd4f59bbb1cc89b6f4ff181b89666d1784bd8fc38dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 198, + 576, + 250 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e264bcef2f9453325657db6676fb1edfdc9e33fc03bf384f5a8ea86c94ee7706.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 250, + 576, + 300 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/1476c31bfcbc733e327d103cc8148e8bbc300b1968115dddab1c68d7bf80e6f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 301, + 576, + 352 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/4b93159d2c873695fdb64eed844b3e72c22c4070dca0151e1897340a52c97462.jpg", + "image_caption": [ + "StyTr2", + "Fig. 5: Qualitative comparison with style transfer methods. For style transfer methods, content images are employed (shown in the bottom right). Despite our method relying solely on text for content, we achieve comparable performance in the fidelity of the content. Furthermore, we excel in preserving style details." + ], + "image_footnote": [], + "bbox": [ + 509, + 353, + 576, + 405 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e41ca3e62ebe14a9808eafe31303259a8263db9d06de58c81dded567f3601f4f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 145, + 645, + 196 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/34009c2c215c55995462b4dc78ce439822f400a8f47f02ecc3754b8173f441ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 198, + 645, + 250 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/1fd7c27b9d3cedce90590f2b1b794c1f6ebdb3bb91f09bbe58e630c76a928466.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 250, + 645, + 300 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/8fdcbdce0135afb31df927f694a23fd3910a085c92de7df6dc37827d04698ab6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 301, + 645, + 352 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3487069969ff08a784ccbb60e4c46f118026d63ab2069534411f7daec87421b4.jpg", + "image_caption": [ + "CCPL" + ], + "image_footnote": [], + "bbox": [ + 578, + 353, + 645, + 405 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/987a6b0916c64ccabaa92e53be4572161cd6d477426dc4baa3004fb4e22a8d70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 145, + 715, + 196 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/045f44cf9b34091bb737841c03f6264b59476797cef469c95c02a9fd0615060d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 198, + 715, + 250 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/02fba50679b86a7063d59cd655d216a36378a742a971c229f1341e5ac30046a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 250, + 715, + 300 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/48e2922a4116385d9a72fd20ddb187d0dce8e687c4feb90b5859d93d60c78ee6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 301, + 715, + 352 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/bfb1dcd8741518b7e4e125fd298e9295f0816ec4af33087a11432480483cff82.jpg", + "image_caption": [ + "CAST" + ], + "image_footnote": [], + "bbox": [ + 648, + 353, + 715, + 405 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/b45d1cc9e7ae0b014b3310e82ec8e6e580f41351184793f9912aa37b1e52385c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 145, + 784, + 196 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/b0c9100974743ffa73e09a8c2b364963320cdada2537265b94bb012071077003.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 199, + 784, + 250 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c88c1c8132436e21e819713e6a68af7e3b7ca7f87a3da6c7d55f384e05abaa90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 250, + 784, + 300 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/026dc34338163e68aa922611d253adf77f91015dfa3e6afaaf49b935b3859adb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 301, + 784, + 352 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/cc3dfe80fe84d7125c55048fa26b61828e98d962b7edeca00acce1a79438501d.jpg", + "image_caption": [ + "AdaAttn" + ], + "image_footnote": [], + "bbox": [ + 718, + 353, + 784, + 405 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Stylized Image Synthesis", + "text_level": 1, + "bbox": [ + 215, + 527, + 465, + 544 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We conduct comparisons of INSTASTYLE against four recent methods following the open-sourced code $^{4,5}$ , i.e., StyleDrop [52], DreamBooth [48], Custom Diffusion [31], and Textual Inversion [10]. Besides, following [53], we also make a comparison with style transfer methods [6,19,21,39,66,72] to further illustrate the superiority of our approach. Notably, the style transfer task takes images as content input, in contrast to our task which only utilizes text prompts for content. To make a comparison between the two tasks, we additionally generate content images based on text prompts for style transfer methods. Although we can compare the two tasks under this setting, the comparison is unfair to our approach because the content images utilized by style transfer methods may provide additional content information compared to our text prompts.", + "bbox": [ + 212, + 556, + 787, + 723 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Qualitative results. As shown in Fig. 4, textual Inversion and Custom Diffusion often yield an unsatisfactory style. Taking the well-known Van Gogh painting style (e.g., row 3 and row 4) as an example, they fail to capture fine-grained style information in the reference image. The fine-tuning-based methods, i.e., DreamBooth and StyleDrop, result in distorted content. For example, they gen", + "bbox": [ + 212, + 724, + 787, + 800 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "X. Cui et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "4 https://github.com/aim-uofa/StyleDrop-PyTorch", + "bbox": [ + 217, + 810, + 584, + 825 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "$^{5}$ https://github.com/huggingface/diffusers", + "bbox": [ + 220, + 825, + 542, + 839 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/b97c23e07180451ef7bc9a887f4f418f38fe01f180ef884b7ec12315e3f35978.jpg", + "table_caption": [ + "Table 1: (a) Quantitative comparison regarding the style consistency score and content fidelity score. (b) User study. The results show the percentage of votes where the comparison method is preferred over ours, ties with ours and is inferior to ours. We use the abbreviation (ST) to represent approaches for style transfer tasks.", + "(a)" + ], + "table_footnote": [], + "table_body": "
MethodStyle (↑)Content (↑)
(ST) AdaAttN [39]0.5870.287
(ST) CCPL [66]0.5420.286
(ST) StyTr2 [6]0.5600.282
(ST) CAST [72]0.5940.285
(ST) QuantArt [21]0.5540.289
(ST) AesPA-Net [19]0.6040.284
Textual Inversion [10]0.5620.278
Custom Diffusion [31]0.6110.280
DreamBooth [48]0.5610.244
StyleDrop [52]0.6020.248
Ours (INSTSTYLE)0.6550.294
", + "bbox": [ + 233, + 234, + 486, + 375 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/2ecf6b175cb53e3bc4fcf708d6c7c25790e9474b2358d467b1c2505d94f0e61d.jpg", + "table_caption": [ + "(b)" + ], + "table_footnote": [], + "table_body": "
MethodPreferenceTieOurs
(ST) AdaAttN [39]0.260.130.61
(ST) CCPL [66]0.300.230.47
(ST) StyTr2 [6]0.270.200.53
(ST) CAST [72]0.310.260.43
(ST) QuantArt [21]0.120.230.65
(ST) AesPA-nNet [19]0.230.220.55
Textual Inversion [10]0.090.130.78
Custom Diffusion [31]0.090.150.76
DreamBooth [48]0.240.170.56
StyleDrop [52]0.270.260.47
", + "bbox": [ + 516, + 234, + 767, + 356 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "erate bicycles with leaked content from the reference image in the last row. In contrast, benefiting from the style signal in the inversion noise, our INSTASTYLE can generate stylized images with fine-grained style details and higher fidelity. Take the bicycle in Fig. 4 as an example, our method can better preserve the color, textures, and brushstroke. Besides, the generated bicycle is more accurate. Fig. 5 presents the qualitative comparisons with style transfer methods. Despite the inherent challenge of utilizing text as content input in contrast to style transfer methods that use the content image as input, our INSTASTYLE demonstrates comparable performance in content fidelity. Additionally, our INSTASTYLE exhibits superior performance in preserving the style details of the reference image, notably in terms of color and brushstroke characteristics. More qualitative results are shown in Figs. 9 to 13 in the Supplementary Material.", + "bbox": [ + 212, + 409, + 787, + 590 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative results. We follow evaluation measures in prior works [10,52]. We utilize 100 objects in CIFAR100 [30] as the target objects. The generated images are evaluated from two aspects. For style consistency, we calculate the CLIP score [44] between the generated image and the reference image. For content fidelity, CLIP score [44] is calculated between the image and the prompt.", + "bbox": [ + 212, + 594, + 787, + 671 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As shown in Tab. 1 (a), INSTASTYLE achieves the highest style consistency and content fidelity, indicating that the images generated by INSTASTYLE exhibit a consistent style with the reference image while retaining its generation capability. For the comparison methods, Textual Inversion and Custom Diffusion have a lower style consistency, falling short of INSTASTYLE in style preservation. As for DreamBooth and StyleDrop, our approach surpasses them in generating target objects with a higher content score. We also compare with style transfer methods [6,19,21,39,66,72] which utilize content images as input. Notably, our approach diverges by only employing text prompts as content input. This brings challenges in terms of generating high-quality content as text prompts contain less information than content images. Despite this distinction, we achieve an", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "INSTASTYLE", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "improved content score. Additionally, our approach excels in preserving style details, exhibiting a higher style consistency score.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "User study. Given the highly subjective nature of stylized generation and the inherent biases in content alignment and style consistency, we conduct a user study following previous works [6, 52]. Each question presents participants with a pair of images: one stylized image generated by our approach and another produced by a selected comparison method. To avoid bias, the images are anonymized and randomized for each question. Participants are tasked with evaluating which result they believed exhibited better stylization effects and content structures. To ensure robustness and reliability, we invite 20 participants for the user study. We compare our method with 10 existing methods, including the style transfer methods [6, 18, 21, 39, 66, 72] and current comparable methods [10, 31, 48, 52]. For each method, 30 generated image pairs are randomly selected. Finally, each participant completed 300 rounds of comparisons, resulting in a total of 6,000 votes across all methods. We count the votes and show the statistical results in Tab. 1 (b). Our approach obtains a higher preference, underscoring its superiority over competitors in both style preservation and content generation.", + "bbox": [ + 212, + 176, + 787, + 402 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ddc98eed017386d9aacf4897124584d45d4e45043821f6481d8d2ab6a5cd0cd3.jpg", + "image_caption": [ + "Fig. 6: Visualization of the combination of two styles. In each case, we show our combination results (the biggest image). For better comparison, we also present the stylized generation results of style1 and style2 in the top left (blue box) and bottom left (red box), respectively. Objects for synthesis are Boat, Truck, Horse, Elephant, Pear, Lion, Plate, Pot, Helicopter, Bicycle, Palace, and Motorcycle." + ], + "image_footnote": [], + "bbox": [ + 215, + 430, + 274, + 604 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/08c824084edf5b72d9e5b220c3f1e0e1226b1893a8afded2055c0d8853171490.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 430, + 415, + 606 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f8d2d26bed1296db83e55389e0b19a7c9fcceefb533625c708b4b4e8069954c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 430, + 537, + 604 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/82e18b05baf7c9fa1db32c84e3cca3ae22b7db7774eaf5f6ed4e8a03996ea6cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 549, + 430, + 661, + 604 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9fbc5de0087b186f10f8abda4b9334c6691931f27bdd6402aaa9477563c972a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 430, + 784, + 604 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Combination of Two Styles", + "text_level": 1, + "bbox": [ + 215, + 753, + 485, + 768 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We present style combination results in Fig. 6, where we set $\\alpha = 0.5$ and $\\beta = 0.5$ to illustrate a more obvious effect of style combination. Specifically, in each case, we also show the stylized generation results of style 1 (the top left) and style 2 (bottom left), respectively. The biggest image on the right is our combination", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "X. Cui et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/90e0317c5d4ede5d33bd0924b11ddf8f1f3121ebb4d02ac49c021304b98895c4.jpg", + "image_caption": [ + "Fig. 7: Visualization of the combination of two styles. The style can be controlled via the noise mix ratio and the prompt mix ratio. Our approach enables continuous style combinations, demonstrating its flexibility and diversity." + ], + "image_footnote": [], + "bbox": [ + 328, + 143, + 674, + 400 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "results, showing the powerful style combination ability of our approach. More visualizations are shown in. Figs. 14 and 15 in Supplementary Material.", + "bbox": [ + 212, + 484, + 784, + 513 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As introduced in Sec. 3.4, both the noise mix ratio $\\alpha$ and the prompt mix ratio $\\beta$ can affect the style. Fig. 7 illustrates the impact of the two parameters, where each row shows a different noise mix ratio ( $\\alpha = 0.1, 0.3, 0.5, 0.7, 0.9$ ) and each column shows a different prompt mix ratio ( $\\beta = 0.1, 0.3, 0.5, 0.7, 0.9$ ). The style is mainly influenced by the noise and the prompt further improves the style details, showing that our approach is flexible and can generate diverse results.", + "bbox": [ + 212, + 515, + 785, + 606 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 630, + 388, + 645 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our approach achieves stylized generation from two perspectives. One is the \"style\" noise obtained via DDIM inversion, which is utilized as the initial noise during inference to provide fine-grained style information. The other is the style token, which is learned via prompt refinement tuning to describe the style precisely. In this section, we conduct the ablation study. The visualization results and quantitative results are shown in Fig. 8 (a) and Fig. 8 (b), respectively. Specifically, each point in Fig. 8 (b) presents the average style consistency score and content fidelity score of 100 generated images for a reference style.", + "bbox": [ + 212, + 657, + 787, + 777 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Impact of inversion noise. A prominent advantage of our INSTASTYLE is that we utilize the inversion noise as the initial image during inference time to preserve the fine-grained style information. Thereby, we first conduct an ablation study by replacing the inversion \"style\" noise with random noise. As shown in", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "INSTASTYLE", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/27548266a07c0e682ddb7c8380909a1d562657bb15c4fac7d54b7559f5a08dc2.jpg", + "image_caption": [ + "Fig. 8: Ablation study. (a) Visualization of ablation study. The reference style image is shown in the bottom right corner. Objects for synthesis are *Dinosaur*, *Bicycle*, *Bee*, and *Boy*, respectively. Without the inversion noise, the stylized performance is seriously degraded. Prompt refinement can further improve style details. (b) Quantitative results of ablation study. Our method lies further along the top right corner, showing better style preservation and content generation capability." + ], + "image_footnote": [], + "bbox": [ + 217, + 141, + 787, + 313 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Fig. 8 (a) (second row), without the inversion noise, the stylized performance is seriously degraded. The quantitative results in Fig. 8 (b) further illustrate that ablating the \"style\" noise (blue dots) will result in a lower style consistency score, harming the style of the generated image.", + "bbox": [ + 212, + 450, + 787, + 513 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Impact of prompt refinement tuning. We ablate the prompt refinement stage to show the necessity of learning a style token to avoid ambiguity and bias in human-writing textual style tokens. Fig. 8 (a) (third row) reveals that our learnable style token can better describe the style without harming the generation ability. The quantitative results in Fig. 8 (b) show that ablating the prompt refinement (green dots) will result in a lower content fidelity score and style score, further proving the importance of the learnable style token.", + "bbox": [ + 212, + 513, + 787, + 619 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 648, + 359, + 666 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our research solves the task of stylized generation by exploring the fine-grained style information in the reference image. Through extensive analysis and empirical investigation, we find that the initial noise with a non-zero signal-to-noise ratio in the diffusion model leans toward a particular style. Therefore, DDIM inversion is introduced to obtain an initial noise containing fine-grained style information. Furthermore, we highlight the challenge caused by the internal bias in textual style tokens and propose to learn a style token adaptively to better describe the style without bias. Our approach produces satisfactory stylized generation results and supports the combination of two styles. Extensive qualitative and quantitative comparisons demonstrate the superiority of our method.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "X. Cui et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement This research is sponsored by National Natural Science Foundation of China (Grant No. 62306041, U21B2045, 62176025), Beijing Nova Program (Grant No. Z211100002121106, 20230484488, 20230484276), Youth Innovation Promotion Association CAS (Grant No.2022132), and Beijing Municipal Science & Technology Commission (Z231100007423015).", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 246, + 323, + 263 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alaluf, Y., Richardson, E., Metzer, G., Cohen-Or, D.: A neural space-time representation for text-to-image personalization. ACM TOG (2023)", + "2. Chen, H., Zhao, L., Wang, Z., Zhang, H., Zuo, Z., Li, A., Xing, W., Lu, D.: Dualast: Dual style-learning networks for artistic style transfer. In: CVPR (2021)", + "3. Cho, J., Nam, G., Kim, S., Yang, H., Kwak, S.: Promptstyler: Prompt-driven style generation for source-free domain generalization. In: ICCV (2023)", + "4. Cui, X., Li, P., Li, Z., Liu, X., Zou, Y., He, Z.: Localize, understand, collaborate: Semantic-aware dragging via intention reasoner. arXiv preprint arXiv:2406.00432 (2024)", + "5. Cui, X., Li, Z., Li, P., Hu, Y., Shi, H., Cao, C., He, Z.: Chatedit: Towards multi-turn interactive facial image editing via dialogue. In: EMNLP (2023)", + "6. Deng, Y., Tang, F., Dong, W., Ma, C., Pan, X., Wang, L., Xu, C.: Stytr2: Image style transfer with transformers. In: CVPR (2022)", + "7. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS (2021)", + "8. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: CVPR (2021)", + "9. Everaert, M.N., Bocchio, M., Arpa, S., Susstrunk, S., Achanta, R.: Diffusion in style. In: ICCV (2023)", + "0. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. In: ICLR (2022)", + "1. Gal, R., Patashnik, O., Maron, H., Bermano, A.H., Chechik, G., Cohen-Or, D.: Stylegan-nada: Clip-guided domain adaptation of image generators. ACM TOG (2022)", + "2. Gatys, L.A., Ecker, A.S., Bethge, M.: Image style transfer using convolutional neural networks. In: CVPR (2016)", + "3. Gatys, L.A., Ecker, A.S., Bethge, M., Hertzmann, A., Shechtman, E.: Controlling perceptual factors in neural style transfer. In: CVPR (2017)", + "4. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. CACM (2020)", + "5. Gu, B., Fan, H., Zhang, L.: Two birds, one stone: A unified framework for joint learning of image and video style transfers. In: ICCV (2023)", + "6. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)", + "7. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPSW (2021)", + "8. Hong, K., Jeon, S., Lee, J., Ahn, N., Kim, K., Lee, P., Kim, D., Uh, Y., Byun, H.: Aespa-net: Aesthetic pattern-aware style transfer networks. In: ICCV (2023)", + "9. Hong, K., Jeon, S., Lee, J., Ahn, N., Kim, K., Lee, P., Kim, D., Uh, Y., Byun, H.: Aespa-net: Aesthetic pattern-aware style transfer networks. In: ICCV (2023)" + ], + "bbox": [ + 225, + 280, + 784, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "INSTASTYLE", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "20. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: ICLR (2021)", + "21. Huang, S., An, J., Wei, D., Luo, J., Pfister, H.: Quantart: Quantizing image style transfer towards high visual fidelity. In: CVPR (2023)", + "22. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: ICCV (2017)", + "23. Jia, G., Li, P., He, R.: Theme-aware aesthetic distribution prediction with full-resolution photographs. TNNLS (2022)", + "24. Jing, Y., Liu, X., Ding, Y., Wang, X., Ding, E., Song, M., Wen, S.: Dynamic instance normalization for arbitrary style transfer. In: AAAI (2020)", + "25. Jing, Y., Liu, Y., Yang, Y., Feng, Z., Yu, Y., Tao, D., Song, M.: Stroke controllable fast style transfer with adaptive receptive fields. In: ECCV (2018)", + "26. Jing, Y., Mao, Y., Yang, Y., Zhan, Y., Song, M., Wang, X., Tao, D.: Learning graph neural networks for image style transfer. In: ECCV (2022)", + "27. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019)", + "28. Ke, Z., Liu, Y., Zhu, L., Zhao, N., Lau, R.W.: Neural preset for color style transfer. In: CVPR (2023)", + "29. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)", + "30. Krizhevsky, A., et al.: Learning multiple layers of features from tiny images. Technical report, University of Toronto (2009)", + "31. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: CVPR (2023)", + "32. Li, P., Hu, Y., He, R., Sun, Z.: Global and local consistent wavelet-domain age synthesis. TIFS (2019)", + "33. Li, P., Liu, X., Huang, J., Xia, D., Yang, J., Lu, Z.: Progressive generation of 3d point clouds with hierarchical consistency. PR (2023)", + "34. Li, P., Wang, R., Huang, H., He, R., He, Z.: Pluralistic aging diffusion autoencoder. In: ICCV (2023)", + "35. Li, P., Wu, X., Hu, Y., He, R., Sun, Z.: M2fpa: A multi-yaw multi-pitch high-quality dataset and benchmark for facial pose analysis. In: ICCV (2019)", + "36. Lin, S., Liu, B., Li, J., Yang, X.: Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891 (2023)", + "37. Lin, T., Ma, Z., Li, F., He, D., Li, X., Ding, E., Wang, N., Li, J., Gao, X.: Drafting and revision: Laplacian pyramid network for fast high-quality artistic style transfer. In: CVPR (2021)", + "38. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: ECCV (2022)", + "39. Liu, S., Lin, T., He, D., Li, F., Wang, M., Li, X., Sun, Z., Li, Q., Ding, E.: Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In: ICCV (2021)", + "40. Mao, J., Wang, X., Aizawa, K.: Guided image synthesis via initial image editing in diffusion model. In: ACM MM (2023)", + "41. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2021)", + "42. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: CVPR (2023)", + "43. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: ICLR (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "X. Cui et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)", + "45. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021)", + "46. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)", + "47. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015)", + "48. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023)", + "49. Ruiz, N., Li, Y., Jampani, V., Wei, W., Hou, T., Pritch, Y., Wadhwa, N., Rubinstein, M., Aberman, K.: Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949 (2023)", + "50. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: ICLR (2021)", + "51. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015)", + "52. Sohn, K., Ruiz, N., Lee, K., Chin, D.C., Blok, I., Chang, H., Barber, J., Jiang, L., Entis, G., Li, Y., et al.: Styledrop: Text-to-image generation in any style. In: NeurIPS (2023)", + "53. Sohn, K., Shaw, A., Hao, Y., Zhang, H., Polania, L., Chang, H., Jiang, L., Essa, I.: Learning disentangled prompts for compositional image synthesis. arXiv preprint arXiv:2306.00763 (2023)", + "54. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2020)", + "55. Tang, H., Liu, S., Lin, T., Huang, S., Li, F., He, D., Wang, X.: Master: Meta style transformer for controllable zero-shot and few-shot artistic style transfer. In: CVPR (2023)", + "56. Tang, J., Shu, X., Qi, G.J., Li, Z., Wang, M., Yan, S., Jain, R.: Tri-clustered tensor completion for social-aware image tag refinement. TPAMI (2016)", + "57. Teng, Q., Wang, R., Cui, X., Li, P., He, Z.: Exploring 3d-aware lifespan face aging via disentangled shape-texture representations. In: ICME (2024)", + "58. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)", + "59. Wang, H., Li, Y., Wang, Y., Hu, H., Yang, M.H.: Collaborative distillation for ultra-resolution universal style transfer. In: CVPR (2020)", + "60. Wang, R., Guo, H., Liu, J., Li, H., Zhao, H., Tang, X., Hu, Y., Tang, H., Li, P.: Stablegarment: Garment-centric generation via stable diffusion. arXiv preprint arXiv:2403.10783 (2024)", + "61. Wang, R., Li, P., Huang, H., Cao, C., He, R., He, Z.: Learning-to-rank meets language: Boosting language-driven ordering alignment for ordinal classification. In: NeurIPS (2023)", + "62. Wang, Z., Zhao, L., Xing, W.: Stylediffusion: Controllable disentangled style transfer via diffusion models. In: ICCV (2023)", + "63. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In: ICCV (2023)", + "64. Wen, L., Gao, C., Zou, C.: Cap-vstnet: Content affinity preserved versatile style transfer. In: CVPR (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "INSTASTYLE", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "65. Wu, X., Hu, Z., Sheng, L., Xu, D.: Styleformer: Real-time arbitrary style transfer via parametric style composition. In: ICCV (2021)", + "66. Wu, Z., Zhu, Z., Du, J., Bai, X.: Ccpl: contrastive coherence preserving loss for versatile style transfer. In: ECCV (2022)", + "67. Xie, X., Li, Y., Huang, H., Fu, H., Wang, W., Guo, Y.: Artistic style discovery with independent components. In: CVPR (2022)", + "68. Xu, W., Long, C., Nie, Y.: Learning dynamic style kernels for artistic style transfer. In: CVPR (2023)", + "69. Xu, Z., Sangineto, E., Sebe, N.: Stylerdalle: Language-guided style transfer using a vector-quantized tokenizer of a large-scale generative model. In: ICCV (2023)", + "70. Yang, S., Hwang, H., Ye, J.C.: Zero-shot contrastive loss for text-guided diffusion image style transfer. In: ICCV (2023)", + "71. Zhang, Y., Huang, N., Tang, F., Huang, H., Ma, C., Dong, W., Xu, C.: Inversion-based style transfer with diffusion models. In: CVPR (2023)", + "72. Zhang, Y., Tang, F., Dong, W., Huang, H., Ma, C., Lee, T.Y., Xu, C.: Domain enhanced arbitrary image style transfer via contrastive learning. In: ACM SIGGRAPH (2022)", + "73. Zhang, Z., Li, B., Nie, X., Han, C., Guo, T., Liu, L.: Towards consistent video editing with text-to-image diffusion models. In: NeurIPS (2023)", + "74. Zhou, Y., Zhang, R., Sun, T., Xu, J.: Enhancing detail preservation for customized text-to-image generation: A regularization-free approach. arXiv preprint arXiv:2305.13579 (2023)", + "75. Zhu, M., He, X., Wang, N., Wang, X., Gao, X.: All-to-key attention for arbitrary style transfer. In: ICCV (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 479 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "X. Cui et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/InstaStyle_ Inversion Noise of a Stylized Image is Secretly a Style Adviser/938fb53e-3bec-4f8a-bc7e-f5b36dd7b576_model.json b/2024/InstaStyle_ Inversion Noise of a Stylized Image is Secretly a Style Adviser/938fb53e-3bec-4f8a-bc7e-f5b36dd7b576_model.json new file mode 100644 index 0000000000000000000000000000000000000000..02e0c66ea2ec82ad4846306858e30763e80504f4 --- /dev/null +++ b/2024/InstaStyle_ Inversion Noise of a Stylized Image is Secretly a Style Adviser/938fb53e-3bec-4f8a-bc7e-f5b36dd7b576_model.json @@ -0,0 +1,3877 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.225, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "INSTASTYLE: Inversion Noise of a Stylized Image is Secretly a Style Adviser" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.212, + 0.764, + 0.245 + ], + "angle": 0, + "content": "Xing Cui\\(^{1}\\) Zekun Li\\(^{2}\\) Peipei Li\\(^{1\\star}\\) Huaibo Huang\\(^{3}\\) Xuannan Liu\\(^{1}\\) Zhaofeng He\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.255, + 0.682, + 0.269 + ], + "angle": 0, + "content": "1 Beijing University of Posts and Telecommunications" + }, + { + "type": "text", + "bbox": [ + 0.364, + 0.269, + 0.641, + 0.283 + ], + "angle": 0, + "content": "2 University of California, Santa Barbara" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.283, + 0.746, + 0.297 + ], + "angle": 0, + "content": "\\(^{3}\\) MAIS & NLPR, Institute of Automation, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.289, + 0.298, + 0.714, + 0.311 + ], + "angle": 0, + "content": "{cuixing, lipeipei, liuxuannan, zhaofenghe}@bupt.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.312, + 0.697, + 0.325 + ], + "angle": 0, + "content": "zekunli@cs.ucsb.edu huaibo.huang@cripac.ia.ac.cn" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.326, + 0.706, + 0.339 + ], + "angle": 0, + "content": "https://cuixing100876.github.io/instastyle.github.io/" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.365, + 0.315, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.316, + 0.366, + 0.385, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.366, + 0.45, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.366, + 0.522, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.478, + 0.456, + 0.488 + ], + "angle": 0, + "content": "(a) Stylized image generation" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.366, + 0.617, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.366, + 0.686, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.366, + 0.757, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.478, + 0.733, + 0.488 + ], + "angle": 0, + "content": "(b) Combination of two styles" + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.49, + 0.315, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.49, + 0.397, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.49, + 0.467, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.49, + 0.536, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.49, + 0.605, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.49, + 0.676, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.49, + 0.739, + 0.5 + ], + "angle": 0, + "content": "Style 2" + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.5, + 0.757, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.558, + 0.622, + 0.569 + ], + "angle": 0, + "content": "(c) Smooth style combination between two styles" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.581, + 0.789, + 0.651 + ], + "angle": 0, + "content": "Fig.1: Visualization of INSTASTYLE. (a) Our method excels at capturing style details and distinguishing between similar styles. (b) The first and third columns show images styled with reference to style 1 and style 2, respectively. The middle column shows images in a combined style. (c) Our method supports adjusting the degree of two styles during combination, dynamically ranging from one style to another." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.677, + 0.741, + 0.816 + ], + "angle": 0, + "content": "Abstract. Stylized text-to-image generation focuses on creating images from textual descriptions while adhering to a style specified by reference images. However, subtle style variations within different reference images can hinder the model from accurately learning the target style. In this paper, we propose INSTASTYLE, a novel approach that excels in generating high-fidelity stylized images with only a single reference image. Our approach is based on the finding that the inversion noise from a stylized reference image inherently carries the style signal, as evidenced by their non-zero signal-to-noise ratio. We employ DDIM inversion to extract this noise from the reference image and leverage a" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.385, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "X. Cui et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.26 + ], + "angle": 0, + "content": "diffusion model to generate new stylized images from the \"style\" noise. Additionally, the inherent ambiguity and bias of textual prompts impede the precise conveying of style during image inversion. To address this, we devise prompt refinement, which learns a style token assisted by human feedback. Qualitative and quantitative experimental results demonstrate that INSTASTYLE achieves superior performance compared to current benchmarks. Furthermore, our approach also showcases its capability in the creative task of style combination with mixed inversion noise." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.273, + 0.744, + 0.303 + ], + "angle": 0, + "content": "Keywords: Stylized image generation \\(\\cdot\\) Inversion noise \\(\\cdot\\) Signal-to-noise ratio \\(\\cdot\\) Prompt refinement" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.333, + 0.376, + 0.349 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.489, + 0.612 + ], + "angle": 0, + "content": "Recently, the advent of DDPMs [16, 51] has ushered in a new era of image generation. Existing text-to-image generation methods [5, 43, 45, 46] can generate images in coarse-grained styles by incorporating style descriptions in the textual prompt. However, the textual prompt is ambiguous which makes it hard to express style precisely [11]. Recently, personalized generation [10,48,63] has been introduced to generate novel concepts, which can be utilized for stylized generation by viewing the reference style as a new concept. These approaches typically necessitate mul" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.734 + ], + "angle": 0, + "content": "tiple images as references. However, the subtle style variations in these multiple images present a challenge for the model to accurately learn and replicate the intended style. As shown in the first column of Fig. 1 (a), although the reference image Houses at Auvers (top) and Thatched Cottages by a Hill (bottom) are all Van Gogh's artworks, the differences in terms of color and texture [2] pose a significant challenge. Some recent approaches [3,53,69] are designed to more accurately describe the target style during generation. However, they may ignore the fine-grained style information in the reference images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we propose INSTASTYLE (Inversion Noise of a Stylized Image is Secretly a Style Adviser) based on the observation that the inversion noise from a stylized reference image inherently contains the style signal. Specifically, our approach only requires a single reference image, which is transformed into a noise via DDIM inversion. The inversion noise, which preserves the style signal, is then used to generate stylized images by utilizing diffusion models. To shed light on this phenomenon, we demonstrate that the inversion noise maintains" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.375, + 0.789, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.491, + 0.789, + 0.589 + ], + "angle": 0, + "content": "Fig. 2: Motivation. Sampling from the inversion noise of a reference image can generate stylized images. However, the optimal style token varies for each case among human-written style tokens. Our learnable style token, i.e., \"\", shows greater universality across various scenarios." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "INSTASTYLE" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "a non-zero signal-to-noise ratio in Sec. 3.2, indicating that it retains essential information (including the style signal) from the reference image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.448 + ], + "angle": 0, + "content": "Nevertheless, directly using human-written prompts to describe styles for inverting the reference image often encounters challenges due to the inherent ambiguity of natural language. For instance, specific style descriptors like \"ink painting\" don't always work effectively with various objects. As shown in Fig. 2, the style \"ink painting\" might suit an object like a \"chair\" but not as well with a \"boy\". This inconsistency can arise because \"ink painting\" is typically associated with landscapes and might not yield optimal results for human subjects. Conversely, when using vague descriptors like \"specific style\", they may fail to provide enough information, leading to unpredictable generation quality (see the second row in Fig. 2). To address this, we propose Prompt Refinement to learn a style token with the assistance of human feedback. As illustrated in the last row of Fig. 2, the style information is boosted by using the learned style token in the noising process to obtain a better \"style\" noise. Therefore, our prompt refinement stands apart from previous methods [3, 53, 69], where style tokens are utilized in the denoising process to generate target images. During prompt refinement, we first collect images via human feedback. Then we optimize the embedding of the style token and the key and value projection matrices in the cross-attention layers of the diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.526 + ], + "angle": 0, + "content": "As shown in Fig. 1 (a), our approach can effectively retain the fine-grained style in the reference image as well as generate new objects in high fidelity. Furthermore, our method supports the creative generation of style combination (Fig. 1 (b) and (c)) and allows adjusting the degree of two styles dynamically (Fig. 1 (c)). To sum up, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.534, + 0.784, + 0.578 + ], + "angle": 0, + "content": "- We find that the inversion noise of a reference image via DDIM inversion potentially retains style information and evidence this observation by the non-zero signal-to-noise ratio of the inversion noise." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.58, + 0.785, + 0.653 + ], + "angle": 0, + "content": "- We propose INSTASTYLE based on our observation, which leverages the style signal in inversion noise. Additionally, to better represent the style during inversion, a prompt refinement scheme is designed to learn the style token. Human feedback is also introduced to enhance model alignment with human preferences during prompt refinement." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.655, + 0.784, + 0.699 + ], + "angle": 0, + "content": "- Qualitative and quantitative experimental results show the superiority of our INSTASTYLE in generating high-quality stylized images. Moreover, it exhibits promising potential in creative scenarios like style combinations." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.534, + 0.785, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.723, + 0.388, + 0.739 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.461, + 0.77 + ], + "angle": 0, + "content": "2.1 Text-to-image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Image synthesis is an essential subject in computer vision [7,23,41,56]. Previous works are based on variational auto-encoder [29, 33] or generative adversarial networks [5,14,32,35,57]. With the advance of pretrained vison-language models [44,61] and diffusion models [34,54,73], text-to-image generation [4,8,43,45," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "X. Cui et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "[60] has been widely studied and shown remarkable generalization ability. Recently, Lin et al. [36] have pointed out that the noise of the last step during noisering in the training process has a non-zero signal-to-noise ratio (SNR), i.e., there exists a signal leak. This results in a misalignment between the training and inference process. Therefore, some works [36, 50] propose to train diffusion models by enforcing the SNR to zero to avoid the signal leak. On the contrary, our approach leverages the leaked signal, which potentially includes the style details, from the reference image for the stylized generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.406 + ], + "angle": 0, + "content": "As these methods ignore the concepts that do not appear in the training set [10], some works [10,48,49] study personalized text-to-image generation which aims to adapt text-to-image models to new concepts given several reference images. For example, Textual Inversion [10] introduces and optimizes a word vector for each new concept. Subsequent works [1,63,74] further enhance the flexibility and adaptiveness of the learning strategy. Some methods [20, 48, 49] propose to finetune the original diffusion model, showing more satisfactory results. However, maintaining fine-grained style while simultaneously generating new objects remains a challenge for current personalized generation methods [52]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.431, + 0.482, + 0.447 + ], + "angle": 0, + "content": "2.2 Stylized Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.788, + 0.673 + ], + "angle": 0, + "content": "Stylized image generation is a new paradigm for image generation which aims to generate content in a specific style given a few reference images. Although it is similar to the neural style transfer task [12, 13, 22, 27, 37, 59] which generates stylized images as well, they are fundamentally different. Style transfer solves an image translation task, focusing on translating a content image to target style [15, 21, 24, 28, 62, 64, 67, 68, 70]. For example, some works explore the global and local information to preserve the content of the source image [6, 18, 25, 26, 55, 65, 71, 75]. On the contrary, stylized image generation is geared towards generating a new image in specific style conditioned on a text. For example, ZDAIS [53] views stylized image generation as a domain adaptive task, where each style belongs to a domain. It learns disentangled prompts to adapt the model to new domains. Some approaches [9, 52] fine-tune the model to capture style properties. Our work differs from theirs by proposing a framework based on our observation that inversion noise contains style information." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.699, + 0.331, + 0.715 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "As shown in Fig. 3, our proposed method involves two main stages. In the first stage, we employ DDIM inversion to transform the reference image into noise. Notably, the inversion noise exhibits a non-zero signal-to-noise ratio, suggesting the presence of style signals from the reference image. Subsequently, we generate \\(M\\) stylized images from the inversion noise conditioned on the given textual prompts. Due to the inherent ambiguity of the textual prompts that describe the style, it is challenging to precisely convey the desired style. Addressing this," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "INSTASTYLE" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.788, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.291, + 0.788, + 0.361 + ], + "angle": 0, + "content": "Fig. 3: The training process of INSTASTYLE. (a) The first stage is initial stylized image generation. The reference image is inverted to noise conditioned on a prompt via DDIM Inversion. Then the inversion noise is utilized to generate initial stylized images. (b) The second stage is prompt refinement which leverages the selected high-quality initial stylized images to learn a style token." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.437 + ], + "angle": 0, + "content": "the second stage involves the incorporation of human feedback to select \\( N \\) high-quality generated images from the first stage. The selected images are then used to learn a style token via prompt refinement." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.458, + 0.373, + 0.473 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.788, + 0.558 + ], + "angle": 0, + "content": "We begin by reviewing the fundamental diffusion model. Subsequently, we introduce Stable Diffusion [46], a key component of our framework. Finally, we discuss both DDIM and DDIM inversion. In our method, DDIM serves the purpose of denoising a noise to generate new images. Concurrently, DDIM inversion is utilized to transform the reference image into a corresponding \"style\" noise." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.558, + 0.788, + 0.604 + ], + "angle": 0, + "content": "Diffusion models. Diffusion models [16, 51] contain a forward process and a backward process. The forward process adds noise to the data according to a predetermined, non-learned variance schedule \\(\\beta_{1},\\ldots,\\beta_{T}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.614, + 0.787, + 0.632 + ], + "angle": 0, + "content": "\\[\nq \\left(z _ {t} \\mid z _ {t - 1}\\right) := \\mathcal {N} \\left(z _ {t}; \\sqrt {1 - \\beta_ {t}} z _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.787, + 0.673 + ], + "angle": 0, + "content": "An important property of the forward process is that we can obtain the latent variable \\( z_{t} \\) directly based on \\( z_{0} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.683, + 0.787, + 0.701 + ], + "angle": 0, + "content": "\\[\nz _ {t} = \\sqrt {\\alpha_ {t}} z _ {0} + \\sqrt {1 - \\alpha_ {t}} \\varepsilon , \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.71, + 0.787, + 0.742 + ], + "angle": 0, + "content": "where \\(\\varepsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\), \\(\\alpha_{t}\\coloneqq \\prod_{i = 1}^{t}(1 - \\beta_{i})\\). Diffusion models restore the information by learning the backward process:" + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.753, + 0.787, + 0.771 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(z _ {t - 1} \\mid z _ {t}\\right) := \\mathcal {N} \\left(z _ {t - 1}; \\mu_ {\\theta} \\left(z _ {t}, t\\right), \\Sigma_ {\\theta} \\left(z _ {t}, t\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Stable Diffusion. In Stable Diffusion, the diffusion process happens in latent space [46]. It utilizes a pretrained autoencoder which consists of an encoder and a decoder. The encoder \\(\\mathcal{E}(\\cdot)\\) maps an image \\(x\\) into a latent code \\(z\\) and the decoder \\(\\mathcal{D}(\\cdot)\\) transforms the latent code back into an image. The denoising model \\(\\varepsilon_{\\theta}(\\cdot)\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.357, + 0.128 + ], + "angle": 0, + "content": "X. Cui et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "is an Unet [47]. During training, artificial noise is added to the sampled data \\( z_0 \\) according to timestamp \\( t \\) based on Eq. (2), resulting in a noised sample \\( z_t \\). Denoising model starts with \\( z_t \\) and is trained to predict the artificial noise:" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.203, + 0.786, + 0.226 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} E _ {z _ {0}, \\varepsilon \\sim N (0, I), t \\sim \\operatorname {U n i f o r m} (1, T)} \\| \\varepsilon - \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C}\\right) \\| _ {2} ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.236, + 0.608, + 0.251 + ], + "angle": 0, + "content": "where \\(\\mathcal{C} = \\psi (\\mathcal{P})\\) is the embedding of text condition \\(\\mathcal{P}\\)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.251, + 0.784, + 0.296 + ], + "angle": 0, + "content": "DDIM. In inference time, given a noise vector \\( z_{T} \\), the noise is gradually removed by sequentially predicting the added noise for \\( T \\) steps. DDIM [54] is one of the denoising approaches with a deterministic process:" + }, + { + "type": "equation", + "bbox": [ + 0.311, + 0.307, + 0.786, + 0.337 + ], + "angle": 0, + "content": "\\[\nz _ {t - 1} = \\sqrt {\\frac {\\alpha_ {t - 1}}{\\alpha_ {t}}} z _ {t} + \\left(\\sqrt {\\frac {1}{\\alpha_ {t - 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\cdot \\tilde {\\varepsilon} _ {\\theta}, \\qquad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.345, + 0.447, + 0.359 + ], + "angle": 0, + "content": "where \\(\\tilde{\\varepsilon}_{\\theta}\\) is the estimated noise." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.36, + 0.784, + 0.391 + ], + "angle": 0, + "content": "DDIM inversion [42, 54] transforms an image to a noise conditioned on a prompt. The diffusion process is performed in the reverse direction, i.e., \\( z_0 \\to z_T \\):" + }, + { + "type": "equation", + "bbox": [ + 0.311, + 0.402, + 0.786, + 0.431 + ], + "angle": 0, + "content": "\\[\nz _ {t + 1} = \\sqrt {\\frac {\\alpha_ {t + 1}}{\\alpha_ {t}}} z _ {t} + \\left(\\sqrt {\\frac {1}{\\alpha_ {t + 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\cdot \\tilde {\\varepsilon} _ {\\theta}. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.45, + 0.538, + 0.466 + ], + "angle": 0, + "content": "3.2 Initial Stylized Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.474, + 0.784, + 0.504 + ], + "angle": 0, + "content": "As shown in Fig. 3 (a), the first stage of our INSTASTYLE involves obtaining the inversion noise via DDIM inversion and sampling images via DDIM." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.505, + 0.786, + 0.626 + ], + "angle": 0, + "content": "In DDIM inversion, the added noise for each step is calculated conditioned on a prompt and gradually incorporated into the reference image following Eq. (6). Specifically, we set the learnable style token as a human-written style description in this stage. We demonstrate that the inversion noise from a stylized reference image inherently carries the style signal from the perspective of the signal-to-noise ratio. As the estimated noise \\(\\varepsilon_{\\theta}(z_t,t,\\mathcal{C})\\) is trained to approximate the artificial noise \\(\\varepsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\), we assume that \\(\\varepsilon_{\\theta}(z_t,t,\\mathcal{C}) \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). According to Eq. (6), \\(z_{t + 1}\\) can be approximated in a closed form:" + }, + { + "type": "equation", + "bbox": [ + 0.282, + 0.638, + 0.786, + 0.67 + ], + "angle": 0, + "content": "\\[\nz _ {t + 1} = \\sqrt {\\frac {\\alpha_ {t + 1}}{\\alpha_ {0}}} z _ {0} + \\sqrt {\\sum_ {i = 0} ^ {t} \\frac {\\alpha_ {t + 1}}{\\alpha_ {i + 1}} \\left(\\sqrt {\\frac {1}{\\alpha_ {i + 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {i}} - 1}\\right) ^ {2}} \\cdot \\bar {\\varepsilon} _ {0}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.679, + 0.36, + 0.695 + ], + "angle": 0, + "content": "where \\(\\bar{\\varepsilon}_0\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.695, + 0.784, + 0.739 + ], + "angle": 0, + "content": "Signal-to-Noise ratio. Signal-to-noise ratio (SNR) is introduced to measure the ratio of signals from the original image preserved in the noise [36,46]. The SNR of the inversion noise can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.75, + 0.786, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S N R} (t) := \\frac {1}{\\sum_ {i = 0} ^ {t} \\frac {\\alpha_ {0}}{\\alpha_ {i + 1}} \\left(\\sqrt {\\frac {1}{\\alpha_ {i + 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {i}} - 1}\\right) ^ {2}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.795, + 0.784, + 0.843 + ], + "angle": 0, + "content": "We present detailed derivations in Sec. B in Supplementary Material. In Stable Diffusion [46], the predetermined variance schedule \\(\\beta_{t} = (\\sqrt{0.00085}\\cdot (1 - j) + \\sqrt{0.012}\\cdot j)^{2}\\), where \\(j = \\frac{t - 1}{T - 1}\\). At the terminal timestep \\(T = 1000\\), the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "INSTASTYLE" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "\\(SNR(T) = 0.015144\\), i.e., \\(z_{T}\\) has a none-zero signal-to-noise ratio. This nonzero signal-to-noise ratio suggests that the noise obtained via DDIM inversion still retains information from the reference image and deviates from white noise. Besides, we conduct qualitative experiments and observe that style information can consistently be preserved at the terminal timestep (refer to Sec. C.3 in Supplementary Material). Therefore, we can generate target stylized images by sampling from the inversion noise based on Eq. (5)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.275, + 0.427, + 0.29 + ], + "angle": 0, + "content": "3.3 Prompt Refinement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.3, + 0.785, + 0.405 + ], + "angle": 0, + "content": "As the natural language prompts may not precisely convey the style (Fig. 2), we propose Prompt Refinement to learn the style tokens as shown in Fig. 3 (b). We utilize the generated stylized images in the first stage to constitute the training data for prompt refinement learning. While these generated images may not always be precise, we observe that most of them successfully retain the style information alongside appropriate content. We manually select images that distinctly embody the reference style and the target object to build the dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.787, + 0.542 + ], + "angle": 0, + "content": "Specifically, we introduce a new token, i.e., “” to represent the style descriptor and learn its embedding. In practice, we initialize the new token with the embeddings of the textual style descriptor of the reference image. The text condition is input to the model via two projection matrices in the cross-attention block of diffusion model \\(\\varepsilon_{\\theta}\\), i.e., \\(W^{k} \\in \\mathbb{R}^{d \\times d'}\\) and \\(W^{v} \\in \\mathbb{R}^{d \\times d'}\\). Therefore, we also fine-tune these two projection matrices. The text feature \\(\\mathbf{c} \\in \\mathbb{R}^{s \\times d}\\) is projected to key \\(K = \\mathbf{c}W^{k}\\) and value \\(V = \\mathbf{c}W^{v}\\). The query matrix \\(W^{q} \\in \\mathbb{R}^{l \\times d'}\\) projects the latent image feature \\(\\mathbf{f} \\in \\mathbb{R}^{(h \\times w) \\times l}\\) into query feature \\(Q = \\mathbf{f}W^{q}\\). Then the cross-attention [58] is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.554, + 0.785, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d ^ {\\prime}}}\\right) V. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.597, + 0.784, + 0.703 + ], + "angle": 0, + "content": "We utilize the LoRA [20] for model fine-tuning. Specifically, for a projection matrix \\( W \\in \\mathbb{R}^{d \\times d'} \\) (i.e., \\( W = W^k \\) or \\( W = W^v \\)) to be fine-tuned, we update a low-rank residual rather than directly fine-tuning \\( W \\). Formally, we denote the fine-tuned projection matrix as \\( W' = W + BA \\), where \\( B \\in \\mathbb{R}^{d \\times r} \\), \\( A \\in \\mathbb{R}^{r \\times d'} \\), and the rank \\( r \\ll \\min(d, d') \\). During training, only \\( A \\) and \\( B \\) are learnable. For \\( y = Wx \\), the modified forward pass is \\( y = Wx + BAx \\). Putting the two stages together, our full algorithm is shown in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.725, + 0.341, + 0.739 + ], + "angle": 0, + "content": "3.4 Inference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Stylized image generation. For stylized image generation, we first learn the style token \"\" to describe the style in the reference image. During inference, the learned style token is integrated into the prompt that inverts image and the prompt that generates the target image. Besides, the diffusion model \\(\\varepsilon_{\\theta}\\) is utilized as the backbone of DDIM inversion and DDIM. Specifically, the reference image is first inverted to a noise \\(z_{T}\\) via DDIM inversion conditioned" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "X. Cui et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.23, + 0.151, + 0.434, + 0.165 + ], + "angle": 0, + "content": "Algorithm 1: INSTASTYLE" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.169, + 0.761, + 0.197 + ], + "angle": 0, + "content": "1. Input: A source prompt embedding \\(\\mathcal{C} = \\psi(\\mathcal{P})\\) and a reference image \\(\\mathcal{I}\\). Initial style token embedding \\(v\\) is included in the prompt embedding \\(\\mathcal{C}\\)." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.198, + 0.71, + 0.212 + ], + "angle": 0, + "content": "2 Output: Optimized style token embedding \\( v^{*} \\) and diffusion model \\( \\varepsilon_{\\theta}^{*} \\)." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.169, + 0.761, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.219, + 0.521, + 0.232 + ], + "angle": 0, + "content": "// Initial Stylized Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.233, + 0.764, + 0.26 + ], + "angle": 0, + "content": "3 Compute the inversion noise \\( z_{T} \\) using DDIM inversion over \\( \\mathcal{I} \\) conditioned on \\( \\mathcal{C} \\) based on Eq. (6);" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.26, + 0.758, + 0.288 + ], + "angle": 0, + "content": "4 Generate \\(M\\) images conditioned on a predefined prompt set based on Eq. (5); // Prompt Refinement" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.288, + 0.758, + 0.316 + ], + "angle": 0, + "content": "5 Select \\( N \\) images from \\( M \\) images with human feedback as the training dataset \\( q(\\mathbf{x}_0, c) \\);" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.233, + 0.764, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.318, + 0.294, + 0.329 + ], + "angle": 0, + "content": "6 repeat" + }, + { + "type": "code", + "bbox": [ + 0.226, + 0.33, + 0.566, + 0.417 + ], + "angle": 0, + "content": "7 \\(\\mathbf{x}_0,c\\sim q(\\mathbf{x}_0,c)\\) \n8 \\(z_{0} = \\mathcal{E}(\\mathbf{x}_{0})\\) \n9 \\(t\\sim \\mathrm{Uniform}(\\{1,\\dots ,T\\})\\) \n10 \\(\\varepsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) \n11 \\(z_{t} = \\sqrt{\\alpha_{t}} z_{0} + \\sqrt{1 - \\alpha_{t}}\\varepsilon\\) \n12 Take gradient step on \\(\\nabla_{\\theta ,v}\\| \\varepsilon -\\varepsilon_{\\theta}(z_t,t,c)\\| _2^2;\\)" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.418, + 0.354, + 0.431 + ], + "angle": 0, + "content": "13 until converged;" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.431, + 0.343, + 0.445 + ], + "angle": 0, + "content": "14 Return \\( v^{*} \\), \\( \\varepsilon_{\\theta}^{*} \\)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.418, + 0.354, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.542 + ], + "angle": 0, + "content": "on a prompt containing the learned style token. Then we sample from the inversion noise conditioned on a target prompt which contains the target content and the learned style token via DDIM. Specifically, the estimated noise is calculated based on the classifier-free guidance [17]:" + }, + { + "type": "equation", + "bbox": [ + 0.263, + 0.55, + 0.786, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\varepsilon} _ {\\theta} \\left(z _ {t}, t, \\mathcal {C}, \\varnothing\\right) = \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\varnothing\\right) + w \\cdot \\left(\\varepsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C}\\right) - \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\varnothing\\right)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.787, + 0.607 + ], + "angle": 0, + "content": "where \\(\\varnothing = \\psi(\")\\) is the embedding of a null text. \\(\\varepsilon_{\\theta}(z_t, t, \\mathcal{C})\\) represents the conditional predictions. \\(w\\) is the guidance scale parameter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.788, + 0.76 + ], + "angle": 0, + "content": "Combination of two styles. To combine two styles, we learn style tokens for each style, i.e., “” and “”. Specifically, we use the selected images of both styles to constitute the training set. Then we jointly optimize the embeddings of the style tokens and fine-tune the key projection and value projection in the cross-attention block. During inference, we individually transform two reference images into their corresponding inversion noise \\( z_{t1} \\in \\mathbb{R}^{H \\times W \\times C} \\) and \\( z_{t2} \\in \\mathbb{R}^{H \\times W \\times C} \\) via DDIM inversion conditioned on a prompt containing their learned style token. As the style can be described by both the “style” noise and style token, we combine the styles from these two perspectives. For the “style” noise, the combined inversion noise \\( z_{t} \\) is obtained based on a masking strategy:" + }, + { + "type": "equation", + "bbox": [ + 0.39, + 0.769, + 0.786, + 0.786 + ], + "angle": 0, + "content": "\\[\nz _ {t} = (\\mathbf {1} - \\mathbf {M}) \\odot z _ {t 1} + \\mathbf {M} \\odot z _ {t 2}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "where \\(\\odot\\) is element-wise multiplication and \\(\\mathbf{1}\\) is a binary mask filled with ones. \\(\\mathbf{M} \\in \\{0, 1\\}^{H \\times W}\\) denotes a random binary mask indicating where to drop out and fill in from two inversion noises. The noise mix ratio between two inversion" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "INSTASTYLE" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "noises is \\(\\alpha\\), representing the percent of \\(\\mathbf{M}\\) that is set to 1 (the rest is set to 0). Finally, we utilize a composed guidance mechanism [38] to estimate the noise:" + }, + { + "type": "equation", + "bbox": [ + 0.304, + 0.186, + 0.786, + 0.244 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {\\varepsilon} _ {\\theta} \\left(z _ {t}, t, \\mathcal {C}, \\varnothing\\right) = \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\varnothing\\right) \\\\ + w \\cdot (1 - \\beta) \\cdot \\left(\\varepsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C} _ {1}\\right) - \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\varnothing\\right)\\right) \\tag {12} \\\\ + w \\cdot \\beta \\cdot \\left(\\varepsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C} _ {2}\\right) - \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\varnothing\\right)\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.252, + 0.784, + 0.313 + ], + "angle": 0, + "content": "where \\(\\beta\\) is a prompt mix ratio. \\(\\mathcal{C}_1\\) and \\(\\mathcal{C}_2\\) are the embeddings of the target prompts for the two styles, respectively. Denote the target object as \\(<\\mathrm{obj}>\\), they can be formulated as \\(\\mathcal{C}_1 = \\psi(\"A < \\mathrm{obj}>\\) in the style of \\(<\\mathrm{style}1>\\)) and \\(\\mathcal{C}_2 = \\psi(\"A < \\mathrm{obj}>\\) in the style of \\(<\\mathrm{style}2>\\))." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.333, + 0.366, + 0.351 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.363, + 0.44, + 0.379 + ], + "angle": 0, + "content": "4.1 Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.385, + 0.788, + 0.507 + ], + "angle": 0, + "content": "We collect 60 images as the reference style dataset. The complete information on these images is listed in Tab. 2 in the Supplementary Material. Our method is implemented using PyTorch and executed on a single NVIDIA GeForce RTX 3090 GPU. The training procedure consists of 500 iterations, employing the Adam optimizer with a learning rate of \\(1 \\times 10^{-5}\\). During sampling, we set the guidance scale \\(w = 2.5\\). The number of generated images \\(M\\) in the first stage is 15 whose details are shown in Sec. C.1 in the Supplementary Material. The number of selected images \\(N\\) in the second stage is set to 5." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.531, + 0.287, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.583, + 0.284, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.633, + 0.283, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.681, + 0.283, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.732, + 0.276, + 0.742 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.532, + 0.395, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.583, + 0.394, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.633, + 0.394, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.681, + 0.396, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.732, + 0.396, + 0.742 + ], + "angle": 0, + "content": "(a)InstaStyle(Ours)" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.532, + 0.493, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.583, + 0.493, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.632, + 0.493, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.681, + 0.493, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.732, + 0.478, + 0.742 + ], + "angle": 0, + "content": "(b) StyleDrop" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.532, + 0.588, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.583, + 0.588, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.632, + 0.588, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.681, + 0.588, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.732, + 0.579, + 0.741 + ], + "angle": 0, + "content": "(c) DreamBooth" + }, + { + "type": "image", + "bbox": [ + 0.592, + 0.532, + 0.686, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.592, + 0.583, + 0.686, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.592, + 0.632, + 0.686, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.592, + 0.681, + 0.686, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.592, + 0.732, + 0.686, + 0.742 + ], + "angle": 0, + "content": "(d) Custom Diffusion" + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.532, + 0.784, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.583, + 0.784, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.632, + 0.784, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.681, + 0.784, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.732, + 0.785, + 0.742 + ], + "angle": 0, + "content": "(e) Textual Inversion" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.767, + 0.787, + 0.81 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparison of stylized image generation on various styles. Objects for synthesis are Bicycle, Sunflowers and Chair. Our method excels at capturing fine-grained style information, such as color, textures, and brushstrokes." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "X. Cui et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.288, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.2, + 0.287, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.25, + 0.287, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.304, + 0.287, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.354, + 0.287, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.408, + 0.278, + 0.417 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.146, + 0.371, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.187, + 0.363, + 0.195 + ], + "angle": 0, + "content": "Elephant" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.2, + 0.372, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.303, + 0.242, + 0.354, + 0.249 + ], + "angle": 0, + "content": "Castle" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.251, + 0.372, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.302, + 0.372, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.354, + 0.372, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.408, + 0.433, + 0.417 + ], + "angle": 0, + "content": "InstaStyle (Ours) AsePA-Net" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.146, + 0.442, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.199, + 0.441, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.251, + 0.441, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.302, + 0.441, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.354, + 0.441, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.38, + 0.408, + 0.433, + 0.417 + ], + "angle": 0, + "content": "AsePA-Net" + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.146, + 0.509, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.199, + 0.508, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.251, + 0.508, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.302, + 0.508, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.354, + 0.508, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.452, + 0.408, + 0.501, + 0.417 + ], + "angle": 0, + "content": "QuantArt" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.146, + 0.578, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.199, + 0.578, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.251, + 0.578, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.302, + 0.578, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.354, + 0.578, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.408, + 0.561, + 0.417 + ], + "angle": 0, + "content": "StyTr2" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.146, + 0.647, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.199, + 0.647, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.251, + 0.647, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.302, + 0.647, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.354, + 0.647, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.408, + 0.629, + 0.417 + ], + "angle": 0, + "content": "CCPL" + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.146, + 0.716, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.199, + 0.716, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.251, + 0.716, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.302, + 0.716, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.354, + 0.716, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.408, + 0.698, + 0.417 + ], + "angle": 0, + "content": "CAST" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.146, + 0.785, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.2, + 0.785, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.251, + 0.785, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.302, + 0.785, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.354, + 0.785, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.73, + 0.408, + 0.772, + 0.417 + ], + "angle": 0, + "content": "AdaAttn" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.441, + 0.788, + 0.497 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparison with style transfer methods. For style transfer methods, content images are employed (shown in the bottom right). Despite our method relying solely on text for content, we achieve comparable performance in the fidelity of the content. Furthermore, we excel in preserving style details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.529, + 0.467, + 0.545 + ], + "angle": 0, + "content": "4.2 Stylized Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.788, + 0.724 + ], + "angle": 0, + "content": "We conduct comparisons of INSTASTYLE against four recent methods following the open-sourced code\\(^{4,5}\\), i.e., StyleDrop [52], DreamBooth [48], Custom Diffusion [31], and Textual Inversion [10]. Besides, following [53], we also make a comparison with style transfer methods [6,19,21,39,66,72] to further illustrate the superiority of our approach. Notably, the style transfer task takes images as content input, in contrast to our task which only utilizes text prompts for content. To make a comparison between the two tasks, we additionally generate content images based on text prompts for style transfer methods. Although we can compare the two tasks under this setting, the comparison is unfair to our approach because the content images utilized by style transfer methods may provide additional content information compared to our text prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.725, + 0.788, + 0.801 + ], + "angle": 0, + "content": "Qualitative results. As shown in Fig. 4, textual Inversion and Custom Diffusion often yield an unsatisfactory style. Taking the well-known Van Gogh painting style (e.g., row 3 and row 4) as an example, they fail to capture fine-grained style information in the reference image. The fine-tuning-based methods, i.e., DreamBooth and StyleDrop, result in distorted content. For example, they gen" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.586, + 0.826 + ], + "angle": 0, + "content": "4 https://github.com/aim-uofa/StyleDrop-PyTorch" + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.827, + 0.544, + 0.84 + ], + "angle": 0, + "content": "\\(^{5}\\) https://github.com/huggingface/diffusers" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.811, + 0.586, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "INSTASTYLE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 1: (a) Quantitative comparison regarding the style consistency score and content fidelity score. (b) User study. The results show the percentage of votes where the comparison method is preferred over ours, ties with ours and is inferior to ours. We use the abbreviation (ST) to represent approaches for style transfer tasks." + }, + { + "type": "table_caption", + "bbox": [ + 0.351, + 0.212, + 0.371, + 0.224 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.235, + 0.487, + 0.375 + ], + "angle": 0, + "content": "
MethodStyle (↑)Content (↑)
(ST) AdaAttN [39]0.5870.287
(ST) CCPL [66]0.5420.286
(ST) StyTr2 [6]0.5600.282
(ST) CAST [72]0.5940.285
(ST) QuantArt [21]0.5540.289
(ST) AesPA-Net [19]0.6040.284
Textual Inversion [10]0.5620.278
Custom Diffusion [31]0.6110.280
DreamBooth [48]0.5610.244
StyleDrop [52]0.6020.248
Ours (INSTSTYLE)0.6550.294
" + }, + { + "type": "table_caption", + "bbox": [ + 0.633, + 0.212, + 0.654, + 0.224 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.235, + 0.769, + 0.357 + ], + "angle": 0, + "content": "
MethodPreferenceTieOurs
(ST) AdaAttN [39]0.260.130.61
(ST) CCPL [66]0.300.230.47
(ST) StyTr2 [6]0.270.200.53
(ST) CAST [72]0.310.260.43
(ST) QuantArt [21]0.120.230.65
(ST) AesPA-nNet [19]0.230.220.55
Textual Inversion [10]0.090.130.78
Custom Diffusion [31]0.090.150.76
DreamBooth [48]0.240.170.56
StyleDrop [52]0.270.260.47
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.788, + 0.592 + ], + "angle": 0, + "content": "erate bicycles with leaked content from the reference image in the last row. In contrast, benefiting from the style signal in the inversion noise, our INSTASTYLE can generate stylized images with fine-grained style details and higher fidelity. Take the bicycle in Fig. 4 as an example, our method can better preserve the color, textures, and brushstroke. Besides, the generated bicycle is more accurate. Fig. 5 presents the qualitative comparisons with style transfer methods. Despite the inherent challenge of utilizing text as content input in contrast to style transfer methods that use the content image as input, our INSTASTYLE demonstrates comparable performance in content fidelity. Additionally, our INSTASTYLE exhibits superior performance in preserving the style details of the reference image, notably in terms of color and brushstroke characteristics. More qualitative results are shown in Figs. 9 to 13 in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.672 + ], + "angle": 0, + "content": "Quantitative results. We follow evaluation measures in prior works [10,52]. We utilize 100 objects in CIFAR100 [30] as the target objects. The generated images are evaluated from two aspects. For style consistency, we calculate the CLIP score [44] between the generated image and the reference image. For content fidelity, CLIP score [44] is calculated between the image and the prompt." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As shown in Tab. 1 (a), INSTASTYLE achieves the highest style consistency and content fidelity, indicating that the images generated by INSTASTYLE exhibit a consistent style with the reference image while retaining its generation capability. For the comparison methods, Textual Inversion and Custom Diffusion have a lower style consistency, falling short of INSTASTYLE in style preservation. As for DreamBooth and StyleDrop, our approach surpasses them in generating target objects with a higher content score. We also compare with style transfer methods [6,19,21,39,66,72] which utilize content images as input. Notably, our approach diverges by only employing text prompts as content input. This brings challenges in terms of generating high-quality content as text prompts contain less information than content images. Despite this distinction, we achieve an" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "X. Cui et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "improved content score. Additionally, our approach excels in preserving style details, exhibiting a higher style consistency score." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.404 + ], + "angle": 0, + "content": "User study. Given the highly subjective nature of stylized generation and the inherent biases in content alignment and style consistency, we conduct a user study following previous works [6, 52]. Each question presents participants with a pair of images: one stylized image generated by our approach and another produced by a selected comparison method. To avoid bias, the images are anonymized and randomized for each question. Participants are tasked with evaluating which result they believed exhibited better stylization effects and content structures. To ensure robustness and reliability, we invite 20 participants for the user study. We compare our method with 10 existing methods, including the style transfer methods [6, 18, 21, 39, 66, 72] and current comparable methods [10, 31, 48, 52]. For each method, 30 generated image pairs are randomly selected. Finally, each participant completed 300 rounds of comparisons, resulting in a total of 6,000 votes across all methods. We count the votes and show the statistical results in Tab. 1 (b). Our approach obtains a higher preference, underscoring its superiority over competitors in both style preservation and content generation." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.431, + 0.275, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.431, + 0.416, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.431, + 0.539, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.431, + 0.662, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.431, + 0.785, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.631, + 0.788, + 0.701 + ], + "angle": 0, + "content": "Fig. 6: Visualization of the combination of two styles. In each case, we show our combination results (the biggest image). For better comparison, we also present the stylized generation results of style1 and style2 in the top left (blue box) and bottom left (red box), respectively. Objects for synthesis are Boat, Truck, Horse, Elephant, Pear, Lion, Plate, Pot, Helicopter, Bicycle, Palace, and Motorcycle." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.754, + 0.486, + 0.77 + ], + "angle": 0, + "content": "4.3 Combination of Two Styles" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We present style combination results in Fig. 6, where we set \\(\\alpha = 0.5\\) and \\(\\beta = 0.5\\) to illustrate a more obvious effect of style combination. Specifically, in each case, we also show the stylized generation results of style 1 (the top left) and style 2 (bottom left), respectively. The biggest image on the right is our combination" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "INSTASTYLE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.144, + 0.675, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.412, + 0.785, + 0.455 + ], + "angle": 0, + "content": "Fig. 7: Visualization of the combination of two styles. The style can be controlled via the noise mix ratio and the prompt mix ratio. Our approach enables continuous style combinations, demonstrating its flexibility and diversity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.785, + 0.514 + ], + "angle": 0, + "content": "results, showing the powerful style combination ability of our approach. More visualizations are shown in. Figs. 14 and 15 in Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.787, + 0.607 + ], + "angle": 0, + "content": "As introduced in Sec. 3.4, both the noise mix ratio \\(\\alpha\\) and the prompt mix ratio \\(\\beta\\) can affect the style. Fig. 7 illustrates the impact of the two parameters, where each row shows a different noise mix ratio (\\(\\alpha = 0.1, 0.3, 0.5, 0.7, 0.9\\)) and each column shows a different prompt mix ratio (\\(\\beta = 0.1, 0.3, 0.5, 0.7, 0.9\\)). The style is mainly influenced by the noise and the prompt further improves the style details, showing that our approach is flexible and can generate diverse results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.631, + 0.389, + 0.646 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Our approach achieves stylized generation from two perspectives. One is the \"style\" noise obtained via DDIM inversion, which is utilized as the initial noise during inference to provide fine-grained style information. The other is the style token, which is learned via prompt refinement tuning to describe the style precisely. In this section, we conduct the ablation study. The visualization results and quantitative results are shown in Fig. 8 (a) and Fig. 8 (b), respectively. Specifically, each point in Fig. 8 (b) presents the average style consistency score and content fidelity score of 100 generated images for a reference style." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Impact of inversion noise. A prominent advantage of our INSTASTYLE is that we utilize the inversion noise as the initial image during inference time to preserve the fine-grained style information. Thereby, we first conduct an ablation study by replacing the inversion \"style\" noise with random noise. As shown in" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "X. Cui et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.142, + 0.788, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.335, + 0.788, + 0.419 + ], + "angle": 0, + "content": "Fig. 8: Ablation study. (a) Visualization of ablation study. The reference style image is shown in the bottom right corner. Objects for synthesis are *Dinosaur*, *Bicycle*, *Bee*, and *Boy*, respectively. Without the inversion noise, the stylized performance is seriously degraded. Prompt refinement can further improve style details. (b) Quantitative results of ablation study. Our method lies further along the top right corner, showing better style preservation and content generation capability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.452, + 0.788, + 0.514 + ], + "angle": 0, + "content": "Fig. 8 (a) (second row), without the inversion noise, the stylized performance is seriously degraded. The quantitative results in Fig. 8 (b) further illustrate that ablating the \"style\" noise (blue dots) will result in a lower style consistency score, harming the style of the generated image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.621 + ], + "angle": 0, + "content": "Impact of prompt refinement tuning. We ablate the prompt refinement stage to show the necessity of learning a style token to avoid ambiguity and bias in human-writing textual style tokens. Fig. 8 (a) (third row) reveals that our learnable style token can better describe the style without harming the generation ability. The quantitative results in Fig. 8 (b) show that ablating the prompt refinement (green dots) will result in a lower content fidelity score and style score, further proving the importance of the learnable style token." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.36, + 0.667 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our research solves the task of stylized generation by exploring the fine-grained style information in the reference image. Through extensive analysis and empirical investigation, we find that the initial noise with a non-zero signal-to-noise ratio in the diffusion model leans toward a particular style. Therefore, DDIM inversion is introduced to obtain an initial noise containing fine-grained style information. Furthermore, we highlight the challenge caused by the internal bias in textual style tokens and propose to learn a style token adaptively to better describe the style without bias. Our approach produces satisfactory stylized generation results and supports the combination of two styles. Extensive qualitative and quantitative comparisons demonstrate the superiority of our method." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "INSTASTYLE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "Acknowledgement This research is sponsored by National Natural Science Foundation of China (Grant No. 62306041, U21B2045, 62176025), Beijing Nova Program (Grant No. Z211100002121106, 20230484488, 20230484276), Youth Innovation Promotion Association CAS (Grant No.2022132), and Beijing Municipal Science & Technology Commission (Z231100007423015)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.247, + 0.324, + 0.264 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.281, + 0.785, + 0.308 + ], + "angle": 0, + "content": "1. Alaluf, Y., Richardson, E., Metzer, G., Cohen-Or, D.: A neural space-time representation for text-to-image personalization. ACM TOG (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.337 + ], + "angle": 0, + "content": "2. Chen, H., Zhao, L., Wang, Z., Zhang, H., Zuo, Z., Li, A., Xing, W., Lu, D.: Dualast: Dual style-learning networks for artistic style transfer. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.785, + 0.365 + ], + "angle": 0, + "content": "3. Cho, J., Nam, G., Kim, S., Yang, H., Kwak, S.: Promptstyler: Prompt-driven style generation for source-free domain generalization. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.785, + 0.406 + ], + "angle": 0, + "content": "4. Cui, X., Li, P., Li, Z., Liu, X., Zou, Y., He, Z.: Localize, understand, collaborate: Semantic-aware dragging via intention reasoner. arXiv preprint arXiv:2406.00432 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.407, + 0.785, + 0.435 + ], + "angle": 0, + "content": "5. Cui, X., Li, Z., Li, P., Hu, Y., Shi, H., Cao, C., He, Z.: Chatedit: Towards multi-turn interactive facial image editing via dialogue. In: EMNLP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.436, + 0.785, + 0.463 + ], + "angle": 0, + "content": "6. Deng, Y., Tang, F., Dong, W., Ma, C., Pan, X., Wang, L., Xu, C.: Stytr2: Image style transfer with transformers. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.464, + 0.785, + 0.49 + ], + "angle": 0, + "content": "7. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.491, + 0.785, + 0.518 + ], + "angle": 0, + "content": "8. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.519, + 0.785, + 0.546 + ], + "angle": 0, + "content": "9. Everaert, M.N., Bocchio, M., Arpa, S., Susstrunk, S., Achanta, R.: Diffusion in style. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.547, + 0.785, + 0.588 + ], + "angle": 0, + "content": "0. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.589, + 0.785, + 0.63 + ], + "angle": 0, + "content": "1. Gal, R., Patashnik, O., Maron, H., Bermano, A.H., Chechik, G., Cohen-Or, D.: Stylegan-nada: Clip-guided domain adaptation of image generators. ACM TOG (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.631, + 0.785, + 0.658 + ], + "angle": 0, + "content": "2. Gatys, L.A., Ecker, A.S., Bethge, M.: Image style transfer using convolutional neural networks. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.785, + 0.686 + ], + "angle": 0, + "content": "3. Gatys, L.A., Ecker, A.S., Bethge, M., Hertzmann, A., Shechtman, E.: Controlling perceptual factors in neural style transfer. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.785, + 0.714 + ], + "angle": 0, + "content": "4. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. CACM (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.785, + 0.743 + ], + "angle": 0, + "content": "5. Gu, B., Fan, H., Zhang, L.: Two birds, one stone: A unified framework for joint learning of image and video style transfers. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "6. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.75, + 0.784 + ], + "angle": 0, + "content": "7. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPSW (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.813 + ], + "angle": 0, + "content": "8. Hong, K., Jeon, S., Lee, J., Ahn, N., Kim, K., Lee, P., Kim, D., Uh, Y., Byun, H.: Aespa-net: Aesthetic pattern-aware style transfer networks. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.841 + ], + "angle": 0, + "content": "9. Hong, K., Jeon, S., Lee, J., Ahn, N., Kim, K., Lee, P., Kim, D., Uh, Y., Byun, H.: Aespa-net: Aesthetic pattern-aware style transfer networks. In: ICCV (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.281, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "X. Cui et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "20. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "21. Huang, S., An, J., Wei, D., Luo, J., Pfister, H.: Quantart: Quantizing image style transfer towards high visual fidelity. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "22. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "23. Jia, G., Li, P., He, R.: Theme-aware aesthetic distribution prediction with full-resolution photographs. TNNLS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "24. Jing, Y., Liu, X., Ding, Y., Wang, X., Ding, E., Song, M., Wen, S.: Dynamic instance normalization for arbitrary style transfer. In: AAAI (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "25. Jing, Y., Liu, Y., Yang, Y., Feng, Z., Yu, Y., Tao, D., Song, M.: Stroke controllable fast style transfer with adaptive receptive fields. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "26. Jing, Y., Mao, Y., Yang, Y., Zhan, Y., Song, M., Wang, X., Tao, D.: Learning graph neural networks for image style transfer. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "27. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.396 + ], + "angle": 0, + "content": "28. Ke, Z., Liu, Y., Zhu, L., Zhao, N., Lau, R.W.: Neural preset for color style transfer. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "29. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "30. Krizhevsky, A., et al.: Learning multiple layers of features from tiny images. Technical report, University of Toronto (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "31. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "32. Li, P., Hu, Y., He, R., Sun, Z.: Global and local consistent wavelet-domain age synthesis. TIFS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "33. Li, P., Liu, X., Huang, J., Xia, D., Yang, J., Lu, Z.: Progressive generation of 3d point clouds with hierarchical consistency. PR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "34. Li, P., Wang, R., Huang, H., He, R., He, Z.: Pluralistic aging diffusion autoencoder. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "35. Li, P., Wu, X., Hu, Y., He, R., Sun, Z.: M2fpa: A multi-yaw multi-pitch high-quality dataset and benchmark for facial pose analysis. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "36. Lin, S., Liu, B., Li, J., Yang, X.: Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "37. Lin, T., Ma, Z., Li, F., He, D., Li, X., Ding, E., Wang, N., Li, J., Gao, X.: Drafting and revision: Laplacian pyramid network for fast high-quality artistic style transfer. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "38. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "39. Liu, S., Lin, T., He, D., Li, F., Wang, M., Li, X., Sun, Z., Li, Q., Ding, E.: Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.785, + 0.743 + ], + "angle": 0, + "content": "40. Mao, J., Wang, X., Aizawa, K.: Guided image synthesis via initial image editing in diffusion model. In: ACM MM (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "41. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "42. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "43. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: ICLR (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "INSTASTYLE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.217 + ], + "angle": 0, + "content": "45. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.786, + 0.244 + ], + "angle": 0, + "content": "46. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.272 + ], + "angle": 0, + "content": "47. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.313 + ], + "angle": 0, + "content": "48. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.356 + ], + "angle": 0, + "content": "49. Ruiz, N., Li, Y., Jampani, V., Wei, W., Hou, T., Pritch, Y., Wadhwa, N., Rubinstein, M., Aberman, K.: Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.383 + ], + "angle": 0, + "content": "50. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "51. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "52. Sohn, K., Ruiz, N., Lee, K., Chin, D.C., Blok, I., Chang, H., Barber, J., Jiang, L., Entis, G., Li, Y., et al.: Styledrop: Text-to-image generation in any style. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "53. Sohn, K., Shaw, A., Hao, Y., Zhang, H., Polania, L., Chang, H., Jiang, L., Essa, I.: Learning disentangled prompts for compositional image synthesis. arXiv preprint arXiv:2306.00763 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.508 + ], + "angle": 0, + "content": "54. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.549 + ], + "angle": 0, + "content": "55. Tang, H., Liu, S., Lin, T., Huang, S., Li, F., He, D., Wang, X.: Master: Meta style transformer for controllable zero-shot and few-shot artistic style transfer. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "56. Tang, J., Shu, X., Qi, G.J., Li, Z., Wang, M., Yan, S., Jain, R.: Tri-clustered tensor completion for social-aware image tag refinement. TPAMI (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.604 + ], + "angle": 0, + "content": "57. Teng, Q., Wang, R., Cui, X., Li, P., He, Z.: Exploring 3d-aware lifespan face aging via disentangled shape-texture representations. In: ICME (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "58. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.785, + 0.66 + ], + "angle": 0, + "content": "59. Wang, H., Li, Y., Wang, Y., Hu, H., Yang, M.H.: Collaborative distillation for ultra-resolution universal style transfer. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "60. Wang, R., Guo, H., Liu, J., Li, H., Zhao, H., Tang, X., Hu, Y., Tang, H., Li, P.: Stablegarment: Garment-centric generation via stable diffusion. arXiv preprint arXiv:2403.10783 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.785, + 0.743 + ], + "angle": 0, + "content": "61. Wang, R., Li, P., Huang, H., Cao, C., He, R., He, Z.: Learning-to-rank meets language: Boosting language-driven ordering alignment for ordinal classification. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "62. Wang, Z., Zhao, L., Xing, W.: Stylediffusion: Controllable disentangled style transfer via diffusion models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "63. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "64. Wen, L., Gao, C., Zou, C.: Cap-vstnet: Content affinity preserved versatile style transfer. In: CVPR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "X. Cui et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "65. Wu, X., Hu, Z., Sheng, L., Xu, D.: Styleformer: Real-time arbitrary style transfer via parametric style composition. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.203 + ], + "angle": 0, + "content": "66. Wu, Z., Zhu, Z., Du, J., Bai, X.: Ccpl: contrastive coherence preserving loss for versatile style transfer. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.231 + ], + "angle": 0, + "content": "67. Xie, X., Li, Y., Huang, H., Fu, H., Wang, W., Guo, Y.: Artistic style discovery with independent components. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.785, + 0.259 + ], + "angle": 0, + "content": "68. Xu, W., Long, C., Nie, Y.: Learning dynamic style kernels for artistic style transfer. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.286 + ], + "angle": 0, + "content": "69. Xu, Z., Sangineto, E., Sebe, N.: Stylerdalle: Language-guided style transfer using a vector-quantized tokenizer of a large-scale generative model. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.286, + 0.785, + 0.314 + ], + "angle": 0, + "content": "70. Yang, S., Hwang, H., Ye, J.C.: Zero-shot contrastive loss for text-guided diffusion image style transfer. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.314, + 0.785, + 0.342 + ], + "angle": 0, + "content": "71. Zhang, Y., Huang, N., Tang, F., Huang, H., Ma, C., Dong, W., Xu, C.: Inversion-based style transfer with diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.383 + ], + "angle": 0, + "content": "72. Zhang, Y., Tang, F., Dong, W., Huang, H., Ma, C., Lee, T.Y., Xu, C.: Domain enhanced arbitrary image style transfer via contrastive learning. In: ACM SIGGRAPH (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.785, + 0.411 + ], + "angle": 0, + "content": "73. Zhang, Z., Li, B., Nie, X., Han, C., Guo, T., Liu, L.: Towards consistent video editing with text-to-image diffusion models. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.452 + ], + "angle": 0, + "content": "74. Zhou, Y., Zhang, R., Sun, T., Xu, J.: Enhancing detail preservation for customized text-to-image generation: A regularization-free approach. arXiv preprint arXiv:2305.13579 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.452, + 0.785, + 0.481 + ], + "angle": 0, + "content": "75. Zhu, M., He, X., Wang, N., Wang, X., Gao, X.: All-to-key attention for arbitrary style transfer. In: ICCV (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.481 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/InstaStyle_ Inversion Noise of a Stylized Image is Secretly a Style Adviser/938fb53e-3bec-4f8a-bc7e-f5b36dd7b576_origin.pdf b/2024/InstaStyle_ Inversion Noise of a Stylized Image is Secretly a Style Adviser/938fb53e-3bec-4f8a-bc7e-f5b36dd7b576_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4daaa31099c823de63522cd667d641c121044400 --- /dev/null +++ b/2024/InstaStyle_ Inversion Noise of a Stylized Image is Secretly a Style Adviser/938fb53e-3bec-4f8a-bc7e-f5b36dd7b576_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f284794e06787ec7c9e9760beedc22ba923a15e1341161c470b6094bd797f8db +size 7832460 diff --git a/2024/InstaStyle_ Inversion Noise of a Stylized Image is Secretly a Style Adviser/full.md b/2024/InstaStyle_ Inversion Noise of a Stylized Image is Secretly a Style Adviser/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f65898235238fe696313075f59ce98e87f47510b --- /dev/null +++ b/2024/InstaStyle_ Inversion Noise of a Stylized Image is Secretly a Style Adviser/full.md @@ -0,0 +1,530 @@ +# INSTASTYLE: Inversion Noise of a Stylized Image is Secretly a Style Adviser + +Xing Cui $^{1}$ Zekun Li $^{2}$ Peipei Li $^{1\star}$ Huaibo Huang $^{3}$ Xuannan Liu $^{1}$ Zhaofeng He $^{1}$ + +1 Beijing University of Posts and Telecommunications + +2 University of California, Santa Barbara + +$^{3}$ MAIS & NLPR, Institute of Automation, Chinese Academy of Sciences + +{cuixing, lipeipei, liuxuannan, zhaofenghe}@bupt.edu.cn + +zekunli@cs.ucsb.edu huaibo.huang@cripac.ia.ac.cn + +https://cuixing100876.github.io/instastyle.github.io/ + +![](images/1e63eab505223bc53f6724f12ef348ed063d982a0d5eed9da5777533207015f9.jpg) +(a) Stylized image generation + +![](images/169ca768357d38194f2bb25e7bf82d15cb71c2e183bd28dbc4ac52b090739872.jpg) + +![](images/69cb3e298b998b86a6927958978f4559c255977f7d7a973f04cf3751e19c297d.jpg) + +![](images/84c629f341a3fd09e2ee3f6cd287e8f9f227a9693baefa5afcb66f43fc32c2af.jpg) + +![](images/b473543400ab34673d2adfbf995d8edb1c31fb93a1762a0942448bb5d951a3dc.jpg) +(b) Combination of two styles + +![](images/1c4d6b7cda07a19095eb7c0f005b26060b39492853b3c4481b232b0aab1effce.jpg) + +![](images/97529d597569d6ed5e4480cfb297ad0f2af082ea4440564d09051db94f66be3f.jpg) + +![](images/9330de21c23d5472d0ff22763d378e135a93505a0d1b824f021dedc8d5609c02.jpg) +Fig.1: Visualization of INSTASTYLE. (a) Our method excels at capturing style details and distinguishing between similar styles. (b) The first and third columns show images styled with reference to style 1 and style 2, respectively. The middle column shows images in a combined style. (c) Our method supports adjusting the degree of two styles during combination, dynamically ranging from one style to another. + +![](images/907bae5c8e0dc894846f9e6f17ff37fd2c55f8afe2aa130360b62c3694a522a6.jpg) +(c) Smooth style combination between two styles + +![](images/55ba7284fe8c72a89735b12aa9a149385830693aae246f036db71da7804d6512.jpg) + +![](images/267eaa01b35a34ab943f9ada63666182b2a808ac49138da21b5861ffa48e9bce.jpg) + +![](images/c07f653d646e4259d232a88b8f4c396b95f600fd4f1138d064b6859e612e65a8.jpg) + +![](images/8f12bad59ba600af0c00a833e3d7ab06fdbc0c2280cf101b86b5e6ab0fce7cc1.jpg) + +![](images/4bcbdf770e23e41f3ada368e2fd9598647f1c8198b5ab2243a224d827bbd285c.jpg) +Style 2 + +Abstract. Stylized text-to-image generation focuses on creating images from textual descriptions while adhering to a style specified by reference images. However, subtle style variations within different reference images can hinder the model from accurately learning the target style. In this paper, we propose INSTASTYLE, a novel approach that excels in generating high-fidelity stylized images with only a single reference image. Our approach is based on the finding that the inversion noise from a stylized reference image inherently carries the style signal, as evidenced by their non-zero signal-to-noise ratio. We employ DDIM inversion to extract this noise from the reference image and leverage a + +diffusion model to generate new stylized images from the "style" noise. Additionally, the inherent ambiguity and bias of textual prompts impede the precise conveying of style during image inversion. To address this, we devise prompt refinement, which learns a style token assisted by human feedback. Qualitative and quantitative experimental results demonstrate that INSTASTYLE achieves superior performance compared to current benchmarks. Furthermore, our approach also showcases its capability in the creative task of style combination with mixed inversion noise. + +Keywords: Stylized image generation $\cdot$ Inversion noise $\cdot$ Signal-to-noise ratio $\cdot$ Prompt refinement + +# 1 Introduction + +Recently, the advent of DDPMs [16, 51] has ushered in a new era of image generation. Existing text-to-image generation methods [5, 43, 45, 46] can generate images in coarse-grained styles by incorporating style descriptions in the textual prompt. However, the textual prompt is ambiguous which makes it hard to express style precisely [11]. Recently, personalized generation [10,48,63] has been introduced to generate novel concepts, which can be utilized for stylized generation by viewing the reference style as a new concept. These approaches typically necessitate mul + +tiple images as references. However, the subtle style variations in these multiple images present a challenge for the model to accurately learn and replicate the intended style. As shown in the first column of Fig. 1 (a), although the reference image Houses at Auvers (top) and Thatched Cottages by a Hill (bottom) are all Van Gogh's artworks, the differences in terms of color and texture [2] pose a significant challenge. Some recent approaches [3,53,69] are designed to more accurately describe the target style during generation. However, they may ignore the fine-grained style information in the reference images. + +In this paper, we propose INSTASTYLE (Inversion Noise of a Stylized Image is Secretly a Style Adviser) based on the observation that the inversion noise from a stylized reference image inherently contains the style signal. Specifically, our approach only requires a single reference image, which is transformed into a noise via DDIM inversion. The inversion noise, which preserves the style signal, is then used to generate stylized images by utilizing diffusion models. To shed light on this phenomenon, we demonstrate that the inversion noise maintains + +![](images/c6e36f47eb5ca9567d1b717d049362c833b3659b33beacdc9ab7cb5b91e60c4c.jpg) +Fig. 2: Motivation. Sampling from the inversion noise of a reference image can generate stylized images. However, the optimal style token varies for each case among human-written style tokens. Our learnable style token, i.e., "", shows greater universality across various scenarios. + +a non-zero signal-to-noise ratio in Sec. 3.2, indicating that it retains essential information (including the style signal) from the reference image. + +Nevertheless, directly using human-written prompts to describe styles for inverting the reference image often encounters challenges due to the inherent ambiguity of natural language. For instance, specific style descriptors like "ink painting" don't always work effectively with various objects. As shown in Fig. 2, the style "ink painting" might suit an object like a "chair" but not as well with a "boy". This inconsistency can arise because "ink painting" is typically associated with landscapes and might not yield optimal results for human subjects. Conversely, when using vague descriptors like "specific style", they may fail to provide enough information, leading to unpredictable generation quality (see the second row in Fig. 2). To address this, we propose Prompt Refinement to learn a style token with the assistance of human feedback. As illustrated in the last row of Fig. 2, the style information is boosted by using the learned style token in the noising process to obtain a better "style" noise. Therefore, our prompt refinement stands apart from previous methods [3, 53, 69], where style tokens are utilized in the denoising process to generate target images. During prompt refinement, we first collect images via human feedback. Then we optimize the embedding of the style token and the key and value projection matrices in the cross-attention layers of the diffusion model. + +As shown in Fig. 1 (a), our approach can effectively retain the fine-grained style in the reference image as well as generate new objects in high fidelity. Furthermore, our method supports the creative generation of style combination (Fig. 1 (b) and (c)) and allows adjusting the degree of two styles dynamically (Fig. 1 (c)). To sum up, our contributions are as follows: + +- We find that the inversion noise of a reference image via DDIM inversion potentially retains style information and evidence this observation by the non-zero signal-to-noise ratio of the inversion noise. +- We propose INSTASTYLE based on our observation, which leverages the style signal in inversion noise. Additionally, to better represent the style during inversion, a prompt refinement scheme is designed to learn the style token. Human feedback is also introduced to enhance model alignment with human preferences during prompt refinement. +- Qualitative and quantitative experimental results show the superiority of our INSTASTYLE in generating high-quality stylized images. Moreover, it exhibits promising potential in creative scenarios like style combinations. + +# 2 Related Work + +# 2.1 Text-to-image Synthesis + +Image synthesis is an essential subject in computer vision [7,23,41,56]. Previous works are based on variational auto-encoder [29, 33] or generative adversarial networks [5,14,32,35,57]. With the advance of pretrained vison-language models [44,61] and diffusion models [34,54,73], text-to-image generation [4,8,43,45, + +[60] has been widely studied and shown remarkable generalization ability. Recently, Lin et al. [36] have pointed out that the noise of the last step during noisering in the training process has a non-zero signal-to-noise ratio (SNR), i.e., there exists a signal leak. This results in a misalignment between the training and inference process. Therefore, some works [36, 50] propose to train diffusion models by enforcing the SNR to zero to avoid the signal leak. On the contrary, our approach leverages the leaked signal, which potentially includes the style details, from the reference image for the stylized generation. + +As these methods ignore the concepts that do not appear in the training set [10], some works [10,48,49] study personalized text-to-image generation which aims to adapt text-to-image models to new concepts given several reference images. For example, Textual Inversion [10] introduces and optimizes a word vector for each new concept. Subsequent works [1,63,74] further enhance the flexibility and adaptiveness of the learning strategy. Some methods [20, 48, 49] propose to finetune the original diffusion model, showing more satisfactory results. However, maintaining fine-grained style while simultaneously generating new objects remains a challenge for current personalized generation methods [52]. + +# 2.2 Stylized Image Generation + +Stylized image generation is a new paradigm for image generation which aims to generate content in a specific style given a few reference images. Although it is similar to the neural style transfer task [12, 13, 22, 27, 37, 59] which generates stylized images as well, they are fundamentally different. Style transfer solves an image translation task, focusing on translating a content image to target style [15, 21, 24, 28, 62, 64, 67, 68, 70]. For example, some works explore the global and local information to preserve the content of the source image [6, 18, 25, 26, 55, 65, 71, 75]. On the contrary, stylized image generation is geared towards generating a new image in specific style conditioned on a text. For example, ZDAIS [53] views stylized image generation as a domain adaptive task, where each style belongs to a domain. It learns disentangled prompts to adapt the model to new domains. Some approaches [9, 52] fine-tune the model to capture style properties. Our work differs from theirs by proposing a framework based on our observation that inversion noise contains style information. + +# 3 Method + +As shown in Fig. 3, our proposed method involves two main stages. In the first stage, we employ DDIM inversion to transform the reference image into noise. Notably, the inversion noise exhibits a non-zero signal-to-noise ratio, suggesting the presence of style signals from the reference image. Subsequently, we generate $M$ stylized images from the inversion noise conditioned on the given textual prompts. Due to the inherent ambiguity of the textual prompts that describe the style, it is challenging to precisely convey the desired style. Addressing this, + +![](images/eb3fe407cfbedfbac602427d450029f373f0a543a3e1a5fda5646613d6d87ca0.jpg) +Fig. 3: The training process of INSTASTYLE. (a) The first stage is initial stylized image generation. The reference image is inverted to noise conditioned on a prompt via DDIM Inversion. Then the inversion noise is utilized to generate initial stylized images. (b) The second stage is prompt refinement which leverages the selected high-quality initial stylized images to learn a style token. + +the second stage involves the incorporation of human feedback to select $N$ high-quality generated images from the first stage. The selected images are then used to learn a style token via prompt refinement. + +# 3.1 Preliminaries + +We begin by reviewing the fundamental diffusion model. Subsequently, we introduce Stable Diffusion [46], a key component of our framework. Finally, we discuss both DDIM and DDIM inversion. In our method, DDIM serves the purpose of denoising a noise to generate new images. Concurrently, DDIM inversion is utilized to transform the reference image into a corresponding "style" noise. + +Diffusion models. Diffusion models [16, 51] contain a forward process and a backward process. The forward process adds noise to the data according to a predetermined, non-learned variance schedule $\beta_{1},\ldots,\beta_{T}$ : + +$$ +q \left(z _ {t} \mid z _ {t - 1}\right) := \mathcal {N} \left(z _ {t}; \sqrt {1 - \beta_ {t}} z _ {t - 1}, \beta_ {t} \mathbf {I}\right). \tag {1} +$$ + +An important property of the forward process is that we can obtain the latent variable $z_{t}$ directly based on $z_{0}$ : + +$$ +z _ {t} = \sqrt {\alpha_ {t}} z _ {0} + \sqrt {1 - \alpha_ {t}} \varepsilon , \tag {2} +$$ + +where $\varepsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ , $\alpha_{t}\coloneqq \prod_{i = 1}^{t}(1 - \beta_{i})$ . Diffusion models restore the information by learning the backward process: + +$$ +p _ {\theta} \left(z _ {t - 1} \mid z _ {t}\right) := \mathcal {N} \left(z _ {t - 1}; \mu_ {\theta} \left(z _ {t}, t\right), \Sigma_ {\theta} \left(z _ {t}, t\right)\right). \tag {3} +$$ + +Stable Diffusion. In Stable Diffusion, the diffusion process happens in latent space [46]. It utilizes a pretrained autoencoder which consists of an encoder and a decoder. The encoder $\mathcal{E}(\cdot)$ maps an image $x$ into a latent code $z$ and the decoder $\mathcal{D}(\cdot)$ transforms the latent code back into an image. The denoising model $\varepsilon_{\theta}(\cdot)$ + +is an Unet [47]. During training, artificial noise is added to the sampled data $z_0$ according to timestamp $t$ based on Eq. (2), resulting in a noised sample $z_t$ . Denoising model starts with $z_t$ and is trained to predict the artificial noise: + +$$ +\min _ {\theta} E _ {z _ {0}, \varepsilon \sim N (0, I), t \sim \operatorname {U n i f o r m} (1, T)} \| \varepsilon - \varepsilon_ {\theta} \left(z _ {t}, t, \mathcal {C}\right) \| _ {2} ^ {2}, \tag {4} +$$ + +where $\mathcal{C} = \psi (\mathcal{P})$ is the embedding of text condition $\mathcal{P}$ + +DDIM. In inference time, given a noise vector $z_{T}$ , the noise is gradually removed by sequentially predicting the added noise for $T$ steps. DDIM [54] is one of the denoising approaches with a deterministic process: + +$$ +z _ {t - 1} = \sqrt {\frac {\alpha_ {t - 1}}{\alpha_ {t}}} z _ {t} + \left(\sqrt {\frac {1}{\alpha_ {t - 1}} - 1} - \sqrt {\frac {1}{\alpha_ {t}} - 1}\right) \cdot \tilde {\varepsilon} _ {\theta}, \qquad (5) +$$ + +where $\tilde{\varepsilon}_{\theta}$ is the estimated noise. + +DDIM inversion [42, 54] transforms an image to a noise conditioned on a prompt. The diffusion process is performed in the reverse direction, i.e., $z_0 \to z_T$ : + +$$ +z _ {t + 1} = \sqrt {\frac {\alpha_ {t + 1}}{\alpha_ {t}}} z _ {t} + \left(\sqrt {\frac {1}{\alpha_ {t + 1}} - 1} - \sqrt {\frac {1}{\alpha_ {t}} - 1}\right) \cdot \tilde {\varepsilon} _ {\theta}. \tag {6} +$$ + +# 3.2 Initial Stylized Image Generation + +As shown in Fig. 3 (a), the first stage of our INSTASTYLE involves obtaining the inversion noise via DDIM inversion and sampling images via DDIM. + +In DDIM inversion, the added noise for each step is calculated conditioned on a prompt and gradually incorporated into the reference image following Eq. (6). Specifically, we set the learnable style token as a human-written style description in this stage. We demonstrate that the inversion noise from a stylized reference image inherently carries the style signal from the perspective of the signal-to-noise ratio. As the estimated noise $\varepsilon_{\theta}(z_t,t,\mathcal{C})$ is trained to approximate the artificial noise $\varepsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ , we assume that $\varepsilon_{\theta}(z_t,t,\mathcal{C}) \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ . According to Eq. (6), $z_{t + 1}$ can be approximated in a closed form: + +$$ +z _ {t + 1} = \sqrt {\frac {\alpha_ {t + 1}}{\alpha_ {0}}} z _ {0} + \sqrt {\sum_ {i = 0} ^ {t} \frac {\alpha_ {t + 1}}{\alpha_ {i + 1}} \left(\sqrt {\frac {1}{\alpha_ {i + 1}} - 1} - \sqrt {\frac {1}{\alpha_ {i}} - 1}\right) ^ {2}} \cdot \bar {\varepsilon} _ {0}, \tag {7} +$$ + +where $\bar{\varepsilon}_0\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ + +Signal-to-Noise ratio. Signal-to-noise ratio (SNR) is introduced to measure the ratio of signals from the original image preserved in the noise [36,46]. The SNR of the inversion noise can be calculated as: + +$$ +\operatorname {S N R} (t) := \frac {1}{\sum_ {i = 0} ^ {t} \frac {\alpha_ {0}}{\alpha_ {i + 1}} \left(\sqrt {\frac {1}{\alpha_ {i + 1}} - 1} - \sqrt {\frac {1}{\alpha_ {i}} - 1}\right) ^ {2}}. \tag {8} +$$ + +We present detailed derivations in Sec. B in Supplementary Material. In Stable Diffusion [46], the predetermined variance schedule $\beta_{t} = (\sqrt{0.00085}\cdot (1 - j) + \sqrt{0.012}\cdot j)^{2}$ , where $j = \frac{t - 1}{T - 1}$ . At the terminal timestep $T = 1000$ , the + +$SNR(T) = 0.015144$ , i.e., $z_{T}$ has a none-zero signal-to-noise ratio. This nonzero signal-to-noise ratio suggests that the noise obtained via DDIM inversion still retains information from the reference image and deviates from white noise. Besides, we conduct qualitative experiments and observe that style information can consistently be preserved at the terminal timestep (refer to Sec. C.3 in Supplementary Material). Therefore, we can generate target stylized images by sampling from the inversion noise based on Eq. (5). + +# 3.3 Prompt Refinement + +As the natural language prompts may not precisely convey the style (Fig. 2), we propose Prompt Refinement to learn the style tokens as shown in Fig. 3 (b). We utilize the generated stylized images in the first stage to constitute the training data for prompt refinement learning. While these generated images may not always be precise, we observe that most of them successfully retain the style information alongside appropriate content. We manually select images that distinctly embody the reference style and the target object to build the dataset. + +Specifically, we introduce a new token, i.e., “” to represent the style descriptor and learn its embedding. In practice, we initialize the new token with the embeddings of the textual style descriptor of the reference image. The text condition is input to the model via two projection matrices in the cross-attention block of diffusion model $\varepsilon_{\theta}$ , i.e., $W^{k} \in \mathbb{R}^{d \times d'}$ and $W^{v} \in \mathbb{R}^{d \times d'}$ . Therefore, we also fine-tune these two projection matrices. The text feature $\mathbf{c} \in \mathbb{R}^{s \times d}$ is projected to key $K = \mathbf{c}W^{k}$ and value $V = \mathbf{c}W^{v}$ . The query matrix $W^{q} \in \mathbb{R}^{l \times d'}$ projects the latent image feature $\mathbf{f} \in \mathbb{R}^{(h \times w) \times l}$ into query feature $Q = \mathbf{f}W^{q}$ . Then the cross-attention [58] is calculated as: + +$$ +\operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {S o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d ^ {\prime}}}\right) V. \tag {9} +$$ + +We utilize the LoRA [20] for model fine-tuning. Specifically, for a projection matrix $W \in \mathbb{R}^{d \times d'}$ (i.e., $W = W^k$ or $W = W^v$ ) to be fine-tuned, we update a low-rank residual rather than directly fine-tuning $W$ . Formally, we denote the fine-tuned projection matrix as $W' = W + BA$ , where $B \in \mathbb{R}^{d \times r}$ , $A \in \mathbb{R}^{r \times d'}$ , and the rank $r \ll \min(d, d')$ . During training, only $A$ and $B$ are learnable. For $y = Wx$ , the modified forward pass is $y = Wx + BAx$ . Putting the two stages together, our full algorithm is shown in Algorithm 1. + +# 3.4 Inference + +Stylized image generation. For stylized image generation, we first learn the style token "" to describe the style in the reference image. During inference, the learned style token is integrated into the prompt that inverts image and the prompt that generates the target image. Besides, the diffusion model $\varepsilon_{\theta}$ is utilized as the backbone of DDIM inversion and DDIM. Specifically, the reference image is first inverted to a noise $z_{T}$ via DDIM inversion conditioned + +1. Input: A source prompt embedding $\mathcal{C} = \psi(\mathcal{P})$ and a reference image $\mathcal{I}$ . Initial style token embedding $v$ is included in the prompt embedding $\mathcal{C}$ . +2 Output: Optimized style token embedding $v^{*}$ and diffusion model $\varepsilon_{\theta}^{*}$ . + +// Initial Stylized Image Generation + +3 Compute the inversion noise $z_{T}$ using DDIM inversion over $\mathcal{I}$ conditioned on $\mathcal{C}$ based on Eq. (6); +4 Generate $M$ images conditioned on a predefined prompt set based on Eq. (5); // Prompt Refinement +5 Select $N$ images from $M$ images with human feedback as the training dataset $q(\mathbf{x}_0, c)$ ; + +6 repeat + +Algorithm 1: INSTASTYLE +```txt +7 $\mathbf{x}_0,c\sim q(\mathbf{x}_0,c)$ +8 $z_{0} = \mathcal{E}(\mathbf{x}_{0})$ +9 $t\sim \mathrm{Uniform}(\{1,\dots ,T\})$ +10 $\varepsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ +11 $z_{t} = \sqrt{\alpha_{t}} z_{0} + \sqrt{1 - \alpha_{t}}\varepsilon$ +12 Take gradient step on $\nabla_{\theta ,v}\| \varepsilon -\varepsilon_{\theta}(z_t,t,c)\| _2^2;$ +``` + +13 until converged; +14 Return $v^{*}$ , $\varepsilon_{\theta}^{*}$ + +on a prompt containing the learned style token. Then we sample from the inversion noise conditioned on a target prompt which contains the target content and the learned style token via DDIM. Specifically, the estimated noise is calculated based on the classifier-free guidance [17]: + +$$ +\tilde {\varepsilon} _ {\theta} \left(z _ {t}, t, \mathcal {C}, \varnothing\right) = \varepsilon_ {\theta} \left(z _ {t}, t, \varnothing\right) + w \cdot \left(\varepsilon_ {\theta} \left(z _ {t}, t, \mathcal {C}\right) - \varepsilon_ {\theta} \left(z _ {t}, t, \varnothing\right)\right), \tag {10} +$$ + +where $\varnothing = \psi(")$ is the embedding of a null text. $\varepsilon_{\theta}(z_t, t, \mathcal{C})$ represents the conditional predictions. $w$ is the guidance scale parameter. + +Combination of two styles. To combine two styles, we learn style tokens for each style, i.e., “” and “”. Specifically, we use the selected images of both styles to constitute the training set. Then we jointly optimize the embeddings of the style tokens and fine-tune the key projection and value projection in the cross-attention block. During inference, we individually transform two reference images into their corresponding inversion noise $z_{t1} \in \mathbb{R}^{H \times W \times C}$ and $z_{t2} \in \mathbb{R}^{H \times W \times C}$ via DDIM inversion conditioned on a prompt containing their learned style token. As the style can be described by both the “style” noise and style token, we combine the styles from these two perspectives. For the “style” noise, the combined inversion noise $z_{t}$ is obtained based on a masking strategy: + +$$ +z _ {t} = (\mathbf {1} - \mathbf {M}) \odot z _ {t 1} + \mathbf {M} \odot z _ {t 2}, \tag {11} +$$ + +where $\odot$ is element-wise multiplication and $\mathbf{1}$ is a binary mask filled with ones. $\mathbf{M} \in \{0, 1\}^{H \times W}$ denotes a random binary mask indicating where to drop out and fill in from two inversion noises. The noise mix ratio between two inversion + +noises is $\alpha$ , representing the percent of $\mathbf{M}$ that is set to 1 (the rest is set to 0). Finally, we utilize a composed guidance mechanism [38] to estimate the noise: + +$$ +\begin{array}{l} \tilde {\varepsilon} _ {\theta} \left(z _ {t}, t, \mathcal {C}, \varnothing\right) = \varepsilon_ {\theta} \left(z _ {t}, t, \varnothing\right) \\ + w \cdot (1 - \beta) \cdot \left(\varepsilon_ {\theta} \left(z _ {t}, t, \mathcal {C} _ {1}\right) - \varepsilon_ {\theta} \left(z _ {t}, t, \varnothing\right)\right) \tag {12} \\ + w \cdot \beta \cdot \left(\varepsilon_ {\theta} \left(z _ {t}, t, \mathcal {C} _ {2}\right) - \varepsilon_ {\theta} \left(z _ {t}, t, \varnothing\right)\right), \\ \end{array} +$$ + +where $\beta$ is a prompt mix ratio. $\mathcal{C}_1$ and $\mathcal{C}_2$ are the embeddings of the target prompts for the two styles, respectively. Denote the target object as $<\mathrm{obj}>$ , they can be formulated as $\mathcal{C}_1 = \psi("A < \mathrm{obj}>$ in the style of $<\mathrm{style}1>$ ) and $\mathcal{C}_2 = \psi("A < \mathrm{obj}>$ in the style of $<\mathrm{style}2>$ ). + +# 4 Experiment + +# 4.1 Experimental Setting + +We collect 60 images as the reference style dataset. The complete information on these images is listed in Tab. 2 in the Supplementary Material. Our method is implemented using PyTorch and executed on a single NVIDIA GeForce RTX 3090 GPU. The training procedure consists of 500 iterations, employing the Adam optimizer with a learning rate of $1 \times 10^{-5}$ . During sampling, we set the guidance scale $w = 2.5$ . The number of generated images $M$ in the first stage is 15 whose details are shown in Sec. C.1 in the Supplementary Material. The number of selected images $N$ in the second stage is set to 5. + +![](images/c80cd58cfac2a6c96bc4d5b69f6c86d483da87d9c4e1d3427d3057bcb40f2567.jpg) + +![](images/62bb4229f8af2972e988e022ba0d231ef3ad458bc58eb5fedee8e8358b7441be.jpg) + +![](images/bd86b46287db8e57f5d00d11245aa8a8a3df65f210ac82d2bf536f74c6df529a.jpg) + +![](images/f76010f551154aba616b502ebb7632c7ed8a8dc21e847c25fe407ec728079064.jpg) +Reference + +![](images/7da6a41e381c596b65db7c7f3decaf9953cc2a7dd0161b4941397773a727d28a.jpg) + +![](images/ad889a0748d7d4b91bc4e81feab077c629a9fff69edfcd9d34d3c23ae55d6273.jpg) + +![](images/ec16abb08643cb1c6a287e592936c53702e6390106467c49c48455a9c122504f.jpg) + +![](images/a4f917a9ca3a35ce6227ed953403d9a6a49b24239d416c7cadfd7b3ae105b922.jpg) +(a)InstaStyle(Ours) + +![](images/820c13bc6a331fdd56672ae44865209d89456ea7e66cc8269458d290bf5066b3.jpg) + +![](images/f4ea39848124db46872acc74bb26bef845260cfadd46797d37f9938aa2eb0ba4.jpg) + +![](images/f30cfcc739e5c8d127bd62253d7bf6fd9854cea2b82a4bef23f836be27277f2a.jpg) + +![](images/55ceaa9fe7cd17f7c80fc91b8bafb11df3a9528725f082866ec54ca2184e4fcc.jpg) +(b) StyleDrop + +![](images/a44d63fc11adf0886ed91375a8e74f7fae3953358ce775aa66cefe1ac9d7695e.jpg) + +![](images/4674897d31eeb69fbacd88345437a172ba213fbd74048934bf294f05392867fb.jpg) + +![](images/0107c43360cc799d70b54fe3f4e7e4e8b60c9f736ad4ba37ad6fb88f2ea1bc34.jpg) + +![](images/cedd09f2c56d58e9316e4c472b9610fbbd96ec4b0292eeb9a36619ac9f8bb852.jpg) +(c) DreamBooth +Fig. 4: Qualitative comparison of stylized image generation on various styles. Objects for synthesis are Bicycle, Sunflowers and Chair. Our method excels at capturing fine-grained style information, such as color, textures, and brushstrokes. + +![](images/002147886c24b0ab81221df2673671661e13f714c9273a98d253f706318af2ce.jpg) + +![](images/653565d4528b7c0528cf9491e47b6acdccc83b98af0dc302b7a61fafab4ca3fe.jpg) + +![](images/6983507207ecd7191875b40ee1de1704c34080d70e3b1344588e9ce3ba303710.jpg) + +![](images/83b87eac91516065c091178a75e7d91cba991330dfc8b756dcd910b4ccddb590.jpg) +(d) Custom Diffusion + +![](images/40a69f8153fe9d2dbf1fbe2c7145211952e9bf6c2a33c76174eb41e0829690d7.jpg) + +![](images/88421700aef09d80582a937676a89610ad3f6f07859c37df56059aa7ebbfea5e.jpg) + +![](images/79e448d2d647579d093015077ce84c9df2da00dec33a313e6c9974613d4136ee.jpg) + +![](images/d5b43c83b2a50bb2f1d961f2fb79cdcb39838166256d1b4b070d8709d0b2ca67.jpg) +(e) Textual Inversion + +![](images/93a947d27a8f62ea37793f74e43126f818fb863980e76daf4c312e7d6151b1a3.jpg) + +![](images/e723e52182e240f95fbbe1b7a875a9b19283d0e8e74851d803ff4d831ec3d572.jpg) + +![](images/49116d4c68b22b28a783b598b4d337418063adbb1b2519c0a2bf9d78da5f3f88.jpg) + +![](images/6c1a1cdaf0f6f75b8d8923bfa18fd5b89f2d51c76d7d7dc52e08972b08ca6aa6.jpg) + +![](images/a8d3e58b66bc7a859d63d2473cbdf9c7094fcc83fa576d996983b959fa348831.jpg) +Reference + +![](images/6308999419633e125169ded48e8b7eff36ed76541bc56a75a94058289b96d5f0.jpg) +Elephant + +![](images/edc3f3da5f4238def5eac7a6ed55fdab14f7825f302830fd0521a3e20a4fd899.jpg) +Castle + +![](images/dd8c65a113b26c9d1a7ab6c21cd6573ca34f853337ef4a1d7a5000375691bd2a.jpg) + +![](images/de7afc33523f9555c9b8ce501cbfdd3109d3fbc90863b3439ab6346959c63a42.jpg) + +![](images/d7c07f514c16acc211d896020f09eae1047f00ab72e06c84dd7ccecb71e2a852.jpg) +InstaStyle (Ours) AsePA-Net + +![](images/1f685ed8d931e04fcb720f28c5e00c1b0b51feb446af776d043f685b254cc413.jpg) + +![](images/330ec054cd93cdd9ed17ca7eb53fd48bddd2eef517b06bc473b1e77534e8c038.jpg) + +![](images/ef77012026b7637a45a696c5a90362734f2fbf4bd85a431bdbf604570b051001.jpg) + +![](images/206cfea526225d6930e12f88a1c78d4df1041164ce2bd67524b1ba578a9e2a94.jpg) + +![](images/d10c581de5238462787cc8f7fff64353771ac95d9828003356f85a3eb8091a98.jpg) + +AsePA-Net + +![](images/73463e740a620daeeb15fcfe9c403a856ba19fd6a67bb97c6da8d3d443e15c3e.jpg) + +![](images/95bfccb1ea87ba82468c96b26daec57cc902ad2b653de2bcfeef06fd6fa3ecd4.jpg) + +![](images/35d60c178be4da672a7f060d94fa71e8479354998dd926f364096d7bbe1c0838.jpg) + +![](images/cbfcc14075106e7f94a26eb32d48e078074cd840d83d012f91adf7a93a083600.jpg) + +![](images/d7180e1caddb0d3a86177ba9e3e5d782cb58390549923bd61eaf4ac902533446.jpg) +QuantArt + +![](images/05598e7175a982d5487822da50fa43e462e996653c5f4523c0726f50f9d32d62.jpg) + +![](images/0c9d46d1c3d0bcf4267c7bd4f59bbb1cc89b6f4ff181b89666d1784bd8fc38dc.jpg) + +![](images/e264bcef2f9453325657db6676fb1edfdc9e33fc03bf384f5a8ea86c94ee7706.jpg) + +![](images/1476c31bfcbc733e327d103cc8148e8bbc300b1968115dddab1c68d7bf80e6f0.jpg) + +![](images/4b93159d2c873695fdb64eed844b3e72c22c4070dca0151e1897340a52c97462.jpg) +StyTr2 +Fig. 5: Qualitative comparison with style transfer methods. For style transfer methods, content images are employed (shown in the bottom right). Despite our method relying solely on text for content, we achieve comparable performance in the fidelity of the content. Furthermore, we excel in preserving style details. + +![](images/e41ca3e62ebe14a9808eafe31303259a8263db9d06de58c81dded567f3601f4f.jpg) + +![](images/34009c2c215c55995462b4dc78ce439822f400a8f47f02ecc3754b8173f441ae.jpg) + +![](images/1fd7c27b9d3cedce90590f2b1b794c1f6ebdb3bb91f09bbe58e630c76a928466.jpg) + +![](images/8fdcbdce0135afb31df927f694a23fd3910a085c92de7df6dc37827d04698ab6.jpg) + +![](images/3487069969ff08a784ccbb60e4c46f118026d63ab2069534411f7daec87421b4.jpg) +CCPL + +![](images/987a6b0916c64ccabaa92e53be4572161cd6d477426dc4baa3004fb4e22a8d70.jpg) + +![](images/045f44cf9b34091bb737841c03f6264b59476797cef469c95c02a9fd0615060d.jpg) + +![](images/02fba50679b86a7063d59cd655d216a36378a742a971c229f1341e5ac30046a6.jpg) + +![](images/48e2922a4116385d9a72fd20ddb187d0dce8e687c4feb90b5859d93d60c78ee6.jpg) + +![](images/bfb1dcd8741518b7e4e125fd298e9295f0816ec4af33087a11432480483cff82.jpg) +CAST + +![](images/b45d1cc9e7ae0b014b3310e82ec8e6e580f41351184793f9912aa37b1e52385c.jpg) + +![](images/b0c9100974743ffa73e09a8c2b364963320cdada2537265b94bb012071077003.jpg) + +![](images/c88c1c8132436e21e819713e6a68af7e3b7ca7f87a3da6c7d55f384e05abaa90.jpg) + +![](images/026dc34338163e68aa922611d253adf77f91015dfa3e6afaaf49b935b3859adb.jpg) + +![](images/cc3dfe80fe84d7125c55048fa26b61828e98d962b7edeca00acce1a79438501d.jpg) +AdaAttn + +# 4.2 Stylized Image Synthesis + +We conduct comparisons of INSTASTYLE against four recent methods following the open-sourced code $^{4,5}$ , i.e., StyleDrop [52], DreamBooth [48], Custom Diffusion [31], and Textual Inversion [10]. Besides, following [53], we also make a comparison with style transfer methods [6,19,21,39,66,72] to further illustrate the superiority of our approach. Notably, the style transfer task takes images as content input, in contrast to our task which only utilizes text prompts for content. To make a comparison between the two tasks, we additionally generate content images based on text prompts for style transfer methods. Although we can compare the two tasks under this setting, the comparison is unfair to our approach because the content images utilized by style transfer methods may provide additional content information compared to our text prompts. + +Qualitative results. As shown in Fig. 4, textual Inversion and Custom Diffusion often yield an unsatisfactory style. Taking the well-known Van Gogh painting style (e.g., row 3 and row 4) as an example, they fail to capture fine-grained style information in the reference image. The fine-tuning-based methods, i.e., DreamBooth and StyleDrop, result in distorted content. For example, they gen + +Table 1: (a) Quantitative comparison regarding the style consistency score and content fidelity score. (b) User study. The results show the percentage of votes where the comparison method is preferred over ours, ties with ours and is inferior to ours. We use the abbreviation (ST) to represent approaches for style transfer tasks. +(a) + +
MethodStyle (↑)Content (↑)
(ST) AdaAttN [39]0.5870.287
(ST) CCPL [66]0.5420.286
(ST) StyTr2 [6]0.5600.282
(ST) CAST [72]0.5940.285
(ST) QuantArt [21]0.5540.289
(ST) AesPA-Net [19]0.6040.284
Textual Inversion [10]0.5620.278
Custom Diffusion [31]0.6110.280
DreamBooth [48]0.5610.244
StyleDrop [52]0.6020.248
Ours (INSTSTYLE)0.6550.294
+ +(b) + +
MethodPreferenceTieOurs
(ST) AdaAttN [39]0.260.130.61
(ST) CCPL [66]0.300.230.47
(ST) StyTr2 [6]0.270.200.53
(ST) CAST [72]0.310.260.43
(ST) QuantArt [21]0.120.230.65
(ST) AesPA-nNet [19]0.230.220.55
Textual Inversion [10]0.090.130.78
Custom Diffusion [31]0.090.150.76
DreamBooth [48]0.240.170.56
StyleDrop [52]0.270.260.47
+ +erate bicycles with leaked content from the reference image in the last row. In contrast, benefiting from the style signal in the inversion noise, our INSTASTYLE can generate stylized images with fine-grained style details and higher fidelity. Take the bicycle in Fig. 4 as an example, our method can better preserve the color, textures, and brushstroke. Besides, the generated bicycle is more accurate. Fig. 5 presents the qualitative comparisons with style transfer methods. Despite the inherent challenge of utilizing text as content input in contrast to style transfer methods that use the content image as input, our INSTASTYLE demonstrates comparable performance in content fidelity. Additionally, our INSTASTYLE exhibits superior performance in preserving the style details of the reference image, notably in terms of color and brushstroke characteristics. More qualitative results are shown in Figs. 9 to 13 in the Supplementary Material. + +Quantitative results. We follow evaluation measures in prior works [10,52]. We utilize 100 objects in CIFAR100 [30] as the target objects. The generated images are evaluated from two aspects. For style consistency, we calculate the CLIP score [44] between the generated image and the reference image. For content fidelity, CLIP score [44] is calculated between the image and the prompt. + +As shown in Tab. 1 (a), INSTASTYLE achieves the highest style consistency and content fidelity, indicating that the images generated by INSTASTYLE exhibit a consistent style with the reference image while retaining its generation capability. For the comparison methods, Textual Inversion and Custom Diffusion have a lower style consistency, falling short of INSTASTYLE in style preservation. As for DreamBooth and StyleDrop, our approach surpasses them in generating target objects with a higher content score. We also compare with style transfer methods [6,19,21,39,66,72] which utilize content images as input. Notably, our approach diverges by only employing text prompts as content input. This brings challenges in terms of generating high-quality content as text prompts contain less information than content images. Despite this distinction, we achieve an + +improved content score. Additionally, our approach excels in preserving style details, exhibiting a higher style consistency score. + +User study. Given the highly subjective nature of stylized generation and the inherent biases in content alignment and style consistency, we conduct a user study following previous works [6, 52]. Each question presents participants with a pair of images: one stylized image generated by our approach and another produced by a selected comparison method. To avoid bias, the images are anonymized and randomized for each question. Participants are tasked with evaluating which result they believed exhibited better stylization effects and content structures. To ensure robustness and reliability, we invite 20 participants for the user study. We compare our method with 10 existing methods, including the style transfer methods [6, 18, 21, 39, 66, 72] and current comparable methods [10, 31, 48, 52]. For each method, 30 generated image pairs are randomly selected. Finally, each participant completed 300 rounds of comparisons, resulting in a total of 6,000 votes across all methods. We count the votes and show the statistical results in Tab. 1 (b). Our approach obtains a higher preference, underscoring its superiority over competitors in both style preservation and content generation. + +![](images/ddc98eed017386d9aacf4897124584d45d4e45043821f6481d8d2ab6a5cd0cd3.jpg) +Fig. 6: Visualization of the combination of two styles. In each case, we show our combination results (the biggest image). For better comparison, we also present the stylized generation results of style1 and style2 in the top left (blue box) and bottom left (red box), respectively. Objects for synthesis are Boat, Truck, Horse, Elephant, Pear, Lion, Plate, Pot, Helicopter, Bicycle, Palace, and Motorcycle. + +![](images/08c824084edf5b72d9e5b220c3f1e0e1226b1893a8afded2055c0d8853171490.jpg) + +![](images/f8d2d26bed1296db83e55389e0b19a7c9fcceefb533625c708b4b4e8069954c2.jpg) + +![](images/82e18b05baf7c9fa1db32c84e3cca3ae22b7db7774eaf5f6ed4e8a03996ea6cb.jpg) + +![](images/9fbc5de0087b186f10f8abda4b9334c6691931f27bdd6402aaa9477563c972a1.jpg) + +# 4.3 Combination of Two Styles + +We present style combination results in Fig. 6, where we set $\alpha = 0.5$ and $\beta = 0.5$ to illustrate a more obvious effect of style combination. Specifically, in each case, we also show the stylized generation results of style 1 (the top left) and style 2 (bottom left), respectively. The biggest image on the right is our combination + +![](images/90e0317c5d4ede5d33bd0924b11ddf8f1f3121ebb4d02ac49c021304b98895c4.jpg) +Fig. 7: Visualization of the combination of two styles. The style can be controlled via the noise mix ratio and the prompt mix ratio. Our approach enables continuous style combinations, demonstrating its flexibility and diversity. + +results, showing the powerful style combination ability of our approach. More visualizations are shown in. Figs. 14 and 15 in Supplementary Material. + +As introduced in Sec. 3.4, both the noise mix ratio $\alpha$ and the prompt mix ratio $\beta$ can affect the style. Fig. 7 illustrates the impact of the two parameters, where each row shows a different noise mix ratio ( $\alpha = 0.1, 0.3, 0.5, 0.7, 0.9$ ) and each column shows a different prompt mix ratio ( $\beta = 0.1, 0.3, 0.5, 0.7, 0.9$ ). The style is mainly influenced by the noise and the prompt further improves the style details, showing that our approach is flexible and can generate diverse results. + +# 4.4 Ablation Study + +Our approach achieves stylized generation from two perspectives. One is the "style" noise obtained via DDIM inversion, which is utilized as the initial noise during inference to provide fine-grained style information. The other is the style token, which is learned via prompt refinement tuning to describe the style precisely. In this section, we conduct the ablation study. The visualization results and quantitative results are shown in Fig. 8 (a) and Fig. 8 (b), respectively. Specifically, each point in Fig. 8 (b) presents the average style consistency score and content fidelity score of 100 generated images for a reference style. + +Impact of inversion noise. A prominent advantage of our INSTASTYLE is that we utilize the inversion noise as the initial image during inference time to preserve the fine-grained style information. Thereby, we first conduct an ablation study by replacing the inversion "style" noise with random noise. As shown in + +![](images/27548266a07c0e682ddb7c8380909a1d562657bb15c4fac7d54b7559f5a08dc2.jpg) +Fig. 8: Ablation study. (a) Visualization of ablation study. The reference style image is shown in the bottom right corner. Objects for synthesis are *Dinosaur*, *Bicycle*, *Bee*, and *Boy*, respectively. Without the inversion noise, the stylized performance is seriously degraded. Prompt refinement can further improve style details. (b) Quantitative results of ablation study. Our method lies further along the top right corner, showing better style preservation and content generation capability. + +Fig. 8 (a) (second row), without the inversion noise, the stylized performance is seriously degraded. The quantitative results in Fig. 8 (b) further illustrate that ablating the "style" noise (blue dots) will result in a lower style consistency score, harming the style of the generated image. + +Impact of prompt refinement tuning. We ablate the prompt refinement stage to show the necessity of learning a style token to avoid ambiguity and bias in human-writing textual style tokens. Fig. 8 (a) (third row) reveals that our learnable style token can better describe the style without harming the generation ability. The quantitative results in Fig. 8 (b) show that ablating the prompt refinement (green dots) will result in a lower content fidelity score and style score, further proving the importance of the learnable style token. + +# 5 Conclusion + +Our research solves the task of stylized generation by exploring the fine-grained style information in the reference image. Through extensive analysis and empirical investigation, we find that the initial noise with a non-zero signal-to-noise ratio in the diffusion model leans toward a particular style. Therefore, DDIM inversion is introduced to obtain an initial noise containing fine-grained style information. Furthermore, we highlight the challenge caused by the internal bias in textual style tokens and propose to learn a style token adaptively to better describe the style without bias. Our approach produces satisfactory stylized generation results and supports the combination of two styles. Extensive qualitative and quantitative comparisons demonstrate the superiority of our method. + +Acknowledgement This research is sponsored by National Natural Science Foundation of China (Grant No. 62306041, U21B2045, 62176025), Beijing Nova Program (Grant No. Z211100002121106, 20230484488, 20230484276), Youth Innovation Promotion Association CAS (Grant No.2022132), and Beijing Municipal Science & Technology Commission (Z231100007423015). + +# References + +1. Alaluf, Y., Richardson, E., Metzer, G., Cohen-Or, D.: A neural space-time representation for text-to-image personalization. ACM TOG (2023) +2. Chen, H., Zhao, L., Wang, Z., Zhang, H., Zuo, Z., Li, A., Xing, W., Lu, D.: Dualast: Dual style-learning networks for artistic style transfer. In: CVPR (2021) +3. Cho, J., Nam, G., Kim, S., Yang, H., Kwak, S.: Promptstyler: Prompt-driven style generation for source-free domain generalization. In: ICCV (2023) +4. Cui, X., Li, P., Li, Z., Liu, X., Zou, Y., He, Z.: Localize, understand, collaborate: Semantic-aware dragging via intention reasoner. arXiv preprint arXiv:2406.00432 (2024) +5. Cui, X., Li, Z., Li, P., Hu, Y., Shi, H., Cao, C., He, Z.: Chatedit: Towards multi-turn interactive facial image editing via dialogue. In: EMNLP (2023) +6. Deng, Y., Tang, F., Dong, W., Ma, C., Pan, X., Wang, L., Xu, C.: Stytr2: Image style transfer with transformers. In: CVPR (2022) +7. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS (2021) +8. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: CVPR (2021) +9. Everaert, M.N., Bocchio, M., Arpa, S., Susstrunk, S., Achanta, R.: Diffusion in style. In: ICCV (2023) +0. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. In: ICLR (2022) +1. Gal, R., Patashnik, O., Maron, H., Bermano, A.H., Chechik, G., Cohen-Or, D.: Stylegan-nada: Clip-guided domain adaptation of image generators. ACM TOG (2022) +2. Gatys, L.A., Ecker, A.S., Bethge, M.: Image style transfer using convolutional neural networks. In: CVPR (2016) +3. Gatys, L.A., Ecker, A.S., Bethge, M., Hertzmann, A., Shechtman, E.: Controlling perceptual factors in neural style transfer. In: CVPR (2017) +4. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. CACM (2020) +5. Gu, B., Fan, H., Zhang, L.: Two birds, one stone: A unified framework for joint learning of image and video style transfers. In: ICCV (2023) +6. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020) +7. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPSW (2021) +8. Hong, K., Jeon, S., Lee, J., Ahn, N., Kim, K., Lee, P., Kim, D., Uh, Y., Byun, H.: Aespa-net: Aesthetic pattern-aware style transfer networks. In: ICCV (2023) +9. Hong, K., Jeon, S., Lee, J., Ahn, N., Kim, K., Lee, P., Kim, D., Uh, Y., Byun, H.: Aespa-net: Aesthetic pattern-aware style transfer networks. In: ICCV (2023) + +20. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: ICLR (2021) +21. Huang, S., An, J., Wei, D., Luo, J., Pfister, H.: Quantart: Quantizing image style transfer towards high visual fidelity. In: CVPR (2023) +22. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: ICCV (2017) +23. Jia, G., Li, P., He, R.: Theme-aware aesthetic distribution prediction with full-resolution photographs. TNNLS (2022) +24. Jing, Y., Liu, X., Ding, Y., Wang, X., Ding, E., Song, M., Wen, S.: Dynamic instance normalization for arbitrary style transfer. In: AAAI (2020) +25. Jing, Y., Liu, Y., Yang, Y., Feng, Z., Yu, Y., Tao, D., Song, M.: Stroke controllable fast style transfer with adaptive receptive fields. In: ECCV (2018) +26. Jing, Y., Mao, Y., Yang, Y., Zhan, Y., Song, M., Wang, X., Tao, D.: Learning graph neural networks for image style transfer. In: ECCV (2022) +27. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019) +28. Ke, Z., Liu, Y., Zhu, L., Zhao, N., Lau, R.W.: Neural preset for color style transfer. In: CVPR (2023) +29. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013) +30. Krizhevsky, A., et al.: Learning multiple layers of features from tiny images. Technical report, University of Toronto (2009) +31. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: CVPR (2023) +32. Li, P., Hu, Y., He, R., Sun, Z.: Global and local consistent wavelet-domain age synthesis. TIFS (2019) +33. Li, P., Liu, X., Huang, J., Xia, D., Yang, J., Lu, Z.: Progressive generation of 3d point clouds with hierarchical consistency. PR (2023) +34. Li, P., Wang, R., Huang, H., He, R., He, Z.: Pluralistic aging diffusion autoencoder. In: ICCV (2023) +35. Li, P., Wu, X., Hu, Y., He, R., Sun, Z.: M2fpa: A multi-yaw multi-pitch high-quality dataset and benchmark for facial pose analysis. In: ICCV (2019) +36. Lin, S., Liu, B., Li, J., Yang, X.: Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891 (2023) +37. Lin, T., Ma, Z., Li, F., He, D., Li, X., Ding, E., Wang, N., Li, J., Gao, X.: Drafting and revision: Laplacian pyramid network for fast high-quality artistic style transfer. In: CVPR (2021) +38. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: ECCV (2022) +39. Liu, S., Lin, T., He, D., Li, F., Wang, M., Li, X., Sun, Z., Li, Q., Ding, E.: Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In: ICCV (2021) +40. Mao, J., Wang, X., Aizawa, K.: Guided image synthesis via initial image editing in diffusion model. In: ACM MM (2023) +41. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2021) +42. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: CVPR (2023) +43. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: ICLR (2022) + +44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) +45. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021) +46. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) +47. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015) +48. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023) +49. Ruiz, N., Li, Y., Jampani, V., Wei, W., Hou, T., Pritch, Y., Wadhwa, N., Rubinstein, M., Aberman, K.: Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949 (2023) +50. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: ICLR (2021) +51. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015) +52. Sohn, K., Ruiz, N., Lee, K., Chin, D.C., Blok, I., Chang, H., Barber, J., Jiang, L., Entis, G., Li, Y., et al.: Styledrop: Text-to-image generation in any style. In: NeurIPS (2023) +53. Sohn, K., Shaw, A., Hao, Y., Zhang, H., Polania, L., Chang, H., Jiang, L., Essa, I.: Learning disentangled prompts for compositional image synthesis. arXiv preprint arXiv:2306.00763 (2023) +54. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2020) +55. Tang, H., Liu, S., Lin, T., Huang, S., Li, F., He, D., Wang, X.: Master: Meta style transformer for controllable zero-shot and few-shot artistic style transfer. In: CVPR (2023) +56. Tang, J., Shu, X., Qi, G.J., Li, Z., Wang, M., Yan, S., Jain, R.: Tri-clustered tensor completion for social-aware image tag refinement. TPAMI (2016) +57. Teng, Q., Wang, R., Cui, X., Li, P., He, Z.: Exploring 3d-aware lifespan face aging via disentangled shape-texture representations. In: ICME (2024) +58. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) +59. Wang, H., Li, Y., Wang, Y., Hu, H., Yang, M.H.: Collaborative distillation for ultra-resolution universal style transfer. In: CVPR (2020) +60. Wang, R., Guo, H., Liu, J., Li, H., Zhao, H., Tang, X., Hu, Y., Tang, H., Li, P.: Stablegarment: Garment-centric generation via stable diffusion. arXiv preprint arXiv:2403.10783 (2024) +61. Wang, R., Li, P., Huang, H., Cao, C., He, R., He, Z.: Learning-to-rank meets language: Boosting language-driven ordering alignment for ordinal classification. In: NeurIPS (2023) +62. Wang, Z., Zhao, L., Xing, W.: Stylediffusion: Controllable disentangled style transfer via diffusion models. In: ICCV (2023) +63. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In: ICCV (2023) +64. Wen, L., Gao, C., Zou, C.: Cap-vstnet: Content affinity preserved versatile style transfer. In: CVPR (2023) + +65. Wu, X., Hu, Z., Sheng, L., Xu, D.: Styleformer: Real-time arbitrary style transfer via parametric style composition. In: ICCV (2021) +66. Wu, Z., Zhu, Z., Du, J., Bai, X.: Ccpl: contrastive coherence preserving loss for versatile style transfer. In: ECCV (2022) +67. Xie, X., Li, Y., Huang, H., Fu, H., Wang, W., Guo, Y.: Artistic style discovery with independent components. In: CVPR (2022) +68. Xu, W., Long, C., Nie, Y.: Learning dynamic style kernels for artistic style transfer. In: CVPR (2023) +69. Xu, Z., Sangineto, E., Sebe, N.: Stylerdalle: Language-guided style transfer using a vector-quantized tokenizer of a large-scale generative model. In: ICCV (2023) +70. Yang, S., Hwang, H., Ye, J.C.: Zero-shot contrastive loss for text-guided diffusion image style transfer. In: ICCV (2023) +71. Zhang, Y., Huang, N., Tang, F., Huang, H., Ma, C., Dong, W., Xu, C.: Inversion-based style transfer with diffusion models. In: CVPR (2023) +72. Zhang, Y., Tang, F., Dong, W., Huang, H., Ma, C., Lee, T.Y., Xu, C.: Domain enhanced arbitrary image style transfer via contrastive learning. In: ACM SIGGRAPH (2022) +73. Zhang, Z., Li, B., Nie, X., Han, C., Guo, T., Liu, L.: Towards consistent video editing with text-to-image diffusion models. In: NeurIPS (2023) +74. Zhou, Y., Zhang, R., Sun, T., Xu, J.: Enhancing detail preservation for customized text-to-image generation: A regularization-free approach. arXiv preprint arXiv:2305.13579 (2023) +75. Zhu, M., He, X., Wang, N., Wang, X., Gao, X.: All-to-key attention for arbitrary style transfer. In: ICCV (2023) \ No newline at end of file diff --git a/2024/InstaStyle_ Inversion Noise of a Stylized Image is Secretly a Style Adviser/images.zip b/2024/InstaStyle_ Inversion Noise of a Stylized Image is Secretly a Style Adviser/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..026f173c673b6ebbbbbcaad1a70db5c0ad3bf56f --- /dev/null +++ b/2024/InstaStyle_ Inversion Noise of a Stylized Image is Secretly a Style Adviser/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e94a5b2fd9715e05632636e762acc748f5b004cf1ab143c3b2fcdbe8e812dc09 +size 861086 diff --git a/2024/InstaStyle_ Inversion Noise of a Stylized Image is Secretly a Style Adviser/layout.json b/2024/InstaStyle_ Inversion Noise of a Stylized Image is Secretly a Style Adviser/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6cde4cc641878d29c6253b05921fe254ec9bab5b --- /dev/null +++ b/2024/InstaStyle_ Inversion Noise of a Stylized Image is Secretly a Style Adviser/layout.json @@ -0,0 +1,14882 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "text", + "content": "INSTASTYLE: Inversion Noise of a Stylized Image is Secretly a Style Adviser" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 167, + 467, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 167, + 467, + 194 + ], + "spans": [ + { + "bbox": [ + 146, + 167, + 467, + 194 + ], + "type": "text", + "content": "Xing Cui" + }, + { + "bbox": [ + 146, + 167, + 467, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 167, + 467, + 194 + ], + "type": "text", + "content": " Zekun Li" + }, + { + "bbox": [ + 146, + 167, + 467, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 167, + 467, + 194 + ], + "type": "text", + "content": " Peipei Li" + }, + { + "bbox": [ + 146, + 167, + 467, + 194 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 146, + 167, + 467, + 194 + ], + "type": "text", + "content": " Huaibo Huang" + }, + { + "bbox": [ + 146, + 167, + 467, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 146, + 167, + 467, + 194 + ], + "type": "text", + "content": " Xuannan Liu" + }, + { + "bbox": [ + 146, + 167, + 467, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 167, + 467, + 194 + ], + "type": "text", + "content": " Zhaofeng He" + }, + { + "bbox": [ + 146, + 167, + 467, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 196, + 201, + 417, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 201, + 417, + 213 + ], + "spans": [ + { + "bbox": [ + 196, + 201, + 417, + 213 + ], + "type": "text", + "content": "1 Beijing University of Posts and Telecommunications" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 222, + 213, + 392, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 213, + 392, + 224 + ], + "spans": [ + { + "bbox": [ + 222, + 213, + 392, + 224 + ], + "type": "text", + "content": "2 University of California, Santa Barbara" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 157, + 224, + 456, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 224, + 456, + 235 + ], + "spans": [ + { + "bbox": [ + 157, + 224, + 456, + 235 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 157, + 224, + 456, + 235 + ], + "type": "text", + "content": " MAIS & NLPR, Institute of Automation, Chinese Academy of Sciences" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 176, + 236, + 436, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 236, + 436, + 246 + ], + "spans": [ + { + "bbox": [ + 176, + 236, + 436, + 246 + ], + "type": "text", + "content": "{cuixing, lipeipei, liuxuannan, zhaofenghe}@bupt.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 187, + 247, + 426, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 247, + 426, + 257 + ], + "spans": [ + { + "bbox": [ + 187, + 247, + 426, + 257 + ], + "type": "text", + "content": "zekunli@cs.ucsb.edu huaibo.huang@cripac.ia.ac.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 181, + 258, + 432, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 258, + 432, + 268 + ], + "spans": [ + { + "bbox": [ + 181, + 258, + 432, + 268 + ], + "type": "text", + "content": "https://cuixing100876.github.io/instastyle.github.io/" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 149, + 289, + 192, + 377 + ], + "blocks": [ + { + "bbox": [ + 149, + 289, + 192, + 377 + ], + "lines": [ + { + "bbox": [ + 149, + 289, + 192, + 377 + ], + "spans": [ + { + "bbox": [ + 149, + 289, + 192, + 377 + ], + "type": "image", + "image_path": "1e63eab505223bc53f6724f12ef348ed063d982a0d5eed9da5777533207015f9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 191, + 378, + 279, + 386 + ], + "lines": [ + { + "bbox": [ + 191, + 378, + 279, + 386 + ], + "spans": [ + { + "bbox": [ + 191, + 378, + 279, + 386 + ], + "type": "text", + "content": "(a) Stylized image generation" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 193, + 289, + 235, + 376 + ], + "blocks": [ + { + "bbox": [ + 193, + 289, + 235, + 376 + ], + "lines": [ + { + "bbox": [ + 193, + 289, + 235, + 376 + ], + "spans": [ + { + "bbox": [ + 193, + 289, + 235, + 376 + ], + "type": "image", + "image_path": "169ca768357d38194f2bb25e7bf82d15cb71c2e183bd28dbc4ac52b090739872.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 240, + 289, + 275, + 376 + ], + "blocks": [ + { + "bbox": [ + 240, + 289, + 275, + 376 + ], + "lines": [ + { + "bbox": [ + 240, + 289, + 275, + 376 + ], + "spans": [ + { + "bbox": [ + 240, + 289, + 275, + 376 + ], + "type": "image", + "image_path": "69cb3e298b998b86a6927958978f4559c255977f7d7a973f04cf3751e19c297d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 278, + 289, + 319, + 376 + ], + "blocks": [ + { + "bbox": [ + 278, + 289, + 319, + 376 + ], + "lines": [ + { + "bbox": [ + 278, + 289, + 319, + 376 + ], + "spans": [ + { + "bbox": [ + 278, + 289, + 319, + 376 + ], + "type": "image", + "image_path": "84c629f341a3fd09e2ee3f6cd287e8f9f227a9693baefa5afcb66f43fc32c2af.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 328, + 289, + 377, + 376 + ], + "blocks": [ + { + "bbox": [ + 328, + 289, + 377, + 376 + ], + "lines": [ + { + "bbox": [ + 328, + 289, + 377, + 376 + ], + "spans": [ + { + "bbox": [ + 328, + 289, + 377, + 376 + ], + "type": "image", + "image_path": "b473543400ab34673d2adfbf995d8edb1c31fb93a1762a0942448bb5d951a3dc.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 378, + 448, + 386 + ], + "lines": [ + { + "bbox": [ + 355, + 378, + 448, + 386 + ], + "spans": [ + { + "bbox": [ + 355, + 378, + 448, + 386 + ], + "type": "text", + "content": "(b) Combination of two styles" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 378, + 289, + 419, + 376 + ], + "blocks": [ + { + "bbox": [ + 378, + 289, + 419, + 376 + ], + "lines": [ + { + "bbox": [ + 378, + 289, + 419, + 376 + ], + "spans": [ + { + "bbox": [ + 378, + 289, + 419, + 376 + ], + "type": "image", + "image_path": "1c4d6b7cda07a19095eb7c0f005b26060b39492853b3c4481b232b0aab1effce.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 421, + 289, + 463, + 376 + ], + "blocks": [ + { + "bbox": [ + 421, + 289, + 463, + 376 + ], + "lines": [ + { + "bbox": [ + 421, + 289, + 463, + 376 + ], + "spans": [ + { + "bbox": [ + 421, + 289, + 463, + 376 + ], + "type": "image", + "image_path": "97529d597569d6ed5e4480cfb297ad0f2af082ea4440564d09051db94f66be3f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 151, + 388, + 192, + 435 + ], + "blocks": [ + { + "bbox": [ + 151, + 388, + 192, + 435 + ], + "lines": [ + { + "bbox": [ + 151, + 388, + 192, + 435 + ], + "spans": [ + { + "bbox": [ + 151, + 388, + 192, + 435 + ], + "type": "image", + "image_path": "9330de21c23d5472d0ff22763d378e135a93505a0d1b824f021dedc8d5609c02.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 460, + 482, + 515 + ], + "lines": [ + { + "bbox": [ + 131, + 460, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 131, + 460, + 482, + 515 + ], + "type": "text", + "content": "Fig.1: Visualization of INSTASTYLE. (a) Our method excels at capturing style details and distinguishing between similar styles. (b) The first and third columns show images styled with reference to style 1 and style 2, respectively. The middle column shows images in a combined style. (c) Our method supports adjusting the degree of two styles during combination, dynamically ranging from one style to another." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 200, + 388, + 242, + 435 + ], + "blocks": [ + { + "bbox": [ + 200, + 388, + 242, + 435 + ], + "lines": [ + { + "bbox": [ + 200, + 388, + 242, + 435 + ], + "spans": [ + { + "bbox": [ + 200, + 388, + 242, + 435 + ], + "type": "image", + "image_path": "907bae5c8e0dc894846f9e6f17ff37fd2c55f8afe2aa130360b62c3694a522a6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 441, + 380, + 450 + ], + "lines": [ + { + "bbox": [ + 233, + 441, + 380, + 450 + ], + "spans": [ + { + "bbox": [ + 233, + 441, + 380, + 450 + ], + "type": "text", + "content": "(c) Smooth style combination between two styles" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 244, + 388, + 285, + 435 + ], + "blocks": [ + { + "bbox": [ + 244, + 388, + 285, + 435 + ], + "lines": [ + { + "bbox": [ + 244, + 388, + 285, + 435 + ], + "spans": [ + { + "bbox": [ + 244, + 388, + 285, + 435 + ], + "type": "image", + "image_path": "55ba7284fe8c72a89735b12aa9a149385830693aae246f036db71da7804d6512.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 286, + 388, + 328, + 435 + ], + "blocks": [ + { + "bbox": [ + 286, + 388, + 328, + 435 + ], + "lines": [ + { + "bbox": [ + 286, + 388, + 328, + 435 + ], + "spans": [ + { + "bbox": [ + 286, + 388, + 328, + 435 + ], + "type": "image", + "image_path": "267eaa01b35a34ab943f9ada63666182b2a808ac49138da21b5861ffa48e9bce.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 328, + 388, + 370, + 435 + ], + "blocks": [ + { + "bbox": [ + 328, + 388, + 370, + 435 + ], + "lines": [ + { + "bbox": [ + 328, + 388, + 370, + 435 + ], + "spans": [ + { + "bbox": [ + 328, + 388, + 370, + 435 + ], + "type": "image", + "image_path": "c07f653d646e4259d232a88b8f4c396b95f600fd4f1138d064b6859e612e65a8.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 371, + 388, + 413, + 435 + ], + "blocks": [ + { + "bbox": [ + 371, + 388, + 413, + 435 + ], + "lines": [ + { + "bbox": [ + 371, + 388, + 413, + 435 + ], + "spans": [ + { + "bbox": [ + 371, + 388, + 413, + 435 + ], + "type": "image", + "image_path": "8f12bad59ba600af0c00a833e3d7ab06fdbc0c2280cf101b86b5e6ab0fce7cc1.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 422, + 396, + 463, + 435 + ], + "blocks": [ + { + "bbox": [ + 432, + 388, + 452, + 396 + ], + "lines": [ + { + "bbox": [ + 432, + 388, + 452, + 396 + ], + "spans": [ + { + "bbox": [ + 432, + 388, + 452, + 396 + ], + "type": "text", + "content": "Style 2" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 422, + 396, + 463, + 435 + ], + "lines": [ + { + "bbox": [ + 422, + 396, + 463, + 435 + ], + "spans": [ + { + "bbox": [ + 422, + 396, + 463, + 435 + ], + "type": "image", + "image_path": "4bcbdf770e23e41f3ada368e2fd9598647f1c8198b5ab2243a224d827bbd285c.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 159, + 536, + 453, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 536, + 453, + 646 + ], + "spans": [ + { + "bbox": [ + 159, + 536, + 453, + 646 + ], + "type": "text", + "content": "Abstract. Stylized text-to-image generation focuses on creating images from textual descriptions while adhering to a style specified by reference images. However, subtle style variations within different reference images can hinder the model from accurately learning the target style. In this paper, we propose INSTASTYLE, a novel approach that excels in generating high-fidelity stylized images with only a single reference image. Our approach is based on the finding that the inversion noise from a stylized reference image inherently carries the style signal, as evidenced by their non-zero signal-to-noise ratio. We employ DDIM inversion to extract this noise from the reference image and leverage a" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 654, + 235, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 235, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 235, + 665 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 205 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 205 + ], + "type": "text", + "content": "diffusion model to generate new stylized images from the \"style\" noise. Additionally, the inherent ambiguity and bias of textual prompts impede the precise conveying of style during image inversion. To address this, we devise prompt refinement, which learns a style token assisted by human feedback. Qualitative and quantitative experimental results demonstrate that INSTASTYLE achieves superior performance compared to current benchmarks. Furthermore, our approach also showcases its capability in the creative task of style combination with mixed inversion noise." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 216, + 455, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 216, + 455, + 239 + ], + "spans": [ + { + "bbox": [ + 159, + 216, + 455, + 239 + ], + "type": "text", + "content": "Keywords: Stylized image generation " + }, + { + "bbox": [ + 159, + 216, + 455, + 239 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 216, + 455, + 239 + ], + "type": "text", + "content": " Inversion noise " + }, + { + "bbox": [ + 159, + 216, + 455, + 239 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 216, + 455, + 239 + ], + "type": "text", + "content": " Signal-to-noise ratio " + }, + { + "bbox": [ + 159, + 216, + 455, + 239 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 216, + 455, + 239 + ], + "type": "text", + "content": " Prompt refinement" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 263, + 230, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 230, + 276 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 230, + 276 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 293, + 299, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 299, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 299, + 484 + ], + "type": "text", + "content": "Recently, the advent of DDPMs [16, 51] has ushered in a new era of image generation. Existing text-to-image generation methods [5, 43, 45, 46] can generate images in coarse-grained styles by incorporating style descriptions in the textual prompt. However, the textual prompt is ambiguous which makes it hard to express style precisely [11]. Recently, personalized generation [10,48,63] has been introduced to generate novel concepts, which can be utilized for stylized generation by viewing the reference style as a new concept. These approaches typically necessitate mul" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "type": "text", + "content": "tiple images as references. However, the subtle style variations in these multiple images present a challenge for the model to accurately learn and replicate the intended style. As shown in the first column of Fig. 1 (a), although the reference image Houses at Auvers (top) and Thatched Cottages by a Hill (bottom) are all Van Gogh's artworks, the differences in terms of color and texture [2] pose a significant challenge. Some recent approaches [3,53,69] are designed to more accurately describe the target style during generation. However, they may ignore the fine-grained style information in the reference images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we propose INSTASTYLE (Inversion Noise of a Stylized Image is Secretly a Style Adviser) based on the observation that the inversion noise from a stylized reference image inherently contains the style signal. Specifically, our approach only requires a single reference image, which is transformed into a noise via DDIM inversion. The inversion noise, which preserves the style signal, is then used to generate stylized images by utilizing diffusion models. To shed light on this phenomenon, we demonstrate that the inversion noise maintains" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 297, + 482, + 381 + ], + "blocks": [ + { + "bbox": [ + 307, + 297, + 482, + 381 + ], + "lines": [ + { + "bbox": [ + 307, + 297, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 482, + 381 + ], + "type": "image", + "image_path": "c6e36f47eb5ca9567d1b717d049362c833b3659b33beacdc9ab7cb5b91e60c4c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 388, + 482, + 466 + ], + "lines": [ + { + "bbox": [ + 304, + 388, + 482, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 482, + 466 + ], + "type": "text", + "content": "Fig. 2: Motivation. Sampling from the inversion noise of a reference image can generate stylized images. However, the optimal style token varies for each case among human-written style tokens. Our learnable style token, i.e., \"\", shows greater universality across various scenarios." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "X. Cui et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "a non-zero signal-to-noise ratio in Sec. 3.2, indicating that it retains essential information (including the style signal) from the reference image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 354 + ], + "type": "text", + "content": "Nevertheless, directly using human-written prompts to describe styles for inverting the reference image often encounters challenges due to the inherent ambiguity of natural language. For instance, specific style descriptors like \"ink painting\" don't always work effectively with various objects. As shown in Fig. 2, the style \"ink painting\" might suit an object like a \"chair\" but not as well with a \"boy\". This inconsistency can arise because \"ink painting\" is typically associated with landscapes and might not yield optimal results for human subjects. Conversely, when using vague descriptors like \"specific style\", they may fail to provide enough information, leading to unpredictable generation quality (see the second row in Fig. 2). To address this, we propose Prompt Refinement to learn a style token with the assistance of human feedback. As illustrated in the last row of Fig. 2, the style information is boosted by using the learned style token in the noising process to obtain a better \"style\" noise. Therefore, our prompt refinement stands apart from previous methods [3, 53, 69], where style tokens are utilized in the denoising process to generate target images. During prompt refinement, we first collect images via human feedback. Then we optimize the embedding of the style token and the key and value projection matrices in the cross-attention layers of the diffusion model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 356, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 416 + ], + "type": "text", + "content": "As shown in Fig. 1 (a), our approach can effectively retain the fine-grained style in the reference image as well as generate new objects in high fidelity. Furthermore, our method supports the creative generation of style combination (Fig. 1 (b) and (c)) and allows adjusting the degree of two styles dynamically (Fig. 1 (c)). To sum up, our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 422, + 480, + 553 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 422, + 479, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 479, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 479, + 457 + ], + "type": "text", + "content": "- We find that the inversion noise of a reference image via DDIM inversion potentially retains style information and evidence this observation by the non-zero signal-to-noise ratio of the inversion noise." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 459, + 480, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 480, + 517 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 480, + 517 + ], + "type": "text", + "content": "- We propose INSTASTYLE based on our observation, which leverages the style signal in inversion noise. Additionally, to better represent the style during inversion, a prompt refinement scheme is designed to learn the style token. Human feedback is also introduced to enhance model alignment with human preferences during prompt refinement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 518, + 479, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 518, + 479, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 518, + 479, + 553 + ], + "type": "text", + "content": "- Qualitative and quantitative experimental results show the superiority of our INSTASTYLE in generating high-quality stylized images. Moreover, it exhibits promising potential in creative scenarios like style combinations." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 572, + 237, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 572, + 237, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 572, + 237, + 585 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 597, + 282, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 282, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 282, + 609 + ], + "type": "text", + "content": "2.1 Text-to-image Synthesis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Image synthesis is an essential subject in computer vision [7,23,41,56]. Previous works are based on variational auto-encoder [29, 33] or generative adversarial networks [5,14,32,35,57]. With the advance of pretrained vison-language models [44,61] and diffusion models [34,54,73], text-to-image generation [4,8,43,45," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "INSTASTYLE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "[60] has been widely studied and shown remarkable generalization ability. Recently, Lin et al. [36] have pointed out that the noise of the last step during noisering in the training process has a non-zero signal-to-noise ratio (SNR), i.e., there exists a signal leak. This results in a misalignment between the training and inference process. Therefore, some works [36, 50] propose to train diffusion models by enforcing the SNR to zero to avoid the signal leak. On the contrary, our approach leverages the leaked signal, which potentially includes the style details, from the reference image for the stylized generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 321 + ], + "type": "text", + "content": "As these methods ignore the concepts that do not appear in the training set [10], some works [10,48,49] study personalized text-to-image generation which aims to adapt text-to-image models to new concepts given several reference images. For example, Textual Inversion [10] introduces and optimizes a word vector for each new concept. Subsequent works [1,63,74] further enhance the flexibility and adaptiveness of the learning strategy. Some methods [20, 48, 49] propose to finetune the original diffusion model, showing more satisfactory results. However, maintaining fine-grained style while simultaneously generating new objects remains a challenge for current personalized generation methods [52]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 341, + 294, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 341, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 131, + 341, + 294, + 354 + ], + "type": "text", + "content": "2.2 Stylized Image Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 364, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 533 + ], + "type": "text", + "content": "Stylized image generation is a new paradigm for image generation which aims to generate content in a specific style given a few reference images. Although it is similar to the neural style transfer task [12, 13, 22, 27, 37, 59] which generates stylized images as well, they are fundamentally different. Style transfer solves an image translation task, focusing on translating a content image to target style [15, 21, 24, 28, 62, 64, 67, 68, 70]. For example, some works explore the global and local information to preserve the content of the source image [6, 18, 25, 26, 55, 65, 71, 75]. On the contrary, stylized image generation is geared towards generating a new image in specific style conditioned on a text. For example, ZDAIS [53] views stylized image generation as a domain adaptive task, where each style belongs to a domain. It learns disentangled prompts to adapt the model to new domains. Some approaches [9, 52] fine-tune the model to capture style properties. Our work differs from theirs by proposing a framework based on our observation that inversion noise contains style information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 553, + 202, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 553, + 202, + 566 + ], + "spans": [ + { + "bbox": [ + 131, + 553, + 202, + 566 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "As shown in Fig. 3, our proposed method involves two main stages. In the first stage, we employ DDIM inversion to transform the reference image into noise. Notably, the inversion noise exhibits a non-zero signal-to-noise ratio, suggesting the presence of style signals from the reference image. Subsequently, we generate " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " stylized images from the inversion noise conditioned on the given textual prompts. Due to the inherent ambiguity of the textual prompts that describe the style, it is challenging to precisely convey the desired style. Addressing this," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "X. Cui et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 482, + 213 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 213 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 213 + ], + "type": "image", + "image_path": "eb3fe407cfbedfbac602427d450029f373f0a543a3e1a5fda5646613d6d87ca0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 230, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 130, + 230, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 482, + 285 + ], + "type": "text", + "content": "Fig. 3: The training process of INSTASTYLE. (a) The first stage is initial stylized image generation. The reference image is inverted to noise conditioned on a prompt via DDIM Inversion. Then the inversion noise is utilized to generate initial stylized images. (b) The second stage is prompt refinement which leverages the selected high-quality initial stylized images to learn a style token." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 308, + 482, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 346 + ], + "type": "text", + "content": "the second stage involves the incorporation of human feedback to select " + }, + { + "bbox": [ + 130, + 308, + 482, + 346 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 308, + 482, + 346 + ], + "type": "text", + "content": " high-quality generated images from the first stage. The selected images are then used to learn a style token via prompt refinement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 362, + 228, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 362, + 228, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 362, + 228, + 374 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 382, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 482, + 441 + ], + "type": "text", + "content": "We begin by reviewing the fundamental diffusion model. Subsequently, we introduce Stable Diffusion [46], a key component of our framework. Finally, we discuss both DDIM and DDIM inversion. In our method, DDIM serves the purpose of denoising a noise to generate new images. Concurrently, DDIM inversion is utilized to transform the reference image into a corresponding \"style\" noise." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 441, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 478 + ], + "type": "text", + "content": "Diffusion models. Diffusion models [16, 51] contain a forward process and a backward process. The forward process adds noise to the data according to a predetermined, non-learned variance schedule " + }, + { + "bbox": [ + 130, + 441, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\beta_{1},\\ldots,\\beta_{T}" + }, + { + "bbox": [ + 130, + 441, + 482, + 478 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 222, + 486, + 481, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 486, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 222, + 486, + 481, + 500 + ], + "type": "interline_equation", + "content": "q \\left(z _ {t} \\mid z _ {t - 1}\\right) := \\mathcal {N} \\left(z _ {t}; \\sqrt {1 - \\beta_ {t}} z _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right). \\tag {1}", + "image_path": "9d07a534a08839ab8047e62a2e4f37e1a849b9188ef2ee22617cbe53042c059b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 508, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 481, + 533 + ], + "type": "text", + "content": "An important property of the forward process is that we can obtain the latent variable " + }, + { + "bbox": [ + 130, + 508, + 481, + 533 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 508, + 481, + 533 + ], + "type": "text", + "content": " directly based on " + }, + { + "bbox": [ + 130, + 508, + 481, + 533 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 130, + 508, + 481, + 533 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 253, + 540, + 481, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 540, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 253, + 540, + 481, + 555 + ], + "type": "interline_equation", + "content": "z _ {t} = \\sqrt {\\alpha_ {t}} z _ {0} + \\sqrt {1 - \\alpha_ {t}} \\varepsilon , \\tag {2}", + "image_path": "5f9e17ca263406314b31f986ad703cd5183fb90fa82454a10150cfaf0f40734c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 562, + 481, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 562, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 130, + 562, + 481, + 587 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 562, + 481, + 587 + ], + "type": "inline_equation", + "content": "\\varepsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 130, + 562, + 481, + 587 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 562, + 481, + 587 + ], + "type": "inline_equation", + "content": "\\alpha_{t}\\coloneqq \\prod_{i = 1}^{t}(1 - \\beta_{i})" + }, + { + "bbox": [ + 130, + 562, + 481, + 587 + ], + "type": "text", + "content": ". Diffusion models restore the information by learning the backward process:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 214, + 596, + 481, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 596, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 214, + 596, + 481, + 610 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(z _ {t - 1} \\mid z _ {t}\\right) := \\mathcal {N} \\left(z _ {t - 1}; \\mu_ {\\theta} \\left(z _ {t}, t\\right), \\Sigma_ {\\theta} \\left(z _ {t}, t\\right)\\right). \\tag {3}", + "image_path": "16b37ae23bcaa0f332cb99f060339d67055255a2011aec4ab8fc02ade7342793.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": "Stable Diffusion. In Stable Diffusion, the diffusion process happens in latent space [46]. It utilizes a pretrained autoencoder which consists of an encoder and a decoder. The encoder " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\cdot)" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " maps an image " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " into a latent code " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " and the decoder " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\cdot)" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " transforms the latent code back into an image. The denoising model " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\varepsilon_{\\theta}(\\cdot)" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "INSTASTYLE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 152 + ], + "type": "text", + "content": "is an Unet [47]. During training, artificial noise is added to the sampled data " + }, + { + "bbox": [ + 132, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 132, + 116, + 479, + 152 + ], + "type": "text", + "content": " according to timestamp " + }, + { + "bbox": [ + 132, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 116, + 479, + 152 + ], + "type": "text", + "content": " based on Eq. (2), resulting in a noised sample " + }, + { + "bbox": [ + 132, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 132, + 116, + 479, + 152 + ], + "type": "text", + "content": ". Denoising model starts with " + }, + { + "bbox": [ + 132, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 132, + 116, + 479, + 152 + ], + "type": "text", + "content": " and is trained to predict the artificial noise:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 202, + 160, + 481, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 160, + 481, + 178 + ], + "spans": [ + { + "bbox": [ + 202, + 160, + 481, + 178 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} E _ {z _ {0}, \\varepsilon \\sim N (0, I), t \\sim \\operatorname {U n i f o r m} (1, T)} \\| \\varepsilon - \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C}\\right) \\| _ {2} ^ {2}, \\tag {4}", + "image_path": "df47a63c1f50991e5ee6c4e065bcc6f2e55ebfddba891aaac47a422bf288b8db.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 186, + 372, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 186, + 372, + 198 + ], + "spans": [ + { + "bbox": [ + 132, + 186, + 372, + 198 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 186, + 372, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\psi (\\mathcal{P})" + }, + { + "bbox": [ + 132, + 186, + 372, + 198 + ], + "type": "text", + "content": " is the embedding of text condition " + }, + { + "bbox": [ + 132, + 186, + 372, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 198, + 479, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 198, + 479, + 234 + ], + "spans": [ + { + "bbox": [ + 132, + 198, + 479, + 234 + ], + "type": "text", + "content": "DDIM. In inference time, given a noise vector " + }, + { + "bbox": [ + 132, + 198, + 479, + 234 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 132, + 198, + 479, + 234 + ], + "type": "text", + "content": ", the noise is gradually removed by sequentially predicting the added noise for " + }, + { + "bbox": [ + 132, + 198, + 479, + 234 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 132, + 198, + 479, + 234 + ], + "type": "text", + "content": " steps. DDIM [54] is one of the denoising approaches with a deterministic process:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 190, + 243, + 481, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 243, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 190, + 243, + 481, + 266 + ], + "type": "interline_equation", + "content": "z _ {t - 1} = \\sqrt {\\frac {\\alpha_ {t - 1}}{\\alpha_ {t}}} z _ {t} + \\left(\\sqrt {\\frac {1}{\\alpha_ {t - 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\cdot \\tilde {\\varepsilon} _ {\\theta}, \\qquad (5)", + "image_path": "47d9bb34503d8671d50a6dc52e5a6d7e32f7c87965f6e719276c181917e01e73.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 273, + 273, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 273, + 273, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 273, + 273, + 284 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 273, + 273, + 284 + ], + "type": "inline_equation", + "content": "\\tilde{\\varepsilon}_{\\theta}" + }, + { + "bbox": [ + 132, + 273, + 273, + 284 + ], + "type": "text", + "content": " is the estimated noise." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 285, + 479, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 479, + 309 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 479, + 309 + ], + "type": "text", + "content": "DDIM inversion [42, 54] transforms an image to a noise conditioned on a prompt. The diffusion process is performed in the reverse direction, i.e., " + }, + { + "bbox": [ + 132, + 285, + 479, + 309 + ], + "type": "inline_equation", + "content": "z_0 \\to z_T" + }, + { + "bbox": [ + 132, + 285, + 479, + 309 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 190, + 318, + 481, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 318, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 190, + 318, + 481, + 341 + ], + "type": "interline_equation", + "content": "z _ {t + 1} = \\sqrt {\\frac {\\alpha_ {t + 1}}{\\alpha_ {t}}} z _ {t} + \\left(\\sqrt {\\frac {1}{\\alpha_ {t + 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\cdot \\tilde {\\varepsilon} _ {\\theta}. \\tag {6}", + "image_path": "5ceea9258a6824180906495b0550612c02a7fb96156105378e56abcb30807467.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 356, + 329, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 356, + 329, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 356, + 329, + 369 + ], + "type": "text", + "content": "3.2 Initial Stylized Image Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 375, + 479, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 479, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 479, + 399 + ], + "type": "text", + "content": "As shown in Fig. 3 (a), the first stage of our INSTASTYLE involves obtaining the inversion noise via DDIM inversion and sampling images via DDIM." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 399, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 399, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 481, + 495 + ], + "type": "text", + "content": "In DDIM inversion, the added noise for each step is calculated conditioned on a prompt and gradually incorporated into the reference image following Eq. (6). Specifically, we set the learnable style token as a human-written style description in this stage. We demonstrate that the inversion noise from a stylized reference image inherently carries the style signal from the perspective of the signal-to-noise ratio. As the estimated noise " + }, + { + "bbox": [ + 132, + 399, + 481, + 495 + ], + "type": "inline_equation", + "content": "\\varepsilon_{\\theta}(z_t,t,\\mathcal{C})" + }, + { + "bbox": [ + 132, + 399, + 481, + 495 + ], + "type": "text", + "content": " is trained to approximate the artificial noise " + }, + { + "bbox": [ + 132, + 399, + 481, + 495 + ], + "type": "inline_equation", + "content": "\\varepsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 132, + 399, + 481, + 495 + ], + "type": "text", + "content": ", we assume that " + }, + { + "bbox": [ + 132, + 399, + 481, + 495 + ], + "type": "inline_equation", + "content": "\\varepsilon_{\\theta}(z_t,t,\\mathcal{C}) \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 132, + 399, + 481, + 495 + ], + "type": "text", + "content": ". According to Eq. (6), " + }, + { + "bbox": [ + 132, + 399, + 481, + 495 + ], + "type": "inline_equation", + "content": "z_{t + 1}" + }, + { + "bbox": [ + 132, + 399, + 481, + 495 + ], + "type": "text", + "content": " can be approximated in a closed form:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 172, + 505, + 481, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 505, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 172, + 505, + 481, + 530 + ], + "type": "interline_equation", + "content": "z _ {t + 1} = \\sqrt {\\frac {\\alpha_ {t + 1}}{\\alpha_ {0}}} z _ {0} + \\sqrt {\\sum_ {i = 0} ^ {t} \\frac {\\alpha_ {t + 1}}{\\alpha_ {i + 1}} \\left(\\sqrt {\\frac {1}{\\alpha_ {i + 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {i}} - 1}\\right) ^ {2}} \\cdot \\bar {\\varepsilon} _ {0}, \\tag {7}", + "image_path": "75363d84f58232d84da632bf5df5bd00d02cd53385ca8da45414bfc83e5a54bf.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 537, + 220, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 220, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 220, + 550 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 537, + 220, + 550 + ], + "type": "inline_equation", + "content": "\\bar{\\varepsilon}_0\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 550, + 479, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 550, + 479, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 550, + 479, + 585 + ], + "type": "text", + "content": "Signal-to-Noise ratio. Signal-to-noise ratio (SNR) is introduced to measure the ratio of signals from the original image preserved in the noise [36,46]. The SNR of the inversion noise can be calculated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 206, + 594, + 481, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 594, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 206, + 594, + 481, + 622 + ], + "type": "interline_equation", + "content": "\\operatorname {S N R} (t) := \\frac {1}{\\sum_ {i = 0} ^ {t} \\frac {\\alpha_ {0}}{\\alpha_ {i + 1}} \\left(\\sqrt {\\frac {1}{\\alpha_ {i + 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {i}} - 1}\\right) ^ {2}}. \\tag {8}", + "image_path": "7e804e2cb136ba1a2af877d3d6d356992e5f9c13bdd434094c13d009806513a6.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 629, + 479, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 479, + 667 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 479, + 667 + ], + "type": "text", + "content": "We present detailed derivations in Sec. B in Supplementary Material. In Stable Diffusion [46], the predetermined variance schedule " + }, + { + "bbox": [ + 132, + 629, + 479, + 667 + ], + "type": "inline_equation", + "content": "\\beta_{t} = (\\sqrt{0.00085}\\cdot (1 - j) + \\sqrt{0.012}\\cdot j)^{2}" + }, + { + "bbox": [ + 132, + 629, + 479, + 667 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 629, + 479, + 667 + ], + "type": "inline_equation", + "content": "j = \\frac{t - 1}{T - 1}" + }, + { + "bbox": [ + 132, + 629, + 479, + 667 + ], + "type": "text", + "content": ". At the terminal timestep " + }, + { + "bbox": [ + 132, + 629, + 479, + 667 + ], + "type": "inline_equation", + "content": "T = 1000" + }, + { + "bbox": [ + 132, + 629, + 479, + 667 + ], + "type": "text", + "content": ", the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 218, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 218, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 218, + 101 + ], + "type": "text", + "content": "X. Cui et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "SNR(T) = 0.015144" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": " has a none-zero signal-to-noise ratio. This nonzero signal-to-noise ratio suggests that the noise obtained via DDIM inversion still retains information from the reference image and deviates from white noise. Besides, we conduct qualitative experiments and observe that style information can consistently be preserved at the terminal timestep (refer to Sec. C.3 in Supplementary Material). Therefore, we can generate target stylized images by sampling from the inversion noise based on Eq. (5)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 217, + 261, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 217, + 261, + 229 + ], + "spans": [ + { + "bbox": [ + 131, + 217, + 261, + 229 + ], + "type": "text", + "content": "3.3 Prompt Refinement" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 237, + 480, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 480, + 320 + ], + "type": "text", + "content": "As the natural language prompts may not precisely convey the style (Fig. 2), we propose Prompt Refinement to learn the style tokens as shown in Fig. 3 (b). We utilize the generated stylized images in the first stage to constitute the training data for prompt refinement learning. While these generated images may not always be precise, we observe that most of them successfully retain the style information alongside appropriate content. We manually select images that distinctly embody the reference style and the target object to build the dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 321, + 481, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 481, + 429 + ], + "type": "text", + "content": "Specifically, we introduce a new token, i.e., “” to represent the style descriptor and learn its embedding. In practice, we initialize the new token with the embeddings of the textual style descriptor of the reference image. The text condition is input to the model via two projection matrices in the cross-attention block of diffusion model " + }, + { + "bbox": [ + 130, + 321, + 481, + 429 + ], + "type": "inline_equation", + "content": "\\varepsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 321, + 481, + 429 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 321, + 481, + 429 + ], + "type": "inline_equation", + "content": "W^{k} \\in \\mathbb{R}^{d \\times d'}" + }, + { + "bbox": [ + 130, + 321, + 481, + 429 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 321, + 481, + 429 + ], + "type": "inline_equation", + "content": "W^{v} \\in \\mathbb{R}^{d \\times d'}" + }, + { + "bbox": [ + 130, + 321, + 481, + 429 + ], + "type": "text", + "content": ". Therefore, we also fine-tune these two projection matrices. The text feature " + }, + { + "bbox": [ + 130, + 321, + 481, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{R}^{s \\times d}" + }, + { + "bbox": [ + 130, + 321, + 481, + 429 + ], + "type": "text", + "content": " is projected to key " + }, + { + "bbox": [ + 130, + 321, + 481, + 429 + ], + "type": "inline_equation", + "content": "K = \\mathbf{c}W^{k}" + }, + { + "bbox": [ + 130, + 321, + 481, + 429 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 130, + 321, + 481, + 429 + ], + "type": "inline_equation", + "content": "V = \\mathbf{c}W^{v}" + }, + { + "bbox": [ + 130, + 321, + 481, + 429 + ], + "type": "text", + "content": ". The query matrix " + }, + { + "bbox": [ + 130, + 321, + 481, + 429 + ], + "type": "inline_equation", + "content": "W^{q} \\in \\mathbb{R}^{l \\times d'}" + }, + { + "bbox": [ + 130, + 321, + 481, + 429 + ], + "type": "text", + "content": " projects the latent image feature " + }, + { + "bbox": [ + 130, + 321, + 481, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{f} \\in \\mathbb{R}^{(h \\times w) \\times l}" + }, + { + "bbox": [ + 130, + 321, + 481, + 429 + ], + "type": "text", + "content": " into query feature " + }, + { + "bbox": [ + 130, + 321, + 481, + 429 + ], + "type": "inline_equation", + "content": "Q = \\mathbf{f}W^{q}" + }, + { + "bbox": [ + 130, + 321, + 481, + 429 + ], + "type": "text", + "content": ". Then the cross-attention [58] is calculated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 218, + 438, + 480, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 438, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 218, + 438, + 480, + 464 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d ^ {\\prime}}}\\right) V. \\tag {9}", + "image_path": "97db405996cdd2036a2b4fc3f9fe6ac94b77e4a899444cde6b425574bbe30256.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "spans": [ + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "text", + "content": "We utilize the LoRA [20] for model fine-tuning. Specifically, for a projection matrix " + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{d \\times d'}" + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "inline_equation", + "content": "W = W^k" + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "inline_equation", + "content": "W = W^v" + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "text", + "content": ") to be fine-tuned, we update a low-rank residual rather than directly fine-tuning " + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "text", + "content": ". Formally, we denote the fine-tuned projection matrix as " + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "inline_equation", + "content": "W' = W + BA" + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{d \\times r}" + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{r \\times d'}" + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "text", + "content": ", and the rank " + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "inline_equation", + "content": "r \\ll \\min(d, d')" + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "text", + "content": ". During training, only " + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "text", + "content": " are learnable. For " + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "inline_equation", + "content": "y = Wx" + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "text", + "content": ", the modified forward pass is " + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "inline_equation", + "content": "y = Wx + BAx" + }, + { + "bbox": [ + 131, + 472, + 479, + 556 + ], + "type": "text", + "content": ". Putting the two stages together, our full algorithm is shown in Algorithm 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 574, + 208, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 208, + 585 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 208, + 585 + ], + "type": "text", + "content": "3.4 Inference" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "content": "Stylized image generation. For stylized image generation, we first learn the style token \"\" to describe the style in the reference image. During inference, the learned style token is integrated into the prompt that inverts image and the prompt that generates the target image. Besides, the diffusion model " + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\varepsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "content": " is utilized as the backbone of DDIM inversion and DDIM. Specifically, the reference image is first inverted to a noise " + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "content": " via DDIM inversion conditioned" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "INSTASTYLE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 133, + 465, + 167 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 138, + 133, + 465, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 133, + 465, + 156 + ], + "spans": [ + { + "bbox": [ + 138, + 133, + 465, + 156 + ], + "type": "text", + "content": "1. Input: A source prompt embedding " + }, + { + "bbox": [ + 138, + 133, + 465, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\psi(\\mathcal{P})" + }, + { + "bbox": [ + 138, + 133, + 465, + 156 + ], + "type": "text", + "content": " and a reference image " + }, + { + "bbox": [ + 138, + 133, + 465, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 138, + 133, + 465, + 156 + ], + "type": "text", + "content": ". Initial style token embedding " + }, + { + "bbox": [ + 138, + 133, + 465, + 156 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 138, + 133, + 465, + 156 + ], + "type": "text", + "content": " is included in the prompt embedding " + }, + { + "bbox": [ + 138, + 133, + 465, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 138, + 133, + 465, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 156, + 434, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 156, + 434, + 167 + ], + "spans": [ + { + "bbox": [ + 138, + 156, + 434, + 167 + ], + "type": "text", + "content": "2 Output: Optimized style token embedding " + }, + { + "bbox": [ + 138, + 156, + 434, + 167 + ], + "type": "inline_equation", + "content": "v^{*}" + }, + { + "bbox": [ + 138, + 156, + 434, + 167 + ], + "type": "text", + "content": " and diffusion model " + }, + { + "bbox": [ + 138, + 156, + 434, + 167 + ], + "type": "inline_equation", + "content": "\\varepsilon_{\\theta}^{*}" + }, + { + "bbox": [ + 138, + 156, + 434, + 167 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 147, + 173, + 318, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 173, + 318, + 183 + ], + "spans": [ + { + "bbox": [ + 147, + 173, + 318, + 183 + ], + "type": "text", + "content": "// Initial Stylized Image Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 184, + 467, + 250 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 138, + 184, + 467, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 467, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 467, + 205 + ], + "type": "text", + "content": "3 Compute the inversion noise " + }, + { + "bbox": [ + 138, + 184, + 467, + 205 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 138, + 184, + 467, + 205 + ], + "type": "text", + "content": " using DDIM inversion over " + }, + { + "bbox": [ + 138, + 184, + 467, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 138, + 184, + 467, + 205 + ], + "type": "text", + "content": " conditioned on " + }, + { + "bbox": [ + 138, + 184, + 467, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 138, + 184, + 467, + 205 + ], + "type": "text", + "content": " based on Eq. (6);" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 205, + 463, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 463, + 228 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 463, + 228 + ], + "type": "text", + "content": "4 Generate " + }, + { + "bbox": [ + 138, + 205, + 463, + 228 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 138, + 205, + 463, + 228 + ], + "type": "text", + "content": " images conditioned on a predefined prompt set based on Eq. (5); // Prompt Refinement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 228, + 463, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 463, + 250 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 463, + 250 + ], + "type": "text", + "content": "5 Select " + }, + { + "bbox": [ + 138, + 228, + 463, + 250 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 138, + 228, + 463, + 250 + ], + "type": "text", + "content": " images from " + }, + { + "bbox": [ + 138, + 228, + 463, + 250 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 138, + 228, + 463, + 250 + ], + "type": "text", + "content": " images with human feedback as the training dataset " + }, + { + "bbox": [ + 138, + 228, + 463, + 250 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_0, c)" + }, + { + "bbox": [ + 138, + 228, + 463, + 250 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 138, + 251, + 179, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 179, + 260 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 179, + 260 + ], + "type": "text", + "content": "6 repeat" + } + ] + } + ], + "index": 11 + }, + { + "type": "code", + "bbox": [ + 138, + 261, + 346, + 330 + ], + "blocks": [ + { + "bbox": [ + 140, + 119, + 265, + 130 + ], + "lines": [ + { + "bbox": [ + 140, + 119, + 265, + 130 + ], + "spans": [ + { + "bbox": [ + 140, + 119, + 265, + 130 + ], + "type": "text", + "content": "Algorithm 1: INSTASTYLE" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 138, + 261, + 346, + 330 + ], + "lines": [ + { + "bbox": [ + 138, + 261, + 346, + 330 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 346, + 330 + ], + "type": "text", + "content": "7 " + }, + { + "bbox": [ + 138, + 261, + 346, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0,c\\sim q(\\mathbf{x}_0,c)" + }, + { + "bbox": [ + 138, + 261, + 346, + 330 + ], + "type": "text", + "content": " \n8 " + }, + { + "bbox": [ + 138, + 261, + 346, + 330 + ], + "type": "inline_equation", + "content": "z_{0} = \\mathcal{E}(\\mathbf{x}_{0})" + }, + { + "bbox": [ + 138, + 261, + 346, + 330 + ], + "type": "text", + "content": " \n9 " + }, + { + "bbox": [ + 138, + 261, + 346, + 330 + ], + "type": "inline_equation", + "content": "t\\sim \\mathrm{Uniform}(\\{1,\\dots ,T\\})" + }, + { + "bbox": [ + 138, + 261, + 346, + 330 + ], + "type": "text", + "content": " \n10 " + }, + { + "bbox": [ + 138, + 261, + 346, + 330 + ], + "type": "inline_equation", + "content": "\\varepsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 138, + 261, + 346, + 330 + ], + "type": "text", + "content": " \n11 " + }, + { + "bbox": [ + 138, + 261, + 346, + 330 + ], + "type": "inline_equation", + "content": "z_{t} = \\sqrt{\\alpha_{t}} z_{0} + \\sqrt{1 - \\alpha_{t}}\\varepsilon" + }, + { + "bbox": [ + 138, + 261, + 346, + 330 + ], + "type": "text", + "content": " \n12 Take gradient step on " + }, + { + "bbox": [ + 138, + 261, + 346, + 330 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta ,v}\\| \\varepsilon -\\varepsilon_{\\theta}(z_t,t,c)\\| _2^2;" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_body" + } + ], + "index": 12, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "bbox": [ + 134, + 331, + 216, + 352 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 134, + 331, + 216, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 331, + 216, + 341 + ], + "spans": [ + { + "bbox": [ + 134, + 331, + 216, + 341 + ], + "type": "text", + "content": "13 until converged;" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 341, + 209, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 341, + 209, + 352 + ], + "spans": [ + { + "bbox": [ + 134, + 341, + 209, + 352 + ], + "type": "text", + "content": "14 Return " + }, + { + "bbox": [ + 134, + 341, + 209, + 352 + ], + "type": "inline_equation", + "content": "v^{*}" + }, + { + "bbox": [ + 134, + 341, + 209, + 352 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 134, + 341, + 209, + 352 + ], + "type": "inline_equation", + "content": "\\varepsilon_{\\theta}^{*}" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 380, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 429 + ], + "type": "text", + "content": "on a prompt containing the learned style token. Then we sample from the inversion noise conditioned on a target prompt which contains the target content and the learned style token via DDIM. Specifically, the estimated noise is calculated based on the classifier-free guidance [17]:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 160, + 435, + 481, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 435, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 160, + 435, + 481, + 449 + ], + "type": "interline_equation", + "content": "\\tilde {\\varepsilon} _ {\\theta} \\left(z _ {t}, t, \\mathcal {C}, \\varnothing\\right) = \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\varnothing\\right) + w \\cdot \\left(\\varepsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C}\\right) - \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\varnothing\\right)\\right), \\tag {10}", + "image_path": "f22c242fdc2d516f9b807139f15315507be861476d9479667683bc3e67c29465.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 457, + 481, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 481, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 481, + 480 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 457, + 481, + 480 + ], + "type": "inline_equation", + "content": "\\varnothing = \\psi(\")" + }, + { + "bbox": [ + 130, + 457, + 481, + 480 + ], + "type": "text", + "content": " is the embedding of a null text. " + }, + { + "bbox": [ + 130, + 457, + 481, + 480 + ], + "type": "inline_equation", + "content": "\\varepsilon_{\\theta}(z_t, t, \\mathcal{C})" + }, + { + "bbox": [ + 130, + 457, + 481, + 480 + ], + "type": "text", + "content": " represents the conditional predictions. " + }, + { + "bbox": [ + 130, + 457, + 481, + 480 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 457, + 481, + 480 + ], + "type": "text", + "content": " is the guidance scale parameter." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 481, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 482, + 601 + ], + "type": "text", + "content": "Combination of two styles. To combine two styles, we learn style tokens for each style, i.e., “” and “”. Specifically, we use the selected images of both styles to constitute the training set. Then we jointly optimize the embeddings of the style tokens and fine-tune the key projection and value projection in the cross-attention block. During inference, we individually transform two reference images into their corresponding inversion noise " + }, + { + "bbox": [ + 130, + 481, + 482, + 601 + ], + "type": "inline_equation", + "content": "z_{t1} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 130, + 481, + 482, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 481, + 482, + 601 + ], + "type": "inline_equation", + "content": "z_{t2} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 130, + 481, + 482, + 601 + ], + "type": "text", + "content": " via DDIM inversion conditioned on a prompt containing their learned style token. As the style can be described by both the “style” noise and style token, we combine the styles from these two perspectives. For the “style” noise, the combined inversion noise " + }, + { + "bbox": [ + 130, + 481, + 482, + 601 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 481, + 482, + 601 + ], + "type": "text", + "content": " is obtained based on a masking strategy:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 238, + 609, + 481, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 609, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 238, + 609, + 481, + 622 + ], + "type": "interline_equation", + "content": "z _ {t} = (\\mathbf {1} - \\mathbf {M}) \\odot z _ {t 1} + \\mathbf {M} \\odot z _ {t 2}, \\tag {11}", + "image_path": "7f8e98e883d0cb2c6edd0c1d53136ca6157d8d06ca6e6fc84805e64d35b8cfb1.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": " is element-wise multiplication and " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{1}" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": " is a binary mask filled with ones. " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\{0, 1\\}^{H \\times W}" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": " denotes a random binary mask indicating where to drop out and fill in from two inversion noises. The noise mix ratio between two inversion" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "X. Cui et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": "noises is " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": ", representing the percent of " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": " that is set to 1 (the rest is set to 0). Finally, we utilize a composed guidance mechanism [38] to estimate the noise:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 186, + 147, + 481, + 193 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 147, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 186, + 147, + 481, + 193 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {\\varepsilon} _ {\\theta} \\left(z _ {t}, t, \\mathcal {C}, \\varnothing\\right) = \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\varnothing\\right) \\\\ + w \\cdot (1 - \\beta) \\cdot \\left(\\varepsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C} _ {1}\\right) - \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\varnothing\\right)\\right) \\tag {12} \\\\ + w \\cdot \\beta \\cdot \\left(\\varepsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {C} _ {2}\\right) - \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\varnothing\\right)\\right), \\\\ \\end{array}", + "image_path": "2c22ed11b6523b4048993f02c2a25fc192299db04ae85abde85aa893cde41f19.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 199, + 479, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 199, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 131, + 199, + 479, + 247 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 199, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 131, + 199, + 479, + 247 + ], + "type": "text", + "content": " is a prompt mix ratio. " + }, + { + "bbox": [ + 131, + 199, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_1" + }, + { + "bbox": [ + 131, + 199, + 479, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 199, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_2" + }, + { + "bbox": [ + 131, + 199, + 479, + 247 + ], + "type": "text", + "content": " are the embeddings of the target prompts for the two styles, respectively. Denote the target object as " + }, + { + "bbox": [ + 131, + 199, + 479, + 247 + ], + "type": "inline_equation", + "content": "<\\mathrm{obj}>" + }, + { + "bbox": [ + 131, + 199, + 479, + 247 + ], + "type": "text", + "content": ", they can be formulated as " + }, + { + "bbox": [ + 131, + 199, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_1 = \\psi(\"A < \\mathrm{obj}>" + }, + { + "bbox": [ + 131, + 199, + 479, + 247 + ], + "type": "text", + "content": " in the style of " + }, + { + "bbox": [ + 131, + 199, + 479, + 247 + ], + "type": "inline_equation", + "content": "<\\mathrm{style}1>" + }, + { + "bbox": [ + 131, + 199, + 479, + 247 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 131, + 199, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_2 = \\psi(\"A < \\mathrm{obj}>" + }, + { + "bbox": [ + 131, + 199, + 479, + 247 + ], + "type": "text", + "content": " in the style of " + }, + { + "bbox": [ + 131, + 199, + 479, + 247 + ], + "type": "inline_equation", + "content": "<\\mathrm{style}2>" + }, + { + "bbox": [ + 131, + 199, + 479, + 247 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 263, + 223, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 223, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 223, + 277 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 287, + 269, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 287, + 269, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 287, + 269, + 300 + ], + "type": "text", + "content": "4.1 Experimental Setting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 304, + 482, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 304, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 131, + 304, + 482, + 401 + ], + "type": "text", + "content": "We collect 60 images as the reference style dataset. The complete information on these images is listed in Tab. 2 in the Supplementary Material. Our method is implemented using PyTorch and executed on a single NVIDIA GeForce RTX 3090 GPU. The training procedure consists of 500 iterations, employing the Adam optimizer with a learning rate of " + }, + { + "bbox": [ + 131, + 304, + 482, + 401 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 131, + 304, + 482, + 401 + ], + "type": "text", + "content": ". During sampling, we set the guidance scale " + }, + { + "bbox": [ + 131, + 304, + 482, + 401 + ], + "type": "inline_equation", + "content": "w = 2.5" + }, + { + "bbox": [ + 131, + 304, + 482, + 401 + ], + "type": "text", + "content": ". The number of generated images " + }, + { + "bbox": [ + 131, + 304, + 482, + 401 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 304, + 482, + 401 + ], + "type": "text", + "content": " in the first stage is 15 whose details are shown in Sec. C.1 in the Supplementary Material. The number of selected images " + }, + { + "bbox": [ + 131, + 304, + 482, + 401 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 304, + 482, + 401 + ], + "type": "text", + "content": " in the second stage is set to 5." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 133, + 420, + 175, + 460 + ], + "blocks": [ + { + "bbox": [ + 133, + 420, + 175, + 460 + ], + "lines": [ + { + "bbox": [ + 133, + 420, + 175, + 460 + ], + "spans": [ + { + "bbox": [ + 133, + 420, + 175, + 460 + ], + "type": "image", + "image_path": "c80cd58cfac2a6c96bc4d5b69f6c86d483da87d9c4e1d3427d3057bcb40f2567.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 133, + 461, + 173, + 499 + ], + "blocks": [ + { + "bbox": [ + 133, + 461, + 173, + 499 + ], + "lines": [ + { + "bbox": [ + 133, + 461, + 173, + 499 + ], + "spans": [ + { + "bbox": [ + 133, + 461, + 173, + 499 + ], + "type": "image", + "image_path": "62bb4229f8af2972e988e022ba0d231ef3ad458bc58eb5fedee8e8358b7441be.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 133, + 501, + 173, + 538 + ], + "blocks": [ + { + "bbox": [ + 133, + 501, + 173, + 538 + ], + "lines": [ + { + "bbox": [ + 133, + 501, + 173, + 538 + ], + "spans": [ + { + "bbox": [ + 133, + 501, + 173, + 538 + ], + "type": "image", + "image_path": "bd86b46287db8e57f5d00d11245aa8a8a3df65f210ac82d2bf536f74c6df529a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 133, + 539, + 173, + 578 + ], + "blocks": [ + { + "bbox": [ + 133, + 539, + 173, + 578 + ], + "lines": [ + { + "bbox": [ + 133, + 539, + 173, + 578 + ], + "spans": [ + { + "bbox": [ + 133, + 539, + 173, + 578 + ], + "type": "image", + "image_path": "f76010f551154aba616b502ebb7632c7ed8a8dc21e847c25fe407ec728079064.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 579, + 168, + 587 + ], + "lines": [ + { + "bbox": [ + 138, + 579, + 168, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 168, + 587 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 184, + 421, + 241, + 460 + ], + "blocks": [ + { + "bbox": [ + 184, + 421, + 241, + 460 + ], + "lines": [ + { + "bbox": [ + 184, + 421, + 241, + 460 + ], + "spans": [ + { + "bbox": [ + 184, + 421, + 241, + 460 + ], + "type": "image", + "image_path": "7da6a41e381c596b65db7c7f3decaf9953cc2a7dd0161b4941397773a727d28a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 184, + 461, + 241, + 499 + ], + "blocks": [ + { + "bbox": [ + 184, + 461, + 241, + 499 + ], + "lines": [ + { + "bbox": [ + 184, + 461, + 241, + 499 + ], + "spans": [ + { + "bbox": [ + 184, + 461, + 241, + 499 + ], + "type": "image", + "image_path": "ad889a0748d7d4b91bc4e81feab077c629a9fff69edfcd9d34d3c23ae55d6273.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 184, + 501, + 241, + 539 + ], + "blocks": [ + { + "bbox": [ + 184, + 501, + 241, + 539 + ], + "lines": [ + { + "bbox": [ + 184, + 501, + 241, + 539 + ], + "spans": [ + { + "bbox": [ + 184, + 501, + 241, + 539 + ], + "type": "image", + "image_path": "ec16abb08643cb1c6a287e592936c53702e6390106467c49c48455a9c122504f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 184, + 539, + 242, + 578 + ], + "blocks": [ + { + "bbox": [ + 184, + 539, + 242, + 578 + ], + "lines": [ + { + "bbox": [ + 184, + 539, + 242, + 578 + ], + "spans": [ + { + "bbox": [ + 184, + 539, + 242, + 578 + ], + "type": "image", + "image_path": "a4f917a9ca3a35ce6227ed953403d9a6a49b24239d416c7cadfd7b3ae105b922.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 579, + 242, + 587 + ], + "lines": [ + { + "bbox": [ + 184, + 579, + 242, + 587 + ], + "spans": [ + { + "bbox": [ + 184, + 579, + 242, + 587 + ], + "type": "text", + "content": "(a)InstaStyle(Ours)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 244, + 421, + 301, + 460 + ], + "blocks": [ + { + "bbox": [ + 244, + 421, + 301, + 460 + ], + "lines": [ + { + "bbox": [ + 244, + 421, + 301, + 460 + ], + "spans": [ + { + "bbox": [ + 244, + 421, + 301, + 460 + ], + "type": "image", + "image_path": "820c13bc6a331fdd56672ae44865209d89456ea7e66cc8269458d290bf5066b3.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 244, + 461, + 301, + 499 + ], + "blocks": [ + { + "bbox": [ + 244, + 461, + 301, + 499 + ], + "lines": [ + { + "bbox": [ + 244, + 461, + 301, + 499 + ], + "spans": [ + { + "bbox": [ + 244, + 461, + 301, + 499 + ], + "type": "image", + "image_path": "f4ea39848124db46872acc74bb26bef845260cfadd46797d37f9938aa2eb0ba4.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 244, + 500, + 301, + 539 + ], + "blocks": [ + { + "bbox": [ + 244, + 500, + 301, + 539 + ], + "lines": [ + { + "bbox": [ + 244, + 500, + 301, + 539 + ], + "spans": [ + { + "bbox": [ + 244, + 500, + 301, + 539 + ], + "type": "image", + "image_path": "f30cfcc739e5c8d127bd62253d7bf6fd9854cea2b82a4bef23f836be27277f2a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 244, + 539, + 301, + 578 + ], + "blocks": [ + { + "bbox": [ + 244, + 539, + 301, + 578 + ], + "lines": [ + { + "bbox": [ + 244, + 539, + 301, + 578 + ], + "spans": [ + { + "bbox": [ + 244, + 539, + 301, + 578 + ], + "type": "image", + "image_path": "55ceaa9fe7cd17f7c80fc91b8bafb11df3a9528725f082866ec54ca2184e4fcc.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 579, + 292, + 587 + ], + "lines": [ + { + "bbox": [ + 252, + 579, + 292, + 587 + ], + "spans": [ + { + "bbox": [ + 252, + 579, + 292, + 587 + ], + "type": "text", + "content": "(b) StyleDrop" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 303, + 421, + 359, + 460 + ], + "blocks": [ + { + "bbox": [ + 303, + 421, + 359, + 460 + ], + "lines": [ + { + "bbox": [ + 303, + 421, + 359, + 460 + ], + "spans": [ + { + "bbox": [ + 303, + 421, + 359, + 460 + ], + "type": "image", + "image_path": "a44d63fc11adf0886ed91375a8e74f7fae3953358ce775aa66cefe1ac9d7695e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 303, + 461, + 359, + 499 + ], + "blocks": [ + { + "bbox": [ + 303, + 461, + 359, + 499 + ], + "lines": [ + { + "bbox": [ + 303, + 461, + 359, + 499 + ], + "spans": [ + { + "bbox": [ + 303, + 461, + 359, + 499 + ], + "type": "image", + "image_path": "4674897d31eeb69fbacd88345437a172ba213fbd74048934bf294f05392867fb.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 303, + 500, + 359, + 539 + ], + "blocks": [ + { + "bbox": [ + 303, + 500, + 359, + 539 + ], + "lines": [ + { + "bbox": [ + 303, + 500, + 359, + 539 + ], + "spans": [ + { + "bbox": [ + 303, + 500, + 359, + 539 + ], + "type": "image", + "image_path": "0107c43360cc799d70b54fe3f4e7e4e8b60c9f736ad4ba37ad6fb88f2ea1bc34.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 303, + 539, + 359, + 578 + ], + "blocks": [ + { + "bbox": [ + 303, + 539, + 359, + 578 + ], + "lines": [ + { + "bbox": [ + 303, + 539, + 359, + 578 + ], + "spans": [ + { + "bbox": [ + 303, + 539, + 359, + 578 + ], + "type": "image", + "image_path": "cedd09f2c56d58e9316e4c472b9610fbbd96ec4b0292eeb9a36619ac9f8bb852.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 579, + 354, + 586 + ], + "lines": [ + { + "bbox": [ + 307, + 579, + 354, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 354, + 586 + ], + "type": "text", + "content": "(c) DreamBooth" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 607, + 481, + 641 + ], + "lines": [ + { + "bbox": [ + 131, + 607, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 131, + 607, + 481, + 641 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparison of stylized image generation on various styles. Objects for synthesis are Bicycle, Sunflowers and Chair. Our method excels at capturing fine-grained style information, such as color, textures, and brushstrokes." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 362, + 421, + 419, + 460 + ], + "blocks": [ + { + "bbox": [ + 362, + 421, + 419, + 460 + ], + "lines": [ + { + "bbox": [ + 362, + 421, + 419, + 460 + ], + "spans": [ + { + "bbox": [ + 362, + 421, + 419, + 460 + ], + "type": "image", + "image_path": "002147886c24b0ab81221df2673671661e13f714c9273a98d253f706318af2ce.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 362, + 461, + 419, + 499 + ], + "blocks": [ + { + "bbox": [ + 362, + 461, + 419, + 499 + ], + "lines": [ + { + "bbox": [ + 362, + 461, + 419, + 499 + ], + "spans": [ + { + "bbox": [ + 362, + 461, + 419, + 499 + ], + "type": "image", + "image_path": "653565d4528b7c0528cf9491e47b6acdccc83b98af0dc302b7a61fafab4ca3fe.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 362, + 500, + 419, + 539 + ], + "blocks": [ + { + "bbox": [ + 362, + 500, + 419, + 539 + ], + "lines": [ + { + "bbox": [ + 362, + 500, + 419, + 539 + ], + "spans": [ + { + "bbox": [ + 362, + 500, + 419, + 539 + ], + "type": "image", + "image_path": "6983507207ecd7191875b40ee1de1704c34080d70e3b1344588e9ce3ba303710.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 362, + 539, + 419, + 578 + ], + "blocks": [ + { + "bbox": [ + 362, + 539, + 419, + 578 + ], + "lines": [ + { + "bbox": [ + 362, + 539, + 419, + 578 + ], + "spans": [ + { + "bbox": [ + 362, + 539, + 419, + 578 + ], + "type": "image", + "image_path": "83b87eac91516065c091178a75e7d91cba991330dfc8b756dcd910b4ccddb590.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 579, + 419, + 587 + ], + "lines": [ + { + "bbox": [ + 362, + 579, + 419, + 587 + ], + "spans": [ + { + "bbox": [ + 362, + 579, + 419, + 587 + ], + "type": "text", + "content": "(d) Custom Diffusion" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 422, + 421, + 479, + 460 + ], + "blocks": [ + { + "bbox": [ + 422, + 421, + 479, + 460 + ], + "lines": [ + { + "bbox": [ + 422, + 421, + 479, + 460 + ], + "spans": [ + { + "bbox": [ + 422, + 421, + 479, + 460 + ], + "type": "image", + "image_path": "40a69f8153fe9d2dbf1fbe2c7145211952e9bf6c2a33c76174eb41e0829690d7.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 422, + 461, + 479, + 499 + ], + "blocks": [ + { + "bbox": [ + 422, + 461, + 479, + 499 + ], + "lines": [ + { + "bbox": [ + 422, + 461, + 479, + 499 + ], + "spans": [ + { + "bbox": [ + 422, + 461, + 479, + 499 + ], + "type": "image", + "image_path": "88421700aef09d80582a937676a89610ad3f6f07859c37df56059aa7ebbfea5e.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 422, + 500, + 479, + 539 + ], + "blocks": [ + { + "bbox": [ + 422, + 500, + 479, + 539 + ], + "lines": [ + { + "bbox": [ + 422, + 500, + 479, + 539 + ], + "spans": [ + { + "bbox": [ + 422, + 500, + 479, + 539 + ], + "type": "image", + "image_path": "79e448d2d647579d093015077ce84c9df2da00dec33a313e6c9974613d4136ee.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 422, + 539, + 479, + 578 + ], + "blocks": [ + { + "bbox": [ + 422, + 539, + 479, + 578 + ], + "lines": [ + { + "bbox": [ + 422, + 539, + 479, + 578 + ], + "spans": [ + { + "bbox": [ + 422, + 539, + 479, + 578 + ], + "type": "image", + "image_path": "d5b43c83b2a50bb2f1d961f2fb79cdcb39838166256d1b4b070d8709d0b2ca67.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 579, + 480, + 587 + ], + "lines": [ + { + "bbox": [ + 422, + 579, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 422, + 579, + 480, + 587 + ], + "type": "text", + "content": "(e) Textual Inversion" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "INSTASTYLE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 176, + 156 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 176, + 156 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 176, + 156 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 176, + 156 + ], + "type": "image", + "image_path": "93a947d27a8f62ea37793f74e43126f818fb863980e76daf4c312e7d6151b1a3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 158, + 175, + 197 + ], + "blocks": [ + { + "bbox": [ + 133, + 158, + 175, + 197 + ], + "lines": [ + { + "bbox": [ + 133, + 158, + 175, + 197 + ], + "spans": [ + { + "bbox": [ + 133, + 158, + 175, + 197 + ], + "type": "image", + "image_path": "e723e52182e240f95fbbe1b7a875a9b19283d0e8e74851d803ff4d831ec3d572.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 198, + 175, + 239 + ], + "blocks": [ + { + "bbox": [ + 133, + 198, + 175, + 239 + ], + "lines": [ + { + "bbox": [ + 133, + 198, + 175, + 239 + ], + "spans": [ + { + "bbox": [ + 133, + 198, + 175, + 239 + ], + "type": "image", + "image_path": "49116d4c68b22b28a783b598b4d337418063adbb1b2519c0a2bf9d78da5f3f88.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 133, + 240, + 175, + 279 + ], + "blocks": [ + { + "bbox": [ + 133, + 240, + 175, + 279 + ], + "lines": [ + { + "bbox": [ + 133, + 240, + 175, + 279 + ], + "spans": [ + { + "bbox": [ + 133, + 240, + 175, + 279 + ], + "type": "image", + "image_path": "6c1a1cdaf0f6f75b8d8923bfa18fd5b89f2d51c76d7d7dc52e08972b08ca6aa6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 280, + 175, + 322 + ], + "blocks": [ + { + "bbox": [ + 133, + 280, + 175, + 322 + ], + "lines": [ + { + "bbox": [ + 133, + 280, + 175, + 322 + ], + "spans": [ + { + "bbox": [ + 133, + 280, + 175, + 322 + ], + "type": "image", + "image_path": "a8d3e58b66bc7a859d63d2473cbdf9c7094fcc83fa576d996983b959fa348831.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 323, + 170, + 330 + ], + "lines": [ + { + "bbox": [ + 140, + 323, + 170, + 330 + ], + "spans": [ + { + "bbox": [ + 140, + 323, + 170, + 330 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 186, + 115, + 227, + 147 + ], + "blocks": [ + { + "bbox": [ + 186, + 115, + 227, + 147 + ], + "lines": [ + { + "bbox": [ + 186, + 115, + 227, + 147 + ], + "spans": [ + { + "bbox": [ + 186, + 115, + 227, + 147 + ], + "type": "image", + "image_path": "6308999419633e125169ded48e8b7eff36ed76541bc56a75a94058289b96d5f0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 148, + 222, + 154 + ], + "lines": [ + { + "bbox": [ + 194, + 148, + 222, + 154 + ], + "spans": [ + { + "bbox": [ + 194, + 148, + 222, + 154 + ], + "type": "text", + "content": "Elephant" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 186, + 158, + 227, + 190 + ], + "blocks": [ + { + "bbox": [ + 186, + 158, + 227, + 190 + ], + "lines": [ + { + "bbox": [ + 186, + 158, + 227, + 190 + ], + "spans": [ + { + "bbox": [ + 186, + 158, + 227, + 190 + ], + "type": "image", + "image_path": "edc3f3da5f4238def5eac7a6ed55fdab14f7825f302830fd0521a3e20a4fd899.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 185, + 191, + 216, + 197 + ], + "lines": [ + { + "bbox": [ + 185, + 191, + 216, + 197 + ], + "spans": [ + { + "bbox": [ + 185, + 191, + 216, + 197 + ], + "type": "text", + "content": "Castle" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 186, + 198, + 227, + 236 + ], + "blocks": [ + { + "bbox": [ + 186, + 198, + 227, + 236 + ], + "lines": [ + { + "bbox": [ + 186, + 198, + 227, + 236 + ], + "spans": [ + { + "bbox": [ + 186, + 198, + 227, + 236 + ], + "type": "image", + "image_path": "dd8c65a113b26c9d1a7ab6c21cd6573ca34f853337ef4a1d7a5000375691bd2a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 185, + 239, + 227, + 279 + ], + "blocks": [ + { + "bbox": [ + 185, + 239, + 227, + 279 + ], + "lines": [ + { + "bbox": [ + 185, + 239, + 227, + 279 + ], + "spans": [ + { + "bbox": [ + 185, + 239, + 227, + 279 + ], + "type": "image", + "image_path": "de7afc33523f9555c9b8ce501cbfdd3109d3fbc90863b3439ab6346959c63a42.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 186, + 280, + 227, + 321 + ], + "blocks": [ + { + "bbox": [ + 186, + 280, + 227, + 321 + ], + "lines": [ + { + "bbox": [ + 186, + 280, + 227, + 321 + ], + "spans": [ + { + "bbox": [ + 186, + 280, + 227, + 321 + ], + "type": "image", + "image_path": "d7c07f514c16acc211d896020f09eae1047f00ab72e06c84dd7ccecb71e2a852.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 323, + 264, + 330 + ], + "lines": [ + { + "bbox": [ + 182, + 323, + 264, + 330 + ], + "spans": [ + { + "bbox": [ + 182, + 323, + 264, + 330 + ], + "type": "text", + "content": "InstaStyle (Ours) AsePA-Net" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 229, + 115, + 270, + 156 + ], + "blocks": [ + { + "bbox": [ + 229, + 115, + 270, + 156 + ], + "lines": [ + { + "bbox": [ + 229, + 115, + 270, + 156 + ], + "spans": [ + { + "bbox": [ + 229, + 115, + 270, + 156 + ], + "type": "image", + "image_path": "1f685ed8d931e04fcb720f28c5e00c1b0b51feb446af776d043f685b254cc413.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 229, + 157, + 269, + 198 + ], + "blocks": [ + { + "bbox": [ + 229, + 157, + 269, + 198 + ], + "lines": [ + { + "bbox": [ + 229, + 157, + 269, + 198 + ], + "spans": [ + { + "bbox": [ + 229, + 157, + 269, + 198 + ], + "type": "image", + "image_path": "330ec054cd93cdd9ed17ca7eb53fd48bddd2eef517b06bc473b1e77534e8c038.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 229, + 198, + 269, + 238 + ], + "blocks": [ + { + "bbox": [ + 229, + 198, + 269, + 238 + ], + "lines": [ + { + "bbox": [ + 229, + 198, + 269, + 238 + ], + "spans": [ + { + "bbox": [ + 229, + 198, + 269, + 238 + ], + "type": "image", + "image_path": "ef77012026b7637a45a696c5a90362734f2fbf4bd85a431bdbf604570b051001.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 229, + 239, + 269, + 279 + ], + "blocks": [ + { + "bbox": [ + 229, + 239, + 269, + 279 + ], + "lines": [ + { + "bbox": [ + 229, + 239, + 269, + 279 + ], + "spans": [ + { + "bbox": [ + 229, + 239, + 269, + 279 + ], + "type": "image", + "image_path": "206cfea526225d6930e12f88a1c78d4df1041164ce2bd67524b1ba578a9e2a94.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 229, + 280, + 269, + 321 + ], + "blocks": [ + { + "bbox": [ + 229, + 280, + 269, + 321 + ], + "lines": [ + { + "bbox": [ + 229, + 280, + 269, + 321 + ], + "spans": [ + { + "bbox": [ + 229, + 280, + 269, + 321 + ], + "type": "image", + "image_path": "d10c581de5238462787cc8f7fff64353771ac95d9828003356f85a3eb8091a98.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 232, + 323, + 264, + 330 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 323, + 264, + 330 + ], + "spans": [ + { + "bbox": [ + 232, + 323, + 264, + 330 + ], + "type": "text", + "content": "AsePA-Net" + } + ] + } + ], + "index": 21, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 271, + 115, + 311, + 156 + ], + "blocks": [ + { + "bbox": [ + 271, + 115, + 311, + 156 + ], + "lines": [ + { + "bbox": [ + 271, + 115, + 311, + 156 + ], + "spans": [ + { + "bbox": [ + 271, + 115, + 311, + 156 + ], + "type": "image", + "image_path": "73463e740a620daeeb15fcfe9c403a856ba19fd6a67bb97c6da8d3d443e15c3e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 271, + 157, + 310, + 198 + ], + "blocks": [ + { + "bbox": [ + 271, + 157, + 310, + 198 + ], + "lines": [ + { + "bbox": [ + 271, + 157, + 310, + 198 + ], + "spans": [ + { + "bbox": [ + 271, + 157, + 310, + 198 + ], + "type": "image", + "image_path": "95bfccb1ea87ba82468c96b26daec57cc902ad2b653de2bcfeef06fd6fa3ecd4.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 271, + 198, + 310, + 238 + ], + "blocks": [ + { + "bbox": [ + 271, + 198, + 310, + 238 + ], + "lines": [ + { + "bbox": [ + 271, + 198, + 310, + 238 + ], + "spans": [ + { + "bbox": [ + 271, + 198, + 310, + 238 + ], + "type": "image", + "image_path": "35d60c178be4da672a7f060d94fa71e8479354998dd926f364096d7bbe1c0838.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 271, + 239, + 310, + 279 + ], + "blocks": [ + { + "bbox": [ + 271, + 239, + 310, + 279 + ], + "lines": [ + { + "bbox": [ + 271, + 239, + 310, + 279 + ], + "spans": [ + { + "bbox": [ + 271, + 239, + 310, + 279 + ], + "type": "image", + "image_path": "cbfcc14075106e7f94a26eb32d48e078074cd840d83d012f91adf7a93a083600.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 271, + 280, + 310, + 321 + ], + "blocks": [ + { + "bbox": [ + 271, + 280, + 310, + 321 + ], + "lines": [ + { + "bbox": [ + 271, + 280, + 310, + 321 + ], + "spans": [ + { + "bbox": [ + 271, + 280, + 310, + 321 + ], + "type": "image", + "image_path": "d7180e1caddb0d3a86177ba9e3e5d782cb58390549923bd61eaf4ac902533446.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 323, + 306, + 330 + ], + "lines": [ + { + "bbox": [ + 276, + 323, + 306, + 330 + ], + "spans": [ + { + "bbox": [ + 276, + 323, + 306, + 330 + ], + "type": "text", + "content": "QuantArt" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 312, + 115, + 353, + 156 + ], + "blocks": [ + { + "bbox": [ + 312, + 115, + 353, + 156 + ], + "lines": [ + { + "bbox": [ + 312, + 115, + 353, + 156 + ], + "spans": [ + { + "bbox": [ + 312, + 115, + 353, + 156 + ], + "type": "image", + "image_path": "05598e7175a982d5487822da50fa43e462e996653c5f4523c0726f50f9d32d62.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 312, + 157, + 353, + 198 + ], + "blocks": [ + { + "bbox": [ + 312, + 157, + 353, + 198 + ], + "lines": [ + { + "bbox": [ + 312, + 157, + 353, + 198 + ], + "spans": [ + { + "bbox": [ + 312, + 157, + 353, + 198 + ], + "type": "image", + "image_path": "0c9d46d1c3d0bcf4267c7bd4f59bbb1cc89b6f4ff181b89666d1784bd8fc38dc.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 312, + 198, + 353, + 238 + ], + "blocks": [ + { + "bbox": [ + 312, + 198, + 353, + 238 + ], + "lines": [ + { + "bbox": [ + 312, + 198, + 353, + 238 + ], + "spans": [ + { + "bbox": [ + 312, + 198, + 353, + 238 + ], + "type": "image", + "image_path": "e264bcef2f9453325657db6676fb1edfdc9e33fc03bf384f5a8ea86c94ee7706.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 312, + 239, + 353, + 279 + ], + "blocks": [ + { + "bbox": [ + 312, + 239, + 353, + 279 + ], + "lines": [ + { + "bbox": [ + 312, + 239, + 353, + 279 + ], + "spans": [ + { + "bbox": [ + 312, + 239, + 353, + 279 + ], + "type": "image", + "image_path": "1476c31bfcbc733e327d103cc8148e8bbc300b1968115dddab1c68d7bf80e6f0.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 312, + 280, + 353, + 321 + ], + "blocks": [ + { + "bbox": [ + 312, + 280, + 353, + 321 + ], + "lines": [ + { + "bbox": [ + 312, + 280, + 353, + 321 + ], + "spans": [ + { + "bbox": [ + 312, + 280, + 353, + 321 + ], + "type": "image", + "image_path": "4b93159d2c873695fdb64eed844b3e72c22c4070dca0151e1897340a52c97462.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 323, + 343, + 330 + ], + "lines": [ + { + "bbox": [ + 323, + 323, + 343, + 330 + ], + "spans": [ + { + "bbox": [ + 323, + 323, + 343, + 330 + ], + "type": "text", + "content": "StyTr2" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 349, + 482, + 393 + ], + "lines": [ + { + "bbox": [ + 132, + 349, + 482, + 393 + ], + "spans": [ + { + "bbox": [ + 132, + 349, + 482, + 393 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparison with style transfer methods. For style transfer methods, content images are employed (shown in the bottom right). Despite our method relying solely on text for content, we achieve comparable performance in the fidelity of the content. Furthermore, we excel in preserving style details." + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 354, + 115, + 395, + 156 + ], + "blocks": [ + { + "bbox": [ + 354, + 115, + 395, + 156 + ], + "lines": [ + { + "bbox": [ + 354, + 115, + 395, + 156 + ], + "spans": [ + { + "bbox": [ + 354, + 115, + 395, + 156 + ], + "type": "image", + "image_path": "e41ca3e62ebe14a9808eafe31303259a8263db9d06de58c81dded567f3601f4f.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 354, + 157, + 395, + 198 + ], + "blocks": [ + { + "bbox": [ + 354, + 157, + 395, + 198 + ], + "lines": [ + { + "bbox": [ + 354, + 157, + 395, + 198 + ], + "spans": [ + { + "bbox": [ + 354, + 157, + 395, + 198 + ], + "type": "image", + "image_path": "34009c2c215c55995462b4dc78ce439822f400a8f47f02ecc3754b8173f441ae.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 354, + 198, + 395, + 238 + ], + "blocks": [ + { + "bbox": [ + 354, + 198, + 395, + 238 + ], + "lines": [ + { + "bbox": [ + 354, + 198, + 395, + 238 + ], + "spans": [ + { + "bbox": [ + 354, + 198, + 395, + 238 + ], + "type": "image", + "image_path": "1fd7c27b9d3cedce90590f2b1b794c1f6ebdb3bb91f09bbe58e630c76a928466.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 354, + 239, + 395, + 279 + ], + "blocks": [ + { + "bbox": [ + 354, + 239, + 395, + 279 + ], + "lines": [ + { + "bbox": [ + 354, + 239, + 395, + 279 + ], + "spans": [ + { + "bbox": [ + 354, + 239, + 395, + 279 + ], + "type": "image", + "image_path": "8fdcbdce0135afb31df927f694a23fd3910a085c92de7df6dc37827d04698ab6.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 354, + 280, + 395, + 321 + ], + "blocks": [ + { + "bbox": [ + 354, + 280, + 395, + 321 + ], + "lines": [ + { + "bbox": [ + 354, + 280, + 395, + 321 + ], + "spans": [ + { + "bbox": [ + 354, + 280, + 395, + 321 + ], + "type": "image", + "image_path": "3487069969ff08a784ccbb60e4c46f118026d63ab2069534411f7daec87421b4.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 323, + 384, + 330 + ], + "lines": [ + { + "bbox": [ + 366, + 323, + 384, + 330 + ], + "spans": [ + { + "bbox": [ + 366, + 323, + 384, + 330 + ], + "type": "text", + "content": "CCPL" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 397, + 115, + 438, + 156 + ], + "blocks": [ + { + "bbox": [ + 397, + 115, + 438, + 156 + ], + "lines": [ + { + "bbox": [ + 397, + 115, + 438, + 156 + ], + "spans": [ + { + "bbox": [ + 397, + 115, + 438, + 156 + ], + "type": "image", + "image_path": "987a6b0916c64ccabaa92e53be4572161cd6d477426dc4baa3004fb4e22a8d70.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 397, + 157, + 438, + 198 + ], + "blocks": [ + { + "bbox": [ + 397, + 157, + 438, + 198 + ], + "lines": [ + { + "bbox": [ + 397, + 157, + 438, + 198 + ], + "spans": [ + { + "bbox": [ + 397, + 157, + 438, + 198 + ], + "type": "image", + "image_path": "045f44cf9b34091bb737841c03f6264b59476797cef469c95c02a9fd0615060d.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 397, + 198, + 438, + 238 + ], + "blocks": [ + { + "bbox": [ + 397, + 198, + 438, + 238 + ], + "lines": [ + { + "bbox": [ + 397, + 198, + 438, + 238 + ], + "spans": [ + { + "bbox": [ + 397, + 198, + 438, + 238 + ], + "type": "image", + "image_path": "02fba50679b86a7063d59cd655d216a36378a742a971c229f1341e5ac30046a6.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 397, + 239, + 438, + 279 + ], + "blocks": [ + { + "bbox": [ + 397, + 239, + 438, + 279 + ], + "lines": [ + { + "bbox": [ + 397, + 239, + 438, + 279 + ], + "spans": [ + { + "bbox": [ + 397, + 239, + 438, + 279 + ], + "type": "image", + "image_path": "48e2922a4116385d9a72fd20ddb187d0dce8e687c4feb90b5859d93d60c78ee6.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 397, + 280, + 438, + 321 + ], + "blocks": [ + { + "bbox": [ + 397, + 280, + 438, + 321 + ], + "lines": [ + { + "bbox": [ + 397, + 280, + 438, + 321 + ], + "spans": [ + { + "bbox": [ + 397, + 280, + 438, + 321 + ], + "type": "image", + "image_path": "bfb1dcd8741518b7e4e125fd298e9295f0816ec4af33087a11432480483cff82.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 323, + 427, + 330 + ], + "lines": [ + { + "bbox": [ + 408, + 323, + 427, + 330 + ], + "spans": [ + { + "bbox": [ + 408, + 323, + 427, + 330 + ], + "type": "text", + "content": "CAST" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 440, + 115, + 480, + 156 + ], + "blocks": [ + { + "bbox": [ + 440, + 115, + 480, + 156 + ], + "lines": [ + { + "bbox": [ + 440, + 115, + 480, + 156 + ], + "spans": [ + { + "bbox": [ + 440, + 115, + 480, + 156 + ], + "type": "image", + "image_path": "b45d1cc9e7ae0b014b3310e82ec8e6e580f41351184793f9912aa37b1e52385c.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 440, + 158, + 480, + 198 + ], + "blocks": [ + { + "bbox": [ + 440, + 158, + 480, + 198 + ], + "lines": [ + { + "bbox": [ + 440, + 158, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 440, + 158, + 480, + 198 + ], + "type": "image", + "image_path": "b0c9100974743ffa73e09a8c2b364963320cdada2537265b94bb012071077003.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 440, + 198, + 480, + 238 + ], + "blocks": [ + { + "bbox": [ + 440, + 198, + 480, + 238 + ], + "lines": [ + { + "bbox": [ + 440, + 198, + 480, + 238 + ], + "spans": [ + { + "bbox": [ + 440, + 198, + 480, + 238 + ], + "type": "image", + "image_path": "c88c1c8132436e21e819713e6a68af7e3b7ca7f87a3da6c7d55f384e05abaa90.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 440, + 239, + 480, + 279 + ], + "blocks": [ + { + "bbox": [ + 440, + 239, + 480, + 279 + ], + "lines": [ + { + "bbox": [ + 440, + 239, + 480, + 279 + ], + "spans": [ + { + "bbox": [ + 440, + 239, + 480, + 279 + ], + "type": "image", + "image_path": "026dc34338163e68aa922611d253adf77f91015dfa3e6afaaf49b935b3859adb.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 440, + 280, + 480, + 321 + ], + "blocks": [ + { + "bbox": [ + 440, + 280, + 480, + 321 + ], + "lines": [ + { + "bbox": [ + 440, + 280, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 440, + 280, + 480, + 321 + ], + "type": "image", + "image_path": "cc3dfe80fe84d7125c55048fa26b61828e98d962b7edeca00acce1a79438501d.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 323, + 472, + 330 + ], + "lines": [ + { + "bbox": [ + 446, + 323, + 472, + 330 + ], + "spans": [ + { + "bbox": [ + 446, + 323, + 472, + 330 + ], + "type": "text", + "content": "AdaAttn" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "bbox": [ + 132, + 418, + 285, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 285, + 431 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 285, + 431 + ], + "type": "text", + "content": "4.2 Stylized Image Synthesis" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 130, + 441, + 482, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 573 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 573 + ], + "type": "text", + "content": "We conduct comparisons of INSTASTYLE against four recent methods following the open-sourced code" + }, + { + "bbox": [ + 130, + 441, + 482, + 573 + ], + "type": "inline_equation", + "content": "^{4,5}" + }, + { + "bbox": [ + 130, + 441, + 482, + 573 + ], + "type": "text", + "content": ", i.e., StyleDrop [52], DreamBooth [48], Custom Diffusion [31], and Textual Inversion [10]. Besides, following [53], we also make a comparison with style transfer methods [6,19,21,39,66,72] to further illustrate the superiority of our approach. Notably, the style transfer task takes images as content input, in contrast to our task which only utilizes text prompts for content. To make a comparison between the two tasks, we additionally generate content images based on text prompts for style transfer methods. Although we can compare the two tasks under this setting, the comparison is unfair to our approach because the content images utilized by style transfer methods may provide additional content information compared to our text prompts." + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 130, + 574, + 482, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 634 + ], + "type": "text", + "content": "Qualitative results. As shown in Fig. 4, textual Inversion and Custom Diffusion often yield an unsatisfactory style. Taking the well-known Van Gogh painting style (e.g., row 3 and row 4) as an example, they fail to capture fine-grained style information in the reference image. The fine-tuning-based methods, i.e., DreamBooth and StyleDrop, result in distorted content. For example, they gen" + } + ] + } + ], + "index": 55 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "X. Cui et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 358, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 358, + 654 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 358, + 654 + ], + "type": "text", + "content": "4 https://github.com/aim-uofa/StyleDrop-PyTorch" + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 135, + 654, + 332, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 654, + 332, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 654, + 332, + 665 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 135, + 654, + 332, + 665 + ], + "type": "text", + "content": " https://github.com/huggingface/diffusers" + } + ] + } + ], + "index": 57 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 186, + 298, + 297 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 1: (a) Quantitative comparison regarding the style consistency score and content fidelity score. (b) User study. The results show the percentage of votes where the comparison method is preferred over ours, ties with ours and is inferior to ours. We use the abbreviation (ST) to represent approaches for style transfer tasks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 214, + 167, + 227, + 177 + ], + "lines": [ + { + "bbox": [ + 214, + 167, + 227, + 177 + ], + "spans": [ + { + "bbox": [ + 214, + 167, + 227, + 177 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 186, + 298, + 297 + ], + "lines": [ + { + "bbox": [ + 143, + 186, + 298, + 297 + ], + "spans": [ + { + "bbox": [ + 143, + 186, + 298, + 297 + ], + "type": "table", + "html": "
MethodStyle (↑)Content (↑)
(ST) AdaAttN [39]0.5870.287
(ST) CCPL [66]0.5420.286
(ST) StyTr2 [6]0.5600.282
(ST) CAST [72]0.5940.285
(ST) QuantArt [21]0.5540.289
(ST) AesPA-Net [19]0.6040.284
Textual Inversion [10]0.5620.278
Custom Diffusion [31]0.6110.280
DreamBooth [48]0.5610.244
StyleDrop [52]0.6020.248
Ours (INSTSTYLE)0.6550.294
", + "image_path": "b97c23e07180451ef7bc9a887f4f418f38fe01f180ef884b7ec12315e3f35978.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 316, + 186, + 470, + 282 + ], + "blocks": [ + { + "bbox": [ + 387, + 167, + 400, + 177 + ], + "lines": [ + { + "bbox": [ + 387, + 167, + 400, + 177 + ], + "spans": [ + { + "bbox": [ + 387, + 167, + 400, + 177 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 186, + 470, + 282 + ], + "lines": [ + { + "bbox": [ + 316, + 186, + 470, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 470, + 282 + ], + "type": "table", + "html": "
MethodPreferenceTieOurs
(ST) AdaAttN [39]0.260.130.61
(ST) CCPL [66]0.300.230.47
(ST) StyTr2 [6]0.270.200.53
(ST) CAST [72]0.310.260.43
(ST) QuantArt [21]0.120.230.65
(ST) AesPA-nNet [19]0.230.220.55
Textual Inversion [10]0.090.130.78
Custom Diffusion [31]0.090.150.76
DreamBooth [48]0.240.170.56
StyleDrop [52]0.270.260.47
", + "image_path": "2ecf6b175cb53e3bc4fcf708d6c7c25790e9474b2358d467b1c2505d94f0e61d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 324, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 468 + ], + "type": "text", + "content": "erate bicycles with leaked content from the reference image in the last row. In contrast, benefiting from the style signal in the inversion noise, our INSTASTYLE can generate stylized images with fine-grained style details and higher fidelity. Take the bicycle in Fig. 4 as an example, our method can better preserve the color, textures, and brushstroke. Besides, the generated bicycle is more accurate. Fig. 5 presents the qualitative comparisons with style transfer methods. Despite the inherent challenge of utilizing text as content input in contrast to style transfer methods that use the content image as input, our INSTASTYLE demonstrates comparable performance in content fidelity. Additionally, our INSTASTYLE exhibits superior performance in preserving the style details of the reference image, notably in terms of color and brushstroke characteristics. More qualitative results are shown in Figs. 9 to 13 in the Supplementary Material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 471, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 532 + ], + "type": "text", + "content": "Quantitative results. We follow evaluation measures in prior works [10,52]. We utilize 100 objects in CIFAR100 [30] as the target objects. The generated images are evaluated from two aspects. For style consistency, we calculate the CLIP score [44] between the generated image and the reference image. For content fidelity, CLIP score [44] is calculated between the image and the prompt." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "As shown in Tab. 1 (a), INSTASTYLE achieves the highest style consistency and content fidelity, indicating that the images generated by INSTASTYLE exhibit a consistent style with the reference image while retaining its generation capability. For the comparison methods, Textual Inversion and Custom Diffusion have a lower style consistency, falling short of INSTASTYLE in style preservation. As for DreamBooth and StyleDrop, our approach surpasses them in generating target objects with a higher content score. We also compare with style transfer methods [6,19,21,39,66,72] which utilize content images as input. Notably, our approach diverges by only employing text prompts as content input. This brings challenges in terms of generating high-quality content as text prompts contain less information than content images. Despite this distinction, we achieve an" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "INSTASTYLE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "improved content score. Additionally, our approach excels in preserving style details, exhibiting a higher style consistency score." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 319 + ], + "type": "text", + "content": "User study. Given the highly subjective nature of stylized generation and the inherent biases in content alignment and style consistency, we conduct a user study following previous works [6, 52]. Each question presents participants with a pair of images: one stylized image generated by our approach and another produced by a selected comparison method. To avoid bias, the images are anonymized and randomized for each question. Participants are tasked with evaluating which result they believed exhibited better stylization effects and content structures. To ensure robustness and reliability, we invite 20 participants for the user study. We compare our method with 10 existing methods, including the style transfer methods [6, 18, 21, 39, 66, 72] and current comparable methods [10, 31, 48, 52]. For each method, 30 generated image pairs are randomly selected. Finally, each participant completed 300 rounds of comparisons, resulting in a total of 6,000 votes across all methods. We count the votes and show the statistical results in Tab. 1 (b). Our approach obtains a higher preference, underscoring its superiority over competitors in both style preservation and content generation." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 132, + 341, + 168, + 479 + ], + "blocks": [ + { + "bbox": [ + 132, + 341, + 168, + 479 + ], + "lines": [ + { + "bbox": [ + 132, + 341, + 168, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 168, + 479 + ], + "type": "image", + "image_path": "ddc98eed017386d9aacf4897124584d45d4e45043821f6481d8d2ab6a5cd0cd3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 499, + 482, + 555 + ], + "lines": [ + { + "bbox": [ + 130, + 499, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 499, + 482, + 555 + ], + "type": "text", + "content": "Fig. 6: Visualization of the combination of two styles. In each case, we show our combination results (the biggest image). For better comparison, we also present the stylized generation results of style1 and style2 in the top left (blue box) and bottom left (red box), respectively. Objects for synthesis are Boat, Truck, Horse, Elephant, Pear, Lion, Plate, Pot, Helicopter, Bicycle, Palace, and Motorcycle." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 186, + 341, + 254, + 480 + ], + "blocks": [ + { + "bbox": [ + 186, + 341, + 254, + 480 + ], + "lines": [ + { + "bbox": [ + 186, + 341, + 254, + 480 + ], + "spans": [ + { + "bbox": [ + 186, + 341, + 254, + 480 + ], + "type": "image", + "image_path": "08c824084edf5b72d9e5b220c3f1e0e1226b1893a8afded2055c0d8853171490.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 261, + 341, + 329, + 479 + ], + "blocks": [ + { + "bbox": [ + 261, + 341, + 329, + 479 + ], + "lines": [ + { + "bbox": [ + 261, + 341, + 329, + 479 + ], + "spans": [ + { + "bbox": [ + 261, + 341, + 329, + 479 + ], + "type": "image", + "image_path": "f8d2d26bed1296db83e55389e0b19a7c9fcceefb533625c708b4b4e8069954c2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 336, + 341, + 405, + 479 + ], + "blocks": [ + { + "bbox": [ + 336, + 341, + 405, + 479 + ], + "lines": [ + { + "bbox": [ + 336, + 341, + 405, + 479 + ], + "spans": [ + { + "bbox": [ + 336, + 341, + 405, + 479 + ], + "type": "image", + "image_path": "82e18b05baf7c9fa1db32c84e3cca3ae22b7db7774eaf5f6ed4e8a03996ea6cb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 411, + 341, + 480, + 479 + ], + "blocks": [ + { + "bbox": [ + 411, + 341, + 480, + 479 + ], + "lines": [ + { + "bbox": [ + 411, + 341, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 411, + 341, + 480, + 479 + ], + "type": "image", + "image_path": "9fbc5de0087b186f10f8abda4b9334c6691931f27bdd6402aaa9477563c972a1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 597, + 297, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 297, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 297, + 609 + ], + "type": "text", + "content": "4.3 Combination of Two Styles" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "We present style combination results in Fig. 6, where we set " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta = 0.5" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " to illustrate a more obvious effect of style combination. Specifically, in each case, we also show the stylized generation results of style 1 (the top left) and style 2 (bottom left), respectively. The biggest image on the right is our combination" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "X. Cui et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 201, + 114, + 413, + 317 + ], + "blocks": [ + { + "bbox": [ + 201, + 114, + 413, + 317 + ], + "lines": [ + { + "bbox": [ + 201, + 114, + 413, + 317 + ], + "spans": [ + { + "bbox": [ + 201, + 114, + 413, + 317 + ], + "type": "image", + "image_path": "90e0317c5d4ede5d33bd0924b11ddf8f1f3121ebb4d02ac49c021304b98895c4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 326, + 480, + 360 + ], + "lines": [ + { + "bbox": [ + 131, + 326, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 131, + 326, + 480, + 360 + ], + "type": "text", + "content": "Fig. 7: Visualization of the combination of two styles. The style can be controlled via the noise mix ratio and the prompt mix ratio. Our approach enables continuous style combinations, demonstrating its flexibility and diversity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 384, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 480, + 407 + ], + "type": "text", + "content": "results, showing the powerful style combination ability of our approach. More visualizations are shown in. Figs. 14 and 15 in Supplementary Material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 408, + 481, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 481, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 481, + 480 + ], + "type": "text", + "content": "As introduced in Sec. 3.4, both the noise mix ratio " + }, + { + "bbox": [ + 130, + 408, + 481, + 480 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 408, + 481, + 480 + ], + "type": "text", + "content": " and the prompt mix ratio " + }, + { + "bbox": [ + 130, + 408, + 481, + 480 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 408, + 481, + 480 + ], + "type": "text", + "content": " can affect the style. Fig. 7 illustrates the impact of the two parameters, where each row shows a different noise mix ratio (" + }, + { + "bbox": [ + 130, + 408, + 481, + 480 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1, 0.3, 0.5, 0.7, 0.9" + }, + { + "bbox": [ + 130, + 408, + 481, + 480 + ], + "type": "text", + "content": ") and each column shows a different prompt mix ratio (" + }, + { + "bbox": [ + 130, + 408, + 481, + 480 + ], + "type": "inline_equation", + "content": "\\beta = 0.1, 0.3, 0.5, 0.7, 0.9" + }, + { + "bbox": [ + 130, + 408, + 481, + 480 + ], + "type": "text", + "content": "). The style is mainly influenced by the noise and the prompt further improves the style details, showing that our approach is flexible and can generate diverse results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 499, + 238, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 499, + 238, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 499, + 238, + 511 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "text", + "content": "Our approach achieves stylized generation from two perspectives. One is the \"style\" noise obtained via DDIM inversion, which is utilized as the initial noise during inference to provide fine-grained style information. The other is the style token, which is learned via prompt refinement tuning to describe the style precisely. In this section, we conduct the ablation study. The visualization results and quantitative results are shown in Fig. 8 (a) and Fig. 8 (b), respectively. Specifically, each point in Fig. 8 (b) presents the average style consistency score and content fidelity score of 100 generated images for a reference style." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Impact of inversion noise. A prominent advantage of our INSTASTYLE is that we utilize the inversion noise as the initial image during inference time to preserve the fine-grained style information. Thereby, we first conduct an ablation study by replacing the inversion \"style\" noise with random noise. As shown in" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "INSTASTYLE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 112, + 482, + 248 + ], + "blocks": [ + { + "bbox": [ + 133, + 112, + 482, + 248 + ], + "lines": [ + { + "bbox": [ + 133, + 112, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 112, + 482, + 248 + ], + "type": "image", + "image_path": "27548266a07c0e682ddb7c8380909a1d562657bb15c4fac7d54b7559f5a08dc2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 265, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 331 + ], + "type": "text", + "content": "Fig. 8: Ablation study. (a) Visualization of ablation study. The reference style image is shown in the bottom right corner. Objects for synthesis are *Dinosaur*, *Bicycle*, *Bee*, and *Boy*, respectively. Without the inversion noise, the stylized performance is seriously degraded. Prompt refinement can further improve style details. (b) Quantitative results of ablation study. Our method lies further along the top right corner, showing better style preservation and content generation capability." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 357, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 407 + ], + "type": "text", + "content": "Fig. 8 (a) (second row), without the inversion noise, the stylized performance is seriously degraded. The quantitative results in Fig. 8 (b) further illustrate that ablating the \"style\" noise (blue dots) will result in a lower style consistency score, harming the style of the generated image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 407, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 491 + ], + "type": "text", + "content": "Impact of prompt refinement tuning. We ablate the prompt refinement stage to show the necessity of learning a style token to avoid ambiguity and bias in human-writing textual style tokens. Fig. 8 (a) (third row) reveals that our learnable style token can better describe the style without harming the generation ability. The quantitative results in Fig. 8 (b) show that ablating the prompt refinement (green dots) will result in a lower content fidelity score and style score, further proving the importance of the learnable style token." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 514, + 220, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 220, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 220, + 528 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Our research solves the task of stylized generation by exploring the fine-grained style information in the reference image. Through extensive analysis and empirical investigation, we find that the initial noise with a non-zero signal-to-noise ratio in the diffusion model leans toward a particular style. Therefore, DDIM inversion is introduced to obtain an initial noise containing fine-grained style information. Furthermore, we highlight the challenge caused by the internal bias in textual style tokens and propose to learn a style token adaptively to better describe the style without bias. Our approach produces satisfactory stylized generation results and supports the combination of two styles. Extensive qualitative and quantitative comparisons demonstrate the superiority of our method." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "X. Cui et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "Acknowledgement This research is sponsored by National Natural Science Foundation of China (Grant No. 62306041, U21B2045, 62176025), Beijing Nova Program (Grant No. Z211100002121106, 20230484488, 20230484276), Youth Innovation Promotion Association CAS (Grant No.2022132), and Beijing Municipal Science & Technology Commission (Z231100007423015)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 195, + 198, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 198, + 209 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 198, + 209 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 222, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 222, + 480, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 222, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 222, + 480, + 243 + ], + "type": "text", + "content": "1. Alaluf, Y., Richardson, E., Metzer, G., Cohen-Or, D.: A neural space-time representation for text-to-image personalization. ACM TOG (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "type": "text", + "content": "2. Chen, H., Zhao, L., Wang, Z., Zhang, H., Zuo, Z., Li, A., Xing, W., Lu, D.: Dualast: Dual style-learning networks for artistic style transfer. In: CVPR (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 267, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 480, + 289 + ], + "type": "text", + "content": "3. Cho, J., Nam, G., Kim, S., Yang, H., Kwak, S.: Promptstyler: Prompt-driven style generation for source-free domain generalization. In: ICCV (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 289, + 480, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 480, + 321 + ], + "type": "text", + "content": "4. Cui, X., Li, P., Li, Z., Liu, X., Zou, Y., He, Z.: Localize, understand, collaborate: Semantic-aware dragging via intention reasoner. arXiv preprint arXiv:2406.00432 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 322, + 480, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 322, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 322, + 480, + 344 + ], + "type": "text", + "content": "5. Cui, X., Li, Z., Li, P., Hu, Y., Shi, H., Cao, C., He, Z.: Chatedit: Towards multi-turn interactive facial image editing via dialogue. In: EMNLP (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 345, + 480, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 345, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 138, + 345, + 480, + 366 + ], + "type": "text", + "content": "6. Deng, Y., Tang, F., Dong, W., Ma, C., Pan, X., Wang, L., Xu, C.: Stytr2: Image style transfer with transformers. In: CVPR (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 367, + 480, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 367, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 367, + 480, + 388 + ], + "type": "text", + "content": "7. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 388, + 480, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 410 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 410 + ], + "type": "text", + "content": "8. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: CVPR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 411, + 480, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 480, + 432 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 480, + 432 + ], + "type": "text", + "content": "9. Everaert, M.N., Bocchio, M., Arpa, S., Susstrunk, S., Achanta, R.: Diffusion in style. In: ICCV (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 433, + 480, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 433, + 480, + 465 + ], + "spans": [ + { + "bbox": [ + 138, + 433, + 480, + 465 + ], + "type": "text", + "content": "0. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. In: ICLR (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 466, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 466, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 466, + 480, + 498 + ], + "type": "text", + "content": "1. Gal, R., Patashnik, O., Maron, H., Bermano, A.H., Chechik, G., Cohen-Or, D.: Stylegan-nada: Clip-guided domain adaptation of image generators. ACM TOG (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 499, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 499, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 499, + 480, + 521 + ], + "type": "text", + "content": "2. Gatys, L.A., Ecker, A.S., Bethge, M.: Image style transfer using convolutional neural networks. In: CVPR (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 521, + 480, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 543 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 543 + ], + "type": "text", + "content": "3. Gatys, L.A., Ecker, A.S., Bethge, M., Hertzmann, A., Shechtman, E.: Controlling perceptual factors in neural style transfer. In: CVPR (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 544, + 480, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 565 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 565 + ], + "type": "text", + "content": "4. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. CACM (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 566, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 480, + 588 + ], + "type": "text", + "content": "5. Gu, B., Fan, H., Zhang, L.: Two birds, one stone: A unified framework for joint learning of image and video style transfers. In: ICCV (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "type": "text", + "content": "6. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 610, + 459, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 459, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 459, + 620 + ], + "type": "text", + "content": "7. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPSW (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "text", + "content": "8. Hong, K., Jeon, S., Lee, J., Ahn, N., Kim, K., Lee, P., Kim, D., Uh, Y., Byun, H.: Aespa-net: Aesthetic pattern-aware style transfer networks. In: ICCV (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 644, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 666 + ], + "type": "text", + "content": "9. Hong, K., Jeon, S., Lee, J., Ahn, N., Kim, K., Lee, P., Kim, D., Uh, Y., Byun, H.: Aespa-net: Aesthetic pattern-aware style transfer networks. In: ICCV (2023)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "INSTASTYLE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "20. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: ICLR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "21. Huang, S., An, J., Wei, D., Luo, J., Pfister, H.: Quantart: Quantizing image style transfer towards high visual fidelity. In: CVPR (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "22. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: ICCV (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "23. Jia, G., Li, P., He, R.: Theme-aware aesthetic distribution prediction with full-resolution photographs. TNNLS (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "24. Jing, Y., Liu, X., Ding, Y., Wang, X., Ding, E., Song, M., Wen, S.: Dynamic instance normalization for arbitrary style transfer. In: AAAI (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "25. Jing, Y., Liu, Y., Yang, Y., Feng, Z., Yu, Y., Tao, D., Song, M.: Stroke controllable fast style transfer with adaptive receptive fields. In: ECCV (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "26. Jing, Y., Mao, Y., Yang, Y., Zhan, Y., Song, M., Wang, X., Tao, D.: Learning graph neural networks for image style transfer. In: ECCV (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "text", + "content": "27. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 293, + 480, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 313 + ], + "type": "text", + "content": "28. Ke, Z., Liu, Y., Zhu, L., Zhao, N., Lau, R.W.: Neural preset for color style transfer. In: CVPR (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "29. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "30. Krizhevsky, A., et al.: Learning multiple layers of features from tiny images. Technical report, University of Toronto (2009)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "31. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: CVPR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "32. Li, P., Hu, Y., He, R., Sun, Z.: Global and local consistent wavelet-domain age synthesis. TIFS (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "33. Li, P., Liu, X., Huang, J., Xia, D., Yang, J., Lu, Z.: Progressive generation of 3d point clouds with hierarchical consistency. PR (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "text", + "content": "34. Li, P., Wang, R., Huang, H., He, R., He, Z.: Pluralistic aging diffusion autoencoder. In: ICCV (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "text", + "content": "35. Li, P., Wu, X., Hu, Y., He, R., Sun, Z.: M2fpa: A multi-yaw multi-pitch high-quality dataset and benchmark for facial pose analysis. In: ICCV (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "36. Lin, S., Liu, B., Li, J., Yang, X.: Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "text", + "content": "37. Lin, T., Ma, Z., Li, F., He, D., Li, X., Ding, E., Wang, N., Li, J., Gao, X.: Drafting and revision: Laplacian pyramid network for fast high-quality artistic style transfer. In: CVPR (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "text", + "content": "38. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: ECCV (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "text", + "content": "39. Liu, S., Lin, T., He, D., Li, F., Wang, M., Li, X., Sun, Z., Li, Q., Ding, E.: Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In: ICCV (2021)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "40. Mao, J., Wang, X., Aizawa, K.: Guided image synthesis via initial image editing in diffusion model. In: ACM MM (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "text", + "content": "41. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2021)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "42. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "43. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: ICLR (2022)" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "X. Cui et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "text", + "content": "45. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "text", + "content": "46. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "text", + "content": "47. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "text", + "content": "48. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "49. Ruiz, N., Li, Y., Jampani, V., Wei, W., Hou, T., Pritch, Y., Wadhwa, N., Rubinstein, M., Aberman, K.: Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "text", + "content": "50. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: ICLR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "51. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "52. Sohn, K., Ruiz, N., Lee, K., Chin, D.C., Blok, I., Chang, H., Barber, J., Jiang, L., Entis, G., Li, Y., et al.: Styledrop: Text-to-image generation in any style. In: NeurIPS (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "text", + "content": "53. Sohn, K., Shaw, A., Hao, Y., Zhang, H., Polania, L., Chang, H., Jiang, L., Essa, I.: Learning disentangled prompts for compositional image synthesis. arXiv preprint arXiv:2306.00763 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 391, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 402 + ], + "type": "text", + "content": "54. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "type": "text", + "content": "55. Tang, H., Liu, S., Lin, T., Huang, S., Li, F., He, D., Wang, X.: Master: Meta style transformer for controllable zero-shot and few-shot artistic style transfer. In: CVPR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "56. Tang, J., Shu, X., Qi, G.J., Li, Z., Wang, M., Yan, S., Jain, R.: Tri-clustered tensor completion for social-aware image tag refinement. TPAMI (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "type": "text", + "content": "57. Teng, Q., Wang, R., Cui, X., Li, P., He, Z.: Exploring 3d-aware lifespan face aging via disentangled shape-texture representations. In: ICME (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "58. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 502, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 480, + 522 + ], + "type": "text", + "content": "59. Wang, H., Li, Y., Wang, Y., Hu, H., Yang, M.H.: Collaborative distillation for ultra-resolution universal style transfer. In: CVPR (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "text", + "content": "60. Wang, R., Guo, H., Liu, J., Li, H., Zhao, H., Tang, X., Hu, Y., Tang, H., Li, P.: Stablegarment: Garment-centric generation via stable diffusion. arXiv preprint arXiv:2403.10783 (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 556, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 480, + 588 + ], + "type": "text", + "content": "61. Wang, R., Li, P., Huang, H., Cao, C., He, R., He, Z.: Learning-to-rank meets language: Boosting language-driven ordering alignment for ordinal classification. In: NeurIPS (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "text", + "content": "62. Wang, Z., Zhao, L., Xing, W.: Stylediffusion: Controllable disentangled style transfer via diffusion models. In: ICCV (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "63. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In: ICCV (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "64. Wen, L., Gao, C., Zou, C.: Cap-vstnet: Content affinity preserved versatile style transfer. In: CVPR (2023)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "INSTASTYLE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 380 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "65. Wu, X., Hu, Z., Sheng, L., Xu, D.: Styleformer: Real-time arbitrary style transfer via parametric style composition. In: ICCV (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "type": "text", + "content": "66. Wu, Z., Zhu, Z., Du, J., Bai, X.: Ccpl: contrastive coherence preserving loss for versatile style transfer. In: ECCV (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "type": "text", + "content": "67. Xie, X., Li, Y., Huang, H., Fu, H., Wang, W., Guo, Y.: Artistic style discovery with independent components. In: CVPR (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "type": "text", + "content": "68. Xu, W., Long, C., Nie, Y.: Learning dynamic style kernels for artistic style transfer. In: CVPR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "69. Xu, Z., Sangineto, E., Sebe, N.: Stylerdalle: Language-guided style transfer using a vector-quantized tokenizer of a large-scale generative model. In: ICCV (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 226, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 226, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 480, + 248 + ], + "type": "text", + "content": "70. Yang, S., Hwang, H., Ye, J.C.: Zero-shot contrastive loss for text-guided diffusion image style transfer. In: ICCV (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 248, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 480, + 270 + ], + "type": "text", + "content": "71. Zhang, Y., Huang, N., Tang, F., Huang, H., Ma, C., Dong, W., Xu, C.: Inversion-based style transfer with diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "type": "text", + "content": "72. Zhang, Y., Tang, F., Dong, W., Huang, H., Ma, C., Lee, T.Y., Xu, C.: Domain enhanced arbitrary image style transfer via contrastive learning. In: ACM SIGGRAPH (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 303, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 480, + 325 + ], + "type": "text", + "content": "73. Zhang, Z., Li, B., Nie, X., Han, C., Guo, T., Liu, L.: Towards consistent video editing with text-to-image diffusion models. In: NeurIPS (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "type": "text", + "content": "74. Zhou, Y., Zhang, R., Sun, T., Xu, J.: Enhancing detail preservation for customized text-to-image generation: A regularization-free approach. arXiv preprint arXiv:2305.13579 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 357, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 480, + 380 + ], + "type": "text", + "content": "75. Zhu, M., He, X., Wang, N., Wang, X., Gao, X.: All-to-key attention for arbitrary style transfer. In: ICCV (2023)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "X. Cui et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation/8698146c-3791-42ec-912e-34881bbacb25_content_list.json b/2024/Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation/8698146c-3791-42ec-912e-34881bbacb25_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7ba6a416887730494ebc00ca1542a8c3e7a0a68c --- /dev/null +++ b/2024/Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation/8698146c-3791-42ec-912e-34881bbacb25_content_list.json @@ -0,0 +1,1582 @@ +[ + { + "type": "text", + "text": "Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation", + "text_level": 1, + "bbox": [ + 238, + 140, + 767, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Arpit Garg1, Cuong Nguyen3, Rafael Felix1, Thanh-Toan Do2, and Gustavo Carneiro3", + "bbox": [ + 236, + 210, + 767, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Australian Institute for Machine Learning, University of Adelaide, SA, Australia $^{2}$ Department of Data Science and AI, Monash University, Australia", + "bbox": [ + 225, + 253, + 774, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Centre for Vision, Speech and Signal Processing, University of Surrey, UK", + "bbox": [ + 245, + 282, + 754, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Deep learning faces a formidable challenge when handling noisy labels, as models tend to overfit samples affected by label noise. This challenge is further compounded by the presence of instance-dependent noise (IDN), a realistic form of label noise arising from ambiguous sample information. To address IDN, Label Noise Learning (LNL) incorporates a sample selection stage to differentiate clean and noisy-label samples. This stage uses an arbitrary criterion and a pre-defined curriculum that initially selects most samples as noisy and gradually decreases this selection rate during training. Such curriculum is sub-optimal since it does not consider the actual label noise rate in the training set. This paper addresses this issue with a new noise-rate estimation method that is easily integrated with most state-of-the-art (SOTA) LNL methods to produce a more effective curriculum. Synthetic and real-world benchmarks' results demonstrate that integrating our approach with SOTA LNL methods improves accuracy in most cases. $^4$", + "bbox": [ + 259, + 335, + 743, + 544 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Noisy-labels $\\cdot$ Instance-dependent noise $\\cdot$ Label noise Learning", + "bbox": [ + 259, + 556, + 743, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 612, + 375, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) have demonstrated their effectiveness in various domains, including vision [47], language [42], and medicine [37]. However, its efficacy is largely dependent on high-quality training data, which can be resource-intensive to obtain [43]. Although cost-effective labelling techniques, such as data mining [11] and crowdsourcing [39], offer cheaper alternatives, they often compromise label quality [39]. Consequently, this can lead to erroneous labels in real-world datasets [21]. This is a relevant issue because even slight inaccuracies in the labels can significantly affect the performance of DNN due to their inherent memorisation capabilities [34, 53]. This has prompted the development of algorithms resilient to noisy-label learning, with the aim of training models", + "bbox": [ + 212, + 643, + 789, + 796 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{4}$ Code is available at https://github.com/arpit2412/NoiseRateLearning. Supported by the Engineering and Physical Sciences Research Council (EPSRC) through grant EP/Y018036/1 and the Australian Research Council (ARC) through grant FT190100525.", + "bbox": [ + 217, + 805, + 787, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "despite inaccuracies in training data labels. Although various strategies exist, our work uniquely proposes an approach centered on the estimation of the label noise rate from the training data, which is notably absent from current methodologies.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are two types of label noise, namely instance-independent noise (IIN) [16] and instance-dependent noise (IDN) [44]. Such type of label noise typically influences the design principles of noisy-label learning algorithms. For example, IIN focuses on mislabellings that are independent of sample information [16], where estimating the underlying label transition matrix is a common way of handling this type of noise [51]. On the other hand, in the more realistic IDN, mislabelling is due to both sample information and true class labels [44], which generally require the combination of many label noise learning techniques, such as robust loss functions [28, 56], and noisy-label sample selection [26, 58]. Of the strategies mentioned above, sample selection approaches that classify training data into clean and noisy samples have produced competitive results on many benchmarks [9, 11, 15, 21, 26]. Such sample selection techniques require the definition of a classification criterion and a selection curriculum. Many studies on this topic focus on developing new sample selection criteria, such as the small-loss hypothesis [26], which states that noisy-label samples have higher loss values than clean-label samples, particularly at the early stage of training [1]. Another criterion type is the feature-based one. An example of such criterion is FINE [21] that discriminates clean and noisy-label samples via the distance to class-specific eigenvectors. In this technique, clean-label samples tend to lie closer to the class-specific dominant eigenvector of the latent representations than noisy-label samples. Another type of criterion is proposed in SSR [11], which introduces a selection criterion based on the K-nearest-neighbor (KNN) classification in the feature space. Furthermore, CC [58] uses a two-stage sampling procedure, including class-level feature clustering followed by a consistency score. An equally important problem in sample selection is the definition of the curriculum for selecting clean training samples, but it has received comparatively less attention.", + "bbox": [ + 212, + 195, + 787, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The sample selection curriculum defines a threshold to be used with one of the criteria listed above to classify the training samples as clean or noisy at each training epoch [45]. For example, the threshold can be fixed to an arbitrary clustering score that separates clean and noisy samples [26], but this strategy does not account for the proportion of label noise in the training set, nor does it consider the dynamics of the selection of noisy-label samples during the training. The consideration of such dynamics has been studied in [16, 49], which defined a curriculum of the noisy-label sampling rate $R(t)$ as a function of the training epoch $t \\in \\{1, \\dots, T\\}$ . The curriculum $R(t)$ defines a sampling rate close to $100\\%$ of the training set at the beginning of the training, which is then reduced to arbitrarily low rates at the end of the training. In practice, the function $R(t)$ is predefined [16] or learned by weighting a set of basis functions [49].", + "bbox": [ + 212, + 593, + 789, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although generally effective, these techniques do not consider the label noise rate estimated from the training set, making them vulnerable to over-fitting (if too many noisy-label samples are classified as clean) or under-fitting (if informative clean-label samples are classified as noisy). It can be argued that the estimation", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "A. Garg et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a180e75bd1f43c8f3c626a7caf5e2fc0fcf82f6c0d858ceb5e6d660de8757ded.jpg", + "image_caption": [ + "(a) Motivation" + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 491, + 289 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7d25d80cae29bd7cfecbfdbad32f62383c9de642ce5c58d654608ce8e8055e3e.jpg", + "image_caption": [ + "(b) Proposed Graphical Model", + "Fig. 1: (a) Comparison of test accuracy $\\%$ (as a function of training epoch) between the original DivideMix [26] (solid, blue curve) and our modified DivideMix (dashed, red curve) that selects the clean and noisy data based on a fixed noise rate $R(t) = 1 - \\epsilon = 50\\%$ using the small-loss criterion on CIFAR100 [22] at 0.5 IDN [44]; (b) The proposed probabilistic graphical model that generates noisy-label $\\hat{Y}$ conditioned on the image $X$ , the latent clean-label $Y$ and noise rate $\\epsilon$ , where forward pass (solid lines) is parameterised by $\\theta_y, \\theta_{\\hat{y}}$ and $\\epsilon$ representing the generation step, and the backward pass (dashed lines) is parameterised by $\\rho$ ." + ], + "image_footnote": [], + "bbox": [ + 511, + 142, + 769, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of the label transition matrix [6,7,48,51] aims to recover the noise rate affecting pairwise label transitions. However, label transition matrix techniques follow a quite different strategy compared to sample selection methods, where their main challenge is the general underconstrained aspect of the matrix estimation, making them sensitive to large noise rates and not scalable to a high number of classes [39]. Although several methods address label noise, none estimates the label noise rate directly from the training data to guide sample selection. Addressing this gap is the primary challenge of our study.", + "bbox": [ + 212, + 474, + 787, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To underscore the importance of using the noise rate for sample selection during training, we experiment with CIFAR100 [22] at an IDN rate $\\epsilon = 50\\%$ [44] (noise rate specifications and other details are explained in Sec. 4). We use DivideMix [26] and replace its sample selection (based on an arbitrary clustering score [10,32,41]) by a thresholding process that classifies the $R(t) = 1 - \\epsilon = 50\\%$ largest loss samples as noisy, and the remaining ones as clean in all training epochs $t \\in \\{1, \\dots, T\\}$ . This sample selection is used for the semi-supervised learning of DivideMix [26]. As shown in Fig. 1a, the new sample selection approach based on the \"manually provided\" noise rate (dashed red curve) improves $6\\%$ in terms of prediction accuracy compared to the original DivideMix [26] (solid blue curve) which relies on arbitrary thresholding. Similar conclusions can be achieved with other methods that apply sample selection strategies to address the noisy-label learning problem, as shown in the experiments.", + "bbox": [ + 212, + 597, + 787, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we introduce a new sample selection strategy centered on estimating the label noise rate of the training set. Our strategy is based on our novel noisy-label learning graphical model illustrated in Fig. 1b. This model can be", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Noisy-Label Learning with Noise-Rate Estimation", + "bbox": [ + 397, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c2fc73a33c98dface34ed4f3415f9bc7caaa303299e41de6b32bf051a54fe249.jpg", + "image_caption": [ + "(a) Curriculum function $R(t)$" + ], + "image_footnote": [], + "bbox": [ + 220, + 146, + 485, + 282 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d989ee52eda15970afc13f9f66dc0f90d99152f959141d0db19c6b74d4948716.jpg", + "image_caption": [ + "(b) Accuracy Comparison", + "Fig. 2: (a) Visual comparison of different $R(t)$ on CIFAR100 [22] at 0.5 IDN [44]: (i) Co-teaching [16] with curriculum based on hyper-parameter $T_{k} \\in \\{5,10\\}$ , where $R(t) = 1 - \\tau \\cdot \\min(t / T_{k}, 1)$ with $\\tau = \\epsilon = 0.5$ being manually set; (ii) S2E [49], where $R(t)$ is estimated with a bi-level optimisation; (iii) SSR [11], where $R(t)$ is based on a relabelling pattern, (iv) CC [58] following the small-loss hypothesis [26] and cosine similarity for $R(t)$ ; and (v) ours. (b) Accuracy of noisy-label learning methods on CIFAR100 [22] at 0.5 IDN [44], including DivideMix [26], FINE [21] and InstanceGM [15], without (left, blue) and (right, orange) integration of our proposed graphical model for estimating the noise rate $\\epsilon$ and sample selection rate based on $\\epsilon$ ." + ], + "image_footnote": [], + "bbox": [ + 509, + 142, + 750, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "seamlessly integrated with state-of-the-art (SOTA) noisy-label learning techniques, providing them with a precise noise rate estimate and subsequently refining the sample selection curriculum. In particular, our model's curriculum is not restrained by predefined $R(t)$ functions [16, 49], but rather relies on a dynamically estimated noise rate sourced directly from the training set, as depicted in Fig. 2a. Our method dynamically estimates training set noise rates to mitigate overfitting and underfitting, seamlessly integrates with existing algorithms, and serves as a robust foundation for future noisy-label learning research. To summarise, our main contributions include:", + "bbox": [ + 212, + 467, + 787, + 603 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- An innovative noisy-label learning graphical model, shown in Fig. 1b, that not only estimates but also leverages the noise rate from the training dataset to produce a refined sample selection curriculum for SOTA LNL methods.", + "- A simple and synergistic integration strategy of our novel graphical model with several SOTA noisy-label learning algorithms, such as DivideMix [26] and SSR [11], to improve their sample selection effectiveness, leading to an increased test accuracy, as shown in Fig. 2b." + ], + "bbox": [ + 223, + 616, + 782, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Empirical evaluations show the role of our proposed sample selection methodology in improving the efficacy of leading noisy-label learning algorithms across various synthetic and real-world IDN benchmarks. Although our primary focus lies in addressing IDN problems, because of its more challenging and realistic nature, we have also applied our model to IIN problems, with detailed results in Tab. 6 and discussion in Appendix 3 of the supplementary material. With our innovative approach, we aim to improve the accuracy of SOTA LNL methods.", + "bbox": [ + 212, + 733, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "A. Garg et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is well-known that DNNs suffer from overfitting when trained with noisy-labels [52], resulting in poor generalisation [28, 55]. To mitigate the problems related to noisy-label learning, several techniques have been developed, including noise-robust loss functions [30], noise-label sample selection [21, 26, 40, 58, 59], meta-learning [57], and re-labeling followed by sample selection [11]. These techniques can handle many types of label noise (e.g., symmetric and asymmetric), but the most challenging type, i.e., IDN [39], tends to be successfully addressed with methods that have a training stage based on sample selection techniques [21, 26, 40, 58, 59]. These sample selection techniques separate clean and noisy-label samples, where clean samples are treated as labelled, and noisy samples are discarded [8, 16, 18, 49] or treated as unlabeled samples [9, 15, 26, 35, 59] for semi-supervised learning [3]. Another solution consists of aligning clean and noisy samples with an information fusion method [19]. A major limitation of these approaches is the overly simplistic design of the curriculum to select clean and noisy-label samples during training. It is either predetermined [16, 18] or learned from a set of predetermined basis functions [49]. We argue that the design of a curriculum based on an estimated noise rate would benefit such methods, as motivated by Fig. 1a, but to the best of our knowledge, such estimation has not been explored by previous methods. The closest idea explored is based on the estimation of the type of noise instead of the noise rate [45]. It is possible to argue that the estimation of the label transition matrices [5, 7, 44, 60], composed of the learned noise rate that affects the transition between pairs of labels, is a way to estimate the label noise rate. However, they portray results of comparatively lower accuracy for large real-world datasets or for IDN problems [39, 44]. A possible reason for these poorer results is that label transition approaches suffer from identifiability issues [14], where any clean-label distribution assignment is acceptable as long as the distribution of observed labels can be reconstructed [13]. This makes the identification of the actual underlying clean-labels challenging. One solution is to consider the use of multiple annotations per sample to help analyse agreements and disagreements for improved identification of clean patterns [14, 31]. However, the annotation of datasets with a single label per training sample is already challenging; the complexity is significantly larger to acquire multiple labels per sample.", + "bbox": [ + 212, + 172, + 787, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An alternative technique to handle IDN problems is based on graphical models that represent the relationship between various observed and latent variables [15,45,50]. The approaches in [15,50] use graphical models that rely on a generative approach to produce noisy-labels from the respective image features and latent clean-labels. However, previous graphical models do not take into account the underlying noise rate parameter during modeling. Our work is the first graphical model approach to estimate the noise rate of the dataset. In addition, a close examination of existing literature reveals a trend: when traditional graphical models are integrated with SOTA noisy-label learning methods, the outcomes often fall short of expectations in terms of accuracy [2,15,50]. This observation underscores a strong point of our work. Our approach is uniquely tailored for", + "bbox": [ + 212, + 674, + 789, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Noisy-Label Learning with Noise-Rate Estimation", + "bbox": [ + 397, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "easy integration with current SOTA noisy-label learning methods for sample selection, which, in turn, improves over the SOTA classification accuracy.", + "bbox": [ + 212, + 146, + 787, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 199, + 330, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we present our new graphical model that estimates the noise rate, which will be used in the sample selection process. Let $\\mathcal{D} = \\{(x_i,\\hat{y}_i)\\}_{i = 1}^N$ be the noisy-label training set containing $d$ -dimensional data vector $x_{i}\\in \\mathcal{X}\\subseteq \\mathbb{R}^{d}$ and its respective $C$ -dimensional one-hot encoded observed (potentially corrupted) label $\\hat{y}_i\\in \\hat{\\mathcal{V}} = \\{\\hat{y}:\\hat{y}\\in \\{0,1\\} ^C\\wedge \\mathbf{1}_C^\\top \\hat{y} = 1\\}$ , where $\\mathbf{1}_C$ is a vector of ones with $C$ dimensions. The aim is to estimate the label noise rate $\\epsilon$ , used for the generation of noisy-label training data from the observed training dataset $\\mathcal{D}$ and integrate this label noise rate into the sample selection strategy.", + "bbox": [ + 212, + 231, + 789, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Graphical Model", + "text_level": 1, + "bbox": [ + 215, + 375, + 403, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We portray the generation of noisy-label via the probabilistic graphical model shown in Fig. 1b. The observed random variables, denoted by shaded circles, are data $X$ and the corresponding noisy-label $\\hat{Y}$ . We also have one latent variable, namely: the clean-label $Y$ . Under our proposed modeling assumption, a noisy-label of a data instance can be generated as follows:", + "bbox": [ + 212, + 398, + 787, + 474 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- sample an instance from $p(X)$ , i.e.: $x \\sim p(X)$", + "- sample a clean-label from the clean-label distribution:" + ], + "bbox": [ + 225, + 486, + 630, + 516 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ny \\sim \\operatorname {C a t e g o r i c a l} (Y; f _ {\\theta_ {y}} (x))\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 517, + 434, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- sample a noisy-label from the noisy-label distribution:", + "bbox": [ + 225, + 532, + 630, + 546 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} \\sim \\operatorname {C a t e g o r i c a l} (\\hat {Y}; \\epsilon \\times f _ {\\theta_ {\\hat {y}}} (x, f _ {\\theta_ {y}} (x)) + (1 - \\epsilon) \\times y),\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 546, + 612, + 564 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\text{Categorical}()$ denotes a categorical distribution, $f_{\\theta_y}: \\mathcal{X} \\to \\Delta_{C-1}$ and $f_{\\theta_{\\hat{y}}}: \\mathcal{X} \\times \\Delta_{C-1} \\to \\Delta_{C-1}$ denote two classifiers for the clean-label $Y$ and noisy-label $\\hat{Y}$ , respectively, with $\\Delta_{C-1} = \\{s : s \\in [0,1]^C \\wedge \\mathbf{1}_C^\\top s = 1\\}$ being the $(C-1)$ -dimensional probability simplex. According to the data generation process shown in Fig. 1b, $\\epsilon$ corresponds to $\\mathbb{E}_{(x,\\hat{y}) \\sim p(X,\\hat{Y})}[P(\\hat{y} \\neq y|x)]$ , which is the label noise rate of the training dataset of interest. Our aim is to infer the parameters $\\theta_y, \\theta_{\\hat{y}}$ and $\\epsilon$ from a noisy-label dataset $\\mathcal{D}$ by maximising the following log-likelihood:", + "bbox": [ + 212, + 573, + 787, + 696 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon} \\mathbb {E} _ {(x _ {i}, \\hat {y} _ {i}) \\sim \\mathcal {D}} [ \\ln p (\\hat {y} _ {i} | x _ {i}; \\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon) ] = \\max _ {\\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon} \\mathbb {E} _ {(x _ {i}, \\hat {y} _ {i}) \\sim \\mathcal {D}} [ \\ln \\sum_ {y _ {i}} p (\\hat {y} _ {i}, y _ {i} | x _ {i}; \\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon) ]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 708, + 787, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Due to the presence of the clean-label $y_{i}$ , it is difficult to directly evaluate the log-likelihood in Eq. (1). Therefore, we employ the expectation-maximisation (EM) algorithm [10] to maximise the log-likelihood. The main idea of the EM algorithm is to (i) construct a tight lower bound of the likelihood in Eq. (1) by estimating the posterior of the latent variable $Y$ (known as expectation step) and (ii) maximise that lower bound (known as maximisation step). Formally, let", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "A. Garg et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$q(y_{i}|x,\\hat{y};\\rho)$ be an arbitrary distribution over a clean-label $y_{i}$ . The evidence lower bound (ELBO) of the log-likelihood in Eq. (1) can be obtained through Jensen's inequality and is presented as follows:", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} Q \\left(\\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon , \\rho\\right) = \\mathbb {E} _ {\\left(x _ {i}, \\hat {y} _ {i}\\right) \\sim \\mathcal {D}} \\left[ \\mathbb {E} _ {q \\left(y _ {i} \\mid x _ {i}; \\hat {y} _ {i}; \\rho\\right)} \\left[ \\ln p \\left(\\hat {y} _ {i} \\mid x _ {i}, y _ {i}; \\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon\\right) \\right] \\right. \\\\ - \\operatorname {K L} \\left[ q \\left(y _ {i} \\mid x _ {i}, \\hat {y} _ {i}; \\rho\\right) \\| p \\left(y _ {i} \\mid x _ {i}; \\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon\\right) \\right] \\\\ = \\mathbb {E} _ {\\left(x _ {i}, \\hat {y} _ {i}\\right) \\sim \\mathcal {D}} \\left[ \\mathbb {E} _ {q \\left(y _ {i} \\mid x _ {i}; \\hat {y} _ {i}; \\rho\\right)} \\left[ \\ln p \\left(\\hat {y} _ {i} \\mid x _ {i}, y _ {i}; \\theta_ {\\hat {y}}, \\epsilon\\right) \\right. \\right. \\\\ \\left. + \\ln p \\left(y _ {i} \\mid x _ {i}; \\theta_ {y}\\right) \\right] + \\mathbb {H} \\left[ q \\left(y _ {i} \\mid x _ {i}, \\hat {y} _ {i}; \\rho\\right) \\right], \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 281, + 203, + 785, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{KL}[q\\| p]$ is the Kullback - Leibler divergence between distributions $q$ and $p$ and $\\mathbb{H}(q)$ is the entropy of the distribution $q$ . The EM algorithm is then carried out iteratively by alternating the following two steps:", + "bbox": [ + 212, + 287, + 782, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "E step: we maximise the ELBO in (2) w.r.t. $q(y_{i}|x_{i},\\hat{y}_{i};\\rho)$ . Theoretically, such an optimisation results in KL $[q(y_{i}|x_{i},\\hat{y}_{i};\\rho)\\| p(y_{i}|x_{i},\\hat{y}_{i})] = 0$ or $q(y_{i}|x_{i},\\hat{y}_{i};\\rho) = p(y_{i}|x_{i},\\hat{y}_{i})$ . This is equivalent to estimating the posterior of the clean-label $y_{i}$ given noisy-label data $(x_{i},\\hat{y}_{i})$ . Obtaining the posterior $p(y_{i}|x_{i},\\hat{y}_{i})$ is, however, intractable for most deep-learning models. To mitigate such an issue, we follow the variational EM approach [33] by employing an approximate posterior $q(y_{i}|x_{i},\\hat{y}_{i};\\rho^{(t)})$ that is the closest to the true posterior $p(y_{i}|x_{i},\\hat{y}_{i})$ , where:", + "bbox": [ + 212, + 347, + 784, + 455 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\rho^ {(t)} = \\arg \\max _ {\\rho} Q \\left(\\theta_ {y} ^ {(t)}, \\theta_ {\\hat {y}} ^ {(t)}, \\epsilon^ {(t)}, \\rho\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 465, + 785, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with the superscript ${}^{(t)}$ denoting the parameters at the $t$ -th iteration. Although this results in a non-tight lower bound of the log-likelihood in Eq. (1), it does increase the variational bound $Q$ .", + "bbox": [ + 212, + 503, + 782, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "M step: we maximise the ELBO in Eq. (2) w.r.t. $\\theta_y, \\theta_{\\hat{y}}$ and $\\epsilon$ given $\\rho^{(t)}$ obtained in the E step:", + "bbox": [ + 212, + 563, + 782, + 594 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {y} ^ {(t + 1)}, \\theta_ {\\hat {y}} ^ {(t + 1)}, \\epsilon^ {(t + 1)} = \\arg \\max _ {\\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon} Q (\\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon , \\rho^ {(t)}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 604, + 785, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The estimated noise rate $\\epsilon$ can then be integrated into certain noisy-label algorithms to train the models of interest as mentioned in Sec. 1. In addition, the inference of noise rate $\\epsilon$ might be associated with the identifiability issue when estimating the clean-label $Y$ [29], i.e., there exists multiple sets of $\\rho$ and $\\theta_y$ , where each set can explain the observed noisy-label data equally well. Such issues are addressed in the following subsection.", + "bbox": [ + 212, + 642, + 784, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Sample Selection", + "text_level": 1, + "bbox": [ + 214, + 753, + 401, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The identifiability issue when inferring the clean-label $Y$ from noisy-label data $(X, \\hat{Y})$ can be mitigated either by acquiring multiple noisy-labels [29] or introducing additional constraints, such as small loss hypothesis [16] or FINE [21]. Since requesting additional noisy-labels per training sample is not always possible, we", + "bbox": [ + 212, + 779, + 782, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Noisy-Label Learning with Noise-Rate Estimation", + "bbox": [ + 397, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Proposed noisy-label learning that relies on the estimation of noise rate $\\epsilon$ to build a sample selection curriculum." + ], + "code_body": "1: procedure NOISE RATE ESTIMATION AND INTEGRATION(D,T,λ) \n2: $\\triangleright$ $\\mathcal{D} = \\{(x_i,\\hat{y}_i)\\}_{i = 1}^N$ training set with noisy-label data \n3: $\\triangleright$ T: number of epochs \n4: $\\triangleright$ λ: a hyper-parameter \n5: $\\triangleright$ Notation: $\\Theta^{(t)} = (\\theta_y^{(t)},\\theta_{\\hat{y}}^{(t)},\\epsilon^{(t)})$ \n6: Initialise $\\theta_y^{(1)},\\theta_{\\hat{y}}^{(1)},\\epsilon^{(1)}$ and $\\rho^{(0)}$ \n7: $\\theta_y^{(1)}\\gets \\mathrm{WARM~UP}(\\mathcal{D},\\theta_y^{(1)})$ \n8: $t\\gets 0$ \n9: for $n_{\\mathrm{epoch}} = 1:T$ do \n10: for each mini-batch $\\mathcal{S}$ in SHUFFLE(D) do \n11: $t\\gets t + 1$ \n12: $\\mathcal{S}_{\\mathrm{clean}},\\mathcal{S}_{\\mathrm{noisy}}\\gets \\mathrm{SAMPLE~SELECTION}(\\mathcal{S},\\theta_y^{(t)},\\epsilon^{(t)})$ \n13: $\\triangleright$ Variational E-step as in Eq. (3) \n14: $\\rho^{(t)}\\gets \\mathrm{E~STEP}(\\mathcal{S},\\theta_y^{(t)},\\theta_{\\hat{y}}^{(t)},\\epsilon^{(t)},\\rho^{(t - 1)})$ \n15: $\\triangleright$ M-step as in Eq. (8) \n16: $\\theta_y^{(t + 1)},\\theta_{\\hat{y}}^{(t + 1)},\\epsilon^{(t + 1)}\\gets \\mathrm{M~STEP}(\\mathcal{S}_{\\mathrm{clean}},\\mathcal{S}_{\\mathrm{noisy}},\\theta_y^{(t)},\\theta_{\\hat{y}}^{(t)},\\epsilon^{(t)},\\rho^{(t)},\\lambda)$ \n17: return $\\theta_y$ \n1 parameter of the clean-label classifier", + "bbox": [ + 215, + 176, + 790, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "follow the latter approach by imposing a constraint (scoring term), denoted as $L(\\theta_y, \\epsilon^{(t)})$ , over $\\theta_y$ in the M step via a sample selection approach based on the estimated noise rate $\\epsilon^{(t)}$ . Formally, we propose a new curriculum when selecting samples as follows:", + "bbox": [ + 215, + 511, + 787, + 570 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nR (t) = 1 - \\epsilon^ {(t)}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 570, + 785, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the simplest case, such as Co-teaching [16] or FINE [21], the constraint for $\\theta_y$ can be written as:", + "bbox": [ + 212, + 593, + 785, + 625 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL \\left(\\theta_ {y}, \\epsilon^ {(t)}\\right) = \\sum_ {\\left(x _ {i}, \\hat {y} _ {i}\\right) \\in \\mathcal {S} _ {\\text {c l e a n}}} \\operatorname {K L} \\left[ \\operatorname {C a t e g o r i c a l} (Y; \\hat {y}) \\| \\operatorname {C a t e g o r i c a l} \\left(Y; f _ {\\theta_ {y}} \\left(x _ {i}\\right)\\right) \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 632, + 787, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where:", + "bbox": [ + 215, + 676, + 266, + 689 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {\\text {c l e a n}} = \\left\\{\\left(x _ {i}, \\hat {y} _ {i}\\right): \\left(x _ {i}, \\hat {y} _ {i}\\right) \\in \\mathcal {D} \\wedge z \\left(x _ {i}, y _ {i}\\right) \\leq m \\right\\},\n$$\n", + "text_format": "latex", + "bbox": [ + 256, + 699, + 609, + 717 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nm \\in \\{m: \\Pr (z \\left(x _ {i}, y _ {i}\\right) \\leq m) \\geq R (t) \\wedge \\Pr (z \\left(x _ {i}, y _ {i}\\right) \\geq m) \\geq 1 - R (t) \\},\n$$\n", + "text_format": "latex", + "bbox": [ + 256, + 718, + 741, + 734 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {\\text {n o i s y}} = \\mathcal {D} \\backslash \\mathcal {S} _ {\\text {c l e a n}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 256, + 737, + 785, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with $R(t)$ defined in (5), and $z(x,y)$ representing the score of a criterion (e.g., loss [26, 58, 59], distance to the largest eigenvectors [21], or KNN scores [11]). Intuitively, the loss in Eq. (6) is simply the cross-entropy loss on the $\\lfloor R(t)\\times N\\rfloor$ samples that have smallest scores (with $\\lfloor .\\rfloor$ being the floor function). One can also extend to other SOTA models by replacing the loss $L$ accordingly. For example,", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "A. Garg et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/08c3a57bb884bedf5284bacc78cd61a6a3ead553af3eb4a53130280ac377ba00.jpg", + "image_caption": [ + "Fig. 3: Our training algorithm uses an existing noisy-label classifier (e.g., DivideMix [26]) parameterised by $\\theta_y$ as the clean-label model $p(Y|X;\\theta_y)$ . The generation of noisy-label (given $X$ and $Y$ ) is carried out by a model parameterised by noise rate $\\epsilon$ and $\\theta_{\\tilde{y}}$ . The noisy-label classifier is based on a sample-selection mechanism that uses a curriculum $R(t) = 1 - \\epsilon^{(t)}$ ." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 772, + 316 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "if DivideMix [26] is used as a base model to constrain $\\theta_y$ , $L$ will include two additional terms: loss on un-labeled data and regularisation using mixup [54].", + "bbox": [ + 214, + 428, + 787, + 460 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 Training and Testing", + "text_level": 1, + "bbox": [ + 215, + 487, + 434, + 503 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Given the sample selection approach in Sec. 3.2, the M step in Eq. (4) is slightly modified by integrating the constraint in Eq. (6), which can be written as", + "bbox": [ + 214, + 518, + 785, + 550 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {y} ^ {(t + 1)}, \\theta_ {\\hat {y}} ^ {(t + 1)}, \\epsilon^ {(t + 1)} = \\arg \\max _ {\\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon} Q (\\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon , \\rho^ {(t)}) - \\lambda L (\\theta_ {y}, \\epsilon^ {(t)}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 264, + 563, + 787, + 590 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyper-parameter and $L$ is defined in Eq. (6). The training procedure is summarised in Algorithm 1 and visualised in Fig. 3. In the implementation, we integrate the proposed method into existing models, such as DivideMix [26] or FINE [21]. Note that the clean-label classifier $f_{\\theta_y}(.)$ is also the clean classifier of the base model.", + "bbox": [ + 214, + 606, + 787, + 680 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 710, + 375, + 728 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We show extensive experiments in several noisy-label synthetic benchmarks with CIFAR100 [22], and real-world benchmarks, including CNWL's red mini-ImageNet [17], Clothing1M [45] and mini-WebVision [27]. Sec. 4.2 describes implementation details. We evaluate our approach by plugging SOTA models into $p(y|x;\\theta_y)$ , defined in Sec. 3, with results being shown in Sec. 4.3, and ablation studies in Sec. 4.4.", + "bbox": [ + 214, + 750, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Noisy-Label Learning with Noise-Rate Estimation", + "bbox": [ + 397, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Dataset Descriptions", + "text_level": 1, + "bbox": [ + 215, + 146, + 433, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We follow the existing literature to generate the IDN labels [44] for CIFAR100 [22]. The dataset contains 50,000 training images and 10,000 testing images of shape $32 \\times 32 \\times 3$ . The dataset is class-balanced with 100 different categories and the IDN rates considered are 0.2, 0.3, 0.4 and 0.5 [44]. Another important IDN benchmark dataset is the red mini-ImageNet, a real-world dataset from CNWL [17]. In this dataset, there are 100 distinct classes, with each class containing 600 images. The images and their corresponding noisy-labels have been crawled from the internet at various controllable label noise rates. For a fair comparison with the existing literature [9, 15, 46], we have resized the images to $32 \\times 32$ pixels from the $84 \\times 84$ original pixel settings. We show results with $40\\%$ , $60\\%$ and $60\\%$ noise rates on this dataset. Clothing1M [45] is a real-world clothing classification dataset that contains 1 million images with an estimated noise rate of $38.5\\%$ . The dataset contains 14 different categories, and the labels are generated from surrounding texts. Images in this dataset are of different sizes, and we follow the resize structure suggested in [21, 26]. This dataset also contains $50k$ manually validated clean training images, $14k$ images for the validation set, and $10k$ testing images. We have not used the clean training, validation set, or any extra training images while training. Only the testing set is used for evaluation. Mini-WebVision [26] contains 65,944 images with their respective labels from the 50 different initial categories from the WebVision [27], with image size of $256 \\times 256$ pixels. Following the evaluation process commonly used in this benchmark, 50 categories from the ILSVRC12 [23] dataset are also used for testing.", + "bbox": [ + 212, + 171, + 789, + 506 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation", + "text_level": 1, + "bbox": [ + 215, + 527, + 392, + 541 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "All methods are implemented in Pytorch [36] and use one NVIDIA RTX 3090 card for training and testing. As mentioned in the original papers, hyperparameter settings are kept the same for the baselines used in the proposed algorithm. All classifier architectures are also kept the same as the baseline models. A random initialisation of noise rate parameter $\\epsilon$ with the sigmoid as its activation function is employed for all experiments to maintain the fairness of the comparisons with other approaches. The value of $\\lambda$ in Eq. (8) is set to 1 for all the cases. We integrate many SOTA approaches [11, 15, 21, 26, 58, 59] into our graphical model, as explained in Sec. 3.3. For CIFAR100 [22] with IDN [44], we integrate DivideMix [26], C2D [59], CC [58], and InstanceGM [15] into our model, given their superior performance across various noise rates. Additionally, we also use F-Dividemix from FINE [21], as shown in Fig. 2b. For red mini-ImageNet [17], we test our proposed approach with and without DINO self-supervision [4]. For the implementation without self-supervision, we use DivideMix [26] and InstanceGM [15], and for the self-supervised version, we only use InstanceGM [15]. The models for Clothing1M [45] and mini-WebVision [26] (validation on ImageNet [23]) are trained using DivideMix [26], SSR [11], C2D [59] and CC [58]. Additional implementation details, empirical analysis, and computational analysis are presented in the supplementary material in Appendices 1, 2 and 4, respectively.", + "bbox": [ + 212, + 551, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "A. Garg et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/4bfd87feb77d72e512e4ee274f0c31901970f3e0ba53f7122e3fe8f58099bdec.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNoise Rates - IDN
0.20.30.40.5
CE [50]30.4224.1521.4514.42
PartT [44]65.3364.5659.7356.80
kMEIDTM [7]69.1666.7663.4659.18
DivideMix [26]77.0376.3370.8058.61
DivideMix-Ours77.4277.2172.4164.02
(+0.39)(+0.88)(+1.61)(+5.41)
C2D [59]78.6178.1872.8963.19
C2D-Ours79.0778.5973.3165.28
(+0.46)(+0.41)(+0.42)(+2.09)
CC [58]79.6177.5676.5863.19
CC-Ours79.7278.7177.3867.53
(+0.11)(+1.15)(+0.80)(+4.34)
InstanceGM [15]79.6979.2178.4777.19
InstanceGM-Ours79.6179.4079.5278.21
(-0.08)(+0.19)(+1.05)(+1.02)
", + "bbox": [ + 225, + 243, + 472, + 458 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b2d58fde9424dd3cccf214c591423fbdd0f7123c4d8ff348a89bcb2192a7e5c3.jpg", + "table_caption": [ + "Table 1: Test accuracy % on (left) CIFAR100 [22] under different IDN [45], and (right) red mini-ImageNet [17]. Other models' results are from [7, 15, 46]. Here, we integrate DivideMix [26], C2D [59], CC [58] and InstanceGM [15] into our proposed model for CIFAR100 [22]. For red mini-ImageNet [17], we present the results with and without self-supervision [4]. We integrate DivideMix [26] and InstanceGM [15] into our model, with the latter tested with and without self-supervision." + ], + "table_footnote": [], + "table_body": "
MethodNoise rate
0.40.60.8
CE [46]42.7037.3029.76
MixUp [54]46.4040.5833.58
MentorMix [17]47.1443.8033.46
FaMUS [46]51.4245.1035.50
DivideMix [26]46.7243.1434.50
DivideMix-Ours50.7045.1137.44
(+3.98)(+1.97)(+2.94)
InstanceGM [15]52.2447.9639.62
InstanceGM-Ours56.6151.4043.83
(+4.37)(+3.44)(+4.21)
With self-supervised learning
PropMix [9]56.2252.8443.42
InstanceGM-SS [15]56.3753.2144.03
InstanceGM-SS-Ours58.2953.6045.47
(+1.92)(+0.39)(+1.44)
", + "bbox": [ + 542, + 239, + 777, + 460 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/284a65786cefd4bf552f156bc2186dde2912e47a7aaa070cc230f159a079983e.jpg", + "table_caption": [ + "Table 2: Final estimated noise rate $\\epsilon$ for the Tab. 1, (left) CIFAR100 [22] under various IDN [44], and (right) red mini-ImageNet [17] with and without self-supervision." + ], + "table_footnote": [], + "table_body": "
Estimated noise rates
MethodActual noise rate
0.20.30.40.5
DivideMix-Ours0.180.340.470.53
C2D-Ours0.190.330.380.52
CC-Ours0.210.340.410.54
InstanceGM-Ours0.230.370.420.47
", + "bbox": [ + 241, + 513, + 426, + 604 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/d433bbea64944dfbb523ed9b62d1e29215db162b39b2c3d74b086baa83a95672.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Estimated noise rates
MethodActual noise rate
0.40.6
DivideMix-Ours0.390.58
InstanceGM-Ours0.380.55
InstanceGM-SS-Ours0.480.53
0.73
", + "bbox": [ + 558, + 518, + 759, + 599 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Comparison with SOTA", + "text_level": 1, + "bbox": [ + 215, + 628, + 460, + 645 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "This section compares our approach with SOTA methods on datasets with the IDN settings and noisy real-world settings. The bold text in tables indicate SOTA results, and our results are in the greyed rows.", + "bbox": [ + 212, + 652, + 787, + 699 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Synthetic Instance-Dependent Noise The comparison between various baselines and our proposed method on CIFAR100 [22] with IDN [44] is shown in Tab. 1 (left) with the noise rate ranging from 0.2 to 0.5. It is worth noting that using our proposed model with DivideMix [26], C2D [59], CC [58] and InstanceGM [15] improve their performance in almost all cases, particularly with large noise rates. Tab. 2 (left) also shows the final noise rate $\\epsilon$ estimated by our model which is reasonably close to the simulated noise rate displayed in the table's header.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Noisy-Label Learning with Noise-Rate Estimation", + "bbox": [ + 397, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/0332020760bcfb9fed8e6553408cbb4a0a7b58c2587196643a6ee2d8aa30f5c8.jpg", + "table_caption": [ + "Table 3: Test accuracy (\\%) of competing methods, and final estimated noise rate $\\epsilon$ on Clothing1M [45]. Also, we have not considered competing models that rely on a clean or validation set whilst training. We integrate DivideMix [26], C2D [59], SSR [11] and CC [58] into our model. DivideMix [26] and CC [58] shows the locally reproduced results. Mentioned noise rate by [45] is 0.385." + ], + "table_footnote": [], + "table_body": "
MethodTest accuracy (%)Estimated noise rate
OT-Filter [12]74.50
ELR+ with C2D [59]74.58
AugDesc [35]75.11
NCE [25]75.30
DivideMix [26]74.32
DivideMix-Ours74.41 (+0.09)0.41
C2D [59]74.58
C2D-Ours74.71 (+0.13)0.41
SSR (class-imbalance) [11]74.12
SSR-Ours74.20 (+0.08)0.42
CC [58]75.24
CC-Ours75.31 (+0.07)0.41
", + "bbox": [ + 316, + 226, + 689, + 367 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Real-World Noise We also evaluate our proposed method on various real-world noisy settings regarding test accuracy and estimated noise rates $\\epsilon$ in Tabs. 1 to 4. Similarly to the synthetic IDN in Tab. 1 (left), the results show that existing noisy-label robust methods can be easily integrated with our model to outperform current SOTA results for real-world noisy-label datasets. Tabs. 1 and 2 (right) shows the results on red mini-ImageNet using two configurations, including cases without self-supervision (top part of the table) and with self-supervision (bottom part of the table). The self-supervision DINO pre-training [4] relies only on images from red mini-ImageNet to enable a fair comparison with existing baselines [9, 15]. Results from Tab. 1 (right) demonstrate that our approach improves the performance of SOTA methods by a considerable margin in all cases. In fact, using estimated noise rate $\\epsilon$ (Tab. 2 (right)) while training InstanceGM [15] without self-supervision shows better performance than existing self-supervised baselines at 0.4 noise rate. Moreover, DivideMix [26], C2D [59], SSR [11], and CC [58] are used as a baselines for Clothing1M [45] as shown in Tab. 3, with results showing slight improvements with the use of our method. For mini-WebVision [26] (validation on ImageNet [23]), we use DivideMix [26], C2D [59], SSR [11], and CC [58] as baselines (Tab. 4). It is worth noting that our results are better than the SOTA for all settings in Tab. 4.", + "bbox": [ + 212, + 395, + 789, + 683 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation", + "text_level": 1, + "bbox": [ + 215, + 707, + 334, + 722 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We show an ablation study with testing accuracy (top) and training time (bottom) of our approach in Tab. 5 (left) on CIFAR100 [22] at 0.3 and 0.5 IDN [44] using DivideMix [26] as baseline. Detailed complexity analysis is shown in the supplementary material, Tab. 10. Initially, the accuracy result of baseline DivideMix [26] is $76.33\\%$ , $58.61\\%$ for 0.3 and 0.5 IDN, respectively. In the second row, we fix the noise rate $\\epsilon$ at 0.3 and 0.5 for DivideMix's sample selection, as explained in Sec. 3.3 (without updating $\\epsilon$ ), then the results improved to $78.14\\%$ and $64.44\\%$ ,", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "A. Garg et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3964453370a30ed11490097d59e6764282009b9af482a1587051d72fc2d23f24.jpg", + "table_caption": [ + "Table 4: Test accuracy (\\%) and final estimated noise rate $\\epsilon$ on mini-WebVision [26] and validation on ImageNet [23]. We integrate DivideMix [26], Contrast-to-Divide (C2D) [59], SSR [11] and CC [58] into our model, whilst models suffixed with -Ours denotes our proposed approach. ImageNet [23] is only considered for validation." + ], + "table_footnote": [], + "table_body": "
Datasetmini-WebVisionImageNetEstimated noise rate
Top-1Top-5Top-1Top-5
BtR [38]80.8892.7675.9692.20
NCE [25]79.5093.8076.3094.10
DivideMix [26]77.3291.6475.2091.64
DivideMix-Ours78.5192.0376.1193.240.43
(+1.19)(+0.39)(+0.91)(+1.60)
C2D [59]79.4292.3278.5793.04
C2D-Ours80.2092.8279.1693.120.43
(+0.78)(+0.50)(+0.59)(+0.08)
SSR [11]80.9292.8075.7691.76
SSR-Ours81.6893.8076.9193.050.43
(+0.76)(+1.00)(+1.15)(+1.29)
CC [58]79.3693.6476.0893.86-
CC-Ours80.0193.7977.1194.210.44
(+0.65)(+0.15)(+1.03)(+0.35)
", + "bbox": [ + 334, + 212, + 671, + 388 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b6a8bf4fbd2ae27fd3fda69734d7d7a474bf226e810f821298f9c5518fc378f5.jpg", + "table_caption": [ + "Table 5: (left) Ablation results on CIFAR100 at 0.3 and 0.5 IDN. We display accuracy results under different configurations and compare the training time of our approach against DivideMix [26]. The implementation settings and detailed computational analysis are described in Sec. 4.2 and Appendix 4 respectively. (right above) Ablation results for initialisation with different $\\epsilon$ values on CIFAR100 [22] at a synthetic noise rate of 0.5 IDN [44] with DivideMix-Ours. (right below) Ablation graph showing the learning pattern of the noise rate parameter estimation under different initialisation values, i.e., $\\epsilon = 0.001$ (blue), $\\epsilon = 0.1$ (orange), $\\epsilon = 0.9$ (red), when trained on DivideMix-ours case at 0.5 IDN [44] on CIFAR100 [22]." + ], + "table_footnote": [], + "table_body": "
ModelsAccuracy (%)
0.30.5
DivideMix [26]76.3358.61
DivideMix with fixed ε78.1464.44
Our method with pre-trained DivideMix75.0152.31
Our method with original DivideMix (no ε)76.2056.30
DivideMix-Ours77.2164.02
Training Time (GPU-hours)
DivideMix [26]7.20
DivideMix-Ours8.50
DivideMix-Ours (Mixed Precision)6.10
", + "bbox": [ + 228, + 561, + 493, + 686 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/aa983ece11a754aafafd52ae1d8727f56159fc784823abeb033ac358abf72881.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 544, + 717, + 696 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "which is the ideal case (i.e., a perfect noise rate estimation) that motivated our work. In the third case, we use the proposed graphical model with pre-trained DivideMix [26] that shows accuracy of $75.01\\%$ and $52.31\\%$ . In the next case, the proposed graphical model is trained together with DivideMix [26] without considering the estimated noise rate $\\epsilon$ for sample selection, which results in accuracy of $76.20\\%$ and $56.30\\%$ . In the last row, we show the training of the proposed model with DivideMix [26], together with the estimation of noise rate $\\epsilon$ , and the", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Noisy-Label Learning with Noise-Rate Estimation", + "bbox": [ + 397, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "selection of samples based on that, with $\\approx 1\\%$ and $\\approx 8\\%$ accuracy improvement to 0.3 and 0.5 IDN, which is quite close to our ideal case (second row). We also study the effect that the initialisation of the noise rate parameter $\\epsilon$ plays in our method. As described in Sec. 4.2, we randomly initialise $\\epsilon$ within the range $(0,1)$ . To test the robustness of our method to this initialisation, we perform an ablation study by initialising $\\epsilon$ with three different values, namely: 0.001, 0.1, and 0.9 on CIFAR100 [24] at 0.5 IDN [44] using the DivideMix-ours approach. Our results, presented in Tab. 5 (right above and below), demonstrate that the initial value of $\\epsilon$ has a limited effect on the model's performance, highlighting the stability of our method across different initialisation of $\\epsilon$ . In the supplementary material, Tab. 7 (left) presents comparative analyses of our model against PartT [44] and SSR [11] on CIFAR100 [22] at 0.5 IDN [44], while Tab. 7 (right) showcases similar comparisons with baselines DivideMix [26] and InstanceGM [15] at a higher noise rate of 0.6 IDN [44]. Additionally, Tab. 8 provides further insights into the performance of DivideMix [26] on red mini-ImageNet [17] at lower noise rates of 0.1, 0.2 and 0.3. Tab. 9 illustrates noise rate estimation with standard deviation for our model integrated with DivideMix [26] and InstanceGM [15] on CIFAR100 [22] under IDN [44] settings at 0.2, 0.3, 0.4, and 0.5. Further statistical analysis is detailed in Appendix 5 with confidence interval presented in Tab. 12. Performance analysis with various architectures (Tab. 13) and lambda values (Tab. 14) are mentioned in Appendix 6. Threshold-based comparisons are in Appendix 7 and Tab. 15.", + "bbox": [ + 212, + 146, + 787, + 479 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 215, + 501, + 508, + 517 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we proposed an innovative graphical model for IDN noisy-label learning, focusing on the estimation of the label noise rate that is leveraged to introduce a robust sample selection curriculum. Given that SOTA IDN noisy-label learning approaches tend to rely on sample selection methods, our method can be seamlessly integrated to them to improve their performance in many synthetic and real-world benchmarks, including CIFAR100 [22], red mini-ImageNet [17], Clothing1M [45], mini-WebVision [26], and ImageNet [23].", + "bbox": [ + 212, + 532, + 787, + 638 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "From a societal perspective, our methodology offers a positive impact by mitigating biases inherent in noisy data, thereby promoting more fair and accurate machine learning outcomes. It also opens the door to future research avenues, including exploring advanced noise rate estimation methodologies and investigating noise-robust loss functions such as GCE [56] and ELR [28] with it. Our choice of using the same network architecture for clean and noisy labels enables compatibility with co-teaching techniques [16,26]. However, investigating different network structures for clean and noisy labels remains a potential avenue for future research. Moreover, we aim to delve into its dynamics, uncovering substantial improvements for certain models while observing more marginal enhancements for others. For more detailed information, please refer to Appendix 8. Our approach also hints at potential breakthroughs in streamlining data annotation challenges.", + "bbox": [ + 212, + 640, + 787, + 821 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "A. Garg et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Arpit, D., Jastrzebski, S., Ballas, N., Krueger, D., Bengio, E., Kanwal, M.S., Maharaj, T., Fischer, A., Courville, A., Bengio, Y., Lacoste-Julien, S.: A closer look at memorization in deep networks. In: International Conference on Machine Learning. vol. 70, pp. 233-242. PMLR (2017)", + "2. Bae, H., Shin, S., Jang, J., Na, B., Song, K., Moon, I.C.: From noisy prediction to true label: Noisy prediction calibration via generative model. In: International Conference on Machine Learning (2022)", + "3. Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.A.: Mixmatch: A holistic approach to semi-supervised learning. In: Advances in Neural Information Processing Systems. vol. 32 (2019)", + "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: International Conference on Computer Vision. pp. 9650-9660 (2021)", + "5. Chen, P., Ye, J., Chen, G., Zhao, J., Heng, P.A.: Beyond class-conditional assumption: A primary attempt to combat instance-dependent label noise. In: AAAI Conference on Artificial Intelligence. pp. 11442–11450 (2021)", + "6. Chen, Z., Song, A., Wang, Y., Huang, X., Kong, Y.: A noise rate estimation method for image classification with label noise. Journal of Physics: Conference Series 2433(1), 012039 (feb 2023). https://doi.org/10.1088/1742-6596/2433/1/012039, https://dx.doi.org/10.1088/1742-6596/2433/1/012039", + "7. Cheng, D., Liu, T., Ning, Y., Wang, N., Han, B., Niu, G., Gao, X., Sugiyama, M.: Instance-dependent label-noise learning with manifold-regularized transition matrix estimation. In: Conference on Computer Vision and Pattern Recognition. pp. 16630-16639 (2022)", + "8. Cheng, H., Zhu, Z., Li, X., Gong, Y., Sun, X., Liu, Y.: Learning with instance-dependent label noise: A sample sieve approach. In: International Conference on Learning Representations (2020)", + "9. Cordeiro, F.R., Belagiannis, V., Reid, I., Carneiro, G.: PropMix: Hard sample filtering and proportional mixup for learning with noisy labels. In: British Machine Vision Conference (2021)", + "0. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the EM algorithm. Journal of the royal statistical society: series B (methodological) 39(1), 1-22 (1977)", + "1. Feng, C., Tzimiropoulos, G., Patras, I.: SSR: An efficient and robust framework for learning with unknown label noise. In: British Machine Vision Conference (2022)", + "2. Feng, C., Ren, Y., Xie, X.: Ot-filter: An optimal transport filter for learning with noisy labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16164-16174 (2023)", + "3. Frénay, B., Kabán, A.: A comprehensive introduction to label noise. In: European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (2014)", + "4. Frénay, B., Verleysen, M.: Classification in the presence of label noise: A survey. IEEE Transactions on Neural Networks and Learning Systems 25(5), 845-869 (2013)", + "5. Garg, A., Nguyen, C., Felix, R., Do, T.T., Carneiro, G.: Instance-dependent noisy label learning via graphical modelling. In: Winter Conference on Applications of Computer Vision. pp. 2288-2298 (2023)" + ], + "bbox": [ + 225, + 178, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Noisy-Label Learning with Noise-Rate Estimation", + "bbox": [ + 397, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Han, B., Yao, Q., Yu, X., Niu, G., Xu, M., Hu, W., Tsang, I., Sugiyama, M.: Co-teaching: Robust training of deep neural networks with extremely noisy labels. In: Advances in Neural Information Processing Systems. vol. 31 (2018)", + "17. Jiang, L., Huang, D., Liu, M., Yang, W.: Beyond synthetic noise: Deep learning on controlled noisy labels. In: International Conference on Machine Learning. pp. 4804-4815. PMLR (2020)", + "18. Jiang, L., Zhou, Z., Leung, T., Li, L.J., Fei-Fei, L.: MentorNet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In: International Conference on Machine Learning. pp. 2304-2313. PMLR (2018)", + "19. Jiang, Z., Zhou, K., Liu, Z., Li, L., Chen, R., Choi, S.H., Hu, X.: An information fusion approach to learning with instance-dependent label noise. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=ecH2FKaARUp", + "20. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in Neural Information Processing Systems 33, 18661-18673 (2020)", + "21. Kim, T., Ko, J., Choi, J., Yun, S.Y.: FINE samples for learning with noisy labels. In: Advances in Neural Information Processing Systems. vol. 34 (2021)", + "22. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. Tech. rep., University of Toronto (2009)", + "23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: ImageNet classification with deep convolutional neural networks. In: Advances in Neural Information Processing Systems. vol. 25 (2012)", + "24. Kye, S.M., Choi, K., Yi, J., Chang, B.: Learning with noisy labels by efficient transition matrix estimation to combat label miscorrection. In: European Conference on Computer Vision. pp. 717-738. Springer (2022)", + "25. Li, J., Li, G., Liu, F., Yu, Y.: Neighborhood collective estimation for noisy label identification and correction. In: European Conference on Computer Vision. pp. 128-145. Springer (2022)", + "26. Li, J., Socher, R., Hoi, S.C.: DivideMix: Learning with noisy labels as semi-supervised learning. In: International Conference on Learning Representations (2020)", + "27. Li, W., Wang, L., Li, W., Agustsson, E., Gool, L.V.: WebVision Database: Visual learning and understanding from web data. CoRR (2017)", + "28. Liu, S., Niles-Weed, J., Razavian, N., Fernandez-Granda, C.: Early-learning regularization prevents memorization of noisy labels. In: Advances in Neural Information Processing Systems. vol. 33, pp. 20331-20342 (2020)", + "29. Liu, Y., Cheng, H., Zhang, K.: Identifiability of label noise transition matrix. In: International Conference on Machine Learning (2023)", + "30. Ma, X., Huang, H., Wang, Y., Romano, S., Erfani, S., Bailey, J.: Normalized loss functions for deep learning with noisy labels. In: International Conference on Machine Learning. pp. 6543-6553. PMLR (2020)", + "31. Malach, E., Shalev-Shwartz, S.: Decoupling \"when to update\" from \"how to update\". In: Advances in Neural Information Processing Systems. vol. 30 (2017)", + "32. McLachlan, G.J., Peel, D.: Finite mixture models. Wiley (1996)", + "33. Neal, R.M., Hinton, G.E.: A view of the EM algorithm that justifies incremental, sparse, and other variants. Learning in graphical models pp. 355-368 (1998)", + "34. Neyshabur, B., Bhojanapalli, S., McAllester, D., Srebro, N.: Exploring generalization in deep learning. In: Advances in Neural Information Processing Systems. vol. 30 (2017)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "A. Garg et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Nishi, K., Ding, Y., Rich, A., Hollerer, T.: Augmentation strategies for learning with noisy labels. In: Conference on Computer Vision and Pattern Recognition. pp. 8022-8031 (2021)", + "36. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: PyTorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems 32 (2019)", + "37. Shamshad, F., Khan, S., Zamir, S.W., Khan, M.H., Hayat, M., Khan, F.S., Fu, H.: Transformers in medical imaging: A survey. Medical Image Analysis p. 102802 (2023)", + "38. Smart, B., Carneiro, G.: Bootstrapping the relationship between images and their clean and noisy labels. In: Winter Conference on Applications of Computer Vision. pp. 5344-5354 (2023)", + "39. Song, H., Kim, M., Park, D., Shin, Y., Lee, J.G.: Learning from noisy labels with deep neural networks: A survey. IEEE Transactions on Neural Networks and Learning Systems (2022)", + "40. Tanaka, D., Ikami, D., Yamasaki, T., Aizawa, K.: Joint optimization framework for learning with noisy labels. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018)", + "41. Titterington, D.M.: On the likelihood of mixtures of distributions. Biometrika 71(3), 511-522 (1984). https://doi.org/10.1093/biomet/71.3.511", + "42. Trummer, I.: From BERT to GPT-3 codex: harnessing the potential of very large language models for data management. VLDB Endowment 15(12), 3770-3773 (2022)", + "43. Tu, H., Menzies, T.: Debtfree: minimizing labeling cost in self-admitted technical debt identification using semi-supervised learning. Empirical Software Engineering 27(4), 80 (2022)", + "44. Xia, X., Liu, T., Han, B., Wang, N., Gong, M., Liu, H., Niu, G., Tao, D., Sugiyama, M.: Part-dependent label noise: Towards instance-dependent label noise. In: Advances in Neural Information Processing Systems. vol. 33, pp. 7597-7610 (2020)", + "45. Xiao, T., Xia, T., Yang, Y., Huang, C., Wang, X.: Learning from massive noisy labeled data for image classification. In: Conference on Computer Vision and Pattern Recognition. pp. 2691-2699 (2015)", + "46. Xu, Y., Zhu, L., Jiang, L., Yang, Y.: Faster meta update strategy for noise-robust deep learning. In: Conference on Computer Vision and Pattern Recognition. pp. 144-153 (June 2021)", + "47. Yang, S., Jiang, L., Liu, Z., Loy, C.C.: Unsupervised image-to-image translation with generative prior. In: Conference on Computer Vision and Pattern Recognition. pp. 18332-18341 (2022)", + "48. Yang, S., Yang, E., Han, B., Liu, Y., Xu, M., Niu, G., Liu, T.: Estimating instance-dependent bayes-label transition matrix using a deep neural network. In: International Conference on Machine Learning. pp. 25302-25312. PMLR (2022)", + "49. Yao, Q., Yang, H., Han, B., Niu, G., Kwok, J.T.Y.: Searching to exploit memorization effect in learning with noisy labels. In: International Conference on Machine Learning. pp. 10789-10798. PMLR (2020)", + "50. Yao, Y., Liu, T., Gong, M., Han, B., Niu, G., Zhang, K.: Instance-dependent label-noise learning under a structural causal model. In: Advances in Neural Information Processing Systems. vol. 34 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Noisy-Label Learning with Noise-Rate Estimation", + "bbox": [ + 397, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "51. Yao, Y., Liu, T., Han, B., Gong, M., Deng, J., Niu, G., Sugiyama, M.: Dual T: Reducing estimation error for transition matrix in label-noise learning. In: Advances in Neural Information Processing Systems. vol. 33, pp. 7260-7271 (2020)", + "52. Zhang, C., Bengio, S., Hardt, M., Recht, B., Vinyals, O.: Understanding deep learning requires rethinking generalization. In: International Conference on Learning Representations (2017)", + "53. Zhang, C., Bengio, S., Hardt, M., Recht, B., Vinyals, O.: Understanding deep learning (still) requires rethinking generalization. Communications of the ACM 64(3), 107-115 (2021)", + "54. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. In: International Conference on Learning Representations (2017)", + "55. Zhang, Y., Niu, G., Sugiyama, M.: Learning noise transition matrix from only noisy labels via total variation regularization. In: International Conference on Machine Learning. pp. 12501-12512. PMLR (2021)", + "56. Zhang, Z., Sabuncu, M.: Generalized cross entropy loss for training deep neural networks with noisy labels. Advances in Neural Information Processing Systems 31 (2018)", + "57. Zhang, Z., Pfister, T.: Learning fast sample re-weighting without reward data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 725-734 (2021)", + "58. Zhao, G., Li, G., Qin, Y., Liu, F., Yu, Y.: Centrality and consistency: two-stage clean samples identification for learning with instance-dependent noisy labels. In: European Conference on Computer Vision. pp. 21-37. Springer (2022)", + "59. Zheltonozhskii, E., Baskin, C., Mendelson, A., Bronstein, A.M., Litany, O.: Contrast to divide: Self-supervised pre-training for learning with noisy labels. In: Winter Conference on Applications of Computer Vision. pp. 1657-1667 (2022)", + "60. Zhu, Z., Liu, T., Liu, Y.: A second-order approach to learning with instance-dependent label noise. In: Conference on Computer Vision and Pattern Recognition. pp. 10113-10123 (June 2021)" + ], + "bbox": [ + 215, + 146, + 787, + 550 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "A. Garg et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation/8698146c-3791-42ec-912e-34881bbacb25_model.json b/2024/Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation/8698146c-3791-42ec-912e-34881bbacb25_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4a47d0cc484d39e84548c1dbdf24a888120fbe9b --- /dev/null +++ b/2024/Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation/8698146c-3791-42ec-912e-34881bbacb25_model.json @@ -0,0 +1,2304 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.239, + 0.141, + 0.768, + 0.187 + ], + "angle": 0, + "content": "Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.212, + 0.768, + 0.245 + ], + "angle": 0, + "content": "Arpit Garg1, Cuong Nguyen3, Rafael Felix1, Thanh-Toan Do2, and Gustavo Carneiro3" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.255, + 0.776, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\) Australian Institute for Machine Learning, University of Adelaide, SA, Australia \\(^{2}\\) Department of Data Science and AI, Monash University, Australia" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.283, + 0.756, + 0.297 + ], + "angle": 0, + "content": "3 Centre for Vision, Speech and Signal Processing, University of Surrey, UK" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.336, + 0.744, + 0.545 + ], + "angle": 0, + "content": "Abstract. Deep learning faces a formidable challenge when handling noisy labels, as models tend to overfit samples affected by label noise. This challenge is further compounded by the presence of instance-dependent noise (IDN), a realistic form of label noise arising from ambiguous sample information. To address IDN, Label Noise Learning (LNL) incorporates a sample selection stage to differentiate clean and noisy-label samples. This stage uses an arbitrary criterion and a pre-defined curriculum that initially selects most samples as noisy and gradually decreases this selection rate during training. Such curriculum is sub-optimal since it does not consider the actual label noise rate in the training set. This paper addresses this issue with a new noise-rate estimation method that is easily integrated with most state-of-the-art (SOTA) LNL methods to produce a more effective curriculum. Synthetic and real-world benchmarks' results demonstrate that integrating our approach with SOTA LNL methods improves accuracy in most cases.\\(^4\\)" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.558, + 0.744, + 0.587 + ], + "angle": 0, + "content": "Keywords: Noisy-labels \\(\\cdot\\) Instance-dependent noise \\(\\cdot\\) Label noise Learning" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.613, + 0.377, + 0.629 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.79, + 0.797 + ], + "angle": 0, + "content": "Deep neural networks (DNNs) have demonstrated their effectiveness in various domains, including vision [47], language [42], and medicine [37]. However, its efficacy is largely dependent on high-quality training data, which can be resource-intensive to obtain [43]. Although cost-effective labelling techniques, such as data mining [11] and crowdsourcing [39], offer cheaper alternatives, they often compromise label quality [39]. Consequently, this can lead to erroneous labels in real-world datasets [21]. This is a relevant issue because even slight inaccuracies in the labels can significantly affect the performance of DNN due to their inherent memorisation capabilities [34, 53]. This has prompted the development of algorithms resilient to noisy-label learning, with the aim of training models" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.806, + 0.788, + 0.84 + ], + "angle": 0, + "content": "\\(^{4}\\) Code is available at https://github.com/arpit2412/NoiseRateLearning. Supported by the Engineering and Physical Sciences Research Council (EPSRC) through grant EP/Y018036/1 and the Australian Research Council (ARC) through grant FT190100525." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "A. Garg et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "despite inaccuracies in training data labels. Although various strategies exist, our work uniquely proposes an approach centered on the estimation of the label noise rate from the training data, which is notably absent from current methodologies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.196, + 0.789, + 0.59 + ], + "angle": 0, + "content": "There are two types of label noise, namely instance-independent noise (IIN) [16] and instance-dependent noise (IDN) [44]. Such type of label noise typically influences the design principles of noisy-label learning algorithms. For example, IIN focuses on mislabellings that are independent of sample information [16], where estimating the underlying label transition matrix is a common way of handling this type of noise [51]. On the other hand, in the more realistic IDN, mislabelling is due to both sample information and true class labels [44], which generally require the combination of many label noise learning techniques, such as robust loss functions [28, 56], and noisy-label sample selection [26, 58]. Of the strategies mentioned above, sample selection approaches that classify training data into clean and noisy samples have produced competitive results on many benchmarks [9, 11, 15, 21, 26]. Such sample selection techniques require the definition of a classification criterion and a selection curriculum. Many studies on this topic focus on developing new sample selection criteria, such as the small-loss hypothesis [26], which states that noisy-label samples have higher loss values than clean-label samples, particularly at the early stage of training [1]. Another criterion type is the feature-based one. An example of such criterion is FINE [21] that discriminates clean and noisy-label samples via the distance to class-specific eigenvectors. In this technique, clean-label samples tend to lie closer to the class-specific dominant eigenvector of the latent representations than noisy-label samples. Another type of criterion is proposed in SSR [11], which introduces a selection criterion based on the K-nearest-neighbor (KNN) classification in the feature space. Furthermore, CC [58] uses a two-stage sampling procedure, including class-level feature clustering followed by a consistency score. An equally important problem in sample selection is the definition of the curriculum for selecting clean training samples, but it has received comparatively less attention." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.79, + 0.777 + ], + "angle": 0, + "content": "The sample selection curriculum defines a threshold to be used with one of the criteria listed above to classify the training samples as clean or noisy at each training epoch [45]. For example, the threshold can be fixed to an arbitrary clustering score that separates clean and noisy samples [26], but this strategy does not account for the proportion of label noise in the training set, nor does it consider the dynamics of the selection of noisy-label samples during the training. The consideration of such dynamics has been studied in [16, 49], which defined a curriculum of the noisy-label sampling rate \\( R(t) \\) as a function of the training epoch \\( t \\in \\{1, \\dots, T\\} \\). The curriculum \\( R(t) \\) defines a sampling rate close to \\( 100\\% \\) of the training set at the beginning of the training, which is then reduced to arbitrarily low rates at the end of the training. In practice, the function \\( R(t) \\) is predefined [16] or learned by weighting a set of basis functions [49]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Although generally effective, these techniques do not consider the label noise rate estimated from the training set, making them vulnerable to over-fitting (if too many noisy-label samples are classified as clean) or under-fitting (if informative clean-label samples are classified as noisy). It can be argued that the estimation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.398, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Noisy-Label Learning with Noise-Rate Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.493, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.309, + 0.403, + 0.32 + ], + "angle": 0, + "content": "(a) Motivation" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.143, + 0.77, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.309, + 0.733, + 0.32 + ], + "angle": 0, + "content": "(b) Proposed Graphical Model" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.443 + ], + "angle": 0, + "content": "Fig. 1: (a) Comparison of test accuracy \\(\\%\\) (as a function of training epoch) between the original DivideMix [26] (solid, blue curve) and our modified DivideMix (dashed, red curve) that selects the clean and noisy data based on a fixed noise rate \\(R(t) = 1 - \\epsilon = 50\\%\\) using the small-loss criterion on CIFAR100 [22] at 0.5 IDN [44]; (b) The proposed probabilistic graphical model that generates noisy-label \\(\\hat{Y}\\) conditioned on the image \\(X\\), the latent clean-label \\(Y\\) and noise rate \\(\\epsilon\\), where forward pass (solid lines) is parameterised by \\(\\theta_y, \\theta_{\\hat{y}}\\) and \\(\\epsilon\\) representing the generation step, and the backward pass (dashed lines) is parameterised by \\(\\rho\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.788, + 0.596 + ], + "angle": 0, + "content": "of the label transition matrix [6,7,48,51] aims to recover the noise rate affecting pairwise label transitions. However, label transition matrix techniques follow a quite different strategy compared to sample selection methods, where their main challenge is the general underconstrained aspect of the matrix estimation, making them sensitive to large noise rates and not scalable to a high number of classes [39]. Although several methods address label noise, none estimates the label noise rate directly from the training data to guide sample selection. Addressing this gap is the primary challenge of our study." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.793 + ], + "angle": 0, + "content": "To underscore the importance of using the noise rate for sample selection during training, we experiment with CIFAR100 [22] at an IDN rate \\(\\epsilon = 50\\%\\) [44] (noise rate specifications and other details are explained in Sec. 4). We use DivideMix [26] and replace its sample selection (based on an arbitrary clustering score [10,32,41]) by a thresholding process that classifies the \\(R(t) = 1 - \\epsilon = 50\\%\\) largest loss samples as noisy, and the remaining ones as clean in all training epochs \\(t \\in \\{1, \\dots, T\\}\\). This sample selection is used for the semi-supervised learning of DivideMix [26]. As shown in Fig. 1a, the new sample selection approach based on the \"manually provided\" noise rate (dashed red curve) improves \\(6\\%\\) in terms of prediction accuracy compared to the original DivideMix [26] (solid blue curve) which relies on arbitrary thresholding. Similar conclusions can be achieved with other methods that apply sample selection strategies to address the noisy-label learning problem, as shown in the experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this paper, we introduce a new sample selection strategy centered on estimating the label noise rate of the training set. Our strategy is based on our novel noisy-label learning graphical model illustrated in Fig. 1b. This model can be" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "A. Garg et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.147, + 0.486, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.289, + 0.443, + 0.302 + ], + "angle": 0, + "content": "(a) Curriculum function \\(R(t)\\)" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.143, + 0.75, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.289, + 0.722, + 0.302 + ], + "angle": 0, + "content": "(b) Accuracy Comparison" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.312, + 0.789, + 0.438 + ], + "angle": 0, + "content": "Fig. 2: (a) Visual comparison of different \\( R(t) \\) on CIFAR100 [22] at 0.5 IDN [44]: (i) Co-teaching [16] with curriculum based on hyper-parameter \\( T_{k} \\in \\{5,10\\} \\), where \\( R(t) = 1 - \\tau \\cdot \\min(t / T_{k}, 1) \\) with \\( \\tau = \\epsilon = 0.5 \\) being manually set; (ii) S2E [49], where \\( R(t) \\) is estimated with a bi-level optimisation; (iii) SSR [11], where \\( R(t) \\) is based on a relabelling pattern, (iv) CC [58] following the small-loss hypothesis [26] and cosine similarity for \\( R(t) \\); and (v) ours. (b) Accuracy of noisy-label learning methods on CIFAR100 [22] at 0.5 IDN [44], including DivideMix [26], FINE [21] and InstanceGM [15], without (left, blue) and (right, orange) integration of our proposed graphical model for estimating the noise rate \\( \\epsilon \\) and sample selection rate based on \\( \\epsilon \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.468, + 0.789, + 0.604 + ], + "angle": 0, + "content": "seamlessly integrated with state-of-the-art (SOTA) noisy-label learning techniques, providing them with a precise noise rate estimate and subsequently refining the sample selection curriculum. In particular, our model's curriculum is not restrained by predefined \\( R(t) \\) functions [16, 49], but rather relies on a dynamically estimated noise rate sourced directly from the training set, as depicted in Fig. 2a. Our method dynamically estimates training set noise rates to mitigate overfitting and underfitting, seamlessly integrates with existing algorithms, and serves as a robust foundation for future noisy-label learning research. To summarise, our main contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.617, + 0.784, + 0.66 + ], + "angle": 0, + "content": "- An innovative noisy-label learning graphical model, shown in Fig. 1b, that not only estimates but also leverages the noise rate from the training dataset to produce a refined sample selection curriculum for SOTA LNL methods." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.663, + 0.784, + 0.72 + ], + "angle": 0, + "content": "- A simple and synergistic integration strategy of our novel graphical model with several SOTA noisy-label learning algorithms, such as DivideMix [26] and SSR [11], to improve their sample selection effectiveness, leading to an increased test accuracy, as shown in Fig. 2b." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.617, + 0.784, + 0.72 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Empirical evaluations show the role of our proposed sample selection methodology in improving the efficacy of leading noisy-label learning algorithms across various synthetic and real-world IDN benchmarks. Although our primary focus lies in addressing IDN problems, because of its more challenging and realistic nature, we have also applied our model to IIN problems, with detailed results in Tab. 6 and discussion in Appendix 3 of the supplementary material. With our innovative approach, we aim to improve the accuracy of SOTA LNL methods." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.398, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Noisy-Label Learning with Noise-Rate Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.789, + 0.674 + ], + "angle": 0, + "content": "It is well-known that DNNs suffer from overfitting when trained with noisy-labels [52], resulting in poor generalisation [28, 55]. To mitigate the problems related to noisy-label learning, several techniques have been developed, including noise-robust loss functions [30], noise-label sample selection [21, 26, 40, 58, 59], meta-learning [57], and re-labeling followed by sample selection [11]. These techniques can handle many types of label noise (e.g., symmetric and asymmetric), but the most challenging type, i.e., IDN [39], tends to be successfully addressed with methods that have a training stage based on sample selection techniques [21, 26, 40, 58, 59]. These sample selection techniques separate clean and noisy-label samples, where clean samples are treated as labelled, and noisy samples are discarded [8, 16, 18, 49] or treated as unlabeled samples [9, 15, 26, 35, 59] for semi-supervised learning [3]. Another solution consists of aligning clean and noisy samples with an information fusion method [19]. A major limitation of these approaches is the overly simplistic design of the curriculum to select clean and noisy-label samples during training. It is either predetermined [16, 18] or learned from a set of predetermined basis functions [49]. We argue that the design of a curriculum based on an estimated noise rate would benefit such methods, as motivated by Fig. 1a, but to the best of our knowledge, such estimation has not been explored by previous methods. The closest idea explored is based on the estimation of the type of noise instead of the noise rate [45]. It is possible to argue that the estimation of the label transition matrices [5, 7, 44, 60], composed of the learned noise rate that affects the transition between pairs of labels, is a way to estimate the label noise rate. However, they portray results of comparatively lower accuracy for large real-world datasets or for IDN problems [39, 44]. A possible reason for these poorer results is that label transition approaches suffer from identifiability issues [14], where any clean-label distribution assignment is acceptable as long as the distribution of observed labels can be reconstructed [13]. This makes the identification of the actual underlying clean-labels challenging. One solution is to consider the use of multiple annotations per sample to help analyse agreements and disagreements for improved identification of clean patterns [14, 31]. However, the annotation of datasets with a single label per training sample is already challenging; the complexity is significantly larger to acquire multiple labels per sample." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.79, + 0.842 + ], + "angle": 0, + "content": "An alternative technique to handle IDN problems is based on graphical models that represent the relationship between various observed and latent variables [15,45,50]. The approaches in [15,50] use graphical models that rely on a generative approach to produce noisy-labels from the respective image features and latent clean-labels. However, previous graphical models do not take into account the underlying noise rate parameter during modeling. Our work is the first graphical model approach to estimate the noise rate of the dataset. In addition, a close examination of existing literature reveals a trend: when traditional graphical models are integrated with SOTA noisy-label learning methods, the outcomes often fall short of expectations in terms of accuracy [2,15,50]. This observation underscores a strong point of our work. Our approach is uniquely tailored for" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "A. Garg et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.178 + ], + "angle": 0, + "content": "easy integration with current SOTA noisy-label learning methods for sample selection, which, in turn, improves over the SOTA classification accuracy." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.2, + 0.331, + 0.216 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.232, + 0.79, + 0.354 + ], + "angle": 0, + "content": "In this section, we present our new graphical model that estimates the noise rate, which will be used in the sample selection process. Let \\(\\mathcal{D} = \\{(x_i,\\hat{y}_i)\\}_{i = 1}^N\\) be the noisy-label training set containing \\(d\\)-dimensional data vector \\(x_{i}\\in \\mathcal{X}\\subseteq \\mathbb{R}^{d}\\) and its respective \\(C\\)-dimensional one-hot encoded observed (potentially corrupted) label \\(\\hat{y}_i\\in \\hat{\\mathcal{V}} = \\{\\hat{y}:\\hat{y}\\in \\{0,1\\} ^C\\wedge \\mathbf{1}_C^\\top \\hat{y} = 1\\}\\), where \\(\\mathbf{1}_C\\) is a vector of ones with \\(C\\) dimensions. The aim is to estimate the label noise rate \\(\\epsilon\\), used for the generation of noisy-label training data from the observed training dataset \\(\\mathcal{D}\\) and integrate this label noise rate into the sample selection strategy." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.375, + 0.405, + 0.39 + ], + "angle": 0, + "content": "3.1 Graphical Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.789, + 0.475 + ], + "angle": 0, + "content": "We portray the generation of noisy-label via the probabilistic graphical model shown in Fig. 1b. The observed random variables, denoted by shaded circles, are data \\( X \\) and the corresponding noisy-label \\( \\hat{Y} \\). We also have one latent variable, namely: the clean-label \\( Y \\). Under our proposed modeling assumption, a noisy-label of a data instance can be generated as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.487, + 0.576, + 0.502 + ], + "angle": 0, + "content": "- sample an instance from \\( p(X) \\), i.e.: \\( x \\sim p(X) \\)" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.502, + 0.631, + 0.517 + ], + "angle": 0, + "content": "- sample a clean-label from the clean-label distribution:" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.487, + 0.631, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.245, + 0.518, + 0.435, + 0.533 + ], + "angle": 0, + "content": "\\[\ny \\sim \\operatorname {C a t e g o r i c a l} (Y; f _ {\\theta_ {y}} (x))\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.533, + 0.632, + 0.547 + ], + "angle": 0, + "content": "- sample a noisy-label from the noisy-label distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.245, + 0.547, + 0.614, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\hat {y} \\sim \\operatorname {C a t e g o r i c a l} (\\hat {Y}; \\epsilon \\times f _ {\\theta_ {\\hat {y}}} (x, f _ {\\theta_ {y}} (x)) + (1 - \\epsilon) \\times y),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.789, + 0.698 + ], + "angle": 0, + "content": "where \\( \\text{Categorical}() \\) denotes a categorical distribution, \\( f_{\\theta_y}: \\mathcal{X} \\to \\Delta_{C-1} \\) and \\( f_{\\theta_{\\hat{y}}}: \\mathcal{X} \\times \\Delta_{C-1} \\to \\Delta_{C-1} \\) denote two classifiers for the clean-label \\( Y \\) and noisy-label \\( \\hat{Y} \\), respectively, with \\( \\Delta_{C-1} = \\{s : s \\in [0,1]^C \\wedge \\mathbf{1}_C^\\top s = 1\\} \\) being the \\( (C-1) \\)-dimensional probability simplex. According to the data generation process shown in Fig. 1b, \\( \\epsilon \\) corresponds to \\( \\mathbb{E}_{(x,\\hat{y}) \\sim p(X,\\hat{Y})}[P(\\hat{y} \\neq y|x)] \\), which is the label noise rate of the training dataset of interest. Our aim is to infer the parameters \\( \\theta_y, \\theta_{\\hat{y}} \\) and \\( \\epsilon \\) from a noisy-label dataset \\( \\mathcal{D} \\) by maximising the following log-likelihood:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.709, + 0.788, + 0.749 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon} \\mathbb {E} _ {(x _ {i}, \\hat {y} _ {i}) \\sim \\mathcal {D}} [ \\ln p (\\hat {y} _ {i} | x _ {i}; \\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon) ] = \\max _ {\\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon} \\mathbb {E} _ {(x _ {i}, \\hat {y} _ {i}) \\sim \\mathcal {D}} [ \\ln \\sum_ {y _ {i}} p (\\hat {y} _ {i}, y _ {i} | x _ {i}; \\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon) ]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Due to the presence of the clean-label \\( y_{i} \\), it is difficult to directly evaluate the log-likelihood in Eq. (1). Therefore, we employ the expectation-maximisation (EM) algorithm [10] to maximise the log-likelihood. The main idea of the EM algorithm is to (i) construct a tight lower bound of the likelihood in Eq. (1) by estimating the posterior of the latent variable \\( Y \\) (known as expectation step) and (ii) maximise that lower bound (known as maximisation step). Formally, let" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.398, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Noisy-Label Learning with Noise-Rate Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "\\(q(y_{i}|x,\\hat{y};\\rho)\\) be an arbitrary distribution over a clean-label \\(y_{i}\\). The evidence lower bound (ELBO) of the log-likelihood in Eq. (1) can be obtained through Jensen's inequality and is presented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.282, + 0.204, + 0.787, + 0.278 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} Q \\left(\\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon , \\rho\\right) = \\mathbb {E} _ {\\left(x _ {i}, \\hat {y} _ {i}\\right) \\sim \\mathcal {D}} \\left[ \\mathbb {E} _ {q \\left(y _ {i} \\mid x _ {i}; \\hat {y} _ {i}; \\rho\\right)} \\left[ \\ln p \\left(\\hat {y} _ {i} \\mid x _ {i}, y _ {i}; \\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon\\right) \\right] \\right. \\\\ - \\operatorname {K L} \\left[ q \\left(y _ {i} \\mid x _ {i}, \\hat {y} _ {i}; \\rho\\right) \\| p \\left(y _ {i} \\mid x _ {i}; \\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon\\right) \\right] \\\\ = \\mathbb {E} _ {\\left(x _ {i}, \\hat {y} _ {i}\\right) \\sim \\mathcal {D}} \\left[ \\mathbb {E} _ {q \\left(y _ {i} \\mid x _ {i}; \\hat {y} _ {i}; \\rho\\right)} \\left[ \\ln p \\left(\\hat {y} _ {i} \\mid x _ {i}, y _ {i}; \\theta_ {\\hat {y}}, \\epsilon\\right) \\right. \\right. \\\\ \\left. + \\ln p \\left(y _ {i} \\mid x _ {i}; \\theta_ {y}\\right) \\right] + \\mathbb {H} \\left[ q \\left(y _ {i} \\mid x _ {i}, \\hat {y} _ {i}; \\rho\\right) \\right], \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.288, + 0.784, + 0.334 + ], + "angle": 0, + "content": "where \\(\\mathrm{KL}[q\\| p]\\) is the Kullback - Leibler divergence between distributions \\(q\\) and \\(p\\) and \\(\\mathbb{H}(q)\\) is the entropy of the distribution \\(q\\). The EM algorithm is then carried out iteratively by alternating the following two steps:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.785, + 0.456 + ], + "angle": 0, + "content": "E step: we maximise the ELBO in (2) w.r.t. \\( q(y_{i}|x_{i},\\hat{y}_{i};\\rho) \\). Theoretically, such an optimisation results in KL \\([q(y_{i}|x_{i},\\hat{y}_{i};\\rho)\\| p(y_{i}|x_{i},\\hat{y}_{i})] = 0\\) or \\( q(y_{i}|x_{i},\\hat{y}_{i};\\rho) = p(y_{i}|x_{i},\\hat{y}_{i}) \\). This is equivalent to estimating the posterior of the clean-label \\( y_{i} \\) given noisy-label data \\((x_{i},\\hat{y}_{i})\\). Obtaining the posterior \\( p(y_{i}|x_{i},\\hat{y}_{i}) \\) is, however, intractable for most deep-learning models. To mitigate such an issue, we follow the variational EM approach [33] by employing an approximate posterior \\( q(y_{i}|x_{i},\\hat{y}_{i};\\rho^{(t)}) \\) that is the closest to the true posterior \\( p(y_{i}|x_{i},\\hat{y}_{i}) \\), where:" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.466, + 0.787, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\rho^ {(t)} = \\arg \\max _ {\\rho} Q \\left(\\theta_ {y} ^ {(t)}, \\theta_ {\\hat {y}} ^ {(t)}, \\epsilon^ {(t)}, \\rho\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.784, + 0.551 + ], + "angle": 0, + "content": "with the superscript \\({}^{(t)}\\) denoting the parameters at the \\(t\\)-th iteration. Although this results in a non-tight lower bound of the log-likelihood in Eq. (1), it does increase the variational bound \\(Q\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.784, + 0.595 + ], + "angle": 0, + "content": "M step: we maximise the ELBO in Eq. (2) w.r.t. \\(\\theta_y, \\theta_{\\hat{y}}\\) and \\(\\epsilon\\) given \\(\\rho^{(t)}\\) obtained in the E step:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.606, + 0.787, + 0.633 + ], + "angle": 0, + "content": "\\[\n\\theta_ {y} ^ {(t + 1)}, \\theta_ {\\hat {y}} ^ {(t + 1)}, \\epsilon^ {(t + 1)} = \\arg \\max _ {\\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon} Q (\\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon , \\rho^ {(t)}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.785, + 0.734 + ], + "angle": 0, + "content": "The estimated noise rate \\(\\epsilon\\) can then be integrated into certain noisy-label algorithms to train the models of interest as mentioned in Sec. 1. In addition, the inference of noise rate \\(\\epsilon\\) might be associated with the identifiability issue when estimating the clean-label \\(Y\\) [29], i.e., there exists multiple sets of \\(\\rho\\) and \\(\\theta_y\\), where each set can explain the observed noisy-label data equally well. Such issues are addressed in the following subsection." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.755, + 0.403, + 0.77 + ], + "angle": 0, + "content": "3.2 Sample Selection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.843 + ], + "angle": 0, + "content": "The identifiability issue when inferring the clean-label \\( Y \\) from noisy-label data \\( (X, \\hat{Y}) \\) can be mitigated either by acquiring multiple noisy-labels [29] or introducing additional constraints, such as small loss hypothesis [16] or FINE [21]. Since requesting additional noisy-labels per training sample is not always possible, we" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "A. Garg et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.216, + 0.146, + 0.787, + 0.176 + ], + "angle": 0, + "content": "Algorithm 1 Proposed noisy-label learning that relies on the estimation of noise rate \\(\\epsilon\\) to build a sample selection curriculum." + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.178, + 0.792, + 0.482 + ], + "angle": 0, + "content": "1: procedure NOISE RATE ESTIMATION AND INTEGRATION(D,T,λ) \n2: \\(\\triangleright\\) \\(\\mathcal{D} = \\{(x_i,\\hat{y}_i)\\}_{i = 1}^N\\) training set with noisy-label data \n3: \\(\\triangleright\\) T: number of epochs \n4: \\(\\triangleright\\) λ: a hyper-parameter \n5: \\(\\triangleright\\) Notation: \\(\\Theta^{(t)} = (\\theta_y^{(t)},\\theta_{\\hat{y}}^{(t)},\\epsilon^{(t)})\\) \n6: Initialise \\(\\theta_y^{(1)},\\theta_{\\hat{y}}^{(1)},\\epsilon^{(1)}\\) and \\(\\rho^{(0)}\\) \n7: \\(\\theta_y^{(1)}\\gets \\mathrm{WARM~UP}(\\mathcal{D},\\theta_y^{(1)})\\) \n8: \\(t\\gets 0\\) \n9: for \\(n_{\\mathrm{epoch}} = 1:T\\) do \n10: for each mini-batch \\(\\mathcal{S}\\) in SHUFFLE(D) do \n11: \\(t\\gets t + 1\\) \n12: \\(\\mathcal{S}_{\\mathrm{clean}},\\mathcal{S}_{\\mathrm{noisy}}\\gets \\mathrm{SAMPLE~SELECTION}(\\mathcal{S},\\theta_y^{(t)},\\epsilon^{(t)})\\) \n13: \\(\\triangleright\\) Variational E-step as in Eq. (3) \n14: \\(\\rho^{(t)}\\gets \\mathrm{E~STEP}(\\mathcal{S},\\theta_y^{(t)},\\theta_{\\hat{y}}^{(t)},\\epsilon^{(t)},\\rho^{(t - 1)})\\) \n15: \\(\\triangleright\\) M-step as in Eq. (8) \n16: \\(\\theta_y^{(t + 1)},\\theta_{\\hat{y}}^{(t + 1)},\\epsilon^{(t + 1)}\\gets \\mathrm{M~STEP}(\\mathcal{S}_{\\mathrm{clean}},\\mathcal{S}_{\\mathrm{noisy}},\\theta_y^{(t)},\\theta_{\\hat{y}}^{(t)},\\epsilon^{(t)},\\rho^{(t)},\\lambda)\\) \n17: return \\(\\theta_y\\) \n1 parameter of the clean-label classifier" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.512, + 0.788, + 0.571 + ], + "angle": 0, + "content": "follow the latter approach by imposing a constraint (scoring term), denoted as \\( L(\\theta_y, \\epsilon^{(t)}) \\), over \\( \\theta_y \\) in the M step via a sample selection approach based on the estimated noise rate \\( \\epsilon^{(t)} \\). Formally, we propose a new curriculum when selecting samples as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.446, + 0.571, + 0.787, + 0.589 + ], + "angle": 0, + "content": "\\[\nR (t) = 1 - \\epsilon^ {(t)}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.787, + 0.625 + ], + "angle": 0, + "content": "In the simplest case, such as Co-teaching [16] or FINE [21], the constraint for \\(\\theta_y\\) can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.633, + 0.788, + 0.668 + ], + "angle": 0, + "content": "\\[\nL \\left(\\theta_ {y}, \\epsilon^ {(t)}\\right) = \\sum_ {\\left(x _ {i}, \\hat {y} _ {i}\\right) \\in \\mathcal {S} _ {\\text {c l e a n}}} \\operatorname {K L} \\left[ \\operatorname {C a t e g o r i c a l} (Y; \\hat {y}) \\| \\operatorname {C a t e g o r i c a l} \\left(Y; f _ {\\theta_ {y}} \\left(x _ {i}\\right)\\right) \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.678, + 0.267, + 0.69 + ], + "angle": 0, + "content": "where:" + }, + { + "type": "equation", + "bbox": [ + 0.257, + 0.7, + 0.61, + 0.718 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {\\text {c l e a n}} = \\left\\{\\left(x _ {i}, \\hat {y} _ {i}\\right): \\left(x _ {i}, \\hat {y} _ {i}\\right) \\in \\mathcal {D} \\wedge z \\left(x _ {i}, y _ {i}\\right) \\leq m \\right\\},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.258, + 0.719, + 0.743, + 0.736 + ], + "angle": 0, + "content": "\\[\nm \\in \\{m: \\Pr (z \\left(x _ {i}, y _ {i}\\right) \\leq m) \\geq R (t) \\wedge \\Pr (z \\left(x _ {i}, y _ {i}\\right) \\geq m) \\geq 1 - R (t) \\},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.258, + 0.738, + 0.787, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {\\text {n o i s y}} = \\mathcal {D} \\backslash \\mathcal {S} _ {\\text {c l e a n}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.843 + ], + "angle": 0, + "content": "with \\( R(t) \\) defined in (5), and \\( z(x,y) \\) representing the score of a criterion (e.g., loss [26, 58, 59], distance to the largest eigenvectors [21], or KNN scores [11]). Intuitively, the loss in Eq. (6) is simply the cross-entropy loss on the \\( \\lfloor R(t)\\times N\\rfloor \\) samples that have smallest scores (with \\( \\lfloor .\\rfloor \\) being the floor function). One can also extend to other SOTA models by replacing the loss \\( L \\) accordingly. For example," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.398, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Noisy-Label Learning with Noise-Rate Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.144, + 0.773, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.327, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Fig. 3: Our training algorithm uses an existing noisy-label classifier (e.g., DivideMix [26]) parameterised by \\(\\theta_y\\) as the clean-label model \\(p(Y|X;\\theta_y)\\). The generation of noisy-label (given \\(X\\) and \\(Y\\)) is carried out by a model parameterised by noise rate \\(\\epsilon\\) and \\(\\theta_{\\tilde{y}}\\). The noisy-label classifier is based on a sample-selection mechanism that uses a curriculum \\(R(t) = 1 - \\epsilon^{(t)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.429, + 0.788, + 0.462 + ], + "angle": 0, + "content": "if DivideMix [26] is used as a base model to constrain \\(\\theta_y\\), \\(L\\) will include two additional terms: loss on un-labeled data and regularisation using mixup [54]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.488, + 0.436, + 0.504 + ], + "angle": 0, + "content": "3.3 Training and Testing" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.52, + 0.787, + 0.551 + ], + "angle": 0, + "content": "Given the sample selection approach in Sec. 3.2, the M step in Eq. (4) is slightly modified by integrating the constraint in Eq. (6), which can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.265, + 0.564, + 0.788, + 0.592 + ], + "angle": 0, + "content": "\\[\n\\theta_ {y} ^ {(t + 1)}, \\theta_ {\\hat {y}} ^ {(t + 1)}, \\epsilon^ {(t + 1)} = \\arg \\max _ {\\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon} Q (\\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon , \\rho^ {(t)}) - \\lambda L (\\theta_ {y}, \\epsilon^ {(t)}), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.607, + 0.788, + 0.681 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyper-parameter and \\(L\\) is defined in Eq. (6). The training procedure is summarised in Algorithm 1 and visualised in Fig. 3. In the implementation, we integrate the proposed method into existing models, such as DivideMix [26] or FINE [21]. Note that the clean-label classifier \\(f_{\\theta_y}(.)\\) is also the clean classifier of the base model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.712, + 0.377, + 0.729 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We show extensive experiments in several noisy-label synthetic benchmarks with CIFAR100 [22], and real-world benchmarks, including CNWL's red mini-ImageNet [17], Clothing1M [45] and mini-WebVision [27]. Sec. 4.2 describes implementation details. We evaluate our approach by plugging SOTA models into \\( p(y|x;\\theta_y) \\), defined in Sec. 3, with results being shown in Sec. 4.3, and ablation studies in Sec. 4.4." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "A. Garg et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.434, + 0.163 + ], + "angle": 0, + "content": "4.1 Dataset Descriptions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.79, + 0.507 + ], + "angle": 0, + "content": "We follow the existing literature to generate the IDN labels [44] for CIFAR100 [22]. The dataset contains 50,000 training images and 10,000 testing images of shape \\(32 \\times 32 \\times 3\\). The dataset is class-balanced with 100 different categories and the IDN rates considered are 0.2, 0.3, 0.4 and 0.5 [44]. Another important IDN benchmark dataset is the red mini-ImageNet, a real-world dataset from CNWL [17]. In this dataset, there are 100 distinct classes, with each class containing 600 images. The images and their corresponding noisy-labels have been crawled from the internet at various controllable label noise rates. For a fair comparison with the existing literature [9, 15, 46], we have resized the images to \\(32 \\times 32\\) pixels from the \\(84 \\times 84\\) original pixel settings. We show results with \\(40\\%\\), \\(60\\%\\) and \\(60\\%\\) noise rates on this dataset. Clothing1M [45] is a real-world clothing classification dataset that contains 1 million images with an estimated noise rate of \\(38.5\\%\\). The dataset contains 14 different categories, and the labels are generated from surrounding texts. Images in this dataset are of different sizes, and we follow the resize structure suggested in [21, 26]. This dataset also contains \\(50k\\) manually validated clean training images, \\(14k\\) images for the validation set, and \\(10k\\) testing images. We have not used the clean training, validation set, or any extra training images while training. Only the testing set is used for evaluation. Mini-WebVision [26] contains 65,944 images with their respective labels from the 50 different initial categories from the WebVision [27], with image size of \\(256 \\times 256\\) pixels. Following the evaluation process commonly used in this benchmark, 50 categories from the ILSVRC12 [23] dataset are also used for testing." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.528, + 0.393, + 0.542 + ], + "angle": 0, + "content": "4.2 Implementation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.789, + 0.842 + ], + "angle": 0, + "content": "All methods are implemented in Pytorch [36] and use one NVIDIA RTX 3090 card for training and testing. As mentioned in the original papers, hyperparameter settings are kept the same for the baselines used in the proposed algorithm. All classifier architectures are also kept the same as the baseline models. A random initialisation of noise rate parameter \\(\\epsilon\\) with the sigmoid as its activation function is employed for all experiments to maintain the fairness of the comparisons with other approaches. The value of \\(\\lambda\\) in Eq. (8) is set to 1 for all the cases. We integrate many SOTA approaches [11, 15, 21, 26, 58, 59] into our graphical model, as explained in Sec. 3.3. For CIFAR100 [22] with IDN [44], we integrate DivideMix [26], C2D [59], CC [58], and InstanceGM [15] into our model, given their superior performance across various noise rates. Additionally, we also use F-Dividemix from FINE [21], as shown in Fig. 2b. For red mini-ImageNet [17], we test our proposed approach with and without DINO self-supervision [4]. For the implementation without self-supervision, we use DivideMix [26] and InstanceGM [15], and for the self-supervised version, we only use InstanceGM [15]. The models for Clothing1M [45] and mini-WebVision [26] (validation on ImageNet [23]) are trained using DivideMix [26], SSR [11], C2D [59] and CC [58]. Additional implementation details, empirical analysis, and computational analysis are presented in the supplementary material in Appendices 1, 2 and 4, respectively." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.398, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Noisy-Label Learning with Noise-Rate Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.144, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 1: Test accuracy % on (left) CIFAR100 [22] under different IDN [45], and (right) red mini-ImageNet [17]. Other models' results are from [7, 15, 46]. Here, we integrate DivideMix [26], C2D [59], CC [58] and InstanceGM [15] into our proposed model for CIFAR100 [22]. For red mini-ImageNet [17], we present the results with and without self-supervision [4]. We integrate DivideMix [26] and InstanceGM [15] into our model, with the latter tested with and without self-supervision." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.244, + 0.473, + 0.459 + ], + "angle": 0, + "content": "
MethodNoise Rates - IDN
0.20.30.40.5
CE [50]30.4224.1521.4514.42
PartT [44]65.3364.5659.7356.80
kMEIDTM [7]69.1666.7663.4659.18
DivideMix [26]77.0376.3370.8058.61
DivideMix-Ours77.4277.2172.4164.02
(+0.39)(+0.88)(+1.61)(+5.41)
C2D [59]78.6178.1872.8963.19
C2D-Ours79.0778.5973.3165.28
(+0.46)(+0.41)(+0.42)(+2.09)
CC [58]79.6177.5676.5863.19
CC-Ours79.7278.7177.3867.53
(+0.11)(+1.15)(+0.80)(+4.34)
InstanceGM [15]79.6979.2178.4777.19
InstanceGM-Ours79.6179.4079.5278.21
(-0.08)(+0.19)(+1.05)(+1.02)
" + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.24, + 0.778, + 0.462 + ], + "angle": 0, + "content": "
MethodNoise rate
0.40.60.8
CE [46]42.7037.3029.76
MixUp [54]46.4040.5833.58
MentorMix [17]47.1443.8033.46
FaMUS [46]51.4245.1035.50
DivideMix [26]46.7243.1434.50
DivideMix-Ours50.7045.1137.44
(+3.98)(+1.97)(+2.94)
InstanceGM [15]52.2447.9639.62
InstanceGM-Ours56.6151.4043.83
(+4.37)(+3.44)(+4.21)
With self-supervised learning
PropMix [9]56.2252.8443.42
InstanceGM-SS [15]56.3753.2144.03
InstanceGM-SS-Ours58.2953.6045.47
(+1.92)(+0.39)(+1.44)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.503 + ], + "angle": 0, + "content": "Table 2: Final estimated noise rate \\(\\epsilon\\) for the Tab. 1, (left) CIFAR100 [22] under various IDN [44], and (right) red mini-ImageNet [17] with and without self-supervision." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.514, + 0.428, + 0.605 + ], + "angle": 0, + "content": "
Estimated noise rates
MethodActual noise rate
0.20.30.40.5
DivideMix-Ours0.180.340.470.53
C2D-Ours0.190.330.380.52
CC-Ours0.210.340.410.54
InstanceGM-Ours0.230.370.420.47
" + }, + { + "type": "table", + "bbox": [ + 0.559, + 0.519, + 0.761, + 0.6 + ], + "angle": 0, + "content": "
Estimated noise rates
MethodActual noise rate
0.40.6
DivideMix-Ours0.390.58
InstanceGM-Ours0.380.55
InstanceGM-SS-Ours0.480.53
0.73
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.63, + 0.461, + 0.646 + ], + "angle": 0, + "content": "4.3 Comparison with SOTA" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.654, + 0.788, + 0.7 + ], + "angle": 0, + "content": "This section compares our approach with SOTA methods on datasets with the IDN settings and noisy real-world settings. The bold text in tables indicate SOTA results, and our results are in the greyed rows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Synthetic Instance-Dependent Noise The comparison between various baselines and our proposed method on CIFAR100 [22] with IDN [44] is shown in Tab. 1 (left) with the noise rate ranging from 0.2 to 0.5. It is worth noting that using our proposed model with DivideMix [26], C2D [59], CC [58] and InstanceGM [15] improve their performance in almost all cases, particularly with large noise rates. Tab. 2 (left) also shows the final noise rate \\(\\epsilon\\) estimated by our model which is reasonably close to the simulated noise rate displayed in the table's header." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "A. Garg et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 3: Test accuracy (\\%) of competing methods, and final estimated noise rate \\(\\epsilon\\) on Clothing1M [45]. Also, we have not considered competing models that rely on a clean or validation set whilst training. We integrate DivideMix [26], C2D [59], SSR [11] and CC [58] into our model. DivideMix [26] and CC [58] shows the locally reproduced results. Mentioned noise rate by [45] is 0.385." + }, + { + "type": "table", + "bbox": [ + 0.318, + 0.227, + 0.69, + 0.368 + ], + "angle": 0, + "content": "
MethodTest accuracy (%)Estimated noise rate
OT-Filter [12]74.50
ELR+ with C2D [59]74.58
AugDesc [35]75.11
NCE [25]75.30
DivideMix [26]74.32
DivideMix-Ours74.41 (+0.09)0.41
C2D [59]74.58
C2D-Ours74.71 (+0.13)0.41
SSR (class-imbalance) [11]74.12
SSR-Ours74.20 (+0.08)0.42
CC [58]75.24
CC-Ours75.31 (+0.07)0.41
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.79, + 0.684 + ], + "angle": 0, + "content": "Real-World Noise We also evaluate our proposed method on various real-world noisy settings regarding test accuracy and estimated noise rates \\(\\epsilon\\) in Tabs. 1 to 4. Similarly to the synthetic IDN in Tab. 1 (left), the results show that existing noisy-label robust methods can be easily integrated with our model to outperform current SOTA results for real-world noisy-label datasets. Tabs. 1 and 2 (right) shows the results on red mini-ImageNet using two configurations, including cases without self-supervision (top part of the table) and with self-supervision (bottom part of the table). The self-supervision DINO pre-training [4] relies only on images from red mini-ImageNet to enable a fair comparison with existing baselines [9, 15]. Results from Tab. 1 (right) demonstrate that our approach improves the performance of SOTA methods by a considerable margin in all cases. In fact, using estimated noise rate \\(\\epsilon\\) (Tab. 2 (right)) while training InstanceGM [15] without self-supervision shows better performance than existing self-supervised baselines at 0.4 noise rate. Moreover, DivideMix [26], C2D [59], SSR [11], and CC [58] are used as a baselines for Clothing1M [45] as shown in Tab. 3, with results showing slight improvements with the use of our method. For mini-WebVision [26] (validation on ImageNet [23]), we use DivideMix [26], C2D [59], SSR [11], and CC [58] as baselines (Tab. 4). It is worth noting that our results are better than the SOTA for all settings in Tab. 4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.708, + 0.335, + 0.723 + ], + "angle": 0, + "content": "4.4 Ablation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We show an ablation study with testing accuracy (top) and training time (bottom) of our approach in Tab. 5 (left) on CIFAR100 [22] at 0.3 and 0.5 IDN [44] using DivideMix [26] as baseline. Detailed complexity analysis is shown in the supplementary material, Tab. 10. Initially, the accuracy result of baseline DivideMix [26] is \\(76.33\\%\\), \\(58.61\\%\\) for 0.3 and 0.5 IDN, respectively. In the second row, we fix the noise rate \\(\\epsilon\\) at 0.3 and 0.5 for DivideMix's sample selection, as explained in Sec. 3.3 (without updating \\(\\epsilon\\)), then the results improved to \\(78.14\\%\\) and \\(64.44\\%\\)," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.398, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Noisy-Label Learning with Noise-Rate Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.144, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 4: Test accuracy (\\%) and final estimated noise rate \\(\\epsilon\\) on mini-WebVision [26] and validation on ImageNet [23]. We integrate DivideMix [26], Contrast-to-Divide (C2D) [59], SSR [11] and CC [58] into our model, whilst models suffixed with -Ours denotes our proposed approach. ImageNet [23] is only considered for validation." + }, + { + "type": "table", + "bbox": [ + 0.336, + 0.213, + 0.672, + 0.389 + ], + "angle": 0, + "content": "
Datasetmini-WebVisionImageNetEstimated noise rate
Top-1Top-5Top-1Top-5
BtR [38]80.8892.7675.9692.20
NCE [25]79.5093.8076.3094.10
DivideMix [26]77.3291.6475.2091.64
DivideMix-Ours78.5192.0376.1193.240.43
(+1.19)(+0.39)(+0.91)(+1.60)
C2D [59]79.4292.3278.5793.04
C2D-Ours80.2092.8279.1693.120.43
(+0.78)(+0.50)(+0.59)(+0.08)
SSR [11]80.9292.8075.7691.76
SSR-Ours81.6893.8076.9193.050.43
(+0.76)(+1.00)(+1.15)(+1.29)
CC [58]79.3693.6476.0893.86-
CC-Ours80.0193.7977.1194.210.44
(+0.65)(+0.15)(+1.03)(+0.35)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.407, + 0.788, + 0.533 + ], + "angle": 0, + "content": "Table 5: (left) Ablation results on CIFAR100 at 0.3 and 0.5 IDN. We display accuracy results under different configurations and compare the training time of our approach against DivideMix [26]. The implementation settings and detailed computational analysis are described in Sec. 4.2 and Appendix 4 respectively. (right above) Ablation results for initialisation with different \\(\\epsilon\\) values on CIFAR100 [22] at a synthetic noise rate of 0.5 IDN [44] with DivideMix-Ours. (right below) Ablation graph showing the learning pattern of the noise rate parameter estimation under different initialisation values, i.e., \\(\\epsilon = 0.001\\) (blue), \\(\\epsilon = 0.1\\) (orange), \\(\\epsilon = 0.9\\) (red), when trained on DivideMix-ours case at 0.5 IDN [44] on CIFAR100 [22]." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.562, + 0.495, + 0.687 + ], + "angle": 0, + "content": "
ModelsAccuracy (%)
0.30.5
DivideMix [26]76.3358.61
DivideMix with fixed ε78.1464.44
Our method with pre-trained DivideMix75.0152.31
Our method with original DivideMix (no ε)76.2056.30
DivideMix-Ours77.2164.02
Training Time (GPU-hours)
DivideMix [26]7.20
DivideMix-Ours8.50
DivideMix-Ours (Mixed Precision)6.10
" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.545, + 0.718, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "which is the ideal case (i.e., a perfect noise rate estimation) that motivated our work. In the third case, we use the proposed graphical model with pre-trained DivideMix [26] that shows accuracy of \\(75.01\\%\\) and \\(52.31\\%\\). In the next case, the proposed graphical model is trained together with DivideMix [26] without considering the estimated noise rate \\(\\epsilon\\) for sample selection, which results in accuracy of \\(76.20\\%\\) and \\(56.30\\%\\). In the last row, we show the training of the proposed model with DivideMix [26], together with the estimation of noise rate \\(\\epsilon\\), and the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "A. Garg et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.481 + ], + "angle": 0, + "content": "selection of samples based on that, with \\(\\approx 1\\%\\) and \\(\\approx 8\\%\\) accuracy improvement to 0.3 and 0.5 IDN, which is quite close to our ideal case (second row). We also study the effect that the initialisation of the noise rate parameter \\(\\epsilon\\) plays in our method. As described in Sec. 4.2, we randomly initialise \\(\\epsilon\\) within the range \\((0,1)\\). To test the robustness of our method to this initialisation, we perform an ablation study by initialising \\(\\epsilon\\) with three different values, namely: 0.001, 0.1, and 0.9 on CIFAR100 [24] at 0.5 IDN [44] using the DivideMix-ours approach. Our results, presented in Tab. 5 (right above and below), demonstrate that the initial value of \\(\\epsilon\\) has a limited effect on the model's performance, highlighting the stability of our method across different initialisation of \\(\\epsilon\\). In the supplementary material, Tab. 7 (left) presents comparative analyses of our model against PartT [44] and SSR [11] on CIFAR100 [22] at 0.5 IDN [44], while Tab. 7 (right) showcases similar comparisons with baselines DivideMix [26] and InstanceGM [15] at a higher noise rate of 0.6 IDN [44]. Additionally, Tab. 8 provides further insights into the performance of DivideMix [26] on red mini-ImageNet [17] at lower noise rates of 0.1, 0.2 and 0.3. Tab. 9 illustrates noise rate estimation with standard deviation for our model integrated with DivideMix [26] and InstanceGM [15] on CIFAR100 [22] under IDN [44] settings at 0.2, 0.3, 0.4, and 0.5. Further statistical analysis is detailed in Appendix 5 with confidence interval presented in Tab. 12. Performance analysis with various architectures (Tab. 13) and lambda values (Tab. 14) are mentioned in Appendix 6. Threshold-based comparisons are in Appendix 7 and Tab. 15." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.502, + 0.509, + 0.518 + ], + "angle": 0, + "content": "5 Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.788, + 0.64 + ], + "angle": 0, + "content": "In this work, we proposed an innovative graphical model for IDN noisy-label learning, focusing on the estimation of the label noise rate that is leveraged to introduce a robust sample selection curriculum. Given that SOTA IDN noisy-label learning approaches tend to rely on sample selection methods, our method can be seamlessly integrated to them to improve their performance in many synthetic and real-world benchmarks, including CIFAR100 [22], red mini-ImageNet [17], Clothing1M [45], mini-WebVision [26], and ImageNet [23]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.789, + 0.823 + ], + "angle": 0, + "content": "From a societal perspective, our methodology offers a positive impact by mitigating biases inherent in noisy data, thereby promoting more fair and accurate machine learning outcomes. It also opens the door to future research avenues, including exploring advanced noise rate estimation methodologies and investigating noise-robust loss functions such as GCE [56] and ELR [28] with it. Our choice of using the same network architecture for clean and noisy labels enables compatibility with co-teaching techniques [16,26]. However, investigating different network structures for clean and noisy labels remains a potential avenue for future research. Moreover, we aim to delve into its dynamics, uncovering substantial improvements for certain models while observing more marginal enhancements for others. For more detailed information, please refer to Appendix 8. Our approach also hints at potential breakthroughs in streamlining data annotation challenges." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.398, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Noisy-Label Learning with Noise-Rate Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.787, + 0.236 + ], + "angle": 0, + "content": "1. Arpit, D., Jastrzebski, S., Ballas, N., Krueger, D., Bengio, E., Kanwal, M.S., Maharaj, T., Fischer, A., Courville, A., Bengio, Y., Lacoste-Julien, S.: A closer look at memorization in deep networks. In: International Conference on Machine Learning. vol. 70, pp. 233-242. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.237, + 0.787, + 0.278 + ], + "angle": 0, + "content": "2. Bae, H., Shin, S., Jang, J., Na, B., Song, K., Moon, I.C.: From noisy prediction to true label: Noisy prediction calibration via generative model. In: International Conference on Machine Learning (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.279, + 0.787, + 0.32 + ], + "angle": 0, + "content": "3. Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.A.: Mixmatch: A holistic approach to semi-supervised learning. In: Advances in Neural Information Processing Systems. vol. 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.321, + 0.787, + 0.362 + ], + "angle": 0, + "content": "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: International Conference on Computer Vision. pp. 9650-9660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.363, + 0.787, + 0.404 + ], + "angle": 0, + "content": "5. Chen, P., Ye, J., Chen, G., Zhao, J., Heng, P.A.: Beyond class-conditional assumption: A primary attempt to combat instance-dependent label noise. In: AAAI Conference on Artificial Intelligence. pp. 11442–11450 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.787, + 0.46 + ], + "angle": 0, + "content": "6. Chen, Z., Song, A., Wang, Y., Huang, X., Kong, Y.: A noise rate estimation method for image classification with label noise. Journal of Physics: Conference Series 2433(1), 012039 (feb 2023). https://doi.org/10.1088/1742-6596/2433/1/012039, https://dx.doi.org/10.1088/1742-6596/2433/1/012039" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.787, + 0.516 + ], + "angle": 0, + "content": "7. Cheng, D., Liu, T., Ning, Y., Wang, N., Han, B., Niu, G., Gao, X., Sugiyama, M.: Instance-dependent label-noise learning with manifold-regularized transition matrix estimation. In: Conference on Computer Vision and Pattern Recognition. pp. 16630-16639 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.517, + 0.787, + 0.559 + ], + "angle": 0, + "content": "8. Cheng, H., Zhu, Z., Li, X., Gong, Y., Sun, X., Liu, Y.: Learning with instance-dependent label noise: A sample sieve approach. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.56, + 0.787, + 0.601 + ], + "angle": 0, + "content": "9. Cordeiro, F.R., Belagiannis, V., Reid, I., Carneiro, G.: PropMix: Hard sample filtering and proportional mixup for learning with noisy labels. In: British Machine Vision Conference (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.602, + 0.787, + 0.643 + ], + "angle": 0, + "content": "0. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the EM algorithm. Journal of the royal statistical society: series B (methodological) 39(1), 1-22 (1977)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.644, + 0.787, + 0.672 + ], + "angle": 0, + "content": "1. Feng, C., Tzimiropoulos, G., Patras, I.: SSR: An efficient and robust framework for learning with unknown label noise. In: British Machine Vision Conference (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.787, + 0.713 + ], + "angle": 0, + "content": "2. Feng, C., Ren, Y., Xie, X.: Ot-filter: An optimal transport filter for learning with noisy labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16164-16174 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.714, + 0.787, + 0.755 + ], + "angle": 0, + "content": "3. Frénay, B., Kabán, A.: A comprehensive introduction to label noise. In: European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "4. Frénay, B., Verleysen, M.: Classification in the presence of label noise: A survey. IEEE Transactions on Neural Networks and Learning Systems 25(5), 845-869 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "5. Garg, A., Nguyen, C., Felix, R., Do, T.T., Carneiro, G.: Instance-dependent noisy label learning via graphical modelling. In: Winter Conference on Applications of Computer Vision. pp. 2288-2298 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "A. Garg et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "16. Han, B., Yao, Q., Yu, X., Niu, G., Xu, M., Hu, W., Tsang, I., Sugiyama, M.: Co-teaching: Robust training of deep neural networks with extremely noisy labels. In: Advances in Neural Information Processing Systems. vol. 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.231 + ], + "angle": 0, + "content": "17. Jiang, L., Huang, D., Liu, M., Yang, W.: Beyond synthetic noise: Deep learning on controlled noisy labels. In: International Conference on Machine Learning. pp. 4804-4815. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.786, + 0.272 + ], + "angle": 0, + "content": "18. Jiang, L., Zhou, Z., Leung, T., Li, L.J., Fei-Fei, L.: MentorNet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In: International Conference on Machine Learning. pp. 2304-2313. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.327 + ], + "angle": 0, + "content": "19. Jiang, Z., Zhou, K., Liu, Z., Li, L., Chen, R., Choi, S.H., Hu, X.: An information fusion approach to learning with instance-dependent label noise. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=ecH2FKaARUp" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "20. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in Neural Information Processing Systems 33, 18661-18673 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.397 + ], + "angle": 0, + "content": "21. Kim, T., Ko, J., Choi, J., Yun, S.Y.: FINE samples for learning with noisy labels. In: Advances in Neural Information Processing Systems. vol. 34 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.424 + ], + "angle": 0, + "content": "22. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. Tech. rep., University of Toronto (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.466 + ], + "angle": 0, + "content": "23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: ImageNet classification with deep convolutional neural networks. In: Advances in Neural Information Processing Systems. vol. 25 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.508 + ], + "angle": 0, + "content": "24. Kye, S.M., Choi, K., Yi, J., Chang, B.: Learning with noisy labels by efficient transition matrix estimation to combat label miscorrection. In: European Conference on Computer Vision. pp. 717-738. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "25. Li, J., Li, G., Liu, F., Yu, Y.: Neighborhood collective estimation for noisy label identification and correction. In: European Conference on Computer Vision. pp. 128-145. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.59 + ], + "angle": 0, + "content": "26. Li, J., Socher, R., Hoi, S.C.: DivideMix: Learning with noisy labels as semi-supervised learning. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.787, + 0.619 + ], + "angle": 0, + "content": "27. Li, W., Wang, L., Li, W., Agustsson, E., Gool, L.V.: WebVision Database: Visual learning and understanding from web data. CoRR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.787, + 0.66 + ], + "angle": 0, + "content": "28. Liu, S., Niles-Weed, J., Razavian, N., Fernandez-Granda, C.: Early-learning regularization prevents memorization of noisy labels. In: Advances in Neural Information Processing Systems. vol. 33, pp. 20331-20342 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.787, + 0.688 + ], + "angle": 0, + "content": "29. Liu, Y., Cheng, H., Zhang, K.: Identifiability of label noise transition matrix. In: International Conference on Machine Learning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.73 + ], + "angle": 0, + "content": "30. Ma, X., Huang, H., Wang, Y., Romano, S., Erfani, S., Bailey, J.: Normalized loss functions for deep learning with noisy labels. In: International Conference on Machine Learning. pp. 6543-6553. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.787, + 0.757 + ], + "angle": 0, + "content": "31. Malach, E., Shalev-Shwartz, S.: Decoupling \"when to update\" from \"how to update\". In: Advances in Neural Information Processing Systems. vol. 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.668, + 0.771 + ], + "angle": 0, + "content": "32. McLachlan, G.J., Peel, D.: Finite mixture models. Wiley (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "33. Neal, R.M., Hinton, G.E.: A view of the EM algorithm that justifies incremental, sparse, and other variants. Learning in graphical models pp. 355-368 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "34. Neyshabur, B., Bhojanapalli, S., McAllester, D., Srebro, N.: Exploring generalization in deep learning. In: Advances in Neural Information Processing Systems. vol. 30 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.398, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Noisy-Label Learning with Noise-Rate Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "35. Nishi, K., Ding, Y., Rich, A., Hollerer, T.: Augmentation strategies for learning with noisy labels. In: Conference on Computer Vision and Pattern Recognition. pp. 8022-8031 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.26 + ], + "angle": 0, + "content": "36. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: PyTorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.261, + 0.788, + 0.302 + ], + "angle": 0, + "content": "37. Shamshad, F., Khan, S., Zamir, S.W., Khan, M.H., Hayat, M., Khan, F.S., Fu, H.: Transformers in medical imaging: A survey. Medical Image Analysis p. 102802 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.303, + 0.788, + 0.345 + ], + "angle": 0, + "content": "38. Smart, B., Carneiro, G.: Bootstrapping the relationship between images and their clean and noisy labels. In: Winter Conference on Applications of Computer Vision. pp. 5344-5354 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.346, + 0.788, + 0.387 + ], + "angle": 0, + "content": "39. Song, H., Kim, M., Park, D., Shin, Y., Lee, J.G.: Learning from noisy labels with deep neural networks: A survey. IEEE Transactions on Neural Networks and Learning Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.388, + 0.788, + 0.43 + ], + "angle": 0, + "content": "40. Tanaka, D., Ikami, D., Yamasaki, T., Aizawa, K.: Joint optimization framework for learning with noisy labels. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.431, + 0.788, + 0.458 + ], + "angle": 0, + "content": "41. Titterington, D.M.: On the likelihood of mixtures of distributions. Biometrika 71(3), 511-522 (1984). https://doi.org/10.1093/biomet/71.3.511" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.459, + 0.788, + 0.5 + ], + "angle": 0, + "content": "42. Trummer, I.: From BERT to GPT-3 codex: harnessing the potential of very large language models for data management. VLDB Endowment 15(12), 3770-3773 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.501, + 0.788, + 0.543 + ], + "angle": 0, + "content": "43. Tu, H., Menzies, T.: Debtfree: minimizing labeling cost in self-admitted technical debt identification using semi-supervised learning. Empirical Software Engineering 27(4), 80 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.544, + 0.788, + 0.586 + ], + "angle": 0, + "content": "44. Xia, X., Liu, T., Han, B., Wang, N., Gong, M., Liu, H., Niu, G., Tao, D., Sugiyama, M.: Part-dependent label noise: Towards instance-dependent label noise. In: Advances in Neural Information Processing Systems. vol. 33, pp. 7597-7610 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.587, + 0.788, + 0.628 + ], + "angle": 0, + "content": "45. Xiao, T., Xia, T., Yang, Y., Huang, C., Wang, X.: Learning from massive noisy labeled data for image classification. In: Conference on Computer Vision and Pattern Recognition. pp. 2691-2699 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.629, + 0.788, + 0.67 + ], + "angle": 0, + "content": "46. Xu, Y., Zhu, L., Jiang, L., Yang, Y.: Faster meta update strategy for noise-robust deep learning. In: Conference on Computer Vision and Pattern Recognition. pp. 144-153 (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.671, + 0.788, + 0.713 + ], + "angle": 0, + "content": "47. Yang, S., Jiang, L., Liu, Z., Loy, C.C.: Unsupervised image-to-image translation with generative prior. In: Conference on Computer Vision and Pattern Recognition. pp. 18332-18341 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.714, + 0.788, + 0.755 + ], + "angle": 0, + "content": "48. Yang, S., Yang, E., Han, B., Liu, Y., Xu, M., Niu, G., Liu, T.: Estimating instance-dependent bayes-label transition matrix using a deep neural network. In: International Conference on Machine Learning. pp. 25302-25312. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.757, + 0.788, + 0.798 + ], + "angle": 0, + "content": "49. Yao, Q., Yang, H., Han, B., Niu, G., Kwok, J.T.Y.: Searching to exploit memorization effect in learning with noisy labels. In: International Conference on Machine Learning. pp. 10789-10798. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.799, + 0.788, + 0.84 + ], + "angle": 0, + "content": "50. Yao, Y., Liu, T., Gong, M., Han, B., Niu, G., Zhang, K.: Instance-dependent label-noise learning under a structural causal model. In: Advances in Neural Information Processing Systems. vol. 34 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "A. Garg et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "51. Yao, Y., Liu, T., Han, B., Gong, M., Deng, J., Niu, G., Sugiyama, M.: Dual T: Reducing estimation error for transition matrix in label-noise learning. In: Advances in Neural Information Processing Systems. vol. 33, pp. 7260-7271 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "52. Zhang, C., Bengio, S., Hardt, M., Recht, B., Vinyals, O.: Understanding deep learning requires rethinking generalization. In: International Conference on Learning Representations (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "53. Zhang, C., Bengio, S., Hardt, M., Recht, B., Vinyals, O.: Understanding deep learning (still) requires rethinking generalization. Communications of the ACM 64(3), 107-115 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "54. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. In: International Conference on Learning Representations (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "55. Zhang, Y., Niu, G., Sugiyama, M.: Learning noise transition matrix from only noisy labels via total variation regularization. In: International Conference on Machine Learning. pp. 12501-12512. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.787, + 0.383 + ], + "angle": 0, + "content": "56. Zhang, Z., Sabuncu, M.: Generalized cross entropy loss for training deep neural networks with noisy labels. Advances in Neural Information Processing Systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.787, + 0.424 + ], + "angle": 0, + "content": "57. Zhang, Z., Pfister, T.: Learning fast sample re-weighting without reward data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 725-734 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.424, + 0.787, + 0.467 + ], + "angle": 0, + "content": "58. Zhao, G., Li, G., Qin, Y., Liu, F., Yu, Y.: Centrality and consistency: two-stage clean samples identification for learning with instance-dependent noisy labels. In: European Conference on Computer Vision. pp. 21-37. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.508 + ], + "angle": 0, + "content": "59. Zheltonozhskii, E., Baskin, C., Mendelson, A., Bronstein, A.M., Litany, O.: Contrast to divide: Self-supervised pre-training for learning with noisy labels. In: Winter Conference on Applications of Computer Vision. pp. 1657-1667 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.787, + 0.551 + ], + "angle": 0, + "content": "60. Zhu, Z., Liu, T., Liu, Y.: A second-order approach to learning with instance-dependent label noise. In: Conference on Computer Vision and Pattern Recognition. pp. 10113-10123 (June 2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.551 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation/8698146c-3791-42ec-912e-34881bbacb25_origin.pdf b/2024/Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation/8698146c-3791-42ec-912e-34881bbacb25_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fa598d3e9643ab4561a26c6ba6084d062ffbbeb6 --- /dev/null +++ b/2024/Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation/8698146c-3791-42ec-912e-34881bbacb25_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:753848272b20b76fbaea9992feff1d0386d6641d6d74a0199d49c67153d8579a +size 1320827 diff --git a/2024/Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation/full.md b/2024/Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4a47417f53d50fa9746742a0f539e40df3548d6d --- /dev/null +++ b/2024/Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation/full.md @@ -0,0 +1,304 @@ +# Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation + +Arpit Garg1, Cuong Nguyen3, Rafael Felix1, Thanh-Toan Do2, and Gustavo Carneiro3 + +$^{1}$ Australian Institute for Machine Learning, University of Adelaide, SA, Australia $^{2}$ Department of Data Science and AI, Monash University, Australia + +3 Centre for Vision, Speech and Signal Processing, University of Surrey, UK + +Abstract. Deep learning faces a formidable challenge when handling noisy labels, as models tend to overfit samples affected by label noise. This challenge is further compounded by the presence of instance-dependent noise (IDN), a realistic form of label noise arising from ambiguous sample information. To address IDN, Label Noise Learning (LNL) incorporates a sample selection stage to differentiate clean and noisy-label samples. This stage uses an arbitrary criterion and a pre-defined curriculum that initially selects most samples as noisy and gradually decreases this selection rate during training. Such curriculum is sub-optimal since it does not consider the actual label noise rate in the training set. This paper addresses this issue with a new noise-rate estimation method that is easily integrated with most state-of-the-art (SOTA) LNL methods to produce a more effective curriculum. Synthetic and real-world benchmarks' results demonstrate that integrating our approach with SOTA LNL methods improves accuracy in most cases. $^4$ + +Keywords: Noisy-labels $\cdot$ Instance-dependent noise $\cdot$ Label noise Learning + +# 1 Introduction + +Deep neural networks (DNNs) have demonstrated their effectiveness in various domains, including vision [47], language [42], and medicine [37]. However, its efficacy is largely dependent on high-quality training data, which can be resource-intensive to obtain [43]. Although cost-effective labelling techniques, such as data mining [11] and crowdsourcing [39], offer cheaper alternatives, they often compromise label quality [39]. Consequently, this can lead to erroneous labels in real-world datasets [21]. This is a relevant issue because even slight inaccuracies in the labels can significantly affect the performance of DNN due to their inherent memorisation capabilities [34, 53]. This has prompted the development of algorithms resilient to noisy-label learning, with the aim of training models + +despite inaccuracies in training data labels. Although various strategies exist, our work uniquely proposes an approach centered on the estimation of the label noise rate from the training data, which is notably absent from current methodologies. + +There are two types of label noise, namely instance-independent noise (IIN) [16] and instance-dependent noise (IDN) [44]. Such type of label noise typically influences the design principles of noisy-label learning algorithms. For example, IIN focuses on mislabellings that are independent of sample information [16], where estimating the underlying label transition matrix is a common way of handling this type of noise [51]. On the other hand, in the more realistic IDN, mislabelling is due to both sample information and true class labels [44], which generally require the combination of many label noise learning techniques, such as robust loss functions [28, 56], and noisy-label sample selection [26, 58]. Of the strategies mentioned above, sample selection approaches that classify training data into clean and noisy samples have produced competitive results on many benchmarks [9, 11, 15, 21, 26]. Such sample selection techniques require the definition of a classification criterion and a selection curriculum. Many studies on this topic focus on developing new sample selection criteria, such as the small-loss hypothesis [26], which states that noisy-label samples have higher loss values than clean-label samples, particularly at the early stage of training [1]. Another criterion type is the feature-based one. An example of such criterion is FINE [21] that discriminates clean and noisy-label samples via the distance to class-specific eigenvectors. In this technique, clean-label samples tend to lie closer to the class-specific dominant eigenvector of the latent representations than noisy-label samples. Another type of criterion is proposed in SSR [11], which introduces a selection criterion based on the K-nearest-neighbor (KNN) classification in the feature space. Furthermore, CC [58] uses a two-stage sampling procedure, including class-level feature clustering followed by a consistency score. An equally important problem in sample selection is the definition of the curriculum for selecting clean training samples, but it has received comparatively less attention. + +The sample selection curriculum defines a threshold to be used with one of the criteria listed above to classify the training samples as clean or noisy at each training epoch [45]. For example, the threshold can be fixed to an arbitrary clustering score that separates clean and noisy samples [26], but this strategy does not account for the proportion of label noise in the training set, nor does it consider the dynamics of the selection of noisy-label samples during the training. The consideration of such dynamics has been studied in [16, 49], which defined a curriculum of the noisy-label sampling rate $R(t)$ as a function of the training epoch $t \in \{1, \dots, T\}$ . The curriculum $R(t)$ defines a sampling rate close to $100\%$ of the training set at the beginning of the training, which is then reduced to arbitrarily low rates at the end of the training. In practice, the function $R(t)$ is predefined [16] or learned by weighting a set of basis functions [49]. + +Although generally effective, these techniques do not consider the label noise rate estimated from the training set, making them vulnerable to over-fitting (if too many noisy-label samples are classified as clean) or under-fitting (if informative clean-label samples are classified as noisy). It can be argued that the estimation + +![](images/a180e75bd1f43c8f3c626a7caf5e2fc0fcf82f6c0d858ceb5e6d660de8757ded.jpg) +(a) Motivation + +![](images/7d25d80cae29bd7cfecbfdbad32f62383c9de642ce5c58d654608ce8e8055e3e.jpg) +(b) Proposed Graphical Model +Fig. 1: (a) Comparison of test accuracy $\%$ (as a function of training epoch) between the original DivideMix [26] (solid, blue curve) and our modified DivideMix (dashed, red curve) that selects the clean and noisy data based on a fixed noise rate $R(t) = 1 - \epsilon = 50\%$ using the small-loss criterion on CIFAR100 [22] at 0.5 IDN [44]; (b) The proposed probabilistic graphical model that generates noisy-label $\hat{Y}$ conditioned on the image $X$ , the latent clean-label $Y$ and noise rate $\epsilon$ , where forward pass (solid lines) is parameterised by $\theta_y, \theta_{\hat{y}}$ and $\epsilon$ representing the generation step, and the backward pass (dashed lines) is parameterised by $\rho$ . + +of the label transition matrix [6,7,48,51] aims to recover the noise rate affecting pairwise label transitions. However, label transition matrix techniques follow a quite different strategy compared to sample selection methods, where their main challenge is the general underconstrained aspect of the matrix estimation, making them sensitive to large noise rates and not scalable to a high number of classes [39]. Although several methods address label noise, none estimates the label noise rate directly from the training data to guide sample selection. Addressing this gap is the primary challenge of our study. + +To underscore the importance of using the noise rate for sample selection during training, we experiment with CIFAR100 [22] at an IDN rate $\epsilon = 50\%$ [44] (noise rate specifications and other details are explained in Sec. 4). We use DivideMix [26] and replace its sample selection (based on an arbitrary clustering score [10,32,41]) by a thresholding process that classifies the $R(t) = 1 - \epsilon = 50\%$ largest loss samples as noisy, and the remaining ones as clean in all training epochs $t \in \{1, \dots, T\}$ . This sample selection is used for the semi-supervised learning of DivideMix [26]. As shown in Fig. 1a, the new sample selection approach based on the "manually provided" noise rate (dashed red curve) improves $6\%$ in terms of prediction accuracy compared to the original DivideMix [26] (solid blue curve) which relies on arbitrary thresholding. Similar conclusions can be achieved with other methods that apply sample selection strategies to address the noisy-label learning problem, as shown in the experiments. + +In this paper, we introduce a new sample selection strategy centered on estimating the label noise rate of the training set. Our strategy is based on our novel noisy-label learning graphical model illustrated in Fig. 1b. This model can be + +![](images/c2fc73a33c98dface34ed4f3415f9bc7caaa303299e41de6b32bf051a54fe249.jpg) +(a) Curriculum function $R(t)$ + +![](images/d989ee52eda15970afc13f9f66dc0f90d99152f959141d0db19c6b74d4948716.jpg) +(b) Accuracy Comparison +Fig. 2: (a) Visual comparison of different $R(t)$ on CIFAR100 [22] at 0.5 IDN [44]: (i) Co-teaching [16] with curriculum based on hyper-parameter $T_{k} \in \{5,10\}$ , where $R(t) = 1 - \tau \cdot \min(t / T_{k}, 1)$ with $\tau = \epsilon = 0.5$ being manually set; (ii) S2E [49], where $R(t)$ is estimated with a bi-level optimisation; (iii) SSR [11], where $R(t)$ is based on a relabelling pattern, (iv) CC [58] following the small-loss hypothesis [26] and cosine similarity for $R(t)$ ; and (v) ours. (b) Accuracy of noisy-label learning methods on CIFAR100 [22] at 0.5 IDN [44], including DivideMix [26], FINE [21] and InstanceGM [15], without (left, blue) and (right, orange) integration of our proposed graphical model for estimating the noise rate $\epsilon$ and sample selection rate based on $\epsilon$ . + +seamlessly integrated with state-of-the-art (SOTA) noisy-label learning techniques, providing them with a precise noise rate estimate and subsequently refining the sample selection curriculum. In particular, our model's curriculum is not restrained by predefined $R(t)$ functions [16, 49], but rather relies on a dynamically estimated noise rate sourced directly from the training set, as depicted in Fig. 2a. Our method dynamically estimates training set noise rates to mitigate overfitting and underfitting, seamlessly integrates with existing algorithms, and serves as a robust foundation for future noisy-label learning research. To summarise, our main contributions include: + +- An innovative noisy-label learning graphical model, shown in Fig. 1b, that not only estimates but also leverages the noise rate from the training dataset to produce a refined sample selection curriculum for SOTA LNL methods. +- A simple and synergistic integration strategy of our novel graphical model with several SOTA noisy-label learning algorithms, such as DivideMix [26] and SSR [11], to improve their sample selection effectiveness, leading to an increased test accuracy, as shown in Fig. 2b. + +Empirical evaluations show the role of our proposed sample selection methodology in improving the efficacy of leading noisy-label learning algorithms across various synthetic and real-world IDN benchmarks. Although our primary focus lies in addressing IDN problems, because of its more challenging and realistic nature, we have also applied our model to IIN problems, with detailed results in Tab. 6 and discussion in Appendix 3 of the supplementary material. With our innovative approach, we aim to improve the accuracy of SOTA LNL methods. + +# 2 Related Work + +It is well-known that DNNs suffer from overfitting when trained with noisy-labels [52], resulting in poor generalisation [28, 55]. To mitigate the problems related to noisy-label learning, several techniques have been developed, including noise-robust loss functions [30], noise-label sample selection [21, 26, 40, 58, 59], meta-learning [57], and re-labeling followed by sample selection [11]. These techniques can handle many types of label noise (e.g., symmetric and asymmetric), but the most challenging type, i.e., IDN [39], tends to be successfully addressed with methods that have a training stage based on sample selection techniques [21, 26, 40, 58, 59]. These sample selection techniques separate clean and noisy-label samples, where clean samples are treated as labelled, and noisy samples are discarded [8, 16, 18, 49] or treated as unlabeled samples [9, 15, 26, 35, 59] for semi-supervised learning [3]. Another solution consists of aligning clean and noisy samples with an information fusion method [19]. A major limitation of these approaches is the overly simplistic design of the curriculum to select clean and noisy-label samples during training. It is either predetermined [16, 18] or learned from a set of predetermined basis functions [49]. We argue that the design of a curriculum based on an estimated noise rate would benefit such methods, as motivated by Fig. 1a, but to the best of our knowledge, such estimation has not been explored by previous methods. The closest idea explored is based on the estimation of the type of noise instead of the noise rate [45]. It is possible to argue that the estimation of the label transition matrices [5, 7, 44, 60], composed of the learned noise rate that affects the transition between pairs of labels, is a way to estimate the label noise rate. However, they portray results of comparatively lower accuracy for large real-world datasets or for IDN problems [39, 44]. A possible reason for these poorer results is that label transition approaches suffer from identifiability issues [14], where any clean-label distribution assignment is acceptable as long as the distribution of observed labels can be reconstructed [13]. This makes the identification of the actual underlying clean-labels challenging. One solution is to consider the use of multiple annotations per sample to help analyse agreements and disagreements for improved identification of clean patterns [14, 31]. However, the annotation of datasets with a single label per training sample is already challenging; the complexity is significantly larger to acquire multiple labels per sample. + +An alternative technique to handle IDN problems is based on graphical models that represent the relationship between various observed and latent variables [15,45,50]. The approaches in [15,50] use graphical models that rely on a generative approach to produce noisy-labels from the respective image features and latent clean-labels. However, previous graphical models do not take into account the underlying noise rate parameter during modeling. Our work is the first graphical model approach to estimate the noise rate of the dataset. In addition, a close examination of existing literature reveals a trend: when traditional graphical models are integrated with SOTA noisy-label learning methods, the outcomes often fall short of expectations in terms of accuracy [2,15,50]. This observation underscores a strong point of our work. Our approach is uniquely tailored for + +easy integration with current SOTA noisy-label learning methods for sample selection, which, in turn, improves over the SOTA classification accuracy. + +# 3 Method + +In this section, we present our new graphical model that estimates the noise rate, which will be used in the sample selection process. Let $\mathcal{D} = \{(x_i,\hat{y}_i)\}_{i = 1}^N$ be the noisy-label training set containing $d$ -dimensional data vector $x_{i}\in \mathcal{X}\subseteq \mathbb{R}^{d}$ and its respective $C$ -dimensional one-hot encoded observed (potentially corrupted) label $\hat{y}_i\in \hat{\mathcal{V}} = \{\hat{y}:\hat{y}\in \{0,1\} ^C\wedge \mathbf{1}_C^\top \hat{y} = 1\}$ , where $\mathbf{1}_C$ is a vector of ones with $C$ dimensions. The aim is to estimate the label noise rate $\epsilon$ , used for the generation of noisy-label training data from the observed training dataset $\mathcal{D}$ and integrate this label noise rate into the sample selection strategy. + +# 3.1 Graphical Model + +We portray the generation of noisy-label via the probabilistic graphical model shown in Fig. 1b. The observed random variables, denoted by shaded circles, are data $X$ and the corresponding noisy-label $\hat{Y}$ . We also have one latent variable, namely: the clean-label $Y$ . Under our proposed modeling assumption, a noisy-label of a data instance can be generated as follows: + +- sample an instance from $p(X)$ , i.e.: $x \sim p(X)$ +- sample a clean-label from the clean-label distribution: + +$$ +y \sim \operatorname {C a t e g o r i c a l} (Y; f _ {\theta_ {y}} (x)) +$$ + +- sample a noisy-label from the noisy-label distribution: + +$$ +\hat {y} \sim \operatorname {C a t e g o r i c a l} (\hat {Y}; \epsilon \times f _ {\theta_ {\hat {y}}} (x, f _ {\theta_ {y}} (x)) + (1 - \epsilon) \times y), +$$ + +where $\text{Categorical}()$ denotes a categorical distribution, $f_{\theta_y}: \mathcal{X} \to \Delta_{C-1}$ and $f_{\theta_{\hat{y}}}: \mathcal{X} \times \Delta_{C-1} \to \Delta_{C-1}$ denote two classifiers for the clean-label $Y$ and noisy-label $\hat{Y}$ , respectively, with $\Delta_{C-1} = \{s : s \in [0,1]^C \wedge \mathbf{1}_C^\top s = 1\}$ being the $(C-1)$ -dimensional probability simplex. According to the data generation process shown in Fig. 1b, $\epsilon$ corresponds to $\mathbb{E}_{(x,\hat{y}) \sim p(X,\hat{Y})}[P(\hat{y} \neq y|x)]$ , which is the label noise rate of the training dataset of interest. Our aim is to infer the parameters $\theta_y, \theta_{\hat{y}}$ and $\epsilon$ from a noisy-label dataset $\mathcal{D}$ by maximising the following log-likelihood: + +$$ +\max _ {\theta_ {y}, \theta_ {\hat {y}}, \epsilon} \mathbb {E} _ {(x _ {i}, \hat {y} _ {i}) \sim \mathcal {D}} [ \ln p (\hat {y} _ {i} | x _ {i}; \theta_ {y}, \theta_ {\hat {y}}, \epsilon) ] = \max _ {\theta_ {y}, \theta_ {\hat {y}}, \epsilon} \mathbb {E} _ {(x _ {i}, \hat {y} _ {i}) \sim \mathcal {D}} [ \ln \sum_ {y _ {i}} p (\hat {y} _ {i}, y _ {i} | x _ {i}; \theta_ {y}, \theta_ {\hat {y}}, \epsilon) ]. \tag {1} +$$ + +Due to the presence of the clean-label $y_{i}$ , it is difficult to directly evaluate the log-likelihood in Eq. (1). Therefore, we employ the expectation-maximisation (EM) algorithm [10] to maximise the log-likelihood. The main idea of the EM algorithm is to (i) construct a tight lower bound of the likelihood in Eq. (1) by estimating the posterior of the latent variable $Y$ (known as expectation step) and (ii) maximise that lower bound (known as maximisation step). Formally, let + +$q(y_{i}|x,\hat{y};\rho)$ be an arbitrary distribution over a clean-label $y_{i}$ . The evidence lower bound (ELBO) of the log-likelihood in Eq. (1) can be obtained through Jensen's inequality and is presented as follows: + +$$ +\begin{array}{l} Q \left(\theta_ {y}, \theta_ {\hat {y}}, \epsilon , \rho\right) = \mathbb {E} _ {\left(x _ {i}, \hat {y} _ {i}\right) \sim \mathcal {D}} \left[ \mathbb {E} _ {q \left(y _ {i} \mid x _ {i}; \hat {y} _ {i}; \rho\right)} \left[ \ln p \left(\hat {y} _ {i} \mid x _ {i}, y _ {i}; \theta_ {y}, \theta_ {\hat {y}}, \epsilon\right) \right] \right. \\ - \operatorname {K L} \left[ q \left(y _ {i} \mid x _ {i}, \hat {y} _ {i}; \rho\right) \| p \left(y _ {i} \mid x _ {i}; \theta_ {y}, \theta_ {\hat {y}}, \epsilon\right) \right] \\ = \mathbb {E} _ {\left(x _ {i}, \hat {y} _ {i}\right) \sim \mathcal {D}} \left[ \mathbb {E} _ {q \left(y _ {i} \mid x _ {i}; \hat {y} _ {i}; \rho\right)} \left[ \ln p \left(\hat {y} _ {i} \mid x _ {i}, y _ {i}; \theta_ {\hat {y}}, \epsilon\right) \right. \right. \\ \left. + \ln p \left(y _ {i} \mid x _ {i}; \theta_ {y}\right) \right] + \mathbb {H} \left[ q \left(y _ {i} \mid x _ {i}, \hat {y} _ {i}; \rho\right) \right], \tag {2} \\ \end{array} +$$ + +where $\mathrm{KL}[q\| p]$ is the Kullback - Leibler divergence between distributions $q$ and $p$ and $\mathbb{H}(q)$ is the entropy of the distribution $q$ . The EM algorithm is then carried out iteratively by alternating the following two steps: + +E step: we maximise the ELBO in (2) w.r.t. $q(y_{i}|x_{i},\hat{y}_{i};\rho)$ . Theoretically, such an optimisation results in KL $[q(y_{i}|x_{i},\hat{y}_{i};\rho)\| p(y_{i}|x_{i},\hat{y}_{i})] = 0$ or $q(y_{i}|x_{i},\hat{y}_{i};\rho) = p(y_{i}|x_{i},\hat{y}_{i})$ . This is equivalent to estimating the posterior of the clean-label $y_{i}$ given noisy-label data $(x_{i},\hat{y}_{i})$ . Obtaining the posterior $p(y_{i}|x_{i},\hat{y}_{i})$ is, however, intractable for most deep-learning models. To mitigate such an issue, we follow the variational EM approach [33] by employing an approximate posterior $q(y_{i}|x_{i},\hat{y}_{i};\rho^{(t)})$ that is the closest to the true posterior $p(y_{i}|x_{i},\hat{y}_{i})$ , where: + +$$ +\rho^ {(t)} = \arg \max _ {\rho} Q \left(\theta_ {y} ^ {(t)}, \theta_ {\hat {y}} ^ {(t)}, \epsilon^ {(t)}, \rho\right), \tag {3} +$$ + +with the superscript ${}^{(t)}$ denoting the parameters at the $t$ -th iteration. Although this results in a non-tight lower bound of the log-likelihood in Eq. (1), it does increase the variational bound $Q$ . + +M step: we maximise the ELBO in Eq. (2) w.r.t. $\theta_y, \theta_{\hat{y}}$ and $\epsilon$ given $\rho^{(t)}$ obtained in the E step: + +$$ +\theta_ {y} ^ {(t + 1)}, \theta_ {\hat {y}} ^ {(t + 1)}, \epsilon^ {(t + 1)} = \arg \max _ {\theta_ {y}, \theta_ {\hat {y}}, \epsilon} Q (\theta_ {y}, \theta_ {\hat {y}}, \epsilon , \rho^ {(t)}). \tag {4} +$$ + +The estimated noise rate $\epsilon$ can then be integrated into certain noisy-label algorithms to train the models of interest as mentioned in Sec. 1. In addition, the inference of noise rate $\epsilon$ might be associated with the identifiability issue when estimating the clean-label $Y$ [29], i.e., there exists multiple sets of $\rho$ and $\theta_y$ , where each set can explain the observed noisy-label data equally well. Such issues are addressed in the following subsection. + +# 3.2 Sample Selection + +The identifiability issue when inferring the clean-label $Y$ from noisy-label data $(X, \hat{Y})$ can be mitigated either by acquiring multiple noisy-labels [29] or introducing additional constraints, such as small loss hypothesis [16] or FINE [21]. Since requesting additional noisy-labels per training sample is not always possible, we + +Algorithm 1 Proposed noisy-label learning that relies on the estimation of noise rate $\epsilon$ to build a sample selection curriculum. +1: procedure NOISE RATE ESTIMATION AND INTEGRATION(D,T,λ) +2: $\triangleright$ $\mathcal{D} = \{(x_i,\hat{y}_i)\}_{i = 1}^N$ training set with noisy-label data +3: $\triangleright$ T: number of epochs +4: $\triangleright$ λ: a hyper-parameter +5: $\triangleright$ Notation: $\Theta^{(t)} = (\theta_y^{(t)},\theta_{\hat{y}}^{(t)},\epsilon^{(t)})$ +6: Initialise $\theta_y^{(1)},\theta_{\hat{y}}^{(1)},\epsilon^{(1)}$ and $\rho^{(0)}$ +7: $\theta_y^{(1)}\gets \mathrm{WARM~UP}(\mathcal{D},\theta_y^{(1)})$ +8: $t\gets 0$ +9: for $n_{\mathrm{epoch}} = 1:T$ do +10: for each mini-batch $\mathcal{S}$ in SHUFFLE(D) do +11: $t\gets t + 1$ +12: $\mathcal{S}_{\mathrm{clean}},\mathcal{S}_{\mathrm{noisy}}\gets \mathrm{SAMPLE~SELECTION}(\mathcal{S},\theta_y^{(t)},\epsilon^{(t)})$ +13: $\triangleright$ Variational E-step as in Eq. (3) +14: $\rho^{(t)}\gets \mathrm{E~STEP}(\mathcal{S},\theta_y^{(t)},\theta_{\hat{y}}^{(t)},\epsilon^{(t)},\rho^{(t - 1)})$ +15: $\triangleright$ M-step as in Eq. (8) +16: $\theta_y^{(t + 1)},\theta_{\hat{y}}^{(t + 1)},\epsilon^{(t + 1)}\gets \mathrm{M~STEP}(\mathcal{S}_{\mathrm{clean}},\mathcal{S}_{\mathrm{noisy}},\theta_y^{(t)},\theta_{\hat{y}}^{(t)},\epsilon^{(t)},\rho^{(t)},\lambda)$ +17: return $\theta_y$ +1 parameter of the clean-label classifier + +follow the latter approach by imposing a constraint (scoring term), denoted as $L(\theta_y, \epsilon^{(t)})$ , over $\theta_y$ in the M step via a sample selection approach based on the estimated noise rate $\epsilon^{(t)}$ . Formally, we propose a new curriculum when selecting samples as follows: + +$$ +R (t) = 1 - \epsilon^ {(t)}. \tag {5} +$$ + +In the simplest case, such as Co-teaching [16] or FINE [21], the constraint for $\theta_y$ can be written as: + +$$ +L \left(\theta_ {y}, \epsilon^ {(t)}\right) = \sum_ {\left(x _ {i}, \hat {y} _ {i}\right) \in \mathcal {S} _ {\text {c l e a n}}} \operatorname {K L} \left[ \operatorname {C a t e g o r i c a l} (Y; \hat {y}) \| \operatorname {C a t e g o r i c a l} \left(Y; f _ {\theta_ {y}} \left(x _ {i}\right)\right) \right], \tag {6} +$$ + +where: + +$$ +\mathcal {S} _ {\text {c l e a n}} = \left\{\left(x _ {i}, \hat {y} _ {i}\right): \left(x _ {i}, \hat {y} _ {i}\right) \in \mathcal {D} \wedge z \left(x _ {i}, y _ {i}\right) \leq m \right\}, +$$ + +$$ +m \in \{m: \Pr (z \left(x _ {i}, y _ {i}\right) \leq m) \geq R (t) \wedge \Pr (z \left(x _ {i}, y _ {i}\right) \geq m) \geq 1 - R (t) \}, +$$ + +$$ +\mathcal {S} _ {\text {n o i s y}} = \mathcal {D} \backslash \mathcal {S} _ {\text {c l e a n}}. \tag {7} +$$ + +with $R(t)$ defined in (5), and $z(x,y)$ representing the score of a criterion (e.g., loss [26, 58, 59], distance to the largest eigenvectors [21], or KNN scores [11]). Intuitively, the loss in Eq. (6) is simply the cross-entropy loss on the $\lfloor R(t)\times N\rfloor$ samples that have smallest scores (with $\lfloor .\rfloor$ being the floor function). One can also extend to other SOTA models by replacing the loss $L$ accordingly. For example, + +![](images/08c3a57bb884bedf5284bacc78cd61a6a3ead553af3eb4a53130280ac377ba00.jpg) +Fig. 3: Our training algorithm uses an existing noisy-label classifier (e.g., DivideMix [26]) parameterised by $\theta_y$ as the clean-label model $p(Y|X;\theta_y)$ . The generation of noisy-label (given $X$ and $Y$ ) is carried out by a model parameterised by noise rate $\epsilon$ and $\theta_{\tilde{y}}$ . The noisy-label classifier is based on a sample-selection mechanism that uses a curriculum $R(t) = 1 - \epsilon^{(t)}$ . + +if DivideMix [26] is used as a base model to constrain $\theta_y$ , $L$ will include two additional terms: loss on un-labeled data and regularisation using mixup [54]. + +# 3.3 Training and Testing + +Given the sample selection approach in Sec. 3.2, the M step in Eq. (4) is slightly modified by integrating the constraint in Eq. (6), which can be written as + +$$ +\theta_ {y} ^ {(t + 1)}, \theta_ {\hat {y}} ^ {(t + 1)}, \epsilon^ {(t + 1)} = \arg \max _ {\theta_ {y}, \theta_ {\hat {y}}, \epsilon} Q (\theta_ {y}, \theta_ {\hat {y}}, \epsilon , \rho^ {(t)}) - \lambda L (\theta_ {y}, \epsilon^ {(t)}), \tag {8} +$$ + +where $\lambda$ is a hyper-parameter and $L$ is defined in Eq. (6). The training procedure is summarised in Algorithm 1 and visualised in Fig. 3. In the implementation, we integrate the proposed method into existing models, such as DivideMix [26] or FINE [21]. Note that the clean-label classifier $f_{\theta_y}(.)$ is also the clean classifier of the base model. + +# 4 Experiments + +We show extensive experiments in several noisy-label synthetic benchmarks with CIFAR100 [22], and real-world benchmarks, including CNWL's red mini-ImageNet [17], Clothing1M [45] and mini-WebVision [27]. Sec. 4.2 describes implementation details. We evaluate our approach by plugging SOTA models into $p(y|x;\theta_y)$ , defined in Sec. 3, with results being shown in Sec. 4.3, and ablation studies in Sec. 4.4. + +# 4.1 Dataset Descriptions + +We follow the existing literature to generate the IDN labels [44] for CIFAR100 [22]. The dataset contains 50,000 training images and 10,000 testing images of shape $32 \times 32 \times 3$ . The dataset is class-balanced with 100 different categories and the IDN rates considered are 0.2, 0.3, 0.4 and 0.5 [44]. Another important IDN benchmark dataset is the red mini-ImageNet, a real-world dataset from CNWL [17]. In this dataset, there are 100 distinct classes, with each class containing 600 images. The images and their corresponding noisy-labels have been crawled from the internet at various controllable label noise rates. For a fair comparison with the existing literature [9, 15, 46], we have resized the images to $32 \times 32$ pixels from the $84 \times 84$ original pixel settings. We show results with $40\%$ , $60\%$ and $60\%$ noise rates on this dataset. Clothing1M [45] is a real-world clothing classification dataset that contains 1 million images with an estimated noise rate of $38.5\%$ . The dataset contains 14 different categories, and the labels are generated from surrounding texts. Images in this dataset are of different sizes, and we follow the resize structure suggested in [21, 26]. This dataset also contains $50k$ manually validated clean training images, $14k$ images for the validation set, and $10k$ testing images. We have not used the clean training, validation set, or any extra training images while training. Only the testing set is used for evaluation. Mini-WebVision [26] contains 65,944 images with their respective labels from the 50 different initial categories from the WebVision [27], with image size of $256 \times 256$ pixels. Following the evaluation process commonly used in this benchmark, 50 categories from the ILSVRC12 [23] dataset are also used for testing. + +# 4.2 Implementation + +All methods are implemented in Pytorch [36] and use one NVIDIA RTX 3090 card for training and testing. As mentioned in the original papers, hyperparameter settings are kept the same for the baselines used in the proposed algorithm. All classifier architectures are also kept the same as the baseline models. A random initialisation of noise rate parameter $\epsilon$ with the sigmoid as its activation function is employed for all experiments to maintain the fairness of the comparisons with other approaches. The value of $\lambda$ in Eq. (8) is set to 1 for all the cases. We integrate many SOTA approaches [11, 15, 21, 26, 58, 59] into our graphical model, as explained in Sec. 3.3. For CIFAR100 [22] with IDN [44], we integrate DivideMix [26], C2D [59], CC [58], and InstanceGM [15] into our model, given their superior performance across various noise rates. Additionally, we also use F-Dividemix from FINE [21], as shown in Fig. 2b. For red mini-ImageNet [17], we test our proposed approach with and without DINO self-supervision [4]. For the implementation without self-supervision, we use DivideMix [26] and InstanceGM [15], and for the self-supervised version, we only use InstanceGM [15]. The models for Clothing1M [45] and mini-WebVision [26] (validation on ImageNet [23]) are trained using DivideMix [26], SSR [11], C2D [59] and CC [58]. Additional implementation details, empirical analysis, and computational analysis are presented in the supplementary material in Appendices 1, 2 and 4, respectively. + +
MethodNoise Rates - IDN
0.20.30.40.5
CE [50]30.4224.1521.4514.42
PartT [44]65.3364.5659.7356.80
kMEIDTM [7]69.1666.7663.4659.18
DivideMix [26]77.0376.3370.8058.61
DivideMix-Ours77.4277.2172.4164.02
(+0.39)(+0.88)(+1.61)(+5.41)
C2D [59]78.6178.1872.8963.19
C2D-Ours79.0778.5973.3165.28
(+0.46)(+0.41)(+0.42)(+2.09)
CC [58]79.6177.5676.5863.19
CC-Ours79.7278.7177.3867.53
(+0.11)(+1.15)(+0.80)(+4.34)
InstanceGM [15]79.6979.2178.4777.19
InstanceGM-Ours79.6179.4079.5278.21
(-0.08)(+0.19)(+1.05)(+1.02)
+ +Table 1: Test accuracy % on (left) CIFAR100 [22] under different IDN [45], and (right) red mini-ImageNet [17]. Other models' results are from [7, 15, 46]. Here, we integrate DivideMix [26], C2D [59], CC [58] and InstanceGM [15] into our proposed model for CIFAR100 [22]. For red mini-ImageNet [17], we present the results with and without self-supervision [4]. We integrate DivideMix [26] and InstanceGM [15] into our model, with the latter tested with and without self-supervision. + +
MethodNoise rate
0.40.60.8
CE [46]42.7037.3029.76
MixUp [54]46.4040.5833.58
MentorMix [17]47.1443.8033.46
FaMUS [46]51.4245.1035.50
DivideMix [26]46.7243.1434.50
DivideMix-Ours50.7045.1137.44
(+3.98)(+1.97)(+2.94)
InstanceGM [15]52.2447.9639.62
InstanceGM-Ours56.6151.4043.83
(+4.37)(+3.44)(+4.21)
With self-supervised learning
PropMix [9]56.2252.8443.42
InstanceGM-SS [15]56.3753.2144.03
InstanceGM-SS-Ours58.2953.6045.47
(+1.92)(+0.39)(+1.44)
+ +Table 2: Final estimated noise rate $\epsilon$ for the Tab. 1, (left) CIFAR100 [22] under various IDN [44], and (right) red mini-ImageNet [17] with and without self-supervision. + +
Estimated noise rates
MethodActual noise rate
0.20.30.40.5
DivideMix-Ours0.180.340.470.53
C2D-Ours0.190.330.380.52
CC-Ours0.210.340.410.54
InstanceGM-Ours0.230.370.420.47
+ +
Estimated noise rates
MethodActual noise rate
0.40.6
DivideMix-Ours0.390.58
InstanceGM-Ours0.380.55
InstanceGM-SS-Ours0.480.53
0.73
+ +# 4.3 Comparison with SOTA + +This section compares our approach with SOTA methods on datasets with the IDN settings and noisy real-world settings. The bold text in tables indicate SOTA results, and our results are in the greyed rows. + +Synthetic Instance-Dependent Noise The comparison between various baselines and our proposed method on CIFAR100 [22] with IDN [44] is shown in Tab. 1 (left) with the noise rate ranging from 0.2 to 0.5. It is worth noting that using our proposed model with DivideMix [26], C2D [59], CC [58] and InstanceGM [15] improve their performance in almost all cases, particularly with large noise rates. Tab. 2 (left) also shows the final noise rate $\epsilon$ estimated by our model which is reasonably close to the simulated noise rate displayed in the table's header. + +Table 3: Test accuracy (\%) of competing methods, and final estimated noise rate $\epsilon$ on Clothing1M [45]. Also, we have not considered competing models that rely on a clean or validation set whilst training. We integrate DivideMix [26], C2D [59], SSR [11] and CC [58] into our model. DivideMix [26] and CC [58] shows the locally reproduced results. Mentioned noise rate by [45] is 0.385. + +
MethodTest accuracy (%)Estimated noise rate
OT-Filter [12]74.50
ELR+ with C2D [59]74.58
AugDesc [35]75.11
NCE [25]75.30
DivideMix [26]74.32
DivideMix-Ours74.41 (+0.09)0.41
C2D [59]74.58
C2D-Ours74.71 (+0.13)0.41
SSR (class-imbalance) [11]74.12
SSR-Ours74.20 (+0.08)0.42
CC [58]75.24
CC-Ours75.31 (+0.07)0.41
+ +Real-World Noise We also evaluate our proposed method on various real-world noisy settings regarding test accuracy and estimated noise rates $\epsilon$ in Tabs. 1 to 4. Similarly to the synthetic IDN in Tab. 1 (left), the results show that existing noisy-label robust methods can be easily integrated with our model to outperform current SOTA results for real-world noisy-label datasets. Tabs. 1 and 2 (right) shows the results on red mini-ImageNet using two configurations, including cases without self-supervision (top part of the table) and with self-supervision (bottom part of the table). The self-supervision DINO pre-training [4] relies only on images from red mini-ImageNet to enable a fair comparison with existing baselines [9, 15]. Results from Tab. 1 (right) demonstrate that our approach improves the performance of SOTA methods by a considerable margin in all cases. In fact, using estimated noise rate $\epsilon$ (Tab. 2 (right)) while training InstanceGM [15] without self-supervision shows better performance than existing self-supervised baselines at 0.4 noise rate. Moreover, DivideMix [26], C2D [59], SSR [11], and CC [58] are used as a baselines for Clothing1M [45] as shown in Tab. 3, with results showing slight improvements with the use of our method. For mini-WebVision [26] (validation on ImageNet [23]), we use DivideMix [26], C2D [59], SSR [11], and CC [58] as baselines (Tab. 4). It is worth noting that our results are better than the SOTA for all settings in Tab. 4. + +# 4.4 Ablation + +We show an ablation study with testing accuracy (top) and training time (bottom) of our approach in Tab. 5 (left) on CIFAR100 [22] at 0.3 and 0.5 IDN [44] using DivideMix [26] as baseline. Detailed complexity analysis is shown in the supplementary material, Tab. 10. Initially, the accuracy result of baseline DivideMix [26] is $76.33\%$ , $58.61\%$ for 0.3 and 0.5 IDN, respectively. In the second row, we fix the noise rate $\epsilon$ at 0.3 and 0.5 for DivideMix's sample selection, as explained in Sec. 3.3 (without updating $\epsilon$ ), then the results improved to $78.14\%$ and $64.44\%$ , + +Table 4: Test accuracy (\%) and final estimated noise rate $\epsilon$ on mini-WebVision [26] and validation on ImageNet [23]. We integrate DivideMix [26], Contrast-to-Divide (C2D) [59], SSR [11] and CC [58] into our model, whilst models suffixed with -Ours denotes our proposed approach. ImageNet [23] is only considered for validation. + +
Datasetmini-WebVisionImageNetEstimated noise rate
Top-1Top-5Top-1Top-5
BtR [38]80.8892.7675.9692.20
NCE [25]79.5093.8076.3094.10
DivideMix [26]77.3291.6475.2091.64
DivideMix-Ours78.5192.0376.1193.240.43
(+1.19)(+0.39)(+0.91)(+1.60)
C2D [59]79.4292.3278.5793.04
C2D-Ours80.2092.8279.1693.120.43
(+0.78)(+0.50)(+0.59)(+0.08)
SSR [11]80.9292.8075.7691.76
SSR-Ours81.6893.8076.9193.050.43
(+0.76)(+1.00)(+1.15)(+1.29)
CC [58]79.3693.6476.0893.86-
CC-Ours80.0193.7977.1194.210.44
(+0.65)(+0.15)(+1.03)(+0.35)
+ +Table 5: (left) Ablation results on CIFAR100 at 0.3 and 0.5 IDN. We display accuracy results under different configurations and compare the training time of our approach against DivideMix [26]. The implementation settings and detailed computational analysis are described in Sec. 4.2 and Appendix 4 respectively. (right above) Ablation results for initialisation with different $\epsilon$ values on CIFAR100 [22] at a synthetic noise rate of 0.5 IDN [44] with DivideMix-Ours. (right below) Ablation graph showing the learning pattern of the noise rate parameter estimation under different initialisation values, i.e., $\epsilon = 0.001$ (blue), $\epsilon = 0.1$ (orange), $\epsilon = 0.9$ (red), when trained on DivideMix-ours case at 0.5 IDN [44] on CIFAR100 [22]. + +
ModelsAccuracy (%)
0.30.5
DivideMix [26]76.3358.61
DivideMix with fixed ε78.1464.44
Our method with pre-trained DivideMix75.0152.31
Our method with original DivideMix (no ε)76.2056.30
DivideMix-Ours77.2164.02
Training Time (GPU-hours)
DivideMix [26]7.20
DivideMix-Ours8.50
DivideMix-Ours (Mixed Precision)6.10
+ +![](images/aa983ece11a754aafafd52ae1d8727f56159fc784823abeb033ac358abf72881.jpg) + +which is the ideal case (i.e., a perfect noise rate estimation) that motivated our work. In the third case, we use the proposed graphical model with pre-trained DivideMix [26] that shows accuracy of $75.01\%$ and $52.31\%$ . In the next case, the proposed graphical model is trained together with DivideMix [26] without considering the estimated noise rate $\epsilon$ for sample selection, which results in accuracy of $76.20\%$ and $56.30\%$ . In the last row, we show the training of the proposed model with DivideMix [26], together with the estimation of noise rate $\epsilon$ , and the + +selection of samples based on that, with $\approx 1\%$ and $\approx 8\%$ accuracy improvement to 0.3 and 0.5 IDN, which is quite close to our ideal case (second row). We also study the effect that the initialisation of the noise rate parameter $\epsilon$ plays in our method. As described in Sec. 4.2, we randomly initialise $\epsilon$ within the range $(0,1)$ . To test the robustness of our method to this initialisation, we perform an ablation study by initialising $\epsilon$ with three different values, namely: 0.001, 0.1, and 0.9 on CIFAR100 [24] at 0.5 IDN [44] using the DivideMix-ours approach. Our results, presented in Tab. 5 (right above and below), demonstrate that the initial value of $\epsilon$ has a limited effect on the model's performance, highlighting the stability of our method across different initialisation of $\epsilon$ . In the supplementary material, Tab. 7 (left) presents comparative analyses of our model against PartT [44] and SSR [11] on CIFAR100 [22] at 0.5 IDN [44], while Tab. 7 (right) showcases similar comparisons with baselines DivideMix [26] and InstanceGM [15] at a higher noise rate of 0.6 IDN [44]. Additionally, Tab. 8 provides further insights into the performance of DivideMix [26] on red mini-ImageNet [17] at lower noise rates of 0.1, 0.2 and 0.3. Tab. 9 illustrates noise rate estimation with standard deviation for our model integrated with DivideMix [26] and InstanceGM [15] on CIFAR100 [22] under IDN [44] settings at 0.2, 0.3, 0.4, and 0.5. Further statistical analysis is detailed in Appendix 5 with confidence interval presented in Tab. 12. Performance analysis with various architectures (Tab. 13) and lambda values (Tab. 14) are mentioned in Appendix 6. Threshold-based comparisons are in Appendix 7 and Tab. 15. + +# 5 Discussion and Conclusion + +In this work, we proposed an innovative graphical model for IDN noisy-label learning, focusing on the estimation of the label noise rate that is leveraged to introduce a robust sample selection curriculum. Given that SOTA IDN noisy-label learning approaches tend to rely on sample selection methods, our method can be seamlessly integrated to them to improve their performance in many synthetic and real-world benchmarks, including CIFAR100 [22], red mini-ImageNet [17], Clothing1M [45], mini-WebVision [26], and ImageNet [23]. + +From a societal perspective, our methodology offers a positive impact by mitigating biases inherent in noisy data, thereby promoting more fair and accurate machine learning outcomes. It also opens the door to future research avenues, including exploring advanced noise rate estimation methodologies and investigating noise-robust loss functions such as GCE [56] and ELR [28] with it. Our choice of using the same network architecture for clean and noisy labels enables compatibility with co-teaching techniques [16,26]. However, investigating different network structures for clean and noisy labels remains a potential avenue for future research. Moreover, we aim to delve into its dynamics, uncovering substantial improvements for certain models while observing more marginal enhancements for others. For more detailed information, please refer to Appendix 8. Our approach also hints at potential breakthroughs in streamlining data annotation challenges. + +# References + +1. Arpit, D., Jastrzebski, S., Ballas, N., Krueger, D., Bengio, E., Kanwal, M.S., Maharaj, T., Fischer, A., Courville, A., Bengio, Y., Lacoste-Julien, S.: A closer look at memorization in deep networks. In: International Conference on Machine Learning. vol. 70, pp. 233-242. PMLR (2017) +2. Bae, H., Shin, S., Jang, J., Na, B., Song, K., Moon, I.C.: From noisy prediction to true label: Noisy prediction calibration via generative model. In: International Conference on Machine Learning (2022) +3. Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.A.: Mixmatch: A holistic approach to semi-supervised learning. In: Advances in Neural Information Processing Systems. vol. 32 (2019) +4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: International Conference on Computer Vision. pp. 9650-9660 (2021) +5. Chen, P., Ye, J., Chen, G., Zhao, J., Heng, P.A.: Beyond class-conditional assumption: A primary attempt to combat instance-dependent label noise. In: AAAI Conference on Artificial Intelligence. pp. 11442–11450 (2021) +6. Chen, Z., Song, A., Wang, Y., Huang, X., Kong, Y.: A noise rate estimation method for image classification with label noise. Journal of Physics: Conference Series 2433(1), 012039 (feb 2023). https://doi.org/10.1088/1742-6596/2433/1/012039, https://dx.doi.org/10.1088/1742-6596/2433/1/012039 +7. Cheng, D., Liu, T., Ning, Y., Wang, N., Han, B., Niu, G., Gao, X., Sugiyama, M.: Instance-dependent label-noise learning with manifold-regularized transition matrix estimation. In: Conference on Computer Vision and Pattern Recognition. pp. 16630-16639 (2022) +8. Cheng, H., Zhu, Z., Li, X., Gong, Y., Sun, X., Liu, Y.: Learning with instance-dependent label noise: A sample sieve approach. In: International Conference on Learning Representations (2020) +9. Cordeiro, F.R., Belagiannis, V., Reid, I., Carneiro, G.: PropMix: Hard sample filtering and proportional mixup for learning with noisy labels. In: British Machine Vision Conference (2021) +0. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the EM algorithm. Journal of the royal statistical society: series B (methodological) 39(1), 1-22 (1977) +1. Feng, C., Tzimiropoulos, G., Patras, I.: SSR: An efficient and robust framework for learning with unknown label noise. In: British Machine Vision Conference (2022) +2. Feng, C., Ren, Y., Xie, X.: Ot-filter: An optimal transport filter for learning with noisy labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16164-16174 (2023) +3. Frénay, B., Kabán, A.: A comprehensive introduction to label noise. In: European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (2014) +4. Frénay, B., Verleysen, M.: Classification in the presence of label noise: A survey. IEEE Transactions on Neural Networks and Learning Systems 25(5), 845-869 (2013) +5. Garg, A., Nguyen, C., Felix, R., Do, T.T., Carneiro, G.: Instance-dependent noisy label learning via graphical modelling. In: Winter Conference on Applications of Computer Vision. pp. 2288-2298 (2023) + +16. Han, B., Yao, Q., Yu, X., Niu, G., Xu, M., Hu, W., Tsang, I., Sugiyama, M.: Co-teaching: Robust training of deep neural networks with extremely noisy labels. In: Advances in Neural Information Processing Systems. vol. 31 (2018) +17. Jiang, L., Huang, D., Liu, M., Yang, W.: Beyond synthetic noise: Deep learning on controlled noisy labels. In: International Conference on Machine Learning. pp. 4804-4815. PMLR (2020) +18. Jiang, L., Zhou, Z., Leung, T., Li, L.J., Fei-Fei, L.: MentorNet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In: International Conference on Machine Learning. pp. 2304-2313. PMLR (2018) +19. Jiang, Z., Zhou, K., Liu, Z., Li, L., Chen, R., Choi, S.H., Hu, X.: An information fusion approach to learning with instance-dependent label noise. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=ecH2FKaARUp +20. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in Neural Information Processing Systems 33, 18661-18673 (2020) +21. Kim, T., Ko, J., Choi, J., Yun, S.Y.: FINE samples for learning with noisy labels. In: Advances in Neural Information Processing Systems. vol. 34 (2021) +22. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. Tech. rep., University of Toronto (2009) +23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: ImageNet classification with deep convolutional neural networks. In: Advances in Neural Information Processing Systems. vol. 25 (2012) +24. Kye, S.M., Choi, K., Yi, J., Chang, B.: Learning with noisy labels by efficient transition matrix estimation to combat label miscorrection. In: European Conference on Computer Vision. pp. 717-738. Springer (2022) +25. Li, J., Li, G., Liu, F., Yu, Y.: Neighborhood collective estimation for noisy label identification and correction. In: European Conference on Computer Vision. pp. 128-145. Springer (2022) +26. Li, J., Socher, R., Hoi, S.C.: DivideMix: Learning with noisy labels as semi-supervised learning. In: International Conference on Learning Representations (2020) +27. Li, W., Wang, L., Li, W., Agustsson, E., Gool, L.V.: WebVision Database: Visual learning and understanding from web data. CoRR (2017) +28. Liu, S., Niles-Weed, J., Razavian, N., Fernandez-Granda, C.: Early-learning regularization prevents memorization of noisy labels. In: Advances in Neural Information Processing Systems. vol. 33, pp. 20331-20342 (2020) +29. Liu, Y., Cheng, H., Zhang, K.: Identifiability of label noise transition matrix. In: International Conference on Machine Learning (2023) +30. Ma, X., Huang, H., Wang, Y., Romano, S., Erfani, S., Bailey, J.: Normalized loss functions for deep learning with noisy labels. In: International Conference on Machine Learning. pp. 6543-6553. PMLR (2020) +31. Malach, E., Shalev-Shwartz, S.: Decoupling "when to update" from "how to update". In: Advances in Neural Information Processing Systems. vol. 30 (2017) +32. McLachlan, G.J., Peel, D.: Finite mixture models. Wiley (1996) +33. Neal, R.M., Hinton, G.E.: A view of the EM algorithm that justifies incremental, sparse, and other variants. Learning in graphical models pp. 355-368 (1998) +34. Neyshabur, B., Bhojanapalli, S., McAllester, D., Srebro, N.: Exploring generalization in deep learning. In: Advances in Neural Information Processing Systems. vol. 30 (2017) + +35. Nishi, K., Ding, Y., Rich, A., Hollerer, T.: Augmentation strategies for learning with noisy labels. In: Conference on Computer Vision and Pattern Recognition. pp. 8022-8031 (2021) +36. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: PyTorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems 32 (2019) +37. Shamshad, F., Khan, S., Zamir, S.W., Khan, M.H., Hayat, M., Khan, F.S., Fu, H.: Transformers in medical imaging: A survey. Medical Image Analysis p. 102802 (2023) +38. Smart, B., Carneiro, G.: Bootstrapping the relationship between images and their clean and noisy labels. In: Winter Conference on Applications of Computer Vision. pp. 5344-5354 (2023) +39. Song, H., Kim, M., Park, D., Shin, Y., Lee, J.G.: Learning from noisy labels with deep neural networks: A survey. IEEE Transactions on Neural Networks and Learning Systems (2022) +40. Tanaka, D., Ikami, D., Yamasaki, T., Aizawa, K.: Joint optimization framework for learning with noisy labels. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) +41. Titterington, D.M.: On the likelihood of mixtures of distributions. Biometrika 71(3), 511-522 (1984). https://doi.org/10.1093/biomet/71.3.511 +42. Trummer, I.: From BERT to GPT-3 codex: harnessing the potential of very large language models for data management. VLDB Endowment 15(12), 3770-3773 (2022) +43. Tu, H., Menzies, T.: Debtfree: minimizing labeling cost in self-admitted technical debt identification using semi-supervised learning. Empirical Software Engineering 27(4), 80 (2022) +44. Xia, X., Liu, T., Han, B., Wang, N., Gong, M., Liu, H., Niu, G., Tao, D., Sugiyama, M.: Part-dependent label noise: Towards instance-dependent label noise. In: Advances in Neural Information Processing Systems. vol. 33, pp. 7597-7610 (2020) +45. Xiao, T., Xia, T., Yang, Y., Huang, C., Wang, X.: Learning from massive noisy labeled data for image classification. In: Conference on Computer Vision and Pattern Recognition. pp. 2691-2699 (2015) +46. Xu, Y., Zhu, L., Jiang, L., Yang, Y.: Faster meta update strategy for noise-robust deep learning. In: Conference on Computer Vision and Pattern Recognition. pp. 144-153 (June 2021) +47. Yang, S., Jiang, L., Liu, Z., Loy, C.C.: Unsupervised image-to-image translation with generative prior. In: Conference on Computer Vision and Pattern Recognition. pp. 18332-18341 (2022) +48. Yang, S., Yang, E., Han, B., Liu, Y., Xu, M., Niu, G., Liu, T.: Estimating instance-dependent bayes-label transition matrix using a deep neural network. In: International Conference on Machine Learning. pp. 25302-25312. PMLR (2022) +49. Yao, Q., Yang, H., Han, B., Niu, G., Kwok, J.T.Y.: Searching to exploit memorization effect in learning with noisy labels. In: International Conference on Machine Learning. pp. 10789-10798. PMLR (2020) +50. Yao, Y., Liu, T., Gong, M., Han, B., Niu, G., Zhang, K.: Instance-dependent label-noise learning under a structural causal model. In: Advances in Neural Information Processing Systems. vol. 34 (2021) + +51. Yao, Y., Liu, T., Han, B., Gong, M., Deng, J., Niu, G., Sugiyama, M.: Dual T: Reducing estimation error for transition matrix in label-noise learning. In: Advances in Neural Information Processing Systems. vol. 33, pp. 7260-7271 (2020) +52. Zhang, C., Bengio, S., Hardt, M., Recht, B., Vinyals, O.: Understanding deep learning requires rethinking generalization. In: International Conference on Learning Representations (2017) +53. Zhang, C., Bengio, S., Hardt, M., Recht, B., Vinyals, O.: Understanding deep learning (still) requires rethinking generalization. Communications of the ACM 64(3), 107-115 (2021) +54. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. In: International Conference on Learning Representations (2017) +55. Zhang, Y., Niu, G., Sugiyama, M.: Learning noise transition matrix from only noisy labels via total variation regularization. In: International Conference on Machine Learning. pp. 12501-12512. PMLR (2021) +56. Zhang, Z., Sabuncu, M.: Generalized cross entropy loss for training deep neural networks with noisy labels. Advances in Neural Information Processing Systems 31 (2018) +57. Zhang, Z., Pfister, T.: Learning fast sample re-weighting without reward data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 725-734 (2021) +58. Zhao, G., Li, G., Qin, Y., Liu, F., Yu, Y.: Centrality and consistency: two-stage clean samples identification for learning with instance-dependent noisy labels. In: European Conference on Computer Vision. pp. 21-37. Springer (2022) +59. Zheltonozhskii, E., Baskin, C., Mendelson, A., Bronstein, A.M., Litany, O.: Contrast to divide: Self-supervised pre-training for learning with noisy labels. In: Winter Conference on Applications of Computer Vision. pp. 1657-1667 (2022) +60. Zhu, Z., Liu, T., Liu, Y.: A second-order approach to learning with instance-dependent label noise. In: Conference on Computer Vision and Pattern Recognition. pp. 10113-10123 (June 2021) \ No newline at end of file diff --git a/2024/Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation/images.zip b/2024/Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..60d521cbc9a7af058b691ea32d8c2864ef2a82f3 --- /dev/null +++ b/2024/Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93bf86638b5299bfed8fe3ebe7e6521eed565469240e75207e41b2d515310b67 +size 406365 diff --git a/2024/Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation/layout.json b/2024/Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..96b8ee51b00ee9be251d8445c22f8c883ad05b8b --- /dev/null +++ b/2024/Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation/layout.json @@ -0,0 +1,10526 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 111, + 470, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 111, + 470, + 148 + ], + "spans": [ + { + "bbox": [ + 146, + 111, + 470, + 148 + ], + "type": "text", + "content": "Instance-dependent Noisy-label Learning with Graphical Model Based Noise-rate Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 167, + 470, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 167, + 470, + 194 + ], + "spans": [ + { + "bbox": [ + 145, + 167, + 470, + 194 + ], + "type": "text", + "content": "Arpit Garg1, Cuong Nguyen3, Rafael Felix1, Thanh-Toan Do2, and Gustavo Carneiro3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 201, + 474, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 201, + 474, + 223 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 474, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 201, + 474, + 223 + ], + "type": "text", + "content": " Australian Institute for Machine Learning, University of Adelaide, SA, Australia " + }, + { + "bbox": [ + 138, + 201, + 474, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 201, + 474, + 223 + ], + "type": "text", + "content": " Department of Data Science and AI, Monash University, Australia" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 150, + 224, + 462, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 224, + 462, + 235 + ], + "spans": [ + { + "bbox": [ + 150, + 224, + 462, + 235 + ], + "type": "text", + "content": "3 Centre for Vision, Speech and Signal Processing, University of Surrey, UK" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 266, + 455, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 266, + 455, + 431 + ], + "spans": [ + { + "bbox": [ + 159, + 266, + 455, + 431 + ], + "type": "text", + "content": "Abstract. Deep learning faces a formidable challenge when handling noisy labels, as models tend to overfit samples affected by label noise. This challenge is further compounded by the presence of instance-dependent noise (IDN), a realistic form of label noise arising from ambiguous sample information. To address IDN, Label Noise Learning (LNL) incorporates a sample selection stage to differentiate clean and noisy-label samples. This stage uses an arbitrary criterion and a pre-defined curriculum that initially selects most samples as noisy and gradually decreases this selection rate during training. Such curriculum is sub-optimal since it does not consider the actual label noise rate in the training set. This paper addresses this issue with a new noise-rate estimation method that is easily integrated with most state-of-the-art (SOTA) LNL methods to produce a more effective curriculum. Synthetic and real-world benchmarks' results demonstrate that integrating our approach with SOTA LNL methods improves accuracy in most cases." + }, + { + "bbox": [ + 159, + 266, + 455, + 431 + ], + "type": "inline_equation", + "content": "^4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 441, + 455, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 441, + 455, + 464 + ], + "spans": [ + { + "bbox": [ + 159, + 441, + 455, + 464 + ], + "type": "text", + "content": "Keywords: Noisy-labels " + }, + { + "bbox": [ + 159, + 441, + 455, + 464 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 441, + 455, + 464 + ], + "type": "text", + "content": " Instance-dependent noise " + }, + { + "bbox": [ + 159, + 441, + 455, + 464 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 441, + 455, + 464 + ], + "type": "text", + "content": " Label noise Learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 485, + 230, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 485, + 230, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 485, + 230, + 498 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 483, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 483, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 483, + 631 + ], + "type": "text", + "content": "Deep neural networks (DNNs) have demonstrated their effectiveness in various domains, including vision [47], language [42], and medicine [37]. However, its efficacy is largely dependent on high-quality training data, which can be resource-intensive to obtain [43]. Although cost-effective labelling techniques, such as data mining [11] and crowdsourcing [39], offer cheaper alternatives, they often compromise label quality [39]. Consequently, this can lead to erroneous labels in real-world datasets [21]. This is a relevant issue because even slight inaccuracies in the labels can significantly affect the performance of DNN due to their inherent memorisation capabilities [34, 53]. This has prompted the development of algorithms resilient to noisy-label learning, with the aim of training models" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 638, + 482, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 638, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 638, + 482, + 665 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 133, + 638, + 482, + 665 + ], + "type": "text", + "content": " Code is available at https://github.com/arpit2412/NoiseRateLearning. Supported by the Engineering and Physical Sciences Research Council (EPSRC) through grant EP/Y018036/1 and the Australian Research Council (ARC) through grant FT190100525." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "despite inaccuracies in training data labels. Although various strategies exist, our work uniquely proposes an approach centered on the estimation of the label noise rate from the training data, which is notably absent from current methodologies." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 155, + 482, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 155, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 155, + 482, + 467 + ], + "type": "text", + "content": "There are two types of label noise, namely instance-independent noise (IIN) [16] and instance-dependent noise (IDN) [44]. Such type of label noise typically influences the design principles of noisy-label learning algorithms. For example, IIN focuses on mislabellings that are independent of sample information [16], where estimating the underlying label transition matrix is a common way of handling this type of noise [51]. On the other hand, in the more realistic IDN, mislabelling is due to both sample information and true class labels [44], which generally require the combination of many label noise learning techniques, such as robust loss functions [28, 56], and noisy-label sample selection [26, 58]. Of the strategies mentioned above, sample selection approaches that classify training data into clean and noisy samples have produced competitive results on many benchmarks [9, 11, 15, 21, 26]. Such sample selection techniques require the definition of a classification criterion and a selection curriculum. Many studies on this topic focus on developing new sample selection criteria, such as the small-loss hypothesis [26], which states that noisy-label samples have higher loss values than clean-label samples, particularly at the early stage of training [1]. Another criterion type is the feature-based one. An example of such criterion is FINE [21] that discriminates clean and noisy-label samples via the distance to class-specific eigenvectors. In this technique, clean-label samples tend to lie closer to the class-specific dominant eigenvector of the latent representations than noisy-label samples. Another type of criterion is proposed in SSR [11], which introduces a selection criterion based on the K-nearest-neighbor (KNN) classification in the feature space. Furthermore, CC [58] uses a two-stage sampling procedure, including class-level feature clustering followed by a consistency score. An equally important problem in sample selection is the definition of the curriculum for selecting clean training samples, but it has received comparatively less attention." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 470, + 483, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 483, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 483, + 615 + ], + "type": "text", + "content": "The sample selection curriculum defines a threshold to be used with one of the criteria listed above to classify the training samples as clean or noisy at each training epoch [45]. For example, the threshold can be fixed to an arbitrary clustering score that separates clean and noisy samples [26], but this strategy does not account for the proportion of label noise in the training set, nor does it consider the dynamics of the selection of noisy-label samples during the training. The consideration of such dynamics has been studied in [16, 49], which defined a curriculum of the noisy-label sampling rate " + }, + { + "bbox": [ + 130, + 470, + 483, + 615 + ], + "type": "inline_equation", + "content": "R(t)" + }, + { + "bbox": [ + 130, + 470, + 483, + 615 + ], + "type": "text", + "content": " as a function of the training epoch " + }, + { + "bbox": [ + 130, + 470, + 483, + 615 + ], + "type": "inline_equation", + "content": "t \\in \\{1, \\dots, T\\}" + }, + { + "bbox": [ + 130, + 470, + 483, + 615 + ], + "type": "text", + "content": ". The curriculum " + }, + { + "bbox": [ + 130, + 470, + 483, + 615 + ], + "type": "inline_equation", + "content": "R(t)" + }, + { + "bbox": [ + 130, + 470, + 483, + 615 + ], + "type": "text", + "content": " defines a sampling rate close to " + }, + { + "bbox": [ + 130, + 470, + 483, + 615 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 470, + 483, + 615 + ], + "type": "text", + "content": " of the training set at the beginning of the training, which is then reduced to arbitrarily low rates at the end of the training. In practice, the function " + }, + { + "bbox": [ + 130, + 470, + 483, + 615 + ], + "type": "inline_equation", + "content": "R(t)" + }, + { + "bbox": [ + 130, + 470, + 483, + 615 + ], + "type": "text", + "content": " is predefined [16] or learned by weighting a set of basis functions [49]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Although generally effective, these techniques do not consider the label noise rate estimated from the training set, making them vulnerable to over-fitting (if too many noisy-label samples are classified as clean) or under-fitting (if informative clean-label samples are classified as noisy). It can be argued that the estimation" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "A. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 301, + 229 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 301, + 229 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 301, + 229 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 301, + 229 + ], + "type": "image", + "image_path": "a180e75bd1f43c8f3c626a7caf5e2fc0fcf82f6c0d858ceb5e6d660de8757ded.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 244, + 246, + 253 + ], + "lines": [ + { + "bbox": [ + 192, + 244, + 246, + 253 + ], + "spans": [ + { + "bbox": [ + 192, + 244, + 246, + 253 + ], + "type": "text", + "content": "(a) Motivation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 113, + 471, + 243 + ], + "blocks": [ + { + "bbox": [ + 313, + 113, + 471, + 243 + ], + "lines": [ + { + "bbox": [ + 313, + 113, + 471, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 113, + 471, + 243 + ], + "type": "image", + "image_path": "7d25d80cae29bd7cfecbfdbad32f62383c9de642ce5c58d654608ce8e8055e3e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 244, + 448, + 253 + ], + "lines": [ + { + "bbox": [ + 339, + 244, + 448, + 253 + ], + "spans": [ + { + "bbox": [ + 339, + 244, + 448, + 253 + ], + "type": "text", + "content": "(b) Proposed Graphical Model" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 262, + 482, + 350 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 350 + ], + "type": "text", + "content": "Fig. 1: (a) Comparison of test accuracy " + }, + { + "bbox": [ + 130, + 262, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 130, + 262, + 482, + 350 + ], + "type": "text", + "content": " (as a function of training epoch) between the original DivideMix [26] (solid, blue curve) and our modified DivideMix (dashed, red curve) that selects the clean and noisy data based on a fixed noise rate " + }, + { + "bbox": [ + 130, + 262, + 482, + 350 + ], + "type": "inline_equation", + "content": "R(t) = 1 - \\epsilon = 50\\%" + }, + { + "bbox": [ + 130, + 262, + 482, + 350 + ], + "type": "text", + "content": " using the small-loss criterion on CIFAR100 [22] at 0.5 IDN [44]; (b) The proposed probabilistic graphical model that generates noisy-label " + }, + { + "bbox": [ + 130, + 262, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 130, + 262, + 482, + 350 + ], + "type": "text", + "content": " conditioned on the image " + }, + { + "bbox": [ + 130, + 262, + 482, + 350 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 262, + 482, + 350 + ], + "type": "text", + "content": ", the latent clean-label " + }, + { + "bbox": [ + 130, + 262, + 482, + 350 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 262, + 482, + 350 + ], + "type": "text", + "content": " and noise rate " + }, + { + "bbox": [ + 130, + 262, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 262, + 482, + 350 + ], + "type": "text", + "content": ", where forward pass (solid lines) is parameterised by " + }, + { + "bbox": [ + 130, + 262, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\theta_y, \\theta_{\\hat{y}}" + }, + { + "bbox": [ + 130, + 262, + 482, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 262, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 262, + 482, + 350 + ], + "type": "text", + "content": " representing the generation step, and the backward pass (dashed lines) is parameterised by " + }, + { + "bbox": [ + 130, + 262, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 262, + 482, + 350 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "content": "of the label transition matrix [6,7,48,51] aims to recover the noise rate affecting pairwise label transitions. However, label transition matrix techniques follow a quite different strategy compared to sample selection methods, where their main challenge is the general underconstrained aspect of the matrix estimation, making them sensitive to large noise rates and not scalable to a high number of classes [39]. Although several methods address label noise, none estimates the label noise rate directly from the training data to guide sample selection. Addressing this gap is the primary challenge of our study." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 473, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 628 + ], + "type": "text", + "content": "To underscore the importance of using the noise rate for sample selection during training, we experiment with CIFAR100 [22] at an IDN rate " + }, + { + "bbox": [ + 130, + 473, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\epsilon = 50\\%" + }, + { + "bbox": [ + 130, + 473, + 482, + 628 + ], + "type": "text", + "content": " [44] (noise rate specifications and other details are explained in Sec. 4). We use DivideMix [26] and replace its sample selection (based on an arbitrary clustering score [10,32,41]) by a thresholding process that classifies the " + }, + { + "bbox": [ + 130, + 473, + 482, + 628 + ], + "type": "inline_equation", + "content": "R(t) = 1 - \\epsilon = 50\\%" + }, + { + "bbox": [ + 130, + 473, + 482, + 628 + ], + "type": "text", + "content": " largest loss samples as noisy, and the remaining ones as clean in all training epochs " + }, + { + "bbox": [ + 130, + 473, + 482, + 628 + ], + "type": "inline_equation", + "content": "t \\in \\{1, \\dots, T\\}" + }, + { + "bbox": [ + 130, + 473, + 482, + 628 + ], + "type": "text", + "content": ". This sample selection is used for the semi-supervised learning of DivideMix [26]. As shown in Fig. 1a, the new sample selection approach based on the \"manually provided\" noise rate (dashed red curve) improves " + }, + { + "bbox": [ + 130, + 473, + 482, + 628 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 130, + 473, + 482, + 628 + ], + "type": "text", + "content": " in terms of prediction accuracy compared to the original DivideMix [26] (solid blue curve) which relies on arbitrary thresholding. Similar conclusions can be achieved with other methods that apply sample selection strategies to address the noisy-label learning problem, as shown in the experiments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "In this paper, we introduce a new sample selection strategy centered on estimating the label noise rate of the training set. Our strategy is based on our novel noisy-label learning graphical model illustrated in Fig. 1b. This model can be" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "type": "text", + "content": "Noisy-Label Learning with Noise-Rate Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 116, + 297, + 224 + ], + "blocks": [ + { + "bbox": [ + 135, + 116, + 297, + 224 + ], + "lines": [ + { + "bbox": [ + 135, + 116, + 297, + 224 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 297, + 224 + ], + "type": "image", + "image_path": "c2fc73a33c98dface34ed4f3415f9bc7caaa303299e41de6b32bf051a54fe249.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 228, + 271, + 239 + ], + "lines": [ + { + "bbox": [ + 165, + 228, + 271, + 239 + ], + "spans": [ + { + "bbox": [ + 165, + 228, + 271, + 239 + ], + "type": "text", + "content": "(a) Curriculum function " + }, + { + "bbox": [ + 165, + 228, + 271, + 239 + ], + "type": "inline_equation", + "content": "R(t)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 113, + 459, + 225 + ], + "blocks": [ + { + "bbox": [ + 312, + 113, + 459, + 225 + ], + "lines": [ + { + "bbox": [ + 312, + 113, + 459, + 225 + ], + "spans": [ + { + "bbox": [ + 312, + 113, + 459, + 225 + ], + "type": "image", + "image_path": "d989ee52eda15970afc13f9f66dc0f90d99152f959141d0db19c6b74d4948716.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 228, + 441, + 239 + ], + "lines": [ + { + "bbox": [ + 348, + 228, + 441, + 239 + ], + "spans": [ + { + "bbox": [ + 348, + 228, + 441, + 239 + ], + "type": "text", + "content": "(b) Accuracy Comparison" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "type": "text", + "content": "Fig. 2: (a) Visual comparison of different " + }, + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "type": "inline_equation", + "content": "R(t)" + }, + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "type": "text", + "content": " on CIFAR100 [22] at 0.5 IDN [44]: (i) Co-teaching [16] with curriculum based on hyper-parameter " + }, + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "type": "inline_equation", + "content": "T_{k} \\in \\{5,10\\}" + }, + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "type": "inline_equation", + "content": "R(t) = 1 - \\tau \\cdot \\min(t / T_{k}, 1)" + }, + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\tau = \\epsilon = 0.5" + }, + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "type": "text", + "content": " being manually set; (ii) S2E [49], where " + }, + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "type": "inline_equation", + "content": "R(t)" + }, + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "type": "text", + "content": " is estimated with a bi-level optimisation; (iii) SSR [11], where " + }, + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "type": "inline_equation", + "content": "R(t)" + }, + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "type": "text", + "content": " is based on a relabelling pattern, (iv) CC [58] following the small-loss hypothesis [26] and cosine similarity for " + }, + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "type": "inline_equation", + "content": "R(t)" + }, + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "type": "text", + "content": "; and (v) ours. (b) Accuracy of noisy-label learning methods on CIFAR100 [22] at 0.5 IDN [44], including DivideMix [26], FINE [21] and InstanceGM [15], without (left, blue) and (right, orange) integration of our proposed graphical model for estimating the noise rate " + }, + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "type": "text", + "content": " and sample selection rate based on " + }, + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 247, + 482, + 346 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 370, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 482, + 478 + ], + "type": "text", + "content": "seamlessly integrated with state-of-the-art (SOTA) noisy-label learning techniques, providing them with a precise noise rate estimate and subsequently refining the sample selection curriculum. In particular, our model's curriculum is not restrained by predefined " + }, + { + "bbox": [ + 130, + 370, + 482, + 478 + ], + "type": "inline_equation", + "content": "R(t)" + }, + { + "bbox": [ + 130, + 370, + 482, + 478 + ], + "type": "text", + "content": " functions [16, 49], but rather relies on a dynamically estimated noise rate sourced directly from the training set, as depicted in Fig. 2a. Our method dynamically estimates training set noise rates to mitigate overfitting and underfitting, seamlessly integrates with existing algorithms, and serves as a robust foundation for future noisy-label learning research. To summarise, our main contributions include:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 488, + 479, + 570 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 137, + 488, + 479, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 488, + 479, + 522 + ], + "spans": [ + { + "bbox": [ + 137, + 488, + 479, + 522 + ], + "type": "text", + "content": "- An innovative noisy-label learning graphical model, shown in Fig. 1b, that not only estimates but also leverages the noise rate from the training dataset to produce a refined sample selection curriculum for SOTA LNL methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 525, + 479, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 525, + 479, + 570 + ], + "spans": [ + { + "bbox": [ + 137, + 525, + 479, + 570 + ], + "type": "text", + "content": "- A simple and synergistic integration strategy of our novel graphical model with several SOTA noisy-label learning algorithms, such as DivideMix [26] and SSR [11], to improve their sample selection effectiveness, leading to an increased test accuracy, as shown in Fig. 2b." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": "Empirical evaluations show the role of our proposed sample selection methodology in improving the efficacy of leading noisy-label learning algorithms across various synthetic and real-world IDN benchmarks. Although our primary focus lies in addressing IDN problems, because of its more challenging and realistic nature, we have also applied our model to IIN problems, with detailed results in Tab. 6 and discussion in Appendix 3 of the supplementary material. With our innovative approach, we aim to improve the accuracy of SOTA LNL methods." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "A. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 533 + ], + "type": "text", + "content": "It is well-known that DNNs suffer from overfitting when trained with noisy-labels [52], resulting in poor generalisation [28, 55]. To mitigate the problems related to noisy-label learning, several techniques have been developed, including noise-robust loss functions [30], noise-label sample selection [21, 26, 40, 58, 59], meta-learning [57], and re-labeling followed by sample selection [11]. These techniques can handle many types of label noise (e.g., symmetric and asymmetric), but the most challenging type, i.e., IDN [39], tends to be successfully addressed with methods that have a training stage based on sample selection techniques [21, 26, 40, 58, 59]. These sample selection techniques separate clean and noisy-label samples, where clean samples are treated as labelled, and noisy samples are discarded [8, 16, 18, 49] or treated as unlabeled samples [9, 15, 26, 35, 59] for semi-supervised learning [3]. Another solution consists of aligning clean and noisy samples with an information fusion method [19]. A major limitation of these approaches is the overly simplistic design of the curriculum to select clean and noisy-label samples during training. It is either predetermined [16, 18] or learned from a set of predetermined basis functions [49]. We argue that the design of a curriculum based on an estimated noise rate would benefit such methods, as motivated by Fig. 1a, but to the best of our knowledge, such estimation has not been explored by previous methods. The closest idea explored is based on the estimation of the type of noise instead of the noise rate [45]. It is possible to argue that the estimation of the label transition matrices [5, 7, 44, 60], composed of the learned noise rate that affects the transition between pairs of labels, is a way to estimate the label noise rate. However, they portray results of comparatively lower accuracy for large real-world datasets or for IDN problems [39, 44]. A possible reason for these poorer results is that label transition approaches suffer from identifiability issues [14], where any clean-label distribution assignment is acceptable as long as the distribution of observed labels can be reconstructed [13]. This makes the identification of the actual underlying clean-labels challenging. One solution is to consider the use of multiple annotations per sample to help analyse agreements and disagreements for improved identification of clean patterns [14, 31]. However, the annotation of datasets with a single label per training sample is already challenging; the complexity is significantly larger to acquire multiple labels per sample." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 534, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 483, + 666 + ], + "type": "text", + "content": "An alternative technique to handle IDN problems is based on graphical models that represent the relationship between various observed and latent variables [15,45,50]. The approaches in [15,50] use graphical models that rely on a generative approach to produce noisy-labels from the respective image features and latent clean-labels. However, previous graphical models do not take into account the underlying noise rate parameter during modeling. Our work is the first graphical model approach to estimate the noise rate of the dataset. In addition, a close examination of existing literature reveals a trend: when traditional graphical models are integrated with SOTA noisy-label learning methods, the outcomes often fall short of expectations in terms of accuracy [2,15,50]. This observation underscores a strong point of our work. Our approach is uniquely tailored for" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "type": "text", + "content": "Noisy-Label Learning with Noise-Rate Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "content": "easy integration with current SOTA noisy-label learning methods for sample selection, which, in turn, improves over the SOTA classification accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 158, + 202, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 158, + 202, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 158, + 202, + 171 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 483, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 483, + 280 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 483, + 280 + ], + "type": "text", + "content": "In this section, we present our new graphical model that estimates the noise rate, which will be used in the sample selection process. Let " + }, + { + "bbox": [ + 130, + 183, + 483, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x_i,\\hat{y}_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 130, + 183, + 483, + 280 + ], + "type": "text", + "content": " be the noisy-label training set containing " + }, + { + "bbox": [ + 130, + 183, + 483, + 280 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 183, + 483, + 280 + ], + "type": "text", + "content": "-dimensional data vector " + }, + { + "bbox": [ + 130, + 183, + 483, + 280 + ], + "type": "inline_equation", + "content": "x_{i}\\in \\mathcal{X}\\subseteq \\mathbb{R}^{d}" + }, + { + "bbox": [ + 130, + 183, + 483, + 280 + ], + "type": "text", + "content": " and its respective " + }, + { + "bbox": [ + 130, + 183, + 483, + 280 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 183, + 483, + 280 + ], + "type": "text", + "content": "-dimensional one-hot encoded observed (potentially corrupted) label " + }, + { + "bbox": [ + 130, + 183, + 483, + 280 + ], + "type": "inline_equation", + "content": "\\hat{y}_i\\in \\hat{\\mathcal{V}} = \\{\\hat{y}:\\hat{y}\\in \\{0,1\\} ^C\\wedge \\mathbf{1}_C^\\top \\hat{y} = 1\\}" + }, + { + "bbox": [ + 130, + 183, + 483, + 280 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 183, + 483, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_C" + }, + { + "bbox": [ + 130, + 183, + 483, + 280 + ], + "type": "text", + "content": " is a vector of ones with " + }, + { + "bbox": [ + 130, + 183, + 483, + 280 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 183, + 483, + 280 + ], + "type": "text", + "content": " dimensions. The aim is to estimate the label noise rate " + }, + { + "bbox": [ + 130, + 183, + 483, + 280 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 183, + 483, + 280 + ], + "type": "text", + "content": ", used for the generation of noisy-label training data from the observed training dataset " + }, + { + "bbox": [ + 130, + 183, + 483, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 183, + 483, + 280 + ], + "type": "text", + "content": " and integrate this label noise rate into the sample selection strategy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 297, + 247, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 247, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 247, + 308 + ], + "type": "text", + "content": "3.1 Graphical Model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 316, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 376 + ], + "type": "text", + "content": "We portray the generation of noisy-label via the probabilistic graphical model shown in Fig. 1b. The observed random variables, denoted by shaded circles, are data " + }, + { + "bbox": [ + 130, + 316, + 482, + 376 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 316, + 482, + 376 + ], + "type": "text", + "content": " and the corresponding noisy-label " + }, + { + "bbox": [ + 130, + 316, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 130, + 316, + 482, + 376 + ], + "type": "text", + "content": ". We also have one latent variable, namely: the clean-label " + }, + { + "bbox": [ + 130, + 316, + 482, + 376 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 316, + 482, + 376 + ], + "type": "text", + "content": ". Under our proposed modeling assumption, a noisy-label of a data instance can be generated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 385, + 386, + 409 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 385, + 352, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 385, + 352, + 397 + ], + "spans": [ + { + "bbox": [ + 138, + 385, + 352, + 397 + ], + "type": "text", + "content": "- sample an instance from " + }, + { + "bbox": [ + 138, + 385, + 352, + 397 + ], + "type": "inline_equation", + "content": "p(X)" + }, + { + "bbox": [ + 138, + 385, + 352, + 397 + ], + "type": "text", + "content": ", i.e.: " + }, + { + "bbox": [ + 138, + 385, + 352, + 397 + ], + "type": "inline_equation", + "content": "x \\sim p(X)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 397, + 386, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 397, + 386, + 409 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 386, + 409 + ], + "type": "text", + "content": "- sample a clean-label from the clean-label distribution:" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 149, + 410, + 266, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 410, + 266, + 422 + ], + "spans": [ + { + "bbox": [ + 149, + 410, + 266, + 422 + ], + "type": "interline_equation", + "content": "y \\sim \\operatorname {C a t e g o r i c a l} (Y; f _ {\\theta_ {y}} (x))", + "image_path": "4f5a96abde3605296013b03504c1a8a3236405b14d8f89fce146cdc0c660862e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 422, + 386, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 386, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 386, + 433 + ], + "type": "text", + "content": "- sample a noisy-label from the noisy-label distribution:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 149, + 433, + 375, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 433, + 375, + 447 + ], + "spans": [ + { + "bbox": [ + 149, + 433, + 375, + 447 + ], + "type": "interline_equation", + "content": "\\hat {y} \\sim \\operatorname {C a t e g o r i c a l} (\\hat {Y}; \\epsilon \\times f _ {\\theta_ {\\hat {y}}} (x, f _ {\\theta_ {y}} (x)) + (1 - \\epsilon) \\times y),", + "image_path": "2da6e089edf988c7a3c1a041570bd3218feeaf63881cc950d15b374323f2d032.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\text{Categorical}()" + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "text", + "content": " denotes a categorical distribution, " + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "inline_equation", + "content": "f_{\\theta_y}: \\mathcal{X} \\to \\Delta_{C-1}" + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "inline_equation", + "content": "f_{\\theta_{\\hat{y}}}: \\mathcal{X} \\times \\Delta_{C-1} \\to \\Delta_{C-1}" + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "text", + "content": " denote two classifiers for the clean-label " + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "text", + "content": " and noisy-label " + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "text", + "content": ", respectively, with " + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\Delta_{C-1} = \\{s : s \\in [0,1]^C \\wedge \\mathbf{1}_C^\\top s = 1\\}" + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "text", + "content": " being the " + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "inline_equation", + "content": "(C-1)" + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "text", + "content": "-dimensional probability simplex. According to the data generation process shown in Fig. 1b, " + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "text", + "content": " corresponds to " + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{(x,\\hat{y}) \\sim p(X,\\hat{Y})}[P(\\hat{y} \\neq y|x)]" + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "text", + "content": ", which is the label noise rate of the training dataset of interest. Our aim is to infer the parameters " + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\theta_y, \\theta_{\\hat{y}}" + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "text", + "content": " from a noisy-label dataset " + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 454, + 482, + 552 + ], + "type": "text", + "content": " by maximising the following log-likelihood:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 561, + 482, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 561, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 133, + 561, + 482, + 593 + ], + "type": "interline_equation", + "content": "\\max _ {\\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon} \\mathbb {E} _ {(x _ {i}, \\hat {y} _ {i}) \\sim \\mathcal {D}} [ \\ln p (\\hat {y} _ {i} | x _ {i}; \\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon) ] = \\max _ {\\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon} \\mathbb {E} _ {(x _ {i}, \\hat {y} _ {i}) \\sim \\mathcal {D}} [ \\ln \\sum_ {y _ {i}} p (\\hat {y} _ {i}, y _ {i} | x _ {i}; \\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon) ]. \\tag {1}", + "image_path": "feb5c2b1fd14da5433f2bb15c7ded9a6820ab0a4cec70de9b230d1385ac98a32.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Due to the presence of the clean-label " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", it is difficult to directly evaluate the log-likelihood in Eq. (1). Therefore, we employ the expectation-maximisation (EM) algorithm [10] to maximise the log-likelihood. The main idea of the EM algorithm is to (i) construct a tight lower bound of the likelihood in Eq. (1) by estimating the posterior of the latent variable " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " (known as expectation step) and (ii) maximise that lower bound (known as maximisation step). Formally, let" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "A. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "q(y_{i}|x,\\hat{y};\\rho)" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " be an arbitrary distribution over a clean-label " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ". The evidence lower bound (ELBO) of the log-likelihood in Eq. (1) can be obtained through Jensen's inequality and is presented as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 172, + 161, + 481, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 161, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 172, + 161, + 481, + 220 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} Q \\left(\\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon , \\rho\\right) = \\mathbb {E} _ {\\left(x _ {i}, \\hat {y} _ {i}\\right) \\sim \\mathcal {D}} \\left[ \\mathbb {E} _ {q \\left(y _ {i} \\mid x _ {i}; \\hat {y} _ {i}; \\rho\\right)} \\left[ \\ln p \\left(\\hat {y} _ {i} \\mid x _ {i}, y _ {i}; \\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon\\right) \\right] \\right. \\\\ - \\operatorname {K L} \\left[ q \\left(y _ {i} \\mid x _ {i}, \\hat {y} _ {i}; \\rho\\right) \\| p \\left(y _ {i} \\mid x _ {i}; \\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon\\right) \\right] \\\\ = \\mathbb {E} _ {\\left(x _ {i}, \\hat {y} _ {i}\\right) \\sim \\mathcal {D}} \\left[ \\mathbb {E} _ {q \\left(y _ {i} \\mid x _ {i}; \\hat {y} _ {i}; \\rho\\right)} \\left[ \\ln p \\left(\\hat {y} _ {i} \\mid x _ {i}, y _ {i}; \\theta_ {\\hat {y}}, \\epsilon\\right) \\right. \\right. \\\\ \\left. + \\ln p \\left(y _ {i} \\mid x _ {i}; \\theta_ {y}\\right) \\right] + \\mathbb {H} \\left[ q \\left(y _ {i} \\mid x _ {i}, \\hat {y} _ {i}; \\rho\\right) \\right], \\tag {2} \\\\ \\end{array}", + "image_path": "a279a575576dd9717d25cf114ed3675671535d4d93276d2211818bec4291b5a8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 228, + 479, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 479, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 479, + 264 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 228, + 479, + 264 + ], + "type": "inline_equation", + "content": "\\mathrm{KL}[q\\| p]" + }, + { + "bbox": [ + 130, + 228, + 479, + 264 + ], + "type": "text", + "content": " is the Kullback - Leibler divergence between distributions " + }, + { + "bbox": [ + 130, + 228, + 479, + 264 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 228, + 479, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 228, + 479, + 264 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 228, + 479, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 228, + 479, + 264 + ], + "type": "inline_equation", + "content": "\\mathbb{H}(q)" + }, + { + "bbox": [ + 130, + 228, + 479, + 264 + ], + "type": "text", + "content": " is the entropy of the distribution " + }, + { + "bbox": [ + 130, + 228, + 479, + 264 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 228, + 479, + 264 + ], + "type": "text", + "content": ". The EM algorithm is then carried out iteratively by alternating the following two steps:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 275, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 480, + 361 + ], + "type": "text", + "content": "E step: we maximise the ELBO in (2) w.r.t. " + }, + { + "bbox": [ + 130, + 275, + 480, + 361 + ], + "type": "inline_equation", + "content": "q(y_{i}|x_{i},\\hat{y}_{i};\\rho)" + }, + { + "bbox": [ + 130, + 275, + 480, + 361 + ], + "type": "text", + "content": ". Theoretically, such an optimisation results in KL " + }, + { + "bbox": [ + 130, + 275, + 480, + 361 + ], + "type": "inline_equation", + "content": "[q(y_{i}|x_{i},\\hat{y}_{i};\\rho)\\| p(y_{i}|x_{i},\\hat{y}_{i})] = 0" + }, + { + "bbox": [ + 130, + 275, + 480, + 361 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 275, + 480, + 361 + ], + "type": "inline_equation", + "content": "q(y_{i}|x_{i},\\hat{y}_{i};\\rho) = p(y_{i}|x_{i},\\hat{y}_{i})" + }, + { + "bbox": [ + 130, + 275, + 480, + 361 + ], + "type": "text", + "content": ". This is equivalent to estimating the posterior of the clean-label " + }, + { + "bbox": [ + 130, + 275, + 480, + 361 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 130, + 275, + 480, + 361 + ], + "type": "text", + "content": " given noisy-label data " + }, + { + "bbox": [ + 130, + 275, + 480, + 361 + ], + "type": "inline_equation", + "content": "(x_{i},\\hat{y}_{i})" + }, + { + "bbox": [ + 130, + 275, + 480, + 361 + ], + "type": "text", + "content": ". Obtaining the posterior " + }, + { + "bbox": [ + 130, + 275, + 480, + 361 + ], + "type": "inline_equation", + "content": "p(y_{i}|x_{i},\\hat{y}_{i})" + }, + { + "bbox": [ + 130, + 275, + 480, + 361 + ], + "type": "text", + "content": " is, however, intractable for most deep-learning models. To mitigate such an issue, we follow the variational EM approach [33] by employing an approximate posterior " + }, + { + "bbox": [ + 130, + 275, + 480, + 361 + ], + "type": "inline_equation", + "content": "q(y_{i}|x_{i},\\hat{y}_{i};\\rho^{(t)})" + }, + { + "bbox": [ + 130, + 275, + 480, + 361 + ], + "type": "text", + "content": " that is the closest to the true posterior " + }, + { + "bbox": [ + 130, + 275, + 480, + 361 + ], + "type": "inline_equation", + "content": "p(y_{i}|x_{i},\\hat{y}_{i})" + }, + { + "bbox": [ + 130, + 275, + 480, + 361 + ], + "type": "text", + "content": ", where:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 231, + 369, + 481, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 369, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 231, + 369, + 481, + 389 + ], + "type": "interline_equation", + "content": "\\rho^ {(t)} = \\arg \\max _ {\\rho} Q \\left(\\theta_ {y} ^ {(t)}, \\theta_ {\\hat {y}} ^ {(t)}, \\epsilon^ {(t)}, \\rho\\right), \\tag {3}", + "image_path": "82ad30224a99e80aff624d784cf1a81298a7e0e348888408463fd9593474c618.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 399, + 479, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 479, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 479, + 436 + ], + "type": "text", + "content": "with the superscript " + }, + { + "bbox": [ + 130, + 399, + 479, + 436 + ], + "type": "inline_equation", + "content": "{}^{(t)}" + }, + { + "bbox": [ + 130, + 399, + 479, + 436 + ], + "type": "text", + "content": " denoting the parameters at the " + }, + { + "bbox": [ + 130, + 399, + 479, + 436 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 399, + 479, + 436 + ], + "type": "text", + "content": "-th iteration. Although this results in a non-tight lower bound of the log-likelihood in Eq. (1), it does increase the variational bound " + }, + { + "bbox": [ + 130, + 399, + 479, + 436 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 399, + 479, + 436 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 446, + 479, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 479, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 479, + 471 + ], + "type": "text", + "content": "M step: we maximise the ELBO in Eq. (2) w.r.t. " + }, + { + "bbox": [ + 130, + 446, + 479, + 471 + ], + "type": "inline_equation", + "content": "\\theta_y, \\theta_{\\hat{y}}" + }, + { + "bbox": [ + 130, + 446, + 479, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 446, + 479, + 471 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 446, + 479, + 471 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 130, + 446, + 479, + 471 + ], + "type": "inline_equation", + "content": "\\rho^{(t)}" + }, + { + "bbox": [ + 130, + 446, + 479, + 471 + ], + "type": "text", + "content": " obtained in the E step:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 200, + 479, + 481, + 501 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 200, + 479, + 481, + 501 + ], + "type": "interline_equation", + "content": "\\theta_ {y} ^ {(t + 1)}, \\theta_ {\\hat {y}} ^ {(t + 1)}, \\epsilon^ {(t + 1)} = \\arg \\max _ {\\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon} Q (\\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon , \\rho^ {(t)}). \\tag {4}", + "image_path": "33bd41aa4a684e853d5654ae7ed042c614d86c38cf603d0679a8f58a663b7fec.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 509, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 480, + 581 + ], + "type": "text", + "content": "The estimated noise rate " + }, + { + "bbox": [ + 130, + 509, + 480, + 581 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 509, + 480, + 581 + ], + "type": "text", + "content": " can then be integrated into certain noisy-label algorithms to train the models of interest as mentioned in Sec. 1. In addition, the inference of noise rate " + }, + { + "bbox": [ + 130, + 509, + 480, + 581 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 509, + 480, + 581 + ], + "type": "text", + "content": " might be associated with the identifiability issue when estimating the clean-label " + }, + { + "bbox": [ + 130, + 509, + 480, + 581 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 509, + 480, + 581 + ], + "type": "text", + "content": " [29], i.e., there exists multiple sets of " + }, + { + "bbox": [ + 130, + 509, + 480, + 581 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 509, + 480, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 509, + 480, + 581 + ], + "type": "inline_equation", + "content": "\\theta_y" + }, + { + "bbox": [ + 130, + 509, + 480, + 581 + ], + "type": "text", + "content": ", where each set can explain the observed noisy-label data equally well. Such issues are addressed in the following subsection." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 597, + 246, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 246, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 246, + 609 + ], + "type": "text", + "content": "3.2 Sample Selection" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 479, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 667 + ], + "type": "text", + "content": "The identifiability issue when inferring the clean-label " + }, + { + "bbox": [ + 130, + 617, + 479, + 667 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 617, + 479, + 667 + ], + "type": "text", + "content": " from noisy-label data " + }, + { + "bbox": [ + 130, + 617, + 479, + 667 + ], + "type": "inline_equation", + "content": "(X, \\hat{Y})" + }, + { + "bbox": [ + 130, + 617, + 479, + 667 + ], + "type": "text", + "content": " can be mitigated either by acquiring multiple noisy-labels [29] or introducing additional constraints, such as small loss hypothesis [16] or FINE [21]. Since requesting additional noisy-labels per training sample is not always possible, we" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "type": "text", + "content": "Noisy-Label Learning with Noise-Rate Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 140, + 484, + 381 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 481, + 139 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 481, + 139 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 481, + 139 + ], + "type": "text", + "content": "Algorithm 1 Proposed noisy-label learning that relies on the estimation of noise rate " + }, + { + "bbox": [ + 132, + 115, + 481, + 139 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 132, + 115, + 481, + 139 + ], + "type": "text", + "content": " to build a sample selection curriculum." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "lines": [ + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "text", + "content": "1: procedure NOISE RATE ESTIMATION AND INTEGRATION(D,T,λ) \n2: " + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x_i,\\hat{y}_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "text", + "content": " training set with noisy-label data \n3: " + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "text", + "content": " T: number of epochs \n4: " + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "text", + "content": " λ: a hyper-parameter \n5: " + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "text", + "content": " Notation: " + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "inline_equation", + "content": "\\Theta^{(t)} = (\\theta_y^{(t)},\\theta_{\\hat{y}}^{(t)},\\epsilon^{(t)})" + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "text", + "content": " \n6: Initialise " + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "inline_equation", + "content": "\\theta_y^{(1)},\\theta_{\\hat{y}}^{(1)},\\epsilon^{(1)}" + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "inline_equation", + "content": "\\rho^{(0)}" + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "inline_equation", + "content": "\\theta_y^{(1)}\\gets \\mathrm{WARM~UP}(\\mathcal{D},\\theta_y^{(1)})" + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "inline_equation", + "content": "t\\gets 0" + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "text", + "content": " \n9: for " + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{epoch}} = 1:T" + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "text", + "content": " do \n10: for each mini-batch " + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "text", + "content": " in SHUFFLE(D) do \n11: " + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "inline_equation", + "content": "t\\gets t + 1" + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "text", + "content": " \n12: " + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_{\\mathrm{clean}},\\mathcal{S}_{\\mathrm{noisy}}\\gets \\mathrm{SAMPLE~SELECTION}(\\mathcal{S},\\theta_y^{(t)},\\epsilon^{(t)})" + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "text", + "content": " \n13: " + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "text", + "content": " Variational E-step as in Eq. (3) \n14: " + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "inline_equation", + "content": "\\rho^{(t)}\\gets \\mathrm{E~STEP}(\\mathcal{S},\\theta_y^{(t)},\\theta_{\\hat{y}}^{(t)},\\epsilon^{(t)},\\rho^{(t - 1)})" + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "text", + "content": " \n15: " + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "text", + "content": " M-step as in Eq. (8) \n16: " + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "inline_equation", + "content": "\\theta_y^{(t + 1)},\\theta_{\\hat{y}}^{(t + 1)},\\epsilon^{(t + 1)}\\gets \\mathrm{M~STEP}(\\mathcal{S}_{\\mathrm{clean}},\\mathcal{S}_{\\mathrm{noisy}},\\theta_y^{(t)},\\theta_{\\hat{y}}^{(t)},\\epsilon^{(t)},\\rho^{(t)},\\lambda)" + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "text", + "content": " \n17: return " + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "inline_equation", + "content": "\\theta_y" + }, + { + "bbox": [ + 132, + 140, + 484, + 381 + ], + "type": "text", + "content": " \n1 parameter of the clean-label classifier" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 132, + 405, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 405, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 132, + 405, + 482, + 452 + ], + "type": "text", + "content": "follow the latter approach by imposing a constraint (scoring term), denoted as " + }, + { + "bbox": [ + 132, + 405, + 482, + 452 + ], + "type": "inline_equation", + "content": "L(\\theta_y, \\epsilon^{(t)})" + }, + { + "bbox": [ + 132, + 405, + 482, + 452 + ], + "type": "text", + "content": ", over " + }, + { + "bbox": [ + 132, + 405, + 482, + 452 + ], + "type": "inline_equation", + "content": "\\theta_y" + }, + { + "bbox": [ + 132, + 405, + 482, + 452 + ], + "type": "text", + "content": " in the M step via a sample selection approach based on the estimated noise rate " + }, + { + "bbox": [ + 132, + 405, + 482, + 452 + ], + "type": "inline_equation", + "content": "\\epsilon^{(t)}" + }, + { + "bbox": [ + 132, + 405, + 482, + 452 + ], + "type": "text", + "content": ". Formally, we propose a new curriculum when selecting samples as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 272, + 452, + 481, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 452, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 272, + 452, + 481, + 466 + ], + "type": "interline_equation", + "content": "R (t) = 1 - \\epsilon^ {(t)}. \\tag {5}", + "image_path": "da740fa7302e97725b22b61d6c63bb5c795b45a5a655104203bc00ce1fae20a5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 470, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 481, + 495 + ], + "type": "text", + "content": "In the simplest case, such as Co-teaching [16] or FINE [21], the constraint for " + }, + { + "bbox": [ + 130, + 470, + 481, + 495 + ], + "type": "inline_equation", + "content": "\\theta_y" + }, + { + "bbox": [ + 130, + 470, + 481, + 495 + ], + "type": "text", + "content": " can be written as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 501, + 482, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 501, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 137, + 501, + 482, + 529 + ], + "type": "interline_equation", + "content": "L \\left(\\theta_ {y}, \\epsilon^ {(t)}\\right) = \\sum_ {\\left(x _ {i}, \\hat {y} _ {i}\\right) \\in \\mathcal {S} _ {\\text {c l e a n}}} \\operatorname {K L} \\left[ \\operatorname {C a t e g o r i c a l} (Y; \\hat {y}) \\| \\operatorname {C a t e g o r i c a l} \\left(Y; f _ {\\theta_ {y}} \\left(x _ {i}\\right)\\right) \\right], \\tag {6}", + "image_path": "963ac6ac447e4fad5cfda1fc9d9463cf465f7a9daa49d0e60321768ea3306586.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 536, + 163, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 163, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 163, + 546 + ], + "type": "text", + "content": "where:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 157, + 554, + 373, + 568 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 554, + 373, + 568 + ], + "spans": [ + { + "bbox": [ + 157, + 554, + 373, + 568 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {\\text {c l e a n}} = \\left\\{\\left(x _ {i}, \\hat {y} _ {i}\\right): \\left(x _ {i}, \\hat {y} _ {i}\\right) \\in \\mathcal {D} \\wedge z \\left(x _ {i}, y _ {i}\\right) \\leq m \\right\\},", + "image_path": "cc39a3cbb1f8b65f3859b6167d663598efd1bc6d1bd1c52214a13f5419bb157b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 157, + 569, + 454, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 569, + 454, + 582 + ], + "spans": [ + { + "bbox": [ + 157, + 569, + 454, + 582 + ], + "type": "interline_equation", + "content": "m \\in \\{m: \\Pr (z \\left(x _ {i}, y _ {i}\\right) \\leq m) \\geq R (t) \\wedge \\Pr (z \\left(x _ {i}, y _ {i}\\right) \\geq m) \\geq 1 - R (t) \\},", + "image_path": "11b186dd1ce8250f2ab212b2dac441a44d57d1f3a159135f8aae16c0afe4c85d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 157, + 584, + 481, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 584, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 157, + 584, + 481, + 597 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {\\text {n o i s y}} = \\mathcal {D} \\backslash \\mathcal {S} _ {\\text {c l e a n}}. \\tag {7}", + "image_path": "eea7db1a74839658b8a7ffd77b425a19a078d146e7fe146fd5ded518076ff2c0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "R(t)" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " defined in (5), and " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "z(x,y)" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " representing the score of a criterion (e.g., loss [26, 58, 59], distance to the largest eigenvectors [21], or KNN scores [11]). Intuitively, the loss in Eq. (6) is simply the cross-entropy loss on the " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\lfloor R(t)\\times N\\rfloor" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " samples that have smallest scores (with " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\lfloor .\\rfloor" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " being the floor function). One can also extend to other SOTA models by replacing the loss " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " accordingly. For example," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "A. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 473, + 251 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 473, + 251 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 473, + 251 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 473, + 251 + ], + "type": "image", + "image_path": "08c3a57bb884bedf5284bacc78cd61a6a3ead553af3eb4a53130280ac377ba00.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 258, + 482, + 314 + ], + "lines": [ + { + "bbox": [ + 131, + 258, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 131, + 258, + 482, + 314 + ], + "type": "text", + "content": "Fig. 3: Our training algorithm uses an existing noisy-label classifier (e.g., DivideMix [26]) parameterised by " + }, + { + "bbox": [ + 131, + 258, + 482, + 314 + ], + "type": "inline_equation", + "content": "\\theta_y" + }, + { + "bbox": [ + 131, + 258, + 482, + 314 + ], + "type": "text", + "content": " as the clean-label model " + }, + { + "bbox": [ + 131, + 258, + 482, + 314 + ], + "type": "inline_equation", + "content": "p(Y|X;\\theta_y)" + }, + { + "bbox": [ + 131, + 258, + 482, + 314 + ], + "type": "text", + "content": ". The generation of noisy-label (given " + }, + { + "bbox": [ + 131, + 258, + 482, + 314 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 131, + 258, + 482, + 314 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 258, + 482, + 314 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 131, + 258, + 482, + 314 + ], + "type": "text", + "content": ") is carried out by a model parameterised by noise rate " + }, + { + "bbox": [ + 131, + 258, + 482, + 314 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 131, + 258, + 482, + 314 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 258, + 482, + 314 + ], + "type": "inline_equation", + "content": "\\theta_{\\tilde{y}}" + }, + { + "bbox": [ + 131, + 258, + 482, + 314 + ], + "type": "text", + "content": ". The noisy-label classifier is based on a sample-selection mechanism that uses a curriculum " + }, + { + "bbox": [ + 131, + 258, + 482, + 314 + ], + "type": "inline_equation", + "content": "R(t) = 1 - \\epsilon^{(t)}" + }, + { + "bbox": [ + 131, + 258, + 482, + 314 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 339, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 339, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 131, + 339, + 482, + 365 + ], + "type": "text", + "content": "if DivideMix [26] is used as a base model to constrain " + }, + { + "bbox": [ + 131, + 339, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\theta_y" + }, + { + "bbox": [ + 131, + 339, + 482, + 365 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 339, + 482, + 365 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 131, + 339, + 482, + 365 + ], + "type": "text", + "content": " will include two additional terms: loss on un-labeled data and regularisation using mixup [54]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 386, + 266, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 266, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 266, + 399 + ], + "type": "text", + "content": "3.3 Training and Testing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 411, + 481, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 411, + 481, + 436 + ], + "spans": [ + { + "bbox": [ + 131, + 411, + 481, + 436 + ], + "type": "text", + "content": "Given the sample selection approach in Sec. 3.2, the M step in Eq. (4) is slightly modified by integrating the constraint in Eq. (6), which can be written as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 162, + 446, + 482, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 446, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 162, + 446, + 482, + 468 + ], + "type": "interline_equation", + "content": "\\theta_ {y} ^ {(t + 1)}, \\theta_ {\\hat {y}} ^ {(t + 1)}, \\epsilon^ {(t + 1)} = \\arg \\max _ {\\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon} Q (\\theta_ {y}, \\theta_ {\\hat {y}}, \\epsilon , \\rho^ {(t)}) - \\lambda L (\\theta_ {y}, \\epsilon^ {(t)}), \\tag {8}", + "image_path": "555e92563122365328e337c716358a985e5b1d0c52895b5414ae27ed10edc27f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 480, + 482, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 480, + 482, + 539 + ], + "spans": [ + { + "bbox": [ + 131, + 480, + 482, + 539 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 480, + 482, + 539 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 131, + 480, + 482, + 539 + ], + "type": "text", + "content": " is a hyper-parameter and " + }, + { + "bbox": [ + 131, + 480, + 482, + 539 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 131, + 480, + 482, + 539 + ], + "type": "text", + "content": " is defined in Eq. (6). The training procedure is summarised in Algorithm 1 and visualised in Fig. 3. In the implementation, we integrate the proposed method into existing models, such as DivideMix [26] or FINE [21]. Note that the clean-label classifier " + }, + { + "bbox": [ + 131, + 480, + 482, + 539 + ], + "type": "inline_equation", + "content": "f_{\\theta_y}(.)" + }, + { + "bbox": [ + 131, + 480, + 482, + 539 + ], + "type": "text", + "content": " is also the clean classifier of the base model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 563, + 230, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 230, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 230, + 577 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 482, + 665 + ], + "type": "text", + "content": "We show extensive experiments in several noisy-label synthetic benchmarks with CIFAR100 [22], and real-world benchmarks, including CNWL's red mini-ImageNet [17], Clothing1M [45] and mini-WebVision [27]. Sec. 4.2 describes implementation details. We evaluate our approach by plugging SOTA models into " + }, + { + "bbox": [ + 131, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "p(y|x;\\theta_y)" + }, + { + "bbox": [ + 131, + 594, + 482, + 665 + ], + "type": "text", + "content": ", defined in Sec. 3, with results being shown in Sec. 4.3, and ablation studies in Sec. 4.4." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "type": "text", + "content": "Noisy-Label Learning with Noise-Rate Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 265, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 265, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 265, + 129 + ], + "type": "text", + "content": "4.1 Dataset Descriptions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "text", + "content": "We follow the existing literature to generate the IDN labels [44] for CIFAR100 [22]. The dataset contains 50,000 training images and 10,000 testing images of shape " + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "inline_equation", + "content": "32 \\times 32 \\times 3" + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "text", + "content": ". The dataset is class-balanced with 100 different categories and the IDN rates considered are 0.2, 0.3, 0.4 and 0.5 [44]. Another important IDN benchmark dataset is the red mini-ImageNet, a real-world dataset from CNWL [17]. In this dataset, there are 100 distinct classes, with each class containing 600 images. The images and their corresponding noisy-labels have been crawled from the internet at various controllable label noise rates. For a fair comparison with the existing literature [9, 15, 46], we have resized the images to " + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "text", + "content": " pixels from the " + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "inline_equation", + "content": "84 \\times 84" + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "text", + "content": " original pixel settings. We show results with " + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "text", + "content": " noise rates on this dataset. Clothing1M [45] is a real-world clothing classification dataset that contains 1 million images with an estimated noise rate of " + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "inline_equation", + "content": "38.5\\%" + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "text", + "content": ". The dataset contains 14 different categories, and the labels are generated from surrounding texts. Images in this dataset are of different sizes, and we follow the resize structure suggested in [21, 26]. This dataset also contains " + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "inline_equation", + "content": "50k" + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "text", + "content": " manually validated clean training images, " + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "inline_equation", + "content": "14k" + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "text", + "content": " images for the validation set, and " + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "inline_equation", + "content": "10k" + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "text", + "content": " testing images. We have not used the clean training, validation set, or any extra training images while training. Only the testing set is used for evaluation. Mini-WebVision [26] contains 65,944 images with their respective labels from the 50 different initial categories from the WebVision [27], with image size of " + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 136, + 483, + 401 + ], + "type": "text", + "content": " pixels. Following the evaluation process commonly used in this benchmark, 50 categories from the ILSVRC12 [23] dataset are also used for testing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 418, + 240, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 240, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 240, + 429 + ], + "type": "text", + "content": "4.2 Implementation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": "All methods are implemented in Pytorch [36] and use one NVIDIA RTX 3090 card for training and testing. As mentioned in the original papers, hyperparameter settings are kept the same for the baselines used in the proposed algorithm. All classifier architectures are also kept the same as the baseline models. A random initialisation of noise rate parameter " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": " with the sigmoid as its activation function is employed for all experiments to maintain the fairness of the comparisons with other approaches. The value of " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": " in Eq. (8) is set to 1 for all the cases. We integrate many SOTA approaches [11, 15, 21, 26, 58, 59] into our graphical model, as explained in Sec. 3.3. For CIFAR100 [22] with IDN [44], we integrate DivideMix [26], C2D [59], CC [58], and InstanceGM [15] into our model, given their superior performance across various noise rates. Additionally, we also use F-Dividemix from FINE [21], as shown in Fig. 2b. For red mini-ImageNet [17], we test our proposed approach with and without DINO self-supervision [4]. For the implementation without self-supervision, we use DivideMix [26] and InstanceGM [15], and for the self-supervised version, we only use InstanceGM [15]. The models for Clothing1M [45] and mini-WebVision [26] (validation on ImageNet [23]) are trained using DivideMix [26], SSR [11], C2D [59] and CC [58]. Additional implementation details, empirical analysis, and computational analysis are presented in the supplementary material in Appendices 1, 2 and 4, respectively." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "A. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 193, + 289, + 363 + ], + "blocks": [ + { + "bbox": [ + 138, + 193, + 289, + 363 + ], + "lines": [ + { + "bbox": [ + 138, + 193, + 289, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 193, + 289, + 363 + ], + "type": "table", + "html": "
MethodNoise Rates - IDN
0.20.30.40.5
CE [50]30.4224.1521.4514.42
PartT [44]65.3364.5659.7356.80
kMEIDTM [7]69.1666.7663.4659.18
DivideMix [26]77.0376.3370.8058.61
DivideMix-Ours77.4277.2172.4164.02
(+0.39)(+0.88)(+1.61)(+5.41)
C2D [59]78.6178.1872.8963.19
C2D-Ours79.0778.5973.3165.28
(+0.46)(+0.41)(+0.42)(+2.09)
CC [58]79.6177.5676.5863.19
CC-Ours79.7278.7177.3867.53
(+0.11)(+1.15)(+0.80)(+4.34)
InstanceGM [15]79.6979.2178.4777.19
InstanceGM-Ours79.6179.4079.5278.21
(-0.08)(+0.19)(+1.05)(+1.02)
", + "image_path": "4bfd87feb77d72e512e4ee274f0c31901970f3e0ba53f7122e3fe8f58099bdec.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 332, + 190, + 476, + 365 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 1: Test accuracy % on (left) CIFAR100 [22] under different IDN [45], and (right) red mini-ImageNet [17]. Other models' results are from [7, 15, 46]. Here, we integrate DivideMix [26], C2D [59], CC [58] and InstanceGM [15] into our proposed model for CIFAR100 [22]. For red mini-ImageNet [17], we present the results with and without self-supervision [4]. We integrate DivideMix [26] and InstanceGM [15] into our model, with the latter tested with and without self-supervision." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 190, + 476, + 365 + ], + "lines": [ + { + "bbox": [ + 332, + 190, + 476, + 365 + ], + "spans": [ + { + "bbox": [ + 332, + 190, + 476, + 365 + ], + "type": "table", + "html": "
MethodNoise rate
0.40.60.8
CE [46]42.7037.3029.76
MixUp [54]46.4040.5833.58
MentorMix [17]47.1443.8033.46
FaMUS [46]51.4245.1035.50
DivideMix [26]46.7243.1434.50
DivideMix-Ours50.7045.1137.44
(+3.98)(+1.97)(+2.94)
InstanceGM [15]52.2447.9639.62
InstanceGM-Ours56.6151.4043.83
(+4.37)(+3.44)(+4.21)
With self-supervised learning
PropMix [9]56.2252.8443.42
InstanceGM-SS [15]56.3753.2144.03
InstanceGM-SS-Ours58.2953.6045.47
(+1.92)(+0.39)(+1.44)
", + "image_path": "b2d58fde9424dd3cccf214c591423fbdd0f7123c4d8ff348a89bcb2192a7e5c3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 148, + 407, + 261, + 479 + ], + "blocks": [ + { + "bbox": [ + 130, + 374, + 482, + 398 + ], + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 398 + ], + "type": "text", + "content": "Table 2: Final estimated noise rate " + }, + { + "bbox": [ + 130, + 374, + 482, + 398 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 374, + 482, + 398 + ], + "type": "text", + "content": " for the Tab. 1, (left) CIFAR100 [22] under various IDN [44], and (right) red mini-ImageNet [17] with and without self-supervision." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 407, + 261, + 479 + ], + "lines": [ + { + "bbox": [ + 148, + 407, + 261, + 479 + ], + "spans": [ + { + "bbox": [ + 148, + 407, + 261, + 479 + ], + "type": "table", + "html": "
Estimated noise rates
MethodActual noise rate
0.20.30.40.5
DivideMix-Ours0.180.340.470.53
C2D-Ours0.190.330.380.52
CC-Ours0.210.340.410.54
InstanceGM-Ours0.230.370.420.47
", + "image_path": "284a65786cefd4bf552f156bc2186dde2912e47a7aaa070cc230f159a079983e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 342, + 411, + 465, + 475 + ], + "blocks": [ + { + "bbox": [ + 342, + 411, + 465, + 475 + ], + "lines": [ + { + "bbox": [ + 342, + 411, + 465, + 475 + ], + "spans": [ + { + "bbox": [ + 342, + 411, + 465, + 475 + ], + "type": "table", + "html": "
Estimated noise rates
MethodActual noise rate
0.40.6
DivideMix-Ours0.390.58
InstanceGM-Ours0.380.55
InstanceGM-SS-Ours0.480.53
0.73
", + "image_path": "d433bbea64944dfbb523ed9b62d1e29215db162b39b2c3d74b086baa83a95672.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 498, + 282, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 282, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 282, + 511 + ], + "type": "text", + "content": "4.3 Comparison with SOTA" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 517, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 554 + ], + "type": "text", + "content": "This section compares our approach with SOTA methods on datasets with the IDN settings and noisy real-world settings. The bold text in tables indicate SOTA results, and our results are in the greyed rows." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Synthetic Instance-Dependent Noise The comparison between various baselines and our proposed method on CIFAR100 [22] with IDN [44] is shown in Tab. 1 (left) with the noise rate ranging from 0.2 to 0.5. It is worth noting that using our proposed model with DivideMix [26], C2D [59], CC [58] and InstanceGM [15] improve their performance in almost all cases, particularly with large noise rates. Tab. 2 (left) also shows the final noise rate " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " estimated by our model which is reasonably close to the simulated noise rate displayed in the table's header." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "type": "text", + "content": "Noisy-Label Learning with Noise-Rate Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 194, + 179, + 422, + 291 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 3: Test accuracy (\\%) of competing methods, and final estimated noise rate " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " on Clothing1M [45]. Also, we have not considered competing models that rely on a clean or validation set whilst training. We integrate DivideMix [26], C2D [59], SSR [11] and CC [58] into our model. DivideMix [26] and CC [58] shows the locally reproduced results. Mentioned noise rate by [45] is 0.385." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 194, + 179, + 422, + 291 + ], + "lines": [ + { + "bbox": [ + 194, + 179, + 422, + 291 + ], + "spans": [ + { + "bbox": [ + 194, + 179, + 422, + 291 + ], + "type": "table", + "html": "
MethodTest accuracy (%)Estimated noise rate
OT-Filter [12]74.50
ELR+ with C2D [59]74.58
AugDesc [35]75.11
NCE [25]75.30
DivideMix [26]74.32
DivideMix-Ours74.41 (+0.09)0.41
C2D [59]74.58
C2D-Ours74.71 (+0.13)0.41
SSR (class-imbalance) [11]74.12
SSR-Ours74.20 (+0.08)0.42
CC [58]75.24
CC-Ours75.31 (+0.07)0.41
", + "image_path": "0332020760bcfb9fed8e6553408cbb4a0a7b58c2587196643a6ee2d8aa30f5c8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 313, + 483, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 483, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 483, + 541 + ], + "type": "text", + "content": "Real-World Noise We also evaluate our proposed method on various real-world noisy settings regarding test accuracy and estimated noise rates " + }, + { + "bbox": [ + 130, + 313, + 483, + 541 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 313, + 483, + 541 + ], + "type": "text", + "content": " in Tabs. 1 to 4. Similarly to the synthetic IDN in Tab. 1 (left), the results show that existing noisy-label robust methods can be easily integrated with our model to outperform current SOTA results for real-world noisy-label datasets. Tabs. 1 and 2 (right) shows the results on red mini-ImageNet using two configurations, including cases without self-supervision (top part of the table) and with self-supervision (bottom part of the table). The self-supervision DINO pre-training [4] relies only on images from red mini-ImageNet to enable a fair comparison with existing baselines [9, 15]. Results from Tab. 1 (right) demonstrate that our approach improves the performance of SOTA methods by a considerable margin in all cases. In fact, using estimated noise rate " + }, + { + "bbox": [ + 130, + 313, + 483, + 541 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 313, + 483, + 541 + ], + "type": "text", + "content": " (Tab. 2 (right)) while training InstanceGM [15] without self-supervision shows better performance than existing self-supervised baselines at 0.4 noise rate. Moreover, DivideMix [26], C2D [59], SSR [11], and CC [58] are used as a baselines for Clothing1M [45] as shown in Tab. 3, with results showing slight improvements with the use of our method. For mini-WebVision [26] (validation on ImageNet [23]), we use DivideMix [26], C2D [59], SSR [11], and CC [58] as baselines (Tab. 4). It is worth noting that our results are better than the SOTA for all settings in Tab. 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 560, + 205, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 560, + 205, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 560, + 205, + 572 + ], + "type": "text", + "content": "4.4 Ablation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "We show an ablation study with testing accuracy (top) and training time (bottom) of our approach in Tab. 5 (left) on CIFAR100 [22] at 0.3 and 0.5 IDN [44] using DivideMix [26] as baseline. Detailed complexity analysis is shown in the supplementary material, Tab. 10. Initially, the accuracy result of baseline DivideMix [26] is " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "76.33\\%" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "58.61\\%" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " for 0.3 and 0.5 IDN, respectively. In the second row, we fix the noise rate " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " at 0.3 and 0.5 for DivideMix's sample selection, as explained in Sec. 3.3 (without updating " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "), then the results improved to " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "78.14\\%" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "64.44\\%" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "A. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 205, + 168, + 411, + 308 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 4: Test accuracy (\\%) and final estimated noise rate " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": " on mini-WebVision [26] and validation on ImageNet [23]. We integrate DivideMix [26], Contrast-to-Divide (C2D) [59], SSR [11] and CC [58] into our model, whilst models suffixed with -Ours denotes our proposed approach. ImageNet [23] is only considered for validation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 205, + 168, + 411, + 308 + ], + "lines": [ + { + "bbox": [ + 205, + 168, + 411, + 308 + ], + "spans": [ + { + "bbox": [ + 205, + 168, + 411, + 308 + ], + "type": "table", + "html": "
Datasetmini-WebVisionImageNetEstimated noise rate
Top-1Top-5Top-1Top-5
BtR [38]80.8892.7675.9692.20
NCE [25]79.5093.8076.3094.10
DivideMix [26]77.3291.6475.2091.64
DivideMix-Ours78.5192.0376.1193.240.43
(+1.19)(+0.39)(+0.91)(+1.60)
C2D [59]79.4292.3278.5793.04
C2D-Ours80.2092.8279.1693.120.43
(+0.78)(+0.50)(+0.59)(+0.08)
SSR [11]80.9292.8075.7691.76
SSR-Ours81.6893.8076.9193.050.43
(+0.76)(+1.00)(+1.15)(+1.29)
CC [58]79.3693.6476.0893.86-
CC-Ours80.0193.7977.1194.210.44
(+0.65)(+0.15)(+1.03)(+0.35)
", + "image_path": "3964453370a30ed11490097d59e6764282009b9af482a1587051d72fc2d23f24.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 140, + 445, + 302, + 544 + ], + "blocks": [ + { + "bbox": [ + 130, + 322, + 482, + 422 + ], + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 422 + ], + "type": "text", + "content": "Table 5: (left) Ablation results on CIFAR100 at 0.3 and 0.5 IDN. We display accuracy results under different configurations and compare the training time of our approach against DivideMix [26]. The implementation settings and detailed computational analysis are described in Sec. 4.2 and Appendix 4 respectively. (right above) Ablation results for initialisation with different " + }, + { + "bbox": [ + 130, + 322, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 322, + 482, + 422 + ], + "type": "text", + "content": " values on CIFAR100 [22] at a synthetic noise rate of 0.5 IDN [44] with DivideMix-Ours. (right below) Ablation graph showing the learning pattern of the noise rate parameter estimation under different initialisation values, i.e., " + }, + { + "bbox": [ + 130, + 322, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.001" + }, + { + "bbox": [ + 130, + 322, + 482, + 422 + ], + "type": "text", + "content": " (blue), " + }, + { + "bbox": [ + 130, + 322, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.1" + }, + { + "bbox": [ + 130, + 322, + 482, + 422 + ], + "type": "text", + "content": " (orange), " + }, + { + "bbox": [ + 130, + 322, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.9" + }, + { + "bbox": [ + 130, + 322, + 482, + 422 + ], + "type": "text", + "content": " (red), when trained on DivideMix-ours case at 0.5 IDN [44] on CIFAR100 [22]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 445, + 302, + 544 + ], + "lines": [ + { + "bbox": [ + 140, + 445, + 302, + 544 + ], + "spans": [ + { + "bbox": [ + 140, + 445, + 302, + 544 + ], + "type": "table", + "html": "
ModelsAccuracy (%)
0.30.5
DivideMix [26]76.3358.61
DivideMix with fixed ε78.1464.44
Our method with pre-trained DivideMix75.0152.31
Our method with original DivideMix (no ε)76.2056.30
DivideMix-Ours77.2164.02
Training Time (GPU-hours)
DivideMix [26]7.20
DivideMix-Ours8.50
DivideMix-Ours (Mixed Precision)6.10
", + "image_path": "b6a8bf4fbd2ae27fd3fda69734d7d7a474bf226e810f821298f9c5518fc378f5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 331, + 431, + 439, + 552 + ], + "blocks": [ + { + "bbox": [ + 331, + 431, + 439, + 552 + ], + "lines": [ + { + "bbox": [ + 331, + 431, + 439, + 552 + ], + "spans": [ + { + "bbox": [ + 331, + 431, + 439, + 552 + ], + "type": "image", + "image_path": "aa983ece11a754aafafd52ae1d8727f56159fc784823abeb033ac358abf72881.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "which is the ideal case (i.e., a perfect noise rate estimation) that motivated our work. In the third case, we use the proposed graphical model with pre-trained DivideMix [26] that shows accuracy of " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "75.01\\%" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "52.31\\%" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ". In the next case, the proposed graphical model is trained together with DivideMix [26] without considering the estimated noise rate " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " for sample selection, which results in accuracy of " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "76.20\\%" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "56.30\\%" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ". In the last row, we show the training of the proposed model with DivideMix [26], together with the estimation of noise rate " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ", and the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "type": "text", + "content": "Noisy-Label Learning with Noise-Rate Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "text", + "content": "selection of samples based on that, with " + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\approx 1\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\approx 8\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "text", + "content": " accuracy improvement to 0.3 and 0.5 IDN, which is quite close to our ideal case (second row). We also study the effect that the initialisation of the noise rate parameter " + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "text", + "content": " plays in our method. As described in Sec. 4.2, we randomly initialise " + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "text", + "content": " within the range " + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "text", + "content": ". To test the robustness of our method to this initialisation, we perform an ablation study by initialising " + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "text", + "content": " with three different values, namely: 0.001, 0.1, and 0.9 on CIFAR100 [24] at 0.5 IDN [44] using the DivideMix-ours approach. Our results, presented in Tab. 5 (right above and below), demonstrate that the initial value of " + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "text", + "content": " has a limited effect on the model's performance, highlighting the stability of our method across different initialisation of " + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "text", + "content": ". In the supplementary material, Tab. 7 (left) presents comparative analyses of our model against PartT [44] and SSR [11] on CIFAR100 [22] at 0.5 IDN [44], while Tab. 7 (right) showcases similar comparisons with baselines DivideMix [26] and InstanceGM [15] at a higher noise rate of 0.6 IDN [44]. Additionally, Tab. 8 provides further insights into the performance of DivideMix [26] on red mini-ImageNet [17] at lower noise rates of 0.1, 0.2 and 0.3. Tab. 9 illustrates noise rate estimation with standard deviation for our model integrated with DivideMix [26] and InstanceGM [15] on CIFAR100 [22] under IDN [44] settings at 0.2, 0.3, 0.4, and 0.5. Further statistical analysis is detailed in Appendix 5 with confidence interval presented in Tab. 12. Performance analysis with various architectures (Tab. 13) and lambda values (Tab. 14) are mentioned in Appendix 6. Threshold-based comparisons are in Appendix 7 and Tab. 15." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 397, + 311, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 311, + 410 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 311, + 410 + ], + "type": "text", + "content": "5 Discussion and Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 422, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 506 + ], + "type": "text", + "content": "In this work, we proposed an innovative graphical model for IDN noisy-label learning, focusing on the estimation of the label noise rate that is leveraged to introduce a robust sample selection curriculum. Given that SOTA IDN noisy-label learning approaches tend to rely on sample selection methods, our method can be seamlessly integrated to them to improve their performance in many synthetic and real-world benchmarks, including CIFAR100 [22], red mini-ImageNet [17], Clothing1M [45], mini-WebVision [26], and ImageNet [23]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 507, + 482, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 651 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 651 + ], + "type": "text", + "content": "From a societal perspective, our methodology offers a positive impact by mitigating biases inherent in noisy data, thereby promoting more fair and accurate machine learning outcomes. It also opens the door to future research avenues, including exploring advanced noise rate estimation methodologies and investigating noise-robust loss functions such as GCE [56] and ELR [28] with it. Our choice of using the same network architecture for clean and noisy labels enables compatibility with co-teaching techniques [16,26]. However, investigating different network structures for clean and noisy labels remains a potential avenue for future research. Moreover, we aim to delve into its dynamics, uncovering substantial improvements for certain models while observing more marginal enhancements for others. For more detailed information, please refer to Appendix 8. Our approach also hints at potential breakthroughs in streamlining data annotation challenges." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "A. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 141, + 481, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 481, + 186 + ], + "type": "text", + "content": "1. Arpit, D., Jastrzebski, S., Ballas, N., Krueger, D., Bengio, E., Kanwal, M.S., Maharaj, T., Fischer, A., Courville, A., Bengio, Y., Lacoste-Julien, S.: A closer look at memorization in deep networks. In: International Conference on Machine Learning. vol. 70, pp. 233-242. PMLR (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 187, + 481, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 187, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 187, + 481, + 220 + ], + "type": "text", + "content": "2. Bae, H., Shin, S., Jang, J., Na, B., Song, K., Moon, I.C.: From noisy prediction to true label: Noisy prediction calibration via generative model. In: International Conference on Machine Learning (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 220, + 481, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 481, + 253 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 481, + 253 + ], + "type": "text", + "content": "3. Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.A.: Mixmatch: A holistic approach to semi-supervised learning. In: Advances in Neural Information Processing Systems. vol. 32 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 254, + 481, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 481, + 286 + ], + "type": "text", + "content": "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: International Conference on Computer Vision. pp. 9650-9660 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 287, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 287, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 287, + 481, + 319 + ], + "type": "text", + "content": "5. Chen, P., Ye, J., Chen, G., Zhao, J., Heng, P.A.: Beyond class-conditional assumption: A primary attempt to combat instance-dependent label noise. In: AAAI Conference on Artificial Intelligence. pp. 11442–11450 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 321, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 481, + 364 + ], + "type": "text", + "content": "6. Chen, Z., Song, A., Wang, Y., Huang, X., Kong, Y.: A noise rate estimation method for image classification with label noise. Journal of Physics: Conference Series 2433(1), 012039 (feb 2023). https://doi.org/10.1088/1742-6596/2433/1/012039, https://dx.doi.org/10.1088/1742-6596/2433/1/012039" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 365, + 481, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 481, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 481, + 408 + ], + "type": "text", + "content": "7. Cheng, D., Liu, T., Ning, Y., Wang, N., Han, B., Niu, G., Gao, X., Sugiyama, M.: Instance-dependent label-noise learning with manifold-regularized transition matrix estimation. In: Conference on Computer Vision and Pattern Recognition. pp. 16630-16639 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 409, + 481, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 409, + 481, + 442 + ], + "spans": [ + { + "bbox": [ + 138, + 409, + 481, + 442 + ], + "type": "text", + "content": "8. Cheng, H., Zhu, Z., Li, X., Gong, Y., Sun, X., Liu, Y.: Learning with instance-dependent label noise: A sample sieve approach. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 443, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 443, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 443, + 481, + 475 + ], + "type": "text", + "content": "9. Cordeiro, F.R., Belagiannis, V., Reid, I., Carneiro, G.: PropMix: Hard sample filtering and proportional mixup for learning with noisy labels. In: British Machine Vision Conference (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 476, + 481, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 476, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 138, + 476, + 481, + 509 + ], + "type": "text", + "content": "0. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the EM algorithm. Journal of the royal statistical society: series B (methodological) 39(1), 1-22 (1977)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 510, + 481, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 481, + 532 + ], + "type": "text", + "content": "1. Feng, C., Tzimiropoulos, G., Patras, I.: SSR: An efficient and robust framework for learning with unknown label noise. In: British Machine Vision Conference (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 533, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 481, + 564 + ], + "type": "text", + "content": "2. Feng, C., Ren, Y., Xie, X.: Ot-filter: An optimal transport filter for learning with noisy labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16164-16174 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 565, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 565, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 565, + 481, + 597 + ], + "type": "text", + "content": "3. Frénay, B., Kabán, A.: A comprehensive introduction to label noise. In: European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 481, + 632 + ], + "type": "text", + "content": "4. Frénay, B., Verleysen, M.: Classification in the presence of label noise: A survey. IEEE Transactions on Neural Networks and Learning Systems 25(5), 845-869 (2013)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "5. Garg, A., Nguyen, C., Felix, R., Do, T.T., Carneiro, G.: Instance-dependent noisy label learning via graphical modelling. In: Winter Conference on Applications of Computer Vision. pp. 2288-2298 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "type": "text", + "content": "Noisy-Label Learning with Noise-Rate Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "16. Han, B., Yao, Q., Yu, X., Niu, G., Xu, M., Hu, W., Tsang, I., Sugiyama, M.: Co-teaching: Robust training of deep neural networks with extremely noisy labels. In: Advances in Neural Information Processing Systems. vol. 31 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "text", + "content": "17. Jiang, L., Huang, D., Liu, M., Yang, W.: Beyond synthetic noise: Deep learning on controlled noisy labels. In: International Conference on Machine Learning. pp. 4804-4815. PMLR (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "text", + "content": "18. Jiang, L., Zhou, Z., Leung, T., Li, L.J., Fei-Fei, L.: MentorNet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In: International Conference on Machine Learning. pp. 2304-2313. PMLR (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 258 + ], + "type": "text", + "content": "19. Jiang, Z., Zhou, K., Liu, Z., Li, L., Chen, R., Choi, S.H., Hu, X.: An information fusion approach to learning with instance-dependent label noise. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=ecH2FKaARUp" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "20. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in Neural Information Processing Systems 33, 18661-18673 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "text", + "content": "21. Kim, T., Ko, J., Choi, J., Yun, S.Y.: FINE samples for learning with noisy labels. In: Advances in Neural Information Processing Systems. vol. 34 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "text", + "content": "22. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. Tech. rep., University of Toronto (2009)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: ImageNet classification with deep convolutional neural networks. In: Advances in Neural Information Processing Systems. vol. 25 (2012)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "24. Kye, S.M., Choi, K., Yi, J., Chang, B.: Learning with noisy labels by efficient transition matrix estimation to combat label miscorrection. In: European Conference on Computer Vision. pp. 717-738. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "25. Li, J., Li, G., Liu, F., Yu, Y.: Neighborhood collective estimation for noisy label identification and correction. In: European Conference on Computer Vision. pp. 128-145. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 481, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 467 + ], + "type": "text", + "content": "26. Li, J., Socher, R., Hoi, S.C.: DivideMix: Learning with noisy labels as semi-supervised learning. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "text", + "content": "27. Li, W., Wang, L., Li, W., Agustsson, E., Gool, L.V.: WebVision Database: Visual learning and understanding from web data. CoRR (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "text", + "content": "28. Liu, S., Niles-Weed, J., Razavian, N., Fernandez-Granda, C.: Early-learning regularization prevents memorization of noisy labels. In: Advances in Neural Information Processing Systems. vol. 33, pp. 20331-20342 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "type": "text", + "content": "29. Liu, Y., Cheng, H., Zhang, K.: Identifiability of label noise transition matrix. In: International Conference on Machine Learning (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "text", + "content": "30. Ma, X., Huang, H., Wang, Y., Romano, S., Erfani, S., Bailey, J.: Normalized loss functions for deep learning with noisy labels. In: International Conference on Machine Learning. pp. 6543-6553. PMLR (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 578, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 599 + ], + "type": "text", + "content": "31. Malach, E., Shalev-Shwartz, S.: Decoupling \"when to update\" from \"how to update\". In: Advances in Neural Information Processing Systems. vol. 30 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 408, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 408, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 408, + 610 + ], + "type": "text", + "content": "32. McLachlan, G.J., Peel, D.: Finite mixture models. Wiley (1996)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "text", + "content": "33. Neal, R.M., Hinton, G.E.: A view of the EM algorithm that justifies incremental, sparse, and other variants. Learning in graphical models pp. 355-368 (1998)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "34. Neyshabur, B., Bhojanapalli, S., McAllester, D., Srebro, N.: Exploring generalization in deep learning. In: Advances in Neural Information Processing Systems. vol. 30 (2017)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "A. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "35. Nishi, K., Ding, Y., Rich, A., Hollerer, T.: Augmentation strategies for learning with noisy labels. In: Conference on Computer Vision and Pattern Recognition. pp. 8022-8031 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 205 + ], + "type": "text", + "content": "36. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: PyTorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems 32 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 206, + 482, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 482, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 482, + 239 + ], + "type": "text", + "content": "37. Shamshad, F., Khan, S., Zamir, S.W., Khan, M.H., Hayat, M., Khan, F.S., Fu, H.: Transformers in medical imaging: A survey. Medical Image Analysis p. 102802 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 239, + 482, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 482, + 273 + ], + "type": "text", + "content": "38. Smart, B., Carneiro, G.: Bootstrapping the relationship between images and their clean and noisy labels. In: Winter Conference on Applications of Computer Vision. pp. 5344-5354 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 274, + 482, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 482, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 482, + 306 + ], + "type": "text", + "content": "39. Song, H., Kim, M., Park, D., Shin, Y., Lee, J.G.: Learning from noisy labels with deep neural networks: A survey. IEEE Transactions on Neural Networks and Learning Systems (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 307, + 482, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 482, + 340 + ], + "type": "text", + "content": "40. Tanaka, D., Ikami, D., Yamasaki, T., Aizawa, K.: Joint optimization framework for learning with noisy labels. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 341, + 482, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 482, + 362 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 482, + 362 + ], + "type": "text", + "content": "41. Titterington, D.M.: On the likelihood of mixtures of distributions. Biometrika 71(3), 511-522 (1984). https://doi.org/10.1093/biomet/71.3.511" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 363, + 482, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 482, + 396 + ], + "type": "text", + "content": "42. Trummer, I.: From BERT to GPT-3 codex: harnessing the potential of very large language models for data management. VLDB Endowment 15(12), 3770-3773 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 396, + 482, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 482, + 430 + ], + "type": "text", + "content": "43. Tu, H., Menzies, T.: Debtfree: minimizing labeling cost in self-admitted technical debt identification using semi-supervised learning. Empirical Software Engineering 27(4), 80 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 430, + 482, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 482, + 464 + ], + "type": "text", + "content": "44. Xia, X., Liu, T., Han, B., Wang, N., Gong, M., Liu, H., Niu, G., Tao, D., Sugiyama, M.: Part-dependent label noise: Towards instance-dependent label noise. In: Advances in Neural Information Processing Systems. vol. 33, pp. 7597-7610 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 464, + 482, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 482, + 497 + ], + "type": "text", + "content": "45. Xiao, T., Xia, T., Yang, Y., Huang, C., Wang, X.: Learning from massive noisy labeled data for image classification. In: Conference on Computer Vision and Pattern Recognition. pp. 2691-2699 (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 498, + 482, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 482, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 482, + 530 + ], + "type": "text", + "content": "46. Xu, Y., Zhu, L., Jiang, L., Yang, Y.: Faster meta update strategy for noise-robust deep learning. In: Conference on Computer Vision and Pattern Recognition. pp. 144-153 (June 2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 531, + 482, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 482, + 564 + ], + "type": "text", + "content": "47. Yang, S., Jiang, L., Liu, Z., Loy, C.C.: Unsupervised image-to-image translation with generative prior. In: Conference on Computer Vision and Pattern Recognition. pp. 18332-18341 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 565, + 482, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 482, + 597 + ], + "type": "text", + "content": "48. Yang, S., Yang, E., Han, B., Liu, Y., Xu, M., Niu, G., Liu, T.: Estimating instance-dependent bayes-label transition matrix using a deep neural network. In: International Conference on Machine Learning. pp. 25302-25312. PMLR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 599, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 482, + 632 + ], + "type": "text", + "content": "49. Yao, Q., Yang, H., Han, B., Niu, G., Kwok, J.T.Y.: Searching to exploit memorization effect in learning with noisy labels. In: International Conference on Machine Learning. pp. 10789-10798. PMLR (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 482, + 665 + ], + "type": "text", + "content": "50. Yao, Y., Liu, T., Gong, M., Han, B., Niu, G., Zhang, K.: Instance-dependent label-noise learning under a structural causal model. In: Advances in Neural Information Processing Systems. vol. 34 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "type": "text", + "content": "Noisy-Label Learning with Noise-Rate Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 436 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "51. Yao, Y., Liu, T., Han, B., Gong, M., Deng, J., Niu, G., Sugiyama, M.: Dual T: Reducing estimation error for transition matrix in label-noise learning. In: Advances in Neural Information Processing Systems. vol. 33, pp. 7260-7271 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "52. Zhang, C., Bengio, S., Hardt, M., Recht, B., Vinyals, O.: Understanding deep learning requires rethinking generalization. In: International Conference on Learning Representations (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "text", + "content": "53. Zhang, C., Bengio, S., Hardt, M., Recht, B., Vinyals, O.: Understanding deep learning (still) requires rethinking generalization. Communications of the ACM 64(3), 107-115 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "54. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. In: International Conference on Learning Representations (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "55. Zhang, Y., Niu, G., Sugiyama, M.: Learning noise transition matrix from only noisy labels via total variation regularization. In: International Conference on Machine Learning. pp. 12501-12512. PMLR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "text", + "content": "56. Zhang, Z., Sabuncu, M.: Generalized cross entropy loss for training deep neural networks with noisy labels. Advances in Neural Information Processing Systems 31 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 303, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 481, + 335 + ], + "type": "text", + "content": "57. Zhang, Z., Pfister, T.: Learning fast sample re-weighting without reward data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 725-734 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 335, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 481, + 369 + ], + "type": "text", + "content": "58. Zhao, G., Li, G., Qin, Y., Liu, F., Yu, Y.: Centrality and consistency: two-stage clean samples identification for learning with instance-dependent noisy labels. In: European Conference on Computer Vision. pp. 21-37. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "59. Zheltonozhskii, E., Baskin, C., Mendelson, A., Bronstein, A.M., Litany, O.: Contrast to divide: Self-supervised pre-training for learning with noisy labels. In: Winter Conference on Applications of Computer Vision. pp. 1657-1667 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 402, + 481, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 436 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 436 + ], + "type": "text", + "content": "60. Zhu, Z., Liu, T., Liu, Y.: A second-order approach to learning with instance-dependent label noise. In: Conference on Computer Vision and Pattern Recognition. pp. 10113-10123 (June 2021)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "A. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Instant 3D Human Avatar Generation using Image Diffusion Models/848a10c5-4ff1-45f7-8010-a0858cf611ea_content_list.json b/2024/Instant 3D Human Avatar Generation using Image Diffusion Models/848a10c5-4ff1-45f7-8010-a0858cf611ea_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7d142a8dfcc5c550c1f4d6d0a707e35a67506731 --- /dev/null +++ b/2024/Instant 3D Human Avatar Generation using Image Diffusion Models/848a10c5-4ff1-45f7-8010-a0858cf611ea_content_list.json @@ -0,0 +1,1907 @@ +[ + { + "type": "text", + "text": "Instant 3D Human Avatar Generation using Image Diffusion Models", + "text_level": 1, + "bbox": [ + 250, + 141, + 751, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nikos Kolotouros, Thiemo Alldieck, Enric Corona, Eduard Gabriel Bazavan, and Cristian Sminchisescu", + "bbox": [ + 243, + 204, + 756, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google Research\\* {kolotouros, alldieck, egbazavan, enriccorona, sminchisescu}@google.com", + "bbox": [ + 241, + 247, + 759, + 273 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/13947c3f92d8dc40fa114453e17f43421b851573c29a2de60d376a787ccfaa73.jpg", + "image_caption": [ + "Fig. 1: We present AvatarPopUp, a new method for the automatic generation of 3D human assets. AvatarPopUp can generate rigged 3D models from text or from single images and has control over body pose and shape. In this example, we show 77 models generated from various text prompts in 12 minutes on a single GPU." + ], + "image_footnote": [], + "bbox": [ + 220, + 289, + 779, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We present AvatarPopUp, a method for fast, high quality 3D human avatar generation from different input modalities, such as images and text prompts and with control over the generated pose and shape. The common theme is the use of diffusion-based image generation networks that are specialized for each particular task, followed by a 3D lifting network. We purposefully decouple the generation from the 3D modeling which allows us to leverage powerful image synthesis priors, trained on billions of text-image pairs. We fine-tune latent diffusion networks with additional image conditioning for image generation and back-view prediction, and to support qualitatively different multiple 3D hypotheses. Our partial fine-tuning approach allows to adapt the networks for each task without inducing catastrophic forgetting. In our experiments, we demonstrate that our method produces accurate, high-quality 3D avatars with diverse appearance that respect the multimodal text, image, and body control signals. Our approach can produce a 3D model in as few as 2 seconds, a four orders of magnitude speedup w.r.t. the vast majority of existing methods, most of which solve only a subset of our tasks, and with fewer controls. AvatarPopUp enables applications that require the controlled 3D generation of human avatars at scale. The project website can be found at https://www.nikoskolot.com/avatarpopulation/.", + "bbox": [ + 259, + 531, + 740, + 809 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Now at Google DeepMind.", + "bbox": [ + 220, + 825, + 411, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 374, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We present AvatarPopUp, a method for instant generation of rigged full-body 3D human avatars, with multimodal controls in the form of text, images, and/or human pose and shape. The remarkable recent progress in image synthesis [15, 24, 54, 56, 58, 63] acted as a catalyst for a wide range of media generation applications. In just a few years, we have witnessed rapid developments in video generation [9, 23, 34, 68, 76], audio synthesis [49, 77] or text-to-3D object generation [40, 41, 51, 53, 65], among others. Pivotal to success of all these methods is their probabilistic nature, however this requires very large training sets. While inspiring efforts have been made [14], training set size is still a problem in many domains, and particularly for 3D. In an attempt to alleviate the need for massive 3D datasets, DreamFusion [51] leverages the rich priors of text-to-image diffusion models in an optimization framework. The influential DreamFusion ideas were also quickly adopted for 3D avatar creation [32, 37], a field previously dominated by image- or video-based reconstruction solutions [6, 8, 60, 61]. Text-to-avatar methods enabled novel creative processes, but came with a significant drawback. While image-based methods typically use pretrained feed-forward networks and create outputs in seconds, existing text-to-avatar solutions are optimization-based and take minutes to several hours to complete, per instance.", + "bbox": [ + 217, + 175, + 785, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we are closing this gap and present, for the first time, a methodology for instant, text-controlled, rigged, full-body 3D human avatar creation. Our AvatarPopUp is purely feed-forward, can be conditioned on images and textual descriptions, allows fine-grained control over the generated body pose and shape, can generate multiple hypotheses, and runs in 2-10 seconds per instance.", + "bbox": [ + 217, + 448, + 785, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Key to success is our pragmatic decoupling of the two stages of probabilistic text-to-image generation and 3D lifting. Decoupling 2D generation and 3D lifting has two major advantages: (1) We can leverage the power of pretrained text-to-image generative networks, which have shown impressive results in modeling complex conditional distributions. Trained with large training sets of images, both generation quality and diversity are very high. (2) We alleviate the need for very large 3D datasets required to train state-of-the-art generative 3D models. Our method generates diverse plausible image configurations that contain rich enough information that can be lifted in 3D with minimal ambiguity. In other words, we distribute the workload between two expert systems: a pretrained probabilistic generation network fine-tuned for our task to produce front and back image views of the person, and a state-of-the-art unimodal, feed-forward image-to-3D model that can be trained using comparably small datasets.", + "bbox": [ + 217, + 523, + 785, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our proposed decoupling strategy allows us to maximally exploit available data sources with different levels of supervision. We first fine-tune a pretrained Latent Diffusion network to generate images of people based on textual descriptions and with additional control over the desired pose and shape. This step does not require any ground truth 3D data for supervision and enables scaling our image generator to web scale data of images of people in various poses. Next, we leverage a small-scale dataset of scanned 3D human assets and fine-tune a second latent diffusion network to learn the distribution of back side views con", + "bbox": [ + 217, + 719, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "N. Kolotouros et al.", + "bbox": [ + 271, + 114, + 403, + 126 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/7b3a3c98dd89b2fa8378124325d793e29af52efddfb4da07794600ca2bc71598.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Control from Text Control from Image Shape Control
BodyShapeGeneratesGeometryTexture
✓✓✓✓✓✓/XRuntime
X✓X✓✓/XXseveral hoursOptimization-based text-to-3D [26,32,37]
✓X✓✓XXsecondsImage-to-3D [4,8,60,61]
XX✓✓✓XminutesHybrid [31]
✓✓✓✓✓✓seconds3D GANs [16,25]
secondsAvatarPopUp (Ours)
", + "bbox": [ + 215, + 143, + 785, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1: AvatarPopUp generates 3D assets with texture from text prompts or input images of a target subject and can be controlled with body pose and shape. In contrast to baselines that require up to hours per prompt, our model takes under five seconds and can de facto by used in interactive applications. In the experimental section we demonstrate the large diversity of the model, and applications in cloth editing.", + "bbox": [ + 212, + 241, + 782, + 310 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ditioned on a front view image of the person. We optionally also condition on a textual description that can naturally complement the evidence available in the front view image. Furthermore, we propose a novel fine-tuning strategy that prevents overfitting to the new datasets. Finally, we design and train a 3D reconstruction network that predicts a textured 3D shape in the form of an implicit signed distance field given the pair of front and back views and optionally 3D body signals. The resulting cascaded method, which we call AvatarPopUp supports a wide range of 3D generation and reconstruction tasks: First, it enables fast and interactive 3D generation of assets at scale, see Fig. 1. Second, we can repurpose parts of the cascade for image-based 3D reconstruction at state-of-the-art quality. Finally, we demonstrate how AvatarPopUp can be used for creative editing tasks exemplified in 3D virtual try-on with body shape preservation. To summarize, our main contributions are:", + "bbox": [ + 212, + 329, + 784, + 525 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A method for controllable 3D human avatar generation, based on multimodal text, pose, shape and image input signals, that outputs a detailed human mesh instance in 2-10 seconds.", + "- We propose a simple yet effective way to fine-tune pretrained diffusion models on small-scale datasets, without inducing catastrophic forgetting.", + "- While not our primary goal, our approach achieves state-of-the-art results in single-image 3D reconstruction and enables 3D creative editing applications." + ], + "bbox": [ + 225, + 537, + 782, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 667, + 387, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1 summarizes the characteristics of AvatarPopUp in comparison to previous work, along several important property axes.", + "bbox": [ + 212, + 702, + 782, + 732 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-3D generation. The success of text-to-image models [55, 57, 59] was quickly followed by a significant amount of work on text-to-3D content generation [38, 45, 51, 66]. Due to limited training data, methods typically use optimization approaches, where a neural representation [47] is optimized per instance by minimizing a distillation loss [51] derived from large text-to-image models. This idea has been extended to generate human avatars [18, 27, 29, 32, 37, 69, 73, 78–80] or heads [19, 36, 39] enabling the text-based creation of 3D human assets that", + "bbox": [ + 212, + 734, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Instant 3D Human Avatar Generation", + "bbox": [ + 475, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "are diverse in terms of shape, appearance, clothing and various accessories. In these works, the optimization process is often regularized using a 3D body model [32,37,78], which also enables animation. However, such approaches generally take hours per instance, and rendering is slow. With the appearance of Gaussian Splatting [30], other works [2,42,82] reduced rendering time at the expense of accurate geometry. In any case, creating an avatar still takes a significant amount of time, making such methods unsuitable for interactive applications. In this work we propose an alternative direction, which also builds upon the success of text-to-image models and combines them with 3D reconstruction pipelines. Also related are 3D human generation methods. AG3D [16] and EVA3D [25] are GAN-based methods learned from 2D data, that allow to sample 3D humans anchored in a 3D body model. CHUPA [31] generates dual normal maps based on text and then fits a body model to obtain a full 3D representation. While generation is similar in spirit to our method, CHUPA requires optimization per instance and does not generate texture.", + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Photorealistic 3D Human Reconstruction. Our framework generates 3D human assets and builds on top of state-of-the-art 3D reconstruction techniques. This has been widely explored in the past and can be roughly categorized by its use of explicit or implicit representations. An important line of work leverages 3D body models [43,50,72] and reconstructs their associated parameters, in some cases extended with vertex offsets to represent some clothing and hair detail [4-7, 48, 85]. Other efforts have considered voxels [67, 84], depth maps [17] and more recently implicit representations [8, 13, 21, 28, 60-62, 70, 71, 74, 75, 83]. Being topology free, the latter allows the representation of loose clothing more easily. They typically provide more detail and enable high-resolution reconstruction, often conditioned on local pixel-aligned features [60]. On the other hand, these methods yield reconstructions with no semantic labels that cannot be easily animated. To solve this problem, some work combined body models with implicit representations [13, 21, 27, 28, 70, 71], but this is prone to errors when the pose is noisy at inference time. In contrast, we drive the synthesis process with guidance from an input body model – sampled or estimated – so that the generated image is well aligned with the body prior. As we show in Sec. 4, this allows rigging the 3D avatar without post-processing and natively supports 3D animation.", + "bbox": [ + 212, + 380, + 787, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a single input image of a person, previous work aims to generate realistic reconstructions for the non-visible parts. However, this often leads to blurry textures and lack of geometric detail, e.g. no wrinkles. Some methods [60, 61, 70, 71] generate back normal maps to enhance details, or consider probabilistic reconstructions [3, 62]. However, all these methods cannot be prompted from text or other modalities and still yield limited diversity. In contrast, we guide the synthesis process by means of generated front and back images, yielding high-quality 3D reconstructions. Another challenge in previous work is limited training data. Most prior methods rely on a few hundred 3D scans, due to the pricey and laborious process of good quality human capture. We alleviate the need for large scale 3D training data by proposing a framework that can quickly generate humans with a given clothing, pose and shape.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "N. Kolotouros et al.", + "bbox": [ + 271, + 114, + 403, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/162ab2d8066f91630fb298720f8e4fd1e5303b92e5df8162c95d919a99d4e1b2.jpg", + "image_caption": [ + "Fig. 2: AvatarPopUp method. (Top) AvatarPopUp builds on the capacity of text-to-image models to generate highly detailed and diverse input images. First, a Latent Diffusion network takes a text prompt and a target body pose and shape $\\mathcal{G}$ , and generates a highly detailed front image $I_{f}$ of a person. Next, a second network generates a consistent back view $I_{b}$ in the same pose and clothing. (Bottom) We perform pixel-aligned 3D reconstruction given the generated front/back views $I_{f}, I_{b}$ and optionally the given 3D body pose and shape $\\mathcal{G}$ . This decoupling enables the generation of 3D avatars from text, images or a combination of the two." + ], + "image_footnote": [], + "bbox": [ + 218, + 141, + 787, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 474, + 330, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We learn a distribution $p(X|c)$ of textured 3D shapes $X$ conditioned on a collection of signals $c$ that we factorize as follows", + "bbox": [ + 212, + 510, + 784, + 541 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np (X | c) = \\int \\int p \\left(X \\mid I _ {f}, I _ {b}, c\\right) \\cdot p \\left(I _ {b} \\mid I _ {f}, c\\right) \\cdot p \\left(I _ {f} \\mid c\\right) d I _ {f} d I _ {b}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 553, + 785, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p(X|I_f, I_b, c)$ is the probability of 3D shape $X$ given $c$ and front and back image observations $I_f$ and $I_b$ respectively, $p(I_b|I_f, c)$ is the probability of the back view image given the front image $I_f$ and conditioning signals $c$ , and $p(I_f|c)$ the conditional probability of front view images of the person given $c$ .", + "bbox": [ + 212, + 595, + 784, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Computing the integral in (1) is intractable, but our goal is to generate samples from the distribution rather than expectations. To do so, we employ ancestral sampling. We first sample a front view $I_{f}$ given $c$ , we then sample a back view $I_{b}$ given $I_{f}$ and $c$ , and last we sample the 3D reconstruction based on the entire context. In practice, $p(I_f|c)$ and $p(I_b|I_f,c)$ are implemented using Latent Diffusion models, whereas $p(X|I_f,I_b,c)$ is a unimodal, neural implicit field generator.", + "bbox": [ + 212, + 657, + 787, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the case of single-image 3D reconstruction the conditioning signal $c$ is $I_{f}$ , and consequently we can omit the first step. For text-based generation, $c$ is a text prompt describing the appearance of the person together with a signal encoding the body pose and shape. The conditioning information $c$ may be extended with additional signals, as in the case of 3D editing, cf. Sec. 4.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Instant 3D Human Avatar Generation", + "bbox": [ + 475, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Controllable Text-to-Image Generator", + "text_level": 1, + "bbox": [ + 215, + 146, + 576, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Recent advances in diffusion-based text-to-image generation networks [56] have enabled synthesizing high-quality images given only a text prompt as input. However, for certain use cases, such as human generation, it is difficult to inject fine-grained, inherently continuous forms of control, like the 3D pose of people or their precise body shape proportions, when generating with text alone.", + "bbox": [ + 212, + 169, + 782, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inspired by ControlNet [81], we propose to add simultaneous control over body pose and shape by augmenting a pretrained Latent Diffusion network with an additional image input that jointly encodes both modalities. For control we use GHUM [72] but other models (e.g. [43]) can be used. Specifically given 3D pose and shape parameters $\\pmb{\\theta}$ and $\\pmb{\\beta}$ respectively, we render the corresponding mesh $\\mathcal{M} = \\mathrm{GHUM}(\\pmb{\\theta},\\pmb{\\beta})$ using GHUM's template coordinates and posed vertex locations as 6D vertex colors, obtaining a dense, pixel-aligned pose- and shape-informed control signal $\\mathcal{G}$ .", + "bbox": [ + 212, + 244, + 784, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To fine-tune the network, we generate a dataset of images of people with corresponding GHUM 3D pose and shape parameters and text annotations. This dataset is comprised of a set of scanned assets [1] that are rendered from different viewpoints, as well as a set of real images scraped from the web. For the synthetic part of the dataset, the pose and shape parameters are obtained by fitting GHUM to 3D scans. We additionally use real images to which we fit GHUM using keypoint optimization in the style of [33]. For all images we obtained text annotations using an off-the shelf image captioning system [12] by prompting it to describe the clothing of the people in the image. Since we are interested in generating 3D human assets, we additionally mask out the background in all images, and train the network to output segmented images. This makes the downstream 3D reconstruction task easier, and improves the reconstruction quality because it encourages the network to focus on human appearance, rather than allocating capacity to model complex backgrounds.", + "bbox": [ + 212, + 366, + 784, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We want to exploit the rich priors learned by text-to-image foundation models by fine-tuning a Latent Diffusion model [56] with the dense GHUM rendering as an additional input. For fine-tuning we propose a simpler and more lightweight method than a standard ControlNet. We pad the weights of the input convolutional layer with additional channels initialized with zeros, and we then fine-tune only the weights of the convolutional layers of the encoder network. All the decoder and attention layers are kept frozen. With this simple strategy, even though our model is trained on a relatively small set of images, it can generalize to unseen types of clothing. At the same time, our strategy is more practical than training a ControlNet, which involves keeping in separate copy of the original network weights in memory, and thus enables fine-tuning large models with moderate hardware utilization. We optimize the encoder of the diffusion model by minimizing the simple variant of the diffusion loss [24]", + "bbox": [ + 212, + 578, + 785, + 773 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(\\psi_ {\\text {e n c}}\\right) = \\mathbb {E} _ {\\mathcal {E} (x), \\epsilon , t, \\tau , \\mathcal {G}} \\left\\| \\epsilon - \\epsilon_ {\\psi} \\left(z _ {t}, t, \\tau , \\mathcal {E} (\\mathcal {G})\\right) \\right\\|, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 784, + 784, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $t \\in \\{1, \\dots, T\\}$ is the diffusion time step, $\\epsilon \\sim \\mathcal{N}(0, I)$ the injected noise, $z_{t} = \\alpha_{t} \\mathcal{E}(x) + \\nu_{t}$ is the noisy image latent, $\\tau$ is the text encoding, $\\mathcal{E}(\\mathcal{G})$ the", + "bbox": [ + 215, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "N. Kolotouros et al.", + "bbox": [ + 271, + 114, + 403, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1db3f8e20ecd95376e3f40ffa57af103cd1f5e34035f9d69a43d3c1d082b2377.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 148, + 331, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/01db6990a2368276c251bea01ea04cb4d3d3bdd0a1562e42583b377a7562c54b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 150, + 455, + 241 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/69caf4714bb22336466c3167fea99aa8a152d2eb57025c327c4df4a0a4e78ad4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 150, + 586, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/581460b48d12d79363f37c039c9742e8218ee53639a8e58fb60d61cc7cd679ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 668, + 150, + 715, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cade929a17ec102891f1cc8fcf8972281ae09fad17a6875f4a71ac34d99e57bd.jpg", + "image_caption": [ + "Fig. 3: Diverse back view hypotheses. Conditioned on the front view, our method is able to generate diverse plausible back views of the person, with different hairstyles, wrinkle patterns, or lighting. Our network can also be controlled with text (second row), to add fine-grained detail to our generated back-side views." + ], + "image_footnote": [], + "bbox": [ + 279, + 258, + 331, + 353 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc4a18690be7fdbe678f416b58d57f4d67372038af4ac46a7e6022b1cd0c7b98.jpg", + "image_caption": [ + "No text conditioning" + ], + "image_footnote": [], + "bbox": [ + 410, + 258, + 460, + 352 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e2a400732757cac23a4a9922daf6f5839c31b9736fdd90dda848c9c8b4941f10.jpg", + "image_caption": [ + "+“gray hair”" + ], + "image_footnote": [], + "bbox": [ + 537, + 258, + 589, + 352 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1daf589e7a6556f159cd04ac5d4f59caf3364c3b6768c12b31433d9c80a1db40.jpg", + "image_caption": [ + "+“with a tattoo”" + ], + "image_footnote": [], + "bbox": [ + 668, + 258, + 718, + 352 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "latent encoding of the dense GHUM signal, and $\\psi_{\\mathrm{enc}}$ the encoder subset of the denoising UNet parameters $\\psi$ .", + "bbox": [ + 212, + 446, + 787, + 478 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Back View Generation", + "text_level": 1, + "bbox": [ + 214, + 500, + 449, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "One could try to lift the generated front views from the previous stage to 3D directly by applying a single-image 3D reconstruction method like PHORHUM [8]. However, because of the inherent ambiguity of the problem this will result in significant loss of geometric detail and blurry textures for the non-visible body surfaces. To avoid this, we propose to fine-tune again a latent diffusion network with the same strategy as in the previous section. This time the additional image conditioning is a front view and optionally a text prompt, and we train the network to learn the distribution of back views conditioned on the front view. The additional text prompt can be used in cases where it is desired to additionally guide the generation by very specific properties. Fig. 3 shows different back sides sampled from the conditional distribution. We also show that the additional text inputs are useful in modulating certain parts of the generation that are not immediately deducible from the front image, such as hairstyles or specific patterns.", + "bbox": [ + 212, + 522, + 787, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 3D Reconstruction Model", + "text_level": 1, + "bbox": [ + 214, + 757, + 477, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our 3D reconstruction network is inspired by PHORHUM [8], and our design choices are informed by the limitations of typical single-image 3D reconstruction methods. Specifically, given a collection of input image signals $\\mathcal{I} = \\{I_f, I_b, \\mathcal{G}\\}$ , we first concatenate them, and then use a convolutional encoder $G$ to compute", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Instant 3D Human Avatar Generation", + "bbox": [ + 475, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "a pixel-aligned feature map $G(\\mathcal{I})$ . The 3D body control signal $\\mathcal{G}$ is optional and may be omitted, e.g. for single-image reconstruction. Then, each point $\\mathbf{x} \\in \\mathbb{R}^3$ in the scene gets projected on this feature map to get pixel-aligned features $\\mathbf{z}_{\\mathbf{x}} = g(\\mathcal{I}, \\mathbf{x}; \\pi) = b(G(\\mathcal{I}), \\pi(\\mathbf{x}))$ using interpolation, where $b(\\cdot)$ is the bilinear sampling operator and $\\pi(\\mathbf{x})$ is the pixel location of the projection of $\\mathbf{x}$ using the camera $\\pi$ . These pixel aligned features are then concatenated with a positional encoding $\\gamma(\\mathbf{x})$ of the 3D point and are fed to an MLP $f$ that outputs the signed distance from the surface $d$ as well as surface color $c$ . Finally, the 3D shape $S$ is represented as the zero-level-set of $d$", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} (\\mathcal {I}) = \\left\\{\\mathbf {x} \\in \\mathbb {R} ^ {3} | f (g (I, \\mathbf {x}; \\pi), \\gamma (\\mathbf {x})) = (0, c) \\right\\}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 292, + 785, + 310 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\mathcal{S}$ can be transformed to a mesh directly, using Marching Cubes [44].", + "bbox": [ + 214, + 319, + 712, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Animation of Generated Avatars", + "text_level": 1, + "bbox": [ + 214, + 356, + 532, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method can generate diverse 3D avatars with various poses, shapes and appearances. Optionally, we may leverage the conditioning body model to rig the estimated 3D shape. As a result of our conditioning strategy, 3D avatars and the conditional body model instances are aligned in 3D. This allows us to anchor the reconstructed 3D shape on the body model surface [10] and re-pose or re-shape it accordingly. Alternatively, we may also transfer just the LBS skeleton and weights from the body model to the scan. This enables importing and animating the generated 3D assets in various rendering engines. See Fig. 8 and Sup. Mat. for examples and videos of animations of the generated 3D assets.", + "bbox": [ + 212, + 377, + 787, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 536, + 375, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data. We use meshes from RenderPeople [1] for training as well as our own captured data, totaling in $\\sim 10\\mathrm{K}$ scans with diverse poses, body shapes, and clothing styles. We render each scan with randomly sampled HDRI background, random cloth color augmentations, and lighting using Blender [11]. During this process, we render both front and back views used to train the different stages of our model. For the front image generation network we also use a set of $10\\mathrm{K}$ real images on which we fitted the GHUM model using 2D keypoints. For testing we defined a split based on subject identity and held out $\\sim 1\\mathrm{K}$ scans.", + "bbox": [ + 212, + 566, + 787, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We provide results for 2 different versions of our model. The standard quality model is generated in 2 seconds by running 5 DDIM [64] steps during inference and Marching Cubes at $256^{3}$ resolution. The high quality model is generated in 10 seconds using 50 DDIM steps and Marching Cubes at $512^{3}$ resolution. The timings were recorded on a single 40 GB A100 GPU. Unless otherwise stated, all results we report are obtained using the high quality model.", + "bbox": [ + 212, + 688, + 787, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Metrics and Baselines. We compare our method numerically for two different problems. First we consider the task of text-to-3D human generation, where we sample 100 different text prompts and compare against representative text-to-3D generation methods. For numerical comparisons we evaluate the text-image", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "N. Kolotouros et al.", + "bbox": [ + 271, + 114, + 403, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8c9a1eba9b64f3018c4f5159767cf74820e47964bd0031a4a402609ea2f6e9e0.jpg", + "image_caption": [ + "Fig. 4: Diversity of our 3D generation. For the same text prompt and the same pose and shape conditioning, our model can generate a diverse set of 3D avatars that respect both the text and the 3D body controls." + ], + "image_footnote": [], + "bbox": [ + 251, + 145, + 754, + 321 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "alignment using CLIP [52]. Specifically, we use the retrieval accuracy using CLIP, as proposed in [51]. We additionally show qualitative results.", + "bbox": [ + 212, + 390, + 784, + 421 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Second, we validate the performance of our 3D reconstruction component against state-of-the-art methods [8,27,60,61,70,71] considering both geometry and texture. Pixel-aligned image features dominate in recent work [8,60,61], but some methods proposed to combine them with body models [70,71], which offer advantages in being able to animate reconstructions. We also rely on pixel-aligned features, yet propose a method that inherently enables animation. We run AvatarPopUp by generating the back side of the subject and applying the reconstruction network. To evaluate 3D geometry, we report bi-directional Chamfer distance $\\times 10^{-3}$ , Normal Consistency (NC ↑), and Volumetric Intersection over Union (IoU ↑) after ICP alignment. However, these metrics do not necessarily correlate with good visual quality, e.g. Chamfer distance is minimized by smooth, non-detailed geometry. To measure the quality of reconstructions, we additionally report FID scores [22] of the front/back views for both geometry and texture.", + "bbox": [ + 212, + 422, + 787, + 633 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Text-to-3D Generation", + "text_level": 1, + "bbox": [ + 215, + 667, + 450, + 681 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In Fig. 4 we generate different avatars given the same text prompt and driving poses. We can see that our model is able to create a very diverse set of assets, a property not observed in the previous text-to-3D generation methods [32].", + "bbox": [ + 212, + 700, + 785, + 747 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In Tab. 2 we use CLIP to evaluate our model against other text-to-3D generation methods. In general, CLIP-based metrics are not indicative of the generated image quality, because they only consider the alignment with the text, and often over-saturated images with extreme details tend to have high CLIP scores. To further demonstrate that our method generates higher quality avatars, we also include a qualitative comparison in Fig. 5.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Instant 3D Human Avatar Generation", + "bbox": [ + 475, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/d5b878beddc754fe7938b70945f5feb2ef3b940e4f77b225bc761548b3f8a018.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ColorColorRuntime ↓
R-Prec. ↑Top-3 ↑R-Prec. ↑Top-3 ↑
0.680.920.040.258hDreamHuman [32]
0.560.820.030.153hTADA [37]
--0.030.083mCHUPA [31]
0.580.730.080.17~2sOurs
0.620.770.110.17~10sOurs (high quality)
", + "bbox": [ + 302, + 133, + 694, + 222 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 2: Numerical comparisons with other text-to-3d human generation methods. We mark the best and second best results. Our method allows to trade-off speed and quality and we report results for two different settings.", + "bbox": [ + 214, + 227, + 785, + 268 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/dca817aad1b5c4cf1303324e456c60a0543ca2f879737932aa8e2ffeef5c575e.jpg", + "image_caption": [ + "A person wearing a black off-shoulder top and a denim skirt", + "Fig. 5: Comparisons with text-to-3d human generation methods. Our method generates high quality results that respect the text prompt well, at a fraction of the others' runtime, cf. Tab. 3. TADA's results appear unnatural; DreamHuman failed for one subject and produces oversaturated colors; CHUPA failed to respect the prompt." + ], + "image_footnote": [], + "bbox": [ + 246, + 280, + 759, + 481 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Single-image 3D Reconstruction", + "text_level": 1, + "bbox": [ + 215, + 583, + 526, + 598 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "While not specifically designed for 3D reconstruction, our method achieves state-of-the-art performance also for this task. The evaluation setup is the following: given an input image $I$ , we draw one random sample from the back view image generator network, and then feed the pair of front/back images to our 3D reconstruction network. For all methods we extract a textured 3D mesh using Marching Cubes and report numerical results in Tab. 3. Furthermore, we show qualitative results in Fig. 6. Notably, our method not only performs on par numerically and qualitatively on reconstructed front views, but also generates highly detailed back view texture and geometry. Finally, we also compare with the optimization-based method TeCH [27] in Fig. 7. TeCH produces detailed front and back geometry but also also exhibits problems at times, rooted in its 3D pose estimation method. Most importantly, TeCH runs for several hours per instance, while ours computes results in a single feed-forward pass, in only a few seconds.", + "bbox": [ + 212, + 628, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "N. Kolotouros et al.", + "bbox": [ + 269, + 114, + 403, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/af7d81b77ae1ecfe223fe1e32c6a416f3f1c426307bb8a859318d6685aa752c5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
3D MetricsFID ↓
Ch. ↓IoU ↑NC ↑Color. FColor B.Normal F.Normal B.
7.10.500.72--29.657.2PIFu [60]
6.20.520.74--26.952.5PIFuHD [61]
7.00.480.72--36.149.0ICON [71]
3.40.580.76--18.336.1ECON [70]
0.840.710.8711.040.811.238.8PHORHUM* [8]
0.910.700.8611.020.012.825.8Ours
", + "bbox": [ + 259, + 143, + 736, + 243 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 3: Numerical comparisons with single-view 3D reconstructions methods and ablations of our method. We mark the best and second best results. All Chamfer metrics are $\\times 10^{-3}$ . Not all methods generate colors. For fair comparisons, we retrained PHORHUM [8] using the same data as AvatarPopUp. We observe comparable results in terms of 3D metrics, however ours performs better on generating more realistic and diverse back views and back normals.", + "bbox": [ + 212, + 250, + 787, + 332 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 3D Virtual Try On", + "text_level": 1, + "bbox": [ + 215, + 349, + 418, + 366 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "An immediate application of our method is the option to perform 3D garment edits for a given identity. Given an input image, we first recover 3D pose and shape parameters of the person in the image. Using a text prompt and the identity preservation strategy introduced below, we can generate updated images of the same person wearing different garments or accessories. To preserve the identity of the person, we first locate the person's head in the source image and use then Repaint [46] to out-paint a novel body for the given head while still conditioning on the estimated pose and shape parameters. With this strategy we ensure that our method not only preserves the facial characteristics but also the overall body proportions. In Fig. 9 we illustrate such editing examples. The generated 3D edits present garment details like wrinkles on both front and back views, and preserve the subjects' facial appearance. Also, note that body shape and identity are well preserved for the subjects, even though only one image is given. While there has been a significant amount of 2D virtual try-on research [20,35,86], our methodology can generate consistent and highly detailed 3D meshes that can be animated and rendered from other viewpoints.", + "bbox": [ + 212, + 377, + 787, + 619 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Animation", + "text_level": 1, + "bbox": [ + 215, + 642, + 351, + 656 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "AvatarPopUp enables animation of the generated assets by design, provided that its generation is conditioned on an underlying body model. In Fig. 8, we show an example of a generated avatar that is rigged automatically.", + "bbox": [ + 212, + 667, + 785, + 714 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 737, + 388, + 752 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Pose and shape encoding in the 3D reconstruction network. We compare the effectiveness of our additional pose and shape encoding inputs $\\mathcal{G}$ to our reconstruction network. To do so, we use the same set of 100 text prompts as before and for each text prompt sample a random pose and shape configuration. For each $(\\tau, \\theta, \\beta)$ triplet we run inferences and compute 2 meshes: one using", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Instant 3D Human Avatar Generation", + "bbox": [ + 475, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/33f7851c4d76bff20fb1c6fd45e6c8df67519d153221809c899f5db5aef1c0c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 133, + 263, + 215 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8e76543a34a53fe154685c8f8fdab35449968cff6ecd28b41931c8a9e6862184.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 218, + 254, + 294 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1a277a21efc4e57daeed803f8b0c412f8ffdb6a95d5267679749727b5d40ff08.jpg", + "image_caption": [ + "Fig. 6: Qualitative comparisons with state-of-the-art single image 3D reconstruction methods. Our method produces front color and normals on par with state-of-the-art and much more detailed back view hypotheses." + ], + "image_footnote": [], + "bbox": [ + 223, + 299, + 254, + 373 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e79dc5f2dedb4b3ffdf3bf42737cd570d004c8ecaa1c8c7ac072c867c4af560b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 136, + 460, + 215 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8ac23b6f8be52bac373b46279c603a437dc6cf667e7ff312e21282fce33fae41.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 218, + 323, + 294 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c005d6f61016cf555a1f4332c9f7c4c534e90b5f546040205e819479751e93b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 299, + 321, + 373 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/577b67f56e47e08f57816e503d5e515d093325fc5537f9cf62057260038e20df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 218, + 408, + 295 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f446eb4225152364a096a181f1464f51327cc608605628e28236310c96cdbe4a.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 331, + 299, + 405, + 375 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "", + "bbox": [ + 418, + 136, + 460, + 215 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/cd0de6d86362a5fa18e9385ccd15422ac54614390d4ab7ec8682c73ee8863a87.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 218, + 452, + 294 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f2ac061c35fa60b0f87b0fb223d1065cf808553e8c4e96565f8bb564ea206bfa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 299, + 452, + 375 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1ae583bb3a87dcbd16e91e626ca5c29c9236a0d206d8a46b537de1b18d5eb1d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 133, + 606, + 212 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/964b0c392b24ab8509e4eaeeb90c3449626aa9910e7945c0e6981bc61f1c079d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 215, + 607, + 294 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1e00fcc6ec50608af2f4641d9ee6804db03edbd9f36657dea13c35cb78157725.jpg", + "image_caption": [ + "PHORHUM" + ], + "image_footnote": [], + "bbox": [ + 501, + 297, + 606, + 375 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8c5920f736078ba8bc63186005d4bedecea707dd22e2bc532eb429d6c33ffd8e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 133, + 663, + 212 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/cc7e1d9099b9174a99c2ac3d57a8518d4fd370fcd087890a32876bc2d53b6234.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 217, + 640, + 292 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/39bb2b5585ebbd15757b5ecbd220da8505d42aa5141e57c8a4c0a52d23f86a0b.jpg", + "image_caption": [ + "" + ], + "image_footnote": [], + "bbox": [ + 612, + 299, + 643, + 375 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/208dab8bc67d85f3a927ccaa1a57637c436ee52f0d30e42e31b6b96ad1e42a62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 135, + 781, + 214 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7d0ed829c20412a073bdd09e851b60956f11e3e14aa9fb83ef930a400ff22c74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 218, + 728, + 295 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a95d4bd010c91addafda3846475302ee32d4172e6e61044918dd164d4b4da8b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 218, + 777, + 295 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/227939bc09641de63cfda27840ce207abcaf506027c3ad240eb91d9fee96c425.jpg", + "image_caption": [ + "PIFuHD" + ], + "image_footnote": [], + "bbox": [ + 699, + 299, + 745, + 375 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "only the front and back images, and another one additionally using the dense GHUM encodings. We then evaluate the Chamfer distance between each mesh and the corresponding GHUM mesh. The model using the additional GHUM signals has an average Chamfer distance of $d_{\\mathrm{with}} = 1.4$ whereas the one without $d_{\\mathrm{without}} = 8.6$ , thus validating our design choice. Not only is the control respected well, but this also allows for animation as discussed previously.", + "bbox": [ + 212, + 460, + 784, + 551 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Partial vs. Complete Network Finetuning. We finetune two Latent Diffusion networks, one in a standard way by optimizing all parameters, and another one using our proposed strategy where we only finetune the convolutional layers of the encoder. Empirically we observed that the network that was finetuned as a whole experienced catastrophic forgetting, and has poor performance when asked to generate types of garment not seen in the training set. Fig. 10 shows a comparison for text prompts not in the training set.", + "bbox": [ + 212, + 553, + 785, + 659 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5 Discussion & Conclusions", + "text_level": 1, + "bbox": [ + 215, + 683, + 500, + 699 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Limitations. AvatarPopUp inherits limitations from other pixel-aligned methods, e.g. with regions parallel to camera rays being not as detailed in the reconstruction. Further, artifacts might be visible in poses or very loose clothing that are underrepresented during training.", + "bbox": [ + 212, + 717, + 784, + 777 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ethical Considerations. We present a generative tool to create 3D human assets, thus reducing the risk of scanning and using real humans for training large scale 3D generative models. AvatarPopUp generates diverse results and can lead to a better coverage of subject distributions.", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "N. Kolotouros et al.", + "bbox": [ + 271, + 114, + 403, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/644893664bb6a0e81b86e5f6ed2631977fd60acfe99dfcc4712156cb815c5dff.jpg", + "image_caption": [ + "Fig. 7: Additional comparisons with TeCH. TeCH is optimization-based and takes several hours to complete for a single reconstruction. Our results are obtained in seconds, yet are as detailed and even less noisy." + ], + "image_footnote": [], + "bbox": [ + 212, + 165, + 272, + 554 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1e5512458355ed8fa858558aa05d8784f3f158ff317bff63f0750b309f745733.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 316, + 166, + 531, + 554 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/add9dc208474d3fb82997f320290655b0af98f036ed2a320767ba0541b6c766f.jpg", + "image_caption": [ + "TeCH" + ], + "image_footnote": [], + "bbox": [ + 576, + 166, + 779, + 554 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/798f16ce9656d504b756b7a0d45247a153e8448229982b1f4b3b01101a7f5eeb.jpg", + "image_caption": [ + "Fig. 8: Reposing example.. We first reconstruct an avatar \"wearing a gray suit\" in the A-pose and then we transfer it to different poses." + ], + "image_footnote": [], + "bbox": [ + 300, + 672, + 689, + 779 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Instant 3D Human Avatar Generation", + "bbox": [ + 475, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ca6e8a5633e19c806361af05dbdef45340e11facaea29dab86c9e17767da3fa7.jpg", + "image_caption": [ + "Fig. 9: Identity preserving 3D avatar editing. Our method allows for editing the clothing, while preserving the identity of the generated person. In each row we're using the image on the left as input, recover the person's pose and shape parameters and then run our method using the shown prompts. The generated avatars share the same identity and respect the target prompts." + ], + "image_footnote": [], + "bbox": [ + 243, + 133, + 759, + 377 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/7fbc485cb9f842c8d7873adff5ae0bea9ddd8f3ed28deed93f6287a79c08f91e.jpg", + "image_caption": [ + "Fig. 10: Partial vs. complete fine-tuning strategy. For each text prompt the first example is generated using our partial fine-tuning strategy whereas the second by fine-tuning the entire network. Empirically, partial fine-tuning is more resilient to catastrophic forgetting, generating more diverse and better text-aligned images." + ], + "image_footnote": [], + "bbox": [ + 290, + 464, + 717, + 563 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have presented AvatarPopUp, a novel framework for generating 3D human avatars controlled by text or images and yielding rigged 3D models in 2-10 seconds. AvatarPopUp is purely feed-forward, allows for fine-grained control over the generated body pose and shape, and can produce multiple qualitatively different hypotheses. AvatarPopUp is composed of a cascade of expert systems, decoupling image generation and 3D lifting. Through this design choice, AvatarPopUp benefits from both web-scale image datasets, ensuring high generation diversity, and from smaller size accurate 3D datasets, resulting in reconstructions with increased detail and precisely controlled based on text and identity specifications. In the future, we would like to explore other 3D construction strategies besides pixel-aligned features. Longer term, we aim to support highly detailed and controllable 3D human model generation for entertainment, education, architecture and art, or medical applications.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "N. Kolotouros et al.", + "bbox": [ + 271, + 114, + 403, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. https://renderpeople.com/", + "2. Abdal, R., Yifan, W., Shi, Z., Xu, Y., Po, R., Kuang, Z., Chen, Q., Yeung, D.Y., Wetzstein, G.: Gaussian shell maps for efficient 3d human generation. arXiv preprint arXiv:2311.17857 (2023)", + "3. AlBahar, B., Saito, S., Tseng, H.Y., Kim, C., Kopf, J., Huang, J.B.: Single-image 3d human digitization with shape-guided diffusion. In: SIGGRAPH Asia (2023)", + "4. Alldieck, T., Magnor, M., Bhatnagar, B.L., Theobalt, C., Pons-Moll, G.: Learning to reconstruct people in clothing from a single rgb camera. In: CVPR (2019)", + "5. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Detailed human avatars from monocular video. In: 3DV (2018)", + "6. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Video based reconstruction of 3D people models. In: CVPR (2018)", + "7. Alldieck, T., Pons-Moll, G., Theobalt, C., Magnor, M.: Tex2shape: Detailed full human body geometry from a single image. In: ICCV (2019)", + "8. Alldieck, T., Zanfir, M., Sminchisescu, C.: Photorealistic monocular 3D reconstruction of humans wearing clothing. In: CVPR (2022)", + "9. Bar-Tal, O., Chefer, H., Tov, O., Herrmann, C., Paiss, R., Zada, S., Ephrat, A., Hur, J., Li, Y., Michaeli, T., et al.: Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945 (2024)", + "0. Bazavan, E.G., Zanfir, A., Zanfir, M., Freeman, W.T., Sukthankar, R., Sminchisescu, C.: Hspace: Synthetic parametric humans animated in complex environments. arXiv (2021)", + "1. Blender Online Community: Blender - a 3D modelling and rendering package. Blender Foundation, Blender Institute, Amsterdam (2020), http://www.blender.org", + "12. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., Kolesnikov, A., Puigcerver, J., Ding, N., Rong, K., Akbari, H., Mishra, G., Xue, L., Thapliyal, A., Bradbury, J., Kuo, W., Seyedhosseini, M., Jia, C., Ayan, B.K., Riquelme, C., Steiner, A., Angelova, A., Zhai, X., Houlsby, N., Soricut, R.: Pali: A jointly-scaled multilingual language-image model (2022). https://doi.org/10.48550/ARXIV.2209.06794, https://arxiv.org/abs/2209.06794", + "13. Corona, E., Zanfir, M., Alldieck, T., Bazavan, E.G., Zanfir, A., Sminchisescu, C.: Structured 3d features for reconstructing relightable and animatable avatars. In: CVPR (2023)", + "14. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3d objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13142-13153 (2023)", + "15. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)", + "16. Dong, Z., Xu Chen, J.Y., J.Black, M., Hilliges, O., Geiger, A.: AG3D: Learning to generate 3D avatars from 2D image collections. In: International Conference on Computer Vision (ICCV) (2023)", + "17. Gabeur, V., Franco, J.S., Martin, X., Schmid, C., Rogez, G.: Moulding humans: Non-parametric 3D human shape estimation from single images. In: ICCV (2019)", + "18. Gong, C., Dai, Y., Li, R., Bao, A., Li, J., Yang, J., Zhang, Y., Li, X.: Text2 avatar: Text to 3d human avatar generation with codebook-driven body controllable attribute. arXiv preprint arXiv:2401.00711 (2024)" + ], + "bbox": [ + 218, + 172, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Instant 3D Human Avatar Generation", + "bbox": [ + 475, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Han, X., Cao, Y., Han, K., Zhu, X., Deng, J., Song, Y.Z., Xiang, T., Wong, K.Y.K.: Headsculpt: Crafting 3d head avatars with text. Advances in Neural Information Processing Systems 36 (2024)", + "20. Han, X., Wu, Z., Wu, Z., Yu, R., Davis, L.S.: Viton: An image-based virtual try-on network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7543-7552 (2018)", + "21. He, T., Xu, Y., Saito, S., Soatto, S., Tung, T.: Arch++: Animation-ready clothed human reconstruction revisited. In: CVPR (2021)", + "22. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)", + "23. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)", + "24. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239 (2020)", + "25. Hong, F., Chen, Z., LAN, Y., Pan, L., Liu, Z.: EVA3d: Compositional 3d human generation from 2d image collections. In: International Conference on Learning Representations (2023), https://openreview.net/forum?id=g7U9jD_2CUr", + "26. Hong, F., Zhang, M., Pan, L., Cai, Z., Yang, L., Liu, Z.: Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. ACM Transactions on Graphics (TOG) 41(4), 1-19 (2022)", + "27. Huang, Y., Yi, H., Xiu, Y., Liao, T., Tang, J., Cai, D., Thies, J.: TeCH: Text-guided Reconstruction of Lifelike Clothed Humans. In: International Conference on 3D Vision (3DV) (2024)", + "28. Huang, Z., Xu, Y., Lassner, C., Li, H., Tung, T.: Arch: Animatable reconstruction of clothed humans. In: CVPR (2020)", + "29. Jiang, R., Wang, C., Zhang, J., Chai, M., He, M., Chen, D., Liao, J.: Avatarcraft: Transforming text into neural human avatars with parameterized shape and pose control. arXiv preprint arXiv:2303.17606 (2023)", + "30. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)", + "31. Kim, B., Kwon, P., Lee, K., Lee, M., Han, S., Kim, D., Joo, H.: Chupa: Carving 3d clothed humans from skinned shape priors using 2d diffusion probabilistic models. arXiv preprint arXiv:2305.11870 (2023)", + "32. Kolotouros, N., Alldieck, T., Zanfir, A., Bazavan, E., Fieraru, M., Sminchisescu, C.: Dreamhuman: Animatable 3d avatars from text. Advances in Neural Information Processing Systems 36 (2024)", + "33. Kolotouros, N., Pavlakos, G., Black, M.J., Daniilidis, K.: Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2252-2261 (2019)", + "34. Kondratyuk, D., Yu, L., Gu, X., Lezama, J., Huang, J., Hornung, R., Adam, H., Akbari, H., Alon, Y., Birodkar, V., et al.: Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125 (2023)", + "35. Lee, S., Gu, G., Park, S., Choi, S., Choo, J.: High-resolution virtual try-on with misalignment and occlusion-handled conditions. In: European Conference on Computer Vision. pp. 204-219. Springer (2022)", + "36. Lei, B., Yu, K., Feng, M., Cui, M., Xie, X.: Diffusiongan3d: Boosting text-guided 3d generation and domain adaption by combining 3d gans and diffusion priors. arXiv preprint arXiv:2312.16837 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "N. Kolotouros et al.", + "bbox": [ + 271, + 114, + 403, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Liao, T., Yi, H., Xiu, Y., Tang, J., Huang, Y., Thies, J., Black, M.J.: Tada! text to animatable digital avatars. In: 3DV (2023)", + "38. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023)", + "39. Liu, H., Wang, X., Wan, Z., Shen, Y., Song, Y., Liao, J., Chen, Q.: Headartist: Text-conditioned 3d head generation with self score distillation. arXiv preprint arXiv:2312.07539 (2023)", + "40. Liu, M., Xu, C., Jin, H., Chen, L., Varma T, M., Xu, Z., Su, H.: One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems 36 (2024)", + "41. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9298-9309 (2023)", + "42. Liu, X., Zhan, X., Tang, J., Shan, Y., Zeng, G., Lin, D., Liu, X., Liu, Z.: Human-gaussian: Text-driven 3d human generation with gaussian splatting. arXiv preprint arXiv:2311.17061 (2023)", + "43. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ToG (2015)", + "44. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3D surface construction algorithm. SIGGRAPH (1987)", + "45. Lorraine, J., Xie, K., Zeng, X., Lin, C.H., Takikawa, T., Sharp, N., Lin, T.Y., Liu, M.Y., Fidler, S., Lucas, J.: Att3d: Amortized text-to-3d object synthesis. arXiv preprint arXiv:2306.07349 (2023)", + "46. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11461-11471 (2022)", + "47. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. ECCV (2020)", + "48. Onizuka, H., Hayirci, Z., Thomas, D., Sugimoto, A., Uchiyama, H., Taniguchi, R.i.: Tetratsdf: 3D human reconstruction from a single image with a tetrahedral outer shell. In: CVPR (2020)", + "49. Oord, A.v.d., Dieleman, S., Zen, H., Simonyan, K., Vinyals, O., Graves, A., Kalchbrenner, N., Senior, A., Kavukcuoglu, K.: Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499 (2016)", + "50. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: CVPR (2019)", + "51. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: Int. Conf. Learn. Represent. (2022)", + "52. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "53. Raj, A., Kaza, S., Poole, B., Niemeyer, M., Mildenhall, B., Ruiz, N., Zada, S., Aberman, K., Rubenstein, M., Barron, J., Li, Y., Jampani, V.: Dreambooth3d: Subject-driven text-to-3d generation. ICCV (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Instant 3D Human Avatar Generation", + "bbox": [ + 475, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "54. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)", + "55. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)", + "56. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10684-10695 (2022)", + "57. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "58. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS. vol. 35, pp. 36479-36494 (2022)", + "59. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)", + "60. Saito, S., Huang, Z., Natsume, R., Morishima, S., Kanazawa, A., Li, H.: Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In: ICCV (2019)", + "61. Saito, S., Simon, T., Saragih, J., Joo, H.: Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3D human digitization. In: CVPR (2020)", + "62. Sengupta, A., Alldieck, T., Kolotouros, N., Corona, E., Zanfir, A., Sminchisescu, C.: Diffhuman: Probabilistic photorealistic 3D reconstruction of humans. In: CVPR (2024)", + "63. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)", + "64. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: Int. Conf. Learn. Represent. (2020)", + "65. Sun, J., Zhang, B., Shao, R., Wang, L., Liu, W., Xie, Z., Liu, Y.: Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior. arXiv preprint arXiv:2310.16818 (2023)", + "66. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)", + "67. Varol, G., Ceylan, D., Russell, B., Yang, J., Yumer, E., Laptev, I., Schmid, C.: Bodynet: Volumetric inference of 3D human body shapes. In: ECCV (2018)", + "68. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399 (2022)", + "69. Wang, J., Liu, Y., Dou, Z., Yu, Z., Liang, Y., Li, X., Wang, W., Xie, R., Song, L.: Disentangled clothed avatar generation from text descriptions. arXiv preprint arXiv:2312.05295 (2023)", + "70. Xiu, Y., Yang, J., Cao, X., Tzionas, D., Black, M.J.: ECON: Explicit Clothed humans Optimized via Normal integration. In: CVPR (June 2023)", + "71. Xiu, Y., Yang, J., Tzionas, D., Black, M.J.: Icon: Implicit clothed humans obtained from normals. In: CVPR (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "N. Kolotouros et al.", + "bbox": [ + 271, + 114, + 403, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "72. Xu, H., Bazavan, E.G., Zanfir, A., Freeman, W.T., Sukthankar, R., Sminchisescu, C.: Ghum & ghuml: Generative 3D human shape and articulated pose models. In: CVPR (2020)", + "73. Xu, Y., Yang, Z., Yang, Y.: Seeavatar: Photorealistic text-to-3d avatar generation with constrained geometry and appearance. arXiv preprint arXiv:2312.08889 (2023)", + "74. Yang, Z., Wang, S., Manivasagam, S., Huang, Z., Ma, W.C., Yan, X., Yumer, E., Urtasun, R.: S3: Neural shape, skeleton, and skinning fields for 3D human modeling. In: CVPR (2021)", + "75. Yanting Chan, K., Lin, G., Zhao, H., Lin, W.: Integratedpifu: Integrated pixel aligned implicit function for single-view human reconstruction. In: ECCV (2022)", + "76. Zeng, Y., Wei, G., Zheng, J., Zou, J., Wei, Y., Zhang, Y., Li, H.: Make pixels dance: High-dynamic video generation. arXiv preprint arXiv:2311.10982 (2023)", + "77. Zhang, C., Zhang, C., Zheng, S., Zhang, M., Qamar, M., Bae, S.H., Kweon, I.S.: Audio diffusion model for speech synthesis: A survey on text to speech and speech enhancement in generative ai. arXiv preprint arXiv:2303.13336 (2023)", + "78. Zhang, H., Feng, Y., Kulits, P., Wen, Y., Thies, J., Black, M.J.: Text-guided generation and editing of compositional 3d avatars. arXiv preprint arXiv:2309.07125 (2023)", + "79. Zhang, H., Chen, B., Yang, H., Qu, L., Wang, X., Chen, L., Long, C., Zhu, F., Du, K., Zheng, M.: Avatarverse: High-quality & stable 3d avatar creation from text and pose. arXiv preprint arXiv:2308.03610 (2023)", + "80. Zhang, J., Zhang, X., Zhang, H., Liew, J.H., Zhang, C., Yang, Y., Feng, J.: Avatarstudio: High-fidelity and animatable 3d avatar creation from text. arXiv preprint arXiv:2311.17917 (2023)", + "81. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. IEEE International Conference on Computer Vision (ICCV) (2023)", + "82. Zhao, Z., Bao, Z., Li, Q., Qiu, G., Liu, K.: Psavatar: A point-based morphable shape model for real-time head avatar creation with 3d gaussian splatting. arXiv preprint arXiv:2401.12900 (2024)", + "83. Zheng, Z., Yu, T., Liu, Y., Dai, Q.: Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. PAMI (2021)", + "84. Zheng, Z., Yu, T., Wei, Y., Dai, Q., Liu, Y.: Deephuman: 3D human reconstruction from a single image. In: ICCV (2019)", + "85. Zhu, H., Zuo, X., Wang, S., Cao, X., Yang, R.: Detailed human shape estimation from a single image by hierarchical mesh deformation. In: CVPR (2019)", + "86. Zhu, L., Yang, D., Zhu, T., Reda, F., Chan, W., Sahara, C., Norouzi, M., Kemelmacher-Shlizerman, I.: Tryondiffusion: A tale of two unets. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4606-4615 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 702 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Instant 3D Human Avatar Generation", + "bbox": [ + 475, + 114, + 730, + 127 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Instant 3D Human Avatar Generation using Image Diffusion Models/848a10c5-4ff1-45f7-8010-a0858cf611ea_model.json b/2024/Instant 3D Human Avatar Generation using Image Diffusion Models/848a10c5-4ff1-45f7-8010-a0858cf611ea_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bcfbf7873c8c6bcafd6e08860bfd061e52f87d6b --- /dev/null +++ b/2024/Instant 3D Human Avatar Generation using Image Diffusion Models/848a10c5-4ff1-45f7-8010-a0858cf611ea_model.json @@ -0,0 +1,2900 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.25, + 0.142, + 0.753, + 0.186 + ], + "angle": 0, + "content": "Instant 3D Human Avatar Generation using Image Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.205, + 0.758, + 0.235 + ], + "angle": 0, + "content": "Nikos Kolotouros, Thiemo Alldieck, Enric Corona, Eduard Gabriel Bazavan, and Cristian Sminchisescu" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.248, + 0.761, + 0.275 + ], + "angle": 0, + "content": "Google Research\\* {kolotouros, alldieck, egbazavan, enriccorona, sminchisescu}@google.com" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.29, + 0.78, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.517 + ], + "angle": 0, + "content": "Fig. 1: We present AvatarPopUp, a new method for the automatic generation of 3D human assets. AvatarPopUp can generate rigged 3D models from text or from single images and has control over body pose and shape. In this example, we show 77 models generated from various text prompts in 12 minutes on a single GPU." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.532, + 0.741, + 0.81 + ], + "angle": 0, + "content": "Abstract. We present AvatarPopUp, a method for fast, high quality 3D human avatar generation from different input modalities, such as images and text prompts and with control over the generated pose and shape. The common theme is the use of diffusion-based image generation networks that are specialized for each particular task, followed by a 3D lifting network. We purposefully decouple the generation from the 3D modeling which allows us to leverage powerful image synthesis priors, trained on billions of text-image pairs. We fine-tune latent diffusion networks with additional image conditioning for image generation and back-view prediction, and to support qualitatively different multiple 3D hypotheses. Our partial fine-tuning approach allows to adapt the networks for each task without inducing catastrophic forgetting. In our experiments, we demonstrate that our method produces accurate, high-quality 3D avatars with diverse appearance that respect the multimodal text, image, and body control signals. Our approach can produce a 3D model in as few as 2 seconds, a four orders of magnitude speedup w.r.t. the vast majority of existing methods, most of which solve only a subset of our tasks, and with fewer controls. AvatarPopUp enables applications that require the controlled 3D generation of human avatars at scale. The project website can be found at https://www.nikoskolot.com/avatarpopulation/." + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.826, + 0.413, + 0.84 + ], + "angle": 0, + "content": "* Now at Google DeepMind." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.405, + 0.127 + ], + "angle": 0, + "content": "N. Kolotouros et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.375, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.176, + 0.786, + 0.447 + ], + "angle": 0, + "content": "We present AvatarPopUp, a method for instant generation of rigged full-body 3D human avatars, with multimodal controls in the form of text, images, and/or human pose and shape. The remarkable recent progress in image synthesis [15, 24, 54, 56, 58, 63] acted as a catalyst for a wide range of media generation applications. In just a few years, we have witnessed rapid developments in video generation [9, 23, 34, 68, 76], audio synthesis [49, 77] or text-to-3D object generation [40, 41, 51, 53, 65], among others. Pivotal to success of all these methods is their probabilistic nature, however this requires very large training sets. While inspiring efforts have been made [14], training set size is still a problem in many domains, and particularly for 3D. In an attempt to alleviate the need for massive 3D datasets, DreamFusion [51] leverages the rich priors of text-to-image diffusion models in an optimization framework. The influential DreamFusion ideas were also quickly adopted for 3D avatar creation [32, 37], a field previously dominated by image- or video-based reconstruction solutions [6, 8, 60, 61]. Text-to-avatar methods enabled novel creative processes, but came with a significant drawback. While image-based methods typically use pretrained feed-forward networks and create outputs in seconds, existing text-to-avatar solutions are optimization-based and take minutes to several hours to complete, per instance." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.449, + 0.786, + 0.523 + ], + "angle": 0, + "content": "In this paper, we are closing this gap and present, for the first time, a methodology for instant, text-controlled, rigged, full-body 3D human avatar creation. Our AvatarPopUp is purely feed-forward, can be conditioned on images and textual descriptions, allows fine-grained control over the generated body pose and shape, can generate multiple hypotheses, and runs in 2-10 seconds per instance." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.524, + 0.786, + 0.719 + ], + "angle": 0, + "content": "Key to success is our pragmatic decoupling of the two stages of probabilistic text-to-image generation and 3D lifting. Decoupling 2D generation and 3D lifting has two major advantages: (1) We can leverage the power of pretrained text-to-image generative networks, which have shown impressive results in modeling complex conditional distributions. Trained with large training sets of images, both generation quality and diversity are very high. (2) We alleviate the need for very large 3D datasets required to train state-of-the-art generative 3D models. Our method generates diverse plausible image configurations that contain rich enough information that can be lifted in 3D with minimal ambiguity. In other words, we distribute the workload between two expert systems: a pretrained probabilistic generation network fine-tuned for our task to produce front and back image views of the person, and a state-of-the-art unimodal, feed-forward image-to-3D model that can be trained using comparably small datasets." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.72, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Our proposed decoupling strategy allows us to maximally exploit available data sources with different levels of supervision. We first fine-tune a pretrained Latent Diffusion network to generate images of people based on textual descriptions and with additional control over the desired pose and shape. This step does not require any ground truth 3D data for supervision and enables scaling our image generator to web scale data of images of people in various poses. Next, we leverage a small-scale dataset of scanned 3D human assets and fine-tune a second latent diffusion network to learn the distribution of back side views con" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.477, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Instant 3D Human Avatar Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.145, + 0.787, + 0.238 + ], + "angle": 0, + "content": "
Control from Text Control from Image Shape Control
BodyShapeGeneratesGeometryTexture
✓✓✓✓✓✓/XRuntime
X✓X✓✓/XXseveral hoursOptimization-based text-to-3D [26,32,37]
✓X✓✓XXsecondsImage-to-3D [4,8,60,61]
XX✓✓✓XminutesHybrid [31]
✓✓✓✓✓✓seconds3D GANs [16,25]
secondsAvatarPopUp (Ours)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.242, + 0.784, + 0.311 + ], + "angle": 0, + "content": "Table 1: AvatarPopUp generates 3D assets with texture from text prompts or input images of a target subject and can be controlled with body pose and shape. In contrast to baselines that require up to hours per prompt, our model takes under five seconds and can de facto by used in interactive applications. In the experimental section we demonstrate the large diversity of the model, and applications in cloth editing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.785, + 0.526 + ], + "angle": 0, + "content": "ditioned on a front view image of the person. We optionally also condition on a textual description that can naturally complement the evidence available in the front view image. Furthermore, we propose a novel fine-tuning strategy that prevents overfitting to the new datasets. Finally, we design and train a 3D reconstruction network that predicts a textured 3D shape in the form of an implicit signed distance field given the pair of front and back views and optionally 3D body signals. The resulting cascaded method, which we call AvatarPopUp supports a wide range of 3D generation and reconstruction tasks: First, it enables fast and interactive 3D generation of assets at scale, see Fig. 1. Second, we can repurpose parts of the cascade for image-based 3D reconstruction at state-of-the-art quality. Finally, we demonstrate how AvatarPopUp can be used for creative editing tasks exemplified in 3D virtual try-on with body shape preservation. To summarize, our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.538, + 0.784, + 0.58 + ], + "angle": 0, + "content": "- A method for controllable 3D human avatar generation, based on multimodal text, pose, shape and image input signals, that outputs a detailed human mesh instance in 2-10 seconds." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.583, + 0.784, + 0.611 + ], + "angle": 0, + "content": "- We propose a simple yet effective way to fine-tune pretrained diffusion models on small-scale datasets, without inducing catastrophic forgetting." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.615, + 0.784, + 0.642 + ], + "angle": 0, + "content": "- While not our primary goal, our approach achieves state-of-the-art results in single-image 3D reconstruction and enables 3D creative editing applications." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.538, + 0.784, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.669, + 0.388, + 0.685 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.783, + 0.733 + ], + "angle": 0, + "content": "Table 1 summarizes the characteristics of AvatarPopUp in comparison to previous work, along several important property axes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Text-to-3D generation. The success of text-to-image models [55, 57, 59] was quickly followed by a significant amount of work on text-to-3D content generation [38, 45, 51, 66]. Due to limited training data, methods typically use optimization approaches, where a neural representation [47] is optimized per instance by minimizing a distillation loss [51] derived from large text-to-image models. This idea has been extended to generate human avatars [18, 27, 29, 32, 37, 69, 73, 78–80] or heads [19, 36, 39] enabling the text-based creation of 3D human assets that" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.405, + 0.128 + ], + "angle": 0, + "content": "N. Kolotouros et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": "are diverse in terms of shape, appearance, clothing and various accessories. In these works, the optimization process is often regularized using a 3D body model [32,37,78], which also enables animation. However, such approaches generally take hours per instance, and rendering is slow. With the appearance of Gaussian Splatting [30], other works [2,42,82] reduced rendering time at the expense of accurate geometry. In any case, creating an avatar still takes a significant amount of time, making such methods unsuitable for interactive applications. In this work we propose an alternative direction, which also builds upon the success of text-to-image models and combines them with 3D reconstruction pipelines. Also related are 3D human generation methods. AG3D [16] and EVA3D [25] are GAN-based methods learned from 2D data, that allow to sample 3D humans anchored in a 3D body model. CHUPA [31] generates dual normal maps based on text and then fits a body model to obtain a full 3D representation. While generation is similar in spirit to our method, CHUPA requires optimization per instance and does not generate texture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.788, + 0.653 + ], + "angle": 0, + "content": "Photorealistic 3D Human Reconstruction. Our framework generates 3D human assets and builds on top of state-of-the-art 3D reconstruction techniques. This has been widely explored in the past and can be roughly categorized by its use of explicit or implicit representations. An important line of work leverages 3D body models [43,50,72] and reconstructs their associated parameters, in some cases extended with vertex offsets to represent some clothing and hair detail [4-7, 48, 85]. Other efforts have considered voxels [67, 84], depth maps [17] and more recently implicit representations [8, 13, 21, 28, 60-62, 70, 71, 74, 75, 83]. Being topology free, the latter allows the representation of loose clothing more easily. They typically provide more detail and enable high-resolution reconstruction, often conditioned on local pixel-aligned features [60]. On the other hand, these methods yield reconstructions with no semantic labels that cannot be easily animated. To solve this problem, some work combined body models with implicit representations [13, 21, 27, 28, 70, 71], but this is prone to errors when the pose is noisy at inference time. In contrast, we drive the synthesis process with guidance from an input body model – sampled or estimated – so that the generated image is well aligned with the body prior. As we show in Sec. 4, this allows rigging the 3D avatar without post-processing and natively supports 3D animation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Given a single input image of a person, previous work aims to generate realistic reconstructions for the non-visible parts. However, this often leads to blurry textures and lack of geometric detail, e.g. no wrinkles. Some methods [60, 61, 70, 71] generate back normal maps to enhance details, or consider probabilistic reconstructions [3, 62]. However, all these methods cannot be prompted from text or other modalities and still yield limited diversity. In contrast, we guide the synthesis process by means of generated front and back images, yielding high-quality 3D reconstructions. Another challenge in previous work is limited training data. Most prior methods rely on a few hundred 3D scans, due to the pricey and laborious process of good quality human capture. We alleviate the need for large scale 3D training data by proposing a framework that can quickly generate humans with a given clothing, pose and shape." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.477, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Instant 3D Human Avatar Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.142, + 0.789, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.455 + ], + "angle": 0, + "content": "Fig. 2: AvatarPopUp method. (Top) AvatarPopUp builds on the capacity of text-to-image models to generate highly detailed and diverse input images. First, a Latent Diffusion network takes a text prompt and a target body pose and shape \\(\\mathcal{G}\\), and generates a highly detailed front image \\(I_{f}\\) of a person. Next, a second network generates a consistent back view \\(I_{b}\\) in the same pose and clothing. (Bottom) We perform pixel-aligned 3D reconstruction given the generated front/back views \\(I_{f}, I_{b}\\) and optionally the given 3D body pose and shape \\(\\mathcal{G}\\). This decoupling enables the generation of 3D avatars from text, images or a combination of the two." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.476, + 0.331, + 0.491 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.785, + 0.542 + ], + "angle": 0, + "content": "We learn a distribution \\( p(X|c) \\) of textured 3D shapes \\( X \\) conditioned on a collection of signals \\( c \\) that we factorize as follows" + }, + { + "type": "equation", + "bbox": [ + 0.302, + 0.554, + 0.787, + 0.587 + ], + "angle": 0, + "content": "\\[\np (X | c) = \\int \\int p \\left(X \\mid I _ {f}, I _ {b}, c\\right) \\cdot p \\left(I _ {b} \\mid I _ {f}, c\\right) \\cdot p \\left(I _ {f} \\mid c\\right) d I _ {f} d I _ {b}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.785, + 0.658 + ], + "angle": 0, + "content": "where \\( p(X|I_f, I_b, c) \\) is the probability of 3D shape \\( X \\) given \\( c \\) and front and back image observations \\( I_f \\) and \\( I_b \\) respectively, \\( p(I_b|I_f, c) \\) is the probability of the back view image given the front image \\( I_f \\) and conditioning signals \\( c \\), and \\( p(I_f|c) \\) the conditional probability of front view images of the person given \\( c \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Computing the integral in (1) is intractable, but our goal is to generate samples from the distribution rather than expectations. To do so, we employ ancestral sampling. We first sample a front view \\( I_{f} \\) given \\( c \\), we then sample a back view \\( I_{b} \\) given \\( I_{f} \\) and \\( c \\), and last we sample the 3D reconstruction based on the entire context. In practice, \\( p(I_f|c) \\) and \\( p(I_b|I_f,c) \\) are implemented using Latent Diffusion models, whereas \\( p(X|I_f,I_b,c) \\) is a unimodal, neural implicit field generator." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In the case of single-image 3D reconstruction the conditioning signal \\( c \\) is \\( I_{f} \\), and consequently we can omit the first step. For text-based generation, \\( c \\) is a text prompt describing the appearance of the person together with a signal encoding the body pose and shape. The conditioning information \\( c \\) may be extended with additional signals, as in the case of 3D editing, cf. Sec. 4." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.405, + 0.128 + ], + "angle": 0, + "content": "N. Kolotouros et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.577, + 0.162 + ], + "angle": 0, + "content": "3.1 Controllable Text-to-Image Generator" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.784, + 0.244 + ], + "angle": 0, + "content": "Recent advances in diffusion-based text-to-image generation networks [56] have enabled synthesizing high-quality images given only a text prompt as input. However, for certain use cases, such as human generation, it is difficult to inject fine-grained, inherently continuous forms of control, like the 3D pose of people or their precise body shape proportions, when generating with text alone." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.246, + 0.785, + 0.365 + ], + "angle": 0, + "content": "Inspired by ControlNet [81], we propose to add simultaneous control over body pose and shape by augmenting a pretrained Latent Diffusion network with an additional image input that jointly encodes both modalities. For control we use GHUM [72] but other models (e.g. [43]) can be used. Specifically given 3D pose and shape parameters \\(\\pmb{\\theta}\\) and \\(\\pmb{\\beta}\\) respectively, we render the corresponding mesh \\(\\mathcal{M} = \\mathrm{GHUM}(\\pmb{\\theta},\\pmb{\\beta})\\) using GHUM's template coordinates and posed vertex locations as 6D vertex colors, obtaining a dense, pixel-aligned pose- and shape-informed control signal \\(\\mathcal{G}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.785, + 0.577 + ], + "angle": 0, + "content": "To fine-tune the network, we generate a dataset of images of people with corresponding GHUM 3D pose and shape parameters and text annotations. This dataset is comprised of a set of scanned assets [1] that are rendered from different viewpoints, as well as a set of real images scraped from the web. For the synthetic part of the dataset, the pose and shape parameters are obtained by fitting GHUM to 3D scans. We additionally use real images to which we fit GHUM using keypoint optimization in the style of [33]. For all images we obtained text annotations using an off-the shelf image captioning system [12] by prompting it to describe the clothing of the people in the image. Since we are interested in generating 3D human assets, we additionally mask out the background in all images, and train the network to output segmented images. This makes the downstream 3D reconstruction task easier, and improves the reconstruction quality because it encourages the network to focus on human appearance, rather than allocating capacity to model complex backgrounds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.787, + 0.775 + ], + "angle": 0, + "content": "We want to exploit the rich priors learned by text-to-image foundation models by fine-tuning a Latent Diffusion model [56] with the dense GHUM rendering as an additional input. For fine-tuning we propose a simpler and more lightweight method than a standard ControlNet. We pad the weights of the input convolutional layer with additional channels initialized with zeros, and we then fine-tune only the weights of the convolutional layers of the encoder network. All the decoder and attention layers are kept frozen. With this simple strategy, even though our model is trained on a relatively small set of images, it can generalize to unseen types of clothing. At the same time, our strategy is more practical than training a ControlNet, which involves keeping in separate copy of the original network weights in memory, and thus enables fine-tuning large models with moderate hardware utilization. We optimize the encoder of the diffusion model by minimizing the simple variant of the diffusion loss [24]" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.785, + 0.785, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(\\psi_ {\\text {e n c}}\\right) = \\mathbb {E} _ {\\mathcal {E} (x), \\epsilon , t, \\tau , \\mathcal {G}} \\left\\| \\epsilon - \\epsilon_ {\\psi} \\left(z _ {t}, t, \\tau , \\mathcal {E} (\\mathcal {G})\\right) \\right\\|, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.786, + 0.842 + ], + "angle": 0, + "content": "where \\(t \\in \\{1, \\dots, T\\}\\) is the diffusion time step, \\(\\epsilon \\sim \\mathcal{N}(0, I)\\) the injected noise, \\(z_{t} = \\alpha_{t} \\mathcal{E}(x) + \\nu_{t}\\) is the noisy image latent, \\(\\tau\\) is the text encoding, \\(\\mathcal{E}(\\mathcal{G})\\) the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.477, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Instant 3D Human Avatar Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.15, + 0.332, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.151, + 0.457, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.151, + 0.587, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.151, + 0.716, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.441, + 0.248, + 0.564, + 0.259 + ], + "angle": 0, + "content": "No text conditioning" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.26, + 0.332, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.26, + 0.461, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.26, + 0.59, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.354, + 0.605, + 0.365 + ], + "angle": 0, + "content": "+“gray hair”" + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.26, + 0.719, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.354, + 0.746, + 0.365 + ], + "angle": 0, + "content": "+“with a tattoo”" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.37, + 0.788, + 0.426 + ], + "angle": 0, + "content": "Fig. 3: Diverse back view hypotheses. Conditioned on the front view, our method is able to generate diverse plausible back views of the person, with different hairstyles, wrinkle patterns, or lighting. Our network can also be controlled with text (second row), to add fine-grained detail to our generated back-side views." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.479 + ], + "angle": 0, + "content": "latent encoding of the dense GHUM signal, and \\(\\psi_{\\mathrm{enc}}\\) the encoder subset of the denoising UNet parameters \\(\\psi\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.501, + 0.45, + 0.515 + ], + "angle": 0, + "content": "3.2 Back View Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.789, + 0.735 + ], + "angle": 0, + "content": "One could try to lift the generated front views from the previous stage to 3D directly by applying a single-image 3D reconstruction method like PHORHUM [8]. However, because of the inherent ambiguity of the problem this will result in significant loss of geometric detail and blurry textures for the non-visible body surfaces. To avoid this, we propose to fine-tune again a latent diffusion network with the same strategy as in the previous section. This time the additional image conditioning is a front view and optionally a text prompt, and we train the network to learn the distribution of back views conditioned on the front view. The additional text prompt can be used in cases where it is desired to additionally guide the generation by very specific properties. Fig. 3 shows different back sides sampled from the conditional distribution. We also show that the additional text inputs are useful in modulating certain parts of the generation that are not immediately deducible from the front image, such as hairstyles or specific patterns." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.758, + 0.478, + 0.772 + ], + "angle": 0, + "content": "3.3 3D Reconstruction Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Our 3D reconstruction network is inspired by PHORHUM [8], and our design choices are informed by the limitations of typical single-image 3D reconstruction methods. Specifically, given a collection of input image signals \\(\\mathcal{I} = \\{I_f, I_b, \\mathcal{G}\\}\\), we first concatenate them, and then use a convolutional encoder \\(G\\) to compute" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.405, + 0.128 + ], + "angle": 0, + "content": "N. Kolotouros et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "a pixel-aligned feature map \\( G(\\mathcal{I}) \\). The 3D body control signal \\( \\mathcal{G} \\) is optional and may be omitted, e.g. for single-image reconstruction. Then, each point \\( \\mathbf{x} \\in \\mathbb{R}^3 \\) in the scene gets projected on this feature map to get pixel-aligned features \\( \\mathbf{z}_{\\mathbf{x}} = g(\\mathcal{I}, \\mathbf{x}; \\pi) = b(G(\\mathcal{I}), \\pi(\\mathbf{x})) \\) using interpolation, where \\( b(\\cdot) \\) is the bilinear sampling operator and \\( \\pi(\\mathbf{x}) \\) is the pixel location of the projection of \\( \\mathbf{x} \\) using the camera \\( \\pi \\). These pixel aligned features are then concatenated with a positional encoding \\( \\gamma(\\mathbf{x}) \\) of the 3D point and are fed to an MLP \\( f \\) that outputs the signed distance from the surface \\( d \\) as well as surface color \\( c \\). Finally, the 3D shape \\( S \\) is represented as the zero-level-set of \\( d \\)" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.294, + 0.786, + 0.311 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} (\\mathcal {I}) = \\left\\{\\mathbf {x} \\in \\mathbb {R} ^ {3} | f (g (I, \\mathbf {x}; \\pi), \\gamma (\\mathbf {x})) = (0, c) \\right\\}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.32, + 0.714, + 0.336 + ], + "angle": 0, + "content": "\\(\\mathcal{S}\\) can be transformed to a mesh directly, using Marching Cubes [44]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.357, + 0.533, + 0.372 + ], + "angle": 0, + "content": "3.4 Animation of Generated Avatars" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.789, + 0.515 + ], + "angle": 0, + "content": "Our method can generate diverse 3D avatars with various poses, shapes and appearances. Optionally, we may leverage the conditioning body model to rig the estimated 3D shape. As a result of our conditioning strategy, 3D avatars and the conditional body model instances are aligned in 3D. This allows us to anchor the reconstructed 3D shape on the body model surface [10] and re-pose or re-shape it accordingly. Alternatively, we may also transfer just the LBS skeleton and weights from the body model to the scan. This enables importing and animating the generated 3D assets in various rendering engines. See Fig. 8 and Sup. Mat. for examples and videos of animations of the generated 3D assets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.537, + 0.377, + 0.554 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.688 + ], + "angle": 0, + "content": "Data. We use meshes from RenderPeople [1] for training as well as our own captured data, totaling in \\(\\sim 10\\mathrm{K}\\) scans with diverse poses, body shapes, and clothing styles. We render each scan with randomly sampled HDRI background, random cloth color augmentations, and lighting using Blender [11]. During this process, we render both front and back views used to train the different stages of our model. For the front image generation network we also use a set of \\(10\\mathrm{K}\\) real images on which we fitted the GHUM model using 2D keypoints. For testing we defined a split based on subject identity and held out \\(\\sim 1\\mathrm{K}\\) scans." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.779 + ], + "angle": 0, + "content": "We provide results for 2 different versions of our model. The standard quality model is generated in 2 seconds by running 5 DDIM [64] steps during inference and Marching Cubes at \\(256^{3}\\) resolution. The high quality model is generated in 10 seconds using 50 DDIM steps and Marching Cubes at \\(512^{3}\\) resolution. The timings were recorded on a single 40 GB A100 GPU. Unless otherwise stated, all results we report are obtained using the high quality model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Metrics and Baselines. We compare our method numerically for two different problems. First we consider the task of text-to-3D human generation, where we sample 100 different text prompts and compare against representative text-to-3D generation methods. For numerical comparisons we evaluate the text-image" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.477, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Instant 3D Human Avatar Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.146, + 0.756, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.324, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Fig. 4: Diversity of our 3D generation. For the same text prompt and the same pose and shape conditioning, our model can generate a diverse set of 3D avatars that respect both the text and the 3D body controls." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.785, + 0.422 + ], + "angle": 0, + "content": "alignment using CLIP [52]. Specifically, we use the retrieval accuracy using CLIP, as proposed in [51]. We additionally show qualitative results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.789, + 0.635 + ], + "angle": 0, + "content": "Second, we validate the performance of our 3D reconstruction component against state-of-the-art methods [8,27,60,61,70,71] considering both geometry and texture. Pixel-aligned image features dominate in recent work [8,60,61], but some methods proposed to combine them with body models [70,71], which offer advantages in being able to animate reconstructions. We also rely on pixel-aligned features, yet propose a method that inherently enables animation. We run AvatarPopUp by generating the back side of the subject and applying the reconstruction network. To evaluate 3D geometry, we report bi-directional Chamfer distance \\(\\times 10^{-3}\\), Normal Consistency (NC ↑), and Volumetric Intersection over Union (IoU ↑) after ICP alignment. However, these metrics do not necessarily correlate with good visual quality, e.g. Chamfer distance is minimized by smooth, non-detailed geometry. To measure the quality of reconstructions, we additionally report FID scores [22] of the front/back views for both geometry and texture." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.668, + 0.452, + 0.683 + ], + "angle": 0, + "content": "4.1 Text-to-3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.702, + 0.787, + 0.748 + ], + "angle": 0, + "content": "In Fig. 4 we generate different avatars given the same text prompt and driving poses. We can see that our model is able to create a very diverse set of assets, a property not observed in the previous text-to-3D generation methods [32]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In Tab. 2 we use CLIP to evaluate our model against other text-to-3D generation methods. In general, CLIP-based metrics are not indicative of the generated image quality, because they only consider the alignment with the text, and often over-saturated images with extreme details tend to have high CLIP scores. To further demonstrate that our method generates higher quality avatars, we also include a qualitative comparison in Fig. 5." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.405, + 0.127 + ], + "angle": 0, + "content": "N. Kolotouros et al." + }, + { + "type": "table", + "bbox": [ + 0.303, + 0.134, + 0.696, + 0.223 + ], + "angle": 0, + "content": "
ColorColorRuntime ↓
R-Prec. ↑Top-3 ↑R-Prec. ↑Top-3 ↑
0.680.920.040.258hDreamHuman [32]
0.560.820.030.153hTADA [37]
--0.030.083mCHUPA [31]
0.580.730.080.17~2sOurs
0.620.770.110.17~10sOurs (high quality)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.228, + 0.787, + 0.27 + ], + "angle": 0, + "content": "Table 2: Numerical comparisons with other text-to-3d human generation methods. We mark the best and second best results. Our method allows to trade-off speed and quality and we report results for two different settings." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.281, + 0.761, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.483, + 0.692, + 0.496 + ], + "angle": 0, + "content": "A person wearing a black off-shoulder top and a denim skirt" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.499, + 0.788, + 0.555 + ], + "angle": 0, + "content": "Fig. 5: Comparisons with text-to-3d human generation methods. Our method generates high quality results that respect the text prompt well, at a fraction of the others' runtime, cf. Tab. 3. TADA's results appear unnatural; DreamHuman failed for one subject and produces oversaturated colors; CHUPA failed to respect the prompt." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.584, + 0.527, + 0.599 + ], + "angle": 0, + "content": "4.2 Single-image 3D Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.84 + ], + "angle": 0, + "content": "While not specifically designed for 3D reconstruction, our method achieves state-of-the-art performance also for this task. The evaluation setup is the following: given an input image \\(I\\), we draw one random sample from the back view image generator network, and then feed the pair of front/back images to our 3D reconstruction network. For all methods we extract a textured 3D mesh using Marching Cubes and report numerical results in Tab. 3. Furthermore, we show qualitative results in Fig. 6. Notably, our method not only performs on par numerically and qualitatively on reconstructed front views, but also generates highly detailed back view texture and geometry. Finally, we also compare with the optimization-based method TeCH [27] in Fig. 7. TeCH produces detailed front and back geometry but also also exhibits problems at times, rooted in its 3D pose estimation method. Most importantly, TeCH runs for several hours per instance, while ours computes results in a single feed-forward pass, in only a few seconds." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.477, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Instant 3D Human Avatar Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.261, + 0.144, + 0.737, + 0.244 + ], + "angle": 0, + "content": "
3D MetricsFID ↓
Ch. ↓IoU ↑NC ↑Color. FColor B.Normal F.Normal B.
7.10.500.72--29.657.2PIFu [60]
6.20.520.74--26.952.5PIFuHD [61]
7.00.480.72--36.149.0ICON [71]
3.40.580.76--18.336.1ECON [70]
0.840.710.8711.040.811.238.8PHORHUM* [8]
0.910.700.8611.020.012.825.8Ours
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.25, + 0.788, + 0.333 + ], + "angle": 0, + "content": "Table 3: Numerical comparisons with single-view 3D reconstructions methods and ablations of our method. We mark the best and second best results. All Chamfer metrics are \\(\\times 10^{-3}\\). Not all methods generate colors. For fair comparisons, we retrained PHORHUM [8] using the same data as AvatarPopUp. We observe comparable results in terms of 3D metrics, however ours performs better on generating more realistic and diverse back views and back normals." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.351, + 0.419, + 0.367 + ], + "angle": 0, + "content": "4.3 3D Virtual Try On" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.789, + 0.62 + ], + "angle": 0, + "content": "An immediate application of our method is the option to perform 3D garment edits for a given identity. Given an input image, we first recover 3D pose and shape parameters of the person in the image. Using a text prompt and the identity preservation strategy introduced below, we can generate updated images of the same person wearing different garments or accessories. To preserve the identity of the person, we first locate the person's head in the source image and use then Repaint [46] to out-paint a novel body for the given head while still conditioning on the estimated pose and shape parameters. With this strategy we ensure that our method not only preserves the facial characteristics but also the overall body proportions. In Fig. 9 we illustrate such editing examples. The generated 3D edits present garment details like wrinkles on both front and back views, and preserve the subjects' facial appearance. Also, note that body shape and identity are well preserved for the subjects, even though only one image is given. While there has been a significant amount of 2D virtual try-on research [20,35,86], our methodology can generate consistent and highly detailed 3D meshes that can be animated and rendered from other viewpoints." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.643, + 0.352, + 0.657 + ], + "angle": 0, + "content": "4.4 Animation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.669, + 0.787, + 0.715 + ], + "angle": 0, + "content": "AvatarPopUp enables animation of the generated assets by design, provided that its generation is conditioned on an underlying body model. In Fig. 8, we show an example of a generated avatar that is rigged automatically." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.738, + 0.39, + 0.753 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Pose and shape encoding in the 3D reconstruction network. We compare the effectiveness of our additional pose and shape encoding inputs \\(\\mathcal{G}\\) to our reconstruction network. To do so, we use the same set of 100 text prompts as before and for each text prompt sample a random pose and shape configuration. For each \\((\\tau, \\theta, \\beta)\\) triplet we run inferences and compute 2 meshes: one using" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.405, + 0.127 + ], + "angle": 0, + "content": "N. Kolotouros et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.135, + 0.264, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.219, + 0.256, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.3, + 0.255, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.137, + 0.461, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.219, + 0.324, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.3, + 0.323, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.219, + 0.409, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.3, + 0.406, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.353, + 0.378, + 0.386, + 0.389 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.137, + 0.461, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.219, + 0.453, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.3, + 0.453, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.134, + 0.607, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.217, + 0.609, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.299, + 0.607, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.377, + 0.613, + 0.388 + ], + "angle": 0, + "content": "PHORHUM" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.135, + 0.665, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.218, + 0.641, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.3, + 0.645, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.613, + 0.377, + 0.643, + 0.388 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.136, + 0.782, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.219, + 0.729, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.219, + 0.778, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.3, + 0.746, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.71, + 0.378, + 0.766, + 0.388 + ], + "angle": 0, + "content": "PIFuHD" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.398, + 0.787, + 0.44 + ], + "angle": 0, + "content": "Fig. 6: Qualitative comparisons with state-of-the-art single image 3D reconstruction methods. Our method produces front color and normals on par with state-of-the-art and much more detailed back view hypotheses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.785, + 0.552 + ], + "angle": 0, + "content": "only the front and back images, and another one additionally using the dense GHUM encodings. We then evaluate the Chamfer distance between each mesh and the corresponding GHUM mesh. The model using the additional GHUM signals has an average Chamfer distance of \\( d_{\\mathrm{with}} = 1.4 \\) whereas the one without \\( d_{\\mathrm{without}} = 8.6 \\), thus validating our design choice. Not only is the control respected well, but this also allows for animation as discussed previously." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.787, + 0.66 + ], + "angle": 0, + "content": "Partial vs. Complete Network Finetuning. We finetune two Latent Diffusion networks, one in a standard way by optimizing all parameters, and another one using our proposed strategy where we only finetune the convolutional layers of the encoder. Empirically we observed that the network that was finetuned as a whole experienced catastrophic forgetting, and has poor performance when asked to generate types of garment not seen in the training set. Fig. 10 shows a comparison for text prompts not in the training set." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.684, + 0.5, + 0.7 + ], + "angle": 0, + "content": "5 Discussion & Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.785, + 0.779 + ], + "angle": 0, + "content": "Limitations. AvatarPopUp inherits limitations from other pixel-aligned methods, e.g. with regions parallel to camera rays being not as detailed in the reconstruction. Further, artifacts might be visible in poses or very loose clothing that are underrepresented during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Ethical Considerations. We present a generative tool to create 3D human assets, thus reducing the risk of scanning and using real humans for training large scale 3D generative models. AvatarPopUp generates diverse results and can lead to a better coverage of subject distributions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.477, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Instant 3D Human Avatar Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.166, + 0.273, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.167, + 0.532, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.556, + 0.447, + 0.567 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.167, + 0.781, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.555, + 0.706, + 0.567 + ], + "angle": 0, + "content": "TeCH" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.572, + 0.788, + 0.614 + ], + "angle": 0, + "content": "Fig. 7: Additional comparisons with TeCH. TeCH is optimization-based and takes several hours to complete for a single reconstruction. Our results are obtained in seconds, yet are as detailed and even less noisy." + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.674, + 0.691, + 0.78 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.788, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Fig. 8: Reposing example.. We first reconstruct an avatar \"wearing a gray suit\" in the A-pose and then we transfer it to different poses." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.405, + 0.127 + ], + "angle": 0, + "content": "N. Kolotouros et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.135, + 0.761, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.455 + ], + "angle": 0, + "content": "Fig. 9: Identity preserving 3D avatar editing. Our method allows for editing the clothing, while preserving the identity of the generated person. In each row we're using the image on the left as input, recover the person's pose and shape parameters and then run our method using the shown prompts. The generated avatars share the same identity and respect the target prompts." + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.465, + 0.718, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.625 + ], + "angle": 0, + "content": "Fig. 10: Partial vs. complete fine-tuning strategy. For each text prompt the first example is generated using our partial fine-tuning strategy whereas the second by fine-tuning the entire network. Empirically, partial fine-tuning is more resilient to catastrophic forgetting, generating more diverse and better text-aligned images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.842 + ], + "angle": 0, + "content": "We have presented AvatarPopUp, a novel framework for generating 3D human avatars controlled by text or images and yielding rigged 3D models in 2-10 seconds. AvatarPopUp is purely feed-forward, allows for fine-grained control over the generated body pose and shape, and can produce multiple qualitatively different hypotheses. AvatarPopUp is composed of a cascade of expert systems, decoupling image generation and 3D lifting. Through this design choice, AvatarPopUp benefits from both web-scale image datasets, ensuring high generation diversity, and from smaller size accurate 3D datasets, resulting in reconstructions with increased detail and precisely controlled based on text and identity specifications. In the future, we would like to explore other 3D construction strategies besides pixel-aligned features. Longer term, we aim to support highly detailed and controllable 3D human model generation for entertainment, education, architecture and art, or medical applications." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.477, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Instant 3D Human Avatar Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.174, + 0.441, + 0.187 + ], + "angle": 0, + "content": "1. https://renderpeople.com/" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.187, + 0.787, + 0.227 + ], + "angle": 0, + "content": "2. Abdal, R., Yifan, W., Shi, Z., Xu, Y., Po, R., Kuang, Z., Chen, Q., Yeung, D.Y., Wetzstein, G.: Gaussian shell maps for efficient 3d human generation. arXiv preprint arXiv:2311.17857 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.228, + 0.786, + 0.255 + ], + "angle": 0, + "content": "3. AlBahar, B., Saito, S., Tseng, H.Y., Kim, C., Kopf, J., Huang, J.B.: Single-image 3d human digitization with shape-guided diffusion. In: SIGGRAPH Asia (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.256, + 0.786, + 0.282 + ], + "angle": 0, + "content": "4. Alldieck, T., Magnor, M., Bhatnagar, B.L., Theobalt, C., Pons-Moll, G.: Learning to reconstruct people in clothing from a single rgb camera. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.786, + 0.309 + ], + "angle": 0, + "content": "5. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Detailed human avatars from monocular video. In: 3DV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.786, + 0.336 + ], + "angle": 0, + "content": "6. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Video based reconstruction of 3D people models. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.337, + 0.786, + 0.363 + ], + "angle": 0, + "content": "7. Alldieck, T., Pons-Moll, G., Theobalt, C., Magnor, M.: Tex2shape: Detailed full human body geometry from a single image. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.364, + 0.786, + 0.39 + ], + "angle": 0, + "content": "8. Alldieck, T., Zanfir, M., Sminchisescu, C.: Photorealistic monocular 3D reconstruction of humans wearing clothing. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.786, + 0.431 + ], + "angle": 0, + "content": "9. Bar-Tal, O., Chefer, H., Tov, O., Herrmann, C., Paiss, R., Zada, S., Ephrat, A., Hur, J., Li, Y., Michaeli, T., et al.: Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.432, + 0.786, + 0.471 + ], + "angle": 0, + "content": "0. Bazavan, E.G., Zanfir, A., Zanfir, M., Freeman, W.T., Sukthankar, R., Sminchisescu, C.: Hspace: Synthetic parametric humans animated in complex environments. arXiv (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.472, + 0.786, + 0.512 + ], + "angle": 0, + "content": "1. Blender Online Community: Blender - a 3D modelling and rendering package. Blender Foundation, Blender Institute, Amsterdam (2020), http://www.blender.org" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.513, + 0.787, + 0.609 + ], + "angle": 0, + "content": "12. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., Kolesnikov, A., Puigcerver, J., Ding, N., Rong, K., Akbari, H., Mishra, G., Xue, L., Thapliyal, A., Bradbury, J., Kuo, W., Seyedhosseini, M., Jia, C., Ayan, B.K., Riquelme, C., Steiner, A., Angelova, A., Zhai, X., Houlsby, N., Soricut, R.: Pali: A jointly-scaled multilingual language-image model (2022). https://doi.org/10.48550/ARXIV.2209.06794, https://arxiv.org/abs/2209.06794" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.61, + 0.787, + 0.649 + ], + "angle": 0, + "content": "13. Corona, E., Zanfir, M., Alldieck, T., Bazavan, E.G., Zanfir, A., Sminchisescu, C.: Structured 3d features for reconstructing relightable and animatable avatars. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.65, + 0.787, + 0.704 + ], + "angle": 0, + "content": "14. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3d objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13142-13153 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.705, + 0.787, + 0.731 + ], + "angle": 0, + "content": "15. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.732, + 0.787, + 0.772 + ], + "angle": 0, + "content": "16. Dong, Z., Xu Chen, J.Y., J.Black, M., Hilliges, O., Geiger, A.: AG3D: Learning to generate 3D avatars from 2D image collections. In: International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.773, + 0.787, + 0.8 + ], + "angle": 0, + "content": "17. Gabeur, V., Franco, J.S., Martin, X., Schmid, C., Rogez, G.: Moulding humans: Non-parametric 3D human shape estimation from single images. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "18. Gong, C., Dai, Y., Li, R., Bao, A., Li, J., Yang, J., Zhang, Y., Li, X.: Text2 avatar: Text to 3d human avatar generation with codebook-driven body controllable attribute. arXiv preprint arXiv:2401.00711 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.174, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.405, + 0.127 + ], + "angle": 0, + "content": "N. Kolotouros et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "19. Han, X., Cao, Y., Han, K., Zhu, X., Deng, J., Song, Y.Z., Xiang, T., Wong, K.Y.K.: Headsculpt: Crafting 3d head avatars with text. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "20. Han, X., Wu, Z., Wu, Z., Yu, R., Davis, L.S.: Viton: An image-based virtual try-on network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7543-7552 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "21. He, T., Xu, Y., Saito, S., Soatto, S., Tung, T.: Arch++: Animation-ready clothed human reconstruction revisited. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "22. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "23. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.37 + ], + "angle": 0, + "content": "24. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.412 + ], + "angle": 0, + "content": "25. Hong, F., Chen, Z., LAN, Y., Pan, L., Liu, Z.: EVA3d: Compositional 3d human generation from 2d image collections. In: International Conference on Learning Representations (2023), https://openreview.net/forum?id=g7U9jD_2CUr" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.453 + ], + "angle": 0, + "content": "26. Hong, F., Zhang, M., Pan, L., Cai, Z., Yang, L., Liu, Z.: Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. ACM Transactions on Graphics (TOG) 41(4), 1-19 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.495 + ], + "angle": 0, + "content": "27. Huang, Y., Yi, H., Xiu, Y., Liao, T., Tang, J., Cai, D., Thies, J.: TeCH: Text-guided Reconstruction of Lifelike Clothed Humans. In: International Conference on 3D Vision (3DV) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "28. Huang, Z., Xu, Y., Lassner, C., Li, H., Tung, T.: Arch: Animatable reconstruction of clothed humans. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "29. Jiang, R., Wang, C., Zhang, J., Chai, M., He, M., Chen, D., Liao, J.: Avatarcraft: Transforming text into neural human avatars with parameterized shape and pose control. arXiv preprint arXiv:2303.17606 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.592 + ], + "angle": 0, + "content": "30. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.633 + ], + "angle": 0, + "content": "31. Kim, B., Kwon, P., Lee, K., Lee, M., Han, S., Kim, D., Joo, H.: Chupa: Carving 3d clothed humans from skinned shape priors using 2d diffusion probabilistic models. arXiv preprint arXiv:2305.11870 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.675 + ], + "angle": 0, + "content": "32. Kolotouros, N., Alldieck, T., Zanfir, A., Bazavan, E., Fieraru, M., Sminchisescu, C.: Dreamhuman: Animatable 3d avatars from text. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.716 + ], + "angle": 0, + "content": "33. Kolotouros, N., Pavlakos, G., Black, M.J., Daniilidis, K.: Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2252-2261 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.758 + ], + "angle": 0, + "content": "34. Kondratyuk, D., Yu, L., Gu, X., Lezama, J., Huang, J., Hornung, R., Adam, H., Akbari, H., Alon, Y., Birodkar, V., et al.: Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "35. Lee, S., Gu, G., Park, S., Choi, S., Choo, J.: High-resolution virtual try-on with misalignment and occlusion-handled conditions. In: European Conference on Computer Vision. pp. 204-219. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "36. Lei, B., Yu, K., Feng, M., Cui, M., Xie, X.: Diffusiongan3d: Boosting text-guided 3d generation and domain adaption by combining 3d gans and diffusion priors. arXiv preprint arXiv:2312.16837 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.477, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Instant 3D Human Avatar Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "37. Liao, T., Yi, H., Xiu, Y., Tang, J., Huang, Y., Thies, J., Black, M.J.: Tada! text to animatable digital avatars. In: 3DV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.231 + ], + "angle": 0, + "content": "38. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.272 + ], + "angle": 0, + "content": "39. Liu, H., Wang, X., Wan, Z., Shen, Y., Song, Y., Liao, J., Chen, Q.: Headartist: Text-conditioned 3d head generation with self score distillation. arXiv preprint arXiv:2312.07539 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.314 + ], + "angle": 0, + "content": "40. Liu, M., Xu, C., Jin, H., Chen, L., Varma T, M., Xu, Z., Su, H.: One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.355 + ], + "angle": 0, + "content": "41. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9298-9309 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.397 + ], + "angle": 0, + "content": "42. Liu, X., Zhan, X., Tang, J., Shan, Y., Zeng, G., Lin, D., Liu, X., Liu, Z.: Human-gaussian: Text-driven 3d human generation with gaussian splatting. arXiv preprint arXiv:2311.17061 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.424 + ], + "angle": 0, + "content": "43. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ToG (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.452 + ], + "angle": 0, + "content": "44. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3D surface construction algorithm. SIGGRAPH (1987)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.452, + 0.788, + 0.493 + ], + "angle": 0, + "content": "45. Lorraine, J., Xie, K., Zeng, X., Lin, C.H., Takikawa, T., Sharp, N., Lin, T.Y., Liu, M.Y., Fidler, S., Lucas, J.: Att3d: Amortized text-to-3d object synthesis. arXiv preprint arXiv:2306.07349 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.549 + ], + "angle": 0, + "content": "46. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11461-11471 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.591 + ], + "angle": 0, + "content": "47. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.591, + 0.788, + 0.632 + ], + "angle": 0, + "content": "48. Onizuka, H., Hayirci, Z., Thomas, D., Sugimoto, A., Uchiyama, H., Taniguchi, R.i.: Tetratsdf: 3D human reconstruction from a single image with a tetrahedral outer shell. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.632, + 0.788, + 0.673 + ], + "angle": 0, + "content": "49. Oord, A.v.d., Dieleman, S., Zen, H., Simonyan, K., Vinyals, O., Graves, A., Kalchbrenner, N., Senior, A., Kavukcuoglu, K.: Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.715 + ], + "angle": 0, + "content": "50. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.715, + 0.788, + 0.743 + ], + "angle": 0, + "content": "51. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: Int. Conf. Learn. Represent. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.743, + 0.788, + 0.798 + ], + "angle": 0, + "content": "52. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.84 + ], + "angle": 0, + "content": "53. Raj, A., Kaza, S., Poole, B., Niemeyer, M., Mildenhall, B., Ruiz, N., Zada, S., Aberman, K., Rubenstein, M., Barron, J., Li, Y., Jampani, V.: Dreambooth3d: Subject-driven text-to-3d generation. ICCV (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.405, + 0.127 + ], + "angle": 0, + "content": "N. Kolotouros et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "54. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "55. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.273 + ], + "angle": 0, + "content": "56. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "57. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.37 + ], + "angle": 0, + "content": "58. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS. vol. 35, pp. 36479-36494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.424 + ], + "angle": 0, + "content": "59. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.467 + ], + "angle": 0, + "content": "60. Saito, S., Huang, Z., Natsume, R., Morishima, S., Kanazawa, A., Li, H.: Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "61. Saito, S., Simon, T., Saragih, J., Joo, H.: Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3D human digitization. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.535 + ], + "angle": 0, + "content": "62. Sengupta, A., Alldieck, T., Kolotouros, N., Corona, E., Zanfir, A., Sminchisescu, C.: Diffhuman: Probabilistic photorealistic 3D reconstruction of humans. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "63. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "64. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: Int. Conf. Learn. Represent. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "65. Sun, J., Zhang, B., Shao, R., Wang, L., Liu, W., Xie, Z., Liu, Y.: Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior. arXiv preprint arXiv:2310.16818 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "66. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "67. Varol, G., Ceylan, D., Russell, B., Yang, J., Yumer, E., Laptev, I., Schmid, C.: Bodynet: Volumetric inference of 3D human body shapes. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.743 + ], + "angle": 0, + "content": "68. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "69. Wang, J., Liu, Y., Dou, Z., Yu, Z., Liang, Y., Li, X., Wang, W., Xie, R., Song, L.: Disentangled clothed avatar generation from text descriptions. arXiv preprint arXiv:2312.05295 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.813 + ], + "angle": 0, + "content": "70. Xiu, Y., Yang, J., Cao, X., Tzionas, D., Black, M.J.: ECON: Explicit Clothed humans Optimized via Normal integration. In: CVPR (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "71. Xiu, Y., Yang, J., Tzionas, D., Black, M.J.: Icon: Implicit clothed humans obtained from normals. In: CVPR (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.477, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Instant 3D Human Avatar Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "72. Xu, H., Bazavan, E.G., Zanfir, A., Freeman, W.T., Sukthankar, R., Sminchisescu, C.: Ghum & ghuml: Generative 3D human shape and articulated pose models. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "73. Xu, Y., Yang, Z., Yang, Y.: Seeavatar: Photorealistic text-to-3d avatar generation with constrained geometry and appearance. arXiv preprint arXiv:2312.08889 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.787, + 0.273 + ], + "angle": 0, + "content": "74. Yang, Z., Wang, S., Manivasagam, S., Huang, Z., Ma, W.C., Yan, X., Yumer, E., Urtasun, R.: S3: Neural shape, skeleton, and skinning fields for 3D human modeling. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "75. Yanting Chan, K., Lin, G., Zhao, H., Lin, W.: Integratedpifu: Integrated pixel aligned implicit function for single-view human reconstruction. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.787, + 0.328 + ], + "angle": 0, + "content": "76. Zeng, Y., Wei, G., Zheng, J., Zou, J., Wei, Y., Zhang, Y., Li, H.: Make pixels dance: High-dynamic video generation. arXiv preprint arXiv:2311.10982 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "77. Zhang, C., Zhang, C., Zheng, S., Zhang, M., Qamar, M., Bae, S.H., Kweon, I.S.: Audio diffusion model for speech synthesis: A survey on text to speech and speech enhancement in generative ai. arXiv preprint arXiv:2303.13336 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.371, + 0.787, + 0.411 + ], + "angle": 0, + "content": "78. Zhang, H., Feng, Y., Kulits, P., Wen, Y., Thies, J., Black, M.J.: Text-guided generation and editing of compositional 3d avatars. arXiv preprint arXiv:2309.07125 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.411, + 0.787, + 0.453 + ], + "angle": 0, + "content": "79. Zhang, H., Chen, B., Yang, H., Qu, L., Wang, X., Chen, L., Long, C., Zhu, F., Du, K., Zheng, M.: Avatarverse: High-quality & stable 3d avatar creation from text and pose. arXiv preprint arXiv:2308.03610 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.453, + 0.787, + 0.495 + ], + "angle": 0, + "content": "80. Zhang, J., Zhang, X., Zhang, H., Liew, J.H., Zhang, C., Yang, Y., Feng, J.: Avatarstudio: High-fidelity and animatable 3d avatar creation from text. arXiv preprint arXiv:2311.17917 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.495, + 0.787, + 0.522 + ], + "angle": 0, + "content": "81. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. IEEE International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.787, + 0.564 + ], + "angle": 0, + "content": "82. Zhao, Z., Bao, Z., Li, Q., Qiu, G., Liu, K.: Psavatar: A point-based morphable shape model for real-time head avatar creation with 3d gaussian splatting. arXiv preprint arXiv:2401.12900 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.787, + 0.592 + ], + "angle": 0, + "content": "83. Zheng, Z., Yu, T., Liu, Y., Dai, Q.: Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. PAMI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.592, + 0.787, + 0.619 + ], + "angle": 0, + "content": "84. Zheng, Z., Yu, T., Wei, Y., Dai, Q., Liu, Y.: Deephuman: 3D human reconstruction from a single image. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.619, + 0.787, + 0.647 + ], + "angle": 0, + "content": "85. Zhu, H., Zuo, X., Wang, S., Cao, X., Yang, R.: Detailed human shape estimation from a single image by hierarchical mesh deformation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.647, + 0.787, + 0.703 + ], + "angle": 0, + "content": "86. Zhu, L., Yang, D., Zhu, T., Reda, F., Chan, W., Sahara, C., Norouzi, M., Kemelmacher-Shlizerman, I.: Tryondiffusion: A tale of two unets. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4606-4615 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.703 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Instant 3D Human Avatar Generation using Image Diffusion Models/848a10c5-4ff1-45f7-8010-a0858cf611ea_origin.pdf b/2024/Instant 3D Human Avatar Generation using Image Diffusion Models/848a10c5-4ff1-45f7-8010-a0858cf611ea_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4599d1f3fef5471eabcf291c0e35adc837b3b3ec --- /dev/null +++ b/2024/Instant 3D Human Avatar Generation using Image Diffusion Models/848a10c5-4ff1-45f7-8010-a0858cf611ea_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ef044226253d6ae22d568099f6a939e59777ac5bcd47d1d8e07dd5b24b58967 +size 26430709 diff --git a/2024/Instant 3D Human Avatar Generation using Image Diffusion Models/full.md b/2024/Instant 3D Human Avatar Generation using Image Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e4af6f7fab1d8aa26bebda2f9c1d796f3ce956c6 --- /dev/null +++ b/2024/Instant 3D Human Avatar Generation using Image Diffusion Models/full.md @@ -0,0 +1,333 @@ +# Instant 3D Human Avatar Generation using Image Diffusion Models + +Nikos Kolotouros, Thiemo Alldieck, Enric Corona, Eduard Gabriel Bazavan, and Cristian Sminchisescu + +Google Research\* {kolotouros, alldieck, egbazavan, enriccorona, sminchisescu}@google.com + +![](images/13947c3f92d8dc40fa114453e17f43421b851573c29a2de60d376a787ccfaa73.jpg) +Fig. 1: We present AvatarPopUp, a new method for the automatic generation of 3D human assets. AvatarPopUp can generate rigged 3D models from text or from single images and has control over body pose and shape. In this example, we show 77 models generated from various text prompts in 12 minutes on a single GPU. + +Abstract. We present AvatarPopUp, a method for fast, high quality 3D human avatar generation from different input modalities, such as images and text prompts and with control over the generated pose and shape. The common theme is the use of diffusion-based image generation networks that are specialized for each particular task, followed by a 3D lifting network. We purposefully decouple the generation from the 3D modeling which allows us to leverage powerful image synthesis priors, trained on billions of text-image pairs. We fine-tune latent diffusion networks with additional image conditioning for image generation and back-view prediction, and to support qualitatively different multiple 3D hypotheses. Our partial fine-tuning approach allows to adapt the networks for each task without inducing catastrophic forgetting. In our experiments, we demonstrate that our method produces accurate, high-quality 3D avatars with diverse appearance that respect the multimodal text, image, and body control signals. Our approach can produce a 3D model in as few as 2 seconds, a four orders of magnitude speedup w.r.t. the vast majority of existing methods, most of which solve only a subset of our tasks, and with fewer controls. AvatarPopUp enables applications that require the controlled 3D generation of human avatars at scale. The project website can be found at https://www.nikoskolot.com/avatarpopulation/. + +# 1 Introduction + +We present AvatarPopUp, a method for instant generation of rigged full-body 3D human avatars, with multimodal controls in the form of text, images, and/or human pose and shape. The remarkable recent progress in image synthesis [15, 24, 54, 56, 58, 63] acted as a catalyst for a wide range of media generation applications. In just a few years, we have witnessed rapid developments in video generation [9, 23, 34, 68, 76], audio synthesis [49, 77] or text-to-3D object generation [40, 41, 51, 53, 65], among others. Pivotal to success of all these methods is their probabilistic nature, however this requires very large training sets. While inspiring efforts have been made [14], training set size is still a problem in many domains, and particularly for 3D. In an attempt to alleviate the need for massive 3D datasets, DreamFusion [51] leverages the rich priors of text-to-image diffusion models in an optimization framework. The influential DreamFusion ideas were also quickly adopted for 3D avatar creation [32, 37], a field previously dominated by image- or video-based reconstruction solutions [6, 8, 60, 61]. Text-to-avatar methods enabled novel creative processes, but came with a significant drawback. While image-based methods typically use pretrained feed-forward networks and create outputs in seconds, existing text-to-avatar solutions are optimization-based and take minutes to several hours to complete, per instance. + +In this paper, we are closing this gap and present, for the first time, a methodology for instant, text-controlled, rigged, full-body 3D human avatar creation. Our AvatarPopUp is purely feed-forward, can be conditioned on images and textual descriptions, allows fine-grained control over the generated body pose and shape, can generate multiple hypotheses, and runs in 2-10 seconds per instance. + +Key to success is our pragmatic decoupling of the two stages of probabilistic text-to-image generation and 3D lifting. Decoupling 2D generation and 3D lifting has two major advantages: (1) We can leverage the power of pretrained text-to-image generative networks, which have shown impressive results in modeling complex conditional distributions. Trained with large training sets of images, both generation quality and diversity are very high. (2) We alleviate the need for very large 3D datasets required to train state-of-the-art generative 3D models. Our method generates diverse plausible image configurations that contain rich enough information that can be lifted in 3D with minimal ambiguity. In other words, we distribute the workload between two expert systems: a pretrained probabilistic generation network fine-tuned for our task to produce front and back image views of the person, and a state-of-the-art unimodal, feed-forward image-to-3D model that can be trained using comparably small datasets. + +Our proposed decoupling strategy allows us to maximally exploit available data sources with different levels of supervision. We first fine-tune a pretrained Latent Diffusion network to generate images of people based on textual descriptions and with additional control over the desired pose and shape. This step does not require any ground truth 3D data for supervision and enables scaling our image generator to web scale data of images of people in various poses. Next, we leverage a small-scale dataset of scanned 3D human assets and fine-tune a second latent diffusion network to learn the distribution of back side views con + +
Control from Text Control from Image Shape Control
BodyShapeGeneratesGeometryTexture
✓✓✓✓✓✓/XRuntime
X✓X✓✓/XXseveral hoursOptimization-based text-to-3D [26,32,37]
✓X✓✓XXsecondsImage-to-3D [4,8,60,61]
XX✓✓✓XminutesHybrid [31]
✓✓✓✓✓✓seconds3D GANs [16,25]
secondsAvatarPopUp (Ours)
+ +Table 1: AvatarPopUp generates 3D assets with texture from text prompts or input images of a target subject and can be controlled with body pose and shape. In contrast to baselines that require up to hours per prompt, our model takes under five seconds and can de facto by used in interactive applications. In the experimental section we demonstrate the large diversity of the model, and applications in cloth editing. + +ditioned on a front view image of the person. We optionally also condition on a textual description that can naturally complement the evidence available in the front view image. Furthermore, we propose a novel fine-tuning strategy that prevents overfitting to the new datasets. Finally, we design and train a 3D reconstruction network that predicts a textured 3D shape in the form of an implicit signed distance field given the pair of front and back views and optionally 3D body signals. The resulting cascaded method, which we call AvatarPopUp supports a wide range of 3D generation and reconstruction tasks: First, it enables fast and interactive 3D generation of assets at scale, see Fig. 1. Second, we can repurpose parts of the cascade for image-based 3D reconstruction at state-of-the-art quality. Finally, we demonstrate how AvatarPopUp can be used for creative editing tasks exemplified in 3D virtual try-on with body shape preservation. To summarize, our main contributions are: + +- A method for controllable 3D human avatar generation, based on multimodal text, pose, shape and image input signals, that outputs a detailed human mesh instance in 2-10 seconds. +- We propose a simple yet effective way to fine-tune pretrained diffusion models on small-scale datasets, without inducing catastrophic forgetting. +- While not our primary goal, our approach achieves state-of-the-art results in single-image 3D reconstruction and enables 3D creative editing applications. + +# 2 Related Work + +Table 1 summarizes the characteristics of AvatarPopUp in comparison to previous work, along several important property axes. + +Text-to-3D generation. The success of text-to-image models [55, 57, 59] was quickly followed by a significant amount of work on text-to-3D content generation [38, 45, 51, 66]. Due to limited training data, methods typically use optimization approaches, where a neural representation [47] is optimized per instance by minimizing a distillation loss [51] derived from large text-to-image models. This idea has been extended to generate human avatars [18, 27, 29, 32, 37, 69, 73, 78–80] or heads [19, 36, 39] enabling the text-based creation of 3D human assets that + +are diverse in terms of shape, appearance, clothing and various accessories. In these works, the optimization process is often regularized using a 3D body model [32,37,78], which also enables animation. However, such approaches generally take hours per instance, and rendering is slow. With the appearance of Gaussian Splatting [30], other works [2,42,82] reduced rendering time at the expense of accurate geometry. In any case, creating an avatar still takes a significant amount of time, making such methods unsuitable for interactive applications. In this work we propose an alternative direction, which also builds upon the success of text-to-image models and combines them with 3D reconstruction pipelines. Also related are 3D human generation methods. AG3D [16] and EVA3D [25] are GAN-based methods learned from 2D data, that allow to sample 3D humans anchored in a 3D body model. CHUPA [31] generates dual normal maps based on text and then fits a body model to obtain a full 3D representation. While generation is similar in spirit to our method, CHUPA requires optimization per instance and does not generate texture. + +Photorealistic 3D Human Reconstruction. Our framework generates 3D human assets and builds on top of state-of-the-art 3D reconstruction techniques. This has been widely explored in the past and can be roughly categorized by its use of explicit or implicit representations. An important line of work leverages 3D body models [43,50,72] and reconstructs their associated parameters, in some cases extended with vertex offsets to represent some clothing and hair detail [4-7, 48, 85]. Other efforts have considered voxels [67, 84], depth maps [17] and more recently implicit representations [8, 13, 21, 28, 60-62, 70, 71, 74, 75, 83]. Being topology free, the latter allows the representation of loose clothing more easily. They typically provide more detail and enable high-resolution reconstruction, often conditioned on local pixel-aligned features [60]. On the other hand, these methods yield reconstructions with no semantic labels that cannot be easily animated. To solve this problem, some work combined body models with implicit representations [13, 21, 27, 28, 70, 71], but this is prone to errors when the pose is noisy at inference time. In contrast, we drive the synthesis process with guidance from an input body model – sampled or estimated – so that the generated image is well aligned with the body prior. As we show in Sec. 4, this allows rigging the 3D avatar without post-processing and natively supports 3D animation. + +Given a single input image of a person, previous work aims to generate realistic reconstructions for the non-visible parts. However, this often leads to blurry textures and lack of geometric detail, e.g. no wrinkles. Some methods [60, 61, 70, 71] generate back normal maps to enhance details, or consider probabilistic reconstructions [3, 62]. However, all these methods cannot be prompted from text or other modalities and still yield limited diversity. In contrast, we guide the synthesis process by means of generated front and back images, yielding high-quality 3D reconstructions. Another challenge in previous work is limited training data. Most prior methods rely on a few hundred 3D scans, due to the pricey and laborious process of good quality human capture. We alleviate the need for large scale 3D training data by proposing a framework that can quickly generate humans with a given clothing, pose and shape. + +![](images/162ab2d8066f91630fb298720f8e4fd1e5303b92e5df8162c95d919a99d4e1b2.jpg) +Fig. 2: AvatarPopUp method. (Top) AvatarPopUp builds on the capacity of text-to-image models to generate highly detailed and diverse input images. First, a Latent Diffusion network takes a text prompt and a target body pose and shape $\mathcal{G}$ , and generates a highly detailed front image $I_{f}$ of a person. Next, a second network generates a consistent back view $I_{b}$ in the same pose and clothing. (Bottom) We perform pixel-aligned 3D reconstruction given the generated front/back views $I_{f}, I_{b}$ and optionally the given 3D body pose and shape $\mathcal{G}$ . This decoupling enables the generation of 3D avatars from text, images or a combination of the two. + +# 3 Method + +We learn a distribution $p(X|c)$ of textured 3D shapes $X$ conditioned on a collection of signals $c$ that we factorize as follows + +$$ +p (X | c) = \int \int p \left(X \mid I _ {f}, I _ {b}, c\right) \cdot p \left(I _ {b} \mid I _ {f}, c\right) \cdot p \left(I _ {f} \mid c\right) d I _ {f} d I _ {b}, \tag {1} +$$ + +where $p(X|I_f, I_b, c)$ is the probability of 3D shape $X$ given $c$ and front and back image observations $I_f$ and $I_b$ respectively, $p(I_b|I_f, c)$ is the probability of the back view image given the front image $I_f$ and conditioning signals $c$ , and $p(I_f|c)$ the conditional probability of front view images of the person given $c$ . + +Computing the integral in (1) is intractable, but our goal is to generate samples from the distribution rather than expectations. To do so, we employ ancestral sampling. We first sample a front view $I_{f}$ given $c$ , we then sample a back view $I_{b}$ given $I_{f}$ and $c$ , and last we sample the 3D reconstruction based on the entire context. In practice, $p(I_f|c)$ and $p(I_b|I_f,c)$ are implemented using Latent Diffusion models, whereas $p(X|I_f,I_b,c)$ is a unimodal, neural implicit field generator. + +In the case of single-image 3D reconstruction the conditioning signal $c$ is $I_{f}$ , and consequently we can omit the first step. For text-based generation, $c$ is a text prompt describing the appearance of the person together with a signal encoding the body pose and shape. The conditioning information $c$ may be extended with additional signals, as in the case of 3D editing, cf. Sec. 4. + +# 3.1 Controllable Text-to-Image Generator + +Recent advances in diffusion-based text-to-image generation networks [56] have enabled synthesizing high-quality images given only a text prompt as input. However, for certain use cases, such as human generation, it is difficult to inject fine-grained, inherently continuous forms of control, like the 3D pose of people or their precise body shape proportions, when generating with text alone. + +Inspired by ControlNet [81], we propose to add simultaneous control over body pose and shape by augmenting a pretrained Latent Diffusion network with an additional image input that jointly encodes both modalities. For control we use GHUM [72] but other models (e.g. [43]) can be used. Specifically given 3D pose and shape parameters $\pmb{\theta}$ and $\pmb{\beta}$ respectively, we render the corresponding mesh $\mathcal{M} = \mathrm{GHUM}(\pmb{\theta},\pmb{\beta})$ using GHUM's template coordinates and posed vertex locations as 6D vertex colors, obtaining a dense, pixel-aligned pose- and shape-informed control signal $\mathcal{G}$ . + +To fine-tune the network, we generate a dataset of images of people with corresponding GHUM 3D pose and shape parameters and text annotations. This dataset is comprised of a set of scanned assets [1] that are rendered from different viewpoints, as well as a set of real images scraped from the web. For the synthetic part of the dataset, the pose and shape parameters are obtained by fitting GHUM to 3D scans. We additionally use real images to which we fit GHUM using keypoint optimization in the style of [33]. For all images we obtained text annotations using an off-the shelf image captioning system [12] by prompting it to describe the clothing of the people in the image. Since we are interested in generating 3D human assets, we additionally mask out the background in all images, and train the network to output segmented images. This makes the downstream 3D reconstruction task easier, and improves the reconstruction quality because it encourages the network to focus on human appearance, rather than allocating capacity to model complex backgrounds. + +We want to exploit the rich priors learned by text-to-image foundation models by fine-tuning a Latent Diffusion model [56] with the dense GHUM rendering as an additional input. For fine-tuning we propose a simpler and more lightweight method than a standard ControlNet. We pad the weights of the input convolutional layer with additional channels initialized with zeros, and we then fine-tune only the weights of the convolutional layers of the encoder network. All the decoder and attention layers are kept frozen. With this simple strategy, even though our model is trained on a relatively small set of images, it can generalize to unseen types of clothing. At the same time, our strategy is more practical than training a ControlNet, which involves keeping in separate copy of the original network weights in memory, and thus enables fine-tuning large models with moderate hardware utilization. We optimize the encoder of the diffusion model by minimizing the simple variant of the diffusion loss [24] + +$$ +\mathcal {L} \left(\psi_ {\text {e n c}}\right) = \mathbb {E} _ {\mathcal {E} (x), \epsilon , t, \tau , \mathcal {G}} \left\| \epsilon - \epsilon_ {\psi} \left(z _ {t}, t, \tau , \mathcal {E} (\mathcal {G})\right) \right\|, \tag {2} +$$ + +where $t \in \{1, \dots, T\}$ is the diffusion time step, $\epsilon \sim \mathcal{N}(0, I)$ the injected noise, $z_{t} = \alpha_{t} \mathcal{E}(x) + \nu_{t}$ is the noisy image latent, $\tau$ is the text encoding, $\mathcal{E}(\mathcal{G})$ the + +![](images/1db3f8e20ecd95376e3f40ffa57af103cd1f5e34035f9d69a43d3c1d082b2377.jpg) + +![](images/01db6990a2368276c251bea01ea04cb4d3d3bdd0a1562e42583b377a7562c54b.jpg) + +![](images/69caf4714bb22336466c3167fea99aa8a152d2eb57025c327c4df4a0a4e78ad4.jpg) + +![](images/581460b48d12d79363f37c039c9742e8218ee53639a8e58fb60d61cc7cd679ab.jpg) + +![](images/cade929a17ec102891f1cc8fcf8972281ae09fad17a6875f4a71ac34d99e57bd.jpg) +Fig. 3: Diverse back view hypotheses. Conditioned on the front view, our method is able to generate diverse plausible back views of the person, with different hairstyles, wrinkle patterns, or lighting. Our network can also be controlled with text (second row), to add fine-grained detail to our generated back-side views. + +![](images/dc4a18690be7fdbe678f416b58d57f4d67372038af4ac46a7e6022b1cd0c7b98.jpg) +No text conditioning + +![](images/e2a400732757cac23a4a9922daf6f5839c31b9736fdd90dda848c9c8b4941f10.jpg) ++“gray hair” + +![](images/1daf589e7a6556f159cd04ac5d4f59caf3364c3b6768c12b31433d9c80a1db40.jpg) ++“with a tattoo” + +latent encoding of the dense GHUM signal, and $\psi_{\mathrm{enc}}$ the encoder subset of the denoising UNet parameters $\psi$ . + +# 3.2 Back View Generation + +One could try to lift the generated front views from the previous stage to 3D directly by applying a single-image 3D reconstruction method like PHORHUM [8]. However, because of the inherent ambiguity of the problem this will result in significant loss of geometric detail and blurry textures for the non-visible body surfaces. To avoid this, we propose to fine-tune again a latent diffusion network with the same strategy as in the previous section. This time the additional image conditioning is a front view and optionally a text prompt, and we train the network to learn the distribution of back views conditioned on the front view. The additional text prompt can be used in cases where it is desired to additionally guide the generation by very specific properties. Fig. 3 shows different back sides sampled from the conditional distribution. We also show that the additional text inputs are useful in modulating certain parts of the generation that are not immediately deducible from the front image, such as hairstyles or specific patterns. + +# 3.3 3D Reconstruction Model + +Our 3D reconstruction network is inspired by PHORHUM [8], and our design choices are informed by the limitations of typical single-image 3D reconstruction methods. Specifically, given a collection of input image signals $\mathcal{I} = \{I_f, I_b, \mathcal{G}\}$ , we first concatenate them, and then use a convolutional encoder $G$ to compute + +a pixel-aligned feature map $G(\mathcal{I})$ . The 3D body control signal $\mathcal{G}$ is optional and may be omitted, e.g. for single-image reconstruction. Then, each point $\mathbf{x} \in \mathbb{R}^3$ in the scene gets projected on this feature map to get pixel-aligned features $\mathbf{z}_{\mathbf{x}} = g(\mathcal{I}, \mathbf{x}; \pi) = b(G(\mathcal{I}), \pi(\mathbf{x}))$ using interpolation, where $b(\cdot)$ is the bilinear sampling operator and $\pi(\mathbf{x})$ is the pixel location of the projection of $\mathbf{x}$ using the camera $\pi$ . These pixel aligned features are then concatenated with a positional encoding $\gamma(\mathbf{x})$ of the 3D point and are fed to an MLP $f$ that outputs the signed distance from the surface $d$ as well as surface color $c$ . Finally, the 3D shape $S$ is represented as the zero-level-set of $d$ + +$$ +\mathcal {S} (\mathcal {I}) = \left\{\mathbf {x} \in \mathbb {R} ^ {3} | f (g (I, \mathbf {x}; \pi), \gamma (\mathbf {x})) = (0, c) \right\}. \tag {3} +$$ + +$\mathcal{S}$ can be transformed to a mesh directly, using Marching Cubes [44]. + +# 3.4 Animation of Generated Avatars + +Our method can generate diverse 3D avatars with various poses, shapes and appearances. Optionally, we may leverage the conditioning body model to rig the estimated 3D shape. As a result of our conditioning strategy, 3D avatars and the conditional body model instances are aligned in 3D. This allows us to anchor the reconstructed 3D shape on the body model surface [10] and re-pose or re-shape it accordingly. Alternatively, we may also transfer just the LBS skeleton and weights from the body model to the scan. This enables importing and animating the generated 3D assets in various rendering engines. See Fig. 8 and Sup. Mat. for examples and videos of animations of the generated 3D assets. + +# 4 Experiments + +Data. We use meshes from RenderPeople [1] for training as well as our own captured data, totaling in $\sim 10\mathrm{K}$ scans with diverse poses, body shapes, and clothing styles. We render each scan with randomly sampled HDRI background, random cloth color augmentations, and lighting using Blender [11]. During this process, we render both front and back views used to train the different stages of our model. For the front image generation network we also use a set of $10\mathrm{K}$ real images on which we fitted the GHUM model using 2D keypoints. For testing we defined a split based on subject identity and held out $\sim 1\mathrm{K}$ scans. + +We provide results for 2 different versions of our model. The standard quality model is generated in 2 seconds by running 5 DDIM [64] steps during inference and Marching Cubes at $256^{3}$ resolution. The high quality model is generated in 10 seconds using 50 DDIM steps and Marching Cubes at $512^{3}$ resolution. The timings were recorded on a single 40 GB A100 GPU. Unless otherwise stated, all results we report are obtained using the high quality model. + +Metrics and Baselines. We compare our method numerically for two different problems. First we consider the task of text-to-3D human generation, where we sample 100 different text prompts and compare against representative text-to-3D generation methods. For numerical comparisons we evaluate the text-image + +![](images/8c9a1eba9b64f3018c4f5159767cf74820e47964bd0031a4a402609ea2f6e9e0.jpg) +Fig. 4: Diversity of our 3D generation. For the same text prompt and the same pose and shape conditioning, our model can generate a diverse set of 3D avatars that respect both the text and the 3D body controls. + +alignment using CLIP [52]. Specifically, we use the retrieval accuracy using CLIP, as proposed in [51]. We additionally show qualitative results. + +Second, we validate the performance of our 3D reconstruction component against state-of-the-art methods [8,27,60,61,70,71] considering both geometry and texture. Pixel-aligned image features dominate in recent work [8,60,61], but some methods proposed to combine them with body models [70,71], which offer advantages in being able to animate reconstructions. We also rely on pixel-aligned features, yet propose a method that inherently enables animation. We run AvatarPopUp by generating the back side of the subject and applying the reconstruction network. To evaluate 3D geometry, we report bi-directional Chamfer distance $\times 10^{-3}$ , Normal Consistency (NC ↑), and Volumetric Intersection over Union (IoU ↑) after ICP alignment. However, these metrics do not necessarily correlate with good visual quality, e.g. Chamfer distance is minimized by smooth, non-detailed geometry. To measure the quality of reconstructions, we additionally report FID scores [22] of the front/back views for both geometry and texture. + +# 4.1 Text-to-3D Generation + +In Fig. 4 we generate different avatars given the same text prompt and driving poses. We can see that our model is able to create a very diverse set of assets, a property not observed in the previous text-to-3D generation methods [32]. + +In Tab. 2 we use CLIP to evaluate our model against other text-to-3D generation methods. In general, CLIP-based metrics are not indicative of the generated image quality, because they only consider the alignment with the text, and often over-saturated images with extreme details tend to have high CLIP scores. To further demonstrate that our method generates higher quality avatars, we also include a qualitative comparison in Fig. 5. + +
ColorColorRuntime ↓
R-Prec. ↑Top-3 ↑R-Prec. ↑Top-3 ↑
0.680.920.040.258hDreamHuman [32]
0.560.820.030.153hTADA [37]
--0.030.083mCHUPA [31]
0.580.730.080.17~2sOurs
0.620.770.110.17~10sOurs (high quality)
+ +Table 2: Numerical comparisons with other text-to-3d human generation methods. We mark the best and second best results. Our method allows to trade-off speed and quality and we report results for two different settings. + +![](images/dca817aad1b5c4cf1303324e456c60a0543ca2f879737932aa8e2ffeef5c575e.jpg) +A person wearing a black off-shoulder top and a denim skirt +Fig. 5: Comparisons with text-to-3d human generation methods. Our method generates high quality results that respect the text prompt well, at a fraction of the others' runtime, cf. Tab. 3. TADA's results appear unnatural; DreamHuman failed for one subject and produces oversaturated colors; CHUPA failed to respect the prompt. + +# 4.2 Single-image 3D Reconstruction + +While not specifically designed for 3D reconstruction, our method achieves state-of-the-art performance also for this task. The evaluation setup is the following: given an input image $I$ , we draw one random sample from the back view image generator network, and then feed the pair of front/back images to our 3D reconstruction network. For all methods we extract a textured 3D mesh using Marching Cubes and report numerical results in Tab. 3. Furthermore, we show qualitative results in Fig. 6. Notably, our method not only performs on par numerically and qualitatively on reconstructed front views, but also generates highly detailed back view texture and geometry. Finally, we also compare with the optimization-based method TeCH [27] in Fig. 7. TeCH produces detailed front and back geometry but also also exhibits problems at times, rooted in its 3D pose estimation method. Most importantly, TeCH runs for several hours per instance, while ours computes results in a single feed-forward pass, in only a few seconds. + +
3D MetricsFID ↓
Ch. ↓IoU ↑NC ↑Color. FColor B.Normal F.Normal B.
7.10.500.72--29.657.2PIFu [60]
6.20.520.74--26.952.5PIFuHD [61]
7.00.480.72--36.149.0ICON [71]
3.40.580.76--18.336.1ECON [70]
0.840.710.8711.040.811.238.8PHORHUM* [8]
0.910.700.8611.020.012.825.8Ours
+ +Table 3: Numerical comparisons with single-view 3D reconstructions methods and ablations of our method. We mark the best and second best results. All Chamfer metrics are $\times 10^{-3}$ . Not all methods generate colors. For fair comparisons, we retrained PHORHUM [8] using the same data as AvatarPopUp. We observe comparable results in terms of 3D metrics, however ours performs better on generating more realistic and diverse back views and back normals. + +# 4.3 3D Virtual Try On + +An immediate application of our method is the option to perform 3D garment edits for a given identity. Given an input image, we first recover 3D pose and shape parameters of the person in the image. Using a text prompt and the identity preservation strategy introduced below, we can generate updated images of the same person wearing different garments or accessories. To preserve the identity of the person, we first locate the person's head in the source image and use then Repaint [46] to out-paint a novel body for the given head while still conditioning on the estimated pose and shape parameters. With this strategy we ensure that our method not only preserves the facial characteristics but also the overall body proportions. In Fig. 9 we illustrate such editing examples. The generated 3D edits present garment details like wrinkles on both front and back views, and preserve the subjects' facial appearance. Also, note that body shape and identity are well preserved for the subjects, even though only one image is given. While there has been a significant amount of 2D virtual try-on research [20,35,86], our methodology can generate consistent and highly detailed 3D meshes that can be animated and rendered from other viewpoints. + +# 4.4 Animation + +AvatarPopUp enables animation of the generated assets by design, provided that its generation is conditioned on an underlying body model. In Fig. 8, we show an example of a generated avatar that is rigged automatically. + +# 4.5 Ablation Study + +Pose and shape encoding in the 3D reconstruction network. We compare the effectiveness of our additional pose and shape encoding inputs $\mathcal{G}$ to our reconstruction network. To do so, we use the same set of 100 text prompts as before and for each text prompt sample a random pose and shape configuration. For each $(\tau, \theta, \beta)$ triplet we run inferences and compute 2 meshes: one using + +![](images/33f7851c4d76bff20fb1c6fd45e6c8df67519d153221809c899f5db5aef1c0c4.jpg) + +![](images/8e76543a34a53fe154685c8f8fdab35449968cff6ecd28b41931c8a9e6862184.jpg) + +![](images/1a277a21efc4e57daeed803f8b0c412f8ffdb6a95d5267679749727b5d40ff08.jpg) +Fig. 6: Qualitative comparisons with state-of-the-art single image 3D reconstruction methods. Our method produces front color and normals on par with state-of-the-art and much more detailed back view hypotheses. + +![](images/e79dc5f2dedb4b3ffdf3bf42737cd570d004c8ecaa1c8c7ac072c867c4af560b.jpg) + +![](images/8ac23b6f8be52bac373b46279c603a437dc6cf667e7ff312e21282fce33fae41.jpg) + +![](images/c005d6f61016cf555a1f4332c9f7c4c534e90b5f546040205e819479751e93b7.jpg) + +![](images/577b67f56e47e08f57816e503d5e515d093325fc5537f9cf62057260038e20df.jpg) + +![](images/f446eb4225152364a096a181f1464f51327cc608605628e28236310c96cdbe4a.jpg) +Ours + +![](images/cd0de6d86362a5fa18e9385ccd15422ac54614390d4ab7ec8682c73ee8863a87.jpg) + +![](images/f2ac061c35fa60b0f87b0fb223d1065cf808553e8c4e96565f8bb564ea206bfa.jpg) + +![](images/1ae583bb3a87dcbd16e91e626ca5c29c9236a0d206d8a46b537de1b18d5eb1d4.jpg) + +![](images/964b0c392b24ab8509e4eaeeb90c3449626aa9910e7945c0e6981bc61f1c079d.jpg) + +![](images/1e00fcc6ec50608af2f4641d9ee6804db03edbd9f36657dea13c35cb78157725.jpg) +PHORHUM + +![](images/8c5920f736078ba8bc63186005d4bedecea707dd22e2bc532eb429d6c33ffd8e.jpg) + +![](images/cc7e1d9099b9174a99c2ac3d57a8518d4fd370fcd087890a32876bc2d53b6234.jpg) + +![](images/39bb2b5585ebbd15757b5ecbd220da8505d42aa5141e57c8a4c0a52d23f86a0b.jpg) + +![](images/208dab8bc67d85f3a927ccaa1a57637c436ee52f0d30e42e31b6b96ad1e42a62.jpg) + +![](images/7d0ed829c20412a073bdd09e851b60956f11e3e14aa9fb83ef930a400ff22c74.jpg) + +![](images/a95d4bd010c91addafda3846475302ee32d4172e6e61044918dd164d4b4da8b8.jpg) + +![](images/227939bc09641de63cfda27840ce207abcaf506027c3ad240eb91d9fee96c425.jpg) +PIFuHD + +only the front and back images, and another one additionally using the dense GHUM encodings. We then evaluate the Chamfer distance between each mesh and the corresponding GHUM mesh. The model using the additional GHUM signals has an average Chamfer distance of $d_{\mathrm{with}} = 1.4$ whereas the one without $d_{\mathrm{without}} = 8.6$ , thus validating our design choice. Not only is the control respected well, but this also allows for animation as discussed previously. + +Partial vs. Complete Network Finetuning. We finetune two Latent Diffusion networks, one in a standard way by optimizing all parameters, and another one using our proposed strategy where we only finetune the convolutional layers of the encoder. Empirically we observed that the network that was finetuned as a whole experienced catastrophic forgetting, and has poor performance when asked to generate types of garment not seen in the training set. Fig. 10 shows a comparison for text prompts not in the training set. + +# 5 Discussion & Conclusions + +Limitations. AvatarPopUp inherits limitations from other pixel-aligned methods, e.g. with regions parallel to camera rays being not as detailed in the reconstruction. Further, artifacts might be visible in poses or very loose clothing that are underrepresented during training. + +Ethical Considerations. We present a generative tool to create 3D human assets, thus reducing the risk of scanning and using real humans for training large scale 3D generative models. AvatarPopUp generates diverse results and can lead to a better coverage of subject distributions. + +![](images/644893664bb6a0e81b86e5f6ed2631977fd60acfe99dfcc4712156cb815c5dff.jpg) +Fig. 7: Additional comparisons with TeCH. TeCH is optimization-based and takes several hours to complete for a single reconstruction. Our results are obtained in seconds, yet are as detailed and even less noisy. + +![](images/1e5512458355ed8fa858558aa05d8784f3f158ff317bff63f0750b309f745733.jpg) +Ours + +![](images/add9dc208474d3fb82997f320290655b0af98f036ed2a320767ba0541b6c766f.jpg) +TeCH + +![](images/798f16ce9656d504b756b7a0d45247a153e8448229982b1f4b3b01101a7f5eeb.jpg) +Fig. 8: Reposing example.. We first reconstruct an avatar "wearing a gray suit" in the A-pose and then we transfer it to different poses. + +![](images/ca6e8a5633e19c806361af05dbdef45340e11facaea29dab86c9e17767da3fa7.jpg) +Fig. 9: Identity preserving 3D avatar editing. Our method allows for editing the clothing, while preserving the identity of the generated person. In each row we're using the image on the left as input, recover the person's pose and shape parameters and then run our method using the shown prompts. The generated avatars share the same identity and respect the target prompts. + +![](images/7fbc485cb9f842c8d7873adff5ae0bea9ddd8f3ed28deed93f6287a79c08f91e.jpg) +Fig. 10: Partial vs. complete fine-tuning strategy. For each text prompt the first example is generated using our partial fine-tuning strategy whereas the second by fine-tuning the entire network. Empirically, partial fine-tuning is more resilient to catastrophic forgetting, generating more diverse and better text-aligned images. + +We have presented AvatarPopUp, a novel framework for generating 3D human avatars controlled by text or images and yielding rigged 3D models in 2-10 seconds. AvatarPopUp is purely feed-forward, allows for fine-grained control over the generated body pose and shape, and can produce multiple qualitatively different hypotheses. AvatarPopUp is composed of a cascade of expert systems, decoupling image generation and 3D lifting. Through this design choice, AvatarPopUp benefits from both web-scale image datasets, ensuring high generation diversity, and from smaller size accurate 3D datasets, resulting in reconstructions with increased detail and precisely controlled based on text and identity specifications. In the future, we would like to explore other 3D construction strategies besides pixel-aligned features. Longer term, we aim to support highly detailed and controllable 3D human model generation for entertainment, education, architecture and art, or medical applications. + +# References + +1. https://renderpeople.com/ +2. Abdal, R., Yifan, W., Shi, Z., Xu, Y., Po, R., Kuang, Z., Chen, Q., Yeung, D.Y., Wetzstein, G.: Gaussian shell maps for efficient 3d human generation. arXiv preprint arXiv:2311.17857 (2023) +3. AlBahar, B., Saito, S., Tseng, H.Y., Kim, C., Kopf, J., Huang, J.B.: Single-image 3d human digitization with shape-guided diffusion. In: SIGGRAPH Asia (2023) +4. Alldieck, T., Magnor, M., Bhatnagar, B.L., Theobalt, C., Pons-Moll, G.: Learning to reconstruct people in clothing from a single rgb camera. In: CVPR (2019) +5. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Detailed human avatars from monocular video. In: 3DV (2018) +6. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Video based reconstruction of 3D people models. In: CVPR (2018) +7. Alldieck, T., Pons-Moll, G., Theobalt, C., Magnor, M.: Tex2shape: Detailed full human body geometry from a single image. In: ICCV (2019) +8. Alldieck, T., Zanfir, M., Sminchisescu, C.: Photorealistic monocular 3D reconstruction of humans wearing clothing. In: CVPR (2022) +9. Bar-Tal, O., Chefer, H., Tov, O., Herrmann, C., Paiss, R., Zada, S., Ephrat, A., Hur, J., Li, Y., Michaeli, T., et al.: Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945 (2024) +0. Bazavan, E.G., Zanfir, A., Zanfir, M., Freeman, W.T., Sukthankar, R., Sminchisescu, C.: Hspace: Synthetic parametric humans animated in complex environments. arXiv (2021) +1. Blender Online Community: Blender - a 3D modelling and rendering package. Blender Foundation, Blender Institute, Amsterdam (2020), http://www.blender.org +12. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., Kolesnikov, A., Puigcerver, J., Ding, N., Rong, K., Akbari, H., Mishra, G., Xue, L., Thapliyal, A., Bradbury, J., Kuo, W., Seyedhosseini, M., Jia, C., Ayan, B.K., Riquelme, C., Steiner, A., Angelova, A., Zhai, X., Houlsby, N., Soricut, R.: Pali: A jointly-scaled multilingual language-image model (2022). https://doi.org/10.48550/ARXIV.2209.06794, https://arxiv.org/abs/2209.06794 +13. Corona, E., Zanfir, M., Alldieck, T., Bazavan, E.G., Zanfir, A., Sminchisescu, C.: Structured 3d features for reconstructing relightable and animatable avatars. In: CVPR (2023) +14. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3d objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13142-13153 (2023) +15. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021) +16. Dong, Z., Xu Chen, J.Y., J.Black, M., Hilliges, O., Geiger, A.: AG3D: Learning to generate 3D avatars from 2D image collections. In: International Conference on Computer Vision (ICCV) (2023) +17. Gabeur, V., Franco, J.S., Martin, X., Schmid, C., Rogez, G.: Moulding humans: Non-parametric 3D human shape estimation from single images. In: ICCV (2019) +18. Gong, C., Dai, Y., Li, R., Bao, A., Li, J., Yang, J., Zhang, Y., Li, X.: Text2 avatar: Text to 3d human avatar generation with codebook-driven body controllable attribute. arXiv preprint arXiv:2401.00711 (2024) + +19. Han, X., Cao, Y., Han, K., Zhu, X., Deng, J., Song, Y.Z., Xiang, T., Wong, K.Y.K.: Headsculpt: Crafting 3d head avatars with text. Advances in Neural Information Processing Systems 36 (2024) +20. Han, X., Wu, Z., Wu, Z., Yu, R., Davis, L.S.: Viton: An image-based virtual try-on network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7543-7552 (2018) +21. He, T., Xu, Y., Saito, S., Soatto, S., Tung, T.: Arch++: Animation-ready clothed human reconstruction revisited. In: CVPR (2021) +22. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017) +23. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022) +24. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239 (2020) +25. Hong, F., Chen, Z., LAN, Y., Pan, L., Liu, Z.: EVA3d: Compositional 3d human generation from 2d image collections. In: International Conference on Learning Representations (2023), https://openreview.net/forum?id=g7U9jD_2CUr +26. Hong, F., Zhang, M., Pan, L., Cai, Z., Yang, L., Liu, Z.: Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. ACM Transactions on Graphics (TOG) 41(4), 1-19 (2022) +27. Huang, Y., Yi, H., Xiu, Y., Liao, T., Tang, J., Cai, D., Thies, J.: TeCH: Text-guided Reconstruction of Lifelike Clothed Humans. In: International Conference on 3D Vision (3DV) (2024) +28. Huang, Z., Xu, Y., Lassner, C., Li, H., Tung, T.: Arch: Animatable reconstruction of clothed humans. In: CVPR (2020) +29. Jiang, R., Wang, C., Zhang, J., Chai, M., He, M., Chen, D., Liao, J.: Avatarcraft: Transforming text into neural human avatars with parameterized shape and pose control. arXiv preprint arXiv:2303.17606 (2023) +30. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023) +31. Kim, B., Kwon, P., Lee, K., Lee, M., Han, S., Kim, D., Joo, H.: Chupa: Carving 3d clothed humans from skinned shape priors using 2d diffusion probabilistic models. arXiv preprint arXiv:2305.11870 (2023) +32. Kolotouros, N., Alldieck, T., Zanfir, A., Bazavan, E., Fieraru, M., Sminchisescu, C.: Dreamhuman: Animatable 3d avatars from text. Advances in Neural Information Processing Systems 36 (2024) +33. Kolotouros, N., Pavlakos, G., Black, M.J., Daniilidis, K.: Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2252-2261 (2019) +34. Kondratyuk, D., Yu, L., Gu, X., Lezama, J., Huang, J., Hornung, R., Adam, H., Akbari, H., Alon, Y., Birodkar, V., et al.: Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125 (2023) +35. Lee, S., Gu, G., Park, S., Choi, S., Choo, J.: High-resolution virtual try-on with misalignment and occlusion-handled conditions. In: European Conference on Computer Vision. pp. 204-219. Springer (2022) +36. Lei, B., Yu, K., Feng, M., Cui, M., Xie, X.: Diffusiongan3d: Boosting text-guided 3d generation and domain adaption by combining 3d gans and diffusion priors. arXiv preprint arXiv:2312.16837 (2023) + +37. Liao, T., Yi, H., Xiu, Y., Tang, J., Huang, Y., Thies, J., Black, M.J.: Tada! text to animatable digital avatars. In: 3DV (2023) +38. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023) +39. Liu, H., Wang, X., Wan, Z., Shen, Y., Song, Y., Liao, J., Chen, Q.: Headartist: Text-conditioned 3d head generation with self score distillation. arXiv preprint arXiv:2312.07539 (2023) +40. Liu, M., Xu, C., Jin, H., Chen, L., Varma T, M., Xu, Z., Su, H.: One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems 36 (2024) +41. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9298-9309 (2023) +42. Liu, X., Zhan, X., Tang, J., Shan, Y., Zeng, G., Lin, D., Liu, X., Liu, Z.: Human-gaussian: Text-driven 3d human generation with gaussian splatting. arXiv preprint arXiv:2311.17061 (2023) +43. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ToG (2015) +44. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3D surface construction algorithm. SIGGRAPH (1987) +45. Lorraine, J., Xie, K., Zeng, X., Lin, C.H., Takikawa, T., Sharp, N., Lin, T.Y., Liu, M.Y., Fidler, S., Lucas, J.: Att3d: Amortized text-to-3d object synthesis. arXiv preprint arXiv:2306.07349 (2023) +46. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11461-11471 (2022) +47. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. ECCV (2020) +48. Onizuka, H., Hayirci, Z., Thomas, D., Sugimoto, A., Uchiyama, H., Taniguchi, R.i.: Tetratsdf: 3D human reconstruction from a single image with a tetrahedral outer shell. In: CVPR (2020) +49. Oord, A.v.d., Dieleman, S., Zen, H., Simonyan, K., Vinyals, O., Graves, A., Kalchbrenner, N., Senior, A., Kavukcuoglu, K.: Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499 (2016) +50. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: CVPR (2019) +51. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: Int. Conf. Learn. Represent. (2022) +52. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +53. Raj, A., Kaza, S., Poole, B., Niemeyer, M., Mildenhall, B., Ruiz, N., Zada, S., Aberman, K., Rubenstein, M., Barron, J., Li, Y., Jampani, V.: Dreambooth3d: Subject-driven text-to-3d generation. ICCV (2023) + +54. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021) +55. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021) +56. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10684-10695 (2022) +57. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +58. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS. vol. 35, pp. 36479-36494 (2022) +59. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) +60. Saito, S., Huang, Z., Natsume, R., Morishima, S., Kanazawa, A., Li, H.: Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In: ICCV (2019) +61. Saito, S., Simon, T., Saragih, J., Joo, H.: Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3D human digitization. In: CVPR (2020) +62. Sengupta, A., Alldieck, T., Kolotouros, N., Corona, E., Zanfir, A., Sminchisescu, C.: Diffhuman: Probabilistic photorealistic 3D reconstruction of humans. In: CVPR (2024) +63. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015) +64. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: Int. Conf. Learn. Represent. (2020) +65. Sun, J., Zhang, B., Shao, R., Wang, L., Liu, W., Xie, Z., Liu, Y.: Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior. arXiv preprint arXiv:2310.16818 (2023) +66. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023) +67. Varol, G., Ceylan, D., Russell, B., Yang, J., Yumer, E., Laptev, I., Schmid, C.: Bodynet: Volumetric inference of 3D human body shapes. In: ECCV (2018) +68. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399 (2022) +69. Wang, J., Liu, Y., Dou, Z., Yu, Z., Liang, Y., Li, X., Wang, W., Xie, R., Song, L.: Disentangled clothed avatar generation from text descriptions. arXiv preprint arXiv:2312.05295 (2023) +70. Xiu, Y., Yang, J., Cao, X., Tzionas, D., Black, M.J.: ECON: Explicit Clothed humans Optimized via Normal integration. In: CVPR (June 2023) +71. Xiu, Y., Yang, J., Tzionas, D., Black, M.J.: Icon: Implicit clothed humans obtained from normals. In: CVPR (2022) + +72. Xu, H., Bazavan, E.G., Zanfir, A., Freeman, W.T., Sukthankar, R., Sminchisescu, C.: Ghum & ghuml: Generative 3D human shape and articulated pose models. In: CVPR (2020) +73. Xu, Y., Yang, Z., Yang, Y.: Seeavatar: Photorealistic text-to-3d avatar generation with constrained geometry and appearance. arXiv preprint arXiv:2312.08889 (2023) +74. Yang, Z., Wang, S., Manivasagam, S., Huang, Z., Ma, W.C., Yan, X., Yumer, E., Urtasun, R.: S3: Neural shape, skeleton, and skinning fields for 3D human modeling. In: CVPR (2021) +75. Yanting Chan, K., Lin, G., Zhao, H., Lin, W.: Integratedpifu: Integrated pixel aligned implicit function for single-view human reconstruction. In: ECCV (2022) +76. Zeng, Y., Wei, G., Zheng, J., Zou, J., Wei, Y., Zhang, Y., Li, H.: Make pixels dance: High-dynamic video generation. arXiv preprint arXiv:2311.10982 (2023) +77. Zhang, C., Zhang, C., Zheng, S., Zhang, M., Qamar, M., Bae, S.H., Kweon, I.S.: Audio diffusion model for speech synthesis: A survey on text to speech and speech enhancement in generative ai. arXiv preprint arXiv:2303.13336 (2023) +78. Zhang, H., Feng, Y., Kulits, P., Wen, Y., Thies, J., Black, M.J.: Text-guided generation and editing of compositional 3d avatars. arXiv preprint arXiv:2309.07125 (2023) +79. Zhang, H., Chen, B., Yang, H., Qu, L., Wang, X., Chen, L., Long, C., Zhu, F., Du, K., Zheng, M.: Avatarverse: High-quality & stable 3d avatar creation from text and pose. arXiv preprint arXiv:2308.03610 (2023) +80. Zhang, J., Zhang, X., Zhang, H., Liew, J.H., Zhang, C., Yang, Y., Feng, J.: Avatarstudio: High-fidelity and animatable 3d avatar creation from text. arXiv preprint arXiv:2311.17917 (2023) +81. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. IEEE International Conference on Computer Vision (ICCV) (2023) +82. Zhao, Z., Bao, Z., Li, Q., Qiu, G., Liu, K.: Psavatar: A point-based morphable shape model for real-time head avatar creation with 3d gaussian splatting. arXiv preprint arXiv:2401.12900 (2024) +83. Zheng, Z., Yu, T., Liu, Y., Dai, Q.: Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. PAMI (2021) +84. Zheng, Z., Yu, T., Wei, Y., Dai, Q., Liu, Y.: Deephuman: 3D human reconstruction from a single image. In: ICCV (2019) +85. Zhu, H., Zuo, X., Wang, S., Cao, X., Yang, R.: Detailed human shape estimation from a single image by hierarchical mesh deformation. In: CVPR (2019) +86. Zhu, L., Yang, D., Zhu, T., Reda, F., Chan, W., Sahara, C., Norouzi, M., Kemelmacher-Shlizerman, I.: Tryondiffusion: A tale of two unets. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4606-4615 (2023) \ No newline at end of file diff --git a/2024/Instant 3D Human Avatar Generation using Image Diffusion Models/images.zip b/2024/Instant 3D Human Avatar Generation using Image Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..14bc559caa8f0ae2dc45d1095070f8cc26269632 --- /dev/null +++ b/2024/Instant 3D Human Avatar Generation using Image Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a0c0edbf818c2d1b1cf0d839bc4f7081b5c2d1724d9b01d29691e5bd96a597d +size 637193 diff --git a/2024/Instant 3D Human Avatar Generation using Image Diffusion Models/layout.json b/2024/Instant 3D Human Avatar Generation using Image Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d700874af1e573cd99acb09f285a501910ec63a9 --- /dev/null +++ b/2024/Instant 3D Human Avatar Generation using Image Diffusion Models/layout.json @@ -0,0 +1,10559 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 112, + 460, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 112, + 460, + 147 + ], + "spans": [ + { + "bbox": [ + 153, + 112, + 460, + 147 + ], + "type": "text", + "content": "Instant 3D Human Avatar Generation using Image Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 149, + 162, + 463, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 162, + 463, + 186 + ], + "spans": [ + { + "bbox": [ + 149, + 162, + 463, + 186 + ], + "type": "text", + "content": "Nikos Kolotouros, Thiemo Alldieck, Enric Corona, Eduard Gabriel Bazavan, and Cristian Sminchisescu" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 148, + 196, + 465, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 196, + 465, + 217 + ], + "spans": [ + { + "bbox": [ + 148, + 196, + 465, + 217 + ], + "type": "text", + "content": "Google Research\\* {kolotouros, alldieck, egbazavan, enriccorona, sminchisescu}@google.com" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 135, + 229, + 477, + 363 + ], + "blocks": [ + { + "bbox": [ + 135, + 229, + 477, + 363 + ], + "lines": [ + { + "bbox": [ + 135, + 229, + 477, + 363 + ], + "spans": [ + { + "bbox": [ + 135, + 229, + 477, + 363 + ], + "type": "image", + "image_path": "13947c3f92d8dc40fa114453e17f43421b851573c29a2de60d376a787ccfaa73.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 365, + 482, + 409 + ], + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 409 + ], + "type": "text", + "content": "Fig. 1: We present AvatarPopUp, a new method for the automatic generation of 3D human assets. AvatarPopUp can generate rigged 3D models from text or from single images and has control over body pose and shape. In this example, we show 77 models generated from various text prompts in 12 minutes on a single GPU." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 421, + 453, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 421, + 453, + 641 + ], + "spans": [ + { + "bbox": [ + 159, + 421, + 453, + 641 + ], + "type": "text", + "content": "Abstract. We present AvatarPopUp, a method for fast, high quality 3D human avatar generation from different input modalities, such as images and text prompts and with control over the generated pose and shape. The common theme is the use of diffusion-based image generation networks that are specialized for each particular task, followed by a 3D lifting network. We purposefully decouple the generation from the 3D modeling which allows us to leverage powerful image synthesis priors, trained on billions of text-image pairs. We fine-tune latent diffusion networks with additional image conditioning for image generation and back-view prediction, and to support qualitatively different multiple 3D hypotheses. Our partial fine-tuning approach allows to adapt the networks for each task without inducing catastrophic forgetting. In our experiments, we demonstrate that our method produces accurate, high-quality 3D avatars with diverse appearance that respect the multimodal text, image, and body control signals. Our approach can produce a 3D model in as few as 2 seconds, a four orders of magnitude speedup w.r.t. the vast majority of existing methods, most of which solve only a subset of our tasks, and with fewer controls. AvatarPopUp enables applications that require the controlled 3D generation of human avatars at scale. The project website can be found at https://www.nikoskolot.com/avatarpopulation/." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 654, + 252, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 654, + 252, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 654, + 252, + 665 + ], + "type": "text", + "content": "* Now at Google DeepMind." + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 354 + ], + "type": "text", + "content": "We present AvatarPopUp, a method for instant generation of rigged full-body 3D human avatars, with multimodal controls in the form of text, images, and/or human pose and shape. The remarkable recent progress in image synthesis [15, 24, 54, 56, 58, 63] acted as a catalyst for a wide range of media generation applications. In just a few years, we have witnessed rapid developments in video generation [9, 23, 34, 68, 76], audio synthesis [49, 77] or text-to-3D object generation [40, 41, 51, 53, 65], among others. Pivotal to success of all these methods is their probabilistic nature, however this requires very large training sets. While inspiring efforts have been made [14], training set size is still a problem in many domains, and particularly for 3D. In an attempt to alleviate the need for massive 3D datasets, DreamFusion [51] leverages the rich priors of text-to-image diffusion models in an optimization framework. The influential DreamFusion ideas were also quickly adopted for 3D avatar creation [32, 37], a field previously dominated by image- or video-based reconstruction solutions [6, 8, 60, 61]. Text-to-avatar methods enabled novel creative processes, but came with a significant drawback. While image-based methods typically use pretrained feed-forward networks and create outputs in seconds, existing text-to-avatar solutions are optimization-based and take minutes to several hours to complete, per instance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 355, + 481, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 355, + 481, + 414 + ], + "spans": [ + { + "bbox": [ + 133, + 355, + 481, + 414 + ], + "type": "text", + "content": "In this paper, we are closing this gap and present, for the first time, a methodology for instant, text-controlled, rigged, full-body 3D human avatar creation. Our AvatarPopUp is purely feed-forward, can be conditioned on images and textual descriptions, allows fine-grained control over the generated body pose and shape, can generate multiple hypotheses, and runs in 2-10 seconds per instance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 415, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 415, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 133, + 415, + 481, + 569 + ], + "type": "text", + "content": "Key to success is our pragmatic decoupling of the two stages of probabilistic text-to-image generation and 3D lifting. Decoupling 2D generation and 3D lifting has two major advantages: (1) We can leverage the power of pretrained text-to-image generative networks, which have shown impressive results in modeling complex conditional distributions. Trained with large training sets of images, both generation quality and diversity are very high. (2) We alleviate the need for very large 3D datasets required to train state-of-the-art generative 3D models. Our method generates diverse plausible image configurations that contain rich enough information that can be lifted in 3D with minimal ambiguity. In other words, we distribute the workload between two expert systems: a pretrained probabilistic generation network fine-tuned for our task to produce front and back image views of the person, and a state-of-the-art unimodal, feed-forward image-to-3D model that can be trained using comparably small datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 570, + 481, + 665 + ], + "type": "text", + "content": "Our proposed decoupling strategy allows us to maximally exploit available data sources with different levels of supervision. We first fine-tune a pretrained Latent Diffusion network to generate images of people based on textual descriptions and with additional control over the desired pose and shape. This step does not require any ground truth 3D data for supervision and enables scaling our image generator to web scale data of images of people in various poses. Next, we leverage a small-scale dataset of scanned 3D human assets and fine-tune a second latent diffusion network to learn the distribution of back side views con" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "type": "text", + "content": "N. Kolotouros et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 481, + 188 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 188 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 188 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 188 + ], + "type": "table", + "html": "
Control from Text Control from Image Shape Control
BodyShapeGeneratesGeometryTexture
✓✓✓✓✓✓/XRuntime
X✓X✓✓/XXseveral hoursOptimization-based text-to-3D [26,32,37]
✓X✓✓XXsecondsImage-to-3D [4,8,60,61]
XX✓✓✓XminutesHybrid [31]
✓✓✓✓✓✓seconds3D GANs [16,25]
secondsAvatarPopUp (Ours)
", + "image_path": "7b3a3c98dd89b2fa8378124325d793e29af52efddfb4da07794600ca2bc71598.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 191, + 479, + 246 + ], + "lines": [ + { + "bbox": [ + 130, + 191, + 479, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 479, + 246 + ], + "type": "text", + "content": "Table 1: AvatarPopUp generates 3D assets with texture from text prompts or input images of a target subject and can be controlled with body pose and shape. In contrast to baselines that require up to hours per prompt, our model takes under five seconds and can de facto by used in interactive applications. In the experimental section we demonstrate the large diversity of the model, and applications in cloth editing." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 261, + 480, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 480, + 416 + ], + "type": "text", + "content": "ditioned on a front view image of the person. We optionally also condition on a textual description that can naturally complement the evidence available in the front view image. Furthermore, we propose a novel fine-tuning strategy that prevents overfitting to the new datasets. Finally, we design and train a 3D reconstruction network that predicts a textured 3D shape in the form of an implicit signed distance field given the pair of front and back views and optionally 3D body signals. The resulting cascaded method, which we call AvatarPopUp supports a wide range of 3D generation and reconstruction tasks: First, it enables fast and interactive 3D generation of assets at scale, see Fig. 1. Second, we can repurpose parts of the cascade for image-based 3D reconstruction at state-of-the-art quality. Finally, we demonstrate how AvatarPopUp can be used for creative editing tasks exemplified in 3D virtual try-on with body shape preservation. To summarize, our main contributions are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 426, + 479, + 508 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 426, + 479, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 479, + 459 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 479, + 459 + ], + "type": "text", + "content": "- A method for controllable 3D human avatar generation, based on multimodal text, pose, shape and image input signals, that outputs a detailed human mesh instance in 2-10 seconds." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 461, + 479, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 461, + 479, + 483 + ], + "spans": [ + { + "bbox": [ + 138, + 461, + 479, + 483 + ], + "type": "text", + "content": "- We propose a simple yet effective way to fine-tune pretrained diffusion models on small-scale datasets, without inducing catastrophic forgetting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 487, + 479, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 479, + 508 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 479, + 508 + ], + "type": "text", + "content": "- While not our primary goal, our approach achieves state-of-the-art results in single-image 3D reconstruction and enables 3D creative editing applications." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 529, + 237, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 529, + 237, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 529, + 237, + 542 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 556, + 479, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 479, + 580 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 479, + 580 + ], + "type": "text", + "content": "Table 1 summarizes the characteristics of AvatarPopUp in comparison to previous work, along several important property axes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "content": "Text-to-3D generation. The success of text-to-image models [55, 57, 59] was quickly followed by a significant amount of work on text-to-3D content generation [38, 45, 51, 66]. Due to limited training data, methods typically use optimization approaches, where a neural representation [47] is optimized per instance by minimizing a distillation loss [51] derived from large text-to-image models. This idea has been extended to generate human avatars [18, 27, 29, 32, 37, 69, 73, 78–80] or heads [19, 36, 39] enabling the text-based creation of 3D human assets that" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "type": "text", + "content": "Instant 3D Human Avatar Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "are diverse in terms of shape, appearance, clothing and various accessories. In these works, the optimization process is often regularized using a 3D body model [32,37,78], which also enables animation. However, such approaches generally take hours per instance, and rendering is slow. With the appearance of Gaussian Splatting [30], other works [2,42,82] reduced rendering time at the expense of accurate geometry. In any case, creating an avatar still takes a significant amount of time, making such methods unsuitable for interactive applications. In this work we propose an alternative direction, which also builds upon the success of text-to-image models and combines them with 3D reconstruction pipelines. Also related are 3D human generation methods. AG3D [16] and EVA3D [25] are GAN-based methods learned from 2D data, that allow to sample 3D humans anchored in a 3D body model. CHUPA [31] generates dual normal maps based on text and then fits a body model to obtain a full 3D representation. While generation is similar in spirit to our method, CHUPA requires optimization per instance and does not generate texture." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 301, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 517 + ], + "type": "text", + "content": "Photorealistic 3D Human Reconstruction. Our framework generates 3D human assets and builds on top of state-of-the-art 3D reconstruction techniques. This has been widely explored in the past and can be roughly categorized by its use of explicit or implicit representations. An important line of work leverages 3D body models [43,50,72] and reconstructs their associated parameters, in some cases extended with vertex offsets to represent some clothing and hair detail [4-7, 48, 85]. Other efforts have considered voxels [67, 84], depth maps [17] and more recently implicit representations [8, 13, 21, 28, 60-62, 70, 71, 74, 75, 83]. Being topology free, the latter allows the representation of loose clothing more easily. They typically provide more detail and enable high-resolution reconstruction, often conditioned on local pixel-aligned features [60]. On the other hand, these methods yield reconstructions with no semantic labels that cannot be easily animated. To solve this problem, some work combined body models with implicit representations [13, 21, 27, 28, 70, 71], but this is prone to errors when the pose is noisy at inference time. In contrast, we drive the synthesis process with guidance from an input body model – sampled or estimated – so that the generated image is well aligned with the body prior. As we show in Sec. 4, this allows rigging the 3D avatar without post-processing and natively supports 3D animation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Given a single input image of a person, previous work aims to generate realistic reconstructions for the non-visible parts. However, this often leads to blurry textures and lack of geometric detail, e.g. no wrinkles. Some methods [60, 61, 70, 71] generate back normal maps to enhance details, or consider probabilistic reconstructions [3, 62]. However, all these methods cannot be prompted from text or other modalities and still yield limited diversity. In contrast, we guide the synthesis process by means of generated front and back images, yielding high-quality 3D reconstructions. Another challenge in previous work is limited training data. Most prior methods rely on a few hundred 3D scans, due to the pricey and laborious process of good quality human capture. We alleviate the need for large scale 3D training data by proposing a framework that can quickly generate humans with a given clothing, pose and shape." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "text", + "content": "N. Kolotouros et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 112, + 482, + 269 + ], + "blocks": [ + { + "bbox": [ + 134, + 112, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 134, + 112, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 134, + 112, + 482, + 269 + ], + "type": "image", + "image_path": "162ab2d8066f91630fb298720f8e4fd1e5303b92e5df8162c95d919a99d4e1b2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 271, + 482, + 360 + ], + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 360 + ], + "type": "text", + "content": "Fig. 2: AvatarPopUp method. (Top) AvatarPopUp builds on the capacity of text-to-image models to generate highly detailed and diverse input images. First, a Latent Diffusion network takes a text prompt and a target body pose and shape " + }, + { + "bbox": [ + 130, + 271, + 482, + 360 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 271, + 482, + 360 + ], + "type": "text", + "content": ", and generates a highly detailed front image " + }, + { + "bbox": [ + 130, + 271, + 482, + 360 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 130, + 271, + 482, + 360 + ], + "type": "text", + "content": " of a person. Next, a second network generates a consistent back view " + }, + { + "bbox": [ + 130, + 271, + 482, + 360 + ], + "type": "inline_equation", + "content": "I_{b}" + }, + { + "bbox": [ + 130, + 271, + 482, + 360 + ], + "type": "text", + "content": " in the same pose and clothing. (Bottom) We perform pixel-aligned 3D reconstruction given the generated front/back views " + }, + { + "bbox": [ + 130, + 271, + 482, + 360 + ], + "type": "inline_equation", + "content": "I_{f}, I_{b}" + }, + { + "bbox": [ + 130, + 271, + 482, + 360 + ], + "type": "text", + "content": " and optionally the given 3D body pose and shape " + }, + { + "bbox": [ + 130, + 271, + 482, + 360 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 271, + 482, + 360 + ], + "type": "text", + "content": ". This decoupling enables the generation of 3D avatars from text, images or a combination of the two." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 376, + 202, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 202, + 388 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 202, + 388 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 404, + 480, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 480, + 429 + ], + "type": "text", + "content": "We learn a distribution " + }, + { + "bbox": [ + 130, + 404, + 480, + 429 + ], + "type": "inline_equation", + "content": "p(X|c)" + }, + { + "bbox": [ + 130, + 404, + 480, + 429 + ], + "type": "text", + "content": " of textured 3D shapes " + }, + { + "bbox": [ + 130, + 404, + 480, + 429 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 404, + 480, + 429 + ], + "type": "text", + "content": " conditioned on a collection of signals " + }, + { + "bbox": [ + 130, + 404, + 480, + 429 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 404, + 480, + 429 + ], + "type": "text", + "content": " that we factorize as follows" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 184, + 438, + 481, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 438, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 184, + 438, + 481, + 464 + ], + "type": "interline_equation", + "content": "p (X | c) = \\int \\int p \\left(X \\mid I _ {f}, I _ {b}, c\\right) \\cdot p \\left(I _ {b} \\mid I _ {f}, c\\right) \\cdot p \\left(I _ {f} \\mid c\\right) d I _ {f} d I _ {b}, \\tag {1}", + "image_path": "2da597e4ea60904cdba0055df6e4cc1d0a40adcd80e314600eb7e9240bfc782f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "inline_equation", + "content": "p(X|I_f, I_b, c)" + }, + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "text", + "content": " is the probability of 3D shape " + }, + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "text", + "content": " and front and back image observations " + }, + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "inline_equation", + "content": "I_f" + }, + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "inline_equation", + "content": "I_b" + }, + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "text", + "content": " respectively, " + }, + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "inline_equation", + "content": "p(I_b|I_f, c)" + }, + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "text", + "content": " is the probability of the back view image given the front image " + }, + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "inline_equation", + "content": "I_f" + }, + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "text", + "content": " and conditioning signals " + }, + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "inline_equation", + "content": "p(I_f|c)" + }, + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "text", + "content": " the conditional probability of front view images of the person given " + }, + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 472, + 480, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": "Computing the integral in (1) is intractable, but our goal is to generate samples from the distribution rather than expectations. To do so, we employ ancestral sampling. We first sample a front view " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": ", we then sample a back view " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "I_{b}" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": ", and last we sample the 3D reconstruction based on the entire context. In practice, " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "p(I_f|c)" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "p(I_b|I_f,c)" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": " are implemented using Latent Diffusion models, whereas " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "p(X|I_f,I_b,c)" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": " is a unimodal, neural implicit field generator." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "In the case of single-image 3D reconstruction the conditioning signal " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", and consequently we can omit the first step. For text-based generation, " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " is a text prompt describing the appearance of the person together with a signal encoding the body pose and shape. The conditioning information " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " may be extended with additional signals, as in the case of 3D editing, cf. Sec. 4." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "type": "text", + "content": "Instant 3D Human Avatar Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 353, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 353, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 353, + 128 + ], + "type": "text", + "content": "3.1 Controllable Text-to-Image Generator" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 479, + 193 + ], + "type": "text", + "content": "Recent advances in diffusion-based text-to-image generation networks [56] have enabled synthesizing high-quality images given only a text prompt as input. However, for certain use cases, such as human generation, it is difficult to inject fine-grained, inherently continuous forms of control, like the 3D pose of people or their precise body shape proportions, when generating with text alone." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 480, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 480, + 289 + ], + "type": "text", + "content": "Inspired by ControlNet [81], we propose to add simultaneous control over body pose and shape by augmenting a pretrained Latent Diffusion network with an additional image input that jointly encodes both modalities. For control we use GHUM [72] but other models (e.g. [43]) can be used. Specifically given 3D pose and shape parameters " + }, + { + "bbox": [ + 130, + 194, + 480, + 289 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 130, + 194, + 480, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 194, + 480, + 289 + ], + "type": "inline_equation", + "content": "\\pmb{\\beta}" + }, + { + "bbox": [ + 130, + 194, + 480, + 289 + ], + "type": "text", + "content": " respectively, we render the corresponding mesh " + }, + { + "bbox": [ + 130, + 194, + 480, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\mathrm{GHUM}(\\pmb{\\theta},\\pmb{\\beta})" + }, + { + "bbox": [ + 130, + 194, + 480, + 289 + ], + "type": "text", + "content": " using GHUM's template coordinates and posed vertex locations as 6D vertex colors, obtaining a dense, pixel-aligned pose- and shape-informed control signal " + }, + { + "bbox": [ + 130, + 194, + 480, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 194, + 480, + 289 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 290, + 480, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 480, + 456 + ], + "type": "text", + "content": "To fine-tune the network, we generate a dataset of images of people with corresponding GHUM 3D pose and shape parameters and text annotations. This dataset is comprised of a set of scanned assets [1] that are rendered from different viewpoints, as well as a set of real images scraped from the web. For the synthetic part of the dataset, the pose and shape parameters are obtained by fitting GHUM to 3D scans. We additionally use real images to which we fit GHUM using keypoint optimization in the style of [33]. For all images we obtained text annotations using an off-the shelf image captioning system [12] by prompting it to describe the clothing of the people in the image. Since we are interested in generating 3D human assets, we additionally mask out the background in all images, and train the network to output segmented images. This makes the downstream 3D reconstruction task easier, and improves the reconstruction quality because it encourages the network to focus on human appearance, rather than allocating capacity to model complex backgrounds." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 458, + 481, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 481, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 481, + 613 + ], + "type": "text", + "content": "We want to exploit the rich priors learned by text-to-image foundation models by fine-tuning a Latent Diffusion model [56] with the dense GHUM rendering as an additional input. For fine-tuning we propose a simpler and more lightweight method than a standard ControlNet. We pad the weights of the input convolutional layer with additional channels initialized with zeros, and we then fine-tune only the weights of the convolutional layers of the encoder network. All the decoder and attention layers are kept frozen. With this simple strategy, even though our model is trained on a relatively small set of images, it can generalize to unseen types of clothing. At the same time, our strategy is more practical than training a ControlNet, which involves keeping in separate copy of the original network weights in memory, and thus enables fine-tuning large models with moderate hardware utilization. We optimize the encoder of the diffusion model by minimizing the simple variant of the diffusion loss [24]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 208, + 621, + 480, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 621, + 480, + 634 + ], + "spans": [ + { + "bbox": [ + 208, + 621, + 480, + 634 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(\\psi_ {\\text {e n c}}\\right) = \\mathbb {E} _ {\\mathcal {E} (x), \\epsilon , t, \\tau , \\mathcal {G}} \\left\\| \\epsilon - \\epsilon_ {\\psi} \\left(z _ {t}, t, \\tau , \\mathcal {E} (\\mathcal {G})\\right) \\right\\|, \\tag {2}", + "image_path": "0bfdc1f762aa3784a141075940a6152b7d4b6c34e34c3a794b598f833ebe31ff.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "t \\in \\{1, \\dots, T\\}" + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": " is the diffusion time step, " + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": " the injected noise, " + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "z_{t} = \\alpha_{t} \\mathcal{E}(x) + \\nu_{t}" + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": " is the noisy image latent, " + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": " is the text encoding, " + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\mathcal{G})" + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": " the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "text", + "content": "N. Kolotouros et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 171, + 118, + 203, + 192 + ], + "blocks": [ + { + "bbox": [ + 171, + 118, + 203, + 192 + ], + "lines": [ + { + "bbox": [ + 171, + 118, + 203, + 192 + ], + "spans": [ + { + "bbox": [ + 171, + 118, + 203, + 192 + ], + "type": "image", + "image_path": "1db3f8e20ecd95376e3f40ffa57af103cd1f5e34035f9d69a43d3c1d082b2377.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 252, + 119, + 279, + 191 + ], + "blocks": [ + { + "bbox": [ + 252, + 119, + 279, + 191 + ], + "lines": [ + { + "bbox": [ + 252, + 119, + 279, + 191 + ], + "spans": [ + { + "bbox": [ + 252, + 119, + 279, + 191 + ], + "type": "image", + "image_path": "01db6990a2368276c251bea01ea04cb4d3d3bdd0a1562e42583b377a7562c54b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 329, + 119, + 359, + 194 + ], + "blocks": [ + { + "bbox": [ + 329, + 119, + 359, + 194 + ], + "lines": [ + { + "bbox": [ + 329, + 119, + 359, + 194 + ], + "spans": [ + { + "bbox": [ + 329, + 119, + 359, + 194 + ], + "type": "image", + "image_path": "69caf4714bb22336466c3167fea99aa8a152d2eb57025c327c4df4a0a4e78ad4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 409, + 119, + 438, + 194 + ], + "blocks": [ + { + "bbox": [ + 409, + 119, + 438, + 194 + ], + "lines": [ + { + "bbox": [ + 409, + 119, + 438, + 194 + ], + "spans": [ + { + "bbox": [ + 409, + 119, + 438, + 194 + ], + "type": "image", + "image_path": "581460b48d12d79363f37c039c9742e8218ee53639a8e58fb60d61cc7cd679ab.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 171, + 205, + 203, + 280 + ], + "blocks": [ + { + "bbox": [ + 171, + 205, + 203, + 280 + ], + "lines": [ + { + "bbox": [ + 171, + 205, + 203, + 280 + ], + "spans": [ + { + "bbox": [ + 171, + 205, + 203, + 280 + ], + "type": "image", + "image_path": "cade929a17ec102891f1cc8fcf8972281ae09fad17a6875f4a71ac34d99e57bd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 293, + 482, + 337 + ], + "lines": [ + { + "bbox": [ + 131, + 293, + 482, + 337 + ], + "spans": [ + { + "bbox": [ + 131, + 293, + 482, + 337 + ], + "type": "text", + "content": "Fig. 3: Diverse back view hypotheses. Conditioned on the front view, our method is able to generate diverse plausible back views of the person, with different hairstyles, wrinkle patterns, or lighting. Our network can also be controlled with text (second row), to add fine-grained detail to our generated back-side views." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 251, + 205, + 282, + 279 + ], + "blocks": [ + { + "bbox": [ + 269, + 196, + 345, + 205 + ], + "lines": [ + { + "bbox": [ + 269, + 196, + 345, + 205 + ], + "spans": [ + { + "bbox": [ + 269, + 196, + 345, + 205 + ], + "type": "text", + "content": "No text conditioning" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 251, + 205, + 282, + 279 + ], + "lines": [ + { + "bbox": [ + 251, + 205, + 282, + 279 + ], + "spans": [ + { + "bbox": [ + 251, + 205, + 282, + 279 + ], + "type": "image", + "image_path": "dc4a18690be7fdbe678f416b58d57f4d67372038af4ac46a7e6022b1cd0c7b98.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 329, + 205, + 361, + 279 + ], + "blocks": [ + { + "bbox": [ + 329, + 205, + 361, + 279 + ], + "lines": [ + { + "bbox": [ + 329, + 205, + 361, + 279 + ], + "spans": [ + { + "bbox": [ + 329, + 205, + 361, + 279 + ], + "type": "image", + "image_path": "e2a400732757cac23a4a9922daf6f5839c31b9736fdd90dda848c9c8b4941f10.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 280, + 370, + 289 + ], + "lines": [ + { + "bbox": [ + 321, + 280, + 370, + 289 + ], + "spans": [ + { + "bbox": [ + 321, + 280, + 370, + 289 + ], + "type": "text", + "content": "+“gray hair”" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 409, + 205, + 440, + 279 + ], + "blocks": [ + { + "bbox": [ + 409, + 205, + 440, + 279 + ], + "lines": [ + { + "bbox": [ + 409, + 205, + 440, + 279 + ], + "spans": [ + { + "bbox": [ + 409, + 205, + 440, + 279 + ], + "type": "image", + "image_path": "1daf589e7a6556f159cd04ac5d4f59caf3364c3b6768c12b31433d9c80a1db40.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 280, + 456, + 289 + ], + "lines": [ + { + "bbox": [ + 392, + 280, + 456, + 289 + ], + "spans": [ + { + "bbox": [ + 392, + 280, + 456, + 289 + ], + "type": "text", + "content": "+“with a tattoo”" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 354, + 482, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 379 + ], + "type": "text", + "content": "latent encoding of the dense GHUM signal, and " + }, + { + "bbox": [ + 130, + 354, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathrm{enc}}" + }, + { + "bbox": [ + 130, + 354, + 482, + 379 + ], + "type": "text", + "content": " the encoder subset of the denoising UNet parameters " + }, + { + "bbox": [ + 130, + 354, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 130, + 354, + 482, + 379 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 396, + 275, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 396, + 275, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 396, + 275, + 407 + ], + "type": "text", + "content": "3.2 Back View Generation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 414, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 582 + ], + "type": "text", + "content": "One could try to lift the generated front views from the previous stage to 3D directly by applying a single-image 3D reconstruction method like PHORHUM [8]. However, because of the inherent ambiguity of the problem this will result in significant loss of geometric detail and blurry textures for the non-visible body surfaces. To avoid this, we propose to fine-tune again a latent diffusion network with the same strategy as in the previous section. This time the additional image conditioning is a front view and optionally a text prompt, and we train the network to learn the distribution of back views conditioned on the front view. The additional text prompt can be used in cases where it is desired to additionally guide the generation by very specific properties. Fig. 3 shows different back sides sampled from the conditional distribution. We also show that the additional text inputs are useful in modulating certain parts of the generation that are not immediately deducible from the front image, such as hairstyles or specific patterns." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 600, + 292, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 600, + 292, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 600, + 292, + 611 + ], + "type": "text", + "content": "3.3 3D Reconstruction Model" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Our 3D reconstruction network is inspired by PHORHUM [8], and our design choices are informed by the limitations of typical single-image 3D reconstruction methods. Specifically, given a collection of input image signals " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{I_f, I_b, \\mathcal{G}\\}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", we first concatenate them, and then use a convolutional encoder " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " to compute" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "type": "text", + "content": "Instant 3D Human Avatar Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "a pixel-aligned feature map " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "G(\\mathcal{I})" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ". The 3D body control signal " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " is optional and may be omitted, e.g. for single-image reconstruction. Then, each point " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " in the scene gets projected on this feature map to get pixel-aligned features " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathbf{x}} = g(\\mathcal{I}, \\mathbf{x}; \\pi) = b(G(\\mathcal{I}), \\pi(\\mathbf{x}))" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " using interpolation, where " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "b(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " is the bilinear sampling operator and " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\pi(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " is the pixel location of the projection of " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " using the camera " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ". These pixel aligned features are then concatenated with a positional encoding " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\gamma(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " of the 3D point and are fed to an MLP " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " that outputs the signed distance from the surface " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " as well as surface color " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ". Finally, the 3D shape " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " is represented as the zero-level-set of " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "d" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 211, + 232, + 481, + 246 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 232, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 211, + 232, + 481, + 246 + ], + "type": "interline_equation", + "content": "\\mathcal {S} (\\mathcal {I}) = \\left\\{\\mathbf {x} \\in \\mathbb {R} ^ {3} | f (g (I, \\mathbf {x}; \\pi), \\gamma (\\mathbf {x})) = (0, c) \\right\\}. \\tag {3}", + "image_path": "c3e5b52a16f33d6c624b6fef78aa3263ee5213add40f1981e2a1b41346470f06.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 253, + 436, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 253, + 436, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 253, + 436, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 131, + 253, + 436, + 266 + ], + "type": "text", + "content": " can be transformed to a mesh directly, using Marching Cubes [44]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 282, + 326, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 282, + 326, + 294 + ], + "spans": [ + { + "bbox": [ + 131, + 282, + 326, + 294 + ], + "type": "text", + "content": "3.4 Animation of Generated Avatars" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 299, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 482, + 407 + ], + "type": "text", + "content": "Our method can generate diverse 3D avatars with various poses, shapes and appearances. Optionally, we may leverage the conditioning body model to rig the estimated 3D shape. As a result of our conditioning strategy, 3D avatars and the conditional body model instances are aligned in 3D. This allows us to anchor the reconstructed 3D shape on the body model surface [10] and re-pose or re-shape it accordingly. Alternatively, we may also transfer just the LBS skeleton and weights from the body model to the scan. This enables importing and animating the generated 3D assets in various rendering engines. See Fig. 8 and Sup. Mat. for examples and videos of animations of the generated 3D assets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 425, + 230, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 230, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 230, + 438 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "content": "Data. We use meshes from RenderPeople [1] for training as well as our own captured data, totaling in " + }, + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\sim 10\\mathrm{K}" + }, + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "content": " scans with diverse poses, body shapes, and clothing styles. We render each scan with randomly sampled HDRI background, random cloth color augmentations, and lighting using Blender [11]. During this process, we render both front and back views used to train the different stages of our model. For the front image generation network we also use a set of " + }, + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "inline_equation", + "content": "10\\mathrm{K}" + }, + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "content": " real images on which we fitted the GHUM model using 2D keypoints. For testing we defined a split based on subject identity and held out " + }, + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\sim 1\\mathrm{K}" + }, + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "content": " scans." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": "We provide results for 2 different versions of our model. The standard quality model is generated in 2 seconds by running 5 DDIM [64] steps during inference and Marching Cubes at " + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "inline_equation", + "content": "256^{3}" + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": " resolution. The high quality model is generated in 10 seconds using 50 DDIM steps and Marching Cubes at " + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "inline_equation", + "content": "512^{3}" + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": " resolution. The timings were recorded on a single 40 GB A100 GPU. Unless otherwise stated, all results we report are obtained using the high quality model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Metrics and Baselines. We compare our method numerically for two different problems. First we consider the task of text-to-3D human generation, where we sample 100 different text prompts and compare against representative text-to-3D generation methods. For numerical comparisons we evaluate the text-image" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "text", + "content": "N. Kolotouros et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 154, + 115, + 462, + 255 + ], + "blocks": [ + { + "bbox": [ + 154, + 115, + 462, + 255 + ], + "lines": [ + { + "bbox": [ + 154, + 115, + 462, + 255 + ], + "spans": [ + { + "bbox": [ + 154, + 115, + 462, + 255 + ], + "type": "image", + "image_path": "8c9a1eba9b64f3018c4f5159767cf74820e47964bd0031a4a402609ea2f6e9e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 256, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 131, + 256, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 131, + 256, + 482, + 291 + ], + "type": "text", + "content": "Fig. 4: Diversity of our 3D generation. For the same text prompt and the same pose and shape conditioning, our model can generate a diverse set of 3D avatars that respect both the text and the 3D body controls." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 309, + 480, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 480, + 334 + ], + "type": "text", + "content": "alignment using CLIP [52]. Specifically, we use the retrieval accuracy using CLIP, as proposed in [51]. We additionally show qualitative results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 335, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 502 + ], + "type": "text", + "content": "Second, we validate the performance of our 3D reconstruction component against state-of-the-art methods [8,27,60,61,70,71] considering both geometry and texture. Pixel-aligned image features dominate in recent work [8,60,61], but some methods proposed to combine them with body models [70,71], which offer advantages in being able to animate reconstructions. We also rely on pixel-aligned features, yet propose a method that inherently enables animation. We run AvatarPopUp by generating the back side of the subject and applying the reconstruction network. To evaluate 3D geometry, we report bi-directional Chamfer distance " + }, + { + "bbox": [ + 130, + 335, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\times 10^{-3}" + }, + { + "bbox": [ + 130, + 335, + 482, + 502 + ], + "type": "text", + "content": ", Normal Consistency (NC ↑), and Volumetric Intersection over Union (IoU ↑) after ICP alignment. However, these metrics do not necessarily correlate with good visual quality, e.g. Chamfer distance is minimized by smooth, non-detailed geometry. To measure the quality of reconstructions, we additionally report FID scores [22] of the front/back views for both geometry and texture." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 529, + 276, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 529, + 276, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 529, + 276, + 540 + ], + "type": "text", + "content": "4.1 Text-to-3D Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 555, + 481, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 481, + 592 + ], + "type": "text", + "content": "In Fig. 4 we generate different avatars given the same text prompt and driving poses. We can see that our model is able to create a very diverse set of assets, a property not observed in the previous text-to-3D generation methods [32]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In Tab. 2 we use CLIP to evaluate our model against other text-to-3D generation methods. In general, CLIP-based metrics are not indicative of the generated image quality, because they only consider the alignment with the text, and often over-saturated images with extreme details tend to have high CLIP scores. To further demonstrate that our method generates higher quality avatars, we also include a qualitative comparison in Fig. 5." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "type": "text", + "content": "Instant 3D Human Avatar Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 185, + 106, + 425, + 176 + ], + "blocks": [ + { + "bbox": [ + 185, + 106, + 425, + 176 + ], + "lines": [ + { + "bbox": [ + 185, + 106, + 425, + 176 + ], + "spans": [ + { + "bbox": [ + 185, + 106, + 425, + 176 + ], + "type": "table", + "html": "
ColorColorRuntime ↓
R-Prec. ↑Top-3 ↑R-Prec. ↑Top-3 ↑
0.680.920.040.258hDreamHuman [32]
0.560.820.030.153hTADA [37]
--0.030.083mCHUPA [31]
0.580.730.080.17~2sOurs
0.620.770.110.17~10sOurs (high quality)
", + "image_path": "d5b878beddc754fe7938b70945f5feb2ef3b940e4f77b225bc761548b3f8a018.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 180, + 481, + 213 + ], + "lines": [ + { + "bbox": [ + 131, + 180, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 131, + 180, + 481, + 213 + ], + "type": "text", + "content": "Table 2: Numerical comparisons with other text-to-3d human generation methods. We mark the best and second best results. Our method allows to trade-off speed and quality and we report results for two different settings." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 151, + 222, + 465, + 381 + ], + "blocks": [ + { + "bbox": [ + 151, + 222, + 465, + 381 + ], + "lines": [ + { + "bbox": [ + 151, + 222, + 465, + 381 + ], + "spans": [ + { + "bbox": [ + 151, + 222, + 465, + 381 + ], + "type": "image", + "image_path": "dca817aad1b5c4cf1303324e456c60a0543ca2f879737932aa8e2ffeef5c575e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 382, + 423, + 392 + ], + "lines": [ + { + "bbox": [ + 180, + 382, + 423, + 392 + ], + "spans": [ + { + "bbox": [ + 180, + 382, + 423, + 392 + ], + "type": "text", + "content": "A person wearing a black off-shoulder top and a denim skirt" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 395, + 482, + 439 + ], + "lines": [ + { + "bbox": [ + 130, + 395, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 395, + 482, + 439 + ], + "type": "text", + "content": "Fig. 5: Comparisons with text-to-3d human generation methods. Our method generates high quality results that respect the text prompt well, at a fraction of the others' runtime, cf. Tab. 3. TADA's results appear unnatural; DreamHuman failed for one subject and produces oversaturated colors; CHUPA failed to respect the prompt." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 462, + 322, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 462, + 322, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 462, + 322, + 474 + ], + "type": "text", + "content": "4.2 Single-image 3D Reconstruction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 665 + ], + "type": "text", + "content": "While not specifically designed for 3D reconstruction, our method achieves state-of-the-art performance also for this task. The evaluation setup is the following: given an input image " + }, + { + "bbox": [ + 130, + 498, + 482, + 665 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 498, + 482, + 665 + ], + "type": "text", + "content": ", we draw one random sample from the back view image generator network, and then feed the pair of front/back images to our 3D reconstruction network. For all methods we extract a textured 3D mesh using Marching Cubes and report numerical results in Tab. 3. Furthermore, we show qualitative results in Fig. 6. Notably, our method not only performs on par numerically and qualitatively on reconstructed front views, but also generates highly detailed back view texture and geometry. Finally, we also compare with the optimization-based method TeCH [27] in Fig. 7. TeCH produces detailed front and back geometry but also also exhibits problems at times, rooted in its 3D pose estimation method. Most importantly, TeCH runs for several hours per instance, while ours computes results in a single feed-forward pass, in only a few seconds." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 247, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 247, + 100 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 247, + 100 + ], + "type": "text", + "content": "N. Kolotouros et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 114, + 451, + 193 + ], + "blocks": [ + { + "bbox": [ + 159, + 114, + 451, + 193 + ], + "lines": [ + { + "bbox": [ + 159, + 114, + 451, + 193 + ], + "spans": [ + { + "bbox": [ + 159, + 114, + 451, + 193 + ], + "type": "table", + "html": "
3D MetricsFID ↓
Ch. ↓IoU ↑NC ↑Color. FColor B.Normal F.Normal B.
7.10.500.72--29.657.2PIFu [60]
6.20.520.74--26.952.5PIFuHD [61]
7.00.480.72--36.149.0ICON [71]
3.40.580.76--18.336.1ECON [70]
0.840.710.8711.040.811.238.8PHORHUM* [8]
0.910.700.8611.020.012.825.8Ours
", + "image_path": "af7d81b77ae1ecfe223fe1e32c6a416f3f1c426307bb8a859318d6685aa752c5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 198, + 482, + 263 + ], + "lines": [ + { + "bbox": [ + 130, + 198, + 482, + 263 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 482, + 263 + ], + "type": "text", + "content": "Table 3: Numerical comparisons with single-view 3D reconstructions methods and ablations of our method. We mark the best and second best results. All Chamfer metrics are " + }, + { + "bbox": [ + 130, + 198, + 482, + 263 + ], + "type": "inline_equation", + "content": "\\times 10^{-3}" + }, + { + "bbox": [ + 130, + 198, + 482, + 263 + ], + "type": "text", + "content": ". Not all methods generate colors. For fair comparisons, we retrained PHORHUM [8] using the same data as AvatarPopUp. We observe comparable results in terms of 3D metrics, however ours performs better on generating more realistic and diverse back views and back normals." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 277, + 256, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 256, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 256, + 290 + ], + "type": "text", + "content": "4.3 3D Virtual Try On" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 299, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 482, + 491 + ], + "type": "text", + "content": "An immediate application of our method is the option to perform 3D garment edits for a given identity. Given an input image, we first recover 3D pose and shape parameters of the person in the image. Using a text prompt and the identity preservation strategy introduced below, we can generate updated images of the same person wearing different garments or accessories. To preserve the identity of the person, we first locate the person's head in the source image and use then Repaint [46] to out-paint a novel body for the given head while still conditioning on the estimated pose and shape parameters. With this strategy we ensure that our method not only preserves the facial characteristics but also the overall body proportions. In Fig. 9 we illustrate such editing examples. The generated 3D edits present garment details like wrinkles on both front and back views, and preserve the subjects' facial appearance. Also, note that body shape and identity are well preserved for the subjects, even though only one image is given. While there has been a significant amount of 2D virtual try-on research [20,35,86], our methodology can generate consistent and highly detailed 3D meshes that can be animated and rendered from other viewpoints." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 509, + 215, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 215, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 215, + 520 + ], + "type": "text", + "content": "4.4 Animation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 529, + 481, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 481, + 566 + ], + "type": "text", + "content": "AvatarPopUp enables animation of the generated assets by design, provided that its generation is conditioned on an underlying body model. In Fig. 8, we show an example of a generated avatar that is rigged automatically." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 584, + 238, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 584, + 238, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 584, + 238, + 596 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Pose and shape encoding in the 3D reconstruction network. We compare the effectiveness of our additional pose and shape encoding inputs " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " to our reconstruction network. To do so, we use the same set of 100 text prompts as before and for each text prompt sample a random pose and shape configuration. For each " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "(\\tau, \\theta, \\beta)" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " triplet we run inferences and compute 2 meshes: one using" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "type": "text", + "content": "Instant 3D Human Avatar Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 106, + 161, + 171 + ], + "blocks": [ + { + "bbox": [ + 132, + 106, + 161, + 171 + ], + "lines": [ + { + "bbox": [ + 132, + 106, + 161, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 106, + 161, + 171 + ], + "type": "image", + "image_path": "33f7851c4d76bff20fb1c6fd45e6c8df67519d153221809c899f5db5aef1c0c4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 136, + 173, + 156, + 233 + ], + "blocks": [ + { + "bbox": [ + 136, + 173, + 156, + 233 + ], + "lines": [ + { + "bbox": [ + 136, + 173, + 156, + 233 + ], + "spans": [ + { + "bbox": [ + 136, + 173, + 156, + 233 + ], + "type": "image", + "image_path": "8e76543a34a53fe154685c8f8fdab35449968cff6ecd28b41931c8a9e6862184.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 137, + 237, + 156, + 296 + ], + "blocks": [ + { + "bbox": [ + 137, + 237, + 156, + 296 + ], + "lines": [ + { + "bbox": [ + 137, + 237, + 156, + 296 + ], + "spans": [ + { + "bbox": [ + 137, + 237, + 156, + 296 + ], + "type": "image", + "image_path": "1a277a21efc4e57daeed803f8b0c412f8ffdb6a95d5267679749727b5d40ff08.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 315, + 481, + 348 + ], + "lines": [ + { + "bbox": [ + 131, + 315, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 131, + 315, + 481, + 348 + ], + "type": "text", + "content": "Fig. 6: Qualitative comparisons with state-of-the-art single image 3D reconstruction methods. Our method produces front color and normals on par with state-of-the-art and much more detailed back view hypotheses." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 175, + 108, + 282, + 171 + ], + "blocks": [ + { + "bbox": [ + 175, + 108, + 282, + 171 + ], + "lines": [ + { + "bbox": [ + 175, + 108, + 282, + 171 + ], + "spans": [ + { + "bbox": [ + 175, + 108, + 282, + 171 + ], + "type": "image", + "image_path": "e79dc5f2dedb4b3ffdf3bf42737cd570d004c8ecaa1c8c7ac072c867c4af560b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 178, + 173, + 198, + 233 + ], + "blocks": [ + { + "bbox": [ + 178, + 173, + 198, + 233 + ], + "lines": [ + { + "bbox": [ + 178, + 173, + 198, + 233 + ], + "spans": [ + { + "bbox": [ + 178, + 173, + 198, + 233 + ], + "type": "image", + "image_path": "8ac23b6f8be52bac373b46279c603a437dc6cf667e7ff312e21282fce33fae41.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 178, + 237, + 197, + 296 + ], + "blocks": [ + { + "bbox": [ + 178, + 237, + 197, + 296 + ], + "lines": [ + { + "bbox": [ + 178, + 237, + 197, + 296 + ], + "spans": [ + { + "bbox": [ + 178, + 237, + 197, + 296 + ], + "type": "image", + "image_path": "c005d6f61016cf555a1f4332c9f7c4c534e90b5f546040205e819479751e93b7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 202, + 173, + 250, + 234 + ], + "blocks": [ + { + "bbox": [ + 202, + 173, + 250, + 234 + ], + "lines": [ + { + "bbox": [ + 202, + 173, + 250, + 234 + ], + "spans": [ + { + "bbox": [ + 202, + 173, + 250, + 234 + ], + "type": "image", + "image_path": "577b67f56e47e08f57816e503d5e515d093325fc5537f9cf62057260038e20df.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 203, + 237, + 248, + 297 + ], + "blocks": [ + { + "bbox": [ + 203, + 237, + 248, + 297 + ], + "lines": [ + { + "bbox": [ + 203, + 237, + 248, + 297 + ], + "spans": [ + { + "bbox": [ + 203, + 237, + 248, + 297 + ], + "type": "image", + "image_path": "f446eb4225152364a096a181f1464f51327cc608605628e28236310c96cdbe4a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 299, + 236, + 308 + ], + "lines": [ + { + "bbox": [ + 216, + 299, + 236, + 308 + ], + "spans": [ + { + "bbox": [ + 216, + 299, + 236, + 308 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 256, + 108, + 282, + 171 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 108, + 282, + 171 + ], + "spans": [ + { + "bbox": [ + 256, + 108, + 282, + 171 + ], + "type": "image", + "image_path": "c814c687d6515aa79a1185de69493943908e9c054c10932080fb5a2a3d6c0c72.jpg" + } + ] + } + ], + "index": 11, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 257, + 173, + 277, + 233 + ], + "blocks": [ + { + "bbox": [ + 257, + 173, + 277, + 233 + ], + "lines": [ + { + "bbox": [ + 257, + 173, + 277, + 233 + ], + "spans": [ + { + "bbox": [ + 257, + 173, + 277, + 233 + ], + "type": "image", + "image_path": "cd0de6d86362a5fa18e9385ccd15422ac54614390d4ab7ec8682c73ee8863a87.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 257, + 237, + 277, + 297 + ], + "blocks": [ + { + "bbox": [ + 257, + 237, + 277, + 297 + ], + "lines": [ + { + "bbox": [ + 257, + 237, + 277, + 297 + ], + "spans": [ + { + "bbox": [ + 257, + 237, + 277, + 297 + ], + "type": "image", + "image_path": "f2ac061c35fa60b0f87b0fb223d1065cf808553e8c4e96565f8bb564ea206bfa.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 305, + 106, + 371, + 168 + ], + "blocks": [ + { + "bbox": [ + 305, + 106, + 371, + 168 + ], + "lines": [ + { + "bbox": [ + 305, + 106, + 371, + 168 + ], + "spans": [ + { + "bbox": [ + 305, + 106, + 371, + 168 + ], + "type": "image", + "image_path": "1ae583bb3a87dcbd16e91e626ca5c29c9236a0d206d8a46b537de1b18d5eb1d4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 171, + 372, + 233 + ], + "blocks": [ + { + "bbox": [ + 307, + 171, + 372, + 233 + ], + "lines": [ + { + "bbox": [ + 307, + 171, + 372, + 233 + ], + "spans": [ + { + "bbox": [ + 307, + 171, + 372, + 233 + ], + "type": "image", + "image_path": "964b0c392b24ab8509e4eaeeb90c3449626aa9910e7945c0e6981bc61f1c079d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 307, + 236, + 371, + 297 + ], + "blocks": [ + { + "bbox": [ + 307, + 236, + 371, + 297 + ], + "lines": [ + { + "bbox": [ + 307, + 236, + 371, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 236, + 371, + 297 + ], + "type": "image", + "image_path": "1e00fcc6ec50608af2f4641d9ee6804db03edbd9f36657dea13c35cb78157725.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 298, + 375, + 307 + ], + "lines": [ + { + "bbox": [ + 326, + 298, + 375, + 307 + ], + "spans": [ + { + "bbox": [ + 326, + 298, + 375, + 307 + ], + "type": "text", + "content": "PHORHUM" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 377, + 106, + 406, + 168 + ], + "blocks": [ + { + "bbox": [ + 377, + 106, + 406, + 168 + ], + "lines": [ + { + "bbox": [ + 377, + 106, + 406, + 168 + ], + "spans": [ + { + "bbox": [ + 377, + 106, + 406, + 168 + ], + "type": "image", + "image_path": "8c5920f736078ba8bc63186005d4bedecea707dd22e2bc532eb429d6c33ffd8e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 374, + 172, + 392, + 232 + ], + "blocks": [ + { + "bbox": [ + 374, + 172, + 392, + 232 + ], + "lines": [ + { + "bbox": [ + 374, + 172, + 392, + 232 + ], + "spans": [ + { + "bbox": [ + 374, + 172, + 392, + 232 + ], + "type": "image", + "image_path": "cc7e1d9099b9174a99c2ac3d57a8518d4fd370fcd087890a32876bc2d53b6234.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 375, + 237, + 394, + 297 + ], + "blocks": [ + { + "bbox": [ + 375, + 237, + 394, + 297 + ], + "lines": [ + { + "bbox": [ + 375, + 237, + 394, + 297 + ], + "spans": [ + { + "bbox": [ + 375, + 237, + 394, + 297 + ], + "type": "image", + "image_path": "39bb2b5585ebbd15757b5ecbd220da8505d42aa5141e57c8a4c0a52d23f86a0b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 298, + 393, + 307 + ], + "lines": [ + { + "bbox": [ + 375, + 298, + 393, + 307 + ], + "spans": [ + { + "bbox": [ + 375, + 298, + 393, + 307 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 424, + 107, + 478, + 170 + ], + "blocks": [ + { + "bbox": [ + 424, + 107, + 478, + 170 + ], + "lines": [ + { + "bbox": [ + 424, + 107, + 478, + 170 + ], + "spans": [ + { + "bbox": [ + 424, + 107, + 478, + 170 + ], + "type": "image", + "image_path": "208dab8bc67d85f3a927ccaa1a57637c436ee52f0d30e42e31b6b96ad1e42a62.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 427, + 173, + 446, + 234 + ], + "blocks": [ + { + "bbox": [ + 427, + 173, + 446, + 234 + ], + "lines": [ + { + "bbox": [ + 427, + 173, + 446, + 234 + ], + "spans": [ + { + "bbox": [ + 427, + 173, + 446, + 234 + ], + "type": "image", + "image_path": "7d0ed829c20412a073bdd09e851b60956f11e3e14aa9fb83ef930a400ff22c74.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 454, + 173, + 476, + 234 + ], + "blocks": [ + { + "bbox": [ + 454, + 173, + 476, + 234 + ], + "lines": [ + { + "bbox": [ + 454, + 173, + 476, + 234 + ], + "spans": [ + { + "bbox": [ + 454, + 173, + 476, + 234 + ], + "type": "image", + "image_path": "a95d4bd010c91addafda3846475302ee32d4172e6e61044918dd164d4b4da8b8.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 428, + 237, + 456, + 297 + ], + "blocks": [ + { + "bbox": [ + 428, + 237, + 456, + 297 + ], + "lines": [ + { + "bbox": [ + 428, + 237, + 456, + 297 + ], + "spans": [ + { + "bbox": [ + 428, + 237, + 456, + 297 + ], + "type": "image", + "image_path": "227939bc09641de63cfda27840ce207abcaf506027c3ad240eb91d9fee96c425.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 434, + 299, + 468, + 307 + ], + "lines": [ + { + "bbox": [ + 434, + 299, + 468, + 307 + ], + "spans": [ + { + "bbox": [ + 434, + 299, + 468, + 307 + ], + "type": "text", + "content": "PIFuHD" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 130, + 365, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 480, + 437 + ], + "type": "text", + "content": "only the front and back images, and another one additionally using the dense GHUM encodings. We then evaluate the Chamfer distance between each mesh and the corresponding GHUM mesh. The model using the additional GHUM signals has an average Chamfer distance of " + }, + { + "bbox": [ + 130, + 365, + 480, + 437 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{with}} = 1.4" + }, + { + "bbox": [ + 130, + 365, + 480, + 437 + ], + "type": "text", + "content": " whereas the one without " + }, + { + "bbox": [ + 130, + 365, + 480, + 437 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{without}} = 8.6" + }, + { + "bbox": [ + 130, + 365, + 480, + 437 + ], + "type": "text", + "content": ", thus validating our design choice. Not only is the control respected well, but this also allows for animation as discussed previously." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 130, + 438, + 481, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 481, + 522 + ], + "type": "text", + "content": "Partial vs. Complete Network Finetuning. We finetune two Latent Diffusion networks, one in a standard way by optimizing all parameters, and another one using our proposed strategy where we only finetune the convolutional layers of the encoder. Empirically we observed that the network that was finetuned as a whole experienced catastrophic forgetting, and has poor performance when asked to generate types of garment not seen in the training set. Fig. 10 shows a comparison for text prompts not in the training set." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 132, + 541, + 306, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 541, + 306, + 554 + ], + "spans": [ + { + "bbox": [ + 132, + 541, + 306, + 554 + ], + "type": "text", + "content": "5 Discussion & Conclusions" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 130, + 568, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 480, + 616 + ], + "type": "text", + "content": "Limitations. AvatarPopUp inherits limitations from other pixel-aligned methods, e.g. with regions parallel to camera rays being not as detailed in the reconstruction. Further, artifacts might be visible in poses or very loose clothing that are underrepresented during training." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "Ethical Considerations. We present a generative tool to create 3D human assets, thus reducing the risk of scanning and using real humans for training large scale 3D generative models. AvatarPopUp generates diverse results and can lead to a better coverage of subject distributions." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "type": "text", + "content": "N. Kolotouros et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 131, + 167, + 439 + ], + "blocks": [ + { + "bbox": [ + 130, + 131, + 167, + 439 + ], + "lines": [ + { + "bbox": [ + 130, + 131, + 167, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 131, + 167, + 439 + ], + "type": "image", + "image_path": "644893664bb6a0e81b86e5f6ed2631977fd60acfe99dfcc4712156cb815c5dff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 453, + 482, + 486 + ], + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 486 + ], + "type": "text", + "content": "Fig. 7: Additional comparisons with TeCH. TeCH is optimization-based and takes several hours to complete for a single reconstruction. Our results are obtained in seconds, yet are as detailed and even less noisy." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 194, + 132, + 325, + 439 + ], + "blocks": [ + { + "bbox": [ + 194, + 132, + 325, + 439 + ], + "lines": [ + { + "bbox": [ + 194, + 132, + 325, + 439 + ], + "spans": [ + { + "bbox": [ + 194, + 132, + 325, + 439 + ], + "type": "image", + "image_path": "1e5512458355ed8fa858558aa05d8784f3f158ff317bff63f0750b309f745733.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 440, + 273, + 449 + ], + "lines": [ + { + "bbox": [ + 250, + 440, + 273, + 449 + ], + "spans": [ + { + "bbox": [ + 250, + 440, + 273, + 449 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 353, + 132, + 477, + 439 + ], + "blocks": [ + { + "bbox": [ + 353, + 132, + 477, + 439 + ], + "lines": [ + { + "bbox": [ + 353, + 132, + 477, + 439 + ], + "spans": [ + { + "bbox": [ + 353, + 132, + 477, + 439 + ], + "type": "image", + "image_path": "add9dc208474d3fb82997f320290655b0af98f036ed2a320767ba0541b6c766f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 439, + 432, + 449 + ], + "lines": [ + { + "bbox": [ + 403, + 439, + 432, + 449 + ], + "spans": [ + { + "bbox": [ + 403, + 439, + 432, + 449 + ], + "type": "text", + "content": "TeCH" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 184, + 533, + 422, + 617 + ], + "blocks": [ + { + "bbox": [ + 184, + 533, + 422, + 617 + ], + "lines": [ + { + "bbox": [ + 184, + 533, + 422, + 617 + ], + "spans": [ + { + "bbox": [ + 184, + 533, + 422, + 617 + ], + "type": "image", + "image_path": "798f16ce9656d504b756b7a0d45247a153e8448229982b1f4b3b01101a7f5eeb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 624, + 482, + 647 + ], + "lines": [ + { + "bbox": [ + 130, + 624, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 482, + 647 + ], + "type": "text", + "content": "Fig. 8: Reposing example.. We first reconstruct an avatar \"wearing a gray suit\" in the A-pose and then we transfer it to different poses." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "type": "text", + "content": "Instant 3D Human Avatar Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 106, + 465, + 299 + ], + "blocks": [ + { + "bbox": [ + 149, + 106, + 465, + 299 + ], + "lines": [ + { + "bbox": [ + 149, + 106, + 465, + 299 + ], + "spans": [ + { + "bbox": [ + 149, + 106, + 465, + 299 + ], + "type": "image", + "image_path": "ca6e8a5633e19c806361af05dbdef45340e11facaea29dab86c9e17767da3fa7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 303, + 482, + 360 + ], + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 360 + ], + "type": "text", + "content": "Fig. 9: Identity preserving 3D avatar editing. Our method allows for editing the clothing, while preserving the identity of the generated person. In each row we're using the image on the left as input, recover the person's pose and shape parameters and then run our method using the shown prompts. The generated avatars share the same identity and respect the target prompts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 178, + 368, + 439, + 446 + ], + "blocks": [ + { + "bbox": [ + 178, + 368, + 439, + 446 + ], + "lines": [ + { + "bbox": [ + 178, + 368, + 439, + 446 + ], + "spans": [ + { + "bbox": [ + 178, + 368, + 439, + 446 + ], + "type": "image", + "image_path": "7fbc485cb9f842c8d7873adff5ae0bea9ddd8f3ed28deed93f6287a79c08f91e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 449, + 482, + 495 + ], + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 495 + ], + "type": "text", + "content": "Fig. 10: Partial vs. complete fine-tuning strategy. For each text prompt the first example is generated using our partial fine-tuning strategy whereas the second by fine-tuning the entire network. Empirically, partial fine-tuning is more resilient to catastrophic forgetting, generating more diverse and better text-aligned images." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "We have presented AvatarPopUp, a novel framework for generating 3D human avatars controlled by text or images and yielding rigged 3D models in 2-10 seconds. AvatarPopUp is purely feed-forward, allows for fine-grained control over the generated body pose and shape, and can produce multiple qualitatively different hypotheses. AvatarPopUp is composed of a cascade of expert systems, decoupling image generation and 3D lifting. Through this design choice, AvatarPopUp benefits from both web-scale image datasets, ensuring high generation diversity, and from smaller size accurate 3D datasets, resulting in reconstructions with increased detail and precisely controlled based on text and identity specifications. In the future, we would like to explore other 3D construction strategies besides pixel-aligned features. Longer term, we aim to support highly detailed and controllable 3D human model generation for entertainment, education, architecture and art, or medical applications." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "type": "text", + "content": "N. Kolotouros et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 137, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 137, + 137, + 269, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 137, + 269, + 148 + ], + "spans": [ + { + "bbox": [ + 137, + 137, + 269, + 148 + ], + "type": "text", + "content": "1. https://renderpeople.com/" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 148, + 481, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 148, + 481, + 179 + ], + "spans": [ + { + "bbox": [ + 137, + 148, + 481, + 179 + ], + "type": "text", + "content": "2. Abdal, R., Yifan, W., Shi, Z., Xu, Y., Po, R., Kuang, Z., Chen, Q., Yeung, D.Y., Wetzstein, G.: Gaussian shell maps for efficient 3d human generation. arXiv preprint arXiv:2311.17857 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 180, + 481, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 180, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 138, + 180, + 481, + 201 + ], + "type": "text", + "content": "3. AlBahar, B., Saito, S., Tseng, H.Y., Kim, C., Kopf, J., Huang, J.B.: Single-image 3d human digitization with shape-guided diffusion. In: SIGGRAPH Asia (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 202, + 481, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 202, + 481, + 223 + ], + "spans": [ + { + "bbox": [ + 138, + 202, + 481, + 223 + ], + "type": "text", + "content": "4. Alldieck, T., Magnor, M., Bhatnagar, B.L., Theobalt, C., Pons-Moll, G.: Learning to reconstruct people in clothing from a single rgb camera. In: CVPR (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 224, + 481, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 481, + 244 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 481, + 244 + ], + "type": "text", + "content": "5. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Detailed human avatars from monocular video. In: 3DV (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 245, + 481, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 481, + 266 + ], + "type": "text", + "content": "6. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Video based reconstruction of 3D people models. In: CVPR (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 266, + 481, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 481, + 287 + ], + "type": "text", + "content": "7. Alldieck, T., Pons-Moll, G., Theobalt, C., Magnor, M.: Tex2shape: Detailed full human body geometry from a single image. In: ICCV (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 288, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 288, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 288, + 481, + 308 + ], + "type": "text", + "content": "8. Alldieck, T., Zanfir, M., Sminchisescu, C.: Photorealistic monocular 3D reconstruction of humans wearing clothing. In: CVPR (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 309, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 481, + 341 + ], + "type": "text", + "content": "9. Bar-Tal, O., Chefer, H., Tov, O., Herrmann, C., Paiss, R., Zada, S., Ephrat, A., Hur, J., Li, Y., Michaeli, T., et al.: Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 342, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 481, + 373 + ], + "type": "text", + "content": "0. Bazavan, E.G., Zanfir, A., Zanfir, M., Freeman, W.T., Sukthankar, R., Sminchisescu, C.: Hspace: Synthetic parametric humans animated in complex environments. arXiv (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 373, + 481, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 373, + 481, + 405 + ], + "spans": [ + { + "bbox": [ + 138, + 373, + 481, + 405 + ], + "type": "text", + "content": "1. Blender Online Community: Blender - a 3D modelling and rendering package. Blender Foundation, Blender Institute, Amsterdam (2020), http://www.blender.org" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 406, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 406, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 134, + 406, + 481, + 482 + ], + "type": "text", + "content": "12. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., Kolesnikov, A., Puigcerver, J., Ding, N., Rong, K., Akbari, H., Mishra, G., Xue, L., Thapliyal, A., Bradbury, J., Kuo, W., Seyedhosseini, M., Jia, C., Ayan, B.K., Riquelme, C., Steiner, A., Angelova, A., Zhai, X., Houlsby, N., Soricut, R.: Pali: A jointly-scaled multilingual language-image model (2022). https://doi.org/10.48550/ARXIV.2209.06794, https://arxiv.org/abs/2209.06794" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 483, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 483, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 134, + 483, + 481, + 514 + ], + "type": "text", + "content": "13. Corona, E., Zanfir, M., Alldieck, T., Bazavan, E.G., Zanfir, A., Sminchisescu, C.: Structured 3d features for reconstructing relightable and animatable avatars. In: CVPR (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 514, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 514, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 134, + 514, + 481, + 557 + ], + "type": "text", + "content": "14. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3d objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13142-13153 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 558, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 558, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 134, + 558, + 481, + 578 + ], + "type": "text", + "content": "15. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 579, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 579, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 134, + 579, + 481, + 611 + ], + "type": "text", + "content": "16. Dong, Z., Xu Chen, J.Y., J.Black, M., Hilliges, O., Geiger, A.: AG3D: Learning to generate 3D avatars from 2D image collections. In: International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 612, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 612, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 134, + 612, + 481, + 633 + ], + "type": "text", + "content": "17. Gabeur, V., Franco, J.S., Martin, X., Schmid, C., Rogez, G.: Moulding humans: Non-parametric 3D human shape estimation from single images. In: ICCV (2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "text", + "content": "18. Gong, C., Dai, Y., Li, R., Bao, A., Li, J., Yang, J., Zhang, Y., Li, X.: Text2 avatar: Text to 3d human avatar generation with codebook-driven body controllable attribute. arXiv preprint arXiv:2401.00711 (2024)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "type": "text", + "content": "Instant 3D Human Avatar Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "19. Han, X., Cao, Y., Han, K., Zhu, X., Deng, J., Song, Y.Z., Xiang, T., Wong, K.Y.K.: Headsculpt: Crafting 3d head avatars with text. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "20. Han, X., Wu, Z., Wu, Z., Yu, R., Davis, L.S.: Viton: An image-based virtual try-on network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7543-7552 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "21. He, T., Xu, Y., Saito, S., Soatto, S., Tung, T.: Arch++: Animation-ready clothed human reconstruction revisited. In: CVPR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "22. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "23. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "text", + "content": "24. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "type": "text", + "content": "25. Hong, F., Chen, Z., LAN, Y., Pan, L., Liu, Z.: EVA3d: Compositional 3d human generation from 2d image collections. In: International Conference on Learning Representations (2023), https://openreview.net/forum?id=g7U9jD_2CUr" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "type": "text", + "content": "26. Hong, F., Zhang, M., Pan, L., Cai, Z., Yang, L., Liu, Z.: Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. ACM Transactions on Graphics (TOG) 41(4), 1-19 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "text", + "content": "27. Huang, Y., Yi, H., Xiu, Y., Liao, T., Tang, J., Cai, D., Thies, J.: TeCH: Text-guided Reconstruction of Lifelike Clothed Humans. In: International Conference on 3D Vision (3DV) (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "28. Huang, Z., Xu, Y., Lassner, C., Li, H., Tung, T.: Arch: Animatable reconstruction of clothed humans. In: CVPR (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "29. Jiang, R., Wang, C., Zhang, J., Chai, M., He, M., Chen, D., Liao, J.: Avatarcraft: Transforming text into neural human avatars with parameterized shape and pose control. arXiv preprint arXiv:2303.17606 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "text", + "content": "30. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "text", + "content": "31. Kim, B., Kwon, P., Lee, K., Lee, M., Han, S., Kim, D., Joo, H.: Chupa: Carving 3d clothed humans from skinned shape priors using 2d diffusion probabilistic models. arXiv preprint arXiv:2305.11870 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "type": "text", + "content": "32. Kolotouros, N., Alldieck, T., Zanfir, A., Bazavan, E., Fieraru, M., Sminchisescu, C.: Dreamhuman: Animatable 3d avatars from text. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 567 + ], + "type": "text", + "content": "33. Kolotouros, N., Pavlakos, G., Black, M.J., Daniilidis, K.: Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2252-2261 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "text", + "content": "34. Kondratyuk, D., Yu, L., Gu, X., Lezama, J., Huang, J., Hornung, R., Adam, H., Akbari, H., Alon, Y., Birodkar, V., et al.: Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "35. Lee, S., Gu, G., Park, S., Choi, S., Choo, J.: High-resolution virtual try-on with misalignment and occlusion-handled conditions. In: European Conference on Computer Vision. pp. 204-219. Springer (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "36. Lei, B., Yu, K., Feng, M., Cui, M., Xie, X.: Diffusiongan3d: Boosting text-guided 3d generation and domain adaption by combining 3d gans and diffusion priors. arXiv preprint arXiv:2312.16837 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "type": "text", + "content": "N. Kolotouros et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "37. Liao, T., Yi, H., Xiu, Y., Tang, J., Huang, Y., Thies, J., Black, M.J.: Tada! text to animatable digital avatars. In: 3DV (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 182 + ], + "type": "text", + "content": "38. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "text", + "content": "39. Liu, H., Wang, X., Wan, Z., Shen, Y., Song, Y., Liao, J., Chen, Q.: Headartist: Text-conditioned 3d head generation with self score distillation. arXiv preprint arXiv:2312.07539 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "type": "text", + "content": "40. Liu, M., Xu, C., Jin, H., Chen, L., Varma T, M., Xu, Z., Su, H.: One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "type": "text", + "content": "41. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9298-9309 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "type": "text", + "content": "42. Liu, X., Zhan, X., Tang, J., Shan, Y., Zeng, G., Lin, D., Liu, X., Liu, Z.: Human-gaussian: Text-driven 3d human generation with gaussian splatting. arXiv preprint arXiv:2311.17061 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 314, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 335 + ], + "type": "text", + "content": "43. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ToG (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 335, + 482, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 357 + ], + "type": "text", + "content": "44. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3D surface construction algorithm. SIGGRAPH (1987)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "type": "text", + "content": "45. Lorraine, J., Xie, K., Zeng, X., Lin, C.H., Takikawa, T., Sharp, N., Lin, T.Y., Liu, M.Y., Fidler, S., Lucas, J.: Att3d: Amortized text-to-3d object synthesis. arXiv preprint arXiv:2306.07349 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 390, + 482, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 434 + ], + "type": "text", + "content": "46. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11461-11471 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 434, + 482, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 468 + ], + "type": "text", + "content": "47. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. ECCV (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 468, + 482, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 500 + ], + "type": "text", + "content": "48. Onizuka, H., Hayirci, Z., Thomas, D., Sugimoto, A., Uchiyama, H., Taniguchi, R.i.: Tetratsdf: 3D human reconstruction from a single image with a tetrahedral outer shell. In: CVPR (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 500, + 482, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 482, + 533 + ], + "type": "text", + "content": "49. Oord, A.v.d., Dieleman, S., Zen, H., Simonyan, K., Vinyals, O., Graves, A., Kalchbrenner, N., Senior, A., Kavukcuoglu, K.: Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 533, + 482, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 566 + ], + "type": "text", + "content": "50. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: CVPR (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 566, + 482, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 566, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 566, + 482, + 588 + ], + "type": "text", + "content": "51. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: Int. Conf. Learn. Represent. (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 588, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 588, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 482, + 632 + ], + "type": "text", + "content": "52. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "type": "text", + "content": "53. Raj, A., Kaza, S., Poole, B., Niemeyer, M., Mildenhall, B., Ruiz, N., Zada, S., Aberman, K., Rubenstein, M., Barron, J., Li, Y., Jampani, V.: Dreambooth3d: Subject-driven text-to-3d generation. ICCV (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "type": "text", + "content": "Instant 3D Human Avatar Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "54. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "55. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 216 + ], + "type": "text", + "content": "56. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "57. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 293 + ], + "type": "text", + "content": "58. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS. vol. 35, pp. 36479-36494 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 335 + ], + "type": "text", + "content": "59. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "60. Saito, S., Huang, Z., Natsume, R., Morishima, S., Kanazawa, A., Li, H.: Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In: ICCV (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "61. Saito, S., Simon, T., Saragih, J., Joo, H.: Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3D human digitization. In: CVPR (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "text", + "content": "62. Sengupta, A., Alldieck, T., Kolotouros, N., Corona, E., Zanfir, A., Sminchisescu, C.: Diffhuman: Probabilistic photorealistic 3D reconstruction of humans. In: CVPR (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "63. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "text", + "content": "64. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: Int. Conf. Learn. Represent. (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "text", + "content": "65. Sun, J., Zhang, B., Shao, R., Wang, L., Liu, W., Xie, Z., Liu, Y.: Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior. arXiv preprint arXiv:2310.16818 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "66. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "text", + "content": "67. Varol, G., Ceylan, D., Russell, B., Yang, J., Yumer, E., Laptev, I., Schmid, C.: Bodynet: Volumetric inference of 3D human body shapes. In: ECCV (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "text", + "content": "68. Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "text", + "content": "69. Wang, J., Liu, Y., Dou, Z., Yu, Z., Liang, Y., Li, X., Wang, W., Xie, R., Song, L.: Disentangled clothed avatar generation from text descriptions. arXiv preprint arXiv:2312.05295 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "text", + "content": "70. Xiu, Y., Yang, J., Cao, X., Tzionas, D., Black, M.J.: ECON: Explicit Clothed humans Optimized via Normal integration. In: CVPR (June 2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "71. Xiu, Y., Yang, J., Tzionas, D., Black, M.J.: Icon: Implicit clothed humans obtained from normals. In: CVPR (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 100 + ], + "type": "text", + "content": "N. Kolotouros et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 556 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "72. Xu, H., Bazavan, E.G., Zanfir, A., Freeman, W.T., Sukthankar, R., Sminchisescu, C.: Ghum & ghuml: Generative 3D human shape and articulated pose models. In: CVPR (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "text", + "content": "73. Xu, Y., Yang, Z., Yang, Y.: Seeavatar: Photorealistic text-to-3d avatar generation with constrained geometry and appearance. arXiv preprint arXiv:2312.08889 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 481, + 216 + ], + "type": "text", + "content": "74. Yang, Z., Wang, S., Manivasagam, S., Huang, Z., Ma, W.C., Yan, X., Yumer, E., Urtasun, R.: S3: Neural shape, skeleton, and skinning fields for 3D human modeling. In: CVPR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 481, + 237 + ], + "type": "text", + "content": "75. Yanting Chan, K., Lin, G., Zhao, H., Lin, W.: Integratedpifu: Integrated pixel aligned implicit function for single-view human reconstruction. In: ECCV (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 481, + 259 + ], + "type": "text", + "content": "76. Zeng, Y., Wei, G., Zheng, J., Zou, J., Wei, Y., Zhang, Y., Li, H.: Make pixels dance: High-dynamic video generation. arXiv preprint arXiv:2311.10982 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 481, + 293 + ], + "type": "text", + "content": "77. Zhang, C., Zhang, C., Zheng, S., Zhang, M., Qamar, M., Bae, S.H., Kweon, I.S.: Audio diffusion model for speech synthesis: A survey on text to speech and speech enhancement in generative ai. arXiv preprint arXiv:2303.13336 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 481, + 325 + ], + "type": "text", + "content": "78. Zhang, H., Feng, Y., Kulits, P., Wen, Y., Thies, J., Black, M.J.: Text-guided generation and editing of compositional 3d avatars. arXiv preprint arXiv:2309.07125 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 325, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 481, + 358 + ], + "type": "text", + "content": "79. Zhang, H., Chen, B., Yang, H., Qu, L., Wang, X., Chen, L., Long, C., Zhu, F., Du, K., Zheng, M.: Avatarverse: High-quality & stable 3d avatar creation from text and pose. arXiv preprint arXiv:2308.03610 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 481, + 392 + ], + "type": "text", + "content": "80. Zhang, J., Zhang, X., Zhang, H., Liew, J.H., Zhang, C., Yang, Y., Feng, J.: Avatarstudio: High-fidelity and animatable 3d avatar creation from text. arXiv preprint arXiv:2311.17917 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 481, + 413 + ], + "type": "text", + "content": "81. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. IEEE International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 413, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 481, + 446 + ], + "type": "text", + "content": "82. Zhao, Z., Bao, Z., Li, Q., Qiu, G., Liu, K.: Psavatar: A point-based morphable shape model for real-time head avatar creation with 3d gaussian splatting. arXiv preprint arXiv:2401.12900 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 446, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 481, + 468 + ], + "type": "text", + "content": "83. Zheng, Z., Yu, T., Liu, Y., Dai, Q.: Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. PAMI (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 481, + 490 + ], + "type": "text", + "content": "84. Zheng, Z., Yu, T., Wei, Y., Dai, Q., Liu, Y.: Deephuman: 3D human reconstruction from a single image. In: ICCV (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 490, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 481, + 512 + ], + "type": "text", + "content": "85. Zhu, H., Zuo, X., Wang, S., Cao, X., Yang, R.: Detailed human shape estimation from a single image by hierarchical mesh deformation. In: CVPR (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 512, + 481, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 481, + 556 + ], + "type": "text", + "content": "86. Zhu, L., Yang, D., Zhu, T., Reda, F., Chan, W., Sahara, C., Norouzi, M., Kemelmacher-Shlizerman, I.: Tryondiffusion: A tale of two unets. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4606-4615 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 291, + 91, + 447, + 101 + ], + "type": "text", + "content": "Instant 3D Human Avatar Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator/6c9ab90c-9fb8-405a-a3f6-eaa68a4cea3f_content_list.json b/2024/Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator/6c9ab90c-9fb8-405a-a3f6-eaa68a4cea3f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..189d8850896bdd3d0c6656d559acd02f9dca7955 --- /dev/null +++ b/2024/Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator/6c9ab90c-9fb8-405a-a3f6-eaa68a4cea3f_content_list.json @@ -0,0 +1,1502 @@ +[ + { + "type": "text", + "text": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator", + "text_level": 1, + "bbox": [ + 217, + 140, + 785, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Niki Amini-Naieni1, Tomas Jakab1, Andrea Vedaldi1, and Ronald Clark1", + "bbox": [ + 225, + 212, + 772, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Oxford", + "bbox": [ + 431, + 239, + 571, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Neural Radiance Fields (NeRFs) have markedly improved novel view synthesis, but accurate uncertainty quantification in their image predictions remains an open problem. The prevailing methods for estimating uncertainty, including the state-of-the-art Density-aware NeRF Ensembles (DANE) [29], quantify uncertainty without calibration. This frequently leads to over- or under-confidence in image predictions, which can undermine their real-world applications. In this paper, we propose a method which, for the first time, achieves calibrated uncertainties for NeRFs. To accomplish this, we overcome a significant challenge in adapting existing calibration techniques to NeRFs: a need to hold out ground truth images from the target scene, reducing the number of images left to train the NeRF. This issue is particularly problematic in sparse-view settings, where we can operate with as few as three images. To address this, we introduce the concept of a meta-calibrator that performs uncertainty calibration for NeRFs with a single forward pass without the need for holding out any images from the target scene. Our meta-calibrator is a neural network that takes as input the NeRF images and uncalibrated uncertainty maps and outputs a scene-specific calibration curve that corrects the NeRF's uncalibrated uncertainties. We show that the meta-calibrator can generalize on unseen scenes and achieves well-calibrated and state-of-the-art uncertainty for NeRFs, significantly beating DANE and other approaches. This opens opportunities to improve applications that rely on accurate NeRF uncertainty estimates such as next-best view planning and potentially more trustworthy image reconstruction for medical diagnosis.", + "bbox": [ + 261, + 296, + 738, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: NeRFs $\\cdot$ Uncertainty calibration $\\cdot$ Few-shot learning", + "bbox": [ + 261, + 657, + 697, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 699, + 375, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in scene representations have led to promising new approaches for novel view synthesis and scene reconstruction. Among these, Neural Radiance Fields (NeRFs) [16] have emerged as a particularly powerful tool, offering unprecedented levels of realism and detail in rendered images. The core idea behind NeRFs is to represent a scene as a continuous vector-valued function, parameterized by a neural network, that maps spatial coordinates and view directions to color and density values. This approach enables the creation", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/931aa5f9312feb8cc27913728d089273215c2cacdf110ac6aad8da7b08aa1d22.jpg", + "image_caption": [ + "Fig. 1: We propose a method for efficiently calibrating the uncertainties from NeRF models. Our approach is based on a meta-calibrator that takes as input features from the rendered NeRF images and uncalibrated uncertainty maps and predicts the calibration function, $R_{\\theta}(\\cdot)$ , for the NeRF model. Our meta-calibrator generalizes across scenes, so it only needs to be trained once, and can predict the calibration function in a single forward pass without any ground truth data from the target scene." + ], + "image_footnote": [], + "bbox": [ + 300, + 142, + 702, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of detailed 3D representations from sets of 2D images, revolutionizing 3D reconstruction.", + "bbox": [ + 212, + 453, + 787, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, despite their impressive capabilities, traditional NeRF models lack an essential component: an accurate measure of uncertainty in their predictions. Accurate uncertainties are crucial for applying NeRFs to safety-critical problems such as MRI image reconstruction from sparse data [7], where unreliable confidence estimates could lead to misdiagnosis. More accurate uncertainties could also enhance practical methods such as uncertainty-guided next-best view planning techniques [9]. Prior approaches have attempted to estimate NeRF uncertainties [5, 14, 26, 27, 29], but they all overlook the problem of calibration. Thus, the uncertainties they output are not as accurate as they could be.", + "bbox": [ + 212, + 501, + 787, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In particular, the state-of-the-art uncertainty estimation method for NeRFs, Density-aware NeRF Ensembles (DANE) [29], produces uncalibrated uncertainties. As a result, the confidence intervals and variances do not match the true confidence, meaning it has limited applicability to real-world problems. This constraint is significant as it restricts the use of NeRFs in safety-critical and sparse-data settings, where knowing the confidence in predictions is crucial.", + "bbox": [ + 212, + 640, + 787, + 732 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The best NeRF methods for the sparse-view setting overlook the problem of calibration as well. FlipNeRF [24], the state-of-the-art technique for sparse-view reconstruction, uses uncertainties from an uncalibrated mixture of Laplacians to enhance its training process. Therefore, the uncertainties it outputs at inference as an artifact of training are not accurate.", + "bbox": [ + 212, + 750, + 787, + 825 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 227, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "N. Amini-Naieni et al.", + "bbox": [ + 271, + 114, + 421, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present a novel approach for obtaining calibrated uncertainties for NeRF models in the sparse-view setting. Our strategy integrates the Laplacian mixture from FlipNeRF [24] with the calibration techniques by Kuleshov et al [12]. However, naively applying the calibration method by Kuleshov et al to FlipNeRF does not work due to a significant challenge of the sparse-view setting: there is a lack of held-out data from the target scene for fitting the calibrator. Specifically, holding out just one image for calibration could decrease the size of the training set by over $30\\%$ , resulting in significant performance degradation of the NeRF. To overcome this, we make use of a unique observation: while calibration curves exhibit significant variation across scenes, they also demonstrate a significant regularity in their structure. Utilizing this insight, we propose the concept of a meta-calibrator that learns a low-dimensional representation of the NeRF calibration curves and infers them from scene features. We motivate and show why this meta-calibrator is necessary and demonstrate that it achieves more accurate uncertainties than DANE [29] without holding out any images from the target scene.", + "bbox": [ + 212, + 146, + 792, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, our contributions are: (1) the first investigation into obtaining calibrated uncertainties from NeRFs, (2) a novel meta-calibrator for fitting the calibration model without using held-out data, and (3) experiments on the real-world Local Light Field Fusion (LLFF) [15] and DTU [8] datasets showing that our meta-calibrator achieves state-of-the-art and well-calibrated uncertainties for real scenes. We also demonstrate that our uncertainties can be leveraged for effective next-best view planning.", + "bbox": [ + 212, + 404, + 789, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 542, + 387, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRFs). NeRFs [16] are a popular method for novel view synthesis. From a set of 2D images, NeRFs learn a neural network representation of a single scene. A trained NeRF model outputs estimates of the volume density and emitted radiance at any 3D location and viewing direction. Novel views can be generated by applying volume rendering [10] to the density and radiance values predicted by the NeRF model for points along rays cast into the scene. Due to their simplicity and impressive performance, NeRFs have become a popular technique for solving a variety of rendering problems.", + "bbox": [ + 212, + 580, + 787, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Over the last few years, several extensions of NeRFs have been explored [11, 18, 21, 28, 31]. These include speeding up training and inference [13, 23, 30], modeling dynamic scenes [3], learning from a sparse set of training views [1, 6, 19, 24, 25], and estimating the uncertainty in NeRF predictions [5, 14, 22, 26, 27, 29]. Sparse NeRF methods aim to accurately render novel views when only a few training views are available from the target scene. NeRF uncertainty estimation techniques strive to accurately predict the confidence in the views rendered. Although uncertainty estimation is particularly important in the sparse-view", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "setting, where NeRF renderings are especially unreliable, the main aim of sparse NeRF methods is not to output accurate uncertainties.", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sparse-view NeRF Methods. Despite this, recent sparse NeRF methods do produce uncertainties as an artifact of their training process. Both MixNeRF [25] and FlipNeRF [24], the state-of-the-art approach, model the RGB color channels given a ray as independent random variables that follow a mixture model. FlipNeRF further uses the uncertainties of the pixel colors to regularize the training process, producing more accurate image reconstructions at inference. However, neither MixNeRF nor FlipNeRF outputs calibrated uncertainties.", + "bbox": [ + 215, + 191, + 785, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method significantly extends sparse NeRF methods to obtain more accurate and well-calibrated uncertainties at inference. To benefit from the superior performance of FlipNeRF [24] at sparse novel view synthesis, we apply the proposed meta-calibrator to the learned distribution from FlipNeRF, producing more accurate uncertainties without sacrificing state-of-the-art image quality. However, our approach can be applied to any NeRF method that outputs uncertainties, so it is distinct from FlipNeRF and MixNeRF. In essence, the proposed meta-calibrator augments sparse NeRF methods to achieve state-of-the-art uncertainty, beating both FlipNeRF and techniques designed explicitly for NeRF uncertainty estimation.", + "bbox": [ + 215, + 311, + 785, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Uncertainty in NeRFs. The growing line of methods specifically designed for accurately estimating NeRF uncertainties [5,14,22,26,27,29] do not address the problem of calibration. The current state-of-the-art uncertainty estimation technique for NeRFs is Density-aware NeRF Ensembles (DANE) [29]. DANE adds an epistemic uncertainty term to a naive ensembles approach with five ensemble members. Thus, DANE is very costly as it requires training five NeRFs to obtain uncertainty estimates. Another NeRF uncertainty estimation approach is Stochastic Neural Radiance Fields (S-NeRF) [27], which learns a probability distribution over all possible radiance fields by modeling the volume density and radiance as random variables that follow a joint distribution. S-NeRF employs variational inference to sample from an approximation to this distribution and uses the variances of the sampled pixel colors as the estimated uncertainties. Conditional-flow NeRF (CF-NeRF) [26] builds on S-NeRF by combining latent variable modeling and conditional normalizing flows to relax the strong constraints S-NeRF imposes over the radiance distribution. Despite the growing number of techniques in this area of study, all prior work does not consider calibration, outputting unreliable uncertainties as a result. Our work achieves more accurate uncertainties than these prior methods by filling the gap of uncertainty calibration for NeRFs and drawing on well-established techniques in calibrated regression [12].", + "bbox": [ + 215, + 477, + 785, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Uncertainty Calibration in Deep Learning. While uncertainty calibration has been studied for Bayesian deep learning methods [4, 12], it has not been adapted for or applied successfully to NeRF uncertainty estimates. This may be", + "bbox": [ + 215, + 794, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "N. Amini-Naieni et al.", + "bbox": [ + 271, + 114, + 421, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "because NeRFs introduce additional complexity in the uncertainty estimation process as the neural network model needs to be trained per-scene.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This makes uncertainty calibration challenging as methods for calibrated regression [12] require either using the training set or held-out data to achieve calibrated uncertainties. Using the training set to fit the calibrator results in severe overfitting (see Sec. 2 of the Supplementary). Holding out data from the target scene for calibration means there is less data to train the NeRF, making it more inaccurate at novel view synthesis (see Sec. 3 of the Supplementary). Thus, a trivial application of [12] to NeRFs would not be satisfactory. In this work, we propose the concept of a meta-calibrator that, in contrast to [12], does not require held-out data and achieves calibrated uncertainty estimates for NeRFs.", + "bbox": [ + 212, + 176, + 784, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 351, + 330, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we present a method that calibrates NeRF uncertainties. To this end, we propose a novel meta-calibrator that accepts uncalibrated NeRF uncertainties and predicted images as inputs and outputs a scene-specific calibration curve, correcting the uncalibrated confidence levels. An overview of our method can be seen in Fig. 1. Crucially, our approach does not require holding out any images from the target scene. Thus, it can be applied to sparse-view settings, where holding out a single image could significantly harm the NeRF's performance. In Sec. 3.1, we explain the necessary background concepts, in Sec. 3.2, we describe how we obtain the corrected uncertainty values, and in Sec. 3.3 we detail our meta-calibrator.", + "bbox": [ + 212, + 383, + 787, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 214, + 556, + 372, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRFs). Neural Radiance Fields (NeRFs) [16] represent a scene as a continuous vector-valued function with inputs a Cartesian point $\\mathbf{x} = (x,y,z)$ and unit viewing direction vector $\\mathbf{d} = (u,v,w)$ and outputs an emitted radiance $\\pmb {c} = (\\nu ,\\varrho ,\\theta)$ and volume density $\\sigma$ . By optimizing the weights $\\Theta$ of a neural network approximation $\\mathbf{F}_{\\Theta}$ to this representation, NeRFs can render the color of any pixel in a synthetic image of the scene. To achieve this, principles from classical volume rendering [10] are applied to the radiance and density values estimated by $\\mathbf{F}_{\\Theta}$ for points along a ray cast from the origin $\\mathbf{x_0}$ of the virtual camera, through the pixel, and into the scene. More specifically, the expected color $\\mathbf{c}(\\mathbf{r})$ of a camera ray $\\mathbf{r}(t) = \\mathbf{x_0} + t\\mathbf{d}$ with near and far bounds $t_n$ and $t_f$ is:", + "bbox": [ + 212, + 582, + 787, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} (\\mathbf {r}) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (\\mathbf {r} (t)) \\boldsymbol {c} (\\mathbf {r} (t), \\mathbf {d}) d t, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 750, + 785, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T(t) = e^{-\\int_{t_n}^t \\sigma(\\mathbf{r}(s)) ds}$ . The integral in Eq. (1) is estimated with numerical quadrature to obtain the colors of the pixels in the synthetic image from the NeRF model outputs. NeRF optimizes $\\Theta$ according to Eq. (1) with gradient", + "bbox": [ + 212, + 791, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 782, + 125 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "descent, since the numerical quadrature is differentiable. Conventional NeRFs do not provide an uncertainty associated with their predictions, so extensions like DANE [29] have been developed.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Base NeRF Uncertainties. To obtain the initial uncertainties, we have two options for our base model: (1) DANE [29], the state-of-the-art method for NeRF uncertainty estimation or (2) FlipNeRF [24], the state-of-the-art method for sparse novel view synthesis. DANE requires training five NeRFs per scene and provides poor image quality in the sparse-view setting, so we choose FlipNeRF. We show in the experiments that applying our meta-calibrator to FlipNeRF results in more accurate and well-calibrated uncertainties than those output by DANE. Our base FlipNeRF uncertainties are inferred from a mixture of Laplacians with location and scale parameters learned during training.", + "bbox": [ + 212, + 200, + 784, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FlipNeRF [24] models the joint distribution of the color $\\mathbf{C} = (R,G,B)$ given a ray $\\mathbf{r}$ with a mixture of Laplacians:", + "bbox": [ + 212, + 352, + 782, + 383 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np (\\mathbf {C} = \\mathbf {c} | \\mathbf {r}) = \\sum_ {j = 1} ^ {M} \\pi_ {j} \\mathcal {F} (\\mathbf {C} = \\mathbf {c}; \\boldsymbol {\\mu} _ {j}, \\boldsymbol {\\beta} _ {j}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 396, + 784, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $M$ is the number of sampled points along the ray $r$ . $\\mathcal{F}(\\mathbf{C} = \\mathbf{c};\\pmb{\\mu}_j,\\pmb{\\beta}_j)$ is the 3D Laplacian probability density with location parameter $\\pmb{\\mu}_{j} = (\\mu_{j}^{R},\\mu_{j}^{G},\\mu_{j}^{B})$ and scale parameter $\\beta_{i} = (\\beta_{j}^{R},\\beta_{j}^{G},\\beta_{j}^{B})$ evaluated at the color $\\mathbf{c}$ . More specifically, the mixing coefficients $\\{\\pi_j\\}_{j = 1}^M$ are the normalized coefficients of the radiance values along a ray in Eq. (2), the location parameters $\\{\\pmb{\\mu}_{j}\\}_{j = 1}^{M}$ are the estimated RGB radiance values, and the scale parameters $\\{\\pmb{\\beta}_{j}\\}_{j = 1}^{M}$ are an additional output of the model. These parameters are optimized by FlipNeRF during training.", + "bbox": [ + 212, + 440, + 784, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We now go into detail about how the location and scale parameters are obtained. FlipNeRF [24] learns these parameters by minimizing the negative log-likelihood of the training set $\\mathcal{D} = \\{(\\mathbf{r_i},\\mathbf{c_i})\\}_{i = 1}^N$ containing the rays $\\mathbf{r_i}$ and colors $\\mathbf{c_i}$ from the pixels in the ground truth images of the scene assuming the distribution in Eq. (2). FlipNeRF additionally minimizes the average scale parameters through an auxiliary uncertainty-aware emptiness loss for reducing floating artifacts. The negative log-likelihood loss and the uncertainty-aware emptiness loss are added to FlipNeRF's total training loss, which incorporates other terms such as the mean squared error. As a result of this training process, the location and scale parameters can be inferred by FlipNeRF at new poses. We obtain the uncalibrated confidence levels of predicted ray colors from these location and scale parameters.", + "bbox": [ + 212, + 568, + 784, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using the distribution in Eq. (2), we can easily compute the confidence level for a given ray $\\mathbf{r_t}$ , which we denote as $F_{t}$ . The CDF for the Laplacian mixture has a closed form expression parameterized by the location and scales output by the pretrained FlipNeRF [24]. Thus, we can use this CDF to predict the confidence level of the ground truth color $\\mathbf{c_t}$ of any ray by evaluating it at the given", + "bbox": [ + 212, + 763, + 782, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "N. Amini-Naieni et al.", + "bbox": [ + 271, + 114, + 421, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "color value, $p_t = F_t(\\mathbf{c_t})$ . However, these initial confidence levels are uncalibrated and, hence, inaccurate.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Calibrated regression. In [12], Kuleshov et al extend calibration methods for classification to regression. They define a forecaster $H: \\mathcal{X} \\to (\\mathcal{Y} \\to [0,1])$ as a function that outputs for each $x_{t} \\in \\mathcal{X}$ , a CDF $F_{t}$ . Given a pretrained forecaster $H$ , they suggest training an auxiliary model $R: [0,1] \\to [0,1]$ by fitting $R$ to a recalibration dataset $D = \\left\\{\\left([H(x_{t})](y_{t}), \\hat{P}([H(x_{t})](y_{t}))\\right)\\right\\}_{t=1}^{T}$ , where", + "bbox": [ + 212, + 203, + 784, + 292 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {P} (p) = \\left| \\left\\{y _ {t}: [ H \\left(x _ {t}\\right) ] \\left(y _ {t}\\right) \\leq p \\text {f o r} t = 1, \\dots , T \\right\\} \\right| / T\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 301, + 679, + 318 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "is the empirical confidence level corresponding to the predicted confidence level $p$ . The fitted $R$ forms the calibration curve that corrects the expected confidence levels. We can now obtain predictive posterior values that closely match the true confidences using $\\hat{F}_t \\equiv R \\circ F_t$ for the test data.", + "bbox": [ + 212, + 325, + 782, + 387 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "However, as mentioned before, there are numerous challenges associated with applying this recalibration procedure. Firstly, note that it requires ground-truth values $(y_{t})$ for the predictions we are recalibrating. This means that in order to prevent overfitting we need to reserve part of the training dataset specifically for fitting the calibrator (which leaves less data for training the NeRF — a significant issue if we only have a few input views). Secondly, it does not actually prescribe how to compute a suitable uncertainty value from the predicted distribution. In Sec. 3.3, we address the former issue, and in Sec. 3.2, we address the latter.", + "bbox": [ + 212, + 401, + 784, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Calculating uncertainty", + "text_level": 1, + "bbox": [ + 214, + 571, + 455, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The predictive posterior in Eq. (2) provides a distribution over likely ray colors for a NeRF model, but it does not inherently offer a straightforward metric for quantifying uncertainty at a specific point in the reconstruction. Intuitively, as the variance of this distribution increases, so does the uncertainty of the model's output at that point. Therefore, the variance or standard deviation is a popular choice for quantifying the uncertainty [24,26,27,29]. However, in the case of the corrected mixture distribution obtained from raymarching, this can be slow to compute especially if it has to be done for each pixel. Therefore, we turn to a metric that can be calculated directly from the calibrated CDF $\\hat{F}^t$ .", + "bbox": [ + 212, + 594, + 784, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We propose to use the interquartile range of each calibrated distribution:", + "bbox": [ + 240, + 744, + 766, + 762 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\kappa^ {C} \\left(\\mathbf {r} _ {\\mathbf {t}}\\right) = \\left[ \\hat {F} _ {t} ^ {C} \\right] ^ {- 1} \\left(\\frac {3}{4}\\right) - \\left[ \\hat {F} _ {t} ^ {C} \\right] ^ {- 1} \\left(\\frac {1}{4}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 768, + 784, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\kappa^C (\\mathbf{r_t})$ represents the uncertainty at a ray $\\mathbf{r_t}$ in the color channel $C$ . This difference provides a measure of the statistical dispersion and thus serves", + "bbox": [ + 214, + 808, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator", + "bbox": [ + 285, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d4266a2b88d6ec9833101b28ab0e4ae1b897b32bf656aaae7ef21df07d7cbd39.jpg", + "image_caption": [ + "(a) Creating a parametric model of the calibration curves." + ], + "image_footnote": [], + "bbox": [ + 233, + 145, + 491, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fc20e2f750952699e6d84c4ea763729a17e3d80e0cab4ff95679774442b6f859.jpg", + "image_caption": [ + "(b) Predicting the curve parameters.", + "Fig. 2: Meta-calibrator design. In stage (a) we fit a low-dimensional parametric model of the calibration curves. The meta-calibrator then predicts these curve parameters from rendered images of the scene and their associated uncalibrated uncertainty maps (b)." + ], + "image_footnote": [], + "bbox": [ + 509, + 183, + 769, + 271 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "as a robust measure of the spread of the output channel. By averaging the interquartile range over the color channels, we obtain a single scalar value that effectively quantifies the uncertainty of the NeRF model for the given ray. As shown in our experiments, this method is very computational efficient, and it provides an accurate estimate of uncertainty.", + "bbox": [ + 212, + 378, + 787, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Meta-calibrator", + "text_level": 1, + "bbox": [ + 214, + 473, + 392, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To overcome the challenge that, especially in the sparse-view setting, there is no held-out data available for fitting the calibrator, we propose a novel metacalibrator that infers the calibration curves from uncalibrated NeRF predictions. To do this, we leverage the insight that the calibration curves demonstrate significant regularity. We posit a low-dimensional model of the calibration curves can be learned and predicted using the images and uncalibrated uncertainty maps inferred by the NeRF, enabling us to estimate the calibration function without evaluating the empirical confidence levels using held-out data from the target scene. We now describe this meta-calibrator (illustrated in Fig. 2) in detail.", + "bbox": [ + 212, + 494, + 787, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A Parametric Model for Calibration Curves. We first fit a low-dimensional representation of the calibration curves using Principal Component Analysis (PCA). To create the training set for learning this representation, we sample held-out images from $K$ scenes and apply the calibration procedure by Kuleshov et al. [12] to form $K$ ground truth calibration curves. To construct the training vector $\\mathbf{v_k} \\in \\mathbb{R}^{1 \\times M}$ for scene $k \\leq K$ , we sample $M$ evenly spaced points along its ground truth calibration curve. We find that fitting the PCA model using only a few scenes (21 in our case) provides a good enough approximation to capture the variation in the test curves (see Sec. 4). Here, $\\mathbf{V} = [\\mathbf{v}_k] \\in \\mathbb{R}^{K \\times M}$ contains the ground truth calibration curves for the training scenes, with $K$ representing the number of curves and $M$ the sample count along each curve. PCA is then used to determine the basis vectors $\\mathbf{U} = (\\mathbf{u_1}, \\mathbf{u_2}, \\dots, \\mathbf{u_n})$ and coefficients $\\theta = (\\alpha_1, \\alpha_2, \\dots, \\alpha_n)$ , so that each calibration curve can be represented as:", + "bbox": [ + 212, + 643, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 227, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "N. Amini-Naieni et al.", + "bbox": [ + 271, + 114, + 421, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\mathbf{v}_k = \\sum_{i=1}^n \\alpha_i \\mathbf{u}_i$ . The parameters $\\theta$ fully describe the calibration functions.", + "bbox": [ + 212, + 145, + 753, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To find the optimal number of components, we compute the explained variance, and find that in our case most of the variance is explained using only $n = 3$ components (see experiments). To ensure the calibrator is monotonically increasing, we derive the final calibration function $R_{\\theta}(\\cdot)$ using isotonic regression applied to the curve approximated by $\\pmb{\\theta} \\cdot \\mathbf{U}$ . The idea is that the low-dimensional representation of the calibration curves encoded in $\\mathbf{U}$ will generalise to new target scenes without any additional scene-specific data.", + "bbox": [ + 212, + 176, + 787, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Predicting Calibration Parameters. What remains is to estimate the calibration parameters, $\\theta$ , for a new scene. As we do not want to use additional held-out data from the target scene, we propose predicting these parameters using scene-specific features computed from the pretrained NeRF outputs. This approach is motivated by the human ability to visually identify inaccuracies in the renderings such as floaters and unnatural artifacts. Specifically, we use a Multi-Layer Perceptron (MLP) with three layers of output size: [128, 128, 3] and Leaky ReLU activations throughout except the last layer as the meta-calibrator to estimate $\\theta$ given features extracted by the DINOv2 model [20] from rendered images $(\\mathbf{f}_I)$ and uncalibrated uncertainty maps $(\\mathbf{f}_{\\kappa})$ . The goal here is to have DINOv2 extract features that describe the rendering imperfections, correlating with the calibration curve. We find that training the MLP model on only a few scenes (30 in our case) allows it to generalize well to new test scenes. Once trained, the meta-calibrator can predict the calibration curve of a new target scene as: $\\theta = MLP([\\mathbf{f}_I,\\mathbf{f}_{\\kappa}])$ , without using any additional ground truth data.", + "bbox": [ + 212, + 297, + 787, + 525 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In summary, the meta-calibrator can correct the confidence levels of the model without requiring ground truth data at any stage, suggesting potential enhancements to applications that rely on uncertainty such as next-best view selection (see Sec. 4.3).", + "bbox": [ + 212, + 540, + 787, + 602 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 625, + 375, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The objective of the experiments is to: 1) validate that our approach achieves more accurate uncertainties (lower negative log-likelihood and calibration error) than state-of-the-art approaches for NeRF uncertainty estimation (Sec. 4.1); 2) demonstrate the meta-calibrator improves the accuracy of the uncalibrated uncertainties (decreases both the negative log-likelihood and calibration error) without requiring any held-out data from the target scene (Sec. 4.2); 3) explain the motivation for certain meta-calibrator design decisions; and 4) show that our uncertainties can be leveraged for applications such as next-best view planning (Sec. 4.3).", + "bbox": [ + 212, + 657, + 787, + 792 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For additional results showing: 1) why the PCA representation of the calibration curves is necessary; 2) that using the training set results in severe overfitting; 3) that holding out data results in poor performance at image reconstruction; 4)", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "the influence of the number of samples along the ray on the uncertainty quality; and 5) the efficiency of our uncertainty metric (Eq. (3)) over other approaches, please refer to Sections 1, 2, 3, 4, and 6 of the Supplementary respectively.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metrics and calibration curves. We use a variant of the calibration error from [12] to evaluate the effectiveness of the meta-calibrator. Specifically, given a test set $\\mathcal{D} = \\{(\\mathbf{r_t},\\mathbf{c_t})\\}_{t = 1}^T = \\{(\\mathbf{r_t},(r_t,g_t,b_t))\\}_{t = 1}^T$ , we report:", + "bbox": [ + 212, + 203, + 784, + 250 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nE R R = \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\left(p _ {t} - \\hat {P} \\left(p _ {t}\\right)\\right) ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 260, + 784, + 299 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $p_t$ is the expected confidence level for data point $(\\mathbf{r_t}, c_t)$ , and $\\hat{P}(p_t)$ is the empirical frequency of data points within that confidence level. More specifically, for each $t \\in \\{1, \\dots, T\\}$ , we set $p_t = M_t^C(c_t)$ and,", + "bbox": [ + 212, + 306, + 782, + 354 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {P} (p) = \\left| \\left\\{c _ {t}: M _ {t} ^ {C} \\left(c _ {t}\\right) \\leq p \\text {f o r} t = 1, \\dots , T \\right\\} \\right| / T, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 357, + 784, + 377 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $M \\equiv F$ for uncalibrated errors and $M \\equiv \\hat{F}$ for calibrated errors, and $C \\in \\{R, G, B\\}$ . Note that this formulation of the calibration error is equivalent to the one in [12] with a confidence level for every unique $p_t$ and weights that more significantly penalize errors from frequently predicted confidence levels.", + "bbox": [ + 212, + 382, + 782, + 441 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Following [12], we plot $\\{(p_t, \\hat{P}(p_t))\\}_{t=1}^T$ before and after calibration for each color channel to generate calibration curves. A perfectly calibrated forecaster would produce the straight line $p_t = \\hat{P}(p_t)$ as each expected confidence level would equal the empirical one. Intuitively, our version of the calibration error is the mean squared vertical distance of points on the calibration curve from a perfectly straight line. If an expected confidence level occurs $N$ times in the test data, its distance is counted $N$ times in the mean. Following [29], we additionally report the negative log-likelihood (NLL) of the test data averaged across all scenes. Following [24], we include PSNR and LPIPS [32] to evaluate image quality.", + "bbox": [ + 212, + 443, + 784, + 594 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. We use 30 scenes from 3 datasets: Realistic Synthetic $360^{\\circ}$ [16], the subset of scenes in DTU [8] used in [24], and LLFF [15] for training the metacalibrator and test it on a hold-out scene from either LLFF or DTU to show it generalizes to new target scenes.", + "bbox": [ + 212, + 604, + 782, + 666 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines We compare our approach against the state-of-the-art method for NeRF uncertainty estimation DANE [29] as well as other methods in Sec. 4.1. In Tab. 1, following [29], we implement the naive ensembles approach and DANE using a public implementation of Instant-NGP [2,17] and 5 ensemble members. In Sec. 4.2, we compare the uncalibrated uncertainties to the uncertainties calibrated by our meta-calibrator.", + "bbox": [ + 212, + 676, + 784, + 768 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Meta-calibrator Design The results guiding our decisions to use 3 Principal Component Analysis (PCA) components to represent the calibration curves, fit the PCA components using 21 training scenes, and train the meta-calibrator on 30 scenes to predict the PCA coefficients are shown in Fig. 3.", + "bbox": [ + 212, + 779, + 782, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "N. Amini-Naieni et al.", + "bbox": [ + 271, + 114, + 421, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3e0ba27d002685edcc41c61c18eb02494e05424fd4462c22a89c8e3c9d4af899.jpg", + "image_caption": [ + "(a) Explained variance plot showing most of the variance of the calibration curves is explained using only 3 PCA components." + ], + "image_footnote": [], + "bbox": [ + 215, + 223, + 390, + 325 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e192eca2cd13d334ea21be4c2534f2750d2e441fd97688027a0d3ee19f0e0413.jpg", + "image_caption": [ + "(b) Graph showing calibration error on new scenes is sufficiently low when using 21 scenes to fit 3-coefficient PCA model (i.e. it's an order of magnitude lower than we can expect from the final calibration, meaning good generalization to unseen scenes)." + ], + "image_footnote": [], + "bbox": [ + 413, + 203, + 589, + 286 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/cae523db36afed04044cba723c427bb76988a41c52a4efc60108223fe8037d2b.jpg", + "image_caption": [ + "(c) Graph showing increasing number of training scenes for meta-calibrator decreases calibration error for the test scene $T$ -Rex from LLFF [15], with 30 scenes resulting in good generalization to the held-out scene and well-calibrated uncertainties.", + "Fig. 3: Meta-calibrator design decisions. Results showing using 3 components for Principal Component Analysis (PCA) model of calibration curves, 21 scenes to fit PCA model, and 30 scenes to train meta-calibrator achieves good generalization to new test scenes." + ], + "image_footnote": [], + "bbox": [ + 611, + 202, + 785, + 285 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/052d8564e43b4738dc859bb16c4b5211d53dd8da70521e6538e95c5ad13b0592.jpg", + "table_caption": [ + "Table 1: Quantitative results on standard sparse NeRF benchmark. Our proposed approach results in significantly better uncertainties and image quality than the state-of-the-art NeRF uncertainty estimation method DANE [29] does on the challenging 3-view LLFF [15] dataset. Specifically, our meta-calibrator reduces the calibration error to $6\\%$ of DANE's calibration error and the negative log-likelihood to be over 100 $\\%$ lower than DANE's. Note: lower calibration error (Cal. Err.) and negative log-likelihood (NLL) values indicate more accurate uncertainties. Results are averaged over all 8 scenes in LLFF." + ], + "table_footnote": [], + "table_body": "
Uncert aintyImageQuality
Cal. Err. (↓)NLL (↓)PSNR (↑)LPIPS (↓)
Naïve Ens.0.05054.3915.190.646
DANE [29]0.04413.7515.190.646
Ours0.0026-0.6819.340.235
", + "bbox": [ + 375, + 699, + 624, + 765 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Comparison to State-of-the-art", + "text_level": 1, + "bbox": [ + 215, + 146, + 517, + 161 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we compare our approach to prior methods for NeRF uncertainty estimation. We achieve more accurate uncertainties (94% reduction in calibration error and over 100% reduction in negative log-likelihood) than those estimated by DANE [29], the state-of-the-art method. These results are shown in Tab. 1 for the challenging 3-view LLFF [15] dataset from prior work on sparse novel view synthesis [19,24,25] and Tab. 2 for the less challenging version of LLFF from prior work on NeRF uncertainty estimation [14,26,27,29]. In Fig. 5a, we compare the calibration curves from our approach to DANE's, illustrating that the meta-calibrator predicts expected confidences that match the true ones while DANE does not.", + "bbox": [ + 212, + 172, + 787, + 323 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4d7c7acee6cd8b344d416a7a078bafb068672e896c6b653369c89db8153bef5a.jpg", + "image_caption": [ + "Fig. 4: Quantitative comparison of uncalibrated and calibrated uncertainties. In (a), we show calibration curves on test data from four scenes in LLFF [15]. The color of each curve indicates the color channel it corresponds to. The calibrated curves are much closer to the ideal calibration (dashed lines), demonstrating that the meta-calibrator works very well. In (b), the average calibration error and negative log-likelihood before and after calibration are reported for LLFF, clearly showing the meta-calibrator improves the accuracy of the uncertainties (lowering calibration error and negative log-likelihood). To test generalization, the meta-calibrator was also applied to held-out scenes in DTU [8], achieving a $70\\%$ reduction in calibration error on average." + ], + "image_footnote": [], + "bbox": [ + 218, + 345, + 782, + 500 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Comparison to Uncalibrated Uncertainties", + "text_level": 1, + "bbox": [ + 215, + 708, + 612, + 723 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we compare our uncalibrated base NeRF uncertainties to our calibrated uncertainties obtained from applying the proposed meta-calibrator. In Fig. 6 we show that the calibrated uncertainties better highlight floaters and other errors in the NeRF renderings. In Fig. 4, we show that the meta-calibrator predicts expected confidence that closely match the true ones, lowering both the calibration error and the negative log-likelihood of the uncalibrated uncertainties.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "N. Amini-Naieni et al.", + "bbox": [ + 271, + 114, + 421, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/98b0e41a417b73ed8991fc2a21a635878910fca5ab3c73964ffc0bec10abf746.jpg", + "table_caption": [ + "Table 2: Quantitative results on standard NeRF uncertainty estimation benchmark. Here, we present results on the LLFF [15] dataset used in prior work [14, 26, 27, 29] on uncertainty estimation for NeRFs. This dataset is less challenging than the one in Tab. 1 since 4-12 views are used for training instead of 3. Our proposed approach results in significantly better uncertainties than prior methods for NeRF uncertainty estimation on all 8 scenes in LLFF. Note: lower negative log-likelihood values indicate more accurate uncertainties. $M$ indicates the number of ensemble members, and MC-DO refers to Monte Carlo Dropout sampling with $M$ sample configurations. This table is Tab. 1 from [29] with our results added as an additional column. Please refer to [29] for further details." + ], + "table_footnote": [], + "table_body": "
Scene# of Train. ViewsNegative Log-likelihood (↓)
MC-DO M=5Naïve Ens. M=5NeRF-W [14]S-NeRF [27]CF-NeRF [26]DANE [29] M=5DANE [29] M=10Ours
Fern44.902.472.162.01—-0.98-1.00-1.41
Orchids55.742.232.241.95—-0.28-0.31-0.84
Leaves52.722.660.790.68—0.970.73-1.19
Flower74.631.631.711.27—1.000.85-2.05
Fortress85.192.291.04-0.03—-1.30-1.30-1.99
Room85.062.134.932.35—-1.35-1.35-2.17
T-Rex114.102.281.911.37—-0.31-0.69-1.49
Horns124.182.170.780.60—-0.55-0.66-2.18
Avg.4.572.231.951.270.57-0.35-0.47-1.67
", + "bbox": [ + 215, + 316, + 795, + 441 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2d07211d4bd062480b310d0bdf06858042dc99302597f89c54a366456d60b3c2.jpg", + "image_caption": [ + "(a) Calibration curves for the Fern scene from 3-view LLFF [15]." + ], + "image_footnote": [], + "bbox": [ + 328, + 505, + 483, + 623 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e3e883abb1668e94b629fb55e7d1037e931ef4f52c6da77c20f6d51dbe06af47.jpg", + "image_caption": [ + "(b) Next-best View Planning on Horns from LLFF [15].", + "Fig. 5: Comparison to DANE [29]. Results comparing our uncertainties to those from the state-of-the-art method DANE. In (a) we show DANE's RGB calibration curves are not closely aligned with the perfectly calibrated lines, meaning it is miscalibrated. It is significantly over-confident for expected confidence levels close to 1 and under-confident for confidence levels close to 0. In comparison, the curves for our approach are extremely close to the ideal calibration (dashed lines), demonstrating that the meta-calibrator works very well, predicting expected confidences that match the true ones. This is also verified by how our calibration error is over two orders of magnitude smaller than DANE's. In (b) we show that our approach results in more efficient performance gains over DANE for next-best view planning." + ], + "image_footnote": [], + "bbox": [ + 509, + 525, + 663, + 633 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/772e986f14d98b9828f9bfaedd02765dd81cdf962ac1fb0f1f0b7934a7d90052.jpg", + "image_caption": [ + "(a) Ground Truth", + "Fig. 6: Qualitative comparison of uncalibrated and calibrated uncertainties from the Flower scene in LLFF [15]. The calibrated uncertainties (d) clearly detect incorrect regions (indicated by the red boxes) better than the uncalibrated uncertainties (c) do. This is apparent by noting that (d) and (e) look more similar than (c) and (e)." + ], + "image_footnote": [], + "bbox": [ + 218, + 142, + 331, + 253 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0396d509886d8986cbbb310048ba4947368c50a9cd34d677c2dc8b942ce16e40.jpg", + "image_caption": [ + "(b) NeRF Image" + ], + "image_footnote": [], + "bbox": [ + 334, + 143, + 444, + 253 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/81cdbc67aa61107e29a81b8c38c286f8543e19698907e8ecf1aec0fba758f609.jpg", + "image_caption": [ + "(c) Uncalibrated Uncertainty" + ], + "image_footnote": [], + "bbox": [ + 446, + 143, + 555, + 253 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ca22ab0ca68246f8540977f906d1c3eb9e1c771bafb219b95e24638c50cc8fae.jpg", + "image_caption": [ + "(d) Calibrated Uncertainty" + ], + "image_footnote": [], + "bbox": [ + 558, + 143, + 671, + 253 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0c476a28b0a02923fae760d9da2b1a27895cd36bf62cc1fb6f6d757335777f28.jpg", + "image_caption": [ + "(e) Actual Error" + ], + "image_footnote": [], + "bbox": [ + 674, + 143, + 784, + 253 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.3 Application: Next-best View Planning", + "text_level": 1, + "bbox": [ + 214, + 377, + 578, + 393 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this section, we show that our uncertainties can be leveraged for next-best view planning. Specifically, we start by training the NeRF model for 2000 iterations on a training set of three images. The next-best view is selected by evaluating the average calibrated pixel uncertainty (obtained using the meta-calibrator) for each of the candidate views, and the view with the highest uncertainty is added to the training set. The average PSNR of the test images is reported after each training iteration. In Fig. 5b we show that using our approach results in greater performance gains (higher PSNRs) than DANE [29] does. In the Fig. 1 of the Supplementary, we also compare the information gain from rays selected according to the highest calibrated uncertainties to the information gain from rays selected according to the highest uncalibrated uncertainties and show our meta-calibrator produces higher average PSNRs on the test set for scenes in DTU [8] than the uncalibrated uncertainties do. Thus, calibration specifically re-orders the pixel uncertainties so that rays more likely to raise the PSNR are picked earlier in next-best view planning. In Sec. 5 of the Supplementary, we include a detailed theoretical example showing that such re-ordering is possible.", + "bbox": [ + 212, + 402, + 787, + 645 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 667, + 359, + 683 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper we addressed the open problem of obtaining calibrated uncertainties from NeRF models. We introduce the concept of a meta-calibrator that infers the calibration curves from scene features, and using this approach achieve state-of-the-art uncertainty without holding out any ground truth data from the target scene. By enabling efficient and accurate calibration of NeRF models without relying on additional data, our method represents a significant step forward in the practical application of NeRF to real-world scenarios and opens up new avenues for the use of NeRF in situations where data is limited and uncertainty is critical.", + "bbox": [ + 212, + 699, + 787, + 834 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "N. Amini-Naieni et al.", + "bbox": [ + 271, + 114, + 421, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. The authors would like to thank Seunghyeon Seo for his support on FlipNeRF, and Oishi Deb for her insightful feedback on uncertainty estimation. Niki Amini-Naieni is funded by an AWS Studentship, the Reuben Foundation, and the AIMS CDT program at the University of Oxford. Tomas Jakab is sponsored by EPSRC VisualAI EP/T028572/1 and Andrea Vedaldi by ERC-CoG UNION 101001212.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 261, + 321, + 277 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV. pp. 5835-5844 (2021)", + "2. Bhalgat, Y.: Hashnerf-pytorch. https://github.com/yashbhalgat/HashNeRF-pytorch/ (2022)", + "3. Gafni, G., Thies, J., Zollhöfer, M., Nießner, M.: Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In: CVPR. pp. 8649-8658 (2021)", + "4. Ghoshal, B., Tucker, A.: On calibrated model uncertainty in deep learning. In: ECML (2022)", + "5. Goli, L., Reading, C., Sellan, S., Jacobson, A., Tagliasacchi, A.: Bayes' Rays: Uncertainty quantification in neural radiance fields. ArXiv abs/2309.03185 (2023)", + "6. Jain, A., Tancik, M., Abbeel, P.: Putting nef on a diet: Semantically consistent few-shot view synthesis. In: ICCV. pp. 5865-5874 (2021)", + "7. Jang, T.J., Hyun, C.M.: Nerf solves undersampled mri reconstruction. ArXiv abs/2402.13226 (2024)", + "8. Jensen, R.R., Dahl, A., Vogiatzis, G., Tola, E., Aanaes, H.: Large scale multi-view stereopsis evaluation. In: CVPR (2014)", + "9. Jin, L., Chen, X., Ruckin, J., Popovi'c, M.: Neu-nbv: Next best view planning using uncertainty estimation in image-based neural rendering. In: IROS (2023)", + "0. Kajiya, J.T., Von Herzen, B.P.: Ray tracing volume densities. In: SIGGRAPH. p. 165-174 (1984)", + "1. Kosiorek, A.R., Strathmann, H., Zoran, D., Moreno, P., Schneider, R., Mokr'a, S., Rezende, D.J.: Nerf-vae: A geometry aware 3d scene generative model. In: ICML (2021)", + "2. Kuleshov, V., Fenner, N., Ermon, S.: Accurate uncertainties for deep learning using calibrated regression. In: ICML. pp. 2796-2804 (2018)", + "3. Liu, L., Gu, J., Lin, K.Z., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. In: NIPS (2020)", + "4. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: CVPR (2021)", + "5. Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. In: TOG (2019)", + "6. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "7. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. In: ACM Trans. Graph. (2022)" + ], + "bbox": [ + 225, + 294, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Neff, T., Stadlbauer, P., Parger, M., Kurz, A., Mueller, J.H., Chaitanya, C.R.A., Kaplanyan, A., Steinberger, M.: Donerf: Towards real-time rendering of compact neural radiance fields using depth oracle networks. In: CGF. pp. 45-59 (2021)", + "19. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S.M., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: CVPR (2022)", + "20. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Howes, R., Huang, P.Y., Xu, H., Sharma, V., Li, S.W., Galuba, W., Rabbat, M., Assran, M., Ballas, N., Synnaeve, G., Misra, I., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023)", + "21. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: ICCV (2021)", + "22. Ran, Y., Zeng, J., He, S., Li, L., Chen, Y., Lee, G.H., Chen, J., Ye, Q.: Neurar: Neural uncertainty for autonomous 3d reconstruction. In: RAL (2023)", + "23. Rebain, D., Jiang, W., Yazdani, S., Li, K., Yi, K.M., Tagliasacchi, A.: Derf: Decomposed radiance fields. In: CVPR. pp. 14148-14156 (2020)", + "24. Seo, S., Chang, Y., Kwak, N.: Flipnerf: Flipped reflection rays for few-shot novel view synthesis. In: ICCV (2023)", + "25. Seo, S., Han, D., Chang, Y., Kwak, N.: Mixnerf: Modeling a ray with mixture density for novel view synthesis from sparse inputs. In: CVPR. pp. 20659-20668 (2023)", + "26. Shen, J., Agudo, A., Moreno-Noguer, F., Ruiz, A.: Conditional-flow nerf: Accurate 3d modelling with reliable uncertainty quantification. In: ECCV (2022)", + "27. Shen, J., Ruiz, A., Agudo, A., Moreno-Noguer, F.: Stochastic neural radiance fields: Quantifying uncertainty in implicit 3d representations. In: 3DV. pp. 972-981 (2021)", + "28. Sitzmann, V., Martel, J.N., Bergman, A.W., Lindell, D.B., Wetzstein, G.: Implicit neural representations with periodic activation functions. In: NIPS (2020)", + "29. Sünderhauf, N., Abou-Chakra, J., Miller, D.: Density-aware nef ensembles: Quantifying predictive uncertainty in neural radiance fields. In: ICRA (2023)", + "30. Wang, P., Liu, Y., Chen, Z., Liu, L., Liu, Z., Komura, T., Theobalt, C., Wang, W.: F2-nerf: Fast neural radiance field training with free camera trajectories. In: CVPR (2023)", + "31. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. ArXiv abs/2010.07492 (2020)", + "32. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + ], + "bbox": [ + 215, + 146, + 785, + 660 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "N. Amini-Naieni et al.", + "bbox": [ + 271, + 114, + 419, + 126 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator/6c9ab90c-9fb8-405a-a3f6-eaa68a4cea3f_model.json b/2024/Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator/6c9ab90c-9fb8-405a-a3f6-eaa68a4cea3f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bee2d7cdbe88b18ac861bcecbc67f5a5130ee1ba --- /dev/null +++ b/2024/Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator/6c9ab90c-9fb8-405a-a3f6-eaa68a4cea3f_model.json @@ -0,0 +1,1981 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.787, + 0.186 + ], + "angle": 0, + "content": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.213, + 0.774, + 0.23 + ], + "angle": 0, + "content": "Niki Amini-Naieni1, Tomas Jakab1, Andrea Vedaldi1, and Ronald Clark1" + }, + { + "type": "text", + "bbox": [ + 0.432, + 0.241, + 0.573, + 0.255 + ], + "angle": 0, + "content": "University of Oxford" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.297, + 0.74, + 0.643 + ], + "angle": 0, + "content": "Abstract. Neural Radiance Fields (NeRFs) have markedly improved novel view synthesis, but accurate uncertainty quantification in their image predictions remains an open problem. The prevailing methods for estimating uncertainty, including the state-of-the-art Density-aware NeRF Ensembles (DANE) [29], quantify uncertainty without calibration. This frequently leads to over- or under-confidence in image predictions, which can undermine their real-world applications. In this paper, we propose a method which, for the first time, achieves calibrated uncertainties for NeRFs. To accomplish this, we overcome a significant challenge in adapting existing calibration techniques to NeRFs: a need to hold out ground truth images from the target scene, reducing the number of images left to train the NeRF. This issue is particularly problematic in sparse-view settings, where we can operate with as few as three images. To address this, we introduce the concept of a meta-calibrator that performs uncertainty calibration for NeRFs with a single forward pass without the need for holding out any images from the target scene. Our meta-calibrator is a neural network that takes as input the NeRF images and uncalibrated uncertainty maps and outputs a scene-specific calibration curve that corrects the NeRF's uncalibrated uncertainties. We show that the meta-calibrator can generalize on unseen scenes and achieves well-calibrated and state-of-the-art uncertainty for NeRFs, significantly beating DANE and other approaches. This opens opportunities to improve applications that rely on accurate NeRF uncertainty estimates such as next-best view planning and potentially more trustworthy image reconstruction for medical diagnosis." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.658, + 0.698, + 0.673 + ], + "angle": 0, + "content": "Keywords: NeRFs \\(\\cdot\\) Uncertainty calibration \\(\\cdot\\) Few-shot learning" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.7, + 0.377, + 0.716 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Recent advancements in scene representations have led to promising new approaches for novel view synthesis and scene reconstruction. Among these, Neural Radiance Fields (NeRFs) [16] have emerged as a particularly powerful tool, offering unprecedented levels of realism and detail in rendered images. The core idea behind NeRFs is to represent a scene as a continuous vector-valued function, parameterized by a neural network, that maps spatial coordinates and view directions to color and density values. This approach enables the creation" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.228, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.128 + ], + "angle": 0, + "content": "N. Amini-Naieni et al." + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.143, + 0.703, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.334, + 0.788, + 0.419 + ], + "angle": 0, + "content": "Fig. 1: We propose a method for efficiently calibrating the uncertainties from NeRF models. Our approach is based on a meta-calibrator that takes as input features from the rendered NeRF images and uncalibrated uncertainty maps and predicts the calibration function, \\( R_{\\theta}(\\cdot) \\), for the NeRF model. Our meta-calibrator generalizes across scenes, so it only needs to be trained once, and can predict the calibration function in a single forward pass without any ground truth data from the target scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.788, + 0.484 + ], + "angle": 0, + "content": "of detailed 3D representations from sets of 2D images, revolutionizing 3D reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.788, + 0.639 + ], + "angle": 0, + "content": "However, despite their impressive capabilities, traditional NeRF models lack an essential component: an accurate measure of uncertainty in their predictions. Accurate uncertainties are crucial for applying NeRFs to safety-critical problems such as MRI image reconstruction from sparse data [7], where unreliable confidence estimates could lead to misdiagnosis. More accurate uncertainties could also enhance practical methods such as uncertainty-guided next-best view planning techniques [9]. Prior approaches have attempted to estimate NeRF uncertainties [5, 14, 26, 27, 29], but they all overlook the problem of calibration. Thus, the uncertainties they output are not as accurate as they could be." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.788, + 0.733 + ], + "angle": 0, + "content": "In particular, the state-of-the-art uncertainty estimation method for NeRFs, Density-aware NeRF Ensembles (DANE) [29], produces uncalibrated uncertainties. As a result, the confidence intervals and variances do not match the true confidence, meaning it has limited applicability to real-world problems. This constraint is significant as it restricts the use of NeRFs in safety-critical and sparse-data settings, where knowing the confidence in predictions is crucial." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.826 + ], + "angle": 0, + "content": "The best NeRF methods for the sparse-view setting overlook the problem of calibration as well. FlipNeRF [24], the state-of-the-art technique for sparse-view reconstruction, uses uncertainties from an uncalibrated mixture of Laplacians to enhance its training process. Therefore, the uncertainties it outputs at inference as an artifact of training are not accurate." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.286, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.793, + 0.39 + ], + "angle": 0, + "content": "In this paper, we present a novel approach for obtaining calibrated uncertainties for NeRF models in the sparse-view setting. Our strategy integrates the Laplacian mixture from FlipNeRF [24] with the calibration techniques by Kuleshov et al [12]. However, naively applying the calibration method by Kuleshov et al to FlipNeRF does not work due to a significant challenge of the sparse-view setting: there is a lack of held-out data from the target scene for fitting the calibrator. Specifically, holding out just one image for calibration could decrease the size of the training set by over \\(30\\%\\) , resulting in significant performance degradation of the NeRF. To overcome this, we make use of a unique observation: while calibration curves exhibit significant variation across scenes, they also demonstrate a significant regularity in their structure. Utilizing this insight, we propose the concept of a meta-calibrator that learns a low-dimensional representation of the NeRF calibration curves and infers them from scene features. We motivate and show why this meta-calibrator is necessary and demonstrate that it achieves more accurate uncertainties than DANE [29] without holding out any images from the target scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.79, + 0.512 + ], + "angle": 0, + "content": "Specifically, our contributions are: (1) the first investigation into obtaining calibrated uncertainties from NeRFs, (2) a novel meta-calibrator for fitting the calibration model without using held-out data, and (3) experiments on the real-world Local Light Field Fusion (LLFF) [15] and DTU [8] datasets showing that our meta-calibrator achieves state-of-the-art and well-calibrated uncertainties for real scenes. We also demonstrate that our uncertainties can be leveraged for effective next-best view planning." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.543, + 0.388, + 0.558 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.789, + 0.704 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRFs). NeRFs [16] are a popular method for novel view synthesis. From a set of 2D images, NeRFs learn a neural network representation of a single scene. A trained NeRF model outputs estimates of the volume density and emitted radiance at any 3D location and viewing direction. Novel views can be generated by applying volume rendering [10] to the density and radiance values predicted by the NeRF model for points along rays cast into the scene. Due to their simplicity and impressive performance, NeRFs have become a popular technique for solving a variety of rendering problems." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Over the last few years, several extensions of NeRFs have been explored [11, 18, 21, 28, 31]. These include speeding up training and inference [13, 23, 30], modeling dynamic scenes [3], learning from a sparse set of training views [1, 6, 19, 24, 25], and estimating the uncertainty in NeRF predictions [5, 14, 22, 26, 27, 29]. Sparse NeRF methods aim to accurately render novel views when only a few training views are available from the target scene. NeRF uncertainty estimation techniques strive to accurately predict the confidence in the views rendered. Although uncertainty estimation is particularly important in the sparse-view" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.128 + ], + "angle": 0, + "content": "N. Amini-Naieni et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.177 + ], + "angle": 0, + "content": "setting, where NeRF renderings are especially unreliable, the main aim of sparse NeRF methods is not to output accurate uncertainties." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.192, + 0.787, + 0.298 + ], + "angle": 0, + "content": "Sparse-view NeRF Methods. Despite this, recent sparse NeRF methods do produce uncertainties as an artifact of their training process. Both MixNeRF [25] and FlipNeRF [24], the state-of-the-art approach, model the RGB color channels given a ray as independent random variables that follow a mixture model. FlipNeRF further uses the uncertainties of the pixel colors to regularize the training process, producing more accurate image reconstructions at inference. However, neither MixNeRF nor FlipNeRF outputs calibrated uncertainties." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.313, + 0.787, + 0.464 + ], + "angle": 0, + "content": "Our method significantly extends sparse NeRF methods to obtain more accurate and well-calibrated uncertainties at inference. To benefit from the superior performance of FlipNeRF [24] at sparse novel view synthesis, we apply the proposed meta-calibrator to the learned distribution from FlipNeRF, producing more accurate uncertainties without sacrificing state-of-the-art image quality. However, our approach can be applied to any NeRF method that outputs uncertainties, so it is distinct from FlipNeRF and MixNeRF. In essence, the proposed meta-calibrator augments sparse NeRF methods to achieve state-of-the-art uncertainty, beating both FlipNeRF and techniques designed explicitly for NeRF uncertainty estimation." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.478, + 0.787, + 0.781 + ], + "angle": 0, + "content": "Uncertainty in NeRFs. The growing line of methods specifically designed for accurately estimating NeRF uncertainties [5,14,22,26,27,29] do not address the problem of calibration. The current state-of-the-art uncertainty estimation technique for NeRFs is Density-aware NeRF Ensembles (DANE) [29]. DANE adds an epistemic uncertainty term to a naive ensembles approach with five ensemble members. Thus, DANE is very costly as it requires training five NeRFs to obtain uncertainty estimates. Another NeRF uncertainty estimation approach is Stochastic Neural Radiance Fields (S-NeRF) [27], which learns a probability distribution over all possible radiance fields by modeling the volume density and radiance as random variables that follow a joint distribution. S-NeRF employs variational inference to sample from an approximation to this distribution and uses the variances of the sampled pixel colors as the estimated uncertainties. Conditional-flow NeRF (CF-NeRF) [26] builds on S-NeRF by combining latent variable modeling and conditional normalizing flows to relax the strong constraints S-NeRF imposes over the radiance distribution. Despite the growing number of techniques in this area of study, all prior work does not consider calibration, outputting unreliable uncertainties as a result. Our work achieves more accurate uncertainties than these prior methods by filling the gap of uncertainty calibration for NeRFs and drawing on well-established techniques in calibrated regression [12]." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Uncertainty Calibration in Deep Learning. While uncertainty calibration has been studied for Bayesian deep learning methods [4, 12], it has not been adapted for or applied successfully to NeRF uncertainty estimates. This may be" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.286, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.784, + 0.126 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "because NeRFs introduce additional complexity in the uncertainty estimation process as the neural network model needs to be trained per-scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.312 + ], + "angle": 0, + "content": "This makes uncertainty calibration challenging as methods for calibrated regression [12] require either using the training set or held-out data to achieve calibrated uncertainties. Using the training set to fit the calibrator results in severe overfitting (see Sec. 2 of the Supplementary). Holding out data from the target scene for calibration means there is less data to train the NeRF, making it more inaccurate at novel view synthesis (see Sec. 3 of the Supplementary). Thus, a trivial application of [12] to NeRFs would not be satisfactory. In this work, we propose the concept of a meta-calibrator that, in contrast to [12], does not require held-out data and achieves calibrated uncertainty estimates for NeRFs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.352, + 0.331, + 0.368 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.788, + 0.535 + ], + "angle": 0, + "content": "In this paper, we present a method that calibrates NeRF uncertainties. To this end, we propose a novel meta-calibrator that accepts uncalibrated NeRF uncertainties and predicted images as inputs and outputs a scene-specific calibration curve, correcting the uncalibrated confidence levels. An overview of our method can be seen in Fig. 1. Crucially, our approach does not require holding out any images from the target scene. Thus, it can be applied to sparse-view settings, where holding out a single image could significantly harm the NeRF's performance. In Sec. 3.1, we explain the necessary background concepts, in Sec. 3.2, we describe how we obtain the corrected uncertainty values, and in Sec. 3.3 we detail our meta-calibrator." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.558, + 0.373, + 0.572 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRFs). Neural Radiance Fields (NeRFs) [16] represent a scene as a continuous vector-valued function with inputs a Cartesian point \\(\\mathbf{x} = (x,y,z)\\) and unit viewing direction vector \\(\\mathbf{d} = (u,v,w)\\) and outputs an emitted radiance \\(\\pmb {c} = (\\nu ,\\varrho ,\\theta)\\) and volume density \\(\\sigma\\) . By optimizing the weights \\(\\Theta\\) of a neural network approximation \\(\\mathbf{F}_{\\Theta}\\) to this representation, NeRFs can render the color of any pixel in a synthetic image of the scene. To achieve this, principles from classical volume rendering [10] are applied to the radiance and density values estimated by \\(\\mathbf{F}_{\\Theta}\\) for points along a ray cast from the origin \\(\\mathbf{x_0}\\) of the virtual camera, through the pixel, and into the scene. More specifically, the expected color \\(\\mathbf{c}(\\mathbf{r})\\) of a camera ray \\(\\mathbf{r}(t) = \\mathbf{x_0} + t\\mathbf{d}\\) with near and far bounds \\(t_n\\) and \\(t_f\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.751, + 0.786, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} (\\mathbf {r}) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (\\mathbf {r} (t)) \\boldsymbol {c} (\\mathbf {r} (t), \\mathbf {d}) d t, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.792, + 0.788, + 0.842 + ], + "angle": 0, + "content": "where \\( T(t) = e^{-\\int_{t_n}^t \\sigma(\\mathbf{r}(s)) ds} \\). The integral in Eq. (1) is estimated with numerical quadrature to obtain the colors of the pixels in the synthetic image from the NeRF model outputs. NeRF optimizes \\( \\Theta \\) according to Eq. (1) with gradient" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.128 + ], + "angle": 0, + "content": "N. Amini-Naieni et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "descent, since the numerical quadrature is differentiable. Conventional NeRFs do not provide an uncertainty associated with their predictions, so extensions like DANE [29] have been developed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.202, + 0.785, + 0.339 + ], + "angle": 0, + "content": "Base NeRF Uncertainties. To obtain the initial uncertainties, we have two options for our base model: (1) DANE [29], the state-of-the-art method for NeRF uncertainty estimation or (2) FlipNeRF [24], the state-of-the-art method for sparse novel view synthesis. DANE requires training five NeRFs per scene and provides poor image quality in the sparse-view setting, so we choose FlipNeRF. We show in the experiments that applying our meta-calibrator to FlipNeRF results in more accurate and well-calibrated uncertainties than those output by DANE. Our base FlipNeRF uncertainties are inferred from a mixture of Laplacians with location and scale parameters learned during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.784, + 0.385 + ], + "angle": 0, + "content": "FlipNeRF [24] models the joint distribution of the color \\(\\mathbf{C} = (R,G,B)\\) given a ray \\(\\mathbf{r}\\) with a mixture of Laplacians:" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.397, + 0.785, + 0.438 + ], + "angle": 0, + "content": "\\[\np (\\mathbf {C} = \\mathbf {c} | \\mathbf {r}) = \\sum_ {j = 1} ^ {M} \\pi_ {j} \\mathcal {F} (\\mathbf {C} = \\mathbf {c}; \\boldsymbol {\\mu} _ {j}, \\boldsymbol {\\beta} _ {j}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.785, + 0.553 + ], + "angle": 0, + "content": "where \\( M \\) is the number of sampled points along the ray \\( r \\). \\( \\mathcal{F}(\\mathbf{C} = \\mathbf{c};\\pmb{\\mu}_j,\\pmb{\\beta}_j) \\) is the 3D Laplacian probability density with location parameter \\( \\pmb{\\mu}_{j} = (\\mu_{j}^{R},\\mu_{j}^{G},\\mu_{j}^{B}) \\) and scale parameter \\( \\beta_{i} = (\\beta_{j}^{R},\\beta_{j}^{G},\\beta_{j}^{B}) \\) evaluated at the color \\( \\mathbf{c} \\). More specifically, the mixing coefficients \\( \\{\\pi_j\\}_{j = 1}^M \\) are the normalized coefficients of the radiance values along a ray in Eq. (2), the location parameters \\( \\{\\pmb{\\mu}_{j}\\}_{j = 1}^{M} \\) are the estimated RGB radiance values, and the scale parameters \\( \\{\\pmb{\\beta}_{j}\\}_{j = 1}^{M} \\) are an additional output of the model. These parameters are optimized by FlipNeRF during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.785, + 0.75 + ], + "angle": 0, + "content": "We now go into detail about how the location and scale parameters are obtained. FlipNeRF [24] learns these parameters by minimizing the negative log-likelihood of the training set \\(\\mathcal{D} = \\{(\\mathbf{r_i},\\mathbf{c_i})\\}_{i = 1}^N\\) containing the rays \\(\\mathbf{r_i}\\) and colors \\(\\mathbf{c_i}\\) from the pixels in the ground truth images of the scene assuming the distribution in Eq. (2). FlipNeRF additionally minimizes the average scale parameters through an auxiliary uncertainty-aware emptiness loss for reducing floating artifacts. The negative log-likelihood loss and the uncertainty-aware emptiness loss are added to FlipNeRF's total training loss, which incorporates other terms such as the mean squared error. As a result of this training process, the location and scale parameters can be inferred by FlipNeRF at new poses. We obtain the uncalibrated confidence levels of predicted ray colors from these location and scale parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Using the distribution in Eq. (2), we can easily compute the confidence level for a given ray \\(\\mathbf{r_t}\\), which we denote as \\(F_{t}\\). The CDF for the Laplacian mixture has a closed form expression parameterized by the location and scales output by the pretrained FlipNeRF [24]. Thus, we can use this CDF to predict the confidence level of the ground truth color \\(\\mathbf{c_t}\\) of any ray by evaluating it at the given" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.286, + 0.115, + 0.731, + 0.128 + ], + "angle": 0, + "content": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "color value, \\( p_t = F_t(\\mathbf{c_t}) \\). However, these initial confidence levels are uncalibrated and, hence, inaccurate." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.204, + 0.785, + 0.293 + ], + "angle": 0, + "content": "Calibrated regression. In [12], Kuleshov et al extend calibration methods for classification to regression. They define a forecaster \\( H: \\mathcal{X} \\to (\\mathcal{Y} \\to [0,1]) \\) as a function that outputs for each \\( x_{t} \\in \\mathcal{X} \\), a CDF \\( F_{t} \\). Given a pretrained forecaster \\( H \\), they suggest training an auxiliary model \\( R: [0,1] \\to [0,1] \\) by fitting \\( R \\) to a recalibration dataset \\( D = \\left\\{\\left([H(x_{t})](y_{t}), \\hat{P}([H(x_{t})](y_{t}))\\right)\\right\\}_{t=1}^{T} \\), where" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.302, + 0.68, + 0.319 + ], + "angle": 0, + "content": "\\[\n\\hat {P} (p) = \\left| \\left\\{y _ {t}: [ H \\left(x _ {t}\\right) ] \\left(y _ {t}\\right) \\leq p \\text {f o r} t = 1, \\dots , T \\right\\} \\right| / T\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.784, + 0.388 + ], + "angle": 0, + "content": "is the empirical confidence level corresponding to the predicted confidence level \\( p \\). The fitted \\( R \\) forms the calibration curve that corrects the expected confidence levels. We can now obtain predictive posterior values that closely match the true confidences using \\( \\hat{F}_t \\equiv R \\circ F_t \\) for the test data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.785, + 0.537 + ], + "angle": 0, + "content": "However, as mentioned before, there are numerous challenges associated with applying this recalibration procedure. Firstly, note that it requires ground-truth values \\((y_{t})\\) for the predictions we are recalibrating. This means that in order to prevent overfitting we need to reserve part of the training dataset specifically for fitting the calibrator (which leaves less data for training the NeRF — a significant issue if we only have a few input views). Secondly, it does not actually prescribe how to compute a suitable uncertainty value from the predicted distribution. In Sec. 3.3, we address the former issue, and in Sec. 3.2, we address the latter." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.573, + 0.457, + 0.589 + ], + "angle": 0, + "content": "3.2 Calculating uncertainty" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.785, + 0.731 + ], + "angle": 0, + "content": "The predictive posterior in Eq. (2) provides a distribution over likely ray colors for a NeRF model, but it does not inherently offer a straightforward metric for quantifying uncertainty at a specific point in the reconstruction. Intuitively, as the variance of this distribution increases, so does the uncertainty of the model's output at that point. Therefore, the variance or standard deviation is a popular choice for quantifying the uncertainty [24,26,27,29]. However, in the case of the corrected mixture distribution obtained from raymarching, this can be slow to compute especially if it has to be done for each pixel. Therefore, we turn to a metric that can be calculated directly from the calibrated CDF \\(\\hat{F}^t\\)." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.746, + 0.767, + 0.763 + ], + "angle": 0, + "content": "We propose to use the interquartile range of each calibrated distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.77, + 0.785, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\kappa^ {C} \\left(\\mathbf {r} _ {\\mathbf {t}}\\right) = \\left[ \\hat {F} _ {t} ^ {C} \\right] ^ {- 1} \\left(\\frac {3}{4}\\right) - \\left[ \\hat {F} _ {t} ^ {C} \\right] ^ {- 1} \\left(\\frac {1}{4}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.809, + 0.784, + 0.84 + ], + "angle": 0, + "content": "where \\(\\kappa^C (\\mathbf{r_t})\\) represents the uncertainty at a ray \\(\\mathbf{r_t}\\) in the color channel \\(C\\). This difference provides a measure of the statistical dispersion and thus serves" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.228, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.128 + ], + "angle": 0, + "content": "N. Amini-Naieni et al." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.146, + 0.493, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.262, + 0.49, + 0.284 + ], + "angle": 0, + "content": "(a) Creating a parametric model of the calibration curves." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.184, + 0.77, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.273, + 0.746, + 0.285 + ], + "angle": 0, + "content": "(b) Predicting the curve parameters." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.296, + 0.789, + 0.354 + ], + "angle": 0, + "content": "Fig. 2: Meta-calibrator design. In stage (a) we fit a low-dimensional parametric model of the calibration curves. The meta-calibrator then predicts these curve parameters from rendered images of the scene and their associated uncalibrated uncertainty maps (b)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.789, + 0.456 + ], + "angle": 0, + "content": "as a robust measure of the spread of the output channel. By averaging the interquartile range over the color channels, we obtain a single scalar value that effectively quantifies the uncertainty of the NeRF model for the given ray. As shown in our experiments, this method is very computational efficient, and it provides an accurate estimate of uncertainty." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.474, + 0.393, + 0.488 + ], + "angle": 0, + "content": "3.3 Meta-calibrator" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.789, + 0.633 + ], + "angle": 0, + "content": "To overcome the challenge that, especially in the sparse-view setting, there is no held-out data available for fitting the calibrator, we propose a novel metacalibrator that infers the calibration curves from uncalibrated NeRF predictions. To do this, we leverage the insight that the calibration curves demonstrate significant regularity. We posit a low-dimensional model of the calibration curves can be learned and predicted using the images and uncalibrated uncertainty maps inferred by the NeRF, enabling us to estimate the calibration function without evaluating the empirical confidence levels using held-out data from the target scene. We now describe this meta-calibrator (illustrated in Fig. 2) in detail." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.843 + ], + "angle": 0, + "content": "A Parametric Model for Calibration Curves. We first fit a low-dimensional representation of the calibration curves using Principal Component Analysis (PCA). To create the training set for learning this representation, we sample held-out images from \\(K\\) scenes and apply the calibration procedure by Kuleshov et al. [12] to form \\(K\\) ground truth calibration curves. To construct the training vector \\(\\mathbf{v_k} \\in \\mathbb{R}^{1 \\times M}\\) for scene \\(k \\leq K\\), we sample \\(M\\) evenly spaced points along its ground truth calibration curve. We find that fitting the PCA model using only a few scenes (21 in our case) provides a good enough approximation to capture the variation in the test curves (see Sec. 4). Here, \\(\\mathbf{V} = [\\mathbf{v}_k] \\in \\mathbb{R}^{K \\times M}\\) contains the ground truth calibration curves for the training scenes, with \\(K\\) representing the number of curves and \\(M\\) the sample count along each curve. PCA is then used to determine the basis vectors \\(\\mathbf{U} = (\\mathbf{u_1}, \\mathbf{u_2}, \\dots, \\mathbf{u_n})\\) and coefficients \\(\\theta = (\\alpha_1, \\alpha_2, \\dots, \\alpha_n)\\), so that each calibration curve can be represented as:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.286, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.754, + 0.163 + ], + "angle": 0, + "content": "\\(\\mathbf{v}_k = \\sum_{i=1}^n \\alpha_i \\mathbf{u}_i\\). The parameters \\(\\theta\\) fully describe the calibration functions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.284 + ], + "angle": 0, + "content": "To find the optimal number of components, we compute the explained variance, and find that in our case most of the variance is explained using only \\( n = 3 \\) components (see experiments). To ensure the calibrator is monotonically increasing, we derive the final calibration function \\( R_{\\theta}(\\cdot) \\) using isotonic regression applied to the curve approximated by \\( \\pmb{\\theta} \\cdot \\mathbf{U} \\). The idea is that the low-dimensional representation of the calibration curves encoded in \\( \\mathbf{U} \\) will generalise to new target scenes without any additional scene-specific data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.788, + 0.526 + ], + "angle": 0, + "content": "Predicting Calibration Parameters. What remains is to estimate the calibration parameters, \\(\\theta\\), for a new scene. As we do not want to use additional held-out data from the target scene, we propose predicting these parameters using scene-specific features computed from the pretrained NeRF outputs. This approach is motivated by the human ability to visually identify inaccuracies in the renderings such as floaters and unnatural artifacts. Specifically, we use a Multi-Layer Perceptron (MLP) with three layers of output size: [128, 128, 3] and Leaky ReLU activations throughout except the last layer as the meta-calibrator to estimate \\(\\theta\\) given features extracted by the DINOv2 model [20] from rendered images \\((\\mathbf{f}_I)\\) and uncalibrated uncertainty maps \\((\\mathbf{f}_{\\kappa})\\). The goal here is to have DINOv2 extract features that describe the rendering imperfections, correlating with the calibration curve. We find that training the MLP model on only a few scenes (30 in our case) allows it to generalize well to new test scenes. Once trained, the meta-calibrator can predict the calibration curve of a new target scene as: \\(\\theta = MLP([\\mathbf{f}_I,\\mathbf{f}_{\\kappa}])\\), without using any additional ground truth data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.541, + 0.788, + 0.603 + ], + "angle": 0, + "content": "In summary, the meta-calibrator can correct the confidence levels of the model without requiring ground truth data at any stage, suggesting potential enhancements to applications that rely on uncertainty such as next-best view selection (see Sec. 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.626, + 0.377, + 0.645 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.794 + ], + "angle": 0, + "content": "The objective of the experiments is to: 1) validate that our approach achieves more accurate uncertainties (lower negative log-likelihood and calibration error) than state-of-the-art approaches for NeRF uncertainty estimation (Sec. 4.1); 2) demonstrate the meta-calibrator improves the accuracy of the uncalibrated uncertainties (decreases both the negative log-likelihood and calibration error) without requiring any held-out data from the target scene (Sec. 4.2); 3) explain the motivation for certain meta-calibrator design decisions; and 4) show that our uncertainties can be leveraged for applications such as next-best view planning (Sec. 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "For additional results showing: 1) why the PCA representation of the calibration curves is necessary; 2) that using the training set results in severe overfitting; 3) that holding out data results in poor performance at image reconstruction; 4)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.128 + ], + "angle": 0, + "content": "N. Amini-Naieni et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "the influence of the number of samples along the ray on the uncertainty quality; and 5) the efficiency of our uncertainty metric (Eq. (3)) over other approaches, please refer to Sections 1, 2, 3, 4, and 6 of the Supplementary respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.204, + 0.785, + 0.251 + ], + "angle": 0, + "content": "Metrics and calibration curves. We use a variant of the calibration error from [12] to evaluate the effectiveness of the meta-calibrator. Specifically, given a test set \\(\\mathcal{D} = \\{(\\mathbf{r_t},\\mathbf{c_t})\\}_{t = 1}^T = \\{(\\mathbf{r_t},(r_t,g_t,b_t))\\}_{t = 1}^T\\), we report:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.261, + 0.785, + 0.3 + ], + "angle": 0, + "content": "\\[\nE R R = \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\left(p _ {t} - \\hat {P} \\left(p _ {t}\\right)\\right) ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.784, + 0.355 + ], + "angle": 0, + "content": "where \\( p_t \\) is the expected confidence level for data point \\( (\\mathbf{r_t}, c_t) \\), and \\( \\hat{P}(p_t) \\) is the empirical frequency of data points within that confidence level. More specifically, for each \\( t \\in \\{1, \\dots, T\\} \\), we set \\( p_t = M_t^C(c_t) \\) and," + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.358, + 0.785, + 0.378 + ], + "angle": 0, + "content": "\\[\n\\hat {P} (p) = \\left| \\left\\{c _ {t}: M _ {t} ^ {C} \\left(c _ {t}\\right) \\leq p \\text {f o r} t = 1, \\dots , T \\right\\} \\right| / T, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.784, + 0.443 + ], + "angle": 0, + "content": "where \\( M \\equiv F \\) for uncalibrated errors and \\( M \\equiv \\hat{F} \\) for calibrated errors, and \\( C \\in \\{R, G, B\\} \\). Note that this formulation of the calibration error is equivalent to the one in [12] with a confidence level for every unique \\( p_t \\) and weights that more significantly penalize errors from frequently predicted confidence levels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.785, + 0.595 + ], + "angle": 0, + "content": "Following [12], we plot \\(\\{(p_t, \\hat{P}(p_t))\\}_{t=1}^T\\) before and after calibration for each color channel to generate calibration curves. A perfectly calibrated forecaster would produce the straight line \\(p_t = \\hat{P}(p_t)\\) as each expected confidence level would equal the empirical one. Intuitively, our version of the calibration error is the mean squared vertical distance of points on the calibration curve from a perfectly straight line. If an expected confidence level occurs \\(N\\) times in the test data, its distance is counted \\(N\\) times in the mean. Following [29], we additionally report the negative log-likelihood (NLL) of the test data averaged across all scenes. Following [24], we include PSNR and LPIPS [32] to evaluate image quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.606, + 0.784, + 0.667 + ], + "angle": 0, + "content": "Datasets. We use 30 scenes from 3 datasets: Realistic Synthetic \\(360^{\\circ}\\) [16], the subset of scenes in DTU [8] used in [24], and LLFF [15] for training the metacalibrator and test it on a hold-out scene from either LLFF or DTU to show it generalizes to new target scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.678, + 0.785, + 0.769 + ], + "angle": 0, + "content": "Baselines We compare our approach against the state-of-the-art method for NeRF uncertainty estimation DANE [29] as well as other methods in Sec. 4.1. In Tab. 1, following [29], we implement the naive ensembles approach and DANE using a public implementation of Instant-NGP [2,17] and 5 ensemble members. In Sec. 4.2, we compare the uncalibrated uncertainties to the uncertainties calibrated by our meta-calibrator." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Meta-calibrator Design The results guiding our decisions to use 3 Principal Component Analysis (PCA) components to represent the calibration curves, fit the PCA components using 21 training scenes, and train the meta-calibrator on 30 scenes to predict the PCA coefficients are shown in Fig. 3." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.287, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.224, + 0.391, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.328, + 0.391, + 0.38 + ], + "angle": 0, + "content": "(a) Explained variance plot showing most of the variance of the calibration curves is explained using only 3 PCA components." + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.204, + 0.59, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.414, + 0.288, + 0.589, + 0.38 + ], + "angle": 0, + "content": "(b) Graph showing calibration error on new scenes is sufficiently low when using 21 scenes to fit 3-coefficient PCA model (i.e. it's an order of magnitude lower than we can expect from the final calibration, meaning good generalization to unseen scenes)." + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.203, + 0.787, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.288, + 0.787, + 0.38 + ], + "angle": 0, + "content": "(c) Graph showing increasing number of training scenes for meta-calibrator decreases calibration error for the test scene \\( T \\)-Rex from LLFF [15], with 30 scenes resulting in good generalization to the held-out scene and well-calibrated uncertainties." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.392, + 0.788, + 0.446 + ], + "angle": 0, + "content": "Fig. 3: Meta-calibrator design decisions. Results showing using 3 components for Principal Component Analysis (PCA) model of calibration curves, 21 scenes to fit PCA model, and 30 scenes to train meta-calibrator achieves good generalization to new test scenes." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.688 + ], + "angle": 0, + "content": "Table 1: Quantitative results on standard sparse NeRF benchmark. Our proposed approach results in significantly better uncertainties and image quality than the state-of-the-art NeRF uncertainty estimation method DANE [29] does on the challenging 3-view LLFF [15] dataset. Specifically, our meta-calibrator reduces the calibration error to \\(6\\%\\) of DANE's calibration error and the negative log-likelihood to be over 100 \\(\\%\\) lower than DANE's. Note: lower calibration error (Cal. Err.) and negative log-likelihood (NLL) values indicate more accurate uncertainties. Results are averaged over all 8 scenes in LLFF." + }, + { + "type": "table", + "bbox": [ + 0.377, + 0.7, + 0.625, + 0.766 + ], + "angle": 0, + "content": "
Uncert aintyImageQuality
Cal. Err. (↓)NLL (↓)PSNR (↑)LPIPS (↓)
Naïve Ens.0.05054.3915.190.646
DANE [29]0.04413.7515.190.646
Ours0.0026-0.6819.340.235
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.128 + ], + "angle": 0, + "content": "N. Amini-Naieni et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.519, + 0.162 + ], + "angle": 0, + "content": "4.1 Comparison to State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.324 + ], + "angle": 0, + "content": "In this section, we compare our approach to prior methods for NeRF uncertainty estimation. We achieve more accurate uncertainties (94% reduction in calibration error and over 100% reduction in negative log-likelihood) than those estimated by DANE [29], the state-of-the-art method. These results are shown in Tab. 1 for the challenging 3-view LLFF [15] dataset from prior work on sparse novel view synthesis [19,24,25] and Tab. 2 for the less challenging version of LLFF from prior work on NeRF uncertainty estimation [14,26,27,29]. In Fig. 5a, we compare the calibration curves from our approach to DANE's, illustrating that the meta-calibrator predicts expected confidences that match the true ones while DANE does not." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.347, + 0.784, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.513, + 0.788, + 0.653 + ], + "angle": 0, + "content": "Fig. 4: Quantitative comparison of uncalibrated and calibrated uncertainties. In (a), we show calibration curves on test data from four scenes in LLFF [15]. The color of each curve indicates the color channel it corresponds to. The calibrated curves are much closer to the ideal calibration (dashed lines), demonstrating that the meta-calibrator works very well. In (b), the average calibration error and negative log-likelihood before and after calibration are reported for LLFF, clearly showing the meta-calibrator improves the accuracy of the uncertainties (lowering calibration error and negative log-likelihood). To test generalization, the meta-calibrator was also applied to held-out scenes in DTU [8], achieving a \\(70\\%\\) reduction in calibration error on average." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.709, + 0.614, + 0.724 + ], + "angle": 0, + "content": "4.2 Comparison to Uncalibrated Uncertainties" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this section, we compare our uncalibrated base NeRF uncertainties to our calibrated uncertainties obtained from applying the proposed meta-calibrator. In Fig. 6 we show that the calibrated uncertainties better highlight floaters and other errors in the NeRF renderings. In Fig. 4, we show that the meta-calibrator predicts expected confidence that closely match the true ones, lowering both the calibration error and the negative log-likelihood of the uncalibrated uncertainties." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.286, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.31 + ], + "angle": 0, + "content": "Table 2: Quantitative results on standard NeRF uncertainty estimation benchmark. Here, we present results on the LLFF [15] dataset used in prior work [14, 26, 27, 29] on uncertainty estimation for NeRFs. This dataset is less challenging than the one in Tab. 1 since 4-12 views are used for training instead of 3. Our proposed approach results in significantly better uncertainties than prior methods for NeRF uncertainty estimation on all 8 scenes in LLFF. Note: lower negative log-likelihood values indicate more accurate uncertainties. \\(M\\) indicates the number of ensemble members, and MC-DO refers to Monte Carlo Dropout sampling with \\(M\\) sample configurations. This table is Tab. 1 from [29] with our results added as an additional column. Please refer to [29] for further details." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.317, + 0.797, + 0.442 + ], + "angle": 0, + "content": "
Scene# of Train. ViewsNegative Log-likelihood (↓)
MC-DO M=5Naïve Ens. M=5NeRF-W [14]S-NeRF [27]CF-NeRF [26]DANE [29] M=5DANE [29] M=10Ours
Fern44.902.472.162.01—-0.98-1.00-1.41
Orchids55.742.232.241.95—-0.28-0.31-0.84
Leaves52.722.660.790.68—0.970.73-1.19
Flower74.631.631.711.27—1.000.85-2.05
Fortress85.192.291.04-0.03—-1.30-1.30-1.99
Room85.062.134.932.35—-1.35-1.35-2.17
T-Rex114.102.281.911.37—-0.31-0.69-1.49
Horns124.182.170.780.60—-0.55-0.66-2.18
Avg.4.572.231.951.270.57-0.35-0.47-1.67
" + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.506, + 0.485, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.628, + 0.501, + 0.66 + ], + "angle": 0, + "content": "(a) Calibration curves for the Fern scene from 3-view LLFF [15]." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.526, + 0.664, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.638, + 0.676, + 0.66 + ], + "angle": 0, + "content": "(b) Next-best View Planning on Horns from LLFF [15]." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.672, + 0.788, + 0.811 + ], + "angle": 0, + "content": "Fig. 5: Comparison to DANE [29]. Results comparing our uncertainties to those from the state-of-the-art method DANE. In (a) we show DANE's RGB calibration curves are not closely aligned with the perfectly calibrated lines, meaning it is miscalibrated. It is significantly over-confident for expected confidence levels close to 1 and under-confident for confidence levels close to 0. In comparison, the curves for our approach are extremely close to the ideal calibration (dashed lines), demonstrating that the meta-calibrator works very well, predicting expected confidences that match the true ones. This is also verified by how our calibration error is over two orders of magnitude smaller than DANE's. In (b) we show that our approach results in more efficient performance gains over DANE for next-best view planning." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.128 + ], + "angle": 0, + "content": "N. Amini-Naieni et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.143, + 0.332, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.255, + 0.276, + 0.262 + ], + "angle": 0, + "content": "(a) Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.144, + 0.445, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.255, + 0.386, + 0.262 + ], + "angle": 0, + "content": "(b) NeRF Image" + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.144, + 0.557, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.448, + 0.256, + 0.537, + 0.263 + ], + "angle": 0, + "content": "(c) Uncalibrated Uncertainty" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.145, + 0.672, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.563, + 0.256, + 0.645, + 0.263 + ], + "angle": 0, + "content": "(d) Calibrated Uncertainty" + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.144, + 0.785, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.255, + 0.725, + 0.262 + ], + "angle": 0, + "content": "(e) Actual Error" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.278, + 0.788, + 0.349 + ], + "angle": 0, + "content": "Fig. 6: Qualitative comparison of uncalibrated and calibrated uncertainties from the Flower scene in LLFF [15]. The calibrated uncertainties (d) clearly detect incorrect regions (indicated by the red boxes) better than the uncalibrated uncertainties (c) do. This is apparent by noting that (d) and (e) look more similar than (c) and (e)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.378, + 0.579, + 0.394 + ], + "angle": 0, + "content": "4.3 Application: Next-best View Planning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.789, + 0.646 + ], + "angle": 0, + "content": "In this section, we show that our uncertainties can be leveraged for next-best view planning. Specifically, we start by training the NeRF model for 2000 iterations on a training set of three images. The next-best view is selected by evaluating the average calibrated pixel uncertainty (obtained using the meta-calibrator) for each of the candidate views, and the view with the highest uncertainty is added to the training set. The average PSNR of the test images is reported after each training iteration. In Fig. 5b we show that using our approach results in greater performance gains (higher PSNRs) than DANE [29] does. In the Fig. 1 of the Supplementary, we also compare the information gain from rays selected according to the highest calibrated uncertainties to the information gain from rays selected according to the highest uncalibrated uncertainties and show our meta-calibrator produces higher average PSNRs on the test set for scenes in DTU [8] than the uncalibrated uncertainties do. Thus, calibration specifically re-orders the pixel uncertainties so that rays more likely to raise the PSNR are picked earlier in next-best view planning. In Sec. 5 of the Supplementary, we include a detailed theoretical example showing that such re-ordering is possible." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.668, + 0.36, + 0.684 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.7, + 0.789, + 0.835 + ], + "angle": 0, + "content": "In this paper we addressed the open problem of obtaining calibrated uncertainties from NeRF models. We introduce the concept of a meta-calibrator that infers the calibration curves from scene features, and using this approach achieve state-of-the-art uncertainty without holding out any ground truth data from the target scene. By enabling efficient and accurate calibration of NeRF models without relying on additional data, our method represents a significant step forward in the practical application of NeRF to real-world scenarios and opens up new avenues for the use of NeRF in situations where data is limited and uncertainty is critical." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.286, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "Acknowledgements. The authors would like to thank Seunghyeon Seo for his support on FlipNeRF, and Oishi Deb for her insightful feedback on uncertainty estimation. Niki Amini-Naieni is funded by an AWS Studentship, the Reuben Foundation, and the AIMS CDT program at the University of Oxford. Tomas Jakab is sponsored by EPSRC VisualAI EP/T028572/1 and Andrea Vedaldi by ERC-CoG UNION 101001212." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.262, + 0.323, + 0.278 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.295, + 0.786, + 0.337 + ], + "angle": 0, + "content": "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV. pp. 5835-5844 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.786, + 0.365 + ], + "angle": 0, + "content": "2. Bhalgat, Y.: Hashnerf-pytorch. https://github.com/yashbhalgat/HashNeRF-pytorch/ (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.366, + 0.786, + 0.393 + ], + "angle": 0, + "content": "3. Gafni, G., Thies, J., Zollhöfer, M., Nießner, M.: Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In: CVPR. pp. 8649-8658 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.394, + 0.786, + 0.421 + ], + "angle": 0, + "content": "4. Ghoshal, B., Tucker, A.: On calibrated model uncertainty in deep learning. In: ECML (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.422, + 0.786, + 0.449 + ], + "angle": 0, + "content": "5. Goli, L., Reading, C., Sellan, S., Jacobson, A., Tagliasacchi, A.: Bayes' Rays: Uncertainty quantification in neural radiance fields. ArXiv abs/2309.03185 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.45, + 0.786, + 0.478 + ], + "angle": 0, + "content": "6. Jain, A., Tancik, M., Abbeel, P.: Putting nef on a diet: Semantically consistent few-shot view synthesis. In: ICCV. pp. 5865-5874 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.479, + 0.786, + 0.505 + ], + "angle": 0, + "content": "7. Jang, T.J., Hyun, C.M.: Nerf solves undersampled mri reconstruction. ArXiv abs/2402.13226 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.506, + 0.786, + 0.533 + ], + "angle": 0, + "content": "8. Jensen, R.R., Dahl, A., Vogiatzis, G., Tola, E., Aanaes, H.: Large scale multi-view stereopsis evaluation. In: CVPR (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.534, + 0.786, + 0.561 + ], + "angle": 0, + "content": "9. Jin, L., Chen, X., Ruckin, J., Popovi'c, M.: Neu-nbv: Next best view planning using uncertainty estimation in image-based neural rendering. In: IROS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.562, + 0.786, + 0.589 + ], + "angle": 0, + "content": "0. Kajiya, J.T., Von Herzen, B.P.: Ray tracing volume densities. In: SIGGRAPH. p. 165-174 (1984)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.59, + 0.786, + 0.631 + ], + "angle": 0, + "content": "1. Kosiorek, A.R., Strathmann, H., Zoran, D., Moreno, P., Schneider, R., Mokr'a, S., Rezende, D.J.: Nerf-vae: A geometry aware 3d scene generative model. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.632, + 0.786, + 0.659 + ], + "angle": 0, + "content": "2. Kuleshov, V., Fenner, N., Ermon, S.: Accurate uncertainties for deep learning using calibrated regression. In: ICML. pp. 2796-2804 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.786, + 0.687 + ], + "angle": 0, + "content": "3. Liu, L., Gu, J., Lin, K.Z., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. In: NIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.786, + 0.728 + ], + "angle": 0, + "content": "4. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.786, + 0.77 + ], + "angle": 0, + "content": "5. Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. In: TOG (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.786, + 0.812 + ], + "angle": 0, + "content": "6. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "7. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. In: ACM Trans. Graph. (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.295, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.421, + 0.127 + ], + "angle": 0, + "content": "N. Amini-Naieni et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "18. Neff, T., Stadlbauer, P., Parger, M., Kurz, A., Mueller, J.H., Chaitanya, C.R.A., Kaplanyan, A., Steinberger, M.: Donerf: Towards real-time rendering of compact neural radiance fields using depth oracle networks. In: CGF. pp. 45-59 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.786, + 0.232 + ], + "angle": 0, + "content": "19. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S.M., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.786, + 0.3 + ], + "angle": 0, + "content": "20. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Howes, R., Huang, P.Y., Xu, H., Sharma, V., Li, S.W., Galuba, W., Rabbat, M., Assran, M., Ballas, N., Synnaeve, G., Misra, I., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.786, + 0.328 + ], + "angle": 0, + "content": "21. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.356 + ], + "angle": 0, + "content": "22. Ran, Y., Zeng, J., He, S., Li, L., Chen, Y., Lee, G.H., Chen, J., Ye, Q.: Neurar: Neural uncertainty for autonomous 3d reconstruction. In: RAL (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.786, + 0.384 + ], + "angle": 0, + "content": "23. Rebain, D., Jiang, W., Yazdani, S., Li, K., Yi, K.M., Tagliasacchi, A.: Derf: Decomposed radiance fields. In: CVPR. pp. 14148-14156 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.786, + 0.411 + ], + "angle": 0, + "content": "24. Seo, S., Chang, Y., Kwak, N.: Flipnerf: Flipped reflection rays for few-shot novel view synthesis. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.786, + 0.453 + ], + "angle": 0, + "content": "25. Seo, S., Han, D., Chang, Y., Kwak, N.: Mixnerf: Modeling a ray with mixture density for novel view synthesis from sparse inputs. In: CVPR. pp. 20659-20668 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.481 + ], + "angle": 0, + "content": "26. Shen, J., Agudo, A., Moreno-Noguer, F., Ruiz, A.: Conditional-flow nerf: Accurate 3d modelling with reliable uncertainty quantification. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.786, + 0.509 + ], + "angle": 0, + "content": "27. Shen, J., Ruiz, A., Agudo, A., Moreno-Noguer, F.: Stochastic neural radiance fields: Quantifying uncertainty in implicit 3d representations. In: 3DV. pp. 972-981 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.786, + 0.536 + ], + "angle": 0, + "content": "28. Sitzmann, V., Martel, J.N., Bergman, A.W., Lindell, D.B., Wetzstein, G.: Implicit neural representations with periodic activation functions. In: NIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.786, + 0.564 + ], + "angle": 0, + "content": "29. Sünderhauf, N., Abou-Chakra, J., Miller, D.: Density-aware nef ensembles: Quantifying predictive uncertainty in neural radiance fields. In: ICRA (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.786, + 0.605 + ], + "angle": 0, + "content": "30. Wang, P., Liu, Y., Chen, Z., Liu, L., Liu, Z., Komura, T., Theobalt, C., Wang, W.: F2-nerf: Fast neural radiance field training with free camera trajectories. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.786, + 0.633 + ], + "angle": 0, + "content": "31. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. ArXiv abs/2010.07492 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.786, + 0.661 + ], + "angle": 0, + "content": "32. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.661 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator/6c9ab90c-9fb8-405a-a3f6-eaa68a4cea3f_origin.pdf b/2024/Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator/6c9ab90c-9fb8-405a-a3f6-eaa68a4cea3f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..eea6f19b8d38189e66d5f24d2e07c463a2f6e091 --- /dev/null +++ b/2024/Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator/6c9ab90c-9fb8-405a-a3f6-eaa68a4cea3f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:569850b1fa0dc64341fbf3f8b8c82023b05b2d5034cfd1ca734fe3164de93380 +size 4091365 diff --git a/2024/Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator/full.md b/2024/Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ce3edf35b97b46a91bed0cb303ad00a37ea124e2 --- /dev/null +++ b/2024/Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator/full.md @@ -0,0 +1,253 @@ +# Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator + +Niki Amini-Naieni1, Tomas Jakab1, Andrea Vedaldi1, and Ronald Clark1 + +University of Oxford + +Abstract. Neural Radiance Fields (NeRFs) have markedly improved novel view synthesis, but accurate uncertainty quantification in their image predictions remains an open problem. The prevailing methods for estimating uncertainty, including the state-of-the-art Density-aware NeRF Ensembles (DANE) [29], quantify uncertainty without calibration. This frequently leads to over- or under-confidence in image predictions, which can undermine their real-world applications. In this paper, we propose a method which, for the first time, achieves calibrated uncertainties for NeRFs. To accomplish this, we overcome a significant challenge in adapting existing calibration techniques to NeRFs: a need to hold out ground truth images from the target scene, reducing the number of images left to train the NeRF. This issue is particularly problematic in sparse-view settings, where we can operate with as few as three images. To address this, we introduce the concept of a meta-calibrator that performs uncertainty calibration for NeRFs with a single forward pass without the need for holding out any images from the target scene. Our meta-calibrator is a neural network that takes as input the NeRF images and uncalibrated uncertainty maps and outputs a scene-specific calibration curve that corrects the NeRF's uncalibrated uncertainties. We show that the meta-calibrator can generalize on unseen scenes and achieves well-calibrated and state-of-the-art uncertainty for NeRFs, significantly beating DANE and other approaches. This opens opportunities to improve applications that rely on accurate NeRF uncertainty estimates such as next-best view planning and potentially more trustworthy image reconstruction for medical diagnosis. + +Keywords: NeRFs $\cdot$ Uncertainty calibration $\cdot$ Few-shot learning + +# 1 Introduction + +Recent advancements in scene representations have led to promising new approaches for novel view synthesis and scene reconstruction. Among these, Neural Radiance Fields (NeRFs) [16] have emerged as a particularly powerful tool, offering unprecedented levels of realism and detail in rendered images. The core idea behind NeRFs is to represent a scene as a continuous vector-valued function, parameterized by a neural network, that maps spatial coordinates and view directions to color and density values. This approach enables the creation + +![](images/931aa5f9312feb8cc27913728d089273215c2cacdf110ac6aad8da7b08aa1d22.jpg) +Fig. 1: We propose a method for efficiently calibrating the uncertainties from NeRF models. Our approach is based on a meta-calibrator that takes as input features from the rendered NeRF images and uncalibrated uncertainty maps and predicts the calibration function, $R_{\theta}(\cdot)$ , for the NeRF model. Our meta-calibrator generalizes across scenes, so it only needs to be trained once, and can predict the calibration function in a single forward pass without any ground truth data from the target scene. + +of detailed 3D representations from sets of 2D images, revolutionizing 3D reconstruction. + +However, despite their impressive capabilities, traditional NeRF models lack an essential component: an accurate measure of uncertainty in their predictions. Accurate uncertainties are crucial for applying NeRFs to safety-critical problems such as MRI image reconstruction from sparse data [7], where unreliable confidence estimates could lead to misdiagnosis. More accurate uncertainties could also enhance practical methods such as uncertainty-guided next-best view planning techniques [9]. Prior approaches have attempted to estimate NeRF uncertainties [5, 14, 26, 27, 29], but they all overlook the problem of calibration. Thus, the uncertainties they output are not as accurate as they could be. + +In particular, the state-of-the-art uncertainty estimation method for NeRFs, Density-aware NeRF Ensembles (DANE) [29], produces uncalibrated uncertainties. As a result, the confidence intervals and variances do not match the true confidence, meaning it has limited applicability to real-world problems. This constraint is significant as it restricts the use of NeRFs in safety-critical and sparse-data settings, where knowing the confidence in predictions is crucial. + +The best NeRF methods for the sparse-view setting overlook the problem of calibration as well. FlipNeRF [24], the state-of-the-art technique for sparse-view reconstruction, uses uncertainties from an uncalibrated mixture of Laplacians to enhance its training process. Therefore, the uncertainties it outputs at inference as an artifact of training are not accurate. + +In this paper, we present a novel approach for obtaining calibrated uncertainties for NeRF models in the sparse-view setting. Our strategy integrates the Laplacian mixture from FlipNeRF [24] with the calibration techniques by Kuleshov et al [12]. However, naively applying the calibration method by Kuleshov et al to FlipNeRF does not work due to a significant challenge of the sparse-view setting: there is a lack of held-out data from the target scene for fitting the calibrator. Specifically, holding out just one image for calibration could decrease the size of the training set by over $30\%$ , resulting in significant performance degradation of the NeRF. To overcome this, we make use of a unique observation: while calibration curves exhibit significant variation across scenes, they also demonstrate a significant regularity in their structure. Utilizing this insight, we propose the concept of a meta-calibrator that learns a low-dimensional representation of the NeRF calibration curves and infers them from scene features. We motivate and show why this meta-calibrator is necessary and demonstrate that it achieves more accurate uncertainties than DANE [29] without holding out any images from the target scene. + +Specifically, our contributions are: (1) the first investigation into obtaining calibrated uncertainties from NeRFs, (2) a novel meta-calibrator for fitting the calibration model without using held-out data, and (3) experiments on the real-world Local Light Field Fusion (LLFF) [15] and DTU [8] datasets showing that our meta-calibrator achieves state-of-the-art and well-calibrated uncertainties for real scenes. We also demonstrate that our uncertainties can be leveraged for effective next-best view planning. + +# 2 Related Work + +Neural Radiance Fields (NeRFs). NeRFs [16] are a popular method for novel view synthesis. From a set of 2D images, NeRFs learn a neural network representation of a single scene. A trained NeRF model outputs estimates of the volume density and emitted radiance at any 3D location and viewing direction. Novel views can be generated by applying volume rendering [10] to the density and radiance values predicted by the NeRF model for points along rays cast into the scene. Due to their simplicity and impressive performance, NeRFs have become a popular technique for solving a variety of rendering problems. + +Over the last few years, several extensions of NeRFs have been explored [11, 18, 21, 28, 31]. These include speeding up training and inference [13, 23, 30], modeling dynamic scenes [3], learning from a sparse set of training views [1, 6, 19, 24, 25], and estimating the uncertainty in NeRF predictions [5, 14, 22, 26, 27, 29]. Sparse NeRF methods aim to accurately render novel views when only a few training views are available from the target scene. NeRF uncertainty estimation techniques strive to accurately predict the confidence in the views rendered. Although uncertainty estimation is particularly important in the sparse-view + +setting, where NeRF renderings are especially unreliable, the main aim of sparse NeRF methods is not to output accurate uncertainties. + +Sparse-view NeRF Methods. Despite this, recent sparse NeRF methods do produce uncertainties as an artifact of their training process. Both MixNeRF [25] and FlipNeRF [24], the state-of-the-art approach, model the RGB color channels given a ray as independent random variables that follow a mixture model. FlipNeRF further uses the uncertainties of the pixel colors to regularize the training process, producing more accurate image reconstructions at inference. However, neither MixNeRF nor FlipNeRF outputs calibrated uncertainties. + +Our method significantly extends sparse NeRF methods to obtain more accurate and well-calibrated uncertainties at inference. To benefit from the superior performance of FlipNeRF [24] at sparse novel view synthesis, we apply the proposed meta-calibrator to the learned distribution from FlipNeRF, producing more accurate uncertainties without sacrificing state-of-the-art image quality. However, our approach can be applied to any NeRF method that outputs uncertainties, so it is distinct from FlipNeRF and MixNeRF. In essence, the proposed meta-calibrator augments sparse NeRF methods to achieve state-of-the-art uncertainty, beating both FlipNeRF and techniques designed explicitly for NeRF uncertainty estimation. + +Uncertainty in NeRFs. The growing line of methods specifically designed for accurately estimating NeRF uncertainties [5,14,22,26,27,29] do not address the problem of calibration. The current state-of-the-art uncertainty estimation technique for NeRFs is Density-aware NeRF Ensembles (DANE) [29]. DANE adds an epistemic uncertainty term to a naive ensembles approach with five ensemble members. Thus, DANE is very costly as it requires training five NeRFs to obtain uncertainty estimates. Another NeRF uncertainty estimation approach is Stochastic Neural Radiance Fields (S-NeRF) [27], which learns a probability distribution over all possible radiance fields by modeling the volume density and radiance as random variables that follow a joint distribution. S-NeRF employs variational inference to sample from an approximation to this distribution and uses the variances of the sampled pixel colors as the estimated uncertainties. Conditional-flow NeRF (CF-NeRF) [26] builds on S-NeRF by combining latent variable modeling and conditional normalizing flows to relax the strong constraints S-NeRF imposes over the radiance distribution. Despite the growing number of techniques in this area of study, all prior work does not consider calibration, outputting unreliable uncertainties as a result. Our work achieves more accurate uncertainties than these prior methods by filling the gap of uncertainty calibration for NeRFs and drawing on well-established techniques in calibrated regression [12]. + +Uncertainty Calibration in Deep Learning. While uncertainty calibration has been studied for Bayesian deep learning methods [4, 12], it has not been adapted for or applied successfully to NeRF uncertainty estimates. This may be + +because NeRFs introduce additional complexity in the uncertainty estimation process as the neural network model needs to be trained per-scene. + +This makes uncertainty calibration challenging as methods for calibrated regression [12] require either using the training set or held-out data to achieve calibrated uncertainties. Using the training set to fit the calibrator results in severe overfitting (see Sec. 2 of the Supplementary). Holding out data from the target scene for calibration means there is less data to train the NeRF, making it more inaccurate at novel view synthesis (see Sec. 3 of the Supplementary). Thus, a trivial application of [12] to NeRFs would not be satisfactory. In this work, we propose the concept of a meta-calibrator that, in contrast to [12], does not require held-out data and achieves calibrated uncertainty estimates for NeRFs. + +# 3 Method + +In this paper, we present a method that calibrates NeRF uncertainties. To this end, we propose a novel meta-calibrator that accepts uncalibrated NeRF uncertainties and predicted images as inputs and outputs a scene-specific calibration curve, correcting the uncalibrated confidence levels. An overview of our method can be seen in Fig. 1. Crucially, our approach does not require holding out any images from the target scene. Thus, it can be applied to sparse-view settings, where holding out a single image could significantly harm the NeRF's performance. In Sec. 3.1, we explain the necessary background concepts, in Sec. 3.2, we describe how we obtain the corrected uncertainty values, and in Sec. 3.3 we detail our meta-calibrator. + +# 3.1 Preliminaries + +Neural Radiance Fields (NeRFs). Neural Radiance Fields (NeRFs) [16] represent a scene as a continuous vector-valued function with inputs a Cartesian point $\mathbf{x} = (x,y,z)$ and unit viewing direction vector $\mathbf{d} = (u,v,w)$ and outputs an emitted radiance $\pmb {c} = (\nu ,\varrho ,\theta)$ and volume density $\sigma$ . By optimizing the weights $\Theta$ of a neural network approximation $\mathbf{F}_{\Theta}$ to this representation, NeRFs can render the color of any pixel in a synthetic image of the scene. To achieve this, principles from classical volume rendering [10] are applied to the radiance and density values estimated by $\mathbf{F}_{\Theta}$ for points along a ray cast from the origin $\mathbf{x_0}$ of the virtual camera, through the pixel, and into the scene. More specifically, the expected color $\mathbf{c}(\mathbf{r})$ of a camera ray $\mathbf{r}(t) = \mathbf{x_0} + t\mathbf{d}$ with near and far bounds $t_n$ and $t_f$ is: + +$$ +\mathbf {c} (\mathbf {r}) = \int_ {t _ {n}} ^ {t _ {f}} T (t) \sigma (\mathbf {r} (t)) \boldsymbol {c} (\mathbf {r} (t), \mathbf {d}) d t, \tag {1} +$$ + +where $T(t) = e^{-\int_{t_n}^t \sigma(\mathbf{r}(s)) ds}$ . The integral in Eq. (1) is estimated with numerical quadrature to obtain the colors of the pixels in the synthetic image from the NeRF model outputs. NeRF optimizes $\Theta$ according to Eq. (1) with gradient + +descent, since the numerical quadrature is differentiable. Conventional NeRFs do not provide an uncertainty associated with their predictions, so extensions like DANE [29] have been developed. + +Base NeRF Uncertainties. To obtain the initial uncertainties, we have two options for our base model: (1) DANE [29], the state-of-the-art method for NeRF uncertainty estimation or (2) FlipNeRF [24], the state-of-the-art method for sparse novel view synthesis. DANE requires training five NeRFs per scene and provides poor image quality in the sparse-view setting, so we choose FlipNeRF. We show in the experiments that applying our meta-calibrator to FlipNeRF results in more accurate and well-calibrated uncertainties than those output by DANE. Our base FlipNeRF uncertainties are inferred from a mixture of Laplacians with location and scale parameters learned during training. + +FlipNeRF [24] models the joint distribution of the color $\mathbf{C} = (R,G,B)$ given a ray $\mathbf{r}$ with a mixture of Laplacians: + +$$ +p (\mathbf {C} = \mathbf {c} | \mathbf {r}) = \sum_ {j = 1} ^ {M} \pi_ {j} \mathcal {F} (\mathbf {C} = \mathbf {c}; \boldsymbol {\mu} _ {j}, \boldsymbol {\beta} _ {j}), \tag {2} +$$ + +where $M$ is the number of sampled points along the ray $r$ . $\mathcal{F}(\mathbf{C} = \mathbf{c};\pmb{\mu}_j,\pmb{\beta}_j)$ is the 3D Laplacian probability density with location parameter $\pmb{\mu}_{j} = (\mu_{j}^{R},\mu_{j}^{G},\mu_{j}^{B})$ and scale parameter $\beta_{i} = (\beta_{j}^{R},\beta_{j}^{G},\beta_{j}^{B})$ evaluated at the color $\mathbf{c}$ . More specifically, the mixing coefficients $\{\pi_j\}_{j = 1}^M$ are the normalized coefficients of the radiance values along a ray in Eq. (2), the location parameters $\{\pmb{\mu}_{j}\}_{j = 1}^{M}$ are the estimated RGB radiance values, and the scale parameters $\{\pmb{\beta}_{j}\}_{j = 1}^{M}$ are an additional output of the model. These parameters are optimized by FlipNeRF during training. + +We now go into detail about how the location and scale parameters are obtained. FlipNeRF [24] learns these parameters by minimizing the negative log-likelihood of the training set $\mathcal{D} = \{(\mathbf{r_i},\mathbf{c_i})\}_{i = 1}^N$ containing the rays $\mathbf{r_i}$ and colors $\mathbf{c_i}$ from the pixels in the ground truth images of the scene assuming the distribution in Eq. (2). FlipNeRF additionally minimizes the average scale parameters through an auxiliary uncertainty-aware emptiness loss for reducing floating artifacts. The negative log-likelihood loss and the uncertainty-aware emptiness loss are added to FlipNeRF's total training loss, which incorporates other terms such as the mean squared error. As a result of this training process, the location and scale parameters can be inferred by FlipNeRF at new poses. We obtain the uncalibrated confidence levels of predicted ray colors from these location and scale parameters. + +Using the distribution in Eq. (2), we can easily compute the confidence level for a given ray $\mathbf{r_t}$ , which we denote as $F_{t}$ . The CDF for the Laplacian mixture has a closed form expression parameterized by the location and scales output by the pretrained FlipNeRF [24]. Thus, we can use this CDF to predict the confidence level of the ground truth color $\mathbf{c_t}$ of any ray by evaluating it at the given + +color value, $p_t = F_t(\mathbf{c_t})$ . However, these initial confidence levels are uncalibrated and, hence, inaccurate. + +Calibrated regression. In [12], Kuleshov et al extend calibration methods for classification to regression. They define a forecaster $H: \mathcal{X} \to (\mathcal{Y} \to [0,1])$ as a function that outputs for each $x_{t} \in \mathcal{X}$ , a CDF $F_{t}$ . Given a pretrained forecaster $H$ , they suggest training an auxiliary model $R: [0,1] \to [0,1]$ by fitting $R$ to a recalibration dataset $D = \left\{\left([H(x_{t})](y_{t}), \hat{P}([H(x_{t})](y_{t}))\right)\right\}_{t=1}^{T}$ , where + +$$ +\hat {P} (p) = \left| \left\{y _ {t}: [ H \left(x _ {t}\right) ] \left(y _ {t}\right) \leq p \text {f o r} t = 1, \dots , T \right\} \right| / T +$$ + +is the empirical confidence level corresponding to the predicted confidence level $p$ . The fitted $R$ forms the calibration curve that corrects the expected confidence levels. We can now obtain predictive posterior values that closely match the true confidences using $\hat{F}_t \equiv R \circ F_t$ for the test data. + +However, as mentioned before, there are numerous challenges associated with applying this recalibration procedure. Firstly, note that it requires ground-truth values $(y_{t})$ for the predictions we are recalibrating. This means that in order to prevent overfitting we need to reserve part of the training dataset specifically for fitting the calibrator (which leaves less data for training the NeRF — a significant issue if we only have a few input views). Secondly, it does not actually prescribe how to compute a suitable uncertainty value from the predicted distribution. In Sec. 3.3, we address the former issue, and in Sec. 3.2, we address the latter. + +# 3.2 Calculating uncertainty + +The predictive posterior in Eq. (2) provides a distribution over likely ray colors for a NeRF model, but it does not inherently offer a straightforward metric for quantifying uncertainty at a specific point in the reconstruction. Intuitively, as the variance of this distribution increases, so does the uncertainty of the model's output at that point. Therefore, the variance or standard deviation is a popular choice for quantifying the uncertainty [24,26,27,29]. However, in the case of the corrected mixture distribution obtained from raymarching, this can be slow to compute especially if it has to be done for each pixel. Therefore, we turn to a metric that can be calculated directly from the calibrated CDF $\hat{F}^t$ . + +We propose to use the interquartile range of each calibrated distribution: + +$$ +\kappa^ {C} \left(\mathbf {r} _ {\mathbf {t}}\right) = \left[ \hat {F} _ {t} ^ {C} \right] ^ {- 1} \left(\frac {3}{4}\right) - \left[ \hat {F} _ {t} ^ {C} \right] ^ {- 1} \left(\frac {1}{4}\right), \tag {3} +$$ + +where $\kappa^C (\mathbf{r_t})$ represents the uncertainty at a ray $\mathbf{r_t}$ in the color channel $C$ . This difference provides a measure of the statistical dispersion and thus serves + +![](images/d4266a2b88d6ec9833101b28ab0e4ae1b897b32bf656aaae7ef21df07d7cbd39.jpg) +(a) Creating a parametric model of the calibration curves. + +![](images/fc20e2f750952699e6d84c4ea763729a17e3d80e0cab4ff95679774442b6f859.jpg) +(b) Predicting the curve parameters. +Fig. 2: Meta-calibrator design. In stage (a) we fit a low-dimensional parametric model of the calibration curves. The meta-calibrator then predicts these curve parameters from rendered images of the scene and their associated uncalibrated uncertainty maps (b). + +as a robust measure of the spread of the output channel. By averaging the interquartile range over the color channels, we obtain a single scalar value that effectively quantifies the uncertainty of the NeRF model for the given ray. As shown in our experiments, this method is very computational efficient, and it provides an accurate estimate of uncertainty. + +# 3.3 Meta-calibrator + +To overcome the challenge that, especially in the sparse-view setting, there is no held-out data available for fitting the calibrator, we propose a novel metacalibrator that infers the calibration curves from uncalibrated NeRF predictions. To do this, we leverage the insight that the calibration curves demonstrate significant regularity. We posit a low-dimensional model of the calibration curves can be learned and predicted using the images and uncalibrated uncertainty maps inferred by the NeRF, enabling us to estimate the calibration function without evaluating the empirical confidence levels using held-out data from the target scene. We now describe this meta-calibrator (illustrated in Fig. 2) in detail. + +A Parametric Model for Calibration Curves. We first fit a low-dimensional representation of the calibration curves using Principal Component Analysis (PCA). To create the training set for learning this representation, we sample held-out images from $K$ scenes and apply the calibration procedure by Kuleshov et al. [12] to form $K$ ground truth calibration curves. To construct the training vector $\mathbf{v_k} \in \mathbb{R}^{1 \times M}$ for scene $k \leq K$ , we sample $M$ evenly spaced points along its ground truth calibration curve. We find that fitting the PCA model using only a few scenes (21 in our case) provides a good enough approximation to capture the variation in the test curves (see Sec. 4). Here, $\mathbf{V} = [\mathbf{v}_k] \in \mathbb{R}^{K \times M}$ contains the ground truth calibration curves for the training scenes, with $K$ representing the number of curves and $M$ the sample count along each curve. PCA is then used to determine the basis vectors $\mathbf{U} = (\mathbf{u_1}, \mathbf{u_2}, \dots, \mathbf{u_n})$ and coefficients $\theta = (\alpha_1, \alpha_2, \dots, \alpha_n)$ , so that each calibration curve can be represented as: + +$\mathbf{v}_k = \sum_{i=1}^n \alpha_i \mathbf{u}_i$ . The parameters $\theta$ fully describe the calibration functions. + +To find the optimal number of components, we compute the explained variance, and find that in our case most of the variance is explained using only $n = 3$ components (see experiments). To ensure the calibrator is monotonically increasing, we derive the final calibration function $R_{\theta}(\cdot)$ using isotonic regression applied to the curve approximated by $\pmb{\theta} \cdot \mathbf{U}$ . The idea is that the low-dimensional representation of the calibration curves encoded in $\mathbf{U}$ will generalise to new target scenes without any additional scene-specific data. + +Predicting Calibration Parameters. What remains is to estimate the calibration parameters, $\theta$ , for a new scene. As we do not want to use additional held-out data from the target scene, we propose predicting these parameters using scene-specific features computed from the pretrained NeRF outputs. This approach is motivated by the human ability to visually identify inaccuracies in the renderings such as floaters and unnatural artifacts. Specifically, we use a Multi-Layer Perceptron (MLP) with three layers of output size: [128, 128, 3] and Leaky ReLU activations throughout except the last layer as the meta-calibrator to estimate $\theta$ given features extracted by the DINOv2 model [20] from rendered images $(\mathbf{f}_I)$ and uncalibrated uncertainty maps $(\mathbf{f}_{\kappa})$ . The goal here is to have DINOv2 extract features that describe the rendering imperfections, correlating with the calibration curve. We find that training the MLP model on only a few scenes (30 in our case) allows it to generalize well to new test scenes. Once trained, the meta-calibrator can predict the calibration curve of a new target scene as: $\theta = MLP([\mathbf{f}_I,\mathbf{f}_{\kappa}])$ , without using any additional ground truth data. + +In summary, the meta-calibrator can correct the confidence levels of the model without requiring ground truth data at any stage, suggesting potential enhancements to applications that rely on uncertainty such as next-best view selection (see Sec. 4.3). + +# 4 Experiments + +The objective of the experiments is to: 1) validate that our approach achieves more accurate uncertainties (lower negative log-likelihood and calibration error) than state-of-the-art approaches for NeRF uncertainty estimation (Sec. 4.1); 2) demonstrate the meta-calibrator improves the accuracy of the uncalibrated uncertainties (decreases both the negative log-likelihood and calibration error) without requiring any held-out data from the target scene (Sec. 4.2); 3) explain the motivation for certain meta-calibrator design decisions; and 4) show that our uncertainties can be leveraged for applications such as next-best view planning (Sec. 4.3). + +For additional results showing: 1) why the PCA representation of the calibration curves is necessary; 2) that using the training set results in severe overfitting; 3) that holding out data results in poor performance at image reconstruction; 4) + +the influence of the number of samples along the ray on the uncertainty quality; and 5) the efficiency of our uncertainty metric (Eq. (3)) over other approaches, please refer to Sections 1, 2, 3, 4, and 6 of the Supplementary respectively. + +Metrics and calibration curves. We use a variant of the calibration error from [12] to evaluate the effectiveness of the meta-calibrator. Specifically, given a test set $\mathcal{D} = \{(\mathbf{r_t},\mathbf{c_t})\}_{t = 1}^T = \{(\mathbf{r_t},(r_t,g_t,b_t))\}_{t = 1}^T$ , we report: + +$$ +E R R = \frac {1}{T} \sum_ {t = 1} ^ {T} \left(p _ {t} - \hat {P} \left(p _ {t}\right)\right) ^ {2}, \tag {4} +$$ + +where $p_t$ is the expected confidence level for data point $(\mathbf{r_t}, c_t)$ , and $\hat{P}(p_t)$ is the empirical frequency of data points within that confidence level. More specifically, for each $t \in \{1, \dots, T\}$ , we set $p_t = M_t^C(c_t)$ and, + +$$ +\hat {P} (p) = \left| \left\{c _ {t}: M _ {t} ^ {C} \left(c _ {t}\right) \leq p \text {f o r} t = 1, \dots , T \right\} \right| / T, \tag {5} +$$ + +where $M \equiv F$ for uncalibrated errors and $M \equiv \hat{F}$ for calibrated errors, and $C \in \{R, G, B\}$ . Note that this formulation of the calibration error is equivalent to the one in [12] with a confidence level for every unique $p_t$ and weights that more significantly penalize errors from frequently predicted confidence levels. + +Following [12], we plot $\{(p_t, \hat{P}(p_t))\}_{t=1}^T$ before and after calibration for each color channel to generate calibration curves. A perfectly calibrated forecaster would produce the straight line $p_t = \hat{P}(p_t)$ as each expected confidence level would equal the empirical one. Intuitively, our version of the calibration error is the mean squared vertical distance of points on the calibration curve from a perfectly straight line. If an expected confidence level occurs $N$ times in the test data, its distance is counted $N$ times in the mean. Following [29], we additionally report the negative log-likelihood (NLL) of the test data averaged across all scenes. Following [24], we include PSNR and LPIPS [32] to evaluate image quality. + +Datasets. We use 30 scenes from 3 datasets: Realistic Synthetic $360^{\circ}$ [16], the subset of scenes in DTU [8] used in [24], and LLFF [15] for training the metacalibrator and test it on a hold-out scene from either LLFF or DTU to show it generalizes to new target scenes. + +Baselines We compare our approach against the state-of-the-art method for NeRF uncertainty estimation DANE [29] as well as other methods in Sec. 4.1. In Tab. 1, following [29], we implement the naive ensembles approach and DANE using a public implementation of Instant-NGP [2,17] and 5 ensemble members. In Sec. 4.2, we compare the uncalibrated uncertainties to the uncertainties calibrated by our meta-calibrator. + +Meta-calibrator Design The results guiding our decisions to use 3 Principal Component Analysis (PCA) components to represent the calibration curves, fit the PCA components using 21 training scenes, and train the meta-calibrator on 30 scenes to predict the PCA coefficients are shown in Fig. 3. + +![](images/3e0ba27d002685edcc41c61c18eb02494e05424fd4462c22a89c8e3c9d4af899.jpg) +(a) Explained variance plot showing most of the variance of the calibration curves is explained using only 3 PCA components. + +![](images/e192eca2cd13d334ea21be4c2534f2750d2e441fd97688027a0d3ee19f0e0413.jpg) +(b) Graph showing calibration error on new scenes is sufficiently low when using 21 scenes to fit 3-coefficient PCA model (i.e. it's an order of magnitude lower than we can expect from the final calibration, meaning good generalization to unseen scenes). + +![](images/cae523db36afed04044cba723c427bb76988a41c52a4efc60108223fe8037d2b.jpg) +(c) Graph showing increasing number of training scenes for meta-calibrator decreases calibration error for the test scene $T$ -Rex from LLFF [15], with 30 scenes resulting in good generalization to the held-out scene and well-calibrated uncertainties. +Fig. 3: Meta-calibrator design decisions. Results showing using 3 components for Principal Component Analysis (PCA) model of calibration curves, 21 scenes to fit PCA model, and 30 scenes to train meta-calibrator achieves good generalization to new test scenes. + +Table 1: Quantitative results on standard sparse NeRF benchmark. Our proposed approach results in significantly better uncertainties and image quality than the state-of-the-art NeRF uncertainty estimation method DANE [29] does on the challenging 3-view LLFF [15] dataset. Specifically, our meta-calibrator reduces the calibration error to $6\%$ of DANE's calibration error and the negative log-likelihood to be over 100 $\%$ lower than DANE's. Note: lower calibration error (Cal. Err.) and negative log-likelihood (NLL) values indicate more accurate uncertainties. Results are averaged over all 8 scenes in LLFF. + +
Uncert aintyImageQuality
Cal. Err. (↓)NLL (↓)PSNR (↑)LPIPS (↓)
Naïve Ens.0.05054.3915.190.646
DANE [29]0.04413.7515.190.646
Ours0.0026-0.6819.340.235
+ +# 4.1 Comparison to State-of-the-art + +In this section, we compare our approach to prior methods for NeRF uncertainty estimation. We achieve more accurate uncertainties (94% reduction in calibration error and over 100% reduction in negative log-likelihood) than those estimated by DANE [29], the state-of-the-art method. These results are shown in Tab. 1 for the challenging 3-view LLFF [15] dataset from prior work on sparse novel view synthesis [19,24,25] and Tab. 2 for the less challenging version of LLFF from prior work on NeRF uncertainty estimation [14,26,27,29]. In Fig. 5a, we compare the calibration curves from our approach to DANE's, illustrating that the meta-calibrator predicts expected confidences that match the true ones while DANE does not. + +![](images/4d7c7acee6cd8b344d416a7a078bafb068672e896c6b653369c89db8153bef5a.jpg) +Fig. 4: Quantitative comparison of uncalibrated and calibrated uncertainties. In (a), we show calibration curves on test data from four scenes in LLFF [15]. The color of each curve indicates the color channel it corresponds to. The calibrated curves are much closer to the ideal calibration (dashed lines), demonstrating that the meta-calibrator works very well. In (b), the average calibration error and negative log-likelihood before and after calibration are reported for LLFF, clearly showing the meta-calibrator improves the accuracy of the uncertainties (lowering calibration error and negative log-likelihood). To test generalization, the meta-calibrator was also applied to held-out scenes in DTU [8], achieving a $70\%$ reduction in calibration error on average. + +# 4.2 Comparison to Uncalibrated Uncertainties + +In this section, we compare our uncalibrated base NeRF uncertainties to our calibrated uncertainties obtained from applying the proposed meta-calibrator. In Fig. 6 we show that the calibrated uncertainties better highlight floaters and other errors in the NeRF renderings. In Fig. 4, we show that the meta-calibrator predicts expected confidence that closely match the true ones, lowering both the calibration error and the negative log-likelihood of the uncalibrated uncertainties. + +Table 2: Quantitative results on standard NeRF uncertainty estimation benchmark. Here, we present results on the LLFF [15] dataset used in prior work [14, 26, 27, 29] on uncertainty estimation for NeRFs. This dataset is less challenging than the one in Tab. 1 since 4-12 views are used for training instead of 3. Our proposed approach results in significantly better uncertainties than prior methods for NeRF uncertainty estimation on all 8 scenes in LLFF. Note: lower negative log-likelihood values indicate more accurate uncertainties. $M$ indicates the number of ensemble members, and MC-DO refers to Monte Carlo Dropout sampling with $M$ sample configurations. This table is Tab. 1 from [29] with our results added as an additional column. Please refer to [29] for further details. + +
Scene# of Train. ViewsNegative Log-likelihood (↓)
MC-DO M=5Naïve Ens. M=5NeRF-W [14]S-NeRF [27]CF-NeRF [26]DANE [29] M=5DANE [29] M=10Ours
Fern44.902.472.162.01—-0.98-1.00-1.41
Orchids55.742.232.241.95—-0.28-0.31-0.84
Leaves52.722.660.790.68—0.970.73-1.19
Flower74.631.631.711.27—1.000.85-2.05
Fortress85.192.291.04-0.03—-1.30-1.30-1.99
Room85.062.134.932.35—-1.35-1.35-2.17
T-Rex114.102.281.911.37—-0.31-0.69-1.49
Horns124.182.170.780.60—-0.55-0.66-2.18
Avg.4.572.231.951.270.57-0.35-0.47-1.67
+ +![](images/2d07211d4bd062480b310d0bdf06858042dc99302597f89c54a366456d60b3c2.jpg) +(a) Calibration curves for the Fern scene from 3-view LLFF [15]. + +![](images/e3e883abb1668e94b629fb55e7d1037e931ef4f52c6da77c20f6d51dbe06af47.jpg) +(b) Next-best View Planning on Horns from LLFF [15]. +Fig. 5: Comparison to DANE [29]. Results comparing our uncertainties to those from the state-of-the-art method DANE. In (a) we show DANE's RGB calibration curves are not closely aligned with the perfectly calibrated lines, meaning it is miscalibrated. It is significantly over-confident for expected confidence levels close to 1 and under-confident for confidence levels close to 0. In comparison, the curves for our approach are extremely close to the ideal calibration (dashed lines), demonstrating that the meta-calibrator works very well, predicting expected confidences that match the true ones. This is also verified by how our calibration error is over two orders of magnitude smaller than DANE's. In (b) we show that our approach results in more efficient performance gains over DANE for next-best view planning. + +![](images/772e986f14d98b9828f9bfaedd02765dd81cdf962ac1fb0f1f0b7934a7d90052.jpg) +(a) Ground Truth +Fig. 6: Qualitative comparison of uncalibrated and calibrated uncertainties from the Flower scene in LLFF [15]. The calibrated uncertainties (d) clearly detect incorrect regions (indicated by the red boxes) better than the uncalibrated uncertainties (c) do. This is apparent by noting that (d) and (e) look more similar than (c) and (e). + +![](images/0396d509886d8986cbbb310048ba4947368c50a9cd34d677c2dc8b942ce16e40.jpg) +(b) NeRF Image + +![](images/81cdbc67aa61107e29a81b8c38c286f8543e19698907e8ecf1aec0fba758f609.jpg) +(c) Uncalibrated Uncertainty + +![](images/ca22ab0ca68246f8540977f906d1c3eb9e1c771bafb219b95e24638c50cc8fae.jpg) +(d) Calibrated Uncertainty + +![](images/0c476a28b0a02923fae760d9da2b1a27895cd36bf62cc1fb6f6d757335777f28.jpg) +(e) Actual Error + +# 4.3 Application: Next-best View Planning + +In this section, we show that our uncertainties can be leveraged for next-best view planning. Specifically, we start by training the NeRF model for 2000 iterations on a training set of three images. The next-best view is selected by evaluating the average calibrated pixel uncertainty (obtained using the meta-calibrator) for each of the candidate views, and the view with the highest uncertainty is added to the training set. The average PSNR of the test images is reported after each training iteration. In Fig. 5b we show that using our approach results in greater performance gains (higher PSNRs) than DANE [29] does. In the Fig. 1 of the Supplementary, we also compare the information gain from rays selected according to the highest calibrated uncertainties to the information gain from rays selected according to the highest uncalibrated uncertainties and show our meta-calibrator produces higher average PSNRs on the test set for scenes in DTU [8] than the uncalibrated uncertainties do. Thus, calibration specifically re-orders the pixel uncertainties so that rays more likely to raise the PSNR are picked earlier in next-best view planning. In Sec. 5 of the Supplementary, we include a detailed theoretical example showing that such re-ordering is possible. + +# 5 Conclusion + +In this paper we addressed the open problem of obtaining calibrated uncertainties from NeRF models. We introduce the concept of a meta-calibrator that infers the calibration curves from scene features, and using this approach achieve state-of-the-art uncertainty without holding out any ground truth data from the target scene. By enabling efficient and accurate calibration of NeRF models without relying on additional data, our method represents a significant step forward in the practical application of NeRF to real-world scenarios and opens up new avenues for the use of NeRF in situations where data is limited and uncertainty is critical. + +Acknowledgements. The authors would like to thank Seunghyeon Seo for his support on FlipNeRF, and Oishi Deb for her insightful feedback on uncertainty estimation. Niki Amini-Naieni is funded by an AWS Studentship, the Reuben Foundation, and the AIMS CDT program at the University of Oxford. Tomas Jakab is sponsored by EPSRC VisualAI EP/T028572/1 and Andrea Vedaldi by ERC-CoG UNION 101001212. + +# References + +1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV. pp. 5835-5844 (2021) +2. Bhalgat, Y.: Hashnerf-pytorch. https://github.com/yashbhalgat/HashNeRF-pytorch/ (2022) +3. Gafni, G., Thies, J., Zollhöfer, M., Nießner, M.: Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In: CVPR. pp. 8649-8658 (2021) +4. Ghoshal, B., Tucker, A.: On calibrated model uncertainty in deep learning. In: ECML (2022) +5. Goli, L., Reading, C., Sellan, S., Jacobson, A., Tagliasacchi, A.: Bayes' Rays: Uncertainty quantification in neural radiance fields. ArXiv abs/2309.03185 (2023) +6. Jain, A., Tancik, M., Abbeel, P.: Putting nef on a diet: Semantically consistent few-shot view synthesis. In: ICCV. pp. 5865-5874 (2021) +7. Jang, T.J., Hyun, C.M.: Nerf solves undersampled mri reconstruction. ArXiv abs/2402.13226 (2024) +8. Jensen, R.R., Dahl, A., Vogiatzis, G., Tola, E., Aanaes, H.: Large scale multi-view stereopsis evaluation. In: CVPR (2014) +9. Jin, L., Chen, X., Ruckin, J., Popovi'c, M.: Neu-nbv: Next best view planning using uncertainty estimation in image-based neural rendering. In: IROS (2023) +0. Kajiya, J.T., Von Herzen, B.P.: Ray tracing volume densities. In: SIGGRAPH. p. 165-174 (1984) +1. Kosiorek, A.R., Strathmann, H., Zoran, D., Moreno, P., Schneider, R., Mokr'a, S., Rezende, D.J.: Nerf-vae: A geometry aware 3d scene generative model. In: ICML (2021) +2. Kuleshov, V., Fenner, N., Ermon, S.: Accurate uncertainties for deep learning using calibrated regression. In: ICML. pp. 2796-2804 (2018) +3. Liu, L., Gu, J., Lin, K.Z., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. In: NIPS (2020) +4. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: CVPR (2021) +5. Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. In: TOG (2019) +6. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +7. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. In: ACM Trans. Graph. (2022) + +18. Neff, T., Stadlbauer, P., Parger, M., Kurz, A., Mueller, J.H., Chaitanya, C.R.A., Kaplanyan, A., Steinberger, M.: Donerf: Towards real-time rendering of compact neural radiance fields using depth oracle networks. In: CGF. pp. 45-59 (2021) +19. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S.M., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: CVPR (2022) +20. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Howes, R., Huang, P.Y., Xu, H., Sharma, V., Li, S.W., Galuba, W., Rabbat, M., Assran, M., Ballas, N., Synnaeve, G., Misra, I., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023) +21. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: ICCV (2021) +22. Ran, Y., Zeng, J., He, S., Li, L., Chen, Y., Lee, G.H., Chen, J., Ye, Q.: Neurar: Neural uncertainty for autonomous 3d reconstruction. In: RAL (2023) +23. Rebain, D., Jiang, W., Yazdani, S., Li, K., Yi, K.M., Tagliasacchi, A.: Derf: Decomposed radiance fields. In: CVPR. pp. 14148-14156 (2020) +24. Seo, S., Chang, Y., Kwak, N.: Flipnerf: Flipped reflection rays for few-shot novel view synthesis. In: ICCV (2023) +25. Seo, S., Han, D., Chang, Y., Kwak, N.: Mixnerf: Modeling a ray with mixture density for novel view synthesis from sparse inputs. In: CVPR. pp. 20659-20668 (2023) +26. Shen, J., Agudo, A., Moreno-Noguer, F., Ruiz, A.: Conditional-flow nerf: Accurate 3d modelling with reliable uncertainty quantification. In: ECCV (2022) +27. Shen, J., Ruiz, A., Agudo, A., Moreno-Noguer, F.: Stochastic neural radiance fields: Quantifying uncertainty in implicit 3d representations. In: 3DV. pp. 972-981 (2021) +28. Sitzmann, V., Martel, J.N., Bergman, A.W., Lindell, D.B., Wetzstein, G.: Implicit neural representations with periodic activation functions. In: NIPS (2020) +29. Sünderhauf, N., Abou-Chakra, J., Miller, D.: Density-aware nef ensembles: Quantifying predictive uncertainty in neural radiance fields. In: ICRA (2023) +30. Wang, P., Liu, Y., Chen, Z., Liu, L., Liu, Z., Komura, T., Theobalt, C., Wang, W.: F2-nerf: Fast neural radiance field training with free camera trajectories. In: CVPR (2023) +31. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. ArXiv abs/2010.07492 (2020) +32. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) \ No newline at end of file diff --git a/2024/Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator/images.zip b/2024/Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e7e032ef9a5cf2247722cbbafbb81eaf2d3c5745 --- /dev/null +++ b/2024/Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7c103e680752e8ed1cb174ba9b6a08732ab393a16a8d451a18f0a0624e6be17 +size 323104 diff --git a/2024/Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator/layout.json b/2024/Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e7e8e3df19f0e31844f1eaa45fa9f3cd886cbf9d --- /dev/null +++ b/2024/Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator/layout.json @@ -0,0 +1,8034 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 481, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 481, + 147 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 481, + 147 + ], + "type": "text", + "content": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 168, + 473, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 168, + 473, + 182 + ], + "spans": [ + { + "bbox": [ + 138, + 168, + 473, + 182 + ], + "type": "text", + "content": "Niki Amini-Naieni1, Tomas Jakab1, Andrea Vedaldi1, and Ronald Clark1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 264, + 190, + 350, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 190, + 350, + 201 + ], + "spans": [ + { + "bbox": [ + 264, + 190, + 350, + 201 + ], + "type": "text", + "content": "University of Oxford" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 235, + 452, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 235, + 452, + 509 + ], + "spans": [ + { + "bbox": [ + 160, + 235, + 452, + 509 + ], + "type": "text", + "content": "Abstract. Neural Radiance Fields (NeRFs) have markedly improved novel view synthesis, but accurate uncertainty quantification in their image predictions remains an open problem. The prevailing methods for estimating uncertainty, including the state-of-the-art Density-aware NeRF Ensembles (DANE) [29], quantify uncertainty without calibration. This frequently leads to over- or under-confidence in image predictions, which can undermine their real-world applications. In this paper, we propose a method which, for the first time, achieves calibrated uncertainties for NeRFs. To accomplish this, we overcome a significant challenge in adapting existing calibration techniques to NeRFs: a need to hold out ground truth images from the target scene, reducing the number of images left to train the NeRF. This issue is particularly problematic in sparse-view settings, where we can operate with as few as three images. To address this, we introduce the concept of a meta-calibrator that performs uncertainty calibration for NeRFs with a single forward pass without the need for holding out any images from the target scene. Our meta-calibrator is a neural network that takes as input the NeRF images and uncalibrated uncertainty maps and outputs a scene-specific calibration curve that corrects the NeRF's uncalibrated uncertainties. We show that the meta-calibrator can generalize on unseen scenes and achieves well-calibrated and state-of-the-art uncertainty for NeRFs, significantly beating DANE and other approaches. This opens opportunities to improve applications that rely on accurate NeRF uncertainty estimates such as next-best view planning and potentially more trustworthy image reconstruction for medical diagnosis." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 521, + 427, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 521, + 427, + 533 + ], + "spans": [ + { + "bbox": [ + 160, + 521, + 427, + 533 + ], + "type": "text", + "content": "Keywords: NeRFs " + }, + { + "bbox": [ + 160, + 521, + 427, + 533 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 521, + 427, + 533 + ], + "type": "text", + "content": " Uncertainty calibration " + }, + { + "bbox": [ + 160, + 521, + 427, + 533 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 521, + 427, + 533 + ], + "type": "text", + "content": " Few-shot learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 554, + 230, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 230, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 230, + 567 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Recent advancements in scene representations have led to promising new approaches for novel view synthesis and scene reconstruction. Among these, Neural Radiance Fields (NeRFs) [16] have emerged as a particularly powerful tool, offering unprecedented levels of realism and detail in rendered images. The core idea behind NeRFs is to represent a scene as a continuous vector-valued function, parameterized by a neural network, that maps spatial coordinates and view directions to color and density values. This approach enables the creation" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 184, + 113, + 430, + 255 + ], + "blocks": [ + { + "bbox": [ + 184, + 113, + 430, + 255 + ], + "lines": [ + { + "bbox": [ + 184, + 113, + 430, + 255 + ], + "spans": [ + { + "bbox": [ + 184, + 113, + 430, + 255 + ], + "type": "image", + "image_path": "931aa5f9312feb8cc27913728d089273215c2cacdf110ac6aad8da7b08aa1d22.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 264, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 130, + 264, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 482, + 331 + ], + "type": "text", + "content": "Fig. 1: We propose a method for efficiently calibrating the uncertainties from NeRF models. Our approach is based on a meta-calibrator that takes as input features from the rendered NeRF images and uncalibrated uncertainty maps and predicts the calibration function, " + }, + { + "bbox": [ + 130, + 264, + 482, + 331 + ], + "type": "inline_equation", + "content": "R_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 130, + 264, + 482, + 331 + ], + "type": "text", + "content": ", for the NeRF model. Our meta-calibrator generalizes across scenes, so it only needs to be trained once, and can predict the calibration function in a single forward pass without any ground truth data from the target scene." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 359, + 482, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 482, + 383 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 482, + 383 + ], + "type": "text", + "content": "of detailed 3D representations from sets of 2D images, revolutionizing 3D reconstruction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 506 + ], + "type": "text", + "content": "However, despite their impressive capabilities, traditional NeRF models lack an essential component: an accurate measure of uncertainty in their predictions. Accurate uncertainties are crucial for applying NeRFs to safety-critical problems such as MRI image reconstruction from sparse data [7], where unreliable confidence estimates could lead to misdiagnosis. More accurate uncertainties could also enhance practical methods such as uncertainty-guided next-best view planning techniques [9]. Prior approaches have attempted to estimate NeRF uncertainties [5, 14, 26, 27, 29], but they all overlook the problem of calibration. Thus, the uncertainties they output are not as accurate as they could be." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 507, + 482, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 580 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 580 + ], + "type": "text", + "content": "In particular, the state-of-the-art uncertainty estimation method for NeRFs, Density-aware NeRF Ensembles (DANE) [29], produces uncalibrated uncertainties. As a result, the confidence intervals and variances do not match the true confidence, meaning it has limited applicability to real-world problems. This constraint is significant as it restricts the use of NeRFs in safety-critical and sparse-data settings, where knowing the confidence in predictions is crucial." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 654 + ], + "type": "text", + "content": "The best NeRF methods for the sparse-view setting overlook the problem of calibration as well. FlipNeRF [24], the state-of-the-art technique for sparse-view reconstruction, uses uncertainties from an uncalibrated mixture of Laplacians to enhance its training process. Therefore, the uncertainties it outputs at inference as an artifact of training are not accurate." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "type": "text", + "content": "N. Amini-Naieni et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 485, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 485, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 485, + 308 + ], + "type": "text", + "content": "In this paper, we present a novel approach for obtaining calibrated uncertainties for NeRF models in the sparse-view setting. Our strategy integrates the Laplacian mixture from FlipNeRF [24] with the calibration techniques by Kuleshov et al [12]. However, naively applying the calibration method by Kuleshov et al to FlipNeRF does not work due to a significant challenge of the sparse-view setting: there is a lack of held-out data from the target scene for fitting the calibrator. Specifically, holding out just one image for calibration could decrease the size of the training set by over " + }, + { + "bbox": [ + 130, + 116, + 485, + 308 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 116, + 485, + 308 + ], + "type": "text", + "content": " , resulting in significant performance degradation of the NeRF. To overcome this, we make use of a unique observation: while calibration curves exhibit significant variation across scenes, they also demonstrate a significant regularity in their structure. Utilizing this insight, we propose the concept of a meta-calibrator that learns a low-dimensional representation of the NeRF calibration curves and infers them from scene features. We motivate and show why this meta-calibrator is necessary and demonstrate that it achieves more accurate uncertainties than DANE [29] without holding out any images from the target scene." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 320, + 483, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 483, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 483, + 405 + ], + "type": "text", + "content": "Specifically, our contributions are: (1) the first investigation into obtaining calibrated uncertainties from NeRFs, (2) a novel meta-calibrator for fitting the calibration model without using held-out data, and (3) experiments on the real-world Local Light Field Fusion (LLFF) [15] and DTU [8] datasets showing that our meta-calibrator achieves state-of-the-art and well-calibrated uncertainties for real scenes. We also demonstrate that our uncertainties can be leveraged for effective next-best view planning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 430, + 237, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 237, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 237, + 441 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 460, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 557 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRFs). NeRFs [16] are a popular method for novel view synthesis. From a set of 2D images, NeRFs learn a neural network representation of a single scene. A trained NeRF model outputs estimates of the volume density and emitted radiance at any 3D location and viewing direction. Novel views can be generated by applying volume rendering [10] to the density and radiance values predicted by the NeRF model for points along rays cast into the scene. Due to their simplicity and impressive performance, NeRFs have become a popular technique for solving a variety of rendering problems." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Over the last few years, several extensions of NeRFs have been explored [11, 18, 21, 28, 31]. These include speeding up training and inference [13, 23, 30], modeling dynamic scenes [3], learning from a sparse set of training views [1, 6, 19, 24, 25], and estimating the uncertainty in NeRF predictions [5, 14, 22, 26, 27, 29]. Sparse NeRF methods aim to accurately render novel views when only a few training views are available from the target scene. NeRF uncertainty estimation techniques strive to accurately predict the confidence in the views rendered. Although uncertainty estimation is particularly important in the sparse-view" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "setting, where NeRF renderings are especially unreliable, the main aim of sparse NeRF methods is not to output accurate uncertainties." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 152, + 481, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 152, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 152, + 481, + 236 + ], + "type": "text", + "content": "Sparse-view NeRF Methods. Despite this, recent sparse NeRF methods do produce uncertainties as an artifact of their training process. Both MixNeRF [25] and FlipNeRF [24], the state-of-the-art approach, model the RGB color channels given a ray as independent random variables that follow a mixture model. FlipNeRF further uses the uncertainties of the pixel colors to regularize the training process, producing more accurate image reconstructions at inference. However, neither MixNeRF nor FlipNeRF outputs calibrated uncertainties." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 247, + 481, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 481, + 367 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 481, + 367 + ], + "type": "text", + "content": "Our method significantly extends sparse NeRF methods to obtain more accurate and well-calibrated uncertainties at inference. To benefit from the superior performance of FlipNeRF [24] at sparse novel view synthesis, we apply the proposed meta-calibrator to the learned distribution from FlipNeRF, producing more accurate uncertainties without sacrificing state-of-the-art image quality. However, our approach can be applied to any NeRF method that outputs uncertainties, so it is distinct from FlipNeRF and MixNeRF. In essence, the proposed meta-calibrator augments sparse NeRF methods to achieve state-of-the-art uncertainty, beating both FlipNeRF and techniques designed explicitly for NeRF uncertainty estimation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 378, + 481, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 378, + 481, + 618 + ], + "spans": [ + { + "bbox": [ + 132, + 378, + 481, + 618 + ], + "type": "text", + "content": "Uncertainty in NeRFs. The growing line of methods specifically designed for accurately estimating NeRF uncertainties [5,14,22,26,27,29] do not address the problem of calibration. The current state-of-the-art uncertainty estimation technique for NeRFs is Density-aware NeRF Ensembles (DANE) [29]. DANE adds an epistemic uncertainty term to a naive ensembles approach with five ensemble members. Thus, DANE is very costly as it requires training five NeRFs to obtain uncertainty estimates. Another NeRF uncertainty estimation approach is Stochastic Neural Radiance Fields (S-NeRF) [27], which learns a probability distribution over all possible radiance fields by modeling the volume density and radiance as random variables that follow a joint distribution. S-NeRF employs variational inference to sample from an approximation to this distribution and uses the variances of the sampled pixel colors as the estimated uncertainties. Conditional-flow NeRF (CF-NeRF) [26] builds on S-NeRF by combining latent variable modeling and conditional normalizing flows to relax the strong constraints S-NeRF imposes over the radiance distribution. Despite the growing number of techniques in this area of study, all prior work does not consider calibration, outputting unreliable uncertainties as a result. Our work achieves more accurate uncertainties than these prior methods by filling the gap of uncertainty calibration for NeRFs and drawing on well-established techniques in calibrated regression [12]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "content": "Uncertainty Calibration in Deep Learning. While uncertainty calibration has been studied for Bayesian deep learning methods [4, 12], it has not been adapted for or applied successfully to NeRF uncertainty estimates. This may be" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "type": "text", + "content": "N. Amini-Naieni et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "because NeRFs introduce additional complexity in the uncertainty estimation process as the neural network model needs to be trained per-scene." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 247 + ], + "type": "text", + "content": "This makes uncertainty calibration challenging as methods for calibrated regression [12] require either using the training set or held-out data to achieve calibrated uncertainties. Using the training set to fit the calibrator results in severe overfitting (see Sec. 2 of the Supplementary). Holding out data from the target scene for calibration means there is less data to train the NeRF, making it more inaccurate at novel view synthesis (see Sec. 3 of the Supplementary). Thus, a trivial application of [12] to NeRFs would not be satisfactory. In this work, we propose the concept of a meta-calibrator that, in contrast to [12], does not require held-out data and achieves calibrated uncertainty estimates for NeRFs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 278, + 202, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 202, + 291 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 202, + 291 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 304, + 482, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 423 + ], + "type": "text", + "content": "In this paper, we present a method that calibrates NeRF uncertainties. To this end, we propose a novel meta-calibrator that accepts uncalibrated NeRF uncertainties and predicted images as inputs and outputs a scene-specific calibration curve, correcting the uncalibrated confidence levels. An overview of our method can be seen in Fig. 1. Crucially, our approach does not require holding out any images from the target scene. Thus, it can be applied to sparse-view settings, where holding out a single image could significantly harm the NeRF's performance. In Sec. 3.1, we explain the necessary background concepts, in Sec. 3.2, we describe how we obtain the corrected uncertainty values, and in Sec. 3.3 we detail our meta-calibrator." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 441, + 228, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 441, + 228, + 453 + ], + "spans": [ + { + "bbox": [ + 131, + 441, + 228, + 453 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRFs). Neural Radiance Fields (NeRFs) [16] represent a scene as a continuous vector-valued function with inputs a Cartesian point " + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = (x,y,z)" + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "text", + "content": " and unit viewing direction vector " + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{d} = (u,v,w)" + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "text", + "content": " and outputs an emitted radiance " + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\pmb {c} = (\\nu ,\\varrho ,\\theta)" + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "text", + "content": " and volume density " + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "text", + "content": " . By optimizing the weights " + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "text", + "content": " of a neural network approximation " + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\Theta}" + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "text", + "content": " to this representation, NeRFs can render the color of any pixel in a synthetic image of the scene. To achieve this, principles from classical volume rendering [10] are applied to the radiance and density values estimated by " + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\Theta}" + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "text", + "content": " for points along a ray cast from the origin " + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{x_0}" + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "text", + "content": " of the virtual camera, through the pixel, and into the scene. More specifically, the expected color " + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{c}(\\mathbf{r})" + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "text", + "content": " of a camera ray " + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(t) = \\mathbf{x_0} + t\\mathbf{d}" + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "text", + "content": " with near and far bounds " + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "inline_equation", + "content": "t_f" + }, + { + "bbox": [ + 130, + 461, + 482, + 594 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 230, + 594, + 481, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 594, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 230, + 594, + 481, + 622 + ], + "type": "interline_equation", + "content": "\\mathbf {c} (\\mathbf {r}) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (\\mathbf {r} (t)) \\boldsymbol {c} (\\mathbf {r} (t), \\mathbf {d}) d t, \\tag {1}", + "image_path": "4b606fc08e4baa0370572b8091f3009ee026cd62e965aa1e970318c9857732a3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 627, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 627, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 627, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 627, + 482, + 666 + ], + "type": "inline_equation", + "content": "T(t) = e^{-\\int_{t_n}^t \\sigma(\\mathbf{r}(s)) ds}" + }, + { + "bbox": [ + 130, + 627, + 482, + 666 + ], + "type": "text", + "content": ". The integral in Eq. (1) is estimated with numerical quadrature to obtain the colors of the pixels in the synthetic image from the NeRF model outputs. NeRF optimizes " + }, + { + "bbox": [ + 130, + 627, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 627, + 482, + 666 + ], + "type": "text", + "content": " according to Eq. (1) with gradient" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 479, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 479, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 479, + 99 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "descent, since the numerical quadrature is differentiable. Conventional NeRFs do not provide an uncertainty associated with their predictions, so extensions like DANE [29] have been developed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 159, + 480, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 159, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 159, + 480, + 268 + ], + "type": "text", + "content": "Base NeRF Uncertainties. To obtain the initial uncertainties, we have two options for our base model: (1) DANE [29], the state-of-the-art method for NeRF uncertainty estimation or (2) FlipNeRF [24], the state-of-the-art method for sparse novel view synthesis. DANE requires training five NeRFs per scene and provides poor image quality in the sparse-view setting, so we choose FlipNeRF. We show in the experiments that applying our meta-calibrator to FlipNeRF results in more accurate and well-calibrated uncertainties than those output by DANE. Our base FlipNeRF uncertainties are inferred from a mixture of Laplacians with location and scale parameters learned during training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 279, + 479, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 479, + 304 + ], + "type": "text", + "content": "FlipNeRF [24] models the joint distribution of the color " + }, + { + "bbox": [ + 130, + 279, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{C} = (R,G,B)" + }, + { + "bbox": [ + 130, + 279, + 479, + 304 + ], + "type": "text", + "content": " given a ray " + }, + { + "bbox": [ + 130, + 279, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 279, + 479, + 304 + ], + "type": "text", + "content": " with a mixture of Laplacians:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 224, + 314, + 480, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 314, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 224, + 314, + 480, + 346 + ], + "type": "interline_equation", + "content": "p (\\mathbf {C} = \\mathbf {c} | \\mathbf {r}) = \\sum_ {j = 1} ^ {M} \\pi_ {j} \\mathcal {F} (\\mathbf {C} = \\mathbf {c}; \\boldsymbol {\\mu} _ {j}, \\boldsymbol {\\beta} _ {j}), \\tag {2}", + "image_path": "18f8e290799b3cfd10b66853470f47b2bd068b7ac5654f98ef69abf0cf2be268.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 349, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 480, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 349, + 480, + 437 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 349, + 480, + 437 + ], + "type": "text", + "content": " is the number of sampled points along the ray " + }, + { + "bbox": [ + 130, + 349, + 480, + 437 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 349, + 480, + 437 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 349, + 480, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{C} = \\mathbf{c};\\pmb{\\mu}_j,\\pmb{\\beta}_j)" + }, + { + "bbox": [ + 130, + 349, + 480, + 437 + ], + "type": "text", + "content": " is the 3D Laplacian probability density with location parameter " + }, + { + "bbox": [ + 130, + 349, + 480, + 437 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{j} = (\\mu_{j}^{R},\\mu_{j}^{G},\\mu_{j}^{B})" + }, + { + "bbox": [ + 130, + 349, + 480, + 437 + ], + "type": "text", + "content": " and scale parameter " + }, + { + "bbox": [ + 130, + 349, + 480, + 437 + ], + "type": "inline_equation", + "content": "\\beta_{i} = (\\beta_{j}^{R},\\beta_{j}^{G},\\beta_{j}^{B})" + }, + { + "bbox": [ + 130, + 349, + 480, + 437 + ], + "type": "text", + "content": " evaluated at the color " + }, + { + "bbox": [ + 130, + 349, + 480, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 349, + 480, + 437 + ], + "type": "text", + "content": ". More specifically, the mixing coefficients " + }, + { + "bbox": [ + 130, + 349, + 480, + 437 + ], + "type": "inline_equation", + "content": "\\{\\pi_j\\}_{j = 1}^M" + }, + { + "bbox": [ + 130, + 349, + 480, + 437 + ], + "type": "text", + "content": " are the normalized coefficients of the radiance values along a ray in Eq. (2), the location parameters " + }, + { + "bbox": [ + 130, + 349, + 480, + 437 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\mu}_{j}\\}_{j = 1}^{M}" + }, + { + "bbox": [ + 130, + 349, + 480, + 437 + ], + "type": "text", + "content": " are the estimated RGB radiance values, and the scale parameters " + }, + { + "bbox": [ + 130, + 349, + 480, + 437 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\beta}_{j}\\}_{j = 1}^{M}" + }, + { + "bbox": [ + 130, + 349, + 480, + 437 + ], + "type": "text", + "content": " are an additional output of the model. These parameters are optimized by FlipNeRF during training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 450, + 480, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 480, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 480, + 594 + ], + "type": "text", + "content": "We now go into detail about how the location and scale parameters are obtained. FlipNeRF [24] learns these parameters by minimizing the negative log-likelihood of the training set " + }, + { + "bbox": [ + 130, + 450, + 480, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(\\mathbf{r_i},\\mathbf{c_i})\\}_{i = 1}^N" + }, + { + "bbox": [ + 130, + 450, + 480, + 594 + ], + "type": "text", + "content": " containing the rays " + }, + { + "bbox": [ + 130, + 450, + 480, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{r_i}" + }, + { + "bbox": [ + 130, + 450, + 480, + 594 + ], + "type": "text", + "content": " and colors " + }, + { + "bbox": [ + 130, + 450, + 480, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{c_i}" + }, + { + "bbox": [ + 130, + 450, + 480, + 594 + ], + "type": "text", + "content": " from the pixels in the ground truth images of the scene assuming the distribution in Eq. (2). FlipNeRF additionally minimizes the average scale parameters through an auxiliary uncertainty-aware emptiness loss for reducing floating artifacts. The negative log-likelihood loss and the uncertainty-aware emptiness loss are added to FlipNeRF's total training loss, which incorporates other terms such as the mean squared error. As a result of this training process, the location and scale parameters can be inferred by FlipNeRF at new poses. We obtain the uncalibrated confidence levels of predicted ray colors from these location and scale parameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": "Using the distribution in Eq. (2), we can easily compute the confidence level for a given ray " + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{r_t}" + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": ", which we denote as " + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": ". The CDF for the Laplacian mixture has a closed form expression parameterized by the location and scales output by the pretrained FlipNeRF [24]. Thus, we can use this CDF to predict the confidence level of the ground truth color " + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{c_t}" + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": " of any ray by evaluating it at the given" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "type": "text", + "content": "N. Amini-Naieni et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "color value, " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "p_t = F_t(\\mathbf{c_t})" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ". However, these initial confidence levels are uncalibrated and, hence, inaccurate." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 480, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 480, + 232 + ], + "type": "text", + "content": "Calibrated regression. In [12], Kuleshov et al extend calibration methods for classification to regression. They define a forecaster " + }, + { + "bbox": [ + 130, + 161, + 480, + 232 + ], + "type": "inline_equation", + "content": "H: \\mathcal{X} \\to (\\mathcal{Y} \\to [0,1])" + }, + { + "bbox": [ + 130, + 161, + 480, + 232 + ], + "type": "text", + "content": " as a function that outputs for each " + }, + { + "bbox": [ + 130, + 161, + 480, + 232 + ], + "type": "inline_equation", + "content": "x_{t} \\in \\mathcal{X}" + }, + { + "bbox": [ + 130, + 161, + 480, + 232 + ], + "type": "text", + "content": ", a CDF " + }, + { + "bbox": [ + 130, + 161, + 480, + 232 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 130, + 161, + 480, + 232 + ], + "type": "text", + "content": ". Given a pretrained forecaster " + }, + { + "bbox": [ + 130, + 161, + 480, + 232 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 161, + 480, + 232 + ], + "type": "text", + "content": ", they suggest training an auxiliary model " + }, + { + "bbox": [ + 130, + 161, + 480, + 232 + ], + "type": "inline_equation", + "content": "R: [0,1] \\to [0,1]" + }, + { + "bbox": [ + 130, + 161, + 480, + 232 + ], + "type": "text", + "content": " by fitting " + }, + { + "bbox": [ + 130, + 161, + 480, + 232 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 161, + 480, + 232 + ], + "type": "text", + "content": " to a recalibration dataset " + }, + { + "bbox": [ + 130, + 161, + 480, + 232 + ], + "type": "inline_equation", + "content": "D = \\left\\{\\left([H(x_{t})](y_{t}), \\hat{P}([H(x_{t})](y_{t}))\\right)\\right\\}_{t=1}^{T}" + }, + { + "bbox": [ + 130, + 161, + 480, + 232 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 200, + 239, + 416, + 252 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 239, + 416, + 252 + ], + "spans": [ + { + "bbox": [ + 200, + 239, + 416, + 252 + ], + "type": "interline_equation", + "content": "\\hat {P} (p) = \\left| \\left\\{y _ {t}: [ H \\left(x _ {t}\\right) ] \\left(y _ {t}\\right) \\leq p \\text {f o r} t = 1, \\dots , T \\right\\} \\right| / T", + "image_path": "470b1d998155827c6606ba7adf4d02dfc25275dee04ac6f2e4d7d2d8f0867d37.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 258, + 479, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 479, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 479, + 307 + ], + "type": "text", + "content": "is the empirical confidence level corresponding to the predicted confidence level " + }, + { + "bbox": [ + 130, + 258, + 479, + 307 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 258, + 479, + 307 + ], + "type": "text", + "content": ". The fitted " + }, + { + "bbox": [ + 130, + 258, + 479, + 307 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 258, + 479, + 307 + ], + "type": "text", + "content": " forms the calibration curve that corrects the expected confidence levels. We can now obtain predictive posterior values that closely match the true confidences using " + }, + { + "bbox": [ + 130, + 258, + 479, + 307 + ], + "type": "inline_equation", + "content": "\\hat{F}_t \\equiv R \\circ F_t" + }, + { + "bbox": [ + 130, + 258, + 479, + 307 + ], + "type": "text", + "content": " for the test data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 318, + 480, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 480, + 425 + ], + "type": "text", + "content": "However, as mentioned before, there are numerous challenges associated with applying this recalibration procedure. Firstly, note that it requires ground-truth values " + }, + { + "bbox": [ + 130, + 318, + 480, + 425 + ], + "type": "inline_equation", + "content": "(y_{t})" + }, + { + "bbox": [ + 130, + 318, + 480, + 425 + ], + "type": "text", + "content": " for the predictions we are recalibrating. This means that in order to prevent overfitting we need to reserve part of the training dataset specifically for fitting the calibrator (which leaves less data for training the NeRF — a significant issue if we only have a few input views). Secondly, it does not actually prescribe how to compute a suitable uncertainty value from the predicted distribution. In Sec. 3.3, we address the former issue, and in Sec. 3.2, we address the latter." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 453, + 279, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 453, + 279, + 466 + ], + "spans": [ + { + "bbox": [ + 131, + 453, + 279, + 466 + ], + "type": "text", + "content": "3.2 Calculating uncertainty" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 471, + 480, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 480, + 578 + ], + "type": "text", + "content": "The predictive posterior in Eq. (2) provides a distribution over likely ray colors for a NeRF model, but it does not inherently offer a straightforward metric for quantifying uncertainty at a specific point in the reconstruction. Intuitively, as the variance of this distribution increases, so does the uncertainty of the model's output at that point. Therefore, the variance or standard deviation is a popular choice for quantifying the uncertainty [24,26,27,29]. However, in the case of the corrected mixture distribution obtained from raymarching, this can be slow to compute especially if it has to be done for each pixel. Therefore, we turn to a metric that can be calculated directly from the calibrated CDF " + }, + { + "bbox": [ + 130, + 471, + 480, + 578 + ], + "type": "inline_equation", + "content": "\\hat{F}^t" + }, + { + "bbox": [ + 130, + 471, + 480, + 578 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 147, + 590, + 469, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 590, + 469, + 604 + ], + "spans": [ + { + "bbox": [ + 147, + 590, + 469, + 604 + ], + "type": "text", + "content": "We propose to use the interquartile range of each calibrated distribution:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 223, + 609, + 480, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 609, + 480, + 635 + ], + "spans": [ + { + "bbox": [ + 223, + 609, + 480, + 635 + ], + "type": "interline_equation", + "content": "\\kappa^ {C} \\left(\\mathbf {r} _ {\\mathbf {t}}\\right) = \\left[ \\hat {F} _ {t} ^ {C} \\right] ^ {- 1} \\left(\\frac {3}{4}\\right) - \\left[ \\hat {F} _ {t} ^ {C} \\right] ^ {- 1} \\left(\\frac {1}{4}\\right), \\tag {3}", + "image_path": "f79c793afd3d47daa1b6e45944c6c26c658117b2b81b96d9fda33c1a8524e431.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 640, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 640, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 640, + 479, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 640, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\kappa^C (\\mathbf{r_t})" + }, + { + "bbox": [ + 131, + 640, + 479, + 665 + ], + "type": "text", + "content": " represents the uncertainty at a ray " + }, + { + "bbox": [ + 131, + 640, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{r_t}" + }, + { + "bbox": [ + 131, + 640, + 479, + 665 + ], + "type": "text", + "content": " in the color channel " + }, + { + "bbox": [ + 131, + 640, + 479, + 665 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 131, + 640, + 479, + 665 + ], + "type": "text", + "content": ". This difference provides a measure of the statistical dispersion and thus serves" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 101 + ], + "type": "text", + "content": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 115, + 301, + 205 + ], + "blocks": [ + { + "bbox": [ + 143, + 115, + 301, + 205 + ], + "lines": [ + { + "bbox": [ + 143, + 115, + 301, + 205 + ], + "spans": [ + { + "bbox": [ + 143, + 115, + 301, + 205 + ], + "type": "image", + "image_path": "d4266a2b88d6ec9833101b28ab0e4ae1b897b32bf656aaae7ef21df07d7cbd39.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 207, + 299, + 224 + ], + "lines": [ + { + "bbox": [ + 143, + 207, + 299, + 224 + ], + "spans": [ + { + "bbox": [ + 143, + 207, + 299, + 224 + ], + "type": "text", + "content": "(a) Creating a parametric model of the calibration curves." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 145, + 471, + 215 + ], + "blocks": [ + { + "bbox": [ + 312, + 145, + 471, + 215 + ], + "lines": [ + { + "bbox": [ + 312, + 145, + 471, + 215 + ], + "spans": [ + { + "bbox": [ + 312, + 145, + 471, + 215 + ], + "type": "image", + "image_path": "fc20e2f750952699e6d84c4ea763729a17e3d80e0cab4ff95679774442b6f859.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 216, + 456, + 225 + ], + "lines": [ + { + "bbox": [ + 326, + 216, + 456, + 225 + ], + "spans": [ + { + "bbox": [ + 326, + 216, + 456, + 225 + ], + "type": "text", + "content": "(b) Predicting the curve parameters." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 234, + 482, + 280 + ], + "lines": [ + { + "bbox": [ + 130, + 234, + 482, + 280 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 482, + 280 + ], + "type": "text", + "content": "Fig. 2: Meta-calibrator design. In stage (a) we fit a low-dimensional parametric model of the calibration curves. The meta-calibrator then predicts these curve parameters from rendered images of the scene and their associated uncalibrated uncertainty maps (b)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 361 + ], + "type": "text", + "content": "as a robust measure of the spread of the output channel. By averaging the interquartile range over the color channels, we obtain a single scalar value that effectively quantifies the uncertainty of the NeRF model for the given ray. As shown in our experiments, this method is very computational efficient, and it provides an accurate estimate of uncertainty." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 375, + 240, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 375, + 240, + 386 + ], + "spans": [ + { + "bbox": [ + 131, + 375, + 240, + 386 + ], + "type": "text", + "content": "3.3 Meta-calibrator" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 392, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 501 + ], + "type": "text", + "content": "To overcome the challenge that, especially in the sparse-view setting, there is no held-out data available for fitting the calibrator, we propose a novel metacalibrator that infers the calibration curves from uncalibrated NeRF predictions. To do this, we leverage the insight that the calibration curves demonstrate significant regularity. We posit a low-dimensional model of the calibration curves can be learned and predicted using the images and uncalibrated uncertainty maps inferred by the NeRF, enabling us to estimate the calibration function without evaluating the empirical confidence levels using held-out data from the target scene. We now describe this meta-calibrator (illustrated in Fig. 2) in detail." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "text", + "content": "A Parametric Model for Calibration Curves. We first fit a low-dimensional representation of the calibration curves using Principal Component Analysis (PCA). To create the training set for learning this representation, we sample held-out images from " + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "text", + "content": " scenes and apply the calibration procedure by Kuleshov et al. [12] to form " + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "text", + "content": " ground truth calibration curves. To construct the training vector " + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{v_k} \\in \\mathbb{R}^{1 \\times M}" + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "text", + "content": " for scene " + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "inline_equation", + "content": "k \\leq K" + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "text", + "content": ", we sample " + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "text", + "content": " evenly spaced points along its ground truth calibration curve. We find that fitting the PCA model using only a few scenes (21 in our case) provides a good enough approximation to capture the variation in the test curves (see Sec. 4). Here, " + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = [\\mathbf{v}_k] \\in \\mathbb{R}^{K \\times M}" + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "text", + "content": " contains the ground truth calibration curves for the training scenes, with " + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "text", + "content": " representing the number of curves and " + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "text", + "content": " the sample count along each curve. PCA is then used to determine the basis vectors " + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{U} = (\\mathbf{u_1}, \\mathbf{u_2}, \\dots, \\mathbf{u_n})" + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "text", + "content": " and coefficients " + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\theta = (\\alpha_1, \\alpha_2, \\dots, \\alpha_n)" + }, + { + "bbox": [ + 130, + 510, + 482, + 667 + ], + "type": "text", + "content": ", so that each calibration curve can be represented as:" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "type": "text", + "content": "N. Amini-Naieni et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 461, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 461, + 129 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 461, + 129 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_k = \\sum_{i=1}^n \\alpha_i \\mathbf{u}_i" + }, + { + "bbox": [ + 130, + 115, + 461, + 129 + ], + "type": "text", + "content": ". The parameters " + }, + { + "bbox": [ + 130, + 115, + 461, + 129 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 115, + 461, + 129 + ], + "type": "text", + "content": " fully describe the calibration functions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": "To find the optimal number of components, we compute the explained variance, and find that in our case most of the variance is explained using only " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "n = 3" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " components (see experiments). To ensure the calibrator is monotonically increasing, we derive the final calibration function " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "R_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " using isotonic regression applied to the curve approximated by " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta} \\cdot \\mathbf{U}" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": ". The idea is that the low-dimensional representation of the calibration curves encoded in " + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": " will generalise to new target scenes without any additional scene-specific data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 416 + ], + "type": "text", + "content": "Predicting Calibration Parameters. What remains is to estimate the calibration parameters, " + }, + { + "bbox": [ + 130, + 236, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 236, + 482, + 416 + ], + "type": "text", + "content": ", for a new scene. As we do not want to use additional held-out data from the target scene, we propose predicting these parameters using scene-specific features computed from the pretrained NeRF outputs. This approach is motivated by the human ability to visually identify inaccuracies in the renderings such as floaters and unnatural artifacts. Specifically, we use a Multi-Layer Perceptron (MLP) with three layers of output size: [128, 128, 3] and Leaky ReLU activations throughout except the last layer as the meta-calibrator to estimate " + }, + { + "bbox": [ + 130, + 236, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 236, + 482, + 416 + ], + "type": "text", + "content": " given features extracted by the DINOv2 model [20] from rendered images " + }, + { + "bbox": [ + 130, + 236, + 482, + 416 + ], + "type": "inline_equation", + "content": "(\\mathbf{f}_I)" + }, + { + "bbox": [ + 130, + 236, + 482, + 416 + ], + "type": "text", + "content": " and uncalibrated uncertainty maps " + }, + { + "bbox": [ + 130, + 236, + 482, + 416 + ], + "type": "inline_equation", + "content": "(\\mathbf{f}_{\\kappa})" + }, + { + "bbox": [ + 130, + 236, + 482, + 416 + ], + "type": "text", + "content": ". The goal here is to have DINOv2 extract features that describe the rendering imperfections, correlating with the calibration curve. We find that training the MLP model on only a few scenes (30 in our case) allows it to generalize well to new test scenes. Once trained, the meta-calibrator can predict the calibration curve of a new target scene as: " + }, + { + "bbox": [ + 130, + 236, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\theta = MLP([\\mathbf{f}_I,\\mathbf{f}_{\\kappa}])" + }, + { + "bbox": [ + 130, + 236, + 482, + 416 + ], + "type": "text", + "content": ", without using any additional ground truth data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 428, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 428, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 428, + 482, + 477 + ], + "type": "text", + "content": "In summary, the meta-calibrator can correct the confidence levels of the model without requiring ground truth data at any stage, suggesting potential enhancements to applications that rely on uncertainty such as next-best view selection (see Sec. 4.3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 495, + 230, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 495, + 230, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 495, + 230, + 510 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": "The objective of the experiments is to: 1) validate that our approach achieves more accurate uncertainties (lower negative log-likelihood and calibration error) than state-of-the-art approaches for NeRF uncertainty estimation (Sec. 4.1); 2) demonstrate the meta-calibrator improves the accuracy of the uncalibrated uncertainties (decreases both the negative log-likelihood and calibration error) without requiring any held-out data from the target scene (Sec. 4.2); 3) explain the motivation for certain meta-calibrator design decisions; and 4) show that our uncertainties can be leveraged for applications such as next-best view planning (Sec. 4.3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "For additional results showing: 1) why the PCA representation of the calibration curves is necessary; 2) that using the training set results in severe overfitting; 3) that holding out data results in poor performance at image reconstruction; 4)" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "the influence of the number of samples along the ray on the uncertainty quality; and 5) the efficiency of our uncertainty metric (Eq. (3)) over other approaches, please refer to Sections 1, 2, 3, 4, and 6 of the Supplementary respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 480, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 480, + 198 + ], + "type": "text", + "content": "Metrics and calibration curves. We use a variant of the calibration error from [12] to evaluate the effectiveness of the meta-calibrator. Specifically, given a test set " + }, + { + "bbox": [ + 130, + 161, + 480, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(\\mathbf{r_t},\\mathbf{c_t})\\}_{t = 1}^T = \\{(\\mathbf{r_t},(r_t,g_t,b_t))\\}_{t = 1}^T" + }, + { + "bbox": [ + 130, + 161, + 480, + 198 + ], + "type": "text", + "content": ", we report:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 244, + 206, + 480, + 237 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 206, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 244, + 206, + 480, + 237 + ], + "type": "interline_equation", + "content": "E R R = \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\left(p _ {t} - \\hat {P} \\left(p _ {t}\\right)\\right) ^ {2}, \\tag {4}", + "image_path": "26c92bc2d6b3c4e6482ccd2abcc05c7b8a00660ae6b22b46d07532a5a2cefeb2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 243, + 479, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 479, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 243, + 479, + 281 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 130, + 243, + 479, + 281 + ], + "type": "text", + "content": " is the expected confidence level for data point " + }, + { + "bbox": [ + 130, + 243, + 479, + 281 + ], + "type": "inline_equation", + "content": "(\\mathbf{r_t}, c_t)" + }, + { + "bbox": [ + 130, + 243, + 479, + 281 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 243, + 479, + 281 + ], + "type": "inline_equation", + "content": "\\hat{P}(p_t)" + }, + { + "bbox": [ + 130, + 243, + 479, + 281 + ], + "type": "text", + "content": " is the empirical frequency of data points within that confidence level. More specifically, for each " + }, + { + "bbox": [ + 130, + 243, + 479, + 281 + ], + "type": "inline_equation", + "content": "t \\in \\{1, \\dots, T\\}" + }, + { + "bbox": [ + 130, + 243, + 479, + 281 + ], + "type": "text", + "content": ", we set " + }, + { + "bbox": [ + 130, + 243, + 479, + 281 + ], + "type": "inline_equation", + "content": "p_t = M_t^C(c_t)" + }, + { + "bbox": [ + 130, + 243, + 479, + 281 + ], + "type": "text", + "content": " and," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 205, + 283, + 480, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 283, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 205, + 283, + 480, + 299 + ], + "type": "interline_equation", + "content": "\\hat {P} (p) = \\left| \\left\\{c _ {t}: M _ {t} ^ {C} \\left(c _ {t}\\right) \\leq p \\text {f o r} t = 1, \\dots , T \\right\\} \\right| / T, \\tag {5}", + "image_path": "acfd91553abee1ed41b5035d7d537db848fb5fcbcff952ba99bcd0e97f26e98c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 303, + 479, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 479, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 479, + 350 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 303, + 479, + 350 + ], + "type": "inline_equation", + "content": "M \\equiv F" + }, + { + "bbox": [ + 130, + 303, + 479, + 350 + ], + "type": "text", + "content": " for uncalibrated errors and " + }, + { + "bbox": [ + 130, + 303, + 479, + 350 + ], + "type": "inline_equation", + "content": "M \\equiv \\hat{F}" + }, + { + "bbox": [ + 130, + 303, + 479, + 350 + ], + "type": "text", + "content": " for calibrated errors, and " + }, + { + "bbox": [ + 130, + 303, + 479, + 350 + ], + "type": "inline_equation", + "content": "C \\in \\{R, G, B\\}" + }, + { + "bbox": [ + 130, + 303, + 479, + 350 + ], + "type": "text", + "content": ". Note that this formulation of the calibration error is equivalent to the one in [12] with a confidence level for every unique " + }, + { + "bbox": [ + 130, + 303, + 479, + 350 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 130, + 303, + 479, + 350 + ], + "type": "text", + "content": " and weights that more significantly penalize errors from frequently predicted confidence levels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 351, + 480, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 480, + 471 + ], + "type": "text", + "content": "Following [12], we plot " + }, + { + "bbox": [ + 130, + 351, + 480, + 471 + ], + "type": "inline_equation", + "content": "\\{(p_t, \\hat{P}(p_t))\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 351, + 480, + 471 + ], + "type": "text", + "content": " before and after calibration for each color channel to generate calibration curves. A perfectly calibrated forecaster would produce the straight line " + }, + { + "bbox": [ + 130, + 351, + 480, + 471 + ], + "type": "inline_equation", + "content": "p_t = \\hat{P}(p_t)" + }, + { + "bbox": [ + 130, + 351, + 480, + 471 + ], + "type": "text", + "content": " as each expected confidence level would equal the empirical one. Intuitively, our version of the calibration error is the mean squared vertical distance of points on the calibration curve from a perfectly straight line. If an expected confidence level occurs " + }, + { + "bbox": [ + 130, + 351, + 480, + 471 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 351, + 480, + 471 + ], + "type": "text", + "content": " times in the test data, its distance is counted " + }, + { + "bbox": [ + 130, + 351, + 480, + 471 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 351, + 480, + 471 + ], + "type": "text", + "content": " times in the mean. Following [29], we additionally report the negative log-likelihood (NLL) of the test data averaged across all scenes. Following [24], we include PSNR and LPIPS [32] to evaluate image quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 479, + 479, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 479, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 479, + 528 + ], + "type": "text", + "content": "Datasets. We use 30 scenes from 3 datasets: Realistic Synthetic " + }, + { + "bbox": [ + 130, + 479, + 479, + 528 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 479, + 479, + 528 + ], + "type": "text", + "content": " [16], the subset of scenes in DTU [8] used in [24], and LLFF [15] for training the metacalibrator and test it on a hold-out scene from either LLFF or DTU to show it generalizes to new target scenes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 536, + 480, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 480, + 609 + ], + "type": "text", + "content": "Baselines We compare our approach against the state-of-the-art method for NeRF uncertainty estimation DANE [29] as well as other methods in Sec. 4.1. In Tab. 1, following [29], we implement the naive ensembles approach and DANE using a public implementation of Instant-NGP [2,17] and 5 ensemble members. In Sec. 4.2, we compare the uncalibrated uncertainties to the uncertainties calibrated by our meta-calibrator." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": "Meta-calibrator Design The results guiding our decisions to use 3 Principal Component Analysis (PCA) components to represent the calibration curves, fit the PCA components using 21 training scenes, and train the meta-calibrator on 30 scenes to predict the PCA coefficients are shown in Fig. 3." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "type": "text", + "content": "N. Amini-Naieni et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 177, + 239, + 258 + ], + "blocks": [ + { + "bbox": [ + 132, + 177, + 239, + 258 + ], + "lines": [ + { + "bbox": [ + 132, + 177, + 239, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 177, + 239, + 258 + ], + "type": "image", + "image_path": "3e0ba27d002685edcc41c61c18eb02494e05424fd4462c22a89c8e3c9d4af899.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 259, + 239, + 300 + ], + "lines": [ + { + "bbox": [ + 132, + 259, + 239, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 239, + 300 + ], + "type": "text", + "content": "(a) Explained variance plot showing most of the variance of the calibration curves is explained using only 3 PCA components." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 253, + 161, + 361, + 227 + ], + "blocks": [ + { + "bbox": [ + 253, + 161, + 361, + 227 + ], + "lines": [ + { + "bbox": [ + 253, + 161, + 361, + 227 + ], + "spans": [ + { + "bbox": [ + 253, + 161, + 361, + 227 + ], + "type": "image", + "image_path": "e192eca2cd13d334ea21be4c2534f2750d2e441fd97688027a0d3ee19f0e0413.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 228, + 360, + 300 + ], + "lines": [ + { + "bbox": [ + 253, + 228, + 360, + 300 + ], + "spans": [ + { + "bbox": [ + 253, + 228, + 360, + 300 + ], + "type": "text", + "content": "(b) Graph showing calibration error on new scenes is sufficiently low when using 21 scenes to fit 3-coefficient PCA model (i.e. it's an order of magnitude lower than we can expect from the final calibration, meaning good generalization to unseen scenes)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 374, + 160, + 481, + 226 + ], + "blocks": [ + { + "bbox": [ + 374, + 160, + 481, + 226 + ], + "lines": [ + { + "bbox": [ + 374, + 160, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 374, + 160, + 481, + 226 + ], + "type": "image", + "image_path": "cae523db36afed04044cba723c427bb76988a41c52a4efc60108223fe8037d2b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 228, + 481, + 300 + ], + "lines": [ + { + "bbox": [ + 374, + 228, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 374, + 228, + 481, + 300 + ], + "type": "text", + "content": "(c) Graph showing increasing number of training scenes for meta-calibrator decreases calibration error for the test scene " + }, + { + "bbox": [ + 374, + 228, + 481, + 300 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 374, + 228, + 481, + 300 + ], + "type": "text", + "content": "-Rex from LLFF [15], with 30 scenes resulting in good generalization to the held-out scene and well-calibrated uncertainties." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 310, + 482, + 353 + ], + "lines": [ + { + "bbox": [ + 130, + 310, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 482, + 353 + ], + "type": "text", + "content": "Fig. 3: Meta-calibrator design decisions. Results showing using 3 components for Principal Component Analysis (PCA) model of calibration curves, 21 scenes to fit PCA model, and 30 scenes to train meta-calibrator achieves good generalization to new test scenes." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 230, + 554, + 382, + 606 + ], + "blocks": [ + { + "bbox": [ + 130, + 456, + 482, + 544 + ], + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 544 + ], + "type": "text", + "content": "Table 1: Quantitative results on standard sparse NeRF benchmark. Our proposed approach results in significantly better uncertainties and image quality than the state-of-the-art NeRF uncertainty estimation method DANE [29] does on the challenging 3-view LLFF [15] dataset. Specifically, our meta-calibrator reduces the calibration error to " + }, + { + "bbox": [ + 130, + 456, + 482, + 544 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 130, + 456, + 482, + 544 + ], + "type": "text", + "content": " of DANE's calibration error and the negative log-likelihood to be over 100 " + }, + { + "bbox": [ + 130, + 456, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 130, + 456, + 482, + 544 + ], + "type": "text", + "content": " lower than DANE's. Note: lower calibration error (Cal. Err.) and negative log-likelihood (NLL) values indicate more accurate uncertainties. Results are averaged over all 8 scenes in LLFF." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 230, + 554, + 382, + 606 + ], + "lines": [ + { + "bbox": [ + 230, + 554, + 382, + 606 + ], + "spans": [ + { + "bbox": [ + 230, + 554, + 382, + 606 + ], + "type": "table", + "html": "
Uncert aintyImageQuality
Cal. Err. (↓)NLL (↓)PSNR (↑)LPIPS (↓)
Naïve Ens.0.05054.3915.190.646
DANE [29]0.04413.7515.190.646
Ours0.0026-0.6819.340.235
", + "image_path": "052d8564e43b4738dc859bb16c4b5211d53dd8da70521e6538e95c5ad13b0592.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 317, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 317, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 317, + 128 + ], + "type": "text", + "content": "4.1 Comparison to State-of-the-art" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 256 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 256 + ], + "type": "text", + "content": "In this section, we compare our approach to prior methods for NeRF uncertainty estimation. We achieve more accurate uncertainties (94% reduction in calibration error and over 100% reduction in negative log-likelihood) than those estimated by DANE [29], the state-of-the-art method. These results are shown in Tab. 1 for the challenging 3-view LLFF [15] dataset from prior work on sparse novel view synthesis [19,24,25] and Tab. 2 for the less challenging version of LLFF from prior work on NeRF uncertainty estimation [14,26,27,29]. In Fig. 5a, we compare the calibration curves from our approach to DANE's, illustrating that the meta-calibrator predicts expected confidences that match the true ones while DANE does not." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 274, + 479, + 396 + ], + "blocks": [ + { + "bbox": [ + 134, + 274, + 479, + 396 + ], + "lines": [ + { + "bbox": [ + 134, + 274, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 134, + 274, + 479, + 396 + ], + "type": "image", + "image_path": "4d7c7acee6cd8b344d416a7a078bafb068672e896c6b653369c89db8153bef5a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 406, + 482, + 517 + ], + "lines": [ + { + "bbox": [ + 130, + 406, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 482, + 517 + ], + "type": "text", + "content": "Fig. 4: Quantitative comparison of uncalibrated and calibrated uncertainties. In (a), we show calibration curves on test data from four scenes in LLFF [15]. The color of each curve indicates the color channel it corresponds to. The calibrated curves are much closer to the ideal calibration (dashed lines), demonstrating that the meta-calibrator works very well. In (b), the average calibration error and negative log-likelihood before and after calibration are reported for LLFF, clearly showing the meta-calibrator improves the accuracy of the uncertainties (lowering calibration error and negative log-likelihood). To test generalization, the meta-calibrator was also applied to held-out scenes in DTU [8], achieving a " + }, + { + "bbox": [ + 130, + 406, + 482, + 517 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 406, + 482, + 517 + ], + "type": "text", + "content": " reduction in calibration error on average." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 561, + 375, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 561, + 375, + 573 + ], + "spans": [ + { + "bbox": [ + 132, + 561, + 375, + 573 + ], + "type": "text", + "content": "4.2 Comparison to Uncalibrated Uncertainties" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "In this section, we compare our uncalibrated base NeRF uncertainties to our calibrated uncertainties obtained from applying the proposed meta-calibrator. In Fig. 6 we show that the calibrated uncertainties better highlight floaters and other errors in the NeRF renderings. In Fig. 4, we show that the meta-calibrator predicts expected confidence that closely match the true ones, lowering both the calibration error and the negative log-likelihood of the uncalibrated uncertainties." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "type": "text", + "content": "N. Amini-Naieni et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 251, + 487, + 350 + ], + "blocks": [ + { + "bbox": [ + 130, + 135, + 482, + 245 + ], + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 245 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 245 + ], + "type": "text", + "content": "Table 2: Quantitative results on standard NeRF uncertainty estimation benchmark. Here, we present results on the LLFF [15] dataset used in prior work [14, 26, 27, 29] on uncertainty estimation for NeRFs. This dataset is less challenging than the one in Tab. 1 since 4-12 views are used for training instead of 3. Our proposed approach results in significantly better uncertainties than prior methods for NeRF uncertainty estimation on all 8 scenes in LLFF. Note: lower negative log-likelihood values indicate more accurate uncertainties. " + }, + { + "bbox": [ + 130, + 135, + 482, + 245 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 135, + 482, + 245 + ], + "type": "text", + "content": " indicates the number of ensemble members, and MC-DO refers to Monte Carlo Dropout sampling with " + }, + { + "bbox": [ + 130, + 135, + 482, + 245 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 135, + 482, + 245 + ], + "type": "text", + "content": " sample configurations. This table is Tab. 1 from [29] with our results added as an additional column. Please refer to [29] for further details." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 251, + 487, + 350 + ], + "lines": [ + { + "bbox": [ + 132, + 251, + 487, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 487, + 350 + ], + "type": "table", + "html": "
Scene# of Train. ViewsNegative Log-likelihood (↓)
MC-DO M=5Naïve Ens. M=5NeRF-W [14]S-NeRF [27]CF-NeRF [26]DANE [29] M=5DANE [29] M=10Ours
Fern44.902.472.162.01—-0.98-1.00-1.41
Orchids55.742.232.241.95—-0.28-0.31-0.84
Leaves52.722.660.790.68—0.970.73-1.19
Flower74.631.631.711.27—1.000.85-2.05
Fortress85.192.291.04-0.03—-1.30-1.30-1.99
Room85.062.134.932.35—-1.35-1.35-2.17
T-Rex114.102.281.911.37—-0.31-0.69-1.49
Horns124.182.170.780.60—-0.55-0.66-2.18
Avg.4.572.231.951.270.57-0.35-0.47-1.67
", + "image_path": "98b0e41a417b73ed8991fc2a21a635878910fca5ab3c73964ffc0bec10abf746.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 201, + 400, + 296, + 494 + ], + "blocks": [ + { + "bbox": [ + 201, + 400, + 296, + 494 + ], + "lines": [ + { + "bbox": [ + 201, + 400, + 296, + 494 + ], + "spans": [ + { + "bbox": [ + 201, + 400, + 296, + 494 + ], + "type": "image", + "image_path": "2d07211d4bd062480b310d0bdf06858042dc99302597f89c54a366456d60b3c2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 497, + 306, + 522 + ], + "lines": [ + { + "bbox": [ + 200, + 497, + 306, + 522 + ], + "spans": [ + { + "bbox": [ + 200, + 497, + 306, + 522 + ], + "type": "text", + "content": "(a) Calibration curves for the Fern scene from 3-view LLFF [15]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 312, + 416, + 406, + 502 + ], + "blocks": [ + { + "bbox": [ + 312, + 416, + 406, + 502 + ], + "lines": [ + { + "bbox": [ + 312, + 416, + 406, + 502 + ], + "spans": [ + { + "bbox": [ + 312, + 416, + 406, + 502 + ], + "type": "image", + "image_path": "e3e883abb1668e94b629fb55e7d1037e931ef4f52c6da77c20f6d51dbe06af47.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 505, + 413, + 522 + ], + "lines": [ + { + "bbox": [ + 307, + 505, + 413, + 522 + ], + "spans": [ + { + "bbox": [ + 307, + 505, + 413, + 522 + ], + "type": "text", + "content": "(b) Next-best View Planning on Horns from LLFF [15]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 532, + 482, + 642 + ], + "lines": [ + { + "bbox": [ + 130, + 532, + 482, + 642 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 482, + 642 + ], + "type": "text", + "content": "Fig. 5: Comparison to DANE [29]. Results comparing our uncertainties to those from the state-of-the-art method DANE. In (a) we show DANE's RGB calibration curves are not closely aligned with the perfectly calibrated lines, meaning it is miscalibrated. It is significantly over-confident for expected confidence levels close to 1 and under-confident for confidence levels close to 0. In comparison, the curves for our approach are extremely close to the ideal calibration (dashed lines), demonstrating that the meta-calibrator works very well, predicting expected confidences that match the true ones. This is also verified by how our calibration error is over two orders of magnitude smaller than DANE's. In (b) we show that our approach results in more efficient performance gains over DANE for next-best view planning." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 113, + 203, + 201 + ], + "blocks": [ + { + "bbox": [ + 134, + 113, + 203, + 201 + ], + "lines": [ + { + "bbox": [ + 134, + 113, + 203, + 201 + ], + "spans": [ + { + "bbox": [ + 134, + 113, + 203, + 201 + ], + "type": "image", + "image_path": "772e986f14d98b9828f9bfaedd02765dd81cdf962ac1fb0f1f0b7934a7d90052.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 201, + 168, + 207 + ], + "lines": [ + { + "bbox": [ + 136, + 201, + 168, + 207 + ], + "spans": [ + { + "bbox": [ + 136, + 201, + 168, + 207 + ], + "type": "text", + "content": "(a) Ground Truth" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "lines": [ + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 482, + 276 + ], + "type": "text", + "content": "Fig. 6: Qualitative comparison of uncalibrated and calibrated uncertainties from the Flower scene in LLFF [15]. The calibrated uncertainties (d) clearly detect incorrect regions (indicated by the red boxes) better than the uncalibrated uncertainties (c) do. This is apparent by noting that (d) and (e) look more similar than (c) and (e)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 205, + 114, + 272, + 201 + ], + "blocks": [ + { + "bbox": [ + 205, + 114, + 272, + 201 + ], + "lines": [ + { + "bbox": [ + 205, + 114, + 272, + 201 + ], + "spans": [ + { + "bbox": [ + 205, + 114, + 272, + 201 + ], + "type": "image", + "image_path": "0396d509886d8986cbbb310048ba4947368c50a9cd34d677c2dc8b942ce16e40.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 201, + 236, + 207 + ], + "lines": [ + { + "bbox": [ + 206, + 201, + 236, + 207 + ], + "spans": [ + { + "bbox": [ + 206, + 201, + 236, + 207 + ], + "type": "text", + "content": "(b) NeRF Image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 273, + 114, + 340, + 201 + ], + "blocks": [ + { + "bbox": [ + 273, + 114, + 340, + 201 + ], + "lines": [ + { + "bbox": [ + 273, + 114, + 340, + 201 + ], + "spans": [ + { + "bbox": [ + 273, + 114, + 340, + 201 + ], + "type": "image", + "image_path": "81cdbc67aa61107e29a81b8c38c286f8543e19698907e8ecf1aec0fba758f609.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 202, + 328, + 208 + ], + "lines": [ + { + "bbox": [ + 274, + 202, + 328, + 208 + ], + "spans": [ + { + "bbox": [ + 274, + 202, + 328, + 208 + ], + "type": "text", + "content": "(c) Uncalibrated Uncertainty" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 342, + 114, + 411, + 201 + ], + "blocks": [ + { + "bbox": [ + 342, + 114, + 411, + 201 + ], + "lines": [ + { + "bbox": [ + 342, + 114, + 411, + 201 + ], + "spans": [ + { + "bbox": [ + 342, + 114, + 411, + 201 + ], + "type": "image", + "image_path": "ca22ab0ca68246f8540977f906d1c3eb9e1c771bafb219b95e24638c50cc8fae.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 344, + 202, + 394, + 208 + ], + "lines": [ + { + "bbox": [ + 344, + 202, + 394, + 208 + ], + "spans": [ + { + "bbox": [ + 344, + 202, + 394, + 208 + ], + "type": "text", + "content": "(d) Calibrated Uncertainty" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 413, + 114, + 480, + 201 + ], + "blocks": [ + { + "bbox": [ + 413, + 114, + 480, + 201 + ], + "lines": [ + { + "bbox": [ + 413, + 114, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 413, + 114, + 480, + 201 + ], + "type": "image", + "image_path": "0c476a28b0a02923fae760d9da2b1a27895cd36bf62cc1fb6f6d757335777f28.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 201, + 443, + 207 + ], + "lines": [ + { + "bbox": [ + 414, + 201, + 443, + 207 + ], + "spans": [ + { + "bbox": [ + 414, + 201, + 443, + 207 + ], + "type": "text", + "content": "(e) Actual Error" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 299, + 354, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 299, + 354, + 312 + ], + "spans": [ + { + "bbox": [ + 131, + 299, + 354, + 312 + ], + "type": "text", + "content": "4.3 Application: Next-best View Planning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 319, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 511 + ], + "type": "text", + "content": "In this section, we show that our uncertainties can be leveraged for next-best view planning. Specifically, we start by training the NeRF model for 2000 iterations on a training set of three images. The next-best view is selected by evaluating the average calibrated pixel uncertainty (obtained using the meta-calibrator) for each of the candidate views, and the view with the highest uncertainty is added to the training set. The average PSNR of the test images is reported after each training iteration. In Fig. 5b we show that using our approach results in greater performance gains (higher PSNRs) than DANE [29] does. In the Fig. 1 of the Supplementary, we also compare the information gain from rays selected according to the highest calibrated uncertainties to the information gain from rays selected according to the highest uncalibrated uncertainties and show our meta-calibrator produces higher average PSNRs on the test set for scenes in DTU [8] than the uncalibrated uncertainties do. Thus, calibration specifically re-orders the pixel uncertainties so that rays more likely to raise the PSNR are picked earlier in next-best view planning. In Sec. 5 of the Supplementary, we include a detailed theoretical example showing that such re-ordering is possible." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 529, + 220, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 529, + 220, + 541 + ], + "spans": [ + { + "bbox": [ + 131, + 529, + 220, + 541 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 554, + 482, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 554, + 482, + 661 + ], + "spans": [ + { + "bbox": [ + 130, + 554, + 482, + 661 + ], + "type": "text", + "content": "In this paper we addressed the open problem of obtaining calibrated uncertainties from NeRF models. We introduce the concept of a meta-calibrator that infers the calibration curves from scene features, and using this approach achieve state-of-the-art uncertainty without holding out any ground truth data from the target scene. By enabling efficient and accurate calibration of NeRF models without relying on additional data, our method represents a significant step forward in the practical application of NeRF to real-world scenarios and opens up new avenues for the use of NeRF in situations where data is limited and uncertainty is critical." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "type": "text", + "content": "N. Amini-Naieni et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "Acknowledgements. The authors would like to thank Seunghyeon Seo for his support on FlipNeRF, and Oishi Deb for her insightful feedback on uncertainty estimation. Niki Amini-Naieni is funded by an AWS Studentship, the Reuben Foundation, and the AIMS CDT program at the University of Oxford. Tomas Jakab is sponsored by EPSRC VisualAI EP/T028572/1 and Andrea Vedaldi by ERC-CoG UNION 101001212." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 207, + 197, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 197, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 197, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 233, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 233, + 481, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 233, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 233, + 481, + 266 + ], + "type": "text", + "content": "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV. pp. 5835-5844 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 267, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 481, + 289 + ], + "type": "text", + "content": "2. Bhalgat, Y.: Hashnerf-pytorch. https://github.com/yashbhalgat/HashNeRF-pytorch/ (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 289, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 481, + 311 + ], + "type": "text", + "content": "3. Gafni, G., Thies, J., Zollhöfer, M., Nießner, M.: Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In: CVPR. pp. 8649-8658 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 312, + 481, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 481, + 333 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 481, + 333 + ], + "type": "text", + "content": "4. Ghoshal, B., Tucker, A.: On calibrated model uncertainty in deep learning. In: ECML (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 334, + 481, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 334, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 334, + 481, + 355 + ], + "type": "text", + "content": "5. Goli, L., Reading, C., Sellan, S., Jacobson, A., Tagliasacchi, A.: Bayes' Rays: Uncertainty quantification in neural radiance fields. ArXiv abs/2309.03185 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 356, + 481, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 356, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 356, + 481, + 378 + ], + "type": "text", + "content": "6. Jain, A., Tancik, M., Abbeel, P.: Putting nef on a diet: Semantically consistent few-shot view synthesis. In: ICCV. pp. 5865-5874 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 379, + 481, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 481, + 399 + ], + "type": "text", + "content": "7. Jang, T.J., Hyun, C.M.: Nerf solves undersampled mri reconstruction. ArXiv abs/2402.13226 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 400, + 481, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 400, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 400, + 481, + 422 + ], + "type": "text", + "content": "8. Jensen, R.R., Dahl, A., Vogiatzis, G., Tola, E., Aanaes, H.: Large scale multi-view stereopsis evaluation. In: CVPR (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 422, + 481, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 481, + 444 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 481, + 444 + ], + "type": "text", + "content": "9. Jin, L., Chen, X., Ruckin, J., Popovi'c, M.: Neu-nbv: Next best view planning using uncertainty estimation in image-based neural rendering. In: IROS (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 445, + 481, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 481, + 466 + ], + "type": "text", + "content": "0. Kajiya, J.T., Von Herzen, B.P.: Ray tracing volume densities. In: SIGGRAPH. p. 165-174 (1984)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 467, + 481, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 481, + 499 + ], + "type": "text", + "content": "1. Kosiorek, A.R., Strathmann, H., Zoran, D., Moreno, P., Schneider, R., Mokr'a, S., Rezende, D.J.: Nerf-vae: A geometry aware 3d scene generative model. In: ICML (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 500, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 500, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 500, + 481, + 521 + ], + "type": "text", + "content": "2. Kuleshov, V., Fenner, N., Ermon, S.: Accurate uncertainties for deep learning using calibrated regression. In: ICML. pp. 2796-2804 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 522, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 481, + 544 + ], + "type": "text", + "content": "3. Liu, L., Gu, J., Lin, K.Z., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. In: NIPS (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 544, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 481, + 576 + ], + "type": "text", + "content": "4. Martin-Brualla, R., Radwan, N., Sajjadi, M.S.M., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: CVPR (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 577, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 481, + 609 + ], + "type": "text", + "content": "5. Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. In: TOG (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 481, + 643 + ], + "type": "text", + "content": "6. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "7. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. In: ACM Trans. Graph. (2022)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Instant Uncertainty Calibration of NeRFs Using a Meta-Calibrator" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 523 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "18. Neff, T., Stadlbauer, P., Parger, M., Kurz, A., Mueller, J.H., Chaitanya, C.R.A., Kaplanyan, A., Steinberger, M.: Donerf: Towards real-time rendering of compact neural radiance fields using depth oracle networks. In: CGF. pp. 45-59 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "19. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S.M., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: CVPR (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 237 + ], + "type": "text", + "content": "20. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Howes, R., Huang, P.Y., Xu, H., Sharma, V., Li, S.W., Galuba, W., Rabbat, M., Assran, M., Ballas, N., Synnaeve, G., Misra, I., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 238, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 259 + ], + "type": "text", + "content": "21. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: ICCV (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "22. Ran, Y., Zeng, J., He, S., Li, L., Chen, Y., Lee, G.H., Chen, J., Ye, Q.: Neurar: Neural uncertainty for autonomous 3d reconstruction. In: RAL (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "type": "text", + "content": "23. Rebain, D., Jiang, W., Yazdani, S., Li, K., Yi, K.M., Tagliasacchi, A.: Derf: Decomposed radiance fields. In: CVPR. pp. 14148-14156 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "24. Seo, S., Chang, Y., Kwak, N.: Flipnerf: Flipped reflection rays for few-shot novel view synthesis. In: ICCV (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "text", + "content": "25. Seo, S., Han, D., Chang, Y., Kwak, N.: Mixnerf: Modeling a ray with mixture density for novel view synthesis from sparse inputs. In: CVPR. pp. 20659-20668 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "type": "text", + "content": "26. Shen, J., Agudo, A., Moreno-Noguer, F., Ruiz, A.: Conditional-flow nerf: Accurate 3d modelling with reliable uncertainty quantification. In: ECCV (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 403 + ], + "type": "text", + "content": "27. Shen, J., Ruiz, A., Agudo, A., Moreno-Noguer, F.: Stochastic neural radiance fields: Quantifying uncertainty in implicit 3d representations. In: 3DV. pp. 972-981 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "text", + "content": "28. Sitzmann, V., Martel, J.N., Bergman, A.W., Lindell, D.B., Wetzstein, G.: Implicit neural representations with periodic activation functions. In: NIPS (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 424, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 446 + ], + "type": "text", + "content": "29. Sünderhauf, N., Abou-Chakra, J., Miller, D.: Density-aware nef ensembles: Quantifying predictive uncertainty in neural radiance fields. In: ICRA (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "30. Wang, P., Liu, Y., Chen, Z., Liu, L., Liu, Z., Komura, T., Theobalt, C., Wang, W.: F2-nerf: Fast neural radiance field training with free camera trajectories. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "text", + "content": "31. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. ArXiv abs/2010.07492 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 501, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 523 + ], + "type": "text", + "content": "32. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 257, + 100 + ], + "type": "text", + "content": "N. Amini-Naieni et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/InstructGIE_ Towards Generalizable Image Editing/d8d7c0e8-e8d9-4aa1-9b02-c4948b5c2541_content_list.json b/2024/InstructGIE_ Towards Generalizable Image Editing/d8d7c0e8-e8d9-4aa1-9b02-c4948b5c2541_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1a87d6c48d0586ae2328fd7bf570a0e730ba9660 --- /dev/null +++ b/2024/InstructGIE_ Towards Generalizable Image Editing/d8d7c0e8-e8d9-4aa1-9b02-c4948b5c2541_content_list.json @@ -0,0 +1,2820 @@ +[ + { + "type": "text", + "text": "InstructGIE: Towards Generalizable Image Editing", + "text_level": 1, + "bbox": [ + 256, + 140, + 746, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zichong Meng\\*, Changdi Yang\\*, Jun Liu\\*, Hao Tang\\*\\*, Pu Zhao\\*\\*, and Yanzhi Wang\\*\\*", + "bbox": [ + 227, + 212, + 776, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Northeastern University, Boston MA 02115, USA", + "bbox": [ + 328, + 253, + 671, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Peking University, Beijing 100871, China &", + "bbox": [ + 346, + 268, + 653, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Carnegie Mellon University, Pittsburgh PA 15213, USA", + "bbox": [ + 316, + 284, + 684, + 297 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1163a20c0ef3ae95a0a9ff40e4ae76f0a4ebc103dbef8ac4214961895c7a95c4.jpg", + "image_caption": [ + "Like in Game", + "Add a Cow", + "Fig. 1: Demo results of the proposed InstructGIE framework on various image manipulation tasks to both humans and scenes. By our proposed method, our model can generalize to generate the desired output with great detail qualities." + ], + "image_footnote": [], + "bbox": [ + 238, + 340, + 764, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent advances in image editing have been driven by the development of denoising diffusion models, marking a significant leap", + "bbox": [ + 261, + 771, + 740, + 800 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contribution", + "bbox": [ + 217, + 811, + 364, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*\\* Corresponding Authors", + "bbox": [ + 215, + 825, + 390, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "forward in this field. Despite these advances, the generalization capabilities of recent image editing approaches remain constrained. In response to this challenge, our study introduces a novel image editing framework with enhanced generalization robustness by boosting in-context learning capability and unifying language instruction. This framework incorporates a module specifically optimized for image editing tasks, leveraging the VMamba block and an editing-shift matching strategy to augment in-context learning. Furthermore, we unveil a selective area-matching technique specifically engineered to address and rectify corrupted details in generated images, such as human facial features, to further improve the quality. Another key innovation of our approach is the integration of a language unification technique, which aligns language embeddings with editing semantics to elevate the quality of image editing. Moreover, we compile the first dataset for image editing with visual prompts and editing instructions that could be used to enhance in-context capability. Trained on this dataset, our methodology not only achieves superior synthesis quality for trained tasks, but also demonstrates robust generalization capability across unseen vision tasks through tailored prompts. Our project page is available at https://cr8br0ze.github.io/InstructGIE.", + "bbox": [ + 261, + 147, + 738, + 410 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: Image Editing $\\cdot$ In-Context Learning $\\cdot$ Diffusion Model", + "bbox": [ + 261, + 425, + 714, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 469, + 375, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a crucial task in computer vision, image editing has witnessed significant improvements enhanced with the increasingly popular denoising stable diffusion techniques in recent years [22,25,38,40,41]. Given a set of text or image prompts as generation constraints or instructions, diffusion-based image editing can follow the instructions and synthesize an edited image. However, since the model does not have the capability to accurately model all possible samples in the conditional distribution space [17], if specific instructions are not included in the training dataset, current diffusion-based image editing methods can hardly generate satisfactory results. Thus, editing performance largely depends on the training dataset without superior generalization capabilities.", + "bbox": [ + 217, + 507, + 785, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, large language models (LLMs) have proven extraordinary abilities to learn from contexts, referred to as in-context learning, which allows LLMs to perform unseen tasks by providing a combination of input-output examples and a query input. Inspired by the potential to enhance the generalization of the model with LLMs, [38,41] explore in-context learning for computer vision tasks, allowing them to perform unseen tasks with novel vision-language prompt designs. However, these methods are not tailored for image editing applications, leading to unsatisfying synthetic qualities with inaccurate or incorrect output and lack of detail. To improve the generalization of image editing with improved synthetic image quality, it is crucial to effectively understand the text & image prompts and specifically control image editing details, which is challenging in the current literature.", + "bbox": [ + 217, + 659, + 785, + 838 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Meng and C. Yang et al.", + "bbox": [ + 271, + 114, + 454, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose InstructGIE, an image editing framework with enhanced generalizability. We improve image editing performance from both visual and text aspects. (i) For the visual information, we incorporate a VMamba-based module to specifically enhance the image editing outputs. As VMamba [26] has proven its better performance in capturing in-context information from inputs with larger receptive fields [26], we leverage VMamba and propose an editing-shift matching strategy to augment in-context learning. Furthermore, since current image editing works do not perform well in generating correct features with accurate details, we unveil a selective area-matching technique specifically engineered to address and rectify corrupted details in generated images, such as human facial features, to further improve the quality. (ii) Another key innovation of our approach is the integration of a language unification technique, which aligns language embeddings with editing semantics to elevate the quality of image editing. Our framework not only achieves superior in-context generation for trained tasks but also demonstrates robust generalization across unseen vision tasks. Moreover, we compile a publicly available image editing dataset with plenty of visual prompts and editing instructions for better generalization evaluation of image editing. Our contributions are summarized as follows:", + "bbox": [ + 217, + 146, + 785, + 417 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an image editing framework, including in-context learning enhancement and language unification strategies, specifically designed to enhance generalization ability from both visual and text domains.", + "- We compile the first dataset for image editing with visual prompts and editing instructions that could be used to enhance in-context capability.", + "- We conduct extensive experiments and achieve great generalization ability in the multiple unseen image editing task, both quantitatively and qualitatively." + ], + "bbox": [ + 225, + 426, + 784, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 561, + 395, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Denoising Stable Diffusion Based Image Editing", + "text_level": 1, + "bbox": [ + 215, + 590, + 656, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denoising Stable Diffusion [16,35,36] based image editing could follow guidance from text or image prompts. With the foundation of text-guided models offering rich generative capabilities, there has been a surge in research aimed at adapting these models for image manipulation tasks from textual descriptions. To steer the image editing process in the desired direction, the use of models like CLIP to fine-tune diffusion models has become a common practice. Although these methods [3,22,41] have shown impressive results, they often involve costly finetuning processes. Recent innovations [14] have introduced techniques that inject cross-attention into the models to more effectively edit specific semantic areas within the spatial feature maps. Further advancements [24] have enhanced these techniques by adding semantic loss or applying attention loss to refine the integration of plugged features, improving the precision and quality of the editing outcomes. [38] proposes a framework that could learn instructions from visual image pairs for more accurate editing and firstly formulate this task as an image inpainting problem.", + "bbox": [ + 217, + 613, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "InstructGIE: Towards Generalizable Image Editing", + "bbox": [ + 390, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Generalization Capability in Visual Tasks", + "text_level": 1, + "bbox": [ + 215, + 146, + 604, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In-context learning is widely applied in the field of natural language processing (NLP), enabling models to adapt to new tasks such as translation, question answering, and complex reasoning. NLP models utilize in-context examples, comprising text and corresponding labels, to tackle tasks they haven't seen before. However, applying in-context learning to the visual domain introduces more challenges and remains less explored. A significant hurdle is the nature of fixed-size input requirements for vision models, as opposed to variable-length text inputs that can be managed by language models. Vision models generally struggle with processing inputs of varying sizes, making it impractical to process multiple image prompts in one-shot for global understanding. Moreover, in intricate visual understanding, specific instructions are often implied from a limited set of image examples rather than explicitly stated, which poses additional difficulties for vision models in identifying and understanding high-level visual relationships.", + "bbox": [ + 212, + 172, + 787, + 368 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent strides in applying masked image modeling have marked a step forward in improving in-context learning for vision models. The method proposed by [41], employing a masked autoencoder-based technique, predicts a missing image within a two-by-two grid, using two images as in-context examples and another as the query. This concept was later expanded by [38] with a multitask framework. Despite their progress, such inpainting methods are limited by the necessity of a fixed number of in-context examples and increased memory demands. Painter, highlighted in [40], exemplifies an inpainting approach tailored for versatility across various vision tasks.", + "bbox": [ + 212, + 369, + 787, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast, inspired by ControlNet [44], [41] refines the framework by adding an additional pair of example images and employing a multitask supervised finetuning method. Prompt diffusion excels in visual in-context learning. However, it faces certain limitations or challenges in its practical applications.", + "bbox": [ + 212, + 503, + 785, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Dataset for Diffusion-based Image Editing", + "text_level": 1, + "bbox": [ + 215, + 587, + 609, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Currently, various types of datasets are used for training in diffusion-based image editing. There are datasets that concentrate on specific domains like CelebA [27] and FFHQ [19] for human face image manipulation, AFHQ [7] for animal face image editing, LSUN [42] for object modification, and WikiArt [29] for style transfer. In-the-wild video datasets could also be leveraged to train image editing tasks. The Scannet dataset [9] encompasses a vast array of more than 1,500 indoor scenes from various settings, such as apartments, offices, and hotels, providing extensive annotations. The LRW dataset [8], tailored for lip reading tasks, includes more than 1000 video utterances of 500 distinct words. The UBC-Fashion dataset [43] features 600 videos spanning various clothing categories, with 500 videos allocated for training and 100 for testing, guaranteeing no repetition of individuals in the training set. The DAVIS dataset [43] (Densely Annotated VVideo Segmentation), a widely recognized benchmark for video object segmentation, contains 150 videos in total. There are also image editing works proposing to generate datasets with editing instructions. InstructPix2pix [3] collects over", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Meng and C. Yang et al.", + "bbox": [ + 271, + 114, + 454, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "450,000 training image pairs. For each pair, given an image with its caption, it first uses a finetuned GPT-3 [4] to generate an editing instruction and an edited image caption. Then it employs Stable Diffusion and the Prompt-to-Prompt algorithm [14] to generate edited image following the caption. However, currently there are no datasets with multiple image pairs under one editing instruction, which is crucial to enhance the generalization ability of image editing.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 263, + 369, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recent advances in generative models have been significantly driven by the emergence of diffusion models, which have set new benchmarks in image creation [10,21,37]. These models have found applications across a broad spectrum of areas [2,5,20,39,45], demonstrating their versatility and effectiveness. The fundamental concept behind diffusion models involves starting with an image that is initially just random noise $\\mathbf{x}_T\\sim \\mathcal{N}(0,\\mathbf{I})$ and progressively refining this image step by step until it becomes a high-quality, realistic image $\\mathbf{x}_0$ . This refinement process involves generating intermediate samples $\\mathbf{x}_t$ (for $t\\in \\{0,\\dots ,T\\}$ ), where each sample is defined as:", + "bbox": [ + 212, + 297, + 787, + 434 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t} = \\sqrt {\\alpha_ {t}} \\mathbf {x} _ {0} + \\sqrt {1 - \\alpha_ {t}} \\boldsymbol {\\epsilon} _ {t}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 446, + 785, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the parameter $\\alpha_{t}$ sets the pace of the diffusion process, ranging from $0 = \\alpha_{T} < \\alpha_{T - 1} < \\dots < \\alpha_{1} < \\alpha_{0} = 1$ , and $\\epsilon_t\\sim \\mathcal{N}(0,\\mathbf{I})$ represents the added noise. The model refines the image by applying a neural network $f_{\\theta}(\\mathbf{x}_t,t)$ to each sample $\\mathbf{x}_t$ , followed by the addition of Gaussian noise to produce the next sample $\\mathbf{x}_{t - 1}$ . This neural network is optimized to achieve a denoising goal, striving for $f_{\\theta}(\\mathbf{x}_t,t)\\approx \\epsilon_t$ , resulting in a generative process that closely mimics the desired image distribution.", + "bbox": [ + 212, + 476, + 787, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Expanding this framework to conditional generative modeling, the process involves conditioning the neural network $f_{\\theta}(\\mathbf{x}_t,t,\\mathbf{y})$ on an additional input $\\mathbf{y}$ , enabling the generation of images from a distribution conditioned on $\\mathbf{y}$ . This conditional input could be anything from a low-resolution image, a category label, or a descriptive text sequence. Leveraging the advancements in LLMs [33] and hybrid vision-language models [31], text-to-image diffusion models are developed. These models allow for the creation of detailed, high-resolution images from mere text descriptions, starting with a low-resolution image generated through the diffusion process, which is subsequently refined into a high-resolution image using additional models.", + "bbox": [ + 212, + 583, + 787, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 The Proposed Method", + "text_level": 1, + "bbox": [ + 215, + 758, + 472, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present our framework pipeline in Fig. 2. For efficient training and better controllability, we adopt a line of techniques with popular architectures such as ControlNet [44] and Stable Diffusion [34] to design a generalizable image editing", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "InstructGIE: Towards Generalizable Image Editing", + "bbox": [ + 390, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/558d61ed0c6aa5cec6856d32f1f5db211c09dff21ca1d6c7eb3b3e53c71e943a.jpg", + "image_caption": [ + "Fig. 2: Overall architecture of InstructGIE. The lower pipeline is for both training and inference processes where the model obtains unified editing instructions outputted by Instruction Unification Module $\\mathcal{U}$ and combines with visual prompted input $\\mathbf{Img}^{\\mathrm{VPcon}}$ to pass through Zero-VMamba integrated Stable Diffusion model with ControlNet for output image. The upper pipeline is for training only which compares output image and training ground truth $\\mathbf{Img}^{\\mathrm{train}}$ and computes editing shift loss $\\mathcal{L}_{es}$ with Editing Shift Module and selective area matching loss $\\mathcal{L}_{sam}$ with Selective Area Matching Module." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 766, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tool with accurate high-quality outputs. Specifically, we introduce enhanced in-context learning both at the architecture level and training level to improve the image quality. Furthermore, language instruction unification is adopted to maximize the generalization ability for unseen editing tasks. Moreover, selective area matching is proposed to further improve the output quality with full details.", + "bbox": [ + 212, + 566, + 787, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Enhanced In-Context Learning", + "text_level": 1, + "bbox": [ + 215, + 667, + 517, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual prompting based on inpainting is an effective visual in-context learning method in various computer vision tasks [1,40], which is applied in image editing tasks [38] recently. However, the methods perform poorly in quality when dealing with unseen image manipulation tasks. Therefore, we propose the enhanced in-context learning specifically tailored for generalizable image editing.", + "bbox": [ + 212, + 694, + 787, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reforming Conditioned Latent Diffusion Model. To improve the generalization ability of image editing, it is crucial for the framework to explicitly capture low-level visual editing contexts. Current diffusion-based image editing", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Meng and C. Yang et al.", + "bbox": [ + 271, + 114, + 454, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d3da15b41880757d9c95fae96f376efce5c9ef400fe4063e1481d597a777a623.jpg", + "image_caption": [ + "ERF of ResNet-50" + ], + "image_footnote": [], + "bbox": [ + 256, + 143, + 410, + 261 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/71e7689d683d74dc0c08647a3d29f4ba934dbf57742ff9c21a5ad312fa416d2d.jpg", + "image_caption": [ + "ERF of VIT-L/14", + "Fig. 3: Effective Reception Field (ERF) of ConvNet, ViT, VMamba based model architectures." + ], + "image_footnote": [], + "bbox": [ + 424, + 143, + 578, + 262 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f9ed6567ff258182b8f8926cccb22cfd59509a8d97c2aa0268ed804cda3d13d8.jpg", + "image_caption": [ + "ERF of VMamba-T" + ], + "image_footnote": [], + "bbox": [ + 591, + 143, + 748, + 262 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods [30, 38] that involve visual prompting either adopt ConvNet [23] or ViT [11] as the vision encoder for visual prompts. However, these methods fail to generalize well as they are not able to capture enough visual editing contexts (see Fig. 3). To address this, we formulate the visual prompted condition as a single image $\\mathbf{Img}^{\\mathrm{VPcon}} = \\{\\mathbf{Img}_0^{\\mathrm{in}}, \\mathbf{Img}_0^{\\mathrm{out}}, \\mathbf{Img}_1^{\\mathrm{in}}, \\mathbf{Grey}\\}$ , as shown in Fig 2, with a global effective receptive field (ERF).", + "bbox": [ + 212, + 344, + 782, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Moreover, we further propose to reform the conditioned latent diffusion model Inspired by the recent visual state space model VMamba [26] which exhibits a better global ERF and also emphasizes shifting boundaries of input's four quadrants, we propose to adopt a vision encoder based on Zero-VMamba to fit our structure. Specifically, the vision encoder $\\mathcal{G}_{ven}$ comprehends the visual prompted condition in latent space $x_{vpc}$ as follows,", + "bbox": [ + 212, + 435, + 784, + 526 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\ne _ {v p c} = \\mathcal {G} _ {v e n} \\left(x _ {v p c}; \\Theta_ {g}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 535, + 784, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $e_{vpc}$ is the processed embedding of the visual prompted condition, and $\\Theta_g$ is the model parameters initialized to zeros.", + "bbox": [ + 212, + 559, + 782, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further improve the performance, after each ControlNet trainable copied modules $\\mathcal{F}_c$ with parameters $\\Theta_c$ , we propose to link and inject the processed visual prompted condition information to the frozen Stable Diffusion model $\\mathcal{F}$ with parameters $\\Theta$ through zero-VMamba layer $\\mathcal{G}$ . We use two instances of VMamba with parameters $\\theta_{g1}$ and $\\theta_{g2}$ respectively. The complete model then computes the following,", + "bbox": [ + 212, + 590, + 784, + 679 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\ny _ {v p c} = \\mathcal {F} (x; \\Theta) + \\mathcal {G} (\\mathcal {F} (x + \\mathcal {G} (x _ {v p c}; \\Theta_ {g 1}); \\Theta_ {c}); \\Theta_ {g 2}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 689, + 784, + 707 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $y_{vpc}$ is the output of our conditioned diffusion model block.", + "bbox": [ + 214, + 714, + 689, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our conditioned latent diffusion model can process all four quadrants in our visual prompted conditions with a global receptive field, while it does not generate random noises during initial training steps.", + "bbox": [ + 212, + 729, + 782, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Editing-Shift Matching. Besides the architecture innovation, we incorporate an editing-shift-matching technique to enhance in-context learning ability in image editing with more accurate detailed outputs.", + "bbox": [ + 214, + 794, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "InstructGIE: Towards Generalizable Image Editing", + "bbox": [ + 390, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In specific, for each training ground truth $\\mathbf{Img}^{\\mathrm{train}} = \\{\\mathbf{Img}_0^{\\mathrm{in}},\\mathbf{Img}_0^{\\mathrm{out}},\\mathbf{Img}_1^{\\mathrm{in}},$ $\\mathbf{Img}_1^{\\mathrm{out}}\\}$ , we calculate a implicit editing shift value using CLIP [32] image embedding:", + "bbox": [ + 212, + 143, + 782, + 190 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} (\\mathbf {I m g}) = \\frac {\\sum_ {i = 0} ^ {1} \\operatorname {C L I P} \\left(\\mathbf {I m g} _ {i} ^ {\\text {i n}}\\right) - \\operatorname {C L I P} \\left(\\mathbf {I m g} _ {i} ^ {\\text {o u t}}\\right)}{2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 191, + 785, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During the training process, after predicting the noise, we use it to reverse the noised input and obtain a pseudo output image $\\mathbf{Img}^{\\mathrm{PO}} = \\{\\mathbf{PO}_0^{\\mathrm{in}},\\mathbf{PO}_0^{\\mathrm{out}},\\mathbf{PO}_1^{\\mathrm{in}},$ $\\mathbf{PO}_1^{\\mathrm{out}}\\}$ . Our framework then calculates the editing transfer value of the pseudo output image and deduces a editing shift loss to optimize during our training via the cosine similarity of the two design transfer values:", + "bbox": [ + 214, + 227, + 782, + 303 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {e s} = 1 - \\frac {\\mathcal {T} (\\mathbf {I m g} ^ {\\mathrm {P O}}) \\cdot \\mathcal {T} (\\mathbf {I m g} ^ {\\mathrm {t r a i n}})}{| | \\mathcal {T} (\\mathbf {I m g} ^ {\\mathrm {P O}}) | | \\times | | \\mathcal {T} (\\mathbf {I m g} ^ {\\mathrm {t r a i n}}) | |} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 314, + 785, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Through editing-shift matching, our model can better comprehend how editing should be done within a visual context level through an implicit editing shift value. Furthermore, this implicit editing shift value can further guide the sampling process creating a controllable editing.", + "bbox": [ + 214, + 359, + 782, + 420 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Language Instruction Unification", + "text_level": 1, + "bbox": [ + 215, + 441, + 532, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Previous works that utilize visual prompting in image editing tend to focus more on visual instructions and believe language instructions can be vague and unstable [38]. We believe that the potential of text prompts is not fully explored in image editing. Language instructions have significant impacts on diffusion model outputs. Language instructions with the same meaning can still result in entirely different outputs due to different processed language embeddings.", + "bbox": [ + 212, + 467, + 784, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To improve image editing and explore language instructions, we propose a novel approach, language instruction unification. During the training process, for each batch of training data, we randomly sample $50\\%$ of training data, collect their language editing instructions $l$ , and process them through a frozen lightweight LLM, Open Llama 3b V2 Quant 4 [13] denoted as $\\mathcal{U}$ . The LLM $\\mathcal{U}$ is fixed prompted with a fixed random seed to uniformly reformulate the language instruction $l$ better for machine-level understanding. The LLM $\\mathcal{U}$ will output a unified language editing instruction $l'$ .", + "bbox": [ + 212, + 559, + 785, + 679 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nl ^ {\\prime} = \\mathcal {U} (l) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 470, + 691, + 784, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We then augment the training data with unified language editing instructions.", + "bbox": [ + 238, + 719, + 782, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During the inference, each language editing instruction $l$ is passed through the frozen LLM $\\mathcal{U}$ for language instruction unification and then sent to our conditioned diffusion model.", + "bbox": [ + 212, + 734, + 782, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By adopting language instruction unification for training data augmentation during the training, our conditioned diffusion model can learn diverse non-uniformed editing instructions to build up the model's knowledge distribution with unified language prompts. Adopting language instruction unification during", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Meng and C. Yang et al.", + "bbox": [ + 271, + 114, + 454, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "inference aligns with the training, therefore, greatly minimizing the possibility of diverse quality in edited outputs and maximizing the ability to generalize to unseen editing tasks.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Selective Area Matching", + "text_level": 1, + "bbox": [ + 214, + 209, + 464, + 224 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Diffusion-based image editing models usually suffer from low quality in specific details and this bottleneck appears to be more crucial in generalizable image editing methods [30, 38]. The details of human and animal images are easily distorted in these methods.", + "bbox": [ + 212, + 229, + 782, + 289 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A naive solution might be utilizing negative prompts [34] like 'do not draw hands with 6 fingers' in general text-to-image tasks. However, in image editing, it is challenging to apply negative prompts. Users typically can not foresee the specific details after-edited outputs, therefore they are not able to construct appropriate negative prompts. Besides, negative prompts may limit the artistic potential of image editing models.", + "bbox": [ + 212, + 290, + 784, + 381 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To properly address this issue for generalizable image editorings, we propose an optimization method, namely selective area matching, that targets the difference in the chosen detailed editing area between the original training ground truth $\\mathbf{Img}^{\\mathrm{train}}$ and the reversed pseudo output $\\mathbf{Img}^{\\mathrm{PO}}$ .", + "bbox": [ + 212, + 381, + 784, + 441 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In particular, during the training process, we incorporate a frozen Mask2Former model [6] $\\mathcal{M}$ to obtain panoptic segmented training image information including segmented masks and class information $C$ .", + "bbox": [ + 212, + 441, + 797, + 486 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s e g m a s k}, C = \\mathcal {M} (\\mathbf {I m g} ^ {\\text {t r a i n}}) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 493, + 784, + 508 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "After that, our framework processes the class information using the same lightweight LLM $\\mathcal{U}$ described in Sec. 4.2 to filter out pre-defined classes including living creatures and humans requiring special attention for addressing the details.", + "bbox": [ + 214, + 515, + 803, + 559 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nC _ {\\text {f i l t e r e d}} = \\left\\{c \\in C | \\mathcal {U} (c, \\text {s e l e c t e d c l a s s e s}) = 1 \\right\\} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 566, + 784, + 583 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Based on selected class labels, the framework then deduces a segmented binary mask for the selected editing area.", + "bbox": [ + 214, + 588, + 782, + 619 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n[ \\text {m a s k} ] _ {i} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f c l s} ([ \\text {s e g m a s k} ] _ {i}) \\in C _ {\\text {f i l t e r e d}} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 625, + 784, + 664 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "During the training process, our framework calculates and optimizes the selective area matching loss by", + "bbox": [ + 214, + 667, + 782, + 698 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s a m} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\left([ \\mathbf {I m g} ^ {\\mathrm {P O}} ] _ {i} \\cdot [ \\operatorname {m a s k} ] _ {i}\\right) - \\left(\\left[ \\mathbf {I m g} ^ {\\mathrm {t r a i n}} \\right] _ {i} \\cdot [ \\operatorname {m a s k} ] _ {i}\\right)\\right) ^ {2} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 282, + 704, + 784, + 744 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $N$ as the total pixel number in the image.", + "bbox": [ + 214, + 750, + 563, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "With selective area matching, image editing does not need negative prompts to deal with distorted details in images, which can make the most of the model's artistic editing capacity to generate high-quality outputs with great details. It is only incorporated during training, which does not change the inference, greatly saving inference efforts compared with negative prompts.", + "bbox": [ + 212, + 765, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "InstructGIE: Towards Generalizable Image Editing", + "bbox": [ + 390, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a76af07fe12c29cb021ee74e300bb7f75a2bdb619b59fe337e94dd6b515d5e4a.jpg", + "image_caption": [ + "Phase 1: Data Generation" + ], + "image_footnote": [], + "bbox": [ + 218, + 157, + 625, + 286 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/9fdfd78cd360bb8a5ac837a7ccc9cfa5b17a5022efc5241c53c6d155295e9246.jpg", + "image_caption": [ + "Phase 2: Data Processing", + "Fig. 4: Dataset generation process Our dataset generation consists of two phases. Data Generation is to generate sets of image pairs under one editing caption. Data Processing is randomly pick image pairs under the same editing instruction and concatenate them together as one input for training." + ], + "image_footnote": [], + "bbox": [ + 633, + 159, + 785, + 284 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Visual Prompt Dataset For Image Editing", + "text_level": 1, + "bbox": [ + 215, + 378, + 658, + 397 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Traditional image editing datasets only contain editing image pairs with a small amount of similar editing instructions. To the best of our knowledge, there is no open-sourced image editing dataset that is explicitly designed for image editing with visual prompting, which utilizes multiple different image pairs for each editing instruction to provide a general demonstration in various cases.", + "bbox": [ + 212, + 407, + 784, + 482 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Therefore, we introduce and open source a new dataset that is designed specifically for image editing tasks utilizing visual prompts. Our dataset generation pipeline is shown in Fig. 4 with 2 phases: Data Generation and Data Processing. In Data Generation phase, we first fine-tune GPT-3 [4] for 2 epochs with 700 groups of human-written edits, each consisting of 5 different pairs of input and edited captions with one editing instruction. Then as shown in step 1, we generate around 3500 groups of editing prompts using the fine-tuned GPT-3 model. In each group, there is one instruction $E_0$ and five pairs of caption and edited caption $(I_0,O_0), (I_1,O_1), \\dots, (I_4,O_4)$ . In step 2, similar to InstructPix2Pix, we then also adopt Prompt-to-Prompt for image generation. For each input and edited caption pair, we generated 50 times each with random noise and followed InstructPix2Pix to keep the best one image pair using CLIP-based metric. In addition, we also make sure for each editing instruction, there are at least two pairs of images. In step 3, we generate more image pair sets using in the same way. With filtering, we obtained around 12,000 images with around 2,000 editing instructions. In data preparing, we randomly choose 2 pairs of images under the same editing instruction and concatenate them for training.", + "bbox": [ + 212, + 483, + 787, + 739 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6 Experiments", + "text_level": 1, + "bbox": [ + 215, + 760, + 375, + 777 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 215, + 787, + 444, + 803 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. To fairly compare our methods with baseline methods, we conduct both qualitative and quantitative experiments on our proposed synthetic dataset", + "bbox": [ + 214, + 809, + 784, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Meng and C. Yang et al.", + "bbox": [ + 269, + 114, + 454, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/601e3632b1419bd570fdf82e47fff0a416040136cb19593fda7313cf7acdae4f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 370, + 219 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8105753172b4fa12fa52f152417a3609475d28b35f69580b83c0fc45d2ec1a57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 143, + 452, + 218 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8f610fcb1aa24437e9cf97a63f06059e6f6b520a00368dd3c7eb9e4639059236.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 220, + 454, + 290 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/252f52df050f96a72baa362cc8403ac3c2d91102cedd4c00333411ed6a20866a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 143, + 537, + 218 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b1cc42a4f3bb0deeaef273bd8d2ae025335e3aeef4af21dc6133bcdf75dd31c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 220, + 537, + 290 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7596bbcd456e64506fbf9988e26a2fad4750635029642a6bbbcba3efa9b137fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 143, + 620, + 218 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/19932ac3b1098543dbf8c6eae2e2f6fe8bd01d7d57c9009e2efa8074b3eb8ba9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 220, + 620, + 290 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2f21625a226f9308eefa402e70bdd9c1e44903c9e46e93df9b74711a3a6dcc8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 143, + 702, + 218 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/00289cd57fb4612c9d0d743ee8443acd4ba57f001ad33cf587add54f890a73bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 220, + 702, + 290 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/03d83ccabe08a072f88ccd06b15dde5330aa25bfd0bd1c821cbd714c68329d88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 143, + 785, + 218 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e682fc074c481e8a2f7286e53702f81b3759c256a3ad0d7068cd9e2f9a692b34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 220, + 785, + 290 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c228565a9f3052fcafaeabc8a841ce408e879f8d87bad1909d0acf7d5987cab1.jpg", + "image_caption": [ + "Crown Him Like A King", + "As a Precious Oil Painting" + ], + "image_footnote": [], + "bbox": [ + 217, + 220, + 370, + 290 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5ffbb102262a4b334bd79d909983d6f262d5bdb4e9f085fbcb5ccf048aafd65a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 228, + 454, + 290 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a312e3dc3cdf2ca508f5ac46676ee92d8a5fdbe7ae34a26f9397a4e441fb9625.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 301, + 454, + 362 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f154617d457968c5fd517493e62732ce0fab753c06c80f38ab7f5c52b96c7f62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 301, + 537, + 362 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ea110f8ab3669df4fcba8d9970e1aaac3ac1f3c7651a38e3a0e2407806ec3baf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 301, + 620, + 362 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2f3cee44554f9e3944a382f77df85585738ffb3deeb196a2df1c1b49421b887f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 301, + 702, + 362 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2303655566728338af3498715949bcab20ee80be9a8f90e506bfe44ac8f31ef1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 301, + 785, + 362 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8412a367a8ae7966e67454a42d393586e4528eca1bd62df9205e660f8d8cc20b.jpg", + "image_caption": [ + "Change to Pink Hair" + ], + "image_footnote": [], + "bbox": [ + 215, + 301, + 370, + 362 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/092c5e22587610c44d5779aeeddae1244c84b0ad345d29f1c447a3f8cf21371b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 375, + 454, + 434 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "", + "bbox": [ + 375, + 446, + 454, + 506 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1a0318c5289ee24aaaa4920ae838cbb51c40a1d2eac342b2a8ed14d2e4ef27ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 375, + 537, + 434 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9d9dfe6188c2416cdb8c88b65ad09bb43e0eaf604c449feba991f0e2f2e67af7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 375, + 620, + 434 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/87f939a61b87c9cbb85a10525bc3d9695301b2a98ac5ef9fe98ec289ae45d322.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 375, + 702, + 434 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/55ca2370efb06ed65873cdc9d01001b5e10be8cfb11fa97c51a41b9df7414233.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 375, + 785, + 434 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7a96eb3508cc78e58a6d0abf721b0b0989b598107e5e8a6313ae0b137df89f63.jpg", + "image_caption": [ + "Light Them On Devastating Fire!" + ], + "image_footnote": [], + "bbox": [ + 215, + 375, + 370, + 434 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ce1808793784b27066bded330f2aff2db3c4677d10a30c13bbb7402d63186cdb.jpg", + "image_caption": [ + "", + "Fig. 5: Qualitative Comparison on our Test Dataset. We conducted experiments on various scenarios, including human, architecture and landscape." + ], + "image_footnote": [], + "bbox": [ + 374, + 446, + 454, + 506 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6ca9d4c176cb6f732c529c1a1bc83f78f3dcd0b851847e57ae8ea5815d6beda8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 446, + 537, + 506 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/844b5f9f0edecc5f02b4b1aa9cc423130025b98c8da87b609b77608f7208a46b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 446, + 620, + 506 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/dd0bdf10146d061430688e52f31198e1f698449455b12aadd213c4d6d5876f91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 446, + 702, + 506 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e7e116bea32dfc9e89eed9714def87901d951f4f346d784696af79b2eef56644.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 446, + 785, + 506 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "that involves over 12,000 filtered images and 2,000 filtered editing instructions. All single input images have a resolution of $512 \\times 512$ and are resized to $256 \\times 256$ for training and testing purposes.", + "bbox": [ + 212, + 594, + 785, + 640 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details. In our approach, we split the dataset with $80\\%$ for training and $20\\%$ for testing. As demonstrated in Fig. 2, two image pairs are concatenated into one $512 \\times 512$ image with the same editing instructions as $\\mathbf{Img}^{\\mathrm{train}}$ and mask the fourth quadrant with a grey color as $\\mathbf{Img}^{\\mathrm{VPcon}}$ . We prepare the in-domain test dataset in the same format. For out-of-domain testings, we ensure that both the visual instruction pairs and the text instructions are not used during training, to best simulate how models perform in real-life image editing generalization scenarios.", + "bbox": [ + 212, + 641, + 787, + 762 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For baselines, we use their original configurations to train their model. In our method, we only fine-tune the additional ControlNet for 5000 steps with a batch size of 1024 and a learning rate at $1 \\times 10^{-4}$ . During training, we set the classifier-free scale the same as the original ControlNet. And we randomly drop $15\\%$ of language or visual editing instructions to further enhance our model's", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "InstructGIE: Towards Generalizable Image Editing", + "bbox": [ + 390, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/dc87bfcc57517dda4642a7636981c23cd71c7c5419abf582b7ac8043d2c72da9.jpg", + "table_caption": [ + "Table 1: Quantitative results comparison between our method and baseline methods. Quantitative results show our method excels the baseline methods in both FID and CLIP directional Similarity score in a great margin." + ], + "table_footnote": [], + "table_body": "
SDEdit (E)InstructPix2PixPromptDiffusionOurs
FID ↓21.6717.8713.757.57
CLIP DirSim ↑0.110.170.210.27
", + "bbox": [ + 218, + 198, + 782, + 256 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "generalization ability. Our implementation utilizes PyTorch and is trained on 4 Tesla A100-40G GPUs with AdamW optimizer.", + "bbox": [ + 215, + 285, + 782, + 315 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison Methods. To evaluate the effectiveness of our work, we compare with other state-of-the-art image editing frameworks, including SDEdit [28], Instruct-Pix2pix [3] and PromptDiffusion [41]. We adopt two quantitative metrics Fr'echet Inception Distance (FID) and CLIP directional Similarity (CLIP DirSim) proposed by Gal et al. [12]. We utilize the FID score to quantitatively represent the image quality of generated editing outputs, and CLIP DirSim to evaluate how well the models follow editing instructions to produce the output.", + "bbox": [ + 215, + 316, + 784, + 421 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.2 State-of-the-Art Comparisons", + "text_level": 1, + "bbox": [ + 215, + 446, + 504, + 463 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative Evaluation. In Fig. 5, we present our qualitative results in the testing set (in domain). The comparison shows that our method surpass previous baseline methods. Our method understands and follows both visual and language editing instructions better, and produces a far more detailed manipulated output especially in human figures. In Fig. 6, we present our qualitative results tested in out-of-domain settings. We include five editing instructions that are considered extremely hard for diffusion-based image editing model [18], including object add, object remove, structure location change, structure action change, and object size change. It is important to note that due to how we generate our training dataset utilizing Prompt-to-Prompt [15], editing images pairs with these types of editing instructions is not feasible to generate our training data. This generalization comparison shows that our method excels baseline methods by a significant margin. Our method shows a great capability to carry out well-detailed quality outputs following these hard editing instructions in diffusion-based image editing models, while other baseline methods all fail to understand the editing instructions well or perform manipulations close to the editing instructions.", + "bbox": [ + 215, + 476, + 785, + 717 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Quantitative Evaluation. In quantitative evaluation, we score 7.57 in FID, better than SDEdit (E), InstructPix2Pix and PromptDiffusion which scores 21.67, 17.87 and 13.75. We achieve 0.27 in CLIP DirSim, better than baselines with $0.11 / 0.17 / 0.21$ of CLIP DirSim scores. These quantitative findings show that our method generates higher-quality images with better detailed qualities and also exactly follows both language and visual editing instructions.", + "bbox": [ + 215, + 718, + 785, + 808 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation study. We conduct an ablation study on each of the four components of our proposed method. Namely, the Reformed Conditioned Latent Diffusion", + "bbox": [ + 215, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Meng and C. Yang et al.", + "bbox": [ + 271, + 114, + 454, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9b1a8b3f5261250f98c783b25eccfbe31d5fd0bfed21837faff2b3367ba6b160.jpg", + "image_caption": [ + "Add Some Snow" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 370, + 218 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/99d1dcfd357672b24d81322483cf5088e5b9be75663d07fdafa2f3e8fbfdd6fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 143, + 454, + 218 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3756a06f72ab93543c972051d6c4be5b0d870abd68255c9bc54205cf688411b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 143, + 537, + 217 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d833da5ab0c4172b98c118ff38364ef7549add0cfe8023f0f8b76fd839b5c085.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 143, + 620, + 217 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/797926c8e1b5c902225aaae83f05ec2f05c95c671c2deec57e67316b32595c79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 143, + 702, + 217 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ab4878c744f07d4e4f4a64fbbd2cfa94e27c40ee5904c2d2dd40db2166e77c14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 143, + 785, + 218 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e1652f09b079c909f62d270f254730f3e158f4f97e83b6d35335c7bb168c829a.jpg", + "image_caption": [ + "Bend His Arm" + ], + "image_footnote": [], + "bbox": [ + 217, + 229, + 370, + 290 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/012f19323ab9e590912012e26ff71cd5d12c176bd3fcba6836f034c062a68d62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 229, + 454, + 290 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9560d3f1592a7e353ba4de48a8fce01dfa0f59ec7242998a42470e332306f668.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 229, + 537, + 290 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d5fea7382be0670f5660fd1e2806dd1d1d3f8fdb86a6816c4f5a0e2baa929e69.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 229, + 620, + 290 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b51ef89bf099d75b3744f636e901729534c40b4f0184e0efc185c7fad304e046.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 229, + 702, + 290 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/eeab39df57bc1ebe4fe2ad600987a9b9bce9d0cdf112ec165e286c50d5c8efdb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 229, + 785, + 290 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e03509a783da7fb55ad8ac482e7831e22ca4a1a857d22dcbd6da06e8cfc56151.jpg", + "image_caption": [ + "Move It to the Right" + ], + "image_footnote": [], + "bbox": [ + 217, + 301, + 370, + 362 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/33096d3c36a7c2fd9f30fbf9d6c2cf4aa7e5cc07ab2a481ec8df1430101aa095.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 301, + 454, + 362 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/943b5fc4af426ec1e5392f045af78087f90ace3b5ecfa58fbda23027e81ca4f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 301, + 537, + 362 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1d708ea75de81e7af6d8087b27c23007baa369a05ce3763d82b503b50ce1f0c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 301, + 620, + 362 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/13c3b8dd5458aec95a6b141b0547b69a115571e79ecb8033179f24eae69bc8ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 301, + 702, + 362 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d8bdd96b446ebe407c0a50ae9a1926dd702a77746fac67ddd6414425054c471f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 301, + 785, + 362 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/32d7e0370e93e733f256b94210909ca4de698e8842d7510fb2f4709c23877235.jpg", + "image_caption": [ + "Keep Only One" + ], + "image_footnote": [], + "bbox": [ + 217, + 375, + 370, + 434 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/45fe24e1b3caf71aeb773019b4b125c7bec1877a3b12cc2f8120e6af8c041857.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 375, + 454, + 434 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/28d0d92e67a7f47c855c1f7a1b861c22c812911bcaf11ae3594a6ecb1d95737e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 375, + 537, + 434 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/95e274f915cfadaf716efec172f2fc007ab44bff7ae0a46e017af779a75cec9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 375, + 620, + 434 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3baac35ba9da06666f5181ae26f7523aefd9c019601c15a98c22b1d0d9c68047.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 375, + 702, + 434 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/061f0f591285299418ae2abf9ee73ef0c8260f77ddfac0dcce9a3ed7f0912d14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 375, + 785, + 434 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1fbff72ce8e03f5cf2beb0e0d2dc12df8f4e820029f2b9f7b312982eab92ad1e.jpg", + "image_caption": [ + "Make It Smaller" + ], + "image_footnote": [], + "bbox": [ + 217, + 446, + 370, + 506 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e765032929ade0c0ad1167c5c5936f9dd7464708df3db3ad6086ea0c3a6d8a36.jpg", + "image_caption": [ + "Fig. 6: Qualitative Comparison on Out-of-Domain Images. We conducted experiments on images and instruct that are not in training dataset." + ], + "image_footnote": [], + "bbox": [ + 374, + 446, + 454, + 506 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ac231f8e32ab2c7cff80dbaf39eb6dc5b6ed2fadca375e354fc56156c29b0e3c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 445, + 537, + 506 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6198413c95eeff30c7552d7b8246dea47d0370a4b84e493102d83b0637063f29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 446, + 620, + 506 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1732f0685612529675eed4fea1e2ac58d6e3c9a2330cc6e2583bb156b79c98f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 446, + 702, + 506 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0c69def8584cf15bfda75e026975622d217c5f78a330ab3a73d694c1ab8180c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 446, + 785, + 506 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Model (RCLDM), Editing Shift Matching (ESM), Language Instruction Unification (LIU) and Selective Area Matching (SAM). We present the qualitative results in Fig. 7. From the qualitative results, we can see that without RCLDM and ESM, the model understands the visual editing instructions much weaker, especially in out-of-domain editing. Without LIU, for two language editing instructions with the same meaning, the model produces two output edited images in different detail and quality. This difference in quality tends to increase in out-of-domain settings. Without SAM, the details of the human face are distorted making the model more vulnerable to producing outputs with worse detailed qualities. We also include the qualitative ablation results on the testing dataset in Tab. 2. From observation, our method performs the best when incorporating all four components. Without SAM or LIU, the FID score increases, meaning those modules enhance the detail quality of the output generated. Without CLDM or ESM, the CLIP DirSim score decreases, showing that those two modules contribute to a better understanding in both language and visual level.", + "bbox": [ + 212, + 612, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "InstructGIE: Towards Generalizable Image Editing", + "bbox": [ + 390, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/082dfed84ce85344d81271af58b8bd5ba51fb090989a06b30344d5a06fa7e781.jpg", + "image_caption": [ + "Make It in Winter Time" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 354, + 210 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4516c3aa8231a82b3e917500fa4ce33387ce852b5d0f995ad628c7cb48115536.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 356, + 143, + 424, + 209 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2ac1ede521065bc6be8185c74650c1a75a75254c1f8064e859553d4c24291b6a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 428, + 143, + 496, + 209 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/dd533610c6ab99385d42506ad2d07b70640f702702acf25cd3a8bd3a5ce910e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 143, + 568, + 209 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ab2f1684e0015fab2c381669c266a6fe00fd3c5a7c04264473c307d94d4c0588.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 143, + 640, + 209 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/be79d7f66060746e7965bfa2e9c8316ca7f0a44f78dc9ba448eefa813a8f19f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 143, + 712, + 209 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b7a9092fa842d1e60c4a0c1be6e98baa5698343b94c1f64ee0f3cea0fd492030.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 143, + 787, + 209 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/dd15423423a2b17de3a6bf03986dc5b55de89166d1cc3ba1599e4901ba993c06.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 218, + 352, + 271 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/deffb905bdb4c184a2a5d6b604bea4a756498e18cb8a53d10a6e32ea5ef4554e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 356, + 218, + 424, + 271 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/55ba8916f9ea4c0cf3c66de6e0083d2f77ff953772a5d41e667c9e2ef34a39c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 428, + 218, + 496, + 271 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/365374894182f105968bc6bd785f343786d8feea6c2b74abe279dd7f7bca8d0c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 218, + 568, + 271 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d16b8cc6e61d2eb5d8e4aec4ec07b31e647678fdda58c1c50509067fa17f3ccb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 218, + 640, + 271 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/14eaf355a89442b992c7ea979fb63fb0b087181a36b25b608fc0c1df99da0c53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 218, + 712, + 271 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/aa8ee905cddf6ff10dfe10116542105c63737916cf609f7d488c63a52ba0ea65.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 218, + 784, + 271 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/af4b6e6cae789bbe9d49613b9f3f3a246c0be79aeffb448fb5cef5823190b852.jpg", + "image_caption": [ + "Turn It to Winter Time", + "Make Emotion Cuter" + ], + "image_footnote": [], + "bbox": [ + 217, + 281, + 352, + 334 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c66ae89a6a8613b54338b96d6e606533f6b8bcf44c27a05754b430e64e9f5707.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 356, + 281, + 424, + 334 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d21adcd1f3ce5b6827141d0ee2983d49cf81ed28c9fd5e8167a645d93af71d19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 428, + 281, + 496, + 334 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/32a37f0cd5345c4398dbc102a13cd0f7acfd4d7e9c3c6bf353a687055fccda08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 281, + 568, + 334 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/6c0dd084fe17ee444ade75f0942d8c7cfa8ffc09a254c6e3d21a31efe6592fb2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 281, + 640, + 334 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/6ac2f3fe89e5754757a57deb6cb3e5f3483ce585f81633c05d512db48f99b5dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 281, + 712, + 334 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/285f2962961afa9dcb0489d9d636db5f96227bfd7070e70f61992e38fe29d0ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 281, + 784, + 334 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/16b6014d961b8b145318bf832d3f61d44de62776020b5c0e888d59fda80e9363.jpg", + "image_caption": [ + "Cuter Emotion", + "Fig. 7: Ablation Study Results for both in-domain (first two rows), and out-of-domain first two rows) image manipulations" + ], + "image_footnote": [], + "bbox": [ + 215, + 344, + 352, + 397 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9956882be333de8fc4bcd4d056c7b61fa64ff3941c5c46d83d214c8c86f52d41.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 356, + 344, + 424, + 397 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1e38f74099ac8b8b755e791e0db43a3e6895b754b502306f5e6e1c993194d62a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 428, + 344, + 496, + 397 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c88305bdd51f8c03e042473ad850b1c3b01e822f218c3f985a850716ee503d09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 344, + 568, + 397 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5e809d829b6b4eb4d973f8ba5ee685c671e5082ac363545ca39790ba525a7eef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 344, + 640, + 397 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0b45f0363aa469f07c35def0283f7e48e83d4046adddca4baab0535749779c79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 344, + 712, + 397 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/924aa9236d63901c0339c7872f03f266a6e61668d908244cd7d8981898549db4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 344, + 784, + 397 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/678c636c857138e1f9b75fc28dc91e8e1e67b970fc974d97ccfa2752bdacc0c9.jpg", + "table_caption": [ + "Table 2: Abalation Study results of comparison between our method with all components and without the Reformed Conditioned Latent Diffusion Model (RCLDM), Editing Shift Matching (ESM), Language Instruction Unification (LIU) and Selective Area Matching (SAM). The ablation study is conducted on the entire test dataset." + ], + "table_footnote": [], + "table_body": "
w/o RCLDMw/o ESMw/o LIUw/o SAMOurs
FID ↓10.159.2310.3711.317.57
CLIP DirSim ↑0.130.150.230.190.27
", + "bbox": [ + 218, + 544, + 782, + 601 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 636, + 357, + 651 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we propose InstructGIE, an image editing framework with enhanced generalization ability, improving performance in both visual and text aspects. We incorporate a VMamba-based module to enhance visual outputs and introduce an editing-shift matching strategy to augment in-context learning. Our selective area-matching technique addresses and rectifies corrupted details, while a language unification technique aligns language embeddings with editing semantics. Additionally, we compile a publicly available dataset for better generalization evaluation. Extensive experiments demonstrate our framework's superior in-context generation performance and robust generalization capability across unseen vision tasks, both quantitative and qualitively.", + "bbox": [ + 212, + 688, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Meng and C. Yang et al.", + "bbox": [ + 271, + 114, + 454, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is partially supported by the Army Research Office/Army Research Laboratory via grant W911-NF-20-1-0167 to Northeastern University, National Science Foundation CCF-1937500, and the Fundamental Research Funds for the Central Universities, Peking University.", + "bbox": [ + 212, + 175, + 785, + 236 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 258, + 321, + 273 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bar, A., Gandelsman, Y., Darrell, T., Globerson, A., Efros, A.: Visual prompting via image inpainting. Advances in Neural Information Processing Systems 35, 25005-25017 (2022)", + "2. Blau, T., Ganz, R., Kawar, B., Bronstein, A., Elad, M.: Threat model-agnostic adversarial defense using diffusion models. arXiv preprint arXiv:2207.08089 (2022)", + "3. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)", + "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)", + "5. Chen, W., Hu, H., Sahara, C., Cohen, W.W.: Re-imagen: Retrieval-augmented text-to-image generator. arXiv preprint arXiv:2209.14491 (2022)", + "6. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290–1299 (2022)", + "7. Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8188-8197 (2020)", + "8. Chung, J.S., Zisserman, A.: Lip reading in the wild. In: Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13. pp. 87-103. Springer (2017)", + "9. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017)", + "0. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)", + "11. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021)", + "12. Gal, R., Patashnik, O., Maron, H., Chechik, G., Cohen-Or, D.: Stylegan-nada: Clipguided domain adaptation of image generators. arXiv preprint arXiv:2108.00946 (2021)", + "13. Geng, X., Liu, H.: Openllama: An open reproduction of llama (May 2023), https://github.com/openlm-research/open_llama", + "14. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022)" + ], + "bbox": [ + 218, + 287, + 784, + 838 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "InstructGIE: Towards Generalizable Image Editing", + "bbox": [ + 390, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022)", + "16. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)", + "17. Huang, Y., Huang, J., Liu, Y., Yan, M., Lv, J., Liu, J., Xiong, W., Zhang, H., Chen, S., Cao, L.: Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525 (2024)", + "18. Huang, Y., Huang, J., Liu, Y., Yan, M., Lv, J., Liu, J., Xiong, W., Zhang, H., Chen, S., Cao, L.: Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525 (2024)", + "19. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019)", + "20. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. Advances in Neural Information Processing Systems 35, 23593-23606 (2022)", + "21. Kawar, B., Ganz, R., Elad, M.: Enhancing diffusion-based image synthesis with robust classifier guidance. arXiv preprint arXiv:2208.08664 (2022)", + "22. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)", + "23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Pereira, F., Burges, C., Bottou, L., Weinberger, K. (eds.) Advances in Neural Information Processing Systems. vol. 25. Curran Associates, Inc. (2012)", + "24. Kwon, G., Ye, J.C.: Diffusion-based image translation using disentangled style and content representation. arXiv preprint arXiv:2209.15264 (2022)", + "25. Liu, X., Park, D.H., Azadi, S., Zhang, G., Chopikyan, A., Hu, Y., Shi, H., Rohrbach, A., Darrell, T.: More control for free! image synthesis with semantic diffusion guidance. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 289-299 (2023)", + "26. Liu, Y., Tian, Y., Zhao, Y., Yu, H., Xie, L., Wang, Y., Ye, Q., Liu, Y.: Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166 (2024)", + "27. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of International Conference on Computer Vision (ICCV) (December 2015)", + "28. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073 (2021)", + "29. Mohammad, S., Kiritchenko, S.: Wikiart emotions: An annotated dataset of emotions evoked by art. In: Proceedings of the eleventh international conference on language resources and evaluation (LREC 2018) (2018)", + "30. Nguyen, T., Li, Y., Ojha, U., Lee, Y.J.: Visual instruction inversion: Image editing via visual prompting. arXiv preprint arXiv:2307.14331 (2023)", + "31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Meng and C. Yang et al.", + "bbox": [ + 271, + 114, + 454, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "33. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)", + "34. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "35. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)", + "36. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems 32 (2019)", + "37. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)", + "38. Sun, Y., Yang, Y., Peng, H., Shen, Y., Yang, Y., Hu, H., Qiu, L., Koike, H.: Imagebrush: Learning visual in-context instructions for exemplar-based image manipulation. arXiv preprint arXiv:2308.00906 (2023)", + "39. Theis, L., Salimans, T., Hoffman, M.D., Mentzer, F.: Lossy compression with gaussian diffusion. arXiv preprint arXiv:2206.08889 (2022)", + "40. Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6830-6839 (2023)", + "41. Wang, Z., Jiang, Y., Lu, Y., He, P., Chen, W., Wang, Z., Zhou, M., et al.: Incontext learning unlocked for diffusion models. Advances in Neural Information Processing Systems 36 (2024)", + "42. Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., Xiao, J.: Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365 (2015)", + "43. Zablotskaia, P., Siarohin, A., Zhao, B., Sigal, L.: Dwnet: Dense warp-based network for pose-guided human video generation. arXiv preprint arXiv:1910.09139 (2019)", + "44. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023)", + "45. Zimmermann, R.S., Schott, L., Song, Y., Dunn, B.A., Klindt, D.A.: Score-based generative classifiers. arXiv preprint arXiv:2110.00473 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 686 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "InstructGIE: Towards Generalizable Image Editing", + "bbox": [ + 390, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/InstructGIE_ Towards Generalizable Image Editing/d8d7c0e8-e8d9-4aa1-9b02-c4948b5c2541_model.json b/2024/InstructGIE_ Towards Generalizable Image Editing/d8d7c0e8-e8d9-4aa1-9b02-c4948b5c2541_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6b026fdbea2e2cdfb70d49b47504f6b1db007518 --- /dev/null +++ b/2024/InstructGIE_ Towards Generalizable Image Editing/d8d7c0e8-e8d9-4aa1-9b02-c4948b5c2541_model.json @@ -0,0 +1,3391 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.258, + 0.141, + 0.747, + 0.187 + ], + "angle": 0, + "content": "InstructGIE: Towards Generalizable Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.213, + 0.777, + 0.245 + ], + "angle": 0, + "content": "Zichong Meng\\*, Changdi Yang\\*, Jun Liu\\*, Hao Tang\\*\\*, Pu Zhao\\*\\*, and Yanzhi Wang\\*\\*" + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.255, + 0.673, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) Northeastern University, Boston MA 02115, USA" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.27, + 0.654, + 0.284 + ], + "angle": 0, + "content": "2 Peking University, Beijing 100871, China &" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.285, + 0.686, + 0.298 + ], + "angle": 0, + "content": "Carnegie Mellon University, Pittsburgh PA 15213, USA" + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.342, + 0.765, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.647, + 0.295, + 0.654 + ], + "angle": 0, + "content": "Like in Game" + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.647, + 0.55, + 0.654 + ], + "angle": 0, + "content": "Add a Cow" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.67, + 0.788, + 0.713 + ], + "angle": 0, + "content": "Fig. 1: Demo results of the proposed InstructGIE framework on various image manipulation tasks to both humans and scenes. By our proposed method, our model can generalize to generate the desired output with great detail qualities." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.772, + 0.741, + 0.801 + ], + "angle": 0, + "content": "Abstract. Recent advances in image editing have been driven by the development of denoising diffusion models, marking a significant leap" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.812, + 0.365, + 0.826 + ], + "angle": 0, + "content": "* Equal Contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.217, + 0.827, + 0.392, + 0.84 + ], + "angle": 0, + "content": "\\*\\* Corresponding Authors" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.812, + 0.392, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.129 + ], + "angle": 0, + "content": "Z. Meng and C. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.148, + 0.74, + 0.411 + ], + "angle": 0, + "content": "forward in this field. Despite these advances, the generalization capabilities of recent image editing approaches remain constrained. In response to this challenge, our study introduces a novel image editing framework with enhanced generalization robustness by boosting in-context learning capability and unifying language instruction. This framework incorporates a module specifically optimized for image editing tasks, leveraging the VMamba block and an editing-shift matching strategy to augment in-context learning. Furthermore, we unveil a selective area-matching technique specifically engineered to address and rectify corrupted details in generated images, such as human facial features, to further improve the quality. Another key innovation of our approach is the integration of a language unification technique, which aligns language embeddings with editing semantics to elevate the quality of image editing. Moreover, we compile the first dataset for image editing with visual prompts and editing instructions that could be used to enhance in-context capability. Trained on this dataset, our methodology not only achieves superior synthesis quality for trained tasks, but also demonstrates robust generalization capability across unseen vision tasks through tailored prompts. Our project page is available at https://cr8br0ze.github.io/InstructGIE." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.426, + 0.715, + 0.44 + ], + "angle": 0, + "content": "Keywords: Image Editing \\(\\cdot\\) In-Context Learning \\(\\cdot\\) Diffusion Model" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.47, + 0.377, + 0.487 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.508, + 0.786, + 0.658 + ], + "angle": 0, + "content": "As a crucial task in computer vision, image editing has witnessed significant improvements enhanced with the increasingly popular denoising stable diffusion techniques in recent years [22,25,38,40,41]. Given a set of text or image prompts as generation constraints or instructions, diffusion-based image editing can follow the instructions and synthesize an edited image. However, since the model does not have the capability to accurately model all possible samples in the conditional distribution space [17], if specific instructions are not included in the training dataset, current diffusion-based image editing methods can hardly generate satisfactory results. Thus, editing performance largely depends on the training dataset without superior generalization capabilities." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.66, + 0.786, + 0.839 + ], + "angle": 0, + "content": "On the other hand, large language models (LLMs) have proven extraordinary abilities to learn from contexts, referred to as in-context learning, which allows LLMs to perform unseen tasks by providing a combination of input-output examples and a query input. Inspired by the potential to enhance the generalization of the model with LLMs, [38,41] explore in-context learning for computer vision tasks, allowing them to perform unseen tasks with novel vision-language prompt designs. However, these methods are not tailored for image editing applications, leading to unsatisfying synthetic qualities with inaccurate or incorrect output and lack of detail. To improve the generalization of image editing with improved synthetic image quality, it is crucial to effectively understand the text & image prompts and specifically control image editing details, which is challenging in the current literature." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.392, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "InstructGIE: Towards Generalizable Image Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.786, + 0.419 + ], + "angle": 0, + "content": "In this work, we propose InstructGIE, an image editing framework with enhanced generalizability. We improve image editing performance from both visual and text aspects. (i) For the visual information, we incorporate a VMamba-based module to specifically enhance the image editing outputs. As VMamba [26] has proven its better performance in capturing in-context information from inputs with larger receptive fields [26], we leverage VMamba and propose an editing-shift matching strategy to augment in-context learning. Furthermore, since current image editing works do not perform well in generating correct features with accurate details, we unveil a selective area-matching technique specifically engineered to address and rectify corrupted details in generated images, such as human facial features, to further improve the quality. (ii) Another key innovation of our approach is the integration of a language unification technique, which aligns language embeddings with editing semantics to elevate the quality of image editing. Our framework not only achieves superior in-context generation for trained tasks but also demonstrates robust generalization across unseen vision tasks. Moreover, we compile a publicly available image editing dataset with plenty of visual prompts and editing instructions for better generalization evaluation of image editing. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.428, + 0.784, + 0.471 + ], + "angle": 0, + "content": "- We propose an image editing framework, including in-context learning enhancement and language unification strategies, specifically designed to enhance generalization ability from both visual and text domains." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.473, + 0.785, + 0.501 + ], + "angle": 0, + "content": "- We compile the first dataset for image editing with visual prompts and editing instructions that could be used to enhance in-context capability." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.502, + 0.785, + 0.531 + ], + "angle": 0, + "content": "- We conduct extensive experiments and achieve great generalization ability in the multiple unseen image editing task, both quantitatively and qualitatively." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.428, + 0.785, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.562, + 0.396, + 0.577 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.592, + 0.657, + 0.607 + ], + "angle": 0, + "content": "2.1 Denoising Stable Diffusion Based Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.614, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Denoising Stable Diffusion [16,35,36] based image editing could follow guidance from text or image prompts. With the foundation of text-guided models offering rich generative capabilities, there has been a surge in research aimed at adapting these models for image manipulation tasks from textual descriptions. To steer the image editing process in the desired direction, the use of models like CLIP to fine-tune diffusion models has become a common practice. Although these methods [3,22,41] have shown impressive results, they often involve costly finetuning processes. Recent innovations [14] have introduced techniques that inject cross-attention into the models to more effectively edit specific semantic areas within the spatial feature maps. Further advancements [24] have enhanced these techniques by adding semantic loss or applying attention loss to refine the integration of plugged features, improving the precision and quality of the editing outcomes. [38] proposes a framework that could learn instructions from visual image pairs for more accurate editing and firstly formulate this task as an image inpainting problem." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.129 + ], + "angle": 0, + "content": "Z. Meng and C. Yang et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.605, + 0.163 + ], + "angle": 0, + "content": "2.2 Generalization Capability in Visual Tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.369 + ], + "angle": 0, + "content": "In-context learning is widely applied in the field of natural language processing (NLP), enabling models to adapt to new tasks such as translation, question answering, and complex reasoning. NLP models utilize in-context examples, comprising text and corresponding labels, to tackle tasks they haven't seen before. However, applying in-context learning to the visual domain introduces more challenges and remains less explored. A significant hurdle is the nature of fixed-size input requirements for vision models, as opposed to variable-length text inputs that can be managed by language models. Vision models generally struggle with processing inputs of varying sizes, making it impractical to process multiple image prompts in one-shot for global understanding. Moreover, in intricate visual understanding, specific instructions are often implied from a limited set of image examples rather than explicitly stated, which poses additional difficulties for vision models in identifying and understanding high-level visual relationships." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.504 + ], + "angle": 0, + "content": "Recent strides in applying masked image modeling have marked a step forward in improving in-context learning for vision models. The method proposed by [41], employing a masked autoencoder-based technique, predicts a missing image within a two-by-two grid, using two images as in-context examples and another as the query. This concept was later expanded by [38] with a multitask framework. Despite their progress, such inpainting methods are limited by the necessity of a fixed number of in-context examples and increased memory demands. Painter, highlighted in [40], exemplifies an inpainting approach tailored for versatility across various vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.787, + 0.566 + ], + "angle": 0, + "content": "In contrast, inspired by ControlNet [44], [41] refines the framework by adding an additional pair of example images and employing a multitask supervised finetuning method. Prompt diffusion excels in visual in-context learning. However, it faces certain limitations or challenges in its practical applications." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.588, + 0.611, + 0.604 + ], + "angle": 0, + "content": "2.3 Dataset for Diffusion-based Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Currently, various types of datasets are used for training in diffusion-based image editing. There are datasets that concentrate on specific domains like CelebA [27] and FFHQ [19] for human face image manipulation, AFHQ [7] for animal face image editing, LSUN [42] for object modification, and WikiArt [29] for style transfer. In-the-wild video datasets could also be leveraged to train image editing tasks. The Scannet dataset [9] encompasses a vast array of more than 1,500 indoor scenes from various settings, such as apartments, offices, and hotels, providing extensive annotations. The LRW dataset [8], tailored for lip reading tasks, includes more than 1000 video utterances of 500 distinct words. The UBC-Fashion dataset [43] features 600 videos spanning various clothing categories, with 500 videos allocated for training and 100 for testing, guaranteeing no repetition of individuals in the training set. The DAVIS dataset [43] (Densely Annotated VVideo Segmentation), a widely recognized benchmark for video object segmentation, contains 150 videos in total. There are also image editing works proposing to generate datasets with editing instructions. InstructPix2pix [3] collects over" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.391, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "InstructGIE: Towards Generalizable Image Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "450,000 training image pairs. For each pair, given an image with its caption, it first uses a finetuned GPT-3 [4] to generate an editing instruction and an edited image caption. Then it employs Stable Diffusion and the Prompt-to-Prompt algorithm [14] to generate edited image following the caption. However, currently there are no datasets with multiple image pairs under one editing instruction, which is crucial to enhance the generalization ability of image editing." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.264, + 0.37, + 0.283 + ], + "angle": 0, + "content": "3 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.789, + 0.435 + ], + "angle": 0, + "content": "Recent advances in generative models have been significantly driven by the emergence of diffusion models, which have set new benchmarks in image creation [10,21,37]. These models have found applications across a broad spectrum of areas [2,5,20,39,45], demonstrating their versatility and effectiveness. The fundamental concept behind diffusion models involves starting with an image that is initially just random noise \\(\\mathbf{x}_T\\sim \\mathcal{N}(0,\\mathbf{I})\\) and progressively refining this image step by step until it becomes a high-quality, realistic image \\(\\mathbf{x}_0\\). This refinement process involves generating intermediate samples \\(\\mathbf{x}_t\\) (for \\(t\\in \\{0,\\dots ,T\\}\\)), where each sample is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.448, + 0.786, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t} = \\sqrt {\\alpha_ {t}} \\mathbf {x} _ {0} + \\sqrt {1 - \\alpha_ {t}} \\boldsymbol {\\epsilon} _ {t}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.789, + 0.583 + ], + "angle": 0, + "content": "where the parameter \\(\\alpha_{t}\\) sets the pace of the diffusion process, ranging from \\(0 = \\alpha_{T} < \\alpha_{T - 1} < \\dots < \\alpha_{1} < \\alpha_{0} = 1\\), and \\(\\epsilon_t\\sim \\mathcal{N}(0,\\mathbf{I})\\) represents the added noise. The model refines the image by applying a neural network \\(f_{\\theta}(\\mathbf{x}_t,t)\\) to each sample \\(\\mathbf{x}_t\\), followed by the addition of Gaussian noise to produce the next sample \\(\\mathbf{x}_{t - 1}\\). This neural network is optimized to achieve a denoising goal, striving for \\(f_{\\theta}(\\mathbf{x}_t,t)\\approx \\epsilon_t\\), resulting in a generative process that closely mimics the desired image distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.735 + ], + "angle": 0, + "content": "Expanding this framework to conditional generative modeling, the process involves conditioning the neural network \\( f_{\\theta}(\\mathbf{x}_t,t,\\mathbf{y}) \\) on an additional input \\( \\mathbf{y} \\), enabling the generation of images from a distribution conditioned on \\( \\mathbf{y} \\). This conditional input could be anything from a low-resolution image, a category label, or a descriptive text sequence. Leveraging the advancements in LLMs [33] and hybrid vision-language models [31], text-to-image diffusion models are developed. These models allow for the creation of detailed, high-resolution images from mere text descriptions, starting with a low-resolution image generated through the diffusion process, which is subsequently refined into a high-resolution image using additional models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.76, + 0.473, + 0.779 + ], + "angle": 0, + "content": "4 The Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.843 + ], + "angle": 0, + "content": "We present our framework pipeline in Fig. 2. For efficient training and better controllability, we adopt a line of techniques with popular architectures such as ControlNet [44] and Stable Diffusion [34] to design a generalizable image editing" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.129 + ], + "angle": 0, + "content": "Z. Meng and C. Yang et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.767, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.538 + ], + "angle": 0, + "content": "Fig. 2: Overall architecture of InstructGIE. The lower pipeline is for both training and inference processes where the model obtains unified editing instructions outputted by Instruction Unification Module \\(\\mathcal{U}\\) and combines with visual prompted input \\(\\mathbf{Img}^{\\mathrm{VPcon}}\\) to pass through Zero-VMamba integrated Stable Diffusion model with ControlNet for output image. The upper pipeline is for training only which compares output image and training ground truth \\(\\mathbf{Img}^{\\mathrm{train}}\\) and computes editing shift loss \\(\\mathcal{L}_{es}\\) with Editing Shift Module and selective area matching loss \\(\\mathcal{L}_{sam}\\) with Selective Area Matching Module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.645 + ], + "angle": 0, + "content": "tool with accurate high-quality outputs. Specifically, we introduce enhanced in-context learning both at the architecture level and training level to improve the image quality. Furthermore, language instruction unification is adopted to maximize the generalization ability for unseen editing tasks. Moreover, selective area matching is proposed to further improve the output quality with full details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.668, + 0.519, + 0.684 + ], + "angle": 0, + "content": "4.1 Enhanced In-Context Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.788, + 0.773 + ], + "angle": 0, + "content": "Visual prompting based on inpainting is an effective visual in-context learning method in various computer vision tasks [1,40], which is applied in image editing tasks [38] recently. However, the methods perform poorly in quality when dealing with unseen image manipulation tasks. Therefore, we propose the enhanced in-context learning specifically tailored for generalizable image editing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Reforming Conditioned Latent Diffusion Model. To improve the generalization ability of image editing, it is crucial for the framework to explicitly capture low-level visual editing contexts. Current diffusion-based image editing" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.392, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "InstructGIE: Towards Generalizable Image Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.144, + 0.411, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.264, + 0.381, + 0.274 + ], + "angle": 0, + "content": "ERF of ResNet-50" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.144, + 0.58, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.457, + 0.264, + 0.546, + 0.274 + ], + "angle": 0, + "content": "ERF of VIT-L/14" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.144, + 0.749, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.264, + 0.72, + 0.274 + ], + "angle": 0, + "content": "ERF of VMamba-T" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.289, + 0.787, + 0.318 + ], + "angle": 0, + "content": "Fig. 3: Effective Reception Field (ERF) of ConvNet, ViT, VMamba based model architectures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.345, + 0.784, + 0.435 + ], + "angle": 0, + "content": "methods [30, 38] that involve visual prompting either adopt ConvNet [23] or ViT [11] as the vision encoder for visual prompts. However, these methods fail to generalize well as they are not able to capture enough visual editing contexts (see Fig. 3). To address this, we formulate the visual prompted condition as a single image \\(\\mathbf{Img}^{\\mathrm{VPcon}} = \\{\\mathbf{Img}_0^{\\mathrm{in}}, \\mathbf{Img}_0^{\\mathrm{out}}, \\mathbf{Img}_1^{\\mathrm{in}}, \\mathbf{Grey}\\}\\), as shown in Fig 2, with a global effective receptive field (ERF)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.785, + 0.527 + ], + "angle": 0, + "content": "Moreover, we further propose to reform the conditioned latent diffusion model Inspired by the recent visual state space model VMamba [26] which exhibits a better global ERF and also emphasizes shifting boundaries of input's four quadrants, we propose to adopt a vision encoder based on Zero-VMamba to fit our structure. Specifically, the vision encoder \\(\\mathcal{G}_{ven}\\) comprehends the visual prompted condition in latent space \\(x_{vpc}\\) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.42, + 0.536, + 0.785, + 0.553 + ], + "angle": 0, + "content": "\\[\ne _ {v p c} = \\mathcal {G} _ {v e n} \\left(x _ {v p c}; \\Theta_ {g}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.56, + 0.784, + 0.59 + ], + "angle": 0, + "content": "where \\( e_{vpc} \\) is the processed embedding of the visual prompted condition, and \\( \\Theta_g \\) is the model parameters initialized to zeros." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.591, + 0.785, + 0.68 + ], + "angle": 0, + "content": "To further improve the performance, after each ControlNet trainable copied modules \\(\\mathcal{F}_c\\) with parameters \\(\\Theta_c\\), we propose to link and inject the processed visual prompted condition information to the frozen Stable Diffusion model \\(\\mathcal{F}\\) with parameters \\(\\Theta\\) through zero-VMamba layer \\(\\mathcal{G}\\). We use two instances of VMamba with parameters \\(\\theta_{g1}\\) and \\(\\theta_{g2}\\) respectively. The complete model then computes the following," + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.69, + 0.785, + 0.708 + ], + "angle": 0, + "content": "\\[\ny _ {v p c} = \\mathcal {F} (x; \\Theta) + \\mathcal {G} (\\mathcal {F} (x + \\mathcal {G} (x _ {v p c}; \\Theta_ {g 1}); \\Theta_ {c}); \\Theta_ {g 2}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.715, + 0.691, + 0.73 + ], + "angle": 0, + "content": "where \\( y_{vpc} \\) is the output of our conditioned diffusion model block." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.73, + 0.784, + 0.775 + ], + "angle": 0, + "content": "Our conditioned latent diffusion model can process all four quadrants in our visual prompted conditions with a global receptive field, while it does not generate random noises during initial training steps." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Editing-Shift Matching. Besides the architecture innovation, we incorporate an editing-shift-matching technique to enhance in-context learning ability in image editing with more accurate detailed outputs." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.129 + ], + "angle": 0, + "content": "Z. Meng and C. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.784, + 0.191 + ], + "angle": 0, + "content": "In specific, for each training ground truth \\(\\mathbf{Img}^{\\mathrm{train}} = \\{\\mathbf{Img}_0^{\\mathrm{in}},\\mathbf{Img}_0^{\\mathrm{out}},\\mathbf{Img}_1^{\\mathrm{in}},\\) \\(\\mathbf{Img}_1^{\\mathrm{out}}\\}\\), we calculate a implicit editing shift value using CLIP [32] image embedding:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.192, + 0.786, + 0.224 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} (\\mathbf {I m g}) = \\frac {\\sum_ {i = 0} ^ {1} \\operatorname {C L I P} \\left(\\mathbf {I m g} _ {i} ^ {\\text {i n}}\\right) - \\operatorname {C L I P} \\left(\\mathbf {I m g} _ {i} ^ {\\text {o u t}}\\right)}{2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.228, + 0.784, + 0.304 + ], + "angle": 0, + "content": "During the training process, after predicting the noise, we use it to reverse the noised input and obtain a pseudo output image \\(\\mathbf{Img}^{\\mathrm{PO}} = \\{\\mathbf{PO}_0^{\\mathrm{in}},\\mathbf{PO}_0^{\\mathrm{out}},\\mathbf{PO}_1^{\\mathrm{in}},\\) \\(\\mathbf{PO}_1^{\\mathrm{out}}\\}\\). Our framework then calculates the editing transfer value of the pseudo output image and deduces a editing shift loss to optimize during our training via the cosine similarity of the two design transfer values:" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.315, + 0.786, + 0.353 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {e s} = 1 - \\frac {\\mathcal {T} (\\mathbf {I m g} ^ {\\mathrm {P O}}) \\cdot \\mathcal {T} (\\mathbf {I m g} ^ {\\mathrm {t r a i n}})}{| | \\mathcal {T} (\\mathbf {I m g} ^ {\\mathrm {P O}}) | | \\times | | \\mathcal {T} (\\mathbf {I m g} ^ {\\mathrm {t r a i n}}) | |} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.36, + 0.784, + 0.421 + ], + "angle": 0, + "content": "Through editing-shift matching, our model can better comprehend how editing should be done within a visual context level through an implicit editing shift value. Furthermore, this implicit editing shift value can further guide the sampling process creating a controllable editing." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.443, + 0.534, + 0.459 + ], + "angle": 0, + "content": "4.2 Language Instruction Unification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.468, + 0.785, + 0.559 + ], + "angle": 0, + "content": "Previous works that utilize visual prompting in image editing tend to focus more on visual instructions and believe language instructions can be vague and unstable [38]. We believe that the potential of text prompts is not fully explored in image editing. Language instructions have significant impacts on diffusion model outputs. Language instructions with the same meaning can still result in entirely different outputs due to different processed language embeddings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.56, + 0.786, + 0.68 + ], + "angle": 0, + "content": "To improve image editing and explore language instructions, we propose a novel approach, language instruction unification. During the training process, for each batch of training data, we randomly sample \\(50\\%\\) of training data, collect their language editing instructions \\(l\\), and process them through a frozen lightweight LLM, Open Llama 3b V2 Quant 4 [13] denoted as \\(\\mathcal{U}\\). The LLM \\(\\mathcal{U}\\) is fixed prompted with a fixed random seed to uniformly reformulate the language instruction \\(l\\) better for machine-level understanding. The LLM \\(\\mathcal{U}\\) will output a unified language editing instruction \\(l'\\)." + }, + { + "type": "equation", + "bbox": [ + 0.471, + 0.693, + 0.785, + 0.708 + ], + "angle": 0, + "content": "\\[\nl ^ {\\prime} = \\mathcal {U} (l) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.72, + 0.784, + 0.735 + ], + "angle": 0, + "content": "We then augment the training data with unified language editing instructions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.784, + 0.779 + ], + "angle": 0, + "content": "During the inference, each language editing instruction \\( l \\) is passed through the frozen LLM \\( \\mathcal{U} \\) for language instruction unification and then sent to our conditioned diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.786, + 0.841 + ], + "angle": 0, + "content": "By adopting language instruction unification for training data augmentation during the training, our conditioned diffusion model can learn diverse non-uniformed editing instructions to build up the model's knowledge distribution with unified language prompts. Adopting language instruction unification during" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.391, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "InstructGIE: Towards Generalizable Image Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "inference aligns with the training, therefore, greatly minimizing the possibility of diverse quality in edited outputs and maximizing the ability to generalize to unseen editing tasks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.21, + 0.465, + 0.226 + ], + "angle": 0, + "content": "4.3 Selective Area Matching" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.23, + 0.784, + 0.29 + ], + "angle": 0, + "content": "Diffusion-based image editing models usually suffer from low quality in specific details and this bottleneck appears to be more crucial in generalizable image editing methods [30, 38]. The details of human and animal images are easily distorted in these methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.291, + 0.785, + 0.382 + ], + "angle": 0, + "content": "A naive solution might be utilizing negative prompts [34] like 'do not draw hands with 6 fingers' in general text-to-image tasks. However, in image editing, it is challenging to apply negative prompts. Users typically can not foresee the specific details after-edited outputs, therefore they are not able to construct appropriate negative prompts. Besides, negative prompts may limit the artistic potential of image editing models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.785, + 0.442 + ], + "angle": 0, + "content": "To properly address this issue for generalizable image editorings, we propose an optimization method, namely selective area matching, that targets the difference in the chosen detailed editing area between the original training ground truth \\(\\mathbf{Img}^{\\mathrm{train}}\\) and the reversed pseudo output \\(\\mathbf{Img}^{\\mathrm{PO}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.798, + 0.487 + ], + "angle": 0, + "content": "In particular, during the training process, we incorporate a frozen Mask2Former model [6] \\(\\mathcal{M}\\) to obtain panoptic segmented training image information including segmented masks and class information \\(C\\)." + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.494, + 0.785, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s e g m a s k}, C = \\mathcal {M} (\\mathbf {I m g} ^ {\\text {t r a i n}}) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.516, + 0.804, + 0.56 + ], + "angle": 0, + "content": "After that, our framework processes the class information using the same lightweight LLM \\(\\mathcal{U}\\) described in Sec. 4.2 to filter out pre-defined classes including living creatures and humans requiring special attention for addressing the details." + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.568, + 0.785, + 0.584 + ], + "angle": 0, + "content": "\\[\nC _ {\\text {f i l t e r e d}} = \\left\\{c \\in C | \\mathcal {U} (c, \\text {s e l e c t e d c l a s s e s}) = 1 \\right\\} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.589, + 0.784, + 0.62 + ], + "angle": 0, + "content": "Based on selected class labels, the framework then deduces a segmented binary mask for the selected editing area." + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.625, + 0.785, + 0.665 + ], + "angle": 0, + "content": "\\[\n[ \\text {m a s k} ] _ {i} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f c l s} ([ \\text {s e g m a s k} ] _ {i}) \\in C _ {\\text {f i l t e r e d}} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.669, + 0.784, + 0.699 + ], + "angle": 0, + "content": "During the training process, our framework calculates and optimizes the selective area matching loss by" + }, + { + "type": "equation", + "bbox": [ + 0.284, + 0.705, + 0.785, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s a m} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\left([ \\mathbf {I m g} ^ {\\mathrm {P O}} ] _ {i} \\cdot [ \\operatorname {m a s k} ] _ {i}\\right) - \\left(\\left[ \\mathbf {I m g} ^ {\\mathrm {t r a i n}} \\right] _ {i} \\cdot [ \\operatorname {m a s k} ] _ {i}\\right)\\right) ^ {2} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.75, + 0.565, + 0.765 + ], + "angle": 0, + "content": "where \\(N\\) as the total pixel number in the image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.784, + 0.84 + ], + "angle": 0, + "content": "With selective area matching, image editing does not need negative prompts to deal with distorted details in images, which can make the most of the model's artistic editing capacity to generate high-quality outputs with great details. It is only incorporated during training, which does not change the inference, greatly saving inference efforts compared with negative prompts." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.455, + 0.129 + ], + "angle": 0, + "content": "Z. Meng and C. Yang et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.363, + 0.146, + 0.471, + 0.155 + ], + "angle": 0, + "content": "Phase 1: Data Generation" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.158, + 0.626, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.147, + 0.765, + 0.157 + ], + "angle": 0, + "content": "Phase 2: Data Processing" + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.16, + 0.787, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.299, + 0.785, + 0.355 + ], + "angle": 0, + "content": "Fig. 4: Dataset generation process Our dataset generation consists of two phases. Data Generation is to generate sets of image pairs under one editing caption. Data Processing is randomly pick image pairs under the same editing instruction and concatenate them together as one input for training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.38, + 0.66, + 0.398 + ], + "angle": 0, + "content": "5 Visual Prompt Dataset For Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.785, + 0.483 + ], + "angle": 0, + "content": "Traditional image editing datasets only contain editing image pairs with a small amount of similar editing instructions. To the best of our knowledge, there is no open-sourced image editing dataset that is explicitly designed for image editing with visual prompting, which utilizes multiple different image pairs for each editing instruction to provide a general demonstration in various cases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.484, + 0.788, + 0.741 + ], + "angle": 0, + "content": "Therefore, we introduce and open source a new dataset that is designed specifically for image editing tasks utilizing visual prompts. Our dataset generation pipeline is shown in Fig. 4 with 2 phases: Data Generation and Data Processing. In Data Generation phase, we first fine-tune GPT-3 [4] for 2 epochs with 700 groups of human-written edits, each consisting of 5 different pairs of input and edited captions with one editing instruction. Then as shown in step 1, we generate around 3500 groups of editing prompts using the fine-tuned GPT-3 model. In each group, there is one instruction \\( E_0 \\) and five pairs of caption and edited caption \\( (I_0,O_0), (I_1,O_1), \\dots, (I_4,O_4) \\). In step 2, similar to InstructPix2Pix, we then also adopt Prompt-to-Prompt for image generation. For each input and edited caption pair, we generated 50 times each with random noise and followed InstructPix2Pix to keep the best one image pair using CLIP-based metric. In addition, we also make sure for each editing instruction, there are at least two pairs of images. In step 3, we generate more image pair sets using in the same way. With filtering, we obtained around 12,000 images with around 2,000 editing instructions. In data preparing, we randomly choose 2 pairs of images under the same editing instruction and concatenate them for training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.761, + 0.376, + 0.778 + ], + "angle": 0, + "content": "6 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.789, + 0.446, + 0.804 + ], + "angle": 0, + "content": "6.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Datasets. To fairly compare our methods with baseline methods, we conduct both qualitative and quantitative experiments on our proposed synthetic dataset" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.392, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "InstructGIE: Towards Generalizable Image Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.372, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.221, + 0.338, + 0.229 + ], + "angle": 0, + "content": "Crown Him Like A King" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.145, + 0.454, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.221, + 0.455, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.145, + 0.538, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.221, + 0.538, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.145, + 0.621, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.221, + 0.621, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.145, + 0.704, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.221, + 0.704, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.145, + 0.787, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.221, + 0.787, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.221, + 0.372, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.292, + 0.341, + 0.3 + ], + "angle": 0, + "content": "As a Precious Oil Painting" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.229, + 0.455, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.302, + 0.455, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.302, + 0.538, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.302, + 0.621, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.302, + 0.704, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.302, + 0.787, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.303, + 0.372, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.365, + 0.331, + 0.372 + ], + "angle": 0, + "content": "Change to Pink Hair" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.375, + 0.455, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.447, + 0.455, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.375, + 0.538, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.375, + 0.621, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.375, + 0.704, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.375, + 0.787, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.375, + 0.372, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.437, + 0.352, + 0.445 + ], + "angle": 0, + "content": "Light Them On Devastating Fire!" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.447, + 0.455, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.322, + 0.509, + 0.376, + 0.516 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.447, + 0.538, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.447, + 0.621, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.447, + 0.704, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.447, + 0.787, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.533, + 0.788, + 0.562 + ], + "angle": 0, + "content": "Fig. 5: Qualitative Comparison on our Test Dataset. We conducted experiments on various scenarios, including human, architecture and landscape." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.787, + 0.641 + ], + "angle": 0, + "content": "that involves over 12,000 filtered images and 2,000 filtered editing instructions. All single input images have a resolution of \\(512 \\times 512\\) and are resized to \\(256 \\times 256\\) for training and testing purposes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Implementation Details. In our approach, we split the dataset with \\(80\\%\\) for training and \\(20\\%\\) for testing. As demonstrated in Fig. 2, two image pairs are concatenated into one \\(512 \\times 512\\) image with the same editing instructions as \\(\\mathbf{Img}^{\\mathrm{train}}\\) and mask the fourth quadrant with a grey color as \\(\\mathbf{Img}^{\\mathrm{VPcon}}\\). We prepare the in-domain test dataset in the same format. For out-of-domain testings, we ensure that both the visual instruction pairs and the text instructions are not used during training, to best simulate how models perform in real-life image editing generalization scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "For baselines, we use their original configurations to train their model. In our method, we only fine-tune the additional ControlNet for 5000 steps with a batch size of 1024 and a learning rate at \\(1 \\times 10^{-4}\\). During training, we set the classifier-free scale the same as the original ControlNet. And we randomly drop \\(15\\%\\) of language or visual editing instructions to further enhance our model's" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.129 + ], + "angle": 0, + "content": "Z. Meng and C. Yang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.187 + ], + "angle": 0, + "content": "Table 1: Quantitative results comparison between our method and baseline methods. Quantitative results show our method excels the baseline methods in both FID and CLIP directional Similarity score in a great margin." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.199, + 0.784, + 0.257 + ], + "angle": 0, + "content": "
SDEdit (E)InstructPix2PixPromptDiffusionOurs
FID ↓21.6717.8713.757.57
CLIP DirSim ↑0.110.170.210.27
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.286, + 0.784, + 0.316 + ], + "angle": 0, + "content": "generalization ability. Our implementation utilizes PyTorch and is trained on 4 Tesla A100-40G GPUs with AdamW optimizer." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.317, + 0.785, + 0.422 + ], + "angle": 0, + "content": "Comparison Methods. To evaluate the effectiveness of our work, we compare with other state-of-the-art image editing frameworks, including SDEdit [28], Instruct-Pix2pix [3] and PromptDiffusion [41]. We adopt two quantitative metrics Fr'echet Inception Distance (FID) and CLIP directional Similarity (CLIP DirSim) proposed by Gal et al. [12]. We utilize the FID score to quantitatively represent the image quality of generated editing outputs, and CLIP DirSim to evaluate how well the models follow editing instructions to produce the output." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.448, + 0.506, + 0.464 + ], + "angle": 0, + "content": "6.2 State-of-the-Art Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.477, + 0.786, + 0.718 + ], + "angle": 0, + "content": "Qualitative Evaluation. In Fig. 5, we present our qualitative results in the testing set (in domain). The comparison shows that our method surpass previous baseline methods. Our method understands and follows both visual and language editing instructions better, and produces a far more detailed manipulated output especially in human figures. In Fig. 6, we present our qualitative results tested in out-of-domain settings. We include five editing instructions that are considered extremely hard for diffusion-based image editing model [18], including object add, object remove, structure location change, structure action change, and object size change. It is important to note that due to how we generate our training dataset utilizing Prompt-to-Prompt [15], editing images pairs with these types of editing instructions is not feasible to generate our training data. This generalization comparison shows that our method excels baseline methods by a significant margin. Our method shows a great capability to carry out well-detailed quality outputs following these hard editing instructions in diffusion-based image editing models, while other baseline methods all fail to understand the editing instructions well or perform manipulations close to the editing instructions." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.719, + 0.786, + 0.809 + ], + "angle": 0, + "content": "Quantitative Evaluation. In quantitative evaluation, we score 7.57 in FID, better than SDEdit (E), InstructPix2Pix and PromptDiffusion which scores 21.67, 17.87 and 13.75. We achieve 0.27 in CLIP DirSim, better than baselines with \\(0.11 / 0.17 / 0.21\\) of CLIP DirSim scores. These quantitative findings show that our method generates higher-quality images with better detailed qualities and also exactly follows both language and visual editing instructions." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.81, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Ablation study. We conduct an ablation study on each of the four components of our proposed method. Namely, the Reformed Conditioned Latent Diffusion" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.392, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "InstructGIE: Towards Generalizable Image Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.372, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.221, + 0.325, + 0.229 + ], + "angle": 0, + "content": "Add Some Snow" + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.144, + 0.455, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.144, + 0.538, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.144, + 0.621, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.144, + 0.704, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.707, + 0.144, + 0.787, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.23, + 0.372, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.292, + 0.321, + 0.3 + ], + "angle": 0, + "content": "Bend His Arm" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.23, + 0.455, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.23, + 0.538, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.23, + 0.621, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.23, + 0.704, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.707, + 0.23, + 0.787, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.303, + 0.372, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.365, + 0.331, + 0.373 + ], + "angle": 0, + "content": "Move It to the Right" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.303, + 0.455, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.303, + 0.538, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.303, + 0.621, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.303, + 0.704, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.707, + 0.303, + 0.787, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.375, + 0.372, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.437, + 0.323, + 0.445 + ], + "angle": 0, + "content": "Keep Only One" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.375, + 0.455, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.375, + 0.538, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.375, + 0.621, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.375, + 0.704, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.707, + 0.375, + 0.787, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.447, + 0.372, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.509, + 0.324, + 0.516 + ], + "angle": 0, + "content": "Make It Smaller" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.447, + 0.455, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.446, + 0.538, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.447, + 0.621, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.447, + 0.704, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.707, + 0.447, + 0.787, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.533, + 0.787, + 0.561 + ], + "angle": 0, + "content": "Fig. 6: Qualitative Comparison on Out-of-Domain Images. We conducted experiments on images and instruct that are not in training dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Model (RCLDM), Editing Shift Matching (ESM), Language Instruction Unification (LIU) and Selective Area Matching (SAM). We present the qualitative results in Fig. 7. From the qualitative results, we can see that without RCLDM and ESM, the model understands the visual editing instructions much weaker, especially in out-of-domain editing. Without LIU, for two language editing instructions with the same meaning, the model produces two output edited images in different detail and quality. This difference in quality tends to increase in out-of-domain settings. Without SAM, the details of the human face are distorted making the model more vulnerable to producing outputs with worse detailed qualities. We also include the qualitative ablation results on the testing dataset in Tab. 2. From observation, our method performs the best when incorporating all four components. Without SAM or LIU, the FID score increases, meaning those modules enhance the detail quality of the output generated. Without CLDM or ESM, the CLIP DirSim score decreases, showing that those two modules contribute to a better understanding in both language and visual level." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.129 + ], + "angle": 0, + "content": "Z. Meng and C. Yang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.355, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.212, + 0.32, + 0.218 + ], + "angle": 0, + "content": "Make It in Winter Time" + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.145, + 0.425, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.145, + 0.497, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.145, + 0.57, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.145, + 0.642, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.145, + 0.713, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.145, + 0.788, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.219, + 0.353, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.274, + 0.319, + 0.281 + ], + "angle": 0, + "content": "Turn It to Winter Time" + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.219, + 0.425, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.219, + 0.497, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.219, + 0.57, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.219, + 0.642, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.219, + 0.713, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.219, + 0.785, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.282, + 0.353, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.337, + 0.317, + 0.343 + ], + "angle": 0, + "content": "Make Emotion Cuter" + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.282, + 0.425, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.282, + 0.497, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.282, + 0.57, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.282, + 0.642, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.282, + 0.713, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.282, + 0.785, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.345, + 0.353, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.263, + 0.4, + 0.307, + 0.406 + ], + "angle": 0, + "content": "Cuter Emotion" + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.345, + 0.425, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.345, + 0.497, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.345, + 0.57, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.345, + 0.642, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.345, + 0.713, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.345, + 0.785, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.425, + 0.785, + 0.454 + ], + "angle": 0, + "content": "Fig. 7: Ablation Study Results for both in-domain (first two rows), and out-of-domain first two rows) image manipulations" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.476, + 0.785, + 0.532 + ], + "angle": 0, + "content": "Table 2: Abalation Study results of comparison between our method with all components and without the Reformed Conditioned Latent Diffusion Model (RCLDM), Editing Shift Matching (ESM), Language Instruction Unification (LIU) and Selective Area Matching (SAM). The ablation study is conducted on the entire test dataset." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.545, + 0.784, + 0.602 + ], + "angle": 0, + "content": "
w/o RCLDMw/o ESMw/o LIUw/o SAMOurs
FID ↓10.159.2310.3711.317.57
CLIP DirSim ↑0.130.150.230.190.27
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.637, + 0.358, + 0.652 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In this work, we propose InstructGIE, an image editing framework with enhanced generalization ability, improving performance in both visual and text aspects. We incorporate a VMamba-based module to enhance visual outputs and introduce an editing-shift matching strategy to augment in-context learning. Our selective area-matching technique addresses and rectifies corrupted details, while a language unification technique aligns language embeddings with editing semantics. Additionally, we compile a publicly available dataset for better generalization evaluation. Extensive experiments demonstrate our framework's superior in-context generation performance and robust generalization capability across unseen vision tasks, both quantitative and qualitively." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.392, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "InstructGIE: Towards Generalizable Image Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.787, + 0.237 + ], + "angle": 0, + "content": "This work is partially supported by the Army Research Office/Army Research Laboratory via grant W911-NF-20-1-0167 to Northeastern University, National Science Foundation CCF-1937500, and the Fundamental Research Funds for the Central Universities, Peking University." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.259, + 0.323, + 0.275 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.331 + ], + "angle": 0, + "content": "1. Bar, A., Gandelsman, Y., Darrell, T., Globerson, A., Efros, A.: Visual prompting via image inpainting. Advances in Neural Information Processing Systems 35, 25005-25017 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.785, + 0.358 + ], + "angle": 0, + "content": "2. Blau, T., Ganz, R., Kawar, B., Bronstein, A., Elad, M.: Threat model-agnostic adversarial defense using diffusion models. arXiv preprint arXiv:2207.08089 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.36, + 0.785, + 0.399 + ], + "angle": 0, + "content": "3. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.401, + 0.785, + 0.441 + ], + "angle": 0, + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.443, + 0.785, + 0.468 + ], + "angle": 0, + "content": "5. Chen, W., Hu, H., Sahara, C., Cohen, W.W.: Re-imagen: Retrieval-augmented text-to-image generator. arXiv preprint arXiv:2209.14491 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.47, + 0.785, + 0.523 + ], + "angle": 0, + "content": "6. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290–1299 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.525, + 0.785, + 0.564 + ], + "angle": 0, + "content": "7. Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8188-8197 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.606 + ], + "angle": 0, + "content": "8. Chung, J.S., Zisserman, A.: Lip reading in the wild. In: Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13. pp. 87-103. Springer (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.608, + 0.785, + 0.647 + ], + "angle": 0, + "content": "9. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.649, + 0.785, + 0.674 + ], + "angle": 0, + "content": "0. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.676, + 0.785, + 0.729 + ], + "angle": 0, + "content": "11. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.731, + 0.785, + 0.771 + ], + "angle": 0, + "content": "12. Gal, R., Patashnik, O., Maron, H., Chechik, G., Cohen-Or, D.: Stylegan-nada: Clipguided domain adaptation of image generators. arXiv preprint arXiv:2108.00946 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.773, + 0.785, + 0.798 + ], + "angle": 0, + "content": "13. Geng, X., Liu, H.: Openllama: An open reproduction of llama (May 2023), https://github.com/openlm-research/open_llama" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.839 + ], + "angle": 0, + "content": "14. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.289, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.129 + ], + "angle": 0, + "content": "Z. Meng and C. Yang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "15. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.219 + ], + "angle": 0, + "content": "16. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.219, + 0.785, + 0.26 + ], + "angle": 0, + "content": "17. Huang, Y., Huang, J., Liu, Y., Yan, M., Lv, J., Liu, J., Xiong, W., Zhang, H., Chen, S., Cao, L.: Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.785, + 0.303 + ], + "angle": 0, + "content": "18. Huang, Y., Huang, J., Liu, Y., Yan, M., Lv, J., Liu, J., Xiong, W., Zhang, H., Chen, S., Cao, L.: Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.785, + 0.346 + ], + "angle": 0, + "content": "19. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.785, + 0.374 + ], + "angle": 0, + "content": "20. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. Advances in Neural Information Processing Systems 35, 23593-23606 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.785, + 0.403 + ], + "angle": 0, + "content": "21. Kawar, B., Ganz, R., Elad, M.: Enhancing diffusion-based image synthesis with robust classifier guidance. arXiv preprint arXiv:2208.08664 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.785, + 0.459 + ], + "angle": 0, + "content": "22. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.785, + 0.515 + ], + "angle": 0, + "content": "23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Pereira, F., Burges, C., Bottou, L., Weinberger, K. (eds.) Advances in Neural Information Processing Systems. vol. 25. Curran Associates, Inc. (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.544 + ], + "angle": 0, + "content": "24. Kwon, G., Ye, J.C.: Diffusion-based image translation using disentangled style and content representation. arXiv preprint arXiv:2209.15264 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.785, + 0.6 + ], + "angle": 0, + "content": "25. Liu, X., Park, D.H., Azadi, S., Zhang, G., Chopikyan, A., Hu, Y., Shi, H., Rohrbach, A., Darrell, T.: More control for free! image synthesis with semantic diffusion guidance. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 289-299 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.785, + 0.628 + ], + "angle": 0, + "content": "26. Liu, Y., Tian, Y., Zhao, Y., Yu, H., Xie, L., Wang, Y., Ye, Q., Liu, Y.: Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.785, + 0.67 + ], + "angle": 0, + "content": "27. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of International Conference on Computer Vision (ICCV) (December 2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.671, + 0.785, + 0.713 + ], + "angle": 0, + "content": "28. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.785, + 0.756 + ], + "angle": 0, + "content": "29. Mohammad, S., Kiritchenko, S.: Wikiart emotions: An annotated dataset of emotions evoked by art. In: Proceedings of the eleventh international conference on language resources and evaluation (LREC 2018) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.785 + ], + "angle": 0, + "content": "30. Nguyen, T., Li, Y., Ojha, U., Lee, Y.J.: Visual instruction inversion: Image editing via visual prompting. arXiv preprint arXiv:2307.14331 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.392, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "InstructGIE: Towards Generalizable Image Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "33. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "34. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "35. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.355 + ], + "angle": 0, + "content": "36. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.786, + 0.397 + ], + "angle": 0, + "content": "37. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "38. Sun, Y., Yang, Y., Peng, H., Shen, Y., Yang, Y., Hu, H., Qiu, L., Koike, H.: Imagebrush: Learning visual in-context instructions for exemplar-based image manipulation. arXiv preprint arXiv:2308.00906 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.786, + 0.466 + ], + "angle": 0, + "content": "39. Theis, L., Salimans, T., Hoffman, M.D., Mentzer, F.: Lossy compression with gaussian diffusion. arXiv preprint arXiv:2206.08889 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.508 + ], + "angle": 0, + "content": "40. Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6830-6839 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "41. Wang, Z., Jiang, Y., Lu, Y., He, P., Chen, W., Wang, Z., Zhou, M., et al.: Incontext learning unlocked for diffusion models. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.591 + ], + "angle": 0, + "content": "42. Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., Xiao, J.: Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.787, + 0.619 + ], + "angle": 0, + "content": "43. Zablotskaia, P., Siarohin, A., Zhao, B., Sigal, L.: Dwnet: Dense warp-based network for pose-guided human video generation. arXiv preprint arXiv:1910.09139 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.787, + 0.66 + ], + "angle": 0, + "content": "44. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.787, + 0.688 + ], + "angle": 0, + "content": "45. Zimmermann, R.S., Schott, L., Song, Y., Dunn, B.A., Klindt, D.A.: Score-based generative classifiers. arXiv preprint arXiv:2110.00473 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.688 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/InstructGIE_ Towards Generalizable Image Editing/d8d7c0e8-e8d9-4aa1-9b02-c4948b5c2541_origin.pdf b/2024/InstructGIE_ Towards Generalizable Image Editing/d8d7c0e8-e8d9-4aa1-9b02-c4948b5c2541_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4aa0fdf1c0e4666733a47575851c2e58cbf5ccbf --- /dev/null +++ b/2024/InstructGIE_ Towards Generalizable Image Editing/d8d7c0e8-e8d9-4aa1-9b02-c4948b5c2541_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26e26a0d8218a7560b957aaf6fb7b3d1cc9b6f8b7a01c1b3b73d4ce9c2e0f061 +size 9586921 diff --git a/2024/InstructGIE_ Towards Generalizable Image Editing/full.md b/2024/InstructGIE_ Towards Generalizable Image Editing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2915531938787a2dbe3437433ed6f83e39f8dccb --- /dev/null +++ b/2024/InstructGIE_ Towards Generalizable Image Editing/full.md @@ -0,0 +1,477 @@ +# InstructGIE: Towards Generalizable Image Editing + +Zichong Meng\*, Changdi Yang\*, Jun Liu\*, Hao Tang\*\*, Pu Zhao\*\*, and Yanzhi Wang\*\* + +$^{1}$ Northeastern University, Boston MA 02115, USA + +2 Peking University, Beijing 100871, China & + +Carnegie Mellon University, Pittsburgh PA 15213, USA + +![](images/1163a20c0ef3ae95a0a9ff40e4ae76f0a4ebc103dbef8ac4214961895c7a95c4.jpg) +Like in Game +Add a Cow +Fig. 1: Demo results of the proposed InstructGIE framework on various image manipulation tasks to both humans and scenes. By our proposed method, our model can generalize to generate the desired output with great detail qualities. + +Abstract. Recent advances in image editing have been driven by the development of denoising diffusion models, marking a significant leap + +forward in this field. Despite these advances, the generalization capabilities of recent image editing approaches remain constrained. In response to this challenge, our study introduces a novel image editing framework with enhanced generalization robustness by boosting in-context learning capability and unifying language instruction. This framework incorporates a module specifically optimized for image editing tasks, leveraging the VMamba block and an editing-shift matching strategy to augment in-context learning. Furthermore, we unveil a selective area-matching technique specifically engineered to address and rectify corrupted details in generated images, such as human facial features, to further improve the quality. Another key innovation of our approach is the integration of a language unification technique, which aligns language embeddings with editing semantics to elevate the quality of image editing. Moreover, we compile the first dataset for image editing with visual prompts and editing instructions that could be used to enhance in-context capability. Trained on this dataset, our methodology not only achieves superior synthesis quality for trained tasks, but also demonstrates robust generalization capability across unseen vision tasks through tailored prompts. Our project page is available at https://cr8br0ze.github.io/InstructGIE. + +Keywords: Image Editing $\cdot$ In-Context Learning $\cdot$ Diffusion Model + +# 1 Introduction + +As a crucial task in computer vision, image editing has witnessed significant improvements enhanced with the increasingly popular denoising stable diffusion techniques in recent years [22,25,38,40,41]. Given a set of text or image prompts as generation constraints or instructions, diffusion-based image editing can follow the instructions and synthesize an edited image. However, since the model does not have the capability to accurately model all possible samples in the conditional distribution space [17], if specific instructions are not included in the training dataset, current diffusion-based image editing methods can hardly generate satisfactory results. Thus, editing performance largely depends on the training dataset without superior generalization capabilities. + +On the other hand, large language models (LLMs) have proven extraordinary abilities to learn from contexts, referred to as in-context learning, which allows LLMs to perform unseen tasks by providing a combination of input-output examples and a query input. Inspired by the potential to enhance the generalization of the model with LLMs, [38,41] explore in-context learning for computer vision tasks, allowing them to perform unseen tasks with novel vision-language prompt designs. However, these methods are not tailored for image editing applications, leading to unsatisfying synthetic qualities with inaccurate or incorrect output and lack of detail. To improve the generalization of image editing with improved synthetic image quality, it is crucial to effectively understand the text & image prompts and specifically control image editing details, which is challenging in the current literature. + +In this work, we propose InstructGIE, an image editing framework with enhanced generalizability. We improve image editing performance from both visual and text aspects. (i) For the visual information, we incorporate a VMamba-based module to specifically enhance the image editing outputs. As VMamba [26] has proven its better performance in capturing in-context information from inputs with larger receptive fields [26], we leverage VMamba and propose an editing-shift matching strategy to augment in-context learning. Furthermore, since current image editing works do not perform well in generating correct features with accurate details, we unveil a selective area-matching technique specifically engineered to address and rectify corrupted details in generated images, such as human facial features, to further improve the quality. (ii) Another key innovation of our approach is the integration of a language unification technique, which aligns language embeddings with editing semantics to elevate the quality of image editing. Our framework not only achieves superior in-context generation for trained tasks but also demonstrates robust generalization across unseen vision tasks. Moreover, we compile a publicly available image editing dataset with plenty of visual prompts and editing instructions for better generalization evaluation of image editing. Our contributions are summarized as follows: + +- We propose an image editing framework, including in-context learning enhancement and language unification strategies, specifically designed to enhance generalization ability from both visual and text domains. +- We compile the first dataset for image editing with visual prompts and editing instructions that could be used to enhance in-context capability. +- We conduct extensive experiments and achieve great generalization ability in the multiple unseen image editing task, both quantitatively and qualitatively. + +# 2 Related Works + +# 2.1 Denoising Stable Diffusion Based Image Editing + +Denoising Stable Diffusion [16,35,36] based image editing could follow guidance from text or image prompts. With the foundation of text-guided models offering rich generative capabilities, there has been a surge in research aimed at adapting these models for image manipulation tasks from textual descriptions. To steer the image editing process in the desired direction, the use of models like CLIP to fine-tune diffusion models has become a common practice. Although these methods [3,22,41] have shown impressive results, they often involve costly finetuning processes. Recent innovations [14] have introduced techniques that inject cross-attention into the models to more effectively edit specific semantic areas within the spatial feature maps. Further advancements [24] have enhanced these techniques by adding semantic loss or applying attention loss to refine the integration of plugged features, improving the precision and quality of the editing outcomes. [38] proposes a framework that could learn instructions from visual image pairs for more accurate editing and firstly formulate this task as an image inpainting problem. + +# 2.2 Generalization Capability in Visual Tasks + +In-context learning is widely applied in the field of natural language processing (NLP), enabling models to adapt to new tasks such as translation, question answering, and complex reasoning. NLP models utilize in-context examples, comprising text and corresponding labels, to tackle tasks they haven't seen before. However, applying in-context learning to the visual domain introduces more challenges and remains less explored. A significant hurdle is the nature of fixed-size input requirements for vision models, as opposed to variable-length text inputs that can be managed by language models. Vision models generally struggle with processing inputs of varying sizes, making it impractical to process multiple image prompts in one-shot for global understanding. Moreover, in intricate visual understanding, specific instructions are often implied from a limited set of image examples rather than explicitly stated, which poses additional difficulties for vision models in identifying and understanding high-level visual relationships. + +Recent strides in applying masked image modeling have marked a step forward in improving in-context learning for vision models. The method proposed by [41], employing a masked autoencoder-based technique, predicts a missing image within a two-by-two grid, using two images as in-context examples and another as the query. This concept was later expanded by [38] with a multitask framework. Despite their progress, such inpainting methods are limited by the necessity of a fixed number of in-context examples and increased memory demands. Painter, highlighted in [40], exemplifies an inpainting approach tailored for versatility across various vision tasks. + +In contrast, inspired by ControlNet [44], [41] refines the framework by adding an additional pair of example images and employing a multitask supervised finetuning method. Prompt diffusion excels in visual in-context learning. However, it faces certain limitations or challenges in its practical applications. + +# 2.3 Dataset for Diffusion-based Image Editing + +Currently, various types of datasets are used for training in diffusion-based image editing. There are datasets that concentrate on specific domains like CelebA [27] and FFHQ [19] for human face image manipulation, AFHQ [7] for animal face image editing, LSUN [42] for object modification, and WikiArt [29] for style transfer. In-the-wild video datasets could also be leveraged to train image editing tasks. The Scannet dataset [9] encompasses a vast array of more than 1,500 indoor scenes from various settings, such as apartments, offices, and hotels, providing extensive annotations. The LRW dataset [8], tailored for lip reading tasks, includes more than 1000 video utterances of 500 distinct words. The UBC-Fashion dataset [43] features 600 videos spanning various clothing categories, with 500 videos allocated for training and 100 for testing, guaranteeing no repetition of individuals in the training set. The DAVIS dataset [43] (Densely Annotated VVideo Segmentation), a widely recognized benchmark for video object segmentation, contains 150 videos in total. There are also image editing works proposing to generate datasets with editing instructions. InstructPix2pix [3] collects over + +450,000 training image pairs. For each pair, given an image with its caption, it first uses a finetuned GPT-3 [4] to generate an editing instruction and an edited image caption. Then it employs Stable Diffusion and the Prompt-to-Prompt algorithm [14] to generate edited image following the caption. However, currently there are no datasets with multiple image pairs under one editing instruction, which is crucial to enhance the generalization ability of image editing. + +# 3 Preliminary + +Recent advances in generative models have been significantly driven by the emergence of diffusion models, which have set new benchmarks in image creation [10,21,37]. These models have found applications across a broad spectrum of areas [2,5,20,39,45], demonstrating their versatility and effectiveness. The fundamental concept behind diffusion models involves starting with an image that is initially just random noise $\mathbf{x}_T\sim \mathcal{N}(0,\mathbf{I})$ and progressively refining this image step by step until it becomes a high-quality, realistic image $\mathbf{x}_0$ . This refinement process involves generating intermediate samples $\mathbf{x}_t$ (for $t\in \{0,\dots ,T\}$ ), where each sample is defined as: + +$$ +\mathbf {x} _ {t} = \sqrt {\alpha_ {t}} \mathbf {x} _ {0} + \sqrt {1 - \alpha_ {t}} \boldsymbol {\epsilon} _ {t}, \tag {1} +$$ + +where the parameter $\alpha_{t}$ sets the pace of the diffusion process, ranging from $0 = \alpha_{T} < \alpha_{T - 1} < \dots < \alpha_{1} < \alpha_{0} = 1$ , and $\epsilon_t\sim \mathcal{N}(0,\mathbf{I})$ represents the added noise. The model refines the image by applying a neural network $f_{\theta}(\mathbf{x}_t,t)$ to each sample $\mathbf{x}_t$ , followed by the addition of Gaussian noise to produce the next sample $\mathbf{x}_{t - 1}$ . This neural network is optimized to achieve a denoising goal, striving for $f_{\theta}(\mathbf{x}_t,t)\approx \epsilon_t$ , resulting in a generative process that closely mimics the desired image distribution. + +Expanding this framework to conditional generative modeling, the process involves conditioning the neural network $f_{\theta}(\mathbf{x}_t,t,\mathbf{y})$ on an additional input $\mathbf{y}$ , enabling the generation of images from a distribution conditioned on $\mathbf{y}$ . This conditional input could be anything from a low-resolution image, a category label, or a descriptive text sequence. Leveraging the advancements in LLMs [33] and hybrid vision-language models [31], text-to-image diffusion models are developed. These models allow for the creation of detailed, high-resolution images from mere text descriptions, starting with a low-resolution image generated through the diffusion process, which is subsequently refined into a high-resolution image using additional models. + +# 4 The Proposed Method + +We present our framework pipeline in Fig. 2. For efficient training and better controllability, we adopt a line of techniques with popular architectures such as ControlNet [44] and Stable Diffusion [34] to design a generalizable image editing + +![](images/558d61ed0c6aa5cec6856d32f1f5db211c09dff21ca1d6c7eb3b3e53c71e943a.jpg) +Fig. 2: Overall architecture of InstructGIE. The lower pipeline is for both training and inference processes where the model obtains unified editing instructions outputted by Instruction Unification Module $\mathcal{U}$ and combines with visual prompted input $\mathbf{Img}^{\mathrm{VPcon}}$ to pass through Zero-VMamba integrated Stable Diffusion model with ControlNet for output image. The upper pipeline is for training only which compares output image and training ground truth $\mathbf{Img}^{\mathrm{train}}$ and computes editing shift loss $\mathcal{L}_{es}$ with Editing Shift Module and selective area matching loss $\mathcal{L}_{sam}$ with Selective Area Matching Module. + +tool with accurate high-quality outputs. Specifically, we introduce enhanced in-context learning both at the architecture level and training level to improve the image quality. Furthermore, language instruction unification is adopted to maximize the generalization ability for unseen editing tasks. Moreover, selective area matching is proposed to further improve the output quality with full details. + +# 4.1 Enhanced In-Context Learning + +Visual prompting based on inpainting is an effective visual in-context learning method in various computer vision tasks [1,40], which is applied in image editing tasks [38] recently. However, the methods perform poorly in quality when dealing with unseen image manipulation tasks. Therefore, we propose the enhanced in-context learning specifically tailored for generalizable image editing. + +Reforming Conditioned Latent Diffusion Model. To improve the generalization ability of image editing, it is crucial for the framework to explicitly capture low-level visual editing contexts. Current diffusion-based image editing + +![](images/d3da15b41880757d9c95fae96f376efce5c9ef400fe4063e1481d597a777a623.jpg) +ERF of ResNet-50 + +![](images/71e7689d683d74dc0c08647a3d29f4ba934dbf57742ff9c21a5ad312fa416d2d.jpg) +ERF of VIT-L/14 +Fig. 3: Effective Reception Field (ERF) of ConvNet, ViT, VMamba based model architectures. + +![](images/f9ed6567ff258182b8f8926cccb22cfd59509a8d97c2aa0268ed804cda3d13d8.jpg) +ERF of VMamba-T + +methods [30, 38] that involve visual prompting either adopt ConvNet [23] or ViT [11] as the vision encoder for visual prompts. However, these methods fail to generalize well as they are not able to capture enough visual editing contexts (see Fig. 3). To address this, we formulate the visual prompted condition as a single image $\mathbf{Img}^{\mathrm{VPcon}} = \{\mathbf{Img}_0^{\mathrm{in}}, \mathbf{Img}_0^{\mathrm{out}}, \mathbf{Img}_1^{\mathrm{in}}, \mathbf{Grey}\}$ , as shown in Fig 2, with a global effective receptive field (ERF). + +Moreover, we further propose to reform the conditioned latent diffusion model Inspired by the recent visual state space model VMamba [26] which exhibits a better global ERF and also emphasizes shifting boundaries of input's four quadrants, we propose to adopt a vision encoder based on Zero-VMamba to fit our structure. Specifically, the vision encoder $\mathcal{G}_{ven}$ comprehends the visual prompted condition in latent space $x_{vpc}$ as follows, + +$$ +e _ {v p c} = \mathcal {G} _ {v e n} \left(x _ {v p c}; \Theta_ {g}\right), \tag {2} +$$ + +where $e_{vpc}$ is the processed embedding of the visual prompted condition, and $\Theta_g$ is the model parameters initialized to zeros. + +To further improve the performance, after each ControlNet trainable copied modules $\mathcal{F}_c$ with parameters $\Theta_c$ , we propose to link and inject the processed visual prompted condition information to the frozen Stable Diffusion model $\mathcal{F}$ with parameters $\Theta$ through zero-VMamba layer $\mathcal{G}$ . We use two instances of VMamba with parameters $\theta_{g1}$ and $\theta_{g2}$ respectively. The complete model then computes the following, + +$$ +y _ {v p c} = \mathcal {F} (x; \Theta) + \mathcal {G} (\mathcal {F} (x + \mathcal {G} (x _ {v p c}; \Theta_ {g 1}); \Theta_ {c}); \Theta_ {g 2}), \tag {3} +$$ + +where $y_{vpc}$ is the output of our conditioned diffusion model block. + +Our conditioned latent diffusion model can process all four quadrants in our visual prompted conditions with a global receptive field, while it does not generate random noises during initial training steps. + +Editing-Shift Matching. Besides the architecture innovation, we incorporate an editing-shift-matching technique to enhance in-context learning ability in image editing with more accurate detailed outputs. + +In specific, for each training ground truth $\mathbf{Img}^{\mathrm{train}} = \{\mathbf{Img}_0^{\mathrm{in}},\mathbf{Img}_0^{\mathrm{out}},\mathbf{Img}_1^{\mathrm{in}},$ $\mathbf{Img}_1^{\mathrm{out}}\}$ , we calculate a implicit editing shift value using CLIP [32] image embedding: + +$$ +\mathcal {T} (\mathbf {I m g}) = \frac {\sum_ {i = 0} ^ {1} \operatorname {C L I P} \left(\mathbf {I m g} _ {i} ^ {\text {i n}}\right) - \operatorname {C L I P} \left(\mathbf {I m g} _ {i} ^ {\text {o u t}}\right)}{2}. \tag {4} +$$ + +During the training process, after predicting the noise, we use it to reverse the noised input and obtain a pseudo output image $\mathbf{Img}^{\mathrm{PO}} = \{\mathbf{PO}_0^{\mathrm{in}},\mathbf{PO}_0^{\mathrm{out}},\mathbf{PO}_1^{\mathrm{in}},$ $\mathbf{PO}_1^{\mathrm{out}}\}$ . Our framework then calculates the editing transfer value of the pseudo output image and deduces a editing shift loss to optimize during our training via the cosine similarity of the two design transfer values: + +$$ +\mathcal {L} _ {e s} = 1 - \frac {\mathcal {T} (\mathbf {I m g} ^ {\mathrm {P O}}) \cdot \mathcal {T} (\mathbf {I m g} ^ {\mathrm {t r a i n}})}{| | \mathcal {T} (\mathbf {I m g} ^ {\mathrm {P O}}) | | \times | | \mathcal {T} (\mathbf {I m g} ^ {\mathrm {t r a i n}}) | |} \tag {5} +$$ + +Through editing-shift matching, our model can better comprehend how editing should be done within a visual context level through an implicit editing shift value. Furthermore, this implicit editing shift value can further guide the sampling process creating a controllable editing. + +# 4.2 Language Instruction Unification + +Previous works that utilize visual prompting in image editing tend to focus more on visual instructions and believe language instructions can be vague and unstable [38]. We believe that the potential of text prompts is not fully explored in image editing. Language instructions have significant impacts on diffusion model outputs. Language instructions with the same meaning can still result in entirely different outputs due to different processed language embeddings. + +To improve image editing and explore language instructions, we propose a novel approach, language instruction unification. During the training process, for each batch of training data, we randomly sample $50\%$ of training data, collect their language editing instructions $l$ , and process them through a frozen lightweight LLM, Open Llama 3b V2 Quant 4 [13] denoted as $\mathcal{U}$ . The LLM $\mathcal{U}$ is fixed prompted with a fixed random seed to uniformly reformulate the language instruction $l$ better for machine-level understanding. The LLM $\mathcal{U}$ will output a unified language editing instruction $l'$ . + +$$ +l ^ {\prime} = \mathcal {U} (l) \tag {6} +$$ + +We then augment the training data with unified language editing instructions. + +During the inference, each language editing instruction $l$ is passed through the frozen LLM $\mathcal{U}$ for language instruction unification and then sent to our conditioned diffusion model. + +By adopting language instruction unification for training data augmentation during the training, our conditioned diffusion model can learn diverse non-uniformed editing instructions to build up the model's knowledge distribution with unified language prompts. Adopting language instruction unification during + +inference aligns with the training, therefore, greatly minimizing the possibility of diverse quality in edited outputs and maximizing the ability to generalize to unseen editing tasks. + +# 4.3 Selective Area Matching + +Diffusion-based image editing models usually suffer from low quality in specific details and this bottleneck appears to be more crucial in generalizable image editing methods [30, 38]. The details of human and animal images are easily distorted in these methods. + +A naive solution might be utilizing negative prompts [34] like 'do not draw hands with 6 fingers' in general text-to-image tasks. However, in image editing, it is challenging to apply negative prompts. Users typically can not foresee the specific details after-edited outputs, therefore they are not able to construct appropriate negative prompts. Besides, negative prompts may limit the artistic potential of image editing models. + +To properly address this issue for generalizable image editorings, we propose an optimization method, namely selective area matching, that targets the difference in the chosen detailed editing area between the original training ground truth $\mathbf{Img}^{\mathrm{train}}$ and the reversed pseudo output $\mathbf{Img}^{\mathrm{PO}}$ . + +In particular, during the training process, we incorporate a frozen Mask2Former model [6] $\mathcal{M}$ to obtain panoptic segmented training image information including segmented masks and class information $C$ . + +$$ +\operatorname {s e g m a s k}, C = \mathcal {M} (\mathbf {I m g} ^ {\text {t r a i n}}) \tag {7} +$$ + +After that, our framework processes the class information using the same lightweight LLM $\mathcal{U}$ described in Sec. 4.2 to filter out pre-defined classes including living creatures and humans requiring special attention for addressing the details. + +$$ +C _ {\text {f i l t e r e d}} = \left\{c \in C | \mathcal {U} (c, \text {s e l e c t e d c l a s s e s}) = 1 \right\} \tag {8} +$$ + +Based on selected class labels, the framework then deduces a segmented binary mask for the selected editing area. + +$$ +[ \text {m a s k} ] _ {i} = \left\{ \begin{array}{l l} 1 & \text {i f c l s} ([ \text {s e g m a s k} ] _ {i}) \in C _ {\text {f i l t e r e d}} \\ 0 & \text {o t h e r w i s e} \end{array} \right. \tag {9} +$$ + +During the training process, our framework calculates and optimizes the selective area matching loss by + +$$ +\mathcal {L} _ {s a m} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left(\left([ \mathbf {I m g} ^ {\mathrm {P O}} ] _ {i} \cdot [ \operatorname {m a s k} ] _ {i}\right) - \left(\left[ \mathbf {I m g} ^ {\mathrm {t r a i n}} \right] _ {i} \cdot [ \operatorname {m a s k} ] _ {i}\right)\right) ^ {2} \tag {10} +$$ + +where $N$ as the total pixel number in the image. + +With selective area matching, image editing does not need negative prompts to deal with distorted details in images, which can make the most of the model's artistic editing capacity to generate high-quality outputs with great details. It is only incorporated during training, which does not change the inference, greatly saving inference efforts compared with negative prompts. + +![](images/a76af07fe12c29cb021ee74e300bb7f75a2bdb619b59fe337e94dd6b515d5e4a.jpg) +Phase 1: Data Generation + +![](images/9fdfd78cd360bb8a5ac837a7ccc9cfa5b17a5022efc5241c53c6d155295e9246.jpg) +Phase 2: Data Processing +Fig. 4: Dataset generation process Our dataset generation consists of two phases. Data Generation is to generate sets of image pairs under one editing caption. Data Processing is randomly pick image pairs under the same editing instruction and concatenate them together as one input for training. + +# 5 Visual Prompt Dataset For Image Editing + +Traditional image editing datasets only contain editing image pairs with a small amount of similar editing instructions. To the best of our knowledge, there is no open-sourced image editing dataset that is explicitly designed for image editing with visual prompting, which utilizes multiple different image pairs for each editing instruction to provide a general demonstration in various cases. + +Therefore, we introduce and open source a new dataset that is designed specifically for image editing tasks utilizing visual prompts. Our dataset generation pipeline is shown in Fig. 4 with 2 phases: Data Generation and Data Processing. In Data Generation phase, we first fine-tune GPT-3 [4] for 2 epochs with 700 groups of human-written edits, each consisting of 5 different pairs of input and edited captions with one editing instruction. Then as shown in step 1, we generate around 3500 groups of editing prompts using the fine-tuned GPT-3 model. In each group, there is one instruction $E_0$ and five pairs of caption and edited caption $(I_0,O_0), (I_1,O_1), \dots, (I_4,O_4)$ . In step 2, similar to InstructPix2Pix, we then also adopt Prompt-to-Prompt for image generation. For each input and edited caption pair, we generated 50 times each with random noise and followed InstructPix2Pix to keep the best one image pair using CLIP-based metric. In addition, we also make sure for each editing instruction, there are at least two pairs of images. In step 3, we generate more image pair sets using in the same way. With filtering, we obtained around 12,000 images with around 2,000 editing instructions. In data preparing, we randomly choose 2 pairs of images under the same editing instruction and concatenate them for training. + +# 6 Experiments + +# 6.1 Experimental Settings + +Datasets. To fairly compare our methods with baseline methods, we conduct both qualitative and quantitative experiments on our proposed synthetic dataset + +![](images/601e3632b1419bd570fdf82e47fff0a416040136cb19593fda7313cf7acdae4f.jpg) + +![](images/8105753172b4fa12fa52f152417a3609475d28b35f69580b83c0fc45d2ec1a57.jpg) + +![](images/8f610fcb1aa24437e9cf97a63f06059e6f6b520a00368dd3c7eb9e4639059236.jpg) + +![](images/252f52df050f96a72baa362cc8403ac3c2d91102cedd4c00333411ed6a20866a.jpg) + +![](images/b1cc42a4f3bb0deeaef273bd8d2ae025335e3aeef4af21dc6133bcdf75dd31c6.jpg) + +![](images/7596bbcd456e64506fbf9988e26a2fad4750635029642a6bbbcba3efa9b137fc.jpg) + +![](images/19932ac3b1098543dbf8c6eae2e2f6fe8bd01d7d57c9009e2efa8074b3eb8ba9.jpg) + +![](images/2f21625a226f9308eefa402e70bdd9c1e44903c9e46e93df9b74711a3a6dcc8b.jpg) + +![](images/00289cd57fb4612c9d0d743ee8443acd4ba57f001ad33cf587add54f890a73bf.jpg) + +![](images/03d83ccabe08a072f88ccd06b15dde5330aa25bfd0bd1c821cbd714c68329d88.jpg) + +![](images/e682fc074c481e8a2f7286e53702f81b3759c256a3ad0d7068cd9e2f9a692b34.jpg) + +![](images/c228565a9f3052fcafaeabc8a841ce408e879f8d87bad1909d0acf7d5987cab1.jpg) +Crown Him Like A King +As a Precious Oil Painting + +![](images/5ffbb102262a4b334bd79d909983d6f262d5bdb4e9f085fbcb5ccf048aafd65a.jpg) + +![](images/a312e3dc3cdf2ca508f5ac46676ee92d8a5fdbe7ae34a26f9397a4e441fb9625.jpg) + +![](images/f154617d457968c5fd517493e62732ce0fab753c06c80f38ab7f5c52b96c7f62.jpg) + +![](images/ea110f8ab3669df4fcba8d9970e1aaac3ac1f3c7651a38e3a0e2407806ec3baf.jpg) + +![](images/2f3cee44554f9e3944a382f77df85585738ffb3deeb196a2df1c1b49421b887f.jpg) + +![](images/2303655566728338af3498715949bcab20ee80be9a8f90e506bfe44ac8f31ef1.jpg) + +![](images/8412a367a8ae7966e67454a42d393586e4528eca1bd62df9205e660f8d8cc20b.jpg) +Change to Pink Hair + +![](images/092c5e22587610c44d5779aeeddae1244c84b0ad345d29f1c447a3f8cf21371b.jpg) + +![](images/1a0318c5289ee24aaaa4920ae838cbb51c40a1d2eac342b2a8ed14d2e4ef27ef.jpg) + +![](images/9d9dfe6188c2416cdb8c88b65ad09bb43e0eaf604c449feba991f0e2f2e67af7.jpg) + +![](images/87f939a61b87c9cbb85a10525bc3d9695301b2a98ac5ef9fe98ec289ae45d322.jpg) + +![](images/55ca2370efb06ed65873cdc9d01001b5e10be8cfb11fa97c51a41b9df7414233.jpg) + +![](images/7a96eb3508cc78e58a6d0abf721b0b0989b598107e5e8a6313ae0b137df89f63.jpg) +Light Them On Devastating Fire! + +![](images/ce1808793784b27066bded330f2aff2db3c4677d10a30c13bbb7402d63186cdb.jpg) + +Fig. 5: Qualitative Comparison on our Test Dataset. We conducted experiments on various scenarios, including human, architecture and landscape. + +![](images/6ca9d4c176cb6f732c529c1a1bc83f78f3dcd0b851847e57ae8ea5815d6beda8.jpg) + +![](images/844b5f9f0edecc5f02b4b1aa9cc423130025b98c8da87b609b77608f7208a46b.jpg) + +![](images/dd0bdf10146d061430688e52f31198e1f698449455b12aadd213c4d6d5876f91.jpg) + +![](images/e7e116bea32dfc9e89eed9714def87901d951f4f346d784696af79b2eef56644.jpg) + +that involves over 12,000 filtered images and 2,000 filtered editing instructions. All single input images have a resolution of $512 \times 512$ and are resized to $256 \times 256$ for training and testing purposes. + +Implementation Details. In our approach, we split the dataset with $80\%$ for training and $20\%$ for testing. As demonstrated in Fig. 2, two image pairs are concatenated into one $512 \times 512$ image with the same editing instructions as $\mathbf{Img}^{\mathrm{train}}$ and mask the fourth quadrant with a grey color as $\mathbf{Img}^{\mathrm{VPcon}}$ . We prepare the in-domain test dataset in the same format. For out-of-domain testings, we ensure that both the visual instruction pairs and the text instructions are not used during training, to best simulate how models perform in real-life image editing generalization scenarios. + +For baselines, we use their original configurations to train their model. In our method, we only fine-tune the additional ControlNet for 5000 steps with a batch size of 1024 and a learning rate at $1 \times 10^{-4}$ . During training, we set the classifier-free scale the same as the original ControlNet. And we randomly drop $15\%$ of language or visual editing instructions to further enhance our model's + +Table 1: Quantitative results comparison between our method and baseline methods. Quantitative results show our method excels the baseline methods in both FID and CLIP directional Similarity score in a great margin. + +
SDEdit (E)InstructPix2PixPromptDiffusionOurs
FID ↓21.6717.8713.757.57
CLIP DirSim ↑0.110.170.210.27
+ +generalization ability. Our implementation utilizes PyTorch and is trained on 4 Tesla A100-40G GPUs with AdamW optimizer. + +Comparison Methods. To evaluate the effectiveness of our work, we compare with other state-of-the-art image editing frameworks, including SDEdit [28], Instruct-Pix2pix [3] and PromptDiffusion [41]. We adopt two quantitative metrics Fr'echet Inception Distance (FID) and CLIP directional Similarity (CLIP DirSim) proposed by Gal et al. [12]. We utilize the FID score to quantitatively represent the image quality of generated editing outputs, and CLIP DirSim to evaluate how well the models follow editing instructions to produce the output. + +# 6.2 State-of-the-Art Comparisons + +Qualitative Evaluation. In Fig. 5, we present our qualitative results in the testing set (in domain). The comparison shows that our method surpass previous baseline methods. Our method understands and follows both visual and language editing instructions better, and produces a far more detailed manipulated output especially in human figures. In Fig. 6, we present our qualitative results tested in out-of-domain settings. We include five editing instructions that are considered extremely hard for diffusion-based image editing model [18], including object add, object remove, structure location change, structure action change, and object size change. It is important to note that due to how we generate our training dataset utilizing Prompt-to-Prompt [15], editing images pairs with these types of editing instructions is not feasible to generate our training data. This generalization comparison shows that our method excels baseline methods by a significant margin. Our method shows a great capability to carry out well-detailed quality outputs following these hard editing instructions in diffusion-based image editing models, while other baseline methods all fail to understand the editing instructions well or perform manipulations close to the editing instructions. + +Quantitative Evaluation. In quantitative evaluation, we score 7.57 in FID, better than SDEdit (E), InstructPix2Pix and PromptDiffusion which scores 21.67, 17.87 and 13.75. We achieve 0.27 in CLIP DirSim, better than baselines with $0.11 / 0.17 / 0.21$ of CLIP DirSim scores. These quantitative findings show that our method generates higher-quality images with better detailed qualities and also exactly follows both language and visual editing instructions. + +Ablation study. We conduct an ablation study on each of the four components of our proposed method. Namely, the Reformed Conditioned Latent Diffusion + +![](images/9b1a8b3f5261250f98c783b25eccfbe31d5fd0bfed21837faff2b3367ba6b160.jpg) +Add Some Snow + +![](images/99d1dcfd357672b24d81322483cf5088e5b9be75663d07fdafa2f3e8fbfdd6fd.jpg) + +![](images/3756a06f72ab93543c972051d6c4be5b0d870abd68255c9bc54205cf688411b5.jpg) + +![](images/d833da5ab0c4172b98c118ff38364ef7549add0cfe8023f0f8b76fd839b5c085.jpg) + +![](images/797926c8e1b5c902225aaae83f05ec2f05c95c671c2deec57e67316b32595c79.jpg) + +![](images/ab4878c744f07d4e4f4a64fbbd2cfa94e27c40ee5904c2d2dd40db2166e77c14.jpg) + +![](images/e1652f09b079c909f62d270f254730f3e158f4f97e83b6d35335c7bb168c829a.jpg) +Bend His Arm + +![](images/012f19323ab9e590912012e26ff71cd5d12c176bd3fcba6836f034c062a68d62.jpg) + +![](images/9560d3f1592a7e353ba4de48a8fce01dfa0f59ec7242998a42470e332306f668.jpg) + +![](images/d5fea7382be0670f5660fd1e2806dd1d1d3f8fdb86a6816c4f5a0e2baa929e69.jpg) + +![](images/b51ef89bf099d75b3744f636e901729534c40b4f0184e0efc185c7fad304e046.jpg) + +![](images/eeab39df57bc1ebe4fe2ad600987a9b9bce9d0cdf112ec165e286c50d5c8efdb.jpg) + +![](images/e03509a783da7fb55ad8ac482e7831e22ca4a1a857d22dcbd6da06e8cfc56151.jpg) +Move It to the Right + +![](images/33096d3c36a7c2fd9f30fbf9d6c2cf4aa7e5cc07ab2a481ec8df1430101aa095.jpg) + +![](images/943b5fc4af426ec1e5392f045af78087f90ace3b5ecfa58fbda23027e81ca4f6.jpg) + +![](images/1d708ea75de81e7af6d8087b27c23007baa369a05ce3763d82b503b50ce1f0c3.jpg) + +![](images/13c3b8dd5458aec95a6b141b0547b69a115571e79ecb8033179f24eae69bc8ca.jpg) + +![](images/d8bdd96b446ebe407c0a50ae9a1926dd702a77746fac67ddd6414425054c471f.jpg) + +![](images/32d7e0370e93e733f256b94210909ca4de698e8842d7510fb2f4709c23877235.jpg) +Keep Only One + +![](images/45fe24e1b3caf71aeb773019b4b125c7bec1877a3b12cc2f8120e6af8c041857.jpg) + +![](images/28d0d92e67a7f47c855c1f7a1b861c22c812911bcaf11ae3594a6ecb1d95737e.jpg) + +![](images/95e274f915cfadaf716efec172f2fc007ab44bff7ae0a46e017af779a75cec9f.jpg) + +![](images/3baac35ba9da06666f5181ae26f7523aefd9c019601c15a98c22b1d0d9c68047.jpg) + +![](images/061f0f591285299418ae2abf9ee73ef0c8260f77ddfac0dcce9a3ed7f0912d14.jpg) + +![](images/1fbff72ce8e03f5cf2beb0e0d2dc12df8f4e820029f2b9f7b312982eab92ad1e.jpg) +Make It Smaller + +![](images/e765032929ade0c0ad1167c5c5936f9dd7464708df3db3ad6086ea0c3a6d8a36.jpg) +Fig. 6: Qualitative Comparison on Out-of-Domain Images. We conducted experiments on images and instruct that are not in training dataset. + +![](images/ac231f8e32ab2c7cff80dbaf39eb6dc5b6ed2fadca375e354fc56156c29b0e3c.jpg) + +![](images/6198413c95eeff30c7552d7b8246dea47d0370a4b84e493102d83b0637063f29.jpg) + +![](images/1732f0685612529675eed4fea1e2ac58d6e3c9a2330cc6e2583bb156b79c98f7.jpg) + +![](images/0c69def8584cf15bfda75e026975622d217c5f78a330ab3a73d694c1ab8180c4.jpg) + +Model (RCLDM), Editing Shift Matching (ESM), Language Instruction Unification (LIU) and Selective Area Matching (SAM). We present the qualitative results in Fig. 7. From the qualitative results, we can see that without RCLDM and ESM, the model understands the visual editing instructions much weaker, especially in out-of-domain editing. Without LIU, for two language editing instructions with the same meaning, the model produces two output edited images in different detail and quality. This difference in quality tends to increase in out-of-domain settings. Without SAM, the details of the human face are distorted making the model more vulnerable to producing outputs with worse detailed qualities. We also include the qualitative ablation results on the testing dataset in Tab. 2. From observation, our method performs the best when incorporating all four components. Without SAM or LIU, the FID score increases, meaning those modules enhance the detail quality of the output generated. Without CLDM or ESM, the CLIP DirSim score decreases, showing that those two modules contribute to a better understanding in both language and visual level. + +![](images/082dfed84ce85344d81271af58b8bd5ba51fb090989a06b30344d5a06fa7e781.jpg) +Make It in Winter Time + +![](images/4516c3aa8231a82b3e917500fa4ce33387ce852b5d0f995ad628c7cb48115536.jpg) + +![](images/2ac1ede521065bc6be8185c74650c1a75a75254c1f8064e859553d4c24291b6a.jpg) + +![](images/dd533610c6ab99385d42506ad2d07b70640f702702acf25cd3a8bd3a5ce910e4.jpg) + +![](images/ab2f1684e0015fab2c381669c266a6fe00fd3c5a7c04264473c307d94d4c0588.jpg) + +![](images/be79d7f66060746e7965bfa2e9c8316ca7f0a44f78dc9ba448eefa813a8f19f9.jpg) + +![](images/b7a9092fa842d1e60c4a0c1be6e98baa5698343b94c1f64ee0f3cea0fd492030.jpg) + +![](images/dd15423423a2b17de3a6bf03986dc5b55de89166d1cc3ba1599e4901ba993c06.jpg) + +![](images/deffb905bdb4c184a2a5d6b604bea4a756498e18cb8a53d10a6e32ea5ef4554e.jpg) + +![](images/55ba8916f9ea4c0cf3c66de6e0083d2f77ff953772a5d41e667c9e2ef34a39c8.jpg) + +![](images/365374894182f105968bc6bd785f343786d8feea6c2b74abe279dd7f7bca8d0c.jpg) + +![](images/d16b8cc6e61d2eb5d8e4aec4ec07b31e647678fdda58c1c50509067fa17f3ccb.jpg) + +![](images/14eaf355a89442b992c7ea979fb63fb0b087181a36b25b608fc0c1df99da0c53.jpg) + +![](images/aa8ee905cddf6ff10dfe10116542105c63737916cf609f7d488c63a52ba0ea65.jpg) + +![](images/af4b6e6cae789bbe9d49613b9f3f3a246c0be79aeffb448fb5cef5823190b852.jpg) +Turn It to Winter Time +Make Emotion Cuter + +![](images/c66ae89a6a8613b54338b96d6e606533f6b8bcf44c27a05754b430e64e9f5707.jpg) + +![](images/d21adcd1f3ce5b6827141d0ee2983d49cf81ed28c9fd5e8167a645d93af71d19.jpg) + +![](images/32a37f0cd5345c4398dbc102a13cd0f7acfd4d7e9c3c6bf353a687055fccda08.jpg) + +![](images/6c0dd084fe17ee444ade75f0942d8c7cfa8ffc09a254c6e3d21a31efe6592fb2.jpg) + +![](images/6ac2f3fe89e5754757a57deb6cb3e5f3483ce585f81633c05d512db48f99b5dc.jpg) + +![](images/285f2962961afa9dcb0489d9d636db5f96227bfd7070e70f61992e38fe29d0ac.jpg) + +![](images/16b6014d961b8b145318bf832d3f61d44de62776020b5c0e888d59fda80e9363.jpg) +Cuter Emotion +Fig. 7: Ablation Study Results for both in-domain (first two rows), and out-of-domain first two rows) image manipulations + +![](images/9956882be333de8fc4bcd4d056c7b61fa64ff3941c5c46d83d214c8c86f52d41.jpg) + +![](images/1e38f74099ac8b8b755e791e0db43a3e6895b754b502306f5e6e1c993194d62a.jpg) + +![](images/c88305bdd51f8c03e042473ad850b1c3b01e822f218c3f985a850716ee503d09.jpg) + +![](images/5e809d829b6b4eb4d973f8ba5ee685c671e5082ac363545ca39790ba525a7eef.jpg) + +![](images/0b45f0363aa469f07c35def0283f7e48e83d4046adddca4baab0535749779c79.jpg) + +![](images/924aa9236d63901c0339c7872f03f266a6e61668d908244cd7d8981898549db4.jpg) + +Table 2: Abalation Study results of comparison between our method with all components and without the Reformed Conditioned Latent Diffusion Model (RCLDM), Editing Shift Matching (ESM), Language Instruction Unification (LIU) and Selective Area Matching (SAM). The ablation study is conducted on the entire test dataset. + +
w/o RCLDMw/o ESMw/o LIUw/o SAMOurs
FID ↓10.159.2310.3711.317.57
CLIP DirSim ↑0.130.150.230.190.27
+ +# 7 Conclusion + +In this work, we propose InstructGIE, an image editing framework with enhanced generalization ability, improving performance in both visual and text aspects. We incorporate a VMamba-based module to enhance visual outputs and introduce an editing-shift matching strategy to augment in-context learning. Our selective area-matching technique addresses and rectifies corrupted details, while a language unification technique aligns language embeddings with editing semantics. Additionally, we compile a publicly available dataset for better generalization evaluation. Extensive experiments demonstrate our framework's superior in-context generation performance and robust generalization capability across unseen vision tasks, both quantitative and qualitively. + +# Acknowledgement + +This work is partially supported by the Army Research Office/Army Research Laboratory via grant W911-NF-20-1-0167 to Northeastern University, National Science Foundation CCF-1937500, and the Fundamental Research Funds for the Central Universities, Peking University. + +# References + +1. Bar, A., Gandelsman, Y., Darrell, T., Globerson, A., Efros, A.: Visual prompting via image inpainting. Advances in Neural Information Processing Systems 35, 25005-25017 (2022) +2. Blau, T., Ganz, R., Kawar, B., Bronstein, A., Elad, M.: Threat model-agnostic adversarial defense using diffusion models. arXiv preprint arXiv:2207.08089 (2022) +3. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023) +4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) +5. Chen, W., Hu, H., Sahara, C., Cohen, W.W.: Re-imagen: Retrieval-augmented text-to-image generator. arXiv preprint arXiv:2209.14491 (2022) +6. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290–1299 (2022) +7. Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8188-8197 (2020) +8. Chung, J.S., Zisserman, A.: Lip reading in the wild. In: Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13. pp. 87-103. Springer (2017) +9. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017) +0. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021) +11. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021) +12. Gal, R., Patashnik, O., Maron, H., Chechik, G., Cohen-Or, D.: Stylegan-nada: Clipguided domain adaptation of image generators. arXiv preprint arXiv:2108.00946 (2021) +13. Geng, X., Liu, H.: Openllama: An open reproduction of llama (May 2023), https://github.com/openlm-research/open_llama +14. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022) + +15. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022) +16. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) +17. Huang, Y., Huang, J., Liu, Y., Yan, M., Lv, J., Liu, J., Xiong, W., Zhang, H., Chen, S., Cao, L.: Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525 (2024) +18. Huang, Y., Huang, J., Liu, Y., Yan, M., Lv, J., Liu, J., Xiong, W., Zhang, H., Chen, S., Cao, L.: Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525 (2024) +19. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019) +20. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. Advances in Neural Information Processing Systems 35, 23593-23606 (2022) +21. Kawar, B., Ganz, R., Elad, M.: Enhancing diffusion-based image synthesis with robust classifier guidance. arXiv preprint arXiv:2208.08664 (2022) +22. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023) +23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Pereira, F., Burges, C., Bottou, L., Weinberger, K. (eds.) Advances in Neural Information Processing Systems. vol. 25. Curran Associates, Inc. (2012) +24. Kwon, G., Ye, J.C.: Diffusion-based image translation using disentangled style and content representation. arXiv preprint arXiv:2209.15264 (2022) +25. Liu, X., Park, D.H., Azadi, S., Zhang, G., Chopikyan, A., Hu, Y., Shi, H., Rohrbach, A., Darrell, T.: More control for free! image synthesis with semantic diffusion guidance. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 289-299 (2023) +26. Liu, Y., Tian, Y., Zhao, Y., Yu, H., Xie, L., Wang, Y., Ye, Q., Liu, Y.: Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166 (2024) +27. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of International Conference on Computer Vision (ICCV) (December 2015) +28. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073 (2021) +29. Mohammad, S., Kiritchenko, S.: Wikiart emotions: An annotated dataset of emotions evoked by art. In: Proceedings of the eleventh international conference on language resources and evaluation (LREC 2018) (2018) +30. Nguyen, T., Li, Y., Ojha, U., Lee, Y.J.: Visual instruction inversion: Image editing via visual prompting. arXiv preprint arXiv:2307.14331 (2023) +31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) + +32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +33. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) +34. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +35. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015) +36. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems 32 (2019) +37. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020) +38. Sun, Y., Yang, Y., Peng, H., Shen, Y., Yang, Y., Hu, H., Qiu, L., Koike, H.: Imagebrush: Learning visual in-context instructions for exemplar-based image manipulation. arXiv preprint arXiv:2308.00906 (2023) +39. Theis, L., Salimans, T., Hoffman, M.D., Mentzer, F.: Lossy compression with gaussian diffusion. arXiv preprint arXiv:2206.08889 (2022) +40. Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6830-6839 (2023) +41. Wang, Z., Jiang, Y., Lu, Y., He, P., Chen, W., Wang, Z., Zhou, M., et al.: Incontext learning unlocked for diffusion models. Advances in Neural Information Processing Systems 36 (2024) +42. Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., Xiao, J.: Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365 (2015) +43. Zablotskaia, P., Siarohin, A., Zhao, B., Sigal, L.: Dwnet: Dense warp-based network for pose-guided human video generation. arXiv preprint arXiv:1910.09139 (2019) +44. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023) +45. Zimmermann, R.S., Schott, L., Song, Y., Dunn, B.A., Klindt, D.A.: Score-based generative classifiers. arXiv preprint arXiv:2110.00473 (2021) \ No newline at end of file diff --git a/2024/InstructGIE_ Towards Generalizable Image Editing/images.zip b/2024/InstructGIE_ Towards Generalizable Image Editing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a4b0a97c809136d1eff8c0ec6c45e9f20c35213c --- /dev/null +++ b/2024/InstructGIE_ Towards Generalizable Image Editing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ece540e50d46aa493a0929565387f3b2740220eead106dccc0d07d845a7aa1a +size 809198 diff --git a/2024/InstructGIE_ Towards Generalizable Image Editing/layout.json b/2024/InstructGIE_ Towards Generalizable Image Editing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..02c199aa669c28622b565d3aeb109e385d298caa --- /dev/null +++ b/2024/InstructGIE_ Towards Generalizable Image Editing/layout.json @@ -0,0 +1,12615 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 157, + 111, + 457, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 111, + 457, + 148 + ], + "spans": [ + { + "bbox": [ + 157, + 111, + 457, + 148 + ], + "type": "text", + "content": "InstructGIE: Towards Generalizable Image Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 139, + 168, + 475, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 168, + 475, + 194 + ], + "spans": [ + { + "bbox": [ + 139, + 168, + 475, + 194 + ], + "type": "text", + "content": "Zichong Meng\\*, Changdi Yang\\*, Jun Liu\\*, Hao Tang\\*\\*, Pu Zhao\\*\\*, and Yanzhi Wang\\*\\*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 201, + 201, + 411, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 201, + 411, + 213 + ], + "spans": [ + { + "bbox": [ + 201, + 201, + 411, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 201, + 201, + 411, + 213 + ], + "type": "text", + "content": " Northeastern University, Boston MA 02115, USA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 212, + 213, + 400, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 213, + 400, + 224 + ], + "spans": [ + { + "bbox": [ + 212, + 213, + 400, + 224 + ], + "type": "text", + "content": "2 Peking University, Beijing 100871, China &" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 194, + 225, + 419, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 225, + 419, + 236 + ], + "spans": [ + { + "bbox": [ + 194, + 225, + 419, + 236 + ], + "type": "text", + "content": "Carnegie Mellon University, Pittsburgh PA 15213, USA" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 146, + 270, + 468, + 511 + ], + "blocks": [ + { + "bbox": [ + 146, + 270, + 468, + 511 + ], + "lines": [ + { + "bbox": [ + 146, + 270, + 468, + 511 + ], + "spans": [ + { + "bbox": [ + 146, + 270, + 468, + 511 + ], + "type": "image", + "image_path": "1163a20c0ef3ae95a0a9ff40e4ae76f0a4ebc103dbef8ac4214961895c7a95c4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 512, + 180, + 517 + ], + "lines": [ + { + "bbox": [ + 147, + 512, + 180, + 517 + ], + "spans": [ + { + "bbox": [ + 147, + 512, + 180, + 517 + ], + "type": "text", + "content": "Like in Game" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 512, + 336, + 517 + ], + "lines": [ + { + "bbox": [ + 309, + 512, + 336, + 517 + ], + "spans": [ + { + "bbox": [ + 309, + 512, + 336, + 517 + ], + "type": "text", + "content": "Add a Cow" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 530, + 482, + 564 + ], + "lines": [ + { + "bbox": [ + 132, + 530, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 530, + 482, + 564 + ], + "type": "text", + "content": "Fig. 1: Demo results of the proposed InstructGIE framework on various image manipulation tasks to both humans and scenes. By our proposed method, our model can generalize to generate the desired output with great detail qualities." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 611, + 453, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 611, + 453, + 634 + ], + "spans": [ + { + "bbox": [ + 160, + 611, + 453, + 634 + ], + "type": "text", + "content": "Abstract. Recent advances in image editing have been driven by the development of denoising diffusion models, marking a significant leap" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 643, + 223, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 223, + 654 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 223, + 654 + ], + "type": "text", + "content": "* Equal Contribution" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 654, + 239, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 654, + 239, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 654, + 239, + 665 + ], + "type": "text", + "content": "\\*\\* Corresponding Authors" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 160, + 117, + 452, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 117, + 452, + 325 + ], + "spans": [ + { + "bbox": [ + 160, + 117, + 452, + 325 + ], + "type": "text", + "content": "forward in this field. Despite these advances, the generalization capabilities of recent image editing approaches remain constrained. In response to this challenge, our study introduces a novel image editing framework with enhanced generalization robustness by boosting in-context learning capability and unifying language instruction. This framework incorporates a module specifically optimized for image editing tasks, leveraging the VMamba block and an editing-shift matching strategy to augment in-context learning. Furthermore, we unveil a selective area-matching technique specifically engineered to address and rectify corrupted details in generated images, such as human facial features, to further improve the quality. Another key innovation of our approach is the integration of a language unification technique, which aligns language embeddings with editing semantics to elevate the quality of image editing. Moreover, we compile the first dataset for image editing with visual prompts and editing instructions that could be used to enhance in-context capability. Trained on this dataset, our methodology not only achieves superior synthesis quality for trained tasks, but also demonstrates robust generalization capability across unseen vision tasks through tailored prompts. Our project page is available at https://cr8br0ze.github.io/InstructGIE." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 337, + 437, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 337, + 437, + 348 + ], + "spans": [ + { + "bbox": [ + 160, + 337, + 437, + 348 + ], + "type": "text", + "content": "Keywords: Image Editing " + }, + { + "bbox": [ + 160, + 337, + 437, + 348 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 337, + 437, + 348 + ], + "type": "text", + "content": " In-Context Learning " + }, + { + "bbox": [ + 160, + 337, + 437, + 348 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 337, + 437, + 348 + ], + "type": "text", + "content": " Diffusion Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 372, + 230, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 372, + 230, + 385 + ], + "spans": [ + { + "bbox": [ + 133, + 372, + 230, + 385 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 402, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 402, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 133, + 402, + 481, + 521 + ], + "type": "text", + "content": "As a crucial task in computer vision, image editing has witnessed significant improvements enhanced with the increasingly popular denoising stable diffusion techniques in recent years [22,25,38,40,41]. Given a set of text or image prompts as generation constraints or instructions, diffusion-based image editing can follow the instructions and synthesize an edited image. However, since the model does not have the capability to accurately model all possible samples in the conditional distribution space [17], if specific instructions are not included in the training dataset, current diffusion-based image editing methods can hardly generate satisfactory results. Thus, editing performance largely depends on the training dataset without superior generalization capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 522, + 481, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 522, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 133, + 522, + 481, + 664 + ], + "type": "text", + "content": "On the other hand, large language models (LLMs) have proven extraordinary abilities to learn from contexts, referred to as in-context learning, which allows LLMs to perform unseen tasks by providing a combination of input-output examples and a query input. Inspired by the potential to enhance the generalization of the model with LLMs, [38,41] explore in-context learning for computer vision tasks, allowing them to perform unseen tasks with novel vision-language prompt designs. However, these methods are not tailored for image editing applications, leading to unsatisfying synthetic qualities with inaccurate or incorrect output and lack of detail. To improve the generalization of image editing with improved synthetic image quality, it is crucial to effectively understand the text & image prompts and specifically control image editing details, which is challenging in the current literature." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "text", + "content": "Z. Meng and C. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 331 + ], + "type": "text", + "content": "In this work, we propose InstructGIE, an image editing framework with enhanced generalizability. We improve image editing performance from both visual and text aspects. (i) For the visual information, we incorporate a VMamba-based module to specifically enhance the image editing outputs. As VMamba [26] has proven its better performance in capturing in-context information from inputs with larger receptive fields [26], we leverage VMamba and propose an editing-shift matching strategy to augment in-context learning. Furthermore, since current image editing works do not perform well in generating correct features with accurate details, we unveil a selective area-matching technique specifically engineered to address and rectify corrupted details in generated images, such as human facial features, to further improve the quality. (ii) Another key innovation of our approach is the integration of a language unification technique, which aligns language embeddings with editing semantics to elevate the quality of image editing. Our framework not only achieves superior in-context generation for trained tasks but also demonstrates robust generalization across unseen vision tasks. Moreover, we compile a publicly available image editing dataset with plenty of visual prompts and editing instructions for better generalization evaluation of image editing. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 338, + 480, + 420 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 338, + 479, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 338, + 479, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 338, + 479, + 373 + ], + "type": "text", + "content": "- We propose an image editing framework, including in-context learning enhancement and language unification strategies, specifically designed to enhance generalization ability from both visual and text domains." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 374, + 480, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 480, + 396 + ], + "type": "text", + "content": "- We compile the first dataset for image editing with visual prompts and editing instructions that could be used to enhance in-context capability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 397, + 480, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 397, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 480, + 420 + ], + "type": "text", + "content": "- We conduct extensive experiments and achieve great generalization ability in the multiple unseen image editing task, both quantitatively and qualitatively." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 445, + 242, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 445, + 242, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 445, + 242, + 456 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 468, + 402, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 402, + 480 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 402, + 480 + ], + "type": "text", + "content": "2.1 Denoising Stable Diffusion Based Image Editing" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 486, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 486, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 486, + 481, + 665 + ], + "type": "text", + "content": "Denoising Stable Diffusion [16,35,36] based image editing could follow guidance from text or image prompts. With the foundation of text-guided models offering rich generative capabilities, there has been a surge in research aimed at adapting these models for image manipulation tasks from textual descriptions. To steer the image editing process in the desired direction, the use of models like CLIP to fine-tune diffusion models has become a common practice. Although these methods [3,22,41] have shown impressive results, they often involve costly finetuning processes. Recent innovations [14] have introduced techniques that inject cross-attention into the models to more effectively edit specific semantic areas within the spatial feature maps. Further advancements [24] have enhanced these techniques by adding semantic loss or applying attention loss to refine the integration of plugged features, improving the precision and quality of the editing outcomes. [38] proposes a framework that could learn instructions from visual image pairs for more accurate editing and firstly formulate this task as an image inpainting problem." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 239, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 239, + 91, + 447, + 102 + ], + "type": "text", + "content": "InstructGIE: Towards Generalizable Image Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 370, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 370, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 370, + 129 + ], + "type": "text", + "content": "2.2 Generalization Capability in Visual Tasks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 292 + ], + "type": "text", + "content": "In-context learning is widely applied in the field of natural language processing (NLP), enabling models to adapt to new tasks such as translation, question answering, and complex reasoning. NLP models utilize in-context examples, comprising text and corresponding labels, to tackle tasks they haven't seen before. However, applying in-context learning to the visual domain introduces more challenges and remains less explored. A significant hurdle is the nature of fixed-size input requirements for vision models, as opposed to variable-length text inputs that can be managed by language models. Vision models generally struggle with processing inputs of varying sizes, making it impractical to process multiple image prompts in one-shot for global understanding. Moreover, in intricate visual understanding, specific instructions are often implied from a limited set of image examples rather than explicitly stated, which poses additional difficulties for vision models in identifying and understanding high-level visual relationships." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 293, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 399 + ], + "type": "text", + "content": "Recent strides in applying masked image modeling have marked a step forward in improving in-context learning for vision models. The method proposed by [41], employing a masked autoencoder-based technique, predicts a missing image within a two-by-two grid, using two images as in-context examples and another as the query. This concept was later expanded by [38] with a multitask framework. Despite their progress, such inpainting methods are limited by the necessity of a fixed number of in-context examples and increased memory demands. Painter, highlighted in [40], exemplifies an inpainting approach tailored for versatility across various vision tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 399, + 481, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 481, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 481, + 448 + ], + "type": "text", + "content": "In contrast, inspired by ControlNet [44], [41] refines the framework by adding an additional pair of example images and employing a multitask supervised finetuning method. Prompt diffusion excels in visual in-context learning. However, it faces certain limitations or challenges in its practical applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 465, + 373, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 465, + 373, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 465, + 373, + 478 + ], + "type": "text", + "content": "2.3 Dataset for Diffusion-based Image Editing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Currently, various types of datasets are used for training in diffusion-based image editing. There are datasets that concentrate on specific domains like CelebA [27] and FFHQ [19] for human face image manipulation, AFHQ [7] for animal face image editing, LSUN [42] for object modification, and WikiArt [29] for style transfer. In-the-wild video datasets could also be leveraged to train image editing tasks. The Scannet dataset [9] encompasses a vast array of more than 1,500 indoor scenes from various settings, such as apartments, offices, and hotels, providing extensive annotations. The LRW dataset [8], tailored for lip reading tasks, includes more than 1000 video utterances of 500 distinct words. The UBC-Fashion dataset [43] features 600 videos spanning various clothing categories, with 500 videos allocated for training and 100 for testing, guaranteeing no repetition of individuals in the training set. The DAVIS dataset [43] (Densely Annotated VVideo Segmentation), a widely recognized benchmark for video object segmentation, contains 150 videos in total. There are also image editing works proposing to generate datasets with editing instructions. InstructPix2pix [3] collects over" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "text", + "content": "Z. Meng and C. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "450,000 training image pairs. For each pair, given an image with its caption, it first uses a finetuned GPT-3 [4] to generate an editing instruction and an edited image caption. Then it employs Stable Diffusion and the Prompt-to-Prompt algorithm [14] to generate edited image following the caption. However, currently there are no datasets with multiple image pairs under one editing instruction, which is crucial to enhance the generalization ability of image editing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 209, + 226, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 209, + 226, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 226, + 224 + ], + "type": "text", + "content": "3 Preliminary" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 344 + ], + "type": "text", + "content": "Recent advances in generative models have been significantly driven by the emergence of diffusion models, which have set new benchmarks in image creation [10,21,37]. These models have found applications across a broad spectrum of areas [2,5,20,39,45], demonstrating their versatility and effectiveness. The fundamental concept behind diffusion models involves starting with an image that is initially just random noise " + }, + { + "bbox": [ + 130, + 236, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T\\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 130, + 236, + 482, + 344 + ], + "type": "text", + "content": " and progressively refining this image step by step until it becomes a high-quality, realistic image " + }, + { + "bbox": [ + 130, + 236, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 236, + 482, + 344 + ], + "type": "text", + "content": ". This refinement process involves generating intermediate samples " + }, + { + "bbox": [ + 130, + 236, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 130, + 236, + 482, + 344 + ], + "type": "text", + "content": " (for " + }, + { + "bbox": [ + 130, + 236, + 482, + 344 + ], + "type": "inline_equation", + "content": "t\\in \\{0,\\dots ,T\\}" + }, + { + "bbox": [ + 130, + 236, + 482, + 344 + ], + "type": "text", + "content": "), where each sample is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 251, + 354, + 481, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 354, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 251, + 354, + 481, + 369 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t} = \\sqrt {\\alpha_ {t}} \\mathbf {x} _ {0} + \\sqrt {1 - \\alpha_ {t}} \\boldsymbol {\\epsilon} _ {t}, \\tag {1}", + "image_path": "4225ad4dbaedab59152de7e1e30fc9ba36e94f9e88bd0263d5ae30f0b7c54909.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "text", + "content": "where the parameter " + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "text", + "content": " sets the pace of the diffusion process, ranging from " + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "inline_equation", + "content": "0 = \\alpha_{T} < \\alpha_{T - 1} < \\dots < \\alpha_{1} < \\alpha_{0} = 1" + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\epsilon_t\\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "text", + "content": " represents the added noise. The model refines the image by applying a neural network " + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\mathbf{x}_t,t)" + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "text", + "content": " to each sample " + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "text", + "content": ", followed by the addition of Gaussian noise to produce the next sample " + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t - 1}" + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "text", + "content": ". This neural network is optimized to achieve a denoising goal, striving for " + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\mathbf{x}_t,t)\\approx \\epsilon_t" + }, + { + "bbox": [ + 130, + 377, + 482, + 461 + ], + "type": "text", + "content": ", resulting in a generative process that closely mimics the desired image distribution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 582 + ], + "type": "text", + "content": "Expanding this framework to conditional generative modeling, the process involves conditioning the neural network " + }, + { + "bbox": [ + 130, + 462, + 482, + 582 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\mathbf{x}_t,t,\\mathbf{y})" + }, + { + "bbox": [ + 130, + 462, + 482, + 582 + ], + "type": "text", + "content": " on an additional input " + }, + { + "bbox": [ + 130, + 462, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 462, + 482, + 582 + ], + "type": "text", + "content": ", enabling the generation of images from a distribution conditioned on " + }, + { + "bbox": [ + 130, + 462, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 462, + 482, + 582 + ], + "type": "text", + "content": ". This conditional input could be anything from a low-resolution image, a category label, or a descriptive text sequence. Leveraging the advancements in LLMs [33] and hybrid vision-language models [31], text-to-image diffusion models are developed. These models allow for the creation of detailed, high-resolution images from mere text descriptions, starting with a low-resolution image generated through the diffusion process, which is subsequently refined into a high-resolution image using additional models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 601, + 289, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 289, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 289, + 616 + ], + "type": "text", + "content": "4 The Proposed Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "We present our framework pipeline in Fig. 2. For efficient training and better controllability, we adopt a line of techniques with popular architectures such as ControlNet [44] and Stable Diffusion [34] to design a generalizable image editing" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 239, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 239, + 91, + 448, + 103 + ], + "type": "text", + "content": "InstructGIE: Towards Generalizable Image Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 469, + 338 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 469, + 338 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 469, + 338 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 469, + 338 + ], + "type": "image", + "image_path": "558d61ed0c6aa5cec6856d32f1f5db211c09dff21ca1d6c7eb3b3e53c71e943a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 347, + 482, + 426 + ], + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 426 + ], + "type": "text", + "content": "Fig. 2: Overall architecture of InstructGIE. The lower pipeline is for both training and inference processes where the model obtains unified editing instructions outputted by Instruction Unification Module " + }, + { + "bbox": [ + 130, + 347, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 130, + 347, + 482, + 426 + ], + "type": "text", + "content": " and combines with visual prompted input " + }, + { + "bbox": [ + 130, + 347, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{Img}^{\\mathrm{VPcon}}" + }, + { + "bbox": [ + 130, + 347, + 482, + 426 + ], + "type": "text", + "content": " to pass through Zero-VMamba integrated Stable Diffusion model with ControlNet for output image. The upper pipeline is for training only which compares output image and training ground truth " + }, + { + "bbox": [ + 130, + 347, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{Img}^{\\mathrm{train}}" + }, + { + "bbox": [ + 130, + 347, + 482, + 426 + ], + "type": "text", + "content": " and computes editing shift loss " + }, + { + "bbox": [ + 130, + 347, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{es}" + }, + { + "bbox": [ + 130, + 347, + 482, + 426 + ], + "type": "text", + "content": " with Editing Shift Module and selective area matching loss " + }, + { + "bbox": [ + 130, + 347, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sam}" + }, + { + "bbox": [ + 130, + 347, + 482, + 426 + ], + "type": "text", + "content": " with Selective Area Matching Module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "text", + "content": "tool with accurate high-quality outputs. Specifically, we introduce enhanced in-context learning both at the architecture level and training level to improve the image quality. Furthermore, language instruction unification is adopted to maximize the generalization ability for unseen editing tasks. Moreover, selective area matching is proposed to further improve the output quality with full details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 529, + 317, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 529, + 317, + 541 + ], + "spans": [ + { + "bbox": [ + 132, + 529, + 317, + 541 + ], + "type": "text", + "content": "4.1 Enhanced In-Context Learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 550, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 482, + 612 + ], + "type": "text", + "content": "Visual prompting based on inpainting is an effective visual in-context learning method in various computer vision tasks [1,40], which is applied in image editing tasks [38] recently. However, the methods perform poorly in quality when dealing with unseen image manipulation tasks. Therefore, we propose the enhanced in-context learning specifically tailored for generalizable image editing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Reforming Conditioned Latent Diffusion Model. To improve the generalization ability of image editing, it is crucial for the framework to explicitly capture low-level visual editing contexts. Current diffusion-based image editing" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "text", + "content": "Z. Meng and C. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 114, + 251, + 207 + ], + "blocks": [ + { + "bbox": [ + 157, + 114, + 251, + 207 + ], + "lines": [ + { + "bbox": [ + 157, + 114, + 251, + 207 + ], + "spans": [ + { + "bbox": [ + 157, + 114, + 251, + 207 + ], + "type": "image", + "image_path": "d3da15b41880757d9c95fae96f376efce5c9ef400fe4063e1481d597a777a623.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 209, + 233, + 217 + ], + "lines": [ + { + "bbox": [ + 174, + 209, + 233, + 217 + ], + "spans": [ + { + "bbox": [ + 174, + 209, + 233, + 217 + ], + "type": "text", + "content": "ERF of ResNet-50" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 260, + 114, + 354, + 208 + ], + "blocks": [ + { + "bbox": [ + 260, + 114, + 354, + 208 + ], + "lines": [ + { + "bbox": [ + 260, + 114, + 354, + 208 + ], + "spans": [ + { + "bbox": [ + 260, + 114, + 354, + 208 + ], + "type": "image", + "image_path": "71e7689d683d74dc0c08647a3d29f4ba934dbf57742ff9c21a5ad312fa416d2d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 209, + 334, + 217 + ], + "lines": [ + { + "bbox": [ + 279, + 209, + 334, + 217 + ], + "spans": [ + { + "bbox": [ + 279, + 209, + 334, + 217 + ], + "type": "text", + "content": "ERF of VIT-L/14" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 228, + 481, + 251 + ], + "lines": [ + { + "bbox": [ + 131, + 228, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 481, + 251 + ], + "type": "text", + "content": "Fig. 3: Effective Reception Field (ERF) of ConvNet, ViT, VMamba based model architectures." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 362, + 114, + 458, + 208 + ], + "blocks": [ + { + "bbox": [ + 362, + 114, + 458, + 208 + ], + "lines": [ + { + "bbox": [ + 362, + 114, + 458, + 208 + ], + "spans": [ + { + "bbox": [ + 362, + 114, + 458, + 208 + ], + "type": "image", + "image_path": "f9ed6567ff258182b8f8926cccb22cfd59509a8d97c2aa0268ed804cda3d13d8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 379, + 209, + 440, + 217 + ], + "lines": [ + { + "bbox": [ + 379, + 209, + 440, + 217 + ], + "spans": [ + { + "bbox": [ + 379, + 209, + 440, + 217 + ], + "type": "text", + "content": "ERF of VMamba-T" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 273, + 479, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 273, + 479, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 479, + 344 + ], + "type": "text", + "content": "methods [30, 38] that involve visual prompting either adopt ConvNet [23] or ViT [11] as the vision encoder for visual prompts. However, these methods fail to generalize well as they are not able to capture enough visual editing contexts (see Fig. 3). To address this, we formulate the visual prompted condition as a single image " + }, + { + "bbox": [ + 130, + 273, + 479, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{Img}^{\\mathrm{VPcon}} = \\{\\mathbf{Img}_0^{\\mathrm{in}}, \\mathbf{Img}_0^{\\mathrm{out}}, \\mathbf{Img}_1^{\\mathrm{in}}, \\mathbf{Grey}\\}" + }, + { + "bbox": [ + 130, + 273, + 479, + 344 + ], + "type": "text", + "content": ", as shown in Fig 2, with a global effective receptive field (ERF)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 345, + 480, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 480, + 417 + ], + "type": "text", + "content": "Moreover, we further propose to reform the conditioned latent diffusion model Inspired by the recent visual state space model VMamba [26] which exhibits a better global ERF and also emphasizes shifting boundaries of input's four quadrants, we propose to adopt a vision encoder based on Zero-VMamba to fit our structure. Specifically, the vision encoder " + }, + { + "bbox": [ + 130, + 345, + 480, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{ven}" + }, + { + "bbox": [ + 130, + 345, + 480, + 417 + ], + "type": "text", + "content": " comprehends the visual prompted condition in latent space " + }, + { + "bbox": [ + 130, + 345, + 480, + 417 + ], + "type": "inline_equation", + "content": "x_{vpc}" + }, + { + "bbox": [ + 130, + 345, + 480, + 417 + ], + "type": "text", + "content": " as follows," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 257, + 424, + 480, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 424, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 257, + 424, + 480, + 437 + ], + "type": "interline_equation", + "content": "e _ {v p c} = \\mathcal {G} _ {v e n} \\left(x _ {v p c}; \\Theta_ {g}\\right), \\tag {2}", + "image_path": "9a5d6c6c38962a70cb3b7cd425f3d320ad32208b9e0e25e0ca6ead8790ed976b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 443, + 479, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 443, + 479, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 479, + 467 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 443, + 479, + 467 + ], + "type": "inline_equation", + "content": "e_{vpc}" + }, + { + "bbox": [ + 130, + 443, + 479, + 467 + ], + "type": "text", + "content": " is the processed embedding of the visual prompted condition, and " + }, + { + "bbox": [ + 130, + 443, + 479, + 467 + ], + "type": "inline_equation", + "content": "\\Theta_g" + }, + { + "bbox": [ + 130, + 443, + 479, + 467 + ], + "type": "text", + "content": " is the model parameters initialized to zeros." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 468, + 480, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 480, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 480, + 538 + ], + "type": "text", + "content": "To further improve the performance, after each ControlNet trainable copied modules " + }, + { + "bbox": [ + 130, + 468, + 480, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_c" + }, + { + "bbox": [ + 130, + 468, + 480, + 538 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 130, + 468, + 480, + 538 + ], + "type": "inline_equation", + "content": "\\Theta_c" + }, + { + "bbox": [ + 130, + 468, + 480, + 538 + ], + "type": "text", + "content": ", we propose to link and inject the processed visual prompted condition information to the frozen Stable Diffusion model " + }, + { + "bbox": [ + 130, + 468, + 480, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 468, + 480, + 538 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 130, + 468, + 480, + 538 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 468, + 480, + 538 + ], + "type": "text", + "content": " through zero-VMamba layer " + }, + { + "bbox": [ + 130, + 468, + 480, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 468, + 480, + 538 + ], + "type": "text", + "content": ". We use two instances of VMamba with parameters " + }, + { + "bbox": [ + 130, + 468, + 480, + 538 + ], + "type": "inline_equation", + "content": "\\theta_{g1}" + }, + { + "bbox": [ + 130, + 468, + 480, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 468, + 480, + 538 + ], + "type": "inline_equation", + "content": "\\theta_{g2}" + }, + { + "bbox": [ + 130, + 468, + 480, + 538 + ], + "type": "text", + "content": " respectively. The complete model then computes the following," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 195, + 546, + 480, + 560 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 546, + 480, + 560 + ], + "spans": [ + { + "bbox": [ + 195, + 546, + 480, + 560 + ], + "type": "interline_equation", + "content": "y _ {v p c} = \\mathcal {F} (x; \\Theta) + \\mathcal {G} (\\mathcal {F} (x + \\mathcal {G} (x _ {v p c}; \\Theta_ {g 1}); \\Theta_ {c}); \\Theta_ {g 2}), \\tag {3}", + "image_path": "e64f4653f86f3a8c013d29615e711f3d05e88eb92441302b60e19c3d4c8d8070.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 566, + 422, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 566, + 422, + 578 + ], + "spans": [ + { + "bbox": [ + 131, + 566, + 422, + 578 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 566, + 422, + 578 + ], + "type": "inline_equation", + "content": "y_{vpc}" + }, + { + "bbox": [ + 131, + 566, + 422, + 578 + ], + "type": "text", + "content": " is the output of our conditioned diffusion model block." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 578, + 479, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 479, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 479, + 613 + ], + "type": "text", + "content": "Our conditioned latent diffusion model can process all four quadrants in our visual prompted conditions with a global receptive field, while it does not generate random noises during initial training steps." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 479, + 665 + ], + "type": "text", + "content": "Editing-Shift Matching. Besides the architecture innovation, we incorporate an editing-shift-matching technique to enhance in-context learning ability in image editing with more accurate detailed outputs." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 239, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 239, + 91, + 448, + 102 + ], + "type": "text", + "content": "InstructGIE: Towards Generalizable Image Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 479, + 151 + ], + "type": "text", + "content": "In specific, for each training ground truth " + }, + { + "bbox": [ + 130, + 114, + 479, + 151 + ], + "type": "inline_equation", + "content": "\\mathbf{Img}^{\\mathrm{train}} = \\{\\mathbf{Img}_0^{\\mathrm{in}},\\mathbf{Img}_0^{\\mathrm{out}},\\mathbf{Img}_1^{\\mathrm{in}}," + }, + { + "bbox": [ + 130, + 114, + 479, + 151 + ], + "type": "inline_equation", + "content": "\\mathbf{Img}_1^{\\mathrm{out}}\\}" + }, + { + "bbox": [ + 130, + 114, + 479, + 151 + ], + "type": "text", + "content": ", we calculate a implicit editing shift value using CLIP [32] image embedding:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 200, + 152, + 481, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 152, + 481, + 177 + ], + "spans": [ + { + "bbox": [ + 200, + 152, + 481, + 177 + ], + "type": "interline_equation", + "content": "\\mathcal {T} (\\mathbf {I m g}) = \\frac {\\sum_ {i = 0} ^ {1} \\operatorname {C L I P} \\left(\\mathbf {I m g} _ {i} ^ {\\text {i n}}\\right) - \\operatorname {C L I P} \\left(\\mathbf {I m g} _ {i} ^ {\\text {o u t}}\\right)}{2}. \\tag {4}", + "image_path": "714ef80bd3b50f65797fadb4579e6be9395bc79e6d4a293904e45a2dbbc02510.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 180, + 479, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 180, + 479, + 240 + ], + "spans": [ + { + "bbox": [ + 131, + 180, + 479, + 240 + ], + "type": "text", + "content": "During the training process, after predicting the noise, we use it to reverse the noised input and obtain a pseudo output image " + }, + { + "bbox": [ + 131, + 180, + 479, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{Img}^{\\mathrm{PO}} = \\{\\mathbf{PO}_0^{\\mathrm{in}},\\mathbf{PO}_0^{\\mathrm{out}},\\mathbf{PO}_1^{\\mathrm{in}}," + }, + { + "bbox": [ + 131, + 180, + 479, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{PO}_1^{\\mathrm{out}}\\}" + }, + { + "bbox": [ + 131, + 180, + 479, + 240 + ], + "type": "text", + "content": ". Our framework then calculates the editing transfer value of the pseudo output image and deduces a editing shift loss to optimize during our training via the cosine similarity of the two design transfer values:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 216, + 249, + 481, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 249, + 481, + 279 + ], + "spans": [ + { + "bbox": [ + 216, + 249, + 481, + 279 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {e s} = 1 - \\frac {\\mathcal {T} (\\mathbf {I m g} ^ {\\mathrm {P O}}) \\cdot \\mathcal {T} (\\mathbf {I m g} ^ {\\mathrm {t r a i n}})}{| | \\mathcal {T} (\\mathbf {I m g} ^ {\\mathrm {P O}}) | | \\times | | \\mathcal {T} (\\mathbf {I m g} ^ {\\mathrm {t r a i n}}) | |} \\tag {5}", + "image_path": "e586dfa3fa09dd58f10346e791696b5ebc5d48c833b936c8248a318e6e6bbfed.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 285, + 479, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 285, + 479, + 333 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 479, + 333 + ], + "type": "text", + "content": "Through editing-shift matching, our model can better comprehend how editing should be done within a visual context level through an implicit editing shift value. Furthermore, this implicit editing shift value can further guide the sampling process creating a controllable editing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 350, + 326, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 350, + 326, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 326, + 363 + ], + "type": "text", + "content": "4.2 Language Instruction Unification" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 370, + 480, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 480, + 442 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 480, + 442 + ], + "type": "text", + "content": "Previous works that utilize visual prompting in image editing tend to focus more on visual instructions and believe language instructions can be vague and unstable [38]. We believe that the potential of text prompts is not fully explored in image editing. Language instructions have significant impacts on diffusion model outputs. Language instructions with the same meaning can still result in entirely different outputs due to different processed language embeddings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 443, + 481, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 443, + 481, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 481, + 538 + ], + "type": "text", + "content": "To improve image editing and explore language instructions, we propose a novel approach, language instruction unification. During the training process, for each batch of training data, we randomly sample " + }, + { + "bbox": [ + 130, + 443, + 481, + 538 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 443, + 481, + 538 + ], + "type": "text", + "content": " of training data, collect their language editing instructions " + }, + { + "bbox": [ + 130, + 443, + 481, + 538 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 443, + 481, + 538 + ], + "type": "text", + "content": ", and process them through a frozen lightweight LLM, Open Llama 3b V2 Quant 4 [13] denoted as " + }, + { + "bbox": [ + 130, + 443, + 481, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 130, + 443, + 481, + 538 + ], + "type": "text", + "content": ". The LLM " + }, + { + "bbox": [ + 130, + 443, + 481, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 130, + 443, + 481, + 538 + ], + "type": "text", + "content": " is fixed prompted with a fixed random seed to uniformly reformulate the language instruction " + }, + { + "bbox": [ + 130, + 443, + 481, + 538 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 443, + 481, + 538 + ], + "type": "text", + "content": " better for machine-level understanding. The LLM " + }, + { + "bbox": [ + 130, + 443, + 481, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 130, + 443, + 481, + 538 + ], + "type": "text", + "content": " will output a unified language editing instruction " + }, + { + "bbox": [ + 130, + 443, + 481, + 538 + ], + "type": "inline_equation", + "content": "l'" + }, + { + "bbox": [ + 130, + 443, + 481, + 538 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 288, + 548, + 480, + 560 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 548, + 480, + 560 + ], + "spans": [ + { + "bbox": [ + 288, + 548, + 480, + 560 + ], + "type": "interline_equation", + "content": "l ^ {\\prime} = \\mathcal {U} (l) \\tag {6}", + "image_path": "bd4dd424f9be2903027edf7213529ac7d66b1b5d2c2b6dad12d67c0cac38aff3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 146, + 570, + 479, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 570, + 479, + 582 + ], + "spans": [ + { + "bbox": [ + 146, + 570, + 479, + 582 + ], + "type": "text", + "content": "We then augment the training data with unified language editing instructions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 582, + 479, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 479, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 479, + 616 + ], + "type": "text", + "content": "During the inference, each language editing instruction " + }, + { + "bbox": [ + 130, + 582, + 479, + 616 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 582, + 479, + 616 + ], + "type": "text", + "content": " is passed through the frozen LLM " + }, + { + "bbox": [ + 130, + 582, + 479, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 130, + 582, + 479, + 616 + ], + "type": "text", + "content": " for language instruction unification and then sent to our conditioned diffusion model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "By adopting language instruction unification for training data augmentation during the training, our conditioned diffusion model can learn diverse non-uniformed editing instructions to build up the model's knowledge distribution with unified language prompts. Adopting language instruction unification during" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "text", + "content": "Z. Meng and C. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "inference aligns with the training, therefore, greatly minimizing the possibility of diverse quality in edited outputs and maximizing the ability to generalize to unseen editing tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 166, + 284, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 166, + 284, + 178 + ], + "spans": [ + { + "bbox": [ + 131, + 166, + 284, + 178 + ], + "type": "text", + "content": "4.3 Selective Area Matching" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 479, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 479, + 229 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 479, + 229 + ], + "type": "text", + "content": "Diffusion-based image editing models usually suffer from low quality in specific details and this bottleneck appears to be more crucial in generalizable image editing methods [30, 38]. The details of human and animal images are easily distorted in these methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 230, + 480, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 230, + 480, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 480, + 302 + ], + "type": "text", + "content": "A naive solution might be utilizing negative prompts [34] like 'do not draw hands with 6 fingers' in general text-to-image tasks. However, in image editing, it is challenging to apply negative prompts. Users typically can not foresee the specific details after-edited outputs, therefore they are not able to construct appropriate negative prompts. Besides, negative prompts may limit the artistic potential of image editing models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 302, + 480, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 480, + 350 + ], + "type": "text", + "content": "To properly address this issue for generalizable image editorings, we propose an optimization method, namely selective area matching, that targets the difference in the chosen detailed editing area between the original training ground truth " + }, + { + "bbox": [ + 130, + 302, + 480, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{Img}^{\\mathrm{train}}" + }, + { + "bbox": [ + 130, + 302, + 480, + 350 + ], + "type": "text", + "content": " and the reversed pseudo output " + }, + { + "bbox": [ + 130, + 302, + 480, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{Img}^{\\mathrm{PO}}" + }, + { + "bbox": [ + 130, + 302, + 480, + 350 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 350, + 488, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 488, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 488, + 385 + ], + "type": "text", + "content": "In particular, during the training process, we incorporate a frozen Mask2Former model [6] " + }, + { + "bbox": [ + 130, + 350, + 488, + 385 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 350, + 488, + 385 + ], + "type": "text", + "content": " to obtain panoptic segmented training image information including segmented masks and class information " + }, + { + "bbox": [ + 130, + 350, + 488, + 385 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 350, + 488, + 385 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 247, + 391, + 480, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 391, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 247, + 391, + 480, + 403 + ], + "type": "interline_equation", + "content": "\\operatorname {s e g m a s k}, C = \\mathcal {M} (\\mathbf {I m g} ^ {\\text {t r a i n}}) \\tag {7}", + "image_path": "a12dd1257017ece1c0254e2ad010b6e0b5c795091bfa352efca8a5a95579d7af.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 408, + 492, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 408, + 492, + 443 + ], + "spans": [ + { + "bbox": [ + 131, + 408, + 492, + 443 + ], + "type": "text", + "content": "After that, our framework processes the class information using the same lightweight LLM " + }, + { + "bbox": [ + 131, + 408, + 492, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 131, + 408, + 492, + 443 + ], + "type": "text", + "content": " described in Sec. 4.2 to filter out pre-defined classes including living creatures and humans requiring special attention for addressing the details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 212, + 449, + 480, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 449, + 480, + 462 + ], + "spans": [ + { + "bbox": [ + 212, + 449, + 480, + 462 + ], + "type": "interline_equation", + "content": "C _ {\\text {f i l t e r e d}} = \\left\\{c \\in C | \\mathcal {U} (c, \\text {s e l e c t e d c l a s s e s}) = 1 \\right\\} \\tag {8}", + "image_path": "e8d0549b6e3a8b8a43412c22f6fd57faed983ad0983160a52c44bc8878498959.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 466, + 479, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 466, + 479, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 466, + 479, + 491 + ], + "type": "text", + "content": "Based on selected class labels, the framework then deduces a segmented binary mask for the selected editing area." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 213, + 495, + 480, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 495, + 480, + 526 + ], + "spans": [ + { + "bbox": [ + 213, + 495, + 480, + 526 + ], + "type": "interline_equation", + "content": "[ \\text {m a s k} ] _ {i} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f c l s} ([ \\text {s e g m a s k} ] _ {i}) \\in C _ {\\text {f i l t e r e d}} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {9}", + "image_path": "6bcab6b57a6a821b9472866ac5aa28b9f7a7672f3be426561cde6eb34331823a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 529, + 479, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 529, + 479, + 553 + ], + "spans": [ + { + "bbox": [ + 131, + 529, + 479, + 553 + ], + "type": "text", + "content": "During the training process, our framework calculates and optimizes the selective area matching loss by" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 173, + 558, + 480, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 558, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 173, + 558, + 480, + 590 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s a m} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\left([ \\mathbf {I m g} ^ {\\mathrm {P O}} ] _ {i} \\cdot [ \\operatorname {m a s k} ] _ {i}\\right) - \\left(\\left[ \\mathbf {I m g} ^ {\\mathrm {t r a i n}} \\right] _ {i} \\cdot [ \\operatorname {m a s k} ] _ {i}\\right)\\right) ^ {2} \\tag {10}", + "image_path": "48b43d27a5ff97a29b00aff05980823e518a427a9678a513d733ae86a29d3d8a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 594, + 345, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 345, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 345, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 594, + 345, + 605 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 594, + 345, + 605 + ], + "type": "text", + "content": " as the total pixel number in the image." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "content": "With selective area matching, image editing does not need negative prompts to deal with distorted details in images, which can make the most of the model's artistic editing capacity to generate high-quality outputs with great details. It is only incorporated during training, which does not change the inference, greatly saving inference efforts compared with negative prompts." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 239, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 239, + 91, + 448, + 103 + ], + "type": "text", + "content": "InstructGIE: Towards Generalizable Image Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 125, + 383, + 227 + ], + "blocks": [ + { + "bbox": [ + 222, + 115, + 288, + 122 + ], + "lines": [ + { + "bbox": [ + 222, + 115, + 288, + 122 + ], + "spans": [ + { + "bbox": [ + 222, + 115, + 288, + 122 + ], + "type": "text", + "content": "Phase 1: Data Generation" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 125, + 383, + 227 + ], + "lines": [ + { + "bbox": [ + 134, + 125, + 383, + 227 + ], + "spans": [ + { + "bbox": [ + 134, + 125, + 383, + 227 + ], + "type": "image", + "image_path": "a76af07fe12c29cb021ee74e300bb7f75a2bdb619b59fe337e94dd6b515d5e4a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 388, + 126, + 481, + 225 + ], + "blocks": [ + { + "bbox": [ + 402, + 116, + 468, + 124 + ], + "lines": [ + { + "bbox": [ + 402, + 116, + 468, + 124 + ], + "spans": [ + { + "bbox": [ + 402, + 116, + 468, + 124 + ], + "type": "text", + "content": "Phase 2: Data Processing" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 388, + 126, + 481, + 225 + ], + "lines": [ + { + "bbox": [ + 388, + 126, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 388, + 126, + 481, + 225 + ], + "type": "image", + "image_path": "9fdfd78cd360bb8a5ac837a7ccc9cfa5b17a5022efc5241c53c6d155295e9246.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 236, + 480, + 281 + ], + "lines": [ + { + "bbox": [ + 131, + 236, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 480, + 281 + ], + "type": "text", + "content": "Fig. 4: Dataset generation process Our dataset generation consists of two phases. Data Generation is to generate sets of image pairs under one editing caption. Data Processing is randomly pick image pairs under the same editing instruction and concatenate them together as one input for training." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 300, + 403, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 403, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 403, + 315 + ], + "type": "text", + "content": "5 Visual Prompt Dataset For Image Editing" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 323, + 480, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 480, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 480, + 382 + ], + "type": "text", + "content": "Traditional image editing datasets only contain editing image pairs with a small amount of similar editing instructions. To the best of our knowledge, there is no open-sourced image editing dataset that is explicitly designed for image editing with visual prompting, which utilizes multiple different image pairs for each editing instruction to provide a general demonstration in various cases." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 383, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 482, + 586 + ], + "type": "text", + "content": "Therefore, we introduce and open source a new dataset that is designed specifically for image editing tasks utilizing visual prompts. Our dataset generation pipeline is shown in Fig. 4 with 2 phases: Data Generation and Data Processing. In Data Generation phase, we first fine-tune GPT-3 [4] for 2 epochs with 700 groups of human-written edits, each consisting of 5 different pairs of input and edited captions with one editing instruction. Then as shown in step 1, we generate around 3500 groups of editing prompts using the fine-tuned GPT-3 model. In each group, there is one instruction " + }, + { + "bbox": [ + 130, + 383, + 482, + 586 + ], + "type": "inline_equation", + "content": "E_0" + }, + { + "bbox": [ + 130, + 383, + 482, + 586 + ], + "type": "text", + "content": " and five pairs of caption and edited caption " + }, + { + "bbox": [ + 130, + 383, + 482, + 586 + ], + "type": "inline_equation", + "content": "(I_0,O_0), (I_1,O_1), \\dots, (I_4,O_4)" + }, + { + "bbox": [ + 130, + 383, + 482, + 586 + ], + "type": "text", + "content": ". In step 2, similar to InstructPix2Pix, we then also adopt Prompt-to-Prompt for image generation. For each input and edited caption pair, we generated 50 times each with random noise and followed InstructPix2Pix to keep the best one image pair using CLIP-based metric. In addition, we also make sure for each editing instruction, there are at least two pairs of images. In step 3, we generate more image pair sets using in the same way. With filtering, we obtained around 12,000 images with around 2,000 editing instructions. In data preparing, we randomly choose 2 pairs of images under the same editing instruction and concatenate them for training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 602, + 230, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 602, + 230, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 602, + 230, + 616 + ], + "type": "text", + "content": "6 Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 624, + 272, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 272, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 272, + 636 + ], + "type": "text", + "content": "6.1 Experimental Settings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "content": "Datasets. To fairly compare our methods with baseline methods, we conduct both qualitative and quantitative experiments on our proposed synthetic dataset" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 278, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 278, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 278, + 102 + ], + "type": "text", + "content": "Z. Meng and C. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 227, + 174 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 227, + 174 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 227, + 174 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 227, + 174 + ], + "type": "image", + "image_path": "601e3632b1419bd570fdf82e47fff0a416040136cb19593fda7313cf7acdae4f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 229, + 114, + 277, + 173 + ], + "blocks": [ + { + "bbox": [ + 229, + 114, + 277, + 173 + ], + "lines": [ + { + "bbox": [ + 229, + 114, + 277, + 173 + ], + "spans": [ + { + "bbox": [ + 229, + 114, + 277, + 173 + ], + "type": "image", + "image_path": "8105753172b4fa12fa52f152417a3609475d28b35f69580b83c0fc45d2ec1a57.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 230, + 175, + 278, + 230 + ], + "blocks": [ + { + "bbox": [ + 230, + 175, + 278, + 230 + ], + "lines": [ + { + "bbox": [ + 230, + 175, + 278, + 230 + ], + "spans": [ + { + "bbox": [ + 230, + 175, + 278, + 230 + ], + "type": "image", + "image_path": "8f610fcb1aa24437e9cf97a63f06059e6f6b520a00368dd3c7eb9e4639059236.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 280, + 114, + 329, + 173 + ], + "blocks": [ + { + "bbox": [ + 280, + 114, + 329, + 173 + ], + "lines": [ + { + "bbox": [ + 280, + 114, + 329, + 173 + ], + "spans": [ + { + "bbox": [ + 280, + 114, + 329, + 173 + ], + "type": "image", + "image_path": "252f52df050f96a72baa362cc8403ac3c2d91102cedd4c00333411ed6a20866a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 280, + 175, + 329, + 230 + ], + "blocks": [ + { + "bbox": [ + 280, + 175, + 329, + 230 + ], + "lines": [ + { + "bbox": [ + 280, + 175, + 329, + 230 + ], + "spans": [ + { + "bbox": [ + 280, + 175, + 329, + 230 + ], + "type": "image", + "image_path": "b1cc42a4f3bb0deeaef273bd8d2ae025335e3aeef4af21dc6133bcdf75dd31c6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 331, + 114, + 380, + 173 + ], + "blocks": [ + { + "bbox": [ + 331, + 114, + 380, + 173 + ], + "lines": [ + { + "bbox": [ + 331, + 114, + 380, + 173 + ], + "spans": [ + { + "bbox": [ + 331, + 114, + 380, + 173 + ], + "type": "image", + "image_path": "7596bbcd456e64506fbf9988e26a2fad4750635029642a6bbbcba3efa9b137fc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 331, + 175, + 380, + 230 + ], + "blocks": [ + { + "bbox": [ + 331, + 175, + 380, + 230 + ], + "lines": [ + { + "bbox": [ + 331, + 175, + 380, + 230 + ], + "spans": [ + { + "bbox": [ + 331, + 175, + 380, + 230 + ], + "type": "image", + "image_path": "19932ac3b1098543dbf8c6eae2e2f6fe8bd01d7d57c9009e2efa8074b3eb8ba9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 382, + 114, + 430, + 173 + ], + "blocks": [ + { + "bbox": [ + 382, + 114, + 430, + 173 + ], + "lines": [ + { + "bbox": [ + 382, + 114, + 430, + 173 + ], + "spans": [ + { + "bbox": [ + 382, + 114, + 430, + 173 + ], + "type": "image", + "image_path": "2f21625a226f9308eefa402e70bdd9c1e44903c9e46e93df9b74711a3a6dcc8b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 382, + 175, + 430, + 230 + ], + "blocks": [ + { + "bbox": [ + 382, + 175, + 430, + 230 + ], + "lines": [ + { + "bbox": [ + 382, + 175, + 430, + 230 + ], + "spans": [ + { + "bbox": [ + 382, + 175, + 430, + 230 + ], + "type": "image", + "image_path": "00289cd57fb4612c9d0d743ee8443acd4ba57f001ad33cf587add54f890a73bf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 433, + 114, + 481, + 173 + ], + "blocks": [ + { + "bbox": [ + 433, + 114, + 481, + 173 + ], + "lines": [ + { + "bbox": [ + 433, + 114, + 481, + 173 + ], + "spans": [ + { + "bbox": [ + 433, + 114, + 481, + 173 + ], + "type": "image", + "image_path": "03d83ccabe08a072f88ccd06b15dde5330aa25bfd0bd1c821cbd714c68329d88.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 433, + 175, + 481, + 230 + ], + "blocks": [ + { + "bbox": [ + 433, + 175, + 481, + 230 + ], + "lines": [ + { + "bbox": [ + 433, + 175, + 481, + 230 + ], + "spans": [ + { + "bbox": [ + 433, + 175, + 481, + 230 + ], + "type": "image", + "image_path": "e682fc074c481e8a2f7286e53702f81b3759c256a3ad0d7068cd9e2f9a692b34.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 133, + 175, + 227, + 230 + ], + "blocks": [ + { + "bbox": [ + 153, + 175, + 206, + 181 + ], + "lines": [ + { + "bbox": [ + 153, + 175, + 206, + 181 + ], + "spans": [ + { + "bbox": [ + 153, + 175, + 206, + 181 + ], + "type": "text", + "content": "Crown Him Like A King" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 175, + 227, + 230 + ], + "lines": [ + { + "bbox": [ + 133, + 175, + 227, + 230 + ], + "spans": [ + { + "bbox": [ + 133, + 175, + 227, + 230 + ], + "type": "image", + "image_path": "c228565a9f3052fcafaeabc8a841ce408e879f8d87bad1909d0acf7d5987cab1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 231, + 208, + 237 + ], + "lines": [ + { + "bbox": [ + 153, + 231, + 208, + 237 + ], + "spans": [ + { + "bbox": [ + 153, + 231, + 208, + 237 + ], + "type": "text", + "content": "As a Precious Oil Painting" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 229, + 181, + 278, + 230 + ], + "blocks": [ + { + "bbox": [ + 229, + 181, + 278, + 230 + ], + "lines": [ + { + "bbox": [ + 229, + 181, + 278, + 230 + ], + "spans": [ + { + "bbox": [ + 229, + 181, + 278, + 230 + ], + "type": "image", + "image_path": "5ffbb102262a4b334bd79d909983d6f262d5bdb4e9f085fbcb5ccf048aafd65a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 230, + 239, + 278, + 287 + ], + "blocks": [ + { + "bbox": [ + 230, + 239, + 278, + 287 + ], + "lines": [ + { + "bbox": [ + 230, + 239, + 278, + 287 + ], + "spans": [ + { + "bbox": [ + 230, + 239, + 278, + 287 + ], + "type": "image", + "image_path": "a312e3dc3cdf2ca508f5ac46676ee92d8a5fdbe7ae34a26f9397a4e441fb9625.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 280, + 239, + 329, + 287 + ], + "blocks": [ + { + "bbox": [ + 280, + 239, + 329, + 287 + ], + "lines": [ + { + "bbox": [ + 280, + 239, + 329, + 287 + ], + "spans": [ + { + "bbox": [ + 280, + 239, + 329, + 287 + ], + "type": "image", + "image_path": "f154617d457968c5fd517493e62732ce0fab753c06c80f38ab7f5c52b96c7f62.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 331, + 239, + 380, + 287 + ], + "blocks": [ + { + "bbox": [ + 331, + 239, + 380, + 287 + ], + "lines": [ + { + "bbox": [ + 331, + 239, + 380, + 287 + ], + "spans": [ + { + "bbox": [ + 331, + 239, + 380, + 287 + ], + "type": "image", + "image_path": "ea110f8ab3669df4fcba8d9970e1aaac3ac1f3c7651a38e3a0e2407806ec3baf.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 382, + 239, + 430, + 287 + ], + "blocks": [ + { + "bbox": [ + 382, + 239, + 430, + 287 + ], + "lines": [ + { + "bbox": [ + 382, + 239, + 430, + 287 + ], + "spans": [ + { + "bbox": [ + 382, + 239, + 430, + 287 + ], + "type": "image", + "image_path": "2f3cee44554f9e3944a382f77df85585738ffb3deeb196a2df1c1b49421b887f.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 433, + 239, + 481, + 287 + ], + "blocks": [ + { + "bbox": [ + 433, + 239, + 481, + 287 + ], + "lines": [ + { + "bbox": [ + 433, + 239, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 433, + 239, + 481, + 287 + ], + "type": "image", + "image_path": "2303655566728338af3498715949bcab20ee80be9a8f90e506bfe44ac8f31ef1.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 132, + 239, + 227, + 287 + ], + "blocks": [ + { + "bbox": [ + 132, + 239, + 227, + 287 + ], + "lines": [ + { + "bbox": [ + 132, + 239, + 227, + 287 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 227, + 287 + ], + "type": "image", + "image_path": "8412a367a8ae7966e67454a42d393586e4528eca1bd62df9205e660f8d8cc20b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 289, + 202, + 294 + ], + "lines": [ + { + "bbox": [ + 159, + 289, + 202, + 294 + ], + "spans": [ + { + "bbox": [ + 159, + 289, + 202, + 294 + ], + "type": "text", + "content": "Change to Pink Hair" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 229, + 297, + 278, + 344 + ], + "blocks": [ + { + "bbox": [ + 229, + 297, + 278, + 344 + ], + "lines": [ + { + "bbox": [ + 229, + 297, + 278, + 344 + ], + "spans": [ + { + "bbox": [ + 229, + 297, + 278, + 344 + ], + "type": "image", + "image_path": "092c5e22587610c44d5779aeeddae1244c84b0ad345d29f1c447a3f8cf21371b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 230, + 354, + 278, + 401 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 354, + 278, + 401 + ], + "spans": [ + { + "bbox": [ + 230, + 354, + 278, + 401 + ], + "type": "image", + "image_path": "2ccc1c96be8e33208037477976121b620cf9d0933d0a975d769f23dc78176f66.jpg" + } + ] + } + ], + "index": 25, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 280, + 297, + 329, + 344 + ], + "blocks": [ + { + "bbox": [ + 280, + 297, + 329, + 344 + ], + "lines": [ + { + "bbox": [ + 280, + 297, + 329, + 344 + ], + "spans": [ + { + "bbox": [ + 280, + 297, + 329, + 344 + ], + "type": "image", + "image_path": "1a0318c5289ee24aaaa4920ae838cbb51c40a1d2eac342b2a8ed14d2e4ef27ef.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 331, + 297, + 380, + 344 + ], + "blocks": [ + { + "bbox": [ + 331, + 297, + 380, + 344 + ], + "lines": [ + { + "bbox": [ + 331, + 297, + 380, + 344 + ], + "spans": [ + { + "bbox": [ + 331, + 297, + 380, + 344 + ], + "type": "image", + "image_path": "9d9dfe6188c2416cdb8c88b65ad09bb43e0eaf604c449feba991f0e2f2e67af7.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 382, + 297, + 430, + 344 + ], + "blocks": [ + { + "bbox": [ + 382, + 297, + 430, + 344 + ], + "lines": [ + { + "bbox": [ + 382, + 297, + 430, + 344 + ], + "spans": [ + { + "bbox": [ + 382, + 297, + 430, + 344 + ], + "type": "image", + "image_path": "87f939a61b87c9cbb85a10525bc3d9695301b2a98ac5ef9fe98ec289ae45d322.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 433, + 297, + 481, + 344 + ], + "blocks": [ + { + "bbox": [ + 433, + 297, + 481, + 344 + ], + "lines": [ + { + "bbox": [ + 433, + 297, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 433, + 297, + 481, + 344 + ], + "type": "image", + "image_path": "55ca2370efb06ed65873cdc9d01001b5e10be8cfb11fa97c51a41b9df7414233.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 132, + 297, + 227, + 344 + ], + "blocks": [ + { + "bbox": [ + 132, + 297, + 227, + 344 + ], + "lines": [ + { + "bbox": [ + 132, + 297, + 227, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 227, + 344 + ], + "type": "image", + "image_path": "7a96eb3508cc78e58a6d0abf721b0b0989b598107e5e8a6313ae0b137df89f63.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 346, + 215, + 352 + ], + "lines": [ + { + "bbox": [ + 145, + 346, + 215, + 352 + ], + "spans": [ + { + "bbox": [ + 145, + 346, + 215, + 352 + ], + "type": "text", + "content": "Light Them On Devastating Fire!" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 229, + 354, + 278, + 401 + ], + "blocks": [ + { + "bbox": [ + 229, + 354, + 278, + 401 + ], + "lines": [ + { + "bbox": [ + 229, + 354, + 278, + 401 + ], + "spans": [ + { + "bbox": [ + 229, + 354, + 278, + 401 + ], + "type": "image", + "image_path": "ce1808793784b27066bded330f2aff2db3c4677d10a30c13bbb7402d63186cdb.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 197, + 403, + 230, + 408 + ], + "lines": [ + { + "bbox": [ + 197, + 403, + 230, + 408 + ], + "spans": [ + { + "bbox": [ + 197, + 403, + 230, + 408 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 422, + 482, + 445 + ], + "lines": [ + { + "bbox": [ + 131, + 422, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 131, + 422, + 482, + 445 + ], + "type": "text", + "content": "Fig. 5: Qualitative Comparison on our Test Dataset. We conducted experiments on various scenarios, including human, architecture and landscape." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 280, + 354, + 329, + 401 + ], + "blocks": [ + { + "bbox": [ + 280, + 354, + 329, + 401 + ], + "lines": [ + { + "bbox": [ + 280, + 354, + 329, + 401 + ], + "spans": [ + { + "bbox": [ + 280, + 354, + 329, + 401 + ], + "type": "image", + "image_path": "6ca9d4c176cb6f732c529c1a1bc83f78f3dcd0b851847e57ae8ea5815d6beda8.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 331, + 354, + 380, + 401 + ], + "blocks": [ + { + "bbox": [ + 331, + 354, + 380, + 401 + ], + "lines": [ + { + "bbox": [ + 331, + 354, + 380, + 401 + ], + "spans": [ + { + "bbox": [ + 331, + 354, + 380, + 401 + ], + "type": "image", + "image_path": "844b5f9f0edecc5f02b4b1aa9cc423130025b98c8da87b609b77608f7208a46b.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 382, + 354, + 430, + 401 + ], + "blocks": [ + { + "bbox": [ + 382, + 354, + 430, + 401 + ], + "lines": [ + { + "bbox": [ + 382, + 354, + 430, + 401 + ], + "spans": [ + { + "bbox": [ + 382, + 354, + 430, + 401 + ], + "type": "image", + "image_path": "dd0bdf10146d061430688e52f31198e1f698449455b12aadd213c4d6d5876f91.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 433, + 354, + 481, + 401 + ], + "blocks": [ + { + "bbox": [ + 433, + 354, + 481, + 401 + ], + "lines": [ + { + "bbox": [ + 433, + 354, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 433, + 354, + 481, + 401 + ], + "type": "image", + "image_path": "e7e116bea32dfc9e89eed9714def87901d951f4f346d784696af79b2eef56644.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "bbox": [ + 130, + 471, + 481, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 481, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 481, + 507 + ], + "type": "text", + "content": "that involves over 12,000 filtered images and 2,000 filtered editing instructions. All single input images have a resolution of " + }, + { + "bbox": [ + 130, + 471, + 481, + 507 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 471, + 481, + 507 + ], + "type": "text", + "content": " and are resized to " + }, + { + "bbox": [ + 130, + 471, + 481, + 507 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 471, + 481, + 507 + ], + "type": "text", + "content": " for training and testing purposes." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 130, + 508, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 604 + ], + "type": "text", + "content": "Implementation Details. In our approach, we split the dataset with " + }, + { + "bbox": [ + 130, + 508, + 482, + 604 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 130, + 508, + 482, + 604 + ], + "type": "text", + "content": " for training and " + }, + { + "bbox": [ + 130, + 508, + 482, + 604 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 508, + 482, + 604 + ], + "type": "text", + "content": " for testing. As demonstrated in Fig. 2, two image pairs are concatenated into one " + }, + { + "bbox": [ + 130, + 508, + 482, + 604 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 508, + 482, + 604 + ], + "type": "text", + "content": " image with the same editing instructions as " + }, + { + "bbox": [ + 130, + 508, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{Img}^{\\mathrm{train}}" + }, + { + "bbox": [ + 130, + 508, + 482, + 604 + ], + "type": "text", + "content": " and mask the fourth quadrant with a grey color as " + }, + { + "bbox": [ + 130, + 508, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{Img}^{\\mathrm{VPcon}}" + }, + { + "bbox": [ + 130, + 508, + 482, + 604 + ], + "type": "text", + "content": ". We prepare the in-domain test dataset in the same format. For out-of-domain testings, we ensure that both the visual instruction pairs and the text instructions are not used during training, to best simulate how models perform in real-life image editing generalization scenarios." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "For baselines, we use their original configurations to train their model. In our method, we only fine-tune the additional ControlNet for 5000 steps with a batch size of 1024 and a learning rate at " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ". During training, we set the classifier-free scale the same as the original ControlNet. And we randomly drop " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " of language or visual editing instructions to further enhance our model's" + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 239, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 239, + 91, + 448, + 102 + ], + "type": "text", + "content": "InstructGIE: Towards Generalizable Image Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 157, + 479, + 203 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "text", + "content": "Table 1: Quantitative results comparison between our method and baseline methods. Quantitative results show our method excels the baseline methods in both FID and CLIP directional Similarity score in a great margin." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 157, + 479, + 203 + ], + "lines": [ + { + "bbox": [ + 134, + 157, + 479, + 203 + ], + "spans": [ + { + "bbox": [ + 134, + 157, + 479, + 203 + ], + "type": "table", + "html": "
SDEdit (E)InstructPix2PixPromptDiffusionOurs
FID ↓21.6717.8713.757.57
CLIP DirSim ↑0.110.170.210.27
", + "image_path": "dc87bfcc57517dda4642a7636981c23cd71c7c5419abf582b7ac8043d2c72da9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 226, + 479, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 226, + 479, + 250 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 479, + 250 + ], + "type": "text", + "content": "generalization ability. Our implementation utilizes PyTorch and is trained on 4 Tesla A100-40G GPUs with AdamW optimizer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 251, + 480, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 480, + 334 + ], + "type": "text", + "content": "Comparison Methods. To evaluate the effectiveness of our work, we compare with other state-of-the-art image editing frameworks, including SDEdit [28], Instruct-Pix2pix [3] and PromptDiffusion [41]. We adopt two quantitative metrics Fr'echet Inception Distance (FID) and CLIP directional Similarity (CLIP DirSim) proposed by Gal et al. [12]. We utilize the FID score to quantitatively represent the image quality of generated editing outputs, and CLIP DirSim to evaluate how well the models follow editing instructions to produce the output." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 354, + 309, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 309, + 367 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 309, + 367 + ], + "type": "text", + "content": "6.2 State-of-the-Art Comparisons" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 377, + 481, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 377, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 377, + 481, + 568 + ], + "type": "text", + "content": "Qualitative Evaluation. In Fig. 5, we present our qualitative results in the testing set (in domain). The comparison shows that our method surpass previous baseline methods. Our method understands and follows both visual and language editing instructions better, and produces a far more detailed manipulated output especially in human figures. In Fig. 6, we present our qualitative results tested in out-of-domain settings. We include five editing instructions that are considered extremely hard for diffusion-based image editing model [18], including object add, object remove, structure location change, structure action change, and object size change. It is important to note that due to how we generate our training dataset utilizing Prompt-to-Prompt [15], editing images pairs with these types of editing instructions is not feasible to generate our training data. This generalization comparison shows that our method excels baseline methods by a significant margin. Our method shows a great capability to carry out well-detailed quality outputs following these hard editing instructions in diffusion-based image editing models, while other baseline methods all fail to understand the editing instructions well or perform manipulations close to the editing instructions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 569, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 481, + 640 + ], + "type": "text", + "content": "Quantitative Evaluation. In quantitative evaluation, we score 7.57 in FID, better than SDEdit (E), InstructPix2Pix and PromptDiffusion which scores 21.67, 17.87 and 13.75. We achieve 0.27 in CLIP DirSim, better than baselines with " + }, + { + "bbox": [ + 132, + 569, + 481, + 640 + ], + "type": "inline_equation", + "content": "0.11 / 0.17 / 0.21" + }, + { + "bbox": [ + 132, + 569, + 481, + 640 + ], + "type": "text", + "content": " of CLIP DirSim scores. These quantitative findings show that our method generates higher-quality images with better detailed qualities and also exactly follows both language and visual editing instructions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": "Ablation study. We conduct an ablation study on each of the four components of our proposed method. Namely, the Reformed Conditioned Latent Diffusion" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "text", + "content": "Z. Meng and C. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 227, + 173 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 227, + 173 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 227, + 173 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 227, + 173 + ], + "type": "image", + "image_path": "9b1a8b3f5261250f98c783b25eccfbe31d5fd0bfed21837faff2b3367ba6b160.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 175, + 198, + 181 + ], + "lines": [ + { + "bbox": [ + 162, + 175, + 198, + 181 + ], + "spans": [ + { + "bbox": [ + 162, + 175, + 198, + 181 + ], + "type": "text", + "content": "Add Some Snow" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 228, + 114, + 278, + 173 + ], + "blocks": [ + { + "bbox": [ + 228, + 114, + 278, + 173 + ], + "lines": [ + { + "bbox": [ + 228, + 114, + 278, + 173 + ], + "spans": [ + { + "bbox": [ + 228, + 114, + 278, + 173 + ], + "type": "image", + "image_path": "99d1dcfd357672b24d81322483cf5088e5b9be75663d07fdafa2f3e8fbfdd6fd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 280, + 114, + 329, + 172 + ], + "blocks": [ + { + "bbox": [ + 280, + 114, + 329, + 172 + ], + "lines": [ + { + "bbox": [ + 280, + 114, + 329, + 172 + ], + "spans": [ + { + "bbox": [ + 280, + 114, + 329, + 172 + ], + "type": "image", + "image_path": "3756a06f72ab93543c972051d6c4be5b0d870abd68255c9bc54205cf688411b5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 331, + 114, + 380, + 172 + ], + "blocks": [ + { + "bbox": [ + 331, + 114, + 380, + 172 + ], + "lines": [ + { + "bbox": [ + 331, + 114, + 380, + 172 + ], + "spans": [ + { + "bbox": [ + 331, + 114, + 380, + 172 + ], + "type": "image", + "image_path": "d833da5ab0c4172b98c118ff38364ef7549add0cfe8023f0f8b76fd839b5c085.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 381, + 114, + 430, + 172 + ], + "blocks": [ + { + "bbox": [ + 381, + 114, + 430, + 172 + ], + "lines": [ + { + "bbox": [ + 381, + 114, + 430, + 172 + ], + "spans": [ + { + "bbox": [ + 381, + 114, + 430, + 172 + ], + "type": "image", + "image_path": "797926c8e1b5c902225aaae83f05ec2f05c95c671c2deec57e67316b32595c79.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 432, + 114, + 481, + 173 + ], + "blocks": [ + { + "bbox": [ + 432, + 114, + 481, + 173 + ], + "lines": [ + { + "bbox": [ + 432, + 114, + 481, + 173 + ], + "spans": [ + { + "bbox": [ + 432, + 114, + 481, + 173 + ], + "type": "image", + "image_path": "ab4878c744f07d4e4f4a64fbbd2cfa94e27c40ee5904c2d2dd40db2166e77c14.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 133, + 182, + 227, + 230 + ], + "blocks": [ + { + "bbox": [ + 133, + 182, + 227, + 230 + ], + "lines": [ + { + "bbox": [ + 133, + 182, + 227, + 230 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 227, + 230 + ], + "type": "image", + "image_path": "e1652f09b079c909f62d270f254730f3e158f4f97e83b6d35335c7bb168c829a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 231, + 196, + 237 + ], + "lines": [ + { + "bbox": [ + 164, + 231, + 196, + 237 + ], + "spans": [ + { + "bbox": [ + 164, + 231, + 196, + 237 + ], + "type": "text", + "content": "Bend His Arm" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 229, + 182, + 278, + 230 + ], + "blocks": [ + { + "bbox": [ + 229, + 182, + 278, + 230 + ], + "lines": [ + { + "bbox": [ + 229, + 182, + 278, + 230 + ], + "spans": [ + { + "bbox": [ + 229, + 182, + 278, + 230 + ], + "type": "image", + "image_path": "012f19323ab9e590912012e26ff71cd5d12c176bd3fcba6836f034c062a68d62.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 280, + 182, + 329, + 230 + ], + "blocks": [ + { + "bbox": [ + 280, + 182, + 329, + 230 + ], + "lines": [ + { + "bbox": [ + 280, + 182, + 329, + 230 + ], + "spans": [ + { + "bbox": [ + 280, + 182, + 329, + 230 + ], + "type": "image", + "image_path": "9560d3f1592a7e353ba4de48a8fce01dfa0f59ec7242998a42470e332306f668.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 331, + 182, + 380, + 230 + ], + "blocks": [ + { + "bbox": [ + 331, + 182, + 380, + 230 + ], + "lines": [ + { + "bbox": [ + 331, + 182, + 380, + 230 + ], + "spans": [ + { + "bbox": [ + 331, + 182, + 380, + 230 + ], + "type": "image", + "image_path": "d5fea7382be0670f5660fd1e2806dd1d1d3f8fdb86a6816c4f5a0e2baa929e69.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 382, + 182, + 430, + 230 + ], + "blocks": [ + { + "bbox": [ + 382, + 182, + 430, + 230 + ], + "lines": [ + { + "bbox": [ + 382, + 182, + 430, + 230 + ], + "spans": [ + { + "bbox": [ + 382, + 182, + 430, + 230 + ], + "type": "image", + "image_path": "b51ef89bf099d75b3744f636e901729534c40b4f0184e0efc185c7fad304e046.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 432, + 182, + 481, + 230 + ], + "blocks": [ + { + "bbox": [ + 432, + 182, + 481, + 230 + ], + "lines": [ + { + "bbox": [ + 432, + 182, + 481, + 230 + ], + "spans": [ + { + "bbox": [ + 432, + 182, + 481, + 230 + ], + "type": "image", + "image_path": "eeab39df57bc1ebe4fe2ad600987a9b9bce9d0cdf112ec165e286c50d5c8efdb.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 133, + 239, + 227, + 287 + ], + "blocks": [ + { + "bbox": [ + 133, + 239, + 227, + 287 + ], + "lines": [ + { + "bbox": [ + 133, + 239, + 227, + 287 + ], + "spans": [ + { + "bbox": [ + 133, + 239, + 227, + 287 + ], + "type": "image", + "image_path": "e03509a783da7fb55ad8ac482e7831e22ca4a1a857d22dcbd6da06e8cfc56151.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 289, + 202, + 295 + ], + "lines": [ + { + "bbox": [ + 159, + 289, + 202, + 295 + ], + "spans": [ + { + "bbox": [ + 159, + 289, + 202, + 295 + ], + "type": "text", + "content": "Move It to the Right" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 229, + 239, + 278, + 287 + ], + "blocks": [ + { + "bbox": [ + 229, + 239, + 278, + 287 + ], + "lines": [ + { + "bbox": [ + 229, + 239, + 278, + 287 + ], + "spans": [ + { + "bbox": [ + 229, + 239, + 278, + 287 + ], + "type": "image", + "image_path": "33096d3c36a7c2fd9f30fbf9d6c2cf4aa7e5cc07ab2a481ec8df1430101aa095.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 280, + 239, + 329, + 287 + ], + "blocks": [ + { + "bbox": [ + 280, + 239, + 329, + 287 + ], + "lines": [ + { + "bbox": [ + 280, + 239, + 329, + 287 + ], + "spans": [ + { + "bbox": [ + 280, + 239, + 329, + 287 + ], + "type": "image", + "image_path": "943b5fc4af426ec1e5392f045af78087f90ace3b5ecfa58fbda23027e81ca4f6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 331, + 239, + 380, + 287 + ], + "blocks": [ + { + "bbox": [ + 331, + 239, + 380, + 287 + ], + "lines": [ + { + "bbox": [ + 331, + 239, + 380, + 287 + ], + "spans": [ + { + "bbox": [ + 331, + 239, + 380, + 287 + ], + "type": "image", + "image_path": "1d708ea75de81e7af6d8087b27c23007baa369a05ce3763d82b503b50ce1f0c3.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 382, + 239, + 430, + 287 + ], + "blocks": [ + { + "bbox": [ + 382, + 239, + 430, + 287 + ], + "lines": [ + { + "bbox": [ + 382, + 239, + 430, + 287 + ], + "spans": [ + { + "bbox": [ + 382, + 239, + 430, + 287 + ], + "type": "image", + "image_path": "13c3b8dd5458aec95a6b141b0547b69a115571e79ecb8033179f24eae69bc8ca.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 432, + 239, + 481, + 287 + ], + "blocks": [ + { + "bbox": [ + 432, + 239, + 481, + 287 + ], + "lines": [ + { + "bbox": [ + 432, + 239, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 432, + 239, + 481, + 287 + ], + "type": "image", + "image_path": "d8bdd96b446ebe407c0a50ae9a1926dd702a77746fac67ddd6414425054c471f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 133, + 297, + 227, + 344 + ], + "blocks": [ + { + "bbox": [ + 133, + 297, + 227, + 344 + ], + "lines": [ + { + "bbox": [ + 133, + 297, + 227, + 344 + ], + "spans": [ + { + "bbox": [ + 133, + 297, + 227, + 344 + ], + "type": "image", + "image_path": "32d7e0370e93e733f256b94210909ca4de698e8842d7510fb2f4709c23877235.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 346, + 197, + 352 + ], + "lines": [ + { + "bbox": [ + 163, + 346, + 197, + 352 + ], + "spans": [ + { + "bbox": [ + 163, + 346, + 197, + 352 + ], + "type": "text", + "content": "Keep Only One" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 229, + 297, + 278, + 344 + ], + "blocks": [ + { + "bbox": [ + 229, + 297, + 278, + 344 + ], + "lines": [ + { + "bbox": [ + 229, + 297, + 278, + 344 + ], + "spans": [ + { + "bbox": [ + 229, + 297, + 278, + 344 + ], + "type": "image", + "image_path": "45fe24e1b3caf71aeb773019b4b125c7bec1877a3b12cc2f8120e6af8c041857.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 280, + 297, + 329, + 344 + ], + "blocks": [ + { + "bbox": [ + 280, + 297, + 329, + 344 + ], + "lines": [ + { + "bbox": [ + 280, + 297, + 329, + 344 + ], + "spans": [ + { + "bbox": [ + 280, + 297, + 329, + 344 + ], + "type": "image", + "image_path": "28d0d92e67a7f47c855c1f7a1b861c22c812911bcaf11ae3594a6ecb1d95737e.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 331, + 297, + 380, + 344 + ], + "blocks": [ + { + "bbox": [ + 331, + 297, + 380, + 344 + ], + "lines": [ + { + "bbox": [ + 331, + 297, + 380, + 344 + ], + "spans": [ + { + "bbox": [ + 331, + 297, + 380, + 344 + ], + "type": "image", + "image_path": "95e274f915cfadaf716efec172f2fc007ab44bff7ae0a46e017af779a75cec9f.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 382, + 297, + 430, + 344 + ], + "blocks": [ + { + "bbox": [ + 382, + 297, + 430, + 344 + ], + "lines": [ + { + "bbox": [ + 382, + 297, + 430, + 344 + ], + "spans": [ + { + "bbox": [ + 382, + 297, + 430, + 344 + ], + "type": "image", + "image_path": "3baac35ba9da06666f5181ae26f7523aefd9c019601c15a98c22b1d0d9c68047.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 432, + 297, + 481, + 344 + ], + "blocks": [ + { + "bbox": [ + 432, + 297, + 481, + 344 + ], + "lines": [ + { + "bbox": [ + 432, + 297, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 432, + 297, + 481, + 344 + ], + "type": "image", + "image_path": "061f0f591285299418ae2abf9ee73ef0c8260f77ddfac0dcce9a3ed7f0912d14.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 133, + 354, + 227, + 401 + ], + "blocks": [ + { + "bbox": [ + 133, + 354, + 227, + 401 + ], + "lines": [ + { + "bbox": [ + 133, + 354, + 227, + 401 + ], + "spans": [ + { + "bbox": [ + 133, + 354, + 227, + 401 + ], + "type": "image", + "image_path": "1fbff72ce8e03f5cf2beb0e0d2dc12df8f4e820029f2b9f7b312982eab92ad1e.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 403, + 198, + 408 + ], + "lines": [ + { + "bbox": [ + 163, + 403, + 198, + 408 + ], + "spans": [ + { + "bbox": [ + 163, + 403, + 198, + 408 + ], + "type": "text", + "content": "Make It Smaller" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 229, + 354, + 278, + 401 + ], + "blocks": [ + { + "bbox": [ + 229, + 354, + 278, + 401 + ], + "lines": [ + { + "bbox": [ + 229, + 354, + 278, + 401 + ], + "spans": [ + { + "bbox": [ + 229, + 354, + 278, + 401 + ], + "type": "image", + "image_path": "e765032929ade0c0ad1167c5c5936f9dd7464708df3db3ad6086ea0c3a6d8a36.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 422, + 481, + 444 + ], + "lines": [ + { + "bbox": [ + 132, + 422, + 481, + 444 + ], + "spans": [ + { + "bbox": [ + 132, + 422, + 481, + 444 + ], + "type": "text", + "content": "Fig. 6: Qualitative Comparison on Out-of-Domain Images. We conducted experiments on images and instruct that are not in training dataset." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 280, + 353, + 329, + 401 + ], + "blocks": [ + { + "bbox": [ + 280, + 353, + 329, + 401 + ], + "lines": [ + { + "bbox": [ + 280, + 353, + 329, + 401 + ], + "spans": [ + { + "bbox": [ + 280, + 353, + 329, + 401 + ], + "type": "image", + "image_path": "ac231f8e32ab2c7cff80dbaf39eb6dc5b6ed2fadca375e354fc56156c29b0e3c.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 331, + 354, + 380, + 401 + ], + "blocks": [ + { + "bbox": [ + 331, + 354, + 380, + 401 + ], + "lines": [ + { + "bbox": [ + 331, + 354, + 380, + 401 + ], + "spans": [ + { + "bbox": [ + 331, + 354, + 380, + 401 + ], + "type": "image", + "image_path": "6198413c95eeff30c7552d7b8246dea47d0370a4b84e493102d83b0637063f29.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 382, + 354, + 430, + 401 + ], + "blocks": [ + { + "bbox": [ + 382, + 354, + 430, + 401 + ], + "lines": [ + { + "bbox": [ + 382, + 354, + 430, + 401 + ], + "spans": [ + { + "bbox": [ + 382, + 354, + 430, + 401 + ], + "type": "image", + "image_path": "1732f0685612529675eed4fea1e2ac58d6e3c9a2330cc6e2583bb156b79c98f7.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 432, + 354, + 481, + 401 + ], + "blocks": [ + { + "bbox": [ + 432, + 354, + 481, + 401 + ], + "lines": [ + { + "bbox": [ + 432, + 354, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 432, + 354, + 481, + 401 + ], + "type": "image", + "image_path": "0c69def8584cf15bfda75e026975622d217c5f78a330ab3a73d694c1ab8180c4.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "bbox": [ + 130, + 485, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 666 + ], + "type": "text", + "content": "Model (RCLDM), Editing Shift Matching (ESM), Language Instruction Unification (LIU) and Selective Area Matching (SAM). We present the qualitative results in Fig. 7. From the qualitative results, we can see that without RCLDM and ESM, the model understands the visual editing instructions much weaker, especially in out-of-domain editing. Without LIU, for two language editing instructions with the same meaning, the model produces two output edited images in different detail and quality. This difference in quality tends to increase in out-of-domain settings. Without SAM, the details of the human face are distorted making the model more vulnerable to producing outputs with worse detailed qualities. We also include the qualitative ablation results on the testing dataset in Tab. 2. From observation, our method performs the best when incorporating all four components. Without SAM or LIU, the FID score increases, meaning those modules enhance the detail quality of the output generated. Without CLDM or ESM, the CLIP DirSim score decreases, showing that those two modules contribute to a better understanding in both language and visual level." + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 239, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 239, + 91, + 448, + 102 + ], + "type": "text", + "content": "InstructGIE: Towards Generalizable Image Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 217, + 167 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 217, + 167 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 217, + 167 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 217, + 167 + ], + "type": "image", + "image_path": "082dfed84ce85344d81271af58b8bd5ba51fb090989a06b30344d5a06fa7e781.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 167, + 195, + 172 + ], + "lines": [ + { + "bbox": [ + 153, + 167, + 195, + 172 + ], + "spans": [ + { + "bbox": [ + 153, + 167, + 195, + 172 + ], + "type": "text", + "content": "Make It in Winter Time" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 218, + 114, + 260, + 166 + ], + "blocks": [ + { + "bbox": [ + 218, + 114, + 260, + 166 + ], + "lines": [ + { + "bbox": [ + 218, + 114, + 260, + 166 + ], + "spans": [ + { + "bbox": [ + 218, + 114, + 260, + 166 + ], + "type": "image", + "image_path": "4516c3aa8231a82b3e917500fa4ce33387ce852b5d0f995ad628c7cb48115536.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 262, + 114, + 304, + 166 + ], + "blocks": [ + { + "bbox": [ + 262, + 114, + 304, + 166 + ], + "lines": [ + { + "bbox": [ + 262, + 114, + 304, + 166 + ], + "spans": [ + { + "bbox": [ + 262, + 114, + 304, + 166 + ], + "type": "image", + "image_path": "2ac1ede521065bc6be8185c74650c1a75a75254c1f8064e859553d4c24291b6a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 114, + 348, + 166 + ], + "blocks": [ + { + "bbox": [ + 306, + 114, + 348, + 166 + ], + "lines": [ + { + "bbox": [ + 306, + 114, + 348, + 166 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 348, + 166 + ], + "type": "image", + "image_path": "dd533610c6ab99385d42506ad2d07b70640f702702acf25cd3a8bd3a5ce910e4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 351, + 114, + 392, + 166 + ], + "blocks": [ + { + "bbox": [ + 351, + 114, + 392, + 166 + ], + "lines": [ + { + "bbox": [ + 351, + 114, + 392, + 166 + ], + "spans": [ + { + "bbox": [ + 351, + 114, + 392, + 166 + ], + "type": "image", + "image_path": "ab2f1684e0015fab2c381669c266a6fe00fd3c5a7c04264473c307d94d4c0588.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 394, + 114, + 436, + 166 + ], + "blocks": [ + { + "bbox": [ + 394, + 114, + 436, + 166 + ], + "lines": [ + { + "bbox": [ + 394, + 114, + 436, + 166 + ], + "spans": [ + { + "bbox": [ + 394, + 114, + 436, + 166 + ], + "type": "image", + "image_path": "be79d7f66060746e7965bfa2e9c8316ca7f0a44f78dc9ba448eefa813a8f19f9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 438, + 114, + 482, + 166 + ], + "blocks": [ + { + "bbox": [ + 438, + 114, + 482, + 166 + ], + "lines": [ + { + "bbox": [ + 438, + 114, + 482, + 166 + ], + "spans": [ + { + "bbox": [ + 438, + 114, + 482, + 166 + ], + "type": "image", + "image_path": "b7a9092fa842d1e60c4a0c1be6e98baa5698343b94c1f64ee0f3cea0fd492030.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 133, + 173, + 216, + 215 + ], + "blocks": [ + { + "bbox": [ + 133, + 173, + 216, + 215 + ], + "lines": [ + { + "bbox": [ + 133, + 173, + 216, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 173, + 216, + 215 + ], + "type": "image", + "image_path": "dd15423423a2b17de3a6bf03986dc5b55de89166d1cc3ba1599e4901ba993c06.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 218, + 173, + 260, + 215 + ], + "blocks": [ + { + "bbox": [ + 218, + 173, + 260, + 215 + ], + "lines": [ + { + "bbox": [ + 218, + 173, + 260, + 215 + ], + "spans": [ + { + "bbox": [ + 218, + 173, + 260, + 215 + ], + "type": "image", + "image_path": "deffb905bdb4c184a2a5d6b604bea4a756498e18cb8a53d10a6e32ea5ef4554e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 262, + 173, + 304, + 215 + ], + "blocks": [ + { + "bbox": [ + 262, + 173, + 304, + 215 + ], + "lines": [ + { + "bbox": [ + 262, + 173, + 304, + 215 + ], + "spans": [ + { + "bbox": [ + 262, + 173, + 304, + 215 + ], + "type": "image", + "image_path": "55ba8916f9ea4c0cf3c66de6e0083d2f77ff953772a5d41e667c9e2ef34a39c8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 306, + 173, + 348, + 215 + ], + "blocks": [ + { + "bbox": [ + 306, + 173, + 348, + 215 + ], + "lines": [ + { + "bbox": [ + 306, + 173, + 348, + 215 + ], + "spans": [ + { + "bbox": [ + 306, + 173, + 348, + 215 + ], + "type": "image", + "image_path": "365374894182f105968bc6bd785f343786d8feea6c2b74abe279dd7f7bca8d0c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 351, + 173, + 392, + 215 + ], + "blocks": [ + { + "bbox": [ + 351, + 173, + 392, + 215 + ], + "lines": [ + { + "bbox": [ + 351, + 173, + 392, + 215 + ], + "spans": [ + { + "bbox": [ + 351, + 173, + 392, + 215 + ], + "type": "image", + "image_path": "d16b8cc6e61d2eb5d8e4aec4ec07b31e647678fdda58c1c50509067fa17f3ccb.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 394, + 173, + 436, + 215 + ], + "blocks": [ + { + "bbox": [ + 394, + 173, + 436, + 215 + ], + "lines": [ + { + "bbox": [ + 394, + 173, + 436, + 215 + ], + "spans": [ + { + "bbox": [ + 394, + 173, + 436, + 215 + ], + "type": "image", + "image_path": "14eaf355a89442b992c7ea979fb63fb0b087181a36b25b608fc0c1df99da0c53.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 438, + 173, + 480, + 215 + ], + "blocks": [ + { + "bbox": [ + 438, + 173, + 480, + 215 + ], + "lines": [ + { + "bbox": [ + 438, + 173, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 438, + 173, + 480, + 215 + ], + "type": "image", + "image_path": "aa8ee905cddf6ff10dfe10116542105c63737916cf609f7d488c63a52ba0ea65.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 133, + 223, + 216, + 265 + ], + "blocks": [ + { + "bbox": [ + 154, + 217, + 195, + 222 + ], + "lines": [ + { + "bbox": [ + 154, + 217, + 195, + 222 + ], + "spans": [ + { + "bbox": [ + 154, + 217, + 195, + 222 + ], + "type": "text", + "content": "Turn It to Winter Time" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 223, + 216, + 265 + ], + "lines": [ + { + "bbox": [ + 133, + 223, + 216, + 265 + ], + "spans": [ + { + "bbox": [ + 133, + 223, + 216, + 265 + ], + "type": "image", + "image_path": "af4b6e6cae789bbe9d49613b9f3f3a246c0be79aeffb448fb5cef5823190b852.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 266, + 194, + 271 + ], + "lines": [ + { + "bbox": [ + 155, + 266, + 194, + 271 + ], + "spans": [ + { + "bbox": [ + 155, + 266, + 194, + 271 + ], + "type": "text", + "content": "Make Emotion Cuter" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 218, + 223, + 260, + 265 + ], + "blocks": [ + { + "bbox": [ + 218, + 223, + 260, + 265 + ], + "lines": [ + { + "bbox": [ + 218, + 223, + 260, + 265 + ], + "spans": [ + { + "bbox": [ + 218, + 223, + 260, + 265 + ], + "type": "image", + "image_path": "c66ae89a6a8613b54338b96d6e606533f6b8bcf44c27a05754b430e64e9f5707.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 262, + 223, + 304, + 265 + ], + "blocks": [ + { + "bbox": [ + 262, + 223, + 304, + 265 + ], + "lines": [ + { + "bbox": [ + 262, + 223, + 304, + 265 + ], + "spans": [ + { + "bbox": [ + 262, + 223, + 304, + 265 + ], + "type": "image", + "image_path": "d21adcd1f3ce5b6827141d0ee2983d49cf81ed28c9fd5e8167a645d93af71d19.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 306, + 223, + 348, + 265 + ], + "blocks": [ + { + "bbox": [ + 306, + 223, + 348, + 265 + ], + "lines": [ + { + "bbox": [ + 306, + 223, + 348, + 265 + ], + "spans": [ + { + "bbox": [ + 306, + 223, + 348, + 265 + ], + "type": "image", + "image_path": "32a37f0cd5345c4398dbc102a13cd0f7acfd4d7e9c3c6bf353a687055fccda08.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 351, + 223, + 392, + 265 + ], + "blocks": [ + { + "bbox": [ + 351, + 223, + 392, + 265 + ], + "lines": [ + { + "bbox": [ + 351, + 223, + 392, + 265 + ], + "spans": [ + { + "bbox": [ + 351, + 223, + 392, + 265 + ], + "type": "image", + "image_path": "6c0dd084fe17ee444ade75f0942d8c7cfa8ffc09a254c6e3d21a31efe6592fb2.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 394, + 223, + 436, + 265 + ], + "blocks": [ + { + "bbox": [ + 394, + 223, + 436, + 265 + ], + "lines": [ + { + "bbox": [ + 394, + 223, + 436, + 265 + ], + "spans": [ + { + "bbox": [ + 394, + 223, + 436, + 265 + ], + "type": "image", + "image_path": "6ac2f3fe89e5754757a57deb6cb3e5f3483ce585f81633c05d512db48f99b5dc.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 438, + 223, + 480, + 265 + ], + "blocks": [ + { + "bbox": [ + 438, + 223, + 480, + 265 + ], + "lines": [ + { + "bbox": [ + 438, + 223, + 480, + 265 + ], + "spans": [ + { + "bbox": [ + 438, + 223, + 480, + 265 + ], + "type": "image", + "image_path": "285f2962961afa9dcb0489d9d636db5f96227bfd7070e70f61992e38fe29d0ac.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 132, + 273, + 216, + 315 + ], + "blocks": [ + { + "bbox": [ + 132, + 273, + 216, + 315 + ], + "lines": [ + { + "bbox": [ + 132, + 273, + 216, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 273, + 216, + 315 + ], + "type": "image", + "image_path": "16b6014d961b8b145318bf832d3f61d44de62776020b5c0e888d59fda80e9363.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 316, + 187, + 321 + ], + "lines": [ + { + "bbox": [ + 160, + 316, + 187, + 321 + ], + "spans": [ + { + "bbox": [ + 160, + 316, + 187, + 321 + ], + "type": "text", + "content": "Cuter Emotion" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 336, + 480, + 359 + ], + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 359 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 359 + ], + "type": "text", + "content": "Fig. 7: Ablation Study Results for both in-domain (first two rows), and out-of-domain first two rows) image manipulations" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 218, + 273, + 260, + 315 + ], + "blocks": [ + { + "bbox": [ + 218, + 273, + 260, + 315 + ], + "lines": [ + { + "bbox": [ + 218, + 273, + 260, + 315 + ], + "spans": [ + { + "bbox": [ + 218, + 273, + 260, + 315 + ], + "type": "image", + "image_path": "9956882be333de8fc4bcd4d056c7b61fa64ff3941c5c46d83d214c8c86f52d41.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 262, + 273, + 304, + 315 + ], + "blocks": [ + { + "bbox": [ + 262, + 273, + 304, + 315 + ], + "lines": [ + { + "bbox": [ + 262, + 273, + 304, + 315 + ], + "spans": [ + { + "bbox": [ + 262, + 273, + 304, + 315 + ], + "type": "image", + "image_path": "1e38f74099ac8b8b755e791e0db43a3e6895b754b502306f5e6e1c993194d62a.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 306, + 273, + 348, + 315 + ], + "blocks": [ + { + "bbox": [ + 306, + 273, + 348, + 315 + ], + "lines": [ + { + "bbox": [ + 306, + 273, + 348, + 315 + ], + "spans": [ + { + "bbox": [ + 306, + 273, + 348, + 315 + ], + "type": "image", + "image_path": "c88305bdd51f8c03e042473ad850b1c3b01e822f218c3f985a850716ee503d09.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 351, + 273, + 392, + 315 + ], + "blocks": [ + { + "bbox": [ + 351, + 273, + 392, + 315 + ], + "lines": [ + { + "bbox": [ + 351, + 273, + 392, + 315 + ], + "spans": [ + { + "bbox": [ + 351, + 273, + 392, + 315 + ], + "type": "image", + "image_path": "5e809d829b6b4eb4d973f8ba5ee685c671e5082ac363545ca39790ba525a7eef.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 394, + 273, + 436, + 315 + ], + "blocks": [ + { + "bbox": [ + 394, + 273, + 436, + 315 + ], + "lines": [ + { + "bbox": [ + 394, + 273, + 436, + 315 + ], + "spans": [ + { + "bbox": [ + 394, + 273, + 436, + 315 + ], + "type": "image", + "image_path": "0b45f0363aa469f07c35def0283f7e48e83d4046adddca4baab0535749779c79.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 438, + 273, + 480, + 315 + ], + "blocks": [ + { + "bbox": [ + 438, + 273, + 480, + 315 + ], + "lines": [ + { + "bbox": [ + 438, + 273, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 438, + 273, + 480, + 315 + ], + "type": "image", + "image_path": "924aa9236d63901c0339c7872f03f266a6e61668d908244cd7d8981898549db4.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "table", + "bbox": [ + 134, + 431, + 479, + 476 + ], + "blocks": [ + { + "bbox": [ + 132, + 376, + 480, + 421 + ], + "lines": [ + { + "bbox": [ + 132, + 376, + 480, + 421 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 480, + 421 + ], + "type": "text", + "content": "Table 2: Abalation Study results of comparison between our method with all components and without the Reformed Conditioned Latent Diffusion Model (RCLDM), Editing Shift Matching (ESM), Language Instruction Unification (LIU) and Selective Area Matching (SAM). The ablation study is conducted on the entire test dataset." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 431, + 479, + 476 + ], + "lines": [ + { + "bbox": [ + 134, + 431, + 479, + 476 + ], + "spans": [ + { + "bbox": [ + 134, + 431, + 479, + 476 + ], + "type": "table", + "html": "
w/o RCLDMw/o ESMw/o LIUw/o SAMOurs
FID ↓10.159.2310.3711.317.57
CLIP DirSim ↑0.130.150.230.190.27
", + "image_path": "678c636c857138e1f9b75fc28dc91e8e1e67b970fc974d97ccfa2752bdacc0c9.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "table_body" + } + ], + "index": 36 + }, + { + "bbox": [ + 132, + 504, + 219, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 219, + 516 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 219, + 516 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": "In this work, we propose InstructGIE, an image editing framework with enhanced generalization ability, improving performance in both visual and text aspects. We incorporate a VMamba-based module to enhance visual outputs and introduce an editing-shift matching strategy to augment in-context learning. Our selective area-matching technique addresses and rectifies corrupted details, while a language unification technique aligns language embeddings with editing semantics. Additionally, we compile a publicly available dataset for better generalization evaluation. Extensive experiments demonstrate our framework's superior in-context generation performance and robust generalization capability across unseen vision tasks, both quantitative and qualitively." + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "text", + "content": "Z. Meng and C. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 481, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 481, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 481, + 187 + ], + "type": "text", + "content": "This work is partially supported by the Army Research Office/Army Research Laboratory via grant W911-NF-20-1-0167 to Northeastern University, National Science Foundation CCF-1937500, and the Fundamental Research Funds for the Central Universities, Peking University." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 228, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "type": "text", + "content": "1. Bar, A., Gandelsman, Y., Darrell, T., Globerson, A., Efros, A.: Visual prompting via image inpainting. Advances in Neural Information Processing Systems 35, 25005-25017 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 262, + 480, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 480, + 283 + ], + "type": "text", + "content": "2. Blau, T., Ganz, R., Kawar, B., Bronstein, A., Elad, M.: Threat model-agnostic adversarial defense using diffusion models. arXiv preprint arXiv:2207.08089 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 285, + 480, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 480, + 316 + ], + "type": "text", + "content": "3. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 317, + 480, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 317, + 480, + 349 + ], + "spans": [ + { + "bbox": [ + 138, + 317, + 480, + 349 + ], + "type": "text", + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 350, + 480, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 480, + 370 + ], + "type": "text", + "content": "5. Chen, W., Hu, H., Sahara, C., Cohen, W.W.: Re-imagen: Retrieval-augmented text-to-image generator. arXiv preprint arXiv:2209.14491 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 372, + 480, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 372, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 138, + 372, + 480, + 414 + ], + "type": "text", + "content": "6. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290–1299 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "type": "text", + "content": "7. Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8188-8197 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "text", + "content": "8. Chung, J.S., Zisserman, A.: Lip reading in the wild. In: Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13. pp. 87-103. Springer (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 481, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 481, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 481, + 480, + 512 + ], + "type": "text", + "content": "9. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 514, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 533 + ], + "type": "text", + "content": "0. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 535, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 535, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 134, + 535, + 480, + 577 + ], + "type": "text", + "content": "11. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 134, + 578, + 480, + 610 + ], + "type": "text", + "content": "12. Gal, R., Patashnik, O., Maron, H., Chechik, G., Cohen-Or, D.: Stylegan-nada: Clipguided domain adaptation of image generators. arXiv preprint arXiv:2108.00946 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "type": "text", + "content": "13. Geng, X., Liu, H.: Openllama: An open reproduction of llama (May 2023), https://github.com/openlm-research/open_llama" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 633, + 480, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 664 + ], + "type": "text", + "content": "14. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 239, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 239, + 91, + 447, + 102 + ], + "type": "text", + "content": "InstructGIE: Towards Generalizable Image Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "15. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 173 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 173 + ], + "type": "text", + "content": "16. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 480, + 205 + ], + "type": "text", + "content": "17. Huang, Y., Huang, J., Liu, Y., Yan, M., Lv, J., Liu, J., Xiong, W., Zhang, H., Chen, S., Cao, L.: Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 206, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 480, + 239 + ], + "type": "text", + "content": "18. Huang, Y., Huang, J., Liu, Y., Yan, M., Lv, J., Liu, J., Xiong, W., Zhang, H., Chen, S., Cao, L.: Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "type": "text", + "content": "19. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 480, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 296 + ], + "type": "text", + "content": "20. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. Advances in Neural Information Processing Systems 35, 23593-23606 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 297, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 319 + ], + "type": "text", + "content": "21. Kawar, B., Ganz, R., Elad, M.: Enhancing diffusion-based image synthesis with robust classifier guidance. arXiv preprint arXiv:2208.08664 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 319, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 480, + 363 + ], + "type": "text", + "content": "22. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 363, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 480, + 407 + ], + "type": "text", + "content": "23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Pereira, F., Burges, C., Bottou, L., Weinberger, K. (eds.) Advances in Neural Information Processing Systems. vol. 25. Curran Associates, Inc. (2012)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 430 + ], + "type": "text", + "content": "24. Kwon, G., Ye, J.C.: Diffusion-based image translation using disentangled style and content representation. arXiv preprint arXiv:2209.15264 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 430, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 480, + 475 + ], + "type": "text", + "content": "25. Liu, X., Park, D.H., Azadi, S., Zhang, G., Chopikyan, A., Hu, Y., Shi, H., Rohrbach, A., Darrell, T.: More control for free! image synthesis with semantic diffusion guidance. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 289-299 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 475, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 497 + ], + "type": "text", + "content": "26. Liu, Y., Tian, Y., Zhao, Y., Yu, H., Xie, L., Wang, Y., Ye, Q., Liu, Y.: Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "type": "text", + "content": "27. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of International Conference on Computer Vision (ICCV) (December 2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 531, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 480, + 564 + ], + "type": "text", + "content": "28. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 565, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 480, + 598 + ], + "type": "text", + "content": "29. Mohammad, S., Kiritchenko, S.: Wikiart emotions: An annotated dataset of emotions evoked by art. In: Proceedings of the eleventh international conference on language resources and evaluation (LREC 2018) (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 599, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 621 + ], + "type": "text", + "content": "30. Nguyen, T., Li, Y., Ojha, U., Lee, Y.J.: Visual instruction inversion: Image editing via visual prompting. arXiv preprint arXiv:2307.14331 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "type": "text", + "content": "31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "text", + "content": "Z. Meng and C. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 544 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "text", + "content": "33. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "34. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "35. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "36. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "text", + "content": "37. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "38. Sun, Y., Yang, Y., Peng, H., Shen, Y., Yang, Y., Hu, H., Qiu, L., Koike, H.: Imagebrush: Learning visual in-context instructions for exemplar-based image manipulation. arXiv preprint arXiv:2308.00906 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "text", + "content": "39. Theis, L., Salimans, T., Hoffman, M.D., Mentzer, F.: Lossy compression with gaussian diffusion. arXiv preprint arXiv:2206.08889 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "40. Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6830-6839 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "41. Wang, Z., Jiang, Y., Lu, Y., He, P., Chen, W., Wang, Z., Zhou, M., et al.: Incontext learning unlocked for diffusion models. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "42. Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., Xiao, J.: Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365 (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "text", + "content": "43. Zablotskaia, P., Siarohin, A., Zhao, B., Sigal, L.: Dwnet: Dense warp-based network for pose-guided human video generation. arXiv preprint arXiv:1910.09139 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "text", + "content": "44. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 522, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 544 + ], + "type": "text", + "content": "45. Zimmermann, R.S., Schott, L., Song, Y., Dunn, B.A., Klindt, D.A.: Score-based generative classifiers. arXiv preprint arXiv:2110.00473 (2021)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 239, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 239, + 91, + 447, + 102 + ], + "type": "text", + "content": "InstructGIE: Towards Generalizable Image Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/InstructIR_ High-Quality Image Restoration Following Human Instructions/86f2cef3-3642-43ee-930e-b6ddf921d714_content_list.json b/2024/InstructIR_ High-Quality Image Restoration Following Human Instructions/86f2cef3-3642-43ee-930e-b6ddf921d714_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..24b97d68d410cc5ce75ee9c96c59f352aa768d22 --- /dev/null +++ b/2024/InstructIR_ High-Quality Image Restoration Following Human Instructions/86f2cef3-3642-43ee-930e-b6ddf921d714_content_list.json @@ -0,0 +1,3289 @@ +[ + { + "type": "text", + "text": "InstructIR: High-Quality Image Restoration Following Human Instructions", + "text_level": 1, + "bbox": [ + 250, + 140, + 751, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Marcos V. Conde1,2, Gregor Geigle1, and Radu Timofte1", + "bbox": [ + 282, + 210, + 717, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Computer Vision Lab, CAIDAS & IFI, University of Würzburg", + "bbox": [ + 282, + 238, + 718, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Sony PlayStation, FTG", + "bbox": [ + 415, + 253, + 586, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/mv-lab/InstructIR", + "bbox": [ + 359, + 268, + 640, + 281 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f042f64cdc8d87bede4f89391118e4f2275a1a2bc840ec5a90bb5597a16bba4d.jpg", + "image_caption": [ + "Can you stabilize this pie? I was running..." + ], + "image_footnote": [], + "bbox": [ + 218, + 303, + 352, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9dd094cb6c5f041e6306f3a5393cf19a191ec10a2123ccb53550381e6f70f876.jpg", + "image_caption": [ + "My image has tiny dots, doesn't look great, fix it" + ], + "image_footnote": [], + "bbox": [ + 352, + 303, + 467, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2d2022e9d107b43ad89835ee2cca6ba1453ba164985093492f391320b734ce7c.jpg", + "image_caption": [ + "My photo is too dark, I cannot see anything" + ], + "image_footnote": [], + "bbox": [ + 468, + 303, + 573, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2bf86199a634bc594aae62f760f80cb338a58b5d2eafb3e6a9711a26a5465b25.jpg", + "image_caption": [ + "Retouch my picture like a photographer" + ], + "image_footnote": [], + "bbox": [ + 573, + 303, + 671, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/38d7d6a663ed5b99ffee6dfc67dad3b49e4ecb144dd689daacad416089543c08.jpg", + "image_caption": [ + "It is a foggy day, can you make it clearer?" + ], + "image_footnote": [], + "bbox": [ + 673, + 303, + 784, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0e811c9713c52e4d60a40020c5aaa1e0c6cd94acd1ed4ed44710756e7e299928.jpg", + "image_caption": [ + "Fig. 1: Given an image and a prompt for how to improve that image, our all-in-one restoration model corrects the image considering the human instruction. InstructIR, can tackle various types and levels of degradation, and it is able to generalize in some real-world scenarios (last three images, from left to right)." + ], + "image_footnote": [], + "bbox": [ + 217, + 406, + 352, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8614232ac2764f401353daa5357f0876ac53be1e56804dca5239d99d4fb733b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 406, + 468, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/56ee4c73e5aa2082c5add852c5d29587a1cec94d89bc2677ab1f5e06aac2d923.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 468, + 406, + 573, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c76e60981c7a373272d75a7079855c5531ee8e13949234ff9e266fbbd61e2105.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 406, + 671, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f033464962ab159b067523ee26c82659a35232c6d93325dac3b5711cd6c7db66.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 406, + 784, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Image restoration is a fundamental problem that involves recovering a high-quality clean image from its degraded observation. All-In-One image restoration models can effectively restore images from various types and levels of degradation using degradation-specific information as prompts to guide the restoration model. In this work, we present the first approach that uses human-written instructions to guide the image restoration model. Given natural language prompts, our model can recover high-quality images from their degraded counterparts, considering multiple degradation types. Our method, InstructIR, achieves state-of-the-art results on several restoration tasks including image denoising, deraining, deblurring, dehazing, and (low-light) image enhancement. InstructIR improves $+1\\mathrm{dB}$ over previous all-in-one restoration methods. Moreover, our dataset and results represent a novel benchmark for new research on text-guided image restoration and enhancement.", + "bbox": [ + 259, + 568, + 740, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 782, + 375, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Images often contain unpleasant effects such as noise, motion blur, haze, and low dynamic range. Such effects are commonly known in low-level computer", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "vision as degradations. These can result from camera limitations or challenging environmental conditions e.g. low light.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image restoration aims to recover a high-quality image from its degraded counterpart. This is a complex inverse problem since multiple different solutions can exist for restoring any given image [13, 16, 38, 52, 90, 91].", + "bbox": [ + 212, + 176, + 782, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some methods focus on specific degradations, for instance reducing noise (denoising) [57, 90, 91], removing blur (deblurring) [51, 92], or clearing haze (dehazing) [13, 58]. Such methods are effective for their specific task, yet they do not generalize well to other types of degradation. Other approaches use a general neural network for diverse tasks [7, 66, 73, 83], yet training the neural network for each specific task independently. Since using a separate model for each possible degradation is resource-intensive, recent approaches propose All-in-One restoration models [36, 53, 54, 89]. These approaches use a single deep blind restoration model considering multiple degradation types and levels. Contemporary works such as PromptIR [54] or ProRes [42] utilize a unified model for blind image restoration using learned guidance vectors, also known as \"prompt embeddings\", in contrast to raw user prompts in text form, which we use in this work.", + "bbox": [ + 212, + 222, + 784, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In parallel, recent works such as InstructPix2Pix [4] show the potential of using text prompts to guide image generation and editing models. However, this method (or recent alternatives) do not tackle inverse problems. Inspired by these works, we argue that text guidance can help to guide blind restoration models better than the image-based degradation classification used in previous works [36,53,89]. Users generally have an idea about what has to be fixed (though they might lack domain-specific vocabulary) so we can use this information to guide the model.", + "bbox": [ + 212, + 402, + 785, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions We propose the first approach that utilizes real human-written instructions to solve multi-task image restoration. Our comprehensive experiments demonstrate the potential of using text guidance for image restoration and enhancement by achieving state-of-the-art performance on various image restoration tasks, including image denoising, deraining, deblurring, dehazing, and low-light image enhancement. Our model, InstructIR, is able to generalize to restoring images using complex human-written instructions. Moreover, our single all-in-one model covers more tasks than many previous works. We show diverse restoration samples of our method in Figure 1.", + "bbox": [ + 212, + 530, + 787, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 688, + 387, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Restoration. Recent deep learning methods [13, 38, 51, 57, 66, 83] have shown consistently better results compared to traditional techniques for blind image restoration [14,25,31,33,47,65]. The proposed neural networks are based on convolutional neural networks (CNNs) and Transformers [68] (or related attention mechanisms). We focus on general-purpose restoration models [7,38,73,83]. For example, SwinIR [38], MAXIM [66] and Uformer [73]. These models can be trained -independently- for diverse tasks such as denoising, deraining or deblurring. Their ability to capture local and global feature interactions, and enhance", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Conde, Marcos et al.", + "bbox": [ + 271, + 114, + 410, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "them, allows the models to achieve great performance consistently across different tasks. For instance, Restormer [83] uses non-local blocks [71] to capture complex features across the image.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "NAFNet [7] is an efficient alternative to complex transformer-based methods. The model uses simplified channel attention, and gating as an alternative to non-linear activations. The building block (NAFBlock) follows a simple metaformer [82] architecture with efficient inverted residual blocks [27]. In this work, we build our InstructIR model using NAFNet as backbone, due to its efficient and simple design, and high performance in several restoration tasks.", + "bbox": [ + 212, + 191, + 784, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "All-in-One Image Restoration. Single degradation (or single task) restoration methods are well-studied, however, their real-world applications are limited due to the required resources i.e. allocating different models, and selecting the adequate model on demand. Moreover, images rarely present a single degradation, for instance, noise and blur are almost ubiquitous in any image capture.", + "bbox": [ + 212, + 292, + 784, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "All-in-One (also known as multi-degradation or multi-task) image restoration is emerging as a new research field in low-level computer vision [36,42,53,54,67, 81,87,88]. These approaches use a single deep blind restoration model to tackle different degradation types and levels.", + "bbox": [ + 212, + 369, + 784, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use as reference AirNet [36], IDR [89] and ADMS [53]. We also consider the contemporary work PromptIR [54]. The methods use different techniques to guide the blind model in the restoration process. For instance, an auxiliary model for degradation classification [36, 53], or multi-dimensional guidance vectors (also known as \"prompts\") [41, 42, 54] that help the model to discriminate the different types of degradation in the image.", + "bbox": [ + 212, + 431, + 784, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-guided Image Manipulation. In recent years, multiple methods have been proposed for text-to-image generation and text-based image editing works [4, 26, 30, 46, 62]. These models use text prompts to describe images or actions, and powerful diffusion-based models for generating the corresponding images. Our main reference is InstructPix2Pix [4], this method enables editing from instructions that tell the model what action to perform, as opposed to text labels, captions or descriptions of the input or output images. Therefore, the user can transmit what to do in natural written text, without requiring to provide further image descriptions or sample reference images.", + "bbox": [ + 212, + 532, + 784, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Image Restoration Following Instructions", + "text_level": 1, + "bbox": [ + 214, + 696, + 650, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We treat instruction-based image restoration as a supervised learning problem similar to previous works [4]. First, we generate over 10000 prompts using GPT-4 based on our own sample instructions. We explain the creation of the prompt dataset in Sec. 3.1. We then build a large paired training dataset of prompts and degraded/clean images. Finally, we train our InstructIR model, and we evaluate it on a wide variety of instructions including real human-written prompts. We explain our text encoder in Sec 3.2, and our complete model in Sec. 3.3.", + "bbox": [ + 212, + 734, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "InstructIR", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 Generating Prompts for Training", + "text_level": 1, + "bbox": [ + 215, + 146, + 535, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Why instructions? Inspired by InstructPix2Pix [4], we adopt human written instructions as the mechanism of control for our model. There is no need for the user to provide additional information, such as example clean images, or descriptions of the visual content. Instructions offer a clear and expressive way to interact, enabling users to pinpoint the unpleasant effects (degradations) in the images. We also consider the language complexity, from ambiguous instructions (e.g. \"fix my image\") to precise instructions (e.g. \"remove the noise\").", + "bbox": [ + 212, + 171, + 787, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Handling free-form user prompts rather than fixed degradation-specific prompts increases the usability of our model for laypeople who lack domain expertise. We thus want our model to be capable of understanding diverse prompts posed by users \"in-the-wild\" e.g. kids, adults, or photographers. To this end, we use a large language model (i.e., GPT-4) to create diverse requests that might be asked by users for the different degradations types. We then filter those generated prompts to remove ambiguous or unclear prompts (e.g., \"Make the image cleaner\", \"improve this image\"). Our final instructions set contains over 10000 different prompts in total, for 7 different tasks. We display some examples in Table 1. As we show in Figure 2 the prompts are sampled randomly depending on the input degradation.", + "bbox": [ + 212, + 285, + 488, + 632 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/dfd9b59d43050c5d32fce70e6bf35ef504cc36c4459fc34524bcb7eb23997a89.jpg", + "table_caption": [ + "Table 1: Examples of our curated GPT4-generated and real user prompts with varying language and domain expertise." + ], + "table_footnote": [], + "table_body": "
Degradation Prompts
DenoisingCan you clean the dots from my image? Fix the grainy parts of this photo Remove the noise from my picture
DeblurringCan you reduce the movement in the image? My picture's not sharp, fix it Deblur my picture, it's too fuzzy
DehazingCan you make this picture clearer? Help, my picture is all cloudy Remove the fog from my photo
DerainingI want my photo to be clear, not rainy Clear the rain from my picture Remove the raindrops from my photo
Super-Res.Make my photo bigger and better Add details to this image Increase the resolution of this photo
Low-lightThe photo is too dark, improve exposure Increase the illumination in this shot My shot has very low dynamic range
EnhancementMake it pop! Adjust the color balance for a natural look Apply a cinematic color grade to the photo
GeneralFix my image please make the image look better
", + "bbox": [ + 496, + 340, + 785, + 621 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9bbcf64a2b9c6ecf65d1245798bab7130f8c4a344ccf15552c000cbef1e1a42b.jpg", + "image_caption": [ + "Fig. 2: We train our blind image restoration models using common image datasets, and prompts generated using GPT-4, note that this is (self-)supervised learning. At inference time, our model generalizes to human-written instructions and restores (or enhances) the images." + ], + "image_footnote": [], + "bbox": [ + 217, + 659, + 787, + 773 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Conde, Marcos et al.", + "bbox": [ + 271, + 114, + 410, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Text Encoder", + "text_level": 1, + "bbox": [ + 215, + 146, + 374, + 159 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Choice of the Text Encoder. A text encoder maps the user prompt to a fixed-size vector representation (a text embedding). The related methods for text-based image generation [59] and manipulation [3,4] often use the text encoder of a CLIP model [55] to encode user prompts as CLIP excels in visual prompts. However, user prompts for degradation contain, in general, little to no visual content (e.g. the use describes the degradation, not the image itself). We opt, instead, to use a pure text-based sentence encoder [56], that is, a smaller model trained to encode sentences in a semantically meaningful embedding space. Sentence encoders - pre-trained with millions of examples - are compact and fast in comparison to CLIP, while being able to encode the semantics of diverse user prompts. For instance, we use the BGE-micro-v2 sentence transformer. We compare text encoders in the supplementary material.", + "bbox": [ + 212, + 175, + 787, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fine-tuning the Text Encoder. We want to adapt the text encoder E for the restoration task to better encode the required information for the restoration model. Training the full text encoder is likely to lead to overfitting on our small training set and lead to loss of generalization. Instead, we freeze the text encoder and train a projection head on top:", + "bbox": [ + 212, + 359, + 787, + 436 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {e} = \\operatorname {n o r m} (\\mathbf {W} \\cdot \\mathrm {E} (t)) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 446, + 784, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $t$ is the text, $\\mathrm{E}(t)$ represents the raw text embedding, $\\mathbf{W} \\in \\mathbb{R}^{d_t \\times d_v}$ is a learned projection from the text dimension $(d_t)$ to the input dimension for the restoration model $(d_v)$ , and norm is the 12-norm.", + "bbox": [ + 212, + 470, + 787, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 3 shows that while the text encoder is capable out-of-the-box to cluster instructions to some extent (Figure 3a), our trained projection yields greatly improved clusters (Figure 3b). We distinguish clearly the clusters for deraining, denoising, dehazing, deblurring, and low-light image enhancement. The instructions for such tasks or degradations are very characteristic. Furthermore, we can appreciate that \"super-res\" and \"enhancement\" tasks are quite spread and between the previous ones, which matches the language logic. For instance \"add details to this image\" could be used for enhancement, deblurring, or denoising. In our experiments, $d_{t} = 384$ , $d_{v} = 256$ and $\\mathbf{W}$ is a linear layer. The representation $\\mathbf{e}$ from the text encoder is shared across the blocks, and each block has a trainable projection $\\mathbf{W}$ .", + "bbox": [ + 212, + 518, + 787, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intent Classification Loss. We propose a guidance loss on the text embedding $\\mathbf{e}$ to improve training and interpretability. Using the degradation types as targets, we train a simple classification head $\\mathcal{C}$ such that $\\mathbf{c} = \\mathcal{C}(\\mathbf{e})$ , where $\\mathbf{c} \\in \\mathbb{R}^{D}$ , being $D$ is the number of degradation classes. The classification head $\\mathcal{C}$ is a simple two-layers MLP. Thus, we only need to train a projection layer $\\mathbf{W}$ and a simple MLP to capture the natural language knowledge. This allows the text model to learn meaningful embeddings as we can appreciate in Figure 3, not just guidance vectors for the main image processing model. We find that the model is able to classify accurately (i.e. over $95\\%$ accuracy) the underlying degradation in the user's prompt after a few epochs.", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "InstructIR", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ea91eaf38deadf1d7358efae1e44401fb00bdf1492e55ecbf948c468ec3a0c07.jpg", + "image_caption": [ + "(a) t-SNE of embeddings before training i.e. frozen text encoder" + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 483, + 227 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e185e23b332ca2daba02817a4bf95b7edc9ba593d262562571d35b704d3849cb.jpg", + "image_caption": [ + "(b) t-SNE of embeddings after training our learned projection", + "Fig. 3: We show t-SNE plots of the text embeddings before/after training InstructIR. Each dot represents a human instruction." + ], + "image_footnote": [], + "bbox": [ + 519, + 146, + 781, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 InstructIR", + "text_level": 1, + "bbox": [ + 215, + 313, + 349, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method InstructIR consists of an image model and a text encoder. We introduced our text encoder in Sec. 3.2. We use NAFNet [7] as the image model, an efficient image restoration model that follows a U-Net architecture [60]. To successfully learn multiple tasks using a single model, we use task routing techniques. Our framework for training and evaluation is illustrated in Figure 2.", + "bbox": [ + 212, + 338, + 782, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text Guidance. The key aspect of InstructIR is the integration of the encoded instruction as a mechanism of control for the image model. Inspired in task routing for many-task learning [11, 61, 63], we propose an \"Instruction Condition Block\" (ICB) to enable task-specific transformations within the model. Conventional task routing [63] applies task-specific binary masks to the channel features. Since our model does not know a priori the degradation, we cannot use this technique directly.", + "bbox": [ + 212, + 428, + 491, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Considering the image features $\\mathcal{F}$ and the encoded instruction $\\mathbf{e}$ , we apply task routing as follows:", + "bbox": [ + 212, + 640, + 491, + 685 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/55a6c6b16f37b8e82e7dab78d358b43fa9e5915cba3093a16b8577cf0966cc5d.jpg", + "image_caption": [ + "Fig. 4: Instruction Condition Block (ICB) using an approximation of task routing [63] for many-tasks learning (See Eq. 2). This mechanism allows the neural network to select and prioritize specific features depending on the instruction, similarly to a Mixture of Experts (MoE)." + ], + "image_footnote": [], + "bbox": [ + 504, + 433, + 777, + 537 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {c} ^ {\\prime} = \\operatorname {B l o c k} \\left(\\mathcal {F} _ {c} \\odot \\mathbf {m} _ {c}\\right) + \\mathcal {F} _ {c} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 696, + 785, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the mask $\\mathbf{m}_c = \\sigma (\\mathbf{W_c}\\cdot \\mathbf{e})$ is produced using a linear layer $\\mathbf{W}_{\\mathbf{c}}$ activated using the Sigmoid function - to produce a set of weights depending on the text embedding $\\mathbf{e}$ . Thus, we obtain a $c$ -dimensional per-channel (soft-)binary mask $\\mathbf{m}_c$ . As [24, 63], task routing is applied as the channel-wise multiplication $\\odot$ for masking features depending on the task. The conditioned features are further enhanced using a convolutional NAFBlock [7] (Block). We illustrate our task-routing ICB block in Figure 4. We use \"regular\" NAFBlocks [7], followed by ICBs to condition the features, at both encoder and decoder blocks. The", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Conde, Marcos et al.", + "bbox": [ + 271, + 114, + 410, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "formulation is $F^{l + 1} = \\mathrm{ICB}(\\mathrm{Block}(F^l))$ where $l$ is the layer. Although we do not condition explicitly the filters of the neural network, as in [63], the mask allows the model to select the most relevant channels depending on the image information and the instruction. Note that this formulation enables differentiable feature masking, and certain interpretability i.e. the features with high weights contribute the most to the restoration process. Indirectly, this also enforces to learn diverse filters and reduce sparsity [11, 63].", + "bbox": [ + 212, + 145, + 782, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Is InstructIR a blind restoration model? The model does not use explicit information about the degradation in the image e.g. noise profiles, blur kernels, or PSFs. Since our model infers the task (degradation) given the image and the instruction, we consider InstructIR a blind image restoration model. Similarly to previous works that use auxiliary image-based degradation classification [36,53].", + "bbox": [ + 212, + 261, + 784, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experimental Results", + "text_level": 1, + "bbox": [ + 214, + 362, + 460, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate our model on 9 well-known benchmarks for different image restoration tasks: image denoising, deblurring, deraining, dehazing, real low-light enhancement, and photo-realistic image enhancement. We present extensive quantitative results in Table 2 and Table 3. We provide extensive comparisons with other all-in-one methods as well as task-specific methods. Our single model successfully restores images considering different degradation types and levels. We provide additional results and ablation studies in the supplementary material.", + "bbox": [ + 212, + 396, + 782, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Implementation Details.", + "text_level": 1, + "bbox": [ + 214, + 526, + 460, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our InstructIR model is end-to-end trainable. The image model does not require pre-training but we use a pre-trained sentence encoder as language model.", + "bbox": [ + 212, + 553, + 782, + 584 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text Encoder. As we discussed in Sec. 3.2, we only need to train the text embedding projection and classification head ( $\\approx$ 100K parameters). We initialize the text encoder with BGE-MICRO-v2 $^{3}$ , a distilled version of BGE-SMALL-EN [76]. The BGE encoders are BERT-like encoders [10] pre-trained on large amounts of supervised and unsupervised data for general-purpose sentence encoding. The BGE-micro model is a 3-layer encoder with 17.3 million parameters, which we freeze during training. We also explore ALL-MINILM-L6-v2 and CLIP encoders, however, we concluded that small models prevent overfitting and provide the best performance while being fast. We provide the ablation study comparing the three text encoders in the supplementary material.", + "bbox": [ + 212, + 593, + 784, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image Model. We use NAFNet [7] as the image model backbone. The architecture consists of a 4-level encoder-decoder, with varying numbers of blocks at each level, specifically [2, 2, 4, 8] for the encoder, and [2, 2, 2, 2] for the decoder, from the level-1 to level-4 respectively. Between the encoder and decoder we use", + "bbox": [ + 212, + 753, + 782, + 814 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "InstructIR", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "3 https://huggingface.co/TaylorAI/bge-micro-v2", + "bbox": [ + 217, + 823, + 576, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 middle blocks to enhance further the features. The decoder implements addition instead of concatenation for the skip connections. We use the Instruction Condition Block (ICB) for task-routing [63] only in the encoder and decoder.", + "bbox": [ + 212, + 146, + 785, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The model is optimized using the $\\mathcal{L}_1$ loss between the ground-truth clean image and the restored one. Additionally, we use the cross-entropy loss $\\mathcal{L}_{ce}$ for the intent classification head of the text encoder. We train using a batch size of 32 and AdamW [32] optimizer with learning rate $5e^{-4}$ for 500 epochs (approximately 1 day using a single NVIDIA A100). We also use cosine annealing learning rate decay. During training, we utilize cropped patches of size $256 \\times 256$ as input, and we use random horizontal and vertical flips as augmentations. Since our model uses as input instruction-image pairs, given an image, and knowing its degradation, we randomly sample instructions from our prompt dataset ( $>10\\mathrm{K}$ samples). Our image model has only 16M parameters, and the learned text projection is just 100k parameters (the language model is 17M parameters), thus, our model can be trained easily on standard GPUs such as NVIDIA RTX 2080Ti or 3090Ti in a couple of days. Furthermore, the inference process also fits in low-computation budgets (e.g. Google Colab T4 16Gb GPU).", + "bbox": [ + 212, + 191, + 787, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Datasets and Benchmarks", + "text_level": 1, + "bbox": [ + 214, + 421, + 477, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Following previous works [36,54,89], we prepare the datasets for different restoration tasks, including real and synthetic datasets.", + "bbox": [ + 212, + 443, + 785, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Image denoising. We use a combination of BSD400 [2] and WED [43] datasets for training. This combined training set contains $\\approx 5000$ images. Using as reference the clean images in the dataset, we generate the noisy images by adding Gaussian noise with different noise levels $\\sigma \\in \\{15, 25, 50\\}$ . We test the models on the well-known BSD68 [45] and Urban100 [28] datasets.", + "bbox": [ + 212, + 470, + 785, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Image deraining. We use the Rain100L [78] dataset, which consists of 200 clean-rainy image pairs for training, and 100 pairs for testing.", + "bbox": [ + 212, + 544, + 785, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Image dehazing. We utilize the Reside (outdoor) SOTS [35] dataset, which contains $\\approx 72\\mathrm{K}$ training images. However, many images are low-quality and unrealistic, thus, we filtered the dataset and selected a random set of 2000 images - also to avoid imbalance w.r.t the other tasks. We use the standard outdoor test set of 500 images.", + "bbox": [ + 212, + 571, + 785, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Image deblurring. We use the GoPro dataset for motion deblurring [50] which consists of 2103 images for training, and 1111 for testing.", + "bbox": [ + 212, + 643, + 785, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Real-world Low-light Image Enhancement. We use the LOL [74] dataset (v1), which contains real-case low/normal-light image pairs. We adopt its official split of 485 training images and 15 testing images.", + "bbox": [ + 212, + 671, + 785, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Real-world Image Enhancement. Extending previous works, we also study photorealistic image enhancement using the MIT5K DSLR dataset [5]. We use 1000 images for training, and the standard split of 500 images for testing (as in [66]).", + "bbox": [ + 212, + 714, + 785, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, as previous works [36,54,89], we combine all the aforementioned training datasets, and we train our unified model for all-in-one restoration. Note that we do not include more real-world datasets because previous works do not provide results (or models) for those. Moreover, previous works were limited to synthetic data, in contrast, InstructIR also tackles real-world image enhancement.", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Conde, Marcos et al.", + "bbox": [ + 271, + 114, + 410, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/65889baca9ae7456dd5002331e32c162aa745ad274ca82d82af410111e95ebd2.jpg", + "table_caption": [ + "Table 2: Quantitative results on five restoration tasks (5D) with state-of-the-art general image restoration and all-in-one methods. We highlight the reference model without text (image only), the best overall results, and the second best results. We also present the ablation study of our multi-task variants (from 5 to 7 tasks — 5D, 6D, 7D). This table is based on Zhang et al. IDR [89]." + ], + "table_footnote": [], + "table_body": "
MethodsDerainingDehazingDenoisingDeblurringLow-light Enh.AverageParams
Rain100L [78]SOTS [35]BSD68 [45]GoPro [50]LOL [74]
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑(M)
HINet [8]35.670.96924.740.93731.000.88126.120.78819.470.80027.400.87588.67
DGUNet [49]36.620.97124.780.94031.100.88327.250.83721.870.82328.320.89117.33
MIRNetV2 [84]33.890.95424.030.92730.970.88126.300.79921.520.81527.340.8755.86
SwinIR [38]30.780.92321.500.89130.590.86824.520.77317.810.72325.040.8350.91
Restormer [83]34.810.96224.090.92731.490.88427.220.82920.410.80627.600.88126.13
NAFNet [7]35.560.96725.230.93931.020.88326.530.80820.490.80927.760.88117.11
DL [17]21.960.76220.540.82623.090.74519.860.67219.830.71221.050.7432.09
Transweather [67]29.430.90521.320.88529.000.84125.120.75721.210.79225.220.83637.93
TAPE [39]29.670.90422.160.86130.180.85524.470.76318.970.62125.090.8011.07
AirNet [36]32.980.95121.040.88430.910.88224.350.78118.180.73525.490.8468.93
InstructIR w/o text35.580.96725.200.93831.090.88326.650.81020.700.82027.840.88417.11
IDR [89]35.630.96525.240.94331.600.88727.870.84621.340.82628.340.89315.34
InstructIR-5D36.840.97327.100.95631.400.88729.400.88623.000.83629.550.90715.8
InstructIR-6D36.800.97327.000.95131.390.88829.730.89222.830.83629.550.90815.8
InstructIR-7D36.750.97226.900.95231.370.88729.700.89222.810.83629.500.90715.8
", + "bbox": [ + 217, + 229, + 782, + 426 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/0e449264ea826cfe3d84c91dc02cb0a421d307be7ef625024682a0269cab6b3c.jpg", + "table_caption": [ + "Table 3: Comparisons of all-in-one restoration models for 3 restoration tasks (3D). We also show an ablation study for image denoising -the fundamental inverse problem-considering different noise levels. We report PSNR/SSIM metrics. Table based on [54]." + ], + "table_footnote": [], + "table_body": "
MethodsDehazing SOTS [35]Deraining Rain100L [17]Denoising ablation study σ = 1529.76/0.836 σ = 2526.34/0.836 σ = 50Average
BRDNet [64]23.23/0.89527.42/0.89532.26/0.89829.76/0.83626.34/0.83627.80/0.843
LPNet [20]20.84/0.82824.88/0.78426.47/0.77824.77/0.74821.26/0.55223.64/0.738
FDGAN [15]24.71/0.92429.89/0.93330.25/0.91028.81/0.86826.43/0.77628.02/0.883
MPRNet [85]25.28/0.95433.57/0.95433.54/0.92730.89/0.88027.56/0.77930.17/0.899
DL [17]26.92/0.93132.62/0.93133.05/0.91430.41/0.86126.90/0.74029.98/0.875
AirNet [36]27.94/0.96234.90/0.96733.92/0.93331.26/0.88828.00/0.79731.20/0.910
PromptIR [54]30.58/0.97436.37/0.97233.98/0.93331.31/0.88828.06/0.79932.06/0.913
InstructIR-3D30.22/0.95937.98/0.97834.15/0.93331.52/0.89028.30/0.80432.43/0.913
InstructIR-5D27.10/0.95636.84/0.97334.00/0.93131.40/0.88728.15/0.79831.50/0.909
InstructIR w/o text26.84/0.94834.02/0.96033.70/0.92930.94/0.88227.78/0.78030.65/0.900
", + "bbox": [ + 248, + 491, + 758, + 623 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Multiple Degradation Results", + "text_level": 1, + "bbox": [ + 215, + 651, + 504, + 667 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We define two initial setups for multi-task restoration:", + "bbox": [ + 215, + 679, + 606, + 694 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- 3D for three-degradation models such as AirNet [36], these tackle image denoising, dehazing and deraining.", + "- 5D for five-degradation models, considering image denoising, deblurring, dehazing, deraining and low-light image enhancement as in [89]." + ], + "bbox": [ + 223, + 705, + 785, + 767 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In Table 2, we show the performance of 5D models. Following Zhang et al. [89], we compare InstructIR with several state-of-the-art methods for general image restoration [7,8,38,83,84], and all-in-one image restoration methods [17,36, 39,67,89]. We can observe that our simple image model (just 16M parameters)", + "bbox": [ + 215, + 779, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "InstructIR", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "can tackle successfully at least five different tasks thanks to the instruction-based guidance and achieves the most competitive results. In Table 3 we can appreciate a similar behavior, when the number of tasks is just three (3D), our model improves further in terms of reconstruction performance.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Based on these results, we pose the following question: How many tasks can we tackle using a single model without losing too much performance? To answer this, we propose the 6D and 7D variants. For the 6D variant, we fine-tune the original 5D to consider also super-resolution as sixth task. Finally, our 7D model includes all previous tasks, and additionally image enhancement (MIT5K photo retouching). We show the performance of these two variants in Table 2.", + "bbox": [ + 212, + 207, + 787, + 297 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Test Instructions. InstructIR requires as input the degraded image and the human-written instruction. Therefore, we also prepare a test set of prompts i.e. instruction-image test pairs. The performance of InstructIR depends on the ambiguity and precision of the instruction. We provide the ablation study in Table 4. InstructIR is quite robust to more/less detailed instructions. However, it is still limited with ambiguous instructions such as \"enhance this image\". We show diverse instructions in Figures 5 and 6.", + "bbox": [ + 212, + 313, + 488, + 523 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 4: Ablation study on the sensitivity of instructions. We report PSNR/SSIM metrics for each task using our 5D base model. We repeat the evaluation on each test set 10 times, each time we sample different prompts for each image, and we report the average results. The \"Real Users $\\dagger$ \" in this study are amateur photographers, thus, the instructions were very precise.", + "bbox": [ + 496, + 328, + 787, + 453 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e38b7e5d009604fdb8842314cffc656bb6f6bc9c319ddc323e97a4f05096ba9e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Language LevelDerainingDenoisingDeblurringLOL
Basic & Precise36.84/0.97331.40/0.88729.47/0.88723.00/0.836
Basic & Ambiguous36.24/0.97031.35/0.88729.21/0.88521.85/0.827
Real Users †36.84/0.97331.40/0.88729.47/0.88723.00/0.836
", + "bbox": [ + 501, + 455, + 782, + 496 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Multi-Task Discussion and Study", + "text_level": 1, + "bbox": [ + 214, + 547, + 571, + 565 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "How does 6D work? Besides the 5 basic tasks -as previous works-, we include single image super-resolution (SISR). For this, we include as training data the DIV2K [1]. Since our model does not perform upsampling, we use the Bicubic degradation model [1,12] for generating the low-resolution images (LR), and the upsampled versions (HR) that are fed into our model to enhance them. Adding this extra task increases the performance on deblurring -a related degradation-, without harming notably the performance on the other tasks.", + "bbox": [ + 212, + 580, + 787, + 686 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "How does $7D$ work? Finally, if we add real image enhancement -a task not related to the previous ones i.e. inverse problems- the performance on the restoration tasks decays slightly. However, the model still achieves state-of-the-art results. Moreover, as we show in Table 5, the performance on this task using the MIT5K [5] dataset is notable, while keeping the performance on the other tasks.", + "bbox": [ + 212, + 700, + 545, + 838 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e8187c98199246e0ef5e185beacb3a8672cce34533ad4fdf316c557607041a1b.jpg", + "table_caption": [ + "Table 5: Real Image Enhancement results on MIT5K [5]." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑ SSIM↑ ΔEab ↓
UPE [69]21.880.85310.80
DPE [21]23.750.9089.34
HDRNet [9]24.320.9128.49
3DLUT [86]25.210.9227.61
InstructIR-7D24.650.9008.20
", + "bbox": [ + 558, + 731, + 782, + 809 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Conde, Marcos et al.", + "bbox": [ + 271, + 114, + 410, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/8bcccb2bada2614507419e35540c265557176728e894c331fa84a74e60ff2ca2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 143, + 359, + 205 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0ef09a6f541d427173452ee88a6f80ad132db1d5bbbe39f24dd49e5e809b8145.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 143, + 500, + 205 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/add9da21428c8935aa2c35ad808f39b92c6d3abc6049c3caeffaa5b200f8d36e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 143, + 638, + 205 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/81c2619cc80a3b4be0258702bf19df7a29061e39ab58b9b7d0516dbab4278f02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 143, + 779, + 205 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7b15fd1272c8a874060da8925a4c3edfc5944c2c860ccfb789898aa0e8fcd7a1.jpg", + "image_caption": [ + "Input", + "\"Retouch this image and improve colors\"", + "Fig. 5: Adversarial and OOD samples for Instruction-based Restoration. InstructIR understands a wide range of instructions for a given task (first row). Given an adversarial or out-of-distribution instruction (second row), the model does not modify the image notably (i.e. performs the identity) – we did not enforce this during training-." + ], + "image_footnote": [], + "bbox": [ + 223, + 236, + 359, + 299 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/cbd76114bdf1ec34600399e82596afbf57b92df835d5baf86ac36ceeb0c88d59.jpg", + "image_caption": [ + "\"Clean up my image, it's too fuzzy.\"", + "\"Reduce the motion in this shot\"" + ], + "image_footnote": [], + "bbox": [ + 362, + 236, + 500, + 299 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/19d4b1a93dff9047f2a26c4a36168ea84fdc642bc6c34fbf644761570f447d6b.jpg", + "image_caption": [ + "\"Get rid of the grain in my photo\"", + "\"Please get rid of the raindrops\"" + ], + "image_footnote": [], + "bbox": [ + 501, + 236, + 638, + 299 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/22af49b9e372b4790a0cd0e27a054aba1f0b8d1cd63a3c96920d6a2be17cb6ce.jpg", + "image_caption": [ + "\"Remove the strange spots on my photo\"", + "\"Reduce the fog in this landmark\"" + ], + "image_footnote": [], + "bbox": [ + 640, + 236, + 779, + 299 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/5259200baaf5aa970ffedb8ea00d91441423c3e1cce3572afc19ad880656d522.jpg", + "table_caption": [ + "Table 7: Quantitative comparisons with state-of-the-art methods on the LOL [74] dataset for Real-World Low-light Enhancement. Note that InstructIR-7D is a multitask method, while the other methods are task-specific. Table based on [72]." + ], + "table_footnote": [], + "table_body": "
MethodLPNet [37]URetinex -Net [75]DeepLPF [48]SCI [44]LIME [22]MF [18]NPE [70]SRIE [19]SDD [23]CDEF [34]InstructIR Ours
PSNR ↑21.4621.3215.2815.8016.7616.9616.9611.8613.3416.3322.81
SSIM ↑0.8020.8350.4730.5270.4440.5050.4810.4930.6350.5830.836
MethodDRBN [79]KinD [93]RUAS [40]FIDE [77]EG [29]MS-RDN [80]Retinex -Net [74]MIRNet [84]IPT [6]Uformer [73]IAGC [72]
PSNR ↑20.1320.8718.2318.2717.4817.2016.7724.1416.2716.3624.53
SSIM ↑0.8300.8000.7200.6650.6500.6400.5600.8300.5040.5070.842
", + "bbox": [ + 220, + 474, + 782, + 588 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We summarize the multi-task ablation study in Table 6. Our model can tackle multiple tasks without losing performance notably thanks to the instruction-based task routing.", + "bbox": [ + 215, + 614, + 486, + 690 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/2e5b74d24ff896c868ad44bef204da2ccb7db19edaff341ac225447b223dcc9e.jpg", + "table_caption": [ + "Table 6: Summary ablation study on the multi-task variants of InstructIR that tackle from 3 to 7 tasks." + ], + "table_footnote": [], + "table_body": "
TasksRainNoise (σ15)BlurLOL
3D37.98/0.97831.52/0.890--
5D36.84/0.97331.40/0.88729.40/0.88623.00/0.836
6D36.80 0.97331.39 0.88829.73/0.89222.83 0.836
7D36.75 0.97231.37 0.88729.70/0.89222.81 0.836
", + "bbox": [ + 501, + 655, + 782, + 718 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison with Task-specific Methods Our main goal is to design a powerful all-in-one model, thus, InstructIR was not designed to be trained for a particular degradation. Nevertheless, we also compare InstructIR with task-specific methods i.e. models tailored and trained for specific tasks.", + "bbox": [ + 215, + 741, + 787, + 803 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We compare with task-specific methods for real-world photography enhancement in Table 5, and for real-world low-light image enhancement in Table 7.", + "bbox": [ + 215, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "InstructIR", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/60416b4c0584616589a2dcc447494d0a3e7d7b6ab89c0de32db686af5e21767c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 143, + 408, + 233 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7634f3a49bcc4952134959d94071342db8ffbd6afcb27d2a8f4dc1d4a8ffe8b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 146, + 588, + 232 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0bf5d9047ce5bb3c306cc50271400608be89ef3463e4d7b1d027b91800bd5ee7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 593, + 148, + 771, + 232 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c8a16fb2224007a58f6b571e89331558607424a72f7a0b2240065ec8608012c1.jpg", + "image_caption": [ + "Input", + "(1) \"Retouch it as a photographer" + ], + "image_footnote": [], + "bbox": [ + 228, + 246, + 406, + 330 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e8192ce76b7575b1fdaf5fbc9e2d3118ecab16261e3f76575b0ab576a63ec396.jpg", + "image_caption": [ + "(1) \"Clear the rain from my picture\" $\\longrightarrow$ (2) \"Make it pop\"", + "$r'' \\longrightarrow 'Can you remove the raindrops?\" \\longrightarrow 'Increase the resolution''$" + ], + "image_footnote": [], + "bbox": [ + 408, + 246, + 586, + 330 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/08ca14f84c6597d74dd2d97a1ac11c654c940a388499bf5a081c51bed8631df4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 593, + 246, + 771, + 330 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7630abf7b0765598023e47eacecf3607c3f0dfdc2ebb8bcf62d0fac41b9d6abc.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 225, + 345, + 406, + 434 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/497093888fc3f303a84d83180981ad284859affdc056a0d8abd189d808097426.jpg", + "image_caption": [ + "(1) \"My image is too dark, fix it\" $\\longrightarrow$ (2) 'Apply a tonemap'" + ], + "image_footnote": [], + "bbox": [ + 408, + 345, + 586, + 433 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b59efc2ccaf1b29ca540a899374a2dbde8cf3960b9e693953a1be1ba9f58c1ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 593, + 345, + 771, + 431 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/97d321d57a875344929996d08e4630d2127fed00ff3a9e03ed2858f91a698e6b.jpg", + "image_caption": [ + "Fig. 6: Control via instructions. We can prompt multiple instructions (in sequence) to restore and enhance the images. This provides additional control. We show two examples of multiple instructions applied to the \"Input\" image -from left to right-.", + "Input (RealSRSet)", + "Fig.7: Comparison with InstructPix2Pix [4] using the prompt \"Remove the noise in this photo\". Real-case image from RealSRSet [38]." + ], + "image_footnote": [], + "bbox": [ + 236, + 515, + 364, + 592 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b4772aa72cdf41c53b2e9426a5b5dccb453f55b02e485e6e045c0a3ffe7356ae.jpg", + "image_caption": [ + "InstructIR", + "InstructPix2Pix #1 InstructPix2Pix #2" + ], + "image_footnote": [], + "bbox": [ + 366, + 516, + 493, + 592 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7c690650ec75ceffd9c5bc7652b1368a9c9997626cfe1e864746a3e621e9a192.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 516, + 624, + 592 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c0e64c0032ad5b926c3e1826b7d2c2278b8fe69dca2c36af265f62c08030eca0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 516, + 756, + 592 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.1 On the Effectiveness of Instructions", + "text_level": 1, + "bbox": [ + 214, + 675, + 555, + 689 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Thanks to our integration of human instructions, users can control how to enhance the images. We provide examples in Figures 5 and 6, where we show the potential of our method to restore and enhance images in a controllable manner.", + "bbox": [ + 212, + 700, + 785, + 744 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "This implies an advancement w.r.t classical (deterministic) image restoration methods. Classical deep restoration methods lead to a unique result, thus, they do not allow to control how the image is processed. We also compare InstructIR with InstructPix2Pix [4] (a diffusion-based generative model) in Figure 7.", + "bbox": [ + 212, + 746, + 787, + 806 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative Results. We provide diverse qualitative results for several tasks, and we compare with all-in-one and task-specific methods. In Figure 10, we show", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Conde, Marcos et al.", + "bbox": [ + 271, + 114, + 410, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Instruction: \"my colors are too off, make it pop so I can use these photos in instagram\"", + "bbox": [ + 240, + 146, + 756, + 157 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f3ad7deaa755c317646cd5f67d5fa8166f280b606740351fb765904cc572b7a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 159, + 362, + 226 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/715d36189c1551feff2188e486e25f156503ba81d5cb9b15e4035e37c789d4a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 159, + 496, + 226 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ad1a1fd835e7d4b8d80b514a23be1881227e3269c969d5515b85af5787881535.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 159, + 629, + 226 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ae72433d5937f559ffab2d268aac0bf187b1113e70afefbb9d5f470060c34dd0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 159, + 764, + 226 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "\"Increase the brightness of my photo please, is it totoro?\" / \"my image is too dark, fix it\"", + "bbox": [ + 230, + 231, + 767, + 243 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/730bb54bbc219579e5f3b13f1de38d9cd1dfbee411cf2ac96fb0f8381126c42b.jpg", + "image_caption": [ + "Fig. 8: Real-world samples of image restoration and enhancement using InstructIR." + ], + "image_footnote": [], + "bbox": [ + 230, + 243, + 362, + 311 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/969ecfbfdb19383e15c0f4ca2f279061cdca9f8f129cfdb964a585ff50bbea38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 243, + 496, + 311 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b1229dc4382b7055869832deb70100f77f37065fe172bbb3547fb03ee33e7c54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 243, + 629, + 311 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9da96949b25277e62dd4107bfc1528f7a062e27fc91a0768ebcbb82dbe645381.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 243, + 764, + 311 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9435cfd5a78f9d5f8d9e944ab690553c8e233ece96a8fbcedaa3fed60567f913.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 351, + 331, + 415 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/46c4466867627c5e860a5254080a700a8b44d9728240c751fa732e7886c2ac20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 352, + 442, + 415 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c70505c6af0166b68dec30cc3d0452d7beb4f09816cd16f6fcef2aa52cef4c15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 352, + 553, + 415 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/00bc1e382402d42498c92573b5951324819509bb0fd0f1006861aaac828b2eb0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 352, + 663, + 415 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/44b8bab9807a75e30954c2f56a8c1931275b0bcb5fdcf3f4275f201d21f3f8f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 352, + 774, + 415 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/be63fd17da14f825879323891e2fc1cc6541d59aee3209696722935c2a5ee5c2.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 222, + 417, + 331, + 465 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c3af3f60b8a75090cffc5fb877d38618b0107097d4fc7ff15f557416d3b7772b.jpg", + "image_caption": [ + "AirNet [36]" + ], + "image_footnote": [], + "bbox": [ + 334, + 419, + 442, + 465 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6233f7c954b3f97a09df14099a8a66a742b89d1a5d1a143adfee17293deffbd5.jpg", + "image_caption": [ + "PromptIR [54]" + ], + "image_footnote": [], + "bbox": [ + 444, + 419, + 553, + 465 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3f1def5a075c6dd824b66ee5514e81b80be6938430cc5366b39aa9652b2643a3.jpg", + "image_caption": [ + "InstructIR", + "Fig. 9: Dehazing comparisons for all-in-one restoration methods on SOTS [35]." + ], + "image_footnote": [], + "bbox": [ + 555, + 419, + 663, + 465 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ec45eabe49713611f28bcee6f6de5a036e840e05f0fa0f6732172f1dafa03b89.jpg", + "image_caption": [ + "Reference" + ], + "image_footnote": [], + "bbox": [ + 666, + 419, + 774, + 465 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "results on the LOL [74] dataset. In Figure 11, we compare methods on the motion deblurring task using the GoPro [50] dataset. In Figures 9 and 12, we compare with different methods for the dehazing task on SOTS (outdoor) [35]. In Figure 13, we compare with image restoration methods for deraining on Rain100L [17]. Finally, we show denoising results in Figure 14. In this qualitative analysis, we use our single InstructIR-5D model to restore all the images. Limitations As with previous all-in-one methods, our model struggles to process images with more than one degradation (i.e. complex real-world images), or unknown out-of-distribution degradations, yet this is a common limitation among the related methods. However, we believe that these limitations can be surpassed with more realistic training data, and scaling the model's complexity.", + "bbox": [ + 212, + 532, + 787, + 703 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 717, + 357, + 733 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We present a novel approach that uses natural human-written instructions to guide the image restoration model. Given a prompt, our multi-task model can recover high-quality images from their degraded counterparts, considering multiple degradations. We achieve state-of-the-art results on several restoration tasks, demonstrating the power and flexibility of instruction guidance. Our results represent a new benchmark for text-guided image restoration.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "InstructIR", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8125aafc4839ebcb4c71b26079513104de1684649f67fede8bc456704dce09db.jpg", + "image_caption": [ + "Low-light Image" + ], + "image_footnote": [], + "bbox": [ + 225, + 152, + 401, + 234 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d9be2b4d2d1f65cda746e00c7d31508865d3810f8c91fa6eb6076d1dbec21560.jpg", + "image_caption": [ + "NAFNet" + ], + "image_footnote": [], + "bbox": [ + 405, + 152, + 477, + 189 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5c3d96742248411cc2e169d206adad6bd16ecddfe2f43b5cc09bf4f4d081eef3.jpg", + "image_caption": [ + "MPRNet" + ], + "image_footnote": [], + "bbox": [ + 478, + 152, + 550, + 189 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d00f684e6d24e1cc588bc98620fe75d5e3d1fa289a4a5597b1407d9a6f908aa3.jpg", + "image_caption": [ + "DGUNet" + ], + "image_footnote": [], + "bbox": [ + 552, + 152, + 622, + 189 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ba46adc333e4cd04c14997084a80ba0c318eb90ef4a644dd93d61ab1f3935bc7.jpg", + "image_caption": [ + "MIRNetV2" + ], + "image_footnote": [], + "bbox": [ + 625, + 152, + 697, + 189 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/031d26c4dd4bf444fb3111eb40b8cb489799984a83f947355212adbc9876a69e.jpg", + "image_caption": [ + "InstructIR" + ], + "image_footnote": [], + "bbox": [ + 697, + 152, + 771, + 189 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/46616ef4e8c83a1a172efef1e99d18b9e5ce30b9caa5d06a2f1153b0791fcdd7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 199, + 401, + 234 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f639bfa8c48a1d8e1a3fbbcf1f54596b496afe186343c5a078c9213f734181b5.jpg", + "image_caption": [ + "TAPE" + ], + "image_footnote": [], + "bbox": [ + 403, + 199, + 477, + 234 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/946712f5ded19400803b8e125945abd58504658f00abf2107c7346d7dda5b212.jpg", + "image_caption": [ + "Restormer" + ], + "image_footnote": [], + "bbox": [ + 478, + 199, + 550, + 234 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5031453b98702aaa860947e198148d8ba4978ee4a37e19c84468a193f34c7db7.jpg", + "image_caption": [ + "AirNet", + "IDR" + ], + "image_footnote": [], + "bbox": [ + 552, + 199, + 697, + 234 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "AirNet", + "bbox": [ + 575, + 234, + 601, + 243 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "", + "bbox": [ + 599, + 199, + 697, + 234 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/442cfc75e98c6466278325f29be28db9caac2fcd26670f4453a5d15d23366cfa.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 700, + 199, + 769, + 234 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/6cdcf2e2c02ae92bd12cdce4be545aab9f1002a983c45ca28cbfa4fc834871f2.jpg", + "image_caption": [ + "Blurry Image", + "Fig. 11: Image Deblurring Results. GoPro [50] dataset." + ], + "image_footnote": [], + "bbox": [ + 225, + 319, + 403, + 402 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/dc4d1d40b9fcbb5f946538adce8a4961a32568684a5ad744d08691a87950b4b9.jpg", + "image_caption": [ + "NAFNet" + ], + "image_footnote": [], + "bbox": [ + 405, + 320, + 477, + 358 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/818281906252ff084602697ddcc1f69d33cfeb657b5b34e1c936993b06906b61.jpg", + "image_caption": [ + "MPRNet" + ], + "image_footnote": [], + "bbox": [ + 478, + 320, + 550, + 358 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b28ff142c1596559d1312fae787d60132f830244de99cba50d27912864f75425.jpg", + "image_caption": [ + "DGUNet" + ], + "image_footnote": [], + "bbox": [ + 552, + 320, + 622, + 358 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/6ddaf809bf7118af0334a9a135487cf824692545ca2a3795eb2599018d5bf463.jpg", + "image_caption": [ + "MIRNetV2" + ], + "image_footnote": [], + "bbox": [ + 625, + 320, + 697, + 358 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9555f65f6a06e4d7532a7edf95901eecea10972e14fe852db4cddc4341590aac.jpg", + "image_caption": [ + "Fig. 10: Real Low-light Enhancement Results. LOL [74] testset (748.png). AirNet [36] and IDR [89] are well-known all-in-one restoration methods. NAFNet [7] is equivalent to InstructIR without text conditions (i.e. our image-only backbone).", + "InstructIR" + ], + "image_footnote": [], + "bbox": [ + 700, + 320, + 771, + 358 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "", + "bbox": [ + 225, + 364, + 403, + 402 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/18bca19ee81a0da59440aaa9c0f131d78f9f26838609b69cdfa21eb5684a1916.jpg", + "image_caption": [ + "TAPE" + ], + "image_footnote": [], + "bbox": [ + 403, + 364, + 477, + 402 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/24df408de42c7c6d838b38dd6a518938fbf8337d48101e1999a6420d6cdcc354.jpg", + "image_caption": [ + "Restormer" + ], + "image_footnote": [], + "bbox": [ + 478, + 364, + 550, + 402 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2784727bb426120f76a3bd782012d65c414effab0f4627b2e48824fb6004fc6d.jpg", + "image_caption": [ + "AirNet" + ], + "image_footnote": [], + "bbox": [ + 552, + 364, + 622, + 402 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b84b3ee66eea74cfdf05cb063e3944b2ad1060ebd9ec20084c4b1a470b0a6e8c.jpg", + "image_caption": [ + "IDR" + ], + "image_footnote": [], + "bbox": [ + 625, + 364, + 697, + 402 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/68a9005f7f24e806fe68cea6ebbf48e42f2a0e6aa4d5617c70bc0d40d4011449.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 700, + 364, + 771, + 402 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0e0be5640cdba7caa33ace574f795c3f509110dd7306bd8879b57b6d96131929.jpg", + "image_caption": [ + "Haze Image", + "TAPE", + "Fig. 12: Image Dehazing Results. SOTS [35] outdoor dataset (0150.jpg)." + ], + "image_footnote": [], + "bbox": [ + 225, + 453, + 403, + 536 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/41c39169c458ddfe91ce3411ba6aec81d3b2eb5477474c2395723a61a7f410ad.jpg", + "image_caption": [ + "NAFNet" + ], + "image_footnote": [], + "bbox": [ + 405, + 454, + 477, + 491 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9b3374cdce6fc4ce038fad4d6cc21f23a865a59b965c6eb10a200630e047791b.jpg", + "image_caption": [ + "MPRNet" + ], + "image_footnote": [], + "bbox": [ + 478, + 454, + 550, + 491 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a411721d191f5b171ee7d820db2e35e04ecaa9be30d9ac53e0c4fa74cb0a83b9.jpg", + "image_caption": [ + "DGUNet" + ], + "image_footnote": [], + "bbox": [ + 552, + 454, + 622, + 491 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/668e91bd950f4a412be7f115edb85eff794294872c104839ee66af610d3f3e3a.jpg", + "image_caption": [ + "MIRNetV2" + ], + "image_footnote": [], + "bbox": [ + 625, + 454, + 697, + 491 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/8846fd0590d30a846e9f4b8cbf3943f3c48be3e85cd79c958401cf49135a57b4.jpg", + "image_caption": [ + "InstructIR" + ], + "image_footnote": [], + "bbox": [ + 700, + 454, + 771, + 491 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "", + "bbox": [ + 225, + 500, + 403, + 536 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/eb25d9f6a35b2302ae0e7b01347306658c0f360aaf810c53ffa994063c3e6e62.jpg", + "image_caption": [ + "Restormer" + ], + "image_footnote": [], + "bbox": [ + 478, + 500, + 550, + 536 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9ddf39bc733b26f72261a9b09c42852b39cea34377683b17dd6131c6a3464f19.jpg", + "image_caption": [ + "AirNet" + ], + "image_footnote": [], + "bbox": [ + 552, + 500, + 622, + 536 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f227c6a65dc0d6e1e23721ed4446cd12fb45d21a2ab0f254b368b238217465b9.jpg", + "image_caption": [ + "IDF" + ], + "image_footnote": [], + "bbox": [ + 625, + 500, + 697, + 536 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/7b045258e36478771730b85a0f0cb35d41b66df6eadeea2a0b2fa5a8be4d995a.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 700, + 500, + 771, + 536 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9dd58982d68cba36318bbd95577bf45ebd505881e9e9d26081984ebd5e877b1a.jpg", + "image_caption": [ + "$\\therefore m = \\frac{3}{11}$ ;" + ], + "image_footnote": [], + "bbox": [ + 225, + 585, + 403, + 643 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5e708036c87c9a42ff105d210ef45578cfbc3df9654ac0d2de78605cebe12193.jpg", + "image_caption": [ + "NAFNet" + ], + "image_footnote": [], + "bbox": [ + 405, + 587, + 477, + 625 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3c8b6a87e36742171897855d274bbe3f36186351476c51be7bb317087b25e4e5.jpg", + "image_caption": [ + "MPRNet" + ], + "image_footnote": [], + "bbox": [ + 478, + 587, + 550, + 625 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/659c360949b5cdd10f7b01f5ecdf1d1ee8c834e51cbb050b43216e8a54c31d85.jpg", + "image_caption": [ + "DGUNet" + ], + "image_footnote": [], + "bbox": [ + 552, + 587, + 622, + 625 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/098b484679f2c6b41b81bd05bd50c5ef916bb4ea978023dfcfcb5570b2c024a7.jpg", + "image_caption": [ + "MIRNetV2" + ], + "image_footnote": [], + "bbox": [ + 625, + 587, + 697, + 625 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/7fe9ec5d9bc72bb886bdf65e0d75b2d48f10c877cd3e4b93facca106bc5468ef.jpg", + "image_caption": [ + "InstructIR" + ], + "image_footnote": [], + "bbox": [ + 700, + 587, + 771, + 625 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/60f5f6186f56dfd1afdd6183dba3ea62e7573afd7a373f7ab25fff5165f6d57a.jpg", + "image_caption": [ + "Rainy Image", + "Fig. 13: Image Deraining Results on Rain100L [17] (035.png)." + ], + "image_footnote": [], + "bbox": [ + 225, + 652, + 403, + 669 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/02c08d5e562af874a657c0ea006a78a3b903a4bfc55c06b888bcf9d43a55d91e.jpg", + "image_caption": [ + "TAPE" + ], + "image_footnote": [], + "bbox": [ + 403, + 652, + 477, + 669 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/27c5e77d6415849da93a7b1145cef8754b19a83814ee73537eac0ef4aefd0e75.jpg", + "image_caption": [ + "Restormer" + ], + "image_footnote": [], + "bbox": [ + 478, + 652, + 550, + 669 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0388fd22cc593cdc8e884516b5ce1f60e5a1883ab944d65ffa31f9304a0f109c.jpg", + "image_caption": [ + "AirNet" + ], + "image_footnote": [], + "bbox": [ + 550, + 652, + 622, + 669 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/454ba8c97256230e526d9ad8d7f0fb2b37361fa75e181c3668006c66cc8be812.jpg", + "image_caption": [ + "IDR" + ], + "image_footnote": [], + "bbox": [ + 625, + 652, + 697, + 669 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f9073ec7db09e4b08fb8174028b30d443515794d6ae6551784d87336206b6ffb.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 700, + 652, + 771, + 669 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c5b2ba6176dfe37e6be5f2c3cf06e6581d67aad0a1b41396d786339dace7444c.jpg", + "image_caption": [ + "Fig. 14: Image Denoising Results on BSD68 [45] (0060.png)." + ], + "image_footnote": [], + "bbox": [ + 225, + 719, + 403, + 803 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Noise Image", + "bbox": [ + 290, + 801, + 339, + 811 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/292b40798c91c11d14544e9d1d0ef2ba539294365767e3b35668d18bd14bbb62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 720, + 477, + 756 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0892a6e504b0f6094241e8703d7dbb15dc6eea22ee3943e4ac95a01838107508.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 720, + 550, + 756 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/23d85597e72292315b869bb39eb28cf3d9bddecf2d62596d2ced5d1bab8f4c7e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 552, + 720, + 622, + 756 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/23359ee7289ec06eb35212328358961fdd6a4507880254dd9f5d2b7ae8e10651.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 720, + 697, + 756 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e25e2fe1fc849443e17bec637366b61d62bdc1811250fdcad65af4220cb804d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 720, + 771, + 756 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "", + "bbox": [ + 225, + 763, + 403, + 801 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Noise Image", + "bbox": [ + 290, + 801, + 339, + 811 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e78430966735a45edb4a887cd0ec3405e0c167185bb9aeda3dcbff327f5a686a.jpg", + "image_caption": [ + "NAFNet", + "TAPE" + ], + "image_footnote": [], + "bbox": [ + 403, + 763, + 477, + 801 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4092a20cb176a696593465c27a6ca80f5977fd10859e6522f51f3fe33b992221.jpg", + "image_caption": [ + "MPRNet", + "Restormer" + ], + "image_footnote": [], + "bbox": [ + 478, + 763, + 550, + 801 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c164dae35b527dc63d1bc60de44a3aa0a6088218fc6d1fd81f45771fd287a0e4.jpg", + "image_caption": [ + "DGUNet", + "AirNet" + ], + "image_footnote": [], + "bbox": [ + 552, + 763, + 622, + 801 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2c658dde79342b0c7f889e82357ba98ef9f5fe05cdea5392fd48f4f389415436.jpg", + "image_caption": [ + "MIRNetV2", + "IDR" + ], + "image_footnote": [], + "bbox": [ + 625, + 763, + 697, + 801 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/8851ecd7763a336e7b162facddfa9332400f1ff5aa192253f535c53dd02ca242.jpg", + "image_caption": [ + "InstructIR", + "GT" + ], + "image_footnote": [], + "bbox": [ + 700, + 763, + 771, + 801 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Conde, Marcos et al.", + "bbox": [ + 271, + 114, + 410, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Agustsson, E., Timofte, R.: NTIRE 2017 challenge on single image superresolution: Dataset and study. In: CVPR Workshops (2017) 10", + "2. Arbelaez, P., Maire, M., Fowlkes, C., Malik, J.: Contour detection and hierarchical image segmentation. TPAMI (2011) 8", + "3. Bai, Y., Wang, C., Xie, S., Dong, C., Yuan, C., Wang, Z.: Textir: A simple framework for text-based editable image restoration. CoRR abs/2302.14736 (2023). https://doi.org/10.48550/ARXIV.2302.14736, https://doi.org/10.48550/arXiv.2302.14736 5", + "4. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023. pp. 18392-18402. IEEE (2023). https://doi.org/10.1109/CVPR52729.2023.01764, https://doi.org/10.1109/CVPR52729.2023.01764 2, 3, 4, 5, 12", + "5. Bychkovsky, V., Paris, S., Chan, E., Durand, F.: Learning photographic global tonal adjustment with a database of input / output image pairs. In: The Twenty-Fourth IEEE Conference on Computer Vision and Pattern Recognition (2011) 8, 10", + "6. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR (2021) 11", + "7. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV (2022) 2, 3, 6, 7, 9, 14", + "8. Chen, L., Lu, X., Zhang, J., Chu, X., Chen, C.: Hinet: Half instance normalization network for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 182-192 (2021) 9", + "9. Chen, Y.S., Wang, Y.C., Kao, M.H., Chuang, Y.Y.: Deep photo enhancer: Unpaired learning for image enhancement from photographs with gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6306-6314 (2018) 10", + "10. Devlin, J., Chang, M., Lee, K., Toutanova, K.: BERT: pre-training of deep bidirectional transformers for language understanding. In: Burstein, J., Doran, C., Solorio, T. (eds.) Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers). pp. 4171-4186. Association for Computational Linguistics (2019). https://doi.org/10.18653/V1/N19-1423, https://doi.org/10.18653/v1/n19-1423 7", + "11. Ding, C., Lu, Z., Wang, S., Cheng, R., Boddeti, V.N.: Mitigating task interference in multi-task learning via explicit task routing with non-learnable primitives. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7756-7765 (2023) 6, 7", + "12. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. TPAMI (2015) 10", + "13. Dong, H., Pan, J., Xiang, L., Hu, Z., Zhang, X., Wang, F., Yang, M.H.: Multi-scale boosted dehazing network with dense feature fusion. In: CVPR (2020) 2", + "14. Dong, W., Zhang, L., Shi, G., Wu, X.: Image deblurring and super-resolution by adaptive sparse domain selection and adaptive regularization. TIP (2011) 2", + "15. Dong, Y., Liu, Y., Zhang, H., Chen, S., Qiao, Y.: Fd-gan: Generative adversarial networks with fusion-discriminator for single image dehazing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 10729-10736 (2020) 9" + ], + "bbox": [ + 218, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "InstructIR", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Elad, M., Feuer, A.: Restoration of a single superresolution image from several blurred, noisy, and undersampled measured images. IEEE transactions on image processing 6(12), 1646-1658 (1997) 2", + "17. Fan, Q., Chen, D., Yuan, L., Hua, G., Yu, N., Chen, B.: A general decoupled learning framework for parameterized image operators. IEEE transactions on pattern analysis and machine intelligence 43(1), 33-47 (2019) 9, 13, 14", + "18. Fu, X., Zeng, D., Huang, Y., Liao, Y., Ding, X., Paisley, J.: A fusion-based enhancing method for weakly illuminated images 129, 82-96 (2016) 11", + "19. Fu, X., Zeng, D., Huang, Y., Zhang, X.P., Ding, X.: A weighted variational model for simultaneous reflectance and illumination estimation. In: CVPR (2016) 11", + "20. Gao, H., Tao, X., Shen, X., Jia, J.: Dynamic scene deblurring with parameter selective sharing and nested skip connections. In: CVPR. pp. 3848-3856 (2019) 9", + "21. Gharbi, M., Chen, J., Barron, J.T., Hasinoff, S.W., Durand, F.: Deep bilateral learning for real-time image enhancement. ACM Transactions on Graphics (TOG) 36(4), 1-12 (2017) 10", + "22. Guo, X., Li, Y., Ling, H.: Lime: Low-light image enhancement via illumination map estimation. IEEE TIP 26(2), 982-993 (2016) 11", + "23. Hao, S., Han, X., Guo, Y., Xu, X., Wang, M.: Low-light image enhancement with semi-decoupled decomposition. IEEE TMM 22(12), 3025-3038 (2020) 11", + "24. He, J., Dong, C., Qiao, Y.: Modulating image restoration with continual levels via adaptive feature modification layers (2019), https://arxiv.org/abs/1904.081186", + "25. He, K., Sun, J., Tang, X.: Single image haze removal using dark channel prior. TPAMI (2010) 2", + "26. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022) 3", + "27. Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al.: Searching for mobilenetv3. In: ICCV (2019) 3", + "28. Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5197-5206 (2015) 8", + "29. Jiang, Y., Gong, X., Liu, D., Cheng, Y., Fang, C., Shen, X., Yang, J., Zhou, P., Wang, Z.: Enlightengan: Deep light enhancement without paired supervision. IEEE TIP 30, 2340-249 (2021) 11", + "30. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023) 3", + "31. Kim, K.I., Kwon, Y.: Single-image super-resolution using sparse regression and natural image prior. TPAMI (2010) 2", + "32. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv:1412.6980 (2014) 8", + "33. Kopf, J., Neubert, B., Chen, B., Cohen, M., Cohen-Or, D., Deussen, O., Uyttendaele, M., Lischinski, D.: Deep photo: Model-based photograph enhancement and viewing. ACM TOG (2008) 2", + "34. Lei, X., Fei, Z., Zhou, W., Zhou, H., Fei, M.: Low-light image enhancement using the cell vibration model. IEEE TMM pp. 1-1 (2022) 11", + "35. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing 28(1), 492-505 (2018) 8, 9, 13, 14" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Conde, Marcos et al.", + "bbox": [ + 271, + 114, + 410, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "36. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR. pp. 17452-17462 (June 2022) 2, 3, 7, 8, 9, 13, 14", + "37. Li, J., Li, J., Fang, F., Li, F., Zhang, G.: Luminance-aware pyramid network for low-light image enhancement. IEEE TMM 23, 3153-3165 (2020) 11", + "38. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: SwinIR: Image restoration using swin transformer. In: ICCV Workshops (2021) 2, 9, 12", + "39. Liu, L., Xie, L., Zhang, X., Yuan, S., Chen, X., Zhou, W., Li, H., Tian, Q.: Tape: Task-agnostic prior embedding for image restoration. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XVIII. pp. 447–464. Springer (2022) 9", + "40. Liu, R., Ma, L., Zhang, J., Fan, X., Luo, Z.: Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In: CVPR (2021) 11", + "41. Liu, Y., Liu, A., Gu, J., Zhang, Z., Wu, W., Qiao, Y., Dong, C.: Discovering distinctive \"semantics\" in super-resolution networks (2022), https://arxiv.org/abs/2108.00406 3", + "42. Ma, J., Cheng, T., Wang, G., Zhang, Q., Wang, X., Zhang, L.: Prores: Exploring degradation-aware visual prompt for universal image restoration. arXiv preprint arXiv:2306.13653 (2023) 2, 3", + "43. Ma, K., Duanmu, Z., Wu, Q., Wang, Z., Yong, H., Li, H., Zhang, L.: Waterloo exploration database: New challenges for image quality assessment models. TIP (2016) 8", + "44. Ma, L., Ma, T., Liu, R., Fan, X., Luo, Z.: Toward fast, flexible, and robust low-light image enhancement. In: CVPR (2022) 11", + "45. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: ICCV (2001) 8, 9, 14", + "46. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073 (2021) 3", + "47. Michaeli, T., Irani, M.: Nonparametric blind super-resolution. In: ICCV (2013) 2", + "48. Moran, S., Marza, P., McDonagh, S., Parisot, S., Slabaugh, G.: Deeplpf: Deep local parametric filters for image enhancement. In: CVPR (2020) 11", + "49. Mou, C., Wang, Q., Zhang, J.: Deep generalized unfolding networks for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17399-17410 (2022) 9", + "50. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: CVPR (2017) 8, 9, 13, 14", + "51. Nah, S., Son, S., Lee, J., Lee, K.M.: Clean images are hard to reblur: Exploiting the ill-posed inverse task for dynamic scene deblurring. In: ICLR (2022) 2", + "52. Nguyen, N., Milanfar, P., Golub, G.: Efficient generalized cross-validation with applications to parametric image restoration and resolution enhancement. IEEE Transactions on image processing 10(9), 1299-1308 (2001) 2", + "53. Park, D., Lee, B.H., Chun, S.Y.: All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5815-5824. IEEE (2023) 2, 3, 7", + "54. Potlapalli, V., Zamir, S.W., Khan, S., Khan, F.S.: Prompt: Prompting for all-in-one blind image restoration. arXiv preprint arXiv:2306.13090 (2023) 2, 3, 8, 9, 13" + ], + "bbox": [ + 212, + 146, + 787, + 837 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "InstructIR", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "55. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event. Proceedings of Machine Learning Research, vol. 139, pp. 8748-8763. PMLR (2021), http://proceedings.mlr.press/v139/radford21a.html 5", + "56. Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert-networks. In: Inui, K., Jiang, J., Ng, V., Wan, X. (eds.) Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019. pp. 3980-3990. Association for Computational Linguistics (2019). https://doi.org/10.18653/V1/D19-1410, https://doi.org/10.18653/v1/D19-1410 5", + "57. Ren, C., He, X., Wang, C., Zhao, Z.: Adaptive consistency prior based deep network for image denoising. In: CVPR (2021) 2", + "58. Ren, W., Pan, J., Zhang, H., Cao, X., Yang, M.H.: Single image dehazing via multi-scale convolutional neural networks with holistic edges. IJCV (2020) 2", + "59. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 10674-10685. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.01042, https://doi.org/10.1109/CVPR52688.2022.01042 5", + "60. Ronneberger, O., Fischer, P., Brox, T.: U-Net: convolutional networks for biomedical image segmentation. In: MICCAI (2015) 6", + "61. Rosenbaum, C., Klinger, T., Riemer, M.: Routing networks: Adaptive selection of non-linear functions for multi-task learning. arXiv preprint arXiv:1711.01239 (2017) 6", + "62. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023) 3", + "63. Strezoski, G., Noord, N.v., Worring, M.: Many task learning with task routing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1375-1384 (2019) 6, 7, 8", + "64. Tian, C., Xu, Y., Zuo, W.: Image denoising using deep cnn with batch renormalization. Neural Networks (2020) 9", + "65. Timofte, R., De Smet, V., Van Gool, L.: Anchored neighborhood regression for fast example-based super-resolution. In: ICCV (2013) 2", + "66. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A., Li, Y.: MAXIM: Multi-axis MLP for image processing. In: CVPR. pp. 5769-5780 (2022) 2, 8", + "67. Valanarasu, J.M.J., Yasarla, R., Patel, V.M.: Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In: CVPR. pp. 2353-2363 (2022) 3, 9", + "68. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 2", + "69. Wang, R., Zhang, Q., Fu, C.W., Shen, X., Zheng, W.S., Jia, J.: Underexposed photo enhancement using deep illumination estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6849-6857 (2019) 10" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Conde, Marcos et al.", + "bbox": [ + 271, + 114, + 410, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "70. Wang, S., Zheng, J., Hu, H.M., Li, B.: Naturalness preserved enhancement algorithm for non-uniform illumination images. IEEE TIP 22(9), 3538-3548 (2013) 11", + "71. Wang, X., Girshick, R., Gupta, A., He, K.: Non-local neural networks. In: CVPR (2018) 3", + "72. Wang, Y., Liu, Z., Liu, J., Xu, S., Liu, S.: Low-light image enhancement with illumination-aware gamma correction and complete image modelling network. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13128-13137 (2023) 11", + "73. Wang, Z., Cun, X., Bao, J., Liu, J.: Uformer: A general u-shaped transformer for image restoration. arXiv:2106.03106 (2021) 2, 11", + "74. Wei, C., Wang, W., Yang, W., Liu, J.: Deep retina decomposition for low-light enhancement. In: British Machine Vision Conference (2018) 8, 9, 11, 13, 14", + "75. Wu, W., Weng, J., Zhang, P., Wang, X., Yang, W., Jiang, J.: Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement. In: CVPR (2022) 11", + "76. Xiao, S., Liu, Z., Zhang, P., Muennighof, N.: C-pack: Packaged resources to advance general chinese embedding. CoRR abs/2309.07597 (2023). https://doi.org/10.48550/ARXIV.2309.07597, https://doi.org/10.48550/arXiv.2309.07597", + "77. Xu, K., Yang, X., Yin, B., Lau, R.W.: Learning to restore low-light images via decomposition-and-enhancement. In: CVPR (2020) 11", + "78. Yang, F., Yang, H., Fu, J., Lu, H., Guo, B.: Learning texture transformer network for image super-resolution. In: CVPR (2020) 8, 9", + "79. Yang, W., Wang, S., Fang, Y., Wang, Y., Liu, J.: Band representation-based semi-supervised low-light image enhancement: bridging the gap between signal fidelity and perceptual quality. IEEE TIP 30, 3461-3473 (2021) 11", + "80. Yang, W., Wang, W., Huang, H., Wang, S., Liu, J.: Sparse gradient regularized deep retina network for robust low-light image enhancement. IEEE TIP 30, 2072-2086 (2021) 11", + "81. Yao, M., Xu, R., Guan, Y., Huang, J., Xiong, Z.: Neural degradation representation learning for all-in-one image restoration. arXiv preprint arXiv:2310.12848 (2023) 3", + "82. Yu, W., Luo, M., Zhou, P., Si, C., Zhou, Y., Wang, X., Feng, J., Yan, S.: Metaformer is actually what you need for vision. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10819-10829 (2022) 3", + "83. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022) 2, 3, 9", + "84. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Learning enriched features for real image restoration and enhancement. In: ECCV (2020) 9, 11", + "85. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR (2021) 9", + "86. Zeng, H., Cai, J., Li, L., Cao, Z., Zhang, L.: Learning image-adaptive 3d lookup tables for high performance photo enhancement in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(4), 2058-2073 (2020) 10", + "87. Zhang, C., Zhu, Y., Yan, Q., Sun, J., Zhang, Y.: All-in-one multi-degradation image restoration network via hierarchical degradation representation. arXiv preprint arXiv:2308.03021 (2023) 3" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "InstructIR", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "88. Zhang, C., Zhu, Y., Yan, Q., Sun, J., Zhang, Y.: All-in-one multi-degradation image restoration network via hierarchical degradation representation. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 2285-2293 (2023) 3", + "89. Zhang, J., Huang, J., Yao, M., Yang, Z., Yu, H., Zhou, M., Zhao, F.: Ingredient-oriented multi-degradation learning for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5825-5835 (2023) 2, 3, 8, 9, 14", + "90. Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing 26(7), 3142-3155 (2017) 2", + "91. Zhang, K., Zuo, W., Gu, S., Zhang, L.: Learning deep CNN denoiser prior for image restoration. In: CVPR (2017) 2", + "92. Zhang, K., Luo, W., Zhong, Y., Ma, L., Stenger, B., Liu, W., Li, H.: Deblurring by realistic blurring. In: CVPR. pp. 2737-2746 (2020) 2", + "93. Zhang, Y., Zhang, J., Guo, X.: Kindling the darkness: A practical low-light image enhancer. In: ACM MM (2019) 11" + ], + "bbox": [ + 215, + 146, + 787, + 369 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "Conde, Marcos et al.", + "bbox": [ + 271, + 114, + 410, + 127 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/InstructIR_ High-Quality Image Restoration Following Human Instructions/86f2cef3-3642-43ee-930e-b6ddf921d714_model.json b/2024/InstructIR_ High-Quality Image Restoration Following Human Instructions/86f2cef3-3642-43ee-930e-b6ddf921d714_model.json new file mode 100644 index 0000000000000000000000000000000000000000..89e1e8bf1d26afdf48d45a12c2b15b480137a2c9 --- /dev/null +++ b/2024/InstructIR_ High-Quality Image Restoration Following Human Instructions/86f2cef3-3642-43ee-930e-b6ddf921d714_model.json @@ -0,0 +1,4970 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.251, + 0.141, + 0.753, + 0.187 + ], + "angle": 0, + "content": "InstructIR: High-Quality Image Restoration Following Human Instructions" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.212, + 0.718, + 0.228 + ], + "angle": 0, + "content": "Marcos V. Conde1,2, Gregor Geigle1, and Radu Timofte1" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.239, + 0.72, + 0.254 + ], + "angle": 0, + "content": "1 Computer Vision Lab, CAIDAS & IFI, University of Würzburg" + }, + { + "type": "text", + "bbox": [ + 0.416, + 0.254, + 0.588, + 0.268 + ], + "angle": 0, + "content": "2 Sony PlayStation, FTG" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.269, + 0.642, + 0.282 + ], + "angle": 0, + "content": "https://github.com/mv-lab/InstructIR" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.304, + 0.353, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.388, + 0.339, + 0.406 + ], + "angle": 0, + "content": "Can you stabilize this pie? I was running..." + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.304, + 0.468, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.388, + 0.461, + 0.406 + ], + "angle": 0, + "content": "My image has tiny dots, doesn't look great, fix it" + }, + { + "type": "image", + "bbox": [ + 0.47, + 0.304, + 0.574, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.476, + 0.388, + 0.566, + 0.406 + ], + "angle": 0, + "content": "My photo is too dark, I cannot see anything" + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.304, + 0.673, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.388, + 0.673, + 0.406 + ], + "angle": 0, + "content": "Retouch my picture like a photographer" + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.304, + 0.785, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.388, + 0.782, + 0.406 + ], + "angle": 0, + "content": "It is a foggy day, can you make it clearer?" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.407, + 0.353, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.407, + 0.469, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.47, + 0.407, + 0.574, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.407, + 0.673, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.407, + 0.785, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.552 + ], + "angle": 0, + "content": "Fig. 1: Given an image and a prompt for how to improve that image, our all-in-one restoration model corrects the image considering the human instruction. InstructIR, can tackle various types and levels of degradation, and it is able to generalize in some real-world scenarios (last three images, from left to right)." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.569, + 0.741, + 0.764 + ], + "angle": 0, + "content": "Abstract. Image restoration is a fundamental problem that involves recovering a high-quality clean image from its degraded observation. All-In-One image restoration models can effectively restore images from various types and levels of degradation using degradation-specific information as prompts to guide the restoration model. In this work, we present the first approach that uses human-written instructions to guide the image restoration model. Given natural language prompts, our model can recover high-quality images from their degraded counterparts, considering multiple degradation types. Our method, InstructIR, achieves state-of-the-art results on several restoration tasks including image denoising, deraining, deblurring, dehazing, and (low-light) image enhancement. InstructIR improves \\(+1\\mathrm{dB}\\) over previous all-in-one restoration methods. Moreover, our dataset and results represent a novel benchmark for new research on text-guided image restoration and enhancement." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.784, + 0.377, + 0.799 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Images often contain unpleasant effects such as noise, motion blur, haze, and low dynamic range. Such effects are commonly known in low-level computer" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.411, + 0.128 + ], + "angle": 0, + "content": "Conde, Marcos et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "vision as degradations. These can result from camera limitations or challenging environmental conditions e.g. low light." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.784, + 0.221 + ], + "angle": 0, + "content": "Image restoration aims to recover a high-quality image from its degraded counterpart. This is a complex inverse problem since multiple different solutions can exist for restoring any given image [13, 16, 38, 52, 90, 91]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.403 + ], + "angle": 0, + "content": "Some methods focus on specific degradations, for instance reducing noise (denoising) [57, 90, 91], removing blur (deblurring) [51, 92], or clearing haze (dehazing) [13, 58]. Such methods are effective for their specific task, yet they do not generalize well to other types of degradation. Other approaches use a general neural network for diverse tasks [7, 66, 73, 83], yet training the neural network for each specific task independently. Since using a separate model for each possible degradation is resource-intensive, recent approaches propose All-in-One restoration models [36, 53, 54, 89]. These approaches use a single deep blind restoration model considering multiple degradation types and levels. Contemporary works such as PromptIR [54] or ProRes [42] utilize a unified model for blind image restoration using learned guidance vectors, also known as \"prompt embeddings\", in contrast to raw user prompts in text form, which we use in this work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.787, + 0.524 + ], + "angle": 0, + "content": "In parallel, recent works such as InstructPix2Pix [4] show the potential of using text prompts to guide image generation and editing models. However, this method (or recent alternatives) do not tackle inverse problems. Inspired by these works, we argue that text guidance can help to guide blind restoration models better than the image-based degradation classification used in previous works [36,53,89]. Users generally have an idea about what has to be fixed (though they might lack domain-specific vocabulary) so we can use this information to guide the model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.788, + 0.668 + ], + "angle": 0, + "content": "Contributions We propose the first approach that utilizes real human-written instructions to solve multi-task image restoration. Our comprehensive experiments demonstrate the potential of using text guidance for image restoration and enhancement by achieving state-of-the-art performance on various image restoration tasks, including image denoising, deraining, deblurring, dehazing, and low-light image enhancement. Our model, InstructIR, is able to generalize to restoring images using complex human-written instructions. Moreover, our single all-in-one model covers more tasks than many previous works. We show diverse restoration samples of our method in Figure 1." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.689, + 0.388, + 0.705 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Image Restoration. Recent deep learning methods [13, 38, 51, 57, 66, 83] have shown consistently better results compared to traditional techniques for blind image restoration [14,25,31,33,47,65]. The proposed neural networks are based on convolutional neural networks (CNNs) and Transformers [68] (or related attention mechanisms). We focus on general-purpose restoration models [7,38,73,83]. For example, SwinIR [38], MAXIM [66] and Uformer [73]. These models can be trained -independently- for diverse tasks such as denoising, deraining or deblurring. Their ability to capture local and global feature interactions, and enhance" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InstructIR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "them, allows the models to achieve great performance consistently across different tasks. For instance, Restormer [83] uses non-local blocks [71] to capture complex features across the image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.284 + ], + "angle": 0, + "content": "NAFNet [7] is an efficient alternative to complex transformer-based methods. The model uses simplified channel attention, and gating as an alternative to non-linear activations. The building block (NAFBlock) follows a simple metaformer [82] architecture with efficient inverted residual blocks [27]. In this work, we build our InstructIR model using NAFNet as backbone, due to its efficient and simple design, and high performance in several restoration tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.294, + 0.785, + 0.37 + ], + "angle": 0, + "content": "All-in-One Image Restoration. Single degradation (or single task) restoration methods are well-studied, however, their real-world applications are limited due to the required resources i.e. allocating different models, and selecting the adequate model on demand. Moreover, images rarely present a single degradation, for instance, noise and blur are almost ubiquitous in any image capture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.785, + 0.431 + ], + "angle": 0, + "content": "All-in-One (also known as multi-degradation or multi-task) image restoration is emerging as a new research field in low-level computer vision [36,42,53,54,67, 81,87,88]. These approaches use a single deep blind restoration model to tackle different degradation types and levels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.785, + 0.523 + ], + "angle": 0, + "content": "We use as reference AirNet [36], IDR [89] and ADMS [53]. We also consider the contemporary work PromptIR [54]. The methods use different techniques to guide the blind model in the restoration process. For instance, an auxiliary model for degradation classification [36, 53], or multi-dimensional guidance vectors (also known as \"prompts\") [41, 42, 54] that help the model to discriminate the different types of degradation in the image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.785, + 0.669 + ], + "angle": 0, + "content": "Text-guided Image Manipulation. In recent years, multiple methods have been proposed for text-to-image generation and text-based image editing works [4, 26, 30, 46, 62]. These models use text prompts to describe images or actions, and powerful diffusion-based models for generating the corresponding images. Our main reference is InstructPix2Pix [4], this method enables editing from instructions that tell the model what action to perform, as opposed to text labels, captions or descriptions of the input or output images. Therefore, the user can transmit what to do in natural written text, without requiring to provide further image descriptions or sample reference images." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.697, + 0.651, + 0.715 + ], + "angle": 0, + "content": "3 Image Restoration Following Instructions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.785, + 0.84 + ], + "angle": 0, + "content": "We treat instruction-based image restoration as a supervised learning problem similar to previous works [4]. First, we generate over 10000 prompts using GPT-4 based on our own sample instructions. We explain the creation of the prompt dataset in Sec. 3.1. We then build a large paired training dataset of prompts and degraded/clean images. Finally, we train our InstructIR model, and we evaluate it on a wide variety of instructions including real human-written prompts. We explain our text encoder in Sec 3.2, and our complete model in Sec. 3.3." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.411, + 0.128 + ], + "angle": 0, + "content": "Conde, Marcos et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.536, + 0.163 + ], + "angle": 0, + "content": "3.1 Generating Prompts for Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.28 + ], + "angle": 0, + "content": "Why instructions? Inspired by InstructPix2Pix [4], we adopt human written instructions as the mechanism of control for our model. There is no need for the user to provide additional information, such as example clean images, or descriptions of the visual content. Instructions offer a clear and expressive way to interact, enabling users to pinpoint the unpleasant effects (degradations) in the images. We also consider the language complexity, from ambiguous instructions (e.g. \"fix my image\") to precise instructions (e.g. \"remove the noise\")." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.286, + 0.489, + 0.633 + ], + "angle": 0, + "content": "Handling free-form user prompts rather than fixed degradation-specific prompts increases the usability of our model for laypeople who lack domain expertise. We thus want our model to be capable of understanding diverse prompts posed by users \"in-the-wild\" e.g. kids, adults, or photographers. To this end, we use a large language model (i.e., GPT-4) to create diverse requests that might be asked by users for the different degradations types. We then filter those generated prompts to remove ambiguous or unclear prompts (e.g., \"Make the image cleaner\", \"improve this image\"). Our final instructions set contains over 10000 different prompts in total, for 7 different tasks. We display some examples in Table 1. As we show in Figure 2 the prompts are sampled randomly depending on the input degradation." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.298, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Table 1: Examples of our curated GPT4-generated and real user prompts with varying language and domain expertise." + }, + { + "type": "table", + "bbox": [ + 0.498, + 0.342, + 0.787, + 0.622 + ], + "angle": 0, + "content": "
Degradation Prompts
DenoisingCan you clean the dots from my image? Fix the grainy parts of this photo Remove the noise from my picture
DeblurringCan you reduce the movement in the image? My picture's not sharp, fix it Deblur my picture, it's too fuzzy
DehazingCan you make this picture clearer? Help, my picture is all cloudy Remove the fog from my photo
DerainingI want my photo to be clear, not rainy Clear the rain from my picture Remove the raindrops from my photo
Super-Res.Make my photo bigger and better Add details to this image Increase the resolution of this photo
Low-lightThe photo is too dark, improve exposure Increase the illumination in this shot My shot has very low dynamic range
EnhancementMake it pop! Adjust the color balance for a natural look Apply a cinematic color grade to the photo
GeneralFix my image please make the image look better
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.66, + 0.789, + 0.774 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.784, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Fig. 2: We train our blind image restoration models using common image datasets, and prompts generated using GPT-4, note that this is (self-)supervised learning. At inference time, our model generalizes to human-written instructions and restores (or enhances) the images." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InstructIR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.375, + 0.16 + ], + "angle": 0, + "content": "3.2 Text Encoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.357 + ], + "angle": 0, + "content": "The Choice of the Text Encoder. A text encoder maps the user prompt to a fixed-size vector representation (a text embedding). The related methods for text-based image generation [59] and manipulation [3,4] often use the text encoder of a CLIP model [55] to encode user prompts as CLIP excels in visual prompts. However, user prompts for degradation contain, in general, little to no visual content (e.g. the use describes the degradation, not the image itself). We opt, instead, to use a pure text-based sentence encoder [56], that is, a smaller model trained to encode sentences in a semantically meaningful embedding space. Sentence encoders - pre-trained with millions of examples - are compact and fast in comparison to CLIP, while being able to encode the semantics of diverse user prompts. For instance, we use the BGE-micro-v2 sentence transformer. We compare text encoders in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.361, + 0.788, + 0.437 + ], + "angle": 0, + "content": "Fine-tuning the Text Encoder. We want to adapt the text encoder E for the restoration task to better encode the required information for the restoration model. Training the full text encoder is likely to lead to overfitting on our small training set and lead to loss of generalization. Instead, we freeze the text encoder and train a projection head on top:" + }, + { + "type": "equation", + "bbox": [ + 0.43, + 0.448, + 0.785, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\mathbf {e} = \\operatorname {n o r m} (\\mathbf {W} \\cdot \\mathrm {E} (t)) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.788, + 0.518 + ], + "angle": 0, + "content": "where \\( t \\) is the text, \\( \\mathrm{E}(t) \\) represents the raw text embedding, \\( \\mathbf{W} \\in \\mathbb{R}^{d_t \\times d_v} \\) is a learned projection from the text dimension \\( (d_t) \\) to the input dimension for the restoration model \\( (d_v) \\), and norm is the 12-norm." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.788, + 0.685 + ], + "angle": 0, + "content": "Figure 3 shows that while the text encoder is capable out-of-the-box to cluster instructions to some extent (Figure 3a), our trained projection yields greatly improved clusters (Figure 3b). We distinguish clearly the clusters for deraining, denoising, dehazing, deblurring, and low-light image enhancement. The instructions for such tasks or degradations are very characteristic. Furthermore, we can appreciate that \"super-res\" and \"enhancement\" tasks are quite spread and between the previous ones, which matches the language logic. For instance \"add details to this image\" could be used for enhancement, deblurring, or denoising. In our experiments, \\( d_{t} = 384 \\), \\( d_{v} = 256 \\) and \\( \\mathbf{W} \\) is a linear layer. The representation \\( \\mathbf{e} \\) from the text encoder is shared across the blocks, and each block has a trainable projection \\( \\mathbf{W} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Intent Classification Loss. We propose a guidance loss on the text embedding \\(\\mathbf{e}\\) to improve training and interpretability. Using the degradation types as targets, we train a simple classification head \\(\\mathcal{C}\\) such that \\(\\mathbf{c} = \\mathcal{C}(\\mathbf{e})\\), where \\(\\mathbf{c} \\in \\mathbb{R}^{D}\\), being \\(D\\) is the number of degradation classes. The classification head \\(\\mathcal{C}\\) is a simple two-layers MLP. Thus, we only need to train a projection layer \\(\\mathbf{W}\\) and a simple MLP to capture the natural language knowledge. This allows the text model to learn meaningful embeddings as we can appreciate in Figure 3, not just guidance vectors for the main image processing model. We find that the model is able to classify accurately (i.e. over \\(95\\%\\) accuracy) the underlying degradation in the user's prompt after a few epochs." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.411, + 0.128 + ], + "angle": 0, + "content": "Conde, Marcos et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.484, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.233, + 0.486, + 0.255 + ], + "angle": 0, + "content": "(a) t-SNE of embeddings before training i.e. frozen text encoder" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.147, + 0.782, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.233, + 0.785, + 0.255 + ], + "angle": 0, + "content": "(b) t-SNE of embeddings after training our learned projection" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.267, + 0.784, + 0.294 + ], + "angle": 0, + "content": "Fig. 3: We show t-SNE plots of the text embeddings before/after training InstructIR. Each dot represents a human instruction." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.314, + 0.351, + 0.328 + ], + "angle": 0, + "content": "3.3 InstructIR" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.784, + 0.415 + ], + "angle": 0, + "content": "Our method InstructIR consists of an image model and a text encoder. We introduced our text encoder in Sec. 3.2. We use NAFNet [7] as the image model, an efficient image restoration model that follows a U-Net architecture [60]. To successfully learn multiple tasks using a single model, we use task routing techniques. Our framework for training and evaluation is illustrated in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.492, + 0.64 + ], + "angle": 0, + "content": "Text Guidance. The key aspect of InstructIR is the integration of the encoded instruction as a mechanism of control for the image model. Inspired in task routing for many-task learning [11, 61, 63], we propose an \"Instruction Condition Block\" (ICB) to enable task-specific transformations within the model. Conventional task routing [63] applies task-specific binary masks to the channel features. Since our model does not know a priori the degradation, we cannot use this technique directly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.493, + 0.686 + ], + "angle": 0, + "content": "Considering the image features \\(\\mathcal{F}\\) and the encoded instruction \\(\\mathbf{e}\\), we apply task routing as follows:" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.434, + 0.779, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.543, + 0.788, + 0.642 + ], + "angle": 0, + "content": "Fig. 4: Instruction Condition Block (ICB) using an approximation of task routing [63] for many-tasks learning (See Eq. 2). This mechanism allows the neural network to select and prioritize specific features depending on the instruction, similarly to a Mixture of Experts (MoE)." + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.697, + 0.786, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {c} ^ {\\prime} = \\operatorname {B l o c k} \\left(\\mathcal {F} _ {c} \\odot \\mathbf {m} _ {c}\\right) + \\mathcal {F} _ {c} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "where the mask \\(\\mathbf{m}_c = \\sigma (\\mathbf{W_c}\\cdot \\mathbf{e})\\) is produced using a linear layer \\(\\mathbf{W}_{\\mathbf{c}}\\) activated using the Sigmoid function - to produce a set of weights depending on the text embedding \\(\\mathbf{e}\\). Thus, we obtain a \\(c\\)-dimensional per-channel (soft-)binary mask \\(\\mathbf{m}_c\\). As [24, 63], task routing is applied as the channel-wise multiplication \\(\\odot\\) for masking features depending on the task. The conditioned features are further enhanced using a convolutional NAFBlock [7] (Block). We illustrate our task-routing ICB block in Figure 4. We use \"regular\" NAFBlocks [7], followed by ICBs to condition the features, at both encoder and decoder blocks. The" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InstructIR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.253 + ], + "angle": 0, + "content": "formulation is \\( F^{l + 1} = \\mathrm{ICB}(\\mathrm{Block}(F^l)) \\) where \\( l \\) is the layer. Although we do not condition explicitly the filters of the neural network, as in [63], the mask allows the model to select the most relevant channels depending on the image information and the instruction. Note that this formulation enables differentiable feature masking, and certain interpretability i.e. the features with high weights contribute the most to the restoration process. Indirectly, this also enforces to learn diverse filters and reduce sparsity [11, 63]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.262, + 0.785, + 0.339 + ], + "angle": 0, + "content": "Is InstructIR a blind restoration model? The model does not use explicit information about the degradation in the image e.g. noise profiles, blur kernels, or PSFs. Since our model infers the task (degradation) given the image and the instruction, we consider InstructIR a blind image restoration model. Similarly to previous works that use auxiliary image-based degradation classification [36,53]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.363, + 0.461, + 0.381 + ], + "angle": 0, + "content": "4 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.784, + 0.503 + ], + "angle": 0, + "content": "We evaluate our model on 9 well-known benchmarks for different image restoration tasks: image denoising, deblurring, deraining, dehazing, real low-light enhancement, and photo-realistic image enhancement. We present extensive quantitative results in Table 2 and Table 3. We provide extensive comparisons with other all-in-one methods as well as task-specific methods. Our single model successfully restores images considering different degradation types and levels. We provide additional results and ablation studies in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.527, + 0.461, + 0.542 + ], + "angle": 0, + "content": "4.1 Implementation Details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.784, + 0.585 + ], + "angle": 0, + "content": "Our InstructIR model is end-to-end trainable. The image model does not require pre-training but we use a pre-trained sentence encoder as language model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.785, + 0.745 + ], + "angle": 0, + "content": "Text Encoder. As we discussed in Sec. 3.2, we only need to train the text embedding projection and classification head (\\(\\approx\\)100K parameters). We initialize the text encoder with BGE-MICRO-v2\\(^{3}\\), a distilled version of BGE-SMALL-EN [76]. The BGE encoders are BERT-like encoders [10] pre-trained on large amounts of supervised and unsupervised data for general-purpose sentence encoding. The BGE-micro model is a 3-layer encoder with 17.3 million parameters, which we freeze during training. We also explore ALL-MINILM-L6-v2 and CLIP encoders, however, we concluded that small models prevent overfitting and provide the best performance while being fast. We provide the ablation study comparing the three text encoders in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.755, + 0.784, + 0.815 + ], + "angle": 0, + "content": "Image Model. We use NAFNet [7] as the image model backbone. The architecture consists of a 4-level encoder-decoder, with varying numbers of blocks at each level, specifically [2, 2, 4, 8] for the encoder, and [2, 2, 2, 2] for the decoder, from the level-1 to level-4 respectively. Between the encoder and decoder we use" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.577, + 0.84 + ], + "angle": 0, + "content": "3 https://huggingface.co/TaylorAI/bge-micro-v2" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.411, + 0.128 + ], + "angle": 0, + "content": "Conde, Marcos et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.786, + 0.192 + ], + "angle": 0, + "content": "4 middle blocks to enhance further the features. The decoder implements addition instead of concatenation for the skip connections. We use the Instruction Condition Block (ICB) for task-routing [63] only in the encoder and decoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.404 + ], + "angle": 0, + "content": "The model is optimized using the \\(\\mathcal{L}_1\\) loss between the ground-truth clean image and the restored one. Additionally, we use the cross-entropy loss \\(\\mathcal{L}_{ce}\\) for the intent classification head of the text encoder. We train using a batch size of 32 and AdamW [32] optimizer with learning rate \\(5e^{-4}\\) for 500 epochs (approximately 1 day using a single NVIDIA A100). We also use cosine annealing learning rate decay. During training, we utilize cropped patches of size \\(256 \\times 256\\) as input, and we use random horizontal and vertical flips as augmentations. Since our model uses as input instruction-image pairs, given an image, and knowing its degradation, we randomly sample instructions from our prompt dataset (\\(>10\\mathrm{K}\\) samples). Our image model has only 16M parameters, and the learned text projection is just 100k parameters (the language model is 17M parameters), thus, our model can be trained easily on standard GPUs such as NVIDIA RTX 2080Ti or 3090Ti in a couple of days. Furthermore, the inference process also fits in low-computation budgets (e.g. Google Colab T4 16Gb GPU)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.422, + 0.478, + 0.436 + ], + "angle": 0, + "content": "4.2 Datasets and Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.786, + 0.472 + ], + "angle": 0, + "content": "Following previous works [36,54,89], we prepare the datasets for different restoration tasks, including real and synthetic datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.786, + 0.545 + ], + "angle": 0, + "content": "Image denoising. We use a combination of BSD400 [2] and WED [43] datasets for training. This combined training set contains \\(\\approx 5000\\) images. Using as reference the clean images in the dataset, we generate the noisy images by adding Gaussian noise with different noise levels \\(\\sigma \\in \\{15, 25, 50\\}\\). We test the models on the well-known BSD68 [45] and Urban100 [28] datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.545, + 0.786, + 0.572 + ], + "angle": 0, + "content": "Image deraining. We use the Rain100L [78] dataset, which consists of 200 clean-rainy image pairs for training, and 100 pairs for testing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.786, + 0.645 + ], + "angle": 0, + "content": "Image dehazing. We utilize the Reside (outdoor) SOTS [35] dataset, which contains \\(\\approx 72\\mathrm{K}\\) training images. However, many images are low-quality and unrealistic, thus, we filtered the dataset and selected a random set of 2000 images - also to avoid imbalance w.r.t the other tasks. We use the standard outdoor test set of 500 images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.786, + 0.672 + ], + "angle": 0, + "content": "Image deblurring. We use the GoPro dataset for motion deblurring [50] which consists of 2103 images for training, and 1111 for testing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.672, + 0.786, + 0.715 + ], + "angle": 0, + "content": "Real-world Low-light Image Enhancement. We use the LOL [74] dataset (v1), which contains real-case low/normal-light image pairs. We adopt its official split of 485 training images and 15 testing images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.715, + 0.786, + 0.759 + ], + "angle": 0, + "content": "Real-world Image Enhancement. Extending previous works, we also study photorealistic image enhancement using the MIT5K DSLR dataset [5]. We use 1000 images for training, and the standard split of 500 images for testing (as in [66])." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Finally, as previous works [36,54,89], we combine all the aforementioned training datasets, and we train our unified model for all-in-one restoration. Note that we do not include more real-world datasets because previous works do not provide results (or models) for those. Moreover, previous works were limited to synthetic data, in contrast, InstructIR also tackles real-world image enhancement." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InstructIR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.224 + ], + "angle": 0, + "content": "Table 2: Quantitative results on five restoration tasks (5D) with state-of-the-art general image restoration and all-in-one methods. We highlight the reference model without text (image only), the best overall results, and the second best results. We also present the ablation study of our multi-task variants (from 5 to 7 tasks — 5D, 6D, 7D). This table is based on Zhang et al. IDR [89]." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.231, + 0.784, + 0.428 + ], + "angle": 0, + "content": "
MethodsDerainingDehazingDenoisingDeblurringLow-light Enh.AverageParams
Rain100L [78]SOTS [35]BSD68 [45]GoPro [50]LOL [74]
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑(M)
HINet [8]35.670.96924.740.93731.000.88126.120.78819.470.80027.400.87588.67
DGUNet [49]36.620.97124.780.94031.100.88327.250.83721.870.82328.320.89117.33
MIRNetV2 [84]33.890.95424.030.92730.970.88126.300.79921.520.81527.340.8755.86
SwinIR [38]30.780.92321.500.89130.590.86824.520.77317.810.72325.040.8350.91
Restormer [83]34.810.96224.090.92731.490.88427.220.82920.410.80627.600.88126.13
NAFNet [7]35.560.96725.230.93931.020.88326.530.80820.490.80927.760.88117.11
DL [17]21.960.76220.540.82623.090.74519.860.67219.830.71221.050.7432.09
Transweather [67]29.430.90521.320.88529.000.84125.120.75721.210.79225.220.83637.93
TAPE [39]29.670.90422.160.86130.180.85524.470.76318.970.62125.090.8011.07
AirNet [36]32.980.95121.040.88430.910.88224.350.78118.180.73525.490.8468.93
InstructIR w/o text35.580.96725.200.93831.090.88326.650.81020.700.82027.840.88417.11
IDR [89]35.630.96525.240.94331.600.88727.870.84621.340.82628.340.89315.34
InstructIR-5D36.840.97327.100.95631.400.88729.400.88623.000.83629.550.90715.8
InstructIR-6D36.800.97327.000.95131.390.88829.730.89222.830.83629.550.90815.8
InstructIR-7D36.750.97226.900.95231.370.88729.700.89222.810.83629.500.90715.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.441, + 0.785, + 0.484 + ], + "angle": 0, + "content": "Table 3: Comparisons of all-in-one restoration models for 3 restoration tasks (3D). We also show an ablation study for image denoising -the fundamental inverse problem-considering different noise levels. We report PSNR/SSIM metrics. Table based on [54]." + }, + { + "type": "table", + "bbox": [ + 0.249, + 0.492, + 0.759, + 0.624 + ], + "angle": 0, + "content": "
MethodsDehazing SOTS [35]Deraining Rain100L [17]Denoising ablation study σ = 1529.76/0.836 σ = 2526.34/0.836 σ = 50Average
BRDNet [64]23.23/0.89527.42/0.89532.26/0.89829.76/0.83626.34/0.83627.80/0.843
LPNet [20]20.84/0.82824.88/0.78426.47/0.77824.77/0.74821.26/0.55223.64/0.738
FDGAN [15]24.71/0.92429.89/0.93330.25/0.91028.81/0.86826.43/0.77628.02/0.883
MPRNet [85]25.28/0.95433.57/0.95433.54/0.92730.89/0.88027.56/0.77930.17/0.899
DL [17]26.92/0.93132.62/0.93133.05/0.91430.41/0.86126.90/0.74029.98/0.875
AirNet [36]27.94/0.96234.90/0.96733.92/0.93331.26/0.88828.00/0.79731.20/0.910
PromptIR [54]30.58/0.97436.37/0.97233.98/0.93331.31/0.88828.06/0.79932.06/0.913
InstructIR-3D30.22/0.95937.98/0.97834.15/0.93331.52/0.89028.30/0.80432.43/0.913
InstructIR-5D27.10/0.95636.84/0.97334.00/0.93131.40/0.88728.15/0.79831.50/0.909
InstructIR w/o text26.84/0.94834.02/0.96033.70/0.92930.94/0.88227.78/0.78030.65/0.900
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.652, + 0.506, + 0.668 + ], + "angle": 0, + "content": "4.3 Multiple Degradation Results" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.68, + 0.607, + 0.695 + ], + "angle": 0, + "content": "We define two initial setups for multi-task restoration:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.707, + 0.787, + 0.737 + ], + "angle": 0, + "content": "- 3D for three-degradation models such as AirNet [36], these tackle image denoising, dehazing and deraining." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.738, + 0.787, + 0.768 + ], + "angle": 0, + "content": "- 5D for five-degradation models, considering image denoising, deblurring, dehazing, deraining and low-light image enhancement as in [89]." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.707, + 0.787, + 0.768 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In Table 2, we show the performance of 5D models. Following Zhang et al. [89], we compare InstructIR with several state-of-the-art methods for general image restoration [7,8,38,83,84], and all-in-one image restoration methods [17,36, 39,67,89]. We can observe that our simple image model (just 16M parameters)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.411, + 0.128 + ], + "angle": 0, + "content": "Conde, Marcos et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "can tackle successfully at least five different tasks thanks to the instruction-based guidance and achieves the most competitive results. In Table 3 we can appreciate a similar behavior, when the number of tasks is just three (3D), our model improves further in terms of reconstruction performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.788, + 0.299 + ], + "angle": 0, + "content": "Based on these results, we pose the following question: How many tasks can we tackle using a single model without losing too much performance? To answer this, we propose the 6D and 7D variants. For the 6D variant, we fine-tune the original 5D to consider also super-resolution as sixth task. Finally, our 7D model includes all previous tasks, and additionally image enhancement (MIT5K photo retouching). We show the performance of these two variants in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.489, + 0.525 + ], + "angle": 0, + "content": "Test Instructions. InstructIR requires as input the degraded image and the human-written instruction. Therefore, we also prepare a test set of prompts i.e. instruction-image test pairs. The performance of InstructIR depends on the ambiguity and precision of the instruction. We provide the ablation study in Table 4. InstructIR is quite robust to more/less detailed instructions. However, it is still limited with ambiguous instructions such as \"enhance this image\". We show diverse instructions in Figures 5 and 6." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.329, + 0.788, + 0.454 + ], + "angle": 0, + "content": "Table 4: Ablation study on the sensitivity of instructions. We report PSNR/SSIM metrics for each task using our 5D base model. We repeat the evaluation on each test set 10 times, each time we sample different prompts for each image, and we report the average results. The \"Real Users \\( \\dagger \\) \" in this study are amateur photographers, thus, the instructions were very precise." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.456, + 0.784, + 0.497 + ], + "angle": 0, + "content": "
Language LevelDerainingDenoisingDeblurringLOL
Basic & Precise36.84/0.97331.40/0.88729.47/0.88723.00/0.836
Basic & Ambiguous36.24/0.97031.35/0.88729.21/0.88521.85/0.827
Real Users †36.84/0.97331.40/0.88729.47/0.88723.00/0.836
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.549, + 0.573, + 0.566 + ], + "angle": 0, + "content": "5 Multi-Task Discussion and Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.788, + 0.687 + ], + "angle": 0, + "content": "How does 6D work? Besides the 5 basic tasks -as previous works-, we include single image super-resolution (SISR). For this, we include as training data the DIV2K [1]. Since our model does not perform upsampling, we use the Bicubic degradation model [1,12] for generating the low-resolution images (LR), and the upsampled versions (HR) that are fed into our model to enhance them. Adding this extra task increases the performance on deblurring -a related degradation-, without harming notably the performance on the other tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.702, + 0.546, + 0.839 + ], + "angle": 0, + "content": "How does \\(7D\\) work? Finally, if we add real image enhancement -a task not related to the previous ones i.e. inverse problems- the performance on the restoration tasks decays slightly. However, the model still achieves state-of-the-art results. Moreover, as we show in Table 5, the performance on this task using the MIT5K [5] dataset is notable, while keeping the performance on the other tasks." + }, + { + "type": "table_caption", + "bbox": [ + 0.555, + 0.703, + 0.787, + 0.731 + ], + "angle": 0, + "content": "Table 5: Real Image Enhancement results on MIT5K [5]." + }, + { + "type": "table", + "bbox": [ + 0.56, + 0.732, + 0.784, + 0.81 + ], + "angle": 0, + "content": "
MethodPSNR↑ SSIM↑ ΔEab ↓
UPE [69]21.880.85310.80
DPE [21]23.750.9089.34
HDRNet [9]24.320.9128.49
3DLUT [86]25.210.9227.61
InstructIR-7D24.650.9008.20
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InstructIR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.144, + 0.361, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.211, + 0.313, + 0.223 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.144, + 0.501, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.364, + 0.21, + 0.501, + 0.237 + ], + "angle": 0, + "content": "\"Clean up my image, it's too fuzzy.\"" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.64, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.21, + 0.64, + 0.237 + ], + "angle": 0, + "content": "\"Get rid of the grain in my photo\"" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.144, + 0.78, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.21, + 0.779, + 0.237 + ], + "angle": 0, + "content": "\"Remove the strange spots on my photo\"" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.237, + 0.361, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.303, + 0.36, + 0.33 + ], + "angle": 0, + "content": "\"Retouch this image and improve colors\"" + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.237, + 0.501, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.303, + 0.497, + 0.33 + ], + "angle": 0, + "content": "\"Reduce the motion in this shot\"" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.237, + 0.64, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.303, + 0.631, + 0.33 + ], + "angle": 0, + "content": "\"Please get rid of the raindrops\"" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.237, + 0.78, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.303, + 0.771, + 0.33 + ], + "angle": 0, + "content": "\"Reduce the fog in this landmark\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.342, + 0.788, + 0.4 + ], + "angle": 0, + "content": "Fig. 5: Adversarial and OOD samples for Instruction-based Restoration. InstructIR understands a wide range of instructions for a given task (first row). Given an adversarial or out-of-distribution instruction (second row), the model does not modify the image notably (i.e. performs the identity) – we did not enforce this during training-." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.428, + 0.789, + 0.47 + ], + "angle": 0, + "content": "Table 7: Quantitative comparisons with state-of-the-art methods on the LOL [74] dataset for Real-World Low-light Enhancement. Note that InstructIR-7D is a multitask method, while the other methods are task-specific. Table based on [72]." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.476, + 0.784, + 0.589 + ], + "angle": 0, + "content": "
MethodLPNet [37]URetinex -Net [75]DeepLPF [48]SCI [44]LIME [22]MF [18]NPE [70]SRIE [19]SDD [23]CDEF [34]InstructIR Ours
PSNR ↑21.4621.3215.2815.8016.7616.9616.9611.8613.3416.3322.81
SSIM ↑0.8020.8350.4730.5270.4440.5050.4810.4930.6350.5830.836
MethodDRBN [79]KinD [93]RUAS [40]FIDE [77]EG [29]MS-RDN [80]Retinex -Net [74]MIRNet [84]IPT [6]Uformer [73]IAGC [72]
PSNR ↑20.1320.8718.2318.2717.4817.2016.7724.1416.2716.3624.53
SSIM ↑0.8300.8000.7200.6650.6500.6400.5600.8300.5040.5070.842
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.615, + 0.487, + 0.691 + ], + "angle": 0, + "content": "We summarize the multi-task ablation study in Table 6. Our model can tackle multiple tasks without losing performance notably thanks to the instruction-based task routing." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.613, + 0.787, + 0.654 + ], + "angle": 0, + "content": "Table 6: Summary ablation study on the multi-task variants of InstructIR that tackle from 3 to 7 tasks." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.656, + 0.783, + 0.719 + ], + "angle": 0, + "content": "
TasksRainNoise (σ15)BlurLOL
3D37.98/0.97831.52/0.890--
5D36.84/0.97331.40/0.88729.40/0.88623.00/0.836
6D36.80 0.97331.39 0.88829.73/0.89222.83 0.836
7D36.75 0.97231.37 0.88729.70/0.89222.81 0.836
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.742, + 0.788, + 0.804 + ], + "angle": 0, + "content": "Comparison with Task-specific Methods Our main goal is to design a powerful all-in-one model, thus, InstructIR was not designed to be trained for a particular degradation. Nevertheless, we also compare InstructIR with task-specific methods i.e. models tailored and trained for specific tasks." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We compare with task-specific methods for real-world photography enhancement in Table 5, and for real-world low-light image enhancement in Table 7." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.411, + 0.128 + ], + "angle": 0, + "content": "Conde, Marcos et al." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.145, + 0.409, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.235, + 0.335, + 0.247 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.147, + 0.589, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.432, + 0.235, + 0.761, + 0.247 + ], + "angle": 0, + "content": "(1) \"Clear the rain from my picture\" \\(\\longrightarrow\\) (2) \"Make it pop\"" + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.149, + 0.772, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.247, + 0.407, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.333, + 0.411, + 0.345 + ], + "angle": 0, + "content": "(1) \"Retouch it as a photographer" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.247, + 0.588, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.41, + 0.333, + 0.779, + 0.345 + ], + "angle": 0, + "content": "\\(r'' \\longrightarrow 'Can you remove the raindrops?\" \\longrightarrow 'Increase the resolution''\\)" + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.247, + 0.772, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.346, + 0.408, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.436, + 0.335, + 0.449 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.346, + 0.588, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.426, + 0.437, + 0.765, + 0.449 + ], + "angle": 0, + "content": "(1) \"My image is too dark, fix it\" \\(\\longrightarrow\\) (2) 'Apply a tonemap'" + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.347, + 0.772, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.461, + 0.788, + 0.504 + ], + "angle": 0, + "content": "Fig. 6: Control via instructions. We can prompt multiple instructions (in sequence) to restore and enhance the images. This provides additional control. We show two examples of multiple instructions applied to the \"Input\" image -from left to right-." + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.516, + 0.366, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.595, + 0.365, + 0.609 + ], + "angle": 0, + "content": "Input (RealSRSet)" + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.517, + 0.495, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.595, + 0.466, + 0.608 + ], + "angle": 0, + "content": "InstructIR" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.517, + 0.625, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.595, + 0.76, + 0.609 + ], + "angle": 0, + "content": "InstructPix2Pix #1 InstructPix2Pix #2" + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.517, + 0.757, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.618, + 0.789, + 0.647 + ], + "angle": 0, + "content": "Fig.7: Comparison with InstructPix2Pix [4] using the prompt \"Remove the noise in this photo\". Real-case image from RealSRSet [38]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.676, + 0.557, + 0.69 + ], + "angle": 0, + "content": "5.1 On the Effectiveness of Instructions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.787, + 0.746 + ], + "angle": 0, + "content": "Thanks to our integration of human instructions, users can control how to enhance the images. We provide examples in Figures 5 and 6, where we show the potential of our method to restore and enhance images in a controllable manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.747, + 0.788, + 0.807 + ], + "angle": 0, + "content": "This implies an advancement w.r.t classical (deterministic) image restoration methods. Classical deep restoration methods lead to a unique result, thus, they do not allow to control how the image is processed. We also compare InstructIR with InstructPix2Pix [4] (a diffusion-based generative model) in Figure 7." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Qualitative Results. We provide diverse qualitative results for several tasks, and we compare with all-in-one and task-specific methods. In Figure 10, we show" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InstructIR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.147, + 0.758, + 0.159 + ], + "angle": 0, + "content": "Instruction: \"my colors are too off, make it pop so I can use these photos in instagram\"" + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.16, + 0.364, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.16, + 0.497, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.16, + 0.63, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.16, + 0.765, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.232, + 0.768, + 0.244 + ], + "angle": 0, + "content": "\"Increase the brightness of my photo please, is it totoro?\" / \"my image is too dark, fix it\"" + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.244, + 0.364, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.244, + 0.497, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.244, + 0.63, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.244, + 0.765, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.324, + 0.784, + 0.338 + ], + "angle": 0, + "content": "Fig. 8: Real-world samples of image restoration and enhancement using InstructIR." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.352, + 0.332, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.353, + 0.443, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.353, + 0.554, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.353, + 0.664, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.353, + 0.776, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.419, + 0.332, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.471, + 0.297, + 0.484 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.42, + 0.443, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.471, + 0.425, + 0.484 + ], + "angle": 0, + "content": "AirNet [36]" + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.42, + 0.554, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.47, + 0.547, + 0.484 + ], + "angle": 0, + "content": "PromptIR [54]" + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.42, + 0.664, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.471, + 0.645, + 0.483 + ], + "angle": 0, + "content": "InstructIR" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.42, + 0.776, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.471, + 0.753, + 0.482 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.489, + 0.775, + 0.504 + ], + "angle": 0, + "content": "Fig. 9: Dehazing comparisons for all-in-one restoration methods on SOTS [35]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.788, + 0.704 + ], + "angle": 0, + "content": "results on the LOL [74] dataset. In Figure 11, we compare methods on the motion deblurring task using the GoPro [50] dataset. In Figures 9 and 12, we compare with different methods for the dehazing task on SOTS (outdoor) [35]. In Figure 13, we compare with image restoration methods for deraining on Rain100L [17]. Finally, we show denoising results in Figure 14. In this qualitative analysis, we use our single InstructIR-5D model to restore all the images. Limitations As with previous all-in-one methods, our model struggles to process images with more than one degradation (i.e. complex real-world images), or unknown out-of-distribution degradations, yet this is a common limitation among the related methods. However, we believe that these limitations can be surpassed with more realistic training data, and scaling the model's complexity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.718, + 0.359, + 0.734 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We present a novel approach that uses natural human-written instructions to guide the image restoration model. Given a prompt, our multi-task model can recover high-quality images from their degraded counterparts, considering multiple degradations. We achieve state-of-the-art results on several restoration tasks, demonstrating the power and flexibility of instruction guidance. Our results represent a new benchmark for text-guided image restoration." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.411, + 0.128 + ], + "angle": 0, + "content": "Conde, Marcos et al." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.153, + 0.403, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.236, + 0.349, + 0.245 + ], + "angle": 0, + "content": "Low-light Image" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.153, + 0.478, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.191, + 0.459, + 0.198 + ], + "angle": 0, + "content": "NAFNet" + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.154, + 0.551, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.191, + 0.532, + 0.198 + ], + "angle": 0, + "content": "MPRNet" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.154, + 0.624, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.191, + 0.607, + 0.198 + ], + "angle": 0, + "content": "DGUNet" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.153, + 0.698, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.191, + 0.685, + 0.198 + ], + "angle": 0, + "content": "MIRNetV2" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.154, + 0.772, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.191, + 0.76, + 0.198 + ], + "angle": 0, + "content": "InstructIR" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.2, + 0.402, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.236, + 0.453, + 0.244 + ], + "angle": 0, + "content": "TAPE" + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.2, + 0.478, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.236, + 0.535, + 0.244 + ], + "angle": 0, + "content": "Restormer" + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.2, + 0.551, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.236, + 0.603, + 0.244 + ], + "angle": 0, + "content": "AirNet" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.2, + 0.698, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.236, + 0.602, + 0.244 + ], + "angle": 0, + "content": "AirNet" + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.2, + 0.698, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.236, + 0.671, + 0.244 + ], + "angle": 0, + "content": "IDR" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.2, + 0.771, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.236, + 0.743, + 0.244 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.253, + 0.787, + 0.296 + ], + "angle": 0, + "content": "Fig. 10: Real Low-light Enhancement Results. LOL [74] testset (748.png). AirNet [36] and IDR [89] are well-known all-in-one restoration methods. NAFNet [7] is equivalent to InstructIR without text conditions (i.e. our image-only backbone)." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.32, + 0.404, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.404, + 0.347, + 0.412 + ], + "angle": 0, + "content": "Blurry Image" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.321, + 0.478, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.359, + 0.46, + 0.366 + ], + "angle": 0, + "content": "NAFNet" + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.321, + 0.551, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.359, + 0.532, + 0.366 + ], + "angle": 0, + "content": "MPRNet" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.321, + 0.624, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.359, + 0.607, + 0.366 + ], + "angle": 0, + "content": "DGUNet" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.321, + 0.699, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.359, + 0.685, + 0.366 + ], + "angle": 0, + "content": "MIRNetV2" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.321, + 0.772, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.359, + 0.76, + 0.366 + ], + "angle": 0, + "content": "InstructIR" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.366, + 0.404, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.405, + 0.366, + 0.478, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.404, + 0.453, + 0.412 + ], + "angle": 0, + "content": "TAPE" + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.366, + 0.551, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.404, + 0.536, + 0.412 + ], + "angle": 0, + "content": "Restormer" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.366, + 0.624, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.404, + 0.603, + 0.412 + ], + "angle": 0, + "content": "AirNet" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.366, + 0.699, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.404, + 0.685, + 0.412 + ], + "angle": 0, + "content": "IDR" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.366, + 0.772, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.404, + 0.743, + 0.412 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.414, + 0.699, + 0.428 + ], + "angle": 0, + "content": "Fig. 11: Image Deblurring Results. GoPro [50] dataset." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.454, + 0.404, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.537, + 0.338, + 0.546 + ], + "angle": 0, + "content": "Haze Image" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.455, + 0.478, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.492, + 0.459, + 0.499 + ], + "angle": 0, + "content": "NAFNet" + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.455, + 0.551, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.492, + 0.532, + 0.499 + ], + "angle": 0, + "content": "MPRNet" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.455, + 0.624, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.492, + 0.607, + 0.499 + ], + "angle": 0, + "content": "DGUNet" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.455, + 0.699, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.492, + 0.685, + 0.5 + ], + "angle": 0, + "content": "MIRNetV2" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.455, + 0.772, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.492, + 0.76, + 0.5 + ], + "angle": 0, + "content": "InstructIR" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.5, + 0.404, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.537, + 0.454, + 0.545 + ], + "angle": 0, + "content": "TAPE" + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.5, + 0.551, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.537, + 0.536, + 0.545 + ], + "angle": 0, + "content": "Restormer" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.5, + 0.624, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.537, + 0.603, + 0.545 + ], + "angle": 0, + "content": "AirNet" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.5, + 0.699, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.537, + 0.67, + 0.545 + ], + "angle": 0, + "content": "IDF" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.5, + 0.772, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.537, + 0.743, + 0.544 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.547, + 0.758, + 0.563 + ], + "angle": 0, + "content": "Fig. 12: Image Dehazing Results. SOTS [35] outdoor dataset (0150.jpg)." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.587, + 0.404, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.644, + 0.341, + 0.652 + ], + "angle": 0, + "content": "\\( \\therefore m = \\frac{3}{11} \\) ;" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.588, + 0.478, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.625, + 0.459, + 0.632 + ], + "angle": 0, + "content": "NAFNet" + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.588, + 0.551, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.625, + 0.532, + 0.632 + ], + "angle": 0, + "content": "MPRNet" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.588, + 0.624, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.625, + 0.607, + 0.632 + ], + "angle": 0, + "content": "DGUNet" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.588, + 0.699, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.625, + 0.685, + 0.632 + ], + "angle": 0, + "content": "MIRNetV2" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.588, + 0.772, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.625, + 0.76, + 0.632 + ], + "angle": 0, + "content": "InstructIR" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.653, + 0.404, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.67, + 0.341, + 0.678 + ], + "angle": 0, + "content": "Rainy Image" + }, + { + "type": "image", + "bbox": [ + 0.405, + 0.653, + 0.478, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.67, + 0.453, + 0.678 + ], + "angle": 0, + "content": "TAPE" + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.653, + 0.551, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.67, + 0.536, + 0.678 + ], + "angle": 0, + "content": "Restormer" + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.653, + 0.624, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.67, + 0.603, + 0.678 + ], + "angle": 0, + "content": "AirNet" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.653, + 0.699, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.67, + 0.671, + 0.678 + ], + "angle": 0, + "content": "IDR" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.653, + 0.772, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.67, + 0.743, + 0.678 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.68, + 0.722, + 0.696 + ], + "angle": 0, + "content": "Fig. 13: Image Deraining Results on Rain100L [17] (035.png)." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.72, + 0.404, + 0.804 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.803, + 0.341, + 0.812 + ], + "angle": 0, + "content": "Noise Image" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.721, + 0.478, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.758, + 0.459, + 0.766 + ], + "angle": 0, + "content": "NAFNet" + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.721, + 0.551, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.758, + 0.532, + 0.765 + ], + "angle": 0, + "content": "MPRNet" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.721, + 0.624, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.758, + 0.607, + 0.765 + ], + "angle": 0, + "content": "DGUNet" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.721, + 0.699, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.758, + 0.685, + 0.766 + ], + "angle": 0, + "content": "MIRNetV2" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.721, + 0.772, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.758, + 0.76, + 0.766 + ], + "angle": 0, + "content": "InstructIR" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.765, + 0.404, + 0.803 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.803, + 0.341, + 0.812 + ], + "angle": 0, + "content": "Noise Image" + }, + { + "type": "image", + "bbox": [ + 0.405, + 0.765, + 0.478, + 0.803 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.803, + 0.453, + 0.811 + ], + "angle": 0, + "content": "TAPE" + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.765, + 0.551, + 0.803 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.803, + 0.535, + 0.811 + ], + "angle": 0, + "content": "Restormer" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.765, + 0.624, + 0.803 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.803, + 0.603, + 0.811 + ], + "angle": 0, + "content": "AirNet" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.765, + 0.699, + 0.803 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.803, + 0.671, + 0.811 + ], + "angle": 0, + "content": "IDR" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.765, + 0.772, + 0.803 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.803, + 0.743, + 0.811 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.814, + 0.717, + 0.829 + ], + "angle": 0, + "content": "Fig. 14: Image Denoising Results on BSD68 [45] (0060.png)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InstructIR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.173, + 0.785, + 0.2 + ], + "angle": 0, + "content": "1. Agustsson, E., Timofte, R.: NTIRE 2017 challenge on single image superresolution: Dataset and study. In: CVPR Workshops (2017) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.201, + 0.785, + 0.227 + ], + "angle": 0, + "content": "2. Arbelaez, P., Maire, M., Fowlkes, C., Malik, J.: Contour detection and hierarchical image segmentation. TPAMI (2011) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.228, + 0.785, + 0.281 + ], + "angle": 0, + "content": "3. Bai, Y., Wang, C., Xie, S., Dong, C., Yuan, C., Wang, Z.: Textir: A simple framework for text-based editable image restoration. CoRR abs/2302.14736 (2023). https://doi.org/10.48550/ARXIV.2302.14736, https://doi.org/10.48550/arXiv.2302.14736 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.282, + 0.785, + 0.35 + ], + "angle": 0, + "content": "4. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023. pp. 18392-18402. IEEE (2023). https://doi.org/10.1109/CVPR52729.2023.01764, https://doi.org/10.1109/CVPR52729.2023.01764 2, 3, 4, 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.785, + 0.404 + ], + "angle": 0, + "content": "5. Bychkovsky, V., Paris, S., Chan, E., Durand, F.: Learning photographic global tonal adjustment with a database of input / output image pairs. In: The Twenty-Fourth IEEE Conference on Computer Vision and Pattern Recognition (2011) 8, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.785, + 0.432 + ], + "angle": 0, + "content": "6. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR (2021) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.785, + 0.458 + ], + "angle": 0, + "content": "7. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV (2022) 2, 3, 6, 7, 9, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.459, + 0.785, + 0.5 + ], + "angle": 0, + "content": "8. Chen, L., Lu, X., Zhang, J., Chu, X., Chen, C.: Hinet: Half instance normalization network for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 182-192 (2021) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.785, + 0.554 + ], + "angle": 0, + "content": "9. Chen, Y.S., Wang, Y.C., Kao, M.H., Chuang, Y.Y.: Deep photo enhancer: Unpaired learning for image enhancement from photographs with gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6306-6314 (2018) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.555, + 0.785, + 0.662 + ], + "angle": 0, + "content": "10. Devlin, J., Chang, M., Lee, K., Toutanova, K.: BERT: pre-training of deep bidirectional transformers for language understanding. In: Burstein, J., Doran, C., Solorio, T. (eds.) Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers). pp. 4171-4186. Association for Computational Linguistics (2019). https://doi.org/10.18653/V1/N19-1423, https://doi.org/10.18653/v1/n19-1423 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.663, + 0.785, + 0.718 + ], + "angle": 0, + "content": "11. Ding, C., Lu, Z., Wang, S., Cheng, R., Boddeti, V.N.: Mitigating task interference in multi-task learning via explicit task routing with non-learnable primitives. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7756-7765 (2023) 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.719, + 0.785, + 0.746 + ], + "angle": 0, + "content": "12. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. TPAMI (2015) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.746, + 0.785, + 0.772 + ], + "angle": 0, + "content": "13. Dong, H., Pan, J., Xiang, L., Hu, Z., Zhang, X., Wang, F., Yang, M.H.: Multi-scale boosted dehazing network with dense feature fusion. In: CVPR (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.773, + 0.785, + 0.799 + ], + "angle": 0, + "content": "14. Dong, W., Zhang, L., Shi, G., Wu, X.: Image deblurring and super-resolution by adaptive sparse domain selection and adaptive regularization. TIP (2011) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "15. Dong, Y., Liu, Y., Zhang, H., Chen, S., Qiao, Y.: Fd-gan: Generative adversarial networks with fusion-discriminator for single image dehazing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 10729-10736 (2020) 9" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.411, + 0.128 + ], + "angle": 0, + "content": "Conde, Marcos et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "16. Elad, M., Feuer, A.: Restoration of a single superresolution image from several blurred, noisy, and undersampled measured images. IEEE transactions on image processing 6(12), 1646-1658 (1997) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "17. Fan, Q., Chen, D., Yuan, L., Hua, G., Yu, N., Chen, B.: A general decoupled learning framework for parameterized image operators. IEEE transactions on pattern analysis and machine intelligence 43(1), 33-47 (2019) 9, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.787, + 0.257 + ], + "angle": 0, + "content": "18. Fu, X., Zeng, D., Huang, Y., Liao, Y., Ding, X., Paisley, J.: A fusion-based enhancing method for weakly illuminated images 129, 82-96 (2016) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.787, + 0.284 + ], + "angle": 0, + "content": "19. Fu, X., Zeng, D., Huang, Y., Zhang, X.P., Ding, X.: A weighted variational model for simultaneous reflectance and illumination estimation. In: CVPR (2016) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.787, + 0.311 + ], + "angle": 0, + "content": "20. Gao, H., Tao, X., Shen, X., Jia, J.: Dynamic scene deblurring with parameter selective sharing and nested skip connections. In: CVPR. pp. 3848-3856 (2019) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.787, + 0.352 + ], + "angle": 0, + "content": "21. Gharbi, M., Chen, J., Barron, J.T., Hasinoff, S.W., Durand, F.: Deep bilateral learning for real-time image enhancement. ACM Transactions on Graphics (TOG) 36(4), 1-12 (2017) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.787, + 0.379 + ], + "angle": 0, + "content": "22. Guo, X., Li, Y., Ling, H.: Lime: Low-light image enhancement via illumination map estimation. IEEE TIP 26(2), 982-993 (2016) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.405 + ], + "angle": 0, + "content": "23. Hao, S., Han, X., Guo, Y., Xu, X., Wang, M.: Low-light image enhancement with semi-decoupled decomposition. IEEE TMM 22(12), 3025-3038 (2020) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.405, + 0.787, + 0.445 + ], + "angle": 0, + "content": "24. He, J., Dong, C., Qiao, Y.: Modulating image restoration with continual levels via adaptive feature modification layers (2019), https://arxiv.org/abs/1904.081186" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.787, + 0.473 + ], + "angle": 0, + "content": "25. He, K., Sun, J., Tang, X.: Single image haze removal using dark channel prior. TPAMI (2010) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "26. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.542 + ], + "angle": 0, + "content": "27. Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al.: Searching for mobilenetv3. In: ICCV (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.787, + 0.582 + ], + "angle": 0, + "content": "28. Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5197-5206 (2015) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.787, + 0.623 + ], + "angle": 0, + "content": "29. Jiang, Y., Gong, X., Liu, D., Cheng, Y., Fang, C., Shen, X., Yang, J., Zhou, P., Wang, Z.: Enlightengan: Deep light enhancement without paired supervision. IEEE TIP 30, 2340-249 (2021) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.787, + 0.678 + ], + "angle": 0, + "content": "30. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.787, + 0.704 + ], + "angle": 0, + "content": "31. Kim, K.I., Kwon, Y.: Single-image super-resolution using sparse regression and natural image prior. TPAMI (2010) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.787, + 0.732 + ], + "angle": 0, + "content": "32. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv:1412.6980 (2014) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.787, + 0.772 + ], + "angle": 0, + "content": "33. Kopf, J., Neubert, B., Chen, B., Cohen, M., Cohen-Or, D., Deussen, O., Uyttendaele, M., Lischinski, D.: Deep photo: Model-based photograph enhancement and viewing. ACM TOG (2008) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "34. Lei, X., Fei, Z., Zhou, W., Zhou, H., Fei, M.: Low-light image enhancement using the cell vibration model. IEEE TMM pp. 1-1 (2022) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "35. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing 28(1), 492-505 (2018) 8, 9, 13, 14" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InstructIR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "36. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR. pp. 17452-17462 (June 2022) 2, 3, 7, 8, 9, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.176, + 0.787, + 0.203 + ], + "angle": 0, + "content": "37. Li, J., Li, J., Fang, F., Li, F., Zhang, G.: Luminance-aware pyramid network for low-light image enhancement. IEEE TMM 23, 3153-3165 (2020) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.204, + 0.786, + 0.231 + ], + "angle": 0, + "content": "38. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: SwinIR: Image restoration using swin transformer. In: ICCV Workshops (2021) 2, 9, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.286 + ], + "angle": 0, + "content": "39. Liu, L., Xie, L., Zhang, X., Yuan, S., Chen, X., Zhou, W., Li, H., Tian, Q.: Tape: Task-agnostic prior embedding for image restoration. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XVIII. pp. 447–464. Springer (2022) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "40. Liu, R., Ma, L., Zhang, J., Fan, X., Luo, Z.: Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In: CVPR (2021) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.368 + ], + "angle": 0, + "content": "41. Liu, Y., Liu, A., Gu, J., Zhang, Z., Wu, W., Qiao, Y., Dong, C.: Discovering distinctive \"semantics\" in super-resolution networks (2022), https://arxiv.org/abs/2108.00406 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.787, + 0.41 + ], + "angle": 0, + "content": "42. Ma, J., Cheng, T., Wang, G., Zhang, Q., Wang, X., Zhang, L.: Prores: Exploring degradation-aware visual prompt for universal image restoration. arXiv preprint arXiv:2306.13653 (2023) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.787, + 0.452 + ], + "angle": 0, + "content": "43. Ma, K., Duanmu, Z., Wu, Q., Wang, Z., Yong, H., Li, H., Zhang, L.: Waterloo exploration database: New challenges for image quality assessment models. TIP (2016) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.479 + ], + "angle": 0, + "content": "44. Ma, L., Ma, T., Liu, R., Fan, X., Luo, Z.: Toward fast, flexible, and robust low-light image enhancement. In: CVPR (2022) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.787, + 0.522 + ], + "angle": 0, + "content": "45. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: ICCV (2001) 8, 9, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.787, + 0.563 + ], + "angle": 0, + "content": "46. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.787, + 0.577 + ], + "angle": 0, + "content": "47. Michaeli, T., Irani, M.: Nonparametric blind super-resolution. In: ICCV (2013) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.605 + ], + "angle": 0, + "content": "48. Moran, S., Marza, P., McDonagh, S., Parisot, S., Slabaugh, G.: Deeplpf: Deep local parametric filters for image enhancement. In: CVPR (2020) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "49. Mou, C., Wang, Q., Zhang, J.: Deep generalized unfolding networks for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17399-17410 (2022) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.673 + ], + "angle": 0, + "content": "50. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: CVPR (2017) 8, 9, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.702 + ], + "angle": 0, + "content": "51. Nah, S., Son, S., Lee, J., Lee, K.M.: Clean images are hard to reblur: Exploiting the ill-posed inverse task for dynamic scene deblurring. In: ICLR (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.787, + 0.743 + ], + "angle": 0, + "content": "52. Nguyen, N., Milanfar, P., Golub, G.: Efficient generalized cross-validation with applications to parametric image restoration and resolution enhancement. IEEE Transactions on image processing 10(9), 1299-1308 (2001) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.799 + ], + "angle": 0, + "content": "53. Park, D., Lee, B.H., Chun, S.Y.: All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5815-5824. IEEE (2023) 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.838 + ], + "angle": 0, + "content": "54. Potlapalli, V., Zamir, S.W., Khan, S., Khan, F.S.: Prompt: Prompting for all-in-one blind image restoration. arXiv preprint arXiv:2306.13090 (2023) 2, 3, 8, 9, 13" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.838 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.411, + 0.128 + ], + "angle": 0, + "content": "Conde, Marcos et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.243 + ], + "angle": 0, + "content": "55. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event. Proceedings of Machine Learning Research, vol. 139, pp. 8748-8763. PMLR (2021), http://proceedings.mlr.press/v139/radford21a.html 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.341 + ], + "angle": 0, + "content": "56. Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert-networks. In: Inui, K., Jiang, J., Ng, V., Wan, X. (eds.) Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019. pp. 3980-3990. Association for Computational Linguistics (2019). https://doi.org/10.18653/V1/D19-1410, https://doi.org/10.18653/v1/D19-1410 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.786, + 0.37 + ], + "angle": 0, + "content": "57. Ren, C., He, X., Wang, C., Zhao, Z.: Adaptive consistency prior based deep network for image denoising. In: CVPR (2021) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.786, + 0.397 + ], + "angle": 0, + "content": "58. Ren, W., Pan, J., Zhang, H., Cao, X., Yang, M.H.: Single image dehazing via multi-scale convolutional neural networks with holistic edges. IJCV (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.786, + 0.466 + ], + "angle": 0, + "content": "59. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 10674-10685. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.01042, https://doi.org/10.1109/CVPR52688.2022.01042 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.786, + 0.494 + ], + "angle": 0, + "content": "60. Ronneberger, O., Fischer, P., Brox, T.: U-Net: convolutional networks for biomedical image segmentation. In: MICCAI (2015) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.786, + 0.535 + ], + "angle": 0, + "content": "61. Rosenbaum, C., Klinger, T., Riemer, M.: Routing networks: Adaptive selection of non-linear functions for multi-task learning. arXiv preprint arXiv:1711.01239 (2017) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.786, + 0.59 + ], + "angle": 0, + "content": "62. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.786, + 0.632 + ], + "angle": 0, + "content": "63. Strezoski, G., Noord, N.v., Worring, M.: Many task learning with task routing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1375-1384 (2019) 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.786, + 0.66 + ], + "angle": 0, + "content": "64. Tian, C., Xu, Y., Zuo, W.: Image denoising using deep cnn with batch renormalization. Neural Networks (2020) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.786, + 0.687 + ], + "angle": 0, + "content": "65. Timofte, R., De Smet, V., Van Gool, L.: Anchored neighborhood regression for fast example-based super-resolution. In: ICCV (2013) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.786, + 0.716 + ], + "angle": 0, + "content": "66. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A., Li, Y.: MAXIM: Multi-axis MLP for image processing. In: CVPR. pp. 5769-5780 (2022) 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.786, + 0.757 + ], + "angle": 0, + "content": "67. Valanarasu, J.M.J., Yasarla, R., Patel, V.M.: Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In: CVPR. pp. 2353-2363 (2022) 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.786, + 0.785 + ], + "angle": 0, + "content": "68. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.786, + 0.84 + ], + "angle": 0, + "content": "69. Wang, R., Zhang, Q., Fu, C.W., Shen, X., Zheng, W.S., Jia, J.: Underexposed photo enhancement using deep illumination estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6849-6857 (2019) 10" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InstructIR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.188 + ], + "angle": 0, + "content": "70. Wang, S., Zheng, J., Hu, H.M., Li, B.: Naturalness preserved enhancement algorithm for non-uniform illumination images. IEEE TIP 22(9), 3538-3548 (2013) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "71. Wang, X., Girshick, R., Gupta, A., He, K.: Non-local neural networks. In: CVPR (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.219, + 0.786, + 0.274 + ], + "angle": 0, + "content": "72. Wang, Y., Liu, Z., Liu, J., Xu, S., Liu, S.: Low-light image enhancement with illumination-aware gamma correction and complete image modelling network. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13128-13137 (2023) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.787, + 0.302 + ], + "angle": 0, + "content": "73. Wang, Z., Cun, X., Bao, J., Liu, J.: Uformer: A general u-shaped transformer for image restoration. arXiv:2106.03106 (2021) 2, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.787, + 0.331 + ], + "angle": 0, + "content": "74. Wei, C., Wang, W., Yang, W., Liu, J.: Deep retina decomposition for low-light enhancement. In: British Machine Vision Conference (2018) 8, 9, 11, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.333, + 0.787, + 0.372 + ], + "angle": 0, + "content": "75. Wu, W., Weng, J., Zhang, P., Wang, X., Yang, W., Jiang, J.: Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement. In: CVPR (2022) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.374, + 0.787, + 0.415 + ], + "angle": 0, + "content": "76. Xiao, S., Liu, Z., Zhang, P., Muennighof, N.: C-pack: Packaged resources to advance general chinese embedding. CoRR abs/2309.07597 (2023). https://doi.org/10.48550/ARXIV.2309.07597, https://doi.org/10.48550/arXiv.2309.07597" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.787, + 0.444 + ], + "angle": 0, + "content": "77. Xu, K., Yang, X., Yin, B., Lau, R.W.: Learning to restore low-light images via decomposition-and-enhancement. In: CVPR (2020) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.787, + 0.472 + ], + "angle": 0, + "content": "78. Yang, F., Yang, H., Fu, J., Lu, H., Guo, B.: Learning texture transformer network for image super-resolution. In: CVPR (2020) 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.787, + 0.515 + ], + "angle": 0, + "content": "79. Yang, W., Wang, S., Fang, Y., Wang, Y., Liu, J.: Band representation-based semi-supervised low-light image enhancement: bridging the gap between signal fidelity and perceptual quality. IEEE TIP 30, 3461-3473 (2021) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.517, + 0.787, + 0.557 + ], + "angle": 0, + "content": "80. Yang, W., Wang, W., Huang, H., Wang, S., Liu, J.: Sparse gradient regularized deep retina network for robust low-light image enhancement. IEEE TIP 30, 2072-2086 (2021) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.787, + 0.598 + ], + "angle": 0, + "content": "81. Yao, M., Xu, R., Guan, Y., Huang, J., Xiong, Z.: Neural degradation representation learning for all-in-one image restoration. arXiv preprint arXiv:2310.12848 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.787, + 0.642 + ], + "angle": 0, + "content": "82. Yu, W., Luo, M., Zhou, P., Si, C., Zhou, Y., Wang, X., Feng, J., Yan, S.: Metaformer is actually what you need for vision. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10819-10829 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.644, + 0.787, + 0.683 + ], + "angle": 0, + "content": "83. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022) 2, 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.787, + 0.727 + ], + "angle": 0, + "content": "84. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Learning enriched features for real image restoration and enhancement. In: ECCV (2020) 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.787, + 0.755 + ], + "angle": 0, + "content": "85. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR (2021) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.797 + ], + "angle": 0, + "content": "86. Zeng, H., Cai, J., Li, L., Cao, Z., Zhang, L.: Learning image-adaptive 3d lookup tables for high performance photo enhancement in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(4), 2058-2073 (2020) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "87. Zhang, C., Zhu, Y., Yan, Q., Sun, J., Zhang, Y.: All-in-one multi-degradation image restoration network via hierarchical degradation representation. arXiv preprint arXiv:2308.03021 (2023) 3" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.411, + 0.128 + ], + "angle": 0, + "content": "Conde, Marcos et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "88. Zhang, C., Zhu, Y., Yan, Q., Sun, J., Zhang, Y.: All-in-one multi-degradation image restoration network via hierarchical degradation representation. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 2285-2293 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.245 + ], + "angle": 0, + "content": "89. Zhang, J., Huang, J., Yao, M., Yang, Z., Yu, H., Zhou, M., Zhao, F.: Ingredient-oriented multi-degradation learning for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5825-5835 (2023) 2, 3, 8, 9, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.286 + ], + "angle": 0, + "content": "90. Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing 26(7), 3142-3155 (2017) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.314 + ], + "angle": 0, + "content": "91. Zhang, K., Zuo, W., Gu, S., Zhang, L.: Learning deep CNN denoiser prior for image restoration. In: CVPR (2017) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.342 + ], + "angle": 0, + "content": "92. Zhang, K., Luo, W., Zhong, Y., Ma, L., Stenger, B., Liu, W., Li, H.: Deblurring by realistic blurring. In: CVPR. pp. 2737-2746 (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.37 + ], + "angle": 0, + "content": "93. Zhang, Y., Zhang, J., Guo, X.: Kindling the darkness: A practical low-light image enhancer. In: ACM MM (2019) 11" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.37 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/InstructIR_ High-Quality Image Restoration Following Human Instructions/86f2cef3-3642-43ee-930e-b6ddf921d714_origin.pdf b/2024/InstructIR_ High-Quality Image Restoration Following Human Instructions/86f2cef3-3642-43ee-930e-b6ddf921d714_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..034eaccc2b8d33b13dc74ce252d6d81ed322aeeb --- /dev/null +++ b/2024/InstructIR_ High-Quality Image Restoration Following Human Instructions/86f2cef3-3642-43ee-930e-b6ddf921d714_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de40efec38c37ec86e3e3e8b0cb13f9b74b23acc60a93de2e8594e8bd1721355 +size 16481615 diff --git a/2024/InstructIR_ High-Quality Image Restoration Following Human Instructions/full.md b/2024/InstructIR_ High-Quality Image Restoration Following Human Instructions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..695bb118881f62702bb4da97b0ec942bafb3c919 --- /dev/null +++ b/2024/InstructIR_ High-Quality Image Restoration Following Human Instructions/full.md @@ -0,0 +1,613 @@ +# InstructIR: High-Quality Image Restoration Following Human Instructions + +Marcos V. Conde1,2, Gregor Geigle1, and Radu Timofte1 + +1 Computer Vision Lab, CAIDAS & IFI, University of Würzburg + +2 Sony PlayStation, FTG + +https://github.com/mv-lab/InstructIR + +![](images/f042f64cdc8d87bede4f89391118e4f2275a1a2bc840ec5a90bb5597a16bba4d.jpg) +Can you stabilize this pie? I was running... + +![](images/9dd094cb6c5f041e6306f3a5393cf19a191ec10a2123ccb53550381e6f70f876.jpg) +My image has tiny dots, doesn't look great, fix it + +![](images/2d2022e9d107b43ad89835ee2cca6ba1453ba164985093492f391320b734ce7c.jpg) +My photo is too dark, I cannot see anything + +![](images/2bf86199a634bc594aae62f760f80cb338a58b5d2eafb3e6a9711a26a5465b25.jpg) +Retouch my picture like a photographer + +![](images/38d7d6a663ed5b99ffee6dfc67dad3b49e4ecb144dd689daacad416089543c08.jpg) +It is a foggy day, can you make it clearer? + +![](images/0e811c9713c52e4d60a40020c5aaa1e0c6cd94acd1ed4ed44710756e7e299928.jpg) +Fig. 1: Given an image and a prompt for how to improve that image, our all-in-one restoration model corrects the image considering the human instruction. InstructIR, can tackle various types and levels of degradation, and it is able to generalize in some real-world scenarios (last three images, from left to right). + +![](images/8614232ac2764f401353daa5357f0876ac53be1e56804dca5239d99d4fb733b0.jpg) + +![](images/56ee4c73e5aa2082c5add852c5d29587a1cec94d89bc2677ab1f5e06aac2d923.jpg) + +![](images/c76e60981c7a373272d75a7079855c5531ee8e13949234ff9e266fbbd61e2105.jpg) + +![](images/f033464962ab159b067523ee26c82659a35232c6d93325dac3b5711cd6c7db66.jpg) + +Abstract. Image restoration is a fundamental problem that involves recovering a high-quality clean image from its degraded observation. All-In-One image restoration models can effectively restore images from various types and levels of degradation using degradation-specific information as prompts to guide the restoration model. In this work, we present the first approach that uses human-written instructions to guide the image restoration model. Given natural language prompts, our model can recover high-quality images from their degraded counterparts, considering multiple degradation types. Our method, InstructIR, achieves state-of-the-art results on several restoration tasks including image denoising, deraining, deblurring, dehazing, and (low-light) image enhancement. InstructIR improves $+1\mathrm{dB}$ over previous all-in-one restoration methods. Moreover, our dataset and results represent a novel benchmark for new research on text-guided image restoration and enhancement. + +# 1 Introduction + +Images often contain unpleasant effects such as noise, motion blur, haze, and low dynamic range. Such effects are commonly known in low-level computer + +vision as degradations. These can result from camera limitations or challenging environmental conditions e.g. low light. + +Image restoration aims to recover a high-quality image from its degraded counterpart. This is a complex inverse problem since multiple different solutions can exist for restoring any given image [13, 16, 38, 52, 90, 91]. + +Some methods focus on specific degradations, for instance reducing noise (denoising) [57, 90, 91], removing blur (deblurring) [51, 92], or clearing haze (dehazing) [13, 58]. Such methods are effective for their specific task, yet they do not generalize well to other types of degradation. Other approaches use a general neural network for diverse tasks [7, 66, 73, 83], yet training the neural network for each specific task independently. Since using a separate model for each possible degradation is resource-intensive, recent approaches propose All-in-One restoration models [36, 53, 54, 89]. These approaches use a single deep blind restoration model considering multiple degradation types and levels. Contemporary works such as PromptIR [54] or ProRes [42] utilize a unified model for blind image restoration using learned guidance vectors, also known as "prompt embeddings", in contrast to raw user prompts in text form, which we use in this work. + +In parallel, recent works such as InstructPix2Pix [4] show the potential of using text prompts to guide image generation and editing models. However, this method (or recent alternatives) do not tackle inverse problems. Inspired by these works, we argue that text guidance can help to guide blind restoration models better than the image-based degradation classification used in previous works [36,53,89]. Users generally have an idea about what has to be fixed (though they might lack domain-specific vocabulary) so we can use this information to guide the model. + +Contributions We propose the first approach that utilizes real human-written instructions to solve multi-task image restoration. Our comprehensive experiments demonstrate the potential of using text guidance for image restoration and enhancement by achieving state-of-the-art performance on various image restoration tasks, including image denoising, deraining, deblurring, dehazing, and low-light image enhancement. Our model, InstructIR, is able to generalize to restoring images using complex human-written instructions. Moreover, our single all-in-one model covers more tasks than many previous works. We show diverse restoration samples of our method in Figure 1. + +# 2 Related Work + +Image Restoration. Recent deep learning methods [13, 38, 51, 57, 66, 83] have shown consistently better results compared to traditional techniques for blind image restoration [14,25,31,33,47,65]. The proposed neural networks are based on convolutional neural networks (CNNs) and Transformers [68] (or related attention mechanisms). We focus on general-purpose restoration models [7,38,73,83]. For example, SwinIR [38], MAXIM [66] and Uformer [73]. These models can be trained -independently- for diverse tasks such as denoising, deraining or deblurring. Their ability to capture local and global feature interactions, and enhance + +them, allows the models to achieve great performance consistently across different tasks. For instance, Restormer [83] uses non-local blocks [71] to capture complex features across the image. + +NAFNet [7] is an efficient alternative to complex transformer-based methods. The model uses simplified channel attention, and gating as an alternative to non-linear activations. The building block (NAFBlock) follows a simple metaformer [82] architecture with efficient inverted residual blocks [27]. In this work, we build our InstructIR model using NAFNet as backbone, due to its efficient and simple design, and high performance in several restoration tasks. + +All-in-One Image Restoration. Single degradation (or single task) restoration methods are well-studied, however, their real-world applications are limited due to the required resources i.e. allocating different models, and selecting the adequate model on demand. Moreover, images rarely present a single degradation, for instance, noise and blur are almost ubiquitous in any image capture. + +All-in-One (also known as multi-degradation or multi-task) image restoration is emerging as a new research field in low-level computer vision [36,42,53,54,67, 81,87,88]. These approaches use a single deep blind restoration model to tackle different degradation types and levels. + +We use as reference AirNet [36], IDR [89] and ADMS [53]. We also consider the contemporary work PromptIR [54]. The methods use different techniques to guide the blind model in the restoration process. For instance, an auxiliary model for degradation classification [36, 53], or multi-dimensional guidance vectors (also known as "prompts") [41, 42, 54] that help the model to discriminate the different types of degradation in the image. + +Text-guided Image Manipulation. In recent years, multiple methods have been proposed for text-to-image generation and text-based image editing works [4, 26, 30, 46, 62]. These models use text prompts to describe images or actions, and powerful diffusion-based models for generating the corresponding images. Our main reference is InstructPix2Pix [4], this method enables editing from instructions that tell the model what action to perform, as opposed to text labels, captions or descriptions of the input or output images. Therefore, the user can transmit what to do in natural written text, without requiring to provide further image descriptions or sample reference images. + +# 3 Image Restoration Following Instructions + +We treat instruction-based image restoration as a supervised learning problem similar to previous works [4]. First, we generate over 10000 prompts using GPT-4 based on our own sample instructions. We explain the creation of the prompt dataset in Sec. 3.1. We then build a large paired training dataset of prompts and degraded/clean images. Finally, we train our InstructIR model, and we evaluate it on a wide variety of instructions including real human-written prompts. We explain our text encoder in Sec 3.2, and our complete model in Sec. 3.3. + +# 3.1 Generating Prompts for Training + +Why instructions? Inspired by InstructPix2Pix [4], we adopt human written instructions as the mechanism of control for our model. There is no need for the user to provide additional information, such as example clean images, or descriptions of the visual content. Instructions offer a clear and expressive way to interact, enabling users to pinpoint the unpleasant effects (degradations) in the images. We also consider the language complexity, from ambiguous instructions (e.g. "fix my image") to precise instructions (e.g. "remove the noise"). + +Handling free-form user prompts rather than fixed degradation-specific prompts increases the usability of our model for laypeople who lack domain expertise. We thus want our model to be capable of understanding diverse prompts posed by users "in-the-wild" e.g. kids, adults, or photographers. To this end, we use a large language model (i.e., GPT-4) to create diverse requests that might be asked by users for the different degradations types. We then filter those generated prompts to remove ambiguous or unclear prompts (e.g., "Make the image cleaner", "improve this image"). Our final instructions set contains over 10000 different prompts in total, for 7 different tasks. We display some examples in Table 1. As we show in Figure 2 the prompts are sampled randomly depending on the input degradation. + +Table 1: Examples of our curated GPT4-generated and real user prompts with varying language and domain expertise. + +
Degradation Prompts
DenoisingCan you clean the dots from my image? Fix the grainy parts of this photo Remove the noise from my picture
DeblurringCan you reduce the movement in the image? My picture's not sharp, fix it Deblur my picture, it's too fuzzy
DehazingCan you make this picture clearer? Help, my picture is all cloudy Remove the fog from my photo
DerainingI want my photo to be clear, not rainy Clear the rain from my picture Remove the raindrops from my photo
Super-Res.Make my photo bigger and better Add details to this image Increase the resolution of this photo
Low-lightThe photo is too dark, improve exposure Increase the illumination in this shot My shot has very low dynamic range
EnhancementMake it pop! Adjust the color balance for a natural look Apply a cinematic color grade to the photo
GeneralFix my image please make the image look better
+ +![](images/9bbcf64a2b9c6ecf65d1245798bab7130f8c4a344ccf15552c000cbef1e1a42b.jpg) +Fig. 2: We train our blind image restoration models using common image datasets, and prompts generated using GPT-4, note that this is (self-)supervised learning. At inference time, our model generalizes to human-written instructions and restores (or enhances) the images. + +# 3.2 Text Encoder + +The Choice of the Text Encoder. A text encoder maps the user prompt to a fixed-size vector representation (a text embedding). The related methods for text-based image generation [59] and manipulation [3,4] often use the text encoder of a CLIP model [55] to encode user prompts as CLIP excels in visual prompts. However, user prompts for degradation contain, in general, little to no visual content (e.g. the use describes the degradation, not the image itself). We opt, instead, to use a pure text-based sentence encoder [56], that is, a smaller model trained to encode sentences in a semantically meaningful embedding space. Sentence encoders - pre-trained with millions of examples - are compact and fast in comparison to CLIP, while being able to encode the semantics of diverse user prompts. For instance, we use the BGE-micro-v2 sentence transformer. We compare text encoders in the supplementary material. + +Fine-tuning the Text Encoder. We want to adapt the text encoder E for the restoration task to better encode the required information for the restoration model. Training the full text encoder is likely to lead to overfitting on our small training set and lead to loss of generalization. Instead, we freeze the text encoder and train a projection head on top: + +$$ +\mathbf {e} = \operatorname {n o r m} (\mathbf {W} \cdot \mathrm {E} (t)) \tag {1} +$$ + +where $t$ is the text, $\mathrm{E}(t)$ represents the raw text embedding, $\mathbf{W} \in \mathbb{R}^{d_t \times d_v}$ is a learned projection from the text dimension $(d_t)$ to the input dimension for the restoration model $(d_v)$ , and norm is the 12-norm. + +Figure 3 shows that while the text encoder is capable out-of-the-box to cluster instructions to some extent (Figure 3a), our trained projection yields greatly improved clusters (Figure 3b). We distinguish clearly the clusters for deraining, denoising, dehazing, deblurring, and low-light image enhancement. The instructions for such tasks or degradations are very characteristic. Furthermore, we can appreciate that "super-res" and "enhancement" tasks are quite spread and between the previous ones, which matches the language logic. For instance "add details to this image" could be used for enhancement, deblurring, or denoising. In our experiments, $d_{t} = 384$ , $d_{v} = 256$ and $\mathbf{W}$ is a linear layer. The representation $\mathbf{e}$ from the text encoder is shared across the blocks, and each block has a trainable projection $\mathbf{W}$ . + +Intent Classification Loss. We propose a guidance loss on the text embedding $\mathbf{e}$ to improve training and interpretability. Using the degradation types as targets, we train a simple classification head $\mathcal{C}$ such that $\mathbf{c} = \mathcal{C}(\mathbf{e})$ , where $\mathbf{c} \in \mathbb{R}^{D}$ , being $D$ is the number of degradation classes. The classification head $\mathcal{C}$ is a simple two-layers MLP. Thus, we only need to train a projection layer $\mathbf{W}$ and a simple MLP to capture the natural language knowledge. This allows the text model to learn meaningful embeddings as we can appreciate in Figure 3, not just guidance vectors for the main image processing model. We find that the model is able to classify accurately (i.e. over $95\%$ accuracy) the underlying degradation in the user's prompt after a few epochs. + +![](images/ea91eaf38deadf1d7358efae1e44401fb00bdf1492e55ecbf948c468ec3a0c07.jpg) +(a) t-SNE of embeddings before training i.e. frozen text encoder + +![](images/e185e23b332ca2daba02817a4bf95b7edc9ba593d262562571d35b704d3849cb.jpg) +(b) t-SNE of embeddings after training our learned projection +Fig. 3: We show t-SNE plots of the text embeddings before/after training InstructIR. Each dot represents a human instruction. + +# 3.3 InstructIR + +Our method InstructIR consists of an image model and a text encoder. We introduced our text encoder in Sec. 3.2. We use NAFNet [7] as the image model, an efficient image restoration model that follows a U-Net architecture [60]. To successfully learn multiple tasks using a single model, we use task routing techniques. Our framework for training and evaluation is illustrated in Figure 2. + +Text Guidance. The key aspect of InstructIR is the integration of the encoded instruction as a mechanism of control for the image model. Inspired in task routing for many-task learning [11, 61, 63], we propose an "Instruction Condition Block" (ICB) to enable task-specific transformations within the model. Conventional task routing [63] applies task-specific binary masks to the channel features. Since our model does not know a priori the degradation, we cannot use this technique directly. + +Considering the image features $\mathcal{F}$ and the encoded instruction $\mathbf{e}$ , we apply task routing as follows: + +![](images/55a6c6b16f37b8e82e7dab78d358b43fa9e5915cba3093a16b8577cf0966cc5d.jpg) +Fig. 4: Instruction Condition Block (ICB) using an approximation of task routing [63] for many-tasks learning (See Eq. 2). This mechanism allows the neural network to select and prioritize specific features depending on the instruction, similarly to a Mixture of Experts (MoE). + +$$ +\mathcal {F} _ {c} ^ {\prime} = \operatorname {B l o c k} \left(\mathcal {F} _ {c} \odot \mathbf {m} _ {c}\right) + \mathcal {F} _ {c} \tag {2} +$$ + +where the mask $\mathbf{m}_c = \sigma (\mathbf{W_c}\cdot \mathbf{e})$ is produced using a linear layer $\mathbf{W}_{\mathbf{c}}$ activated using the Sigmoid function - to produce a set of weights depending on the text embedding $\mathbf{e}$ . Thus, we obtain a $c$ -dimensional per-channel (soft-)binary mask $\mathbf{m}_c$ . As [24, 63], task routing is applied as the channel-wise multiplication $\odot$ for masking features depending on the task. The conditioned features are further enhanced using a convolutional NAFBlock [7] (Block). We illustrate our task-routing ICB block in Figure 4. We use "regular" NAFBlocks [7], followed by ICBs to condition the features, at both encoder and decoder blocks. The + +formulation is $F^{l + 1} = \mathrm{ICB}(\mathrm{Block}(F^l))$ where $l$ is the layer. Although we do not condition explicitly the filters of the neural network, as in [63], the mask allows the model to select the most relevant channels depending on the image information and the instruction. Note that this formulation enables differentiable feature masking, and certain interpretability i.e. the features with high weights contribute the most to the restoration process. Indirectly, this also enforces to learn diverse filters and reduce sparsity [11, 63]. + +Is InstructIR a blind restoration model? The model does not use explicit information about the degradation in the image e.g. noise profiles, blur kernels, or PSFs. Since our model infers the task (degradation) given the image and the instruction, we consider InstructIR a blind image restoration model. Similarly to previous works that use auxiliary image-based degradation classification [36,53]. + +# 4 Experimental Results + +We evaluate our model on 9 well-known benchmarks for different image restoration tasks: image denoising, deblurring, deraining, dehazing, real low-light enhancement, and photo-realistic image enhancement. We present extensive quantitative results in Table 2 and Table 3. We provide extensive comparisons with other all-in-one methods as well as task-specific methods. Our single model successfully restores images considering different degradation types and levels. We provide additional results and ablation studies in the supplementary material. + +# 4.1 Implementation Details. + +Our InstructIR model is end-to-end trainable. The image model does not require pre-training but we use a pre-trained sentence encoder as language model. + +Text Encoder. As we discussed in Sec. 3.2, we only need to train the text embedding projection and classification head ( $\approx$ 100K parameters). We initialize the text encoder with BGE-MICRO-v2 $^{3}$ , a distilled version of BGE-SMALL-EN [76]. The BGE encoders are BERT-like encoders [10] pre-trained on large amounts of supervised and unsupervised data for general-purpose sentence encoding. The BGE-micro model is a 3-layer encoder with 17.3 million parameters, which we freeze during training. We also explore ALL-MINILM-L6-v2 and CLIP encoders, however, we concluded that small models prevent overfitting and provide the best performance while being fast. We provide the ablation study comparing the three text encoders in the supplementary material. + +Image Model. We use NAFNet [7] as the image model backbone. The architecture consists of a 4-level encoder-decoder, with varying numbers of blocks at each level, specifically [2, 2, 4, 8] for the encoder, and [2, 2, 2, 2] for the decoder, from the level-1 to level-4 respectively. Between the encoder and decoder we use + +4 middle blocks to enhance further the features. The decoder implements addition instead of concatenation for the skip connections. We use the Instruction Condition Block (ICB) for task-routing [63] only in the encoder and decoder. + +The model is optimized using the $\mathcal{L}_1$ loss between the ground-truth clean image and the restored one. Additionally, we use the cross-entropy loss $\mathcal{L}_{ce}$ for the intent classification head of the text encoder. We train using a batch size of 32 and AdamW [32] optimizer with learning rate $5e^{-4}$ for 500 epochs (approximately 1 day using a single NVIDIA A100). We also use cosine annealing learning rate decay. During training, we utilize cropped patches of size $256 \times 256$ as input, and we use random horizontal and vertical flips as augmentations. Since our model uses as input instruction-image pairs, given an image, and knowing its degradation, we randomly sample instructions from our prompt dataset ( $>10\mathrm{K}$ samples). Our image model has only 16M parameters, and the learned text projection is just 100k parameters (the language model is 17M parameters), thus, our model can be trained easily on standard GPUs such as NVIDIA RTX 2080Ti or 3090Ti in a couple of days. Furthermore, the inference process also fits in low-computation budgets (e.g. Google Colab T4 16Gb GPU). + +# 4.2 Datasets and Benchmarks + +Following previous works [36,54,89], we prepare the datasets for different restoration tasks, including real and synthetic datasets. + +Image denoising. We use a combination of BSD400 [2] and WED [43] datasets for training. This combined training set contains $\approx 5000$ images. Using as reference the clean images in the dataset, we generate the noisy images by adding Gaussian noise with different noise levels $\sigma \in \{15, 25, 50\}$ . We test the models on the well-known BSD68 [45] and Urban100 [28] datasets. + +Image deraining. We use the Rain100L [78] dataset, which consists of 200 clean-rainy image pairs for training, and 100 pairs for testing. + +Image dehazing. We utilize the Reside (outdoor) SOTS [35] dataset, which contains $\approx 72\mathrm{K}$ training images. However, many images are low-quality and unrealistic, thus, we filtered the dataset and selected a random set of 2000 images - also to avoid imbalance w.r.t the other tasks. We use the standard outdoor test set of 500 images. + +Image deblurring. We use the GoPro dataset for motion deblurring [50] which consists of 2103 images for training, and 1111 for testing. + +Real-world Low-light Image Enhancement. We use the LOL [74] dataset (v1), which contains real-case low/normal-light image pairs. We adopt its official split of 485 training images and 15 testing images. + +Real-world Image Enhancement. Extending previous works, we also study photorealistic image enhancement using the MIT5K DSLR dataset [5]. We use 1000 images for training, and the standard split of 500 images for testing (as in [66]). + +Finally, as previous works [36,54,89], we combine all the aforementioned training datasets, and we train our unified model for all-in-one restoration. Note that we do not include more real-world datasets because previous works do not provide results (or models) for those. Moreover, previous works were limited to synthetic data, in contrast, InstructIR also tackles real-world image enhancement. + +Table 2: Quantitative results on five restoration tasks (5D) with state-of-the-art general image restoration and all-in-one methods. We highlight the reference model without text (image only), the best overall results, and the second best results. We also present the ablation study of our multi-task variants (from 5 to 7 tasks — 5D, 6D, 7D). This table is based on Zhang et al. IDR [89]. + +
MethodsDerainingDehazingDenoisingDeblurringLow-light Enh.AverageParams
Rain100L [78]SOTS [35]BSD68 [45]GoPro [50]LOL [74]
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑(M)
HINet [8]35.670.96924.740.93731.000.88126.120.78819.470.80027.400.87588.67
DGUNet [49]36.620.97124.780.94031.100.88327.250.83721.870.82328.320.89117.33
MIRNetV2 [84]33.890.95424.030.92730.970.88126.300.79921.520.81527.340.8755.86
SwinIR [38]30.780.92321.500.89130.590.86824.520.77317.810.72325.040.8350.91
Restormer [83]34.810.96224.090.92731.490.88427.220.82920.410.80627.600.88126.13
NAFNet [7]35.560.96725.230.93931.020.88326.530.80820.490.80927.760.88117.11
DL [17]21.960.76220.540.82623.090.74519.860.67219.830.71221.050.7432.09
Transweather [67]29.430.90521.320.88529.000.84125.120.75721.210.79225.220.83637.93
TAPE [39]29.670.90422.160.86130.180.85524.470.76318.970.62125.090.8011.07
AirNet [36]32.980.95121.040.88430.910.88224.350.78118.180.73525.490.8468.93
InstructIR w/o text35.580.96725.200.93831.090.88326.650.81020.700.82027.840.88417.11
IDR [89]35.630.96525.240.94331.600.88727.870.84621.340.82628.340.89315.34
InstructIR-5D36.840.97327.100.95631.400.88729.400.88623.000.83629.550.90715.8
InstructIR-6D36.800.97327.000.95131.390.88829.730.89222.830.83629.550.90815.8
InstructIR-7D36.750.97226.900.95231.370.88729.700.89222.810.83629.500.90715.8
+ +Table 3: Comparisons of all-in-one restoration models for 3 restoration tasks (3D). We also show an ablation study for image denoising -the fundamental inverse problem-considering different noise levels. We report PSNR/SSIM metrics. Table based on [54]. + +
MethodsDehazing SOTS [35]Deraining Rain100L [17]Denoising ablation study σ = 1529.76/0.836 σ = 2526.34/0.836 σ = 50Average
BRDNet [64]23.23/0.89527.42/0.89532.26/0.89829.76/0.83626.34/0.83627.80/0.843
LPNet [20]20.84/0.82824.88/0.78426.47/0.77824.77/0.74821.26/0.55223.64/0.738
FDGAN [15]24.71/0.92429.89/0.93330.25/0.91028.81/0.86826.43/0.77628.02/0.883
MPRNet [85]25.28/0.95433.57/0.95433.54/0.92730.89/0.88027.56/0.77930.17/0.899
DL [17]26.92/0.93132.62/0.93133.05/0.91430.41/0.86126.90/0.74029.98/0.875
AirNet [36]27.94/0.96234.90/0.96733.92/0.93331.26/0.88828.00/0.79731.20/0.910
PromptIR [54]30.58/0.97436.37/0.97233.98/0.93331.31/0.88828.06/0.79932.06/0.913
InstructIR-3D30.22/0.95937.98/0.97834.15/0.93331.52/0.89028.30/0.80432.43/0.913
InstructIR-5D27.10/0.95636.84/0.97334.00/0.93131.40/0.88728.15/0.79831.50/0.909
InstructIR w/o text26.84/0.94834.02/0.96033.70/0.92930.94/0.88227.78/0.78030.65/0.900
+ +# 4.3 Multiple Degradation Results + +We define two initial setups for multi-task restoration: + +- 3D for three-degradation models such as AirNet [36], these tackle image denoising, dehazing and deraining. +- 5D for five-degradation models, considering image denoising, deblurring, dehazing, deraining and low-light image enhancement as in [89]. + +In Table 2, we show the performance of 5D models. Following Zhang et al. [89], we compare InstructIR with several state-of-the-art methods for general image restoration [7,8,38,83,84], and all-in-one image restoration methods [17,36, 39,67,89]. We can observe that our simple image model (just 16M parameters) + +can tackle successfully at least five different tasks thanks to the instruction-based guidance and achieves the most competitive results. In Table 3 we can appreciate a similar behavior, when the number of tasks is just three (3D), our model improves further in terms of reconstruction performance. + +Based on these results, we pose the following question: How many tasks can we tackle using a single model without losing too much performance? To answer this, we propose the 6D and 7D variants. For the 6D variant, we fine-tune the original 5D to consider also super-resolution as sixth task. Finally, our 7D model includes all previous tasks, and additionally image enhancement (MIT5K photo retouching). We show the performance of these two variants in Table 2. + +Test Instructions. InstructIR requires as input the degraded image and the human-written instruction. Therefore, we also prepare a test set of prompts i.e. instruction-image test pairs. The performance of InstructIR depends on the ambiguity and precision of the instruction. We provide the ablation study in Table 4. InstructIR is quite robust to more/less detailed instructions. However, it is still limited with ambiguous instructions such as "enhance this image". We show diverse instructions in Figures 5 and 6. + +Table 4: Ablation study on the sensitivity of instructions. We report PSNR/SSIM metrics for each task using our 5D base model. We repeat the evaluation on each test set 10 times, each time we sample different prompts for each image, and we report the average results. The "Real Users $\dagger$ " in this study are amateur photographers, thus, the instructions were very precise. + +
Language LevelDerainingDenoisingDeblurringLOL
Basic & Precise36.84/0.97331.40/0.88729.47/0.88723.00/0.836
Basic & Ambiguous36.24/0.97031.35/0.88729.21/0.88521.85/0.827
Real Users †36.84/0.97331.40/0.88729.47/0.88723.00/0.836
+ +# 5 Multi-Task Discussion and Study + +How does 6D work? Besides the 5 basic tasks -as previous works-, we include single image super-resolution (SISR). For this, we include as training data the DIV2K [1]. Since our model does not perform upsampling, we use the Bicubic degradation model [1,12] for generating the low-resolution images (LR), and the upsampled versions (HR) that are fed into our model to enhance them. Adding this extra task increases the performance on deblurring -a related degradation-, without harming notably the performance on the other tasks. + +How does $7D$ work? Finally, if we add real image enhancement -a task not related to the previous ones i.e. inverse problems- the performance on the restoration tasks decays slightly. However, the model still achieves state-of-the-art results. Moreover, as we show in Table 5, the performance on this task using the MIT5K [5] dataset is notable, while keeping the performance on the other tasks. + +Table 5: Real Image Enhancement results on MIT5K [5]. + +
MethodPSNR↑ SSIM↑ ΔEab ↓
UPE [69]21.880.85310.80
DPE [21]23.750.9089.34
HDRNet [9]24.320.9128.49
3DLUT [86]25.210.9227.61
InstructIR-7D24.650.9008.20
+ +![](images/8bcccb2bada2614507419e35540c265557176728e894c331fa84a74e60ff2ca2.jpg) + +![](images/0ef09a6f541d427173452ee88a6f80ad132db1d5bbbe39f24dd49e5e809b8145.jpg) + +![](images/add9da21428c8935aa2c35ad808f39b92c6d3abc6049c3caeffaa5b200f8d36e.jpg) + +![](images/81c2619cc80a3b4be0258702bf19df7a29061e39ab58b9b7d0516dbab4278f02.jpg) + +![](images/7b15fd1272c8a874060da8925a4c3edfc5944c2c860ccfb789898aa0e8fcd7a1.jpg) +Input +"Retouch this image and improve colors" +Fig. 5: Adversarial and OOD samples for Instruction-based Restoration. InstructIR understands a wide range of instructions for a given task (first row). Given an adversarial or out-of-distribution instruction (second row), the model does not modify the image notably (i.e. performs the identity) – we did not enforce this during training-. + +![](images/cbd76114bdf1ec34600399e82596afbf57b92df835d5baf86ac36ceeb0c88d59.jpg) +"Clean up my image, it's too fuzzy." +"Reduce the motion in this shot" + +![](images/19d4b1a93dff9047f2a26c4a36168ea84fdc642bc6c34fbf644761570f447d6b.jpg) +"Get rid of the grain in my photo" +"Please get rid of the raindrops" + +![](images/22af49b9e372b4790a0cd0e27a054aba1f0b8d1cd63a3c96920d6a2be17cb6ce.jpg) +"Remove the strange spots on my photo" +"Reduce the fog in this landmark" + +Table 7: Quantitative comparisons with state-of-the-art methods on the LOL [74] dataset for Real-World Low-light Enhancement. Note that InstructIR-7D is a multitask method, while the other methods are task-specific. Table based on [72]. + +
MethodLPNet [37]URetinex -Net [75]DeepLPF [48]SCI [44]LIME [22]MF [18]NPE [70]SRIE [19]SDD [23]CDEF [34]InstructIR Ours
PSNR ↑21.4621.3215.2815.8016.7616.9616.9611.8613.3416.3322.81
SSIM ↑0.8020.8350.4730.5270.4440.5050.4810.4930.6350.5830.836
MethodDRBN [79]KinD [93]RUAS [40]FIDE [77]EG [29]MS-RDN [80]Retinex -Net [74]MIRNet [84]IPT [6]Uformer [73]IAGC [72]
PSNR ↑20.1320.8718.2318.2717.4817.2016.7724.1416.2716.3624.53
SSIM ↑0.8300.8000.7200.6650.6500.6400.5600.8300.5040.5070.842
+ +We summarize the multi-task ablation study in Table 6. Our model can tackle multiple tasks without losing performance notably thanks to the instruction-based task routing. + +Table 6: Summary ablation study on the multi-task variants of InstructIR that tackle from 3 to 7 tasks. + +
TasksRainNoise (σ15)BlurLOL
3D37.98/0.97831.52/0.890--
5D36.84/0.97331.40/0.88729.40/0.88623.00/0.836
6D36.80 0.97331.39 0.88829.73/0.89222.83 0.836
7D36.75 0.97231.37 0.88729.70/0.89222.81 0.836
+ +Comparison with Task-specific Methods Our main goal is to design a powerful all-in-one model, thus, InstructIR was not designed to be trained for a particular degradation. Nevertheless, we also compare InstructIR with task-specific methods i.e. models tailored and trained for specific tasks. + +We compare with task-specific methods for real-world photography enhancement in Table 5, and for real-world low-light image enhancement in Table 7. + +![](images/60416b4c0584616589a2dcc447494d0a3e7d7b6ab89c0de32db686af5e21767c.jpg) + +![](images/7634f3a49bcc4952134959d94071342db8ffbd6afcb27d2a8f4dc1d4a8ffe8b4.jpg) + +![](images/0bf5d9047ce5bb3c306cc50271400608be89ef3463e4d7b1d027b91800bd5ee7.jpg) + +![](images/c8a16fb2224007a58f6b571e89331558607424a72f7a0b2240065ec8608012c1.jpg) +Input +(1) "Retouch it as a photographer + +![](images/e8192ce76b7575b1fdaf5fbc9e2d3118ecab16261e3f76575b0ab576a63ec396.jpg) +(1) "Clear the rain from my picture" $\longrightarrow$ (2) "Make it pop" +$r'' \longrightarrow 'Can you remove the raindrops?" \longrightarrow 'Increase the resolution''$ + +![](images/08ca14f84c6597d74dd2d97a1ac11c654c940a388499bf5a081c51bed8631df4.jpg) + +![](images/7630abf7b0765598023e47eacecf3607c3f0dfdc2ebb8bcf62d0fac41b9d6abc.jpg) +Input + +![](images/497093888fc3f303a84d83180981ad284859affdc056a0d8abd189d808097426.jpg) +(1) "My image is too dark, fix it" $\longrightarrow$ (2) 'Apply a tonemap' + +![](images/b59efc2ccaf1b29ca540a899374a2dbde8cf3960b9e693953a1be1ba9f58c1ab.jpg) + +![](images/97d321d57a875344929996d08e4630d2127fed00ff3a9e03ed2858f91a698e6b.jpg) +Fig. 6: Control via instructions. We can prompt multiple instructions (in sequence) to restore and enhance the images. This provides additional control. We show two examples of multiple instructions applied to the "Input" image -from left to right-. +Input (RealSRSet) +Fig.7: Comparison with InstructPix2Pix [4] using the prompt "Remove the noise in this photo". Real-case image from RealSRSet [38]. + +![](images/b4772aa72cdf41c53b2e9426a5b5dccb453f55b02e485e6e045c0a3ffe7356ae.jpg) +InstructIR +InstructPix2Pix #1 InstructPix2Pix #2 + +![](images/7c690650ec75ceffd9c5bc7652b1368a9c9997626cfe1e864746a3e621e9a192.jpg) + +![](images/c0e64c0032ad5b926c3e1826b7d2c2278b8fe69dca2c36af265f62c08030eca0.jpg) + +# 5.1 On the Effectiveness of Instructions + +Thanks to our integration of human instructions, users can control how to enhance the images. We provide examples in Figures 5 and 6, where we show the potential of our method to restore and enhance images in a controllable manner. + +This implies an advancement w.r.t classical (deterministic) image restoration methods. Classical deep restoration methods lead to a unique result, thus, they do not allow to control how the image is processed. We also compare InstructIR with InstructPix2Pix [4] (a diffusion-based generative model) in Figure 7. + +Qualitative Results. We provide diverse qualitative results for several tasks, and we compare with all-in-one and task-specific methods. In Figure 10, we show + +Instruction: "my colors are too off, make it pop so I can use these photos in instagram" + +![](images/f3ad7deaa755c317646cd5f67d5fa8166f280b606740351fb765904cc572b7a1.jpg) + +![](images/715d36189c1551feff2188e486e25f156503ba81d5cb9b15e4035e37c789d4a9.jpg) + +![](images/ad1a1fd835e7d4b8d80b514a23be1881227e3269c969d5515b85af5787881535.jpg) + +![](images/ae72433d5937f559ffab2d268aac0bf187b1113e70afefbb9d5f470060c34dd0.jpg) + +"Increase the brightness of my photo please, is it totoro?" / "my image is too dark, fix it" + +![](images/730bb54bbc219579e5f3b13f1de38d9cd1dfbee411cf2ac96fb0f8381126c42b.jpg) +Fig. 8: Real-world samples of image restoration and enhancement using InstructIR. + +![](images/969ecfbfdb19383e15c0f4ca2f279061cdca9f8f129cfdb964a585ff50bbea38.jpg) + +![](images/b1229dc4382b7055869832deb70100f77f37065fe172bbb3547fb03ee33e7c54.jpg) + +![](images/9da96949b25277e62dd4107bfc1528f7a062e27fc91a0768ebcbb82dbe645381.jpg) + +![](images/9435cfd5a78f9d5f8d9e944ab690553c8e233ece96a8fbcedaa3fed60567f913.jpg) + +![](images/46c4466867627c5e860a5254080a700a8b44d9728240c751fa732e7886c2ac20.jpg) + +![](images/c70505c6af0166b68dec30cc3d0452d7beb4f09816cd16f6fcef2aa52cef4c15.jpg) + +![](images/00bc1e382402d42498c92573b5951324819509bb0fd0f1006861aaac828b2eb0.jpg) + +![](images/44b8bab9807a75e30954c2f56a8c1931275b0bcb5fdcf3f4275f201d21f3f8f4.jpg) + +![](images/be63fd17da14f825879323891e2fc1cc6541d59aee3209696722935c2a5ee5c2.jpg) +Input + +![](images/c3af3f60b8a75090cffc5fb877d38618b0107097d4fc7ff15f557416d3b7772b.jpg) +AirNet [36] + +![](images/6233f7c954b3f97a09df14099a8a66a742b89d1a5d1a143adfee17293deffbd5.jpg) +PromptIR [54] + +![](images/3f1def5a075c6dd824b66ee5514e81b80be6938430cc5366b39aa9652b2643a3.jpg) +InstructIR +Fig. 9: Dehazing comparisons for all-in-one restoration methods on SOTS [35]. + +![](images/ec45eabe49713611f28bcee6f6de5a036e840e05f0fa0f6732172f1dafa03b89.jpg) +Reference + +results on the LOL [74] dataset. In Figure 11, we compare methods on the motion deblurring task using the GoPro [50] dataset. In Figures 9 and 12, we compare with different methods for the dehazing task on SOTS (outdoor) [35]. In Figure 13, we compare with image restoration methods for deraining on Rain100L [17]. Finally, we show denoising results in Figure 14. In this qualitative analysis, we use our single InstructIR-5D model to restore all the images. Limitations As with previous all-in-one methods, our model struggles to process images with more than one degradation (i.e. complex real-world images), or unknown out-of-distribution degradations, yet this is a common limitation among the related methods. However, we believe that these limitations can be surpassed with more realistic training data, and scaling the model's complexity. + +# 6 Conclusion + +We present a novel approach that uses natural human-written instructions to guide the image restoration model. Given a prompt, our multi-task model can recover high-quality images from their degraded counterparts, considering multiple degradations. We achieve state-of-the-art results on several restoration tasks, demonstrating the power and flexibility of instruction guidance. Our results represent a new benchmark for text-guided image restoration. + +![](images/8125aafc4839ebcb4c71b26079513104de1684649f67fede8bc456704dce09db.jpg) +Low-light Image + +![](images/d9be2b4d2d1f65cda746e00c7d31508865d3810f8c91fa6eb6076d1dbec21560.jpg) +NAFNet + +![](images/5c3d96742248411cc2e169d206adad6bd16ecddfe2f43b5cc09bf4f4d081eef3.jpg) +MPRNet + +![](images/d00f684e6d24e1cc588bc98620fe75d5e3d1fa289a4a5597b1407d9a6f908aa3.jpg) +DGUNet + +![](images/ba46adc333e4cd04c14997084a80ba0c318eb90ef4a644dd93d61ab1f3935bc7.jpg) +MIRNetV2 + +![](images/031d26c4dd4bf444fb3111eb40b8cb489799984a83f947355212adbc9876a69e.jpg) +InstructIR + +![](images/46616ef4e8c83a1a172efef1e99d18b9e5ce30b9caa5d06a2f1153b0791fcdd7.jpg) + +![](images/f639bfa8c48a1d8e1a3fbbcf1f54596b496afe186343c5a078c9213f734181b5.jpg) +TAPE + +![](images/946712f5ded19400803b8e125945abd58504658f00abf2107c7346d7dda5b212.jpg) +Restormer + +![](images/5031453b98702aaa860947e198148d8ba4978ee4a37e19c84468a193f34c7db7.jpg) +AirNet +IDR + +AirNet + +![](images/442cfc75e98c6466278325f29be28db9caac2fcd26670f4453a5d15d23366cfa.jpg) +GT + +![](images/6cdcf2e2c02ae92bd12cdce4be545aab9f1002a983c45ca28cbfa4fc834871f2.jpg) +Blurry Image +Fig. 11: Image Deblurring Results. GoPro [50] dataset. + +![](images/dc4d1d40b9fcbb5f946538adce8a4961a32568684a5ad744d08691a87950b4b9.jpg) +NAFNet + +![](images/818281906252ff084602697ddcc1f69d33cfeb657b5b34e1c936993b06906b61.jpg) +MPRNet + +![](images/b28ff142c1596559d1312fae787d60132f830244de99cba50d27912864f75425.jpg) +DGUNet + +![](images/6ddaf809bf7118af0334a9a135487cf824692545ca2a3795eb2599018d5bf463.jpg) +MIRNetV2 + +![](images/9555f65f6a06e4d7532a7edf95901eecea10972e14fe852db4cddc4341590aac.jpg) +Fig. 10: Real Low-light Enhancement Results. LOL [74] testset (748.png). AirNet [36] and IDR [89] are well-known all-in-one restoration methods. NAFNet [7] is equivalent to InstructIR without text conditions (i.e. our image-only backbone). +InstructIR + +![](images/18bca19ee81a0da59440aaa9c0f131d78f9f26838609b69cdfa21eb5684a1916.jpg) +TAPE + +![](images/24df408de42c7c6d838b38dd6a518938fbf8337d48101e1999a6420d6cdcc354.jpg) +Restormer + +![](images/2784727bb426120f76a3bd782012d65c414effab0f4627b2e48824fb6004fc6d.jpg) +AirNet + +![](images/b84b3ee66eea74cfdf05cb063e3944b2ad1060ebd9ec20084c4b1a470b0a6e8c.jpg) +IDR + +![](images/68a9005f7f24e806fe68cea6ebbf48e42f2a0e6aa4d5617c70bc0d40d4011449.jpg) +GT + +![](images/0e0be5640cdba7caa33ace574f795c3f509110dd7306bd8879b57b6d96131929.jpg) +Haze Image +TAPE +Fig. 12: Image Dehazing Results. SOTS [35] outdoor dataset (0150.jpg). + +![](images/41c39169c458ddfe91ce3411ba6aec81d3b2eb5477474c2395723a61a7f410ad.jpg) +NAFNet + +![](images/9b3374cdce6fc4ce038fad4d6cc21f23a865a59b965c6eb10a200630e047791b.jpg) +MPRNet + +![](images/a411721d191f5b171ee7d820db2e35e04ecaa9be30d9ac53e0c4fa74cb0a83b9.jpg) +DGUNet + +![](images/668e91bd950f4a412be7f115edb85eff794294872c104839ee66af610d3f3e3a.jpg) +MIRNetV2 + +![](images/8846fd0590d30a846e9f4b8cbf3943f3c48be3e85cd79c958401cf49135a57b4.jpg) +InstructIR + +![](images/eb25d9f6a35b2302ae0e7b01347306658c0f360aaf810c53ffa994063c3e6e62.jpg) +Restormer + +![](images/9ddf39bc733b26f72261a9b09c42852b39cea34377683b17dd6131c6a3464f19.jpg) +AirNet + +![](images/f227c6a65dc0d6e1e23721ed4446cd12fb45d21a2ab0f254b368b238217465b9.jpg) +IDF + +![](images/7b045258e36478771730b85a0f0cb35d41b66df6eadeea2a0b2fa5a8be4d995a.jpg) +GT + +![](images/9dd58982d68cba36318bbd95577bf45ebd505881e9e9d26081984ebd5e877b1a.jpg) +$\therefore m = \frac{3}{11}$ ; + +![](images/5e708036c87c9a42ff105d210ef45578cfbc3df9654ac0d2de78605cebe12193.jpg) +NAFNet + +![](images/3c8b6a87e36742171897855d274bbe3f36186351476c51be7bb317087b25e4e5.jpg) +MPRNet + +![](images/659c360949b5cdd10f7b01f5ecdf1d1ee8c834e51cbb050b43216e8a54c31d85.jpg) +DGUNet + +![](images/098b484679f2c6b41b81bd05bd50c5ef916bb4ea978023dfcfcb5570b2c024a7.jpg) +MIRNetV2 + +![](images/7fe9ec5d9bc72bb886bdf65e0d75b2d48f10c877cd3e4b93facca106bc5468ef.jpg) +InstructIR + +![](images/60f5f6186f56dfd1afdd6183dba3ea62e7573afd7a373f7ab25fff5165f6d57a.jpg) +Rainy Image +Fig. 13: Image Deraining Results on Rain100L [17] (035.png). + +![](images/02c08d5e562af874a657c0ea006a78a3b903a4bfc55c06b888bcf9d43a55d91e.jpg) +TAPE + +![](images/27c5e77d6415849da93a7b1145cef8754b19a83814ee73537eac0ef4aefd0e75.jpg) +Restormer + +![](images/0388fd22cc593cdc8e884516b5ce1f60e5a1883ab944d65ffa31f9304a0f109c.jpg) +AirNet + +![](images/454ba8c97256230e526d9ad8d7f0fb2b37361fa75e181c3668006c66cc8be812.jpg) +IDR + +![](images/f9073ec7db09e4b08fb8174028b30d443515794d6ae6551784d87336206b6ffb.jpg) +GT + +![](images/c5b2ba6176dfe37e6be5f2c3cf06e6581d67aad0a1b41396d786339dace7444c.jpg) +Fig. 14: Image Denoising Results on BSD68 [45] (0060.png). + +Noise Image + +![](images/292b40798c91c11d14544e9d1d0ef2ba539294365767e3b35668d18bd14bbb62.jpg) + +![](images/0892a6e504b0f6094241e8703d7dbb15dc6eea22ee3943e4ac95a01838107508.jpg) + +![](images/23d85597e72292315b869bb39eb28cf3d9bddecf2d62596d2ced5d1bab8f4c7e.jpg) + +![](images/23359ee7289ec06eb35212328358961fdd6a4507880254dd9f5d2b7ae8e10651.jpg) + +![](images/e25e2fe1fc849443e17bec637366b61d62bdc1811250fdcad65af4220cb804d7.jpg) + +Noise Image + +![](images/e78430966735a45edb4a887cd0ec3405e0c167185bb9aeda3dcbff327f5a686a.jpg) +NAFNet +TAPE + +![](images/4092a20cb176a696593465c27a6ca80f5977fd10859e6522f51f3fe33b992221.jpg) +MPRNet +Restormer + +![](images/c164dae35b527dc63d1bc60de44a3aa0a6088218fc6d1fd81f45771fd287a0e4.jpg) +DGUNet +AirNet + +![](images/2c658dde79342b0c7f889e82357ba98ef9f5fe05cdea5392fd48f4f389415436.jpg) +MIRNetV2 +IDR + +![](images/8851ecd7763a336e7b162facddfa9332400f1ff5aa192253f535c53dd02ca242.jpg) +InstructIR +GT + +# References + +1. Agustsson, E., Timofte, R.: NTIRE 2017 challenge on single image superresolution: Dataset and study. In: CVPR Workshops (2017) 10 +2. Arbelaez, P., Maire, M., Fowlkes, C., Malik, J.: Contour detection and hierarchical image segmentation. TPAMI (2011) 8 +3. Bai, Y., Wang, C., Xie, S., Dong, C., Yuan, C., Wang, Z.: Textir: A simple framework for text-based editable image restoration. CoRR abs/2302.14736 (2023). https://doi.org/10.48550/ARXIV.2302.14736, https://doi.org/10.48550/arXiv.2302.14736 5 +4. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023. pp. 18392-18402. IEEE (2023). https://doi.org/10.1109/CVPR52729.2023.01764, https://doi.org/10.1109/CVPR52729.2023.01764 2, 3, 4, 5, 12 +5. Bychkovsky, V., Paris, S., Chan, E., Durand, F.: Learning photographic global tonal adjustment with a database of input / output image pairs. In: The Twenty-Fourth IEEE Conference on Computer Vision and Pattern Recognition (2011) 8, 10 +6. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR (2021) 11 +7. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV (2022) 2, 3, 6, 7, 9, 14 +8. Chen, L., Lu, X., Zhang, J., Chu, X., Chen, C.: Hinet: Half instance normalization network for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 182-192 (2021) 9 +9. Chen, Y.S., Wang, Y.C., Kao, M.H., Chuang, Y.Y.: Deep photo enhancer: Unpaired learning for image enhancement from photographs with gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6306-6314 (2018) 10 +10. Devlin, J., Chang, M., Lee, K., Toutanova, K.: BERT: pre-training of deep bidirectional transformers for language understanding. In: Burstein, J., Doran, C., Solorio, T. (eds.) Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers). pp. 4171-4186. Association for Computational Linguistics (2019). https://doi.org/10.18653/V1/N19-1423, https://doi.org/10.18653/v1/n19-1423 7 +11. Ding, C., Lu, Z., Wang, S., Cheng, R., Boddeti, V.N.: Mitigating task interference in multi-task learning via explicit task routing with non-learnable primitives. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7756-7765 (2023) 6, 7 +12. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. TPAMI (2015) 10 +13. Dong, H., Pan, J., Xiang, L., Hu, Z., Zhang, X., Wang, F., Yang, M.H.: Multi-scale boosted dehazing network with dense feature fusion. In: CVPR (2020) 2 +14. Dong, W., Zhang, L., Shi, G., Wu, X.: Image deblurring and super-resolution by adaptive sparse domain selection and adaptive regularization. TIP (2011) 2 +15. Dong, Y., Liu, Y., Zhang, H., Chen, S., Qiao, Y.: Fd-gan: Generative adversarial networks with fusion-discriminator for single image dehazing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 10729-10736 (2020) 9 + +16. Elad, M., Feuer, A.: Restoration of a single superresolution image from several blurred, noisy, and undersampled measured images. IEEE transactions on image processing 6(12), 1646-1658 (1997) 2 +17. Fan, Q., Chen, D., Yuan, L., Hua, G., Yu, N., Chen, B.: A general decoupled learning framework for parameterized image operators. IEEE transactions on pattern analysis and machine intelligence 43(1), 33-47 (2019) 9, 13, 14 +18. Fu, X., Zeng, D., Huang, Y., Liao, Y., Ding, X., Paisley, J.: A fusion-based enhancing method for weakly illuminated images 129, 82-96 (2016) 11 +19. Fu, X., Zeng, D., Huang, Y., Zhang, X.P., Ding, X.: A weighted variational model for simultaneous reflectance and illumination estimation. In: CVPR (2016) 11 +20. Gao, H., Tao, X., Shen, X., Jia, J.: Dynamic scene deblurring with parameter selective sharing and nested skip connections. In: CVPR. pp. 3848-3856 (2019) 9 +21. Gharbi, M., Chen, J., Barron, J.T., Hasinoff, S.W., Durand, F.: Deep bilateral learning for real-time image enhancement. ACM Transactions on Graphics (TOG) 36(4), 1-12 (2017) 10 +22. Guo, X., Li, Y., Ling, H.: Lime: Low-light image enhancement via illumination map estimation. IEEE TIP 26(2), 982-993 (2016) 11 +23. Hao, S., Han, X., Guo, Y., Xu, X., Wang, M.: Low-light image enhancement with semi-decoupled decomposition. IEEE TMM 22(12), 3025-3038 (2020) 11 +24. He, J., Dong, C., Qiao, Y.: Modulating image restoration with continual levels via adaptive feature modification layers (2019), https://arxiv.org/abs/1904.081186 +25. He, K., Sun, J., Tang, X.: Single image haze removal using dark channel prior. TPAMI (2010) 2 +26. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022) 3 +27. Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al.: Searching for mobilenetv3. In: ICCV (2019) 3 +28. Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5197-5206 (2015) 8 +29. Jiang, Y., Gong, X., Liu, D., Cheng, Y., Fang, C., Shen, X., Yang, J., Zhou, P., Wang, Z.: Enlightengan: Deep light enhancement without paired supervision. IEEE TIP 30, 2340-249 (2021) 11 +30. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023) 3 +31. Kim, K.I., Kwon, Y.: Single-image super-resolution using sparse regression and natural image prior. TPAMI (2010) 2 +32. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv:1412.6980 (2014) 8 +33. Kopf, J., Neubert, B., Chen, B., Cohen, M., Cohen-Or, D., Deussen, O., Uyttendaele, M., Lischinski, D.: Deep photo: Model-based photograph enhancement and viewing. ACM TOG (2008) 2 +34. Lei, X., Fei, Z., Zhou, W., Zhou, H., Fei, M.: Low-light image enhancement using the cell vibration model. IEEE TMM pp. 1-1 (2022) 11 +35. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing 28(1), 492-505 (2018) 8, 9, 13, 14 + +36. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR. pp. 17452-17462 (June 2022) 2, 3, 7, 8, 9, 13, 14 +37. Li, J., Li, J., Fang, F., Li, F., Zhang, G.: Luminance-aware pyramid network for low-light image enhancement. IEEE TMM 23, 3153-3165 (2020) 11 +38. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: SwinIR: Image restoration using swin transformer. In: ICCV Workshops (2021) 2, 9, 12 +39. Liu, L., Xie, L., Zhang, X., Yuan, S., Chen, X., Zhou, W., Li, H., Tian, Q.: Tape: Task-agnostic prior embedding for image restoration. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XVIII. pp. 447–464. Springer (2022) 9 +40. Liu, R., Ma, L., Zhang, J., Fan, X., Luo, Z.: Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In: CVPR (2021) 11 +41. Liu, Y., Liu, A., Gu, J., Zhang, Z., Wu, W., Qiao, Y., Dong, C.: Discovering distinctive "semantics" in super-resolution networks (2022), https://arxiv.org/abs/2108.00406 3 +42. Ma, J., Cheng, T., Wang, G., Zhang, Q., Wang, X., Zhang, L.: Prores: Exploring degradation-aware visual prompt for universal image restoration. arXiv preprint arXiv:2306.13653 (2023) 2, 3 +43. Ma, K., Duanmu, Z., Wu, Q., Wang, Z., Yong, H., Li, H., Zhang, L.: Waterloo exploration database: New challenges for image quality assessment models. TIP (2016) 8 +44. Ma, L., Ma, T., Liu, R., Fan, X., Luo, Z.: Toward fast, flexible, and robust low-light image enhancement. In: CVPR (2022) 11 +45. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: ICCV (2001) 8, 9, 14 +46. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073 (2021) 3 +47. Michaeli, T., Irani, M.: Nonparametric blind super-resolution. In: ICCV (2013) 2 +48. Moran, S., Marza, P., McDonagh, S., Parisot, S., Slabaugh, G.: Deeplpf: Deep local parametric filters for image enhancement. In: CVPR (2020) 11 +49. Mou, C., Wang, Q., Zhang, J.: Deep generalized unfolding networks for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17399-17410 (2022) 9 +50. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: CVPR (2017) 8, 9, 13, 14 +51. Nah, S., Son, S., Lee, J., Lee, K.M.: Clean images are hard to reblur: Exploiting the ill-posed inverse task for dynamic scene deblurring. In: ICLR (2022) 2 +52. Nguyen, N., Milanfar, P., Golub, G.: Efficient generalized cross-validation with applications to parametric image restoration and resolution enhancement. IEEE Transactions on image processing 10(9), 1299-1308 (2001) 2 +53. Park, D., Lee, B.H., Chun, S.Y.: All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5815-5824. IEEE (2023) 2, 3, 7 +54. Potlapalli, V., Zamir, S.W., Khan, S., Khan, F.S.: Prompt: Prompting for all-in-one blind image restoration. arXiv preprint arXiv:2306.13090 (2023) 2, 3, 8, 9, 13 + +55. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event. Proceedings of Machine Learning Research, vol. 139, pp. 8748-8763. PMLR (2021), http://proceedings.mlr.press/v139/radford21a.html 5 +56. Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert-networks. In: Inui, K., Jiang, J., Ng, V., Wan, X. (eds.) Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019. pp. 3980-3990. Association for Computational Linguistics (2019). https://doi.org/10.18653/V1/D19-1410, https://doi.org/10.18653/v1/D19-1410 5 +57. Ren, C., He, X., Wang, C., Zhao, Z.: Adaptive consistency prior based deep network for image denoising. In: CVPR (2021) 2 +58. Ren, W., Pan, J., Zhang, H., Cao, X., Yang, M.H.: Single image dehazing via multi-scale convolutional neural networks with holistic edges. IJCV (2020) 2 +59. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 10674-10685. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.01042, https://doi.org/10.1109/CVPR52688.2022.01042 5 +60. Ronneberger, O., Fischer, P., Brox, T.: U-Net: convolutional networks for biomedical image segmentation. In: MICCAI (2015) 6 +61. Rosenbaum, C., Klinger, T., Riemer, M.: Routing networks: Adaptive selection of non-linear functions for multi-task learning. arXiv preprint arXiv:1711.01239 (2017) 6 +62. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023) 3 +63. Strezoski, G., Noord, N.v., Worring, M.: Many task learning with task routing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1375-1384 (2019) 6, 7, 8 +64. Tian, C., Xu, Y., Zuo, W.: Image denoising using deep cnn with batch renormalization. Neural Networks (2020) 9 +65. Timofte, R., De Smet, V., Van Gool, L.: Anchored neighborhood regression for fast example-based super-resolution. In: ICCV (2013) 2 +66. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A., Li, Y.: MAXIM: Multi-axis MLP for image processing. In: CVPR. pp. 5769-5780 (2022) 2, 8 +67. Valanarasu, J.M.J., Yasarla, R., Patel, V.M.: Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In: CVPR. pp. 2353-2363 (2022) 3, 9 +68. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 2 +69. Wang, R., Zhang, Q., Fu, C.W., Shen, X., Zheng, W.S., Jia, J.: Underexposed photo enhancement using deep illumination estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6849-6857 (2019) 10 + +70. Wang, S., Zheng, J., Hu, H.M., Li, B.: Naturalness preserved enhancement algorithm for non-uniform illumination images. IEEE TIP 22(9), 3538-3548 (2013) 11 +71. Wang, X., Girshick, R., Gupta, A., He, K.: Non-local neural networks. In: CVPR (2018) 3 +72. Wang, Y., Liu, Z., Liu, J., Xu, S., Liu, S.: Low-light image enhancement with illumination-aware gamma correction and complete image modelling network. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13128-13137 (2023) 11 +73. Wang, Z., Cun, X., Bao, J., Liu, J.: Uformer: A general u-shaped transformer for image restoration. arXiv:2106.03106 (2021) 2, 11 +74. Wei, C., Wang, W., Yang, W., Liu, J.: Deep retina decomposition for low-light enhancement. In: British Machine Vision Conference (2018) 8, 9, 11, 13, 14 +75. Wu, W., Weng, J., Zhang, P., Wang, X., Yang, W., Jiang, J.: Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement. In: CVPR (2022) 11 +76. Xiao, S., Liu, Z., Zhang, P., Muennighof, N.: C-pack: Packaged resources to advance general chinese embedding. CoRR abs/2309.07597 (2023). https://doi.org/10.48550/ARXIV.2309.07597, https://doi.org/10.48550/arXiv.2309.07597 +77. Xu, K., Yang, X., Yin, B., Lau, R.W.: Learning to restore low-light images via decomposition-and-enhancement. In: CVPR (2020) 11 +78. Yang, F., Yang, H., Fu, J., Lu, H., Guo, B.: Learning texture transformer network for image super-resolution. In: CVPR (2020) 8, 9 +79. Yang, W., Wang, S., Fang, Y., Wang, Y., Liu, J.: Band representation-based semi-supervised low-light image enhancement: bridging the gap between signal fidelity and perceptual quality. IEEE TIP 30, 3461-3473 (2021) 11 +80. Yang, W., Wang, W., Huang, H., Wang, S., Liu, J.: Sparse gradient regularized deep retina network for robust low-light image enhancement. IEEE TIP 30, 2072-2086 (2021) 11 +81. Yao, M., Xu, R., Guan, Y., Huang, J., Xiong, Z.: Neural degradation representation learning for all-in-one image restoration. arXiv preprint arXiv:2310.12848 (2023) 3 +82. Yu, W., Luo, M., Zhou, P., Si, C., Zhou, Y., Wang, X., Feng, J., Yan, S.: Metaformer is actually what you need for vision. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10819-10829 (2022) 3 +83. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022) 2, 3, 9 +84. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Learning enriched features for real image restoration and enhancement. In: ECCV (2020) 9, 11 +85. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR (2021) 9 +86. Zeng, H., Cai, J., Li, L., Cao, Z., Zhang, L.: Learning image-adaptive 3d lookup tables for high performance photo enhancement in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(4), 2058-2073 (2020) 10 +87. Zhang, C., Zhu, Y., Yan, Q., Sun, J., Zhang, Y.: All-in-one multi-degradation image restoration network via hierarchical degradation representation. arXiv preprint arXiv:2308.03021 (2023) 3 + +88. Zhang, C., Zhu, Y., Yan, Q., Sun, J., Zhang, Y.: All-in-one multi-degradation image restoration network via hierarchical degradation representation. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 2285-2293 (2023) 3 +89. Zhang, J., Huang, J., Yao, M., Yang, Z., Yu, H., Zhou, M., Zhao, F.: Ingredient-oriented multi-degradation learning for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5825-5835 (2023) 2, 3, 8, 9, 14 +90. Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing 26(7), 3142-3155 (2017) 2 +91. Zhang, K., Zuo, W., Gu, S., Zhang, L.: Learning deep CNN denoiser prior for image restoration. In: CVPR (2017) 2 +92. Zhang, K., Luo, W., Zhong, Y., Ma, L., Stenger, B., Liu, W., Li, H.: Deblurring by realistic blurring. In: CVPR. pp. 2737-2746 (2020) 2 +93. Zhang, Y., Zhang, J., Guo, X.: Kindling the darkness: A practical low-light image enhancer. In: ACM MM (2019) 11 \ No newline at end of file diff --git a/2024/InstructIR_ High-Quality Image Restoration Following Human Instructions/images.zip b/2024/InstructIR_ High-Quality Image Restoration Following Human Instructions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a291e3dc6569e69d64518b0c5b54462c22044fa9 --- /dev/null +++ b/2024/InstructIR_ High-Quality Image Restoration Following Human Instructions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd0ff361a740d7e6765413b05466637fda80f83d667a226eccf8c9934e07093f +size 1029047 diff --git a/2024/InstructIR_ High-Quality Image Restoration Following Human Instructions/layout.json b/2024/InstructIR_ High-Quality Image Restoration Following Human Instructions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..52a8f1dc572372263fac0b6d065ea7854f5259f7 --- /dev/null +++ b/2024/InstructIR_ High-Quality Image Restoration Following Human Instructions/layout.json @@ -0,0 +1,17135 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 111, + 460, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 111, + 460, + 148 + ], + "spans": [ + { + "bbox": [ + 153, + 111, + 460, + 148 + ], + "type": "text", + "content": "InstructIR: High-Quality Image Restoration Following Human Instructions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 173, + 167, + 439, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 167, + 439, + 180 + ], + "spans": [ + { + "bbox": [ + 173, + 167, + 439, + 180 + ], + "type": "text", + "content": "Marcos V. Conde1,2, Gregor Geigle1, and Radu Timofte1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 173, + 189, + 440, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 189, + 440, + 201 + ], + "spans": [ + { + "bbox": [ + 173, + 189, + 440, + 201 + ], + "type": "text", + "content": "1 Computer Vision Lab, CAIDAS & IFI, University of Würzburg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 254, + 201, + 359, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 201, + 359, + 212 + ], + "spans": [ + { + "bbox": [ + 254, + 201, + 359, + 212 + ], + "type": "text", + "content": "2 Sony PlayStation, FTG" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 220, + 213, + 392, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 213, + 392, + 223 + ], + "spans": [ + { + "bbox": [ + 220, + 213, + 392, + 223 + ], + "type": "text", + "content": "https://github.com/mv-lab/InstructIR" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 134, + 240, + 216, + 306 + ], + "blocks": [ + { + "bbox": [ + 134, + 240, + 216, + 306 + ], + "lines": [ + { + "bbox": [ + 134, + 240, + 216, + 306 + ], + "spans": [ + { + "bbox": [ + 134, + 240, + 216, + 306 + ], + "type": "image", + "image_path": "f042f64cdc8d87bede4f89391118e4f2275a1a2bc840ec5a90bb5597a16bba4d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 307, + 207, + 321 + ], + "lines": [ + { + "bbox": [ + 140, + 307, + 207, + 321 + ], + "spans": [ + { + "bbox": [ + 140, + 307, + 207, + 321 + ], + "type": "text", + "content": "Can you stabilize this pie? I was running..." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 216, + 240, + 286, + 306 + ], + "blocks": [ + { + "bbox": [ + 216, + 240, + 286, + 306 + ], + "lines": [ + { + "bbox": [ + 216, + 240, + 286, + 306 + ], + "spans": [ + { + "bbox": [ + 216, + 240, + 286, + 306 + ], + "type": "image", + "image_path": "9dd094cb6c5f041e6306f3a5393cf19a191ec10a2123ccb53550381e6f70f876.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 307, + 282, + 321 + ], + "lines": [ + { + "bbox": [ + 220, + 307, + 282, + 321 + ], + "spans": [ + { + "bbox": [ + 220, + 307, + 282, + 321 + ], + "type": "text", + "content": "My image has tiny dots, doesn't look great, fix it" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 287, + 240, + 351, + 306 + ], + "blocks": [ + { + "bbox": [ + 287, + 240, + 351, + 306 + ], + "lines": [ + { + "bbox": [ + 287, + 240, + 351, + 306 + ], + "spans": [ + { + "bbox": [ + 287, + 240, + 351, + 306 + ], + "type": "image", + "image_path": "2d2022e9d107b43ad89835ee2cca6ba1453ba164985093492f391320b734ce7c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 307, + 346, + 321 + ], + "lines": [ + { + "bbox": [ + 291, + 307, + 346, + 321 + ], + "spans": [ + { + "bbox": [ + 291, + 307, + 346, + 321 + ], + "type": "text", + "content": "My photo is too dark, I cannot see anything" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 351, + 240, + 411, + 306 + ], + "blocks": [ + { + "bbox": [ + 351, + 240, + 411, + 306 + ], + "lines": [ + { + "bbox": [ + 351, + 240, + 411, + 306 + ], + "spans": [ + { + "bbox": [ + 351, + 240, + 411, + 306 + ], + "type": "image", + "image_path": "2bf86199a634bc594aae62f760f80cb338a58b5d2eafb3e6a9711a26a5465b25.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 307, + 411, + 321 + ], + "lines": [ + { + "bbox": [ + 351, + 307, + 411, + 321 + ], + "spans": [ + { + "bbox": [ + 351, + 307, + 411, + 321 + ], + "type": "text", + "content": "Retouch my picture like a photographer" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 412, + 240, + 480, + 306 + ], + "blocks": [ + { + "bbox": [ + 412, + 240, + 480, + 306 + ], + "lines": [ + { + "bbox": [ + 412, + 240, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 412, + 240, + 480, + 306 + ], + "type": "image", + "image_path": "38d7d6a663ed5b99ffee6dfc67dad3b49e4ecb144dd689daacad416089543c08.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 307, + 478, + 321 + ], + "lines": [ + { + "bbox": [ + 414, + 307, + 478, + 321 + ], + "spans": [ + { + "bbox": [ + 414, + 307, + 478, + 321 + ], + "type": "text", + "content": "It is a foggy day, can you make it clearer?" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 133, + 322, + 216, + 388 + ], + "blocks": [ + { + "bbox": [ + 133, + 322, + 216, + 388 + ], + "lines": [ + { + "bbox": [ + 133, + 322, + 216, + 388 + ], + "spans": [ + { + "bbox": [ + 133, + 322, + 216, + 388 + ], + "type": "image", + "image_path": "0e811c9713c52e4d60a40020c5aaa1e0c6cd94acd1ed4ed44710756e7e299928.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 392, + 482, + 437 + ], + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 437 + ], + "type": "text", + "content": "Fig. 1: Given an image and a prompt for how to improve that image, our all-in-one restoration model corrects the image considering the human instruction. InstructIR, can tackle various types and levels of degradation, and it is able to generalize in some real-world scenarios (last three images, from left to right)." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 216, + 322, + 287, + 388 + ], + "blocks": [ + { + "bbox": [ + 216, + 322, + 287, + 388 + ], + "lines": [ + { + "bbox": [ + 216, + 322, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 216, + 322, + 287, + 388 + ], + "type": "image", + "image_path": "8614232ac2764f401353daa5357f0876ac53be1e56804dca5239d99d4fb733b0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 287, + 322, + 351, + 388 + ], + "blocks": [ + { + "bbox": [ + 287, + 322, + 351, + 388 + ], + "lines": [ + { + "bbox": [ + 287, + 322, + 351, + 388 + ], + "spans": [ + { + "bbox": [ + 287, + 322, + 351, + 388 + ], + "type": "image", + "image_path": "56ee4c73e5aa2082c5add852c5d29587a1cec94d89bc2677ab1f5e06aac2d923.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 351, + 322, + 411, + 388 + ], + "blocks": [ + { + "bbox": [ + 351, + 322, + 411, + 388 + ], + "lines": [ + { + "bbox": [ + 351, + 322, + 411, + 388 + ], + "spans": [ + { + "bbox": [ + 351, + 322, + 411, + 388 + ], + "type": "image", + "image_path": "c76e60981c7a373272d75a7079855c5531ee8e13949234ff9e266fbbd61e2105.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 412, + 322, + 480, + 388 + ], + "blocks": [ + { + "bbox": [ + 412, + 322, + 480, + 388 + ], + "lines": [ + { + "bbox": [ + 412, + 322, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 412, + 322, + 480, + 388 + ], + "type": "image", + "image_path": "f033464962ab159b067523ee26c82659a35232c6d93325dac3b5711cd6c7db66.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 159, + 450, + 453, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 450, + 453, + 605 + ], + "spans": [ + { + "bbox": [ + 159, + 450, + 453, + 605 + ], + "type": "text", + "content": "Abstract. Image restoration is a fundamental problem that involves recovering a high-quality clean image from its degraded observation. All-In-One image restoration models can effectively restore images from various types and levels of degradation using degradation-specific information as prompts to guide the restoration model. In this work, we present the first approach that uses human-written instructions to guide the image restoration model. Given natural language prompts, our model can recover high-quality images from their degraded counterparts, considering multiple degradation types. Our method, InstructIR, achieves state-of-the-art results on several restoration tasks including image denoising, deraining, deblurring, dehazing, and (low-light) image enhancement. InstructIR improves " + }, + { + "bbox": [ + 159, + 450, + 453, + 605 + ], + "type": "inline_equation", + "content": "+1\\mathrm{dB}" + }, + { + "bbox": [ + 159, + 450, + 453, + 605 + ], + "type": "text", + "content": " over previous all-in-one restoration methods. Moreover, our dataset and results represent a novel benchmark for new research on text-guided image restoration and enhancement." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 620, + 230, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 230, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 230, + 632 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Images often contain unpleasant effects such as noise, motion blur, haze, and low dynamic range. Such effects are commonly known in low-level computer" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "vision as degradations. These can result from camera limitations or challenging environmental conditions e.g. low light." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 479, + 175 + ], + "type": "text", + "content": "Image restoration aims to recover a high-quality image from its degraded counterpart. This is a complex inverse problem since multiple different solutions can exist for restoring any given image [13, 16, 38, 52, 90, 91]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 176, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 319 + ], + "type": "text", + "content": "Some methods focus on specific degradations, for instance reducing noise (denoising) [57, 90, 91], removing blur (deblurring) [51, 92], or clearing haze (dehazing) [13, 58]. Such methods are effective for their specific task, yet they do not generalize well to other types of degradation. Other approaches use a general neural network for diverse tasks [7, 66, 73, 83], yet training the neural network for each specific task independently. Since using a separate model for each possible degradation is resource-intensive, recent approaches propose All-in-One restoration models [36, 53, 54, 89]. These approaches use a single deep blind restoration model considering multiple degradation types and levels. Contemporary works such as PromptIR [54] or ProRes [42] utilize a unified model for blind image restoration using learned guidance vectors, also known as \"prompt embeddings\", in contrast to raw user prompts in text form, which we use in this work." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 481, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 415 + ], + "type": "text", + "content": "In parallel, recent works such as InstructPix2Pix [4] show the potential of using text prompts to guide image generation and editing models. However, this method (or recent alternatives) do not tackle inverse problems. Inspired by these works, we argue that text guidance can help to guide blind restoration models better than the image-based degradation classification used in previous works [36,53,89]. Users generally have an idea about what has to be fixed (though they might lack domain-specific vocabulary) so we can use this information to guide the model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "text", + "content": "Contributions We propose the first approach that utilizes real human-written instructions to solve multi-task image restoration. Our comprehensive experiments demonstrate the potential of using text guidance for image restoration and enhancement by achieving state-of-the-art performance on various image restoration tasks, including image denoising, deraining, deblurring, dehazing, and low-light image enhancement. Our model, InstructIR, is able to generalize to restoring images using complex human-written instructions. Moreover, our single all-in-one model covers more tasks than many previous works. We show diverse restoration samples of our method in Figure 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 545, + 237, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 545, + 237, + 558 + ], + "spans": [ + { + "bbox": [ + 131, + 545, + 237, + 558 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Image Restoration. Recent deep learning methods [13, 38, 51, 57, 66, 83] have shown consistently better results compared to traditional techniques for blind image restoration [14,25,31,33,47,65]. The proposed neural networks are based on convolutional neural networks (CNNs) and Transformers [68] (or related attention mechanisms). We focus on general-purpose restoration models [7,38,73,83]. For example, SwinIR [38], MAXIM [66] and Uformer [73]. These models can be trained -independently- for diverse tasks such as denoising, deraining or deblurring. Their ability to capture local and global feature interactions, and enhance" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "text", + "content": "Conde, Marcos et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "them, allows the models to achieve great performance consistently across different tasks. For instance, Restormer [83] uses non-local blocks [71] to capture complex features across the image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 224 + ], + "type": "text", + "content": "NAFNet [7] is an efficient alternative to complex transformer-based methods. The model uses simplified channel attention, and gating as an alternative to non-linear activations. The building block (NAFBlock) follows a simple metaformer [82] architecture with efficient inverted residual blocks [27]. In this work, we build our InstructIR model using NAFNet as backbone, due to its efficient and simple design, and high performance in several restoration tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 232, + 480, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 480, + 293 + ], + "type": "text", + "content": "All-in-One Image Restoration. Single degradation (or single task) restoration methods are well-studied, however, their real-world applications are limited due to the required resources i.e. allocating different models, and selecting the adequate model on demand. Moreover, images rarely present a single degradation, for instance, noise and blur are almost ubiquitous in any image capture." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 293, + 480, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 480, + 341 + ], + "type": "text", + "content": "All-in-One (also known as multi-degradation or multi-task) image restoration is emerging as a new research field in low-level computer vision [36,42,53,54,67, 81,87,88]. These approaches use a single deep blind restoration model to tackle different degradation types and levels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 342, + 480, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 480, + 414 + ], + "type": "text", + "content": "We use as reference AirNet [36], IDR [89] and ADMS [53]. We also consider the contemporary work PromptIR [54]. The methods use different techniques to guide the blind model in the restoration process. For instance, an auxiliary model for degradation classification [36, 53], or multi-dimensional guidance vectors (also known as \"prompts\") [41, 42, 54] that help the model to discriminate the different types of degradation in the image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 422, + 480, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 480, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 480, + 529 + ], + "type": "text", + "content": "Text-guided Image Manipulation. In recent years, multiple methods have been proposed for text-to-image generation and text-based image editing works [4, 26, 30, 46, 62]. These models use text prompts to describe images or actions, and powerful diffusion-based models for generating the corresponding images. Our main reference is InstructPix2Pix [4], this method enables editing from instructions that tell the model what action to perform, as opposed to text labels, captions or descriptions of the input or output images. Therefore, the user can transmit what to do in natural written text, without requiring to provide further image descriptions or sample reference images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 552, + 398, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 398, + 566 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 398, + 566 + ], + "type": "text", + "content": "3 Image Restoration Following Instructions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "type": "text", + "content": "We treat instruction-based image restoration as a supervised learning problem similar to previous works [4]. First, we generate over 10000 prompts using GPT-4 based on our own sample instructions. We explain the creation of the prompt dataset in Sec. 3.1. We then build a large paired training dataset of prompts and degraded/clean images. Finally, we train our InstructIR model, and we evaluate it on a wide variety of instructions including real human-written prompts. We explain our text encoder in Sec 3.2, and our complete model in Sec. 3.3." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "InstructIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 328, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 328, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 328, + 129 + ], + "type": "text", + "content": "3.1 Generating Prompts for Training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 221 + ], + "type": "text", + "content": "Why instructions? Inspired by InstructPix2Pix [4], we adopt human written instructions as the mechanism of control for our model. There is no need for the user to provide additional information, such as example clean images, or descriptions of the visual content. Instructions offer a clear and expressive way to interact, enabling users to pinpoint the unpleasant effects (degradations) in the images. We also consider the language complexity, from ambiguous instructions (e.g. \"fix my image\") to precise instructions (e.g. \"remove the noise\")." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 226, + 299, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 299, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 299, + 501 + ], + "type": "text", + "content": "Handling free-form user prompts rather than fixed degradation-specific prompts increases the usability of our model for laypeople who lack domain expertise. We thus want our model to be capable of understanding diverse prompts posed by users \"in-the-wild\" e.g. kids, adults, or photographers. To this end, we use a large language model (i.e., GPT-4) to create diverse requests that might be asked by users for the different degradations types. We then filter those generated prompts to remove ambiguous or unclear prompts (e.g., \"Make the image cleaner\", \"improve this image\"). Our final instructions set contains over 10000 different prompts in total, for 7 different tasks. We display some examples in Table 1. As we show in Figure 2 the prompts are sampled randomly depending on the input degradation." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 304, + 270, + 481, + 492 + ], + "blocks": [ + { + "bbox": [ + 304, + 236, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 304, + 236, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 482, + 269 + ], + "type": "text", + "content": "Table 1: Examples of our curated GPT4-generated and real user prompts with varying language and domain expertise." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 304, + 270, + 481, + 492 + ], + "lines": [ + { + "bbox": [ + 304, + 270, + 481, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 481, + 492 + ], + "type": "table", + "html": "
Degradation Prompts
DenoisingCan you clean the dots from my image? Fix the grainy parts of this photo Remove the noise from my picture
DeblurringCan you reduce the movement in the image? My picture's not sharp, fix it Deblur my picture, it's too fuzzy
DehazingCan you make this picture clearer? Help, my picture is all cloudy Remove the fog from my photo
DerainingI want my photo to be clear, not rainy Clear the rain from my picture Remove the raindrops from my photo
Super-Res.Make my photo bigger and better Add details to this image Increase the resolution of this photo
Low-lightThe photo is too dark, improve exposure Increase the illumination in this shot My shot has very low dynamic range
EnhancementMake it pop! Adjust the color balance for a natural look Apply a cinematic color grade to the photo
GeneralFix my image please make the image look better
", + "image_path": "dfd9b59d43050c5d32fce70e6bf35ef504cc36c4459fc34524bcb7eb23997a89.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 133, + 522, + 482, + 613 + ], + "blocks": [ + { + "bbox": [ + 133, + 522, + 482, + 613 + ], + "lines": [ + { + "bbox": [ + 133, + 522, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 133, + 522, + 482, + 613 + ], + "type": "image", + "image_path": "9bbcf64a2b9c6ecf65d1245798bab7130f8c4a344ccf15552c000cbef1e1a42b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 620, + 482, + 665 + ], + "lines": [ + { + "bbox": [ + 130, + 620, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 620, + 482, + 665 + ], + "type": "text", + "content": "Fig. 2: We train our blind image restoration models using common image datasets, and prompts generated using GPT-4, note that this is (self-)supervised learning. At inference time, our model generalizes to human-written instructions and restores (or enhances) the images." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "text", + "content": "Conde, Marcos et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 229, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 229, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 229, + 126 + ], + "type": "text", + "content": "3.2 Text Encoder" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 282 + ], + "type": "text", + "content": "The Choice of the Text Encoder. A text encoder maps the user prompt to a fixed-size vector representation (a text embedding). The related methods for text-based image generation [59] and manipulation [3,4] often use the text encoder of a CLIP model [55] to encode user prompts as CLIP excels in visual prompts. However, user prompts for degradation contain, in general, little to no visual content (e.g. the use describes the degradation, not the image itself). We opt, instead, to use a pure text-based sentence encoder [56], that is, a smaller model trained to encode sentences in a semantically meaningful embedding space. Sentence encoders - pre-trained with millions of examples - are compact and fast in comparison to CLIP, while being able to encode the semantics of diverse user prompts. For instance, we use the BGE-micro-v2 sentence transformer. We compare text encoders in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 285, + 482, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 346 + ], + "type": "text", + "content": "Fine-tuning the Text Encoder. We want to adapt the text encoder E for the restoration task to better encode the required information for the restoration model. Training the full text encoder is likely to lead to overfitting on our small training set and lead to loss of generalization. Instead, we freeze the text encoder and train a projection head on top:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 263, + 354, + 480, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 354, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 263, + 354, + 480, + 366 + ], + "type": "interline_equation", + "content": "\\mathbf {e} = \\operatorname {n o r m} (\\mathbf {W} \\cdot \\mathrm {E} (t)) \\tag {1}", + "image_path": "74ce76a9ee5e02dc7d7c8ae0aeb4dfd9070cdbf763c122920b1ad86f1f9d6799.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 373, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 410 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 373, + 482, + 410 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 373, + 482, + 410 + ], + "type": "text", + "content": " is the text, " + }, + { + "bbox": [ + 130, + 373, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\mathrm{E}(t)" + }, + { + "bbox": [ + 130, + 373, + 482, + 410 + ], + "type": "text", + "content": " represents the raw text embedding, " + }, + { + "bbox": [ + 130, + 373, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{d_t \\times d_v}" + }, + { + "bbox": [ + 130, + 373, + 482, + 410 + ], + "type": "text", + "content": " is a learned projection from the text dimension " + }, + { + "bbox": [ + 130, + 373, + 482, + 410 + ], + "type": "inline_equation", + "content": "(d_t)" + }, + { + "bbox": [ + 130, + 373, + 482, + 410 + ], + "type": "text", + "content": " to the input dimension for the restoration model " + }, + { + "bbox": [ + 130, + 373, + 482, + 410 + ], + "type": "inline_equation", + "content": "(d_v)" + }, + { + "bbox": [ + 130, + 373, + 482, + 410 + ], + "type": "text", + "content": ", and norm is the 12-norm." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 411, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 542 + ], + "type": "text", + "content": "Figure 3 shows that while the text encoder is capable out-of-the-box to cluster instructions to some extent (Figure 3a), our trained projection yields greatly improved clusters (Figure 3b). We distinguish clearly the clusters for deraining, denoising, dehazing, deblurring, and low-light image enhancement. The instructions for such tasks or degradations are very characteristic. Furthermore, we can appreciate that \"super-res\" and \"enhancement\" tasks are quite spread and between the previous ones, which matches the language logic. For instance \"add details to this image\" could be used for enhancement, deblurring, or denoising. In our experiments, " + }, + { + "bbox": [ + 130, + 411, + 482, + 542 + ], + "type": "inline_equation", + "content": "d_{t} = 384" + }, + { + "bbox": [ + 130, + 411, + 482, + 542 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 411, + 482, + 542 + ], + "type": "inline_equation", + "content": "d_{v} = 256" + }, + { + "bbox": [ + 130, + 411, + 482, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 411, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 130, + 411, + 482, + 542 + ], + "type": "text", + "content": " is a linear layer. The representation " + }, + { + "bbox": [ + 130, + 411, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 130, + 411, + 482, + 542 + ], + "type": "text", + "content": " from the text encoder is shared across the blocks, and each block has a trainable projection " + }, + { + "bbox": [ + 130, + 411, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 130, + 411, + 482, + 542 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "Intent Classification Loss. We propose a guidance loss on the text embedding " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " to improve training and interpretability. Using the degradation types as targets, we train a simple classification head " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{c} = \\mathcal{C}(\\mathbf{e})" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": ", being " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " is the number of degradation classes. The classification head " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " is a simple two-layers MLP. Thus, we only need to train a projection layer " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " and a simple MLP to capture the natural language knowledge. This allows the text model to learn meaningful embeddings as we can appreciate in Figure 3, not just guidance vectors for the main image processing model. We find that the model is able to classify accurately (i.e. over " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " accuracy) the underlying degradation in the user's prompt after a few epochs." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "InstructIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 296, + 180 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 296, + 180 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 296, + 180 + ], + "type": "image", + "image_path": "ea91eaf38deadf1d7358efae1e44401fb00bdf1492e55ecbf948c468ec3a0c07.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 184, + 297, + 201 + ], + "lines": [ + { + "bbox": [ + 132, + 184, + 297, + 201 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 297, + 201 + ], + "type": "text", + "content": "(a) t-SNE of embeddings before training i.e. frozen text encoder" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 318, + 116, + 478, + 180 + ], + "blocks": [ + { + "bbox": [ + 318, + 116, + 478, + 180 + ], + "lines": [ + { + "bbox": [ + 318, + 116, + 478, + 180 + ], + "spans": [ + { + "bbox": [ + 318, + 116, + 478, + 180 + ], + "type": "image", + "image_path": "e185e23b332ca2daba02817a4bf95b7edc9ba593d262562571d35b704d3849cb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 184, + 480, + 201 + ], + "lines": [ + { + "bbox": [ + 315, + 184, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 315, + 184, + 480, + 201 + ], + "type": "text", + "content": "(b) t-SNE of embeddings after training our learned projection" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 211, + 479, + 232 + ], + "lines": [ + { + "bbox": [ + 131, + 211, + 479, + 232 + ], + "spans": [ + { + "bbox": [ + 131, + 211, + 479, + 232 + ], + "type": "text", + "content": "Fig. 3: We show t-SNE plots of the text embeddings before/after training InstructIR. Each dot represents a human instruction." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 248, + 214, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 214, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 214, + 259 + ], + "type": "text", + "content": "3.3 InstructIR" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 268, + 479, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 479, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 479, + 328 + ], + "type": "text", + "content": "Our method InstructIR consists of an image model and a text encoder. We introduced our text encoder in Sec. 3.2. We use NAFNet [7] as the image model, an efficient image restoration model that follows a U-Net architecture [60]. To successfully learn multiple tasks using a single model, we use task routing techniques. Our framework for training and evaluation is illustrated in Figure 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 339, + 301, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 301, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 301, + 506 + ], + "type": "text", + "content": "Text Guidance. The key aspect of InstructIR is the integration of the encoded instruction as a mechanism of control for the image model. Inspired in task routing for many-task learning [11, 61, 63], we propose an \"Instruction Condition Block\" (ICB) to enable task-specific transformations within the model. Conventional task routing [63] applies task-specific binary masks to the channel features. Since our model does not know a priori the degradation, we cannot use this technique directly." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 507, + 301, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 301, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 301, + 543 + ], + "type": "text", + "content": "Considering the image features " + }, + { + "bbox": [ + 130, + 507, + 301, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 507, + 301, + 543 + ], + "type": "text", + "content": " and the encoded instruction " + }, + { + "bbox": [ + 130, + 507, + 301, + 543 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 130, + 507, + 301, + 543 + ], + "type": "text", + "content": ", we apply task routing as follows:" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 343, + 476, + 426 + ], + "blocks": [ + { + "bbox": [ + 309, + 343, + 476, + 426 + ], + "lines": [ + { + "bbox": [ + 309, + 343, + 476, + 426 + ], + "spans": [ + { + "bbox": [ + 309, + 343, + 476, + 426 + ], + "type": "image", + "image_path": "55a6c6b16f37b8e82e7dab78d358b43fa9e5915cba3093a16b8577cf0966cc5d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 430, + 482, + 508 + ], + "lines": [ + { + "bbox": [ + 306, + 430, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 482, + 508 + ], + "type": "text", + "content": "Fig. 4: Instruction Condition Block (ICB) using an approximation of task routing [63] for many-tasks learning (See Eq. 2). This mechanism allows the neural network to select and prioritize specific features depending on the instruction, similarly to a Mixture of Experts (MoE)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 244, + 552, + 481, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 552, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 244, + 552, + 481, + 564 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {c} ^ {\\prime} = \\operatorname {B l o c k} \\left(\\mathcal {F} _ {c} \\odot \\mathbf {m} _ {c}\\right) + \\mathcal {F} _ {c} \\tag {2}", + "image_path": "6bb61b9f152730f846ea709a388ec98307211091c23781d02630a192f2cd8472.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "where the mask " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_c = \\sigma (\\mathbf{W_c}\\cdot \\mathbf{e})" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " is produced using a linear layer " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{c}}" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " activated using the Sigmoid function - to produce a set of weights depending on the text embedding " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ". Thus, we obtain a " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "-dimensional per-channel (soft-)binary mask " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_c" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ". As [24, 63], task routing is applied as the channel-wise multiplication " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " for masking features depending on the task. The conditioned features are further enhanced using a convolutional NAFBlock [7] (Block). We illustrate our task-routing ICB block in Figure 4. We use \"regular\" NAFBlocks [7], followed by ICBs to condition the features, at both encoder and decoder blocks. The" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "text", + "content": "Conde, Marcos et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "text", + "content": "formulation is " + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "inline_equation", + "content": "F^{l + 1} = \\mathrm{ICB}(\\mathrm{Block}(F^l))" + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 115, + 479, + 200 + ], + "type": "text", + "content": " is the layer. Although we do not condition explicitly the filters of the neural network, as in [63], the mask allows the model to select the most relevant channels depending on the image information and the instruction. Note that this formulation enables differentiable feature masking, and certain interpretability i.e. the features with high weights contribute the most to the restoration process. Indirectly, this also enforces to learn diverse filters and reduce sparsity [11, 63]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 207, + 480, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 207, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 480, + 268 + ], + "type": "text", + "content": "Is InstructIR a blind restoration model? The model does not use explicit information about the degradation in the image e.g. noise profiles, blur kernels, or PSFs. Since our model infers the task (degradation) given the image and the instruction, we consider InstructIR a blind image restoration model. Similarly to previous works that use auxiliary image-based degradation classification [36,53]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 287, + 282, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 287, + 282, + 301 + ], + "spans": [ + { + "bbox": [ + 131, + 287, + 282, + 301 + ], + "type": "text", + "content": "4 Experimental Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 314, + 479, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 479, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 479, + 398 + ], + "type": "text", + "content": "We evaluate our model on 9 well-known benchmarks for different image restoration tasks: image denoising, deblurring, deraining, dehazing, real low-light enhancement, and photo-realistic image enhancement. We present extensive quantitative results in Table 2 and Table 3. We provide extensive comparisons with other all-in-one methods as well as task-specific methods. Our single model successfully restores images considering different degradation types and levels. We provide additional results and ablation studies in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 417, + 282, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 417, + 282, + 429 + ], + "spans": [ + { + "bbox": [ + 131, + 417, + 282, + 429 + ], + "type": "text", + "content": "4.1 Implementation Details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 438, + 479, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 479, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 479, + 463 + ], + "type": "text", + "content": "Our InstructIR model is end-to-end trainable. The image model does not require pre-training but we use a pre-trained sentence encoder as language model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 470, + 480, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 480, + 590 + ], + "type": "text", + "content": "Text Encoder. As we discussed in Sec. 3.2, we only need to train the text embedding projection and classification head (" + }, + { + "bbox": [ + 130, + 470, + 480, + 590 + ], + "type": "inline_equation", + "content": "\\approx" + }, + { + "bbox": [ + 130, + 470, + 480, + 590 + ], + "type": "text", + "content": "100K parameters). We initialize the text encoder with BGE-MICRO-v2" + }, + { + "bbox": [ + 130, + 470, + 480, + 590 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 130, + 470, + 480, + 590 + ], + "type": "text", + "content": ", a distilled version of BGE-SMALL-EN [76]. The BGE encoders are BERT-like encoders [10] pre-trained on large amounts of supervised and unsupervised data for general-purpose sentence encoding. The BGE-micro model is a 3-layer encoder with 17.3 million parameters, which we freeze during training. We also explore ALL-MINILM-L6-v2 and CLIP encoders, however, we concluded that small models prevent overfitting and provide the best performance while being fast. We provide the ablation study comparing the three text encoders in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 597, + 479, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 597, + 479, + 645 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 479, + 645 + ], + "type": "text", + "content": "Image Model. We use NAFNet [7] as the image model backbone. The architecture consists of a 4-level encoder-decoder, with varying numbers of blocks at each level, specifically [2, 2, 4, 8] for the encoder, and [2, 2, 2, 2] for the decoder, from the level-1 to level-4 respectively. Between the encoder and decoder we use" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "InstructIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 652, + 353, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 353, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 353, + 665 + ], + "type": "text", + "content": "3 https://huggingface.co/TaylorAI/bge-micro-v2" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "content": "4 middle blocks to enhance further the features. The decoder implements addition instead of concatenation for the skip connections. We use the Instruction Condition Block (ICB) for task-routing [63] only in the encoder and decoder." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "type": "text", + "content": "The model is optimized using the " + }, + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "type": "text", + "content": " loss between the ground-truth clean image and the restored one. Additionally, we use the cross-entropy loss " + }, + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "type": "text", + "content": " for the intent classification head of the text encoder. We train using a batch size of 32 and AdamW [32] optimizer with learning rate " + }, + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "type": "inline_equation", + "content": "5e^{-4}" + }, + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "type": "text", + "content": " for 500 epochs (approximately 1 day using a single NVIDIA A100). We also use cosine annealing learning rate decay. During training, we utilize cropped patches of size " + }, + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "type": "text", + "content": " as input, and we use random horizontal and vertical flips as augmentations. Since our model uses as input instruction-image pairs, given an image, and knowing its degradation, we randomly sample instructions from our prompt dataset (" + }, + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "type": "inline_equation", + "content": ">10\\mathrm{K}" + }, + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "type": "text", + "content": " samples). Our image model has only 16M parameters, and the learned text projection is just 100k parameters (the language model is 17M parameters), thus, our model can be trained easily on standard GPUs such as NVIDIA RTX 2080Ti or 3090Ti in a couple of days. Furthermore, the inference process also fits in low-computation budgets (e.g. Google Colab T4 16Gb GPU)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 334, + 292, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 334, + 292, + 345 + ], + "spans": [ + { + "bbox": [ + 131, + 334, + 292, + 345 + ], + "type": "text", + "content": "4.2 Datasets and Benchmarks" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 351, + 481, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 481, + 373 + ], + "type": "text", + "content": "Following previous works [36,54,89], we prepare the datasets for different restoration tasks, including real and synthetic datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 373, + 481, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 481, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 481, + 431 + ], + "type": "text", + "content": "Image denoising. We use a combination of BSD400 [2] and WED [43] datasets for training. This combined training set contains " + }, + { + "bbox": [ + 130, + 373, + 481, + 431 + ], + "type": "inline_equation", + "content": "\\approx 5000" + }, + { + "bbox": [ + 130, + 373, + 481, + 431 + ], + "type": "text", + "content": " images. Using as reference the clean images in the dataset, we generate the noisy images by adding Gaussian noise with different noise levels " + }, + { + "bbox": [ + 130, + 373, + 481, + 431 + ], + "type": "inline_equation", + "content": "\\sigma \\in \\{15, 25, 50\\}" + }, + { + "bbox": [ + 130, + 373, + 481, + 431 + ], + "type": "text", + "content": ". We test the models on the well-known BSD68 [45] and Urban100 [28] datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 431, + 481, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 481, + 453 + ], + "type": "text", + "content": "Image deraining. We use the Rain100L [78] dataset, which consists of 200 clean-rainy image pairs for training, and 100 pairs for testing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 453, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 481, + 510 + ], + "type": "text", + "content": "Image dehazing. We utilize the Reside (outdoor) SOTS [35] dataset, which contains " + }, + { + "bbox": [ + 130, + 453, + 481, + 510 + ], + "type": "inline_equation", + "content": "\\approx 72\\mathrm{K}" + }, + { + "bbox": [ + 130, + 453, + 481, + 510 + ], + "type": "text", + "content": " training images. However, many images are low-quality and unrealistic, thus, we filtered the dataset and selected a random set of 2000 images - also to avoid imbalance w.r.t the other tasks. We use the standard outdoor test set of 500 images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 481, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 532 + ], + "type": "text", + "content": "Image deblurring. We use the GoPro dataset for motion deblurring [50] which consists of 2103 images for training, and 1111 for testing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 532, + 481, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 481, + 566 + ], + "type": "text", + "content": "Real-world Low-light Image Enhancement. We use the LOL [74] dataset (v1), which contains real-case low/normal-light image pairs. We adopt its official split of 485 training images and 15 testing images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 566, + 481, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 566, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 566, + 481, + 601 + ], + "type": "text", + "content": "Real-world Image Enhancement. Extending previous works, we also study photorealistic image enhancement using the MIT5K DSLR dataset [5]. We use 1000 images for training, and the standard split of 500 images for testing (as in [66])." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Finally, as previous works [36,54,89], we combine all the aforementioned training datasets, and we train our unified model for all-in-one restoration. Note that we do not include more real-world datasets because previous works do not provide results (or models) for those. Moreover, previous works were limited to synthetic data, in contrast, InstructIR also tackles real-world image enhancement." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "text", + "content": "Conde, Marcos et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 182, + 479, + 338 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 177 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 177 + ], + "type": "text", + "content": "Table 2: Quantitative results on five restoration tasks (5D) with state-of-the-art general image restoration and all-in-one methods. We highlight the reference model without text (image only), the best overall results, and the second best results. We also present the ablation study of our multi-task variants (from 5 to 7 tasks — 5D, 6D, 7D). This table is based on Zhang et al. IDR [89]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 182, + 479, + 338 + ], + "lines": [ + { + "bbox": [ + 133, + 182, + 479, + 338 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 479, + 338 + ], + "type": "table", + "html": "
MethodsDerainingDehazingDenoisingDeblurringLow-light Enh.AverageParams
Rain100L [78]SOTS [35]BSD68 [45]GoPro [50]LOL [74]
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑(M)
HINet [8]35.670.96924.740.93731.000.88126.120.78819.470.80027.400.87588.67
DGUNet [49]36.620.97124.780.94031.100.88327.250.83721.870.82328.320.89117.33
MIRNetV2 [84]33.890.95424.030.92730.970.88126.300.79921.520.81527.340.8755.86
SwinIR [38]30.780.92321.500.89130.590.86824.520.77317.810.72325.040.8350.91
Restormer [83]34.810.96224.090.92731.490.88427.220.82920.410.80627.600.88126.13
NAFNet [7]35.560.96725.230.93931.020.88326.530.80820.490.80927.760.88117.11
DL [17]21.960.76220.540.82623.090.74519.860.67219.830.71221.050.7432.09
Transweather [67]29.430.90521.320.88529.000.84125.120.75721.210.79225.220.83637.93
TAPE [39]29.670.90422.160.86130.180.85524.470.76318.970.62125.090.8011.07
AirNet [36]32.980.95121.040.88430.910.88224.350.78118.180.73525.490.8468.93
InstructIR w/o text35.580.96725.200.93831.090.88326.650.81020.700.82027.840.88417.11
IDR [89]35.630.96525.240.94331.600.88727.870.84621.340.82628.340.89315.34
InstructIR-5D36.840.97327.100.95631.400.88729.400.88623.000.83629.550.90715.8
InstructIR-6D36.800.97327.000.95131.390.88829.730.89222.830.83629.550.90815.8
InstructIR-7D36.750.97226.900.95231.370.88729.700.89222.810.83629.500.90715.8
", + "image_path": "65889baca9ae7456dd5002331e32c162aa745ad274ca82d82af410111e95ebd2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 152, + 389, + 464, + 494 + ], + "blocks": [ + { + "bbox": [ + 132, + 349, + 480, + 383 + ], + "lines": [ + { + "bbox": [ + 132, + 349, + 480, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 349, + 480, + 383 + ], + "type": "text", + "content": "Table 3: Comparisons of all-in-one restoration models for 3 restoration tasks (3D). We also show an ablation study for image denoising -the fundamental inverse problem-considering different noise levels. We report PSNR/SSIM metrics. Table based on [54]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 152, + 389, + 464, + 494 + ], + "lines": [ + { + "bbox": [ + 152, + 389, + 464, + 494 + ], + "spans": [ + { + "bbox": [ + 152, + 389, + 464, + 494 + ], + "type": "table", + "html": "
MethodsDehazing SOTS [35]Deraining Rain100L [17]Denoising ablation study σ = 1529.76/0.836 σ = 2526.34/0.836 σ = 50Average
BRDNet [64]23.23/0.89527.42/0.89532.26/0.89829.76/0.83626.34/0.83627.80/0.843
LPNet [20]20.84/0.82824.88/0.78426.47/0.77824.77/0.74821.26/0.55223.64/0.738
FDGAN [15]24.71/0.92429.89/0.93330.25/0.91028.81/0.86826.43/0.77628.02/0.883
MPRNet [85]25.28/0.95433.57/0.95433.54/0.92730.89/0.88027.56/0.77930.17/0.899
DL [17]26.92/0.93132.62/0.93133.05/0.91430.41/0.86126.90/0.74029.98/0.875
AirNet [36]27.94/0.96234.90/0.96733.92/0.93331.26/0.88828.00/0.79731.20/0.910
PromptIR [54]30.58/0.97436.37/0.97233.98/0.93331.31/0.88828.06/0.79932.06/0.913
InstructIR-3D30.22/0.95937.98/0.97834.15/0.93331.52/0.89028.30/0.80432.43/0.913
InstructIR-5D27.10/0.95636.84/0.97334.00/0.93131.40/0.88728.15/0.79831.50/0.909
InstructIR w/o text26.84/0.94834.02/0.96033.70/0.92930.94/0.88227.78/0.78030.65/0.900
", + "image_path": "0e449264ea826cfe3d84c91dc02cb0a421d307be7ef625024682a0269cab6b3c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 516, + 309, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 516, + 309, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 516, + 309, + 529 + ], + "type": "text", + "content": "4.3 Multiple Degradation Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 538, + 371, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 538, + 371, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 538, + 371, + 550 + ], + "type": "text", + "content": "We define two initial setups for multi-task restoration:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 559, + 481, + 608 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 137, + 559, + 481, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 559, + 481, + 583 + ], + "spans": [ + { + "bbox": [ + 137, + 559, + 481, + 583 + ], + "type": "text", + "content": "- 3D for three-degradation models such as AirNet [36], these tackle image denoising, dehazing and deraining." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 584, + 481, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 584, + 481, + 608 + ], + "spans": [ + { + "bbox": [ + 137, + 584, + 481, + 608 + ], + "type": "text", + "content": "- 5D for five-degradation models, considering image denoising, deblurring, dehazing, deraining and low-light image enhancement as in [89]." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "text", + "content": "In Table 2, we show the performance of 5D models. Following Zhang et al. [89], we compare InstructIR with several state-of-the-art methods for general image restoration [7,8,38,83,84], and all-in-one image restoration methods [17,36, 39,67,89]. We can observe that our simple image model (just 16M parameters)" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "InstructIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "can tackle successfully at least five different tasks thanks to the instruction-based guidance and achieves the most competitive results. In Table 3 we can appreciate a similar behavior, when the number of tasks is just three (3D), our model improves further in terms of reconstruction performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 482, + 236 + ], + "type": "text", + "content": "Based on these results, we pose the following question: How many tasks can we tackle using a single model without losing too much performance? To answer this, we propose the 6D and 7D variants. For the 6D variant, we fine-tune the original 5D to consider also super-resolution as sixth task. Finally, our 7D model includes all previous tasks, and additionally image enhancement (MIT5K photo retouching). We show the performance of these two variants in Table 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 248, + 299, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 299, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 299, + 415 + ], + "type": "text", + "content": "Test Instructions. InstructIR requires as input the degraded image and the human-written instruction. Therefore, we also prepare a test set of prompts i.e. instruction-image test pairs. The performance of InstructIR depends on the ambiguity and precision of the instruction. We provide the ablation study in Table 4. InstructIR is quite robust to more/less detailed instructions. However, it is still limited with ambiguous instructions such as \"enhance this image\". We show diverse instructions in Figures 5 and 6." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 260, + 482, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 260, + 482, + 359 + ], + "spans": [ + { + "bbox": [ + 304, + 260, + 482, + 359 + ], + "type": "text", + "content": "Table 4: Ablation study on the sensitivity of instructions. We report PSNR/SSIM metrics for each task using our 5D base model. We repeat the evaluation on each test set 10 times, each time we sample different prompts for each image, and we report the average results. The \"Real Users " + }, + { + "bbox": [ + 304, + 260, + 482, + 359 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 304, + 260, + 482, + 359 + ], + "type": "text", + "content": " \" in this study are amateur photographers, thus, the instructions were very precise." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 361, + 479, + 393 + ], + "blocks": [ + { + "bbox": [ + 307, + 361, + 479, + 393 + ], + "lines": [ + { + "bbox": [ + 307, + 361, + 479, + 393 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 479, + 393 + ], + "type": "table", + "html": "
Language LevelDerainingDenoisingDeblurringLOL
Basic & Precise36.84/0.97331.40/0.88729.47/0.88723.00/0.836
Basic & Ambiguous36.24/0.97031.35/0.88729.21/0.88521.85/0.827
Real Users †36.84/0.97331.40/0.88729.47/0.88723.00/0.836
", + "image_path": "e38b7e5d009604fdb8842314cffc656bb6f6bc9c319ddc323e97a4f05096ba9e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 434, + 350, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 434, + 350, + 448 + ], + "spans": [ + { + "bbox": [ + 131, + 434, + 350, + 448 + ], + "type": "text", + "content": "5 Multi-Task Discussion and Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 460, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 544 + ], + "type": "text", + "content": "How does 6D work? Besides the 5 basic tasks -as previous works-, we include single image super-resolution (SISR). For this, we include as training data the DIV2K [1]. Since our model does not perform upsampling, we use the Bicubic degradation model [1,12] for generating the low-resolution images (LR), and the upsampled versions (HR) that are fed into our model to enhance them. Adding this extra task increases the performance on deblurring -a related degradation-, without harming notably the performance on the other tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 555, + 334, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 334, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 334, + 664 + ], + "type": "text", + "content": "How does " + }, + { + "bbox": [ + 130, + 555, + 334, + 664 + ], + "type": "inline_equation", + "content": "7D" + }, + { + "bbox": [ + 130, + 555, + 334, + 664 + ], + "type": "text", + "content": " work? Finally, if we add real image enhancement -a task not related to the previous ones i.e. inverse problems- the performance on the restoration tasks decays slightly. However, the model still achieves state-of-the-art results. Moreover, as we show in Table 5, the performance on this task using the MIT5K [5] dataset is notable, while keeping the performance on the other tasks." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 342, + 579, + 479, + 641 + ], + "blocks": [ + { + "bbox": [ + 339, + 556, + 481, + 578 + ], + "lines": [ + { + "bbox": [ + 339, + 556, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 339, + 556, + 481, + 578 + ], + "type": "text", + "content": "Table 5: Real Image Enhancement results on MIT5K [5]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 342, + 579, + 479, + 641 + ], + "lines": [ + { + "bbox": [ + 342, + 579, + 479, + 641 + ], + "spans": [ + { + "bbox": [ + 342, + 579, + 479, + 641 + ], + "type": "table", + "html": "
MethodPSNR↑ SSIM↑ ΔEab ↓
UPE [69]21.880.85310.80
DPE [21]23.750.9089.34
HDRNet [9]24.320.9128.49
3DLUT [86]25.210.9227.61
InstructIR-7D24.650.9008.20
", + "image_path": "e8187c98199246e0ef5e185beacb3a8672cce34533ad4fdf316c557607041a1b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "text", + "content": "Conde, Marcos et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 220, + 163 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 220, + 163 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 220, + 163 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 220, + 163 + ], + "type": "image", + "image_path": "8bcccb2bada2614507419e35540c265557176728e894c331fa84a74e60ff2ca2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 222, + 114, + 306, + 163 + ], + "blocks": [ + { + "bbox": [ + 222, + 114, + 306, + 163 + ], + "lines": [ + { + "bbox": [ + 222, + 114, + 306, + 163 + ], + "spans": [ + { + "bbox": [ + 222, + 114, + 306, + 163 + ], + "type": "image", + "image_path": "0ef09a6f541d427173452ee88a6f80ad132db1d5bbbe39f24dd49e5e809b8145.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 391, + 163 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 391, + 163 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 391, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 391, + 163 + ], + "type": "image", + "image_path": "add9da21428c8935aa2c35ad808f39b92c6d3abc6049c3caeffaa5b200f8d36e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 392, + 114, + 477, + 163 + ], + "blocks": [ + { + "bbox": [ + 392, + 114, + 477, + 163 + ], + "lines": [ + { + "bbox": [ + 392, + 114, + 477, + 163 + ], + "spans": [ + { + "bbox": [ + 392, + 114, + 477, + 163 + ], + "type": "image", + "image_path": "81c2619cc80a3b4be0258702bf19df7a29061e39ab58b9b7d0516dbab4278f02.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 137, + 187, + 220, + 237 + ], + "blocks": [ + { + "bbox": [ + 167, + 167, + 191, + 176 + ], + "lines": [ + { + "bbox": [ + 167, + 167, + 191, + 176 + ], + "spans": [ + { + "bbox": [ + 167, + 167, + 191, + 176 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 187, + 220, + 237 + ], + "lines": [ + { + "bbox": [ + 137, + 187, + 220, + 237 + ], + "spans": [ + { + "bbox": [ + 137, + 187, + 220, + 237 + ], + "type": "image", + "image_path": "7b15fd1272c8a874060da8925a4c3edfc5944c2c860ccfb789898aa0e8fcd7a1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 239, + 220, + 261 + ], + "lines": [ + { + "bbox": [ + 137, + 239, + 220, + 261 + ], + "spans": [ + { + "bbox": [ + 137, + 239, + 220, + 261 + ], + "type": "text", + "content": "\"Retouch this image and improve colors\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 270, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 132, + 270, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 482, + 316 + ], + "type": "text", + "content": "Fig. 5: Adversarial and OOD samples for Instruction-based Restoration. InstructIR understands a wide range of instructions for a given task (first row). Given an adversarial or out-of-distribution instruction (second row), the model does not modify the image notably (i.e. performs the identity) – we did not enforce this during training-." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 222, + 187, + 306, + 237 + ], + "blocks": [ + { + "bbox": [ + 222, + 166, + 306, + 187 + ], + "lines": [ + { + "bbox": [ + 222, + 166, + 306, + 187 + ], + "spans": [ + { + "bbox": [ + 222, + 166, + 306, + 187 + ], + "type": "text", + "content": "\"Clean up my image, it's too fuzzy.\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 222, + 187, + 306, + 237 + ], + "lines": [ + { + "bbox": [ + 222, + 187, + 306, + 237 + ], + "spans": [ + { + "bbox": [ + 222, + 187, + 306, + 237 + ], + "type": "image", + "image_path": "cbd76114bdf1ec34600399e82596afbf57b92df835d5baf86ac36ceeb0c88d59.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 239, + 304, + 261 + ], + "lines": [ + { + "bbox": [ + 225, + 239, + 304, + 261 + ], + "spans": [ + { + "bbox": [ + 225, + 239, + 304, + 261 + ], + "type": "text", + "content": "\"Reduce the motion in this shot\"" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 187, + 391, + 237 + ], + "blocks": [ + { + "bbox": [ + 307, + 166, + 391, + 187 + ], + "lines": [ + { + "bbox": [ + 307, + 166, + 391, + 187 + ], + "spans": [ + { + "bbox": [ + 307, + 166, + 391, + 187 + ], + "type": "text", + "content": "\"Get rid of the grain in my photo\"" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 187, + 391, + 237 + ], + "lines": [ + { + "bbox": [ + 307, + 187, + 391, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 187, + 391, + 237 + ], + "type": "image", + "image_path": "19d4b1a93dff9047f2a26c4a36168ea84fdc642bc6c34fbf644761570f447d6b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 239, + 386, + 261 + ], + "lines": [ + { + "bbox": [ + 313, + 239, + 386, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 386, + 261 + ], + "type": "text", + "content": "\"Please get rid of the raindrops\"" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 392, + 187, + 477, + 237 + ], + "blocks": [ + { + "bbox": [ + 393, + 166, + 476, + 187 + ], + "lines": [ + { + "bbox": [ + 393, + 166, + 476, + 187 + ], + "spans": [ + { + "bbox": [ + 393, + 166, + 476, + 187 + ], + "type": "text", + "content": "\"Remove the strange spots on my photo\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 392, + 187, + 477, + 237 + ], + "lines": [ + { + "bbox": [ + 392, + 187, + 477, + 237 + ], + "spans": [ + { + "bbox": [ + 392, + 187, + 477, + 237 + ], + "type": "image", + "image_path": "22af49b9e372b4790a0cd0e27a054aba1f0b8d1cd63a3c96920d6a2be17cb6ce.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 239, + 471, + 261 + ], + "lines": [ + { + "bbox": [ + 398, + 239, + 471, + 261 + ], + "spans": [ + { + "bbox": [ + 398, + 239, + 471, + 261 + ], + "type": "text", + "content": "\"Reduce the fog in this landmark\"" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 135, + 376, + 479, + 466 + ], + "blocks": [ + { + "bbox": [ + 132, + 338, + 482, + 372 + ], + "lines": [ + { + "bbox": [ + 132, + 338, + 482, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 338, + 482, + 372 + ], + "type": "text", + "content": "Table 7: Quantitative comparisons with state-of-the-art methods on the LOL [74] dataset for Real-World Low-light Enhancement. Note that InstructIR-7D is a multitask method, while the other methods are task-specific. Table based on [72]." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 376, + 479, + 466 + ], + "lines": [ + { + "bbox": [ + 135, + 376, + 479, + 466 + ], + "spans": [ + { + "bbox": [ + 135, + 376, + 479, + 466 + ], + "type": "table", + "html": "
MethodLPNet [37]URetinex -Net [75]DeepLPF [48]SCI [44]LIME [22]MF [18]NPE [70]SRIE [19]SDD [23]CDEF [34]InstructIR Ours
PSNR ↑21.4621.3215.2815.8016.7616.9616.9611.8613.3416.3322.81
SSIM ↑0.8020.8350.4730.5270.4440.5050.4810.4930.6350.5830.836
MethodDRBN [79]KinD [93]RUAS [40]FIDE [77]EG [29]MS-RDN [80]Retinex -Net [74]MIRNet [84]IPT [6]Uformer [73]IAGC [72]
PSNR ↑20.1320.8718.2318.2717.4817.2016.7724.1416.2716.3624.53
SSIM ↑0.8300.8000.7200.6650.6500.6400.5600.8300.5040.5070.842
", + "image_path": "5259200baaf5aa970ffedb8ea00d91441423c3e1cce3572afc19ad880656d522.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 487, + 298, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 298, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 298, + 547 + ], + "type": "text", + "content": "We summarize the multi-task ablation study in Table 6. Our model can tackle multiple tasks without losing performance notably thanks to the instruction-based task routing." + } + ] + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 307, + 519, + 479, + 569 + ], + "blocks": [ + { + "bbox": [ + 305, + 485, + 481, + 517 + ], + "lines": [ + { + "bbox": [ + 305, + 485, + 481, + 517 + ], + "spans": [ + { + "bbox": [ + 305, + 485, + 481, + 517 + ], + "type": "text", + "content": "Table 6: Summary ablation study on the multi-task variants of InstructIR that tackle from 3 to 7 tasks." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 519, + 479, + 569 + ], + "lines": [ + { + "bbox": [ + 307, + 519, + 479, + 569 + ], + "spans": [ + { + "bbox": [ + 307, + 519, + 479, + 569 + ], + "type": "table", + "html": "
TasksRainNoise (σ15)BlurLOL
3D37.98/0.97831.52/0.890--
5D36.84/0.97331.40/0.88729.40/0.88623.00/0.836
6D36.80 0.97331.39 0.88829.73/0.89222.83 0.836
7D36.75 0.97231.37 0.88729.70/0.89222.81 0.836
", + "image_path": "2e5b74d24ff896c868ad44bef204da2ccb7db19edaff341ac225447b223dcc9e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 587, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 482, + 636 + ], + "type": "text", + "content": "Comparison with Task-specific Methods Our main goal is to design a powerful all-in-one model, thus, InstructIR was not designed to be trained for a particular degradation. Nevertheless, we also compare InstructIR with task-specific methods i.e. models tailored and trained for specific tasks." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": "We compare with task-specific methods for real-world photography enhancement in Table 5, and for real-world low-light image enhancement in Table 7." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "InstructIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 250, + 185 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 250, + 185 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 250, + 185 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 250, + 185 + ], + "type": "image", + "image_path": "60416b4c0584616589a2dcc447494d0a3e7d7b6ab89c0de32db686af5e21767c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 116, + 360, + 184 + ], + "blocks": [ + { + "bbox": [ + 250, + 116, + 360, + 184 + ], + "lines": [ + { + "bbox": [ + 250, + 116, + 360, + 184 + ], + "spans": [ + { + "bbox": [ + 250, + 116, + 360, + 184 + ], + "type": "image", + "image_path": "7634f3a49bcc4952134959d94071342db8ffbd6afcb27d2a8f4dc1d4a8ffe8b4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 363, + 118, + 472, + 184 + ], + "blocks": [ + { + "bbox": [ + 363, + 118, + 472, + 184 + ], + "lines": [ + { + "bbox": [ + 363, + 118, + 472, + 184 + ], + "spans": [ + { + "bbox": [ + 363, + 118, + 472, + 184 + ], + "type": "image", + "image_path": "0bf5d9047ce5bb3c306cc50271400608be89ef3463e4d7b1d027b91800bd5ee7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 140, + 195, + 249, + 262 + ], + "blocks": [ + { + "bbox": [ + 182, + 186, + 205, + 195 + ], + "lines": [ + { + "bbox": [ + 182, + 186, + 205, + 195 + ], + "spans": [ + { + "bbox": [ + 182, + 186, + 205, + 195 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 140, + 195, + 249, + 262 + ], + "lines": [ + { + "bbox": [ + 140, + 195, + 249, + 262 + ], + "spans": [ + { + "bbox": [ + 140, + 195, + 249, + 262 + ], + "type": "image", + "image_path": "c8a16fb2224007a58f6b571e89331558607424a72f7a0b2240065ec8608012c1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 263, + 251, + 273 + ], + "lines": [ + { + "bbox": [ + 140, + 263, + 251, + 273 + ], + "spans": [ + { + "bbox": [ + 140, + 263, + 251, + 273 + ], + "type": "text", + "content": "(1) \"Retouch it as a photographer" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 250, + 195, + 359, + 262 + ], + "blocks": [ + { + "bbox": [ + 264, + 186, + 465, + 195 + ], + "lines": [ + { + "bbox": [ + 264, + 186, + 465, + 195 + ], + "spans": [ + { + "bbox": [ + 264, + 186, + 465, + 195 + ], + "type": "text", + "content": "(1) \"Clear the rain from my picture\" " + }, + { + "bbox": [ + 264, + 186, + 465, + 195 + ], + "type": "inline_equation", + "content": "\\longrightarrow" + }, + { + "bbox": [ + 264, + 186, + 465, + 195 + ], + "type": "text", + "content": " (2) \"Make it pop\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 250, + 195, + 359, + 262 + ], + "lines": [ + { + "bbox": [ + 250, + 195, + 359, + 262 + ], + "spans": [ + { + "bbox": [ + 250, + 195, + 359, + 262 + ], + "type": "image", + "image_path": "e8192ce76b7575b1fdaf5fbc9e2d3118ecab16261e3f76575b0ab576a63ec396.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 263, + 476, + 273 + ], + "lines": [ + { + "bbox": [ + 250, + 263, + 476, + 273 + ], + "spans": [ + { + "bbox": [ + 250, + 263, + 476, + 273 + ], + "type": "inline_equation", + "content": "r'' \\longrightarrow 'Can you remove the raindrops?\" \\longrightarrow 'Increase the resolution''" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 363, + 195, + 472, + 262 + ], + "blocks": [ + { + "bbox": [ + 363, + 195, + 472, + 262 + ], + "lines": [ + { + "bbox": [ + 363, + 195, + 472, + 262 + ], + "spans": [ + { + "bbox": [ + 363, + 195, + 472, + 262 + ], + "type": "image", + "image_path": "08ca14f84c6597d74dd2d97a1ac11c654c940a388499bf5a081c51bed8631df4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 138, + 274, + 249, + 344 + ], + "blocks": [ + { + "bbox": [ + 138, + 274, + 249, + 344 + ], + "lines": [ + { + "bbox": [ + 138, + 274, + 249, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 249, + 344 + ], + "type": "image", + "image_path": "7630abf7b0765598023e47eacecf3607c3f0dfdc2ebb8bcf62d0fac41b9d6abc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 345, + 205, + 355 + ], + "lines": [ + { + "bbox": [ + 182, + 345, + 205, + 355 + ], + "spans": [ + { + "bbox": [ + 182, + 345, + 205, + 355 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 250, + 274, + 359, + 343 + ], + "blocks": [ + { + "bbox": [ + 250, + 274, + 359, + 343 + ], + "lines": [ + { + "bbox": [ + 250, + 274, + 359, + 343 + ], + "spans": [ + { + "bbox": [ + 250, + 274, + 359, + 343 + ], + "type": "image", + "image_path": "497093888fc3f303a84d83180981ad284859affdc056a0d8abd189d808097426.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 346, + 468, + 355 + ], + "lines": [ + { + "bbox": [ + 260, + 346, + 468, + 355 + ], + "spans": [ + { + "bbox": [ + 260, + 346, + 468, + 355 + ], + "type": "text", + "content": "(1) \"My image is too dark, fix it\" " + }, + { + "bbox": [ + 260, + 346, + 468, + 355 + ], + "type": "inline_equation", + "content": "\\longrightarrow" + }, + { + "bbox": [ + 260, + 346, + 468, + 355 + ], + "type": "text", + "content": " (2) 'Apply a tonemap'" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 363, + 274, + 472, + 342 + ], + "blocks": [ + { + "bbox": [ + 363, + 274, + 472, + 342 + ], + "lines": [ + { + "bbox": [ + 363, + 274, + 472, + 342 + ], + "spans": [ + { + "bbox": [ + 363, + 274, + 472, + 342 + ], + "type": "image", + "image_path": "b59efc2ccaf1b29ca540a899374a2dbde8cf3960b9e693953a1be1ba9f58c1ab.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 145, + 408, + 223, + 469 + ], + "blocks": [ + { + "bbox": [ + 131, + 365, + 482, + 399 + ], + "lines": [ + { + "bbox": [ + 131, + 365, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 131, + 365, + 482, + 399 + ], + "type": "text", + "content": "Fig. 6: Control via instructions. We can prompt multiple instructions (in sequence) to restore and enhance the images. This provides additional control. We show two examples of multiple instructions applied to the \"Input\" image -from left to right-." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 145, + 408, + 223, + 469 + ], + "lines": [ + { + "bbox": [ + 145, + 408, + 223, + 469 + ], + "spans": [ + { + "bbox": [ + 145, + 408, + 223, + 469 + ], + "type": "image", + "image_path": "97d321d57a875344929996d08e4630d2127fed00ff3a9e03ed2858f91a698e6b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 471, + 223, + 482 + ], + "lines": [ + { + "bbox": [ + 146, + 471, + 223, + 482 + ], + "spans": [ + { + "bbox": [ + 146, + 471, + 223, + 482 + ], + "type": "text", + "content": "Input (RealSRSet)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 489, + 482, + 512 + ], + "lines": [ + { + "bbox": [ + 131, + 489, + 482, + 512 + ], + "spans": [ + { + "bbox": [ + 131, + 489, + 482, + 512 + ], + "type": "text", + "content": "Fig.7: Comparison with InstructPix2Pix [4] using the prompt \"Remove the noise in this photo\". Real-case image from RealSRSet [38]." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 224, + 409, + 302, + 469 + ], + "blocks": [ + { + "bbox": [ + 224, + 409, + 302, + 469 + ], + "lines": [ + { + "bbox": [ + 224, + 409, + 302, + 469 + ], + "spans": [ + { + "bbox": [ + 224, + 409, + 302, + 469 + ], + "type": "image", + "image_path": "b4772aa72cdf41c53b2e9426a5b5dccb453f55b02e485e6e045c0a3ffe7356ae.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 471, + 285, + 481 + ], + "lines": [ + { + "bbox": [ + 241, + 471, + 285, + 481 + ], + "spans": [ + { + "bbox": [ + 241, + 471, + 285, + 481 + ], + "type": "text", + "content": "InstructIR" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 471, + 465, + 482 + ], + "lines": [ + { + "bbox": [ + 302, + 471, + 465, + 482 + ], + "spans": [ + { + "bbox": [ + 302, + 471, + 465, + 482 + ], + "type": "text", + "content": "InstructPix2Pix #1 InstructPix2Pix #2" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 303, + 409, + 382, + 469 + ], + "blocks": [ + { + "bbox": [ + 303, + 409, + 382, + 469 + ], + "lines": [ + { + "bbox": [ + 303, + 409, + 382, + 469 + ], + "spans": [ + { + "bbox": [ + 303, + 409, + 382, + 469 + ], + "type": "image", + "image_path": "7c690650ec75ceffd9c5bc7652b1368a9c9997626cfe1e864746a3e621e9a192.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 385, + 409, + 463, + 469 + ], + "blocks": [ + { + "bbox": [ + 385, + 409, + 463, + 469 + ], + "lines": [ + { + "bbox": [ + 385, + 409, + 463, + 469 + ], + "spans": [ + { + "bbox": [ + 385, + 409, + 463, + 469 + ], + "type": "image", + "image_path": "c0e64c0032ad5b926c3e1826b7d2c2278b8fe69dca2c36af265f62c08030eca0.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 131, + 535, + 340, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 535, + 340, + 546 + ], + "spans": [ + { + "bbox": [ + 131, + 535, + 340, + 546 + ], + "type": "text", + "content": "5.1 On the Effectiveness of Instructions" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 130, + 555, + 481, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 481, + 590 + ], + "type": "text", + "content": "Thanks to our integration of human instructions, users can control how to enhance the images. We provide examples in Figures 5 and 6, where we show the potential of our method to restore and enhance images in a controllable manner." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 130, + 591, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 591, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 591, + 482, + 639 + ], + "type": "text", + "content": "This implies an advancement w.r.t classical (deterministic) image restoration methods. Classical deep restoration methods lead to a unique result, thus, they do not allow to control how the image is processed. We also compare InstructIR with InstructPix2Pix [4] (a diffusion-based generative model) in Figure 7." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Qualitative Results. We provide diverse qualitative results for several tasks, and we compare with all-in-one and task-specific methods. In Figure 10, we show" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "text", + "content": "Conde, Marcos et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 147, + 116, + 463, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 463, + 125 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 463, + 125 + ], + "type": "text", + "content": "Instruction: \"my colors are too off, make it pop so I can use these photos in instagram\"" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 141, + 126, + 222, + 179 + ], + "blocks": [ + { + "bbox": [ + 141, + 126, + 222, + 179 + ], + "lines": [ + { + "bbox": [ + 141, + 126, + 222, + 179 + ], + "spans": [ + { + "bbox": [ + 141, + 126, + 222, + 179 + ], + "type": "image", + "image_path": "f3ad7deaa755c317646cd5f67d5fa8166f280b606740351fb765904cc572b7a1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 223, + 126, + 304, + 179 + ], + "blocks": [ + { + "bbox": [ + 223, + 126, + 304, + 179 + ], + "lines": [ + { + "bbox": [ + 223, + 126, + 304, + 179 + ], + "spans": [ + { + "bbox": [ + 223, + 126, + 304, + 179 + ], + "type": "image", + "image_path": "715d36189c1551feff2188e486e25f156503ba81d5cb9b15e4035e37c789d4a9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 304, + 126, + 385, + 179 + ], + "blocks": [ + { + "bbox": [ + 304, + 126, + 385, + 179 + ], + "lines": [ + { + "bbox": [ + 304, + 126, + 385, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 126, + 385, + 179 + ], + "type": "image", + "image_path": "ad1a1fd835e7d4b8d80b514a23be1881227e3269c969d5515b85af5787881535.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 386, + 126, + 468, + 179 + ], + "blocks": [ + { + "bbox": [ + 386, + 126, + 468, + 179 + ], + "lines": [ + { + "bbox": [ + 386, + 126, + 468, + 179 + ], + "spans": [ + { + "bbox": [ + 386, + 126, + 468, + 179 + ], + "type": "image", + "image_path": "ae72433d5937f559ffab2d268aac0bf187b1113e70afefbb9d5f470060c34dd0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 183, + 470, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 183, + 470, + 193 + ], + "spans": [ + { + "bbox": [ + 141, + 183, + 470, + 193 + ], + "type": "text", + "content": "\"Increase the brightness of my photo please, is it totoro?\" / \"my image is too dark, fix it\"" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 141, + 193, + 222, + 247 + ], + "blocks": [ + { + "bbox": [ + 141, + 193, + 222, + 247 + ], + "lines": [ + { + "bbox": [ + 141, + 193, + 222, + 247 + ], + "spans": [ + { + "bbox": [ + 141, + 193, + 222, + 247 + ], + "type": "image", + "image_path": "730bb54bbc219579e5f3b13f1de38d9cd1dfbee411cf2ac96fb0f8381126c42b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 256, + 479, + 267 + ], + "lines": [ + { + "bbox": [ + 132, + 256, + 479, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 256, + 479, + 267 + ], + "type": "text", + "content": "Fig. 8: Real-world samples of image restoration and enhancement using InstructIR." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 223, + 193, + 304, + 247 + ], + "blocks": [ + { + "bbox": [ + 223, + 193, + 304, + 247 + ], + "lines": [ + { + "bbox": [ + 223, + 193, + 304, + 247 + ], + "spans": [ + { + "bbox": [ + 223, + 193, + 304, + 247 + ], + "type": "image", + "image_path": "969ecfbfdb19383e15c0f4ca2f279061cdca9f8f129cfdb964a585ff50bbea38.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 304, + 193, + 385, + 247 + ], + "blocks": [ + { + "bbox": [ + 304, + 193, + 385, + 247 + ], + "lines": [ + { + "bbox": [ + 304, + 193, + 385, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 385, + 247 + ], + "type": "image", + "image_path": "b1229dc4382b7055869832deb70100f77f37065fe172bbb3547fb03ee33e7c54.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 386, + 193, + 468, + 247 + ], + "blocks": [ + { + "bbox": [ + 386, + 193, + 468, + 247 + ], + "lines": [ + { + "bbox": [ + 386, + 193, + 468, + 247 + ], + "spans": [ + { + "bbox": [ + 386, + 193, + 468, + 247 + ], + "type": "image", + "image_path": "9da96949b25277e62dd4107bfc1528f7a062e27fc91a0768ebcbb82dbe645381.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 136, + 278, + 203, + 329 + ], + "blocks": [ + { + "bbox": [ + 136, + 278, + 203, + 329 + ], + "lines": [ + { + "bbox": [ + 136, + 278, + 203, + 329 + ], + "spans": [ + { + "bbox": [ + 136, + 278, + 203, + 329 + ], + "type": "image", + "image_path": "9435cfd5a78f9d5f8d9e944ab690553c8e233ece96a8fbcedaa3fed60567f913.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 204, + 279, + 271, + 329 + ], + "blocks": [ + { + "bbox": [ + 204, + 279, + 271, + 329 + ], + "lines": [ + { + "bbox": [ + 204, + 279, + 271, + 329 + ], + "spans": [ + { + "bbox": [ + 204, + 279, + 271, + 329 + ], + "type": "image", + "image_path": "46c4466867627c5e860a5254080a700a8b44d9728240c751fa732e7886c2ac20.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 272, + 279, + 339, + 329 + ], + "blocks": [ + { + "bbox": [ + 272, + 279, + 339, + 329 + ], + "lines": [ + { + "bbox": [ + 272, + 279, + 339, + 329 + ], + "spans": [ + { + "bbox": [ + 272, + 279, + 339, + 329 + ], + "type": "image", + "image_path": "c70505c6af0166b68dec30cc3d0452d7beb4f09816cd16f6fcef2aa52cef4c15.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 340, + 279, + 406, + 329 + ], + "blocks": [ + { + "bbox": [ + 340, + 279, + 406, + 329 + ], + "lines": [ + { + "bbox": [ + 340, + 279, + 406, + 329 + ], + "spans": [ + { + "bbox": [ + 340, + 279, + 406, + 329 + ], + "type": "image", + "image_path": "00bc1e382402d42498c92573b5951324819509bb0fd0f1006861aaac828b2eb0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 408, + 279, + 474, + 329 + ], + "blocks": [ + { + "bbox": [ + 408, + 279, + 474, + 329 + ], + "lines": [ + { + "bbox": [ + 408, + 279, + 474, + 329 + ], + "spans": [ + { + "bbox": [ + 408, + 279, + 474, + 329 + ], + "type": "image", + "image_path": "44b8bab9807a75e30954c2f56a8c1931275b0bcb5fdcf3f4275f201d21f3f8f4.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 136, + 331, + 203, + 369 + ], + "blocks": [ + { + "bbox": [ + 136, + 331, + 203, + 369 + ], + "lines": [ + { + "bbox": [ + 136, + 331, + 203, + 369 + ], + "spans": [ + { + "bbox": [ + 136, + 331, + 203, + 369 + ], + "type": "image", + "image_path": "be63fd17da14f825879323891e2fc1cc6541d59aee3209696722935c2a5ee5c2.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 373, + 181, + 383 + ], + "lines": [ + { + "bbox": [ + 158, + 373, + 181, + 383 + ], + "spans": [ + { + "bbox": [ + 158, + 373, + 181, + 383 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 205, + 332, + 271, + 369 + ], + "blocks": [ + { + "bbox": [ + 205, + 332, + 271, + 369 + ], + "lines": [ + { + "bbox": [ + 205, + 332, + 271, + 369 + ], + "spans": [ + { + "bbox": [ + 205, + 332, + 271, + 369 + ], + "type": "image", + "image_path": "c3af3f60b8a75090cffc5fb877d38618b0107097d4fc7ff15f557416d3b7772b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 373, + 260, + 383 + ], + "lines": [ + { + "bbox": [ + 214, + 373, + 260, + 383 + ], + "spans": [ + { + "bbox": [ + 214, + 373, + 260, + 383 + ], + "type": "text", + "content": "AirNet [36]" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 272, + 332, + 339, + 369 + ], + "blocks": [ + { + "bbox": [ + 272, + 332, + 339, + 369 + ], + "lines": [ + { + "bbox": [ + 272, + 332, + 339, + 369 + ], + "spans": [ + { + "bbox": [ + 272, + 332, + 339, + 369 + ], + "type": "image", + "image_path": "6233f7c954b3f97a09df14099a8a66a742b89d1a5d1a143adfee17293deffbd5.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 372, + 334, + 383 + ], + "lines": [ + { + "bbox": [ + 276, + 372, + 334, + 383 + ], + "spans": [ + { + "bbox": [ + 276, + 372, + 334, + 383 + ], + "type": "text", + "content": "PromptIR [54]" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 340, + 332, + 406, + 369 + ], + "blocks": [ + { + "bbox": [ + 340, + 332, + 406, + 369 + ], + "lines": [ + { + "bbox": [ + 340, + 332, + 406, + 369 + ], + "spans": [ + { + "bbox": [ + 340, + 332, + 406, + 369 + ], + "type": "image", + "image_path": "3f1def5a075c6dd824b66ee5514e81b80be6938430cc5366b39aa9652b2643a3.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 373, + 394, + 382 + ], + "lines": [ + { + "bbox": [ + 351, + 373, + 394, + 382 + ], + "spans": [ + { + "bbox": [ + 351, + 373, + 394, + 382 + ], + "type": "text", + "content": "InstructIR" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 387, + 474, + 399 + ], + "lines": [ + { + "bbox": [ + 138, + 387, + 474, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 387, + 474, + 399 + ], + "type": "text", + "content": "Fig. 9: Dehazing comparisons for all-in-one restoration methods on SOTS [35]." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 408, + 332, + 474, + 369 + ], + "blocks": [ + { + "bbox": [ + 408, + 332, + 474, + 369 + ], + "lines": [ + { + "bbox": [ + 408, + 332, + 474, + 369 + ], + "spans": [ + { + "bbox": [ + 408, + 332, + 474, + 369 + ], + "type": "image", + "image_path": "ec45eabe49713611f28bcee6f6de5a036e840e05f0fa0f6732172f1dafa03b89.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 373, + 460, + 381 + ], + "lines": [ + { + "bbox": [ + 421, + 373, + 460, + 381 + ], + "spans": [ + { + "bbox": [ + 421, + 373, + 460, + 381 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 130, + 422, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 557 + ], + "type": "text", + "content": "results on the LOL [74] dataset. In Figure 11, we compare methods on the motion deblurring task using the GoPro [50] dataset. In Figures 9 and 12, we compare with different methods for the dehazing task on SOTS (outdoor) [35]. In Figure 13, we compare with image restoration methods for deraining on Rain100L [17]. Finally, we show denoising results in Figure 14. In this qualitative analysis, we use our single InstructIR-5D model to restore all the images. Limitations As with previous all-in-one methods, our model struggles to process images with more than one degradation (i.e. complex real-world images), or unknown out-of-distribution degradations, yet this is a common limitation among the related methods. However, we believe that these limitations can be surpassed with more realistic training data, and scaling the model's complexity." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 132, + 568, + 219, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 219, + 581 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 219, + 581 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "We present a novel approach that uses natural human-written instructions to guide the image restoration model. Given a prompt, our multi-task model can recover high-quality images from their degraded counterparts, considering multiple degradations. We achieve state-of-the-art results on several restoration tasks, demonstrating the power and flexibility of instruction guidance. Our results represent a new benchmark for text-guided image restoration." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "InstructIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 121, + 246, + 186 + ], + "blocks": [ + { + "bbox": [ + 138, + 121, + 246, + 186 + ], + "lines": [ + { + "bbox": [ + 138, + 121, + 246, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 121, + 246, + 186 + ], + "type": "image", + "image_path": "8125aafc4839ebcb4c71b26079513104de1684649f67fede8bc456704dce09db.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 186, + 213, + 194 + ], + "lines": [ + { + "bbox": [ + 171, + 186, + 213, + 194 + ], + "spans": [ + { + "bbox": [ + 171, + 186, + 213, + 194 + ], + "type": "text", + "content": "Low-light Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 248, + 121, + 292, + 150 + ], + "blocks": [ + { + "bbox": [ + 248, + 121, + 292, + 150 + ], + "lines": [ + { + "bbox": [ + 248, + 121, + 292, + 150 + ], + "spans": [ + { + "bbox": [ + 248, + 121, + 292, + 150 + ], + "type": "image", + "image_path": "d9be2b4d2d1f65cda746e00c7d31508865d3810f8c91fa6eb6076d1dbec21560.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 151, + 280, + 156 + ], + "lines": [ + { + "bbox": [ + 260, + 151, + 280, + 156 + ], + "spans": [ + { + "bbox": [ + 260, + 151, + 280, + 156 + ], + "type": "text", + "content": "NAFNet" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 293, + 121, + 337, + 150 + ], + "blocks": [ + { + "bbox": [ + 293, + 121, + 337, + 150 + ], + "lines": [ + { + "bbox": [ + 293, + 121, + 337, + 150 + ], + "spans": [ + { + "bbox": [ + 293, + 121, + 337, + 150 + ], + "type": "image", + "image_path": "5c3d96742248411cc2e169d206adad6bd16ecddfe2f43b5cc09bf4f4d081eef3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 151, + 325, + 156 + ], + "lines": [ + { + "bbox": [ + 304, + 151, + 325, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 151, + 325, + 156 + ], + "type": "text", + "content": "MPRNet" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 338, + 121, + 381, + 150 + ], + "blocks": [ + { + "bbox": [ + 338, + 121, + 381, + 150 + ], + "lines": [ + { + "bbox": [ + 338, + 121, + 381, + 150 + ], + "spans": [ + { + "bbox": [ + 338, + 121, + 381, + 150 + ], + "type": "image", + "image_path": "d00f684e6d24e1cc588bc98620fe75d5e3d1fa289a4a5597b1407d9a6f908aa3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 151, + 371, + 156 + ], + "lines": [ + { + "bbox": [ + 350, + 151, + 371, + 156 + ], + "spans": [ + { + "bbox": [ + 350, + 151, + 371, + 156 + ], + "type": "text", + "content": "DGUNet" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 383, + 121, + 427, + 150 + ], + "blocks": [ + { + "bbox": [ + 383, + 121, + 427, + 150 + ], + "lines": [ + { + "bbox": [ + 383, + 121, + 427, + 150 + ], + "spans": [ + { + "bbox": [ + 383, + 121, + 427, + 150 + ], + "type": "image", + "image_path": "ba46adc333e4cd04c14997084a80ba0c318eb90ef4a644dd93d61ab1f3935bc7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 151, + 419, + 156 + ], + "lines": [ + { + "bbox": [ + 392, + 151, + 419, + 156 + ], + "spans": [ + { + "bbox": [ + 392, + 151, + 419, + 156 + ], + "type": "text", + "content": "MIRNetV2" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 427, + 121, + 472, + 150 + ], + "blocks": [ + { + "bbox": [ + 427, + 121, + 472, + 150 + ], + "lines": [ + { + "bbox": [ + 427, + 121, + 472, + 150 + ], + "spans": [ + { + "bbox": [ + 427, + 121, + 472, + 150 + ], + "type": "image", + "image_path": "031d26c4dd4bf444fb3111eb40b8cb489799984a83f947355212adbc9876a69e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 151, + 465, + 156 + ], + "lines": [ + { + "bbox": [ + 435, + 151, + 465, + 156 + ], + "spans": [ + { + "bbox": [ + 435, + 151, + 465, + 156 + ], + "type": "text", + "content": "InstructIR" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 135, + 158, + 246, + 186 + ], + "blocks": [ + { + "bbox": [ + 135, + 158, + 246, + 186 + ], + "lines": [ + { + "bbox": [ + 135, + 158, + 246, + 186 + ], + "spans": [ + { + "bbox": [ + 135, + 158, + 246, + 186 + ], + "type": "image", + "image_path": "46616ef4e8c83a1a172efef1e99d18b9e5ce30b9caa5d06a2f1153b0791fcdd7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 247, + 158, + 292, + 186 + ], + "blocks": [ + { + "bbox": [ + 262, + 186, + 277, + 193 + ], + "lines": [ + { + "bbox": [ + 262, + 186, + 277, + 193 + ], + "spans": [ + { + "bbox": [ + 262, + 186, + 277, + 193 + ], + "type": "text", + "content": "TAPE" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 247, + 158, + 292, + 186 + ], + "lines": [ + { + "bbox": [ + 247, + 158, + 292, + 186 + ], + "spans": [ + { + "bbox": [ + 247, + 158, + 292, + 186 + ], + "type": "image", + "image_path": "f639bfa8c48a1d8e1a3fbbcf1f54596b496afe186343c5a078c9213f734181b5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 293, + 158, + 337, + 186 + ], + "blocks": [ + { + "bbox": [ + 302, + 186, + 327, + 193 + ], + "lines": [ + { + "bbox": [ + 302, + 186, + 327, + 193 + ], + "spans": [ + { + "bbox": [ + 302, + 186, + 327, + 193 + ], + "type": "text", + "content": "Restormer" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 293, + 158, + 337, + 186 + ], + "lines": [ + { + "bbox": [ + 293, + 158, + 337, + 186 + ], + "spans": [ + { + "bbox": [ + 293, + 158, + 337, + 186 + ], + "type": "image", + "image_path": "946712f5ded19400803b8e125945abd58504658f00abf2107c7346d7dda5b212.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 338, + 158, + 427, + 186 + ], + "blocks": [ + { + "bbox": [ + 352, + 186, + 369, + 193 + ], + "lines": [ + { + "bbox": [ + 352, + 186, + 369, + 193 + ], + "spans": [ + { + "bbox": [ + 352, + 186, + 369, + 193 + ], + "type": "text", + "content": "AirNet" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 338, + 158, + 427, + 186 + ], + "lines": [ + { + "bbox": [ + 338, + 158, + 427, + 186 + ], + "spans": [ + { + "bbox": [ + 338, + 158, + 427, + 186 + ], + "type": "image", + "image_path": "5031453b98702aaa860947e198148d8ba4978ee4a37e19c84468a193f34c7db7.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 186, + 410, + 193 + ], + "lines": [ + { + "bbox": [ + 400, + 186, + 410, + 193 + ], + "spans": [ + { + "bbox": [ + 400, + 186, + 410, + 193 + ], + "type": "text", + "content": "IDR" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 352, + 186, + 368, + 193 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 186, + 368, + 193 + ], + "spans": [ + { + "bbox": [ + 352, + 186, + 368, + 193 + ], + "type": "text", + "content": "AirNet" + } + ] + } + ], + "index": 21, + "type": "text" + }, + { + "bbox": [ + 367, + 158, + 427, + 186 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 158, + 427, + 186 + ], + "spans": [ + { + "bbox": [ + 367, + 158, + 427, + 186 + ], + "type": "image", + "image_path": "732735b2aeee3411c3815bb2357b162cf06ef286fbfa837a6ec74199f9fa3dc6.jpg" + } + ] + } + ], + "index": 22, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 429, + 158, + 471, + 186 + ], + "blocks": [ + { + "bbox": [ + 429, + 158, + 471, + 186 + ], + "lines": [ + { + "bbox": [ + 429, + 158, + 471, + 186 + ], + "spans": [ + { + "bbox": [ + 429, + 158, + 471, + 186 + ], + "type": "image", + "image_path": "442cfc75e98c6466278325f29be28db9caac2fcd26670f4453a5d15d23366cfa.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 186, + 454, + 193 + ], + "lines": [ + { + "bbox": [ + 446, + 186, + 454, + 193 + ], + "spans": [ + { + "bbox": [ + 446, + 186, + 454, + 193 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 138, + 253, + 247, + 319 + ], + "blocks": [ + { + "bbox": [ + 138, + 253, + 247, + 319 + ], + "lines": [ + { + "bbox": [ + 138, + 253, + 247, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 253, + 247, + 319 + ], + "type": "image", + "image_path": "6cdcf2e2c02ae92bd12cdce4be545aab9f1002a983c45ca28cbfa4fc834871f2.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 319, + 212, + 326 + ], + "lines": [ + { + "bbox": [ + 180, + 319, + 212, + 326 + ], + "spans": [ + { + "bbox": [ + 180, + 319, + 212, + 326 + ], + "type": "text", + "content": "Blurry Image" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 184, + 327, + 427, + 338 + ], + "lines": [ + { + "bbox": [ + 184, + 327, + 427, + 338 + ], + "spans": [ + { + "bbox": [ + 184, + 327, + 427, + 338 + ], + "type": "text", + "content": "Fig. 11: Image Deblurring Results. GoPro [50] dataset." + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 248, + 254, + 292, + 284 + ], + "blocks": [ + { + "bbox": [ + 248, + 254, + 292, + 284 + ], + "lines": [ + { + "bbox": [ + 248, + 254, + 292, + 284 + ], + "spans": [ + { + "bbox": [ + 248, + 254, + 292, + 284 + ], + "type": "image", + "image_path": "dc4d1d40b9fcbb5f946538adce8a4961a32568684a5ad744d08691a87950b4b9.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 284, + 281, + 289 + ], + "lines": [ + { + "bbox": [ + 260, + 284, + 281, + 289 + ], + "spans": [ + { + "bbox": [ + 260, + 284, + 281, + 289 + ], + "type": "text", + "content": "NAFNet" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 293, + 254, + 337, + 284 + ], + "blocks": [ + { + "bbox": [ + 293, + 254, + 337, + 284 + ], + "lines": [ + { + "bbox": [ + 293, + 254, + 337, + 284 + ], + "spans": [ + { + "bbox": [ + 293, + 254, + 337, + 284 + ], + "type": "image", + "image_path": "818281906252ff084602697ddcc1f69d33cfeb657b5b34e1c936993b06906b61.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 284, + 325, + 289 + ], + "lines": [ + { + "bbox": [ + 304, + 284, + 325, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 325, + 289 + ], + "type": "text", + "content": "MPRNet" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 338, + 254, + 381, + 284 + ], + "blocks": [ + { + "bbox": [ + 338, + 254, + 381, + 284 + ], + "lines": [ + { + "bbox": [ + 338, + 254, + 381, + 284 + ], + "spans": [ + { + "bbox": [ + 338, + 254, + 381, + 284 + ], + "type": "image", + "image_path": "b28ff142c1596559d1312fae787d60132f830244de99cba50d27912864f75425.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 284, + 371, + 289 + ], + "lines": [ + { + "bbox": [ + 349, + 284, + 371, + 289 + ], + "spans": [ + { + "bbox": [ + 349, + 284, + 371, + 289 + ], + "type": "text", + "content": "DGUNet" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 383, + 254, + 427, + 284 + ], + "blocks": [ + { + "bbox": [ + 383, + 254, + 427, + 284 + ], + "lines": [ + { + "bbox": [ + 383, + 254, + 427, + 284 + ], + "spans": [ + { + "bbox": [ + 383, + 254, + 427, + 284 + ], + "type": "image", + "image_path": "6ddaf809bf7118af0334a9a135487cf824692545ca2a3795eb2599018d5bf463.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 284, + 419, + 289 + ], + "lines": [ + { + "bbox": [ + 392, + 284, + 419, + 289 + ], + "spans": [ + { + "bbox": [ + 392, + 284, + 419, + 289 + ], + "type": "text", + "content": "MIRNetV2" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 429, + 254, + 472, + 284 + ], + "blocks": [ + { + "bbox": [ + 131, + 200, + 481, + 234 + ], + "lines": [ + { + "bbox": [ + 131, + 200, + 481, + 234 + ], + "spans": [ + { + "bbox": [ + 131, + 200, + 481, + 234 + ], + "type": "text", + "content": "Fig. 10: Real Low-light Enhancement Results. LOL [74] testset (748.png). AirNet [36] and IDR [89] are well-known all-in-one restoration methods. NAFNet [7] is equivalent to InstructIR without text conditions (i.e. our image-only backbone)." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 429, + 254, + 472, + 284 + ], + "lines": [ + { + "bbox": [ + 429, + 254, + 472, + 284 + ], + "spans": [ + { + "bbox": [ + 429, + 254, + 472, + 284 + ], + "type": "image", + "image_path": "9555f65f6a06e4d7532a7edf95901eecea10972e14fe852db4cddc4341590aac.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 284, + 465, + 289 + ], + "lines": [ + { + "bbox": [ + 435, + 284, + 465, + 289 + ], + "spans": [ + { + "bbox": [ + 435, + 284, + 465, + 289 + ], + "type": "text", + "content": "InstructIR" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "bbox": [ + 138, + 289, + 247, + 319 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 247, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 247, + 319 + ], + "type": "image", + "image_path": "775dacf94276c8efeee73a024932c3af700d6761924dba6f53756122075da86d.jpg" + } + ] + } + ], + "index": 39, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 247, + 289, + 292, + 319 + ], + "blocks": [ + { + "bbox": [ + 247, + 289, + 292, + 319 + ], + "lines": [ + { + "bbox": [ + 247, + 289, + 292, + 319 + ], + "spans": [ + { + "bbox": [ + 247, + 289, + 292, + 319 + ], + "type": "image", + "image_path": "18bca19ee81a0da59440aaa9c0f131d78f9f26838609b69cdfa21eb5684a1916.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 319, + 277, + 326 + ], + "lines": [ + { + "bbox": [ + 263, + 319, + 277, + 326 + ], + "spans": [ + { + "bbox": [ + 263, + 319, + 277, + 326 + ], + "type": "text", + "content": "TAPE" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 293, + 289, + 337, + 319 + ], + "blocks": [ + { + "bbox": [ + 293, + 289, + 337, + 319 + ], + "lines": [ + { + "bbox": [ + 293, + 289, + 337, + 319 + ], + "spans": [ + { + "bbox": [ + 293, + 289, + 337, + 319 + ], + "type": "image", + "image_path": "24df408de42c7c6d838b38dd6a518938fbf8337d48101e1999a6420d6cdcc354.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 319, + 328, + 326 + ], + "lines": [ + { + "bbox": [ + 302, + 319, + 328, + 326 + ], + "spans": [ + { + "bbox": [ + 302, + 319, + 328, + 326 + ], + "type": "text", + "content": "Restormer" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 338, + 289, + 381, + 319 + ], + "blocks": [ + { + "bbox": [ + 338, + 289, + 381, + 319 + ], + "lines": [ + { + "bbox": [ + 338, + 289, + 381, + 319 + ], + "spans": [ + { + "bbox": [ + 338, + 289, + 381, + 319 + ], + "type": "image", + "image_path": "2784727bb426120f76a3bd782012d65c414effab0f4627b2e48824fb6004fc6d.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 319, + 369, + 326 + ], + "lines": [ + { + "bbox": [ + 351, + 319, + 369, + 326 + ], + "spans": [ + { + "bbox": [ + 351, + 319, + 369, + 326 + ], + "type": "text", + "content": "AirNet" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 383, + 289, + 427, + 319 + ], + "blocks": [ + { + "bbox": [ + 383, + 289, + 427, + 319 + ], + "lines": [ + { + "bbox": [ + 383, + 289, + 427, + 319 + ], + "spans": [ + { + "bbox": [ + 383, + 289, + 427, + 319 + ], + "type": "image", + "image_path": "b84b3ee66eea74cfdf05cb063e3944b2ad1060ebd9ec20084c4b1a470b0a6e8c.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 319, + 419, + 326 + ], + "lines": [ + { + "bbox": [ + 400, + 319, + 419, + 326 + ], + "spans": [ + { + "bbox": [ + 400, + 319, + 419, + 326 + ], + "type": "text", + "content": "IDR" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 429, + 289, + 472, + 319 + ], + "blocks": [ + { + "bbox": [ + 429, + 289, + 472, + 319 + ], + "lines": [ + { + "bbox": [ + 429, + 289, + 472, + 319 + ], + "spans": [ + { + "bbox": [ + 429, + 289, + 472, + 319 + ], + "type": "image", + "image_path": "68a9005f7f24e806fe68cea6ebbf48e42f2a0e6aa4d5617c70bc0d40d4011449.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 319, + 454, + 326 + ], + "lines": [ + { + "bbox": [ + 446, + 319, + 454, + 326 + ], + "spans": [ + { + "bbox": [ + 446, + 319, + 454, + 326 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 138, + 359, + 247, + 425 + ], + "blocks": [ + { + "bbox": [ + 138, + 359, + 247, + 425 + ], + "lines": [ + { + "bbox": [ + 138, + 359, + 247, + 425 + ], + "spans": [ + { + "bbox": [ + 138, + 359, + 247, + 425 + ], + "type": "image", + "image_path": "0e0be5640cdba7caa33ace574f795c3f509110dd7306bd8879b57b6d96131929.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 425, + 206, + 432 + ], + "lines": [ + { + "bbox": [ + 178, + 425, + 206, + 432 + ], + "spans": [ + { + "bbox": [ + 178, + 425, + 206, + 432 + ], + "type": "text", + "content": "Haze Image" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 262, + 425, + 277, + 431 + ], + "lines": [ + { + "bbox": [ + 262, + 425, + 277, + 431 + ], + "spans": [ + { + "bbox": [ + 262, + 425, + 277, + 431 + ], + "type": "text", + "content": "TAPE" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 148, + 433, + 463, + 445 + ], + "lines": [ + { + "bbox": [ + 148, + 433, + 463, + 445 + ], + "spans": [ + { + "bbox": [ + 148, + 433, + 463, + 445 + ], + "type": "text", + "content": "Fig. 12: Image Dehazing Results. SOTS [35] outdoor dataset (0150.jpg)." + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_caption" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 248, + 360, + 292, + 389 + ], + "blocks": [ + { + "bbox": [ + 248, + 360, + 292, + 389 + ], + "lines": [ + { + "bbox": [ + 248, + 360, + 292, + 389 + ], + "spans": [ + { + "bbox": [ + 248, + 360, + 292, + 389 + ], + "type": "image", + "image_path": "41c39169c458ddfe91ce3411ba6aec81d3b2eb5477474c2395723a61a7f410ad.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 389, + 280, + 395 + ], + "lines": [ + { + "bbox": [ + 260, + 389, + 280, + 395 + ], + "spans": [ + { + "bbox": [ + 260, + 389, + 280, + 395 + ], + "type": "text", + "content": "NAFNet" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 293, + 360, + 337, + 389 + ], + "blocks": [ + { + "bbox": [ + 293, + 360, + 337, + 389 + ], + "lines": [ + { + "bbox": [ + 293, + 360, + 337, + 389 + ], + "spans": [ + { + "bbox": [ + 293, + 360, + 337, + 389 + ], + "type": "image", + "image_path": "9b3374cdce6fc4ce038fad4d6cc21f23a865a59b965c6eb10a200630e047791b.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 389, + 325, + 395 + ], + "lines": [ + { + "bbox": [ + 304, + 389, + 325, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 325, + 395 + ], + "type": "text", + "content": "MPRNet" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 338, + 360, + 381, + 389 + ], + "blocks": [ + { + "bbox": [ + 338, + 360, + 381, + 389 + ], + "lines": [ + { + "bbox": [ + 338, + 360, + 381, + 389 + ], + "spans": [ + { + "bbox": [ + 338, + 360, + 381, + 389 + ], + "type": "image", + "image_path": "a411721d191f5b171ee7d820db2e35e04ecaa9be30d9ac53e0c4fa74cb0a83b9.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 389, + 371, + 395 + ], + "lines": [ + { + "bbox": [ + 350, + 389, + 371, + 395 + ], + "spans": [ + { + "bbox": [ + 350, + 389, + 371, + 395 + ], + "type": "text", + "content": "DGUNet" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_caption" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 383, + 360, + 427, + 389 + ], + "blocks": [ + { + "bbox": [ + 383, + 360, + 427, + 389 + ], + "lines": [ + { + "bbox": [ + 383, + 360, + 427, + 389 + ], + "spans": [ + { + "bbox": [ + 383, + 360, + 427, + 389 + ], + "type": "image", + "image_path": "668e91bd950f4a412be7f115edb85eff794294872c104839ee66af610d3f3e3a.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 389, + 419, + 396 + ], + "lines": [ + { + "bbox": [ + 392, + 389, + 419, + 396 + ], + "spans": [ + { + "bbox": [ + 392, + 389, + 419, + 396 + ], + "type": "text", + "content": "MIRNetV2" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 429, + 360, + 472, + 389 + ], + "blocks": [ + { + "bbox": [ + 429, + 360, + 472, + 389 + ], + "lines": [ + { + "bbox": [ + 429, + 360, + 472, + 389 + ], + "spans": [ + { + "bbox": [ + 429, + 360, + 472, + 389 + ], + "type": "image", + "image_path": "8846fd0590d30a846e9f4b8cbf3943f3c48be3e85cd79c958401cf49135a57b4.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 389, + 465, + 396 + ], + "lines": [ + { + "bbox": [ + 435, + 389, + 465, + 396 + ], + "spans": [ + { + "bbox": [ + 435, + 389, + 465, + 396 + ], + "type": "text", + "content": "InstructIR" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_caption" + } + ], + "index": 61 + }, + { + "bbox": [ + 138, + 396, + 247, + 425 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 247, + 425 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 247, + 425 + ], + "type": "image", + "image_path": "1da8d3afc46437464dc0037396ee53f9175678dcf091c338d535b696529b14d0.jpg" + } + ] + } + ], + "index": 63, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 293, + 396, + 337, + 425 + ], + "blocks": [ + { + "bbox": [ + 293, + 396, + 337, + 425 + ], + "lines": [ + { + "bbox": [ + 293, + 396, + 337, + 425 + ], + "spans": [ + { + "bbox": [ + 293, + 396, + 337, + 425 + ], + "type": "image", + "image_path": "eb25d9f6a35b2302ae0e7b01347306658c0f360aaf810c53ffa994063c3e6e62.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 425, + 328, + 431 + ], + "lines": [ + { + "bbox": [ + 302, + 425, + 328, + 431 + ], + "spans": [ + { + "bbox": [ + 302, + 425, + 328, + 431 + ], + "type": "text", + "content": "Restormer" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_caption" + } + ], + "index": 65 + }, + { + "type": "image", + "bbox": [ + 338, + 396, + 381, + 425 + ], + "blocks": [ + { + "bbox": [ + 338, + 396, + 381, + 425 + ], + "lines": [ + { + "bbox": [ + 338, + 396, + 381, + 425 + ], + "spans": [ + { + "bbox": [ + 338, + 396, + 381, + 425 + ], + "type": "image", + "image_path": "9ddf39bc733b26f72261a9b09c42852b39cea34377683b17dd6131c6a3464f19.jpg" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 425, + 369, + 431 + ], + "lines": [ + { + "bbox": [ + 352, + 425, + 369, + 431 + ], + "spans": [ + { + "bbox": [ + 352, + 425, + 369, + 431 + ], + "type": "text", + "content": "AirNet" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_caption" + } + ], + "index": 67 + }, + { + "type": "image", + "bbox": [ + 383, + 396, + 427, + 425 + ], + "blocks": [ + { + "bbox": [ + 383, + 396, + 427, + 425 + ], + "lines": [ + { + "bbox": [ + 383, + 396, + 427, + 425 + ], + "spans": [ + { + "bbox": [ + 383, + 396, + 427, + 425 + ], + "type": "image", + "image_path": "f227c6a65dc0d6e1e23721ed4446cd12fb45d21a2ab0f254b368b238217465b9.jpg" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 425, + 410, + 431 + ], + "lines": [ + { + "bbox": [ + 400, + 425, + 410, + 431 + ], + "spans": [ + { + "bbox": [ + 400, + 425, + 410, + 431 + ], + "type": "text", + "content": "IDF" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_caption" + } + ], + "index": 69 + }, + { + "type": "image", + "bbox": [ + 429, + 396, + 472, + 425 + ], + "blocks": [ + { + "bbox": [ + 429, + 396, + 472, + 425 + ], + "lines": [ + { + "bbox": [ + 429, + 396, + 472, + 425 + ], + "spans": [ + { + "bbox": [ + 429, + 396, + 472, + 425 + ], + "type": "image", + "image_path": "7b045258e36478771730b85a0f0cb35d41b66df6eadeea2a0b2fa5a8be4d995a.jpg" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 425, + 454, + 430 + ], + "lines": [ + { + "bbox": [ + 446, + 425, + 454, + 430 + ], + "spans": [ + { + "bbox": [ + 446, + 425, + 454, + 430 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_caption" + } + ], + "index": 71 + }, + { + "type": "image", + "bbox": [ + 138, + 464, + 247, + 510 + ], + "blocks": [ + { + "bbox": [ + 138, + 464, + 247, + 510 + ], + "lines": [ + { + "bbox": [ + 138, + 464, + 247, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 464, + 247, + 510 + ], + "type": "image", + "image_path": "9dd58982d68cba36318bbd95577bf45ebd505881e9e9d26081984ebd5e877b1a.jpg" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 510, + 208, + 516 + ], + "lines": [ + { + "bbox": [ + 178, + 510, + 208, + 516 + ], + "spans": [ + { + "bbox": [ + 178, + 510, + 208, + 516 + ], + "type": "inline_equation", + "content": "\\therefore m = \\frac{3}{11}" + }, + { + "bbox": [ + 178, + 510, + 208, + 516 + ], + "type": "text", + "content": " ;" + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "image_caption" + } + ], + "index": 74 + }, + { + "type": "image", + "bbox": [ + 248, + 465, + 292, + 495 + ], + "blocks": [ + { + "bbox": [ + 248, + 465, + 292, + 495 + ], + "lines": [ + { + "bbox": [ + 248, + 465, + 292, + 495 + ], + "spans": [ + { + "bbox": [ + 248, + 465, + 292, + 495 + ], + "type": "image", + "image_path": "5e708036c87c9a42ff105d210ef45578cfbc3df9654ac0d2de78605cebe12193.jpg" + } + ] + } + ], + "index": 76, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 495, + 280, + 500 + ], + "lines": [ + { + "bbox": [ + 260, + 495, + 280, + 500 + ], + "spans": [ + { + "bbox": [ + 260, + 495, + 280, + 500 + ], + "type": "text", + "content": "NAFNet" + } + ] + } + ], + "index": 77, + "angle": 0, + "type": "image_caption" + } + ], + "index": 76 + }, + { + "type": "image", + "bbox": [ + 293, + 465, + 337, + 495 + ], + "blocks": [ + { + "bbox": [ + 293, + 465, + 337, + 495 + ], + "lines": [ + { + "bbox": [ + 293, + 465, + 337, + 495 + ], + "spans": [ + { + "bbox": [ + 293, + 465, + 337, + 495 + ], + "type": "image", + "image_path": "3c8b6a87e36742171897855d274bbe3f36186351476c51be7bb317087b25e4e5.jpg" + } + ] + } + ], + "index": 78, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 495, + 325, + 500 + ], + "lines": [ + { + "bbox": [ + 304, + 495, + 325, + 500 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 325, + 500 + ], + "type": "text", + "content": "MPRNet" + } + ] + } + ], + "index": 79, + "angle": 0, + "type": "image_caption" + } + ], + "index": 78 + }, + { + "type": "image", + "bbox": [ + 338, + 465, + 381, + 495 + ], + "blocks": [ + { + "bbox": [ + 338, + 465, + 381, + 495 + ], + "lines": [ + { + "bbox": [ + 338, + 465, + 381, + 495 + ], + "spans": [ + { + "bbox": [ + 338, + 465, + 381, + 495 + ], + "type": "image", + "image_path": "659c360949b5cdd10f7b01f5ecdf1d1ee8c834e51cbb050b43216e8a54c31d85.jpg" + } + ] + } + ], + "index": 80, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 495, + 371, + 500 + ], + "lines": [ + { + "bbox": [ + 350, + 495, + 371, + 500 + ], + "spans": [ + { + "bbox": [ + 350, + 495, + 371, + 500 + ], + "type": "text", + "content": "DGUNet" + } + ] + } + ], + "index": 81, + "angle": 0, + "type": "image_caption" + } + ], + "index": 80 + }, + { + "type": "image", + "bbox": [ + 383, + 465, + 427, + 495 + ], + "blocks": [ + { + "bbox": [ + 383, + 465, + 427, + 495 + ], + "lines": [ + { + "bbox": [ + 383, + 465, + 427, + 495 + ], + "spans": [ + { + "bbox": [ + 383, + 465, + 427, + 495 + ], + "type": "image", + "image_path": "098b484679f2c6b41b81bd05bd50c5ef916bb4ea978023dfcfcb5570b2c024a7.jpg" + } + ] + } + ], + "index": 82, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 495, + 419, + 500 + ], + "lines": [ + { + "bbox": [ + 392, + 495, + 419, + 500 + ], + "spans": [ + { + "bbox": [ + 392, + 495, + 419, + 500 + ], + "type": "text", + "content": "MIRNetV2" + } + ] + } + ], + "index": 83, + "angle": 0, + "type": "image_caption" + } + ], + "index": 82 + }, + { + "type": "image", + "bbox": [ + 429, + 465, + 472, + 495 + ], + "blocks": [ + { + "bbox": [ + 429, + 465, + 472, + 495 + ], + "lines": [ + { + "bbox": [ + 429, + 465, + 472, + 495 + ], + "spans": [ + { + "bbox": [ + 429, + 465, + 472, + 495 + ], + "type": "image", + "image_path": "7fe9ec5d9bc72bb886bdf65e0d75b2d48f10c877cd3e4b93facca106bc5468ef.jpg" + } + ] + } + ], + "index": 84, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 495, + 465, + 500 + ], + "lines": [ + { + "bbox": [ + 435, + 495, + 465, + 500 + ], + "spans": [ + { + "bbox": [ + 435, + 495, + 465, + 500 + ], + "type": "text", + "content": "InstructIR" + } + ] + } + ], + "index": 85, + "angle": 0, + "type": "image_caption" + } + ], + "index": 84 + }, + { + "type": "image", + "bbox": [ + 138, + 517, + 247, + 530 + ], + "blocks": [ + { + "bbox": [ + 138, + 517, + 247, + 530 + ], + "lines": [ + { + "bbox": [ + 138, + 517, + 247, + 530 + ], + "spans": [ + { + "bbox": [ + 138, + 517, + 247, + 530 + ], + "type": "image", + "image_path": "60f5f6186f56dfd1afdd6183dba3ea62e7573afd7a373f7ab25fff5165f6d57a.jpg" + } + ] + } + ], + "index": 86, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 530, + 208, + 536 + ], + "lines": [ + { + "bbox": [ + 178, + 530, + 208, + 536 + ], + "spans": [ + { + "bbox": [ + 178, + 530, + 208, + 536 + ], + "type": "text", + "content": "Rainy Image" + } + ] + } + ], + "index": 87, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 171, + 538, + 441, + 551 + ], + "lines": [ + { + "bbox": [ + 171, + 538, + 441, + 551 + ], + "spans": [ + { + "bbox": [ + 171, + 538, + 441, + 551 + ], + "type": "text", + "content": "Fig. 13: Image Deraining Results on Rain100L [17] (035.png)." + } + ] + } + ], + "index": 98, + "angle": 0, + "type": "image_caption" + } + ], + "index": 86 + }, + { + "type": "image", + "bbox": [ + 247, + 517, + 292, + 530 + ], + "blocks": [ + { + "bbox": [ + 247, + 517, + 292, + 530 + ], + "lines": [ + { + "bbox": [ + 247, + 517, + 292, + 530 + ], + "spans": [ + { + "bbox": [ + 247, + 517, + 292, + 530 + ], + "type": "image", + "image_path": "02c08d5e562af874a657c0ea006a78a3b903a4bfc55c06b888bcf9d43a55d91e.jpg" + } + ] + } + ], + "index": 88, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 262, + 530, + 277, + 536 + ], + "lines": [ + { + "bbox": [ + 262, + 530, + 277, + 536 + ], + "spans": [ + { + "bbox": [ + 262, + 530, + 277, + 536 + ], + "type": "text", + "content": "TAPE" + } + ] + } + ], + "index": 89, + "angle": 0, + "type": "image_caption" + } + ], + "index": 88 + }, + { + "type": "image", + "bbox": [ + 293, + 517, + 337, + 530 + ], + "blocks": [ + { + "bbox": [ + 293, + 517, + 337, + 530 + ], + "lines": [ + { + "bbox": [ + 293, + 517, + 337, + 530 + ], + "spans": [ + { + "bbox": [ + 293, + 517, + 337, + 530 + ], + "type": "image", + "image_path": "27c5e77d6415849da93a7b1145cef8754b19a83814ee73537eac0ef4aefd0e75.jpg" + } + ] + } + ], + "index": 90, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 530, + 328, + 536 + ], + "lines": [ + { + "bbox": [ + 302, + 530, + 328, + 536 + ], + "spans": [ + { + "bbox": [ + 302, + 530, + 328, + 536 + ], + "type": "text", + "content": "Restormer" + } + ] + } + ], + "index": 91, + "angle": 0, + "type": "image_caption" + } + ], + "index": 90 + }, + { + "type": "image", + "bbox": [ + 337, + 517, + 381, + 530 + ], + "blocks": [ + { + "bbox": [ + 337, + 517, + 381, + 530 + ], + "lines": [ + { + "bbox": [ + 337, + 517, + 381, + 530 + ], + "spans": [ + { + "bbox": [ + 337, + 517, + 381, + 530 + ], + "type": "image", + "image_path": "0388fd22cc593cdc8e884516b5ce1f60e5a1883ab944d65ffa31f9304a0f109c.jpg" + } + ] + } + ], + "index": 92, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 530, + 369, + 536 + ], + "lines": [ + { + "bbox": [ + 351, + 530, + 369, + 536 + ], + "spans": [ + { + "bbox": [ + 351, + 530, + 369, + 536 + ], + "type": "text", + "content": "AirNet" + } + ] + } + ], + "index": 93, + "angle": 0, + "type": "image_caption" + } + ], + "index": 92 + }, + { + "type": "image", + "bbox": [ + 383, + 517, + 427, + 530 + ], + "blocks": [ + { + "bbox": [ + 383, + 517, + 427, + 530 + ], + "lines": [ + { + "bbox": [ + 383, + 517, + 427, + 530 + ], + "spans": [ + { + "bbox": [ + 383, + 517, + 427, + 530 + ], + "type": "image", + "image_path": "454ba8c97256230e526d9ad8d7f0fb2b37361fa75e181c3668006c66cc8be812.jpg" + } + ] + } + ], + "index": 94, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 530, + 410, + 536 + ], + "lines": [ + { + "bbox": [ + 400, + 530, + 410, + 536 + ], + "spans": [ + { + "bbox": [ + 400, + 530, + 410, + 536 + ], + "type": "text", + "content": "IDR" + } + ] + } + ], + "index": 95, + "angle": 0, + "type": "image_caption" + } + ], + "index": 94 + }, + { + "type": "image", + "bbox": [ + 429, + 517, + 472, + 530 + ], + "blocks": [ + { + "bbox": [ + 429, + 517, + 472, + 530 + ], + "lines": [ + { + "bbox": [ + 429, + 517, + 472, + 530 + ], + "spans": [ + { + "bbox": [ + 429, + 517, + 472, + 530 + ], + "type": "image", + "image_path": "f9073ec7db09e4b08fb8174028b30d443515794d6ae6551784d87336206b6ffb.jpg" + } + ] + } + ], + "index": 96, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 530, + 454, + 536 + ], + "lines": [ + { + "bbox": [ + 446, + 530, + 454, + 536 + ], + "spans": [ + { + "bbox": [ + 446, + 530, + 454, + 536 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 97, + "angle": 0, + "type": "image_caption" + } + ], + "index": 96 + }, + { + "type": "image", + "bbox": [ + 138, + 570, + 247, + 636 + ], + "blocks": [ + { + "bbox": [ + 138, + 570, + 247, + 636 + ], + "lines": [ + { + "bbox": [ + 138, + 570, + 247, + 636 + ], + "spans": [ + { + "bbox": [ + 138, + 570, + 247, + 636 + ], + "type": "image", + "image_path": "c5b2ba6176dfe37e6be5f2c3cf06e6581d67aad0a1b41396d786339dace7444c.jpg" + } + ] + } + ], + "index": 99, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 644, + 438, + 656 + ], + "lines": [ + { + "bbox": [ + 173, + 644, + 438, + 656 + ], + "spans": [ + { + "bbox": [ + 173, + 644, + 438, + 656 + ], + "type": "text", + "content": "Fig. 14: Image Denoising Results on BSD68 [45] (0060.png)." + } + ] + } + ], + "index": 123, + "angle": 0, + "type": "image_caption" + } + ], + "index": 99 + }, + { + "bbox": [ + 178, + 635, + 208, + 643 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 635, + 208, + 643 + ], + "spans": [ + { + "bbox": [ + 178, + 635, + 208, + 643 + ], + "type": "text", + "content": "Noise Image" + } + ] + } + ], + "index": 100, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 248, + 571, + 292, + 599 + ], + "blocks": [ + { + "bbox": [ + 248, + 571, + 292, + 599 + ], + "lines": [ + { + "bbox": [ + 248, + 571, + 292, + 599 + ], + "spans": [ + { + "bbox": [ + 248, + 571, + 292, + 599 + ], + "type": "image", + "image_path": "292b40798c91c11d14544e9d1d0ef2ba539294365767e3b35668d18bd14bbb62.jpg" + } + ] + } + ], + "index": 101, + "angle": 0, + "type": "image_body" + } + ], + "index": 101 + }, + { + "type": "image", + "bbox": [ + 293, + 571, + 337, + 599 + ], + "blocks": [ + { + "bbox": [ + 293, + 571, + 337, + 599 + ], + "lines": [ + { + "bbox": [ + 293, + 571, + 337, + 599 + ], + "spans": [ + { + "bbox": [ + 293, + 571, + 337, + 599 + ], + "type": "image", + "image_path": "0892a6e504b0f6094241e8703d7dbb15dc6eea22ee3943e4ac95a01838107508.jpg" + } + ] + } + ], + "index": 103, + "angle": 0, + "type": "image_body" + } + ], + "index": 103 + }, + { + "type": "image", + "bbox": [ + 338, + 571, + 381, + 599 + ], + "blocks": [ + { + "bbox": [ + 338, + 571, + 381, + 599 + ], + "lines": [ + { + "bbox": [ + 338, + 571, + 381, + 599 + ], + "spans": [ + { + "bbox": [ + 338, + 571, + 381, + 599 + ], + "type": "image", + "image_path": "23d85597e72292315b869bb39eb28cf3d9bddecf2d62596d2ced5d1bab8f4c7e.jpg" + } + ] + } + ], + "index": 105, + "angle": 0, + "type": "image_body" + } + ], + "index": 105 + }, + { + "type": "image", + "bbox": [ + 383, + 571, + 427, + 599 + ], + "blocks": [ + { + "bbox": [ + 383, + 571, + 427, + 599 + ], + "lines": [ + { + "bbox": [ + 383, + 571, + 427, + 599 + ], + "spans": [ + { + "bbox": [ + 383, + 571, + 427, + 599 + ], + "type": "image", + "image_path": "23359ee7289ec06eb35212328358961fdd6a4507880254dd9f5d2b7ae8e10651.jpg" + } + ] + } + ], + "index": 107, + "angle": 0, + "type": "image_body" + } + ], + "index": 107 + }, + { + "type": "image", + "bbox": [ + 429, + 571, + 472, + 599 + ], + "blocks": [ + { + "bbox": [ + 429, + 571, + 472, + 599 + ], + "lines": [ + { + "bbox": [ + 429, + 571, + 472, + 599 + ], + "spans": [ + { + "bbox": [ + 429, + 571, + 472, + 599 + ], + "type": "image", + "image_path": "e25e2fe1fc849443e17bec637366b61d62bdc1811250fdcad65af4220cb804d7.jpg" + } + ] + } + ], + "index": 109, + "angle": 0, + "type": "image_body" + } + ], + "index": 109 + }, + { + "bbox": [ + 138, + 605, + 247, + 635 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 605, + 247, + 635 + ], + "spans": [ + { + "bbox": [ + 138, + 605, + 247, + 635 + ], + "type": "image", + "image_path": "4d7d99aa8872a57150d4dd1268e39ac9b329f3e0e8ec888800d3bcf4e61a14f3.jpg" + } + ] + } + ], + "index": 111, + "type": "text" + }, + { + "bbox": [ + 178, + 635, + 208, + 643 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 635, + 208, + 643 + ], + "spans": [ + { + "bbox": [ + 178, + 635, + 208, + 643 + ], + "type": "text", + "content": "Noise Image" + } + ] + } + ], + "index": 112, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 247, + 605, + 292, + 635 + ], + "blocks": [ + { + "bbox": [ + 260, + 600, + 280, + 606 + ], + "lines": [ + { + "bbox": [ + 260, + 600, + 280, + 606 + ], + "spans": [ + { + "bbox": [ + 260, + 600, + 280, + 606 + ], + "type": "text", + "content": "NAFNet" + } + ] + } + ], + "index": 102, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 247, + 605, + 292, + 635 + ], + "lines": [ + { + "bbox": [ + 247, + 605, + 292, + 635 + ], + "spans": [ + { + "bbox": [ + 247, + 605, + 292, + 635 + ], + "type": "image", + "image_path": "e78430966735a45edb4a887cd0ec3405e0c167185bb9aeda3dcbff327f5a686a.jpg" + } + ] + } + ], + "index": 113, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 262, + 635, + 277, + 642 + ], + "lines": [ + { + "bbox": [ + 262, + 635, + 277, + 642 + ], + "spans": [ + { + "bbox": [ + 262, + 635, + 277, + 642 + ], + "type": "text", + "content": "TAPE" + } + ] + } + ], + "index": 114, + "angle": 0, + "type": "image_caption" + } + ], + "index": 113 + }, + { + "type": "image", + "bbox": [ + 293, + 605, + 337, + 635 + ], + "blocks": [ + { + "bbox": [ + 304, + 600, + 325, + 605 + ], + "lines": [ + { + "bbox": [ + 304, + 600, + 325, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 600, + 325, + 605 + ], + "type": "text", + "content": "MPRNet" + } + ] + } + ], + "index": 104, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 293, + 605, + 337, + 635 + ], + "lines": [ + { + "bbox": [ + 293, + 605, + 337, + 635 + ], + "spans": [ + { + "bbox": [ + 293, + 605, + 337, + 635 + ], + "type": "image", + "image_path": "4092a20cb176a696593465c27a6ca80f5977fd10859e6522f51f3fe33b992221.jpg" + } + ] + } + ], + "index": 115, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 635, + 327, + 642 + ], + "lines": [ + { + "bbox": [ + 302, + 635, + 327, + 642 + ], + "spans": [ + { + "bbox": [ + 302, + 635, + 327, + 642 + ], + "type": "text", + "content": "Restormer" + } + ] + } + ], + "index": 116, + "angle": 0, + "type": "image_caption" + } + ], + "index": 115 + }, + { + "type": "image", + "bbox": [ + 338, + 605, + 381, + 635 + ], + "blocks": [ + { + "bbox": [ + 350, + 600, + 371, + 605 + ], + "lines": [ + { + "bbox": [ + 350, + 600, + 371, + 605 + ], + "spans": [ + { + "bbox": [ + 350, + 600, + 371, + 605 + ], + "type": "text", + "content": "DGUNet" + } + ] + } + ], + "index": 106, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 338, + 605, + 381, + 635 + ], + "lines": [ + { + "bbox": [ + 338, + 605, + 381, + 635 + ], + "spans": [ + { + "bbox": [ + 338, + 605, + 381, + 635 + ], + "type": "image", + "image_path": "c164dae35b527dc63d1bc60de44a3aa0a6088218fc6d1fd81f45771fd287a0e4.jpg" + } + ] + } + ], + "index": 117, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 635, + 369, + 642 + ], + "lines": [ + { + "bbox": [ + 351, + 635, + 369, + 642 + ], + "spans": [ + { + "bbox": [ + 351, + 635, + 369, + 642 + ], + "type": "text", + "content": "AirNet" + } + ] + } + ], + "index": 118, + "angle": 0, + "type": "image_caption" + } + ], + "index": 117 + }, + { + "type": "image", + "bbox": [ + 383, + 605, + 427, + 635 + ], + "blocks": [ + { + "bbox": [ + 392, + 600, + 419, + 606 + ], + "lines": [ + { + "bbox": [ + 392, + 600, + 419, + 606 + ], + "spans": [ + { + "bbox": [ + 392, + 600, + 419, + 606 + ], + "type": "text", + "content": "MIRNetV2" + } + ] + } + ], + "index": 108, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 383, + 605, + 427, + 635 + ], + "lines": [ + { + "bbox": [ + 383, + 605, + 427, + 635 + ], + "spans": [ + { + "bbox": [ + 383, + 605, + 427, + 635 + ], + "type": "image", + "image_path": "2c658dde79342b0c7f889e82357ba98ef9f5fe05cdea5392fd48f4f389415436.jpg" + } + ] + } + ], + "index": 119, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 635, + 410, + 642 + ], + "lines": [ + { + "bbox": [ + 399, + 635, + 410, + 642 + ], + "spans": [ + { + "bbox": [ + 399, + 635, + 410, + 642 + ], + "type": "text", + "content": "IDR" + } + ] + } + ], + "index": 120, + "angle": 0, + "type": "image_caption" + } + ], + "index": 119 + }, + { + "type": "image", + "bbox": [ + 429, + 605, + 472, + 635 + ], + "blocks": [ + { + "bbox": [ + 435, + 600, + 465, + 606 + ], + "lines": [ + { + "bbox": [ + 435, + 600, + 465, + 606 + ], + "spans": [ + { + "bbox": [ + 435, + 600, + 465, + 606 + ], + "type": "text", + "content": "InstructIR" + } + ] + } + ], + "index": 110, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 429, + 605, + 472, + 635 + ], + "lines": [ + { + "bbox": [ + 429, + 605, + 472, + 635 + ], + "spans": [ + { + "bbox": [ + 429, + 605, + 472, + 635 + ], + "type": "image", + "image_path": "8851ecd7763a336e7b162facddfa9332400f1ff5aa192253f535c53dd02ca242.jpg" + } + ] + } + ], + "index": 121, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 635, + 454, + 642 + ], + "lines": [ + { + "bbox": [ + 446, + 635, + 454, + 642 + ], + "spans": [ + { + "bbox": [ + 446, + 635, + 454, + 642 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 122, + "angle": 0, + "type": "image_caption" + } + ], + "index": 121 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "text", + "content": "Conde, Marcos et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 137, + 137, + 480, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 137, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 137, + 137, + 480, + 158 + ], + "type": "text", + "content": "1. Agustsson, E., Timofte, R.: NTIRE 2017 challenge on single image superresolution: Dataset and study. In: CVPR Workshops (2017) 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 159, + 480, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 159, + 480, + 179 + ], + "spans": [ + { + "bbox": [ + 137, + 159, + 480, + 179 + ], + "type": "text", + "content": "2. Arbelaez, P., Maire, M., Fowlkes, C., Malik, J.: Contour detection and hierarchical image segmentation. TPAMI (2011) 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 180, + 480, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 180, + 480, + 222 + ], + "spans": [ + { + "bbox": [ + 138, + 180, + 480, + 222 + ], + "type": "text", + "content": "3. Bai, Y., Wang, C., Xie, S., Dong, C., Yuan, C., Wang, Z.: Textir: A simple framework for text-based editable image restoration. CoRR abs/2302.14736 (2023). https://doi.org/10.48550/ARXIV.2302.14736, https://doi.org/10.48550/arXiv.2302.14736 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 223, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 223, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 223, + 480, + 277 + ], + "type": "text", + "content": "4. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023. pp. 18392-18402. IEEE (2023). https://doi.org/10.1109/CVPR52729.2023.01764, https://doi.org/10.1109/CVPR52729.2023.01764 2, 3, 4, 5, 12" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 277, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 480, + 319 + ], + "type": "text", + "content": "5. Bychkovsky, V., Paris, S., Chan, E., Durand, F.: Learning photographic global tonal adjustment with a database of input / output image pairs. In: The Twenty-Fourth IEEE Conference on Computer Vision and Pattern Recognition (2011) 8, 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 320, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 480, + 342 + ], + "type": "text", + "content": "6. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR (2021) 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 342, + 480, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 362 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 362 + ], + "type": "text", + "content": "7. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV (2022) 2, 3, 6, 7, 9, 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 363, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 363, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 363, + 480, + 396 + ], + "type": "text", + "content": "8. Chen, L., Lu, X., Zhang, J., Chu, X., Chen, C.: Hinet: Half instance normalization network for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 182-192 (2021) 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 396, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 438 + ], + "type": "text", + "content": "9. Chen, Y.S., Wang, Y.C., Kao, M.H., Chuang, Y.Y.: Deep photo enhancer: Unpaired learning for image enhancement from photographs with gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6306-6314 (2018) 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 439, + 480, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 439, + 480, + 524 + ], + "spans": [ + { + "bbox": [ + 134, + 439, + 480, + 524 + ], + "type": "text", + "content": "10. Devlin, J., Chang, M., Lee, K., Toutanova, K.: BERT: pre-training of deep bidirectional transformers for language understanding. In: Burstein, J., Doran, C., Solorio, T. (eds.) Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers). pp. 4171-4186. Association for Computational Linguistics (2019). https://doi.org/10.18653/V1/N19-1423, https://doi.org/10.18653/v1/n19-1423 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 525, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 525, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 134, + 525, + 480, + 568 + ], + "type": "text", + "content": "11. Ding, C., Lu, Z., Wang, S., Cheng, R., Boddeti, V.N.: Mitigating task interference in multi-task learning via explicit task routing with non-learnable primitives. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7756-7765 (2023) 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 569, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 569, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 134, + 569, + 480, + 590 + ], + "type": "text", + "content": "12. Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. TPAMI (2015) 10" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 134, + 590, + 480, + 611 + ], + "type": "text", + "content": "13. Dong, H., Pan, J., Xiang, L., Hu, Z., Zhang, X., Wang, F., Yang, M.H.: Multi-scale boosted dehazing network with dense feature fusion. In: CVPR (2020) 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "type": "text", + "content": "14. Dong, W., Zhang, L., Shi, G., Wu, X.: Image deblurring and super-resolution by adaptive sparse domain selection and adaptive regularization. TIP (2011) 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "15. Dong, Y., Liu, Y., Zhang, H., Chen, S., Qiao, Y.: Fd-gan: Generative adversarial networks with fusion-discriminator for single image dehazing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 10729-10736 (2020) 9" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "InstructIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "16. Elad, M., Feuer, A.: Restoration of a single superresolution image from several blurred, noisy, and undersampled measured images. IEEE transactions on image processing 6(12), 1646-1658 (1997) 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "text", + "content": "17. Fan, Q., Chen, D., Yuan, L., Hua, G., Yu, N., Chen, B.: A general decoupled learning framework for parameterized image operators. IEEE transactions on pattern analysis and machine intelligence 43(1), 33-47 (2019) 9, 13, 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 203 + ], + "type": "text", + "content": "18. Fu, X., Zeng, D., Huang, Y., Liao, Y., Ding, X., Paisley, J.: A fusion-based enhancing method for weakly illuminated images 129, 82-96 (2016) 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 481, + 224 + ], + "type": "text", + "content": "19. Fu, X., Zeng, D., Huang, Y., Zhang, X.P., Ding, X.: A weighted variational model for simultaneous reflectance and illumination estimation. In: CVPR (2016) 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 481, + 246 + ], + "type": "text", + "content": "20. Gao, H., Tao, X., Shen, X., Jia, J.: Dynamic scene deblurring with parameter selective sharing and nested skip connections. In: CVPR. pp. 3848-3856 (2019) 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "text", + "content": "21. Gharbi, M., Chen, J., Barron, J.T., Hasinoff, S.W., Durand, F.: Deep bilateral learning for real-time image enhancement. ACM Transactions on Graphics (TOG) 36(4), 1-12 (2017) 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 300 + ], + "type": "text", + "content": "22. Guo, X., Li, Y., Ling, H.: Lime: Low-light image enhancement via illumination map estimation. IEEE TIP 26(2), 982-993 (2016) 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 300, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 320 + ], + "type": "text", + "content": "23. Hao, S., Han, X., Guo, Y., Xu, X., Wang, M.: Low-light image enhancement with semi-decoupled decomposition. IEEE TMM 22(12), 3025-3038 (2020) 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 320, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 481, + 352 + ], + "type": "text", + "content": "24. He, J., Dong, C., Qiao, Y.: Modulating image restoration with continual levels via adaptive feature modification layers (2019), https://arxiv.org/abs/1904.081186" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 353, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 481, + 374 + ], + "type": "text", + "content": "25. He, K., Sun, J., Tang, X.: Single image haze removal using dark channel prior. TPAMI (2010) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "26. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 407, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 429 + ], + "type": "text", + "content": "27. Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al.: Searching for mobilenetv3. In: ICCV (2019) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 429, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 481, + 460 + ], + "type": "text", + "content": "28. Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5197-5206 (2015) 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "type": "text", + "content": "29. Jiang, Y., Gong, X., Liu, D., Cheng, Y., Fang, C., Shen, X., Yang, J., Zhou, P., Wang, Z.: Enlightengan: Deep light enhancement without paired supervision. IEEE TIP 30, 2340-249 (2021) 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 493, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 481, + 536 + ], + "type": "text", + "content": "30. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "type": "text", + "content": "31. Kim, K.I., Kwon, Y.: Single-image super-resolution using sparse regression and natural image prior. TPAMI (2010) 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 557, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 579 + ], + "type": "text", + "content": "32. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv:1412.6980 (2014) 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 579, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 481, + 611 + ], + "type": "text", + "content": "33. Kopf, J., Neubert, B., Chen, B., Cohen, M., Cohen-Or, D., Deussen, O., Uyttendaele, M., Lischinski, D.: Deep photo: Model-based photograph enhancement and viewing. ACM TOG (2008) 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "text", + "content": "34. Lei, X., Fei, Z., Zhou, W., Zhou, H., Fei, M.: Low-light image enhancement using the cell vibration model. IEEE TMM pp. 1-1 (2022) 11" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "35. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing 28(1), 492-505 (2018) 8, 9, 13, 14" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "text", + "content": "Conde, Marcos et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 663 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "36. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR. pp. 17452-17462 (June 2022) 2, 3, 7, 8, 9, 13, 14" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 481, + 160 + ], + "type": "text", + "content": "37. Li, J., Li, J., Fang, F., Li, F., Zhang, G.: Luminance-aware pyramid network for low-light image enhancement. IEEE TMM 23, 3153-3165 (2020) 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "type": "text", + "content": "38. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: SwinIR: Image restoration using swin transformer. In: ICCV Workshops (2021) 2, 9, 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "text", + "content": "39. Liu, L., Xie, L., Zhang, X., Yuan, S., Chen, X., Zhou, W., Li, H., Tian, Q.: Tape: Task-agnostic prior embedding for image restoration. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XVIII. pp. 447–464. Springer (2022) 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "40. Liu, R., Ma, L., Zhang, J., Fan, X., Luo, Z.: Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In: CVPR (2021) 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 291 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 291 + ], + "type": "text", + "content": "41. Liu, Y., Liu, A., Gu, J., Zhang, Z., Wu, W., Qiao, Y., Dong, C.: Discovering distinctive \"semantics\" in super-resolution networks (2022), https://arxiv.org/abs/2108.00406 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 324 + ], + "type": "text", + "content": "42. Ma, J., Cheng, T., Wang, G., Zhang, Q., Wang, X., Zhang, L.: Prores: Exploring degradation-aware visual prompt for universal image restoration. arXiv preprint arXiv:2306.13653 (2023) 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "text", + "content": "43. Ma, K., Duanmu, Z., Wu, Q., Wang, Z., Yong, H., Li, H., Zhang, L.: Waterloo exploration database: New challenges for image quality assessment models. TIP (2016) 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "text", + "content": "44. Ma, L., Ma, T., Liu, R., Fan, X., Luo, Z.: Toward fast, flexible, and robust low-light image enhancement. In: CVPR (2022) 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "45. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: ICCV (2001) 8, 9, 14" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "text", + "content": "46. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073 (2021) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 447, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 456 + ], + "type": "text", + "content": "47. Michaeli, T., Irani, M.: Nonparametric blind super-resolution. In: ICCV (2013) 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "text", + "content": "48. Moran, S., Marza, P., McDonagh, S., Parisot, S., Slabaugh, G.: Deeplpf: Deep local parametric filters for image enhancement. In: CVPR (2020) 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "49. Mou, C., Wang, Q., Zhang, J.: Deep generalized unfolding networks for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17399-17410 (2022) 9" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "50. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: CVPR (2017) 8, 9, 13, 14" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "text", + "content": "51. Nah, S., Son, S., Lee, J., Lee, K.M.: Clean images are hard to reblur: Exploiting the ill-posed inverse task for dynamic scene deblurring. In: ICLR (2022) 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 556, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 481, + 588 + ], + "type": "text", + "content": "52. Nguyen, N., Milanfar, P., Golub, G.: Efficient generalized cross-validation with applications to parametric image restoration and resolution enhancement. IEEE Transactions on image processing 10(9), 1299-1308 (2001) 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "text", + "content": "53. Park, D., Lee, B.H., Chun, S.Y.: All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5815-5824. IEEE (2023) 2, 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 633, + 481, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 663 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 663 + ], + "type": "text", + "content": "54. Potlapalli, V., Zamir, S.W., Khan, S., Khan, F.S.: Prompt: Prompting for all-in-one blind image restoration. arXiv preprint arXiv:2306.13090 (2023) 2, 3, 8, 9, 13" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "InstructIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 192 + ], + "type": "text", + "content": "55. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event. Proceedings of Machine Learning Research, vol. 139, pp. 8748-8763. PMLR (2021), http://proceedings.mlr.press/v139/radford21a.html 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 194, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 270 + ], + "type": "text", + "content": "56. Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert-networks. In: Inui, K., Jiang, J., Ng, V., Wan, X. (eds.) Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019. pp. 3980-3990. Association for Computational Linguistics (2019). https://doi.org/10.18653/V1/D19-1410, https://doi.org/10.18653/v1/D19-1410 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "text", + "content": "57. Ren, C., He, X., Wang, C., Zhao, Z.: Adaptive consistency prior based deep network for image denoising. In: CVPR (2021) 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "text", + "content": "58. Ren, W., Pan, J., Zhang, H., Cao, X., Yang, M.H.: Single image dehazing via multi-scale convolutional neural networks with holistic edges. IJCV (2020) 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 315, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 369 + ], + "type": "text", + "content": "59. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 10674-10685. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.01042, https://doi.org/10.1109/CVPR52688.2022.01042 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "text", + "content": "60. Ronneberger, O., Fischer, P., Brox, T.: U-Net: convolutional networks for biomedical image segmentation. In: MICCAI (2015) 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "text", + "content": "61. Rosenbaum, C., Klinger, T., Riemer, M.: Routing networks: Adaptive selection of non-linear functions for multi-task learning. arXiv preprint arXiv:1711.01239 (2017) 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 424, + 481, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 467 + ], + "type": "text", + "content": "62. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "text", + "content": "63. Strezoski, G., Noord, N.v., Worring, M.: Many task learning with task routing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1375-1384 (2019) 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 501, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 522 + ], + "type": "text", + "content": "64. Tian, C., Xu, Y., Zuo, W.: Image denoising using deep cnn with batch renormalization. Neural Networks (2020) 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "type": "text", + "content": "65. Timofte, R., De Smet, V., Van Gool, L.: Anchored neighborhood regression for fast example-based super-resolution. In: ICCV (2013) 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 544, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 567 + ], + "type": "text", + "content": "66. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A., Li, Y.: MAXIM: Multi-axis MLP for image processing. In: CVPR. pp. 5769-5780 (2022) 2, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "67. Valanarasu, J.M.J., Yasarla, R., Patel, V.M.: Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In: CVPR. pp. 2353-2363 (2022) 3, 9" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "type": "text", + "content": "68. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "text", + "content": "69. Wang, R., Zhang, Q., Fu, C.W., Shen, X., Zheng, W.S., Jia, J.: Underexposed photo enhancement using deep illumination estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6849-6857 (2019) 10" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "text", + "content": "Conde, Marcos et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 148 + ], + "type": "text", + "content": "70. Wang, S., Zheng, J., Hu, H.M., Li, B.: Naturalness preserved enhancement algorithm for non-uniform illumination images. IEEE TIP 22(9), 3538-3548 (2013) 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "text", + "content": "71. Wang, X., Girshick, R., Gupta, A., He, K.: Non-local neural networks. In: CVPR (2018) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 173, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 481, + 217 + ], + "type": "text", + "content": "72. Wang, Y., Liu, Z., Liu, J., Xu, S., Liu, S.: Low-light image enhancement with illumination-aware gamma correction and complete image modelling network. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13128-13137 (2023) 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "text", + "content": "73. Wang, Z., Cun, X., Bao, J., Liu, J.: Uformer: A general u-shaped transformer for image restoration. arXiv:2106.03106 (2021) 2, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 262 + ], + "type": "text", + "content": "74. Wei, C., Wang, W., Yang, W., Liu, J.: Deep retina decomposition for low-light enhancement. In: British Machine Vision Conference (2018) 8, 9, 11, 13, 14" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 263, + 481, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 481, + 294 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 481, + 294 + ], + "type": "text", + "content": "75. Wu, W., Weng, J., Zhang, P., Wang, X., Yang, W., Jiang, J.: Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement. In: CVPR (2022) 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 296, + 481, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 296, + 481, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 481, + 328 + ], + "type": "text", + "content": "76. Xiao, S., Liu, Z., Zhang, P., Muennighof, N.: C-pack: Packaged resources to advance general chinese embedding. CoRR abs/2309.07597 (2023). https://doi.org/10.48550/ARXIV.2309.07597, https://doi.org/10.48550/arXiv.2309.07597" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 330, + 481, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 351 + ], + "type": "text", + "content": "77. Xu, K., Yang, X., Yin, B., Lau, R.W.: Learning to restore low-light images via decomposition-and-enhancement. In: CVPR (2020) 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 353, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 481, + 373 + ], + "type": "text", + "content": "78. Yang, F., Yang, H., Fu, J., Lu, H., Guo, B.: Learning texture transformer network for image super-resolution. In: CVPR (2020) 8, 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "text", + "content": "79. Yang, W., Wang, S., Fang, Y., Wang, Y., Liu, J.: Band representation-based semi-supervised low-light image enhancement: bridging the gap between signal fidelity and perceptual quality. IEEE TIP 30, 3461-3473 (2021) 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 409, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 481, + 441 + ], + "type": "text", + "content": "80. Yang, W., Wang, W., Huang, H., Wang, S., Liu, J.: Sparse gradient regularized deep retina network for robust low-light image enhancement. IEEE TIP 30, 2072-2086 (2021) 11" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 442, + 481, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 473 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 473 + ], + "type": "text", + "content": "81. Yao, M., Xu, R., Guan, Y., Huang, J., Xiong, Z.: Neural degradation representation learning for all-in-one image restoration. arXiv preprint arXiv:2310.12848 (2023) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "type": "text", + "content": "82. Yu, W., Luo, M., Zhou, P., Si, C., Zhou, Y., Wang, X., Feng, J., Yan, S.: Metaformer is actually what you need for vision. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10819-10829 (2022) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 510, + 481, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 481, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 481, + 540 + ], + "type": "text", + "content": "83. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022) 2, 3, 9" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "type": "text", + "content": "84. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Learning enriched features for real image restoration and enhancement. In: ECCV (2020) 9, 11" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "type": "text", + "content": "85. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR (2021) 9" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 599, + 481, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 631 + ], + "type": "text", + "content": "86. Zeng, H., Cai, J., Li, L., Cao, Z., Zhang, L.: Learning image-adaptive 3d lookup tables for high performance photo enhancement in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(4), 2058-2073 (2020) 10" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "87. Zhang, C., Zhu, Y., Yan, Q., Sun, J., Zhang, Y.: All-in-one multi-degradation image restoration network via hierarchical degradation representation. arXiv preprint arXiv:2308.03021 (2023) 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "InstructIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 293 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "88. Zhang, C., Zhu, Y., Yan, Q., Sun, J., Zhang, Y.: All-in-one multi-degradation image restoration network via hierarchical degradation representation. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 2285-2293 (2023) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 194 + ], + "type": "text", + "content": "89. Zhang, J., Huang, J., Yao, M., Yang, Z., Yu, H., Zhou, M., Zhao, F.: Ingredient-oriented multi-degradation learning for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5825-5835 (2023) 2, 3, 8, 9, 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "90. Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing 26(7), 3142-3155 (2017) 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "text", + "content": "91. Zhang, K., Zuo, W., Gu, S., Zhang, L.: Learning deep CNN denoiser prior for image restoration. In: CVPR (2017) 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "text", + "content": "92. Zhang, K., Luo, W., Zhong, Y., Ma, L., Stenger, B., Liu, W., Li, H.: Deblurring by realistic blurring. In: CVPR. pp. 2737-2746 (2020) 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "text", + "content": "93. Zhang, Y., Zhang, J., Guo, X.: Kindling the darkness: A practical low-light image enhancer. In: ACM MM (2019) 11" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 251, + 101 + ], + "type": "text", + "content": "Conde, Marcos et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Instruction Tuning-free Visual Token Complement for Multimodal LLMs/95b8eb87-0839-45ae-a393-d76dd91f7ccf_content_list.json b/2024/Instruction Tuning-free Visual Token Complement for Multimodal LLMs/95b8eb87-0839-45ae-a393-d76dd91f7ccf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f8e79897a224fda62d1a916e0455b1f83aa3f51c --- /dev/null +++ b/2024/Instruction Tuning-free Visual Token Complement for Multimodal LLMs/95b8eb87-0839-45ae-a393-d76dd91f7ccf_content_list.json @@ -0,0 +1,1549 @@ +[ + { + "type": "text", + "text": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs", + "text_level": 1, + "bbox": [ + 290, + 140, + 712, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dongsheng Wang $^{1}$ , Jiequan Cui $^{2}$ , Miaoge Li $^{3}$ , Wang Lin $^{4}$ , Bo Chen $^{5}$ , and Hanwang Zhang $^{2}$", + "bbox": [ + 269, + 210, + 732, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Shenzhen University, Shenzhen, 518052, P.R. China, dongshengwang@szu.edu.cn", + "bbox": [ + 318, + 253, + 681, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Nanyang Technological University, 50 Nanyang Avenue, 639798, Singapore \n $^{3}$ Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong", + "bbox": [ + 245, + 282, + 754, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Zhejiang University, Hangzhou 310058, P.R. China", + "bbox": [ + 323, + 310, + 676, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ Xidian University, Shaanxi, 710126, P.R. China", + "bbox": [ + 336, + 324, + 666, + 337 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. As the open community of large language models (LLMs) matures, multimodal LLMs (MLLMs) have promised an elegant bridge between vision and language. However, current research is inherently constrained by challenges such as the need for high-quality instruction pairs and the loss of visual information in image-to-text training objectives. To this end, we propose a Visual Token Complement framework (VTC) that helps MLLMs regain the missing visual features and thus improve response accuracy. Specifically, our VTC integrates text-to-image generation as a guide to identifying the text-irrelevant features, and a visual selector is then developed to generate complementary visual tokens to enrich the original visual input. Moreover, an iterative strategy is further designed to extract more visual information by iteratively using the visual selector without any additional training. Notably, the training pipeline requires no additional image-text pairs, resulting in a desired instruction tuning-free property. Both qualitative and quantitative experiments demonstrate the superiority and efficiency of our VTC.", + "bbox": [ + 259, + 377, + 743, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 626, + 375, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large language models (LLMs) have demonstrated extraordinary reasoning ability across various applications [3, 20, 29, 30, 35, 43]. This sparks a research trend that focuses on equipping LLMs with visual prompts to enhance vision-language understanding. To complete various multimodal tasks, we can feed unified vision-language prompts, such as “ An image that shows” or “ What color is the dog?”, into an LLM to generate the required answers such as image captioning or VQA [6, 18, 19, 46]. We designate these multimodal-enhanced Large Language Models as Multimodal LLMs (MLLMs), as our focus is solely on multimodal understanding, rather than on both understanding and/or generation. In particular, the latter is known as Large Multimodal Models (LMMs) [2, 5, 7, 16, 32, 45], which, unfortunately, have been shown to possess a lower understanding capability compared to MLLMs [6, 8, 39].", + "bbox": [ + 212, + 657, + 787, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a80f5e753ca884386f1db8726d574824d1505f510a0e71eb285ea17866aac96e.jpg", + "image_caption": [ + "Fig. 1: Comparisons of our proposed Reconstruction prior (c) for training Visual Token Complement (VTC), Caption prior (a), and Instruction prior (b) for training Visual Prompt Generator (VPG). The attention maps of VPG and VTC are obtained by averaging the query-patch attention weights(Sec. 3.1)." + ], + "image_footnote": [], + "bbox": [ + 263, + 143, + 743, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As \"an image is more than a thousand words\", it is efficient to represent visual prompts as continuous embeddings but not discrete language tokens. For example, BLIP-2 [18] and MiniGPT4 [46] use Q-Former to transform image patches into a fixed number of visual tokens; while LLaVA [19] demonstrates that a simple linear projection can be equally effective instead of using Q-Former. These visual tokens are expected to encode the visual contents in the image, providing informative context for LLMs to generate responses.", + "bbox": [ + 212, + 354, + 787, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The crux of making MLLMs effective is to train a Vision Prompt Generator (VPG) that can generate visual tokens, recognizable by LLMs, from an input image. As VPG is essentially a visual-to-semantic translator, we need a visual-semantic alignment training task to align the visual tokens and their semantics written in language output by a frozen LLM. There are two representative tasks that introduce different visual priors into VPG:", + "bbox": [ + 212, + 460, + 787, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Caption prior: Image captioning is a straightforward alignment task. For instance, the Q-Former VPG in BLIP-2 [18] and MiniGPT4 [46] is trained on large-scale image-caption pairs [27]: the frozen LLM should output the ground-truth caption when taking the visual tokens of the corresponding image as input. However, this task inevitably introduces an inductive bias losing significant details, that is, the VPG tends to only capture the visual semantics which is just enough for captioning the image contents according to the prescribed annotation details in training. As illustrated in Fig. 1(a), as the VPG only focuses on the main object pomelos but neglects the price tag, the MLLM fails to answer the question about price.", + "bbox": [ + 212, + 551, + 787, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instruction prior: Inspired by instruction tuning [22, 34, 42], which turns the general language prior of LLMs—next-token prediction—into more powerful task solvers, InstructBLIP [6] and LLaVA [19] aim to train VPG for task-aware visual tokens. They curate a variety of vision-language tasks such as image caption, visual grounding and VQA into the same template, which is fed into a frozen LLM for training VPG. However, curating instructions at scale and with diversity is particularly costly for vision-language tasks, so the VPG is usually biased to the instruction prior due to overfitted tuning. As shown in Fig. 1(b), the VPG cannot distinguish the price tags of different fruits", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "D. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a9d3857c2d1fdef3ae0446ebde46f70cb9344fbd4458ab5a5883b2cf0d85a6e9.jpg", + "image_caption": [ + "Fig. 2: Iterative complement at the inference stage. Once trained, Our VTC can be reused multiple times to re-pick up the visual concepts omitted by the previous VPG. We showcase three times complements with increasingly detailed output captions." + ], + "image_footnote": [], + "bbox": [ + 271, + 145, + 733, + 291 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "because such price inquiry of multiple objects is rare in the multi-task datasets used in InstructBLIP [6].", + "bbox": [ + 212, + 376, + 782, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we use Reconstruction prior to recover the missing visual semantics by VPG with caption/instruction prior. As shown in Fig. 1(c), we propose a trainable module called Visual Token Complement (VTC) that attends to visual regions missed by VPG. By feeding them to VPG, we can obtain the complementary visual tokens encoding visual semantics not yet captured by VPG. Together with the existing visual tokens, how well the complementary tokens complete the image contents is measured by the image reconstruction loss. We highlight the following three advantages of the proposed VTC:", + "bbox": [ + 212, + 407, + 784, + 526 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Instruction tuning-free. Training VTC is unsupervised. All we need to calculate the reconstruction loss is the input images and three frozen modules: VPG, LLM, and image decoder. No caption pairs are needed to train our VTC.", + "- Completeness on-demand. As illustrated in Fig. 2, thanks to the simple implementation of VTC (Sec. 3.1), we have the flexibility to iteratively generate additional complementary tokens as required. As the VPG is frozen, the iteration can always make up for the visual semantics missed by the VPG in the previous reconstruction. We show that two iterations are sufficient in our experiments (Sec. 4.3).", + "- Vision-language disentanglement. Thanks to the above advantages, if we have instruction tuning data, we can fix the trained VTC and VPG and only tune the LLM for instruction comprehension. As the visual tokens are already completed, such disentanglement can mitigate the vision-language spurious correlation in training data, and utilize the valuable instruction data more efficiently—LLM is more specialized in understanding instructions. We show promising results in Sec. 4.3." + ], + "bbox": [ + 225, + 539, + 784, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To evaluate the proposed VTC, we conducted extensive comparisons on LVLM-eHub [37], MME [8], and DEMON [17] multimodal benchmarks, covering", + "bbox": [ + 215, + 809, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs", + "bbox": [ + 246, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "22 diverse vision tasks in the zero-shot setting. By completing the missing information, our VTC shows significant improvements over its baselines, especially in visual dialogue (showing $45\\%$ improvement over InstructBLIP [6]). Besides, the visualizations of the appended visual tokens, provide an interpretable tool to analyze the complementary semantics.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 243, + 388, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, research focused on extending large language models (LLMs) to accommodate image inputs has garnered increasing attention, primarily due to their significant improvements in various vision-language tasks [9, 17, 21, 36]. They aim to learn cross-modal reasoning from large-scale multimodal data. Pioneering studies like SimVLM [33], Frozen [31], and GiT [32] view image as a foreign language and have achieved exceptional fine-tuning performance under the autoregressive framework. In-context learning series [2, 4, 16, 44] are proposed to align a pre-trained vision encoder with LLMs by the design of multimodal prompt, showing remarkable potential in zero-shot and few-shot tasks. Following that, various VPGs are developed to efficiently align visual features with the language model, e.g., Linear mapping [19] and Q-Former [18]. The follow-up instruction-tuned methods are proposed from various assumptions to help VPGs make full use of prior knowledge implied in LLMs. For instance, LLaVA [19] collects high-quality instruction pairs and finetunes the linear projection layer between the VPG and LLM. InstructBLIP [6] introduces instruction-aware Q-Former and extracts informative features according to the given instruction. Several attempts have also been made to enable MLLMs to understand videos and audiios. They share similar techniques with image-language alignment, yielding impressive performance. The scope of this paper is image-language modalities.", + "bbox": [ + 212, + 273, + 787, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, most existing VPGs align the vision and language domains through autoregressive training objectives, e.g. the image-to-caption loss, which may result in incomplete alignment due to the semantic gaps and noisy matchings in multimodal data [45]. Our approach, in contrast, adopts a text-to-image strategy to guide the learning of visual tokens so that they can reconstruct the input image as closely as possible, resulting in instruction tuning-free visual token complement.", + "bbox": [ + 212, + 561, + 789, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 The Proposed Method", + "text_level": 1, + "bbox": [ + 215, + 688, + 472, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a given image $\\pmb{x}$ , VPGs aim to generate a soft vision prompt $\\pmb{y} = \\{\\pmb{y}_i\\}_{i=1}^N$ containing $N$ visual tokens, where $\\pmb{y}_i \\in \\mathbb{R}^d$ is the $n$ -th token vector and $d$ is the embedding dimension. Without loss of generality, we adopt the Q-Former as the VPG in the following discussion. Q-Former utilizes a fixed number of $N$ query vectors $\\pmb{q}$ to interact with image features $\\pmb{h} = f(\\pmb{x})$ by multiple cross-attention layers, where $f$ is a frozen vision encoder (e.g., ViT-G). The output queries are used as visual tokens $\\pmb{y}$ , which are prepended in the input text embedding of the instruction. To train a VPG, an autoregressive loss is often employed on", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "D. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6001c0f08b2f15d07a4ee027c83d6192f3f68e89061dfdeba58664bcb6659e78.jpg", + "image_caption": [ + "Fig. 3: (a) Training pipeline of the proposed VTC, and (b) structure of VTC." + ], + "image_footnote": [], + "bbox": [ + 254, + 143, + 751, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the predicted text: $p(\\pmb{w}|\\pmb{y}, \\pmb{w}_{\\text{instruct}})$ , where $\\pmb{w}_{\\text{instruct}}$ represents the textual input and $\\pmb{w}$ is the predicted answers. Caption prior-based VPGs typically specify $\\pmb{w}_{\\text{instruct}}$ with the caption template: “ An image that shows”; while the instruction prior-based VPGs tailor various templates to specific tasks, e.g., “ Question: {Question}. Short answer:” for VQA and “ Given the image, generate a question whose answer is: {Answer}. Question:” for visual grounding. VPGs serve as an intermediary between the image and the frozen LLM, with their core idea being to find the optimal $\\pmb{y}$ that provides efficient visual semantics about $\\pmb{x}$ , thereby enabling the LLM to produce accurate responses.", + "bbox": [ + 215, + 356, + 787, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Visual Token Complement", + "text_level": 1, + "bbox": [ + 215, + 529, + 478, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we focus on addressing the issue of information loss during transforming image features into visual tokens, and propose the Visual Token Complement (VTC) framework for existing VPGs. As discussed above, VPGs based on caption/instruction prior often struggle with incompleteness issues, where the generated vision tokens $\\pmb{y}$ are prone to emphasizing main objects at the expense of finer image details. To complete $\\pmb{y}$ and provide comprehensive visual details into LLM, VTC introduces an additional set of vision tokens $\\pmb{y}^{\\prime} = \\{\\pmb{y}_{i}^{\\prime}\\}_{i = 1}^{N}$ by reusing the VPG:", + "bbox": [ + 212, + 553, + 787, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} ^ {\\prime} = \\operatorname {V P G} (\\boldsymbol {c}, \\boldsymbol {q}), \\quad \\boldsymbol {c} = \\operatorname {V T C} (\\boldsymbol {h}, \\boldsymbol {y}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 374, + 685, + 785, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where VTC takes image feature $\\pmb{h}$ , the generated vision tokens $\\pmb{y}$ as inputs and outputs the complementary feature $\\pmb{c}$ . We will introduce the detailed structure of VTC later. $\\pmb{c}$ captures the visual details missed by $\\pmb{y}$ and is fed into the VPG to output the complementary visual tokens. Mathematically, our VTC can be applied with most existing VPGs, e.g., Q-Former and linear projection in LLaVA [19]", + "bbox": [ + 212, + 710, + 787, + 801 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs", + "bbox": [ + 246, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "6 We empirically adopt the Q-Former as our VPG. LLaVA generates 256 tokens for each image, which is too long for training our VTC.", + "bbox": [ + 217, + 810, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once the complementary vision tokens $\\pmb{y}^{\\prime}$ are generated, the final completed vision tokens are obtained as:", + "bbox": [ + 215, + 146, + 787, + 175 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {y}} = \\operatorname {C o n c a t} \\left(\\boldsymbol {y}, \\boldsymbol {y} ^ {\\prime}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 436, + 184, + 785, + 199 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathrm{Concat}(\\cdot)$ denotes applying the concatenating at the token dimension. Like previous MLLMs, $\\hat{\\pmb{y}}$ in VTC will be inserted into the position of the virtual image tokens to form the LLM's input according to different instruction templates. Crucially, the completed $\\hat{\\pmb{y}}$ combines the original caption-aware $\\pmb{y}$ and the newly generated tokens $\\pmb{y}'$ , where the former tends to capture the main visual objects and the latter focuses on the missing details. Together, they offer comprehensive visual evidence and help LLM capture more vision semantics.", + "bbox": [ + 212, + 205, + 782, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, moving beyond the patch embeddings at the last layer of $f$ , VTC aims to explore multi-level semantics and provide hierarchical vision descriptions for LLMs. Let $\\pmb{h}^{(l)}$ denote the patch embeddings at $l$ -th layer of $f$ , VTC pickups the missing features as:", + "bbox": [ + 212, + 311, + 782, + 371 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n[ \\boldsymbol {c} ^ {(l)}, \\boldsymbol {o} ] = \\operatorname {S e l f - A t t} (\\operatorname {C o n c a t} (\\boldsymbol {h} ^ {(l)}, \\boldsymbol {y})), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 378, + 785, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where Self-Att denotes a standard Self-Attention layer, which stacks the self-attention, feeding forward, and residual layers (shown as Fig. 3(b)). $\\pmb{c}^{(l)}$ denotes the complementary feature at layer $l$ . Naturally, we can obtain the multi-layer complementary features $\\pmb{h}$ in Eq. 3 by concatenating $\\pmb{c}^{(l)}$ at different layers, resulting in more comprehensive visual tokens $\\hat{\\pmb{y}}$ .", + "bbox": [ + 212, + 401, + 782, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Reconstruction-prior Training", + "text_level": 1, + "bbox": [ + 215, + 494, + 509, + 510 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inspired by autoencoders [10, 13] that perform representation learning by reconstructing the input, we here introduce the reconstruction-prior training to learn the optimal $\\hat{\\pmb{y}}$ . Specifically, we view $\\hat{\\pmb{y}}$ as the latent embeddings and learn VTC with the image reconstruction task. To complete the encoder-decoder framework, we employ a frozen stable diffusion (SD) [24, 28] as the image decoder $g(\\pmb{x}|\\hat{\\pmb{y}})$ , mapping $\\hat{\\pmb{y}}$ back to the pixel space due to its impressive text-to-image results in the recent AI-generated content field [14, 25, 41].", + "bbox": [ + 212, + 515, + 782, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given $(\\pmb{x},\\hat{\\pmb{y}})$ , SD views $\\hat{\\pmb{y}}$ as the textual condition, and updates it by solving the following Gaussian noise regression as in training the reversed diffusion steps (with $T$ timesteps):", + "bbox": [ + 212, + 621, + 782, + 666 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\hat {\\boldsymbol {y}}} {\\arg \\min } \\mathbb {E} _ {\\boldsymbol {x} _ {0}, \\epsilon , t} [ \\| \\boldsymbol {\\epsilon} - \\epsilon (\\boldsymbol {x} _ {t}, \\hat {\\boldsymbol {y}} \\mathbf {W}) \\| ^ {2} ], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 672, + 785, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\pmb{x}_0 = \\pmb{x}$ , $\\epsilon \\sim \\mathcal{N}(0, I)$ is the Gaussian noise, and $t \\sim [0, T]$ is the sampled timestep. $\\epsilon(\\cdot)$ is the frozen noise prediction network, conditionally on the completed $\\hat{\\pmb{y}}$ , $\\pmb{x}_t$ is noised image at timestep $t$ : $\\pmb{x}_t \\sim \\mathcal{N}(\\sqrt{1 - \\beta_t} \\pmb{x}_{t-1}, \\beta_t \\pmb{I})$ with $\\beta_t \\in (0, 1)$ being the variance. $\\mathbf{W} \\in \\mathbb{R}^{d \\times d_c}$ is a linear mapping that projects the $d$ -dimensional LLM embedding into the $d_c$ -dimensional CLIP space. We pre-calculate $\\mathbf{W}$ via the least squares with orthogonality constraint [23]:", + "bbox": [ + 212, + 702, + 782, + 792 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\min _ {\\mathbf {W}} \\mathbb {E} _ {v \\in V _ {\\mathrm {C L I P}} \\cap V _ {\\mathrm {L L M}}} \\| \\boldsymbol {v} _ {\\mathrm {C L I P}} - \\boldsymbol {v} _ {\\mathrm {L L M}} \\mathbf {W} \\| ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 800, + 785, + 828 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns. t., \\mathbf {W} ^ {T} \\mathbf {W} = \\mathbf {I},\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 825, + 619, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "D. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $V_{\\mathrm{CLIP}}$ and $V_{\\mathrm{LLM}}$ denote the vocabulary sets of CLIP and LLM, respectively, and $\\pmb{v}_{\\mathrm{CLIP}}$ and $\\pmb{v}_{\\mathrm{LLM}}$ are the corresponding CLIP and LLM embedding vector of the shared token $v$ . The orthogonality constraint ensures the learned projection matrix $\\mathbf{W}$ implicitly preserves the structure of the LLM's embedding and regulars the learning of $\\mathbf{W}$ focus on the semantic translation [23, 26].", + "bbox": [ + 212, + 146, + 782, + 222 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unlike previous VPGs that optimize the vision tokens $\\pmb{y}$ by minimizing the image-to-text loss, - with $\\pmb{y}$ inserted into the \"\" part of the instruction. Eq. 4 gives us an elegant solution to learn the optimal $\\hat{\\pmb{y}}$ based on the reconstruction prior. The idea is simple: the vision tokens are complete if they can reconstruct the input image. Leveraging the reconstruction loss, the proposed VTC is trained in an unsupervised manner. Given the frozen image encoder, VPG and SD, VTC is the only to-be-learned module to regain the missing features for better reconstruction. This disentangles the vision-language learning. Since the visual tokens $\\hat{\\pmb{y}}$ are already completed, we can fine-tune the LLM for instruction comprehension on instruction pairs, which is the same as the pre-training stage of LLM, where LLM is more specialized.", + "bbox": [ + 212, + 222, + 784, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Iterative Inference", + "text_level": 1, + "bbox": [ + 215, + 409, + 415, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Once trained with Eq. 4, our VTC in Eq. 3 has the ability to identify the missing features by comparing the generated tokens $\\pmb{y}$ and the image embedding $\\pmb{h}$ . This makes it possible to reuse VTC and then iteratively generate additional complementary tokens as required. Specifically, let $\\pmb{y}^k$ and $\\hat{\\pmb{y}}^k$ denote the generated vision tokens and the completed vision tokens at iteration $k$ , respectively, e.g., $\\pmb{y}^1 = \\pmb{y}'$ and $\\hat{\\pmb{y}}^1 = \\hat{\\pmb{y}}$ . Thus the complementary tokens $\\pmb{y}^{k+1}$ can be obtained as:", + "bbox": [ + 212, + 431, + 782, + 525 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} ^ {k + 1} = \\operatorname {V P G} \\left(\\boldsymbol {c} ^ {k}, \\boldsymbol {q}\\right), \\quad \\boldsymbol {c} ^ {k} = \\operatorname {V T C} \\left(\\boldsymbol {h}, \\hat {\\boldsymbol {y}} ^ {k}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 532, + 784, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\pmb{y}^{k + 1}$ captures the visual details missed by $\\hat{\\pmb{y}}^k$ . Benefiting from the flexible structure of VTC, we can train VTC at $k = 1$ and iteratively reuse it to squeeze the visual message. Empirically, we find two iterations are sufficient, accounting for the computational cost.", + "bbox": [ + 212, + 559, + 782, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 643, + 375, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct extensive experiments on several benchmark multimodal datasets to evaluate the zero-shot performance of VTC against the state-of-the-art MLLMs.", + "bbox": [ + 212, + 674, + 782, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 215, + 724, + 444, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Three multimodal benchmarks. In the experiment, we evaluate the diverse and robust performance of our proposed VTC on three zero-shot multimodal benchmarks, $LVLM-eHub$ [37], $MME$ [8] and $DEMON$ [17]. They measure the visual-language comprehension ability of MLLM from different perspectives. $LVLM-eHub$ is the first multimodal benchmark, which focuses on visual knowledge acquisition, visual reasoning, embodied intelligence, and other 6 categories", + "bbox": [ + 212, + 750, + 784, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs", + "bbox": [ + 245, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of multimodal capabilities. MME is another recently proposed benchmark that measures both the visual perception and cognition abilities of MLLMs across 14 datasets. DEMON focuses on more complex scenarios, where the demonstrative instructions may contain multiple, interleaved, and multimodal inputs. It covers 29 tasks across 7 categories, including multimodal dialogue, visual storytelling, and other multimodal understanding scenarios.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baselines. We compare our VTC with the recent advanced MLLMs on three widely used benchmarks. For all methods, we choose versions with parameter sizes less than 10B, Otter [16], which aligns the image and text under the in-context learning framework; mPLUG-Owl [38] tunes the vision encoder and LLM (with a LoRA module [11]) at different stages; VPGTrans [40], which designs a projector warm-up and vanilla fine-tuning to transfer an existing VPG from one MLLM to the target MLLM; And VPG based MLLMs: LLaVA [19], BLIP2 [18], MiniGPT-4 [46], and InstructBLIP [42], they share similar ideas but varying VPGs and training data. For all methods, we select versions with parameter sizes under 10B. Please refer to the Appendix for details.", + "bbox": [ + 212, + 238, + 787, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training Data. We train our VTC in an unsupervised manner, and we adopt the ImageNet1K as our training dataset.", + "bbox": [ + 212, + 390, + 785, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation Details. We chose three VPG-based MLLMs as our base model: BLIP2, MiniGPT4, and InstructBLIP, and we denote the corresponding variants as VTC(BLIP2), VTC(MiniGPT4), and VTC(InstructBLIP). We follow the same implementation details as the base model. For the SD model, we load the checkpoints of Stable-Diffusion-v1-5. VTC is trained on ImageNet1K with 2 epochs. For the learning rate, we first conduct a linear warm-up from $1e - 6$ to $1e - 4$ , and then use a cosine learning rate schedule with the minimal $lr = 1e - 5$ . We set $l = [12,24,36]$ in Eq. 3 to capture the multi-level visual semantics.", + "bbox": [ + 212, + 421, + 787, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Quantitative Analysis", + "text_level": 1, + "bbox": [ + 215, + 568, + 441, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on LVLM-eHub [37]. We first evaluate the vision-language understanding of our VTC on LVLM-eHub. We here focus on the Visual Reasoning task due to it requires a comprehensive understanding of images and texts. It includes visual question answering (VQA, e.g., DocVQA, TextVQA, STVQA, OCR-VQA, OKVQA, GQA, and IconQA), visual spatial reasoning (VSR), visual dialog (e.g., Visdial), and knowledge-grounded image description (e.g., ScienceQA and VizWiz). We report the evaluation results at Table 1. We have the following remarks about the results. Overall, our proposed VTC outperforms the others consistently on all datasets, except the case where ours is the second on the DocVQA dataset. Moreover, compared with our base models (BLIP2, MiniGPT4, and InstructBLIP), our model performs better in general, especially in visual dialog cases. This is not surprising, as our VTC acts as a complementary module and provides LLM with useful image details, which are usually missed by VPGs. By concatenating the original tokens with the complementary tokens and freezing the whole MLLM, VTC improves VPGs by making full use of the prior knowledge.", + "bbox": [ + 212, + 598, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "D. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/78d9733e7031c8230acfc18ba9e0667fa3b10d564b52b193d7d27aa9a5f4f9be.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsDocTextSTVOCROKGQAVisdialIconVSRWHOOPSScienceQAVizWiz
mPLUG-Owl2.2438.7612.108.8422.8914.0213.3411.6424.7420.702.8038.99
Otter3.4421.5215.2319.5049.0138.1211.6726.776.4015.1427.2250.04
VPGTrans3.5321.9817.1321.7144.5132.999.7038.2248.7715.8836.9953.23
LLaVA6.2638.9228.4023.4054.3641.3014.6642.9551.2424.3949.3362.42
BLIP24.7531.9820.9838.8544.9345.5310.7360.8263.6324.8760.7365.44
VTC(BLIP2)5.3232.025.6738.8946.0848.9777.2062.3163.7726.6563.1066.30
MiniGPT-42.6519.4013.5516.8537.4830.8210.3137.5941.5617.9125.4347.48
VTC (MiniGPT-4)3.0422.7515.5019.5441.7535.1176.6239.0343.7518.5337.8552.10
InstructBLIP5.8939.6028.3060.2060.5249.9645.2056.2541.2830.1346.2665.31
VTC (Instruct)6.1440.5230.7060.2460.9651.1265.7856.8443.1033.7446.2166.31
", + "bbox": [ + 218, + 143, + 789, + 276 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/4642b0a58a0dae2ead3a8cf5c729fe325e5776435852c1ac60582e9cda6a047c.jpg", + "table_caption": [ + "Table 1: Comparison of Zero-shot Performance for various MLLMs on general VQA tasks." + ], + "table_footnote": [], + "table_body": "
DatasetsmPLUG-OwlOtterVPGTransLLaVABLIP2VTC(B)MiniGPT4VTC(M)InstructBLIPVTC(I)
Existence120.00195.0070.0050.00160.00170.0068.3372.00185.00188.00
Count50.0088.3385.0050.00135.00138.3955.0055.00143.33144.01
Position50.0086.6763.3350.0077.3380.5143.3345.7666.6770.84
Color55.00113.3373.3355.00148.33152.9075.0075.00153.33154.51
Poster136.06138.7884.0150.00141.85147.3241.8444.00123.81127.04
Celebrity100.29172.6553.5348.82105.59112.6554.4157.63101.18108.74
Scene135.50158.75141.7550.00145.25150.0071.7575.34153.00155.00
Landmark159.25137.2564.7550.00138.00141.2854.0054.0079.7580.34
Artwork96.25129.0077.2549.00136.50137.4460.5064.39134.25134.97
OCR65.0072.5077.5050.00110.00110.0057.5060.0072.5077.50
Commonsense78.57106.4364.2957.14110.00110.9559.2962.55129.29130.03
Numerical60.0072.5050.0050.0040.0042.0045.0045.0040.0042.00
Text Translation80.0057.5077.5057.5065.0065.005.5060.0065.0067.50
Code Reasoning57.5070.0057.5050.0075.0070.5040.0040.0057.5060.00
", + "bbox": [ + 218, + 333, + 792, + 496 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 2: Evaluation results (\\%) of perception and recognition on MME. The results are calculated by summing ACC and $\\mathrm{ACC + }$ in MME [8]. We shorten VTC variants by keeping the first letter of the corresponding base models.", + "bbox": [ + 215, + 497, + 784, + 537 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Results on MME [8]. MME is another recently proposed benchmark measuring both the visual perception and cognition abilities of MLLMs across 14 datasets. Given a test image, MME designs two prompts, and each of them consists of two parts, including a visual question and an instruction \"Please answer yes or no\". This makes it convenient to report the quantitative performance of accuracy. We report the accuracy score at Table. 2. First, We find that Otter, VTC(B), and VTC(I) achieve 5/14, 3/14, and 3/14 best results, respectively. Otter aligns the vision-language modalities under the in-context learning framework on massive multimodal datasets, while our approach is trained from a single modality. These competitive results indicate our unsupervised training paradigms effectively complete the missing semantics, showing great potential in overcoming the inherent limitation of VPGs. Second, compared to mPLUG-Owl (with 2/14 best results), which finetunes the LLM for multimodal alignments, our approach shows better performance in terms of both perception and cognition. This indicates the efficiency of our image reconstruction loss, which helps VTC complete the image details so that the LLM has sufficient visual evidence for the correct response.", + "bbox": [ + 212, + 582, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs", + "bbox": [ + 245, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e7673d037ef1940d785a4a6f3f00e147bac7d4f7a673bf57337993892c5e9299.jpg", + "image_caption": [ + "Fig. 4: Zero-shot evaluation results on DEMON. We only report our VTC(InstructBLIP) for clear representation." + ], + "image_footnote": [], + "bbox": [ + 223, + 157, + 483, + 297 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/5e9ecb33007ac3188a5bb664d7df4ef3788ea856c81898dcadeb492d7901a7da.jpg", + "image_caption": [ + "Fig. 5: ChatGPT 4(V) evaluation results of various MLLMs with scores ranging from [1-5]." + ], + "image_footnote": [], + "bbox": [ + 509, + 185, + 787, + 272 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results on DEMON [17]. We adopt DEMON as our third vision-instruction understanding benchmark. Different from the first two evaluations, DEMON designs more complex demonstrative instructions, including interleaved, and multimodal inputs. We report the evaluation score in Fig. 4. First, our approach surrounds other baseline methods in all directions, demonstrating robust multimodal understanding. Moreover, our approach obtains $20\\%$ improvement over InstructBLIP in visual relation inference. We attribute this to our reconstruction prior, which offers helpful guidance to help visual tokens memorize the location information of the targets. Lastly, compared to Otter, which shows higher performance on MME, our VTC(InstructBLIP) achieves significant improvements on all tasks. This demonstrates our superior vision-language understanding in complex scenarios, where the reconstruction prior still provides fine-grained guidance, while the captions fail to capture such detailed supervision.", + "bbox": [ + 212, + 398, + 789, + 597 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Further Analysis", + "text_level": 1, + "bbox": [ + 215, + 618, + 401, + 633 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ChatGPT 4(V) evaluation. Besides the above three benchmark evaluations, we further leverage ChatGPT 4(V) [21] to have a systematic measurement of the quality of the proposed model. Specifically, motivated by previous works [1,19] we first randomly select 50 image-instruction pairs from the MIMIC-IT datasets [15] and collect the answers from various MLLMs. ChatGPT 4(V) is then asked to evaluate the quality of answers based on the input image and output the corresponding scores (1-5), where a higher score indicates better overall performance. We report the results in Fig. 5. We find that our VTC(InstructBLIP) is the only model with a score larger than 4, and the second model is our VTC(MiniGPT4). We hope this evaluation results serve as strong evidence for supporting our performance. The evaluation prompt in this section can be found in the Appendix. Information completion of VTC. One of the core ideas behind VTC is to learn complementary visual tokens that capture visual features missed by VPG.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "D. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/ef2b2bfc3320f7586b6d0eec14bb4fce6ad6dc240d178ce0d09593918e821d9e.jpg", + "image_caption": [ + "(a) VTC (MiniGPT4)" + ], + "image_footnote": [], + "bbox": [ + 261, + 143, + 379, + 215 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1c7c0e7c2b40649ca4f710b779a08f4641957a5997681b62babd1ecd82ffeca7.jpg", + "image_caption": [ + "(b) VTC (InstructBLIP)", + "Fig. 6: (a-b) Improvement trends on OKVQA, IconQA, and GQA datasets with different numbers of iterations. (c-d) tag-token nearest neighbor semantic distance $(\\times 100)$ on Flicker dataset with different bumbers of iterations ( $x = 0$ means the original VPG)." + ], + "image_footnote": [], + "bbox": [ + 379, + 143, + 496, + 215 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4365fbf6653016dddedd3bacddd0a9ccc68e593913d213442ce6f6753a22fda1.jpg", + "image_caption": [ + "(c) VTC (MiniGPT4)" + ], + "image_footnote": [], + "bbox": [ + 504, + 143, + 620, + 214 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/471baf434cf71b215f4f745f4f9d255815dce047102259896a155abb7d99123a.jpg", + "image_caption": [ + "(d) VTC (InstructBLIP)" + ], + "image_footnote": [], + "bbox": [ + 620, + 143, + 738, + 214 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/17130b78a29963f704dfca14909692926160e510a0ea2adb73b7c95d730e4169.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsVTC(BLIP2)VTC(InstructBLIP)
BLIP2BLIP2(64)VTCInstructBLIPInstructBLIP(64)VTC
Perception133.79133.95138.23121.42121.86124.31
Cognition73.9473.9675.5473.0273.4475.92
", + "bbox": [ + 274, + 297, + 733, + 358 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 3: Abalation results of fine-tuning with LoRA and the number of visual tokens. We finetune LLMs in our model and baselines with LoRA to test the performance of finetuning with the instruction dataset. (64) means we add 32 visual queries to the original model, thus resulting visual tokens have the same length as our VTC.", + "bbox": [ + 212, + 358, + 785, + 412 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To evaluate such information completion, we here propose the tag-token distance $d$ , which measures the nearest neighbor semantic distance between the image tags and generated visual tokens: $d = \\frac{1}{M} \\sum_{i=1}^{M} \\cosine(\\pmb{t}_i, \\pmb{y}_i)$ , where $M$ is the number of tags, the visual token $\\pmb{y}_i$ is the nearest neighbor of tag $i$ , and $\\cosine(\\cdot)$ is the cosine distance. We adopt the tagged Flicker dataset [12] to measure the semantic distance, where the average number of tags per image is 8.94. We report the results of VTC(MiniGPT4) and VTC(InstructBLIP) in Fig. 6(c-d), where the $x$ -axis denotes the number we reuse the VTC module, and $x = 0$ means the original VPG. We find that VTC significantly reduces the semantic distance. That is, the newly generated tokens are able to capture visual concepts missed by the previous tokens, which chimes with our motivations. We also visualize the vision tokens in Fig. 7, which will be introduced later.", + "bbox": [ + 212, + 460, + 787, + 641 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Impact of the number of iterations. Due to the flexible design of our VTC, we can reuse the VTC module iteratively at the test stage, once the VTC is trained with the first iteration. This makes it possible to increasingly complete the visual tokens on demand. Yet more iterations mean higher computational costs. We report the impact of the number of iterations in Fig. 6. Fig. 6(a-b) denote the normalized VQA results on OKVQA, IconQV, and GQA datasets with different iterations(e.g.,1,2,3), and Fig. 6(c-d) denote the tag-token distance on Flick dataset. Overall, increasing the number of iterations can improve the performance in terms of VQA and tag-token retrievals. This demonstrates that the newly generated complementary tokens indeed capture more visual concepts, enhancing the LLMs. Besides, we observe that two iterations are sufficient and after two, the gain is not very significant. This is reasonable since our frozen SD model forces the visual tokens $\\hat{\\pmb{y}}$ to capture as many features as possible to", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs", + "bbox": [ + 246, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/dcf342fb6d81be695275f18728d7e50d47ce31460614e62d2a46f94f9071b51f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
InstructBLIPTokens:tags:man, grass field, frisbeeCaptions:The image features a shirtless man catching a frisbee in a grassy field. The man is in the air, ready to catch the frisbee and throw it into the air. There are several other people present in the scene, including a woman, a man...
VTC(InstructBLIP)-1Complementary tokens:tags:man, grass field, frisbee, outdoor,jumping, park, carCaptions:The image features a man catching a frisbee in a grassy field. The shirtless man is in the air, preparing to throw the frisbee into the air. There are several other people present in the scene, including a man in a red shirt and a man in a blue shirt. There are several cars in the background, including a red car and a white car...
VTC(InstructBLIP)-2Complementary tokens:tags:man, grass field, frisbee, sunny, tree, sky,outdoor, jumping, parkCaptions:The image features a man catching a frisbee in a grassy field. The shirtless man is in the air, preparing to throw the frisbee into the air. There is a cloudy sky in the distance. The man appears to be enjoying a game of frisbee in a park, with a frisbee in his hands. There are several other people present in the scene, including a woman, a man. There are several cars in the background, including ...
", + "bbox": [ + 220, + 143, + 782, + 281 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2791c93a99b99b0a189e52d227f7b61be06c0ab90531594a0a96619c8cb8dda5.jpg", + "image_caption": [ + "Fig. 7: Visualization of the complementary tokens in the first iteration (VTC(InstructBLIP)-1) and second iteration (VTC(InstructBLIP)-2). We also provide the retrieved tags and predicted captions.", + "Fig.8: Ablation results of the reconstruction quality. The first row denotes the raw input images. $K = 0$ denotes the original InstructBLIP and $K = 1$ denotes we use VTC once to complete the visual tokens." + ], + "image_footnote": [], + "bbox": [ + 274, + 354, + 730, + 520 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "reconstruct the input image in the first iteration, and therefore there is little detail to pick up in the following iterations.", + "bbox": [ + 212, + 609, + 785, + 641 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Vision-language disentanglement. After being trained with reconstruction prior, the completed tokens $\\hat{y}$ provide LLMs with lossless image messages. Thus we can fix the VTC and VPG and only finetune the LLM for instruction learning. This mitigates the spurious correlations between vision translating and instruction understanding, which are two important components of MLLMs and often entangled with each other in previous VPGs. As the visual tokens are already completed, we can focus on the latter by finetuning LLMs. Specifically, We finetune our LLM and baselines with LoRA [11] on the LLaVA Visual Instruct 150K dataset [19] and report the evaluation results on MME benchmark at Table. 3. We find that fine-tuning LLM improves performance considerably further. Recalling that BLIP2 is trained with image-caption pairs and InstuctBLIP tunes BLIP2 with massive instruction pairs, our VTC(BLIP2) with finetuning outperforms VTC(InstructBLIP) on both the perception and cognition tasks. This", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "D. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/44360e4497db92cb0580233778ef9e245ea29b7f658054dabd93a0d301c2f3c7.jpg", + "image_caption": [ + "Fig.9: Examples of the test dialogs. VTC shows impressive vision-language understanding in image captioning, human common sense, and visual reasoning, across single and multi-turns." + ], + "image_footnote": [], + "bbox": [ + 272, + 143, + 725, + 648 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "demonstrates that our approach uses the instruction pairs more efficiently. Besides, we also test the performance of visual query length. Our VTC learns 32 additional tokens in the visual prompt, we thus add 32 visual queries to the original baselines and jointly train those vectors on the instruction pairs. We find that our VTC achieves higher scores on MME. We attribute this to our reconstruction-aware training, which effectively learns missing concepts.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs", + "bbox": [ + 246, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation studies on reconstruction quality. One of the core ideas behind VTC is to learn reconstruction-aware tokens for comprehensive image representation. The proposed VTC utilizes a frozen SD as the image decoder to complete the visual tokens. To measure the performance of the completed visual tokens, We here aim to test the improvements by measuring the quality of the reconstructed images. Specifically, we report the FID scores on the COCO2014 validation dataset. A lower FID score means closer semantics between the reconstructed images and the original images. From Fig. 8 we find that VTC captures more image details, such as the number of persons, the number of airplanes, and the accurate sky background. The higher reconstruction quality demonstrates the efficiency of the newly generated tokens, thus contributing to the responses of LLMs.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.4 Visualization & Case Studies", + "text_level": 1, + "bbox": [ + 215, + 349, + 501, + 364 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Recall that the cross-attention weights between the query embeddings $\\mathbf{q}$ and the patch embeddings $\\mathbf{h}$ in VPG provide us with an interpretable tool to visualize the complementary tokens. Specifically, we average the attention weight over all tokens and visualize the attention map in Fig. 7. In addition to the attention map, we also provide the retrieved tags and the corresponding predicted caption. We find that the complementary tokens successfully attend to the missing regions and different iterations cover diverse visual concepts. This meets with our complement motivation, resulting in more accurate tags and detailed captions.", + "bbox": [ + 212, + 376, + 787, + 496 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We also showcase several test dialogs in Fig. 9. Those examples cover human common sense, reasoning, and knowledge retrieval, across single and multiple turns. For example, our VTC has the ability to reason with human common sense, identify the funny aspect of connecting VGA with smartphones, and the unusual cases in a street picture. Moreover, VTC is able to infer the winner of a tic-tac-toe game and the correct number in a Sudoku puzzle based on pre-trained human knowledge. Those results demonstrate the impressive visual perception ability of our approach.", + "bbox": [ + 212, + 497, + 787, + 617 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 641, + 359, + 657 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We in this paper propose VTC, an instruction tuning-free framework to learn complementary visual tokens that capture image details missed by VPG. VTC leverages text-to-image generation to guide the acquisition of reconstruction-aware vision tokens. The enhanced vision prompt is formed by merging these new tokens with the original ones. Additionally, we have developed an iterative strategy to progressively reclaim missing visual information on demand during the testing phase. Extensive experiments are conducted on three widely used multimodal benchmarks to evaluate the proposed model. A ChatGPT 4(V) evaluation is also included to have a systematic measurement. Visualization and case studies demonstrate the interpretability and impressive vision-language understanding of our model.", + "bbox": [ + 212, + 672, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "D. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant U21B2006; in part by Shaanxi Youth Innovation Team Project; in part by the Fundamental Research Funds for the Central Universities QTZX24003 and QTZX22160; in part by the 111 Project under Grant B18039.", + "bbox": [ + 212, + 176, + 784, + 236 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 258, + 321, + 273 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality (March 2023)", + "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)", + "3. Anil, R., Dai, A.M., First, O., Johnson, M., Lepikhin, D., Passos, A., Shakeri, S., Taropa, E., Bailey, P., Chen, Z., et al.: Palm 2 technical report. arXiv preprint arXiv:2305.10403 (2023)", + "4. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., et al.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)", + "5. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022)", + "6. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.C.H.: Instructclip: Towards general-purpose vision-language models with instruction tuning. CoRR abs/2305.06500 (2023)", + "7. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023)", + "8. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., Li, K., Sun, X., Ji, R.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)", + "9. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023)", + "10. He, K., Chen, X., Xie, S., Li, Y., Dólar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)", + "1. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)", + "2. Huiskes, M.J., Lew, M.S.: The mir flickr retrieval evaluation. In: Proceedings of the 1st ACM international conference on Multimedia information retrieval. pp. 39-43 (2008)", + "3. Kenton, J.D.M.W.C., Toutanova, L.K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of naacL-HLT. vol. 1, p. 2 (2019)" + ], + "bbox": [ + 225, + 287, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs", + "bbox": [ + 245, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439 (2023)", + "15. Li, B., Zhang, Y., Chen, L., Wang, J., Pu, F., Yang, J., Li, C., Liu, Z.: Mimic-it: Multi-modal in-context instruction tuning (2023)", + "16. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023)", + "17. Li, J., Pan, K., Ge, Z., Gao, M., Zhang, H., Ji, W., Zhang, W., Chua, T.S., Tang, S., Zhuang, Y.: Fine-tuning multimodal llms to follow zero-shot demonstrative instructions. arXiv preprint arXiv:2308.04152 (2023)", + "18. Li, J., Li, D., Savarese, S., Hoi, S.C.H.: BLIP-2: bootstrapping language-image pretraining with frozen image encoders and large language models. In: International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA. vol. 202, pp. 19730-19742. PMLR (2023)", + "19. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023)", + "20. OpenAI: GPT-4 technical report. CoRR abs/2303.08774 (2023)", + "21. OpenAI: GPT-4V(ision) system card (2023)", + "22. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback, 2022 13 (2022)", + "23. Paischer, F., Adler, T., Hofmarcher, M., Hochreiter, S.: Sitta: A semantic image-text alignment for image captioning. arXiv preprint arXiv:2307.05591 (2023)", + "24. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "25. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)", + "26. Schonemann, P.H.: A generalized solution of the orthogonal procrustes problem. Psychometrika 31(1), 1-10 (1966)", + "27. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)", + "28. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (2021)", + "29. Surís, D., Menon, S., Vondrick, C.: Vipergpt: Visual inference via python execution for reasoning. arXiv preprint arXiv:2303.08128 (2023)", + "30. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)", + "31. Tsimpoukelli, M., Menick, J.L., Cabi, S., Eslami, S., Vinyals, O., Hill, F.: Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems 34, 200-212 (2021)", + "32. Wang, J., Yang, Z., Hu, X., Li, L., Lin, K., Gan, Z., Liu, Z., Liu, C., Wang, L.: Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100 (2022)", + "33. Wang, Z., Yu, J., Yu, A.W., Dai, Z., Tsvetkov, Y., Cao, Y.: Simvlm: Simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904 (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "D. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Wei, J., Bosma, M., Zhao, V.Y., Guu, K., Yu, A.W., Lester, B., Du, N., Dai, A.M., Le, Q.V.: Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652 (2021)", + "35. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)", + "36. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023)", + "37. Xu, P., Shao, W., Zhang, K., Gao, P., Liu, S., Lei, M., Meng, F., Huang, S., Qiao, Y., Luo, P.: Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. arXiv preprint arXiv:2306.09265 (2023)", + "38. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)", + "39. Yin, S., Fu, C., Zhao, S., Li, K., Sun, X., Xu, T., Chen, E.: A survey on multimodal large language models. arXiv preprint arXiv:2306.13549 (2023)", + "40. Zhang, A., Fei, H., Yao, Y., Ji, W., Li, L., Liu, Z., Chua, T.S.: Transfer visual prompt generator across llms. arXiv preprint arXiv:2305.01278 (2023)", + "41. Zhang, C., Zhang, C., Zhang, M., Kweon, I.S.: Text-to-image diffusion model in generative ai: A survey. arXiv preprint arXiv:2303.07909 (2023)", + "42. Zhang, S., Dong, L., Li, X., Zhang, S., Sun, X., Wang, S., Li, J., Hu, R., Zhang, T., Wu, F., et al.: Instruction tuning for large language models: A survey. arXiv preprint arXiv:2308.10792 (2023)", + "43. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)", + "44. Zhao, H., Cai, Z., Si, S., Ma, X., An, K., Chen, L., Liu, Z., Wang, S., Han, W., Chang, B.: Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915 (2023)", + "45. Zhao, X., Liu, B., Liu, Q., Shi, G., Wu, X.M.: Making multimodal generation easier: When diffusion models meet llms. arXiv preprint arXiv:2310.08949 (2023)", + "46. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 631 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs", + "bbox": [ + 245, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Instruction Tuning-free Visual Token Complement for Multimodal LLMs/95b8eb87-0839-45ae-a393-d76dd91f7ccf_model.json b/2024/Instruction Tuning-free Visual Token Complement for Multimodal LLMs/95b8eb87-0839-45ae-a393-d76dd91f7ccf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..68e8dab11db7ede076c66581452e414f9dd7eb66 --- /dev/null +++ b/2024/Instruction Tuning-free Visual Token Complement for Multimodal LLMs/95b8eb87-0839-45ae-a393-d76dd91f7ccf_model.json @@ -0,0 +1,2137 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.291, + 0.141, + 0.713, + 0.187 + ], + "angle": 0, + "content": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.212, + 0.733, + 0.245 + ], + "angle": 0, + "content": "Dongsheng Wang\\(^{1}\\), Jiequan Cui\\(^{2}\\), Miaoge Li\\(^{3}\\), Wang Lin\\(^{4}\\), Bo Chen\\(^{5}\\), and Hanwang Zhang\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.32, + 0.255, + 0.682, + 0.282 + ], + "angle": 0, + "content": "1 Shenzhen University, Shenzhen, 518052, P.R. China, dongshengwang@szu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.283, + 0.756, + 0.309 + ], + "angle": 0, + "content": "\\(^{2}\\) Nanyang Technological University, 50 Nanyang Avenue, 639798, Singapore \n\\(^{3}\\) Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.311, + 0.678, + 0.325 + ], + "angle": 0, + "content": "4 Zhejiang University, Hangzhou 310058, P.R. China" + }, + { + "type": "text", + "bbox": [ + 0.337, + 0.325, + 0.667, + 0.338 + ], + "angle": 0, + "content": "\\(^{5}\\) Xidian University, Shaanxi, 710126, P.R. China" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.378, + 0.744, + 0.6 + ], + "angle": 0, + "content": "Abstract. As the open community of large language models (LLMs) matures, multimodal LLMs (MLLMs) have promised an elegant bridge between vision and language. However, current research is inherently constrained by challenges such as the need for high-quality instruction pairs and the loss of visual information in image-to-text training objectives. To this end, we propose a Visual Token Complement framework (VTC) that helps MLLMs regain the missing visual features and thus improve response accuracy. Specifically, our VTC integrates text-to-image generation as a guide to identifying the text-irrelevant features, and a visual selector is then developed to generate complementary visual tokens to enrich the original visual input. Moreover, an iterative strategy is further designed to extract more visual information by iteratively using the visual selector without any additional training. Notably, the training pipeline requires no additional image-text pairs, resulting in a desired instruction tuning-free property. Both qualitative and quantitative experiments demonstrate the superiority and efficiency of our VTC." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.627, + 0.376, + 0.643 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Large language models (LLMs) have demonstrated extraordinary reasoning ability across various applications [3, 20, 29, 30, 35, 43]. This sparks a research trend that focuses on equipping LLMs with visual prompts to enhance vision-language understanding. To complete various multimodal tasks, we can feed unified vision-language prompts, such as “ An image that shows” or “ What color is the dog?”, into an LLM to generate the required answers such as image captioning or VQA [6, 18, 19, 46]. We designate these multimodal-enhanced Large Language Models as Multimodal LLMs (MLLMs), as our focus is solely on multimodal understanding, rather than on both understanding and/or generation. In particular, the latter is known as Large Multimodal Models (LMMs) [2, 5, 7, 16, 32, 45], which, unfortunately, have been shown to possess a lower understanding capability compared to MLLMs [6, 8, 39]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.129 + ], + "angle": 0, + "content": "D. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.144, + 0.744, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.325 + ], + "angle": 0, + "content": "Fig. 1: Comparisons of our proposed Reconstruction prior (c) for training Visual Token Complement (VTC), Caption prior (a), and Instruction prior (b) for training Visual Prompt Generator (VPG). The attention maps of VPG and VTC are obtained by averaging the query-patch attention weights(Sec. 3.1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.788, + 0.461 + ], + "angle": 0, + "content": "As \"an image is more than a thousand words\", it is efficient to represent visual prompts as continuous embeddings but not discrete language tokens. For example, BLIP-2 [18] and MiniGPT4 [46] use Q-Former to transform image patches into a fixed number of visual tokens; while LLaVA [19] demonstrates that a simple linear projection can be equally effective instead of using Q-Former. These visual tokens are expected to encode the visual contents in the image, providing informative context for LLMs to generate responses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.552 + ], + "angle": 0, + "content": "The crux of making MLLMs effective is to train a Vision Prompt Generator (VPG) that can generate visual tokens, recognizable by LLMs, from an input image. As VPG is essentially a visual-to-semantic translator, we need a visual-semantic alignment training task to align the visual tokens and their semantics written in language output by a frozen LLM. There are two representative tasks that introduce different visual priors into VPG:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Caption prior: Image captioning is a straightforward alignment task. For instance, the Q-Former VPG in BLIP-2 [18] and MiniGPT4 [46] is trained on large-scale image-caption pairs [27]: the frozen LLM should output the ground-truth caption when taking the visual tokens of the corresponding image as input. However, this task inevitably introduces an inductive bias losing significant details, that is, the VPG tends to only capture the visual semantics which is just enough for captioning the image contents according to the prescribed annotation details in training. As illustrated in Fig. 1(a), as the VPG only focuses on the main object pomelos but neglects the price tag, the MLLM fails to answer the question about price." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Instruction prior: Inspired by instruction tuning [22, 34, 42], which turns the general language prior of LLMs—next-token prediction—into more powerful task solvers, InstructBLIP [6] and LLaVA [19] aim to train VPG for task-aware visual tokens. They curate a variety of vision-language tasks such as image caption, visual grounding and VQA into the same template, which is fed into a frozen LLM for training VPG. However, curating instructions at scale and with diversity is particularly costly for vision-language tasks, so the VPG is usually biased to the instruction prior due to overfitted tuning. As shown in Fig. 1(b), the VPG cannot distinguish the price tags of different fruits" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.247, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.146, + 0.734, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.303, + 0.785, + 0.346 + ], + "angle": 0, + "content": "Fig. 2: Iterative complement at the inference stage. Once trained, Our VTC can be reused multiple times to re-pick up the visual concepts omitted by the previous VPG. We showcase three times complements with increasingly detailed output captions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.784, + 0.405 + ], + "angle": 0, + "content": "because such price inquiry of multiple objects is rare in the multi-task datasets used in InstructBLIP [6]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.785, + 0.527 + ], + "angle": 0, + "content": "In this paper, we use Reconstruction prior to recover the missing visual semantics by VPG with caption/instruction prior. As shown in Fig. 1(c), we propose a trainable module called Visual Token Complement (VTC) that attends to visual regions missed by VPG. By feeding them to VPG, we can obtain the complementary visual tokens encoding visual semantics not yet captured by VPG. Together with the existing visual tokens, how well the complementary tokens complete the image contents is measured by the image reconstruction loss. We highlight the following three advantages of the proposed VTC:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.54, + 0.784, + 0.599 + ], + "angle": 0, + "content": "- Instruction tuning-free. Training VTC is unsupervised. All we need to calculate the reconstruction loss is the input images and three frozen modules: VPG, LLM, and image decoder. No caption pairs are needed to train our VTC." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.601, + 0.785, + 0.691 + ], + "angle": 0, + "content": "- Completeness on-demand. As illustrated in Fig. 2, thanks to the simple implementation of VTC (Sec. 3.1), we have the flexibility to iteratively generate additional complementary tokens as required. As the VPG is frozen, the iteration can always make up for the visual semantics missed by the VPG in the previous reconstruction. We show that two iterations are sufficient in our experiments (Sec. 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.693, + 0.785, + 0.798 + ], + "angle": 0, + "content": "- Vision-language disentanglement. Thanks to the above advantages, if we have instruction tuning data, we can fix the trained VTC and VPG and only tune the LLM for instruction comprehension. As the visual tokens are already completed, such disentanglement can mitigate the vision-language spurious correlation in training data, and utilize the valuable instruction data more efficiently—LLM is more specialized in understanding instructions. We show promising results in Sec. 4.3." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.54, + 0.785, + 0.798 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "To evaluate the proposed VTC, we conducted extensive comparisons on LVLM-eHub [37], MME [8], and DEMON [17] multimodal benchmarks, covering" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "D. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "22 diverse vision tasks in the zero-shot setting. By completing the missing information, our VTC shows significant improvements over its baselines, especially in visual dialogue (showing \\(45\\%\\) improvement over InstructBLIP [6]). Besides, the visualizations of the appended visual tokens, provide an interpretable tool to analyze the complementary semantics." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.244, + 0.39, + 0.26 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.275, + 0.789, + 0.562 + ], + "angle": 0, + "content": "Recently, research focused on extending large language models (LLMs) to accommodate image inputs has garnered increasing attention, primarily due to their significant improvements in various vision-language tasks [9, 17, 21, 36]. They aim to learn cross-modal reasoning from large-scale multimodal data. Pioneering studies like SimVLM [33], Frozen [31], and GiT [32] view image as a foreign language and have achieved exceptional fine-tuning performance under the autoregressive framework. In-context learning series [2, 4, 16, 44] are proposed to align a pre-trained vision encoder with LLMs by the design of multimodal prompt, showing remarkable potential in zero-shot and few-shot tasks. Following that, various VPGs are developed to efficiently align visual features with the language model, e.g., Linear mapping [19] and Q-Former [18]. The follow-up instruction-tuned methods are proposed from various assumptions to help VPGs make full use of prior knowledge implied in LLMs. For instance, LLaVA [19] collects high-quality instruction pairs and finetunes the linear projection layer between the VPG and LLM. InstructBLIP [6] introduces instruction-aware Q-Former and extracts informative features according to the given instruction. Several attempts have also been made to enable MLLMs to understand videos and audiios. They share similar techniques with image-language alignment, yielding impressive performance. The scope of this paper is image-language modalities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.79, + 0.668 + ], + "angle": 0, + "content": "However, most existing VPGs align the vision and language domains through autoregressive training objectives, e.g. the image-to-caption loss, which may result in incomplete alignment due to the semantic gaps and noisy matchings in multimodal data [45]. Our approach, in contrast, adopts a text-to-image strategy to guide the learning of visual tokens so that they can reconstruct the input image as closely as possible, resulting in instruction tuning-free visual token complement." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.689, + 0.473, + 0.708 + ], + "angle": 0, + "content": "3 The Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.789, + 0.841 + ], + "angle": 0, + "content": "For a given image \\( \\pmb{x} \\), VPGs aim to generate a soft vision prompt \\( \\pmb{y} = \\{\\pmb{y}_i\\}_{i=1}^N \\) containing \\( N \\) visual tokens, where \\( \\pmb{y}_i \\in \\mathbb{R}^d \\) is the \\( n \\)-th token vector and \\( d \\) is the embedding dimension. Without loss of generality, we adopt the Q-Former as the VPG in the following discussion. Q-Former utilizes a fixed number of \\( N \\) query vectors \\( \\pmb{q} \\) to interact with image features \\( \\pmb{h} = f(\\pmb{x}) \\) by multiple cross-attention layers, where \\( f \\) is a frozen vision encoder (e.g., ViT-G). The output queries are used as visual tokens \\( \\pmb{y} \\), which are prepended in the input text embedding of the instruction. To train a VPG, an autoregressive loss is often employed on" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.247, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.145, + 0.752, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.313, + 0.761, + 0.328 + ], + "angle": 0, + "content": "Fig. 3: (a) Training pipeline of the proposed VTC, and (b) structure of VTC." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.357, + 0.788, + 0.508 + ], + "angle": 0, + "content": "the predicted text: \\( p(\\pmb{w}|\\pmb{y}, \\pmb{w}_{\\text{instruct}}) \\), where \\( \\pmb{w}_{\\text{instruct}} \\) represents the textual input and \\( \\pmb{w} \\) is the predicted answers. Caption prior-based VPGs typically specify \\( \\pmb{w}_{\\text{instruct}} \\) with the caption template: “ An image that shows”; while the instruction prior-based VPGs tailor various templates to specific tasks, e.g., “ Question: {Question}. Short answer:” for VQA and “ Given the image, generate a question whose answer is: {Answer}. Question:” for visual grounding. VPGs serve as an intermediary between the image and the frozen LLM, with their core idea being to find the optimal \\( \\pmb{y} \\) that provides efficient visual semantics about \\( \\pmb{x} \\), thereby enabling the LLM to produce accurate responses." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.53, + 0.48, + 0.545 + ], + "angle": 0, + "content": "3.1 Visual Token Complement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.674 + ], + "angle": 0, + "content": "In this paper, we focus on addressing the issue of information loss during transforming image features into visual tokens, and propose the Visual Token Complement (VTC) framework for existing VPGs. As discussed above, VPGs based on caption/instruction prior often struggle with incompleteness issues, where the generated vision tokens \\(\\pmb{y}\\) are prone to emphasizing main objects at the expense of finer image details. To complete \\(\\pmb{y}\\) and provide comprehensive visual details into LLM, VTC introduces an additional set of vision tokens \\(\\pmb{y}^{\\prime} = \\{\\pmb{y}_{i}^{\\prime}\\}_{i = 1}^{N}\\) by reusing the VPG:" + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.686, + 0.787, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} ^ {\\prime} = \\operatorname {V P G} (\\boldsymbol {c}, \\boldsymbol {q}), \\quad \\boldsymbol {c} = \\operatorname {V T C} (\\boldsymbol {h}, \\boldsymbol {y}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.788, + 0.803 + ], + "angle": 0, + "content": "where VTC takes image feature \\( \\pmb{h} \\), the generated vision tokens \\( \\pmb{y} \\) as inputs and outputs the complementary feature \\( \\pmb{c} \\). We will introduce the detailed structure of VTC later. \\( \\pmb{c} \\) captures the visual details missed by \\( \\pmb{y} \\) and is fed into the VPG to output the complementary visual tokens. Mathematically, our VTC can be applied with most existing VPGs, e.g., Q-Former and linear projection in LLaVA [19]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "6 We empirically adopt the Q-Former as our VPG. LLaVA generates 256 tokens for each image, which is too long for training our VTC." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.129 + ], + "angle": 0, + "content": "D. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.176 + ], + "angle": 0, + "content": "Once the complementary vision tokens \\(\\pmb{y}^{\\prime}\\) are generated, the final completed vision tokens are obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.437, + 0.185, + 0.786, + 0.2 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {y}} = \\operatorname {C o n c a t} \\left(\\boldsymbol {y}, \\boldsymbol {y} ^ {\\prime}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.206, + 0.784, + 0.311 + ], + "angle": 0, + "content": "where \\(\\mathrm{Concat}(\\cdot)\\) denotes applying the concatenating at the token dimension. Like previous MLLMs, \\(\\hat{\\pmb{y}}\\) in VTC will be inserted into the position of the virtual image tokens to form the LLM's input according to different instruction templates. Crucially, the completed \\(\\hat{\\pmb{y}}\\) combines the original caption-aware \\(\\pmb{y}\\) and the newly generated tokens \\(\\pmb{y}'\\), where the former tends to capture the main visual objects and the latter focuses on the missing details. Together, they offer comprehensive visual evidence and help LLM capture more vision semantics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.784, + 0.372 + ], + "angle": 0, + "content": "In addition, moving beyond the patch embeddings at the last layer of \\( f \\), VTC aims to explore multi-level semantics and provide hierarchical vision descriptions for LLMs. Let \\( \\pmb{h}^{(l)} \\) denote the patch embeddings at \\( l \\)-th layer of \\( f \\), VTC pickups the missing features as:" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.379, + 0.786, + 0.396 + ], + "angle": 0, + "content": "\\[\n[ \\boldsymbol {c} ^ {(l)}, \\boldsymbol {o} ] = \\operatorname {S e l f - A t t} (\\operatorname {C o n c a t} (\\boldsymbol {h} ^ {(l)}, \\boldsymbol {y})), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.784, + 0.478 + ], + "angle": 0, + "content": "where Self-Att denotes a standard Self-Attention layer, which stacks the self-attention, feeding forward, and residual layers (shown as Fig. 3(b)). \\( \\pmb{c}^{(l)} \\) denotes the complementary feature at layer \\( l \\). Naturally, we can obtain the multi-layer complementary features \\( \\pmb{h} \\) in Eq. 3 by concatenating \\( \\pmb{c}^{(l)} \\) at different layers, resulting in more comprehensive visual tokens \\( \\hat{\\pmb{y}} \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.495, + 0.51, + 0.511 + ], + "angle": 0, + "content": "3.2 Reconstruction-prior Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.784, + 0.621 + ], + "angle": 0, + "content": "Inspired by autoencoders [10, 13] that perform representation learning by reconstructing the input, we here introduce the reconstruction-prior training to learn the optimal \\(\\hat{\\pmb{y}}\\). Specifically, we view \\(\\hat{\\pmb{y}}\\) as the latent embeddings and learn VTC with the image reconstruction task. To complete the encoder-decoder framework, we employ a frozen stable diffusion (SD) [24, 28] as the image decoder \\(g(\\pmb{x}|\\hat{\\pmb{y}})\\), mapping \\(\\hat{\\pmb{y}}\\) back to the pixel space due to its impressive text-to-image results in the recent AI-generated content field [14, 25, 41]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.784, + 0.667 + ], + "angle": 0, + "content": "Given \\((\\pmb{x},\\hat{\\pmb{y}})\\), SD views \\(\\hat{\\pmb{y}}\\) as the textual condition, and updates it by solving the following Gaussian noise regression as in training the reversed diffusion steps (with \\(T\\) timesteps):" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.674, + 0.786, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\underset {\\hat {\\boldsymbol {y}}} {\\arg \\min } \\mathbb {E} _ {\\boldsymbol {x} _ {0}, \\epsilon , t} [ \\| \\boldsymbol {\\epsilon} - \\epsilon (\\boldsymbol {x} _ {t}, \\hat {\\boldsymbol {y}} \\mathbf {W}) \\| ^ {2} ], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.784, + 0.794 + ], + "angle": 0, + "content": "where \\(\\pmb{x}_0 = \\pmb{x}\\), \\(\\epsilon \\sim \\mathcal{N}(0, I)\\) is the Gaussian noise, and \\(t \\sim [0, T]\\) is the sampled timestep. \\(\\epsilon(\\cdot)\\) is the frozen noise prediction network, conditionally on the completed \\(\\hat{\\pmb{y}}\\), \\(\\pmb{x}_t\\) is noised image at timestep \\(t\\): \\(\\pmb{x}_t \\sim \\mathcal{N}(\\sqrt{1 - \\beta_t} \\pmb{x}_{t-1}, \\beta_t \\pmb{I})\\) with \\(\\beta_t \\in (0, 1)\\) being the variance. \\(\\mathbf{W} \\in \\mathbb{R}^{d \\times d_c}\\) is a linear mapping that projects the \\(d\\)-dimensional LLM embedding into the \\(d_c\\)-dimensional CLIP space. We pre-calculate \\(\\mathbf{W}\\) via the least squares with orthogonality constraint [23]:" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.801, + 0.786, + 0.829 + ], + "angle": 0, + "content": "\\[\n\\arg \\min _ {\\mathbf {W}} \\mathbb {E} _ {v \\in V _ {\\mathrm {C L I P}} \\cap V _ {\\mathrm {L L M}}} \\| \\boldsymbol {v} _ {\\mathrm {C L I P}} - \\boldsymbol {v} _ {\\mathrm {L L M}} \\mathbf {W} \\| ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.827, + 0.62, + 0.842 + ], + "angle": 0, + "content": "\\[\ns. t., \\mathbf {W} ^ {T} \\mathbf {W} = \\mathbf {I},\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.223 + ], + "angle": 0, + "content": "where \\( V_{\\mathrm{CLIP}} \\) and \\( V_{\\mathrm{LLM}} \\) denote the vocabulary sets of CLIP and LLM, respectively, and \\( \\pmb{v}_{\\mathrm{CLIP}} \\) and \\( \\pmb{v}_{\\mathrm{LLM}} \\) are the corresponding CLIP and LLM embedding vector of the shared token \\( v \\). The orthogonality constraint ensures the learned projection matrix \\( \\mathbf{W} \\) implicitly preserves the structure of the LLM's embedding and regulars the learning of \\( \\mathbf{W} \\) focus on the semantic translation [23, 26]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.389 + ], + "angle": 0, + "content": "Unlike previous VPGs that optimize the vision tokens \\(\\pmb{y}\\) by minimizing the image-to-text loss, - with \\(\\pmb{y}\\) inserted into the \"\" part of the instruction. Eq. 4 gives us an elegant solution to learn the optimal \\(\\hat{\\pmb{y}}\\) based on the reconstruction prior. The idea is simple: the vision tokens are complete if they can reconstruct the input image. Leveraging the reconstruction loss, the proposed VTC is trained in an unsupervised manner. Given the frozen image encoder, VPG and SD, VTC is the only to-be-learned module to regain the missing features for better reconstruction. This disentangles the vision-language learning. Since the visual tokens \\(\\hat{\\pmb{y}}\\) are already completed, we can fine-tune the LLM for instruction comprehension on instruction pairs, which is the same as the pre-training stage of LLM, where LLM is more specialized." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.41, + 0.416, + 0.423 + ], + "angle": 0, + "content": "3.3 Iterative Inference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.784, + 0.526 + ], + "angle": 0, + "content": "Once trained with Eq. 4, our VTC in Eq. 3 has the ability to identify the missing features by comparing the generated tokens \\(\\pmb{y}\\) and the image embedding \\(\\pmb{h}\\). This makes it possible to reuse VTC and then iteratively generate additional complementary tokens as required. Specifically, let \\(\\pmb{y}^k\\) and \\(\\hat{\\pmb{y}}^k\\) denote the generated vision tokens and the completed vision tokens at iteration \\(k\\), respectively, e.g., \\(\\pmb{y}^1 = \\pmb{y}'\\) and \\(\\hat{\\pmb{y}}^1 = \\hat{\\pmb{y}}\\). Thus the complementary tokens \\(\\pmb{y}^{k+1}\\) can be obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.534, + 0.785, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} ^ {k + 1} = \\operatorname {V P G} \\left(\\boldsymbol {c} ^ {k}, \\boldsymbol {q}\\right), \\quad \\boldsymbol {c} ^ {k} = \\operatorname {V T C} \\left(\\boldsymbol {h}, \\hat {\\boldsymbol {y}} ^ {k}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.56, + 0.784, + 0.621 + ], + "angle": 0, + "content": "\\(\\pmb{y}^{k + 1}\\) captures the visual details missed by \\(\\hat{\\pmb{y}}^k\\). Benefiting from the flexible structure of VTC, we can train VTC at \\(k = 1\\) and iteratively reuse it to squeeze the visual message. Empirically, we find two iterations are sufficient, accounting for the computational cost." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.644, + 0.376, + 0.661 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.784, + 0.705 + ], + "angle": 0, + "content": "We conduct extensive experiments on several benchmark multimodal datasets to evaluate the zero-shot performance of VTC against the state-of-the-art MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.726, + 0.446, + 0.741 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Three multimodal benchmarks. In the experiment, we evaluate the diverse and robust performance of our proposed VTC on three zero-shot multimodal benchmarks, \\(LVLM-eHub\\) [37], \\(MME\\) [8] and \\(DEMON\\) [17]. They measure the visual-language comprehension ability of MLLM from different perspectives. \\(LVLM-eHub\\) is the first multimodal benchmark, which focuses on visual knowledge acquisition, visual reasoning, embodied intelligence, and other 6 categories" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "D. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "of multimodal capabilities. MME is another recently proposed benchmark that measures both the visual perception and cognition abilities of MLLMs across 14 datasets. DEMON focuses on more complex scenarios, where the demonstrative instructions may contain multiple, interleaved, and multimodal inputs. It covers 29 tasks across 7 categories, including multimodal dialogue, visual storytelling, and other multimodal understanding scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.788, + 0.389 + ], + "angle": 0, + "content": "Baselines. We compare our VTC with the recent advanced MLLMs on three widely used benchmarks. For all methods, we choose versions with parameter sizes less than 10B, Otter [16], which aligns the image and text under the in-context learning framework; mPLUG-Owl [38] tunes the vision encoder and LLM (with a LoRA module [11]) at different stages; VPGTrans [40], which designs a projector warm-up and vanilla fine-tuning to transfer an existing VPG from one MLLM to the target MLLM; And VPG based MLLMs: LLaVA [19], BLIP2 [18], MiniGPT-4 [46], and InstructBLIP [42], they share similar ideas but varying VPGs and training data. For all methods, we select versions with parameter sizes under 10B. Please refer to the Appendix for details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.787, + 0.42 + ], + "angle": 0, + "content": "Training Data. We train our VTC in an unsupervised manner, and we adopt the ImageNet1K as our training dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.788, + 0.543 + ], + "angle": 0, + "content": "Implementation Details. We chose three VPG-based MLLMs as our base model: BLIP2, MiniGPT4, and InstructBLIP, and we denote the corresponding variants as VTC(BLIP2), VTC(MiniGPT4), and VTC(InstructBLIP). We follow the same implementation details as the base model. For the SD model, we load the checkpoints of Stable-Diffusion-v1-5. VTC is trained on ImageNet1K with 2 epochs. For the learning rate, we first conduct a linear warm-up from \\( 1e - 6 \\) to \\( 1e - 4 \\), and then use a cosine learning rate schedule with the minimal \\( lr = 1e - 5 \\). We set \\( l = [12,24,36] \\) in Eq. 3 to capture the multi-level visual semantics." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.569, + 0.442, + 0.584 + ], + "angle": 0, + "content": "4.2 Quantitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Results on LVLM-eHub [37]. We first evaluate the vision-language understanding of our VTC on LVLM-eHub. We here focus on the Visual Reasoning task due to it requires a comprehensive understanding of images and texts. It includes visual question answering (VQA, e.g., DocVQA, TextVQA, STVQA, OCR-VQA, OKVQA, GQA, and IconQA), visual spatial reasoning (VSR), visual dialog (e.g., Visdial), and knowledge-grounded image description (e.g., ScienceQA and VizWiz). We report the evaluation results at Table 1. We have the following remarks about the results. Overall, our proposed VTC outperforms the others consistently on all datasets, except the case where ours is the second on the DocVQA dataset. Moreover, compared with our base models (BLIP2, MiniGPT4, and InstructBLIP), our model performs better in general, especially in visual dialog cases. This is not surprising, as our VTC acts as a complementary module and provides LLM with useful image details, which are usually missed by VPGs. By concatenating the original tokens with the complementary tokens and freezing the whole MLLM, VTC improves VPGs by making full use of the prior knowledge." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.145, + 0.79, + 0.277 + ], + "angle": 0, + "content": "
MethodsDocTextSTVOCROKGQAVisdialIconVSRWHOOPSScienceQAVizWiz
mPLUG-Owl2.2438.7612.108.8422.8914.0213.3411.6424.7420.702.8038.99
Otter3.4421.5215.2319.5049.0138.1211.6726.776.4015.1427.2250.04
VPGTrans3.5321.9817.1321.7144.5132.999.7038.2248.7715.8836.9953.23
LLaVA6.2638.9228.4023.4054.3641.3014.6642.9551.2424.3949.3362.42
BLIP24.7531.9820.9838.8544.9345.5310.7360.8263.6324.8760.7365.44
VTC(BLIP2)5.3232.025.6738.8946.0848.9777.2062.3163.7726.6563.1066.30
MiniGPT-42.6519.4013.5516.8537.4830.8210.3137.5941.5617.9125.4347.48
VTC (MiniGPT-4)3.0422.7515.5019.5441.7535.1176.6239.0343.7518.5337.8552.10
InstructBLIP5.8939.6028.3060.2060.5249.9645.2056.2541.2830.1346.2665.31
VTC (Instruct)6.1440.5230.7060.2460.9651.1265.7856.8443.1033.7446.2166.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.278, + 0.785, + 0.303 + ], + "angle": 0, + "content": "Table 1: Comparison of Zero-shot Performance for various MLLMs on general VQA tasks." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.334, + 0.793, + 0.497 + ], + "angle": 0, + "content": "
DatasetsmPLUG-OwlOtterVPGTransLLaVABLIP2VTC(B)MiniGPT4VTC(M)InstructBLIPVTC(I)
Existence120.00195.0070.0050.00160.00170.0068.3372.00185.00188.00
Count50.0088.3385.0050.00135.00138.3955.0055.00143.33144.01
Position50.0086.6763.3350.0077.3380.5143.3345.7666.6770.84
Color55.00113.3373.3355.00148.33152.9075.0075.00153.33154.51
Poster136.06138.7884.0150.00141.85147.3241.8444.00123.81127.04
Celebrity100.29172.6553.5348.82105.59112.6554.4157.63101.18108.74
Scene135.50158.75141.7550.00145.25150.0071.7575.34153.00155.00
Landmark159.25137.2564.7550.00138.00141.2854.0054.0079.7580.34
Artwork96.25129.0077.2549.00136.50137.4460.5064.39134.25134.97
OCR65.0072.5077.5050.00110.00110.0057.5060.0072.5077.50
Commonsense78.57106.4364.2957.14110.00110.9559.2962.55129.29130.03
Numerical60.0072.5050.0050.0040.0042.0045.0045.0040.0042.00
Text Translation80.0057.5077.5057.5065.0065.005.5060.0065.0067.50
Code Reasoning57.5070.0057.5050.0075.0070.5040.0040.0057.5060.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.498, + 0.785, + 0.539 + ], + "angle": 0, + "content": "Table 2: Evaluation results (\\%) of perception and recognition on MME. The results are calculated by summing ACC and \\(\\mathrm{ACC + }\\) in MME [8]. We shorten VTC variants by keeping the first letter of the corresponding base models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Results on MME [8]. MME is another recently proposed benchmark measuring both the visual perception and cognition abilities of MLLMs across 14 datasets. Given a test image, MME designs two prompts, and each of them consists of two parts, including a visual question and an instruction \"Please answer yes or no\". This makes it convenient to report the quantitative performance of accuracy. We report the accuracy score at Table. 2. First, We find that Otter, VTC(B), and VTC(I) achieve 5/14, 3/14, and 3/14 best results, respectively. Otter aligns the vision-language modalities under the in-context learning framework on massive multimodal datasets, while our approach is trained from a single modality. These competitive results indicate our unsupervised training paradigms effectively complete the missing semantics, showing great potential in overcoming the inherent limitation of VPGs. Second, compared to mPLUG-Owl (with 2/14 best results), which finetunes the LLM for multimodal alignments, our approach shows better performance in terms of both perception and cognition. This indicates the efficiency of our image reconstruction loss, which helps VTC complete the image details so that the LLM has sufficient visual evidence for the correct response." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "D. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.158, + 0.485, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.315, + 0.495, + 0.37 + ], + "angle": 0, + "content": "Fig. 4: Zero-shot evaluation results on DEMON. We only report our VTC(InstructBLIP) for clear representation." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.186, + 0.789, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.285, + 0.789, + 0.329 + ], + "angle": 0, + "content": "Fig. 5: ChatGPT 4(V) evaluation results of various MLLMs with scores ranging from [1-5]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.79, + 0.598 + ], + "angle": 0, + "content": "Results on DEMON [17]. We adopt DEMON as our third vision-instruction understanding benchmark. Different from the first two evaluations, DEMON designs more complex demonstrative instructions, including interleaved, and multimodal inputs. We report the evaluation score in Fig. 4. First, our approach surrounds other baseline methods in all directions, demonstrating robust multimodal understanding. Moreover, our approach obtains \\(20\\%\\) improvement over InstructBLIP in visual relation inference. We attribute this to our reconstruction prior, which offers helpful guidance to help visual tokens memorize the location information of the targets. Lastly, compared to Otter, which shows higher performance on MME, our VTC(InstructBLIP) achieves significant improvements on all tasks. This demonstrates our superior vision-language understanding in complex scenarios, where the reconstruction prior still provides fine-grained guidance, while the captions fail to capture such detailed supervision." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.619, + 0.403, + 0.635 + ], + "angle": 0, + "content": "4.3 Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.842 + ], + "angle": 0, + "content": "ChatGPT 4(V) evaluation. Besides the above three benchmark evaluations, we further leverage ChatGPT 4(V) [21] to have a systematic measurement of the quality of the proposed model. Specifically, motivated by previous works [1,19] we first randomly select 50 image-instruction pairs from the MIMIC-IT datasets [15] and collect the answers from various MLLMs. ChatGPT 4(V) is then asked to evaluate the quality of answers based on the input image and output the corresponding scores (1-5), where a higher score indicates better overall performance. We report the results in Fig. 5. We find that our VTC(InstructBLIP) is the only model with a score larger than 4, and the second model is our VTC(MiniGPT4). We hope this evaluation results serve as strong evidence for supporting our performance. The evaluation prompt in this section can be found in the Appendix. Information completion of VTC. One of the core ideas behind VTC is to learn complementary visual tokens that capture visual features missed by VPG." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.247, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.145, + 0.38, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.292, + 0.217, + 0.369, + 0.227 + ], + "angle": 0, + "content": "(a) VTC (MiniGPT4)" + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.145, + 0.498, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.406, + 0.217, + 0.493, + 0.227 + ], + "angle": 0, + "content": "(b) VTC (InstructBLIP)" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.144, + 0.622, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.217, + 0.612, + 0.227 + ], + "angle": 0, + "content": "(c) VTC (MiniGPT4)" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.144, + 0.74, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.217, + 0.735, + 0.227 + ], + "angle": 0, + "content": "(d) VTC (InstructBLIP)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.238, + 0.784, + 0.28 + ], + "angle": 0, + "content": "Fig. 6: (a-b) Improvement trends on OKVQA, IconQA, and GQA datasets with different numbers of iterations. (c-d) tag-token nearest neighbor semantic distance \\((\\times 100)\\) on Flicker dataset with different bumbers of iterations (\\(x = 0\\) means the original VPG)." + }, + { + "type": "table", + "bbox": [ + 0.275, + 0.298, + 0.734, + 0.359 + ], + "angle": 0, + "content": "
DatasetsVTC(BLIP2)VTC(InstructBLIP)
BLIP2BLIP2(64)VTCInstructBLIPInstructBLIP(64)VTC
Perception133.79133.95138.23121.42121.86124.31
Cognition73.9473.9675.5473.0273.4475.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.359, + 0.787, + 0.414 + ], + "angle": 0, + "content": "Table 3: Abalation results of fine-tuning with LoRA and the number of visual tokens. We finetune LLMs in our model and baselines with LoRA to test the performance of finetuning with the instruction dataset. (64) means we add 32 visual queries to the original model, thus resulting visual tokens have the same length as our VTC." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.642 + ], + "angle": 0, + "content": "To evaluate such information completion, we here propose the tag-token distance \\( d \\), which measures the nearest neighbor semantic distance between the image tags and generated visual tokens: \\( d = \\frac{1}{M} \\sum_{i=1}^{M} \\cosine(\\pmb{t}_i, \\pmb{y}_i) \\), where \\( M \\) is the number of tags, the visual token \\( \\pmb{y}_i \\) is the nearest neighbor of tag \\( i \\), and \\( \\cosine(\\cdot) \\) is the cosine distance. We adopt the tagged Flicker dataset [12] to measure the semantic distance, where the average number of tags per image is 8.94. We report the results of VTC(MiniGPT4) and VTC(InstructBLIP) in Fig. 6(c-d), where the \\( x \\)-axis denotes the number we reuse the VTC module, and \\( x = 0 \\) means the original VPG. We find that VTC significantly reduces the semantic distance. That is, the newly generated tokens are able to capture visual concepts missed by the previous tokens, which chimes with our motivations. We also visualize the vision tokens in Fig. 7, which will be introduced later." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Impact of the number of iterations. Due to the flexible design of our VTC, we can reuse the VTC module iteratively at the test stage, once the VTC is trained with the first iteration. This makes it possible to increasingly complete the visual tokens on demand. Yet more iterations mean higher computational costs. We report the impact of the number of iterations in Fig. 6. Fig. 6(a-b) denote the normalized VQA results on OKVQA, IconQV, and GQA datasets with different iterations(e.g.,1,2,3), and Fig. 6(c-d) denote the tag-token distance on Flick dataset. Overall, increasing the number of iterations can improve the performance in terms of VQA and tag-token retrievals. This demonstrates that the newly generated complementary tokens indeed capture more visual concepts, enhancing the LLMs. Besides, we observe that two iterations are sufficient and after two, the gain is not very significant. This is reasonable since our frozen SD model forces the visual tokens \\(\\hat{\\pmb{y}}\\) to capture as many features as possible to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "D. Wang et al." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.145, + 0.784, + 0.282 + ], + "angle": 0, + "content": "
InstructBLIPTokens:tags:man, grass field, frisbeeCaptions:The image features a shirtless man catching a frisbee in a grassy field. The man is in the air, ready to catch the frisbee and throw it into the air. There are several other people present in the scene, including a woman, a man...
VTC(InstructBLIP)-1Complementary tokens:tags:man, grass field, frisbee, outdoor,jumping, park, carCaptions:The image features a man catching a frisbee in a grassy field. The shirtless man is in the air, preparing to throw the frisbee into the air. There are several other people present in the scene, including a man in a red shirt and a man in a blue shirt. There are several cars in the background, including a red car and a white car...
VTC(InstructBLIP)-2Complementary tokens:tags:man, grass field, frisbee, sunny, tree, sky,outdoor, jumping, parkCaptions:The image features a man catching a frisbee in a grassy field. The shirtless man is in the air, preparing to throw the frisbee into the air. There is a cloudy sky in the distance. The man appears to be enjoying a game of frisbee in a park, with a frisbee in his hands. There are several other people present in the scene, including a woman, a man. There are several cars in the background, including ...
" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.294, + 0.788, + 0.337 + ], + "angle": 0, + "content": "Fig. 7: Visualization of the complementary tokens in the first iteration (VTC(InstructBLIP)-1) and second iteration (VTC(InstructBLIP)-2). We also provide the retrieved tags and predicted captions." + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.356, + 0.731, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.533, + 0.788, + 0.577 + ], + "angle": 0, + "content": "Fig.8: Ablation results of the reconstruction quality. The first row denotes the raw input images. \\( K = 0 \\) denotes the original InstructBLIP and \\( K = 1 \\) denotes we use VTC once to complete the visual tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.787, + 0.642 + ], + "angle": 0, + "content": "reconstruct the input image in the first iteration, and therefore there is little detail to pick up in the following iterations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Vision-language disentanglement. After being trained with reconstruction prior, the completed tokens \\(\\hat{y}\\) provide LLMs with lossless image messages. Thus we can fix the VTC and VPG and only finetune the LLM for instruction learning. This mitigates the spurious correlations between vision translating and instruction understanding, which are two important components of MLLMs and often entangled with each other in previous VPGs. As the visual tokens are already completed, we can focus on the latter by finetuning LLMs. Specifically, We finetune our LLM and baselines with LoRA [11] on the LLaVA Visual Instruct 150K dataset [19] and report the evaluation results on MME benchmark at Table. 3. We find that fine-tuning LLM improves performance considerably further. Recalling that BLIP2 is trained with image-caption pairs and InstuctBLIP tunes BLIP2 with massive instruction pairs, our VTC(BLIP2) with finetuning outperforms VTC(InstructBLIP) on both the perception and cognition tasks. This" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.248, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.145, + 0.727, + 0.649 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.661, + 0.788, + 0.703 + ], + "angle": 0, + "content": "Fig.9: Examples of the test dialogs. VTC shows impressive vision-language understanding in image captioning, human common sense, and visual reasoning, across single and multi-turns." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "demonstrates that our approach uses the instruction pairs more efficiently. Besides, we also test the performance of visual query length. Our VTC learns 32 additional tokens in the visual prompt, we thus add 32 visual queries to the original baselines and jointly train those vectors on the instruction pairs. We find that our VTC achieves higher scores on MME. We attribute this to our reconstruction-aware training, which effectively learns missing concepts." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "D. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "Ablation studies on reconstruction quality. One of the core ideas behind VTC is to learn reconstruction-aware tokens for comprehensive image representation. The proposed VTC utilizes a frozen SD as the image decoder to complete the visual tokens. To measure the performance of the completed visual tokens, We here aim to test the improvements by measuring the quality of the reconstructed images. Specifically, we report the FID scores on the COCO2014 validation dataset. A lower FID score means closer semantics between the reconstructed images and the original images. From Fig. 8 we find that VTC captures more image details, such as the number of persons, the number of airplanes, and the accurate sky background. The higher reconstruction quality demonstrates the efficiency of the newly generated tokens, thus contributing to the responses of LLMs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.351, + 0.502, + 0.366 + ], + "angle": 0, + "content": "4.4 Visualization & Case Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.788, + 0.497 + ], + "angle": 0, + "content": "Recall that the cross-attention weights between the query embeddings \\( \\mathbf{q} \\) and the patch embeddings \\( \\mathbf{h} \\) in VPG provide us with an interpretable tool to visualize the complementary tokens. Specifically, we average the attention weight over all tokens and visualize the attention map in Fig. 7. In addition to the attention map, we also provide the retrieved tags and the corresponding predicted caption. We find that the complementary tokens successfully attend to the missing regions and different iterations cover diverse visual concepts. This meets with our complement motivation, resulting in more accurate tags and detailed captions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.788, + 0.618 + ], + "angle": 0, + "content": "We also showcase several test dialogs in Fig. 9. Those examples cover human common sense, reasoning, and knowledge retrieval, across single and multiple turns. For example, our VTC has the ability to reason with human common sense, identify the funny aspect of connecting VGA with smartphones, and the unusual cases in a street picture. Moreover, VTC is able to infer the winner of a tic-tac-toe game and the correct number in a Sudoku puzzle based on pre-trained human knowledge. Those results demonstrate the impressive visual perception ability of our approach." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.642, + 0.36, + 0.658 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We in this paper propose VTC, an instruction tuning-free framework to learn complementary visual tokens that capture image details missed by VPG. VTC leverages text-to-image generation to guide the acquisition of reconstruction-aware vision tokens. The enhanced vision prompt is formed by merging these new tokens with the original ones. Additionally, we have developed an iterative strategy to progressively reclaim missing visual information on demand during the testing phase. Extensive experiments are conducted on three widely used multimodal benchmarks to evaluate the proposed model. A ChatGPT 4(V) evaluation is also included to have a systematic measurement. Visualization and case studies demonstrate the interpretability and impressive vision-language understanding of our model." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.785, + 0.237 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant U21B2006; in part by Shaanxi Youth Innovation Team Project; in part by the Fundamental Research Funds for the Central Universities QTZX24003 and QTZX22160; in part by the 111 Project under Grant B18039." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.259, + 0.323, + 0.275 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.317 + ], + "angle": 0, + "content": "1. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality (March 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.372 + ], + "angle": 0, + "content": "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.373, + 0.785, + 0.413 + ], + "angle": 0, + "content": "3. Anil, R., Dai, A.M., First, O., Johnson, M., Lepikhin, D., Passos, A., Shakeri, S., Taropa, E., Bailey, P., Chen, Z., et al.: Palm 2 technical report. arXiv preprint arXiv:2305.10403 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.414, + 0.785, + 0.468 + ], + "angle": 0, + "content": "4. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., et al.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.469, + 0.785, + 0.51 + ], + "angle": 0, + "content": "5. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.785, + 0.551 + ], + "angle": 0, + "content": "6. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.C.H.: Instructclip: Towards general-purpose vision-language models with instruction tuning. CoRR abs/2305.06500 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.552, + 0.785, + 0.593 + ], + "angle": 0, + "content": "7. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.594, + 0.785, + 0.635 + ], + "angle": 0, + "content": "8. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., Li, K., Sun, X., Ji, R.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.635, + 0.785, + 0.675 + ], + "angle": 0, + "content": "9. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.676, + 0.785, + 0.716 + ], + "angle": 0, + "content": "10. He, K., Chen, X., Xie, S., Li, Y., Dólar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.758 + ], + "angle": 0, + "content": "1. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "2. Huiskes, M.J., Lew, M.S.: The mir flickr retrieval evaluation. In: Proceedings of the 1st ACM international conference on Multimedia information retrieval. pp. 39-43 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "3. Kenton, J.D.M.W.C., Toutanova, L.K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of naacL-HLT. vol. 1, p. 2 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "D. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "14. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "15. Li, B., Zhang, Y., Chen, L., Wang, J., Pu, F., Yang, J., Li, C., Liu, Z.: Mimic-it: Multi-modal in-context instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.785, + 0.243 + ], + "angle": 0, + "content": "16. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.785, + 0.284 + ], + "angle": 0, + "content": "17. Li, J., Pan, K., Ge, Z., Gao, M., Zhang, H., Ji, W., Zhang, W., Chua, T.S., Tang, S., Zhuang, Y.: Fine-tuning multimodal llms to follow zero-shot demonstrative instructions. arXiv preprint arXiv:2308.04152 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.284, + 0.785, + 0.338 + ], + "angle": 0, + "content": "18. Li, J., Li, D., Savarese, S., Hoi, S.C.H.: BLIP-2: bootstrapping language-image pretraining with frozen image encoders and large language models. In: International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA. vol. 202, pp. 19730-19742. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.339, + 0.688, + 0.352 + ], + "angle": 0, + "content": "19. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.352, + 0.681, + 0.366 + ], + "angle": 0, + "content": "20. OpenAI: GPT-4 technical report. CoRR abs/2303.08774 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.366, + 0.537, + 0.378 + ], + "angle": 0, + "content": "21. OpenAI: GPT-4V(ision) system card (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.378, + 0.785, + 0.418 + ], + "angle": 0, + "content": "22. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback, 2022 13 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.419, + 0.785, + 0.446 + ], + "angle": 0, + "content": "23. Paischer, F., Adler, T., Hofmarcher, M., Hochreiter, S.: Sitta: A semantic image-text alignment for image captioning. arXiv preprint arXiv:2307.05591 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.446, + 0.785, + 0.487 + ], + "angle": 0, + "content": "24. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.487, + 0.785, + 0.541 + ], + "angle": 0, + "content": "25. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.541, + 0.785, + 0.568 + ], + "angle": 0, + "content": "26. Schonemann, P.H.: A generalized solution of the orthogonal procrustes problem. Psychometrika 31(1), 1-10 (1966)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.568, + 0.785, + 0.609 + ], + "angle": 0, + "content": "27. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.609, + 0.785, + 0.651 + ], + "angle": 0, + "content": "28. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.651, + 0.785, + 0.677 + ], + "angle": 0, + "content": "29. Surís, D., Menon, S., Vondrick, C.: Vipergpt: Visual inference via python execution for reasoning. arXiv preprint arXiv:2303.08128 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.677, + 0.785, + 0.717 + ], + "angle": 0, + "content": "30. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "31. Tsimpoukelli, M., Menick, J.L., Cabi, S., Eslami, S., Vinyals, O., Hill, F.: Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems 34, 200-212 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "32. Wang, J., Yang, Z., Hu, X., Li, L., Lin, K., Gan, Z., Liu, Z., Liu, C., Wang, L.: Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "33. Wang, Z., Yu, J., Yu, A.W., Dai, Z., Tsvetkov, Y., Cao, Y.: Simvlm: Simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "34. Wei, J., Bosma, M., Zhao, V.Y., Guu, K., Yu, A.W., Lester, B., Du, N., Dai, A.M., Le, Q.V.: Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "35. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.785, + 0.272 + ], + "angle": 0, + "content": "36. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.272, + 0.785, + 0.314 + ], + "angle": 0, + "content": "37. Xu, P., Shao, W., Zhang, K., Gao, P., Liu, S., Lei, M., Meng, F., Huang, S., Qiao, Y., Luo, P.: Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. arXiv preprint arXiv:2306.09265 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.314, + 0.785, + 0.355 + ], + "angle": 0, + "content": "38. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.383 + ], + "angle": 0, + "content": "39. Yin, S., Fu, C., Zhao, S., Li, K., Sun, X., Xu, T., Chen, E.: A survey on multimodal large language models. arXiv preprint arXiv:2306.13549 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.785, + 0.411 + ], + "angle": 0, + "content": "40. Zhang, A., Fei, H., Yao, Y., Ji, W., Li, L., Liu, Z., Chua, T.S.: Transfer visual prompt generator across llms. arXiv preprint arXiv:2305.01278 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.439 + ], + "angle": 0, + "content": "41. Zhang, C., Zhang, C., Zhang, M., Kweon, I.S.: Text-to-image diffusion model in generative ai: A survey. arXiv preprint arXiv:2303.07909 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "42. Zhang, S., Dong, L., Li, X., Zhang, S., Sun, X., Wang, S., Li, J., Hu, R., Zhang, T., Wu, F., et al.: Instruction tuning for large language models: A survey. arXiv preprint arXiv:2308.10792 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "43. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "44. Zhao, H., Cai, Z., Si, S., Ma, X., An, K., Chen, L., Liu, Z., Wang, S., Han, W., Chang, B.: Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.592 + ], + "angle": 0, + "content": "45. Zhao, X., Liu, B., Liu, Q., Shi, G., Wu, X.M.: Making multimodal generation easier: When diffusion models meet llms. arXiv preprint arXiv:2310.08949 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.632 + ], + "angle": 0, + "content": "46. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.632 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Instruction Tuning-free Visual Token Complement for Multimodal LLMs/95b8eb87-0839-45ae-a393-d76dd91f7ccf_origin.pdf b/2024/Instruction Tuning-free Visual Token Complement for Multimodal LLMs/95b8eb87-0839-45ae-a393-d76dd91f7ccf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4f1ca92207ea36703bf826bfbbde82b37aad8e84 --- /dev/null +++ b/2024/Instruction Tuning-free Visual Token Complement for Multimodal LLMs/95b8eb87-0839-45ae-a393-d76dd91f7ccf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a4f13e0d3ff07e1167ee2068a8276236cc92978d852042e90551fb4a34c0138 +size 3996440 diff --git a/2024/Instruction Tuning-free Visual Token Complement for Multimodal LLMs/full.md b/2024/Instruction Tuning-free Visual Token Complement for Multimodal LLMs/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d0539aab57903d2a02f07bf2011b078b7013cd04 --- /dev/null +++ b/2024/Instruction Tuning-free Visual Token Complement for Multimodal LLMs/full.md @@ -0,0 +1,264 @@ +# Instruction Tuning-free Visual Token Complement for Multimodal LLMs + +Dongsheng Wang $^{1}$ , Jiequan Cui $^{2}$ , Miaoge Li $^{3}$ , Wang Lin $^{4}$ , Bo Chen $^{5}$ , and Hanwang Zhang $^{2}$ + +1 Shenzhen University, Shenzhen, 518052, P.R. China, dongshengwang@szu.edu.cn + +$^{2}$ Nanyang Technological University, 50 Nanyang Avenue, 639798, Singapore + $^{3}$ Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong + +4 Zhejiang University, Hangzhou 310058, P.R. China + +$^{5}$ Xidian University, Shaanxi, 710126, P.R. China + +Abstract. As the open community of large language models (LLMs) matures, multimodal LLMs (MLLMs) have promised an elegant bridge between vision and language. However, current research is inherently constrained by challenges such as the need for high-quality instruction pairs and the loss of visual information in image-to-text training objectives. To this end, we propose a Visual Token Complement framework (VTC) that helps MLLMs regain the missing visual features and thus improve response accuracy. Specifically, our VTC integrates text-to-image generation as a guide to identifying the text-irrelevant features, and a visual selector is then developed to generate complementary visual tokens to enrich the original visual input. Moreover, an iterative strategy is further designed to extract more visual information by iteratively using the visual selector without any additional training. Notably, the training pipeline requires no additional image-text pairs, resulting in a desired instruction tuning-free property. Both qualitative and quantitative experiments demonstrate the superiority and efficiency of our VTC. + +# 1 Introduction + +Large language models (LLMs) have demonstrated extraordinary reasoning ability across various applications [3, 20, 29, 30, 35, 43]. This sparks a research trend that focuses on equipping LLMs with visual prompts to enhance vision-language understanding. To complete various multimodal tasks, we can feed unified vision-language prompts, such as “ An image that shows” or “ What color is the dog?”, into an LLM to generate the required answers such as image captioning or VQA [6, 18, 19, 46]. We designate these multimodal-enhanced Large Language Models as Multimodal LLMs (MLLMs), as our focus is solely on multimodal understanding, rather than on both understanding and/or generation. In particular, the latter is known as Large Multimodal Models (LMMs) [2, 5, 7, 16, 32, 45], which, unfortunately, have been shown to possess a lower understanding capability compared to MLLMs [6, 8, 39]. + +![](images/a80f5e753ca884386f1db8726d574824d1505f510a0e71eb285ea17866aac96e.jpg) +Fig. 1: Comparisons of our proposed Reconstruction prior (c) for training Visual Token Complement (VTC), Caption prior (a), and Instruction prior (b) for training Visual Prompt Generator (VPG). The attention maps of VPG and VTC are obtained by averaging the query-patch attention weights(Sec. 3.1). + +As "an image is more than a thousand words", it is efficient to represent visual prompts as continuous embeddings but not discrete language tokens. For example, BLIP-2 [18] and MiniGPT4 [46] use Q-Former to transform image patches into a fixed number of visual tokens; while LLaVA [19] demonstrates that a simple linear projection can be equally effective instead of using Q-Former. These visual tokens are expected to encode the visual contents in the image, providing informative context for LLMs to generate responses. + +The crux of making MLLMs effective is to train a Vision Prompt Generator (VPG) that can generate visual tokens, recognizable by LLMs, from an input image. As VPG is essentially a visual-to-semantic translator, we need a visual-semantic alignment training task to align the visual tokens and their semantics written in language output by a frozen LLM. There are two representative tasks that introduce different visual priors into VPG: + +Caption prior: Image captioning is a straightforward alignment task. For instance, the Q-Former VPG in BLIP-2 [18] and MiniGPT4 [46] is trained on large-scale image-caption pairs [27]: the frozen LLM should output the ground-truth caption when taking the visual tokens of the corresponding image as input. However, this task inevitably introduces an inductive bias losing significant details, that is, the VPG tends to only capture the visual semantics which is just enough for captioning the image contents according to the prescribed annotation details in training. As illustrated in Fig. 1(a), as the VPG only focuses on the main object pomelos but neglects the price tag, the MLLM fails to answer the question about price. + +Instruction prior: Inspired by instruction tuning [22, 34, 42], which turns the general language prior of LLMs—next-token prediction—into more powerful task solvers, InstructBLIP [6] and LLaVA [19] aim to train VPG for task-aware visual tokens. They curate a variety of vision-language tasks such as image caption, visual grounding and VQA into the same template, which is fed into a frozen LLM for training VPG. However, curating instructions at scale and with diversity is particularly costly for vision-language tasks, so the VPG is usually biased to the instruction prior due to overfitted tuning. As shown in Fig. 1(b), the VPG cannot distinguish the price tags of different fruits + +![](images/a9d3857c2d1fdef3ae0446ebde46f70cb9344fbd4458ab5a5883b2cf0d85a6e9.jpg) +Fig. 2: Iterative complement at the inference stage. Once trained, Our VTC can be reused multiple times to re-pick up the visual concepts omitted by the previous VPG. We showcase three times complements with increasingly detailed output captions. + +because such price inquiry of multiple objects is rare in the multi-task datasets used in InstructBLIP [6]. + +In this paper, we use Reconstruction prior to recover the missing visual semantics by VPG with caption/instruction prior. As shown in Fig. 1(c), we propose a trainable module called Visual Token Complement (VTC) that attends to visual regions missed by VPG. By feeding them to VPG, we can obtain the complementary visual tokens encoding visual semantics not yet captured by VPG. Together with the existing visual tokens, how well the complementary tokens complete the image contents is measured by the image reconstruction loss. We highlight the following three advantages of the proposed VTC: + +- Instruction tuning-free. Training VTC is unsupervised. All we need to calculate the reconstruction loss is the input images and three frozen modules: VPG, LLM, and image decoder. No caption pairs are needed to train our VTC. +- Completeness on-demand. As illustrated in Fig. 2, thanks to the simple implementation of VTC (Sec. 3.1), we have the flexibility to iteratively generate additional complementary tokens as required. As the VPG is frozen, the iteration can always make up for the visual semantics missed by the VPG in the previous reconstruction. We show that two iterations are sufficient in our experiments (Sec. 4.3). +- Vision-language disentanglement. Thanks to the above advantages, if we have instruction tuning data, we can fix the trained VTC and VPG and only tune the LLM for instruction comprehension. As the visual tokens are already completed, such disentanglement can mitigate the vision-language spurious correlation in training data, and utilize the valuable instruction data more efficiently—LLM is more specialized in understanding instructions. We show promising results in Sec. 4.3. + +To evaluate the proposed VTC, we conducted extensive comparisons on LVLM-eHub [37], MME [8], and DEMON [17] multimodal benchmarks, covering + +22 diverse vision tasks in the zero-shot setting. By completing the missing information, our VTC shows significant improvements over its baselines, especially in visual dialogue (showing $45\%$ improvement over InstructBLIP [6]). Besides, the visualizations of the appended visual tokens, provide an interpretable tool to analyze the complementary semantics. + +# 2 Related Work + +Recently, research focused on extending large language models (LLMs) to accommodate image inputs has garnered increasing attention, primarily due to their significant improvements in various vision-language tasks [9, 17, 21, 36]. They aim to learn cross-modal reasoning from large-scale multimodal data. Pioneering studies like SimVLM [33], Frozen [31], and GiT [32] view image as a foreign language and have achieved exceptional fine-tuning performance under the autoregressive framework. In-context learning series [2, 4, 16, 44] are proposed to align a pre-trained vision encoder with LLMs by the design of multimodal prompt, showing remarkable potential in zero-shot and few-shot tasks. Following that, various VPGs are developed to efficiently align visual features with the language model, e.g., Linear mapping [19] and Q-Former [18]. The follow-up instruction-tuned methods are proposed from various assumptions to help VPGs make full use of prior knowledge implied in LLMs. For instance, LLaVA [19] collects high-quality instruction pairs and finetunes the linear projection layer between the VPG and LLM. InstructBLIP [6] introduces instruction-aware Q-Former and extracts informative features according to the given instruction. Several attempts have also been made to enable MLLMs to understand videos and audiios. They share similar techniques with image-language alignment, yielding impressive performance. The scope of this paper is image-language modalities. + +However, most existing VPGs align the vision and language domains through autoregressive training objectives, e.g. the image-to-caption loss, which may result in incomplete alignment due to the semantic gaps and noisy matchings in multimodal data [45]. Our approach, in contrast, adopts a text-to-image strategy to guide the learning of visual tokens so that they can reconstruct the input image as closely as possible, resulting in instruction tuning-free visual token complement. + +# 3 The Proposed Method + +For a given image $\pmb{x}$ , VPGs aim to generate a soft vision prompt $\pmb{y} = \{\pmb{y}_i\}_{i=1}^N$ containing $N$ visual tokens, where $\pmb{y}_i \in \mathbb{R}^d$ is the $n$ -th token vector and $d$ is the embedding dimension. Without loss of generality, we adopt the Q-Former as the VPG in the following discussion. Q-Former utilizes a fixed number of $N$ query vectors $\pmb{q}$ to interact with image features $\pmb{h} = f(\pmb{x})$ by multiple cross-attention layers, where $f$ is a frozen vision encoder (e.g., ViT-G). The output queries are used as visual tokens $\pmb{y}$ , which are prepended in the input text embedding of the instruction. To train a VPG, an autoregressive loss is often employed on + +![](images/6001c0f08b2f15d07a4ee027c83d6192f3f68e89061dfdeba58664bcb6659e78.jpg) +Fig. 3: (a) Training pipeline of the proposed VTC, and (b) structure of VTC. + +the predicted text: $p(\pmb{w}|\pmb{y}, \pmb{w}_{\text{instruct}})$ , where $\pmb{w}_{\text{instruct}}$ represents the textual input and $\pmb{w}$ is the predicted answers. Caption prior-based VPGs typically specify $\pmb{w}_{\text{instruct}}$ with the caption template: “ An image that shows”; while the instruction prior-based VPGs tailor various templates to specific tasks, e.g., “ Question: {Question}. Short answer:” for VQA and “ Given the image, generate a question whose answer is: {Answer}. Question:” for visual grounding. VPGs serve as an intermediary between the image and the frozen LLM, with their core idea being to find the optimal $\pmb{y}$ that provides efficient visual semantics about $\pmb{x}$ , thereby enabling the LLM to produce accurate responses. + +# 3.1 Visual Token Complement + +In this paper, we focus on addressing the issue of information loss during transforming image features into visual tokens, and propose the Visual Token Complement (VTC) framework for existing VPGs. As discussed above, VPGs based on caption/instruction prior often struggle with incompleteness issues, where the generated vision tokens $\pmb{y}$ are prone to emphasizing main objects at the expense of finer image details. To complete $\pmb{y}$ and provide comprehensive visual details into LLM, VTC introduces an additional set of vision tokens $\pmb{y}^{\prime} = \{\pmb{y}_{i}^{\prime}\}_{i = 1}^{N}$ by reusing the VPG: + +$$ +\boldsymbol {y} ^ {\prime} = \operatorname {V P G} (\boldsymbol {c}, \boldsymbol {q}), \quad \boldsymbol {c} = \operatorname {V T C} (\boldsymbol {h}, \boldsymbol {y}), \tag {1} +$$ + +where VTC takes image feature $\pmb{h}$ , the generated vision tokens $\pmb{y}$ as inputs and outputs the complementary feature $\pmb{c}$ . We will introduce the detailed structure of VTC later. $\pmb{c}$ captures the visual details missed by $\pmb{y}$ and is fed into the VPG to output the complementary visual tokens. Mathematically, our VTC can be applied with most existing VPGs, e.g., Q-Former and linear projection in LLaVA [19] + +Once the complementary vision tokens $\pmb{y}^{\prime}$ are generated, the final completed vision tokens are obtained as: + +$$ +\hat {\boldsymbol {y}} = \operatorname {C o n c a t} \left(\boldsymbol {y}, \boldsymbol {y} ^ {\prime}\right), \tag {2} +$$ + +where $\mathrm{Concat}(\cdot)$ denotes applying the concatenating at the token dimension. Like previous MLLMs, $\hat{\pmb{y}}$ in VTC will be inserted into the position of the virtual image tokens to form the LLM's input according to different instruction templates. Crucially, the completed $\hat{\pmb{y}}$ combines the original caption-aware $\pmb{y}$ and the newly generated tokens $\pmb{y}'$ , where the former tends to capture the main visual objects and the latter focuses on the missing details. Together, they offer comprehensive visual evidence and help LLM capture more vision semantics. + +In addition, moving beyond the patch embeddings at the last layer of $f$ , VTC aims to explore multi-level semantics and provide hierarchical vision descriptions for LLMs. Let $\pmb{h}^{(l)}$ denote the patch embeddings at $l$ -th layer of $f$ , VTC pickups the missing features as: + +$$ +[ \boldsymbol {c} ^ {(l)}, \boldsymbol {o} ] = \operatorname {S e l f - A t t} (\operatorname {C o n c a t} (\boldsymbol {h} ^ {(l)}, \boldsymbol {y})), \tag {3} +$$ + +where Self-Att denotes a standard Self-Attention layer, which stacks the self-attention, feeding forward, and residual layers (shown as Fig. 3(b)). $\pmb{c}^{(l)}$ denotes the complementary feature at layer $l$ . Naturally, we can obtain the multi-layer complementary features $\pmb{h}$ in Eq. 3 by concatenating $\pmb{c}^{(l)}$ at different layers, resulting in more comprehensive visual tokens $\hat{\pmb{y}}$ . + +# 3.2 Reconstruction-prior Training + +Inspired by autoencoders [10, 13] that perform representation learning by reconstructing the input, we here introduce the reconstruction-prior training to learn the optimal $\hat{\pmb{y}}$ . Specifically, we view $\hat{\pmb{y}}$ as the latent embeddings and learn VTC with the image reconstruction task. To complete the encoder-decoder framework, we employ a frozen stable diffusion (SD) [24, 28] as the image decoder $g(\pmb{x}|\hat{\pmb{y}})$ , mapping $\hat{\pmb{y}}$ back to the pixel space due to its impressive text-to-image results in the recent AI-generated content field [14, 25, 41]. + +Given $(\pmb{x},\hat{\pmb{y}})$ , SD views $\hat{\pmb{y}}$ as the textual condition, and updates it by solving the following Gaussian noise regression as in training the reversed diffusion steps (with $T$ timesteps): + +$$ +\underset {\hat {\boldsymbol {y}}} {\arg \min } \mathbb {E} _ {\boldsymbol {x} _ {0}, \epsilon , t} [ \| \boldsymbol {\epsilon} - \epsilon (\boldsymbol {x} _ {t}, \hat {\boldsymbol {y}} \mathbf {W}) \| ^ {2} ], \tag {4} +$$ + +where $\pmb{x}_0 = \pmb{x}$ , $\epsilon \sim \mathcal{N}(0, I)$ is the Gaussian noise, and $t \sim [0, T]$ is the sampled timestep. $\epsilon(\cdot)$ is the frozen noise prediction network, conditionally on the completed $\hat{\pmb{y}}$ , $\pmb{x}_t$ is noised image at timestep $t$ : $\pmb{x}_t \sim \mathcal{N}(\sqrt{1 - \beta_t} \pmb{x}_{t-1}, \beta_t \pmb{I})$ with $\beta_t \in (0, 1)$ being the variance. $\mathbf{W} \in \mathbb{R}^{d \times d_c}$ is a linear mapping that projects the $d$ -dimensional LLM embedding into the $d_c$ -dimensional CLIP space. We pre-calculate $\mathbf{W}$ via the least squares with orthogonality constraint [23]: + +$$ +\arg \min _ {\mathbf {W}} \mathbb {E} _ {v \in V _ {\mathrm {C L I P}} \cap V _ {\mathrm {L L M}}} \| \boldsymbol {v} _ {\mathrm {C L I P}} - \boldsymbol {v} _ {\mathrm {L L M}} \mathbf {W} \| ^ {2}, \tag {5} +$$ + +$$ +s. t., \mathbf {W} ^ {T} \mathbf {W} = \mathbf {I}, +$$ + +where $V_{\mathrm{CLIP}}$ and $V_{\mathrm{LLM}}$ denote the vocabulary sets of CLIP and LLM, respectively, and $\pmb{v}_{\mathrm{CLIP}}$ and $\pmb{v}_{\mathrm{LLM}}$ are the corresponding CLIP and LLM embedding vector of the shared token $v$ . The orthogonality constraint ensures the learned projection matrix $\mathbf{W}$ implicitly preserves the structure of the LLM's embedding and regulars the learning of $\mathbf{W}$ focus on the semantic translation [23, 26]. + +Unlike previous VPGs that optimize the vision tokens $\pmb{y}$ by minimizing the image-to-text loss, - with $\pmb{y}$ inserted into the "" part of the instruction. Eq. 4 gives us an elegant solution to learn the optimal $\hat{\pmb{y}}$ based on the reconstruction prior. The idea is simple: the vision tokens are complete if they can reconstruct the input image. Leveraging the reconstruction loss, the proposed VTC is trained in an unsupervised manner. Given the frozen image encoder, VPG and SD, VTC is the only to-be-learned module to regain the missing features for better reconstruction. This disentangles the vision-language learning. Since the visual tokens $\hat{\pmb{y}}$ are already completed, we can fine-tune the LLM for instruction comprehension on instruction pairs, which is the same as the pre-training stage of LLM, where LLM is more specialized. + +# 3.3 Iterative Inference + +Once trained with Eq. 4, our VTC in Eq. 3 has the ability to identify the missing features by comparing the generated tokens $\pmb{y}$ and the image embedding $\pmb{h}$ . This makes it possible to reuse VTC and then iteratively generate additional complementary tokens as required. Specifically, let $\pmb{y}^k$ and $\hat{\pmb{y}}^k$ denote the generated vision tokens and the completed vision tokens at iteration $k$ , respectively, e.g., $\pmb{y}^1 = \pmb{y}'$ and $\hat{\pmb{y}}^1 = \hat{\pmb{y}}$ . Thus the complementary tokens $\pmb{y}^{k+1}$ can be obtained as: + +$$ +\boldsymbol {y} ^ {k + 1} = \operatorname {V P G} \left(\boldsymbol {c} ^ {k}, \boldsymbol {q}\right), \quad \boldsymbol {c} ^ {k} = \operatorname {V T C} \left(\boldsymbol {h}, \hat {\boldsymbol {y}} ^ {k}\right). \tag {6} +$$ + +$\pmb{y}^{k + 1}$ captures the visual details missed by $\hat{\pmb{y}}^k$ . Benefiting from the flexible structure of VTC, we can train VTC at $k = 1$ and iteratively reuse it to squeeze the visual message. Empirically, we find two iterations are sufficient, accounting for the computational cost. + +# 4 Experiments + +We conduct extensive experiments on several benchmark multimodal datasets to evaluate the zero-shot performance of VTC against the state-of-the-art MLLMs. + +# 4.1 Experimental Settings + +Three multimodal benchmarks. In the experiment, we evaluate the diverse and robust performance of our proposed VTC on three zero-shot multimodal benchmarks, $LVLM-eHub$ [37], $MME$ [8] and $DEMON$ [17]. They measure the visual-language comprehension ability of MLLM from different perspectives. $LVLM-eHub$ is the first multimodal benchmark, which focuses on visual knowledge acquisition, visual reasoning, embodied intelligence, and other 6 categories + +of multimodal capabilities. MME is another recently proposed benchmark that measures both the visual perception and cognition abilities of MLLMs across 14 datasets. DEMON focuses on more complex scenarios, where the demonstrative instructions may contain multiple, interleaved, and multimodal inputs. It covers 29 tasks across 7 categories, including multimodal dialogue, visual storytelling, and other multimodal understanding scenarios. + +Baselines. We compare our VTC with the recent advanced MLLMs on three widely used benchmarks. For all methods, we choose versions with parameter sizes less than 10B, Otter [16], which aligns the image and text under the in-context learning framework; mPLUG-Owl [38] tunes the vision encoder and LLM (with a LoRA module [11]) at different stages; VPGTrans [40], which designs a projector warm-up and vanilla fine-tuning to transfer an existing VPG from one MLLM to the target MLLM; And VPG based MLLMs: LLaVA [19], BLIP2 [18], MiniGPT-4 [46], and InstructBLIP [42], they share similar ideas but varying VPGs and training data. For all methods, we select versions with parameter sizes under 10B. Please refer to the Appendix for details. + +Training Data. We train our VTC in an unsupervised manner, and we adopt the ImageNet1K as our training dataset. + +Implementation Details. We chose three VPG-based MLLMs as our base model: BLIP2, MiniGPT4, and InstructBLIP, and we denote the corresponding variants as VTC(BLIP2), VTC(MiniGPT4), and VTC(InstructBLIP). We follow the same implementation details as the base model. For the SD model, we load the checkpoints of Stable-Diffusion-v1-5. VTC is trained on ImageNet1K with 2 epochs. For the learning rate, we first conduct a linear warm-up from $1e - 6$ to $1e - 4$ , and then use a cosine learning rate schedule with the minimal $lr = 1e - 5$ . We set $l = [12,24,36]$ in Eq. 3 to capture the multi-level visual semantics. + +# 4.2 Quantitative Analysis + +Results on LVLM-eHub [37]. We first evaluate the vision-language understanding of our VTC on LVLM-eHub. We here focus on the Visual Reasoning task due to it requires a comprehensive understanding of images and texts. It includes visual question answering (VQA, e.g., DocVQA, TextVQA, STVQA, OCR-VQA, OKVQA, GQA, and IconQA), visual spatial reasoning (VSR), visual dialog (e.g., Visdial), and knowledge-grounded image description (e.g., ScienceQA and VizWiz). We report the evaluation results at Table 1. We have the following remarks about the results. Overall, our proposed VTC outperforms the others consistently on all datasets, except the case where ours is the second on the DocVQA dataset. Moreover, compared with our base models (BLIP2, MiniGPT4, and InstructBLIP), our model performs better in general, especially in visual dialog cases. This is not surprising, as our VTC acts as a complementary module and provides LLM with useful image details, which are usually missed by VPGs. By concatenating the original tokens with the complementary tokens and freezing the whole MLLM, VTC improves VPGs by making full use of the prior knowledge. + +
MethodsDocTextSTVOCROKGQAVisdialIconVSRWHOOPSScienceQAVizWiz
mPLUG-Owl2.2438.7612.108.8422.8914.0213.3411.6424.7420.702.8038.99
Otter3.4421.5215.2319.5049.0138.1211.6726.776.4015.1427.2250.04
VPGTrans3.5321.9817.1321.7144.5132.999.7038.2248.7715.8836.9953.23
LLaVA6.2638.9228.4023.4054.3641.3014.6642.9551.2424.3949.3362.42
BLIP24.7531.9820.9838.8544.9345.5310.7360.8263.6324.8760.7365.44
VTC(BLIP2)5.3232.025.6738.8946.0848.9777.2062.3163.7726.6563.1066.30
MiniGPT-42.6519.4013.5516.8537.4830.8210.3137.5941.5617.9125.4347.48
VTC (MiniGPT-4)3.0422.7515.5019.5441.7535.1176.6239.0343.7518.5337.8552.10
InstructBLIP5.8939.6028.3060.2060.5249.9645.2056.2541.2830.1346.2665.31
VTC (Instruct)6.1440.5230.7060.2460.9651.1265.7856.8443.1033.7446.2166.31
+ +Table 1: Comparison of Zero-shot Performance for various MLLMs on general VQA tasks. + +
DatasetsmPLUG-OwlOtterVPGTransLLaVABLIP2VTC(B)MiniGPT4VTC(M)InstructBLIPVTC(I)
Existence120.00195.0070.0050.00160.00170.0068.3372.00185.00188.00
Count50.0088.3385.0050.00135.00138.3955.0055.00143.33144.01
Position50.0086.6763.3350.0077.3380.5143.3345.7666.6770.84
Color55.00113.3373.3355.00148.33152.9075.0075.00153.33154.51
Poster136.06138.7884.0150.00141.85147.3241.8444.00123.81127.04
Celebrity100.29172.6553.5348.82105.59112.6554.4157.63101.18108.74
Scene135.50158.75141.7550.00145.25150.0071.7575.34153.00155.00
Landmark159.25137.2564.7550.00138.00141.2854.0054.0079.7580.34
Artwork96.25129.0077.2549.00136.50137.4460.5064.39134.25134.97
OCR65.0072.5077.5050.00110.00110.0057.5060.0072.5077.50
Commonsense78.57106.4364.2957.14110.00110.9559.2962.55129.29130.03
Numerical60.0072.5050.0050.0040.0042.0045.0045.0040.0042.00
Text Translation80.0057.5077.5057.5065.0065.005.5060.0065.0067.50
Code Reasoning57.5070.0057.5050.0075.0070.5040.0040.0057.5060.00
+ +Table 2: Evaluation results (\%) of perception and recognition on MME. The results are calculated by summing ACC and $\mathrm{ACC + }$ in MME [8]. We shorten VTC variants by keeping the first letter of the corresponding base models. + +Results on MME [8]. MME is another recently proposed benchmark measuring both the visual perception and cognition abilities of MLLMs across 14 datasets. Given a test image, MME designs two prompts, and each of them consists of two parts, including a visual question and an instruction "Please answer yes or no". This makes it convenient to report the quantitative performance of accuracy. We report the accuracy score at Table. 2. First, We find that Otter, VTC(B), and VTC(I) achieve 5/14, 3/14, and 3/14 best results, respectively. Otter aligns the vision-language modalities under the in-context learning framework on massive multimodal datasets, while our approach is trained from a single modality. These competitive results indicate our unsupervised training paradigms effectively complete the missing semantics, showing great potential in overcoming the inherent limitation of VPGs. Second, compared to mPLUG-Owl (with 2/14 best results), which finetunes the LLM for multimodal alignments, our approach shows better performance in terms of both perception and cognition. This indicates the efficiency of our image reconstruction loss, which helps VTC complete the image details so that the LLM has sufficient visual evidence for the correct response. + +![](images/e7673d037ef1940d785a4a6f3f00e147bac7d4f7a673bf57337993892c5e9299.jpg) +Fig. 4: Zero-shot evaluation results on DEMON. We only report our VTC(InstructBLIP) for clear representation. + +![](images/5e9ecb33007ac3188a5bb664d7df4ef3788ea856c81898dcadeb492d7901a7da.jpg) +Fig. 5: ChatGPT 4(V) evaluation results of various MLLMs with scores ranging from [1-5]. + +Results on DEMON [17]. We adopt DEMON as our third vision-instruction understanding benchmark. Different from the first two evaluations, DEMON designs more complex demonstrative instructions, including interleaved, and multimodal inputs. We report the evaluation score in Fig. 4. First, our approach surrounds other baseline methods in all directions, demonstrating robust multimodal understanding. Moreover, our approach obtains $20\%$ improvement over InstructBLIP in visual relation inference. We attribute this to our reconstruction prior, which offers helpful guidance to help visual tokens memorize the location information of the targets. Lastly, compared to Otter, which shows higher performance on MME, our VTC(InstructBLIP) achieves significant improvements on all tasks. This demonstrates our superior vision-language understanding in complex scenarios, where the reconstruction prior still provides fine-grained guidance, while the captions fail to capture such detailed supervision. + +# 4.3 Further Analysis + +ChatGPT 4(V) evaluation. Besides the above three benchmark evaluations, we further leverage ChatGPT 4(V) [21] to have a systematic measurement of the quality of the proposed model. Specifically, motivated by previous works [1,19] we first randomly select 50 image-instruction pairs from the MIMIC-IT datasets [15] and collect the answers from various MLLMs. ChatGPT 4(V) is then asked to evaluate the quality of answers based on the input image and output the corresponding scores (1-5), where a higher score indicates better overall performance. We report the results in Fig. 5. We find that our VTC(InstructBLIP) is the only model with a score larger than 4, and the second model is our VTC(MiniGPT4). We hope this evaluation results serve as strong evidence for supporting our performance. The evaluation prompt in this section can be found in the Appendix. Information completion of VTC. One of the core ideas behind VTC is to learn complementary visual tokens that capture visual features missed by VPG. + +![](images/ef2b2bfc3320f7586b6d0eec14bb4fce6ad6dc240d178ce0d09593918e821d9e.jpg) +(a) VTC (MiniGPT4) + +![](images/1c7c0e7c2b40649ca4f710b779a08f4641957a5997681b62babd1ecd82ffeca7.jpg) +(b) VTC (InstructBLIP) +Fig. 6: (a-b) Improvement trends on OKVQA, IconQA, and GQA datasets with different numbers of iterations. (c-d) tag-token nearest neighbor semantic distance $(\times 100)$ on Flicker dataset with different bumbers of iterations ( $x = 0$ means the original VPG). + +![](images/4365fbf6653016dddedd3bacddd0a9ccc68e593913d213442ce6f6753a22fda1.jpg) +(c) VTC (MiniGPT4) + +![](images/471baf434cf71b215f4f745f4f9d255815dce047102259896a155abb7d99123a.jpg) +(d) VTC (InstructBLIP) + +
DatasetsVTC(BLIP2)VTC(InstructBLIP)
BLIP2BLIP2(64)VTCInstructBLIPInstructBLIP(64)VTC
Perception133.79133.95138.23121.42121.86124.31
Cognition73.9473.9675.5473.0273.4475.92
+ +Table 3: Abalation results of fine-tuning with LoRA and the number of visual tokens. We finetune LLMs in our model and baselines with LoRA to test the performance of finetuning with the instruction dataset. (64) means we add 32 visual queries to the original model, thus resulting visual tokens have the same length as our VTC. + +To evaluate such information completion, we here propose the tag-token distance $d$ , which measures the nearest neighbor semantic distance between the image tags and generated visual tokens: $d = \frac{1}{M} \sum_{i=1}^{M} \cosine(\pmb{t}_i, \pmb{y}_i)$ , where $M$ is the number of tags, the visual token $\pmb{y}_i$ is the nearest neighbor of tag $i$ , and $\cosine(\cdot)$ is the cosine distance. We adopt the tagged Flicker dataset [12] to measure the semantic distance, where the average number of tags per image is 8.94. We report the results of VTC(MiniGPT4) and VTC(InstructBLIP) in Fig. 6(c-d), where the $x$ -axis denotes the number we reuse the VTC module, and $x = 0$ means the original VPG. We find that VTC significantly reduces the semantic distance. That is, the newly generated tokens are able to capture visual concepts missed by the previous tokens, which chimes with our motivations. We also visualize the vision tokens in Fig. 7, which will be introduced later. + +Impact of the number of iterations. Due to the flexible design of our VTC, we can reuse the VTC module iteratively at the test stage, once the VTC is trained with the first iteration. This makes it possible to increasingly complete the visual tokens on demand. Yet more iterations mean higher computational costs. We report the impact of the number of iterations in Fig. 6. Fig. 6(a-b) denote the normalized VQA results on OKVQA, IconQV, and GQA datasets with different iterations(e.g.,1,2,3), and Fig. 6(c-d) denote the tag-token distance on Flick dataset. Overall, increasing the number of iterations can improve the performance in terms of VQA and tag-token retrievals. This demonstrates that the newly generated complementary tokens indeed capture more visual concepts, enhancing the LLMs. Besides, we observe that two iterations are sufficient and after two, the gain is not very significant. This is reasonable since our frozen SD model forces the visual tokens $\hat{\pmb{y}}$ to capture as many features as possible to + +
InstructBLIPTokens:tags:man, grass field, frisbeeCaptions:The image features a shirtless man catching a frisbee in a grassy field. The man is in the air, ready to catch the frisbee and throw it into the air. There are several other people present in the scene, including a woman, a man...
VTC(InstructBLIP)-1Complementary tokens:tags:man, grass field, frisbee, outdoor,jumping, park, carCaptions:The image features a man catching a frisbee in a grassy field. The shirtless man is in the air, preparing to throw the frisbee into the air. There are several other people present in the scene, including a man in a red shirt and a man in a blue shirt. There are several cars in the background, including a red car and a white car...
VTC(InstructBLIP)-2Complementary tokens:tags:man, grass field, frisbee, sunny, tree, sky,outdoor, jumping, parkCaptions:The image features a man catching a frisbee in a grassy field. The shirtless man is in the air, preparing to throw the frisbee into the air. There is a cloudy sky in the distance. The man appears to be enjoying a game of frisbee in a park, with a frisbee in his hands. There are several other people present in the scene, including a woman, a man. There are several cars in the background, including ...
+ +![](images/2791c93a99b99b0a189e52d227f7b61be06c0ab90531594a0a96619c8cb8dda5.jpg) +Fig. 7: Visualization of the complementary tokens in the first iteration (VTC(InstructBLIP)-1) and second iteration (VTC(InstructBLIP)-2). We also provide the retrieved tags and predicted captions. +Fig.8: Ablation results of the reconstruction quality. The first row denotes the raw input images. $K = 0$ denotes the original InstructBLIP and $K = 1$ denotes we use VTC once to complete the visual tokens. + +reconstruct the input image in the first iteration, and therefore there is little detail to pick up in the following iterations. + +Vision-language disentanglement. After being trained with reconstruction prior, the completed tokens $\hat{y}$ provide LLMs with lossless image messages. Thus we can fix the VTC and VPG and only finetune the LLM for instruction learning. This mitigates the spurious correlations between vision translating and instruction understanding, which are two important components of MLLMs and often entangled with each other in previous VPGs. As the visual tokens are already completed, we can focus on the latter by finetuning LLMs. Specifically, We finetune our LLM and baselines with LoRA [11] on the LLaVA Visual Instruct 150K dataset [19] and report the evaluation results on MME benchmark at Table. 3. We find that fine-tuning LLM improves performance considerably further. Recalling that BLIP2 is trained with image-caption pairs and InstuctBLIP tunes BLIP2 with massive instruction pairs, our VTC(BLIP2) with finetuning outperforms VTC(InstructBLIP) on both the perception and cognition tasks. This + +![](images/44360e4497db92cb0580233778ef9e245ea29b7f658054dabd93a0d301c2f3c7.jpg) +Fig.9: Examples of the test dialogs. VTC shows impressive vision-language understanding in image captioning, human common sense, and visual reasoning, across single and multi-turns. + +demonstrates that our approach uses the instruction pairs more efficiently. Besides, we also test the performance of visual query length. Our VTC learns 32 additional tokens in the visual prompt, we thus add 32 visual queries to the original baselines and jointly train those vectors on the instruction pairs. We find that our VTC achieves higher scores on MME. We attribute this to our reconstruction-aware training, which effectively learns missing concepts. + +Ablation studies on reconstruction quality. One of the core ideas behind VTC is to learn reconstruction-aware tokens for comprehensive image representation. The proposed VTC utilizes a frozen SD as the image decoder to complete the visual tokens. To measure the performance of the completed visual tokens, We here aim to test the improvements by measuring the quality of the reconstructed images. Specifically, we report the FID scores on the COCO2014 validation dataset. A lower FID score means closer semantics between the reconstructed images and the original images. From Fig. 8 we find that VTC captures more image details, such as the number of persons, the number of airplanes, and the accurate sky background. The higher reconstruction quality demonstrates the efficiency of the newly generated tokens, thus contributing to the responses of LLMs. + +# 4.4 Visualization & Case Studies + +Recall that the cross-attention weights between the query embeddings $\mathbf{q}$ and the patch embeddings $\mathbf{h}$ in VPG provide us with an interpretable tool to visualize the complementary tokens. Specifically, we average the attention weight over all tokens and visualize the attention map in Fig. 7. In addition to the attention map, we also provide the retrieved tags and the corresponding predicted caption. We find that the complementary tokens successfully attend to the missing regions and different iterations cover diverse visual concepts. This meets with our complement motivation, resulting in more accurate tags and detailed captions. + +We also showcase several test dialogs in Fig. 9. Those examples cover human common sense, reasoning, and knowledge retrieval, across single and multiple turns. For example, our VTC has the ability to reason with human common sense, identify the funny aspect of connecting VGA with smartphones, and the unusual cases in a street picture. Moreover, VTC is able to infer the winner of a tic-tac-toe game and the correct number in a Sudoku puzzle based on pre-trained human knowledge. Those results demonstrate the impressive visual perception ability of our approach. + +# 5 Conclusion + +We in this paper propose VTC, an instruction tuning-free framework to learn complementary visual tokens that capture image details missed by VPG. VTC leverages text-to-image generation to guide the acquisition of reconstruction-aware vision tokens. The enhanced vision prompt is formed by merging these new tokens with the original ones. Additionally, we have developed an iterative strategy to progressively reclaim missing visual information on demand during the testing phase. Extensive experiments are conducted on three widely used multimodal benchmarks to evaluate the proposed model. A ChatGPT 4(V) evaluation is also included to have a systematic measurement. Visualization and case studies demonstrate the interpretability and impressive vision-language understanding of our model. + +# Acknowledgement + +This work was supported in part by the National Natural Science Foundation of China under Grant U21B2006; in part by Shaanxi Youth Innovation Team Project; in part by the Fundamental Research Funds for the Central Universities QTZX24003 and QTZX22160; in part by the 111 Project under Grant B18039. + +# References + +1. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality (March 2023) +2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022) +3. Anil, R., Dai, A.M., First, O., Johnson, M., Lepikhin, D., Passos, A., Shakeri, S., Taropa, E., Bailey, P., Chen, Z., et al.: Palm 2 technical report. arXiv preprint arXiv:2305.10403 (2023) +4. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., et al.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023) +5. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022) +6. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.C.H.: Instructclip: Towards general-purpose vision-language models with instruction tuning. CoRR abs/2305.06500 (2023) +7. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023) +8. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., Li, K., Sun, X., Ji, R.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023) +9. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023) +10. He, K., Chen, X., Xie, S., Li, Y., Dólar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022) +1. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) +2. Huiskes, M.J., Lew, M.S.: The mir flickr retrieval evaluation. In: Proceedings of the 1st ACM international conference on Multimedia information retrieval. pp. 39-43 (2008) +3. Kenton, J.D.M.W.C., Toutanova, L.K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of naacL-HLT. vol. 1, p. 2 (2019) + +14. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439 (2023) +15. Li, B., Zhang, Y., Chen, L., Wang, J., Pu, F., Yang, J., Li, C., Liu, Z.: Mimic-it: Multi-modal in-context instruction tuning (2023) +16. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023) +17. Li, J., Pan, K., Ge, Z., Gao, M., Zhang, H., Ji, W., Zhang, W., Chua, T.S., Tang, S., Zhuang, Y.: Fine-tuning multimodal llms to follow zero-shot demonstrative instructions. arXiv preprint arXiv:2308.04152 (2023) +18. Li, J., Li, D., Savarese, S., Hoi, S.C.H.: BLIP-2: bootstrapping language-image pretraining with frozen image encoders and large language models. In: International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA. vol. 202, pp. 19730-19742. PMLR (2023) +19. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023) +20. OpenAI: GPT-4 technical report. CoRR abs/2303.08774 (2023) +21. OpenAI: GPT-4V(ision) system card (2023) +22. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback, 2022 13 (2022) +23. Paischer, F., Adler, T., Hofmarcher, M., Hochreiter, S.: Sitta: A semantic image-text alignment for image captioning. arXiv preprint arXiv:2307.05591 (2023) +24. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +25. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) +26. Schonemann, P.H.: A generalized solution of the orthogonal procrustes problem. Psychometrika 31(1), 1-10 (1966) +27. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021) +28. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (2021) +29. Surís, D., Menon, S., Vondrick, C.: Vipergpt: Visual inference via python execution for reasoning. arXiv preprint arXiv:2303.08128 (2023) +30. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) +31. Tsimpoukelli, M., Menick, J.L., Cabi, S., Eslami, S., Vinyals, O., Hill, F.: Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems 34, 200-212 (2021) +32. Wang, J., Yang, Z., Hu, X., Li, L., Lin, K., Gan, Z., Liu, Z., Liu, C., Wang, L.: Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100 (2022) +33. Wang, Z., Yu, J., Yu, A.W., Dai, Z., Tsvetkov, Y., Cao, Y.: Simvlm: Simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904 (2021) + +34. Wei, J., Bosma, M., Zhao, V.Y., Guu, K., Yu, A.W., Lester, B., Du, N., Dai, A.M., Le, Q.V.: Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652 (2021) +35. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022) +36. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023) +37. Xu, P., Shao, W., Zhang, K., Gao, P., Liu, S., Lei, M., Meng, F., Huang, S., Qiao, Y., Luo, P.: Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. arXiv preprint arXiv:2306.09265 (2023) +38. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023) +39. Yin, S., Fu, C., Zhao, S., Li, K., Sun, X., Xu, T., Chen, E.: A survey on multimodal large language models. arXiv preprint arXiv:2306.13549 (2023) +40. Zhang, A., Fei, H., Yao, Y., Ji, W., Li, L., Liu, Z., Chua, T.S.: Transfer visual prompt generator across llms. arXiv preprint arXiv:2305.01278 (2023) +41. Zhang, C., Zhang, C., Zhang, M., Kweon, I.S.: Text-to-image diffusion model in generative ai: A survey. arXiv preprint arXiv:2303.07909 (2023) +42. Zhang, S., Dong, L., Li, X., Zhang, S., Sun, X., Wang, S., Li, J., Hu, R., Zhang, T., Wu, F., et al.: Instruction tuning for large language models: A survey. arXiv preprint arXiv:2308.10792 (2023) +43. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022) +44. Zhao, H., Cai, Z., Si, S., Ma, X., An, K., Chen, L., Liu, Z., Wang, S., Han, W., Chang, B.: Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915 (2023) +45. Zhao, X., Liu, B., Liu, Q., Shi, G., Wu, X.M.: Making multimodal generation easier: When diffusion models meet llms. arXiv preprint arXiv:2310.08949 (2023) +46. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) \ No newline at end of file diff --git a/2024/Instruction Tuning-free Visual Token Complement for Multimodal LLMs/images.zip b/2024/Instruction Tuning-free Visual Token Complement for Multimodal LLMs/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..293b71c5ee959fdfd7a9ce8cec4a0c96def19abb --- /dev/null +++ b/2024/Instruction Tuning-free Visual Token Complement for Multimodal LLMs/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:31a1147096c06a7640ac39c35c099e0b027b227fbd4d8fe5da02026c56d723e3 +size 670251 diff --git a/2024/Instruction Tuning-free Visual Token Complement for Multimodal LLMs/layout.json b/2024/Instruction Tuning-free Visual Token Complement for Multimodal LLMs/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fe0e486398064a4ef0737ee30519f5f7ef2defab --- /dev/null +++ b/2024/Instruction Tuning-free Visual Token Complement for Multimodal LLMs/layout.json @@ -0,0 +1,8950 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 178, + 111, + 436, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 111, + 436, + 148 + ], + "spans": [ + { + "bbox": [ + 178, + 111, + 436, + 148 + ], + "type": "text", + "content": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 167, + 448, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 167, + 448, + 194 + ], + "spans": [ + { + "bbox": [ + 165, + 167, + 448, + 194 + ], + "type": "text", + "content": "Dongsheng Wang" + }, + { + "bbox": [ + 165, + 167, + 448, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 165, + 167, + 448, + 194 + ], + "type": "text", + "content": ", Jiequan Cui" + }, + { + "bbox": [ + 165, + 167, + 448, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 165, + 167, + 448, + 194 + ], + "type": "text", + "content": ", Miaoge Li" + }, + { + "bbox": [ + 165, + 167, + 448, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 165, + 167, + 448, + 194 + ], + "type": "text", + "content": ", Wang Lin" + }, + { + "bbox": [ + 165, + 167, + 448, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 165, + 167, + 448, + 194 + ], + "type": "text", + "content": ", Bo Chen" + }, + { + "bbox": [ + 165, + 167, + 448, + 194 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 165, + 167, + 448, + 194 + ], + "type": "text", + "content": ", and Hanwang Zhang" + }, + { + "bbox": [ + 165, + 167, + 448, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 195, + 201, + 417, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 201, + 417, + 223 + ], + "spans": [ + { + "bbox": [ + 195, + 201, + 417, + 223 + ], + "type": "text", + "content": "1 Shenzhen University, Shenzhen, 518052, P.R. China, dongshengwang@szu.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 150, + 224, + 462, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 224, + 462, + 244 + ], + "spans": [ + { + "bbox": [ + 150, + 224, + 462, + 244 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 150, + 224, + 462, + 244 + ], + "type": "text", + "content": " Nanyang Technological University, 50 Nanyang Avenue, 639798, Singapore \n" + }, + { + "bbox": [ + 150, + 224, + 462, + 244 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 150, + 224, + 462, + 244 + ], + "type": "text", + "content": " Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 198, + 246, + 414, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 246, + 414, + 257 + ], + "spans": [ + { + "bbox": [ + 198, + 246, + 414, + 257 + ], + "type": "text", + "content": "4 Zhejiang University, Hangzhou 310058, P.R. China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 206, + 257, + 408, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 257, + 408, + 267 + ], + "spans": [ + { + "bbox": [ + 206, + 257, + 408, + 267 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 206, + 257, + 408, + 267 + ], + "type": "text", + "content": " Xidian University, Shaanxi, 710126, P.R. China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 299, + 455, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 299, + 455, + 475 + ], + "spans": [ + { + "bbox": [ + 159, + 299, + 455, + 475 + ], + "type": "text", + "content": "Abstract. As the open community of large language models (LLMs) matures, multimodal LLMs (MLLMs) have promised an elegant bridge between vision and language. However, current research is inherently constrained by challenges such as the need for high-quality instruction pairs and the loss of visual information in image-to-text training objectives. To this end, we propose a Visual Token Complement framework (VTC) that helps MLLMs regain the missing visual features and thus improve response accuracy. Specifically, our VTC integrates text-to-image generation as a guide to identifying the text-irrelevant features, and a visual selector is then developed to generate complementary visual tokens to enrich the original visual input. Moreover, an iterative strategy is further designed to extract more visual information by iteratively using the visual selector without any additional training. Notably, the training pipeline requires no additional image-text pairs, resulting in a desired instruction tuning-free property. Both qualitative and quantitative experiments demonstrate the superiority and efficiency of our VTC." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 496, + 230, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 496, + 230, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 496, + 230, + 509 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": "Large language models (LLMs) have demonstrated extraordinary reasoning ability across various applications [3, 20, 29, 30, 35, 43]. This sparks a research trend that focuses on equipping LLMs with visual prompts to enhance vision-language understanding. To complete various multimodal tasks, we can feed unified vision-language prompts, such as “ An image that shows” or “ What color is the dog?”, into an LLM to generate the required answers such as image captioning or VQA [6, 18, 19, 46]. We designate these multimodal-enhanced Large Language Models as Multimodal LLMs (MLLMs), as our focus is solely on multimodal understanding, rather than on both understanding and/or generation. In particular, the latter is known as Large Multimodal Models (LMMs) [2, 5, 7, 16, 32, 45], which, unfortunately, have been shown to possess a lower understanding capability compared to MLLMs [6, 8, 39]." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 114, + 455, + 205 + ], + "blocks": [ + { + "bbox": [ + 161, + 114, + 455, + 205 + ], + "lines": [ + { + "bbox": [ + 161, + 114, + 455, + 205 + ], + "spans": [ + { + "bbox": [ + 161, + 114, + 455, + 205 + ], + "type": "image", + "image_path": "a80f5e753ca884386f1db8726d574824d1505f510a0e71eb285ea17866aac96e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 212, + 482, + 257 + ], + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 257 + ], + "type": "text", + "content": "Fig. 1: Comparisons of our proposed Reconstruction prior (c) for training Visual Token Complement (VTC), Caption prior (a), and Instruction prior (b) for training Visual Prompt Generator (VPG). The attention maps of VPG and VTC are obtained by averaging the query-patch attention weights(Sec. 3.1)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 281, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 365 + ], + "type": "text", + "content": "As \"an image is more than a thousand words\", it is efficient to represent visual prompts as continuous embeddings but not discrete language tokens. For example, BLIP-2 [18] and MiniGPT4 [46] use Q-Former to transform image patches into a fixed number of visual tokens; while LLaVA [19] demonstrates that a simple linear projection can be equally effective instead of using Q-Former. These visual tokens are expected to encode the visual contents in the image, providing informative context for LLMs to generate responses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "content": "The crux of making MLLMs effective is to train a Vision Prompt Generator (VPG) that can generate visual tokens, recognizable by LLMs, from an input image. As VPG is essentially a visual-to-semantic translator, we need a visual-semantic alignment training task to align the visual tokens and their semantics written in language output by a frozen LLM. There are two representative tasks that introduce different visual priors into VPG:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": "Caption prior: Image captioning is a straightforward alignment task. For instance, the Q-Former VPG in BLIP-2 [18] and MiniGPT4 [46] is trained on large-scale image-caption pairs [27]: the frozen LLM should output the ground-truth caption when taking the visual tokens of the corresponding image as input. However, this task inevitably introduces an inductive bias losing significant details, that is, the VPG tends to only capture the visual semantics which is just enough for captioning the image contents according to the prescribed annotation details in training. As illustrated in Fig. 1(a), as the VPG only focuses on the main object pomelos but neglects the price tag, the MLLM fails to answer the question about price." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Instruction prior: Inspired by instruction tuning [22, 34, 42], which turns the general language prior of LLMs—next-token prediction—into more powerful task solvers, InstructBLIP [6] and LLaVA [19] aim to train VPG for task-aware visual tokens. They curate a variety of vision-language tasks such as image caption, visual grounding and VQA into the same template, which is fed into a frozen LLM for training VPG. However, curating instructions at scale and with diversity is particularly costly for vision-language tasks, so the VPG is usually biased to the instruction prior due to overfitted tuning. As shown in Fig. 1(b), the VPG cannot distinguish the price tags of different fruits" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "D. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 166, + 115, + 449, + 231 + ], + "blocks": [ + { + "bbox": [ + 166, + 115, + 449, + 231 + ], + "lines": [ + { + "bbox": [ + 166, + 115, + 449, + 231 + ], + "spans": [ + { + "bbox": [ + 166, + 115, + 449, + 231 + ], + "type": "image", + "image_path": "a9d3857c2d1fdef3ae0446ebde46f70cb9344fbd4458ab5a5883b2cf0d85a6e9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 239, + 480, + 274 + ], + "lines": [ + { + "bbox": [ + 131, + 239, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 131, + 239, + 480, + 274 + ], + "type": "text", + "content": "Fig. 2: Iterative complement at the inference stage. Once trained, Our VTC can be reused multiple times to re-pick up the visual concepts omitted by the previous VPG. We showcase three times complements with increasingly detailed output captions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 298, + 479, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 479, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 479, + 320 + ], + "type": "text", + "content": "because such price inquiry of multiple objects is rare in the multi-task datasets used in InstructBLIP [6]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 323, + 480, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 480, + 417 + ], + "type": "text", + "content": "In this paper, we use Reconstruction prior to recover the missing visual semantics by VPG with caption/instruction prior. As shown in Fig. 1(c), we propose a trainable module called Visual Token Complement (VTC) that attends to visual regions missed by VPG. By feeding them to VPG, we can obtain the complementary visual tokens encoding visual semantics not yet captured by VPG. Together with the existing visual tokens, how well the complementary tokens complete the image contents is measured by the image reconstruction loss. We highlight the following three advantages of the proposed VTC:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 427, + 480, + 632 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 427, + 479, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 427, + 479, + 474 + ], + "spans": [ + { + "bbox": [ + 138, + 427, + 479, + 474 + ], + "type": "text", + "content": "- Instruction tuning-free. Training VTC is unsupervised. All we need to calculate the reconstruction loss is the input images and three frozen modules: VPG, LLM, and image decoder. No caption pairs are needed to train our VTC." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 475, + 480, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 475, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 138, + 475, + 480, + 547 + ], + "type": "text", + "content": "- Completeness on-demand. As illustrated in Fig. 2, thanks to the simple implementation of VTC (Sec. 3.1), we have the flexibility to iteratively generate additional complementary tokens as required. As the VPG is frozen, the iteration can always make up for the visual semantics missed by the VPG in the previous reconstruction. We show that two iterations are sufficient in our experiments (Sec. 4.3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 548, + 480, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 548, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 548, + 480, + 632 + ], + "type": "text", + "content": "- Vision-language disentanglement. Thanks to the above advantages, if we have instruction tuning data, we can fix the trained VTC and VPG and only tune the LLM for instruction comprehension. As the visual tokens are already completed, such disentanglement can mitigate the vision-language spurious correlation in training data, and utilize the valuable instruction data more efficiently—LLM is more specialized in understanding instructions. We show promising results in Sec. 4.3." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 479, + 665 + ], + "type": "text", + "content": "To evaluate the proposed VTC, we conducted extensive comparisons on LVLM-eHub [37], MME [8], and DEMON [17] multimodal benchmarks, covering" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "text", + "content": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "22 diverse vision tasks in the zero-shot setting. By completing the missing information, our VTC shows significant improvements over its baselines, especially in visual dialogue (showing " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " improvement over InstructBLIP [6]). Besides, the visualizations of the appended visual tokens, provide an interpretable tool to analyze the complementary semantics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 193, + 238, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 238, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 238, + 205 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 217, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 445 + ], + "type": "text", + "content": "Recently, research focused on extending large language models (LLMs) to accommodate image inputs has garnered increasing attention, primarily due to their significant improvements in various vision-language tasks [9, 17, 21, 36]. They aim to learn cross-modal reasoning from large-scale multimodal data. Pioneering studies like SimVLM [33], Frozen [31], and GiT [32] view image as a foreign language and have achieved exceptional fine-tuning performance under the autoregressive framework. In-context learning series [2, 4, 16, 44] are proposed to align a pre-trained vision encoder with LLMs by the design of multimodal prompt, showing remarkable potential in zero-shot and few-shot tasks. Following that, various VPGs are developed to efficiently align visual features with the language model, e.g., Linear mapping [19] and Q-Former [18]. The follow-up instruction-tuned methods are proposed from various assumptions to help VPGs make full use of prior knowledge implied in LLMs. For instance, LLaVA [19] collects high-quality instruction pairs and finetunes the linear projection layer between the VPG and LLM. InstructBLIP [6] introduces instruction-aware Q-Former and extracts informative features according to the given instruction. Several attempts have also been made to enable MLLMs to understand videos and audiios. They share similar techniques with image-language alignment, yielding impressive performance. The scope of this paper is image-language modalities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 445, + 483, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 483, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 483, + 529 + ], + "type": "text", + "content": "However, most existing VPGs align the vision and language domains through autoregressive training objectives, e.g. the image-to-caption loss, which may result in incomplete alignment due to the semantic gaps and noisy matchings in multimodal data [45]. Our approach, in contrast, adopts a text-to-image strategy to guide the learning of visual tokens so that they can reconstruct the input image as closely as possible, resulting in instruction tuning-free visual token complement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 545, + 289, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 289, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 289, + 560 + ], + "type": "text", + "content": "3 The Proposed Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "For a given image " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": ", VPGs aim to generate a soft vision prompt " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{y} = \\{\\pmb{y}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " containing " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " visual tokens, where " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{y}_i \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "-th token vector and " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " is the embedding dimension. Without loss of generality, we adopt the Q-Former as the VPG in the following discussion. Q-Former utilizes a fixed number of " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " query vectors " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " to interact with image features " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{h} = f(\\pmb{x})" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " by multiple cross-attention layers, where " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " is a frozen vision encoder (e.g., ViT-G). The output queries are used as visual tokens " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": ", which are prepended in the input text embedding of the instruction. To train a VPG, an autoregressive loss is often employed on" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "D. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 156, + 114, + 460, + 236 + ], + "blocks": [ + { + "bbox": [ + 156, + 114, + 460, + 236 + ], + "lines": [ + { + "bbox": [ + 156, + 114, + 460, + 236 + ], + "spans": [ + { + "bbox": [ + 156, + 114, + 460, + 236 + ], + "type": "image", + "image_path": "6001c0f08b2f15d07a4ee027c83d6192f3f68e89061dfdeba58664bcb6659e78.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 247, + 465, + 259 + ], + "lines": [ + { + "bbox": [ + 147, + 247, + 465, + 259 + ], + "spans": [ + { + "bbox": [ + 147, + 247, + 465, + 259 + ], + "type": "text", + "content": "Fig. 3: (a) Training pipeline of the proposed VTC, and (b) structure of VTC." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 282, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 482, + 402 + ], + "type": "text", + "content": "the predicted text: " + }, + { + "bbox": [ + 132, + 282, + 482, + 402 + ], + "type": "inline_equation", + "content": "p(\\pmb{w}|\\pmb{y}, \\pmb{w}_{\\text{instruct}})" + }, + { + "bbox": [ + 132, + 282, + 482, + 402 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 132, + 282, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\pmb{w}_{\\text{instruct}}" + }, + { + "bbox": [ + 132, + 282, + 482, + 402 + ], + "type": "text", + "content": " represents the textual input and " + }, + { + "bbox": [ + 132, + 282, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 132, + 282, + 482, + 402 + ], + "type": "text", + "content": " is the predicted answers. Caption prior-based VPGs typically specify " + }, + { + "bbox": [ + 132, + 282, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\pmb{w}_{\\text{instruct}}" + }, + { + "bbox": [ + 132, + 282, + 482, + 402 + ], + "type": "text", + "content": " with the caption template: “ An image that shows”; while the instruction prior-based VPGs tailor various templates to specific tasks, e.g., “ Question: {Question}. Short answer:” for VQA and “ Given the image, generate a question whose answer is: {Answer}. Question:” for visual grounding. VPGs serve as an intermediary between the image and the frozen LLM, with their core idea being to find the optimal " + }, + { + "bbox": [ + 132, + 282, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 132, + 282, + 482, + 402 + ], + "type": "text", + "content": " that provides efficient visual semantics about " + }, + { + "bbox": [ + 132, + 282, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 132, + 282, + 482, + 402 + ], + "type": "text", + "content": ", thereby enabling the LLM to produce accurate responses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 419, + 293, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 293, + 431 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 293, + 431 + ], + "type": "text", + "content": "3.1 Visual Token Complement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 438, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 533 + ], + "type": "text", + "content": "In this paper, we focus on addressing the issue of information loss during transforming image features into visual tokens, and propose the Visual Token Complement (VTC) framework for existing VPGs. As discussed above, VPGs based on caption/instruction prior often struggle with incompleteness issues, where the generated vision tokens " + }, + { + "bbox": [ + 130, + 438, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 438, + 482, + 533 + ], + "type": "text", + "content": " are prone to emphasizing main objects at the expense of finer image details. To complete " + }, + { + "bbox": [ + 130, + 438, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 438, + 482, + 533 + ], + "type": "text", + "content": " and provide comprehensive visual details into LLM, VTC introduces an additional set of vision tokens " + }, + { + "bbox": [ + 130, + 438, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{y}^{\\prime} = \\{\\pmb{y}_{i}^{\\prime}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 130, + 438, + 482, + 533 + ], + "type": "text", + "content": " by reusing the VPG:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 229, + 543, + 481, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 543, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 229, + 543, + 481, + 556 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} ^ {\\prime} = \\operatorname {V P G} (\\boldsymbol {c}, \\boldsymbol {q}), \\quad \\boldsymbol {c} = \\operatorname {V T C} (\\boldsymbol {h}, \\boldsymbol {y}), \\tag {1}", + "image_path": "2ee2dc56e3371165cb02b1b438414591ee85cf36a8dc16f6e73027eff94fc47a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "text", + "content": "where VTC takes image feature " + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\pmb{h}" + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "text", + "content": ", the generated vision tokens " + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "text", + "content": " as inputs and outputs the complementary feature " + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\pmb{c}" + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "text", + "content": ". We will introduce the detailed structure of VTC later. " + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\pmb{c}" + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "text", + "content": " captures the visual details missed by " + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "text", + "content": " and is fed into the VPG to output the complementary visual tokens. Mathematically, our VTC can be applied with most existing VPGs, e.g., Q-Former and linear projection in LLaVA [19]" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "type": "text", + "content": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": "6 We empirically adopt the Q-Former as our VPG. LLaVA generates 256 tokens for each image, which is too long for training our VTC." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 139 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 139 + ], + "type": "text", + "content": "Once the complementary vision tokens " + }, + { + "bbox": [ + 132, + 116, + 482, + 139 + ], + "type": "inline_equation", + "content": "\\pmb{y}^{\\prime}" + }, + { + "bbox": [ + 132, + 116, + 482, + 139 + ], + "type": "text", + "content": " are generated, the final completed vision tokens are obtained as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 267, + 146, + 481, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 146, + 481, + 158 + ], + "spans": [ + { + "bbox": [ + 267, + 146, + 481, + 158 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {y}} = \\operatorname {C o n c a t} \\left(\\boldsymbol {y}, \\boldsymbol {y} ^ {\\prime}\\right), \\tag {2}", + "image_path": "482440580eb4fced6f1e75420ddfd91bd7408f701ad3d16e6ae4a04fc69a2b02.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 163, + 479, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 163, + 479, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 163, + 479, + 246 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 163, + 479, + 246 + ], + "type": "inline_equation", + "content": "\\mathrm{Concat}(\\cdot)" + }, + { + "bbox": [ + 130, + 163, + 479, + 246 + ], + "type": "text", + "content": " denotes applying the concatenating at the token dimension. Like previous MLLMs, " + }, + { + "bbox": [ + 130, + 163, + 479, + 246 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 163, + 479, + 246 + ], + "type": "text", + "content": " in VTC will be inserted into the position of the virtual image tokens to form the LLM's input according to different instruction templates. Crucially, the completed " + }, + { + "bbox": [ + 130, + 163, + 479, + 246 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 163, + 479, + 246 + ], + "type": "text", + "content": " combines the original caption-aware " + }, + { + "bbox": [ + 130, + 163, + 479, + 246 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 163, + 479, + 246 + ], + "type": "text", + "content": " and the newly generated tokens " + }, + { + "bbox": [ + 130, + 163, + 479, + 246 + ], + "type": "inline_equation", + "content": "\\pmb{y}'" + }, + { + "bbox": [ + 130, + 163, + 479, + 246 + ], + "type": "text", + "content": ", where the former tends to capture the main visual objects and the latter focuses on the missing details. Together, they offer comprehensive visual evidence and help LLM capture more vision semantics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 247, + 479, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 479, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 479, + 294 + ], + "type": "text", + "content": "In addition, moving beyond the patch embeddings at the last layer of " + }, + { + "bbox": [ + 130, + 247, + 479, + 294 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 247, + 479, + 294 + ], + "type": "text", + "content": ", VTC aims to explore multi-level semantics and provide hierarchical vision descriptions for LLMs. Let " + }, + { + "bbox": [ + 130, + 247, + 479, + 294 + ], + "type": "inline_equation", + "content": "\\pmb{h}^{(l)}" + }, + { + "bbox": [ + 130, + 247, + 479, + 294 + ], + "type": "text", + "content": " denote the patch embeddings at " + }, + { + "bbox": [ + 130, + 247, + 479, + 294 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 247, + 479, + 294 + ], + "type": "text", + "content": "-th layer of " + }, + { + "bbox": [ + 130, + 247, + 479, + 294 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 247, + 479, + 294 + ], + "type": "text", + "content": ", VTC pickups the missing features as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 231, + 300, + 481, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 300, + 481, + 313 + ], + "spans": [ + { + "bbox": [ + 231, + 300, + 481, + 313 + ], + "type": "interline_equation", + "content": "[ \\boldsymbol {c} ^ {(l)}, \\boldsymbol {o} ] = \\operatorname {S e l f - A t t} (\\operatorname {C o n c a t} (\\boldsymbol {h} ^ {(l)}, \\boldsymbol {y})), \\tag {3}", + "image_path": "d0823919f59c0a669ea1c9cd96dd46174a4f8dc87d3406437985671077c0ffe4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 318, + 479, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 479, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 479, + 378 + ], + "type": "text", + "content": "where Self-Att denotes a standard Self-Attention layer, which stacks the self-attention, feeding forward, and residual layers (shown as Fig. 3(b)). " + }, + { + "bbox": [ + 130, + 318, + 479, + 378 + ], + "type": "inline_equation", + "content": "\\pmb{c}^{(l)}" + }, + { + "bbox": [ + 130, + 318, + 479, + 378 + ], + "type": "text", + "content": " denotes the complementary feature at layer " + }, + { + "bbox": [ + 130, + 318, + 479, + 378 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 318, + 479, + 378 + ], + "type": "text", + "content": ". Naturally, we can obtain the multi-layer complementary features " + }, + { + "bbox": [ + 130, + 318, + 479, + 378 + ], + "type": "inline_equation", + "content": "\\pmb{h}" + }, + { + "bbox": [ + 130, + 318, + 479, + 378 + ], + "type": "text", + "content": " in Eq. 3 by concatenating " + }, + { + "bbox": [ + 130, + 318, + 479, + 378 + ], + "type": "inline_equation", + "content": "\\pmb{c}^{(l)}" + }, + { + "bbox": [ + 130, + 318, + 479, + 378 + ], + "type": "text", + "content": " at different layers, resulting in more comprehensive visual tokens " + }, + { + "bbox": [ + 130, + 318, + 479, + 378 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 318, + 479, + 378 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 392, + 312, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 312, + 404 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 312, + 404 + ], + "type": "text", + "content": "3.2 Reconstruction-prior Training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 408, + 479, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 479, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 479, + 491 + ], + "type": "text", + "content": "Inspired by autoencoders [10, 13] that perform representation learning by reconstructing the input, we here introduce the reconstruction-prior training to learn the optimal " + }, + { + "bbox": [ + 130, + 408, + 479, + 491 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 408, + 479, + 491 + ], + "type": "text", + "content": ". Specifically, we view " + }, + { + "bbox": [ + 130, + 408, + 479, + 491 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 408, + 479, + 491 + ], + "type": "text", + "content": " as the latent embeddings and learn VTC with the image reconstruction task. To complete the encoder-decoder framework, we employ a frozen stable diffusion (SD) [24, 28] as the image decoder " + }, + { + "bbox": [ + 130, + 408, + 479, + 491 + ], + "type": "inline_equation", + "content": "g(\\pmb{x}|\\hat{\\pmb{y}})" + }, + { + "bbox": [ + 130, + 408, + 479, + 491 + ], + "type": "text", + "content": ", mapping " + }, + { + "bbox": [ + 130, + 408, + 479, + 491 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 408, + 479, + 491 + ], + "type": "text", + "content": " back to the pixel space due to its impressive text-to-image results in the recent AI-generated content field [14, 25, 41]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 492, + 479, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 479, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 479, + 528 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 130, + 492, + 479, + 528 + ], + "type": "inline_equation", + "content": "(\\pmb{x},\\hat{\\pmb{y}})" + }, + { + "bbox": [ + 130, + 492, + 479, + 528 + ], + "type": "text", + "content": ", SD views " + }, + { + "bbox": [ + 130, + 492, + 479, + 528 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 492, + 479, + 528 + ], + "type": "text", + "content": " as the textual condition, and updates it by solving the following Gaussian noise regression as in training the reversed diffusion steps (with " + }, + { + "bbox": [ + 130, + 492, + 479, + 528 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 492, + 479, + 528 + ], + "type": "text", + "content": " timesteps):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 233, + 533, + 481, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 533, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 233, + 533, + 481, + 552 + ], + "type": "interline_equation", + "content": "\\underset {\\hat {\\boldsymbol {y}}} {\\arg \\min } \\mathbb {E} _ {\\boldsymbol {x} _ {0}, \\epsilon , t} [ \\| \\boldsymbol {\\epsilon} - \\epsilon (\\boldsymbol {x} _ {t}, \\hat {\\boldsymbol {y}} \\mathbf {W}) \\| ^ {2} ], \\tag {4}", + "image_path": "6917acb41bf1b501c2fd009d70fdb8bd5e84c7cb6356e334535aa6fba5943d98.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0 = \\pmb{x}" + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "text", + "content": " is the Gaussian noise, and " + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "inline_equation", + "content": "t \\sim [0, T]" + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "text", + "content": " is the sampled timestep. " + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "inline_equation", + "content": "\\epsilon(\\cdot)" + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "text", + "content": " is the frozen noise prediction network, conditionally on the completed " + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "text", + "content": " is noised image at timestep " + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t \\sim \\mathcal{N}(\\sqrt{1 - \\beta_t} \\pmb{x}_{t-1}, \\beta_t \\pmb{I})" + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "inline_equation", + "content": "\\beta_t \\in (0, 1)" + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "text", + "content": " being the variance. " + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{d \\times d_c}" + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "text", + "content": " is a linear mapping that projects the " + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "text", + "content": "-dimensional LLM embedding into the " + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "inline_equation", + "content": "d_c" + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "text", + "content": "-dimensional CLIP space. We pre-calculate " + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 130, + 556, + 479, + 628 + ], + "type": "text", + "content": " via the least squares with orthogonality constraint [23]:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 213, + 634, + 481, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 634, + 481, + 656 + ], + "spans": [ + { + "bbox": [ + 213, + 634, + 481, + 656 + ], + "type": "interline_equation", + "content": "\\arg \\min _ {\\mathbf {W}} \\mathbb {E} _ {v \\in V _ {\\mathrm {C L I P}} \\cap V _ {\\mathrm {L L M}}} \\| \\boldsymbol {v} _ {\\mathrm {C L I P}} - \\boldsymbol {v} _ {\\mathrm {L L M}} \\mathbf {W} \\| ^ {2}, \\tag {5}", + "image_path": "2774184005c956ea6e46003dea452d09037fdabcbb6ed45c9178f4e3d1b74a1d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 310, + 654, + 379, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 654, + 379, + 666 + ], + "spans": [ + { + "bbox": [ + 310, + 654, + 379, + 666 + ], + "type": "interline_equation", + "content": "s. t., \\mathbf {W} ^ {T} \\mathbf {W} = \\mathbf {I},", + "image_path": "3919bbc779aa3991c791825333152e12938468f6155fd77fb26e6d263cbf0e6c.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "D. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{LLM}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": " denote the vocabulary sets of CLIP and LLM, respectively, and " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{\\mathrm{LLM}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": " are the corresponding CLIP and LLM embedding vector of the shared token " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": ". The orthogonality constraint ensures the learned projection matrix " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": " implicitly preserves the structure of the LLM's embedding and regulars the learning of " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": " focus on the semantic translation [23, 26]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 308 + ], + "type": "text", + "content": "Unlike previous VPGs that optimize the vision tokens " + }, + { + "bbox": [ + 130, + 176, + 480, + 308 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 176, + 480, + 308 + ], + "type": "text", + "content": " by minimizing the image-to-text loss, - with " + }, + { + "bbox": [ + 130, + 176, + 480, + 308 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 176, + 480, + 308 + ], + "type": "text", + "content": " inserted into the \"\" part of the instruction. Eq. 4 gives us an elegant solution to learn the optimal " + }, + { + "bbox": [ + 130, + 176, + 480, + 308 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 176, + 480, + 308 + ], + "type": "text", + "content": " based on the reconstruction prior. The idea is simple: the vision tokens are complete if they can reconstruct the input image. Leveraging the reconstruction loss, the proposed VTC is trained in an unsupervised manner. Given the frozen image encoder, VPG and SD, VTC is the only to-be-learned module to regain the missing features for better reconstruction. This disentangles the vision-language learning. Since the visual tokens " + }, + { + "bbox": [ + 130, + 176, + 480, + 308 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 176, + 480, + 308 + ], + "type": "text", + "content": " are already completed, we can fine-tune the LLM for instruction comprehension on instruction pairs, which is the same as the pre-training stage of LLM, where LLM is more specialized." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 324, + 254, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 324, + 254, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 324, + 254, + 335 + ], + "type": "text", + "content": "3.3 Iterative Inference" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 479, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 479, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 479, + 416 + ], + "type": "text", + "content": "Once trained with Eq. 4, our VTC in Eq. 3 has the ability to identify the missing features by comparing the generated tokens " + }, + { + "bbox": [ + 130, + 342, + 479, + 416 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 342, + 479, + 416 + ], + "type": "text", + "content": " and the image embedding " + }, + { + "bbox": [ + 130, + 342, + 479, + 416 + ], + "type": "inline_equation", + "content": "\\pmb{h}" + }, + { + "bbox": [ + 130, + 342, + 479, + 416 + ], + "type": "text", + "content": ". This makes it possible to reuse VTC and then iteratively generate additional complementary tokens as required. Specifically, let " + }, + { + "bbox": [ + 130, + 342, + 479, + 416 + ], + "type": "inline_equation", + "content": "\\pmb{y}^k" + }, + { + "bbox": [ + 130, + 342, + 479, + 416 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 479, + 416 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}^k" + }, + { + "bbox": [ + 130, + 342, + 479, + 416 + ], + "type": "text", + "content": " denote the generated vision tokens and the completed vision tokens at iteration " + }, + { + "bbox": [ + 130, + 342, + 479, + 416 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 342, + 479, + 416 + ], + "type": "text", + "content": ", respectively, e.g., " + }, + { + "bbox": [ + 130, + 342, + 479, + 416 + ], + "type": "inline_equation", + "content": "\\pmb{y}^1 = \\pmb{y}'" + }, + { + "bbox": [ + 130, + 342, + 479, + 416 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 479, + 416 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}^1 = \\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 342, + 479, + 416 + ], + "type": "text", + "content": ". Thus the complementary tokens " + }, + { + "bbox": [ + 130, + 342, + 479, + 416 + ], + "type": "inline_equation", + "content": "\\pmb{y}^{k+1}" + }, + { + "bbox": [ + 130, + 342, + 479, + 416 + ], + "type": "text", + "content": " can be obtained as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 218, + 422, + 480, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 422, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 218, + 422, + 480, + 436 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} ^ {k + 1} = \\operatorname {V P G} \\left(\\boldsymbol {c} ^ {k}, \\boldsymbol {q}\\right), \\quad \\boldsymbol {c} ^ {k} = \\operatorname {V T C} \\left(\\boldsymbol {h}, \\hat {\\boldsymbol {y}} ^ {k}\\right). \\tag {6}", + "image_path": "8f98ae4aa3bced5c6286fd9bc2e9ff9ac40f388f0b9e0ab59dd0c25e8b5bb781.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 443, + 479, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 443, + 479, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 479, + 491 + ], + "type": "inline_equation", + "content": "\\pmb{y}^{k + 1}" + }, + { + "bbox": [ + 130, + 443, + 479, + 491 + ], + "type": "text", + "content": " captures the visual details missed by " + }, + { + "bbox": [ + 130, + 443, + 479, + 491 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}^k" + }, + { + "bbox": [ + 130, + 443, + 479, + 491 + ], + "type": "text", + "content": ". Benefiting from the flexible structure of VTC, we can train VTC at " + }, + { + "bbox": [ + 130, + 443, + 479, + 491 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 130, + 443, + 479, + 491 + ], + "type": "text", + "content": " and iteratively reuse it to squeeze the visual message. Empirically, we find two iterations are sufficient, accounting for the computational cost." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 510, + 230, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 230, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 230, + 523 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 534, + 479, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 479, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 479, + 558 + ], + "type": "text", + "content": "We conduct extensive experiments on several benchmark multimodal datasets to evaluate the zero-shot performance of VTC against the state-of-the-art MLLMs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 574, + 272, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 272, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 272, + 586 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "content": "Three multimodal benchmarks. In the experiment, we evaluate the diverse and robust performance of our proposed VTC on three zero-shot multimodal benchmarks, " + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "inline_equation", + "content": "LVLM-eHub" + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "content": " [37], " + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "inline_equation", + "content": "MME" + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "content": " [8] and " + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "inline_equation", + "content": "DEMON" + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "content": " [17]. They measure the visual-language comprehension ability of MLLM from different perspectives. " + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "inline_equation", + "content": "LVLM-eHub" + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "content": " is the first multimodal benchmark, which focuses on visual knowledge acquisition, visual reasoning, embodied intelligence, and other 6 categories" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "type": "text", + "content": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "of multimodal capabilities. MME is another recently proposed benchmark that measures both the visual perception and cognition abilities of MLLMs across 14 datasets. DEMON focuses on more complex scenarios, where the demonstrative instructions may contain multiple, interleaved, and multimodal inputs. It covers 29 tasks across 7 categories, including multimodal dialogue, visual storytelling, and other multimodal understanding scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 189, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 482, + 308 + ], + "type": "text", + "content": "Baselines. We compare our VTC with the recent advanced MLLMs on three widely used benchmarks. For all methods, we choose versions with parameter sizes less than 10B, Otter [16], which aligns the image and text under the in-context learning framework; mPLUG-Owl [38] tunes the vision encoder and LLM (with a LoRA module [11]) at different stages; VPGTrans [40], which designs a projector warm-up and vanilla fine-tuning to transfer an existing VPG from one MLLM to the target MLLM; And VPG based MLLMs: LLaVA [19], BLIP2 [18], MiniGPT-4 [46], and InstructBLIP [42], they share similar ideas but varying VPGs and training data. For all methods, we select versions with parameter sizes under 10B. Please refer to the Appendix for details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 309, + 481, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 481, + 332 + ], + "type": "text", + "content": "Training Data. We train our VTC in an unsupervised manner, and we adopt the ImageNet1K as our training dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 334, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 430 + ], + "type": "text", + "content": "Implementation Details. We chose three VPG-based MLLMs as our base model: BLIP2, MiniGPT4, and InstructBLIP, and we denote the corresponding variants as VTC(BLIP2), VTC(MiniGPT4), and VTC(InstructBLIP). We follow the same implementation details as the base model. For the SD model, we load the checkpoints of Stable-Diffusion-v1-5. VTC is trained on ImageNet1K with 2 epochs. For the learning rate, we first conduct a linear warm-up from " + }, + { + "bbox": [ + 130, + 334, + 482, + 430 + ], + "type": "inline_equation", + "content": "1e - 6" + }, + { + "bbox": [ + 130, + 334, + 482, + 430 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 334, + 482, + 430 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 130, + 334, + 482, + 430 + ], + "type": "text", + "content": ", and then use a cosine learning rate schedule with the minimal " + }, + { + "bbox": [ + 130, + 334, + 482, + 430 + ], + "type": "inline_equation", + "content": "lr = 1e - 5" + }, + { + "bbox": [ + 130, + 334, + 482, + 430 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 130, + 334, + 482, + 430 + ], + "type": "inline_equation", + "content": "l = [12,24,36]" + }, + { + "bbox": [ + 130, + 334, + 482, + 430 + ], + "type": "text", + "content": " in Eq. 3 to capture the multi-level visual semantics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 450, + 270, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 270, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 270, + 462 + ], + "type": "text", + "content": "4.2 Quantitative Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": "Results on LVLM-eHub [37]. We first evaluate the vision-language understanding of our VTC on LVLM-eHub. We here focus on the Visual Reasoning task due to it requires a comprehensive understanding of images and texts. It includes visual question answering (VQA, e.g., DocVQA, TextVQA, STVQA, OCR-VQA, OKVQA, GQA, and IconQA), visual spatial reasoning (VSR), visual dialog (e.g., Visdial), and knowledge-grounded image description (e.g., ScienceQA and VizWiz). We report the evaluation results at Table 1. We have the following remarks about the results. Overall, our proposed VTC outperforms the others consistently on all datasets, except the case where ours is the second on the DocVQA dataset. Moreover, compared with our base models (BLIP2, MiniGPT4, and InstructBLIP), our model performs better in general, especially in visual dialog cases. This is not surprising, as our VTC acts as a complementary module and provides LLM with useful image details, which are usually missed by VPGs. By concatenating the original tokens with the complementary tokens and freezing the whole MLLM, VTC improves VPGs by making full use of the prior knowledge." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "D. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 483, + 219 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 483, + 219 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 483, + 219 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 483, + 219 + ], + "type": "table", + "html": "
MethodsDocTextSTVOCROKGQAVisdialIconVSRWHOOPSScienceQAVizWiz
mPLUG-Owl2.2438.7612.108.8422.8914.0213.3411.6424.7420.702.8038.99
Otter3.4421.5215.2319.5049.0138.1211.6726.776.4015.1427.2250.04
VPGTrans3.5321.9817.1321.7144.5132.999.7038.2248.7715.8836.9953.23
LLaVA6.2638.9228.4023.4054.3641.3014.6642.9551.2424.3949.3362.42
BLIP24.7531.9820.9838.8544.9345.5310.7360.8263.6324.8760.7365.44
VTC(BLIP2)5.3232.025.6738.8946.0848.9777.2062.3163.7726.6563.1066.30
MiniGPT-42.6519.4013.5516.8537.4830.8210.3137.5941.5617.9125.4347.48
VTC (MiniGPT-4)3.0422.7515.5019.5441.7535.1176.6239.0343.7518.5337.8552.10
InstructBLIP5.8939.6028.3060.2060.5249.9645.2056.2541.2830.1346.2665.31
VTC (Instruct)6.1440.5230.7060.2460.9651.1265.7856.8443.1033.7446.2166.31
", + "image_path": "78d9733e7031c8230acfc18ba9e0667fa3b10d564b52b193d7d27aa9a5f4f9be.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 134, + 264, + 485, + 393 + ], + "blocks": [ + { + "bbox": [ + 132, + 220, + 480, + 239 + ], + "lines": [ + { + "bbox": [ + 132, + 220, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 480, + 239 + ], + "type": "text", + "content": "Table 1: Comparison of Zero-shot Performance for various MLLMs on general VQA tasks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 264, + 485, + 393 + ], + "lines": [ + { + "bbox": [ + 134, + 264, + 485, + 393 + ], + "spans": [ + { + "bbox": [ + 134, + 264, + 485, + 393 + ], + "type": "table", + "html": "
DatasetsmPLUG-OwlOtterVPGTransLLaVABLIP2VTC(B)MiniGPT4VTC(M)InstructBLIPVTC(I)
Existence120.00195.0070.0050.00160.00170.0068.3372.00185.00188.00
Count50.0088.3385.0050.00135.00138.3955.0055.00143.33144.01
Position50.0086.6763.3350.0077.3380.5143.3345.7666.6770.84
Color55.00113.3373.3355.00148.33152.9075.0075.00153.33154.51
Poster136.06138.7884.0150.00141.85147.3241.8444.00123.81127.04
Celebrity100.29172.6553.5348.82105.59112.6554.4157.63101.18108.74
Scene135.50158.75141.7550.00145.25150.0071.7575.34153.00155.00
Landmark159.25137.2564.7550.00138.00141.2854.0054.0079.7580.34
Artwork96.25129.0077.2549.00136.50137.4460.5064.39134.25134.97
OCR65.0072.5077.5050.00110.00110.0057.5060.0072.5077.50
Commonsense78.57106.4364.2957.14110.00110.9559.2962.55129.29130.03
Numerical60.0072.5050.0050.0040.0042.0045.0045.0040.0042.00
Text Translation80.0057.5077.5057.5065.0065.005.5060.0065.0067.50
Code Reasoning57.5070.0057.5050.0075.0070.5040.0040.0057.5060.00
", + "image_path": "4642b0a58a0dae2ead3a8cf5c729fe325e5776435852c1ac60582e9cda6a047c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 394, + 480, + 426 + ], + "lines": [ + { + "bbox": [ + 132, + 394, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 132, + 394, + 480, + 426 + ], + "type": "text", + "content": "Table 2: Evaluation results (\\%) of perception and recognition on MME. The results are calculated by summing ACC and " + }, + { + "bbox": [ + 132, + 394, + 480, + 426 + ], + "type": "inline_equation", + "content": "\\mathrm{ACC + }" + }, + { + "bbox": [ + 132, + 394, + 480, + 426 + ], + "type": "text", + "content": " in MME [8]. We shorten VTC variants by keeping the first letter of the corresponding base models." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 461, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 481, + 666 + ], + "type": "text", + "content": "Results on MME [8]. MME is another recently proposed benchmark measuring both the visual perception and cognition abilities of MLLMs across 14 datasets. Given a test image, MME designs two prompts, and each of them consists of two parts, including a visual question and an instruction \"Please answer yes or no\". This makes it convenient to report the quantitative performance of accuracy. We report the accuracy score at Table. 2. First, We find that Otter, VTC(B), and VTC(I) achieve 5/14, 3/14, and 3/14 best results, respectively. Otter aligns the vision-language modalities under the in-context learning framework on massive multimodal datasets, while our approach is trained from a single modality. These competitive results indicate our unsupervised training paradigms effectively complete the missing semantics, showing great potential in overcoming the inherent limitation of VPGs. Second, compared to mPLUG-Owl (with 2/14 best results), which finetunes the LLM for multimodal alignments, our approach shows better performance in terms of both perception and cognition. This indicates the efficiency of our image reconstruction loss, which helps VTC complete the image details so that the LLM has sufficient visual evidence for the correct response." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "text", + "content": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 125, + 296, + 236 + ], + "blocks": [ + { + "bbox": [ + 137, + 125, + 296, + 236 + ], + "lines": [ + { + "bbox": [ + 137, + 125, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 137, + 125, + 296, + 236 + ], + "type": "image", + "image_path": "e7673d037ef1940d785a4a6f3f00e147bac7d4f7a673bf57337993892c5e9299.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 249, + 302, + 293 + ], + "lines": [ + { + "bbox": [ + 131, + 249, + 302, + 293 + ], + "spans": [ + { + "bbox": [ + 131, + 249, + 302, + 293 + ], + "type": "text", + "content": "Fig. 4: Zero-shot evaluation results on DEMON. We only report our VTC(InstructBLIP) for clear representation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 147, + 482, + 216 + ], + "blocks": [ + { + "bbox": [ + 312, + 147, + 482, + 216 + ], + "lines": [ + { + "bbox": [ + 312, + 147, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 312, + 147, + 482, + 216 + ], + "type": "image", + "image_path": "5e9ecb33007ac3188a5bb664d7df4ef3788ea856c81898dcadeb492d7901a7da.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 225, + 482, + 260 + ], + "lines": [ + { + "bbox": [ + 311, + 225, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 311, + 225, + 482, + 260 + ], + "type": "text", + "content": "Fig. 5: ChatGPT 4(V) evaluation results of various MLLMs with scores ranging from [1-5]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 316, + 483, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 483, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 483, + 473 + ], + "type": "text", + "content": "Results on DEMON [17]. We adopt DEMON as our third vision-instruction understanding benchmark. Different from the first two evaluations, DEMON designs more complex demonstrative instructions, including interleaved, and multimodal inputs. We report the evaluation score in Fig. 4. First, our approach surrounds other baseline methods in all directions, demonstrating robust multimodal understanding. Moreover, our approach obtains " + }, + { + "bbox": [ + 130, + 316, + 483, + 473 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 316, + 483, + 473 + ], + "type": "text", + "content": " improvement over InstructBLIP in visual relation inference. We attribute this to our reconstruction prior, which offers helpful guidance to help visual tokens memorize the location information of the targets. Lastly, compared to Otter, which shows higher performance on MME, our VTC(InstructBLIP) achieves significant improvements on all tasks. This demonstrates our superior vision-language understanding in complex scenarios, where the reconstruction prior still provides fine-grained guidance, while the captions fail to capture such detailed supervision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 490, + 246, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 246, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 246, + 502 + ], + "type": "text", + "content": "4.3 Further Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "ChatGPT 4(V) evaluation. Besides the above three benchmark evaluations, we further leverage ChatGPT 4(V) [21] to have a systematic measurement of the quality of the proposed model. Specifically, motivated by previous works [1,19] we first randomly select 50 image-instruction pairs from the MIMIC-IT datasets [15] and collect the answers from various MLLMs. ChatGPT 4(V) is then asked to evaluate the quality of answers based on the input image and output the corresponding scores (1-5), where a higher score indicates better overall performance. We report the results in Fig. 5. We find that our VTC(InstructBLIP) is the only model with a score larger than 4, and the second model is our VTC(MiniGPT4). We hope this evaluation results serve as strong evidence for supporting our performance. The evaluation prompt in this section can be found in the Appendix. Information completion of VTC. One of the core ideas behind VTC is to learn complementary visual tokens that capture visual features missed by VPG." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "D. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 114, + 232, + 171 + ], + "blocks": [ + { + "bbox": [ + 160, + 114, + 232, + 171 + ], + "lines": [ + { + "bbox": [ + 160, + 114, + 232, + 171 + ], + "spans": [ + { + "bbox": [ + 160, + 114, + 232, + 171 + ], + "type": "image", + "image_path": "ef2b2bfc3320f7586b6d0eec14bb4fce6ad6dc240d178ce0d09593918e821d9e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 171, + 225, + 179 + ], + "lines": [ + { + "bbox": [ + 178, + 171, + 225, + 179 + ], + "spans": [ + { + "bbox": [ + 178, + 171, + 225, + 179 + ], + "type": "text", + "content": "(a) VTC (MiniGPT4)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 232, + 114, + 304, + 171 + ], + "blocks": [ + { + "bbox": [ + 232, + 114, + 304, + 171 + ], + "lines": [ + { + "bbox": [ + 232, + 114, + 304, + 171 + ], + "spans": [ + { + "bbox": [ + 232, + 114, + 304, + 171 + ], + "type": "image", + "image_path": "1c7c0e7c2b40649ca4f710b779a08f4641957a5997681b62babd1ecd82ffeca7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 248, + 171, + 301, + 179 + ], + "lines": [ + { + "bbox": [ + 248, + 171, + 301, + 179 + ], + "spans": [ + { + "bbox": [ + 248, + 171, + 301, + 179 + ], + "type": "text", + "content": "(b) VTC (InstructBLIP)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 188, + 479, + 221 + ], + "lines": [ + { + "bbox": [ + 130, + 188, + 479, + 221 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 479, + 221 + ], + "type": "text", + "content": "Fig. 6: (a-b) Improvement trends on OKVQA, IconQA, and GQA datasets with different numbers of iterations. (c-d) tag-token nearest neighbor semantic distance " + }, + { + "bbox": [ + 130, + 188, + 479, + 221 + ], + "type": "inline_equation", + "content": "(\\times 100)" + }, + { + "bbox": [ + 130, + 188, + 479, + 221 + ], + "type": "text", + "content": " on Flicker dataset with different bumbers of iterations (" + }, + { + "bbox": [ + 130, + 188, + 479, + 221 + ], + "type": "inline_equation", + "content": "x = 0" + }, + { + "bbox": [ + 130, + 188, + 479, + 221 + ], + "type": "text", + "content": " means the original VPG)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 380, + 170 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 380, + 170 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 380, + 170 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 380, + 170 + ], + "type": "image", + "image_path": "4365fbf6653016dddedd3bacddd0a9ccc68e593913d213442ce6f6753a22fda1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 171, + 374, + 179 + ], + "lines": [ + { + "bbox": [ + 328, + 171, + 374, + 179 + ], + "spans": [ + { + "bbox": [ + 328, + 171, + 374, + 179 + ], + "type": "text", + "content": "(c) VTC (MiniGPT4)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 380, + 114, + 452, + 170 + ], + "blocks": [ + { + "bbox": [ + 380, + 114, + 452, + 170 + ], + "lines": [ + { + "bbox": [ + 380, + 114, + 452, + 170 + ], + "spans": [ + { + "bbox": [ + 380, + 114, + 452, + 170 + ], + "type": "image", + "image_path": "471baf434cf71b215f4f745f4f9d255815dce047102259896a155abb7d99123a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 171, + 449, + 179 + ], + "lines": [ + { + "bbox": [ + 397, + 171, + 449, + 179 + ], + "spans": [ + { + "bbox": [ + 397, + 171, + 449, + 179 + ], + "type": "text", + "content": "(d) VTC (InstructBLIP)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 168, + 236, + 449, + 284 + ], + "blocks": [ + { + "bbox": [ + 168, + 236, + 449, + 284 + ], + "lines": [ + { + "bbox": [ + 168, + 236, + 449, + 284 + ], + "spans": [ + { + "bbox": [ + 168, + 236, + 449, + 284 + ], + "type": "table", + "html": "
DatasetsVTC(BLIP2)VTC(InstructBLIP)
BLIP2BLIP2(64)VTCInstructBLIPInstructBLIP(64)VTC
Perception133.79133.95138.23121.42121.86124.31
Cognition73.9473.9675.5473.0273.4475.92
", + "image_path": "17130b78a29963f704dfca14909692926160e510a0ea2adb73b7c95d730e4169.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 284, + 481, + 327 + ], + "lines": [ + { + "bbox": [ + 130, + 284, + 481, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 481, + 327 + ], + "type": "text", + "content": "Table 3: Abalation results of fine-tuning with LoRA and the number of visual tokens. We finetune LLMs in our model and baselines with LoRA to test the performance of finetuning with the instruction dataset. (64) means we add 32 visual queries to the original model, thus resulting visual tokens have the same length as our VTC." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "text", + "content": "To evaluate such information completion, we here propose the tag-token distance " + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "text", + "content": ", which measures the nearest neighbor semantic distance between the image tags and generated visual tokens: " + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "inline_equation", + "content": "d = \\frac{1}{M} \\sum_{i=1}^{M} \\cosine(\\pmb{t}_i, \\pmb{y}_i)" + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "text", + "content": " is the number of tags, the visual token " + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "inline_equation", + "content": "\\pmb{y}_i" + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "text", + "content": " is the nearest neighbor of tag " + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "inline_equation", + "content": "\\cosine(\\cdot)" + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "text", + "content": " is the cosine distance. We adopt the tagged Flicker dataset [12] to measure the semantic distance, where the average number of tags per image is 8.94. We report the results of VTC(MiniGPT4) and VTC(InstructBLIP) in Fig. 6(c-d), where the " + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "text", + "content": "-axis denotes the number we reuse the VTC module, and " + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "inline_equation", + "content": "x = 0" + }, + { + "bbox": [ + 130, + 365, + 482, + 508 + ], + "type": "text", + "content": " means the original VPG. We find that VTC significantly reduces the semantic distance. That is, the newly generated tokens are able to capture visual concepts missed by the previous tokens, which chimes with our motivations. We also visualize the vision tokens in Fig. 7, which will be introduced later." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Impact of the number of iterations. Due to the flexible design of our VTC, we can reuse the VTC module iteratively at the test stage, once the VTC is trained with the first iteration. This makes it possible to increasingly complete the visual tokens on demand. Yet more iterations mean higher computational costs. We report the impact of the number of iterations in Fig. 6. Fig. 6(a-b) denote the normalized VQA results on OKVQA, IconQV, and GQA datasets with different iterations(e.g.,1,2,3), and Fig. 6(c-d) denote the tag-token distance on Flick dataset. Overall, increasing the number of iterations can improve the performance in terms of VQA and tag-token retrievals. This demonstrates that the newly generated complementary tokens indeed capture more visual concepts, enhancing the LLMs. Besides, we observe that two iterations are sufficient and after two, the gain is not very significant. This is reasonable since our frozen SD model forces the visual tokens " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " to capture as many features as possible to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "text", + "content": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 114, + 479, + 223 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 479, + 223 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 479, + 223 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 479, + 223 + ], + "type": "table", + "html": "
InstructBLIPTokens:tags:man, grass field, frisbeeCaptions:The image features a shirtless man catching a frisbee in a grassy field. The man is in the air, ready to catch the frisbee and throw it into the air. There are several other people present in the scene, including a woman, a man...
VTC(InstructBLIP)-1Complementary tokens:tags:man, grass field, frisbee, outdoor,jumping, park, carCaptions:The image features a man catching a frisbee in a grassy field. The shirtless man is in the air, preparing to throw the frisbee into the air. There are several other people present in the scene, including a man in a red shirt and a man in a blue shirt. There are several cars in the background, including a red car and a white car...
VTC(InstructBLIP)-2Complementary tokens:tags:man, grass field, frisbee, sunny, tree, sky,outdoor, jumping, parkCaptions:The image features a man catching a frisbee in a grassy field. The shirtless man is in the air, preparing to throw the frisbee into the air. There is a cloudy sky in the distance. The man appears to be enjoying a game of frisbee in a park, with a frisbee in his hands. There are several other people present in the scene, including a woman, a man. There are several cars in the background, including ...
", + "image_path": "dcf342fb6d81be695275f18728d7e50d47ce31460614e62d2a46f94f9071b51f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 281, + 447, + 412 + ], + "blocks": [ + { + "bbox": [ + 132, + 232, + 482, + 266 + ], + "lines": [ + { + "bbox": [ + 132, + 232, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 482, + 266 + ], + "type": "text", + "content": "Fig. 7: Visualization of the complementary tokens in the first iteration (VTC(InstructBLIP)-1) and second iteration (VTC(InstructBLIP)-2). We also provide the retrieved tags and predicted captions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 168, + 281, + 447, + 412 + ], + "lines": [ + { + "bbox": [ + 168, + 281, + 447, + 412 + ], + "spans": [ + { + "bbox": [ + 168, + 281, + 447, + 412 + ], + "type": "image", + "image_path": "2791c93a99b99b0a189e52d227f7b61be06c0ab90531594a0a96619c8cb8dda5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 422, + 482, + 456 + ], + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 456 + ], + "type": "text", + "content": "Fig.8: Ablation results of the reconstruction quality. The first row denotes the raw input images. " + }, + { + "bbox": [ + 130, + 422, + 482, + 456 + ], + "type": "inline_equation", + "content": "K = 0" + }, + { + "bbox": [ + 130, + 422, + 482, + 456 + ], + "type": "text", + "content": " denotes the original InstructBLIP and " + }, + { + "bbox": [ + 130, + 422, + 482, + 456 + ], + "type": "inline_equation", + "content": "K = 1" + }, + { + "bbox": [ + 130, + 422, + 482, + 456 + ], + "type": "text", + "content": " denotes we use VTC once to complete the visual tokens." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 483, + 481, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 481, + 508 + ], + "type": "text", + "content": "reconstruct the input image in the first iteration, and therefore there is little detail to pick up in the following iterations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Vision-language disentanglement. After being trained with reconstruction prior, the completed tokens " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " provide LLMs with lossless image messages. Thus we can fix the VTC and VPG and only finetune the LLM for instruction learning. This mitigates the spurious correlations between vision translating and instruction understanding, which are two important components of MLLMs and often entangled with each other in previous VPGs. As the visual tokens are already completed, we can focus on the latter by finetuning LLMs. Specifically, We finetune our LLM and baselines with LoRA [11] on the LLaVA Visual Instruct 150K dataset [19] and report the evaluation results on MME benchmark at Table. 3. We find that fine-tuning LLM improves performance considerably further. Recalling that BLIP2 is trained with image-caption pairs and InstuctBLIP tunes BLIP2 with massive instruction pairs, our VTC(BLIP2) with finetuning outperforms VTC(InstructBLIP) on both the perception and cognition tasks. This" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "D. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 114, + 444, + 514 + ], + "blocks": [ + { + "bbox": [ + 167, + 114, + 444, + 514 + ], + "lines": [ + { + "bbox": [ + 167, + 114, + 444, + 514 + ], + "spans": [ + { + "bbox": [ + 167, + 114, + 444, + 514 + ], + "type": "image", + "image_path": "44360e4497db92cb0580233778ef9e245ea29b7f658054dabd93a0d301c2f3c7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 523, + 482, + 556 + ], + "lines": [ + { + "bbox": [ + 130, + 523, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 523, + 482, + 556 + ], + "type": "text", + "content": "Fig.9: Examples of the test dialogs. VTC shows impressive vision-language understanding in image captioning, human common sense, and visual reasoning, across single and multi-turns." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "demonstrates that our approach uses the instruction pairs more efficiently. Besides, we also test the performance of visual query length. Our VTC learns 32 additional tokens in the visual prompt, we thus add 32 visual queries to the original baselines and jointly train those vectors on the instruction pairs. We find that our VTC achieves higher scores on MME. We attribute this to our reconstruction-aware training, which effectively learns missing concepts." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "type": "text", + "content": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "Ablation studies on reconstruction quality. One of the core ideas behind VTC is to learn reconstruction-aware tokens for comprehensive image representation. The proposed VTC utilizes a frozen SD as the image decoder to complete the visual tokens. To measure the performance of the completed visual tokens, We here aim to test the improvements by measuring the quality of the reconstructed images. Specifically, we report the FID scores on the COCO2014 validation dataset. A lower FID score means closer semantics between the reconstructed images and the original images. From Fig. 8 we find that VTC captures more image details, such as the number of persons, the number of airplanes, and the accurate sky background. The higher reconstruction quality demonstrates the efficiency of the newly generated tokens, thus contributing to the responses of LLMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 277, + 307, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 307, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 307, + 289 + ], + "type": "text", + "content": "4.4 Visualization & Case Studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 298, + 482, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 393 + ], + "type": "text", + "content": "Recall that the cross-attention weights between the query embeddings " + }, + { + "bbox": [ + 130, + 298, + 482, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 130, + 298, + 482, + 393 + ], + "type": "text", + "content": " and the patch embeddings " + }, + { + "bbox": [ + 130, + 298, + 482, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 130, + 298, + 482, + 393 + ], + "type": "text", + "content": " in VPG provide us with an interpretable tool to visualize the complementary tokens. Specifically, we average the attention weight over all tokens and visualize the attention map in Fig. 7. In addition to the attention map, we also provide the retrieved tags and the corresponding predicted caption. We find that the complementary tokens successfully attend to the missing regions and different iterations cover diverse visual concepts. This meets with our complement motivation, resulting in more accurate tags and detailed captions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 394, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 489 + ], + "type": "text", + "content": "We also showcase several test dialogs in Fig. 9. Those examples cover human common sense, reasoning, and knowledge retrieval, across single and multiple turns. For example, our VTC has the ability to reason with human common sense, identify the funny aspect of connecting VGA with smartphones, and the unusual cases in a street picture. Moreover, VTC is able to infer the winner of a tic-tac-toe game and the correct number in a Sudoku puzzle based on pre-trained human knowledge. Those results demonstrate the impressive visual perception ability of our approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 508, + 220, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 508, + 220, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 220, + 521 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": "We in this paper propose VTC, an instruction tuning-free framework to learn complementary visual tokens that capture image details missed by VPG. VTC leverages text-to-image generation to guide the acquisition of reconstruction-aware vision tokens. The enhanced vision prompt is formed by merging these new tokens with the original ones. Additionally, we have developed an iterative strategy to progressively reclaim missing visual information on demand during the testing phase. Extensive experiments are conducted on three widely used multimodal benchmarks to evaluate the proposed model. A ChatGPT 4(V) evaluation is also included to have a systematic measurement. Visualization and case studies demonstrate the interpretability and impressive vision-language understanding of our model." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "D. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant U21B2006; in part by Shaanxi Youth Innovation Team Project; in part by the Fundamental Research Funds for the Central Universities QTZX24003 and QTZX22160; in part by the 111 Project under Grant B18039." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 228, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "type": "text", + "content": "1. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "type": "text", + "content": " * chatgpt quality (March 2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 480, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 294 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 294 + ], + "type": "text", + "content": "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 295, + 480, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 295, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 295, + 480, + 327 + ], + "type": "text", + "content": "3. Anil, R., Dai, A.M., First, O., Johnson, M., Lepikhin, D., Passos, A., Shakeri, S., Taropa, E., Bailey, P., Chen, Z., et al.: Palm 2 technical report. arXiv preprint arXiv:2305.10403 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 327, + 480, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 480, + 370 + ], + "type": "text", + "content": "4. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., et al.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 371, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 371, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 371, + 480, + 403 + ], + "type": "text", + "content": "5. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 404, + 480, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 480, + 436 + ], + "type": "text", + "content": "6. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.C.H.: Instructclip: Towards general-purpose vision-language models with instruction tuning. CoRR abs/2305.06500 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 437, + 480, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 480, + 469 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 480, + 469 + ], + "type": "text", + "content": "7. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 470, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 470, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 470, + 480, + 502 + ], + "type": "text", + "content": "8. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., Li, K., Sun, X., Ji, R.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 502, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 534 + ], + "type": "text", + "content": "9. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 535, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 535, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 535, + 480, + 567 + ], + "type": "text", + "content": "10. He, K., Chen, X., Xie, S., Li, Y., Dólar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 600 + ], + "type": "text", + "content": "1. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "text", + "content": "2. Huiskes, M.J., Lew, M.S.: The mir flickr retrieval evaluation. In: Proceedings of the 1st ACM international conference on Multimedia information retrieval. pp. 39-43 (2008)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "3. Kenton, J.D.M.W.C., Toutanova, L.K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of naacL-HLT. vol. 1, p. 2 (2019)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "text", + "content": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "14. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "type": "text", + "content": "15. Li, B., Zhang, Y., Chen, L., Wang, J., Pu, F., Yang, J., Li, C., Liu, Z.: Mimic-it: Multi-modal in-context instruction tuning (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "type": "text", + "content": "16. Li, B., Zhang, Y., Chen, L., Wang, J., Yang, J., Liu, Z.: Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 480, + 224 + ], + "type": "text", + "content": "17. Li, J., Pan, K., Ge, Z., Gao, M., Zhang, H., Ji, W., Zhang, W., Chua, T.S., Tang, S., Zhuang, Y.: Fine-tuning multimodal llms to follow zero-shot demonstrative instructions. arXiv preprint arXiv:2308.04152 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 224, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 224, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 224, + 480, + 267 + ], + "type": "text", + "content": "18. Li, J., Li, D., Savarese, S., Hoi, S.C.H.: BLIP-2: bootstrapping language-image pretraining with frozen image encoders and large language models. In: International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA. vol. 202, pp. 19730-19742. PMLR (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 268, + 421, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 268, + 421, + 278 + ], + "spans": [ + { + "bbox": [ + 133, + 268, + 421, + 278 + ], + "type": "text", + "content": "19. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 278, + 416, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 278, + 416, + 289 + ], + "spans": [ + { + "bbox": [ + 133, + 278, + 416, + 289 + ], + "type": "text", + "content": "20. OpenAI: GPT-4 technical report. CoRR abs/2303.08774 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 289, + 328, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 289, + 328, + 299 + ], + "spans": [ + { + "bbox": [ + 133, + 289, + 328, + 299 + ], + "type": "text", + "content": "21. OpenAI: GPT-4V(ision) system card (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 299, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 299, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 299, + 480, + 331 + ], + "type": "text", + "content": "22. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback, 2022 13 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 331, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 353 + ], + "type": "text", + "content": "23. Paischer, F., Adler, T., Hofmarcher, M., Hochreiter, S.: Sitta: A semantic image-text alignment for image captioning. arXiv preprint arXiv:2307.05591 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 353, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 353, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 133, + 353, + 480, + 385 + ], + "type": "text", + "content": "24. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 385, + 480, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 385, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 133, + 385, + 480, + 428 + ], + "type": "text", + "content": "25. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 428, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 428, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 133, + 428, + 480, + 449 + ], + "type": "text", + "content": "26. Schonemann, P.H.: A generalized solution of the orthogonal procrustes problem. Psychometrika 31(1), 1-10 (1966)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 449, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 449, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 133, + 449, + 480, + 482 + ], + "type": "text", + "content": "27. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 482, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 482, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 133, + 482, + 480, + 515 + ], + "type": "text", + "content": "28. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 515, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 515, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 133, + 515, + 480, + 536 + ], + "type": "text", + "content": "29. Surís, D., Menon, S., Vondrick, C.: Vipergpt: Visual inference via python execution for reasoning. arXiv preprint arXiv:2303.08128 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 133, + 536, + 480, + 567 + ], + "type": "text", + "content": "30. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 133, + 568, + 480, + 601 + ], + "type": "text", + "content": "31. Tsimpoukelli, M., Menick, J.L., Cabi, S., Eslami, S., Vinyals, O., Hill, F.: Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems 34, 200-212 (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 133, + 601, + 480, + 633 + ], + "type": "text", + "content": "32. Wang, J., Yang, Z., Hu, X., Li, L., Lin, K., Gan, Z., Liu, Z., Liu, C., Wang, L.: Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100 (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "text", + "content": "33. Wang, Z., Yu, J., Yu, A.W., Dai, Z., Tsvetkov, Y., Cao, Y.: Simvlm: Simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904 (2021)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "D. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 500 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "34. Wei, J., Bosma, M., Zhao, V.Y., Guu, K., Yu, A.W., Lester, B., Du, N., Dai, A.M., Le, Q.V.: Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "text", + "content": "35. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "type": "text", + "content": "36. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 215, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 215, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 215, + 480, + 248 + ], + "type": "text", + "content": "37. Xu, P., Shao, W., Zhang, K., Gao, P., Liu, S., Lei, M., Meng, F., Huang, S., Qiao, Y., Luo, P.: Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. arXiv preprint arXiv:2306.09265 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 248, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 480, + 281 + ], + "type": "text", + "content": "38. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "text", + "content": "39. Yin, S., Fu, C., Zhao, S., Li, K., Sun, X., Xu, T., Chen, E.: A survey on multimodal large language models. arXiv preprint arXiv:2306.13549 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 303, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 480, + 325 + ], + "type": "text", + "content": "40. Zhang, A., Fei, H., Yao, Y., Ji, W., Li, L., Liu, Z., Chua, T.S.: Transfer visual prompt generator across llms. arXiv preprint arXiv:2305.01278 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "type": "text", + "content": "41. Zhang, C., Zhang, C., Zhang, M., Kweon, I.S.: Text-to-image diffusion model in generative ai: A survey. arXiv preprint arXiv:2303.07909 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "42. Zhang, S., Dong, L., Li, X., Zhang, S., Sun, X., Wang, S., Li, J., Hu, R., Zhang, T., Wu, F., et al.: Instruction tuning for large language models: A survey. arXiv preprint arXiv:2308.10792 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "43. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "44. Zhao, H., Cai, Z., Si, S., Ma, X., An, K., Chen, L., Liu, Z., Wang, S., Han, W., Chang, B.: Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "text", + "content": "45. Zhao, X., Liu, B., Liu, Q., Shi, G., Wu, X.M.: Making multimodal generation easier: When diffusion models meet llms. arXiv preprint arXiv:2310.08949 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "46. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "text", + "content": "Instruction Tuning-free Visual Token Complement for Multimodal LLMs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Integer-Valued Training and Spike-driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection/b44a5b42-9423-4159-817b-8bb17aae60da_content_list.json b/2024/Integer-Valued Training and Spike-driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection/b44a5b42-9423-4159-817b-8bb17aae60da_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..521e7105747c7a4d1255a425504e4e6f873d0f6b --- /dev/null +++ b/2024/Integer-Valued Training and Spike-driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection/b44a5b42-9423-4159-817b-8bb17aae60da_content_list.json @@ -0,0 +1,1845 @@ +[ + { + "type": "text", + "text": "Integer-Valued Training and Spike-Driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection", + "text_level": 1, + "bbox": [ + 238, + 140, + 764, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinhao Luo $^{1*}$ , Man Yao $^{1*}$ , Yuhong Chou $^{2}$ , Bo Xu $^{1}$ , and Guoqi Li $^{1}$", + "bbox": [ + 220, + 256, + 781, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Automation, Chinese Academy of Sciences {luoxinhao2023, man.yao, guoqi.li}@ia.ac.cn $^{2}$ Xi'an Jiaotong University", + "bbox": [ + 310, + 284, + 691, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Brain-inspired Spiking Neural Networks (SNNs) have bioplausibility and low-power advantages over Artificial Neural Networks (ANNs). Applications of SNNs are currently limited to simple classification tasks because of their poor performance. In this work, we focus on bridging the performance gap between ANNs and SNNs on object detection. Our design revolves around network architecture and spiking neuron. First, the overly complex module design causes spike degradation when the YOLO series is converted to the corresponding spiking version. We design a SpikeYOLO architecture to solve this problem by simplifying the vanilla YOLO and incorporating meta SNN blocks. Second, object detection is more sensitive to quantization errors in the conversion of membrane potentials into binary spikes by spiking neurons. To address this challenge, we design a new spiking neuron that activates Integer values during training while maintaining spike-driven by extending virtual timesteps during inference. The proposed method is validated on both static and neuromorphic object detection datasets. On the static COCO dataset, we obtain $66.2\\%$ mAP@50 and $48.9\\%$ mAP@50:95, which is $+15.0\\%$ and $+18.7\\%$ higher than the prior state-of-the-art SNN, respectively. On the neuromorphic Gen1 dataset, we achieve $67.2\\%$ mAP@50, which is $+2.5\\%$ greater than the ANN with equivalent architecture, and the energy efficiency is improved by $5.7\\times$ . Code: https://github.com/BICLab/SpikeYOLO", + "bbox": [ + 259, + 356, + 743, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Spiking neural network $\\cdot$ Object detection $\\cdot$ Spike-driven $\\cdot$ Neuromorphic vision $\\cdot$ Neuromorphic computing", + "bbox": [ + 259, + 674, + 740, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 744, + 375, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Brain-inspired SNNs are known for their low power consumption [46, 48]. Spiking neurons incorporate spatio-temporal information and emit spikes when the", + "bbox": [ + 212, + 773, + 787, + 805 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*Equal contribution. $(\\boxtimes)$ Corresponding author.", + "bbox": [ + 238, + 811, + 428, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "membrane potentials exceed a threshold [36]. Thus, spiking neurons trigger sparse additions only when they receive a spike and are otherwise idle. This spike-driven enables SNNs to exhibit obvious low-power advantages over ANNs when deployed on neuromorphic chips [7,37,41,61]. However, the negative impact of complex neuronal dynamics and spike-driven nature is that SNNs are difficult to train and have limited task performance and application scenarios [13].", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For example, most applications of SNN algorithms in computer vision are limited to simple image classification tasks [9, 14, 20, 25, 27, 31, 33, 42, 54, 60]. Another more commonly used and challenging computer vision task, object detection, is rarely explored in SNNs. In 2020, Spiking-YOLO [29] provided the first object detection model in deep SNNs, exploiting the method of converting ANN to SNN with thousands of timesteps. In 2023, EMS-YOLO [50] became the first work to use direct training SNNs to handle object detection. Recently, the direct training Meta-SpikeFormer [58] can process the object detection in a pre-training and fine-tuning manner for the first time. However, the performance gap between these works and ANNs is significant. In this work, we aim to bridge this gap and demonstrate the low-power of SNNs and their unique advantages in neuromorphic applications. We achieve this goal through two efforts.", + "bbox": [ + 212, + 237, + 789, + 417 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "First, we design a new architecture, SpikeYOLO, which combines the macro design of YOLO with the micro design of the SNN module. Simply replacing the artificial neurons in the YOLO series [1,43,52] with spiking neurons generally does not work. Existing solutions include establishing the equivalence between ANN activation and spike firing rate [29], or improving the residual design [50]. We argue that another potential reason is that the module design in the YOLO series is too complex, which is effective in ANNs, but not suitable for SNNs. We observe that after the complex YOLO modules are converted into the corresponding spiking versions, there is a phenomenon of spike degradation in which the deep layers almost no longer emit spikes. Therefore, we tend to simplify the design in SpikeYOLO. We only retain the macro architecture in YOLOv83 while using the simple meta SNN block in Meta-SpikeFormer [58] as the basic module and performing micro design.", + "bbox": [ + 212, + 417, + 789, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, we design a novel spiking neuron, Integer Leaky Integrate-and-Fire (I-LIF), to reduce the quantization error of SNNs. Spike-driven is the key to low-power, while exploiting only binary signals will drop performance, especially in challenging object detection. Numerous studies have attempted to mitigate this problem, such as attention-based membrane optimization [63], ternary spike [18], information max [19], converting quantized ANNs [26], optimal ANN2SNN [2]. However, direct training based on optimization strategies can only alleviate errors; the optimal approximation of ANN2SNN requires large timesteps, and it is difficult to exploit the temporal information. In contrast, the idea of the proposed I-LIF is to use integer-valued activations to drop quantization errors and convert them into spikes during inference by extending the virtual timesteps. The features of I-LIF are: 1) Integer-valued training improves performance and is easy to train; 2) Integer activation during training can be equivalent to spike-driven", + "bbox": [ + 212, + 614, + 789, + 811 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Luo et al.", + "bbox": [ + 271, + 114, + 336, + 126 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "3https://github.com/ultralytics/ultralytics", + "bbox": [ + 236, + 824, + 524, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in inference that there is only sparse addition; 3) The temporal dynamics of LIF are reserved, capable of processing neuromorphic object detection.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed method is validated on both static COCO [35] and neuromorphic Gen1 [8] object detection datasets. The performance we report on both datasets far exceeds the existing best models in SNNs. In addition, we explore the performance changes after the mutual conversion of ANN and SNN with the same architecture. The main contributions of this work are:", + "bbox": [ + 212, + 176, + 785, + 251 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- SpikeYOLO. We explore suitable architectures in SNNs for handling object detection tasks and propose SpikeYOLO, which simplifies YOLOv8 and incorporates meta SNN blocks. This inspires us that the complex modules in ANN may not be suitable for SNN architecture design.", + "- I-LIF Spiking Neuron. We propose an I-LIF spiking neuron that combines integer-valued training with spike-driven inference. The former is used to reduce quantization errors in spiking neurons, and the latter is the basis of the low-power nature of SNNs.", + "- Performance. The proposed method achieves outstanding accuracy with low power consumption on object detection datasets, demonstrating the potential of SNNs in complex vision tasks. On the COCO dataset, we obtain $66.2\\%$ mAP@50 and $48.9\\%$ mAP@50:95, which is $+15.0\\%$ and $+18.7\\%$ higher than the prior state-of-the-art SNN, respectively. On the Gen1 dataset, SpikeYOLO is $+2.5\\%$ better than ANN models with $5.7\\times$ energy efficiency" + ], + "bbox": [ + 225, + 261, + 784, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 506, + 395, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SNN Training Method. Training methods have restricted the development of SNN for a long time. To make SNNs deeper, two training methods have been developed. ANN2SNN substitutes the ReLU function with spiking neurons, aiming to mimic the continuous activation by controlling the firing rate of spiking neurons [3, 10, 49]. It can achieve high performance but often requires long timesteps and is difficult to process sequence tasks that exploit the spatiotemporal dynamic nature of SNNs. In contrast, another directly training an SNN leverage gradient surrogate to circumvent the non-differentiability of binary spikes [39, 55]. Direct training is more flexible and requires fewer timesteps, but its performance usually suffers compared to ANNs of the same architecture. In this work, we focus on utilizing directly trained SNNs to process object detection due to its more flexibility in architectural design.", + "bbox": [ + 212, + 537, + 785, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SNN Architecture Design. The architecture of SNNs can be roughly divided into two categories: CNN-based and Transformer-based SNNs. Spiking ResNet has long dominated the SNN field because residual learning [22] can address the performance degradation of SNNs as they become deeper. Typical spiking ResNet includes vanilla spiking ResNet [68], SEW-ResNet [14], and MS-ResNet [27]. The main difference between them is the location of shortcuts and the ability to achieve identity mapping [23]. Recently, the Transformer [12, 51] architecture has become popular [21, 38, 58, 60, 66, 67, 67, 69] in the SNN field and", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SpikeYOLO", + "bbox": [ + 648, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "has refreshed the performance upper bound. The SpikeYOLO we designed draws on the idea of the meta SNN block in Meta-SpikeFormer [58] and merges it with the YOLOv8 architecture.", + "bbox": [ + 212, + 146, + 784, + 190 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Quantization Errors in SNNs. Spiking neurons quantize continuous membrane potentials with complex spatio-temporal dynamics into binary spikes. Obviously, quantization error will limit the performance of SNNs. In ANN2SNN, the reduction of quantization error is relatively simple, which can be achieved by increasing the timestep [2]. In direct training, adjusting the relationship between membrane potential distribution and threshold is the main solution, such as attention mechanism [63] optimizes membrane potential distribution to reduce noise spikes; information max [19] is derived from the information entropy angle to optimize spike firing. However, optimization cannot change the inherently quantized error nature of binary spikes. The solution of this work is to use integer values to reduce quantization errors during training and convert them to binary spikes to ensure spike-driven inference.", + "bbox": [ + 212, + 191, + 787, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object Detection. Object detection is a challenging yet pivotal task. Existing object detection frameworks can be simply categorized into: two-stage frameworks (RCNN series [16, 17, 45]) and single-stage frameworks (YOLO [1, 43, 52], Detr [4, 70] series). The former generate region proposals before determining precise object locations and classifications, whereas the latter directly ascertain these elements, offering a swifter, more streamlined solution. The object detection task has always been difficult for SNNs, and there are few related works. Current works include ANN2SNN-based YOLO [29, 65] and directly-trained spiking YOLO [50, 58]. Their performance is poor, and it isn't easy to meet the needs of real scenarios.", + "bbox": [ + 212, + 373, + 787, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methods", + "text_level": 1, + "bbox": [ + 215, + 547, + 338, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We exploit SpikeYOLO to process both static and neuromorphic object detection datasets. We first introduce how network inputs are unified. Then, the details of SpikeYOLO architecture and I-LIF spiking neuron are presented, respectively.", + "bbox": [ + 212, + 579, + 785, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Network Input", + "text_level": 1, + "bbox": [ + 215, + 647, + 385, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The input of SNNs can be denoted as $X \\in \\mathbb{R}^{T \\times C \\times H \\times W}$ , where $T$ is the timestep, $C$ is the channel, $H \\times W$ denote the spatial resolution.", + "bbox": [ + 212, + 672, + 785, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Static Image. To leverage the spatio-temporal capabilities of SNNs, it is common practice that static images are repeated and utilized as input for each timestep $T$ . This is called direct input encoding [30, 56], where the first layer of spiking neurons in the network encodes the continuous values of the input into spike signals.", + "bbox": [ + 212, + 704, + 785, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Neuromorphic Event Stream. Neuromorphic data (also known as event-based vision) are generated by a Dynamic Vision Sensor (DVS), which only generates spikes when the logarithmic change in light intensity at a pixel surpasses a predefined threshold. An event-based stream is characterized as $(x_{n},y_{n},t_{n},p_{n})$ ,", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Luo et al.", + "bbox": [ + 271, + 114, + 338, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cb459888c0d1bf6d011f1523043b9ae7175ca846f9f9d09964bdadd597f18124.jpg", + "image_caption": [ + "Fig. 1: The overall architecture of SpikeYOLO. We designed two SNN blocks, SNN-Block-1 and SNN-Block-2, and kept other architectures remain as YOLOv8. SNN-Block-1 employs standard convolution within its ChannelConv ( $\\cdot$ ) component, whereas SNN-Block-2 utilizes re-parameterization convolution. That is, the difference between the two is the channel mixer module. In the low and high stages, we use SNN-Block-1 and SNN-Block-2, respectively. The spiking neuron is I-LIF, which activates integer values during training while converting them to binary spikes during inference." + ], + "image_footnote": [], + "bbox": [ + 261, + 143, + 743, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with each event capturing spatial coordinates $(x,y)$ , timestamp $t$ and polarity $p$ , where $p \\in \\{-1,1\\}$ indicates whether light intensity has increased or decreased. Neuromorphic vision offers several advantages [15, 61], such as low resource requirements, high temporal resolution, and high robustness. The spike-driven nature of SNN makes it naturally suitable for processing event streams. The general strategy for neuromorphic pre-processing is to aggregate the event stream within a fixed time window into a frame format [57, 59, 62]. In this work, we follow this operation. Specifically, the total input window length is $T \\times dt$ , where $dt$ and $T$ are temporal resolution and timestep, respectively.", + "bbox": [ + 212, + 547, + 787, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 SpikeYOLO Architecture", + "text_level": 1, + "bbox": [ + 214, + 707, + 470, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overview. SpikeYOLO integrates the macro design of the YOLOv8 with the micro design of Meta-SpikeFormer [58]. The motivation is that we observe complex computations within YOLO's modules result in spike degradation [27] upon direct conversion to the SNN version. Consequently, we maintain the overarching design principles of the YOLO architecture while incorporating the inverted residual structure [47] and re-parameterization convolution [11] design of the Meta-SpikeFormer for detailed aspects.", + "bbox": [ + 212, + 733, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SpikeYOLO", + "bbox": [ + 648, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network Output. In object detection, the network outputs the class and position of each object based on the input image sequence $X = \\{X_{t}\\}_{t = 1}^{T}$ . Suppose the input image sequence has $N$ goals, the output $B = \\{B_{n}\\}_{n = 1}^{N}$ can be calculated:", + "bbox": [ + 215, + 146, + 789, + 193 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nB = \\operatorname {M o d e l} (X), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 205, + 787, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where each $B_{n} = \\{f_{n},c_{n},x_{n},y_{n},w_{n},h_{n}\\}$ contains information about degree of confidence $f_{n}$ , class $C_n$ , center coordinates $(x_{n},y_{n})$ and target size $(w_{n},h_{n})$ . Model $(\\cdot)$ refers to the proposed SpikeYOLO architecture.", + "bbox": [ + 214, + 227, + 789, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Macro Design. Fig. 1 shows the overview of SpikeYOLO, a variation of the YOLO framework that is more suitable for the feature extraction scheme of SNNs. Specifically, YOLOv8 is a classic single-stage detection framework that partitions the image into numerous grids, with each grid responsible for predicting a target independently. Some classic designs, such as the feature pyramid network [34] in YOLOv8, play a crucial role in facilitating efficient feature extraction and fusion. By contrast, its feature extraction module, such as C2F, performs repeated feature extraction from the same set of feature maps. This module can enhance feature extraction in ANNs but does not work well in SNNs. As a compromise, we preserve the classic Backbone/Neck/Head architecture in YOLOv8 while incorporating strategies from the meta SNN block in Meta-SpikeFormer.", + "bbox": [ + 214, + 292, + 789, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Micro Design. Meta-SpikeFormer [58] is the current state-of-the-art architecture in SNNs, which explores the meta design of SNN and consists of CNN-based and Transformer-based SNN blocks. The meta block comprises a token mixer module and a channel mixer module. The difference between CNN-based and Transformer-based SNN blocks lies in the token mixer, which are spike-driven convolution and spike-driven self-attention, respectively. In this work, we mainly redesign the channel mixer module for the object detection task. As shown in Fig. 1, SNN-Block-1 and SNN-Block-2 are designed to extract low-stage and high-stage features, respectively.", + "bbox": [ + 214, + 479, + 787, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The meta SNN block in [58] can be written as:", + "bbox": [ + 238, + 617, + 580, + 631 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nU ^ {\\prime} = U + \\operatorname {S e p C o n v} (U), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 641, + 787, + 659 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nU ^ {\\prime \\prime} = U ^ {\\prime} + \\operatorname {C h a n n e l C o n v} \\left(U ^ {\\prime}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 681, + 787, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $U \\in \\mathbb{R}^{T \\times C \\times H \\times W}$ is the layer input, $\\operatorname{SepConv}(\\cdot)$ is an inverted separable convolution module [47] with $7 \\times 7$ kernel size in MobileNetv2 to capture global features, followed by a $3 \\times 3$ depthwise convolution for further spatial feature fusion. Sepconv ( $\\cdot$ ) can be expressed as:", + "bbox": [ + 214, + 703, + 787, + 765 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S e p C o n v} (U) = \\operatorname {C o n v} _ {\\mathrm {d w} 2} \\left(\\operatorname {C o n v} _ {\\mathrm {p w} 2} \\left(S N \\left(\\operatorname {C o n v} _ {\\mathrm {d w} 1} \\left(S N \\left(\\operatorname {C o n v} _ {\\mathrm {p w} 1} \\left(S N (U)\\right)\\right)\\right)\\right)\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 786, + 787, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathrm{Conv}_{\\mathrm{dw1}}(\\cdot)$ and $\\mathrm{Conv}_{\\mathrm{dw1}}(\\cdot)$ are depthwise convolutions, $\\mathrm{Conv}_{\\mathrm{pw1}}(\\cdot)$ and $\\mathrm{Conv}_{\\mathrm{pw1}}(\\cdot)$ are pointwise convolutions [5]. $SN(\\cdot)$ is the spiking neuron layer.", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Luo et al.", + "bbox": [ + 271, + 114, + 338, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ChannelConv $(\\cdot)$ is the channel mixer, which facilitates inter-channel information fusion. We redesign ChannelConv $(\\cdot)$ in object detection task.", + "bbox": [ + 214, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For SNN-Block-1, it is written as:", + "bbox": [ + 238, + 176, + 485, + 191 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C h a n n e l C o n v 1} \\left(U ^ {\\prime}\\right) = \\operatorname {C o n v} \\left(S N \\left(\\operatorname {C o n v} \\left(S N \\left(U ^ {\\prime}\\right)\\right)\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 312, + 199, + 787, + 215 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{Conv}(\\cdot)$ is a standard convolution with expansion ratio $r = 4$ . In contrast, addressing high-stage features, SNN-Block-2 employs a re-parameterization convolution to minimize parameter count, which can be described as:", + "bbox": [ + 214, + 224, + 782, + 268 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C h a n n e l C o n v 2} \\left(U ^ {\\prime}\\right) = \\operatorname {R e p C o n v} \\left(S N \\left(\\operatorname {R e p C o n v} \\left(S N \\left(U ^ {\\prime}\\right)\\right)\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 282, + 282, + 787, + 300 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {R e p C o n v} \\left(U ^ {\\prime}\\right) = \\operatorname {C o n v} _ {\\mathrm {p w} 2} \\left(\\operatorname {C o n v} _ {\\mathrm {d w} 1} \\left(\\operatorname {C o n v} _ {\\mathrm {p w} 1} \\left(U ^ {\\prime}\\right)\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 319, + 785, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{RepConv}(\\cdot)$ is the re-parameterization convolution [11] with kernel size $3\\times 3$ , it can be re-parameterized to a standard convolution during inference.", + "bbox": [ + 212, + 340, + 782, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 I-LIF Spiking Neuron", + "text_level": 1, + "bbox": [ + 214, + 388, + 441, + 404 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Spiking neurons propagate information in both spatial and temporal domains, and they mimic the spiking communication scheme of biological neurons. However, there are inherent quantization errors in converting the membrane potential of spiking neurons into binary spikes, which severely limits the representation of the model. Recently, Fast-SNN [26] achieves high-performance conversion with small timesteps by converting quantized ANNs into SNNs. This inspires us to \"why not train directly with integer values\", which can significantly reduce the quantization error. We just need to be careful to ensure that the inference is spike-driven. So, we came up with the idea of I-LIF.", + "bbox": [ + 212, + 412, + 782, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LIF. Leaky Integrate-and-Fire (LIF) spiking neuron [36] is the most popular neuron to construct SNNs due to its balance between bio-plausibility and computing complexity. The dynamics of LIF with soft reset is:", + "bbox": [ + 214, + 566, + 782, + 612 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nU [ t ] = H [ t - 1 ] + X [ t ], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 627, + 785, + 645 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nS [ t ] = \\Theta (U [ t ] - V _ {t h}), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 662, + 785, + 680 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nH [ t ] = \\beta (U [ t ] - S [ t ]) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 698, + 785, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $t$ denotes the timestep, $U[t]$ is the membrane potential that integrates the temporal information $H[t - 1]$ and spatial information $X[t]$ . $\\Theta(\\cdot)$ is the Heaviside step function which equals 1 for $x \\geq 0$ and 0 otherwise. If $U[t]$ exceeds the firing threshold $V_{th}$ , spiking neuron fire a spike $S[t]$ and $U[t]$ will subtract it subsequently. Otherwise, $H[t]$ will remain unchanged. And, $U[t]$ decays to $H[t]$ by a factor of $\\beta$ , which denotes the decay constant. For simplicity, we focus on Eq.9 and denote the spiking neuron layer as $SN(\\cdot)$ , with its input as membrane potential tensor $U$ and its output as spike tensor $S$ .", + "bbox": [ + 212, + 719, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SpikeYOLO", + "bbox": [ + 647, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c04ec5cb61c3d5c74c39dc427d0bc5643b41658648c405cb0cc440bbba48030.jpg", + "image_caption": [ + "Fig. 2: Comparison of I-LIF and LIF. Binary spikes are emitted by LIF during both training and inference processes, which results in quantization errors. I-LIF emits integer values during the training process to reduce quantization errors, and converts them into binary spikes during inference to make the network only perform sparse addition." + ], + "image_footnote": [], + "bbox": [ + 263, + 143, + 738, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "I-LIF. We propose the Integer Leaky Integrate-and-Fire (I-LIF) neuron to reduce the quantization error. As shown in Fig. 2, I-LIF emits integer values while training, and converts them into $0/1$ spikes when inference. Specifically, in I-LIF, Eq.9 is rewritten as:", + "bbox": [ + 212, + 453, + 787, + 513 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nS [ t ] = \\operatorname {C l i p} (r o u n d (U [ t ]), 0, D), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 523, + 787, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\text{round}(\\cdot)$ is a round symbol, $\\text{Clip}(x, \\min, \\max)$ denotes that clipping $x$ to $[\\min, \\max]$ , $D$ is a hyperparameter indicating the maximum emitted integer value by I-LIF.", + "bbox": [ + 212, + 547, + 787, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training Stage. Eq.11 is not a continuous function, making its derivative a step function, potentially causing training instability. Previous studies have introduced several surrogate gradient functions, which primarily address binary spike outputs. We consistently utilize rectangular windows as the surrogate function. For simplicity, We retain gradients solely for neurons activated in the $[0,D]$ range, nullifying all others.", + "bbox": [ + 212, + 594, + 787, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inference Stage. Introducing integer value necessitates additional MACs (Multiply-Accumulation operations), potentially diminishing the energy efficiency of SNNs. Thus, converting integer values to binary spikes is essential. Fig. 3 shows an example of how integer values convert to binary spikes by extending virtual timesteps during inference. Specifically, the input to the spiking neuron at $l + 1$ layer can be described as $X^{l + 1}[t] = W^l S^l [t]$ . We extend the $T$ time step to $T\\times D$ , and convert the integer value $S^l [t]$ to a spike sequence $\\{S^l [t,d]\\}_{d = 1}^D$ , which satisfied:", + "bbox": [ + 212, + 684, + 789, + 803 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {d = 1} ^ {D} S ^ {l} [ t, d ] = S ^ {l} [ t ]. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 801, + 787, + 843 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Luo et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8f44748fd2c812e1cdec7d4cef98af5b8d983eb8e80bf33ca8a9ab624a9c6a76.jpg", + "image_caption": [ + "Fig. 3: An example of how the proposed I-LIF works. We assume $T = 3, D = 2$ , and show the corresponding binary spike sequences of integer value during inference. The membrane potential in [0.5, 1.5) are quantized to 1, while those in [1.5, 2.5) are quantized to 2. membrane potential that $> 2.5$ are also quantized to 2 due to the maximum integer value $D = 2$ . Subsequently, the membrane potential will be subtracted from the integer value. The training spike with a value of 2 will be converted into two binary spikes by extending virtual timesteps during inference." + ], + "image_footnote": [], + "bbox": [ + 259, + 141, + 743, + 308 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Thus, the neuron's input at $l + 1$ layer is reformulated as:", + "bbox": [ + 215, + 443, + 632, + 458 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nX ^ {l} [ t ] = W ^ {l} \\sum_ {d = 1} ^ {D} S ^ {l} [ t, d ]. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 465, + 787, + 507 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Given that matrix multiplication functions as linear operators, we establish:", + "bbox": [ + 215, + 515, + 761, + 531 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nW ^ {l} \\sum_ {d = 1} ^ {D} S ^ {l} [ t, d ] = \\sum_ {d = 1} ^ {D} \\left(W ^ {l} S ^ {l} [ t, d ]\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 537, + 787, + 579 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Therefore, the input of the neuron at $l + 1$ layer can be computed by:", + "bbox": [ + 215, + 585, + 712, + 602 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nX ^ {l} [ t ] = \\sum_ {d = 1} ^ {D} \\left(W ^ {l} S ^ {l} [ t, d ]\\right). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 609, + 785, + 650 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The spike sequence $S^l[t,d]$ only contains $0/1$ , so all MACs can be converted into sparse ACs(Accumulation operations), which can ensure spike-driven when inference.", + "bbox": [ + 214, + 659, + 787, + 705 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 726, + 375, + 743 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate the proposed method on COCO 2017 val [35] and neuromorphic Gen1 [8] datasets. The mean Average Precision(mAP) at IOU=0.5(mAP@50), the average mAP between 0.5 and 0.95(mAP@50:95), and energy cost are reported for each model. Specifically, the power of ANNs and SNNs can be calculated as:", + "bbox": [ + 214, + 753, + 787, + 816 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nE _ {A N N} = O ^ {2} \\times C _ {i n} \\times C _ {o u t} \\times k ^ {2} \\times E _ {M A C}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 823, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SpikeYOLO", + "bbox": [ + 648, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/2741f760f7575439ad0768b762cf5ffa8ee26ba0dd0e02b5f9bd7eb4434b1c51.jpg", + "table_caption": [ + "Table 1: Results on COCO 2017 val [35]. $T \\times D$ means that we set up $T$ timesteps, and each timestep is expanded $D$ times. In prior SNNs and ANNs, $D$ defaults to 1." + ], + "table_footnote": [], + "table_body": "
ArchitectureModelParam (M)Power (mJ)T × DmAP@ 50(%)mAP@ 50:95(%)
ANNPVT [53]32.9520.3159.236.7
DETR [4]41.0197.8162.442.0
YOLOv5 421.2112.5164.145.4
ANN2SNNSpiking-Yolo [29]10.2-3500-25.7
Bayesian Optim [28]10.2-5000-25.9
Spike Calib [32]17.1-51245.4-
Directly-trained SNNEMS-YOLO [50]26.929.0450.130.1
Meta-SpikeFormer (MaskRCNN) [58]34.949.5144.0-
75.0140.8151.2-
Meta-SpikeFormer (YOLO) [58]16.834.8145.0-
16.870.7450.3-
SpikeYOLO(Ours)13.223.11 × 459.242.5
23.118.41 × 152.736.1
23.134.61 × 462.345.5
23.167.64 × 155.738.7
23.1134.74 × 463.346.3
48.168.51 × 464.647.4
68.884.21 × 466.248.9
", + "bbox": [ + 238, + 184, + 766, + 479 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nE _ {S N N} = (T \\times D) \\times f r \\times O ^ {2} \\times C _ {i n} \\times C _ {o u t} \\times k ^ {2} \\times E _ {A C}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 521, + 787, + 539 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $O$ is the feature output size, $C_{in}$ and $C_{out}$ denotes the number of input channel and output channel, $k$ is the kernel size, $fr$ denotes the average spike firing rate, $T$ is the timestep, $D$ is the upper limit of integer activation during training. We follow the most commonly used energy consumption evaluation method in the SNN field [40,63,64]. All operations assume a 32-bit floating-point implementation on $45\\mathrm{nm}$ technology, where $E_{MAC} = 4.6pJ$ and $E_{AC} = 0.9pJ$ [24]. As can be seen from Eq. 16 and 17, SNN's low power comes from its sparse addition operation. The fewer spikes, the sparser the computation.", + "bbox": [ + 212, + 545, + 787, + 667 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 COCO 2017 val Dataset", + "text_level": 1, + "bbox": [ + 214, + 688, + 460, + 702 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Experimental Setup. As a predominant static dataset for object detection, COCO 2017 val [35] comprises 80 classes split into 118K training and 5K validating images. In all experiments, we set decay factor $\\beta = 0.25$ , learning rate to 0.01, and adopt SGD optimizer. The models are trained for 300 epochs with a batch size of 40 on 4 NVIDIA V100 GPUs. Mosaic data augmentation [1] technique is employed. The network structure is given in the supplementary material. Note,", + "bbox": [ + 212, + 713, + 787, + 806 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Luo et al.", + "bbox": [ + 271, + 114, + 336, + 126 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "4https://github.com/ultralytics/yolov5", + "bbox": [ + 236, + 824, + 500, + 840 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/35c4c9d4985b1dd93f0352d17af5f12eddba21d03fed6cfa38f6e6f442efdd8f.jpg", + "table_caption": [ + "Table 2: Ablation studies of architectural design. We first convert YOLOv8 directly into the corresponding spiking version. Then we convert the SpikeYOLO designed in this work into the corresponding ANN version." + ], + "table_footnote": [], + "table_body": "
ArchitectureModelParam (M)Power (mJ)T × DmAP@ 50(%)mAP@ 50:95(%)
ANN → SNNYOLOv825.8183.5167.250.2
Spiking YOLOv825.87.21 × 146.831.3
SNN → ANNSpikeYOLO (Ours)23.118.41 × 152.736.1
YOLO23.1314.1165.048.1
", + "bbox": [ + 250, + 198, + 753, + 299 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "in our method, inference timestep is reported as $T \\times D$ , e.g., $1 \\times 4$ denotes $T = 1, D = 4$ .", + "bbox": [ + 212, + 325, + 782, + 354 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Main Results are shown in Table 1. The proposed SpikeYOLO significantly improves the performance upper bound of the COCO dataset in SNNs. We obtain $66.2\\%$ mAP@50 and $48.9\\%$ mAP@50:95, which is $+15.0\\%$ and $+18.7\\%$ higher than the prior state-of-the-art SNN [58], respectively. SpikeYOLO also has significant advantages over existing SNNs in terms of parameters and power: SpikeYOLO vs. EMS-YOLO [50]: Param, 23.1M vs. 26.9M; mAP@50, $62.3\\%$ vs. $50.1\\%$ ; mAP@50:95, $45.5\\%$ vs. $30.1\\%$ ; Power, $33.2\\mathbf{mJ}$ vs. $29.0\\mathrm{mJ}$ . Moreover, the performance gap between SNNs and ANNs is significantly narrowed. For example, under similar parameters, the performance of SpikeYOLO and YOLO v5 are comparable, and the energy efficiency is $3.3\\times$ .", + "bbox": [ + 212, + 356, + 787, + 507 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ablation Studies of Architectural Design. We simplify YOLOv8 for SNN and incorporate meta SNN blocks. As shown in Table 1, this architectural improvement enables the accuracy of spikeYOLO at $T = 1$ and $D = 1$ to reach $52.7\\%$ , better than the prior state-of-the-art SNN. We are also interested in the question \"whether the architectures in SNNs and ANNs can be used directly interchangeably?\" We conduct the experiments in Table 2. We observe that directly converting the ANN architecture into the corresponding SNN brings significant performance degradation. The special architectural design of SNNs can improve its representation.", + "bbox": [ + 212, + 507, + 787, + 642 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ablation Studies of Quantization Error. Integer-valued training is designed to reduce quantization error in SNNs. The larger $D$ is, the smaller the quantization error is. In Table 1, we fixed the parameters to 23.1M. When $T = 1$ and $T = 4$ , we expand $D = 1$ to $D = 4$ , respectively, and the accuracies of mAP@50 are increased by $+9.6\\%$ and $+7.6\\%$ . In contrast, if we fix $D = 1$ and increase $T = 1$ to $T = 4$ , the performance improvement of mAP@50 is only $3\\%$ . These results show that quantization error has a greater impact on performance than the setting of timesteps. And, in terms of power, increasing $D$ is more cost-effective than increasing $T$ . For instance, when $1 \\times 1$ changes to $1 \\times 4$ , power increases by $88\\%$ ; while $1 \\times 1$ changes to $4 \\times 1$ , power improves by $267\\%$ .", + "bbox": [ + 212, + 643, + 787, + 792 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Gen1 Automotive Detection Dataset", + "text_level": 1, + "bbox": [ + 214, + 815, + 562, + 829 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SpikeYOLO", + "bbox": [ + 647, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/16afe8a08478fb9313be8f681fcb9517c1796ad3ec5d2233b684e07ee932e1f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 145, + 385, + 222 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/bd2aabd570bffd763fdb08c07642b82465fe3c210297322b628aff711960c945.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 146, + 500, + 220 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/32abed312d109b59cdb0623658e92f9023bbf8b2a080975ff7108931d20043c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 146, + 614, + 220 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b4d530beb57dbeeeb266f98201b1250c1aab75660ec67d6f123556976a89cfe0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 146, + 727, + 220 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/fe6c64c5af7da18aa5399b7ab012a7fc0b88ee2541e107bf5f7ee0394f2e4140.jpg", + "image_caption": [ + "Fig. 4: The object detection results on the COCO dataset. The first two columns compare the effect of maximum integer value $D$ on performance for the same structure. The second and third columns compare the effect of the size of the model on performance." + ], + "image_footnote": [], + "bbox": [ + 274, + 222, + 385, + 345 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/57972062b0a46e1512e7b6aa87dcb4123b92292af0b8e911b4aeb21f685dfc65.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 222, + 500, + 345 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4b0ec6a5b167585a656232354c002b1eb31a37f80a9333b9f1174b747abf5357.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 222, + 612, + 345 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b4f3de35749ccced1ac59aec97410af3c58f9a76913687f1bf3214afe8eb50c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 222, + 727, + 345 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Experimental Setup. As a large neuromorphic object detection dataset, Gen1 [8] encompasses 39 hours of open road and various driving scenarios, captured using an ATIS sensor with a resolution of $304 \\times 240$ pixels. The dataset is organized into training, validation, and testing subsets. The bounding box annotations of pedestrians and cars (over 255,000) were manually labeled. For each annotation, we process the event-based stream 2.5 seconds before its occurrence, dividing it into $T$ slices for model input. We train the model for 50 epochs and maintain other hyperparameters same as the COCO 2017 dataset.", + "bbox": [ + 212, + 431, + 787, + 551 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Main Results on Gen1 dataset are shown in Table 3. The proposed SpikeYOLO notably elevates the performance benchmark for the Gen1 dataset in SNNs. We achieve $67.2\\%$ mAP@50 with 23.1M parameters, which outperforms the prior state-of-the-art SNN model by $+8.2\\%$ . For example, when $T = 5$ , SpikeYOLO vs. EMS-YOLO [50]: Param, 13.2M vs. 14.4M; mAP@50, $66.0\\%$ vs. $59.0\\%$ ; mAP@50:95, $38.5\\%$ vs. $31.0\\%$ . In contrast to the COCO dataset, Gen1 contains temporal information that is more suitable for SNN processing. We conduct experiments on the performance of SNN and ANN with the same architecture. We observe that SpikeYOLO's mAP@50 accuracy is $+2.5\\%$ higher than the corresponding ANN, and shows a $5.7\\times$ energy efficiency. This indicates that SNN has attractive potential in processing neuromorphic data.", + "bbox": [ + 212, + 551, + 787, + 718 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation Studies of Quantization Error. Both $T$ and $D$ significantly influence outcomes when processing neuromorphic datasets. Table 4 gives a comprehensive ablation study on SpikeYOLO with 23.1M parameters that evaluate the effects of varying $T$ and $D$ . We observe some interesting experimental results. First, boosting the timestep $T$ will bring improvement in accuracy and power. For instance, with the set of $D = 1$ , extending $T = 1$ to $T = 4$ yields a $+6.7\\%$ increase in mAP@50, and the power will increase by $3.7\\times$ . But further extending $T = 4$ to $T = 8$ results in a marginal increase of only $+0.7\\%$ and significantly", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Luo et al.", + "bbox": [ + 271, + 114, + 336, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/85b2871f573a0247eb1d9983e51a4f8d3f2922d4f0f96eaf246b264ad82c8955.jpg", + "table_caption": [ + "Table 3: Results on the Gen1 dataset [8]. * We convert 23.1M of SpikeYOLO into ANN with the same architecture." + ], + "table_footnote": [], + "table_body": "
ArchitectureModelParam (M)Power (mJ)T × DmAP@ 50(%)mAP@ 50:95(%)
ANNYOLOv3-tiny [44]10.25.1144.5-
SpikeYOLO*23.178.5164.739.7
SNNEMS-YOLO [50]6.21.2554.726.7
9.32.0556.528.6
14.43.4559.031.0
VGG-11+SDD12.611.11-17.4
MobileNet-64+SSD24.35.71-14.7
DenseNet121-24+SSD [6]8.23.91-18.9
Spiking-Yolo [29]7.9102.350044.2-
Tr-Spiking-Yolo [65]7.90.9545.3-
SpikeYOLO(Ours)13.211.05 × 166.038.5
23.119.75 × 166.438.9
23.112.94 × 267.240.4
", + "bbox": [ + 236, + 184, + 767, + 409 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0079e7d018e7539d81ef4a45d9786cabc5cf8afaf89c001478378d9c769445a9.jpg", + "table_caption": [ + "Table 4: The influence of $T$ and $D$ on Gen1. We set SpikeYOLO (23.1M) as the baseline and vary $T$ and $D$ for each study." + ], + "table_footnote": [], + "table_body": "
MethodT × DPower(mJ)mAP@50(%)mAP@50:95(%)
SpikeYOLO1 × 14.059.333.1
1 × 43.9 (-0.1)65.1 (+5.8)38.9 (+5.8)
2 × 18.163.636.5
2 × 27.8 (-0.3)66.1 (+2.5)39.0 (+2.5)
2 × 47.1 (-1.0)67.0 (+3.4)40.1 (+3.6)
4 × 114.866.038.4
4 × 212.9 (-1.9)67.2 (+1.2)40.4 (+2.0)
8 × 127.066.739.3
", + "bbox": [ + 292, + 463, + 710, + 606 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "increases energy cost. Second, We were surprised to see that by expanding $D$ at a fixed $T$ , the performance will be improved, while the power will be dropped. For example, $2 \\times 1$ vs. $2 \\times 2$ vs. $2 \\times 4$ : mAP@50, $63.6\\%$ vs. $66.1\\%$ vs. $67.0\\%$ ; Power, $8.1\\mathrm{mJ}$ vs. $7.8\\mathrm{mJ}$ vs. $7.1\\mathrm{mJ}$ . This trend is completely different from SpikeYOLO's performance in COCO, where the increase of $D$ will bring more energy cost. We argue that this phenomenon is because SNN exhibits various spike firing for dense/sparse data.", + "bbox": [ + 212, + 635, + 787, + 742 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Architecture Ablation Experiments", + "text_level": 1, + "bbox": [ + 214, + 766, + 553, + 782 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Re-parameterization Design. As shown in Table 5, if we remove re-parameterization by adding neurons into inverted separable convolutions, the mAP@50 and mAP@50:95 will decrease $1.7\\%$ and $1.8\\%$ respectively.", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SpikeYOLO", + "bbox": [ + 648, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/c0e0124565300da2d5435a82229de818368d500729735e40abc8051754823caa.jpg", + "table_caption": [ + "Table 5: Ablation studies of architecture design. We set $T \\times D = 1 \\times 4$ and modify just one point of baseline to test how the parameters, power and performance vary." + ], + "table_footnote": [], + "table_body": "
MethodParam(M)mAP@50(%)mAP@50:95(%)
SpikeYOLO(Baseline)23.162.345.5
Remove re-parameterization23.160.643.7
SNN-Block-1 → SNN-Block-219.961.244.7
SNN-Block-2 → Transformer Block24.561.044.2
Anchor-free head → Anchor-based head21.259.539.7
", + "bbox": [ + 235, + 184, + 767, + 286 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "SNN Block Design. Including a $3 \\times 3$ standard convolution within the initial stages of convolution blocks is crucial. As shown in Table 5, substituting SNN-Block-1 for SNN-Block-2 leads to a reduction in performance of around $1\\%$ . Moreover, we try to replace high-stage SNN-Block-2 with meta Transformer-based SNN block, just like Meta-SpikeFormer [58]. We find that there is little to no performance gain by doing this and that the parameters increase. Therefore, only spiking CNN blocks are exploited in our SpikeYOLO.", + "bbox": [ + 212, + 313, + 787, + 419 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Detection Head. The detection mechanisms within YOLO are categorized into anchor-based heads(e.g.,YOLOv5) and anchor-free heads(e.g.,YOLOv8). The former directly predicts each bounding box's dimensions, whereas the latter estimates the probability distribution of each bounding box. Previous EMSYOLO [50] and Meta-SpikeFormer [58] employ anchor-based heads. SpikeYOLO exploits the anchor-free head because of its higher accuracy (see Table 5).", + "bbox": [ + 212, + 419, + 789, + 511 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 532, + 359, + 549 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work significantly narrows the performance gap between SNNs and ANNs on object detection tasks. We achieve this through network architecture and spiking neuron design. The proposed SpikeYOLO architecture abandons the complex module design in the vanilla YOLO series and exploits simple meta spike blocks to build the model. Then, the I-LIF spiking neuron capable of integer-valued training and spike-driven inference is proposed to drop quantization errors. We improve the upper bound of the SNN domain's performance on the COCO dataset by $+15.0\\%$ (mAP@50) and $+18.7\\%$ (mAP@50:95), respectively. On the neuromorphic Gen1 dataset, SpikeYOLO achieves better performance and lower power than ANN of the same architecture. Furthermore, we investigate the performance of equivalent architecture ANNs and SNNs in different datasets, and the results show that the redesigned SNN architecture performed better. This work enables SNNs to handle complex object detection and can inspire the application of SNNs in more visual scenarios.", + "bbox": [ + 212, + 565, + 789, + 777 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Luo et al.", + "bbox": [ + 271, + 114, + 336, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was partially supported by National Distinguished Young Scholars (62325603), and National Natural Science Foundation of China (62236009, U22A20103,62441606), Beijing Natural Science Foundation for Distinguished Young Scholars (JQ21015), China Postdoctoral Science Foundation (GZB20240824, 2024M753497), and CAAI-MindSpore Open Fund, developed on OpenI Community.", + "bbox": [ + 212, + 175, + 792, + 265 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 287, + 321, + 303 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bochkovskiy, A., Wang, C.Y., Liao, H.Y.M.: Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934 (2020)", + "2. Bu, T., Fang, W., Ding, J., DAI, P., Yu, Z., Huang, T.: Optimal ANN-SNN conversion for high-accuracy and ultra-low-latency spiking neural networks. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=7B3IJMM1k_M", + "3. Cao, Y., Chen, Y., Khosla, D.: Spiking deep convolutional neural networks for energy-efficient object recognition. International Journal of Computer Vision 113, 54-66 (2015)", + "4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European Conference on Computer Vision. pp. 213-229. Springer (2020)", + "5. Chollet, F.: Xception: Deep learning with depthwise separable convolutions. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1800-1807 (2017)", + "6. Cordone, L., Miramond, B., Thierion, P.: Object detection with spiking neural networks on automotive event data. In: 2022 International Joint Conference on Neural Networks. pp. 1-8. IEEE (2022)", + "7. Davies, M., Srinivasa, N., Lin, T.H., Chinya, G., Cao, Y., Choday, S.H., Dimou, G., Joshi, P., Imam, N., Jain, S., et al.: Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro 38(1), 82-99 (2018)", + "8. De Tournemire, P., Nitti, D., Perot, E., Migliore, D., Sironi, A.: A large scale event-based detection dataset for automotive. arXiv preprint arXiv:2001.08499 (2020)", + "9. Deng, S., Li, Y., Zhang, S., Gu, S.: Temporal efficient training of spiking neural network via gradient re-weighting. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=_XNtisL32jv", + "10. Diehl, P.U., Neil, D., Binas, J., Cook, M., Liu, S.C., Pfeiffer, M.: Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In: 2015 International Joint Conference on Neural Networks. pp. 1-8. IEEE (2015)", + "11. Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 13733-13742 (2021)", + "12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=YicbFdNTTy" + ], + "bbox": [ + 218, + 316, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SpikeYOLO", + "bbox": [ + 648, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Eshraghian, J.K., Ward, M., Neftci, E.O., Wang, X., Lenz, G., Dwivedi, G., Bennamoun, M., Jeong, D.S., Lu, W.D.: Training spiking neural networks using lessons from deep learning. Proceedings of the IEEE 111(9), 1016-1054 (2023)", + "14. Fang, W., Yu, Z., Chen, Y., Huang, T., Masquelier, T., Tian, Y.: Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems 34, 21056-21069 (2021)", + "15. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(1), 154-180 (2020)", + "16. Girshick, R.: Fast r-cnn. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1440-1448 (2015)", + "17. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 580-587 (2014)", + "18. Guo, Y., Chen, Y., Liu, X., Peng, W., Zhang, Y., Huang, X., Ma, Z.: Ternary spike: Learning ternary spikes for spiking neural networks. arXiv preprint arXiv:2312.06372 (2023)", + "19. Guo, Y., Chen, Y., Zhang, L., Liu, X., Wang, Y., Huang, X., Ma, Z.: Im-loss: information maximization loss for spiking neural networks. Advances in Neural Information Processing Systems 35, 156-166 (2022)", + "20. Guo, Y., Liu, X., Chen, Y., Zhang, L., Peng, W., Zhang, Y., Huang, X., Ma, Z.: Rmp-loss: Regularizing membrane potential distribution for spiking neural networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17391-17401 (2023)", + "21. Han, M., Wang, Q., Zhang, T., Wang, Y., Zhang, D., Xu, B.: Complex dynamic neurons improved spiking transformer network for efficient automatic speech recognition. Thirty-Seventh AAAI Conference on Artificial Intelligence (AAAI 2023) (2023)", + "22. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)", + "23. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) Computer Vision - ECCV 2016. pp. 630-645. Springer International Publishing, Cham (2016)", + "24. Horowitz, M.: 1.1 computing's energy problem (and what we can do about it). In: 2014 IEEE International Solid-State Sircuits Conference digest of technical papers. pp. 10-14. IEEE (2014)", + " 25. Hu, J., Yao, M., Qiu, X., Chou, Y., Cai, Y., Qiao, N., Tian, Y., XU, B., Li, G.: High-performance temporal reversible spiking neural networks with $\\mathcal{O}\\{0\\}(1)$ training memory and $\\mathcal{M} \\mathcal{H} \\mathcal{C} \\mathcal{A}\\{0\\}(1)\\$ inference cost. In: Forty-first International Conference on Machine Learning (2024)", + "26. Hu, Y., Zheng, Q., Jiang, X., Pan, G.: Fast-snn: Fast spiking neural network by converting quantized ann. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(12), 14546-14562 (2023)", + "27. Hu, Y., Deng, L., Wu, Y., Yao, M., Li, G.: Advancing spiking neural networks toward deep residual learning. IEEE Transactions on Neural Networks and Learning Systems pp. 1-15 (2024)", + "28. Kim, S., Park, S., Na, B., Kim, J., Yoon, S.: Towards fast and accurate object detection in bio-inspired spiking neural networks through bayesian optimization. IEEE Access 9, 2633-2643 (2020)" + ], + "bbox": [ + 217, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Luo et al.", + "bbox": [ + 271, + 114, + 336, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Kim, S., Park, S., Na, B., Yoon, S.: Spiking-yolo: spiking neural network for energy-efficient object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11270-11277 (2020)", + "30. Kim, Y., Park, H., Moitra, A., Bhattacharjee, A., Venkatesha, Y., Panda, P.: Rate coding or direct coding: Which one is better for accurate, robust, and energy-efficient spiking neural networks? In: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 71-75. IEEE (2022)", + "31. Li, C., Jones, E.G., Furber, S.: Unleashing the potential of spiking neural networks with dynamic confidence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13304-13314 (2023)", + "32. Li, Y., He, X., Dong, Y., Kong, Q., Zeng, Y.: Spike calibration: Fast and accurate conversion of spiking neural network for object detection and segmentation. arXiv preprint arXiv:2207.02702 (2022)", + "33. Li, Y., Geller, T., Kim, Y., Panda, P.: Seenn: Towards temporal spiking early exit neural networks. Advances in Neural Information Processing Systems 36 (2024)", + "34. Lin, T.Y., Dólár, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2117-2125 (2017)", + "35. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "36. Maass, W.: Networks of spiking neurons: the third generation of neural network models. Neural Networks 10(9), 1659-1671 (1997)", + "37. Merolla, P.A., Arthur, J.V., Alvarez-Icaza, R., Cassidy, A.S., Sawada, J., Akopyan, F., Jackson, B.L., Imam, N., Guo, C., Nakamura, Y., et al.: A million spiking-neuron integrated circuit with a scalable communication network and interface. Science 345(6197), 668-673 (2014)", + "38. Mueller, E., Studenyak, V., Auge, D., Knoll, A.: Spiking transformer networks: A rate coded approach for processing sequential data. In: 2021 7th International Conference on Systems and Informatics (ICSAI). pp. 1-5. IEEE (2021)", + "39. Neftci, E.O., Mostafa, H., Zenke, F.: Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine 36(6), 51-63 (2019)", + "40. Panda, P., Aketi, S.A., Roy, K.: Toward scalable, efficient, and accurate deep spiking neural networks with backward residual connections, stochastic softmax, and hybridization. Frontiers in Neuroscience 14, 653 (2020)", + "41. Pei, J., Deng, L., Song, S., Zhao, M., Zhang, Y., Wu, S., Wang, G., Zou, Z., Wu, Z., He, W., et al.: Towards artificial general intelligence with hybrid tianjic chip architecture. Nature 572(7767), 106-111 (2019)", + "42. Qiu, X., Zhu, R.J., Chou, Y., Wang, Z., Deng, L.j., Li, G.: Gated attention coding for training high-performance and efficient spiking neural networks. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 601-610 (2024)", + "43. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 779-788 (2016)", + "44. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018)", + "45. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems 28 (2015)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SpikeYOLO", + "bbox": [ + 648, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "46. Roy, K., Jaiswal, A., Panda, P.: Towards spike-based machine intelligence with neuromorphic computing. Nature 575(7784), 607-617 (2019)", + "47. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: Mobilenetv2: Inverted residuals and linear bottlenecks. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 4510-4520 (2018)", + "48. Schuman, C.D., Kulkarni, S.R., Parsa, M., Mitchell, J.P., Kay, B., et al.: Opportunities for neuromorphic computing algorithms and applications. Nature Computational Science 2(1), 10-19 (2022)", + "49. Sengupta, A., Ye, Y., Wang, R., Liu, C., Roy, K.: Going deeper in spiking neural networks: Vgg and residual architectures. Frontiers in Neuroscience 13, 95 (2019)", + "50. Su, Q., Chou, Y., Hu, Y., Li, J., Mei, S., Zhang, Z., Li, G.: Deep directly-trained spiking neural networks for object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6555-6565 (2023)", + "51. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems. pp. 5998-6008 (2017)", + "52. Wang, C.Y., Bochkovskiy, A., Liao, H.Y.M.: Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7464-7475 (2023)", + "53. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 568-578 (2021)", + "54. Wang, Z., Fang, Y., Cao, J., Zhang, Q., Wang, Z., Xu, R.: Masked spiking transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1761-1771 (2023)", + "55. Wu, Y., Deng, L., Li, G., Zhu, J., Shi, L.: Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in Neuroscience 12 331 (2018)", + "56. Wu, Y., Deng, L., Li, G., Zhu, J., Xie, Y., Shi, L.: Direct training for spiking neural networks: Faster, larger, better. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 1311-1318 (2019)", + "57. Yao, M., Gao, H., Zhao, G., Wang, D., Lin, Y., Yang, Z., Li, G.: Temporal-wise attention spiking neural networks for event streams classification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10221-10230 (2021)", + "58. Yao, M., Hu, J., Hu, T., Xu, Y., Zhou, Z., Tian, Y., XU, B., Li, G.: Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=1SIBN5Xyw7", + "59. Yao, M., Hu, J., Zhao, G., Wang, Y., Zhang, Z., Xu, B., Li, G.: Inherent redundancy in spiking neural networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16924-16934 (2023)", + "60. Yao, M., Hu, J., Zhou, Z., Yuan, L., Tian, Y., Xu, B., Li, G.: Spike-driven transformer. Advances in neural information processing systems 36 (2024)", + "61. Yao, M., Richter, O., Zhao, G., Qiao, N., Xing, Y., Wang, D., Hu, T., Fang, W., Demirci, T., De Marchi, M., Deng, L., Yan, T., Nielsen, C., Sheik, S., Wu, C., Tian, Y., Xu, B., Li, G.: Spike-based dynamic computing with asynchronous sensing-computing neuromorphic chip. Nature Communications 15(1), 4464 (May 2024)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Luo et al.", + "bbox": [ + 271, + 114, + 336, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "62. Yao, M., Zhang, H., Zhao, G., Zhang, X., Wang, D., Cao, G., Li, G.: Sparser spiking activity can be better: Feature refine-and-mask spiking neural network for event-based visual recognition. Neural Networks 166, 410-423 (2023)", + "63. Yao, M., Zhao, G., Zhang, H., Hu, Y., Deng, L., Tian, Y., Xu, B., Li, G.: Attention spiking neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(8), 9393-9410 (2023)", + "64. Yin, B., Corradi, F., Bohte, S.M.: Accurate and efficient time-domain classification with adaptive spiking recurrent neural networks. Nature Machine Intelligence 3(10), 905-913 (2021)", + "65. Yuan, M., Zhang, C., Wang, Z., Liu, H., Pan, G., Tang, H.: Trainable spiking-yolo for low-latency and high-performance object detection. Neural Networks 172, 106092 (2024)", + "66. Zhang, J., Dong, B., Zhang, H., Ding, J., Heide, F., Yin, B., Yang, X.: Spiking transformers for event-based single object tracking. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 8801-8810 (2022)", + "67. Zhang, J., Tang, L., Yu, Z., Lu, J., Huang, T.: Spike transformer: Monocular depth estimation for spiking camera. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII. pp. 34-52. Springer (2022)", + "68. Zheng, H., Wu, Y., Deng, L., Hu, Y., Li, G.: Going deeper with directly-trained larger spiking neural networks. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 11062-11070 (2021)", + "69. Zhou, Z., Zhu, Y., He, C., Wang, Y., YAN, S., Tian, Y., Yuan, L.: Spikformer: When spiking neural network meets transformer. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=frE4fUwz_h", + "70. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable $\\{\\mathrm{detr}\\}$ : Deformable transformers for end-to-end object detection. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=gZ9hCDWe6ke" + ], + "bbox": [ + 212, + 146, + 787, + 550 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "SpikeYOLO", + "bbox": [ + 648, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Integer-Valued Training and Spike-driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection/b44a5b42-9423-4159-817b-8bb17aae60da_model.json b/2024/Integer-Valued Training and Spike-driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection/b44a5b42-9423-4159-817b-8bb17aae60da_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4fca90c04441e597926754562ef916a364978af1 --- /dev/null +++ b/2024/Integer-Valued Training and Spike-driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection/b44a5b42-9423-4159-817b-8bb17aae60da_model.json @@ -0,0 +1,2614 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.239, + 0.141, + 0.765, + 0.23 + ], + "angle": 0, + "content": "Integer-Valued Training and Spike-Driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.257, + 0.782, + 0.274 + ], + "angle": 0, + "content": "Xinhao Luo\\(^{1*}\\), Man Yao\\(^{1*}\\), Yuhong Chou\\(^{2}\\), Bo Xu\\(^{1}\\), and Guoqi Li\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.311, + 0.285, + 0.692, + 0.327 + ], + "angle": 0, + "content": "\\(^{1}\\) Institute of Automation, Chinese Academy of Sciences {luoxinhao2023, man.yao, guoqi.li}@ia.ac.cn \\(^{2}\\) Xi'an Jiaotong University" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.357, + 0.744, + 0.663 + ], + "angle": 0, + "content": "Abstract. Brain-inspired Spiking Neural Networks (SNNs) have bioplausibility and low-power advantages over Artificial Neural Networks (ANNs). Applications of SNNs are currently limited to simple classification tasks because of their poor performance. In this work, we focus on bridging the performance gap between ANNs and SNNs on object detection. Our design revolves around network architecture and spiking neuron. First, the overly complex module design causes spike degradation when the YOLO series is converted to the corresponding spiking version. We design a SpikeYOLO architecture to solve this problem by simplifying the vanilla YOLO and incorporating meta SNN blocks. Second, object detection is more sensitive to quantization errors in the conversion of membrane potentials into binary spikes by spiking neurons. To address this challenge, we design a new spiking neuron that activates Integer values during training while maintaining spike-driven by extending virtual timesteps during inference. The proposed method is validated on both static and neuromorphic object detection datasets. On the static COCO dataset, we obtain \\(66.2\\%\\) mAP@50 and \\(48.9\\%\\) mAP@50:95, which is \\(+15.0\\%\\) and \\(+18.7\\%\\) higher than the prior state-of-the-art SNN, respectively. On the neuromorphic Gen1 dataset, we achieve \\(67.2\\%\\) mAP@50, which is \\(+2.5\\%\\) greater than the ANN with equivalent architecture, and the energy efficiency is improved by \\(5.7\\times\\). Code: https://github.com/BICLab/SpikeYOLO" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.675, + 0.741, + 0.703 + ], + "angle": 0, + "content": "Keywords: Spiking neural network \\(\\cdot\\) Object detection \\(\\cdot\\) Spike-driven \\(\\cdot\\) Neuromorphic vision \\(\\cdot\\) Neuromorphic computing" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.746, + 0.376, + 0.761 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.774, + 0.788, + 0.806 + ], + "angle": 0, + "content": "Brain-inspired SNNs are known for their low power consumption [46, 48]. Spiking neurons incorporate spatio-temporal information and emit spikes when the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.239, + 0.812, + 0.429, + 0.841 + ], + "angle": 0, + "content": "\\*Equal contribution. \\((\\boxtimes)\\) Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.338, + 0.127 + ], + "angle": 0, + "content": "Luo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "membrane potentials exceed a threshold [36]. Thus, spiking neurons trigger sparse additions only when they receive a spike and are otherwise idle. This spike-driven enables SNNs to exhibit obvious low-power advantages over ANNs when deployed on neuromorphic chips [7,37,41,61]. However, the negative impact of complex neuronal dynamics and spike-driven nature is that SNNs are difficult to train and have limited task performance and application scenarios [13]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.79, + 0.418 + ], + "angle": 0, + "content": "For example, most applications of SNN algorithms in computer vision are limited to simple image classification tasks [9, 14, 20, 25, 27, 31, 33, 42, 54, 60]. Another more commonly used and challenging computer vision task, object detection, is rarely explored in SNNs. In 2020, Spiking-YOLO [29] provided the first object detection model in deep SNNs, exploiting the method of converting ANN to SNN with thousands of timesteps. In 2023, EMS-YOLO [50] became the first work to use direct training SNNs to handle object detection. Recently, the direct training Meta-SpikeFormer [58] can process the object detection in a pre-training and fine-tuning manner for the first time. However, the performance gap between these works and ANNs is significant. In this work, we aim to bridge this gap and demonstrate the low-power of SNNs and their unique advantages in neuromorphic applications. We achieve this goal through two efforts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.79, + 0.614 + ], + "angle": 0, + "content": "First, we design a new architecture, SpikeYOLO, which combines the macro design of YOLO with the micro design of the SNN module. Simply replacing the artificial neurons in the YOLO series [1,43,52] with spiking neurons generally does not work. Existing solutions include establishing the equivalence between ANN activation and spike firing rate [29], or improving the residual design [50]. We argue that another potential reason is that the module design in the YOLO series is too complex, which is effective in ANNs, but not suitable for SNNs. We observe that after the complex YOLO modules are converted into the corresponding spiking versions, there is a phenomenon of spike degradation in which the deep layers almost no longer emit spikes. Therefore, we tend to simplify the design in SpikeYOLO. We only retain the macro architecture in YOLOv83 while using the simple meta SNN block in Meta-SpikeFormer [58] as the basic module and performing micro design." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.79, + 0.812 + ], + "angle": 0, + "content": "Second, we design a novel spiking neuron, Integer Leaky Integrate-and-Fire (I-LIF), to reduce the quantization error of SNNs. Spike-driven is the key to low-power, while exploiting only binary signals will drop performance, especially in challenging object detection. Numerous studies have attempted to mitigate this problem, such as attention-based membrane optimization [63], ternary spike [18], information max [19], converting quantized ANNs [26], optimal ANN2SNN [2]. However, direct training based on optimization strategies can only alleviate errors; the optimal approximation of ANN2SNN requires large timesteps, and it is difficult to exploit the temporal information. In contrast, the idea of the proposed I-LIF is to use integer-valued activations to drop quantization errors and convert them into spikes during inference by extending the virtual timesteps. The features of I-LIF are: 1) Integer-valued training improves performance and is easy to train; 2) Integer activation during training can be equivalent to spike-driven" + }, + { + "type": "page_footnote", + "bbox": [ + 0.238, + 0.825, + 0.526, + 0.84 + ], + "angle": 0, + "content": "3https://github.com/ultralytics/ultralytics" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.649, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SpikeYOLO" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "in inference that there is only sparse addition; 3) The temporal dynamics of LIF are reserved, capable of processing neuromorphic object detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.252 + ], + "angle": 0, + "content": "The proposed method is validated on both static COCO [35] and neuromorphic Gen1 [8] object detection datasets. The performance we report on both datasets far exceeds the existing best models in SNNs. In addition, we explore the performance changes after the mutual conversion of ANN and SNN with the same architecture. The main contributions of this work are:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.262, + 0.784, + 0.321 + ], + "angle": 0, + "content": "- SpikeYOLO. We explore suitable architectures in SNNs for handling object detection tasks and propose SpikeYOLO, which simplifies YOLOv8 and incorporates meta SNN blocks. This inspires us that the complex modules in ANN may not be suitable for SNN architecture design." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.323, + 0.785, + 0.38 + ], + "angle": 0, + "content": "- I-LIF Spiking Neuron. We propose an I-LIF spiking neuron that combines integer-valued training with spike-driven inference. The former is used to reduce quantization errors in spiking neurons, and the latter is the basis of the low-power nature of SNNs." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.384, + 0.785, + 0.47 + ], + "angle": 0, + "content": "- Performance. The proposed method achieves outstanding accuracy with low power consumption on object detection datasets, demonstrating the potential of SNNs in complex vision tasks. On the COCO dataset, we obtain \\(66.2\\%\\) mAP@50 and \\(48.9\\%\\) mAP@50:95, which is \\(+15.0\\%\\) and \\(+18.7\\%\\) higher than the prior state-of-the-art SNN, respectively. On the Gen1 dataset, SpikeYOLO is \\(+2.5\\%\\) better than ANN models with \\(5.7\\times\\) energy efficiency" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.262, + 0.785, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.507, + 0.396, + 0.523 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.787, + 0.719 + ], + "angle": 0, + "content": "SNN Training Method. Training methods have restricted the development of SNN for a long time. To make SNNs deeper, two training methods have been developed. ANN2SNN substitutes the ReLU function with spiking neurons, aiming to mimic the continuous activation by controlling the firing rate of spiking neurons [3, 10, 49]. It can achieve high performance but often requires long timesteps and is difficult to process sequence tasks that exploit the spatiotemporal dynamic nature of SNNs. In contrast, another directly training an SNN leverage gradient surrogate to circumvent the non-differentiability of binary spikes [39, 55]. Direct training is more flexible and requires fewer timesteps, but its performance usually suffers compared to ANNs of the same architecture. In this work, we focus on utilizing directly trained SNNs to process object detection due to its more flexibility in architectural design." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "SNN Architecture Design. The architecture of SNNs can be roughly divided into two categories: CNN-based and Transformer-based SNNs. Spiking ResNet has long dominated the SNN field because residual learning [22] can address the performance degradation of SNNs as they become deeper. Typical spiking ResNet includes vanilla spiking ResNet [68], SEW-ResNet [14], and MS-ResNet [27]. The main difference between them is the location of shortcuts and the ability to achieve identity mapping [23]. Recently, the Transformer [12, 51] architecture has become popular [21, 38, 58, 60, 66, 67, 67, 69] in the SNN field and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.339, + 0.127 + ], + "angle": 0, + "content": "Luo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.191 + ], + "angle": 0, + "content": "has refreshed the performance upper bound. The SpikeYOLO we designed draws on the idea of the meta SNN block in Meta-SpikeFormer [58] and merges it with the YOLOv8 architecture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.788, + 0.373 + ], + "angle": 0, + "content": "Quantization Errors in SNNs. Spiking neurons quantize continuous membrane potentials with complex spatio-temporal dynamics into binary spikes. Obviously, quantization error will limit the performance of SNNs. In ANN2SNN, the reduction of quantization error is relatively simple, which can be achieved by increasing the timestep [2]. In direct training, adjusting the relationship between membrane potential distribution and threshold is the main solution, such as attention mechanism [63] optimizes membrane potential distribution to reduce noise spikes; information max [19] is derived from the information entropy angle to optimize spike firing. However, optimization cannot change the inherently quantized error nature of binary spikes. The solution of this work is to use integer values to reduce quantization errors during training and convert them to binary spikes to ensure spike-driven inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.525 + ], + "angle": 0, + "content": "Object Detection. Object detection is a challenging yet pivotal task. Existing object detection frameworks can be simply categorized into: two-stage frameworks (RCNN series [16, 17, 45]) and single-stage frameworks (YOLO [1, 43, 52], Detr [4, 70] series). The former generate region proposals before determining precise object locations and classifications, whereas the latter directly ascertain these elements, offering a swifter, more streamlined solution. The object detection task has always been difficult for SNNs, and there are few related works. Current works include ANN2SNN-based YOLO [29, 65] and directly-trained spiking YOLO [50, 58]. Their performance is poor, and it isn't easy to meet the needs of real scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.548, + 0.339, + 0.563 + ], + "angle": 0, + "content": "3 Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.787, + 0.626 + ], + "angle": 0, + "content": "We exploit SpikeYOLO to process both static and neuromorphic object detection datasets. We first introduce how network inputs are unified. Then, the details of SpikeYOLO architecture and I-LIF spiking neuron are presented, respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.648, + 0.387, + 0.664 + ], + "angle": 0, + "content": "3.1 Network Input" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.787, + 0.703 + ], + "angle": 0, + "content": "The input of SNNs can be denoted as \\( X \\in \\mathbb{R}^{T \\times C \\times H \\times W} \\), where \\( T \\) is the timestep, \\( C \\) is the channel, \\( H \\times W \\) denote the spatial resolution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.78 + ], + "angle": 0, + "content": "Static Image. To leverage the spatio-temporal capabilities of SNNs, it is common practice that static images are repeated and utilized as input for each timestep \\( T \\). This is called direct input encoding [30, 56], where the first layer of spiking neurons in the network encodes the continuous values of the input into spike signals." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Neuromorphic Event Stream. Neuromorphic data (also known as event-based vision) are generated by a Dynamic Vision Sensor (DVS), which only generates spikes when the logarithmic change in light intensity at a pixel surpasses a predefined threshold. An event-based stream is characterized as \\((x_{n},y_{n},t_{n},p_{n})\\)," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.649, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SpikeYOLO" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.145, + 0.744, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.42, + 0.788, + 0.517 + ], + "angle": 0, + "content": "Fig. 1: The overall architecture of SpikeYOLO. We designed two SNN blocks, SNN-Block-1 and SNN-Block-2, and kept other architectures remain as YOLOv8. SNN-Block-1 employs standard convolution within its ChannelConv (\\(\\cdot\\)) component, whereas SNN-Block-2 utilizes re-parameterization convolution. That is, the difference between the two is the channel mixer module. In the low and high stages, we use SNN-Block-1 and SNN-Block-2, respectively. The spiking neuron is I-LIF, which activates integer values during training while converting them to binary spikes during inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.684 + ], + "angle": 0, + "content": "with each event capturing spatial coordinates \\((x,y)\\), timestamp \\(t\\) and polarity \\(p\\), where \\(p \\in \\{-1,1\\}\\) indicates whether light intensity has increased or decreased. Neuromorphic vision offers several advantages [15, 61], such as low resource requirements, high temporal resolution, and high robustness. The spike-driven nature of SNN makes it naturally suitable for processing event streams. The general strategy for neuromorphic pre-processing is to aggregate the event stream within a fixed time window into a frame format [57, 59, 62]. In this work, we follow this operation. Specifically, the total input window length is \\(T \\times dt\\), where \\(dt\\) and \\(T\\) are temporal resolution and timestep, respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.708, + 0.471, + 0.723 + ], + "angle": 0, + "content": "3.2 SpikeYOLO Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Overview. SpikeYOLO integrates the macro design of the YOLOv8 with the micro design of Meta-SpikeFormer [58]. The motivation is that we observe complex computations within YOLO's modules result in spike degradation [27] upon direct conversion to the SNN version. Consequently, we maintain the overarching design principles of the YOLO architecture while incorporating the inverted residual structure [47] and re-parameterization convolution [11] design of the Meta-SpikeFormer for detailed aspects." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.339, + 0.128 + ], + "angle": 0, + "content": "Luo et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.79, + 0.194 + ], + "angle": 0, + "content": "Network Output. In object detection, the network outputs the class and position of each object based on the input image sequence \\( X = \\{X_{t}\\}_{t = 1}^{T} \\). Suppose the input image sequence has \\( N \\) goals, the output \\( B = \\{B_{n}\\}_{n = 1}^{N} \\) can be calculated:" + }, + { + "type": "equation", + "bbox": [ + 0.44, + 0.206, + 0.788, + 0.224 + ], + "angle": 0, + "content": "\\[\nB = \\operatorname {M o d e l} (X), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.228, + 0.79, + 0.275 + ], + "angle": 0, + "content": "where each \\( B_{n} = \\{f_{n},c_{n},x_{n},y_{n},w_{n},h_{n}\\} \\) contains information about degree of confidence \\( f_{n} \\), class \\( C_n \\), center coordinates \\( (x_{n},y_{n}) \\) and target size \\( (w_{n},h_{n}) \\). Model \\( (\\cdot) \\) refers to the proposed SpikeYOLO architecture." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.294, + 0.79, + 0.462 + ], + "angle": 0, + "content": "Macro Design. Fig. 1 shows the overview of SpikeYOLO, a variation of the YOLO framework that is more suitable for the feature extraction scheme of SNNs. Specifically, YOLOv8 is a classic single-stage detection framework that partitions the image into numerous grids, with each grid responsible for predicting a target independently. Some classic designs, such as the feature pyramid network [34] in YOLOv8, play a crucial role in facilitating efficient feature extraction and fusion. By contrast, its feature extraction module, such as C2F, performs repeated feature extraction from the same set of feature maps. This module can enhance feature extraction in ANNs but does not work well in SNNs. As a compromise, we preserve the classic Backbone/Neck/Head architecture in YOLOv8 while incorporating strategies from the meta SNN block in Meta-SpikeFormer." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.481, + 0.789, + 0.617 + ], + "angle": 0, + "content": "Micro Design. Meta-SpikeFormer [58] is the current state-of-the-art architecture in SNNs, which explores the meta design of SNN and consists of CNN-based and Transformer-based SNN blocks. The meta block comprises a token mixer module and a channel mixer module. The difference between CNN-based and Transformer-based SNN blocks lies in the token mixer, which are spike-driven convolution and spike-driven self-attention, respectively. In this work, we mainly redesign the channel mixer module for the object detection task. As shown in Fig. 1, SNN-Block-1 and SNN-Block-2 are designed to extract low-stage and high-stage features, respectively." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.618, + 0.581, + 0.632 + ], + "angle": 0, + "content": "The meta SNN block in [58] can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.642, + 0.788, + 0.66 + ], + "angle": 0, + "content": "\\[\nU ^ {\\prime} = U + \\operatorname {S e p C o n v} (U), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.39, + 0.682, + 0.788, + 0.7 + ], + "angle": 0, + "content": "\\[\nU ^ {\\prime \\prime} = U ^ {\\prime} + \\operatorname {C h a n n e l C o n v} \\left(U ^ {\\prime}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.704, + 0.789, + 0.766 + ], + "angle": 0, + "content": "where \\(U \\in \\mathbb{R}^{T \\times C \\times H \\times W}\\) is the layer input, \\(\\operatorname{SepConv}(\\cdot)\\) is an inverted separable convolution module [47] with \\(7 \\times 7\\) kernel size in MobileNetv2 to capture global features, followed by a \\(3 \\times 3\\) depthwise convolution for further spatial feature fusion. Sepconv (\\(\\cdot\\)) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.787, + 0.788, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S e p C o n v} (U) = \\operatorname {C o n v} _ {\\mathrm {d w} 2} \\left(\\operatorname {C o n v} _ {\\mathrm {p w} 2} \\left(S N \\left(\\operatorname {C o n v} _ {\\mathrm {d w} 1} \\left(S N \\left(\\operatorname {C o n v} _ {\\mathrm {p w} 1} \\left(S N (U)\\right)\\right)\\right)\\right)\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "where \\(\\mathrm{Conv}_{\\mathrm{dw1}}(\\cdot)\\) and \\(\\mathrm{Conv}_{\\mathrm{dw1}}(\\cdot)\\) are depthwise convolutions, \\(\\mathrm{Conv}_{\\mathrm{pw1}}(\\cdot)\\) and \\(\\mathrm{Conv}_{\\mathrm{pw1}}(\\cdot)\\) are pointwise convolutions [5]. \\(SN(\\cdot)\\) is the spiking neuron layer." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.648, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SpikeYOLO" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "ChannelConv \\((\\cdot)\\) is the channel mixer, which facilitates inter-channel information fusion. We redesign ChannelConv \\((\\cdot)\\) in object detection task." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.178, + 0.486, + 0.192 + ], + "angle": 0, + "content": "For SNN-Block-1, it is written as:" + }, + { + "type": "equation", + "bbox": [ + 0.313, + 0.2, + 0.788, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C h a n n e l C o n v 1} \\left(U ^ {\\prime}\\right) = \\operatorname {C o n v} \\left(S N \\left(\\operatorname {C o n v} \\left(S N \\left(U ^ {\\prime}\\right)\\right)\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.225, + 0.784, + 0.27 + ], + "angle": 0, + "content": "where \\(\\mathrm{Conv}(\\cdot)\\) is a standard convolution with expansion ratio \\(r = 4\\). In contrast, addressing high-stage features, SNN-Block-2 employs a re-parameterization convolution to minimize parameter count, which can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.284, + 0.284, + 0.788, + 0.301 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C h a n n e l C o n v 2} \\left(U ^ {\\prime}\\right) = \\operatorname {R e p C o n v} \\left(S N \\left(\\operatorname {R e p C o n v} \\left(S N \\left(U ^ {\\prime}\\right)\\right)\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.311, + 0.32, + 0.787, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\operatorname {R e p C o n v} \\left(U ^ {\\prime}\\right) = \\operatorname {C o n v} _ {\\mathrm {p w} 2} \\left(\\operatorname {C o n v} _ {\\mathrm {d w} 1} \\left(\\operatorname {C o n v} _ {\\mathrm {p w} 1} \\left(U ^ {\\prime}\\right)\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.784, + 0.371 + ], + "angle": 0, + "content": "where \\(\\mathrm{RepConv}(\\cdot)\\) is the re-parameterization convolution [11] with kernel size \\(3\\times 3\\), it can be re-parameterized to a standard convolution during inference." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.39, + 0.442, + 0.405 + ], + "angle": 0, + "content": "3.3 I-LIF Spiking Neuron" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.784, + 0.549 + ], + "angle": 0, + "content": "Spiking neurons propagate information in both spatial and temporal domains, and they mimic the spiking communication scheme of biological neurons. However, there are inherent quantization errors in converting the membrane potential of spiking neurons into binary spikes, which severely limits the representation of the model. Recently, Fast-SNN [26] achieves high-performance conversion with small timesteps by converting quantized ANNs into SNNs. This inspires us to \"why not train directly with integer values\", which can significantly reduce the quantization error. We just need to be careful to ensure that the inference is spike-driven. So, we came up with the idea of I-LIF." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.568, + 0.784, + 0.613 + ], + "angle": 0, + "content": "LIF. Leaky Integrate-and-Fire (LIF) spiking neuron [36] is the most popular neuron to construct SNNs due to its balance between bio-plausibility and computing complexity. The dynamics of LIF with soft reset is:" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.628, + 0.787, + 0.646 + ], + "angle": 0, + "content": "\\[\nU [ t ] = H [ t - 1 ] + X [ t ], \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.663, + 0.787, + 0.681 + ], + "angle": 0, + "content": "\\[\nS [ t ] = \\Theta (U [ t ] - V _ {t h}), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.699, + 0.787, + 0.716 + ], + "angle": 0, + "content": "\\[\nH [ t ] = \\beta (U [ t ] - S [ t ]) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.84 + ], + "angle": 0, + "content": "where \\( t \\) denotes the timestep, \\( U[t] \\) is the membrane potential that integrates the temporal information \\( H[t - 1] \\) and spatial information \\( X[t] \\). \\( \\Theta(\\cdot) \\) is the Heaviside step function which equals 1 for \\( x \\geq 0 \\) and 0 otherwise. If \\( U[t] \\) exceeds the firing threshold \\( V_{th} \\), spiking neuron fire a spike \\( S[t] \\) and \\( U[t] \\) will subtract it subsequently. Otherwise, \\( H[t] \\) will remain unchanged. And, \\( U[t] \\) decays to \\( H[t] \\) by a factor of \\( \\beta \\), which denotes the decay constant. For simplicity, we focus on Eq.9 and denote the spiking neuron layer as \\( SN(\\cdot) \\), with its input as membrane potential tensor \\( U \\) and its output as spike tensor \\( S \\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.34, + 0.127 + ], + "angle": 0, + "content": "Luo et al." + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.145, + 0.739, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.371, + 0.789, + 0.429 + ], + "angle": 0, + "content": "Fig. 2: Comparison of I-LIF and LIF. Binary spikes are emitted by LIF during both training and inference processes, which results in quantization errors. I-LIF emits integer values during the training process to reduce quantization errors, and converts them into binary spikes during inference to make the network only perform sparse addition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.789, + 0.515 + ], + "angle": 0, + "content": "I-LIF. We propose the Integer Leaky Integrate-and-Fire (I-LIF) neuron to reduce the quantization error. As shown in Fig. 2, I-LIF emits integer values while training, and converts them into \\(0/1\\) spikes when inference. Specifically, in I-LIF, Eq.9 is rewritten as:" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.524, + 0.788, + 0.542 + ], + "angle": 0, + "content": "\\[\nS [ t ] = \\operatorname {C l i p} (r o u n d (U [ t ]), 0, D), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.789, + 0.594 + ], + "angle": 0, + "content": "where \\( \\text{round}(\\cdot) \\) is a round symbol, \\( \\text{Clip}(x, \\min, \\max) \\) denotes that clipping \\( x \\) to \\( [\\min, \\max] \\), \\( D \\) is a hyperparameter indicating the maximum emitted integer value by I-LIF." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.789, + 0.685 + ], + "angle": 0, + "content": "Training Stage. Eq.11 is not a continuous function, making its derivative a step function, potentially causing training instability. Previous studies have introduced several surrogate gradient functions, which primarily address binary spike outputs. We consistently utilize rectangular windows as the surrogate function. For simplicity, We retain gradients solely for neurons activated in the \\([0,D]\\) range, nullifying all others." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.685, + 0.79, + 0.804 + ], + "angle": 0, + "content": "Inference Stage. Introducing integer value necessitates additional MACs (Multiply-Accumulation operations), potentially diminishing the energy efficiency of SNNs. Thus, converting integer values to binary spikes is essential. Fig. 3 shows an example of how integer values convert to binary spikes by extending virtual timesteps during inference. Specifically, the input to the spiking neuron at \\( l + 1 \\) layer can be described as \\( X^{l + 1}[t] = W^l S^l [t] \\). We extend the \\( T \\) time step to \\( T\\times D \\), and convert the integer value \\( S^l [t] \\) to a spike sequence \\( \\{S^l [t,d]\\}_{d = 1}^D \\), which satisfied:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.803, + 0.788, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\sum_ {d = 1} ^ {D} S ^ {l} [ t, d ] = S ^ {l} [ t ]. \\tag {12}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.649, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SpikeYOLO" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.142, + 0.744, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.319, + 0.788, + 0.417 + ], + "angle": 0, + "content": "Fig. 3: An example of how the proposed I-LIF works. We assume \\( T = 3, D = 2 \\), and show the corresponding binary spike sequences of integer value during inference. The membrane potential in [0.5, 1.5) are quantized to 1, while those in [1.5, 2.5) are quantized to 2. membrane potential that \\( > 2.5 \\) are also quantized to 2 due to the maximum integer value \\( D = 2 \\). Subsequently, the membrane potential will be subtracted from the integer value. The training spike with a value of 2 will be converted into two binary spikes by extending virtual timesteps during inference." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.444, + 0.633, + 0.459 + ], + "angle": 0, + "content": "Thus, the neuron's input at \\( l + 1 \\) layer is reformulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.415, + 0.467, + 0.788, + 0.508 + ], + "angle": 0, + "content": "\\[\nX ^ {l} [ t ] = W ^ {l} \\sum_ {d = 1} ^ {D} S ^ {l} [ t, d ]. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.516, + 0.763, + 0.532 + ], + "angle": 0, + "content": "Given that matrix multiplication functions as linear operators, we establish:" + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.538, + 0.788, + 0.58 + ], + "angle": 0, + "content": "\\[\nW ^ {l} \\sum_ {d = 1} ^ {D} S ^ {l} [ t, d ] = \\sum_ {d = 1} ^ {D} \\left(W ^ {l} S ^ {l} [ t, d ]\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.587, + 0.713, + 0.603 + ], + "angle": 0, + "content": "Therefore, the input of the neuron at \\( l + 1 \\) layer can be computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.412, + 0.61, + 0.787, + 0.651 + ], + "angle": 0, + "content": "\\[\nX ^ {l} [ t ] = \\sum_ {d = 1} ^ {D} \\left(W ^ {l} S ^ {l} [ t, d ]\\right). \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.66, + 0.788, + 0.706 + ], + "angle": 0, + "content": "The spike sequence \\( S^l[t,d] \\) only contains \\( 0/1 \\), so all MACs can be converted into sparse ACs(Accumulation operations), which can ensure spike-driven when inference." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.727, + 0.376, + 0.744 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.755, + 0.788, + 0.817 + ], + "angle": 0, + "content": "We evaluate the proposed method on COCO 2017 val [35] and neuromorphic Gen1 [8] datasets. The mean Average Precision(mAP) at IOU=0.5(mAP@50), the average mAP between 0.5 and 0.95(mAP@50:95), and energy cost are reported for each model. Specifically, the power of ANNs and SNNs can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.824, + 0.787, + 0.842 + ], + "angle": 0, + "content": "\\[\nE _ {A N N} = O ^ {2} \\times C _ {i n} \\times C _ {o u t} \\times k ^ {2} \\times E _ {M A C}, \\tag {16}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.127 + ], + "angle": 0, + "content": "Luo et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 1: Results on COCO 2017 val [35]. \\( T \\times D \\) means that we set up \\( T \\) timesteps, and each timestep is expanded \\( D \\) times. In prior SNNs and ANNs, \\( D \\) defaults to 1." + }, + { + "type": "table", + "bbox": [ + 0.239, + 0.185, + 0.767, + 0.48 + ], + "angle": 0, + "content": "
ArchitectureModelParam (M)Power (mJ)T × DmAP@ 50(%)mAP@ 50:95(%)
ANNPVT [53]32.9520.3159.236.7
DETR [4]41.0197.8162.442.0
YOLOv5 421.2112.5164.145.4
ANN2SNNSpiking-Yolo [29]10.2-3500-25.7
Bayesian Optim [28]10.2-5000-25.9
Spike Calib [32]17.1-51245.4-
Directly-trained SNNEMS-YOLO [50]26.929.0450.130.1
Meta-SpikeFormer (MaskRCNN) [58]34.949.5144.0-
75.0140.8151.2-
Meta-SpikeFormer (YOLO) [58]16.834.8145.0-
16.870.7450.3-
SpikeYOLO(Ours)13.223.11 × 459.242.5
23.118.41 × 152.736.1
23.134.61 × 462.345.5
23.167.64 × 155.738.7
23.1134.74 × 463.346.3
48.168.51 × 464.647.4
68.884.21 × 466.248.9
" + }, + { + "type": "equation", + "bbox": [ + 0.302, + 0.522, + 0.788, + 0.54 + ], + "angle": 0, + "content": "\\[\nE _ {S N N} = (T \\times D) \\times f r \\times O ^ {2} \\times C _ {i n} \\times C _ {o u t} \\times k ^ {2} \\times E _ {A C}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.789, + 0.668 + ], + "angle": 0, + "content": "where \\( O \\) is the feature output size, \\( C_{in} \\) and \\( C_{out} \\) denotes the number of input channel and output channel, \\( k \\) is the kernel size, \\( fr \\) denotes the average spike firing rate, \\( T \\) is the timestep, \\( D \\) is the upper limit of integer activation during training. We follow the most commonly used energy consumption evaluation method in the SNN field [40,63,64]. All operations assume a 32-bit floating-point implementation on \\( 45\\mathrm{nm} \\) technology, where \\( E_{MAC} = 4.6pJ \\) and \\( E_{AC} = 0.9pJ \\) [24]. As can be seen from Eq. 16 and 17, SNN's low power comes from its sparse addition operation. The fewer spikes, the sparser the computation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.689, + 0.462, + 0.703 + ], + "angle": 0, + "content": "4.1 COCO 2017 val Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.714, + 0.789, + 0.807 + ], + "angle": 0, + "content": "Experimental Setup. As a predominant static dataset for object detection, COCO 2017 val [35] comprises 80 classes split into 118K training and 5K validating images. In all experiments, we set decay factor \\(\\beta = 0.25\\), learning rate to 0.01, and adopt SGD optimizer. The models are trained for 300 epochs with a batch size of 40 on 4 NVIDIA V100 GPUs. Mosaic data augmentation [1] technique is employed. The network structure is given in the supplementary material. Note," + }, + { + "type": "page_footnote", + "bbox": [ + 0.238, + 0.825, + 0.501, + 0.841 + ], + "angle": 0, + "content": "4https://github.com/ultralytics/yolov5" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.648, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SpikeYOLO" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Ablation studies of architectural design. We first convert YOLOv8 directly into the corresponding spiking version. Then we convert the SpikeYOLO designed in this work into the corresponding ANN version." + }, + { + "type": "table", + "bbox": [ + 0.25, + 0.199, + 0.754, + 0.3 + ], + "angle": 0, + "content": "
ArchitectureModelParam (M)Power (mJ)T × DmAP@ 50(%)mAP@ 50:95(%)
ANN → SNNYOLOv825.8183.5167.250.2
Spiking YOLOv825.87.21 × 146.831.3
SNN → ANNSpikeYOLO (Ours)23.118.41 × 152.736.1
YOLO23.1314.1165.048.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.784, + 0.355 + ], + "angle": 0, + "content": "in our method, inference timestep is reported as \\( T \\times D \\), e.g., \\( 1 \\times 4 \\) denotes \\( T = 1, D = 4 \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.508 + ], + "angle": 0, + "content": "Main Results are shown in Table 1. The proposed SpikeYOLO significantly improves the performance upper bound of the COCO dataset in SNNs. We obtain \\(66.2\\%\\) mAP@50 and \\(48.9\\%\\) mAP@50:95, which is \\(+15.0\\%\\) and \\(+18.7\\%\\) higher than the prior state-of-the-art SNN [58], respectively. SpikeYOLO also has significant advantages over existing SNNs in terms of parameters and power: SpikeYOLO vs. EMS-YOLO [50]: Param, 23.1M vs. 26.9M; mAP@50, \\(62.3\\%\\) vs. \\(50.1\\%\\); mAP@50:95, \\(45.5\\%\\) vs. \\(30.1\\%\\); Power, \\(33.2\\mathbf{mJ}\\) vs. \\(29.0\\mathrm{mJ}\\). Moreover, the performance gap between SNNs and ANNs is significantly narrowed. For example, under similar parameters, the performance of SpikeYOLO and YOLO v5 are comparable, and the energy efficiency is \\(3.3\\times\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.643 + ], + "angle": 0, + "content": "Ablation Studies of Architectural Design. We simplify YOLOv8 for SNN and incorporate meta SNN blocks. As shown in Table 1, this architectural improvement enables the accuracy of spikeYOLO at \\( T = 1 \\) and \\( D = 1 \\) to reach \\( 52.7\\% \\), better than the prior state-of-the-art SNN. We are also interested in the question \"whether the architectures in SNNs and ANNs can be used directly interchangeably?\" We conduct the experiments in Table 2. We observe that directly converting the ANN architecture into the corresponding SNN brings significant performance degradation. The special architectural design of SNNs can improve its representation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.794 + ], + "angle": 0, + "content": "Ablation Studies of Quantization Error. Integer-valued training is designed to reduce quantization error in SNNs. The larger \\( D \\) is, the smaller the quantization error is. In Table 1, we fixed the parameters to 23.1M. When \\( T = 1 \\) and \\( T = 4 \\), we expand \\( D = 1 \\) to \\( D = 4 \\), respectively, and the accuracies of mAP@50 are increased by \\( +9.6\\% \\) and \\( +7.6\\% \\). In contrast, if we fix \\( D = 1 \\) and increase \\( T = 1 \\) to \\( T = 4 \\), the performance improvement of mAP@50 is only \\( 3\\% \\). These results show that quantization error has a greater impact on performance than the setting of timesteps. And, in terms of power, increasing \\( D \\) is more cost-effective than increasing \\( T \\). For instance, when \\( 1 \\times 1 \\) changes to \\( 1 \\times 4 \\), power increases by \\( 88\\% \\); while \\( 1 \\times 1 \\) changes to \\( 4 \\times 1 \\), power improves by \\( 267\\% \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.816, + 0.563, + 0.83 + ], + "angle": 0, + "content": "4.2 Gen1 Automotive Detection Dataset" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.338, + 0.127 + ], + "angle": 0, + "content": "Luo et al." + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.146, + 0.387, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.147, + 0.501, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.147, + 0.615, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.147, + 0.728, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.223, + 0.387, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.223, + 0.501, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.223, + 0.614, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.223, + 0.728, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.401 + ], + "angle": 0, + "content": "Fig. 4: The object detection results on the COCO dataset. The first two columns compare the effect of maximum integer value \\( D \\) on performance for the same structure. The second and third columns compare the effect of the size of the model on performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.552 + ], + "angle": 0, + "content": "Experimental Setup. As a large neuromorphic object detection dataset, Gen1 [8] encompasses 39 hours of open road and various driving scenarios, captured using an ATIS sensor with a resolution of \\(304 \\times 240\\) pixels. The dataset is organized into training, validation, and testing subsets. The bounding box annotations of pedestrians and cars (over 255,000) were manually labeled. For each annotation, we process the event-based stream 2.5 seconds before its occurrence, dividing it into \\(T\\) slices for model input. We train the model for 50 epochs and maintain other hyperparameters same as the COCO 2017 dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Main Results on Gen1 dataset are shown in Table 3. The proposed SpikeYOLO notably elevates the performance benchmark for the Gen1 dataset in SNNs. We achieve \\(67.2\\%\\) mAP@50 with 23.1M parameters, which outperforms the prior state-of-the-art SNN model by \\(+8.2\\%\\). For example, when \\(T = 5\\), SpikeYOLO vs. EMS-YOLO [50]: Param, 13.2M vs. 14.4M; mAP@50, \\(66.0\\%\\) vs. \\(59.0\\%\\); mAP@50:95, \\(38.5\\%\\) vs. \\(31.0\\%\\). In contrast to the COCO dataset, Gen1 contains temporal information that is more suitable for SNN processing. We conduct experiments on the performance of SNN and ANN with the same architecture. We observe that SpikeYOLO's mAP@50 accuracy is \\(+2.5\\%\\) higher than the corresponding ANN, and shows a \\(5.7\\times\\) energy efficiency. This indicates that SNN has attractive potential in processing neuromorphic data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Ablation Studies of Quantization Error. Both \\( T \\) and \\( D \\) significantly influence outcomes when processing neuromorphic datasets. Table 4 gives a comprehensive ablation study on SpikeYOLO with 23.1M parameters that evaluate the effects of varying \\( T \\) and \\( D \\). We observe some interesting experimental results. First, boosting the timestep \\( T \\) will bring improvement in accuracy and power. For instance, with the set of \\( D = 1 \\), extending \\( T = 1 \\) to \\( T = 4 \\) yields a \\( +6.7\\% \\) increase in mAP@50, and the power will increase by \\( 3.7\\times \\). But further extending \\( T = 4 \\) to \\( T = 8 \\) results in a marginal increase of only \\( +0.7\\% \\) and significantly" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.649, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SpikeYOLO" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.144, + 0.788, + 0.173 + ], + "angle": 0, + "content": "Table 3: Results on the Gen1 dataset [8]. * We convert 23.1M of SpikeYOLO into ANN with the same architecture." + }, + { + "type": "table", + "bbox": [ + 0.237, + 0.185, + 0.768, + 0.41 + ], + "angle": 0, + "content": "
ArchitectureModelParam (M)Power (mJ)T × DmAP@ 50(%)mAP@ 50:95(%)
ANNYOLOv3-tiny [44]10.25.1144.5-
SpikeYOLO*23.178.5164.739.7
SNNEMS-YOLO [50]6.21.2554.726.7
9.32.0556.528.6
14.43.4559.031.0
VGG-11+SDD12.611.11-17.4
MobileNet-64+SSD24.35.71-14.7
DenseNet121-24+SSD [6]8.23.91-18.9
Spiking-Yolo [29]7.9102.350044.2-
Tr-Spiking-Yolo [65]7.90.9545.3-
SpikeYOLO(Ours)13.211.05 × 166.038.5
23.119.75 × 166.438.9
23.112.94 × 267.240.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.424, + 0.788, + 0.454 + ], + "angle": 0, + "content": "Table 4: The influence of \\( T \\) and \\( D \\) on Gen1. We set SpikeYOLO (23.1M) as the baseline and vary \\( T \\) and \\( D \\) for each study." + }, + { + "type": "table", + "bbox": [ + 0.293, + 0.464, + 0.712, + 0.607 + ], + "angle": 0, + "content": "
MethodT × DPower(mJ)mAP@50(%)mAP@50:95(%)
SpikeYOLO1 × 14.059.333.1
1 × 43.9 (-0.1)65.1 (+5.8)38.9 (+5.8)
2 × 18.163.636.5
2 × 27.8 (-0.3)66.1 (+2.5)39.0 (+2.5)
2 × 47.1 (-1.0)67.0 (+3.4)40.1 (+3.6)
4 × 114.866.038.4
4 × 212.9 (-1.9)67.2 (+1.2)40.4 (+2.0)
8 × 127.066.739.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.788, + 0.743 + ], + "angle": 0, + "content": "increases energy cost. Second, We were surprised to see that by expanding \\( D \\) at a fixed \\( T \\), the performance will be improved, while the power will be dropped. For example, \\( 2 \\times 1 \\) vs. \\( 2 \\times 2 \\) vs. \\( 2 \\times 4 \\): mAP@50, \\( 63.6\\% \\) vs. \\( 66.1\\% \\) vs. \\( 67.0\\% \\); Power, \\( 8.1\\mathrm{mJ} \\) vs. \\( 7.8\\mathrm{mJ} \\) vs. \\( 7.1\\mathrm{mJ} \\). This trend is completely different from SpikeYOLO's performance in COCO, where the increase of \\( D \\) will bring more energy cost. We argue that this phenomenon is because SNN exhibits various spike firing for dense/sparse data." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.767, + 0.555, + 0.784 + ], + "angle": 0, + "content": "4.3 Architecture Ablation Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Re-parameterization Design. As shown in Table 5, if we remove re-parameterization by adding neurons into inverted separable convolutions, the mAP@50 and mAP@50:95 will decrease \\(1.7\\%\\) and \\(1.8\\%\\) respectively." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.127 + ], + "angle": 0, + "content": "Luo et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.175 + ], + "angle": 0, + "content": "Table 5: Ablation studies of architecture design. We set \\( T \\times D = 1 \\times 4 \\) and modify just one point of baseline to test how the parameters, power and performance vary." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.185, + 0.769, + 0.287 + ], + "angle": 0, + "content": "
MethodParam(M)mAP@50(%)mAP@50:95(%)
SpikeYOLO(Baseline)23.162.345.5
Remove re-parameterization23.160.643.7
SNN-Block-1 → SNN-Block-219.961.244.7
SNN-Block-2 → Transformer Block24.561.044.2
Anchor-free head → Anchor-based head21.259.539.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.789, + 0.42 + ], + "angle": 0, + "content": "SNN Block Design. Including a \\(3 \\times 3\\) standard convolution within the initial stages of convolution blocks is crucial. As shown in Table 5, substituting SNN-Block-1 for SNN-Block-2 leads to a reduction in performance of around \\(1\\%\\). Moreover, we try to replace high-stage SNN-Block-2 with meta Transformer-based SNN block, just like Meta-SpikeFormer [58]. We find that there is little to no performance gain by doing this and that the parameters increase. Therefore, only spiking CNN blocks are exploited in our SpikeYOLO." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.79, + 0.512 + ], + "angle": 0, + "content": "Detection Head. The detection mechanisms within YOLO are categorized into anchor-based heads(e.g.,YOLOv5) and anchor-free heads(e.g.,YOLOv8). The former directly predicts each bounding box's dimensions, whereas the latter estimates the probability distribution of each bounding box. Previous EMSYOLO [50] and Meta-SpikeFormer [58] employ anchor-based heads. SpikeYOLO exploits the anchor-free head because of its higher accuracy (see Table 5)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.534, + 0.36, + 0.55 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.79, + 0.779 + ], + "angle": 0, + "content": "This work significantly narrows the performance gap between SNNs and ANNs on object detection tasks. We achieve this through network architecture and spiking neuron design. The proposed SpikeYOLO architecture abandons the complex module design in the vanilla YOLO series and exploits simple meta spike blocks to build the model. Then, the I-LIF spiking neuron capable of integer-valued training and spike-driven inference is proposed to drop quantization errors. We improve the upper bound of the SNN domain's performance on the COCO dataset by \\(+15.0\\%\\) (mAP@50) and \\(+18.7\\%\\) (mAP@50:95), respectively. On the neuromorphic Gen1 dataset, SpikeYOLO achieves better performance and lower power than ANN of the same architecture. Furthermore, we investigate the performance of equivalent architecture ANNs and SNNs in different datasets, and the results show that the redesigned SNN architecture performed better. This work enables SNNs to handle complex object detection and can inspire the application of SNNs in more visual scenarios." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.649, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SpikeYOLO" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.793, + 0.266 + ], + "angle": 0, + "content": "This work was partially supported by National Distinguished Young Scholars (62325603), and National Natural Science Foundation of China (62236009, U22A20103,62441606), Beijing Natural Science Foundation for Distinguished Young Scholars (JQ21015), China Postdoctoral Science Foundation (GZB20240824, 2024M753497), and CAAI-MindSpore Open Fund, developed on OpenI Community." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.288, + 0.323, + 0.304 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.345 + ], + "angle": 0, + "content": "1. Bochkovskiy, A., Wang, C.Y., Liao, H.Y.M.: Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.347, + 0.787, + 0.4 + ], + "angle": 0, + "content": "2. Bu, T., Fang, W., Ding, J., DAI, P., Yu, Z., Huang, T.: Optimal ANN-SNN conversion for high-accuracy and ultra-low-latency spiking neural networks. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=7B3IJMM1k_M" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.402, + 0.787, + 0.441 + ], + "angle": 0, + "content": "3. Cao, Y., Chen, Y., Khosla, D.: Spiking deep convolutional neural networks for energy-efficient object recognition. International Journal of Computer Vision 113, 54-66 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.443, + 0.787, + 0.482 + ], + "angle": 0, + "content": "4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European Conference on Computer Vision. pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.484, + 0.787, + 0.524 + ], + "angle": 0, + "content": "5. Chollet, F.: Xception: Deep learning with depthwise separable convolutions. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1800-1807 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.526, + 0.787, + 0.565 + ], + "angle": 0, + "content": "6. Cordone, L., Miramond, B., Thierion, P.: Object detection with spiking neural networks on automotive event data. In: 2022 International Joint Conference on Neural Networks. pp. 1-8. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.567, + 0.787, + 0.607 + ], + "angle": 0, + "content": "7. Davies, M., Srinivasa, N., Lin, T.H., Chinya, G., Cao, Y., Choday, S.H., Dimou, G., Joshi, P., Imam, N., Jain, S., et al.: Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro 38(1), 82-99 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.608, + 0.787, + 0.647 + ], + "angle": 0, + "content": "8. De Tournemire, P., Nitti, D., Perot, E., Migliore, D., Sironi, A.: A large scale event-based detection dataset for automotive. arXiv preprint arXiv:2001.08499 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.649, + 0.787, + 0.689 + ], + "angle": 0, + "content": "9. Deng, S., Li, Y., Zhang, S., Gu, S.: Temporal efficient training of spiking neural network via gradient re-weighting. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=_XNtisL32jv" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.69, + 0.787, + 0.731 + ], + "angle": 0, + "content": "10. Diehl, P.U., Neil, D., Binas, J., Cook, M., Liu, S.C., Pfeiffer, M.: Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In: 2015 International Joint Conference on Neural Networks. pp. 1-8. IEEE (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.731, + 0.787, + 0.771 + ], + "angle": 0, + "content": "11. Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 13733-13742 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.772, + 0.787, + 0.84 + ], + "angle": 0, + "content": "12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=YicbFdNTTy" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.318, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.127 + ], + "angle": 0, + "content": "Luo et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "13. Eshraghian, J.K., Ward, M., Neftci, E.O., Wang, X., Lenz, G., Dwivedi, G., Bennamoun, M., Jeong, D.S., Lu, W.D.: Training spiking neural networks using lessons from deep learning. Proceedings of the IEEE 111(9), 1016-1054 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "14. Fang, W., Yu, Z., Chen, Y., Huang, T., Masquelier, T., Tian, Y.: Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems 34, 21056-21069 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.787, + 0.285 + ], + "angle": 0, + "content": "15. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(1), 154-180 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.285, + 0.787, + 0.311 + ], + "angle": 0, + "content": "16. Girshick, R.: Fast r-cnn. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1440-1448 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.311, + 0.787, + 0.352 + ], + "angle": 0, + "content": "17. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 580-587 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.352, + 0.787, + 0.392 + ], + "angle": 0, + "content": "18. Guo, Y., Chen, Y., Liu, X., Peng, W., Zhang, Y., Huang, X., Ma, Z.: Ternary spike: Learning ternary spikes for spiking neural networks. arXiv preprint arXiv:2312.06372 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.392, + 0.787, + 0.433 + ], + "angle": 0, + "content": "19. Guo, Y., Chen, Y., Zhang, L., Liu, X., Wang, Y., Huang, X., Ma, Z.: Im-loss: information maximization loss for spiking neural networks. Advances in Neural Information Processing Systems 35, 156-166 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.433, + 0.787, + 0.488 + ], + "angle": 0, + "content": "20. Guo, Y., Liu, X., Chen, Y., Zhang, L., Peng, W., Zhang, Y., Huang, X., Ma, Z.: Rmp-loss: Regularizing membrane potential distribution for spiking neural networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17391-17401 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.488, + 0.787, + 0.542 + ], + "angle": 0, + "content": "21. Han, M., Wang, Q., Zhang, T., Wang, Y., Zhang, D., Xu, B.: Complex dynamic neurons improved spiking transformer network for efficient automatic speech recognition. Thirty-Seventh AAAI Conference on Artificial Intelligence (AAAI 2023) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.542, + 0.787, + 0.583 + ], + "angle": 0, + "content": "22. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.583, + 0.787, + 0.624 + ], + "angle": 0, + "content": "23. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) Computer Vision - ECCV 2016. pp. 630-645. Springer International Publishing, Cham (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.624, + 0.787, + 0.664 + ], + "angle": 0, + "content": "24. Horowitz, M.: 1.1 computing's energy problem (and what we can do about it). In: 2014 IEEE International Solid-State Sircuits Conference digest of technical papers. pp. 10-14. IEEE (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.664, + 0.787, + 0.718 + ], + "angle": 0, + "content": " 25. Hu, J., Yao, M., Qiu, X., Chou, Y., Cai, Y., Qiao, N., Tian, Y., XU, B., Li, G.: High-performance temporal reversible spiking neural networks with $\\mathcal{O}\\{0\\}(1)$ training memory and $\\mathcal{M} \\mathcal{H} \\mathcal{C} \\mathcal{A}\\{0\\}(1)\\$ inference cost. In: Forty-first International Conference on Machine Learning (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "26. Hu, Y., Zheng, Q., Jiang, X., Pan, G.: Fast-snn: Fast spiking neural network by converting quantized ann. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(12), 14546-14562 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "27. Hu, Y., Deng, L., Wu, Y., Yao, M., Li, G.: Advancing spiking neural networks toward deep residual learning. IEEE Transactions on Neural Networks and Learning Systems pp. 1-15 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "28. Kim, S., Park, S., Na, B., Kim, J., Yoon, S.: Towards fast and accurate object detection in bio-inspired spiking neural networks through bayesian optimization. IEEE Access 9, 2633-2643 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.649, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SpikeYOLO" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "29. Kim, S., Park, S., Na, B., Yoon, S.: Spiking-yolo: spiking neural network for energy-efficient object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11270-11277 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.243 + ], + "angle": 0, + "content": "30. Kim, Y., Park, H., Moitra, A., Bhattacharjee, A., Venkatesha, Y., Panda, P.: Rate coding or direct coding: Which one is better for accurate, robust, and energy-efficient spiking neural networks? In: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 71-75. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.243, + 0.788, + 0.285 + ], + "angle": 0, + "content": "31. Li, C., Jones, E.G., Furber, S.: Unleashing the potential of spiking neural networks with dynamic confidence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13304-13314 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.285, + 0.788, + 0.325 + ], + "angle": 0, + "content": "32. Li, Y., He, X., Dong, Y., Kong, Q., Zeng, Y.: Spike calibration: Fast and accurate conversion of spiking neural network for object detection and segmentation. arXiv preprint arXiv:2207.02702 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.352 + ], + "angle": 0, + "content": "33. Li, Y., Geller, T., Kim, Y., Panda, P.: Seenn: Towards temporal spiking early exit neural networks. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.393 + ], + "angle": 0, + "content": "34. Lin, T.Y., Dólár, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2117-2125 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.393, + 0.788, + 0.447 + ], + "angle": 0, + "content": "35. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.473 + ], + "angle": 0, + "content": "36. Maass, W.: Networks of spiking neurons: the third generation of neural network models. Neural Networks 10(9), 1659-1671 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.528 + ], + "angle": 0, + "content": "37. Merolla, P.A., Arthur, J.V., Alvarez-Icaza, R., Cassidy, A.S., Sawada, J., Akopyan, F., Jackson, B.L., Imam, N., Guo, C., Nakamura, Y., et al.: A million spiking-neuron integrated circuit with a scalable communication network and interface. Science 345(6197), 668-673 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.569 + ], + "angle": 0, + "content": "38. Mueller, E., Studenyak, V., Auge, D., Knoll, A.: Spiking transformer networks: A rate coded approach for processing sequential data. In: 2021 7th International Conference on Systems and Informatics (ICSAI). pp. 1-5. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.61 + ], + "angle": 0, + "content": "39. Neftci, E.O., Mostafa, H., Zenke, F.: Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine 36(6), 51-63 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.61, + 0.788, + 0.651 + ], + "angle": 0, + "content": "40. Panda, P., Aketi, S.A., Roy, K.: Toward scalable, efficient, and accurate deep spiking neural networks with backward residual connections, stochastic softmax, and hybridization. Frontiers in Neuroscience 14, 653 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.691 + ], + "angle": 0, + "content": "41. Pei, J., Deng, L., Song, S., Zhao, M., Zhang, Y., Wu, S., Wang, G., Zou, Z., Wu, Z., He, W., et al.: Towards artificial general intelligence with hybrid tianjic chip architecture. Nature 572(7767), 106-111 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.691, + 0.788, + 0.732 + ], + "angle": 0, + "content": "42. Qiu, X., Zhu, R.J., Chou, Y., Wang, Z., Deng, L.j., Li, G.: Gated attention coding for training high-performance and efficient spiking neural networks. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 601-610 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.732, + 0.788, + 0.772 + ], + "angle": 0, + "content": "43. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 779-788 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.772, + 0.788, + 0.799 + ], + "angle": 0, + "content": "44. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.841 + ], + "angle": 0, + "content": "45. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems 28 (2015)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.127 + ], + "angle": 0, + "content": "Luo et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "46. Roy, K., Jaiswal, A., Panda, P.: Towards spike-based machine intelligence with neuromorphic computing. Nature 575(7784), 607-617 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.217 + ], + "angle": 0, + "content": "47. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: Mobilenetv2: Inverted residuals and linear bottlenecks. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 4510-4520 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "48. Schuman, C.D., Kulkarni, S.R., Parsa, M., Mitchell, J.P., Kay, B., et al.: Opportunities for neuromorphic computing algorithms and applications. Nature Computational Science 2(1), 10-19 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "49. Sengupta, A., Ye, Y., Wang, R., Liu, C., Roy, K.: Going deeper in spiking neural networks: Vgg and residual architectures. Frontiers in Neuroscience 13, 95 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.328 + ], + "angle": 0, + "content": "50. Su, Q., Chou, Y., Hu, Y., Li, J., Mei, S., Zhang, Z., Li, G.: Deep directly-trained spiking neural networks for object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6555-6565 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.785, + 0.37 + ], + "angle": 0, + "content": "51. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems. pp. 5998-6008 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.412 + ], + "angle": 0, + "content": "52. Wang, C.Y., Bochkovskiy, A., Liao, H.Y.M.: Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7464-7475 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.467 + ], + "angle": 0, + "content": "53. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 568-578 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.785, + 0.509 + ], + "angle": 0, + "content": "54. Wang, Z., Fang, Y., Cao, J., Zhang, Q., Wang, Z., Xu, R.: Masked spiking transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1761-1771 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "55. Wu, Y., Deng, L., Li, G., Zhu, J., Shi, L.: Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in Neuroscience 12 331 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.592 + ], + "angle": 0, + "content": "56. Wu, Y., Deng, L., Li, G., Zhu, J., Xie, Y., Shi, L.: Direct training for spiking neural networks: Faster, larger, better. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 1311-1318 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.647 + ], + "angle": 0, + "content": "57. Yao, M., Gao, H., Zhao, G., Wang, D., Lin, Y., Yang, Z., Li, G.: Temporal-wise attention spiking neural networks for event streams classification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10221-10230 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.702 + ], + "angle": 0, + "content": "58. Yao, M., Hu, J., Hu, T., Xu, Y., Zhou, Z., Tian, Y., XU, B., Li, G.: Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=1SIBN5Xyw7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.743 + ], + "angle": 0, + "content": "59. Yao, M., Hu, J., Zhao, G., Wang, Y., Zhang, Z., Xu, B., Li, G.: Inherent redundancy in spiking neural networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16924-16934 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.772 + ], + "angle": 0, + "content": "60. Yao, M., Hu, J., Zhou, Z., Yuan, L., Tian, Y., Xu, B., Li, G.: Spike-driven transformer. Advances in neural information processing systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.84 + ], + "angle": 0, + "content": "61. Yao, M., Richter, O., Zhao, G., Qiao, N., Xing, Y., Wang, D., Hu, T., Fang, W., Demirci, T., De Marchi, M., Deng, L., Yan, T., Nielsen, C., Sheik, S., Wu, C., Tian, Y., Xu, B., Li, G.: Spike-based dynamic computing with asynchronous sensing-computing neuromorphic chip. Nature Communications 15(1), 4464 (May 2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.649, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "SpikeYOLO" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "62. Yao, M., Zhang, H., Zhao, G., Zhang, X., Wang, D., Cao, G., Li, G.: Sparser spiking activity can be better: Feature refine-and-mask spiking neural network for event-based visual recognition. Neural Networks 166, 410-423 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "63. Yao, M., Zhao, G., Zhang, H., Hu, Y., Deng, L., Tian, Y., Xu, B., Li, G.: Attention spiking neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(8), 9393-9410 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.272 + ], + "angle": 0, + "content": "64. Yin, B., Corradi, F., Bohte, S.M.: Accurate and efficient time-domain classification with adaptive spiking recurrent neural networks. Nature Machine Intelligence 3(10), 905-913 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.313 + ], + "angle": 0, + "content": "65. Yuan, M., Zhang, C., Wang, Z., Liu, H., Pan, G., Tang, H.: Trainable spiking-yolo for low-latency and high-performance object detection. Neural Networks 172, 106092 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.357 + ], + "angle": 0, + "content": "66. Zhang, J., Dong, B., Zhang, H., Ding, J., Heide, F., Yin, B., Yang, X.: Spiking transformers for event-based single object tracking. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 8801-8810 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.412 + ], + "angle": 0, + "content": "67. Zhang, J., Tang, L., Yu, Z., Lu, J., Huang, T.: Spike transformer: Monocular depth estimation for spiking camera. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII. pp. 34-52. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.412, + 0.788, + 0.453 + ], + "angle": 0, + "content": "68. Zheng, H., Wu, Y., Deng, L., Hu, Y., Li, G.: Going deeper with directly-trained larger spiking neural networks. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 11062-11070 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.507 + ], + "angle": 0, + "content": "69. Zhou, Z., Zhu, Y., He, C., Wang, Y., YAN, S., Tian, Y., Yuan, L.: Spikformer: When spiking neural network meets transformer. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=frE4fUwz_h" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.551 + ], + "angle": 0, + "content": "70. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable \\(\\{\\mathrm{detr}\\}\\): Deformable transformers for end-to-end object detection. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=gZ9hCDWe6ke" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.551 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Integer-Valued Training and Spike-driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection/b44a5b42-9423-4159-817b-8bb17aae60da_origin.pdf b/2024/Integer-Valued Training and Spike-driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection/b44a5b42-9423-4159-817b-8bb17aae60da_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9735be37d240b6eb9f5adc50d012d338ef65dd57 --- /dev/null +++ b/2024/Integer-Valued Training and Spike-driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection/b44a5b42-9423-4159-817b-8bb17aae60da_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b6bab45352932241530fcbc8e9402979adc88dd507f8f473738fbd003c31fb0 +size 2917507 diff --git a/2024/Integer-Valued Training and Spike-driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection/full.md b/2024/Integer-Valued Training and Spike-driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..86e1b6d17aca9f45f42d287ffd9f271184838321 --- /dev/null +++ b/2024/Integer-Valued Training and Spike-driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection/full.md @@ -0,0 +1,340 @@ +# Integer-Valued Training and Spike-Driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection + +Xinhao Luo $^{1*}$ , Man Yao $^{1*}$ , Yuhong Chou $^{2}$ , Bo Xu $^{1}$ , and Guoqi Li $^{1}$ + +$^{1}$ Institute of Automation, Chinese Academy of Sciences {luoxinhao2023, man.yao, guoqi.li}@ia.ac.cn $^{2}$ Xi'an Jiaotong University + +Abstract. Brain-inspired Spiking Neural Networks (SNNs) have bioplausibility and low-power advantages over Artificial Neural Networks (ANNs). Applications of SNNs are currently limited to simple classification tasks because of their poor performance. In this work, we focus on bridging the performance gap between ANNs and SNNs on object detection. Our design revolves around network architecture and spiking neuron. First, the overly complex module design causes spike degradation when the YOLO series is converted to the corresponding spiking version. We design a SpikeYOLO architecture to solve this problem by simplifying the vanilla YOLO and incorporating meta SNN blocks. Second, object detection is more sensitive to quantization errors in the conversion of membrane potentials into binary spikes by spiking neurons. To address this challenge, we design a new spiking neuron that activates Integer values during training while maintaining spike-driven by extending virtual timesteps during inference. The proposed method is validated on both static and neuromorphic object detection datasets. On the static COCO dataset, we obtain $66.2\%$ mAP@50 and $48.9\%$ mAP@50:95, which is $+15.0\%$ and $+18.7\%$ higher than the prior state-of-the-art SNN, respectively. On the neuromorphic Gen1 dataset, we achieve $67.2\%$ mAP@50, which is $+2.5\%$ greater than the ANN with equivalent architecture, and the energy efficiency is improved by $5.7\times$ . Code: https://github.com/BICLab/SpikeYOLO + +Keywords: Spiking neural network $\cdot$ Object detection $\cdot$ Spike-driven $\cdot$ Neuromorphic vision $\cdot$ Neuromorphic computing + +# 1 Introduction + +Brain-inspired SNNs are known for their low power consumption [46, 48]. Spiking neurons incorporate spatio-temporal information and emit spikes when the + +membrane potentials exceed a threshold [36]. Thus, spiking neurons trigger sparse additions only when they receive a spike and are otherwise idle. This spike-driven enables SNNs to exhibit obvious low-power advantages over ANNs when deployed on neuromorphic chips [7,37,41,61]. However, the negative impact of complex neuronal dynamics and spike-driven nature is that SNNs are difficult to train and have limited task performance and application scenarios [13]. + +For example, most applications of SNN algorithms in computer vision are limited to simple image classification tasks [9, 14, 20, 25, 27, 31, 33, 42, 54, 60]. Another more commonly used and challenging computer vision task, object detection, is rarely explored in SNNs. In 2020, Spiking-YOLO [29] provided the first object detection model in deep SNNs, exploiting the method of converting ANN to SNN with thousands of timesteps. In 2023, EMS-YOLO [50] became the first work to use direct training SNNs to handle object detection. Recently, the direct training Meta-SpikeFormer [58] can process the object detection in a pre-training and fine-tuning manner for the first time. However, the performance gap between these works and ANNs is significant. In this work, we aim to bridge this gap and demonstrate the low-power of SNNs and their unique advantages in neuromorphic applications. We achieve this goal through two efforts. + +First, we design a new architecture, SpikeYOLO, which combines the macro design of YOLO with the micro design of the SNN module. Simply replacing the artificial neurons in the YOLO series [1,43,52] with spiking neurons generally does not work. Existing solutions include establishing the equivalence between ANN activation and spike firing rate [29], or improving the residual design [50]. We argue that another potential reason is that the module design in the YOLO series is too complex, which is effective in ANNs, but not suitable for SNNs. We observe that after the complex YOLO modules are converted into the corresponding spiking versions, there is a phenomenon of spike degradation in which the deep layers almost no longer emit spikes. Therefore, we tend to simplify the design in SpikeYOLO. We only retain the macro architecture in YOLOv83 while using the simple meta SNN block in Meta-SpikeFormer [58] as the basic module and performing micro design. + +Second, we design a novel spiking neuron, Integer Leaky Integrate-and-Fire (I-LIF), to reduce the quantization error of SNNs. Spike-driven is the key to low-power, while exploiting only binary signals will drop performance, especially in challenging object detection. Numerous studies have attempted to mitigate this problem, such as attention-based membrane optimization [63], ternary spike [18], information max [19], converting quantized ANNs [26], optimal ANN2SNN [2]. However, direct training based on optimization strategies can only alleviate errors; the optimal approximation of ANN2SNN requires large timesteps, and it is difficult to exploit the temporal information. In contrast, the idea of the proposed I-LIF is to use integer-valued activations to drop quantization errors and convert them into spikes during inference by extending the virtual timesteps. The features of I-LIF are: 1) Integer-valued training improves performance and is easy to train; 2) Integer activation during training can be equivalent to spike-driven + +in inference that there is only sparse addition; 3) The temporal dynamics of LIF are reserved, capable of processing neuromorphic object detection. + +The proposed method is validated on both static COCO [35] and neuromorphic Gen1 [8] object detection datasets. The performance we report on both datasets far exceeds the existing best models in SNNs. In addition, we explore the performance changes after the mutual conversion of ANN and SNN with the same architecture. The main contributions of this work are: + +- SpikeYOLO. We explore suitable architectures in SNNs for handling object detection tasks and propose SpikeYOLO, which simplifies YOLOv8 and incorporates meta SNN blocks. This inspires us that the complex modules in ANN may not be suitable for SNN architecture design. +- I-LIF Spiking Neuron. We propose an I-LIF spiking neuron that combines integer-valued training with spike-driven inference. The former is used to reduce quantization errors in spiking neurons, and the latter is the basis of the low-power nature of SNNs. +- Performance. The proposed method achieves outstanding accuracy with low power consumption on object detection datasets, demonstrating the potential of SNNs in complex vision tasks. On the COCO dataset, we obtain $66.2\%$ mAP@50 and $48.9\%$ mAP@50:95, which is $+15.0\%$ and $+18.7\%$ higher than the prior state-of-the-art SNN, respectively. On the Gen1 dataset, SpikeYOLO is $+2.5\%$ better than ANN models with $5.7\times$ energy efficiency + +# 2 Related Works + +SNN Training Method. Training methods have restricted the development of SNN for a long time. To make SNNs deeper, two training methods have been developed. ANN2SNN substitutes the ReLU function with spiking neurons, aiming to mimic the continuous activation by controlling the firing rate of spiking neurons [3, 10, 49]. It can achieve high performance but often requires long timesteps and is difficult to process sequence tasks that exploit the spatiotemporal dynamic nature of SNNs. In contrast, another directly training an SNN leverage gradient surrogate to circumvent the non-differentiability of binary spikes [39, 55]. Direct training is more flexible and requires fewer timesteps, but its performance usually suffers compared to ANNs of the same architecture. In this work, we focus on utilizing directly trained SNNs to process object detection due to its more flexibility in architectural design. + +SNN Architecture Design. The architecture of SNNs can be roughly divided into two categories: CNN-based and Transformer-based SNNs. Spiking ResNet has long dominated the SNN field because residual learning [22] can address the performance degradation of SNNs as they become deeper. Typical spiking ResNet includes vanilla spiking ResNet [68], SEW-ResNet [14], and MS-ResNet [27]. The main difference between them is the location of shortcuts and the ability to achieve identity mapping [23]. Recently, the Transformer [12, 51] architecture has become popular [21, 38, 58, 60, 66, 67, 67, 69] in the SNN field and + +has refreshed the performance upper bound. The SpikeYOLO we designed draws on the idea of the meta SNN block in Meta-SpikeFormer [58] and merges it with the YOLOv8 architecture. + +Quantization Errors in SNNs. Spiking neurons quantize continuous membrane potentials with complex spatio-temporal dynamics into binary spikes. Obviously, quantization error will limit the performance of SNNs. In ANN2SNN, the reduction of quantization error is relatively simple, which can be achieved by increasing the timestep [2]. In direct training, adjusting the relationship between membrane potential distribution and threshold is the main solution, such as attention mechanism [63] optimizes membrane potential distribution to reduce noise spikes; information max [19] is derived from the information entropy angle to optimize spike firing. However, optimization cannot change the inherently quantized error nature of binary spikes. The solution of this work is to use integer values to reduce quantization errors during training and convert them to binary spikes to ensure spike-driven inference. + +Object Detection. Object detection is a challenging yet pivotal task. Existing object detection frameworks can be simply categorized into: two-stage frameworks (RCNN series [16, 17, 45]) and single-stage frameworks (YOLO [1, 43, 52], Detr [4, 70] series). The former generate region proposals before determining precise object locations and classifications, whereas the latter directly ascertain these elements, offering a swifter, more streamlined solution. The object detection task has always been difficult for SNNs, and there are few related works. Current works include ANN2SNN-based YOLO [29, 65] and directly-trained spiking YOLO [50, 58]. Their performance is poor, and it isn't easy to meet the needs of real scenarios. + +# 3 Methods + +We exploit SpikeYOLO to process both static and neuromorphic object detection datasets. We first introduce how network inputs are unified. Then, the details of SpikeYOLO architecture and I-LIF spiking neuron are presented, respectively. + +# 3.1 Network Input + +The input of SNNs can be denoted as $X \in \mathbb{R}^{T \times C \times H \times W}$ , where $T$ is the timestep, $C$ is the channel, $H \times W$ denote the spatial resolution. + +Static Image. To leverage the spatio-temporal capabilities of SNNs, it is common practice that static images are repeated and utilized as input for each timestep $T$ . This is called direct input encoding [30, 56], where the first layer of spiking neurons in the network encodes the continuous values of the input into spike signals. + +Neuromorphic Event Stream. Neuromorphic data (also known as event-based vision) are generated by a Dynamic Vision Sensor (DVS), which only generates spikes when the logarithmic change in light intensity at a pixel surpasses a predefined threshold. An event-based stream is characterized as $(x_{n},y_{n},t_{n},p_{n})$ , + +![](images/cb459888c0d1bf6d011f1523043b9ae7175ca846f9f9d09964bdadd597f18124.jpg) +Fig. 1: The overall architecture of SpikeYOLO. We designed two SNN blocks, SNN-Block-1 and SNN-Block-2, and kept other architectures remain as YOLOv8. SNN-Block-1 employs standard convolution within its ChannelConv ( $\cdot$ ) component, whereas SNN-Block-2 utilizes re-parameterization convolution. That is, the difference between the two is the channel mixer module. In the low and high stages, we use SNN-Block-1 and SNN-Block-2, respectively. The spiking neuron is I-LIF, which activates integer values during training while converting them to binary spikes during inference. + +with each event capturing spatial coordinates $(x,y)$ , timestamp $t$ and polarity $p$ , where $p \in \{-1,1\}$ indicates whether light intensity has increased or decreased. Neuromorphic vision offers several advantages [15, 61], such as low resource requirements, high temporal resolution, and high robustness. The spike-driven nature of SNN makes it naturally suitable for processing event streams. The general strategy for neuromorphic pre-processing is to aggregate the event stream within a fixed time window into a frame format [57, 59, 62]. In this work, we follow this operation. Specifically, the total input window length is $T \times dt$ , where $dt$ and $T$ are temporal resolution and timestep, respectively. + +# 3.2 SpikeYOLO Architecture + +Overview. SpikeYOLO integrates the macro design of the YOLOv8 with the micro design of Meta-SpikeFormer [58]. The motivation is that we observe complex computations within YOLO's modules result in spike degradation [27] upon direct conversion to the SNN version. Consequently, we maintain the overarching design principles of the YOLO architecture while incorporating the inverted residual structure [47] and re-parameterization convolution [11] design of the Meta-SpikeFormer for detailed aspects. + +Network Output. In object detection, the network outputs the class and position of each object based on the input image sequence $X = \{X_{t}\}_{t = 1}^{T}$ . Suppose the input image sequence has $N$ goals, the output $B = \{B_{n}\}_{n = 1}^{N}$ can be calculated: + +$$ +B = \operatorname {M o d e l} (X), \tag {1} +$$ + +where each $B_{n} = \{f_{n},c_{n},x_{n},y_{n},w_{n},h_{n}\}$ contains information about degree of confidence $f_{n}$ , class $C_n$ , center coordinates $(x_{n},y_{n})$ and target size $(w_{n},h_{n})$ . Model $(\cdot)$ refers to the proposed SpikeYOLO architecture. + +Macro Design. Fig. 1 shows the overview of SpikeYOLO, a variation of the YOLO framework that is more suitable for the feature extraction scheme of SNNs. Specifically, YOLOv8 is a classic single-stage detection framework that partitions the image into numerous grids, with each grid responsible for predicting a target independently. Some classic designs, such as the feature pyramid network [34] in YOLOv8, play a crucial role in facilitating efficient feature extraction and fusion. By contrast, its feature extraction module, such as C2F, performs repeated feature extraction from the same set of feature maps. This module can enhance feature extraction in ANNs but does not work well in SNNs. As a compromise, we preserve the classic Backbone/Neck/Head architecture in YOLOv8 while incorporating strategies from the meta SNN block in Meta-SpikeFormer. + +Micro Design. Meta-SpikeFormer [58] is the current state-of-the-art architecture in SNNs, which explores the meta design of SNN and consists of CNN-based and Transformer-based SNN blocks. The meta block comprises a token mixer module and a channel mixer module. The difference between CNN-based and Transformer-based SNN blocks lies in the token mixer, which are spike-driven convolution and spike-driven self-attention, respectively. In this work, we mainly redesign the channel mixer module for the object detection task. As shown in Fig. 1, SNN-Block-1 and SNN-Block-2 are designed to extract low-stage and high-stage features, respectively. + +The meta SNN block in [58] can be written as: + +$$ +U ^ {\prime} = U + \operatorname {S e p C o n v} (U), \tag {2} +$$ + +$$ +U ^ {\prime \prime} = U ^ {\prime} + \operatorname {C h a n n e l C o n v} \left(U ^ {\prime}\right), \tag {3} +$$ + +where $U \in \mathbb{R}^{T \times C \times H \times W}$ is the layer input, $\operatorname{SepConv}(\cdot)$ is an inverted separable convolution module [47] with $7 \times 7$ kernel size in MobileNetv2 to capture global features, followed by a $3 \times 3$ depthwise convolution for further spatial feature fusion. Sepconv ( $\cdot$ ) can be expressed as: + +$$ +\operatorname {S e p C o n v} (U) = \operatorname {C o n v} _ {\mathrm {d w} 2} \left(\operatorname {C o n v} _ {\mathrm {p w} 2} \left(S N \left(\operatorname {C o n v} _ {\mathrm {d w} 1} \left(S N \left(\operatorname {C o n v} _ {\mathrm {p w} 1} \left(S N (U)\right)\right)\right)\right)\right)\right), \tag {4} +$$ + +where $\mathrm{Conv}_{\mathrm{dw1}}(\cdot)$ and $\mathrm{Conv}_{\mathrm{dw1}}(\cdot)$ are depthwise convolutions, $\mathrm{Conv}_{\mathrm{pw1}}(\cdot)$ and $\mathrm{Conv}_{\mathrm{pw1}}(\cdot)$ are pointwise convolutions [5]. $SN(\cdot)$ is the spiking neuron layer. + +ChannelConv $(\cdot)$ is the channel mixer, which facilitates inter-channel information fusion. We redesign ChannelConv $(\cdot)$ in object detection task. + +For SNN-Block-1, it is written as: + +$$ +\operatorname {C h a n n e l C o n v 1} \left(U ^ {\prime}\right) = \operatorname {C o n v} \left(S N \left(\operatorname {C o n v} \left(S N \left(U ^ {\prime}\right)\right)\right)\right), \tag {5} +$$ + +where $\mathrm{Conv}(\cdot)$ is a standard convolution with expansion ratio $r = 4$ . In contrast, addressing high-stage features, SNN-Block-2 employs a re-parameterization convolution to minimize parameter count, which can be described as: + +$$ +\operatorname {C h a n n e l C o n v 2} \left(U ^ {\prime}\right) = \operatorname {R e p C o n v} \left(S N \left(\operatorname {R e p C o n v} \left(S N \left(U ^ {\prime}\right)\right)\right)\right), \tag {6} +$$ + +$$ +\operatorname {R e p C o n v} \left(U ^ {\prime}\right) = \operatorname {C o n v} _ {\mathrm {p w} 2} \left(\operatorname {C o n v} _ {\mathrm {d w} 1} \left(\operatorname {C o n v} _ {\mathrm {p w} 1} \left(U ^ {\prime}\right)\right)\right), \tag {7} +$$ + +where $\mathrm{RepConv}(\cdot)$ is the re-parameterization convolution [11] with kernel size $3\times 3$ , it can be re-parameterized to a standard convolution during inference. + +# 3.3 I-LIF Spiking Neuron + +Spiking neurons propagate information in both spatial and temporal domains, and they mimic the spiking communication scheme of biological neurons. However, there are inherent quantization errors in converting the membrane potential of spiking neurons into binary spikes, which severely limits the representation of the model. Recently, Fast-SNN [26] achieves high-performance conversion with small timesteps by converting quantized ANNs into SNNs. This inspires us to "why not train directly with integer values", which can significantly reduce the quantization error. We just need to be careful to ensure that the inference is spike-driven. So, we came up with the idea of I-LIF. + +LIF. Leaky Integrate-and-Fire (LIF) spiking neuron [36] is the most popular neuron to construct SNNs due to its balance between bio-plausibility and computing complexity. The dynamics of LIF with soft reset is: + +$$ +U [ t ] = H [ t - 1 ] + X [ t ], \tag {8} +$$ + +$$ +S [ t ] = \Theta (U [ t ] - V _ {t h}), \tag {9} +$$ + +$$ +H [ t ] = \beta (U [ t ] - S [ t ]) \tag {10} +$$ + +where $t$ denotes the timestep, $U[t]$ is the membrane potential that integrates the temporal information $H[t - 1]$ and spatial information $X[t]$ . $\Theta(\cdot)$ is the Heaviside step function which equals 1 for $x \geq 0$ and 0 otherwise. If $U[t]$ exceeds the firing threshold $V_{th}$ , spiking neuron fire a spike $S[t]$ and $U[t]$ will subtract it subsequently. Otherwise, $H[t]$ will remain unchanged. And, $U[t]$ decays to $H[t]$ by a factor of $\beta$ , which denotes the decay constant. For simplicity, we focus on Eq.9 and denote the spiking neuron layer as $SN(\cdot)$ , with its input as membrane potential tensor $U$ and its output as spike tensor $S$ . + +![](images/0c04ec5cb61c3d5c74c39dc427d0bc5643b41658648c405cb0cc440bbba48030.jpg) +Fig. 2: Comparison of I-LIF and LIF. Binary spikes are emitted by LIF during both training and inference processes, which results in quantization errors. I-LIF emits integer values during the training process to reduce quantization errors, and converts them into binary spikes during inference to make the network only perform sparse addition. + +I-LIF. We propose the Integer Leaky Integrate-and-Fire (I-LIF) neuron to reduce the quantization error. As shown in Fig. 2, I-LIF emits integer values while training, and converts them into $0/1$ spikes when inference. Specifically, in I-LIF, Eq.9 is rewritten as: + +$$ +S [ t ] = \operatorname {C l i p} (r o u n d (U [ t ]), 0, D), \tag {11} +$$ + +where $\text{round}(\cdot)$ is a round symbol, $\text{Clip}(x, \min, \max)$ denotes that clipping $x$ to $[\min, \max]$ , $D$ is a hyperparameter indicating the maximum emitted integer value by I-LIF. + +Training Stage. Eq.11 is not a continuous function, making its derivative a step function, potentially causing training instability. Previous studies have introduced several surrogate gradient functions, which primarily address binary spike outputs. We consistently utilize rectangular windows as the surrogate function. For simplicity, We retain gradients solely for neurons activated in the $[0,D]$ range, nullifying all others. + +Inference Stage. Introducing integer value necessitates additional MACs (Multiply-Accumulation operations), potentially diminishing the energy efficiency of SNNs. Thus, converting integer values to binary spikes is essential. Fig. 3 shows an example of how integer values convert to binary spikes by extending virtual timesteps during inference. Specifically, the input to the spiking neuron at $l + 1$ layer can be described as $X^{l + 1}[t] = W^l S^l [t]$ . We extend the $T$ time step to $T\times D$ , and convert the integer value $S^l [t]$ to a spike sequence $\{S^l [t,d]\}_{d = 1}^D$ , which satisfied: + +$$ +\sum_ {d = 1} ^ {D} S ^ {l} [ t, d ] = S ^ {l} [ t ]. \tag {12} +$$ + +![](images/8f44748fd2c812e1cdec7d4cef98af5b8d983eb8e80bf33ca8a9ab624a9c6a76.jpg) +Fig. 3: An example of how the proposed I-LIF works. We assume $T = 3, D = 2$ , and show the corresponding binary spike sequences of integer value during inference. The membrane potential in [0.5, 1.5) are quantized to 1, while those in [1.5, 2.5) are quantized to 2. membrane potential that $> 2.5$ are also quantized to 2 due to the maximum integer value $D = 2$ . Subsequently, the membrane potential will be subtracted from the integer value. The training spike with a value of 2 will be converted into two binary spikes by extending virtual timesteps during inference. + +Thus, the neuron's input at $l + 1$ layer is reformulated as: + +$$ +X ^ {l} [ t ] = W ^ {l} \sum_ {d = 1} ^ {D} S ^ {l} [ t, d ]. \tag {13} +$$ + +Given that matrix multiplication functions as linear operators, we establish: + +$$ +W ^ {l} \sum_ {d = 1} ^ {D} S ^ {l} [ t, d ] = \sum_ {d = 1} ^ {D} \left(W ^ {l} S ^ {l} [ t, d ]\right). \tag {14} +$$ + +Therefore, the input of the neuron at $l + 1$ layer can be computed by: + +$$ +X ^ {l} [ t ] = \sum_ {d = 1} ^ {D} \left(W ^ {l} S ^ {l} [ t, d ]\right). \tag {15} +$$ + +The spike sequence $S^l[t,d]$ only contains $0/1$ , so all MACs can be converted into sparse ACs(Accumulation operations), which can ensure spike-driven when inference. + +# 4 Experiments + +We evaluate the proposed method on COCO 2017 val [35] and neuromorphic Gen1 [8] datasets. The mean Average Precision(mAP) at IOU=0.5(mAP@50), the average mAP between 0.5 and 0.95(mAP@50:95), and energy cost are reported for each model. Specifically, the power of ANNs and SNNs can be calculated as: + +$$ +E _ {A N N} = O ^ {2} \times C _ {i n} \times C _ {o u t} \times k ^ {2} \times E _ {M A C}, \tag {16} +$$ + +Table 1: Results on COCO 2017 val [35]. $T \times D$ means that we set up $T$ timesteps, and each timestep is expanded $D$ times. In prior SNNs and ANNs, $D$ defaults to 1. + +
ArchitectureModelParam (M)Power (mJ)T × DmAP@ 50(%)mAP@ 50:95(%)
ANNPVT [53]32.9520.3159.236.7
DETR [4]41.0197.8162.442.0
YOLOv5 421.2112.5164.145.4
ANN2SNNSpiking-Yolo [29]10.2-3500-25.7
Bayesian Optim [28]10.2-5000-25.9
Spike Calib [32]17.1-51245.4-
Directly-trained SNNEMS-YOLO [50]26.929.0450.130.1
Meta-SpikeFormer (MaskRCNN) [58]34.949.5144.0-
75.0140.8151.2-
Meta-SpikeFormer (YOLO) [58]16.834.8145.0-
16.870.7450.3-
SpikeYOLO(Ours)13.223.11 × 459.242.5
23.118.41 × 152.736.1
23.134.61 × 462.345.5
23.167.64 × 155.738.7
23.1134.74 × 463.346.3
48.168.51 × 464.647.4
68.884.21 × 466.248.9
+ +$$ +E _ {S N N} = (T \times D) \times f r \times O ^ {2} \times C _ {i n} \times C _ {o u t} \times k ^ {2} \times E _ {A C}, \tag {17} +$$ + +where $O$ is the feature output size, $C_{in}$ and $C_{out}$ denotes the number of input channel and output channel, $k$ is the kernel size, $fr$ denotes the average spike firing rate, $T$ is the timestep, $D$ is the upper limit of integer activation during training. We follow the most commonly used energy consumption evaluation method in the SNN field [40,63,64]. All operations assume a 32-bit floating-point implementation on $45\mathrm{nm}$ technology, where $E_{MAC} = 4.6pJ$ and $E_{AC} = 0.9pJ$ [24]. As can be seen from Eq. 16 and 17, SNN's low power comes from its sparse addition operation. The fewer spikes, the sparser the computation. + +# 4.1 COCO 2017 val Dataset + +Experimental Setup. As a predominant static dataset for object detection, COCO 2017 val [35] comprises 80 classes split into 118K training and 5K validating images. In all experiments, we set decay factor $\beta = 0.25$ , learning rate to 0.01, and adopt SGD optimizer. The models are trained for 300 epochs with a batch size of 40 on 4 NVIDIA V100 GPUs. Mosaic data augmentation [1] technique is employed. The network structure is given in the supplementary material. Note, + +Table 2: Ablation studies of architectural design. We first convert YOLOv8 directly into the corresponding spiking version. Then we convert the SpikeYOLO designed in this work into the corresponding ANN version. + +
ArchitectureModelParam (M)Power (mJ)T × DmAP@ 50(%)mAP@ 50:95(%)
ANN → SNNYOLOv825.8183.5167.250.2
Spiking YOLOv825.87.21 × 146.831.3
SNN → ANNSpikeYOLO (Ours)23.118.41 × 152.736.1
YOLO23.1314.1165.048.1
+ +in our method, inference timestep is reported as $T \times D$ , e.g., $1 \times 4$ denotes $T = 1, D = 4$ . + +Main Results are shown in Table 1. The proposed SpikeYOLO significantly improves the performance upper bound of the COCO dataset in SNNs. We obtain $66.2\%$ mAP@50 and $48.9\%$ mAP@50:95, which is $+15.0\%$ and $+18.7\%$ higher than the prior state-of-the-art SNN [58], respectively. SpikeYOLO also has significant advantages over existing SNNs in terms of parameters and power: SpikeYOLO vs. EMS-YOLO [50]: Param, 23.1M vs. 26.9M; mAP@50, $62.3\%$ vs. $50.1\%$ ; mAP@50:95, $45.5\%$ vs. $30.1\%$ ; Power, $33.2\mathbf{mJ}$ vs. $29.0\mathrm{mJ}$ . Moreover, the performance gap between SNNs and ANNs is significantly narrowed. For example, under similar parameters, the performance of SpikeYOLO and YOLO v5 are comparable, and the energy efficiency is $3.3\times$ . + +Ablation Studies of Architectural Design. We simplify YOLOv8 for SNN and incorporate meta SNN blocks. As shown in Table 1, this architectural improvement enables the accuracy of spikeYOLO at $T = 1$ and $D = 1$ to reach $52.7\%$ , better than the prior state-of-the-art SNN. We are also interested in the question "whether the architectures in SNNs and ANNs can be used directly interchangeably?" We conduct the experiments in Table 2. We observe that directly converting the ANN architecture into the corresponding SNN brings significant performance degradation. The special architectural design of SNNs can improve its representation. + +Ablation Studies of Quantization Error. Integer-valued training is designed to reduce quantization error in SNNs. The larger $D$ is, the smaller the quantization error is. In Table 1, we fixed the parameters to 23.1M. When $T = 1$ and $T = 4$ , we expand $D = 1$ to $D = 4$ , respectively, and the accuracies of mAP@50 are increased by $+9.6\%$ and $+7.6\%$ . In contrast, if we fix $D = 1$ and increase $T = 1$ to $T = 4$ , the performance improvement of mAP@50 is only $3\%$ . These results show that quantization error has a greater impact on performance than the setting of timesteps. And, in terms of power, increasing $D$ is more cost-effective than increasing $T$ . For instance, when $1 \times 1$ changes to $1 \times 4$ , power increases by $88\%$ ; while $1 \times 1$ changes to $4 \times 1$ , power improves by $267\%$ . + +# 4.2 Gen1 Automotive Detection Dataset + +![](images/16afe8a08478fb9313be8f681fcb9517c1796ad3ec5d2233b684e07ee932e1f5.jpg) + +![](images/bd2aabd570bffd763fdb08c07642b82465fe3c210297322b628aff711960c945.jpg) + +![](images/32abed312d109b59cdb0623658e92f9023bbf8b2a080975ff7108931d20043c5.jpg) + +![](images/b4d530beb57dbeeeb266f98201b1250c1aab75660ec67d6f123556976a89cfe0.jpg) + +![](images/fe6c64c5af7da18aa5399b7ab012a7fc0b88ee2541e107bf5f7ee0394f2e4140.jpg) +Fig. 4: The object detection results on the COCO dataset. The first two columns compare the effect of maximum integer value $D$ on performance for the same structure. The second and third columns compare the effect of the size of the model on performance. + +![](images/57972062b0a46e1512e7b6aa87dcb4123b92292af0b8e911b4aeb21f685dfc65.jpg) + +![](images/4b0ec6a5b167585a656232354c002b1eb31a37f80a9333b9f1174b747abf5357.jpg) + +![](images/b4f3de35749ccced1ac59aec97410af3c58f9a76913687f1bf3214afe8eb50c9.jpg) + +Experimental Setup. As a large neuromorphic object detection dataset, Gen1 [8] encompasses 39 hours of open road and various driving scenarios, captured using an ATIS sensor with a resolution of $304 \times 240$ pixels. The dataset is organized into training, validation, and testing subsets. The bounding box annotations of pedestrians and cars (over 255,000) were manually labeled. For each annotation, we process the event-based stream 2.5 seconds before its occurrence, dividing it into $T$ slices for model input. We train the model for 50 epochs and maintain other hyperparameters same as the COCO 2017 dataset. + +Main Results on Gen1 dataset are shown in Table 3. The proposed SpikeYOLO notably elevates the performance benchmark for the Gen1 dataset in SNNs. We achieve $67.2\%$ mAP@50 with 23.1M parameters, which outperforms the prior state-of-the-art SNN model by $+8.2\%$ . For example, when $T = 5$ , SpikeYOLO vs. EMS-YOLO [50]: Param, 13.2M vs. 14.4M; mAP@50, $66.0\%$ vs. $59.0\%$ ; mAP@50:95, $38.5\%$ vs. $31.0\%$ . In contrast to the COCO dataset, Gen1 contains temporal information that is more suitable for SNN processing. We conduct experiments on the performance of SNN and ANN with the same architecture. We observe that SpikeYOLO's mAP@50 accuracy is $+2.5\%$ higher than the corresponding ANN, and shows a $5.7\times$ energy efficiency. This indicates that SNN has attractive potential in processing neuromorphic data. + +Ablation Studies of Quantization Error. Both $T$ and $D$ significantly influence outcomes when processing neuromorphic datasets. Table 4 gives a comprehensive ablation study on SpikeYOLO with 23.1M parameters that evaluate the effects of varying $T$ and $D$ . We observe some interesting experimental results. First, boosting the timestep $T$ will bring improvement in accuracy and power. For instance, with the set of $D = 1$ , extending $T = 1$ to $T = 4$ yields a $+6.7\%$ increase in mAP@50, and the power will increase by $3.7\times$ . But further extending $T = 4$ to $T = 8$ results in a marginal increase of only $+0.7\%$ and significantly + +Table 3: Results on the Gen1 dataset [8]. * We convert 23.1M of SpikeYOLO into ANN with the same architecture. + +
ArchitectureModelParam (M)Power (mJ)T × DmAP@ 50(%)mAP@ 50:95(%)
ANNYOLOv3-tiny [44]10.25.1144.5-
SpikeYOLO*23.178.5164.739.7
SNNEMS-YOLO [50]6.21.2554.726.7
9.32.0556.528.6
14.43.4559.031.0
VGG-11+SDD12.611.11-17.4
MobileNet-64+SSD24.35.71-14.7
DenseNet121-24+SSD [6]8.23.91-18.9
Spiking-Yolo [29]7.9102.350044.2-
Tr-Spiking-Yolo [65]7.90.9545.3-
SpikeYOLO(Ours)13.211.05 × 166.038.5
23.119.75 × 166.438.9
23.112.94 × 267.240.4
+ +Table 4: The influence of $T$ and $D$ on Gen1. We set SpikeYOLO (23.1M) as the baseline and vary $T$ and $D$ for each study. + +
MethodT × DPower(mJ)mAP@50(%)mAP@50:95(%)
SpikeYOLO1 × 14.059.333.1
1 × 43.9 (-0.1)65.1 (+5.8)38.9 (+5.8)
2 × 18.163.636.5
2 × 27.8 (-0.3)66.1 (+2.5)39.0 (+2.5)
2 × 47.1 (-1.0)67.0 (+3.4)40.1 (+3.6)
4 × 114.866.038.4
4 × 212.9 (-1.9)67.2 (+1.2)40.4 (+2.0)
8 × 127.066.739.3
+ +increases energy cost. Second, We were surprised to see that by expanding $D$ at a fixed $T$ , the performance will be improved, while the power will be dropped. For example, $2 \times 1$ vs. $2 \times 2$ vs. $2 \times 4$ : mAP@50, $63.6\%$ vs. $66.1\%$ vs. $67.0\%$ ; Power, $8.1\mathrm{mJ}$ vs. $7.8\mathrm{mJ}$ vs. $7.1\mathrm{mJ}$ . This trend is completely different from SpikeYOLO's performance in COCO, where the increase of $D$ will bring more energy cost. We argue that this phenomenon is because SNN exhibits various spike firing for dense/sparse data. + +# 4.3 Architecture Ablation Experiments + +Re-parameterization Design. As shown in Table 5, if we remove re-parameterization by adding neurons into inverted separable convolutions, the mAP@50 and mAP@50:95 will decrease $1.7\%$ and $1.8\%$ respectively. + +Table 5: Ablation studies of architecture design. We set $T \times D = 1 \times 4$ and modify just one point of baseline to test how the parameters, power and performance vary. + +
MethodParam(M)mAP@50(%)mAP@50:95(%)
SpikeYOLO(Baseline)23.162.345.5
Remove re-parameterization23.160.643.7
SNN-Block-1 → SNN-Block-219.961.244.7
SNN-Block-2 → Transformer Block24.561.044.2
Anchor-free head → Anchor-based head21.259.539.7
+ +SNN Block Design. Including a $3 \times 3$ standard convolution within the initial stages of convolution blocks is crucial. As shown in Table 5, substituting SNN-Block-1 for SNN-Block-2 leads to a reduction in performance of around $1\%$ . Moreover, we try to replace high-stage SNN-Block-2 with meta Transformer-based SNN block, just like Meta-SpikeFormer [58]. We find that there is little to no performance gain by doing this and that the parameters increase. Therefore, only spiking CNN blocks are exploited in our SpikeYOLO. + +Detection Head. The detection mechanisms within YOLO are categorized into anchor-based heads(e.g.,YOLOv5) and anchor-free heads(e.g.,YOLOv8). The former directly predicts each bounding box's dimensions, whereas the latter estimates the probability distribution of each bounding box. Previous EMSYOLO [50] and Meta-SpikeFormer [58] employ anchor-based heads. SpikeYOLO exploits the anchor-free head because of its higher accuracy (see Table 5). + +# 5 Conclusion + +This work significantly narrows the performance gap between SNNs and ANNs on object detection tasks. We achieve this through network architecture and spiking neuron design. The proposed SpikeYOLO architecture abandons the complex module design in the vanilla YOLO series and exploits simple meta spike blocks to build the model. Then, the I-LIF spiking neuron capable of integer-valued training and spike-driven inference is proposed to drop quantization errors. We improve the upper bound of the SNN domain's performance on the COCO dataset by $+15.0\%$ (mAP@50) and $+18.7\%$ (mAP@50:95), respectively. On the neuromorphic Gen1 dataset, SpikeYOLO achieves better performance and lower power than ANN of the same architecture. Furthermore, we investigate the performance of equivalent architecture ANNs and SNNs in different datasets, and the results show that the redesigned SNN architecture performed better. This work enables SNNs to handle complex object detection and can inspire the application of SNNs in more visual scenarios. + +# Acknowledgements + +This work was partially supported by National Distinguished Young Scholars (62325603), and National Natural Science Foundation of China (62236009, U22A20103,62441606), Beijing Natural Science Foundation for Distinguished Young Scholars (JQ21015), China Postdoctoral Science Foundation (GZB20240824, 2024M753497), and CAAI-MindSpore Open Fund, developed on OpenI Community. + +# References + +1. Bochkovskiy, A., Wang, C.Y., Liao, H.Y.M.: Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934 (2020) +2. Bu, T., Fang, W., Ding, J., DAI, P., Yu, Z., Huang, T.: Optimal ANN-SNN conversion for high-accuracy and ultra-low-latency spiking neural networks. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=7B3IJMM1k_M +3. Cao, Y., Chen, Y., Khosla, D.: Spiking deep convolutional neural networks for energy-efficient object recognition. International Journal of Computer Vision 113, 54-66 (2015) +4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European Conference on Computer Vision. pp. 213-229. Springer (2020) +5. Chollet, F.: Xception: Deep learning with depthwise separable convolutions. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1800-1807 (2017) +6. Cordone, L., Miramond, B., Thierion, P.: Object detection with spiking neural networks on automotive event data. In: 2022 International Joint Conference on Neural Networks. pp. 1-8. IEEE (2022) +7. Davies, M., Srinivasa, N., Lin, T.H., Chinya, G., Cao, Y., Choday, S.H., Dimou, G., Joshi, P., Imam, N., Jain, S., et al.: Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro 38(1), 82-99 (2018) +8. De Tournemire, P., Nitti, D., Perot, E., Migliore, D., Sironi, A.: A large scale event-based detection dataset for automotive. arXiv preprint arXiv:2001.08499 (2020) +9. Deng, S., Li, Y., Zhang, S., Gu, S.: Temporal efficient training of spiking neural network via gradient re-weighting. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=_XNtisL32jv +10. Diehl, P.U., Neil, D., Binas, J., Cook, M., Liu, S.C., Pfeiffer, M.: Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In: 2015 International Joint Conference on Neural Networks. pp. 1-8. IEEE (2015) +11. Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 13733-13742 (2021) +12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=YicbFdNTTy + +13. Eshraghian, J.K., Ward, M., Neftci, E.O., Wang, X., Lenz, G., Dwivedi, G., Bennamoun, M., Jeong, D.S., Lu, W.D.: Training spiking neural networks using lessons from deep learning. Proceedings of the IEEE 111(9), 1016-1054 (2023) +14. Fang, W., Yu, Z., Chen, Y., Huang, T., Masquelier, T., Tian, Y.: Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems 34, 21056-21069 (2021) +15. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(1), 154-180 (2020) +16. Girshick, R.: Fast r-cnn. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1440-1448 (2015) +17. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 580-587 (2014) +18. Guo, Y., Chen, Y., Liu, X., Peng, W., Zhang, Y., Huang, X., Ma, Z.: Ternary spike: Learning ternary spikes for spiking neural networks. arXiv preprint arXiv:2312.06372 (2023) +19. Guo, Y., Chen, Y., Zhang, L., Liu, X., Wang, Y., Huang, X., Ma, Z.: Im-loss: information maximization loss for spiking neural networks. Advances in Neural Information Processing Systems 35, 156-166 (2022) +20. Guo, Y., Liu, X., Chen, Y., Zhang, L., Peng, W., Zhang, Y., Huang, X., Ma, Z.: Rmp-loss: Regularizing membrane potential distribution for spiking neural networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17391-17401 (2023) +21. Han, M., Wang, Q., Zhang, T., Wang, Y., Zhang, D., Xu, B.: Complex dynamic neurons improved spiking transformer network for efficient automatic speech recognition. Thirty-Seventh AAAI Conference on Artificial Intelligence (AAAI 2023) (2023) +22. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016) +23. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) Computer Vision - ECCV 2016. pp. 630-645. Springer International Publishing, Cham (2016) +24. Horowitz, M.: 1.1 computing's energy problem (and what we can do about it). In: 2014 IEEE International Solid-State Sircuits Conference digest of technical papers. pp. 10-14. IEEE (2014) + 25. Hu, J., Yao, M., Qiu, X., Chou, Y., Cai, Y., Qiao, N., Tian, Y., XU, B., Li, G.: High-performance temporal reversible spiking neural networks with $\mathcal{O}\{0\}(1)$ training memory and $\mathcal{M} \mathcal{H} \mathcal{C} \mathcal{A}\{0\}(1)\$ inference cost. In: Forty-first International Conference on Machine Learning (2024) +26. Hu, Y., Zheng, Q., Jiang, X., Pan, G.: Fast-snn: Fast spiking neural network by converting quantized ann. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(12), 14546-14562 (2023) +27. Hu, Y., Deng, L., Wu, Y., Yao, M., Li, G.: Advancing spiking neural networks toward deep residual learning. IEEE Transactions on Neural Networks and Learning Systems pp. 1-15 (2024) +28. Kim, S., Park, S., Na, B., Kim, J., Yoon, S.: Towards fast and accurate object detection in bio-inspired spiking neural networks through bayesian optimization. IEEE Access 9, 2633-2643 (2020) + +29. Kim, S., Park, S., Na, B., Yoon, S.: Spiking-yolo: spiking neural network for energy-efficient object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11270-11277 (2020) +30. Kim, Y., Park, H., Moitra, A., Bhattacharjee, A., Venkatesha, Y., Panda, P.: Rate coding or direct coding: Which one is better for accurate, robust, and energy-efficient spiking neural networks? In: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 71-75. IEEE (2022) +31. Li, C., Jones, E.G., Furber, S.: Unleashing the potential of spiking neural networks with dynamic confidence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13304-13314 (2023) +32. Li, Y., He, X., Dong, Y., Kong, Q., Zeng, Y.: Spike calibration: Fast and accurate conversion of spiking neural network for object detection and segmentation. arXiv preprint arXiv:2207.02702 (2022) +33. Li, Y., Geller, T., Kim, Y., Panda, P.: Seenn: Towards temporal spiking early exit neural networks. Advances in Neural Information Processing Systems 36 (2024) +34. Lin, T.Y., Dólár, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2117-2125 (2017) +35. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +36. Maass, W.: Networks of spiking neurons: the third generation of neural network models. Neural Networks 10(9), 1659-1671 (1997) +37. Merolla, P.A., Arthur, J.V., Alvarez-Icaza, R., Cassidy, A.S., Sawada, J., Akopyan, F., Jackson, B.L., Imam, N., Guo, C., Nakamura, Y., et al.: A million spiking-neuron integrated circuit with a scalable communication network and interface. Science 345(6197), 668-673 (2014) +38. Mueller, E., Studenyak, V., Auge, D., Knoll, A.: Spiking transformer networks: A rate coded approach for processing sequential data. In: 2021 7th International Conference on Systems and Informatics (ICSAI). pp. 1-5. IEEE (2021) +39. Neftci, E.O., Mostafa, H., Zenke, F.: Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine 36(6), 51-63 (2019) +40. Panda, P., Aketi, S.A., Roy, K.: Toward scalable, efficient, and accurate deep spiking neural networks with backward residual connections, stochastic softmax, and hybridization. Frontiers in Neuroscience 14, 653 (2020) +41. Pei, J., Deng, L., Song, S., Zhao, M., Zhang, Y., Wu, S., Wang, G., Zou, Z., Wu, Z., He, W., et al.: Towards artificial general intelligence with hybrid tianjic chip architecture. Nature 572(7767), 106-111 (2019) +42. Qiu, X., Zhu, R.J., Chou, Y., Wang, Z., Deng, L.j., Li, G.: Gated attention coding for training high-performance and efficient spiking neural networks. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 601-610 (2024) +43. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 779-788 (2016) +44. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018) +45. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems 28 (2015) + +46. Roy, K., Jaiswal, A., Panda, P.: Towards spike-based machine intelligence with neuromorphic computing. Nature 575(7784), 607-617 (2019) +47. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: Mobilenetv2: Inverted residuals and linear bottlenecks. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 4510-4520 (2018) +48. Schuman, C.D., Kulkarni, S.R., Parsa, M., Mitchell, J.P., Kay, B., et al.: Opportunities for neuromorphic computing algorithms and applications. Nature Computational Science 2(1), 10-19 (2022) +49. Sengupta, A., Ye, Y., Wang, R., Liu, C., Roy, K.: Going deeper in spiking neural networks: Vgg and residual architectures. Frontiers in Neuroscience 13, 95 (2019) +50. Su, Q., Chou, Y., Hu, Y., Li, J., Mei, S., Zhang, Z., Li, G.: Deep directly-trained spiking neural networks for object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6555-6565 (2023) +51. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems. pp. 5998-6008 (2017) +52. Wang, C.Y., Bochkovskiy, A., Liao, H.Y.M.: Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7464-7475 (2023) +53. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 568-578 (2021) +54. Wang, Z., Fang, Y., Cao, J., Zhang, Q., Wang, Z., Xu, R.: Masked spiking transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1761-1771 (2023) +55. Wu, Y., Deng, L., Li, G., Zhu, J., Shi, L.: Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in Neuroscience 12 331 (2018) +56. Wu, Y., Deng, L., Li, G., Zhu, J., Xie, Y., Shi, L.: Direct training for spiking neural networks: Faster, larger, better. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 1311-1318 (2019) +57. Yao, M., Gao, H., Zhao, G., Wang, D., Lin, Y., Yang, Z., Li, G.: Temporal-wise attention spiking neural networks for event streams classification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10221-10230 (2021) +58. Yao, M., Hu, J., Hu, T., Xu, Y., Zhou, Z., Tian, Y., XU, B., Li, G.: Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=1SIBN5Xyw7 +59. Yao, M., Hu, J., Zhao, G., Wang, Y., Zhang, Z., Xu, B., Li, G.: Inherent redundancy in spiking neural networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16924-16934 (2023) +60. Yao, M., Hu, J., Zhou, Z., Yuan, L., Tian, Y., Xu, B., Li, G.: Spike-driven transformer. Advances in neural information processing systems 36 (2024) +61. Yao, M., Richter, O., Zhao, G., Qiao, N., Xing, Y., Wang, D., Hu, T., Fang, W., Demirci, T., De Marchi, M., Deng, L., Yan, T., Nielsen, C., Sheik, S., Wu, C., Tian, Y., Xu, B., Li, G.: Spike-based dynamic computing with asynchronous sensing-computing neuromorphic chip. Nature Communications 15(1), 4464 (May 2024) + +62. Yao, M., Zhang, H., Zhao, G., Zhang, X., Wang, D., Cao, G., Li, G.: Sparser spiking activity can be better: Feature refine-and-mask spiking neural network for event-based visual recognition. Neural Networks 166, 410-423 (2023) +63. Yao, M., Zhao, G., Zhang, H., Hu, Y., Deng, L., Tian, Y., Xu, B., Li, G.: Attention spiking neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(8), 9393-9410 (2023) +64. Yin, B., Corradi, F., Bohte, S.M.: Accurate and efficient time-domain classification with adaptive spiking recurrent neural networks. Nature Machine Intelligence 3(10), 905-913 (2021) +65. Yuan, M., Zhang, C., Wang, Z., Liu, H., Pan, G., Tang, H.: Trainable spiking-yolo for low-latency and high-performance object detection. Neural Networks 172, 106092 (2024) +66. Zhang, J., Dong, B., Zhang, H., Ding, J., Heide, F., Yin, B., Yang, X.: Spiking transformers for event-based single object tracking. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 8801-8810 (2022) +67. Zhang, J., Tang, L., Yu, Z., Lu, J., Huang, T.: Spike transformer: Monocular depth estimation for spiking camera. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII. pp. 34-52. Springer (2022) +68. Zheng, H., Wu, Y., Deng, L., Hu, Y., Li, G.: Going deeper with directly-trained larger spiking neural networks. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 11062-11070 (2021) +69. Zhou, Z., Zhu, Y., He, C., Wang, Y., YAN, S., Tian, Y., Yuan, L.: Spikformer: When spiking neural network meets transformer. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=frE4fUwz_h +70. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable $\{\mathrm{detr}\}$ : Deformable transformers for end-to-end object detection. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=gZ9hCDWe6ke \ No newline at end of file diff --git a/2024/Integer-Valued Training and Spike-driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection/images.zip b/2024/Integer-Valued Training and Spike-driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..42fc399060093b6a62411f7ee3fb4dd3321cc5bb --- /dev/null +++ b/2024/Integer-Valued Training and Spike-driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df2a8b2c057148ecad261e1b806080efc3b18526ce8bd7f7ab58d57da6a674a2 +size 587520 diff --git a/2024/Integer-Valued Training and Spike-driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection/layout.json b/2024/Integer-Valued Training and Spike-driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..789503295b1e49cbacf5e69cfd818e7a8ef18b15 --- /dev/null +++ b/2024/Integer-Valued Training and Spike-driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection/layout.json @@ -0,0 +1,12037 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 111, + 468, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 111, + 468, + 182 + ], + "spans": [ + { + "bbox": [ + 146, + 111, + 468, + 182 + ], + "type": "text", + "content": "Integer-Valued Training and Spike-Driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 135, + 203, + 478, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 203, + 478, + 217 + ], + "spans": [ + { + "bbox": [ + 135, + 203, + 478, + 217 + ], + "type": "text", + "content": "Xinhao Luo" + }, + { + "bbox": [ + 135, + 203, + 478, + 217 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 135, + 203, + 478, + 217 + ], + "type": "text", + "content": ", Man Yao" + }, + { + "bbox": [ + 135, + 203, + 478, + 217 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 135, + 203, + 478, + 217 + ], + "type": "text", + "content": ", Yuhong Chou" + }, + { + "bbox": [ + 135, + 203, + 478, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 203, + 478, + 217 + ], + "type": "text", + "content": ", Bo Xu" + }, + { + "bbox": [ + 135, + 203, + 478, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 203, + 478, + 217 + ], + "type": "text", + "content": ", and Guoqi Li" + }, + { + "bbox": [ + 135, + 203, + 478, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 190, + 225, + 423, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 225, + 423, + 258 + ], + "spans": [ + { + "bbox": [ + 190, + 225, + 423, + 258 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 190, + 225, + 423, + 258 + ], + "type": "text", + "content": " Institute of Automation, Chinese Academy of Sciences {luoxinhao2023, man.yao, guoqi.li}@ia.ac.cn " + }, + { + "bbox": [ + 190, + 225, + 423, + 258 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 190, + 225, + 423, + 258 + ], + "type": "text", + "content": " Xi'an Jiaotong University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 282, + 455, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 282, + 455, + 525 + ], + "spans": [ + { + "bbox": [ + 159, + 282, + 455, + 525 + ], + "type": "text", + "content": "Abstract. Brain-inspired Spiking Neural Networks (SNNs) have bioplausibility and low-power advantages over Artificial Neural Networks (ANNs). Applications of SNNs are currently limited to simple classification tasks because of their poor performance. In this work, we focus on bridging the performance gap between ANNs and SNNs on object detection. Our design revolves around network architecture and spiking neuron. First, the overly complex module design causes spike degradation when the YOLO series is converted to the corresponding spiking version. We design a SpikeYOLO architecture to solve this problem by simplifying the vanilla YOLO and incorporating meta SNN blocks. Second, object detection is more sensitive to quantization errors in the conversion of membrane potentials into binary spikes by spiking neurons. To address this challenge, we design a new spiking neuron that activates Integer values during training while maintaining spike-driven by extending virtual timesteps during inference. The proposed method is validated on both static and neuromorphic object detection datasets. On the static COCO dataset, we obtain " + }, + { + "bbox": [ + 159, + 282, + 455, + 525 + ], + "type": "inline_equation", + "content": "66.2\\%" + }, + { + "bbox": [ + 159, + 282, + 455, + 525 + ], + "type": "text", + "content": " mAP@50 and " + }, + { + "bbox": [ + 159, + 282, + 455, + 525 + ], + "type": "inline_equation", + "content": "48.9\\%" + }, + { + "bbox": [ + 159, + 282, + 455, + 525 + ], + "type": "text", + "content": " mAP@50:95, which is " + }, + { + "bbox": [ + 159, + 282, + 455, + 525 + ], + "type": "inline_equation", + "content": "+15.0\\%" + }, + { + "bbox": [ + 159, + 282, + 455, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 159, + 282, + 455, + 525 + ], + "type": "inline_equation", + "content": "+18.7\\%" + }, + { + "bbox": [ + 159, + 282, + 455, + 525 + ], + "type": "text", + "content": " higher than the prior state-of-the-art SNN, respectively. On the neuromorphic Gen1 dataset, we achieve " + }, + { + "bbox": [ + 159, + 282, + 455, + 525 + ], + "type": "inline_equation", + "content": "67.2\\%" + }, + { + "bbox": [ + 159, + 282, + 455, + 525 + ], + "type": "text", + "content": " mAP@50, which is " + }, + { + "bbox": [ + 159, + 282, + 455, + 525 + ], + "type": "inline_equation", + "content": "+2.5\\%" + }, + { + "bbox": [ + 159, + 282, + 455, + 525 + ], + "type": "text", + "content": " greater than the ANN with equivalent architecture, and the energy efficiency is improved by " + }, + { + "bbox": [ + 159, + 282, + 455, + 525 + ], + "type": "inline_equation", + "content": "5.7\\times" + }, + { + "bbox": [ + 159, + 282, + 455, + 525 + ], + "type": "text", + "content": ". Code: https://github.com/BICLab/SpikeYOLO" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 534, + 453, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 534, + 453, + 556 + ], + "spans": [ + { + "bbox": [ + 159, + 534, + 453, + 556 + ], + "type": "text", + "content": "Keywords: Spiking neural network " + }, + { + "bbox": [ + 159, + 534, + 453, + 556 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 534, + 453, + 556 + ], + "type": "text", + "content": " Object detection " + }, + { + "bbox": [ + 159, + 534, + 453, + 556 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 534, + 453, + 556 + ], + "type": "text", + "content": " Spike-driven " + }, + { + "bbox": [ + 159, + 534, + 453, + 556 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 534, + 453, + 556 + ], + "type": "text", + "content": " Neuromorphic vision " + }, + { + "bbox": [ + 159, + 534, + 453, + 556 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 534, + 453, + 556 + ], + "type": "text", + "content": " Neuromorphic computing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 590, + 230, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 230, + 602 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 230, + 602 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 613, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 613, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 613, + 482, + 638 + ], + "type": "text", + "content": "Brain-inspired SNNs are known for their low power consumption [46, 48]. Spiking neurons incorporate spatio-temporal information and emit spikes when the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 643, + 262, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 643, + 262, + 666 + ], + "spans": [ + { + "bbox": [ + 146, + 643, + 262, + 666 + ], + "type": "text", + "content": "\\*Equal contribution. " + }, + { + "bbox": [ + 146, + 643, + 262, + 666 + ], + "type": "inline_equation", + "content": "(\\boxtimes)" + }, + { + "bbox": [ + 146, + 643, + 262, + 666 + ], + "type": "text", + "content": " Corresponding author." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "membrane potentials exceed a threshold [36]. Thus, spiking neurons trigger sparse additions only when they receive a spike and are otherwise idle. This spike-driven enables SNNs to exhibit obvious low-power advantages over ANNs when deployed on neuromorphic chips [7,37,41,61]. However, the negative impact of complex neuronal dynamics and spike-driven nature is that SNNs are difficult to train and have limited task performance and application scenarios [13]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 483, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 483, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 483, + 331 + ], + "type": "text", + "content": "For example, most applications of SNN algorithms in computer vision are limited to simple image classification tasks [9, 14, 20, 25, 27, 31, 33, 42, 54, 60]. Another more commonly used and challenging computer vision task, object detection, is rarely explored in SNNs. In 2020, Spiking-YOLO [29] provided the first object detection model in deep SNNs, exploiting the method of converting ANN to SNN with thousands of timesteps. In 2023, EMS-YOLO [50] became the first work to use direct training SNNs to handle object detection. Recently, the direct training Meta-SpikeFormer [58] can process the object detection in a pre-training and fine-tuning manner for the first time. However, the performance gap between these works and ANNs is significant. In this work, we aim to bridge this gap and demonstrate the low-power of SNNs and their unique advantages in neuromorphic applications. We achieve this goal through two efforts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 331, + 483, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 483, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 483, + 486 + ], + "type": "text", + "content": "First, we design a new architecture, SpikeYOLO, which combines the macro design of YOLO with the micro design of the SNN module. Simply replacing the artificial neurons in the YOLO series [1,43,52] with spiking neurons generally does not work. Existing solutions include establishing the equivalence between ANN activation and spike firing rate [29], or improving the residual design [50]. We argue that another potential reason is that the module design in the YOLO series is too complex, which is effective in ANNs, but not suitable for SNNs. We observe that after the complex YOLO modules are converted into the corresponding spiking versions, there is a phenomenon of spike degradation in which the deep layers almost no longer emit spikes. Therefore, we tend to simplify the design in SpikeYOLO. We only retain the macro architecture in YOLOv83 while using the simple meta SNN block in Meta-SpikeFormer [58] as the basic module and performing micro design." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 487, + 483, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 483, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 483, + 643 + ], + "type": "text", + "content": "Second, we design a novel spiking neuron, Integer Leaky Integrate-and-Fire (I-LIF), to reduce the quantization error of SNNs. Spike-driven is the key to low-power, while exploiting only binary signals will drop performance, especially in challenging object detection. Numerous studies have attempted to mitigate this problem, such as attention-based membrane optimization [63], ternary spike [18], information max [19], converting quantized ANNs [26], optimal ANN2SNN [2]. However, direct training based on optimization strategies can only alleviate errors; the optimal approximation of ANN2SNN requires large timesteps, and it is difficult to exploit the temporal information. In contrast, the idea of the proposed I-LIF is to use integer-valued activations to drop quantization errors and convert them into spikes during inference by extending the virtual timesteps. The features of I-LIF are: 1) Integer-valued training improves performance and is easy to train; 2) Integer activation during training can be equivalent to spike-driven" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "text", + "content": "Luo et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 653, + 321, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 653, + 321, + 665 + ], + "spans": [ + { + "bbox": [ + 145, + 653, + 321, + 665 + ], + "type": "text", + "content": "3https://github.com/ultralytics/ultralytics" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "in inference that there is only sparse addition; 3) The temporal dynamics of LIF are reserved, capable of processing neuromorphic object detection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 199 + ], + "type": "text", + "content": "The proposed method is validated on both static COCO [35] and neuromorphic Gen1 [8] object detection datasets. The performance we report on both datasets far exceeds the existing best models in SNNs. In addition, we explore the performance changes after the mutual conversion of ANN and SNN with the same architecture. The main contributions of this work are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 207, + 480, + 372 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 207, + 479, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 207, + 479, + 254 + ], + "spans": [ + { + "bbox": [ + 138, + 207, + 479, + 254 + ], + "type": "text", + "content": "- SpikeYOLO. We explore suitable architectures in SNNs for handling object detection tasks and propose SpikeYOLO, which simplifies YOLOv8 and incorporates meta SNN blocks. This inspires us that the complex modules in ANN may not be suitable for SNN architecture design." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 255, + 480, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 255, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 480, + 300 + ], + "type": "text", + "content": "- I-LIF Spiking Neuron. We propose an I-LIF spiking neuron that combines integer-valued training with spike-driven inference. The former is used to reduce quantization errors in spiking neurons, and the latter is the basis of the low-power nature of SNNs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 304, + 480, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 372 + ], + "type": "text", + "content": "- Performance. The proposed method achieves outstanding accuracy with low power consumption on object detection datasets, demonstrating the potential of SNNs in complex vision tasks. On the COCO dataset, we obtain " + }, + { + "bbox": [ + 138, + 304, + 480, + 372 + ], + "type": "inline_equation", + "content": "66.2\\%" + }, + { + "bbox": [ + 138, + 304, + 480, + 372 + ], + "type": "text", + "content": " mAP@50 and " + }, + { + "bbox": [ + 138, + 304, + 480, + 372 + ], + "type": "inline_equation", + "content": "48.9\\%" + }, + { + "bbox": [ + 138, + 304, + 480, + 372 + ], + "type": "text", + "content": " mAP@50:95, which is " + }, + { + "bbox": [ + 138, + 304, + 480, + 372 + ], + "type": "inline_equation", + "content": "+15.0\\%" + }, + { + "bbox": [ + 138, + 304, + 480, + 372 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 304, + 480, + 372 + ], + "type": "inline_equation", + "content": "+18.7\\%" + }, + { + "bbox": [ + 138, + 304, + 480, + 372 + ], + "type": "text", + "content": " higher than the prior state-of-the-art SNN, respectively. On the Gen1 dataset, SpikeYOLO is " + }, + { + "bbox": [ + 138, + 304, + 480, + 372 + ], + "type": "inline_equation", + "content": "+2.5\\%" + }, + { + "bbox": [ + 138, + 304, + 480, + 372 + ], + "type": "text", + "content": " better than ANN models with " + }, + { + "bbox": [ + 138, + 304, + 480, + 372 + ], + "type": "inline_equation", + "content": "5.7\\times" + }, + { + "bbox": [ + 138, + 304, + 480, + 372 + ], + "type": "text", + "content": " energy efficiency" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 401, + 242, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 401, + 242, + 414 + ], + "spans": [ + { + "bbox": [ + 132, + 401, + 242, + 414 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 426, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 569 + ], + "type": "text", + "content": "SNN Training Method. Training methods have restricted the development of SNN for a long time. To make SNNs deeper, two training methods have been developed. ANN2SNN substitutes the ReLU function with spiking neurons, aiming to mimic the continuous activation by controlling the firing rate of spiking neurons [3, 10, 49]. It can achieve high performance but often requires long timesteps and is difficult to process sequence tasks that exploit the spatiotemporal dynamic nature of SNNs. In contrast, another directly training an SNN leverage gradient surrogate to circumvent the non-differentiability of binary spikes [39, 55]. Direct training is more flexible and requires fewer timesteps, but its performance usually suffers compared to ANNs of the same architecture. In this work, we focus on utilizing directly trained SNNs to process object detection due to its more flexibility in architectural design." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "SNN Architecture Design. The architecture of SNNs can be roughly divided into two categories: CNN-based and Transformer-based SNNs. Spiking ResNet has long dominated the SNN field because residual learning [22] can address the performance degradation of SNNs as they become deeper. Typical spiking ResNet includes vanilla spiking ResNet [68], SEW-ResNet [14], and MS-ResNet [27]. The main difference between them is the location of shortcuts and the ability to achieve identity mapping [23]. Recently, the Transformer [12, 51] architecture has become popular [21, 38, 58, 60, 66, 67, 67, 69] in the SNN field and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 397, + 91, + 447, + 102 + ], + "type": "text", + "content": "SpikeYOLO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 151 + ], + "type": "text", + "content": "has refreshed the performance upper bound. The SpikeYOLO we designed draws on the idea of the meta SNN block in Meta-SpikeFormer [58] and merges it with the YOLOv8 architecture." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 295 + ], + "type": "text", + "content": "Quantization Errors in SNNs. Spiking neurons quantize continuous membrane potentials with complex spatio-temporal dynamics into binary spikes. Obviously, quantization error will limit the performance of SNNs. In ANN2SNN, the reduction of quantization error is relatively simple, which can be achieved by increasing the timestep [2]. In direct training, adjusting the relationship between membrane potential distribution and threshold is the main solution, such as attention mechanism [63] optimizes membrane potential distribution to reduce noise spikes; information max [19] is derived from the information entropy angle to optimize spike firing. However, optimization cannot change the inherently quantized error nature of binary spikes. The solution of this work is to use integer values to reduce quantization errors during training and convert them to binary spikes to ensure spike-driven inference." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 415 + ], + "type": "text", + "content": "Object Detection. Object detection is a challenging yet pivotal task. Existing object detection frameworks can be simply categorized into: two-stage frameworks (RCNN series [16, 17, 45]) and single-stage frameworks (YOLO [1, 43, 52], Detr [4, 70] series). The former generate region proposals before determining precise object locations and classifications, whereas the latter directly ascertain these elements, offering a swifter, more streamlined solution. The object detection task has always been difficult for SNNs, and there are few related works. Current works include ANN2SNN-based YOLO [29, 65] and directly-trained spiking YOLO [50, 58]. Their performance is poor, and it isn't easy to meet the needs of real scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 434, + 207, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 207, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 207, + 445 + ], + "type": "text", + "content": "3 Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 459, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 481, + 495 + ], + "type": "text", + "content": "We exploit SpikeYOLO to process both static and neuromorphic object detection datasets. We first introduce how network inputs are unified. Then, the details of SpikeYOLO architecture and I-LIF spiking neuron are presented, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 513, + 236, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 513, + 236, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 236, + 525 + ], + "type": "text", + "content": "3.1 Network Input" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 481, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 556 + ], + "type": "text", + "content": "The input of SNNs can be denoted as " + }, + { + "bbox": [ + 130, + 533, + 481, + 556 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{T \\times C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 533, + 481, + 556 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 533, + 481, + 556 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 533, + 481, + 556 + ], + "type": "text", + "content": " is the timestep, " + }, + { + "bbox": [ + 130, + 533, + 481, + 556 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 533, + 481, + 556 + ], + "type": "text", + "content": " is the channel, " + }, + { + "bbox": [ + 130, + 533, + 481, + 556 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 130, + 533, + 481, + 556 + ], + "type": "text", + "content": " denote the spatial resolution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 617 + ], + "type": "text", + "content": "Static Image. To leverage the spatio-temporal capabilities of SNNs, it is common practice that static images are repeated and utilized as input for each timestep " + }, + { + "bbox": [ + 130, + 558, + 481, + 617 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 558, + 481, + 617 + ], + "type": "text", + "content": ". This is called direct input encoding [30, 56], where the first layer of spiking neurons in the network encodes the continuous values of the input into spike signals." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Neuromorphic Event Stream. Neuromorphic data (also known as event-based vision) are generated by a Dynamic Vision Sensor (DVS), which only generates spikes when the logarithmic change in light intensity at a pixel surpasses a predefined threshold. An event-based stream is characterized as " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "(x_{n},y_{n},t_{n},p_{n})" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 207, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 207, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 207, + 100 + ], + "type": "text", + "content": "Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 114, + 455, + 323 + ], + "blocks": [ + { + "bbox": [ + 160, + 114, + 455, + 323 + ], + "lines": [ + { + "bbox": [ + 160, + 114, + 455, + 323 + ], + "spans": [ + { + "bbox": [ + 160, + 114, + 455, + 323 + ], + "type": "image", + "image_path": "cb459888c0d1bf6d011f1523043b9ae7175ca846f9f9d09964bdadd597f18124.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 332, + 482, + 409 + ], + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 409 + ], + "type": "text", + "content": "Fig. 1: The overall architecture of SpikeYOLO. We designed two SNN blocks, SNN-Block-1 and SNN-Block-2, and kept other architectures remain as YOLOv8. SNN-Block-1 employs standard convolution within its ChannelConv (" + }, + { + "bbox": [ + 130, + 332, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 130, + 332, + 482, + 409 + ], + "type": "text", + "content": ") component, whereas SNN-Block-2 utilizes re-parameterization convolution. That is, the difference between the two is the channel mixer module. In the low and high stages, we use SNN-Block-1 and SNN-Block-2, respectively. The spiking neuron is I-LIF, which activates integer values during training while converting them to binary spikes during inference." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 434, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 541 + ], + "type": "text", + "content": "with each event capturing spatial coordinates " + }, + { + "bbox": [ + 130, + 434, + 482, + 541 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 130, + 434, + 482, + 541 + ], + "type": "text", + "content": ", timestamp " + }, + { + "bbox": [ + 130, + 434, + 482, + 541 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 434, + 482, + 541 + ], + "type": "text", + "content": " and polarity " + }, + { + "bbox": [ + 130, + 434, + 482, + 541 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 434, + 482, + 541 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 434, + 482, + 541 + ], + "type": "inline_equation", + "content": "p \\in \\{-1,1\\}" + }, + { + "bbox": [ + 130, + 434, + 482, + 541 + ], + "type": "text", + "content": " indicates whether light intensity has increased or decreased. Neuromorphic vision offers several advantages [15, 61], such as low resource requirements, high temporal resolution, and high robustness. The spike-driven nature of SNN makes it naturally suitable for processing event streams. The general strategy for neuromorphic pre-processing is to aggregate the event stream within a fixed time window into a frame format [57, 59, 62]. In this work, we follow this operation. Specifically, the total input window length is " + }, + { + "bbox": [ + 130, + 434, + 482, + 541 + ], + "type": "inline_equation", + "content": "T \\times dt" + }, + { + "bbox": [ + 130, + 434, + 482, + 541 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 434, + 482, + 541 + ], + "type": "inline_equation", + "content": "dt" + }, + { + "bbox": [ + 130, + 434, + 482, + 541 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 434, + 482, + 541 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 434, + 482, + 541 + ], + "type": "text", + "content": " are temporal resolution and timestep, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 560, + 288, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 560, + 288, + 572 + ], + "spans": [ + { + "bbox": [ + 131, + 560, + 288, + 572 + ], + "type": "text", + "content": "3.2 SpikeYOLO Architecture" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": "Overview. SpikeYOLO integrates the macro design of the YOLOv8 with the micro design of Meta-SpikeFormer [58]. The motivation is that we observe complex computations within YOLO's modules result in spike degradation [27] upon direct conversion to the SNN version. Consequently, we maintain the overarching design principles of the YOLO architecture while incorporating the inverted residual structure [47] and re-parameterization convolution [11] design of the Meta-SpikeFormer for detailed aspects." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 397, + 91, + 447, + 102 + ], + "type": "text", + "content": "SpikeYOLO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 483, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 483, + 153 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 483, + 153 + ], + "type": "text", + "content": "Network Output. In object detection, the network outputs the class and position of each object based on the input image sequence " + }, + { + "bbox": [ + 132, + 116, + 483, + 153 + ], + "type": "inline_equation", + "content": "X = \\{X_{t}\\}_{t = 1}^{T}" + }, + { + "bbox": [ + 132, + 116, + 483, + 153 + ], + "type": "text", + "content": ". Suppose the input image sequence has " + }, + { + "bbox": [ + 132, + 116, + 483, + 153 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 132, + 116, + 483, + 153 + ], + "type": "text", + "content": " goals, the output " + }, + { + "bbox": [ + 132, + 116, + 483, + 153 + ], + "type": "inline_equation", + "content": "B = \\{B_{n}\\}_{n = 1}^{N}" + }, + { + "bbox": [ + 132, + 116, + 483, + 153 + ], + "type": "text", + "content": " can be calculated:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 269, + 163, + 482, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 163, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 269, + 163, + 482, + 177 + ], + "type": "interline_equation", + "content": "B = \\operatorname {M o d e l} (X), \\tag {1}", + "image_path": "9a7e646724d2a494f1f6b24e1ea5d52320ca84e706ff904a1fbb15db1afd3f14.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 180, + 483, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 180, + 483, + 217 + ], + "spans": [ + { + "bbox": [ + 131, + 180, + 483, + 217 + ], + "type": "text", + "content": "where each " + }, + { + "bbox": [ + 131, + 180, + 483, + 217 + ], + "type": "inline_equation", + "content": "B_{n} = \\{f_{n},c_{n},x_{n},y_{n},w_{n},h_{n}\\}" + }, + { + "bbox": [ + 131, + 180, + 483, + 217 + ], + "type": "text", + "content": " contains information about degree of confidence " + }, + { + "bbox": [ + 131, + 180, + 483, + 217 + ], + "type": "inline_equation", + "content": "f_{n}" + }, + { + "bbox": [ + 131, + 180, + 483, + 217 + ], + "type": "text", + "content": ", class " + }, + { + "bbox": [ + 131, + 180, + 483, + 217 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 131, + 180, + 483, + 217 + ], + "type": "text", + "content": ", center coordinates " + }, + { + "bbox": [ + 131, + 180, + 483, + 217 + ], + "type": "inline_equation", + "content": "(x_{n},y_{n})" + }, + { + "bbox": [ + 131, + 180, + 483, + 217 + ], + "type": "text", + "content": " and target size " + }, + { + "bbox": [ + 131, + 180, + 483, + 217 + ], + "type": "inline_equation", + "content": "(w_{n},h_{n})" + }, + { + "bbox": [ + 131, + 180, + 483, + 217 + ], + "type": "text", + "content": ". Model " + }, + { + "bbox": [ + 131, + 180, + 483, + 217 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 131, + 180, + 483, + 217 + ], + "type": "text", + "content": " refers to the proposed SpikeYOLO architecture." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 232, + 483, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 232, + 483, + 365 + ], + "spans": [ + { + "bbox": [ + 131, + 232, + 483, + 365 + ], + "type": "text", + "content": "Macro Design. Fig. 1 shows the overview of SpikeYOLO, a variation of the YOLO framework that is more suitable for the feature extraction scheme of SNNs. Specifically, YOLOv8 is a classic single-stage detection framework that partitions the image into numerous grids, with each grid responsible for predicting a target independently. Some classic designs, such as the feature pyramid network [34] in YOLOv8, play a crucial role in facilitating efficient feature extraction and fusion. By contrast, its feature extraction module, such as C2F, performs repeated feature extraction from the same set of feature maps. This module can enhance feature extraction in ANNs but does not work well in SNNs. As a compromise, we preserve the classic Backbone/Neck/Head architecture in YOLOv8 while incorporating strategies from the meta SNN block in Meta-SpikeFormer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 380, + 482, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 380, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 482, + 488 + ], + "type": "text", + "content": "Micro Design. Meta-SpikeFormer [58] is the current state-of-the-art architecture in SNNs, which explores the meta design of SNN and consists of CNN-based and Transformer-based SNN blocks. The meta block comprises a token mixer module and a channel mixer module. The difference between CNN-based and Transformer-based SNN blocks lies in the token mixer, which are spike-driven convolution and spike-driven self-attention, respectively. In this work, we mainly redesign the channel mixer module for the object detection task. As shown in Fig. 1, SNN-Block-1 and SNN-Block-2 are designed to extract low-stage and high-stage features, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 146, + 489, + 355, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 489, + 355, + 500 + ], + "spans": [ + { + "bbox": [ + 146, + 489, + 355, + 500 + ], + "type": "text", + "content": "The meta SNN block in [58] can be written as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 252, + 508, + 482, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 508, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 252, + 508, + 482, + 522 + ], + "type": "interline_equation", + "content": "U ^ {\\prime} = U + \\operatorname {S e p C o n v} (U), \\tag {2}", + "image_path": "7a0ef443adecc6868fbc853cba29ceed5d4dce8e2b7e8a06f52269651fa59c7b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 238, + 540, + 482, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 540, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 238, + 540, + 482, + 554 + ], + "type": "interline_equation", + "content": "U ^ {\\prime \\prime} = U ^ {\\prime} + \\operatorname {C h a n n e l C o n v} \\left(U ^ {\\prime}\\right), \\tag {3}", + "image_path": "496f8d61e87ee40ee46cf8cec644fe0c342b5048c3638cc1b4a1ebd8e7e68a3b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 557, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 557, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 131, + 557, + 482, + 606 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 557, + 482, + 606 + ], + "type": "inline_equation", + "content": "U \\in \\mathbb{R}^{T \\times C \\times H \\times W}" + }, + { + "bbox": [ + 131, + 557, + 482, + 606 + ], + "type": "text", + "content": " is the layer input, " + }, + { + "bbox": [ + 131, + 557, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\operatorname{SepConv}(\\cdot)" + }, + { + "bbox": [ + 131, + 557, + 482, + 606 + ], + "type": "text", + "content": " is an inverted separable convolution module [47] with " + }, + { + "bbox": [ + 131, + 557, + 482, + 606 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 131, + 557, + 482, + 606 + ], + "type": "text", + "content": " kernel size in MobileNetv2 to capture global features, followed by a " + }, + { + "bbox": [ + 131, + 557, + 482, + 606 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 131, + 557, + 482, + 606 + ], + "type": "text", + "content": " depthwise convolution for further spatial feature fusion. Sepconv (" + }, + { + "bbox": [ + 131, + 557, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 131, + 557, + 482, + 606 + ], + "type": "text", + "content": ") can be expressed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 140, + 623, + 482, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 623, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 140, + 623, + 482, + 635 + ], + "type": "interline_equation", + "content": "\\operatorname {S e p C o n v} (U) = \\operatorname {C o n v} _ {\\mathrm {d w} 2} \\left(\\operatorname {C o n v} _ {\\mathrm {p w} 2} \\left(S N \\left(\\operatorname {C o n v} _ {\\mathrm {d w} 1} \\left(S N \\left(\\operatorname {C o n v} _ {\\mathrm {p w} 1} \\left(S N (U)\\right)\\right)\\right)\\right)\\right)\\right), \\tag {4}", + "image_path": "fea32ce892cfbcb5488b39d4e40f495dcbd9f0c71fd16f339beb7abdd3655c31.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}_{\\mathrm{dw1}}(\\cdot)" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}_{\\mathrm{dw1}}(\\cdot)" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " are depthwise convolutions, " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}_{\\mathrm{pw1}}(\\cdot)" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}_{\\mathrm{pw1}}(\\cdot)" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " are pointwise convolutions [5]. " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "SN(\\cdot)" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " is the spiking neuron layer." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 207, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 207, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 207, + 101 + ], + "type": "text", + "content": "Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": "ChannelConv " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": " is the channel mixer, which facilitates inter-channel information fusion. We redesign ChannelConv " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": " in object detection task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 140, + 297, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 140, + 297, + 152 + ], + "spans": [ + { + "bbox": [ + 146, + 140, + 297, + 152 + ], + "type": "text", + "content": "For SNN-Block-1, it is written as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 191, + 158, + 482, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 191, + 158, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 191, + 158, + 482, + 171 + ], + "type": "interline_equation", + "content": "\\operatorname {C h a n n e l C o n v 1} \\left(U ^ {\\prime}\\right) = \\operatorname {C o n v} \\left(S N \\left(\\operatorname {C o n v} \\left(S N \\left(U ^ {\\prime}\\right)\\right)\\right)\\right), \\tag {5}", + "image_path": "5958c1d1268df2693f5c825b2be79b27c8765e9b3d5a41994f1807cac2d8e65d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 178, + 479, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 178, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 131, + 178, + 479, + 213 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 178, + 479, + 213 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}(\\cdot)" + }, + { + "bbox": [ + 131, + 178, + 479, + 213 + ], + "type": "text", + "content": " is a standard convolution with expansion ratio " + }, + { + "bbox": [ + 131, + 178, + 479, + 213 + ], + "type": "inline_equation", + "content": "r = 4" + }, + { + "bbox": [ + 131, + 178, + 479, + 213 + ], + "type": "text", + "content": ". In contrast, addressing high-stage features, SNN-Block-2 employs a re-parameterization convolution to minimize parameter count, which can be described as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 173, + 224, + 482, + 238 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 224, + 482, + 238 + ], + "spans": [ + { + "bbox": [ + 173, + 224, + 482, + 238 + ], + "type": "interline_equation", + "content": "\\operatorname {C h a n n e l C o n v 2} \\left(U ^ {\\prime}\\right) = \\operatorname {R e p C o n v} \\left(S N \\left(\\operatorname {R e p C o n v} \\left(S N \\left(U ^ {\\prime}\\right)\\right)\\right)\\right), \\tag {6}", + "image_path": "88f86482ddc7585be7c1e92dfbb9e9956dfea72345a330de43ea127ab6426690.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 190, + 253, + 481, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 253, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 190, + 253, + 481, + 266 + ], + "type": "interline_equation", + "content": "\\operatorname {R e p C o n v} \\left(U ^ {\\prime}\\right) = \\operatorname {C o n v} _ {\\mathrm {p w} 2} \\left(\\operatorname {C o n v} _ {\\mathrm {d w} 1} \\left(\\operatorname {C o n v} _ {\\mathrm {p w} 1} \\left(U ^ {\\prime}\\right)\\right)\\right), \\tag {7}", + "image_path": "ecad54a7b0ed8b2d5ad7c9ab5c60c82405466da91860d13174e308d649d6a183.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 270, + 479, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 479, + 293 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 270, + 479, + 293 + ], + "type": "inline_equation", + "content": "\\mathrm{RepConv}(\\cdot)" + }, + { + "bbox": [ + 130, + 270, + 479, + 293 + ], + "type": "text", + "content": " is the re-parameterization convolution [11] with kernel size " + }, + { + "bbox": [ + 130, + 270, + 479, + 293 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 130, + 270, + 479, + 293 + ], + "type": "text", + "content": ", it can be re-parameterized to a standard convolution during inference." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 308, + 270, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 308, + 270, + 320 + ], + "spans": [ + { + "bbox": [ + 131, + 308, + 270, + 320 + ], + "type": "text", + "content": "3.3 I-LIF Spiking Neuron" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 327, + 479, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 479, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 479, + 434 + ], + "type": "text", + "content": "Spiking neurons propagate information in both spatial and temporal domains, and they mimic the spiking communication scheme of biological neurons. However, there are inherent quantization errors in converting the membrane potential of spiking neurons into binary spikes, which severely limits the representation of the model. Recently, Fast-SNN [26] achieves high-performance conversion with small timesteps by converting quantized ANNs into SNNs. This inspires us to \"why not train directly with integer values\", which can significantly reduce the quantization error. We just need to be careful to ensure that the inference is spike-driven. So, we came up with the idea of I-LIF." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 449, + 479, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 449, + 479, + 485 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 479, + 485 + ], + "type": "text", + "content": "LIF. Leaky Integrate-and-Fire (LIF) spiking neuron [36] is the most popular neuron to construct SNNs due to its balance between bio-plausibility and computing complexity. The dynamics of LIF with soft reset is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 251, + 497, + 481, + 511 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 497, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 251, + 497, + 481, + 511 + ], + "type": "interline_equation", + "content": "U [ t ] = H [ t - 1 ] + X [ t ], \\tag {8}", + "image_path": "529e3a05359d909d56a010affb69e39cc92597623e0d7503b13c2aeadca97605.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 256, + 525, + 481, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 525, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 256, + 525, + 481, + 539 + ], + "type": "interline_equation", + "content": "S [ t ] = \\Theta (U [ t ] - V _ {t h}), \\tag {9}", + "image_path": "cbe3ae8f27e5e11d7969bf34cfbc1b948b7c724fcb5918efbc48bfcf1cf9929d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 254, + 553, + 481, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 553, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 254, + 553, + 481, + 567 + ], + "type": "interline_equation", + "content": "H [ t ] = \\beta (U [ t ] - S [ t ]) \\tag {10}", + "image_path": "fe30928fcc8034279ed003736bb149ac6b98722657b3db441cbf2ed81e0b2fa6.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": " denotes the timestep, " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "U[t]" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": " is the membrane potential that integrates the temporal information " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "H[t - 1]" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": " and spatial information " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "X[t]" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\Theta(\\cdot)" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": " is the Heaviside step function which equals 1 for " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "x \\geq 0" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": " and 0 otherwise. If " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "U[t]" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": " exceeds the firing threshold " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "V_{th}" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": ", spiking neuron fire a spike " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "S[t]" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "U[t]" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": " will subtract it subsequently. Otherwise, " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "H[t]" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": " will remain unchanged. And, " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "U[t]" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": " decays to " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "H[t]" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": " by a factor of " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": ", which denotes the decay constant. For simplicity, we focus on Eq.9 and denote the spiking neuron layer as " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "SN(\\cdot)" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": ", with its input as membrane potential tensor " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": " and its output as spike tensor " + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 396, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 396, + 91, + 447, + 102 + ], + "type": "text", + "content": "SpikeYOLO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 114, + 452, + 286 + ], + "blocks": [ + { + "bbox": [ + 161, + 114, + 452, + 286 + ], + "lines": [ + { + "bbox": [ + 161, + 114, + 452, + 286 + ], + "spans": [ + { + "bbox": [ + 161, + 114, + 452, + 286 + ], + "type": "image", + "image_path": "0c04ec5cb61c3d5c74c39dc427d0bc5643b41658648c405cb0cc440bbba48030.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 293, + 482, + 339 + ], + "lines": [ + { + "bbox": [ + 131, + 293, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 131, + 293, + 482, + 339 + ], + "type": "text", + "content": "Fig. 2: Comparison of I-LIF and LIF. Binary spikes are emitted by LIF during both training and inference processes, which results in quantization errors. I-LIF emits integer values during the training process to reduce quantization errors, and converts them into binary spikes during inference to make the network only perform sparse addition." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 359, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 482, + 407 + ], + "type": "text", + "content": "I-LIF. We propose the Integer Leaky Integrate-and-Fire (I-LIF) neuron to reduce the quantization error. As shown in Fig. 2, I-LIF emits integer values while training, and converts them into " + }, + { + "bbox": [ + 130, + 359, + 482, + 407 + ], + "type": "inline_equation", + "content": "0/1" + }, + { + "bbox": [ + 130, + 359, + 482, + 407 + ], + "type": "text", + "content": " spikes when inference. Specifically, in I-LIF, Eq.9 is rewritten as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 235, + 415, + 482, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 415, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 235, + 415, + 482, + 429 + ], + "type": "interline_equation", + "content": "S [ t ] = \\operatorname {C l i p} (r o u n d (U [ t ]), 0, D), \\tag {11}", + "image_path": "4b0072cbb4fdb34498b7d55f2181df4b85963c6ea1de261fe6fc20e8bd542819.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 434, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 470 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 434, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\text{round}(\\cdot)" + }, + { + "bbox": [ + 130, + 434, + 482, + 470 + ], + "type": "text", + "content": " is a round symbol, " + }, + { + "bbox": [ + 130, + 434, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\text{Clip}(x, \\min, \\max)" + }, + { + "bbox": [ + 130, + 434, + 482, + 470 + ], + "type": "text", + "content": " denotes that clipping " + }, + { + "bbox": [ + 130, + 434, + 482, + 470 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 434, + 482, + 470 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 434, + 482, + 470 + ], + "type": "inline_equation", + "content": "[\\min, \\max]" + }, + { + "bbox": [ + 130, + 434, + 482, + 470 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 434, + 482, + 470 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 434, + 482, + 470 + ], + "type": "text", + "content": " is a hyperparameter indicating the maximum emitted integer value by I-LIF." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 471, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 542 + ], + "type": "text", + "content": "Training Stage. Eq.11 is not a continuous function, making its derivative a step function, potentially causing training instability. Previous studies have introduced several surrogate gradient functions, which primarily address binary spike outputs. We consistently utilize rectangular windows as the surrogate function. For simplicity, We retain gradients solely for neurons activated in the " + }, + { + "bbox": [ + 130, + 471, + 482, + 542 + ], + "type": "inline_equation", + "content": "[0,D]" + }, + { + "bbox": [ + 130, + 471, + 482, + 542 + ], + "type": "text", + "content": " range, nullifying all others." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 542, + 483, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 542, + 483, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 542, + 483, + 636 + ], + "type": "text", + "content": "Inference Stage. Introducing integer value necessitates additional MACs (Multiply-Accumulation operations), potentially diminishing the energy efficiency of SNNs. Thus, converting integer values to binary spikes is essential. Fig. 3 shows an example of how integer values convert to binary spikes by extending virtual timesteps during inference. Specifically, the input to the spiking neuron at " + }, + { + "bbox": [ + 130, + 542, + 483, + 636 + ], + "type": "inline_equation", + "content": "l + 1" + }, + { + "bbox": [ + 130, + 542, + 483, + 636 + ], + "type": "text", + "content": " layer can be described as " + }, + { + "bbox": [ + 130, + 542, + 483, + 636 + ], + "type": "inline_equation", + "content": "X^{l + 1}[t] = W^l S^l [t]" + }, + { + "bbox": [ + 130, + 542, + 483, + 636 + ], + "type": "text", + "content": ". We extend the " + }, + { + "bbox": [ + 130, + 542, + 483, + 636 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 542, + 483, + 636 + ], + "type": "text", + "content": " time step to " + }, + { + "bbox": [ + 130, + 542, + 483, + 636 + ], + "type": "inline_equation", + "content": "T\\times D" + }, + { + "bbox": [ + 130, + 542, + 483, + 636 + ], + "type": "text", + "content": ", and convert the integer value " + }, + { + "bbox": [ + 130, + 542, + 483, + 636 + ], + "type": "inline_equation", + "content": "S^l [t]" + }, + { + "bbox": [ + 130, + 542, + 483, + 636 + ], + "type": "text", + "content": " to a spike sequence " + }, + { + "bbox": [ + 130, + 542, + 483, + 636 + ], + "type": "inline_equation", + "content": "\\{S^l [t,d]\\}_{d = 1}^D" + }, + { + "bbox": [ + 130, + 542, + 483, + 636 + ], + "type": "text", + "content": ", which satisfied:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 261, + 635, + 482, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 635, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 261, + 635, + 482, + 668 + ], + "type": "interline_equation", + "content": "\\sum_ {d = 1} ^ {D} S ^ {l} [ t, d ] = S ^ {l} [ t ]. \\tag {12}", + "image_path": "42ead3b2972a08973023cc587ef9c93e2c15097b06f3f42a4ef35acd14d9b477.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 112, + 455, + 244 + ], + "blocks": [ + { + "bbox": [ + 159, + 112, + 455, + 244 + ], + "lines": [ + { + "bbox": [ + 159, + 112, + 455, + 244 + ], + "spans": [ + { + "bbox": [ + 159, + 112, + 455, + 244 + ], + "type": "image", + "image_path": "8f44748fd2c812e1cdec7d4cef98af5b8d983eb8e80bf33ca8a9ab624a9c6a76.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 252, + 482, + 330 + ], + "lines": [ + { + "bbox": [ + 131, + 252, + 482, + 330 + ], + "spans": [ + { + "bbox": [ + 131, + 252, + 482, + 330 + ], + "type": "text", + "content": "Fig. 3: An example of how the proposed I-LIF works. We assume " + }, + { + "bbox": [ + 131, + 252, + 482, + 330 + ], + "type": "inline_equation", + "content": "T = 3, D = 2" + }, + { + "bbox": [ + 131, + 252, + 482, + 330 + ], + "type": "text", + "content": ", and show the corresponding binary spike sequences of integer value during inference. The membrane potential in [0.5, 1.5) are quantized to 1, while those in [1.5, 2.5) are quantized to 2. membrane potential that " + }, + { + "bbox": [ + 131, + 252, + 482, + 330 + ], + "type": "inline_equation", + "content": "> 2.5" + }, + { + "bbox": [ + 131, + 252, + 482, + 330 + ], + "type": "text", + "content": " are also quantized to 2 due to the maximum integer value " + }, + { + "bbox": [ + 131, + 252, + 482, + 330 + ], + "type": "inline_equation", + "content": "D = 2" + }, + { + "bbox": [ + 131, + 252, + 482, + 330 + ], + "type": "text", + "content": ". Subsequently, the membrane potential will be subtracted from the integer value. The training spike with a value of 2 will be converted into two binary spikes by extending virtual timesteps during inference." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 351, + 387, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 387, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 387, + 363 + ], + "type": "text", + "content": "Thus, the neuron's input at " + }, + { + "bbox": [ + 132, + 351, + 387, + 363 + ], + "type": "inline_equation", + "content": "l + 1" + }, + { + "bbox": [ + 132, + 351, + 387, + 363 + ], + "type": "text", + "content": " layer is reformulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 253, + 369, + 482, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 369, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 253, + 369, + 482, + 402 + ], + "type": "interline_equation", + "content": "X ^ {l} [ t ] = W ^ {l} \\sum_ {d = 1} ^ {D} S ^ {l} [ t, d ]. \\tag {13}", + "image_path": "60f6440d30d2693c7bd06a45e3ef9b42cb46d0a689b363c85ceba103a94e1352.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 408, + 466, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 466, + 421 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 466, + 421 + ], + "type": "text", + "content": "Given that matrix multiplication functions as linear operators, we establish:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 232, + 426, + 482, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 426, + 482, + 459 + ], + "spans": [ + { + "bbox": [ + 232, + 426, + 482, + 459 + ], + "type": "interline_equation", + "content": "W ^ {l} \\sum_ {d = 1} ^ {D} S ^ {l} [ t, d ] = \\sum_ {d = 1} ^ {D} \\left(W ^ {l} S ^ {l} [ t, d ]\\right). \\tag {14}", + "image_path": "bd89f1466ca79c2ab4bc01e687ffb15e58cef462c6350d75ce3061040272550a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 464, + 436, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 436, + 477 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 436, + 477 + ], + "type": "text", + "content": "Therefore, the input of the neuron at " + }, + { + "bbox": [ + 132, + 464, + 436, + 477 + ], + "type": "inline_equation", + "content": "l + 1" + }, + { + "bbox": [ + 132, + 464, + 436, + 477 + ], + "type": "text", + "content": " layer can be computed by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 252, + 483, + 481, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 483, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 252, + 483, + 481, + 515 + ], + "type": "interline_equation", + "content": "X ^ {l} [ t ] = \\sum_ {d = 1} ^ {D} \\left(W ^ {l} S ^ {l} [ t, d ]\\right). \\tag {15}", + "image_path": "f2bb31456356abfbe92048e5ce6e4386300c39474446c29ca847f44274ea953b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 522, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 522, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 131, + 522, + 482, + 559 + ], + "type": "text", + "content": "The spike sequence " + }, + { + "bbox": [ + 131, + 522, + 482, + 559 + ], + "type": "inline_equation", + "content": "S^l[t,d]" + }, + { + "bbox": [ + 131, + 522, + 482, + 559 + ], + "type": "text", + "content": " only contains " + }, + { + "bbox": [ + 131, + 522, + 482, + 559 + ], + "type": "inline_equation", + "content": "0/1" + }, + { + "bbox": [ + 131, + 522, + 482, + 559 + ], + "type": "text", + "content": ", so all MACs can be converted into sparse ACs(Accumulation operations), which can ensure spike-driven when inference." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 575, + 230, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 575, + 230, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 575, + 230, + 589 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 597, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 482, + 647 + ], + "type": "text", + "content": "We evaluate the proposed method on COCO 2017 val [35] and neuromorphic Gen1 [8] datasets. The mean Average Precision(mAP) at IOU=0.5(mAP@50), the average mAP between 0.5 and 0.95(mAP@50:95), and energy cost are reported for each model. Specifically, the power of ANNs and SNNs can be calculated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 215, + 652, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 652, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 215, + 652, + 481, + 666 + ], + "type": "interline_equation", + "content": "E _ {A N N} = O ^ {2} \\times C _ {i n} \\times C _ {o u t} \\times k ^ {2} \\times E _ {M A C}, \\tag {16}", + "image_path": "b86a39ba4f649591344f1e11e8ad231cb77f8ff1f5342c74f9fda08e98062249.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 397, + 91, + 447, + 102 + ], + "type": "text", + "content": "SpikeYOLO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 146, + 146, + 469, + 380 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 1: Results on COCO 2017 val [35]. " + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "T \\times D" + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": " means that we set up " + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": " timesteps, and each timestep is expanded " + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": " times. In prior SNNs and ANNs, " + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": " defaults to 1." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 146, + 146, + 469, + 380 + ], + "lines": [ + { + "bbox": [ + 146, + 146, + 469, + 380 + ], + "spans": [ + { + "bbox": [ + 146, + 146, + 469, + 380 + ], + "type": "table", + "html": "
ArchitectureModelParam (M)Power (mJ)T × DmAP@ 50(%)mAP@ 50:95(%)
ANNPVT [53]32.9520.3159.236.7
DETR [4]41.0197.8162.442.0
YOLOv5 421.2112.5164.145.4
ANN2SNNSpiking-Yolo [29]10.2-3500-25.7
Bayesian Optim [28]10.2-5000-25.9
Spike Calib [32]17.1-51245.4-
Directly-trained SNNEMS-YOLO [50]26.929.0450.130.1
Meta-SpikeFormer (MaskRCNN) [58]34.949.5144.0-
75.0140.8151.2-
Meta-SpikeFormer (YOLO) [58]16.834.8145.0-
16.870.7450.3-
SpikeYOLO(Ours)13.223.11 × 459.242.5
23.118.41 × 152.736.1
23.134.61 × 462.345.5
23.167.64 × 155.738.7
23.1134.74 × 463.346.3
48.168.51 × 464.647.4
68.884.21 × 466.248.9
", + "image_path": "2741f760f7575439ad0768b762cf5ffa8ee26ba0dd0e02b5f9bd7eb4434b1c51.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 184, + 413, + 482, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 413, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 184, + 413, + 482, + 427 + ], + "type": "interline_equation", + "content": "E _ {S N N} = (T \\times D) \\times f r \\times O ^ {2} \\times C _ {i n} \\times C _ {o u t} \\times k ^ {2} \\times E _ {A C}, \\tag {17}", + "image_path": "8c3b465bb6ce153294288c3c31c9f57df15b956cd3e0fb6d4d7a86147e5d6a8e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "text", + "content": " is the feature output size, " + }, + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "inline_equation", + "content": "C_{in}" + }, + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "inline_equation", + "content": "C_{out}" + }, + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "text", + "content": " denotes the number of input channel and output channel, " + }, + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "text", + "content": " is the kernel size, " + }, + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "inline_equation", + "content": "fr" + }, + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "text", + "content": " denotes the average spike firing rate, " + }, + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "text", + "content": " is the timestep, " + }, + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "text", + "content": " is the upper limit of integer activation during training. We follow the most commonly used energy consumption evaluation method in the SNN field [40,63,64]. All operations assume a 32-bit floating-point implementation on " + }, + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "inline_equation", + "content": "45\\mathrm{nm}" + }, + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "text", + "content": " technology, where " + }, + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "inline_equation", + "content": "E_{MAC} = 4.6pJ" + }, + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "inline_equation", + "content": "E_{AC} = 0.9pJ" + }, + { + "bbox": [ + 130, + 432, + 482, + 529 + ], + "type": "text", + "content": " [24]. As can be seen from Eq. 16 and 17, SNN's low power comes from its sparse addition operation. The fewer spikes, the sparser the computation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 545, + 282, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 545, + 282, + 556 + ], + "spans": [ + { + "bbox": [ + 131, + 545, + 282, + 556 + ], + "type": "text", + "content": "4.1 COCO 2017 val Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 565, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 565, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 565, + 482, + 639 + ], + "type": "text", + "content": "Experimental Setup. As a predominant static dataset for object detection, COCO 2017 val [35] comprises 80 classes split into 118K training and 5K validating images. In all experiments, we set decay factor " + }, + { + "bbox": [ + 130, + 565, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\beta = 0.25" + }, + { + "bbox": [ + 130, + 565, + 482, + 639 + ], + "type": "text", + "content": ", learning rate to 0.01, and adopt SGD optimizer. The models are trained for 300 epochs with a batch size of 40 on 4 NVIDIA V100 GPUs. Mosaic data augmentation [1] technique is employed. The network structure is given in the supplementary material. Note," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "text", + "content": "Luo et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 653, + 306, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 653, + 306, + 666 + ], + "spans": [ + { + "bbox": [ + 145, + 653, + 306, + 666 + ], + "type": "text", + "content": "4https://github.com/ultralytics/yolov5" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 153, + 157, + 461, + 237 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Ablation studies of architectural design. We first convert YOLOv8 directly into the corresponding spiking version. Then we convert the SpikeYOLO designed in this work into the corresponding ANN version." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 157, + 461, + 237 + ], + "lines": [ + { + "bbox": [ + 153, + 157, + 461, + 237 + ], + "spans": [ + { + "bbox": [ + 153, + 157, + 461, + 237 + ], + "type": "table", + "html": "
ArchitectureModelParam (M)Power (mJ)T × DmAP@ 50(%)mAP@ 50:95(%)
ANN → SNNYOLOv825.8183.5167.250.2
Spiking YOLOv825.87.21 × 146.831.3
SNN → ANNSpikeYOLO (Ours)23.118.41 × 152.736.1
YOLO23.1314.1165.048.1
", + "image_path": "35c4c9d4985b1dd93f0352d17af5f12eddba21d03fed6cfa38f6e6f442efdd8f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 479, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 479, + 281 + ], + "type": "text", + "content": "in our method, inference timestep is reported as " + }, + { + "bbox": [ + 130, + 258, + 479, + 281 + ], + "type": "inline_equation", + "content": "T \\times D" + }, + { + "bbox": [ + 130, + 258, + 479, + 281 + ], + "type": "text", + "content": ", e.g., " + }, + { + "bbox": [ + 130, + 258, + 479, + 281 + ], + "type": "inline_equation", + "content": "1 \\times 4" + }, + { + "bbox": [ + 130, + 258, + 479, + 281 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 130, + 258, + 479, + 281 + ], + "type": "inline_equation", + "content": "T = 1, D = 4" + }, + { + "bbox": [ + 130, + 258, + 479, + 281 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "text", + "content": "Main Results are shown in Table 1. The proposed SpikeYOLO significantly improves the performance upper bound of the COCO dataset in SNNs. We obtain " + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "inline_equation", + "content": "66.2\\%" + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "text", + "content": " mAP@50 and " + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "inline_equation", + "content": "48.9\\%" + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "text", + "content": " mAP@50:95, which is " + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "inline_equation", + "content": "+15.0\\%" + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "inline_equation", + "content": "+18.7\\%" + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "text", + "content": " higher than the prior state-of-the-art SNN [58], respectively. SpikeYOLO also has significant advantages over existing SNNs in terms of parameters and power: SpikeYOLO vs. EMS-YOLO [50]: Param, 23.1M vs. 26.9M; mAP@50, " + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "inline_equation", + "content": "62.3\\%" + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "inline_equation", + "content": "50.1\\%" + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "text", + "content": "; mAP@50:95, " + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "inline_equation", + "content": "45.5\\%" + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "inline_equation", + "content": "30.1\\%" + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "text", + "content": "; Power, " + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "inline_equation", + "content": "33.2\\mathbf{mJ}" + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "inline_equation", + "content": "29.0\\mathrm{mJ}" + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "text", + "content": ". Moreover, the performance gap between SNNs and ANNs is significantly narrowed. For example, under similar parameters, the performance of SpikeYOLO and YOLO v5 are comparable, and the energy efficiency is " + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "inline_equation", + "content": "3.3\\times" + }, + { + "bbox": [ + 130, + 282, + 482, + 402 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "type": "text", + "content": "Ablation Studies of Architectural Design. We simplify YOLOv8 for SNN and incorporate meta SNN blocks. As shown in Table 1, this architectural improvement enables the accuracy of spikeYOLO at " + }, + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "type": "inline_equation", + "content": "T = 1" + }, + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "type": "inline_equation", + "content": "D = 1" + }, + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "type": "text", + "content": " to reach " + }, + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "type": "inline_equation", + "content": "52.7\\%" + }, + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "type": "text", + "content": ", better than the prior state-of-the-art SNN. We are also interested in the question \"whether the architectures in SNNs and ANNs can be used directly interchangeably?\" We conduct the experiments in Table 2. We observe that directly converting the ANN architecture into the corresponding SNN brings significant performance degradation. The special architectural design of SNNs can improve its representation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "content": "Ablation Studies of Quantization Error. Integer-valued training is designed to reduce quantization error in SNNs. The larger " + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "content": " is, the smaller the quantization error is. In Table 1, we fixed the parameters to 23.1M. When " + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "inline_equation", + "content": "T = 1" + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "inline_equation", + "content": "T = 4" + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "content": ", we expand " + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "inline_equation", + "content": "D = 1" + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "inline_equation", + "content": "D = 4" + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "content": ", respectively, and the accuracies of mAP@50 are increased by " + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "inline_equation", + "content": "+9.6\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "inline_equation", + "content": "+7.6\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "content": ". In contrast, if we fix " + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "inline_equation", + "content": "D = 1" + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "content": " and increase " + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "inline_equation", + "content": "T = 1" + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "inline_equation", + "content": "T = 4" + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "content": ", the performance improvement of mAP@50 is only " + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "content": ". These results show that quantization error has a greater impact on performance than the setting of timesteps. And, in terms of power, increasing " + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "content": " is more cost-effective than increasing " + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "content": ". For instance, when " + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "content": " changes to " + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "inline_equation", + "content": "1 \\times 4" + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "content": ", power increases by " + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "inline_equation", + "content": "88\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "content": "; while " + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "content": " changes to " + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "inline_equation", + "content": "4 \\times 1" + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "content": ", power improves by " + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "inline_equation", + "content": "267\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 646, + 344, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 646, + 344, + 657 + ], + "spans": [ + { + "bbox": [ + 131, + 646, + 344, + 657 + ], + "type": "text", + "content": "4.2 Gen1 Automotive Detection Dataset" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 396, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 396, + 91, + 447, + 102 + ], + "type": "text", + "content": "SpikeYOLO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 115, + 236, + 176 + ], + "blocks": [ + { + "bbox": [ + 167, + 115, + 236, + 176 + ], + "lines": [ + { + "bbox": [ + 167, + 115, + 236, + 176 + ], + "spans": [ + { + "bbox": [ + 167, + 115, + 236, + 176 + ], + "type": "image", + "image_path": "16afe8a08478fb9313be8f681fcb9517c1796ad3ec5d2233b684e07ee932e1f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 238, + 116, + 306, + 175 + ], + "blocks": [ + { + "bbox": [ + 238, + 116, + 306, + 175 + ], + "lines": [ + { + "bbox": [ + 238, + 116, + 306, + 175 + ], + "spans": [ + { + "bbox": [ + 238, + 116, + 306, + 175 + ], + "type": "image", + "image_path": "bd2aabd570bffd763fdb08c07642b82465fe3c210297322b628aff711960c945.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 307, + 116, + 376, + 175 + ], + "blocks": [ + { + "bbox": [ + 307, + 116, + 376, + 175 + ], + "lines": [ + { + "bbox": [ + 307, + 116, + 376, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 116, + 376, + 175 + ], + "type": "image", + "image_path": "32abed312d109b59cdb0623658e92f9023bbf8b2a080975ff7108931d20043c5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 376, + 116, + 445, + 175 + ], + "blocks": [ + { + "bbox": [ + 376, + 116, + 445, + 175 + ], + "lines": [ + { + "bbox": [ + 376, + 116, + 445, + 175 + ], + "spans": [ + { + "bbox": [ + 376, + 116, + 445, + 175 + ], + "type": "image", + "image_path": "b4d530beb57dbeeeb266f98201b1250c1aab75660ec67d6f123556976a89cfe0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 168, + 176, + 236, + 274 + ], + "blocks": [ + { + "bbox": [ + 168, + 176, + 236, + 274 + ], + "lines": [ + { + "bbox": [ + 168, + 176, + 236, + 274 + ], + "spans": [ + { + "bbox": [ + 168, + 176, + 236, + 274 + ], + "type": "image", + "image_path": "fe6c64c5af7da18aa5399b7ab012a7fc0b88ee2541e107bf5f7ee0394f2e4140.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 284, + 482, + 317 + ], + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 317 + ], + "type": "text", + "content": "Fig. 4: The object detection results on the COCO dataset. The first two columns compare the effect of maximum integer value " + }, + { + "bbox": [ + 130, + 284, + 482, + 317 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 284, + 482, + 317 + ], + "type": "text", + "content": " on performance for the same structure. The second and third columns compare the effect of the size of the model on performance." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 238, + 176, + 306, + 274 + ], + "blocks": [ + { + "bbox": [ + 238, + 176, + 306, + 274 + ], + "lines": [ + { + "bbox": [ + 238, + 176, + 306, + 274 + ], + "spans": [ + { + "bbox": [ + 238, + 176, + 306, + 274 + ], + "type": "image", + "image_path": "57972062b0a46e1512e7b6aa87dcb4123b92292af0b8e911b4aeb21f685dfc65.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 176, + 375, + 274 + ], + "blocks": [ + { + "bbox": [ + 307, + 176, + 375, + 274 + ], + "lines": [ + { + "bbox": [ + 307, + 176, + 375, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 176, + 375, + 274 + ], + "type": "image", + "image_path": "4b0ec6a5b167585a656232354c002b1eb31a37f80a9333b9f1174b747abf5357.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 376, + 176, + 445, + 274 + ], + "blocks": [ + { + "bbox": [ + 376, + 176, + 445, + 274 + ], + "lines": [ + { + "bbox": [ + 376, + 176, + 445, + 274 + ], + "spans": [ + { + "bbox": [ + 376, + 176, + 445, + 274 + ], + "type": "image", + "image_path": "b4f3de35749ccced1ac59aec97410af3c58f9a76913687f1bf3214afe8eb50c9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 342, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 437 + ], + "type": "text", + "content": "Experimental Setup. As a large neuromorphic object detection dataset, Gen1 [8] encompasses 39 hours of open road and various driving scenarios, captured using an ATIS sensor with a resolution of " + }, + { + "bbox": [ + 130, + 342, + 482, + 437 + ], + "type": "inline_equation", + "content": "304 \\times 240" + }, + { + "bbox": [ + 130, + 342, + 482, + 437 + ], + "type": "text", + "content": " pixels. The dataset is organized into training, validation, and testing subsets. The bounding box annotations of pedestrians and cars (over 255,000) were manually labeled. For each annotation, we process the event-based stream 2.5 seconds before its occurrence, dividing it into " + }, + { + "bbox": [ + 130, + 342, + 482, + 437 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 342, + 482, + 437 + ], + "type": "text", + "content": " slices for model input. We train the model for 50 epochs and maintain other hyperparameters same as the COCO 2017 dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "text", + "content": "Main Results on Gen1 dataset are shown in Table 3. The proposed SpikeYOLO notably elevates the performance benchmark for the Gen1 dataset in SNNs. We achieve " + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "inline_equation", + "content": "67.2\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "text", + "content": " mAP@50 with 23.1M parameters, which outperforms the prior state-of-the-art SNN model by " + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "inline_equation", + "content": "+8.2\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "text", + "content": ". For example, when " + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "inline_equation", + "content": "T = 5" + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "text", + "content": ", SpikeYOLO vs. EMS-YOLO [50]: Param, 13.2M vs. 14.4M; mAP@50, " + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "inline_equation", + "content": "66.0\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "inline_equation", + "content": "59.0\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "text", + "content": "; mAP@50:95, " + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "inline_equation", + "content": "38.5\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "inline_equation", + "content": "31.0\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "text", + "content": ". In contrast to the COCO dataset, Gen1 contains temporal information that is more suitable for SNN processing. We conduct experiments on the performance of SNN and ANN with the same architecture. We observe that SpikeYOLO's mAP@50 accuracy is " + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "inline_equation", + "content": "+2.5\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "text", + "content": " higher than the corresponding ANN, and shows a " + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "inline_equation", + "content": "5.7\\times" + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "text", + "content": " energy efficiency. This indicates that SNN has attractive potential in processing neuromorphic data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Ablation Studies of Quantization Error. Both " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " significantly influence outcomes when processing neuromorphic datasets. Table 4 gives a comprehensive ablation study on SpikeYOLO with 23.1M parameters that evaluate the effects of varying " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". We observe some interesting experimental results. First, boosting the timestep " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " will bring improvement in accuracy and power. For instance, with the set of " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "D = 1" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", extending " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "T = 1" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "T = 4" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " yields a " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "+6.7\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " increase in mAP@50, and the power will increase by " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "3.7\\times" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". But further extending " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "T = 4" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "T = 8" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " results in a marginal increase of only " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "+0.7\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and significantly" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "text", + "content": "Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 145, + 146, + 470, + 324 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 3: Results on the Gen1 dataset [8]. * We convert 23.1M of SpikeYOLO into ANN with the same architecture." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 145, + 146, + 470, + 324 + ], + "lines": [ + { + "bbox": [ + 145, + 146, + 470, + 324 + ], + "spans": [ + { + "bbox": [ + 145, + 146, + 470, + 324 + ], + "type": "table", + "html": "
ArchitectureModelParam (M)Power (mJ)T × DmAP@ 50(%)mAP@ 50:95(%)
ANNYOLOv3-tiny [44]10.25.1144.5-
SpikeYOLO*23.178.5164.739.7
SNNEMS-YOLO [50]6.21.2554.726.7
9.32.0556.528.6
14.43.4559.031.0
VGG-11+SDD12.611.11-17.4
MobileNet-64+SSD24.35.71-14.7
DenseNet121-24+SSD [6]8.23.91-18.9
Spiking-Yolo [29]7.9102.350044.2-
Tr-Spiking-Yolo [65]7.90.9545.3-
SpikeYOLO(Ours)13.211.05 × 166.038.5
23.119.75 × 166.438.9
23.112.94 × 267.240.4
", + "image_path": "85b2871f573a0247eb1d9983e51a4f8d3f2922d4f0f96eaf246b264ad82c8955.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 179, + 367, + 435, + 480 + ], + "blocks": [ + { + "bbox": [ + 131, + 335, + 482, + 359 + ], + "lines": [ + { + "bbox": [ + 131, + 335, + 482, + 359 + ], + "spans": [ + { + "bbox": [ + 131, + 335, + 482, + 359 + ], + "type": "text", + "content": "Table 4: The influence of " + }, + { + "bbox": [ + 131, + 335, + 482, + 359 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 335, + 482, + 359 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 335, + 482, + 359 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 131, + 335, + 482, + 359 + ], + "type": "text", + "content": " on Gen1. We set SpikeYOLO (23.1M) as the baseline and vary " + }, + { + "bbox": [ + 131, + 335, + 482, + 359 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 335, + 482, + 359 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 335, + 482, + 359 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 131, + 335, + 482, + 359 + ], + "type": "text", + "content": " for each study." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 179, + 367, + 435, + 480 + ], + "lines": [ + { + "bbox": [ + 179, + 367, + 435, + 480 + ], + "spans": [ + { + "bbox": [ + 179, + 367, + 435, + 480 + ], + "type": "table", + "html": "
MethodT × DPower(mJ)mAP@50(%)mAP@50:95(%)
SpikeYOLO1 × 14.059.333.1
1 × 43.9 (-0.1)65.1 (+5.8)38.9 (+5.8)
2 × 18.163.636.5
2 × 27.8 (-0.3)66.1 (+2.5)39.0 (+2.5)
2 × 47.1 (-1.0)67.0 (+3.4)40.1 (+3.6)
4 × 114.866.038.4
4 × 212.9 (-1.9)67.2 (+1.2)40.4 (+2.0)
8 × 127.066.739.3
", + "image_path": "0079e7d018e7539d81ef4a45d9786cabc5cf8afaf89c001478378d9c769445a9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "text", + "content": "increases energy cost. Second, We were surprised to see that by expanding " + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "text", + "content": " at a fixed " + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "text", + "content": ", the performance will be improved, while the power will be dropped. For example, " + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "inline_equation", + "content": "2 \\times 1" + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "inline_equation", + "content": "2 \\times 4" + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "text", + "content": ": mAP@50, " + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "inline_equation", + "content": "63.6\\%" + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "inline_equation", + "content": "66.1\\%" + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "inline_equation", + "content": "67.0\\%" + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "text", + "content": "; Power, " + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "inline_equation", + "content": "8.1\\mathrm{mJ}" + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "inline_equation", + "content": "7.8\\mathrm{mJ}" + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "inline_equation", + "content": "7.1\\mathrm{mJ}" + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "text", + "content": ". This trend is completely different from SpikeYOLO's performance in COCO, where the increase of " + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 503, + 482, + 588 + ], + "type": "text", + "content": " will bring more energy cost. We argue that this phenomenon is because SNN exhibits various spike firing for dense/sparse data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 607, + 339, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 607, + 339, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 607, + 339, + 620 + ], + "type": "text", + "content": "4.3 Architecture Ablation Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Re-parameterization Design. As shown in Table 5, if we remove re-parameterization by adding neurons into inverted separable convolutions, the mAP@50 and mAP@50:95 will decrease " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 397, + 91, + 447, + 102 + ], + "type": "text", + "content": "SpikeYOLO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 144, + 146, + 470, + 227 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 138 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 138 + ], + "type": "text", + "content": "Table 5: Ablation studies of architecture design. We set " + }, + { + "bbox": [ + 130, + 114, + 482, + 138 + ], + "type": "inline_equation", + "content": "T \\times D = 1 \\times 4" + }, + { + "bbox": [ + 130, + 114, + 482, + 138 + ], + "type": "text", + "content": " and modify just one point of baseline to test how the parameters, power and performance vary." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 146, + 470, + 227 + ], + "lines": [ + { + "bbox": [ + 144, + 146, + 470, + 227 + ], + "spans": [ + { + "bbox": [ + 144, + 146, + 470, + 227 + ], + "type": "table", + "html": "
MethodParam(M)mAP@50(%)mAP@50:95(%)
SpikeYOLO(Baseline)23.162.345.5
Remove re-parameterization23.160.643.7
SNN-Block-1 → SNN-Block-219.961.244.7
SNN-Block-2 → Transformer Block24.561.044.2
Anchor-free head → Anchor-based head21.259.539.7
", + "image_path": "c0e0124565300da2d5435a82229de818368d500729735e40abc8051754823caa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 248, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 332 + ], + "type": "text", + "content": "SNN Block Design. Including a " + }, + { + "bbox": [ + 130, + 248, + 482, + 332 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 248, + 482, + 332 + ], + "type": "text", + "content": " standard convolution within the initial stages of convolution blocks is crucial. As shown in Table 5, substituting SNN-Block-1 for SNN-Block-2 leads to a reduction in performance of around " + }, + { + "bbox": [ + 130, + 248, + 482, + 332 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 248, + 482, + 332 + ], + "type": "text", + "content": ". Moreover, we try to replace high-stage SNN-Block-2 with meta Transformer-based SNN block, just like Meta-SpikeFormer [58]. We find that there is little to no performance gain by doing this and that the parameters increase. Therefore, only spiking CNN blocks are exploited in our SpikeYOLO." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 332, + 483, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 483, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 483, + 405 + ], + "type": "text", + "content": "Detection Head. The detection mechanisms within YOLO are categorized into anchor-based heads(e.g.,YOLOv5) and anchor-free heads(e.g.,YOLOv8). The former directly predicts each bounding box's dimensions, whereas the latter estimates the probability distribution of each bounding box. Previous EMSYOLO [50] and Meta-SpikeFormer [58] employ anchor-based heads. SpikeYOLO exploits the anchor-free head because of its higher accuracy (see Table 5)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 422, + 220, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 422, + 220, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 422, + 220, + 435 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 448, + 483, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 483, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 483, + 616 + ], + "type": "text", + "content": "This work significantly narrows the performance gap between SNNs and ANNs on object detection tasks. We achieve this through network architecture and spiking neuron design. The proposed SpikeYOLO architecture abandons the complex module design in the vanilla YOLO series and exploits simple meta spike blocks to build the model. Then, the I-LIF spiking neuron capable of integer-valued training and spike-driven inference is proposed to drop quantization errors. We improve the upper bound of the SNN domain's performance on the COCO dataset by " + }, + { + "bbox": [ + 130, + 448, + 483, + 616 + ], + "type": "inline_equation", + "content": "+15.0\\%" + }, + { + "bbox": [ + 130, + 448, + 483, + 616 + ], + "type": "text", + "content": " (mAP@50) and " + }, + { + "bbox": [ + 130, + 448, + 483, + 616 + ], + "type": "inline_equation", + "content": "+18.7\\%" + }, + { + "bbox": [ + 130, + 448, + 483, + 616 + ], + "type": "text", + "content": " (mAP@50:95), respectively. On the neuromorphic Gen1 dataset, SpikeYOLO achieves better performance and lower power than ANN of the same architecture. Furthermore, we investigate the performance of equivalent architecture ANNs and SNNs in different datasets, and the results show that the redesigned SNN architecture performed better. This work enables SNNs to handle complex object detection and can inspire the application of SNNs in more visual scenarios." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "text", + "content": "Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 485, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 485, + 210 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 485, + 210 + ], + "type": "text", + "content": "This work was partially supported by National Distinguished Young Scholars (62325603), and National Natural Science Foundation of China (62236009, U22A20103,62441606), Beijing Natural Science Foundation for Distinguished Young Scholars (JQ21015), China Postdoctoral Science Foundation (GZB20240824, 2024M753497), and CAAI-MindSpore Open Fund, developed on OpenI Community." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 228, + 197, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 228, + 197, + 240 + ], + "spans": [ + { + "bbox": [ + 133, + 228, + 197, + 240 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 251, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 251, + 480, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 273 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 273 + ], + "type": "text", + "content": "1. Bochkovskiy, A., Wang, C.Y., Liao, H.Y.M.: Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 274, + 481, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 481, + 316 + ], + "type": "text", + "content": "2. Bu, T., Fang, W., Ding, J., DAI, P., Yu, Z., Huang, T.: Optimal ANN-SNN conversion for high-accuracy and ultra-low-latency spiking neural networks. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=7B3IJMM1k_M" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 318, + 481, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 318, + 481, + 349 + ], + "spans": [ + { + "bbox": [ + 138, + 318, + 481, + 349 + ], + "type": "text", + "content": "3. Cao, Y., Chen, Y., Khosla, D.: Spiking deep convolutional neural networks for energy-efficient object recognition. International Journal of Computer Vision 113, 54-66 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 350, + 481, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 481, + 381 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 481, + 381 + ], + "type": "text", + "content": "4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European Conference on Computer Vision. pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 383, + 481, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 383, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 138, + 383, + 481, + 415 + ], + "type": "text", + "content": "5. Chollet, F.: Xception: Deep learning with depthwise separable convolutions. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1800-1807 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 416, + 481, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 416, + 481, + 447 + ], + "spans": [ + { + "bbox": [ + 138, + 416, + 481, + 447 + ], + "type": "text", + "content": "6. Cordone, L., Miramond, B., Thierion, P.: Object detection with spiking neural networks on automotive event data. In: 2022 International Joint Conference on Neural Networks. pp. 1-8. IEEE (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 449, + 481, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 481, + 480 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 481, + 480 + ], + "type": "text", + "content": "7. Davies, M., Srinivasa, N., Lin, T.H., Chinya, G., Cao, Y., Choday, S.H., Dimou, G., Joshi, P., Imam, N., Jain, S., et al.: Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro 38(1), 82-99 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 481, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 481, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 481, + 481, + 512 + ], + "type": "text", + "content": "8. De Tournemire, P., Nitti, D., Perot, E., Migliore, D., Sironi, A.: A large scale event-based detection dataset for automotive. arXiv preprint arXiv:2001.08499 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 514, + 481, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 481, + 545 + ], + "type": "text", + "content": "9. Deng, S., Li, Y., Zhang, S., Gu, S.: Temporal efficient training of spiking neural network via gradient re-weighting. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=_XNtisL32jv" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 546, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 546, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 134, + 546, + 481, + 578 + ], + "type": "text", + "content": "10. Diehl, P.U., Neil, D., Binas, J., Cook, M., Liu, S.C., Pfeiffer, M.: Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In: 2015 International Joint Conference on Neural Networks. pp. 1-8. IEEE (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 134, + 578, + 481, + 610 + ], + "type": "text", + "content": "11. Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 13733-13742 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 611, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 611, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 611, + 481, + 665 + ], + "type": "text", + "content": "12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=YicbFdNTTy" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 397, + 91, + 447, + 102 + ], + "type": "text", + "content": "SpikeYOLO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "13. Eshraghian, J.K., Ward, M., Neftci, E.O., Wang, X., Lenz, G., Dwivedi, G., Bennamoun, M., Jeong, D.S., Lu, W.D.: Training spiking neural networks using lessons from deep learning. Proceedings of the IEEE 111(9), 1016-1054 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "text", + "content": "14. Fang, W., Yu, Z., Chen, Y., Huang, T., Masquelier, T., Tian, Y.: Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems 34, 21056-21069 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 481, + 225 + ], + "type": "text", + "content": "15. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(1), 154-180 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 225, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 225, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 225, + 481, + 246 + ], + "type": "text", + "content": "16. Girshick, R.: Fast r-cnn. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1440-1448 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 246, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 246, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 133, + 246, + 481, + 278 + ], + "type": "text", + "content": "17. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 580-587 (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 278, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 278, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 133, + 278, + 481, + 310 + ], + "type": "text", + "content": "18. Guo, Y., Chen, Y., Liu, X., Peng, W., Zhang, Y., Huang, X., Ma, Z.: Ternary spike: Learning ternary spikes for spiking neural networks. arXiv preprint arXiv:2312.06372 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 310, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 310, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 133, + 310, + 481, + 342 + ], + "type": "text", + "content": "19. Guo, Y., Chen, Y., Zhang, L., Liu, X., Wang, Y., Huang, X., Ma, Z.: Im-loss: information maximization loss for spiking neural networks. Advances in Neural Information Processing Systems 35, 156-166 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 342, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 342, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 133, + 342, + 481, + 386 + ], + "type": "text", + "content": "20. Guo, Y., Liu, X., Chen, Y., Zhang, L., Peng, W., Zhang, Y., Huang, X., Ma, Z.: Rmp-loss: Regularizing membrane potential distribution for spiking neural networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17391-17401 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 386, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 386, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 133, + 386, + 481, + 429 + ], + "type": "text", + "content": "21. Han, M., Wang, Q., Zhang, T., Wang, Y., Zhang, D., Xu, B.: Complex dynamic neurons improved spiking transformer network for efficient automatic speech recognition. Thirty-Seventh AAAI Conference on Artificial Intelligence (AAAI 2023) (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 429, + 481, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 429, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 133, + 429, + 481, + 461 + ], + "type": "text", + "content": "22. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 461, + 481, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 461, + 481, + 494 + ], + "spans": [ + { + "bbox": [ + 133, + 461, + 481, + 494 + ], + "type": "text", + "content": "23. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) Computer Vision - ECCV 2016. pp. 630-645. Springer International Publishing, Cham (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 494, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 494, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 133, + 494, + 481, + 525 + ], + "type": "text", + "content": "24. Horowitz, M.: 1.1 computing's energy problem (and what we can do about it). In: 2014 IEEE International Solid-State Sircuits Conference digest of technical papers. pp. 10-14. IEEE (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 525, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 525, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 133, + 525, + 481, + 568 + ], + "type": "text", + "content": " 25. Hu, J., Yao, M., Qiu, X., Chou, Y., Cai, Y., Qiao, N., Tian, Y., XU, B., Li, G.: High-performance temporal reversible spiking neural networks with $\\mathcal{O}\\{0\\}(1)$ training memory and $\\mathcal{M} \\mathcal{H} \\mathcal{C} \\mathcal{A}\\{0\\}(1)\\$ inference cost. In: Forty-first International Conference on Machine Learning (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 133, + 568, + 481, + 601 + ], + "type": "text", + "content": "26. Hu, Y., Zheng, Q., Jiang, X., Pan, G.: Fast-snn: Fast spiking neural network by converting quantized ann. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(12), 14546-14562 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 133, + 601, + 481, + 633 + ], + "type": "text", + "content": "27. Hu, Y., Deng, L., Wu, Y., Yao, M., Li, G.: Advancing spiking neural networks toward deep residual learning. IEEE Transactions on Neural Networks and Learning Systems pp. 1-15 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 481, + 666 + ], + "type": "text", + "content": "28. Kim, S., Park, S., Na, B., Kim, J., Yoon, S.: Towards fast and accurate object detection in bio-inspired spiking neural networks through bayesian optimization. IEEE Access 9, 2633-2643 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "text", + "content": "Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "29. Kim, S., Park, S., Na, B., Yoon, S.: Spiking-yolo: spiking neural network for energy-efficient object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11270-11277 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 192 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 192 + ], + "type": "text", + "content": "30. Kim, Y., Park, H., Moitra, A., Bhattacharjee, A., Venkatesha, Y., Panda, P.: Rate coding or direct coding: Which one is better for accurate, robust, and energy-efficient spiking neural networks? In: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 71-75. IEEE (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 192, + 482, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 192, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 482, + 225 + ], + "type": "text", + "content": "31. Li, C., Jones, E.G., Furber, S.: Unleashing the potential of spiking neural networks with dynamic confidence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13304-13314 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 225, + 482, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 257 + ], + "type": "text", + "content": "32. Li, Y., He, X., Dong, Y., Kong, Q., Zeng, Y.: Spike calibration: Fast and accurate conversion of spiking neural network for object detection and segmentation. arXiv preprint arXiv:2207.02702 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 257, + 482, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 278 + ], + "type": "text", + "content": "33. Li, Y., Geller, T., Kim, Y., Panda, P.: Seenn: Towards temporal spiking early exit neural networks. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 278, + 482, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 311 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 311 + ], + "type": "text", + "content": "34. Lin, T.Y., Dólár, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2117-2125 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 311, + 482, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 354 + ], + "type": "text", + "content": "35. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 354, + 482, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 374 + ], + "type": "text", + "content": "36. Maass, W.: Networks of spiking neurons: the third generation of neural network models. Neural Networks 10(9), 1659-1671 (1997)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 374, + 482, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 418 + ], + "type": "text", + "content": "37. Merolla, P.A., Arthur, J.V., Alvarez-Icaza, R., Cassidy, A.S., Sawada, J., Akopyan, F., Jackson, B.L., Imam, N., Guo, C., Nakamura, Y., et al.: A million spiking-neuron integrated circuit with a scalable communication network and interface. Science 345(6197), 668-673 (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 418, + 482, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 450 + ], + "type": "text", + "content": "38. Mueller, E., Studenyak, V., Auge, D., Knoll, A.: Spiking transformer networks: A rate coded approach for processing sequential data. In: 2021 7th International Conference on Systems and Informatics (ICSAI). pp. 1-5. IEEE (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 450, + 482, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 483 + ], + "type": "text", + "content": "39. Neftci, E.O., Mostafa, H., Zenke, F.: Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine 36(6), 51-63 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 483, + 482, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 515 + ], + "type": "text", + "content": "40. Panda, P., Aketi, S.A., Roy, K.: Toward scalable, efficient, and accurate deep spiking neural networks with backward residual connections, stochastic softmax, and hybridization. Frontiers in Neuroscience 14, 653 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 515, + 482, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 547 + ], + "type": "text", + "content": "41. Pei, J., Deng, L., Song, S., Zhao, M., Zhang, Y., Wu, S., Wang, G., Zou, Z., Wu, Z., He, W., et al.: Towards artificial general intelligence with hybrid tianjic chip architecture. Nature 572(7767), 106-111 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 547, + 482, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 547, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 547, + 482, + 579 + ], + "type": "text", + "content": "42. Qiu, X., Zhu, R.J., Chou, Y., Wang, Z., Deng, L.j., Li, G.: Gated attention coding for training high-performance and efficient spiking neural networks. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 601-610 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 579, + 482, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 482, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 482, + 611 + ], + "type": "text", + "content": "43. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 779-788 (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 611, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 482, + 632 + ], + "type": "text", + "content": "44. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "45. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems 28 (2015)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 397, + 91, + 447, + 102 + ], + "type": "text", + "content": "SpikeYOLO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "46. Roy, K., Jaiswal, A., Panda, P.: Towards spike-based machine intelligence with neuromorphic computing. Nature 575(7784), 607-617 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "text", + "content": "47. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: Mobilenetv2: Inverted residuals and linear bottlenecks. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 4510-4520 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "text", + "content": "48. Schuman, C.D., Kulkarni, S.R., Parsa, M., Mitchell, J.P., Kay, B., et al.: Opportunities for neuromorphic computing algorithms and applications. Nature Computational Science 2(1), 10-19 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "49. Sengupta, A., Ye, Y., Wang, R., Liu, C., Roy, K.: Going deeper in spiking neural networks: Vgg and residual architectures. Frontiers in Neuroscience 13, 95 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "type": "text", + "content": "50. Su, Q., Chou, Y., Hu, Y., Li, J., Mei, S., Zhang, Z., Li, G.: Deep directly-trained spiking neural networks for object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6555-6565 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 260, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 480, + 293 + ], + "type": "text", + "content": "51. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems. pp. 5998-6008 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "type": "text", + "content": "52. Wang, C.Y., Bochkovskiy, A., Liao, H.Y.M.: Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7464-7475 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 369 + ], + "type": "text", + "content": "53. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 568-578 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 370, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 480, + 403 + ], + "type": "text", + "content": "54. Wang, Z., Fang, Y., Cao, J., Zhang, Q., Wang, Z., Xu, R.: Masked spiking transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1761-1771 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "55. Wu, Y., Deng, L., Li, G., Zhu, J., Shi, L.: Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in Neuroscience 12 331 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "56. Wu, Y., Deng, L., Li, G., Zhu, J., Xie, Y., Shi, L.: Direct training for spiking neural networks: Faster, larger, better. In: Proceedings of the AAAI conference on artificial intelligence. vol. 33, pp. 1311-1318 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 512 + ], + "type": "text", + "content": "57. Yao, M., Gao, H., Zhao, G., Wang, D., Lin, Y., Yang, Z., Li, G.: Temporal-wise attention spiking neural networks for event streams classification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10221-10230 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 555 + ], + "type": "text", + "content": "58. Yao, M., Hu, J., Hu, T., Xu, Y., Zhou, Z., Tian, Y., XU, B., Li, G.: Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=1SIBN5Xyw7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "text", + "content": "59. Yao, M., Hu, J., Zhao, G., Wang, Y., Zhang, Z., Xu, B., Li, G.: Inherent redundancy in spiking neural networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16924-16934 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 589, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 611 + ], + "type": "text", + "content": "60. Yao, M., Hu, J., Zhou, Z., Yuan, L., Tian, Y., Xu, B., Li, G.: Spike-driven transformer. Advances in neural information processing systems 36 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 611, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 665 + ], + "type": "text", + "content": "61. Yao, M., Richter, O., Zhao, G., Qiao, N., Xing, Y., Wang, D., Hu, T., Fang, W., Demirci, T., De Marchi, M., Deng, L., Yan, T., Nielsen, C., Sheik, S., Wu, C., Tian, Y., Xu, B., Li, G.: Spike-based dynamic computing with asynchronous sensing-computing neuromorphic chip. Nature Communications 15(1), 4464 (May 2024)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "text", + "content": "Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 436 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "62. Yao, M., Zhang, H., Zhao, G., Zhang, X., Wang, D., Cao, G., Li, G.: Sparser spiking activity can be better: Feature refine-and-mask spiking neural network for event-based visual recognition. Neural Networks 166, 410-423 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "63. Yao, M., Zhao, G., Zhang, H., Hu, Y., Deng, L., Tian, Y., Xu, B., Li, G.: Attention spiking neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(8), 9393-9410 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "text", + "content": "64. Yin, B., Corradi, F., Bohte, S.M.: Accurate and efficient time-domain classification with adaptive spiking recurrent neural networks. Nature Machine Intelligence 3(10), 905-913 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 247 + ], + "type": "text", + "content": "65. Yuan, M., Zhang, C., Wang, Z., Liu, H., Pan, G., Tang, H.: Trainable spiking-yolo for low-latency and high-performance object detection. Neural Networks 172, 106092 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 249, + 482, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 282 + ], + "type": "text", + "content": "66. Zhang, J., Dong, B., Zhang, H., Ding, J., Heide, F., Yin, B., Yang, X.: Spiking transformers for event-based single object tracking. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 8801-8810 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 282, + 482, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 326 + ], + "type": "text", + "content": "67. Zhang, J., Tang, L., Yu, Z., Lu, J., Huang, T.: Spike transformer: Monocular depth estimation for spiking camera. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII. pp. 34-52. Springer (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 326, + 482, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 358 + ], + "type": "text", + "content": "68. Zheng, H., Wu, Y., Deng, L., Hu, Y., Li, G.: Going deeper with directly-trained larger spiking neural networks. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 11062-11070 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 358, + 482, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 401 + ], + "type": "text", + "content": "69. Zhou, Z., Zhu, Y., He, C., Wang, Y., YAN, S., Tian, Y., Yuan, L.: Spikformer: When spiking neural network meets transformer. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=frE4fUwz_h" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 402, + 482, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 436 + ], + "type": "text", + "content": "70. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable " + }, + { + "bbox": [ + 130, + 402, + 482, + 436 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{detr}\\}" + }, + { + "bbox": [ + 130, + 402, + 482, + 436 + ], + "type": "text", + "content": ": Deformable transformers for end-to-end object detection. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=gZ9hCDWe6ke" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 397, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 397, + 91, + 447, + 102 + ], + "type": "text", + "content": "SpikeYOLO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization/bb833fed-6970-4db4-b3da-646e63e5493d_content_list.json b/2024/Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization/bb833fed-6970-4db4-b3da-646e63e5493d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2a251c4a65fa4d66bdd8d4a8b3bd3135aba08a25 --- /dev/null +++ b/2024/Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization/bb833fed-6970-4db4-b3da-646e63e5493d_content_list.json @@ -0,0 +1,1687 @@ +[ + { + "type": "text", + "text": "Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization", + "text_level": 1, + "bbox": [ + 251, + 140, + 751, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Naiyu Yin $^{1}$ , Hanjing Wang $^{1}$ , Yue Yu $^{2}$ , Tian Gao $^{3}$ , Amit Dhurandhar $^{3}$ , and Qiang Ji $^{1}$", + "bbox": [ + 217, + 233, + 782, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Rensselaer Polytechnic Institute, Troy NY 12180, USA {yinn2, wangh36, jiq}@rpi.edu", + "bbox": [ + 308, + 276, + 691, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Lehigh University, Bethlehem PA 18015, USA {yuy214}@lehigh.edu", + "bbox": [ + 341, + 305, + 661, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 IBM Research, Yorktown Heights NJ 10598, USA {tgao, adhuran}@us.ibm.com", + "bbox": [ + 328, + 333, + 673, + 361 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Identifying low-dimensional, semantic latent causal representations for high-dimensional data has become a dynamic field in computer vision and machine learning. Causal domain generalization methods aim to identify latent causal variables that generate input data and build invariant causal mechanisms for prediction tasks, thereby improving out-of-distribution (OOD) prediction performance. However, there is no consensus on the best approach for selecting causal variables for prediction. Existing methods typically choose causal or anti-causal variables, excluding other invariant, discriminative features. In this paper, we propose using Markov Blanket features due to their property of being the minimal set that possesses the maximum mutual information with the target. To achieve this, we establish a Causal Markov Blanket Representation Learning (CMBRL) framework, which allows for Markov Blanket discovery in the latent space. We then construct an invariant prediction mechanism using the identified Markov Blanket features, making it suitable for predictions across domains. Compared to state-of-the-art domain generalization methods, our approach exhibits robustness and adaptability under distribution shifts.", + "bbox": [ + 261, + 398, + 738, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Causal Representation Learning $\\cdot$ Markov Blanket Discovery $\\cdot$ Domain Generalization", + "bbox": [ + 259, + 662, + 740, + 689 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 717, + 375, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modern deep-learning models have significantly improved performance across various machine learning and computer vision applications in recent decades. However, they are also known for their limitations, including poor generalization, a lack of interpretability, and fairness concerns. [35] pinpointed the primary cause of the poor generalization of modern deep-learning models stems from the spurious correlations between irrelevant features and the prediction", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "task through theoretical analysis and empirical evidence. These spurious correlations are caused by data biases and vary when distribution shifts, rendering the deep-learning models unreliable for the prediction of data from unseen domains. Hence, causal domain generalization methods propose to formulate prediction tasks from a causal perspective and exploit the robust, domain-invariant causality instead of correlations to improve the Out-of-distribution performance.", + "bbox": [ + 212, + 146, + 782, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent research in causal domain generalization [22,31-34,39,41,52] utilizes structural causal models (SCMs) to delve into the fundamental mechanisms underlying data generation. SCMs effectively capture intrinsic, stable, and interpretable causal relationships within a data distribution. These studies leverage SCMs to extract latent representations that are causally related to the target variable and subsequently develop invariant prediction mechanisms based on these representations. Key differences among existing methods stem from their choices of causally related features and their approaches to learning or selecting these features. However, there is no consensus on the optimal types of causally related features to be learned or selected. While many approaches [22,32-34,52] focus on identifying parent variables (causal features) relative to the target, [31] contends that anti-causal features (child variables) may offer greater robustness and predictive power, particularly in vision tasks. This paper theoretically examines the best causally related variables for prediction tasks and proposes an effective, potentially efficient method for identifying these variables.", + "bbox": [ + 212, + 239, + 787, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "According to [12], the Markov Blanket (MB) of a variable is theoretically proven to be optimal for prediction tasks, as it is the minimal set that contains the maximum information about the target. Consequently, MB features have been extensively used for feature selection across various prediction tasks [4,38, 47]. Recent research [55] explores the use of MB features in the latent space for domain generalization. We find that, within the framework of our proposed SCM, the influence of spurious features is effectively mitigated by conditioning on the MB features. Specifically, the MB set can block all paths between the domain variable $U$ and the target $Y$ , as illustrated in Figure 1. This characteristic makes MB features theoretically optimal for domain generalization prediction tasks.", + "bbox": [ + 212, + 470, + 787, + 623 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a framework for selecting Markov Blanket features from latent high-level variables that generate high-dimensional data. We refer to the chosen MB latent variables as Causal Markov Blanket (CMB) representations, which include the parents, children, and spouses of the target variable in the latent space. To obtain the CMB representations, we first obtain a set of identifiable latent variables using existing identifiable variational autoencoder framework. Then we propose an efficient MB discovery approach to identify the CMB representations collectively. A prediction model built using CMB representations is theoretically guaranteed to be both invariant and informative for predicting the target variable.", + "bbox": [ + 212, + 625, + 787, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this paper are as follows: 1) We introduce a novel and general SCM to elucidate the data generation mechanisms underlying prediction tasks. 2) We establish a framework for discovering Markov Blanket features in the latent space. 3) We propose a three-phase algorithm for do", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "N. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "main generalization prediction. We validate the effectiveness of our algorithm on benchmark datasets with distribution shifts, demonstrating improved generalization performance in the presence of these shifts.", + "bbox": [ + 218, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 218, + 213, + 395, + 228 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we will review previous studies focusing on two key areas: Markov Blanket discovery and domain generalization (DG) prediction.", + "bbox": [ + 218, + 242, + 782, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Markov Blanket Discovery and Feature Selection. The Markov Blanket set of a variable consists of its immediate neighbors (parents, children, and spouses) in a causal graph. Popular MB discovery methods, including KS [21], HITON-MB [5], IPCMB [10], IAMB [48], and STMB [11], identify the Markov Blanket set through a series of independence tests, making MB discovery performance dependent on the accuracy of these tests. In terms of applications, multiple works [4, 12, 38, 47] have utilized MB discovery in the input space as a feature selection strategy, employing MB features for prediction. However, directly applying these methods to high-dimensional data, such as images, videos, or texts, is computationally expensive. Moreover, as noted by [32], causal and spurious features may not be disentangled in the input space since each dimension of the input data could be influenced by both types of features.", + "bbox": [ + 218, + 285, + 784, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Domain Generalization Approaches. We aim to employ Causal Markov Blanket representations to enhance domain generalization performance. To this end, we compare our method both theoretically and empirically with existing domain generalization approaches. Specifically, in a domain generalization setting, we use data from one or multiple source domains to make predictions on unseen target domains. Recent domain adaptation works [22,26,46] that require target domain data and intervention targets during training are beyond the scope of this paper. Popular domain generalization approaches include disentangled representation learning [19, 28, 29, 42], data augmentation [33], \"mix-up\" strategies [14, 56, 57], adversarial training [50, 51], and meta-learning [24, 43]. These methods typically do not involve causality.", + "bbox": [ + 218, + 479, + 784, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Causal Domain Generalization Approaches. Compared to non-causal DG approaches, causal approaches make assumptions about data generation via SCMs and derive their algorithms accordingly. Generally, causal approaches can be categorized by their use of intervention. Methods without interventions include stable representation learning [7,8,17,18] and invariant feature learning [1-3, 6,23,40]. Stable representation learning methods acquire causal or anti-causal features through strategies like covariate balancing or by using SCM as a form of regularization. Invariant feature learning methods aim to identify domain-invariant causal features from multi-environmental data based on certain invariance criteria derived from causal theorems. One popular invariant feature learning approach is invariant risk minimization (IRM), which identifies invariant predictors corresponding to the parent variables of the target variable. In", + "bbox": [ + 218, + 657, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Causal Markov Blanket Representation Learning", + "bbox": [ + 406, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "particular, [55] formulates an invariant causal Markov Blanket representation learning framework that is trained with causal constraints derived from the proposed SCM. Recently, a line of work [19, 20, 32] employs identifiable variational autoencoders (iVAEs) to obtain a set of latent variables with various degrees of identifiable guarantees, using the parent variables within this set for domain generalization prediction. We follow this approach and propose to learn the MB set of the target variable from the identifiable latent representations. Methods with interventions include robust feature learning through data augmentation [33] and interventional inference-guided methods [27, 34, 52]. These methods primarily address latent confounding issues between input data and the target. While our method generalizes across domains with different biases, it cannot yet effectively handle spurious correlations resulting from latent confounders. Extending our algorithm to handle latent confounders is an interesting future direction but is beyond the scope of this paper.", + "bbox": [ + 212, + 146, + 787, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Causal Analysis for Prediction Tasks", + "text_level": 1, + "bbox": [ + 214, + 388, + 607, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminary on Markov Blanket Features", + "text_level": 1, + "bbox": [ + 214, + 426, + 602, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We outline the definition of Markov Blanket in Definition 1.", + "bbox": [ + 214, + 460, + 661, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 1. ([37]) A Markov Blanket of a target variable $T$ within the variable set $\\mathbf{V}$ , $\\mathbf{MB}_T$ , is the minimal set of nodes conditioned on which all other nodes are independent of $T$ , denoted as $O \\perp T|\\mathbf{MB}_T, \\forall O \\in \\{\\mathbf{V} \\backslash T \\backslash \\mathbf{MB}_T\\}$ .", + "bbox": [ + 214, + 489, + 787, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The $\\mathbf{MB}_T$ consists of the parent, child, and spouse variables of the target variable $T$ . According to the Theorem 1-3 in [12], the MB features of the target possess the following properties: they constitute the minimal set of features holding the maximal information about the target variable, and they ensure the least Bayes errors when predicting the target.", + "bbox": [ + 212, + 547, + 787, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike prior works that use MB discovery for feature selection in the input space, we focus on MB selection in the latent space, guided by a structural causal model that we introduce for modeling the data generation process. To distinguish between them, we denote the MB set in the latent representation space as Causal Markov Blanket (CMB) representations. CMB representations are crucial for identifying the informative variables for predicting targets with higher accuracy despite distribution shifts.", + "bbox": [ + 212, + 625, + 787, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Assumptions on Structural Causal Model", + "text_level": 1, + "bbox": [ + 214, + 760, + 604, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We formulate the prediction task under distribution shifts from a causal perspective. Using a structural causal model, we discern the underlying causal mechanisms of the data generation process, as illustrated in Figure 1.", + "bbox": [ + 212, + 792, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "N. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\pmb{X} \\in \\mathbb{R}^{d}$ represents the inputs and $Y$ the prediction target. We denote $\\pmb{Z} = \\{Z_{1}, Z_{2}, \\dots, Z_{N}\\} \\in \\mathbb{R}^{N}$ as the latent factors for generating input $\\pmb{X}$ . To better account for the distribution shifts, we introduce a domain variable $U$ to encode domain-specific information4. Domain variable $U$ is the common cause for some latent factors in $\\pmb{Z}$ and hence can change the appearance of input $\\pmb{X}$ . Judged by their relations to the $Y$ , $\\pmb{Z}$ can be further categorized into four types: parent variables $\\pmb{Z}_{p} = \\{Z_{p_{1}}, Z_{p_{2}}, \\dots\\}$ , child variables $\\pmb{Z}_{c} = \\{Z_{c_{1}}, Z_{c_{2}}, \\dots\\}$ , spouse variables $\\pmb{Z}_{s} = \\{Z_{s_{1}}, Z_{s_{2}}, \\dots\\}$ , and spurious variables $\\pmb{Z}_{o} = \\{Z_{o_{1}}, Z_{o_{2}}, \\dots\\}$ . $\\pmb{Z}_{p}, \\pmb{Z}_{c}$ and $\\pmb{Z}_{s}$ are the direct causes, direct effects, and the causes of the direct effects. They collectively form the", + "bbox": [ + 215, + 145, + 540, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CMB representations $Z_{cmb}$ . $Z_o$ are the variables that are spuriously correlated to target $Y$ via $Z_{cmb}$ variables or domain variable $U$ . To generalize our SCM, we allow for arbitrary causal relations within $Z$ as long as the causal graph over $\\{X, Y, Z, U\\}$ is a directed acyclic graph (DAG). We summarize the detailed assumptions in Assumption 1 of Appendix A.1. We show that our assumptions is practical and generally hold in real-world applications in Appendix A.2. Moreover, the SCM is Figure 1, combined with the assumptions, covers most scenarios from prior works [3, 32] and hence is a flexible model for performing causal analysis on prediction tasks. The semantic meanings of $Z_o$ and $Z_{cmb}$ are context-dependent and vary with the data. For instance, in CMNIST, $Z_o$ represents background information like color, whereas $Z_{cmb}$ denotes foreground information such as digit shape.", + "bbox": [ + 212, + 402, + 787, + 584 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2f950604a24249e491a85377114c2a8ce5377d99e34a25e2c679f69e768557aa.jpg", + "image_caption": [ + "Fig.1: Proposed SCM over $\\{X,U,Z,Y\\}$ . Undirected links between $Z_{o}, Z_{p}$ and $Z_{s}$ indicate that both edge directions are permissible without violating the DAG constraint." + ], + "image_footnote": [], + "bbox": [ + 550, + 147, + 785, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Invariant Predictive Mechanism", + "text_level": 1, + "bbox": [ + 215, + 613, + 526, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The conditional distribution $p(Y|\\mathbf{X})$ , typically captured by a traditional classifier, is influenced by the domain variable $U$ , meaning that $U \\not\\perp Y|\\mathbf{X}$ . Consequently, the domain-variant distribution $p(Y|\\mathbf{X})$ is unsuitable for prediction in unseen domains. Prior causal approaches such as [22, 33, 34] and other content-style approaches make strong assumptions, assuming there is only a single type of causal feature (usually parent variables $Z_{p}$ ) within the SCM. Due to the simplicity of their SCMs, these methods fail when encountering scenarios where other types of causal features exist, as their selection of causal features cannot block all paths from $U$ to $Y$ . For example, if the SCM only considers the existence of parent variables $Z_{p}$ while $Z_{c}$ also exists in practice, the domain variable $U$ can reach $Y$ via $U \\to Z_{o} \\to X \\to Z_{c} \\gets Y$ when $X$ is given.", + "bbox": [ + 212, + 643, + 787, + 811 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Causal Markov Blanket Representation Learning", + "bbox": [ + 405, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "4 This is a standard setting in previous works [32, 39]", + "bbox": [ + 217, + 824, + 576, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We advocate for using the Causal Markov Blanket features, denoted as $\\mathbf{Z}_{cmb} = \\{\\mathbf{Z}_p,\\mathbf{Z}_c,\\mathbf{Z}_s\\}$ , to predict the target in unseen domains. From Figure 1, we infer that $Y \\perp U|\\{\\mathbf{Z}_p,\\mathbf{Z}_c,\\mathbf{Z}_s\\}$ , implying that $p(Y|\\mathbf{Z}_{cmb})$ remains invariant across domains. Furthermore, we assert that the CMB features constitute the minimal sufficient set required for achieving invariant predictions across varying domains, as adding or removing variables from this set may affect its invariance. We demonstrate how to identify the CMB latent variables $\\mathbf{Z}_{cmb}$ from the given data distribution and construct the invariant prediction mechanism $p(Y|\\mathbf{Z}_{cmb})$ for out-of-distribution prediction in Section 4.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Causal Markov Blanket Representation Learning", + "text_level": 1, + "bbox": [ + 214, + 306, + 723, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We refer to our proposed method as CMBRL, representing Causal Markov Blanket Representation Learning. The training and inference procedures are introduced in Section 4.1 and Section 4.2, respectively.", + "bbox": [ + 212, + 340, + 787, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Training Procedure", + "text_level": 1, + "bbox": [ + 214, + 410, + 423, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We illustrate the training procedure of the proposed algorithm in Figure 2. We", + "bbox": [ + 214, + 436, + 785, + 454 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c91c943988c8d4d70045d70246392ede36e4dacb50c09e29971a61ca5f15662e.jpg", + "image_caption": [ + "Fig. 2: The illustration of the three-phase training procedure of CMBRL." + ], + "image_footnote": [], + "bbox": [ + 218, + 477, + 787, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "propose a three-phase training procedure. Initially, we utilize a VAE framework to learn a set of latent variables, which are proven to be component-wise identifiable. Subsequently, we conduct a CMB search using our novel mutual information quantification method to select the latent CMB representations. Finally, we construct a predictor using the selected CMB representations.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "N. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Phase I: Identifiable Latent Variables Learning. We aim to obtain a set of latent variables $\\mathbf{Z}$ . To do so, we employ an existing (identifiable) variational autoencoder (VAE) framework [32]. We posit a prior distribution on $\\mathbf{Z}$ that is consistent with our Assumption 1(b) and belongs to a general exponential family, i.e.,", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\boldsymbol {T}, \\boldsymbol {\\lambda}} (\\boldsymbol {Z} | Y, U) = \\frac {\\mathcal {Q} (\\boldsymbol {Z})}{\\mathcal {C} (Y , U)} \\exp [ \\boldsymbol {T} (\\boldsymbol {Z}) ^ {T} \\boldsymbol {\\lambda} (Y, U) ]\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 220, + 643, + 242 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{Q}$ is the base measure. $\\mathcal{C}$ is the normalizing constant. $\\lambda$ is the arbitrary function. $T$ is the sufficient statistics. To manage the increasing optimization challenges caused by additional parameters in the prior, we train the VAE framework using a standard protocol from prior work. [16,49]. As described in Eq. (1), the training objective comprises an ELBO loss $\\mathcal{L}_{\\mathrm{ELBO}}$ and a score matching loss $\\mathcal{L}_{\\mathrm{SM}}$ . The ELBO loss $\\mathcal{L}_{\\mathrm{ELBO}}$ optimizes over encoder and decoder parameters $(\\theta, \\phi)$ , while the score matching loss $\\mathcal{L}_{\\mathrm{SM}}$ minimizes over prior parameters $(T, \\lambda)$ . The parameters $T, \\lambda$ are constants in $\\mathcal{L}_{\\mathrm{ELBO}}$ , and $\\theta, \\phi$ are constants in $\\mathcal{L}_{\\mathrm{SM}}$ .", + "bbox": [ + 212, + 248, + 784, + 383 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {o b j}} (\\boldsymbol {\\theta}, \\phi , \\boldsymbol {T}, \\lambda) := \\mathcal {L} _ {\\mathrm {E L B O}} (\\boldsymbol {\\theta}, \\phi , \\hat {\\boldsymbol {T}}, \\hat {\\lambda}) + \\mathcal {L} _ {\\mathrm {S M}} (\\hat {\\boldsymbol {\\theta}}, \\hat {\\phi}, \\boldsymbol {T}, \\lambda) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 385, + 785, + 404 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{\\mathrm{ELBO}}$ and $\\mathcal{L}_{\\mathrm{SM}}$ are outlined in Eq. (2). $p_{\\mathcal{D}}$ is the training data distribution.", + "bbox": [ + 214, + 411, + 766, + 426 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {E L B O}} := - \\mathbb {E} _ {p _ {\\mathcal {D}}} \\left[ \\mathbb {E} _ {q _ {\\boldsymbol {\\theta}} (\\boldsymbol {z} | \\boldsymbol {x}, y, u)} [ \\log p _ {\\phi} (\\boldsymbol {x} | \\boldsymbol {z}) + \\log p _ {\\boldsymbol {T}, \\boldsymbol {\\lambda}} (\\boldsymbol {z} | y, u) - \\log q _ {\\boldsymbol {\\theta}} (\\boldsymbol {z} | \\boldsymbol {x}, y, u) ] \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 439, + 785, + 460 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S M}} := \\mathbb {E} _ {p _ {\\mathcal {D}}} \\left[ \\mathbb {E} _ {q _ {\\boldsymbol {\\theta}} (\\boldsymbol {z} | \\boldsymbol {x}, \\boldsymbol {y}, u)} [ \\| \\nabla_ {\\boldsymbol {z}} \\log q _ {\\boldsymbol {\\theta}} (\\boldsymbol {z} | \\boldsymbol {x}, \\boldsymbol {y}, u) - \\nabla_ {\\boldsymbol {z}} \\log p _ {\\boldsymbol {T}, \\boldsymbol {\\lambda}} (\\boldsymbol {z} | \\boldsymbol {y}, u) \\| ^ {2} ] \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 459, + 712, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As can be seen from the joint distribution $p(\\pmb{x}, y, u) \\propto \\int_{\\pmb{z}} p(\\pmb{x} | \\pmb{z}) p(\\pmb{z} | y, u) \\mathrm{d}\\pmb{z}$ , the prior $p(\\pmb{z} | y, u)$ is consistent with our SCM. In fact, based on our SCM, $p(\\pmb{z} | y, u)$ can be further decomposed into the product of conditional probabilities, which would result in further sparsity in $\\lambda$ . This property makes our learnt model differ from existed work based on different SCMs [19,20,32], as elaborated in Appendix B.4. However, the causal graph is generally unknown a priori, and hence we cannot pre-define the sparsity of $\\lambda$ . Therefore, in our learning algorithm we treat $p(\\pmb{z} | y, u)$ as a generic form of prior that satisfies Assumption 1(b). Without a fully specified causal graph, we use this generic prior $p(\\pmb{z} | y, u)$ to constrain the learning of VAE to obtain the $Z$ without knowing their causal identities. We summarize the assumptions and identifiability results in Theorem 1.", + "bbox": [ + 212, + 488, + 787, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Theorem 1. Assume the data is sampled from a generative model described by", + "bbox": [ + 214, + 666, + 787, + 681 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\boldsymbol {\\xi} = (\\boldsymbol {\\phi}, \\boldsymbol {T}, \\boldsymbol {\\lambda})} (\\boldsymbol {X}, \\boldsymbol {Z} | Y, U) = p _ {\\boldsymbol {\\phi}} (\\boldsymbol {X} | \\boldsymbol {Z}) p _ {\\boldsymbol {T}, \\boldsymbol {\\lambda}} (\\boldsymbol {Z} | Y, U), p _ {\\boldsymbol {\\phi}} (\\boldsymbol {X} | \\boldsymbol {Z}) = p _ {\\boldsymbol {\\epsilon}} (\\boldsymbol {X} - g _ {\\boldsymbol {\\phi}} (\\boldsymbol {Z}))\n$$\n", + "text_format": "latex", + "bbox": [ + 250, + 695, + 746, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We have the following: (1) Under certain assumptions regarding domain variability, the model parameter $\\xi$ is identifiable up to a permutation and component-wise transformation. (2) If (1) is true and $q_{\\theta}(z|\\mathbf{x},y,u)$ is flexible and positive everywhere, then our framework learns the true parameters $\\xi^{*}$ and true latent variables $\\mathbf{Z}^{*}$ up to a permutation and component-wise transformation.", + "bbox": [ + 214, + 724, + 787, + 800 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Causal Markov Blanket Representation Learning", + "bbox": [ + 405, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "5 Arbitrary function $\\lambda$ and sufficient statistics $\\pmb{T}$ are modeled by neural networks with ReLU activation due to their universal approximation ability.", + "bbox": [ + 217, + 810, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We provide the detailed assumptions for (1) to hold and proof in Appendix B.26. These detailed assumptions are common in identifiable VAE methods and can impact the performance of the CMB discovery in Phase II. The component-wise identifiability of $Z$ ensures that the CMB variables and spurious variables are separable, which is a crucial prerequisite for performing CMB discovery in Phase II, as verified by empirical results in Appendix F.1. While these assumptions may be violated in real-world applications, strict adherence to them is not always necessary. This motivates our investigation into identifiability performance on the CMNIST dataset in Section 5.1. Empirical results demonstrate that the VAE approach we employ can achieve a decent level of identifiability, even when certain assumptions do not hold.", + "bbox": [ + 212, + 145, + 787, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Phase II: Identify CMB Set. In this phase, we aim to identify the CMB representations relevant to $Y$ among the latent variables $Z$ we learned from the VAE framework. Traditional approaches to MB discovery involve revealing the local causal structure to the target through a series of (conditional) independence (CI) tests. However, conducting MB discovery in latent space encounters two major challenges: 1) the large dimension of the latent variable space, especially with complex and challenging input data, and 2) the mixture of variable types. These challenges can compromise the efficiency and accuracy of CMB discovery results. To tackle these two issues, we introduce an efficient search strategy and a more practical and general approach to conducting mutual-information-based (MI) independence tests. It is worth noting that we obtain CMB variables collectively by detecting and removing spurious variables with designed independence tests. We do not distinguish between the parent, child, and spouse variables within the CMB set.", + "bbox": [ + 212, + 335, + 787, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Search Strategy: We find that learning the exact local graph of the target variable $Y$ is unnecessary for identifying the CMB set. Instead of distinguishing among the CMB set, our objective is to separate spurious variables $Z_{o}$ from $Z_{cmb}$ . In particularly, we discover that spurious variables $Z_{o}$ and CMB variables $Z_{cmb}$ possess different relations concerning $Y$ conditioned on the remaining variables. If $Z_{i} \\in Z_{o}$ , we have $Z_{i} \\perp Y|\\{Z_{\\backslash i}, U\\}$ , where $Z_{\\backslash i} = \\{Z \\backslash Z_{i}\\}$ . If $Z_{i} \\in Z_{cmb}$ , we have $Z_{i} \\nsubseteq Y|\\{Z_{\\backslash i}, U\\}$ . We employ mutual information as measurement for independence tests. Theoretically, we can detect spurious variables and separate them from CMB set by quantifying $\\mathcal{I}(Z_{i}; Y|Z_{\\backslash i}, U), \\forall Z_{i} \\in Z$ . Actually, for real-world scenarios where $U$ is unknown, it is challenging to estimate $\\mathcal{I}(Z_{i}; Y|Z_{\\backslash i}, U)$ . However, subject to our SCM structural assumptions, we have $\\mathcal{I}(Z_{i}; Y|Z_{\\backslash i}) = \\mathcal{I}(Z_{i}; Y|Z_{\\backslash i}, U) = 0$ , if $Z_{i} \\in Z_{o}$ . We provide a proof in Appendix C.1. Meanwhile, if $Z_{i} \\in Z_{cmb}$ , $\\mathcal{I}(Z_{i}; Y|Z_{\\backslash i}) > 0$ . Therefore, we can separate $Z_{o}$ and $Z_{cmb}$ by performing $\\mathcal{I}(Z_{i}; Y|Z_{\\backslash i}), \\forall Z_{i} \\in Z$ . We summarize such theoretical findings in Proposition 1.", + "bbox": [ + 212, + 547, + 787, + 776 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "N. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "6 Our proof follows a similar approach to that of [32]. We aim to demonstrate that our assumptions on SCM do not violate the assumptions in the proof, thereby ensuring that the identifiable results hold under our framework.", + "bbox": [ + 217, + 796, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Proposition 1. If the causal graph in Figure 1 and Assumption 1 holds, we have that: 1) for $Z_{i} \\in \\mathbf{Z}_{o}$ , $Z_{i} \\perp Y|Z_{\\backslash i}$ , $\\mathcal{I}(Z_{i};Y|Z_{\\backslash i}) = 0$ ; 2) for $Z_{i} \\in \\mathbf{Z}_{cmb}$ , $Z_{i} \\nsubseteq Y|Z_{\\backslash i}$ , $\\mathcal{I}(Z_{i};Y|Z_{\\backslash i}) > 0$ .", + "bbox": [ + 214, + 146, + 787, + 194 + ], + "page_idx": 8 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Causal Markov Blanket Set Search Strategy" + ], + "code_body": "Input: M observations of Z and Y, i.e., $\\{z_{1}(j), z_{2}(j), \\dots, z_{N}(j), y(j)\\}_{j=1}^{M}$ \nOutput: CMB set $Z_{cmb}$ \nStep 1(Fast-Search): find the identity for each variable $Z_{i} \\in \\mathbf{Z}$ \nInitial two sets $Z_{cmb} \\gets \\emptyset$ , $Z_{o} \\gets \\emptyset$ \nfor $Z_{i} \\in \\mathbf{Z}, i = 1, 2, \\dots, N$ , do \n Perform Indtest $(Z_{i}; Y|Z_{\\backslash i})$ \n if $Z_{i} \\nsubseteq Y|Z_{\\backslash i}$ then \n $Z_{cmb} \\gets Z_{cmb} \\cup Z_{i}$ \n else \n $Z_{o} \\gets Z_{o} \\cup Z_{i}$ \nend if \nend for \nStep 2 (Forward-Backward Search): verify the obtained CMB set. \nrepeat \n for $Z_{l} \\in Z_{cmb}$ do \n ▷ Backward procedure \n Perform Indtest $(Z_{l}; Y|\\{Z_{cmb}\\backslash Z_{l}\\})$ \n if $Z_{l} \\perp Y|\\{Z_{cmb}\\backslash Z_{l}\\}$ then \n $Z_{cmb} \\gets Z_{cmb} \\backslash Z_{l}$ \n end if \nend for \nfor $Z_{j} \\in Z_{o}$ do \n ▷ Forward procedure \n Perform Indtest $(Z_{j}; Y|Z_{cmb})$ \n if $Z_{j} \\nsubseteq Y|Z_{cmb}$ then \n $Z_{cmb} \\gets Z_{cmb} \\cup Z_{j}$ \n end if \nend for \nuntil Converge", + "bbox": [ + 215, + 252, + 787, + 633 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We provide the detailed proof of Proposition 1 in Appendix C.2.", + "bbox": [ + 215, + 661, + 694, + 676 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We hence propose a fast-search strategy, which performs CI tests $\\mathcal{I}(Z_i;Y|\\mathbf{Z}_{\\backslash i})$ for each variable $Z_{i}\\in \\mathbf{Z}$ . After performing the fast-search and obtaining the initial set of CMB, we verify the initial CMB set by repeatedly checking whether $\\forall Z_{j}\\in \\mathbf{Z}_{o}$ , $Z_{j}\\perp Y|Z_{cmb}$ and $\\forall Z_{k}\\in \\mathbf{Z}_{cmb}$ , $Z_{k}\\nsubseteq Y|\\{\\mathbf{Z}_{cmb}\\backslash Z_{k}\\}$ . We outline our CMB search strategy in Algorithm 1. As illustrated by Proposition 2, our algorithm can correctly identify the CMB variables if the assumptions hold. We provide the detailed proof for Proposition 2 in Appendix C.3.", + "bbox": [ + 212, + 676, + 787, + 782 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Proposition 2. Assume independence test results are correct and Assumption 1 for the proposed SCM holds, the ground truth Causal Markov Blanket set can be identified with Algorithm 1.", + "bbox": [ + 214, + 794, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Causal Markov Blanket Representation Learning", + "bbox": [ + 405, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In Step 1, we leverage the structural properties of CMB features to efficiently distinguish them from spurious features. However, since we condition on $N - 1$ variables, which is a larger set than CMB, in step 1, we may have false positive detections and wrongly include spurious features. Thus, we employ Step 2, a forward-backward procedure, to eliminate possible errors in Step 1.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Complexity Analysis of CMB Search: Assuming it takes $k$ iterations, with $k << N$ as observed in practice, for the verification to converge, our CMB strategy requires performing $(k + 1)N$ CI tests and has a linear complexity of $\\mathcal{O}(N)$ . Compared to traditional MB discovery methods like IAMB [48], which directly execute the forward-backward search, our fast-search step reduces the search space and improves learning efficiency, often achieving convergence in just $k = 1$ or 2 iterations in practice.", + "bbox": [ + 212, + 220, + 787, + 327 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "CI Tests: We use the MI-based, denoted as $\\mathcal{I}$ , independence tests, with a significance level $\\alpha$ . If $\\mathcal{I} < \\alpha$ , we declare independence. In Phase II, we need to compute the following mutual information $\\mathcal{I}(Y; Z_i | \\mathcal{C})$ , where $\\mathcal{C}$ is the set of variables to condition on. We propose a practical approach that quantifies the MI using a trained predictor and is applicable for both classification tasks and regression tasks. Calculating mutual information for classification tasks is particularly challenging due to the mixture of variable types (where $Y$ is discrete and variables in $\\mathbf{Z}$ are continuous). We illustrate our approach to estimating the MI in Eq. (3) and Eq. (4). By definition of mutual information and conditional entropy, we have", + "bbox": [ + 212, + 328, + 787, + 477 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {I} (y; z _ {i} | \\mathcal {C}) = \\mathcal {H} [ y | \\mathcal {C} ] - \\mathcal {H} [ y | \\{\\mathcal {C}, z _ {i} \\} ] \\\\ = \\mathbb {E} _ {p (\\mathcal {C})} \\left[ \\mathcal {H} [ p (y | \\mathcal {C}) ] \\right] - \\mathbb {E} _ {p (\\mathcal {C}, z _ {i})} \\left[ \\mathcal {H} [ p (y | \\mathcal {C}, z _ {i}) ] \\right] \\tag {3} \\\\ = \\mathbb {E} _ {p (\\mathcal {C})} \\left[ \\mathcal {H} \\left[ \\mathbb {E} _ {p (\\boldsymbol {z} \\backslash \\mathcal {C} | \\mathcal {C})} [ p (y | \\boldsymbol {z}) ] \\right] \\right] - \\mathbb {E} _ {p (\\mathcal {C}, z _ {i})} \\left[ \\mathcal {H} \\left[ \\mathbb {E} _ {p (\\boldsymbol {z} \\backslash \\{\\mathcal {C}, z _ {i} \\} | \\{\\mathcal {C}, z _ {i} \\})} [ p (y | \\boldsymbol {z}) ] \\right] \\right] \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 266, + 484, + 784, + 545 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "According to Eq. (3), we can calculate $\\mathcal{I}(y;z_i|\\mathcal{C})$ using a predictor distribution $p(Y|\\mathbf{Z})$ and samples of $\\mathbf{Z}$ . For one input and its label from the training data, i.e., $\\boldsymbol{x}(j), y(j), u(j) \\sim p_{\\mathcal{D}}$ , we use the mean of learned encoder $q_{\\hat{\\theta}}(\\mathbf{Z}|\\mathbf{X} = \\boldsymbol{x}(j), Y = y(j), U = u(j))$ as the sample for $\\mathbf{Z}$ , denoted as $\\boldsymbol{z}(j) = [z_1(j), z_2(j), \\dots, z_N(j)]$ . After obtaining a total of $M$ observations for $\\mathbf{Z}$ , i.e., $\\{\\boldsymbol{z}(j)\\}_{j=1}^{M}$ , we first learn a predictor $p(Y|\\mathbf{Z})$ with $\\{\\boldsymbol{z}(j), y(j)\\}_{j=1}^{M}$ . Substituting the samples and learned distribution into Eq. (3), we have", + "bbox": [ + 212, + 551, + 787, + 660 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {I} (y, z _ {i} | \\mathcal {C}) := - \\frac {1}{M} \\sum_ {j = 1} ^ {M} \\sum_ {l = 0} ^ {L - 1} \\left(\\frac {\\sum_ {k = 1} ^ {M} p \\big (y = l | \\mathcal {C} (j) , \\{z \\backslash \\mathcal {C} \\} (k) \\big)}{M}\\right) \\log \\left(\\frac {\\sum_ {k = 1} ^ {M} p \\big (y = l | \\mathcal {C} (j) , \\{z \\backslash \\mathcal {C} \\} (k) \\big)}{M}\\right) + \\\\ \\frac {1}{M} \\sum_ {j = 1} ^ {M} \\sum_ {l = 0} ^ {L - 1} \\left(\\frac {\\sum_ {k = 1} ^ {M} p (y = l | \\{\\mathcal {C} , z _ {i} \\} (j) , \\{z \\backslash (\\mathcal {C} , z _ {i}) \\} (k))}{M}\\right) \\log \\left(\\frac {\\sum_ {k = 1} ^ {M} p (y = l | \\{\\mathcal {C} , z _ {i} \\} (j) , \\{z \\backslash (\\mathcal {C} , z _ {i}) \\} (k))}{M}\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 269, + 667, + 730, + 724 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $L$ is the number of classes for $Y$ . $l$ denotes the value of $Y$ and $l \\in \\{0,1,\\dots ,L - 1\\}$ . Especially, for the fast-search strategy, $\\mathcal{C} = \\mathbf{z}_{\\backslash i}$ , we have:", + "bbox": [ + 214, + 731, + 785, + 763 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {I} (y; z _ {i} | \\boldsymbol {z} _ {\\backslash i}) = - \\frac {1}{M} \\sum_ {j = 1} ^ {M} \\sum_ {l = 0} ^ {L - 1} \\frac {\\sum_ {k = 1} ^ {M} p \\left(y = l | \\boldsymbol {z} _ {\\backslash i} (j) , z _ {i} (k)\\right)}{M} \\log \\frac {\\sum_ {k = 1} ^ {M} p \\left(y = l | \\boldsymbol {z} _ {\\backslash i} (j) , z _ {i} (k)\\right)}{M} \\\\ + \\frac {1}{M} \\sum_ {j = 1} ^ {M} \\sum_ {l = 0} ^ {L - 1} p (y = l | z (j)) \\log p (y = l | z (j)) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 246, + 772, + 733, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "N. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The detailed derivation can be found in Appendix C.4. Our CMB discovery method assumes the accuracy of the IndTest. In practice, our approach, like other statistical tests, faces challenges with accuracy when the condition set is large and data is insufficient. Future studies aimed at improving the accuracy of independence tests could benefit from utilizing Bayesian mutual information [9].", + "bbox": [ + 214, + 146, + 787, + 223 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Phase III: Construct Invariant Predictor. We utilize the obtained CMB variables $Z_{cmb}$ and train an invariant predictor $p(Y|Z_{cmb})$ . Our training procedure is outlined in Eq. (6).", + "bbox": [ + 214, + 239, + 787, + 287 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\psi} = \\arg \\min _ {\\boldsymbol {\\psi}} \\frac {1}{M} \\sum_ {j = 1} ^ {M} \\mathcal {L} _ {\\operatorname {p r e d}} \\left(y (j), p _ {\\boldsymbol {\\psi}} \\left(\\boldsymbol {Y} \\mid \\boldsymbol {Z} _ {c m b} = \\boldsymbol {z} _ {c m b} (j)\\right)\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 292, + 785, + 316 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $\\psi$ are the parameters of the predictor. $\\mathcal{L}_{\\mathrm{pred}}(\\cdot)$ represents the cross-entropy loss for classification tasks and the mean squared loss for regression tasks.", + "bbox": [ + 214, + 323, + 785, + 354 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Inference Procedure", + "text_level": 1, + "bbox": [ + 215, + 373, + 429, + 387 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In the inference procedure, we predict labels for data from an unseen test domain, $p_{\\mathcal{D}^{test}}$ . One can infer from Figure 1 that $p(\\pmb{X}|\\pmb{Z})$ is domain-invariant since $X \\perp U|Z$ . Therefore, inferring the latent variable values for inputs from an unseen test domain using $p(\\pmb{X}|\\pmb{Z})$ is reasonable. We adopt the standard inference procedure from [32, 45]. Given an input from the test domain, $\\pmb{x}^t \\sim p_{\\mathcal{D}^{test}}$ , we obtain the optimal values of $Z_{cmb}$ , denoted as $z_{cmb}^{*}$ , by solving the optimization problem outlined in Eq. (7). Here, $\\lambda_1$ and $\\lambda_2$ are hyperparameters used to control the scales of the learned $Z_{cmb}$ and $Z_o$ .", + "bbox": [ + 214, + 393, + 787, + 516 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {c m b} ^ {*}, \\boldsymbol {z} _ {o} ^ {*} = \\arg \\min _ {\\boldsymbol {z} _ {c m b}, \\boldsymbol {z} _ {o}} p _ {\\hat {\\phi}} (\\boldsymbol {x} ^ {t} | \\boldsymbol {z} _ {c m b}, \\boldsymbol {z} _ {o}) + \\lambda_ {1} \\| \\boldsymbol {z} _ {c m b} \\| _ {2} ^ {2} + \\lambda_ {2} \\| \\boldsymbol {z} _ {o} \\| _ {2} ^ {2} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 267, + 521, + 785, + 542 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Then we can infer the label using the constructed invariant predictor, i.e.,", + "bbox": [ + 215, + 546, + 748, + 561 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} = \\arg \\max _ {y} p _ {\\hat {\\psi}} (y | \\boldsymbol {z} _ {c m b} ^ {*}) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 568, + 785, + 590 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We provide the complexity analysis of the optimization in the inference procedure in Appendix D and the ablation study of runtime comparison in Appendix F.3.", + "bbox": [ + 214, + 598, + 785, + 628 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 648, + 375, + 666 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We validate our CMBRL method for OOD prediction using synthetic and real benchmark distribution shift datasets against state-of-the-art domain generalization baselines. We experiment on four datasets: CMNIST, CelebA, PACS, and VLCS. Our experiments, conducted across 5 trials, are summarized in tables, reporting the mean and standard deviation of accuracy. The details of datasets and implementations can be found in Appendix E.1 and E.2. Additionally, detailed ablation studies on the selection of hyperparameters are available in Appendix F.", + "bbox": [ + 214, + 676, + 787, + 796 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Identifiability of Latent Variables", + "text_level": 1, + "bbox": [ + 215, + 815, + 537, + 830 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Causal Markov Blanket Representation Learning", + "bbox": [ + 405, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "First, we verify that Phase I of our proposed CMBRL yields identifiable $\\mathbf{Z}$ . We compare it with VAE [30], which lacks identifiability guarantees, and iVAE [19], which uses a conditionally factorized prior. Following standard procedures [19, 20, 22, 32], we evaluate identifiability on CMNIST data by computing the average mean correlation coefficient (MCC) between latent variables recovered by different models with various random initializations. Higher MCC scores indicate stronger identifiability, as shown in Figure 37. Using the same prior and training process as [32], we refer to our method as NF-", + "bbox": [ + 212, + 146, + 537, + 356 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b777fe0796829863924bd9511afa0e31472150f31cb974729e662398569b3db8.jpg", + "image_caption": [ + "Fig.3: MCC on CMNIST" + ], + "image_footnote": [], + "bbox": [ + 558, + 155, + 784, + 295 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "iVAE. Figure 3 demonstrates that NF-iVAE recovers latent variables with better identifiability. This identifiability impacts the disentanglement of spurious and CMB variables, influencing the accuracy of the estimated CMB set. Our ablation study in Appendix F.1 empirically shows that better identifiability leads to a more accurate CMB set.", + "bbox": [ + 212, + 357, + 787, + 434 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Out-of-distribution Prediction Accuracy", + "text_level": 1, + "bbox": [ + 214, + 455, + 591, + 472 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We present both the in-distribution and OOD prediction accuracies of the CMNIST dataset in Table 1. First, we compare with the ERM trained on the entire training domain. Then, we compare with the popular causal invariant feature learning method, IRM, and its two variants: F-IRM (where $\\varPhi$ is fixed to the identity) and V-IRM (where $\\varPhi$ is variable). We also include the Robust MIN MAX method, which minimizes the maximum loss across multiple domains. Additionally, we compare with other approaches that leverage SCM knowledge to identify causal features", + "bbox": [ + 212, + 483, + 493, + 710 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/da141dafa74ea058053b9d07c2a05769a93e464ac39b3a85b063c5678585f0e2.jpg", + "table_caption": [ + "Table 1: Results on Colored MNIST in terms of accuracy $(\\%)$" + ], + "table_footnote": [], + "table_body": "
MethodsPrediction Acc (%)
In-distributionOOD
ERM85.7±0.510.3±0.2
Robust MIN MAX84.3±0.410.9±0.5
F-IRM GAME63.4±1.160.0±2.7
V-IRM GAME64.0±1.049.2±3.4
IRM59.3±4.462.8±9.6
Causalrep70.1±1.568.6±5.5
CTrans76.9±0.872.5±1.1
iCaRL870.6±0.868.8±1.5
CMBRL(ours)76.4±0.874.1±1.5
", + "bbox": [ + 516, + 556, + 772, + 680 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "of the target, including Causalrep [53], CTrans [34], and iCaRL [32]. CTrans, in particular, considers latent confounding effects between spurious representation and target and employs causal inference via intervention.", + "bbox": [ + 212, + 710, + 784, + 755 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "As shown in Table 1, our CMBRL method achieves optimal OOD performance while minimally compromising in-distribution accuracy. In contrast, the ERM method, which minimizes losses over the entire training domain, tends to", + "bbox": [ + 212, + 755, + 785, + 801 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "N. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "page_footnote", + "text": "7 We verify our NF-iVAE implementation's correctness by demonstrating similar identifiability performance to iCaRL in Figure 3", + "bbox": [ + 217, + 810, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "capture spurious statistical correlations between labels and colors. Although it achieves the highest in-distribution accuracy, there is a significant drop in prediction accuracy on OOD data. Among the multiple domain-invariant causal representation learning methods, including IRM, F-IRM GAME, and V-IRM GAME, there are notable improvements. However, the limited number of available domains (only 2) may constrain the performance of the IRM methods. Conversely, causal representation learning methods operating under various Structural Causal Models (SCMs), including our CMBRL, demonstrate superior empirical performance.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We further investigate the distinctions between the feature sets selected by iCaRL and our CMBRL. Following the approach described in [32], we first learn the causal graph between the estimated latent variables $Z$ from NF-iVAE and $Y$ using the PC algorithm. We then select the parent set with additional independent tests. The selected parent set yields an OOD prediction accuracy of $69.8\\%$ , consistent with the reported results in Table 1. By applying a suitable significance level $\\alpha$ , we can select a CMB set that includes all the variables present in the selected parent set. The OOD prediction accuracy using such a CMB set increased to $74.1\\%$ . This finding supports our assumptions on the proposed SCM and suggests the presence of variables not categorized as parent variables of the target, which can further enhance OOD prediction performance. Moreover, CTrans [34] and [22] assume there is an invariant latent confounder between $Y$ and $Z_{o}$ , leading to spurious correlations between $Y$ and $X$ . For example, in an image with a water bird and sea background, the confounder could include temperature and altitude, influencing the generation of the sea (spurious variable $Z_{o}$ ) and bird $(Y)$ . Our proposed SCM does not explicitly model latent confounders. Therefore, we cannot guarantee the generalization of our CMBRL method to scenarios involving unobserved confounders between input $X$ and target $Y$ . While addressing the latent confounding issue falls outside the scope of this paper, it presents an intriguing avenue for future research. We sum", + "bbox": [ + 212, + 284, + 787, + 587 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/613f9e9cd4062edd19658c1c6d6ace01f967c3d9881818843afd5c8ff84be822.jpg", + "table_caption": [ + "Table 2: Empirical results on CelebA in terms of the OOD prediction accuracy $(\\%)$" + ], + "table_footnote": [], + "table_body": "
MethodsGroupDROMLDGCORALMMDDACNNMixupERMCTransCausalrepiCaRLCMBRL(ours)
Avg Acc65.165.666.766.268.769.560.766.360.471.576.2
", + "bbox": [ + 218, + 643, + 784, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "marize empirical results on average OOD prediction accuracy over 7 settings for the CelebA dataset in Table 2. Detailed empirical results for each pair of settings can be found in Appendix H. We compare the results of our CMBRL with the ERM method and state-of-the-art causal and non-causal representation learning methods. As illustrated in Table 2, our method demonstrates optimal performance, surpassing the second-best baseline by an average margin of $4.7\\%$ across all settings. This outcome indicates the efficacy of our CMBRL methods in excluding spurious variables and avoiding reliance on spurious correlations for prediction in computer vision applications.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Causal Markov Blanket Representation Learning", + "bbox": [ + 406, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1a3c955f21a5f4e2efab8ce1d0d54609b0fcc80b7d099dbe62010973af0e98c4.jpg", + "table_caption": [ + "Table 3: Empirical results on VLCS and PACS datasets in terms of OOD prediction accuracy (\\%) with a backbone of ResNet50." + ], + "table_footnote": [], + "table_body": "
AlgorithmsVLCSPACS
CLSVAvgACPSAvg
ERM98.0±0.462.6±0.970.8 ±1.977.5 ±1.977.284.8±1.376.4±1.196.7±0.676.1±1.083.5
GroupDRO98.1±0.366.4 ±0.971.0 ±0.376.1 ±1.477.983.5±0.979.1±0.696.7±0.378.3±2.084.4
MLDG98.5±0.361.7±1.273.6±1.875.0±0.877.285.5±1.480.1±1.797.4±0.376.6±1.184.9
CORAL96.9±0.965.7±1.273.3±0.778.7±0.878.788.3±0.280.0±0.797.5±0.378.8±1.386.2
MMD98.3±0.165.6±0.769.7±1.075.7±0.977.386.1±1.479.4±0.996.6±0.276.5±0.784.6
RSC97.5±0.663.1±1.273.0±1.376.2±0.577.585.4±0.879.7±1.897.6±0.378.2±1.285.2
Mixup98.4±0.363.4±0.772.9±0.876.1±1.277.786.1±0.778.9±0.897.6±0.175.8±1.884.6
DANN98.5±1.364.9±1.372.6±1.478.7±1.778.286.4±0.877.4±0.897.3±0.473.5±2.383.6
CDANN97.6±0.665.2±0.873.4±1.476.9±0.578.384.6±1.875.5±0.996.8±0.373.5±0.682.6
MTL97.6±0.660.6±1.371.0±1.277.2±0.776.687.5±0.877.1±0.796.4±0.877.3±1.884.6
ARM97.2±0.562.7±1.70.6±0.675.8±0.976.686.8±0.676.8±0.797.4±0.379.3±1.285.1
IRM98.6±0.166.0 ±0.972.3 ±0.677.3 ±0.978.584.7±0.480.0±0.697.2±0.379.3±1.085.5
SagNet97.3±0.461.6±0.873.4±1.977.6±0.477.587.4±1.080.7±0.697.1±0.180.0±0.486.3
iCaRL----81.8----88.7
CMBRL (ours)98.7±0.271.2±0.177.1±0.482.1±0.182.389.2±0.785.3±1.297.7±0.584.1±0.689.1
", + "bbox": [ + 218, + 185, + 792, + 372 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We present the empirical results for PACS and VLCS regarding OOD prediction accuracy in Table 3. These datasets serve as benchmarks for domain generalization tasks. For a comprehensive comparison, we evaluate not only against causal-inspired methods but also against various other DG methods, including GroupDRO [41], MLDG [25], CORAL [44], RSC [15], Mixup [54], SageNet [36], and others. We adhere to the experimental settings outlined in [13]. We report the OOD performance for each pair of train and test domains and the average accuracy across all combinations. As depicted in Table 3, our CMBRL method establishes state-of-the-art performance compared to popular alternatives in the domain generalization landscape.", + "bbox": [ + 212, + 402, + 787, + 555 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 580, + 359, + 595 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we investigate the Causal Markov Blanket discovery for high-dimensional data. We establish a framework guided by an SCM describing the data generation process, allowing for the Causal Markov Blanket discovery in latent space. We then construct an invariant prediction mechanism utilizing the CMB representations, which is suitable for OOD prediction. This framework can be further employed for causal analysis, reasoning, and intervention. We propose a three-phase algorithm to disentangle CMB features from spurious ones, reducing the risk of the prediction model relying on spurious correlations. It is worth noting that our method requires the satisfaction of assumptions on SCM and the assurance of the identifiability of latent variables. It may also suffer from inefficiencies due to the three training phases and optimization during inference. However, our method demonstrates its effectiveness by achieving significant OOD prediction performance, surpassing state-of-the-art causal representation learning methods and domain generalization methods on multiple benchmark distribution shift datasets.", + "bbox": [ + 212, + 613, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "N. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by the National Science Foundation award IIS 2236026 and in part by IBM through the IBM-Rensselaer Future of Computing Research Collaboration.", + "bbox": [ + 215, + 172, + 785, + 218 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 239, + 321, + 255 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahuja, K., Caballero, E., Zhang, D., Bengio, Y., Mitlagkas, I., Rish, I.: Invariance principle meets information bottleneck for out-of-distribution generalization. In: Advances in Neural Inf. Proc. Systems (2021) 3", + "2. Ahuja, K., Shanmugam, K., Varshney, K., Dhurandhar, A.: Invariant risk minimization game. In: International Conference on Machine Learning (2020) 3", + "3. Ahuja, K., Wang, J., Dhurandhar, A., Shanmugam, K., Varshney, K.R.: Empirical or invariant risk minimization? a sample complexity perspective. In: International Conference on Learning Representations (2021) 3, 5", + "4. Aliferis, C.F., Statnikov, A., Tsamardinos, I., Mani, S., Koutsoukos, X.D.: Local causal and markov blanket induction for causal discovery and feature selection for classification part i: algorithms and empirical evaluation. Journal of Machine Learning Research 11(1) (2010) 2, 3", + "5. Aliferis, C.F., Tsamardinos, I., Statnikov, A.: Hiton: a novel markov blanket algorithm for optimal variable selection. In: AMIA annual symposium proceedings. vol. 2003, p. 21. American Medical Informatics Association (2003) 3", + "6. Arjovsky, M., Bottou, L., Gulrajani, I., Lopez-Paz, D.: Invariant risk minimization. arXiv preprint arXiv:1907.02893 (2019) 3", + "7. Cui, P., Athey, S.: Stable learning establishes some common ground between causal inference and machine learning. Nature Machine Intelligence 4(2), 110-115 (2022) 3", + "8. Cui, P., Shen, Z., Li, S., Yao, L., Li, Y., Chu, Z., Gao, J.: Causal inference meets machine learning. In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. pp. 3527-3528 (2020) 3", + "9. Cui, Z., Yin, N., Wang, Y., Ji, Q.: Empirical bayesian approaches for robust constraint-based causal discovery under insufficient data. International Joint Conference on Artificial Intelligence (2022) 11", + "0. Fu, S., Desmarais, M.C.: Fast markov blanket discovery algorithm via local learning within single pass. In: Advances in Artificial Intelligence: 21st Conference of the Canadian Society for Computational Studies of Intelligence, Canadian AI 2008 Windsor, Canada, May 28-30, 2008 Proceedings 21. pp. 96-107. Springer (2008) 3", + "1. Gao, T., Ji, Q.: Local causal discovery of direct causes and effects. Advances in Neural Information Processing Systems 28 (2015) 3", + "2. Gao, T., Wang, Z., Ji, Q.: Structured feature selection. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (December 2015) 2, 3, 4", + "3. Gulrajani, I., Lopez-Paz, D.: In search of lost domain generalization. arXiv preprint arXiv:2007.01434 (2020) 14", + "4. Hendrycks, D., Mu, N., Cubuk, E.D., Zoph, B., Gilmer, J., Lakshminarayanan, B.: Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781 (2019) 3", + "5. Huang, Z., Wang, H., Xing, E.P., Huang, D.: Self-challenging improves cross-domain generalization. In: European Conference on Computer Vision. pp. 124-140. Springer (2020) 14" + ], + "bbox": [ + 225, + 267, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Causal Markov Blanket Representation Learning", + "bbox": [ + 406, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Hyvarinen, A., Dayan, P.: Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research 6(4) (2005) 7", + "17. Janzing, D.: Causal regularization. Advances in Neural Information Processing Systems 32 (2019) 3", + "18. Jiang, Y., Veitch, V.: Invariant and transportable representations for anti-causal domain shifts. arXiv preprint arXiv:2207.01603 (2022) 3", + "19. Khemakhem, I., Kingma, D., Monti, R., Hyvarinen, A.: Variational autoencoders and nonlinear ica: A unifying framework. In: International Conference on Artificial Intelligence and Statistics. pp. 2207-2217. PMLR (2020) 3, 4, 7, 12", + "20. Khemakhem, I., Monti, R., Kingma, D., Hyvarinen, A.: Ice-beem: Identifiable conditional energy-based deep models based on nonlinear ica. Advances in Neural Information Processing Systems 33, 12768-12778 (2020) 4, 7, 12", + "21. Koller, D., Sahami, M.: Toward optimal feature selection. Tech. rep., Stanford InfoLab (1996) 3", + "22. Kong, L., Xie, S., Yao, W., Zheng, Y., Chen, G., Stojanov, P., Akinwande, V., Zhang, K.: Partial identifiability for domain adaptation. arXiv preprint arXiv:2306.06510 (2023) 2, 3, 5, 12, 13", + "23. Koyama, M., Yamaguchi, S.: When is invariance useful in an out-of-distribution generalization problem? arXiv preprint arXiv:2008.01883 (2020) 3", + "24. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.: Learning to generalize: Meta-learning for domain generalization. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018) 3", + "25. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.M.: Deeper, broader and artier domain generalization. In: Proceedings of the IEEE international conference on computer vision. pp. 5542-5550 (2017) 14", + "26. Li, Z., Cai, R., Chen, G., Sun, B., Hao, Z., Zhang, K.: Subspace identification for multi-source domain adaptation. Advances in Neural Information Processing Systems 36 (2024) 3", + "27. Liu, B., Wang, D., Yang, X., Zhou, Y., Yao, R., Shao, Z., Zhao, J.: Show, deconfound and tell: Image captioning with causal inference. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18041-18050 (2022) 4", + "28. Locatello, F., Bauer, S., Lucic, M., Raetsch, G., Gelly, S., Schölkopf, B., Bachem, O.: Challenging common assumptions in the unsupervised learning of disentangled representations. In: international conference on machine learning. pp. 4114-4124. PMLR (2019) 3", + "29. Locatello, F., Poole, B., Ratsch, G., Scholkopf, B., Bachem, O., Tschannen, M.: Weakly-supervised disentanglement without compromises. In: International Conference on Machine Learning. pp. 6348-6359. PMLR (2020) 3", + "30. Lopez, R., Regier, J., Jordan, M.I., Yosef, N.: Information constraints on auto-encoding variational bayes. Advances in neural information processing systems 31 (2018) 12", + "31. Lopez-Paz, D., Nishihara, R., Chintala, S., Scholkopf, B., Bottou, L.: Discovering causal signals in images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6979-6987 (2017) 2", + "32. Lu, C., Wu, Y., Hernández-Lobato, J.M., Schölkopf, B.: Invariant causal representation learning for out-of-distribution generalization. In: International Conference on Learning Representations (2021) 2, 3, 4, 5, 7, 8, 11, 12, 13", + "33. Mao, C., Cha, A., Gupta, A., Wang, H., Yang, J., Vondrick, C.: Generative interventions for causal learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3947-3956 (2021) 2, 3, 4, 5" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "N. Yin et al.", + "bbox": [ + 271, + 114, + 356, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Mao, C., Xia, K., Wang, J., Wang, H., Yang, J., Baireinboim, E., Vondrick, C.: Causal transportability for visual recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7521-7531 (2022) 2, 4, 5, 12, 13", + "35. Nagarajan, V., Andreassen, A., Neyshabur, B.: Understanding the failure modes of out-of-distribution generalization. arXiv preprint arXiv:2010.15775 (2020) 1", + "36. Nam, H., Lee, H., Park, J., Yoon, W., Yoo, D.: Reducing domain gap by reducing style bias. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8690-8699 (2021) 14", + "37. Pearl, J.: Probabilistic reasoning in intelligent systems: networks of plausible inference. Morgan Kaufmann (1988) 4", + "38. Pellet, J.P., Elisseeff, A.: Using markov blankets for causal structure learning. Journal of Machine Learning Research 9(7) (2008) 2, 3", + "39. Peters, J., Buhlmann, P., Meinshausen, N.: Causal inference using invariant prediction: identification and confidence intervals. arxiv. Methodology (2015) 2, 5", + "40. Rosenfeld, E., Ravikumar, P., Risteski, A.: The risks of invariant risk minimization. In: International Conference on Learning Representations (2021) 3", + "41. Sagawa, S., Koh, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731 (2019) 2, 14", + "42. Shen, X., Liu, F., Dong, H., Lian, Q., Chen, Z., Zhang, T.: Weakly supervised disentangled generative causal representation learning. The Journal of Machine Learning Research 23(1), 10994-11048 (2022) 3", + "43. Shu, Y., Cao, Z., Wang, C., Wang, J., Long, M.: Open domain generalization with domain-augmented meta-learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9624-9633 (2021) 3", + "44. Sun, B., Saenko, K.: Deep coral: Correlation alignment for deep domain adaptation. In: Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14. pp. 443-450. Springer (2016) 14", + "45. Sun, X., Wu, B., Zheng, X., Liu, C., Chen, W., Qin, T., Liu, T.y.: Latent causal invariant model. arXiv preprint arXiv:2011.02203 (2020) 11", + "46. Talon, D., Lippe, P., James, S., Del Bue, A., Magliacane, S.: Towards the reusability and compositionality of causal representations. arXiv preprint arXiv:2403.09830 (2024) 3", + "47. Tan, Y., Liu, Z.: Feature selection and prediction with a markov blanket structure learning algorithm. In: BMC bioinformatics. vol. 14, pp. 1-3. BioMed Central (2013) 2, 3", + "48. Tsamardinos, I., Aliferis, C.F., Statnikov, A.R., Statnikov, E.: Algorithms for large scale markov blanket discovery. In: FLAIRS conference. vol. 2, pp. 376-380. St. Augustine, FL (2003) 3, 10", + "49. Vincent, P.: A connection between score matching and denoising autoencoders. Neural computation 23(7), 1661-1674 (2011) 7", + "50. Volpi, R., Namkoong, H., Sener, O., Duchi, J.C., Murino, V., Savarese, S.: Generalizing to unseen domains via adversarial data augmentation. Advances in neural information processing systems 31 (2018) 3", + "51. Wang, H., Xiao, C., Kossaifi, J., Yu, Z., Anandkumar, A., Wang, Z.: Augmax: Adversarial composition of random augmentations for robust training. Advances in neural information processing systems 34, 237-250 (2021) 3" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Causal Markov Blanket Representation Learning", + "bbox": [ + 406, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "52. Wang, T., Huang, J., Zhang, H., Sun, Q.: Visual commonsense r-cnn. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10760-10770 (2020) 2, 4", + "53. Wang, Y., Jordan, M.I.: Desiderata for representation learning: A causal perspective. arXiv preprint arXiv:2109.03795 (2021) 12", + "54. Yan, S., Song, H., Li, N., Zou, L., Ren, L.: Improve unsupervised domain adaptation with mixup training. arXiv preprint arXiv:2001.00677 (2020) 14", + "55. Yin, N., Wang, H., Gao, T., Dhurandhar, A., Ji, Q.: Causal markov blanket representation learning for out-of-distribution generalization. In: Causal Representation Learning Workshop at NeurIPS 2023 (2023) 2, 4", + "56. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6023-6032 (2019) 3", + "57. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017) 3" + ], + "bbox": [ + 215, + 146, + 784, + 354 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "N. Yin et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization/bb833fed-6970-4db4-b3da-646e63e5493d_model.json b/2024/Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization/bb833fed-6970-4db4-b3da-646e63e5493d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..74f717f745a82f12fd7dabd8590ac8fa2ff26b33 --- /dev/null +++ b/2024/Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization/bb833fed-6970-4db4-b3da-646e63e5493d_model.json @@ -0,0 +1,2304 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.252, + 0.141, + 0.753, + 0.207 + ], + "angle": 0, + "content": "Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.234, + 0.784, + 0.266 + ], + "angle": 0, + "content": "Naiyu Yin\\(^{1}\\), Hanjing Wang\\(^{1}\\), Yue Yu\\(^{2}\\), Tian Gao\\(^{3}\\), Amit Dhurandhar\\(^{3}\\), and Qiang Ji\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.277, + 0.692, + 0.306 + ], + "angle": 0, + "content": "1 Rensselaer Polytechnic Institute, Troy NY 12180, USA {yinn2, wangh36, jiq}@rpi.edu" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.306, + 0.662, + 0.334 + ], + "angle": 0, + "content": "2 Lehigh University, Bethlehem PA 18015, USA {yuy214}@lehigh.edu" + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.334, + 0.674, + 0.362 + ], + "angle": 0, + "content": "3 IBM Research, Yorktown Heights NJ 10598, USA {tgao, adhuran}@us.ibm.com" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.4, + 0.74, + 0.65 + ], + "angle": 0, + "content": "Abstract. Identifying low-dimensional, semantic latent causal representations for high-dimensional data has become a dynamic field in computer vision and machine learning. Causal domain generalization methods aim to identify latent causal variables that generate input data and build invariant causal mechanisms for prediction tasks, thereby improving out-of-distribution (OOD) prediction performance. However, there is no consensus on the best approach for selecting causal variables for prediction. Existing methods typically choose causal or anti-causal variables, excluding other invariant, discriminative features. In this paper, we propose using Markov Blanket features due to their property of being the minimal set that possesses the maximum mutual information with the target. To achieve this, we establish a Causal Markov Blanket Representation Learning (CMBRL) framework, which allows for Markov Blanket discovery in the latent space. We then construct an invariant prediction mechanism using the identified Markov Blanket features, making it suitable for predictions across domains. Compared to state-of-the-art domain generalization methods, our approach exhibits robustness and adaptability under distribution shifts." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.663, + 0.741, + 0.69 + ], + "angle": 0, + "content": "Keywords: Causal Representation Learning \\(\\cdot\\) Markov Blanket Discovery \\(\\cdot\\) Domain Generalization" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.718, + 0.376, + 0.733 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Modern deep-learning models have significantly improved performance across various machine learning and computer vision applications in recent decades. However, they are also known for their limitations, including poor generalization, a lack of interpretability, and fairness concerns. [35] pinpointed the primary cause of the poor generalization of modern deep-learning models stems from the spurious correlations between irrelevant features and the prediction" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "N. Yin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.236 + ], + "angle": 0, + "content": "task through theoretical analysis and empirical evidence. These spurious correlations are caused by data biases and vary when distribution shifts, rendering the deep-learning models unreliable for the prediction of data from unseen domains. Hence, causal domain generalization methods propose to formulate prediction tasks from a causal perspective and exploit the robust, domain-invariant causality instead of correlations to improve the Out-of-distribution performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.241, + 0.788, + 0.468 + ], + "angle": 0, + "content": "Recent research in causal domain generalization [22,31-34,39,41,52] utilizes structural causal models (SCMs) to delve into the fundamental mechanisms underlying data generation. SCMs effectively capture intrinsic, stable, and interpretable causal relationships within a data distribution. These studies leverage SCMs to extract latent representations that are causally related to the target variable and subsequently develop invariant prediction mechanisms based on these representations. Key differences among existing methods stem from their choices of causally related features and their approaches to learning or selecting these features. However, there is no consensus on the optimal types of causally related features to be learned or selected. While many approaches [22,32-34,52] focus on identifying parent variables (causal features) relative to the target, [31] contends that anti-causal features (child variables) may offer greater robustness and predictive power, particularly in vision tasks. This paper theoretically examines the best causally related variables for prediction tasks and proposes an effective, potentially efficient method for identifying these variables." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.788, + 0.624 + ], + "angle": 0, + "content": "According to [12], the Markov Blanket (MB) of a variable is theoretically proven to be optimal for prediction tasks, as it is the minimal set that contains the maximum information about the target. Consequently, MB features have been extensively used for feature selection across various prediction tasks [4,38, 47]. Recent research [55] explores the use of MB features in the latent space for domain generalization. We find that, within the framework of our proposed SCM, the influence of spurious features is effectively mitigated by conditioning on the MB features. Specifically, the MB set can block all paths between the domain variable \\(U\\) and the target \\(Y\\), as illustrated in Figure 1. This characteristic makes MB features theoretically optimal for domain generalization prediction tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.788, + 0.777 + ], + "angle": 0, + "content": "We propose a framework for selecting Markov Blanket features from latent high-level variables that generate high-dimensional data. We refer to the chosen MB latent variables as Causal Markov Blanket (CMB) representations, which include the parents, children, and spouses of the target variable in the latent space. To obtain the CMB representations, we first obtain a set of identifiable latent variables using existing identifiable variational autoencoder framework. Then we propose an efficient MB discovery approach to identify the CMB representations collectively. A prediction model built using CMB representations is theoretically guaranteed to be both invariant and informative for predicting the target variable." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The main contributions of this paper are as follows: 1) We introduce a novel and general SCM to elucidate the data generation mechanisms underlying prediction tasks. 2) We establish a framework for discovering Markov Blanket features in the latent space. 3) We propose a three-phase algorithm for do" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.407, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Causal Markov Blanket Representation Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "main generalization prediction. We validate the effectiveness of our algorithm on benchmark datasets with distribution shifts, demonstrating improved generalization performance in the presence of these shifts." + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.214, + 0.396, + 0.229 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.243, + 0.784, + 0.273 + ], + "angle": 0, + "content": "In this section, we will review previous studies focusing on two key areas: Markov Blanket discovery and domain generalization (DG) prediction." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.286, + 0.785, + 0.467 + ], + "angle": 0, + "content": "Markov Blanket Discovery and Feature Selection. The Markov Blanket set of a variable consists of its immediate neighbors (parents, children, and spouses) in a causal graph. Popular MB discovery methods, including KS [21], HITON-MB [5], IPCMB [10], IAMB [48], and STMB [11], identify the Markov Blanket set through a series of independence tests, making MB discovery performance dependent on the accuracy of these tests. In terms of applications, multiple works [4, 12, 38, 47] have utilized MB discovery in the input space as a feature selection strategy, employing MB features for prediction. However, directly applying these methods to high-dimensional data, such as images, videos, or texts, is computationally expensive. Moreover, as noted by [32], causal and spurious features may not be disentangled in the input space since each dimension of the input data could be influenced by both types of features." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.48, + 0.785, + 0.646 + ], + "angle": 0, + "content": "Domain Generalization Approaches. We aim to employ Causal Markov Blanket representations to enhance domain generalization performance. To this end, we compare our method both theoretically and empirically with existing domain generalization approaches. Specifically, in a domain generalization setting, we use data from one or multiple source domains to make predictions on unseen target domains. Recent domain adaptation works [22,26,46] that require target domain data and intervention targets during training are beyond the scope of this paper. Popular domain generalization approaches include disentangled representation learning [19, 28, 29, 42], data augmentation [33], \"mix-up\" strategies [14, 56, 57], adversarial training [50, 51], and meta-learning [24, 43]. These methods typically do not involve causality." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.659, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Causal Domain Generalization Approaches. Compared to non-causal DG approaches, causal approaches make assumptions about data generation via SCMs and derive their algorithms accordingly. Generally, causal approaches can be categorized by their use of intervention. Methods without interventions include stable representation learning [7,8,17,18] and invariant feature learning [1-3, 6,23,40]. Stable representation learning methods acquire causal or anti-causal features through strategies like covariate balancing or by using SCM as a form of regularization. Invariant feature learning methods aim to identify domain-invariant causal features from multi-environmental data based on certain invariance criteria derived from causal theorems. One popular invariant feature learning approach is invariant risk minimization (IRM), which identifies invariant predictors corresponding to the parent variables of the target variable. In" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "N. Yin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.36 + ], + "angle": 0, + "content": "particular, [55] formulates an invariant causal Markov Blanket representation learning framework that is trained with causal constraints derived from the proposed SCM. Recently, a line of work [19, 20, 32] employs identifiable variational autoencoders (iVAEs) to obtain a set of latent variables with various degrees of identifiable guarantees, using the parent variables within this set for domain generalization prediction. We follow this approach and propose to learn the MB set of the target variable from the identifiable latent representations. Methods with interventions include robust feature learning through data augmentation [33] and interventional inference-guided methods [27, 34, 52]. These methods primarily address latent confounding issues between input data and the target. While our method generalizes across domains with different biases, it cannot yet effectively handle spurious correlations resulting from latent confounders. Extending our algorithm to handle latent confounders is an interesting future direction but is beyond the scope of this paper." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.389, + 0.609, + 0.406 + ], + "angle": 0, + "content": "3 Causal Analysis for Prediction Tasks" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.428, + 0.603, + 0.443 + ], + "angle": 0, + "content": "3.1 Preliminary on Markov Blanket Features" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.461, + 0.662, + 0.476 + ], + "angle": 0, + "content": "We outline the definition of Markov Blanket in Definition 1." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.49, + 0.788, + 0.538 + ], + "angle": 0, + "content": "Definition 1. ([37]) A Markov Blanket of a target variable \\(T\\) within the variable set \\(\\mathbf{V}\\), \\(\\mathbf{MB}_T\\), is the minimal set of nodes conditioned on which all other nodes are independent of \\(T\\), denoted as \\(O \\perp T|\\mathbf{MB}_T, \\forall O \\in \\{\\mathbf{V} \\backslash T \\backslash \\mathbf{MB}_T\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.626 + ], + "angle": 0, + "content": "The \\(\\mathbf{MB}_T\\) consists of the parent, child, and spouse variables of the target variable \\(T\\). According to the Theorem 1-3 in [12], the MB features of the target possess the following properties: they constitute the minimal set of features holding the maximal information about the target variable, and they ensure the least Bayes errors when predicting the target." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.788, + 0.732 + ], + "angle": 0, + "content": "Unlike prior works that use MB discovery for feature selection in the input space, we focus on MB selection in the latent space, guided by a structural causal model that we introduce for modeling the data generation process. To distinguish between them, we denote the MB set in the latent representation space as Causal Markov Blanket (CMB) representations. CMB representations are crucial for identifying the informative variables for predicting targets with higher accuracy despite distribution shifts." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.761, + 0.605, + 0.777 + ], + "angle": 0, + "content": "3.2 Assumptions on Structural Causal Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.794, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We formulate the prediction task under distribution shifts from a causal perspective. Using a structural causal model, we discern the underlying causal mechanisms of the data generation process, as illustrated in Figure 1." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.406, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Causal Markov Blanket Representation Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.146, + 0.541, + 0.403 + ], + "angle": 0, + "content": "\\(\\pmb{X} \\in \\mathbb{R}^{d}\\) represents the inputs and \\(Y\\) the prediction target. We denote \\(\\pmb{Z} = \\{Z_{1}, Z_{2}, \\dots, Z_{N}\\} \\in \\mathbb{R}^{N}\\) as the latent factors for generating input \\(\\pmb{X}\\). To better account for the distribution shifts, we introduce a domain variable \\(U\\) to encode domain-specific information4. Domain variable \\(U\\) is the common cause for some latent factors in \\(\\pmb{Z}\\) and hence can change the appearance of input \\(\\pmb{X}\\). Judged by their relations to the \\(Y\\), \\(\\pmb{Z}\\) can be further categorized into four types: parent variables \\(\\pmb{Z}_{p} = \\{Z_{p_{1}}, Z_{p_{2}}, \\dots\\}\\), child variables \\(\\pmb{Z}_{c} = \\{Z_{c_{1}}, Z_{c_{2}}, \\dots\\}\\), spouse variables \\(\\pmb{Z}_{s} = \\{Z_{s_{1}}, Z_{s_{2}}, \\dots\\}\\), and spurious variables \\(\\pmb{Z}_{o} = \\{Z_{o_{1}}, Z_{o_{2}}, \\dots\\}\\). \\(\\pmb{Z}_{p}, \\pmb{Z}_{c}\\) and \\(\\pmb{Z}_{s}\\) are the direct causes, direct effects, and the causes of the direct effects. They collectively form the" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.788, + 0.585 + ], + "angle": 0, + "content": "CMB representations \\( Z_{cmb} \\). \\( Z_o \\) are the variables that are spuriously correlated to target \\( Y \\) via \\( Z_{cmb} \\) variables or domain variable \\( U \\). To generalize our SCM, we allow for arbitrary causal relations within \\( Z \\) as long as the causal graph over \\( \\{X, Y, Z, U\\} \\) is a directed acyclic graph (DAG). We summarize the detailed assumptions in Assumption 1 of Appendix A.1. We show that our assumptions is practical and generally hold in real-world applications in Appendix A.2. Moreover, the SCM is Figure 1, combined with the assumptions, covers most scenarios from prior works [3, 32] and hence is a flexible model for performing causal analysis on prediction tasks. The semantic meanings of \\( Z_o \\) and \\( Z_{cmb} \\) are context-dependent and vary with the data. For instance, in CMNIST, \\( Z_o \\) represents background information like color, whereas \\( Z_{cmb} \\) denotes foreground information such as digit shape." + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.148, + 0.787, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.291, + 0.788, + 0.374 + ], + "angle": 0, + "content": "Fig.1: Proposed SCM over \\(\\{X,U,Z,Y\\}\\). Undirected links between \\(Z_{o}, Z_{p}\\) and \\(Z_{s}\\) indicate that both edge directions are permissible without violating the DAG constraint." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.614, + 0.527, + 0.628 + ], + "angle": 0, + "content": "3.3 Invariant Predictive Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.812 + ], + "angle": 0, + "content": "The conditional distribution \\( p(Y|\\mathbf{X}) \\), typically captured by a traditional classifier, is influenced by the domain variable \\( U \\), meaning that \\( U \\not\\perp Y|\\mathbf{X} \\). Consequently, the domain-variant distribution \\( p(Y|\\mathbf{X}) \\) is unsuitable for prediction in unseen domains. Prior causal approaches such as [22, 33, 34] and other content-style approaches make strong assumptions, assuming there is only a single type of causal feature (usually parent variables \\( Z_{p} \\)) within the SCM. Due to the simplicity of their SCMs, these methods fail when encountering scenarios where other types of causal features exist, as their selection of causal features cannot block all paths from \\( U \\) to \\( Y \\). For example, if the SCM only considers the existence of parent variables \\( Z_{p} \\) while \\( Z_{c} \\) also exists in practice, the domain variable \\( U \\) can reach \\( Y \\) via \\( U \\to Z_{o} \\to X \\to Z_{c} \\gets Y \\) when \\( X \\) is given." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.577, + 0.841 + ], + "angle": 0, + "content": "4 This is a standard setting in previous works [32, 39]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "N. Yin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "We advocate for using the Causal Markov Blanket features, denoted as \\( \\mathbf{Z}_{cmb} = \\{\\mathbf{Z}_p,\\mathbf{Z}_c,\\mathbf{Z}_s\\} \\), to predict the target in unseen domains. From Figure 1, we infer that \\( Y \\perp U|\\{\\mathbf{Z}_p,\\mathbf{Z}_c,\\mathbf{Z}_s\\} \\), implying that \\( p(Y|\\mathbf{Z}_{cmb}) \\) remains invariant across domains. Furthermore, we assert that the CMB features constitute the minimal sufficient set required for achieving invariant predictions across varying domains, as adding or removing variables from this set may affect its invariance. We demonstrate how to identify the CMB latent variables \\( \\mathbf{Z}_{cmb} \\) from the given data distribution and construct the invariant prediction mechanism \\( p(Y|\\mathbf{Z}_{cmb}) \\) for out-of-distribution prediction in Section 4." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.308, + 0.725, + 0.326 + ], + "angle": 0, + "content": "4 Causal Markov Blanket Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.388 + ], + "angle": 0, + "content": "We refer to our proposed method as CMBRL, representing Causal Markov Blanket Representation Learning. The training and inference procedures are introduced in Section 4.1 and Section 4.2, respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.411, + 0.424, + 0.427 + ], + "angle": 0, + "content": "4.1 Training Procedure" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.438, + 0.787, + 0.455 + ], + "angle": 0, + "content": "We illustrate the training procedure of the proposed algorithm in Figure 2. We" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.478, + 0.789, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.719, + 0.747, + 0.735 + ], + "angle": 0, + "content": "Fig. 2: The illustration of the three-phase training procedure of CMBRL." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "propose a three-phase training procedure. Initially, we utilize a VAE framework to learn a set of latent variables, which are proven to be component-wise identifiable. Subsequently, we conduct a CMB search using our novel mutual information quantification method to select the latent CMB representations. Finally, we construct a predictor using the selected CMB representations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.406, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Causal Markov Blanket Representation Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.222 + ], + "angle": 0, + "content": "Phase I: Identifiable Latent Variables Learning. We aim to obtain a set of latent variables \\( \\mathbf{Z} \\). To do so, we employ an existing (identifiable) variational autoencoder (VAE) framework [32]. We posit a prior distribution on \\( \\mathbf{Z} \\) that is consistent with our Assumption 1(b) and belongs to a general exponential family, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.222, + 0.645, + 0.243 + ], + "angle": 0, + "content": "\\[\np _ {\\boldsymbol {T}, \\boldsymbol {\\lambda}} (\\boldsymbol {Z} | Y, U) = \\frac {\\mathcal {Q} (\\boldsymbol {Z})}{\\mathcal {C} (Y , U)} \\exp [ \\boldsymbol {T} (\\boldsymbol {Z}) ^ {T} \\boldsymbol {\\lambda} (Y, U) ]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.249, + 0.785, + 0.385 + ], + "angle": 0, + "content": "where \\(\\mathcal{Q}\\) is the base measure. \\(\\mathcal{C}\\) is the normalizing constant. \\(\\lambda\\) is the arbitrary function. \\(T\\) is the sufficient statistics. To manage the increasing optimization challenges caused by additional parameters in the prior, we train the VAE framework using a standard protocol from prior work. [16,49]. As described in Eq. (1), the training objective comprises an ELBO loss \\(\\mathcal{L}_{\\mathrm{ELBO}}\\) and a score matching loss \\(\\mathcal{L}_{\\mathrm{SM}}\\). The ELBO loss \\(\\mathcal{L}_{\\mathrm{ELBO}}\\) optimizes over encoder and decoder parameters \\((\\theta, \\phi)\\), while the score matching loss \\(\\mathcal{L}_{\\mathrm{SM}}\\) minimizes over prior parameters \\((T, \\lambda)\\). The parameters \\(T, \\lambda\\) are constants in \\(\\mathcal{L}_{\\mathrm{ELBO}}\\), and \\(\\theta, \\phi\\) are constants in \\(\\mathcal{L}_{\\mathrm{SM}}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.311, + 0.386, + 0.786, + 0.405 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {o b j}} (\\boldsymbol {\\theta}, \\phi , \\boldsymbol {T}, \\lambda) := \\mathcal {L} _ {\\mathrm {E L B O}} (\\boldsymbol {\\theta}, \\phi , \\hat {\\boldsymbol {T}}, \\hat {\\lambda}) + \\mathcal {L} _ {\\mathrm {S M}} (\\hat {\\boldsymbol {\\theta}}, \\hat {\\phi}, \\boldsymbol {T}, \\lambda) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.412, + 0.767, + 0.428 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{\\mathrm{ELBO}}\\) and \\(\\mathcal{L}_{\\mathrm{SM}}\\) are outlined in Eq. (2). \\(p_{\\mathcal{D}}\\) is the training data distribution." + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.44, + 0.786, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {E L B O}} := - \\mathbb {E} _ {p _ {\\mathcal {D}}} \\left[ \\mathbb {E} _ {q _ {\\boldsymbol {\\theta}} (\\boldsymbol {z} | \\boldsymbol {x}, y, u)} [ \\log p _ {\\phi} (\\boldsymbol {x} | \\boldsymbol {z}) + \\log p _ {\\boldsymbol {T}, \\boldsymbol {\\lambda}} (\\boldsymbol {z} | y, u) - \\log q _ {\\boldsymbol {\\theta}} (\\boldsymbol {z} | \\boldsymbol {x}, y, u) ] \\right] \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.245, + 0.46, + 0.714, + 0.476 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S M}} := \\mathbb {E} _ {p _ {\\mathcal {D}}} \\left[ \\mathbb {E} _ {q _ {\\boldsymbol {\\theta}} (\\boldsymbol {z} | \\boldsymbol {x}, \\boldsymbol {y}, u)} [ \\| \\nabla_ {\\boldsymbol {z}} \\log q _ {\\boldsymbol {\\theta}} (\\boldsymbol {z} | \\boldsymbol {x}, \\boldsymbol {y}, u) - \\nabla_ {\\boldsymbol {z}} \\log p _ {\\boldsymbol {T}, \\boldsymbol {\\lambda}} (\\boldsymbol {z} | \\boldsymbol {y}, u) \\| ^ {2} ] \\right]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.655 + ], + "angle": 0, + "content": "As can be seen from the joint distribution \\( p(\\pmb{x}, y, u) \\propto \\int_{\\pmb{z}} p(\\pmb{x} | \\pmb{z}) p(\\pmb{z} | y, u) \\mathrm{d}\\pmb{z} \\), the prior \\( p(\\pmb{z} | y, u) \\) is consistent with our SCM. In fact, based on our SCM, \\( p(\\pmb{z} | y, u) \\) can be further decomposed into the product of conditional probabilities, which would result in further sparsity in \\( \\lambda \\). This property makes our learnt model differ from existed work based on different SCMs [19,20,32], as elaborated in Appendix B.4. However, the causal graph is generally unknown a priori, and hence we cannot pre-define the sparsity of \\( \\lambda \\). Therefore, in our learning algorithm we treat \\( p(\\pmb{z} | y, u) \\) as a generic form of prior that satisfies Assumption 1(b). Without a fully specified causal graph, we use this generic prior \\( p(\\pmb{z} | y, u) \\) to constrain the learning of VAE to obtain the \\( Z \\) without knowing their causal identities. We summarize the assumptions and identifiability results in Theorem 1." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.667, + 0.788, + 0.683 + ], + "angle": 0, + "content": "Theorem 1. Assume the data is sampled from a generative model described by" + }, + { + "type": "equation", + "bbox": [ + 0.251, + 0.696, + 0.748, + 0.713 + ], + "angle": 0, + "content": "\\[\np _ {\\boldsymbol {\\xi} = (\\boldsymbol {\\phi}, \\boldsymbol {T}, \\boldsymbol {\\lambda})} (\\boldsymbol {X}, \\boldsymbol {Z} | Y, U) = p _ {\\boldsymbol {\\phi}} (\\boldsymbol {X} | \\boldsymbol {Z}) p _ {\\boldsymbol {T}, \\boldsymbol {\\lambda}} (\\boldsymbol {Z} | Y, U), p _ {\\boldsymbol {\\phi}} (\\boldsymbol {X} | \\boldsymbol {Z}) = p _ {\\boldsymbol {\\epsilon}} (\\boldsymbol {X} - g _ {\\boldsymbol {\\phi}} (\\boldsymbol {Z}))\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.725, + 0.788, + 0.801 + ], + "angle": 0, + "content": "We have the following: (1) Under certain assumptions regarding domain variability, the model parameter \\(\\xi\\) is identifiable up to a permutation and component-wise transformation. (2) If (1) is true and \\(q_{\\theta}(z|\\mathbf{x},y,u)\\) is flexible and positive everywhere, then our framework learns the true parameters \\(\\xi^{*}\\) and true latent variables \\(\\mathbf{Z}^{*}\\) up to a permutation and component-wise transformation." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "5 Arbitrary function \\(\\lambda\\) and sufficient statistics \\(\\pmb{T}\\) are modeled by neural networks with ReLU activation due to their universal approximation ability." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "N. Yin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.314 + ], + "angle": 0, + "content": "We provide the detailed assumptions for (1) to hold and proof in Appendix B.26. These detailed assumptions are common in identifiable VAE methods and can impact the performance of the CMB discovery in Phase II. The component-wise identifiability of \\( Z \\) ensures that the CMB variables and spurious variables are separable, which is a crucial prerequisite for performing CMB discovery in Phase II, as verified by empirical results in Appendix F.1. While these assumptions may be violated in real-world applications, strict adherence to them is not always necessary. This motivates our investigation into identifiability performance on the CMNIST dataset in Section 5.1. Empirical results demonstrate that the VAE approach we employ can achieve a decent level of identifiability, even when certain assumptions do not hold." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.788, + 0.547 + ], + "angle": 0, + "content": "Phase II: Identify CMB Set. In this phase, we aim to identify the CMB representations relevant to \\( Y \\) among the latent variables \\( Z \\) we learned from the VAE framework. Traditional approaches to MB discovery involve revealing the local causal structure to the target through a series of (conditional) independence (CI) tests. However, conducting MB discovery in latent space encounters two major challenges: 1) the large dimension of the latent variable space, especially with complex and challenging input data, and 2) the mixture of variable types. These challenges can compromise the efficiency and accuracy of CMB discovery results. To tackle these two issues, we introduce an efficient search strategy and a more practical and general approach to conducting mutual-information-based (MI) independence tests. It is worth noting that we obtain CMB variables collectively by detecting and removing spurious variables with designed independence tests. We do not distinguish between the parent, child, and spouse variables within the CMB set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.777 + ], + "angle": 0, + "content": "Search Strategy: We find that learning the exact local graph of the target variable \\( Y \\) is unnecessary for identifying the CMB set. Instead of distinguishing among the CMB set, our objective is to separate spurious variables \\( Z_{o} \\) from \\( Z_{cmb} \\). In particularly, we discover that spurious variables \\( Z_{o} \\) and CMB variables \\( Z_{cmb} \\) possess different relations concerning \\( Y \\) conditioned on the remaining variables. If \\( Z_{i} \\in Z_{o} \\), we have \\( Z_{i} \\perp Y|\\{Z_{\\backslash i}, U\\} \\), where \\( Z_{\\backslash i} = \\{Z \\backslash Z_{i}\\} \\). If \\( Z_{i} \\in Z_{cmb} \\), we have \\( Z_{i} \\nsubseteq Y|\\{Z_{\\backslash i}, U\\} \\). We employ mutual information as measurement for independence tests. Theoretically, we can detect spurious variables and separate them from CMB set by quantifying \\( \\mathcal{I}(Z_{i}; Y|Z_{\\backslash i}, U), \\forall Z_{i} \\in Z \\). Actually, for real-world scenarios where \\( U \\) is unknown, it is challenging to estimate \\( \\mathcal{I}(Z_{i}; Y|Z_{\\backslash i}, U) \\). However, subject to our SCM structural assumptions, we have \\( \\mathcal{I}(Z_{i}; Y|Z_{\\backslash i}) = \\mathcal{I}(Z_{i}; Y|Z_{\\backslash i}, U) = 0 \\), if \\( Z_{i} \\in Z_{o} \\). We provide a proof in Appendix C.1. Meanwhile, if \\( Z_{i} \\in Z_{cmb} \\), \\( \\mathcal{I}(Z_{i}; Y|Z_{\\backslash i}) > 0 \\). Therefore, we can separate \\( Z_{o} \\) and \\( Z_{cmb} \\) by performing \\( \\mathcal{I}(Z_{i}; Y|Z_{\\backslash i}), \\forall Z_{i} \\in Z \\). We summarize such theoretical findings in Proposition 1." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.797, + 0.788, + 0.84 + ], + "angle": 0, + "content": "6 Our proof follows a similar approach to that of [32]. We aim to demonstrate that our assumptions on SCM do not violate the assumptions in the proof, thereby ensuring that the identifiable results hold under our framework." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.406, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Causal Markov Blanket Representation Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.195 + ], + "angle": 0, + "content": "Proposition 1. If the causal graph in Figure 1 and Assumption 1 holds, we have that: 1) for \\( Z_{i} \\in \\mathbf{Z}_{o} \\), \\( Z_{i} \\perp Y|Z_{\\backslash i} \\), \\( \\mathcal{I}(Z_{i};Y|Z_{\\backslash i}) = 0 \\); 2) for \\( Z_{i} \\in \\mathbf{Z}_{cmb} \\), \\( Z_{i} \\nsubseteq Y|Z_{\\backslash i} \\), \\( \\mathcal{I}(Z_{i};Y|Z_{\\backslash i}) > 0 \\)." + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.234, + 0.64, + 0.251 + ], + "angle": 0, + "content": "Algorithm 1 Causal Markov Blanket Set Search Strategy" + }, + { + "type": "algorithm", + "bbox": [ + 0.216, + 0.253, + 0.789, + 0.635 + ], + "angle": 0, + "content": "Input: M observations of Z and Y, i.e., \\(\\{z_{1}(j), z_{2}(j), \\dots, z_{N}(j), y(j)\\}_{j=1}^{M}\\) \nOutput: CMB set \\(Z_{cmb}\\) \nStep 1(Fast-Search): find the identity for each variable \\(Z_{i} \\in \\mathbf{Z}\\) \nInitial two sets \\(Z_{cmb} \\gets \\emptyset\\), \\(Z_{o} \\gets \\emptyset\\) \nfor \\(Z_{i} \\in \\mathbf{Z}, i = 1, 2, \\dots, N\\), do \n Perform Indtest \\((Z_{i}; Y|Z_{\\backslash i})\\) \n if \\(Z_{i} \\nsubseteq Y|Z_{\\backslash i}\\) then \n \\(Z_{cmb} \\gets Z_{cmb} \\cup Z_{i}\\) \n else \n \\(Z_{o} \\gets Z_{o} \\cup Z_{i}\\) \nend if \nend for \nStep 2 (Forward-Backward Search): verify the obtained CMB set. \nrepeat \n for \\(Z_{l} \\in Z_{cmb}\\) do \n ▷ Backward procedure \n Perform Indtest \\((Z_{l}; Y|\\{Z_{cmb}\\backslash Z_{l}\\})\\) \n if \\(Z_{l} \\perp Y|\\{Z_{cmb}\\backslash Z_{l}\\}\\) then \n \\(Z_{cmb} \\gets Z_{cmb} \\backslash Z_{l}\\) \n end if \nend for \nfor \\(Z_{j} \\in Z_{o}\\) do \n ▷ Forward procedure \n Perform Indtest \\((Z_{j}; Y|Z_{cmb})\\) \n if \\(Z_{j} \\nsubseteq Y|Z_{cmb}\\) then \n \\(Z_{cmb} \\gets Z_{cmb} \\cup Z_{j}\\) \n end if \nend for \nuntil Converge" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.662, + 0.696, + 0.677 + ], + "angle": 0, + "content": "We provide the detailed proof of Proposition 1 in Appendix C.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.678, + 0.788, + 0.784 + ], + "angle": 0, + "content": "We hence propose a fast-search strategy, which performs CI tests \\(\\mathcal{I}(Z_i;Y|\\mathbf{Z}_{\\backslash i})\\) for each variable \\(Z_{i}\\in \\mathbf{Z}\\). After performing the fast-search and obtaining the initial set of CMB, we verify the initial CMB set by repeatedly checking whether \\(\\forall Z_{j}\\in \\mathbf{Z}_{o}\\), \\(Z_{j}\\perp Y|Z_{cmb}\\) and \\(\\forall Z_{k}\\in \\mathbf{Z}_{cmb}\\), \\(Z_{k}\\nsubseteq Y|\\{\\mathbf{Z}_{cmb}\\backslash Z_{k}\\}\\). We outline our CMB search strategy in Algorithm 1. As illustrated by Proposition 2, our algorithm can correctly identify the CMB variables if the assumptions hold. We provide the detailed proof for Proposition 2 in Appendix C.3." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Proposition 2. Assume independence test results are correct and Assumption 1 for the proposed SCM holds, the ground truth Causal Markov Blanket set can be identified with Algorithm 1." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "N. Yin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "In Step 1, we leverage the structural properties of CMB features to efficiently distinguish them from spurious features. However, since we condition on \\( N - 1 \\) variables, which is a larger set than CMB, in step 1, we may have false positive detections and wrongly include spurious features. Thus, we employ Step 2, a forward-backward procedure, to eliminate possible errors in Step 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.788, + 0.328 + ], + "angle": 0, + "content": "Complexity Analysis of CMB Search: Assuming it takes \\( k \\) iterations, with \\( k << N \\) as observed in practice, for the verification to converge, our CMB strategy requires performing \\( (k + 1)N \\) CI tests and has a linear complexity of \\( \\mathcal{O}(N) \\). Compared to traditional MB discovery methods like IAMB [48], which directly execute the forward-backward search, our fast-search step reduces the search space and improves learning efficiency, often achieving convergence in just \\( k = 1 \\) or 2 iterations in practice." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.478 + ], + "angle": 0, + "content": "CI Tests: We use the MI-based, denoted as \\(\\mathcal{I}\\), independence tests, with a significance level \\(\\alpha\\). If \\(\\mathcal{I} < \\alpha\\), we declare independence. In Phase II, we need to compute the following mutual information \\(\\mathcal{I}(Y; Z_i | \\mathcal{C})\\), where \\(\\mathcal{C}\\) is the set of variables to condition on. We propose a practical approach that quantifies the MI using a trained predictor and is applicable for both classification tasks and regression tasks. Calculating mutual information for classification tasks is particularly challenging due to the mixture of variable types (where \\(Y\\) is discrete and variables in \\(\\mathbf{Z}\\) are continuous). We illustrate our approach to estimating the MI in Eq. (3) and Eq. (4). By definition of mutual information and conditional entropy, we have" + }, + { + "type": "equation", + "bbox": [ + 0.267, + 0.485, + 0.785, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {I} (y; z _ {i} | \\mathcal {C}) = \\mathcal {H} [ y | \\mathcal {C} ] - \\mathcal {H} [ y | \\{\\mathcal {C}, z _ {i} \\} ] \\\\ = \\mathbb {E} _ {p (\\mathcal {C})} \\left[ \\mathcal {H} [ p (y | \\mathcal {C}) ] \\right] - \\mathbb {E} _ {p (\\mathcal {C}, z _ {i})} \\left[ \\mathcal {H} [ p (y | \\mathcal {C}, z _ {i}) ] \\right] \\tag {3} \\\\ = \\mathbb {E} _ {p (\\mathcal {C})} \\left[ \\mathcal {H} \\left[ \\mathbb {E} _ {p (\\boldsymbol {z} \\backslash \\mathcal {C} | \\mathcal {C})} [ p (y | \\boldsymbol {z}) ] \\right] \\right] - \\mathbb {E} _ {p (\\mathcal {C}, z _ {i})} \\left[ \\mathcal {H} \\left[ \\mathbb {E} _ {p (\\boldsymbol {z} \\backslash \\{\\mathcal {C}, z _ {i} \\} | \\{\\mathcal {C}, z _ {i} \\})} [ p (y | \\boldsymbol {z}) ] \\right] \\right] \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.661 + ], + "angle": 0, + "content": "According to Eq. (3), we can calculate \\(\\mathcal{I}(y;z_i|\\mathcal{C})\\) using a predictor distribution \\(p(Y|\\mathbf{Z})\\) and samples of \\(\\mathbf{Z}\\). For one input and its label from the training data, i.e., \\(\\boldsymbol{x}(j), y(j), u(j) \\sim p_{\\mathcal{D}}\\), we use the mean of learned encoder \\(q_{\\hat{\\theta}}(\\mathbf{Z}|\\mathbf{X} = \\boldsymbol{x}(j), Y = y(j), U = u(j))\\) as the sample for \\(\\mathbf{Z}\\), denoted as \\(\\boldsymbol{z}(j) = [z_1(j), z_2(j), \\dots, z_N(j)]\\). After obtaining a total of \\(M\\) observations for \\(\\mathbf{Z}\\), i.e., \\(\\{\\boldsymbol{z}(j)\\}_{j=1}^{M}\\), we first learn a predictor \\(p(Y|\\mathbf{Z})\\) with \\(\\{\\boldsymbol{z}(j), y(j)\\}_{j=1}^{M}\\). Substituting the samples and learned distribution into Eq. (3), we have" + }, + { + "type": "equation", + "bbox": [ + 0.27, + 0.668, + 0.732, + 0.725 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {I} (y, z _ {i} | \\mathcal {C}) := - \\frac {1}{M} \\sum_ {j = 1} ^ {M} \\sum_ {l = 0} ^ {L - 1} \\left(\\frac {\\sum_ {k = 1} ^ {M} p \\big (y = l | \\mathcal {C} (j) , \\{z \\backslash \\mathcal {C} \\} (k) \\big)}{M}\\right) \\log \\left(\\frac {\\sum_ {k = 1} ^ {M} p \\big (y = l | \\mathcal {C} (j) , \\{z \\backslash \\mathcal {C} \\} (k) \\big)}{M}\\right) + \\\\ \\frac {1}{M} \\sum_ {j = 1} ^ {M} \\sum_ {l = 0} ^ {L - 1} \\left(\\frac {\\sum_ {k = 1} ^ {M} p (y = l | \\{\\mathcal {C} , z _ {i} \\} (j) , \\{z \\backslash (\\mathcal {C} , z _ {i}) \\} (k))}{M}\\right) \\log \\left(\\frac {\\sum_ {k = 1} ^ {M} p (y = l | \\{\\mathcal {C} , z _ {i} \\} (j) , \\{z \\backslash (\\mathcal {C} , z _ {i}) \\} (k))}{M}\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.732, + 0.787, + 0.765 + ], + "angle": 0, + "content": "where \\( L \\) is the number of classes for \\( Y \\). \\( l \\) denotes the value of \\( Y \\) and \\( l \\in \\{0,1,\\dots ,L - 1\\} \\). Especially, for the fast-search strategy, \\( \\mathcal{C} = \\mathbf{z}_{\\backslash i} \\), we have:" + }, + { + "type": "equation", + "bbox": [ + 0.247, + 0.773, + 0.735, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {I} (y; z _ {i} | \\boldsymbol {z} _ {\\backslash i}) = - \\frac {1}{M} \\sum_ {j = 1} ^ {M} \\sum_ {l = 0} ^ {L - 1} \\frac {\\sum_ {k = 1} ^ {M} p \\left(y = l | \\boldsymbol {z} _ {\\backslash i} (j) , z _ {i} (k)\\right)}{M} \\log \\frac {\\sum_ {k = 1} ^ {M} p \\left(y = l | \\boldsymbol {z} _ {\\backslash i} (j) , z _ {i} (k)\\right)}{M} \\\\ + \\frac {1}{M} \\sum_ {j = 1} ^ {M} \\sum_ {l = 0} ^ {L - 1} p (y = l | z (j)) \\log p (y = l | z (j)) \\\\ \\end{array}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.406, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Causal Markov Blanket Representation Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "The detailed derivation can be found in Appendix C.4. Our CMB discovery method assumes the accuracy of the IndTest. In practice, our approach, like other statistical tests, faces challenges with accuracy when the condition set is large and data is insufficient. Future studies aimed at improving the accuracy of independence tests could benefit from utilizing Bayesian mutual information [9]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.241, + 0.788, + 0.288 + ], + "angle": 0, + "content": "Phase III: Construct Invariant Predictor. We utilize the obtained CMB variables \\( Z_{cmb} \\) and train an invariant predictor \\( p(Y|Z_{cmb}) \\). Our training procedure is outlined in Eq. (6)." + }, + { + "type": "equation", + "bbox": [ + 0.297, + 0.294, + 0.786, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\hat {\\psi} = \\arg \\min _ {\\boldsymbol {\\psi}} \\frac {1}{M} \\sum_ {j = 1} ^ {M} \\mathcal {L} _ {\\operatorname {p r e d}} \\left(y (j), p _ {\\boldsymbol {\\psi}} \\left(\\boldsymbol {Y} \\mid \\boldsymbol {Z} _ {c m b} = \\boldsymbol {z} _ {c m b} (j)\\right)\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.324, + 0.787, + 0.355 + ], + "angle": 0, + "content": "where \\(\\psi\\) are the parameters of the predictor. \\(\\mathcal{L}_{\\mathrm{pred}}(\\cdot)\\) represents the cross-entropy loss for classification tasks and the mean squared loss for regression tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.374, + 0.431, + 0.388 + ], + "angle": 0, + "content": "4.2 Inference Procedure" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.395, + 0.788, + 0.517 + ], + "angle": 0, + "content": "In the inference procedure, we predict labels for data from an unseen test domain, \\( p_{\\mathcal{D}^{test}} \\). One can infer from Figure 1 that \\( p(\\pmb{X}|\\pmb{Z}) \\) is domain-invariant since \\( X \\perp U|Z \\). Therefore, inferring the latent variable values for inputs from an unseen test domain using \\( p(\\pmb{X}|\\pmb{Z}) \\) is reasonable. We adopt the standard inference procedure from [32, 45]. Given an input from the test domain, \\( \\pmb{x}^t \\sim p_{\\mathcal{D}^{test}} \\), we obtain the optimal values of \\( Z_{cmb} \\), denoted as \\( z_{cmb}^{*} \\), by solving the optimization problem outlined in Eq. (7). Here, \\( \\lambda_1 \\) and \\( \\lambda_2 \\) are hyperparameters used to control the scales of the learned \\( Z_{cmb} \\) and \\( Z_o \\)." + }, + { + "type": "equation", + "bbox": [ + 0.268, + 0.522, + 0.786, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {c m b} ^ {*}, \\boldsymbol {z} _ {o} ^ {*} = \\arg \\min _ {\\boldsymbol {z} _ {c m b}, \\boldsymbol {z} _ {o}} p _ {\\hat {\\phi}} (\\boldsymbol {x} ^ {t} | \\boldsymbol {z} _ {c m b}, \\boldsymbol {z} _ {o}) + \\lambda_ {1} \\| \\boldsymbol {z} _ {c m b} \\| _ {2} ^ {2} + \\lambda_ {2} \\| \\boldsymbol {z} _ {o} \\| _ {2} ^ {2} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.547, + 0.749, + 0.563 + ], + "angle": 0, + "content": "Then we can infer the label using the constructed invariant predictor, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.569, + 0.786, + 0.592 + ], + "angle": 0, + "content": "\\[\n\\hat {y} = \\arg \\max _ {y} p _ {\\hat {\\psi}} (y | \\boldsymbol {z} _ {c m b} ^ {*}) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.599, + 0.787, + 0.63 + ], + "angle": 0, + "content": "We provide the complexity analysis of the optimization in the inference procedure in Appendix D and the ablation study of runtime comparison in Appendix F.3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.649, + 0.377, + 0.667 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.677, + 0.788, + 0.797 + ], + "angle": 0, + "content": "We validate our CMBRL method for OOD prediction using synthetic and real benchmark distribution shift datasets against state-of-the-art domain generalization baselines. We experiment on four datasets: CMNIST, CelebA, PACS, and VLCS. Our experiments, conducted across 5 trials, are summarized in tables, reporting the mean and standard deviation of accuracy. The details of datasets and implementations can be found in Appendix E.1 and E.2. Additionally, detailed ablation studies on the selection of hyperparameters are available in Appendix F." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.816, + 0.538, + 0.832 + ], + "angle": 0, + "content": "5.1 Identifiability of Latent Variables" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "N. Yin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.538, + 0.357 + ], + "angle": 0, + "content": "First, we verify that Phase I of our proposed CMBRL yields identifiable \\( \\mathbf{Z} \\). We compare it with VAE [30], which lacks identifiability guarantees, and iVAE [19], which uses a conditionally factorized prior. Following standard procedures [19, 20, 22, 32], we evaluate identifiability on CMNIST data by computing the average mean correlation coefficient (MCC) between latent variables recovered by different models with various random initializations. Higher MCC scores indicate stronger identifiability, as shown in Figure 37. Using the same prior and training process as [32], we refer to our method as NF-" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.156, + 0.785, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.579, + 0.312, + 0.756, + 0.326 + ], + "angle": 0, + "content": "Fig.3: MCC on CMNIST" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.435 + ], + "angle": 0, + "content": "iVAE. Figure 3 demonstrates that NF-iVAE recovers latent variables with better identifiability. This identifiability impacts the disentanglement of spurious and CMB variables, influencing the accuracy of the estimated CMB set. Our ablation study in Appendix F.1 empirically shows that better identifiability leads to a more accurate CMB set." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.457, + 0.593, + 0.473 + ], + "angle": 0, + "content": "5.2 Out-of-distribution Prediction Accuracy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.484, + 0.494, + 0.711 + ], + "angle": 0, + "content": "We present both the in-distribution and OOD prediction accuracies of the CMNIST dataset in Table 1. First, we compare with the ERM trained on the entire training domain. Then, we compare with the popular causal invariant feature learning method, IRM, and its two variants: F-IRM (where \\(\\varPhi\\) is fixed to the identity) and V-IRM (where \\(\\varPhi\\) is variable). We also include the Robust MIN MAX method, which minimizes the maximum loss across multiple domains. Additionally, we compare with other approaches that leverage SCM knowledge to identify causal features" + }, + { + "type": "table_caption", + "bbox": [ + 0.503, + 0.528, + 0.787, + 0.556 + ], + "angle": 0, + "content": "Table 1: Results on Colored MNIST in terms of accuracy \\((\\%)\\)" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.558, + 0.774, + 0.681 + ], + "angle": 0, + "content": "
MethodsPrediction Acc (%)
In-distributionOOD
ERM85.7±0.510.3±0.2
Robust MIN MAX84.3±0.410.9±0.5
F-IRM GAME63.4±1.160.0±2.7
V-IRM GAME64.0±1.049.2±3.4
IRM59.3±4.462.8±9.6
Causalrep70.1±1.568.6±5.5
CTrans76.9±0.872.5±1.1
iCaRL870.6±0.868.8±1.5
CMBRL(ours)76.4±0.874.1±1.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.711, + 0.785, + 0.756 + ], + "angle": 0, + "content": "of the target, including Causalrep [53], CTrans [34], and iCaRL [32]. CTrans, in particular, considers latent confounding effects between spurious representation and target and employs causal inference via intervention." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.756, + 0.787, + 0.803 + ], + "angle": 0, + "content": "As shown in Table 1, our CMBRL method achieves optimal OOD performance while minimally compromising in-distribution accuracy. In contrast, the ERM method, which minimizes losses over the entire training domain, tends to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.841 + ], + "angle": 0, + "content": "7 We verify our NF-iVAE implementation's correctness by demonstrating similar identifiability performance to iCaRL in Figure 3" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.407, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Causal Markov Blanket Representation Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "capture spurious statistical correlations between labels and colors. Although it achieves the highest in-distribution accuracy, there is a significant drop in prediction accuracy on OOD data. Among the multiple domain-invariant causal representation learning methods, including IRM, F-IRM GAME, and V-IRM GAME, there are notable improvements. However, the limited number of available domains (only 2) may constrain the performance of the IRM methods. Conversely, causal representation learning methods operating under various Structural Causal Models (SCMs), including our CMBRL, demonstrate superior empirical performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.285, + 0.789, + 0.588 + ], + "angle": 0, + "content": "We further investigate the distinctions between the feature sets selected by iCaRL and our CMBRL. Following the approach described in [32], we first learn the causal graph between the estimated latent variables \\( Z \\) from NF-iVAE and \\( Y \\) using the PC algorithm. We then select the parent set with additional independent tests. The selected parent set yields an OOD prediction accuracy of \\( 69.8\\% \\), consistent with the reported results in Table 1. By applying a suitable significance level \\( \\alpha \\), we can select a CMB set that includes all the variables present in the selected parent set. The OOD prediction accuracy using such a CMB set increased to \\( 74.1\\% \\). This finding supports our assumptions on the proposed SCM and suggests the presence of variables not categorized as parent variables of the target, which can further enhance OOD prediction performance. Moreover, CTrans [34] and [22] assume there is an invariant latent confounder between \\( Y \\) and \\( Z_{o} \\), leading to spurious correlations between \\( Y \\) and \\( X \\). For example, in an image with a water bird and sea background, the confounder could include temperature and altitude, influencing the generation of the sea (spurious variable \\( Z_{o} \\)) and bird \\( (Y) \\). Our proposed SCM does not explicitly model latent confounders. Therefore, we cannot guarantee the generalization of our CMBRL method to scenarios involving unobserved confounders between input \\( X \\) and target \\( Y \\). While addressing the latent confounding issue falls outside the scope of this paper, it presents an intriguing avenue for future research. We sum" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.616, + 0.784, + 0.632 + ], + "angle": 0, + "content": "Table 2: Empirical results on CelebA in terms of the OOD prediction accuracy \\((\\%)\\)" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.644, + 0.785, + 0.673 + ], + "angle": 0, + "content": "
MethodsGroupDROMLDGCORALMMDDACNNMixupERMCTransCausalrepiCaRLCMBRL(ours)
Avg Acc65.165.666.766.268.769.560.766.360.471.576.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "marize empirical results on average OOD prediction accuracy over 7 settings for the CelebA dataset in Table 2. Detailed empirical results for each pair of settings can be found in Appendix H. We compare the results of our CMBRL with the ERM method and state-of-the-art causal and non-causal representation learning methods. As illustrated in Table 2, our method demonstrates optimal performance, surpassing the second-best baseline by an average margin of \\(4.7\\%\\) across all settings. This outcome indicates the efficacy of our CMBRL methods in excluding spurious variables and avoiding reliance on spurious correlations for prediction in computer vision applications." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "N. Yin et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 3: Empirical results on VLCS and PACS datasets in terms of OOD prediction accuracy (\\%) with a backbone of ResNet50." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.186, + 0.794, + 0.373 + ], + "angle": 0, + "content": "
AlgorithmsVLCSPACS
CLSVAvgACPSAvg
ERM98.0±0.462.6±0.970.8 ±1.977.5 ±1.977.284.8±1.376.4±1.196.7±0.676.1±1.083.5
GroupDRO98.1±0.366.4 ±0.971.0 ±0.376.1 ±1.477.983.5±0.979.1±0.696.7±0.378.3±2.084.4
MLDG98.5±0.361.7±1.273.6±1.875.0±0.877.285.5±1.480.1±1.797.4±0.376.6±1.184.9
CORAL96.9±0.965.7±1.273.3±0.778.7±0.878.788.3±0.280.0±0.797.5±0.378.8±1.386.2
MMD98.3±0.165.6±0.769.7±1.075.7±0.977.386.1±1.479.4±0.996.6±0.276.5±0.784.6
RSC97.5±0.663.1±1.273.0±1.376.2±0.577.585.4±0.879.7±1.897.6±0.378.2±1.285.2
Mixup98.4±0.363.4±0.772.9±0.876.1±1.277.786.1±0.778.9±0.897.6±0.175.8±1.884.6
DANN98.5±1.364.9±1.372.6±1.478.7±1.778.286.4±0.877.4±0.897.3±0.473.5±2.383.6
CDANN97.6±0.665.2±0.873.4±1.476.9±0.578.384.6±1.875.5±0.996.8±0.373.5±0.682.6
MTL97.6±0.660.6±1.371.0±1.277.2±0.776.687.5±0.877.1±0.796.4±0.877.3±1.884.6
ARM97.2±0.562.7±1.70.6±0.675.8±0.976.686.8±0.676.8±0.797.4±0.379.3±1.285.1
IRM98.6±0.166.0 ±0.972.3 ±0.677.3 ±0.978.584.7±0.480.0±0.697.2±0.379.3±1.085.5
SagNet97.3±0.461.6±0.873.4±1.977.6±0.477.587.4±1.080.7±0.697.1±0.180.0±0.486.3
iCaRL----81.8----88.7
CMBRL (ours)98.7±0.271.2±0.177.1±0.482.1±0.182.389.2±0.785.3±1.297.7±0.584.1±0.689.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.788, + 0.556 + ], + "angle": 0, + "content": "We present the empirical results for PACS and VLCS regarding OOD prediction accuracy in Table 3. These datasets serve as benchmarks for domain generalization tasks. For a comprehensive comparison, we evaluate not only against causal-inspired methods but also against various other DG methods, including GroupDRO [41], MLDG [25], CORAL [44], RSC [15], Mixup [54], SageNet [36], and others. We adhere to the experimental settings outlined in [13]. We report the OOD performance for each pair of train and test domains and the average accuracy across all combinations. As depicted in Table 3, our CMBRL method establishes state-of-the-art performance compared to popular alternatives in the domain generalization landscape." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.581, + 0.36, + 0.597 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.84 + ], + "angle": 0, + "content": "In this work, we investigate the Causal Markov Blanket discovery for high-dimensional data. We establish a framework guided by an SCM describing the data generation process, allowing for the Causal Markov Blanket discovery in latent space. We then construct an invariant prediction mechanism utilizing the CMB representations, which is suitable for OOD prediction. This framework can be further employed for causal analysis, reasoning, and intervention. We propose a three-phase algorithm to disentangle CMB features from spurious ones, reducing the risk of the prediction model relying on spurious correlations. It is worth noting that our method requires the satisfaction of assumptions on SCM and the assurance of the identifiability of latent variables. It may also suffer from inefficiencies due to the three training phases and optimization during inference. However, our method demonstrates its effectiveness by achieving significant OOD prediction performance, surpassing state-of-the-art causal representation learning methods and domain generalization methods on multiple benchmark distribution shift datasets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.407, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Causal Markov Blanket Representation Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.174, + 0.787, + 0.219 + ], + "angle": 0, + "content": "This work was supported in part by the National Science Foundation award IIS 2236026 and in part by IBM through the IBM-Rensselaer Future of Computing Research Collaboration." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.241, + 0.323, + 0.256 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.268, + 0.786, + 0.31 + ], + "angle": 0, + "content": "1. Ahuja, K., Caballero, E., Zhang, D., Bengio, Y., Mitlagkas, I., Rish, I.: Invariance principle meets information bottleneck for out-of-distribution generalization. In: Advances in Neural Inf. Proc. Systems (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.786, + 0.337 + ], + "angle": 0, + "content": "2. Ahuja, K., Shanmugam, K., Varshney, K., Dhurandhar, A.: Invariant risk minimization game. In: International Conference on Machine Learning (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.337, + 0.786, + 0.378 + ], + "angle": 0, + "content": "3. Ahuja, K., Wang, J., Dhurandhar, A., Shanmugam, K., Varshney, K.R.: Empirical or invariant risk minimization? a sample complexity perspective. In: International Conference on Learning Representations (2021) 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.786, + 0.433 + ], + "angle": 0, + "content": "4. Aliferis, C.F., Statnikov, A., Tsamardinos, I., Mani, S., Koutsoukos, X.D.: Local causal and markov blanket induction for causal discovery and feature selection for classification part i: algorithms and empirical evaluation. Journal of Machine Learning Research 11(1) (2010) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.786, + 0.473 + ], + "angle": 0, + "content": "5. Aliferis, C.F., Tsamardinos, I., Statnikov, A.: Hiton: a novel markov blanket algorithm for optimal variable selection. In: AMIA annual symposium proceedings. vol. 2003, p. 21. American Medical Informatics Association (2003) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.786, + 0.5 + ], + "angle": 0, + "content": "6. Arjovsky, M., Bottou, L., Gulrajani, I., Lopez-Paz, D.: Invariant risk minimization. arXiv preprint arXiv:1907.02893 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.786, + 0.539 + ], + "angle": 0, + "content": "7. Cui, P., Athey, S.: Stable learning establishes some common ground between causal inference and machine learning. Nature Machine Intelligence 4(2), 110-115 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.54, + 0.786, + 0.582 + ], + "angle": 0, + "content": "8. Cui, P., Shen, Z., Li, S., Yao, L., Li, Y., Chu, Z., Gao, J.: Causal inference meets machine learning. In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. pp. 3527-3528 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.786, + 0.623 + ], + "angle": 0, + "content": "9. Cui, Z., Yin, N., Wang, Y., Ji, Q.: Empirical bayesian approaches for robust constraint-based causal discovery under insufficient data. International Joint Conference on Artificial Intelligence (2022) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.786, + 0.678 + ], + "angle": 0, + "content": "0. Fu, S., Desmarais, M.C.: Fast markov blanket discovery algorithm via local learning within single pass. In: Advances in Artificial Intelligence: 21st Conference of the Canadian Society for Computational Studies of Intelligence, Canadian AI 2008 Windsor, Canada, May 28-30, 2008 Proceedings 21. pp. 96-107. Springer (2008) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.678, + 0.786, + 0.704 + ], + "angle": 0, + "content": "1. Gao, T., Ji, Q.: Local causal discovery of direct causes and effects. Advances in Neural Information Processing Systems 28 (2015) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.705, + 0.786, + 0.731 + ], + "angle": 0, + "content": "2. Gao, T., Wang, Z., Ji, Q.: Structured feature selection. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (December 2015) 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.786, + 0.758 + ], + "angle": 0, + "content": "3. Gulrajani, I., Lopez-Paz, D.: In search of lost domain generalization. arXiv preprint arXiv:2007.01434 (2020) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.786, + 0.799 + ], + "angle": 0, + "content": "4. Hendrycks, D., Mu, N., Cubuk, E.D., Zoph, B., Gilmer, J., Lakshminarayanan, B.: Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "5. Huang, Z., Wang, H., Xing, E.P., Huang, D.: Self-challenging improves cross-domain generalization. In: European Conference on Computer Vision. pp. 124-140. Springer (2020) 14" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.268, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.357, + 0.127 + ], + "angle": 0, + "content": "N. Yin et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Hyvarinen, A., Dayan, P.: Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research 6(4) (2005) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.202 + ], + "angle": 0, + "content": "17. Janzing, D.: Causal regularization. Advances in Neural Information Processing Systems 32 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.228 + ], + "angle": 0, + "content": "18. Jiang, Y., Veitch, V.: Invariant and transportable representations for anti-causal domain shifts. arXiv preprint arXiv:2207.01603 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.269 + ], + "angle": 0, + "content": "19. Khemakhem, I., Kingma, D., Monti, R., Hyvarinen, A.: Variational autoencoders and nonlinear ica: A unifying framework. In: International Conference on Artificial Intelligence and Statistics. pp. 2207-2217. PMLR (2020) 3, 4, 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.31 + ], + "angle": 0, + "content": "20. Khemakhem, I., Monti, R., Kingma, D., Hyvarinen, A.: Ice-beem: Identifiable conditional energy-based deep models based on nonlinear ica. Advances in Neural Information Processing Systems 33, 12768-12778 (2020) 4, 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.312, + 0.785, + 0.337 + ], + "angle": 0, + "content": "21. Koller, D., Sahami, M.: Toward optimal feature selection. Tech. rep., Stanford InfoLab (1996) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.377 + ], + "angle": 0, + "content": "22. Kong, L., Xie, S., Yao, W., Zheng, Y., Chen, G., Stojanov, P., Akinwande, V., Zhang, K.: Partial identifiability for domain adaptation. arXiv preprint arXiv:2306.06510 (2023) 2, 3, 5, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.404 + ], + "angle": 0, + "content": "23. Koyama, M., Yamaguchi, S.: When is invariance useful in an out-of-distribution generalization problem? arXiv preprint arXiv:2008.01883 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.785, + 0.445 + ], + "angle": 0, + "content": "24. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.: Learning to generalize: Meta-learning for domain generalization. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.486 + ], + "angle": 0, + "content": "25. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.M.: Deeper, broader and artier domain generalization. In: Proceedings of the IEEE international conference on computer vision. pp. 5542-5550 (2017) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.526 + ], + "angle": 0, + "content": "26. Li, Z., Cai, R., Chen, G., Sun, B., Hao, Z., Zhang, K.: Subspace identification for multi-source domain adaptation. Advances in Neural Information Processing Systems 36 (2024) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.58 + ], + "angle": 0, + "content": "27. Liu, B., Wang, D., Yang, X., Zhou, Y., Yao, R., Shao, Z., Zhao, J.: Show, deconfound and tell: Image captioning with causal inference. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18041-18050 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.635 + ], + "angle": 0, + "content": "28. Locatello, F., Bauer, S., Lucic, M., Raetsch, G., Gelly, S., Schölkopf, B., Bachem, O.: Challenging common assumptions in the unsupervised learning of disentangled representations. In: international conference on machine learning. pp. 4114-4124. PMLR (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.676 + ], + "angle": 0, + "content": "29. Locatello, F., Poole, B., Ratsch, G., Scholkopf, B., Bachem, O., Tschannen, M.: Weakly-supervised disentanglement without compromises. In: International Conference on Machine Learning. pp. 6348-6359. PMLR (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.717 + ], + "angle": 0, + "content": "30. Lopez, R., Regier, J., Jordan, M.I., Yosef, N.: Information constraints on auto-encoding variational bayes. Advances in neural information processing systems 31 (2018) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.719, + 0.785, + 0.758 + ], + "angle": 0, + "content": "31. Lopez-Paz, D., Nishihara, R., Chintala, S., Scholkopf, B., Bottou, L.: Discovering causal signals in images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6979-6987 (2017) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.76, + 0.785, + 0.798 + ], + "angle": 0, + "content": "32. Lu, C., Wu, Y., Hernández-Lobato, J.M., Schölkopf, B.: Invariant causal representation learning for out-of-distribution generalization. In: International Conference on Learning Representations (2021) 2, 3, 4, 5, 7, 8, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "33. Mao, C., Cha, A., Gupta, A., Wang, H., Yang, J., Vondrick, C.: Generative interventions for causal learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3947-3956 (2021) 2, 3, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.407, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Causal Markov Blanket Representation Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "34. Mao, C., Xia, K., Wang, J., Wang, H., Yang, J., Baireinboim, E., Vondrick, C.: Causal transportability for visual recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7521-7531 (2022) 2, 4, 5, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.787, + 0.232 + ], + "angle": 0, + "content": "35. Nagarajan, V., Andreassen, A., Neyshabur, B.: Understanding the failure modes of out-of-distribution generalization. arXiv preprint arXiv:2010.15775 (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.233, + 0.786, + 0.274 + ], + "angle": 0, + "content": "36. Nam, H., Lee, H., Park, J., Yoon, W., Yoo, D.: Reducing domain gap by reducing style bias. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8690-8699 (2021) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.276, + 0.786, + 0.302 + ], + "angle": 0, + "content": "37. Pearl, J.: Probabilistic reasoning in intelligent systems: networks of plausible inference. Morgan Kaufmann (1988) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.304, + 0.786, + 0.331 + ], + "angle": 0, + "content": "38. Pellet, J.P., Elisseeff, A.: Using markov blankets for causal structure learning. Journal of Machine Learning Research 9(7) (2008) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.333, + 0.786, + 0.36 + ], + "angle": 0, + "content": "39. Peters, J., Buhlmann, P., Meinshausen, N.: Causal inference using invariant prediction: identification and confidence intervals. arxiv. Methodology (2015) 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.361, + 0.786, + 0.388 + ], + "angle": 0, + "content": "40. Rosenfeld, E., Ravikumar, P., Risteski, A.: The risks of invariant risk minimization. In: International Conference on Learning Representations (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.39, + 0.786, + 0.431 + ], + "angle": 0, + "content": "41. Sagawa, S., Koh, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731 (2019) 2, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.432, + 0.786, + 0.472 + ], + "angle": 0, + "content": "42. Shen, X., Liu, F., Dong, H., Lian, Q., Chen, Z., Zhang, T.: Weakly supervised disentangled generative causal representation learning. The Journal of Machine Learning Research 23(1), 10994-11048 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.474, + 0.786, + 0.515 + ], + "angle": 0, + "content": "43. Shu, Y., Cao, Z., Wang, C., Wang, J., Long, M.: Open domain generalization with domain-augmented meta-learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9624-9633 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.517, + 0.786, + 0.571 + ], + "angle": 0, + "content": "44. Sun, B., Saenko, K.: Deep coral: Correlation alignment for deep domain adaptation. In: Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14. pp. 443-450. Springer (2016) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.573, + 0.786, + 0.6 + ], + "angle": 0, + "content": "45. Sun, X., Wu, B., Zheng, X., Liu, C., Chen, W., Qin, T., Liu, T.y.: Latent causal invariant model. arXiv preprint arXiv:2011.02203 (2020) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.601, + 0.786, + 0.642 + ], + "angle": 0, + "content": "46. Talon, D., Lippe, P., James, S., Del Bue, A., Magliacane, S.: Towards the reusability and compositionality of causal representations. arXiv preprint arXiv:2403.09830 (2024) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.643, + 0.786, + 0.684 + ], + "angle": 0, + "content": "47. Tan, Y., Liu, Z.: Feature selection and prediction with a markov blanket structure learning algorithm. In: BMC bioinformatics. vol. 14, pp. 1-3. BioMed Central (2013) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.686, + 0.786, + 0.727 + ], + "angle": 0, + "content": "48. Tsamardinos, I., Aliferis, C.F., Statnikov, A.R., Statnikov, E.: Algorithms for large scale markov blanket discovery. In: FLAIRS conference. vol. 2, pp. 376-380. St. Augustine, FL (2003) 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.728, + 0.786, + 0.755 + ], + "angle": 0, + "content": "49. Vincent, P.: A connection between score matching and denoising autoencoders. Neural computation 23(7), 1661-1674 (2011) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.757, + 0.786, + 0.798 + ], + "angle": 0, + "content": "50. Volpi, R., Namkoong, H., Sener, O., Duchi, J.C., Murino, V., Savarese, S.: Generalizing to unseen domains via adversarial data augmentation. Advances in neural information processing systems 31 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "51. Wang, H., Xiao, C., Kossaifi, J., Yu, Z., Anandkumar, A., Wang, Z.: Augmax: Adversarial composition of random augmentations for robust training. Advances in neural information processing systems 34, 237-250 (2021) 3" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "N. Yin et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "52. Wang, T., Huang, J., Zhang, H., Sun, Q.: Visual commonsense r-cnn. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10760-10770 (2020) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.217 + ], + "angle": 0, + "content": "53. Wang, Y., Jordan, M.I.: Desiderata for representation learning: A causal perspective. arXiv preprint arXiv:2109.03795 (2021) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "54. Yan, S., Song, H., Li, N., Zou, L., Ren, L.: Improve unsupervised domain adaptation with mixup training. arXiv preprint arXiv:2001.00677 (2020) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.286 + ], + "angle": 0, + "content": "55. Yin, N., Wang, H., Gao, T., Dhurandhar, A., Ji, Q.: Causal markov blanket representation learning for out-of-distribution generalization. In: Causal Representation Learning Workshop at NeurIPS 2023 (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.328 + ], + "angle": 0, + "content": "56. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6023-6032 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.785, + 0.356 + ], + "angle": 0, + "content": "57. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017) 3" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.356 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization/bb833fed-6970-4db4-b3da-646e63e5493d_origin.pdf b/2024/Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization/bb833fed-6970-4db4-b3da-646e63e5493d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d09a4ee2c6db42d0f0322b8c6589fcf37e0bd02f --- /dev/null +++ b/2024/Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization/bb833fed-6970-4db4-b3da-646e63e5493d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69244be8b35d0f1efc415c3aaf1045aceb036dc7bf1668c8e59cd77edb28432d +size 868088 diff --git a/2024/Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization/full.md b/2024/Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..54ea756da1c1f29330d82b72fd9cff49966fd4e7 --- /dev/null +++ b/2024/Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization/full.md @@ -0,0 +1,320 @@ +# Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization + +Naiyu Yin $^{1}$ , Hanjing Wang $^{1}$ , Yue Yu $^{2}$ , Tian Gao $^{3}$ , Amit Dhurandhar $^{3}$ , and Qiang Ji $^{1}$ + +1 Rensselaer Polytechnic Institute, Troy NY 12180, USA {yinn2, wangh36, jiq}@rpi.edu + +2 Lehigh University, Bethlehem PA 18015, USA {yuy214}@lehigh.edu + +3 IBM Research, Yorktown Heights NJ 10598, USA {tgao, adhuran}@us.ibm.com + +Abstract. Identifying low-dimensional, semantic latent causal representations for high-dimensional data has become a dynamic field in computer vision and machine learning. Causal domain generalization methods aim to identify latent causal variables that generate input data and build invariant causal mechanisms for prediction tasks, thereby improving out-of-distribution (OOD) prediction performance. However, there is no consensus on the best approach for selecting causal variables for prediction. Existing methods typically choose causal or anti-causal variables, excluding other invariant, discriminative features. In this paper, we propose using Markov Blanket features due to their property of being the minimal set that possesses the maximum mutual information with the target. To achieve this, we establish a Causal Markov Blanket Representation Learning (CMBRL) framework, which allows for Markov Blanket discovery in the latent space. We then construct an invariant prediction mechanism using the identified Markov Blanket features, making it suitable for predictions across domains. Compared to state-of-the-art domain generalization methods, our approach exhibits robustness and adaptability under distribution shifts. + +Keywords: Causal Representation Learning $\cdot$ Markov Blanket Discovery $\cdot$ Domain Generalization + +# 1 Introduction + +Modern deep-learning models have significantly improved performance across various machine learning and computer vision applications in recent decades. However, they are also known for their limitations, including poor generalization, a lack of interpretability, and fairness concerns. [35] pinpointed the primary cause of the poor generalization of modern deep-learning models stems from the spurious correlations between irrelevant features and the prediction + +task through theoretical analysis and empirical evidence. These spurious correlations are caused by data biases and vary when distribution shifts, rendering the deep-learning models unreliable for the prediction of data from unseen domains. Hence, causal domain generalization methods propose to formulate prediction tasks from a causal perspective and exploit the robust, domain-invariant causality instead of correlations to improve the Out-of-distribution performance. + +Recent research in causal domain generalization [22,31-34,39,41,52] utilizes structural causal models (SCMs) to delve into the fundamental mechanisms underlying data generation. SCMs effectively capture intrinsic, stable, and interpretable causal relationships within a data distribution. These studies leverage SCMs to extract latent representations that are causally related to the target variable and subsequently develop invariant prediction mechanisms based on these representations. Key differences among existing methods stem from their choices of causally related features and their approaches to learning or selecting these features. However, there is no consensus on the optimal types of causally related features to be learned or selected. While many approaches [22,32-34,52] focus on identifying parent variables (causal features) relative to the target, [31] contends that anti-causal features (child variables) may offer greater robustness and predictive power, particularly in vision tasks. This paper theoretically examines the best causally related variables for prediction tasks and proposes an effective, potentially efficient method for identifying these variables. + +According to [12], the Markov Blanket (MB) of a variable is theoretically proven to be optimal for prediction tasks, as it is the minimal set that contains the maximum information about the target. Consequently, MB features have been extensively used for feature selection across various prediction tasks [4,38, 47]. Recent research [55] explores the use of MB features in the latent space for domain generalization. We find that, within the framework of our proposed SCM, the influence of spurious features is effectively mitigated by conditioning on the MB features. Specifically, the MB set can block all paths between the domain variable $U$ and the target $Y$ , as illustrated in Figure 1. This characteristic makes MB features theoretically optimal for domain generalization prediction tasks. + +We propose a framework for selecting Markov Blanket features from latent high-level variables that generate high-dimensional data. We refer to the chosen MB latent variables as Causal Markov Blanket (CMB) representations, which include the parents, children, and spouses of the target variable in the latent space. To obtain the CMB representations, we first obtain a set of identifiable latent variables using existing identifiable variational autoencoder framework. Then we propose an efficient MB discovery approach to identify the CMB representations collectively. A prediction model built using CMB representations is theoretically guaranteed to be both invariant and informative for predicting the target variable. + +The main contributions of this paper are as follows: 1) We introduce a novel and general SCM to elucidate the data generation mechanisms underlying prediction tasks. 2) We establish a framework for discovering Markov Blanket features in the latent space. 3) We propose a three-phase algorithm for do + +main generalization prediction. We validate the effectiveness of our algorithm on benchmark datasets with distribution shifts, demonstrating improved generalization performance in the presence of these shifts. + +# 2 Related Works + +In this section, we will review previous studies focusing on two key areas: Markov Blanket discovery and domain generalization (DG) prediction. + +Markov Blanket Discovery and Feature Selection. The Markov Blanket set of a variable consists of its immediate neighbors (parents, children, and spouses) in a causal graph. Popular MB discovery methods, including KS [21], HITON-MB [5], IPCMB [10], IAMB [48], and STMB [11], identify the Markov Blanket set through a series of independence tests, making MB discovery performance dependent on the accuracy of these tests. In terms of applications, multiple works [4, 12, 38, 47] have utilized MB discovery in the input space as a feature selection strategy, employing MB features for prediction. However, directly applying these methods to high-dimensional data, such as images, videos, or texts, is computationally expensive. Moreover, as noted by [32], causal and spurious features may not be disentangled in the input space since each dimension of the input data could be influenced by both types of features. + +Domain Generalization Approaches. We aim to employ Causal Markov Blanket representations to enhance domain generalization performance. To this end, we compare our method both theoretically and empirically with existing domain generalization approaches. Specifically, in a domain generalization setting, we use data from one or multiple source domains to make predictions on unseen target domains. Recent domain adaptation works [22,26,46] that require target domain data and intervention targets during training are beyond the scope of this paper. Popular domain generalization approaches include disentangled representation learning [19, 28, 29, 42], data augmentation [33], "mix-up" strategies [14, 56, 57], adversarial training [50, 51], and meta-learning [24, 43]. These methods typically do not involve causality. + +Causal Domain Generalization Approaches. Compared to non-causal DG approaches, causal approaches make assumptions about data generation via SCMs and derive their algorithms accordingly. Generally, causal approaches can be categorized by their use of intervention. Methods without interventions include stable representation learning [7,8,17,18] and invariant feature learning [1-3, 6,23,40]. Stable representation learning methods acquire causal or anti-causal features through strategies like covariate balancing or by using SCM as a form of regularization. Invariant feature learning methods aim to identify domain-invariant causal features from multi-environmental data based on certain invariance criteria derived from causal theorems. One popular invariant feature learning approach is invariant risk minimization (IRM), which identifies invariant predictors corresponding to the parent variables of the target variable. In + +particular, [55] formulates an invariant causal Markov Blanket representation learning framework that is trained with causal constraints derived from the proposed SCM. Recently, a line of work [19, 20, 32] employs identifiable variational autoencoders (iVAEs) to obtain a set of latent variables with various degrees of identifiable guarantees, using the parent variables within this set for domain generalization prediction. We follow this approach and propose to learn the MB set of the target variable from the identifiable latent representations. Methods with interventions include robust feature learning through data augmentation [33] and interventional inference-guided methods [27, 34, 52]. These methods primarily address latent confounding issues between input data and the target. While our method generalizes across domains with different biases, it cannot yet effectively handle spurious correlations resulting from latent confounders. Extending our algorithm to handle latent confounders is an interesting future direction but is beyond the scope of this paper. + +# 3 Causal Analysis for Prediction Tasks + +# 3.1 Preliminary on Markov Blanket Features + +We outline the definition of Markov Blanket in Definition 1. + +Definition 1. ([37]) A Markov Blanket of a target variable $T$ within the variable set $\mathbf{V}$ , $\mathbf{MB}_T$ , is the minimal set of nodes conditioned on which all other nodes are independent of $T$ , denoted as $O \perp T|\mathbf{MB}_T, \forall O \in \{\mathbf{V} \backslash T \backslash \mathbf{MB}_T\}$ . + +The $\mathbf{MB}_T$ consists of the parent, child, and spouse variables of the target variable $T$ . According to the Theorem 1-3 in [12], the MB features of the target possess the following properties: they constitute the minimal set of features holding the maximal information about the target variable, and they ensure the least Bayes errors when predicting the target. + +Unlike prior works that use MB discovery for feature selection in the input space, we focus on MB selection in the latent space, guided by a structural causal model that we introduce for modeling the data generation process. To distinguish between them, we denote the MB set in the latent representation space as Causal Markov Blanket (CMB) representations. CMB representations are crucial for identifying the informative variables for predicting targets with higher accuracy despite distribution shifts. + +# 3.2 Assumptions on Structural Causal Model + +We formulate the prediction task under distribution shifts from a causal perspective. Using a structural causal model, we discern the underlying causal mechanisms of the data generation process, as illustrated in Figure 1. + +$\pmb{X} \in \mathbb{R}^{d}$ represents the inputs and $Y$ the prediction target. We denote $\pmb{Z} = \{Z_{1}, Z_{2}, \dots, Z_{N}\} \in \mathbb{R}^{N}$ as the latent factors for generating input $\pmb{X}$ . To better account for the distribution shifts, we introduce a domain variable $U$ to encode domain-specific information4. Domain variable $U$ is the common cause for some latent factors in $\pmb{Z}$ and hence can change the appearance of input $\pmb{X}$ . Judged by their relations to the $Y$ , $\pmb{Z}$ can be further categorized into four types: parent variables $\pmb{Z}_{p} = \{Z_{p_{1}}, Z_{p_{2}}, \dots\}$ , child variables $\pmb{Z}_{c} = \{Z_{c_{1}}, Z_{c_{2}}, \dots\}$ , spouse variables $\pmb{Z}_{s} = \{Z_{s_{1}}, Z_{s_{2}}, \dots\}$ , and spurious variables $\pmb{Z}_{o} = \{Z_{o_{1}}, Z_{o_{2}}, \dots\}$ . $\pmb{Z}_{p}, \pmb{Z}_{c}$ and $\pmb{Z}_{s}$ are the direct causes, direct effects, and the causes of the direct effects. They collectively form the + +CMB representations $Z_{cmb}$ . $Z_o$ are the variables that are spuriously correlated to target $Y$ via $Z_{cmb}$ variables or domain variable $U$ . To generalize our SCM, we allow for arbitrary causal relations within $Z$ as long as the causal graph over $\{X, Y, Z, U\}$ is a directed acyclic graph (DAG). We summarize the detailed assumptions in Assumption 1 of Appendix A.1. We show that our assumptions is practical and generally hold in real-world applications in Appendix A.2. Moreover, the SCM is Figure 1, combined with the assumptions, covers most scenarios from prior works [3, 32] and hence is a flexible model for performing causal analysis on prediction tasks. The semantic meanings of $Z_o$ and $Z_{cmb}$ are context-dependent and vary with the data. For instance, in CMNIST, $Z_o$ represents background information like color, whereas $Z_{cmb}$ denotes foreground information such as digit shape. + +![](images/2f950604a24249e491a85377114c2a8ce5377d99e34a25e2c679f69e768557aa.jpg) +Fig.1: Proposed SCM over $\{X,U,Z,Y\}$ . Undirected links between $Z_{o}, Z_{p}$ and $Z_{s}$ indicate that both edge directions are permissible without violating the DAG constraint. + +# 3.3 Invariant Predictive Mechanism + +The conditional distribution $p(Y|\mathbf{X})$ , typically captured by a traditional classifier, is influenced by the domain variable $U$ , meaning that $U \not\perp Y|\mathbf{X}$ . Consequently, the domain-variant distribution $p(Y|\mathbf{X})$ is unsuitable for prediction in unseen domains. Prior causal approaches such as [22, 33, 34] and other content-style approaches make strong assumptions, assuming there is only a single type of causal feature (usually parent variables $Z_{p}$ ) within the SCM. Due to the simplicity of their SCMs, these methods fail when encountering scenarios where other types of causal features exist, as their selection of causal features cannot block all paths from $U$ to $Y$ . For example, if the SCM only considers the existence of parent variables $Z_{p}$ while $Z_{c}$ also exists in practice, the domain variable $U$ can reach $Y$ via $U \to Z_{o} \to X \to Z_{c} \gets Y$ when $X$ is given. + +We advocate for using the Causal Markov Blanket features, denoted as $\mathbf{Z}_{cmb} = \{\mathbf{Z}_p,\mathbf{Z}_c,\mathbf{Z}_s\}$ , to predict the target in unseen domains. From Figure 1, we infer that $Y \perp U|\{\mathbf{Z}_p,\mathbf{Z}_c,\mathbf{Z}_s\}$ , implying that $p(Y|\mathbf{Z}_{cmb})$ remains invariant across domains. Furthermore, we assert that the CMB features constitute the minimal sufficient set required for achieving invariant predictions across varying domains, as adding or removing variables from this set may affect its invariance. We demonstrate how to identify the CMB latent variables $\mathbf{Z}_{cmb}$ from the given data distribution and construct the invariant prediction mechanism $p(Y|\mathbf{Z}_{cmb})$ for out-of-distribution prediction in Section 4. + +# 4 Causal Markov Blanket Representation Learning + +We refer to our proposed method as CMBRL, representing Causal Markov Blanket Representation Learning. The training and inference procedures are introduced in Section 4.1 and Section 4.2, respectively. + +# 4.1 Training Procedure + +We illustrate the training procedure of the proposed algorithm in Figure 2. We + +![](images/c91c943988c8d4d70045d70246392ede36e4dacb50c09e29971a61ca5f15662e.jpg) +Fig. 2: The illustration of the three-phase training procedure of CMBRL. + +propose a three-phase training procedure. Initially, we utilize a VAE framework to learn a set of latent variables, which are proven to be component-wise identifiable. Subsequently, we conduct a CMB search using our novel mutual information quantification method to select the latent CMB representations. Finally, we construct a predictor using the selected CMB representations. + +Phase I: Identifiable Latent Variables Learning. We aim to obtain a set of latent variables $\mathbf{Z}$ . To do so, we employ an existing (identifiable) variational autoencoder (VAE) framework [32]. We posit a prior distribution on $\mathbf{Z}$ that is consistent with our Assumption 1(b) and belongs to a general exponential family, i.e., + +$$ +p _ {\boldsymbol {T}, \boldsymbol {\lambda}} (\boldsymbol {Z} | Y, U) = \frac {\mathcal {Q} (\boldsymbol {Z})}{\mathcal {C} (Y , U)} \exp [ \boldsymbol {T} (\boldsymbol {Z}) ^ {T} \boldsymbol {\lambda} (Y, U) ] +$$ + +where $\mathcal{Q}$ is the base measure. $\mathcal{C}$ is the normalizing constant. $\lambda$ is the arbitrary function. $T$ is the sufficient statistics. To manage the increasing optimization challenges caused by additional parameters in the prior, we train the VAE framework using a standard protocol from prior work. [16,49]. As described in Eq. (1), the training objective comprises an ELBO loss $\mathcal{L}_{\mathrm{ELBO}}$ and a score matching loss $\mathcal{L}_{\mathrm{SM}}$ . The ELBO loss $\mathcal{L}_{\mathrm{ELBO}}$ optimizes over encoder and decoder parameters $(\theta, \phi)$ , while the score matching loss $\mathcal{L}_{\mathrm{SM}}$ minimizes over prior parameters $(T, \lambda)$ . The parameters $T, \lambda$ are constants in $\mathcal{L}_{\mathrm{ELBO}}$ , and $\theta, \phi$ are constants in $\mathcal{L}_{\mathrm{SM}}$ . + +$$ +\mathcal {L} _ {\mathrm {o b j}} (\boldsymbol {\theta}, \phi , \boldsymbol {T}, \lambda) := \mathcal {L} _ {\mathrm {E L B O}} (\boldsymbol {\theta}, \phi , \hat {\boldsymbol {T}}, \hat {\lambda}) + \mathcal {L} _ {\mathrm {S M}} (\hat {\boldsymbol {\theta}}, \hat {\phi}, \boldsymbol {T}, \lambda) \tag {1} +$$ + +$\mathcal{L}_{\mathrm{ELBO}}$ and $\mathcal{L}_{\mathrm{SM}}$ are outlined in Eq. (2). $p_{\mathcal{D}}$ is the training data distribution. + +$$ +\mathcal {L} _ {\text {E L B O}} := - \mathbb {E} _ {p _ {\mathcal {D}}} \left[ \mathbb {E} _ {q _ {\boldsymbol {\theta}} (\boldsymbol {z} | \boldsymbol {x}, y, u)} [ \log p _ {\phi} (\boldsymbol {x} | \boldsymbol {z}) + \log p _ {\boldsymbol {T}, \boldsymbol {\lambda}} (\boldsymbol {z} | y, u) - \log q _ {\boldsymbol {\theta}} (\boldsymbol {z} | \boldsymbol {x}, y, u) ] \right] \tag {2} +$$ + +$$ +\mathcal {L} _ {\mathrm {S M}} := \mathbb {E} _ {p _ {\mathcal {D}}} \left[ \mathbb {E} _ {q _ {\boldsymbol {\theta}} (\boldsymbol {z} | \boldsymbol {x}, \boldsymbol {y}, u)} [ \| \nabla_ {\boldsymbol {z}} \log q _ {\boldsymbol {\theta}} (\boldsymbol {z} | \boldsymbol {x}, \boldsymbol {y}, u) - \nabla_ {\boldsymbol {z}} \log p _ {\boldsymbol {T}, \boldsymbol {\lambda}} (\boldsymbol {z} | \boldsymbol {y}, u) \| ^ {2} ] \right] +$$ + +As can be seen from the joint distribution $p(\pmb{x}, y, u) \propto \int_{\pmb{z}} p(\pmb{x} | \pmb{z}) p(\pmb{z} | y, u) \mathrm{d}\pmb{z}$ , the prior $p(\pmb{z} | y, u)$ is consistent with our SCM. In fact, based on our SCM, $p(\pmb{z} | y, u)$ can be further decomposed into the product of conditional probabilities, which would result in further sparsity in $\lambda$ . This property makes our learnt model differ from existed work based on different SCMs [19,20,32], as elaborated in Appendix B.4. However, the causal graph is generally unknown a priori, and hence we cannot pre-define the sparsity of $\lambda$ . Therefore, in our learning algorithm we treat $p(\pmb{z} | y, u)$ as a generic form of prior that satisfies Assumption 1(b). Without a fully specified causal graph, we use this generic prior $p(\pmb{z} | y, u)$ to constrain the learning of VAE to obtain the $Z$ without knowing their causal identities. We summarize the assumptions and identifiability results in Theorem 1. + +Theorem 1. Assume the data is sampled from a generative model described by + +$$ +p _ {\boldsymbol {\xi} = (\boldsymbol {\phi}, \boldsymbol {T}, \boldsymbol {\lambda})} (\boldsymbol {X}, \boldsymbol {Z} | Y, U) = p _ {\boldsymbol {\phi}} (\boldsymbol {X} | \boldsymbol {Z}) p _ {\boldsymbol {T}, \boldsymbol {\lambda}} (\boldsymbol {Z} | Y, U), p _ {\boldsymbol {\phi}} (\boldsymbol {X} | \boldsymbol {Z}) = p _ {\boldsymbol {\epsilon}} (\boldsymbol {X} - g _ {\boldsymbol {\phi}} (\boldsymbol {Z})) +$$ + +We have the following: (1) Under certain assumptions regarding domain variability, the model parameter $\xi$ is identifiable up to a permutation and component-wise transformation. (2) If (1) is true and $q_{\theta}(z|\mathbf{x},y,u)$ is flexible and positive everywhere, then our framework learns the true parameters $\xi^{*}$ and true latent variables $\mathbf{Z}^{*}$ up to a permutation and component-wise transformation. + +We provide the detailed assumptions for (1) to hold and proof in Appendix B.26. These detailed assumptions are common in identifiable VAE methods and can impact the performance of the CMB discovery in Phase II. The component-wise identifiability of $Z$ ensures that the CMB variables and spurious variables are separable, which is a crucial prerequisite for performing CMB discovery in Phase II, as verified by empirical results in Appendix F.1. While these assumptions may be violated in real-world applications, strict adherence to them is not always necessary. This motivates our investigation into identifiability performance on the CMNIST dataset in Section 5.1. Empirical results demonstrate that the VAE approach we employ can achieve a decent level of identifiability, even when certain assumptions do not hold. + +Phase II: Identify CMB Set. In this phase, we aim to identify the CMB representations relevant to $Y$ among the latent variables $Z$ we learned from the VAE framework. Traditional approaches to MB discovery involve revealing the local causal structure to the target through a series of (conditional) independence (CI) tests. However, conducting MB discovery in latent space encounters two major challenges: 1) the large dimension of the latent variable space, especially with complex and challenging input data, and 2) the mixture of variable types. These challenges can compromise the efficiency and accuracy of CMB discovery results. To tackle these two issues, we introduce an efficient search strategy and a more practical and general approach to conducting mutual-information-based (MI) independence tests. It is worth noting that we obtain CMB variables collectively by detecting and removing spurious variables with designed independence tests. We do not distinguish between the parent, child, and spouse variables within the CMB set. + +Search Strategy: We find that learning the exact local graph of the target variable $Y$ is unnecessary for identifying the CMB set. Instead of distinguishing among the CMB set, our objective is to separate spurious variables $Z_{o}$ from $Z_{cmb}$ . In particularly, we discover that spurious variables $Z_{o}$ and CMB variables $Z_{cmb}$ possess different relations concerning $Y$ conditioned on the remaining variables. If $Z_{i} \in Z_{o}$ , we have $Z_{i} \perp Y|\{Z_{\backslash i}, U\}$ , where $Z_{\backslash i} = \{Z \backslash Z_{i}\}$ . If $Z_{i} \in Z_{cmb}$ , we have $Z_{i} \nsubseteq Y|\{Z_{\backslash i}, U\}$ . We employ mutual information as measurement for independence tests. Theoretically, we can detect spurious variables and separate them from CMB set by quantifying $\mathcal{I}(Z_{i}; Y|Z_{\backslash i}, U), \forall Z_{i} \in Z$ . Actually, for real-world scenarios where $U$ is unknown, it is challenging to estimate $\mathcal{I}(Z_{i}; Y|Z_{\backslash i}, U)$ . However, subject to our SCM structural assumptions, we have $\mathcal{I}(Z_{i}; Y|Z_{\backslash i}) = \mathcal{I}(Z_{i}; Y|Z_{\backslash i}, U) = 0$ , if $Z_{i} \in Z_{o}$ . We provide a proof in Appendix C.1. Meanwhile, if $Z_{i} \in Z_{cmb}$ , $\mathcal{I}(Z_{i}; Y|Z_{\backslash i}) > 0$ . Therefore, we can separate $Z_{o}$ and $Z_{cmb}$ by performing $\mathcal{I}(Z_{i}; Y|Z_{\backslash i}), \forall Z_{i} \in Z$ . We summarize such theoretical findings in Proposition 1. + +Proposition 1. If the causal graph in Figure 1 and Assumption 1 holds, we have that: 1) for $Z_{i} \in \mathbf{Z}_{o}$ , $Z_{i} \perp Y|Z_{\backslash i}$ , $\mathcal{I}(Z_{i};Y|Z_{\backslash i}) = 0$ ; 2) for $Z_{i} \in \mathbf{Z}_{cmb}$ , $Z_{i} \nsubseteq Y|Z_{\backslash i}$ , $\mathcal{I}(Z_{i};Y|Z_{\backslash i}) > 0$ . + +Algorithm 1 Causal Markov Blanket Set Search Strategy +Input: M observations of Z and Y, i.e., $\{z_{1}(j), z_{2}(j), \dots, z_{N}(j), y(j)\}_{j=1}^{M}$ +Output: CMB set $Z_{cmb}$ +Step 1(Fast-Search): find the identity for each variable $Z_{i} \in \mathbf{Z}$ +Initial two sets $Z_{cmb} \gets \emptyset$ , $Z_{o} \gets \emptyset$ +for $Z_{i} \in \mathbf{Z}, i = 1, 2, \dots, N$ , do + Perform Indtest $(Z_{i}; Y|Z_{\backslash i})$ + if $Z_{i} \nsubseteq Y|Z_{\backslash i}$ then + $Z_{cmb} \gets Z_{cmb} \cup Z_{i}$ + else + $Z_{o} \gets Z_{o} \cup Z_{i}$ +end if +end for +Step 2 (Forward-Backward Search): verify the obtained CMB set. +repeat + for $Z_{l} \in Z_{cmb}$ do + ▷ Backward procedure + Perform Indtest $(Z_{l}; Y|\{Z_{cmb}\backslash Z_{l}\})$ + if $Z_{l} \perp Y|\{Z_{cmb}\backslash Z_{l}\}$ then + $Z_{cmb} \gets Z_{cmb} \backslash Z_{l}$ + end if +end for +for $Z_{j} \in Z_{o}$ do + ▷ Forward procedure + Perform Indtest $(Z_{j}; Y|Z_{cmb})$ + if $Z_{j} \nsubseteq Y|Z_{cmb}$ then + $Z_{cmb} \gets Z_{cmb} \cup Z_{j}$ + end if +end for +until Converge + +We provide the detailed proof of Proposition 1 in Appendix C.2. + +We hence propose a fast-search strategy, which performs CI tests $\mathcal{I}(Z_i;Y|\mathbf{Z}_{\backslash i})$ for each variable $Z_{i}\in \mathbf{Z}$ . After performing the fast-search and obtaining the initial set of CMB, we verify the initial CMB set by repeatedly checking whether $\forall Z_{j}\in \mathbf{Z}_{o}$ , $Z_{j}\perp Y|Z_{cmb}$ and $\forall Z_{k}\in \mathbf{Z}_{cmb}$ , $Z_{k}\nsubseteq Y|\{\mathbf{Z}_{cmb}\backslash Z_{k}\}$ . We outline our CMB search strategy in Algorithm 1. As illustrated by Proposition 2, our algorithm can correctly identify the CMB variables if the assumptions hold. We provide the detailed proof for Proposition 2 in Appendix C.3. + +Proposition 2. Assume independence test results are correct and Assumption 1 for the proposed SCM holds, the ground truth Causal Markov Blanket set can be identified with Algorithm 1. + +In Step 1, we leverage the structural properties of CMB features to efficiently distinguish them from spurious features. However, since we condition on $N - 1$ variables, which is a larger set than CMB, in step 1, we may have false positive detections and wrongly include spurious features. Thus, we employ Step 2, a forward-backward procedure, to eliminate possible errors in Step 1. + +Complexity Analysis of CMB Search: Assuming it takes $k$ iterations, with $k << N$ as observed in practice, for the verification to converge, our CMB strategy requires performing $(k + 1)N$ CI tests and has a linear complexity of $\mathcal{O}(N)$ . Compared to traditional MB discovery methods like IAMB [48], which directly execute the forward-backward search, our fast-search step reduces the search space and improves learning efficiency, often achieving convergence in just $k = 1$ or 2 iterations in practice. + +CI Tests: We use the MI-based, denoted as $\mathcal{I}$ , independence tests, with a significance level $\alpha$ . If $\mathcal{I} < \alpha$ , we declare independence. In Phase II, we need to compute the following mutual information $\mathcal{I}(Y; Z_i | \mathcal{C})$ , where $\mathcal{C}$ is the set of variables to condition on. We propose a practical approach that quantifies the MI using a trained predictor and is applicable for both classification tasks and regression tasks. Calculating mutual information for classification tasks is particularly challenging due to the mixture of variable types (where $Y$ is discrete and variables in $\mathbf{Z}$ are continuous). We illustrate our approach to estimating the MI in Eq. (3) and Eq. (4). By definition of mutual information and conditional entropy, we have + +$$ +\begin{array}{l} \mathcal {I} (y; z _ {i} | \mathcal {C}) = \mathcal {H} [ y | \mathcal {C} ] - \mathcal {H} [ y | \{\mathcal {C}, z _ {i} \} ] \\ = \mathbb {E} _ {p (\mathcal {C})} \left[ \mathcal {H} [ p (y | \mathcal {C}) ] \right] - \mathbb {E} _ {p (\mathcal {C}, z _ {i})} \left[ \mathcal {H} [ p (y | \mathcal {C}, z _ {i}) ] \right] \tag {3} \\ = \mathbb {E} _ {p (\mathcal {C})} \left[ \mathcal {H} \left[ \mathbb {E} _ {p (\boldsymbol {z} \backslash \mathcal {C} | \mathcal {C})} [ p (y | \boldsymbol {z}) ] \right] \right] - \mathbb {E} _ {p (\mathcal {C}, z _ {i})} \left[ \mathcal {H} \left[ \mathbb {E} _ {p (\boldsymbol {z} \backslash \{\mathcal {C}, z _ {i} \} | \{\mathcal {C}, z _ {i} \})} [ p (y | \boldsymbol {z}) ] \right] \right] \\ \end{array} +$$ + +According to Eq. (3), we can calculate $\mathcal{I}(y;z_i|\mathcal{C})$ using a predictor distribution $p(Y|\mathbf{Z})$ and samples of $\mathbf{Z}$ . For one input and its label from the training data, i.e., $\boldsymbol{x}(j), y(j), u(j) \sim p_{\mathcal{D}}$ , we use the mean of learned encoder $q_{\hat{\theta}}(\mathbf{Z}|\mathbf{X} = \boldsymbol{x}(j), Y = y(j), U = u(j))$ as the sample for $\mathbf{Z}$ , denoted as $\boldsymbol{z}(j) = [z_1(j), z_2(j), \dots, z_N(j)]$ . After obtaining a total of $M$ observations for $\mathbf{Z}$ , i.e., $\{\boldsymbol{z}(j)\}_{j=1}^{M}$ , we first learn a predictor $p(Y|\mathbf{Z})$ with $\{\boldsymbol{z}(j), y(j)\}_{j=1}^{M}$ . Substituting the samples and learned distribution into Eq. (3), we have + +$$ +\begin{array}{l} \mathcal {I} (y, z _ {i} | \mathcal {C}) := - \frac {1}{M} \sum_ {j = 1} ^ {M} \sum_ {l = 0} ^ {L - 1} \left(\frac {\sum_ {k = 1} ^ {M} p \big (y = l | \mathcal {C} (j) , \{z \backslash \mathcal {C} \} (k) \big)}{M}\right) \log \left(\frac {\sum_ {k = 1} ^ {M} p \big (y = l | \mathcal {C} (j) , \{z \backslash \mathcal {C} \} (k) \big)}{M}\right) + \\ \frac {1}{M} \sum_ {j = 1} ^ {M} \sum_ {l = 0} ^ {L - 1} \left(\frac {\sum_ {k = 1} ^ {M} p (y = l | \{\mathcal {C} , z _ {i} \} (j) , \{z \backslash (\mathcal {C} , z _ {i}) \} (k))}{M}\right) \log \left(\frac {\sum_ {k = 1} ^ {M} p (y = l | \{\mathcal {C} , z _ {i} \} (j) , \{z \backslash (\mathcal {C} , z _ {i}) \} (k))}{M}\right) \\ \end{array} +$$ + +where $L$ is the number of classes for $Y$ . $l$ denotes the value of $Y$ and $l \in \{0,1,\dots ,L - 1\}$ . Especially, for the fast-search strategy, $\mathcal{C} = \mathbf{z}_{\backslash i}$ , we have: + +$$ +\begin{array}{l} \mathcal {I} (y; z _ {i} | \boldsymbol {z} _ {\backslash i}) = - \frac {1}{M} \sum_ {j = 1} ^ {M} \sum_ {l = 0} ^ {L - 1} \frac {\sum_ {k = 1} ^ {M} p \left(y = l | \boldsymbol {z} _ {\backslash i} (j) , z _ {i} (k)\right)}{M} \log \frac {\sum_ {k = 1} ^ {M} p \left(y = l | \boldsymbol {z} _ {\backslash i} (j) , z _ {i} (k)\right)}{M} \\ + \frac {1}{M} \sum_ {j = 1} ^ {M} \sum_ {l = 0} ^ {L - 1} p (y = l | z (j)) \log p (y = l | z (j)) \\ \end{array} +$$ + +The detailed derivation can be found in Appendix C.4. Our CMB discovery method assumes the accuracy of the IndTest. In practice, our approach, like other statistical tests, faces challenges with accuracy when the condition set is large and data is insufficient. Future studies aimed at improving the accuracy of independence tests could benefit from utilizing Bayesian mutual information [9]. + +Phase III: Construct Invariant Predictor. We utilize the obtained CMB variables $Z_{cmb}$ and train an invariant predictor $p(Y|Z_{cmb})$ . Our training procedure is outlined in Eq. (6). + +$$ +\hat {\psi} = \arg \min _ {\boldsymbol {\psi}} \frac {1}{M} \sum_ {j = 1} ^ {M} \mathcal {L} _ {\operatorname {p r e d}} \left(y (j), p _ {\boldsymbol {\psi}} \left(\boldsymbol {Y} \mid \boldsymbol {Z} _ {c m b} = \boldsymbol {z} _ {c m b} (j)\right)\right) \tag {6} +$$ + +where $\psi$ are the parameters of the predictor. $\mathcal{L}_{\mathrm{pred}}(\cdot)$ represents the cross-entropy loss for classification tasks and the mean squared loss for regression tasks. + +# 4.2 Inference Procedure + +In the inference procedure, we predict labels for data from an unseen test domain, $p_{\mathcal{D}^{test}}$ . One can infer from Figure 1 that $p(\pmb{X}|\pmb{Z})$ is domain-invariant since $X \perp U|Z$ . Therefore, inferring the latent variable values for inputs from an unseen test domain using $p(\pmb{X}|\pmb{Z})$ is reasonable. We adopt the standard inference procedure from [32, 45]. Given an input from the test domain, $\pmb{x}^t \sim p_{\mathcal{D}^{test}}$ , we obtain the optimal values of $Z_{cmb}$ , denoted as $z_{cmb}^{*}$ , by solving the optimization problem outlined in Eq. (7). Here, $\lambda_1$ and $\lambda_2$ are hyperparameters used to control the scales of the learned $Z_{cmb}$ and $Z_o$ . + +$$ +\boldsymbol {z} _ {c m b} ^ {*}, \boldsymbol {z} _ {o} ^ {*} = \arg \min _ {\boldsymbol {z} _ {c m b}, \boldsymbol {z} _ {o}} p _ {\hat {\phi}} (\boldsymbol {x} ^ {t} | \boldsymbol {z} _ {c m b}, \boldsymbol {z} _ {o}) + \lambda_ {1} \| \boldsymbol {z} _ {c m b} \| _ {2} ^ {2} + \lambda_ {2} \| \boldsymbol {z} _ {o} \| _ {2} ^ {2} \tag {7} +$$ + +Then we can infer the label using the constructed invariant predictor, i.e., + +$$ +\hat {y} = \arg \max _ {y} p _ {\hat {\psi}} (y | \boldsymbol {z} _ {c m b} ^ {*}) \tag {8} +$$ + +We provide the complexity analysis of the optimization in the inference procedure in Appendix D and the ablation study of runtime comparison in Appendix F.3. + +# 5 Experiments + +We validate our CMBRL method for OOD prediction using synthetic and real benchmark distribution shift datasets against state-of-the-art domain generalization baselines. We experiment on four datasets: CMNIST, CelebA, PACS, and VLCS. Our experiments, conducted across 5 trials, are summarized in tables, reporting the mean and standard deviation of accuracy. The details of datasets and implementations can be found in Appendix E.1 and E.2. Additionally, detailed ablation studies on the selection of hyperparameters are available in Appendix F. + +# 5.1 Identifiability of Latent Variables + +First, we verify that Phase I of our proposed CMBRL yields identifiable $\mathbf{Z}$ . We compare it with VAE [30], which lacks identifiability guarantees, and iVAE [19], which uses a conditionally factorized prior. Following standard procedures [19, 20, 22, 32], we evaluate identifiability on CMNIST data by computing the average mean correlation coefficient (MCC) between latent variables recovered by different models with various random initializations. Higher MCC scores indicate stronger identifiability, as shown in Figure 37. Using the same prior and training process as [32], we refer to our method as NF- + +![](images/b777fe0796829863924bd9511afa0e31472150f31cb974729e662398569b3db8.jpg) +Fig.3: MCC on CMNIST + +iVAE. Figure 3 demonstrates that NF-iVAE recovers latent variables with better identifiability. This identifiability impacts the disentanglement of spurious and CMB variables, influencing the accuracy of the estimated CMB set. Our ablation study in Appendix F.1 empirically shows that better identifiability leads to a more accurate CMB set. + +# 5.2 Out-of-distribution Prediction Accuracy + +We present both the in-distribution and OOD prediction accuracies of the CMNIST dataset in Table 1. First, we compare with the ERM trained on the entire training domain. Then, we compare with the popular causal invariant feature learning method, IRM, and its two variants: F-IRM (where $\varPhi$ is fixed to the identity) and V-IRM (where $\varPhi$ is variable). We also include the Robust MIN MAX method, which minimizes the maximum loss across multiple domains. Additionally, we compare with other approaches that leverage SCM knowledge to identify causal features + +Table 1: Results on Colored MNIST in terms of accuracy $(\%)$ + +
MethodsPrediction Acc (%)
In-distributionOOD
ERM85.7±0.510.3±0.2
Robust MIN MAX84.3±0.410.9±0.5
F-IRM GAME63.4±1.160.0±2.7
V-IRM GAME64.0±1.049.2±3.4
IRM59.3±4.462.8±9.6
Causalrep70.1±1.568.6±5.5
CTrans76.9±0.872.5±1.1
iCaRL870.6±0.868.8±1.5
CMBRL(ours)76.4±0.874.1±1.5
+ +of the target, including Causalrep [53], CTrans [34], and iCaRL [32]. CTrans, in particular, considers latent confounding effects between spurious representation and target and employs causal inference via intervention. + +As shown in Table 1, our CMBRL method achieves optimal OOD performance while minimally compromising in-distribution accuracy. In contrast, the ERM method, which minimizes losses over the entire training domain, tends to + +capture spurious statistical correlations between labels and colors. Although it achieves the highest in-distribution accuracy, there is a significant drop in prediction accuracy on OOD data. Among the multiple domain-invariant causal representation learning methods, including IRM, F-IRM GAME, and V-IRM GAME, there are notable improvements. However, the limited number of available domains (only 2) may constrain the performance of the IRM methods. Conversely, causal representation learning methods operating under various Structural Causal Models (SCMs), including our CMBRL, demonstrate superior empirical performance. + +We further investigate the distinctions between the feature sets selected by iCaRL and our CMBRL. Following the approach described in [32], we first learn the causal graph between the estimated latent variables $Z$ from NF-iVAE and $Y$ using the PC algorithm. We then select the parent set with additional independent tests. The selected parent set yields an OOD prediction accuracy of $69.8\%$ , consistent with the reported results in Table 1. By applying a suitable significance level $\alpha$ , we can select a CMB set that includes all the variables present in the selected parent set. The OOD prediction accuracy using such a CMB set increased to $74.1\%$ . This finding supports our assumptions on the proposed SCM and suggests the presence of variables not categorized as parent variables of the target, which can further enhance OOD prediction performance. Moreover, CTrans [34] and [22] assume there is an invariant latent confounder between $Y$ and $Z_{o}$ , leading to spurious correlations between $Y$ and $X$ . For example, in an image with a water bird and sea background, the confounder could include temperature and altitude, influencing the generation of the sea (spurious variable $Z_{o}$ ) and bird $(Y)$ . Our proposed SCM does not explicitly model latent confounders. Therefore, we cannot guarantee the generalization of our CMBRL method to scenarios involving unobserved confounders between input $X$ and target $Y$ . While addressing the latent confounding issue falls outside the scope of this paper, it presents an intriguing avenue for future research. We sum + +Table 2: Empirical results on CelebA in terms of the OOD prediction accuracy $(\%)$ + +
MethodsGroupDROMLDGCORALMMDDACNNMixupERMCTransCausalrepiCaRLCMBRL(ours)
Avg Acc65.165.666.766.268.769.560.766.360.471.576.2
+ +marize empirical results on average OOD prediction accuracy over 7 settings for the CelebA dataset in Table 2. Detailed empirical results for each pair of settings can be found in Appendix H. We compare the results of our CMBRL with the ERM method and state-of-the-art causal and non-causal representation learning methods. As illustrated in Table 2, our method demonstrates optimal performance, surpassing the second-best baseline by an average margin of $4.7\%$ across all settings. This outcome indicates the efficacy of our CMBRL methods in excluding spurious variables and avoiding reliance on spurious correlations for prediction in computer vision applications. + +Table 3: Empirical results on VLCS and PACS datasets in terms of OOD prediction accuracy (\%) with a backbone of ResNet50. + +
AlgorithmsVLCSPACS
CLSVAvgACPSAvg
ERM98.0±0.462.6±0.970.8 ±1.977.5 ±1.977.284.8±1.376.4±1.196.7±0.676.1±1.083.5
GroupDRO98.1±0.366.4 ±0.971.0 ±0.376.1 ±1.477.983.5±0.979.1±0.696.7±0.378.3±2.084.4
MLDG98.5±0.361.7±1.273.6±1.875.0±0.877.285.5±1.480.1±1.797.4±0.376.6±1.184.9
CORAL96.9±0.965.7±1.273.3±0.778.7±0.878.788.3±0.280.0±0.797.5±0.378.8±1.386.2
MMD98.3±0.165.6±0.769.7±1.075.7±0.977.386.1±1.479.4±0.996.6±0.276.5±0.784.6
RSC97.5±0.663.1±1.273.0±1.376.2±0.577.585.4±0.879.7±1.897.6±0.378.2±1.285.2
Mixup98.4±0.363.4±0.772.9±0.876.1±1.277.786.1±0.778.9±0.897.6±0.175.8±1.884.6
DANN98.5±1.364.9±1.372.6±1.478.7±1.778.286.4±0.877.4±0.897.3±0.473.5±2.383.6
CDANN97.6±0.665.2±0.873.4±1.476.9±0.578.384.6±1.875.5±0.996.8±0.373.5±0.682.6
MTL97.6±0.660.6±1.371.0±1.277.2±0.776.687.5±0.877.1±0.796.4±0.877.3±1.884.6
ARM97.2±0.562.7±1.70.6±0.675.8±0.976.686.8±0.676.8±0.797.4±0.379.3±1.285.1
IRM98.6±0.166.0 ±0.972.3 ±0.677.3 ±0.978.584.7±0.480.0±0.697.2±0.379.3±1.085.5
SagNet97.3±0.461.6±0.873.4±1.977.6±0.477.587.4±1.080.7±0.697.1±0.180.0±0.486.3
iCaRL----81.8----88.7
CMBRL (ours)98.7±0.271.2±0.177.1±0.482.1±0.182.389.2±0.785.3±1.297.7±0.584.1±0.689.1
+ +We present the empirical results for PACS and VLCS regarding OOD prediction accuracy in Table 3. These datasets serve as benchmarks for domain generalization tasks. For a comprehensive comparison, we evaluate not only against causal-inspired methods but also against various other DG methods, including GroupDRO [41], MLDG [25], CORAL [44], RSC [15], Mixup [54], SageNet [36], and others. We adhere to the experimental settings outlined in [13]. We report the OOD performance for each pair of train and test domains and the average accuracy across all combinations. As depicted in Table 3, our CMBRL method establishes state-of-the-art performance compared to popular alternatives in the domain generalization landscape. + +# 6 Conclusion + +In this work, we investigate the Causal Markov Blanket discovery for high-dimensional data. We establish a framework guided by an SCM describing the data generation process, allowing for the Causal Markov Blanket discovery in latent space. We then construct an invariant prediction mechanism utilizing the CMB representations, which is suitable for OOD prediction. This framework can be further employed for causal analysis, reasoning, and intervention. We propose a three-phase algorithm to disentangle CMB features from spurious ones, reducing the risk of the prediction model relying on spurious correlations. It is worth noting that our method requires the satisfaction of assumptions on SCM and the assurance of the identifiability of latent variables. It may also suffer from inefficiencies due to the three training phases and optimization during inference. However, our method demonstrates its effectiveness by achieving significant OOD prediction performance, surpassing state-of-the-art causal representation learning methods and domain generalization methods on multiple benchmark distribution shift datasets. + +# Acknowledgements + +This work was supported in part by the National Science Foundation award IIS 2236026 and in part by IBM through the IBM-Rensselaer Future of Computing Research Collaboration. + +# References + +1. Ahuja, K., Caballero, E., Zhang, D., Bengio, Y., Mitlagkas, I., Rish, I.: Invariance principle meets information bottleneck for out-of-distribution generalization. In: Advances in Neural Inf. Proc. Systems (2021) 3 +2. Ahuja, K., Shanmugam, K., Varshney, K., Dhurandhar, A.: Invariant risk minimization game. In: International Conference on Machine Learning (2020) 3 +3. Ahuja, K., Wang, J., Dhurandhar, A., Shanmugam, K., Varshney, K.R.: Empirical or invariant risk minimization? a sample complexity perspective. In: International Conference on Learning Representations (2021) 3, 5 +4. Aliferis, C.F., Statnikov, A., Tsamardinos, I., Mani, S., Koutsoukos, X.D.: Local causal and markov blanket induction for causal discovery and feature selection for classification part i: algorithms and empirical evaluation. Journal of Machine Learning Research 11(1) (2010) 2, 3 +5. Aliferis, C.F., Tsamardinos, I., Statnikov, A.: Hiton: a novel markov blanket algorithm for optimal variable selection. In: AMIA annual symposium proceedings. vol. 2003, p. 21. American Medical Informatics Association (2003) 3 +6. Arjovsky, M., Bottou, L., Gulrajani, I., Lopez-Paz, D.: Invariant risk minimization. arXiv preprint arXiv:1907.02893 (2019) 3 +7. Cui, P., Athey, S.: Stable learning establishes some common ground between causal inference and machine learning. Nature Machine Intelligence 4(2), 110-115 (2022) 3 +8. Cui, P., Shen, Z., Li, S., Yao, L., Li, Y., Chu, Z., Gao, J.: Causal inference meets machine learning. In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. pp. 3527-3528 (2020) 3 +9. Cui, Z., Yin, N., Wang, Y., Ji, Q.: Empirical bayesian approaches for robust constraint-based causal discovery under insufficient data. International Joint Conference on Artificial Intelligence (2022) 11 +0. Fu, S., Desmarais, M.C.: Fast markov blanket discovery algorithm via local learning within single pass. In: Advances in Artificial Intelligence: 21st Conference of the Canadian Society for Computational Studies of Intelligence, Canadian AI 2008 Windsor, Canada, May 28-30, 2008 Proceedings 21. pp. 96-107. Springer (2008) 3 +1. Gao, T., Ji, Q.: Local causal discovery of direct causes and effects. Advances in Neural Information Processing Systems 28 (2015) 3 +2. Gao, T., Wang, Z., Ji, Q.: Structured feature selection. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (December 2015) 2, 3, 4 +3. Gulrajani, I., Lopez-Paz, D.: In search of lost domain generalization. arXiv preprint arXiv:2007.01434 (2020) 14 +4. Hendrycks, D., Mu, N., Cubuk, E.D., Zoph, B., Gilmer, J., Lakshminarayanan, B.: Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781 (2019) 3 +5. Huang, Z., Wang, H., Xing, E.P., Huang, D.: Self-challenging improves cross-domain generalization. In: European Conference on Computer Vision. pp. 124-140. Springer (2020) 14 + +16. Hyvarinen, A., Dayan, P.: Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research 6(4) (2005) 7 +17. Janzing, D.: Causal regularization. Advances in Neural Information Processing Systems 32 (2019) 3 +18. Jiang, Y., Veitch, V.: Invariant and transportable representations for anti-causal domain shifts. arXiv preprint arXiv:2207.01603 (2022) 3 +19. Khemakhem, I., Kingma, D., Monti, R., Hyvarinen, A.: Variational autoencoders and nonlinear ica: A unifying framework. In: International Conference on Artificial Intelligence and Statistics. pp. 2207-2217. PMLR (2020) 3, 4, 7, 12 +20. Khemakhem, I., Monti, R., Kingma, D., Hyvarinen, A.: Ice-beem: Identifiable conditional energy-based deep models based on nonlinear ica. Advances in Neural Information Processing Systems 33, 12768-12778 (2020) 4, 7, 12 +21. Koller, D., Sahami, M.: Toward optimal feature selection. Tech. rep., Stanford InfoLab (1996) 3 +22. Kong, L., Xie, S., Yao, W., Zheng, Y., Chen, G., Stojanov, P., Akinwande, V., Zhang, K.: Partial identifiability for domain adaptation. arXiv preprint arXiv:2306.06510 (2023) 2, 3, 5, 12, 13 +23. Koyama, M., Yamaguchi, S.: When is invariance useful in an out-of-distribution generalization problem? arXiv preprint arXiv:2008.01883 (2020) 3 +24. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.: Learning to generalize: Meta-learning for domain generalization. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018) 3 +25. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.M.: Deeper, broader and artier domain generalization. In: Proceedings of the IEEE international conference on computer vision. pp. 5542-5550 (2017) 14 +26. Li, Z., Cai, R., Chen, G., Sun, B., Hao, Z., Zhang, K.: Subspace identification for multi-source domain adaptation. Advances in Neural Information Processing Systems 36 (2024) 3 +27. Liu, B., Wang, D., Yang, X., Zhou, Y., Yao, R., Shao, Z., Zhao, J.: Show, deconfound and tell: Image captioning with causal inference. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18041-18050 (2022) 4 +28. Locatello, F., Bauer, S., Lucic, M., Raetsch, G., Gelly, S., Schölkopf, B., Bachem, O.: Challenging common assumptions in the unsupervised learning of disentangled representations. In: international conference on machine learning. pp. 4114-4124. PMLR (2019) 3 +29. Locatello, F., Poole, B., Ratsch, G., Scholkopf, B., Bachem, O., Tschannen, M.: Weakly-supervised disentanglement without compromises. In: International Conference on Machine Learning. pp. 6348-6359. PMLR (2020) 3 +30. Lopez, R., Regier, J., Jordan, M.I., Yosef, N.: Information constraints on auto-encoding variational bayes. Advances in neural information processing systems 31 (2018) 12 +31. Lopez-Paz, D., Nishihara, R., Chintala, S., Scholkopf, B., Bottou, L.: Discovering causal signals in images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6979-6987 (2017) 2 +32. Lu, C., Wu, Y., Hernández-Lobato, J.M., Schölkopf, B.: Invariant causal representation learning for out-of-distribution generalization. In: International Conference on Learning Representations (2021) 2, 3, 4, 5, 7, 8, 11, 12, 13 +33. Mao, C., Cha, A., Gupta, A., Wang, H., Yang, J., Vondrick, C.: Generative interventions for causal learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3947-3956 (2021) 2, 3, 4, 5 + +34. Mao, C., Xia, K., Wang, J., Wang, H., Yang, J., Baireinboim, E., Vondrick, C.: Causal transportability for visual recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7521-7531 (2022) 2, 4, 5, 12, 13 +35. Nagarajan, V., Andreassen, A., Neyshabur, B.: Understanding the failure modes of out-of-distribution generalization. arXiv preprint arXiv:2010.15775 (2020) 1 +36. Nam, H., Lee, H., Park, J., Yoon, W., Yoo, D.: Reducing domain gap by reducing style bias. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8690-8699 (2021) 14 +37. Pearl, J.: Probabilistic reasoning in intelligent systems: networks of plausible inference. Morgan Kaufmann (1988) 4 +38. Pellet, J.P., Elisseeff, A.: Using markov blankets for causal structure learning. Journal of Machine Learning Research 9(7) (2008) 2, 3 +39. Peters, J., Buhlmann, P., Meinshausen, N.: Causal inference using invariant prediction: identification and confidence intervals. arxiv. Methodology (2015) 2, 5 +40. Rosenfeld, E., Ravikumar, P., Risteski, A.: The risks of invariant risk minimization. In: International Conference on Learning Representations (2021) 3 +41. Sagawa, S., Koh, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731 (2019) 2, 14 +42. Shen, X., Liu, F., Dong, H., Lian, Q., Chen, Z., Zhang, T.: Weakly supervised disentangled generative causal representation learning. The Journal of Machine Learning Research 23(1), 10994-11048 (2022) 3 +43. Shu, Y., Cao, Z., Wang, C., Wang, J., Long, M.: Open domain generalization with domain-augmented meta-learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9624-9633 (2021) 3 +44. Sun, B., Saenko, K.: Deep coral: Correlation alignment for deep domain adaptation. In: Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14. pp. 443-450. Springer (2016) 14 +45. Sun, X., Wu, B., Zheng, X., Liu, C., Chen, W., Qin, T., Liu, T.y.: Latent causal invariant model. arXiv preprint arXiv:2011.02203 (2020) 11 +46. Talon, D., Lippe, P., James, S., Del Bue, A., Magliacane, S.: Towards the reusability and compositionality of causal representations. arXiv preprint arXiv:2403.09830 (2024) 3 +47. Tan, Y., Liu, Z.: Feature selection and prediction with a markov blanket structure learning algorithm. In: BMC bioinformatics. vol. 14, pp. 1-3. BioMed Central (2013) 2, 3 +48. Tsamardinos, I., Aliferis, C.F., Statnikov, A.R., Statnikov, E.: Algorithms for large scale markov blanket discovery. In: FLAIRS conference. vol. 2, pp. 376-380. St. Augustine, FL (2003) 3, 10 +49. Vincent, P.: A connection between score matching and denoising autoencoders. Neural computation 23(7), 1661-1674 (2011) 7 +50. Volpi, R., Namkoong, H., Sener, O., Duchi, J.C., Murino, V., Savarese, S.: Generalizing to unseen domains via adversarial data augmentation. Advances in neural information processing systems 31 (2018) 3 +51. Wang, H., Xiao, C., Kossaifi, J., Yu, Z., Anandkumar, A., Wang, Z.: Augmax: Adversarial composition of random augmentations for robust training. Advances in neural information processing systems 34, 237-250 (2021) 3 + +52. Wang, T., Huang, J., Zhang, H., Sun, Q.: Visual commonsense r-cnn. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10760-10770 (2020) 2, 4 +53. Wang, Y., Jordan, M.I.: Desiderata for representation learning: A causal perspective. arXiv preprint arXiv:2109.03795 (2021) 12 +54. Yan, S., Song, H., Li, N., Zou, L., Ren, L.: Improve unsupervised domain adaptation with mixup training. arXiv preprint arXiv:2001.00677 (2020) 14 +55. Yin, N., Wang, H., Gao, T., Dhurandhar, A., Ji, Q.: Causal markov blanket representation learning for out-of-distribution generalization. In: Causal Representation Learning Workshop at NeurIPS 2023 (2023) 2, 4 +56. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6023-6032 (2019) 3 +57. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017) 3 \ No newline at end of file diff --git a/2024/Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization/images.zip b/2024/Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9c1b9611610adf7a5765357256ccf119f864198a --- /dev/null +++ b/2024/Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6cfa9b88fb586824479c58afc2b3fbe25d94065abfca8e060a93463f5daf5f82 +size 315066 diff --git a/2024/Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization/layout.json b/2024/Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1af41b0749725356b342d617bb7422b6cd25efa0 --- /dev/null +++ b/2024/Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization/layout.json @@ -0,0 +1,11314 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 154, + 111, + 460, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 111, + 460, + 163 + ], + "spans": [ + { + "bbox": [ + 154, + 111, + 460, + 163 + ], + "type": "text", + "content": "Integrating Markov Blanket Discovery into Causal Representation Learning for Domain Generalization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "text", + "content": "Naiyu Yin" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "text", + "content": ", Hanjing Wang" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "text", + "content": ", Yue Yu" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "text", + "content": ", Tian Gao" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "text", + "content": ", Amit Dhurandhar" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "text", + "content": ", and Qiang Ji" + }, + { + "bbox": [ + 133, + 185, + 479, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 189, + 219, + 423, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 219, + 423, + 242 + ], + "spans": [ + { + "bbox": [ + 189, + 219, + 423, + 242 + ], + "type": "text", + "content": "1 Rensselaer Polytechnic Institute, Troy NY 12180, USA {yinn2, wangh36, jiq}@rpi.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 209, + 242, + 405, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 242, + 405, + 264 + ], + "spans": [ + { + "bbox": [ + 209, + 242, + 405, + 264 + ], + "type": "text", + "content": "2 Lehigh University, Bethlehem PA 18015, USA {yuy214}@lehigh.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 201, + 264, + 412, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 264, + 412, + 286 + ], + "spans": [ + { + "bbox": [ + 201, + 264, + 412, + 286 + ], + "type": "text", + "content": "3 IBM Research, Yorktown Heights NJ 10598, USA {tgao, adhuran}@us.ibm.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 316, + 452, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 316, + 452, + 514 + ], + "spans": [ + { + "bbox": [ + 160, + 316, + 452, + 514 + ], + "type": "text", + "content": "Abstract. Identifying low-dimensional, semantic latent causal representations for high-dimensional data has become a dynamic field in computer vision and machine learning. Causal domain generalization methods aim to identify latent causal variables that generate input data and build invariant causal mechanisms for prediction tasks, thereby improving out-of-distribution (OOD) prediction performance. However, there is no consensus on the best approach for selecting causal variables for prediction. Existing methods typically choose causal or anti-causal variables, excluding other invariant, discriminative features. In this paper, we propose using Markov Blanket features due to their property of being the minimal set that possesses the maximum mutual information with the target. To achieve this, we establish a Causal Markov Blanket Representation Learning (CMBRL) framework, which allows for Markov Blanket discovery in the latent space. We then construct an invariant prediction mechanism using the identified Markov Blanket features, making it suitable for predictions across domains. Compared to state-of-the-art domain generalization methods, our approach exhibits robustness and adaptability under distribution shifts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 525, + 453, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 525, + 453, + 546 + ], + "spans": [ + { + "bbox": [ + 159, + 525, + 453, + 546 + ], + "type": "text", + "content": "Keywords: Causal Representation Learning " + }, + { + "bbox": [ + 159, + 525, + 453, + 546 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 525, + 453, + 546 + ], + "type": "text", + "content": " Markov Blanket Discovery " + }, + { + "bbox": [ + 159, + 525, + 453, + 546 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 525, + 453, + 546 + ], + "type": "text", + "content": " Domain Generalization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 568, + 230, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 230, + 580 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 230, + 580 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Modern deep-learning models have significantly improved performance across various machine learning and computer vision applications in recent decades. However, they are also known for their limitations, including poor generalization, a lack of interpretability, and fairness concerns. [35] pinpointed the primary cause of the poor generalization of modern deep-learning models stems from the spurious correlations between irrelevant features and the prediction" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": "task through theoretical analysis and empirical evidence. These spurious correlations are caused by data biases and vary when distribution shifts, rendering the deep-learning models unreliable for the prediction of data from unseen domains. Hence, causal domain generalization methods propose to formulate prediction tasks from a causal perspective and exploit the robust, domain-invariant causality instead of correlations to improve the Out-of-distribution performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 190, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 482, + 370 + ], + "type": "text", + "content": "Recent research in causal domain generalization [22,31-34,39,41,52] utilizes structural causal models (SCMs) to delve into the fundamental mechanisms underlying data generation. SCMs effectively capture intrinsic, stable, and interpretable causal relationships within a data distribution. These studies leverage SCMs to extract latent representations that are causally related to the target variable and subsequently develop invariant prediction mechanisms based on these representations. Key differences among existing methods stem from their choices of causally related features and their approaches to learning or selecting these features. However, there is no consensus on the optimal types of causally related features to be learned or selected. While many approaches [22,32-34,52] focus on identifying parent variables (causal features) relative to the target, [31] contends that anti-causal features (child variables) may offer greater robustness and predictive power, particularly in vision tasks. This paper theoretically examines the best causally related variables for prediction tasks and proposes an effective, potentially efficient method for identifying these variables." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 373, + 482, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 494 + ], + "type": "text", + "content": "According to [12], the Markov Blanket (MB) of a variable is theoretically proven to be optimal for prediction tasks, as it is the minimal set that contains the maximum information about the target. Consequently, MB features have been extensively used for feature selection across various prediction tasks [4,38, 47]. Recent research [55] explores the use of MB features in the latent space for domain generalization. We find that, within the framework of our proposed SCM, the influence of spurious features is effectively mitigated by conditioning on the MB features. Specifically, the MB set can block all paths between the domain variable " + }, + { + "bbox": [ + 130, + 373, + 482, + 494 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 373, + 482, + 494 + ], + "type": "text", + "content": " and the target " + }, + { + "bbox": [ + 130, + 373, + 482, + 494 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 373, + 482, + 494 + ], + "type": "text", + "content": ", as illustrated in Figure 1. This characteristic makes MB features theoretically optimal for domain generalization prediction tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 495, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 615 + ], + "type": "text", + "content": "We propose a framework for selecting Markov Blanket features from latent high-level variables that generate high-dimensional data. We refer to the chosen MB latent variables as Causal Markov Blanket (CMB) representations, which include the parents, children, and spouses of the target variable in the latent space. To obtain the CMB representations, we first obtain a set of identifiable latent variables using existing identifiable variational autoencoder framework. Then we propose an efficient MB discovery approach to identify the CMB representations collectively. A prediction model built using CMB representations is theoretically guaranteed to be both invariant and informative for predicting the target variable." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "The main contributions of this paper are as follows: 1) We introduce a novel and general SCM to elucidate the data generation mechanisms underlying prediction tasks. 2) We establish a framework for discovering Markov Blanket features in the latent space. 3) We propose a three-phase algorithm for do" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "N. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 479, + 152 + ], + "type": "text", + "content": "main generalization prediction. We validate the effectiveness of our algorithm on benchmark datasets with distribution shifts, demonstrating improved generalization performance in the presence of these shifts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 169, + 242, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 169, + 242, + 181 + ], + "spans": [ + { + "bbox": [ + 134, + 169, + 242, + 181 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 192, + 479, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 192, + 479, + 216 + ], + "spans": [ + { + "bbox": [ + 134, + 192, + 479, + 216 + ], + "type": "text", + "content": "In this section, we will review previous studies focusing on two key areas: Markov Blanket discovery and domain generalization (DG) prediction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 226, + 480, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 226, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 134, + 226, + 480, + 369 + ], + "type": "text", + "content": "Markov Blanket Discovery and Feature Selection. The Markov Blanket set of a variable consists of its immediate neighbors (parents, children, and spouses) in a causal graph. Popular MB discovery methods, including KS [21], HITON-MB [5], IPCMB [10], IAMB [48], and STMB [11], identify the Markov Blanket set through a series of independence tests, making MB discovery performance dependent on the accuracy of these tests. In terms of applications, multiple works [4, 12, 38, 47] have utilized MB discovery in the input space as a feature selection strategy, employing MB features for prediction. However, directly applying these methods to high-dimensional data, such as images, videos, or texts, is computationally expensive. Moreover, as noted by [32], causal and spurious features may not be disentangled in the input space since each dimension of the input data could be influenced by both types of features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 380, + 480, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 380, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 134, + 380, + 480, + 511 + ], + "type": "text", + "content": "Domain Generalization Approaches. We aim to employ Causal Markov Blanket representations to enhance domain generalization performance. To this end, we compare our method both theoretically and empirically with existing domain generalization approaches. Specifically, in a domain generalization setting, we use data from one or multiple source domains to make predictions on unseen target domains. Recent domain adaptation works [22,26,46] that require target domain data and intervention targets during training are beyond the scope of this paper. Popular domain generalization approaches include disentangled representation learning [19, 28, 29, 42], data augmentation [33], \"mix-up\" strategies [14, 56, 57], adversarial training [50, 51], and meta-learning [24, 43]. These methods typically do not involve causality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 521, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 521, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 521, + 480, + 665 + ], + "type": "text", + "content": "Causal Domain Generalization Approaches. Compared to non-causal DG approaches, causal approaches make assumptions about data generation via SCMs and derive their algorithms accordingly. Generally, causal approaches can be categorized by their use of intervention. Methods without interventions include stable representation learning [7,8,17,18] and invariant feature learning [1-3, 6,23,40]. Stable representation learning methods acquire causal or anti-causal features through strategies like covariate balancing or by using SCM as a form of regularization. Invariant feature learning methods aim to identify domain-invariant causal features from multi-environmental data based on certain invariance criteria derived from causal theorems. One popular invariant feature learning approach is invariant risk minimization (IRM), which identifies invariant predictors corresponding to the parent variables of the target variable. In" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 249, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 249, + 91, + 447, + 102 + ], + "type": "text", + "content": "Causal Markov Blanket Representation Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": "particular, [55] formulates an invariant causal Markov Blanket representation learning framework that is trained with causal constraints derived from the proposed SCM. Recently, a line of work [19, 20, 32] employs identifiable variational autoencoders (iVAEs) to obtain a set of latent variables with various degrees of identifiable guarantees, using the parent variables within this set for domain generalization prediction. We follow this approach and propose to learn the MB set of the target variable from the identifiable latent representations. Methods with interventions include robust feature learning through data augmentation [33] and interventional inference-guided methods [27, 34, 52]. These methods primarily address latent confounding issues between input data and the target. While our method generalizes across domains with different biases, it cannot yet effectively handle spurious correlations resulting from latent confounders. Extending our algorithm to handle latent confounders is an interesting future direction but is beyond the scope of this paper." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 308, + 372, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 308, + 372, + 321 + ], + "spans": [ + { + "bbox": [ + 131, + 308, + 372, + 321 + ], + "type": "text", + "content": "3 Causal Analysis for Prediction Tasks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 338, + 369, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 338, + 369, + 350 + ], + "spans": [ + { + "bbox": [ + 131, + 338, + 369, + 350 + ], + "type": "text", + "content": "3.1 Preliminary on Markov Blanket Features" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 365, + 405, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 365, + 405, + 376 + ], + "spans": [ + { + "bbox": [ + 131, + 365, + 405, + 376 + ], + "type": "text", + "content": "We outline the definition of Markov Blanket in Definition 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 388, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 388, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 388, + 482, + 426 + ], + "type": "text", + "content": "Definition 1. ([37]) A Markov Blanket of a target variable " + }, + { + "bbox": [ + 131, + 388, + 482, + 426 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 388, + 482, + 426 + ], + "type": "text", + "content": " within the variable set " + }, + { + "bbox": [ + 131, + 388, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 131, + 388, + 482, + 426 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 388, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{MB}_T" + }, + { + "bbox": [ + 131, + 388, + 482, + 426 + ], + "type": "text", + "content": ", is the minimal set of nodes conditioned on which all other nodes are independent of " + }, + { + "bbox": [ + 131, + 388, + 482, + 426 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 388, + 482, + 426 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 131, + 388, + 482, + 426 + ], + "type": "inline_equation", + "content": "O \\perp T|\\mathbf{MB}_T, \\forall O \\in \\{\\mathbf{V} \\backslash T \\backslash \\mathbf{MB}_T\\}" + }, + { + "bbox": [ + 131, + 388, + 482, + 426 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 434, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 495 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 130, + 434, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{MB}_T" + }, + { + "bbox": [ + 130, + 434, + 482, + 495 + ], + "type": "text", + "content": " consists of the parent, child, and spouse variables of the target variable " + }, + { + "bbox": [ + 130, + 434, + 482, + 495 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 434, + 482, + 495 + ], + "type": "text", + "content": ". According to the Theorem 1-3 in [12], the MB features of the target possess the following properties: they constitute the minimal set of features holding the maximal information about the target variable, and they ensure the least Bayes errors when predicting the target." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 495, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 579 + ], + "type": "text", + "content": "Unlike prior works that use MB discovery for feature selection in the input space, we focus on MB selection in the latent space, guided by a structural causal model that we introduce for modeling the data generation process. To distinguish between them, we denote the MB set in the latent representation space as Causal Markov Blanket (CMB) representations. CMB representations are crucial for identifying the informative variables for predicting targets with higher accuracy despite distribution shifts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 602, + 370, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 602, + 370, + 615 + ], + "spans": [ + { + "bbox": [ + 131, + 602, + 370, + 615 + ], + "type": "text", + "content": "3.2 Assumptions on Structural Causal Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 628, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 628, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 481, + 665 + ], + "type": "text", + "content": "We formulate the prediction task under distribution shifts from a causal perspective. Using a structural causal model, we discern the underlying causal mechanisms of the data generation process, as illustrated in Figure 1." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "N. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "inline_equation", + "content": "\\pmb{X} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "text", + "content": " represents the inputs and " + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "text", + "content": " the prediction target. We denote " + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "inline_equation", + "content": "\\pmb{Z} = \\{Z_{1}, Z_{2}, \\dots, Z_{N}\\} \\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "text", + "content": " as the latent factors for generating input " + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "inline_equation", + "content": "\\pmb{X}" + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "text", + "content": ". To better account for the distribution shifts, we introduce a domain variable " + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "text", + "content": " to encode domain-specific information4. Domain variable " + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "text", + "content": " is the common cause for some latent factors in " + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "inline_equation", + "content": "\\pmb{Z}" + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "text", + "content": " and hence can change the appearance of input " + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "inline_equation", + "content": "\\pmb{X}" + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "text", + "content": ". Judged by their relations to the " + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "inline_equation", + "content": "\\pmb{Z}" + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "text", + "content": " can be further categorized into four types: parent variables " + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "inline_equation", + "content": "\\pmb{Z}_{p} = \\{Z_{p_{1}}, Z_{p_{2}}, \\dots\\}" + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "text", + "content": ", child variables " + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "inline_equation", + "content": "\\pmb{Z}_{c} = \\{Z_{c_{1}}, Z_{c_{2}}, \\dots\\}" + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "text", + "content": ", spouse variables " + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "inline_equation", + "content": "\\pmb{Z}_{s} = \\{Z_{s_{1}}, Z_{s_{2}}, \\dots\\}" + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "text", + "content": ", and spurious variables " + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "inline_equation", + "content": "\\pmb{Z}_{o} = \\{Z_{o_{1}}, Z_{o_{2}}, \\dots\\}" + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "inline_equation", + "content": "\\pmb{Z}_{p}, \\pmb{Z}_{c}" + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "inline_equation", + "content": "\\pmb{Z}_{s}" + }, + { + "bbox": [ + 132, + 115, + 331, + 319 + ], + "type": "text", + "content": " are the direct causes, direct effects, and the causes of the direct effects. They collectively form the" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": "CMB representations " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "Z_{cmb}" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "Z_o" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " are the variables that are spuriously correlated to target " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "Z_{cmb}" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " variables or domain variable " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": ". To generalize our SCM, we allow for arbitrary causal relations within " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " as long as the causal graph over " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "\\{X, Y, Z, U\\}" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " is a directed acyclic graph (DAG). We summarize the detailed assumptions in Assumption 1 of Appendix A.1. We show that our assumptions is practical and generally hold in real-world applications in Appendix A.2. Moreover, the SCM is Figure 1, combined with the assumptions, covers most scenarios from prior works [3, 32] and hence is a flexible model for performing causal analysis on prediction tasks. The semantic meanings of " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "Z_o" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "Z_{cmb}" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " are context-dependent and vary with the data. For instance, in CMNIST, " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "Z_o" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " represents background information like color, whereas " + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "inline_equation", + "content": "Z_{cmb}" + }, + { + "bbox": [ + 130, + 319, + 482, + 463 + ], + "type": "text", + "content": " denotes foreground information such as digit shape." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 337, + 117, + 481, + 222 + ], + "blocks": [ + { + "bbox": [ + 337, + 117, + 481, + 222 + ], + "lines": [ + { + "bbox": [ + 337, + 117, + 481, + 222 + ], + "spans": [ + { + "bbox": [ + 337, + 117, + 481, + 222 + ], + "type": "image", + "image_path": "2f950604a24249e491a85377114c2a8ce5377d99e34a25e2c679f69e768557aa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 230, + 482, + 296 + ], + "lines": [ + { + "bbox": [ + 335, + 230, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 335, + 230, + 482, + 296 + ], + "type": "text", + "content": "Fig.1: Proposed SCM over " + }, + { + "bbox": [ + 335, + 230, + 482, + 296 + ], + "type": "inline_equation", + "content": "\\{X,U,Z,Y\\}" + }, + { + "bbox": [ + 335, + 230, + 482, + 296 + ], + "type": "text", + "content": ". Undirected links between " + }, + { + "bbox": [ + 335, + 230, + 482, + 296 + ], + "type": "inline_equation", + "content": "Z_{o}, Z_{p}" + }, + { + "bbox": [ + 335, + 230, + 482, + 296 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 335, + 230, + 482, + 296 + ], + "type": "inline_equation", + "content": "Z_{s}" + }, + { + "bbox": [ + 335, + 230, + 482, + 296 + ], + "type": "text", + "content": " indicate that both edge directions are permissible without violating the DAG constraint." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 486, + 322, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 322, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 322, + 497 + ], + "type": "text", + "content": "3.3 Invariant Predictive Mechanism" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "text", + "content": "The conditional distribution " + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "inline_equation", + "content": "p(Y|\\mathbf{X})" + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "text", + "content": ", typically captured by a traditional classifier, is influenced by the domain variable " + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "text", + "content": ", meaning that " + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "inline_equation", + "content": "U \\not\\perp Y|\\mathbf{X}" + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "text", + "content": ". Consequently, the domain-variant distribution " + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "inline_equation", + "content": "p(Y|\\mathbf{X})" + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "text", + "content": " is unsuitable for prediction in unseen domains. Prior causal approaches such as [22, 33, 34] and other content-style approaches make strong assumptions, assuming there is only a single type of causal feature (usually parent variables " + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "inline_equation", + "content": "Z_{p}" + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "text", + "content": ") within the SCM. Due to the simplicity of their SCMs, these methods fail when encountering scenarios where other types of causal features exist, as their selection of causal features cannot block all paths from " + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "text", + "content": ". For example, if the SCM only considers the existence of parent variables " + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "inline_equation", + "content": "Z_{p}" + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "text", + "content": " while " + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "inline_equation", + "content": "Z_{c}" + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "text", + "content": " also exists in practice, the domain variable " + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "text", + "content": " can reach " + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "inline_equation", + "content": "U \\to Z_{o} \\to X \\to Z_{c} \\gets Y" + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 510, + 482, + 643 + ], + "type": "text", + "content": " is given." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 248, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 248, + 91, + 448, + 102 + ], + "type": "text", + "content": "Causal Markov Blanket Representation Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 353, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 353, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 353, + 666 + ], + "type": "text", + "content": "4 This is a standard setting in previous works [32, 39]" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "We advocate for using the Causal Markov Blanket features, denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{cmb} = \\{\\mathbf{Z}_p,\\mathbf{Z}_c,\\mathbf{Z}_s\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ", to predict the target in unseen domains. From Figure 1, we infer that " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "Y \\perp U|\\{\\mathbf{Z}_p,\\mathbf{Z}_c,\\mathbf{Z}_s\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ", implying that " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "p(Y|\\mathbf{Z}_{cmb})" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " remains invariant across domains. Furthermore, we assert that the CMB features constitute the minimal sufficient set required for achieving invariant predictions across varying domains, as adding or removing variables from this set may affect its invariance. We demonstrate how to identify the CMB latent variables " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{cmb}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " from the given data distribution and construct the invariant prediction mechanism " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "p(Y|\\mathbf{Z}_{cmb})" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " for out-of-distribution prediction in Section 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 243, + 443, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 243, + 443, + 258 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 443, + 258 + ], + "type": "text", + "content": "4 Causal Markov Blanket Representation Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 270, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 307 + ], + "type": "text", + "content": "We refer to our proposed method as CMBRL, representing Causal Markov Blanket Representation Learning. The training and inference procedures are introduced in Section 4.1 and Section 4.2, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 325, + 259, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 325, + 259, + 338 + ], + "spans": [ + { + "bbox": [ + 131, + 325, + 259, + 338 + ], + "type": "text", + "content": "4.1 Training Procedure" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 346, + 481, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 346, + 481, + 360 + ], + "spans": [ + { + "bbox": [ + 131, + 346, + 481, + 360 + ], + "type": "text", + "content": "We illustrate the training procedure of the proposed algorithm in Figure 2. We" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 134, + 378, + 482, + 563 + ], + "blocks": [ + { + "bbox": [ + 134, + 378, + 482, + 563 + ], + "lines": [ + { + "bbox": [ + 134, + 378, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 134, + 378, + 482, + 563 + ], + "type": "image", + "image_path": "c91c943988c8d4d70045d70246392ede36e4dacb50c09e29971a61ca5f15662e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 569, + 457, + 582 + ], + "lines": [ + { + "bbox": [ + 155, + 569, + 457, + 582 + ], + "spans": [ + { + "bbox": [ + 155, + 569, + 457, + 582 + ], + "type": "text", + "content": "Fig. 2: The illustration of the three-phase training procedure of CMBRL." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "propose a three-phase training procedure. Initially, we utilize a VAE framework to learn a set of latent variables, which are proven to be component-wise identifiable. Subsequently, we conduct a CMB search using our novel mutual information quantification method to select the latent CMB representations. Finally, we construct a predictor using the selected CMB representations." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "N. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "Phase I: Identifiable Latent Variables Learning. We aim to obtain a set of latent variables " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": ". To do so, we employ an existing (identifiable) variational autoencoder (VAE) framework [32]. We posit a prior distribution on " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " that is consistent with our Assumption 1(b) and belongs to a general exponential family, i.e.," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 217, + 175, + 394, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 175, + 394, + 192 + ], + "spans": [ + { + "bbox": [ + 217, + 175, + 394, + 192 + ], + "type": "interline_equation", + "content": "p _ {\\boldsymbol {T}, \\boldsymbol {\\lambda}} (\\boldsymbol {Z} | Y, U) = \\frac {\\mathcal {Q} (\\boldsymbol {Z})}{\\mathcal {C} (Y , U)} \\exp [ \\boldsymbol {T} (\\boldsymbol {Z}) ^ {T} \\boldsymbol {\\lambda} (Y, U) ]", + "image_path": "fa34d805024926dace66d5c66204a5461a6c51684ede4c830531c08d75bc89ce.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "text", + "content": " is the base measure. " + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "text", + "content": " is the normalizing constant. " + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "text", + "content": " is the arbitrary function. " + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "text", + "content": " is the sufficient statistics. To manage the increasing optimization challenges caused by additional parameters in the prior, we train the VAE framework using a standard protocol from prior work. [16,49]. As described in Eq. (1), the training objective comprises an ELBO loss " + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ELBO}}" + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "text", + "content": " and a score matching loss " + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SM}}" + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "text", + "content": ". The ELBO loss " + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ELBO}}" + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "text", + "content": " optimizes over encoder and decoder parameters " + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "inline_equation", + "content": "(\\theta, \\phi)" + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "text", + "content": ", while the score matching loss " + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SM}}" + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "text", + "content": " minimizes over prior parameters " + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "inline_equation", + "content": "(T, \\lambda)" + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "text", + "content": ". The parameters " + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "inline_equation", + "content": "T, \\lambda" + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "text", + "content": " are constants in " + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ELBO}}" + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "inline_equation", + "content": "\\theta, \\phi" + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "text", + "content": " are constants in " + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SM}}" + }, + { + "bbox": [ + 130, + 197, + 480, + 304 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 190, + 305, + 481, + 320 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 305, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 190, + 305, + 481, + 320 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {o b j}} (\\boldsymbol {\\theta}, \\phi , \\boldsymbol {T}, \\lambda) := \\mathcal {L} _ {\\mathrm {E L B O}} (\\boldsymbol {\\theta}, \\phi , \\hat {\\boldsymbol {T}}, \\hat {\\lambda}) + \\mathcal {L} _ {\\mathrm {S M}} (\\hat {\\boldsymbol {\\theta}}, \\hat {\\phi}, \\boldsymbol {T}, \\lambda) \\tag {1}", + "image_path": "377acebaae7fb0190a6ea98a83756eb46f984857d4ea125dde536ba3fc46a2e9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 326, + 469, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 326, + 469, + 338 + ], + "spans": [ + { + "bbox": [ + 131, + 326, + 469, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ELBO}}" + }, + { + "bbox": [ + 131, + 326, + 469, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 326, + 469, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SM}}" + }, + { + "bbox": [ + 131, + 326, + 469, + 338 + ], + "type": "text", + "content": " are outlined in Eq. (2). " + }, + { + "bbox": [ + 131, + 326, + 469, + 338 + ], + "type": "inline_equation", + "content": "p_{\\mathcal{D}}" + }, + { + "bbox": [ + 131, + 326, + 469, + 338 + ], + "type": "text", + "content": " is the training data distribution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 348, + 481, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 481, + 365 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {E L B O}} := - \\mathbb {E} _ {p _ {\\mathcal {D}}} \\left[ \\mathbb {E} _ {q _ {\\boldsymbol {\\theta}} (\\boldsymbol {z} | \\boldsymbol {x}, y, u)} [ \\log p _ {\\phi} (\\boldsymbol {x} | \\boldsymbol {z}) + \\log p _ {\\boldsymbol {T}, \\boldsymbol {\\lambda}} (\\boldsymbol {z} | y, u) - \\log q _ {\\boldsymbol {\\theta}} (\\boldsymbol {z} | \\boldsymbol {x}, y, u) ] \\right] \\tag {2}", + "image_path": "c5a7a28f965bcfe44c64d87f8865c6113792cb650edce567865c7fe28a7011b5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 149, + 364, + 436, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 364, + 436, + 376 + ], + "spans": [ + { + "bbox": [ + 149, + 364, + 436, + 376 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S M}} := \\mathbb {E} _ {p _ {\\mathcal {D}}} \\left[ \\mathbb {E} _ {q _ {\\boldsymbol {\\theta}} (\\boldsymbol {z} | \\boldsymbol {x}, \\boldsymbol {y}, u)} [ \\| \\nabla_ {\\boldsymbol {z}} \\log q _ {\\boldsymbol {\\theta}} (\\boldsymbol {z} | \\boldsymbol {x}, \\boldsymbol {y}, u) - \\nabla_ {\\boldsymbol {z}} \\log p _ {\\boldsymbol {T}, \\boldsymbol {\\lambda}} (\\boldsymbol {z} | \\boldsymbol {y}, u) \\| ^ {2} ] \\right]", + "image_path": "3f676fdfd954f345b3533a5d976ff02c0ba32ab709e57e8d013fcb40be060ad2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "text", + "content": "As can be seen from the joint distribution " + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "inline_equation", + "content": "p(\\pmb{x}, y, u) \\propto \\int_{\\pmb{z}} p(\\pmb{x} | \\pmb{z}) p(\\pmb{z} | y, u) \\mathrm{d}\\pmb{z}" + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "text", + "content": ", the prior " + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "inline_equation", + "content": "p(\\pmb{z} | y, u)" + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "text", + "content": " is consistent with our SCM. In fact, based on our SCM, " + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "inline_equation", + "content": "p(\\pmb{z} | y, u)" + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "text", + "content": " can be further decomposed into the product of conditional probabilities, which would result in further sparsity in " + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "text", + "content": ". This property makes our learnt model differ from existed work based on different SCMs [19,20,32], as elaborated in Appendix B.4. However, the causal graph is generally unknown a priori, and hence we cannot pre-define the sparsity of " + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "text", + "content": ". Therefore, in our learning algorithm we treat " + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "inline_equation", + "content": "p(\\pmb{z} | y, u)" + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "text", + "content": " as a generic form of prior that satisfies Assumption 1(b). Without a fully specified causal graph, we use this generic prior " + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "inline_equation", + "content": "p(\\pmb{z} | y, u)" + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "text", + "content": " to constrain the learning of VAE to obtain the " + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 387, + 482, + 518 + ], + "type": "text", + "content": " without knowing their causal identities. We summarize the assumptions and identifiability results in Theorem 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 528, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 528, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 131, + 528, + 482, + 540 + ], + "type": "text", + "content": "Theorem 1. Assume the data is sampled from a generative model described by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 153, + 551, + 457, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 551, + 457, + 564 + ], + "spans": [ + { + "bbox": [ + 153, + 551, + 457, + 564 + ], + "type": "interline_equation", + "content": "p _ {\\boldsymbol {\\xi} = (\\boldsymbol {\\phi}, \\boldsymbol {T}, \\boldsymbol {\\lambda})} (\\boldsymbol {X}, \\boldsymbol {Z} | Y, U) = p _ {\\boldsymbol {\\phi}} (\\boldsymbol {X} | \\boldsymbol {Z}) p _ {\\boldsymbol {T}, \\boldsymbol {\\lambda}} (\\boldsymbol {Z} | Y, U), p _ {\\boldsymbol {\\phi}} (\\boldsymbol {X} | \\boldsymbol {Z}) = p _ {\\boldsymbol {\\epsilon}} (\\boldsymbol {X} - g _ {\\boldsymbol {\\phi}} (\\boldsymbol {Z}))", + "image_path": "9641d0630554c214e7ce7efc86b74be2be1f19bb5d60ecd2532ec096c060ec9e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 574, + 482, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 482, + 634 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 482, + 634 + ], + "type": "text", + "content": "We have the following: (1) Under certain assumptions regarding domain variability, the model parameter " + }, + { + "bbox": [ + 131, + 574, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 131, + 574, + 482, + 634 + ], + "type": "text", + "content": " is identifiable up to a permutation and component-wise transformation. (2) If (1) is true and " + }, + { + "bbox": [ + 131, + 574, + 482, + 634 + ], + "type": "inline_equation", + "content": "q_{\\theta}(z|\\mathbf{x},y,u)" + }, + { + "bbox": [ + 131, + 574, + 482, + 634 + ], + "type": "text", + "content": " is flexible and positive everywhere, then our framework learns the true parameters " + }, + { + "bbox": [ + 131, + 574, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\xi^{*}" + }, + { + "bbox": [ + 131, + 574, + 482, + 634 + ], + "type": "text", + "content": " and true latent variables " + }, + { + "bbox": [ + 131, + 574, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{*}" + }, + { + "bbox": [ + 131, + 574, + 482, + 634 + ], + "type": "text", + "content": " up to a permutation and component-wise transformation." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 248, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 248, + 91, + 448, + 103 + ], + "type": "text", + "content": "Causal Markov Blanket Representation Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": "5 Arbitrary function " + }, + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": " and sufficient statistics " + }, + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{T}" + }, + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": " are modeled by neural networks with ReLU activation due to their universal approximation ability." + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": "We provide the detailed assumptions for (1) to hold and proof in Appendix B.26. These detailed assumptions are common in identifiable VAE methods and can impact the performance of the CMB discovery in Phase II. The component-wise identifiability of " + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": " ensures that the CMB variables and spurious variables are separable, which is a crucial prerequisite for performing CMB discovery in Phase II, as verified by empirical results in Appendix F.1. While these assumptions may be violated in real-world applications, strict adherence to them is not always necessary. This motivates our investigation into identifiability performance on the CMNIST dataset in Section 5.1. Empirical results demonstrate that the VAE approach we employ can achieve a decent level of identifiability, even when certain assumptions do not hold." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 266, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 433 + ], + "type": "text", + "content": "Phase II: Identify CMB Set. In this phase, we aim to identify the CMB representations relevant to " + }, + { + "bbox": [ + 130, + 266, + 482, + 433 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 266, + 482, + 433 + ], + "type": "text", + "content": " among the latent variables " + }, + { + "bbox": [ + 130, + 266, + 482, + 433 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 266, + 482, + 433 + ], + "type": "text", + "content": " we learned from the VAE framework. Traditional approaches to MB discovery involve revealing the local causal structure to the target through a series of (conditional) independence (CI) tests. However, conducting MB discovery in latent space encounters two major challenges: 1) the large dimension of the latent variable space, especially with complex and challenging input data, and 2) the mixture of variable types. These challenges can compromise the efficiency and accuracy of CMB discovery results. To tackle these two issues, we introduce an efficient search strategy and a more practical and general approach to conducting mutual-information-based (MI) independence tests. It is worth noting that we obtain CMB variables collectively by detecting and removing spurious variables with designed independence tests. We do not distinguish between the parent, child, and spouse variables within the CMB set." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": "Search Strategy: We find that learning the exact local graph of the target variable " + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": " is unnecessary for identifying the CMB set. Instead of distinguishing among the CMB set, our objective is to separate spurious variables " + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "inline_equation", + "content": "Z_{o}" + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "inline_equation", + "content": "Z_{cmb}" + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": ". In particularly, we discover that spurious variables " + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "inline_equation", + "content": "Z_{o}" + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": " and CMB variables " + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "inline_equation", + "content": "Z_{cmb}" + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": " possess different relations concerning " + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": " conditioned on the remaining variables. If " + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "inline_equation", + "content": "Z_{i} \\in Z_{o}" + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "inline_equation", + "content": "Z_{i} \\perp Y|\\{Z_{\\backslash i}, U\\}" + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "inline_equation", + "content": "Z_{\\backslash i} = \\{Z \\backslash Z_{i}\\}" + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "inline_equation", + "content": "Z_{i} \\in Z_{cmb}" + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "inline_equation", + "content": "Z_{i} \\nsubseteq Y|\\{Z_{\\backslash i}, U\\}" + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": ". We employ mutual information as measurement for independence tests. Theoretically, we can detect spurious variables and separate them from CMB set by quantifying " + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(Z_{i}; Y|Z_{\\backslash i}, U), \\forall Z_{i} \\in Z" + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": ". Actually, for real-world scenarios where " + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": " is unknown, it is challenging to estimate " + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(Z_{i}; Y|Z_{\\backslash i}, U)" + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": ". However, subject to our SCM structural assumptions, we have " + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(Z_{i}; Y|Z_{\\backslash i}) = \\mathcal{I}(Z_{i}; Y|Z_{\\backslash i}, U) = 0" + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": ", if " + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "inline_equation", + "content": "Z_{i} \\in Z_{o}" + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": ". We provide a proof in Appendix C.1. Meanwhile, if " + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "inline_equation", + "content": "Z_{i} \\in Z_{cmb}" + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(Z_{i}; Y|Z_{\\backslash i}) > 0" + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": ". Therefore, we can separate " + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "inline_equation", + "content": "Z_{o}" + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "inline_equation", + "content": "Z_{cmb}" + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": " by performing " + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(Z_{i}; Y|Z_{\\backslash i}), \\forall Z_{i} \\in Z" + }, + { + "bbox": [ + 130, + 434, + 482, + 615 + ], + "type": "text", + "content": ". We summarize such theoretical findings in Proposition 1." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "N. Yin et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 631, + 482, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 631, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 631, + 482, + 665 + ], + "type": "text", + "content": "6 Our proof follows a similar approach to that of [32]. We aim to demonstrate that our assumptions on SCM do not violate the assumptions in the proof, thereby ensuring that the identifiable results hold under our framework." + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 154 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 154 + ], + "type": "text", + "content": "Proposition 1. If the causal graph in Figure 1 and Assumption 1 holds, we have that: 1) for " + }, + { + "bbox": [ + 131, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "Z_{i} \\in \\mathbf{Z}_{o}" + }, + { + "bbox": [ + 131, + 116, + 482, + 154 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "Z_{i} \\perp Y|Z_{\\backslash i}" + }, + { + "bbox": [ + 131, + 116, + 482, + 154 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(Z_{i};Y|Z_{\\backslash i}) = 0" + }, + { + "bbox": [ + 131, + 116, + 482, + 154 + ], + "type": "text", + "content": "; 2) for " + }, + { + "bbox": [ + 131, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "Z_{i} \\in \\mathbf{Z}_{cmb}" + }, + { + "bbox": [ + 131, + 116, + 482, + 154 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "Z_{i} \\nsubseteq Y|Z_{\\backslash i}" + }, + { + "bbox": [ + 131, + 116, + 482, + 154 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(Z_{i};Y|Z_{\\backslash i}) > 0" + }, + { + "bbox": [ + 131, + 116, + 482, + 154 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "type": "code", + "bbox": [ + 132, + 200, + 482, + 502 + ], + "blocks": [ + { + "bbox": [ + 133, + 185, + 391, + 198 + ], + "lines": [ + { + "bbox": [ + 133, + 185, + 391, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 391, + 198 + ], + "type": "text", + "content": "Algorithm 1 Causal Markov Blanket Set Search Strategy" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "lines": [ + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "text", + "content": "Input: M observations of Z and Y, i.e., " + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "inline_equation", + "content": "\\{z_{1}(j), z_{2}(j), \\dots, z_{N}(j), y(j)\\}_{j=1}^{M}" + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "text", + "content": " \nOutput: CMB set " + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "inline_equation", + "content": "Z_{cmb}" + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "text", + "content": " \nStep 1(Fast-Search): find the identity for each variable " + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "inline_equation", + "content": "Z_{i} \\in \\mathbf{Z}" + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "text", + "content": " \nInitial two sets " + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "inline_equation", + "content": "Z_{cmb} \\gets \\emptyset" + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "inline_equation", + "content": "Z_{o} \\gets \\emptyset" + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "text", + "content": " \nfor " + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "inline_equation", + "content": "Z_{i} \\in \\mathbf{Z}, i = 1, 2, \\dots, N" + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "text", + "content": ", do \n Perform Indtest " + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "inline_equation", + "content": "(Z_{i}; Y|Z_{\\backslash i})" + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "text", + "content": " \n if " + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "inline_equation", + "content": "Z_{i} \\nsubseteq Y|Z_{\\backslash i}" + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "text", + "content": " then \n " + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "inline_equation", + "content": "Z_{cmb} \\gets Z_{cmb} \\cup Z_{i}" + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "text", + "content": " \n else \n " + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "inline_equation", + "content": "Z_{o} \\gets Z_{o} \\cup Z_{i}" + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "text", + "content": " \nend if \nend for \nStep 2 (Forward-Backward Search): verify the obtained CMB set. \nrepeat \n for " + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "inline_equation", + "content": "Z_{l} \\in Z_{cmb}" + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "text", + "content": " do \n ▷ Backward procedure \n Perform Indtest " + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "inline_equation", + "content": "(Z_{l}; Y|\\{Z_{cmb}\\backslash Z_{l}\\})" + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "text", + "content": " \n if " + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "inline_equation", + "content": "Z_{l} \\perp Y|\\{Z_{cmb}\\backslash Z_{l}\\}" + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "text", + "content": " then \n " + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "inline_equation", + "content": "Z_{cmb} \\gets Z_{cmb} \\backslash Z_{l}" + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "text", + "content": " \n end if \nend for \nfor " + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "inline_equation", + "content": "Z_{j} \\in Z_{o}" + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "text", + "content": " do \n ▷ Forward procedure \n Perform Indtest " + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "inline_equation", + "content": "(Z_{j}; Y|Z_{cmb})" + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "text", + "content": " \n if " + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "inline_equation", + "content": "Z_{j} \\nsubseteq Y|Z_{cmb}" + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "text", + "content": " then \n " + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "inline_equation", + "content": "Z_{cmb} \\gets Z_{cmb} \\cup Z_{j}" + }, + { + "bbox": [ + 132, + 200, + 482, + 502 + ], + "type": "text", + "content": " \n end if \nend for \nuntil Converge" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_body" + } + ], + "index": 4, + "sub_type": "algorithm" + }, + { + "bbox": [ + 132, + 524, + 425, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 524, + 425, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 524, + 425, + 536 + ], + "type": "text", + "content": "We provide the detailed proof of Proposition 1 in Appendix C.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 536, + 482, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 620 + ], + "type": "text", + "content": "We hence propose a fast-search strategy, which performs CI tests " + }, + { + "bbox": [ + 130, + 536, + 482, + 620 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(Z_i;Y|\\mathbf{Z}_{\\backslash i})" + }, + { + "bbox": [ + 130, + 536, + 482, + 620 + ], + "type": "text", + "content": " for each variable " + }, + { + "bbox": [ + 130, + 536, + 482, + 620 + ], + "type": "inline_equation", + "content": "Z_{i}\\in \\mathbf{Z}" + }, + { + "bbox": [ + 130, + 536, + 482, + 620 + ], + "type": "text", + "content": ". After performing the fast-search and obtaining the initial set of CMB, we verify the initial CMB set by repeatedly checking whether " + }, + { + "bbox": [ + 130, + 536, + 482, + 620 + ], + "type": "inline_equation", + "content": "\\forall Z_{j}\\in \\mathbf{Z}_{o}" + }, + { + "bbox": [ + 130, + 536, + 482, + 620 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 536, + 482, + 620 + ], + "type": "inline_equation", + "content": "Z_{j}\\perp Y|Z_{cmb}" + }, + { + "bbox": [ + 130, + 536, + 482, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 536, + 482, + 620 + ], + "type": "inline_equation", + "content": "\\forall Z_{k}\\in \\mathbf{Z}_{cmb}" + }, + { + "bbox": [ + 130, + 536, + 482, + 620 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 536, + 482, + 620 + ], + "type": "inline_equation", + "content": "Z_{k}\\nsubseteq Y|\\{\\mathbf{Z}_{cmb}\\backslash Z_{k}\\}" + }, + { + "bbox": [ + 130, + 536, + 482, + 620 + ], + "type": "text", + "content": ". We outline our CMB search strategy in Algorithm 1. As illustrated by Proposition 2, our algorithm can correctly identify the CMB variables if the assumptions hold. We provide the detailed proof for Proposition 2 in Appendix C.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "text", + "content": "Proposition 2. Assume independence test results are correct and Assumption 1 for the proposed SCM holds, the ground truth Causal Markov Blanket set can be identified with Algorithm 1." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 248, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 248, + 91, + 448, + 102 + ], + "type": "text", + "content": "Causal Markov Blanket Representation Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "In Step 1, we leverage the structural properties of CMB features to efficiently distinguish them from spurious features. However, since we condition on " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": " variables, which is a larger set than CMB, in step 1, we may have false positive detections and wrongly include spurious features. Thus, we employ Step 2, a forward-backward procedure, to eliminate possible errors in Step 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 175, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 482, + 259 + ], + "type": "text", + "content": "Complexity Analysis of CMB Search: Assuming it takes " + }, + { + "bbox": [ + 130, + 175, + 482, + 259 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 175, + 482, + 259 + ], + "type": "text", + "content": " iterations, with " + }, + { + "bbox": [ + 130, + 175, + 482, + 259 + ], + "type": "inline_equation", + "content": "k << N" + }, + { + "bbox": [ + 130, + 175, + 482, + 259 + ], + "type": "text", + "content": " as observed in practice, for the verification to converge, our CMB strategy requires performing " + }, + { + "bbox": [ + 130, + 175, + 482, + 259 + ], + "type": "inline_equation", + "content": "(k + 1)N" + }, + { + "bbox": [ + 130, + 175, + 482, + 259 + ], + "type": "text", + "content": " CI tests and has a linear complexity of " + }, + { + "bbox": [ + 130, + 175, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(N)" + }, + { + "bbox": [ + 130, + 175, + 482, + 259 + ], + "type": "text", + "content": ". Compared to traditional MB discovery methods like IAMB [48], which directly execute the forward-backward search, our fast-search step reduces the search space and improves learning efficiency, often achieving convergence in just " + }, + { + "bbox": [ + 130, + 175, + 482, + 259 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 130, + 175, + 482, + 259 + ], + "type": "text", + "content": " or 2 iterations in practice." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 260, + 482, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 378 + ], + "type": "text", + "content": "CI Tests: We use the MI-based, denoted as " + }, + { + "bbox": [ + 130, + 260, + 482, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 260, + 482, + 378 + ], + "type": "text", + "content": ", independence tests, with a significance level " + }, + { + "bbox": [ + 130, + 260, + 482, + 378 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 260, + 482, + 378 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 130, + 260, + 482, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{I} < \\alpha" + }, + { + "bbox": [ + 130, + 260, + 482, + 378 + ], + "type": "text", + "content": ", we declare independence. In Phase II, we need to compute the following mutual information " + }, + { + "bbox": [ + 130, + 260, + 482, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(Y; Z_i | \\mathcal{C})" + }, + { + "bbox": [ + 130, + 260, + 482, + 378 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 260, + 482, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 260, + 482, + 378 + ], + "type": "text", + "content": " is the set of variables to condition on. We propose a practical approach that quantifies the MI using a trained predictor and is applicable for both classification tasks and regression tasks. Calculating mutual information for classification tasks is particularly challenging due to the mixture of variable types (where " + }, + { + "bbox": [ + 130, + 260, + 482, + 378 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 260, + 482, + 378 + ], + "type": "text", + "content": " is discrete and variables in " + }, + { + "bbox": [ + 130, + 260, + 482, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 130, + 260, + 482, + 378 + ], + "type": "text", + "content": " are continuous). We illustrate our approach to estimating the MI in Eq. (3) and Eq. (4). By definition of mutual information and conditional entropy, we have" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 163, + 384, + 480, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 163, + 384, + 480, + 432 + ], + "spans": [ + { + "bbox": [ + 163, + 384, + 480, + 432 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {I} (y; z _ {i} | \\mathcal {C}) = \\mathcal {H} [ y | \\mathcal {C} ] - \\mathcal {H} [ y | \\{\\mathcal {C}, z _ {i} \\} ] \\\\ = \\mathbb {E} _ {p (\\mathcal {C})} \\left[ \\mathcal {H} [ p (y | \\mathcal {C}) ] \\right] - \\mathbb {E} _ {p (\\mathcal {C}, z _ {i})} \\left[ \\mathcal {H} [ p (y | \\mathcal {C}, z _ {i}) ] \\right] \\tag {3} \\\\ = \\mathbb {E} _ {p (\\mathcal {C})} \\left[ \\mathcal {H} \\left[ \\mathbb {E} _ {p (\\boldsymbol {z} \\backslash \\mathcal {C} | \\mathcal {C})} [ p (y | \\boldsymbol {z}) ] \\right] \\right] - \\mathbb {E} _ {p (\\mathcal {C}, z _ {i})} \\left[ \\mathcal {H} \\left[ \\mathbb {E} _ {p (\\boldsymbol {z} \\backslash \\{\\mathcal {C}, z _ {i} \\} | \\{\\mathcal {C}, z _ {i} \\})} [ p (y | \\boldsymbol {z}) ] \\right] \\right] \\\\ \\end{array}", + "image_path": "fffd8fd36009cae2a8086982e838e3d1505e75ce1814f3fcc12df06cb7cbb26c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "text", + "content": "According to Eq. (3), we can calculate " + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(y;z_i|\\mathcal{C})" + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "text", + "content": " using a predictor distribution " + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "inline_equation", + "content": "p(Y|\\mathbf{Z})" + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "text", + "content": " and samples of " + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "text", + "content": ". For one input and its label from the training data, i.e., " + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}(j), y(j), u(j) \\sim p_{\\mathcal{D}}" + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "text", + "content": ", we use the mean of learned encoder " + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "inline_equation", + "content": "q_{\\hat{\\theta}}(\\mathbf{Z}|\\mathbf{X} = \\boldsymbol{x}(j), Y = y(j), U = u(j))" + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "text", + "content": " as the sample for " + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}(j) = [z_1(j), z_2(j), \\dots, z_N(j)]" + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "text", + "content": ". After obtaining a total of " + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "text", + "content": " observations for " + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "inline_equation", + "content": "\\{\\boldsymbol{z}(j)\\}_{j=1}^{M}" + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "text", + "content": ", we first learn a predictor " + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "inline_equation", + "content": "p(Y|\\mathbf{Z})" + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "inline_equation", + "content": "\\{\\boldsymbol{z}(j), y(j)\\}_{j=1}^{M}" + }, + { + "bbox": [ + 130, + 437, + 482, + 523 + ], + "type": "text", + "content": ". Substituting the samples and learned distribution into Eq. (3), we have" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 165, + 529, + 447, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 529, + 447, + 574 + ], + "spans": [ + { + "bbox": [ + 165, + 529, + 447, + 574 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {I} (y, z _ {i} | \\mathcal {C}) := - \\frac {1}{M} \\sum_ {j = 1} ^ {M} \\sum_ {l = 0} ^ {L - 1} \\left(\\frac {\\sum_ {k = 1} ^ {M} p \\big (y = l | \\mathcal {C} (j) , \\{z \\backslash \\mathcal {C} \\} (k) \\big)}{M}\\right) \\log \\left(\\frac {\\sum_ {k = 1} ^ {M} p \\big (y = l | \\mathcal {C} (j) , \\{z \\backslash \\mathcal {C} \\} (k) \\big)}{M}\\right) + \\\\ \\frac {1}{M} \\sum_ {j = 1} ^ {M} \\sum_ {l = 0} ^ {L - 1} \\left(\\frac {\\sum_ {k = 1} ^ {M} p (y = l | \\{\\mathcal {C} , z _ {i} \\} (j) , \\{z \\backslash (\\mathcal {C} , z _ {i}) \\} (k))}{M}\\right) \\log \\left(\\frac {\\sum_ {k = 1} ^ {M} p (y = l | \\{\\mathcal {C} , z _ {i} \\} (j) , \\{z \\backslash (\\mathcal {C} , z _ {i}) \\} (k))}{M}\\right) \\\\ \\end{array}", + "image_path": "92ee650b177cad4e9da99920950cbeda42c567ac65765ed779ef2d92e8862564.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 579, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 579, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 579, + 481, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 579, + 481, + 605 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 131, + 579, + 481, + 605 + ], + "type": "text", + "content": " is the number of classes for " + }, + { + "bbox": [ + 131, + 579, + 481, + 605 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 131, + 579, + 481, + 605 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 579, + 481, + 605 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 131, + 579, + 481, + 605 + ], + "type": "text", + "content": " denotes the value of " + }, + { + "bbox": [ + 131, + 579, + 481, + 605 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 131, + 579, + 481, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 579, + 481, + 605 + ], + "type": "inline_equation", + "content": "l \\in \\{0,1,\\dots ,L - 1\\}" + }, + { + "bbox": [ + 131, + 579, + 481, + 605 + ], + "type": "text", + "content": ". Especially, for the fast-search strategy, " + }, + { + "bbox": [ + 131, + 579, + 481, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\mathbf{z}_{\\backslash i}" + }, + { + "bbox": [ + 131, + 579, + 481, + 605 + ], + "type": "text", + "content": ", we have:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 612, + 449, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 612, + 449, + 666 + ], + "spans": [ + { + "bbox": [ + 151, + 612, + 449, + 666 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {I} (y; z _ {i} | \\boldsymbol {z} _ {\\backslash i}) = - \\frac {1}{M} \\sum_ {j = 1} ^ {M} \\sum_ {l = 0} ^ {L - 1} \\frac {\\sum_ {k = 1} ^ {M} p \\left(y = l | \\boldsymbol {z} _ {\\backslash i} (j) , z _ {i} (k)\\right)}{M} \\log \\frac {\\sum_ {k = 1} ^ {M} p \\left(y = l | \\boldsymbol {z} _ {\\backslash i} (j) , z _ {i} (k)\\right)}{M} \\\\ + \\frac {1}{M} \\sum_ {j = 1} ^ {M} \\sum_ {l = 0} ^ {L - 1} p (y = l | z (j)) \\log p (y = l | z (j)) \\\\ \\end{array}", + "image_path": "21bde330d1187823d73d8328164c743e7a85456d693951655302d14fb567481f.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "N. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "content": "The detailed derivation can be found in Appendix C.4. Our CMB discovery method assumes the accuracy of the IndTest. In practice, our approach, like other statistical tests, faces challenges with accuracy when the condition set is large and data is insufficient. Future studies aimed at improving the accuracy of independence tests could benefit from utilizing Bayesian mutual information [9]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 190, + 482, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 190, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 131, + 190, + 482, + 228 + ], + "type": "text", + "content": "Phase III: Construct Invariant Predictor. We utilize the obtained CMB variables " + }, + { + "bbox": [ + 131, + 190, + 482, + 228 + ], + "type": "inline_equation", + "content": "Z_{cmb}" + }, + { + "bbox": [ + 131, + 190, + 482, + 228 + ], + "type": "text", + "content": " and train an invariant predictor " + }, + { + "bbox": [ + 131, + 190, + 482, + 228 + ], + "type": "inline_equation", + "content": "p(Y|Z_{cmb})" + }, + { + "bbox": [ + 131, + 190, + 482, + 228 + ], + "type": "text", + "content": ". Our training procedure is outlined in Eq. (6)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 181, + 232, + 481, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 232, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 181, + 232, + 481, + 251 + ], + "type": "interline_equation", + "content": "\\hat {\\psi} = \\arg \\min _ {\\boldsymbol {\\psi}} \\frac {1}{M} \\sum_ {j = 1} ^ {M} \\mathcal {L} _ {\\operatorname {p r e d}} \\left(y (j), p _ {\\boldsymbol {\\psi}} \\left(\\boldsymbol {Y} \\mid \\boldsymbol {Z} _ {c m b} = \\boldsymbol {z} _ {c m b} (j)\\right)\\right) \\tag {6}", + "image_path": "407b0cc636dbc99c14a17939312e02a1e1fd82a8a1bc3eec287f5d62ab636b1e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 256, + 481, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 256, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 131, + 256, + 481, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 256, + 481, + 281 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 131, + 256, + 481, + 281 + ], + "type": "text", + "content": " are the parameters of the predictor. " + }, + { + "bbox": [ + 131, + 256, + 481, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pred}}(\\cdot)" + }, + { + "bbox": [ + 131, + 256, + 481, + 281 + ], + "type": "text", + "content": " represents the cross-entropy loss for classification tasks and the mean squared loss for regression tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 296, + 263, + 307 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 296, + 263, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 263, + 307 + ], + "type": "text", + "content": "4.2 Inference Procedure" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "text", + "content": "In the inference procedure, we predict labels for data from an unseen test domain, " + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "p_{\\mathcal{D}^{test}}" + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "text", + "content": ". One can infer from Figure 1 that " + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "p(\\pmb{X}|\\pmb{Z})" + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "text", + "content": " is domain-invariant since " + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "X \\perp U|Z" + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "text", + "content": ". Therefore, inferring the latent variable values for inputs from an unseen test domain using " + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "p(\\pmb{X}|\\pmb{Z})" + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "text", + "content": " is reasonable. We adopt the standard inference procedure from [32, 45]. Given an input from the test domain, " + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\pmb{x}^t \\sim p_{\\mathcal{D}^{test}}" + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "text", + "content": ", we obtain the optimal values of " + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "Z_{cmb}" + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "z_{cmb}^{*}" + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "text", + "content": ", by solving the optimization problem outlined in Eq. (7). Here, " + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\lambda_1" + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\lambda_2" + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "text", + "content": " are hyperparameters used to control the scales of the learned " + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "Z_{cmb}" + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "inline_equation", + "content": "Z_o" + }, + { + "bbox": [ + 131, + 312, + 482, + 409 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 164, + 413, + 481, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 413, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 164, + 413, + 481, + 430 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {c m b} ^ {*}, \\boldsymbol {z} _ {o} ^ {*} = \\arg \\min _ {\\boldsymbol {z} _ {c m b}, \\boldsymbol {z} _ {o}} p _ {\\hat {\\phi}} (\\boldsymbol {x} ^ {t} | \\boldsymbol {z} _ {c m b}, \\boldsymbol {z} _ {o}) + \\lambda_ {1} \\| \\boldsymbol {z} _ {c m b} \\| _ {2} ^ {2} + \\lambda_ {2} \\| \\boldsymbol {z} _ {o} \\| _ {2} ^ {2} \\tag {7}", + "image_path": "df102faacae34368717fae933fec35c7249e5d22f2160d4105908a1836b38d48.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 433, + 458, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 433, + 458, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 433, + 458, + 445 + ], + "type": "text", + "content": "Then we can infer the label using the constructed invariant predictor, i.e.," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 253, + 450, + 481, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 450, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 253, + 450, + 481, + 468 + ], + "type": "interline_equation", + "content": "\\hat {y} = \\arg \\max _ {y} p _ {\\hat {\\psi}} (y | \\boldsymbol {z} _ {c m b} ^ {*}) \\tag {8}", + "image_path": "80b38d9d32ea9432332e12302bdedcbf65b8b3bea77041d354f3d334aac19724.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 474, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 474, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 131, + 474, + 481, + 498 + ], + "type": "text", + "content": "We provide the complexity analysis of the optimization in the inference procedure in Appendix D and the ablation study of runtime comparison in Appendix F.3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 536, + 482, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 536, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 131, + 536, + 482, + 631 + ], + "type": "text", + "content": "We validate our CMBRL method for OOD prediction using synthetic and real benchmark distribution shift datasets against state-of-the-art domain generalization baselines. We experiment on four datasets: CMNIST, CelebA, PACS, and VLCS. Our experiments, conducted across 5 trials, are summarized in tables, reporting the mean and standard deviation of accuracy. The details of datasets and implementations can be found in Appendix E.1 and E.2. Additionally, detailed ablation studies on the selection of hyperparameters are available in Appendix F." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 646, + 329, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 646, + 329, + 658 + ], + "spans": [ + { + "bbox": [ + 132, + 646, + 329, + 658 + ], + "type": "text", + "content": "5.1 Identifiability of Latent Variables" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 248, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 248, + 91, + 448, + 102 + ], + "type": "text", + "content": "Causal Markov Blanket Representation Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 329, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 329, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 329, + 282 + ], + "type": "text", + "content": "First, we verify that Phase I of our proposed CMBRL yields identifiable " + }, + { + "bbox": [ + 130, + 116, + 329, + 282 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 130, + 116, + 329, + 282 + ], + "type": "text", + "content": ". We compare it with VAE [30], which lacks identifiability guarantees, and iVAE [19], which uses a conditionally factorized prior. Following standard procedures [19, 20, 22, 32], we evaluate identifiability on CMNIST data by computing the average mean correlation coefficient (MCC) between latent variables recovered by different models with various random initializations. Higher MCC scores indicate stronger identifiability, as shown in Figure 37. Using the same prior and training process as [32], we refer to our method as NF-" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 342, + 123, + 480, + 234 + ], + "blocks": [ + { + "bbox": [ + 342, + 123, + 480, + 234 + ], + "lines": [ + { + "bbox": [ + 342, + 123, + 480, + 234 + ], + "spans": [ + { + "bbox": [ + 342, + 123, + 480, + 234 + ], + "type": "image", + "image_path": "b777fe0796829863924bd9511afa0e31472150f31cb974729e662398569b3db8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 247, + 462, + 258 + ], + "lines": [ + { + "bbox": [ + 354, + 247, + 462, + 258 + ], + "spans": [ + { + "bbox": [ + 354, + 247, + 462, + 258 + ], + "type": "text", + "content": "Fig.3: MCC on CMNIST" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "type": "text", + "content": "iVAE. Figure 3 demonstrates that NF-iVAE recovers latent variables with better identifiability. This identifiability impacts the disentanglement of spurious and CMB variables, influencing the accuracy of the estimated CMB set. Our ablation study in Appendix F.1 empirically shows that better identifiability leads to a more accurate CMB set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 361, + 362, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 361, + 362, + 374 + ], + "spans": [ + { + "bbox": [ + 131, + 361, + 362, + 374 + ], + "type": "text", + "content": "5.2 Out-of-distribution Prediction Accuracy" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 383, + 302, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 302, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 302, + 563 + ], + "type": "text", + "content": "We present both the in-distribution and OOD prediction accuracies of the CMNIST dataset in Table 1. First, we compare with the ERM trained on the entire training domain. Then, we compare with the popular causal invariant feature learning method, IRM, and its two variants: F-IRM (where " + }, + { + "bbox": [ + 130, + 383, + 302, + 563 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 130, + 383, + 302, + 563 + ], + "type": "text", + "content": " is fixed to the identity) and V-IRM (where " + }, + { + "bbox": [ + 130, + 383, + 302, + 563 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 130, + 383, + 302, + 563 + ], + "type": "text", + "content": " is variable). We also include the Robust MIN MAX method, which minimizes the maximum loss across multiple domains. Additionally, we compare with other approaches that leverage SCM knowledge to identify causal features" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 441, + 473, + 539 + ], + "blocks": [ + { + "bbox": [ + 307, + 418, + 481, + 440 + ], + "lines": [ + { + "bbox": [ + 307, + 418, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 307, + 418, + 481, + 440 + ], + "type": "text", + "content": "Table 1: Results on Colored MNIST in terms of accuracy " + }, + { + "bbox": [ + 307, + 418, + 481, + 440 + ], + "type": "inline_equation", + "content": "(\\%)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 441, + 473, + 539 + ], + "lines": [ + { + "bbox": [ + 316, + 441, + 473, + 539 + ], + "spans": [ + { + "bbox": [ + 316, + 441, + 473, + 539 + ], + "type": "table", + "html": "
MethodsPrediction Acc (%)
In-distributionOOD
ERM85.7±0.510.3±0.2
Robust MIN MAX84.3±0.410.9±0.5
F-IRM GAME63.4±1.160.0±2.7
V-IRM GAME64.0±1.049.2±3.4
IRM59.3±4.462.8±9.6
Causalrep70.1±1.568.6±5.5
CTrans76.9±0.872.5±1.1
iCaRL870.6±0.868.8±1.5
CMBRL(ours)76.4±0.874.1±1.5
", + "image_path": "da141dafa74ea058053b9d07c2a05769a93e464ac39b3a85b063c5678585f0e2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 563, + 480, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 480, + 598 + ], + "type": "text", + "content": "of the target, including Causalrep [53], CTrans [34], and iCaRL [32]. CTrans, in particular, considers latent confounding effects between spurious representation and target and employs causal inference via intervention." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 598, + 481, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 598, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 598, + 481, + 635 + ], + "type": "text", + "content": "As shown in Table 1, our CMBRL method achieves optimal OOD performance while minimally compromising in-distribution accuracy. In contrast, the ERM method, which minimizes losses over the entire training domain, tends to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "N. Yin et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "text", + "content": "7 We verify our NF-iVAE implementation's correctness by demonstrating similar identifiability performance to iCaRL in Figure 3" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "capture spurious statistical correlations between labels and colors. Although it achieves the highest in-distribution accuracy, there is a significant drop in prediction accuracy on OOD data. Among the multiple domain-invariant causal representation learning methods, including IRM, F-IRM GAME, and V-IRM GAME, there are notable improvements. However, the limited number of available domains (only 2) may constrain the performance of the IRM methods. Conversely, causal representation learning methods operating under various Structural Causal Models (SCMs), including our CMBRL, demonstrate superior empirical performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "text", + "content": "We further investigate the distinctions between the feature sets selected by iCaRL and our CMBRL. Following the approach described in [32], we first learn the causal graph between the estimated latent variables " + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "text", + "content": " from NF-iVAE and " + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "text", + "content": " using the PC algorithm. We then select the parent set with additional independent tests. The selected parent set yields an OOD prediction accuracy of " + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "inline_equation", + "content": "69.8\\%" + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "text", + "content": ", consistent with the reported results in Table 1. By applying a suitable significance level " + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "text", + "content": ", we can select a CMB set that includes all the variables present in the selected parent set. The OOD prediction accuracy using such a CMB set increased to " + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "inline_equation", + "content": "74.1\\%" + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "text", + "content": ". This finding supports our assumptions on the proposed SCM and suggests the presence of variables not categorized as parent variables of the target, which can further enhance OOD prediction performance. Moreover, CTrans [34] and [22] assume there is an invariant latent confounder between " + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "inline_equation", + "content": "Z_{o}" + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "text", + "content": ", leading to spurious correlations between " + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "text", + "content": ". For example, in an image with a water bird and sea background, the confounder could include temperature and altitude, influencing the generation of the sea (spurious variable " + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "inline_equation", + "content": "Z_{o}" + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "text", + "content": ") and bird " + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "inline_equation", + "content": "(Y)" + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "text", + "content": ". Our proposed SCM does not explicitly model latent confounders. Therefore, we cannot guarantee the generalization of our CMBRL method to scenarios involving unobserved confounders between input " + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "text", + "content": " and target " + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 225, + 482, + 465 + ], + "type": "text", + "content": ". While addressing the latent confounding issue falls outside the scope of this paper, it presents an intriguing avenue for future research. We sum" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 134, + 510, + 480, + 533 + ], + "blocks": [ + { + "bbox": [ + 133, + 487, + 479, + 500 + ], + "lines": [ + { + "bbox": [ + 133, + 487, + 479, + 500 + ], + "spans": [ + { + "bbox": [ + 133, + 487, + 479, + 500 + ], + "type": "text", + "content": "Table 2: Empirical results on CelebA in terms of the OOD prediction accuracy " + }, + { + "bbox": [ + 133, + 487, + 479, + 500 + ], + "type": "inline_equation", + "content": "(\\%)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 510, + 480, + 533 + ], + "lines": [ + { + "bbox": [ + 134, + 510, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 134, + 510, + 480, + 533 + ], + "type": "table", + "html": "
MethodsGroupDROMLDGCORALMMDDACNNMixupERMCTransCausalrepiCaRLCMBRL(ours)
Avg Acc65.165.666.766.268.769.560.766.360.471.576.2
", + "image_path": "613f9e9cd4062edd19658c1c6d6ace01f967c3d9881818843afd5c8ff84be822.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "marize empirical results on average OOD prediction accuracy over 7 settings for the CelebA dataset in Table 2. Detailed empirical results for each pair of settings can be found in Appendix H. We compare the results of our CMBRL with the ERM method and state-of-the-art causal and non-causal representation learning methods. As illustrated in Table 2, our method demonstrates optimal performance, surpassing the second-best baseline by an average margin of " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "4.7\\%" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " across all settings. This outcome indicates the efficacy of our CMBRL methods in excluding spurious variables and avoiding reliance on spurious correlations for prediction in computer vision applications." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 249, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 249, + 91, + 447, + 102 + ], + "type": "text", + "content": "Causal Markov Blanket Representation Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 147, + 485, + 295 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 3: Empirical results on VLCS and PACS datasets in terms of OOD prediction accuracy (\\%) with a backbone of ResNet50." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 147, + 485, + 295 + ], + "lines": [ + { + "bbox": [ + 134, + 147, + 485, + 295 + ], + "spans": [ + { + "bbox": [ + 134, + 147, + 485, + 295 + ], + "type": "table", + "html": "
AlgorithmsVLCSPACS
CLSVAvgACPSAvg
ERM98.0±0.462.6±0.970.8 ±1.977.5 ±1.977.284.8±1.376.4±1.196.7±0.676.1±1.083.5
GroupDRO98.1±0.366.4 ±0.971.0 ±0.376.1 ±1.477.983.5±0.979.1±0.696.7±0.378.3±2.084.4
MLDG98.5±0.361.7±1.273.6±1.875.0±0.877.285.5±1.480.1±1.797.4±0.376.6±1.184.9
CORAL96.9±0.965.7±1.273.3±0.778.7±0.878.788.3±0.280.0±0.797.5±0.378.8±1.386.2
MMD98.3±0.165.6±0.769.7±1.075.7±0.977.386.1±1.479.4±0.996.6±0.276.5±0.784.6
RSC97.5±0.663.1±1.273.0±1.376.2±0.577.585.4±0.879.7±1.897.6±0.378.2±1.285.2
Mixup98.4±0.363.4±0.772.9±0.876.1±1.277.786.1±0.778.9±0.897.6±0.175.8±1.884.6
DANN98.5±1.364.9±1.372.6±1.478.7±1.778.286.4±0.877.4±0.897.3±0.473.5±2.383.6
CDANN97.6±0.665.2±0.873.4±1.476.9±0.578.384.6±1.875.5±0.996.8±0.373.5±0.682.6
MTL97.6±0.660.6±1.371.0±1.277.2±0.776.687.5±0.877.1±0.796.4±0.877.3±1.884.6
ARM97.2±0.562.7±1.70.6±0.675.8±0.976.686.8±0.676.8±0.797.4±0.379.3±1.285.1
IRM98.6±0.166.0 ±0.972.3 ±0.677.3 ±0.978.584.7±0.480.0±0.697.2±0.379.3±1.085.5
SagNet97.3±0.461.6±0.873.4±1.977.6±0.477.587.4±1.080.7±0.697.1±0.180.0±0.486.3
iCaRL----81.8----88.7
CMBRL (ours)98.7±0.271.2±0.177.1±0.482.1±0.182.389.2±0.785.3±1.297.7±0.584.1±0.689.1
", + "image_path": "1a3c955f21a5f4e2efab8ce1d0d54609b0fcc80b7d099dbe62010973af0e98c4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 319, + 482, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 440 + ], + "type": "text", + "content": "We present the empirical results for PACS and VLCS regarding OOD prediction accuracy in Table 3. These datasets serve as benchmarks for domain generalization tasks. For a comprehensive comparison, we evaluate not only against causal-inspired methods but also against various other DG methods, including GroupDRO [41], MLDG [25], CORAL [44], RSC [15], Mixup [54], SageNet [36], and others. We adhere to the experimental settings outlined in [13]. We report the OOD performance for each pair of train and test domains and the average accuracy across all combinations. As depicted in Table 3, our CMBRL method establishes state-of-the-art performance compared to popular alternatives in the domain generalization landscape." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 460, + 220, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 220, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 220, + 472 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 665 + ], + "type": "text", + "content": "In this work, we investigate the Causal Markov Blanket discovery for high-dimensional data. We establish a framework guided by an SCM describing the data generation process, allowing for the Causal Markov Blanket discovery in latent space. We then construct an invariant prediction mechanism utilizing the CMB representations, which is suitable for OOD prediction. This framework can be further employed for causal analysis, reasoning, and intervention. We propose a three-phase algorithm to disentangle CMB features from spurious ones, reducing the risk of the prediction model relying on spurious correlations. It is worth noting that our method requires the satisfaction of assumptions on SCM and the assurance of the identifiability of latent variables. It may also suffer from inefficiencies due to the three training phases and optimization during inference. However, our method demonstrates its effectiveness by achieving significant OOD prediction performance, surpassing state-of-the-art causal representation learning methods and domain generalization methods on multiple benchmark distribution shift datasets." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "N. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 481, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 481, + 173 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 481, + 173 + ], + "type": "text", + "content": "This work was supported in part by the National Science Foundation award IIS 2236026 and in part by IBM through the IBM-Rensselaer Future of Computing Research Collaboration." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 212, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 212, + 481, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 212, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 212, + 481, + 245 + ], + "type": "text", + "content": "1. Ahuja, K., Caballero, E., Zhang, D., Bengio, Y., Mitlagkas, I., Rish, I.: Invariance principle meets information bottleneck for out-of-distribution generalization. In: Advances in Neural Inf. Proc. Systems (2021) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 245, + 481, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 481, + 266 + ], + "type": "text", + "content": "2. Ahuja, K., Shanmugam, K., Varshney, K., Dhurandhar, A.: Invariant risk minimization game. In: International Conference on Machine Learning (2020) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 481, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 481, + 299 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 481, + 299 + ], + "type": "text", + "content": "3. Ahuja, K., Wang, J., Dhurandhar, A., Shanmugam, K., Varshney, K.R.: Empirical or invariant risk minimization? a sample complexity perspective. In: International Conference on Learning Representations (2021) 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 299, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 481, + 342 + ], + "type": "text", + "content": "4. Aliferis, C.F., Statnikov, A., Tsamardinos, I., Mani, S., Koutsoukos, X.D.: Local causal and markov blanket induction for causal discovery and feature selection for classification part i: algorithms and empirical evaluation. Journal of Machine Learning Research 11(1) (2010) 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 342, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 481, + 374 + ], + "type": "text", + "content": "5. Aliferis, C.F., Tsamardinos, I., Statnikov, A.: Hiton: a novel markov blanket algorithm for optimal variable selection. In: AMIA annual symposium proceedings. vol. 2003, p. 21. American Medical Informatics Association (2003) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 374, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 481, + 396 + ], + "type": "text", + "content": "6. Arjovsky, M., Bottou, L., Gulrajani, I., Lopez-Paz, D.: Invariant risk minimization. arXiv preprint arXiv:1907.02893 (2019) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 396, + 481, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 481, + 426 + ], + "type": "text", + "content": "7. Cui, P., Athey, S.: Stable learning establishes some common ground between causal inference and machine learning. Nature Machine Intelligence 4(2), 110-115 (2022) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 427, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 427, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 427, + 481, + 460 + ], + "type": "text", + "content": "8. Cui, P., Shen, Z., Li, S., Yao, L., Li, Y., Chu, Z., Gao, J.: Causal inference meets machine learning. In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. pp. 3527-3528 (2020) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 460, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 481, + 493 + ], + "type": "text", + "content": "9. Cui, Z., Yin, N., Wang, Y., Ji, Q.: Empirical bayesian approaches for robust constraint-based causal discovery under insufficient data. International Joint Conference on Artificial Intelligence (2022) 11" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 493, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 481, + 536 + ], + "type": "text", + "content": "0. Fu, S., Desmarais, M.C.: Fast markov blanket discovery algorithm via local learning within single pass. In: Advances in Artificial Intelligence: 21st Conference of the Canadian Society for Computational Studies of Intelligence, Canadian AI 2008 Windsor, Canada, May 28-30, 2008 Proceedings 21. pp. 96-107. Springer (2008) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 536, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 481, + 557 + ], + "type": "text", + "content": "1. Gao, T., Ji, Q.: Local causal discovery of direct causes and effects. Advances in Neural Information Processing Systems 28 (2015) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 558, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 558, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 558, + 481, + 578 + ], + "type": "text", + "content": "2. Gao, T., Wang, Z., Ji, Q.: Structured feature selection. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (December 2015) 2, 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 578, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 600 + ], + "type": "text", + "content": "3. Gulrajani, I., Lopez-Paz, D.: In search of lost domain generalization. arXiv preprint arXiv:2007.01434 (2020) 14" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "type": "text", + "content": "4. Hendrycks, D., Mu, N., Cubuk, E.D., Zoph, B., Gilmer, J., Lakshminarayanan, B.: Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781 (2019) 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "5. Huang, Z., Wang, H., Xing, E.P., Huang, D.: Self-challenging improves cross-domain generalization. In: European Conference on Computer Vision. pp. 124-140. Springer (2020) 14" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 249, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 249, + 91, + 447, + 102 + ], + "type": "text", + "content": "Causal Markov Blanket Representation Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "16. Hyvarinen, A., Dayan, P.: Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research 6(4) (2005) 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 159 + ], + "type": "text", + "content": "17. Janzing, D.: Causal regularization. Advances in Neural Information Processing Systems 32 (2019) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 180 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 180 + ], + "type": "text", + "content": "18. Jiang, Y., Veitch, V.: Invariant and transportable representations for anti-causal domain shifts. arXiv preprint arXiv:2207.01603 (2022) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 182, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 213 + ], + "type": "text", + "content": "19. Khemakhem, I., Kingma, D., Monti, R., Hyvarinen, A.: Variational autoencoders and nonlinear ica: A unifying framework. In: International Conference on Artificial Intelligence and Statistics. pp. 2207-2217. PMLR (2020) 3, 4, 7, 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 214, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 245 + ], + "type": "text", + "content": "20. Khemakhem, I., Monti, R., Kingma, D., Hyvarinen, A.: Ice-beem: Identifiable conditional energy-based deep models based on nonlinear ica. Advances in Neural Information Processing Systems 33, 12768-12778 (2020) 4, 7, 12" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 247, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 480, + 266 + ], + "type": "text", + "content": "21. Koller, D., Sahami, M.: Toward optimal feature selection. Tech. rep., Stanford InfoLab (1996) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 268, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 298 + ], + "type": "text", + "content": "22. Kong, L., Xie, S., Yao, W., Zheng, Y., Chen, G., Stojanov, P., Akinwande, V., Zhang, K.: Partial identifiability for domain adaptation. arXiv preprint arXiv:2306.06510 (2023) 2, 3, 5, 12, 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 300, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 319 + ], + "type": "text", + "content": "23. Koyama, M., Yamaguchi, S.: When is invariance useful in an out-of-distribution generalization problem? arXiv preprint arXiv:2008.01883 (2020) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 322, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 480, + 352 + ], + "type": "text", + "content": "24. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.: Learning to generalize: Meta-learning for domain generalization. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 354, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 384 + ], + "type": "text", + "content": "25. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.M.: Deeper, broader and artier domain generalization. In: Proceedings of the IEEE international conference on computer vision. pp. 5542-5550 (2017) 14" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 386, + 480, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 416 + ], + "type": "text", + "content": "26. Li, Z., Cai, R., Chen, G., Sun, B., Hao, Z., Zhang, K.: Subspace identification for multi-source domain adaptation. Advances in Neural Information Processing Systems 36 (2024) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 418, + 480, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 459 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 459 + ], + "type": "text", + "content": "27. Liu, B., Wang, D., Yang, X., Zhou, Y., Yao, R., Shao, Z., Zhao, J.: Show, deconfound and tell: Image captioning with causal inference. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18041-18050 (2022) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 461, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 502 + ], + "type": "text", + "content": "28. Locatello, F., Bauer, S., Lucic, M., Raetsch, G., Gelly, S., Schölkopf, B., Bachem, O.: Challenging common assumptions in the unsupervised learning of disentangled representations. In: international conference on machine learning. pp. 4114-4124. PMLR (2019) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 504, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 535 + ], + "type": "text", + "content": "29. Locatello, F., Poole, B., Ratsch, G., Scholkopf, B., Bachem, O., Tschannen, M.: Weakly-supervised disentanglement without compromises. In: International Conference on Machine Learning. pp. 6348-6359. PMLR (2020) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 567 + ], + "type": "text", + "content": "30. Lopez, R., Regier, J., Jordan, M.I., Yosef, N.: Information constraints on auto-encoding variational bayes. Advances in neural information processing systems 31 (2018) 12" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 569, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 480, + 600 + ], + "type": "text", + "content": "31. Lopez-Paz, D., Nishihara, R., Chintala, S., Scholkopf, B., Bottou, L.: Discovering causal signals in images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6979-6987 (2017) 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "text", + "content": "32. Lu, C., Wu, Y., Hernández-Lobato, J.M., Schölkopf, B.: Invariant causal representation learning for out-of-distribution generalization. In: International Conference on Learning Representations (2021) 2, 3, 4, 5, 7, 8, 11, 12, 13" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "33. Mao, C., Cha, A., Gupta, A., Wang, H., Yang, J., Vondrick, C.: Generative interventions for causal learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3947-3956 (2021) 2, 3, 4, 5" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 218, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 218, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 218, + 100 + ], + "type": "text", + "content": "N. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "34. Mao, C., Xia, K., Wang, J., Wang, H., Yang, J., Baireinboim, E., Vondrick, C.: Causal transportability for visual recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7521-7531 (2022) 2, 4, 5, 12, 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 481, + 183 + ], + "type": "text", + "content": "35. Nagarajan, V., Andreassen, A., Neyshabur, B.: Understanding the failure modes of out-of-distribution generalization. arXiv preprint arXiv:2010.15775 (2020) 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "text", + "content": "36. Nam, H., Lee, H., Park, J., Yoon, W., Yoo, D.: Reducing domain gap by reducing style bias. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8690-8699 (2021) 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "text", + "content": "37. Pearl, J.: Probabilistic reasoning in intelligent systems: networks of plausible inference. Morgan Kaufmann (1988) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 262 + ], + "type": "text", + "content": "38. Pellet, J.P., Elisseeff, A.: Using markov blankets for causal structure learning. Journal of Machine Learning Research 9(7) (2008) 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 263, + 481, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 481, + 285 + ], + "type": "text", + "content": "39. Peters, J., Buhlmann, P., Meinshausen, N.: Causal inference using invariant prediction: identification and confidence intervals. arxiv. Methodology (2015) 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 285, + 481, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 307 + ], + "type": "text", + "content": "40. Rosenfeld, E., Ravikumar, P., Risteski, A.: The risks of invariant risk minimization. In: International Conference on Learning Representations (2021) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "text", + "content": "41. Sagawa, S., Koh, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731 (2019) 2, 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 342, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 373 + ], + "type": "text", + "content": "42. Shen, X., Liu, F., Dong, H., Lian, Q., Chen, Z., Zhang, T.: Weakly supervised disentangled generative causal representation learning. The Journal of Machine Learning Research 23(1), 10994-11048 (2022) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "text", + "content": "43. Shu, Y., Cao, Z., Wang, C., Wang, J., Long, M.: Open domain generalization with domain-augmented meta-learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9624-9633 (2021) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 409, + 481, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 481, + 452 + ], + "type": "text", + "content": "44. Sun, B., Saenko, K.: Deep coral: Correlation alignment for deep domain adaptation. In: Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14. pp. 443-450. Springer (2016) 14" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 453, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 475 + ], + "type": "text", + "content": "45. Sun, X., Wu, B., Zheng, X., Liu, C., Chen, W., Qin, T., Liu, T.y.: Latent causal invariant model. arXiv preprint arXiv:2011.02203 (2020) 11" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 508 + ], + "type": "text", + "content": "46. Talon, D., Lippe, P., James, S., Del Bue, A., Magliacane, S.: Towards the reusability and compositionality of causal representations. arXiv preprint arXiv:2403.09830 (2024) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 509, + 481, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 481, + 541 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 481, + 541 + ], + "type": "text", + "content": "47. Tan, Y., Liu, Z.: Feature selection and prediction with a markov blanket structure learning algorithm. In: BMC bioinformatics. vol. 14, pp. 1-3. BioMed Central (2013) 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 575 + ], + "type": "text", + "content": "48. Tsamardinos, I., Aliferis, C.F., Statnikov, A.R., Statnikov, E.: Algorithms for large scale markov blanket discovery. In: FLAIRS conference. vol. 2, pp. 376-380. St. Augustine, FL (2003) 3, 10" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "type": "text", + "content": "49. Vincent, P.: A connection between score matching and denoising autoencoders. Neural computation 23(7), 1661-1674 (2011) 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "text", + "content": "50. Volpi, R., Namkoong, H., Sener, O., Duchi, J.C., Murino, V., Savarese, S.: Generalizing to unseen domains via adversarial data augmentation. Advances in neural information processing systems 31 (2018) 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "51. Wang, H., Xiao, C., Kossaifi, J., Yu, Z., Anandkumar, A., Wang, Z.: Augmax: Adversarial composition of random augmentations for robust training. Advances in neural information processing systems 34, 237-250 (2021) 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 249, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 249, + 91, + 447, + 102 + ], + "type": "text", + "content": "Causal Markov Blanket Representation Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 281 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "52. Wang, T., Huang, J., Zhang, H., Sun, Q.: Visual commonsense r-cnn. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10760-10770 (2020) 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "text", + "content": "53. Wang, Y., Jordan, M.I.: Desiderata for representation learning: A causal perspective. arXiv preprint arXiv:2109.03795 (2021) 12" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "54. Yan, S., Song, H., Li, N., Zou, L., Ren, L.: Improve unsupervised domain adaptation with mixup training. arXiv preprint arXiv:2001.00677 (2020) 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "55. Yin, N., Wang, H., Gao, T., Dhurandhar, A., Ji, Q.: Causal markov blanket representation learning for out-of-distribution generalization. In: Causal Representation Learning Workshop at NeurIPS 2023 (2023) 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "type": "text", + "content": "56. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6023-6032 (2019) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "type": "text", + "content": "57. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017) 3" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "N. Yin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Integration of Global and Local Representations for Fine-grained Cross-modal Alignment/35329370-3365-4c03-a82d-856d1e6d798d_content_list.json b/2024/Integration of Global and Local Representations for Fine-grained Cross-modal Alignment/35329370-3365-4c03-a82d-856d1e6d798d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..deb8eb74c53e860144d77cb18159bb8b65b156a2 --- /dev/null +++ b/2024/Integration of Global and Local Representations for Fine-grained Cross-modal Alignment/35329370-3365-4c03-a82d-856d1e6d798d_content_list.json @@ -0,0 +1,1918 @@ +[ + { + "type": "text", + "text": "Integration of Global and Local Representations for Fine-grained Cross-modal Alignment", + "text_level": 1, + "bbox": [ + 225, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seungwan Jin $^{\\text{®}}$ , Hoyoung Choi $^{\\text{®}}$ , Taehyung Noh $^{\\text{®}}$ , and Kyungsik Han $^{\\star}$", + "bbox": [ + 238, + 213, + 764, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hanyang University, South Korea {seungwanjin, choiho00, yestaehyung, kyungsikhan}@hanyang.ac.kr", + "bbox": [ + 267, + 239, + 733, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Fashion is one of the representative domains of fine-grained Vision-Language Pre-training (VLP) involving a large number of images and text. Previous fashion VLP research has proposed various pretraining tasks to account for fine-grained details in multimodal fusion. However, fashion VLP research has not yet addressed the need to focus on (1) uni-modal embeddings that reflect fine-grained features and (2) hard negative samples to improve the performance of fine-grained $\\mathrm{V + L}$ retrieval tasks. In this paper, we propose Fashion-FINE (Fashion VLP with Fine-grained Cross-modal Alignment using the INtegrated representations of global and local patch Embeddings), which consists of three key modules. First, a modality-agnostic adapter (MAA) learns uni-modal integrated representations and reflects fine-grained details contained in local patches. Second, hard negative mining with focal loss (HNM-F) performs cross-modal alignment using the integrated representations, focusing on hard negatives to boost the learning of fine-grained cross-modal alignment. Third, comprehensive cross-modal alignment (C-CmA) extracts low- and high-level fashion information from the text and learns the semantic alignment to encourage disentangled embedding of the integrated image representations. Fashion-FINE achieved state-of-the-art performance on two representative public benchmarks (i.e., FashionGen and FashionIQ) in three representative $\\mathrm{V + L}$ retrieval tasks, demonstrating its effectiveness in learning fine-grained features.", + "bbox": [ + 261, + 311, + 743, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Vision-Language Pre-training $\\cdot$ Fine-grained Representation Learning $\\cdot$ Fashion", + "bbox": [ + 259, + 630, + 740, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 686, + 375, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With recent advances in general VLP research that aims to learn common sense knowledge, there has been increasing attention to fine-grained cross-modal representation learning. Fashion is one of the most representative VLP research domains that applies the concept of cross-modal representations based on large-scale image and text information, about fashion items for fine-grained $\\mathrm{V + L}$ retrieval tasks such as cross-modal retrieval and text-guided image retrieval.", + "bbox": [ + 212, + 723, + 787, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 217, + 824, + 382, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/35c804a7d8757f0a0dbc7b0b951d6267a454a85987cc436f555f1daed0677069.jpg", + "image_caption": [ + "(a) Image-Text Pairs in Fashion Domain" + ], + "image_footnote": [], + "bbox": [ + 223, + 160, + 282, + 215 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Long sleeve down parka-style coat in olive green. Genuine raccon fur trim at hood. Two-way zip closure at front concealed by button placket. Drawstring at collar. Welt pockets and flap pockets at front. Logo patch at upper sleeve. Pleated at back panel. Fully lined. Tonal stitching. Elasticized sleeve cuffs.", + "bbox": [ + 290, + 167, + 411, + 212 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/efcdb87db5dbb4f15cdd7c8095403e784913d8d1aaac7dbe37e7a1828c86590a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 217, + 282, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Long sleeve down parka-style coat in olive green. Removable genuine raccoon fur trim at hood. Two-way zip closure at front concealed by press-stud packet. Welt pockets and flap pockets at front. Logo patch at upper sleeve. Drawstring at waist. Fully lined. Tonal stitching. Inset ribbed knit sleeve cuff.", + "bbox": [ + 290, + 220, + 411, + 267 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/242c51ded600a09d98c7c5d4a4be58b03926717afd1c560269b33747bc77b06f.jpg", + "image_caption": [ + "Fig. 1: Our motivation for this research. (a) shows an example of three image-text pair cases (i.e., positive, hard negative, and easy negative pairs) in fashion products, where the colored areas specifically indicate fashion attributes that distinguish positives from hard negatives. (b) shows that fashion $\\mathrm{V + L}$ tasks (e.g., cross-modal retrieval) are essential for distinguishing positives from hard negatives given anchors by capturing the fine-grained features. Our motivation is that it is difficult to distinguish between positives and hard negatives using only global representations, but this can be compensated by integrating local representations that reflect fine-grained attributes." + ], + "image_footnote": [], + "bbox": [ + 225, + 271, + 282, + 323 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Long sleeve down parka in black. Removable hood with cinch-belt detail at crown and wired removable fur trim. Tonal bungee drawstrings at hood. Stand collar. Two-way zip closure at front concealed by press-stud placket. Tonal logo applique at exterior arm. Welt and flap pockets at front with tonal signature patch...", + "bbox": [ + 289, + 273, + 413, + 321 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e52ebc137c4d3d3587b667d757ac8748b62cc05fa004e9f943fd3e6d6fe8b6da.jpg", + "image_caption": [ + "(b) Our Motivation for Fine-grained Cross-modal Alignment" + ], + "image_footnote": [], + "bbox": [ + 433, + 160, + 782, + 328 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most fashion VLP research [7, 8, 10, 11, 25, 45] has been evolved based on the general VLP architectures by applying self-supervised pre-training tasks, such as image-text contrastive learning (ITC) and image-text matching (ITM). Recently, employing the dual-stream+fusion [13, 39] structure, fashion VLP research [7, 10, 11, 45] proposed various pre-training tasks (e.g., masked patch modeling) to address the fine-grained features using the fusion encoder. FashionSAP [11], the current state-of-the-art model, proposed high-level multimodal semantic alignment and used the relative similarity between the global representations (i.e., [CLS] embedding) obtained before fusion to perform $\\mathrm{V} + \\mathrm{L}$ retrieval tasks instead of the ITM score calculated in the fusion encoder. However, using only global representations has limitations in learning fine-grained features in the cross-modal alignment and poses challenges in distinguishing between positive and hard negatives for a given anchor during the test phase (Figure 1).", + "bbox": [ + 212, + 502, + 787, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this, conventional methods have focused on learning to embed positive image-text pairs closer and negative pairs further using loss functions such as InfoNCE [26]. However, in self-supervised VLP, since only the pre-matched image-text pairs are considered positive and all others are considered negative, a severe class imbalance problem may arise due to the difference between the number of positive and negative pairs. Moreover, the overwhelming number of easy negative pairs compared to hard negative pairs leads to a dominance of the loss scale for easy negatives over hard negatives, thereby limiting the learning of the model to distinguish between positive and hard negative pairs.", + "bbox": [ + 212, + 704, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Jin et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose Fashion-FINE (Fashion VLP with Fine-grained cross-modal alignment using the INtegrated representations of global and local patch Embeddings), which learns cross-modal alignment in three phases.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The first phase is to learn the importance of each local patch embedding for images and text using a modality-agnostic adapter (MAA). For each modality, we obtain a local representation that reflects the feature importance calculated by MAA at each patch location, and this local representation is then combined with the global representation, specifically the [CLS] embedding, to finally form the integrated representation. The second phase is the hard negative mining with focal loss (HNM-F), which integrates ITC and focal loss to mitigate the class and loss imbalance issues in ITC and to focus on learning to distinguish between positive and hard negative pairs. The third phase is the comprehensive cross-modal alignment (C-CmA), which semantically aligns the low-level and high-level fashion information in text descriptions with the integrated image representations. Through this, Fashion-FINE learns to encourage disentangled embedding of the image representations in the multimodal space, taking into account the comprehensive fashion text information.", + "bbox": [ + 212, + 191, + 784, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 238, + 405, + 566, + 419 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Fashion-FINE, which performs fine-grained cross-modal alignment by integrating global and local representations of the two modalities. Fashion-FINE has achieved state-of-the-art performance in all fashion V+L retrieval tasks.", + "- We have improved the architecture for fashion VLP research by integrating Fashion-FINE with existing fusion encoder-based pre-training tasks, and we have verified its applicability through performance improvements.", + "- Our study is the first to quantitatively and qualitatively validate the importance of reflecting local patch embeddings in fine-grained cross-modal alignment and the boosting effect of hard negative mining with focal loss." + ], + "bbox": [ + 225, + 431, + 784, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 609, + 387, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Vision-Language Pre-training (VLP)", + "text_level": 1, + "bbox": [ + 214, + 645, + 563, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Many VLP methods have achieved significant performance in various downstream $\\mathrm{V + L}$ tasks. The most successful VLP architectures fall into two categories: dual-stream and dual-stream $^+$ fusion [13, 39]. The dual-stream architecture uses two modality-specific encoders to perform pre-training tasks (e.g., ITC) for semantic alignment between the two modalities. For example, CLIP [31] and ALIGN [14] have used ITC to train image-text alignment on large-scale multimodal data, achieving remarkable performance in general downstream $\\mathrm{V + L}$ tasks. While this dual-stream is advantageous for achieving high performance, it requires extremely large-scale data during pre-training. On the other hand, the dual-stream $^+$ fusion architecture places multimodal fusion encoders on top of two modality-specific encoders to perform ITC followed by cross-modal learning (e.g.,", + "bbox": [ + 212, + 672, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Fashion-FINE", + "bbox": [ + 635, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MLM and ITM). ALBEF [20] and TCL [43], based on the dual-stream+fusion, have achieved superior performance compared to dual-stream models.", + "bbox": [ + 212, + 145, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing VLP methods have been trained on noisy and large multimodal Web data due to the limited amount of high-quality human-annotated image-text pair data [14]. Recently, BLIP [19] and BLIP-2 [18] have achieved superior performance compared to ALBEF and TCL by expanding the dual-stream+fusion structure by proposing a decoder to bootstrap noisy text captions and Q-Former to learn soft visual prompt for bootstrapping a decoder-based large language model (LLM), respectively. Rather than focusing on human-annotated image-text pairs, BLIP-based methods tackle the inherent challenges in processing the noisy and large-scale multimodal datasets collected from the Web.", + "bbox": [ + 212, + 176, + 785, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The public benchmark datasets (e.g., FashionGen [33], FashionIQ [41]) addressed in recent fashion VLP research [10, 11, 25, 40] consist of multimodal datasets annotated by humans who manually describe fine-grained fashion attributes and contain no (or less) noise, unlike Web data. Thus, we set a pure dual-stream+fusion architecture without decoders as the baseline of FashionFINE, which is consistent with recent fashion VLP. By integrating our proposed modules with the pre-training tasks proposed in the prior work, we advance the fashion VLP architecture for the fine-grained $\\mathrm{V + L}$ retrieval tasks.", + "bbox": [ + 212, + 311, + 787, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Hard Negative Mining in VLP", + "text_level": 1, + "bbox": [ + 214, + 454, + 514, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hard negative mining is an important technique used in machine learning, especially in fields such as computer vision and natural language processing, to improve model performance [9,32,34,42]. It involves focused sampling and training on the sampled hard negatives, which are semantically similar to positive data but are negative. In VLP, hard negative mining has been recently proposed to learn to distinguish between positive and hard negative samples in the image-text matching (ITM) task by extracting hard negative samples based on the calculated relative similarity during the cross-modal alignment phase [20, 43]. To enhance the effectiveness of the hard negative sampling, GRIT-VLP [1] proposed a grouped mini-batch sampling strategy to improve in-batch hard negative sampling in dual-stream $^+$ fusion architecture. However, the previous studies still rely solely on the use of global representations, which are limited in reflecting fine-grained details when calculating relative similarity for negative sampling.", + "bbox": [ + 212, + 478, + 787, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To the best of our knowledge, no research has been aimed at learning fine-grained local features to distinguish positive from hard negative in the cross-modal alignment to improve fine-grained $\\mathrm{V + L}$ tasks.", + "bbox": [ + 212, + 674, + 785, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 VLP in Fashion $\\mathbf{V} + \\mathbf{L}$ tasks", + "text_level": 1, + "bbox": [ + 214, + 739, + 485, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared to general VLP that uses coarse-grained common sense knowledge [16, 22, 27, 29, 36], fashion VLP should focus more on fine-grained attributes in the multimodal fashion data [2]. FashionBERT [7] is the first cross-modal retrieval method targeting the fashion domain. It uses a transformer-based fusion encoder and introduces masked patch modeling to enable contextual learning of images", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Jin et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/922f80b5d608b76d42000ecfb65054f99e871e91ccf9babdcae8617606bafecd.jpg", + "image_caption": [ + "(a) Fine-grained Cross-modal Alignment before Multimodal Fusion", + "Fig. 2: An overview of Fashion-FINE. (a): Fine-grained cross-modal alignment through our proposed three modules (i.e., MAA, HNM-F, and C-CmA). (b): Sampling hard negatives from the mini-batch for each modality (i.e., image and text). (c): Fusion encoder-based pre-training tasks (i.e., image-text matching (ITM), prompt token prediction (PTP), and token replacement prediction (TRP)) for the given positive pair and the two hard negative pairs." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 486, + 273 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cde9a834051775f83ca8a1de094ac306314dee4deb4dc039a10d52fbbfff807e.jpg", + "image_caption": [ + "(b) Hard Negative Sampling" + ], + "image_footnote": [], + "bbox": [ + 495, + 143, + 625, + 272 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/765dbfd8cbae699c228ebeeac816e710342282bee85a93e44cee67eea4fb1952.jpg", + "image_caption": [ + "(c) Pre-training Tasks in Fusion Encoder" + ], + "image_footnote": [], + "bbox": [ + 625, + 143, + 781, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and text. KaleidoBERT [45] is a method for semantic alignment with word tokens after extracting image patches through various patch methods (e.g., Rotation recognition, Jigsaw puzzle solving). In fashion e-commerce, words may be used differently from their common meaning. For example, \"diesel\" and \"canada goose\" are the brand names, not a type of gas for vehicles or an animal in Canada. EICLIP [25] proposed semantic alignment methods that consider the homonymous nature of e-commerce language. However, the previous studies still had to rely on the general image-text contrastive learning method for cross-modal alignment.", + "bbox": [ + 212, + 449, + 787, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recently, representation learning methods applied to the fashion domain have been proposed using the dual-stream+fusion architecture of general VLP methods. FashionViL [10] introduced masked patch feature classification (MPFC), which learns characteristics of image patches based on discrete codes from a predefined codebook. FashionSAP [11] proposed pre-training tasks using masked text modeling (i.e., prompt token prediction, token replacement prediction) and achieved state-of-the-art performance in fashion $\\mathrm{V} + \\mathrm{L}$ tasks. However, both methods base their proposed pre-training tasks on the use of a fusion encoder and still do not consider the fine-grained features of images and text in the process of deriving unimodal image and text representations for cross-modal alignment.", + "bbox": [ + 212, + 584, + 787, + 737 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we propose Fashion-FINE to enhance fine-grained cross-modal alignment by incorporating global and local representations before multimodal fusion and by encouraging more focus on the loss of hard negatives in ITC, which has not been addressed in previous studies. In Sections 4.2 and 4.3, we demonstrate the effectiveness of our Fashion-FINE on the representative fashion $\\mathrm{V + L}$ tasks through quantitative and qualitative analyses.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Fashion-FINE", + "bbox": [ + 635, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Fashion-FINE", + "text_level": 1, + "bbox": [ + 215, + 143, + 388, + 160 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Baseline Architecture", + "text_level": 1, + "bbox": [ + 215, + 176, + 441, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As illustrated in Figure 2, Fashion-FINE is based on the dual-stream+fusion structure [13, 39] effective for VLP, consisting of one image encoder, one text encoder, and one fusion encoder. We utilize the 12-layers of ViT-B/16 [6] as the image encoder, the first six layers of the BERTbase [5] as the text encoder, and the last six layers of the BERTbase as the fusion encoder. The input image $I$ is encoded into a set of global and local patch embeddings $\\{\\pmb{v}_{cls}, \\pmb{v}_1, \\dots, \\pmb{v}_N\\} \\in \\mathbb{R}^{(N + 1) \\times d_e}$ by the image encoder, where $\\pmb{v}_{cls}$ represents the [CLS] embedding of the input image. The input text $T$ is encoded into a set of global and tokenized local patch embeddings $\\{\\pmb{t}_{cls}, \\pmb{t}_1, \\dots, \\pmb{t}_M\\} \\in \\mathbb{R}^{(M + 1) \\times d_e}$ by the text encoder, where $\\pmb{t}_{cls}$ represents the [CLS] embedding of the input text. The fusion encoder takes the image and text embeddings and performs cross-attention. We maintain two queues to store the representations (i.e., $\\hat{\\pmb{v}}_{cls}, \\hat{\\pmb{t}}_{cls}$ ) of momentum models for conducting large-scale contrastive learning. The architecture of the momentum model remains identical to the original baseline architecture.", + "bbox": [ + 212, + 200, + 787, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Modality-Agnostic Adapter (MAA)", + "text_level": 1, + "bbox": [ + 215, + 434, + 555, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our proposed MAA aims to integrate the global and local representations of images and text by learning the importance of local patch features for each modality through an attention network with shared parameters.", + "bbox": [ + 212, + 459, + 787, + 506 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {V} = \\left\\{\\boldsymbol {v} _ {1}, \\dots , \\boldsymbol {v} _ {N} \\right\\}, \\quad \\boldsymbol {T} = \\left\\{\\boldsymbol {t} _ {1}, \\dots , \\boldsymbol {t} _ {M} \\right\\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 516, + 785, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{V} \\in \\mathbb{R}^{N \\times d_e}$ and $\\mathbf{T} \\in \\mathbb{R}^{M \\times d_e}$ are the sets of local patch embeddings obtained from the image and text encoder outputs, respectively. We introduce modality-agnostic attention to allow the model to learn how to assign patch importance to each local patch embedding for each modality based on the context of the input image-text pair. The local representations $\\mathbf{v}_l \\in \\mathbb{R}^{d_e}$ and $\\mathbf{t}_l \\in \\mathbb{R}^{d_e}$ are computed as follows.", + "bbox": [ + 212, + 544, + 787, + 637 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {l} = \\sum_ {i = 1} ^ {N} \\alpha_ {i} \\boldsymbol {v} _ {i}, \\quad \\boldsymbol {t} _ {l} = \\sum_ {i = 1} ^ {M} \\beta_ {i} \\boldsymbol {t} _ {i} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 647, + 785, + 688 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {u} _ {i} ^ {v} = \\tanh (\\boldsymbol {W} _ {w} \\boldsymbol {v} _ {i} + \\boldsymbol {b} _ {w}), \\quad \\boldsymbol {u} _ {i} ^ {t} = \\tanh (\\boldsymbol {W} _ {w} \\boldsymbol {t} _ {i} + \\boldsymbol {b} _ {w}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 325, + 698, + 785, + 715 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} = \\frac {\\exp \\left(\\boldsymbol {u} _ {i} ^ {v} \\boldsymbol {u} _ {w} ^ {\\top}\\right)}{\\sum_ {k = 1} ^ {N} \\exp \\left(\\boldsymbol {u} _ {k} ^ {v} \\boldsymbol {u} _ {w} ^ {\\top}\\right)}, \\quad \\beta_ {i} = \\frac {\\exp \\left(\\boldsymbol {u} _ {i} ^ {t} \\boldsymbol {u} _ {w} ^ {\\top}\\right)}{\\sum_ {k = 1} ^ {M} \\exp \\left(\\boldsymbol {u} _ {k} ^ {t} \\boldsymbol {u} _ {w} ^ {\\top}\\right)} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 731, + 785, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha_{i}$ and $\\beta_{i}$ measure the importance score of the $i^{th}$ image and text patch embeddings, respectively. Note that the learning parameters (i.e., $\\pmb{W}_{w},\\pmb{b}_{w}$ , and $\\pmb{u}_{w}$ ) are shared across modalities, and this approach encourages interactions between modalities, as demonstrated in how Chen et al. [4] uses shared learning", + "bbox": [ + 215, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Jin et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "parameters as discrete tokens to ground the features of both modalities in a shared manifold space. We compute the integrated representation $\\mathbf{v}_{int}$ and $\\mathbf{t}_{int}$ through the weighted sum of the global and local representations as follows.", + "bbox": [ + 214, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {i n t} = \\lambda_ {I} \\times \\boldsymbol {v} _ {c l s} + (1 - \\lambda_ {I}) \\times \\boldsymbol {v} _ {l} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 205, + 784, + 219 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {t} _ {i n t} = \\lambda_ {T} \\times \\boldsymbol {t} _ {c l s} + (1 - \\lambda_ {T}) \\times \\boldsymbol {t} _ {l} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 223, + 784, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We have observed the performance changes according to the variation of $\\lambda_I$ and $\\lambda_T$ and our model achieved the highest performance when $\\lambda_I = 0.95$ and $\\lambda_T = 0.95$ . Using the integrated image and text representations, our model performs HNM-F (Section 3.3) and C-CmA (Section 3.4), and gradients are propagated not only to the [CLS] embedding, but also to the local patch embeddings. Thus, the use of integrated representations in HNM-F and C-CmA encourages the learning of local patch embeddings for fine-grained cross-modal alignment.", + "bbox": [ + 214, + 250, + 787, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Hard Negative Mining w/ Focal Loss (HNM-F)", + "text_level": 1, + "bbox": [ + 215, + 377, + 653, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our proposed HNM-F aims to improve the distinction learning between positive and hard negatives for a given anchor through the application of focal loss in ITC using integrated image and text representations. To account for the large scale of negatives during HNM-F, we maintain two large queues that store the most recently generated $Q$ integrated representations obtained from the momentum model. We calculate the relative similarity of $i^{th}$ positive image-text and text-image pairs in the mini-batch, respectively, as follows.", + "bbox": [ + 212, + 402, + 782, + 507 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np _ {i 2 t} ^ {(i, +)} = \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {v} _ {i n t} ^ {(i)} , \\hat {\\boldsymbol {t}} _ {i n t} ^ {(+)}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {B + Q} \\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {v} _ {i n t} ^ {(i)} , \\hat {\\boldsymbol {t}} _ {i n t} ^ {(j)}\\right) / \\tau\\right)} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 520, + 784, + 566 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np _ {t 2 i} ^ {(i, +)} = \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {t} _ {\\text {i n t}} ^ {(i)} , \\boldsymbol {v} _ {\\text {i n t}} ^ {(+)}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {B + Q} \\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {t} _ {\\text {i n t}} ^ {(i)} , \\boldsymbol {v} _ {\\text {i n t}} ^ {(j)}\\right) / \\tau\\right)} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 575, + 784, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\pmb{v}_{int}$ and $\\pmb{t}_{int}$ are the integrated representations calculated in Equations 5 and 6, respectively, $B$ denotes a mini-batch size, and $\\tau$ denotes a temperature parameter [12]. $sim(\\pmb{v}_{int},\\hat{\\pmb{t}}_{int})$ and $sim(\\pmb{t}_{int},\\hat{\\pmb{v}}_{int})$ are calculated as $g_v(\\pmb{v}_{int})^\\top g_t'(\\hat{\\pmb{t}}_{int})$ and $g_t(\\pmb{t}_{int})^\\top g_v'(\\hat{\\pmb{v}}_{int})$ in operations within the batch and queues. $g_v$ and $g_t$ are the norm(Adp(\\cdot)) to obtain $\\pmb{v}_{int}' \\in \\mathbb{R}^d$ and $\\pmb{t}_{int}' \\in \\mathbb{R}^d$ , which are the representations output from the original model. $g_v'$ and $g_t'$ are the norm(Adp'(\\cdot)) to obtain $\\hat{\\pmb{v}}_{int}' \\in \\mathbb{R}^d$ and $\\hat{\\pmb{t}}_{int}' \\in \\mathbb{R}^d$ , which are the representations output from the momentum model. Adp(\\cdot) and Adp'(\\cdot) are the modality-specific adaptive layers that project to $d$ dimensions and norm is the normalization function.", + "bbox": [ + 214, + 626, + 782, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "One of the primary characteristics of fashion data is the significant presence of hard negative pairs that share most of the fine-grained attributes with positive pairs for a given anchor. Thus, training ITC by setting the one-hot encoded ground truth similarity for positive and negative pairs to 1 and 0, respectively, can lead to overfitting on the training data. To address this challenge, we adopted", + "bbox": [ + 212, + 765, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Fashion-FINE", + "bbox": [ + 635, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "soft labels by combining pseudo-targets, which are calculated from the momentum image and text encoders, with the one-hot encoded hard labels through a weighted sum. We define the soft labels $\\pmb{q}_{i2t}$ and $\\pmb{q}_{t2i}$ as follows.", + "bbox": [ + 212, + 146, + 782, + 193 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {q} _ {i 2 t} = (1 - \\eta) \\times \\boldsymbol {y} _ {i 2 t} + \\eta \\times \\hat {\\boldsymbol {y}} _ {i 2 t} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 202, + 785, + 217 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {q} _ {t 2 i} = (1 - \\eta) \\times \\boldsymbol {y} _ {t 2 i} + \\eta \\times \\hat {\\boldsymbol {y}} _ {t 2 i} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 220, + 785, + 236 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\pmb{y}_{i2t} \\in \\mathbb{R}^{B \\times (B + Q)}$ and $\\pmb{y}_{t2i} \\in \\mathbb{R}^{B \\times (B + Q)}$ are both represent rectangular diagonal matrices with one-hot encoding, where the main diagonal elements are 1, and all other elements are 0. $\\hat{\\pmb{y}}_{i2t} \\in \\mathbb{R}^{B \\times (B + Q)}$ and $\\hat{\\pmb{y}}_{t2i} \\in \\mathbb{R}^{B \\times (B + Q)}$ are pseudo-targets calculated from the dot product values between the momentum image and text encoders. We experimentally found that the optimal $\\eta$ is 0.4. As demonstrated in [20], minimizing Kullback - Leibler $(\\pmb{p}_{i2t}, \\pmb{q}_{i2t})$ is equivalent to minimizing $-\\sum_{i=1}^{B} q_{i2t}^{(i, +)} \\log p_{i2t}^{(i, +)}$ and maximizing mutual information between $\\pmb{v}_{int}^{(i)}$ and $\\pmb{t}_{int}^{(+)}$ for text that shares similar semantic meaning with the image $I_i$ . Finally, we propose the HNM-F loss with the soft labels as follows.", + "bbox": [ + 215, + 243, + 787, + 388 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {h n m - f} = - \\frac {\\alpha}{2 B} \\sum_ {i = 1} ^ {B} [ (1 - p _ {i 2 t} ^ {(i, +)}) ^ {\\gamma} q _ {i 2 t} ^ {(i, +)} \\log p _ {i 2 t} ^ {(i, +)} \\\\ \\left. + \\left(1 - p _ {t 2 i} ^ {(i, +)}\\right) ^ {\\gamma} q _ {t 2 i} ^ {(i, +)} \\log p _ {t 2 i} ^ {(i, +)} \\right] \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 333, + 396, + 784, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\alpha \\in [0,1]$ is a weighting factor to mitigate the learning issues of the class imbalance between positive and negative pair classes. $(1 - p_{i2t}^{(i, + )})^{\\gamma}$ and $(1 - p_{t2i}^{(i, + )})^{\\gamma}$ are modulating factors based on the relative similarity of positive pairs. $\\gamma \\geq 0$ represents a tunable focusing parameter. A higher $\\gamma$ value promotes increasing the similarity of positive pairs with low similarity and decreasing the relative similarity between positive and hard negatives in HNM-F.", + "bbox": [ + 212, + 465, + 787, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Comprehensive Cross-modal Alignment (C-CmA)", + "text_level": 1, + "bbox": [ + 214, + 579, + 671, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C-CmA aims to enhance fine-grained semantic alignment between text and image representations by learning hierarchical text features of both low-level and high-level features that are difficult to capture with the integrated text representation (Equation 6) alone. First, to extract low-level features in the form of both words and phrases (e.g., \"placket,\" \"cinch belt,\" \"removable fur trim\"), we applied a 1-D convolution to local patch embeddings according to three kernel sizes (i.e., uni-gram, bi-gram, and tri-gram).", + "bbox": [ + 212, + 602, + 787, + 708 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {t} _ {c _ {k}} = f _ {k} (\\boldsymbol {T}), \\quad k \\in \\{1, 2, 3 \\} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 717, + 785, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $f_{k}$ denotes the 1-D convolution layers with kernel size $k$ . $t_{c_k} \\in \\mathbb{R}^{M \\times d_e}$ is the embedding obtained by applying the 1-D convolution with kernel size $k$ . The derivation of the low-level representation $t_c$ is as follows.", + "bbox": [ + 212, + 742, + 787, + 787 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {t} _ {c} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\operatorname {m a x p o o l} \\left(\\left[ \\boldsymbol {t} _ {c _ {1}}, \\boldsymbol {t} _ {c _ {2}}, \\boldsymbol {t} _ {c _ {3}} \\right]\\right), \\quad \\boldsymbol {t} _ {c} \\in \\mathbb {R} ^ {d _ {e}} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 797, + 785, + 837 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Jin et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Next, we added a new fashion symbol token to the text input to obtain a high-level representation of fashion symbols (Figure 2). The text encoder of Fashion-FINE embeds the high-level representation, $\\pmb{t}_s \\in \\mathbb{R}^{d_e}$ , for the nine pre-defined fashion symbols (i.e., tops, dresses, skirts, coats, pants, shoes, bags, accessories, and others). These symbols are pre-defined in FashionSAP based on the categories in FashionGen [33]. The total set of uni-modal embeddings from the text encoder outputs, $T_{total} \\in \\mathbb{R}^{(M + 2) \\times d_e}$ , and the calculated multi-level text representation, $t_m \\in \\mathbb{R}^{d_e}$ , are as follows.", + "bbox": [ + 212, + 146, + 782, + 267 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {T} _ {\\text {t o t a l}} = \\left\\{\\boldsymbol {t} _ {\\text {c l s}}, \\boldsymbol {t} _ {s}, \\boldsymbol {t} _ {1}, \\dots , \\boldsymbol {t} _ {M} \\right\\} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 277, + 784, + 294 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {t} _ {m} = \\delta \\times \\boldsymbol {t} _ {s} + (1 - \\delta) \\times \\boldsymbol {t} _ {c} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 297, + 784, + 313 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In C-CmA, note that we do not use the integrated text representation calculated in Equation 6, but instead use a text representation that integrates both low-level and high-level attributes. We have observed the performance changes according to the variation of $\\delta$ and achieved the highest performance when $\\delta = 0.95$ . The loss function of C-CmA is as follows.", + "bbox": [ + 212, + 321, + 782, + 397 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {i n t} ^ {\\prime} = \\operatorname {n o r m} \\left(\\operatorname {A d p} \\left(\\boldsymbol {v} _ {i n t}\\right)\\right), \\quad \\boldsymbol {t} _ {m} ^ {\\prime} = \\operatorname {n o r m} \\left(\\operatorname {A d p} \\left(\\boldsymbol {t} _ {m}\\right)\\right) \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 407, + 784, + 425 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c - c m a} = \\frac {1}{B} \\left[ 1 - \\sum_ {b = 1} ^ {B} \\frac {1}{2} \\left[ v _ {i n t} ^ {\\prime b} \\left(\\boldsymbol {t} _ {m} ^ {\\prime b}\\right) ^ {\\top} + 1 \\right] \\right] \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 429, + 784, + 469 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\pmb{v}_{int}^{\\prime}$ is a calculated adapted integrated image representation in Section 3.3 and $\\pmb{t}_m^\\prime$ are the adapted multi-level text representation. $\\mathrm{Adp}(\\cdot)$ and norm are the same as the used in Section 3.3, and $B$ denotes the mini-batch size.", + "bbox": [ + 214, + 478, + 782, + 523 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Hard Negative Sampling", + "text_level": 1, + "bbox": [ + 214, + 544, + 465, + 560 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For the interactive V+L retrieval tasks (e.g., text-guided image retrieval), it is important to learn to distinguish the representations of positive and hard negative pairs. For each image-text pair in the mini-batch, we set the image and text as respective anchors and sampled one hard negative for each anchor based on the relative similarity calculated in HNM-F. The sampled two hard negatives for each modality are used to perform ITM in the fusion encoder. Thus, improving hard negative mining through our Fashion-FINE can benefit not only the cross-modal retrieval but also the text-guided image retrieval tasks, which rely on the [CLS] embedding output of the fusion encoder. We have quantitatively and qualitatively verified the effectiveness in Sections 4.2 and 4.3. This section introduces the three fusion encoder-based pre-training tasks used in the previous fashion VLP research [10, 11]. We set the pre-training tasks introduced in this section as the baseline.", + "bbox": [ + 212, + 568, + 784, + 765 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.6 Image-Text Matching (ITM)", + "text_level": 1, + "bbox": [ + 214, + 785, + 496, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In Fashion-FINE, ITM corresponds to a binary classification problem that takes the output [CLS] embeddings based on the cross-attention in the fusion encoder", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Fashion-FINE", + "bbox": [ + 635, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/0fa3c56a632c16a3ffc88c684d29713e8892b0c2d1eb33ef3aa9061d4ae3422c.jpg", + "table_caption": [ + "Table 1: Cross-modal retrieval results on FashionGen with the protocol used in previous fashion VLP research [7,10,11,25,45]." + ], + "table_footnote": [], + "table_body": "
MethodsITRTIRSumR
R@1R@5R@10R@1R@5R@10
VL-BERT [37]19.2639.9046.0522.6336.4848.52212.84
ViLBERT [24]20.9740.4948.2121.1237.2350.11239.18
Image-BERT [30]22.7641.8950.7724.7845.2055.90241.30
OscAR [21]23.3944.6752.5525.1049.1456.68251.53
FashionBERT [7]23.9646.3152.1226.7546.4855.74251.36
KaleidoBERT [45]27.9960.0968.3733.8860.6068.59319.52
CommerceMM [44]41.6064.0072.8039.6061.5072.70352.20
EI-CLIP [25]38.7072.2084.2540.0671.9982.90390.10
ALBEF [20]63.9788.9294.4160.5284.9991.45484.26
FashionVIL [10]65.5491.3496.3061.8887.3293.22495.60
FashionSAP [11]73.1492.8096.8770.1291.7696.38521.07
Fashion-FINE (Ours)74.8093.9497.0070.7091.9096.18524.54
", + "bbox": [ + 328, + 176, + 669, + 323 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "to classify as positive or negative. The ITM loss is as follows.", + "bbox": [ + 215, + 339, + 656, + 354 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i t m} = \\mathbb {E} _ {(\\boldsymbol {I}, \\boldsymbol {T}) \\sim D} \\operatorname {C E} \\left(\\boldsymbol {y} ^ {i t m}, \\boldsymbol {p} ^ {i t m} \\left(\\boldsymbol {H} _ {i t m}\\right)\\right) \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 361, + 364, + 785, + 385 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $D$ denotes the entire dataset comprised of image-text pairs, and CE denotes the cross-entropy loss. $\\pmb{H}_{itm} = [\\pmb{H}_v\\oplus \\pmb{H}_t]$ , and $\\pmb{H}_{v}$ and $\\pmb{H}_{t}$ denote the uni-modal image and text representations, respectively. $\\pmb{p}^{itm}(\\pmb{H}_{itm})$ is the predictive probability distribution for the ITM task, and $\\pmb{y}^{itm}$ is the ground truth.", + "bbox": [ + 214, + 395, + 785, + 455 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3.7 Pre-training Objective", + "text_level": 1, + "bbox": [ + 215, + 493, + 447, + 508 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For learning prompt token prediction $(L_{ptp})$ and token replacement prediction $(L_{trp})$ in the fusion encoder, we employ the same prompt input strategies as FashionSAP [11] (see supplementary material for more details). The total loss for Fashion-FINE is computed as follows. Fashion-FINE is optimized end-to-end for the pre-training datasets by minimizing $L$ .", + "bbox": [ + 214, + 518, + 785, + 595 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {h n m - f} + L _ {c - c m a} + L _ {i t m} + L _ {p t p} + L _ {t r p} \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 606, + 785, + 625 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 646, + 375, + 662 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 215, + 678, + 446, + 694 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset Following previous research [7, 10, 11, 25, 40, 45], we used two representative public benchmarks for pre-training and fashion $\\mathrm{V + L}$ retrieval tasks: FashionGen [33] and FashionIQ [41]. FashionGen contains 260K image-text pairs for the training dataset and 35K for test dataset. FashionIQ contains 18K training triplets (i.e., reference image, modifying text, and target image) and 6K validation triplets across three categories: dress, shirt, and toptee. For fair experiments, we pre-trained our model using FashionGen, consistent with previous studies. For the downstream tasks, we fine-tuned our model on FashionGen for cross-modal retrieval and on FashionIQ for text-guided image retrieval.", + "bbox": [ + 212, + 703, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Jin et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/cec4478470228318b64c07c432f45a22951c53fe48ee9b9583b37bf400517e26.jpg", + "table_caption": [ + "FashionGen with full evaluation." + ], + "table_footnote": [], + "table_body": "
MethodsITRTIRSumR
R@1R@5R@10R@1R@5R@10
EI-CLIP [25]25.7054.5066.8028.4057.1069.40301.90
ALBEF [20]41.6867.3975.5050.9575.3684.15395.03
FashionKLIP [40]47.1669.2775.3954.6075.0681.39402.87
FashionViL [10]42.8871.5780.5551.3475.4284.57406.33
FashionSAP [11]54.4377.3083.1562.8283.9690.16451.82
Fashion-FINE (Ours)54.2977.7483.8765.2385.6191.37458.11
", + "bbox": [ + 218, + 189, + 496, + 268 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/0eee957202a4c31564cce61395702fb52512e66dffc3c91cf3fa6418d7ac9355.jpg", + "table_caption": [ + "Table 3: Text-guided image retrieval re", + "Table 2: Cross-modal retrieval results on sutls on FashionIQ." + ], + "table_footnote": [], + "table_body": "
MethodsDressToptieShirtSumR
R@10R@50R@10R@50R@10R@50
FiLM [28]14.2333.3417.3037.6815.0434.09151.68
CIRPLANT [23]17.4540.4121.6445.3817.5338.81181.22
VAL [3]22.5344.0027.5351.6822.3844.15212.27
TIRG(bert) [38]27.1753.2527.8457.1122.2845.58233.23
CosMo [17]25.6450.3029.2157.4624.9049.18236.69
DCNet [15]28.9556.7030.4458.2923.9547.30245.63
FashionVIL [10]33.4759.9434.9860.7925.1750.39264.74
FashionVLP [8]32.4260.2938.5168.7931.8958.44290.34
FashionSAP [11]33.7160.4341.9170.9333.1761.33301.48
Fashion-FINE (Ours)36.6463.4646.8174.0938.7663.30323.06
", + "bbox": [ + 503, + 176, + 777, + 281 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/dbe38161ba97ec9dc269444a2918a3a490ba071e2c44c0264552ea9ae8e03879.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 228, + 282, + 498, + 444 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/48bea441cc2904aa28dbdf16c695c6a7022613daa5e88cfeb6b052275009b11a.jpg", + "image_caption": [ + "(b)", + "Fig. 3: Qualitative analysis results. (a) TIR: the green text in the query indicates the distinguished attributes of the primary ground truth. (b) TGIR: the green text in the modifying text indicates attributes that need to be modified in the reference image and included in the target image. Each row shows the top-10 retrieved results for each model (far left: top-1, far right: top-10), and the green box represents the ground truth." + ], + "image_footnote": [], + "bbox": [ + 504, + 282, + 777, + 444 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details We trained the proposed modules with randomly initialized weights, and the baseline encoders were initialized with the same pre-trained weights as in FashionSAP [11]. Through hyperparameter search experiments, we found that the optimal values for HNM-F are $\\tau = 0.07$ , $\\alpha = 0.25$ , and $\\gamma = 2.0$ (see supplementary material for more details).", + "bbox": [ + 212, + 566, + 787, + 643 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Cross-modal Retrieval", + "text_level": 1, + "bbox": [ + 215, + 664, + 446, + 679 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative Analysis Following the previous study [11], we evaluated cross-modal retrieval performance based on the calculated relative similarity between the integrated uni-modal representations. Thus, we fine-tuned our model using MAA, HNM-F, and C-CmA. ITR aims to retrieve matched text for an input query image, while TIR aims to retrieve a matched image for an input query text. We evaluated our Fashion-FINE using the following two evaluation methods.", + "bbox": [ + 212, + 688, + 787, + 777 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "First, following previous research [7, 10, 11, 25, 45], we evaluated the average performance result of five randomly sampled retrieval test sets containing 1,000 unique image and text queries. For each query in the test sets, only one candidate is matched (positive), while the other 100 candidates are mismatched (negative)", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Fashion-FINE", + "bbox": [ + 635, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ba351f8ad192219823c14ca9b8c1271300e7bce4f7a1e3768e9688ffad03c671.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Attributesturtleneckbomberjacketclosurestraightdenimdistresslongsleevearrow
Baseline(Fashion-FINEw/o maa, hmm-f,c-cma)
Fashion-FINE
", + "bbox": [ + 218, + 143, + 785, + 268 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Fig. 4: Qualitative analysis of the fine-grained alignment in the pre-training phase. The visualized results are the 1st layer of the cross-attention map in the fusion encoder.", + "bbox": [ + 215, + 277, + 787, + 308 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "and randomly sampled from other products within the same subcategory. As a result, Fashion-FINE achieved state-of-the-art performance in ITR and TIR based on the SumR metric, as shown in Table 1. Second, we used the full test dataset to thoroughly evaluate the performance and reproducibility of Fashion-FINE, consistent with previous research [10,11,25,40]. As a result, Fashion-FINE also achieved state-of-the-art performance in ITR and TIR, as shown in Table 2.", + "bbox": [ + 212, + 345, + 787, + 439 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative Analysis We analyzed the cases where the ground truth retrieval rank increased in the TIR results through the ablation study (Figure 3a). The baseline model retrieved hard negatives that share the fine-grained attributes (e.g., 'plaid jacket,' 'spread collar,' and 'single-button barrel cuffs') with the ground truth at a higher rank than the ground truth. Fashion-FINE (w/o C-CmA) ranked the ground truth in 4th place, yet it ranked hard negatives-sharing the same attributes as the ground truth but differing in location or having slight color variations-ahead of the ground truth. On the other hand, Fashion-FINE ranked all ground truths between 1st and 4th place. These results represent superior qualitative achievements not captured by the $R@k$ metric presented in Table 2, which considers a retrieval correct if the retrieved top-k list includes at least one correct answer. Through the analyses, we have qualitatively demonstrated the effectiveness of Fashion-FINE for fine-grained cross-modal retrieval.", + "bbox": [ + 212, + 459, + 787, + 657 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Text-Guided Image Retrieval", + "text_level": 1, + "bbox": [ + 215, + 693, + 503, + 709 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Quantitative Analysis We fine-tuned our model using contrastive learning based on the similarity between the output [CLS] embedding of the fusion encoder for the input query image-text pair and the output [CLS] embedding of the image encoder in Fashion-FINE for the target image. This task aims to retrieve a target image by referencing an image-text pair as a query, where the text modifies certain attributes of the image. We compared our model with fashion VLP methods under the original protocol used by FashionIQ. As a result, Fashion-FINE achieved state-of-the-art performance in TGIR, as shown in Table 3.", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Jin et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/71fce061dc5cd1f1ecc736eea5b95abb37969948a766842c7ce3e760d54bd7d2.jpg", + "table_caption": [ + "Table 4: Ablation study results for our proposed modules on FashionGen and FashionIQ. We used the baseline architecture (Section 3.1) in this ablation study." + ], + "table_footnote": [], + "table_body": "
ArchitectureProposed ModulesCross-modal RetrievalText-Guided Image Retrieval
MAA HNM-F C-CmAITRTIRSumRDressTopteeShirt
R@1R@5R@10R@1R@5R@10R@10R@50R@10R@50
BaselineXXX54.4377.3083.1562.8283.9690.16451.8233.7160.4333.17
Baseline✓XX54.3377.4283.7564.6284.0790.18454.3732.9760.7835.43
(w/ Ours)✓✓X54.3077.8184.0064.4085.4290.74456.6736.2462.3238.03
✓✓✓54.2977.7483.8765.2385.6191.37458.1136.6463.4638.76
", + "bbox": [ + 218, + 176, + 782, + 262 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/9db9766806f6c90f56a91202c83f5fdbbc90a0aa8d5649c3059109fe49c3214a.jpg", + "table_caption": [ + "Table 5: Experimental results on the applicability of Fashion-FINE to the foundation model (i.e., BLIP [19]) on FashionGen." + ], + "table_footnote": [], + "table_body": "
ArchitectureProposed ModulesCross-modal Retrieval
MAAHNM-FC-CmAITRTIR
R@1R@5R@10R@1R@5
BLIPXXX55.2778.3985.0661.6382.46
BLIP (w/ Ours)✓XX54.6078.6085.1265.0585.65
✓✓X57.2480.2086.5363.3586.41
✓✓✓59.8182.4888.1268.4987.98
", + "bbox": [ + 316, + 297, + 679, + 383 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Qualitative Analysis We qualitatively analyzed the cases where the rank of the retrieved ground truth increased in the TGIR task according to the ablation study (Figure 3b). The baseline model failed to retrieve the ground truth in the top-10, highly ranking hard negatives that contain \"no sleeves\" or \"two different colors.\" On the other hand, we observed that as more modules of Fashion-FINE were incorporated, the rank of the ground truth increased. This also qualitatively demonstrates the effectiveness of Fashion-FINE in learning fine-grained details, which is not captured in Table 3. Through the analyses, we have demonstrated that Fashion-FINE effectively learned not only the representation of the target image, but also the fine-grained alignment in the fusion encoder to obtain representations of the input reference image-modifying text pairs.", + "bbox": [ + 212, + 400, + 787, + 566 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Fine-grained Alignment Analysis", + "text_level": 1, + "bbox": [ + 215, + 588, + 532, + 603 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To verify the fine-grained alignment in the fusion encoder, following [11], we chose the four instances from FashionGen and visualized cross-attention maps using Grad-CAM [35]. In Figure 4, Fashion-FINE more accurately highlights the regions of fine-grained attributes (e.g., \"arrow,\" \"distress\") while minimizing the highlight on the background compared to the baseline. We also found that Fashion-FINE effectively highlights the region of attributes for both single words (e.g. \"closure,\" \"arrow,\" \"straight,\" \"denim,\" \"distress\") and phrases (e.g., \"turtle neck,\" \"bomber jacket,\" \"long sleeve\") in the form of fine-grained attributes.", + "bbox": [ + 212, + 612, + 787, + 733 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 755, + 387, + 770 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We conducted an ablation study to evaluate the effectiveness of the proposed modules. For fair experiments, the experimental settings are consistent in the same ablation series. Compared to the baseline, our methods achieved significant performance improvements in the $\\mathrm{V + L}$ retrieval tasks (i.e., TIR and TGIR).", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Fashion-FINE", + "bbox": [ + 635, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "These results indicate that our methods were effective in learning to distinguish the positives from the hard negatives among the image candidates for given input query text or query image-text pairs in fine-grained V+L retrieval tasks. This is supported by the significant performance improvement results in TIR and TGIR tasks, where the disentangled embedding of images is important. The notable performance improvement in TGIR resulting from the inclusion of the HNM-F module indicates that our approach effectively learned not only the disentangled embedding of the input image but also the representation of the input image-text pair (i.e., reference image and modifying text) in the fusion encoder.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.6 Applicability of Fashion-FINE to the Foundation Model", + "text_level": 1, + "bbox": [ + 214, + 300, + 722, + 316 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We conducted experiments to verify the applicability of our Fashion-FINE to foundational models. For the experiments, we performed an ablation study of our modules based on the dual-stream+fusion foundational model architecture, BLIP [19]. Since BLIP uses bootstrapping on noisy data during the pre-training phase with a decoder, we applied our modules in a fine-tuning phase only, which was designed without a decoder, to the pre-trained BLIP model to evaluate the pure performance of our modules. The encoder configuration for the ITC and ITM tasks in the BLIP architecture is the same as that of Fashion-FINE, so we applied Fashion-FINE to BLIP in the same way as the baseline architecture described in Section 3. In the fine-tuning phase for the $\\mathrm{V + L}$ tasks (Sections 4.2 and 4.3), since we applied the modules of Fashion-FINE only for ITR and TIR and not for TGIR, the application of Fashion-FINE to BLIP was also done only for ITR and TIR. In Table 5, we have demonstrated that the modules of Fashion-FINE can be successfully applied to the BLIP architecture, a foundational model.", + "bbox": [ + 212, + 321, + 787, + 534 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 553, + 359, + 569 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we presented Fashion-FINE, which aims to improve the learning of fine-grained cross-modal alignment. Fashion-FINE addresses the challenge of distinguishing positives from hard negatives based on MAA, which integrates global features with the fine-grained features of local representations, and on HNM-F and C-CmA, which enhance the learning effects of hard negative mining and refine the integrated image representations. Unlike previous fashion VLP research, Fashion-FINE focuses on cross-modal alignment before multimodal fusion, taking into account fine-grained details. Through quantitative and qualitative analyses, we have validated that Fashion-FINE can be successfully integrated with fusion encoder-based pre-training tasks. With this work, we advance the fashion VLP architecture and expect that our approach will contribute to the research on the development of fine-grained foundation methods.", + "bbox": [ + 212, + 580, + 787, + 763 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 782, + 403, + 799 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported by the National Research Foundation of Korea (NRF) grant (No.2021M3A9E4080780, No.2018R1A5A7059549) and by the Institute of", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Jin et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean government (MSIT) (No.RS-2020-II201373, Artificial Intelligence Graduate School Program (Hanyang University)).", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 212, + 321, + 227 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Byun, J., Hwang, T., Fu, J., Moon, T.: Grit-vlp: Grouped mini-batch sampling for efficient vision and language pre-training. In: European Conference on Computer Vision. pp. 395-412. Springer (2022)", + "2. Chen, Q., Huang, J., Feris, R., Brown, L.M., Dong, J., Yan, S.: Deep domain adaptation for describing people based on fine-grained clothing attributes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5315-5324 (2015)", + "3. Chen, Y., Gong, S., Bazzani, L.: Image search with text feedback by visiolinguistic attention learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3001-3011 (2020)", + "4. Chen, Y., Yuan, J., Tian, Y., Geng, S., Li, X., Zhou, D., Metaxas, D.N., Yang, H.: Revisiting multimodal representation in contrastive learning: from patch and token embeddings to finite discrete tokens. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15095-15104 (2023)", + "5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)", + "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "7. Gao, D., Jin, L., Chen, B., Qiu, M., Li, P., Wei, Y., Hu, Y., Wang, H.: Fashion-bert: Text and image matching with adaptive loss for cross-modal retrieval. In: Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 2251-2260 (2020)", + "8. Goenka, S., Zheng, Z., Jaiswal, A., Chada, R., Wu, Y., Hedau, V., Natarajan, P.: Fashionvlp: Vision language transformer for fashion retrieval with feedback. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14105-14115 (2022)", + "9. Gunel, B., Du, J., Conneau, A., Stoyanov, V.: Supervised contrastive learning for pre-trained language model fine-tuning. International Conference on Learning Representations (2021)", + "10. Han, X., Yu, L., Zhu, X., Zhang, L., Song, Y.Z., Xiang, T.: Fashionvil: Fashion-focused vision-and-language representation learning. In: European Conference on Computer Vision. pp. 634-651. Springer (2022)", + "11. Han, Y., Zhang, L., Chen, Q., Chen, Z., Li, Z., Yang, J., Cao, Z.: Fashionsap: Symbols and attributes prompt for fine-grained fashion vision-language pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15028-15038 (2023)", + "12. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)", + "13. Jang, J., Kong, C., Jeon, D., Kim, S., Kwak, N.: Unifying vision-language representation space with single-tower transformer. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 980-988 (2023)" + ], + "bbox": [ + 218, + 239, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Fashion-FINE", + "bbox": [ + 635, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)", + "15. Kim, J., Yu, Y., Kim, H., Kim, G.: Dual compositional learning in interactive image retrieval. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1771-1779 (2021)", + "16. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017)", + "17. Lee, S., Kim, D., Han, B.: Cosmo: Content-style modulation for image retrieval with text feedback. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 802-812 (2021)", + "18. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: bootstrapping language-image pretraining with frozen image encoders and large language models. In: ICML (2023)", + "19. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)", + "20. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021)", + "21. Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., et al.: Oscar: Object-semantics aligned pre-training for vision-language tasks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16. pp. 121-137. Springer (2020)", + "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "23. Liu, Z., Rodriguez-Opazo, C., Teney, D., Gould, S.: Image retrieval on real-life images with pre-trained vision-and-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2125-2134 (2021)", + "24. Lu, J., Batra, D., Parikh, D., Lee, S.: Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems 32 (2019)", + "25. Ma, H., Zhao, H., Lin, Z., Kale, A., Wang, Z., Yu, T., Gu, J., Choudhary, S., Xie, X.: Ei-clip: Entity-aware interventional contrastive learning for e-commerce cross-modal retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18051-18061 (2022)", + "26. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)", + "27. Ordonez, V., Kulkarni, G., Berg, T.: Im2text: Describing images using 1 million captioned photographs. Advances in neural information processing systems 24 (2011)", + "28. Perez, E., Strub, F., De Vries, H., Dumoulin, V., Courville, A.: Film: Visual reasoning with a general conditioning layer. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018)", + "29. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Jin et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image-to-sentence models. In: Proceedings of the IEEE international conference on computer vision. pp. 2641-2649 (2015)", + "30. Qi, D., Su, L., Song, J., Cui, E., Bharti, T., Sacheti, A.: Imagebert: Cross-modal pre-training with large-scale weak-supervised image-text data. arXiv preprint arXiv:2001.07966 (2020)", + "31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "32. Robinson, J., Chuang, C.Y., Sra, S., Jegelka, S.: Contrastive learning with hard negative samples. International Conference on Learning Representations (2021)", + "33. Rostamzadeh, N., Hosseini, S., Boquet, T., Stokowiec, W., Zhang, Y., Jauvin, C., Pal, C.: Fashion-gen: The generative fashion dataset and challenge. arXiv preprint arXiv:1806.08317 (2018)", + "34. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015)", + "35. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)", + "36. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2556-2565 (2018)", + "37. Su, W., Zhu, X., Cao, Y., Li, B., Lu, L., Wei, F., Dai, J.: Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530 (2019)", + "38. Vo, N., Jiang, L., Sun, C., Murphy, K., Li, L.J., Fei-Fei, L., Hays, J.: Composing text and image for image retrieval—an empirical odyssey. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6439–6448 (2019)", + "39. Wang, J., Zhou, P., Shou, M.Z., Yan, S.: Position-guided text prompt for vision-language pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23242-23251 (2023)", + "40. Wang, X., Wang, C., Li, L., Li, Z., Chen, B., Jin, L., Huang, J., Xiao, Y., Gao, M.: Fashionclip: Enhancing e-commerce image-text retrieval with fashion multimodal conceptual knowledge graph. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 5: Industry Track). pp. 149-158 (2023)", + "41. Wu, H., Gao, Y., Guo, X., Al-Halah, Z., Rennie, S., Grauman, K., Feris, R.: Fashion iq: A new dataset towards retrieving images by natural language feedback. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 11307-11317 (2021)", + "42. Wu, M., Mosse, M., Zhuang, C., Yamins, D., Goodman, N.: Conditional negative sampling for contrastive learning of visual representations. In: International Conference on Learning Representations (2020)", + "43. Yang, J., Duan, J., Tran, S., Xu, Y., Chanda, S., Chen, L., Zeng, B., Chilimbi, T., Huang, J.: Vision-language pre-training with triple contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15671-15680 (2022)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Fashion-FINE", + "bbox": [ + 635, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "44. Yu, L., Chen, J., Sinha, A., Wang, M., Chen, Y., Berg, T.L., Zhang, N.: Commercemm: Large-scale commerce multimodal representation learning with omni retrieval. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 4433-4442 (2022)", + "45. Zhuge, M., Gao, D., Fan, D.P., Jin, L., Chen, B., Zhou, H., Qiu, M., Shao, L.: Kaleido-bert: Vision-language pre-training on fashion domain. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12647-12657 (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 258 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "S. Jin et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Integration of Global and Local Representations for Fine-grained Cross-modal Alignment/35329370-3365-4c03-a82d-856d1e6d798d_model.json b/2024/Integration of Global and Local Representations for Fine-grained Cross-modal Alignment/35329370-3365-4c03-a82d-856d1e6d798d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b51c85526aa0e458469faa6d3820552e2061940f --- /dev/null +++ b/2024/Integration of Global and Local Representations for Fine-grained Cross-modal Alignment/35329370-3365-4c03-a82d-856d1e6d798d_model.json @@ -0,0 +1,2502 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.227, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "Integration of Global and Local Representations for Fine-grained Cross-modal Alignment" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.214, + 0.765, + 0.229 + ], + "angle": 0, + "content": "Seungwan Jin\\(^{\\text{®}}\\), Hoyoung Choi\\(^{\\text{®}}\\), Taehyung Noh\\(^{\\text{®}}\\), and Kyungsik Han\\(^{\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.241, + 0.734, + 0.27 + ], + "angle": 0, + "content": "Hanyang University, South Korea {seungwanjin, choiho00, yestaehyung, kyungsikhan}@hanyang.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.312, + 0.744, + 0.617 + ], + "angle": 0, + "content": "Abstract. Fashion is one of the representative domains of fine-grained Vision-Language Pre-training (VLP) involving a large number of images and text. Previous fashion VLP research has proposed various pretraining tasks to account for fine-grained details in multimodal fusion. However, fashion VLP research has not yet addressed the need to focus on (1) uni-modal embeddings that reflect fine-grained features and (2) hard negative samples to improve the performance of fine-grained \\(\\mathrm{V + L}\\) retrieval tasks. In this paper, we propose Fashion-FINE (Fashion VLP with Fine-grained Cross-modal Alignment using the INtegrated representations of global and local patch Embeddings), which consists of three key modules. First, a modality-agnostic adapter (MAA) learns uni-modal integrated representations and reflects fine-grained details contained in local patches. Second, hard negative mining with focal loss (HNM-F) performs cross-modal alignment using the integrated representations, focusing on hard negatives to boost the learning of fine-grained cross-modal alignment. Third, comprehensive cross-modal alignment (C-CmA) extracts low- and high-level fashion information from the text and learns the semantic alignment to encourage disentangled embedding of the integrated image representations. Fashion-FINE achieved state-of-the-art performance on two representative public benchmarks (i.e., FashionGen and FashionIQ) in three representative \\(\\mathrm{V + L}\\) retrieval tasks, demonstrating its effectiveness in learning fine-grained features." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.631, + 0.741, + 0.659 + ], + "angle": 0, + "content": "Keywords: Vision-Language Pre-training \\(\\cdot\\) Fine-grained Representation Learning \\(\\cdot\\) Fashion" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.688, + 0.377, + 0.705 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.724, + 0.788, + 0.816 + ], + "angle": 0, + "content": "With recent advances in general VLP research that aims to learn common sense knowledge, there has been increasing attention to fine-grained cross-modal representation learning. Fashion is one of the most representative VLP research domains that applies the concept of cross-modal representations based on large-scale image and text information, about fashion items for fine-grained \\(\\mathrm{V + L}\\) retrieval tasks such as cross-modal retrieval and text-guided image retrieval." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.383, + 0.841 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "S. Jin et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.148, + 0.406, + 0.159 + ], + "angle": 0, + "content": "(a) Image-Text Pairs in Fashion Domain" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.161, + 0.284, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.168, + 0.413, + 0.213 + ], + "angle": 0, + "content": "Long sleeve down parka-style coat in olive green. Genuine raccon fur trim at hood. Two-way zip closure at front concealed by button placket. Drawstring at collar. Welt pockets and flap pockets at front. Logo patch at upper sleeve. Pleated at back panel. Fully lined. Tonal stitching. Elasticized sleeve cuffs." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.218, + 0.283, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.221, + 0.412, + 0.268 + ], + "angle": 0, + "content": "Long sleeve down parka-style coat in olive green. Removable genuine raccoon fur trim at hood. Two-way zip closure at front concealed by press-stud packet. Welt pockets and flap pockets at front. Logo patch at upper sleeve. Drawstring at waist. Fully lined. Tonal stitching. Inset ribbed knit sleeve cuff." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.272, + 0.283, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.274, + 0.414, + 0.322 + ], + "angle": 0, + "content": "Long sleeve down parka in black. Removable hood with cinch-belt detail at crown and wired removable fur trim. Tonal bungee drawstrings at hood. Stand collar. Two-way zip closure at front concealed by press-stud placket. Tonal logo applique at exterior arm. Welt and flap pockets at front with tonal signature patch..." + }, + { + "type": "image_caption", + "bbox": [ + 0.434, + 0.148, + 0.71, + 0.159 + ], + "angle": 0, + "content": "(b) Our Motivation for Fine-grained Cross-modal Alignment" + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.161, + 0.783, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.341, + 0.788, + 0.453 + ], + "angle": 0, + "content": "Fig. 1: Our motivation for this research. (a) shows an example of three image-text pair cases (i.e., positive, hard negative, and easy negative pairs) in fashion products, where the colored areas specifically indicate fashion attributes that distinguish positives from hard negatives. (b) shows that fashion \\(\\mathrm{V + L}\\) tasks (e.g., cross-modal retrieval) are essential for distinguishing positives from hard negatives given anchors by capturing the fine-grained features. Our motivation is that it is difficult to distinguish between positives and hard negatives using only global representations, but this can be compensated by integrating local representations that reflect fine-grained attributes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.789, + 0.7 + ], + "angle": 0, + "content": "Most fashion VLP research [7, 8, 10, 11, 25, 45] has been evolved based on the general VLP architectures by applying self-supervised pre-training tasks, such as image-text contrastive learning (ITC) and image-text matching (ITM). Recently, employing the dual-stream+fusion [13, 39] structure, fashion VLP research [7, 10, 11, 45] proposed various pre-training tasks (e.g., masked patch modeling) to address the fine-grained features using the fusion encoder. FashionSAP [11], the current state-of-the-art model, proposed high-level multimodal semantic alignment and used the relative similarity between the global representations (i.e., [CLS] embedding) obtained before fusion to perform \\(\\mathrm{V} + \\mathrm{L}\\) retrieval tasks instead of the ITM score calculated in the fusion encoder. However, using only global representations has limitations in learning fine-grained features in the cross-modal alignment and poses challenges in distinguishing between positive and hard negatives for a given anchor during the test phase (Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.79, + 0.841 + ], + "angle": 0, + "content": "To address this, conventional methods have focused on learning to embed positive image-text pairs closer and negative pairs further using loss functions such as InfoNCE [26]. However, in self-supervised VLP, since only the pre-matched image-text pairs are considered positive and all others are considered negative, a severe class imbalance problem may arise due to the difference between the number of positive and negative pairs. Moreover, the overwhelming number of easy negative pairs compared to hard negative pairs leads to a dominance of the loss scale for easy negatives over hard negatives, thereby limiting the learning of the model to distinguish between positive and hard negative pairs." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.636, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Fashion-FINE" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "In this paper, we propose Fashion-FINE (Fashion VLP with Fine-grained cross-modal alignment using the INtegrated representations of global and local patch Embeddings), which learns cross-modal alignment in three phases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.403 + ], + "angle": 0, + "content": "The first phase is to learn the importance of each local patch embedding for images and text using a modality-agnostic adapter (MAA). For each modality, we obtain a local representation that reflects the feature importance calculated by MAA at each patch location, and this local representation is then combined with the global representation, specifically the [CLS] embedding, to finally form the integrated representation. The second phase is the hard negative mining with focal loss (HNM-F), which integrates ITC and focal loss to mitigate the class and loss imbalance issues in ITC and to focus on learning to distinguish between positive and hard negative pairs. The third phase is the comprehensive cross-modal alignment (C-CmA), which semantically aligns the low-level and high-level fashion information in text descriptions with the integrated image representations. Through this, Fashion-FINE learns to encourage disentangled embedding of the image representations in the multimodal space, taking into account the comprehensive fashion text information." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.406, + 0.568, + 0.42 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.432, + 0.784, + 0.491 + ], + "angle": 0, + "content": "- We propose Fashion-FINE, which performs fine-grained cross-modal alignment by integrating global and local representations of the two modalities. Fashion-FINE has achieved state-of-the-art performance in all fashion V+L retrieval tasks." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.494, + 0.785, + 0.539 + ], + "angle": 0, + "content": "- We have improved the architecture for fashion VLP research by integrating Fashion-FINE with existing fusion encoder-based pre-training tasks, and we have verified its applicability through performance improvements." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.54, + 0.785, + 0.585 + ], + "angle": 0, + "content": "- Our study is the first to quantitatively and qualitatively validate the importance of reflecting local patch embeddings in fine-grained cross-modal alignment and the boosting effect of hard negative mining with focal loss." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.432, + 0.785, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.611, + 0.388, + 0.627 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.646, + 0.564, + 0.663 + ], + "angle": 0, + "content": "2.1 Vision-Language Pre-training (VLP)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Many VLP methods have achieved significant performance in various downstream \\(\\mathrm{V + L}\\) tasks. The most successful VLP architectures fall into two categories: dual-stream and dual-stream \\(^+\\) fusion [13, 39]. The dual-stream architecture uses two modality-specific encoders to perform pre-training tasks (e.g., ITC) for semantic alignment between the two modalities. For example, CLIP [31] and ALIGN [14] have used ITC to train image-text alignment on large-scale multimodal data, achieving remarkable performance in general downstream \\(\\mathrm{V + L}\\) tasks. While this dual-stream is advantageous for achieving high performance, it requires extremely large-scale data during pre-training. On the other hand, the dual-stream \\(^+\\) fusion architecture places multimodal fusion encoders on top of two modality-specific encoders to perform ITC followed by cross-modal learning (e.g.," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "S. Jin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.177 + ], + "angle": 0, + "content": "MLM and ITM). ALBEF [20] and TCL [43], based on the dual-stream+fusion, have achieved superior performance compared to dual-stream models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.312 + ], + "angle": 0, + "content": "Existing VLP methods have been trained on noisy and large multimodal Web data due to the limited amount of high-quality human-annotated image-text pair data [14]. Recently, BLIP [19] and BLIP-2 [18] have achieved superior performance compared to ALBEF and TCL by expanding the dual-stream+fusion structure by proposing a decoder to bootstrap noisy text captions and Q-Former to learn soft visual prompt for bootstrapping a decoder-based large language model (LLM), respectively. Rather than focusing on human-annotated image-text pairs, BLIP-based methods tackle the inherent challenges in processing the noisy and large-scale multimodal datasets collected from the Web." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.435 + ], + "angle": 0, + "content": "The public benchmark datasets (e.g., FashionGen [33], FashionIQ [41]) addressed in recent fashion VLP research [10, 11, 25, 40] consist of multimodal datasets annotated by humans who manually describe fine-grained fashion attributes and contain no (or less) noise, unlike Web data. Thus, we set a pure dual-stream+fusion architecture without decoders as the baseline of FashionFINE, which is consistent with recent fashion VLP. By integrating our proposed modules with the pre-training tasks proposed in the prior work, we advance the fashion VLP architecture for the fine-grained \\(\\mathrm{V + L}\\) retrieval tasks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.455, + 0.515, + 0.47 + ], + "angle": 0, + "content": "2.2 Hard Negative Mining in VLP" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.675 + ], + "angle": 0, + "content": "Hard negative mining is an important technique used in machine learning, especially in fields such as computer vision and natural language processing, to improve model performance [9,32,34,42]. It involves focused sampling and training on the sampled hard negatives, which are semantically similar to positive data but are negative. In VLP, hard negative mining has been recently proposed to learn to distinguish between positive and hard negative samples in the image-text matching (ITM) task by extracting hard negative samples based on the calculated relative similarity during the cross-modal alignment phase [20, 43]. To enhance the effectiveness of the hard negative sampling, GRIT-VLP [1] proposed a grouped mini-batch sampling strategy to improve in-batch hard negative sampling in dual-stream \\(^+\\) fusion architecture. However, the previous studies still rely solely on the use of global representations, which are limited in reflecting fine-grained details when calculating relative similarity for negative sampling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.787, + 0.721 + ], + "angle": 0, + "content": "To the best of our knowledge, no research has been aimed at learning fine-grained local features to distinguish positive from hard negative in the cross-modal alignment to improve fine-grained \\(\\mathrm{V + L}\\) tasks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.741, + 0.486, + 0.756 + ], + "angle": 0, + "content": "2.3 VLP in Fashion \\(\\mathbf{V} + \\mathbf{L}\\) tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Compared to general VLP that uses coarse-grained common sense knowledge [16, 22, 27, 29, 36], fashion VLP should focus more on fine-grained attributes in the multimodal fashion data [2]. FashionBERT [7] is the first cross-modal retrieval method targeting the fashion domain. It uses a transformer-based fusion encoder and introduces masked patch modeling to enable contextual learning of images" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.636, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Fashion-FINE" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.488, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.275, + 0.477, + 0.284 + ], + "angle": 0, + "content": "(a) Fine-grained Cross-modal Alignment before Multimodal Fusion" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.144, + 0.626, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.276, + 0.612, + 0.285 + ], + "angle": 0, + "content": "(b) Hard Negative Sampling" + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.144, + 0.782, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.275, + 0.78, + 0.284 + ], + "angle": 0, + "content": "(c) Pre-training Tasks in Fusion Encoder" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.299, + 0.785, + 0.384 + ], + "angle": 0, + "content": "Fig. 2: An overview of Fashion-FINE. (a): Fine-grained cross-modal alignment through our proposed three modules (i.e., MAA, HNM-F, and C-CmA). (b): Sampling hard negatives from the mini-batch for each modality (i.e., image and text). (c): Fusion encoder-based pre-training tasks (i.e., image-text matching (ITM), prompt token prediction (PTP), and token replacement prediction (TRP)) for the given positive pair and the two hard negative pairs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.572 + ], + "angle": 0, + "content": "and text. KaleidoBERT [45] is a method for semantic alignment with word tokens after extracting image patches through various patch methods (e.g., Rotation recognition, Jigsaw puzzle solving). In fashion e-commerce, words may be used differently from their common meaning. For example, \"diesel\" and \"canada goose\" are the brand names, not a type of gas for vehicles or an animal in Canada. EICLIP [25] proposed semantic alignment methods that consider the homonymous nature of e-commerce language. However, the previous studies still had to rely on the general image-text contrastive learning method for cross-modal alignment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.788, + 0.738 + ], + "angle": 0, + "content": "Recently, representation learning methods applied to the fashion domain have been proposed using the dual-stream+fusion architecture of general VLP methods. FashionViL [10] introduced masked patch feature classification (MPFC), which learns characteristics of image patches based on discrete codes from a predefined codebook. FashionSAP [11] proposed pre-training tasks using masked text modeling (i.e., prompt token prediction, token replacement prediction) and achieved state-of-the-art performance in fashion \\(\\mathrm{V} + \\mathrm{L}\\) tasks. However, both methods base their proposed pre-training tasks on the use of a fusion encoder and still do not consider the fine-grained features of images and text in the process of deriving unimodal image and text representations for cross-modal alignment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we propose Fashion-FINE to enhance fine-grained cross-modal alignment by incorporating global and local representations before multimodal fusion and by encouraging more focus on the loss of hard negatives in ITC, which has not been addressed in previous studies. In Sections 4.2 and 4.3, we demonstrate the effectiveness of our Fashion-FINE on the representative fashion \\(\\mathrm{V + L}\\) tasks through quantitative and qualitative analyses." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "S. Jin et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.39, + 0.161 + ], + "angle": 0, + "content": "3 Fashion-FINE" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.177, + 0.442, + 0.192 + ], + "angle": 0, + "content": "3.1 Baseline Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.202, + 0.789, + 0.414 + ], + "angle": 0, + "content": "As illustrated in Figure 2, Fashion-FINE is based on the dual-stream+fusion structure [13, 39] effective for VLP, consisting of one image encoder, one text encoder, and one fusion encoder. We utilize the 12-layers of ViT-B/16 [6] as the image encoder, the first six layers of the BERTbase [5] as the text encoder, and the last six layers of the BERTbase as the fusion encoder. The input image \\(I\\) is encoded into a set of global and local patch embeddings \\(\\{\\pmb{v}_{cls}, \\pmb{v}_1, \\dots, \\pmb{v}_N\\} \\in \\mathbb{R}^{(N + 1) \\times d_e}\\) by the image encoder, where \\(\\pmb{v}_{cls}\\) represents the [CLS] embedding of the input image. The input text \\(T\\) is encoded into a set of global and tokenized local patch embeddings \\(\\{\\pmb{t}_{cls}, \\pmb{t}_1, \\dots, \\pmb{t}_M\\} \\in \\mathbb{R}^{(M + 1) \\times d_e}\\) by the text encoder, where \\(\\pmb{t}_{cls}\\) represents the [CLS] embedding of the input text. The fusion encoder takes the image and text embeddings and performs cross-attention. We maintain two queues to store the representations (i.e., \\(\\hat{\\pmb{v}}_{cls}, \\hat{\\pmb{t}}_{cls}\\)) of momentum models for conducting large-scale contrastive learning. The architecture of the momentum model remains identical to the original baseline architecture." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.435, + 0.557, + 0.452 + ], + "angle": 0, + "content": "3.2 Modality-Agnostic Adapter (MAA)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.789, + 0.507 + ], + "angle": 0, + "content": "Our proposed MAA aims to integrate the global and local representations of images and text by learning the importance of local patch features for each modality through an attention network with shared parameters." + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.517, + 0.787, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {V} = \\left\\{\\boldsymbol {v} _ {1}, \\dots , \\boldsymbol {v} _ {N} \\right\\}, \\quad \\boldsymbol {T} = \\left\\{\\boldsymbol {t} _ {1}, \\dots , \\boldsymbol {t} _ {M} \\right\\} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.545, + 0.788, + 0.638 + ], + "angle": 0, + "content": "where \\(\\mathbf{V} \\in \\mathbb{R}^{N \\times d_e}\\) and \\(\\mathbf{T} \\in \\mathbb{R}^{M \\times d_e}\\) are the sets of local patch embeddings obtained from the image and text encoder outputs, respectively. We introduce modality-agnostic attention to allow the model to learn how to assign patch importance to each local patch embedding for each modality based on the context of the input image-text pair. The local representations \\(\\mathbf{v}_l \\in \\mathbb{R}^{d_e}\\) and \\(\\mathbf{t}_l \\in \\mathbb{R}^{d_e}\\) are computed as follows." + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.648, + 0.787, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {l} = \\sum_ {i = 1} ^ {N} \\alpha_ {i} \\boldsymbol {v} _ {i}, \\quad \\boldsymbol {t} _ {l} = \\sum_ {i = 1} ^ {M} \\beta_ {i} \\boldsymbol {t} _ {i} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.326, + 0.699, + 0.786, + 0.716 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {u} _ {i} ^ {v} = \\tanh (\\boldsymbol {W} _ {w} \\boldsymbol {v} _ {i} + \\boldsymbol {b} _ {w}), \\quad \\boldsymbol {u} _ {i} ^ {t} = \\tanh (\\boldsymbol {W} _ {w} \\boldsymbol {t} _ {i} + \\boldsymbol {b} _ {w}) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.732, + 0.786, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} = \\frac {\\exp \\left(\\boldsymbol {u} _ {i} ^ {v} \\boldsymbol {u} _ {w} ^ {\\top}\\right)}{\\sum_ {k = 1} ^ {N} \\exp \\left(\\boldsymbol {u} _ {k} ^ {v} \\boldsymbol {u} _ {w} ^ {\\top}\\right)}, \\quad \\beta_ {i} = \\frac {\\exp \\left(\\boldsymbol {u} _ {i} ^ {t} \\boldsymbol {u} _ {w} ^ {\\top}\\right)}{\\sum_ {k = 1} ^ {M} \\exp \\left(\\boldsymbol {u} _ {k} ^ {t} \\boldsymbol {u} _ {w} ^ {\\top}\\right)} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "where \\(\\alpha_{i}\\) and \\(\\beta_{i}\\) measure the importance score of the \\(i^{th}\\) image and text patch embeddings, respectively. Note that the learning parameters (i.e., \\(\\pmb{W}_{w},\\pmb{b}_{w}\\), and \\(\\pmb{u}_{w}\\)) are shared across modalities, and this approach encourages interactions between modalities, as demonstrated in how Chen et al. [4] uses shared learning" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.636, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Fashion-FINE" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "parameters as discrete tokens to ground the features of both modalities in a shared manifold space. We compute the integrated representation \\( \\mathbf{v}_{int} \\) and \\( \\mathbf{t}_{int} \\) through the weighted sum of the global and local representations as follows." + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.206, + 0.785, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {i n t} = \\lambda_ {I} \\times \\boldsymbol {v} _ {c l s} + (1 - \\lambda_ {I}) \\times \\boldsymbol {v} _ {l} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.224, + 0.785, + 0.24 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {t} _ {i n t} = \\lambda_ {T} \\times \\boldsymbol {t} _ {c l s} + (1 - \\lambda_ {T}) \\times \\boldsymbol {t} _ {l} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.251, + 0.788, + 0.357 + ], + "angle": 0, + "content": "We have observed the performance changes according to the variation of \\(\\lambda_I\\) and \\(\\lambda_T\\) and our model achieved the highest performance when \\(\\lambda_I = 0.95\\) and \\(\\lambda_T = 0.95\\). Using the integrated image and text representations, our model performs HNM-F (Section 3.3) and C-CmA (Section 3.4), and gradients are propagated not only to the [CLS] embedding, but also to the local patch embeddings. Thus, the use of integrated representations in HNM-F and C-CmA encourages the learning of local patch embeddings for fine-grained cross-modal alignment." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.378, + 0.654, + 0.394 + ], + "angle": 0, + "content": "3.3 Hard Negative Mining w/ Focal Loss (HNM-F)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.784, + 0.508 + ], + "angle": 0, + "content": "Our proposed HNM-F aims to improve the distinction learning between positive and hard negatives for a given anchor through the application of focal loss in ITC using integrated image and text representations. To account for the large scale of negatives during HNM-F, we maintain two large queues that store the most recently generated \\(Q\\) integrated representations obtained from the momentum model. We calculate the relative similarity of \\(i^{th}\\) positive image-text and text-image pairs in the mini-batch, respectively, as follows." + }, + { + "type": "equation", + "bbox": [ + 0.366, + 0.521, + 0.785, + 0.567 + ], + "angle": 0, + "content": "\\[\np _ {i 2 t} ^ {(i, +)} = \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {v} _ {i n t} ^ {(i)} , \\hat {\\boldsymbol {t}} _ {i n t} ^ {(+)}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {B + Q} \\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {v} _ {i n t} ^ {(i)} , \\hat {\\boldsymbol {t}} _ {i n t} ^ {(j)}\\right) / \\tau\\right)} \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.366, + 0.576, + 0.785, + 0.617 + ], + "angle": 0, + "content": "\\[\np _ {t 2 i} ^ {(i, +)} = \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {t} _ {\\text {i n t}} ^ {(i)} , \\boldsymbol {v} _ {\\text {i n t}} ^ {(+)}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {B + Q} \\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {t} _ {\\text {i n t}} ^ {(i)} , \\boldsymbol {v} _ {\\text {i n t}} ^ {(j)}\\right) / \\tau\\right)} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.627, + 0.784, + 0.764 + ], + "angle": 0, + "content": "where \\(\\pmb{v}_{int}\\) and \\(\\pmb{t}_{int}\\) are the integrated representations calculated in Equations 5 and 6, respectively, \\(B\\) denotes a mini-batch size, and \\(\\tau\\) denotes a temperature parameter [12]. \\(sim(\\pmb{v}_{int},\\hat{\\pmb{t}}_{int})\\) and \\(sim(\\pmb{t}_{int},\\hat{\\pmb{v}}_{int})\\) are calculated as \\(g_v(\\pmb{v}_{int})^\\top g_t'(\\hat{\\pmb{t}}_{int})\\) and \\(g_t(\\pmb{t}_{int})^\\top g_v'(\\hat{\\pmb{v}}_{int})\\) in operations within the batch and queues. \\(g_v\\) and \\(g_t\\) are the norm(Adp(\\cdot)) to obtain \\(\\pmb{v}_{int}' \\in \\mathbb{R}^d\\) and \\(\\pmb{t}_{int}' \\in \\mathbb{R}^d\\), which are the representations output from the original model. \\(g_v'\\) and \\(g_t'\\) are the norm(Adp'(\\cdot)) to obtain \\(\\hat{\\pmb{v}}_{int}' \\in \\mathbb{R}^d\\) and \\(\\hat{\\pmb{t}}_{int}' \\in \\mathbb{R}^d\\), which are the representations output from the momentum model. Adp(\\cdot) and Adp'(\\cdot) are the modality-specific adaptive layers that project to \\(d\\) dimensions and norm is the normalization function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.785, + 0.84 + ], + "angle": 0, + "content": "One of the primary characteristics of fashion data is the significant presence of hard negative pairs that share most of the fine-grained attributes with positive pairs for a given anchor. Thus, training ITC by setting the one-hot encoded ground truth similarity for positive and negative pairs to 1 and 0, respectively, can lead to overfitting on the training data. To address this challenge, we adopted" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "S. Jin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.194 + ], + "angle": 0, + "content": "soft labels by combining pseudo-targets, which are calculated from the momentum image and text encoders, with the one-hot encoded hard labels through a weighted sum. We define the soft labels \\( \\pmb{q}_{i2t} \\) and \\( \\pmb{q}_{t2i} \\) as follows." + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.203, + 0.786, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {q} _ {i 2 t} = (1 - \\eta) \\times \\boldsymbol {y} _ {i 2 t} + \\eta \\times \\hat {\\boldsymbol {y}} _ {i 2 t} \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.222, + 0.786, + 0.237 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {q} _ {t 2 i} = (1 - \\eta) \\times \\boldsymbol {y} _ {t 2 i} + \\eta \\times \\hat {\\boldsymbol {y}} _ {t 2 i} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.244, + 0.788, + 0.389 + ], + "angle": 0, + "content": "where \\(\\pmb{y}_{i2t} \\in \\mathbb{R}^{B \\times (B + Q)}\\) and \\(\\pmb{y}_{t2i} \\in \\mathbb{R}^{B \\times (B + Q)}\\) are both represent rectangular diagonal matrices with one-hot encoding, where the main diagonal elements are 1, and all other elements are 0. \\(\\hat{\\pmb{y}}_{i2t} \\in \\mathbb{R}^{B \\times (B + Q)}\\) and \\(\\hat{\\pmb{y}}_{t2i} \\in \\mathbb{R}^{B \\times (B + Q)}\\) are pseudo-targets calculated from the dot product values between the momentum image and text encoders. We experimentally found that the optimal \\(\\eta\\) is 0.4. As demonstrated in [20], minimizing Kullback - Leibler \\((\\pmb{p}_{i2t}, \\pmb{q}_{i2t})\\) is equivalent to minimizing \\(-\\sum_{i=1}^{B} q_{i2t}^{(i, +)} \\log p_{i2t}^{(i, +)}\\) and maximizing mutual information between \\(\\pmb{v}_{int}^{(i)}\\) and \\(\\pmb{t}_{int}^{(+)}\\) for text that shares similar semantic meaning with the image \\(I_i\\). Finally, we propose the HNM-F loss with the soft labels as follows." + }, + { + "type": "equation", + "bbox": [ + 0.334, + 0.397, + 0.785, + 0.46 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {h n m - f} = - \\frac {\\alpha}{2 B} \\sum_ {i = 1} ^ {B} [ (1 - p _ {i 2 t} ^ {(i, +)}) ^ {\\gamma} q _ {i 2 t} ^ {(i, +)} \\log p _ {i 2 t} ^ {(i, +)} \\\\ \\left. + \\left(1 - p _ {t 2 i} ^ {(i, +)}\\right) ^ {\\gamma} q _ {t 2 i} ^ {(i, +)} \\log p _ {t 2 i} ^ {(i, +)} \\right] \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.561 + ], + "angle": 0, + "content": "where \\(\\alpha \\in [0,1]\\) is a weighting factor to mitigate the learning issues of the class imbalance between positive and negative pair classes. \\((1 - p_{i2t}^{(i, + )})^{\\gamma}\\) and \\((1 - p_{t2i}^{(i, + )})^{\\gamma}\\) are modulating factors based on the relative similarity of positive pairs. \\(\\gamma \\geq 0\\) represents a tunable focusing parameter. A higher \\(\\gamma\\) value promotes increasing the similarity of positive pairs with low similarity and decreasing the relative similarity between positive and hard negatives in HNM-F." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.58, + 0.673, + 0.596 + ], + "angle": 0, + "content": "3.4 Comprehensive Cross-modal Alignment (C-CmA)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.603, + 0.788, + 0.709 + ], + "angle": 0, + "content": "C-CmA aims to enhance fine-grained semantic alignment between text and image representations by learning hierarchical text features of both low-level and high-level features that are difficult to capture with the integrated text representation (Equation 6) alone. First, to extract low-level features in the form of both words and phrases (e.g., \"placket,\" \"cinch belt,\" \"removable fur trim\"), we applied a 1-D convolution to local patch embeddings according to three kernel sizes (i.e., uni-gram, bi-gram, and tri-gram)." + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.718, + 0.786, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {t} _ {c _ {k}} = f _ {k} (\\boldsymbol {T}), \\quad k \\in \\{1, 2, 3 \\} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.743, + 0.788, + 0.789 + ], + "angle": 0, + "content": "where \\( f_{k} \\) denotes the 1-D convolution layers with kernel size \\( k \\). \\( t_{c_k} \\in \\mathbb{R}^{M \\times d_e} \\) is the embedding obtained by applying the 1-D convolution with kernel size \\( k \\). The derivation of the low-level representation \\( t_c \\) is as follows." + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.798, + 0.786, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {t} _ {c} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\operatorname {m a x p o o l} \\left(\\left[ \\boldsymbol {t} _ {c _ {1}}, \\boldsymbol {t} _ {c _ {2}}, \\boldsymbol {t} _ {c _ {3}} \\right]\\right), \\quad \\boldsymbol {t} _ {c} \\in \\mathbb {R} ^ {d _ {e}} \\tag {13}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.636, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Fashion-FINE" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.268 + ], + "angle": 0, + "content": "Next, we added a new fashion symbol token to the text input to obtain a high-level representation of fashion symbols (Figure 2). The text encoder of Fashion-FINE embeds the high-level representation, \\( \\pmb{t}_s \\in \\mathbb{R}^{d_e} \\), for the nine pre-defined fashion symbols (i.e., tops, dresses, skirts, coats, pants, shoes, bags, accessories, and others). These symbols are pre-defined in FashionSAP based on the categories in FashionGen [33]. The total set of uni-modal embeddings from the text encoder outputs, \\( T_{total} \\in \\mathbb{R}^{(M + 2) \\times d_e} \\), and the calculated multi-level text representation, \\( t_m \\in \\mathbb{R}^{d_e} \\), are as follows." + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.279, + 0.785, + 0.295 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {T} _ {\\text {t o t a l}} = \\left\\{\\boldsymbol {t} _ {\\text {c l s}}, \\boldsymbol {t} _ {s}, \\boldsymbol {t} _ {1}, \\dots , \\boldsymbol {t} _ {M} \\right\\} \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.298, + 0.785, + 0.314 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {t} _ {m} = \\delta \\times \\boldsymbol {t} _ {s} + (1 - \\delta) \\times \\boldsymbol {t} _ {c} \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.784, + 0.398 + ], + "angle": 0, + "content": "In C-CmA, note that we do not use the integrated text representation calculated in Equation 6, but instead use a text representation that integrates both low-level and high-level attributes. We have observed the performance changes according to the variation of \\(\\delta\\) and achieved the highest performance when \\(\\delta = 0.95\\). The loss function of C-CmA is as follows." + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.409, + 0.785, + 0.426 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {i n t} ^ {\\prime} = \\operatorname {n o r m} \\left(\\operatorname {A d p} \\left(\\boldsymbol {v} _ {i n t}\\right)\\right), \\quad \\boldsymbol {t} _ {m} ^ {\\prime} = \\operatorname {n o r m} \\left(\\operatorname {A d p} \\left(\\boldsymbol {t} _ {m}\\right)\\right) \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.43, + 0.785, + 0.47 + ], + "angle": 0, + "content": "\\[\nL _ {c - c m a} = \\frac {1}{B} \\left[ 1 - \\sum_ {b = 1} ^ {B} \\frac {1}{2} \\left[ v _ {i n t} ^ {\\prime b} \\left(\\boldsymbol {t} _ {m} ^ {\\prime b}\\right) ^ {\\top} + 1 \\right] \\right] \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.479, + 0.784, + 0.525 + ], + "angle": 0, + "content": "where \\(\\pmb{v}_{int}^{\\prime}\\) is a calculated adapted integrated image representation in Section 3.3 and \\(\\pmb{t}_m^\\prime\\) are the adapted multi-level text representation. \\(\\mathrm{Adp}(\\cdot)\\) and norm are the same as the used in Section 3.3, and \\(B\\) denotes the mini-batch size." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.545, + 0.467, + 0.561 + ], + "angle": 0, + "content": "3.5 Hard Negative Sampling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.785, + 0.766 + ], + "angle": 0, + "content": "For the interactive V+L retrieval tasks (e.g., text-guided image retrieval), it is important to learn to distinguish the representations of positive and hard negative pairs. For each image-text pair in the mini-batch, we set the image and text as respective anchors and sampled one hard negative for each anchor based on the relative similarity calculated in HNM-F. The sampled two hard negatives for each modality are used to perform ITM in the fusion encoder. Thus, improving hard negative mining through our Fashion-FINE can benefit not only the cross-modal retrieval but also the text-guided image retrieval tasks, which rely on the [CLS] embedding output of the fusion encoder. We have quantitatively and qualitatively verified the effectiveness in Sections 4.2 and 4.3. This section introduces the three fusion encoder-based pre-training tasks used in the previous fashion VLP research [10, 11]. We set the pre-training tasks introduced in this section as the baseline." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.786, + 0.498, + 0.802 + ], + "angle": 0, + "content": "3.6 Image-Text Matching (ITM)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In Fashion-FINE, ITM corresponds to a binary classification problem that takes the output [CLS] embeddings based on the cross-attention in the fusion encoder" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "S. Jin et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 1: Cross-modal retrieval results on FashionGen with the protocol used in previous fashion VLP research [7,10,11,25,45]." + }, + { + "type": "table", + "bbox": [ + 0.33, + 0.178, + 0.67, + 0.324 + ], + "angle": 0, + "content": "
MethodsITRTIRSumR
R@1R@5R@10R@1R@5R@10
VL-BERT [37]19.2639.9046.0522.6336.4848.52212.84
ViLBERT [24]20.9740.4948.2121.1237.2350.11239.18
Image-BERT [30]22.7641.8950.7724.7845.2055.90241.30
OscAR [21]23.3944.6752.5525.1049.1456.68251.53
FashionBERT [7]23.9646.3152.1226.7546.4855.74251.36
KaleidoBERT [45]27.9960.0968.3733.8860.6068.59319.52
CommerceMM [44]41.6064.0072.8039.6061.5072.70352.20
EI-CLIP [25]38.7072.2084.2540.0671.9982.90390.10
ALBEF [20]63.9788.9294.4160.5284.9991.45484.26
FashionVIL [10]65.5491.3496.3061.8887.3293.22495.60
FashionSAP [11]73.1492.8096.8770.1291.7696.38521.07
Fashion-FINE (Ours)74.8093.9497.0070.7091.9096.18524.54
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.34, + 0.657, + 0.355 + ], + "angle": 0, + "content": "to classify as positive or negative. The ITM loss is as follows." + }, + { + "type": "equation", + "bbox": [ + 0.362, + 0.366, + 0.786, + 0.386 + ], + "angle": 0, + "content": "\\[\nL _ {i t m} = \\mathbb {E} _ {(\\boldsymbol {I}, \\boldsymbol {T}) \\sim D} \\operatorname {C E} \\left(\\boldsymbol {y} ^ {i t m}, \\boldsymbol {p} ^ {i t m} \\left(\\boldsymbol {H} _ {i t m}\\right)\\right) \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.396, + 0.787, + 0.456 + ], + "angle": 0, + "content": "where \\(D\\) denotes the entire dataset comprised of image-text pairs, and CE denotes the cross-entropy loss. \\(\\pmb{H}_{itm} = [\\pmb{H}_v\\oplus \\pmb{H}_t]\\), and \\(\\pmb{H}_{v}\\) and \\(\\pmb{H}_{t}\\) denote the uni-modal image and text representations, respectively. \\(\\pmb{p}^{itm}(\\pmb{H}_{itm})\\) is the predictive probability distribution for the ITM task, and \\(\\pmb{y}^{itm}\\) is the ground truth." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.494, + 0.449, + 0.51 + ], + "angle": 0, + "content": "3.7 Pre-training Objective" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.52, + 0.787, + 0.596 + ], + "angle": 0, + "content": "For learning prompt token prediction \\((L_{ptp})\\) and token replacement prediction \\((L_{trp})\\) in the fusion encoder, we employ the same prompt input strategies as FashionSAP [11] (see supplementary material for more details). The total loss for Fashion-FINE is computed as follows. Fashion-FINE is optimized end-to-end for the pre-training datasets by minimizing \\(L\\)." + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.607, + 0.786, + 0.626 + ], + "angle": 0, + "content": "\\[\nL = L _ {h n m - f} + L _ {c - c m a} + L _ {i t m} + L _ {p t p} + L _ {t r p} \\tag {19}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.647, + 0.377, + 0.664 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.679, + 0.447, + 0.695 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Dataset Following previous research [7, 10, 11, 25, 40, 45], we used two representative public benchmarks for pre-training and fashion \\(\\mathrm{V + L}\\) retrieval tasks: FashionGen [33] and FashionIQ [41]. FashionGen contains 260K image-text pairs for the training dataset and 35K for test dataset. FashionIQ contains 18K training triplets (i.e., reference image, modifying text, and target image) and 6K validation triplets across three categories: dress, shirt, and toptee. For fair experiments, we pre-trained our model using FashionGen, consistent with previous studies. For the downstream tasks, we fine-tuned our model on FashionGen for cross-modal retrieval and on FashionIQ for text-guided image retrieval." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.636, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Fashion-FINE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.145, + 0.781, + 0.159 + ], + "angle": 0, + "content": "Table 3: Text-guided image retrieval re" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.159, + 0.631, + 0.173 + ], + "angle": 0, + "content": "Table 2: Cross-modal retrieval results on sutls on FashionIQ." + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.173, + 0.436, + 0.186 + ], + "angle": 0, + "content": "FashionGen with full evaluation." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.19, + 0.497, + 0.27 + ], + "angle": 0, + "content": "
MethodsITRTIRSumR
R@1R@5R@10R@1R@5R@10
EI-CLIP [25]25.7054.5066.8028.4057.1069.40301.90
ALBEF [20]41.6867.3975.5050.9575.3684.15395.03
FashionKLIP [40]47.1669.2775.3954.6075.0681.39402.87
FashionViL [10]42.8871.5780.5551.3475.4284.57406.33
FashionSAP [11]54.4377.3083.1562.8283.9690.16451.82
Fashion-FINE (Ours)54.2977.7483.8765.2385.6191.37458.11
" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.177, + 0.779, + 0.282 + ], + "angle": 0, + "content": "
MethodsDressToptieShirtSumR
R@10R@50R@10R@50R@10R@50
FiLM [28]14.2333.3417.3037.6815.0434.09151.68
CIRPLANT [23]17.4540.4121.6445.3817.5338.81181.22
VAL [3]22.5344.0027.5351.6822.3844.15212.27
TIRG(bert) [38]27.1753.2527.8457.1122.2845.58233.23
CosMo [17]25.6450.3029.2157.4624.9049.18236.69
DCNet [15]28.9556.7030.4458.2923.9547.30245.63
FashionVIL [10]33.4759.9434.9860.7925.1750.39264.74
FashionVLP [8]32.4260.2938.5168.7931.8958.44290.34
FashionSAP [11]33.7160.4341.9170.9333.1761.33301.48
Fashion-FINE (Ours)36.6463.4646.8174.0938.7663.30323.06
" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.284, + 0.499, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.447, + 0.374, + 0.459 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.284, + 0.779, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.446, + 0.651, + 0.458 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.469, + 0.788, + 0.541 + ], + "angle": 0, + "content": "Fig. 3: Qualitative analysis results. (a) TIR: the green text in the query indicates the distinguished attributes of the primary ground truth. (b) TGIR: the green text in the modifying text indicates attributes that need to be modified in the reference image and included in the target image. Each row shows the top-10 retrieved results for each model (far left: top-1, far right: top-10), and the green box represents the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.645 + ], + "angle": 0, + "content": "Implementation Details We trained the proposed modules with randomly initialized weights, and the baseline encoders were initialized with the same pre-trained weights as in FashionSAP [11]. Through hyperparameter search experiments, we found that the optimal values for HNM-F are \\(\\tau = 0.07\\), \\(\\alpha = 0.25\\), and \\(\\gamma = 2.0\\) (see supplementary material for more details)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.665, + 0.447, + 0.68 + ], + "angle": 0, + "content": "4.2 Cross-modal Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Quantitative Analysis Following the previous study [11], we evaluated cross-modal retrieval performance based on the calculated relative similarity between the integrated uni-modal representations. Thus, we fine-tuned our model using MAA, HNM-F, and C-CmA. ITR aims to retrieve matched text for an input query image, while TIR aims to retrieve a matched image for an input query text. We evaluated our Fashion-FINE using the following two evaluation methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "First, following previous research [7, 10, 11, 25, 45], we evaluated the average performance result of five randomly sampled retrieval test sets containing 1,000 unique image and text queries. For each query in the test sets, only one candidate is matched (positive), while the other 100 candidates are mismatched (negative)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "S. Jin et al." + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.144, + 0.787, + 0.27 + ], + "angle": 0, + "content": "
Attributesturtleneckbomberjacketclosurestraightdenimdistresslongsleevearrow
Baseline(Fashion-FINEw/o maa, hmm-f,c-cma)
Fashion-FINE
" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.279, + 0.788, + 0.309 + ], + "angle": 0, + "content": "Fig. 4: Qualitative analysis of the fine-grained alignment in the pre-training phase. The visualized results are the 1st layer of the cross-attention map in the fusion encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.347, + 0.788, + 0.44 + ], + "angle": 0, + "content": "and randomly sampled from other products within the same subcategory. As a result, Fashion-FINE achieved state-of-the-art performance in ITR and TIR based on the SumR metric, as shown in Table 1. Second, we used the full test dataset to thoroughly evaluate the performance and reproducibility of Fashion-FINE, consistent with previous research [10,11,25,40]. As a result, Fashion-FINE also achieved state-of-the-art performance in ITR and TIR, as shown in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.789, + 0.658 + ], + "angle": 0, + "content": "Qualitative Analysis We analyzed the cases where the ground truth retrieval rank increased in the TIR results through the ablation study (Figure 3a). The baseline model retrieved hard negatives that share the fine-grained attributes (e.g., 'plaid jacket,' 'spread collar,' and 'single-button barrel cuffs') with the ground truth at a higher rank than the ground truth. Fashion-FINE (w/o C-CmA) ranked the ground truth in 4th place, yet it ranked hard negatives-sharing the same attributes as the ground truth but differing in location or having slight color variations-ahead of the ground truth. On the other hand, Fashion-FINE ranked all ground truths between 1st and 4th place. These results represent superior qualitative achievements not captured by the \\( R@k \\) metric presented in Table 2, which considers a retrieval correct if the retrieved top-k list includes at least one correct answer. Through the analyses, we have qualitatively demonstrated the effectiveness of Fashion-FINE for fine-grained cross-modal retrieval." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.694, + 0.504, + 0.71 + ], + "angle": 0, + "content": "4.3 Text-Guided Image Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Quantitative Analysis We fine-tuned our model using contrastive learning based on the similarity between the output [CLS] embedding of the fusion encoder for the input query image-text pair and the output [CLS] embedding of the image encoder in Fashion-FINE for the target image. This task aims to retrieve a target image by referencing an image-text pair as a query, where the text modifies certain attributes of the image. We compared our model with fashion VLP methods under the original protocol used by FashionIQ. As a result, Fashion-FINE achieved state-of-the-art performance in TGIR, as shown in Table 3." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.636, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Fashion-FINE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 4: Ablation study results for our proposed modules on FashionGen and FashionIQ. We used the baseline architecture (Section 3.1) in this ablation study." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.178, + 0.784, + 0.263 + ], + "angle": 0, + "content": "
ArchitectureProposed ModulesCross-modal RetrievalText-Guided Image Retrieval
MAA HNM-F C-CmAITRTIRSumRDressTopteeShirt
R@1R@5R@10R@1R@5R@10R@10R@50R@10R@50
BaselineXXX54.4377.3083.1562.8283.9690.16451.8233.7160.4333.17
Baseline✓XX54.3377.4283.7564.6284.0790.18454.3732.9760.7835.43
(w/ Ours)✓✓X54.3077.8184.0064.4085.4290.74456.6736.2462.3238.03
✓✓✓54.2977.7483.8765.2385.6191.37458.1136.6463.4638.76
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.265, + 0.784, + 0.293 + ], + "angle": 0, + "content": "Table 5: Experimental results on the applicability of Fashion-FINE to the foundation model (i.e., BLIP [19]) on FashionGen." + }, + { + "type": "table", + "bbox": [ + 0.317, + 0.298, + 0.681, + 0.384 + ], + "angle": 0, + "content": "
ArchitectureProposed ModulesCross-modal Retrieval
MAAHNM-FC-CmAITRTIR
R@1R@5R@10R@1R@5
BLIPXXX55.2778.3985.0661.6382.46
BLIP (w/ Ours)✓XX54.6078.6085.1265.0585.65
✓✓X57.2480.2086.5363.3586.41
✓✓✓59.8182.4888.1268.4987.98
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.401, + 0.788, + 0.568 + ], + "angle": 0, + "content": "Qualitative Analysis We qualitatively analyzed the cases where the rank of the retrieved ground truth increased in the TGIR task according to the ablation study (Figure 3b). The baseline model failed to retrieve the ground truth in the top-10, highly ranking hard negatives that contain \"no sleeves\" or \"two different colors.\" On the other hand, we observed that as more modules of Fashion-FINE were incorporated, the rank of the ground truth increased. This also qualitatively demonstrates the effectiveness of Fashion-FINE in learning fine-grained details, which is not captured in Table 3. Through the analyses, we have demonstrated that Fashion-FINE effectively learned not only the representation of the target image, but also the fine-grained alignment in the fusion encoder to obtain representations of the input reference image-modifying text pairs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.589, + 0.533, + 0.604 + ], + "angle": 0, + "content": "4.4 Fine-grained Alignment Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.734 + ], + "angle": 0, + "content": "To verify the fine-grained alignment in the fusion encoder, following [11], we chose the four instances from FashionGen and visualized cross-attention maps using Grad-CAM [35]. In Figure 4, Fashion-FINE more accurately highlights the regions of fine-grained attributes (e.g., \"arrow,\" \"distress\") while minimizing the highlight on the background compared to the baseline. We also found that Fashion-FINE effectively highlights the region of attributes for both single words (e.g. \"closure,\" \"arrow,\" \"straight,\" \"denim,\" \"distress\") and phrases (e.g., \"turtle neck,\" \"bomber jacket,\" \"long sleeve\") in the form of fine-grained attributes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.756, + 0.388, + 0.771 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We conducted an ablation study to evaluate the effectiveness of the proposed modules. For fair experiments, the experimental settings are consistent in the same ablation series. Compared to the baseline, our methods achieved significant performance improvements in the \\(\\mathrm{V + L}\\) retrieval tasks (i.e., TIR and TGIR)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "S. Jin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "These results indicate that our methods were effective in learning to distinguish the positives from the hard negatives among the image candidates for given input query text or query image-text pairs in fine-grained V+L retrieval tasks. This is supported by the significant performance improvement results in TIR and TGIR tasks, where the disentangled embedding of images is important. The notable performance improvement in TGIR resulting from the inclusion of the HNM-F module indicates that our approach effectively learned not only the disentangled embedding of the input image but also the representation of the input image-text pair (i.e., reference image and modifying text) in the fusion encoder." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.301, + 0.723, + 0.317 + ], + "angle": 0, + "content": "4.6 Applicability of Fashion-FINE to the Foundation Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.789, + 0.535 + ], + "angle": 0, + "content": "We conducted experiments to verify the applicability of our Fashion-FINE to foundational models. For the experiments, we performed an ablation study of our modules based on the dual-stream+fusion foundational model architecture, BLIP [19]. Since BLIP uses bootstrapping on noisy data during the pre-training phase with a decoder, we applied our modules in a fine-tuning phase only, which was designed without a decoder, to the pre-trained BLIP model to evaluate the pure performance of our modules. The encoder configuration for the ITC and ITM tasks in the BLIP architecture is the same as that of Fashion-FINE, so we applied Fashion-FINE to BLIP in the same way as the baseline architecture described in Section 3. In the fine-tuning phase for the \\(\\mathrm{V + L}\\) tasks (Sections 4.2 and 4.3), since we applied the modules of Fashion-FINE only for ITR and TIR and not for TGIR, the application of Fashion-FINE to BLIP was also done only for ITR and TIR. In Table 5, we have demonstrated that the modules of Fashion-FINE can be successfully applied to the BLIP architecture, a foundational model." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.554, + 0.36, + 0.57 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.789, + 0.764 + ], + "angle": 0, + "content": "In this paper, we presented Fashion-FINE, which aims to improve the learning of fine-grained cross-modal alignment. Fashion-FINE addresses the challenge of distinguishing positives from hard negatives based on MAA, which integrates global features with the fine-grained features of local representations, and on HNM-F and C-CmA, which enhance the learning effects of hard negative mining and refine the integrated image representations. Unlike previous fashion VLP research, Fashion-FINE focuses on cross-modal alignment before multimodal fusion, taking into account fine-grained details. Through quantitative and qualitative analyses, we have validated that Fashion-FINE can be successfully integrated with fusion encoder-based pre-training tasks. With this work, we advance the fashion VLP architecture and expect that our approach will contribute to the research on the development of fine-grained foundation methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.783, + 0.405, + 0.8 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "This work was supported by the National Research Foundation of Korea (NRF) grant (No.2021M3A9E4080780, No.2018R1A5A7059549) and by the Institute of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.636, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Fashion-FINE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean government (MSIT) (No.RS-2020-II201373, Artificial Intelligence Graduate School Program (Hanyang University))." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.213, + 0.323, + 0.228 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.24, + 0.785, + 0.281 + ], + "angle": 0, + "content": "1. Byun, J., Hwang, T., Fu, J., Moon, T.: Grit-vlp: Grouped mini-batch sampling for efficient vision and language pre-training. In: European Conference on Computer Vision. pp. 395-412. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.785, + 0.335 + ], + "angle": 0, + "content": "2. Chen, Q., Huang, J., Feris, R., Brown, L.M., Dong, J., Yan, S.: Deep domain adaptation for describing people based on fine-grained clothing attributes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5315-5324 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.337, + 0.785, + 0.376 + ], + "angle": 0, + "content": "3. Chen, Y., Gong, S., Bazzani, L.: Image search with text feedback by visiolinguistic attention learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3001-3011 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.431 + ], + "angle": 0, + "content": "4. Chen, Y., Yuan, J., Tian, Y., Geng, S., Li, X., Zhou, D., Metaxas, D.N., Yang, H.: Revisiting multimodal representation in contrastive learning: from patch and token embeddings to finite discrete tokens. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15095-15104 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.785, + 0.471 + ], + "angle": 0, + "content": "5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.785, + 0.526 + ], + "angle": 0, + "content": "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.785, + 0.58 + ], + "angle": 0, + "content": "7. Gao, D., Jin, L., Chen, B., Qiu, M., Li, P., Wei, Y., Hu, Y., Wang, H.: Fashion-bert: Text and image matching with adaptive loss for cross-modal retrieval. In: Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 2251-2260 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.785, + 0.635 + ], + "angle": 0, + "content": "8. Goenka, S., Zheng, Z., Jaiswal, A., Chada, R., Wu, Y., Hedau, V., Natarajan, P.: Fashionvlp: Vision language transformer for fashion retrieval with feedback. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14105-14115 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.637, + 0.785, + 0.676 + ], + "angle": 0, + "content": "9. Gunel, B., Du, J., Conneau, A., Stoyanov, V.: Supervised contrastive learning for pre-trained language model fine-tuning. International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.678, + 0.785, + 0.717 + ], + "angle": 0, + "content": "10. Han, X., Yu, L., Zhu, X., Zhang, L., Song, Y.Z., Xiang, T.: Fashionvil: Fashion-focused vision-and-language representation learning. In: European Conference on Computer Vision. pp. 634-651. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.719, + 0.785, + 0.771 + ], + "angle": 0, + "content": "11. Han, Y., Zhang, L., Chen, Q., Chen, Z., Li, Z., Yang, J., Cao, Z.: Fashionsap: Symbols and attributes prompt for fine-grained fashion vision-language pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15028-15038 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.773, + 0.785, + 0.799 + ], + "angle": 0, + "content": "12. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.801, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Jang, J., Kong, C., Jeon, D., Kim, S., Kwak, N.: Unifying vision-language representation space with single-tower transformer. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 980-988 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.24, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "S. Jin et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "14. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "15. Kim, J., Yu, Y., Kim, H., Kim, G.: Dual compositional learning in interactive image retrieval. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1771-1779 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.245, + 0.787, + 0.3 + ], + "angle": 0, + "content": "16. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "17. Lee, S., Kim, D., Han, B.: Cosmo: Content-style modulation for image retrieval with text feedback. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 802-812 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.342, + 0.787, + 0.37 + ], + "angle": 0, + "content": "18. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: bootstrapping language-image pretraining with frozen image encoders and large language models. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.371, + 0.787, + 0.412 + ], + "angle": 0, + "content": "19. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.412, + 0.787, + 0.452 + ], + "angle": 0, + "content": "20. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.453, + 0.787, + 0.508 + ], + "angle": 0, + "content": "21. Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., et al.: Oscar: Object-semantics aligned pre-training for vision-language tasks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16. pp. 121-137. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.508, + 0.787, + 0.564 + ], + "angle": 0, + "content": "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.564, + 0.787, + 0.605 + ], + "angle": 0, + "content": "23. Liu, Z., Rodriguez-Opazo, C., Teney, D., Gould, S.: Image retrieval on real-life images with pre-trained vision-and-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2125-2134 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.605, + 0.787, + 0.646 + ], + "angle": 0, + "content": "24. Lu, J., Batra, D., Parikh, D., Lee, S.: Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.646, + 0.787, + 0.702 + ], + "angle": 0, + "content": "25. Ma, H., Zhao, H., Lin, Z., Kale, A., Wang, Z., Yu, T., Gu, J., Choudhary, S., Xie, X.: Ei-clip: Entity-aware interventional contrastive learning for e-commerce cross-modal retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18051-18061 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.702, + 0.787, + 0.73 + ], + "angle": 0, + "content": "26. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.73, + 0.787, + 0.771 + ], + "angle": 0, + "content": "27. Ordonez, V., Kulkarni, G., Berg, T.: Im2text: Describing images using 1 million captioned photographs. Advances in neural information processing systems 24 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.771, + 0.787, + 0.813 + ], + "angle": 0, + "content": "28. Perez, E., Strub, F., De Vries, H., Dumoulin, V., Courville, A.: Film: Visual reasoning with a general conditioning layer. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "29. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer" + }, + { + "type": "list", + "bbox": [ + 0.216, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.636, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Fashion-FINE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.148, + 0.786, + 0.175 + ], + "angle": 0, + "content": "image-to-sentence models. In: Proceedings of the IEEE international conference on computer vision. pp. 2641-2649 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.787, + 0.217 + ], + "angle": 0, + "content": "30. Qi, D., Su, L., Song, J., Cui, E., Bharti, T., Sacheti, A.: Imagebert: Cross-modal pre-training with large-scale weak-supervised image-text data. arXiv preprint arXiv:2001.07966 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.787, + 0.272 + ], + "angle": 0, + "content": "31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "32. Robinson, J., Chuang, C.Y., Sra, S., Jegelka, S.: Contrastive learning with hard negative samples. International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "33. Rostamzadeh, N., Hosseini, S., Boquet, T., Stokowiec, W., Zhang, Y., Jauvin, C., Pal, C.: Fashion-gen: The generative fashion dataset and challenge. arXiv preprint arXiv:1806.08317 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.383 + ], + "angle": 0, + "content": "34. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.438 + ], + "angle": 0, + "content": "35. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.494 + ], + "angle": 0, + "content": "36. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2556-2565 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.522 + ], + "angle": 0, + "content": "37. Su, W., Zhu, X., Cao, Y., Li, B., Lu, L., Wei, F., Dai, J.: Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.787, + 0.577 + ], + "angle": 0, + "content": "38. Vo, N., Jiang, L., Sun, C., Murphy, K., Li, L.J., Fei-Fei, L., Hays, J.: Composing text and image for image retrieval—an empirical odyssey. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6439–6448 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.619 + ], + "angle": 0, + "content": "39. Wang, J., Zhou, P., Shou, M.Z., Yan, S.: Position-guided text prompt for vision-language pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23242-23251 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.787, + 0.687 + ], + "angle": 0, + "content": "40. Wang, X., Wang, C., Li, L., Li, Z., Chen, B., Jin, L., Huang, J., Xiao, Y., Gao, M.: Fashionclip: Enhancing e-commerce image-text retrieval with fashion multimodal conceptual knowledge graph. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 5: Industry Track). pp. 149-158 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.787, + 0.743 + ], + "angle": 0, + "content": "41. Wu, H., Gao, Y., Guo, X., Al-Halah, Z., Rennie, S., Grauman, K., Feris, R.: Fashion iq: A new dataset towards retrieving images by natural language feedback. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 11307-11317 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.785 + ], + "angle": 0, + "content": "42. Wu, M., Mosse, M., Zhuang, C., Yamins, D., Goodman, N.: Conditional negative sampling for contrastive learning of visual representations. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "43. Yang, J., Duan, J., Tran, S., Xu, Y., Chanda, S., Chen, L., Zeng, B., Chilimbi, T., Huang, J.: Vision-language pre-training with triple contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15671-15680 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "S. Jin et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "44. Yu, L., Chen, J., Sinha, A., Wang, M., Chen, Y., Berg, T.L., Zhang, N.: Commercemm: Large-scale commerce multimodal representation learning with omni retrieval. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 4433-4442 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.26 + ], + "angle": 0, + "content": "45. Zhuge, M., Gao, D., Fan, D.P., Jin, L., Chen, B., Zhou, H., Qiu, M., Shao, L.: Kaleido-bert: Vision-language pre-training on fashion domain. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12647-12657 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.26 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Integration of Global and Local Representations for Fine-grained Cross-modal Alignment/35329370-3365-4c03-a82d-856d1e6d798d_origin.pdf b/2024/Integration of Global and Local Representations for Fine-grained Cross-modal Alignment/35329370-3365-4c03-a82d-856d1e6d798d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..54c2b3f8fae15857d2139d0b22fd1cab087e89ee --- /dev/null +++ b/2024/Integration of Global and Local Representations for Fine-grained Cross-modal Alignment/35329370-3365-4c03-a82d-856d1e6d798d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d9216c6f8dfdac3572bd3953025de362ff310f72a0df7aa1cf11a08a29b6a6d +size 2730007 diff --git a/2024/Integration of Global and Local Representations for Fine-grained Cross-modal Alignment/full.md b/2024/Integration of Global and Local Representations for Fine-grained Cross-modal Alignment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a3a02501027070e101ffb907adaca0cd33a18d1f --- /dev/null +++ b/2024/Integration of Global and Local Representations for Fine-grained Cross-modal Alignment/full.md @@ -0,0 +1,350 @@ +# Integration of Global and Local Representations for Fine-grained Cross-modal Alignment + +Seungwan Jin $^{\text{®}}$ , Hoyoung Choi $^{\text{®}}$ , Taehyung Noh $^{\text{®}}$ , and Kyungsik Han $^{\star}$ + +Hanyang University, South Korea {seungwanjin, choiho00, yestaehyung, kyungsikhan}@hanyang.ac.kr + +Abstract. Fashion is one of the representative domains of fine-grained Vision-Language Pre-training (VLP) involving a large number of images and text. Previous fashion VLP research has proposed various pretraining tasks to account for fine-grained details in multimodal fusion. However, fashion VLP research has not yet addressed the need to focus on (1) uni-modal embeddings that reflect fine-grained features and (2) hard negative samples to improve the performance of fine-grained $\mathrm{V + L}$ retrieval tasks. In this paper, we propose Fashion-FINE (Fashion VLP with Fine-grained Cross-modal Alignment using the INtegrated representations of global and local patch Embeddings), which consists of three key modules. First, a modality-agnostic adapter (MAA) learns uni-modal integrated representations and reflects fine-grained details contained in local patches. Second, hard negative mining with focal loss (HNM-F) performs cross-modal alignment using the integrated representations, focusing on hard negatives to boost the learning of fine-grained cross-modal alignment. Third, comprehensive cross-modal alignment (C-CmA) extracts low- and high-level fashion information from the text and learns the semantic alignment to encourage disentangled embedding of the integrated image representations. Fashion-FINE achieved state-of-the-art performance on two representative public benchmarks (i.e., FashionGen and FashionIQ) in three representative $\mathrm{V + L}$ retrieval tasks, demonstrating its effectiveness in learning fine-grained features. + +Keywords: Vision-Language Pre-training $\cdot$ Fine-grained Representation Learning $\cdot$ Fashion + +# 1 Introduction + +With recent advances in general VLP research that aims to learn common sense knowledge, there has been increasing attention to fine-grained cross-modal representation learning. Fashion is one of the most representative VLP research domains that applies the concept of cross-modal representations based on large-scale image and text information, about fashion items for fine-grained $\mathrm{V + L}$ retrieval tasks such as cross-modal retrieval and text-guided image retrieval. + +![](images/35c804a7d8757f0a0dbc7b0b951d6267a454a85987cc436f555f1daed0677069.jpg) +(a) Image-Text Pairs in Fashion Domain + +Long sleeve down parka-style coat in olive green. Genuine raccon fur trim at hood. Two-way zip closure at front concealed by button placket. Drawstring at collar. Welt pockets and flap pockets at front. Logo patch at upper sleeve. Pleated at back panel. Fully lined. Tonal stitching. Elasticized sleeve cuffs. + +![](images/efcdb87db5dbb4f15cdd7c8095403e784913d8d1aaac7dbe37e7a1828c86590a.jpg) + +Long sleeve down parka-style coat in olive green. Removable genuine raccoon fur trim at hood. Two-way zip closure at front concealed by press-stud packet. Welt pockets and flap pockets at front. Logo patch at upper sleeve. Drawstring at waist. Fully lined. Tonal stitching. Inset ribbed knit sleeve cuff. + +![](images/242c51ded600a09d98c7c5d4a4be58b03926717afd1c560269b33747bc77b06f.jpg) +Fig. 1: Our motivation for this research. (a) shows an example of three image-text pair cases (i.e., positive, hard negative, and easy negative pairs) in fashion products, where the colored areas specifically indicate fashion attributes that distinguish positives from hard negatives. (b) shows that fashion $\mathrm{V + L}$ tasks (e.g., cross-modal retrieval) are essential for distinguishing positives from hard negatives given anchors by capturing the fine-grained features. Our motivation is that it is difficult to distinguish between positives and hard negatives using only global representations, but this can be compensated by integrating local representations that reflect fine-grained attributes. + +Long sleeve down parka in black. Removable hood with cinch-belt detail at crown and wired removable fur trim. Tonal bungee drawstrings at hood. Stand collar. Two-way zip closure at front concealed by press-stud placket. Tonal logo applique at exterior arm. Welt and flap pockets at front with tonal signature patch... + +![](images/e52ebc137c4d3d3587b667d757ac8748b62cc05fa004e9f943fd3e6d6fe8b6da.jpg) +(b) Our Motivation for Fine-grained Cross-modal Alignment + +Most fashion VLP research [7, 8, 10, 11, 25, 45] has been evolved based on the general VLP architectures by applying self-supervised pre-training tasks, such as image-text contrastive learning (ITC) and image-text matching (ITM). Recently, employing the dual-stream+fusion [13, 39] structure, fashion VLP research [7, 10, 11, 45] proposed various pre-training tasks (e.g., masked patch modeling) to address the fine-grained features using the fusion encoder. FashionSAP [11], the current state-of-the-art model, proposed high-level multimodal semantic alignment and used the relative similarity between the global representations (i.e., [CLS] embedding) obtained before fusion to perform $\mathrm{V} + \mathrm{L}$ retrieval tasks instead of the ITM score calculated in the fusion encoder. However, using only global representations has limitations in learning fine-grained features in the cross-modal alignment and poses challenges in distinguishing between positive and hard negatives for a given anchor during the test phase (Figure 1). + +To address this, conventional methods have focused on learning to embed positive image-text pairs closer and negative pairs further using loss functions such as InfoNCE [26]. However, in self-supervised VLP, since only the pre-matched image-text pairs are considered positive and all others are considered negative, a severe class imbalance problem may arise due to the difference between the number of positive and negative pairs. Moreover, the overwhelming number of easy negative pairs compared to hard negative pairs leads to a dominance of the loss scale for easy negatives over hard negatives, thereby limiting the learning of the model to distinguish between positive and hard negative pairs. + +In this paper, we propose Fashion-FINE (Fashion VLP with Fine-grained cross-modal alignment using the INtegrated representations of global and local patch Embeddings), which learns cross-modal alignment in three phases. + +The first phase is to learn the importance of each local patch embedding for images and text using a modality-agnostic adapter (MAA). For each modality, we obtain a local representation that reflects the feature importance calculated by MAA at each patch location, and this local representation is then combined with the global representation, specifically the [CLS] embedding, to finally form the integrated representation. The second phase is the hard negative mining with focal loss (HNM-F), which integrates ITC and focal loss to mitigate the class and loss imbalance issues in ITC and to focus on learning to distinguish between positive and hard negative pairs. The third phase is the comprehensive cross-modal alignment (C-CmA), which semantically aligns the low-level and high-level fashion information in text descriptions with the integrated image representations. Through this, Fashion-FINE learns to encourage disentangled embedding of the image representations in the multimodal space, taking into account the comprehensive fashion text information. + +In summary, our contributions are as follows: + +- We propose Fashion-FINE, which performs fine-grained cross-modal alignment by integrating global and local representations of the two modalities. Fashion-FINE has achieved state-of-the-art performance in all fashion V+L retrieval tasks. +- We have improved the architecture for fashion VLP research by integrating Fashion-FINE with existing fusion encoder-based pre-training tasks, and we have verified its applicability through performance improvements. +- Our study is the first to quantitatively and qualitatively validate the importance of reflecting local patch embeddings in fine-grained cross-modal alignment and the boosting effect of hard negative mining with focal loss. + +# 2 Related Work + +# 2.1 Vision-Language Pre-training (VLP) + +Many VLP methods have achieved significant performance in various downstream $\mathrm{V + L}$ tasks. The most successful VLP architectures fall into two categories: dual-stream and dual-stream $^+$ fusion [13, 39]. The dual-stream architecture uses two modality-specific encoders to perform pre-training tasks (e.g., ITC) for semantic alignment between the two modalities. For example, CLIP [31] and ALIGN [14] have used ITC to train image-text alignment on large-scale multimodal data, achieving remarkable performance in general downstream $\mathrm{V + L}$ tasks. While this dual-stream is advantageous for achieving high performance, it requires extremely large-scale data during pre-training. On the other hand, the dual-stream $^+$ fusion architecture places multimodal fusion encoders on top of two modality-specific encoders to perform ITC followed by cross-modal learning (e.g., + +MLM and ITM). ALBEF [20] and TCL [43], based on the dual-stream+fusion, have achieved superior performance compared to dual-stream models. + +Existing VLP methods have been trained on noisy and large multimodal Web data due to the limited amount of high-quality human-annotated image-text pair data [14]. Recently, BLIP [19] and BLIP-2 [18] have achieved superior performance compared to ALBEF and TCL by expanding the dual-stream+fusion structure by proposing a decoder to bootstrap noisy text captions and Q-Former to learn soft visual prompt for bootstrapping a decoder-based large language model (LLM), respectively. Rather than focusing on human-annotated image-text pairs, BLIP-based methods tackle the inherent challenges in processing the noisy and large-scale multimodal datasets collected from the Web. + +The public benchmark datasets (e.g., FashionGen [33], FashionIQ [41]) addressed in recent fashion VLP research [10, 11, 25, 40] consist of multimodal datasets annotated by humans who manually describe fine-grained fashion attributes and contain no (or less) noise, unlike Web data. Thus, we set a pure dual-stream+fusion architecture without decoders as the baseline of FashionFINE, which is consistent with recent fashion VLP. By integrating our proposed modules with the pre-training tasks proposed in the prior work, we advance the fashion VLP architecture for the fine-grained $\mathrm{V + L}$ retrieval tasks. + +# 2.2 Hard Negative Mining in VLP + +Hard negative mining is an important technique used in machine learning, especially in fields such as computer vision and natural language processing, to improve model performance [9,32,34,42]. It involves focused sampling and training on the sampled hard negatives, which are semantically similar to positive data but are negative. In VLP, hard negative mining has been recently proposed to learn to distinguish between positive and hard negative samples in the image-text matching (ITM) task by extracting hard negative samples based on the calculated relative similarity during the cross-modal alignment phase [20, 43]. To enhance the effectiveness of the hard negative sampling, GRIT-VLP [1] proposed a grouped mini-batch sampling strategy to improve in-batch hard negative sampling in dual-stream $^+$ fusion architecture. However, the previous studies still rely solely on the use of global representations, which are limited in reflecting fine-grained details when calculating relative similarity for negative sampling. + +To the best of our knowledge, no research has been aimed at learning fine-grained local features to distinguish positive from hard negative in the cross-modal alignment to improve fine-grained $\mathrm{V + L}$ tasks. + +# 2.3 VLP in Fashion $\mathbf{V} + \mathbf{L}$ tasks + +Compared to general VLP that uses coarse-grained common sense knowledge [16, 22, 27, 29, 36], fashion VLP should focus more on fine-grained attributes in the multimodal fashion data [2]. FashionBERT [7] is the first cross-modal retrieval method targeting the fashion domain. It uses a transformer-based fusion encoder and introduces masked patch modeling to enable contextual learning of images + +![](images/922f80b5d608b76d42000ecfb65054f99e871e91ccf9babdcae8617606bafecd.jpg) +(a) Fine-grained Cross-modal Alignment before Multimodal Fusion +Fig. 2: An overview of Fashion-FINE. (a): Fine-grained cross-modal alignment through our proposed three modules (i.e., MAA, HNM-F, and C-CmA). (b): Sampling hard negatives from the mini-batch for each modality (i.e., image and text). (c): Fusion encoder-based pre-training tasks (i.e., image-text matching (ITM), prompt token prediction (PTP), and token replacement prediction (TRP)) for the given positive pair and the two hard negative pairs. + +![](images/cde9a834051775f83ca8a1de094ac306314dee4deb4dc039a10d52fbbfff807e.jpg) +(b) Hard Negative Sampling + +![](images/765dbfd8cbae699c228ebeeac816e710342282bee85a93e44cee67eea4fb1952.jpg) +(c) Pre-training Tasks in Fusion Encoder + +and text. KaleidoBERT [45] is a method for semantic alignment with word tokens after extracting image patches through various patch methods (e.g., Rotation recognition, Jigsaw puzzle solving). In fashion e-commerce, words may be used differently from their common meaning. For example, "diesel" and "canada goose" are the brand names, not a type of gas for vehicles or an animal in Canada. EICLIP [25] proposed semantic alignment methods that consider the homonymous nature of e-commerce language. However, the previous studies still had to rely on the general image-text contrastive learning method for cross-modal alignment. + +Recently, representation learning methods applied to the fashion domain have been proposed using the dual-stream+fusion architecture of general VLP methods. FashionViL [10] introduced masked patch feature classification (MPFC), which learns characteristics of image patches based on discrete codes from a predefined codebook. FashionSAP [11] proposed pre-training tasks using masked text modeling (i.e., prompt token prediction, token replacement prediction) and achieved state-of-the-art performance in fashion $\mathrm{V} + \mathrm{L}$ tasks. However, both methods base their proposed pre-training tasks on the use of a fusion encoder and still do not consider the fine-grained features of images and text in the process of deriving unimodal image and text representations for cross-modal alignment. + +In this paper, we propose Fashion-FINE to enhance fine-grained cross-modal alignment by incorporating global and local representations before multimodal fusion and by encouraging more focus on the loss of hard negatives in ITC, which has not been addressed in previous studies. In Sections 4.2 and 4.3, we demonstrate the effectiveness of our Fashion-FINE on the representative fashion $\mathrm{V + L}$ tasks through quantitative and qualitative analyses. + +# 3 Fashion-FINE + +# 3.1 Baseline Architecture + +As illustrated in Figure 2, Fashion-FINE is based on the dual-stream+fusion structure [13, 39] effective for VLP, consisting of one image encoder, one text encoder, and one fusion encoder. We utilize the 12-layers of ViT-B/16 [6] as the image encoder, the first six layers of the BERTbase [5] as the text encoder, and the last six layers of the BERTbase as the fusion encoder. The input image $I$ is encoded into a set of global and local patch embeddings $\{\pmb{v}_{cls}, \pmb{v}_1, \dots, \pmb{v}_N\} \in \mathbb{R}^{(N + 1) \times d_e}$ by the image encoder, where $\pmb{v}_{cls}$ represents the [CLS] embedding of the input image. The input text $T$ is encoded into a set of global and tokenized local patch embeddings $\{\pmb{t}_{cls}, \pmb{t}_1, \dots, \pmb{t}_M\} \in \mathbb{R}^{(M + 1) \times d_e}$ by the text encoder, where $\pmb{t}_{cls}$ represents the [CLS] embedding of the input text. The fusion encoder takes the image and text embeddings and performs cross-attention. We maintain two queues to store the representations (i.e., $\hat{\pmb{v}}_{cls}, \hat{\pmb{t}}_{cls}$ ) of momentum models for conducting large-scale contrastive learning. The architecture of the momentum model remains identical to the original baseline architecture. + +# 3.2 Modality-Agnostic Adapter (MAA) + +Our proposed MAA aims to integrate the global and local representations of images and text by learning the importance of local patch features for each modality through an attention network with shared parameters. + +$$ +\boldsymbol {V} = \left\{\boldsymbol {v} _ {1}, \dots , \boldsymbol {v} _ {N} \right\}, \quad \boldsymbol {T} = \left\{\boldsymbol {t} _ {1}, \dots , \boldsymbol {t} _ {M} \right\} \tag {1} +$$ + +where $\mathbf{V} \in \mathbb{R}^{N \times d_e}$ and $\mathbf{T} \in \mathbb{R}^{M \times d_e}$ are the sets of local patch embeddings obtained from the image and text encoder outputs, respectively. We introduce modality-agnostic attention to allow the model to learn how to assign patch importance to each local patch embedding for each modality based on the context of the input image-text pair. The local representations $\mathbf{v}_l \in \mathbb{R}^{d_e}$ and $\mathbf{t}_l \in \mathbb{R}^{d_e}$ are computed as follows. + +$$ +\boldsymbol {v} _ {l} = \sum_ {i = 1} ^ {N} \alpha_ {i} \boldsymbol {v} _ {i}, \quad \boldsymbol {t} _ {l} = \sum_ {i = 1} ^ {M} \beta_ {i} \boldsymbol {t} _ {i} \tag {2} +$$ + +$$ +\boldsymbol {u} _ {i} ^ {v} = \tanh (\boldsymbol {W} _ {w} \boldsymbol {v} _ {i} + \boldsymbol {b} _ {w}), \quad \boldsymbol {u} _ {i} ^ {t} = \tanh (\boldsymbol {W} _ {w} \boldsymbol {t} _ {i} + \boldsymbol {b} _ {w}) \tag {3} +$$ + +$$ +\alpha_ {i} = \frac {\exp \left(\boldsymbol {u} _ {i} ^ {v} \boldsymbol {u} _ {w} ^ {\top}\right)}{\sum_ {k = 1} ^ {N} \exp \left(\boldsymbol {u} _ {k} ^ {v} \boldsymbol {u} _ {w} ^ {\top}\right)}, \quad \beta_ {i} = \frac {\exp \left(\boldsymbol {u} _ {i} ^ {t} \boldsymbol {u} _ {w} ^ {\top}\right)}{\sum_ {k = 1} ^ {M} \exp \left(\boldsymbol {u} _ {k} ^ {t} \boldsymbol {u} _ {w} ^ {\top}\right)} \tag {4} +$$ + +where $\alpha_{i}$ and $\beta_{i}$ measure the importance score of the $i^{th}$ image and text patch embeddings, respectively. Note that the learning parameters (i.e., $\pmb{W}_{w},\pmb{b}_{w}$ , and $\pmb{u}_{w}$ ) are shared across modalities, and this approach encourages interactions between modalities, as demonstrated in how Chen et al. [4] uses shared learning + +parameters as discrete tokens to ground the features of both modalities in a shared manifold space. We compute the integrated representation $\mathbf{v}_{int}$ and $\mathbf{t}_{int}$ through the weighted sum of the global and local representations as follows. + +$$ +\boldsymbol {v} _ {i n t} = \lambda_ {I} \times \boldsymbol {v} _ {c l s} + (1 - \lambda_ {I}) \times \boldsymbol {v} _ {l} \tag {5} +$$ + +$$ +\boldsymbol {t} _ {i n t} = \lambda_ {T} \times \boldsymbol {t} _ {c l s} + (1 - \lambda_ {T}) \times \boldsymbol {t} _ {l} \tag {6} +$$ + +We have observed the performance changes according to the variation of $\lambda_I$ and $\lambda_T$ and our model achieved the highest performance when $\lambda_I = 0.95$ and $\lambda_T = 0.95$ . Using the integrated image and text representations, our model performs HNM-F (Section 3.3) and C-CmA (Section 3.4), and gradients are propagated not only to the [CLS] embedding, but also to the local patch embeddings. Thus, the use of integrated representations in HNM-F and C-CmA encourages the learning of local patch embeddings for fine-grained cross-modal alignment. + +# 3.3 Hard Negative Mining w/ Focal Loss (HNM-F) + +Our proposed HNM-F aims to improve the distinction learning between positive and hard negatives for a given anchor through the application of focal loss in ITC using integrated image and text representations. To account for the large scale of negatives during HNM-F, we maintain two large queues that store the most recently generated $Q$ integrated representations obtained from the momentum model. We calculate the relative similarity of $i^{th}$ positive image-text and text-image pairs in the mini-batch, respectively, as follows. + +$$ +p _ {i 2 t} ^ {(i, +)} = \frac {\exp \left(\operatorname {s i m} \left(\boldsymbol {v} _ {i n t} ^ {(i)} , \hat {\boldsymbol {t}} _ {i n t} ^ {(+)}\right) / \tau\right)}{\sum_ {j = 1} ^ {B + Q} \exp \left(\operatorname {s i m} \left(\boldsymbol {v} _ {i n t} ^ {(i)} , \hat {\boldsymbol {t}} _ {i n t} ^ {(j)}\right) / \tau\right)} \tag {7} +$$ + +$$ +p _ {t 2 i} ^ {(i, +)} = \frac {\exp \left(\operatorname {s i m} \left(\boldsymbol {t} _ {\text {i n t}} ^ {(i)} , \boldsymbol {v} _ {\text {i n t}} ^ {(+)}\right) / \tau\right)}{\sum_ {j = 1} ^ {B + Q} \exp \left(\operatorname {s i m} \left(\boldsymbol {t} _ {\text {i n t}} ^ {(i)} , \boldsymbol {v} _ {\text {i n t}} ^ {(j)}\right) / \tau\right)} \tag {8} +$$ + +where $\pmb{v}_{int}$ and $\pmb{t}_{int}$ are the integrated representations calculated in Equations 5 and 6, respectively, $B$ denotes a mini-batch size, and $\tau$ denotes a temperature parameter [12]. $sim(\pmb{v}_{int},\hat{\pmb{t}}_{int})$ and $sim(\pmb{t}_{int},\hat{\pmb{v}}_{int})$ are calculated as $g_v(\pmb{v}_{int})^\top g_t'(\hat{\pmb{t}}_{int})$ and $g_t(\pmb{t}_{int})^\top g_v'(\hat{\pmb{v}}_{int})$ in operations within the batch and queues. $g_v$ and $g_t$ are the norm(Adp(\cdot)) to obtain $\pmb{v}_{int}' \in \mathbb{R}^d$ and $\pmb{t}_{int}' \in \mathbb{R}^d$ , which are the representations output from the original model. $g_v'$ and $g_t'$ are the norm(Adp'(\cdot)) to obtain $\hat{\pmb{v}}_{int}' \in \mathbb{R}^d$ and $\hat{\pmb{t}}_{int}' \in \mathbb{R}^d$ , which are the representations output from the momentum model. Adp(\cdot) and Adp'(\cdot) are the modality-specific adaptive layers that project to $d$ dimensions and norm is the normalization function. + +One of the primary characteristics of fashion data is the significant presence of hard negative pairs that share most of the fine-grained attributes with positive pairs for a given anchor. Thus, training ITC by setting the one-hot encoded ground truth similarity for positive and negative pairs to 1 and 0, respectively, can lead to overfitting on the training data. To address this challenge, we adopted + +soft labels by combining pseudo-targets, which are calculated from the momentum image and text encoders, with the one-hot encoded hard labels through a weighted sum. We define the soft labels $\pmb{q}_{i2t}$ and $\pmb{q}_{t2i}$ as follows. + +$$ +\boldsymbol {q} _ {i 2 t} = (1 - \eta) \times \boldsymbol {y} _ {i 2 t} + \eta \times \hat {\boldsymbol {y}} _ {i 2 t} \tag {9} +$$ + +$$ +\boldsymbol {q} _ {t 2 i} = (1 - \eta) \times \boldsymbol {y} _ {t 2 i} + \eta \times \hat {\boldsymbol {y}} _ {t 2 i} \tag {10} +$$ + +where $\pmb{y}_{i2t} \in \mathbb{R}^{B \times (B + Q)}$ and $\pmb{y}_{t2i} \in \mathbb{R}^{B \times (B + Q)}$ are both represent rectangular diagonal matrices with one-hot encoding, where the main diagonal elements are 1, and all other elements are 0. $\hat{\pmb{y}}_{i2t} \in \mathbb{R}^{B \times (B + Q)}$ and $\hat{\pmb{y}}_{t2i} \in \mathbb{R}^{B \times (B + Q)}$ are pseudo-targets calculated from the dot product values between the momentum image and text encoders. We experimentally found that the optimal $\eta$ is 0.4. As demonstrated in [20], minimizing Kullback - Leibler $(\pmb{p}_{i2t}, \pmb{q}_{i2t})$ is equivalent to minimizing $-\sum_{i=1}^{B} q_{i2t}^{(i, +)} \log p_{i2t}^{(i, +)}$ and maximizing mutual information between $\pmb{v}_{int}^{(i)}$ and $\pmb{t}_{int}^{(+)}$ for text that shares similar semantic meaning with the image $I_i$ . Finally, we propose the HNM-F loss with the soft labels as follows. + +$$ +\begin{array}{l} L _ {h n m - f} = - \frac {\alpha}{2 B} \sum_ {i = 1} ^ {B} [ (1 - p _ {i 2 t} ^ {(i, +)}) ^ {\gamma} q _ {i 2 t} ^ {(i, +)} \log p _ {i 2 t} ^ {(i, +)} \\ \left. + \left(1 - p _ {t 2 i} ^ {(i, +)}\right) ^ {\gamma} q _ {t 2 i} ^ {(i, +)} \log p _ {t 2 i} ^ {(i, +)} \right] \tag {11} \\ \end{array} +$$ + +where $\alpha \in [0,1]$ is a weighting factor to mitigate the learning issues of the class imbalance between positive and negative pair classes. $(1 - p_{i2t}^{(i, + )})^{\gamma}$ and $(1 - p_{t2i}^{(i, + )})^{\gamma}$ are modulating factors based on the relative similarity of positive pairs. $\gamma \geq 0$ represents a tunable focusing parameter. A higher $\gamma$ value promotes increasing the similarity of positive pairs with low similarity and decreasing the relative similarity between positive and hard negatives in HNM-F. + +# 3.4 Comprehensive Cross-modal Alignment (C-CmA) + +C-CmA aims to enhance fine-grained semantic alignment between text and image representations by learning hierarchical text features of both low-level and high-level features that are difficult to capture with the integrated text representation (Equation 6) alone. First, to extract low-level features in the form of both words and phrases (e.g., "placket," "cinch belt," "removable fur trim"), we applied a 1-D convolution to local patch embeddings according to three kernel sizes (i.e., uni-gram, bi-gram, and tri-gram). + +$$ +\boldsymbol {t} _ {c _ {k}} = f _ {k} (\boldsymbol {T}), \quad k \in \{1, 2, 3 \} \tag {12} +$$ + +where $f_{k}$ denotes the 1-D convolution layers with kernel size $k$ . $t_{c_k} \in \mathbb{R}^{M \times d_e}$ is the embedding obtained by applying the 1-D convolution with kernel size $k$ . The derivation of the low-level representation $t_c$ is as follows. + +$$ +\boldsymbol {t} _ {c} = \frac {1}{M} \sum_ {i = 1} ^ {M} \operatorname {m a x p o o l} \left(\left[ \boldsymbol {t} _ {c _ {1}}, \boldsymbol {t} _ {c _ {2}}, \boldsymbol {t} _ {c _ {3}} \right]\right), \quad \boldsymbol {t} _ {c} \in \mathbb {R} ^ {d _ {e}} \tag {13} +$$ + +Next, we added a new fashion symbol token to the text input to obtain a high-level representation of fashion symbols (Figure 2). The text encoder of Fashion-FINE embeds the high-level representation, $\pmb{t}_s \in \mathbb{R}^{d_e}$ , for the nine pre-defined fashion symbols (i.e., tops, dresses, skirts, coats, pants, shoes, bags, accessories, and others). These symbols are pre-defined in FashionSAP based on the categories in FashionGen [33]. The total set of uni-modal embeddings from the text encoder outputs, $T_{total} \in \mathbb{R}^{(M + 2) \times d_e}$ , and the calculated multi-level text representation, $t_m \in \mathbb{R}^{d_e}$ , are as follows. + +$$ +\boldsymbol {T} _ {\text {t o t a l}} = \left\{\boldsymbol {t} _ {\text {c l s}}, \boldsymbol {t} _ {s}, \boldsymbol {t} _ {1}, \dots , \boldsymbol {t} _ {M} \right\} \tag {14} +$$ + +$$ +\boldsymbol {t} _ {m} = \delta \times \boldsymbol {t} _ {s} + (1 - \delta) \times \boldsymbol {t} _ {c} \tag {15} +$$ + +In C-CmA, note that we do not use the integrated text representation calculated in Equation 6, but instead use a text representation that integrates both low-level and high-level attributes. We have observed the performance changes according to the variation of $\delta$ and achieved the highest performance when $\delta = 0.95$ . The loss function of C-CmA is as follows. + +$$ +\boldsymbol {v} _ {i n t} ^ {\prime} = \operatorname {n o r m} \left(\operatorname {A d p} \left(\boldsymbol {v} _ {i n t}\right)\right), \quad \boldsymbol {t} _ {m} ^ {\prime} = \operatorname {n o r m} \left(\operatorname {A d p} \left(\boldsymbol {t} _ {m}\right)\right) \tag {16} +$$ + +$$ +L _ {c - c m a} = \frac {1}{B} \left[ 1 - \sum_ {b = 1} ^ {B} \frac {1}{2} \left[ v _ {i n t} ^ {\prime b} \left(\boldsymbol {t} _ {m} ^ {\prime b}\right) ^ {\top} + 1 \right] \right] \tag {17} +$$ + +where $\pmb{v}_{int}^{\prime}$ is a calculated adapted integrated image representation in Section 3.3 and $\pmb{t}_m^\prime$ are the adapted multi-level text representation. $\mathrm{Adp}(\cdot)$ and norm are the same as the used in Section 3.3, and $B$ denotes the mini-batch size. + +# 3.5 Hard Negative Sampling + +For the interactive V+L retrieval tasks (e.g., text-guided image retrieval), it is important to learn to distinguish the representations of positive and hard negative pairs. For each image-text pair in the mini-batch, we set the image and text as respective anchors and sampled one hard negative for each anchor based on the relative similarity calculated in HNM-F. The sampled two hard negatives for each modality are used to perform ITM in the fusion encoder. Thus, improving hard negative mining through our Fashion-FINE can benefit not only the cross-modal retrieval but also the text-guided image retrieval tasks, which rely on the [CLS] embedding output of the fusion encoder. We have quantitatively and qualitatively verified the effectiveness in Sections 4.2 and 4.3. This section introduces the three fusion encoder-based pre-training tasks used in the previous fashion VLP research [10, 11]. We set the pre-training tasks introduced in this section as the baseline. + +# 3.6 Image-Text Matching (ITM) + +In Fashion-FINE, ITM corresponds to a binary classification problem that takes the output [CLS] embeddings based on the cross-attention in the fusion encoder + +Table 1: Cross-modal retrieval results on FashionGen with the protocol used in previous fashion VLP research [7,10,11,25,45]. + +
MethodsITRTIRSumR
R@1R@5R@10R@1R@5R@10
VL-BERT [37]19.2639.9046.0522.6336.4848.52212.84
ViLBERT [24]20.9740.4948.2121.1237.2350.11239.18
Image-BERT [30]22.7641.8950.7724.7845.2055.90241.30
OscAR [21]23.3944.6752.5525.1049.1456.68251.53
FashionBERT [7]23.9646.3152.1226.7546.4855.74251.36
KaleidoBERT [45]27.9960.0968.3733.8860.6068.59319.52
CommerceMM [44]41.6064.0072.8039.6061.5072.70352.20
EI-CLIP [25]38.7072.2084.2540.0671.9982.90390.10
ALBEF [20]63.9788.9294.4160.5284.9991.45484.26
FashionVIL [10]65.5491.3496.3061.8887.3293.22495.60
FashionSAP [11]73.1492.8096.8770.1291.7696.38521.07
Fashion-FINE (Ours)74.8093.9497.0070.7091.9096.18524.54
+ +to classify as positive or negative. The ITM loss is as follows. + +$$ +L _ {i t m} = \mathbb {E} _ {(\boldsymbol {I}, \boldsymbol {T}) \sim D} \operatorname {C E} \left(\boldsymbol {y} ^ {i t m}, \boldsymbol {p} ^ {i t m} \left(\boldsymbol {H} _ {i t m}\right)\right) \tag {18} +$$ + +where $D$ denotes the entire dataset comprised of image-text pairs, and CE denotes the cross-entropy loss. $\pmb{H}_{itm} = [\pmb{H}_v\oplus \pmb{H}_t]$ , and $\pmb{H}_{v}$ and $\pmb{H}_{t}$ denote the uni-modal image and text representations, respectively. $\pmb{p}^{itm}(\pmb{H}_{itm})$ is the predictive probability distribution for the ITM task, and $\pmb{y}^{itm}$ is the ground truth. + +# 3.7 Pre-training Objective + +For learning prompt token prediction $(L_{ptp})$ and token replacement prediction $(L_{trp})$ in the fusion encoder, we employ the same prompt input strategies as FashionSAP [11] (see supplementary material for more details). The total loss for Fashion-FINE is computed as follows. Fashion-FINE is optimized end-to-end for the pre-training datasets by minimizing $L$ . + +$$ +L = L _ {h n m - f} + L _ {c - c m a} + L _ {i t m} + L _ {p t p} + L _ {t r p} \tag {19} +$$ + +# 4 Experiments + +# 4.1 Experimental Settings + +Dataset Following previous research [7, 10, 11, 25, 40, 45], we used two representative public benchmarks for pre-training and fashion $\mathrm{V + L}$ retrieval tasks: FashionGen [33] and FashionIQ [41]. FashionGen contains 260K image-text pairs for the training dataset and 35K for test dataset. FashionIQ contains 18K training triplets (i.e., reference image, modifying text, and target image) and 6K validation triplets across three categories: dress, shirt, and toptee. For fair experiments, we pre-trained our model using FashionGen, consistent with previous studies. For the downstream tasks, we fine-tuned our model on FashionGen for cross-modal retrieval and on FashionIQ for text-guided image retrieval. + +FashionGen with full evaluation. + +
MethodsITRTIRSumR
R@1R@5R@10R@1R@5R@10
EI-CLIP [25]25.7054.5066.8028.4057.1069.40301.90
ALBEF [20]41.6867.3975.5050.9575.3684.15395.03
FashionKLIP [40]47.1669.2775.3954.6075.0681.39402.87
FashionViL [10]42.8871.5780.5551.3475.4284.57406.33
FashionSAP [11]54.4377.3083.1562.8283.9690.16451.82
Fashion-FINE (Ours)54.2977.7483.8765.2385.6191.37458.11
+ +Table 3: Text-guided image retrieval re +Table 2: Cross-modal retrieval results on sutls on FashionIQ. + +
MethodsDressToptieShirtSumR
R@10R@50R@10R@50R@10R@50
FiLM [28]14.2333.3417.3037.6815.0434.09151.68
CIRPLANT [23]17.4540.4121.6445.3817.5338.81181.22
VAL [3]22.5344.0027.5351.6822.3844.15212.27
TIRG(bert) [38]27.1753.2527.8457.1122.2845.58233.23
CosMo [17]25.6450.3029.2157.4624.9049.18236.69
DCNet [15]28.9556.7030.4458.2923.9547.30245.63
FashionVIL [10]33.4759.9434.9860.7925.1750.39264.74
FashionVLP [8]32.4260.2938.5168.7931.8958.44290.34
FashionSAP [11]33.7160.4341.9170.9333.1761.33301.48
Fashion-FINE (Ours)36.6463.4646.8174.0938.7663.30323.06
+ +![](images/dbe38161ba97ec9dc269444a2918a3a490ba071e2c44c0264552ea9ae8e03879.jpg) +(a) + +![](images/48bea441cc2904aa28dbdf16c695c6a7022613daa5e88cfeb6b052275009b11a.jpg) +(b) +Fig. 3: Qualitative analysis results. (a) TIR: the green text in the query indicates the distinguished attributes of the primary ground truth. (b) TGIR: the green text in the modifying text indicates attributes that need to be modified in the reference image and included in the target image. Each row shows the top-10 retrieved results for each model (far left: top-1, far right: top-10), and the green box represents the ground truth. + +Implementation Details We trained the proposed modules with randomly initialized weights, and the baseline encoders were initialized with the same pre-trained weights as in FashionSAP [11]. Through hyperparameter search experiments, we found that the optimal values for HNM-F are $\tau = 0.07$ , $\alpha = 0.25$ , and $\gamma = 2.0$ (see supplementary material for more details). + +# 4.2 Cross-modal Retrieval + +Quantitative Analysis Following the previous study [11], we evaluated cross-modal retrieval performance based on the calculated relative similarity between the integrated uni-modal representations. Thus, we fine-tuned our model using MAA, HNM-F, and C-CmA. ITR aims to retrieve matched text for an input query image, while TIR aims to retrieve a matched image for an input query text. We evaluated our Fashion-FINE using the following two evaluation methods. + +First, following previous research [7, 10, 11, 25, 45], we evaluated the average performance result of five randomly sampled retrieval test sets containing 1,000 unique image and text queries. For each query in the test sets, only one candidate is matched (positive), while the other 100 candidates are mismatched (negative) + +
Attributesturtleneckbomberjacketclosurestraightdenimdistresslongsleevearrow
Baseline(Fashion-FINEw/o maa, hmm-f,c-cma)
Fashion-FINE
+ +Fig. 4: Qualitative analysis of the fine-grained alignment in the pre-training phase. The visualized results are the 1st layer of the cross-attention map in the fusion encoder. + +and randomly sampled from other products within the same subcategory. As a result, Fashion-FINE achieved state-of-the-art performance in ITR and TIR based on the SumR metric, as shown in Table 1. Second, we used the full test dataset to thoroughly evaluate the performance and reproducibility of Fashion-FINE, consistent with previous research [10,11,25,40]. As a result, Fashion-FINE also achieved state-of-the-art performance in ITR and TIR, as shown in Table 2. + +Qualitative Analysis We analyzed the cases where the ground truth retrieval rank increased in the TIR results through the ablation study (Figure 3a). The baseline model retrieved hard negatives that share the fine-grained attributes (e.g., 'plaid jacket,' 'spread collar,' and 'single-button barrel cuffs') with the ground truth at a higher rank than the ground truth. Fashion-FINE (w/o C-CmA) ranked the ground truth in 4th place, yet it ranked hard negatives-sharing the same attributes as the ground truth but differing in location or having slight color variations-ahead of the ground truth. On the other hand, Fashion-FINE ranked all ground truths between 1st and 4th place. These results represent superior qualitative achievements not captured by the $R@k$ metric presented in Table 2, which considers a retrieval correct if the retrieved top-k list includes at least one correct answer. Through the analyses, we have qualitatively demonstrated the effectiveness of Fashion-FINE for fine-grained cross-modal retrieval. + +# 4.3 Text-Guided Image Retrieval + +Quantitative Analysis We fine-tuned our model using contrastive learning based on the similarity between the output [CLS] embedding of the fusion encoder for the input query image-text pair and the output [CLS] embedding of the image encoder in Fashion-FINE for the target image. This task aims to retrieve a target image by referencing an image-text pair as a query, where the text modifies certain attributes of the image. We compared our model with fashion VLP methods under the original protocol used by FashionIQ. As a result, Fashion-FINE achieved state-of-the-art performance in TGIR, as shown in Table 3. + +Table 4: Ablation study results for our proposed modules on FashionGen and FashionIQ. We used the baseline architecture (Section 3.1) in this ablation study. + +
ArchitectureProposed ModulesCross-modal RetrievalText-Guided Image Retrieval
MAA HNM-F C-CmAITRTIRSumRDressTopteeShirt
R@1R@5R@10R@1R@5R@10R@10R@50R@10R@50
BaselineXXX54.4377.3083.1562.8283.9690.16451.8233.7160.4333.17
Baseline✓XX54.3377.4283.7564.6284.0790.18454.3732.9760.7835.43
(w/ Ours)✓✓X54.3077.8184.0064.4085.4290.74456.6736.2462.3238.03
✓✓✓54.2977.7483.8765.2385.6191.37458.1136.6463.4638.76
+ +Table 5: Experimental results on the applicability of Fashion-FINE to the foundation model (i.e., BLIP [19]) on FashionGen. + +
ArchitectureProposed ModulesCross-modal Retrieval
MAAHNM-FC-CmAITRTIR
R@1R@5R@10R@1R@5
BLIPXXX55.2778.3985.0661.6382.46
BLIP (w/ Ours)✓XX54.6078.6085.1265.0585.65
✓✓X57.2480.2086.5363.3586.41
✓✓✓59.8182.4888.1268.4987.98
+ +Qualitative Analysis We qualitatively analyzed the cases where the rank of the retrieved ground truth increased in the TGIR task according to the ablation study (Figure 3b). The baseline model failed to retrieve the ground truth in the top-10, highly ranking hard negatives that contain "no sleeves" or "two different colors." On the other hand, we observed that as more modules of Fashion-FINE were incorporated, the rank of the ground truth increased. This also qualitatively demonstrates the effectiveness of Fashion-FINE in learning fine-grained details, which is not captured in Table 3. Through the analyses, we have demonstrated that Fashion-FINE effectively learned not only the representation of the target image, but also the fine-grained alignment in the fusion encoder to obtain representations of the input reference image-modifying text pairs. + +# 4.4 Fine-grained Alignment Analysis + +To verify the fine-grained alignment in the fusion encoder, following [11], we chose the four instances from FashionGen and visualized cross-attention maps using Grad-CAM [35]. In Figure 4, Fashion-FINE more accurately highlights the regions of fine-grained attributes (e.g., "arrow," "distress") while minimizing the highlight on the background compared to the baseline. We also found that Fashion-FINE effectively highlights the region of attributes for both single words (e.g. "closure," "arrow," "straight," "denim," "distress") and phrases (e.g., "turtle neck," "bomber jacket," "long sleeve") in the form of fine-grained attributes. + +# 4.5 Ablation Study + +We conducted an ablation study to evaluate the effectiveness of the proposed modules. For fair experiments, the experimental settings are consistent in the same ablation series. Compared to the baseline, our methods achieved significant performance improvements in the $\mathrm{V + L}$ retrieval tasks (i.e., TIR and TGIR). + +These results indicate that our methods were effective in learning to distinguish the positives from the hard negatives among the image candidates for given input query text or query image-text pairs in fine-grained V+L retrieval tasks. This is supported by the significant performance improvement results in TIR and TGIR tasks, where the disentangled embedding of images is important. The notable performance improvement in TGIR resulting from the inclusion of the HNM-F module indicates that our approach effectively learned not only the disentangled embedding of the input image but also the representation of the input image-text pair (i.e., reference image and modifying text) in the fusion encoder. + +# 4.6 Applicability of Fashion-FINE to the Foundation Model + +We conducted experiments to verify the applicability of our Fashion-FINE to foundational models. For the experiments, we performed an ablation study of our modules based on the dual-stream+fusion foundational model architecture, BLIP [19]. Since BLIP uses bootstrapping on noisy data during the pre-training phase with a decoder, we applied our modules in a fine-tuning phase only, which was designed without a decoder, to the pre-trained BLIP model to evaluate the pure performance of our modules. The encoder configuration for the ITC and ITM tasks in the BLIP architecture is the same as that of Fashion-FINE, so we applied Fashion-FINE to BLIP in the same way as the baseline architecture described in Section 3. In the fine-tuning phase for the $\mathrm{V + L}$ tasks (Sections 4.2 and 4.3), since we applied the modules of Fashion-FINE only for ITR and TIR and not for TGIR, the application of Fashion-FINE to BLIP was also done only for ITR and TIR. In Table 5, we have demonstrated that the modules of Fashion-FINE can be successfully applied to the BLIP architecture, a foundational model. + +# 5 Conclusion + +In this paper, we presented Fashion-FINE, which aims to improve the learning of fine-grained cross-modal alignment. Fashion-FINE addresses the challenge of distinguishing positives from hard negatives based on MAA, which integrates global features with the fine-grained features of local representations, and on HNM-F and C-CmA, which enhance the learning effects of hard negative mining and refine the integrated image representations. Unlike previous fashion VLP research, Fashion-FINE focuses on cross-modal alignment before multimodal fusion, taking into account fine-grained details. Through quantitative and qualitative analyses, we have validated that Fashion-FINE can be successfully integrated with fusion encoder-based pre-training tasks. With this work, we advance the fashion VLP architecture and expect that our approach will contribute to the research on the development of fine-grained foundation methods. + +# Acknowledgements + +This work was supported by the National Research Foundation of Korea (NRF) grant (No.2021M3A9E4080780, No.2018R1A5A7059549) and by the Institute of + +Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean government (MSIT) (No.RS-2020-II201373, Artificial Intelligence Graduate School Program (Hanyang University)). + +# References + +1. Byun, J., Hwang, T., Fu, J., Moon, T.: Grit-vlp: Grouped mini-batch sampling for efficient vision and language pre-training. In: European Conference on Computer Vision. pp. 395-412. Springer (2022) +2. Chen, Q., Huang, J., Feris, R., Brown, L.M., Dong, J., Yan, S.: Deep domain adaptation for describing people based on fine-grained clothing attributes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5315-5324 (2015) +3. Chen, Y., Gong, S., Bazzani, L.: Image search with text feedback by visiolinguistic attention learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3001-3011 (2020) +4. Chen, Y., Yuan, J., Tian, Y., Geng, S., Li, X., Zhou, D., Metaxas, D.N., Yang, H.: Revisiting multimodal representation in contrastive learning: from patch and token embeddings to finite discrete tokens. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15095-15104 (2023) +5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) +6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +7. Gao, D., Jin, L., Chen, B., Qiu, M., Li, P., Wei, Y., Hu, Y., Wang, H.: Fashion-bert: Text and image matching with adaptive loss for cross-modal retrieval. In: Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 2251-2260 (2020) +8. Goenka, S., Zheng, Z., Jaiswal, A., Chada, R., Wu, Y., Hedau, V., Natarajan, P.: Fashionvlp: Vision language transformer for fashion retrieval with feedback. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14105-14115 (2022) +9. Gunel, B., Du, J., Conneau, A., Stoyanov, V.: Supervised contrastive learning for pre-trained language model fine-tuning. International Conference on Learning Representations (2021) +10. Han, X., Yu, L., Zhu, X., Zhang, L., Song, Y.Z., Xiang, T.: Fashionvil: Fashion-focused vision-and-language representation learning. In: European Conference on Computer Vision. pp. 634-651. Springer (2022) +11. Han, Y., Zhang, L., Chen, Q., Chen, Z., Li, Z., Yang, J., Cao, Z.: Fashionsap: Symbols and attributes prompt for fine-grained fashion vision-language pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15028-15038 (2023) +12. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) +13. Jang, J., Kong, C., Jeon, D., Kim, S., Kwak, N.: Unifying vision-language representation space with single-tower transformer. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 980-988 (2023) + +14. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021) +15. Kim, J., Yu, Y., Kim, H., Kim, G.: Dual compositional learning in interactive image retrieval. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1771-1779 (2021) +16. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017) +17. Lee, S., Kim, D., Han, B.: Cosmo: Content-style modulation for image retrieval with text feedback. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 802-812 (2021) +18. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: bootstrapping language-image pretraining with frozen image encoders and large language models. In: ICML (2023) +19. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) +20. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021) +21. Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., et al.: Oscar: Object-semantics aligned pre-training for vision-language tasks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16. pp. 121-137. Springer (2020) +22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +23. Liu, Z., Rodriguez-Opazo, C., Teney, D., Gould, S.: Image retrieval on real-life images with pre-trained vision-and-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2125-2134 (2021) +24. Lu, J., Batra, D., Parikh, D., Lee, S.: Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems 32 (2019) +25. Ma, H., Zhao, H., Lin, Z., Kale, A., Wang, Z., Yu, T., Gu, J., Choudhary, S., Xie, X.: Ei-clip: Entity-aware interventional contrastive learning for e-commerce cross-modal retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18051-18061 (2022) +26. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) +27. Ordonez, V., Kulkarni, G., Berg, T.: Im2text: Describing images using 1 million captioned photographs. Advances in neural information processing systems 24 (2011) +28. Perez, E., Strub, F., De Vries, H., Dumoulin, V., Courville, A.: Film: Visual reasoning with a general conditioning layer. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018) +29. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer + +image-to-sentence models. In: Proceedings of the IEEE international conference on computer vision. pp. 2641-2649 (2015) +30. Qi, D., Su, L., Song, J., Cui, E., Bharti, T., Sacheti, A.: Imagebert: Cross-modal pre-training with large-scale weak-supervised image-text data. arXiv preprint arXiv:2001.07966 (2020) +31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +32. Robinson, J., Chuang, C.Y., Sra, S., Jegelka, S.: Contrastive learning with hard negative samples. International Conference on Learning Representations (2021) +33. Rostamzadeh, N., Hosseini, S., Boquet, T., Stokowiec, W., Zhang, Y., Jauvin, C., Pal, C.: Fashion-gen: The generative fashion dataset and challenge. arXiv preprint arXiv:1806.08317 (2018) +34. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015) +35. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017) +36. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2556-2565 (2018) +37. Su, W., Zhu, X., Cao, Y., Li, B., Lu, L., Wei, F., Dai, J.: Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530 (2019) +38. Vo, N., Jiang, L., Sun, C., Murphy, K., Li, L.J., Fei-Fei, L., Hays, J.: Composing text and image for image retrieval—an empirical odyssey. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6439–6448 (2019) +39. Wang, J., Zhou, P., Shou, M.Z., Yan, S.: Position-guided text prompt for vision-language pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23242-23251 (2023) +40. Wang, X., Wang, C., Li, L., Li, Z., Chen, B., Jin, L., Huang, J., Xiao, Y., Gao, M.: Fashionclip: Enhancing e-commerce image-text retrieval with fashion multimodal conceptual knowledge graph. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 5: Industry Track). pp. 149-158 (2023) +41. Wu, H., Gao, Y., Guo, X., Al-Halah, Z., Rennie, S., Grauman, K., Feris, R.: Fashion iq: A new dataset towards retrieving images by natural language feedback. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 11307-11317 (2021) +42. Wu, M., Mosse, M., Zhuang, C., Yamins, D., Goodman, N.: Conditional negative sampling for contrastive learning of visual representations. In: International Conference on Learning Representations (2020) +43. Yang, J., Duan, J., Tran, S., Xu, Y., Chanda, S., Chen, L., Zeng, B., Chilimbi, T., Huang, J.: Vision-language pre-training with triple contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15671-15680 (2022) + +44. Yu, L., Chen, J., Sinha, A., Wang, M., Chen, Y., Berg, T.L., Zhang, N.: Commercemm: Large-scale commerce multimodal representation learning with omni retrieval. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 4433-4442 (2022) +45. Zhuge, M., Gao, D., Fan, D.P., Jin, L., Chen, B., Zhou, H., Qiu, M., Shao, L.: Kaleido-bert: Vision-language pre-training on fashion domain. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12647-12657 (2021) \ No newline at end of file diff --git a/2024/Integration of Global and Local Representations for Fine-grained Cross-modal Alignment/images.zip b/2024/Integration of Global and Local Representations for Fine-grained Cross-modal Alignment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..692e1027a922b8e20a5b226c2c20638530eea3e8 --- /dev/null +++ b/2024/Integration of Global and Local Representations for Fine-grained Cross-modal Alignment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b60f02e4751bfe3d77badd2376c2e23542d52054a5dd25f41135a4f730b865ab +size 543791 diff --git a/2024/Integration of Global and Local Representations for Fine-grained Cross-modal Alignment/layout.json b/2024/Integration of Global and Local Representations for Fine-grained Cross-modal Alignment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..18ee649c264690ae7b5ac6c2fdff8be8312d484d --- /dev/null +++ b/2024/Integration of Global and Local Representations for Fine-grained Cross-modal Alignment/layout.json @@ -0,0 +1,9876 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "text", + "content": "Integration of Global and Local Representations for Fine-grained Cross-modal Alignment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 169, + 468, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 169, + 468, + 181 + ], + "spans": [ + { + "bbox": [ + 146, + 169, + 468, + 181 + ], + "type": "text", + "content": "Seungwan Jin" + }, + { + "bbox": [ + 146, + 169, + 468, + 181 + ], + "type": "inline_equation", + "content": "^{\\text{®}}" + }, + { + "bbox": [ + 146, + 169, + 468, + 181 + ], + "type": "text", + "content": ", Hoyoung Choi" + }, + { + "bbox": [ + 146, + 169, + 468, + 181 + ], + "type": "inline_equation", + "content": "^{\\text{®}}" + }, + { + "bbox": [ + 146, + 169, + 468, + 181 + ], + "type": "text", + "content": ", Taehyung Noh" + }, + { + "bbox": [ + 146, + 169, + 468, + 181 + ], + "type": "inline_equation", + "content": "^{\\text{®}}" + }, + { + "bbox": [ + 146, + 169, + 468, + 181 + ], + "type": "text", + "content": ", and Kyungsik Han" + }, + { + "bbox": [ + 146, + 169, + 468, + 181 + ], + "type": "inline_equation", + "content": "^{\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 164, + 190, + 449, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 190, + 449, + 213 + ], + "spans": [ + { + "bbox": [ + 164, + 190, + 449, + 213 + ], + "type": "text", + "content": "Hanyang University, South Korea {seungwanjin, choiho00, yestaehyung, kyungsikhan}@hanyang.ac.kr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 247, + 455, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 247, + 455, + 488 + ], + "spans": [ + { + "bbox": [ + 160, + 247, + 455, + 488 + ], + "type": "text", + "content": "Abstract. Fashion is one of the representative domains of fine-grained Vision-Language Pre-training (VLP) involving a large number of images and text. Previous fashion VLP research has proposed various pretraining tasks to account for fine-grained details in multimodal fusion. However, fashion VLP research has not yet addressed the need to focus on (1) uni-modal embeddings that reflect fine-grained features and (2) hard negative samples to improve the performance of fine-grained " + }, + { + "bbox": [ + 160, + 247, + 455, + 488 + ], + "type": "inline_equation", + "content": "\\mathrm{V + L}" + }, + { + "bbox": [ + 160, + 247, + 455, + 488 + ], + "type": "text", + "content": " retrieval tasks. In this paper, we propose Fashion-FINE (Fashion VLP with Fine-grained Cross-modal Alignment using the INtegrated representations of global and local patch Embeddings), which consists of three key modules. First, a modality-agnostic adapter (MAA) learns uni-modal integrated representations and reflects fine-grained details contained in local patches. Second, hard negative mining with focal loss (HNM-F) performs cross-modal alignment using the integrated representations, focusing on hard negatives to boost the learning of fine-grained cross-modal alignment. Third, comprehensive cross-modal alignment (C-CmA) extracts low- and high-level fashion information from the text and learns the semantic alignment to encourage disentangled embedding of the integrated image representations. Fashion-FINE achieved state-of-the-art performance on two representative public benchmarks (i.e., FashionGen and FashionIQ) in three representative " + }, + { + "bbox": [ + 160, + 247, + 455, + 488 + ], + "type": "inline_equation", + "content": "\\mathrm{V + L}" + }, + { + "bbox": [ + 160, + 247, + 455, + 488 + ], + "type": "text", + "content": " retrieval tasks, demonstrating its effectiveness in learning fine-grained features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 499, + 453, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 499, + 453, + 521 + ], + "spans": [ + { + "bbox": [ + 159, + 499, + 453, + 521 + ], + "type": "text", + "content": "Keywords: Vision-Language Pre-training " + }, + { + "bbox": [ + 159, + 499, + 453, + 521 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 499, + 453, + 521 + ], + "type": "text", + "content": " Fine-grained Representation Learning " + }, + { + "bbox": [ + 159, + 499, + 453, + 521 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 499, + 453, + 521 + ], + "type": "text", + "content": " Fashion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 544, + 230, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 230, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 230, + 558 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 573, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 573, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 573, + 482, + 646 + ], + "type": "text", + "content": "With recent advances in general VLP research that aims to learn common sense knowledge, there has been increasing attention to fine-grained cross-modal representation learning. Fashion is one of the most representative VLP research domains that applies the concept of cross-modal representations based on large-scale image and text information, about fashion items for fine-grained " + }, + { + "bbox": [ + 130, + 573, + 482, + 646 + ], + "type": "inline_equation", + "content": "\\mathrm{V + L}" + }, + { + "bbox": [ + 130, + 573, + 482, + 646 + ], + "type": "text", + "content": " retrieval tasks such as cross-modal retrieval and text-guided image retrieval." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 234, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 234, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 234, + 666 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 127, + 173, + 171 + ], + "blocks": [ + { + "bbox": [ + 137, + 117, + 248, + 125 + ], + "lines": [ + { + "bbox": [ + 137, + 117, + 248, + 125 + ], + "spans": [ + { + "bbox": [ + 137, + 117, + 248, + 125 + ], + "type": "text", + "content": "(a) Image-Text Pairs in Fashion Domain" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 127, + 173, + 171 + ], + "lines": [ + { + "bbox": [ + 137, + 127, + 173, + 171 + ], + "spans": [ + { + "bbox": [ + 137, + 127, + 173, + 171 + ], + "type": "image", + "image_path": "35c804a7d8757f0a0dbc7b0b951d6267a454a85987cc436f555f1daed0677069.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 178, + 133, + 252, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 133, + 252, + 168 + ], + "spans": [ + { + "bbox": [ + 178, + 133, + 252, + 168 + ], + "type": "text", + "content": "Long sleeve down parka-style coat in olive green. Genuine raccon fur trim at hood. Two-way zip closure at front concealed by button placket. Drawstring at collar. Welt pockets and flap pockets at front. Logo patch at upper sleeve. Pleated at back panel. Fully lined. Tonal stitching. Elasticized sleeve cuffs." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 138, + 172, + 173, + 213 + ], + "blocks": [ + { + "bbox": [ + 138, + 172, + 173, + 213 + ], + "lines": [ + { + "bbox": [ + 138, + 172, + 173, + 213 + ], + "spans": [ + { + "bbox": [ + 138, + 172, + 173, + 213 + ], + "type": "image", + "image_path": "efcdb87db5dbb4f15cdd7c8095403e784913d8d1aaac7dbe37e7a1828c86590a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 178, + 175, + 252, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 175, + 252, + 212 + ], + "spans": [ + { + "bbox": [ + 178, + 175, + 252, + 212 + ], + "type": "text", + "content": "Long sleeve down parka-style coat in olive green. Removable genuine raccoon fur trim at hood. Two-way zip closure at front concealed by press-stud packet. Welt pockets and flap pockets at front. Logo patch at upper sleeve. Drawstring at waist. Fully lined. Tonal stitching. Inset ribbed knit sleeve cuff." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 138, + 215, + 173, + 256 + ], + "blocks": [ + { + "bbox": [ + 138, + 215, + 173, + 256 + ], + "lines": [ + { + "bbox": [ + 138, + 215, + 173, + 256 + ], + "spans": [ + { + "bbox": [ + 138, + 215, + 173, + 256 + ], + "type": "image", + "image_path": "242c51ded600a09d98c7c5d4a4be58b03926717afd1c560269b33747bc77b06f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 270, + 482, + 358 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 358 + ], + "type": "text", + "content": "Fig. 1: Our motivation for this research. (a) shows an example of three image-text pair cases (i.e., positive, hard negative, and easy negative pairs) in fashion products, where the colored areas specifically indicate fashion attributes that distinguish positives from hard negatives. (b) shows that fashion " + }, + { + "bbox": [ + 130, + 270, + 482, + 358 + ], + "type": "inline_equation", + "content": "\\mathrm{V + L}" + }, + { + "bbox": [ + 130, + 270, + 482, + 358 + ], + "type": "text", + "content": " tasks (e.g., cross-modal retrieval) are essential for distinguishing positives from hard negatives given anchors by capturing the fine-grained features. Our motivation is that it is difficult to distinguish between positives and hard negatives using only global representations, but this can be compensated by integrating local representations that reflect fine-grained attributes." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 177, + 217, + 253, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 217, + 253, + 255 + ], + "spans": [ + { + "bbox": [ + 177, + 217, + 253, + 255 + ], + "type": "text", + "content": "Long sleeve down parka in black. Removable hood with cinch-belt detail at crown and wired removable fur trim. Tonal bungee drawstrings at hood. Stand collar. Two-way zip closure at front concealed by press-stud placket. Tonal logo applique at exterior arm. Welt and flap pockets at front with tonal signature patch..." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 265, + 127, + 479, + 260 + ], + "blocks": [ + { + "bbox": [ + 265, + 117, + 434, + 125 + ], + "lines": [ + { + "bbox": [ + 265, + 117, + 434, + 125 + ], + "spans": [ + { + "bbox": [ + 265, + 117, + 434, + 125 + ], + "type": "text", + "content": "(b) Our Motivation for Fine-grained Cross-modal Alignment" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 265, + 127, + 479, + 260 + ], + "lines": [ + { + "bbox": [ + 265, + 127, + 479, + 260 + ], + "spans": [ + { + "bbox": [ + 265, + 127, + 479, + 260 + ], + "type": "image", + "image_path": "e52ebc137c4d3d3587b667d757ac8748b62cc05fa004e9f943fd3e6d6fe8b6da.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 398, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 482, + 554 + ], + "type": "text", + "content": "Most fashion VLP research [7, 8, 10, 11, 25, 45] has been evolved based on the general VLP architectures by applying self-supervised pre-training tasks, such as image-text contrastive learning (ITC) and image-text matching (ITM). Recently, employing the dual-stream+fusion [13, 39] structure, fashion VLP research [7, 10, 11, 45] proposed various pre-training tasks (e.g., masked patch modeling) to address the fine-grained features using the fusion encoder. FashionSAP [11], the current state-of-the-art model, proposed high-level multimodal semantic alignment and used the relative similarity between the global representations (i.e., [CLS] embedding) obtained before fusion to perform " + }, + { + "bbox": [ + 130, + 398, + 482, + 554 + ], + "type": "inline_equation", + "content": "\\mathrm{V} + \\mathrm{L}" + }, + { + "bbox": [ + 130, + 398, + 482, + 554 + ], + "type": "text", + "content": " retrieval tasks instead of the ITM score calculated in the fusion encoder. However, using only global representations has limitations in learning fine-grained features in the cross-modal alignment and poses challenges in distinguishing between positive and hard negatives for a given anchor during the test phase (Figure 1)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 558, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 483, + 666 + ], + "type": "text", + "content": "To address this, conventional methods have focused on learning to embed positive image-text pairs closer and negative pairs further using loss functions such as InfoNCE [26]. However, in self-supervised VLP, since only the pre-matched image-text pairs are considered positive and all others are considered negative, a severe class imbalance problem may arise due to the difference between the number of positive and negative pairs. Moreover, the overwhelming number of easy negative pairs compared to hard negative pairs leads to a dominance of the loss scale for easy negatives over hard negatives, thereby limiting the learning of the model to distinguish between positive and hard negative pairs." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "S. Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "In this paper, we propose Fashion-FINE (Fashion VLP with Fine-grained cross-modal alignment using the INtegrated representations of global and local patch Embeddings), which learns cross-modal alignment in three phases." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 319 + ], + "type": "text", + "content": "The first phase is to learn the importance of each local patch embedding for images and text using a modality-agnostic adapter (MAA). For each modality, we obtain a local representation that reflects the feature importance calculated by MAA at each patch location, and this local representation is then combined with the global representation, specifically the [CLS] embedding, to finally form the integrated representation. The second phase is the hard negative mining with focal loss (HNM-F), which integrates ITC and focal loss to mitigate the class and loss imbalance issues in ITC and to focus on learning to distinguish between positive and hard negative pairs. The third phase is the comprehensive cross-modal alignment (C-CmA), which semantically aligns the low-level and high-level fashion information in text descriptions with the integrated image representations. Through this, Fashion-FINE learns to encourage disentangled embedding of the image representations in the multimodal space, taking into account the comprehensive fashion text information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 321, + 347, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 321, + 347, + 332 + ], + "spans": [ + { + "bbox": [ + 146, + 321, + 347, + 332 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 342, + 480, + 463 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 342, + 479, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 479, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 479, + 388 + ], + "type": "text", + "content": "- We propose Fashion-FINE, which performs fine-grained cross-modal alignment by integrating global and local representations of the two modalities. Fashion-FINE has achieved state-of-the-art performance in all fashion V+L retrieval tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 391, + 480, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 391, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 138, + 391, + 480, + 426 + ], + "type": "text", + "content": "- We have improved the architecture for fashion VLP research by integrating Fashion-FINE with existing fusion encoder-based pre-training tasks, and we have verified its applicability through performance improvements." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 427, + 480, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 427, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 138, + 427, + 480, + 463 + ], + "type": "text", + "content": "- Our study is the first to quantitatively and qualitatively validate the importance of reflecting local patch embeddings in fine-grained cross-modal alignment and the boosting effect of hard negative mining with focal loss." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 483, + 237, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 237, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 237, + 496 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 511, + 345, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 511, + 345, + 525 + ], + "spans": [ + { + "bbox": [ + 131, + 511, + 345, + 525 + ], + "type": "text", + "content": "2.1 Vision-Language Pre-training (VLP)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": "Many VLP methods have achieved significant performance in various downstream " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{V + L}" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " tasks. The most successful VLP architectures fall into two categories: dual-stream and dual-stream " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " fusion [13, 39]. The dual-stream architecture uses two modality-specific encoders to perform pre-training tasks (e.g., ITC) for semantic alignment between the two modalities. For example, CLIP [31] and ALIGN [14] have used ITC to train image-text alignment on large-scale multimodal data, achieving remarkable performance in general downstream " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{V + L}" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " tasks. While this dual-stream is advantageous for achieving high performance, it requires extremely large-scale data during pre-training. On the other hand, the dual-stream " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " fusion architecture places multimodal fusion encoders on top of two modality-specific encoders to perform ITC followed by cross-modal learning (e.g.," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "text", + "content": "Fashion-FINE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": "MLM and ITM). ALBEF [20] and TCL [43], based on the dual-stream+fusion, have achieved superior performance compared to dual-stream models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 247 + ], + "type": "text", + "content": "Existing VLP methods have been trained on noisy and large multimodal Web data due to the limited amount of high-quality human-annotated image-text pair data [14]. Recently, BLIP [19] and BLIP-2 [18] have achieved superior performance compared to ALBEF and TCL by expanding the dual-stream+fusion structure by proposing a decoder to bootstrap noisy text captions and Q-Former to learn soft visual prompt for bootstrapping a decoder-based large language model (LLM), respectively. Rather than focusing on human-annotated image-text pairs, BLIP-based methods tackle the inherent challenges in processing the noisy and large-scale multimodal datasets collected from the Web." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 344 + ], + "type": "text", + "content": "The public benchmark datasets (e.g., FashionGen [33], FashionIQ [41]) addressed in recent fashion VLP research [10, 11, 25, 40] consist of multimodal datasets annotated by humans who manually describe fine-grained fashion attributes and contain no (or less) noise, unlike Web data. Thus, we set a pure dual-stream+fusion architecture without decoders as the baseline of FashionFINE, which is consistent with recent fashion VLP. By integrating our proposed modules with the pre-training tasks proposed in the prior work, we advance the fashion VLP architecture for the fine-grained " + }, + { + "bbox": [ + 130, + 247, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\mathrm{V + L}" + }, + { + "bbox": [ + 130, + 247, + 482, + 344 + ], + "type": "text", + "content": " retrieval tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 360, + 315, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 360, + 315, + 372 + ], + "spans": [ + { + "bbox": [ + 131, + 360, + 315, + 372 + ], + "type": "text", + "content": "2.2 Hard Negative Mining in VLP" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 379, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 534 + ], + "type": "text", + "content": "Hard negative mining is an important technique used in machine learning, especially in fields such as computer vision and natural language processing, to improve model performance [9,32,34,42]. It involves focused sampling and training on the sampled hard negatives, which are semantically similar to positive data but are negative. In VLP, hard negative mining has been recently proposed to learn to distinguish between positive and hard negative samples in the image-text matching (ITM) task by extracting hard negative samples based on the calculated relative similarity during the cross-modal alignment phase [20, 43]. To enhance the effectiveness of the hard negative sampling, GRIT-VLP [1] proposed a grouped mini-batch sampling strategy to improve in-batch hard negative sampling in dual-stream " + }, + { + "bbox": [ + 130, + 379, + 482, + 534 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 130, + 379, + 482, + 534 + ], + "type": "text", + "content": " fusion architecture. However, the previous studies still rely solely on the use of global representations, which are limited in reflecting fine-grained details when calculating relative similarity for negative sampling." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 534, + 481, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 481, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 481, + 571 + ], + "type": "text", + "content": "To the best of our knowledge, no research has been aimed at learning fine-grained local features to distinguish positive from hard negative in the cross-modal alignment to improve fine-grained " + }, + { + "bbox": [ + 130, + 534, + 481, + 571 + ], + "type": "inline_equation", + "content": "\\mathrm{V + L}" + }, + { + "bbox": [ + 130, + 534, + 481, + 571 + ], + "type": "text", + "content": " tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 586, + 297, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 297, + 598 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 297, + 598 + ], + "type": "text", + "content": "2.3 VLP in Fashion " + }, + { + "bbox": [ + 131, + 586, + 297, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{V} + \\mathbf{L}" + }, + { + "bbox": [ + 131, + 586, + 297, + 598 + ], + "type": "text", + "content": " tasks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "Compared to general VLP that uses coarse-grained common sense knowledge [16, 22, 27, 29, 36], fashion VLP should focus more on fine-grained attributes in the multimodal fashion data [2]. FashionBERT [7] is the first cross-modal retrieval method targeting the fashion domain. It uses a transformer-based fusion encoder and introduces masked patch modeling to enable contextual learning of images" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "S. Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 298, + 217 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 298, + 217 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 298, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 298, + 217 + ], + "type": "image", + "image_path": "922f80b5d608b76d42000ecfb65054f99e871e91ccf9babdcae8617606bafecd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 217, + 291, + 224 + ], + "lines": [ + { + "bbox": [ + 140, + 217, + 291, + 224 + ], + "spans": [ + { + "bbox": [ + 140, + 217, + 291, + 224 + ], + "type": "text", + "content": "(a) Fine-grained Cross-modal Alignment before Multimodal Fusion" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 236, + 480, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 236, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 480, + 304 + ], + "type": "text", + "content": "Fig. 2: An overview of Fashion-FINE. (a): Fine-grained cross-modal alignment through our proposed three modules (i.e., MAA, HNM-F, and C-CmA). (b): Sampling hard negatives from the mini-batch for each modality (i.e., image and text). (c): Fusion encoder-based pre-training tasks (i.e., image-text matching (ITM), prompt token prediction (PTP), and token replacement prediction (TRP)) for the given positive pair and the two hard negative pairs." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 303, + 114, + 383, + 216 + ], + "blocks": [ + { + "bbox": [ + 303, + 114, + 383, + 216 + ], + "lines": [ + { + "bbox": [ + 303, + 114, + 383, + 216 + ], + "spans": [ + { + "bbox": [ + 303, + 114, + 383, + 216 + ], + "type": "image", + "image_path": "cde9a834051775f83ca8a1de094ac306314dee4deb4dc039a10d52fbbfff807e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 218, + 374, + 225 + ], + "lines": [ + { + "bbox": [ + 310, + 218, + 374, + 225 + ], + "spans": [ + { + "bbox": [ + 310, + 218, + 374, + 225 + ], + "type": "text", + "content": "(b) Hard Negative Sampling" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 383, + 114, + 478, + 216 + ], + "blocks": [ + { + "bbox": [ + 383, + 114, + 478, + 216 + ], + "lines": [ + { + "bbox": [ + 383, + 114, + 478, + 216 + ], + "spans": [ + { + "bbox": [ + 383, + 114, + 478, + 216 + ], + "type": "image", + "image_path": "765dbfd8cbae699c228ebeeac816e710342282bee85a93e44cee67eea4fb1952.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 217, + 477, + 224 + ], + "lines": [ + { + "bbox": [ + 387, + 217, + 477, + 224 + ], + "spans": [ + { + "bbox": [ + 387, + 217, + 477, + 224 + ], + "type": "text", + "content": "(c) Pre-training Tasks in Fusion Encoder" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 453 + ], + "type": "text", + "content": "and text. KaleidoBERT [45] is a method for semantic alignment with word tokens after extracting image patches through various patch methods (e.g., Rotation recognition, Jigsaw puzzle solving). In fashion e-commerce, words may be used differently from their common meaning. For example, \"diesel\" and \"canada goose\" are the brand names, not a type of gas for vehicles or an animal in Canada. EICLIP [25] proposed semantic alignment methods that consider the homonymous nature of e-commerce language. However, the previous studies still had to rely on the general image-text contrastive learning method for cross-modal alignment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 463, + 482, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 584 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 584 + ], + "type": "text", + "content": "Recently, representation learning methods applied to the fashion domain have been proposed using the dual-stream+fusion architecture of general VLP methods. FashionViL [10] introduced masked patch feature classification (MPFC), which learns characteristics of image patches based on discrete codes from a predefined codebook. FashionSAP [11] proposed pre-training tasks using masked text modeling (i.e., prompt token prediction, token replacement prediction) and achieved state-of-the-art performance in fashion " + }, + { + "bbox": [ + 130, + 463, + 482, + 584 + ], + "type": "inline_equation", + "content": "\\mathrm{V} + \\mathrm{L}" + }, + { + "bbox": [ + 130, + 463, + 482, + 584 + ], + "type": "text", + "content": " tasks. However, both methods base their proposed pre-training tasks on the use of a fusion encoder and still do not consider the fine-grained features of images and text in the process of deriving unimodal image and text representations for cross-modal alignment." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we propose Fashion-FINE to enhance fine-grained cross-modal alignment by incorporating global and local representations before multimodal fusion and by encouraging more focus on the loss of hard negatives in ITC, which has not been addressed in previous studies. In Sections 4.2 and 4.3, we demonstrate the effectiveness of our Fashion-FINE on the representative fashion " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{V + L}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " tasks through quantitative and qualitative analyses." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "text", + "content": "Fashion-FINE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "text", + "content": "3 Fashion-FINE" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 270, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 270, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 270, + 152 + ], + "type": "text", + "content": "3.1 Baseline Architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 159, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 159, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 159, + 482, + 327 + ], + "type": "text", + "content": "As illustrated in Figure 2, Fashion-FINE is based on the dual-stream+fusion structure [13, 39] effective for VLP, consisting of one image encoder, one text encoder, and one fusion encoder. We utilize the 12-layers of ViT-B/16 [6] as the image encoder, the first six layers of the BERTbase [5] as the text encoder, and the last six layers of the BERTbase as the fusion encoder. The input image " + }, + { + "bbox": [ + 130, + 159, + 482, + 327 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 159, + 482, + 327 + ], + "type": "text", + "content": " is encoded into a set of global and local patch embeddings " + }, + { + "bbox": [ + 130, + 159, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\{\\pmb{v}_{cls}, \\pmb{v}_1, \\dots, \\pmb{v}_N\\} \\in \\mathbb{R}^{(N + 1) \\times d_e}" + }, + { + "bbox": [ + 130, + 159, + 482, + 327 + ], + "type": "text", + "content": " by the image encoder, where " + }, + { + "bbox": [ + 130, + 159, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{cls}" + }, + { + "bbox": [ + 130, + 159, + 482, + 327 + ], + "type": "text", + "content": " represents the [CLS] embedding of the input image. The input text " + }, + { + "bbox": [ + 130, + 159, + 482, + 327 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 159, + 482, + 327 + ], + "type": "text", + "content": " is encoded into a set of global and tokenized local patch embeddings " + }, + { + "bbox": [ + 130, + 159, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\{\\pmb{t}_{cls}, \\pmb{t}_1, \\dots, \\pmb{t}_M\\} \\in \\mathbb{R}^{(M + 1) \\times d_e}" + }, + { + "bbox": [ + 130, + 159, + 482, + 327 + ], + "type": "text", + "content": " by the text encoder, where " + }, + { + "bbox": [ + 130, + 159, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\pmb{t}_{cls}" + }, + { + "bbox": [ + 130, + 159, + 482, + 327 + ], + "type": "text", + "content": " represents the [CLS] embedding of the input text. The fusion encoder takes the image and text embeddings and performs cross-attention. We maintain two queues to store the representations (i.e., " + }, + { + "bbox": [ + 130, + 159, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{v}}_{cls}, \\hat{\\pmb{t}}_{cls}" + }, + { + "bbox": [ + 130, + 159, + 482, + 327 + ], + "type": "text", + "content": ") of momentum models for conducting large-scale contrastive learning. The architecture of the momentum model remains identical to the original baseline architecture." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 344, + 340, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 344, + 340, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 344, + 340, + 357 + ], + "type": "text", + "content": "3.2 Modality-Agnostic Adapter (MAA)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 364, + 482, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 401 + ], + "type": "text", + "content": "Our proposed MAA aims to integrate the global and local representations of images and text by learning the importance of local patch features for each modality through an attention network with shared parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 231, + 409, + 481, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 409, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 231, + 409, + 481, + 423 + ], + "type": "interline_equation", + "content": "\\boldsymbol {V} = \\left\\{\\boldsymbol {v} _ {1}, \\dots , \\boldsymbol {v} _ {N} \\right\\}, \\quad \\boldsymbol {T} = \\left\\{\\boldsymbol {t} _ {1}, \\dots , \\boldsymbol {t} _ {M} \\right\\} \\tag {1}", + "image_path": "7a18d863803be94ee242eaf9cadd0fc0402ce94bdf578ada4e64e33bdc40b190.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 431, + 482, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 482, + 505 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 482, + 505 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 431, + 482, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{N \\times d_e}" + }, + { + "bbox": [ + 130, + 431, + 482, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 431, + 482, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{T} \\in \\mathbb{R}^{M \\times d_e}" + }, + { + "bbox": [ + 130, + 431, + 482, + 505 + ], + "type": "text", + "content": " are the sets of local patch embeddings obtained from the image and text encoder outputs, respectively. We introduce modality-agnostic attention to allow the model to learn how to assign patch importance to each local patch embedding for each modality based on the context of the input image-text pair. The local representations " + }, + { + "bbox": [ + 130, + 431, + 482, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_l \\in \\mathbb{R}^{d_e}" + }, + { + "bbox": [ + 130, + 431, + 482, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 431, + 482, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_l \\in \\mathbb{R}^{d_e}" + }, + { + "bbox": [ + 130, + 431, + 482, + 505 + ], + "type": "text", + "content": " are computed as follows." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 246, + 513, + 481, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 513, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 246, + 513, + 481, + 545 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {l} = \\sum_ {i = 1} ^ {N} \\alpha_ {i} \\boldsymbol {v} _ {i}, \\quad \\boldsymbol {t} _ {l} = \\sum_ {i = 1} ^ {M} \\beta_ {i} \\boldsymbol {t} _ {i} \\tag {2}", + "image_path": "19d2b65db8e6a9464a676cfbd7ca4edc8b0e253189b4812b4a39d831e697c82e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 199, + 553, + 481, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 553, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 199, + 553, + 481, + 567 + ], + "type": "interline_equation", + "content": "\\boldsymbol {u} _ {i} ^ {v} = \\tanh (\\boldsymbol {W} _ {w} \\boldsymbol {v} _ {i} + \\boldsymbol {b} _ {w}), \\quad \\boldsymbol {u} _ {i} ^ {t} = \\tanh (\\boldsymbol {W} _ {w} \\boldsymbol {t} _ {i} + \\boldsymbol {b} _ {w}) \\tag {3}", + "image_path": "ebea5dab8dd8469b573ad27da67ee50b321508c04e9877a3a0c52fed2973c3d6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 200, + 579, + 481, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 579, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 200, + 579, + 481, + 609 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} = \\frac {\\exp \\left(\\boldsymbol {u} _ {i} ^ {v} \\boldsymbol {u} _ {w} ^ {\\top}\\right)}{\\sum_ {k = 1} ^ {N} \\exp \\left(\\boldsymbol {u} _ {k} ^ {v} \\boldsymbol {u} _ {w} ^ {\\top}\\right)}, \\quad \\beta_ {i} = \\frac {\\exp \\left(\\boldsymbol {u} _ {i} ^ {t} \\boldsymbol {u} _ {w} ^ {\\top}\\right)}{\\sum_ {k = 1} ^ {M} \\exp \\left(\\boldsymbol {u} _ {k} ^ {t} \\boldsymbol {u} _ {w} ^ {\\top}\\right)} \\tag {4}", + "image_path": "a948e9f361dc78ada45c532809febf477070279c0b32be1a6e88603c6cd83442.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta_{i}" + }, + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "text", + "content": " measure the importance score of the " + }, + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "text", + "content": " image and text patch embeddings, respectively. Note that the learning parameters (i.e., " + }, + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{W}_{w},\\pmb{b}_{w}" + }, + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{u}_{w}" + }, + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "text", + "content": ") are shared across modalities, and this approach encourages interactions between modalities, as demonstrated in how Chen et al. [4] uses shared learning" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "S. Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "text", + "content": "parameters as discrete tokens to ground the features of both modalities in a shared manifold space. We compute the integrated representation " + }, + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{int}" + }, + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{int}" + }, + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "text", + "content": " through the weighted sum of the global and local representations as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 239, + 163, + 480, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 163, + 480, + 174 + ], + "spans": [ + { + "bbox": [ + 239, + 163, + 480, + 174 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {i n t} = \\lambda_ {I} \\times \\boldsymbol {v} _ {c l s} + (1 - \\lambda_ {I}) \\times \\boldsymbol {v} _ {l} \\tag {5}", + "image_path": "31112f8e45a5108ff546c14f1959dd4602a5fe0a7fbc453f2809b2d367ad245e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 240, + 177, + 480, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 177, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 240, + 177, + 480, + 190 + ], + "type": "interline_equation", + "content": "\\boldsymbol {t} _ {i n t} = \\lambda_ {T} \\times \\boldsymbol {t} _ {c l s} + (1 - \\lambda_ {T}) \\times \\boldsymbol {t} _ {l} \\tag {6}", + "image_path": "a779d9fca2fa249ae08edd2a8e3e65f17a194369662f3552f21b5c002dcb611a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 198, + 482, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 198, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 131, + 198, + 482, + 282 + ], + "type": "text", + "content": "We have observed the performance changes according to the variation of " + }, + { + "bbox": [ + 131, + 198, + 482, + 282 + ], + "type": "inline_equation", + "content": "\\lambda_I" + }, + { + "bbox": [ + 131, + 198, + 482, + 282 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 198, + 482, + 282 + ], + "type": "inline_equation", + "content": "\\lambda_T" + }, + { + "bbox": [ + 131, + 198, + 482, + 282 + ], + "type": "text", + "content": " and our model achieved the highest performance when " + }, + { + "bbox": [ + 131, + 198, + 482, + 282 + ], + "type": "inline_equation", + "content": "\\lambda_I = 0.95" + }, + { + "bbox": [ + 131, + 198, + 482, + 282 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 198, + 482, + 282 + ], + "type": "inline_equation", + "content": "\\lambda_T = 0.95" + }, + { + "bbox": [ + 131, + 198, + 482, + 282 + ], + "type": "text", + "content": ". Using the integrated image and text representations, our model performs HNM-F (Section 3.3) and C-CmA (Section 3.4), and gradients are propagated not only to the [CLS] embedding, but also to the local patch embeddings. Thus, the use of integrated representations in HNM-F and C-CmA encourages the learning of local patch embeddings for fine-grained cross-modal alignment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 299, + 400, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 299, + 400, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 299, + 400, + 312 + ], + "type": "text", + "content": "3.3 Hard Negative Mining w/ Focal Loss (HNM-F)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 319, + 479, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 479, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 479, + 402 + ], + "type": "text", + "content": "Our proposed HNM-F aims to improve the distinction learning between positive and hard negatives for a given anchor through the application of focal loss in ITC using integrated image and text representations. To account for the large scale of negatives during HNM-F, we maintain two large queues that store the most recently generated " + }, + { + "bbox": [ + 130, + 319, + 479, + 402 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 319, + 479, + 402 + ], + "type": "text", + "content": " integrated representations obtained from the momentum model. We calculate the relative similarity of " + }, + { + "bbox": [ + 130, + 319, + 479, + 402 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 319, + 479, + 402 + ], + "type": "text", + "content": " positive image-text and text-image pairs in the mini-batch, respectively, as follows." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 223, + 412, + 480, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 412, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 223, + 412, + 480, + 449 + ], + "type": "interline_equation", + "content": "p _ {i 2 t} ^ {(i, +)} = \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {v} _ {i n t} ^ {(i)} , \\hat {\\boldsymbol {t}} _ {i n t} ^ {(+)}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {B + Q} \\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {v} _ {i n t} ^ {(i)} , \\hat {\\boldsymbol {t}} _ {i n t} ^ {(j)}\\right) / \\tau\\right)} \\tag {7}", + "image_path": "6c31c53bea76a7d466cdb2ceeb8d2c3810df33891f246ebd9381db5936ed3983.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 223, + 456, + 480, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 456, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 223, + 456, + 480, + 488 + ], + "type": "interline_equation", + "content": "p _ {t 2 i} ^ {(i, +)} = \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {t} _ {\\text {i n t}} ^ {(i)} , \\boldsymbol {v} _ {\\text {i n t}} ^ {(+)}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {B + Q} \\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {t} _ {\\text {i n t}} ^ {(i)} , \\boldsymbol {v} _ {\\text {i n t}} ^ {(j)}\\right) / \\tau\\right)} \\tag {8}", + "image_path": "33d4e2d088f3d4f713c7efd36da6668c82b92e7b7e350ed9b0ac87237217f2f6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{int}" + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{t}_{int}" + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "text", + "content": " are the integrated representations calculated in Equations 5 and 6, respectively, " + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "text", + "content": " denotes a mini-batch size, and " + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "text", + "content": " denotes a temperature parameter [12]. " + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "inline_equation", + "content": "sim(\\pmb{v}_{int},\\hat{\\pmb{t}}_{int})" + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "inline_equation", + "content": "sim(\\pmb{t}_{int},\\hat{\\pmb{v}}_{int})" + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "text", + "content": " are calculated as " + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "inline_equation", + "content": "g_v(\\pmb{v}_{int})^\\top g_t'(\\hat{\\pmb{t}}_{int})" + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "inline_equation", + "content": "g_t(\\pmb{t}_{int})^\\top g_v'(\\hat{\\pmb{v}}_{int})" + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "text", + "content": " in operations within the batch and queues. " + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "inline_equation", + "content": "g_v" + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "inline_equation", + "content": "g_t" + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "text", + "content": " are the norm(Adp(\\cdot)) to obtain " + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{int}' \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{t}_{int}' \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "text", + "content": ", which are the representations output from the original model. " + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "inline_equation", + "content": "g_v'" + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "inline_equation", + "content": "g_t'" + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "text", + "content": " are the norm(Adp'(\\cdot)) to obtain " + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{v}}_{int}' \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{t}}_{int}' \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "text", + "content": ", which are the representations output from the momentum model. Adp(\\cdot) and Adp'(\\cdot) are the modality-specific adaptive layers that project to " + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 131, + 496, + 479, + 605 + ], + "type": "text", + "content": " dimensions and norm is the normalization function." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "text", + "content": "One of the primary characteristics of fashion data is the significant presence of hard negative pairs that share most of the fine-grained attributes with positive pairs for a given anchor. Thus, training ITC by setting the one-hot encoded ground truth similarity for positive and negative pairs to 1 and 0, respectively, can lead to overfitting on the training data. To address this challenge, we adopted" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "text", + "content": "Fashion-FINE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "text", + "content": "soft labels by combining pseudo-targets, which are calculated from the momentum image and text encoders, with the one-hot encoded hard labels through a weighted sum. We define the soft labels " + }, + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "inline_equation", + "content": "\\pmb{q}_{i2t}" + }, + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "inline_equation", + "content": "\\pmb{q}_{t2i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 153 + ], + "type": "text", + "content": " as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 238, + 160, + 481, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 160, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 238, + 160, + 481, + 172 + ], + "type": "interline_equation", + "content": "\\boldsymbol {q} _ {i 2 t} = (1 - \\eta) \\times \\boldsymbol {y} _ {i 2 t} + \\eta \\times \\hat {\\boldsymbol {y}} _ {i 2 t} \\tag {9}", + "image_path": "61996fe3e09bd796d34664a429d173a719f4dd451f4c94e73ed41051b1d6f243.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 238, + 175, + 481, + 187 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 175, + 481, + 187 + ], + "spans": [ + { + "bbox": [ + 238, + 175, + 481, + 187 + ], + "type": "interline_equation", + "content": "\\boldsymbol {q} _ {t 2 i} = (1 - \\eta) \\times \\boldsymbol {y} _ {t 2 i} + \\eta \\times \\hat {\\boldsymbol {y}} _ {t 2 i} \\tag {10}", + "image_path": "46f5abcc6e62fb24aa2299415c5124372131a06e112b15aebc00ac6d902c37ae.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\pmb{y}_{i2t} \\in \\mathbb{R}^{B \\times (B + Q)}" + }, + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\pmb{y}_{t2i} \\in \\mathbb{R}^{B \\times (B + Q)}" + }, + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "text", + "content": " are both represent rectangular diagonal matrices with one-hot encoding, where the main diagonal elements are 1, and all other elements are 0. " + }, + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_{i2t} \\in \\mathbb{R}^{B \\times (B + Q)}" + }, + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_{t2i} \\in \\mathbb{R}^{B \\times (B + Q)}" + }, + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "text", + "content": " are pseudo-targets calculated from the dot product values between the momentum image and text encoders. We experimentally found that the optimal " + }, + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "text", + "content": " is 0.4. As demonstrated in [20], minimizing Kullback - Leibler " + }, + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "inline_equation", + "content": "(\\pmb{p}_{i2t}, \\pmb{q}_{i2t})" + }, + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "text", + "content": " is equivalent to minimizing " + }, + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "inline_equation", + "content": "-\\sum_{i=1}^{B} q_{i2t}^{(i, +)} \\log p_{i2t}^{(i, +)}" + }, + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "text", + "content": " and maximizing mutual information between " + }, + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{int}^{(i)}" + }, + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\pmb{t}_{int}^{(+)}" + }, + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "text", + "content": " for text that shares similar semantic meaning with the image " + }, + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 132, + 193, + 482, + 308 + ], + "type": "text", + "content": ". Finally, we propose the HNM-F loss with the soft labels as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 204, + 314, + 480, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 314, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 204, + 314, + 480, + 364 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {h n m - f} = - \\frac {\\alpha}{2 B} \\sum_ {i = 1} ^ {B} [ (1 - p _ {i 2 t} ^ {(i, +)}) ^ {\\gamma} q _ {i 2 t} ^ {(i, +)} \\log p _ {i 2 t} ^ {(i, +)} \\\\ \\left. + \\left(1 - p _ {t 2 i} ^ {(i, +)}\\right) ^ {\\gamma} q _ {t 2 i} ^ {(i, +)} \\log p _ {t 2 i} ^ {(i, +)} \\right] \\tag {11} \\\\ \\end{array}", + "image_path": "2e4ff17fa61a3affc0114530ccfe8310d793757da46b0606b6b8a0b223837ae8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 369, + 482, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 444 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 369, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 130, + 369, + 482, + 444 + ], + "type": "text", + "content": " is a weighting factor to mitigate the learning issues of the class imbalance between positive and negative pair classes. " + }, + { + "bbox": [ + 130, + 369, + 482, + 444 + ], + "type": "inline_equation", + "content": "(1 - p_{i2t}^{(i, + )})^{\\gamma}" + }, + { + "bbox": [ + 130, + 369, + 482, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 369, + 482, + 444 + ], + "type": "inline_equation", + "content": "(1 - p_{t2i}^{(i, + )})^{\\gamma}" + }, + { + "bbox": [ + 130, + 369, + 482, + 444 + ], + "type": "text", + "content": " are modulating factors based on the relative similarity of positive pairs. " + }, + { + "bbox": [ + 130, + 369, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\gamma \\geq 0" + }, + { + "bbox": [ + 130, + 369, + 482, + 444 + ], + "type": "text", + "content": " represents a tunable focusing parameter. A higher " + }, + { + "bbox": [ + 130, + 369, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 369, + 482, + 444 + ], + "type": "text", + "content": " value promotes increasing the similarity of positive pairs with low similarity and decreasing the relative similarity between positive and hard negatives in HNM-F." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 459, + 411, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 459, + 411, + 472 + ], + "spans": [ + { + "bbox": [ + 131, + 459, + 411, + 472 + ], + "type": "text", + "content": "3.4 Comprehensive Cross-modal Alignment (C-CmA)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 477, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 477, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 477, + 482, + 561 + ], + "type": "text", + "content": "C-CmA aims to enhance fine-grained semantic alignment between text and image representations by learning hierarchical text features of both low-level and high-level features that are difficult to capture with the integrated text representation (Equation 6) alone. First, to extract low-level features in the form of both words and phrases (e.g., \"placket,\" \"cinch belt,\" \"removable fur trim\"), we applied a 1-D convolution to local patch embeddings according to three kernel sizes (i.e., uni-gram, bi-gram, and tri-gram)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 249, + 568, + 481, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 568, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 249, + 568, + 481, + 581 + ], + "type": "interline_equation", + "content": "\\boldsymbol {t} _ {c _ {k}} = f _ {k} (\\boldsymbol {T}), \\quad k \\in \\{1, 2, 3 \\} \\tag {12}", + "image_path": "267e4ddf1ff0fa31a65fc08945d5f87cb94d6896a796970e8ced48aea0199e9b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 588, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 588, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 482, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 588, + 482, + 624 + ], + "type": "inline_equation", + "content": "f_{k}" + }, + { + "bbox": [ + 130, + 588, + 482, + 624 + ], + "type": "text", + "content": " denotes the 1-D convolution layers with kernel size " + }, + { + "bbox": [ + 130, + 588, + 482, + 624 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 588, + 482, + 624 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 588, + 482, + 624 + ], + "type": "inline_equation", + "content": "t_{c_k} \\in \\mathbb{R}^{M \\times d_e}" + }, + { + "bbox": [ + 130, + 588, + 482, + 624 + ], + "type": "text", + "content": " is the embedding obtained by applying the 1-D convolution with kernel size " + }, + { + "bbox": [ + 130, + 588, + 482, + 624 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 588, + 482, + 624 + ], + "type": "text", + "content": ". The derivation of the low-level representation " + }, + { + "bbox": [ + 130, + 588, + 482, + 624 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 130, + 588, + 482, + 624 + ], + "type": "text", + "content": " is as follows." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 208, + 632, + 481, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 632, + 481, + 663 + ], + "spans": [ + { + "bbox": [ + 208, + 632, + 481, + 663 + ], + "type": "interline_equation", + "content": "\\boldsymbol {t} _ {c} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\operatorname {m a x p o o l} \\left(\\left[ \\boldsymbol {t} _ {c _ {1}}, \\boldsymbol {t} _ {c _ {2}}, \\boldsymbol {t} _ {c _ {3}} \\right]\\right), \\quad \\boldsymbol {t} _ {c} \\in \\mathbb {R} ^ {d _ {e}} \\tag {13}", + "image_path": "24602261b72feab070447653e453343df2b0174cd3a0ac3a2b9ca81beb5f43c3.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "S. Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 212 + ], + "type": "text", + "content": "Next, we added a new fashion symbol token to the text input to obtain a high-level representation of fashion symbols (Figure 2). The text encoder of Fashion-FINE embeds the high-level representation, " + }, + { + "bbox": [ + 130, + 116, + 479, + 212 + ], + "type": "inline_equation", + "content": "\\pmb{t}_s \\in \\mathbb{R}^{d_e}" + }, + { + "bbox": [ + 130, + 116, + 479, + 212 + ], + "type": "text", + "content": ", for the nine pre-defined fashion symbols (i.e., tops, dresses, skirts, coats, pants, shoes, bags, accessories, and others). These symbols are pre-defined in FashionSAP based on the categories in FashionGen [33]. The total set of uni-modal embeddings from the text encoder outputs, " + }, + { + "bbox": [ + 130, + 116, + 479, + 212 + ], + "type": "inline_equation", + "content": "T_{total} \\in \\mathbb{R}^{(M + 2) \\times d_e}" + }, + { + "bbox": [ + 130, + 116, + 479, + 212 + ], + "type": "text", + "content": ", and the calculated multi-level text representation, " + }, + { + "bbox": [ + 130, + 116, + 479, + 212 + ], + "type": "inline_equation", + "content": "t_m \\in \\mathbb{R}^{d_e}" + }, + { + "bbox": [ + 130, + 116, + 479, + 212 + ], + "type": "text", + "content": ", are as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 249, + 220, + 480, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 220, + 480, + 233 + ], + "spans": [ + { + "bbox": [ + 249, + 220, + 480, + 233 + ], + "type": "interline_equation", + "content": "\\boldsymbol {T} _ {\\text {t o t a l}} = \\left\\{\\boldsymbol {t} _ {\\text {c l s}}, \\boldsymbol {t} _ {s}, \\boldsymbol {t} _ {1}, \\dots , \\boldsymbol {t} _ {M} \\right\\} \\tag {14}", + "image_path": "e140de59bc82ea8dfc487b1983f18719b383f93c5e13d2eeed3b541a66efcdae.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 254, + 236, + 480, + 248 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 236, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 254, + 236, + 480, + 248 + ], + "type": "interline_equation", + "content": "\\boldsymbol {t} _ {m} = \\delta \\times \\boldsymbol {t} _ {s} + (1 - \\delta) \\times \\boldsymbol {t} _ {c} \\tag {15}", + "image_path": "06fefd0eaf1394624b66c193a93e94f8af6a112597ff1fec35d08827a6ee3875.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 255, + 479, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 479, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 479, + 315 + ], + "type": "text", + "content": "In C-CmA, note that we do not use the integrated text representation calculated in Equation 6, but instead use a text representation that integrates both low-level and high-level attributes. We have observed the performance changes according to the variation of " + }, + { + "bbox": [ + 130, + 255, + 479, + 315 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 255, + 479, + 315 + ], + "type": "text", + "content": " and achieved the highest performance when " + }, + { + "bbox": [ + 130, + 255, + 479, + 315 + ], + "type": "inline_equation", + "content": "\\delta = 0.95" + }, + { + "bbox": [ + 130, + 255, + 479, + 315 + ], + "type": "text", + "content": ". The loss function of C-CmA is as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 202, + 323, + 480, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 323, + 480, + 337 + ], + "spans": [ + { + "bbox": [ + 202, + 323, + 480, + 337 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {i n t} ^ {\\prime} = \\operatorname {n o r m} \\left(\\operatorname {A d p} \\left(\\boldsymbol {v} _ {i n t}\\right)\\right), \\quad \\boldsymbol {t} _ {m} ^ {\\prime} = \\operatorname {n o r m} \\left(\\operatorname {A d p} \\left(\\boldsymbol {t} _ {m}\\right)\\right) \\tag {16}", + "image_path": "6a93ac8b89f22971f254950dd8be18e494b3c30d08a78a01a235203f7e2eafec.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 239, + 340, + 480, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 340, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 239, + 340, + 480, + 372 + ], + "type": "interline_equation", + "content": "L _ {c - c m a} = \\frac {1}{B} \\left[ 1 - \\sum_ {b = 1} ^ {B} \\frac {1}{2} \\left[ v _ {i n t} ^ {\\prime b} \\left(\\boldsymbol {t} _ {m} ^ {\\prime b}\\right) ^ {\\top} + 1 \\right] \\right] \\tag {17}", + "image_path": "78ec2b4f10c20b1c8ee2b7886c1a8704da914495e9e3ecbddb2f3c44341fb939.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 379, + 479, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 379, + 479, + 415 + ], + "spans": [ + { + "bbox": [ + 131, + 379, + 479, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 379, + 479, + 415 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{int}^{\\prime}" + }, + { + "bbox": [ + 131, + 379, + 479, + 415 + ], + "type": "text", + "content": " is a calculated adapted integrated image representation in Section 3.3 and " + }, + { + "bbox": [ + 131, + 379, + 479, + 415 + ], + "type": "inline_equation", + "content": "\\pmb{t}_m^\\prime" + }, + { + "bbox": [ + 131, + 379, + 479, + 415 + ], + "type": "text", + "content": " are the adapted multi-level text representation. " + }, + { + "bbox": [ + 131, + 379, + 479, + 415 + ], + "type": "inline_equation", + "content": "\\mathrm{Adp}(\\cdot)" + }, + { + "bbox": [ + 131, + 379, + 479, + 415 + ], + "type": "text", + "content": " and norm are the same as the used in Section 3.3, and " + }, + { + "bbox": [ + 131, + 379, + 479, + 415 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 131, + 379, + 479, + 415 + ], + "type": "text", + "content": " denotes the mini-batch size." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 431, + 285, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 431, + 285, + 444 + ], + "spans": [ + { + "bbox": [ + 131, + 431, + 285, + 444 + ], + "type": "text", + "content": "3.5 Hard Negative Sampling" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 450, + 480, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 480, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 480, + 606 + ], + "type": "text", + "content": "For the interactive V+L retrieval tasks (e.g., text-guided image retrieval), it is important to learn to distinguish the representations of positive and hard negative pairs. For each image-text pair in the mini-batch, we set the image and text as respective anchors and sampled one hard negative for each anchor based on the relative similarity calculated in HNM-F. The sampled two hard negatives for each modality are used to perform ITM in the fusion encoder. Thus, improving hard negative mining through our Fashion-FINE can benefit not only the cross-modal retrieval but also the text-guided image retrieval tasks, which rely on the [CLS] embedding output of the fusion encoder. We have quantitatively and qualitatively verified the effectiveness in Sections 4.2 and 4.3. This section introduces the three fusion encoder-based pre-training tasks used in the previous fashion VLP research [10, 11]. We set the pre-training tasks introduced in this section as the baseline." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 622, + 304, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 622, + 304, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 622, + 304, + 635 + ], + "type": "text", + "content": "3.6 Image-Text Matching (ITM)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "In Fashion-FINE, ITM corresponds to a binary classification problem that takes the output [CLS] embeddings based on the cross-attention in the fusion encoder" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "text", + "content": "Fashion-FINE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 201, + 140, + 410, + 256 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 1: Cross-modal retrieval results on FashionGen with the protocol used in previous fashion VLP research [7,10,11,25,45]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 201, + 140, + 410, + 256 + ], + "lines": [ + { + "bbox": [ + 201, + 140, + 410, + 256 + ], + "spans": [ + { + "bbox": [ + 201, + 140, + 410, + 256 + ], + "type": "table", + "html": "
MethodsITRTIRSumR
R@1R@5R@10R@1R@5R@10
VL-BERT [37]19.2639.9046.0522.6336.4848.52212.84
ViLBERT [24]20.9740.4948.2121.1237.2350.11239.18
Image-BERT [30]22.7641.8950.7724.7845.2055.90241.30
OscAR [21]23.3944.6752.5525.1049.1456.68251.53
FashionBERT [7]23.9646.3152.1226.7546.4855.74251.36
KaleidoBERT [45]27.9960.0968.3733.8860.6068.59319.52
CommerceMM [44]41.6064.0072.8039.6061.5072.70352.20
EI-CLIP [25]38.7072.2084.2540.0671.9982.90390.10
ALBEF [20]63.9788.9294.4160.5284.9991.45484.26
FashionVIL [10]65.5491.3496.3061.8887.3293.22495.60
FashionSAP [11]73.1492.8096.8770.1291.7696.38521.07
Fashion-FINE (Ours)74.8093.9497.0070.7091.9096.18524.54
", + "image_path": "0fa3c56a632c16a3ffc88c684d29713e8892b0c2d1eb33ef3aa9061d4ae3422c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 269, + 402, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 269, + 402, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 269, + 402, + 281 + ], + "type": "text", + "content": "to classify as positive or negative. The ITM loss is as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 221, + 289, + 481, + 305 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 289, + 481, + 305 + ], + "spans": [ + { + "bbox": [ + 221, + 289, + 481, + 305 + ], + "type": "interline_equation", + "content": "L _ {i t m} = \\mathbb {E} _ {(\\boldsymbol {I}, \\boldsymbol {T}) \\sim D} \\operatorname {C E} \\left(\\boldsymbol {y} ^ {i t m}, \\boldsymbol {p} ^ {i t m} \\left(\\boldsymbol {H} _ {i t m}\\right)\\right) \\tag {18}", + "image_path": "ecfae1d87f2b50b8705f5fb8ef7cc3d079ecb4e16d62138f9dd7cccbe6d1ad07.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 313, + 481, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 313, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 313, + 481, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 313, + 481, + 361 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 131, + 313, + 481, + 361 + ], + "type": "text", + "content": " denotes the entire dataset comprised of image-text pairs, and CE denotes the cross-entropy loss. " + }, + { + "bbox": [ + 131, + 313, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{H}_{itm} = [\\pmb{H}_v\\oplus \\pmb{H}_t]" + }, + { + "bbox": [ + 131, + 313, + 481, + 361 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 313, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{H}_{v}" + }, + { + "bbox": [ + 131, + 313, + 481, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 313, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{H}_{t}" + }, + { + "bbox": [ + 131, + 313, + 481, + 361 + ], + "type": "text", + "content": " denote the uni-modal image and text representations, respectively. " + }, + { + "bbox": [ + 131, + 313, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{itm}(\\pmb{H}_{itm})" + }, + { + "bbox": [ + 131, + 313, + 481, + 361 + ], + "type": "text", + "content": " is the predictive probability distribution for the ITM task, and " + }, + { + "bbox": [ + 131, + 313, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{y}^{itm}" + }, + { + "bbox": [ + 131, + 313, + 481, + 361 + ], + "type": "text", + "content": " is the ground truth." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 391, + 274, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 274, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 274, + 403 + ], + "type": "text", + "content": "3.7 Pre-training Objective" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 411, + 481, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 411, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 131, + 411, + 481, + 472 + ], + "type": "text", + "content": "For learning prompt token prediction " + }, + { + "bbox": [ + 131, + 411, + 481, + 472 + ], + "type": "inline_equation", + "content": "(L_{ptp})" + }, + { + "bbox": [ + 131, + 411, + 481, + 472 + ], + "type": "text", + "content": " and token replacement prediction " + }, + { + "bbox": [ + 131, + 411, + 481, + 472 + ], + "type": "inline_equation", + "content": "(L_{trp})" + }, + { + "bbox": [ + 131, + 411, + 481, + 472 + ], + "type": "text", + "content": " in the fusion encoder, we employ the same prompt input strategies as FashionSAP [11] (see supplementary material for more details). The total loss for Fashion-FINE is computed as follows. Fashion-FINE is optimized end-to-end for the pre-training datasets by minimizing " + }, + { + "bbox": [ + 131, + 411, + 481, + 472 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 131, + 411, + 481, + 472 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 208, + 480, + 481, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 480, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 208, + 480, + 481, + 495 + ], + "type": "interline_equation", + "content": "L = L _ {h n m - f} + L _ {c - c m a} + L _ {i t m} + L _ {p t p} + L _ {t r p} \\tag {19}", + "image_path": "62c72b2c1abcc3a6de7387ed6729db5a922333d42a7f32ea0a2743bf1ccd7a9b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 512, + 230, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 230, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 230, + 525 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 537, + 273, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 273, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 273, + 550 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 557, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 666 + ], + "type": "text", + "content": "Dataset Following previous research [7, 10, 11, 25, 40, 45], we used two representative public benchmarks for pre-training and fashion " + }, + { + "bbox": [ + 130, + 557, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{V + L}" + }, + { + "bbox": [ + 130, + 557, + 481, + 666 + ], + "type": "text", + "content": " retrieval tasks: FashionGen [33] and FashionIQ [41]. FashionGen contains 260K image-text pairs for the training dataset and 35K for test dataset. FashionIQ contains 18K training triplets (i.e., reference image, modifying text, and target image) and 6K validation triplets across three categories: dress, shirt, and toptee. For fair experiments, we pre-trained our model using FashionGen, consistent with previous studies. For the downstream tasks, we fine-tuned our model on FashionGen for cross-modal retrieval and on FashionIQ for text-guided image retrieval." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "S. Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 150, + 304, + 213 + ], + "blocks": [ + { + "bbox": [ + 133, + 137, + 266, + 147 + ], + "lines": [ + { + "bbox": [ + 133, + 137, + 266, + 147 + ], + "spans": [ + { + "bbox": [ + 133, + 137, + 266, + 147 + ], + "type": "text", + "content": "FashionGen with full evaluation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 150, + 304, + 213 + ], + "lines": [ + { + "bbox": [ + 134, + 150, + 304, + 213 + ], + "spans": [ + { + "bbox": [ + 134, + 150, + 304, + 213 + ], + "type": "table", + "html": "
MethodsITRTIRSumR
R@1R@5R@10R@1R@5R@10
EI-CLIP [25]25.7054.5066.8028.4057.1069.40301.90
ALBEF [20]41.6867.3975.5050.9575.3684.15395.03
FashionKLIP [40]47.1669.2775.3954.6075.0681.39402.87
FashionViL [10]42.8871.5780.5551.3475.4284.57406.33
FashionSAP [11]54.4377.3083.1562.8283.9690.16451.82
Fashion-FINE (Ours)54.2977.7483.8765.2385.6191.37458.11
", + "image_path": "cec4478470228318b64c07c432f45a22951c53fe48ee9b9583b37bf400517e26.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 308, + 140, + 476, + 223 + ], + "blocks": [ + { + "bbox": [ + 304, + 114, + 477, + 125 + ], + "lines": [ + { + "bbox": [ + 304, + 114, + 477, + 125 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 477, + 125 + ], + "type": "text", + "content": "Table 3: Text-guided image retrieval re" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 125, + 386, + 137 + ], + "lines": [ + { + "bbox": [ + 133, + 125, + 386, + 137 + ], + "spans": [ + { + "bbox": [ + 133, + 125, + 386, + 137 + ], + "type": "text", + "content": "Table 2: Cross-modal retrieval results on sutls on FashionIQ." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 140, + 476, + 223 + ], + "lines": [ + { + "bbox": [ + 308, + 140, + 476, + 223 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 476, + 223 + ], + "type": "table", + "html": "
MethodsDressToptieShirtSumR
R@10R@50R@10R@50R@10R@50
FiLM [28]14.2333.3417.3037.6815.0434.09151.68
CIRPLANT [23]17.4540.4121.6445.3817.5338.81181.22
VAL [3]22.5344.0027.5351.6822.3844.15212.27
TIRG(bert) [38]27.1753.2527.8457.1122.2845.58233.23
CosMo [17]25.6450.3029.2157.4624.9049.18236.69
DCNet [15]28.9556.7030.4458.2923.9547.30245.63
FashionVIL [10]33.4759.9434.9860.7925.1750.39264.74
FashionVLP [8]32.4260.2938.5168.7931.8958.44290.34
FashionSAP [11]33.7160.4341.9170.9333.1761.33301.48
Fashion-FINE (Ours)36.6463.4646.8174.0938.7663.30323.06
", + "image_path": "0eee957202a4c31564cce61395702fb52512e66dffc3c91cf3fa6418d7ac9355.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 140, + 224, + 305, + 352 + ], + "blocks": [ + { + "bbox": [ + 140, + 224, + 305, + 352 + ], + "lines": [ + { + "bbox": [ + 140, + 224, + 305, + 352 + ], + "spans": [ + { + "bbox": [ + 140, + 224, + 305, + 352 + ], + "type": "image", + "image_path": "dbe38161ba97ec9dc269444a2918a3a490ba071e2c44c0264552ea9ae8e03879.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 354, + 228, + 363 + ], + "lines": [ + { + "bbox": [ + 216, + 354, + 228, + 363 + ], + "spans": [ + { + "bbox": [ + 216, + 354, + 228, + 363 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 224, + 476, + 352 + ], + "blocks": [ + { + "bbox": [ + 309, + 224, + 476, + 352 + ], + "lines": [ + { + "bbox": [ + 309, + 224, + 476, + 352 + ], + "spans": [ + { + "bbox": [ + 309, + 224, + 476, + 352 + ], + "type": "image", + "image_path": "48bea441cc2904aa28dbdf16c695c6a7022613daa5e88cfeb6b052275009b11a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 353, + 398, + 362 + ], + "lines": [ + { + "bbox": [ + 385, + 353, + 398, + 362 + ], + "spans": [ + { + "bbox": [ + 385, + 353, + 398, + 362 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 371, + 482, + 428 + ], + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 428 + ], + "type": "text", + "content": "Fig. 3: Qualitative analysis results. (a) TIR: the green text in the query indicates the distinguished attributes of the primary ground truth. (b) TGIR: the green text in the modifying text indicates attributes that need to be modified in the reference image and included in the target image. Each row shows the top-10 retrieved results for each model (far left: top-1, far right: top-10), and the green box represents the ground truth." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "text", + "content": "Implementation Details We trained the proposed modules with randomly initialized weights, and the baseline encoders were initialized with the same pre-trained weights as in FashionSAP [11]. Through hyperparameter search experiments, we found that the optimal values for HNM-F are " + }, + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\tau = 0.07" + }, + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\alpha = 0.25" + }, + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\gamma = 2.0" + }, + { + "bbox": [ + 130, + 449, + 482, + 510 + ], + "type": "text", + "content": " (see supplementary material for more details)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 526, + 273, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 273, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 273, + 538 + ], + "type": "text", + "content": "4.2 Cross-modal Retrieval" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": "Quantitative Analysis Following the previous study [11], we evaluated cross-modal retrieval performance based on the calculated relative similarity between the integrated uni-modal representations. Thus, we fine-tuned our model using MAA, HNM-F, and C-CmA. ITR aims to retrieve matched text for an input query image, while TIR aims to retrieve a matched image for an input query text. We evaluated our Fashion-FINE using the following two evaluation methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "First, following previous research [7, 10, 11, 25, 45], we evaluated the average performance result of five randomly sampled retrieval test sets containing 1,000 unique image and text queries. For each query in the test sets, only one candidate is matched (positive), while the other 100 candidates are mismatched (negative)" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "text", + "content": "Fashion-FINE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 481, + 213 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 481, + 213 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 481, + 213 + ], + "type": "table", + "html": "
Attributesturtleneckbomberjacketclosurestraightdenimdistresslongsleevearrow
Baseline(Fashion-FINEw/o maa, hmm-f,c-cma)
Fashion-FINE
", + "image_path": "ba351f8ad192219823c14ca9b8c1271300e7bce4f7a1e3768e9688ffad03c671.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 220, + 482, + 244 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 220, + 482, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 482, + 244 + ], + "type": "text", + "content": "Fig. 4: Qualitative analysis of the fine-grained alignment in the pre-training phase. The visualized results are the 1st layer of the cross-attention map in the fusion encoder." + } + ] + } + ], + "index": 3, + "type": "text" + }, + { + "bbox": [ + 130, + 274, + 482, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 348 + ], + "type": "text", + "content": "and randomly sampled from other products within the same subcategory. As a result, Fashion-FINE achieved state-of-the-art performance in ITR and TIR based on the SumR metric, as shown in Table 1. Second, we used the full test dataset to thoroughly evaluate the performance and reproducibility of Fashion-FINE, consistent with previous research [10,11,25,40]. As a result, Fashion-FINE also achieved state-of-the-art performance in ITR and TIR, as shown in Table 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 364, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 521 + ], + "type": "text", + "content": "Qualitative Analysis We analyzed the cases where the ground truth retrieval rank increased in the TIR results through the ablation study (Figure 3a). The baseline model retrieved hard negatives that share the fine-grained attributes (e.g., 'plaid jacket,' 'spread collar,' and 'single-button barrel cuffs') with the ground truth at a higher rank than the ground truth. Fashion-FINE (w/o C-CmA) ranked the ground truth in 4th place, yet it ranked hard negatives-sharing the same attributes as the ground truth but differing in location or having slight color variations-ahead of the ground truth. On the other hand, Fashion-FINE ranked all ground truths between 1st and 4th place. These results represent superior qualitative achievements not captured by the " + }, + { + "bbox": [ + 130, + 364, + 482, + 521 + ], + "type": "inline_equation", + "content": "R@k" + }, + { + "bbox": [ + 130, + 364, + 482, + 521 + ], + "type": "text", + "content": " metric presented in Table 2, which considers a retrieval correct if the retrieved top-k list includes at least one correct answer. Through the analyses, we have qualitatively demonstrated the effectiveness of Fashion-FINE for fine-grained cross-modal retrieval." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 549, + 308, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 549, + 308, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 308, + 562 + ], + "type": "text", + "content": "4.3 Text-Guided Image Retrieval" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Quantitative Analysis We fine-tuned our model using contrastive learning based on the similarity between the output [CLS] embedding of the fusion encoder for the input query image-text pair and the output [CLS] embedding of the image encoder in Fashion-FINE for the target image. This task aims to retrieve a target image by referencing an image-text pair as a query, where the text modifies certain attributes of the image. We compared our model with fashion VLP methods under the original protocol used by FashionIQ. As a result, Fashion-FINE achieved state-of-the-art performance in TGIR, as shown in Table 3." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "S. Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 140, + 479, + 208 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 4: Ablation study results for our proposed modules on FashionGen and FashionIQ. We used the baseline architecture (Section 3.1) in this ablation study." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 140, + 479, + 208 + ], + "lines": [ + { + "bbox": [ + 134, + 140, + 479, + 208 + ], + "spans": [ + { + "bbox": [ + 134, + 140, + 479, + 208 + ], + "type": "table", + "html": "
ArchitectureProposed ModulesCross-modal RetrievalText-Guided Image Retrieval
MAA HNM-F C-CmAITRTIRSumRDressTopteeShirt
R@1R@5R@10R@1R@5R@10R@10R@50R@10R@50
BaselineXXX54.4377.3083.1562.8283.9690.16451.8233.7160.4333.17
Baseline✓XX54.3377.4283.7564.6284.0790.18454.3732.9760.7835.43
(w/ Ours)✓✓X54.3077.8184.0064.4085.4290.74456.6736.2462.3238.03
✓✓✓54.2977.7483.8765.2385.6191.37458.1136.6463.4638.76
", + "image_path": "71fce061dc5cd1f1ecc736eea5b95abb37969948a766842c7ce3e760d54bd7d2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 194, + 236, + 416, + 304 + ], + "blocks": [ + { + "bbox": [ + 132, + 209, + 479, + 232 + ], + "lines": [ + { + "bbox": [ + 132, + 209, + 479, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 479, + 232 + ], + "type": "text", + "content": "Table 5: Experimental results on the applicability of Fashion-FINE to the foundation model (i.e., BLIP [19]) on FashionGen." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 194, + 236, + 416, + 304 + ], + "lines": [ + { + "bbox": [ + 194, + 236, + 416, + 304 + ], + "spans": [ + { + "bbox": [ + 194, + 236, + 416, + 304 + ], + "type": "table", + "html": "
ArchitectureProposed ModulesCross-modal Retrieval
MAAHNM-FC-CmAITRTIR
R@1R@5R@10R@1R@5
BLIPXXX55.2778.3985.0661.6382.46
BLIP (w/ Ours)✓XX54.6078.6085.1265.0585.65
✓✓X57.2480.2086.5363.3586.41
✓✓✓59.8182.4888.1268.4987.98
", + "image_path": "9db9766806f6c90f56a91202c83f5fdbbc90a0aa8d5649c3059109fe49c3214a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 317, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 317, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 482, + 449 + ], + "type": "text", + "content": "Qualitative Analysis We qualitatively analyzed the cases where the rank of the retrieved ground truth increased in the TGIR task according to the ablation study (Figure 3b). The baseline model failed to retrieve the ground truth in the top-10, highly ranking hard negatives that contain \"no sleeves\" or \"two different colors.\" On the other hand, we observed that as more modules of Fashion-FINE were incorporated, the rank of the ground truth increased. This also qualitatively demonstrates the effectiveness of Fashion-FINE in learning fine-grained details, which is not captured in Table 3. Through the analyses, we have demonstrated that Fashion-FINE effectively learned not only the representation of the target image, but also the fine-grained alignment in the fusion encoder to obtain representations of the input reference image-modifying text pairs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 466, + 326, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 466, + 326, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 466, + 326, + 478 + ], + "type": "text", + "content": "4.4 Fine-grained Alignment Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "type": "text", + "content": "To verify the fine-grained alignment in the fusion encoder, following [11], we chose the four instances from FashionGen and visualized cross-attention maps using Grad-CAM [35]. In Figure 4, Fashion-FINE more accurately highlights the regions of fine-grained attributes (e.g., \"arrow,\" \"distress\") while minimizing the highlight on the background compared to the baseline. We also found that Fashion-FINE effectively highlights the region of attributes for both single words (e.g. \"closure,\" \"arrow,\" \"straight,\" \"denim,\" \"distress\") and phrases (e.g., \"turtle neck,\" \"bomber jacket,\" \"long sleeve\") in the form of fine-grained attributes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 598, + 237, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 237, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 237, + 610 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "We conducted an ablation study to evaluate the effectiveness of the proposed modules. For fair experiments, the experimental settings are consistent in the same ablation series. Compared to the baseline, our methods achieved significant performance improvements in the " + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{V + L}" + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": " retrieval tasks (i.e., TIR and TGIR)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "text", + "content": "Fashion-FINE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "These results indicate that our methods were effective in learning to distinguish the positives from the hard negatives among the image candidates for given input query text or query image-text pairs in fine-grained V+L retrieval tasks. This is supported by the significant performance improvement results in TIR and TGIR tasks, where the disentangled embedding of images is important. The notable performance improvement in TGIR resulting from the inclusion of the HNM-F module indicates that our approach effectively learned not only the disentangled embedding of the input image but also the representation of the input image-text pair (i.e., reference image and modifying text) in the fusion encoder." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 238, + 442, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 238, + 442, + 251 + ], + "spans": [ + { + "bbox": [ + 131, + 238, + 442, + 251 + ], + "type": "text", + "content": "4.6 Applicability of Fashion-FINE to the Foundation Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 255, + 482, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 423 + ], + "type": "text", + "content": "We conducted experiments to verify the applicability of our Fashion-FINE to foundational models. For the experiments, we performed an ablation study of our modules based on the dual-stream+fusion foundational model architecture, BLIP [19]. Since BLIP uses bootstrapping on noisy data during the pre-training phase with a decoder, we applied our modules in a fine-tuning phase only, which was designed without a decoder, to the pre-trained BLIP model to evaluate the pure performance of our modules. The encoder configuration for the ITC and ITM tasks in the BLIP architecture is the same as that of Fashion-FINE, so we applied Fashion-FINE to BLIP in the same way as the baseline architecture described in Section 3. In the fine-tuning phase for the " + }, + { + "bbox": [ + 130, + 255, + 482, + 423 + ], + "type": "inline_equation", + "content": "\\mathrm{V + L}" + }, + { + "bbox": [ + 130, + 255, + 482, + 423 + ], + "type": "text", + "content": " tasks (Sections 4.2 and 4.3), since we applied the modules of Fashion-FINE only for ITR and TIR and not for TGIR, the application of Fashion-FINE to BLIP was also done only for ITR and TIR. In Table 5, we have demonstrated that the modules of Fashion-FINE can be successfully applied to the BLIP architecture, a foundational model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 438, + 220, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 438, + 220, + 451 + ], + "spans": [ + { + "bbox": [ + 131, + 438, + 220, + 451 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 460, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 605 + ], + "type": "text", + "content": "In this paper, we presented Fashion-FINE, which aims to improve the learning of fine-grained cross-modal alignment. Fashion-FINE addresses the challenge of distinguishing positives from hard negatives based on MAA, which integrates global features with the fine-grained features of local representations, and on HNM-F and C-CmA, which enhance the learning effects of hard negative mining and refine the integrated image representations. Unlike previous fashion VLP research, Fashion-FINE focuses on cross-modal alignment before multimodal fusion, taking into account fine-grained details. Through quantitative and qualitative analyses, we have validated that Fashion-FINE can be successfully integrated with fusion encoder-based pre-training tasks. With this work, we advance the fashion VLP architecture and expect that our approach will contribute to the research on the development of fine-grained foundation methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 620, + 247, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 247, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 247, + 633 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "This work was supported by the National Research Foundation of Korea (NRF) grant (No.2021M3A9E4080780, No.2018R1A5A7059549) and by the Institute of" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "S. Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean government (MSIT) (No.RS-2020-II201373, Artificial Intelligence Graduate School Program (Hanyang University))." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 168, + 197, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 168, + 197, + 180 + ], + "spans": [ + { + "bbox": [ + 133, + 168, + 197, + 180 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 190, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 190, + 480, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 190, + 480, + 222 + ], + "spans": [ + { + "bbox": [ + 138, + 190, + 480, + 222 + ], + "type": "text", + "content": "1. Byun, J., Hwang, T., Fu, J., Moon, T.: Grit-vlp: Grouped mini-batch sampling for efficient vision and language pre-training. In: European Conference on Computer Vision. pp. 395-412. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 224, + 480, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 480, + 265 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 480, + 265 + ], + "type": "text", + "content": "2. Chen, Q., Huang, J., Feris, R., Brown, L.M., Dong, J., Yan, S.: Deep domain adaptation for describing people based on fine-grained clothing attributes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5315-5324 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 266, + 480, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 480, + 297 + ], + "type": "text", + "content": "3. Chen, Y., Gong, S., Bazzani, L.: Image search with text feedback by visiolinguistic attention learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3001-3011 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 299, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 341 + ], + "type": "text", + "content": "4. Chen, Y., Yuan, J., Tian, Y., Geng, S., Li, X., Zhou, D., Metaxas, D.N., Yang, H.: Revisiting multimodal representation in contrastive learning: from patch and token embeddings to finite discrete tokens. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15095-15104 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "type": "text", + "content": "5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 374, + 480, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 480, + 416 + ], + "type": "text", + "content": "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 418, + 480, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 480, + 459 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 480, + 459 + ], + "type": "text", + "content": "7. Gao, D., Jin, L., Chen, B., Qiu, M., Li, P., Wei, Y., Hu, Y., Wang, H.: Fashion-bert: Text and image matching with adaptive loss for cross-modal retrieval. In: Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 2251-2260 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 460, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 502 + ], + "type": "text", + "content": "8. Goenka, S., Zheng, Z., Jaiswal, A., Chada, R., Wu, Y., Hedau, V., Natarajan, P.: Fashionvlp: Vision language transformer for fashion retrieval with feedback. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14105-14115 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 504, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 480, + 535 + ], + "type": "text", + "content": "9. Gunel, B., Du, J., Conneau, A., Stoyanov, V.: Supervised contrastive learning for pre-trained language model fine-tuning. International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 134, + 536, + 480, + 567 + ], + "type": "text", + "content": "10. Han, X., Yu, L., Zhu, X., Zhang, L., Song, Y.Z., Xiang, T.: Fashionvil: Fashion-focused vision-and-language representation learning. In: European Conference on Computer Vision. pp. 634-651. Springer (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 569, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 569, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 134, + 569, + 480, + 610 + ], + "type": "text", + "content": "11. Han, Y., Zhang, L., Chen, Q., Chen, Z., Li, Z., Yang, J., Cao, Z.: Fashionsap: Symbols and attributes prompt for fine-grained fashion vision-language pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15028-15038 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "type": "text", + "content": "12. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 634, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 634, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 634, + 480, + 665 + ], + "type": "text", + "content": "13. Jang, J., Kong, C., Jeon, D., Kim, S., Kwak, N.: Unifying vision-language representation space with single-tower transformer. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 980-988 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "text", + "content": "Fashion-FINE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "14. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "text", + "content": "15. Kim, J., Yu, Y., Kim, H., Kim, G.: Dual compositional learning in interactive image retrieval. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1771-1779 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "type": "text", + "content": "16. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 481, + 270 + ], + "type": "text", + "content": "17. Lee, S., Kim, D., Han, B.: Cosmo: Content-style modulation for image retrieval with text feedback. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 802-812 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "text", + "content": "18. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: bootstrapping language-image pretraining with frozen image encoders and large language models. In: ICML (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "text", + "content": "19. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "text", + "content": "20. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "text", + "content": "21. Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., et al.: Oscar: Object-semantics aligned pre-training for vision-language tasks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16. pp. 121-137. Springer (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 402, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 446 + ], + "type": "text", + "content": "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "23. Liu, Z., Rodriguez-Opazo, C., Teney, D., Gould, S.: Image retrieval on real-life images with pre-trained vision-and-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2125-2134 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "24. Lu, J., Batra, D., Parikh, D., Lee, S.: Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 511, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 481, + 555 + ], + "type": "text", + "content": "25. Ma, H., Zhao, H., Lin, Z., Kale, A., Wang, Z., Yu, T., Gu, J., Choudhary, S., Xie, X.: Ei-clip: Entity-aware interventional contrastive learning for e-commerce cross-modal retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18051-18061 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "type": "text", + "content": "26. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "text", + "content": "27. Ordonez, V., Kulkarni, G., Berg, T.: Im2text: Describing images using 1 million captioned photographs. Advances in neural information processing systems 24 (2011)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "text", + "content": "28. Perez, E., Strub, F., De Vries, H., Dumoulin, V., Courville, A.: Film: Visual reasoning with a general conditioning layer. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "29. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "S. Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 149, + 117, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 117, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 117, + 481, + 138 + ], + "type": "text", + "content": "image-to-sentence models. In: Proceedings of the IEEE international conference on computer vision. pp. 2641-2649 (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "text", + "content": "30. Qi, D., Su, L., Song, J., Cui, E., Bharti, T., Sacheti, A.: Imagebert: Cross-modal pre-training with large-scale weak-supervised image-text data. arXiv preprint arXiv:2001.07966 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 215 + ], + "type": "text", + "content": "31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "32. Robinson, J., Chuang, C.Y., Sra, S., Jegelka, S.: Contrastive learning with hard negative samples. International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "text", + "content": "33. Rostamzadeh, N., Hosseini, S., Boquet, T., Stokowiec, W., Zhang, Y., Jauvin, C., Pal, C.: Fashion-gen: The generative fashion dataset and challenge. arXiv preprint arXiv:1806.08317 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "34. Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 815-823 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "type": "text", + "content": "35. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 391 + ], + "type": "text", + "content": "36. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2556-2565 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "text", + "content": "37. Su, W., Zhu, X., Cao, Y., Li, B., Lu, L., Wei, F., Dai, J.: Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 414, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 456 + ], + "type": "text", + "content": "38. Vo, N., Jiang, L., Sun, C., Murphy, K., Li, L.J., Fei-Fei, L., Hays, J.: Composing text and image for image retrieval—an empirical odyssey. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6439–6448 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "text", + "content": "39. Wang, J., Zhou, P., Shou, M.Z., Yan, S.: Position-guided text prompt for vision-language pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23242-23251 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 491, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 544 + ], + "type": "text", + "content": "40. Wang, X., Wang, C., Li, L., Li, Z., Chen, B., Jin, L., Huang, J., Xiao, Y., Gao, M.: Fashionclip: Enhancing e-commerce image-text retrieval with fashion multimodal conceptual knowledge graph. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 5: Industry Track). pp. 149-158 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 544, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 588 + ], + "type": "text", + "content": "41. Wu, H., Gao, Y., Guo, X., Al-Halah, Z., Rennie, S., Grauman, K., Feris, R.: Fashion iq: A new dataset towards retrieving images by natural language feedback. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 11307-11317 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "text", + "content": "42. Wu, M., Mosse, M., Zhuang, C., Yamins, D., Goodman, N.: Conditional negative sampling for contrastive learning of visual representations. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "text", + "content": "43. Yang, J., Duan, J., Tran, S., Xu, Y., Chanda, S., Chen, L., Zeng, B., Chilimbi, T., Huang, J.: Vision-language pre-training with triple contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15671-15680 (2022)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "text", + "content": "Fashion-FINE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 205 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "type": "text", + "content": "44. Yu, L., Chen, J., Sinha, A., Wang, M., Chen, Y., Berg, T.L., Zhang, N.: Commercemm: Large-scale commerce multimodal representation learning with omni retrieval. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 4433-4442 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 205 + ], + "type": "text", + "content": "45. Zhuge, M., Gao, D., Fan, D.P., Jin, L., Chen, B., Zhou, H., Qiu, M., Shao, L.: Kaleido-bert: Vision-language pre-training on fashion domain. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12647-12657 (2021)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "S. Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks/5bfa27b2-acad-4e1e-b905-e3f3b1d14c81_content_list.json b/2024/Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks/5bfa27b2-acad-4e1e-b905-e3f3b1d14c81_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..29446e36339181521fa08b500d8b3bb7c496a52f --- /dev/null +++ b/2024/Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks/5bfa27b2-acad-4e1e-b905-e3f3b1d14c81_content_list.json @@ -0,0 +1,1773 @@ +[ + { + "type": "text", + "text": "Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks", + "text_level": 1, + "bbox": [ + 248, + 140, + 754, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lingzhuang Meng $^{1}$ , Mingwen Shao $^{1\\star}$ , Yuanjian Qiao $^{1}$ , and Wenjie Liu $^{1}$", + "bbox": [ + 218, + 212, + 782, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Computer Science and Technology, China University of Petroleum (East", + "bbox": [ + 227, + 239, + 774, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "China), Qingdao, 266580, China", + "bbox": [ + 393, + 253, + 609, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lzhmeng1688@163.com, smw278@126.com, yjqiao@s.upc.edu.cn,", + "bbox": [ + 279, + 268, + 718, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "liuwenjie_82@163.com", + "bbox": [ + 421, + 282, + 580, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In black-box attacks based on substitute training, the similarity of the substitute model to the target model is critical for successful attacks. However, existing schemes merely train the substitute model to mimic the outputs of the target model without fully simulating the decision space, resulting in the adversarial samples generated by the substitute model being classified into the non-target class by the target model. To alleviate this issue, we propose a novel Inter-Class Topology Alignment (ICTA) scheme to more comprehensively simulate the target model by aligning the inter-class positional relationships of two models in the decision space. Specifically, we first design the Position Exploration Sample (PES) to more thoroughly explore the relative positional relationships between classes in the decision space of the target model. Subsequently, we align the inter-class topology between the two models by utilizing the PES to constrain the inter-class relative position of the substitute model in different directions. In this way, the substitute model is more consistent with the target model in the decision space, so that the generated adversarial samples will be more successful in misleading the target model to classify them into the target class. The experimental results demonstrate that our ICTA significantly improves attack success rate in various scenarios compared to existing substitute training methods, particularly performing efficiently in target attacks.", + "bbox": [ + 261, + 333, + 743, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Black-box attacks $\\cdot$ Substitute attacks $\\cdot$ Position exploration sample $\\cdot$ Inter-class topology alignment", + "bbox": [ + 259, + 638, + 740, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 691, + 375, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Convolutional Neural Networks (CNNs) have shown impressive performance in various fields and are widely applied in daily life. However, it has been found that CNNs exhibit poor robustness to well-designed perturbations [3, 36, 38], which poses a great threat to their safety and reliability. For example, in the field of autonomous driving, if the network makes a wrong judgment, it will lead to potential safety hazards [13, 14, 26]. Therefore, it is especially urgent to", + "bbox": [ + 212, + 724, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 217, + 824, + 382, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bdcc90cce1ae60b4488e67de5b5fff51696c7e9b51cff2b913366763d0fc8a33.jpg", + "image_caption": [ + "Fig. 1: (a) Comparison of different methods and (b) Attack success rate. Traditional approaches only focus on intra-class consistency, resulting in the adversarial samples generated by the substitute model failing to mislead the target model to classify them into the target class. In contrast, our ICTA better simulates the target model by aligning the inter-class topology in the decision space, resulting in a successful target attack. The performance of our ICTA is significantly improved across multiple datasets, particularly in target attacks." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 781, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "research deeply into the adversarial attacks [4,30,38] and defense [17,22] schemes to enhance the reliability and security of CNNs in practical applications.", + "bbox": [ + 212, + 429, + 784, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing adversarial attacks can be categorized into white-box and black-box attacks according to the knowledge of the attacker about the target model. In white-box attacks [3-5,11,12,16,29], the attacker possesses complete information about the target model (such as network structure, gradient, etc.), thus can generate highly deceptive adversarial samples based on this information. By contrast, in black-box attacks [2,9,21,24,31,35], the attacker can only utilize the outputs of the network (probabilities or labels) to craft adversarial samples, making the task more challenging and realistic. In such scenario, the substitute attacks [9,27,33,35,37] is proposed, which focus on training a substitute model to emulate the target model, so that the adversarial samples generated by the substitute model can effectively attack the target model.", + "bbox": [ + 212, + 460, + 787, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current substitute attack methods [9, 21, 27, 28, 33, 35, 37] adhere to the paradigm of knowledge distillation, encouraging the substitute model to mimic the outputs of the target model. Some schemes utilized simple $L_{2}$ loss [21, 28, 33, 37] or KL-divergence [9] to constrain the substitute model. As an improvement, DST [27] leveraged a graph structure consisting of multiple outputs to optimize the substitute model. Nevertheless, these strategies only focus on learning the simplistic outputs of the target model, without adequately considering the decision space of the network. Subsequently, Zhang et al. [35] considered aligning the decision boundary between the substitute model and the target model to bridge the gap between the two models. However, this restriction of decision boundary only aims at the intra-class decision consistency and ignores the inter-class relationship consistency of the two models in the decision space. As a result, the adversarial samples generated by the substitute model are difficult to be classified as target class by the target model, as Fig. 1(a).", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "L. Meng et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To alleviate the above issue, we propose a novel Inter-Class Topology Alignment (ICTA) scheme to align the inter-class positional relationships of the two models in decision space, thus better simulating the target model and achieving more accurate target attacks. Specifically, we first design the Position Exploration Sample (PES) to explore the relative positional relationships among all classes of the target model by introducing adversarial perturbations of different directions and strengths. Subsequently, we constrain the inter-class positional consistency of the substitute model to align the inter-class topology between the two models. Thanks to the better alignment of the two models in the decision space, the adversarial samples generated by the substitute model can be more successful deceiving the target model to classify them into the target class, as shown in Fig. 1(a). Experiments demonstrate that our ICTA achieves higher attack success rate on different datasets, and especially more pronounced in target attacks, as Fig. 1(b). Our primary contributions are as follows:", + "bbox": [ + 212, + 146, + 787, + 358 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose ICTA to innovatively align the substitute model with the target model in terms of inter-class topology in the decision space, which effectively improves the success rate of black-box substitute attack.", + "- We propose a novel PES to more comprehensively explore the inter-class relative position of the target model, which is utilized to align the inter-class topology between the substitute model and the target model.", + "- Experiments illustrate that our ICTA achieves significant performance gains in a variety of scenarios and outperforms on target attacks compared to the existing substitute attack schemes." + ], + "bbox": [ + 223, + 367, + 782, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 525, + 395, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 White-box attacks", + "text_level": 1, + "bbox": [ + 215, + 556, + 413, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In white-box attacks, the attacker is fully aware of the information about the target model, such as the structure and parameters [5,11,12,16], and highly deceptive adversarial samples can be generated based on this information. Classical white-box adversarial sample generation schemes utilized gradient information of network to generate adversarial samples, such as FGSM [5], BIM [11], and PGD [16]. Moreover, Croce et al. [4] proposed parameter-free attack combination AutoAttack by integrating multiple attack methods for testing adversarial robustness. However, white-box attacks may lack practicality in real-world scenarios due to the requirement for extensive knowledge about the target model.", + "bbox": [ + 212, + 582, + 787, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Black-box attacks", + "text_level": 1, + "bbox": [ + 215, + 739, + 408, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, black-box attacks are more challenging and more in line with realistic scenarios, as the attacker can generate adversarial samples only using the outputs of the target model, including transfer attacks, query attacks, and substitute attacks. Transfer attacks [1, 3, 7, 29, 31] aim to maximize the transferability of existing adversarial samples, enabling them to successfully attack the target", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Topology Alignment for Substitute Attacks", + "bbox": [ + 442, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "model. For example, Zhu et al. [38] proposed modifying the distribution of the original images to match the distribution of the target class, thereby obtaining adversarial samples with high transferability. Query attacks [2, 24, 30] obtain adversarial samples by continuously querying the target model using specific inputs until a desired output is obtained. Substitute attacks [21, 35, 37] focus on training a substitute model that is similar to the target model, so that the adversarial samples generated by the substitute model can successfully attack the target model. Compared to other schemes, the substitute attack is able to generate more effective adversarial samples due to the ability to simulate the decision process of the target model.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Most of the existing substitute attack schemes used $L_{2}$ loss [21, 28, 33, 37] or KL-divergence [9] as the constraint to train substitute models. For example, DaST [37] used Generative Adversarial Networks (GANs) to generate data for substitute training and leveraged the $L_{2}$ loss as a constraint to align the outputs of the substitute model with the target model. Kariyappa et al. [9] optimized the generator with the help of zeroth-order gradient estimate and trained the substitute training model with a simple KL-divergence. Shao et al. [21] proposed LCA to utilize a diffusion model to generate the data required for substitute training, which greatly improved the training efficiency compared to traditional GANs-based methods. However, the target attack success rate is not impressive as it only used $L_{2}$ loss to constrain the substitute model. In order to simulate the target model more precisely, Wang et al. [27] proposed DST to optimize the substitute model from a graph with multiple outputs of the target model instead of a single output. Nevertheless, these schemes only learn the simple outputs of the target model without fully considering the complex decision space attributes, resulting in poor performance. Subsequently, Zhang et al. [35] encouraged a high alignment of decision boundary between the substitute model and the target model, which improved the attack success rate.", + "bbox": [ + 212, + 297, + 787, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the aforementioned schemes only focused on the intra-class consistency and neglected the inter-class relational consistency, resulting in the adversarial samples generated by the substitute model being difficult to be classified as target class by the target model. In this paper, we further consider the inter-class positional relationships in the decision space and better simulate the target model by aligning the inter-class topology for a more effective target attack.", + "bbox": [ + 212, + 571, + 787, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 690, + 380, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 727, + 362, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a black-box target model $T$ , existing research on substitute attacks typically trains a substitute model $S$ similar to the $T$ through a knowledge distillation paradigm. Formally, the substitute model can be obtained by substitute training:", + "bbox": [ + 212, + 757, + 785, + 805 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\theta} {\\arg \\min } \\mathbb {E} _ {x \\sim \\mathcal {X}} d [ S (x), T (x) ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 821, + 785, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "L. Meng et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d$ denotes the distance metric function, $x \\sim \\mathcal{X}$ denotes the training data, and $\\theta$ is the parameters of $S$ . The goal of substitute training is to optimize the substitute model to resemble the target model as similar as possible.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, the adversarial samples can be generated using white-box adversarial sample generation methods based on the substitute model. We take BIM [11] as an instance, which perturbs the input data along the gradient direction of the target class through multiple iterations to make the output of the model close to the target class in the decision space, as Fig. 2. The generation of adversarial samples for target attacks can be represented as follows:", + "bbox": [ + 212, + 191, + 442, + 433 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/804353b81756ac88ffbd8ccde680b1e847d06b4ca4de33b509c58dd0a63b686f.jpg", + "image_caption": [ + "Fig. 2: Position exploration sample vs. Traditional adversarial sample. Traditional adversarial samples are closer to and clustered at the decision center after several iterations of minor adjustments. While our PES has a much broader coverage and is able to explore the decision space more comprehensively." + ], + "image_footnote": [], + "bbox": [ + 460, + 222, + 782, + 315 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {0} ^ {a d v} = x, x _ {N + 1} ^ {a d v} = C l i p _ {x, \\epsilon} \\left\\{x _ {N} ^ {a d v} - \\alpha s i g n \\left(\\nabla_ {x} L \\left(S \\left(x _ {N + 1} ^ {a d v}\\right), y _ {t a r}\\right)\\right) \\right\\}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 271, + 444, + 784, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y_{tar}$ is the target class, $\\nabla_x$ denotes the gradient of the substitute model, $sign$ denotes the sign function, $\\alpha$ represents the step size during the iteration process, and $Clip$ is the clip function based on the perturbation boundary $\\epsilon$ .", + "bbox": [ + 212, + 474, + 782, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the generated adversarial samples are used to attack the target model. Clearly, the similarity between the substitute model and the target model is a key factor for successful attacks in substitute attacks.", + "bbox": [ + 212, + 518, + 782, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Aligning the Inter-Class Topology in the Decision Space", + "text_level": 1, + "bbox": [ + 214, + 587, + 723, + 603 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing substitute training methods encourage the substitute model to simply mimic the outputs of the target model without fully simulating the decision space of the target model. Resulting in the adversarial samples generated by the substitute model are difficult to be classified into the target class by the target model. Moreover, there is also a lack of research on what aspects the substitute model needs to resemble the target model. In this paper, we propose to understand and improve the substitute training from the perspective of the inter-class topology in the decision space, which refers to the positional relationships (including directions and distances) among all classes in the decision space.", + "bbox": [ + 212, + 613, + 784, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "According to Eq. (2), modifying the adversarial samples along the direction of the gradient $\\nabla_{x}$ in the decision space can bring the decision of the model closer to the target class, as shown in Fig. 2. Therefore, if the relative positions of all classes are consistent between the substitute model and the target model, then the adversarial samples generated by the substitute model can be successfully classified into the target class by the target model. This suggests that aligning", + "bbox": [ + 212, + 750, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Topology Alignment for Substitute Attacks", + "bbox": [ + 442, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fe668db758a260ad7998573604edf96c783bf422356bc2f6f5c918caeb60661f.jpg", + "image_caption": [ + "Fig. 3: The framework of our Inter-Class Topology Alignment (ICTA). ICTA explores the inter-class relative position in the decision space using position exploration sample in different directions, and aligns the inter-class topology of the two models by fitting the inter-class relative position." + ], + "image_footnote": [], + "bbox": [ + 228, + 143, + 777, + 291 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the inter-class topology between the two models in the decision space is crucial for substitute attacks.", + "bbox": [ + 212, + 390, + 787, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the sample $x_{i}$ , we consider $S(x_{i})$ and $T(x_{i})$ as nodes in the two decision spaces, and $e_{S_{ij}} = \\langle S(x_i), S(x_j) \\rangle$ , $i \\neq j$ and $e_{T_{ij}} = \\langle T(x_i), T(x_j) \\rangle$ , $i \\neq j$ to represent the edges of the topology in the two decision spaces. Topology consistency can be achieved by aligning all nodes and edges:", + "bbox": [ + 212, + 421, + 787, + 481 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min \\sum_ {i = 1} ^ {n} \\mathcal {D} _ {p} \\left(S \\left(x _ {i}\\right), T \\left(x _ {i}\\right)\\right) + \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {n} \\mathcal {D} _ {e} \\left(e _ {S _ {i j}}, e _ {T _ {i j}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 325, + 493, + 785, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_p$ and $\\mathcal{D}_e$ denote the distance metrics of points and edges, respectively.", + "bbox": [ + 212, + 545, + 782, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, aligning the edges among all the samples in the decision space is extremely complex. Therefore, we propose a relative position exploration and alignment strategy for matching the inter-class topology of the two models.", + "bbox": [ + 212, + 561, + 787, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Position Exploration Sample", + "text_level": 1, + "bbox": [ + 214, + 630, + 500, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "According to Fig. 3, the relative position of the clean and adversarial samples in the decision space can similarly be used to represent the relative position of the two classes in the decision space. In this case, the adversarial perturbations used to generate the adversarial samples can reflect the distance and direction between the two classes. Therefore, under the assumption that the target model is white-boxed, the inter-class relative position can be estimated using the positional relationships between the clean and adversarial samples.", + "bbox": [ + 212, + 657, + 787, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The traditional adversarial samples aim to progressively bring the sample closer to the decision center of the target class through multiple iterations of minor adjustments. While the sample has a relatively high success rate in the attack, it does not benefit for thoroughly exploring the entire decision space. To alleviate this limitation, we design the Position Exploration Sample (PES)", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "L. Meng et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to more comprehensively explore the inter-class relative position of the target model in the decision space, which can be represented as:", + "bbox": [ + 214, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nx _ {p e s} = x - \\phi \\nabla_ {x} L (T (x), y _ {t a r}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 188, + 785, + 204 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\phi$ is the perturbation strength. We generate the PES using the full gradient, not just the gradient direction, without iteration and without being constrained by the pixel range. Thus, the gradient $\\nabla_{x}$ and the perturbation strength $\\phi$ determine its direction and distance relative to the clean sample in the decision space. We represent the inter-class relative position based on the relative position of the PES to the clean sample and explore the decision space more thoroughly over a wider coverage by controlling $y_{tar}$ and $\\phi$ , as Fig. 2.", + "bbox": [ + 212, + 213, + 782, + 319 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "However, it is impossible to generate the PES directly due to the gradient of the target model is inaccessible. Therefore, we retreat to generating PES using the substitute model, as shown in Fig. 3, which can be represented as:", + "bbox": [ + 212, + 319, + 784, + 364 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nx _ {p e s} = x - \\phi \\nabla_ {x} L (S (x), y _ {a r b}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 376, + 785, + 392 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $y_{arb}$ represents an arbitrary class different from ground truth. In this way, we can explore the decision space from different directions and obtain the complete inter-class topology in the decision space. Moreover, the PES generated by the substitute model has an effect on exploring the decision space of the target model, since our goal is to make the two models as similar as possible.", + "bbox": [ + 212, + 401, + 782, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Topology Alignment Based on PES", + "text_level": 1, + "bbox": [ + 214, + 497, + 553, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In order to align the inter-class topology of the two models, we simplify the topology consistency in Eq. (3) to the consistency of the inter-class relative position. This method allows us to capture the inter-class topology through the position of two nodes without the need to optimize the edges directly, as demonstrated in Fig. 3. We can achieve this alignment by simultaneously optimizing the knowledge distillation loss $L_{KD}$ and the relative position loss $L_{RP}$ :", + "bbox": [ + 212, + 521, + 782, + 612 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {K D} = \\mathcal {D} _ {k d} (T (x), S (x)), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 623, + 784, + 638 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {R P} = \\mathcal {D} _ {r p} \\left(T \\left(x _ {p e s}\\right), S \\left(x _ {p e s}\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 642, + 784, + 657 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {T} = \\lambda_ {1} L _ {K D} + \\lambda_ {2} L _ {R P}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 661, + 784, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_{kd}$ and $\\mathcal{D}_{rp}$ are distance measure formulas, $\\lambda_{1}$ and $\\lambda_{2}$ are hyperparameters. The inter-class relations of the two models can be aligned by ensuring intra-class consistency through $L_{KD}$ and inter-class consistency through $L_{RP}$ .", + "bbox": [ + 212, + 686, + 782, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, we can constrain the substitute model to be inter-class consistent with the target model in different directions and distances based on the PES of a clean sample in different directions and at different distances. At this point, the loss of topology consistency can be expressed as:", + "bbox": [ + 212, + 733, + 782, + 792 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {T} = \\lambda_ {1} L _ {K D} + \\lambda_ {2} \\sum_ {i = 1} ^ {M} L _ {R P} ^ {i}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 801, + 785, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Topology Alignment for Substitute Attacks", + "bbox": [ + 441, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/60264219e6ad8edf2df369696aa6dacc898e90f0c439dc059db11c0a17f18ee1.jpg", + "table_caption": [ + "Table 1: The accuracy of target model. We select four models with different architectures as target models on each of the three datasets." + ], + "table_footnote": [], + "table_body": "
DatasetCIFAR10
Target Model AccuracyResNet34 94.53Vgg16 92.82GoogleNet 95.31ShuffleNetV2 94.53
DatasetCIFAR100
Target Model AccuracyResNet50 79.3MobileNetV2 71.61WideResNet 79.69ResNeXt50 77.56
DatasetTinyImageNet
Target Model AccuracyResNet50 62.81Vgg19 59.66MobileNetV2 62.81ResNeXt50 66.33
", + "bbox": [ + 305, + 184, + 697, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $M$ is the number of $x_{pes}$ generated for a same clean sample.", + "bbox": [ + 212, + 332, + 694, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition, incorporating the original categorical loss into this process is often beneficial, which is usually the cross-entropy loss $L_{CE}$ . The overall loss can be expressed as:", + "bbox": [ + 212, + 348, + 782, + 391 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {a l l} = L _ {T} + \\lambda_ {3} L _ {C E}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 393, + 784, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda_3$ is a hyperparameter. The parameters of the substitute model are optimized by back propagation to better simulate the target model.", + "bbox": [ + 212, + 412, + 782, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experimental", + "text_level": 1, + "bbox": [ + 214, + 465, + 383, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experimental Details", + "text_level": 1, + "bbox": [ + 214, + 494, + 437, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset and Target Model. We choose CIFAR10 [10] (with size of $32 \\times 32$ and 10 classes), CIFAR100 [10] (with size of $32 \\times 32$ and 100 classes), TinyImageNet [19] (with size of $64 \\times 64$ and 200 classes) as the training dataset. And we select VGG(16, 19) [23], ResNet(34, 50) [6], MobileNetV2 [20], GoogleNet [25], WideResNet [34], ShuffleNetV2 [15], and ResNeXt50 [32] as the target model. The accuracy of the target models is shown in Tab. 1.", + "bbox": [ + 212, + 517, + 784, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training Setup. During training, we set the perturbation strength in Eq. (5) to $\\phi = |Y|$ , where $Y$ obeys the Gaussian distribution $Y \\sim N(0,3)$ . The number of PES in Eq. (9) is set to $M = 1$ . The hyperparameters are empirically set to $\\lambda_1 = \\lambda_3 = 1, \\lambda_2 = 10$ . We standardize the query budget for substitute training across all baselines to ensure fairness, setting it to 2 million for CIFAR10 and 3 million for CIFAR100 and TinyImageNet.", + "bbox": [ + 212, + 628, + 784, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, we conduct experiments in two scenarios: full-data and data-free. In the data-free scenario, we utilize Stable Diffusion [18] as a generator and use the LCA [21] strategy to produce the data needed for substitute training. In the full-data scenario, the whole training data is available for substitute training. We perform this part of the experiment to validate the performance of the scheme without the influence of the generator. The algorithm of our ICTA and other experimental details are described in the supplement. The source code will be released upon acceptance of the paper.", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "L. Meng et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e383bb70d65f2234b6ac87be9b4f443c32988967539e39d01befb545d39c5ebd.jpg", + "table_caption": [ + "Table 2: Comparison of different white-box adversarial sample generation schemes. We choose ResNet34 on CIFAR10, ResNet50 on CIFAR100 and VGG19 in TinyImageNet as target models. The best results are marked as bolded." + ], + "table_footnote": [], + "table_body": "
Dataset MethodsCIFAR10CIFAR100TinyImageNet
FGSMBIMPGDFGSMBIMPGDFGSMBIMPGD
Non-TargetMAZE [9]38.6457.9868.9438.2041.5543.079.189.7911.10
DaST [37]35.1956.2164.2240.0142.5447.409.719.269.81
DST [27]40.8367.6571.2646.4549.2159.4611.6712.4733.52
Zhang et al. [35]60.1194.2197.5483.6994.5391.1838.5848.2165.01
DiffusionLCA [21]68.2995.7397.3082.0395.5896.4951.5552.3570.58
ICTA (Ours)75.6098.7199.1987.4299.0199.2583.7492.7393.39
TargetMAZE [9]6.8413.8316.656.2012.7810.010.502.502.54
DaST [37]7.8515.4118.615.0112.4811.410.203.025.20
DST [27]10.2616.1219.3310.2216.3724.651.883.055.87
Zhang et al. [35]28.1772.0564.057.7458.4554.782.3010.8815.01
DiffusionLCA [21]24.3176.9079.984.6145.1951.501.8412.9515.45
ICTA (Ours)29.4187.6391.407.8069.0274.496.8729.3133.41
", + "bbox": [ + 220, + 198, + 781, + 381 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Setup and Process. During testing, we choose three classical methods, including FGSM [5], BIM [11], and PGD [16], to generate adversarial samples, with parameters uniformly set to perturbation boundary $\\epsilon = 8 / 255$ and step size $\\alpha = 2 / 255$ . Our hardware utilizes an RTX 2080Ti GPU, and all schemes are retrained under the same environment.", + "bbox": [ + 215, + 409, + 784, + 484 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We train substitute model by our ICTA and then generate adversarial samples based on the substitute models using a white-box adversarial sample generation scheme. The success rate of these adversarial samples in attacking the target model is used as the final evaluation metric.", + "bbox": [ + 215, + 484, + 784, + 546 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Comparative Experiments in Data-free Scenario", + "text_level": 1, + "bbox": [ + 217, + 566, + 656, + 583 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Comparison of White-box Adversarial Sample Generation Schemes. We compare different white-box adversarial sample generation schemes, including FGSM, BIM, and PGD. As illustrated in Tab. 2, our ICTA realizes significant performance improvements across different white-box adversarial sample generation methods. On the one hand, our ICTA achieves more noticeable performance enhancement on large datasets TinyImageNet than on small datasets CIFAR10. On the other hand, ICTA has superior performance improvement on target attacks than non-target attacks. The results indicate that our ICTA more effectively simulates the decision space of the target model compared to traditional substitute training methods [9,21,27,35,37].", + "bbox": [ + 215, + 590, + 785, + 742 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Comparison on Different Target Models. According to the results in Tab. 3, our ICTA demonstrates a significant performance improvement on different target models compared to the GANs-based approaches [9,27,35,37] and the diffusion-based method [21]. Specifically, ICTA achieves the most superior performance in target attacks, especially on TinyImageNet. Additionally, ICTA", + "bbox": [ + 215, + 763, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Topology Alignment for Substitute Attacks", + "bbox": [ + 442, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/132387563955c009fc3d588063fcf533cb1de0886d98fc38c9164958fde89ea9.jpg", + "table_caption": [ + "Table 3: Comparison on different target models. We choose PGD as the default white-box adversarial sample generation scheme. The best results are marked as bolded." + ], + "table_footnote": [], + "table_body": "
Dataset Target ModelCIFAR10CIFAR100TinyImageNet
VGG16ResNet34GoogleNetMobileNetResNet50ResNet50VGG19
Non-TargetMAZE [9]47.4968.9427.1351.7343.0715.5411.10
DaST [37]53.2964.2226.4857.9447.4016.119.81
DST [27]55.1571.2638.6465.8559.4632.5133.52
Zhang et al. [35]95.8497.5467.2497.7891.1878.8265.01
DiffusionLCA [21]97.0197.3094.8297.0196.4967.5670.58
ICTA (Ours)96.4899.1997.3299.6399.2594.8093.39
TargetMAZE [9]11.3916.6512.9811.0310.013.182.54
DaST [37]15.3918.615.319.5811.412.215.20
DST [27]17.0719.3318.8918.4624.658.555.87
Zhang et al. [35]57.4764.0539.8344.1154.7820.0115.01
DiffusionLCA [21]75.3979.9878.1851.6151.5012.4515.45
ICTA (Ours)79.5791.4084.5770.5674.4943.0233.41
", + "bbox": [ + 220, + 184, + 782, + 356 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/0f9ccf569e335699319bf0abef7cce73584b7cb4c9756c1d41edf7bfe1fb4a96.jpg", + "table_caption": [ + "Table 4: Comparison of attack success rate in label-based and probability-based scenarios. The best results are marked as bolded." + ], + "table_footnote": [], + "table_body": "
methodProbability-basedLabel-based
cifar10cifar100cifar10cifar100
Non-TargetGANsDaST [37]64.2247.4067.7236.57
DST [27]71.2659.4680.4248.50
Zhang et al. [35]97.5491.1898.9890.20
DiffusionLCA [21]97.3096.4998.4894.55
ICTA (Ours)99.1999.2599.4998.53
TargetGANsDaST [37]18.6111.4119.285.20
DST [27]19.3324.6523.854.59
Zhang et al. [35]64.0554.7869.2540.17
DiffusionLCA [21]79.9851.5082.1849.85
ICTA (Ours)91.4074.4993.1061.71
", + "bbox": [ + 305, + 410, + 697, + 569 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "consistently enhances performance across different target models, showcasing the stability of the performance in various scenarios. In summary, our ICTA is able to simulate different target models in black-box attack scenarios and exhibits substantial performance improvements on large-scale datasets.", + "bbox": [ + 215, + 599, + 784, + 660 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparison in Probability-based and Label-based Scenarios. We compare the attack success rate in the probability-based and label-based scenarios, as displayed in Tab. 4. In the label-based scenario, only the output labels of the target model are accessible, while in the probability-based scenario, both the labels and probabilities of the target model are available. The experiment results indicate that our ICTA achieves the optimal performance in both probability-based and label-based scenarios and shows significant improvement over the existing substitute attack methods [21, 27, 35, 37]. This highlights the comprehensive superiority of our ICTA, demonstrating outstanding attack performance even in scenarios where only label outputs are accessible.", + "bbox": [ + 215, + 686, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "L. Meng et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2a75e89d1c9924e5f3372ea70f15115ca70d513532a468a743499effdd66d29f.jpg", + "image_caption": [ + "(a) Attack success rate" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 498, + 296 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5920324477c26a0d460724b5af97e45e144a5bdbf6f7ffb3736a08273d777f78.jpg", + "image_caption": [ + "(b) Accuracy of substitute model", + "Fig. 4: Comparison of attack success rate and accuracy in full-data scenario. Comparison schemes include the substitute training schemes DST [27] and Zhang et al. [35] as well as a distillation method DIST [8]. The ResNet50 on CIFAR100 is used as the target model." + ], + "image_footnote": [], + "bbox": [ + 504, + 145, + 785, + 297 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Comparative Experiments in Full-data Scenario", + "text_level": 1, + "bbox": [ + 214, + 409, + 653, + 425 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Performance of Different Substitute Training Strategies. In Fig. 4, we present the attack success rate (both target and non-target attacks) and the accuracy of the substitute training during the training process. It can be observed that our ICTA demonstrates excellent performance in both target and nontarget attacks and shows higher efficiency in target attacks. There is a significant improvement in non-target attacks compared to conventional methods DST [27] and Zhang et al. [35], which highlights the unique advantages of our ICTA in improving the effectiveness of target attacks. In addition, we compare an excellent knowledge distillation scheme, DIST [8], which also shows the superior performance of our ICTA in attack scenarios.", + "bbox": [ + 212, + 436, + 784, + 585 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Notably, we find that the accuracy of the substitute model does not fully reflect the effectiveness of the substitute training. Despite having a relatively low accuracy in our ICTA, the substitute model exhibits excellent performance in attacks. This indicates that our ICTA better simulates the target model in terms of inter-class topology, rather than just replicating its outputs.", + "bbox": [ + 212, + 588, + 784, + 664 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison of Relative Performance with White-box Attacks. We further compare the attack success rate of our ICTA with the corresponding white-box attacks, as demonstrated in Tab. 5, where the Relative Performance (RP) is used as a metric to measure the performance relative to the white-box attacks. Our ICTA achieves comparable performance to the corresponding white-box attacks in non-target attack, when the query budgets is set to 2M (for CIFAR10) and 3M (for CIFAR100 and TinyImageNet). Moreover, it is worth noting that the target attack success rate exceeds $83\\%$ on the CIFAR10 dataset, $67\\%$ on the CIFAR100 dataset, and $44\\%$ on the TinyImageNet dataset compared to the corresponding white-box attack. As can be seen, the proposed ICTA achieves quite", + "bbox": [ + 212, + 688, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Topology Alignment for Substitute Attacks", + "bbox": [ + 442, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ba839a0069a9da0f974cc6fd5db7be3f977fb2b69424e89941b901ccfe8f4955.jpg", + "table_caption": [ + "Table 5: Comparison with direct white-box attacks on target model. We utilize Relative Performance (RP) to evaluate the attack success of our black-box attack method relative to the white-box attack (in red)." + ], + "table_footnote": [], + "table_body": "
ResNet34Vgg16GoogleNetShuffleNetV2
white-boxOursRPwhite-boxOursRPwhite-boxOursRPwhite-boxOursRP
CIFAR10Non-TargetFGSM78.5483.421.0685.5979.460.9363.8477.731.2284.9783.970.99
BIM100.0099.591.00100.0097.260.9799.9397.610.98100.0098.860.99
PGD100.0099.791.00100.0098.080.98100.0098.260.98100.0099.200.99
TargetFGSM28.7934.691.2019.6630.251.5424.5526.991.1032.9331.780.97
BIM99.8494.580.9584.8181.190.9699.5582.580.83100.0086.010.86
PGD100.0096.630.9798.2185.670.8799.9786.290.86100.0089.030.89
ResNet50WideResNetMobileNetV2ResNeXt50
CIFAR100Non-TargetFGSM84.7786.761.0274.2783.401.1290.1888.200.9887.2487.511.00
BIM100.0099.220.9998.1196.830.9999.9899.090.9999.9899.160.99
PGD100.0099.500.9999.5797.540.98100.0099.460.99100.0099.470.99
TargetFGSM3.058.522.793.879.422.434.397.351.673.838.792.30
BIM95.2272.790.7689.4465.900.7496.3064.790.6797.0673.290.76
PGD99.2579.340.8098.3873.220.7499.6670.330.7199.7080.150.80
ResNet50Vgg19MobileNetV2ResNeXt50
TinyImageNetNon-TargetFGSM95.8286.070.9092.8286.740.9398.4288.000.8996.6283.650.87
BIM99.9796.260.9699.6495.940.96100.0097.210.9799.9894.780.95
PGD99.9896.960.9799.9196.560.97100.0097.780.9899.9895.860.96
TargetFGSM10.2510.040.984.797.991.6715.718.740.5613.188.680.66
BIM98.9652.300.5385.6639.310.4699.7245.920.4699.5343.560.44
PGD99.7457.850.5897.8146.310.4799.9251.180.5199.9650.300.50
", + "bbox": [ + 220, + 198, + 782, + 468 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "satisfactory performance in black-box attack scenarios, demonstrating robust adaptability in target attacks across different datasets.", + "bbox": [ + 212, + 494, + 784, + 523 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Analysis of Decision Boundary Similarity", + "text_level": 1, + "bbox": [ + 215, + 544, + 601, + 560 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The t-SNE Visualization of Decision Space. We", + "text_level": 1, + "bbox": [ + 212, + 568, + 429, + 595 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "exhibit the t-Distributed Stochastic Neighbor Embedding (t-SNE) visualization of the decision space of the models, as exhibited in Fig. 5. In t-SNE visualization, each data point is typically labeled with its corresponding class. Therefore, it can depict the interclass relative position in the decision space, as well as the inter-class topology. The visualization illustrates that the inter-class topology in the decision space of the substitute model trained by our ICTA", + "bbox": [ + 212, + 598, + 452, + 839 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/858835a3adecdc067bdc922f7096def4005dcf635c8895ad22e3101213641a91.jpg", + "image_caption": [ + "Fig. 5: The t-SNE visualization of the decision space for (a) the target model, (b) our ICTA, (c) Zhang et al. [35] and (d) DST [27]." + ], + "image_footnote": [], + "bbox": [ + 446, + 599, + 535, + 787 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f555365d8168529095cbbd02daa21736f37dbab9611910fc00331bb0728ad1e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 599, + 617, + 789 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/dde4a897cb1c2ab0fc82ec9083b1f4c263a684303c331ef937b98f483f520ad4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 599, + 699, + 789 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4d2c5c7420637432cbb064d5533859f023b02a8944afe5a2ebc0f095488de350.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 599, + 782, + 789 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "L. Meng et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2dbc627c5a13e3cd73661b1b7ae055e3d1bdd9f361a0a3d5b9e81f17580a9931.jpg", + "image_caption": [ + "(a) Attack success rate" + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 498, + 291 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/43060eba2b4b6e0c13b209fddffb14ee2b1b88839822769456a3a46907a24a47.jpg", + "image_caption": [ + "(b) Accuracy of substitute model", + "Fig. 6: Ablation experiments on PES. The experiments include baseline (no relative position loss) and strategies for exploring inter-class relative position through traditional adversarial samples, clean samples, and position exploration sample." + ], + "image_footnote": [], + "bbox": [ + 504, + 146, + 785, + 290 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "method is remarkably similar to the topology of the target model. On the contrary, both the output alignment-focused DST [27] and the decision boundary-constrained approach by Zhang et al. [35] exhibit greater differences from the target model in t-SNE visualization. This confirms that our ICTA is more effective in aligning the substitute model to match the target model.", + "bbox": [ + 212, + 388, + 784, + 465 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Quantitative Analysis. We explore random directions of arbitrary samples, and then quantitatively assess the Decision Boundary Similarity (DBS) based on the overlap between target and substitute models. As in Tab. 6, our ICTA has more consistent decision boundaries with the target model.", + "bbox": [ + 214, + 488, + 486, + 623 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/770909530de46121111cac7f2fee19b55722b528e15be1d92ee3b7aa16668309.jpg", + "table_caption": [ + "Table 6: Quantitative analysis of decision boundary similarity." + ], + "table_footnote": [], + "table_body": "
MethodICTA (Ours)Zhang et al.DST
CIFAR100.930.870.80
CIFAR1000.880.770.71
TinyImageNet0.820.570.55
", + "bbox": [ + 503, + 561, + 785, + 612 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 647, + 387, + 662 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation on Position Exploration Sample. In this paper, we design PES to explore the inter-class relative position in the decision space through their relative position relationships with respect to the original clean samples. However, as the position of any two samples in the decision space inherently depict the inter-class relative position, traditional adversarial sample or even a clean sample has the effect of exploring the decision space. In this section, we validate the effectiveness of PES by comparing it with traditional adversarial sample generated by BIM and clean sample, as illustrated in Fig. 6.", + "bbox": [ + 212, + 672, + 784, + 792 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "It is worth noting that while all three approaches contribute to explore the inter-class relative position in the decision space, PES stands out as particularly advantageous. This is due to the fact that both clean samples and the adversarial", + "bbox": [ + 212, + 795, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Topology Alignment for Substitute Attacks", + "bbox": [ + 442, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e9fc07eb7c6ff812749579ff02bb0272bd49885bd91b52daa61c9453af8c266f.jpg", + "table_caption": [ + "Table 7: Ablation comparison of the loss function and the distance metric. We select two distance metrics, MSE and DIST [8], and use different combinations of distance metrics in the loss function. The best result are marked as bolded, and the sub-optimal results are marked as underlined." + ], + "table_footnote": [], + "table_body": "
DkdMSE DISTDrpMSE DISTTargetNon-Target
FGSMBIMPGDFGSMBIMPGD
✓4.3142.1446.1479.4895.3796.38
✓4.1155.5257.6782.8598.3098.80
✓✓7.6464.2071.7084.3598.4399.06
✓✓7.8273.4279.0986.7399.0499.61
✓✓8.5272.7979.3486.7699.2299.50
", + "bbox": [ + 305, + 212, + 697, + 305 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "samples generated by BIM are closer to the decision center, resulting in limited exploration in the decision space. In contrast, the proposed PES is more diverse by employing the full-gradient and unconstrained setting, enabling exploration across a wider range of the decision space. The ablation experiments for the parameters of PES and the number of PES are provided in the supplement.", + "bbox": [ + 212, + 334, + 787, + 411 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ablation on Loss and Metrics. We conduct an ablation study on the loss functions and distance metrics of substitute training in Tab. 7, focusing on the distillation loss $L_{KD}$ and relative position loss $L_{RP}$ . Here, the term 'DIST' refers to the distance metric function in the distillation scheme DIST [8]. It can be seen that when DIST [8] is used as the distance metric in $L_{KD}$ , there is a certain improvement in the attack success rate. And when $L_{RP}$ is added, there is a significant improvement in the attack success rate, which underscores the indispensability of the relative position loss. Notably, our ICTA employs DIST as the distance metric in both losses, as it further improves attack performance.", + "bbox": [ + 212, + 433, + 787, + 570 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 595, + 359, + 611 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose a novel ICTA scheme to better align the substitute model to the target model for efficient black-box substitute attacks. Compared to the existing schemes that only align intra-class consistency, we simulate the target model in the decision space more comprehensively through inter-class topology alignment. Specifically, we first explore the inter-class relative position of the target model in the decision space by the designed PES. Subsequently, we achieve inter-class topology alignment of the two models by constraining the inter-class relative position of the substitute model in different directions. Benefiting from the better alignment of the two models in the decision space, the adversarial samples generated by the substitute model are more successful in misleading the target model to classify them into the target class. Experimental results demonstrate the effectiveness of ICTA, which achieves better simulation of the target model and significantly improves the attack success rate beyond existing substitute attack approaches.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "L. Meng et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the National Key Research and Development Program of China (Grant No. 2021YFA1000102), the National Natural Science Foundation of China (Grant Nos. 62376285, 62272375, and 61673396), and Natural Science Foundation of Shandong Province, China (Grant No. ZR2022MF260).", + "bbox": [ + 212, + 176, + 787, + 236 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 258, + 321, + 273 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Andriushchenko, M., Croce, F., Flammarion, N., Hein, M.: Square Attack: a query-efficient black-box adversarial attack via random search. In: Proceedings of the European Conference on Computer Vision. pp. 484-501 (2020)", + "2. Brendel, W., Rauber, J., Bethge, M.: Decision-Based Adversarial Attacks: Reliable Attacks Against Black-Box Machine Learning Models. In: International Conference on Learning Representations. arXiv preprint arXiv:1712.04248 (2018)", + "3. Chen, S., He, Z., Sun, C., Yang, J., Huang, X.: Universal Adversarial Attack on Attention and the Resulting Dataset DImageNet. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(4), 2188-2197 (2022)", + "4. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: International Conference on Machine Learning. vol. 119, pp. 2206-2216 (2020)", + "5. Goodfellow, I., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014)", + "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)", + "7. Huang, H., Chen, Z., Chen, H., Wang, Y., Zhang, K.: T-SEA: Transfer-based Self-Ensemble Attack on Object Detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20514-20523 (2023)", + "8. Huang, T., You, S., Wang, F., Qian, C., Xu, C.: Knowledge distillation from a stronger teacher. In: Advances in Neural Information Processing Systems. vol. 35, pp. 33716-33727 (2022)", + "9. Kariyappa, S., Prakash, A., Qureshi, M.K.: Maze: Data-free model stealing attack using zeroth-order gradient estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13814-13823 (2021)", + "0. Krizhevsky, A.: Learning Multiple Layers of Features from Tiny Images. Master's thesis, University of Tront (2009)", + "1. Kurakin, A., Goodfellow, I., Bengio, S.: Adversarial examples in the physical world. arXiv preprint arXiv:1607.02533 (2016)", + "2. Lee, B.K., Kim, J.: Mitigating Adversarial Vulnerability through Causal Parameter Estimation by Adversarial Double Machine Learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4499-4509 (2023)", + "3. Li, L., Lian, Q., Chen, Y.C.: Adv3D: Generating 3D Adversarial Examples in Driving Scenarios with NeRF. arXiv preprint arXiv:2309.01351 (2023)", + "4. Liu, J., Lu, B., Xiong, M., Zhang, T., Xiong, H.: Adversarial Attack with Raindrops. arXiv preprint arXiv:2302.14267 (2023)", + "5. Ma, N., Zhang, X., Zheng, H.T., Sun, J.: ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design. In: Proceedings of the European Conference on Computer Vision. pp. 122-138 (2018)" + ], + "bbox": [ + 225, + 287, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Topology Alignment for Substitute Attacks", + "bbox": [ + 442, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2018)", + "17. Ren, K., Zheng, T., Qin, Z., Liu, X.: Adversarial Attacks and Defenses in Deep Learning. Engineering 6(3), 346-360 (2020)", + "18. Robin, R., Andreas, B., Dominik, L., Patrick, E., Bj\\''orn, O.: High-Resolution Image Synthesis with Latent Diffusion Models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10674-10685 (2022)", + "19. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision 115(3), 211-252 (2015)", + "20. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: MobileNetV2: Inverted Residuals and Linear Bottlenecks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4510-4520 (2018)", + "21. Shao, M., Meng, L., Qiao, Y., Zhang, L., Zuo, W.: Data-free black-box attack based on diffusion model. arXiv preprint arXiv:2307.12872 (2023)", + "22. Sheng, L., Liang, J., He, R., Wang, Z., Tan, T.: AdaptGuard: Defending Against Universal Attacks for Model Adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19093-19103 (2023)", + "23. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)", + "24. Su, J., Vargas, D.V., Sakurai, K.: One Pixel Attack for Fooling Deep Neural Networks. IEEE Transactions on Evolutionary Computation 23(5), 828-841 (2019)", + "25. Szegedy, C., Wei Liu, Yangqing Jia, Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1-9 (2015)", + "26. Wang, N., Luo, Y.: Does Physical Adversarial Example Really Matter to Autonomous Driving? Towards System-Level Effect of Adversarial Object Evasion Attack. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4412-4423 (2023)", + "27. Wang, W., Qian, X., Fu, Y., Xue, X.: Dst: Dynamic substitute training for data-free black-box attack. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14361-14370 (2022)", + "28. Wang, W., Yin, B., Yao, T., Zhang, L., Fu, Y., Ding, S., Li, J., Huang, F., Xue, X.: Delving into data: Effectively substitute training for black-box attack. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4761-4770 (2021)", + "29. Wang, X., Zhang, Z.: Structure Invariant Transformation for better Adversarial Transferability. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4607-4619 (2023)", + "30. Wang, X., Zhang, Z., Tong, K., Gong, D., He, K., Li, Z., Liu, W.: Triangle Attack: A Query-efficient Decision-based Adversarial Attack. In: Proceedings of the European Conference on Computer Vision. pp. 156-174 (2022)", + "31. Wang, Z., Yang, H., Feng, Y., Sun, P., Guo, H., Zhang, Z., Ren, K.: Towards Transferable Targeted Adversarial Examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)", + "32. Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K.: Aggregated Residual Transformations for Deep Neural Networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5987-5995 (2017)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "L. Meng et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Yu, M., Sun, S.: FE-DaST: Fast and effective data-free substitute training for black-box adversarial attacks. Computers & Security 113, 102555 (2022)", + "34. Zagoruyko, S., Komodakis, N.: Wide Residual Networks. In: Proceedings of the British Machine Vision Conference. pp. 87.1-87.12 (2016)", + "35. Zhang, J., Li, B., Xu, J., Wu, S., Ding, S., Zhang, L., Wu, C.: Towards efficient data free black-box adversarial attack. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15094-15104 (2022)", + "36. Zheng, X., Fan, Y., Wu, B., Zhang, Y., Wang, J., Pan, S.: Robust Physical-World Attacks on Face Recognition. Pattern Recognition 133, 109009 (2023)", + "37. Zhou, M., Wu, J., Liu, Y., Liu, S., Zhu, C.: Dast: Data-free substitute training for adversarial attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 234-243 (2020)", + "38. Zhu, Y., Chen, Y., Li, X., Chen, K., He, Y., Tian, X., Zheng, B., Chen, Y., Huang, Q.: Toward Understanding and Boosting Adversarial Transferability From a Distribution Perspective. IEEE Transactions on Image Processing 31, 6487-6501 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 356 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Topology Alignment for Substitute Attacks", + "bbox": [ + 442, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks/5bfa27b2-acad-4e1e-b905-e3f3b1d14c81_model.json b/2024/Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks/5bfa27b2-acad-4e1e-b905-e3f3b1d14c81_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8264476f4855e6e2333d9962de51ac7cb902e154 --- /dev/null +++ b/2024/Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks/5bfa27b2-acad-4e1e-b905-e3f3b1d14c81_model.json @@ -0,0 +1,2291 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.249, + 0.141, + 0.756, + 0.187 + ], + "angle": 0, + "content": "Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.213, + 0.784, + 0.229 + ], + "angle": 0, + "content": "Lingzhuang Meng\\(^{1}\\), Mingwen Shao\\(^{1\\star}\\), Yuanjian Qiao\\(^{1}\\), and Wenjie Liu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.24, + 0.776, + 0.254 + ], + "angle": 0, + "content": "School of Computer Science and Technology, China University of Petroleum (East" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.255, + 0.61, + 0.268 + ], + "angle": 0, + "content": "China), Qingdao, 266580, China" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.269, + 0.72, + 0.282 + ], + "angle": 0, + "content": "lzhmeng1688@163.com, smw278@126.com, yjqiao@s.upc.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.423, + 0.284, + 0.581, + 0.296 + ], + "angle": 0, + "content": "liuwenjie_82@163.com" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.334, + 0.744, + 0.626 + ], + "angle": 0, + "content": "Abstract. In black-box attacks based on substitute training, the similarity of the substitute model to the target model is critical for successful attacks. However, existing schemes merely train the substitute model to mimic the outputs of the target model without fully simulating the decision space, resulting in the adversarial samples generated by the substitute model being classified into the non-target class by the target model. To alleviate this issue, we propose a novel Inter-Class Topology Alignment (ICTA) scheme to more comprehensively simulate the target model by aligning the inter-class positional relationships of two models in the decision space. Specifically, we first design the Position Exploration Sample (PES) to more thoroughly explore the relative positional relationships between classes in the decision space of the target model. Subsequently, we align the inter-class topology between the two models by utilizing the PES to constrain the inter-class relative position of the substitute model in different directions. In this way, the substitute model is more consistent with the target model in the decision space, so that the generated adversarial samples will be more successful in misleading the target model to classify them into the target class. The experimental results demonstrate that our ICTA significantly improves attack success rate in various scenarios compared to existing substitute training methods, particularly performing efficiently in target attacks." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.639, + 0.741, + 0.668 + ], + "angle": 0, + "content": "Keywords: Black-box attacks \\(\\cdot\\) Substitute attacks \\(\\cdot\\) Position exploration sample \\(\\cdot\\) Inter-class topology alignment" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.693, + 0.377, + 0.709 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.725, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Convolutional Neural Networks (CNNs) have shown impressive performance in various fields and are widely applied in daily life. However, it has been found that CNNs exhibit poor robustness to well-designed perturbations [3, 36, 38], which poses a great threat to their safety and reliability. For example, in the field of autonomous driving, if the network makes a wrong judgment, it will lead to potential safety hazards [13, 14, 26]. Therefore, it is especially urgent to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.383, + 0.841 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "L. Meng et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.782, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.301, + 0.788, + 0.399 + ], + "angle": 0, + "content": "Fig. 1: (a) Comparison of different methods and (b) Attack success rate. Traditional approaches only focus on intra-class consistency, resulting in the adversarial samples generated by the substitute model failing to mislead the target model to classify them into the target class. In contrast, our ICTA better simulates the target model by aligning the inter-class topology in the decision space, resulting in a successful target attack. The performance of our ICTA is significantly improved across multiple datasets, particularly in target attacks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.785, + 0.461 + ], + "angle": 0, + "content": "research deeply into the adversarial attacks [4,30,38] and defense [17,22] schemes to enhance the reliability and security of CNNs in practical applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.627 + ], + "angle": 0, + "content": "Existing adversarial attacks can be categorized into white-box and black-box attacks according to the knowledge of the attacker about the target model. In white-box attacks [3-5,11,12,16,29], the attacker possesses complete information about the target model (such as network structure, gradient, etc.), thus can generate highly deceptive adversarial samples based on this information. By contrast, in black-box attacks [2,9,21,24,31,35], the attacker can only utilize the outputs of the network (probabilities or labels) to craft adversarial samples, making the task more challenging and realistic. In such scenario, the substitute attacks [9,27,33,35,37] is proposed, which focus on training a substitute model to emulate the target model, so that the adversarial samples generated by the substitute model can effectively attack the target model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Current substitute attack methods [9, 21, 27, 28, 33, 35, 37] adhere to the paradigm of knowledge distillation, encouraging the substitute model to mimic the outputs of the target model. Some schemes utilized simple \\( L_{2} \\) loss [21, 28, 33, 37] or KL-divergence [9] to constrain the substitute model. As an improvement, DST [27] leveraged a graph structure consisting of multiple outputs to optimize the substitute model. Nevertheless, these strategies only focus on learning the simplistic outputs of the target model, without adequately considering the decision space of the network. Subsequently, Zhang et al. [35] considered aligning the decision boundary between the substitute model and the target model to bridge the gap between the two models. However, this restriction of decision boundary only aims at the intra-class decision consistency and ignores the inter-class relationship consistency of the two models in the decision space. As a result, the adversarial samples generated by the substitute model are difficult to be classified as target class by the target model, as Fig. 1(a)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.443, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Topology Alignment for Substitute Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.359 + ], + "angle": 0, + "content": "To alleviate the above issue, we propose a novel Inter-Class Topology Alignment (ICTA) scheme to align the inter-class positional relationships of the two models in decision space, thus better simulating the target model and achieving more accurate target attacks. Specifically, we first design the Position Exploration Sample (PES) to explore the relative positional relationships among all classes of the target model by introducing adversarial perturbations of different directions and strengths. Subsequently, we constrain the inter-class positional consistency of the substitute model to align the inter-class topology between the two models. Thanks to the better alignment of the two models in the decision space, the adversarial samples generated by the substitute model can be more successful deceiving the target model to classify them into the target class, as shown in Fig. 1(a). Experiments demonstrate that our ICTA achieves higher attack success rate on different datasets, and especially more pronounced in target attacks, as Fig. 1(b). Our primary contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.368, + 0.784, + 0.41 + ], + "angle": 0, + "content": "- We propose ICTA to innovatively align the substitute model with the target model in terms of inter-class topology in the decision space, which effectively improves the success rate of black-box substitute attack." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.413, + 0.784, + 0.455 + ], + "angle": 0, + "content": "- We propose a novel PES to more comprehensively explore the inter-class relative position of the target model, which is utilized to align the inter-class topology between the substitute model and the target model." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.459, + 0.784, + 0.5 + ], + "angle": 0, + "content": "- Experiments illustrate that our ICTA achieves significant performance gains in a variety of scenarios and outperforms on target attacks compared to the existing substitute attack schemes." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.368, + 0.784, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.526, + 0.396, + 0.542 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.558, + 0.414, + 0.572 + ], + "angle": 0, + "content": "2.1 White-box attacks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.719 + ], + "angle": 0, + "content": "In white-box attacks, the attacker is fully aware of the information about the target model, such as the structure and parameters [5,11,12,16], and highly deceptive adversarial samples can be generated based on this information. Classical white-box adversarial sample generation schemes utilized gradient information of network to generate adversarial samples, such as FGSM [5], BIM [11], and PGD [16]. Moreover, Croce et al. [4] proposed parameter-free attack combination AutoAttack by integrating multiple attack methods for testing adversarial robustness. However, white-box attacks may lack practicality in real-world scenarios due to the requirement for extensive knowledge about the target model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.74, + 0.409, + 0.754 + ], + "angle": 0, + "content": "2.2 Black-box attacks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In contrast, black-box attacks are more challenging and more in line with realistic scenarios, as the attacker can generate adversarial samples only using the outputs of the target model, including transfer attacks, query attacks, and substitute attacks. Transfer attacks [1, 3, 7, 29, 31] aim to maximize the transferability of existing adversarial samples, enabling them to successfully attack the target" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "L. Meng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "model. For example, Zhu et al. [38] proposed modifying the distribution of the original images to match the distribution of the target class, thereby obtaining adversarial samples with high transferability. Query attacks [2, 24, 30] obtain adversarial samples by continuously querying the target model using specific inputs until a desired output is obtained. Substitute attacks [21, 35, 37] focus on training a substitute model that is similar to the target model, so that the adversarial samples generated by the substitute model can successfully attack the target model. Compared to other schemes, the substitute attack is able to generate more effective adversarial samples due to the ability to simulate the decision process of the target model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.788, + 0.57 + ], + "angle": 0, + "content": "Most of the existing substitute attack schemes used \\( L_{2} \\) loss [21, 28, 33, 37] or KL-divergence [9] as the constraint to train substitute models. For example, DaST [37] used Generative Adversarial Networks (GANs) to generate data for substitute training and leveraged the \\( L_{2} \\) loss as a constraint to align the outputs of the substitute model with the target model. Kariyappa et al. [9] optimized the generator with the help of zeroth-order gradient estimate and trained the substitute training model with a simple KL-divergence. Shao et al. [21] proposed LCA to utilize a diffusion model to generate the data required for substitute training, which greatly improved the training efficiency compared to traditional GANs-based methods. However, the target attack success rate is not impressive as it only used \\( L_{2} \\) loss to constrain the substitute model. In order to simulate the target model more precisely, Wang et al. [27] proposed DST to optimize the substitute model from a graph with multiple outputs of the target model instead of a single output. Nevertheless, these schemes only learn the simple outputs of the target model without fully considering the complex decision space attributes, resulting in poor performance. Subsequently, Zhang et al. [35] encouraged a high alignment of decision boundary between the substitute model and the target model, which improved the attack success rate." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.788, + 0.664 + ], + "angle": 0, + "content": "However, the aforementioned schemes only focused on the intra-class consistency and neglected the inter-class relational consistency, resulting in the adversarial samples generated by the substitute model being difficult to be classified as target class by the target model. In this paper, we further consider the inter-class positional relationships in the decision space and better simulate the target model by aligning the inter-class topology for a more effective target attack." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.691, + 0.381, + 0.71 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.728, + 0.363, + 0.745 + ], + "angle": 0, + "content": "3.1 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.758, + 0.787, + 0.806 + ], + "angle": 0, + "content": "For a black-box target model \\( T \\), existing research on substitute attacks typically trains a substitute model \\( S \\) similar to the \\( T \\) through a knowledge distillation paradigm. Formally, the substitute model can be obtained by substitute training:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.822, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\underset {\\theta} {\\arg \\min } \\mathbb {E} _ {x \\sim \\mathcal {X}} d [ S (x), T (x) ], \\tag {1}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.443, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Topology Alignment for Substitute Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "where \\(d\\) denotes the distance metric function, \\(x \\sim \\mathcal{X}\\) denotes the training data, and \\(\\theta\\) is the parameters of \\(S\\). The goal of substitute training is to optimize the substitute model to resemble the target model as similar as possible." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.443, + 0.434 + ], + "angle": 0, + "content": "Subsequently, the adversarial samples can be generated using white-box adversarial sample generation methods based on the substitute model. We take BIM [11] as an instance, which perturbs the input data along the gradient direction of the target class through multiple iterations to make the output of the model close to the target class in the decision space, as Fig. 2. The generation of adversarial samples for target attacks can be represented as follows:" + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.223, + 0.784, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.452, + 0.327, + 0.788, + 0.424 + ], + "angle": 0, + "content": "Fig. 2: Position exploration sample vs. Traditional adversarial sample. Traditional adversarial samples are closer to and clustered at the decision center after several iterations of minor adjustments. While our PES has a much broader coverage and is able to explore the decision space more comprehensively." + }, + { + "type": "equation", + "bbox": [ + 0.272, + 0.445, + 0.785, + 0.466 + ], + "angle": 0, + "content": "\\[\nx _ {0} ^ {a d v} = x, x _ {N + 1} ^ {a d v} = C l i p _ {x, \\epsilon} \\left\\{x _ {N} ^ {a d v} - \\alpha s i g n \\left(\\nabla_ {x} L \\left(S \\left(x _ {N + 1} ^ {a d v}\\right), y _ {t a r}\\right)\\right) \\right\\}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.784, + 0.519 + ], + "angle": 0, + "content": "where \\( y_{tar} \\) is the target class, \\( \\nabla_x \\) denotes the gradient of the substitute model, \\( sign \\) denotes the sign function, \\( \\alpha \\) represents the step size during the iteration process, and \\( Clip \\) is the clip function based on the perturbation boundary \\( \\epsilon \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.784, + 0.566 + ], + "angle": 0, + "content": "Finally, the generated adversarial samples are used to attack the target model. Clearly, the similarity between the substitute model and the target model is a key factor for successful attacks in substitute attacks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.588, + 0.725, + 0.604 + ], + "angle": 0, + "content": "3.2 Aligning the Inter-Class Topology in the Decision Space" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.785, + 0.75 + ], + "angle": 0, + "content": "Existing substitute training methods encourage the substitute model to simply mimic the outputs of the target model without fully simulating the decision space of the target model. Resulting in the adversarial samples generated by the substitute model are difficult to be classified into the target class by the target model. Moreover, there is also a lack of research on what aspects the substitute model needs to resemble the target model. In this paper, we propose to understand and improve the substitute training from the perspective of the inter-class topology in the decision space, which refers to the positional relationships (including directions and distances) among all classes in the decision space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.841 + ], + "angle": 0, + "content": "According to Eq. (2), modifying the adversarial samples along the direction of the gradient \\(\\nabla_{x}\\) in the decision space can bring the decision of the model closer to the target class, as shown in Fig. 2. Therefore, if the relative positions of all classes are consistent between the substitute model and the target model, then the adversarial samples generated by the substitute model can be successfully classified into the target class by the target model. This suggests that aligning" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "L. Meng et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.145, + 0.778, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.303, + 0.789, + 0.361 + ], + "angle": 0, + "content": "Fig. 3: The framework of our Inter-Class Topology Alignment (ICTA). ICTA explores the inter-class relative position in the decision space using position exploration sample in different directions, and aligns the inter-class topology of the two models by fitting the inter-class relative position." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.789, + 0.42 + ], + "angle": 0, + "content": "the inter-class topology between the two models in the decision space is crucial for substitute attacks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.789, + 0.482 + ], + "angle": 0, + "content": "For the sample \\( x_{i} \\), we consider \\( S(x_{i}) \\) and \\( T(x_{i}) \\) as nodes in the two decision spaces, and \\( e_{S_{ij}} = \\langle S(x_i), S(x_j) \\rangle \\), \\( i \\neq j \\) and \\( e_{T_{ij}} = \\langle T(x_i), T(x_j) \\rangle \\), \\( i \\neq j \\) to represent the edges of the topology in the two decision spaces. Topology consistency can be achieved by aligning all nodes and edges:" + }, + { + "type": "equation", + "bbox": [ + 0.326, + 0.494, + 0.787, + 0.534 + ], + "angle": 0, + "content": "\\[\n\\min \\sum_ {i = 1} ^ {n} \\mathcal {D} _ {p} \\left(S \\left(x _ {i}\\right), T \\left(x _ {i}\\right)\\right) + \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {n} \\mathcal {D} _ {e} \\left(e _ {S _ {i j}}, e _ {T _ {i j}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.784, + 0.562 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_p\\) and \\(\\mathcal{D}_e\\) denote the distance metrics of points and edges, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.789, + 0.607 + ], + "angle": 0, + "content": "However, aligning the edges among all the samples in the decision space is extremely complex. Therefore, we propose a relative position exploration and alignment strategy for matching the inter-class topology of the two models." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.631, + 0.5, + 0.648 + ], + "angle": 0, + "content": "3.3 Position Exploration Sample" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.764 + ], + "angle": 0, + "content": "According to Fig. 3, the relative position of the clean and adversarial samples in the decision space can similarly be used to represent the relative position of the two classes in the decision space. In this case, the adversarial perturbations used to generate the adversarial samples can reflect the distance and direction between the two classes. Therefore, under the assumption that the target model is white-boxed, the inter-class relative position can be estimated using the positional relationships between the clean and adversarial samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.843 + ], + "angle": 0, + "content": "The traditional adversarial samples aim to progressively bring the sample closer to the decision center of the target class through multiple iterations of minor adjustments. While the sample has a relatively high success rate in the attack, it does not benefit for thoroughly exploring the entire decision space. To alleviate this limitation, we design the Position Exploration Sample (PES)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.442, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Topology Alignment for Substitute Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "to more comprehensively explore the inter-class relative position of the target model in the decision space, which can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.189, + 0.786, + 0.205 + ], + "angle": 0, + "content": "\\[\nx _ {p e s} = x - \\phi \\nabla_ {x} L (T (x), y _ {t a r}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.214, + 0.784, + 0.32 + ], + "angle": 0, + "content": "where \\(\\phi\\) is the perturbation strength. We generate the PES using the full gradient, not just the gradient direction, without iteration and without being constrained by the pixel range. Thus, the gradient \\(\\nabla_{x}\\) and the perturbation strength \\(\\phi\\) determine its direction and distance relative to the clean sample in the decision space. We represent the inter-class relative position based on the relative position of the PES to the clean sample and explore the decision space more thoroughly over a wider coverage by controlling \\(y_{tar}\\) and \\(\\phi\\), as Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.32, + 0.785, + 0.366 + ], + "angle": 0, + "content": "However, it is impossible to generate the PES directly due to the gradient of the target model is inaccessible. Therefore, we retreat to generating PES using the substitute model, as shown in Fig. 3, which can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.377, + 0.786, + 0.393 + ], + "angle": 0, + "content": "\\[\nx _ {p e s} = x - \\phi \\nabla_ {x} L (S (x), y _ {a r b}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.784, + 0.478 + ], + "angle": 0, + "content": "where \\( y_{arb} \\) represents an arbitrary class different from ground truth. In this way, we can explore the decision space from different directions and obtain the complete inter-class topology in the decision space. Moreover, the PES generated by the substitute model has an effect on exploring the decision space of the target model, since our goal is to make the two models as similar as possible." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.498, + 0.554, + 0.515 + ], + "angle": 0, + "content": "3.4 Topology Alignment Based on PES" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.784, + 0.613 + ], + "angle": 0, + "content": "In order to align the inter-class topology of the two models, we simplify the topology consistency in Eq. (3) to the consistency of the inter-class relative position. This method allows us to capture the inter-class topology through the position of two nodes without the need to optimize the edges directly, as demonstrated in Fig. 3. We can achieve this alignment by simultaneously optimizing the knowledge distillation loss \\( L_{KD} \\) and the relative position loss \\( L_{RP} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.624, + 0.785, + 0.64 + ], + "angle": 0, + "content": "\\[\nL _ {K D} = \\mathcal {D} _ {k d} (T (x), S (x)), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.643, + 0.785, + 0.659 + ], + "angle": 0, + "content": "\\[\nL _ {R P} = \\mathcal {D} _ {r p} \\left(T \\left(x _ {p e s}\\right), S \\left(x _ {p e s}\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.662, + 0.785, + 0.678 + ], + "angle": 0, + "content": "\\[\nL _ {T} = \\lambda_ {1} L _ {K D} + \\lambda_ {2} L _ {R P}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.784, + 0.734 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_{kd}\\) and \\(\\mathcal{D}_{rp}\\) are distance measure formulas, \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are hyperparameters. The inter-class relations of the two models can be aligned by ensuring intra-class consistency through \\(L_{KD}\\) and inter-class consistency through \\(L_{RP}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.784, + 0.793 + ], + "angle": 0, + "content": "Furthermore, we can constrain the substitute model to be inter-class consistent with the target model in different directions and distances based on the PES of a clean sample in different directions and at different distances. At this point, the loss of topology consistency can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.404, + 0.803, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\nL _ {T} = \\lambda_ {1} L _ {K D} + \\lambda_ {2} \\sum_ {i = 1} ^ {M} L _ {R P} ^ {i}, \\tag {9}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "L. Meng et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 1: The accuracy of target model. We select four models with different architectures as target models on each of the three datasets." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.185, + 0.699, + 0.307 + ], + "angle": 0, + "content": "
DatasetCIFAR10
Target Model AccuracyResNet34 94.53Vgg16 92.82GoogleNet 95.31ShuffleNetV2 94.53
DatasetCIFAR100
Target Model AccuracyResNet50 79.3MobileNetV2 71.61WideResNet 79.69ResNeXt50 77.56
DatasetTinyImageNet
Target Model AccuracyResNet50 62.81Vgg19 59.66MobileNetV2 62.81ResNeXt50 66.33
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.695, + 0.349 + ], + "angle": 0, + "content": "where \\(M\\) is the number of \\(x_{pes}\\) generated for a same clean sample." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.784, + 0.392 + ], + "angle": 0, + "content": "In addition, incorporating the original categorical loss into this process is often beneficial, which is usually the cross-entropy loss \\( L_{CE} \\). The overall loss can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.394, + 0.785, + 0.409 + ], + "angle": 0, + "content": "\\[\nL _ {a l l} = L _ {T} + \\lambda_ {3} L _ {C E}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.784, + 0.445 + ], + "angle": 0, + "content": "where \\(\\lambda_3\\) is a hyperparameter. The parameters of the substitute model are optimized by back propagation to better simulate the target model." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.466, + 0.384, + 0.483 + ], + "angle": 0, + "content": "4 Experimental" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.495, + 0.438, + 0.51 + ], + "angle": 0, + "content": "4.1 Experimental Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.785, + 0.609 + ], + "angle": 0, + "content": "Dataset and Target Model. We choose CIFAR10 [10] (with size of \\(32 \\times 32\\) and 10 classes), CIFAR100 [10] (with size of \\(32 \\times 32\\) and 100 classes), TinyImageNet [19] (with size of \\(64 \\times 64\\) and 200 classes) as the training dataset. And we select VGG(16, 19) [23], ResNet(34, 50) [6], MobileNetV2 [20], GoogleNet [25], WideResNet [34], ShuffleNetV2 [15], and ResNeXt50 [32] as the target model. The accuracy of the target models is shown in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.785, + 0.719 + ], + "angle": 0, + "content": "Training Setup. During training, we set the perturbation strength in Eq. (5) to \\(\\phi = |Y|\\), where \\(Y\\) obeys the Gaussian distribution \\(Y \\sim N(0,3)\\). The number of PES in Eq. (9) is set to \\(M = 1\\). The hyperparameters are empirically set to \\(\\lambda_1 = \\lambda_3 = 1, \\lambda_2 = 10\\). We standardize the query budget for substitute training across all baselines to ensure fairness, setting it to 2 million for CIFAR10 and 3 million for CIFAR100 and TinyImageNet." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Additionally, we conduct experiments in two scenarios: full-data and data-free. In the data-free scenario, we utilize Stable Diffusion [18] as a generator and use the LCA [21] strategy to produce the data needed for substitute training. In the full-data scenario, the whole training data is available for substitute training. We perform this part of the experiment to validate the performance of the scheme without the influence of the generator. The algorithm of our ICTA and other experimental details are described in the supplement. The source code will be released upon acceptance of the paper." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.443, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Topology Alignment for Substitute Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.186 + ], + "angle": 0, + "content": "Table 2: Comparison of different white-box adversarial sample generation schemes. We choose ResNet34 on CIFAR10, ResNet50 on CIFAR100 and VGG19 in TinyImageNet as target models. The best results are marked as bolded." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.199, + 0.782, + 0.382 + ], + "angle": 0, + "content": "
Dataset MethodsCIFAR10CIFAR100TinyImageNet
FGSMBIMPGDFGSMBIMPGDFGSMBIMPGD
Non-TargetMAZE [9]38.6457.9868.9438.2041.5543.079.189.7911.10
DaST [37]35.1956.2164.2240.0142.5447.409.719.269.81
DST [27]40.8367.6571.2646.4549.2159.4611.6712.4733.52
Zhang et al. [35]60.1194.2197.5483.6994.5391.1838.5848.2165.01
DiffusionLCA [21]68.2995.7397.3082.0395.5896.4951.5552.3570.58
ICTA (Ours)75.6098.7199.1987.4299.0199.2583.7492.7393.39
TargetMAZE [9]6.8413.8316.656.2012.7810.010.502.502.54
DaST [37]7.8515.4118.615.0112.4811.410.203.025.20
DST [27]10.2616.1219.3310.2216.3724.651.883.055.87
Zhang et al. [35]28.1772.0564.057.7458.4554.782.3010.8815.01
DiffusionLCA [21]24.3176.9079.984.6145.1951.501.8412.9515.45
ICTA (Ours)29.4187.6391.407.8069.0274.496.8729.3133.41
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.41, + 0.785, + 0.485 + ], + "angle": 0, + "content": "Evaluation Setup and Process. During testing, we choose three classical methods, including FGSM [5], BIM [11], and PGD [16], to generate adversarial samples, with parameters uniformly set to perturbation boundary \\(\\epsilon = 8 / 255\\) and step size \\(\\alpha = 2 / 255\\). Our hardware utilizes an RTX 2080Ti GPU, and all schemes are retrained under the same environment." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.486, + 0.785, + 0.547 + ], + "angle": 0, + "content": "We train substitute model by our ICTA and then generate adversarial samples based on the substitute models using a white-box adversarial sample generation scheme. The success rate of these adversarial samples in attacking the target model is used as the final evaluation metric." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.568, + 0.657, + 0.584 + ], + "angle": 0, + "content": "4.2 Comparative Experiments in Data-free Scenario" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.592, + 0.786, + 0.743 + ], + "angle": 0, + "content": "Comparison of White-box Adversarial Sample Generation Schemes. We compare different white-box adversarial sample generation schemes, including FGSM, BIM, and PGD. As illustrated in Tab. 2, our ICTA realizes significant performance improvements across different white-box adversarial sample generation methods. On the one hand, our ICTA achieves more noticeable performance enhancement on large datasets TinyImageNet than on small datasets CIFAR10. On the other hand, ICTA has superior performance improvement on target attacks than non-target attacks. The results indicate that our ICTA more effectively simulates the decision space of the target model compared to traditional substitute training methods [9,21,27,35,37]." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.765, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Comparison on Different Target Models. According to the results in Tab. 3, our ICTA demonstrates a significant performance improvement on different target models compared to the GANs-based approaches [9,27,35,37] and the diffusion-based method [21]. Specifically, ICTA achieves the most superior performance in target attacks, especially on TinyImageNet. Additionally, ICTA" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "L. Meng et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 3: Comparison on different target models. We choose PGD as the default white-box adversarial sample generation scheme. The best results are marked as bolded." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.185, + 0.784, + 0.357 + ], + "angle": 0, + "content": "
Dataset Target ModelCIFAR10CIFAR100TinyImageNet
VGG16ResNet34GoogleNetMobileNetResNet50ResNet50VGG19
Non-TargetMAZE [9]47.4968.9427.1351.7343.0715.5411.10
DaST [37]53.2964.2226.4857.9447.4016.119.81
DST [27]55.1571.2638.6465.8559.4632.5133.52
Zhang et al. [35]95.8497.5467.2497.7891.1878.8265.01
DiffusionLCA [21]97.0197.3094.8297.0196.4967.5670.58
ICTA (Ours)96.4899.1997.3299.6399.2594.8093.39
TargetMAZE [9]11.3916.6512.9811.0310.013.182.54
DaST [37]15.3918.615.319.5811.412.215.20
DST [27]17.0719.3318.8918.4624.658.555.87
Zhang et al. [35]57.4764.0539.8344.1154.7820.0115.01
DiffusionLCA [21]75.3979.9878.1851.6151.5012.4515.45
ICTA (Ours)79.5791.4084.5770.5674.4943.0233.41
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.398 + ], + "angle": 0, + "content": "Table 4: Comparison of attack success rate in label-based and probability-based scenarios. The best results are marked as bolded." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.411, + 0.698, + 0.57 + ], + "angle": 0, + "content": "
methodProbability-basedLabel-based
cifar10cifar100cifar10cifar100
Non-TargetGANsDaST [37]64.2247.4067.7236.57
DST [27]71.2659.4680.4248.50
Zhang et al. [35]97.5491.1898.9890.20
DiffusionLCA [21]97.3096.4998.4894.55
ICTA (Ours)99.1999.2599.4998.53
TargetGANsDaST [37]18.6111.4119.285.20
DST [27]19.3324.6523.854.59
Zhang et al. [35]64.0554.7869.2540.17
DiffusionLCA [21]79.9851.5082.1849.85
ICTA (Ours)91.4074.4993.1061.71
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.6, + 0.785, + 0.661 + ], + "angle": 0, + "content": "consistently enhances performance across different target models, showcasing the stability of the performance in various scenarios. In summary, our ICTA is able to simulate different target models in black-box attack scenarios and exhibits substantial performance improvements on large-scale datasets." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.688, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Comparison in Probability-based and Label-based Scenarios. We compare the attack success rate in the probability-based and label-based scenarios, as displayed in Tab. 4. In the label-based scenario, only the output labels of the target model are accessible, while in the probability-based scenario, both the labels and probabilities of the target model are available. The experiment results indicate that our ICTA achieves the optimal performance in both probability-based and label-based scenarios and shows significant improvement over the existing substitute attack methods [21, 27, 35, 37]. This highlights the comprehensive superiority of our ICTA, demonstrating outstanding attack performance even in scenarios where only label outputs are accessible." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.443, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Topology Alignment for Substitute Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.499, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.3, + 0.425, + 0.31 + ], + "angle": 0, + "content": "(a) Attack success rate" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.146, + 0.787, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.3, + 0.743, + 0.311 + ], + "angle": 0, + "content": "(b) Accuracy of substitute model" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.323, + 0.785, + 0.379 + ], + "angle": 0, + "content": "Fig. 4: Comparison of attack success rate and accuracy in full-data scenario. Comparison schemes include the substitute training schemes DST [27] and Zhang et al. [35] as well as a distillation method DIST [8]. The ResNet50 on CIFAR100 is used as the target model." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.41, + 0.655, + 0.426 + ], + "angle": 0, + "content": "4.3 Comparative Experiments in Full-data Scenario" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.785, + 0.587 + ], + "angle": 0, + "content": "Performance of Different Substitute Training Strategies. In Fig. 4, we present the attack success rate (both target and non-target attacks) and the accuracy of the substitute training during the training process. It can be observed that our ICTA demonstrates excellent performance in both target and nontarget attacks and shows higher efficiency in target attacks. There is a significant improvement in non-target attacks compared to conventional methods DST [27] and Zhang et al. [35], which highlights the unique advantages of our ICTA in improving the effectiveness of target attacks. In addition, we compare an excellent knowledge distillation scheme, DIST [8], which also shows the superior performance of our ICTA in attack scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.589, + 0.785, + 0.665 + ], + "angle": 0, + "content": "Notably, we find that the accuracy of the substitute model does not fully reflect the effectiveness of the substitute training. Despite having a relatively low accuracy in our ICTA, the substitute model exhibits excellent performance in attacks. This indicates that our ICTA better simulates the target model in terms of inter-class topology, rather than just replicating its outputs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Comparison of Relative Performance with White-box Attacks. We further compare the attack success rate of our ICTA with the corresponding white-box attacks, as demonstrated in Tab. 5, where the Relative Performance (RP) is used as a metric to measure the performance relative to the white-box attacks. Our ICTA achieves comparable performance to the corresponding white-box attacks in non-target attack, when the query budgets is set to 2M (for CIFAR10) and 3M (for CIFAR100 and TinyImageNet). Moreover, it is worth noting that the target attack success rate exceeds \\(83\\%\\) on the CIFAR10 dataset, \\(67\\%\\) on the CIFAR100 dataset, and \\(44\\%\\) on the TinyImageNet dataset compared to the corresponding white-box attack. As can be seen, the proposed ICTA achieves quite" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "L. Meng et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 5: Comparison with direct white-box attacks on target model. We utilize Relative Performance (RP) to evaluate the attack success of our black-box attack method relative to the white-box attack (in red)." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.199, + 0.784, + 0.469 + ], + "angle": 0, + "content": "
ResNet34Vgg16GoogleNetShuffleNetV2
white-boxOursRPwhite-boxOursRPwhite-boxOursRPwhite-boxOursRP
CIFAR10Non-TargetFGSM78.5483.421.0685.5979.460.9363.8477.731.2284.9783.970.99
BIM100.0099.591.00100.0097.260.9799.9397.610.98100.0098.860.99
PGD100.0099.791.00100.0098.080.98100.0098.260.98100.0099.200.99
TargetFGSM28.7934.691.2019.6630.251.5424.5526.991.1032.9331.780.97
BIM99.8494.580.9584.8181.190.9699.5582.580.83100.0086.010.86
PGD100.0096.630.9798.2185.670.8799.9786.290.86100.0089.030.89
ResNet50WideResNetMobileNetV2ResNeXt50
CIFAR100Non-TargetFGSM84.7786.761.0274.2783.401.1290.1888.200.9887.2487.511.00
BIM100.0099.220.9998.1196.830.9999.9899.090.9999.9899.160.99
PGD100.0099.500.9999.5797.540.98100.0099.460.99100.0099.470.99
TargetFGSM3.058.522.793.879.422.434.397.351.673.838.792.30
BIM95.2272.790.7689.4465.900.7496.3064.790.6797.0673.290.76
PGD99.2579.340.8098.3873.220.7499.6670.330.7199.7080.150.80
ResNet50Vgg19MobileNetV2ResNeXt50
TinyImageNetNon-TargetFGSM95.8286.070.9092.8286.740.9398.4288.000.8996.6283.650.87
BIM99.9796.260.9699.6495.940.96100.0097.210.9799.9894.780.95
PGD99.9896.960.9799.9196.560.97100.0097.780.9899.9895.860.96
TargetFGSM10.2510.040.984.797.991.6715.718.740.5613.188.680.66
BIM98.9652.300.5385.6639.310.4699.7245.920.4699.5343.560.44
PGD99.7457.850.5897.8146.310.4799.9251.180.5199.9650.300.50
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.785, + 0.525 + ], + "angle": 0, + "content": "satisfactory performance in black-box attack scenarios, demonstrating robust adaptability in target attacks across different datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.545, + 0.602, + 0.561 + ], + "angle": 0, + "content": "4.4 Analysis of Decision Boundary Similarity" + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.569, + 0.431, + 0.597 + ], + "angle": 0, + "content": "The t-SNE Visualization of Decision Space. We" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.454, + 0.84 + ], + "angle": 0, + "content": "exhibit the t-Distributed Stochastic Neighbor Embedding (t-SNE) visualization of the decision space of the models, as exhibited in Fig. 5. In t-SNE visualization, each data point is typically labeled with its corresponding class. Therefore, it can depict the interclass relative position in the decision space, as well as the inter-class topology. The visualization illustrates that the inter-class topology in the decision space of the substitute model trained by our ICTA" + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.6, + 0.536, + 0.789 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.6, + 0.618, + 0.79 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.6, + 0.7, + 0.79 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.6, + 0.783, + 0.79 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.801, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Fig. 5: The t-SNE visualization of the decision space for (a) the target model, (b) our ICTA, (c) Zhang et al. [35] and (d) DST [27]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.443, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Topology Alignment for Substitute Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.145, + 0.499, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.294, + 0.426, + 0.305 + ], + "angle": 0, + "content": "(a) Attack success rate" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.147, + 0.787, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.293, + 0.743, + 0.305 + ], + "angle": 0, + "content": "(b) Accuracy of substitute model" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.317, + 0.784, + 0.359 + ], + "angle": 0, + "content": "Fig. 6: Ablation experiments on PES. The experiments include baseline (no relative position loss) and strategies for exploring inter-class relative position through traditional adversarial samples, clean samples, and position exploration sample." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.785, + 0.466 + ], + "angle": 0, + "content": "method is remarkably similar to the topology of the target model. On the contrary, both the output alignment-focused DST [27] and the decision boundary-constrained approach by Zhang et al. [35] exhibit greater differences from the target model in t-SNE visualization. This confirms that our ICTA is more effective in aligning the substitute model to match the target model." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.489, + 0.487, + 0.624 + ], + "angle": 0, + "content": "Quantitative Analysis. We explore random directions of arbitrary samples, and then quantitatively assess the Decision Boundary Similarity (DBS) based on the overlap between target and substitute models. As in Tab. 6, our ICTA has more consistent decision boundaries with the target model." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.533, + 0.785, + 0.56 + ], + "angle": 0, + "content": "Table 6: Quantitative analysis of decision boundary similarity." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.562, + 0.787, + 0.613 + ], + "angle": 0, + "content": "
MethodICTA (Ours)Zhang et al.DST
CIFAR100.930.870.80
CIFAR1000.880.770.71
TinyImageNet0.820.570.55
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.648, + 0.388, + 0.664 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.785, + 0.794 + ], + "angle": 0, + "content": "Ablation on Position Exploration Sample. In this paper, we design PES to explore the inter-class relative position in the decision space through their relative position relationships with respect to the original clean samples. However, as the position of any two samples in the decision space inherently depict the inter-class relative position, traditional adversarial sample or even a clean sample has the effect of exploring the decision space. In this section, we validate the effectiveness of PES by comparing it with traditional adversarial sample generated by BIM and clean sample, as illustrated in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.785, + 0.84 + ], + "angle": 0, + "content": "It is worth noting that while all three approaches contribute to explore the inter-class relative position in the decision space, PES stands out as particularly advantageous. This is due to the fact that both clean samples and the adversarial" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "L. Meng et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 7: Ablation comparison of the loss function and the distance metric. We select two distance metrics, MSE and DIST [8], and use different combinations of distance metrics in the loss function. The best result are marked as bolded, and the sub-optimal results are marked as underlined." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.213, + 0.699, + 0.306 + ], + "angle": 0, + "content": "
DkdMSE DISTDrpMSE DISTTargetNon-Target
FGSMBIMPGDFGSMBIMPGD
✓4.3142.1446.1479.4895.3796.38
✓4.1155.5257.6782.8598.3098.80
✓✓7.6464.2071.7084.3598.4399.06
✓✓7.8273.4279.0986.7399.0499.61
✓✓8.5272.7979.3486.7699.2299.50
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.788, + 0.412 + ], + "angle": 0, + "content": "samples generated by BIM are closer to the decision center, resulting in limited exploration in the decision space. In contrast, the proposed PES is more diverse by employing the full-gradient and unconstrained setting, enabling exploration across a wider range of the decision space. The ablation experiments for the parameters of PES and the number of PES are provided in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.571 + ], + "angle": 0, + "content": "Ablation on Loss and Metrics. We conduct an ablation study on the loss functions and distance metrics of substitute training in Tab. 7, focusing on the distillation loss \\( L_{KD} \\) and relative position loss \\( L_{RP} \\). Here, the term 'DIST' refers to the distance metric function in the distillation scheme DIST [8]. It can be seen that when DIST [8] is used as the distance metric in \\( L_{KD} \\), there is a certain improvement in the attack success rate. And when \\( L_{RP} \\) is added, there is a significant improvement in the attack success rate, which underscores the indispensability of the relative position loss. Notably, our ICTA employs DIST as the distance metric in both losses, as it further improves attack performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.596, + 0.36, + 0.612 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we propose a novel ICTA scheme to better align the substitute model to the target model for efficient black-box substitute attacks. Compared to the existing schemes that only align intra-class consistency, we simulate the target model in the decision space more comprehensively through inter-class topology alignment. Specifically, we first explore the inter-class relative position of the target model in the decision space by the designed PES. Subsequently, we achieve inter-class topology alignment of the two models by constraining the inter-class relative position of the substitute model in different directions. Benefiting from the better alignment of the two models in the decision space, the adversarial samples generated by the substitute model are more successful in misleading the target model to classify them into the target class. Experimental results demonstrate the effectiveness of ICTA, which achieves better simulation of the target model and significantly improves the attack success rate beyond existing substitute attack approaches." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.443, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Topology Alignment for Substitute Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.237 + ], + "angle": 0, + "content": "This work was supported by the National Key Research and Development Program of China (Grant No. 2021YFA1000102), the National Natural Science Foundation of China (Grant Nos. 62376285, 62272375, and 61673396), and Natural Science Foundation of Shandong Province, China (Grant No. ZR2022MF260)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.259, + 0.323, + 0.275 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.331 + ], + "angle": 0, + "content": "1. Andriushchenko, M., Croce, F., Flammarion, N., Hein, M.: Square Attack: a query-efficient black-box adversarial attack via random search. In: Proceedings of the European Conference on Computer Vision. pp. 484-501 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.785, + 0.372 + ], + "angle": 0, + "content": "2. Brendel, W., Rauber, J., Bethge, M.: Decision-Based Adversarial Attacks: Reliable Attacks Against Black-Box Machine Learning Models. In: International Conference on Learning Representations. arXiv preprint arXiv:1712.04248 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.373, + 0.785, + 0.413 + ], + "angle": 0, + "content": "3. Chen, S., He, Z., Sun, C., Yang, J., Huang, X.: Universal Adversarial Attack on Attention and the Resulting Dataset DImageNet. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(4), 2188-2197 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.415, + 0.785, + 0.455 + ], + "angle": 0, + "content": "4. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: International Conference on Machine Learning. vol. 119, pp. 2206-2216 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.457, + 0.785, + 0.482 + ], + "angle": 0, + "content": "5. Goodfellow, I., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.484, + 0.785, + 0.524 + ], + "angle": 0, + "content": "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.525, + 0.785, + 0.565 + ], + "angle": 0, + "content": "7. Huang, H., Chen, Z., Chen, H., Wang, Y., Zhang, K.: T-SEA: Transfer-based Self-Ensemble Attack on Object Detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20514-20523 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.567, + 0.785, + 0.606 + ], + "angle": 0, + "content": "8. Huang, T., You, S., Wang, F., Qian, C., Xu, C.: Knowledge distillation from a stronger teacher. In: Advances in Neural Information Processing Systems. vol. 35, pp. 33716-33727 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.607, + 0.785, + 0.647 + ], + "angle": 0, + "content": "9. Kariyappa, S., Prakash, A., Qureshi, M.K.: Maze: Data-free model stealing attack using zeroth-order gradient estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13814-13823 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.649, + 0.785, + 0.675 + ], + "angle": 0, + "content": "0. Krizhevsky, A.: Learning Multiple Layers of Features from Tiny Images. Master's thesis, University of Tront (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.785, + 0.702 + ], + "angle": 0, + "content": "1. Kurakin, A., Goodfellow, I., Bengio, S.: Adversarial examples in the physical world. arXiv preprint arXiv:1607.02533 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.704, + 0.785, + 0.743 + ], + "angle": 0, + "content": "2. Lee, B.K., Kim, J.: Mitigating Adversarial Vulnerability through Causal Parameter Estimation by Adversarial Double Machine Learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4499-4509 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.771 + ], + "angle": 0, + "content": "3. Li, L., Lian, Q., Chen, Y.C.: Adv3D: Generating 3D Adversarial Examples in Driving Scenarios with NeRF. arXiv preprint arXiv:2309.01351 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "4. Liu, J., Lu, B., Xiong, M., Zhang, T., Xiong, H.: Adversarial Attack with Raindrops. arXiv preprint arXiv:2302.14267 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Ma, N., Zhang, X., Zheng, H.T., Sun, J.: ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design. In: Proceedings of the European Conference on Computer Vision. pp. 122-138 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "L. Meng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "17. Ren, K., Zheng, T., Qin, Z., Liu, X.: Adversarial Attacks and Defenses in Deep Learning. Engineering 6(3), 346-360 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.244 + ], + "angle": 0, + "content": "18. Robin, R., Andreas, B., Dominik, L., Patrick, E., Bj\\''orn, O.: High-Resolution Image Synthesis with Latent Diffusion Models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10674-10685 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.245, + 0.785, + 0.299 + ], + "angle": 0, + "content": "19. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision 115(3), 211-252 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "20. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: MobileNetV2: Inverted Residuals and Linear Bottlenecks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4510-4520 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "21. Shao, M., Meng, L., Qiao, Y., Zhang, L., Zuo, W.: Data-free black-box attack based on diffusion model. arXiv preprint arXiv:2307.12872 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "22. Sheng, L., Liang, J., He, R., Wang, Z., Tan, T.: AdaptGuard: Defending Against Universal Attacks for Model Adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19093-19103 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.438 + ], + "angle": 0, + "content": "23. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "24. Su, J., Vargas, D.V., Sakurai, K.: One Pixel Attack for Fooling Deep Neural Networks. IEEE Transactions on Evolutionary Computation 23(5), 828-841 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.467, + 0.785, + 0.522 + ], + "angle": 0, + "content": "25. Szegedy, C., Wei Liu, Yangqing Jia, Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1-9 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.523, + 0.785, + 0.577 + ], + "angle": 0, + "content": "26. Wang, N., Luo, Y.: Does Physical Adversarial Example Really Matter to Autonomous Driving? Towards System-Level Effect of Adversarial Object Evasion Attack. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4412-4423 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.619 + ], + "angle": 0, + "content": "27. Wang, W., Qian, X., Fu, Y., Xue, X.: Dst: Dynamic substitute training for data-free black-box attack. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14361-14370 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.62, + 0.785, + 0.674 + ], + "angle": 0, + "content": "28. Wang, W., Yin, B., Yao, T., Zhang, L., Fu, Y., Ding, S., Li, J., Huang, F., Xue, X.: Delving into data: Effectively substitute training for black-box attack. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4761-4770 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "29. Wang, X., Zhang, Z.: Structure Invariant Transformation for better Adversarial Transferability. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4607-4619 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "30. Wang, X., Zhang, Z., Tong, K., Gong, D., He, K., Li, Z., Liu, W.: Triangle Attack: A Query-efficient Decision-based Adversarial Attack. In: Proceedings of the European Conference on Computer Vision. pp. 156-174 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "31. Wang, Z., Yang, H., Feng, Y., Sun, P., Guo, H., Zhang, Z., Ren, K.: Towards Transferable Targeted Adversarial Examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K.: Aggregated Residual Transformations for Deep Neural Networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5987-5995 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.443, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Topology Alignment for Substitute Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "33. Yu, M., Sun, S.: FE-DaST: Fast and effective data-free substitute training for black-box adversarial attacks. Computers & Security 113, 102555 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.203 + ], + "angle": 0, + "content": "34. Zagoruyko, S., Komodakis, N.: Wide Residual Networks. In: Proceedings of the British Machine Vision Conference. pp. 87.1-87.12 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.788, + 0.245 + ], + "angle": 0, + "content": "35. Zhang, J., Li, B., Xu, J., Wu, S., Ding, S., Zhang, L., Wu, C.: Towards efficient data free black-box adversarial attack. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15094-15104 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.788, + 0.273 + ], + "angle": 0, + "content": "36. Zheng, X., Fan, Y., Wu, B., Zhang, Y., Wang, J., Pan, S.: Robust Physical-World Attacks on Face Recognition. Pattern Recognition 133, 109009 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.314 + ], + "angle": 0, + "content": "37. Zhou, M., Wu, J., Liu, Y., Liu, S., Zhu, C.: Dast: Data-free substitute training for adversarial attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 234-243 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.357 + ], + "angle": 0, + "content": "38. Zhu, Y., Chen, Y., Li, X., Chen, K., He, Y., Tian, X., Zheng, B., Chen, Y., Huang, Q.: Toward Understanding and Boosting Adversarial Transferability From a Distribution Perspective. IEEE Transactions on Image Processing 31, 6487-6501 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.357 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks/5bfa27b2-acad-4e1e-b905-e3f3b1d14c81_origin.pdf b/2024/Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks/5bfa27b2-acad-4e1e-b905-e3f3b1d14c81_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..79a3d9d9005771ffed78376c3af29d8ffecb5926 --- /dev/null +++ b/2024/Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks/5bfa27b2-acad-4e1e-b905-e3f3b1d14c81_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e2b756dc2ee62859f94ab61d6423c82a6a62b8f2dc77540d9d75b1c7d7056f5 +size 4118921 diff --git a/2024/Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks/full.md b/2024/Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0c717a64e1e4bbc7bd39607ae75278c6c1314b3a --- /dev/null +++ b/2024/Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks/full.md @@ -0,0 +1,308 @@ +# Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks + +Lingzhuang Meng $^{1}$ , Mingwen Shao $^{1\star}$ , Yuanjian Qiao $^{1}$ , and Wenjie Liu $^{1}$ + +School of Computer Science and Technology, China University of Petroleum (East + +China), Qingdao, 266580, China + +lzhmeng1688@163.com, smw278@126.com, yjqiao@s.upc.edu.cn, + +liuwenjie_82@163.com + +Abstract. In black-box attacks based on substitute training, the similarity of the substitute model to the target model is critical for successful attacks. However, existing schemes merely train the substitute model to mimic the outputs of the target model without fully simulating the decision space, resulting in the adversarial samples generated by the substitute model being classified into the non-target class by the target model. To alleviate this issue, we propose a novel Inter-Class Topology Alignment (ICTA) scheme to more comprehensively simulate the target model by aligning the inter-class positional relationships of two models in the decision space. Specifically, we first design the Position Exploration Sample (PES) to more thoroughly explore the relative positional relationships between classes in the decision space of the target model. Subsequently, we align the inter-class topology between the two models by utilizing the PES to constrain the inter-class relative position of the substitute model in different directions. In this way, the substitute model is more consistent with the target model in the decision space, so that the generated adversarial samples will be more successful in misleading the target model to classify them into the target class. The experimental results demonstrate that our ICTA significantly improves attack success rate in various scenarios compared to existing substitute training methods, particularly performing efficiently in target attacks. + +Keywords: Black-box attacks $\cdot$ Substitute attacks $\cdot$ Position exploration sample $\cdot$ Inter-class topology alignment + +# 1 Introduction + +Convolutional Neural Networks (CNNs) have shown impressive performance in various fields and are widely applied in daily life. However, it has been found that CNNs exhibit poor robustness to well-designed perturbations [3, 36, 38], which poses a great threat to their safety and reliability. For example, in the field of autonomous driving, if the network makes a wrong judgment, it will lead to potential safety hazards [13, 14, 26]. Therefore, it is especially urgent to + +![](images/bdcc90cce1ae60b4488e67de5b5fff51696c7e9b51cff2b913366763d0fc8a33.jpg) +Fig. 1: (a) Comparison of different methods and (b) Attack success rate. Traditional approaches only focus on intra-class consistency, resulting in the adversarial samples generated by the substitute model failing to mislead the target model to classify them into the target class. In contrast, our ICTA better simulates the target model by aligning the inter-class topology in the decision space, resulting in a successful target attack. The performance of our ICTA is significantly improved across multiple datasets, particularly in target attacks. + +research deeply into the adversarial attacks [4,30,38] and defense [17,22] schemes to enhance the reliability and security of CNNs in practical applications. + +Existing adversarial attacks can be categorized into white-box and black-box attacks according to the knowledge of the attacker about the target model. In white-box attacks [3-5,11,12,16,29], the attacker possesses complete information about the target model (such as network structure, gradient, etc.), thus can generate highly deceptive adversarial samples based on this information. By contrast, in black-box attacks [2,9,21,24,31,35], the attacker can only utilize the outputs of the network (probabilities or labels) to craft adversarial samples, making the task more challenging and realistic. In such scenario, the substitute attacks [9,27,33,35,37] is proposed, which focus on training a substitute model to emulate the target model, so that the adversarial samples generated by the substitute model can effectively attack the target model. + +Current substitute attack methods [9, 21, 27, 28, 33, 35, 37] adhere to the paradigm of knowledge distillation, encouraging the substitute model to mimic the outputs of the target model. Some schemes utilized simple $L_{2}$ loss [21, 28, 33, 37] or KL-divergence [9] to constrain the substitute model. As an improvement, DST [27] leveraged a graph structure consisting of multiple outputs to optimize the substitute model. Nevertheless, these strategies only focus on learning the simplistic outputs of the target model, without adequately considering the decision space of the network. Subsequently, Zhang et al. [35] considered aligning the decision boundary between the substitute model and the target model to bridge the gap between the two models. However, this restriction of decision boundary only aims at the intra-class decision consistency and ignores the inter-class relationship consistency of the two models in the decision space. As a result, the adversarial samples generated by the substitute model are difficult to be classified as target class by the target model, as Fig. 1(a). + +To alleviate the above issue, we propose a novel Inter-Class Topology Alignment (ICTA) scheme to align the inter-class positional relationships of the two models in decision space, thus better simulating the target model and achieving more accurate target attacks. Specifically, we first design the Position Exploration Sample (PES) to explore the relative positional relationships among all classes of the target model by introducing adversarial perturbations of different directions and strengths. Subsequently, we constrain the inter-class positional consistency of the substitute model to align the inter-class topology between the two models. Thanks to the better alignment of the two models in the decision space, the adversarial samples generated by the substitute model can be more successful deceiving the target model to classify them into the target class, as shown in Fig. 1(a). Experiments demonstrate that our ICTA achieves higher attack success rate on different datasets, and especially more pronounced in target attacks, as Fig. 1(b). Our primary contributions are as follows: + +- We propose ICTA to innovatively align the substitute model with the target model in terms of inter-class topology in the decision space, which effectively improves the success rate of black-box substitute attack. +- We propose a novel PES to more comprehensively explore the inter-class relative position of the target model, which is utilized to align the inter-class topology between the substitute model and the target model. +- Experiments illustrate that our ICTA achieves significant performance gains in a variety of scenarios and outperforms on target attacks compared to the existing substitute attack schemes. + +# 2 Related Works + +# 2.1 White-box attacks + +In white-box attacks, the attacker is fully aware of the information about the target model, such as the structure and parameters [5,11,12,16], and highly deceptive adversarial samples can be generated based on this information. Classical white-box adversarial sample generation schemes utilized gradient information of network to generate adversarial samples, such as FGSM [5], BIM [11], and PGD [16]. Moreover, Croce et al. [4] proposed parameter-free attack combination AutoAttack by integrating multiple attack methods for testing adversarial robustness. However, white-box attacks may lack practicality in real-world scenarios due to the requirement for extensive knowledge about the target model. + +# 2.2 Black-box attacks + +In contrast, black-box attacks are more challenging and more in line with realistic scenarios, as the attacker can generate adversarial samples only using the outputs of the target model, including transfer attacks, query attacks, and substitute attacks. Transfer attacks [1, 3, 7, 29, 31] aim to maximize the transferability of existing adversarial samples, enabling them to successfully attack the target + +model. For example, Zhu et al. [38] proposed modifying the distribution of the original images to match the distribution of the target class, thereby obtaining adversarial samples with high transferability. Query attacks [2, 24, 30] obtain adversarial samples by continuously querying the target model using specific inputs until a desired output is obtained. Substitute attacks [21, 35, 37] focus on training a substitute model that is similar to the target model, so that the adversarial samples generated by the substitute model can successfully attack the target model. Compared to other schemes, the substitute attack is able to generate more effective adversarial samples due to the ability to simulate the decision process of the target model. + +Most of the existing substitute attack schemes used $L_{2}$ loss [21, 28, 33, 37] or KL-divergence [9] as the constraint to train substitute models. For example, DaST [37] used Generative Adversarial Networks (GANs) to generate data for substitute training and leveraged the $L_{2}$ loss as a constraint to align the outputs of the substitute model with the target model. Kariyappa et al. [9] optimized the generator with the help of zeroth-order gradient estimate and trained the substitute training model with a simple KL-divergence. Shao et al. [21] proposed LCA to utilize a diffusion model to generate the data required for substitute training, which greatly improved the training efficiency compared to traditional GANs-based methods. However, the target attack success rate is not impressive as it only used $L_{2}$ loss to constrain the substitute model. In order to simulate the target model more precisely, Wang et al. [27] proposed DST to optimize the substitute model from a graph with multiple outputs of the target model instead of a single output. Nevertheless, these schemes only learn the simple outputs of the target model without fully considering the complex decision space attributes, resulting in poor performance. Subsequently, Zhang et al. [35] encouraged a high alignment of decision boundary between the substitute model and the target model, which improved the attack success rate. + +However, the aforementioned schemes only focused on the intra-class consistency and neglected the inter-class relational consistency, resulting in the adversarial samples generated by the substitute model being difficult to be classified as target class by the target model. In this paper, we further consider the inter-class positional relationships in the decision space and better simulate the target model by aligning the inter-class topology for a more effective target attack. + +# 3 Methodology + +# 3.1 Preliminary + +For a black-box target model $T$ , existing research on substitute attacks typically trains a substitute model $S$ similar to the $T$ through a knowledge distillation paradigm. Formally, the substitute model can be obtained by substitute training: + +$$ +\underset {\theta} {\arg \min } \mathbb {E} _ {x \sim \mathcal {X}} d [ S (x), T (x) ], \tag {1} +$$ + +where $d$ denotes the distance metric function, $x \sim \mathcal{X}$ denotes the training data, and $\theta$ is the parameters of $S$ . The goal of substitute training is to optimize the substitute model to resemble the target model as similar as possible. + +Subsequently, the adversarial samples can be generated using white-box adversarial sample generation methods based on the substitute model. We take BIM [11] as an instance, which perturbs the input data along the gradient direction of the target class through multiple iterations to make the output of the model close to the target class in the decision space, as Fig. 2. The generation of adversarial samples for target attacks can be represented as follows: + +![](images/804353b81756ac88ffbd8ccde680b1e847d06b4ca4de33b509c58dd0a63b686f.jpg) +Fig. 2: Position exploration sample vs. Traditional adversarial sample. Traditional adversarial samples are closer to and clustered at the decision center after several iterations of minor adjustments. While our PES has a much broader coverage and is able to explore the decision space more comprehensively. + +$$ +x _ {0} ^ {a d v} = x, x _ {N + 1} ^ {a d v} = C l i p _ {x, \epsilon} \left\{x _ {N} ^ {a d v} - \alpha s i g n \left(\nabla_ {x} L \left(S \left(x _ {N + 1} ^ {a d v}\right), y _ {t a r}\right)\right) \right\}, \tag {2} +$$ + +where $y_{tar}$ is the target class, $\nabla_x$ denotes the gradient of the substitute model, $sign$ denotes the sign function, $\alpha$ represents the step size during the iteration process, and $Clip$ is the clip function based on the perturbation boundary $\epsilon$ . + +Finally, the generated adversarial samples are used to attack the target model. Clearly, the similarity between the substitute model and the target model is a key factor for successful attacks in substitute attacks. + +# 3.2 Aligning the Inter-Class Topology in the Decision Space + +Existing substitute training methods encourage the substitute model to simply mimic the outputs of the target model without fully simulating the decision space of the target model. Resulting in the adversarial samples generated by the substitute model are difficult to be classified into the target class by the target model. Moreover, there is also a lack of research on what aspects the substitute model needs to resemble the target model. In this paper, we propose to understand and improve the substitute training from the perspective of the inter-class topology in the decision space, which refers to the positional relationships (including directions and distances) among all classes in the decision space. + +According to Eq. (2), modifying the adversarial samples along the direction of the gradient $\nabla_{x}$ in the decision space can bring the decision of the model closer to the target class, as shown in Fig. 2. Therefore, if the relative positions of all classes are consistent between the substitute model and the target model, then the adversarial samples generated by the substitute model can be successfully classified into the target class by the target model. This suggests that aligning + +![](images/fe668db758a260ad7998573604edf96c783bf422356bc2f6f5c918caeb60661f.jpg) +Fig. 3: The framework of our Inter-Class Topology Alignment (ICTA). ICTA explores the inter-class relative position in the decision space using position exploration sample in different directions, and aligns the inter-class topology of the two models by fitting the inter-class relative position. + +the inter-class topology between the two models in the decision space is crucial for substitute attacks. + +For the sample $x_{i}$ , we consider $S(x_{i})$ and $T(x_{i})$ as nodes in the two decision spaces, and $e_{S_{ij}} = \langle S(x_i), S(x_j) \rangle$ , $i \neq j$ and $e_{T_{ij}} = \langle T(x_i), T(x_j) \rangle$ , $i \neq j$ to represent the edges of the topology in the two decision spaces. Topology consistency can be achieved by aligning all nodes and edges: + +$$ +\min \sum_ {i = 1} ^ {n} \mathcal {D} _ {p} \left(S \left(x _ {i}\right), T \left(x _ {i}\right)\right) + \sum_ {i = 1} ^ {n} \sum_ {j = 1} ^ {n} \mathcal {D} _ {e} \left(e _ {S _ {i j}}, e _ {T _ {i j}}\right), \tag {3} +$$ + +where $\mathcal{D}_p$ and $\mathcal{D}_e$ denote the distance metrics of points and edges, respectively. + +However, aligning the edges among all the samples in the decision space is extremely complex. Therefore, we propose a relative position exploration and alignment strategy for matching the inter-class topology of the two models. + +# 3.3 Position Exploration Sample + +According to Fig. 3, the relative position of the clean and adversarial samples in the decision space can similarly be used to represent the relative position of the two classes in the decision space. In this case, the adversarial perturbations used to generate the adversarial samples can reflect the distance and direction between the two classes. Therefore, under the assumption that the target model is white-boxed, the inter-class relative position can be estimated using the positional relationships between the clean and adversarial samples. + +The traditional adversarial samples aim to progressively bring the sample closer to the decision center of the target class through multiple iterations of minor adjustments. While the sample has a relatively high success rate in the attack, it does not benefit for thoroughly exploring the entire decision space. To alleviate this limitation, we design the Position Exploration Sample (PES) + +to more comprehensively explore the inter-class relative position of the target model in the decision space, which can be represented as: + +$$ +x _ {p e s} = x - \phi \nabla_ {x} L (T (x), y _ {t a r}), \tag {4} +$$ + +where $\phi$ is the perturbation strength. We generate the PES using the full gradient, not just the gradient direction, without iteration and without being constrained by the pixel range. Thus, the gradient $\nabla_{x}$ and the perturbation strength $\phi$ determine its direction and distance relative to the clean sample in the decision space. We represent the inter-class relative position based on the relative position of the PES to the clean sample and explore the decision space more thoroughly over a wider coverage by controlling $y_{tar}$ and $\phi$ , as Fig. 2. + +However, it is impossible to generate the PES directly due to the gradient of the target model is inaccessible. Therefore, we retreat to generating PES using the substitute model, as shown in Fig. 3, which can be represented as: + +$$ +x _ {p e s} = x - \phi \nabla_ {x} L (S (x), y _ {a r b}), \tag {5} +$$ + +where $y_{arb}$ represents an arbitrary class different from ground truth. In this way, we can explore the decision space from different directions and obtain the complete inter-class topology in the decision space. Moreover, the PES generated by the substitute model has an effect on exploring the decision space of the target model, since our goal is to make the two models as similar as possible. + +# 3.4 Topology Alignment Based on PES + +In order to align the inter-class topology of the two models, we simplify the topology consistency in Eq. (3) to the consistency of the inter-class relative position. This method allows us to capture the inter-class topology through the position of two nodes without the need to optimize the edges directly, as demonstrated in Fig. 3. We can achieve this alignment by simultaneously optimizing the knowledge distillation loss $L_{KD}$ and the relative position loss $L_{RP}$ : + +$$ +L _ {K D} = \mathcal {D} _ {k d} (T (x), S (x)), \tag {6} +$$ + +$$ +L _ {R P} = \mathcal {D} _ {r p} \left(T \left(x _ {p e s}\right), S \left(x _ {p e s}\right)\right), \tag {7} +$$ + +$$ +L _ {T} = \lambda_ {1} L _ {K D} + \lambda_ {2} L _ {R P}, \tag {8} +$$ + +where $\mathcal{D}_{kd}$ and $\mathcal{D}_{rp}$ are distance measure formulas, $\lambda_{1}$ and $\lambda_{2}$ are hyperparameters. The inter-class relations of the two models can be aligned by ensuring intra-class consistency through $L_{KD}$ and inter-class consistency through $L_{RP}$ . + +Furthermore, we can constrain the substitute model to be inter-class consistent with the target model in different directions and distances based on the PES of a clean sample in different directions and at different distances. At this point, the loss of topology consistency can be expressed as: + +$$ +L _ {T} = \lambda_ {1} L _ {K D} + \lambda_ {2} \sum_ {i = 1} ^ {M} L _ {R P} ^ {i}, \tag {9} +$$ + +Table 1: The accuracy of target model. We select four models with different architectures as target models on each of the three datasets. + +
DatasetCIFAR10
Target Model AccuracyResNet34 94.53Vgg16 92.82GoogleNet 95.31ShuffleNetV2 94.53
DatasetCIFAR100
Target Model AccuracyResNet50 79.3MobileNetV2 71.61WideResNet 79.69ResNeXt50 77.56
DatasetTinyImageNet
Target Model AccuracyResNet50 62.81Vgg19 59.66MobileNetV2 62.81ResNeXt50 66.33
+ +where $M$ is the number of $x_{pes}$ generated for a same clean sample. + +In addition, incorporating the original categorical loss into this process is often beneficial, which is usually the cross-entropy loss $L_{CE}$ . The overall loss can be expressed as: + +$$ +L _ {a l l} = L _ {T} + \lambda_ {3} L _ {C E}, \tag {10} +$$ + +where $\lambda_3$ is a hyperparameter. The parameters of the substitute model are optimized by back propagation to better simulate the target model. + +# 4 Experimental + +# 4.1 Experimental Details + +Dataset and Target Model. We choose CIFAR10 [10] (with size of $32 \times 32$ and 10 classes), CIFAR100 [10] (with size of $32 \times 32$ and 100 classes), TinyImageNet [19] (with size of $64 \times 64$ and 200 classes) as the training dataset. And we select VGG(16, 19) [23], ResNet(34, 50) [6], MobileNetV2 [20], GoogleNet [25], WideResNet [34], ShuffleNetV2 [15], and ResNeXt50 [32] as the target model. The accuracy of the target models is shown in Tab. 1. + +Training Setup. During training, we set the perturbation strength in Eq. (5) to $\phi = |Y|$ , where $Y$ obeys the Gaussian distribution $Y \sim N(0,3)$ . The number of PES in Eq. (9) is set to $M = 1$ . The hyperparameters are empirically set to $\lambda_1 = \lambda_3 = 1, \lambda_2 = 10$ . We standardize the query budget for substitute training across all baselines to ensure fairness, setting it to 2 million for CIFAR10 and 3 million for CIFAR100 and TinyImageNet. + +Additionally, we conduct experiments in two scenarios: full-data and data-free. In the data-free scenario, we utilize Stable Diffusion [18] as a generator and use the LCA [21] strategy to produce the data needed for substitute training. In the full-data scenario, the whole training data is available for substitute training. We perform this part of the experiment to validate the performance of the scheme without the influence of the generator. The algorithm of our ICTA and other experimental details are described in the supplement. The source code will be released upon acceptance of the paper. + +Table 2: Comparison of different white-box adversarial sample generation schemes. We choose ResNet34 on CIFAR10, ResNet50 on CIFAR100 and VGG19 in TinyImageNet as target models. The best results are marked as bolded. + +
Dataset MethodsCIFAR10CIFAR100TinyImageNet
FGSMBIMPGDFGSMBIMPGDFGSMBIMPGD
Non-TargetMAZE [9]38.6457.9868.9438.2041.5543.079.189.7911.10
DaST [37]35.1956.2164.2240.0142.5447.409.719.269.81
DST [27]40.8367.6571.2646.4549.2159.4611.6712.4733.52
Zhang et al. [35]60.1194.2197.5483.6994.5391.1838.5848.2165.01
DiffusionLCA [21]68.2995.7397.3082.0395.5896.4951.5552.3570.58
ICTA (Ours)75.6098.7199.1987.4299.0199.2583.7492.7393.39
TargetMAZE [9]6.8413.8316.656.2012.7810.010.502.502.54
DaST [37]7.8515.4118.615.0112.4811.410.203.025.20
DST [27]10.2616.1219.3310.2216.3724.651.883.055.87
Zhang et al. [35]28.1772.0564.057.7458.4554.782.3010.8815.01
DiffusionLCA [21]24.3176.9079.984.6145.1951.501.8412.9515.45
ICTA (Ours)29.4187.6391.407.8069.0274.496.8729.3133.41
+ +Evaluation Setup and Process. During testing, we choose three classical methods, including FGSM [5], BIM [11], and PGD [16], to generate adversarial samples, with parameters uniformly set to perturbation boundary $\epsilon = 8 / 255$ and step size $\alpha = 2 / 255$ . Our hardware utilizes an RTX 2080Ti GPU, and all schemes are retrained under the same environment. + +We train substitute model by our ICTA and then generate adversarial samples based on the substitute models using a white-box adversarial sample generation scheme. The success rate of these adversarial samples in attacking the target model is used as the final evaluation metric. + +# 4.2 Comparative Experiments in Data-free Scenario + +Comparison of White-box Adversarial Sample Generation Schemes. We compare different white-box adversarial sample generation schemes, including FGSM, BIM, and PGD. As illustrated in Tab. 2, our ICTA realizes significant performance improvements across different white-box adversarial sample generation methods. On the one hand, our ICTA achieves more noticeable performance enhancement on large datasets TinyImageNet than on small datasets CIFAR10. On the other hand, ICTA has superior performance improvement on target attacks than non-target attacks. The results indicate that our ICTA more effectively simulates the decision space of the target model compared to traditional substitute training methods [9,21,27,35,37]. + +Comparison on Different Target Models. According to the results in Tab. 3, our ICTA demonstrates a significant performance improvement on different target models compared to the GANs-based approaches [9,27,35,37] and the diffusion-based method [21]. Specifically, ICTA achieves the most superior performance in target attacks, especially on TinyImageNet. Additionally, ICTA + +Table 3: Comparison on different target models. We choose PGD as the default white-box adversarial sample generation scheme. The best results are marked as bolded. + +
Dataset Target ModelCIFAR10CIFAR100TinyImageNet
VGG16ResNet34GoogleNetMobileNetResNet50ResNet50VGG19
Non-TargetMAZE [9]47.4968.9427.1351.7343.0715.5411.10
DaST [37]53.2964.2226.4857.9447.4016.119.81
DST [27]55.1571.2638.6465.8559.4632.5133.52
Zhang et al. [35]95.8497.5467.2497.7891.1878.8265.01
DiffusionLCA [21]97.0197.3094.8297.0196.4967.5670.58
ICTA (Ours)96.4899.1997.3299.6399.2594.8093.39
TargetMAZE [9]11.3916.6512.9811.0310.013.182.54
DaST [37]15.3918.615.319.5811.412.215.20
DST [27]17.0719.3318.8918.4624.658.555.87
Zhang et al. [35]57.4764.0539.8344.1154.7820.0115.01
DiffusionLCA [21]75.3979.9878.1851.6151.5012.4515.45
ICTA (Ours)79.5791.4084.5770.5674.4943.0233.41
+ +Table 4: Comparison of attack success rate in label-based and probability-based scenarios. The best results are marked as bolded. + +
methodProbability-basedLabel-based
cifar10cifar100cifar10cifar100
Non-TargetGANsDaST [37]64.2247.4067.7236.57
DST [27]71.2659.4680.4248.50
Zhang et al. [35]97.5491.1898.9890.20
DiffusionLCA [21]97.3096.4998.4894.55
ICTA (Ours)99.1999.2599.4998.53
TargetGANsDaST [37]18.6111.4119.285.20
DST [27]19.3324.6523.854.59
Zhang et al. [35]64.0554.7869.2540.17
DiffusionLCA [21]79.9851.5082.1849.85
ICTA (Ours)91.4074.4993.1061.71
+ +consistently enhances performance across different target models, showcasing the stability of the performance in various scenarios. In summary, our ICTA is able to simulate different target models in black-box attack scenarios and exhibits substantial performance improvements on large-scale datasets. + +Comparison in Probability-based and Label-based Scenarios. We compare the attack success rate in the probability-based and label-based scenarios, as displayed in Tab. 4. In the label-based scenario, only the output labels of the target model are accessible, while in the probability-based scenario, both the labels and probabilities of the target model are available. The experiment results indicate that our ICTA achieves the optimal performance in both probability-based and label-based scenarios and shows significant improvement over the existing substitute attack methods [21, 27, 35, 37]. This highlights the comprehensive superiority of our ICTA, demonstrating outstanding attack performance even in scenarios where only label outputs are accessible. + +![](images/2a75e89d1c9924e5f3372ea70f15115ca70d513532a468a743499effdd66d29f.jpg) +(a) Attack success rate + +![](images/5920324477c26a0d460724b5af97e45e144a5bdbf6f7ffb3736a08273d777f78.jpg) +(b) Accuracy of substitute model +Fig. 4: Comparison of attack success rate and accuracy in full-data scenario. Comparison schemes include the substitute training schemes DST [27] and Zhang et al. [35] as well as a distillation method DIST [8]. The ResNet50 on CIFAR100 is used as the target model. + +# 4.3 Comparative Experiments in Full-data Scenario + +Performance of Different Substitute Training Strategies. In Fig. 4, we present the attack success rate (both target and non-target attacks) and the accuracy of the substitute training during the training process. It can be observed that our ICTA demonstrates excellent performance in both target and nontarget attacks and shows higher efficiency in target attacks. There is a significant improvement in non-target attacks compared to conventional methods DST [27] and Zhang et al. [35], which highlights the unique advantages of our ICTA in improving the effectiveness of target attacks. In addition, we compare an excellent knowledge distillation scheme, DIST [8], which also shows the superior performance of our ICTA in attack scenarios. + +Notably, we find that the accuracy of the substitute model does not fully reflect the effectiveness of the substitute training. Despite having a relatively low accuracy in our ICTA, the substitute model exhibits excellent performance in attacks. This indicates that our ICTA better simulates the target model in terms of inter-class topology, rather than just replicating its outputs. + +Comparison of Relative Performance with White-box Attacks. We further compare the attack success rate of our ICTA with the corresponding white-box attacks, as demonstrated in Tab. 5, where the Relative Performance (RP) is used as a metric to measure the performance relative to the white-box attacks. Our ICTA achieves comparable performance to the corresponding white-box attacks in non-target attack, when the query budgets is set to 2M (for CIFAR10) and 3M (for CIFAR100 and TinyImageNet). Moreover, it is worth noting that the target attack success rate exceeds $83\%$ on the CIFAR10 dataset, $67\%$ on the CIFAR100 dataset, and $44\%$ on the TinyImageNet dataset compared to the corresponding white-box attack. As can be seen, the proposed ICTA achieves quite + +Table 5: Comparison with direct white-box attacks on target model. We utilize Relative Performance (RP) to evaluate the attack success of our black-box attack method relative to the white-box attack (in red). + +
ResNet34Vgg16GoogleNetShuffleNetV2
white-boxOursRPwhite-boxOursRPwhite-boxOursRPwhite-boxOursRP
CIFAR10Non-TargetFGSM78.5483.421.0685.5979.460.9363.8477.731.2284.9783.970.99
BIM100.0099.591.00100.0097.260.9799.9397.610.98100.0098.860.99
PGD100.0099.791.00100.0098.080.98100.0098.260.98100.0099.200.99
TargetFGSM28.7934.691.2019.6630.251.5424.5526.991.1032.9331.780.97
BIM99.8494.580.9584.8181.190.9699.5582.580.83100.0086.010.86
PGD100.0096.630.9798.2185.670.8799.9786.290.86100.0089.030.89
ResNet50WideResNetMobileNetV2ResNeXt50
CIFAR100Non-TargetFGSM84.7786.761.0274.2783.401.1290.1888.200.9887.2487.511.00
BIM100.0099.220.9998.1196.830.9999.9899.090.9999.9899.160.99
PGD100.0099.500.9999.5797.540.98100.0099.460.99100.0099.470.99
TargetFGSM3.058.522.793.879.422.434.397.351.673.838.792.30
BIM95.2272.790.7689.4465.900.7496.3064.790.6797.0673.290.76
PGD99.2579.340.8098.3873.220.7499.6670.330.7199.7080.150.80
ResNet50Vgg19MobileNetV2ResNeXt50
TinyImageNetNon-TargetFGSM95.8286.070.9092.8286.740.9398.4288.000.8996.6283.650.87
BIM99.9796.260.9699.6495.940.96100.0097.210.9799.9894.780.95
PGD99.9896.960.9799.9196.560.97100.0097.780.9899.9895.860.96
TargetFGSM10.2510.040.984.797.991.6715.718.740.5613.188.680.66
BIM98.9652.300.5385.6639.310.4699.7245.920.4699.5343.560.44
PGD99.7457.850.5897.8146.310.4799.9251.180.5199.9650.300.50
+ +satisfactory performance in black-box attack scenarios, demonstrating robust adaptability in target attacks across different datasets. + +# 4.4 Analysis of Decision Boundary Similarity + +# The t-SNE Visualization of Decision Space. We + +exhibit the t-Distributed Stochastic Neighbor Embedding (t-SNE) visualization of the decision space of the models, as exhibited in Fig. 5. In t-SNE visualization, each data point is typically labeled with its corresponding class. Therefore, it can depict the interclass relative position in the decision space, as well as the inter-class topology. The visualization illustrates that the inter-class topology in the decision space of the substitute model trained by our ICTA + +![](images/858835a3adecdc067bdc922f7096def4005dcf635c8895ad22e3101213641a91.jpg) +Fig. 5: The t-SNE visualization of the decision space for (a) the target model, (b) our ICTA, (c) Zhang et al. [35] and (d) DST [27]. + +![](images/f555365d8168529095cbbd02daa21736f37dbab9611910fc00331bb0728ad1e3.jpg) + +![](images/dde4a897cb1c2ab0fc82ec9083b1f4c263a684303c331ef937b98f483f520ad4.jpg) + +![](images/4d2c5c7420637432cbb064d5533859f023b02a8944afe5a2ebc0f095488de350.jpg) + +![](images/2dbc627c5a13e3cd73661b1b7ae055e3d1bdd9f361a0a3d5b9e81f17580a9931.jpg) +(a) Attack success rate + +![](images/43060eba2b4b6e0c13b209fddffb14ee2b1b88839822769456a3a46907a24a47.jpg) +(b) Accuracy of substitute model +Fig. 6: Ablation experiments on PES. The experiments include baseline (no relative position loss) and strategies for exploring inter-class relative position through traditional adversarial samples, clean samples, and position exploration sample. + +method is remarkably similar to the topology of the target model. On the contrary, both the output alignment-focused DST [27] and the decision boundary-constrained approach by Zhang et al. [35] exhibit greater differences from the target model in t-SNE visualization. This confirms that our ICTA is more effective in aligning the substitute model to match the target model. + +Quantitative Analysis. We explore random directions of arbitrary samples, and then quantitatively assess the Decision Boundary Similarity (DBS) based on the overlap between target and substitute models. As in Tab. 6, our ICTA has more consistent decision boundaries with the target model. + +Table 6: Quantitative analysis of decision boundary similarity. + +
MethodICTA (Ours)Zhang et al.DST
CIFAR100.930.870.80
CIFAR1000.880.770.71
TinyImageNet0.820.570.55
+ +# 4.5 Ablation Study + +Ablation on Position Exploration Sample. In this paper, we design PES to explore the inter-class relative position in the decision space through their relative position relationships with respect to the original clean samples. However, as the position of any two samples in the decision space inherently depict the inter-class relative position, traditional adversarial sample or even a clean sample has the effect of exploring the decision space. In this section, we validate the effectiveness of PES by comparing it with traditional adversarial sample generated by BIM and clean sample, as illustrated in Fig. 6. + +It is worth noting that while all three approaches contribute to explore the inter-class relative position in the decision space, PES stands out as particularly advantageous. This is due to the fact that both clean samples and the adversarial + +Table 7: Ablation comparison of the loss function and the distance metric. We select two distance metrics, MSE and DIST [8], and use different combinations of distance metrics in the loss function. The best result are marked as bolded, and the sub-optimal results are marked as underlined. + +
DkdMSE DISTDrpMSE DISTTargetNon-Target
FGSMBIMPGDFGSMBIMPGD
✓4.3142.1446.1479.4895.3796.38
✓4.1155.5257.6782.8598.3098.80
✓✓7.6464.2071.7084.3598.4399.06
✓✓7.8273.4279.0986.7399.0499.61
✓✓8.5272.7979.3486.7699.2299.50
+ +samples generated by BIM are closer to the decision center, resulting in limited exploration in the decision space. In contrast, the proposed PES is more diverse by employing the full-gradient and unconstrained setting, enabling exploration across a wider range of the decision space. The ablation experiments for the parameters of PES and the number of PES are provided in the supplement. + +Ablation on Loss and Metrics. We conduct an ablation study on the loss functions and distance metrics of substitute training in Tab. 7, focusing on the distillation loss $L_{KD}$ and relative position loss $L_{RP}$ . Here, the term 'DIST' refers to the distance metric function in the distillation scheme DIST [8]. It can be seen that when DIST [8] is used as the distance metric in $L_{KD}$ , there is a certain improvement in the attack success rate. And when $L_{RP}$ is added, there is a significant improvement in the attack success rate, which underscores the indispensability of the relative position loss. Notably, our ICTA employs DIST as the distance metric in both losses, as it further improves attack performance. + +# 5 Conclusion + +In this paper, we propose a novel ICTA scheme to better align the substitute model to the target model for efficient black-box substitute attacks. Compared to the existing schemes that only align intra-class consistency, we simulate the target model in the decision space more comprehensively through inter-class topology alignment. Specifically, we first explore the inter-class relative position of the target model in the decision space by the designed PES. Subsequently, we achieve inter-class topology alignment of the two models by constraining the inter-class relative position of the substitute model in different directions. Benefiting from the better alignment of the two models in the decision space, the adversarial samples generated by the substitute model are more successful in misleading the target model to classify them into the target class. Experimental results demonstrate the effectiveness of ICTA, which achieves better simulation of the target model and significantly improves the attack success rate beyond existing substitute attack approaches. + +# Acknowledgements + +This work was supported by the National Key Research and Development Program of China (Grant No. 2021YFA1000102), the National Natural Science Foundation of China (Grant Nos. 62376285, 62272375, and 61673396), and Natural Science Foundation of Shandong Province, China (Grant No. ZR2022MF260). + +# References + +1. Andriushchenko, M., Croce, F., Flammarion, N., Hein, M.: Square Attack: a query-efficient black-box adversarial attack via random search. In: Proceedings of the European Conference on Computer Vision. pp. 484-501 (2020) +2. Brendel, W., Rauber, J., Bethge, M.: Decision-Based Adversarial Attacks: Reliable Attacks Against Black-Box Machine Learning Models. In: International Conference on Learning Representations. arXiv preprint arXiv:1712.04248 (2018) +3. Chen, S., He, Z., Sun, C., Yang, J., Huang, X.: Universal Adversarial Attack on Attention and the Resulting Dataset DImageNet. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(4), 2188-2197 (2022) +4. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: International Conference on Machine Learning. vol. 119, pp. 2206-2216 (2020) +5. Goodfellow, I., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014) +6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016) +7. Huang, H., Chen, Z., Chen, H., Wang, Y., Zhang, K.: T-SEA: Transfer-based Self-Ensemble Attack on Object Detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20514-20523 (2023) +8. Huang, T., You, S., Wang, F., Qian, C., Xu, C.: Knowledge distillation from a stronger teacher. In: Advances in Neural Information Processing Systems. vol. 35, pp. 33716-33727 (2022) +9. Kariyappa, S., Prakash, A., Qureshi, M.K.: Maze: Data-free model stealing attack using zeroth-order gradient estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13814-13823 (2021) +0. Krizhevsky, A.: Learning Multiple Layers of Features from Tiny Images. Master's thesis, University of Tront (2009) +1. Kurakin, A., Goodfellow, I., Bengio, S.: Adversarial examples in the physical world. arXiv preprint arXiv:1607.02533 (2016) +2. Lee, B.K., Kim, J.: Mitigating Adversarial Vulnerability through Causal Parameter Estimation by Adversarial Double Machine Learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4499-4509 (2023) +3. Li, L., Lian, Q., Chen, Y.C.: Adv3D: Generating 3D Adversarial Examples in Driving Scenarios with NeRF. arXiv preprint arXiv:2309.01351 (2023) +4. Liu, J., Lu, B., Xiong, M., Zhang, T., Xiong, H.: Adversarial Attack with Raindrops. arXiv preprint arXiv:2302.14267 (2023) +5. Ma, N., Zhang, X., Zheng, H.T., Sun, J.: ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design. In: Proceedings of the European Conference on Computer Vision. pp. 122-138 (2018) + +16. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2018) +17. Ren, K., Zheng, T., Qin, Z., Liu, X.: Adversarial Attacks and Defenses in Deep Learning. Engineering 6(3), 346-360 (2020) +18. Robin, R., Andreas, B., Dominik, L., Patrick, E., Bj\''orn, O.: High-Resolution Image Synthesis with Latent Diffusion Models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10674-10685 (2022) +19. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision 115(3), 211-252 (2015) +20. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: MobileNetV2: Inverted Residuals and Linear Bottlenecks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4510-4520 (2018) +21. Shao, M., Meng, L., Qiao, Y., Zhang, L., Zuo, W.: Data-free black-box attack based on diffusion model. arXiv preprint arXiv:2307.12872 (2023) +22. Sheng, L., Liang, J., He, R., Wang, Z., Tan, T.: AdaptGuard: Defending Against Universal Attacks for Model Adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19093-19103 (2023) +23. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) +24. Su, J., Vargas, D.V., Sakurai, K.: One Pixel Attack for Fooling Deep Neural Networks. IEEE Transactions on Evolutionary Computation 23(5), 828-841 (2019) +25. Szegedy, C., Wei Liu, Yangqing Jia, Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1-9 (2015) +26. Wang, N., Luo, Y.: Does Physical Adversarial Example Really Matter to Autonomous Driving? Towards System-Level Effect of Adversarial Object Evasion Attack. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4412-4423 (2023) +27. Wang, W., Qian, X., Fu, Y., Xue, X.: Dst: Dynamic substitute training for data-free black-box attack. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14361-14370 (2022) +28. Wang, W., Yin, B., Yao, T., Zhang, L., Fu, Y., Ding, S., Li, J., Huang, F., Xue, X.: Delving into data: Effectively substitute training for black-box attack. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4761-4770 (2021) +29. Wang, X., Zhang, Z.: Structure Invariant Transformation for better Adversarial Transferability. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4607-4619 (2023) +30. Wang, X., Zhang, Z., Tong, K., Gong, D., He, K., Li, Z., Liu, W.: Triangle Attack: A Query-efficient Decision-based Adversarial Attack. In: Proceedings of the European Conference on Computer Vision. pp. 156-174 (2022) +31. Wang, Z., Yang, H., Feng, Y., Sun, P., Guo, H., Zhang, Z., Ren, K.: Towards Transferable Targeted Adversarial Examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) +32. Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K.: Aggregated Residual Transformations for Deep Neural Networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5987-5995 (2017) + +33. Yu, M., Sun, S.: FE-DaST: Fast and effective data-free substitute training for black-box adversarial attacks. Computers & Security 113, 102555 (2022) +34. Zagoruyko, S., Komodakis, N.: Wide Residual Networks. In: Proceedings of the British Machine Vision Conference. pp. 87.1-87.12 (2016) +35. Zhang, J., Li, B., Xu, J., Wu, S., Ding, S., Zhang, L., Wu, C.: Towards efficient data free black-box adversarial attack. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15094-15104 (2022) +36. Zheng, X., Fan, Y., Wu, B., Zhang, Y., Wang, J., Pan, S.: Robust Physical-World Attacks on Face Recognition. Pattern Recognition 133, 109009 (2023) +37. Zhou, M., Wu, J., Liu, Y., Liu, S., Zhu, C.: Dast: Data-free substitute training for adversarial attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 234-243 (2020) +38. Zhu, Y., Chen, Y., Li, X., Chen, K., He, Y., Tian, X., Zheng, B., Chen, Y., Huang, Q.: Toward Understanding and Boosting Adversarial Transferability From a Distribution Perspective. IEEE Transactions on Image Processing 31, 6487-6501 (2022) \ No newline at end of file diff --git a/2024/Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks/images.zip b/2024/Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2cb5b82344e969840360a325a9e53a1f42d79bf1 --- /dev/null +++ b/2024/Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e084e45ba0dad72c635fd38c3d6d0ba8c8ba7c1ffc262eb1223f99abcaffa6b +size 773653 diff --git a/2024/Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks/layout.json b/2024/Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..155ab5bfab5704d2f580e41913a50312afd2e1a7 --- /dev/null +++ b/2024/Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks/layout.json @@ -0,0 +1,8485 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 152, + 111, + 462, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 111, + 462, + 148 + ], + "spans": [ + { + "bbox": [ + 152, + 111, + 462, + 148 + ], + "type": "text", + "content": "Inter-Class Topology Alignment for Efficient Black-Box Substitute Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 134, + 168, + 479, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 168, + 479, + 181 + ], + "spans": [ + { + "bbox": [ + 134, + 168, + 479, + 181 + ], + "type": "text", + "content": "Lingzhuang Meng" + }, + { + "bbox": [ + 134, + 168, + 479, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 168, + 479, + 181 + ], + "type": "text", + "content": ", Mingwen Shao" + }, + { + "bbox": [ + 134, + 168, + 479, + 181 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 134, + 168, + 479, + 181 + ], + "type": "text", + "content": ", Yuanjian Qiao" + }, + { + "bbox": [ + 134, + 168, + 479, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 168, + 479, + 181 + ], + "type": "text", + "content": ", and Wenjie Liu" + }, + { + "bbox": [ + 134, + 168, + 479, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 139, + 190, + 474, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 190, + 474, + 201 + ], + "spans": [ + { + "bbox": [ + 139, + 190, + 474, + 201 + ], + "type": "text", + "content": "School of Computer Science and Technology, China University of Petroleum (East" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 241, + 201, + 373, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 201, + 373, + 212 + ], + "spans": [ + { + "bbox": [ + 241, + 201, + 373, + 212 + ], + "type": "text", + "content": "China), Qingdao, 266580, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 171, + 213, + 440, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 213, + 440, + 223 + ], + "spans": [ + { + "bbox": [ + 171, + 213, + 440, + 223 + ], + "type": "text", + "content": "lzhmeng1688@163.com, smw278@126.com, yjqiao@s.upc.edu.cn," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 258, + 224, + 355, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 224, + 355, + 234 + ], + "spans": [ + { + "bbox": [ + 258, + 224, + 355, + 234 + ], + "type": "text", + "content": "liuwenjie_82@163.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 264, + 455, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 264, + 455, + 495 + ], + "spans": [ + { + "bbox": [ + 160, + 264, + 455, + 495 + ], + "type": "text", + "content": "Abstract. In black-box attacks based on substitute training, the similarity of the substitute model to the target model is critical for successful attacks. However, existing schemes merely train the substitute model to mimic the outputs of the target model without fully simulating the decision space, resulting in the adversarial samples generated by the substitute model being classified into the non-target class by the target model. To alleviate this issue, we propose a novel Inter-Class Topology Alignment (ICTA) scheme to more comprehensively simulate the target model by aligning the inter-class positional relationships of two models in the decision space. Specifically, we first design the Position Exploration Sample (PES) to more thoroughly explore the relative positional relationships between classes in the decision space of the target model. Subsequently, we align the inter-class topology between the two models by utilizing the PES to constrain the inter-class relative position of the substitute model in different directions. In this way, the substitute model is more consistent with the target model in the decision space, so that the generated adversarial samples will be more successful in misleading the target model to classify them into the target class. The experimental results demonstrate that our ICTA significantly improves attack success rate in various scenarios compared to existing substitute training methods, particularly performing efficiently in target attacks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "spans": [ + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "text", + "content": "Keywords: Black-box attacks " + }, + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "text", + "content": " Substitute attacks " + }, + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "text", + "content": " Position exploration sample " + }, + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "text", + "content": " Inter-class topology alignment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 548, + 230, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 230, + 561 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 230, + 561 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": "Convolutional Neural Networks (CNNs) have shown impressive performance in various fields and are widely applied in daily life. However, it has been found that CNNs exhibit poor robustness to well-designed perturbations [3, 36, 38], which poses a great threat to their safety and reliability. For example, in the field of autonomous driving, if the network makes a wrong judgment, it will lead to potential safety hazards [13, 14, 26]. Therefore, it is especially urgent to" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 234, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 234, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 234, + 666 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 478, + 229 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 478, + 229 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 478, + 229 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 478, + 229 + ], + "type": "image", + "image_path": "bdcc90cce1ae60b4488e67de5b5fff51696c7e9b51cff2b913366763d0fc8a33.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 238, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 130, + 238, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 482, + 316 + ], + "type": "text", + "content": "Fig. 1: (a) Comparison of different methods and (b) Attack success rate. Traditional approaches only focus on intra-class consistency, resulting in the adversarial samples generated by the substitute model failing to mislead the target model to classify them into the target class. In contrast, our ICTA better simulates the target model by aligning the inter-class topology in the decision space, resulting in a successful target attack. The performance of our ICTA is significantly improved across multiple datasets, particularly in target attacks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 340, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 480, + 365 + ], + "type": "text", + "content": "research deeply into the adversarial attacks [4,30,38] and defense [17,22] schemes to enhance the reliability and security of CNNs in practical applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 496 + ], + "type": "text", + "content": "Existing adversarial attacks can be categorized into white-box and black-box attacks according to the knowledge of the attacker about the target model. In white-box attacks [3-5,11,12,16,29], the attacker possesses complete information about the target model (such as network structure, gradient, etc.), thus can generate highly deceptive adversarial samples based on this information. By contrast, in black-box attacks [2,9,21,24,31,35], the attacker can only utilize the outputs of the network (probabilities or labels) to craft adversarial samples, making the task more challenging and realistic. In such scenario, the substitute attacks [9,27,33,35,37] is proposed, which focus on training a substitute model to emulate the target model, so that the adversarial samples generated by the substitute model can effectively attack the target model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Current substitute attack methods [9, 21, 27, 28, 33, 35, 37] adhere to the paradigm of knowledge distillation, encouraging the substitute model to mimic the outputs of the target model. Some schemes utilized simple " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " loss [21, 28, 33, 37] or KL-divergence [9] to constrain the substitute model. As an improvement, DST [27] leveraged a graph structure consisting of multiple outputs to optimize the substitute model. Nevertheless, these strategies only focus on learning the simplistic outputs of the target model, without adequately considering the decision space of the network. Subsequently, Zhang et al. [35] considered aligning the decision boundary between the substitute model and the target model to bridge the gap between the two models. However, this restriction of decision boundary only aims at the intra-class decision consistency and ignores the inter-class relationship consistency of the two models in the decision space. As a result, the adversarial samples generated by the substitute model are difficult to be classified as target class by the target model, as Fig. 1(a)." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "L. Meng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": "To alleviate the above issue, we propose a novel Inter-Class Topology Alignment (ICTA) scheme to align the inter-class positional relationships of the two models in decision space, thus better simulating the target model and achieving more accurate target attacks. Specifically, we first design the Position Exploration Sample (PES) to explore the relative positional relationships among all classes of the target model by introducing adversarial perturbations of different directions and strengths. Subsequently, we constrain the inter-class positional consistency of the substitute model to align the inter-class topology between the two models. Thanks to the better alignment of the two models in the decision space, the adversarial samples generated by the substitute model can be more successful deceiving the target model to classify them into the target class, as shown in Fig. 1(a). Experiments demonstrate that our ICTA achieves higher attack success rate on different datasets, and especially more pronounced in target attacks, as Fig. 1(b). Our primary contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 291, + 479, + 396 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 291, + 479, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 291, + 479, + 324 + ], + "spans": [ + { + "bbox": [ + 138, + 291, + 479, + 324 + ], + "type": "text", + "content": "- We propose ICTA to innovatively align the substitute model with the target model in terms of inter-class topology in the decision space, which effectively improves the success rate of black-box substitute attack." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 327, + 479, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 327, + 479, + 360 + ], + "spans": [ + { + "bbox": [ + 137, + 327, + 479, + 360 + ], + "type": "text", + "content": "- We propose a novel PES to more comprehensively explore the inter-class relative position of the target model, which is utilized to align the inter-class topology between the substitute model and the target model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 363, + 479, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 363, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 137, + 363, + 479, + 396 + ], + "type": "text", + "content": "- Experiments illustrate that our ICTA achieves significant performance gains in a variety of scenarios and outperforms on target attacks compared to the existing substitute attack schemes." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 416, + 242, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 416, + 242, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 416, + 242, + 429 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 441, + 253, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 253, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 253, + 453 + ], + "type": "text", + "content": "2.1 White-box attacks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 461, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 569 + ], + "type": "text", + "content": "In white-box attacks, the attacker is fully aware of the information about the target model, such as the structure and parameters [5,11,12,16], and highly deceptive adversarial samples can be generated based on this information. Classical white-box adversarial sample generation schemes utilized gradient information of network to generate adversarial samples, such as FGSM [5], BIM [11], and PGD [16]. Moreover, Croce et al. [4] proposed parameter-free attack combination AutoAttack by integrating multiple attack methods for testing adversarial robustness. However, white-box attacks may lack practicality in real-world scenarios due to the requirement for extensive knowledge about the target model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 586, + 250, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 250, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 250, + 597 + ], + "type": "text", + "content": "2.2 Black-box attacks" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "In contrast, black-box attacks are more challenging and more in line with realistic scenarios, as the attacker can generate adversarial samples only using the outputs of the target model, including transfer attacks, query attacks, and substitute attacks. Transfer attacks [1, 3, 7, 29, 31] aim to maximize the transferability of existing adversarial samples, enabling them to successfully attack the target" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "text", + "content": "Topology Alignment for Substitute Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "model. For example, Zhu et al. [38] proposed modifying the distribution of the original images to match the distribution of the target class, thereby obtaining adversarial samples with high transferability. Query attacks [2, 24, 30] obtain adversarial samples by continuously querying the target model using specific inputs until a desired output is obtained. Substitute attacks [21, 35, 37] focus on training a substitute model that is similar to the target model, so that the adversarial samples generated by the substitute model can successfully attack the target model. Compared to other schemes, the substitute attack is able to generate more effective adversarial samples due to the ability to simulate the decision process of the target model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 451 + ], + "type": "text", + "content": "Most of the existing substitute attack schemes used " + }, + { + "bbox": [ + 130, + 236, + 482, + 451 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 236, + 482, + 451 + ], + "type": "text", + "content": " loss [21, 28, 33, 37] or KL-divergence [9] as the constraint to train substitute models. For example, DaST [37] used Generative Adversarial Networks (GANs) to generate data for substitute training and leveraged the " + }, + { + "bbox": [ + 130, + 236, + 482, + 451 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 236, + 482, + 451 + ], + "type": "text", + "content": " loss as a constraint to align the outputs of the substitute model with the target model. Kariyappa et al. [9] optimized the generator with the help of zeroth-order gradient estimate and trained the substitute training model with a simple KL-divergence. Shao et al. [21] proposed LCA to utilize a diffusion model to generate the data required for substitute training, which greatly improved the training efficiency compared to traditional GANs-based methods. However, the target attack success rate is not impressive as it only used " + }, + { + "bbox": [ + 130, + 236, + 482, + 451 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 236, + 482, + 451 + ], + "type": "text", + "content": " loss to constrain the substitute model. In order to simulate the target model more precisely, Wang et al. [27] proposed DST to optimize the substitute model from a graph with multiple outputs of the target model instead of a single output. Nevertheless, these schemes only learn the simple outputs of the target model without fully considering the complex decision space attributes, resulting in poor performance. Subsequently, Zhang et al. [35] encouraged a high alignment of decision boundary between the substitute model and the target model, which improved the attack success rate." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 453, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 525 + ], + "type": "text", + "content": "However, the aforementioned schemes only focused on the intra-class consistency and neglected the inter-class relational consistency, resulting in the adversarial samples generated by the substitute model being difficult to be classified as target class by the target model. In this paper, we further consider the inter-class positional relationships in the decision space and better simulate the target model by aligning the inter-class topology for a more effective target attack." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 547, + 233, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 233, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 233, + 562 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 576, + 222, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 222, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 222, + 590 + ], + "type": "text", + "content": "3.1 Preliminary" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 600, + 481, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 481, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 481, + 638 + ], + "type": "text", + "content": "For a black-box target model " + }, + { + "bbox": [ + 130, + 600, + 481, + 638 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 600, + 481, + 638 + ], + "type": "text", + "content": ", existing research on substitute attacks typically trains a substitute model " + }, + { + "bbox": [ + 130, + 600, + 481, + 638 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 600, + 481, + 638 + ], + "type": "text", + "content": " similar to the " + }, + { + "bbox": [ + 130, + 600, + 481, + 638 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 600, + 481, + 638 + ], + "type": "text", + "content": " through a knowledge distillation paradigm. Formally, the substitute model can be obtained by substitute training:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 247, + 651, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 651, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 247, + 651, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\underset {\\theta} {\\arg \\min } \\mathbb {E} _ {x \\sim \\mathcal {X}} d [ S (x), T (x) ], \\tag {1}", + "image_path": "d2156bed2963ae618a5576f935b9ceee9e7ab4d66ce5a2ca25a0f51239079d95.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "L. Meng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " denotes the distance metric function, " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "x \\sim \\mathcal{X}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " denotes the training data, and " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " is the parameters of " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ". The goal of substitute training is to optimize the substitute model to resemble the target model as similar as possible." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 271, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 271, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 271, + 343 + ], + "type": "text", + "content": "Subsequently, the adversarial samples can be generated using white-box adversarial sample generation methods based on the substitute model. We take BIM [11] as an instance, which perturbs the input data along the gradient direction of the target class through multiple iterations to make the output of the model close to the target class in the decision space, as Fig. 2. The generation of adversarial samples for target attacks can be represented as follows:" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 282, + 176, + 479, + 250 + ], + "blocks": [ + { + "bbox": [ + 282, + 176, + 479, + 250 + ], + "lines": [ + { + "bbox": [ + 282, + 176, + 479, + 250 + ], + "spans": [ + { + "bbox": [ + 282, + 176, + 479, + 250 + ], + "type": "image", + "image_path": "804353b81756ac88ffbd8ccde680b1e847d06b4ca4de33b509c58dd0a63b686f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 258, + 482, + 335 + ], + "lines": [ + { + "bbox": [ + 276, + 258, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 276, + 258, + 482, + 335 + ], + "type": "text", + "content": "Fig. 2: Position exploration sample vs. Traditional adversarial sample. Traditional adversarial samples are closer to and clustered at the decision center after several iterations of minor adjustments. While our PES has a much broader coverage and is able to explore the decision space more comprehensively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 166, + 352, + 480, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 352, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 166, + 352, + 480, + 369 + ], + "type": "interline_equation", + "content": "x _ {0} ^ {a d v} = x, x _ {N + 1} ^ {a d v} = C l i p _ {x, \\epsilon} \\left\\{x _ {N} ^ {a d v} - \\alpha s i g n \\left(\\nabla_ {x} L \\left(S \\left(x _ {N + 1} ^ {a d v}\\right), y _ {t a r}\\right)\\right) \\right\\}, \\tag {2}", + "image_path": "2a643be9cfd5a5aaad98f3cf2bb6fc10cf49a1d3034d8fc07e1bc2e6a7b1ab0b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 376, + 479, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 479, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 479, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 376, + 479, + 411 + ], + "type": "inline_equation", + "content": "y_{tar}" + }, + { + "bbox": [ + 130, + 376, + 479, + 411 + ], + "type": "text", + "content": " is the target class, " + }, + { + "bbox": [ + 130, + 376, + 479, + 411 + ], + "type": "inline_equation", + "content": "\\nabla_x" + }, + { + "bbox": [ + 130, + 376, + 479, + 411 + ], + "type": "text", + "content": " denotes the gradient of the substitute model, " + }, + { + "bbox": [ + 130, + 376, + 479, + 411 + ], + "type": "inline_equation", + "content": "sign" + }, + { + "bbox": [ + 130, + 376, + 479, + 411 + ], + "type": "text", + "content": " denotes the sign function, " + }, + { + "bbox": [ + 130, + 376, + 479, + 411 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 376, + 479, + 411 + ], + "type": "text", + "content": " represents the step size during the iteration process, and " + }, + { + "bbox": [ + 130, + 376, + 479, + 411 + ], + "type": "inline_equation", + "content": "Clip" + }, + { + "bbox": [ + 130, + 376, + 479, + 411 + ], + "type": "text", + "content": " is the clip function based on the perturbation boundary " + }, + { + "bbox": [ + 130, + 376, + 479, + 411 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 376, + 479, + 411 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 411, + 479, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 479, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 479, + 448 + ], + "type": "text", + "content": "Finally, the generated adversarial samples are used to attack the target model. Clearly, the similarity between the substitute model and the target model is a key factor for successful attacks in substitute attacks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 465, + 443, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 465, + 443, + 478 + ], + "spans": [ + { + "bbox": [ + 131, + 465, + 443, + 478 + ], + "type": "text", + "content": "3.2 Aligning the Inter-Class Topology in the Decision Space" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 486, + 480, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 480, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 480, + 594 + ], + "type": "text", + "content": "Existing substitute training methods encourage the substitute model to simply mimic the outputs of the target model without fully simulating the decision space of the target model. Resulting in the adversarial samples generated by the substitute model are difficult to be classified into the target class by the target model. Moreover, there is also a lack of research on what aspects the substitute model needs to resemble the target model. In this paper, we propose to understand and improve the substitute training from the perspective of the inter-class topology in the decision space, which refers to the positional relationships (including directions and distances) among all classes in the decision space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "content": "According to Eq. (2), modifying the adversarial samples along the direction of the gradient " + }, + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\nabla_{x}" + }, + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "content": " in the decision space can bring the decision of the model closer to the target class, as shown in Fig. 2. Therefore, if the relative positions of all classes are consistent between the substitute model and the target model, then the adversarial samples generated by the substitute model can be successfully classified into the target class by the target model. This suggests that aligning" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 271, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 271, + 91, + 448, + 102 + ], + "type": "text", + "content": "Topology Alignment for Substitute Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 476, + 231 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 476, + 231 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 476, + 231 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 476, + 231 + ], + "type": "image", + "image_path": "fe668db758a260ad7998573604edf96c783bf422356bc2f6f5c918caeb60661f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 239, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 285 + ], + "type": "text", + "content": "Fig. 3: The framework of our Inter-Class Topology Alignment (ICTA). ICTA explores the inter-class relative position in the decision space using position exploration sample in different directions, and aligns the inter-class topology of the two models by fitting the inter-class relative position." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 309, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 482, + 332 + ], + "type": "text", + "content": "the inter-class topology between the two models in the decision space is crucial for substitute attacks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 334, + 482, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 381 + ], + "type": "text", + "content": "For the sample " + }, + { + "bbox": [ + 130, + 334, + 482, + 381 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 334, + 482, + 381 + ], + "type": "text", + "content": ", we consider " + }, + { + "bbox": [ + 130, + 334, + 482, + 381 + ], + "type": "inline_equation", + "content": "S(x_{i})" + }, + { + "bbox": [ + 130, + 334, + 482, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 334, + 482, + 381 + ], + "type": "inline_equation", + "content": "T(x_{i})" + }, + { + "bbox": [ + 130, + 334, + 482, + 381 + ], + "type": "text", + "content": " as nodes in the two decision spaces, and " + }, + { + "bbox": [ + 130, + 334, + 482, + 381 + ], + "type": "inline_equation", + "content": "e_{S_{ij}} = \\langle S(x_i), S(x_j) \\rangle" + }, + { + "bbox": [ + 130, + 334, + 482, + 381 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 334, + 482, + 381 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 130, + 334, + 482, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 334, + 482, + 381 + ], + "type": "inline_equation", + "content": "e_{T_{ij}} = \\langle T(x_i), T(x_j) \\rangle" + }, + { + "bbox": [ + 130, + 334, + 482, + 381 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 334, + 482, + 381 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 130, + 334, + 482, + 381 + ], + "type": "text", + "content": " to represent the edges of the topology in the two decision spaces. Topology consistency can be achieved by aligning all nodes and edges:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 199, + 391, + 481, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 391, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 199, + 391, + 481, + 422 + ], + "type": "interline_equation", + "content": "\\min \\sum_ {i = 1} ^ {n} \\mathcal {D} _ {p} \\left(S \\left(x _ {i}\\right), T \\left(x _ {i}\\right)\\right) + \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {n} \\mathcal {D} _ {e} \\left(e _ {S _ {i j}}, e _ {T _ {i j}}\\right), \\tag {3}", + "image_path": "4fdd807d78b50ad1604b6a8405bbc7a783d9fe4fe935ebfd5c31cc53f67fdf46.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 432, + 479, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 479, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 479, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 432, + 479, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_p" + }, + { + "bbox": [ + 130, + 432, + 479, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 432, + 479, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_e" + }, + { + "bbox": [ + 130, + 432, + 479, + 445 + ], + "type": "text", + "content": " denote the distance metrics of points and edges, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 445, + 482, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 480 + ], + "type": "text", + "content": "However, aligning the edges among all the samples in the decision space is extremely complex. Therefore, we propose a relative position exploration and alignment strategy for matching the inter-class topology of the two models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 499, + 306, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 499, + 306, + 513 + ], + "spans": [ + { + "bbox": [ + 131, + 499, + 306, + 513 + ], + "type": "text", + "content": "3.3 Position Exploration Sample" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": "According to Fig. 3, the relative position of the clean and adversarial samples in the decision space can similarly be used to represent the relative position of the two classes in the decision space. In this case, the adversarial perturbations used to generate the adversarial samples can reflect the distance and direction between the two classes. Therefore, under the assumption that the target model is white-boxed, the inter-class relative position can be estimated using the positional relationships between the clean and adversarial samples." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "The traditional adversarial samples aim to progressively bring the sample closer to the decision center of the target class through multiple iterations of minor adjustments. While the sample has a relatively high success rate in the attack, it does not benefit for thoroughly exploring the entire decision space. To alleviate this limitation, we design the Position Exploration Sample (PES)" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "L. Meng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": "to more comprehensively explore the inter-class relative position of the target model in the decision space, which can be represented as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 243, + 149, + 481, + 162 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 149, + 481, + 162 + ], + "spans": [ + { + "bbox": [ + 243, + 149, + 481, + 162 + ], + "type": "interline_equation", + "content": "x _ {p e s} = x - \\phi \\nabla_ {x} L (T (x), y _ {t a r}), \\tag {4}", + "image_path": "b6c03954c04f9b8115606c94cdb1111d757c28154c1aec5a0cb74bd9e147959e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 169, + 479, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 169, + 479, + 253 + ], + "spans": [ + { + "bbox": [ + 130, + 169, + 479, + 253 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 169, + 479, + 253 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 169, + 479, + 253 + ], + "type": "text", + "content": " is the perturbation strength. We generate the PES using the full gradient, not just the gradient direction, without iteration and without being constrained by the pixel range. Thus, the gradient " + }, + { + "bbox": [ + 130, + 169, + 479, + 253 + ], + "type": "inline_equation", + "content": "\\nabla_{x}" + }, + { + "bbox": [ + 130, + 169, + 479, + 253 + ], + "type": "text", + "content": " and the perturbation strength " + }, + { + "bbox": [ + 130, + 169, + 479, + 253 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 169, + 479, + 253 + ], + "type": "text", + "content": " determine its direction and distance relative to the clean sample in the decision space. We represent the inter-class relative position based on the relative position of the PES to the clean sample and explore the decision space more thoroughly over a wider coverage by controlling " + }, + { + "bbox": [ + 130, + 169, + 479, + 253 + ], + "type": "inline_equation", + "content": "y_{tar}" + }, + { + "bbox": [ + 130, + 169, + 479, + 253 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 169, + 479, + 253 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 169, + 479, + 253 + ], + "type": "text", + "content": ", as Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 253, + 480, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 253, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 480, + 289 + ], + "type": "text", + "content": "However, it is impossible to generate the PES directly due to the gradient of the target model is inaccessible. Therefore, we retreat to generating PES using the substitute model, as shown in Fig. 3, which can be represented as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 243, + 298, + 481, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 298, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 243, + 298, + 481, + 311 + ], + "type": "interline_equation", + "content": "x _ {p e s} = x - \\phi \\nabla_ {x} L (S (x), y _ {a r b}), \\tag {5}", + "image_path": "fd8e7ebc92a2bfbb665cc4ba75fb747b206a845bd2c3f2988d8f66fc0dcd9fb5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 318, + 479, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 479, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 479, + 378 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 318, + 479, + 378 + ], + "type": "inline_equation", + "content": "y_{arb}" + }, + { + "bbox": [ + 130, + 318, + 479, + 378 + ], + "type": "text", + "content": " represents an arbitrary class different from ground truth. In this way, we can explore the decision space from different directions and obtain the complete inter-class topology in the decision space. Moreover, the PES generated by the substitute model has an effect on exploring the decision space of the target model, since our goal is to make the two models as similar as possible." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 394, + 339, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 394, + 339, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 394, + 339, + 407 + ], + "type": "text", + "content": "3.4 Topology Alignment Based on PES" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 413, + 479, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 479, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 479, + 485 + ], + "type": "text", + "content": "In order to align the inter-class topology of the two models, we simplify the topology consistency in Eq. (3) to the consistency of the inter-class relative position. This method allows us to capture the inter-class topology through the position of two nodes without the need to optimize the edges directly, as demonstrated in Fig. 3. We can achieve this alignment by simultaneously optimizing the knowledge distillation loss " + }, + { + "bbox": [ + 130, + 413, + 479, + 485 + ], + "type": "inline_equation", + "content": "L_{KD}" + }, + { + "bbox": [ + 130, + 413, + 479, + 485 + ], + "type": "text", + "content": " and the relative position loss " + }, + { + "bbox": [ + 130, + 413, + 479, + 485 + ], + "type": "inline_equation", + "content": "L_{RP}" + }, + { + "bbox": [ + 130, + 413, + 479, + 485 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 254, + 494, + 480, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 494, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 254, + 494, + 480, + 506 + ], + "type": "interline_equation", + "content": "L _ {K D} = \\mathcal {D} _ {k d} (T (x), S (x)), \\tag {6}", + "image_path": "9ef7a86e8a74047f43ca573773fab369b2d158568ee444165d65c970aa6cce9d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 242, + 509, + 480, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 509, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 242, + 509, + 480, + 521 + ], + "type": "interline_equation", + "content": "L _ {R P} = \\mathcal {D} _ {r p} \\left(T \\left(x _ {p e s}\\right), S \\left(x _ {p e s}\\right)\\right), \\tag {7}", + "image_path": "26661defefe393c728e5072eab867f0af162e11bfd9872a1627bbb9a6a3aeb23.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 256, + 524, + 480, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 524, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 256, + 524, + 480, + 536 + ], + "type": "interline_equation", + "content": "L _ {T} = \\lambda_ {1} L _ {K D} + \\lambda_ {2} L _ {R P}, \\tag {8}", + "image_path": "2143a1f0ef2fc9891afdc17be0d0eae00435fc6959c1e24e8778f39fed355e99.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 544, + 479, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 479, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 479, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 544, + 479, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{kd}" + }, + { + "bbox": [ + 130, + 544, + 479, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 544, + 479, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{rp}" + }, + { + "bbox": [ + 130, + 544, + 479, + 581 + ], + "type": "text", + "content": " are distance measure formulas, " + }, + { + "bbox": [ + 130, + 544, + 479, + 581 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 544, + 479, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 544, + 479, + 581 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 544, + 479, + 581 + ], + "type": "text", + "content": " are hyperparameters. The inter-class relations of the two models can be aligned by ensuring intra-class consistency through " + }, + { + "bbox": [ + 130, + 544, + 479, + 581 + ], + "type": "inline_equation", + "content": "L_{KD}" + }, + { + "bbox": [ + 130, + 544, + 479, + 581 + ], + "type": "text", + "content": " and inter-class consistency through " + }, + { + "bbox": [ + 130, + 544, + 479, + 581 + ], + "type": "inline_equation", + "content": "L_{RP}" + }, + { + "bbox": [ + 130, + 544, + 479, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 581, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 479, + 628 + ], + "type": "text", + "content": "Furthermore, we can constrain the substitute model to be inter-class consistent with the target model in different directions and distances based on the PES of a clean sample in different directions and at different distances. At this point, the loss of topology consistency can be expressed as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 247, + 635, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 635, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 247, + 635, + 481, + 667 + ], + "type": "interline_equation", + "content": "L _ {T} = \\lambda_ {1} L _ {K D} + \\lambda_ {2} \\sum_ {i = 1} ^ {M} L _ {R P} ^ {i}, \\tag {9}", + "image_path": "844ff594ee0bea8d73fd45be44839c8bd3865917fc091cb3648b15a56f2a9284.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 270, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 270, + 91, + 448, + 102 + ], + "type": "text", + "content": "Topology Alignment for Substitute Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 187, + 146, + 427, + 243 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 1: The accuracy of target model. We select four models with different architectures as target models on each of the three datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 146, + 427, + 243 + ], + "lines": [ + { + "bbox": [ + 187, + 146, + 427, + 243 + ], + "spans": [ + { + "bbox": [ + 187, + 146, + 427, + 243 + ], + "type": "table", + "html": "
DatasetCIFAR10
Target Model AccuracyResNet34 94.53Vgg16 92.82GoogleNet 95.31ShuffleNetV2 94.53
DatasetCIFAR100
Target Model AccuracyResNet50 79.3MobileNetV2 71.61WideResNet 79.69ResNeXt50 77.56
DatasetTinyImageNet
Target Model AccuracyResNet50 62.81Vgg19 59.66MobileNetV2 62.81ResNeXt50 66.33
", + "image_path": "60264219e6ad8edf2df369696aa6dacc898e90f0c439dc059db11c0a17f18ee1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 263, + 425, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 425, + 276 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 425, + 276 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 263, + 425, + 276 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 263, + 425, + 276 + ], + "type": "text", + "content": " is the number of " + }, + { + "bbox": [ + 130, + 263, + 425, + 276 + ], + "type": "inline_equation", + "content": "x_{pes}" + }, + { + "bbox": [ + 130, + 263, + 425, + 276 + ], + "type": "text", + "content": " generated for a same clean sample." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 276, + 479, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 479, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 479, + 310 + ], + "type": "text", + "content": "In addition, incorporating the original categorical loss into this process is often beneficial, which is usually the cross-entropy loss " + }, + { + "bbox": [ + 130, + 276, + 479, + 310 + ], + "type": "inline_equation", + "content": "L_{CE}" + }, + { + "bbox": [ + 130, + 276, + 479, + 310 + ], + "type": "text", + "content": ". The overall loss can be expressed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 261, + 312, + 480, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 312, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 261, + 312, + 480, + 323 + ], + "type": "interline_equation", + "content": "L _ {a l l} = L _ {T} + \\lambda_ {3} L _ {C E}, \\tag {10}", + "image_path": "ea13859fd58293fbb4602d33521cab7c382181bc535b5f4774a012bd63e6841f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 327, + 479, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 479, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 479, + 352 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 327, + 479, + 352 + ], + "type": "inline_equation", + "content": "\\lambda_3" + }, + { + "bbox": [ + 130, + 327, + 479, + 352 + ], + "type": "text", + "content": " is a hyperparameter. The parameters of the substitute model are optimized by back propagation to better simulate the target model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 369, + 235, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 369, + 235, + 382 + ], + "spans": [ + { + "bbox": [ + 131, + 369, + 235, + 382 + ], + "type": "text", + "content": "4 Experimental" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 392, + 268, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 392, + 268, + 403 + ], + "spans": [ + { + "bbox": [ + 131, + 392, + 268, + 403 + ], + "type": "text", + "content": "4.1 Experimental Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 410, + 480, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 480, + 482 + ], + "type": "text", + "content": "Dataset and Target Model. We choose CIFAR10 [10] (with size of " + }, + { + "bbox": [ + 130, + 410, + 480, + 482 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 130, + 410, + 480, + 482 + ], + "type": "text", + "content": " and 10 classes), CIFAR100 [10] (with size of " + }, + { + "bbox": [ + 130, + 410, + 480, + 482 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 130, + 410, + 480, + 482 + ], + "type": "text", + "content": " and 100 classes), TinyImageNet [19] (with size of " + }, + { + "bbox": [ + 130, + 410, + 480, + 482 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 130, + 410, + 480, + 482 + ], + "type": "text", + "content": " and 200 classes) as the training dataset. And we select VGG(16, 19) [23], ResNet(34, 50) [6], MobileNetV2 [20], GoogleNet [25], WideResNet [34], ShuffleNetV2 [15], and ResNeXt50 [32] as the target model. The accuracy of the target models is shown in Tab. 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 498, + 480, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 480, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 480, + 569 + ], + "type": "text", + "content": "Training Setup. During training, we set the perturbation strength in Eq. (5) to " + }, + { + "bbox": [ + 130, + 498, + 480, + 569 + ], + "type": "inline_equation", + "content": "\\phi = |Y|" + }, + { + "bbox": [ + 130, + 498, + 480, + 569 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 498, + 480, + 569 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 498, + 480, + 569 + ], + "type": "text", + "content": " obeys the Gaussian distribution " + }, + { + "bbox": [ + 130, + 498, + 480, + 569 + ], + "type": "inline_equation", + "content": "Y \\sim N(0,3)" + }, + { + "bbox": [ + 130, + 498, + 480, + 569 + ], + "type": "text", + "content": ". The number of PES in Eq. (9) is set to " + }, + { + "bbox": [ + 130, + 498, + 480, + 569 + ], + "type": "inline_equation", + "content": "M = 1" + }, + { + "bbox": [ + 130, + 498, + 480, + 569 + ], + "type": "text", + "content": ". The hyperparameters are empirically set to " + }, + { + "bbox": [ + 130, + 498, + 480, + 569 + ], + "type": "inline_equation", + "content": "\\lambda_1 = \\lambda_3 = 1, \\lambda_2 = 10" + }, + { + "bbox": [ + 130, + 498, + 480, + 569 + ], + "type": "text", + "content": ". We standardize the query budget for substitute training across all baselines to ensure fairness, setting it to 2 million for CIFAR10 and 3 million for CIFAR100 and TinyImageNet." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "Additionally, we conduct experiments in two scenarios: full-data and data-free. In the data-free scenario, we utilize Stable Diffusion [18] as a generator and use the LCA [21] strategy to produce the data needed for substitute training. In the full-data scenario, the whole training data is available for substitute training. We perform this part of the experiment to validate the performance of the scheme without the influence of the generator. The algorithm of our ICTA and other experimental details are described in the supplement. The source code will be released upon acceptance of the paper." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "L. Meng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 157, + 478, + 302 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 147 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 147 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 147 + ], + "type": "text", + "content": "Table 2: Comparison of different white-box adversarial sample generation schemes. We choose ResNet34 on CIFAR10, ResNet50 on CIFAR100 and VGG19 in TinyImageNet as target models. The best results are marked as bolded." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 157, + 478, + 302 + ], + "lines": [ + { + "bbox": [ + 135, + 157, + 478, + 302 + ], + "spans": [ + { + "bbox": [ + 135, + 157, + 478, + 302 + ], + "type": "table", + "html": "
Dataset MethodsCIFAR10CIFAR100TinyImageNet
FGSMBIMPGDFGSMBIMPGDFGSMBIMPGD
Non-TargetMAZE [9]38.6457.9868.9438.2041.5543.079.189.7911.10
DaST [37]35.1956.2164.2240.0142.5447.409.719.269.81
DST [27]40.8367.6571.2646.4549.2159.4611.6712.4733.52
Zhang et al. [35]60.1194.2197.5483.6994.5391.1838.5848.2165.01
DiffusionLCA [21]68.2995.7397.3082.0395.5896.4951.5552.3570.58
ICTA (Ours)75.6098.7199.1987.4299.0199.2583.7492.7393.39
TargetMAZE [9]6.8413.8316.656.2012.7810.010.502.502.54
DaST [37]7.8515.4118.615.0112.4811.410.203.025.20
DST [27]10.2616.1219.3310.2216.3724.651.883.055.87
Zhang et al. [35]28.1772.0564.057.7458.4554.782.3010.8815.01
DiffusionLCA [21]24.3176.9079.984.6145.1951.501.8412.9515.45
ICTA (Ours)29.4187.6391.407.8069.0274.496.8729.3133.41
", + "image_path": "e383bb70d65f2234b6ac87be9b4f443c32988967539e39d01befb545d39c5ebd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 324, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 324, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 324, + 480, + 384 + ], + "type": "text", + "content": "Evaluation Setup and Process. During testing, we choose three classical methods, including FGSM [5], BIM [11], and PGD [16], to generate adversarial samples, with parameters uniformly set to perturbation boundary " + }, + { + "bbox": [ + 132, + 324, + 480, + 384 + ], + "type": "inline_equation", + "content": "\\epsilon = 8 / 255" + }, + { + "bbox": [ + 132, + 324, + 480, + 384 + ], + "type": "text", + "content": " and step size " + }, + { + "bbox": [ + 132, + 324, + 480, + 384 + ], + "type": "inline_equation", + "content": "\\alpha = 2 / 255" + }, + { + "bbox": [ + 132, + 324, + 480, + 384 + ], + "type": "text", + "content": ". Our hardware utilizes an RTX 2080Ti GPU, and all schemes are retrained under the same environment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 384, + 480, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 480, + 433 + ], + "type": "text", + "content": "We train substitute model by our ICTA and then generate adversarial samples based on the substitute models using a white-box adversarial sample generation scheme. The success rate of these adversarial samples in attacking the target model is used as the final evaluation metric." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 449, + 402, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 449, + 402, + 462 + ], + "spans": [ + { + "bbox": [ + 133, + 449, + 402, + 462 + ], + "type": "text", + "content": "4.2 Comparative Experiments in Data-free Scenario" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 468, + 481, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 588 + ], + "type": "text", + "content": "Comparison of White-box Adversarial Sample Generation Schemes. We compare different white-box adversarial sample generation schemes, including FGSM, BIM, and PGD. As illustrated in Tab. 2, our ICTA realizes significant performance improvements across different white-box adversarial sample generation methods. On the one hand, our ICTA achieves more noticeable performance enhancement on large datasets TinyImageNet than on small datasets CIFAR10. On the other hand, ICTA has superior performance improvement on target attacks than non-target attacks. The results indicate that our ICTA more effectively simulates the decision space of the target model compared to traditional substitute training methods [9,21,27,35,37]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 481, + 665 + ], + "type": "text", + "content": "Comparison on Different Target Models. According to the results in Tab. 3, our ICTA demonstrates a significant performance improvement on different target models compared to the GANs-based approaches [9,27,35,37] and the diffusion-based method [21]. Specifically, ICTA achieves the most superior performance in target attacks, especially on TinyImageNet. Additionally, ICTA" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "text", + "content": "Topology Alignment for Substitute Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 146, + 479, + 282 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 3: Comparison on different target models. We choose PGD as the default white-box adversarial sample generation scheme. The best results are marked as bolded." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 146, + 479, + 282 + ], + "lines": [ + { + "bbox": [ + 135, + 146, + 479, + 282 + ], + "spans": [ + { + "bbox": [ + 135, + 146, + 479, + 282 + ], + "type": "table", + "html": "
Dataset Target ModelCIFAR10CIFAR100TinyImageNet
VGG16ResNet34GoogleNetMobileNetResNet50ResNet50VGG19
Non-TargetMAZE [9]47.4968.9427.1351.7343.0715.5411.10
DaST [37]53.2964.2226.4857.9447.4016.119.81
DST [27]55.1571.2638.6465.8559.4632.5133.52
Zhang et al. [35]95.8497.5467.2497.7891.1878.8265.01
DiffusionLCA [21]97.0197.3094.8297.0196.4967.5670.58
ICTA (Ours)96.4899.1997.3299.6399.2594.8093.39
TargetMAZE [9]11.3916.6512.9811.0310.013.182.54
DaST [37]15.3918.615.319.5811.412.215.20
DST [27]17.0719.3318.8918.4624.658.555.87
Zhang et al. [35]57.4764.0539.8344.1154.7820.0115.01
DiffusionLCA [21]75.3979.9878.1851.6151.5012.4515.45
ICTA (Ours)79.5791.4084.5770.5674.4943.0233.41
", + "image_path": "132387563955c009fc3d588063fcf533cb1de0886d98fc38c9164958fde89ea9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 187, + 325, + 427, + 451 + ], + "blocks": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "type": "text", + "content": "Table 4: Comparison of attack success rate in label-based and probability-based scenarios. The best results are marked as bolded." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 325, + 427, + 451 + ], + "lines": [ + { + "bbox": [ + 187, + 325, + 427, + 451 + ], + "spans": [ + { + "bbox": [ + 187, + 325, + 427, + 451 + ], + "type": "table", + "html": "
methodProbability-basedLabel-based
cifar10cifar100cifar10cifar100
Non-TargetGANsDaST [37]64.2247.4067.7236.57
DST [27]71.2659.4680.4248.50
Zhang et al. [35]97.5491.1898.9890.20
DiffusionLCA [21]97.3096.4998.4894.55
ICTA (Ours)99.1999.2599.4998.53
TargetGANsDaST [37]18.6111.4119.285.20
DST [27]19.3324.6523.854.59
Zhang et al. [35]64.0554.7869.2540.17
DiffusionLCA [21]79.9851.5082.1849.85
ICTA (Ours)91.4074.4993.1061.71
", + "image_path": "0f9ccf569e335699319bf0abef7cce73584b7cb4c9756c1d41edf7bfe1fb4a96.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 475, + 480, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 523 + ], + "type": "text", + "content": "consistently enhances performance across different target models, showcasing the stability of the performance in various scenarios. In summary, our ICTA is able to simulate different target models in black-box attack scenarios and exhibits substantial performance improvements on large-scale datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 544, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 665 + ], + "type": "text", + "content": "Comparison in Probability-based and Label-based Scenarios. We compare the attack success rate in the probability-based and label-based scenarios, as displayed in Tab. 4. In the label-based scenario, only the output labels of the target model are accessible, while in the probability-based scenario, both the labels and probabilities of the target model are available. The experiment results indicate that our ICTA achieves the optimal performance in both probability-based and label-based scenarios and shows significant improvement over the existing substitute attack methods [21, 27, 35, 37]. This highlights the comprehensive superiority of our ICTA, demonstrating outstanding attack performance even in scenarios where only label outputs are accessible." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "L. Meng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 305, + 235 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 305, + 235 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 305, + 235 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 305, + 235 + ], + "type": "image", + "image_path": "2a75e89d1c9924e5f3372ea70f15115ca70d513532a468a743499effdd66d29f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 237, + 260, + 245 + ], + "lines": [ + { + "bbox": [ + 178, + 237, + 260, + 245 + ], + "spans": [ + { + "bbox": [ + 178, + 237, + 260, + 245 + ], + "type": "text", + "content": "(a) Attack success rate" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 115, + 481, + 236 + ], + "blocks": [ + { + "bbox": [ + 309, + 115, + 481, + 236 + ], + "lines": [ + { + "bbox": [ + 309, + 115, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 309, + 115, + 481, + 236 + ], + "type": "image", + "image_path": "5920324477c26a0d460724b5af97e45e144a5bdbf6f7ffb3736a08273d777f78.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 237, + 454, + 246 + ], + "lines": [ + { + "bbox": [ + 336, + 237, + 454, + 246 + ], + "spans": [ + { + "bbox": [ + 336, + 237, + 454, + 246 + ], + "type": "text", + "content": "(b) Accuracy of substitute model" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 255, + 480, + 300 + ], + "lines": [ + { + "bbox": [ + 130, + 255, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 480, + 300 + ], + "type": "text", + "content": "Fig. 4: Comparison of attack success rate and accuracy in full-data scenario. Comparison schemes include the substitute training schemes DST [27] and Zhang et al. [35] as well as a distillation method DIST [8]. The ResNet50 on CIFAR100 is used as the target model." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 324, + 400, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 324, + 400, + 337 + ], + "spans": [ + { + "bbox": [ + 131, + 324, + 400, + 337 + ], + "type": "text", + "content": "4.3 Comparative Experiments in Full-data Scenario" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 346, + 480, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 480, + 464 + ], + "type": "text", + "content": "Performance of Different Substitute Training Strategies. In Fig. 4, we present the attack success rate (both target and non-target attacks) and the accuracy of the substitute training during the training process. It can be observed that our ICTA demonstrates excellent performance in both target and nontarget attacks and shows higher efficiency in target attacks. There is a significant improvement in non-target attacks compared to conventional methods DST [27] and Zhang et al. [35], which highlights the unique advantages of our ICTA in improving the effectiveness of target attacks. In addition, we compare an excellent knowledge distillation scheme, DIST [8], which also shows the superior performance of our ICTA in attack scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 466, + 480, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 466, + 480, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 466, + 480, + 526 + ], + "type": "text", + "content": "Notably, we find that the accuracy of the substitute model does not fully reflect the effectiveness of the substitute training. Despite having a relatively low accuracy in our ICTA, the substitute model exhibits excellent performance in attacks. This indicates that our ICTA better simulates the target model in terms of inter-class topology, rather than just replicating its outputs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": "Comparison of Relative Performance with White-box Attacks. We further compare the attack success rate of our ICTA with the corresponding white-box attacks, as demonstrated in Tab. 5, where the Relative Performance (RP) is used as a metric to measure the performance relative to the white-box attacks. Our ICTA achieves comparable performance to the corresponding white-box attacks in non-target attack, when the query budgets is set to 2M (for CIFAR10) and 3M (for CIFAR100 and TinyImageNet). Moreover, it is worth noting that the target attack success rate exceeds " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "83\\%" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " on the CIFAR10 dataset, " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "67\\%" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " on the CIFAR100 dataset, and " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "44\\%" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " on the TinyImageNet dataset compared to the corresponding white-box attack. As can be seen, the proposed ICTA achieves quite" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "text", + "content": "Topology Alignment for Substitute Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 157, + 479, + 371 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 5: Comparison with direct white-box attacks on target model. We utilize Relative Performance (RP) to evaluate the attack success of our black-box attack method relative to the white-box attack (in red)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 157, + 479, + 371 + ], + "lines": [ + { + "bbox": [ + 135, + 157, + 479, + 371 + ], + "spans": [ + { + "bbox": [ + 135, + 157, + 479, + 371 + ], + "type": "table", + "html": "
ResNet34Vgg16GoogleNetShuffleNetV2
white-boxOursRPwhite-boxOursRPwhite-boxOursRPwhite-boxOursRP
CIFAR10Non-TargetFGSM78.5483.421.0685.5979.460.9363.8477.731.2284.9783.970.99
BIM100.0099.591.00100.0097.260.9799.9397.610.98100.0098.860.99
PGD100.0099.791.00100.0098.080.98100.0098.260.98100.0099.200.99
TargetFGSM28.7934.691.2019.6630.251.5424.5526.991.1032.9331.780.97
BIM99.8494.580.9584.8181.190.9699.5582.580.83100.0086.010.86
PGD100.0096.630.9798.2185.670.8799.9786.290.86100.0089.030.89
ResNet50WideResNetMobileNetV2ResNeXt50
CIFAR100Non-TargetFGSM84.7786.761.0274.2783.401.1290.1888.200.9887.2487.511.00
BIM100.0099.220.9998.1196.830.9999.9899.090.9999.9899.160.99
PGD100.0099.500.9999.5797.540.98100.0099.460.99100.0099.470.99
TargetFGSM3.058.522.793.879.422.434.397.351.673.838.792.30
BIM95.2272.790.7689.4465.900.7496.3064.790.6797.0673.290.76
PGD99.2579.340.8098.3873.220.7499.6670.330.7199.7080.150.80
ResNet50Vgg19MobileNetV2ResNeXt50
TinyImageNetNon-TargetFGSM95.8286.070.9092.8286.740.9398.4288.000.8996.6283.650.87
BIM99.9796.260.9699.6495.940.96100.0097.210.9799.9894.780.95
PGD99.9896.960.9799.9196.560.97100.0097.780.9899.9895.860.96
TargetFGSM10.2510.040.984.797.991.6715.718.740.5613.188.680.66
BIM98.9652.300.5385.6639.310.4699.7245.920.4699.5343.560.44
PGD99.7457.850.5897.8146.310.4799.9251.180.5199.9650.300.50
", + "image_path": "ba839a0069a9da0f974cc6fd5db7be3f977fb2b69424e89941b901ccfe8f4955.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 392, + 480, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 480, + 415 + ], + "type": "text", + "content": "satisfactory performance in black-box attack scenarios, demonstrating robust adaptability in target attacks across different datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 431, + 368, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 368, + 444 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 368, + 444 + ], + "type": "text", + "content": "4.4 Analysis of Decision Boundary Similarity" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 450, + 263, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 263, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 263, + 472 + ], + "type": "text", + "content": "The t-SNE Visualization of Decision Space. We" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 277, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 277, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 277, + 665 + ], + "type": "text", + "content": "exhibit the t-Distributed Stochastic Neighbor Embedding (t-SNE) visualization of the decision space of the models, as exhibited in Fig. 5. In t-SNE visualization, each data point is typically labeled with its corresponding class. Therefore, it can depict the interclass relative position in the decision space, as well as the inter-class topology. The visualization illustrates that the inter-class topology in the decision space of the substitute model trained by our ICTA" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 273, + 475, + 328, + 624 + ], + "blocks": [ + { + "bbox": [ + 273, + 475, + 328, + 624 + ], + "lines": [ + { + "bbox": [ + 273, + 475, + 328, + 624 + ], + "spans": [ + { + "bbox": [ + 273, + 475, + 328, + 624 + ], + "type": "image", + "image_path": "858835a3adecdc067bdc922f7096def4005dcf635c8895ad22e3101213641a91.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 634, + 482, + 668 + ], + "lines": [ + { + "bbox": [ + 270, + 634, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 270, + 634, + 482, + 668 + ], + "type": "text", + "content": "Fig. 5: The t-SNE visualization of the decision space for (a) the target model, (b) our ICTA, (c) Zhang et al. [35] and (d) DST [27]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 329, + 475, + 378, + 625 + ], + "blocks": [ + { + "bbox": [ + 329, + 475, + 378, + 625 + ], + "lines": [ + { + "bbox": [ + 329, + 475, + 378, + 625 + ], + "spans": [ + { + "bbox": [ + 329, + 475, + 378, + 625 + ], + "type": "image", + "image_path": "f555365d8168529095cbbd02daa21736f37dbab9611910fc00331bb0728ad1e3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 380, + 475, + 428, + 625 + ], + "blocks": [ + { + "bbox": [ + 380, + 475, + 428, + 625 + ], + "lines": [ + { + "bbox": [ + 380, + 475, + 428, + 625 + ], + "spans": [ + { + "bbox": [ + 380, + 475, + 428, + 625 + ], + "type": "image", + "image_path": "dde4a897cb1c2ab0fc82ec9083b1f4c263a684303c331ef937b98f483f520ad4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 430, + 475, + 479, + 625 + ], + "blocks": [ + { + "bbox": [ + 430, + 475, + 479, + 625 + ], + "lines": [ + { + "bbox": [ + 430, + 475, + 479, + 625 + ], + "spans": [ + { + "bbox": [ + 430, + 475, + 479, + 625 + ], + "type": "image", + "image_path": "4d2c5c7420637432cbb064d5533859f023b02a8944afe5a2ebc0f095488de350.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "L. Meng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 305, + 231 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 305, + 231 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 305, + 231 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 305, + 231 + ], + "type": "image", + "image_path": "2dbc627c5a13e3cd73661b1b7ae055e3d1bdd9f361a0a3d5b9e81f17580a9931.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 177, + 232, + 260, + 241 + ], + "lines": [ + { + "bbox": [ + 177, + 232, + 260, + 241 + ], + "spans": [ + { + "bbox": [ + 177, + 232, + 260, + 241 + ], + "type": "text", + "content": "(a) Attack success rate" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 116, + 481, + 230 + ], + "blocks": [ + { + "bbox": [ + 309, + 116, + 481, + 230 + ], + "lines": [ + { + "bbox": [ + 309, + 116, + 481, + 230 + ], + "spans": [ + { + "bbox": [ + 309, + 116, + 481, + 230 + ], + "type": "image", + "image_path": "43060eba2b4b6e0c13b209fddffb14ee2b1b88839822769456a3a46907a24a47.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 232, + 454, + 241 + ], + "lines": [ + { + "bbox": [ + 335, + 232, + 454, + 241 + ], + "spans": [ + { + "bbox": [ + 335, + 232, + 454, + 241 + ], + "type": "text", + "content": "(b) Accuracy of substitute model" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 251, + 479, + 284 + ], + "lines": [ + { + "bbox": [ + 131, + 251, + 479, + 284 + ], + "spans": [ + { + "bbox": [ + 131, + 251, + 479, + 284 + ], + "type": "text", + "content": "Fig. 6: Ablation experiments on PES. The experiments include baseline (no relative position loss) and strategies for exploring inter-class relative position through traditional adversarial samples, clean samples, and position exploration sample." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 308, + 480, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 480, + 369 + ], + "type": "text", + "content": "method is remarkably similar to the topology of the target model. On the contrary, both the output alignment-focused DST [27] and the decision boundary-constrained approach by Zhang et al. [35] exhibit greater differences from the target model in t-SNE visualization. This confirms that our ICTA is more effective in aligning the substitute model to match the target model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 387, + 298, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 387, + 298, + 494 + ], + "spans": [ + { + "bbox": [ + 131, + 387, + 298, + 494 + ], + "type": "text", + "content": "Quantitative Analysis. We explore random directions of arbitrary samples, and then quantitatively assess the Decision Boundary Similarity (DBS) based on the overlap between target and substitute models. As in Tab. 6, our ICTA has more consistent decision boundaries with the target model." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 445, + 481, + 485 + ], + "blocks": [ + { + "bbox": [ + 306, + 422, + 480, + 443 + ], + "lines": [ + { + "bbox": [ + 306, + 422, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 480, + 443 + ], + "type": "text", + "content": "Table 6: Quantitative analysis of decision boundary similarity." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 445, + 481, + 485 + ], + "lines": [ + { + "bbox": [ + 308, + 445, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 481, + 485 + ], + "type": "table", + "html": "
MethodICTA (Ours)Zhang et al.DST
CIFAR100.930.870.80
CIFAR1000.880.770.71
TinyImageNet0.820.570.55
", + "image_path": "770909530de46121111cac7f2fee19b55722b528e15be1d92ee3b7aa16668309.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 513, + 237, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 513, + 237, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 237, + 525 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 533, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 480, + 628 + ], + "type": "text", + "content": "Ablation on Position Exploration Sample. In this paper, we design PES to explore the inter-class relative position in the decision space through their relative position relationships with respect to the original clean samples. However, as the position of any two samples in the decision space inherently depict the inter-class relative position, traditional adversarial sample or even a clean sample has the effect of exploring the decision space. In this section, we validate the effectiveness of PES by comparing it with traditional adversarial sample generated by BIM and clean sample, as illustrated in Fig. 6." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "type": "text", + "content": "It is worth noting that while all three approaches contribute to explore the inter-class relative position in the decision space, PES stands out as particularly advantageous. This is due to the fact that both clean samples and the adversarial" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "text", + "content": "Topology Alignment for Substitute Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 187, + 168, + 427, + 242 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 7: Ablation comparison of the loss function and the distance metric. We select two distance metrics, MSE and DIST [8], and use different combinations of distance metrics in the loss function. The best result are marked as bolded, and the sub-optimal results are marked as underlined." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 168, + 427, + 242 + ], + "lines": [ + { + "bbox": [ + 187, + 168, + 427, + 242 + ], + "spans": [ + { + "bbox": [ + 187, + 168, + 427, + 242 + ], + "type": "table", + "html": "
DkdMSE DISTDrpMSE DISTTargetNon-Target
FGSMBIMPGDFGSMBIMPGD
✓4.3142.1446.1479.4895.3796.38
✓4.1155.5257.6782.8598.3098.80
✓✓7.6464.2071.7084.3598.4399.06
✓✓7.8273.4279.0986.7399.0499.61
✓✓8.5272.7979.3486.7699.2299.50
", + "image_path": "e9fc07eb7c6ff812749579ff02bb0272bd49885bd91b52daa61c9453af8c266f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 265, + 482, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 326 + ], + "type": "text", + "content": "samples generated by BIM are closer to the decision center, resulting in limited exploration in the decision space. In contrast, the proposed PES is more diverse by employing the full-gradient and unconstrained setting, enabling exploration across a wider range of the decision space. The ablation experiments for the parameters of PES and the number of PES are provided in the supplement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 343, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 452 + ], + "type": "text", + "content": "Ablation on Loss and Metrics. We conduct an ablation study on the loss functions and distance metrics of substitute training in Tab. 7, focusing on the distillation loss " + }, + { + "bbox": [ + 130, + 343, + 482, + 452 + ], + "type": "inline_equation", + "content": "L_{KD}" + }, + { + "bbox": [ + 130, + 343, + 482, + 452 + ], + "type": "text", + "content": " and relative position loss " + }, + { + "bbox": [ + 130, + 343, + 482, + 452 + ], + "type": "inline_equation", + "content": "L_{RP}" + }, + { + "bbox": [ + 130, + 343, + 482, + 452 + ], + "type": "text", + "content": ". Here, the term 'DIST' refers to the distance metric function in the distillation scheme DIST [8]. It can be seen that when DIST [8] is used as the distance metric in " + }, + { + "bbox": [ + 130, + 343, + 482, + 452 + ], + "type": "inline_equation", + "content": "L_{KD}" + }, + { + "bbox": [ + 130, + 343, + 482, + 452 + ], + "type": "text", + "content": ", there is a certain improvement in the attack success rate. And when " + }, + { + "bbox": [ + 130, + 343, + 482, + 452 + ], + "type": "inline_equation", + "content": "L_{RP}" + }, + { + "bbox": [ + 130, + 343, + 482, + 452 + ], + "type": "text", + "content": " is added, there is a significant improvement in the attack success rate, which underscores the indispensability of the relative position loss. Notably, our ICTA employs DIST as the distance metric in both losses, as it further improves attack performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 472, + 220, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 220, + 484 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 220, + 484 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we propose a novel ICTA scheme to better align the substitute model to the target model for efficient black-box substitute attacks. Compared to the existing schemes that only align intra-class consistency, we simulate the target model in the decision space more comprehensively through inter-class topology alignment. Specifically, we first explore the inter-class relative position of the target model in the decision space by the designed PES. Subsequently, we achieve inter-class topology alignment of the two models by constraining the inter-class relative position of the substitute model in different directions. Benefiting from the better alignment of the two models in the decision space, the adversarial samples generated by the substitute model are more successful in misleading the target model to classify them into the target class. Experimental results demonstrate the effectiveness of ICTA, which achieves better simulation of the target model and significantly improves the attack success rate beyond existing substitute attack approaches." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "L. Meng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 187 + ], + "type": "text", + "content": "This work was supported by the National Key Research and Development Program of China (Grant No. 2021YFA1000102), the National Natural Science Foundation of China (Grant Nos. 62376285, 62272375, and 61673396), and Natural Science Foundation of Shandong Province, China (Grant No. ZR2022MF260)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 228, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "type": "text", + "content": "1. Andriushchenko, M., Croce, F., Flammarion, N., Hein, M.: Square Attack: a query-efficient black-box adversarial attack via random search. In: Proceedings of the European Conference on Computer Vision. pp. 484-501 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 262, + 480, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 480, + 294 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 480, + 294 + ], + "type": "text", + "content": "2. Brendel, W., Rauber, J., Bethge, M.: Decision-Based Adversarial Attacks: Reliable Attacks Against Black-Box Machine Learning Models. In: International Conference on Learning Representations. arXiv preprint arXiv:1712.04248 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 295, + 480, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 295, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 295, + 480, + 327 + ], + "type": "text", + "content": "3. Chen, S., He, Z., Sun, C., Yang, J., Huang, X.: Universal Adversarial Attack on Attention and the Resulting Dataset DImageNet. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(4), 2188-2197 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 328, + 480, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 328, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 328, + 480, + 360 + ], + "type": "text", + "content": "4. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: International Conference on Machine Learning. vol. 119, pp. 2206-2216 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 361, + 480, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 480, + 381 + ], + "type": "text", + "content": "5. Goodfellow, I., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 383, + 480, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 383, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 138, + 383, + 480, + 415 + ], + "type": "text", + "content": "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 415, + 480, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 480, + 447 + ], + "type": "text", + "content": "7. Huang, H., Chen, Z., Chen, H., Wang, Y., Zhang, K.: T-SEA: Transfer-based Self-Ensemble Attack on Object Detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20514-20523 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 449, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 480, + 479 + ], + "type": "text", + "content": "8. Huang, T., You, S., Wang, F., Qian, C., Xu, C.: Knowledge distillation from a stronger teacher. In: Advances in Neural Information Processing Systems. vol. 35, pp. 33716-33727 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 480, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 480, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 480, + 480, + 512 + ], + "type": "text", + "content": "9. Kariyappa, S., Prakash, A., Qureshi, M.K.: Maze: Data-free model stealing attack using zeroth-order gradient estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13814-13823 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 514, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 534 + ], + "type": "text", + "content": "0. Krizhevsky, A.: Learning Multiple Layers of Features from Tiny Images. Master's thesis, University of Tront (2009)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 536, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 555 + ], + "type": "text", + "content": "1. Kurakin, A., Goodfellow, I., Bengio, S.: Adversarial examples in the physical world. arXiv preprint arXiv:1607.02533 (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 557, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 480, + 588 + ], + "type": "text", + "content": "2. Lee, B.K., Kim, J.: Mitigating Adversarial Vulnerability through Causal Parameter Estimation by Adversarial Double Machine Learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4499-4509 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "type": "text", + "content": "3. Li, L., Lian, Q., Chen, Y.C.: Adv3D: Generating 3D Adversarial Examples in Driving Scenarios with NeRF. arXiv preprint arXiv:2309.01351 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "text", + "content": "4. Liu, J., Lu, B., Xiong, M., Zhang, T., Xiong, H.: Adversarial Attack with Raindrops. arXiv preprint arXiv:2302.14267 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "5. Ma, N., Zhang, X., Zheng, H.T., Sun, J.: ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design. In: Proceedings of the European Conference on Computer Vision. pp. 122-138 (2018)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "text", + "content": "Topology Alignment for Substitute Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "16. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "17. Ren, K., Zheng, T., Qin, Z., Liu, X.: Adversarial Attacks and Defenses in Deep Learning. Engineering 6(3), 346-360 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 193 + ], + "type": "text", + "content": "18. Robin, R., Andreas, B., Dominik, L., Patrick, E., Bj\\''orn, O.: High-Resolution Image Synthesis with Latent Diffusion Models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10674-10685 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 236 + ], + "type": "text", + "content": "19. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision 115(3), 211-252 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "20. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: MobileNetV2: Inverted Residuals and Linear Bottlenecks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4510-4520 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "text", + "content": "21. Shao, M., Meng, L., Qiao, Y., Zhang, L., Zuo, W.: Data-free black-box attack based on diffusion model. arXiv preprint arXiv:2307.12872 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "22. Sheng, L., Liang, J., He, R., Wang, Z., Tan, T.: AdaptGuard: Defending Against Universal Attacks for Model Adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19093-19103 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "text", + "content": "23. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "24. Su, J., Vargas, D.V., Sakurai, K.: One Pixel Attack for Fooling Deep Neural Networks. IEEE Transactions on Evolutionary Computation 23(5), 828-841 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 369, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 369, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 133, + 369, + 480, + 413 + ], + "type": "text", + "content": "25. Szegedy, C., Wei Liu, Yangqing Jia, Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1-9 (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 414, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 414, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 133, + 414, + 480, + 456 + ], + "type": "text", + "content": "26. Wang, N., Luo, Y.: Does Physical Adversarial Example Really Matter to Autonomous Driving? Towards System-Level Effect of Adversarial Object Evasion Attack. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4412-4423 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 490 + ], + "type": "text", + "content": "27. Wang, W., Qian, X., Fu, Y., Xue, X.: Dst: Dynamic substitute training for data-free black-box attack. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14361-14370 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 491, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 491, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 133, + 491, + 480, + 533 + ], + "type": "text", + "content": "28. Wang, W., Yin, B., Yao, T., Zhang, L., Fu, Y., Ding, S., Li, J., Huang, F., Xue, X.: Delving into data: Effectively substitute training for black-box attack. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4761-4770 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "29. Wang, X., Zhang, Z.: Structure Invariant Transformation for better Adversarial Transferability. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4607-4619 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 133, + 567, + 480, + 599 + ], + "type": "text", + "content": "30. Wang, X., Zhang, Z., Tong, K., Gong, D., He, K., Li, Z., Liu, W.: Triangle Attack: A Query-efficient Decision-based Adversarial Attack. In: Proceedings of the European Conference on Computer Vision. pp. 156-174 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "31. Wang, Z., Yang, H., Feng, Y., Sun, P., Guo, H., Zhang, Z., Ren, K.: Towards Transferable Targeted Adversarial Examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "text", + "content": "32. Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K.: Aggregated Residual Transformations for Deep Neural Networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5987-5995 (2017)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "L. Meng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "33. Yu, M., Sun, S.: FE-DaST: Fast and effective data-free substitute training for black-box adversarial attacks. Computers & Security 113, 102555 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "text", + "content": "34. Zagoruyko, S., Komodakis, N.: Wide Residual Networks. In: Proceedings of the British Machine Vision Conference. pp. 87.1-87.12 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 194 + ], + "type": "text", + "content": "35. Zhang, J., Li, B., Xu, J., Wu, S., Ding, S., Zhang, L., Wu, C.: Towards efficient data free black-box adversarial attack. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15094-15104 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 194, + 482, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 216 + ], + "type": "text", + "content": "36. Zheng, X., Fan, Y., Wu, B., Zhang, Y., Wang, J., Pan, S.: Robust Physical-World Attacks on Face Recognition. Pattern Recognition 133, 109009 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "type": "text", + "content": "37. Zhou, M., Wu, J., Liu, Y., Liu, S., Zhu, C.: Dast: Data-free substitute training for adversarial attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 234-243 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 248, + 482, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 282 + ], + "type": "text", + "content": "38. Zhu, Y., Chen, Y., Li, X., Chen, K., He, Y., Tian, X., Zheng, B., Chen, Y., Huang, Q.: Toward Understanding and Boosting Adversarial Transferability From a Distribution Perspective. IEEE Transactions on Image Processing 31, 6487-6501 (2022)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "text", + "content": "Topology Alignment for Substitute Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/InterFusion_ Text-Driven Generation of 3D Human-Object Interaction/d1b8e0cb-6c9e-44c9-80dd-158238285207_content_list.json b/2024/InterFusion_ Text-Driven Generation of 3D Human-Object Interaction/d1b8e0cb-6c9e-44c9-80dd-158238285207_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8f4666731104378f97974787079d30490e1aff3d --- /dev/null +++ b/2024/InterFusion_ Text-Driven Generation of 3D Human-Object Interaction/d1b8e0cb-6c9e-44c9-80dd-158238285207_content_list.json @@ -0,0 +1,1754 @@ +[ + { + "type": "text", + "text": "InterFusion: Text-Driven Generation of 3D Human-Object Interaction", + "text_level": 1, + "bbox": [ + 277, + 140, + 728, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sisi Dai $^{1}$ Wenhao Li $^{1}$ Haowen Sun $^{2}$ Haibin Huang $^{3}$ Chongyang Ma $^{3}$", + "bbox": [ + 233, + 210, + 767, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hui Huang² Kai Xu1* Ruizhen Hu2*", + "bbox": [ + 357, + 227, + 643, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National University of Defense Technology", + "bbox": [ + 351, + 253, + 651, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Shenzhen University", + "bbox": [ + 424, + 268, + 576, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Kuaishou Technology", + "bbox": [ + 421, + 282, + 580, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://sisidai.github.io/InterFusion/", + "bbox": [ + 352, + 297, + 648, + 310 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/50527545e07cf197e0ea4b8ec32da26788a54af9738ffc8ea344664b715fc370.jpg", + "image_caption": [ + "a person in a military uniform sitting on a chair" + ], + "image_footnote": [], + "bbox": [ + 241, + 352, + 390, + 411 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17c5444b41efb96aeca0e190d47d2f4f8ad4a5923438856ad2850b3acf4e5bf0.jpg", + "image_caption": [ + "a female gymnastics athlete riding a motorcycle" + ], + "image_footnote": [], + "bbox": [ + 411, + 352, + 591, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2e56b4cc995146aa1aa82ecfa544cee682f2d8f82ad89f82701320f6f7d20fc.jpg", + "image_caption": [ + "a person in a fleece pullover and hiking pants carrying a backpack" + ], + "image_footnote": [], + "bbox": [ + 612, + 352, + 767, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f195b07182d8df6deab02e6d56a0f0f2cb08cc8c01932f8366d5baa3597109c4.jpg", + "image_caption": [ + "a man in a quilted gilet and moleskin trousers playing the guitar" + ], + "image_footnote": [], + "bbox": [ + 236, + 441, + 392, + 503 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/92890147ea2c39a7a17782e400d6fe8d58673aa4c49fcdd0b53c740c8d90279f.jpg", + "image_caption": [ + "a man wearing a WORLD lettered T-shirt riding a bike" + ], + "image_footnote": [], + "bbox": [ + 411, + 441, + 586, + 507 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aa89e5e1a2d9857f9c1d6945dcb75bc3475d0d357ba44dd64349d0dc978bdd56.jpg", + "image_caption": [ + "a person in a cropped hoodie and bike shorts jumping a skateboard" + ], + "image_footnote": [], + "bbox": [ + 602, + 443, + 772, + 503 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3024fe4eb00435fcd76f639eb90deadeddb99818b0799aba7148089b8aec6496.jpg", + "image_caption": [ + "a man in a velvet smoking jacket and silk pajama pants playing the saxophone" + ], + "image_footnote": [], + "bbox": [ + 243, + 534, + 393, + 598 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/88157f0a186ebce2852d8e61b2d4a37c57c6b6dd5834d1407aed4f5d88f50332.jpg", + "image_caption": [ + "a man in a paisley waistcoat and bowler hat riding a horse" + ], + "image_footnote": [], + "bbox": [ + 415, + 535, + 581, + 598 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fcdaaf3478512b2e9be0025e1da06574883663a508d0a923d8a75af0d79ccf7d.jpg", + "image_caption": [ + "a man with blond hair wearing a brown leather jacket reading a laptop", + "Fig. 1: Given a text prompt, our method InterFusion can generate diverse 3D scenes of a person interacting with an object." + ], + "image_footnote": [], + "bbox": [ + 602, + 534, + 767, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this study, we tackle the complex task of generating 3D human-object interactions (HOI) from textual descriptions in a zero-shot text-to-3D manner. We identify and address two key challenges: the unsatisfactory outcomes of direct text-to-3D methods in HOI, largely due to the lack of paired text-interaction data, and the inherent difficulties in simultaneously generating multiple concepts with complex spatial relationships. To effectively address these issues, we present InterFusion,", + "bbox": [ + 259, + 719, + 740, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding authors: kevin.kai.xu@gmail.com; ruizhen.hu@gmail.com.", + "bbox": [ + 218, + 825, + 712, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a two-stage framework specifically designed for HOI generation. InterFusion involves human pose estimations derived from text as geometric priors, which simplifies the text-to-3D conversion process and introduces additional constraints for accurate object generation. At the first stage, InterFusion extracts 3D human poses from a synthesized image dataset depicting a wide range of interactions, subsequently mapping these poses to interaction descriptions. The second stage of InterFusion capitalizes on the latest developments in text-to-3D generation, enabling the production of realistic and high-quality 3D HOI scenes. This is achieved through a local-global optimization process, where the generation of human body and object is optimized separately, and jointly refined with a global optimization of the entire scene, ensuring a seamless and contextually coherent integration. Our experimental results affirm that InterFusion significantly outperforms existing state-of-the-art methods in 3D HOI generation.", + "bbox": [ + 259, + 147, + 743, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: Text-Driven Generation $\\cdot$ Zero-Shot Generation $\\cdot$ 3D Human-Object Interaction Generation", + "bbox": [ + 261, + 369, + 750, + 397 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 426, + 375, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The generation of 3D human-object interactions (HOI) stands as a critical challenge in the fields of computer vision and computer graphics, with far-reaching implications in virtual reality, augmented reality, animation, and embodied AI [41, 49]. This task entails the creation of realistic 3D scenes where human figures interact with objects in ways that are not only physically plausible but also contextually relevant.", + "bbox": [ + 212, + 460, + 787, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite its potential, the field has faced significant obstacles, primarily due to the scarcity of large-scale interaction data. Traditional approaches have predominantly relied on motion capture (mocap) datasets or physics-based simulations for generating these interactions. Mocap datasets [15,26,55] are limited by the specific scenarios they capture and are both costly and labor-intensive to produce. These limitations have resulted in a notable gap in generating diverse and contextually rich HOI scenes, especially for novel or complex interactions. Conversely, recent advancements have introduced text-to-3D methods [40,56], marking a significant shift in the field. These methods harness the power of textual descriptions to generate 3D objects without direct 3D supervision, presenting a novel approach to 3D content creation.", + "bbox": [ + 212, + 551, + 787, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this study, we explore text-to-3D method in HOI task within a zero-shot manner, i.e., generating 3D scenes from textual descriptions using 2D diffusion models. Our key observations are two-fold: first of all, a direct application of text-to-3D method in HOI often leads to unsatisfied results like blurry textures and incorrect interactions, which is caused by the relative lack of paired text-interaction data during training and the difficulties of generating multiple concepts for diffusion-based methods [23]. Secondly, while collecting extensive interaction data is challenging, estimating human poses based on described in", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Dai et al.", + "bbox": [ + 271, + 114, + 336, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "teractions is more feasible. These pose estimations can serve as geometric priors in the HOI generation process. By integrating pose estimation, our method substantially simplifies the text-to-3D process, particularly in the geometry optimization stage for human body generation. It also provides additional constraints for object generation, ensuring accurate placement and alignment with the text's semantic content. More importantly, we can separate the generation of human body and object, and jointly refine the details with the estimated pose, allowing for a more coherent and detailed synthesis of the interaction scene.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by these observations, we introduce InterFusion, a novel two-stage framework designed for HOI generation. Specifically, in the first stage, rather than relying on precise 3D interaction data, our approach instead collects a comprehensive dataset of synthesized images that depict a wide range of interactions. From these images, we employ advanced 3D pose estimation technique [8] to extract 3D human poses. Building upon these image-pose pairs, InterFusion develops a sophisticated codebook that establishes a mapping between interaction descriptions and 3D human poses, with the integration of the CLIP (Contrastive Language-Image Pretraining) embedding [43]. By leveraging these embeddings, our framework is able to interpret the nuances of interaction descriptions and translate them into accurate 3D pose representations. In the subsequent stage, InterFusion capitalizes on recent advancements in text-to-3D generation [40], as well as neural radiance fields [30], using the estimated human poses to produce 3D HOI scenes with realistic appearances and high-quality geometry. This stage operates in a 'local-global' manner. At the local level, the generation of the human body (SDS-H) and objects (SDS-O) is separately optimized, with the poses serving as additional constraints for the SDS. At the global level, the generation of the entire scene (SDS-I) is also guided by the integrated description and jointly optimized with SDS-H and SDS-O, ensuring a cohesive and contextually accurate representation of the HOI. Our experiments show that the quality of generation can be improved by a large margin and our approach outperforms state-of-the-art methods in HOI generation.", + "bbox": [ + 212, + 276, + 787, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, our contributions are as follows:", + "bbox": [ + 240, + 616, + 583, + 631 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel two-stage framework InterFusion, for zero-shot 3D human-object interaction generation from text, incorporating 3D pose estimation as geometry priors.", + "- InterFusion leverages text-to-3D generation with a local-global optimization process. This strategy ensures seamless integration of human bodies and objects, producing realistic and high-quality 3D HOI scenes.", + "- InterFusion demonstrates significant improvements over existing methods in 3D HOI generation, showcasing its effectiveness in creating detailed, and contextually rich 3D interactions." + ], + "bbox": [ + 225, + 686, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "InterFusion", + "bbox": [ + 651, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 388, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Human-Object Interaction Synthesis", + "text_level": 1, + "bbox": [ + 215, + 180, + 565, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D human-object interaction (HOI) generation is a challenging problem that has been studied widely by the computer vision and graphics community. Shape2Pose [18] generates plausible 3D human poses interacting with a given 3D object model by learning an affordance model from synthetic data. PiGraphs [47] learns the distribution between human poses and object arrangements from a collected dataset with 3D scene scans and RGB-D videos, generating the interaction snapshots given action specifications and object models. Recently, the parametric human body models such as SMPL [25,37,45] are employed in interaction synthesis to overcome the lack of realism due to human body representations. Benefiting from the PROX dataset [12] which consists of fitted SMPL models in captured 3D scenes, the more specific human-scene interaction has become an active research direction. Given a 3D scene, PSI [63] and PLACE [62] generate the 3D human body mesh represented as SMPL parameters through a conditional variational autoencoder [19, 51]. POSA [13] learns pose-specific priors to generate contacts conditioned by the given posed human, which can further guide the placement of the body mesh in a scene. COINS [64] enables semantic control on interaction synthesis by embedding the action label together with the interacted object as the condition of the generative model. More recently, fine-grained 3D interaction datasets [1, 7, 53] are captured to promote this field. Relying on these datasets, some methods [6, 38, 58], concurrent to this work, are proposed to explore text-guided 3D HOI generation. However, they remain constrained by distributions within the datasets. While previous methods need ground truth 3D interaction data as supervision, our work, for the first time, attempts to break through the limitation of data requirement. We generate a wider range of realistic and detailed 3D HOI scenes, including both indoors and outdoors.", + "bbox": [ + 212, + 209, + 787, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Text-to-3D Content Synthesis", + "text_level": 1, + "bbox": [ + 215, + 614, + 509, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Early methods [3, 16, 24] for text-to-3D shape generation require paired data of 3D data and the corresponding textual descriptions to learn the joint embedding space of shape and text for supervision, which limits their generality to unseen object categories. Benefiting from large pre-trained text-to-image models and differentiable rendering techniques, breakthroughs in text-to-3D content generation have been achieved. For example, DreamFields [17] and PureCLIPN-eRF [20] combine CLIP [43] with neural radiance fields (NeRF) [30], demonstrating the potential for zero-shot NeRF optimization. Meanwhile, CLIP-mesh [32] and Text2Mesh [28] incorporate CLIP to optimize the 3D mesh representation, starting from an initial sphere mesh and an input base mesh, respectively. Recently, DreamFusion [40] and SJC [56] enable NeRF optimization with guidance from pre-trained text-to-image diffusion models [44,46] in place of CLIP, achieving more impressive results. To improve DreamFusion, Magic3D [21] proposes", + "bbox": [ + 212, + 643, + 789, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Dai et al.", + "bbox": [ + 271, + 114, + 336, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a coarse-to-fine pipeline to generate the fine-grained mesh. TextMesh [54] extends the geometry representation from NeRF to an SDF framework, thereby enhancing detailed mesh extraction and photorealistic rendering. Among follow-up works, Latent-NeRF [27] and Vox-E [48] utilize explicit 3D shapes to provide additional training signals for NeRF optimization. While Latent-NeRF utilizes a rough untextured object for shape sculpture, Vox-E takes multiview images of a fine-grained textured object to edit geometry and appearance. All the methods mentioned above focus on the generation or edition of a single subject, while ignoring the interaction between different subjects.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3 Compositional Scene Generation", + "text_level": 1, + "bbox": [ + 214, + 306, + 532, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Representing scenes as compositions of object representations facilitates enhanced controllability. Numerous techniques incorporate additional information at the object level to perform compositional modeling of scenes, effectively separating object representations from the overall scene imagery. For instance, some methods incorporate 2D semantic information such as segmentation labels [65], instance masks [59,61], or features from a pre-trained vision-language model [31]. Some other methods [10, 36, 52] use 3D layout information by object-centric bounding boxes with canonical coordinates. When it comes to scene generation, [34, 35, 60] use compositional representations to generate scenes in a controllable manner. More recent approaches [4, 22, 39] generate compositional 3D scenes from input text prompts. Different from existing methods, our HOI scene is generated automatically without the requirement of input layout. Moreover, the spatial relationship between human and object during interaction is much more complex and cannot be simply characterized using their bounding boxes.", + "bbox": [ + 212, + 333, + 787, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Our Method", + "text_level": 1, + "bbox": [ + 214, + 568, + 375, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 214, + 601, + 341, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we formally introduce InterFusion with a focus on the zero-shot and text-driven generation of 3D human-object interactions (HOI). Specifically, the input is a triplet of text descriptions, $T = \\{T^H, T^O, T^I\\}$ , specifying the desired human style, object style, and interaction type. The goal is to generate a detailed 3D scene, $\\psi = \\{\\psi^H, \\psi^O\\}$ , comprising a human model and an object model that not only adhere to the specified appearance styles but also exhibit a tailored spatial relationship to accurately reflect the described interaction.", + "bbox": [ + 212, + 628, + 787, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Figure 2, our method consists of two primary stages: anchor pose generation and anchor pose guided HOI generation. We first generate an interaction pose based on the input text, termed an anchored pose. This pose then serves as a geometric constraint, guiding the subsequent generation of detailed HOI. In the second phase, the human and object models are optimized separately and refined simultaneously with a global context, ensuring a cohesive and accurate representation of the interaction as described by the input text.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "InterFusion", + "bbox": [ + 651, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/53294ededa3526b22fbae85de4df0dcb00b770483fc2042d3e9e7a401d299b72.jpg", + "image_caption": [ + "Fig. 2: InterFusion is a two-stage framework that transforms textual descriptions into detailed 3D human-object interactions, initially synthesizing anchor poses (upper left) and then optimizing the human model (upper right) and object model (bottom) with constraints from estimated pose and textual prompts." + ], + "image_footnote": [], + "bbox": [ + 222, + 148, + 782, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 474, + 372, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SDS. Score Distillation Sampling (SDS) has been introduced by DreamFusion [40]. While $x = g(\\psi, \\zeta)$ , $x$ is the 2D image rendered by a differentiable renderer $g$ with model parameters $\\psi$ (the volumetric renderer and MLPs correspondingly in NeRF), under a desired camera pose $\\zeta$ . By injecting the sampled noise $\\epsilon$ into $x$ at a time step $t$ , the noisy image $x_{t}$ is produced. The pre-trained 2D text-to-image diffusion model $\\phi$ provides a denoising network $\\hat{\\epsilon}_{\\phi}(x_{t};y,t)$ that predicts the noise $\\hat{\\epsilon}$ given the noisy image $x_{t}$ , time step $t$ , and text embedding $y$ . SDS then optimizes the model parameters $\\psi$ by minimizing the difference between the predicted noise and the added noise:", + "bbox": [ + 212, + 503, + 787, + 640 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\psi} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , x) = \\mathbb {E} _ {t, \\epsilon} [ w (t) (\\hat {\\epsilon} _ {\\phi} (x _ {t}; y, t) - \\epsilon) \\frac {\\partial x}{\\partial \\psi} ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 333, + 651, + 785, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $w(t)$ is the weighting term at the time step $t$ .", + "bbox": [ + 215, + 695, + 589, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Anchor Pose Generation", + "text_level": 1, + "bbox": [ + 215, + 734, + 465, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "InterFusion starts by generating the anchor pose from the input text $T^I$ , which is a non-trivial problem and highly limited by the available pose datasets. Existing datasets (e.g., HumanML3D [9], BABEL [42]) are mocap based datasets and are typically conducted in a controlled laboratory environment, leading to lacks in action diversity and spatial coverage.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Dai et al.", + "bbox": [ + 271, + 114, + 336, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Recently, text-to-image diffusion models have shown the versatility and effectiveness of generating realistic and diverse images, through the integration of visual and linguistic feature spaces. We thus instead utilize advancements in this technology, specifically models like Stable Diffusion, to create a large-scale dataset of Human-Object Interaction (HOI) images. These images depict a wide range of human interactions. To ensure the diversity of interaction types, we utilize ChatGPT to generate prompts about human daily events or actions forming \"verb-ing a/an/the object\". A total of 235 result prompts are generated, covering most interactions in daily life. After filtering synthesized images without humans, we then estimate 3D human poses using the pre-trained PixIE [8] model, creating a comprehensive Syn-HOI pose dataset consisting of a total 55K 3D pseudo-SMPL poses.", + "bbox": [ + 212, + 146, + 782, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While SMPL model maps pose and shape parameters to a triangulated mesh, to align these poses with our text-to-interaction generation task, we render images from multiple perspectives and use CLIP's image encoder to derive pose feature embeddings. The average feature across these multiple perspective images represents the pose feature and attaches the dataset with pairs of averaged CLIP embedding and pose parameters. We further construct a codebook using K-Means clustering to identify key pose centroids based on the feature embeddings. 2,048 cluster centroids are clustered compositing our pose code-book, where each cluster comprises a subset of poses that represents similar interaction as the key poses of the centroids.", + "bbox": [ + 212, + 330, + 784, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Once the codebook is built, for a given input text, we extract its feature embedding using CLIP's text encoder. This serves as a query to retrieve the most similar poses from the codebook, based on feature similarity. Specifically, given query text $T^{I}$ , top $k$ poses $\\theta_{k}^{T^{I}}$ could be matched by pose embeddings $\\theta_{E}$ , the averaged CLIP embeddings among rendered images from key poses:", + "bbox": [ + 212, + 486, + 784, + 564 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {k} ^ {T ^ {I}} = \\operatorname {T O P} _ {k} \\left(f _ {\\text {t e x t}} \\left(T ^ {I}\\right), \\theta_ {E}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 582, + 784, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $f_{text}$ is the text encoder of CLIP and $\\mathrm{TOP}_k(X,Y)$ returns top $k$ poses with the top $k$ highest cosine similarities, and we use $k = 7$ for suitable poses in our experiments.", + "bbox": [ + 212, + 619, + 782, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We then utilize GPT-4V to select the most precise pose as the final queried key pose. Depending on the requirement, we can instead select the poses sampled in the cluster, corresponding to the key pose, to guide the generation process for diverse results. This approach ensures rich and contextually aligned anchor poses for our text-to-interaction generation task.", + "bbox": [ + 212, + 670, + 782, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To restrict the optimization of geometry and appearance using human structural priors from the acquired pose, we further incorporate COAP [29], a neural occupancy representation of the articulated human body based on SMPL parameters [25]. Given a pose $\\theta$ and a shape $\\beta$ , COAP offers an occupancy prediction network $f(x; \\beta, \\theta)$ that maps a 3D query point $p$ to an occupancy value, directly indicating whether the spatial point resides within the 3D body.", + "bbox": [ + 212, + 750, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "InterFusion", + "bbox": [ + 651, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Pose-Guided HOI Generation", + "text_level": 1, + "bbox": [ + 215, + 146, + 504, + 161 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Once get the interaction pose, we further use it along with the input text to guide the generation of a detailed 3D HOI scene. In this phase, the estimated pose serves both as spatial constraints for the scene's geometry and as an anchor that aligns the human and object models. This approach ensures that the human model and the object model can be generated separately, while they are cohesively aligned to form the final 3D HOI scene.", + "bbox": [ + 212, + 175, + 787, + 265 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The anchor pose provides specific spatial constraints for each component in the scene. For the human model, it establishes a basic geometric structure, while for the object model, it defines the areas that should remain unoccupied. This clear distinction is essential for rendering the scene both accurately and realistically. Simultaneously, the input text undergoes a complex processing procedure to offer distinct semantic guidance for each component. This is accomplished through various text conditioning techniques applied in SDS with the pre-trained DeepFloyd model [5]. The text is carefully crafted to direct the generation of both the human and object models to ensure all elements are in harmony with the semantic nuances of input descriptions.", + "bbox": [ + 212, + 266, + 787, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, we introduce a novel camera tracing module to enhance the optimization process under varying text conditions. This module adaptively adjusts the camera pose, focusing on relevant elements within the scene at each optimization stage. This adaptive camera positioning is instrumental in ensuring that each aspect of the scene is optimally rendered according to the text descriptions, resulting in a more dynamic and contextually accurate 3D HOI scene.", + "bbox": [ + 212, + 417, + 787, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We now provide details of each module in this stage, including the Neural Radiance Field representations for the human model (H-NeRF) and object model (O-NeRF), the camera tracing mechanism, and the guided optimization process.", + "bbox": [ + 214, + 523, + 787, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "H-NeRF. We use NeRF to represent the human model, noted as H-NeRF. The generation of H-NeRF is guided by the text description specifying the human style, conducting pose-specific human avatar generation with SDS. To enhance the quality of our renderings, particularly in terms of resolution, we incorporated a specialized optimization process that focuses on the head region of the human avatar. The location of the head for any given pose is determined with COAP [29]. This allows us to augment the text prompt specifically for the head region, using the notation * the head of *, to ensure that the head receives detailed attention during the generation process. The loss function for this optimization process is as follows, designed to balance the fidelity of the head region with the overall pose and style of the human figure.", + "bbox": [ + 214, + 595, + 787, + 763 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\psi^ {H}} \\mathcal {L} _ {\\mathrm {S D S}} = \\mathbb {E} _ {t, \\epsilon} [ w (t) (\\hat {\\epsilon} _ {\\phi} \\left(x _ {t} ^ {H}; y ^ {H}, t\\right) - \\epsilon) \\frac {\\partial x ^ {H}}{\\partial \\psi^ {H}} ] \\tag {3} \\\\ + \\mathbb {E} _ {t, \\epsilon} [ w (t) (\\hat {\\epsilon} _ {\\phi} (x _ {t} ^ {H, h}; y ^ {H, h}, t) - \\epsilon) \\frac {\\partial x ^ {H , h}}{\\partial \\psi^ {H}} ], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 776, + 785, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Dai et al.", + "bbox": [ + 271, + 114, + 336, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In shaping H-NeRF, our geometric constraint ensures that the human model evolves directly from the anchor pose. Points within the anchor are required to be occupied, establishing a firm base for the model. Meanwhile, points outside the anchor can also be occupied, but with probabilities that decrease as they move away from the anchor's surface. This approach allows for the addition of geometric details over the anchor, ensuring the model aligns with the human style described in the text while maintaining a coherent structure rooted in the anchor pose. The loss function used is as follows:", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {g e o}} ^ {H} = C E _ {p _ {i} \\in \\mathbb {P} _ {\\mathrm {i n}}} \\left(\\alpha_ {i}, f (p _ {i})\\right) \\\\ + C E _ {p _ {j} \\in \\mathbb {P} _ {\\text {o u t}}} \\left(\\alpha_ {j}, f (p _ {j})\\right) \\left(1 - e ^ {- \\frac {d}{2 \\eta^ {2}}}\\right), \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 277, + 784, + 321 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $d$ represents the point distance from the anchor surface, and $\\eta$ is the hyperparameter to control the extent of decaying, similar as in [27].", + "bbox": [ + 212, + 325, + 785, + 357 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "O-NeRF. We also employ NeRF to model the object component, referred to as O-NeRF. This model is designed to interact seamlessly with the human model and to embody the desired style as specified by the input text. To guide the generation of both the interaction type and the object style, we utilize SDS with tailored text prompts:", + "bbox": [ + 214, + 378, + 787, + 454 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\psi^ {O}} \\mathcal {L} _ {\\mathrm {S D S}} = \\mathbb {E} _ {t, \\epsilon} [ w (t) (\\hat {\\epsilon} _ {\\phi} \\left(x _ {t} ^ {I}; y ^ {I}, t\\right) - \\epsilon) \\frac {\\partial x ^ {I}}{\\partial \\psi^ {O}} ] \\tag {5} \\\\ + \\mathbb {E} _ {t, \\epsilon} [ w (t) (\\hat {\\epsilon} _ {\\phi} (x _ {t} ^ {O}; y ^ {O}, t) - \\epsilon) \\frac {\\partial x ^ {O}}{\\partial \\psi^ {O}} ], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 464, + 784, + 531 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For the interaction scene $x^{I}$ generation, we use an alpha-composed rendering to integrate H-NeRF and O-NeRF. This method calculates an alpha value from the density at each point, determining its contribution to the scene's color. Higher alpha values indicate a greater influence on the rendering, allowing for a nuanced and realistic integration of the human and object models in the final interaction image:", + "bbox": [ + 212, + 542, + 787, + 635 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {I} = \\sum_ {i} w _ {i} ^ {I} c _ {i} ^ {I}, w _ {i} ^ {I} = \\alpha_ {i} ^ {I} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j} ^ {I}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 374, + 647, + 784, + 691 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nc _ {i} ^ {I} = \\frac {\\alpha_ {i} ^ {H}}{\\alpha_ {i} ^ {H} + \\alpha_ {i} ^ {O}} c _ {i} ^ {H} + \\frac {\\alpha_ {i} ^ {O}}{\\alpha_ {i} ^ {H} + \\alpha_ {i} ^ {O}} c _ {i} ^ {I}.\n$$\n", + "text_format": "latex", + "bbox": [ + 374, + 691, + 604, + 724 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In our NeRF-based approach, alpha values, capped at 1, are derived from density for composite rendering. This setup allows semantic guidance gradients, conditioned by the interaction image, to optimize the density and color of both the human and object models. However, we noted a tendency for the model to prioritize object generation at the expense of the human model's quality. To counteract this, we implemented gradient truncation towards the human model to maintain a balanced optimization between the two components.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "InterFusion", + "bbox": [ + 651, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For O-NeRF's geometric constraints, we aim to prevent occupancy of points within the anchor pose, denoted as $x_{i} \\in \\mathbb{X}_{in}$ , by the object model. Inspired by the physical collision prevention concept, we define a specific loss function that ensures these points remain unoccupied by O-NeRF, effectively preventing overlap between the human and object models in the 3D space:", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {g e o}} ^ {O} = C E _ {p _ {i} \\in \\mathbb {P} _ {\\mathrm {i n}}} \\left(\\alpha_ {i}, 1 - f \\left(p _ {i}\\right)\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 234, + 785, + 253 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "By minimizing the above loss term, model parameters are optimized to enhance the geometric consistency between H-NeRF and O-NeRF.", + "bbox": [ + 214, + 263, + 785, + 294 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Camera tracing. To enhance the generation of O-NeRF, we introduce a dynamic camera tracing module within the SDS framework. This module automatically adjusts the camera pose to focus on the target's center, either the entire interaction scene or just the object during object generation. The camera aligns with the average position of voxels with an occupancy probability above 0.5, ensuring a consistent focus on the most significant parts of the scene or object for optimal detail capture and realism.", + "bbox": [ + 212, + 319, + 787, + 426 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Optimization. The total loss is formulated as follows:", + "bbox": [ + 214, + 450, + 617, + 465 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {S D S}} ^ {H} + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {S D S}} ^ {O} + \\mathcal {L} _ {\\mathrm {g e o}} ^ {H} + \\lambda_ {2} \\mathcal {L} _ {\\mathrm {g e o}} ^ {O} + \\lambda_ {3} L _ {\\text {r e g}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 477, + 785, + 496 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\lambda_1, \\lambda_2$ and $\\lambda_3$ are the corresponding loss weights. Figure 2 (left bottom) shows several intermediate states of the object model during the generation steps and we can see that the object is gradually generated with fine details. Weight annealing is adopted during the guided optimization process. We leave the details in the supplement.", + "bbox": [ + 212, + 507, + 787, + 583 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 609, + 375, + 626 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Figure 3 (left) shows results with various interaction poses, showcasing the strength of our method. Our method can support diverse interaction poses within a single type. Meanwhile, with interaction type and pose fixed, our method can further generate more numerous results under different human styles or object styles, as shown in Figure 3 (right). Moreover, our InterFusion also supports controllable text-conditioned editing, providing users more control over the already generated 3D models, which are presented in the supplementary materials.", + "bbox": [ + 212, + 643, + 787, + 748 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We both qualitatively and quantitatively evaluate our method against alternative baseline methods. To verify the effectiveness of individual components in our method, we conduct ablation studies. Furthermore, we discuss the application potential, limitations and future work of our InterFusion. The implementation details, more evaluations, and further discussions are presented in the supplementary materials.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Dai et al.", + "bbox": [ + 271, + 114, + 336, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2d8c7dc580e85d742db240cc65d5e6505a8946110158bada2f459e393530a939.jpg", + "image_caption": [ + "Fig. 3: More results generated by InterFusion. Diverse integration poses are supported. Numerous human and object styles are also supported." + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 782, + 348 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation prompts and criteria. Similar as the first stage for generating prompts about various interaction types, we also use ChatGPT to randomly generate prompts about human styles and collect the text prompts for evaluation. We select 61 distinct and diverse prompts, ensuring the classes of Human (e.g., a policewoman, a teenager) with various styles (e.g., in a graphic tee and ripped jeans), Objects (e.g., a motorcycle, a guitar) and Interactions (e.g., riding, playing) are reasonable and evenly distributed. In total, we have 13 different types of interactions, covering contact areas across the whole body, which demonstrates the effectiveness of our method on diverse types of interactions compared to baseline methods. We calculate the CLIP scores between the input text prompts and different views of generated 3D human-object interactions, and then compare the means based on different evaluation prompts. The CLIP score measures the similarity between a prompt text for an image and the actual content of the image. We also provide an assessment using GPT-4V for selection, named GPT-4V select, which evaluates the completeness of objects and correctness of physical interactions across multiview rendered images. We refer to the supplementary material for assessment details and additional evaluation criteria.", + "bbox": [ + 212, + 419, + 787, + 676 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baseline approaches. To the best of our knowledge, the proposed method is among the first to generate 3D human-object interactions based on text inputs in a zero-shot manner. We thus compare our method to alternative text-to-3D methods, including DreamFusion [40], Magic3D [21], and TextMesh [54]. Additional comparisons with MVDream [50] and ProlificDreamer [57] are presented in the supplementary material. To justify our key idea of using human pose to guide the generation, we design an object-centric baseline (Ours-OC) as a variant of our method. Instead of using human pose as the prior, we retrieve an object with the given semantic category from ShapeNet dataset [2] and use it as", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "InterFusion", + "bbox": [ + 651, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/c565535e5634524a64053c169f87922ed970c26c1c641f526171b2691c22777b.jpg", + "table_caption": [ + "Table 1: Quantitative evaluation of CLIP score and GPT-4V select." + ], + "table_footnote": [], + "table_body": "
MethodDreamFusion [40] Magic3D [21] TextMesh [54] Ours-OCOurs-HC
CLIP score0.3027 0.3179 0.2761 0.32030.3308
GPT-4V select(%)8.20 11.48 1.64 13.1165.57
", + "bbox": [ + 233, + 171, + 769, + 222 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "geometry constraints to guide both human and object generation. As official implementations are unavailable for some of these baselines, we use the third-party re-implementations provided by threestudio [11] for a fair comparison. Note that all re-implementations use multi-resolution hash-grid [33] for 3D representation and DeepFloyd [5] for guidance.", + "bbox": [ + 212, + 250, + 787, + 327 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.1 Comparison Results", + "text_level": 1, + "bbox": [ + 215, + 349, + 429, + 366 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative evaluations. Some representative visual comparisons are shown in Figure 4, and additional comparisons including those with our object-centric baseline are presented in the supplementary material. The overall results of the baselines reveal common uncertainties in both geometry and appearance attributed to the confusion introduced by multi-concept guidance. Specifically, baseline models may lean towards one specific target, as seen in the example of \"a man wearing a red baseball cap\" with only the red baseball cap generated, the example of \"a policeman\" with only the upper half of the human body generated, or the example of \"a man in a puffer jacket and wool beanie\" where the shopping cart is failed to be generated. Even with a relatively uniform attention distribution, the baseline model may encounter challenges in producing complete results, as demonstrated in the saxophone and violin examples (the 3rd and 4th column in Figure 4), which exhibit deficiencies such as incomplete human body parts, and mixing of human legs. Moreover, the baseline model sometimes struggles to effectively choose the focus of generation when the input text describes a less common target like \"a person in a military uniform\", resulting in poor outcomes.", + "bbox": [ + 212, + 377, + 787, + 633 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Our approach overcomes these issues by conducting optimization in an explicit decomposed way and intelligently guiding attention from SDS jointly in spatial and semantic aspects. Therefore, our method achieves more stable and higher-quality 3D results under multiple-concept guidance.", + "bbox": [ + 212, + 633, + 787, + 696 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Quantitative evaluations. The comparison results are shown in Table 1. We can see that our method (Ours-HC) achieves the best performance compared to baselines, showcasing our results exhibit more 3D plausibility with given text prompts. Among the baselines, Ours-OC gets the best performance. Though with object priors, the object-centric approach still does not provide sufficient body priors for human generation, resulting in its inability to achieve complete interaction generation. Magic3D gets the next best performance among the baselines, as it usually generates the whole interaction scene comparing to other two", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Dai et al.", + "bbox": [ + 271, + 114, + 336, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/bb4bcc47b4835dbb2190e3855df9e24f2aa58e104660583f8d42fc662c834f1b.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparison results with baselines. InterFusion generates more stable and higher-quality results and is more consistent with input interaction descriptions." + ], + "image_footnote": [], + "bbox": [ + 225, + 148, + 774, + 372 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d09f2531a0a0aa6b4ea91ea0a7822a369b44105fbe3ec9892cb227a36719aa6e.jpg", + "image_caption": [ + "Fig. 5: Qualitative results of ablation studies." + ], + "image_footnote": [], + "bbox": [ + 225, + 429, + 781, + 583 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "basielines, however, the results are somewhat vague as shown in Figure 4, thus there is still a large performance gap comparing to our method.", + "bbox": [ + 212, + 640, + 784, + 670 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The significant enhancement in performance can be attributed to our method's effective concurrent generation of both human and object, along with cohesive interactive information.", + "bbox": [ + 212, + 672, + 790, + 715 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.2 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 747, + 400, + 760 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We conduct several ablation studies both qualitatively and quantitatively to show the importance of our design as well as introduced loss terms. Results are presented in Figure 5a, Figure 5b and Table 2. We refer to the supplementary material for additional results and details.", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "InterFusion", + "bbox": [ + 651, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/58757538265b0785170cf6db4d62ba1ed6a1c22505e3faf2501350105e6d6ced.jpg", + "table_caption": [ + "Table 2: Quantitative results of ablation studies." + ], + "table_footnote": [], + "table_body": "
Settingsw/o ISDSw/o L0SDSw/o L0geoOurs
CLIP score0.31640.32930.31710.3308
GPT-4V select(%)1.644.9216.3977.05
", + "bbox": [ + 305, + 171, + 697, + 226 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Syn-HOI-pose. The anchor pose obtained at the first stage provides the geometry constraints for our HOI generation. We thus first demonstrate the effectiveness of synthesizing anchoring poses using our pseudo-pose dataset. We compare the pose generated with the existing approach AvatarCLIP [14], which utilizes CLIP only to retrieve the queried pose from the codebook constructed from the mocap dataset AMASS [26]. Figure 5a shows that the proposed pose generation stage enables a better-fitted anchor pose for input interaction types, which demonstrates our pseudo-poses, reconstructed from synthesized images, has the potential for more diverse poses' requirements than existing mocap datasets.", + "bbox": [ + 212, + 253, + 787, + 391 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Semantic guidance. To better illustrate the efficacy of semantic guidance in our NeRF-based generation, we conduct ablations separately on SDS from object and interaction. As shown in Figure 5b and Table 2, in the absence of SDS from object, semantic consistency for the object is compromised, resulting in noisy outcomes influenced by the human, thus the contact region can not be extracted correctly. SDS from interaction plays the most crucial role in the performance, making results more semantically and visually plausible. The absence of SDS from interaction leads to generated objects being confined to the remaining space, but without interaction with the human.", + "bbox": [ + 212, + 412, + 787, + 549 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Geometric constraint. Relying solely on semantic guidance is inadequate for achieving the final objectives, as the object should be generated in spaces outside the human body, with its generation targets not positioned at the origin. As illustrated in Figure 5b and Table 2, depending only on semantic guidance may yield results that appear spatially conflict with the human body. Without the spatial constraint, the object generation at the origin would additionally have a conflict with the interaction objective, thus resulting in degenerated objects and final interactions.", + "bbox": [ + 212, + 571, + 787, + 691 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 717, + 359, + 732 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In conclusion, our work presents InterFusion, a novel framework for zero-shot 3D human-object interaction generation. InterFusion tackles the challenges of limited 3D interaction data and the complexity of generating multiple concepts simultaneously. Our two-stage approach, which synthesizes 3D interaction poses from text and then uses these poses as geometric anchors for detailed HOI scene generation, has demonstrated significant improvements over existing methods.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Dai et al.", + "bbox": [ + 271, + 114, + 336, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. We thank Zejia Su and Yunfan Ye for helpful discussions and paper proofreading. We also thank the anonymous reviewers for their valuable comments. This project was supported by the NSFC (62325211, 62132021, 62322207), Guangdong Natural Science Foundation (2021B1515020085), Shenzhen Science and Technology Program (RCYX20210609103121030), and the Major Program of Xiangjiang Laboratory (23XJ01009).", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 258, + 323, + 273 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bhatnagar, B.L., Xie, X., Petrov, I.A., Sminchisescu, C., Theobalt, C., Pons-Moll, G.: Behave: Dataset and method for tracking human object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15935-15946 (2022)", + "2. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)", + "3. Chen, K., Choy, C.B., Savva, M., Chang, A.X., Funkhouser, T., Savarese, S.: Text2shape: Generating shapes from natural language by learning joint embeddings. In: Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part III 14. pp. 100-116. Springer (2019)", + "4. Cohen-Bar, D., Richardson, E., Metzer, G., Giryes, R., Cohen-Or, D.: Set-the-scene: Global-local training for generating controllable nef scenes. arXiv preprint arXiv:2303.13450 (2023)", + "5. Deepfloyd IF. https://github.com/deep-floyd/IF, 2023", + "6. Diller, C., Dai, A.: Cg-hoi: Contact-guided 3d human-object interaction generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19888-19901 (2024)", + "7. Fan, Z., Taheri, O., Tzionas, D., Kocabas, M., Kaufmann, M., Black, M.J., Hilliges, O.: Articulated objects in free-form hand interaction. arXiv preprint arXiv:2204.13662 (2022)", + "8. Feng, Y., Choutas, V., Bolkart, T., Tzionas, D., Black, M.J.: Collaborative regression of expressive bodies using moderation. In: 2021 International Conference on 3D Vision (3DV). pp. 792-804. IEEE (2021)", + "9. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5152-5161 (2022)", + "0. Guo, M., Fathi, A., Wu, J., Funkhouser, T.: Object-centric neural scene rendering. arXiv preprint arXiv:2012.08503 (2020)", + "11. Guo, Y.C., Liu, Y.T., Shao, R., Laforte, C., Voleti, V., Luo, G., Chen, C.H., Zou, Z.X., Wang, C., Cao, Y.P., Zhang, S.H.: threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio (2023)", + "12. Hassan, M., Choutas, V., Tzionas, D., Black, M.J.: Resolving 3d human pose ambiguities with 3d scene constraints. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2282-2292 (2019)", + "13. Hassan, M., Ghosh, P., Tesch, J., Tzionas, D., Black, M.J.: Populating 3d scenes by learning human-scene interaction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14708-14718 (2021)" + ], + "bbox": [ + 218, + 287, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "InterFusion", + "bbox": [ + 651, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Hong, F., Zhang, M., Pan, L., Cai, Z., Yang, L., Liu, Z.: Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535 (2022)", + "15. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence 36(7), 1325-1339 (2013)", + "16. Jahan, T., Guan, Y., Van Kaick, O.: Semantics-guided latent space exploration for shape generation. In: Computer Graphics Forum. vol. 40, pp. 115-126. Wiley Online Library (2021)", + "17. Jain, A., Mildenhall, B., Barron, J.T., Abbeel, P., Poole, B.: Zero-shot text-guided object generation with dream fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 867-876 (2022)", + "18. Kim, V.G., Chaudhuri, S., Guibas, L., Funkhouser, T.: Shape2Pose: Human-centric shape analysis. ACM Transactions on Graphics (Proc. SIGGRAPH) (Aug 2014)", + "19. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)", + "20. Lee, H.H., Chang, A.X.: Understanding pure clip guidance for voxel grid nerf models. arXiv preprint arXiv:2209.15172 (2022)", + "21. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440 (2022)", + "22. Lin, Y., Bai, H., Li, S., Lu, H., Lin, X., Xiong, H., Wang, L.: Componerf: Text-guided multi-object compositional nerf with editable 3d scene layout. arXiv preprint arXiv:2303.13843 (2023)", + "23. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: European Conference on Computer Vision. pp. 423-439. Springer (2022)", + "24. Liu, Z., Wang, Y., Qi, X., Fu, C.W.: Towards implicit text-guided 3d shape generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17896-17906 (2022)", + "25. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Spl: A skinned multi-person linear model. ACM transactions on graphics (TOG) 34(6), 1-16 (2015)", + "26. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5442-5451 (2019)", + "27. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latentnerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600 (2022)", + "28. Michel, O., Bar-On, R., Liu, R., Benaim, S., Hanocka, R.: Text2mesh: Text-driven neural stylization for meshes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13492-13502 (2022)", + "29. Mihajlovic, M., Saito, S., Bansal, A., Zollhoefer, M., Tang, S.: Coap: Compositional articulated occupancy of people. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13201-13210 (2022)", + "30. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Dai et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Mirzaei, A., Kant, Y., Kelly, J., Gilitschenski, I.: Laterf: Label and text driven object radiance fields. In: European Conference on Computer Vision. pp. 20-36. Springer (2022)", + "32. Mohammad Khalid, N., Xie, T., Belilovsky, E., Popa, T.: Clip-mesh: Generating textured meshes from text using pretrained image-text models. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-8 (2022)", + "33. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)", + "34. Nguyen-Phuoc, T.H., Richardt, C., Mai, L., Yang, Y., Mitra, N.: Blockgan: Learning 3d object-aware scene representations from unlabelled images. Advances in neural information processing systems 33, 6767-6778 (2020)", + "35. Niemeyer, M., Geiger, A.: Giraffe: Representing scenes as compositional generative neural feature fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11453-11464 (2021)", + "36. Ost, J., Mannan, F., Thuerey, N., Knodt, J., Heide, F.: Neural scene graphs for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2856-2865 (2021)", + "37. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10975-10985 (2019)", + "38. Peng, X., Xie, Y., Wu, Z., Jampani, V., Sun, D., Jiang, H.: Hoi-diff: Text-driven synthesis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553 (2023)", + "39. Po, R., Wetzstein, G.: Compositional 3d scene generation using locally conditioned diffusion. arXiv preprint arXiv:2303.12218 (2023)", + "40. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)", + "41. Puig, X., Undersander, E., Szot, A., Cote, M.D., Yang, T.Y., Partsey, R., Desai, R., Clegg, A.W., Hlavac, M., Min, S.Y., et al.: Habitat 3.0: A co-habitat for humans, avatars and robots. arXiv preprint arXiv:2310.13724 (2023)", + "42. Punnakkal, A.R., Chandrasekaran, A., Athanasiou, N., Quiros-Ramirez, A., Black, M.J.: Babel: Bodies, action and behavior with english labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 722-731 (2021)", + "43. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022)", + "45. Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610 (2022)", + "46. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "InterFusion", + "bbox": [ + 651, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Savva, M., Chang, A.X., Hanrahan, P., Fisher, M., Nießner, M.: Pigraphs: learning interaction snapshots from observations. ACM Transactions on Graphics (TOG) 35(4), 1-12 (2016)", + "48. Sella, E., Fiebelman, G., Hedman, P., Averbuch-Elor, H.: Vox-e: Text-guided voxel editing of 3d objects. arXiv preprint arXiv:2303.12048 (2023)", + "49. Sheridan, T.B.: Human-robot interaction: status and challenges. Human factors 58(4), 525-532 (2016)", + "50. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023)", + "51. Sohn, K., Lee, H., Yan, X.: Learning structured output representation using deep conditional generative models. Advances in neural information processing systems 28 (2015)", + "52. Song, Y., Kong, C., Lee, S., Kwak, N., Lee, J.: Towards efficient neural scene graphs by learning consistency fields. arXiv preprint arXiv:2210.04127 (2022)", + "53. Taheri, O., Ghorbani, N., Black, M.J., Tzionas, D.: Grab: A dataset of whole-body human grasping of objects. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 581-600. Springer (2020)", + "54. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023)", + "55. Von Marcard, T., Henschel, R., Black, M.J., Rosenhahn, B., Pons-Moll, G.: Recovering accurate 3d human pose in the wild using imus and a moving camera. In: Proceedings of the European conference on computer vision (ECCV). pp. 601-617 (2018)", + "56. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774 (2022)", + "57. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems 36 (2024)", + "58. Wu, Q., Shi, Y., Huang, X., Yu, J., Xu, L., Wang, J.: Thor: Text to human-object interaction diffusion via relation intervention. arXiv preprint arXiv:2403.11208 (2024)", + "59. Wu, Q., Liu, X., Chen, Y., Li, K., Zheng, C., Cai, J., Zheng, J.: Object-compositional neural implicit surfaces. In: European Conference on Computer Vision. pp. 197-213. Springer (2022)", + "60. Xu, Y., Chai, M., Shi, Z., Peng, S., Skorokhodov, I., Siarohin, A., Yang, C., Shen, Y., Lee, H.Y., Zhou, B., et al.: Discoscope: Spatially disentangled generative radiance fields for controllable 3d-aware scene synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4402-4412 (2023)", + "61. Yang, B., Zhang, Y., Xu, Y., Li, Y., Zhou, H., Bao, H., Zhang, G., Cui, Z.: Learning object-compositional neural radiance field for editable scene rendering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13779-13788 (2021)", + "62. Zhang, S., Zhang, Y., Ma, Q., Black, M.J., Tang, S.: Place: Proximity learning of articulation and contact in 3d environments. In: 2020 International Conference on 3D Vision (3DV). pp. 642-651. IEEE (2020)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Dai et al.", + "bbox": [ + 271, + 114, + 336, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "63. Zhang, Y., Hassan, M., Neumann, H., Black, M.J., Tang, S.: Generating 3d people in scenes without people. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6194-6204 (2020)", + "64. Zhao, K., Wang, S., Zhang, Y., Beeler, T., Tang, S.: Compositional human-scene interaction synthesis with semantic control. In: European Conference on Computer Vision. pp. 311-327. Springer (2022)", + "65. Zhi, S., Laidlow, T., Leutenegger, S., Davison, A.J.: In-place scene labelling and understanding with implicit scene representation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15838-15847 (2021)" + ], + "bbox": [ + 215, + 146, + 787, + 272 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "InterFusion", + "bbox": [ + 651, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/InterFusion_ Text-Driven Generation of 3D Human-Object Interaction/d1b8e0cb-6c9e-44c9-80dd-158238285207_model.json b/2024/InterFusion_ Text-Driven Generation of 3D Human-Object Interaction/d1b8e0cb-6c9e-44c9-80dd-158238285207_model.json new file mode 100644 index 0000000000000000000000000000000000000000..980099d48ef71f4791f6be708a6a76afc1b3e448 --- /dev/null +++ b/2024/InterFusion_ Text-Driven Generation of 3D Human-Object Interaction/d1b8e0cb-6c9e-44c9-80dd-158238285207_model.json @@ -0,0 +1,2548 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.278, + 0.141, + 0.73, + 0.187 + ], + "angle": 0, + "content": "InterFusion: Text-Driven Generation of 3D Human-Object Interaction" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.212, + 0.768, + 0.228 + ], + "angle": 0, + "content": "Sisi Dai\\(^{1}\\) Wenhao Li\\(^{1}\\) Haowen Sun\\(^{2}\\) Haibin Huang\\(^{3}\\) Chongyang Ma\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.358, + 0.228, + 0.645, + 0.243 + ], + "angle": 0, + "content": "Hui Huang² Kai Xu1* Ruizhen Hu2*" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.255, + 0.652, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) National University of Defense Technology" + }, + { + "type": "text", + "bbox": [ + 0.425, + 0.269, + 0.578, + 0.282 + ], + "angle": 0, + "content": "2 Shenzhen University" + }, + { + "type": "text", + "bbox": [ + 0.423, + 0.283, + 0.581, + 0.297 + ], + "angle": 0, + "content": "3 Kuaishou Technology" + }, + { + "type": "text", + "bbox": [ + 0.353, + 0.298, + 0.649, + 0.311 + ], + "angle": 0, + "content": "https://sisidai.github.io/InterFusion/" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.353, + 0.391, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.247, + 0.416, + 0.383, + 0.437 + ], + "angle": 0, + "content": "a person in a military uniform sitting on a chair" + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.353, + 0.593, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.437, + 0.416, + 0.563, + 0.437 + ], + "angle": 0, + "content": "a female gymnastics athlete riding a motorcycle" + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.353, + 0.768, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.416, + 0.762, + 0.437 + ], + "angle": 0, + "content": "a person in a fleece pullover and hiking pants carrying a backpack" + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.443, + 0.393, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.508, + 0.396, + 0.529 + ], + "angle": 0, + "content": "a man in a quilted gilet and moleskin trousers playing the guitar" + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.442, + 0.588, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.437, + 0.509, + 0.564, + 0.529 + ], + "angle": 0, + "content": "a man wearing a WORLD lettered T-shirt riding a bike" + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.444, + 0.773, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.508, + 0.762, + 0.529 + ], + "angle": 0, + "content": "a person in a cropped hoodie and bike shorts jumping a skateboard" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.535, + 0.395, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.602, + 0.406, + 0.622 + ], + "angle": 0, + "content": "a man in a velvet smoking jacket and silk pajama pants playing the saxophone" + }, + { + "type": "image", + "bbox": [ + 0.416, + 0.536, + 0.582, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.601, + 0.567, + 0.622 + ], + "angle": 0, + "content": "a man in a paisley waistcoat and bowler hat riding a horse" + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.535, + 0.768, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.601, + 0.772, + 0.622 + ], + "angle": 0, + "content": "a man with blond hair wearing a brown leather jacket reading a laptop" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.641, + 0.788, + 0.669 + ], + "angle": 0, + "content": "Fig. 1: Given a text prompt, our method InterFusion can generate diverse 3D scenes of a person interacting with an object." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.72, + 0.741, + 0.818 + ], + "angle": 0, + "content": "Abstract. In this study, we tackle the complex task of generating 3D human-object interactions (HOI) from textual descriptions in a zero-shot text-to-3D manner. We identify and address two key challenges: the unsatisfactory outcomes of direct text-to-3D methods in HOI, largely due to the lack of paired text-interaction data, and the inherent difficulties in simultaneously generating multiple concepts with complex spatial relationships. To effectively address these issues, we present InterFusion," + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.714, + 0.84 + ], + "angle": 0, + "content": "* Corresponding authors: kevin.kai.xu@gmail.com; ruizhen.hu@gmail.com." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.337, + 0.128 + ], + "angle": 0, + "content": "Dai et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.356 + ], + "angle": 0, + "content": "a two-stage framework specifically designed for HOI generation. InterFusion involves human pose estimations derived from text as geometric priors, which simplifies the text-to-3D conversion process and introduces additional constraints for accurate object generation. At the first stage, InterFusion extracts 3D human poses from a synthesized image dataset depicting a wide range of interactions, subsequently mapping these poses to interaction descriptions. The second stage of InterFusion capitalizes on the latest developments in text-to-3D generation, enabling the production of realistic and high-quality 3D HOI scenes. This is achieved through a local-global optimization process, where the generation of human body and object is optimized separately, and jointly refined with a global optimization of the entire scene, ensuring a seamless and contextually coherent integration. Our experimental results affirm that InterFusion significantly outperforms existing state-of-the-art methods in 3D HOI generation." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.37, + 0.75, + 0.398 + ], + "angle": 0, + "content": "Keywords: Text-Driven Generation \\(\\cdot\\) Zero-Shot Generation \\(\\cdot\\) 3D Human-Object Interaction Generation" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.427, + 0.377, + 0.442 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.552 + ], + "angle": 0, + "content": "The generation of 3D human-object interactions (HOI) stands as a critical challenge in the fields of computer vision and computer graphics, with far-reaching implications in virtual reality, augmented reality, animation, and embodied AI [41, 49]. This task entails the creation of realistic 3D scenes where human figures interact with objects in ways that are not only physically plausible but also contextually relevant." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Despite its potential, the field has faced significant obstacles, primarily due to the scarcity of large-scale interaction data. Traditional approaches have predominantly relied on motion capture (mocap) datasets or physics-based simulations for generating these interactions. Mocap datasets [15,26,55] are limited by the specific scenarios they capture and are both costly and labor-intensive to produce. These limitations have resulted in a notable gap in generating diverse and contextually rich HOI scenes, especially for novel or complex interactions. Conversely, recent advancements have introduced text-to-3D methods [40,56], marking a significant shift in the field. These methods harness the power of textual descriptions to generate 3D objects without direct 3D supervision, presenting a novel approach to 3D content creation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this study, we explore text-to-3D method in HOI task within a zero-shot manner, i.e., generating 3D scenes from textual descriptions using 2D diffusion models. Our key observations are two-fold: first of all, a direct application of text-to-3D method in HOI often leads to unsatisfied results like blurry textures and incorrect interactions, which is caused by the relative lack of paired text-interaction data during training and the difficulties of generating multiple concepts for diffusion-based methods [23]. Secondly, while collecting extensive interaction data is challenging, estimating human poses based on described in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InterFusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "teractions is more feasible. These pose estimations can serve as geometric priors in the HOI generation process. By integrating pose estimation, our method substantially simplifies the text-to-3D process, particularly in the geometry optimization stage for human body generation. It also provides additional constraints for object generation, ensuring accurate placement and alignment with the text's semantic content. More importantly, we can separate the generation of human body and object, and jointly refine the details with the estimated pose, allowing for a more coherent and detailed synthesis of the interaction scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.277, + 0.789, + 0.61 + ], + "angle": 0, + "content": "Inspired by these observations, we introduce InterFusion, a novel two-stage framework designed for HOI generation. Specifically, in the first stage, rather than relying on precise 3D interaction data, our approach instead collects a comprehensive dataset of synthesized images that depict a wide range of interactions. From these images, we employ advanced 3D pose estimation technique [8] to extract 3D human poses. Building upon these image-pose pairs, InterFusion develops a sophisticated codebook that establishes a mapping between interaction descriptions and 3D human poses, with the integration of the CLIP (Contrastive Language-Image Pretraining) embedding [43]. By leveraging these embeddings, our framework is able to interpret the nuances of interaction descriptions and translate them into accurate 3D pose representations. In the subsequent stage, InterFusion capitalizes on recent advancements in text-to-3D generation [40], as well as neural radiance fields [30], using the estimated human poses to produce 3D HOI scenes with realistic appearances and high-quality geometry. This stage operates in a 'local-global' manner. At the local level, the generation of the human body (SDS-H) and objects (SDS-O) is separately optimized, with the poses serving as additional constraints for the SDS. At the global level, the generation of the entire scene (SDS-I) is also guided by the integrated description and jointly optimized with SDS-H and SDS-O, ensuring a cohesive and contextually accurate representation of the HOI. Our experiments show that the quality of generation can be improved by a large margin and our approach outperforms state-of-the-art methods in HOI generation." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.617, + 0.584, + 0.632 + ], + "angle": 0, + "content": "To summarize, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.687, + 0.787, + 0.735 + ], + "angle": 0, + "content": "- We introduce a novel two-stage framework InterFusion, for zero-shot 3D human-object interaction generation from text, incorporating 3D pose estimation as geometry priors." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.741, + 0.787, + 0.788 + ], + "angle": 0, + "content": "- InterFusion leverages text-to-3D generation with a local-global optimization process. This strategy ensures seamless integration of human bodies and objects, producing realistic and high-quality 3D HOI scenes." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- InterFusion demonstrates significant improvements over existing methods in 3D HOI generation, showcasing its effectiveness in creating detailed, and contextually rich 3D interactions." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.687, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.337, + 0.129 + ], + "angle": 0, + "content": "Dai et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.39, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.181, + 0.566, + 0.198 + ], + "angle": 0, + "content": "2.1 Human-Object Interaction Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.21, + 0.789, + 0.589 + ], + "angle": 0, + "content": "3D human-object interaction (HOI) generation is a challenging problem that has been studied widely by the computer vision and graphics community. Shape2Pose [18] generates plausible 3D human poses interacting with a given 3D object model by learning an affordance model from synthetic data. PiGraphs [47] learns the distribution between human poses and object arrangements from a collected dataset with 3D scene scans and RGB-D videos, generating the interaction snapshots given action specifications and object models. Recently, the parametric human body models such as SMPL [25,37,45] are employed in interaction synthesis to overcome the lack of realism due to human body representations. Benefiting from the PROX dataset [12] which consists of fitted SMPL models in captured 3D scenes, the more specific human-scene interaction has become an active research direction. Given a 3D scene, PSI [63] and PLACE [62] generate the 3D human body mesh represented as SMPL parameters through a conditional variational autoencoder [19, 51]. POSA [13] learns pose-specific priors to generate contacts conditioned by the given posed human, which can further guide the placement of the body mesh in a scene. COINS [64] enables semantic control on interaction synthesis by embedding the action label together with the interacted object as the condition of the generative model. More recently, fine-grained 3D interaction datasets [1, 7, 53] are captured to promote this field. Relying on these datasets, some methods [6, 38, 58], concurrent to this work, are proposed to explore text-guided 3D HOI generation. However, they remain constrained by distributions within the datasets. While previous methods need ground truth 3D interaction data as supervision, our work, for the first time, attempts to break through the limitation of data requirement. We generate a wider range of realistic and detailed 3D HOI scenes, including both indoors and outdoors." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.615, + 0.511, + 0.631 + ], + "angle": 0, + "content": "2.2 Text-to-3D Content Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.79, + 0.844 + ], + "angle": 0, + "content": "Early methods [3, 16, 24] for text-to-3D shape generation require paired data of 3D data and the corresponding textual descriptions to learn the joint embedding space of shape and text for supervision, which limits their generality to unseen object categories. Benefiting from large pre-trained text-to-image models and differentiable rendering techniques, breakthroughs in text-to-3D content generation have been achieved. For example, DreamFields [17] and PureCLIPN-eRF [20] combine CLIP [43] with neural radiance fields (NeRF) [30], demonstrating the potential for zero-shot NeRF optimization. Meanwhile, CLIP-mesh [32] and Text2Mesh [28] incorporate CLIP to optimize the 3D mesh representation, starting from an initial sphere mesh and an input base mesh, respectively. Recently, DreamFusion [40] and SJC [56] enable NeRF optimization with guidance from pre-trained text-to-image diffusion models [44,46] in place of CLIP, achieving more impressive results. To improve DreamFusion, Magic3D [21] proposes" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InterFusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "a coarse-to-fine pipeline to generate the fine-grained mesh. TextMesh [54] extends the geometry representation from NeRF to an SDF framework, thereby enhancing detailed mesh extraction and photorealistic rendering. Among follow-up works, Latent-NeRF [27] and Vox-E [48] utilize explicit 3D shapes to provide additional training signals for NeRF optimization. While Latent-NeRF utilizes a rough untextured object for shape sculpture, Vox-E takes multiview images of a fine-grained textured object to edit geometry and appearance. All the methods mentioned above focus on the generation or edition of a single subject, while ignoring the interaction between different subjects." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.307, + 0.533, + 0.323 + ], + "angle": 0, + "content": "2.3 Compositional Scene Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.334, + 0.789, + 0.545 + ], + "angle": 0, + "content": "Representing scenes as compositions of object representations facilitates enhanced controllability. Numerous techniques incorporate additional information at the object level to perform compositional modeling of scenes, effectively separating object representations from the overall scene imagery. For instance, some methods incorporate 2D semantic information such as segmentation labels [65], instance masks [59,61], or features from a pre-trained vision-language model [31]. Some other methods [10, 36, 52] use 3D layout information by object-centric bounding boxes with canonical coordinates. When it comes to scene generation, [34, 35, 60] use compositional representations to generate scenes in a controllable manner. More recent approaches [4, 22, 39] generate compositional 3D scenes from input text prompts. Different from existing methods, our HOI scene is generated automatically without the requirement of input layout. Moreover, the spatial relationship between human and object during interaction is much more complex and cannot be simply characterized using their bounding boxes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.569, + 0.377, + 0.585 + ], + "angle": 0, + "content": "3 Our Method" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.602, + 0.342, + 0.617 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.734 + ], + "angle": 0, + "content": "In this section, we formally introduce InterFusion with a focus on the zero-shot and text-driven generation of 3D human-object interactions (HOI). Specifically, the input is a triplet of text descriptions, \\( T = \\{T^H, T^O, T^I\\} \\), specifying the desired human style, object style, and interaction type. The goal is to generate a detailed 3D scene, \\( \\psi = \\{\\psi^H, \\psi^O\\} \\), comprising a human model and an object model that not only adhere to the specified appearance styles but also exhibit a tailored spatial relationship to accurately reflect the described interaction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "As illustrated in Figure 2, our method consists of two primary stages: anchor pose generation and anchor pose guided HOI generation. We first generate an interaction pose based on the input text, termed an anchored pose. This pose then serves as a geometric constraint, guiding the subsequent generation of detailed HOI. In the second phase, the human and object models are optimized separately and refined simultaneously with a global context, ensuring a cohesive and accurate representation of the interaction as described by the input text." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.128 + ], + "angle": 0, + "content": "Dai et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.149, + 0.784, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.447 + ], + "angle": 0, + "content": "Fig. 2: InterFusion is a two-stage framework that transforms textual descriptions into detailed 3D human-object interactions, initially synthesizing anchor poses (upper left) and then optimizing the human model (upper right) and object model (bottom) with constraints from estimated pose and textual prompts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.476, + 0.373, + 0.491 + ], + "angle": 0, + "content": "3.2 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.789, + 0.641 + ], + "angle": 0, + "content": "SDS. Score Distillation Sampling (SDS) has been introduced by DreamFusion [40]. While \\( x = g(\\psi, \\zeta) \\), \\( x \\) is the 2D image rendered by a differentiable renderer \\( g \\) with model parameters \\( \\psi \\) (the volumetric renderer and MLPs correspondingly in NeRF), under a desired camera pose \\( \\zeta \\). By injecting the sampled noise \\( \\epsilon \\) into \\( x \\) at a time step \\( t \\), the noisy image \\( x_{t} \\) is produced. The pre-trained 2D text-to-image diffusion model \\( \\phi \\) provides a denoising network \\( \\hat{\\epsilon}_{\\phi}(x_{t};y,t) \\) that predicts the noise \\( \\hat{\\epsilon} \\) given the noisy image \\( x_{t} \\), time step \\( t \\), and text embedding \\( y \\). SDS then optimizes the model parameters \\( \\psi \\) by minimizing the difference between the predicted noise and the added noise:" + }, + { + "type": "equation", + "bbox": [ + 0.334, + 0.652, + 0.787, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\psi} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , x) = \\mathbb {E} _ {t, \\epsilon} [ w (t) (\\hat {\\epsilon} _ {\\phi} (x _ {t}; y, t) - \\epsilon) \\frac {\\partial x}{\\partial \\psi} ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.696, + 0.591, + 0.712 + ], + "angle": 0, + "content": "where \\( w(t) \\) is the weighting term at the time step \\( t \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.736, + 0.466, + 0.751 + ], + "angle": 0, + "content": "3.3 Anchor Pose Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.841 + ], + "angle": 0, + "content": "InterFusion starts by generating the anchor pose from the input text \\( T^I \\), which is a non-trivial problem and highly limited by the available pose datasets. Existing datasets (e.g., HumanML3D [9], BABEL [42]) are mocap based datasets and are typically conducted in a controlled laboratory environment, leading to lacks in action diversity and spatial coverage." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InterFusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.328 + ], + "angle": 0, + "content": "Recently, text-to-image diffusion models have shown the versatility and effectiveness of generating realistic and diverse images, through the integration of visual and linguistic feature spaces. We thus instead utilize advancements in this technology, specifically models like Stable Diffusion, to create a large-scale dataset of Human-Object Interaction (HOI) images. These images depict a wide range of human interactions. To ensure the diversity of interaction types, we utilize ChatGPT to generate prompts about human daily events or actions forming \"verb-ing a/an/the object\". A total of 235 result prompts are generated, covering most interactions in daily life. After filtering synthesized images without humans, we then estimate 3D human poses using the pre-trained PixIE [8] model, creating a comprehensive Syn-HOI pose dataset consisting of a total 55K 3D pseudo-SMPL poses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.785, + 0.483 + ], + "angle": 0, + "content": "While SMPL model maps pose and shape parameters to a triangulated mesh, to align these poses with our text-to-interaction generation task, we render images from multiple perspectives and use CLIP's image encoder to derive pose feature embeddings. The average feature across these multiple perspective images represents the pose feature and attaches the dataset with pairs of averaged CLIP embedding and pose parameters. We further construct a codebook using K-Means clustering to identify key pose centroids based on the feature embeddings. 2,048 cluster centroids are clustered compositing our pose code-book, where each cluster comprises a subset of poses that represents similar interaction as the key poses of the centroids." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.785, + 0.565 + ], + "angle": 0, + "content": "Once the codebook is built, for a given input text, we extract its feature embedding using CLIP's text encoder. This serves as a query to retrieve the most similar poses from the codebook, based on feature similarity. Specifically, given query text \\( T^{I} \\), top \\( k \\) poses \\( \\theta_{k}^{T^{I}} \\) could be matched by pose embeddings \\( \\theta_{E} \\), the averaged CLIP embeddings among rendered images from key poses:" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.583, + 0.785, + 0.603 + ], + "angle": 0, + "content": "\\[\n\\theta_ {k} ^ {T ^ {I}} = \\operatorname {T O P} _ {k} \\left(f _ {\\text {t e x t}} \\left(T ^ {I}\\right), \\theta_ {E}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.784, + 0.666 + ], + "angle": 0, + "content": "where \\( f_{text} \\) is the text encoder of CLIP and \\( \\mathrm{TOP}_k(X,Y) \\) returns top \\( k \\) poses with the top \\( k \\) highest cosine similarities, and we use \\( k = 7 \\) for suitable poses in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.784, + 0.746 + ], + "angle": 0, + "content": "We then utilize GPT-4V to select the most precise pose as the final queried key pose. Depending on the requirement, we can instead select the poses sampled in the cluster, corresponding to the key pose, to guide the generation process for diverse results. This approach ensures rich and contextually aligned anchor poses for our text-to-interaction generation task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.784, + 0.84 + ], + "angle": 0, + "content": "To restrict the optimization of geometry and appearance using human structural priors from the acquired pose, we further incorporate COAP [29], a neural occupancy representation of the articulated human body based on SMPL parameters [25]. Given a pose \\(\\theta\\) and a shape \\(\\beta\\), COAP offers an occupancy prediction network \\(f(x; \\beta, \\theta)\\) that maps a 3D query point \\(p\\) to an occupancy value, directly indicating whether the spatial point resides within the 3D body." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.337, + 0.128 + ], + "angle": 0, + "content": "Dai et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.506, + 0.162 + ], + "angle": 0, + "content": "3.4 Pose-Guided HOI Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.266 + ], + "angle": 0, + "content": "Once get the interaction pose, we further use it along with the input text to guide the generation of a detailed 3D HOI scene. In this phase, the estimated pose serves both as spatial constraints for the scene's geometry and as an anchor that aligns the human and object models. This approach ensures that the human model and the object model can be generated separately, while they are cohesively aligned to form the final 3D HOI scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.267, + 0.788, + 0.418 + ], + "angle": 0, + "content": "The anchor pose provides specific spatial constraints for each component in the scene. For the human model, it establishes a basic geometric structure, while for the object model, it defines the areas that should remain unoccupied. This clear distinction is essential for rendering the scene both accurately and realistically. Simultaneously, the input text undergoes a complex processing procedure to offer distinct semantic guidance for each component. This is accomplished through various text conditioning techniques applied in SDS with the pre-trained DeepFloyd model [5]. The text is carefully crafted to direct the generation of both the human and object models to ensure all elements are in harmony with the semantic nuances of input descriptions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Additionally, we introduce a novel camera tracing module to enhance the optimization process under varying text conditions. This module adaptively adjusts the camera pose, focusing on relevant elements within the scene at each optimization stage. This adaptive camera positioning is instrumental in ensuring that each aspect of the scene is optimally rendered according to the text descriptions, resulting in a more dynamic and contextually accurate 3D HOI scene." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.524, + 0.788, + 0.572 + ], + "angle": 0, + "content": "We now provide details of each module in this stage, including the Neural Radiance Field representations for the human model (H-NeRF) and object model (O-NeRF), the camera tracing mechanism, and the guided optimization process." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.596, + 0.789, + 0.764 + ], + "angle": 0, + "content": "H-NeRF. We use NeRF to represent the human model, noted as H-NeRF. The generation of H-NeRF is guided by the text description specifying the human style, conducting pose-specific human avatar generation with SDS. To enhance the quality of our renderings, particularly in terms of resolution, we incorporated a specialized optimization process that focuses on the head region of the human avatar. The location of the head for any given pose is determined with COAP [29]. This allows us to augment the text prompt specifically for the head region, using the notation * the head of *, to ensure that the head receives detailed attention during the generation process. The loss function for this optimization process is as follows, designed to balance the fidelity of the head region with the overall pose and style of the human figure." + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.777, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\psi^ {H}} \\mathcal {L} _ {\\mathrm {S D S}} = \\mathbb {E} _ {t, \\epsilon} [ w (t) (\\hat {\\epsilon} _ {\\phi} \\left(x _ {t} ^ {H}; y ^ {H}, t\\right) - \\epsilon) \\frac {\\partial x ^ {H}}{\\partial \\psi^ {H}} ] \\tag {3} \\\\ + \\mathbb {E} _ {t, \\epsilon} [ w (t) (\\hat {\\epsilon} _ {\\phi} (x _ {t} ^ {H, h}; y ^ {H, h}, t) - \\epsilon) \\frac {\\partial x ^ {H , h}}{\\partial \\psi^ {H}} ], \\\\ \\end{array}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InterFusion" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "In shaping H-NeRF, our geometric constraint ensures that the human model evolves directly from the anchor pose. Points within the anchor are required to be occupied, establishing a firm base for the model. Meanwhile, points outside the anchor can also be occupied, but with probabilities that decrease as they move away from the anchor's surface. This approach allows for the addition of geometric details over the anchor, ensuring the model aligns with the human style described in the text while maintaining a coherent structure rooted in the anchor pose. The loss function used is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.279, + 0.785, + 0.322 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {g e o}} ^ {H} = C E _ {p _ {i} \\in \\mathbb {P} _ {\\mathrm {i n}}} \\left(\\alpha_ {i}, f (p _ {i})\\right) \\\\ + C E _ {p _ {j} \\in \\mathbb {P} _ {\\text {o u t}}} \\left(\\alpha_ {j}, f (p _ {j})\\right) \\left(1 - e ^ {- \\frac {d}{2 \\eta^ {2}}}\\right), \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.786, + 0.358 + ], + "angle": 0, + "content": "where \\(d\\) represents the point distance from the anchor surface, and \\(\\eta\\) is the hyperparameter to control the extent of decaying, similar as in [27]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.379, + 0.788, + 0.455 + ], + "angle": 0, + "content": "O-NeRF. We also employ NeRF to model the object component, referred to as O-NeRF. This model is designed to interact seamlessly with the human model and to embody the desired style as specified by the input text. To guide the generation of both the interaction type and the object style, we utilize SDS with tailored text prompts:" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.465, + 0.785, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\psi^ {O}} \\mathcal {L} _ {\\mathrm {S D S}} = \\mathbb {E} _ {t, \\epsilon} [ w (t) (\\hat {\\epsilon} _ {\\phi} \\left(x _ {t} ^ {I}; y ^ {I}, t\\right) - \\epsilon) \\frac {\\partial x ^ {I}}{\\partial \\psi^ {O}} ] \\tag {5} \\\\ + \\mathbb {E} _ {t, \\epsilon} [ w (t) (\\hat {\\epsilon} _ {\\phi} (x _ {t} ^ {O}; y ^ {O}, t) - \\epsilon) \\frac {\\partial x ^ {O}}{\\partial \\psi^ {O}} ], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.788, + 0.636 + ], + "angle": 0, + "content": "For the interaction scene \\( x^{I} \\) generation, we use an alpha-composed rendering to integrate H-NeRF and O-NeRF. This method calculates an alpha value from the density at each point, determining its contribution to the scene's color. Higher alpha values indicate a greater influence on the rendering, allowing for a nuanced and realistic integration of the human and object models in the final interaction image:" + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.648, + 0.785, + 0.693 + ], + "angle": 0, + "content": "\\[\nx ^ {I} = \\sum_ {i} w _ {i} ^ {I} c _ {i} ^ {I}, w _ {i} ^ {I} = \\alpha_ {i} ^ {I} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j} ^ {I}), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.692, + 0.606, + 0.726 + ], + "angle": 0, + "content": "\\[\nc _ {i} ^ {I} = \\frac {\\alpha_ {i} ^ {H}}{\\alpha_ {i} ^ {H} + \\alpha_ {i} ^ {O}} c _ {i} ^ {H} + \\frac {\\alpha_ {i} ^ {O}}{\\alpha_ {i} ^ {H} + \\alpha_ {i} ^ {O}} c _ {i} ^ {I}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In our NeRF-based approach, alpha values, capped at 1, are derived from density for composite rendering. This setup allows semantic guidance gradients, conditioned by the interaction image, to optimize the density and color of both the human and object models. However, we noted a tendency for the model to prioritize object generation at the expense of the human model's quality. To counteract this, we implemented gradient truncation towards the human model to maintain a balanced optimization between the two components." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.337, + 0.128 + ], + "angle": 0, + "content": "Dai et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "For O-NeRF's geometric constraints, we aim to prevent occupancy of points within the anchor pose, denoted as \\( x_{i} \\in \\mathbb{X}_{in} \\), by the object model. Inspired by the physical collision prevention concept, we define a specific loss function that ensures these points remain unoccupied by O-NeRF, effectively preventing overlap between the human and object models in the 3D space:" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.235, + 0.786, + 0.254 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {g e o}} ^ {O} = C E _ {p _ {i} \\in \\mathbb {P} _ {\\mathrm {i n}}} \\left(\\alpha_ {i}, 1 - f \\left(p _ {i}\\right)\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.265, + 0.787, + 0.295 + ], + "angle": 0, + "content": "By minimizing the above loss term, model parameters are optimized to enhance the geometric consistency between H-NeRF and O-NeRF." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.32, + 0.788, + 0.427 + ], + "angle": 0, + "content": "Camera tracing. To enhance the generation of O-NeRF, we introduce a dynamic camera tracing module within the SDS framework. This module automatically adjusts the camera pose to focus on the target's center, either the entire interaction scene or just the object during object generation. The camera aligns with the average position of voxels with an occupancy probability above 0.5, ensuring a consistent focus on the most significant parts of the scene or object for optimal detail capture and realism." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.451, + 0.618, + 0.466 + ], + "angle": 0, + "content": "Optimization. The total loss is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.336, + 0.478, + 0.786, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {S D S}} ^ {H} + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {S D S}} ^ {O} + \\mathcal {L} _ {\\mathrm {g e o}} ^ {H} + \\lambda_ {2} \\mathcal {L} _ {\\mathrm {g e o}} ^ {O} + \\lambda_ {3} L _ {\\text {r e g}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.584 + ], + "angle": 0, + "content": "where \\(\\lambda_1, \\lambda_2\\) and \\(\\lambda_3\\) are the corresponding loss weights. Figure 2 (left bottom) shows several intermediate states of the object model during the generation steps and we can see that the object is gradually generated with fine details. Weight annealing is adopted during the guided optimization process. We leave the details in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.61, + 0.376, + 0.627 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Figure 3 (left) shows results with various interaction poses, showcasing the strength of our method. Our method can support diverse interaction poses within a single type. Meanwhile, with interaction type and pose fixed, our method can further generate more numerous results under different human styles or object styles, as shown in Figure 3 (right). Moreover, our InterFusion also supports controllable text-conditioned editing, providing users more control over the already generated 3D models, which are presented in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We both qualitatively and quantitatively evaluate our method against alternative baseline methods. To verify the effectiveness of individual components in our method, we conduct ablation studies. Furthermore, we discuss the application potential, limitations and future work of our InterFusion. The implementation details, more evaluations, and further discussions are presented in the supplementary materials." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InterFusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.145, + 0.784, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.36, + 0.785, + 0.389 + ], + "angle": 0, + "content": "Fig. 3: More results generated by InterFusion. Diverse integration poses are supported. Numerous human and object styles are also supported." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.789, + 0.678 + ], + "angle": 0, + "content": "Evaluation prompts and criteria. Similar as the first stage for generating prompts about various interaction types, we also use ChatGPT to randomly generate prompts about human styles and collect the text prompts for evaluation. We select 61 distinct and diverse prompts, ensuring the classes of Human (e.g., a policewoman, a teenager) with various styles (e.g., in a graphic tee and ripped jeans), Objects (e.g., a motorcycle, a guitar) and Interactions (e.g., riding, playing) are reasonable and evenly distributed. In total, we have 13 different types of interactions, covering contact areas across the whole body, which demonstrates the effectiveness of our method on diverse types of interactions compared to baseline methods. We calculate the CLIP scores between the input text prompts and different views of generated 3D human-object interactions, and then compare the means based on different evaluation prompts. The CLIP score measures the similarity between a prompt text for an image and the actual content of the image. We also provide an assessment using GPT-4V for selection, named GPT-4V select, which evaluates the completeness of objects and correctness of physical interactions across multiview rendered images. We refer to the supplementary material for assessment details and additional evaluation criteria." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Baseline approaches. To the best of our knowledge, the proposed method is among the first to generate 3D human-object interactions based on text inputs in a zero-shot manner. We thus compare our method to alternative text-to-3D methods, including DreamFusion [40], Magic3D [21], and TextMesh [54]. Additional comparisons with MVDream [50] and ProlificDreamer [57] are presented in the supplementary material. To justify our key idea of using human pose to guide the generation, we design an object-centric baseline (Ours-OC) as a variant of our method. Instead of using human pose as the prior, we retrieve an object with the given semantic category from ShapeNet dataset [2] and use it as" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.128 + ], + "angle": 0, + "content": "Dai et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.272, + 0.145, + 0.731, + 0.16 + ], + "angle": 0, + "content": "Table 1: Quantitative evaluation of CLIP score and GPT-4V select." + }, + { + "type": "table", + "bbox": [ + 0.234, + 0.172, + 0.77, + 0.223 + ], + "angle": 0, + "content": "
MethodDreamFusion [40] Magic3D [21] TextMesh [54] Ours-OCOurs-HC
CLIP score0.3027 0.3179 0.2761 0.32030.3308
GPT-4V select(%)8.20 11.48 1.64 13.1165.57
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.251, + 0.788, + 0.328 + ], + "angle": 0, + "content": "geometry constraints to guide both human and object generation. As official implementations are unavailable for some of these baselines, we use the third-party re-implementations provided by threestudio [11] for a fair comparison. Note that all re-implementations use multi-resolution hash-grid [33] for 3D representation and DeepFloyd [5] for guidance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.351, + 0.43, + 0.367 + ], + "angle": 0, + "content": "4.1 Comparison Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.789, + 0.635 + ], + "angle": 0, + "content": "Qualitative evaluations. Some representative visual comparisons are shown in Figure 4, and additional comparisons including those with our object-centric baseline are presented in the supplementary material. The overall results of the baselines reveal common uncertainties in both geometry and appearance attributed to the confusion introduced by multi-concept guidance. Specifically, baseline models may lean towards one specific target, as seen in the example of \"a man wearing a red baseball cap\" with only the red baseball cap generated, the example of \"a policeman\" with only the upper half of the human body generated, or the example of \"a man in a puffer jacket and wool beanie\" where the shopping cart is failed to be generated. Even with a relatively uniform attention distribution, the baseline model may encounter challenges in producing complete results, as demonstrated in the saxophone and violin examples (the 3rd and 4th column in Figure 4), which exhibit deficiencies such as incomplete human body parts, and mixing of human legs. Moreover, the baseline model sometimes struggles to effectively choose the focus of generation when the input text describes a less common target like \"a person in a military uniform\", resulting in poor outcomes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.635, + 0.788, + 0.697 + ], + "angle": 0, + "content": "Our approach overcomes these issues by conducting optimization in an explicit decomposed way and intelligently guiding attention from SDS jointly in spatial and semantic aspects. Therefore, our method achieves more stable and higher-quality 3D results under multiple-concept guidance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Quantitative evaluations. The comparison results are shown in Table 1. We can see that our method (Ours-HC) achieves the best performance compared to baselines, showcasing our results exhibit more 3D plausibility with given text prompts. Among the baselines, Ours-OC gets the best performance. Though with object priors, the object-centric approach still does not provide sufficient body priors for human generation, resulting in its inability to achieve complete interaction generation. Magic3D gets the next best performance among the baselines, as it usually generates the whole interaction scene comparing to other two" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InterFusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.149, + 0.775, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.384, + 0.785, + 0.413 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparison results with baselines. InterFusion generates more stable and higher-quality results and is more consistent with input interaction descriptions." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.43, + 0.782, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.594, + 0.655, + 0.608 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results of ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.785, + 0.671 + ], + "angle": 0, + "content": "basielines, however, the results are somewhat vague as shown in Figure 4, thus there is still a large performance gap comparing to our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.791, + 0.717 + ], + "angle": 0, + "content": "The significant enhancement in performance can be attributed to our method's effective concurrent generation of both human and object, along with cohesive interactive information." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.748, + 0.401, + 0.761 + ], + "angle": 0, + "content": "4.2 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We conduct several ablation studies both qualitatively and quantitatively to show the importance of our design as well as introduced loss terms. Results are presented in Figure 5a, Figure 5b and Table 2. We refer to the supplementary material for additional results and details." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.337, + 0.128 + ], + "angle": 0, + "content": "Dai et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.335, + 0.145, + 0.668, + 0.16 + ], + "angle": 0, + "content": "Table 2: Quantitative results of ablation studies." + }, + { + "type": "table", + "bbox": [ + 0.307, + 0.172, + 0.699, + 0.227 + ], + "angle": 0, + "content": "
Settingsw/o ISDSw/o L0SDSw/o L0geoOurs
CLIP score0.31640.32930.31710.3308
GPT-4V select(%)1.644.9216.3977.05
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.254, + 0.788, + 0.392 + ], + "angle": 0, + "content": "Syn-HOI-pose. The anchor pose obtained at the first stage provides the geometry constraints for our HOI generation. We thus first demonstrate the effectiveness of synthesizing anchoring poses using our pseudo-pose dataset. We compare the pose generated with the existing approach AvatarCLIP [14], which utilizes CLIP only to retrieve the queried pose from the codebook constructed from the mocap dataset AMASS [26]. Figure 5a shows that the proposed pose generation stage enables a better-fitted anchor pose for input interaction types, which demonstrates our pseudo-poses, reconstructed from synthesized images, has the potential for more diverse poses' requirements than existing mocap datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.788, + 0.55 + ], + "angle": 0, + "content": "Semantic guidance. To better illustrate the efficacy of semantic guidance in our NeRF-based generation, we conduct ablations separately on SDS from object and interaction. As shown in Figure 5b and Table 2, in the absence of SDS from object, semantic consistency for the object is compromised, resulting in noisy outcomes influenced by the human, thus the contact region can not be extracted correctly. SDS from interaction plays the most crucial role in the performance, making results more semantically and visually plausible. The absence of SDS from interaction leads to generated objects being confined to the remaining space, but without interaction with the human." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.788, + 0.693 + ], + "angle": 0, + "content": "Geometric constraint. Relying solely on semantic guidance is inadequate for achieving the final objectives, as the object should be generated in spaces outside the human body, with its generation targets not positioned at the origin. As illustrated in Figure 5b and Table 2, depending only on semantic guidance may yield results that appear spatially conflict with the human body. Without the spatial constraint, the object generation at the origin would additionally have a conflict with the interaction objective, thus resulting in degenerated objects and final interactions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.718, + 0.36, + 0.733 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In conclusion, our work presents InterFusion, a novel framework for zero-shot 3D human-object interaction generation. InterFusion tackles the challenges of limited 3D interaction data and the complexity of generating multiple concepts simultaneously. Our two-stage approach, which synthesizes 3D interaction poses from text and then uses these poses as geometric anchors for detailed HOI scene generation, has demonstrated significant improvements over existing methods." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InterFusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Acknowledgements. We thank Zejia Su and Yunfan Ye for helpful discussions and paper proofreading. We also thank the anonymous reviewers for their valuable comments. This project was supported by the NSFC (62325211, 62132021, 62322207), Guangdong Natural Science Foundation (2021B1515020085), Shenzhen Science and Technology Program (RCYX20210609103121030), and the Major Program of Xiangjiang Laboratory (23XJ01009)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.26, + 0.324, + 0.275 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.345 + ], + "angle": 0, + "content": "1. Bhatnagar, B.L., Xie, X., Petrov, I.A., Sminchisescu, C., Theobalt, C., Pons-Moll, G.: Behave: Dataset and method for tracking human object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15935-15946 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.346, + 0.785, + 0.387 + ], + "angle": 0, + "content": "2. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.388, + 0.785, + 0.456 + ], + "angle": 0, + "content": "3. Chen, K., Choy, C.B., Savva, M., Chang, A.X., Funkhouser, T., Savarese, S.: Text2shape: Generating shapes from natural language by learning joint embeddings. In: Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part III 14. pp. 100-116. Springer (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.457, + 0.785, + 0.497 + ], + "angle": 0, + "content": "4. Cohen-Bar, D., Richardson, E., Metzer, G., Giryes, R., Cohen-Or, D.: Set-the-scene: Global-local training for generating controllable nef scenes. arXiv preprint arXiv:2303.13450 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.497, + 0.604, + 0.51 + ], + "angle": 0, + "content": "5. Deepfloyd IF. https://github.com/deep-floyd/IF, 2023" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.785, + 0.551 + ], + "angle": 0, + "content": "6. Diller, C., Dai, A.: Cg-hoi: Contact-guided 3d human-object interaction generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19888-19901 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.552, + 0.785, + 0.593 + ], + "angle": 0, + "content": "7. Fan, Z., Taheri, O., Tzionas, D., Kocabas, M., Kaufmann, M., Black, M.J., Hilliges, O.: Articulated objects in free-form hand interaction. arXiv preprint arXiv:2204.13662 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.594, + 0.785, + 0.635 + ], + "angle": 0, + "content": "8. Feng, Y., Choutas, V., Bolkart, T., Tzionas, D., Black, M.J.: Collaborative regression of expressive bodies using moderation. In: 2021 International Conference on 3D Vision (3DV). pp. 792-804. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.635, + 0.785, + 0.675 + ], + "angle": 0, + "content": "9. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5152-5161 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.676, + 0.785, + 0.703 + ], + "angle": 0, + "content": "0. Guo, M., Fathi, A., Wu, J., Funkhouser, T.: Object-centric neural scene rendering. arXiv preprint arXiv:2012.08503 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.704, + 0.785, + 0.758 + ], + "angle": 0, + "content": "11. Guo, Y.C., Liu, Y.T., Shao, R., Laforte, C., Voleti, V., Luo, G., Chen, C.H., Zou, Z.X., Wang, C., Cao, Y.P., Zhang, S.H.: threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "12. Hassan, M., Choutas, V., Tzionas, D., Black, M.J.: Resolving 3d human pose ambiguities with 3d scene constraints. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2282-2292 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Hassan, M., Ghosh, P., Tesch, J., Tzionas, D., Black, M.J.: Populating 3d scenes by learning human-scene interaction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14708-14718 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.289, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Dai et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "14. Hong, F., Zhang, M., Pan, L., Cai, Z., Yang, L., Liu, Z.: Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "15. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence 36(7), 1325-1339 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.785, + 0.288 + ], + "angle": 0, + "content": "16. Jahan, T., Guan, Y., Van Kaick, O.: Semantics-guided latent space exploration for shape generation. In: Computer Graphics Forum. vol. 40, pp. 115-126. Wiley Online Library (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.785, + 0.331 + ], + "angle": 0, + "content": "17. Jain, A., Mildenhall, B., Barron, J.T., Abbeel, P., Poole, B.: Zero-shot text-guided object generation with dream fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 867-876 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.785, + 0.36 + ], + "angle": 0, + "content": "18. Kim, V.G., Chaudhuri, S., Guibas, L., Funkhouser, T.: Shape2Pose: Human-centric shape analysis. ACM Transactions on Graphics (Proc. SIGGRAPH) (Aug 2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.785, + 0.387 + ], + "angle": 0, + "content": "19. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.785, + 0.416 + ], + "angle": 0, + "content": "20. Lee, H.H., Chang, A.X.: Understanding pure clip guidance for voxel grid nerf models. arXiv preprint arXiv:2209.15172 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.418, + 0.785, + 0.459 + ], + "angle": 0, + "content": "21. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.501 + ], + "angle": 0, + "content": "22. Lin, Y., Bai, H., Li, S., Lu, H., Lin, X., Xiong, H., Wang, L.: Componerf: Text-guided multi-object compositional nerf with editable 3d scene layout. arXiv preprint arXiv:2303.13843 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.544 + ], + "angle": 0, + "content": "23. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: European Conference on Computer Vision. pp. 423-439. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.545, + 0.785, + 0.585 + ], + "angle": 0, + "content": "24. Liu, Z., Wang, Y., Qi, X., Fu, C.W.: Towards implicit text-guided 3d shape generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17896-17906 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.785, + 0.628 + ], + "angle": 0, + "content": "25. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Spl: A skinned multi-person linear model. ACM transactions on graphics (TOG) 34(6), 1-16 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.63, + 0.785, + 0.67 + ], + "angle": 0, + "content": "26. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5442-5451 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.785, + 0.712 + ], + "angle": 0, + "content": "27. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latentnerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.785, + 0.755 + ], + "angle": 0, + "content": "28. Michel, O., Bar-On, R., Liu, R., Benaim, S., Hanocka, R.: Text2mesh: Text-driven neural stylization for meshes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13492-13502 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.797 + ], + "angle": 0, + "content": "29. Mihajlovic, M., Saito, S., Bansal, A., Zollhoefer, M., Tang, S.: Coap: Compositional articulated occupancy of people. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13201-13210 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "30. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InterFusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "31. Mirzaei, A., Kant, Y., Kelly, J., Gilitschenski, I.: Laterf: Label and text driven object radiance fields. In: European Conference on Computer Vision. pp. 20-36. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.191, + 0.787, + 0.233 + ], + "angle": 0, + "content": "32. Mohammad Khalid, N., Xie, T., Belilovsky, E., Popa, T.: Clip-mesh: Generating textured meshes from text using pretrained image-text models. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-8 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.787, + 0.274 + ], + "angle": 0, + "content": "33. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.787, + 0.317 + ], + "angle": 0, + "content": "34. Nguyen-Phuoc, T.H., Richardt, C., Mai, L., Yang, Y., Mitra, N.: Blockgan: Learning 3d object-aware scene representations from unlabelled images. Advances in neural information processing systems 33, 6767-6778 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.787, + 0.36 + ], + "angle": 0, + "content": "35. Niemeyer, M., Geiger, A.: Giraffe: Representing scenes as compositional generative neural feature fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11453-11464 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.787, + 0.402 + ], + "angle": 0, + "content": "36. Ost, J., Mannan, F., Thuerey, N., Knodt, J., Heide, F.: Neural scene graphs for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2856-2865 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.787, + 0.459 + ], + "angle": 0, + "content": "37. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10975-10985 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.787, + 0.5 + ], + "angle": 0, + "content": "38. Peng, X., Xie, Y., Wu, Z., Jampani, V., Sun, D., Jiang, H.: Hoi-diff: Text-driven synthesis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.529 + ], + "angle": 0, + "content": "39. Po, R., Wetzstein, G.: Compositional 3d scene generation using locally conditioned diffusion. arXiv preprint arXiv:2303.12218 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.787, + 0.558 + ], + "angle": 0, + "content": "40. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.787, + 0.601 + ], + "angle": 0, + "content": "41. Puig, X., Undersander, E., Szot, A., Cote, M.D., Yang, T.Y., Partsey, R., Desai, R., Clegg, A.W., Hlavac, M., Min, S.Y., et al.: Habitat 3.0: A co-habitat for humans, avatars and robots. arXiv preprint arXiv:2310.13724 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.602, + 0.787, + 0.657 + ], + "angle": 0, + "content": "42. Punnakkal, A.R., Chandrasekaran, A., Athanasiou, N., Quiros-Ramirez, A., Black, M.J.: Babel: Bodies, action and behavior with english labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 722-731 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.787, + 0.713 + ], + "angle": 0, + "content": "43. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.787, + 0.756 + ], + "angle": 0, + "content": "44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.784 + ], + "angle": 0, + "content": "45. Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "46. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.337, + 0.128 + ], + "angle": 0, + "content": "Dai et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "47. Savva, M., Chang, A.X., Hanrahan, P., Fisher, M., Nießner, M.: Pigraphs: learning interaction snapshots from observations. ACM Transactions on Graphics (TOG) 35(4), 1-12 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.191, + 0.785, + 0.218 + ], + "angle": 0, + "content": "48. Sella, E., Fiebelman, G., Hedman, P., Averbuch-Elor, H.: Vox-e: Text-guided voxel editing of 3d objects. arXiv preprint arXiv:2303.12048 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.22, + 0.785, + 0.246 + ], + "angle": 0, + "content": "49. Sheridan, T.B.: Human-robot interaction: status and challenges. Human factors 58(4), 525-532 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.249, + 0.785, + 0.275 + ], + "angle": 0, + "content": "50. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.277, + 0.785, + 0.317 + ], + "angle": 0, + "content": "51. Sohn, K., Lee, H., Yan, X.: Learning structured output representation using deep conditional generative models. Advances in neural information processing systems 28 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.319, + 0.785, + 0.346 + ], + "angle": 0, + "content": "52. Song, Y., Kong, C., Lee, S., Kwak, N., Lee, J.: Towards efficient neural scene graphs by learning consistency fields. arXiv preprint arXiv:2210.04127 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.348, + 0.785, + 0.403 + ], + "angle": 0, + "content": "53. Taheri, O., Ghorbani, N., Black, M.J., Tzionas, D.: Grab: A dataset of whole-body human grasping of objects. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 581-600. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.405, + 0.785, + 0.445 + ], + "angle": 0, + "content": "54. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.5 + ], + "angle": 0, + "content": "55. Von Marcard, T., Henschel, R., Black, M.J., Rosenhahn, B., Pons-Moll, G.: Recovering accurate 3d human pose in the wild using imus and a moving camera. In: Proceedings of the European conference on computer vision (ECCV). pp. 601-617 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.503, + 0.785, + 0.543 + ], + "angle": 0, + "content": "56. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.546, + 0.785, + 0.586 + ], + "angle": 0, + "content": "57. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.588, + 0.785, + 0.628 + ], + "angle": 0, + "content": "58. Wu, Q., Shi, Y., Huang, X., Yu, J., Xu, L., Wang, J.: Thor: Text to human-object interaction diffusion via relation intervention. arXiv preprint arXiv:2403.11208 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.63, + 0.785, + 0.671 + ], + "angle": 0, + "content": "59. Wu, Q., Liu, X., Chen, Y., Li, K., Zheng, C., Cai, J., Zheng, J.: Object-compositional neural implicit surfaces. In: European Conference on Computer Vision. pp. 197-213. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.673, + 0.785, + 0.74 + ], + "angle": 0, + "content": "60. Xu, Y., Chai, M., Shi, Z., Peng, S., Skorokhodov, I., Siarohin, A., Yang, C., Shen, Y., Lee, H.Y., Zhou, B., et al.: Discoscope: Spatially disentangled generative radiance fields for controllable 3d-aware scene synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4402-4412 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.798 + ], + "angle": 0, + "content": "61. Yang, B., Zhang, Y., Xu, Y., Li, Y., Zhou, H., Bao, H., Zhang, G., Cui, Z.: Learning object-compositional neural radiance field for editable scene rendering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13779-13788 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "62. Zhang, S., Zhang, Y., Ma, Q., Black, M.J., Tang, S.: Place: Proximity learning of articulation and contact in 3d environments. In: 2020 International Conference on 3D Vision (3DV). pp. 642-651. IEEE (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.653, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InterFusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "63. Zhang, Y., Hassan, M., Neumann, H., Black, M.J., Tang, S.: Generating 3d people in scenes without people. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6194-6204 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "64. Zhao, K., Wang, S., Zhang, Y., Beeler, T., Tang, S.: Compositional human-scene interaction synthesis with semantic control. In: European Conference on Computer Vision. pp. 311-327. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.788, + 0.273 + ], + "angle": 0, + "content": "65. Zhi, S., Laidlow, T., Leutenegger, S., Davison, A.J.: In-place scene labelling and understanding with implicit scene representation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15838-15847 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.273 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/InterFusion_ Text-Driven Generation of 3D Human-Object Interaction/d1b8e0cb-6c9e-44c9-80dd-158238285207_origin.pdf b/2024/InterFusion_ Text-Driven Generation of 3D Human-Object Interaction/d1b8e0cb-6c9e-44c9-80dd-158238285207_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b946b007138621c07b847ed123fffef617429137 --- /dev/null +++ b/2024/InterFusion_ Text-Driven Generation of 3D Human-Object Interaction/d1b8e0cb-6c9e-44c9-80dd-158238285207_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05a5166dfb9a825280a57e340cdaec3b2a63d901b6398f7850e9e8aedf95f449 +size 25456040 diff --git a/2024/InterFusion_ Text-Driven Generation of 3D Human-Object Interaction/full.md b/2024/InterFusion_ Text-Driven Generation of 3D Human-Object Interaction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..29e44e2550251f8122c571ec768b65a857795de4 --- /dev/null +++ b/2024/InterFusion_ Text-Driven Generation of 3D Human-Object Interaction/full.md @@ -0,0 +1,311 @@ +# InterFusion: Text-Driven Generation of 3D Human-Object Interaction + +Sisi Dai $^{1}$ Wenhao Li $^{1}$ Haowen Sun $^{2}$ Haibin Huang $^{3}$ Chongyang Ma $^{3}$ + +Hui Huang² Kai Xu1* Ruizhen Hu2* + +$^{1}$ National University of Defense Technology + +2 Shenzhen University + +3 Kuaishou Technology + +https://sisidai.github.io/InterFusion/ + +![](images/50527545e07cf197e0ea4b8ec32da26788a54af9738ffc8ea344664b715fc370.jpg) +a person in a military uniform sitting on a chair + +![](images/17c5444b41efb96aeca0e190d47d2f4f8ad4a5923438856ad2850b3acf4e5bf0.jpg) +a female gymnastics athlete riding a motorcycle + +![](images/a2e56b4cc995146aa1aa82ecfa544cee682f2d8f82ad89f82701320f6f7d20fc.jpg) +a person in a fleece pullover and hiking pants carrying a backpack + +![](images/f195b07182d8df6deab02e6d56a0f0f2cb08cc8c01932f8366d5baa3597109c4.jpg) +a man in a quilted gilet and moleskin trousers playing the guitar + +![](images/92890147ea2c39a7a17782e400d6fe8d58673aa4c49fcdd0b53c740c8d90279f.jpg) +a man wearing a WORLD lettered T-shirt riding a bike + +![](images/aa89e5e1a2d9857f9c1d6945dcb75bc3475d0d357ba44dd64349d0dc978bdd56.jpg) +a person in a cropped hoodie and bike shorts jumping a skateboard + +![](images/3024fe4eb00435fcd76f639eb90deadeddb99818b0799aba7148089b8aec6496.jpg) +a man in a velvet smoking jacket and silk pajama pants playing the saxophone + +![](images/88157f0a186ebce2852d8e61b2d4a37c57c6b6dd5834d1407aed4f5d88f50332.jpg) +a man in a paisley waistcoat and bowler hat riding a horse + +![](images/fcdaaf3478512b2e9be0025e1da06574883663a508d0a923d8a75af0d79ccf7d.jpg) +a man with blond hair wearing a brown leather jacket reading a laptop +Fig. 1: Given a text prompt, our method InterFusion can generate diverse 3D scenes of a person interacting with an object. + +Abstract. In this study, we tackle the complex task of generating 3D human-object interactions (HOI) from textual descriptions in a zero-shot text-to-3D manner. We identify and address two key challenges: the unsatisfactory outcomes of direct text-to-3D methods in HOI, largely due to the lack of paired text-interaction data, and the inherent difficulties in simultaneously generating multiple concepts with complex spatial relationships. To effectively address these issues, we present InterFusion, + +a two-stage framework specifically designed for HOI generation. InterFusion involves human pose estimations derived from text as geometric priors, which simplifies the text-to-3D conversion process and introduces additional constraints for accurate object generation. At the first stage, InterFusion extracts 3D human poses from a synthesized image dataset depicting a wide range of interactions, subsequently mapping these poses to interaction descriptions. The second stage of InterFusion capitalizes on the latest developments in text-to-3D generation, enabling the production of realistic and high-quality 3D HOI scenes. This is achieved through a local-global optimization process, where the generation of human body and object is optimized separately, and jointly refined with a global optimization of the entire scene, ensuring a seamless and contextually coherent integration. Our experimental results affirm that InterFusion significantly outperforms existing state-of-the-art methods in 3D HOI generation. + +Keywords: Text-Driven Generation $\cdot$ Zero-Shot Generation $\cdot$ 3D Human-Object Interaction Generation + +# 1 Introduction + +The generation of 3D human-object interactions (HOI) stands as a critical challenge in the fields of computer vision and computer graphics, with far-reaching implications in virtual reality, augmented reality, animation, and embodied AI [41, 49]. This task entails the creation of realistic 3D scenes where human figures interact with objects in ways that are not only physically plausible but also contextually relevant. + +Despite its potential, the field has faced significant obstacles, primarily due to the scarcity of large-scale interaction data. Traditional approaches have predominantly relied on motion capture (mocap) datasets or physics-based simulations for generating these interactions. Mocap datasets [15,26,55] are limited by the specific scenarios they capture and are both costly and labor-intensive to produce. These limitations have resulted in a notable gap in generating diverse and contextually rich HOI scenes, especially for novel or complex interactions. Conversely, recent advancements have introduced text-to-3D methods [40,56], marking a significant shift in the field. These methods harness the power of textual descriptions to generate 3D objects without direct 3D supervision, presenting a novel approach to 3D content creation. + +In this study, we explore text-to-3D method in HOI task within a zero-shot manner, i.e., generating 3D scenes from textual descriptions using 2D diffusion models. Our key observations are two-fold: first of all, a direct application of text-to-3D method in HOI often leads to unsatisfied results like blurry textures and incorrect interactions, which is caused by the relative lack of paired text-interaction data during training and the difficulties of generating multiple concepts for diffusion-based methods [23]. Secondly, while collecting extensive interaction data is challenging, estimating human poses based on described in + +teractions is more feasible. These pose estimations can serve as geometric priors in the HOI generation process. By integrating pose estimation, our method substantially simplifies the text-to-3D process, particularly in the geometry optimization stage for human body generation. It also provides additional constraints for object generation, ensuring accurate placement and alignment with the text's semantic content. More importantly, we can separate the generation of human body and object, and jointly refine the details with the estimated pose, allowing for a more coherent and detailed synthesis of the interaction scene. + +Inspired by these observations, we introduce InterFusion, a novel two-stage framework designed for HOI generation. Specifically, in the first stage, rather than relying on precise 3D interaction data, our approach instead collects a comprehensive dataset of synthesized images that depict a wide range of interactions. From these images, we employ advanced 3D pose estimation technique [8] to extract 3D human poses. Building upon these image-pose pairs, InterFusion develops a sophisticated codebook that establishes a mapping between interaction descriptions and 3D human poses, with the integration of the CLIP (Contrastive Language-Image Pretraining) embedding [43]. By leveraging these embeddings, our framework is able to interpret the nuances of interaction descriptions and translate them into accurate 3D pose representations. In the subsequent stage, InterFusion capitalizes on recent advancements in text-to-3D generation [40], as well as neural radiance fields [30], using the estimated human poses to produce 3D HOI scenes with realistic appearances and high-quality geometry. This stage operates in a 'local-global' manner. At the local level, the generation of the human body (SDS-H) and objects (SDS-O) is separately optimized, with the poses serving as additional constraints for the SDS. At the global level, the generation of the entire scene (SDS-I) is also guided by the integrated description and jointly optimized with SDS-H and SDS-O, ensuring a cohesive and contextually accurate representation of the HOI. Our experiments show that the quality of generation can be improved by a large margin and our approach outperforms state-of-the-art methods in HOI generation. + +To summarize, our contributions are as follows: + +- We introduce a novel two-stage framework InterFusion, for zero-shot 3D human-object interaction generation from text, incorporating 3D pose estimation as geometry priors. +- InterFusion leverages text-to-3D generation with a local-global optimization process. This strategy ensures seamless integration of human bodies and objects, producing realistic and high-quality 3D HOI scenes. +- InterFusion demonstrates significant improvements over existing methods in 3D HOI generation, showcasing its effectiveness in creating detailed, and contextually rich 3D interactions. + +# 2 Related Work + +# 2.1 Human-Object Interaction Synthesis + +3D human-object interaction (HOI) generation is a challenging problem that has been studied widely by the computer vision and graphics community. Shape2Pose [18] generates plausible 3D human poses interacting with a given 3D object model by learning an affordance model from synthetic data. PiGraphs [47] learns the distribution between human poses and object arrangements from a collected dataset with 3D scene scans and RGB-D videos, generating the interaction snapshots given action specifications and object models. Recently, the parametric human body models such as SMPL [25,37,45] are employed in interaction synthesis to overcome the lack of realism due to human body representations. Benefiting from the PROX dataset [12] which consists of fitted SMPL models in captured 3D scenes, the more specific human-scene interaction has become an active research direction. Given a 3D scene, PSI [63] and PLACE [62] generate the 3D human body mesh represented as SMPL parameters through a conditional variational autoencoder [19, 51]. POSA [13] learns pose-specific priors to generate contacts conditioned by the given posed human, which can further guide the placement of the body mesh in a scene. COINS [64] enables semantic control on interaction synthesis by embedding the action label together with the interacted object as the condition of the generative model. More recently, fine-grained 3D interaction datasets [1, 7, 53] are captured to promote this field. Relying on these datasets, some methods [6, 38, 58], concurrent to this work, are proposed to explore text-guided 3D HOI generation. However, they remain constrained by distributions within the datasets. While previous methods need ground truth 3D interaction data as supervision, our work, for the first time, attempts to break through the limitation of data requirement. We generate a wider range of realistic and detailed 3D HOI scenes, including both indoors and outdoors. + +# 2.2 Text-to-3D Content Synthesis + +Early methods [3, 16, 24] for text-to-3D shape generation require paired data of 3D data and the corresponding textual descriptions to learn the joint embedding space of shape and text for supervision, which limits their generality to unseen object categories. Benefiting from large pre-trained text-to-image models and differentiable rendering techniques, breakthroughs in text-to-3D content generation have been achieved. For example, DreamFields [17] and PureCLIPN-eRF [20] combine CLIP [43] with neural radiance fields (NeRF) [30], demonstrating the potential for zero-shot NeRF optimization. Meanwhile, CLIP-mesh [32] and Text2Mesh [28] incorporate CLIP to optimize the 3D mesh representation, starting from an initial sphere mesh and an input base mesh, respectively. Recently, DreamFusion [40] and SJC [56] enable NeRF optimization with guidance from pre-trained text-to-image diffusion models [44,46] in place of CLIP, achieving more impressive results. To improve DreamFusion, Magic3D [21] proposes + +a coarse-to-fine pipeline to generate the fine-grained mesh. TextMesh [54] extends the geometry representation from NeRF to an SDF framework, thereby enhancing detailed mesh extraction and photorealistic rendering. Among follow-up works, Latent-NeRF [27] and Vox-E [48] utilize explicit 3D shapes to provide additional training signals for NeRF optimization. While Latent-NeRF utilizes a rough untextured object for shape sculpture, Vox-E takes multiview images of a fine-grained textured object to edit geometry and appearance. All the methods mentioned above focus on the generation or edition of a single subject, while ignoring the interaction between different subjects. + +# 2.3 Compositional Scene Generation + +Representing scenes as compositions of object representations facilitates enhanced controllability. Numerous techniques incorporate additional information at the object level to perform compositional modeling of scenes, effectively separating object representations from the overall scene imagery. For instance, some methods incorporate 2D semantic information such as segmentation labels [65], instance masks [59,61], or features from a pre-trained vision-language model [31]. Some other methods [10, 36, 52] use 3D layout information by object-centric bounding boxes with canonical coordinates. When it comes to scene generation, [34, 35, 60] use compositional representations to generate scenes in a controllable manner. More recent approaches [4, 22, 39] generate compositional 3D scenes from input text prompts. Different from existing methods, our HOI scene is generated automatically without the requirement of input layout. Moreover, the spatial relationship between human and object during interaction is much more complex and cannot be simply characterized using their bounding boxes. + +# 3 Our Method + +# 3.1 Overview + +In this section, we formally introduce InterFusion with a focus on the zero-shot and text-driven generation of 3D human-object interactions (HOI). Specifically, the input is a triplet of text descriptions, $T = \{T^H, T^O, T^I\}$ , specifying the desired human style, object style, and interaction type. The goal is to generate a detailed 3D scene, $\psi = \{\psi^H, \psi^O\}$ , comprising a human model and an object model that not only adhere to the specified appearance styles but also exhibit a tailored spatial relationship to accurately reflect the described interaction. + +As illustrated in Figure 2, our method consists of two primary stages: anchor pose generation and anchor pose guided HOI generation. We first generate an interaction pose based on the input text, termed an anchored pose. This pose then serves as a geometric constraint, guiding the subsequent generation of detailed HOI. In the second phase, the human and object models are optimized separately and refined simultaneously with a global context, ensuring a cohesive and accurate representation of the interaction as described by the input text. + +![](images/53294ededa3526b22fbae85de4df0dcb00b770483fc2042d3e9e7a401d299b72.jpg) +Fig. 2: InterFusion is a two-stage framework that transforms textual descriptions into detailed 3D human-object interactions, initially synthesizing anchor poses (upper left) and then optimizing the human model (upper right) and object model (bottom) with constraints from estimated pose and textual prompts. + +# 3.2 Preliminaries + +SDS. Score Distillation Sampling (SDS) has been introduced by DreamFusion [40]. While $x = g(\psi, \zeta)$ , $x$ is the 2D image rendered by a differentiable renderer $g$ with model parameters $\psi$ (the volumetric renderer and MLPs correspondingly in NeRF), under a desired camera pose $\zeta$ . By injecting the sampled noise $\epsilon$ into $x$ at a time step $t$ , the noisy image $x_{t}$ is produced. The pre-trained 2D text-to-image diffusion model $\phi$ provides a denoising network $\hat{\epsilon}_{\phi}(x_{t};y,t)$ that predicts the noise $\hat{\epsilon}$ given the noisy image $x_{t}$ , time step $t$ , and text embedding $y$ . SDS then optimizes the model parameters $\psi$ by minimizing the difference between the predicted noise and the added noise: + +$$ +\nabla_ {\psi} \mathcal {L} _ {\mathrm {S D S}} (\phi , x) = \mathbb {E} _ {t, \epsilon} [ w (t) (\hat {\epsilon} _ {\phi} (x _ {t}; y, t) - \epsilon) \frac {\partial x}{\partial \psi} ], \tag {1} +$$ + +where $w(t)$ is the weighting term at the time step $t$ . + +# 3.3 Anchor Pose Generation + +InterFusion starts by generating the anchor pose from the input text $T^I$ , which is a non-trivial problem and highly limited by the available pose datasets. Existing datasets (e.g., HumanML3D [9], BABEL [42]) are mocap based datasets and are typically conducted in a controlled laboratory environment, leading to lacks in action diversity and spatial coverage. + +Recently, text-to-image diffusion models have shown the versatility and effectiveness of generating realistic and diverse images, through the integration of visual and linguistic feature spaces. We thus instead utilize advancements in this technology, specifically models like Stable Diffusion, to create a large-scale dataset of Human-Object Interaction (HOI) images. These images depict a wide range of human interactions. To ensure the diversity of interaction types, we utilize ChatGPT to generate prompts about human daily events or actions forming "verb-ing a/an/the object". A total of 235 result prompts are generated, covering most interactions in daily life. After filtering synthesized images without humans, we then estimate 3D human poses using the pre-trained PixIE [8] model, creating a comprehensive Syn-HOI pose dataset consisting of a total 55K 3D pseudo-SMPL poses. + +While SMPL model maps pose and shape parameters to a triangulated mesh, to align these poses with our text-to-interaction generation task, we render images from multiple perspectives and use CLIP's image encoder to derive pose feature embeddings. The average feature across these multiple perspective images represents the pose feature and attaches the dataset with pairs of averaged CLIP embedding and pose parameters. We further construct a codebook using K-Means clustering to identify key pose centroids based on the feature embeddings. 2,048 cluster centroids are clustered compositing our pose code-book, where each cluster comprises a subset of poses that represents similar interaction as the key poses of the centroids. + +Once the codebook is built, for a given input text, we extract its feature embedding using CLIP's text encoder. This serves as a query to retrieve the most similar poses from the codebook, based on feature similarity. Specifically, given query text $T^{I}$ , top $k$ poses $\theta_{k}^{T^{I}}$ could be matched by pose embeddings $\theta_{E}$ , the averaged CLIP embeddings among rendered images from key poses: + +$$ +\theta_ {k} ^ {T ^ {I}} = \operatorname {T O P} _ {k} \left(f _ {\text {t e x t}} \left(T ^ {I}\right), \theta_ {E}\right) \tag {2} +$$ + +where $f_{text}$ is the text encoder of CLIP and $\mathrm{TOP}_k(X,Y)$ returns top $k$ poses with the top $k$ highest cosine similarities, and we use $k = 7$ for suitable poses in our experiments. + +We then utilize GPT-4V to select the most precise pose as the final queried key pose. Depending on the requirement, we can instead select the poses sampled in the cluster, corresponding to the key pose, to guide the generation process for diverse results. This approach ensures rich and contextually aligned anchor poses for our text-to-interaction generation task. + +To restrict the optimization of geometry and appearance using human structural priors from the acquired pose, we further incorporate COAP [29], a neural occupancy representation of the articulated human body based on SMPL parameters [25]. Given a pose $\theta$ and a shape $\beta$ , COAP offers an occupancy prediction network $f(x; \beta, \theta)$ that maps a 3D query point $p$ to an occupancy value, directly indicating whether the spatial point resides within the 3D body. + +# 3.4 Pose-Guided HOI Generation + +Once get the interaction pose, we further use it along with the input text to guide the generation of a detailed 3D HOI scene. In this phase, the estimated pose serves both as spatial constraints for the scene's geometry and as an anchor that aligns the human and object models. This approach ensures that the human model and the object model can be generated separately, while they are cohesively aligned to form the final 3D HOI scene. + +The anchor pose provides specific spatial constraints for each component in the scene. For the human model, it establishes a basic geometric structure, while for the object model, it defines the areas that should remain unoccupied. This clear distinction is essential for rendering the scene both accurately and realistically. Simultaneously, the input text undergoes a complex processing procedure to offer distinct semantic guidance for each component. This is accomplished through various text conditioning techniques applied in SDS with the pre-trained DeepFloyd model [5]. The text is carefully crafted to direct the generation of both the human and object models to ensure all elements are in harmony with the semantic nuances of input descriptions. + +Additionally, we introduce a novel camera tracing module to enhance the optimization process under varying text conditions. This module adaptively adjusts the camera pose, focusing on relevant elements within the scene at each optimization stage. This adaptive camera positioning is instrumental in ensuring that each aspect of the scene is optimally rendered according to the text descriptions, resulting in a more dynamic and contextually accurate 3D HOI scene. + +We now provide details of each module in this stage, including the Neural Radiance Field representations for the human model (H-NeRF) and object model (O-NeRF), the camera tracing mechanism, and the guided optimization process. + +H-NeRF. We use NeRF to represent the human model, noted as H-NeRF. The generation of H-NeRF is guided by the text description specifying the human style, conducting pose-specific human avatar generation with SDS. To enhance the quality of our renderings, particularly in terms of resolution, we incorporated a specialized optimization process that focuses on the head region of the human avatar. The location of the head for any given pose is determined with COAP [29]. This allows us to augment the text prompt specifically for the head region, using the notation * the head of *, to ensure that the head receives detailed attention during the generation process. The loss function for this optimization process is as follows, designed to balance the fidelity of the head region with the overall pose and style of the human figure. + +$$ +\begin{array}{l} \nabla_ {\psi^ {H}} \mathcal {L} _ {\mathrm {S D S}} = \mathbb {E} _ {t, \epsilon} [ w (t) (\hat {\epsilon} _ {\phi} \left(x _ {t} ^ {H}; y ^ {H}, t\right) - \epsilon) \frac {\partial x ^ {H}}{\partial \psi^ {H}} ] \tag {3} \\ + \mathbb {E} _ {t, \epsilon} [ w (t) (\hat {\epsilon} _ {\phi} (x _ {t} ^ {H, h}; y ^ {H, h}, t) - \epsilon) \frac {\partial x ^ {H , h}}{\partial \psi^ {H}} ], \\ \end{array} +$$ + +In shaping H-NeRF, our geometric constraint ensures that the human model evolves directly from the anchor pose. Points within the anchor are required to be occupied, establishing a firm base for the model. Meanwhile, points outside the anchor can also be occupied, but with probabilities that decrease as they move away from the anchor's surface. This approach allows for the addition of geometric details over the anchor, ensuring the model aligns with the human style described in the text while maintaining a coherent structure rooted in the anchor pose. The loss function used is as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {g e o}} ^ {H} = C E _ {p _ {i} \in \mathbb {P} _ {\mathrm {i n}}} \left(\alpha_ {i}, f (p _ {i})\right) \\ + C E _ {p _ {j} \in \mathbb {P} _ {\text {o u t}}} \left(\alpha_ {j}, f (p _ {j})\right) \left(1 - e ^ {- \frac {d}{2 \eta^ {2}}}\right), \tag {4} \\ \end{array} +$$ + +where $d$ represents the point distance from the anchor surface, and $\eta$ is the hyperparameter to control the extent of decaying, similar as in [27]. + +O-NeRF. We also employ NeRF to model the object component, referred to as O-NeRF. This model is designed to interact seamlessly with the human model and to embody the desired style as specified by the input text. To guide the generation of both the interaction type and the object style, we utilize SDS with tailored text prompts: + +$$ +\begin{array}{l} \nabla_ {\psi^ {O}} \mathcal {L} _ {\mathrm {S D S}} = \mathbb {E} _ {t, \epsilon} [ w (t) (\hat {\epsilon} _ {\phi} \left(x _ {t} ^ {I}; y ^ {I}, t\right) - \epsilon) \frac {\partial x ^ {I}}{\partial \psi^ {O}} ] \tag {5} \\ + \mathbb {E} _ {t, \epsilon} [ w (t) (\hat {\epsilon} _ {\phi} (x _ {t} ^ {O}; y ^ {O}, t) - \epsilon) \frac {\partial x ^ {O}}{\partial \psi^ {O}} ], \\ \end{array} +$$ + +For the interaction scene $x^{I}$ generation, we use an alpha-composed rendering to integrate H-NeRF and O-NeRF. This method calculates an alpha value from the density at each point, determining its contribution to the scene's color. Higher alpha values indicate a greater influence on the rendering, allowing for a nuanced and realistic integration of the human and object models in the final interaction image: + +$$ +x ^ {I} = \sum_ {i} w _ {i} ^ {I} c _ {i} ^ {I}, w _ {i} ^ {I} = \alpha_ {i} ^ {I} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j} ^ {I}), \tag {6} +$$ + +$$ +c _ {i} ^ {I} = \frac {\alpha_ {i} ^ {H}}{\alpha_ {i} ^ {H} + \alpha_ {i} ^ {O}} c _ {i} ^ {H} + \frac {\alpha_ {i} ^ {O}}{\alpha_ {i} ^ {H} + \alpha_ {i} ^ {O}} c _ {i} ^ {I}. +$$ + +In our NeRF-based approach, alpha values, capped at 1, are derived from density for composite rendering. This setup allows semantic guidance gradients, conditioned by the interaction image, to optimize the density and color of both the human and object models. However, we noted a tendency for the model to prioritize object generation at the expense of the human model's quality. To counteract this, we implemented gradient truncation towards the human model to maintain a balanced optimization between the two components. + +For O-NeRF's geometric constraints, we aim to prevent occupancy of points within the anchor pose, denoted as $x_{i} \in \mathbb{X}_{in}$ , by the object model. Inspired by the physical collision prevention concept, we define a specific loss function that ensures these points remain unoccupied by O-NeRF, effectively preventing overlap between the human and object models in the 3D space: + +$$ +\mathcal {L} _ {\mathrm {g e o}} ^ {O} = C E _ {p _ {i} \in \mathbb {P} _ {\mathrm {i n}}} \left(\alpha_ {i}, 1 - f \left(p _ {i}\right)\right) \tag {7} +$$ + +By minimizing the above loss term, model parameters are optimized to enhance the geometric consistency between H-NeRF and O-NeRF. + +Camera tracing. To enhance the generation of O-NeRF, we introduce a dynamic camera tracing module within the SDS framework. This module automatically adjusts the camera pose to focus on the target's center, either the entire interaction scene or just the object during object generation. The camera aligns with the average position of voxels with an occupancy probability above 0.5, ensuring a consistent focus on the most significant parts of the scene or object for optimal detail capture and realism. + +Optimization. The total loss is formulated as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {S D S}} ^ {H} + \lambda_ {1} \mathcal {L} _ {\mathrm {S D S}} ^ {O} + \mathcal {L} _ {\mathrm {g e o}} ^ {H} + \lambda_ {2} \mathcal {L} _ {\mathrm {g e o}} ^ {O} + \lambda_ {3} L _ {\text {r e g}}, \tag {8} +$$ + +where $\lambda_1, \lambda_2$ and $\lambda_3$ are the corresponding loss weights. Figure 2 (left bottom) shows several intermediate states of the object model during the generation steps and we can see that the object is gradually generated with fine details. Weight annealing is adopted during the guided optimization process. We leave the details in the supplement. + +# 4 Experiments + +Figure 3 (left) shows results with various interaction poses, showcasing the strength of our method. Our method can support diverse interaction poses within a single type. Meanwhile, with interaction type and pose fixed, our method can further generate more numerous results under different human styles or object styles, as shown in Figure 3 (right). Moreover, our InterFusion also supports controllable text-conditioned editing, providing users more control over the already generated 3D models, which are presented in the supplementary materials. + +We both qualitatively and quantitatively evaluate our method against alternative baseline methods. To verify the effectiveness of individual components in our method, we conduct ablation studies. Furthermore, we discuss the application potential, limitations and future work of our InterFusion. The implementation details, more evaluations, and further discussions are presented in the supplementary materials. + +![](images/2d8c7dc580e85d742db240cc65d5e6505a8946110158bada2f459e393530a939.jpg) +Fig. 3: More results generated by InterFusion. Diverse integration poses are supported. Numerous human and object styles are also supported. + +Evaluation prompts and criteria. Similar as the first stage for generating prompts about various interaction types, we also use ChatGPT to randomly generate prompts about human styles and collect the text prompts for evaluation. We select 61 distinct and diverse prompts, ensuring the classes of Human (e.g., a policewoman, a teenager) with various styles (e.g., in a graphic tee and ripped jeans), Objects (e.g., a motorcycle, a guitar) and Interactions (e.g., riding, playing) are reasonable and evenly distributed. In total, we have 13 different types of interactions, covering contact areas across the whole body, which demonstrates the effectiveness of our method on diverse types of interactions compared to baseline methods. We calculate the CLIP scores between the input text prompts and different views of generated 3D human-object interactions, and then compare the means based on different evaluation prompts. The CLIP score measures the similarity between a prompt text for an image and the actual content of the image. We also provide an assessment using GPT-4V for selection, named GPT-4V select, which evaluates the completeness of objects and correctness of physical interactions across multiview rendered images. We refer to the supplementary material for assessment details and additional evaluation criteria. + +Baseline approaches. To the best of our knowledge, the proposed method is among the first to generate 3D human-object interactions based on text inputs in a zero-shot manner. We thus compare our method to alternative text-to-3D methods, including DreamFusion [40], Magic3D [21], and TextMesh [54]. Additional comparisons with MVDream [50] and ProlificDreamer [57] are presented in the supplementary material. To justify our key idea of using human pose to guide the generation, we design an object-centric baseline (Ours-OC) as a variant of our method. Instead of using human pose as the prior, we retrieve an object with the given semantic category from ShapeNet dataset [2] and use it as + +Table 1: Quantitative evaluation of CLIP score and GPT-4V select. + +
MethodDreamFusion [40] Magic3D [21] TextMesh [54] Ours-OCOurs-HC
CLIP score0.3027 0.3179 0.2761 0.32030.3308
GPT-4V select(%)8.20 11.48 1.64 13.1165.57
+ +geometry constraints to guide both human and object generation. As official implementations are unavailable for some of these baselines, we use the third-party re-implementations provided by threestudio [11] for a fair comparison. Note that all re-implementations use multi-resolution hash-grid [33] for 3D representation and DeepFloyd [5] for guidance. + +# 4.1 Comparison Results + +Qualitative evaluations. Some representative visual comparisons are shown in Figure 4, and additional comparisons including those with our object-centric baseline are presented in the supplementary material. The overall results of the baselines reveal common uncertainties in both geometry and appearance attributed to the confusion introduced by multi-concept guidance. Specifically, baseline models may lean towards one specific target, as seen in the example of "a man wearing a red baseball cap" with only the red baseball cap generated, the example of "a policeman" with only the upper half of the human body generated, or the example of "a man in a puffer jacket and wool beanie" where the shopping cart is failed to be generated. Even with a relatively uniform attention distribution, the baseline model may encounter challenges in producing complete results, as demonstrated in the saxophone and violin examples (the 3rd and 4th column in Figure 4), which exhibit deficiencies such as incomplete human body parts, and mixing of human legs. Moreover, the baseline model sometimes struggles to effectively choose the focus of generation when the input text describes a less common target like "a person in a military uniform", resulting in poor outcomes. + +Our approach overcomes these issues by conducting optimization in an explicit decomposed way and intelligently guiding attention from SDS jointly in spatial and semantic aspects. Therefore, our method achieves more stable and higher-quality 3D results under multiple-concept guidance. + +Quantitative evaluations. The comparison results are shown in Table 1. We can see that our method (Ours-HC) achieves the best performance compared to baselines, showcasing our results exhibit more 3D plausibility with given text prompts. Among the baselines, Ours-OC gets the best performance. Though with object priors, the object-centric approach still does not provide sufficient body priors for human generation, resulting in its inability to achieve complete interaction generation. Magic3D gets the next best performance among the baselines, as it usually generates the whole interaction scene comparing to other two + +![](images/bb4bcc47b4835dbb2190e3855df9e24f2aa58e104660583f8d42fc662c834f1b.jpg) +Fig. 4: Qualitative comparison results with baselines. InterFusion generates more stable and higher-quality results and is more consistent with input interaction descriptions. + +![](images/d09f2531a0a0aa6b4ea91ea0a7822a369b44105fbe3ec9892cb227a36719aa6e.jpg) +Fig. 5: Qualitative results of ablation studies. + +basielines, however, the results are somewhat vague as shown in Figure 4, thus there is still a large performance gap comparing to our method. + +The significant enhancement in performance can be attributed to our method's effective concurrent generation of both human and object, along with cohesive interactive information. + +# 4.2 Ablation Studies + +We conduct several ablation studies both qualitatively and quantitatively to show the importance of our design as well as introduced loss terms. Results are presented in Figure 5a, Figure 5b and Table 2. We refer to the supplementary material for additional results and details. + +Table 2: Quantitative results of ablation studies. + +
Settingsw/o ISDSw/o L0SDSw/o L0geoOurs
CLIP score0.31640.32930.31710.3308
GPT-4V select(%)1.644.9216.3977.05
+ +Syn-HOI-pose. The anchor pose obtained at the first stage provides the geometry constraints for our HOI generation. We thus first demonstrate the effectiveness of synthesizing anchoring poses using our pseudo-pose dataset. We compare the pose generated with the existing approach AvatarCLIP [14], which utilizes CLIP only to retrieve the queried pose from the codebook constructed from the mocap dataset AMASS [26]. Figure 5a shows that the proposed pose generation stage enables a better-fitted anchor pose for input interaction types, which demonstrates our pseudo-poses, reconstructed from synthesized images, has the potential for more diverse poses' requirements than existing mocap datasets. + +Semantic guidance. To better illustrate the efficacy of semantic guidance in our NeRF-based generation, we conduct ablations separately on SDS from object and interaction. As shown in Figure 5b and Table 2, in the absence of SDS from object, semantic consistency for the object is compromised, resulting in noisy outcomes influenced by the human, thus the contact region can not be extracted correctly. SDS from interaction plays the most crucial role in the performance, making results more semantically and visually plausible. The absence of SDS from interaction leads to generated objects being confined to the remaining space, but without interaction with the human. + +Geometric constraint. Relying solely on semantic guidance is inadequate for achieving the final objectives, as the object should be generated in spaces outside the human body, with its generation targets not positioned at the origin. As illustrated in Figure 5b and Table 2, depending only on semantic guidance may yield results that appear spatially conflict with the human body. Without the spatial constraint, the object generation at the origin would additionally have a conflict with the interaction objective, thus resulting in degenerated objects and final interactions. + +# 5 Conclusion + +In conclusion, our work presents InterFusion, a novel framework for zero-shot 3D human-object interaction generation. InterFusion tackles the challenges of limited 3D interaction data and the complexity of generating multiple concepts simultaneously. Our two-stage approach, which synthesizes 3D interaction poses from text and then uses these poses as geometric anchors for detailed HOI scene generation, has demonstrated significant improvements over existing methods. + +Acknowledgements. We thank Zejia Su and Yunfan Ye for helpful discussions and paper proofreading. We also thank the anonymous reviewers for their valuable comments. This project was supported by the NSFC (62325211, 62132021, 62322207), Guangdong Natural Science Foundation (2021B1515020085), Shenzhen Science and Technology Program (RCYX20210609103121030), and the Major Program of Xiangjiang Laboratory (23XJ01009). + +# References + +1. Bhatnagar, B.L., Xie, X., Petrov, I.A., Sminchisescu, C., Theobalt, C., Pons-Moll, G.: Behave: Dataset and method for tracking human object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15935-15946 (2022) +2. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015) +3. Chen, K., Choy, C.B., Savva, M., Chang, A.X., Funkhouser, T., Savarese, S.: Text2shape: Generating shapes from natural language by learning joint embeddings. In: Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part III 14. pp. 100-116. Springer (2019) +4. Cohen-Bar, D., Richardson, E., Metzer, G., Giryes, R., Cohen-Or, D.: Set-the-scene: Global-local training for generating controllable nef scenes. arXiv preprint arXiv:2303.13450 (2023) +5. Deepfloyd IF. https://github.com/deep-floyd/IF, 2023 +6. Diller, C., Dai, A.: Cg-hoi: Contact-guided 3d human-object interaction generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19888-19901 (2024) +7. Fan, Z., Taheri, O., Tzionas, D., Kocabas, M., Kaufmann, M., Black, M.J., Hilliges, O.: Articulated objects in free-form hand interaction. arXiv preprint arXiv:2204.13662 (2022) +8. Feng, Y., Choutas, V., Bolkart, T., Tzionas, D., Black, M.J.: Collaborative regression of expressive bodies using moderation. In: 2021 International Conference on 3D Vision (3DV). pp. 792-804. IEEE (2021) +9. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5152-5161 (2022) +0. Guo, M., Fathi, A., Wu, J., Funkhouser, T.: Object-centric neural scene rendering. arXiv preprint arXiv:2012.08503 (2020) +11. Guo, Y.C., Liu, Y.T., Shao, R., Laforte, C., Voleti, V., Luo, G., Chen, C.H., Zou, Z.X., Wang, C., Cao, Y.P., Zhang, S.H.: threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio (2023) +12. Hassan, M., Choutas, V., Tzionas, D., Black, M.J.: Resolving 3d human pose ambiguities with 3d scene constraints. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2282-2292 (2019) +13. Hassan, M., Ghosh, P., Tesch, J., Tzionas, D., Black, M.J.: Populating 3d scenes by learning human-scene interaction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14708-14718 (2021) + +14. Hong, F., Zhang, M., Pan, L., Cai, Z., Yang, L., Liu, Z.: Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535 (2022) +15. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence 36(7), 1325-1339 (2013) +16. Jahan, T., Guan, Y., Van Kaick, O.: Semantics-guided latent space exploration for shape generation. In: Computer Graphics Forum. vol. 40, pp. 115-126. Wiley Online Library (2021) +17. Jain, A., Mildenhall, B., Barron, J.T., Abbeel, P., Poole, B.: Zero-shot text-guided object generation with dream fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 867-876 (2022) +18. Kim, V.G., Chaudhuri, S., Guibas, L., Funkhouser, T.: Shape2Pose: Human-centric shape analysis. ACM Transactions on Graphics (Proc. SIGGRAPH) (Aug 2014) +19. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013) +20. Lee, H.H., Chang, A.X.: Understanding pure clip guidance for voxel grid nerf models. arXiv preprint arXiv:2209.15172 (2022) +21. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440 (2022) +22. Lin, Y., Bai, H., Li, S., Lu, H., Lin, X., Xiong, H., Wang, L.: Componerf: Text-guided multi-object compositional nerf with editable 3d scene layout. arXiv preprint arXiv:2303.13843 (2023) +23. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: European Conference on Computer Vision. pp. 423-439. Springer (2022) +24. Liu, Z., Wang, Y., Qi, X., Fu, C.W.: Towards implicit text-guided 3d shape generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17896-17906 (2022) +25. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Spl: A skinned multi-person linear model. ACM transactions on graphics (TOG) 34(6), 1-16 (2015) +26. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5442-5451 (2019) +27. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latentnerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600 (2022) +28. Michel, O., Bar-On, R., Liu, R., Benaim, S., Hanocka, R.: Text2mesh: Text-driven neural stylization for meshes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13492-13502 (2022) +29. Mihajlovic, M., Saito, S., Bansal, A., Zollhoefer, M., Tang, S.: Coap: Compositional articulated occupancy of people. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13201-13210 (2022) +30. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) + +31. Mirzaei, A., Kant, Y., Kelly, J., Gilitschenski, I.: Laterf: Label and text driven object radiance fields. In: European Conference on Computer Vision. pp. 20-36. Springer (2022) +32. Mohammad Khalid, N., Xie, T., Belilovsky, E., Popa, T.: Clip-mesh: Generating textured meshes from text using pretrained image-text models. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-8 (2022) +33. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022) +34. Nguyen-Phuoc, T.H., Richardt, C., Mai, L., Yang, Y., Mitra, N.: Blockgan: Learning 3d object-aware scene representations from unlabelled images. Advances in neural information processing systems 33, 6767-6778 (2020) +35. Niemeyer, M., Geiger, A.: Giraffe: Representing scenes as compositional generative neural feature fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11453-11464 (2021) +36. Ost, J., Mannan, F., Thuerey, N., Knodt, J., Heide, F.: Neural scene graphs for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2856-2865 (2021) +37. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10975-10985 (2019) +38. Peng, X., Xie, Y., Wu, Z., Jampani, V., Sun, D., Jiang, H.: Hoi-diff: Text-driven synthesis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553 (2023) +39. Po, R., Wetzstein, G.: Compositional 3d scene generation using locally conditioned diffusion. arXiv preprint arXiv:2303.12218 (2023) +40. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022) +41. Puig, X., Undersander, E., Szot, A., Cote, M.D., Yang, T.Y., Partsey, R., Desai, R., Clegg, A.W., Hlavac, M., Min, S.Y., et al.: Habitat 3.0: A co-habitat for humans, avatars and robots. arXiv preprint arXiv:2310.13724 (2023) +42. Punnakkal, A.R., Chandrasekaran, A., Athanasiou, N., Quiros-Ramirez, A., Black, M.J.: Babel: Bodies, action and behavior with english labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 722-731 (2021) +43. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022) +45. Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610 (2022) +46. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) + +47. Savva, M., Chang, A.X., Hanrahan, P., Fisher, M., Nießner, M.: Pigraphs: learning interaction snapshots from observations. ACM Transactions on Graphics (TOG) 35(4), 1-12 (2016) +48. Sella, E., Fiebelman, G., Hedman, P., Averbuch-Elor, H.: Vox-e: Text-guided voxel editing of 3d objects. arXiv preprint arXiv:2303.12048 (2023) +49. Sheridan, T.B.: Human-robot interaction: status and challenges. Human factors 58(4), 525-532 (2016) +50. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023) +51. Sohn, K., Lee, H., Yan, X.: Learning structured output representation using deep conditional generative models. Advances in neural information processing systems 28 (2015) +52. Song, Y., Kong, C., Lee, S., Kwak, N., Lee, J.: Towards efficient neural scene graphs by learning consistency fields. arXiv preprint arXiv:2210.04127 (2022) +53. Taheri, O., Ghorbani, N., Black, M.J., Tzionas, D.: Grab: A dataset of whole-body human grasping of objects. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 581-600. Springer (2020) +54. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023) +55. Von Marcard, T., Henschel, R., Black, M.J., Rosenhahn, B., Pons-Moll, G.: Recovering accurate 3d human pose in the wild using imus and a moving camera. In: Proceedings of the European conference on computer vision (ECCV). pp. 601-617 (2018) +56. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774 (2022) +57. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems 36 (2024) +58. Wu, Q., Shi, Y., Huang, X., Yu, J., Xu, L., Wang, J.: Thor: Text to human-object interaction diffusion via relation intervention. arXiv preprint arXiv:2403.11208 (2024) +59. Wu, Q., Liu, X., Chen, Y., Li, K., Zheng, C., Cai, J., Zheng, J.: Object-compositional neural implicit surfaces. In: European Conference on Computer Vision. pp. 197-213. Springer (2022) +60. Xu, Y., Chai, M., Shi, Z., Peng, S., Skorokhodov, I., Siarohin, A., Yang, C., Shen, Y., Lee, H.Y., Zhou, B., et al.: Discoscope: Spatially disentangled generative radiance fields for controllable 3d-aware scene synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4402-4412 (2023) +61. Yang, B., Zhang, Y., Xu, Y., Li, Y., Zhou, H., Bao, H., Zhang, G., Cui, Z.: Learning object-compositional neural radiance field for editable scene rendering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13779-13788 (2021) +62. Zhang, S., Zhang, Y., Ma, Q., Black, M.J., Tang, S.: Place: Proximity learning of articulation and contact in 3d environments. In: 2020 International Conference on 3D Vision (3DV). pp. 642-651. IEEE (2020) + +63. Zhang, Y., Hassan, M., Neumann, H., Black, M.J., Tang, S.: Generating 3d people in scenes without people. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6194-6204 (2020) +64. Zhao, K., Wang, S., Zhang, Y., Beeler, T., Tang, S.: Compositional human-scene interaction synthesis with semantic control. In: European Conference on Computer Vision. pp. 311-327. Springer (2022) +65. Zhi, S., Laidlow, T., Leutenegger, S., Davison, A.J.: In-place scene labelling and understanding with implicit scene representation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15838-15847 (2021) \ No newline at end of file diff --git a/2024/InterFusion_ Text-Driven Generation of 3D Human-Object Interaction/images.zip b/2024/InterFusion_ Text-Driven Generation of 3D Human-Object Interaction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..14c9228a409ee8afcda213ca79a0cf9681d53d07 --- /dev/null +++ b/2024/InterFusion_ Text-Driven Generation of 3D Human-Object Interaction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2184fcdfa1dd229f0e8f3f84111ea70076d938bace78624c3bb867522a33fa5 +size 416293 diff --git a/2024/InterFusion_ Text-Driven Generation of 3D Human-Object Interaction/layout.json b/2024/InterFusion_ Text-Driven Generation of 3D Human-Object Interaction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..beb8983462d9d6610477d5ee3db7bb02d0c78921 --- /dev/null +++ b/2024/InterFusion_ Text-Driven Generation of 3D Human-Object Interaction/layout.json @@ -0,0 +1,8756 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 170, + 111, + 446, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 111, + 446, + 148 + ], + "spans": [ + { + "bbox": [ + 170, + 111, + 446, + 148 + ], + "type": "text", + "content": "InterFusion: Text-Driven Generation of 3D Human-Object Interaction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 143, + 167, + 470, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 167, + 470, + 180 + ], + "spans": [ + { + "bbox": [ + 143, + 167, + 470, + 180 + ], + "type": "text", + "content": "Sisi Dai" + }, + { + "bbox": [ + 143, + 167, + 470, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 167, + 470, + 180 + ], + "type": "text", + "content": " Wenhao Li" + }, + { + "bbox": [ + 143, + 167, + 470, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 167, + 470, + 180 + ], + "type": "text", + "content": " Haowen Sun" + }, + { + "bbox": [ + 143, + 167, + 470, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 167, + 470, + 180 + ], + "type": "text", + "content": " Haibin Huang" + }, + { + "bbox": [ + 143, + 167, + 470, + 180 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 143, + 167, + 470, + 180 + ], + "type": "text", + "content": " Chongyang Ma" + }, + { + "bbox": [ + 143, + 167, + 470, + 180 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 219, + 180, + 394, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 180, + 394, + 192 + ], + "spans": [ + { + "bbox": [ + 219, + 180, + 394, + 192 + ], + "type": "text", + "content": "Hui Huang² Kai Xu1* Ruizhen Hu2*" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 215, + 201, + 399, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 201, + 399, + 213 + ], + "spans": [ + { + "bbox": [ + 215, + 201, + 399, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 215, + 201, + 399, + 213 + ], + "type": "text", + "content": " National University of Defense Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 260, + 213, + 353, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 213, + 353, + 223 + ], + "spans": [ + { + "bbox": [ + 260, + 213, + 353, + 223 + ], + "type": "text", + "content": "2 Shenzhen University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 258, + 224, + 355, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 224, + 355, + 235 + ], + "spans": [ + { + "bbox": [ + 258, + 224, + 355, + 235 + ], + "type": "text", + "content": "3 Kuaishou Technology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 216, + 236, + 397, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 236, + 397, + 246 + ], + "spans": [ + { + "bbox": [ + 216, + 236, + 397, + 246 + ], + "type": "text", + "content": "https://sisidai.github.io/InterFusion/" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 148, + 279, + 239, + 326 + ], + "blocks": [ + { + "bbox": [ + 148, + 279, + 239, + 326 + ], + "lines": [ + { + "bbox": [ + 148, + 279, + 239, + 326 + ], + "spans": [ + { + "bbox": [ + 148, + 279, + 239, + 326 + ], + "type": "image", + "image_path": "50527545e07cf197e0ea4b8ec32da26788a54af9738ffc8ea344664b715fc370.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 329, + 234, + 346 + ], + "lines": [ + { + "bbox": [ + 151, + 329, + 234, + 346 + ], + "spans": [ + { + "bbox": [ + 151, + 329, + 234, + 346 + ], + "type": "text", + "content": "a person in a military uniform sitting on a chair" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 252, + 279, + 362, + 328 + ], + "blocks": [ + { + "bbox": [ + 252, + 279, + 362, + 328 + ], + "lines": [ + { + "bbox": [ + 252, + 279, + 362, + 328 + ], + "spans": [ + { + "bbox": [ + 252, + 279, + 362, + 328 + ], + "type": "image", + "image_path": "17c5444b41efb96aeca0e190d47d2f4f8ad4a5923438856ad2850b3acf4e5bf0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 267, + 329, + 344, + 346 + ], + "lines": [ + { + "bbox": [ + 267, + 329, + 344, + 346 + ], + "spans": [ + { + "bbox": [ + 267, + 329, + 344, + 346 + ], + "type": "text", + "content": "a female gymnastics athlete riding a motorcycle" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 375, + 279, + 470, + 328 + ], + "blocks": [ + { + "bbox": [ + 375, + 279, + 470, + 328 + ], + "lines": [ + { + "bbox": [ + 375, + 279, + 470, + 328 + ], + "spans": [ + { + "bbox": [ + 375, + 279, + 470, + 328 + ], + "type": "image", + "image_path": "a2e56b4cc995146aa1aa82ecfa544cee682f2d8f82ad89f82701320f6f7d20fc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 329, + 466, + 346 + ], + "lines": [ + { + "bbox": [ + 374, + 329, + 466, + 346 + ], + "spans": [ + { + "bbox": [ + 374, + 329, + 466, + 346 + ], + "type": "text", + "content": "a person in a fleece pullover and hiking pants carrying a backpack" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 145, + 350, + 240, + 399 + ], + "blocks": [ + { + "bbox": [ + 145, + 350, + 240, + 399 + ], + "lines": [ + { + "bbox": [ + 145, + 350, + 240, + 399 + ], + "spans": [ + { + "bbox": [ + 145, + 350, + 240, + 399 + ], + "type": "image", + "image_path": "f195b07182d8df6deab02e6d56a0f0f2cb08cc8c01932f8366d5baa3597109c4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 402, + 242, + 418 + ], + "lines": [ + { + "bbox": [ + 143, + 402, + 242, + 418 + ], + "spans": [ + { + "bbox": [ + 143, + 402, + 242, + 418 + ], + "type": "text", + "content": "a man in a quilted gilet and moleskin trousers playing the guitar" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 252, + 350, + 359, + 402 + ], + "blocks": [ + { + "bbox": [ + 252, + 350, + 359, + 402 + ], + "lines": [ + { + "bbox": [ + 252, + 350, + 359, + 402 + ], + "spans": [ + { + "bbox": [ + 252, + 350, + 359, + 402 + ], + "type": "image", + "image_path": "92890147ea2c39a7a17782e400d6fe8d58673aa4c49fcdd0b53c740c8d90279f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 267, + 403, + 345, + 418 + ], + "lines": [ + { + "bbox": [ + 267, + 403, + 345, + 418 + ], + "spans": [ + { + "bbox": [ + 267, + 403, + 345, + 418 + ], + "type": "text", + "content": "a man wearing a WORLD lettered T-shirt riding a bike" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 369, + 351, + 473, + 399 + ], + "blocks": [ + { + "bbox": [ + 369, + 351, + 473, + 399 + ], + "lines": [ + { + "bbox": [ + 369, + 351, + 473, + 399 + ], + "spans": [ + { + "bbox": [ + 369, + 351, + 473, + 399 + ], + "type": "image", + "image_path": "aa89e5e1a2d9857f9c1d6945dcb75bc3475d0d357ba44dd64349d0dc978bdd56.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 402, + 466, + 418 + ], + "lines": [ + { + "bbox": [ + 374, + 402, + 466, + 418 + ], + "spans": [ + { + "bbox": [ + 374, + 402, + 466, + 418 + ], + "type": "text", + "content": "a person in a cropped hoodie and bike shorts jumping a skateboard" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 149, + 423, + 241, + 474 + ], + "blocks": [ + { + "bbox": [ + 149, + 423, + 241, + 474 + ], + "lines": [ + { + "bbox": [ + 149, + 423, + 241, + 474 + ], + "spans": [ + { + "bbox": [ + 149, + 423, + 241, + 474 + ], + "type": "image", + "image_path": "3024fe4eb00435fcd76f639eb90deadeddb99818b0799aba7148089b8aec6496.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 476, + 248, + 492 + ], + "lines": [ + { + "bbox": [ + 137, + 476, + 248, + 492 + ], + "spans": [ + { + "bbox": [ + 137, + 476, + 248, + 492 + ], + "type": "text", + "content": "a man in a velvet smoking jacket and silk pajama pants playing the saxophone" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 254, + 424, + 356, + 474 + ], + "blocks": [ + { + "bbox": [ + 254, + 424, + 356, + 474 + ], + "lines": [ + { + "bbox": [ + 254, + 424, + 356, + 474 + ], + "spans": [ + { + "bbox": [ + 254, + 424, + 356, + 474 + ], + "type": "image", + "image_path": "88157f0a186ebce2852d8e61b2d4a37c57c6b6dd5834d1407aed4f5d88f50332.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 266, + 475, + 347, + 492 + ], + "lines": [ + { + "bbox": [ + 266, + 475, + 347, + 492 + ], + "spans": [ + { + "bbox": [ + 266, + 475, + 347, + 492 + ], + "type": "text", + "content": "a man in a paisley waistcoat and bowler hat riding a horse" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 369, + 423, + 470, + 472 + ], + "blocks": [ + { + "bbox": [ + 369, + 423, + 470, + 472 + ], + "lines": [ + { + "bbox": [ + 369, + 423, + 470, + 472 + ], + "spans": [ + { + "bbox": [ + 369, + 423, + 470, + 472 + ], + "type": "image", + "image_path": "fcdaaf3478512b2e9be0025e1da06574883663a508d0a923d8a75af0d79ccf7d.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 475, + 472, + 492 + ], + "lines": [ + { + "bbox": [ + 369, + 475, + 472, + 492 + ], + "spans": [ + { + "bbox": [ + 369, + 475, + 472, + 492 + ], + "type": "text", + "content": "a man with blond hair wearing a brown leather jacket reading a laptop" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 507, + 482, + 529 + ], + "lines": [ + { + "bbox": [ + 132, + 507, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 507, + 482, + 529 + ], + "type": "text", + "content": "Fig. 1: Given a text prompt, our method InterFusion can generate diverse 3D scenes of a person interacting with an object." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 159, + 570, + 453, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 570, + 453, + 647 + ], + "spans": [ + { + "bbox": [ + 159, + 570, + 453, + 647 + ], + "type": "text", + "content": "Abstract. In this study, we tackle the complex task of generating 3D human-object interactions (HOI) from textual descriptions in a zero-shot text-to-3D manner. We identify and address two key challenges: the unsatisfactory outcomes of direct text-to-3D methods in HOI, largely due to the lack of paired text-interaction data, and the inherent difficulties in simultaneously generating multiple concepts with complex spatial relationships. To effectively address these issues, we present InterFusion," + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 654, + 436, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 436, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 436, + 665 + ], + "type": "text", + "content": "* Corresponding authors: kevin.kai.xu@gmail.com; ruizhen.hu@gmail.com." + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 281 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 281 + ], + "type": "text", + "content": "a two-stage framework specifically designed for HOI generation. InterFusion involves human pose estimations derived from text as geometric priors, which simplifies the text-to-3D conversion process and introduces additional constraints for accurate object generation. At the first stage, InterFusion extracts 3D human poses from a synthesized image dataset depicting a wide range of interactions, subsequently mapping these poses to interaction descriptions. The second stage of InterFusion capitalizes on the latest developments in text-to-3D generation, enabling the production of realistic and high-quality 3D HOI scenes. This is achieved through a local-global optimization process, where the generation of human body and object is optimized separately, and jointly refined with a global optimization of the entire scene, ensuring a seamless and contextually coherent integration. Our experimental results affirm that InterFusion significantly outperforms existing state-of-the-art methods in 3D HOI generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 293, + 459, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 293, + 459, + 315 + ], + "spans": [ + { + "bbox": [ + 160, + 293, + 459, + 315 + ], + "type": "text", + "content": "Keywords: Text-Driven Generation " + }, + { + "bbox": [ + 160, + 293, + 459, + 315 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 293, + 459, + 315 + ], + "type": "text", + "content": " Zero-Shot Generation " + }, + { + "bbox": [ + 160, + 293, + 459, + 315 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 293, + 459, + 315 + ], + "type": "text", + "content": " 3D Human-Object Interaction Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 338, + 230, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 338, + 230, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 338, + 230, + 350 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "content": "The generation of 3D human-object interactions (HOI) stands as a critical challenge in the fields of computer vision and computer graphics, with far-reaching implications in virtual reality, augmented reality, animation, and embodied AI [41, 49]. This task entails the creation of realistic 3D scenes where human figures interact with objects in ways that are not only physically plausible but also contextually relevant." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 568 + ], + "type": "text", + "content": "Despite its potential, the field has faced significant obstacles, primarily due to the scarcity of large-scale interaction data. Traditional approaches have predominantly relied on motion capture (mocap) datasets or physics-based simulations for generating these interactions. Mocap datasets [15,26,55] are limited by the specific scenarios they capture and are both costly and labor-intensive to produce. These limitations have resulted in a notable gap in generating diverse and contextually rich HOI scenes, especially for novel or complex interactions. Conversely, recent advancements have introduced text-to-3D methods [40,56], marking a significant shift in the field. These methods harness the power of textual descriptions to generate 3D objects without direct 3D supervision, presenting a novel approach to 3D content creation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "In this study, we explore text-to-3D method in HOI task within a zero-shot manner, i.e., generating 3D scenes from textual descriptions using 2D diffusion models. Our key observations are two-fold: first of all, a direct application of text-to-3D method in HOI often leads to unsatisfied results like blurry textures and incorrect interactions, which is caused by the relative lack of paired text-interaction data during training and the difficulties of generating multiple concepts for diffusion-based methods [23]. Secondly, while collecting extensive interaction data is challenging, estimating human poses based on described in" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "text", + "content": "Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "teractions is more feasible. These pose estimations can serve as geometric priors in the HOI generation process. By integrating pose estimation, our method substantially simplifies the text-to-3D process, particularly in the geometry optimization stage for human body generation. It also provides additional constraints for object generation, ensuring accurate placement and alignment with the text's semantic content. More importantly, we can separate the generation of human body and object, and jointly refine the details with the estimated pose, allowing for a more coherent and detailed synthesis of the interaction scene." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 219, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 219, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 482, + 483 + ], + "type": "text", + "content": "Inspired by these observations, we introduce InterFusion, a novel two-stage framework designed for HOI generation. Specifically, in the first stage, rather than relying on precise 3D interaction data, our approach instead collects a comprehensive dataset of synthesized images that depict a wide range of interactions. From these images, we employ advanced 3D pose estimation technique [8] to extract 3D human poses. Building upon these image-pose pairs, InterFusion develops a sophisticated codebook that establishes a mapping between interaction descriptions and 3D human poses, with the integration of the CLIP (Contrastive Language-Image Pretraining) embedding [43]. By leveraging these embeddings, our framework is able to interpret the nuances of interaction descriptions and translate them into accurate 3D pose representations. In the subsequent stage, InterFusion capitalizes on recent advancements in text-to-3D generation [40], as well as neural radiance fields [30], using the estimated human poses to produce 3D HOI scenes with realistic appearances and high-quality geometry. This stage operates in a 'local-global' manner. At the local level, the generation of the human body (SDS-H) and objects (SDS-O) is separately optimized, with the poses serving as additional constraints for the SDS. At the global level, the generation of the entire scene (SDS-I) is also guided by the integrated description and jointly optimized with SDS-H and SDS-O, ensuring a cohesive and contextually accurate representation of the HOI. Our experiments show that the quality of generation can be improved by a large margin and our approach outperforms state-of-the-art methods in HOI generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 488, + 357, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 488, + 357, + 500 + ], + "spans": [ + { + "bbox": [ + 147, + 488, + 357, + 500 + ], + "type": "text", + "content": "To summarize, our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 544, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 544, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 481, + 582 + ], + "type": "text", + "content": "- We introduce a novel two-stage framework InterFusion, for zero-shot 3D human-object interaction generation from text, incorporating 3D pose estimation as geometry priors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 586, + 481, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 586, + 481, + 624 + ], + "spans": [ + { + "bbox": [ + 138, + 586, + 481, + 624 + ], + "type": "text", + "content": "- InterFusion leverages text-to-3D generation with a local-global optimization process. This strategy ensures seamless integration of human bodies and objects, producing realistic and high-quality 3D HOI scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 629, + 481, + 665 + ], + "type": "text", + "content": "- InterFusion demonstrates significant improvements over existing methods in 3D HOI generation, showcasing its effectiveness in creating detailed, and contextually rich 3D interactions." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "text", + "content": "InterFusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 143, + 346, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 143, + 346, + 156 + ], + "spans": [ + { + "bbox": [ + 132, + 143, + 346, + 156 + ], + "type": "text", + "content": "2.1 Human-Object Interaction Synthesis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 166, + 482, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 166, + 482, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 166, + 482, + 466 + ], + "type": "text", + "content": "3D human-object interaction (HOI) generation is a challenging problem that has been studied widely by the computer vision and graphics community. Shape2Pose [18] generates plausible 3D human poses interacting with a given 3D object model by learning an affordance model from synthetic data. PiGraphs [47] learns the distribution between human poses and object arrangements from a collected dataset with 3D scene scans and RGB-D videos, generating the interaction snapshots given action specifications and object models. Recently, the parametric human body models such as SMPL [25,37,45] are employed in interaction synthesis to overcome the lack of realism due to human body representations. Benefiting from the PROX dataset [12] which consists of fitted SMPL models in captured 3D scenes, the more specific human-scene interaction has become an active research direction. Given a 3D scene, PSI [63] and PLACE [62] generate the 3D human body mesh represented as SMPL parameters through a conditional variational autoencoder [19, 51]. POSA [13] learns pose-specific priors to generate contacts conditioned by the given posed human, which can further guide the placement of the body mesh in a scene. COINS [64] enables semantic control on interaction synthesis by embedding the action label together with the interacted object as the condition of the generative model. More recently, fine-grained 3D interaction datasets [1, 7, 53] are captured to promote this field. Relying on these datasets, some methods [6, 38, 58], concurrent to this work, are proposed to explore text-guided 3D HOI generation. However, they remain constrained by distributions within the datasets. While previous methods need ground truth 3D interaction data as supervision, our work, for the first time, attempts to break through the limitation of data requirement. We generate a wider range of realistic and detailed 3D HOI scenes, including both indoors and outdoors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 487, + 312, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 312, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 312, + 499 + ], + "type": "text", + "content": "2.2 Text-to-3D Content Synthesis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 483, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 483, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 483, + 668 + ], + "type": "text", + "content": "Early methods [3, 16, 24] for text-to-3D shape generation require paired data of 3D data and the corresponding textual descriptions to learn the joint embedding space of shape and text for supervision, which limits their generality to unseen object categories. Benefiting from large pre-trained text-to-image models and differentiable rendering techniques, breakthroughs in text-to-3D content generation have been achieved. For example, DreamFields [17] and PureCLIPN-eRF [20] combine CLIP [43] with neural radiance fields (NeRF) [30], demonstrating the potential for zero-shot NeRF optimization. Meanwhile, CLIP-mesh [32] and Text2Mesh [28] incorporate CLIP to optimize the 3D mesh representation, starting from an initial sphere mesh and an input base mesh, respectively. Recently, DreamFusion [40] and SJC [56] enable NeRF optimization with guidance from pre-trained text-to-image diffusion models [44,46] in place of CLIP, achieving more impressive results. To improve DreamFusion, Magic3D [21] proposes" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "type": "text", + "content": "Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "a coarse-to-fine pipeline to generate the fine-grained mesh. TextMesh [54] extends the geometry representation from NeRF to an SDF framework, thereby enhancing detailed mesh extraction and photorealistic rendering. Among follow-up works, Latent-NeRF [27] and Vox-E [48] utilize explicit 3D shapes to provide additional training signals for NeRF optimization. While Latent-NeRF utilizes a rough untextured object for shape sculpture, Vox-E takes multiview images of a fine-grained textured object to edit geometry and appearance. All the methods mentioned above focus on the generation or edition of a single subject, while ignoring the interaction between different subjects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 243, + 326, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 243, + 326, + 255 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 326, + 255 + ], + "type": "text", + "content": "2.3 Compositional Scene Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 264, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 264, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 482, + 431 + ], + "type": "text", + "content": "Representing scenes as compositions of object representations facilitates enhanced controllability. Numerous techniques incorporate additional information at the object level to perform compositional modeling of scenes, effectively separating object representations from the overall scene imagery. For instance, some methods incorporate 2D semantic information such as segmentation labels [65], instance masks [59,61], or features from a pre-trained vision-language model [31]. Some other methods [10, 36, 52] use 3D layout information by object-centric bounding boxes with canonical coordinates. When it comes to scene generation, [34, 35, 60] use compositional representations to generate scenes in a controllable manner. More recent approaches [4, 22, 39] generate compositional 3D scenes from input text prompts. Different from existing methods, our HOI scene is generated automatically without the requirement of input layout. Moreover, the spatial relationship between human and object during interaction is much more complex and cannot be simply characterized using their bounding boxes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 450, + 230, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 450, + 230, + 463 + ], + "spans": [ + { + "bbox": [ + 131, + 450, + 230, + 463 + ], + "type": "text", + "content": "3 Our Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 476, + 209, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 476, + 209, + 488 + ], + "spans": [ + { + "bbox": [ + 131, + 476, + 209, + 488 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "text", + "content": "In this section, we formally introduce InterFusion with a focus on the zero-shot and text-driven generation of 3D human-object interactions (HOI). Specifically, the input is a triplet of text descriptions, " + }, + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "inline_equation", + "content": "T = \\{T^H, T^O, T^I\\}" + }, + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "text", + "content": ", specifying the desired human style, object style, and interaction type. The goal is to generate a detailed 3D scene, " + }, + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\psi = \\{\\psi^H, \\psi^O\\}" + }, + { + "bbox": [ + 130, + 498, + 482, + 581 + ], + "type": "text", + "content": ", comprising a human model and an object model that not only adhere to the specified appearance styles but also exhibit a tailored spatial relationship to accurately reflect the described interaction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "As illustrated in Figure 2, our method consists of two primary stages: anchor pose generation and anchor pose guided HOI generation. We first generate an interaction pose based on the input text, termed an anchored pose. This pose then serves as a geometric constraint, guiding the subsequent generation of detailed HOI. In the second phase, the human and object models are optimized separately and refined simultaneously with a global context, ensuring a cohesive and accurate representation of the interaction as described by the input text." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "text", + "content": "InterFusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 118, + 479, + 297 + ], + "blocks": [ + { + "bbox": [ + 136, + 118, + 479, + 297 + ], + "lines": [ + { + "bbox": [ + 136, + 118, + 479, + 297 + ], + "spans": [ + { + "bbox": [ + 136, + 118, + 479, + 297 + ], + "type": "image", + "image_path": "53294ededa3526b22fbae85de4df0dcb00b770483fc2042d3e9e7a401d299b72.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 308, + 482, + 354 + ], + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 354 + ], + "type": "text", + "content": "Fig. 2: InterFusion is a two-stage framework that transforms textual descriptions into detailed 3D human-object interactions, initially synthesizing anchor poses (upper left) and then optimizing the human model (upper right) and object model (bottom) with constraints from estimated pose and textual prompts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 376, + 228, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 228, + 388 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 228, + 388 + ], + "type": "text", + "content": "3.2 Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "text", + "content": "SDS. Score Distillation Sampling (SDS) has been introduced by DreamFusion [40]. While " + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "inline_equation", + "content": "x = g(\\psi, \\zeta)" + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "text", + "content": " is the 2D image rendered by a differentiable renderer " + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "text", + "content": " with model parameters " + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "text", + "content": " (the volumetric renderer and MLPs correspondingly in NeRF), under a desired camera pose " + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "text", + "content": ". By injecting the sampled noise " + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "text", + "content": " at a time step " + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "text", + "content": ", the noisy image " + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "text", + "content": " is produced. The pre-trained 2D text-to-image diffusion model " + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "text", + "content": " provides a denoising network " + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\phi}(x_{t};y,t)" + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "text", + "content": " that predicts the noise " + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}" + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "text", + "content": " given the noisy image " + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "text", + "content": ", time step " + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "text", + "content": ", and text embedding " + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "text", + "content": ". SDS then optimizes the model parameters " + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 130, + 399, + 482, + 507 + ], + "type": "text", + "content": " by minimizing the difference between the predicted noise and the added noise:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 204, + 516, + 481, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 516, + 481, + 541 + ], + "spans": [ + { + "bbox": [ + 204, + 516, + 481, + 541 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\psi} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , x) = \\mathbb {E} _ {t, \\epsilon} [ w (t) (\\hat {\\epsilon} _ {\\phi} (x _ {t}; y, t) - \\epsilon) \\frac {\\partial x}{\\partial \\psi} ], \\tag {1}", + "image_path": "60b740223c23501a470a10b4b139f61eaf1f8e5bdc504b8bb26a052dccd30c59.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 551, + 361, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 551, + 361, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 551, + 361, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 551, + 361, + 563 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 132, + 551, + 361, + 563 + ], + "type": "text", + "content": " is the weighting term at the time step " + }, + { + "bbox": [ + 132, + 551, + 361, + 563 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 551, + 361, + 563 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 582, + 285, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 285, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 285, + 594 + ], + "type": "text", + "content": "3.3 Anchor Pose Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "InterFusion starts by generating the anchor pose from the input text " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "T^I" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", which is a non-trivial problem and highly limited by the available pose datasets. Existing datasets (e.g., HumanML3D [9], BABEL [42]) are mocap based datasets and are typically conducted in a controlled laboratory environment, leading to lacks in action diversity and spatial coverage." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "text", + "content": "Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 259 + ], + "type": "text", + "content": "Recently, text-to-image diffusion models have shown the versatility and effectiveness of generating realistic and diverse images, through the integration of visual and linguistic feature spaces. We thus instead utilize advancements in this technology, specifically models like Stable Diffusion, to create a large-scale dataset of Human-Object Interaction (HOI) images. These images depict a wide range of human interactions. To ensure the diversity of interaction types, we utilize ChatGPT to generate prompts about human daily events or actions forming \"verb-ing a/an/the object\". A total of 235 result prompts are generated, covering most interactions in daily life. After filtering synthesized images without humans, we then estimate 3D human poses using the pre-trained PixIE [8] model, creating a comprehensive Syn-HOI pose dataset consisting of a total 55K 3D pseudo-SMPL poses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 262, + 480, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 480, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 480, + 382 + ], + "type": "text", + "content": "While SMPL model maps pose and shape parameters to a triangulated mesh, to align these poses with our text-to-interaction generation task, we render images from multiple perspectives and use CLIP's image encoder to derive pose feature embeddings. The average feature across these multiple perspective images represents the pose feature and attaches the dataset with pairs of averaged CLIP embedding and pose parameters. We further construct a codebook using K-Means clustering to identify key pose centroids based on the feature embeddings. 2,048 cluster centroids are clustered compositing our pose code-book, where each cluster comprises a subset of poses that represents similar interaction as the key poses of the centroids." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 385, + 480, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 480, + 447 + ], + "type": "text", + "content": "Once the codebook is built, for a given input text, we extract its feature embedding using CLIP's text encoder. This serves as a query to retrieve the most similar poses from the codebook, based on feature similarity. Specifically, given query text " + }, + { + "bbox": [ + 130, + 385, + 480, + 447 + ], + "type": "inline_equation", + "content": "T^{I}" + }, + { + "bbox": [ + 130, + 385, + 480, + 447 + ], + "type": "text", + "content": ", top " + }, + { + "bbox": [ + 130, + 385, + 480, + 447 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 385, + 480, + 447 + ], + "type": "text", + "content": " poses " + }, + { + "bbox": [ + 130, + 385, + 480, + 447 + ], + "type": "inline_equation", + "content": "\\theta_{k}^{T^{I}}" + }, + { + "bbox": [ + 130, + 385, + 480, + 447 + ], + "type": "text", + "content": " could be matched by pose embeddings " + }, + { + "bbox": [ + 130, + 385, + 480, + 447 + ], + "type": "inline_equation", + "content": "\\theta_{E}" + }, + { + "bbox": [ + 130, + 385, + 480, + 447 + ], + "type": "text", + "content": ", the averaged CLIP embeddings among rendered images from key poses:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 246, + 461, + 480, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 461, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 246, + 461, + 480, + 477 + ], + "type": "interline_equation", + "content": "\\theta_ {k} ^ {T ^ {I}} = \\operatorname {T O P} _ {k} \\left(f _ {\\text {t e x t}} \\left(T ^ {I}\\right), \\theta_ {E}\\right) \\tag {2}", + "image_path": "77cda589cdd13d6f9984de830f33c7d5ac2f5949d5a0d2345b7aba1437abdaaf.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 491, + 479, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 479, + 527 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 479, + 527 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 491, + 479, + 527 + ], + "type": "inline_equation", + "content": "f_{text}" + }, + { + "bbox": [ + 130, + 491, + 479, + 527 + ], + "type": "text", + "content": " is the text encoder of CLIP and " + }, + { + "bbox": [ + 130, + 491, + 479, + 527 + ], + "type": "inline_equation", + "content": "\\mathrm{TOP}_k(X,Y)" + }, + { + "bbox": [ + 130, + 491, + 479, + 527 + ], + "type": "text", + "content": " returns top " + }, + { + "bbox": [ + 130, + 491, + 479, + 527 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 491, + 479, + 527 + ], + "type": "text", + "content": " poses with the top " + }, + { + "bbox": [ + 130, + 491, + 479, + 527 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 491, + 479, + 527 + ], + "type": "text", + "content": " highest cosine similarities, and we use " + }, + { + "bbox": [ + 130, + 491, + 479, + 527 + ], + "type": "inline_equation", + "content": "k = 7" + }, + { + "bbox": [ + 130, + 491, + 479, + 527 + ], + "type": "text", + "content": " for suitable poses in our experiments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 531, + 479, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 479, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 479, + 590 + ], + "type": "text", + "content": "We then utilize GPT-4V to select the most precise pose as the final queried key pose. Depending on the requirement, we can instead select the poses sampled in the cluster, corresponding to the key pose, to guide the generation process for diverse results. This approach ensures rich and contextually aligned anchor poses for our text-to-interaction generation task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "content": "To restrict the optimization of geometry and appearance using human structural priors from the acquired pose, we further incorporate COAP [29], a neural occupancy representation of the articulated human body based on SMPL parameters [25]. Given a pose " + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "content": " and a shape " + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "content": ", COAP offers an occupancy prediction network " + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "inline_equation", + "content": "f(x; \\beta, \\theta)" + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "content": " that maps a 3D query point " + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "content": " to an occupancy value, directly indicating whether the spatial point resides within the 3D body." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "text", + "content": "InterFusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 309, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 309, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 309, + 128 + ], + "type": "text", + "content": "3.4 Pose-Guided HOI Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 210 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 210 + ], + "type": "text", + "content": "Once get the interaction pose, we further use it along with the input text to guide the generation of a detailed 3D HOI scene. In this phase, the estimated pose serves both as spatial constraints for the scene's geometry and as an anchor that aligns the human and object models. This approach ensures that the human model and the object model can be generated separately, while they are cohesively aligned to form the final 3D HOI scene." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "type": "text", + "content": "The anchor pose provides specific spatial constraints for each component in the scene. For the human model, it establishes a basic geometric structure, while for the object model, it defines the areas that should remain unoccupied. This clear distinction is essential for rendering the scene both accurately and realistically. Simultaneously, the input text undergoes a complex processing procedure to offer distinct semantic guidance for each component. This is accomplished through various text conditioning techniques applied in SDS with the pre-trained DeepFloyd model [5]. The text is carefully crafted to direct the generation of both the human and object models to ensure all elements are in harmony with the semantic nuances of input descriptions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 331, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 413 + ], + "type": "text", + "content": "Additionally, we introduce a novel camera tracing module to enhance the optimization process under varying text conditions. This module adaptively adjusts the camera pose, focusing on relevant elements within the scene at each optimization stage. This adaptive camera positioning is instrumental in ensuring that each aspect of the scene is optimally rendered according to the text descriptions, resulting in a more dynamic and contextually accurate 3D HOI scene." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 415, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 415, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 482, + 453 + ], + "type": "text", + "content": "We now provide details of each module in this stage, including the Neural Radiance Field representations for the human model (H-NeRF) and object model (O-NeRF), the camera tracing mechanism, and the guided optimization process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 472, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 472, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 472, + 482, + 605 + ], + "type": "text", + "content": "H-NeRF. We use NeRF to represent the human model, noted as H-NeRF. The generation of H-NeRF is guided by the text description specifying the human style, conducting pose-specific human avatar generation with SDS. To enhance the quality of our renderings, particularly in terms of resolution, we incorporated a specialized optimization process that focuses on the head region of the human avatar. The location of the head for any given pose is determined with COAP [29]. This allows us to augment the text prompt specifically for the head region, using the notation * the head of *, to ensure that the head receives detailed attention during the generation process. The loss function for this optimization process is as follows, designed to balance the fidelity of the head region with the overall pose and style of the human figure." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 197, + 615, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 615, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 197, + 615, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\psi^ {H}} \\mathcal {L} _ {\\mathrm {S D S}} = \\mathbb {E} _ {t, \\epsilon} [ w (t) (\\hat {\\epsilon} _ {\\phi} \\left(x _ {t} ^ {H}; y ^ {H}, t\\right) - \\epsilon) \\frac {\\partial x ^ {H}}{\\partial \\psi^ {H}} ] \\tag {3} \\\\ + \\mathbb {E} _ {t, \\epsilon} [ w (t) (\\hat {\\epsilon} _ {\\phi} (x _ {t} ^ {H, h}; y ^ {H, h}, t) - \\epsilon) \\frac {\\partial x ^ {H , h}}{\\partial \\psi^ {H}} ], \\\\ \\end{array}", + "image_path": "62fb7e4bde0431b0faf0cc778e5d3bea5917d6914570456308cef97663347a79.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "text", + "content": "Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "In shaping H-NeRF, our geometric constraint ensures that the human model evolves directly from the anchor pose. Points within the anchor are required to be occupied, establishing a firm base for the model. Meanwhile, points outside the anchor can also be occupied, but with probabilities that decrease as they move away from the anchor's surface. This approach allows for the addition of geometric details over the anchor, ensuring the model aligns with the human style described in the text while maintaining a coherent structure rooted in the anchor pose. The loss function used is as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 220, + 220, + 480, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 220, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 220, + 220, + 480, + 255 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {g e o}} ^ {H} = C E _ {p _ {i} \\in \\mathbb {P} _ {\\mathrm {i n}}} \\left(\\alpha_ {i}, f (p _ {i})\\right) \\\\ + C E _ {p _ {j} \\in \\mathbb {P} _ {\\text {o u t}}} \\left(\\alpha_ {j}, f (p _ {j})\\right) \\left(1 - e ^ {- \\frac {d}{2 \\eta^ {2}}}\\right), \\tag {4} \\\\ \\end{array}", + "image_path": "3bea006854b952ce08ff543e3978f11ae9d803c82a1ed40657971c88061ddd67.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 481, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 481, + 283 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 258, + 481, + 283 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 258, + 481, + 283 + ], + "type": "text", + "content": " represents the point distance from the anchor surface, and " + }, + { + "bbox": [ + 130, + 258, + 481, + 283 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 258, + 481, + 283 + ], + "type": "text", + "content": " is the hyperparameter to control the extent of decaying, similar as in [27]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 300, + 482, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 300, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 131, + 300, + 482, + 360 + ], + "type": "text", + "content": "O-NeRF. We also employ NeRF to model the object component, referred to as O-NeRF. This model is designed to interact seamlessly with the human model and to embody the desired style as specified by the input text. To guide the generation of both the interaction type and the object style, we utilize SDS with tailored text prompts:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 208, + 368, + 480, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 368, + 480, + 421 + ], + "spans": [ + { + "bbox": [ + 208, + 368, + 480, + 421 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\psi^ {O}} \\mathcal {L} _ {\\mathrm {S D S}} = \\mathbb {E} _ {t, \\epsilon} [ w (t) (\\hat {\\epsilon} _ {\\phi} \\left(x _ {t} ^ {I}; y ^ {I}, t\\right) - \\epsilon) \\frac {\\partial x ^ {I}}{\\partial \\psi^ {O}} ] \\tag {5} \\\\ + \\mathbb {E} _ {t, \\epsilon} [ w (t) (\\hat {\\epsilon} _ {\\phi} (x _ {t} ^ {O}; y ^ {O}, t) - \\epsilon) \\frac {\\partial x ^ {O}}{\\partial \\psi^ {O}} ], \\\\ \\end{array}", + "image_path": "7592f158f16ec4efb8f460dc4219529fb66d3fc5168b1d55733c5feab30dd070.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 430, + 482, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 503 + ], + "type": "text", + "content": "For the interaction scene " + }, + { + "bbox": [ + 130, + 430, + 482, + 503 + ], + "type": "inline_equation", + "content": "x^{I}" + }, + { + "bbox": [ + 130, + 430, + 482, + 503 + ], + "type": "text", + "content": " generation, we use an alpha-composed rendering to integrate H-NeRF and O-NeRF. This method calculates an alpha value from the density at each point, determining its contribution to the scene's color. Higher alpha values indicate a greater influence on the rendering, allowing for a nuanced and realistic integration of the human and object models in the final interaction image:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 229, + 513, + 480, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 513, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 229, + 513, + 480, + 548 + ], + "type": "interline_equation", + "content": "x ^ {I} = \\sum_ {i} w _ {i} ^ {I} c _ {i} ^ {I}, w _ {i} ^ {I} = \\alpha_ {i} ^ {I} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j} ^ {I}), \\tag {6}", + "image_path": "3a12a0dd01a90fa5a49865dea351356aea94a97ac11a6e9ae3c690a4524aa13a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 229, + 548, + 370, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 548, + 370, + 574 + ], + "spans": [ + { + "bbox": [ + 229, + 548, + 370, + 574 + ], + "type": "interline_equation", + "content": "c _ {i} ^ {I} = \\frac {\\alpha_ {i} ^ {H}}{\\alpha_ {i} ^ {H} + \\alpha_ {i} ^ {O}} c _ {i} ^ {H} + \\frac {\\alpha_ {i} ^ {O}}{\\alpha_ {i} ^ {H} + \\alpha_ {i} ^ {O}} c _ {i} ^ {I}.", + "image_path": "f003892434522124de83002c29e57e996b1c2ad43f604c205cc6398c4981eabc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "In our NeRF-based approach, alpha values, capped at 1, are derived from density for composite rendering. This setup allows semantic guidance gradients, conditioned by the interaction image, to optimize the density and color of both the human and object models. However, we noted a tendency for the model to prioritize object generation at the expense of the human model's quality. To counteract this, we implemented gradient truncation towards the human model to maintain a balanced optimization between the two components." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "text", + "content": "InterFusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "For O-NeRF's geometric constraints, we aim to prevent occupancy of points within the anchor pose, denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "x_{i} \\in \\mathbb{X}_{in}" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ", by the object model. Inspired by the physical collision prevention concept, we define a specific loss function that ensures these points remain unoccupied by O-NeRF, effectively preventing overlap between the human and object models in the 3D space:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 240, + 186, + 481, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 186, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 240, + 186, + 481, + 201 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {g e o}} ^ {O} = C E _ {p _ {i} \\in \\mathbb {P} _ {\\mathrm {i n}}} \\left(\\alpha_ {i}, 1 - f \\left(p _ {i}\\right)\\right) \\tag {7}", + "image_path": "1b0fdd2193cace3bbffc23d6a7160e1d2da96c1429db8297ddb3e76e6e03b566.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 209, + 481, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 209, + 481, + 233 + ], + "spans": [ + { + "bbox": [ + 131, + 209, + 481, + 233 + ], + "type": "text", + "content": "By minimizing the above loss term, model parameters are optimized to enhance the geometric consistency between H-NeRF and O-NeRF." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 253, + 482, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 253, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 482, + 338 + ], + "type": "text", + "content": "Camera tracing. To enhance the generation of O-NeRF, we introduce a dynamic camera tracing module within the SDS framework. This module automatically adjusts the camera pose to focus on the target's center, either the entire interaction scene or just the object during object generation. The camera aligns with the average position of voxels with an occupancy probability above 0.5, ensuring a consistent focus on the most significant parts of the scene or object for optimal detail capture and realism." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 357, + 378, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 357, + 378, + 369 + ], + "spans": [ + { + "bbox": [ + 131, + 357, + 378, + 369 + ], + "type": "text", + "content": "Optimization. The total loss is formulated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 205, + 378, + 481, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 378, + 481, + 393 + ], + "spans": [ + { + "bbox": [ + 205, + 378, + 481, + 393 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {S D S}} ^ {H} + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {S D S}} ^ {O} + \\mathcal {L} _ {\\mathrm {g e o}} ^ {H} + \\lambda_ {2} \\mathcal {L} _ {\\mathrm {g e o}} ^ {O} + \\lambda_ {3} L _ {\\text {r e g}}, \\tag {8}", + "image_path": "7ac13b7a32f3fd1ed2125f3aeabab3c1cd7623aed63f3948540b77103830294c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2" + }, + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "type": "inline_equation", + "content": "\\lambda_3" + }, + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "type": "text", + "content": " are the corresponding loss weights. Figure 2 (left bottom) shows several intermediate states of the object model during the generation steps and we can see that the object is gradually generated with fine details. Weight annealing is adopted during the guided optimization process. We leave the details in the supplement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 483, + 230, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 483, + 230, + 496 + ], + "spans": [ + { + "bbox": [ + 131, + 483, + 230, + 496 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 593 + ], + "type": "text", + "content": "Figure 3 (left) shows results with various interaction poses, showcasing the strength of our method. Our method can support diverse interaction poses within a single type. Meanwhile, with interaction type and pose fixed, our method can further generate more numerous results under different human styles or object styles, as shown in Figure 3 (right). Moreover, our InterFusion also supports controllable text-conditioned editing, providing users more control over the already generated 3D models, which are presented in the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We both qualitatively and quantitatively evaluate our method against alternative baseline methods. To verify the effectiveness of individual components in our method, we conduct ablation studies. Furthermore, we discuss the application potential, limitations and future work of our InterFusion. The implementation details, more evaluations, and further discussions are presented in the supplementary materials." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "text", + "content": "Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 479, + 276 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 479, + 276 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 479, + 276 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 479, + 276 + ], + "type": "image", + "image_path": "2d8c7dc580e85d742db240cc65d5e6505a8946110158bada2f459e393530a939.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 285, + 480, + 308 + ], + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 308 + ], + "type": "text", + "content": "Fig. 3: More results generated by InterFusion. Diverse integration poses are supported. Numerous human and object styles are also supported." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 332, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 536 + ], + "type": "text", + "content": "Evaluation prompts and criteria. Similar as the first stage for generating prompts about various interaction types, we also use ChatGPT to randomly generate prompts about human styles and collect the text prompts for evaluation. We select 61 distinct and diverse prompts, ensuring the classes of Human (e.g., a policewoman, a teenager) with various styles (e.g., in a graphic tee and ripped jeans), Objects (e.g., a motorcycle, a guitar) and Interactions (e.g., riding, playing) are reasonable and evenly distributed. In total, we have 13 different types of interactions, covering contact areas across the whole body, which demonstrates the effectiveness of our method on diverse types of interactions compared to baseline methods. We calculate the CLIP scores between the input text prompts and different views of generated 3D human-object interactions, and then compare the means based on different evaluation prompts. The CLIP score measures the similarity between a prompt text for an image and the actual content of the image. We also provide an assessment using GPT-4V for selection, named GPT-4V select, which evaluates the completeness of objects and correctness of physical interactions across multiview rendered images. We refer to the supplementary material for assessment details and additional evaluation criteria." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Baseline approaches. To the best of our knowledge, the proposed method is among the first to generate 3D human-object interactions based on text inputs in a zero-shot manner. We thus compare our method to alternative text-to-3D methods, including DreamFusion [40], Magic3D [21], and TextMesh [54]. Additional comparisons with MVDream [50] and ProlificDreamer [57] are presented in the supplementary material. To justify our key idea of using human pose to guide the generation, we design an object-centric baseline (Ours-OC) as a variant of our method. Instead of using human pose as the prior, we retrieve an object with the given semantic category from ShapeNet dataset [2] and use it as" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "text", + "content": "InterFusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 136, + 471, + 176 + ], + "blocks": [ + { + "bbox": [ + 166, + 114, + 447, + 126 + ], + "lines": [ + { + "bbox": [ + 166, + 114, + 447, + 126 + ], + "spans": [ + { + "bbox": [ + 166, + 114, + 447, + 126 + ], + "type": "text", + "content": "Table 1: Quantitative evaluation of CLIP score and GPT-4V select." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 136, + 471, + 176 + ], + "lines": [ + { + "bbox": [ + 143, + 136, + 471, + 176 + ], + "spans": [ + { + "bbox": [ + 143, + 136, + 471, + 176 + ], + "type": "table", + "html": "
MethodDreamFusion [40] Magic3D [21] TextMesh [54] Ours-OCOurs-HC
CLIP score0.3027 0.3179 0.2761 0.32030.3308
GPT-4V select(%)8.20 11.48 1.64 13.1165.57
", + "image_path": "c565535e5634524a64053c169f87922ed970c26c1c641f526171b2691c22777b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 198, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 198, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 482, + 259 + ], + "type": "text", + "content": "geometry constraints to guide both human and object generation. As official implementations are unavailable for some of these baselines, we use the third-party re-implementations provided by threestudio [11] for a fair comparison. Note that all re-implementations use multi-resolution hash-grid [33] for 3D representation and DeepFloyd [5] for guidance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 277, + 263, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 263, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 263, + 290 + ], + "type": "text", + "content": "4.1 Comparison Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 299, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 482, + 502 + ], + "type": "text", + "content": "Qualitative evaluations. Some representative visual comparisons are shown in Figure 4, and additional comparisons including those with our object-centric baseline are presented in the supplementary material. The overall results of the baselines reveal common uncertainties in both geometry and appearance attributed to the confusion introduced by multi-concept guidance. Specifically, baseline models may lean towards one specific target, as seen in the example of \"a man wearing a red baseball cap\" with only the red baseball cap generated, the example of \"a policeman\" with only the upper half of the human body generated, or the example of \"a man in a puffer jacket and wool beanie\" where the shopping cart is failed to be generated. Even with a relatively uniform attention distribution, the baseline model may encounter challenges in producing complete results, as demonstrated in the saxophone and violin examples (the 3rd and 4th column in Figure 4), which exhibit deficiencies such as incomplete human body parts, and mixing of human legs. Moreover, the baseline model sometimes struggles to effectively choose the focus of generation when the input text describes a less common target like \"a person in a military uniform\", resulting in poor outcomes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 502, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 552 + ], + "type": "text", + "content": "Our approach overcomes these issues by conducting optimization in an explicit decomposed way and intelligently guiding attention from SDS jointly in spatial and semantic aspects. Therefore, our method achieves more stable and higher-quality 3D results under multiple-concept guidance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Quantitative evaluations. The comparison results are shown in Table 1. We can see that our method (Ours-HC) achieves the best performance compared to baselines, showcasing our results exhibit more 3D plausibility with given text prompts. Among the baselines, Ours-OC gets the best performance. Though with object priors, the object-centric approach still does not provide sufficient body priors for human generation, resulting in its inability to achieve complete interaction generation. Magic3D gets the next best performance among the baselines, as it usually generates the whole interaction scene comparing to other two" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "text", + "content": "Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 118, + 474, + 295 + ], + "blocks": [ + { + "bbox": [ + 138, + 118, + 474, + 295 + ], + "lines": [ + { + "bbox": [ + 138, + 118, + 474, + 295 + ], + "spans": [ + { + "bbox": [ + 138, + 118, + 474, + 295 + ], + "type": "image", + "image_path": "bb4bcc47b4835dbb2190e3855df9e24f2aa58e104660583f8d42fc662c834f1b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 304, + 480, + 327 + ], + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 327 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparison results with baselines. InterFusion generates more stable and higher-quality results and is more consistent with input interaction descriptions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 340, + 478, + 462 + ], + "blocks": [ + { + "bbox": [ + 138, + 340, + 478, + 462 + ], + "lines": [ + { + "bbox": [ + 138, + 340, + 478, + 462 + ], + "spans": [ + { + "bbox": [ + 138, + 340, + 478, + 462 + ], + "type": "image", + "image_path": "d09f2531a0a0aa6b4ea91ea0a7822a369b44105fbe3ec9892cb227a36719aa6e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 470, + 400, + 481 + ], + "lines": [ + { + "bbox": [ + 212, + 470, + 400, + 481 + ], + "spans": [ + { + "bbox": [ + 212, + 470, + 400, + 481 + ], + "type": "text", + "content": "Fig. 5: Qualitative results of ablation studies." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 507, + 480, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 480, + 531 + ], + "type": "text", + "content": "basielines, however, the results are somewhat vague as shown in Figure 4, thus there is still a large performance gap comparing to our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 484, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 484, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 484, + 567 + ], + "type": "text", + "content": "The significant enhancement in performance can be attributed to our method's effective concurrent generation of both human and object, along with cohesive interactive information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 592, + 245, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 592, + 245, + 602 + ], + "spans": [ + { + "bbox": [ + 132, + 592, + 245, + 602 + ], + "type": "text", + "content": "4.2 Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "We conduct several ablation studies both qualitatively and quantitatively to show the importance of our design as well as introduced loss terms. Results are presented in Figure 5a, Figure 5b and Table 2. We refer to the supplementary material for additional results and details." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "text", + "content": "InterFusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 187, + 136, + 427, + 179 + ], + "blocks": [ + { + "bbox": [ + 205, + 114, + 408, + 126 + ], + "lines": [ + { + "bbox": [ + 205, + 114, + 408, + 126 + ], + "spans": [ + { + "bbox": [ + 205, + 114, + 408, + 126 + ], + "type": "text", + "content": "Table 2: Quantitative results of ablation studies." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 136, + 427, + 179 + ], + "lines": [ + { + "bbox": [ + 187, + 136, + 427, + 179 + ], + "spans": [ + { + "bbox": [ + 187, + 136, + 427, + 179 + ], + "type": "table", + "html": "
Settingsw/o ISDSw/o L0SDSw/o L0geoOurs
CLIP score0.31640.32930.31710.3308
GPT-4V select(%)1.644.9216.3977.05
", + "image_path": "58757538265b0785170cf6db4d62ba1ed6a1c22505e3faf2501350105e6d6ced.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 201, + 482, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 310 + ], + "type": "text", + "content": "Syn-HOI-pose. The anchor pose obtained at the first stage provides the geometry constraints for our HOI generation. We thus first demonstrate the effectiveness of synthesizing anchoring poses using our pseudo-pose dataset. We compare the pose generated with the existing approach AvatarCLIP [14], which utilizes CLIP only to retrieve the queried pose from the codebook constructed from the mocap dataset AMASS [26]. Figure 5a shows that the proposed pose generation stage enables a better-fitted anchor pose for input interaction types, which demonstrates our pseudo-poses, reconstructed from synthesized images, has the potential for more diverse poses' requirements than existing mocap datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 327, + 482, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 435 + ], + "type": "text", + "content": "Semantic guidance. To better illustrate the efficacy of semantic guidance in our NeRF-based generation, we conduct ablations separately on SDS from object and interaction. As shown in Figure 5b and Table 2, in the absence of SDS from object, semantic consistency for the object is compromised, resulting in noisy outcomes influenced by the human, thus the contact region can not be extracted correctly. SDS from interaction plays the most crucial role in the performance, making results more semantically and visually plausible. The absence of SDS from interaction leads to generated objects being confined to the remaining space, but without interaction with the human." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "text", + "content": "Geometric constraint. Relying solely on semantic guidance is inadequate for achieving the final objectives, as the object should be generated in spaces outside the human body, with its generation targets not positioned at the origin. As illustrated in Figure 5b and Table 2, depending only on semantic guidance may yield results that appear spatially conflict with the human body. Without the spatial constraint, the object generation at the origin would additionally have a conflict with the interaction objective, thus resulting in degenerated objects and final interactions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 568, + 220, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 220, + 580 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 220, + 580 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In conclusion, our work presents InterFusion, a novel framework for zero-shot 3D human-object interaction generation. InterFusion tackles the challenges of limited 3D interaction data and the complexity of generating multiple concepts simultaneously. Our two-stage approach, which synthesizes 3D interaction poses from text and then uses these poses as geometric anchors for detailed HOI scene generation, has demonstrated significant improvements over existing methods." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "text", + "content": "Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Acknowledgements. We thank Zejia Su and Yunfan Ye for helpful discussions and paper proofreading. We also thank the anonymous reviewers for their valuable comments. This project was supported by the NSFC (62325211, 62132021, 62322207), Guangdong Natural Science Foundation (2021B1515020085), Shenzhen Science and Technology Program (RCYX20210609103121030), and the Major Program of Xiangjiang Laboratory (23XJ01009)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 205, + 198, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 198, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 198, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 228, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 228, + 480, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 273 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 273 + ], + "type": "text", + "content": "1. Bhatnagar, B.L., Xie, X., Petrov, I.A., Sminchisescu, C., Theobalt, C., Pons-Moll, G.: Behave: Dataset and method for tracking human object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15935-15946 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 274, + 480, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 480, + 306 + ], + "type": "text", + "content": "2. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 307, + 480, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 307, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 138, + 307, + 480, + 361 + ], + "type": "text", + "content": "3. Chen, K., Choy, C.B., Savva, M., Chang, A.X., Funkhouser, T., Savarese, S.: Text2shape: Generating shapes from natural language by learning joint embeddings. In: Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part III 14. pp. 100-116. Springer (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 361, + 480, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 480, + 393 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 480, + 393 + ], + "type": "text", + "content": "4. Cohen-Bar, D., Richardson, E., Metzer, G., Giryes, R., Cohen-Or, D.: Set-the-scene: Global-local training for generating controllable nef scenes. arXiv preprint arXiv:2303.13450 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 393, + 369, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 393, + 369, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 393, + 369, + 403 + ], + "type": "text", + "content": "5. Deepfloyd IF. https://github.com/deep-floyd/IF, 2023" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 404, + 480, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 480, + 436 + ], + "type": "text", + "content": "6. Diller, C., Dai, A.: Cg-hoi: Contact-guided 3d human-object interaction generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19888-19901 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 437, + 480, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 480, + 469 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 480, + 469 + ], + "type": "text", + "content": "7. Fan, Z., Taheri, O., Tzionas, D., Kocabas, M., Kaufmann, M., Black, M.J., Hilliges, O.: Articulated objects in free-form hand interaction. arXiv preprint arXiv:2204.13662 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 470, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 470, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 470, + 480, + 502 + ], + "type": "text", + "content": "8. Feng, Y., Choutas, V., Bolkart, T., Tzionas, D., Black, M.J.: Collaborative regression of expressive bodies using moderation. In: 2021 International Conference on 3D Vision (3DV). pp. 792-804. IEEE (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 502, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 534 + ], + "type": "text", + "content": "9. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5152-5161 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 535, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 535, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 138, + 535, + 480, + 556 + ], + "type": "text", + "content": "0. Guo, M., Fathi, A., Wu, J., Funkhouser, T.: Object-centric neural scene rendering. arXiv preprint arXiv:2012.08503 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 557, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 557, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 134, + 557, + 480, + 600 + ], + "type": "text", + "content": "11. Guo, Y.C., Liu, Y.T., Shao, R., Laforte, C., Voleti, V., Luo, G., Chen, C.H., Zou, Z.X., Wang, C., Cao, Y.P., Zhang, S.H.: threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "text", + "content": "12. Hassan, M., Choutas, V., Tzionas, D., Black, M.J.: Resolving 3d human pose ambiguities with 3d scene constraints. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2282-2292 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "13. Hassan, M., Ghosh, P., Tesch, J., Tzionas, D., Black, M.J.: Populating 3d scenes by learning human-scene interaction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14708-14718 (2021)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "text", + "content": "InterFusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "14. Hong, F., Zhang, M., Pan, L., Cai, Z., Yang, L., Liu, Z.: Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "15. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence 36(7), 1325-1339 (2013)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 195, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 480, + 228 + ], + "type": "text", + "content": "16. Jahan, T., Guan, Y., Van Kaick, O.: Semantics-guided latent space exploration for shape generation. In: Computer Graphics Forum. vol. 40, pp. 115-126. Wiley Online Library (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 229, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 480, + 262 + ], + "type": "text", + "content": "17. Jain, A., Mildenhall, B., Barron, J.T., Abbeel, P., Poole, B.: Zero-shot text-guided object generation with dream fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 867-876 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 480, + 285 + ], + "type": "text", + "content": "18. Kim, V.G., Chaudhuri, S., Guibas, L., Funkhouser, T.: Shape2Pose: Human-centric shape analysis. ACM Transactions on Graphics (Proc. SIGGRAPH) (Aug 2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 480, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 306 + ], + "type": "text", + "content": "19. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 480, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 329 + ], + "type": "text", + "content": "20. Lee, H.H., Chang, A.X.: Understanding pure clip guidance for voxel grid nerf models. arXiv preprint arXiv:2209.15172 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "type": "text", + "content": "21. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "text", + "content": "22. Lin, Y., Bai, H., Li, S., Lu, H., Lin, X., Xiong, H., Wang, L.: Componerf: Text-guided multi-object compositional nerf with editable 3d scene layout. arXiv preprint arXiv:2303.13843 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "type": "text", + "content": "23. Liu, N., Li, S., Du, Y., Torralba, A., Tenenbaum, J.B.: Compositional visual generation with composable diffusion models. In: European Conference on Computer Vision. pp. 423-439. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 431, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 480, + 463 + ], + "type": "text", + "content": "24. Liu, Z., Wang, Y., Qi, X., Fu, C.W.: Towards implicit text-guided 3d shape generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17896-17906 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 464, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 497 + ], + "type": "text", + "content": "25. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Spl: A skinned multi-person linear model. ACM transactions on graphics (TOG) 34(6), 1-16 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "type": "text", + "content": "26. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5442-5451 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 532, + 480, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 480, + 563 + ], + "type": "text", + "content": "27. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latentnerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "type": "text", + "content": "28. Michel, O., Bar-On, R., Liu, R., Benaim, S., Hanocka, R.: Text2mesh: Text-driven neural stylization for meshes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13492-13502 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 599, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 631 + ], + "type": "text", + "content": "29. Mihajlovic, M., Saito, S., Bansal, A., Zollhoefer, M., Tang, S.: Coap: Compositional articulated occupancy of people. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13201-13210 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "30. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "31. Mirzaei, A., Kant, Y., Kelly, J., Gilitschenski, I.: Laterf: Label and text driven object radiance fields. In: European Conference on Computer Vision. pp. 20-36. Springer (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "type": "text", + "content": "32. Mohammad Khalid, N., Xie, T., Belilovsky, E., Popa, T.: Clip-mesh: Generating textured meshes from text using pretrained image-text models. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-8 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "type": "text", + "content": "33. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "type": "text", + "content": "34. Nguyen-Phuoc, T.H., Richardt, C., Mai, L., Yang, Y., Mitra, N.: Blockgan: Learning 3d object-aware scene representations from unlabelled images. Advances in neural information processing systems 33, 6767-6778 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 481, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 285 + ], + "type": "text", + "content": "35. Niemeyer, M., Geiger, A.: Giraffe: Representing scenes as compositional generative neural feature fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11453-11464 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "type": "text", + "content": "36. Ost, J., Mannan, F., Thuerey, N., Knodt, J., Heide, F.: Neural scene graphs for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2856-2865 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 319, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 363 + ], + "type": "text", + "content": "37. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10975-10985 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "type": "text", + "content": "38. Peng, X., Xie, Y., Wu, Z., Jampani, V., Sun, D., Jiang, H.: Hoi-diff: Text-driven synthesis of 3d human-object interactions using diffusion models. arXiv preprint arXiv:2312.06553 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "type": "text", + "content": "39. Po, R., Wetzstein, G.: Compositional 3d scene generation using locally conditioned diffusion. arXiv preprint arXiv:2303.12218 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "type": "text", + "content": "40. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "text", + "content": "41. Puig, X., Undersander, E., Szot, A., Cote, M.D., Yang, T.Y., Partsey, R., Desai, R., Clegg, A.W., Hlavac, M., Min, S.Y., et al.: Habitat 3.0: A co-habitat for humans, avatars and robots. arXiv preprint arXiv:2310.13724 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 476, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 481, + 520 + ], + "type": "text", + "content": "42. Punnakkal, A.R., Chandrasekaran, A., Athanasiou, N., Quiros-Ramirez, A., Black, M.J.: Babel: Bodies, action and behavior with english labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 722-731 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 521, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 564 + ], + "type": "text", + "content": "43. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 565, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 598 + ], + "type": "text", + "content": "44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 599, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 620 + ], + "type": "text", + "content": "45. Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "46. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "text", + "content": "InterFusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "47. Savva, M., Chang, A.X., Hanrahan, P., Fisher, M., Nießner, M.: Pigraphs: learning interaction snapshots from observations. ACM Transactions on Graphics (TOG) 35(4), 1-12 (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 151, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 151, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 151, + 480, + 172 + ], + "type": "text", + "content": "48. Sella, E., Fiebelman, G., Hedman, P., Averbuch-Elor, H.: Vox-e: Text-guided voxel editing of 3d objects. arXiv preprint arXiv:2303.12048 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 174, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 174, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 174, + 480, + 194 + ], + "type": "text", + "content": "49. Sheridan, T.B.: Human-robot interaction: status and challenges. Human factors 58(4), 525-532 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 197, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 197, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 197, + 480, + 217 + ], + "type": "text", + "content": "50. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 219, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 219, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 219, + 480, + 251 + ], + "type": "text", + "content": "51. Sohn, K., Lee, H., Yan, X.: Learning structured output representation using deep conditional generative models. Advances in neural information processing systems 28 (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 252, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 252, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 480, + 274 + ], + "type": "text", + "content": "52. Song, Y., Kong, C., Lee, S., Kwak, N., Lee, J.: Towards efficient neural scene graphs by learning consistency fields. arXiv preprint arXiv:2210.04127 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 275, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 275, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 275, + 480, + 319 + ], + "type": "text", + "content": "53. Taheri, O., Ghorbani, N., Black, M.J., Tzionas, D.: Grab: A dataset of whole-body human grasping of objects. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 581-600. Springer (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 320, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 480, + 352 + ], + "type": "text", + "content": "54. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 354, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 396 + ], + "type": "text", + "content": "55. Von Marcard, T., Henschel, R., Black, M.J., Rosenhahn, B., Pons-Moll, G.: Recovering accurate 3d human pose in the wild using imus and a moving camera. In: Proceedings of the European conference on computer vision (ECCV). pp. 601-617 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 398, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 480, + 430 + ], + "type": "text", + "content": "56. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 432, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 432, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 432, + 480, + 464 + ], + "type": "text", + "content": "57. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 465, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 465, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 465, + 480, + 497 + ], + "type": "text", + "content": "58. Wu, Q., Shi, Y., Huang, X., Yu, J., Xu, L., Wang, J.: Thor: Text to human-object interaction diffusion via relation intervention. arXiv preprint arXiv:2403.11208 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 498, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 531 + ], + "type": "text", + "content": "59. Wu, Q., Liu, X., Chen, Y., Li, K., Zheng, C., Cai, J., Zheng, J.: Object-compositional neural implicit surfaces. In: European Conference on Computer Vision. pp. 197-213. Springer (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 533, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 586 + ], + "type": "text", + "content": "60. Xu, Y., Chai, M., Shi, Z., Peng, S., Skorokhodov, I., Siarohin, A., Yang, C., Shen, Y., Lee, H.Y., Zhou, B., et al.: Discoscope: Spatially disentangled generative radiance fields for controllable 3d-aware scene synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4402-4412 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 588, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 632 + ], + "type": "text", + "content": "61. Yang, B., Zhang, Y., Xu, Y., Li, Y., Zhou, H., Bao, H., Zhang, G., Cui, Z.: Learning object-compositional neural radiance field for editable scene rendering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13779-13788 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "62. Zhang, S., Zhang, Y., Ma, Q., Black, M.J., Tang, S.: Place: Proximity learning of articulation and contact in 3d environments. In: 2020 International Conference on 3D Vision (3DV). pp. 642-651. IEEE (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "text", + "content": "Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 216 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "63. Zhang, Y., Hassan, M., Neumann, H., Black, M.J., Tang, S.: Generating 3d people in scenes without people. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6194-6204 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "type": "text", + "content": "64. Zhao, K., Wang, S., Zhang, Y., Beeler, T., Tang, S.: Compositional human-scene interaction synthesis with semantic control. In: European Conference on Computer Vision. pp. 311-327. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 482, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 482, + 216 + ], + "type": "text", + "content": "65. Zhi, S., Laidlow, T., Leutenegger, S., Davison, A.J.: In-place scene labelling and understanding with implicit scene representation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15838-15847 (2021)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 399, + 91, + 447, + 100 + ], + "type": "text", + "content": "InterFusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition/0a79d390-9d09-4042-af06-6b62e335816f_content_list.json b/2024/Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition/0a79d390-9d09-4042-af06-6b62e335816f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7ead527cc970a252373a3760a854724b85aaeb35 --- /dev/null +++ b/2024/Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition/0a79d390-9d09-4042-af06-6b62e335816f_content_list.json @@ -0,0 +1,1638 @@ +[ + { + "type": "text", + "text": "Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition", + "text_level": 1, + "bbox": [ + 295, + 140, + 709, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yisong Wang $^{1*}$ , Nan Xi $^{2*}$ , Jingjing Meng $^{3}$ , and Junsong Yuan $^{2}$", + "bbox": [ + 241, + 243, + 759, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of EECS, Peking University", + "bbox": [ + 375, + 270, + 627, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Computer Science and Engineering, University at Buffalo", + "bbox": [ + 250, + 285, + 750, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Amazon Inc.", + "bbox": [ + 450, + 299, + 552, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wys05y98@gmail.com, {nanxi, jsyuany}@buffalo.edu, jingjing.meng1@gmail.com", + "bbox": [ + 217, + 314, + 784, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Understanding human-object interaction (HOI) in videos represents a fundamental yet intricate challenge in computer vision, requiring perception and reasoning across both spatial and temporal domains. Despite previous success of object detection and tracking, multi-person video HOI recognition still faces two major challenges: (1) the three facets of HOI (human, objects, and the interactions that bind them) exhibit interconnectedness and exert mutual influence upon one another. (2) the complexity of multi-person multi-object combinations in spatio-temporal interaction. To address them, we design a spatio-temporal context fuser to better model the interactions among persons and objects in videos. Furthermore, to equip the model with temporal reasoning capacity, we propose an interaction state reasoner module on top of context fuser. Considering the interaction is a key element to bind human and object, we propose an interaction-centric hypersphere in the feature embedding space to model each category of interaction. It helps to learn the distribution of HOI samples belonging to the same interactions on the hypersphere. After training, each interaction prototype sphere will fit the testing HOI sample to determine the HOI classification result. Empirical results on multi-person video HOI dataset MPHOI-72 indicate that our method remarkably surpasses state-of-the-art (SOTA) method by more than $22\\%$ $F_{1}$ score. At the same time, on single-person datasets Bimanual Actions (single-human two-hand HOI) and CAD-120 (single-human HOI), our method achieves on par or even better results compared with SOTA methods. Source code is released at the following link: https://github.com/southnx/IcH-Vid-HOI.", + "bbox": [ + 259, + 391, + 743, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: interaction-centric $\\cdot$ spatio-temporal context reasoning $\\cdot$ multi-person human-object interaction", + "bbox": [ + 259, + 753, + 756, + 781 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution; ≔ Corresponding author", + "bbox": [ + 232, + 824, + 545, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4b34982c1398083e87608267fe526441237b01c8e66eac06865a957e2a210330.jpg", + "image_caption": [ + "Multi-person Video Human-Object Interaction" + ], + "image_footnote": [], + "bbox": [ + 250, + 161, + 408, + 213 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1f3ae3a97705067a470cd5512afbe464c84545b57759bc5c4d7834d8278d6108.jpg", + "image_caption": [ + "Fig. 1: Examples of multi-person video HOI. Three categories of interactions are presented, including co-working, cheering, hair_cutting." + ], + "image_footnote": [], + "bbox": [ + 421, + 161, + 580, + 213 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3048e4587dd2851d296f2750bfb7e5b0b660ac7cf162769eba993122e9f1f017.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 161, + 751, + 213 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 280, + 375, + 296 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video-based Human-Object Interaction (HOI) recognition aims to identify the interactions occurring between human and object entities within video frames. Precisely recognizing HOIs in real world scenarios is essential for a bunch of applications, such as assisting patients by recognizing daily activities and predicting pedestrian movements to avoid accidents for autonomous vehicles.", + "bbox": [ + 212, + 311, + 785, + 385 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most existing HOI recognition research focuses on static images [19, 36], with much less attention on video-based HOI recognition. Video-based HOI recognition is more demanding than image-based HOI recognition due to the necessity of comprehending complex spatio-temporal dynamics and reasoning about multi-human and multi-object interaction combinations. The complexity is further exacerbated when dealing with multi-person circumstances, as examples shown in Fig. 1. In such cases, various human and object entities mutually influence each other, resulting in intricate interdependencies within the scene. Additionally, the three components (human, object and interactions that the entities are involved in) of HOI exhibit interwind structures, e.g., the possible interactions that can occur within a scene given a human and an object. However, current video HOI recognition methods [23, 27] do not fully explore such inherent structural nature of HOI components. Instead, they often opt for disentangled representations for each component, which may bring suboptimal representation capabilities.", + "bbox": [ + 212, + 387, + 785, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the aforementioned limitations, we introduce an interaction-centric spatio-temporal context reasoning approach for representing video HOIs. For the purpose of enhancing the awareness of complex HOI structures in our representations, we introduce the Context Fuser (CF), which encodes both entity representations and interaction representations. Moreover, to empower our model with the ability to reason about interaction state transitions across video frames, we propose the Interaction State Reasoner (ISR) module for generating interaction representations. In addition, we employ a bidirectional Gated Recurrent Unit (BiGRU) to model temporal dynamics across video frames. This multi-level representation learning framework not only facilitates effective exploration of the interdependencies among structured HOI components but also empowers the model with interaction reasoning capabilities in both spatial and temporal domains. To compute the probability of an HOI in the scene, we introduce an interaction-centric hypersphere. This approach leverages the concept of a hypersphere, where the interaction is located at its center, while the human and object entities involved in that interaction are situated on the hypersphere's", + "bbox": [ + 212, + 599, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Wang et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "surface. We assume that human-object entities belonging to each interaction class are located on their respective hyperspheres, specific to that interaction class.", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concretely, we generate context-rich and reasoning-aware video HOI representations through three key components: the Context Fuser (CF), Interaction State Reasoner (ISR) and Bidirectional GRU (BiGRU). The CF module integrates context information from human-object entities and interactions. It comprises three fuser blocks for humans, objects, and interactions. The Object Fuser Block processes local video frame data, enhancing object features. The Interaction Fuser Block combines human and object representations with interaction-specific characteristics. Additionally, in multi-person scenes, the Human Fuser Block captures human representations influenced by others. This approach fosters comprehensive HOI representations via effective context fusion. To facilitate interaction reasoning, we place the ISR module on top of the context fuser module, yielding entity representations capable of capturing interaction transition dynamics. These entity representations are then input into the BiGRU module to model temporal dynamics across video frames, thereby ensuring a comprehensive understanding of the evolving context and interactions within the video data. Finally, we determine interaction classes in each frame with the interaction-centric hypersphere, computing the probability of human-object entities belonging to specific interaction classes.", + "bbox": [ + 212, + 191, + 787, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We assess our model's performance on three video-based HOI datasets: MPHOI-72 [27] (multi-person), Bimanual Actions [4] and CAD-12 [12] (single-person). Our results highlight our model's superiority in multi-person scenarios, achieving an impressive over $22\\%$ $F_{1}$ score improvement over the current state-of-the-art (SOTA). In single-person scenarios, our method delivers on par or even better performance compared to the current SOTA method. Our major contributions are summarized as follows:", + "bbox": [ + 212, + 463, + 799, + 568 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To represent inherent HOI manifold structures, we propose an interaction-centric hypersphere representation scheme. This scheme explicitly guide the process of learning intrinsic structural nature of HOI and elucidates the interdependencies among its components.", + "- To learn context-rich and reasoning-aware entity representations, we introduce context fuser and interaction state reasoning modules. This enhancement results in entity representations that are highly suitable for video-based HOI tasks.", + "- Extensive experiment results showcase that our method achieves SOTA performance with a huge improvement of more than $22\\%$ $F_{1}$ score over existing methods in multi-person scenario. Additionally, our model achieves competitive results in single-person cases compared to SOTA method." + ], + "bbox": [ + 225, + 577, + 785, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related works", + "text_level": 1, + "bbox": [ + 215, + 779, + 392, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "HOI detection in images: HOI detection in images aims at understanding interactions in images between humans and objects. Different methods have been", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Video HOI Interaction-centric context reasoning", + "bbox": [ + 408, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "proposed in previous studies. Some works propose Convolutional Neural Networks (CNN)-based methods which can be further divided into one-stage methods [9,17,37] and two-stage methods [5,7,16,33]. However, these methods usually lack of ability to capture global context information. Recently, Transformer-based models [8,10,31,36] became the main approach for the HOI task. Following the architecture of DETR, these models achieved superior performance on HOI detection. Moreover, some works also utilize graph [25] and interactiveness field [19] to achieve better performance. These various approaches to image HOI provide the fundamentals for video HOI recognition.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "HOI recognition in videos: Video HOI recognition is the foundation of many real-world applications, including understanding surgical activities in the operation room [34] and guiding 3D human reconstruction [20, 21]. Video-based HOI recognitions have to deal with both spatial and temporal reasoning. Before the use of neural networks, some early studies formulated this task using the Markov model [12] to utilize temporal cues. In [24], HOI hotspots in videos are learned in a novel approach, with two networks trained jointly to capture spatial regions where actions happen. While [6] focuses on visual relation prediction in open-vocabulary with pretrained visual-language models. Recent works have used Recurrent Neural Networks (RNN) combined with Graph Neural Networks (GNN) [23, 26, 27, 30] to predict human-object relations in videos. Inspired by ViT [3], some works also propose Transformer-based methods to reason spatial relations better [32]. However, RNN-based models usually require complex training strategies or long training time in order to achieve the best performance. Moreover, when multiple persons are involved in an activity jointly, these methods lack the ability to model their collaboration, resulting in poor performance when the interactions are performed by multiple persons.", + "bbox": [ + 212, + 282, + 787, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hyperspheres for class representation: Hyperspheres have been demonstrated to be an effective approach to model class representation [2, 22]. Geometrical modeling strategy has been proposed in [2], where the effectiveness has been confirmed. This approach proves advantageous for capturing and representing enriched class-level information, particularly well-suited for creating measurements in Euclidean space. Consequently, it is naturally adaptable to structured prediction tasks.", + "bbox": [ + 212, + 539, + 787, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Motivation", + "text_level": 1, + "bbox": [ + 215, + 670, + 359, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video HOI recognition task involves identifying both the human and object entities engaged in an interaction across a sequence of video frames. This task encompasses spatial and temporal aspects, as it requires understanding the relationships between humans, objects, and their interactions over time. However, current methods for video HOI recognition often neglect this crucial dependency structure, resulting in the separation of learned representations associated with humans, objects, and their interactions, ultimately compromising their representational accuracy. To address this issue, we propose a novel approach featuring a context fuser and an interaction state reasoner to capture spatio-temporal", + "bbox": [ + 212, + 703, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Wang et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4a17d980bda599dd9d63d0945d0cbe99e4d88b1839ba6ff64e60f5fdafcea4fd.jpg", + "image_caption": [ + "Fig. 2: Model Overview. Each video frame is applied with a backbone for feature extraction. Subsequently, a context fuser and interaction state reasoner is employed for learning interaction representations $z_{I}$ and human-object entity representations $\\{\\hat{z}_E^t\\}_{t=1}^T$ . Bidirectional GRU is further utilized to model temporal dynamics across video frames. The final predicted interaction probability $\\{u^t\\}_{t=1}^T$ is computed from the interaction-centric hypersphere. The dimension of each hypersphere is $d$ , which is the same with the feature dimension in our model." + ], + "image_footnote": [], + "bbox": [ + 230, + 141, + 774, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "contextual representations. Additionally, we introduce an interaction-centric hypersphere to represent HOI manifold structure and model the inter-dependency among predicted humans, objects, and interactions.", + "bbox": [ + 212, + 575, + 787, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 645, + 330, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Problem formulation", + "text_level": 1, + "bbox": [ + 215, + 676, + 433, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a video dataset $\\mathcal{V}$ , given a video clip $V\\in \\mathcal{V}$ containing $T$ video frames $\\{v_{1},\\dots,v_{T}\\}$ , video HOI recognition aims to predict the temporal segmentation of interactions between human and object entities across all the video frames. Formally, we aim to learn an HOI recognition model $\\mathcal{M}$ that outputs the segmentation of human's sub-activity $\\{s_n\\}_{n = 1}^N$ in each frame, where $N$ is the number of human sub-activity segments. Each segment $s_n$ is represented as an interval from its start time $t_n$ to end time $t_{n + 1}$ : $s_n = [t_n,t_{n + 1})$ . The start and end time of each segment are determined from interaction probability prediction $\\{u^t\\}_{t = 1}^T$ of each frame, where $u^{t}\\in \\mathbb{R}^{K}$ , $K$ is the number of possible interaction classes.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Video HOI Interaction-centric context reasoning", + "bbox": [ + 408, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2 Model Design", + "text_level": 1, + "bbox": [ + 215, + 146, + 377, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the following section, we introduce our interaction-centric hypersphere reasoning model for video HOI recognition in detail. As shown in Fig. 2, our major idea is to construct a hypersphere to represent each HOI in the scene. For each hypersphere, the interaction locates at the center of the hypersphere, while the corresponding human-object entity belongs to that interaction locates at the surface of the hypersphere. We construct Context Fuser (CF) module to learn context-rich human-object entity representations. For the aim of enabling model with reasoning ability over interaction state transitions, we propose Interaction State Reasoner (ISR) to reason on whether the current interaction will be continued or stopped. To model the temporal dynamics of HOI in videos, we update human-object entity representations $\\{\\hat{z}_E^t\\}_{t=1}^T$ along the temporal domain with bidirectional GRU (BiGRU). Finally, predicted interaction class probability $u^t$ of each frame is computed from the interaction-centric hypersphere.", + "bbox": [ + 212, + 169, + 787, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Context fuser For a sequence of video frames $\\{v_{t}\\}_{t = 1}^{T}$ , we follow 2G-GCN [27] to extract feature of humans and objects from backbone network. The extracted human features $z_{H}\\in \\mathbb{R}^{d}$ contain both bounding box information and skeleton keypoint information, where $d$ indicates the feature dimension. Object features $z_{O}\\in \\mathbb{R}^{d}$ contain only bounding box information.", + "bbox": [ + 212, + 385, + 785, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We design a context fuser (CF) module shown in Fig. 3 to generate human-object entity representations $\\{z_E^t\\}_{t=1}^T$ based on human, object and contextual information. In multi-person circumstances, CF contains object fuser block, interaction fuser block and human fuser block sequentially. First, we design an object fuser block to incorporate object representations in local temporal regions into current object representations, generating learned object representation $\\hat{z}_{O_1}$ and $\\hat{z}_{O_2}$ :", + "bbox": [ + 212, + 460, + 787, + 568 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {O _ {1}} = F F N \\left(S A \\left(Q = z _ {O _ {1}} ^ {t}; K, V = z _ {O _ {1}} ^ {\\bar {t}}\\right) + z _ {O _ {1}} ^ {t}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 575, + 785, + 603 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {O _ {2}} = F F N (S A (Q = z _ {O _ {2}} ^ {t}; K, V = z _ {O _ {2}} ^ {t}) + z _ {O _ {2}} ^ {t}),\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 599, + 668, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $SA$ indicates self-attention, $FFN$ is feed forward network, $z_{O_1}^t \\in \\mathbb{R}^d$ and $z_{O_2}^t \\in \\mathbb{R}^d$ are the initial object features, while $z_{O_1}^{\\bar{t}} \\in \\mathbb{R}^{20d}$ and $z_{O_2}^{\\bar{t}} \\in \\mathbb{R}^{20d}$ are stacked object features from a local time window, $\\bar{t} \\in [t - 10,t + 10)$ . Subsequently, for all the $K$ possible interactions $\\{I_i\\}_{i = 1}^K$ as shown in Fig. 3, each interaction class $I_i$ is prompted as a sentence $s =$ \"The human is [interact]ing in the scene.\", where [interact] indicates the specific interaction class. Then the sentence is applied with the text encoder $(\\mathcal{F}_T)$ of large-scale vision-language model CLIP [28] to initialize the interaction feature $z_I = \\mathcal{F}_T(s) \\in \\mathbb{R}^d$ . We also generate context feature $z_C \\in \\mathbb{R}^d$ to represent the semantic information of each frame. Specifically, we extract frame caption $c_i$ from video frame $i(i = 1,\\dots,T)$ with BLIP [15] model and apply the caption with CLIP model to extract text embedding. In order to adapt the human $(z_{H_1},z_{H_2})$ features, object $(z_{O_1},z_{O_2})$ features and context feature $(z_C)$ to the specific interaction feature $z_I$ , we construct an interaction fuser block that contains a cross-attention (CA) module", + "bbox": [ + 212, + 625, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Wang et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a86e4cc1836c9ef67305448c9de403998b65e0af9da4a47905a32c0086c54aff.jpg", + "image_caption": [ + "Fig.3: Context fuser. In multi-person scenario (left), the context fuser consists of object fuser block, interaction fuser block and human fuser block. In single-person case (right), only object fuser block and interaction fuser block are employed. Detailed descriptions of the architecture and annotations can be found in Sec. 4.2." + ], + "image_footnote": [], + "bbox": [ + 230, + 143, + 594, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/924c78e11e15a31f724848add983b0b0814c26e3bd70240c942f225ab99bc0c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 143, + 772, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "followed by a Feed Forward Network (FFN), generating interaction-aware human features $(z_{H_1}' \\in \\mathbb{R}^d, z_{H_2}' \\in \\mathbb{R}^d)$ , object features $(z_{O_1}' \\in \\mathbb{R}^d, z_{O_2}' \\in \\mathbb{R}^d)$ and context feature $(z_C' \\in \\mathbb{R}^d)$ . Furthermore, to model the influence between the two humans in multi-person scenarios, we construct a human fuser block, featuring the same architecture with interaction fuser block. For the first human (Human1), the updated human feature $\\hat{z}_{H_1} \\in \\mathbb{R}^d$ is generated as:", + "bbox": [ + 212, + 505, + 787, + 599 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {H _ {1}} = F F N (C A (Q = S A (z _ {H _ {1}} ^ {\\prime}); K, V = z _ {H _ {2}} ^ {\\prime}) + S A (z _ {H _ {1}} ^ {\\prime})). \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 292, + 609, + 785, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human2 feature $\\hat{z}_{H_2} \\in \\mathbb{R}^d$ is generated in the same way as Eq. 2. Finally, the human-object entity representation $z_{E_1} \\in \\mathbb{R}^d$ of Human1 is computed by max-pooling operation over all the $d$ dimensions of the four representations shown in Eq. 3:", + "bbox": [ + 212, + 638, + 787, + 699 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nz _ {E _ {1}} = M a x P o o l (\\hat {z} _ {H _ {1}}, z _ {O _ {1}} ^ {\\prime}, z _ {O _ {2}} ^ {\\prime}, z _ {C} ^ {\\prime}). \\qquad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 699, + 785, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The human-object entity representation $z_{E_2} \\in \\mathbb{R}^d$ of Human2 is computed with similar approach as Eq. 3. The CF module for single-person cases are similar with multi-person, except that the human fuser block is removed and there is only one human feature as query to be fed into the interaction fuser block.", + "bbox": [ + 212, + 724, + 787, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Interaction State Reasoner To augment the model's ability in interaction state transition reasoning, we introduce an Interaction State Reasoner (ISR)", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Video HOI Interaction-centric context reasoning", + "bbox": [ + 408, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "module following CF module. ISR module explicitly empowers the model to determine whether the current interaction should persist or transit to another interaction. Specifically, as shown in Fig. 4, at each time $t$ , the two possible states state1 and state2 represent \"continue\" or \"stop\" of an interaction, respectively. Each state is prompted as one sentence, where $s_1 =$ \"This interaction is going to continue.\" and $s_2 =$ \"This interaction is going to stop and change to another interaction.\" Then the embeddings of the two states $z_{state_1}$ and $z_{state_2}$ are generated from CLIP [28] text encoder $\\mathcal{F}_T$ : $z_{state_1} = \\mathcal{F}_T(s_1) \\in \\mathbb{R}^d$ , $z_{state_2} = \\mathcal{F}_T(s_2) \\in \\mathbb{R}^d$ . Interaction state embeddings $z_{state_1}$ and $z_{state_2}$ are further fed to a reasoner block (shown in Fig. 4) together with the interaction embedding $z_t^t \\in \\mathbb{R}^d$ at time $t$ , generating state-informed interaction embeddings $\\hat{z}_{state_1}$ and $\\hat{z}_{state_2}$ . The reasoner block contains a FFN and a State Interpolation (SI) module. The SI module generates the weights $\\omega_1, \\omega_2$ for the two interaction states in the following approach:", + "bbox": [ + 212, + 146, + 787, + 358 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {1}, \\omega_ {2} = S o f t m a x (F F N (z _ {E} ^ {t}) \\cdot [ z _ {s t a t e _ {1}} ^ {\\top}, z _ {s t a t e _ {2}} ^ {\\top} ]), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 364, + 785, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\top$ indicates transpose operation. Subsequently, the final human-object entity representation $\\hat{z}_E^t\\in \\mathbb{R}^d$ at time $t$ is generated by interpolate over the current entity representation $z_{E}^{t}$ at time $t$ and the entity representation $z_{E}^{t - 1}$ at time $t - 1$ :", + "bbox": [ + 212, + 393, + 787, + 450 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {E} ^ {t} = \\omega_ {1} \\cdot z _ {E} ^ {t - 1} + \\omega_ {2} \\cdot z _ {E} ^ {t}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 452, + 785, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Consequently, the generated human-object entity representation $\\hat{z}_E^t$ is able to reason on the possible future interaction transitions.", + "bbox": [ + 212, + 474, + 787, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Interaction-centric Hypersphere With the above generated human-object entity representation $\\hat{z}_E^t$ and interaction representation $z_I$ , we need to calculate the probability of a human-object entity $E$ categorizing into the interaction class $I_i$ . To that end, we design an interaction-centric hypersphere with interaction at the center of hypersphere and human-object entity at its surface. We call it \"hypersphere\" because this is a sphere in the $\\mathbb{R}^d$ space, where $d$ is the feature dimension in our model, which is obviously larger than 3. This hypersphere design models the manifold structure of HOI, which is in charge of the specific interaction class. Concretely, we employ a hyperspherical measurement:", + "bbox": [ + 212, + 540, + 609, + 737 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ae395d494ebb19a6d17892720c327e86089036fbc4fdea5550decaa21338a7eb.jpg", + "image_caption": [ + "Fig. 4: Interaction State Reasoner. $\\odot$ indicates inner product." + ], + "image_footnote": [], + "bbox": [ + 638, + 535, + 767, + 676 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {U} (E, I _ {i}) = \\frac {\\exp \\left(- \\left[ | | z _ {I _ {i}} - \\hat {z} _ {E} | | _ {2} - \\lambda \\right] _ {+}\\right)}{\\sum_ {j = 1} ^ {K} \\exp \\left(- \\left[ | | z _ {I _ {j}} - \\hat {z} _ {E} | | _ {2} - \\lambda \\right] _ {+}\\right)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 744, + 785, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $z_{I_i}$ and $z_{I_j}$ indicate the interaction representations of interaction class $i$ and $j$ , respectively. $\\hat{z}_E$ denotes the human-object entity representation. $[s]_+ \\triangleq$", + "bbox": [ + 212, + 806, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Wang et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$max(0,s)$ . $\\lambda$ indicates the radius of the hypersphere, which is set to be a constant in our model. The higher value of $\\mathcal{U}(E,I_i)$ suggests that the human and object entity $E$ is more likely to be categorized into $I_{i}$ .", + "bbox": [ + 212, + 146, + 787, + 194 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Learning Objective", + "text_level": 1, + "bbox": [ + 215, + 210, + 421, + 227 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The learning objective of our model contains two parts: (i) focal loss $\\mathcal{L}_{cls}$ for interaction classification; (ii) interaction feature loss $\\mathcal{L}_{fea}$ that controls the smoothness of interaction features in local region.", + "bbox": [ + 212, + 232, + 785, + 277 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Focal loss $\\mathcal{L}_{cls}$ : We employ focal loss [18] for interaction classification, mitigating the interaction class imbalance problem on model performance. For each video frame $v_{i}(i = 1,\\dots,T)$ , our model predicts the probability $\\hat{y}_i\\in \\mathbb{R}^K$ of all the interaction classes. The corresponding ground-truth of interaction class $y_{i}\\in \\mathbb{R}^{K}$ is a binary vector. For each intraction class $k$ , the focal loss $\\mathcal{L}_{cls}^{k}$ is formulated as: $\\mathcal{L}_{cls}^{k} = -(1 - p_{k})^{\\gamma}\\log (p_{k})$ , where $\\gamma$ is a hyperparameter to control the focusing extent, $p_k$ is defined as: $\\{p_k = \\hat{y}_i^k$ , if $y_i^k = 1;p_k = 1 - \\hat{y}_i^k$ , otherwise\\}. Subsequently, the focal loss $\\mathcal{L}_{cls}$ of each video frame is obtained by combining the focal loss of each individual interaction class $k$ : $\\mathcal{L}_{cls} = \\Sigma_{k = 1}^{K}\\mathcal{L}_{cls}^{k}$ .", + "bbox": [ + 212, + 279, + 785, + 415 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Interaction feature loss $\\mathcal{L}_{fea}$ : We introduce interaction feature loss $\\mathcal{L}_{fea}$ to control the temporal smoothness of interaction features. Our model outputs the feature of each human-object entity $E$ in each frame. Inspired by [1], in order to improve the continuity, we minimize the feature distance in the same segment and maximize the distance between different segments for each subject. Denote $u_{E}^{t}$ as whether the interaction will continue or change to another action for entity $E$ at time $t$ . $u_{E}^{t} = 1$ indicates the interaction will stop and change to another action at time $t$ for entity $E$ and $u_{E}^{t} = 0$ otherwise. We minimize", + "bbox": [ + 212, + 415, + 785, + 537 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f e a} = \\frac {1}{2} \\sum_ {t = 0} ^ {T - 1} \\left[ (1 - u _ {E} ^ {t}) \\left(\\left| \\left| z _ {E} ^ {t} - z _ {E} ^ {t + 1} \\right| \\right| _ {2}\\right) ^ {2} + u _ {E} ^ {t} \\left(\\max \\left(L - \\left| \\left| z _ {E} ^ {t} - z _ {E} ^ {t + 1} \\right| \\right| _ {2}, 0\\right)\\right) ^ {2} \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 542, + 785, + 584 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $L$ is a threshold that controls the minimal feature distance when interaction changes. In total, the overall loss is written as:", + "bbox": [ + 214, + 590, + 785, + 621 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c l s} + \\alpha \\mathcal {L} _ {f e a}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 433, + 628, + 785, + 646 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\alpha$ is a hyperparameter to control the weight of each loss.", + "bbox": [ + 214, + 652, + 669, + 669 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Model Inference During model inference, we compute the interaction probability $\\hat{y}_i\\in \\mathbb{R}^K$ for each video frame. The interaction class $a$ with the highest probability is chosen as the predicted interaction for that frame: $a = \\arg \\max \\hat{y}_i$ .", + "bbox": [ + 214, + 686, + 785, + 734 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 758, + 375, + 776 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Experimental setup", + "text_level": 1, + "bbox": [ + 215, + 787, + 426, + 803 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets: We evaluate our method on MPHOI-72, Bimanual Actions, and CAD-120 datasets, each representing multi-person collaboration, one person with two", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Video HOI Interaction-centric context reasoning", + "bbox": [ + 408, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "hands and a single hand respectively.", + "bbox": [ + 212, + 146, + 483, + 160 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(I) MPHOI-72 dataset is proposed in [27], which consists of multiple humans and objects in the scene. The dataset comprises 72 videos featuring 3 human subjects and 6 objects. Within each video, 2 individuals are paired to engage in 3 distinct activities, encompassing a total of 13 sub-activities, while utilizing 2 to 4 objects", + "(II) Bimanual Actions dataset [4] is the first HOI dataset to include two hands for subjects to perform interactions which is common in reality. There are 540 videos with one person performing activities with both hands. There are 6 subjects performing 9 different activities with 10 repetitions. There are a total of 14 action labels assigned to each hand, and entity-level annotations are provided on a per-frame basis within the video.", + "(III) CAD-120 dataset [13] is popular for HOI recognition. It contains 120 videos with 10 activities performed by 4 participants. There are 10 human sub-activities labeled per frame." + ], + "bbox": [ + 212, + 161, + 782, + 369 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metric: We report $F_{1}$ at $k$ metric [14] with thresholds $k = 10\\%$ , $25\\%$ , and $50\\%$ . Compared to frame-based metrics which evaluate prediction on every single frame, this metric could measure prediction continuity in action segments because it views each predicted action segment as correct only when it has the Interaction over Union (IoU) with the corresponding ground truth over the threshold $k$ .", + "bbox": [ + 212, + 372, + 782, + 460 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Implementation details", + "text_level": 1, + "bbox": [ + 214, + 484, + 454, + 500 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In the experiment, we use three layers of context fuser for Bimanual and two for CAD-120 and MPHOI. The features of humans and objects are extracted from [27] and their dimension is mapped to 768, 256, and 512 for MPHOI, Bimanual, and CAD-120 respectively. More implementation details can be found in the supplementary material.", + "bbox": [ + 212, + 508, + 782, + 583 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.3 Quantitative results", + "text_level": 1, + "bbox": [ + 214, + 604, + 428, + 619 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Multi-person HOI recognition The quantitative results of joined segmentation and label recognition of sub-activity on MPHOI-72 in Tab. 1 show the performance of our method in multi-person HOI circumstance. Our method outperforms SOTA method 2G-GCN [27] by a large margin in all the three evaluation metrics. For $\\mathrm{F}_1@10$ , $\\mathrm{F}_1@25$ and $\\mathrm{F}_1@50$ scores, our method surpasses 2G-GCN $23.0\\%$ , $23.7\\%$ and $22.5\\%$ , respectively. The significant improvement achieved by our method indicates that the human fuser block in the CF module effectively improves the context-aware human representation learning under multi-person scenarios.", + "bbox": [ + 212, + 628, + 782, + 760 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Single-person HOI recognition We show the quantitative results for single-person HOI recognition in Tab. 2 and Tab. 3, which are performed on CAD-120 and Bimanual Actions datasets, respectively. Results in Tab. 2 show that our method performs slightly better on CAD-120 dataset compared to SOTA method 2G-GCN, with around $1\\%$ improvement over 2G-GCN on all the three metrics.", + "bbox": [ + 212, + 763, + 782, + 837 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Wang et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "$\\Delta \\mathrm{CF}$ : removing context fuser module; $\\Delta \\mathrm{ISR}$ : removing interaction state reasoner module; $\\Delta \\mathcal{L}_{fea}$ : removing interaction feature loss; $\\lambda = 0$ : ablating the hypersphere by replacing it with Euclidean distance; $\\Delta \\mathrm{CLIP} + \\mathrm{BLIP}$ : ablating CLIP and BLIP model; Traditional Classifier: ablating the interaction-centric hypersphere by replacing it with a traditional classifier constructed with MLP and Softmax classifier. The improvements of our method over current SOTA method is highlighted with upward arrows.", + "bbox": [ + 212, + 172, + 784, + 255 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/d72c28bd8d4aa7d3ab66f8db5d022d180487cd6651d91f317f2f3479c5f55b65.jpg", + "table_caption": [ + "Table 1: The results of joined segmentation and label recognition of sub-activity on MPHOI-72." + ], + "table_footnote": [], + "table_body": "
ModelF1@10F1@25F1@50
ASSIGN [23]59.1 ± 12.151.0 ± 16.733.2 ± 14.0
2G-GCN [27]68.6 ± 10.460.8 ± 10.345.2 ± 6.5
Ours91.6 ± 0.9(↑ 23.0)84.5 ± 2.6(↑ 23.7)67.7 ± 2.2(↑ 22.5)
Ours (ΔCF)65.8 ± 12.457.6 ± 14.039.2 ± 12.6
Ours (ΔISR)80.1 ± 5.573.0 ± 8.255.6 ± 6.1
Ours (ΔLfea)73.5 ± 15.769.7 ± 13.348.8 ± 13.0
Ours (λ = 0)81.2 ± 0.774.8 ± 4.253.2 ± 0.3
Ours (ΔCLIP+BLIP)80.0 ± 6.773.0 ± 9.955.6 ± 7.4
Ours (Traditional Classifier)84.6 ± 7.274.7 ± 10.554.6 ± 13.7
", + "bbox": [ + 248, + 279, + 759, + 406 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For the Bimanual Actions dataset, our method performs as good as 2G-GCN in $F_{1}@10$ while achieves $0.9\\%$ and $5.0\\%$ higher than 2G-GCN in $F_{1}@25$ and $F_{1}@50$ score, respectively. These results indicate that our method achieves generally on par or even better performance on single-person video HOI recognition task.", + "bbox": [ + 212, + 446, + 784, + 508 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/312e6ddeff50af8d8f53b23f51178b6bc41ff4155fa4c3f36cb1e8029d9a5d92.jpg", + "image_caption": [ + "Fig. 5: Qualitative ablation study results on MPHOI-72 dataset. Major prediction errors are highlighted in red dashed boxes." + ], + "image_footnote": [], + "bbox": [ + 222, + 536, + 782, + 630 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.4 Ablation studies", + "text_level": 1, + "bbox": [ + 214, + 722, + 397, + 736 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we ablate the CF module and the ISR module for validating the effectiveness of these proposed components. As shown in Tab. 1, removing CF module results in more than $20\\%$ $F_{1}$ score drop of the three metrics in MPHOI-72 dataset, indicating the essential improvement of CF module in learning context-rich representations. Visualization results of temporal segmentation of interactions in Fig. 5 indicates that removing CF module results in incorrect", + "bbox": [ + 212, + 750, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Video HOI Interaction-centric context reasoning", + "bbox": [ + 408, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/0806dc041cd145d92e2250855149e34a5edc790aecf574f6a8de13ff6d575b9e.jpg", + "table_caption": [ + "Table 2: The results of joined segmentation and label recognition of sub-activity on CAD-120. The notations are the same with Tab. 1." + ], + "table_footnote": [], + "table_body": "
ModelF1@10F1@25F1@50
rCRF [29]65.6 ± 3.261.5 ± 4.147.1 ± 4.3
Independent BiRNN [27]70.2 ± 5.564.1 ± 5.348.9 ± 6.8
ATCRF [11]72.0 ± 2.868.9 ± 3.653.5 ± 4.3
Relational BiRNN [27]79.2 ± 2.575.2 ± 3.562.5 ± 5.5
ACoLP [35]90.2 ± 2.687.4 ± 1.476.8 ± 2.3
ASSIGN [23]88.0 ± 1.884.8 ± 3.073.8 ± 5.8
2G-GCN [27]89.5 ± 1.687.1 ± 1.876.2 ± 2.8
Ours90.7 ± 2.9(↑ 0.5)88.1 ± 2.8(↑ 0.7)77.6 ± 4.7(↑ 0.8)
Ours (ΔCF)81.1 ± 4.077.0 ± 4.865.2 ± 5.6
Ours (ΔISR)88.5 ± 3.785.5 ± 3.673.9 ± 5.7
Ours (ΔLfea)89.3 ± 1.985.6 ± 2.175.9 ± 4.4
Ours (λ = 0)72.0 ± 4.465.0 ± 6.948.6 ± 6.3
Ours (ΔCLIP+BLIP)89.4 ± 2.385.5 ± 3.974.9 ± 5.7
Ours (Traditional Classifier)79.5 ± 11.073.9 ± 11.456.6 ± 12.5
", + "bbox": [ + 256, + 196, + 750, + 385 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/205d20fe6fba57ba3413e8035d40578c6d108c0f2141c14c0e415090831424b9.jpg", + "image_caption": [ + "Fig. 6: Qualitative ablation study results on CAD-120 dataset. Major prediction errors are highlighted in red dashed boxes." + ], + "image_footnote": [], + "bbox": [ + 223, + 412, + 781, + 452 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "interaction predictions in both humans (highlighted in red in Fig. 5). Similarly, in CAD-120 and Bimanual Actions datasets, deleting CF module also results in massive $F_{1}$ score drop. The visualization results in Fig. 6 suggests an incorrect prediction segment when removing CF module.", + "bbox": [ + 212, + 521, + 782, + 580 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Furthermore, we ablate hypersphere by replacing it with Euclidean distance, where $\\lambda = 0$ in Eq. 6. Results in Tab. 1, 2 and 3 indicate that utilizing Euclidean distance results in at least $7\\%$ , $10\\%$ and $3\\%$ $F_{1}$ scores drop in MPHOI, CAD-120 and Bimanual Actions dataset, respectively. Therefore, we conclude that Euclidean distance do not introduce HOI structure priors, ignoring valuable structure information of HOI for guiding predictions. Subsequently, we ablate CLIP and BLIP models by randomly initialize interaction features and context features. Results in Tab. 1, 2 and 3 indicate that removing CLIP and BLIP models results in some drop of model performance, but is still on par with or better than SOTA methods. Thus, it is the intricately designed structure of our model that substantiates the substantial enhancement in performance. Finally, we ablate the interaction-centric hypersphere by replacing it with a traditional classifier constructed with multi-layer perceptron (MLP) and Softmax classifier. The outcomes, as presented in Tab. 1, 2 and 3, reveal a notable decline of over $7\\%$ in the $F_{1}$ score within the MPHOI dataset when employing the traditional classifier. Likewise, in the CAD-120 dataset, the traditional classifier results in a substantial decrease of more than $11\\%$ in the $F_{1}$ score. Additionally, within", + "bbox": [ + 212, + 583, + 787, + 843 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Wang et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/7e934948a28f71d787f292856f328d8af3d94d1e8a51c04c7b51be967645edc4.jpg", + "table_caption": [ + "Table 3: The results of joined segmentation and label recognition of sub-activity on Bimanual Actions. The notations are the same with Tab. 1." + ], + "table_footnote": [], + "table_body": "
ModelF1@10F1@25F1@50
Dreher et al. [4]40.6 ± 7.234.8 ± 7.122.2 ± 5.7
Independet BiRNN [27]74.7 ± 7.072.0 ± 7.061.8 ± 7.3
Relational BiRNN [27]77.7 ± 3.975.0 ± 4.264.8 ± 5.3
ASSIGN [23]84.0 ± 2.081.2 ± 2.068.5 ± 3.3
2G-GCN [27]85.0 ± 2.282.0 ± 2.669.2 ± 3.1
Ours85.0 ± 2.582.9 ± 2.9(↑ 0.9)74.2 ± 4.3(↑ 5.0)
Ours (ΔCF)82.5 ± 5.080.5 ± 5.571.1 ± 7.0
Ours (ΔISR)84.1 ± 2.381.8 ± 2.873.0 ± 3.7
Ours (ΔLfea)84.5 ± 4.682.0 ± 5.271.8 ± 6.9
Ours (λ = 0)76.7 ± 5.274.3 ± 6.065.2 ± 6.3
Ours (ΔCLIP+BLIP)84.3 ± 1.481.8 ± 1.873.2 ± 2.7
Ours (Traditional Classifier)82.0 ± 3.679.8 ± 4.171.0 ± 5.6
", + "bbox": [ + 256, + 196, + 750, + 375 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the Bimanual Actions dataset, the traditional classifier induces a decline exceeding $2\\%$ in the $F_{1}$ score. These findings unanimously underscore the efficacy of modeling HOI manifold structures by the hypersphere module.", + "bbox": [ + 212, + 416, + 787, + 463 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c21d56e15216e69ceabf377e4c562f2d66908b903faba48b5f59ccb6fcc1371d.jpg", + "image_caption": [ + "Fig. 7: (A) Visualization of interaction-centric hyperspheres, the learned interaction representations (large points locate at the center of hyperspheres) and human-object entity representations (small points surrounding the hyperspheres) in embedding space. Selected frame samples are shown for each interaction-centric hypersphere. (B) 2D visualization of interaction-centric hyperspheres." + ], + "image_footnote": [], + "bbox": [ + 218, + 492, + 493, + 585 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/448c3a360fa59d726c6c2503f919a4330a2a42cb743008341e1ff2d6e57b3d10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 491, + 805, + 570 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.5 Qualitative results", + "text_level": 1, + "bbox": [ + 214, + 720, + 415, + 736 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We show some visualization results on MPHOI in Fig 8 to compare our method with SOTA method 2G-GCN. The red highlighted boxes indicate that 2G-GCN tend to generate unreasonable interaction predictions, while our method generates more reasonable interaction predictions. The visualization results in Fig. 9 show similar prediction pattern where 2G-GCN predicted some unreasonable short segments (highlighted in red boxes) while our method predicts more accu", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Video HOI Interaction-centric context reasoning", + "bbox": [ + 408, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "rately. We also visualize the interaction-centric hypersphere, the learned interaction representations and the human-object entity representations in embedding space in Fig. 7. Results in Fig. 7 (A) show that human-object entity representations (small dots in Fig. 7 (A)) belonging to the specific interaction class locates near the surface of the corresponding hypersphere. Similar with Fig. 7 (A), we also visualize the intraction-centric hyerspheres in 2D sapce in Fig. 7 (B). Results in Fig. 7 (B) indicate that HOIs distribute around the sphere of different interaction classes. The distributions of HOIs on each sphere also differ a lot, where the sphere belongs to the solve interaction class captures HOIs on the upper part of its sphere, while the sphere belongs to the cut interaction class captures HOIs on the lower part of its sphere. These results suggest that our model successfully model the manifold structure of HOI.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e655df80783c4250729bb875fb971ec4912271b21a525674760f66faadeac719.jpg", + "image_caption": [ + "Fig. 8: Visualization results on MPHOI-72 dataset. Major prediction errors are highlighted in red dashed boxes." + ], + "image_footnote": [], + "bbox": [ + 222, + 356, + 782, + 448 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/50faa1b4a7cee1191ab92636440d64e83d5d4fdc20b17c9ee3ac123369f10cf3.jpg", + "image_caption": [ + "Fig. 9: Visualization results on CAD-120 dataset. Major prediction errors are highlighted in red dashed boxes." + ], + "image_footnote": [], + "bbox": [ + 222, + 542, + 782, + 580 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 669, + 359, + 686 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we propose an interaction-centric spatio-temporal context reasoning network for multi-person video HOI recognition. Specifically, we propose a context fuser and an interaction state reasoner to learn spatio-temporal context-rich and reasoning-aware entity representations. We further represent HOI components with an interaction-centric hypersphere for HOI classification. Experiment results show that our method outperforms SOTA method by more than $22\\%$ $F_{1}$ score in multi-person scenarios, and achieves competitive results on single-person cases compared to SOTA methods.", + "bbox": [ + 212, + 700, + 787, + 823 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Wang et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This material is based upon work supported under the AI Research Institutes program by National Science Foundation and the Institute of Education Sciences, U.S. Department of Education through Award #2229873 - AI Institute for Transforming Education for Children with Speech and Language Processing Challenges. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation, the Institute of Education Sciences, or the U.S. Department of Education.", + "bbox": [ + 212, + 180, + 787, + 301 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 328, + 323, + 344 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Chopra, S., Hadsell, R., LeCun, Y.: Learning a similarity metric discriminatively, with application to face verification. In: 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05). vol. 1, pp. 539-546 vol. 1 (2005). https://doi.org/10.1109/CVPR.2005.2029", + "2. Deng, S., Zhang, N., Chen, H., Tan, C., Huang, F., Xu, C., Chen, H.: Low-resource extraction with knowledge-aware pairwise prototype learning. Knowledge-Based Systems 235, 107584 (2022) 4", + "3. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 4", + "4. Dreher, C.R.G., Wächter, M., Asfour, T.: Learning object-action relations from bimanual human demonstration using graph networks. IEEE Robotics and Automation Letters (RA-L) 5(1), 187-194 (2020). https://doi.org/10.1109/LRA.2019.2949221 3, 10, 13", + "5. Gao, C., Xu, J., Zou, Y., Huang, J.B.: Drg: Dual relation graph for human-object interaction detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16. pp. 696-712. Springer (2020) 4", + "6. Gao, K., Chen, L., Zhang, H., Xiao, J., Sun, Q.: Compositional prompt tuning with motion cues for open-vocabulary video relation detection. In: The Eleventh International Conference on Learning Representations (2022) 4", + "7. Gupta, T., Schwing, A., Hoiem, D.: No-frills human-object interaction detection: Factorization, layout encodings, and training techniques. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9677-9685 (2019) 4", + "8. Iftekhar, A., Chen, H., Kundu, K., Li, X., Tighe, J., Modolo, D.: What to look at and where: Semantic and spatial refined transformer for detecting human-object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5353-5363 (2022) 4", + "9. Kim, B., Choi, T., Kang, J., Kim, H.J.: Uniondet: Union-level detector towards real-time human-object interaction detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 498-514. Springer (2020) 4" + ], + "bbox": [ + 222, + 362, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Video HOI Interaction-centric context reasoning", + "bbox": [ + 408, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "10. Kim, B., Lee, J., Kang, J., Kim, E.S., Kim, H.J.: Hotr: End-to-end human-object interaction detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 74-83 (June 2021) 4", + "11. Koppula, H.S., Saxena, A.: Anticipating human activities using object affordances for reactive robotic response. IEEE transactions on pattern analysis and machine intelligence 38(1), 14-29 (2015) 12", + "12. Koppula, H.S., Gupta, R., Saxena, A.: Learning human activities and object affordances from rgb-d videos. The International journal of robotics research 32(8), 951-970 (2013) 3, 4", + "13. Koppula, H.S., Gupta, R., Saxena, A.: Learning human activities and object affordances from rgb-d videos. Int. J. Rob. Res. 32(8), 951-970 (jul 2013). https://doi.org/10.1177/0278364913478446, https://doi.org/10.1177/0278364913478446 10", + "14. Lea, C., Flynn, M.D., Vidal, R., Reiter, A., Hager, G.D.: Temporal convolutional networks for action segmentation and detection. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 156-165 (2017) 10", + "15. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) 6", + "16. Li, Y.L., Zhou, S., Huang, X., Xu, L., Ma, Z., Fang, H.S., Wang, Y., Lu, C.: Transferable interactiveness knowledge for human-object interaction detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3585-3594 (2019) 4", + "17. Liao, Y., Liu, S., Wang, F., Chen, Y., Qian, C., Feng, J.: Ppdm: Parallel point detection and matching for real-time human-object interaction detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 482-490 (2020) 4", + "18. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017) 9", + "19. Liu, X., Li, Y.L., Wu, X., Tai, Y.W., Lu, C., Tang, C.K.: Interactiveness field in human-object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20113-20122 (2022) 2, 4", + "20. Luan, T., Wang, Y., Zhang, J., Wang, Z., Zhou, Z., Qiao, Y.: Pc-hmr: Pose calibration for 3d human mesh recovery from 2d images/videos. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 2269-2276 (2021) 4", + "21. Luan, T., Zhai, Y., Meng, J., Li, Z., Chen, Z., Xu, Y., Yuan, J.: High fidelity 3d hand shape reconstruction via scalable graph frequency decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16795-16804 (2023) 4", + "22. Mettes, P., Van der Pol, E., Snoek, C.: Hyperspherical prototype networks. Advances in neural information processing systems 32 (2019) 4", + "23. Morais, R., Le, V., Venkatesh, S., Tran, T.: Learning asynchronous and sparse human-object interaction in videos. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16041-16050 (2021) 2, 4, 11, 12, 13", + "24. Nagarajan, T., Feichtenhofer, C., Grauman, K.: Grounded human-object interaction hotspots from video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8688-8697 (2019) 4" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Wang et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "25. Park, J., Park, J.W., Lee, J.S.: Viplo: Vision transformer based pose-conditioned self-loop graph for human-object interaction detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17152-17162 (June 2023) 4", + "26. Qi, S., Wang, W., Jia, B., Shen, J., Zhu, S.C.: Learning human-object interactions by graph parsing neural networks. In: Proceedings of the European Conference on Computer Vision (ECCV) (September 2018) 4", + "27. Qiao, T., Men, Q., Li, F.W., Kubotani, Y., Morishima, S., Shum, H.P.: Geometric features informed multi-person human-object interaction recognition in videos. In: European Conference on Computer Vision. pp. 474-491. Springer (2022) 2, 3, 4, 6, 10, 11, 12, 13", + "28. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 6, 8", + "29. Sener, O., Saxena, A.: rcrf: Recursive belief estimation over crfs in rgb-d activity videos. In: Robotics: Science and systems (2015) 12", + "30. Sunkesula, S.P.R., Dabral, R., Ramakrishnan, G.: Lighten: Learning interactions with graph and hierarchical temporal networks for hoi in videos. In: Proceedings of the 28th ACM international conference on multimedia. pp. 691-699 (2020) 4", + "31. Tamura, M., Ohashi, H., Yoshinaga, T.: Qpic: Query-based pairwise human-object interaction detection with image-wide contextual information. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10410-10419 (June 2021) 4", + "32. Tu, D., Sun, W., Min, X., Zhai, G., Shen, W.: Video-based human-object interaction detection from tubelet tokens. Advances in Neural Information Processing Systems 35, 23345-23357 (2022) 4", + "33. Wang, T., Anwer, R.M., Khan, M.H., Khan, F.S., Pang, Y., Shao, L., Laaksonen, J.: Deep contextual attention for human-object interaction detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5694-5702 (2019) 4", + "34. Xi, N., Meng, J., Yuan, J.: Chain-of-look prompting for verb-centric surgical triplet recognition in endoscopic videos. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 5007-5016 (2023) 4", + "35. Xi, N., Meng, J., Yuan, J.: Open set video hoi detection from action-centric chain-of-look prompting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3079-3089 (2023) 12", + "36. Zhang, Y., Pan, Y., Yao, T., Huang, R., Mei, T., Chen, C.W.: Exploring structure-aware transformer over interaction proposals for human-object interaction detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19548-19557 (2022) 2, 4", + "37. Zhong, X., Qu, X., Ding, C., Tao, D.: Glance and gaze: Inferring action-aware points for one-stage human-object interaction detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13234-13243 (2021) 4" + ], + "bbox": [ + 212, + 146, + 787, + 771 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Video HOI Interaction-centric context reasoning", + "bbox": [ + 408, + 114, + 732, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition/0a79d390-9d09-4042-af06-6b62e335816f_model.json b/2024/Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition/0a79d390-9d09-4042-af06-6b62e335816f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..12515065dd1e5b2dbfa37a6e2f70db4bed0a03a7 --- /dev/null +++ b/2024/Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition/0a79d390-9d09-4042-af06-6b62e335816f_model.json @@ -0,0 +1,2148 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.296, + 0.141, + 0.71, + 0.21 + ], + "angle": 0, + "content": "Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.244, + 0.761, + 0.261 + ], + "angle": 0, + "content": "Yisong Wang\\(^{1*}\\), Nan Xi\\(^{2*}\\), Jingjing Meng\\(^{3}\\), and Junsong Yuan\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.377, + 0.271, + 0.629, + 0.286 + ], + "angle": 0, + "content": "\\(^{1}\\) School of EECS, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.286, + 0.751, + 0.3 + ], + "angle": 0, + "content": "\\(^{2}\\) Department of Computer Science and Engineering, University at Buffalo" + }, + { + "type": "text", + "bbox": [ + 0.451, + 0.3, + 0.553, + 0.312 + ], + "angle": 0, + "content": "3 Amazon Inc." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.315, + 0.785, + 0.329 + ], + "angle": 0, + "content": "wys05y98@gmail.com, {nanxi, jsyuany}@buffalo.edu, jingjing.meng1@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.392, + 0.744, + 0.739 + ], + "angle": 0, + "content": "Abstract. Understanding human-object interaction (HOI) in videos represents a fundamental yet intricate challenge in computer vision, requiring perception and reasoning across both spatial and temporal domains. Despite previous success of object detection and tracking, multi-person video HOI recognition still faces two major challenges: (1) the three facets of HOI (human, objects, and the interactions that bind them) exhibit interconnectedness and exert mutual influence upon one another. (2) the complexity of multi-person multi-object combinations in spatio-temporal interaction. To address them, we design a spatio-temporal context fuser to better model the interactions among persons and objects in videos. Furthermore, to equip the model with temporal reasoning capacity, we propose an interaction state reasoner module on top of context fuser. Considering the interaction is a key element to bind human and object, we propose an interaction-centric hypersphere in the feature embedding space to model each category of interaction. It helps to learn the distribution of HOI samples belonging to the same interactions on the hypersphere. After training, each interaction prototype sphere will fit the testing HOI sample to determine the HOI classification result. Empirical results on multi-person video HOI dataset MPHOI-72 indicate that our method remarkably surpasses state-of-the-art (SOTA) method by more than \\(22\\%\\) \\(F_{1}\\) score. At the same time, on single-person datasets Bimanual Actions (single-human two-hand HOI) and CAD-120 (single-human HOI), our method achieves on par or even better results compared with SOTA methods. Source code is released at the following link: https://github.com/southnx/IcH-Vid-HOI." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.754, + 0.757, + 0.782 + ], + "angle": 0, + "content": "Keywords: interaction-centric \\(\\cdot\\) spatio-temporal context reasoning \\(\\cdot\\) multi-person human-object interaction" + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.825, + 0.546, + 0.841 + ], + "angle": 0, + "content": "* Equal contribution; ≔ Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Wang et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.386, + 0.147, + 0.593, + 0.159 + ], + "angle": 0, + "content": "Multi-person Video Human-Object Interaction" + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.162, + 0.41, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.162, + 0.581, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.162, + 0.752, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.224, + 0.785, + 0.253 + ], + "angle": 0, + "content": "Fig. 1: Examples of multi-person video HOI. Three categories of interactions are presented, including co-working, cheering, hair_cutting." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.281, + 0.376, + 0.297 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.787, + 0.386 + ], + "angle": 0, + "content": "Video-based Human-Object Interaction (HOI) recognition aims to identify the interactions occurring between human and object entities within video frames. Precisely recognizing HOIs in real world scenarios is essential for a bunch of applications, such as assisting patients by recognizing daily activities and predicting pedestrian movements to avoid accidents for autonomous vehicles." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.787, + 0.599 + ], + "angle": 0, + "content": "Most existing HOI recognition research focuses on static images [19, 36], with much less attention on video-based HOI recognition. Video-based HOI recognition is more demanding than image-based HOI recognition due to the necessity of comprehending complex spatio-temporal dynamics and reasoning about multi-human and multi-object interaction combinations. The complexity is further exacerbated when dealing with multi-person circumstances, as examples shown in Fig. 1. In such cases, various human and object entities mutually influence each other, resulting in intricate interdependencies within the scene. Additionally, the three components (human, object and interactions that the entities are involved in) of HOI exhibit interwind structures, e.g., the possible interactions that can occur within a scene given a human and an object. However, current video HOI recognition methods [23, 27] do not fully explore such inherent structural nature of HOI components. Instead, they often opt for disentangled representations for each component, which may bring suboptimal representation capabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.787, + 0.841 + ], + "angle": 0, + "content": "To overcome the aforementioned limitations, we introduce an interaction-centric spatio-temporal context reasoning approach for representing video HOIs. For the purpose of enhancing the awareness of complex HOI structures in our representations, we introduce the Context Fuser (CF), which encodes both entity representations and interaction representations. Moreover, to empower our model with the ability to reason about interaction state transitions across video frames, we propose the Interaction State Reasoner (ISR) module for generating interaction representations. In addition, we employ a bidirectional Gated Recurrent Unit (BiGRU) to model temporal dynamics across video frames. This multi-level representation learning framework not only facilitates effective exploration of the interdependencies among structured HOI components but also empowers the model with interaction reasoning capabilities in both spatial and temporal domains. To compute the probability of an HOI in the scene, we introduce an interaction-centric hypersphere. This approach leverages the concept of a hypersphere, where the interaction is located at its center, while the human and object entities involved in that interaction are situated on the hypersphere's" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.41, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Video HOI Interaction-centric context reasoning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "surface. We assume that human-object entities belonging to each interaction class are located on their respective hyperspheres, specific to that interaction class." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.788, + 0.463 + ], + "angle": 0, + "content": "Concretely, we generate context-rich and reasoning-aware video HOI representations through three key components: the Context Fuser (CF), Interaction State Reasoner (ISR) and Bidirectional GRU (BiGRU). The CF module integrates context information from human-object entities and interactions. It comprises three fuser blocks for humans, objects, and interactions. The Object Fuser Block processes local video frame data, enhancing object features. The Interaction Fuser Block combines human and object representations with interaction-specific characteristics. Additionally, in multi-person scenes, the Human Fuser Block captures human representations influenced by others. This approach fosters comprehensive HOI representations via effective context fusion. To facilitate interaction reasoning, we place the ISR module on top of the context fuser module, yielding entity representations capable of capturing interaction transition dynamics. These entity representations are then input into the BiGRU module to model temporal dynamics across video frames, thereby ensuring a comprehensive understanding of the evolving context and interactions within the video data. Finally, we determine interaction classes in each frame with the interaction-centric hypersphere, computing the probability of human-object entities belonging to specific interaction classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.8, + 0.569 + ], + "angle": 0, + "content": "We assess our model's performance on three video-based HOI datasets: MPHOI-72 [27] (multi-person), Bimanual Actions [4] and CAD-12 [12] (single-person). Our results highlight our model's superiority in multi-person scenarios, achieving an impressive over \\(22\\%\\) \\(F_{1}\\) score improvement over the current state-of-the-art (SOTA). In single-person scenarios, our method delivers on par or even better performance compared to the current SOTA method. Our major contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.578, + 0.787, + 0.638 + ], + "angle": 0, + "content": "- To represent inherent HOI manifold structures, we propose an interaction-centric hypersphere representation scheme. This scheme explicitly guide the process of learning intrinsic structural nature of HOI and elucidates the interdependencies among its components." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.638, + 0.787, + 0.696 + ], + "angle": 0, + "content": "- To learn context-rich and reasoning-aware entity representations, we introduce context fuser and interaction state reasoning modules. This enhancement results in entity representations that are highly suitable for video-based HOI tasks." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.698, + 0.787, + 0.758 + ], + "angle": 0, + "content": "- Extensive experiment results showcase that our method achieves SOTA performance with a huge improvement of more than \\(22\\%\\) \\(F_{1}\\) score over existing methods in multi-person scenario. Additionally, our model achieves competitive results in single-person cases compared to SOTA method." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.578, + 0.787, + 0.758 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.78, + 0.393, + 0.795 + ], + "angle": 0, + "content": "2 Related works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "HOI detection in images: HOI detection in images aims at understanding interactions in images between humans and objects. Different methods have been" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "proposed in previous studies. Some works propose Convolutional Neural Networks (CNN)-based methods which can be further divided into one-stage methods [9,17,37] and two-stage methods [5,7,16,33]. However, these methods usually lack of ability to capture global context information. Recently, Transformer-based models [8,10,31,36] became the main approach for the HOI task. Following the architecture of DETR, these models achieved superior performance on HOI detection. Moreover, some works also utilize graph [25] and interactiveness field [19] to achieve better performance. These various approaches to image HOI provide the fundamentals for video HOI recognition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.539 + ], + "angle": 0, + "content": "HOI recognition in videos: Video HOI recognition is the foundation of many real-world applications, including understanding surgical activities in the operation room [34] and guiding 3D human reconstruction [20, 21]. Video-based HOI recognitions have to deal with both spatial and temporal reasoning. Before the use of neural networks, some early studies formulated this task using the Markov model [12] to utilize temporal cues. In [24], HOI hotspots in videos are learned in a novel approach, with two networks trained jointly to capture spatial regions where actions happen. While [6] focuses on visual relation prediction in open-vocabulary with pretrained visual-language models. Recent works have used Recurrent Neural Networks (RNN) combined with Graph Neural Networks (GNN) [23, 26, 27, 30] to predict human-object relations in videos. Inspired by ViT [3], some works also propose Transformer-based methods to reason spatial relations better [32]. However, RNN-based models usually require complex training strategies or long training time in order to achieve the best performance. Moreover, when multiple persons are involved in an activity jointly, these methods lack the ability to model their collaboration, resulting in poor performance when the interactions are performed by multiple persons." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.788, + 0.646 + ], + "angle": 0, + "content": "Hyperspheres for class representation: Hyperspheres have been demonstrated to be an effective approach to model class representation [2, 22]. Geometrical modeling strategy has been proposed in [2], where the effectiveness has been confirmed. This approach proves advantageous for capturing and representing enriched class-level information, particularly well-suited for creating measurements in Euclidean space. Consequently, it is naturally adaptable to structured prediction tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.671, + 0.36, + 0.687 + ], + "angle": 0, + "content": "3 Motivation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Video HOI recognition task involves identifying both the human and object entities engaged in an interaction across a sequence of video frames. This task encompasses spatial and temporal aspects, as it requires understanding the relationships between humans, objects, and their interactions over time. However, current methods for video HOI recognition often neglect this crucial dependency structure, resulting in the separation of learned representations associated with humans, objects, and their interactions, ultimately compromising their representational accuracy. To address this issue, we propose a novel approach featuring a context fuser and an interaction state reasoner to capture spatio-temporal" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.41, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Video HOI Interaction-centric context reasoning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.142, + 0.775, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.545 + ], + "angle": 0, + "content": "Fig. 2: Model Overview. Each video frame is applied with a backbone for feature extraction. Subsequently, a context fuser and interaction state reasoner is employed for learning interaction representations \\( z_{I} \\) and human-object entity representations \\( \\{\\hat{z}_E^t\\}_{t=1}^T \\). Bidirectional GRU is further utilized to model temporal dynamics across video frames. The final predicted interaction probability \\( \\{u^t\\}_{t=1}^T \\) is computed from the interaction-centric hypersphere. The dimension of each hypersphere is \\( d \\), which is the same with the feature dimension in our model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.576, + 0.788, + 0.622 + ], + "angle": 0, + "content": "contextual representations. Additionally, we introduce an interaction-centric hypersphere to represent HOI manifold structure and model the inter-dependency among predicted humans, objects, and interactions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.646, + 0.331, + 0.661 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.678, + 0.434, + 0.693 + ], + "angle": 0, + "content": "4.1 Problem formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "For a video dataset \\(\\mathcal{V}\\), given a video clip \\(V\\in \\mathcal{V}\\) containing \\(T\\) video frames \\(\\{v_{1},\\dots,v_{T}\\}\\), video HOI recognition aims to predict the temporal segmentation of interactions between human and object entities across all the video frames. Formally, we aim to learn an HOI recognition model \\(\\mathcal{M}\\) that outputs the segmentation of human's sub-activity \\(\\{s_n\\}_{n = 1}^N\\) in each frame, where \\(N\\) is the number of human sub-activity segments. Each segment \\(s_n\\) is represented as an interval from its start time \\(t_n\\) to end time \\(t_{n + 1}\\): \\(s_n = [t_n,t_{n + 1})\\). The start and end time of each segment are determined from interaction probability prediction \\(\\{u^t\\}_{t = 1}^T\\) of each frame, where \\(u^{t}\\in \\mathbb{R}^{K}\\), \\(K\\) is the number of possible interaction classes." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Wang et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.378, + 0.163 + ], + "angle": 0, + "content": "4.2 Model Design" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.368 + ], + "angle": 0, + "content": "In the following section, we introduce our interaction-centric hypersphere reasoning model for video HOI recognition in detail. As shown in Fig. 2, our major idea is to construct a hypersphere to represent each HOI in the scene. For each hypersphere, the interaction locates at the center of the hypersphere, while the corresponding human-object entity belongs to that interaction locates at the surface of the hypersphere. We construct Context Fuser (CF) module to learn context-rich human-object entity representations. For the aim of enabling model with reasoning ability over interaction state transitions, we propose Interaction State Reasoner (ISR) to reason on whether the current interaction will be continued or stopped. To model the temporal dynamics of HOI in videos, we update human-object entity representations \\(\\{\\hat{z}_E^t\\}_{t=1}^T\\) along the temporal domain with bidirectional GRU (BiGRU). Finally, predicted interaction class probability \\(u^t\\) of each frame is computed from the interaction-centric hypersphere." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.787, + 0.461 + ], + "angle": 0, + "content": "Context fuser For a sequence of video frames \\(\\{v_{t}\\}_{t = 1}^{T}\\), we follow 2G-GCN [27] to extract feature of humans and objects from backbone network. The extracted human features \\(z_{H}\\in \\mathbb{R}^{d}\\) contain both bounding box information and skeleton keypoint information, where \\(d\\) indicates the feature dimension. Object features \\(z_{O}\\in \\mathbb{R}^{d}\\) contain only bounding box information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.569 + ], + "angle": 0, + "content": "We design a context fuser (CF) module shown in Fig. 3 to generate human-object entity representations \\(\\{z_E^t\\}_{t=1}^T\\) based on human, object and contextual information. In multi-person circumstances, CF contains object fuser block, interaction fuser block and human fuser block sequentially. First, we design an object fuser block to incorporate object representations in local temporal regions into current object representations, generating learned object representation \\(\\hat{z}_{O_1}\\) and \\(\\hat{z}_{O_2}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.577, + 0.786, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {O _ {1}} = F F N \\left(S A \\left(Q = z _ {O _ {1}} ^ {t}; K, V = z _ {O _ {1}} ^ {\\bar {t}}\\right) + z _ {O _ {1}} ^ {t}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.336, + 0.601, + 0.669, + 0.618 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {O _ {2}} = F F N (S A (Q = z _ {O _ {2}} ^ {t}; K, V = z _ {O _ {2}} ^ {t}) + z _ {O _ {2}} ^ {t}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.789, + 0.841 + ], + "angle": 0, + "content": "where \\( SA \\) indicates self-attention, \\( FFN \\) is feed forward network, \\( z_{O_1}^t \\in \\mathbb{R}^d \\) and \\( z_{O_2}^t \\in \\mathbb{R}^d \\) are the initial object features, while \\( z_{O_1}^{\\bar{t}} \\in \\mathbb{R}^{20d} \\) and \\( z_{O_2}^{\\bar{t}} \\in \\mathbb{R}^{20d} \\) are stacked object features from a local time window, \\( \\bar{t} \\in [t - 10,t + 10) \\). Subsequently, for all the \\( K \\) possible interactions \\( \\{I_i\\}_{i = 1}^K \\) as shown in Fig. 3, each interaction class \\( I_i \\) is prompted as a sentence \\( s = \\) \"The human is [interact]ing in the scene.\", where [interact] indicates the specific interaction class. Then the sentence is applied with the text encoder \\( (\\mathcal{F}_T) \\) of large-scale vision-language model CLIP [28] to initialize the interaction feature \\( z_I = \\mathcal{F}_T(s) \\in \\mathbb{R}^d \\). We also generate context feature \\( z_C \\in \\mathbb{R}^d \\) to represent the semantic information of each frame. Specifically, we extract frame caption \\( c_i \\) from video frame \\( i(i = 1,\\dots,T) \\) with BLIP [15] model and apply the caption with CLIP model to extract text embedding. In order to adapt the human \\( (z_{H_1},z_{H_2}) \\) features, object \\( (z_{O_1},z_{O_2}) \\) features and context feature \\( (z_C) \\) to the specific interaction feature \\( z_I \\), we construct an interaction fuser block that contains a cross-attention (CA) module" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.41, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Video HOI Interaction-centric context reasoning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.144, + 0.596, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.144, + 0.773, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.475 + ], + "angle": 0, + "content": "Fig.3: Context fuser. In multi-person scenario (left), the context fuser consists of object fuser block, interaction fuser block and human fuser block. In single-person case (right), only object fuser block and interaction fuser block are employed. Detailed descriptions of the architecture and annotations can be found in Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.6 + ], + "angle": 0, + "content": "followed by a Feed Forward Network (FFN), generating interaction-aware human features \\((z_{H_1}' \\in \\mathbb{R}^d, z_{H_2}' \\in \\mathbb{R}^d)\\), object features \\((z_{O_1}' \\in \\mathbb{R}^d, z_{O_2}' \\in \\mathbb{R}^d)\\) and context feature \\((z_C' \\in \\mathbb{R}^d)\\). Furthermore, to model the influence between the two humans in multi-person scenarios, we construct a human fuser block, featuring the same architecture with interaction fuser block. For the first human (Human1), the updated human feature \\(\\hat{z}_{H_1} \\in \\mathbb{R}^d\\) is generated as:" + }, + { + "type": "equation", + "bbox": [ + 0.294, + 0.611, + 0.787, + 0.63 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {H _ {1}} = F F N (C A (Q = S A (z _ {H _ {1}} ^ {\\prime}); K, V = z _ {H _ {2}} ^ {\\prime}) + S A (z _ {H _ {1}} ^ {\\prime})). \\qquad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.639, + 0.788, + 0.7 + ], + "angle": 0, + "content": "Human2 feature \\(\\hat{z}_{H_2} \\in \\mathbb{R}^d\\) is generated in the same way as Eq. 2. Finally, the human-object entity representation \\(z_{E_1} \\in \\mathbb{R}^d\\) of Human1 is computed by max-pooling operation over all the \\(d\\) dimensions of the four representations shown in Eq. 3:" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.7, + 0.787, + 0.719 + ], + "angle": 0, + "content": "\\[\nz _ {E _ {1}} = M a x P o o l (\\hat {z} _ {H _ {1}}, z _ {O _ {1}} ^ {\\prime}, z _ {O _ {2}} ^ {\\prime}, z _ {C} ^ {\\prime}). \\qquad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.726, + 0.788, + 0.787 + ], + "angle": 0, + "content": "The human-object entity representation \\( z_{E_2} \\in \\mathbb{R}^d \\) of Human2 is computed with similar approach as Eq. 3. The CF module for single-person cases are similar with multi-person, except that the human fuser block is removed and there is only one human feature as query to be fed into the interaction fuser block." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Interaction State Reasoner To augment the model's ability in interaction state transition reasoning, we introduce an Interaction State Reasoner (ISR)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.359 + ], + "angle": 0, + "content": "module following CF module. ISR module explicitly empowers the model to determine whether the current interaction should persist or transit to another interaction. Specifically, as shown in Fig. 4, at each time \\( t \\), the two possible states state1 and state2 represent \"continue\" or \"stop\" of an interaction, respectively. Each state is prompted as one sentence, where \\( s_1 = \\) \"This interaction is going to continue.\" and \\( s_2 = \\) \"This interaction is going to stop and change to another interaction.\" Then the embeddings of the two states \\( z_{state_1} \\) and \\( z_{state_2} \\) are generated from CLIP [28] text encoder \\( \\mathcal{F}_T \\): \\( z_{state_1} = \\mathcal{F}_T(s_1) \\in \\mathbb{R}^d \\), \\( z_{state_2} = \\mathcal{F}_T(s_2) \\in \\mathbb{R}^d \\). Interaction state embeddings \\( z_{state_1} \\) and \\( z_{state_2} \\) are further fed to a reasoner block (shown in Fig. 4) together with the interaction embedding \\( z_t^t \\in \\mathbb{R}^d \\) at time \\( t \\), generating state-informed interaction embeddings \\( \\hat{z}_{state_1} \\) and \\( \\hat{z}_{state_2} \\). The reasoner block contains a FFN and a State Interpolation (SI) module. The SI module generates the weights \\( \\omega_1, \\omega_2 \\) for the two interaction states in the following approach:" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.366, + 0.786, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\omega_ {1}, \\omega_ {2} = S o f t m a x (F F N (z _ {E} ^ {t}) \\cdot [ z _ {s t a t e _ {1}} ^ {\\top}, z _ {s t a t e _ {2}} ^ {\\top} ]), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.788, + 0.451 + ], + "angle": 0, + "content": "where \\(\\top\\) indicates transpose operation. Subsequently, the final human-object entity representation \\(\\hat{z}_E^t\\in \\mathbb{R}^d\\) at time \\(t\\) is generated by interpolate over the current entity representation \\(z_{E}^{t}\\) at time \\(t\\) and the entity representation \\(z_{E}^{t - 1}\\) at time \\(t - 1\\):" + }, + { + "type": "equation", + "bbox": [ + 0.412, + 0.453, + 0.786, + 0.471 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {E} ^ {t} = \\omega_ {1} \\cdot z _ {E} ^ {t - 1} + \\omega_ {2} \\cdot z _ {E} ^ {t}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.788, + 0.506 + ], + "angle": 0, + "content": "Consequently, the generated human-object entity representation \\(\\hat{z}_E^t\\) is able to reason on the possible future interaction transitions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.541, + 0.611, + 0.738 + ], + "angle": 0, + "content": "Interaction-centric Hypersphere With the above generated human-object entity representation \\(\\hat{z}_E^t\\) and interaction representation \\(z_I\\), we need to calculate the probability of a human-object entity \\(E\\) categorizing into the interaction class \\(I_i\\). To that end, we design an interaction-centric hypersphere with interaction at the center of hypersphere and human-object entity at its surface. We call it \"hypersphere\" because this is a sphere in the \\(\\mathbb{R}^d\\) space, where \\(d\\) is the feature dimension in our model, which is obviously larger than 3. This hypersphere design models the manifold structure of HOI, which is in charge of the specific interaction class. Concretely, we employ a hyperspherical measurement:" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.536, + 0.769, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.687, + 0.788, + 0.731 + ], + "angle": 0, + "content": "Fig. 4: Interaction State Reasoner. \\(\\odot\\) indicates inner product." + }, + { + "type": "equation", + "bbox": [ + 0.329, + 0.745, + 0.787, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\mathcal {U} (E, I _ {i}) = \\frac {\\exp \\left(- \\left[ | | z _ {I _ {i}} - \\hat {z} _ {E} | | _ {2} - \\lambda \\right] _ {+}\\right)}{\\sum_ {j = 1} ^ {K} \\exp \\left(- \\left[ | | z _ {I _ {j}} - \\hat {z} _ {E} | | _ {2} - \\lambda \\right] _ {+}\\right)}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.808, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\( z_{I_i} \\) and \\( z_{I_j} \\) indicate the interaction representations of interaction class \\( i \\) and \\( j \\), respectively. \\( \\hat{z}_E \\) denotes the human-object entity representation. \\( [s]_+ \\triangleq \\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.41, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Video HOI Interaction-centric context reasoning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.195 + ], + "angle": 0, + "content": "\\(max(0,s)\\). \\(\\lambda\\) indicates the radius of the hypersphere, which is set to be a constant in our model. The higher value of \\(\\mathcal{U}(E,I_i)\\) suggests that the human and object entity \\(E\\) is more likely to be categorized into \\(I_{i}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.212, + 0.423, + 0.228 + ], + "angle": 0, + "content": "4.3 Learning Objective" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.233, + 0.786, + 0.279 + ], + "angle": 0, + "content": "The learning objective of our model contains two parts: (i) focal loss \\(\\mathcal{L}_{cls}\\) for interaction classification; (ii) interaction feature loss \\(\\mathcal{L}_{fea}\\) that controls the smoothness of interaction features in local region." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.28, + 0.787, + 0.416 + ], + "angle": 0, + "content": "Focal loss \\(\\mathcal{L}_{cls}\\): We employ focal loss [18] for interaction classification, mitigating the interaction class imbalance problem on model performance. For each video frame \\(v_{i}(i = 1,\\dots,T)\\), our model predicts the probability \\(\\hat{y}_i\\in \\mathbb{R}^K\\) of all the interaction classes. The corresponding ground-truth of interaction class \\(y_{i}\\in \\mathbb{R}^{K}\\) is a binary vector. For each intraction class \\(k\\), the focal loss \\(\\mathcal{L}_{cls}^{k}\\) is formulated as: \\(\\mathcal{L}_{cls}^{k} = -(1 - p_{k})^{\\gamma}\\log (p_{k})\\), where \\(\\gamma\\) is a hyperparameter to control the focusing extent, \\(p_k\\) is defined as: \\(\\{p_k = \\hat{y}_i^k\\), if \\(y_i^k = 1;p_k = 1 - \\hat{y}_i^k\\), otherwise\\}. Subsequently, the focal loss \\(\\mathcal{L}_{cls}\\) of each video frame is obtained by combining the focal loss of each individual interaction class \\(k\\): \\(\\mathcal{L}_{cls} = \\Sigma_{k = 1}^{K}\\mathcal{L}_{cls}^{k}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.787, + 0.538 + ], + "angle": 0, + "content": "Interaction feature loss \\(\\mathcal{L}_{fea}\\): We introduce interaction feature loss \\(\\mathcal{L}_{fea}\\) to control the temporal smoothness of interaction features. Our model outputs the feature of each human-object entity \\(E\\) in each frame. Inspired by [1], in order to improve the continuity, we minimize the feature distance in the same segment and maximize the distance between different segments for each subject. Denote \\(u_{E}^{t}\\) as whether the interaction will continue or change to another action for entity \\(E\\) at time \\(t\\). \\(u_{E}^{t} = 1\\) indicates the interaction will stop and change to another action at time \\(t\\) for entity \\(E\\) and \\(u_{E}^{t} = 0\\) otherwise. We minimize" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.544, + 0.786, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f e a} = \\frac {1}{2} \\sum_ {t = 0} ^ {T - 1} \\left[ (1 - u _ {E} ^ {t}) \\left(\\left| \\left| z _ {E} ^ {t} - z _ {E} ^ {t + 1} \\right| \\right| _ {2}\\right) ^ {2} + u _ {E} ^ {t} \\left(\\max \\left(L - \\left| \\left| z _ {E} ^ {t} - z _ {E} ^ {t + 1} \\right| \\right| _ {2}, 0\\right)\\right) ^ {2} \\right], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.592, + 0.786, + 0.622 + ], + "angle": 0, + "content": "where \\( L \\) is a threshold that controls the minimal feature distance when interaction changes. In total, the overall loss is written as:" + }, + { + "type": "equation", + "bbox": [ + 0.434, + 0.63, + 0.786, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c l s} + \\alpha \\mathcal {L} _ {f e a}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.654, + 0.671, + 0.67 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a hyperparameter to control the weight of each loss." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.688, + 0.787, + 0.735 + ], + "angle": 0, + "content": "Model Inference During model inference, we compute the interaction probability \\(\\hat{y}_i\\in \\mathbb{R}^K\\) for each video frame. The interaction class \\(a\\) with the highest probability is chosen as the predicted interaction for that frame: \\(a = \\arg \\max \\hat{y}_i\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.76, + 0.377, + 0.777 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.427, + 0.804 + ], + "angle": 0, + "content": "5.1 Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Datasets: We evaluate our method on MPHOI-72, Bimanual Actions, and CAD-120 datasets, each representing multi-person collaboration, one person with two" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.484, + 0.161 + ], + "angle": 0, + "content": "hands and a single hand respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.162, + 0.784, + 0.235 + ], + "angle": 0, + "content": "(I) MPHOI-72 dataset is proposed in [27], which consists of multiple humans and objects in the scene. The dataset comprises 72 videos featuring 3 human subjects and 6 objects. Within each video, 2 individuals are paired to engage in 3 distinct activities, encompassing a total of 13 sub-activities, while utilizing 2 to 4 objects" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.784, + 0.325 + ], + "angle": 0, + "content": "(II) Bimanual Actions dataset [4] is the first HOI dataset to include two hands for subjects to perform interactions which is common in reality. There are 540 videos with one person performing activities with both hands. There are 6 subjects performing 9 different activities with 10 repetitions. There are a total of 14 action labels assigned to each hand, and entity-level annotations are provided on a per-frame basis within the video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.784, + 0.37 + ], + "angle": 0, + "content": "(III) CAD-120 dataset [13] is popular for HOI recognition. It contains 120 videos with 10 activities performed by 4 participants. There are 10 human sub-activities labeled per frame." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.162, + 0.784, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.784, + 0.461 + ], + "angle": 0, + "content": "Evaluation Metric: We report \\( F_{1} \\) at \\( k \\) metric [14] with thresholds \\( k = 10\\% \\), \\( 25\\% \\), and \\( 50\\% \\). Compared to frame-based metrics which evaluate prediction on every single frame, this metric could measure prediction continuity in action segments because it views each predicted action segment as correct only when it has the Interaction over Union (IoU) with the corresponding ground truth over the threshold \\( k \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.485, + 0.455, + 0.5 + ], + "angle": 0, + "content": "5.2 Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.784, + 0.584 + ], + "angle": 0, + "content": "In the experiment, we use three layers of context fuser for Bimanual and two for CAD-120 and MPHOI. The features of humans and objects are extracted from [27] and their dimension is mapped to 768, 256, and 512 for MPHOI, Bimanual, and CAD-120 respectively. More implementation details can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.606, + 0.429, + 0.621 + ], + "angle": 0, + "content": "5.3 Quantitative results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.784, + 0.761 + ], + "angle": 0, + "content": "Multi-person HOI recognition The quantitative results of joined segmentation and label recognition of sub-activity on MPHOI-72 in Tab. 1 show the performance of our method in multi-person HOI circumstance. Our method outperforms SOTA method 2G-GCN [27] by a large margin in all the three evaluation metrics. For \\(\\mathrm{F}_1@10\\), \\(\\mathrm{F}_1@25\\) and \\(\\mathrm{F}_1@50\\) scores, our method surpasses 2G-GCN \\(23.0\\%\\), \\(23.7\\%\\) and \\(22.5\\%\\), respectively. The significant improvement achieved by our method indicates that the human fuser block in the CF module effectively improves the context-aware human representation learning under multi-person scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.838 + ], + "angle": 0, + "content": "Single-person HOI recognition We show the quantitative results for single-person HOI recognition in Tab. 2 and Tab. 3, which are performed on CAD-120 and Bimanual Actions datasets, respectively. Results in Tab. 2 show that our method performs slightly better on CAD-120 dataset compared to SOTA method 2G-GCN, with around \\(1\\%\\) improvement over 2G-GCN on all the three metrics." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.41, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Video HOI Interaction-centric context reasoning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.784, + 0.171 + ], + "angle": 0, + "content": "Table 1: The results of joined segmentation and label recognition of sub-activity on MPHOI-72." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.785, + 0.256 + ], + "angle": 0, + "content": "\\(\\Delta \\mathrm{CF}\\): removing context fuser module; \\(\\Delta \\mathrm{ISR}\\): removing interaction state reasoner module; \\(\\Delta \\mathcal{L}_{fea}\\): removing interaction feature loss; \\(\\lambda = 0\\): ablating the hypersphere by replacing it with Euclidean distance; \\(\\Delta \\mathrm{CLIP} + \\mathrm{BLIP}\\): ablating CLIP and BLIP model; Traditional Classifier: ablating the interaction-centric hypersphere by replacing it with a traditional classifier constructed with MLP and Softmax classifier. The improvements of our method over current SOTA method is highlighted with upward arrows." + }, + { + "type": "table", + "bbox": [ + 0.249, + 0.28, + 0.761, + 0.407 + ], + "angle": 0, + "content": "
ModelF1@10F1@25F1@50
ASSIGN [23]59.1 ± 12.151.0 ± 16.733.2 ± 14.0
2G-GCN [27]68.6 ± 10.460.8 ± 10.345.2 ± 6.5
Ours91.6 ± 0.9(↑ 23.0)84.5 ± 2.6(↑ 23.7)67.7 ± 2.2(↑ 22.5)
Ours (ΔCF)65.8 ± 12.457.6 ± 14.039.2 ± 12.6
Ours (ΔISR)80.1 ± 5.573.0 ± 8.255.6 ± 6.1
Ours (ΔLfea)73.5 ± 15.769.7 ± 13.348.8 ± 13.0
Ours (λ = 0)81.2 ± 0.774.8 ± 4.253.2 ± 0.3
Ours (ΔCLIP+BLIP)80.0 ± 6.773.0 ± 9.955.6 ± 7.4
Ours (Traditional Classifier)84.6 ± 7.274.7 ± 10.554.6 ± 13.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.785, + 0.509 + ], + "angle": 0, + "content": "For the Bimanual Actions dataset, our method performs as good as 2G-GCN in \\( F_{1}@10 \\) while achieves \\( 0.9\\% \\) and \\( 5.0\\% \\) higher than 2G-GCN in \\( F_{1}@25 \\) and \\( F_{1}@50 \\) score, respectively. These results indicate that our method achieves generally on par or even better performance on single-person video HOI recognition task." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.537, + 0.784, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.639, + 0.784, + 0.668 + ], + "angle": 0, + "content": "Fig. 5: Qualitative ablation study results on MPHOI-72 dataset. Major prediction errors are highlighted in red dashed boxes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.723, + 0.398, + 0.737 + ], + "angle": 0, + "content": "5.4 Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.84 + ], + "angle": 0, + "content": "In this section, we ablate the CF module and the ISR module for validating the effectiveness of these proposed components. As shown in Tab. 1, removing CF module results in more than \\(20\\%\\) \\(F_{1}\\) score drop of the three metrics in MPHOI-72 dataset, indicating the essential improvement of CF module in learning context-rich representations. Visualization results of temporal segmentation of interactions in Fig. 5 indicates that removing CF module results in incorrect" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.173 + ], + "angle": 0, + "content": "Table 2: The results of joined segmentation and label recognition of sub-activity on CAD-120. The notations are the same with Tab. 1." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.197, + 0.751, + 0.386 + ], + "angle": 0, + "content": "
ModelF1@10F1@25F1@50
rCRF [29]65.6 ± 3.261.5 ± 4.147.1 ± 4.3
Independent BiRNN [27]70.2 ± 5.564.1 ± 5.348.9 ± 6.8
ATCRF [11]72.0 ± 2.868.9 ± 3.653.5 ± 4.3
Relational BiRNN [27]79.2 ± 2.575.2 ± 3.562.5 ± 5.5
ACoLP [35]90.2 ± 2.687.4 ± 1.476.8 ± 2.3
ASSIGN [23]88.0 ± 1.884.8 ± 3.073.8 ± 5.8
2G-GCN [27]89.5 ± 1.687.1 ± 1.876.2 ± 2.8
Ours90.7 ± 2.9(↑ 0.5)88.1 ± 2.8(↑ 0.7)77.6 ± 4.7(↑ 0.8)
Ours (ΔCF)81.1 ± 4.077.0 ± 4.865.2 ± 5.6
Ours (ΔISR)88.5 ± 3.785.5 ± 3.673.9 ± 5.7
Ours (ΔLfea)89.3 ± 1.985.6 ± 2.175.9 ± 4.4
Ours (λ = 0)72.0 ± 4.465.0 ± 6.948.6 ± 6.3
Ours (ΔCLIP+BLIP)89.4 ± 2.385.5 ± 3.974.9 ± 5.7
Ours (Traditional Classifier)79.5 ± 11.073.9 ± 11.456.6 ± 12.5
" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.414, + 0.782, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.461, + 0.788, + 0.489 + ], + "angle": 0, + "content": "Fig. 6: Qualitative ablation study results on CAD-120 dataset. Major prediction errors are highlighted in red dashed boxes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.784, + 0.581 + ], + "angle": 0, + "content": "interaction predictions in both humans (highlighted in red in Fig. 5). Similarly, in CAD-120 and Bimanual Actions datasets, deleting CF module also results in massive \\( F_{1} \\) score drop. The visualization results in Fig. 6 suggests an incorrect prediction segment when removing CF module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Furthermore, we ablate hypersphere by replacing it with Euclidean distance, where \\(\\lambda = 0\\) in Eq. 6. Results in Tab. 1, 2 and 3 indicate that utilizing Euclidean distance results in at least \\(7\\%\\), \\(10\\%\\) and \\(3\\%\\) \\(F_{1}\\) scores drop in MPHOI, CAD-120 and Bimanual Actions dataset, respectively. Therefore, we conclude that Euclidean distance do not introduce HOI structure priors, ignoring valuable structure information of HOI for guiding predictions. Subsequently, we ablate CLIP and BLIP models by randomly initialize interaction features and context features. Results in Tab. 1, 2 and 3 indicate that removing CLIP and BLIP models results in some drop of model performance, but is still on par with or better than SOTA methods. Thus, it is the intricately designed structure of our model that substantiates the substantial enhancement in performance. Finally, we ablate the interaction-centric hypersphere by replacing it with a traditional classifier constructed with multi-layer perceptron (MLP) and Softmax classifier. The outcomes, as presented in Tab. 1, 2 and 3, reveal a notable decline of over \\(7\\%\\) in the \\(F_{1}\\) score within the MPHOI dataset when employing the traditional classifier. Likewise, in the CAD-120 dataset, the traditional classifier results in a substantial decrease of more than \\(11\\%\\) in the \\(F_{1}\\) score. Additionally, within" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.41, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Video HOI Interaction-centric context reasoning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 3: The results of joined segmentation and label recognition of sub-activity on Bimanual Actions. The notations are the same with Tab. 1." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.197, + 0.751, + 0.376 + ], + "angle": 0, + "content": "
ModelF1@10F1@25F1@50
Dreher et al. [4]40.6 ± 7.234.8 ± 7.122.2 ± 5.7
Independet BiRNN [27]74.7 ± 7.072.0 ± 7.061.8 ± 7.3
Relational BiRNN [27]77.7 ± 3.975.0 ± 4.264.8 ± 5.3
ASSIGN [23]84.0 ± 2.081.2 ± 2.068.5 ± 3.3
2G-GCN [27]85.0 ± 2.282.0 ± 2.669.2 ± 3.1
Ours85.0 ± 2.582.9 ± 2.9(↑ 0.9)74.2 ± 4.3(↑ 5.0)
Ours (ΔCF)82.5 ± 5.080.5 ± 5.571.1 ± 7.0
Ours (ΔISR)84.1 ± 2.381.8 ± 2.873.0 ± 3.7
Ours (ΔLfea)84.5 ± 4.682.0 ± 5.271.8 ± 6.9
Ours (λ = 0)76.7 ± 5.274.3 ± 6.065.2 ± 6.3
Ours (ΔCLIP+BLIP)84.3 ± 1.481.8 ± 1.873.2 ± 2.7
Ours (Traditional Classifier)82.0 ± 3.679.8 ± 4.171.0 ± 5.6
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.464 + ], + "angle": 0, + "content": "the Bimanual Actions dataset, the traditional classifier induces a decline exceeding \\(2\\%\\) in the \\(F_{1}\\) score. These findings unanimously underscore the efficacy of modeling HOI manifold structures by the hypersphere module." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.493, + 0.495, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.492, + 0.807, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.669 + ], + "angle": 0, + "content": "Fig. 7: (A) Visualization of interaction-centric hyperspheres, the learned interaction representations (large points locate at the center of hyperspheres) and human-object entity representations (small points surrounding the hyperspheres) in embedding space. Selected frame samples are shown for each interaction-centric hypersphere. (B) 2D visualization of interaction-centric hyperspheres." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.722, + 0.416, + 0.737 + ], + "angle": 0, + "content": "5.5 Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We show some visualization results on MPHOI in Fig 8 to compare our method with SOTA method 2G-GCN. The red highlighted boxes indicate that 2G-GCN tend to generate unreasonable interaction predictions, while our method generates more reasonable interaction predictions. The visualization results in Fig. 9 show similar prediction pattern where 2G-GCN predicted some unreasonable short segments (highlighted in red boxes) while our method predicts more accu" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "rately. We also visualize the interaction-centric hypersphere, the learned interaction representations and the human-object entity representations in embedding space in Fig. 7. Results in Fig. 7 (A) show that human-object entity representations (small dots in Fig. 7 (A)) belonging to the specific interaction class locates near the surface of the corresponding hypersphere. Similar with Fig. 7 (A), we also visualize the intraction-centric hyerspheres in 2D sapce in Fig. 7 (B). Results in Fig. 7 (B) indicate that HOIs distribute around the sphere of different interaction classes. The distributions of HOIs on each sphere also differ a lot, where the sphere belongs to the solve interaction class captures HOIs on the upper part of its sphere, while the sphere belongs to the cut interaction class captures HOIs on the lower part of its sphere. These results suggest that our model successfully model the manifold structure of HOI." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.357, + 0.784, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.456, + 0.788, + 0.486 + ], + "angle": 0, + "content": "Fig. 8: Visualization results on MPHOI-72 dataset. Major prediction errors are highlighted in red dashed boxes." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.543, + 0.784, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.588, + 0.788, + 0.618 + ], + "angle": 0, + "content": "Fig. 9: Visualization results on CAD-120 dataset. Major prediction errors are highlighted in red dashed boxes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.67, + 0.36, + 0.687 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.702, + 0.788, + 0.824 + ], + "angle": 0, + "content": "In this work, we propose an interaction-centric spatio-temporal context reasoning network for multi-person video HOI recognition. Specifically, we propose a context fuser and an interaction state reasoner to learn spatio-temporal context-rich and reasoning-aware entity representations. We further represent HOI components with an interaction-centric hypersphere for HOI classification. Experiment results show that our method outperforms SOTA method by more than \\(22\\%\\) \\(F_{1}\\) score in multi-person scenarios, and achieves competitive results on single-person cases compared to SOTA methods." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.41, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Video HOI Interaction-centric context reasoning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.302 + ], + "angle": 0, + "content": "This material is based upon work supported under the AI Research Institutes program by National Science Foundation and the Institute of Education Sciences, U.S. Department of Education through Award #2229873 - AI Institute for Transforming Education for Children with Speech and Language Processing Challenges. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation, the Institute of Education Sciences, or the U.S. Department of Education." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.329, + 0.325, + 0.345 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.363, + 0.787, + 0.419 + ], + "angle": 0, + "content": "1. Chopra, S., Hadsell, R., LeCun, Y.: Learning a similarity metric discriminatively, with application to face verification. In: 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05). vol. 1, pp. 539-546 vol. 1 (2005). https://doi.org/10.1109/CVPR.2005.2029" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.421, + 0.787, + 0.46 + ], + "angle": 0, + "content": "2. Deng, S., Zhang, N., Chen, H., Tan, C., Huang, F., Xu, C., Chen, H.: Low-resource extraction with knowledge-aware pairwise prototype learning. Knowledge-Based Systems 235, 107584 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.462, + 0.787, + 0.517 + ], + "angle": 0, + "content": "3. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.518, + 0.787, + 0.572 + ], + "angle": 0, + "content": "4. Dreher, C.R.G., Wächter, M., Asfour, T.: Learning object-action relations from bimanual human demonstration using graph networks. IEEE Robotics and Automation Letters (RA-L) 5(1), 187-194 (2020). https://doi.org/10.1109/LRA.2019.2949221 3, 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.787, + 0.63 + ], + "angle": 0, + "content": "5. Gao, C., Xu, J., Zou, Y., Huang, J.B.: Drg: Dual relation graph for human-object interaction detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16. pp. 696-712. Springer (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.631, + 0.787, + 0.672 + ], + "angle": 0, + "content": "6. Gao, K., Chen, L., Zhang, H., Xiao, J., Sun, Q.: Compositional prompt tuning with motion cues for open-vocabulary video relation detection. In: The Eleventh International Conference on Learning Representations (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.787, + 0.727 + ], + "angle": 0, + "content": "7. Gupta, T., Schwing, A., Hoiem, D.: No-frills human-object interaction detection: Factorization, layout encodings, and training techniques. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9677-9685 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.728, + 0.787, + 0.784 + ], + "angle": 0, + "content": "8. Iftekhar, A., Chen, H., Kundu, K., Li, X., Tighe, J., Modolo, D.: What to look at and where: Semantic and spatial refined transformer for detecting human-object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5353-5363 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "9. Kim, B., Choi, T., Kang, J., Kim, H.J.: Uniondet: Union-level detector towards real-time human-object interaction detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 498-514. Springer (2020) 4" + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.363, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "10. Kim, B., Lee, J., Kang, J., Kim, E.S., Kim, H.J.: Hotr: End-to-end human-object interaction detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 74-83 (June 2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "11. Koppula, H.S., Saxena, A.: Anticipating human activities using object affordances for reactive robotic response. IEEE transactions on pattern analysis and machine intelligence 38(1), 14-29 (2015) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.285 + ], + "angle": 0, + "content": "12. Koppula, H.S., Gupta, R., Saxena, A.: Learning human activities and object affordances from rgb-d videos. The International journal of robotics research 32(8), 951-970 (2013) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.34 + ], + "angle": 0, + "content": "13. Koppula, H.S., Gupta, R., Saxena, A.: Learning human activities and object affordances from rgb-d videos. Int. J. Rob. Res. 32(8), 951-970 (jul 2013). https://doi.org/10.1177/0278364913478446, https://doi.org/10.1177/0278364913478446 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.787, + 0.383 + ], + "angle": 0, + "content": "14. Lea, C., Flynn, M.D., Vidal, R., Reiter, A., Hager, G.D.: Temporal convolutional networks for action segmentation and detection. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 156-165 (2017) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.424 + ], + "angle": 0, + "content": "15. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.48 + ], + "angle": 0, + "content": "16. Li, Y.L., Zhou, S., Huang, X., Xu, L., Ma, Z., Fang, H.S., Wang, Y., Lu, C.: Transferable interactiveness knowledge for human-object interaction detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3585-3594 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.535 + ], + "angle": 0, + "content": "17. Liao, Y., Liu, S., Wang, F., Chen, Y., Qian, C., Feng, J.: Ppdm: Parallel point detection and matching for real-time human-object interaction detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 482-490 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.577 + ], + "angle": 0, + "content": "18. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.619 + ], + "angle": 0, + "content": "19. Liu, X., Li, Y.L., Wu, X., Tai, Y.W., Lu, C., Tang, C.K.: Interactiveness field in human-object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20113-20122 (2022) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.787, + 0.66 + ], + "angle": 0, + "content": "20. Luan, T., Wang, Y., Zhang, J., Wang, Z., Zhou, Z., Qiao, Y.: Pc-hmr: Pose calibration for 3d human mesh recovery from 2d images/videos. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 2269-2276 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.787, + 0.715 + ], + "angle": 0, + "content": "21. Luan, T., Zhai, Y., Meng, J., Li, Z., Chen, Z., Xu, Y., Yuan, J.: High fidelity 3d hand shape reconstruction via scalable graph frequency decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16795-16804 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.743 + ], + "angle": 0, + "content": "22. Mettes, P., Van der Pol, E., Snoek, C.: Hyperspherical prototype networks. Advances in neural information processing systems 32 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.797 + ], + "angle": 0, + "content": "23. Morais, R., Le, V., Venkatesh, S., Tran, T.: Learning asynchronous and sparse human-object interaction in videos. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16041-16050 (2021) 2, 4, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "24. Nagarajan, T., Feichtenhofer, C., Grauman, K.: Grounded human-object interaction hotspots from video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8688-8697 (2019) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.41, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Video HOI Interaction-centric context reasoning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.204 + ], + "angle": 0, + "content": "25. Park, J., Park, J.W., Lee, J.S.: Viplo: Vision transformer based pose-conditioned self-loop graph for human-object interaction detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17152-17162 (June 2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.205, + 0.788, + 0.245 + ], + "angle": 0, + "content": "26. Qi, S., Wang, W., Jia, B., Shen, J., Zhu, S.C.: Learning human-object interactions by graph parsing neural networks. In: Proceedings of the European Conference on Computer Vision (ECCV) (September 2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.246, + 0.788, + 0.3 + ], + "angle": 0, + "content": "27. Qiao, T., Men, Q., Li, F.W., Kubotani, Y., Morishima, S., Shum, H.P.: Geometric features informed multi-person human-object interaction recognition in videos. In: European Conference on Computer Vision. pp. 474-491. Springer (2022) 2, 3, 4, 6, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.357 + ], + "angle": 0, + "content": "28. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.384 + ], + "angle": 0, + "content": "29. Sener, O., Saxena, A.: rcrf: Recursive belief estimation over crfs in rgb-d activity videos. In: Robotics: Science and systems (2015) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.425 + ], + "angle": 0, + "content": "30. Sunkesula, S.P.R., Dabral, R., Ramakrishnan, G.: Lighten: Learning interactions with graph and hierarchical temporal networks for hoi in videos. In: Proceedings of the 28th ACM international conference on multimedia. pp. 691-699 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.481 + ], + "angle": 0, + "content": "31. Tamura, M., Ohashi, H., Yoshinaga, T.: Qpic: Query-based pairwise human-object interaction detection with image-wide contextual information. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10410-10419 (June 2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.522 + ], + "angle": 0, + "content": "32. Tu, D., Sun, W., Min, X., Zhai, G., Shen, W.: Video-based human-object interaction detection from tubelet tokens. Advances in Neural Information Processing Systems 35, 23345-23357 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.578 + ], + "angle": 0, + "content": "33. Wang, T., Anwer, R.M., Khan, M.H., Khan, F.S., Pang, Y., Shao, L., Laaksonen, J.: Deep contextual attention for human-object interaction detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5694-5702 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.578, + 0.788, + 0.619 + ], + "angle": 0, + "content": "34. Xi, N., Meng, J., Yuan, J.: Chain-of-look prompting for verb-centric surgical triplet recognition in endoscopic videos. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 5007-5016 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.66 + ], + "angle": 0, + "content": "35. Xi, N., Meng, J., Yuan, J.: Open set video hoi detection from action-centric chain-of-look prompting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3079-3089 (2023) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.716 + ], + "angle": 0, + "content": "36. Zhang, Y., Pan, Y., Yao, T., Huang, R., Mei, T., Chen, C.W.: Exploring structure-aware transformer over interaction proposals for human-object interaction detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19548-19557 (2022) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.716, + 0.788, + 0.772 + ], + "angle": 0, + "content": "37. Zhong, X., Qu, X., Ding, C., Tao, D.: Glance and gaze: Inferring action-aware points for one-stage human-object interaction detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13234-13243 (2021) 4" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.772 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition/0a79d390-9d09-4042-af06-6b62e335816f_origin.pdf b/2024/Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition/0a79d390-9d09-4042-af06-6b62e335816f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5e4db23dc8eb7aa542a9b6df4dde6394408373e9 --- /dev/null +++ b/2024/Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition/0a79d390-9d09-4042-af06-6b62e335816f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c14cd5dd1f2ea6e0bf035717b1d04c8ab0d51af6ac7dff56c813b7e3a2350ba9 +size 6665388 diff --git a/2024/Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition/full.md b/2024/Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1f1ec2f7a888f31e1beb69708dc0314f8402650b --- /dev/null +++ b/2024/Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition/full.md @@ -0,0 +1,277 @@ +# Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition + +Yisong Wang $^{1*}$ , Nan Xi $^{2*}$ , Jingjing Meng $^{3}$ , and Junsong Yuan $^{2}$ + +$^{1}$ School of EECS, Peking University + +$^{2}$ Department of Computer Science and Engineering, University at Buffalo + +3 Amazon Inc. + +wys05y98@gmail.com, {nanxi, jsyuany}@buffalo.edu, jingjing.meng1@gmail.com + +Abstract. Understanding human-object interaction (HOI) in videos represents a fundamental yet intricate challenge in computer vision, requiring perception and reasoning across both spatial and temporal domains. Despite previous success of object detection and tracking, multi-person video HOI recognition still faces two major challenges: (1) the three facets of HOI (human, objects, and the interactions that bind them) exhibit interconnectedness and exert mutual influence upon one another. (2) the complexity of multi-person multi-object combinations in spatio-temporal interaction. To address them, we design a spatio-temporal context fuser to better model the interactions among persons and objects in videos. Furthermore, to equip the model with temporal reasoning capacity, we propose an interaction state reasoner module on top of context fuser. Considering the interaction is a key element to bind human and object, we propose an interaction-centric hypersphere in the feature embedding space to model each category of interaction. It helps to learn the distribution of HOI samples belonging to the same interactions on the hypersphere. After training, each interaction prototype sphere will fit the testing HOI sample to determine the HOI classification result. Empirical results on multi-person video HOI dataset MPHOI-72 indicate that our method remarkably surpasses state-of-the-art (SOTA) method by more than $22\%$ $F_{1}$ score. At the same time, on single-person datasets Bimanual Actions (single-human two-hand HOI) and CAD-120 (single-human HOI), our method achieves on par or even better results compared with SOTA methods. Source code is released at the following link: https://github.com/southnx/IcH-Vid-HOI. + +Keywords: interaction-centric $\cdot$ spatio-temporal context reasoning $\cdot$ multi-person human-object interaction + +![](images/4b34982c1398083e87608267fe526441237b01c8e66eac06865a957e2a210330.jpg) +Multi-person Video Human-Object Interaction + +![](images/1f3ae3a97705067a470cd5512afbe464c84545b57759bc5c4d7834d8278d6108.jpg) +Fig. 1: Examples of multi-person video HOI. Three categories of interactions are presented, including co-working, cheering, hair_cutting. + +![](images/3048e4587dd2851d296f2750bfb7e5b0b660ac7cf162769eba993122e9f1f017.jpg) + +# 1 Introduction + +Video-based Human-Object Interaction (HOI) recognition aims to identify the interactions occurring between human and object entities within video frames. Precisely recognizing HOIs in real world scenarios is essential for a bunch of applications, such as assisting patients by recognizing daily activities and predicting pedestrian movements to avoid accidents for autonomous vehicles. + +Most existing HOI recognition research focuses on static images [19, 36], with much less attention on video-based HOI recognition. Video-based HOI recognition is more demanding than image-based HOI recognition due to the necessity of comprehending complex spatio-temporal dynamics and reasoning about multi-human and multi-object interaction combinations. The complexity is further exacerbated when dealing with multi-person circumstances, as examples shown in Fig. 1. In such cases, various human and object entities mutually influence each other, resulting in intricate interdependencies within the scene. Additionally, the three components (human, object and interactions that the entities are involved in) of HOI exhibit interwind structures, e.g., the possible interactions that can occur within a scene given a human and an object. However, current video HOI recognition methods [23, 27] do not fully explore such inherent structural nature of HOI components. Instead, they often opt for disentangled representations for each component, which may bring suboptimal representation capabilities. + +To overcome the aforementioned limitations, we introduce an interaction-centric spatio-temporal context reasoning approach for representing video HOIs. For the purpose of enhancing the awareness of complex HOI structures in our representations, we introduce the Context Fuser (CF), which encodes both entity representations and interaction representations. Moreover, to empower our model with the ability to reason about interaction state transitions across video frames, we propose the Interaction State Reasoner (ISR) module for generating interaction representations. In addition, we employ a bidirectional Gated Recurrent Unit (BiGRU) to model temporal dynamics across video frames. This multi-level representation learning framework not only facilitates effective exploration of the interdependencies among structured HOI components but also empowers the model with interaction reasoning capabilities in both spatial and temporal domains. To compute the probability of an HOI in the scene, we introduce an interaction-centric hypersphere. This approach leverages the concept of a hypersphere, where the interaction is located at its center, while the human and object entities involved in that interaction are situated on the hypersphere's + +surface. We assume that human-object entities belonging to each interaction class are located on their respective hyperspheres, specific to that interaction class. + +Concretely, we generate context-rich and reasoning-aware video HOI representations through three key components: the Context Fuser (CF), Interaction State Reasoner (ISR) and Bidirectional GRU (BiGRU). The CF module integrates context information from human-object entities and interactions. It comprises three fuser blocks for humans, objects, and interactions. The Object Fuser Block processes local video frame data, enhancing object features. The Interaction Fuser Block combines human and object representations with interaction-specific characteristics. Additionally, in multi-person scenes, the Human Fuser Block captures human representations influenced by others. This approach fosters comprehensive HOI representations via effective context fusion. To facilitate interaction reasoning, we place the ISR module on top of the context fuser module, yielding entity representations capable of capturing interaction transition dynamics. These entity representations are then input into the BiGRU module to model temporal dynamics across video frames, thereby ensuring a comprehensive understanding of the evolving context and interactions within the video data. Finally, we determine interaction classes in each frame with the interaction-centric hypersphere, computing the probability of human-object entities belonging to specific interaction classes. + +We assess our model's performance on three video-based HOI datasets: MPHOI-72 [27] (multi-person), Bimanual Actions [4] and CAD-12 [12] (single-person). Our results highlight our model's superiority in multi-person scenarios, achieving an impressive over $22\%$ $F_{1}$ score improvement over the current state-of-the-art (SOTA). In single-person scenarios, our method delivers on par or even better performance compared to the current SOTA method. Our major contributions are summarized as follows: + +- To represent inherent HOI manifold structures, we propose an interaction-centric hypersphere representation scheme. This scheme explicitly guide the process of learning intrinsic structural nature of HOI and elucidates the interdependencies among its components. +- To learn context-rich and reasoning-aware entity representations, we introduce context fuser and interaction state reasoning modules. This enhancement results in entity representations that are highly suitable for video-based HOI tasks. +- Extensive experiment results showcase that our method achieves SOTA performance with a huge improvement of more than $22\%$ $F_{1}$ score over existing methods in multi-person scenario. Additionally, our model achieves competitive results in single-person cases compared to SOTA method. + +# 2 Related works + +HOI detection in images: HOI detection in images aims at understanding interactions in images between humans and objects. Different methods have been + +proposed in previous studies. Some works propose Convolutional Neural Networks (CNN)-based methods which can be further divided into one-stage methods [9,17,37] and two-stage methods [5,7,16,33]. However, these methods usually lack of ability to capture global context information. Recently, Transformer-based models [8,10,31,36] became the main approach for the HOI task. Following the architecture of DETR, these models achieved superior performance on HOI detection. Moreover, some works also utilize graph [25] and interactiveness field [19] to achieve better performance. These various approaches to image HOI provide the fundamentals for video HOI recognition. + +HOI recognition in videos: Video HOI recognition is the foundation of many real-world applications, including understanding surgical activities in the operation room [34] and guiding 3D human reconstruction [20, 21]. Video-based HOI recognitions have to deal with both spatial and temporal reasoning. Before the use of neural networks, some early studies formulated this task using the Markov model [12] to utilize temporal cues. In [24], HOI hotspots in videos are learned in a novel approach, with two networks trained jointly to capture spatial regions where actions happen. While [6] focuses on visual relation prediction in open-vocabulary with pretrained visual-language models. Recent works have used Recurrent Neural Networks (RNN) combined with Graph Neural Networks (GNN) [23, 26, 27, 30] to predict human-object relations in videos. Inspired by ViT [3], some works also propose Transformer-based methods to reason spatial relations better [32]. However, RNN-based models usually require complex training strategies or long training time in order to achieve the best performance. Moreover, when multiple persons are involved in an activity jointly, these methods lack the ability to model their collaboration, resulting in poor performance when the interactions are performed by multiple persons. + +Hyperspheres for class representation: Hyperspheres have been demonstrated to be an effective approach to model class representation [2, 22]. Geometrical modeling strategy has been proposed in [2], where the effectiveness has been confirmed. This approach proves advantageous for capturing and representing enriched class-level information, particularly well-suited for creating measurements in Euclidean space. Consequently, it is naturally adaptable to structured prediction tasks. + +# 3 Motivation + +Video HOI recognition task involves identifying both the human and object entities engaged in an interaction across a sequence of video frames. This task encompasses spatial and temporal aspects, as it requires understanding the relationships between humans, objects, and their interactions over time. However, current methods for video HOI recognition often neglect this crucial dependency structure, resulting in the separation of learned representations associated with humans, objects, and their interactions, ultimately compromising their representational accuracy. To address this issue, we propose a novel approach featuring a context fuser and an interaction state reasoner to capture spatio-temporal + +![](images/4a17d980bda599dd9d63d0945d0cbe99e4d88b1839ba6ff64e60f5fdafcea4fd.jpg) +Fig. 2: Model Overview. Each video frame is applied with a backbone for feature extraction. Subsequently, a context fuser and interaction state reasoner is employed for learning interaction representations $z_{I}$ and human-object entity representations $\{\hat{z}_E^t\}_{t=1}^T$ . Bidirectional GRU is further utilized to model temporal dynamics across video frames. The final predicted interaction probability $\{u^t\}_{t=1}^T$ is computed from the interaction-centric hypersphere. The dimension of each hypersphere is $d$ , which is the same with the feature dimension in our model. + +contextual representations. Additionally, we introduce an interaction-centric hypersphere to represent HOI manifold structure and model the inter-dependency among predicted humans, objects, and interactions. + +# 4 Method + +# 4.1 Problem formulation + +For a video dataset $\mathcal{V}$ , given a video clip $V\in \mathcal{V}$ containing $T$ video frames $\{v_{1},\dots,v_{T}\}$ , video HOI recognition aims to predict the temporal segmentation of interactions between human and object entities across all the video frames. Formally, we aim to learn an HOI recognition model $\mathcal{M}$ that outputs the segmentation of human's sub-activity $\{s_n\}_{n = 1}^N$ in each frame, where $N$ is the number of human sub-activity segments. Each segment $s_n$ is represented as an interval from its start time $t_n$ to end time $t_{n + 1}$ : $s_n = [t_n,t_{n + 1})$ . The start and end time of each segment are determined from interaction probability prediction $\{u^t\}_{t = 1}^T$ of each frame, where $u^{t}\in \mathbb{R}^{K}$ , $K$ is the number of possible interaction classes. + +# 4.2 Model Design + +In the following section, we introduce our interaction-centric hypersphere reasoning model for video HOI recognition in detail. As shown in Fig. 2, our major idea is to construct a hypersphere to represent each HOI in the scene. For each hypersphere, the interaction locates at the center of the hypersphere, while the corresponding human-object entity belongs to that interaction locates at the surface of the hypersphere. We construct Context Fuser (CF) module to learn context-rich human-object entity representations. For the aim of enabling model with reasoning ability over interaction state transitions, we propose Interaction State Reasoner (ISR) to reason on whether the current interaction will be continued or stopped. To model the temporal dynamics of HOI in videos, we update human-object entity representations $\{\hat{z}_E^t\}_{t=1}^T$ along the temporal domain with bidirectional GRU (BiGRU). Finally, predicted interaction class probability $u^t$ of each frame is computed from the interaction-centric hypersphere. + +Context fuser For a sequence of video frames $\{v_{t}\}_{t = 1}^{T}$ , we follow 2G-GCN [27] to extract feature of humans and objects from backbone network. The extracted human features $z_{H}\in \mathbb{R}^{d}$ contain both bounding box information and skeleton keypoint information, where $d$ indicates the feature dimension. Object features $z_{O}\in \mathbb{R}^{d}$ contain only bounding box information. + +We design a context fuser (CF) module shown in Fig. 3 to generate human-object entity representations $\{z_E^t\}_{t=1}^T$ based on human, object and contextual information. In multi-person circumstances, CF contains object fuser block, interaction fuser block and human fuser block sequentially. First, we design an object fuser block to incorporate object representations in local temporal regions into current object representations, generating learned object representation $\hat{z}_{O_1}$ and $\hat{z}_{O_2}$ : + +$$ +\hat {z} _ {O _ {1}} = F F N \left(S A \left(Q = z _ {O _ {1}} ^ {t}; K, V = z _ {O _ {1}} ^ {\bar {t}}\right) + z _ {O _ {1}} ^ {t}\right), \tag {1} +$$ + +$$ +\hat {z} _ {O _ {2}} = F F N (S A (Q = z _ {O _ {2}} ^ {t}; K, V = z _ {O _ {2}} ^ {t}) + z _ {O _ {2}} ^ {t}), +$$ + +where $SA$ indicates self-attention, $FFN$ is feed forward network, $z_{O_1}^t \in \mathbb{R}^d$ and $z_{O_2}^t \in \mathbb{R}^d$ are the initial object features, while $z_{O_1}^{\bar{t}} \in \mathbb{R}^{20d}$ and $z_{O_2}^{\bar{t}} \in \mathbb{R}^{20d}$ are stacked object features from a local time window, $\bar{t} \in [t - 10,t + 10)$ . Subsequently, for all the $K$ possible interactions $\{I_i\}_{i = 1}^K$ as shown in Fig. 3, each interaction class $I_i$ is prompted as a sentence $s =$ "The human is [interact]ing in the scene.", where [interact] indicates the specific interaction class. Then the sentence is applied with the text encoder $(\mathcal{F}_T)$ of large-scale vision-language model CLIP [28] to initialize the interaction feature $z_I = \mathcal{F}_T(s) \in \mathbb{R}^d$ . We also generate context feature $z_C \in \mathbb{R}^d$ to represent the semantic information of each frame. Specifically, we extract frame caption $c_i$ from video frame $i(i = 1,\dots,T)$ with BLIP [15] model and apply the caption with CLIP model to extract text embedding. In order to adapt the human $(z_{H_1},z_{H_2})$ features, object $(z_{O_1},z_{O_2})$ features and context feature $(z_C)$ to the specific interaction feature $z_I$ , we construct an interaction fuser block that contains a cross-attention (CA) module + +![](images/a86e4cc1836c9ef67305448c9de403998b65e0af9da4a47905a32c0086c54aff.jpg) +Fig.3: Context fuser. In multi-person scenario (left), the context fuser consists of object fuser block, interaction fuser block and human fuser block. In single-person case (right), only object fuser block and interaction fuser block are employed. Detailed descriptions of the architecture and annotations can be found in Sec. 4.2. + +![](images/924c78e11e15a31f724848add983b0b0814c26e3bd70240c942f225ab99bc0c3.jpg) + +followed by a Feed Forward Network (FFN), generating interaction-aware human features $(z_{H_1}' \in \mathbb{R}^d, z_{H_2}' \in \mathbb{R}^d)$ , object features $(z_{O_1}' \in \mathbb{R}^d, z_{O_2}' \in \mathbb{R}^d)$ and context feature $(z_C' \in \mathbb{R}^d)$ . Furthermore, to model the influence between the two humans in multi-person scenarios, we construct a human fuser block, featuring the same architecture with interaction fuser block. For the first human (Human1), the updated human feature $\hat{z}_{H_1} \in \mathbb{R}^d$ is generated as: + +$$ +\hat {z} _ {H _ {1}} = F F N (C A (Q = S A (z _ {H _ {1}} ^ {\prime}); K, V = z _ {H _ {2}} ^ {\prime}) + S A (z _ {H _ {1}} ^ {\prime})). \qquad (2) +$$ + +Human2 feature $\hat{z}_{H_2} \in \mathbb{R}^d$ is generated in the same way as Eq. 2. Finally, the human-object entity representation $z_{E_1} \in \mathbb{R}^d$ of Human1 is computed by max-pooling operation over all the $d$ dimensions of the four representations shown in Eq. 3: + +$$ +z _ {E _ {1}} = M a x P o o l (\hat {z} _ {H _ {1}}, z _ {O _ {1}} ^ {\prime}, z _ {O _ {2}} ^ {\prime}, z _ {C} ^ {\prime}). \qquad (3) +$$ + +The human-object entity representation $z_{E_2} \in \mathbb{R}^d$ of Human2 is computed with similar approach as Eq. 3. The CF module for single-person cases are similar with multi-person, except that the human fuser block is removed and there is only one human feature as query to be fed into the interaction fuser block. + +Interaction State Reasoner To augment the model's ability in interaction state transition reasoning, we introduce an Interaction State Reasoner (ISR) + +module following CF module. ISR module explicitly empowers the model to determine whether the current interaction should persist or transit to another interaction. Specifically, as shown in Fig. 4, at each time $t$ , the two possible states state1 and state2 represent "continue" or "stop" of an interaction, respectively. Each state is prompted as one sentence, where $s_1 =$ "This interaction is going to continue." and $s_2 =$ "This interaction is going to stop and change to another interaction." Then the embeddings of the two states $z_{state_1}$ and $z_{state_2}$ are generated from CLIP [28] text encoder $\mathcal{F}_T$ : $z_{state_1} = \mathcal{F}_T(s_1) \in \mathbb{R}^d$ , $z_{state_2} = \mathcal{F}_T(s_2) \in \mathbb{R}^d$ . Interaction state embeddings $z_{state_1}$ and $z_{state_2}$ are further fed to a reasoner block (shown in Fig. 4) together with the interaction embedding $z_t^t \in \mathbb{R}^d$ at time $t$ , generating state-informed interaction embeddings $\hat{z}_{state_1}$ and $\hat{z}_{state_2}$ . The reasoner block contains a FFN and a State Interpolation (SI) module. The SI module generates the weights $\omega_1, \omega_2$ for the two interaction states in the following approach: + +$$ +\omega_ {1}, \omega_ {2} = S o f t m a x (F F N (z _ {E} ^ {t}) \cdot [ z _ {s t a t e _ {1}} ^ {\top}, z _ {s t a t e _ {2}} ^ {\top} ]), \tag {4} +$$ + +where $\top$ indicates transpose operation. Subsequently, the final human-object entity representation $\hat{z}_E^t\in \mathbb{R}^d$ at time $t$ is generated by interpolate over the current entity representation $z_{E}^{t}$ at time $t$ and the entity representation $z_{E}^{t - 1}$ at time $t - 1$ : + +$$ +\hat {z} _ {E} ^ {t} = \omega_ {1} \cdot z _ {E} ^ {t - 1} + \omega_ {2} \cdot z _ {E} ^ {t}. \tag {5} +$$ + +Consequently, the generated human-object entity representation $\hat{z}_E^t$ is able to reason on the possible future interaction transitions. + +Interaction-centric Hypersphere With the above generated human-object entity representation $\hat{z}_E^t$ and interaction representation $z_I$ , we need to calculate the probability of a human-object entity $E$ categorizing into the interaction class $I_i$ . To that end, we design an interaction-centric hypersphere with interaction at the center of hypersphere and human-object entity at its surface. We call it "hypersphere" because this is a sphere in the $\mathbb{R}^d$ space, where $d$ is the feature dimension in our model, which is obviously larger than 3. This hypersphere design models the manifold structure of HOI, which is in charge of the specific interaction class. Concretely, we employ a hyperspherical measurement: + +![](images/ae395d494ebb19a6d17892720c327e86089036fbc4fdea5550decaa21338a7eb.jpg) +Fig. 4: Interaction State Reasoner. $\odot$ indicates inner product. + +$$ +\mathcal {U} (E, I _ {i}) = \frac {\exp \left(- \left[ | | z _ {I _ {i}} - \hat {z} _ {E} | | _ {2} - \lambda \right] _ {+}\right)}{\sum_ {j = 1} ^ {K} \exp \left(- \left[ | | z _ {I _ {j}} - \hat {z} _ {E} | | _ {2} - \lambda \right] _ {+}\right)}, \tag {6} +$$ + +where $z_{I_i}$ and $z_{I_j}$ indicate the interaction representations of interaction class $i$ and $j$ , respectively. $\hat{z}_E$ denotes the human-object entity representation. $[s]_+ \triangleq$ + +$max(0,s)$ . $\lambda$ indicates the radius of the hypersphere, which is set to be a constant in our model. The higher value of $\mathcal{U}(E,I_i)$ suggests that the human and object entity $E$ is more likely to be categorized into $I_{i}$ . + +# 4.3 Learning Objective + +The learning objective of our model contains two parts: (i) focal loss $\mathcal{L}_{cls}$ for interaction classification; (ii) interaction feature loss $\mathcal{L}_{fea}$ that controls the smoothness of interaction features in local region. + +Focal loss $\mathcal{L}_{cls}$ : We employ focal loss [18] for interaction classification, mitigating the interaction class imbalance problem on model performance. For each video frame $v_{i}(i = 1,\dots,T)$ , our model predicts the probability $\hat{y}_i\in \mathbb{R}^K$ of all the interaction classes. The corresponding ground-truth of interaction class $y_{i}\in \mathbb{R}^{K}$ is a binary vector. For each intraction class $k$ , the focal loss $\mathcal{L}_{cls}^{k}$ is formulated as: $\mathcal{L}_{cls}^{k} = -(1 - p_{k})^{\gamma}\log (p_{k})$ , where $\gamma$ is a hyperparameter to control the focusing extent, $p_k$ is defined as: $\{p_k = \hat{y}_i^k$ , if $y_i^k = 1;p_k = 1 - \hat{y}_i^k$ , otherwise\}. Subsequently, the focal loss $\mathcal{L}_{cls}$ of each video frame is obtained by combining the focal loss of each individual interaction class $k$ : $\mathcal{L}_{cls} = \Sigma_{k = 1}^{K}\mathcal{L}_{cls}^{k}$ . + +Interaction feature loss $\mathcal{L}_{fea}$ : We introduce interaction feature loss $\mathcal{L}_{fea}$ to control the temporal smoothness of interaction features. Our model outputs the feature of each human-object entity $E$ in each frame. Inspired by [1], in order to improve the continuity, we minimize the feature distance in the same segment and maximize the distance between different segments for each subject. Denote $u_{E}^{t}$ as whether the interaction will continue or change to another action for entity $E$ at time $t$ . $u_{E}^{t} = 1$ indicates the interaction will stop and change to another action at time $t$ for entity $E$ and $u_{E}^{t} = 0$ otherwise. We minimize + +$$ +\mathcal {L} _ {f e a} = \frac {1}{2} \sum_ {t = 0} ^ {T - 1} \left[ (1 - u _ {E} ^ {t}) \left(\left| \left| z _ {E} ^ {t} - z _ {E} ^ {t + 1} \right| \right| _ {2}\right) ^ {2} + u _ {E} ^ {t} \left(\max \left(L - \left| \left| z _ {E} ^ {t} - z _ {E} ^ {t + 1} \right| \right| _ {2}, 0\right)\right) ^ {2} \right], \tag {7} +$$ + +where $L$ is a threshold that controls the minimal feature distance when interaction changes. In total, the overall loss is written as: + +$$ +\mathcal {L} = \mathcal {L} _ {c l s} + \alpha \mathcal {L} _ {f e a}, \tag {8} +$$ + +where $\alpha$ is a hyperparameter to control the weight of each loss. + +Model Inference During model inference, we compute the interaction probability $\hat{y}_i\in \mathbb{R}^K$ for each video frame. The interaction class $a$ with the highest probability is chosen as the predicted interaction for that frame: $a = \arg \max \hat{y}_i$ . + +# 5 Experiments + +# 5.1 Experimental setup + +Datasets: We evaluate our method on MPHOI-72, Bimanual Actions, and CAD-120 datasets, each representing multi-person collaboration, one person with two + +hands and a single hand respectively. + +(I) MPHOI-72 dataset is proposed in [27], which consists of multiple humans and objects in the scene. The dataset comprises 72 videos featuring 3 human subjects and 6 objects. Within each video, 2 individuals are paired to engage in 3 distinct activities, encompassing a total of 13 sub-activities, while utilizing 2 to 4 objects +(II) Bimanual Actions dataset [4] is the first HOI dataset to include two hands for subjects to perform interactions which is common in reality. There are 540 videos with one person performing activities with both hands. There are 6 subjects performing 9 different activities with 10 repetitions. There are a total of 14 action labels assigned to each hand, and entity-level annotations are provided on a per-frame basis within the video. +(III) CAD-120 dataset [13] is popular for HOI recognition. It contains 120 videos with 10 activities performed by 4 participants. There are 10 human sub-activities labeled per frame. + +Evaluation Metric: We report $F_{1}$ at $k$ metric [14] with thresholds $k = 10\%$ , $25\%$ , and $50\%$ . Compared to frame-based metrics which evaluate prediction on every single frame, this metric could measure prediction continuity in action segments because it views each predicted action segment as correct only when it has the Interaction over Union (IoU) with the corresponding ground truth over the threshold $k$ . + +# 5.2 Implementation details + +In the experiment, we use three layers of context fuser for Bimanual and two for CAD-120 and MPHOI. The features of humans and objects are extracted from [27] and their dimension is mapped to 768, 256, and 512 for MPHOI, Bimanual, and CAD-120 respectively. More implementation details can be found in the supplementary material. + +# 5.3 Quantitative results + +Multi-person HOI recognition The quantitative results of joined segmentation and label recognition of sub-activity on MPHOI-72 in Tab. 1 show the performance of our method in multi-person HOI circumstance. Our method outperforms SOTA method 2G-GCN [27] by a large margin in all the three evaluation metrics. For $\mathrm{F}_1@10$ , $\mathrm{F}_1@25$ and $\mathrm{F}_1@50$ scores, our method surpasses 2G-GCN $23.0\%$ , $23.7\%$ and $22.5\%$ , respectively. The significant improvement achieved by our method indicates that the human fuser block in the CF module effectively improves the context-aware human representation learning under multi-person scenarios. + +Single-person HOI recognition We show the quantitative results for single-person HOI recognition in Tab. 2 and Tab. 3, which are performed on CAD-120 and Bimanual Actions datasets, respectively. Results in Tab. 2 show that our method performs slightly better on CAD-120 dataset compared to SOTA method 2G-GCN, with around $1\%$ improvement over 2G-GCN on all the three metrics. + +$\Delta \mathrm{CF}$ : removing context fuser module; $\Delta \mathrm{ISR}$ : removing interaction state reasoner module; $\Delta \mathcal{L}_{fea}$ : removing interaction feature loss; $\lambda = 0$ : ablating the hypersphere by replacing it with Euclidean distance; $\Delta \mathrm{CLIP} + \mathrm{BLIP}$ : ablating CLIP and BLIP model; Traditional Classifier: ablating the interaction-centric hypersphere by replacing it with a traditional classifier constructed with MLP and Softmax classifier. The improvements of our method over current SOTA method is highlighted with upward arrows. + +Table 1: The results of joined segmentation and label recognition of sub-activity on MPHOI-72. + +
ModelF1@10F1@25F1@50
ASSIGN [23]59.1 ± 12.151.0 ± 16.733.2 ± 14.0
2G-GCN [27]68.6 ± 10.460.8 ± 10.345.2 ± 6.5
Ours91.6 ± 0.9(↑ 23.0)84.5 ± 2.6(↑ 23.7)67.7 ± 2.2(↑ 22.5)
Ours (ΔCF)65.8 ± 12.457.6 ± 14.039.2 ± 12.6
Ours (ΔISR)80.1 ± 5.573.0 ± 8.255.6 ± 6.1
Ours (ΔLfea)73.5 ± 15.769.7 ± 13.348.8 ± 13.0
Ours (λ = 0)81.2 ± 0.774.8 ± 4.253.2 ± 0.3
Ours (ΔCLIP+BLIP)80.0 ± 6.773.0 ± 9.955.6 ± 7.4
Ours (Traditional Classifier)84.6 ± 7.274.7 ± 10.554.6 ± 13.7
+ +For the Bimanual Actions dataset, our method performs as good as 2G-GCN in $F_{1}@10$ while achieves $0.9\%$ and $5.0\%$ higher than 2G-GCN in $F_{1}@25$ and $F_{1}@50$ score, respectively. These results indicate that our method achieves generally on par or even better performance on single-person video HOI recognition task. + +![](images/312e6ddeff50af8d8f53b23f51178b6bc41ff4155fa4c3f36cb1e8029d9a5d92.jpg) +Fig. 5: Qualitative ablation study results on MPHOI-72 dataset. Major prediction errors are highlighted in red dashed boxes. + +# 5.4 Ablation studies + +In this section, we ablate the CF module and the ISR module for validating the effectiveness of these proposed components. As shown in Tab. 1, removing CF module results in more than $20\%$ $F_{1}$ score drop of the three metrics in MPHOI-72 dataset, indicating the essential improvement of CF module in learning context-rich representations. Visualization results of temporal segmentation of interactions in Fig. 5 indicates that removing CF module results in incorrect + +Table 2: The results of joined segmentation and label recognition of sub-activity on CAD-120. The notations are the same with Tab. 1. + +
ModelF1@10F1@25F1@50
rCRF [29]65.6 ± 3.261.5 ± 4.147.1 ± 4.3
Independent BiRNN [27]70.2 ± 5.564.1 ± 5.348.9 ± 6.8
ATCRF [11]72.0 ± 2.868.9 ± 3.653.5 ± 4.3
Relational BiRNN [27]79.2 ± 2.575.2 ± 3.562.5 ± 5.5
ACoLP [35]90.2 ± 2.687.4 ± 1.476.8 ± 2.3
ASSIGN [23]88.0 ± 1.884.8 ± 3.073.8 ± 5.8
2G-GCN [27]89.5 ± 1.687.1 ± 1.876.2 ± 2.8
Ours90.7 ± 2.9(↑ 0.5)88.1 ± 2.8(↑ 0.7)77.6 ± 4.7(↑ 0.8)
Ours (ΔCF)81.1 ± 4.077.0 ± 4.865.2 ± 5.6
Ours (ΔISR)88.5 ± 3.785.5 ± 3.673.9 ± 5.7
Ours (ΔLfea)89.3 ± 1.985.6 ± 2.175.9 ± 4.4
Ours (λ = 0)72.0 ± 4.465.0 ± 6.948.6 ± 6.3
Ours (ΔCLIP+BLIP)89.4 ± 2.385.5 ± 3.974.9 ± 5.7
Ours (Traditional Classifier)79.5 ± 11.073.9 ± 11.456.6 ± 12.5
+ +![](images/205d20fe6fba57ba3413e8035d40578c6d108c0f2141c14c0e415090831424b9.jpg) +Fig. 6: Qualitative ablation study results on CAD-120 dataset. Major prediction errors are highlighted in red dashed boxes. + +interaction predictions in both humans (highlighted in red in Fig. 5). Similarly, in CAD-120 and Bimanual Actions datasets, deleting CF module also results in massive $F_{1}$ score drop. The visualization results in Fig. 6 suggests an incorrect prediction segment when removing CF module. + +Furthermore, we ablate hypersphere by replacing it with Euclidean distance, where $\lambda = 0$ in Eq. 6. Results in Tab. 1, 2 and 3 indicate that utilizing Euclidean distance results in at least $7\%$ , $10\%$ and $3\%$ $F_{1}$ scores drop in MPHOI, CAD-120 and Bimanual Actions dataset, respectively. Therefore, we conclude that Euclidean distance do not introduce HOI structure priors, ignoring valuable structure information of HOI for guiding predictions. Subsequently, we ablate CLIP and BLIP models by randomly initialize interaction features and context features. Results in Tab. 1, 2 and 3 indicate that removing CLIP and BLIP models results in some drop of model performance, but is still on par with or better than SOTA methods. Thus, it is the intricately designed structure of our model that substantiates the substantial enhancement in performance. Finally, we ablate the interaction-centric hypersphere by replacing it with a traditional classifier constructed with multi-layer perceptron (MLP) and Softmax classifier. The outcomes, as presented in Tab. 1, 2 and 3, reveal a notable decline of over $7\%$ in the $F_{1}$ score within the MPHOI dataset when employing the traditional classifier. Likewise, in the CAD-120 dataset, the traditional classifier results in a substantial decrease of more than $11\%$ in the $F_{1}$ score. Additionally, within + +Table 3: The results of joined segmentation and label recognition of sub-activity on Bimanual Actions. The notations are the same with Tab. 1. + +
ModelF1@10F1@25F1@50
Dreher et al. [4]40.6 ± 7.234.8 ± 7.122.2 ± 5.7
Independet BiRNN [27]74.7 ± 7.072.0 ± 7.061.8 ± 7.3
Relational BiRNN [27]77.7 ± 3.975.0 ± 4.264.8 ± 5.3
ASSIGN [23]84.0 ± 2.081.2 ± 2.068.5 ± 3.3
2G-GCN [27]85.0 ± 2.282.0 ± 2.669.2 ± 3.1
Ours85.0 ± 2.582.9 ± 2.9(↑ 0.9)74.2 ± 4.3(↑ 5.0)
Ours (ΔCF)82.5 ± 5.080.5 ± 5.571.1 ± 7.0
Ours (ΔISR)84.1 ± 2.381.8 ± 2.873.0 ± 3.7
Ours (ΔLfea)84.5 ± 4.682.0 ± 5.271.8 ± 6.9
Ours (λ = 0)76.7 ± 5.274.3 ± 6.065.2 ± 6.3
Ours (ΔCLIP+BLIP)84.3 ± 1.481.8 ± 1.873.2 ± 2.7
Ours (Traditional Classifier)82.0 ± 3.679.8 ± 4.171.0 ± 5.6
+ +the Bimanual Actions dataset, the traditional classifier induces a decline exceeding $2\%$ in the $F_{1}$ score. These findings unanimously underscore the efficacy of modeling HOI manifold structures by the hypersphere module. + +![](images/c21d56e15216e69ceabf377e4c562f2d66908b903faba48b5f59ccb6fcc1371d.jpg) +Fig. 7: (A) Visualization of interaction-centric hyperspheres, the learned interaction representations (large points locate at the center of hyperspheres) and human-object entity representations (small points surrounding the hyperspheres) in embedding space. Selected frame samples are shown for each interaction-centric hypersphere. (B) 2D visualization of interaction-centric hyperspheres. + +![](images/448c3a360fa59d726c6c2503f919a4330a2a42cb743008341e1ff2d6e57b3d10.jpg) + +# 5.5 Qualitative results + +We show some visualization results on MPHOI in Fig 8 to compare our method with SOTA method 2G-GCN. The red highlighted boxes indicate that 2G-GCN tend to generate unreasonable interaction predictions, while our method generates more reasonable interaction predictions. The visualization results in Fig. 9 show similar prediction pattern where 2G-GCN predicted some unreasonable short segments (highlighted in red boxes) while our method predicts more accu + +rately. We also visualize the interaction-centric hypersphere, the learned interaction representations and the human-object entity representations in embedding space in Fig. 7. Results in Fig. 7 (A) show that human-object entity representations (small dots in Fig. 7 (A)) belonging to the specific interaction class locates near the surface of the corresponding hypersphere. Similar with Fig. 7 (A), we also visualize the intraction-centric hyerspheres in 2D sapce in Fig. 7 (B). Results in Fig. 7 (B) indicate that HOIs distribute around the sphere of different interaction classes. The distributions of HOIs on each sphere also differ a lot, where the sphere belongs to the solve interaction class captures HOIs on the upper part of its sphere, while the sphere belongs to the cut interaction class captures HOIs on the lower part of its sphere. These results suggest that our model successfully model the manifold structure of HOI. + +![](images/e655df80783c4250729bb875fb971ec4912271b21a525674760f66faadeac719.jpg) +Fig. 8: Visualization results on MPHOI-72 dataset. Major prediction errors are highlighted in red dashed boxes. + +![](images/50faa1b4a7cee1191ab92636440d64e83d5d4fdc20b17c9ee3ac123369f10cf3.jpg) +Fig. 9: Visualization results on CAD-120 dataset. Major prediction errors are highlighted in red dashed boxes. + +# 6 Conclusion + +In this work, we propose an interaction-centric spatio-temporal context reasoning network for multi-person video HOI recognition. Specifically, we propose a context fuser and an interaction state reasoner to learn spatio-temporal context-rich and reasoning-aware entity representations. We further represent HOI components with an interaction-centric hypersphere for HOI classification. Experiment results show that our method outperforms SOTA method by more than $22\%$ $F_{1}$ score in multi-person scenarios, and achieves competitive results on single-person cases compared to SOTA methods. + +# Acknowledgements + +This material is based upon work supported under the AI Research Institutes program by National Science Foundation and the Institute of Education Sciences, U.S. Department of Education through Award #2229873 - AI Institute for Transforming Education for Children with Speech and Language Processing Challenges. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation, the Institute of Education Sciences, or the U.S. Department of Education. + +# References + +1. Chopra, S., Hadsell, R., LeCun, Y.: Learning a similarity metric discriminatively, with application to face verification. In: 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05). vol. 1, pp. 539-546 vol. 1 (2005). https://doi.org/10.1109/CVPR.2005.2029 +2. Deng, S., Zhang, N., Chen, H., Tan, C., Huang, F., Xu, C., Chen, H.: Low-resource extraction with knowledge-aware pairwise prototype learning. Knowledge-Based Systems 235, 107584 (2022) 4 +3. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 4 +4. Dreher, C.R.G., Wächter, M., Asfour, T.: Learning object-action relations from bimanual human demonstration using graph networks. IEEE Robotics and Automation Letters (RA-L) 5(1), 187-194 (2020). https://doi.org/10.1109/LRA.2019.2949221 3, 10, 13 +5. Gao, C., Xu, J., Zou, Y., Huang, J.B.: Drg: Dual relation graph for human-object interaction detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16. pp. 696-712. Springer (2020) 4 +6. Gao, K., Chen, L., Zhang, H., Xiao, J., Sun, Q.: Compositional prompt tuning with motion cues for open-vocabulary video relation detection. In: The Eleventh International Conference on Learning Representations (2022) 4 +7. Gupta, T., Schwing, A., Hoiem, D.: No-frills human-object interaction detection: Factorization, layout encodings, and training techniques. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9677-9685 (2019) 4 +8. Iftekhar, A., Chen, H., Kundu, K., Li, X., Tighe, J., Modolo, D.: What to look at and where: Semantic and spatial refined transformer for detecting human-object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5353-5363 (2022) 4 +9. Kim, B., Choi, T., Kang, J., Kim, H.J.: Uniondet: Union-level detector towards real-time human-object interaction detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 498-514. Springer (2020) 4 + +10. Kim, B., Lee, J., Kang, J., Kim, E.S., Kim, H.J.: Hotr: End-to-end human-object interaction detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 74-83 (June 2021) 4 +11. Koppula, H.S., Saxena, A.: Anticipating human activities using object affordances for reactive robotic response. IEEE transactions on pattern analysis and machine intelligence 38(1), 14-29 (2015) 12 +12. Koppula, H.S., Gupta, R., Saxena, A.: Learning human activities and object affordances from rgb-d videos. The International journal of robotics research 32(8), 951-970 (2013) 3, 4 +13. Koppula, H.S., Gupta, R., Saxena, A.: Learning human activities and object affordances from rgb-d videos. Int. J. Rob. Res. 32(8), 951-970 (jul 2013). https://doi.org/10.1177/0278364913478446, https://doi.org/10.1177/0278364913478446 10 +14. Lea, C., Flynn, M.D., Vidal, R., Reiter, A., Hager, G.D.: Temporal convolutional networks for action segmentation and detection. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 156-165 (2017) 10 +15. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) 6 +16. Li, Y.L., Zhou, S., Huang, X., Xu, L., Ma, Z., Fang, H.S., Wang, Y., Lu, C.: Transferable interactiveness knowledge for human-object interaction detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3585-3594 (2019) 4 +17. Liao, Y., Liu, S., Wang, F., Chen, Y., Qian, C., Feng, J.: Ppdm: Parallel point detection and matching for real-time human-object interaction detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 482-490 (2020) 4 +18. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017) 9 +19. Liu, X., Li, Y.L., Wu, X., Tai, Y.W., Lu, C., Tang, C.K.: Interactiveness field in human-object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20113-20122 (2022) 2, 4 +20. Luan, T., Wang, Y., Zhang, J., Wang, Z., Zhou, Z., Qiao, Y.: Pc-hmr: Pose calibration for 3d human mesh recovery from 2d images/videos. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 2269-2276 (2021) 4 +21. Luan, T., Zhai, Y., Meng, J., Li, Z., Chen, Z., Xu, Y., Yuan, J.: High fidelity 3d hand shape reconstruction via scalable graph frequency decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16795-16804 (2023) 4 +22. Mettes, P., Van der Pol, E., Snoek, C.: Hyperspherical prototype networks. Advances in neural information processing systems 32 (2019) 4 +23. Morais, R., Le, V., Venkatesh, S., Tran, T.: Learning asynchronous and sparse human-object interaction in videos. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16041-16050 (2021) 2, 4, 11, 12, 13 +24. Nagarajan, T., Feichtenhofer, C., Grauman, K.: Grounded human-object interaction hotspots from video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8688-8697 (2019) 4 + +25. Park, J., Park, J.W., Lee, J.S.: Viplo: Vision transformer based pose-conditioned self-loop graph for human-object interaction detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17152-17162 (June 2023) 4 +26. Qi, S., Wang, W., Jia, B., Shen, J., Zhu, S.C.: Learning human-object interactions by graph parsing neural networks. In: Proceedings of the European Conference on Computer Vision (ECCV) (September 2018) 4 +27. Qiao, T., Men, Q., Li, F.W., Kubotani, Y., Morishima, S., Shum, H.P.: Geometric features informed multi-person human-object interaction recognition in videos. In: European Conference on Computer Vision. pp. 474-491. Springer (2022) 2, 3, 4, 6, 10, 11, 12, 13 +28. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 6, 8 +29. Sener, O., Saxena, A.: rcrf: Recursive belief estimation over crfs in rgb-d activity videos. In: Robotics: Science and systems (2015) 12 +30. Sunkesula, S.P.R., Dabral, R., Ramakrishnan, G.: Lighten: Learning interactions with graph and hierarchical temporal networks for hoi in videos. In: Proceedings of the 28th ACM international conference on multimedia. pp. 691-699 (2020) 4 +31. Tamura, M., Ohashi, H., Yoshinaga, T.: Qpic: Query-based pairwise human-object interaction detection with image-wide contextual information. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10410-10419 (June 2021) 4 +32. Tu, D., Sun, W., Min, X., Zhai, G., Shen, W.: Video-based human-object interaction detection from tubelet tokens. Advances in Neural Information Processing Systems 35, 23345-23357 (2022) 4 +33. Wang, T., Anwer, R.M., Khan, M.H., Khan, F.S., Pang, Y., Shao, L., Laaksonen, J.: Deep contextual attention for human-object interaction detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5694-5702 (2019) 4 +34. Xi, N., Meng, J., Yuan, J.: Chain-of-look prompting for verb-centric surgical triplet recognition in endoscopic videos. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 5007-5016 (2023) 4 +35. Xi, N., Meng, J., Yuan, J.: Open set video hoi detection from action-centric chain-of-look prompting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3079-3089 (2023) 12 +36. Zhang, Y., Pan, Y., Yao, T., Huang, R., Mei, T., Chen, C.W.: Exploring structure-aware transformer over interaction proposals for human-object interaction detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19548-19557 (2022) 2, 4 +37. Zhong, X., Qu, X., Ding, C., Tao, D.: Glance and gaze: Inferring action-aware points for one-stage human-object interaction detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13234-13243 (2021) 4 \ No newline at end of file diff --git a/2024/Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition/images.zip b/2024/Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a8e3c06d4e4348e29c0c519af1bd43d085177451 --- /dev/null +++ b/2024/Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d5c1d6f6db81f97f6a29a709e8ec45ec5c6b2fe1cd355f9b330fe709866110a +size 509841 diff --git a/2024/Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition/layout.json b/2024/Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e2786b4eb2beab903b8437531b75555337b83ee3 --- /dev/null +++ b/2024/Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition/layout.json @@ -0,0 +1,10118 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 181, + 111, + 434, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 111, + 434, + 166 + ], + "spans": [ + { + "bbox": [ + 181, + 111, + 434, + 166 + ], + "type": "text", + "content": "Interaction-centric Spatio-Temporal Context Reasoning for Multi-Person Video HOI Recognition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 148, + 193, + 465, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 193, + 465, + 206 + ], + "spans": [ + { + "bbox": [ + 148, + 193, + 465, + 206 + ], + "type": "text", + "content": "Yisong Wang" + }, + { + "bbox": [ + 148, + 193, + 465, + 206 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 148, + 193, + 465, + 206 + ], + "type": "text", + "content": ", Nan Xi" + }, + { + "bbox": [ + 148, + 193, + 465, + 206 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 148, + 193, + 465, + 206 + ], + "type": "text", + "content": ", Jingjing Meng" + }, + { + "bbox": [ + 148, + 193, + 465, + 206 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 148, + 193, + 465, + 206 + ], + "type": "text", + "content": ", and Junsong Yuan" + }, + { + "bbox": [ + 148, + 193, + 465, + 206 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 230, + 214, + 384, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 214, + 384, + 226 + ], + "spans": [ + { + "bbox": [ + 230, + 214, + 384, + 226 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 230, + 214, + 384, + 226 + ], + "type": "text", + "content": " School of EECS, Peking University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 153, + 226, + 459, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 226, + 459, + 237 + ], + "spans": [ + { + "bbox": [ + 153, + 226, + 459, + 237 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 153, + 226, + 459, + 237 + ], + "type": "text", + "content": " Department of Computer Science and Engineering, University at Buffalo" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 276, + 237, + 338, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 237, + 338, + 247 + ], + "spans": [ + { + "bbox": [ + 276, + 237, + 338, + 247 + ], + "type": "text", + "content": "3 Amazon Inc." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 249, + 480, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 249, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 133, + 249, + 480, + 260 + ], + "type": "text", + "content": "wys05y98@gmail.com, {nanxi, jsyuany}@buffalo.edu, jingjing.meng1@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 310, + 455, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 310, + 455, + 585 + ], + "spans": [ + { + "bbox": [ + 159, + 310, + 455, + 585 + ], + "type": "text", + "content": "Abstract. Understanding human-object interaction (HOI) in videos represents a fundamental yet intricate challenge in computer vision, requiring perception and reasoning across both spatial and temporal domains. Despite previous success of object detection and tracking, multi-person video HOI recognition still faces two major challenges: (1) the three facets of HOI (human, objects, and the interactions that bind them) exhibit interconnectedness and exert mutual influence upon one another. (2) the complexity of multi-person multi-object combinations in spatio-temporal interaction. To address them, we design a spatio-temporal context fuser to better model the interactions among persons and objects in videos. Furthermore, to equip the model with temporal reasoning capacity, we propose an interaction state reasoner module on top of context fuser. Considering the interaction is a key element to bind human and object, we propose an interaction-centric hypersphere in the feature embedding space to model each category of interaction. It helps to learn the distribution of HOI samples belonging to the same interactions on the hypersphere. After training, each interaction prototype sphere will fit the testing HOI sample to determine the HOI classification result. Empirical results on multi-person video HOI dataset MPHOI-72 indicate that our method remarkably surpasses state-of-the-art (SOTA) method by more than " + }, + { + "bbox": [ + 159, + 310, + 455, + 585 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 159, + 310, + 455, + 585 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 159, + 310, + 455, + 585 + ], + "type": "text", + "content": " score. At the same time, on single-person datasets Bimanual Actions (single-human two-hand HOI) and CAD-120 (single-human HOI), our method achieves on par or even better results compared with SOTA methods. Source code is released at the following link: https://github.com/southnx/IcH-Vid-HOI." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 597, + 463, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 597, + 463, + 619 + ], + "spans": [ + { + "bbox": [ + 159, + 597, + 463, + 619 + ], + "type": "text", + "content": "Keywords: interaction-centric " + }, + { + "bbox": [ + 159, + 597, + 463, + 619 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 597, + 463, + 619 + ], + "type": "text", + "content": " spatio-temporal context reasoning " + }, + { + "bbox": [ + 159, + 597, + 463, + 619 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 597, + 463, + 619 + ], + "type": "text", + "content": " multi-person human-object interaction" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 653, + 334, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 653, + 334, + 666 + ], + "spans": [ + { + "bbox": [ + 142, + 653, + 334, + 666 + ], + "type": "text", + "content": "* Equal contribution; ≔ Corresponding author" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 128, + 250, + 169 + ], + "blocks": [ + { + "bbox": [ + 236, + 116, + 362, + 125 + ], + "lines": [ + { + "bbox": [ + 236, + 116, + 362, + 125 + ], + "spans": [ + { + "bbox": [ + 236, + 116, + 362, + 125 + ], + "type": "text", + "content": "Multi-person Video Human-Object Interaction" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 153, + 128, + 250, + 169 + ], + "lines": [ + { + "bbox": [ + 153, + 128, + 250, + 169 + ], + "spans": [ + { + "bbox": [ + 153, + 128, + 250, + 169 + ], + "type": "image", + "image_path": "4b34982c1398083e87608267fe526441237b01c8e66eac06865a957e2a210330.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 258, + 128, + 355, + 169 + ], + "blocks": [ + { + "bbox": [ + 258, + 128, + 355, + 169 + ], + "lines": [ + { + "bbox": [ + 258, + 128, + 355, + 169 + ], + "spans": [ + { + "bbox": [ + 258, + 128, + 355, + 169 + ], + "type": "image", + "image_path": "1f3ae3a97705067a470cd5512afbe464c84545b57759bc5c4d7834d8278d6108.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 177, + 480, + 200 + ], + "lines": [ + { + "bbox": [ + 130, + 177, + 480, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 480, + 200 + ], + "type": "text", + "content": "Fig. 1: Examples of multi-person video HOI. Three categories of interactions are presented, including co-working, cheering, hair_cutting." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 362, + 128, + 460, + 169 + ], + "blocks": [ + { + "bbox": [ + 362, + 128, + 460, + 169 + ], + "lines": [ + { + "bbox": [ + 362, + 128, + 460, + 169 + ], + "spans": [ + { + "bbox": [ + 362, + 128, + 460, + 169 + ], + "type": "image", + "image_path": "3048e4587dd2851d296f2750bfb7e5b0b660ac7cf162769eba993122e9f1f017.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 222, + 230, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 222, + 230, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 222, + 230, + 235 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 247, + 481, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 481, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 481, + 305 + ], + "type": "text", + "content": "Video-based Human-Object Interaction (HOI) recognition aims to identify the interactions occurring between human and object entities within video frames. Precisely recognizing HOIs in real world scenarios is essential for a bunch of applications, such as assisting patients by recognizing daily activities and predicting pedestrian movements to avoid accidents for autonomous vehicles." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 307, + 481, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 481, + 474 + ], + "type": "text", + "content": "Most existing HOI recognition research focuses on static images [19, 36], with much less attention on video-based HOI recognition. Video-based HOI recognition is more demanding than image-based HOI recognition due to the necessity of comprehending complex spatio-temporal dynamics and reasoning about multi-human and multi-object interaction combinations. The complexity is further exacerbated when dealing with multi-person circumstances, as examples shown in Fig. 1. In such cases, various human and object entities mutually influence each other, resulting in intricate interdependencies within the scene. Additionally, the three components (human, object and interactions that the entities are involved in) of HOI exhibit interwind structures, e.g., the possible interactions that can occur within a scene given a human and an object. However, current video HOI recognition methods [23, 27] do not fully explore such inherent structural nature of HOI components. Instead, they often opt for disentangled representations for each component, which may bring suboptimal representation capabilities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 475, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 481, + 666 + ], + "type": "text", + "content": "To overcome the aforementioned limitations, we introduce an interaction-centric spatio-temporal context reasoning approach for representing video HOIs. For the purpose of enhancing the awareness of complex HOI structures in our representations, we introduce the Context Fuser (CF), which encodes both entity representations and interaction representations. Moreover, to empower our model with the ability to reason about interaction state transitions across video frames, we propose the Interaction State Reasoner (ISR) module for generating interaction representations. In addition, we employ a bidirectional Gated Recurrent Unit (BiGRU) to model temporal dynamics across video frames. This multi-level representation learning framework not only facilitates effective exploration of the interdependencies among structured HOI components but also empowers the model with interaction reasoning capabilities in both spatial and temporal domains. To compute the probability of an HOI in the scene, we introduce an interaction-centric hypersphere. This approach leverages the concept of a hypersphere, where the interaction is located at its center, while the human and object entities involved in that interaction are situated on the hypersphere's" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "surface. We assume that human-object entities belonging to each interaction class are located on their respective hyperspheres, specific to that interaction class." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 366 + ], + "type": "text", + "content": "Concretely, we generate context-rich and reasoning-aware video HOI representations through three key components: the Context Fuser (CF), Interaction State Reasoner (ISR) and Bidirectional GRU (BiGRU). The CF module integrates context information from human-object entities and interactions. It comprises three fuser blocks for humans, objects, and interactions. The Object Fuser Block processes local video frame data, enhancing object features. The Interaction Fuser Block combines human and object representations with interaction-specific characteristics. Additionally, in multi-person scenes, the Human Fuser Block captures human representations influenced by others. This approach fosters comprehensive HOI representations via effective context fusion. To facilitate interaction reasoning, we place the ISR module on top of the context fuser module, yielding entity representations capable of capturing interaction transition dynamics. These entity representations are then input into the BiGRU module to model temporal dynamics across video frames, thereby ensuring a comprehensive understanding of the evolving context and interactions within the video data. Finally, we determine interaction classes in each frame with the interaction-centric hypersphere, computing the probability of human-object entities belonging to specific interaction classes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 367, + 489, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 489, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 489, + 450 + ], + "type": "text", + "content": "We assess our model's performance on three video-based HOI datasets: MPHOI-72 [27] (multi-person), Bimanual Actions [4] and CAD-12 [12] (single-person). Our results highlight our model's superiority in multi-person scenarios, achieving an impressive over " + }, + { + "bbox": [ + 130, + 367, + 489, + 450 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 130, + 367, + 489, + 450 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 130, + 367, + 489, + 450 + ], + "type": "text", + "content": " score improvement over the current state-of-the-art (SOTA). In single-person scenarios, our method delivers on par or even better performance compared to the current SOTA method. Our major contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 457, + 481, + 600 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 457, + 481, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 481, + 505 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 481, + 505 + ], + "type": "text", + "content": "- To represent inherent HOI manifold structures, we propose an interaction-centric hypersphere representation scheme. This scheme explicitly guide the process of learning intrinsic structural nature of HOI and elucidates the interdependencies among its components." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 505, + 481, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 505, + 481, + 551 + ], + "spans": [ + { + "bbox": [ + 138, + 505, + 481, + 551 + ], + "type": "text", + "content": "- To learn context-rich and reasoning-aware entity representations, we introduce context fuser and interaction state reasoning modules. This enhancement results in entity representations that are highly suitable for video-based HOI tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 552, + 481, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 552, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 552, + 481, + 600 + ], + "type": "text", + "content": "- Extensive experiment results showcase that our method achieves SOTA performance with a huge improvement of more than " + }, + { + "bbox": [ + 138, + 552, + 481, + 600 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 138, + 552, + 481, + 600 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 138, + 552, + 481, + 600 + ], + "type": "text", + "content": " score over existing methods in multi-person scenario. Additionally, our model achieves competitive results in single-person cases compared to SOTA method." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 617, + 240, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 240, + 629 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 240, + 629 + ], + "type": "text", + "content": "2 Related works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "HOI detection in images: HOI detection in images aims at understanding interactions in images between humans and objects. Different methods have been" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "text", + "content": "Video HOI Interaction-centric context reasoning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "proposed in previous studies. Some works propose Convolutional Neural Networks (CNN)-based methods which can be further divided into one-stage methods [9,17,37] and two-stage methods [5,7,16,33]. However, these methods usually lack of ability to capture global context information. Recently, Transformer-based models [8,10,31,36] became the main approach for the HOI task. Following the architecture of DETR, these models achieved superior performance on HOI detection. Moreover, some works also utilize graph [25] and interactiveness field [19] to achieve better performance. These various approaches to image HOI provide the fundamentals for video HOI recognition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 426 + ], + "type": "text", + "content": "HOI recognition in videos: Video HOI recognition is the foundation of many real-world applications, including understanding surgical activities in the operation room [34] and guiding 3D human reconstruction [20, 21]. Video-based HOI recognitions have to deal with both spatial and temporal reasoning. Before the use of neural networks, some early studies formulated this task using the Markov model [12] to utilize temporal cues. In [24], HOI hotspots in videos are learned in a novel approach, with two networks trained jointly to capture spatial regions where actions happen. While [6] focuses on visual relation prediction in open-vocabulary with pretrained visual-language models. Recent works have used Recurrent Neural Networks (RNN) combined with Graph Neural Networks (GNN) [23, 26, 27, 30] to predict human-object relations in videos. Inspired by ViT [3], some works also propose Transformer-based methods to reason spatial relations better [32]. However, RNN-based models usually require complex training strategies or long training time in order to achieve the best performance. Moreover, when multiple persons are involved in an activity jointly, these methods lack the ability to model their collaboration, resulting in poor performance when the interactions are performed by multiple persons." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 427, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 482, + 511 + ], + "type": "text", + "content": "Hyperspheres for class representation: Hyperspheres have been demonstrated to be an effective approach to model class representation [2, 22]. Geometrical modeling strategy has been proposed in [2], where the effectiveness has been confirmed. This approach proves advantageous for capturing and representing enriched class-level information, particularly well-suited for creating measurements in Euclidean space. Consequently, it is naturally adaptable to structured prediction tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 531, + 220, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 220, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 220, + 544 + ], + "type": "text", + "content": "3 Motivation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": "Video HOI recognition task involves identifying both the human and object entities engaged in an interaction across a sequence of video frames. This task encompasses spatial and temporal aspects, as it requires understanding the relationships between humans, objects, and their interactions over time. However, current methods for video HOI recognition often neglect this crucial dependency structure, resulting in the separation of learned representations associated with humans, objects, and their interactions, ultimately compromising their representational accuracy. To address this issue, we propose a novel approach featuring a context fuser and an interaction state reasoner to capture spatio-temporal" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 112, + 474, + 346 + ], + "blocks": [ + { + "bbox": [ + 141, + 112, + 474, + 346 + ], + "lines": [ + { + "bbox": [ + 141, + 112, + 474, + 346 + ], + "spans": [ + { + "bbox": [ + 141, + 112, + 474, + 346 + ], + "type": "image", + "image_path": "4a17d980bda599dd9d63d0945d0cbe99e4d88b1839ba6ff64e60f5fdafcea4fd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 354, + 482, + 431 + ], + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 431 + ], + "type": "text", + "content": "Fig. 2: Model Overview. Each video frame is applied with a backbone for feature extraction. Subsequently, a context fuser and interaction state reasoner is employed for learning interaction representations " + }, + { + "bbox": [ + 130, + 354, + 482, + 431 + ], + "type": "inline_equation", + "content": "z_{I}" + }, + { + "bbox": [ + 130, + 354, + 482, + 431 + ], + "type": "text", + "content": " and human-object entity representations " + }, + { + "bbox": [ + 130, + 354, + 482, + 431 + ], + "type": "inline_equation", + "content": "\\{\\hat{z}_E^t\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 354, + 482, + 431 + ], + "type": "text", + "content": ". Bidirectional GRU is further utilized to model temporal dynamics across video frames. The final predicted interaction probability " + }, + { + "bbox": [ + 130, + 354, + 482, + 431 + ], + "type": "inline_equation", + "content": "\\{u^t\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 354, + 482, + 431 + ], + "type": "text", + "content": " is computed from the interaction-centric hypersphere. The dimension of each hypersphere is " + }, + { + "bbox": [ + 130, + 354, + 482, + 431 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 354, + 482, + 431 + ], + "type": "text", + "content": ", which is the same with the feature dimension in our model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 456, + 482, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 492 + ], + "type": "text", + "content": "contextual representations. Additionally, we introduce an interaction-centric hypersphere to represent HOI manifold structure and model the inter-dependency among predicted humans, objects, and interactions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 511, + 202, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 202, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 202, + 523 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 536, + 265, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 265, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 265, + 548 + ], + "type": "text", + "content": "4.1 Problem formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "For a video dataset " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ", given a video clip " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "V\\in \\mathcal{V}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " containing " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " video frames " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{v_{1},\\dots,v_{T}\\}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ", video HOI recognition aims to predict the temporal segmentation of interactions between human and object entities across all the video frames. Formally, we aim to learn an HOI recognition model " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " that outputs the segmentation of human's sub-activity " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{s_n\\}_{n = 1}^N" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " in each frame, where " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " is the number of human sub-activity segments. Each segment " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "s_n" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " is represented as an interval from its start time " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " to end time " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "t_{n + 1}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "s_n = [t_n,t_{n + 1})" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ". The start and end time of each segment are determined from interaction probability prediction " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{u^t\\}_{t = 1}^T" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " of each frame, where " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "u^{t}\\in \\mathbb{R}^{K}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " is the number of possible interaction classes." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "text", + "content": "Video HOI Interaction-centric context reasoning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 231, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 231, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 231, + 129 + ], + "type": "text", + "content": "4.2 Model Design" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 291 + ], + "type": "text", + "content": "In the following section, we introduce our interaction-centric hypersphere reasoning model for video HOI recognition in detail. As shown in Fig. 2, our major idea is to construct a hypersphere to represent each HOI in the scene. For each hypersphere, the interaction locates at the center of the hypersphere, while the corresponding human-object entity belongs to that interaction locates at the surface of the hypersphere. We construct Context Fuser (CF) module to learn context-rich human-object entity representations. For the aim of enabling model with reasoning ability over interaction state transitions, we propose Interaction State Reasoner (ISR) to reason on whether the current interaction will be continued or stopped. To model the temporal dynamics of HOI in videos, we update human-object entity representations " + }, + { + "bbox": [ + 130, + 134, + 482, + 291 + ], + "type": "inline_equation", + "content": "\\{\\hat{z}_E^t\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 134, + 482, + 291 + ], + "type": "text", + "content": " along the temporal domain with bidirectional GRU (BiGRU). Finally, predicted interaction class probability " + }, + { + "bbox": [ + 130, + 134, + 482, + 291 + ], + "type": "inline_equation", + "content": "u^t" + }, + { + "bbox": [ + 130, + 134, + 482, + 291 + ], + "type": "text", + "content": " of each frame is computed from the interaction-centric hypersphere." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 305, + 481, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 481, + 365 + ], + "type": "text", + "content": "Context fuser For a sequence of video frames " + }, + { + "bbox": [ + 130, + 305, + 481, + 365 + ], + "type": "inline_equation", + "content": "\\{v_{t}\\}_{t = 1}^{T}" + }, + { + "bbox": [ + 130, + 305, + 481, + 365 + ], + "type": "text", + "content": ", we follow 2G-GCN [27] to extract feature of humans and objects from backbone network. The extracted human features " + }, + { + "bbox": [ + 130, + 305, + 481, + 365 + ], + "type": "inline_equation", + "content": "z_{H}\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 130, + 305, + 481, + 365 + ], + "type": "text", + "content": " contain both bounding box information and skeleton keypoint information, where " + }, + { + "bbox": [ + 130, + 305, + 481, + 365 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 305, + 481, + 365 + ], + "type": "text", + "content": " indicates the feature dimension. Object features " + }, + { + "bbox": [ + 130, + 305, + 481, + 365 + ], + "type": "inline_equation", + "content": "z_{O}\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 130, + 305, + 481, + 365 + ], + "type": "text", + "content": " contain only bounding box information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "text", + "content": "We design a context fuser (CF) module shown in Fig. 3 to generate human-object entity representations " + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\{z_E^t\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "text", + "content": " based on human, object and contextual information. In multi-person circumstances, CF contains object fuser block, interaction fuser block and human fuser block sequentially. First, we design an object fuser block to incorporate object representations in local temporal regions into current object representations, generating learned object representation " + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\hat{z}_{O_1}" + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\hat{z}_{O_2}" + }, + { + "bbox": [ + 130, + 365, + 482, + 450 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 205, + 456, + 481, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 456, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 205, + 456, + 481, + 478 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {O _ {1}} = F F N \\left(S A \\left(Q = z _ {O _ {1}} ^ {t}; K, V = z _ {O _ {1}} ^ {\\bar {t}}\\right) + z _ {O _ {1}} ^ {t}\\right), \\tag {1}", + "image_path": "056a7d9fa497c6381f28b4da268f699e36e107902c0f0ea8a1104f51151ac104.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 205, + 475, + 409, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 475, + 409, + 489 + ], + "spans": [ + { + "bbox": [ + 205, + 475, + 409, + 489 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {O _ {2}} = F F N (S A (Q = z _ {O _ {2}} ^ {t}; K, V = z _ {O _ {2}} ^ {t}) + z _ {O _ {2}} ^ {t}),", + "image_path": "d167a975247a20d218f15e1b4c9616835180b17bc4cdfa6d3e20d53fce87a5a6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "inline_equation", + "content": "SA" + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "content": " indicates self-attention, " + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "inline_equation", + "content": "FFN" + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "content": " is feed forward network, " + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "inline_equation", + "content": "z_{O_1}^t \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "inline_equation", + "content": "z_{O_2}^t \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "content": " are the initial object features, while " + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "inline_equation", + "content": "z_{O_1}^{\\bar{t}} \\in \\mathbb{R}^{20d}" + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "inline_equation", + "content": "z_{O_2}^{\\bar{t}} \\in \\mathbb{R}^{20d}" + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "content": " are stacked object features from a local time window, " + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\bar{t} \\in [t - 10,t + 10)" + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "content": ". Subsequently, for all the " + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "content": " possible interactions " + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{I_i\\}_{i = 1}^K" + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "content": " as shown in Fig. 3, each interaction class " + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "content": " is prompted as a sentence " + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "inline_equation", + "content": "s =" + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "content": " \"The human is [interact]ing in the scene.\", where [interact] indicates the specific interaction class. Then the sentence is applied with the text encoder " + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{F}_T)" + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "content": " of large-scale vision-language model CLIP [28] to initialize the interaction feature " + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "inline_equation", + "content": "z_I = \\mathcal{F}_T(s) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "content": ". We also generate context feature " + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "inline_equation", + "content": "z_C \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "content": " to represent the semantic information of each frame. Specifically, we extract frame caption " + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "content": " from video frame " + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "inline_equation", + "content": "i(i = 1,\\dots,T)" + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "content": " with BLIP [15] model and apply the caption with CLIP model to extract text embedding. In order to adapt the human " + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "inline_equation", + "content": "(z_{H_1},z_{H_2})" + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "content": " features, object " + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "inline_equation", + "content": "(z_{O_1},z_{O_2})" + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "content": " features and context feature " + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "inline_equation", + "content": "(z_C)" + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "content": " to the specific interaction feature " + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "inline_equation", + "content": "z_I" + }, + { + "bbox": [ + 130, + 495, + 482, + 666 + ], + "type": "text", + "content": ", we construct an interaction fuser block that contains a cross-attention (CA) module" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 114, + 364, + 323 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 364, + 323 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 364, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 364, + 323 + ], + "type": "image", + "image_path": "a86e4cc1836c9ef67305448c9de403998b65e0af9da4a47905a32c0086c54aff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 331, + 482, + 376 + ], + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 376 + ], + "type": "text", + "content": "Fig.3: Context fuser. In multi-person scenario (left), the context fuser consists of object fuser block, interaction fuser block and human fuser block. In single-person case (right), only object fuser block and interaction fuser block are employed. Detailed descriptions of the architecture and annotations can be found in Sec. 4.2." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 365, + 114, + 473, + 323 + ], + "blocks": [ + { + "bbox": [ + 365, + 114, + 473, + 323 + ], + "lines": [ + { + "bbox": [ + 365, + 114, + 473, + 323 + ], + "spans": [ + { + "bbox": [ + 365, + 114, + 473, + 323 + ], + "type": "image", + "image_path": "924c78e11e15a31f724848add983b0b0814c26e3bd70240c942f225ab99bc0c3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 400, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 475 + ], + "type": "text", + "content": "followed by a Feed Forward Network (FFN), generating interaction-aware human features " + }, + { + "bbox": [ + 130, + 400, + 482, + 475 + ], + "type": "inline_equation", + "content": "(z_{H_1}' \\in \\mathbb{R}^d, z_{H_2}' \\in \\mathbb{R}^d)" + }, + { + "bbox": [ + 130, + 400, + 482, + 475 + ], + "type": "text", + "content": ", object features " + }, + { + "bbox": [ + 130, + 400, + 482, + 475 + ], + "type": "inline_equation", + "content": "(z_{O_1}' \\in \\mathbb{R}^d, z_{O_2}' \\in \\mathbb{R}^d)" + }, + { + "bbox": [ + 130, + 400, + 482, + 475 + ], + "type": "text", + "content": " and context feature " + }, + { + "bbox": [ + 130, + 400, + 482, + 475 + ], + "type": "inline_equation", + "content": "(z_C' \\in \\mathbb{R}^d)" + }, + { + "bbox": [ + 130, + 400, + 482, + 475 + ], + "type": "text", + "content": ". Furthermore, to model the influence between the two humans in multi-person scenarios, we construct a human fuser block, featuring the same architecture with interaction fuser block. For the first human (Human1), the updated human feature " + }, + { + "bbox": [ + 130, + 400, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\hat{z}_{H_1} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 400, + 482, + 475 + ], + "type": "text", + "content": " is generated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 179, + 483, + 481, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 483, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 179, + 483, + 481, + 498 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {H _ {1}} = F F N (C A (Q = S A (z _ {H _ {1}} ^ {\\prime}); K, V = z _ {H _ {2}} ^ {\\prime}) + S A (z _ {H _ {1}} ^ {\\prime})). \\qquad (2)", + "image_path": "b26de30e781ac73cbae64349fde629c5ac0ec42224cfe89da3da3b27c73ae36f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 506, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 554 + ], + "type": "text", + "content": "Human2 feature " + }, + { + "bbox": [ + 130, + 506, + 482, + 554 + ], + "type": "inline_equation", + "content": "\\hat{z}_{H_2} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 506, + 482, + 554 + ], + "type": "text", + "content": " is generated in the same way as Eq. 2. Finally, the human-object entity representation " + }, + { + "bbox": [ + 130, + 506, + 482, + 554 + ], + "type": "inline_equation", + "content": "z_{E_1} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 506, + 482, + 554 + ], + "type": "text", + "content": " of Human1 is computed by max-pooling operation over all the " + }, + { + "bbox": [ + 130, + 506, + 482, + 554 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 506, + 482, + 554 + ], + "type": "text", + "content": " dimensions of the four representations shown in Eq. 3:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 228, + 554, + 481, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 554, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 228, + 554, + 481, + 569 + ], + "type": "interline_equation", + "content": "z _ {E _ {1}} = M a x P o o l (\\hat {z} _ {H _ {1}}, z _ {O _ {1}} ^ {\\prime}, z _ {O _ {2}} ^ {\\prime}, z _ {C} ^ {\\prime}). \\qquad (3)", + "image_path": "fb3bb809a5ef5124f50f779ad4a6d08cd76025f06cf439b67f6d54647a4ab646.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 574, + 482, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 623 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 623 + ], + "type": "text", + "content": "The human-object entity representation " + }, + { + "bbox": [ + 130, + 574, + 482, + 623 + ], + "type": "inline_equation", + "content": "z_{E_2} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 574, + 482, + 623 + ], + "type": "text", + "content": " of Human2 is computed with similar approach as Eq. 3. The CF module for single-person cases are similar with multi-person, except that the human fuser block is removed and there is only one human feature as query to be fed into the interaction fuser block." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Interaction State Reasoner To augment the model's ability in interaction state transition reasoning, we introduce an Interaction State Reasoner (ISR)" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "text", + "content": "Video HOI Interaction-centric context reasoning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": "module following CF module. ISR module explicitly empowers the model to determine whether the current interaction should persist or transit to another interaction. Specifically, as shown in Fig. 4, at each time " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": ", the two possible states state1 and state2 represent \"continue\" or \"stop\" of an interaction, respectively. Each state is prompted as one sentence, where " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "s_1 =" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " \"This interaction is going to continue.\" and " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "s_2 =" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " \"This interaction is going to stop and change to another interaction.\" Then the embeddings of the two states " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "z_{state_1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "z_{state_2}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " are generated from CLIP [28] text encoder " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_T" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "z_{state_1} = \\mathcal{F}_T(s_1) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "z_{state_2} = \\mathcal{F}_T(s_2) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": ". Interaction state embeddings " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "z_{state_1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "z_{state_2}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " are further fed to a reasoner block (shown in Fig. 4) together with the interaction embedding " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "z_t^t \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": ", generating state-informed interaction embeddings " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\hat{z}_{state_1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\hat{z}_{state_2}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": ". The reasoner block contains a FFN and a State Interpolation (SI) module. The SI module generates the weights " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\omega_1, \\omega_2" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " for the two interaction states in the following approach:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 201, + 289, + 481, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 289, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 201, + 289, + 481, + 306 + ], + "type": "interline_equation", + "content": "\\omega_ {1}, \\omega_ {2} = S o f t m a x (F F N (z _ {E} ^ {t}) \\cdot [ z _ {s t a t e _ {1}} ^ {\\top}, z _ {s t a t e _ {2}} ^ {\\top} ]), \\tag {4}", + "image_path": "67ae8be54b668161a67e693c1fa23e22be95f4d22eba158db754710352cbd753.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\top" + }, + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "type": "text", + "content": " indicates transpose operation. Subsequently, the final human-object entity representation " + }, + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\hat{z}_E^t\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "type": "text", + "content": " is generated by interpolate over the current entity representation " + }, + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "type": "inline_equation", + "content": "z_{E}^{t}" + }, + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "type": "text", + "content": " and the entity representation " + }, + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "type": "inline_equation", + "content": "z_{E}^{t - 1}" + }, + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 252, + 358, + 481, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 358, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 252, + 358, + 481, + 373 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {E} ^ {t} = \\omega_ {1} \\cdot z _ {E} ^ {t - 1} + \\omega_ {2} \\cdot z _ {E} ^ {t}. \\tag {5}", + "image_path": "aee58fb1989abcb5a8536e45eb8efcd5cfd6a9cc6828d379adae0492e5dbde23.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 376, + 482, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 400 + ], + "type": "text", + "content": "Consequently, the generated human-object entity representation " + }, + { + "bbox": [ + 130, + 376, + 482, + 400 + ], + "type": "inline_equation", + "content": "\\hat{z}_E^t" + }, + { + "bbox": [ + 130, + 376, + 482, + 400 + ], + "type": "text", + "content": " is able to reason on the possible future interaction transitions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 428, + 373, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 428, + 373, + 584 + ], + "spans": [ + { + "bbox": [ + 130, + 428, + 373, + 584 + ], + "type": "text", + "content": "Interaction-centric Hypersphere With the above generated human-object entity representation " + }, + { + "bbox": [ + 130, + 428, + 373, + 584 + ], + "type": "inline_equation", + "content": "\\hat{z}_E^t" + }, + { + "bbox": [ + 130, + 428, + 373, + 584 + ], + "type": "text", + "content": " and interaction representation " + }, + { + "bbox": [ + 130, + 428, + 373, + 584 + ], + "type": "inline_equation", + "content": "z_I" + }, + { + "bbox": [ + 130, + 428, + 373, + 584 + ], + "type": "text", + "content": ", we need to calculate the probability of a human-object entity " + }, + { + "bbox": [ + 130, + 428, + 373, + 584 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 428, + 373, + 584 + ], + "type": "text", + "content": " categorizing into the interaction class " + }, + { + "bbox": [ + 130, + 428, + 373, + 584 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 130, + 428, + 373, + 584 + ], + "type": "text", + "content": ". To that end, we design an interaction-centric hypersphere with interaction at the center of hypersphere and human-object entity at its surface. We call it \"hypersphere\" because this is a sphere in the " + }, + { + "bbox": [ + 130, + 428, + 373, + 584 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 428, + 373, + 584 + ], + "type": "text", + "content": " space, where " + }, + { + "bbox": [ + 130, + 428, + 373, + 584 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 428, + 373, + 584 + ], + "type": "text", + "content": " is the feature dimension in our model, which is obviously larger than 3. This hypersphere design models the manifold structure of HOI, which is in charge of the specific interaction class. Concretely, we employ a hyperspherical measurement:" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 391, + 424, + 470, + 536 + ], + "blocks": [ + { + "bbox": [ + 391, + 424, + 470, + 536 + ], + "lines": [ + { + "bbox": [ + 391, + 424, + 470, + 536 + ], + "spans": [ + { + "bbox": [ + 391, + 424, + 470, + 536 + ], + "type": "image", + "image_path": "ae395d494ebb19a6d17892720c327e86089036fbc4fdea5550decaa21338a7eb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 544, + 482, + 578 + ], + "lines": [ + { + "bbox": [ + 378, + 544, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 378, + 544, + 482, + 578 + ], + "type": "text", + "content": "Fig. 4: Interaction State Reasoner. " + }, + { + "bbox": [ + 378, + 544, + 482, + 578 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 378, + 544, + 482, + 578 + ], + "type": "text", + "content": " indicates inner product." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 201, + 590, + 481, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 590, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 201, + 590, + 481, + 634 + ], + "type": "interline_equation", + "content": "\\mathcal {U} (E, I _ {i}) = \\frac {\\exp \\left(- \\left[ | | z _ {I _ {i}} - \\hat {z} _ {E} | | _ {2} - \\lambda \\right] _ {+}\\right)}{\\sum_ {j = 1} ^ {K} \\exp \\left(- \\left[ | | z _ {I _ {j}} - \\hat {z} _ {E} | | _ {2} - \\lambda \\right] _ {+}\\right)}, \\tag {6}", + "image_path": "ad8760f0b5f28e856345d06c8024d180943ce626f5958694b460d1b0650b6181.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "inline_equation", + "content": "z_{I_i}" + }, + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "inline_equation", + "content": "z_{I_j}" + }, + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "text", + "content": " indicate the interaction representations of interaction class " + }, + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{z}_E" + }, + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "text", + "content": " denotes the human-object entity representation. " + }, + { + "bbox": [ + 130, + 639, + 482, + 667 + ], + "type": "inline_equation", + "content": "[s]_+ \\triangleq" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "max(0,s)" + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": " indicates the radius of the hypersphere, which is set to be a constant in our model. The higher value of " + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(E,I_i)" + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": " suggests that the human and object entity " + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": " is more likely to be categorized into " + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 167, + 258, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 258, + 180 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 258, + 180 + ], + "type": "text", + "content": "4.3 Learning Objective" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 184, + 481, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 184, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 130, + 184, + 481, + 220 + ], + "type": "text", + "content": "The learning objective of our model contains two parts: (i) focal loss " + }, + { + "bbox": [ + 130, + 184, + 481, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 130, + 184, + 481, + 220 + ], + "type": "text", + "content": " for interaction classification; (ii) interaction feature loss " + }, + { + "bbox": [ + 130, + 184, + 481, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{fea}" + }, + { + "bbox": [ + 130, + 184, + 481, + 220 + ], + "type": "text", + "content": " that controls the smoothness of interaction features in local region." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "text", + "content": "Focal loss " + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "text", + "content": ": We employ focal loss [18] for interaction classification, mitigating the interaction class imbalance problem on model performance. For each video frame " + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "inline_equation", + "content": "v_{i}(i = 1,\\dots,T)" + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "text", + "content": ", our model predicts the probability " + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "inline_equation", + "content": "\\hat{y}_i\\in \\mathbb{R}^K" + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "text", + "content": " of all the interaction classes. The corresponding ground-truth of interaction class " + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "inline_equation", + "content": "y_{i}\\in \\mathbb{R}^{K}" + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "text", + "content": " is a binary vector. For each intraction class " + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "text", + "content": ", the focal loss " + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}^{k}" + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "text", + "content": " is formulated as: " + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}^{k} = -(1 - p_{k})^{\\gamma}\\log (p_{k})" + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "text", + "content": " is a hyperparameter to control the focusing extent, " + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "text", + "content": " is defined as: " + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "inline_equation", + "content": "\\{p_k = \\hat{y}_i^k" + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "text", + "content": ", if " + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "inline_equation", + "content": "y_i^k = 1;p_k = 1 - \\hat{y}_i^k" + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "text", + "content": ", otherwise\\}. Subsequently, the focal loss " + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "text", + "content": " of each video frame is obtained by combining the focal loss of each individual interaction class " + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls} = \\Sigma_{k = 1}^{K}\\mathcal{L}_{cls}^{k}" + }, + { + "bbox": [ + 130, + 221, + 481, + 329 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "text", + "content": "Interaction feature loss " + }, + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{fea}" + }, + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "text", + "content": ": We introduce interaction feature loss " + }, + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{fea}" + }, + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "text", + "content": " to control the temporal smoothness of interaction features. Our model outputs the feature of each human-object entity " + }, + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "text", + "content": " in each frame. Inspired by [1], in order to improve the continuity, we minimize the feature distance in the same segment and maximize the distance between different segments for each subject. Denote " + }, + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "inline_equation", + "content": "u_{E}^{t}" + }, + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "text", + "content": " as whether the interaction will continue or change to another action for entity " + }, + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "inline_equation", + "content": "u_{E}^{t} = 1" + }, + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "text", + "content": " indicates the interaction will stop and change to another action at time " + }, + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "text", + "content": " for entity " + }, + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "inline_equation", + "content": "u_{E}^{t} = 0" + }, + { + "bbox": [ + 130, + 329, + 481, + 426 + ], + "type": "text", + "content": " otherwise. We minimize" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 430, + 481, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 430, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 136, + 430, + 481, + 463 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f e a} = \\frac {1}{2} \\sum_ {t = 0} ^ {T - 1} \\left[ (1 - u _ {E} ^ {t}) \\left(\\left| \\left| z _ {E} ^ {t} - z _ {E} ^ {t + 1} \\right| \\right| _ {2}\\right) ^ {2} + u _ {E} ^ {t} \\left(\\max \\left(L - \\left| \\left| z _ {E} ^ {t} - z _ {E} ^ {t + 1} \\right| \\right| _ {2}, 0\\right)\\right) ^ {2} \\right], \\tag {7}", + "image_path": "bd1238f830963ea784c9345a43a5f88362a9f3674b0f4d662d44bbaa3d51a7b1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 468, + 481, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 468, + 481, + 492 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 481, + 492 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 468, + 481, + 492 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 131, + 468, + 481, + 492 + ], + "type": "text", + "content": " is a threshold that controls the minimal feature distance when interaction changes. In total, the overall loss is written as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 265, + 498, + 481, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 498, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 265, + 498, + 481, + 512 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c l s} + \\alpha \\mathcal {L} _ {f e a}, \\tag {8}", + "image_path": "ffdd5f461df506689a79673527b66275428a9976a0e2df3c3e1ee082828e5259.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 517, + 410, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 517, + 410, + 530 + ], + "spans": [ + { + "bbox": [ + 131, + 517, + 410, + 530 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 517, + 410, + 530 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 517, + 410, + 530 + ], + "type": "text", + "content": " is a hyperparameter to control the weight of each loss." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 544, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 544, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 131, + 544, + 481, + 582 + ], + "type": "text", + "content": "Model Inference During model inference, we compute the interaction probability " + }, + { + "bbox": [ + 131, + 544, + 481, + 582 + ], + "type": "inline_equation", + "content": "\\hat{y}_i\\in \\mathbb{R}^K" + }, + { + "bbox": [ + 131, + 544, + 481, + 582 + ], + "type": "text", + "content": " for each video frame. The interaction class " + }, + { + "bbox": [ + 131, + 544, + 481, + 582 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 131, + 544, + 481, + 582 + ], + "type": "text", + "content": " with the highest probability is chosen as the predicted interaction for that frame: " + }, + { + "bbox": [ + 131, + 544, + 481, + 582 + ], + "type": "inline_equation", + "content": "a = \\arg \\max \\hat{y}_i" + }, + { + "bbox": [ + 131, + 544, + 481, + 582 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 601, + 230, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 230, + 615 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 230, + 615 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 624, + 261, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 261, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 261, + 636 + ], + "type": "text", + "content": "5.1 Experimental setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Datasets: We evaluate our method on MPHOI-72, Bimanual Actions, and CAD-120 datasets, each representing multi-person collaboration, one person with two" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 448, + 103 + ], + "type": "text", + "content": "Video HOI Interaction-centric context reasoning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 296, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 296, + 127 + ], + "type": "text", + "content": "hands and a single hand respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 128, + 479, + 293 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 130, + 128, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 128, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 128, + 479, + 186 + ], + "type": "text", + "content": "(I) MPHOI-72 dataset is proposed in [27], which consists of multiple humans and objects in the scene. The dataset comprises 72 videos featuring 3 human subjects and 6 objects. Within each video, 2 individuals are paired to engage in 3 distinct activities, encompassing a total of 13 sub-activities, while utilizing 2 to 4 objects" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 188, + 479, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 479, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 479, + 257 + ], + "type": "text", + "content": "(II) Bimanual Actions dataset [4] is the first HOI dataset to include two hands for subjects to perform interactions which is common in reality. There are 540 videos with one person performing activities with both hands. There are 6 subjects performing 9 different activities with 10 repetitions. There are a total of 14 action labels assigned to each hand, and entity-level annotations are provided on a per-frame basis within the video." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 259, + 479, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 479, + 293 + ], + "type": "text", + "content": "(III) CAD-120 dataset [13] is popular for HOI recognition. It contains 120 videos with 10 activities performed by 4 participants. There are 10 human sub-activities labeled per frame." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 295, + 479, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 479, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 479, + 365 + ], + "type": "text", + "content": "Evaluation Metric: We report " + }, + { + "bbox": [ + 130, + 295, + 479, + 365 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 130, + 295, + 479, + 365 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 130, + 295, + 479, + 365 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 295, + 479, + 365 + ], + "type": "text", + "content": " metric [14] with thresholds " + }, + { + "bbox": [ + 130, + 295, + 479, + 365 + ], + "type": "inline_equation", + "content": "k = 10\\%" + }, + { + "bbox": [ + 130, + 295, + 479, + 365 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 295, + 479, + 365 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 130, + 295, + 479, + 365 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 295, + 479, + 365 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 295, + 479, + 365 + ], + "type": "text", + "content": ". Compared to frame-based metrics which evaluate prediction on every single frame, this metric could measure prediction continuity in action segments because it views each predicted action segment as correct only when it has the Interaction over Union (IoU) with the corresponding ground truth over the threshold " + }, + { + "bbox": [ + 130, + 295, + 479, + 365 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 295, + 479, + 365 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 384, + 278, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 384, + 278, + 396 + ], + "spans": [ + { + "bbox": [ + 131, + 384, + 278, + 396 + ], + "type": "text", + "content": "5.2 Implementation details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 403, + 479, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 479, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 479, + 462 + ], + "type": "text", + "content": "In the experiment, we use three layers of context fuser for Bimanual and two for CAD-120 and MPHOI. The features of humans and objects are extracted from [27] and their dimension is mapped to 768, 256, and 512 for MPHOI, Bimanual, and CAD-120 respectively. More implementation details can be found in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 479, + 262, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 262, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 262, + 491 + ], + "type": "text", + "content": "5.3 Quantitative results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 498, + 479, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 479, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 479, + 602 + ], + "type": "text", + "content": "Multi-person HOI recognition The quantitative results of joined segmentation and label recognition of sub-activity on MPHOI-72 in Tab. 1 show the performance of our method in multi-person HOI circumstance. Our method outperforms SOTA method 2G-GCN [27] by a large margin in all the three evaluation metrics. For " + }, + { + "bbox": [ + 130, + 498, + 479, + 602 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_1@10" + }, + { + "bbox": [ + 130, + 498, + 479, + 602 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 498, + 479, + 602 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_1@25" + }, + { + "bbox": [ + 130, + 498, + 479, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 498, + 479, + 602 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_1@50" + }, + { + "bbox": [ + 130, + 498, + 479, + 602 + ], + "type": "text", + "content": " scores, our method surpasses 2G-GCN " + }, + { + "bbox": [ + 130, + 498, + 479, + 602 + ], + "type": "inline_equation", + "content": "23.0\\%" + }, + { + "bbox": [ + 130, + 498, + 479, + 602 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 498, + 479, + 602 + ], + "type": "inline_equation", + "content": "23.7\\%" + }, + { + "bbox": [ + 130, + 498, + 479, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 498, + 479, + 602 + ], + "type": "inline_equation", + "content": "22.5\\%" + }, + { + "bbox": [ + 130, + 498, + 479, + 602 + ], + "type": "text", + "content": ", respectively. The significant improvement achieved by our method indicates that the human fuser block in the CF module effectively improves the context-aware human representation learning under multi-person scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 479, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 663 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 663 + ], + "type": "text", + "content": "Single-person HOI recognition We show the quantitative results for single-person HOI recognition in Tab. 2 and Tab. 3, which are performed on CAD-120 and Bimanual Actions datasets, respectively. Results in Tab. 2 show that our method performs slightly better on CAD-120 dataset compared to SOTA method 2G-GCN, with around " + }, + { + "bbox": [ + 130, + 605, + 479, + 663 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 605, + 479, + 663 + ], + "type": "text", + "content": " improvement over 2G-GCN on all the three metrics." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 137, + 480, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 480, + 202 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{CF}" + }, + { + "bbox": [ + 130, + 137, + 480, + 202 + ], + "type": "text", + "content": ": removing context fuser module; " + }, + { + "bbox": [ + 130, + 137, + 480, + 202 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{ISR}" + }, + { + "bbox": [ + 130, + 137, + 480, + 202 + ], + "type": "text", + "content": ": removing interaction state reasoner module; " + }, + { + "bbox": [ + 130, + 137, + 480, + 202 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{L}_{fea}" + }, + { + "bbox": [ + 130, + 137, + 480, + 202 + ], + "type": "text", + "content": ": removing interaction feature loss; " + }, + { + "bbox": [ + 130, + 137, + 480, + 202 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 130, + 137, + 480, + 202 + ], + "type": "text", + "content": ": ablating the hypersphere by replacing it with Euclidean distance; " + }, + { + "bbox": [ + 130, + 137, + 480, + 202 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{CLIP} + \\mathrm{BLIP}" + }, + { + "bbox": [ + 130, + 137, + 480, + 202 + ], + "type": "text", + "content": ": ablating CLIP and BLIP model; Traditional Classifier: ablating the interaction-centric hypersphere by replacing it with a traditional classifier constructed with MLP and Softmax classifier. The improvements of our method over current SOTA method is highlighted with upward arrows." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 152, + 221, + 465, + 322 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 479, + 135 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 479, + 135 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 479, + 135 + ], + "type": "text", + "content": "Table 1: The results of joined segmentation and label recognition of sub-activity on MPHOI-72." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 152, + 221, + 465, + 322 + ], + "lines": [ + { + "bbox": [ + 152, + 221, + 465, + 322 + ], + "spans": [ + { + "bbox": [ + 152, + 221, + 465, + 322 + ], + "type": "table", + "html": "
ModelF1@10F1@25F1@50
ASSIGN [23]59.1 ± 12.151.0 ± 16.733.2 ± 14.0
2G-GCN [27]68.6 ± 10.460.8 ± 10.345.2 ± 6.5
Ours91.6 ± 0.9(↑ 23.0)84.5 ± 2.6(↑ 23.7)67.7 ± 2.2(↑ 22.5)
Ours (ΔCF)65.8 ± 12.457.6 ± 14.039.2 ± 12.6
Ours (ΔISR)80.1 ± 5.573.0 ± 8.255.6 ± 6.1
Ours (ΔLfea)73.5 ± 15.769.7 ± 13.348.8 ± 13.0
Ours (λ = 0)81.2 ± 0.774.8 ± 4.253.2 ± 0.3
Ours (ΔCLIP+BLIP)80.0 ± 6.773.0 ± 9.955.6 ± 7.4
Ours (Traditional Classifier)84.6 ± 7.274.7 ± 10.554.6 ± 13.7
", + "image_path": "d72c28bd8d4aa7d3ab66f8db5d022d180487cd6651d91f317f2f3479c5f55b65.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "type": "text", + "content": "For the Bimanual Actions dataset, our method performs as good as 2G-GCN in " + }, + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "type": "inline_equation", + "content": "F_{1}@10" + }, + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "type": "text", + "content": " while achieves " + }, + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "type": "inline_equation", + "content": "5.0\\%" + }, + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "type": "text", + "content": " higher than 2G-GCN in " + }, + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "type": "inline_equation", + "content": "F_{1}@25" + }, + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "type": "inline_equation", + "content": "F_{1}@50" + }, + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "type": "text", + "content": " score, respectively. These results indicate that our method achieves generally on par or even better performance on single-person video HOI recognition task." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 136, + 425, + 479, + 499 + ], + "blocks": [ + { + "bbox": [ + 136, + 425, + 479, + 499 + ], + "lines": [ + { + "bbox": [ + 136, + 425, + 479, + 499 + ], + "spans": [ + { + "bbox": [ + 136, + 425, + 479, + 499 + ], + "type": "image", + "image_path": "312e6ddeff50af8d8f53b23f51178b6bc41ff4155fa4c3f36cb1e8029d9a5d92.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 506, + 479, + 529 + ], + "lines": [ + { + "bbox": [ + 130, + 506, + 479, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 479, + 529 + ], + "type": "text", + "content": "Fig. 5: Qualitative ablation study results on MPHOI-72 dataset. Major prediction errors are highlighted in red dashed boxes." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 572, + 243, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 572, + 243, + 583 + ], + "spans": [ + { + "bbox": [ + 131, + 572, + 243, + 583 + ], + "type": "text", + "content": "5.4 Ablation studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "content": "In this section, we ablate the CF module and the ISR module for validating the effectiveness of these proposed components. As shown in Tab. 1, removing CF module results in more than " + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "content": " score drop of the three metrics in MPHOI-72 dataset, indicating the essential improvement of CF module in learning context-rich representations. Visualization results of temporal segmentation of interactions in Fig. 5 indicates that removing CF module results in incorrect" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "text", + "content": "Video HOI Interaction-centric context reasoning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 157, + 156, + 459, + 305 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: The results of joined segmentation and label recognition of sub-activity on CAD-120. The notations are the same with Tab. 1." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 156, + 459, + 305 + ], + "lines": [ + { + "bbox": [ + 157, + 156, + 459, + 305 + ], + "spans": [ + { + "bbox": [ + 157, + 156, + 459, + 305 + ], + "type": "table", + "html": "
ModelF1@10F1@25F1@50
rCRF [29]65.6 ± 3.261.5 ± 4.147.1 ± 4.3
Independent BiRNN [27]70.2 ± 5.564.1 ± 5.348.9 ± 6.8
ATCRF [11]72.0 ± 2.868.9 ± 3.653.5 ± 4.3
Relational BiRNN [27]79.2 ± 2.575.2 ± 3.562.5 ± 5.5
ACoLP [35]90.2 ± 2.687.4 ± 1.476.8 ± 2.3
ASSIGN [23]88.0 ± 1.884.8 ± 3.073.8 ± 5.8
2G-GCN [27]89.5 ± 1.687.1 ± 1.876.2 ± 2.8
Ours90.7 ± 2.9(↑ 0.5)88.1 ± 2.8(↑ 0.7)77.6 ± 4.7(↑ 0.8)
Ours (ΔCF)81.1 ± 4.077.0 ± 4.865.2 ± 5.6
Ours (ΔISR)88.5 ± 3.785.5 ± 3.673.9 ± 5.7
Ours (ΔLfea)89.3 ± 1.985.6 ± 2.175.9 ± 4.4
Ours (λ = 0)72.0 ± 4.465.0 ± 6.948.6 ± 6.3
Ours (ΔCLIP+BLIP)89.4 ± 2.385.5 ± 3.974.9 ± 5.7
Ours (Traditional Classifier)79.5 ± 11.073.9 ± 11.456.6 ± 12.5
", + "image_path": "0806dc041cd145d92e2250855149e34a5edc790aecf574f6a8de13ff6d575b9e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 137, + 327, + 478, + 358 + ], + "blocks": [ + { + "bbox": [ + 137, + 327, + 478, + 358 + ], + "lines": [ + { + "bbox": [ + 137, + 327, + 478, + 358 + ], + "spans": [ + { + "bbox": [ + 137, + 327, + 478, + 358 + ], + "type": "image", + "image_path": "205d20fe6fba57ba3413e8035d40578c6d108c0f2141c14c0e415090831424b9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 365, + 482, + 387 + ], + "lines": [ + { + "bbox": [ + 132, + 365, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 482, + 387 + ], + "type": "text", + "content": "Fig. 6: Qualitative ablation study results on CAD-120 dataset. Major prediction errors are highlighted in red dashed boxes." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 413, + 479, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 479, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 479, + 460 + ], + "type": "text", + "content": "interaction predictions in both humans (highlighted in red in Fig. 5). Similarly, in CAD-120 and Bimanual Actions datasets, deleting CF module also results in massive " + }, + { + "bbox": [ + 130, + 413, + 479, + 460 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 130, + 413, + 479, + 460 + ], + "type": "text", + "content": " score drop. The visualization results in Fig. 6 suggests an incorrect prediction segment when removing CF module." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": "Furthermore, we ablate hypersphere by replacing it with Euclidean distance, where " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " in Eq. 6. Results in Tab. 1, 2 and 3 indicate that utilizing Euclidean distance results in at least " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " scores drop in MPHOI, CAD-120 and Bimanual Actions dataset, respectively. Therefore, we conclude that Euclidean distance do not introduce HOI structure priors, ignoring valuable structure information of HOI for guiding predictions. Subsequently, we ablate CLIP and BLIP models by randomly initialize interaction features and context features. Results in Tab. 1, 2 and 3 indicate that removing CLIP and BLIP models results in some drop of model performance, but is still on par with or better than SOTA methods. Thus, it is the intricately designed structure of our model that substantiates the substantial enhancement in performance. Finally, we ablate the interaction-centric hypersphere by replacing it with a traditional classifier constructed with multi-layer perceptron (MLP) and Softmax classifier. The outcomes, as presented in Tab. 1, 2 and 3, reveal a notable decline of over " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " score within the MPHOI dataset when employing the traditional classifier. Likewise, in the CAD-120 dataset, the traditional classifier results in a substantial decrease of more than " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " score. Additionally, within" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 157, + 156, + 459, + 297 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 3: The results of joined segmentation and label recognition of sub-activity on Bimanual Actions. The notations are the same with Tab. 1." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 156, + 459, + 297 + ], + "lines": [ + { + "bbox": [ + 157, + 156, + 459, + 297 + ], + "spans": [ + { + "bbox": [ + 157, + 156, + 459, + 297 + ], + "type": "table", + "html": "
ModelF1@10F1@25F1@50
Dreher et al. [4]40.6 ± 7.234.8 ± 7.122.2 ± 5.7
Independet BiRNN [27]74.7 ± 7.072.0 ± 7.061.8 ± 7.3
Relational BiRNN [27]77.7 ± 3.975.0 ± 4.264.8 ± 5.3
ASSIGN [23]84.0 ± 2.081.2 ± 2.068.5 ± 3.3
2G-GCN [27]85.0 ± 2.282.0 ± 2.669.2 ± 3.1
Ours85.0 ± 2.582.9 ± 2.9(↑ 0.9)74.2 ± 4.3(↑ 5.0)
Ours (ΔCF)82.5 ± 5.080.5 ± 5.571.1 ± 7.0
Ours (ΔISR)84.1 ± 2.381.8 ± 2.873.0 ± 3.7
Ours (ΔLfea)84.5 ± 4.682.0 ± 5.271.8 ± 6.9
Ours (λ = 0)76.7 ± 5.274.3 ± 6.065.2 ± 6.3
Ours (ΔCLIP+BLIP)84.3 ± 1.481.8 ± 1.873.2 ± 2.7
Ours (Traditional Classifier)82.0 ± 3.679.8 ± 4.171.0 ± 5.6
", + "image_path": "7e934948a28f71d787f292856f328d8af3d94d1e8a51c04c7b51be967645edc4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 330, + 482, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 367 + ], + "type": "text", + "content": "the Bimanual Actions dataset, the traditional classifier induces a decline exceeding " + }, + { + "bbox": [ + 130, + 330, + 482, + 367 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 330, + 482, + 367 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 330, + 482, + 367 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 130, + 330, + 482, + 367 + ], + "type": "text", + "content": " score. These findings unanimously underscore the efficacy of modeling HOI manifold structures by the hypersphere module." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 134, + 390, + 302, + 464 + ], + "blocks": [ + { + "bbox": [ + 134, + 390, + 302, + 464 + ], + "lines": [ + { + "bbox": [ + 134, + 390, + 302, + 464 + ], + "spans": [ + { + "bbox": [ + 134, + 390, + 302, + 464 + ], + "type": "image", + "image_path": "c21d56e15216e69ceabf377e4c562f2d66908b903faba48b5f59ccb6fcc1371d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 472, + 482, + 529 + ], + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 529 + ], + "type": "text", + "content": "Fig. 7: (A) Visualization of interaction-centric hyperspheres, the learned interaction representations (large points locate at the center of hyperspheres) and human-object entity representations (small points surrounding the hyperspheres) in embedding space. Selected frame samples are shown for each interaction-centric hypersphere. (B) 2D visualization of interaction-centric hyperspheres." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 389, + 493, + 452 + ], + "blocks": [ + { + "bbox": [ + 309, + 389, + 493, + 452 + ], + "lines": [ + { + "bbox": [ + 309, + 389, + 493, + 452 + ], + "spans": [ + { + "bbox": [ + 309, + 389, + 493, + 452 + ], + "type": "image", + "image_path": "448c3a360fa59d726c6c2503f919a4330a2a42cb743008341e1ff2d6e57b3d10.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 571, + 254, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 571, + 254, + 583 + ], + "spans": [ + { + "bbox": [ + 131, + 571, + 254, + 583 + ], + "type": "text", + "content": "5.5 Qualitative results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We show some visualization results on MPHOI in Fig 8 to compare our method with SOTA method 2G-GCN. The red highlighted boxes indicate that 2G-GCN tend to generate unreasonable interaction predictions, while our method generates more reasonable interaction predictions. The visualization results in Fig. 9 show similar prediction pattern where 2G-GCN predicted some unreasonable short segments (highlighted in red boxes) while our method predicts more accu" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "text", + "content": "Video HOI Interaction-centric context reasoning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "rately. We also visualize the interaction-centric hypersphere, the learned interaction representations and the human-object entity representations in embedding space in Fig. 7. Results in Fig. 7 (A) show that human-object entity representations (small dots in Fig. 7 (A)) belonging to the specific interaction class locates near the surface of the corresponding hypersphere. Similar with Fig. 7 (A), we also visualize the intraction-centric hyerspheres in 2D sapce in Fig. 7 (B). Results in Fig. 7 (B) indicate that HOIs distribute around the sphere of different interaction classes. The distributions of HOIs on each sphere also differ a lot, where the sphere belongs to the solve interaction class captures HOIs on the upper part of its sphere, while the sphere belongs to the cut interaction class captures HOIs on the lower part of its sphere. These results suggest that our model successfully model the manifold structure of HOI." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 136, + 282, + 479, + 355 + ], + "blocks": [ + { + "bbox": [ + 136, + 282, + 479, + 355 + ], + "lines": [ + { + "bbox": [ + 136, + 282, + 479, + 355 + ], + "spans": [ + { + "bbox": [ + 136, + 282, + 479, + 355 + ], + "type": "image", + "image_path": "e655df80783c4250729bb875fb971ec4912271b21a525674760f66faadeac719.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 361, + 482, + 384 + ], + "lines": [ + { + "bbox": [ + 131, + 361, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 361, + 482, + 384 + ], + "type": "text", + "content": "Fig. 8: Visualization results on MPHOI-72 dataset. Major prediction errors are highlighted in red dashed boxes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 430, + 479, + 460 + ], + "blocks": [ + { + "bbox": [ + 136, + 430, + 479, + 460 + ], + "lines": [ + { + "bbox": [ + 136, + 430, + 479, + 460 + ], + "spans": [ + { + "bbox": [ + 136, + 430, + 479, + 460 + ], + "type": "image", + "image_path": "50faa1b4a7cee1191ab92636440d64e83d5d4fdc20b17c9ee3ac123369f10cf3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 465, + 482, + 489 + ], + "lines": [ + { + "bbox": [ + 131, + 465, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 131, + 465, + 482, + 489 + ], + "type": "text", + "content": "Fig. 9: Visualization results on CAD-120 dataset. Major prediction errors are highlighted in red dashed boxes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 530, + 220, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 530, + 220, + 544 + ], + "spans": [ + { + "bbox": [ + 131, + 530, + 220, + 544 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 555, + 482, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 652 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 652 + ], + "type": "text", + "content": "In this work, we propose an interaction-centric spatio-temporal context reasoning network for multi-person video HOI recognition. Specifically, we propose a context fuser and an interaction state reasoner to learn spatio-temporal context-rich and reasoning-aware entity representations. We further represent HOI components with an interaction-centric hypersphere for HOI classification. Experiment results show that our method outperforms SOTA method by more than " + }, + { + "bbox": [ + 130, + 555, + 482, + 652 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 130, + 555, + 482, + 652 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 130, + 555, + 482, + 652 + ], + "type": "text", + "content": " score in multi-person scenarios, and achieves competitive results on single-person cases compared to SOTA methods." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 239 + ], + "type": "text", + "content": "This material is based upon work supported under the AI Research Institutes program by National Science Foundation and the Institute of Education Sciences, U.S. Department of Education through Award #2229873 - AI Institute for Transforming Education for Children with Speech and Language Processing Challenges. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation, the Institute of Education Sciences, or the U.S. Department of Education." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 260, + 198, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 198, + 273 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 198, + 273 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 287, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 136, + 287, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 287, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 136, + 287, + 481, + 331 + ], + "type": "text", + "content": "1. Chopra, S., Hadsell, R., LeCun, Y.: Learning a similarity metric discriminatively, with application to face verification. In: 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05). vol. 1, pp. 539-546 vol. 1 (2005). https://doi.org/10.1109/CVPR.2005.2029" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 333, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 333, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 136, + 333, + 481, + 364 + ], + "type": "text", + "content": "2. Deng, S., Zhang, N., Chen, H., Tan, C., Huang, F., Xu, C., Chen, H.: Low-resource extraction with knowledge-aware pairwise prototype learning. Knowledge-Based Systems 235, 107584 (2022) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 365, + 481, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 365, + 481, + 409 + ], + "spans": [ + { + "bbox": [ + 137, + 365, + 481, + 409 + ], + "type": "text", + "content": "3. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 410, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 410, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 410, + 481, + 453 + ], + "type": "text", + "content": "4. Dreher, C.R.G., Wächter, M., Asfour, T.: Learning object-action relations from bimanual human demonstration using graph networks. IEEE Robotics and Automation Letters (RA-L) 5(1), 187-194 (2020). https://doi.org/10.1109/LRA.2019.2949221 3, 10, 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 454, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 481, + 498 + ], + "type": "text", + "content": "5. Gao, C., Xu, J., Zou, Y., Huang, J.B.: Drg: Dual relation graph for human-object interaction detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16. pp. 696-712. Springer (2020) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 499, + 481, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 499, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 499, + 481, + 532 + ], + "type": "text", + "content": "6. Gao, K., Chen, L., Zhang, H., Xiao, J., Sun, Q.: Compositional prompt tuning with motion cues for open-vocabulary video relation detection. In: The Eleventh International Conference on Learning Representations (2022) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 533, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 481, + 575 + ], + "type": "text", + "content": "7. Gupta, T., Schwing, A., Hoiem, D.: No-frills human-object interaction detection: Factorization, layout encodings, and training techniques. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9677-9685 (2019) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 576, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 576, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 576, + 481, + 620 + ], + "type": "text", + "content": "8. Iftekhar, A., Chen, H., Kundu, K., Li, X., Tighe, J., Modolo, D.: What to look at and where: Semantic and spatial refined transformer for detecting human-object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5353-5363 (2022) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "text", + "content": "9. Kim, B., Choi, T., Kang, J., Kim, H.J.: Uniondet: Union-level detector towards real-time human-object interaction detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 498-514. Springer (2020) 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "text", + "content": "Video HOI Interaction-centric context reasoning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "10. Kim, B., Lee, J., Kang, J., Kim, E.S., Kim, H.J.: Hotr: End-to-end human-object interaction detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 74-83 (June 2021) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "text", + "content": "11. Koppula, H.S., Saxena, A.: Anticipating human activities using object affordances for reactive robotic response. IEEE transactions on pattern analysis and machine intelligence 38(1), 14-29 (2015) 12" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 225 + ], + "type": "text", + "content": "12. Koppula, H.S., Gupta, R., Saxena, A.: Learning human activities and object affordances from rgb-d videos. The International journal of robotics research 32(8), 951-970 (2013) 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 269 + ], + "type": "text", + "content": "13. Koppula, H.S., Gupta, R., Saxena, A.: Learning human activities and object affordances from rgb-d videos. Int. J. Rob. Res. 32(8), 951-970 (jul 2013). https://doi.org/10.1177/0278364913478446, https://doi.org/10.1177/0278364913478446 10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "text", + "content": "14. Lea, C., Flynn, M.D., Vidal, R., Reiter, A., Hager, G.D.: Temporal convolutional networks for action segmentation and detection. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 156-165 (2017) 10" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "15. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 380 + ], + "type": "text", + "content": "16. Li, Y.L., Zhou, S., Huang, X., Xu, L., Ma, Z., Fang, H.S., Wang, Y., Lu, C.: Transferable interactiveness knowledge for human-object interaction detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3585-3594 (2019) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 423 + ], + "type": "text", + "content": "17. Liao, Y., Liu, S., Wang, F., Chen, Y., Qian, C., Feng, J.: Ppdm: Parallel point detection and matching for real-time human-object interaction detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 482-490 (2020) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "text", + "content": "18. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017) 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "text", + "content": "19. Liu, X., Li, Y.L., Wu, X., Tai, Y.W., Lu, C., Tang, C.K.: Interactiveness field in human-object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20113-20122 (2022) 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "text", + "content": "20. Luan, T., Wang, Y., Zhang, J., Wang, Z., Zhou, Z., Qiao, Y.: Pc-hmr: Pose calibration for 3d human mesh recovery from 2d images/videos. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 2269-2276 (2021) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 523, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 566 + ], + "type": "text", + "content": "21. Luan, T., Zhai, Y., Meng, J., Li, Z., Chen, Z., Xu, Y., Yuan, J.: High fidelity 3d hand shape reconstruction via scalable graph frequency decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16795-16804 (2023) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "type": "text", + "content": "22. Mettes, P., Van der Pol, E., Snoek, C.: Hyperspherical prototype networks. Advances in neural information processing systems 32 (2019) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 589, + 481, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 631 + ], + "type": "text", + "content": "23. Morais, R., Le, V., Venkatesh, S., Tran, T.: Learning asynchronous and sparse human-object interaction in videos. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16041-16050 (2021) 2, 4, 11, 12, 13" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "24. Nagarajan, T., Feichtenhofer, C., Grauman, K.: Grounded human-object interaction hotspots from video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8688-8697 (2019) 4" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 611 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "text", + "content": "25. Park, J., Park, J.W., Lee, J.S.: Viplo: Vision transformer based pose-conditioned self-loop graph for human-object interaction detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17152-17162 (June 2023) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 162, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 162, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 482, + 194 + ], + "type": "text", + "content": "26. Qi, S., Wang, W., Jia, B., Shen, J., Zhu, S.C.: Learning human-object interactions by graph parsing neural networks. In: Proceedings of the European Conference on Computer Vision (ECCV) (September 2018) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 237 + ], + "type": "text", + "content": "27. Qiao, T., Men, Q., Li, F.W., Kubotani, Y., Morishima, S., Shum, H.P.: Geometric features informed multi-person human-object interaction recognition in videos. In: European Conference on Computer Vision. pp. 474-491. Springer (2022) 2, 3, 4, 6, 10, 11, 12, 13" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 237, + 482, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 282 + ], + "type": "text", + "content": "28. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 282, + 482, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 304 + ], + "type": "text", + "content": "29. Sener, O., Saxena, A.: rcrf: Recursive belief estimation over crfs in rgb-d activity videos. In: Robotics: Science and systems (2015) 12" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 304, + 482, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 336 + ], + "type": "text", + "content": "30. Sunkesula, S.P.R., Dabral, R., Ramakrishnan, G.: Lighten: Learning interactions with graph and hierarchical temporal networks for hoi in videos. In: Proceedings of the 28th ACM international conference on multimedia. pp. 691-699 (2020) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 336, + 482, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 380 + ], + "type": "text", + "content": "31. Tamura, M., Ohashi, H., Yoshinaga, T.: Qpic: Query-based pairwise human-object interaction detection with image-wide contextual information. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10410-10419 (June 2021) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "text", + "content": "32. Tu, D., Sun, W., Min, X., Zhai, G., Shen, W.: Video-based human-object interaction detection from tubelet tokens. Advances in Neural Information Processing Systems 35, 23345-23357 (2022) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 413, + 482, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 457 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 457 + ], + "type": "text", + "content": "33. Wang, T., Anwer, R.M., Khan, M.H., Khan, F.S., Pang, Y., Shao, L., Laaksonen, J.: Deep contextual attention for human-object interaction detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5694-5702 (2019) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 457, + 482, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 482, + 490 + ], + "type": "text", + "content": "34. Xi, N., Meng, J., Yuan, J.: Chain-of-look prompting for verb-centric surgical triplet recognition in endoscopic videos. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 5007-5016 (2023) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 490, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 522 + ], + "type": "text", + "content": "35. Xi, N., Meng, J., Yuan, J.: Open set video hoi detection from action-centric chain-of-look prompting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3079-3089 (2023) 12" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 522, + 482, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 567 + ], + "type": "text", + "content": "36. Zhang, Y., Pan, Y., Yao, T., Huang, R., Mei, T., Chen, C.W.: Exploring structure-aware transformer over interaction proposals for human-object interaction detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19548-19557 (2022) 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 567, + 482, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 611 + ], + "type": "text", + "content": "37. Zhong, X., Qu, X., Ding, C., Tao, D.: Glance and gaze: Inferring action-aware points for one-stage human-object interaction detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13234-13243 (2021) 4" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "text", + "content": "Video HOI Interaction-centric context reasoning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Interactive 3D Object Detection with Prompts/20a1b205-7505-40a6-b499-29fffaad19e8_content_list.json b/2024/Interactive 3D Object Detection with Prompts/20a1b205-7505-40a6-b499-29fffaad19e8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1eff9c28fb2569d40951feb59e7c562d52d17330 --- /dev/null +++ b/2024/Interactive 3D Object Detection with Prompts/20a1b205-7505-40a6-b499-29fffaad19e8_content_list.json @@ -0,0 +1,1965 @@ +[ + { + "type": "text", + "text": "Interactive 3D Object Detection with Prompts", + "text_level": 1, + "bbox": [ + 233, + 141, + 769, + 165 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruifei Zhang $^{1,2\\star}$ , Xiangru Lin $^{4\\star}$ , Wei Zhang $^{4}$ , Jincheng Lu $^{4}$ , Xuekuan Wang $^{4}$ , Xiao Tan $^{4}$ , Yingying Li $^{4}$ , Errui Ding $^{4}$ , Jingdong Wang $^{4}$ , and Guanbin Li $^{3,5\\star\\star}$", + "bbox": [ + 218, + 188, + 781, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The Chinese University of Hong Kong, Shenzhen, China", + "bbox": [ + 305, + 231, + 696, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Shenzhen Research Institute of Big Data, Shenzhen, China", + "bbox": [ + 297, + 246, + 704, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Sun Yat-sen University, Guangzhou, China", + "bbox": [ + 349, + 258, + 651, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Department of Computer Vision Technology (VIS), Baidu Inc., China", + "bbox": [ + 261, + 273, + 738, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Peng Cheng Laboratory", + "bbox": [ + 413, + 287, + 589, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ruifeizhang@link.cuhk.edu.cn, liguanbin@mail.sysu.edu.cn,", + "bbox": [ + 279, + 301, + 720, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{linxiangru,zhangwei99,lujincheng01,wangxuekuan,tanxiao01,", + "bbox": [ + 279, + 316, + 722, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "liyingying05,dingerrui,wangjingdong}@baidu.com", + "bbox": [ + 321, + 330, + 679, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The evolution of 3D object detection hinges not only on advanced models but also on effective and efficient annotation strategies. Despite this progress, the labor-intensive nature of 3D object annotation remains a bottleneck, hindering further development in the field. This paper introduces a novel approach, incorporated with \"prompt in 2D, detect in 3D\" and \"detect in 3D, refine in 3D\" strategies, to 3D object annotation: multi-modal interactive 3D object detection. Firstly, by allowing users to engage with simpler 2D interaction prompts (e.g., clicks or boxes on a camera image or a bird's eye view), we bridge the complexity gap between 2D and 3D spaces, reimagining the annotation workflow. Besides, Our framework also supports flexible iterative refinement to the initial 3D annotations, further assisting annotators in achieving satisfying results. Evaluation on the nuScenes dataset demonstrates the effectiveness of our method. And thanks to the prompt-driven and interactive designs, our approach also exhibits outstanding performance in open-set scenarios. This work not only offers a potential solution to the 3D object annotation problem but also paves the way for further innovations in the 3D object detection community.", + "bbox": [ + 259, + 378, + 743, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Interactive 3D Object Detection $\\cdot$ Prompt Learning", + "bbox": [ + 261, + 641, + 696, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 680, + 375, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D object detection is pivotal for autonomous driving, gaining significant traction in recent years. Current research often derives 3D object information from sources such as monocular [24, 39, 51, 54] or multi-view camera images [11, 12, 18, 21], point cloud [17, 46, 53], and multi-modal sensors [1, 5, 45]. This has led to remarkable progress in detection performance. A major catalyst behind these advancements is the availability of high-quality 3D datasets like KITTI [9],", + "bbox": [ + 212, + 710, + 787, + 804 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 215, + 811, + 364, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*\\* Corresponding author.", + "bbox": [ + 215, + 825, + 383, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d7c34366f328e5c3679cc1fce84e33fa557cbd43240bb26a903fc1f25d9c43cf.jpg", + "image_caption": [ + "Traditional Way of 3D Detection Data Annotation", + "Our proposed Way of 3D Detection Data Annotation" + ], + "image_footnote": [], + "bbox": [ + 276, + 157, + 727, + 219 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e778823b744d52fa4b4764e650138b8d133fd6521d8cb6e1b56110d5cf50896a.jpg", + "image_caption": [ + "Fig. 1: We introduce an interactive 3D object detection concept. Rather than the traditional method where a model produces 3D object detection boxes followed by user refinement in 3D space, our strategy is \"prompt in 2D, detect in 3D, refine in 3D\". This method streamlines the annotation process, reducing the need for user refinements." + ], + "image_footnote": [], + "bbox": [ + 276, + 231, + 725, + 328 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "nuScenes [2], and Waymo [36]. These datasets fuel the progression of data-driven algorithms in 3D object detection. However, the intricate nature of 3D object annotation often necessitates significant human resources and incurs high labor costs. On average, annotating an hour of driving data takes hundreds of hours [27]. This annotation challenge potentially restricts the scale and diversity of datasets, possibly hindering further progression in this field.", + "bbox": [ + 212, + 429, + 787, + 520 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Addressing this annotation challenge is the central theme of our paper. While weakly-supervised [25, 27-29] and semi-supervised [38, 43, 50] object detection methods exist, they typically suffer from reduced performance, especially in practical scenarios. This does not substantially address the overarching issue of the data-hungry nature of the field. With the recent emergence of SAM [13], both industry and academia are now recognizing the potential of interactive annotation approaches.", + "bbox": [ + 212, + 521, + 787, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work aims to integrate this concept into 3D object detection, thus simplifying the annotation process, as presented in Fig. 1. Specifically, we observe that within the traditional process of annotating 3D detection data, annotators tend to concentrate on addressing two challenges stemming from imperfect pretrained 3D detectors: 1) supplementing missing closed-set objects and incorporating additional required open-set ones; 2) rectifying existing erroneous annotations. Both scenarios necessitate annotators to operate within the 3D point cloud space, which is extremely time-consuming and labor-intensive.", + "bbox": [ + 212, + 628, + 787, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To circumvent this, we create an interactive 3D object detection system enhanced by two strategies: prompt in $2D$ , detect in $3D$ , and detect in $3D$ , refine in $3D$ . The first principle enables users to interact with a 2D camera image or bird's eye view (BEV) of the point cloud, prompting the model to generate corresponding 3D bounding boxes, instead of encountering the 3D point cloud", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "R. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "directly. Furthermore, benefiting from the prompt-driven annotation paradigm, our method inherently possesses the capability to annotate open-set objects. The latter ensures annotators refine the initial 3D bounding box at the fastest pace with minimal effort, which significantly streamlines operations in the 3D space.", + "bbox": [ + 212, + 146, + 782, + 219 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In light of these principles, we present our multi-modal interactive 3D object detection task and algorithm. Firstly, it employs prompts in the form of clicks or boxes on a 2D camera or BEV view, guiding the model to generate the relevant 3D bounding box. Our framework is reminiscent of the DETR-series detectors [3, 20, 26, 40] but comes with distinctive features. We integrate user inputs as prompt tokens and design a learnable component that investigates multi-modal features of indicated objects, informed by the prompts. Furthermore, we incorporate a global-to-local strategy to supply fine-grained object features, enhancing detection accuracy. Secondly, we empower our model with iterative refinement capability by incorporating the predicted 3D bounding box encoding as a new prompt into the query sequence. For those inaccurate 3D predictions, our framework allows annotators to adjust a few wrong attributes, e.g. dragging the center point position or adjusting the orientation angle of the 3D bounding box, and then perform self-correction on other characters.", + "bbox": [ + 212, + 222, + 784, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To rigorously evaluate our approach, we devise two evaluation configurations for practical annotation scenarios. The first, regarding Annotation from Raw Data, leverages user 2D prompts and produces corresponding 3D predictions. Besides, our framework is also compatible with starting from the results of existing State-of-the-Art (SOTA) 3D detection models, termed as Annotation from SOTA Model and iteratively refining them. Experimental results on the nuScenes dataset highlight the effectiveness of our method in various scenarios.", + "bbox": [ + 212, + 435, + 784, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, this paper has the following contributions:", + "bbox": [ + 238, + 541, + 640, + 555 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a new concept of multi-modal interactive 3D object detection, proposing a substantial solution to prevailing annotation challenges. Our work is expected to provide critical insights for the 3D object detection community across academic and industrial domains.", + "- We unveil a prompt-driven interactive 3D object detection methodology with self-correction capability, leveraging it to establish an interactive system tailored for practical annotation scenarios.", + "- Comprehensive experiments in both annotation scenarios and evaluation settings attest to our method's efficacy." + ], + "bbox": [ + 223, + 568, + 782, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 731, + 387, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 3D Object Detection", + "text_level": 1, + "bbox": [ + 214, + 766, + 433, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D object detection techniques have been developed across a range of input modalities. Our approach emphasizes the advantages of interactive inputs, setting it apart from traditional methods:", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Interactive 3D Object Detection with Prompts", + "bbox": [ + 419, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Camera-only. Monocular 3D object detection techniques [24, 39, 51, 54] enjoy advantages such as low cost and ease of deployment. However, it also faces the inherent challenge of limited depth information. Multi-view methods utilize data from multiple camera perspectives [11, 12, 18, 21, 22, 30, 41], which enhances the model's ability to perceive the 3D environment, thereby improving detection accuracy.", + "bbox": [ + 212, + 146, + 785, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LiDAR-only. Benefiting from the robustness and high precision, numerous studies utilize LiDAR data for 3D object detection. Point-based methods [17, 33,48] like PointNet [31] directly engage with raw LiDAR data, while grid-based strategies structure the data into 3D voxels [46,53] or feature pillars [14].", + "bbox": [ + 212, + 237, + 785, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-modality. Capitalizing on the strengths of camera images and point clouds, multi-modal 3D object detection has seen significant advancements [1, 5, 45]. Early methods combine camera images and point clouds at the input or result stages [5, 34]. While, current models, like TransFusion [1] and CMT [45], emphasize more intricate feature-level fusion. Uniquely, our method embeds interactive capacities, ensuring richer, user-tailored annotations.", + "bbox": [ + 212, + 297, + 785, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Interactive Object Detection/Segmentation", + "text_level": 1, + "bbox": [ + 214, + 409, + 619, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Interactive methods have historically aimed to reduce the annotation workload while maintaining, or even enhancing, model performance:", + "bbox": [ + 212, + 431, + 785, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Interactive Object Detection. Early attempts, like the incremental learning strategy [49], require user corrections to update detectors. More recent models like C3Det [15] have substantially eased the annotation process. Tools such as LATTE [37] and iDet3D [7] further simplify the 3D object annotation process. Our framework, however, brings in the versatility of prompt formats and modalities, offering users a broadened interaction spectrum coupled with robust detection capabilities.", + "bbox": [ + 212, + 463, + 785, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Interactive Object Segmentation. Interactive segmentation has matured significantly over the years [4,35,44]. Models like SAM [13] have empowered users with tools like clicks, boxes, and referring expressions to guide the segmentation process. Diverging from interactive segmentation confined to 2D spaces, our work addresses the challenge of integrating 2D interactive prompts with 3D object detection, mitigating the spatial discrepancy between 2D annotations and 3D environments.", + "bbox": [ + 212, + 569, + 785, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Weakly-supervised Object Detection", + "text_level": 1, + "bbox": [ + 214, + 695, + 563, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The notion of harnessing weak signals, like clicks [29] or extreme-points [25, 28], for supervising detectors has been a topic of intrigue. Some methods [27] have extended this idea to 3D object detection, using center-click based strategies. These approaches, while commendable for their balance between performance and annotation overhead, sometimes fail to harness the full potential of the detectors. Our methodology, on the other hand, incorporates strong 3D bounding box supervisions in an interactive setting, streamlining the data acquisition process without compromising on detection performance.", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "R. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/711e38e5001e2e2f615a72da8466f938568c45427604ddda6d7dacce6293f55d.jpg", + "image_caption": [ + "Fig. 2: An overview of our proposed interactive 3D object detection system/method. Taking multi-modal data with user prompts as input, we explicitly add 3D information $\\mathbf{P}_{im},\\mathbf{P}_{pc}$ into features $\\mathbf{F}_{im},\\mathbf{F}_{pc}$ by 3D Position Encoding. Object queries consist of a learnable token $\\mathbf{Q}_l$ and prompt ones $\\mathbf{Q}_p$ , which are generated by Prompt-Guided Query generation. We also introduce a Global-to-Local enhancement strategy to dynamically adjust features integrated with queries from global representations to local RoI details. Besides, our framework allows users to refine any of 3D attributes directly, and encode the refined 3D bounding box $b_{L}^{r^{\\prime}}$ and category information $c_{L}^{r^{\\prime}}$ , along with attribute refinement record $p_a^r$ and $p_c^r$ , into a new prompt token $\\mathbf{Q}_{3D}$ by 3D Attribute Encoding. $g,w,l,h,\\theta$ denotes the gravity center, width, length, height and yaw angle respectively. This iterative refinement propels our model to automatically rectify other attributes, ultimately delivering satisfactory results." + ], + "image_footnote": [], + "bbox": [ + 233, + 145, + 772, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 482, + 380, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem Definition", + "text_level": 1, + "bbox": [ + 215, + 512, + 419, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present an interactive multi-modal 3D object detection algorithm. Given a multi-view set of camera images and associated point cloud data, the objective is to predict both the class labels and the 3D bounding boxes, driven by user prompts either in the form of a point or a box. Further, our model also supports iteratively self-correction with limited human assistance. To supervise the iteration process, the Ground Truth $(c, b)$ is employed. Here, $c$ signifies the category label, and $b$ represents the 3D attributes. We adopt a deep supervision strategy for our multi-layer Transformer decoder, which guides the predictions across each decoder layer of each iteration, denoted as $\\{(c_i^r, b_i^r), i = 1, 2, \\dots, L, r = 0, 1, \\dots, M\\}$ , with $L$ and $M$ being the total number of decoder layers and iterative refinements.", + "bbox": [ + 212, + 537, + 787, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Overview", + "text_level": 1, + "bbox": [ + 215, + 709, + 339, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The architecture underpinning our interactive 3D object detection method is visualized in Fig. 2, which could be roughly decomposed into two stages, i.e. \"prompt in $2D$ , detect in $3D$ and \"detect in $3D$ , refine in $3D$ \". In the first phase, two separate backbones are utilized to extract multi-view image features $\\mathbf{F}_{im} = \\{\\mathbf{F}_{im}^{t}, t = 1,2,\\dots,T\\}$ and point cloud features $\\mathbf{F}_{pc}$ respectively. To enhance these features with 3D-awareness, following PETR [21] and CMT [45], we introduce 3D position encoding. Specifically, for the multi-view images, this encoding is", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Interactive 3D Object Detection with Prompts", + "bbox": [ + 418, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "represented as $\\mathbf{P}_{im} = \\{\\mathbf{P}_{im}^t, t = 1,2,\\dots,T\\}$ , and for the point cloud, it is denoted as $\\mathbf{P}_{pc}$ . For object queries, denoted as $\\mathbf{Q}$ , we utilize a unique blending of a learnable token, $\\mathbf{Q}_l$ , and prompt-specific tokens, $\\mathbf{Q}_p$ , for every detected object. The prompt queries, $\\mathbf{Q}_p$ , play a pivotal role in guiding the learnable token, $\\mathbf{Q}_l$ , by leveraging the insights from the multi-modal features. Furthermore, to enhance the detection accuracy across various object scales, we incorporate a global-to-local strategy within the decoder. This strategy ensures that the object queries are continually enriched by granular local features, enabling finer detections. Ultimately, the learnable token predicts the object's class and its 3D bounding box $(c_L^0,b_L^0)$ .", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Subsequently, the model enters an interactive refinement stage. This stage reuses the multi-modal features but updates the object queries $\\mathbf{Q}$ by inserting the encoding of the refined 3D prediction from the current iteration. Specifically, in the $r$ -th iteration, we enable annotators to fine-tune the current prediction $(c_{L}^{r}, b_{L}^{r})$ and encode the refined result $(c_{L}^{r'}, b_{L}^{r'})$ and corresponding attribute refinement record $(p_{c}^{r}, p_{a}^{r})$ into a new prompt $\\mathbf{Q}_{3D}$ . $\\mathbf{Q}_{3D}$ and $\\mathbf{Q}_p$ collectively drive $\\mathbf{Q}_l$ to generate more accurate predictions in the next iteration. Note that we omit the iteration superscript of $\\mathbf{Q}_{3D}$ and $\\mathbf{Q}$ for simplicity.", + "bbox": [ + 212, + 297, + 787, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 3D Position Encoding", + "text_level": 1, + "bbox": [ + 215, + 439, + 444, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-view Camera Images. The core idea of embedding 3D position for multi-view image features includes two steps: 1) modeling each pixel coordinate as a series of points along a ray in the camera frustum space, and 2) projecting camera frustum coordinates into 3D space. Specifically, for a given camera image feature, denoted as $\\mathbf{F}_{im}^{t}\\in \\mathbb{R}^{W_{im}\\times H_{im}\\times C}$ , each pixel can be visualized as tracing a ray within the camera frustum space. This perspective allows us to sample $D$ discrete points along the depth dimension. This process is mathematically captured by:", + "bbox": [ + 212, + 463, + 787, + 584 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {c f} (u, v, j) = \\left(u \\times d _ {j}, v \\times d _ {j}, d _ {j}, 1\\right) ^ {T}, j = 1, 2, \\dots , D \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 595, + 784, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\mathbf{P}_{cf}$ indicates the camera's frustum coordinates. Pixel coordinates in the image are denoted by $(u,v)$ , and $d_{j}$ represents the $j$ -th depth value.", + "bbox": [ + 212, + 621, + 785, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To relate these coordinates to a 3D context, we employ an inverse 3D projection method:", + "bbox": [ + 212, + 652, + 785, + 680 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {3 d} ^ {t} (u, v, j) = \\mathbf {K} _ {t} ^ {3 D} \\mathbf {K} _ {t} ^ {- 1} \\mathbf {P} _ {c f} (u, v, j) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 681, + 784, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{K}_t^{3D} \\in \\mathbb{R}^{4 \\times 4}$ represents the transformation matrix that links 3D space with the $t$ -th camera's coordinate space. $\\mathbf{K}_t \\in \\mathbb{R}^{4 \\times 4}$ is the intrinsic matrix of $t$ -th camera. To ensure consistent value scales, we normalize the 3D coordinates to the range [0,1]. The final 3D position encoding for the $t$ -th camera image, represented as $\\mathbf{P}_{im}^t \\in \\mathbb{R}^{W_{im} \\times H_{im} \\times C}$ , is derived via a multi-layer perceptron (MLP): $\\mathbf{P}_{im}^t = \\mathrm{MLP}_{im}(\\mathbf{P}_{3d}^t)$ .", + "bbox": [ + 212, + 703, + 787, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Point Cloud. In contrast to the approach for camera image features, where points are assembled along the depth axis, the LiDAR point set necessitates sampling along the height axis. Specifically, given a BEV feature map, denoted as", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "R. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathbf{F}_{pc} \\in \\mathbb{R}^{W_{pc} \\times H_{pc} \\times C}$ , we can compute the corresponding 3D position information, represented as $\\mathbf{P}_{pc} \\in \\mathbb{R}^{W_{pc} \\times H_{pc} \\times C}$ , using the following steps: First, we determine the 3D position based on the feature map's coordinates:", + "bbox": [ + 212, + 145, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {3 d} (u, v, j) = \\left(u \\times u _ {s}, v \\times v _ {s}, h _ {j}, 1\\right) ^ {T}, j = 1, 2, \\dots , H \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 200, + 785, + 217 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, $(u,v)$ specifies the coordinate within the point cloud feature map. The variables $(u_{s},v_{s})$ represent the dimensions of the feature grid. $H$ is the number of sampled points along the height axis and $h_j$ is the $j$ -th height value. Next, we utilize a multi-layer perceptron (MLP) to compute the 3D position encoding from the derived 3D position: $\\mathbf{P}_{pc} = \\mathrm{MLP}_{pc}(\\mathbf{P}_{3d})$ . Note that as in CMT, only one point is sampled along the height axis, thus simplifying the process into standard 2D position encoding. Consequently, in our methodology, we have adopted a sinusoidal embedding to cater to this behavior.", + "bbox": [ + 212, + 226, + 784, + 347 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Prompt-guided Query Generation", + "text_level": 1, + "bbox": [ + 214, + 367, + 540, + 382 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our approach provides an avenue for user interaction, distinct from traditional DETR-based detectors that often rely on multiple queries to discern image objects. By merging learnable tokens, $\\mathbf{Q}_l$ , with user-directed prompt tokens, $\\mathbf{Q}_p$ , we achieve object detection that is not only efficient but also offers a new angle on object localization.", + "bbox": [ + 212, + 390, + 782, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Learnable Query $\\mathbf{Q}_l$ : Every object is associated with a unique learnable token that facilitates feature probing, and subsequent prediction of class labels and 3D bounding boxes. In line with prior research [20, 40], we conceptualize the learnable query as an anchor point, initiating it with the positional prior of the user prompt. The initialization process unfolds over three phases:", + "bbox": [ + 212, + 465, + 784, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "1). 2D Center Position Computation with User Prompt. A user might provide a bounding box $box = (x_1, y_1, x_2, y_2)$ or a singular point $point = (x, y)$ . Depending on the input type, the 2D center position $(c_x, c_y)$ is determined as:", + "bbox": [ + 212, + 541, + 782, + 587 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left(c _ {x}, c _ {y}\\right) = \\left\\{ \\begin{array}{l l} \\left(\\frac {x _ {1} + x _ {2}}{2}, \\frac {y _ {1} + y _ {2}}{2}\\right) & \\text {i f b o u n d i n g b o x} \\\\ (x, y) & \\text {i f p o i n t} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 595, + 784, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "2). 2D Center Transformation into 3D Coordinate. Given a user prompt on the BEV map, the 2D space naturally omits height data. We address this by designating $c_{z} = 0.5$ , the median value of height, resulting in a 3D center labeled as center $= (c_{x},c_{y},c_{z})$ . Alternatively, for a user prompt within 2D multi-view images, we transform the center coordinates $(c_{x},c_{y})$ into a set of 3D points along the depth axis, using Eqs. (1) and (2). The median of the ray is then selected as the 3D center, represented as center $= (c_x',c_y',c_z')$ . The 3D center serves as a reference point for bounding box prediction and we prioritize the center from BEV if the user provides prompts in both modalities.", + "bbox": [ + 212, + 643, + 784, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3). 3D Position Encoding Extraction from Both Modalities. The 3D center is projected onto both the camera image and BEV plane, yielding their 3D position encodings $\\mathbf{Q}_{im} \\in \\mathbb{R}^{1 \\times C}$ and $\\mathbf{Q}_{pc} \\in \\mathbb{R}^{1 \\times C}$ , respectively. The learnable query is initialized by combining information from both modalities: $\\mathbf{Q}_l = \\mathbf{Q}_{im} + \\mathbf{Q}_{pc}$ .", + "bbox": [ + 212, + 779, + 784, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Interactive 3D Object Detection with Prompts", + "bbox": [ + 418, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/63ec824553a1b3a32dc38049373b5970cda6f01738b8f24593fdf1941ceb0991.jpg", + "image_caption": [ + "Fig. 3: Illustration of Global-to-Local Enhancement. Our queries interact with unmodified multi-modal features and its object's local features respectively in the global and local phases. Self-Attention is confined within queries of identical objects in both stages." + ], + "image_footnote": [], + "bbox": [ + 289, + 146, + 712, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt Query $\\mathbf{Q}_p$ : It is designed to embed user guidance by merging position encoding and context embedding: $\\mathbf{Q}_p = \\mathbf{Q}_{pe} + \\mathbf{Q}_{ce}$ . This synthesis is constructed for both box and point prompt formats: For the box format, the box position is delineated using its center point. From our 3D position encoding, the position encoding for box center, $\\mathbf{Q}_{pe} \\in \\mathbb{R}^{1 \\times C}$ , is derived. The context embedding, $\\mathbf{Q}_{ce} \\in \\mathbb{R}^{1 \\times C}$ , is secured via average pooling of the feature region bounded by the prompt box. For the point format, the same technique determines the position encoding, $\\mathbf{Q}_{pe} \\in \\mathbb{R}^{1 \\times C}$ . The feature at the indicated prompt point gives the context embedding, $\\mathbf{Q}_{ce} \\in \\mathbb{R}^{1 \\times C}$ .", + "bbox": [ + 212, + 353, + 787, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Global-to-Local Enhancement", + "text_level": 1, + "bbox": [ + 214, + 508, + 506, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given multi-modal features $\\mathbf{F}_{pc}$ and $\\mathbf{F}_{im}$ with their respective 3D position embeddings $\\mathbf{P}_{pc}$ and $\\mathbf{P}_{im}$ , along with object queries $\\mathbf{Q} \\in \\mathbb{R}^{N \\times C}$ , where $N$ represents the cumulative number of queries from $\\mathbf{Q}_l$ and $\\mathbf{Q}_p$ across all objects, we employ a Transformer-based decoder to amalgamate these features for prediction.", + "bbox": [ + 212, + 532, + 787, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Self-Attention: The formulation for the decoder's self-attention mechanism is:", + "bbox": [ + 214, + 593, + 785, + 608 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {i} ^ {\\prime} = \\operatorname {s o f t m a x} \\left(\\mathbf {M} _ {s a} + \\mathbf {Q} _ {i} \\mathbf {Q} _ {i} ^ {T}\\right) \\mathbf {Q} _ {i} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 621, + 785, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{Q}_i$ corresponds to the queries from the $i$ -th layer, and $\\mathbf{Q}_0 = \\mathbf{Q}$ . The attention mask $\\mathbf{M}_{sa}$ ensures interaction is confined within queries of identical objects:", + "bbox": [ + 212, + 643, + 787, + 689 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {s a} \\left(q _ {1}, q _ {2}\\right) = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} q _ {1}, q _ {2} \\text {b e l o n g t o t h e s a m e o b j e c t ,} \\\\ - \\infty , & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 290, + 696, + 785, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cross-Attention: For cross-attention in the decoder, we express it as:", + "bbox": [ + 214, + 744, + 730, + 760 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {i + 1} = \\operatorname {s o f t m a x} (\\mathbf {M} _ {c a} + \\mathbf {Q} _ {i} ^ {\\prime} \\mathbf {K} ^ {T}) \\mathbf {V} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 768, + 785, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{K}$ , $\\mathbf{V}$ come from multi-modal features, $\\mathbf{M}_{ca}$ is attention mask for cross-attention. Influenced by traditional two-stage detectors [32] and recent DETR-centric techniques [6, 52], we adopt a global-to-local enhancement strategy.", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "R. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This approach adjusts $\\mathbf{K}$ and $\\mathbf{V}$ dynamically to augment our model's local perceptiveness, as illustrated in Fig. 3.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For the initial $Z$ layers among a total of $L$ decoder layers, queries access unmodified multi-modal features:", + "bbox": [ + 212, + 176, + 784, + 205 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K} = \\left[ f _ {k} \\left(\\mathbf {F} _ {p c} + \\mathbf {P} _ {p c}\\right): f _ {k} \\left(\\mathbf {F} _ {i m} + \\mathbf {P} _ {i m}\\right) \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 218, + 784, + 234 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} = \\left[ f _ {v} \\left(\\mathbf {F} _ {p c}\\right): f _ {v} \\left(\\mathbf {F} _ {i m}\\right) \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 237, + 784, + 253 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{K},\\mathbf{V}\\in \\mathbb{R}^{(H_{pc}W_{pc} + H_{im}W_{im}T)\\times C}$ are original multi-modal features under transformation $f_{k}(\\cdot),f_{v}(\\cdot)$ and concatenation $[\\cdot :\\cdot ]$ . $\\mathbf{M}_{ca} = 0$ during the global cross-attention phase.", + "bbox": [ + 212, + 263, + 782, + 309 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the succeeding $L - Z$ layers, local region-of-interest (RoI) features are extracted by RoI Align [10] via either the user-provided prompt box or the predicted 2D bounding box based on prompt point:", + "bbox": [ + 212, + 310, + 782, + 356 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K} = \\left[ f _ {k} \\left(\\operatorname {R o I} \\left(\\mathbf {F} _ {p c} + \\mathbf {P} _ {p c}\\right)\\right): f _ {k} \\left(\\operatorname {R o I} \\left(\\mathbf {F} _ {i m} + \\mathbf {P} _ {i m}\\right)\\right) \\right], \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 367, + 784, + 383 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} = \\left[ f _ {v} \\left(\\operatorname {R o I} \\left(\\mathbf {F} _ {p c}\\right)\\right): f _ {v} \\left(\\operatorname {R o I} \\left(\\mathbf {F} _ {i m}\\right)\\right) \\right], \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 386, + 784, + 402 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{K},\\mathbf{V}\\in \\mathbb{R}^{(H_oW_o + H_oW_o)\\times C}$ now signify local multi-modal features. $(H_{o},W_{o})$ is the output size of RoI Align. With $\\mathbf{M}_{ca}$ defined similarly to $\\mathbf{M}_{sa}$ , it ensures each query is exclusively attentive to its object's local features. Following the decoding process, we derive the object query $\\mathbf{Q}_L$ from the last decoding layer, and subsequently produce the initial prediction via two feed-forward networks (FFNs):", + "bbox": [ + 212, + 412, + 784, + 503 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nc _ {L} ^ {0} = \\phi_ {c} \\left(\\mathbf {Q} _ {L}\\right), \\quad b _ {L} ^ {0} = \\phi_ {b} \\left(\\mathbf {Q} _ {L}\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 503, + 784, + 520 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Here, $c_{L}^{0}$ and $b_{L}^{0}$ represent the class and 3D bounding box predictions, respectively. While $\\phi_{c}$ and $\\phi_{b}$ are FFNs designated for classification and regression tasks. At this point, the first phase of \"prompt in 2D, detect in 3D\" is completed.", + "bbox": [ + 212, + 525, + 782, + 585 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.6 3D Attribute Encoding", + "text_level": 1, + "bbox": [ + 214, + 607, + 454, + 623 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For those inaccurate results of the initial predictions, we endow our model with iteratively self-correction capability under limited human interaction. Specifically, we decompose the 3D bounding box into (gravity center, width, length, height and yaw angle) five attributes, along with category information, and assume users to adjust a few wrong attributes, e.g. \"dragging the center point of the bounding box\". Given the prediction $b_{L}^{r}$ and $c_{L}^{r}$ of the $r$ -th iteration, the adjusted 3D bounding box and category are denoted as $b_{L}^{r'}$ and $c_{L}^{r'}$ . We also introduce two binary attribute refinement records $p_a^r \\in \\mathbb{R}^5$ and $p_c^r \\in \\mathbb{R}^1$ to explicitly aid the model in distinguishing between corrected attributes and those yet to be predicted.", + "bbox": [ + 212, + 632, + 784, + 782 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Finally, the 3D attribute query $Q_{3D}$ for each object is formed by the combination of both attribute and record embedding:", + "bbox": [ + 212, + 782, + 782, + 813 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {3 D} = \\mathrm {M L P} _ {a t t} \\left(b _ {L} ^ {r ^ {\\prime}}\\right) + \\mathrm {M L P} _ {r e c 1} \\left(p _ {a} ^ {r}\\right) + \\mathrm {M L P} _ {c a t} \\left(c _ {L} ^ {r ^ {\\prime}}\\right) + \\mathrm {M L P} _ {r e c 2} \\left(p _ {c} ^ {r}\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 251, + 824, + 784, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Interactive 3D Object Detection with Prompts", + "bbox": [ + 418, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "And $\\mathbf{Q}_{3D}$ is inserted into the corresponding query sequence of the respective object, collaborating with the original prompt tokens to jointly drive the model to provide more precise predictions. Note that in the following decoder process, $\\mathbf{M}_{sa}$ and $\\mathbf{M}_{ca}$ are also dynamically adapted according to Eq. (6).", + "bbox": [ + 212, + 146, + 787, + 208 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3.7 Label Assignment and Losses", + "text_level": 1, + "bbox": [ + 214, + 228, + 503, + 242 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Due to the intrinsic one-to-one correspondence between our learnable queries and prompt objects, our proposed interactive 3D object detection task does not require label assignment like Hungarian Algorithm. To this end, following the methodologies in [21,45], we employ the Focal loss [19] for the classification task and the L1 loss for regression. To bolster the training of individual decoder layers, a deep supervision strategy is implemented. The 3D loss function amalgamates these predictions and is represented as:", + "bbox": [ + 212, + 251, + 787, + 358 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {3 D} = \\sum_ {r = 0} ^ {M} \\sum_ {i = 1} ^ {L} \\lambda_ {c l s} \\mathcal {L} _ {c l s} \\left(c _ {i} ^ {r}, c\\right) + \\lambda_ {r e g} \\mathcal {L} _ {r e g} \\left(b _ {i} ^ {r}, b\\right) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 367, + 785, + 407 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Here, $c$ and $b$ are the ground truths, $c_i^r$ and $b_i^r$ denote the predictions of the $i$ -th decoder layer and $r$ -th iteration, which are generated by the shared MLP as Eq. (12). $\\lambda_{cls}$ and $\\lambda_{reg}$ are weights for balancing loss items. Besides, we also utilize L1 loss to supervise the 2D bounding box prediction on the image or BEV modality, which is produced by another MLP and serves as the local region for point prompts to perform the global-to-local strategy. The overarching loss function is the combination of both 3D and 2D losses.", + "bbox": [ + 212, + 416, + 787, + 521 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 544, + 375, + 561 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experiment Settings", + "text_level": 1, + "bbox": [ + 214, + 574, + 433, + 590 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset splits: The nuScenes dataset [2] is adopted in our work for training and evaluation. This dataset encompasses 1000 driving scenes, capturing data via six cameras, one LiDAR, and five radars. Following the official partition, we employ 750 scenes for training and 150 for validation. Contrasting prior work, our interactive 3D object detection approach leverages user prompts to guide the model's predictions. Similar to common interactive algorithms [4, 47], we reconstruct the evaluation benchmark by simulating user prompts and refinements from Ground Truth. Notably, since the nuScenes test set's Ground Truth remains inaccessible, we exclude it from our evaluations.", + "bbox": [ + 212, + 598, + 787, + 733 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training Protocol: Throughout the training process, we leverage Ground Truth to simulate practical user prompts and iterative refinements. Detailed training procedures for Prompt in 2D, Detect in 3D and Detect in 3D, Refine in 3D phases are provided in the supplementary materials.", + "bbox": [ + 212, + 734, + 787, + 794 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Protocol: We evaluate our method through human annotation comparisons and simulated experiments, under two distinct annotation scenarios: (1) Annotation from Raw Data and (2) Annotation from SOTA", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "R. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/35ceb3558ff8e0652ef322f94a2fc84f831b12d3537aab7983025d5c56c5f1ad.jpg", + "table_caption": [ + "Table 1: Comparisons of the average and standard deviation time cost required for annotating an image extracted from the nuScenes validation set, with four annotation strategies, i.e. manual-only, Ours, SOTA model CMT with manual correction (CMT+M) and Ours (CMT+Ours) respectively." + ], + "table_footnote": [], + "table_body": "
Time cost (s)Annotation from Raw DataAnnotation from SOTA Model
Manual-onlyOursCMT+MCMT + Ours
nuScenes-val124±2855±1740±1224±7
", + "bbox": [ + 290, + 213, + 715, + 258 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Model. We describe the simulated experimental settings as follows: in scenario (1), we employ 2D projections of 3D Ground Truth with perturbations to simulate user prompts, while in scenario (2), SOTA 3D detection models provide initial prediction results to initiate our algorithm. The attributes from Ground Truth are utilized to mimic subsequent user refinements in both scenarios. We explore various orders of adjustments to bounding box attributes and assess our model both without and with corrected category priors, denoted as Ours and Ours*, respectively. * indicates the pre-calibration of category data. Unless specified, we prioritize using 2D bounding boxes as prompts and following the adjustment sequence of [gravity center, yaw angle, height, width, length]. Experimental results for alternative prompt formats (e.g., point prompts) and refinement orders are detailed in the supplementary materials. Furthermore, in addition to conventional closed-set evaluation, we also conduct extended experiments on challenging open-set scenarios. The nuScenes official metrics, i.e. nuScenes Detection Score (NDS), mean Average Precision (mAP) are adopted to evaluate the approaches.", + "bbox": [ + 212, + 297, + 787, + 541 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Human Annotation Comparisons", + "text_level": 1, + "bbox": [ + 214, + 560, + 535, + 575 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To showcase the efficacy of our proposed method, we perform annotation comparisons across two scenarios employing four strategies: Manual-only, Ours, $\\mathrm{CMT + M}$ (manual correction), and $\\mathrm{CMT + Ours}$ . Ten experienced annotators are recruited to annotate the same set of twenty randomly selected nuScenes validation scenes using each of these strategies. Statistical analysis in Table 1 demonstrates that benefiting from the interactive design, our method achieves a speedup of approximately $56\\%$ and $40\\%$ compared to manual-only and $\\mathrm{CMT + M}$ , respectively.", + "bbox": [ + 212, + 584, + 789, + 691 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Closed-Set 3D Object Detection", + "text_level": 1, + "bbox": [ + 214, + 710, + 529, + 726 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Annotation from SOTA Model Evaluation. Results from our experiments on the nuScenes validation set are tabulated in Table 2. In practical applications, annotators often streamline their efforts by initiating annotations from the 3D predictions generated by SOTA models, such as CMT. Our model seamlessly integrates into this workflow, harnessing its robust capabilities to effectively aid in annotation tasks. Specifically, given the 3D predictions from CMT, Table 2 illustrates the interactive refinement results, without employing category", + "bbox": [ + 212, + 734, + 787, + 842 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Interactive 3D Object Detection with Prompts", + "bbox": [ + 418, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a87d97c9c96d4556d3039d1c7489bfb96c644b90f89081e3c2571fc3948d880f.jpg", + "table_caption": [ + "Table 2: Comparisons of annotation performance on nuScenes val set. \"M\" means manual correction. (Ri) denotes the $i$ -th iteration refinement. \"C\" and \"L\" are camera and LiDAR modalities respectively." + ], + "table_footnote": [], + "table_body": "
MethodModalityNDS↑mAP↑
Traditional 3D Detectors
UVTR [16]CL70.265.4
BEVFusion [23]CL71.468.5
MetaBEV [8]CL71.568.0
SpaseFusion [42]CL72.870.4
CMT [45]CL72.970.3
Annotation from SOTA Model
CMT+M (R1)CL81.381.6
CMT+M (R2)CL83.881.6
CMT+M (R3)CL84.381.6
CMT+M (R4)CL85.381.6
CMT+M (R5)CL86.381.6
CMT+Ours (R1)CL83.4 (+2.1)86.3 (+4.7)
CMT+Ours (R2)CL86.3 (+2.5)86.3 (+4.7)
CMT+Ours (R3)CL87.4 (+3.1)87.0 (+5.4)
CMT+Ours (R4)CL89.3 (+4.0)89.1 (+7.5)
CMT+Ours (R5)CL90.4 (+4.1)89.4 (+7.8)
Annotation from Raw Data
OursCL75.076.2
Ours (R1)CL84.4 (+9.4)88.0 (+11.8)
Ours (R2)CL87.2 (+12.2)88.0 (+11.8)
Ours (R3)CL88.2 (+13.2)88.4 (+12.2)
Ours (R4)CL89.5 (+14.5)89.5 (+13.3)
Ours (R5)CL90.5 (+15.5)89.6 (+13.4)
", + "bbox": [ + 334, + 198, + 674, + 481 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "correction priors. As a comparative reference, we perform identical manual attribute adjustments on the initial results predicted by CMT. It is evident that, in five consecutive iterations, our approach consistently outperforms $\\mathrm{CMT + M}$ by margins of $2.1\\%$ , $2.5\\%$ , $3.1\\%$ , $4.0\\%$ , and $4.1\\%$ , respectively, in terms of NDS. Additionally, CMT maintains a constant mAP after the initial refinement of the center position, lacking self-correction for object categories. In contrast, our approach iteratively refines predictions with minimal human intervention, resulting in sustained performance improvements. In addition, we also provide comparisons with corrected category priors, denoted as $\\mathrm{CMT + M^*}$ and $\\mathrm{CMT + Ours^*}$ in Fig. 4. Through the rectification of object categories, our model exhibits a significantly enhanced self-correction capability, outperforming $\\mathrm{CMT + M^*}$ by a considerable margin.", + "bbox": [ + 212, + 526, + 787, + 708 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Annotation from Raw Data Evaluation. In cases involving missed detections by existing 3D detectors, our method is expected to have the capability of annotating from raw data. As shown in Table 2, by leveraging a box as the prompt format, our model registers an NDS of $75.0\\%$ and an mAP of $76.2\\%$ . Subsequent iterations of refinements contribute to continued performance gains, and the final model attains an NDS score of $90.5\\%$ and mAP of $89.6\\%$ . Some visualization results are presented in Fig. 5.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "R. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/68ea3277f58d0d6a10100bbb409211c0c9f0c2cf903e37f4c8e9556d6405e2c7.jpg", + "image_caption": [ + "Fig. 4: Performance comparisons of annotation from SOTA model. R1-R5 follows a fixed refinement order. * indicates the utilization of category correction." + ], + "image_footnote": [], + "bbox": [ + 361, + 161, + 622, + 297 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/35a5b82eaecad245bbb3667d9c6dfaa4949333068227b5cebf2a2c298909694f.jpg", + "table_caption": [ + "Table 3: Categories and quantities of open-set objects extracted from nuScenes validation set." + ], + "table_footnote": [], + "table_body": "
CategoryNumCategoryNum
personal_mobility15stroller103
animal32bicycle_rac246
pushableifiable2925debris514
ambulance30police64
", + "bbox": [ + 220, + 402, + 460, + 470 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/ed337ed98536e8a10e87d21a0db382980ea3dac63e35957a953b2ae09c9f9558.jpg", + "table_caption": [ + "Table 5: Ablation study of prompt queries. \"PE\", \"CE\" are position encoding and context encoding. \"It.\" denotes the iteration." + ], + "table_footnote": [], + "table_body": "
It.QtQPPEQpCEQ3DbL'NDS↑mAP↑
0✓73.575.5
0✓✓74.575.7
0✓✓✓75.076.2
1✓✓✓✓83.285.9
1✓✓✓✓84.488.0
", + "bbox": [ + 256, + 551, + 437, + 637 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/86b1e606b054c8564ce98fae3b569903de82dae16714759a23e839afba640608.jpg", + "table_caption": [ + "Table 4: Open-set evaluation." + ], + "table_footnote": [], + "table_body": "
MethodModalityNDS↑mAP↑
Annotation from Raw Data
Ours*CL89.391.9
Ours* (R1)CL94.199.5
Ours* (R2)CL96.699.5
Ours* (R3)CL97.399.5
Ours* (R4)CL98.199.5
Ours* (R5)CL99.299.5
", + "bbox": [ + 540, + 375, + 733, + 470 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/9e61dc1fb6cccaf6374dd583b11d23d9876054ae93a404c308775b1463516cb1.jpg", + "table_caption": [ + "Table 6: Ablation study of Global-to-Local strategy (G2L). \"L\", \"G\" denotes utilizing local RoI or global multi-modal features to perform Cross-Attention." + ], + "table_footnote": [], + "table_body": "
It.AttentionNDS↑mAP↑
0G73.775.1
0L74.375.2
0G2L75.076.2
", + "bbox": [ + 560, + 565, + 714, + 625 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Open-Set 3D Object Detection", + "text_level": 1, + "bbox": [ + 214, + 664, + 516, + 680 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To delve deeper into the 3D object annotation capabilities guided by 2D prompts, we extract eight novel categories from the nuScenes dataset, which are not employed in the training process. Table 3 outlines their distribution on the validation set. Since our primary goal is to validate the model's generalization ability based on user-provided prompts, as opposed to venturing into open-vocabulary recognition, we ignore the class prediction and assume that the class information is provided manually together with 2D prompts, denoted as Ours*. The experimental results in Table 4 demonstrate that our model possesses the capability to detect 3D objects based on 2D priors. Furthermore, with manual interaction for refinement, the detection performance improves progressively.", + "bbox": [ + 212, + 688, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Interactive 3D Object Detection with Prompts", + "bbox": [ + 419, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/58d504a583bb570b5a5c5bfd8e5dc20090a6c802ad9351fcc59b8e6cc5d17c9e.jpg", + "image_caption": [ + "Fig. 5: Visualizations of the 3D detection results from our method are presented. Ground Truth and our predictions are exhibited in green and blue, respectively. The outcomes highlighted by red circles illustrate that, with a single round of manual adjustment focusing solely on the gravity center, our model automatically rectified other attributes such as yaw angles, thus yielding more accurate 3D results." + ], + "image_footnote": [], + "bbox": [ + 246, + 143, + 754, + 286 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 396, + 388, + 411 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ablation Study of Object Queries. The pivotal role of prompt tokens is evident from the results in Table 5. When solely introducing prompt queries with position encoding, the model attains a nuanced understanding of the object's spatial location, registering an NDS of $74.5\\%$ —a significant boost of $1.0\\%$ . Moreover, integrating context embedding, which encapsulates appearance attributes of objects, further escalates performance, culminating in a gain of $0.5\\%$ . Besides, we also delve into an analysis of the effectiveness of $Q_{3D}$ during the first iteration refinement stage. Clear performance improvements are evident with the explicit inclusion of the current attribute refinement record $p_a^r$ .", + "bbox": [ + 212, + 421, + 787, + 556 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ablation Study of Global-to-Local Strategy. The proposed global-to-local strategy's efficacy is substantiated by the results displayed in Table 6. Contrasted with models relying exclusively on either global or local feature extraction, our hybrid approach ensures a comprehensive global perspective while meticulously capturing local nuances, leading to superior performance outcomes.", + "bbox": [ + 212, + 558, + 787, + 633 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 657, + 369, + 672 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper introduces a new method for 3D object detection that tackles the daunting task of 3D object annotation. In consideration of the necessity for either creating annotations from raw data or refining existing imperfect annotations within practice annotation scenarios, two dedicated strategies are proposed. \"Prompt in 2D, detect in 3D\" principle leverages straightforward 2D interactions, like clicks or boxes, to streamline the transition between 2D images and 3D object annotations. \"Detect in 3D, refine in 3D\" strategy further endows our model with self-correction capability. Evaluations on the nuScenes dataset validate our approach's superiority. Beyond a mere advancement, our work can serve as a cornerstone for future 3D object detection endeavors.", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "R. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China (NO. 62322608), in part by the Fundamental Research Funds for the Central Universities under Grant 22lgqb25, and in part by the Open Project Program of the Key Laboratory of Artificial Intelligence for Perception and Understanding, Liaoning Province (AIPU, No. 20230003).", + "bbox": [ + 212, + 180, + 787, + 256 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 282, + 321, + 299 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1090-1099 (2022)", + "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11621-11631 (2020)", + "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European Conference on Computer Vision. pp. 213-229. Springer (2020)", + "4. Chen, X., Zhao, Z., Zhang, Y., Duan, M., Qi, D., Zhao, H.: Focalclick: Towards practical interactive image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1300-1309 (2022)", + "5. Chen, X., Ma, H., Wan, J., Li, B., Xia, T.: Multi-view 3d object detection network for autonomous driving. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1907-1915 (2017)", + "6. Chen, Y., Chen, Q., Sun, P., Chen, S., Wang, J., Cheng, J.: Enhancing your trained detrs with box refinement. arXiv preprint arXiv:2307.11828 (2023)", + "7. Choi, D., Cho, W., Kim, K., Choo, J.: idet3d: Towards efficient interactive object detection for lidar point clouds. arXiv preprint arXiv:2312.15449 (2023)", + "8. Ge, C., Chen, J., Xie, E., Wang, Z., Hong, L., Lu, H., Li, Z., Luo, P.: Metabev: Solving sensor failures for bev detection and map segmentation. arXiv preprint arXiv:2304.09801 (2023)", + "9. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition. pp. 3354-3361. IEEE (2012)", + "0. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2961-2969 (2017)", + "1. Huang, J., Huang, G.: Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054 (2022)", + "2. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021)", + "13. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + ], + "bbox": [ + 220, + 316, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Interactive 3D Object Detection with Prompts", + "bbox": [ + 419, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12697-12705 (2019)", + "15. Lee, C., Park, S., Song, H., Ryu, J., Kim, S., Kim, H., Pereira, S., Yoo, D.: Interactive multi-class tiny-object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14136-14145 (2022)", + "16. Li, Y., Chen, Y., Qi, X., Li, Z., Sun, J., Jia, J.: Unifying voxel-based representation with transformer for 3d object detection. Advances in Neural Information Processing Systems 35, 18442-18455 (2022)", + "17. Li, Z., Wang, F., Wang, N.: Lidar r-cnn: An efficient and universal 3d object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7546-7555 (2021)", + "18. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European Conference on Computer Vision. pp. 1-18. Springer (2022)", + "19. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2980-2988 (2017)", + "20. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329 (2022)", + "21. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. In: European Conference on Computer Vision. pp. 531-548. Springer (2022)", + "22. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X., Sun, J.: Petrov2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256 (2022)", + "23. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2774-2781. IEEE (2023)", + "24. Lu, Y., Ma, X., Yang, L., Zhang, T., Liu, Y., Chu, Q., Yan, J., Ouyang, W.: Geometry uncertainty projection network for monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3111-3121 (2021)", + "25. Maninis, K.K., Caelles, S., Pont-Tuset, J., Van Gool, L.: Deep extreme cut: From extreme points to object segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 616-625 (2018)", + "26. Meng, D., Chen, X., Fan, Z., Zeng, G., Li, H., Yuan, Y., Sun, L., Wang, J.: Conditional detr for fast training convergence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3651-3660 (2021)", + "27. Meng, Q., Wang, W., Zhou, T., Shen, J., Jia, Y., Van Gool, L.: Towards a weakly supervised framework for 3d point cloud object detection and annotation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(8), 4454-4468 (2021)", + "28. Papadopoulos, D.P., Uijlings, J.R., Keller, F., Ferrari, V.: Extreme clicking for efficient object annotation. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4930-4939 (2017)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "R. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Papadopoulos, D.P., Uijlings, J.R., Keller, F., Ferrari, V.: Training object class detectors with click supervision. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6374-6383 (2017)", + "30. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 194-210. Springer (2020)", + "31. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017)", + "32. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems 28 (2015)", + "33. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-779 (2019)", + "34. Shin, K., Kwon, Y.P., Tomizuka, M.: Roarnet: A robust 3d object detection based on region approximation refinement. In: 2019 IEEE Intelligent Vehicles Symposium (IV). pp. 2510-2515. IEEE (2019)", + "35. Sofiuk, K., Petrov, I.A., Konushin, A.: Reviving iterative training with mask guidance for interactive segmentation. In: 2022 IEEE International Conference on Image Processing (ICIP). pp. 3141-3145. IEEE (2022)", + "36. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2446-2454 (2020)", + "37. Wang, B., Wu, V., Wu, B., Keutzer, K.: Latte: accelerating lidar point cloud annotation via sensor fusion, one-click annotation, and tracking. In: 2019 IEEE Intelligent Transportation Systems Conference (ITSC). pp. 265-272. IEEE (2019)", + "38. Wang, K., Zhuang, J., Li, G., Fang, C., Cheng, L., Lin, L., Zhou, F.: De-biased teacher: Rethinking iou matching for semi-supervised object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2573-2580 (2023)", + "39. Wang, T., Zhu, X., Pang, J., Lin, D.: Fcos3d: Fully convolutional one-stage monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 913-922 (2021)", + "40. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2567-2575 (2022)", + "41. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning. pp. 180–191. PMLR (2022)", + "42. Xie, Y., Xu, C., Rakotosaona, M.J., Rim, P., Tombari, F., Keutzer, K., Tomizuka, M., Zhan, W.: Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection. arXiv preprint arXiv:2304.14340 (2023)", + "43. Xu, M., Zhang, Z., Hu, H., Wang, J., Wang, L., Wei, F., Bai, X., Liu, Z.: End-to-end semi-supervised object detection with soft teacher. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3060-3069 (2021)", + "44. Xu, N., Price, B., Cohen, S., Yang, J., Huang, T.S.: Deep interactive object selection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 373-381 (2016)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Interactive 3D Object Detection with Prompts", + "bbox": [ + 418, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "45. Yan, J., Liu, Y., Sun, J., Jia, F., Li, S., Wang, T., Zhang, X.: Cross modal transformer via coordinates encoding for 3d object detection. arXiv preprint arXiv:2301.01283 (2023)", + "46. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)", + "47. Yang, J., Zeng, A., Li, F., Liu, S., Zhang, R., Zhang, L.: Neural interactive keypoint detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15122-15132 (2023)", + "48. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3dssd: Point-based 3d single stage object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11040-11048 (2020)", + "49. Yao, A., Gall, J., Leistner, C., Van Gool, L.: Interactive object detection. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition. pp. 3242-3249. IEEE (2012)", + "50. Zhang, J., Lin, X., Zhang, W., Wang, K., Tan, X., Han, J., Ding, E., Wang, J., Li, G.: Semi-detr: Semi-supervised object detection with detection transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23809-23818 (2023)", + "51. Zhang, R., Qiu, H., Wang, T., Guo, Z., Xu, X., Qiao, Y., Gao, P., Li, H.: Monodetr: depth-guided transformer for monocular 3d object detection. arXiv preprint arXiv:2203.13310 (2022)", + "52. Zhao, J., Sun, L., Li, Q.: Recursivelydet: End-to-end region-based recursive object detection. arXiv preprint arXiv:2307.13619 (2023)", + "53. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4490-4499 (2018)", + "54. Zhou, Y., Zhu, H., Liu, Q., Chang, S., Guo, M.: Monoatt: Online monocular 3d object detection with adaptive token transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17493-17503 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 550 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "R. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Interactive 3D Object Detection with Prompts/20a1b205-7505-40a6-b499-29fffaad19e8_model.json b/2024/Interactive 3D Object Detection with Prompts/20a1b205-7505-40a6-b499-29fffaad19e8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1fac4e627b4131b293ae876203cf14a59f6bda1e --- /dev/null +++ b/2024/Interactive 3D Object Detection with Prompts/20a1b205-7505-40a6-b499-29fffaad19e8_model.json @@ -0,0 +1,2623 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.235, + 0.142, + 0.77, + 0.166 + ], + "angle": 0, + "content": "Interactive 3D Object Detection with Prompts" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.189, + 0.782, + 0.222 + ], + "angle": 0, + "content": "Ruifei Zhang\\(^{1,2\\star}\\), Xiangru Lin\\(^{4\\star}\\), Wei Zhang\\(^{4}\\), Jincheng Lu\\(^{4}\\), Xuekuan Wang\\(^{4}\\), Xiao Tan\\(^{4}\\), Yingying Li\\(^{4}\\), Errui Ding\\(^{4}\\), Jingdong Wang\\(^{4}\\), and Guanbin Li\\(^{3,5\\star\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.232, + 0.697, + 0.247 + ], + "angle": 0, + "content": "1 The Chinese University of Hong Kong, Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.247, + 0.705, + 0.26 + ], + "angle": 0, + "content": "2 Shenzhen Research Institute of Big Data, Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.351, + 0.26, + 0.653, + 0.274 + ], + "angle": 0, + "content": "3 Sun Yat-sen University, Guangzhou, China" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.274, + 0.74, + 0.289 + ], + "angle": 0, + "content": "\\(^{4}\\) Department of Computer Vision Technology (VIS), Baidu Inc., China" + }, + { + "type": "text", + "bbox": [ + 0.414, + 0.289, + 0.59, + 0.303 + ], + "angle": 0, + "content": "Peng Cheng Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.303, + 0.721, + 0.317 + ], + "angle": 0, + "content": "ruifeizhang@link.cuhk.edu.cn, liguanbin@mail.sysu.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.317, + 0.723, + 0.331 + ], + "angle": 0, + "content": "{linxiangru,zhangwei99,lujincheng01,wangxuekuan,tanxiao01," + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.331, + 0.68, + 0.344 + ], + "angle": 0, + "content": "liyingying05,dingerrui,wangjingdong}@baidu.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.379, + 0.744, + 0.629 + ], + "angle": 0, + "content": "Abstract. The evolution of 3D object detection hinges not only on advanced models but also on effective and efficient annotation strategies. Despite this progress, the labor-intensive nature of 3D object annotation remains a bottleneck, hindering further development in the field. This paper introduces a novel approach, incorporated with \"prompt in 2D, detect in 3D\" and \"detect in 3D, refine in 3D\" strategies, to 3D object annotation: multi-modal interactive 3D object detection. Firstly, by allowing users to engage with simpler 2D interaction prompts (e.g., clicks or boxes on a camera image or a bird's eye view), we bridge the complexity gap between 2D and 3D spaces, reimagining the annotation workflow. Besides, Our framework also supports flexible iterative refinement to the initial 3D annotations, further assisting annotators in achieving satisfying results. Evaluation on the nuScenes dataset demonstrates the effectiveness of our method. And thanks to the prompt-driven and interactive designs, our approach also exhibits outstanding performance in open-set scenarios. This work not only offers a potential solution to the 3D object annotation problem but also paves the way for further innovations in the 3D object detection community." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.642, + 0.697, + 0.656 + ], + "angle": 0, + "content": "Keywords: Interactive 3D Object Detection \\(\\cdot\\) Prompt Learning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.681, + 0.377, + 0.697 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.788, + 0.805 + ], + "angle": 0, + "content": "3D object detection is pivotal for autonomous driving, gaining significant traction in recent years. Current research often derives 3D object information from sources such as monocular [24, 39, 51, 54] or multi-view camera images [11, 12, 18, 21], point cloud [17, 46, 53], and multi-modal sensors [1, 5, 45]. This has led to remarkable progress in detection performance. A major catalyst behind these advancements is the availability of high-quality 3D datasets like KITTI [9]," + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.812, + 0.365, + 0.825 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.826, + 0.384, + 0.84 + ], + "angle": 0, + "content": "\\*\\* Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.216, + 0.812, + 0.384, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "R. Zhang et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.146, + 0.615, + 0.156 + ], + "angle": 0, + "content": "Traditional Way of 3D Detection Data Annotation" + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.158, + 0.728, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.221, + 0.621, + 0.231 + ], + "angle": 0, + "content": "Our proposed Way of 3D Detection Data Annotation" + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.232, + 0.727, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.398 + ], + "angle": 0, + "content": "Fig. 1: We introduce an interactive 3D object detection concept. Rather than the traditional method where a model produces 3D object detection boxes followed by user refinement in 3D space, our strategy is \"prompt in 2D, detect in 3D, refine in 3D\". This method streamlines the annotation process, reducing the need for user refinements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.521 + ], + "angle": 0, + "content": "nuScenes [2], and Waymo [36]. These datasets fuel the progression of data-driven algorithms in 3D object detection. However, the intricate nature of 3D object annotation often necessitates significant human resources and incurs high labor costs. On average, annotating an hour of driving data takes hundreds of hours [27]. This annotation challenge potentially restricts the scale and diversity of datasets, possibly hindering further progression in this field." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.627 + ], + "angle": 0, + "content": "Addressing this annotation challenge is the central theme of our paper. While weakly-supervised [25, 27-29] and semi-supervised [38, 43, 50] object detection methods exist, they typically suffer from reduced performance, especially in practical scenarios. This does not substantially address the overarching issue of the data-hungry nature of the field. With the recent emergence of SAM [13], both industry and academia are now recognizing the potential of interactive annotation approaches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Our work aims to integrate this concept into 3D object detection, thus simplifying the annotation process, as presented in Fig. 1. Specifically, we observe that within the traditional process of annotating 3D detection data, annotators tend to concentrate on addressing two challenges stemming from imperfect pretrained 3D detectors: 1) supplementing missing closed-set objects and incorporating additional required open-set ones; 2) rectifying existing erroneous annotations. Both scenarios necessitate annotators to operate within the 3D point cloud space, which is extremely time-consuming and labor-intensive." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To circumvent this, we create an interactive 3D object detection system enhanced by two strategies: prompt in \\(2D\\), detect in \\(3D\\), and detect in \\(3D\\), refine in \\(3D\\). The first principle enables users to interact with a 2D camera image or bird's eye view (BEV) of the point cloud, prompting the model to generate corresponding 3D bounding boxes, instead of encountering the 3D point cloud" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.42, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Interactive 3D Object Detection with Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.22 + ], + "angle": 0, + "content": "directly. Furthermore, benefiting from the prompt-driven annotation paradigm, our method inherently possesses the capability to annotate open-set objects. The latter ensures annotators refine the initial 3D bounding box at the fastest pace with minimal effort, which significantly streamlines operations in the 3D space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.433 + ], + "angle": 0, + "content": "In light of these principles, we present our multi-modal interactive 3D object detection task and algorithm. Firstly, it employs prompts in the form of clicks or boxes on a 2D camera or BEV view, guiding the model to generate the relevant 3D bounding box. Our framework is reminiscent of the DETR-series detectors [3, 20, 26, 40] but comes with distinctive features. We integrate user inputs as prompt tokens and design a learnable component that investigates multi-modal features of indicated objects, informed by the prompts. Furthermore, we incorporate a global-to-local strategy to supply fine-grained object features, enhancing detection accuracy. Secondly, we empower our model with iterative refinement capability by incorporating the predicted 3D bounding box encoding as a new prompt into the query sequence. For those inaccurate 3D predictions, our framework allows annotators to adjust a few wrong attributes, e.g. dragging the center point position or adjusting the orientation angle of the 3D bounding box, and then perform self-correction on other characters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.785, + 0.54 + ], + "angle": 0, + "content": "To rigorously evaluate our approach, we devise two evaluation configurations for practical annotation scenarios. The first, regarding Annotation from Raw Data, leverages user 2D prompts and produces corresponding 3D predictions. Besides, our framework is also compatible with starting from the results of existing State-of-the-Art (SOTA) 3D detection models, termed as Annotation from SOTA Model and iteratively refining them. Experimental results on the nuScenes dataset highlight the effectiveness of our method in various scenarios." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.542, + 0.641, + 0.556 + ], + "angle": 0, + "content": "In summary, this paper has the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.569, + 0.783, + 0.626 + ], + "angle": 0, + "content": "- We introduce a new concept of multi-modal interactive 3D object detection, proposing a substantial solution to prevailing annotation challenges. Our work is expected to provide critical insights for the 3D object detection community across academic and industrial domains." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.63, + 0.783, + 0.671 + ], + "angle": 0, + "content": "- We unveil a prompt-driven interactive 3D object detection methodology with self-correction capability, leveraging it to establish an interactive system tailored for practical annotation scenarios." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.676, + 0.783, + 0.702 + ], + "angle": 0, + "content": "- Comprehensive experiments in both annotation scenarios and evaluation settings attest to our method's efficacy." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.569, + 0.783, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.732, + 0.388, + 0.748 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.767, + 0.434, + 0.782 + ], + "angle": 0, + "content": "2.1 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.783, + 0.84 + ], + "angle": 0, + "content": "3D object detection techniques have been developed across a range of input modalities. Our approach emphasizes the advantages of interactive inputs, setting it apart from traditional methods:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "R. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.236 + ], + "angle": 0, + "content": "Camera-only. Monocular 3D object detection techniques [24, 39, 51, 54] enjoy advantages such as low cost and ease of deployment. However, it also faces the inherent challenge of limited depth information. Multi-view methods utilize data from multiple camera perspectives [11, 12, 18, 21, 22, 30, 41], which enhances the model's ability to perceive the 3D environment, thereby improving detection accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.787, + 0.298 + ], + "angle": 0, + "content": "LiDAR-only. Benefiting from the robustness and high precision, numerous studies utilize LiDAR data for 3D object detection. Point-based methods [17, 33,48] like PointNet [31] directly engage with raw LiDAR data, while grid-based strategies structure the data into 3D voxels [46,53] or feature pillars [14]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.787, + 0.389 + ], + "angle": 0, + "content": "Multi-modality. Capitalizing on the strengths of camera images and point clouds, multi-modal 3D object detection has seen significant advancements [1, 5, 45]. Early methods combine camera images and point clouds at the input or result stages [5, 34]. While, current models, like TransFusion [1] and CMT [45], emphasize more intricate feature-level fusion. Uniquely, our method embeds interactive capacities, ensuring richer, user-tailored annotations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.41, + 0.62, + 0.426 + ], + "angle": 0, + "content": "2.2 Interactive Object Detection/Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.787, + 0.463 + ], + "angle": 0, + "content": "Interactive methods have historically aimed to reduce the annotation workload while maintaining, or even enhancing, model performance:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.787, + 0.569 + ], + "angle": 0, + "content": "Interactive Object Detection. Early attempts, like the incremental learning strategy [49], require user corrections to update detectors. More recent models like C3Det [15] have substantially eased the annotation process. Tools such as LATTE [37] and iDet3D [7] further simplify the 3D object annotation process. Our framework, however, brings in the versatility of prompt formats and modalities, offering users a broadened interaction spectrum coupled with robust detection capabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.787, + 0.675 + ], + "angle": 0, + "content": "Interactive Object Segmentation. Interactive segmentation has matured significantly over the years [4,35,44]. Models like SAM [13] have empowered users with tools like clicks, boxes, and referring expressions to guide the segmentation process. Diverging from interactive segmentation confined to 2D spaces, our work addresses the challenge of integrating 2D interactive prompts with 3D object detection, mitigating the spatial discrepancy between 2D annotations and 3D environments." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.696, + 0.565, + 0.712 + ], + "angle": 0, + "content": "2.3 Weakly-supervised Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The notion of harnessing weak signals, like clicks [29] or extreme-points [25, 28], for supervising detectors has been a topic of intrigue. Some methods [27] have extended this idea to 3D object detection, using center-click based strategies. These approaches, while commendable for their balance between performance and annotation overhead, sometimes fail to harness the full potential of the detectors. Our methodology, on the other hand, incorporates strong 3D bounding box supervisions in an interactive setting, streamlining the data acquisition process without compromising on detection performance." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.419, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Interactive 3D Object Detection with Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.146, + 0.773, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.289, + 0.788, + 0.456 + ], + "angle": 0, + "content": "Fig. 2: An overview of our proposed interactive 3D object detection system/method. Taking multi-modal data with user prompts as input, we explicitly add 3D information \\(\\mathbf{P}_{im},\\mathbf{P}_{pc}\\) into features \\(\\mathbf{F}_{im},\\mathbf{F}_{pc}\\) by 3D Position Encoding. Object queries consist of a learnable token \\(\\mathbf{Q}_l\\) and prompt ones \\(\\mathbf{Q}_p\\), which are generated by Prompt-Guided Query generation. We also introduce a Global-to-Local enhancement strategy to dynamically adjust features integrated with queries from global representations to local RoI details. Besides, our framework allows users to refine any of 3D attributes directly, and encode the refined 3D bounding box \\(b_{L}^{r^{\\prime}}\\) and category information \\(c_{L}^{r^{\\prime}}\\), along with attribute refinement record \\(p_a^r\\) and \\(p_c^r\\), into a new prompt token \\(\\mathbf{Q}_{3D}\\) by 3D Attribute Encoding. \\(g,w,l,h,\\theta\\) denotes the gravity center, width, length, height and yaw angle respectively. This iterative refinement propels our model to automatically rectify other attributes, ultimately delivering satisfactory results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.483, + 0.381, + 0.5 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.513, + 0.421, + 0.528 + ], + "angle": 0, + "content": "3.1 Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.691 + ], + "angle": 0, + "content": "We present an interactive multi-modal 3D object detection algorithm. Given a multi-view set of camera images and associated point cloud data, the objective is to predict both the class labels and the 3D bounding boxes, driven by user prompts either in the form of a point or a box. Further, our model also supports iteratively self-correction with limited human assistance. To supervise the iteration process, the Ground Truth \\((c, b)\\) is employed. Here, \\(c\\) signifies the category label, and \\(b\\) represents the 3D attributes. We adopt a deep supervision strategy for our multi-layer Transformer decoder, which guides the predictions across each decoder layer of each iteration, denoted as \\(\\{(c_i^r, b_i^r), i = 1, 2, \\dots, L, r = 0, 1, \\dots, M\\}\\), with \\(L\\) and \\(M\\) being the total number of decoder layers and iterative refinements." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.71, + 0.341, + 0.724 + ], + "angle": 0, + "content": "3.2 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The architecture underpinning our interactive 3D object detection method is visualized in Fig. 2, which could be roughly decomposed into two stages, i.e. \"prompt in \\(2D\\), detect in \\(3D\\) and \"detect in \\(3D\\), refine in \\(3D\\)\". In the first phase, two separate backbones are utilized to extract multi-view image features \\(\\mathbf{F}_{im} = \\{\\mathbf{F}_{im}^{t}, t = 1,2,\\dots,T\\}\\) and point cloud features \\(\\mathbf{F}_{pc}\\) respectively. To enhance these features with 3D-awareness, following PETR [21] and CMT [45], we introduce 3D position encoding. Specifically, for the multi-view images, this encoding is" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "R. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "represented as \\(\\mathbf{P}_{im} = \\{\\mathbf{P}_{im}^t, t = 1,2,\\dots,T\\}\\), and for the point cloud, it is denoted as \\(\\mathbf{P}_{pc}\\). For object queries, denoted as \\(\\mathbf{Q}\\), we utilize a unique blending of a learnable token, \\(\\mathbf{Q}_l\\), and prompt-specific tokens, \\(\\mathbf{Q}_p\\), for every detected object. The prompt queries, \\(\\mathbf{Q}_p\\), play a pivotal role in guiding the learnable token, \\(\\mathbf{Q}_l\\), by leveraging the insights from the multi-modal features. Furthermore, to enhance the detection accuracy across various object scales, we incorporate a global-to-local strategy within the decoder. This strategy ensures that the object queries are continually enriched by granular local features, enabling finer detections. Ultimately, the learnable token predicts the object's class and its 3D bounding box \\((c_L^0,b_L^0)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.788, + 0.42 + ], + "angle": 0, + "content": "Subsequently, the model enters an interactive refinement stage. This stage reuses the multi-modal features but updates the object queries \\(\\mathbf{Q}\\) by inserting the encoding of the refined 3D prediction from the current iteration. Specifically, in the \\(r\\)-th iteration, we enable annotators to fine-tune the current prediction \\((c_{L}^{r}, b_{L}^{r})\\) and encode the refined result \\((c_{L}^{r'}, b_{L}^{r'})\\) and corresponding attribute refinement record \\((p_{c}^{r}, p_{a}^{r})\\) into a new prompt \\(\\mathbf{Q}_{3D}\\). \\(\\mathbf{Q}_{3D}\\) and \\(\\mathbf{Q}_p\\) collectively drive \\(\\mathbf{Q}_l\\) to generate more accurate predictions in the next iteration. Note that we omit the iteration superscript of \\(\\mathbf{Q}_{3D}\\) and \\(\\mathbf{Q}\\) for simplicity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.44, + 0.445, + 0.456 + ], + "angle": 0, + "content": "3.3 3D Position Encoding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.788, + 0.585 + ], + "angle": 0, + "content": "Multi-view Camera Images. The core idea of embedding 3D position for multi-view image features includes two steps: 1) modeling each pixel coordinate as a series of points along a ray in the camera frustum space, and 2) projecting camera frustum coordinates into 3D space. Specifically, for a given camera image feature, denoted as \\(\\mathbf{F}_{im}^{t}\\in \\mathbb{R}^{W_{im}\\times H_{im}\\times C}\\), each pixel can be visualized as tracing a ray within the camera frustum space. This perspective allows us to sample \\(D\\) discrete points along the depth dimension. This process is mathematically captured by:" + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.596, + 0.785, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {c f} (u, v, j) = \\left(u \\times d _ {j}, v \\times d _ {j}, d _ {j}, 1\\right) ^ {T}, j = 1, 2, \\dots , D \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.787, + 0.652 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{P}_{cf}\\) indicates the camera's frustum coordinates. Pixel coordinates in the image are denoted by \\((u,v)\\), and \\(d_{j}\\) represents the \\(j\\)-th depth value." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.653, + 0.787, + 0.681 + ], + "angle": 0, + "content": "To relate these coordinates to a 3D context, we employ an inverse 3D projection method:" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.682, + 0.785, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {3 d} ^ {t} (u, v, j) = \\mathbf {K} _ {t} ^ {3 D} \\mathbf {K} _ {t} ^ {- 1} \\mathbf {P} _ {c f} (u, v, j) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.796 + ], + "angle": 0, + "content": "where \\(\\mathbf{K}_t^{3D} \\in \\mathbb{R}^{4 \\times 4}\\) represents the transformation matrix that links 3D space with the \\(t\\)-th camera's coordinate space. \\(\\mathbf{K}_t \\in \\mathbb{R}^{4 \\times 4}\\) is the intrinsic matrix of \\(t\\)-th camera. To ensure consistent value scales, we normalize the 3D coordinates to the range [0,1]. The final 3D position encoding for the \\(t\\)-th camera image, represented as \\(\\mathbf{P}_{im}^t \\in \\mathbb{R}^{W_{im} \\times H_{im} \\times C}\\), is derived via a multi-layer perceptron (MLP): \\(\\mathbf{P}_{im}^t = \\mathrm{MLP}_{im}(\\mathbf{P}_{3d}^t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Point Cloud. In contrast to the approach for camera image features, where points are assembled along the depth axis, the LiDAR point set necessitates sampling along the height axis. Specifically, given a BEV feature map, denoted as" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.419, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Interactive 3D Object Detection with Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.193 + ], + "angle": 0, + "content": "\\(\\mathbf{F}_{pc} \\in \\mathbb{R}^{W_{pc} \\times H_{pc} \\times C}\\), we can compute the corresponding 3D position information, represented as \\(\\mathbf{P}_{pc} \\in \\mathbb{R}^{W_{pc} \\times H_{pc} \\times C}\\), using the following steps: First, we determine the 3D position based on the feature map's coordinates:" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.201, + 0.786, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {3 d} (u, v, j) = \\left(u \\times u _ {s}, v \\times v _ {s}, h _ {j}, 1\\right) ^ {T}, j = 1, 2, \\dots , H \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.227, + 0.785, + 0.348 + ], + "angle": 0, + "content": "Here, \\((u,v)\\) specifies the coordinate within the point cloud feature map. The variables \\((u_{s},v_{s})\\) represent the dimensions of the feature grid. \\(H\\) is the number of sampled points along the height axis and \\(h_j\\) is the \\(j\\)-th height value. Next, we utilize a multi-layer perceptron (MLP) to compute the 3D position encoding from the derived 3D position: \\(\\mathbf{P}_{pc} = \\mathrm{MLP}_{pc}(\\mathbf{P}_{3d})\\). Note that as in CMT, only one point is sampled along the height axis, thus simplifying the process into standard 2D position encoding. Consequently, in our methodology, we have adopted a sinusoidal embedding to cater to this behavior." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.368, + 0.542, + 0.383 + ], + "angle": 0, + "content": "3.4 Prompt-guided Query Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.784, + 0.464 + ], + "angle": 0, + "content": "Our approach provides an avenue for user interaction, distinct from traditional DETR-based detectors that often rely on multiple queries to discern image objects. By merging learnable tokens, \\(\\mathbf{Q}_l\\), with user-directed prompt tokens, \\(\\mathbf{Q}_p\\), we achieve object detection that is not only efficient but also offers a new angle on object localization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.785, + 0.542 + ], + "angle": 0, + "content": "Learnable Query \\(\\mathbf{Q}_l\\): Every object is associated with a unique learnable token that facilitates feature probing, and subsequent prediction of class labels and 3D bounding boxes. In line with prior research [20, 40], we conceptualize the learnable query as an anchor point, initiating it with the positional prior of the user prompt. The initialization process unfolds over three phases:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.784, + 0.588 + ], + "angle": 0, + "content": "1). 2D Center Position Computation with User Prompt. A user might provide a bounding box \\(box = (x_1, y_1, x_2, y_2)\\) or a singular point \\(point = (x, y)\\). Depending on the input type, the 2D center position \\((c_x, c_y)\\) is determined as:" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.597, + 0.785, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\left(c _ {x}, c _ {y}\\right) = \\left\\{ \\begin{array}{l l} \\left(\\frac {x _ {1} + x _ {2}}{2}, \\frac {y _ {1} + y _ {2}}{2}\\right) & \\text {i f b o u n d i n g b o x} \\\\ (x, y) & \\text {i f p o i n t} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.785, + 0.779 + ], + "angle": 0, + "content": "2). 2D Center Transformation into 3D Coordinate. Given a user prompt on the BEV map, the 2D space naturally omits height data. We address this by designating \\( c_{z} = 0.5 \\), the median value of height, resulting in a 3D center labeled as center \\( = (c_{x},c_{y},c_{z}) \\). Alternatively, for a user prompt within 2D multi-view images, we transform the center coordinates \\( (c_{x},c_{y}) \\) into a set of 3D points along the depth axis, using Eqs. (1) and (2). The median of the ray is then selected as the 3D center, represented as center \\( = (c_x',c_y',c_z') \\). The 3D center serves as a reference point for bounding box prediction and we prioritize the center from BEV if the user provides prompts in both modalities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.843 + ], + "angle": 0, + "content": "3). 3D Position Encoding Extraction from Both Modalities. The 3D center is projected onto both the camera image and BEV plane, yielding their 3D position encodings \\(\\mathbf{Q}_{im} \\in \\mathbb{R}^{1 \\times C}\\) and \\(\\mathbf{Q}_{pc} \\in \\mathbb{R}^{1 \\times C}\\), respectively. The learnable query is initialized by combining information from both modalities: \\(\\mathbf{Q}_l = \\mathbf{Q}_{im} + \\mathbf{Q}_{pc}\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "R. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.147, + 0.713, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.27, + 0.788, + 0.327 + ], + "angle": 0, + "content": "Fig. 3: Illustration of Global-to-Local Enhancement. Our queries interact with unmodified multi-modal features and its object's local features respectively in the global and local phases. Self-Attention is confined within queries of identical objects in both stages." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.789, + 0.492 + ], + "angle": 0, + "content": "Prompt Query \\(\\mathbf{Q}_p\\): It is designed to embed user guidance by merging position encoding and context embedding: \\(\\mathbf{Q}_p = \\mathbf{Q}_{pe} + \\mathbf{Q}_{ce}\\). This synthesis is constructed for both box and point prompt formats: For the box format, the box position is delineated using its center point. From our 3D position encoding, the position encoding for box center, \\(\\mathbf{Q}_{pe} \\in \\mathbb{R}^{1 \\times C}\\), is derived. The context embedding, \\(\\mathbf{Q}_{ce} \\in \\mathbb{R}^{1 \\times C}\\), is secured via average pooling of the feature region bounded by the prompt box. For the point format, the same technique determines the position encoding, \\(\\mathbf{Q}_{pe} \\in \\mathbb{R}^{1 \\times C}\\). The feature at the indicated prompt point gives the context embedding, \\(\\mathbf{Q}_{ce} \\in \\mathbb{R}^{1 \\times C}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.51, + 0.507, + 0.525 + ], + "angle": 0, + "content": "3.5 Global-to-Local Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.788, + 0.594 + ], + "angle": 0, + "content": "Given multi-modal features \\(\\mathbf{F}_{pc}\\) and \\(\\mathbf{F}_{im}\\) with their respective 3D position embeddings \\(\\mathbf{P}_{pc}\\) and \\(\\mathbf{P}_{im}\\), along with object queries \\(\\mathbf{Q} \\in \\mathbb{R}^{N \\times C}\\), where \\(N\\) represents the cumulative number of queries from \\(\\mathbf{Q}_l\\) and \\(\\mathbf{Q}_p\\) across all objects, we employ a Transformer-based decoder to amalgamate these features for prediction." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.594, + 0.787, + 0.609 + ], + "angle": 0, + "content": "Self-Attention: The formulation for the decoder's self-attention mechanism is:" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.622, + 0.787, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {i} ^ {\\prime} = \\operatorname {s o f t m a x} \\left(\\mathbf {M} _ {s a} + \\mathbf {Q} _ {i} \\mathbf {Q} _ {i} ^ {T}\\right) \\mathbf {Q} _ {i} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.69 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q}_i\\) corresponds to the queries from the \\(i\\)-th layer, and \\(\\mathbf{Q}_0 = \\mathbf{Q}\\). The attention mask \\(\\mathbf{M}_{sa}\\) ensures interaction is confined within queries of identical objects:" + }, + { + "type": "equation", + "bbox": [ + 0.291, + 0.698, + 0.787, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {s a} \\left(q _ {1}, q _ {2}\\right) = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} q _ {1}, q _ {2} \\text {b e l o n g t o t h e s a m e o b j e c t ,} \\\\ - \\infty , & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.746, + 0.731, + 0.761 + ], + "angle": 0, + "content": "Cross-Attention: For cross-attention in the decoder, we express it as:" + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.769, + 0.787, + 0.788 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {i + 1} = \\operatorname {s o f t m a x} (\\mathbf {M} _ {c a} + \\mathbf {Q} _ {i} ^ {\\prime} \\mathbf {K} ^ {T}) \\mathbf {V} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\(\\mathbf{K}\\), \\(\\mathbf{V}\\) come from multi-modal features, \\(\\mathbf{M}_{ca}\\) is attention mask for cross-attention. Influenced by traditional two-stage detectors [32] and recent DETR-centric techniques [6, 52], we adopt a global-to-local enhancement strategy." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.419, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Interactive 3D Object Detection with Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "This approach adjusts \\(\\mathbf{K}\\) and \\(\\mathbf{V}\\) dynamically to augment our model's local perceptiveness, as illustrated in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.207 + ], + "angle": 0, + "content": "For the initial \\( Z \\) layers among a total of \\( L \\) decoder layers, queries access unmodified multi-modal features:" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.219, + 0.785, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K} = \\left[ f _ {k} \\left(\\mathbf {F} _ {p c} + \\mathbf {P} _ {p c}\\right): f _ {k} \\left(\\mathbf {F} _ {i m} + \\mathbf {P} _ {i m}\\right) \\right], \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.238, + 0.785, + 0.254 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} = \\left[ f _ {v} \\left(\\mathbf {F} _ {p c}\\right): f _ {v} \\left(\\mathbf {F} _ {i m}\\right) \\right], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.264, + 0.784, + 0.31 + ], + "angle": 0, + "content": "where \\(\\mathbf{K},\\mathbf{V}\\in \\mathbb{R}^{(H_{pc}W_{pc} + H_{im}W_{im}T)\\times C}\\) are original multi-modal features under transformation \\(f_{k}(\\cdot),f_{v}(\\cdot)\\) and concatenation \\([\\cdot :\\cdot ]\\) . \\(\\mathbf{M}_{ca} = 0\\) during the global cross-attention phase." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.784, + 0.357 + ], + "angle": 0, + "content": "In the succeeding \\( L - Z \\) layers, local region-of-interest (RoI) features are extracted by RoI Align [10] via either the user-provided prompt box or the predicted 2D bounding box based on prompt point:" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.368, + 0.785, + 0.384 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K} = \\left[ f _ {k} \\left(\\operatorname {R o I} \\left(\\mathbf {F} _ {p c} + \\mathbf {P} _ {p c}\\right)\\right): f _ {k} \\left(\\operatorname {R o I} \\left(\\mathbf {F} _ {i m} + \\mathbf {P} _ {i m}\\right)\\right) \\right], \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.387, + 0.785, + 0.403 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} = \\left[ f _ {v} \\left(\\operatorname {R o I} \\left(\\mathbf {F} _ {p c}\\right)\\right): f _ {v} \\left(\\operatorname {R o I} \\left(\\mathbf {F} _ {i m}\\right)\\right) \\right], \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.785, + 0.505 + ], + "angle": 0, + "content": "where \\(\\mathbf{K},\\mathbf{V}\\in \\mathbb{R}^{(H_oW_o + H_oW_o)\\times C}\\) now signify local multi-modal features. \\((H_{o},W_{o})\\) is the output size of RoI Align. With \\(\\mathbf{M}_{ca}\\) defined similarly to \\(\\mathbf{M}_{sa}\\), it ensures each query is exclusively attentive to its object's local features. Following the decoding process, we derive the object query \\(\\mathbf{Q}_L\\) from the last decoding layer, and subsequently produce the initial prediction via two feed-forward networks (FFNs):" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.505, + 0.785, + 0.521 + ], + "angle": 0, + "content": "\\[\nc _ {L} ^ {0} = \\phi_ {c} \\left(\\mathbf {Q} _ {L}\\right), \\quad b _ {L} ^ {0} = \\phi_ {b} \\left(\\mathbf {Q} _ {L}\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.784, + 0.587 + ], + "angle": 0, + "content": "Here, \\( c_{L}^{0} \\) and \\( b_{L}^{0} \\) represent the class and 3D bounding box predictions, respectively. While \\( \\phi_{c} \\) and \\( \\phi_{b} \\) are FFNs designated for classification and regression tasks. At this point, the first phase of \"prompt in 2D, detect in 3D\" is completed." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.608, + 0.455, + 0.624 + ], + "angle": 0, + "content": "3.6 3D Attribute Encoding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.785, + 0.783 + ], + "angle": 0, + "content": "For those inaccurate results of the initial predictions, we endow our model with iteratively self-correction capability under limited human interaction. Specifically, we decompose the 3D bounding box into (gravity center, width, length, height and yaw angle) five attributes, along with category information, and assume users to adjust a few wrong attributes, e.g. \"dragging the center point of the bounding box\". Given the prediction \\( b_{L}^{r} \\) and \\( c_{L}^{r} \\) of the \\( r \\)-th iteration, the adjusted 3D bounding box and category are denoted as \\( b_{L}^{r'} \\) and \\( c_{L}^{r'} \\). We also introduce two binary attribute refinement records \\( p_a^r \\in \\mathbb{R}^5 \\) and \\( p_c^r \\in \\mathbb{R}^1 \\) to explicitly aid the model in distinguishing between corrected attributes and those yet to be predicted." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.784, + 0.784, + 0.814 + ], + "angle": 0, + "content": "Finally, the 3D attribute query \\( Q_{3D} \\) for each object is formed by the combination of both attribute and record embedding:" + }, + { + "type": "equation", + "bbox": [ + 0.252, + 0.825, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {3 D} = \\mathrm {M L P} _ {a t t} \\left(b _ {L} ^ {r ^ {\\prime}}\\right) + \\mathrm {M L P} _ {r e c 1} \\left(p _ {a} ^ {r}\\right) + \\mathrm {M L P} _ {c a t} \\left(c _ {L} ^ {r ^ {\\prime}}\\right) + \\mathrm {M L P} _ {r e c 2} \\left(p _ {c} ^ {r}\\right) \\tag {13}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "R. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.209 + ], + "angle": 0, + "content": "And \\(\\mathbf{Q}_{3D}\\) is inserted into the corresponding query sequence of the respective object, collaborating with the original prompt tokens to jointly drive the model to provide more precise predictions. Note that in the following decoder process, \\(\\mathbf{M}_{sa}\\) and \\(\\mathbf{M}_{ca}\\) are also dynamically adapted according to Eq. (6)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.229, + 0.504, + 0.243 + ], + "angle": 0, + "content": "3.7 Label Assignment and Losses" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.252, + 0.788, + 0.359 + ], + "angle": 0, + "content": "Due to the intrinsic one-to-one correspondence between our learnable queries and prompt objects, our proposed interactive 3D object detection task does not require label assignment like Hungarian Algorithm. To this end, following the methodologies in [21,45], we employ the Focal loss [19] for the classification task and the L1 loss for regression. To bolster the training of individual decoder layers, a deep supervision strategy is implemented. The 3D loss function amalgamates these predictions and is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.368, + 0.786, + 0.409 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {3 D} = \\sum_ {r = 0} ^ {M} \\sum_ {i = 1} ^ {L} \\lambda_ {c l s} \\mathcal {L} _ {c l s} \\left(c _ {i} ^ {r}, c\\right) + \\lambda_ {r e g} \\mathcal {L} _ {r e g} \\left(b _ {i} ^ {r}, b\\right) \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Here, \\( c \\) and \\( b \\) are the ground truths, \\( c_i^r \\) and \\( b_i^r \\) denote the predictions of the \\( i \\)-th decoder layer and \\( r \\)-th iteration, which are generated by the shared MLP as Eq. (12). \\( \\lambda_{cls} \\) and \\( \\lambda_{reg} \\) are weights for balancing loss items. Besides, we also utilize L1 loss to supervise the 2D bounding box prediction on the image or BEV modality, which is produced by another MLP and serves as the local region for point prompts to perform the global-to-local strategy. The overarching loss function is the combination of both 3D and 2D losses." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.545, + 0.377, + 0.563 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.575, + 0.434, + 0.591 + ], + "angle": 0, + "content": "4.1 Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Dataset splits: The nuScenes dataset [2] is adopted in our work for training and evaluation. This dataset encompasses 1000 driving scenes, capturing data via six cameras, one LiDAR, and five radars. Following the official partition, we employ 750 scenes for training and 150 for validation. Contrasting prior work, our interactive 3D object detection approach leverages user prompts to guide the model's predictions. Similar to common interactive algorithms [4, 47], we reconstruct the evaluation benchmark by simulating user prompts and refinements from Ground Truth. Notably, since the nuScenes test set's Ground Truth remains inaccessible, we exclude it from our evaluations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Training Protocol: Throughout the training process, we leverage Ground Truth to simulate practical user prompts and iterative refinements. Detailed training procedures for Prompt in 2D, Detect in 3D and Detect in 3D, Refine in 3D phases are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Evaluation Protocol: We evaluate our method through human annotation comparisons and simulated experiments, under two distinct annotation scenarios: (1) Annotation from Raw Data and (2) Annotation from SOTA" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.419, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Interactive 3D Object Detection with Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 1: Comparisons of the average and standard deviation time cost required for annotating an image extracted from the nuScenes validation set, with four annotation strategies, i.e. manual-only, Ours, SOTA model CMT with manual correction (CMT+M) and Ours (CMT+Ours) respectively." + }, + { + "type": "table", + "bbox": [ + 0.292, + 0.214, + 0.717, + 0.259 + ], + "angle": 0, + "content": "
Time cost (s)Annotation from Raw DataAnnotation from SOTA Model
Manual-onlyOursCMT+MCMT + Ours
nuScenes-val124±2855±1740±1224±7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.789, + 0.542 + ], + "angle": 0, + "content": "Model. We describe the simulated experimental settings as follows: in scenario (1), we employ 2D projections of 3D Ground Truth with perturbations to simulate user prompts, while in scenario (2), SOTA 3D detection models provide initial prediction results to initiate our algorithm. The attributes from Ground Truth are utilized to mimic subsequent user refinements in both scenarios. We explore various orders of adjustments to bounding box attributes and assess our model both without and with corrected category priors, denoted as Ours and Ours*, respectively. * indicates the pre-calibration of category data. Unless specified, we prioritize using 2D bounding boxes as prompts and following the adjustment sequence of [gravity center, yaw angle, height, width, length]. Experimental results for alternative prompt formats (e.g., point prompts) and refinement orders are detailed in the supplementary materials. Furthermore, in addition to conventional closed-set evaluation, we also conduct extended experiments on challenging open-set scenarios. The nuScenes official metrics, i.e. nuScenes Detection Score (NDS), mean Average Precision (mAP) are adopted to evaluate the approaches." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.561, + 0.536, + 0.577 + ], + "angle": 0, + "content": "4.2 Human Annotation Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.79, + 0.693 + ], + "angle": 0, + "content": "To showcase the efficacy of our proposed method, we perform annotation comparisons across two scenarios employing four strategies: Manual-only, Ours, \\(\\mathrm{CMT + M}\\) (manual correction), and \\(\\mathrm{CMT + Ours}\\). Ten experienced annotators are recruited to annotate the same set of twenty randomly selected nuScenes validation scenes using each of these strategies. Statistical analysis in Table 1 demonstrates that benefiting from the interactive design, our method achieves a speedup of approximately \\(56\\%\\) and \\(40\\%\\) compared to manual-only and \\(\\mathrm{CMT + M}\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.711, + 0.53, + 0.727 + ], + "angle": 0, + "content": "4.3 Closed-Set 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Annotation from SOTA Model Evaluation. Results from our experiments on the nuScenes validation set are tabulated in Table 2. In practical applications, annotators often streamline their efforts by initiating annotations from the 3D predictions generated by SOTA models, such as CMT. Our model seamlessly integrates into this workflow, harnessing its robust capabilities to effectively aid in annotation tasks. Specifically, given the 3D predictions from CMT, Table 2 illustrates the interactive refinement results, without employing category" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "R. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Comparisons of annotation performance on nuScenes val set. \"M\" means manual correction. (Ri) denotes the \\( i \\)-th iteration refinement. \"C\" and \"L\" are camera and LiDAR modalities respectively." + }, + { + "type": "table", + "bbox": [ + 0.336, + 0.199, + 0.675, + 0.482 + ], + "angle": 0, + "content": "
MethodModalityNDS↑mAP↑
Traditional 3D Detectors
UVTR [16]CL70.265.4
BEVFusion [23]CL71.468.5
MetaBEV [8]CL71.568.0
SpaseFusion [42]CL72.870.4
CMT [45]CL72.970.3
Annotation from SOTA Model
CMT+M (R1)CL81.381.6
CMT+M (R2)CL83.881.6
CMT+M (R3)CL84.381.6
CMT+M (R4)CL85.381.6
CMT+M (R5)CL86.381.6
CMT+Ours (R1)CL83.4 (+2.1)86.3 (+4.7)
CMT+Ours (R2)CL86.3 (+2.5)86.3 (+4.7)
CMT+Ours (R3)CL87.4 (+3.1)87.0 (+5.4)
CMT+Ours (R4)CL89.3 (+4.0)89.1 (+7.5)
CMT+Ours (R5)CL90.4 (+4.1)89.4 (+7.8)
Annotation from Raw Data
OursCL75.076.2
Ours (R1)CL84.4 (+9.4)88.0 (+11.8)
Ours (R2)CL87.2 (+12.2)88.0 (+11.8)
Ours (R3)CL88.2 (+13.2)88.4 (+12.2)
Ours (R4)CL89.5 (+14.5)89.5 (+13.3)
Ours (R5)CL90.5 (+15.5)89.6 (+13.4)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.788, + 0.709 + ], + "angle": 0, + "content": "correction priors. As a comparative reference, we perform identical manual attribute adjustments on the initial results predicted by CMT. It is evident that, in five consecutive iterations, our approach consistently outperforms \\(\\mathrm{CMT + M}\\) by margins of \\(2.1\\%\\), \\(2.5\\%\\), \\(3.1\\%\\), \\(4.0\\%\\), and \\(4.1\\%\\), respectively, in terms of NDS. Additionally, CMT maintains a constant mAP after the initial refinement of the center position, lacking self-correction for object categories. In contrast, our approach iteratively refines predictions with minimal human intervention, resulting in sustained performance improvements. In addition, we also provide comparisons with corrected category priors, denoted as \\(\\mathrm{CMT + M^*}\\) and \\(\\mathrm{CMT + Ours^*}\\) in Fig. 4. Through the rectification of object categories, our model exhibits a significantly enhanced self-correction capability, outperforming \\(\\mathrm{CMT + M^*}\\) by a considerable margin." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Annotation from Raw Data Evaluation. In cases involving missed detections by existing 3D detectors, our method is expected to have the capability of annotating from raw data. As shown in Table 2, by leveraging a box as the prompt format, our model registers an NDS of \\(75.0\\%\\) and an mAP of \\(76.2\\%\\). Subsequent iterations of refinements contribute to continued performance gains, and the final model attains an NDS score of \\(90.5\\%\\) and mAP of \\(89.6\\%\\). Some visualization results are presented in Fig. 5." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.42, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Interactive 3D Object Detection with Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.162, + 0.624, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.306, + 0.785, + 0.334 + ], + "angle": 0, + "content": "Fig. 4: Performance comparisons of annotation from SOTA model. R1-R5 follows a fixed refinement order. * indicates the utilization of category correction." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.349, + 0.475, + 0.389 + ], + "angle": 0, + "content": "Table 3: Categories and quantities of open-set objects extracted from nuScenes validation set." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.403, + 0.462, + 0.471 + ], + "angle": 0, + "content": "
CategoryNumCategoryNum
personal_mobility15stroller103
animal32bicycle_rac246
pushableifiable2925debris514
ambulance30police64
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.484, + 0.476, + 0.539 + ], + "angle": 0, + "content": "Table 5: Ablation study of prompt queries. \"PE\", \"CE\" are position encoding and context encoding. \"It.\" denotes the iteration." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.552, + 0.438, + 0.638 + ], + "angle": 0, + "content": "
It.QtQPPEQpCEQ3DbL'NDS↑mAP↑
0✓73.575.5
0✓✓74.575.7
0✓✓✓75.076.2
1✓✓✓✓83.285.9
1✓✓✓✓84.488.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.533, + 0.349, + 0.737, + 0.363 + ], + "angle": 0, + "content": "Table 4: Open-set evaluation." + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.375, + 0.734, + 0.471 + ], + "angle": 0, + "content": "
MethodModalityNDS↑mAP↑
Annotation from Raw Data
Ours*CL89.391.9
Ours* (R1)CL94.199.5
Ours* (R2)CL96.699.5
Ours* (R3)CL97.399.5
Ours* (R4)CL98.199.5
Ours* (R5)CL99.299.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.505, + 0.484, + 0.766, + 0.552 + ], + "angle": 0, + "content": "Table 6: Ablation study of Global-to-Local strategy (G2L). \"L\", \"G\" denotes utilizing local RoI or global multi-modal features to perform Cross-Attention." + }, + { + "type": "table", + "bbox": [ + 0.561, + 0.566, + 0.715, + 0.626 + ], + "angle": 0, + "content": "
It.AttentionNDS↑mAP↑
0G73.775.1
0L74.375.2
0G2L75.076.2
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.665, + 0.517, + 0.681 + ], + "angle": 0, + "content": "4.4 Open-Set 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.841 + ], + "angle": 0, + "content": "To delve deeper into the 3D object annotation capabilities guided by 2D prompts, we extract eight novel categories from the nuScenes dataset, which are not employed in the training process. Table 3 outlines their distribution on the validation set. Since our primary goal is to validate the model's generalization ability based on user-provided prompts, as opposed to venturing into open-vocabulary recognition, we ignore the class prediction and assume that the class information is provided manually together with 2D prompts, denoted as Ours*. The experimental results in Table 4 demonstrate that our model possesses the capability to detect 3D objects based on 2D priors. Furthermore, with manual interaction for refinement, the detection performance improves progressively." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "R. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.145, + 0.756, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Fig. 5: Visualizations of the 3D detection results from our method are presented. Ground Truth and our predictions are exhibited in green and blue, respectively. The outcomes highlighted by red circles illustrate that, with a single round of manual adjustment focusing solely on the gravity center, our model automatically rectified other attributes such as yaw angles, thus yielding more accurate 3D results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.397, + 0.39, + 0.412 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.788, + 0.558 + ], + "angle": 0, + "content": "Ablation Study of Object Queries. The pivotal role of prompt tokens is evident from the results in Table 5. When solely introducing prompt queries with position encoding, the model attains a nuanced understanding of the object's spatial location, registering an NDS of \\(74.5\\%\\) —a significant boost of \\(1.0\\%\\). Moreover, integrating context embedding, which encapsulates appearance attributes of objects, further escalates performance, culminating in a gain of \\(0.5\\%\\). Besides, we also delve into an analysis of the effectiveness of \\(Q_{3D}\\) during the first iteration refinement stage. Clear performance improvements are evident with the explicit inclusion of the current attribute refinement record \\(p_a^r\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.559, + 0.788, + 0.634 + ], + "angle": 0, + "content": "Ablation Study of Global-to-Local Strategy. The proposed global-to-local strategy's efficacy is substantiated by the results displayed in Table 6. Contrasted with models relying exclusively on either global or local feature extraction, our hybrid approach ensures a comprehensive global perspective while meticulously capturing local nuances, leading to superior performance outcomes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.658, + 0.37, + 0.673 + ], + "angle": 0, + "content": "5 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.84 + ], + "angle": 0, + "content": "This paper introduces a new method for 3D object detection that tackles the daunting task of 3D object annotation. In consideration of the necessity for either creating annotations from raw data or refining existing imperfect annotations within practice annotation scenarios, two dedicated strategies are proposed. \"Prompt in 2D, detect in 3D\" principle leverages straightforward 2D interactions, like clicks or boxes, to streamline the transition between 2D images and 3D object annotations. \"Detect in 3D, refine in 3D\" strategy further endows our model with self-correction capability. Evaluations on the nuScenes dataset validate our approach's superiority. Beyond a mere advancement, our work can serve as a cornerstone for future 3D object detection endeavors." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.42, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Interactive 3D Object Detection with Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.257 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China (NO. 62322608), in part by the Fundamental Research Funds for the Central Universities under Grant 22lgqb25, and in part by the Open Project Program of the Key Laboratory of Artificial Intelligence for Perception and Understanding, Liaoning Province (AIPU, No. 20230003)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.284, + 0.323, + 0.3 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.374 + ], + "angle": 0, + "content": "1. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1090-1099 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.376, + 0.785, + 0.431 + ], + "angle": 0, + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11621-11631 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.785, + 0.472 + ], + "angle": 0, + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European Conference on Computer Vision. pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.474, + 0.785, + 0.514 + ], + "angle": 0, + "content": "4. Chen, X., Zhao, Z., Zhang, Y., Duan, M., Qi, D., Zhao, H.: Focalclick: Towards practical interactive image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1300-1309 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.517, + 0.785, + 0.557 + ], + "angle": 0, + "content": "5. Chen, X., Ma, H., Wan, J., Li, B., Xia, T.: Multi-view 3d object detection network for autonomous driving. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1907-1915 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.559, + 0.785, + 0.585 + ], + "angle": 0, + "content": "6. Chen, Y., Chen, Q., Sun, P., Chen, S., Wang, J., Cheng, J.: Enhancing your trained detrs with box refinement. arXiv preprint arXiv:2307.11828 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.587, + 0.785, + 0.613 + ], + "angle": 0, + "content": "7. Choi, D., Cho, W., Kim, K., Choo, J.: idet3d: Towards efficient interactive object detection for lidar point clouds. arXiv preprint arXiv:2312.15449 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.616, + 0.785, + 0.655 + ], + "angle": 0, + "content": "8. Ge, C., Chen, J., Xie, E., Wang, Z., Hong, L., Lu, H., Li, Z., Luo, P.: Metabev: Solving sensor failures for bev detection and map segmentation. arXiv preprint arXiv:2304.09801 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.657, + 0.785, + 0.697 + ], + "angle": 0, + "content": "9. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition. pp. 3354-3361. IEEE (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.785, + 0.727 + ], + "angle": 0, + "content": "0. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2961-2969 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.785, + 0.755 + ], + "angle": 0, + "content": "1. Huang, J., Huang, G.: Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.797 + ], + "angle": 0, + "content": "2. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.221, + 0.318, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "R. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.19 + ], + "angle": 0, + "content": "14. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12697-12705 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.233 + ], + "angle": 0, + "content": "15. Lee, C., Park, S., Song, H., Ryu, J., Kim, S., Kim, H., Pereira, S., Yoo, D.: Interactive multi-class tiny-object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14136-14145 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.787, + 0.275 + ], + "angle": 0, + "content": "16. Li, Y., Chen, Y., Qi, X., Li, Z., Sun, J., Jia, J.: Unifying voxel-based representation with transformer for 3d object detection. Advances in Neural Information Processing Systems 35, 18442-18455 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.787, + 0.318 + ], + "angle": 0, + "content": "17. Li, Z., Wang, F., Wang, N.: Lidar r-cnn: An efficient and universal 3d object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7546-7555 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.787, + 0.374 + ], + "angle": 0, + "content": "18. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European Conference on Computer Vision. pp. 1-18. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.374, + 0.787, + 0.416 + ], + "angle": 0, + "content": "19. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2980-2988 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.416, + 0.787, + 0.459 + ], + "angle": 0, + "content": "20. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.787, + 0.502 + ], + "angle": 0, + "content": "21. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. In: European Conference on Computer Vision. pp. 531-548. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.787, + 0.544 + ], + "angle": 0, + "content": "22. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X., Sun, J.: Petrov2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.787, + 0.6 + ], + "angle": 0, + "content": "23. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2774-2781. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.6, + 0.787, + 0.657 + ], + "angle": 0, + "content": "24. Lu, Y., Ma, X., Yang, L., Zhang, T., Liu, Y., Chu, Q., Yan, J., Ouyang, W.: Geometry uncertainty projection network for monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3111-3121 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.787, + 0.699 + ], + "angle": 0, + "content": "25. Maninis, K.K., Caelles, S., Pont-Tuset, J., Van Gool, L.: Deep extreme cut: From extreme points to object segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 616-625 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.699, + 0.787, + 0.742 + ], + "angle": 0, + "content": "26. Meng, D., Chen, X., Fan, Z., Zeng, G., Li, H., Yuan, Y., Sun, L., Wang, J.: Conditional detr for fast training convergence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3651-3660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.787, + 0.799 + ], + "angle": 0, + "content": "27. Meng, Q., Wang, W., Zhou, T., Shen, J., Jia, Y., Van Gool, L.: Towards a weakly supervised framework for 3d point cloud object detection and annotation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(8), 4454-4468 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.842 + ], + "angle": 0, + "content": "28. Papadopoulos, D.P., Uijlings, J.R., Keller, F., Ferrari, V.: Extreme clicking for efficient object annotation. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4930-4939 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.419, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Interactive 3D Object Detection with Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "29. Papadopoulos, D.P., Uijlings, J.R., Keller, F., Ferrari, V.: Training object class detectors with click supervision. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6374-6383 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.244 + ], + "angle": 0, + "content": "30. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 194-210. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.244, + 0.788, + 0.284 + ], + "angle": 0, + "content": "31. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.284, + 0.788, + 0.325 + ], + "angle": 0, + "content": "32. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems 28 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.365 + ], + "angle": 0, + "content": "33. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-779 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.406 + ], + "angle": 0, + "content": "34. Shin, K., Kwon, Y.P., Tomizuka, M.: Roarnet: A robust 3d object detection based on region approximation refinement. In: 2019 IEEE Intelligent Vehicles Symposium (IV). pp. 2510-2515. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.406, + 0.788, + 0.447 + ], + "angle": 0, + "content": "35. Sofiuk, K., Petrov, I.A., Konushin, A.: Reviving iterative training with mask guidance for interactive segmentation. In: 2022 IEEE International Conference on Image Processing (ICIP). pp. 3141-3145. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.501 + ], + "angle": 0, + "content": "36. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2446-2454 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.501, + 0.788, + 0.542 + ], + "angle": 0, + "content": "37. Wang, B., Wu, V., Wu, B., Keutzer, K.: Latte: accelerating lidar point cloud annotation via sensor fusion, one-click annotation, and tracking. In: 2019 IEEE Intelligent Transportation Systems Conference (ITSC). pp. 265-272. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.596 + ], + "angle": 0, + "content": "38. Wang, K., Zhuang, J., Li, G., Fang, C., Cheng, L., Lin, L., Zhou, F.: De-biased teacher: Rethinking iou matching for semi-supervised object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2573-2580 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.637 + ], + "angle": 0, + "content": "39. Wang, T., Zhu, X., Pang, J., Lin, D.: Fcos3d: Fully convolutional one-stage monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 913-922 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.678 + ], + "angle": 0, + "content": "40. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2567-2575 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.678, + 0.788, + 0.718 + ], + "angle": 0, + "content": "41. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning. pp. 180–191. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.759 + ], + "angle": 0, + "content": "42. Xie, Y., Xu, C., Rakotosaona, M.J., Rim, P., Tombari, F., Keutzer, K., Tomizuka, M., Zhan, W.: Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection. arXiv preprint arXiv:2304.14340 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.8 + ], + "angle": 0, + "content": "43. Xu, M., Zhang, Z., Hu, H., Wang, J., Wang, L., Wei, F., Bai, X., Liu, Z.: End-to-end semi-supervised object detection with soft teacher. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3060-3069 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.8, + 0.788, + 0.84 + ], + "angle": 0, + "content": "44. Xu, N., Price, B., Cohen, S., Yang, J., Huang, T.S.: Deep interactive object selection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 373-381 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "R. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "45. Yan, J., Liu, Y., Sun, J., Jia, F., Li, S., Wang, T., Zhang, X.: Cross modal transformer via coordinates encoding for 3d object detection. arXiv preprint arXiv:2301.01283 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.786, + 0.218 + ], + "angle": 0, + "content": "46. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "47. Yang, J., Zeng, A., Li, F., Liu, S., Zhang, R., Zhang, L.: Neural interactive keypoint detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15122-15132 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.3 + ], + "angle": 0, + "content": "48. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3dssd: Point-based 3d single stage object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11040-11048 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "49. Yao, A., Gall, J., Leistner, C., Van Gool, L.: Interactive object detection. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition. pp. 3242-3249. IEEE (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.787, + 0.397 + ], + "angle": 0, + "content": "50. Zhang, J., Lin, X., Zhang, W., Wang, K., Tan, X., Han, J., Ding, E., Wang, J., Li, G.: Semi-detr: Semi-supervised object detection with detection transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23809-23818 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.397, + 0.787, + 0.438 + ], + "angle": 0, + "content": "51. Zhang, R., Qiu, H., Wang, T., Guo, Z., Xu, X., Qiao, Y., Gao, P., Li, H.: Monodetr: depth-guided transformer for monocular 3d object detection. arXiv preprint arXiv:2203.13310 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.438, + 0.787, + 0.467 + ], + "angle": 0, + "content": "52. Zhao, J., Sun, L., Li, Q.: Recursivelydet: End-to-end region-based recursive object detection. arXiv preprint arXiv:2307.13619 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.508 + ], + "angle": 0, + "content": "53. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4490-4499 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.787, + 0.551 + ], + "angle": 0, + "content": "54. Zhou, Y., Zhu, H., Liu, Q., Chang, S., Guo, M.: Monoatt: Online monocular 3d object detection with adaptive token transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17493-17503 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.551 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Interactive 3D Object Detection with Prompts/20a1b205-7505-40a6-b499-29fffaad19e8_origin.pdf b/2024/Interactive 3D Object Detection with Prompts/20a1b205-7505-40a6-b499-29fffaad19e8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..17ce679817b8d9defd8953a73b3dfa60ded6b006 --- /dev/null +++ b/2024/Interactive 3D Object Detection with Prompts/20a1b205-7505-40a6-b499-29fffaad19e8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f2cc6b9450a09b7f81e064b2f791021b5f37922881a1c8ee26d733822228bad +size 4457075 diff --git a/2024/Interactive 3D Object Detection with Prompts/full.md b/2024/Interactive 3D Object Detection with Prompts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0143d539bd64e6a48735c6de50a80cdef0fa16da --- /dev/null +++ b/2024/Interactive 3D Object Detection with Prompts/full.md @@ -0,0 +1,354 @@ +# Interactive 3D Object Detection with Prompts + +Ruifei Zhang $^{1,2\star}$ , Xiangru Lin $^{4\star}$ , Wei Zhang $^{4}$ , Jincheng Lu $^{4}$ , Xuekuan Wang $^{4}$ , Xiao Tan $^{4}$ , Yingying Li $^{4}$ , Errui Ding $^{4}$ , Jingdong Wang $^{4}$ , and Guanbin Li $^{3,5\star\star}$ + +1 The Chinese University of Hong Kong, Shenzhen, China + +2 Shenzhen Research Institute of Big Data, Shenzhen, China + +3 Sun Yat-sen University, Guangzhou, China + +$^{4}$ Department of Computer Vision Technology (VIS), Baidu Inc., China + +Peng Cheng Laboratory + +ruifeizhang@link.cuhk.edu.cn, liguanbin@mail.sysu.edu.cn, + +{linxiangru,zhangwei99,lujincheng01,wangxuekuan,tanxiao01, + +liyingying05,dingerrui,wangjingdong}@baidu.com + +Abstract. The evolution of 3D object detection hinges not only on advanced models but also on effective and efficient annotation strategies. Despite this progress, the labor-intensive nature of 3D object annotation remains a bottleneck, hindering further development in the field. This paper introduces a novel approach, incorporated with "prompt in 2D, detect in 3D" and "detect in 3D, refine in 3D" strategies, to 3D object annotation: multi-modal interactive 3D object detection. Firstly, by allowing users to engage with simpler 2D interaction prompts (e.g., clicks or boxes on a camera image or a bird's eye view), we bridge the complexity gap between 2D and 3D spaces, reimagining the annotation workflow. Besides, Our framework also supports flexible iterative refinement to the initial 3D annotations, further assisting annotators in achieving satisfying results. Evaluation on the nuScenes dataset demonstrates the effectiveness of our method. And thanks to the prompt-driven and interactive designs, our approach also exhibits outstanding performance in open-set scenarios. This work not only offers a potential solution to the 3D object annotation problem but also paves the way for further innovations in the 3D object detection community. + +Keywords: Interactive 3D Object Detection $\cdot$ Prompt Learning + +# 1 Introduction + +3D object detection is pivotal for autonomous driving, gaining significant traction in recent years. Current research often derives 3D object information from sources such as monocular [24, 39, 51, 54] or multi-view camera images [11, 12, 18, 21], point cloud [17, 46, 53], and multi-modal sensors [1, 5, 45]. This has led to remarkable progress in detection performance. A major catalyst behind these advancements is the availability of high-quality 3D datasets like KITTI [9], + +![](images/d7c34366f328e5c3679cc1fce84e33fa557cbd43240bb26a903fc1f25d9c43cf.jpg) +Traditional Way of 3D Detection Data Annotation +Our proposed Way of 3D Detection Data Annotation + +![](images/e778823b744d52fa4b4764e650138b8d133fd6521d8cb6e1b56110d5cf50896a.jpg) +Fig. 1: We introduce an interactive 3D object detection concept. Rather than the traditional method where a model produces 3D object detection boxes followed by user refinement in 3D space, our strategy is "prompt in 2D, detect in 3D, refine in 3D". This method streamlines the annotation process, reducing the need for user refinements. + +nuScenes [2], and Waymo [36]. These datasets fuel the progression of data-driven algorithms in 3D object detection. However, the intricate nature of 3D object annotation often necessitates significant human resources and incurs high labor costs. On average, annotating an hour of driving data takes hundreds of hours [27]. This annotation challenge potentially restricts the scale and diversity of datasets, possibly hindering further progression in this field. + +Addressing this annotation challenge is the central theme of our paper. While weakly-supervised [25, 27-29] and semi-supervised [38, 43, 50] object detection methods exist, they typically suffer from reduced performance, especially in practical scenarios. This does not substantially address the overarching issue of the data-hungry nature of the field. With the recent emergence of SAM [13], both industry and academia are now recognizing the potential of interactive annotation approaches. + +Our work aims to integrate this concept into 3D object detection, thus simplifying the annotation process, as presented in Fig. 1. Specifically, we observe that within the traditional process of annotating 3D detection data, annotators tend to concentrate on addressing two challenges stemming from imperfect pretrained 3D detectors: 1) supplementing missing closed-set objects and incorporating additional required open-set ones; 2) rectifying existing erroneous annotations. Both scenarios necessitate annotators to operate within the 3D point cloud space, which is extremely time-consuming and labor-intensive. + +To circumvent this, we create an interactive 3D object detection system enhanced by two strategies: prompt in $2D$ , detect in $3D$ , and detect in $3D$ , refine in $3D$ . The first principle enables users to interact with a 2D camera image or bird's eye view (BEV) of the point cloud, prompting the model to generate corresponding 3D bounding boxes, instead of encountering the 3D point cloud + +directly. Furthermore, benefiting from the prompt-driven annotation paradigm, our method inherently possesses the capability to annotate open-set objects. The latter ensures annotators refine the initial 3D bounding box at the fastest pace with minimal effort, which significantly streamlines operations in the 3D space. + +In light of these principles, we present our multi-modal interactive 3D object detection task and algorithm. Firstly, it employs prompts in the form of clicks or boxes on a 2D camera or BEV view, guiding the model to generate the relevant 3D bounding box. Our framework is reminiscent of the DETR-series detectors [3, 20, 26, 40] but comes with distinctive features. We integrate user inputs as prompt tokens and design a learnable component that investigates multi-modal features of indicated objects, informed by the prompts. Furthermore, we incorporate a global-to-local strategy to supply fine-grained object features, enhancing detection accuracy. Secondly, we empower our model with iterative refinement capability by incorporating the predicted 3D bounding box encoding as a new prompt into the query sequence. For those inaccurate 3D predictions, our framework allows annotators to adjust a few wrong attributes, e.g. dragging the center point position or adjusting the orientation angle of the 3D bounding box, and then perform self-correction on other characters. + +To rigorously evaluate our approach, we devise two evaluation configurations for practical annotation scenarios. The first, regarding Annotation from Raw Data, leverages user 2D prompts and produces corresponding 3D predictions. Besides, our framework is also compatible with starting from the results of existing State-of-the-Art (SOTA) 3D detection models, termed as Annotation from SOTA Model and iteratively refining them. Experimental results on the nuScenes dataset highlight the effectiveness of our method in various scenarios. + +In summary, this paper has the following contributions: + +- We introduce a new concept of multi-modal interactive 3D object detection, proposing a substantial solution to prevailing annotation challenges. Our work is expected to provide critical insights for the 3D object detection community across academic and industrial domains. +- We unveil a prompt-driven interactive 3D object detection methodology with self-correction capability, leveraging it to establish an interactive system tailored for practical annotation scenarios. +- Comprehensive experiments in both annotation scenarios and evaluation settings attest to our method's efficacy. + +# 2 Related Work + +# 2.1 3D Object Detection + +3D object detection techniques have been developed across a range of input modalities. Our approach emphasizes the advantages of interactive inputs, setting it apart from traditional methods: + +Camera-only. Monocular 3D object detection techniques [24, 39, 51, 54] enjoy advantages such as low cost and ease of deployment. However, it also faces the inherent challenge of limited depth information. Multi-view methods utilize data from multiple camera perspectives [11, 12, 18, 21, 22, 30, 41], which enhances the model's ability to perceive the 3D environment, thereby improving detection accuracy. + +LiDAR-only. Benefiting from the robustness and high precision, numerous studies utilize LiDAR data for 3D object detection. Point-based methods [17, 33,48] like PointNet [31] directly engage with raw LiDAR data, while grid-based strategies structure the data into 3D voxels [46,53] or feature pillars [14]. + +Multi-modality. Capitalizing on the strengths of camera images and point clouds, multi-modal 3D object detection has seen significant advancements [1, 5, 45]. Early methods combine camera images and point clouds at the input or result stages [5, 34]. While, current models, like TransFusion [1] and CMT [45], emphasize more intricate feature-level fusion. Uniquely, our method embeds interactive capacities, ensuring richer, user-tailored annotations. + +# 2.2 Interactive Object Detection/Segmentation + +Interactive methods have historically aimed to reduce the annotation workload while maintaining, or even enhancing, model performance: + +Interactive Object Detection. Early attempts, like the incremental learning strategy [49], require user corrections to update detectors. More recent models like C3Det [15] have substantially eased the annotation process. Tools such as LATTE [37] and iDet3D [7] further simplify the 3D object annotation process. Our framework, however, brings in the versatility of prompt formats and modalities, offering users a broadened interaction spectrum coupled with robust detection capabilities. + +Interactive Object Segmentation. Interactive segmentation has matured significantly over the years [4,35,44]. Models like SAM [13] have empowered users with tools like clicks, boxes, and referring expressions to guide the segmentation process. Diverging from interactive segmentation confined to 2D spaces, our work addresses the challenge of integrating 2D interactive prompts with 3D object detection, mitigating the spatial discrepancy between 2D annotations and 3D environments. + +# 2.3 Weakly-supervised Object Detection + +The notion of harnessing weak signals, like clicks [29] or extreme-points [25, 28], for supervising detectors has been a topic of intrigue. Some methods [27] have extended this idea to 3D object detection, using center-click based strategies. These approaches, while commendable for their balance between performance and annotation overhead, sometimes fail to harness the full potential of the detectors. Our methodology, on the other hand, incorporates strong 3D bounding box supervisions in an interactive setting, streamlining the data acquisition process without compromising on detection performance. + +![](images/711e38e5001e2e2f615a72da8466f938568c45427604ddda6d7dacce6293f55d.jpg) +Fig. 2: An overview of our proposed interactive 3D object detection system/method. Taking multi-modal data with user prompts as input, we explicitly add 3D information $\mathbf{P}_{im},\mathbf{P}_{pc}$ into features $\mathbf{F}_{im},\mathbf{F}_{pc}$ by 3D Position Encoding. Object queries consist of a learnable token $\mathbf{Q}_l$ and prompt ones $\mathbf{Q}_p$ , which are generated by Prompt-Guided Query generation. We also introduce a Global-to-Local enhancement strategy to dynamically adjust features integrated with queries from global representations to local RoI details. Besides, our framework allows users to refine any of 3D attributes directly, and encode the refined 3D bounding box $b_{L}^{r^{\prime}}$ and category information $c_{L}^{r^{\prime}}$ , along with attribute refinement record $p_a^r$ and $p_c^r$ , into a new prompt token $\mathbf{Q}_{3D}$ by 3D Attribute Encoding. $g,w,l,h,\theta$ denotes the gravity center, width, length, height and yaw angle respectively. This iterative refinement propels our model to automatically rectify other attributes, ultimately delivering satisfactory results. + +# 3 Methodology + +# 3.1 Problem Definition + +We present an interactive multi-modal 3D object detection algorithm. Given a multi-view set of camera images and associated point cloud data, the objective is to predict both the class labels and the 3D bounding boxes, driven by user prompts either in the form of a point or a box. Further, our model also supports iteratively self-correction with limited human assistance. To supervise the iteration process, the Ground Truth $(c, b)$ is employed. Here, $c$ signifies the category label, and $b$ represents the 3D attributes. We adopt a deep supervision strategy for our multi-layer Transformer decoder, which guides the predictions across each decoder layer of each iteration, denoted as $\{(c_i^r, b_i^r), i = 1, 2, \dots, L, r = 0, 1, \dots, M\}$ , with $L$ and $M$ being the total number of decoder layers and iterative refinements. + +# 3.2 Overview + +The architecture underpinning our interactive 3D object detection method is visualized in Fig. 2, which could be roughly decomposed into two stages, i.e. "prompt in $2D$ , detect in $3D$ and "detect in $3D$ , refine in $3D$ ". In the first phase, two separate backbones are utilized to extract multi-view image features $\mathbf{F}_{im} = \{\mathbf{F}_{im}^{t}, t = 1,2,\dots,T\}$ and point cloud features $\mathbf{F}_{pc}$ respectively. To enhance these features with 3D-awareness, following PETR [21] and CMT [45], we introduce 3D position encoding. Specifically, for the multi-view images, this encoding is + +represented as $\mathbf{P}_{im} = \{\mathbf{P}_{im}^t, t = 1,2,\dots,T\}$ , and for the point cloud, it is denoted as $\mathbf{P}_{pc}$ . For object queries, denoted as $\mathbf{Q}$ , we utilize a unique blending of a learnable token, $\mathbf{Q}_l$ , and prompt-specific tokens, $\mathbf{Q}_p$ , for every detected object. The prompt queries, $\mathbf{Q}_p$ , play a pivotal role in guiding the learnable token, $\mathbf{Q}_l$ , by leveraging the insights from the multi-modal features. Furthermore, to enhance the detection accuracy across various object scales, we incorporate a global-to-local strategy within the decoder. This strategy ensures that the object queries are continually enriched by granular local features, enabling finer detections. Ultimately, the learnable token predicts the object's class and its 3D bounding box $(c_L^0,b_L^0)$ . + +Subsequently, the model enters an interactive refinement stage. This stage reuses the multi-modal features but updates the object queries $\mathbf{Q}$ by inserting the encoding of the refined 3D prediction from the current iteration. Specifically, in the $r$ -th iteration, we enable annotators to fine-tune the current prediction $(c_{L}^{r}, b_{L}^{r})$ and encode the refined result $(c_{L}^{r'}, b_{L}^{r'})$ and corresponding attribute refinement record $(p_{c}^{r}, p_{a}^{r})$ into a new prompt $\mathbf{Q}_{3D}$ . $\mathbf{Q}_{3D}$ and $\mathbf{Q}_p$ collectively drive $\mathbf{Q}_l$ to generate more accurate predictions in the next iteration. Note that we omit the iteration superscript of $\mathbf{Q}_{3D}$ and $\mathbf{Q}$ for simplicity. + +# 3.3 3D Position Encoding + +Multi-view Camera Images. The core idea of embedding 3D position for multi-view image features includes two steps: 1) modeling each pixel coordinate as a series of points along a ray in the camera frustum space, and 2) projecting camera frustum coordinates into 3D space. Specifically, for a given camera image feature, denoted as $\mathbf{F}_{im}^{t}\in \mathbb{R}^{W_{im}\times H_{im}\times C}$ , each pixel can be visualized as tracing a ray within the camera frustum space. This perspective allows us to sample $D$ discrete points along the depth dimension. This process is mathematically captured by: + +$$ +\mathbf {P} _ {c f} (u, v, j) = \left(u \times d _ {j}, v \times d _ {j}, d _ {j}, 1\right) ^ {T}, j = 1, 2, \dots , D \tag {1} +$$ + +Here, $\mathbf{P}_{cf}$ indicates the camera's frustum coordinates. Pixel coordinates in the image are denoted by $(u,v)$ , and $d_{j}$ represents the $j$ -th depth value. + +To relate these coordinates to a 3D context, we employ an inverse 3D projection method: + +$$ +\mathbf {P} _ {3 d} ^ {t} (u, v, j) = \mathbf {K} _ {t} ^ {3 D} \mathbf {K} _ {t} ^ {- 1} \mathbf {P} _ {c f} (u, v, j) \tag {2} +$$ + +where $\mathbf{K}_t^{3D} \in \mathbb{R}^{4 \times 4}$ represents the transformation matrix that links 3D space with the $t$ -th camera's coordinate space. $\mathbf{K}_t \in \mathbb{R}^{4 \times 4}$ is the intrinsic matrix of $t$ -th camera. To ensure consistent value scales, we normalize the 3D coordinates to the range [0,1]. The final 3D position encoding for the $t$ -th camera image, represented as $\mathbf{P}_{im}^t \in \mathbb{R}^{W_{im} \times H_{im} \times C}$ , is derived via a multi-layer perceptron (MLP): $\mathbf{P}_{im}^t = \mathrm{MLP}_{im}(\mathbf{P}_{3d}^t)$ . + +Point Cloud. In contrast to the approach for camera image features, where points are assembled along the depth axis, the LiDAR point set necessitates sampling along the height axis. Specifically, given a BEV feature map, denoted as + +$\mathbf{F}_{pc} \in \mathbb{R}^{W_{pc} \times H_{pc} \times C}$ , we can compute the corresponding 3D position information, represented as $\mathbf{P}_{pc} \in \mathbb{R}^{W_{pc} \times H_{pc} \times C}$ , using the following steps: First, we determine the 3D position based on the feature map's coordinates: + +$$ +\mathbf {P} _ {3 d} (u, v, j) = \left(u \times u _ {s}, v \times v _ {s}, h _ {j}, 1\right) ^ {T}, j = 1, 2, \dots , H \tag {3} +$$ + +Here, $(u,v)$ specifies the coordinate within the point cloud feature map. The variables $(u_{s},v_{s})$ represent the dimensions of the feature grid. $H$ is the number of sampled points along the height axis and $h_j$ is the $j$ -th height value. Next, we utilize a multi-layer perceptron (MLP) to compute the 3D position encoding from the derived 3D position: $\mathbf{P}_{pc} = \mathrm{MLP}_{pc}(\mathbf{P}_{3d})$ . Note that as in CMT, only one point is sampled along the height axis, thus simplifying the process into standard 2D position encoding. Consequently, in our methodology, we have adopted a sinusoidal embedding to cater to this behavior. + +# 3.4 Prompt-guided Query Generation + +Our approach provides an avenue for user interaction, distinct from traditional DETR-based detectors that often rely on multiple queries to discern image objects. By merging learnable tokens, $\mathbf{Q}_l$ , with user-directed prompt tokens, $\mathbf{Q}_p$ , we achieve object detection that is not only efficient but also offers a new angle on object localization. + +Learnable Query $\mathbf{Q}_l$ : Every object is associated with a unique learnable token that facilitates feature probing, and subsequent prediction of class labels and 3D bounding boxes. In line with prior research [20, 40], we conceptualize the learnable query as an anchor point, initiating it with the positional prior of the user prompt. The initialization process unfolds over three phases: + +1). 2D Center Position Computation with User Prompt. A user might provide a bounding box $box = (x_1, y_1, x_2, y_2)$ or a singular point $point = (x, y)$ . Depending on the input type, the 2D center position $(c_x, c_y)$ is determined as: + +$$ +\left(c _ {x}, c _ {y}\right) = \left\{ \begin{array}{l l} \left(\frac {x _ {1} + x _ {2}}{2}, \frac {y _ {1} + y _ {2}}{2}\right) & \text {i f b o u n d i n g b o x} \\ (x, y) & \text {i f p o i n t} \end{array} \right. \tag {4} +$$ + +2). 2D Center Transformation into 3D Coordinate. Given a user prompt on the BEV map, the 2D space naturally omits height data. We address this by designating $c_{z} = 0.5$ , the median value of height, resulting in a 3D center labeled as center $= (c_{x},c_{y},c_{z})$ . Alternatively, for a user prompt within 2D multi-view images, we transform the center coordinates $(c_{x},c_{y})$ into a set of 3D points along the depth axis, using Eqs. (1) and (2). The median of the ray is then selected as the 3D center, represented as center $= (c_x',c_y',c_z')$ . The 3D center serves as a reference point for bounding box prediction and we prioritize the center from BEV if the user provides prompts in both modalities. + +3). 3D Position Encoding Extraction from Both Modalities. The 3D center is projected onto both the camera image and BEV plane, yielding their 3D position encodings $\mathbf{Q}_{im} \in \mathbb{R}^{1 \times C}$ and $\mathbf{Q}_{pc} \in \mathbb{R}^{1 \times C}$ , respectively. The learnable query is initialized by combining information from both modalities: $\mathbf{Q}_l = \mathbf{Q}_{im} + \mathbf{Q}_{pc}$ . + +![](images/63ec824553a1b3a32dc38049373b5970cda6f01738b8f24593fdf1941ceb0991.jpg) +Fig. 3: Illustration of Global-to-Local Enhancement. Our queries interact with unmodified multi-modal features and its object's local features respectively in the global and local phases. Self-Attention is confined within queries of identical objects in both stages. + +Prompt Query $\mathbf{Q}_p$ : It is designed to embed user guidance by merging position encoding and context embedding: $\mathbf{Q}_p = \mathbf{Q}_{pe} + \mathbf{Q}_{ce}$ . This synthesis is constructed for both box and point prompt formats: For the box format, the box position is delineated using its center point. From our 3D position encoding, the position encoding for box center, $\mathbf{Q}_{pe} \in \mathbb{R}^{1 \times C}$ , is derived. The context embedding, $\mathbf{Q}_{ce} \in \mathbb{R}^{1 \times C}$ , is secured via average pooling of the feature region bounded by the prompt box. For the point format, the same technique determines the position encoding, $\mathbf{Q}_{pe} \in \mathbb{R}^{1 \times C}$ . The feature at the indicated prompt point gives the context embedding, $\mathbf{Q}_{ce} \in \mathbb{R}^{1 \times C}$ . + +# 3.5 Global-to-Local Enhancement + +Given multi-modal features $\mathbf{F}_{pc}$ and $\mathbf{F}_{im}$ with their respective 3D position embeddings $\mathbf{P}_{pc}$ and $\mathbf{P}_{im}$ , along with object queries $\mathbf{Q} \in \mathbb{R}^{N \times C}$ , where $N$ represents the cumulative number of queries from $\mathbf{Q}_l$ and $\mathbf{Q}_p$ across all objects, we employ a Transformer-based decoder to amalgamate these features for prediction. + +Self-Attention: The formulation for the decoder's self-attention mechanism is: + +$$ +\mathbf {Q} _ {i} ^ {\prime} = \operatorname {s o f t m a x} \left(\mathbf {M} _ {s a} + \mathbf {Q} _ {i} \mathbf {Q} _ {i} ^ {T}\right) \mathbf {Q} _ {i} \tag {5} +$$ + +where $\mathbf{Q}_i$ corresponds to the queries from the $i$ -th layer, and $\mathbf{Q}_0 = \mathbf{Q}$ . The attention mask $\mathbf{M}_{sa}$ ensures interaction is confined within queries of identical objects: + +$$ +\mathbf {M} _ {s a} \left(q _ {1}, q _ {2}\right) = \left\{ \begin{array}{l l} 0, & \text {i f} q _ {1}, q _ {2} \text {b e l o n g t o t h e s a m e o b j e c t ,} \\ - \infty , & \text {o t h e r w i s e .} \end{array} \right. \tag {6} +$$ + +Cross-Attention: For cross-attention in the decoder, we express it as: + +$$ +\mathbf {Q} _ {i + 1} = \operatorname {s o f t m a x} (\mathbf {M} _ {c a} + \mathbf {Q} _ {i} ^ {\prime} \mathbf {K} ^ {T}) \mathbf {V} \tag {7} +$$ + +where $\mathbf{K}$ , $\mathbf{V}$ come from multi-modal features, $\mathbf{M}_{ca}$ is attention mask for cross-attention. Influenced by traditional two-stage detectors [32] and recent DETR-centric techniques [6, 52], we adopt a global-to-local enhancement strategy. + +This approach adjusts $\mathbf{K}$ and $\mathbf{V}$ dynamically to augment our model's local perceptiveness, as illustrated in Fig. 3. + +For the initial $Z$ layers among a total of $L$ decoder layers, queries access unmodified multi-modal features: + +$$ +\mathbf {K} = \left[ f _ {k} \left(\mathbf {F} _ {p c} + \mathbf {P} _ {p c}\right): f _ {k} \left(\mathbf {F} _ {i m} + \mathbf {P} _ {i m}\right) \right], \tag {8} +$$ + +$$ +\mathbf {V} = \left[ f _ {v} \left(\mathbf {F} _ {p c}\right): f _ {v} \left(\mathbf {F} _ {i m}\right) \right], \tag {9} +$$ + +where $\mathbf{K},\mathbf{V}\in \mathbb{R}^{(H_{pc}W_{pc} + H_{im}W_{im}T)\times C}$ are original multi-modal features under transformation $f_{k}(\cdot),f_{v}(\cdot)$ and concatenation $[\cdot :\cdot ]$ . $\mathbf{M}_{ca} = 0$ during the global cross-attention phase. + +In the succeeding $L - Z$ layers, local region-of-interest (RoI) features are extracted by RoI Align [10] via either the user-provided prompt box or the predicted 2D bounding box based on prompt point: + +$$ +\mathbf {K} = \left[ f _ {k} \left(\operatorname {R o I} \left(\mathbf {F} _ {p c} + \mathbf {P} _ {p c}\right)\right): f _ {k} \left(\operatorname {R o I} \left(\mathbf {F} _ {i m} + \mathbf {P} _ {i m}\right)\right) \right], \tag {10} +$$ + +$$ +\mathbf {V} = \left[ f _ {v} \left(\operatorname {R o I} \left(\mathbf {F} _ {p c}\right)\right): f _ {v} \left(\operatorname {R o I} \left(\mathbf {F} _ {i m}\right)\right) \right], \tag {11} +$$ + +where $\mathbf{K},\mathbf{V}\in \mathbb{R}^{(H_oW_o + H_oW_o)\times C}$ now signify local multi-modal features. $(H_{o},W_{o})$ is the output size of RoI Align. With $\mathbf{M}_{ca}$ defined similarly to $\mathbf{M}_{sa}$ , it ensures each query is exclusively attentive to its object's local features. Following the decoding process, we derive the object query $\mathbf{Q}_L$ from the last decoding layer, and subsequently produce the initial prediction via two feed-forward networks (FFNs): + +$$ +c _ {L} ^ {0} = \phi_ {c} \left(\mathbf {Q} _ {L}\right), \quad b _ {L} ^ {0} = \phi_ {b} \left(\mathbf {Q} _ {L}\right) \tag {12} +$$ + +Here, $c_{L}^{0}$ and $b_{L}^{0}$ represent the class and 3D bounding box predictions, respectively. While $\phi_{c}$ and $\phi_{b}$ are FFNs designated for classification and regression tasks. At this point, the first phase of "prompt in 2D, detect in 3D" is completed. + +# 3.6 3D Attribute Encoding + +For those inaccurate results of the initial predictions, we endow our model with iteratively self-correction capability under limited human interaction. Specifically, we decompose the 3D bounding box into (gravity center, width, length, height and yaw angle) five attributes, along with category information, and assume users to adjust a few wrong attributes, e.g. "dragging the center point of the bounding box". Given the prediction $b_{L}^{r}$ and $c_{L}^{r}$ of the $r$ -th iteration, the adjusted 3D bounding box and category are denoted as $b_{L}^{r'}$ and $c_{L}^{r'}$ . We also introduce two binary attribute refinement records $p_a^r \in \mathbb{R}^5$ and $p_c^r \in \mathbb{R}^1$ to explicitly aid the model in distinguishing between corrected attributes and those yet to be predicted. + +Finally, the 3D attribute query $Q_{3D}$ for each object is formed by the combination of both attribute and record embedding: + +$$ +\mathbf {Q} _ {3 D} = \mathrm {M L P} _ {a t t} \left(b _ {L} ^ {r ^ {\prime}}\right) + \mathrm {M L P} _ {r e c 1} \left(p _ {a} ^ {r}\right) + \mathrm {M L P} _ {c a t} \left(c _ {L} ^ {r ^ {\prime}}\right) + \mathrm {M L P} _ {r e c 2} \left(p _ {c} ^ {r}\right) \tag {13} +$$ + +And $\mathbf{Q}_{3D}$ is inserted into the corresponding query sequence of the respective object, collaborating with the original prompt tokens to jointly drive the model to provide more precise predictions. Note that in the following decoder process, $\mathbf{M}_{sa}$ and $\mathbf{M}_{ca}$ are also dynamically adapted according to Eq. (6). + +# 3.7 Label Assignment and Losses + +Due to the intrinsic one-to-one correspondence between our learnable queries and prompt objects, our proposed interactive 3D object detection task does not require label assignment like Hungarian Algorithm. To this end, following the methodologies in [21,45], we employ the Focal loss [19] for the classification task and the L1 loss for regression. To bolster the training of individual decoder layers, a deep supervision strategy is implemented. The 3D loss function amalgamates these predictions and is represented as: + +$$ +\mathcal {L} _ {3 D} = \sum_ {r = 0} ^ {M} \sum_ {i = 1} ^ {L} \lambda_ {c l s} \mathcal {L} _ {c l s} \left(c _ {i} ^ {r}, c\right) + \lambda_ {r e g} \mathcal {L} _ {r e g} \left(b _ {i} ^ {r}, b\right) \tag {14} +$$ + +Here, $c$ and $b$ are the ground truths, $c_i^r$ and $b_i^r$ denote the predictions of the $i$ -th decoder layer and $r$ -th iteration, which are generated by the shared MLP as Eq. (12). $\lambda_{cls}$ and $\lambda_{reg}$ are weights for balancing loss items. Besides, we also utilize L1 loss to supervise the 2D bounding box prediction on the image or BEV modality, which is produced by another MLP and serves as the local region for point prompts to perform the global-to-local strategy. The overarching loss function is the combination of both 3D and 2D losses. + +# 4 Experiments + +# 4.1 Experiment Settings + +Dataset splits: The nuScenes dataset [2] is adopted in our work for training and evaluation. This dataset encompasses 1000 driving scenes, capturing data via six cameras, one LiDAR, and five radars. Following the official partition, we employ 750 scenes for training and 150 for validation. Contrasting prior work, our interactive 3D object detection approach leverages user prompts to guide the model's predictions. Similar to common interactive algorithms [4, 47], we reconstruct the evaluation benchmark by simulating user prompts and refinements from Ground Truth. Notably, since the nuScenes test set's Ground Truth remains inaccessible, we exclude it from our evaluations. + +Training Protocol: Throughout the training process, we leverage Ground Truth to simulate practical user prompts and iterative refinements. Detailed training procedures for Prompt in 2D, Detect in 3D and Detect in 3D, Refine in 3D phases are provided in the supplementary materials. + +Evaluation Protocol: We evaluate our method through human annotation comparisons and simulated experiments, under two distinct annotation scenarios: (1) Annotation from Raw Data and (2) Annotation from SOTA + +Table 1: Comparisons of the average and standard deviation time cost required for annotating an image extracted from the nuScenes validation set, with four annotation strategies, i.e. manual-only, Ours, SOTA model CMT with manual correction (CMT+M) and Ours (CMT+Ours) respectively. + +
Time cost (s)Annotation from Raw DataAnnotation from SOTA Model
Manual-onlyOursCMT+MCMT + Ours
nuScenes-val124±2855±1740±1224±7
+ +Model. We describe the simulated experimental settings as follows: in scenario (1), we employ 2D projections of 3D Ground Truth with perturbations to simulate user prompts, while in scenario (2), SOTA 3D detection models provide initial prediction results to initiate our algorithm. The attributes from Ground Truth are utilized to mimic subsequent user refinements in both scenarios. We explore various orders of adjustments to bounding box attributes and assess our model both without and with corrected category priors, denoted as Ours and Ours*, respectively. * indicates the pre-calibration of category data. Unless specified, we prioritize using 2D bounding boxes as prompts and following the adjustment sequence of [gravity center, yaw angle, height, width, length]. Experimental results for alternative prompt formats (e.g., point prompts) and refinement orders are detailed in the supplementary materials. Furthermore, in addition to conventional closed-set evaluation, we also conduct extended experiments on challenging open-set scenarios. The nuScenes official metrics, i.e. nuScenes Detection Score (NDS), mean Average Precision (mAP) are adopted to evaluate the approaches. + +# 4.2 Human Annotation Comparisons + +To showcase the efficacy of our proposed method, we perform annotation comparisons across two scenarios employing four strategies: Manual-only, Ours, $\mathrm{CMT + M}$ (manual correction), and $\mathrm{CMT + Ours}$ . Ten experienced annotators are recruited to annotate the same set of twenty randomly selected nuScenes validation scenes using each of these strategies. Statistical analysis in Table 1 demonstrates that benefiting from the interactive design, our method achieves a speedup of approximately $56\%$ and $40\%$ compared to manual-only and $\mathrm{CMT + M}$ , respectively. + +# 4.3 Closed-Set 3D Object Detection + +Annotation from SOTA Model Evaluation. Results from our experiments on the nuScenes validation set are tabulated in Table 2. In practical applications, annotators often streamline their efforts by initiating annotations from the 3D predictions generated by SOTA models, such as CMT. Our model seamlessly integrates into this workflow, harnessing its robust capabilities to effectively aid in annotation tasks. Specifically, given the 3D predictions from CMT, Table 2 illustrates the interactive refinement results, without employing category + +Table 2: Comparisons of annotation performance on nuScenes val set. "M" means manual correction. (Ri) denotes the $i$ -th iteration refinement. "C" and "L" are camera and LiDAR modalities respectively. + +
MethodModalityNDS↑mAP↑
Traditional 3D Detectors
UVTR [16]CL70.265.4
BEVFusion [23]CL71.468.5
MetaBEV [8]CL71.568.0
SpaseFusion [42]CL72.870.4
CMT [45]CL72.970.3
Annotation from SOTA Model
CMT+M (R1)CL81.381.6
CMT+M (R2)CL83.881.6
CMT+M (R3)CL84.381.6
CMT+M (R4)CL85.381.6
CMT+M (R5)CL86.381.6
CMT+Ours (R1)CL83.4 (+2.1)86.3 (+4.7)
CMT+Ours (R2)CL86.3 (+2.5)86.3 (+4.7)
CMT+Ours (R3)CL87.4 (+3.1)87.0 (+5.4)
CMT+Ours (R4)CL89.3 (+4.0)89.1 (+7.5)
CMT+Ours (R5)CL90.4 (+4.1)89.4 (+7.8)
Annotation from Raw Data
OursCL75.076.2
Ours (R1)CL84.4 (+9.4)88.0 (+11.8)
Ours (R2)CL87.2 (+12.2)88.0 (+11.8)
Ours (R3)CL88.2 (+13.2)88.4 (+12.2)
Ours (R4)CL89.5 (+14.5)89.5 (+13.3)
Ours (R5)CL90.5 (+15.5)89.6 (+13.4)
+ +correction priors. As a comparative reference, we perform identical manual attribute adjustments on the initial results predicted by CMT. It is evident that, in five consecutive iterations, our approach consistently outperforms $\mathrm{CMT + M}$ by margins of $2.1\%$ , $2.5\%$ , $3.1\%$ , $4.0\%$ , and $4.1\%$ , respectively, in terms of NDS. Additionally, CMT maintains a constant mAP after the initial refinement of the center position, lacking self-correction for object categories. In contrast, our approach iteratively refines predictions with minimal human intervention, resulting in sustained performance improvements. In addition, we also provide comparisons with corrected category priors, denoted as $\mathrm{CMT + M^*}$ and $\mathrm{CMT + Ours^*}$ in Fig. 4. Through the rectification of object categories, our model exhibits a significantly enhanced self-correction capability, outperforming $\mathrm{CMT + M^*}$ by a considerable margin. + +Annotation from Raw Data Evaluation. In cases involving missed detections by existing 3D detectors, our method is expected to have the capability of annotating from raw data. As shown in Table 2, by leveraging a box as the prompt format, our model registers an NDS of $75.0\%$ and an mAP of $76.2\%$ . Subsequent iterations of refinements contribute to continued performance gains, and the final model attains an NDS score of $90.5\%$ and mAP of $89.6\%$ . Some visualization results are presented in Fig. 5. + +![](images/68ea3277f58d0d6a10100bbb409211c0c9f0c2cf903e37f4c8e9556d6405e2c7.jpg) +Fig. 4: Performance comparisons of annotation from SOTA model. R1-R5 follows a fixed refinement order. * indicates the utilization of category correction. + +Table 3: Categories and quantities of open-set objects extracted from nuScenes validation set. + +
CategoryNumCategoryNum
personal_mobility15stroller103
animal32bicycle_rac246
pushableifiable2925debris514
ambulance30police64
+ +Table 5: Ablation study of prompt queries. "PE", "CE" are position encoding and context encoding. "It." denotes the iteration. + +
It.QtQPPEQpCEQ3DbL'NDS↑mAP↑
0✓73.575.5
0✓✓74.575.7
0✓✓✓75.076.2
1✓✓✓✓83.285.9
1✓✓✓✓84.488.0
+ +Table 4: Open-set evaluation. + +
MethodModalityNDS↑mAP↑
Annotation from Raw Data
Ours*CL89.391.9
Ours* (R1)CL94.199.5
Ours* (R2)CL96.699.5
Ours* (R3)CL97.399.5
Ours* (R4)CL98.199.5
Ours* (R5)CL99.299.5
+ +Table 6: Ablation study of Global-to-Local strategy (G2L). "L", "G" denotes utilizing local RoI or global multi-modal features to perform Cross-Attention. + +
It.AttentionNDS↑mAP↑
0G73.775.1
0L74.375.2
0G2L75.076.2
+ +# 4.4 Open-Set 3D Object Detection + +To delve deeper into the 3D object annotation capabilities guided by 2D prompts, we extract eight novel categories from the nuScenes dataset, which are not employed in the training process. Table 3 outlines their distribution on the validation set. Since our primary goal is to validate the model's generalization ability based on user-provided prompts, as opposed to venturing into open-vocabulary recognition, we ignore the class prediction and assume that the class information is provided manually together with 2D prompts, denoted as Ours*. The experimental results in Table 4 demonstrate that our model possesses the capability to detect 3D objects based on 2D priors. Furthermore, with manual interaction for refinement, the detection performance improves progressively. + +![](images/58d504a583bb570b5a5c5bfd8e5dc20090a6c802ad9351fcc59b8e6cc5d17c9e.jpg) +Fig. 5: Visualizations of the 3D detection results from our method are presented. Ground Truth and our predictions are exhibited in green and blue, respectively. The outcomes highlighted by red circles illustrate that, with a single round of manual adjustment focusing solely on the gravity center, our model automatically rectified other attributes such as yaw angles, thus yielding more accurate 3D results. + +# 4.5 Ablation Study + +Ablation Study of Object Queries. The pivotal role of prompt tokens is evident from the results in Table 5. When solely introducing prompt queries with position encoding, the model attains a nuanced understanding of the object's spatial location, registering an NDS of $74.5\%$ —a significant boost of $1.0\%$ . Moreover, integrating context embedding, which encapsulates appearance attributes of objects, further escalates performance, culminating in a gain of $0.5\%$ . Besides, we also delve into an analysis of the effectiveness of $Q_{3D}$ during the first iteration refinement stage. Clear performance improvements are evident with the explicit inclusion of the current attribute refinement record $p_a^r$ . + +Ablation Study of Global-to-Local Strategy. The proposed global-to-local strategy's efficacy is substantiated by the results displayed in Table 6. Contrasted with models relying exclusively on either global or local feature extraction, our hybrid approach ensures a comprehensive global perspective while meticulously capturing local nuances, leading to superior performance outcomes. + +# 5 Conclusions + +This paper introduces a new method for 3D object detection that tackles the daunting task of 3D object annotation. In consideration of the necessity for either creating annotations from raw data or refining existing imperfect annotations within practice annotation scenarios, two dedicated strategies are proposed. "Prompt in 2D, detect in 3D" principle leverages straightforward 2D interactions, like clicks or boxes, to streamline the transition between 2D images and 3D object annotations. "Detect in 3D, refine in 3D" strategy further endows our model with self-correction capability. Evaluations on the nuScenes dataset validate our approach's superiority. Beyond a mere advancement, our work can serve as a cornerstone for future 3D object detection endeavors. + +# Acknowledgements + +This work was supported in part by the National Natural Science Foundation of China (NO. 62322608), in part by the Fundamental Research Funds for the Central Universities under Grant 22lgqb25, and in part by the Open Project Program of the Key Laboratory of Artificial Intelligence for Perception and Understanding, Liaoning Province (AIPU, No. 20230003). + +# References + +1. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1090-1099 (2022) +2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11621-11631 (2020) +3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European Conference on Computer Vision. pp. 213-229. Springer (2020) +4. Chen, X., Zhao, Z., Zhang, Y., Duan, M., Qi, D., Zhao, H.: Focalclick: Towards practical interactive image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1300-1309 (2022) +5. Chen, X., Ma, H., Wan, J., Li, B., Xia, T.: Multi-view 3d object detection network for autonomous driving. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1907-1915 (2017) +6. Chen, Y., Chen, Q., Sun, P., Chen, S., Wang, J., Cheng, J.: Enhancing your trained detrs with box refinement. arXiv preprint arXiv:2307.11828 (2023) +7. Choi, D., Cho, W., Kim, K., Choo, J.: idet3d: Towards efficient interactive object detection for lidar point clouds. arXiv preprint arXiv:2312.15449 (2023) +8. Ge, C., Chen, J., Xie, E., Wang, Z., Hong, L., Lu, H., Li, Z., Luo, P.: Metabev: Solving sensor failures for bev detection and map segmentation. arXiv preprint arXiv:2304.09801 (2023) +9. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition. pp. 3354-3361. IEEE (2012) +0. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2961-2969 (2017) +1. Huang, J., Huang, G.: Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054 (2022) +2. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021) +13. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) + +14. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12697-12705 (2019) +15. Lee, C., Park, S., Song, H., Ryu, J., Kim, S., Kim, H., Pereira, S., Yoo, D.: Interactive multi-class tiny-object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14136-14145 (2022) +16. Li, Y., Chen, Y., Qi, X., Li, Z., Sun, J., Jia, J.: Unifying voxel-based representation with transformer for 3d object detection. Advances in Neural Information Processing Systems 35, 18442-18455 (2022) +17. Li, Z., Wang, F., Wang, N.: Lidar r-cnn: An efficient and universal 3d object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7546-7555 (2021) +18. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European Conference on Computer Vision. pp. 1-18. Springer (2022) +19. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2980-2988 (2017) +20. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329 (2022) +21. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. In: European Conference on Computer Vision. pp. 531-548. Springer (2022) +22. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X., Sun, J.: Petrov2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256 (2022) +23. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2774-2781. IEEE (2023) +24. Lu, Y., Ma, X., Yang, L., Zhang, T., Liu, Y., Chu, Q., Yan, J., Ouyang, W.: Geometry uncertainty projection network for monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3111-3121 (2021) +25. Maninis, K.K., Caelles, S., Pont-Tuset, J., Van Gool, L.: Deep extreme cut: From extreme points to object segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 616-625 (2018) +26. Meng, D., Chen, X., Fan, Z., Zeng, G., Li, H., Yuan, Y., Sun, L., Wang, J.: Conditional detr for fast training convergence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3651-3660 (2021) +27. Meng, Q., Wang, W., Zhou, T., Shen, J., Jia, Y., Van Gool, L.: Towards a weakly supervised framework for 3d point cloud object detection and annotation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(8), 4454-4468 (2021) +28. Papadopoulos, D.P., Uijlings, J.R., Keller, F., Ferrari, V.: Extreme clicking for efficient object annotation. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4930-4939 (2017) + +29. Papadopoulos, D.P., Uijlings, J.R., Keller, F., Ferrari, V.: Training object class detectors with click supervision. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6374-6383 (2017) +30. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 194-210. Springer (2020) +31. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017) +32. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems 28 (2015) +33. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-779 (2019) +34. Shin, K., Kwon, Y.P., Tomizuka, M.: Roarnet: A robust 3d object detection based on region approximation refinement. In: 2019 IEEE Intelligent Vehicles Symposium (IV). pp. 2510-2515. IEEE (2019) +35. Sofiuk, K., Petrov, I.A., Konushin, A.: Reviving iterative training with mask guidance for interactive segmentation. In: 2022 IEEE International Conference on Image Processing (ICIP). pp. 3141-3145. IEEE (2022) +36. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2446-2454 (2020) +37. Wang, B., Wu, V., Wu, B., Keutzer, K.: Latte: accelerating lidar point cloud annotation via sensor fusion, one-click annotation, and tracking. In: 2019 IEEE Intelligent Transportation Systems Conference (ITSC). pp. 265-272. IEEE (2019) +38. Wang, K., Zhuang, J., Li, G., Fang, C., Cheng, L., Lin, L., Zhou, F.: De-biased teacher: Rethinking iou matching for semi-supervised object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2573-2580 (2023) +39. Wang, T., Zhu, X., Pang, J., Lin, D.: Fcos3d: Fully convolutional one-stage monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 913-922 (2021) +40. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2567-2575 (2022) +41. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning. pp. 180–191. PMLR (2022) +42. Xie, Y., Xu, C., Rakotosaona, M.J., Rim, P., Tombari, F., Keutzer, K., Tomizuka, M., Zhan, W.: Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection. arXiv preprint arXiv:2304.14340 (2023) +43. Xu, M., Zhang, Z., Hu, H., Wang, J., Wang, L., Wei, F., Bai, X., Liu, Z.: End-to-end semi-supervised object detection with soft teacher. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3060-3069 (2021) +44. Xu, N., Price, B., Cohen, S., Yang, J., Huang, T.S.: Deep interactive object selection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 373-381 (2016) + +45. Yan, J., Liu, Y., Sun, J., Jia, F., Li, S., Wang, T., Zhang, X.: Cross modal transformer via coordinates encoding for 3d object detection. arXiv preprint arXiv:2301.01283 (2023) +46. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018) +47. Yang, J., Zeng, A., Li, F., Liu, S., Zhang, R., Zhang, L.: Neural interactive keypoint detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15122-15132 (2023) +48. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3dssd: Point-based 3d single stage object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11040-11048 (2020) +49. Yao, A., Gall, J., Leistner, C., Van Gool, L.: Interactive object detection. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition. pp. 3242-3249. IEEE (2012) +50. Zhang, J., Lin, X., Zhang, W., Wang, K., Tan, X., Han, J., Ding, E., Wang, J., Li, G.: Semi-detr: Semi-supervised object detection with detection transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23809-23818 (2023) +51. Zhang, R., Qiu, H., Wang, T., Guo, Z., Xu, X., Qiao, Y., Gao, P., Li, H.: Monodetr: depth-guided transformer for monocular 3d object detection. arXiv preprint arXiv:2203.13310 (2022) +52. Zhao, J., Sun, L., Li, Q.: Recursivelydet: End-to-end region-based recursive object detection. arXiv preprint arXiv:2307.13619 (2023) +53. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4490-4499 (2018) +54. Zhou, Y., Zhu, H., Liu, Q., Chang, S., Guo, M.: Monoatt: Online monocular 3d object detection with adaptive token transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17493-17503 (2023) \ No newline at end of file diff --git a/2024/Interactive 3D Object Detection with Prompts/images.zip b/2024/Interactive 3D Object Detection with Prompts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b5da48129e2126915007351bbecfd838c16aa78f --- /dev/null +++ b/2024/Interactive 3D Object Detection with Prompts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e629dc6e8740a91d263051fbde894a2813a491574d43b1eb566ed5903952fb25 +size 400892 diff --git a/2024/Interactive 3D Object Detection with Prompts/layout.json b/2024/Interactive 3D Object Detection with Prompts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8ede5fb0582a09b9c7672e51a99ca234fdb6c145 --- /dev/null +++ b/2024/Interactive 3D Object Detection with Prompts/layout.json @@ -0,0 +1,11425 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 112, + 471, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 112, + 471, + 131 + ], + "spans": [ + { + "bbox": [ + 143, + 112, + 471, + 131 + ], + "type": "text", + "content": "Interactive 3D Object Detection with Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "spans": [ + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "type": "text", + "content": "Ruifei Zhang" + }, + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{1,2\\star}" + }, + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "type": "text", + "content": ", Xiangru Lin" + }, + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{4\\star}" + }, + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "type": "text", + "content": ", Wei Zhang" + }, + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "type": "text", + "content": ", Jincheng Lu" + }, + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "type": "text", + "content": ", Xuekuan Wang" + }, + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "type": "text", + "content": ", Xiao Tan" + }, + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "type": "text", + "content": ", Yingying Li" + }, + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "type": "text", + "content": ", Errui Ding" + }, + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "type": "text", + "content": ", Jingdong Wang" + }, + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "type": "text", + "content": ", and Guanbin Li" + }, + { + "bbox": [ + 134, + 149, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{3,5\\star\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 187, + 183, + 426, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 183, + 426, + 195 + ], + "spans": [ + { + "bbox": [ + 187, + 183, + 426, + 195 + ], + "type": "text", + "content": "1 The Chinese University of Hong Kong, Shenzhen, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 182, + 195, + 431, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 195, + 431, + 205 + ], + "spans": [ + { + "bbox": [ + 182, + 195, + 431, + 205 + ], + "type": "text", + "content": "2 Shenzhen Research Institute of Big Data, Shenzhen, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 214, + 205, + 399, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 205, + 399, + 217 + ], + "spans": [ + { + "bbox": [ + 214, + 205, + 399, + 217 + ], + "type": "text", + "content": "3 Sun Yat-sen University, Guangzhou, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 217, + 452, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 217, + 452, + 228 + ], + "spans": [ + { + "bbox": [ + 160, + 217, + 452, + 228 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 160, + 217, + 452, + 228 + ], + "type": "text", + "content": " Department of Computer Vision Technology (VIS), Baidu Inc., China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 253, + 228, + 361, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 228, + 361, + 239 + ], + "spans": [ + { + "bbox": [ + 253, + 228, + 361, + 239 + ], + "type": "text", + "content": "Peng Cheng Laboratory" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 171, + 239, + 441, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 239, + 441, + 251 + ], + "spans": [ + { + "bbox": [ + 171, + 239, + 441, + 251 + ], + "type": "text", + "content": "ruifeizhang@link.cuhk.edu.cn, liguanbin@mail.sysu.edu.cn," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 171, + 251, + 442, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 251, + 442, + 262 + ], + "spans": [ + { + "bbox": [ + 171, + 251, + 442, + 262 + ], + "type": "text", + "content": "{linxiangru,zhangwei99,lujincheng01,wangxuekuan,tanxiao01," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 197, + 262, + 416, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 262, + 416, + 272 + ], + "spans": [ + { + "bbox": [ + 197, + 262, + 416, + 272 + ], + "type": "text", + "content": "liyingying05,dingerrui,wangjingdong}@baidu.com" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 159, + 300, + 455, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 300, + 455, + 498 + ], + "spans": [ + { + "bbox": [ + 159, + 300, + 455, + 498 + ], + "type": "text", + "content": "Abstract. The evolution of 3D object detection hinges not only on advanced models but also on effective and efficient annotation strategies. Despite this progress, the labor-intensive nature of 3D object annotation remains a bottleneck, hindering further development in the field. This paper introduces a novel approach, incorporated with \"prompt in 2D, detect in 3D\" and \"detect in 3D, refine in 3D\" strategies, to 3D object annotation: multi-modal interactive 3D object detection. Firstly, by allowing users to engage with simpler 2D interaction prompts (e.g., clicks or boxes on a camera image or a bird's eye view), we bridge the complexity gap between 2D and 3D spaces, reimagining the annotation workflow. Besides, Our framework also supports flexible iterative refinement to the initial 3D annotations, further assisting annotators in achieving satisfying results. Evaluation on the nuScenes dataset demonstrates the effectiveness of our method. And thanks to the prompt-driven and interactive designs, our approach also exhibits outstanding performance in open-set scenarios. This work not only offers a potential solution to the 3D object annotation problem but also paves the way for further innovations in the 3D object detection community." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 160, + 508, + 426, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 508, + 426, + 519 + ], + "spans": [ + { + "bbox": [ + 160, + 508, + 426, + 519 + ], + "type": "text", + "content": "Keywords: Interactive 3D Object Detection " + }, + { + "bbox": [ + 160, + 508, + 426, + 519 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 508, + 426, + 519 + ], + "type": "text", + "content": " Prompt Learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 539, + 230, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 539, + 230, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 539, + 230, + 552 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 563, + 482, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 637 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 637 + ], + "type": "text", + "content": "3D object detection is pivotal for autonomous driving, gaining significant traction in recent years. Current research often derives 3D object information from sources such as monocular [24, 39, 51, 54] or multi-view camera images [11, 12, 18, 21], point cloud [17, 46, 53], and multi-modal sensors [1, 5, 45]. This has led to remarkable progress in detection performance. A major catalyst behind these advancements is the availability of high-quality 3D datasets like KITTI [9]," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 643, + 223, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 223, + 653 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 223, + 653 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 654, + 235, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 654, + 235, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 654, + 235, + 665 + ], + "type": "text", + "content": "\\*\\* Corresponding author." + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 169, + 125, + 445, + 174 + ], + "blocks": [ + { + "bbox": [ + 254, + 115, + 376, + 123 + ], + "lines": [ + { + "bbox": [ + 254, + 115, + 376, + 123 + ], + "spans": [ + { + "bbox": [ + 254, + 115, + 376, + 123 + ], + "type": "text", + "content": "Traditional Way of 3D Detection Data Annotation" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 169, + 125, + 445, + 174 + ], + "lines": [ + { + "bbox": [ + 169, + 125, + 445, + 174 + ], + "spans": [ + { + "bbox": [ + 169, + 125, + 445, + 174 + ], + "type": "image", + "image_path": "d7c34366f328e5c3679cc1fce84e33fa557cbd43240bb26a903fc1f25d9c43cf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 175, + 380, + 182 + ], + "lines": [ + { + "bbox": [ + 251, + 175, + 380, + 182 + ], + "spans": [ + { + "bbox": [ + 251, + 175, + 380, + 182 + ], + "type": "text", + "content": "Our proposed Way of 3D Detection Data Annotation" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 169, + 183, + 444, + 260 + ], + "blocks": [ + { + "bbox": [ + 169, + 183, + 444, + 260 + ], + "lines": [ + { + "bbox": [ + 169, + 183, + 444, + 260 + ], + "spans": [ + { + "bbox": [ + 169, + 183, + 444, + 260 + ], + "type": "image", + "image_path": "e778823b744d52fa4b4764e650138b8d133fd6521d8cb6e1b56110d5cf50896a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 270, + 482, + 315 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 315 + ], + "type": "text", + "content": "Fig. 1: We introduce an interactive 3D object detection concept. Rather than the traditional method where a model produces 3D object detection boxes followed by user refinement in 3D space, our strategy is \"prompt in 2D, detect in 3D, refine in 3D\". This method streamlines the annotation process, reducing the need for user refinements." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 340, + 482, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 412 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 412 + ], + "type": "text", + "content": "nuScenes [2], and Waymo [36]. These datasets fuel the progression of data-driven algorithms in 3D object detection. However, the intricate nature of 3D object annotation often necessitates significant human resources and incurs high labor costs. On average, annotating an hour of driving data takes hundreds of hours [27]. This annotation challenge potentially restricts the scale and diversity of datasets, possibly hindering further progression in this field." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 413, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 496 + ], + "type": "text", + "content": "Addressing this annotation challenge is the central theme of our paper. While weakly-supervised [25, 27-29] and semi-supervised [38, 43, 50] object detection methods exist, they typically suffer from reduced performance, especially in practical scenarios. This does not substantially address the overarching issue of the data-hungry nature of the field. With the recent emergence of SAM [13], both industry and academia are now recognizing the potential of interactive annotation approaches." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 604 + ], + "type": "text", + "content": "Our work aims to integrate this concept into 3D object detection, thus simplifying the annotation process, as presented in Fig. 1. Specifically, we observe that within the traditional process of annotating 3D detection data, annotators tend to concentrate on addressing two challenges stemming from imperfect pretrained 3D detectors: 1) supplementing missing closed-set objects and incorporating additional required open-set ones; 2) rectifying existing erroneous annotations. Both scenarios necessitate annotators to operate within the 3D point cloud space, which is extremely time-consuming and labor-intensive." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "To circumvent this, we create an interactive 3D object detection system enhanced by two strategies: prompt in " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "2D" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", detect in " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "3D" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", and detect in " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "3D" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", refine in " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "3D" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ". The first principle enables users to interact with a 2D camera image or bird's eye view (BEV) of the point cloud, prompting the model to generate corresponding 3D bounding boxes, instead of encountering the 3D point cloud" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "R. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "type": "text", + "content": "directly. Furthermore, benefiting from the prompt-driven annotation paradigm, our method inherently possesses the capability to annotate open-set objects. The latter ensures annotators refine the initial 3D bounding box at the fastest pace with minimal effort, which significantly streamlines operations in the 3D space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 342 + ], + "type": "text", + "content": "In light of these principles, we present our multi-modal interactive 3D object detection task and algorithm. Firstly, it employs prompts in the form of clicks or boxes on a 2D camera or BEV view, guiding the model to generate the relevant 3D bounding box. Our framework is reminiscent of the DETR-series detectors [3, 20, 26, 40] but comes with distinctive features. We integrate user inputs as prompt tokens and design a learnable component that investigates multi-modal features of indicated objects, informed by the prompts. Furthermore, we incorporate a global-to-local strategy to supply fine-grained object features, enhancing detection accuracy. Secondly, we empower our model with iterative refinement capability by incorporating the predicted 3D bounding box encoding as a new prompt into the query sequence. For those inaccurate 3D predictions, our framework allows annotators to adjust a few wrong attributes, e.g. dragging the center point position or adjusting the orientation angle of the 3D bounding box, and then perform self-correction on other characters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 345, + 480, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 480, + 427 + ], + "type": "text", + "content": "To rigorously evaluate our approach, we devise two evaluation configurations for practical annotation scenarios. The first, regarding Annotation from Raw Data, leverages user 2D prompts and produces corresponding 3D predictions. Besides, our framework is also compatible with starting from the results of existing State-of-the-Art (SOTA) 3D detection models, termed as Annotation from SOTA Model and iteratively refining them. Experimental results on the nuScenes dataset highlight the effectiveness of our method in various scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 429, + 392, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 429, + 392, + 440 + ], + "spans": [ + { + "bbox": [ + 146, + 429, + 392, + 440 + ], + "type": "text", + "content": "In summary, this paper has the following contributions:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 450, + 479, + 555 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 450, + 479, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 450, + 479, + 495 + ], + "spans": [ + { + "bbox": [ + 138, + 450, + 479, + 495 + ], + "type": "text", + "content": "- We introduce a new concept of multi-modal interactive 3D object detection, proposing a substantial solution to prevailing annotation challenges. Our work is expected to provide critical insights for the 3D object detection community across academic and industrial domains." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 498, + 479, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 498, + 479, + 531 + ], + "spans": [ + { + "bbox": [ + 137, + 498, + 479, + 531 + ], + "type": "text", + "content": "- We unveil a prompt-driven interactive 3D object detection methodology with self-correction capability, leveraging it to establish an interactive system tailored for practical annotation scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 535, + 479, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 535, + 479, + 555 + ], + "spans": [ + { + "bbox": [ + 137, + 535, + 479, + 555 + ], + "type": "text", + "content": "- Comprehensive experiments in both annotation scenarios and evaluation settings attest to our method's efficacy." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 579, + 237, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 237, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 237, + 592 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 607, + 265, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 607, + 265, + 619 + ], + "spans": [ + { + "bbox": [ + 131, + 607, + 265, + 619 + ], + "type": "text", + "content": "2.1 3D Object Detection" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "3D object detection techniques have been developed across a range of input modalities. Our approach emphasizes the advantages of interactive inputs, setting it apart from traditional methods:" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 257, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 257, + 91, + 447, + 102 + ], + "type": "text", + "content": "Interactive 3D Object Detection with Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 481, + 186 + ], + "type": "text", + "content": "Camera-only. Monocular 3D object detection techniques [24, 39, 51, 54] enjoy advantages such as low cost and ease of deployment. However, it also faces the inherent challenge of limited depth information. Multi-view methods utilize data from multiple camera perspectives [11, 12, 18, 21, 22, 30, 41], which enhances the model's ability to perceive the 3D environment, thereby improving detection accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 481, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 481, + 236 + ], + "type": "text", + "content": "LiDAR-only. Benefiting from the robustness and high precision, numerous studies utilize LiDAR data for 3D object detection. Point-based methods [17, 33,48] like PointNet [31] directly engage with raw LiDAR data, while grid-based strategies structure the data into 3D voxels [46,53] or feature pillars [14]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 481, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 481, + 308 + ], + "type": "text", + "content": "Multi-modality. Capitalizing on the strengths of camera images and point clouds, multi-modal 3D object detection has seen significant advancements [1, 5, 45]. Early methods combine camera images and point clouds at the input or result stages [5, 34]. While, current models, like TransFusion [1] and CMT [45], emphasize more intricate feature-level fusion. Uniquely, our method embeds interactive capacities, ensuring richer, user-tailored annotations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 324, + 379, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 324, + 379, + 337 + ], + "spans": [ + { + "bbox": [ + 131, + 324, + 379, + 337 + ], + "type": "text", + "content": "2.2 Interactive Object Detection/Segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 342, + 481, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 481, + 366 + ], + "type": "text", + "content": "Interactive methods have historically aimed to reduce the annotation workload while maintaining, or even enhancing, model performance:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 367, + 481, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 481, + 450 + ], + "type": "text", + "content": "Interactive Object Detection. Early attempts, like the incremental learning strategy [49], require user corrections to update detectors. More recent models like C3Det [15] have substantially eased the annotation process. Tools such as LATTE [37] and iDet3D [7] further simplify the 3D object annotation process. Our framework, however, brings in the versatility of prompt formats and modalities, offering users a broadened interaction spectrum coupled with robust detection capabilities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 451, + 481, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 481, + 534 + ], + "type": "text", + "content": "Interactive Object Segmentation. Interactive segmentation has matured significantly over the years [4,35,44]. Models like SAM [13] have empowered users with tools like clicks, boxes, and referring expressions to guide the segmentation process. Diverging from interactive segmentation confined to 2D spaces, our work addresses the challenge of integrating 2D interactive prompts with 3D object detection, mitigating the spatial discrepancy between 2D annotations and 3D environments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 551, + 345, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 551, + 345, + 563 + ], + "spans": [ + { + "bbox": [ + 131, + 551, + 345, + 563 + ], + "type": "text", + "content": "2.3 Weakly-supervised Object Detection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "The notion of harnessing weak signals, like clicks [29] or extreme-points [25, 28], for supervising detectors has been a topic of intrigue. Some methods [27] have extended this idea to 3D object detection, using center-click based strategies. These approaches, while commendable for their balance between performance and annotation overhead, sometimes fail to harness the full potential of the detectors. Our methodology, on the other hand, incorporates strong 3D bounding box supervisions in an interactive setting, streamlining the data acquisition process without compromising on detection performance." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "R. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 115, + 473, + 219 + ], + "blocks": [ + { + "bbox": [ + 143, + 115, + 473, + 219 + ], + "lines": [ + { + "bbox": [ + 143, + 115, + 473, + 219 + ], + "spans": [ + { + "bbox": [ + 143, + 115, + 473, + 219 + ], + "type": "image", + "image_path": "711e38e5001e2e2f615a72da8466f938568c45427604ddda6d7dacce6293f55d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "type": "text", + "content": "Fig. 2: An overview of our proposed interactive 3D object detection system/method. Taking multi-modal data with user prompts as input, we explicitly add 3D information " + }, + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{im},\\mathbf{P}_{pc}" + }, + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "type": "text", + "content": " into features " + }, + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{im},\\mathbf{F}_{pc}" + }, + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "type": "text", + "content": " by 3D Position Encoding. Object queries consist of a learnable token " + }, + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_l" + }, + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "type": "text", + "content": " and prompt ones " + }, + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_p" + }, + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "type": "text", + "content": ", which are generated by Prompt-Guided Query generation. We also introduce a Global-to-Local enhancement strategy to dynamically adjust features integrated with queries from global representations to local RoI details. Besides, our framework allows users to refine any of 3D attributes directly, and encode the refined 3D bounding box " + }, + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "type": "inline_equation", + "content": "b_{L}^{r^{\\prime}}" + }, + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "type": "text", + "content": " and category information " + }, + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "type": "inline_equation", + "content": "c_{L}^{r^{\\prime}}" + }, + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "type": "text", + "content": ", along with attribute refinement record " + }, + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "type": "inline_equation", + "content": "p_a^r" + }, + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "type": "inline_equation", + "content": "p_c^r" + }, + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "type": "text", + "content": ", into a new prompt token " + }, + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{3D}" + }, + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "type": "text", + "content": " by 3D Attribute Encoding. " + }, + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "type": "inline_equation", + "content": "g,w,l,h,\\theta" + }, + { + "bbox": [ + 130, + 228, + 482, + 361 + ], + "type": "text", + "content": " denotes the gravity center, width, length, height and yaw angle respectively. This iterative refinement propels our model to automatically rectify other attributes, ultimately delivering satisfactory results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 382, + 233, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 382, + 233, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 382, + 233, + 396 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 406, + 257, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 406, + 257, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 406, + 257, + 418 + ], + "type": "text", + "content": "3.1 Problem Definition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "text", + "content": "We present an interactive multi-modal 3D object detection algorithm. Given a multi-view set of camera images and associated point cloud data, the objective is to predict both the class labels and the 3D bounding boxes, driven by user prompts either in the form of a point or a box. Further, our model also supports iteratively self-correction with limited human assistance. To supervise the iteration process, the Ground Truth " + }, + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "inline_equation", + "content": "(c, b)" + }, + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "text", + "content": " is employed. Here, " + }, + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "text", + "content": " signifies the category label, and " + }, + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "text", + "content": " represents the 3D attributes. We adopt a deep supervision strategy for our multi-layer Transformer decoder, which guides the predictions across each decoder layer of each iteration, denoted as " + }, + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\{(c_i^r, b_i^r), i = 1, 2, \\dots, L, r = 0, 1, \\dots, M\\}" + }, + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 426, + 482, + 547 + ], + "type": "text", + "content": " being the total number of decoder layers and iterative refinements." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 562, + 208, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 208, + 573 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 208, + 573 + ], + "type": "text", + "content": "3.2 Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "The architecture underpinning our interactive 3D object detection method is visualized in Fig. 2, which could be roughly decomposed into two stages, i.e. \"prompt in " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "2D" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ", detect in " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "3D" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " and \"detect in " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "3D" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ", refine in " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "3D" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "\". In the first phase, two separate backbones are utilized to extract multi-view image features " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{im} = \\{\\mathbf{F}_{im}^{t}, t = 1,2,\\dots,T\\}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " and point cloud features " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{pc}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " respectively. To enhance these features with 3D-awareness, following PETR [21] and CMT [45], we introduce 3D position encoding. Specifically, for the multi-view images, this encoding is" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 256, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 256, + 91, + 448, + 102 + ], + "type": "text", + "content": "Interactive 3D Object Detection with Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "represented as " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{im} = \\{\\mathbf{P}_{im}^t, t = 1,2,\\dots,T\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", and for the point cloud, it is denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{pc}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ". For object queries, denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", we utilize a unique blending of a learnable token, " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_l" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", and prompt-specific tokens, " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_p" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", for every detected object. The prompt queries, " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_p" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", play a pivotal role in guiding the learnable token, " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_l" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", by leveraging the insights from the multi-modal features. Furthermore, to enhance the detection accuracy across various object scales, we incorporate a global-to-local strategy within the decoder. This strategy ensures that the object queries are continually enriched by granular local features, enabling finer detections. Ultimately, the learnable token predicts the object's class and its 3D bounding box " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "(c_L^0,b_L^0)" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "text", + "content": "Subsequently, the model enters an interactive refinement stage. This stage reuses the multi-modal features but updates the object queries " + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "text", + "content": " by inserting the encoding of the refined 3D prediction from the current iteration. Specifically, in the " + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "text", + "content": "-th iteration, we enable annotators to fine-tune the current prediction " + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "inline_equation", + "content": "(c_{L}^{r}, b_{L}^{r})" + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "text", + "content": " and encode the refined result " + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "inline_equation", + "content": "(c_{L}^{r'}, b_{L}^{r'})" + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "text", + "content": " and corresponding attribute refinement record " + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "inline_equation", + "content": "(p_{c}^{r}, p_{a}^{r})" + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "text", + "content": " into a new prompt " + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{3D}" + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{3D}" + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_p" + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "text", + "content": " collectively drive " + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_l" + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "text", + "content": " to generate more accurate predictions in the next iteration. Note that we omit the iteration superscript of " + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{3D}" + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 130, + 236, + 482, + 332 + ], + "type": "text", + "content": " for simplicity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 348, + 272, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 272, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 272, + 361 + ], + "type": "text", + "content": "3.3 3D Position Encoding" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 367, + 482, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 463 + ], + "type": "text", + "content": "Multi-view Camera Images. The core idea of embedding 3D position for multi-view image features includes two steps: 1) modeling each pixel coordinate as a series of points along a ray in the camera frustum space, and 2) projecting camera frustum coordinates into 3D space. Specifically, for a given camera image feature, denoted as " + }, + { + "bbox": [ + 130, + 367, + 482, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{im}^{t}\\in \\mathbb{R}^{W_{im}\\times H_{im}\\times C}" + }, + { + "bbox": [ + 130, + 367, + 482, + 463 + ], + "type": "text", + "content": ", each pixel can be visualized as tracing a ray within the camera frustum space. This perspective allows us to sample " + }, + { + "bbox": [ + 130, + 367, + 482, + 463 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 367, + 482, + 463 + ], + "type": "text", + "content": " discrete points along the depth dimension. This process is mathematically captured by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 197, + 472, + 480, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 472, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 197, + 472, + 480, + 485 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {c f} (u, v, j) = \\left(u \\times d _ {j}, v \\times d _ {j}, d _ {j}, 1\\right) ^ {T}, j = 1, 2, \\dots , D \\tag {1}", + "image_path": "3c8609fcf36a68c309febf1c8742086d9ea2743770df16fe2045b52f4ccbcc49.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 492, + 481, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 481, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 481, + 516 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 492, + 481, + 516 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{cf}" + }, + { + "bbox": [ + 130, + 492, + 481, + 516 + ], + "type": "text", + "content": " indicates the camera's frustum coordinates. Pixel coordinates in the image are denoted by " + }, + { + "bbox": [ + 130, + 492, + 481, + 516 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 130, + 492, + 481, + 516 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 492, + 481, + 516 + ], + "type": "inline_equation", + "content": "d_{j}" + }, + { + "bbox": [ + 130, + 492, + 481, + 516 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 130, + 492, + 481, + 516 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 492, + 481, + 516 + ], + "type": "text", + "content": "-th depth value." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 517, + 481, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 481, + 539 + ], + "type": "text", + "content": "To relate these coordinates to a 3D context, we employ an inverse 3D projection method:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 231, + 540, + 480, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 540, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 231, + 540, + 480, + 554 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {3 d} ^ {t} (u, v, j) = \\mathbf {K} _ {t} ^ {3 D} \\mathbf {K} _ {t} ^ {- 1} \\mathbf {P} _ {c f} (u, v, j) \\tag {2}", + "image_path": "f4584b8f3c7b9eced6d90ff3442eb2a5a600e19ef7c6fbe6de1ff88e804e2c00.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_t^{3D} \\in \\mathbb{R}^{4 \\times 4}" + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "text", + "content": " represents the transformation matrix that links 3D space with the " + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "text", + "content": "-th camera's coordinate space. " + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_t \\in \\mathbb{R}^{4 \\times 4}" + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "text", + "content": " is the intrinsic matrix of " + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "text", + "content": "-th camera. To ensure consistent value scales, we normalize the 3D coordinates to the range [0,1]. The final 3D position encoding for the " + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "text", + "content": "-th camera image, represented as " + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{im}^t \\in \\mathbb{R}^{W_{im} \\times H_{im} \\times C}" + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "text", + "content": ", is derived via a multi-layer perceptron (MLP): " + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{im}^t = \\mathrm{MLP}_{im}(\\mathbf{P}_{3d}^t)" + }, + { + "bbox": [ + 130, + 557, + 482, + 630 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Point Cloud. In contrast to the approach for camera image features, where points are assembled along the depth axis, the LiDAR point set necessitates sampling along the height axis. Specifically, given a BEV feature map, denoted as" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "R. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{pc} \\in \\mathbb{R}^{W_{pc} \\times H_{pc} \\times C}" + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": ", we can compute the corresponding 3D position information, represented as " + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{pc} \\in \\mathbb{R}^{W_{pc} \\times H_{pc} \\times C}" + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": ", using the following steps: First, we determine the 3D position based on the feature map's coordinates:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 195, + 159, + 481, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 159, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 195, + 159, + 481, + 172 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {3 d} (u, v, j) = \\left(u \\times u _ {s}, v \\times v _ {s}, h _ {j}, 1\\right) ^ {T}, j = 1, 2, \\dots , H \\tag {3}", + "image_path": "237e8b1400879ef89f4248c890767d9c0e33d8c813b0af3121ccc595cb8ccab5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 179, + 480, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 179, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 179, + 480, + 275 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 179, + 480, + 275 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 130, + 179, + 480, + 275 + ], + "type": "text", + "content": " specifies the coordinate within the point cloud feature map. The variables " + }, + { + "bbox": [ + 130, + 179, + 480, + 275 + ], + "type": "inline_equation", + "content": "(u_{s},v_{s})" + }, + { + "bbox": [ + 130, + 179, + 480, + 275 + ], + "type": "text", + "content": " represent the dimensions of the feature grid. " + }, + { + "bbox": [ + 130, + 179, + 480, + 275 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 179, + 480, + 275 + ], + "type": "text", + "content": " is the number of sampled points along the height axis and " + }, + { + "bbox": [ + 130, + 179, + 480, + 275 + ], + "type": "inline_equation", + "content": "h_j" + }, + { + "bbox": [ + 130, + 179, + 480, + 275 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 179, + 480, + 275 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 179, + 480, + 275 + ], + "type": "text", + "content": "-th height value. Next, we utilize a multi-layer perceptron (MLP) to compute the 3D position encoding from the derived 3D position: " + }, + { + "bbox": [ + 130, + 179, + 480, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{pc} = \\mathrm{MLP}_{pc}(\\mathbf{P}_{3d})" + }, + { + "bbox": [ + 130, + 179, + 480, + 275 + ], + "type": "text", + "content": ". Note that as in CMT, only one point is sampled along the height axis, thus simplifying the process into standard 2D position encoding. Consequently, in our methodology, we have adopted a sinusoidal embedding to cater to this behavior." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 291, + 331, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 291, + 331, + 303 + ], + "spans": [ + { + "bbox": [ + 131, + 291, + 331, + 303 + ], + "type": "text", + "content": "3.4 Prompt-guided Query Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 309, + 479, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 479, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 479, + 367 + ], + "type": "text", + "content": "Our approach provides an avenue for user interaction, distinct from traditional DETR-based detectors that often rely on multiple queries to discern image objects. By merging learnable tokens, " + }, + { + "bbox": [ + 130, + 309, + 479, + 367 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_l" + }, + { + "bbox": [ + 130, + 309, + 479, + 367 + ], + "type": "text", + "content": ", with user-directed prompt tokens, " + }, + { + "bbox": [ + 130, + 309, + 479, + 367 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_p" + }, + { + "bbox": [ + 130, + 309, + 479, + 367 + ], + "type": "text", + "content": ", we achieve object detection that is not only efficient but also offers a new angle on object localization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 369, + 480, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 480, + 429 + ], + "type": "text", + "content": "Learnable Query " + }, + { + "bbox": [ + 130, + 369, + 480, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_l" + }, + { + "bbox": [ + 130, + 369, + 480, + 429 + ], + "type": "text", + "content": ": Every object is associated with a unique learnable token that facilitates feature probing, and subsequent prediction of class labels and 3D bounding boxes. In line with prior research [20, 40], we conceptualize the learnable query as an anchor point, initiating it with the positional prior of the user prompt. The initialization process unfolds over three phases:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 429, + 479, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 479, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 479, + 465 + ], + "type": "text", + "content": "1). 2D Center Position Computation with User Prompt. A user might provide a bounding box " + }, + { + "bbox": [ + 130, + 429, + 479, + 465 + ], + "type": "inline_equation", + "content": "box = (x_1, y_1, x_2, y_2)" + }, + { + "bbox": [ + 130, + 429, + 479, + 465 + ], + "type": "text", + "content": " or a singular point " + }, + { + "bbox": [ + 130, + 429, + 479, + 465 + ], + "type": "inline_equation", + "content": "point = (x, y)" + }, + { + "bbox": [ + 130, + 429, + 479, + 465 + ], + "type": "text", + "content": ". Depending on the input type, the 2D center position " + }, + { + "bbox": [ + 130, + 429, + 479, + 465 + ], + "type": "inline_equation", + "content": "(c_x, c_y)" + }, + { + "bbox": [ + 130, + 429, + 479, + 465 + ], + "type": "text", + "content": " is determined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 210, + 472, + 480, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 472, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 210, + 472, + 480, + 503 + ], + "type": "interline_equation", + "content": "\\left(c _ {x}, c _ {y}\\right) = \\left\\{ \\begin{array}{l l} \\left(\\frac {x _ {1} + x _ {2}}{2}, \\frac {y _ {1} + y _ {2}}{2}\\right) & \\text {i f b o u n d i n g b o x} \\\\ (x, y) & \\text {i f p o i n t} \\end{array} \\right. \\tag {4}", + "image_path": "3f760048b675461232df83735828dd8e0567bfc7f2a1f292b27d9828dac5ad75.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 480, + 616 + ], + "type": "text", + "content": "2). 2D Center Transformation into 3D Coordinate. Given a user prompt on the BEV map, the 2D space naturally omits height data. We address this by designating " + }, + { + "bbox": [ + 130, + 510, + 480, + 616 + ], + "type": "inline_equation", + "content": "c_{z} = 0.5" + }, + { + "bbox": [ + 130, + 510, + 480, + 616 + ], + "type": "text", + "content": ", the median value of height, resulting in a 3D center labeled as center " + }, + { + "bbox": [ + 130, + 510, + 480, + 616 + ], + "type": "inline_equation", + "content": "= (c_{x},c_{y},c_{z})" + }, + { + "bbox": [ + 130, + 510, + 480, + 616 + ], + "type": "text", + "content": ". Alternatively, for a user prompt within 2D multi-view images, we transform the center coordinates " + }, + { + "bbox": [ + 130, + 510, + 480, + 616 + ], + "type": "inline_equation", + "content": "(c_{x},c_{y})" + }, + { + "bbox": [ + 130, + 510, + 480, + 616 + ], + "type": "text", + "content": " into a set of 3D points along the depth axis, using Eqs. (1) and (2). The median of the ray is then selected as the 3D center, represented as center " + }, + { + "bbox": [ + 130, + 510, + 480, + 616 + ], + "type": "inline_equation", + "content": "= (c_x',c_y',c_z')" + }, + { + "bbox": [ + 130, + 510, + 480, + 616 + ], + "type": "text", + "content": ". The 3D center serves as a reference point for bounding box prediction and we prioritize the center from BEV if the user provides prompts in both modalities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 480, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 667 + ], + "type": "text", + "content": "3). 3D Position Encoding Extraction from Both Modalities. The 3D center is projected onto both the camera image and BEV plane, yielding their 3D position encodings " + }, + { + "bbox": [ + 130, + 617, + 480, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{im} \\in \\mathbb{R}^{1 \\times C}" + }, + { + "bbox": [ + 130, + 617, + 480, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 480, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{pc} \\in \\mathbb{R}^{1 \\times C}" + }, + { + "bbox": [ + 130, + 617, + 480, + 667 + ], + "type": "text", + "content": ", respectively. The learnable query is initialized by combining information from both modalities: " + }, + { + "bbox": [ + 130, + 617, + 480, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_l = \\mathbf{Q}_{im} + \\mathbf{Q}_{pc}" + }, + { + "bbox": [ + 130, + 617, + 480, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 256, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 256, + 91, + 448, + 103 + ], + "type": "text", + "content": "Interactive 3D Object Detection with Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 177, + 116, + 436, + 208 + ], + "blocks": [ + { + "bbox": [ + 177, + 116, + 436, + 208 + ], + "lines": [ + { + "bbox": [ + 177, + 116, + 436, + 208 + ], + "spans": [ + { + "bbox": [ + 177, + 116, + 436, + 208 + ], + "type": "image", + "image_path": "63ec824553a1b3a32dc38049373b5970cda6f01738b8f24593fdf1941ceb0991.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 213, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 258 + ], + "type": "text", + "content": "Fig. 3: Illustration of Global-to-Local Enhancement. Our queries interact with unmodified multi-modal features and its object's local features respectively in the global and local phases. Self-Attention is confined within queries of identical objects in both stages." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 280, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 389 + ], + "type": "text", + "content": "Prompt Query " + }, + { + "bbox": [ + 130, + 280, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_p" + }, + { + "bbox": [ + 130, + 280, + 482, + 389 + ], + "type": "text", + "content": ": It is designed to embed user guidance by merging position encoding and context embedding: " + }, + { + "bbox": [ + 130, + 280, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_p = \\mathbf{Q}_{pe} + \\mathbf{Q}_{ce}" + }, + { + "bbox": [ + 130, + 280, + 482, + 389 + ], + "type": "text", + "content": ". This synthesis is constructed for both box and point prompt formats: For the box format, the box position is delineated using its center point. From our 3D position encoding, the position encoding for box center, " + }, + { + "bbox": [ + 130, + 280, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{pe} \\in \\mathbb{R}^{1 \\times C}" + }, + { + "bbox": [ + 130, + 280, + 482, + 389 + ], + "type": "text", + "content": ", is derived. The context embedding, " + }, + { + "bbox": [ + 130, + 280, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{ce} \\in \\mathbb{R}^{1 \\times C}" + }, + { + "bbox": [ + 130, + 280, + 482, + 389 + ], + "type": "text", + "content": ", is secured via average pooling of the feature region bounded by the prompt box. For the point format, the same technique determines the position encoding, " + }, + { + "bbox": [ + 130, + 280, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{pe} \\in \\mathbb{R}^{1 \\times C}" + }, + { + "bbox": [ + 130, + 280, + 482, + 389 + ], + "type": "text", + "content": ". The feature at the indicated prompt point gives the context embedding, " + }, + { + "bbox": [ + 130, + 280, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{ce} \\in \\mathbb{R}^{1 \\times C}" + }, + { + "bbox": [ + 130, + 280, + 482, + 389 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 403, + 310, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 403, + 310, + 415 + ], + "spans": [ + { + "bbox": [ + 131, + 403, + 310, + 415 + ], + "type": "text", + "content": "3.5 Global-to-Local Enhancement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 422, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 470 + ], + "type": "text", + "content": "Given multi-modal features " + }, + { + "bbox": [ + 130, + 422, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{pc}" + }, + { + "bbox": [ + 130, + 422, + 482, + 470 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 422, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{im}" + }, + { + "bbox": [ + 130, + 422, + 482, + 470 + ], + "type": "text", + "content": " with their respective 3D position embeddings " + }, + { + "bbox": [ + 130, + 422, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{pc}" + }, + { + "bbox": [ + 130, + 422, + 482, + 470 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 422, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{im}" + }, + { + "bbox": [ + 130, + 422, + 482, + 470 + ], + "type": "text", + "content": ", along with object queries " + }, + { + "bbox": [ + 130, + 422, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 130, + 422, + 482, + 470 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 422, + 482, + 470 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 422, + 482, + 470 + ], + "type": "text", + "content": " represents the cumulative number of queries from " + }, + { + "bbox": [ + 130, + 422, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_l" + }, + { + "bbox": [ + 130, + 422, + 482, + 470 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 422, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_p" + }, + { + "bbox": [ + 130, + 422, + 482, + 470 + ], + "type": "text", + "content": " across all objects, we employ a Transformer-based decoder to amalgamate these features for prediction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 470, + 481, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 470, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 131, + 470, + 481, + 482 + ], + "type": "text", + "content": "Self-Attention: The formulation for the decoder's self-attention mechanism is:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 235, + 492, + 481, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 492, + 481, + 507 + ], + "spans": [ + { + "bbox": [ + 235, + 492, + 481, + 507 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {i} ^ {\\prime} = \\operatorname {s o f t m a x} \\left(\\mathbf {M} _ {s a} + \\mathbf {Q} _ {i} \\mathbf {Q} _ {i} ^ {T}\\right) \\mathbf {Q} _ {i} \\tag {5}", + "image_path": "2103f91b66160d81bf2f39b53b92ae151600c7fe8e0534e089393a4b142dec2f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 546 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 546 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 510, + 482, + 546 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i" + }, + { + "bbox": [ + 130, + 510, + 482, + 546 + ], + "type": "text", + "content": " corresponds to the queries from the " + }, + { + "bbox": [ + 130, + 510, + 482, + 546 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 510, + 482, + 546 + ], + "type": "text", + "content": "-th layer, and " + }, + { + "bbox": [ + 130, + 510, + 482, + 546 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_0 = \\mathbf{Q}" + }, + { + "bbox": [ + 130, + 510, + 482, + 546 + ], + "type": "text", + "content": ". The attention mask " + }, + { + "bbox": [ + 130, + 510, + 482, + 546 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{sa}" + }, + { + "bbox": [ + 130, + 510, + 482, + 546 + ], + "type": "text", + "content": " ensures interaction is confined within queries of identical objects:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 178, + 552, + 481, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 552, + 481, + 585 + ], + "spans": [ + { + "bbox": [ + 178, + 552, + 481, + 585 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {s a} \\left(q _ {1}, q _ {2}\\right) = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} q _ {1}, q _ {2} \\text {b e l o n g t o t h e s a m e o b j e c t ,} \\\\ - \\infty , & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {6}", + "image_path": "d99c7fb40e8e69c640dfed4e95ced970b5b1840aa00c44d641a714900e1c0e3d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 590, + 447, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 590, + 447, + 602 + ], + "spans": [ + { + "bbox": [ + 131, + 590, + 447, + 602 + ], + "type": "text", + "content": "Cross-Attention: For cross-attention in the decoder, we express it as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 232, + 609, + 481, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 609, + 481, + 624 + ], + "spans": [ + { + "bbox": [ + 232, + 609, + 481, + 624 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {i + 1} = \\operatorname {s o f t m a x} (\\mathbf {M} _ {c a} + \\mathbf {Q} _ {i} ^ {\\prime} \\mathbf {K} ^ {T}) \\mathbf {V} \\tag {7}", + "image_path": "4336f383d4b048f619494b25c6e855c79ac7330644e5f09fb6297fb3a5b35e84.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " come from multi-modal features, " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ca}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " is attention mask for cross-attention. Influenced by traditional two-stage detectors [32] and recent DETR-centric techniques [6, 52], we adopt a global-to-local enhancement strategy." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "R. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "This approach adjusts " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " dynamically to augment our model's local perceptiveness, as illustrated in Fig. 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "type": "text", + "content": "For the initial " + }, + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "type": "text", + "content": " layers among a total of " + }, + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "type": "text", + "content": " decoder layers, queries access unmodified multi-modal features:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 224, + 173, + 480, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 173, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 224, + 173, + 480, + 186 + ], + "type": "interline_equation", + "content": "\\mathbf {K} = \\left[ f _ {k} \\left(\\mathbf {F} _ {p c} + \\mathbf {P} _ {p c}\\right): f _ {k} \\left(\\mathbf {F} _ {i m} + \\mathbf {P} _ {i m}\\right) \\right], \\tag {8}", + "image_path": "1485092b2f54f9bd81b71a7597da9ab46e1bb89ebf965bc23c9f7eea6d6a450b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 225, + 188, + 480, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 188, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 225, + 188, + 480, + 201 + ], + "type": "interline_equation", + "content": "\\mathbf {V} = \\left[ f _ {v} \\left(\\mathbf {F} _ {p c}\\right): f _ {v} \\left(\\mathbf {F} _ {i m}\\right) \\right], \\tag {9}", + "image_path": "cae6a3c38ded4001462fb34f0d0380ee3d96bd8673bd0c22bd7bd2402d5f8e32.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 209, + 479, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 479, + 245 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 479, + 245 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 209, + 479, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{K},\\mathbf{V}\\in \\mathbb{R}^{(H_{pc}W_{pc} + H_{im}W_{im}T)\\times C}" + }, + { + "bbox": [ + 130, + 209, + 479, + 245 + ], + "type": "text", + "content": " are original multi-modal features under transformation " + }, + { + "bbox": [ + 130, + 209, + 479, + 245 + ], + "type": "inline_equation", + "content": "f_{k}(\\cdot),f_{v}(\\cdot)" + }, + { + "bbox": [ + 130, + 209, + 479, + 245 + ], + "type": "text", + "content": " and concatenation " + }, + { + "bbox": [ + 130, + 209, + 479, + 245 + ], + "type": "inline_equation", + "content": "[\\cdot :\\cdot ]" + }, + { + "bbox": [ + 130, + 209, + 479, + 245 + ], + "type": "text", + "content": " . " + }, + { + "bbox": [ + 130, + 209, + 479, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ca} = 0" + }, + { + "bbox": [ + 130, + 209, + 479, + 245 + ], + "type": "text", + "content": " during the global cross-attention phase." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 246, + 479, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 479, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 479, + 282 + ], + "type": "text", + "content": "In the succeeding " + }, + { + "bbox": [ + 130, + 246, + 479, + 282 + ], + "type": "inline_equation", + "content": "L - Z" + }, + { + "bbox": [ + 130, + 246, + 479, + 282 + ], + "type": "text", + "content": " layers, local region-of-interest (RoI) features are extracted by RoI Align [10] via either the user-provided prompt box or the predicted 2D bounding box based on prompt point:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 200, + 291, + 480, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 291, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 200, + 291, + 480, + 304 + ], + "type": "interline_equation", + "content": "\\mathbf {K} = \\left[ f _ {k} \\left(\\operatorname {R o I} \\left(\\mathbf {F} _ {p c} + \\mathbf {P} _ {p c}\\right)\\right): f _ {k} \\left(\\operatorname {R o I} \\left(\\mathbf {F} _ {i m} + \\mathbf {P} _ {i m}\\right)\\right) \\right], \\tag {10}", + "image_path": "5d9b29857f88c2a82fd4d9d2f0f47d943ec41c121e09319bff89415b6f030a8e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 200, + 306, + 480, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 306, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 200, + 306, + 480, + 319 + ], + "type": "interline_equation", + "content": "\\mathbf {V} = \\left[ f _ {v} \\left(\\operatorname {R o I} \\left(\\mathbf {F} _ {p c}\\right)\\right): f _ {v} \\left(\\operatorname {R o I} \\left(\\mathbf {F} _ {i m}\\right)\\right) \\right], \\tag {11}", + "image_path": "086079e986df876dad243a747a5d081586259f8e5d48b5a369b32b585e645929.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{K},\\mathbf{V}\\in \\mathbb{R}^{(H_oW_o + H_oW_o)\\times C}" + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "text", + "content": " now signify local multi-modal features. " + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "inline_equation", + "content": "(H_{o},W_{o})" + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "text", + "content": " is the output size of RoI Align. With " + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ca}" + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "text", + "content": " defined similarly to " + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{sa}" + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "text", + "content": ", it ensures each query is exclusively attentive to its object's local features. Following the decoding process, we derive the object query " + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_L" + }, + { + "bbox": [ + 130, + 327, + 480, + 399 + ], + "type": "text", + "content": " from the last decoding layer, and subsequently produce the initial prediction via two feed-forward networks (FFNs):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 244, + 399, + 480, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 399, + 480, + 412 + ], + "spans": [ + { + "bbox": [ + 244, + 399, + 480, + 412 + ], + "type": "interline_equation", + "content": "c _ {L} ^ {0} = \\phi_ {c} \\left(\\mathbf {Q} _ {L}\\right), \\quad b _ {L} ^ {0} = \\phi_ {b} \\left(\\mathbf {Q} _ {L}\\right) \\tag {12}", + "image_path": "44c5d63d0709ad445af89683f8c4b070b92e30eba0dc05d6bc47132edbb568cc.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 416, + 479, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 479, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 479, + 464 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 416, + 479, + 464 + ], + "type": "inline_equation", + "content": "c_{L}^{0}" + }, + { + "bbox": [ + 130, + 416, + 479, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 416, + 479, + 464 + ], + "type": "inline_equation", + "content": "b_{L}^{0}" + }, + { + "bbox": [ + 130, + 416, + 479, + 464 + ], + "type": "text", + "content": " represent the class and 3D bounding box predictions, respectively. While " + }, + { + "bbox": [ + 130, + 416, + 479, + 464 + ], + "type": "inline_equation", + "content": "\\phi_{c}" + }, + { + "bbox": [ + 130, + 416, + 479, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 416, + 479, + 464 + ], + "type": "inline_equation", + "content": "\\phi_{b}" + }, + { + "bbox": [ + 130, + 416, + 479, + 464 + ], + "type": "text", + "content": " are FFNs designated for classification and regression tasks. At this point, the first phase of \"prompt in 2D, detect in 3D\" is completed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 481, + 278, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 481, + 278, + 494 + ], + "spans": [ + { + "bbox": [ + 131, + 481, + 278, + 494 + ], + "type": "text", + "content": "3.6 3D Attribute Encoding" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 501, + 480, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 480, + 620 + ], + "type": "text", + "content": "For those inaccurate results of the initial predictions, we endow our model with iteratively self-correction capability under limited human interaction. Specifically, we decompose the 3D bounding box into (gravity center, width, length, height and yaw angle) five attributes, along with category information, and assume users to adjust a few wrong attributes, e.g. \"dragging the center point of the bounding box\". Given the prediction " + }, + { + "bbox": [ + 130, + 501, + 480, + 620 + ], + "type": "inline_equation", + "content": "b_{L}^{r}" + }, + { + "bbox": [ + 130, + 501, + 480, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 501, + 480, + 620 + ], + "type": "inline_equation", + "content": "c_{L}^{r}" + }, + { + "bbox": [ + 130, + 501, + 480, + 620 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 130, + 501, + 480, + 620 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 501, + 480, + 620 + ], + "type": "text", + "content": "-th iteration, the adjusted 3D bounding box and category are denoted as " + }, + { + "bbox": [ + 130, + 501, + 480, + 620 + ], + "type": "inline_equation", + "content": "b_{L}^{r'}" + }, + { + "bbox": [ + 130, + 501, + 480, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 501, + 480, + 620 + ], + "type": "inline_equation", + "content": "c_{L}^{r'}" + }, + { + "bbox": [ + 130, + 501, + 480, + 620 + ], + "type": "text", + "content": ". We also introduce two binary attribute refinement records " + }, + { + "bbox": [ + 130, + 501, + 480, + 620 + ], + "type": "inline_equation", + "content": "p_a^r \\in \\mathbb{R}^5" + }, + { + "bbox": [ + 130, + 501, + 480, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 501, + 480, + 620 + ], + "type": "inline_equation", + "content": "p_c^r \\in \\mathbb{R}^1" + }, + { + "bbox": [ + 130, + 501, + 480, + 620 + ], + "type": "text", + "content": " to explicitly aid the model in distinguishing between corrected attributes and those yet to be predicted." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 620, + 479, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 620, + 479, + 644 + ], + "spans": [ + { + "bbox": [ + 130, + 620, + 479, + 644 + ], + "type": "text", + "content": "Finally, the 3D attribute query " + }, + { + "bbox": [ + 130, + 620, + 479, + 644 + ], + "type": "inline_equation", + "content": "Q_{3D}" + }, + { + "bbox": [ + 130, + 620, + 479, + 644 + ], + "type": "text", + "content": " for each object is formed by the combination of both attribute and record embedding:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 154, + 653, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 653, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 154, + 653, + 480, + 666 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {3 D} = \\mathrm {M L P} _ {a t t} \\left(b _ {L} ^ {r ^ {\\prime}}\\right) + \\mathrm {M L P} _ {r e c 1} \\left(p _ {a} ^ {r}\\right) + \\mathrm {M L P} _ {c a t} \\left(c _ {L} ^ {r ^ {\\prime}}\\right) + \\mathrm {M L P} _ {r e c 2} \\left(p _ {c} ^ {r}\\right) \\tag {13}", + "image_path": "89d3bc037162a46e24f7fe3f3cf823d8fbd2f5b7633e71b2a46228c9138bf31a.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 256, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 256, + 91, + 448, + 102 + ], + "type": "text", + "content": "Interactive 3D Object Detection with Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": "And " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{3D}" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " is inserted into the corresponding query sequence of the respective object, collaborating with the original prompt tokens to jointly drive the model to provide more precise predictions. Note that in the following decoder process, " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{sa}" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ca}" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " are also dynamically adapted according to Eq. (6)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 181, + 308, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 181, + 308, + 192 + ], + "spans": [ + { + "bbox": [ + 131, + 181, + 308, + 192 + ], + "type": "text", + "content": "3.7 Label Assignment and Losses" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 199, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 199, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 199, + 482, + 284 + ], + "type": "text", + "content": "Due to the intrinsic one-to-one correspondence between our learnable queries and prompt objects, our proposed interactive 3D object detection task does not require label assignment like Hungarian Algorithm. To this end, following the methodologies in [21,45], we employ the Focal loss [19] for the classification task and the L1 loss for regression. To bolster the training of individual decoder layers, a deep supervision strategy is implemented. The 3D loss function amalgamates these predictions and is represented as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 208, + 291, + 481, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 291, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 208, + 291, + 481, + 323 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {3 D} = \\sum_ {r = 0} ^ {M} \\sum_ {i = 1} ^ {L} \\lambda_ {c l s} \\mathcal {L} _ {c l s} \\left(c _ {i} ^ {r}, c\\right) + \\lambda_ {r e g} \\mathcal {L} _ {r e g} \\left(b _ {i} ^ {r}, b\\right) \\tag {14}", + "image_path": "540486cbf8f9e794c862222b61841da5c274ef502b33057b8f372a5ba99623e5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "type": "text", + "content": " are the ground truths, " + }, + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "type": "inline_equation", + "content": "c_i^r" + }, + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "type": "inline_equation", + "content": "b_i^r" + }, + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "type": "text", + "content": " denote the predictions of the " + }, + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "type": "text", + "content": "-th decoder layer and " + }, + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "type": "text", + "content": "-th iteration, which are generated by the shared MLP as Eq. (12). " + }, + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\lambda_{cls}" + }, + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\lambda_{reg}" + }, + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "type": "text", + "content": " are weights for balancing loss items. Besides, we also utilize L1 loss to supervise the 2D bounding box prediction on the image or BEV modality, which is produced by another MLP and serves as the local region for point prompts to perform the global-to-local strategy. The overarching loss function is the combination of both 3D and 2D losses." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 431, + 230, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 431, + 230, + 445 + ], + "spans": [ + { + "bbox": [ + 131, + 431, + 230, + 445 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 455, + 265, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 455, + 265, + 468 + ], + "spans": [ + { + "bbox": [ + 131, + 455, + 265, + 468 + ], + "type": "text", + "content": "4.1 Experiment Settings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "content": "Dataset splits: The nuScenes dataset [2] is adopted in our work for training and evaluation. This dataset encompasses 1000 driving scenes, capturing data via six cameras, one LiDAR, and five radars. Following the official partition, we employ 750 scenes for training and 150 for validation. Contrasting prior work, our interactive 3D object detection approach leverages user prompts to guide the model's predictions. Similar to common interactive algorithms [4, 47], we reconstruct the evaluation benchmark by simulating user prompts and refinements from Ground Truth. Notably, since the nuScenes test set's Ground Truth remains inaccessible, we exclude it from our evaluations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "text", + "content": "Training Protocol: Throughout the training process, we leverage Ground Truth to simulate practical user prompts and iterative refinements. Detailed training procedures for Prompt in 2D, Detect in 3D and Detect in 3D, Refine in 3D phases are provided in the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Evaluation Protocol: We evaluate our method through human annotation comparisons and simulated experiments, under two distinct annotation scenarios: (1) Annotation from Raw Data and (2) Annotation from SOTA" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "R. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 178, + 169, + 438, + 205 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 1: Comparisons of the average and standard deviation time cost required for annotating an image extracted from the nuScenes validation set, with four annotation strategies, i.e. manual-only, Ours, SOTA model CMT with manual correction (CMT+M) and Ours (CMT+Ours) respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 178, + 169, + 438, + 205 + ], + "lines": [ + { + "bbox": [ + 178, + 169, + 438, + 205 + ], + "spans": [ + { + "bbox": [ + 178, + 169, + 438, + 205 + ], + "type": "table", + "html": "
Time cost (s)Annotation from Raw DataAnnotation from SOTA Model
Manual-onlyOursCMT+MCMT + Ours
nuScenes-val124±2855±1740±1224±7
", + "image_path": "35ceb3558ff8e0652ef322f94a2fc84f831b12d3537aab7983025d5c56c5f1ad.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 429 + ], + "type": "text", + "content": "Model. We describe the simulated experimental settings as follows: in scenario (1), we employ 2D projections of 3D Ground Truth with perturbations to simulate user prompts, while in scenario (2), SOTA 3D detection models provide initial prediction results to initiate our algorithm. The attributes from Ground Truth are utilized to mimic subsequent user refinements in both scenarios. We explore various orders of adjustments to bounding box attributes and assess our model both without and with corrected category priors, denoted as Ours and Ours*, respectively. * indicates the pre-calibration of category data. Unless specified, we prioritize using 2D bounding boxes as prompts and following the adjustment sequence of [gravity center, yaw angle, height, width, length]. Experimental results for alternative prompt formats (e.g., point prompts) and refinement orders are detailed in the supplementary materials. Furthermore, in addition to conventional closed-set evaluation, we also conduct extended experiments on challenging open-set scenarios. The nuScenes official metrics, i.e. nuScenes Detection Score (NDS), mean Average Precision (mAP) are adopted to evaluate the approaches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 444, + 328, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 444, + 328, + 456 + ], + "spans": [ + { + "bbox": [ + 131, + 444, + 328, + 456 + ], + "type": "text", + "content": "4.2 Human Annotation Comparisons" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 463, + 483, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 483, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 483, + 548 + ], + "type": "text", + "content": "To showcase the efficacy of our proposed method, we perform annotation comparisons across two scenarios employing four strategies: Manual-only, Ours, " + }, + { + "bbox": [ + 130, + 463, + 483, + 548 + ], + "type": "inline_equation", + "content": "\\mathrm{CMT + M}" + }, + { + "bbox": [ + 130, + 463, + 483, + 548 + ], + "type": "text", + "content": " (manual correction), and " + }, + { + "bbox": [ + 130, + 463, + 483, + 548 + ], + "type": "inline_equation", + "content": "\\mathrm{CMT + Ours}" + }, + { + "bbox": [ + 130, + 463, + 483, + 548 + ], + "type": "text", + "content": ". Ten experienced annotators are recruited to annotate the same set of twenty randomly selected nuScenes validation scenes using each of these strategies. Statistical analysis in Table 1 demonstrates that benefiting from the interactive design, our method achieves a speedup of approximately " + }, + { + "bbox": [ + 130, + 463, + 483, + 548 + ], + "type": "inline_equation", + "content": "56\\%" + }, + { + "bbox": [ + 130, + 463, + 483, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 463, + 483, + 548 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 463, + 483, + 548 + ], + "type": "text", + "content": " compared to manual-only and " + }, + { + "bbox": [ + 130, + 463, + 483, + 548 + ], + "type": "inline_equation", + "content": "\\mathrm{CMT + M}" + }, + { + "bbox": [ + 130, + 463, + 483, + 548 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 563, + 324, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 563, + 324, + 575 + ], + "spans": [ + { + "bbox": [ + 131, + 563, + 324, + 575 + ], + "type": "text", + "content": "4.3 Closed-Set 3D Object Detection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": "Annotation from SOTA Model Evaluation. Results from our experiments on the nuScenes validation set are tabulated in Table 2. In practical applications, annotators often streamline their efforts by initiating annotations from the 3D predictions generated by SOTA models, such as CMT. Our model seamlessly integrates into this workflow, harnessing its robust capabilities to effectively aid in annotation tasks. Specifically, given the 3D predictions from CMT, Table 2 illustrates the interactive refinement results, without employing category" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 256, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 256, + 91, + 448, + 102 + ], + "type": "text", + "content": "Interactive 3D Object Detection with Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 205, + 157, + 413, + 381 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Comparisons of annotation performance on nuScenes val set. \"M\" means manual correction. (Ri) denotes the " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "-th iteration refinement. \"C\" and \"L\" are camera and LiDAR modalities respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 205, + 157, + 413, + 381 + ], + "lines": [ + { + "bbox": [ + 205, + 157, + 413, + 381 + ], + "spans": [ + { + "bbox": [ + 205, + 157, + 413, + 381 + ], + "type": "table", + "html": "
MethodModalityNDS↑mAP↑
Traditional 3D Detectors
UVTR [16]CL70.265.4
BEVFusion [23]CL71.468.5
MetaBEV [8]CL71.568.0
SpaseFusion [42]CL72.870.4
CMT [45]CL72.970.3
Annotation from SOTA Model
CMT+M (R1)CL81.381.6
CMT+M (R2)CL83.881.6
CMT+M (R3)CL84.381.6
CMT+M (R4)CL85.381.6
CMT+M (R5)CL86.381.6
CMT+Ours (R1)CL83.4 (+2.1)86.3 (+4.7)
CMT+Ours (R2)CL86.3 (+2.5)86.3 (+4.7)
CMT+Ours (R3)CL87.4 (+3.1)87.0 (+5.4)
CMT+Ours (R4)CL89.3 (+4.0)89.1 (+7.5)
CMT+Ours (R5)CL90.4 (+4.1)89.4 (+7.8)
Annotation from Raw Data
OursCL75.076.2
Ours (R1)CL84.4 (+9.4)88.0 (+11.8)
Ours (R2)CL87.2 (+12.2)88.0 (+11.8)
Ours (R3)CL88.2 (+13.2)88.4 (+12.2)
Ours (R4)CL89.5 (+14.5)89.5 (+13.3)
Ours (R5)CL90.5 (+15.5)89.6 (+13.4)
", + "image_path": "a87d97c9c96d4556d3039d1c7489bfb96c644b90f89081e3c2571fc3948d880f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "text", + "content": "correction priors. As a comparative reference, we perform identical manual attribute adjustments on the initial results predicted by CMT. It is evident that, in five consecutive iterations, our approach consistently outperforms " + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\mathrm{CMT + M}" + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "text", + "content": " by margins of " + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "inline_equation", + "content": "4.0\\%" + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "text", + "content": ", respectively, in terms of NDS. Additionally, CMT maintains a constant mAP after the initial refinement of the center position, lacking self-correction for object categories. In contrast, our approach iteratively refines predictions with minimal human intervention, resulting in sustained performance improvements. In addition, we also provide comparisons with corrected category priors, denoted as " + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\mathrm{CMT + M^*}" + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\mathrm{CMT + Ours^*}" + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "text", + "content": " in Fig. 4. Through the rectification of object categories, our model exhibits a significantly enhanced self-correction capability, outperforming " + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\mathrm{CMT + M^*}" + }, + { + "bbox": [ + 130, + 417, + 482, + 561 + ], + "type": "text", + "content": " by a considerable margin." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Annotation from Raw Data Evaluation. In cases involving missed detections by existing 3D detectors, our method is expected to have the capability of annotating from raw data. As shown in Table 2, by leveraging a box as the prompt format, our model registers an NDS of " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "75.0\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " and an mAP of " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "76.2\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ". Subsequent iterations of refinements contribute to continued performance gains, and the final model attains an NDS score of " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "90.5\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " and mAP of " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "89.6\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ". Some visualization results are presented in Fig. 5." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "R. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 221, + 128, + 381, + 236 + ], + "blocks": [ + { + "bbox": [ + 221, + 128, + 381, + 236 + ], + "lines": [ + { + "bbox": [ + 221, + 128, + 381, + 236 + ], + "spans": [ + { + "bbox": [ + 221, + 128, + 381, + 236 + ], + "type": "image", + "image_path": "68ea3277f58d0d6a10100bbb409211c0c9f0c2cf903e37f4c8e9556d6405e2c7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 242, + 480, + 264 + ], + "lines": [ + { + "bbox": [ + 131, + 242, + 480, + 264 + ], + "spans": [ + { + "bbox": [ + 131, + 242, + 480, + 264 + ], + "type": "text", + "content": "Fig. 4: Performance comparisons of annotation from SOTA model. R1-R5 follows a fixed refinement order. * indicates the utilization of category correction." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 135, + 319, + 282, + 373 + ], + "blocks": [ + { + "bbox": [ + 131, + 276, + 290, + 308 + ], + "lines": [ + { + "bbox": [ + 131, + 276, + 290, + 308 + ], + "spans": [ + { + "bbox": [ + 131, + 276, + 290, + 308 + ], + "type": "text", + "content": "Table 3: Categories and quantities of open-set objects extracted from nuScenes validation set." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 319, + 282, + 373 + ], + "lines": [ + { + "bbox": [ + 135, + 319, + 282, + 373 + ], + "spans": [ + { + "bbox": [ + 135, + 319, + 282, + 373 + ], + "type": "table", + "html": "
CategoryNumCategoryNum
personal_mobility15stroller103
animal32bicycle_rac246
pushableifiable2925debris514
ambulance30police64
", + "image_path": "35a5b82eaecad245bbb3667d9c6dfaa4949333068227b5cebf2a2c298909694f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 157, + 437, + 268, + 505 + ], + "blocks": [ + { + "bbox": [ + 131, + 383, + 291, + 426 + ], + "lines": [ + { + "bbox": [ + 131, + 383, + 291, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 383, + 291, + 426 + ], + "type": "text", + "content": "Table 5: Ablation study of prompt queries. \"PE\", \"CE\" are position encoding and context encoding. \"It.\" denotes the iteration." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 437, + 268, + 505 + ], + "lines": [ + { + "bbox": [ + 157, + 437, + 268, + 505 + ], + "spans": [ + { + "bbox": [ + 157, + 437, + 268, + 505 + ], + "type": "table", + "html": "
It.QtQPPEQpCEQ3DbL'NDS↑mAP↑
0✓73.575.5
0✓✓74.575.7
0✓✓✓75.076.2
1✓✓✓✓83.285.9
1✓✓✓✓84.488.0
", + "image_path": "ed337ed98536e8a10e87d21a0db382980ea3dac63e35957a953b2ae09c9f9558.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 331, + 297, + 449, + 373 + ], + "blocks": [ + { + "bbox": [ + 326, + 276, + 451, + 287 + ], + "lines": [ + { + "bbox": [ + 326, + 276, + 451, + 287 + ], + "spans": [ + { + "bbox": [ + 326, + 276, + 451, + 287 + ], + "type": "text", + "content": "Table 4: Open-set evaluation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 297, + 449, + 373 + ], + "lines": [ + { + "bbox": [ + 331, + 297, + 449, + 373 + ], + "spans": [ + { + "bbox": [ + 331, + 297, + 449, + 373 + ], + "type": "table", + "html": "
MethodModalityNDS↑mAP↑
Annotation from Raw Data
Ours*CL89.391.9
Ours* (R1)CL94.199.5
Ours* (R2)CL96.699.5
Ours* (R3)CL97.399.5
Ours* (R4)CL98.199.5
Ours* (R5)CL99.299.5
", + "image_path": "86b1e606b054c8564ce98fae3b569903de82dae16714759a23e839afba640608.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 343, + 448, + 437, + 495 + ], + "blocks": [ + { + "bbox": [ + 309, + 383, + 468, + 437 + ], + "lines": [ + { + "bbox": [ + 309, + 383, + 468, + 437 + ], + "spans": [ + { + "bbox": [ + 309, + 383, + 468, + 437 + ], + "type": "text", + "content": "Table 6: Ablation study of Global-to-Local strategy (G2L). \"L\", \"G\" denotes utilizing local RoI or global multi-modal features to perform Cross-Attention." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 343, + 448, + 437, + 495 + ], + "lines": [ + { + "bbox": [ + 343, + 448, + 437, + 495 + ], + "spans": [ + { + "bbox": [ + 343, + 448, + 437, + 495 + ], + "type": "table", + "html": "
It.AttentionNDS↑mAP↑
0G73.775.1
0L74.375.2
0G2L75.076.2
", + "image_path": "9e61dc1fb6cccaf6374dd583b11d23d9876054ae93a404c308775b1463516cb1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 526, + 316, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 526, + 316, + 539 + ], + "spans": [ + { + "bbox": [ + 131, + 526, + 316, + 539 + ], + "type": "text", + "content": "4.4 Open-Set 3D Object Detection" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": "To delve deeper into the 3D object annotation capabilities guided by 2D prompts, we extract eight novel categories from the nuScenes dataset, which are not employed in the training process. Table 3 outlines their distribution on the validation set. Since our primary goal is to validate the model's generalization ability based on user-provided prompts, as opposed to venturing into open-vocabulary recognition, we ignore the class prediction and assume that the class information is provided manually together with 2D prompts, denoted as Ours*. The experimental results in Table 4 demonstrate that our model possesses the capability to detect 3D objects based on 2D priors. Furthermore, with manual interaction for refinement, the detection performance improves progressively." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 257, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 257, + 91, + 447, + 102 + ], + "type": "text", + "content": "Interactive 3D Object Detection with Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 114, + 462, + 227 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 462, + 227 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 462, + 227 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 462, + 227 + ], + "type": "image", + "image_path": "58d504a583bb570b5a5c5bfd8e5dc20090a6c802ad9351fcc59b8e6cc5d17c9e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 236, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 291 + ], + "type": "text", + "content": "Fig. 5: Visualizations of the 3D detection results from our method are presented. Ground Truth and our predictions are exhibited in green and blue, respectively. The outcomes highlighted by red circles illustrate that, with a single round of manual adjustment focusing solely on the gravity center, our model automatically rectified other attributes such as yaw angles, thus yielding more accurate 3D results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 314, + 238, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 238, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 238, + 326 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 334, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 441 + ], + "type": "text", + "content": "Ablation Study of Object Queries. The pivotal role of prompt tokens is evident from the results in Table 5. When solely introducing prompt queries with position encoding, the model attains a nuanced understanding of the object's spatial location, registering an NDS of " + }, + { + "bbox": [ + 130, + 334, + 482, + 441 + ], + "type": "inline_equation", + "content": "74.5\\%" + }, + { + "bbox": [ + 130, + 334, + 482, + 441 + ], + "type": "text", + "content": " —a significant boost of " + }, + { + "bbox": [ + 130, + 334, + 482, + 441 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 130, + 334, + 482, + 441 + ], + "type": "text", + "content": ". Moreover, integrating context embedding, which encapsulates appearance attributes of objects, further escalates performance, culminating in a gain of " + }, + { + "bbox": [ + 130, + 334, + 482, + 441 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 130, + 334, + 482, + 441 + ], + "type": "text", + "content": ". Besides, we also delve into an analysis of the effectiveness of " + }, + { + "bbox": [ + 130, + 334, + 482, + 441 + ], + "type": "inline_equation", + "content": "Q_{3D}" + }, + { + "bbox": [ + 130, + 334, + 482, + 441 + ], + "type": "text", + "content": " during the first iteration refinement stage. Clear performance improvements are evident with the explicit inclusion of the current attribute refinement record " + }, + { + "bbox": [ + 130, + 334, + 482, + 441 + ], + "type": "inline_equation", + "content": "p_a^r" + }, + { + "bbox": [ + 130, + 334, + 482, + 441 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 442, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 442, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 442, + 482, + 502 + ], + "type": "text", + "content": "Ablation Study of Global-to-Local Strategy. The proposed global-to-local strategy's efficacy is substantiated by the results displayed in Table 6. Contrasted with models relying exclusively on either global or local feature extraction, our hybrid approach ensures a comprehensive global perspective while meticulously capturing local nuances, leading to superior performance outcomes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 521, + 226, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 226, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 226, + 533 + ], + "type": "text", + "content": "5 Conclusions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "This paper introduces a new method for 3D object detection that tackles the daunting task of 3D object annotation. In consideration of the necessity for either creating annotations from raw data or refining existing imperfect annotations within practice annotation scenarios, two dedicated strategies are proposed. \"Prompt in 2D, detect in 3D\" principle leverages straightforward 2D interactions, like clicks or boxes, to streamline the transition between 2D images and 3D object annotations. \"Detect in 3D, refine in 3D\" strategy further endows our model with self-correction capability. Evaluations on the nuScenes dataset validate our approach's superiority. Beyond a mere advancement, our work can serve as a cornerstone for future 3D object detection endeavors." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "R. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 203 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 203 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China (NO. 62322608), in part by the Fundamental Research Funds for the Central Universities under Grant 22lgqb25, and in part by the Open Project Program of the Key Laboratory of Artificial Intelligence for Perception and Understanding, Liaoning Province (AIPU, No. 20230003)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 224, + 197, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 197, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 197, + 237 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 251, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 251, + 480, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 296 + ], + "type": "text", + "content": "1. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1090-1099 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 297, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 297, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 138, + 297, + 480, + 341 + ], + "type": "text", + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11621-11631 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "type": "text", + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European Conference on Computer Vision. pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 375, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 375, + 480, + 407 + ], + "type": "text", + "content": "4. Chen, X., Zhao, Z., Zhang, Y., Duan, M., Qi, D., Zhao, H.: Focalclick: Towards practical interactive image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1300-1309 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 409, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 409, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 138, + 409, + 480, + 441 + ], + "type": "text", + "content": "5. Chen, X., Ma, H., Wan, J., Li, B., Xia, T.: Multi-view 3d object detection network for autonomous driving. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1907-1915 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 442, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 442, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 138, + 442, + 480, + 463 + ], + "type": "text", + "content": "6. Chen, Y., Chen, Q., Sun, P., Chen, S., Wang, J., Cheng, J.: Enhancing your trained detrs with box refinement. arXiv preprint arXiv:2307.11828 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 464, + 480, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 464, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 138, + 464, + 480, + 485 + ], + "type": "text", + "content": "7. Choi, D., Cho, W., Kim, K., Choo, J.: idet3d: Towards efficient interactive object detection for lidar point clouds. arXiv preprint arXiv:2312.15449 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 487, + 480, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 480, + 518 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 480, + 518 + ], + "type": "text", + "content": "8. Ge, C., Chen, J., Xie, E., Wang, Z., Hong, L., Lu, H., Li, Z., Luo, P.: Metabev: Solving sensor failures for bev detection and map segmentation. arXiv preprint arXiv:2304.09801 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 520, + 480, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 520, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 138, + 520, + 480, + 552 + ], + "type": "text", + "content": "9. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition. pp. 3354-3361. IEEE (2012)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 554, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 480, + 575 + ], + "type": "text", + "content": "0. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2961-2969 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 577, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 597 + ], + "type": "text", + "content": "1. Huang, J., Huang, G.: Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 599, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 631 + ], + "type": "text", + "content": "2. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 135, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 632, + 480, + 665 + ], + "type": "text", + "content": "13. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 257, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 257, + 91, + 447, + 102 + ], + "type": "text", + "content": "Interactive 3D Object Detection with Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 150 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 150 + ], + "type": "text", + "content": "14. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12697-12705 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 184 + ], + "type": "text", + "content": "15. Lee, C., Park, S., Song, H., Ryu, J., Kim, S., Kim, H., Pereira, S., Yoo, D.: Interactive multi-class tiny-object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14136-14145 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "text", + "content": "16. Li, Y., Chen, Y., Qi, X., Li, Z., Sun, J., Jia, J.: Unifying voxel-based representation with transformer for 3d object detection. Advances in Neural Information Processing Systems 35, 18442-18455 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "type": "text", + "content": "17. Li, Z., Wang, F., Wang, N.: Lidar r-cnn: An efficient and universal 3d object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7546-7555 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 296 + ], + "type": "text", + "content": "18. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: European Conference on Computer Vision. pp. 1-18. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 296, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 296, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 481, + 329 + ], + "type": "text", + "content": "19. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2980-2988 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 329, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 329, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 329, + 481, + 363 + ], + "type": "text", + "content": "20. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 363, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 481, + 397 + ], + "type": "text", + "content": "21. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. In: European Conference on Computer Vision. pp. 531-548. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "type": "text", + "content": "22. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X., Sun, J.: Petrov2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 430, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 475 + ], + "type": "text", + "content": "23. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2774-2781. IEEE (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 475, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 520 + ], + "type": "text", + "content": "24. Lu, Y., Ma, X., Yang, L., Zhang, T., Liu, Y., Chu, Q., Yan, J., Ouyang, W.: Geometry uncertainty projection network for monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3111-3121 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "text", + "content": "25. Maninis, K.K., Caelles, S., Pont-Tuset, J., Van Gool, L.: Deep extreme cut: From extreme points to object segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 616-625 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 553, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 553, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 553, + 481, + 587 + ], + "type": "text", + "content": "26. Meng, D., Chen, X., Fan, Z., Zeng, G., Li, H., Yuan, Y., Sun, L., Wang, J.: Conditional detr for fast training convergence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3651-3660 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "type": "text", + "content": "27. Meng, Q., Wang, W., Zhou, T., Shen, J., Jia, Y., Van Gool, L.: Towards a weakly supervised framework for 3d point cloud object detection and annotation. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(8), 4454-4468 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "28. Papadopoulos, D.P., Uijlings, J.R., Keller, F., Ferrari, V.: Extreme clicking for efficient object annotation. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4930-4939 (2017)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "R. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "29. Papadopoulos, D.P., Uijlings, J.R., Keller, F., Ferrari, V.: Training object class detectors with click supervision. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6374-6383 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 193 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 193 + ], + "type": "text", + "content": "30. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 194-210. Springer (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 193, + 482, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 482, + 224 + ], + "type": "text", + "content": "31. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 224, + 482, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 257 + ], + "type": "text", + "content": "32. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems 28 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "type": "text", + "content": "33. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-779 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 289, + 482, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 321 + ], + "type": "text", + "content": "34. Shin, K., Kwon, Y.P., Tomizuka, M.: Roarnet: A robust 3d object detection based on region approximation refinement. In: 2019 IEEE Intelligent Vehicles Symposium (IV). pp. 2510-2515. IEEE (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 321, + 482, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 354 + ], + "type": "text", + "content": "35. Sofiuk, K., Petrov, I.A., Konushin, A.: Reviving iterative training with mask guidance for interactive segmentation. In: 2022 IEEE International Conference on Image Processing (ICIP). pp. 3141-3145. IEEE (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 354, + 482, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 396 + ], + "type": "text", + "content": "36. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2446-2454 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 396, + 482, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 429 + ], + "type": "text", + "content": "37. Wang, B., Wu, V., Wu, B., Keutzer, K.: Latte: accelerating lidar point cloud annotation via sensor fusion, one-click annotation, and tracking. In: 2019 IEEE Intelligent Transportation Systems Conference (ITSC). pp. 265-272. IEEE (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 429, + 482, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 472 + ], + "type": "text", + "content": "38. Wang, K., Zhuang, J., Li, G., Fang, C., Cheng, L., Lin, L., Zhou, F.: De-biased teacher: Rethinking iou matching for semi-supervised object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2573-2580 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 472, + 482, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 504 + ], + "type": "text", + "content": "39. Wang, T., Zhu, X., Pang, J., Lin, D.: Fcos3d: Fully convolutional one-stage monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 913-922 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 504, + 482, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 536 + ], + "type": "text", + "content": "40. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor detr: Query design for transformer-based detector. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2567-2575 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "type": "text", + "content": "41. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning. pp. 180–191. PMLR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "type": "text", + "content": "42. Xie, Y., Xu, C., Rakotosaona, M.J., Rim, P., Tombari, F., Keutzer, K., Tomizuka, M., Zhan, W.: Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection. arXiv preprint arXiv:2304.14340 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "text", + "content": "43. Xu, M., Zhang, Z., Hu, H., Wang, J., Wang, L., Wei, F., Bai, X., Liu, Z.: End-to-end semi-supervised object detection with soft teacher. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3060-3069 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 633, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 633, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 633, + 482, + 665 + ], + "type": "text", + "content": "44. Xu, N., Price, B., Cohen, S., Yang, J., Huang, T.S.: Deep interactive object selection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 373-381 (2016)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 256, + 91, + 447, + 102 + ], + "type": "text", + "content": "Interactive 3D Object Detection with Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 436 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "45. Yan, J., Liu, Y., Sun, J., Jia, F., Li, S., Wang, T., Zhang, X.: Cross modal transformer via coordinates encoding for 3d object detection. arXiv preprint arXiv:2301.01283 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "text", + "content": "46. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "47. Yang, J., Zeng, A., Li, F., Liu, S., Zhang, R., Zhang, L.: Neural interactive keypoint detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15122-15132 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "48. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3dssd: Point-based 3d single stage object detector. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11040-11048 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "49. Yao, A., Gall, J., Leistner, C., Van Gool, L.: Interactive object detection. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition. pp. 3242-3249. IEEE (2012)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 314 + ], + "type": "text", + "content": "50. Zhang, J., Lin, X., Zhang, W., Wang, K., Tan, X., Han, J., Ding, E., Wang, J., Li, G.: Semi-detr: Semi-supervised object detection with detection transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23809-23818 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 314, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 481, + 346 + ], + "type": "text", + "content": "51. Zhang, R., Qiu, H., Wang, T., Guo, Z., Xu, X., Qiao, Y., Gao, P., Li, H.: Monodetr: depth-guided transformer for monocular 3d object detection. arXiv preprint arXiv:2203.13310 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 346, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 346, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 346, + 481, + 369 + ], + "type": "text", + "content": "52. Zhao, J., Sun, L., Li, Q.: Recursivelydet: End-to-end region-based recursive object detection. arXiv preprint arXiv:2307.13619 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "53. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4490-4499 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 402, + 481, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 436 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 436 + ], + "type": "text", + "content": "54. Zhou, Y., Zhu, H., Liu, Q., Chang, S., Guo, M.: Monoatt: Online monocular 3d object detection with adaptive token transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17493-17503 (2023)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "R. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection/0ff966a8-e3c5-432a-8beb-e808413549e5_content_list.json b/2024/Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection/0ff966a8-e3c5-432a-8beb-e808413549e5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..31b81ad22ff13ce050b3372faa6c99cc1eda8044 --- /dev/null +++ b/2024/Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection/0ff966a8-e3c5-432a-8beb-e808413549e5_content_list.json @@ -0,0 +1,1804 @@ +[ + { + "type": "text", + "text": "Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection", + "text_level": 1, + "bbox": [ + 233, + 140, + 769, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yongwei Nie $^{1}$ , Hao Huang $^{1}$ , Chengjiang Long $^{2}$ , Qing Zhang $^{3}$ , Pradipta", + "bbox": [ + 223, + 233, + 781, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Maji $^{4}$ , and Hongmin Cai $^{1}$", + "bbox": [ + 390, + 253, + 612, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ South China University of Technology, China", + "bbox": [ + 341, + 282, + 658, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meta Reality Labs, USA", + "bbox": [ + 410, + 297, + 591, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Sun Yat-sen University, China", + "bbox": [ + 393, + 311, + 609, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Indian Statistical Institute, India", + "bbox": [ + 383, + 325, + 619, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Video Anomaly Detection (VAD) has been extensively studied under the settings of One-Class Classification (OCC) and Weakly-Supervised learning (WS), which however both require laborious human-annotated normal/abnormal labels. In this paper, we study Unsupervised VAD (UVAD) that does not depend on any label by combining OCC and WS into a unified training framework. Specifically, we extend OCC to weighted OCC (wOCC) and propose a wOCC-WS interleaving training module, where the two models automatically generate pseudo-labels for each other. We face two challenges to make the combination effective: (1) Models' performance fluctuates occasionally during the training process due to the inevitable randomness of the pseudo labels. (2) Thresholds are needed to divide pseudo labels, making the training depend on the accuracy of user intervention. For the first problem, we propose to use wOCC requiring soft labels instead of OCC trained with hard zero/one labels, as soft labels exhibit high consistency throughout different training cycles while hard labels are prone to sudden changes. For the second problem, we repeat the interleaving training module multiple times, during which we propose an adaptive thresholding strategy that can progressively refine a rough threshold to a relatively optimal threshold, which reduces the influence of user interaction. A benefit of employing OCC and WS methods to compose a UVAD method is that we can incorporate the most recent OCC or WS model into our framework. Experiments demonstrate the effectiveness of the proposed UVAD framework. Our code is available at https://github.com/benedictstar/Joint-VAD.", + "bbox": [ + 259, + 366, + 743, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Video anomaly detection $\\cdot$ One-class classification $\\cdot$ Weakly-supervised learning $\\cdot$ Unsupervised learning", + "bbox": [ + 259, + 710, + 740, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 760, + 375, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video Anomaly Detection (VAD) is a task that identifies abnormal events in a video, where the abnormal event could be a fire alarm, a flaw in an industrial", + "bbox": [ + 212, + 787, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Hongmin Cai is the corresponding author. Email: hmcai@scut.edu.cn", + "bbox": [ + 217, + 825, + 692, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "product, or a traffic accident, etc. Most previous VAD methods fall in two categories: One-Class Classification (OCC) methods [12,13,18,19,33,39] and Weakly-Supervised (WS) approaches [21,32,34,38,43,49]. Both kinds of approaches require human-annotated labels for training1. Supervised VAD methods have two prominent problems. First, abnormal events are sparse and difficult to recognize, making the annotation of the training dataset a highly labor-intensive task. Secondly, the categories of abnormal events are unbounded, while human can only collect a limited scope of them, yielding the risk that the trained VAD methods cannot handle unobserved abnormal events.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Considering the above problems, this paper studies Unsupervised VAD (UVAD) that does not rely on labels. Despite the importance of UVAD, few methods have been developed except [47] and its subsequent works [2, 37] to the best of our knowledge. To achieve unsupervised learning, the method of [47] adversarially trains a generator and a discriminator, where the generator is an autoencoder-based reconstruction model and the discriminator is a fully-connected (FC) binary classifier. While the proposed network architecture is sophisticated, the adopted autoencoder and FC networks are relatively weak which limit the performance of the approach.", + "bbox": [ + 212, + 282, + 795, + 419 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a new UVAD framework. Our key idea is that the techniques of OCC and WS approaches are rapidly developed, while a UVAD method is usually implemented by training two VAD models alternately generating pseudo labels for each other. This motivates us to propose a UVAD framework that directly interleaves the training of a pair of OCC and WS models. This allows us to incorporate the recent advances in the two hot research fields to implement a UVAD method. We have also attempted to combine two OCC or two WS models, but found two homogeneous models are less effective.", + "bbox": [ + 212, + 420, + 787, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While combining OCC and WS methods, we encounter two problems preventing it from effective. First, we observe the OCC or WS model's performance fluctuates occasionally during the training process, which affects the final accuracy they can achieve. This is partly because the pseudo labels generated by each other are changed frequently, making the training not stable. To solve this problem, we extend OCC to weighted OCC (wOCC), and propose a wOCC-WS interleaving training module. Previous OCC model is trained on normal data, while our wOCC model is trained with soft labels in the range of [0,1]. The soft labels are more consistent across adjacent training cycles, making the performance of wOCC more stable than OCC under changing pseudo labels, thus suppressing the fluctuation effect. Second, we still have to set a threshold to partition normal and abnormal pseudo labels for the WS model, making our method susceptible to a user-provided hyperparameter. To reduce the influence of user interaction, we repeat the interleaving training module multiple times, during which we propose an adaptive threshold mechanism that finds an optimal threshold given just a rough initial value from the user.", + "bbox": [ + 212, + 541, + 787, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To conclude, the contributions of this work include:", + "bbox": [ + 238, + 784, + 612, + 799 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Nie et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1 OCC models are trained on normal data only, which requires human labor to exclude abnormal events in preparing the training dataset.", + "bbox": [ + 217, + 810, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new UVAD framework that interleaves the training of a pair of OCC and WS models.", + "- We improve OCC to wOCC which stabilizes the training of the UVAD framework, and also propose an adaptive thresholding strategy to alleviate the influence of user interaction.", + "- Our UVAD framework is flexible to employ different pairs of OCC and WS models. Experiments on three OCC and two WS models demonstrate the effectiveness of our method." + ], + "bbox": [ + 223, + 146, + 784, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 292, + 387, + 308 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unsupervised VAD (UVAD). Unsupervised VAD requires identifying anomalies from data containing both normal and abnormal samples without any annotation. This is a challenging new task that is almost unexplored in the literature. Although significant progress has been made in OCC and WS VAD, directly applying them independently to address UVAD does not yield good results. Zaheer et al. [47] firstly raise this task and propose to solve UVAD by generative cooperative learning (GCL). Differently, our proposed method modifies both the OCC and WS VAD models and combines the modified two models together to learn from unmarked training data.", + "bbox": [ + 212, + 325, + 785, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One-Class Classification VAD (OCC). OCC-based VAD approaches only have access to normal data and try to model normal data to identify behaviors that are significantly different from normal behaviors as anomalies. Early works use hand-crafted features to help detect anomalies [5,23,40,51]. With the rapid development of deep learning, recent methods turn to normal representations extracted by using a deep neural network [26, 30, 36, 41]. Some methods identify normal patterns by using the reconstruction/prediction model to reconstruct the representations [6,12,14,18,19,24,25,27,31,33,45,46,52]. These models may lead to well-reconstructed anomalies, thereby limiting the performance of the OCC-based methods. Other OCC-based methods turn to identify normal representations by using proxy tasks [4,11,20,29,39]. What's more, recently, Hirschorn and Avidan [13] propose to build a multivariate Gaussian distribution for normal data and detect instances deviating from this distribution as anomalies. However, the above methods treat all samples equally, ignoring the potential differences in importance between normal samples.", + "bbox": [ + 212, + 462, + 785, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Weakly-Supervised VAD (WS). In weakly-supervised VAD, video-level annotations are available in the training stage. Sultani et al [32] first propose to use the video-level labels and solve WS VAD based on multi-instance learning (MIL) framework [3, 7]. Since then, many works [21, 28, 38, 43, 49, 50] have viewed the WS VAD task as a MIL problem. Tian et al. [34] develop to extend MIL to Top- $k$ MIL method by training with a robust temporal feature magnitude (RTFM) loss function. The methods mentioned above directly train in the supervision of video-level labels and the coarse-grained supervision limits the accuracy of WS models. Recently, two-stage self-training methods [8, 17, 44, 53] adopt a two-stage pipeline to use more fine-grained labels to supervise the networks more", + "bbox": [ + 212, + 688, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD", + "bbox": [ + 238, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fc004a50af94976ab7d38dd19d88eb911212310d85aa5cd2ff8131f8cc6bb500.jpg", + "image_caption": [ + "Fig. 1: Pipeline of our proposed interleaved framework. The wOCC-WS interleaving training module comprises a weighted OCC (wOCC) model and a WS model, using the output of each model as the pseudo-labels to supervise the other model. Further, the training module is repeated multiple times with the adaptive threshold until the stopping criterion we set. To start the first training module, we randomly sample $w_{X}$ to train the wOCC model and assign a large value as the threshold for the WS model. Once a training module converges, the last WS model of it supplies $w_{X}$ to start the next training module and adaptive thresholding takes the outputs of all previous wOCC models into consideration for updating the threshold." + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 782, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "strictly. The WS model in our framework shares a similar idea with [8] and is tailored to use finer-grained snippet-level labels.", + "bbox": [ + 212, + 486, + 785, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 540, + 330, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 571, + 341, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our proposed UVAD framework interleaves the training of a pair of OCC and WS models. In theory, any OCC or WS methods can be incorporated into our framework. Without loss of generality, we take STG-NF [13] (a recent OCC model) and RTFM [34] (a recent WS model) as examples to introduce our method in the following sections. We test many other OCC/WS models in Section 4.5.", + "bbox": [ + 212, + 597, + 785, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 1 shows an overview of our framework. The basic buildingblock of our method is the wOCC-WS interleaving training module. In the module, the outputs of each model are used as pseudo labels for defining the loss function training the other model. After one module converges, we re-initialize another module to run the interleaving training again with a difference that the threshold of the WS model is updated by an adaptive threshold adjustment mechanism based on wOCC models trained in previous modules. Since wOCC has more stable behaviors than WS, we always initialize the wOCC model to start the training of a module. For the first module, we use a random initialization based on Beta distribution sampling. For other modules, we use the results of WS of the preceding module for the initialization.", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Nie et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 wOCC-WS Interleaving Training Module", + "text_level": 1, + "bbox": [ + 215, + 146, + 602, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first introduce the wOCC-WS interleaving training module which comprises a wOCC model and a WS model. At the very beginning, we initialize the wOCC model to start the interleaving training. Then, pseudo labels created using wOCC are used to train the WS model. Next, pseudo labels created by the WS model are used to improve the wOCC model. The whole training does not use human annotations.", + "bbox": [ + 212, + 171, + 787, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "wOCC Model (STG-NF). Typically, OCC models are trained only on normal data. Directly using OCC in our UVAD framework may make the training fluctuate, i.e., the performance of the trained models sometimes gets better, but sometimes worse, finally converging to an inferior point. We use 0 to indicate normal samples, and 1 abnormal samples, where 0, 1 are hard binary labels. During training, we need to identify 0-labelled samples to train the OCC model. The hard partitioned labels may be suddenly changed from 0 to 1 or from 1 to 0, yielding the frequent change of the normal samples used to train the OCC model which is the source of the unstable training. To this end, we propose wOCC which relies on soft labels instead of hard binary labels. Soft labels are in the range of [0, 1]. For example, the soft label 0.7 may be changed to 0.6 but not 0 suddenly. This consistency reduces the fluctuation of the interleaving training.", + "bbox": [ + 212, + 261, + 787, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Taking STG-NF [13] (an OCC model) as an example, we introduce how to improve it to wOCC. STG-NF extracts a set of objects $\\mathcal{X} = \\{X_1,\\dots ,X_i,\\dots \\}$ from the training videos, where $X_{i}$ is a sequence of poses of a person. Each object $X_{i}$ is the basic building block processed by STG-NF. Originally, STG-NF minimizes the negative log-likelihood across the normal data:", + "bbox": [ + 212, + 443, + 787, + 518 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {o c c} = - \\log \\left(p _ {Z} \\left(f _ {S T G - N F} \\left(X _ {i} ^ {+}\\right)\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 531, + 784, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $X_{i}^{+} \\in \\mathcal{X}^{+}$ is sampled from normal data, $Z = f_{STG-NF}(X_{i}^{+})$ is the feature extracted from $X_{i}^{+}$ by the STG-NF, $p_{Z}$ is the established distribution of the normal data. We upgrade STG-NF by introducing the soft labels $w_{X} = \\{w_{X_{1}}, \\dots, w_{X_{i}}, \\dots\\}$ as the weight to supervise its training on both normal and abnormal data. Formally, for the wOCC model improved from STG-NF, we minimize the following weighted negative log-likelihood across all the data:", + "bbox": [ + 214, + 556, + 787, + 650 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {w o c c} = - \\left(1 - w _ {X _ {i}}\\right) \\log \\left(p _ {Z} \\left(f _ {S T G - N F} \\left(X _ {i}\\right)\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 662, + 784, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $X_{i} \\in \\mathcal{X}$ is sampled from the whole training dataset rather than the normal part and $w_{X_i} \\in [0,1]$ is the weight of the object $X_{i}$ indicating the anomaly degree of the object which serves as the soft label. The object with smaller $w_{X_i}$ means more normal which is better captured by the learned distribution of wOCC.", + "bbox": [ + 212, + 688, + 787, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pseudo Labels $l_{S}$ from wOCC. Now we use the trained wOCC model to generate pseudo labels $l_{S} = \\{l_{S_{1}},\\dots ,l_{S_{i}},\\dots \\}$ for training the WS model. \"Snippet\" is the basic element processed by WS models. We first split the training videos into snippets, obtaining a set of snippets $S = \\{S_1,\\dots ,S_i,\\dots \\}$ . Our aim is to compute the label $l_{S_i}\\in \\{0,1\\}$ for each snippet $S_{i}$ . Let $x_{i}$ be the anomaly score of the object $X_{i}$ computed by the trained wOCC model. The anomaly score $\\hat{s}_i$", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD", + "bbox": [ + 238, + 113, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of a video snippet $S_{i}$ is computed by averaging all the anomaly scores of objects in the snippet. To generate pseudo label $l_{S_i}$ , one way is to check whether $\\hat{s}_i$ is larger than a threshold $(l_{S_i} = 1)$ or not $(l_{S_i} = 0)$ . However, as the training goes, the anomaly scores of snippets may vary in a large range, making the absolute threshold invalid. Instead, we sort the snippets according to their anomaly scores in descending order, and get the index of the snippet $S_{i}$ by $\\mathrm{Rank}(\\hat{s}_i)$ in the sorted list. Finally, the label $l_{S_i}$ of snippet $S_{i}$ is computed by", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nl _ {S _ {i}} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\quad \\operatorname {R a n k} \\left(\\hat {s} _ {i}\\right) < T _ {w s}, \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 262, + 785, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $T_{ws}$ is the relative threshold not affected by concrete anomaly scores.", + "bbox": [ + 214, + 308, + 761, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "WS Model (RTFM). With snippet-level pseudo annotations $l_{S}$ (obtained in Eq. 3), we now define positive and negative bags and use them to train a WS model. Previously, approaches [32,34] usually treat snippets of the same video as a bag based on video-level labels. Differently, thanks to the snippet-level labels provided by the wOCC model, we can compose positive and negative bags more flexibly. In our method, a positive bag $B^{+} = \\{S_{i}^{+}\\}_{i=1}^{C}$ is composed of randomly selected $C = 16$ abnormal snippets. Similarly, a negative bag $B^{-} = \\{S_{j}^{-}\\}_{j=1}^{C}$ is composed of randomly selected $C$ normal snippets. Taking RTFM [34] as an example, the WS method employs multiple instance learning, requiring the average Top- $k$ maximum magnitudes of features in positive bags exceeds that in negative bags:", + "bbox": [ + 212, + 324, + 787, + 489 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {w s} = \\max (0, m - d _ {R T F M} (B ^ {+}) + d _ {R T F M} (B ^ {-})) + \\tag {4} \\\\ \\mathcal {L} _ {B C E} \\left(B ^ {+}\\right) + \\mathcal {L} _ {B C E} \\left(B ^ {-}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 325, + 500, + 785, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $m$ is the margin, and $d_{RTFM}(B)$ (where $B$ can be $B^{+}$ or $B^{-}$ ) returns the average of the Top- $k$ maximum magnitudes of features. Finally, $\\mathcal{L}_{BCE}$ is the binary cross-entropy-based classification loss:", + "bbox": [ + 212, + 544, + 787, + 589 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {B C E} (B) = - Y \\log \\left(f _ {R T F M} (B)\\right) + (1 - Y) \\log \\left(1 - f _ {R T F M} (B)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 276, + 599, + 785, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $Y$ is the label of the bag $B$ and $f_{RTFM}(B)$ returns the average of the Top- $k$ maximum anomaly scores. Note that $f_{RTFM}(B)$ is different from $d_{RTFM}(B)$ , where $f_{RTFM}(B)$ returns the anomaly scores while $d_{RTFM}(B)$ returns the magnitude of the features.", + "bbox": [ + 212, + 623, + 787, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pseudo Labels $w_{X}$ from WS. wOCC model needs soft labels $w_{X}$ . With the trained WS model, we compute the anomaly score $\\hat{x}_i$ of object $X_i$ by simply setting it as the anomaly score of the snippet containing the object $X_i$ . Then, the soft label $w_{X_i}$ is directly set as:", + "bbox": [ + 212, + 684, + 787, + 746 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nw _ {X _ {i}} = \\hat {x} _ {i}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 467, + 756, + 785, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\hat{x}_i$ is already in the range of $[0,1]$ , so we do not need to normalize $w_{X_i}$ .", + "bbox": [ + 214, + 779, + 725, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Initializing the Interleaving Training Module. As discussed above, in our proposed interleaving training module, wOCC training relies on pseudo labels generated by the WS model, and conversely, WS training requires pseudo", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Nie et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "labels derived from the wOCC. This egg and chicken problem needs to be solved at the very beginning. Our solution is to provide the wOCC model with randomly generated soft labels to start the interleaving training. Usually, the training dataset contains much more normal samples than abnormal ones, which means most soft labels are around 0 while a small portion of them are near 1. Therefore, we randomly sample the weight $w_{X}$ from the Beta distribution, i.e., $w_{X} \\sim \\mathrm{Beta}(\\alpha, \\beta)$ , where $\\alpha = 1$ and $\\beta = 5$ . The sampled weights $w_{X}$ are mostly around 0, and a small portion of them is close to 1, satisfying the prior about the distribution of normal and abnormal data.", + "bbox": [ + 212, + 146, + 782, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Convergence Analysis. We have discussed the mechanism and the way to start the interleaving training module above. Then, in the following, we empirically analyze its convergence, which is validated by the experiment in Figure 6. At the very beginning, we randomly initialize the weight $w_{X}$ to train our wOCC model. Although the random initialization is not precise, the wOCC model can still learn normal patterns from the training data, as we assume there is more normal data than abnormal data in the training dataset. Therefore, the wOCC model can produce relatively reliable pseudo labels for the WS model. This is very important as the WS model has a high requirement on the quality of the pseudo labels. In turn, with the trained WS model, we can provide better pseudo supervision to train a stronger wOCC model than the random initialization. In this way, we achieve mutual improvement between two models until convergence.", + "bbox": [ + 212, + 282, + 784, + 464 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The above analysis implies the reason why we do not choose to interleave the training of two OCC models or two WS models. The OCC model can provide reliable initialization, while the WS model explicitly enlarges the gap between normal and abnormal samples. Two OCC models fail to utilize the abnormal data, while for two WS models it is not known how to provide a reliable initialization to start the training.", + "bbox": [ + 212, + 465, + 784, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Repeating Procedure with Adaptive Thresholding", + "text_level": 1, + "bbox": [ + 214, + 583, + 671, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the above, the wOCC and WS models need to provide pseudo labels to supervise the other. As the wOCC model relies on soft labels $w_{X}$ as defined in Eq. 6, its pseudo labels can be directly provided by the WS model without thresholding. However, as defined in Eq. 3, we need to specify a threshold for the WS model since it requires binary hard partition labels to construct positive and negative bags. To alleviate the influence of human interaction, we propose a repeating procedure of the interleaving training module which can adaptively refine a user-specified rough threshold to a relatively optimal one.", + "bbox": [ + 212, + 612, + 782, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Starting and running the next interleaving training module. As shown in Figure 1, we run several interleaving training modules one after another. Recall that we provide $w_{X}$ by random initialization to start the first interleaving training module as discussed in Section 3.2. To start the next module, we use the final WS model in the last module to provide the labels $w_{X}$ . For the wOCC and WS models in a new interleaving training module, we train them from scratch, not inheriting parameters from the last module.", + "bbox": [ + 212, + 734, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD", + "bbox": [ + 238, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adaptive Thresholding. The threshold is fixed in each module. Between two modules, we employ an adaptive thresholding mechanism to refine the threshold. As the repeating procedure runs, we obtain a series of thresholds $\\{T_{ws}^{1}, T_{ws}^{2}, \\dots, T_{ws}^{i}, \\dots\\}$ corresponding to the interleaving training modules. The key idea of adaptive thresholding is to ensure:", + "bbox": [ + 215, + 146, + 787, + 223 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nT _ {w s} ^ {1} \\geq T _ {w s} ^ {2} \\geq \\dots \\geq T _ {w s} ^ {i} \\geq \\dots . \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 234, + 785, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Starting from a sufficiently large (larger than the optimal threshold) value of $T_{ws}^{1}$ , there must be one threshold $T_{ws}^{*} \\in \\{T_{ws}^{1}, T_{ws}^{2}, \\dots, T_{ws}^{i}, \\dots\\}$ that is closest to the optimal threshold. The problem is how to enforce the monotonically decreasing of the threshold and how to stop the repeating procedure at the optimal point.", + "bbox": [ + 212, + 265, + 784, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "First, we initialize $T_{ws}^{1} = R\\% \\times N$ , where $R$ is a user-specific hyper-parameter, e.g. $R = 30$ , and $N$ is the total number of snippets in the training dataset. This setting makes $T_{ws}^{1}$ a large enough threshold, since generally the ratio of abnormal data in the training dataset is less than $30\\%$ . We test different $R$ (e.g., $35\\%$ ) in the experimental section and find that it does not influence our final results much, demonstrating the robustness of our method to the user interaction.", + "bbox": [ + 212, + 325, + 784, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Then, to ensure Eq. 7, our adaptive thresholding method determines the threshold used in the current module based on wOCC models trained in all previous modules. Each training module produces multiple wOCC models, obtaining one after each loop between wOCC and WS. Let $M_{i}$ be the number of wOCC models trained from module 1 to module $i$ . When computing $T_{ws}^{i+1}$ for module $i+1$ with $i \\geq 1$ , we take all the $M_{i}$ wOCC models into consideration. Specifically, for each wOCC, e.g., the $j^{th}$ one where $j \\in [1, M_{i}]$ , we use it to identify $R\\%$ of snippets in the training dataset that have high anomaly scores, and denote the resulting set of snippets as $A_{j}$ . We then compute intersection between all sets and count the number of the resulting intersection as the threshold:", + "bbox": [ + 212, + 417, + 787, + 570 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nT _ {w s} ^ {i + 1} = \\operatorname {N u m} \\left(A _ {1} \\cap A _ {2} \\cap \\dots \\cap A _ {M _ {i}}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 582, + 784, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathrm{Num}(\\cdot)$ counts the number of elements in the intersection. The operation of computing intersection ensures that the obtained threshold is monotonically decreased. In essence, the threshold computed in this way means the number of snippets viewed as abnormal by all wOCC models. At the early stages, the number of wOCC models is small, therefore the size of the intersection is large, yielding large thresholds. As more wOCC models involved, it is more difficult to form a consensus, thus producing smaller thresholds.", + "bbox": [ + 212, + 612, + 787, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To elucidate the effect of the adaptive thresholding, we visualize how WS models are gradually improved as the repeating procedure runs in Figure 2. In the figure, we show two examples. Taking the one on the left as an example, it shows a video with an abnormal event in the middle. At the beginning when $T_{ws} = 1643$ , the WS model wrongly identifies the abnormal event as normal. As the repeating procedure runs, $T_{ws}$ becomes smaller and smaller, which is finally reduced to 834 after repeating 5 times, and the WS model begins to identify the anomaly event progressively.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Nie et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/860735d1d9744c90ccc3a68a5c408ca563845b9f3d80e1d6e2fbe130fac10d2f.jpg", + "image_caption": [ + "Fig. 2: As repeating procedure goes, $T_{ws}$ monotonically decreases, improving the WS model gradually. Left: example from ShanghaiTech. Right: example from UBnormal." + ], + "image_footnote": [], + "bbox": [ + 305, + 143, + 496, + 306 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/662a2edb42455b1676dcec073ef668028956d895e7fcafe2b037cf985cbf2f0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 143, + 694, + 306 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Stopping Criterion. As our goal is to find the threshold $T_{ws}^{*}$ close to the optimal threshold, we need to stop the repeating procedure at the right time. Experimentally, we find that $T_{ws}$ drops fast at the first few interleaving training modules, then the rate of change of $T_{ws}$ becomes smaller which indicates that the wOCCs achieve a consensus about anomalies. The size of the consensus set is probably the actual number of abnormal snippets in the dataset. Based on this observation, once change rate of $T_{ws}$ between two consecutive training modules is less than $Q\\%$ of the first change rate at the very beginning, we stop the repeating procedure. Please see the high correlation between the change rate of $T_{ws}$ and the accuracy of the WS model in Figure 4.", + "bbox": [ + 212, + 375, + 787, + 525 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training Time Analysis. Though we need to repeat the interleaving training module multiple times, our training is as fast as the original wOCC or WS method. First, the repeating procedure usually stops with fewer than 6 modules thanks to the stopping criterion. Furthermore, we train wOCC (or WS) for just one epoch, and then exchange to train WS (or wOCC) for another epoch. If we have 6 modules, each running an average of 5 loops of interleaving training, the wOCC and WS models are trained for just 30 epochs each. The original wOCC or WS models are also trained for the similar number of epochs in their default experimental settings. The detailed experiments are in Section 4.5.", + "bbox": [ + 212, + 525, + 787, + 662 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Inference", + "text_level": 1, + "bbox": [ + 214, + 681, + 339, + 695 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "At the inference stage, both wOCC and WS models can be used to detect anomalies. We provide results of both models when compared with previous methods.", + "bbox": [ + 214, + 704, + 785, + 736 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 756, + 377, + 773 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 214, + 786, + 532, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ShanghaiTech. The ShanghaiTech dataset [18] contains 437 videos and was created as a benchmark for OCC with only normal videos available in the training", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD", + "bbox": [ + 238, + 113, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "set. Zhong et al. [53] reorganized the dataset to enable the training of WS systems. The new split contains 63 abnormal videos and 175 normal videos, while there are 44 anomalous and 155 normal videos in the new testing split. In our unsupervised setting, we follow the split for WS approaches but do not provide video-level annotations in the training stage.", + "bbox": [ + 212, + 146, + 782, + 219 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "UBnormal. The UBnormal dataset [1] is a synthetic open-set benchmark, containing 268 training videos, 64 validation videos and 211 test videos. Normal and abnormal videos are mixed in the three subsets. The dataset is challenging because of disjoint sets of anomaly types in training and testing sets. We follow its original organization but train our proposed UVAD without any annotations.", + "bbox": [ + 212, + 222, + 782, + 296 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following prior VAD arts [13, 34], we use the frame-level area under ROC curve (AUC) to measure VAD's accuracy for both datasets.", + "bbox": [ + 214, + 297, + 782, + 327 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 wOCC and WS models Employed", + "text_level": 1, + "bbox": [ + 215, + 352, + 539, + 368 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As stated, our method is a flexible UVAD framework that can use different wOCC and WS models. We test improving three OCC models to wOCC, including the AE model used in [47], Jigsaw [39], and STG-NF [13]. The wOCC losses for $AE$ and Jigsaw are put into supplemental material.", + "bbox": [ + 212, + 378, + 784, + 436 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) AE. GCL [47] proposes to use an AutoEncoder that reconstructs the features extracted from videos as their OCC model. We use it here too.", + "(2) Jigsaw. The OCC model Jigsaw [39] addresses VAD by solving a pretext task: spatio-temporal jigsaw puzzles. It divides the spatio-temporal space of a video into smaller cubes, then shuffles the positions of the cubes, and finally tries to restore the original positions of the cubes.", + "(3) STG-NF. STG-NF [13] extracts people's pose sequences, and extends Glow [16] to build multivariate Gaussian distribution of normal action sequences.", + "For WS, we test the method proposed in Sultani et al. [32] and RTFM [34].", + "(1) Sultani et al. [32]. Sultani et al. [32] propose the first MIL VAD model that maximizes the separability between a positive bag containing snippets of an abnormal video and a negative bag with snippets of a normal video.", + "(2) RTFM. RTFM [34] extends Sultani et al. [32] to compare the Top- $k$ largest-magnitude snippets between positive and negative bags." + ], + "bbox": [ + 214, + 440, + 782, + 651 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 676, + 454, + 691 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We implement our method in PyTorch. For STG-NF and RTFM (implementation details of other wOCC/WS models are put in supplemental material), they are optimized by Adam optimizer with $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999$ , and a weight decay of 0.0005. The batch size of STG-NF is 256, and that of RTFM is 32. The learning rate is $5e - 4$ and $1e - 3$ for STG-NF and RTFM respectively. For RTFM, the margin $m$ is set to 100. During interleaving training, we train a model for one epoch and then exchange to train the other model for another epoch.", + "bbox": [ + 212, + 703, + 782, + 806 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our method has two hyperparameters shared by all types of wOCC or WS models. One is $R\\%$ which determines the initial threshold $T_{ws}^{1}$ . By default, we set", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Nie et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/73fe892bf66a361199e8f434b1eecaf3fad7858fb132253057e8c607845c95a5.jpg", + "table_caption": [ + "Table 1: Comparison with previous approaches. For wOCC, we use STG-NF [13]. For WS, we use RTFM [34]. We test using C3D [35] and I3D [15] to extract features for RTFM. Our $w_{OCC}$ gives the AUC of the wOCC model. Our $WS$ gives the AUC of the WS model. The methods split by the dashed line are trained as a whole in our framework. * means we reimplement the method by taking I3D as the feature. ** means our method degenerated to the supervised OCC or WS method. *** means the results are reproduced by us using the official code." + ], + "table_footnote": [], + "table_body": "
UnsupervisedOne-Class ClassificationWeakly Supervised
MethodFeaturesSTech AUC %UB AUC %MethodSTech AUC %UB AUC %MethodFeaturesSTech AUC %UB AUC %
AEAllDataI3D60.51-MemAE [12]71.20-GCN [53]C3D76.44-
Frame Prediction [18]73.40-GCN [53]TSN84.44-
STG-NFAllData [13]-80.2970.48Markovitz et al. [22]76.1052.00Zhang et al. [50]C3D82.50-
HF2-VAD [19]76.20-Sultani et al.*** [32]I3D84.5354.12
GCL* [47]I3D76.14-CT-D2GAN [9]77.70-AR-Net [38]I3D85.38-
CAC [42]79.30-CLAWS [48]C3D89.67-
OurwOCC-81.5071.67SSMTL [10]82.40-MIST [8]I3D94.83-
OurwSC3D85.4359.05Georgescu et al. [11]82.7059.30Li et al. [17]I3D96.08-
SSMTL++ [4]83.8062.10RTFM*** [34]C3D91.5162.30
OurwOCC-82.5774.76Jigsaw [39]84.3056.40RTFM*** [34]I3D96.1066.83
STG-NF [13]85.9071.80S3R [43]I3D97.48-
OurwSI3D88.1863.10OurwOCC (STG-NF)86.3772.81OurwS (RTFM)I3D96.3367.42
", + "bbox": [ + 225, + 253, + 785, + 386 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/50e40612f53917ec5fef2fbd678112f2589477766aaea2a58d175ea207bce192.jpg", + "table_caption": [ + "Table 2: Ablation study on wOCC im- Table 3: Ablation study on using dif proved from different OCC models, always ferent WS models in our method, always with RTFM as the WS model. with STG-NF as the wOCC model." + ], + "table_footnote": [], + "table_body": "
OCC ModelOurwOCCOurWS(RTFM)GCLClassifierWS ModelOurwOCC(STG-NF)OurWS
AE70.9978.9076.14Sultani et al. [32]81.9277.41
Jigsaw [39]81.2385.35-RTFM [13]82.5788.18
STG-NF [13]82.5788.18-
", + "bbox": [ + 223, + 454, + 766, + 500 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "it to $15\\%$ for the ShanghaiTech dataset which contains a small ratio of abnormal data. The UBnormal dataset contains relatively more abnormal data and we set $R\\%$ to $30\\%$ . The other parameter is $Q\\%$ used to stop the repeating procedure. For both datasets, we set $Q\\%$ to $10\\%$ . Analysis of $Q\\%$ is put into Supp.", + "bbox": [ + 212, + 530, + 787, + 590 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Comparison with Previous Approaches", + "text_level": 1, + "bbox": [ + 214, + 614, + 584, + 631 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "From left to right, Table 1 shows the comparison between our and previous approaches under Unsupervised, One-Class Classification and Weakly-Supervised. Since the unsupervised methods are scarce, besides GCL [47], we compare with $\\mathrm{AE}_{\\text{AllData}}$ which means training the AE model on the whole training dataset containing both normal and abnormal data. We also compare with STG-NF $_{\\text{AllData}}$ . For our method, we adopt STG-NF [13] as wOCC, and RTFM [34] as WS.", + "bbox": [ + 212, + 643, + 787, + 733 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "First, Our $wOCC$ (trained along with RTFM (I3D)) outperforms STG-NFAllData (see the table on the left). This validates the benefit of using our method to distinguish the normal from abnormal data in the training dataset. On ShanghaiTech, our proposed wOCC improves STG-NF from $80.29\\%$ to $82.57\\%$ . In contrast, the improvement is from $70.48\\%$ to $74.76\\%$ on UBnormal. The improvement on UBnormal is larger as there is more abnormal data in UBnormal which causes STG-NFAllData to perform worse.", + "bbox": [ + 212, + 734, + 797, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD", + "bbox": [ + 238, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/e32d1ffcb223522326a9df540ca20ced88ae56082bd13795e972d20db48f3de8.jpg", + "table_caption": [ + "Table 4: Comparison with fixed thresholding." + ], + "table_footnote": [], + "table_body": "
Weighted OCCWS with Adaptive ThresholdRTFM (AUC %)STG-NF (AUC %)
XX82.0680.52
✓X83.4881.78
X✓85.8681.94
✓✓88.1882.57
", + "bbox": [ + 240, + 170, + 758, + 231 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Second, please compare our method with GCL. Our $W_S$ achieves an AUC of 88.18%, while that of GCL is 76.14% (this score is reported by the classifier of GCL, thus we use our WS model for the comparison). This significant improvement is partly due to the reason that we use a stronger OCC model, i.e., STG-NF, than GCL. If using the same OCC model, our method outperforms GCL by 2.76%, as shown in Table 2.", + "bbox": [ + 212, + 258, + 785, + 349 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Our UVAD method can be degenerated into a supervised OCC method. For example, we can perform our interleaving training on the normal data only. The obtained wOCC model in this setting is denoted by $\\mathrm{Our}_{wOCC}^{**}$ which outperforms existing supervised OCC approaches (see the table in the middle). This validates the effectiveness of the proposed wOCC with weighted importance, as even among the normal data there is data that is more normal than other data, and the wOCC model can better identify their difference.", + "bbox": [ + 212, + 349, + 785, + 455 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To degenerate our method into a supervised WS method, we allow our WS model to know video-level labels. Specifically, snippets of normal videos are treated as normal. Since snippets of abnormal videos can be either normal or abnormal, they are processed by our adaptive thresholding approach. As seen in Table 1 on the right, Our $^{**}_{WS}$ (RTFM) outperforms the baseline RTFM [34]. On UBnormal, our method performs best among all the compared WS methods.", + "bbox": [ + 212, + 455, + 785, + 547 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Ablation Study and Discussions", + "text_level": 1, + "bbox": [ + 215, + 571, + 524, + 585 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Next, we perform ablation studies on our method. Without extra explanations, experiments are conducted by taking STG-NF as wOCC and RTFM as WS.", + "bbox": [ + 212, + 597, + 784, + 627 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison between OCC over wOCC. In Figure 3, we compare the performance fluctuation of using wOCC or OCC (STG-NF) in our framework in the first two interleaving training modules' training. As shown, directly using OCC leads to performance fluctuation of both OCC and WS models and converges to an unsatisfactory point. In comparison, our proposed wOCC stabilizes the training and improves the performance of both models.", + "bbox": [ + 212, + 628, + 785, + 717 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Using Different OCC/WS Models. In Table 2, we conduct experiments of upgrading different types of OCC models to wOCC in our method. We test AE [47], Jigsaw [39], and STG-NF [13], while always using RTFM [34] as the WS model. As seen, the effectiveness of our method is essentially affected by wOCC model. Similarly in Table 3, the same wOCC model is used but with different WS models. Overall, a better wOCC or WS model yields better results.", + "bbox": [ + 212, + 718, + 785, + 808 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison with Fixed Thresholding. In Table 4, the first row shows results of using fixed thresholding for both OCC and WS models. The second and", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Nie et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e49694ac09a0b54bf3227bc03b60835fad9bbad271e57af2fb697a0a57382436.jpg", + "image_caption": [ + "Fig. 3: Performance fluctuation in the first Fig. 4: Left: ShanghaiTech. Right: UBnor and second modules. Red line: AUC of mal. Top: $T_{ws}$ in each interleaving trainwOCC (or WS) when interleaving wOCC ing module. Bottom: The AUC of the WS and WS. Blue line: AUC of OCC (or WS) model in each module. Stars in top figures: when interleaving OCC and WS. position where achieves stopping criterion." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 493, + 262 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/87c312a702868bd884db20a804a0401d1327f3b87656714014a95d732e2ce808.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 143, + 638, + 258 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/62b28728f4ba29b746f6044d05fe6177e72bd51e49af8b1fbee6991a62dbe5f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 143, + 779, + 258 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/6efe5da69c35f42f6ec420a733397b6431cbf42fd893bb7dd14fb45d1841e923.jpg", + "table_caption": [ + "Table 5: Training time on ShanghaiTech." + ], + "table_footnote": [], + "table_body": "
ModelEpoch Each StepTraining StepsTraining EpochsTraining TimeSTG-NF AUC %RTFM AUC %
Our UVAD1172×1×172h28m82.5788.18
2122×2×122h42m82.5387.95
3112×3×113h31m82.2786.36
STG-NF [13]--810m85.90-
RTFM [34]--502h32m-97.21
", + "bbox": [ + 277, + 387, + 720, + 460 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "third rows use fixed thresholding for either OCC or WS models. The fourth row shows our method with wOCC and adaptive thresholding. As seen, our method outperforms all the other variants.", + "bbox": [ + 212, + 489, + 784, + 532 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of the Stopping Criterion. In Figure 4, we show that with the current stopping criterion, our method can stop when it achieves the best AUC (i.e., the best VAD accuracy). As the repeating procedure runs, $T_{ws}$ drops and AUC increases. $T_{ws}$ drops very fast at the very beginning, and then the drop rate becomes slower. Simultaneously, AUC first increases and then decreases. We stop our method when the change rate of $T_{ws}$ is about $10\\%$ of its original change rate, which finds the best stopping positions indicated by the stars in top figures.", + "bbox": [ + 212, + 536, + 785, + 642 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Robustness to $R\\%$ . Our method is robust to the parameter $R\\%$ , as demonstrated in Figure 5. On the top, we show that the threshold $T_{ws}$ converges to similar values after 6 interleaving training modules when using different $R\\%$ . At the bottom, we show that different $R\\%$ yield similar AUC on both training datasets. The experiments demonstrate that our method is not influenced much by this parameter for each dataset. For different datasets, we need to set different $R\\%$ that roughly match with the ratio of anomaly data in the datasets.", + "bbox": [ + 212, + 642, + 787, + 747 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Training Loss Curve. In Figure 6, we show training loss curves of the wOCC (STG-NF) model in our framework. As seen, the loss drops smoothly within each interleaving training module. Since we re-train the wOCC model from scratch in the next module, the loss suddenly increases at the beginning of the next module, but the peak magnitude of the loss is lower than that of the previous module. Please check the zoomed-in curve showing the loss curve of the", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD", + "bbox": [ + 238, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/894b7525f4f435893aebefee2473d7936a0170ca0ab8d3170d020cf0945d1248.jpg", + "image_caption": [ + "Fig.5: Top: different $R\\%$ converge to Fig.6: Bottom: training loss curve of the similar $T_{ws}$ (on ShanghaiTech). Bottom: wOCC model during the whole repeating different $R\\%$ yield similar AUC (Left: on procedure. Top: zoomed-in training loss ShanghaiTech, Right: on UBnormal). curve in the third module." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 769, + 292 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "wOCC model in the third module. Even though the model is trained alternately with another model, its loss can still decrease smoothly, not affected by the other model. The WS model's training loss curve is put into the supplemental material.", + "bbox": [ + 212, + 390, + 784, + 436 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Training Time. As mentioned in Section 3.3, although we need to train the wOCC and WS models alternately many times in multiple modules, our method is fast in training. That is because, we just train a model for one epoch and then exchange to train the other model for another epoch, which form a training step of our method. Our interleaving training converges very fast. As shown in Table 5, there are only 17 training steps during the whole training, yielding a total of $34 = 2(\\mathrm{wOCC}$ and $\\mathrm{WS})\\times 1(\\mathrm{epoch})\\times 17$ (all training steps) epochs. It takes around 2.5 hours to train our method on ShanghaiTech. We have also trained the wOCC or WS model for 2 or 3 epochs in each training step, but obtained worse results, probably because training with more epochs would enhance the supervision of wrong pseudo labels in early training modules.", + "bbox": [ + 212, + 436, + 787, + 604 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 633, + 359, + 648 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we interleave the One-Classification and Weakly-Supervised models with adaptive thresholding to tackle the unsupervised video anomaly detection problem without human annotation. We face the challenges of performance fluctuation and the demanding of accurate threshold when designing the framework. To alleviate performance fluctuation, We propose the wOCC model which requires soft labels but not binary labels. To obtain a relative optimal threshold, we repeat our proposed interleaving training module, during which we propose a monotonically threshold decreasing mechanism to refine a rough threshold progressively. Extensive experiments demonstrate the effectiveness of our method. Remarkably, our method can be upgraded with the most recent development in OCC and WS VAD fields.", + "bbox": [ + 212, + 672, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Nie et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was partly supported by the Natural Science Foundation of China (62072191, U21A20520, 62325204), the National Key Research and Development Program of China (2022YFE0112200), the Key-Area Research and Development Program of Guangzhou City (202206030009), and the Guangdong Basic and Applied Basic Research Fund (2023A1515030002, 2024A1515011995).", + "bbox": [ + 215, + 176, + 785, + 253 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 277, + 321, + 292 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Acsintoae, A., Florescu, A., Georgescu, M., Mare, T., Sumedrea, P., Ionescu, R.T., Khan, F.S., Shah, M.: Unbnormal: New benchmark for supervised open-set video anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2022)", + "2. Al-lahham, A., Tastan, N., Zaheer, M.Z., Nandakumar, K.: A coarse-to-fine pseudolabeling (c2fpl) framework for unsupervised video anomaly detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6793-6802 (2024)", + "3. Andrews, S., Tsochantaridis, I., Hofmann, T.: Support vector machines for multiple-instance learning. Advances in neural information processing systems 15 (2002)", + "4. Barbalau, A., Ionescu, R.T., Georgescu, M.I., Dueholm, J., Ramachandra, B., Nasrollahi, K., Khan, F.S., Moeslund, T.B., Shah, M.: Ssmtl++: Revisiting self-supervised multi-task learning for video anomaly detection. Computer Vision and Image Understanding 229, 103656 (2023)", + "5. Basharat, A., Gritai, A., Shah, M.: Learning object motion patterns for anomaly detection and improved object detection. In: 2008 IEEE conference on computer vision and pattern recognition. pp. 1-8. IEEE (2008)", + "6. Burlina, P., Joshi, N., Wang, I., et al.: Where's wally now? deep generative and discriminative embeddings for novelty detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11507-11516 (2019)", + "7. Dietterich, T.G., Lathrop, R.H., Lozano-Pérez, T.: Solving the multiple instance problem with axis-parallel rectangles. Artificial intelligence 89(1-2), 31-71 (1997)", + "8. Feng, J.C., Hong, F.T., Zheng, W.S.: Mist: Multiple instance self-training framework for video anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14009-14018 (2021)", + "9. Feng, X., Song, D., Chen, Y., Chen, Z., Ni, J., Chen, H.: Convolutional transformer based dual discriminator generative adversarial networks for video anomaly detection. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 5546-5554 (2021)", + "10. Georgescu, M.I., Barbalau, A., Ionescu, R.T., Khan, F.S., Popescu, M., Shah, M.: Anomaly detection in video via self-supervised and multi-task learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12742-12752 (2021)", + "11. Georgescu, M.I., Ionescu, R.T., Khan, F.S., Popescu, M., Shah, M.: A background-agnostic framework with adversarial training for abnormal event detection in video. IEEE transactions on pattern analysis and machine intelligence 44(9), 4505-4523 (2021)" + ], + "bbox": [ + 218, + 310, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD", + "bbox": [ + 238, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "12. Gong, D., Liu, L., Le, V., Saha, B., Mansour, M.R., Venkatesh, S., Hengel, A.v.d.: Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1705-1714 (2019)", + "13. Hirschorn, O., Avidan, S.: Normalizing flows for human pose anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13545-13554 (2023)", + "14. Ionescu, R.T., Khan, F.S., Georgescu, M.I., Shao, L.: Object-centric auto-encoders and dummy anomalies for abnormal event detection in video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7842-7851 (2019)", + "15. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017)", + "16. Kingma, D.P., Dhariwal, P.: Glow: Generative flow with invertible 1x1 convolutions. Advances in neural information processing systems 31 (2018)", + "17. Li, S., Liu, F., Jiao, L.: Self-training multi-sequence learning with transformer for weakly supervised video anomaly detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1395-1403 (2022)", + "18. Liu, W., Luo, W., Lian, D., Gao, S.: Future frame prediction for anomaly detection-- a new baseline. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6536-6545 (2018)", + "19. Liu, Z., Nie, Y., Long, C., Zhang, Q., Li, G.: A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13588-13597 (2021)", + "20. Lorre, G., Rabarisoa, J., Orcesi, A., Ainouz, S., Canu, S.: Temporal contrastive pretraining for video action recognition. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 662-670 (2020)", + "21. Lv, H., Yue, Z., Sun, Q., Luo, B., Cui, Z., Zhang, H.: Unbiased multiple instance learning for weakly supervised video anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8022-8031 (June 2023)", + "22. Markovitz, A., Sharir, G., Friedman, I., Zelnik-Manor, L., Avidan, S.: Graph embedded pose clustering for anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10539-10547 (2020)", + "23. Medioni, G., Cohen, I., Brémond, F., Hongeng, S., Nevatia, R.: Event detection and analysis from video streams. IEEE Transactions on pattern analysis and machine intelligence 23(8), 873-889 (2001)", + "24. Morais, R., Le, V., Tran, T., Saha, B., Mansour, M., Venkatesh, S.: Learning regularity in skeleton trajectories for anomaly detection in videos. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11996-12004 (2019)", + "25. Nguyen, T.N., Meunier, J.: Anomaly detection in video sequence with appearance-motion correspondence. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1273-1283 (2019)", + "26. Pang, G., Yan, C., Shen, C., Hengel, A.v.d., Bai, X.: Self-trained deep ordinal regression for end-to-end video anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12173-12182 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Nie et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "27. Park, H., Noh, J., Ham, B.: Learning memory-guided normality for anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14372-14381 (2020)", + "28. Sapkota, H., Yu, Q.: Bayesian nonparametric submodular video partition for robust anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3212-3221 (2022)", + "29. Shi, C., Sun, C., Wu, Y., Jia, Y.: Video anomaly detection via sequentially learning multiple pretext tasks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10330-10340 (2023)", + "30. Smeureanu, S., Ionescu, R.T., Popescu, M., Alexe, B.: Deep appearance features for abnormal behavior detection in video. In: Image Analysis and Processing-ICIAP 2017: 19th International Conference, Catania, Italy, September 11-15, 2017, Proceedings, Part II 19. pp. 779-789. Springer (2017)", + "31. Sohrab, F., Raitoharju, J., Gabbouj, M., Iosifidis, A.: Subspace support vector data description. In: 2018 24th International Conference on Pattern Recognition (ICPR). pp. 722-727. IEEE (2018)", + "32. Sultani, W., Chen, C., Shah, M.: Real-world anomaly detection in surveillance videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6479-6488 (2018)", + "33. Sun, S., Gong, X.: Hierarchical semantic contrast for scene-aware video anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22846-22856 (June 2023)", + "34. Tian, Y., Pang, G., Chen, Y., Singh, R., Verjans, J.W., Carneiro, G.: Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4975-4986 (2021)", + "35. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015)", + "36. Tudor Ionescu, R., Smeureanu, S., Alexe, B., Popescu, M.: Unmasking the abnormal events in video. In: Proceedings of the IEEE international conference on computer vision. pp. 2895-2903 (2017)", + "37. Tur, A.O., Dall'Asen, N., Bryan, C., Ricci, E.: Exploring diffusion models for unsupervised video anomaly detection. In: 2023 IEEE International Conference on Image Processing (ICIP). pp. 2540-2544. IEEE (2023)", + "38. Wan, B., Fang, Y., Xia, X., Mei, J.: Weakly supervised video anomaly detection via center-guided discriminative learning. In: 2020 IEEE international conference on multimedia and expo (ICME). pp. 1-6. IEEE (2020)", + "39. Wang, G., Wang, Y., Qin, J., Zhang, D., Bao, X., Huang, D.: Video anomaly detection by solving decoupled spatio-temporal jigsaw puzzles. In: European Conference on Computer Vision. pp. 494-511. Springer (2022)", + "40. Wang, J., Song, Y., Leung, T., Rosenberg, C., Wang, J., Philbin, J., Chen, B., Wu, Y.: Learning fine-grained image similarity with deep ranking. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1386-1393 (2014)", + "41. Wang, J., Cherian, A.: Gods: Generalized one-class discriminative subspaces for anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8201-8211 (2019)", + "42. Wang, Z., Zou, Y., Zhang, Z.: Cluster attention contrast for video anomaly detection. In: Proceedings of the 28th ACM international conference on multimedia. pp. 2463-2471 (2020)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD", + "bbox": [ + 238, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Wu, J.C., Hsieh, H.Y., Chen, D.J., Fuh, C.S., Liu, T.L.: Self-supervised sparse representation for video anomaly detection. In: European Conference on Computer Vision. pp. 729-745. Springer (2022)", + "44. Wu, P., Liu, J., Shi, Y., Sun, Y., Shao, F., Wu, Z., Yang, Z.: Not only look, but also listen: Learning multimodal violence detection under weak supervision. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16. pp. 322-339. Springer (2020)", + "45. Yan, C., Zhang, S., Liu, Y., Pang, G., Wang, W.: Feature prediction diffusion model for video anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5527-5537 (October 2023)", + "46. Yang, Z., Liu, J., Wu, Z., Wu, P., Liu, X.: Video event restoration based on keyframes for video anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14592-14601 (June 2023)", + "47. Zaheer, M.Z., Mahmood, A., Khan, M.H., Segu, M., Yu, F., Lee, S.I.: Generative cooperative learning for unsupervised video anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14744-14754 (2022)", + "48. Zaheer, M.Z., Mahmood, A., Astrid, M., Lee, S.I.: Claws: Clustering assisted weakly supervised learning with normalcy suppression for anomalous event detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 358-376. Springer (2020)", + "49. Zhang, C., Li, G., Qi, Y., Wang, S., Qing, L., Huang, Q., Yang, M.H.: Exploiting completeness and uncertainty of pseudo labels for weakly supervised video anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16271-16280 (June 2023)", + "50. Zhang, J., Qing, L., Miao, J.: Temporal convolutional network with complementary inner bag loss for weakly supervised anomaly detection. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 4030-4034. IEEE (2019)", + "51. Zhang, T., Lu, H., Li, S.Z.: Learning semantic scene models by object classification and trajectory clustering. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1940-1947. IEEE (2009)", + "52. Zhao, Y., Deng, B., Shen, C., Liu, Y., Lu, H., Hua, X.S.: Spatio-temporal autoencoder for video anomaly detection. In: Proceedings of the 25th ACM international conference on Multimedia. pp. 1933-1941 (2017)", + "53. Zhong, J.X., Li, N., Kong, W., Liu, S., Li, T.H., Li, G.: Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1237-1246 (2019)" + ], + "bbox": [ + 215, + 146, + 784, + 686 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Nie et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection/0ff966a8-e3c5-432a-8beb-e808413549e5_model.json b/2024/Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection/0ff966a8-e3c5-432a-8beb-e808413549e5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8eb158e8fcd5ac4cffd37654b09af8af99d1fdd6 --- /dev/null +++ b/2024/Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection/0ff966a8-e3c5-432a-8beb-e808413549e5_model.json @@ -0,0 +1,2469 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.235, + 0.141, + 0.77, + 0.21 + ], + "angle": 0, + "content": "Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.234, + 0.782, + 0.253 + ], + "angle": 0, + "content": "Yongwei Nie\\(^{1}\\), Hao Huang\\(^{1}\\), Chengjiang Long\\(^{2}\\), Qing Zhang\\(^{3}\\), Pradipta" + }, + { + "type": "text", + "bbox": [ + 0.392, + 0.255, + 0.613, + 0.274 + ], + "angle": 0, + "content": "Maji\\(^{4}\\), and Hongmin Cai\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.284, + 0.66, + 0.299 + ], + "angle": 0, + "content": "\\(^{1}\\) South China University of Technology, China" + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.299, + 0.593, + 0.312 + ], + "angle": 0, + "content": "Meta Reality Labs, USA" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.312, + 0.61, + 0.326 + ], + "angle": 0, + "content": "3 Sun Yat-sen University, China" + }, + { + "type": "text", + "bbox": [ + 0.384, + 0.326, + 0.62, + 0.34 + ], + "angle": 0, + "content": "\\(^{4}\\) Indian Statistical Institute, India" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.367, + 0.744, + 0.701 + ], + "angle": 0, + "content": "Abstract. Video Anomaly Detection (VAD) has been extensively studied under the settings of One-Class Classification (OCC) and Weakly-Supervised learning (WS), which however both require laborious human-annotated normal/abnormal labels. In this paper, we study Unsupervised VAD (UVAD) that does not depend on any label by combining OCC and WS into a unified training framework. Specifically, we extend OCC to weighted OCC (wOCC) and propose a wOCC-WS interleaving training module, where the two models automatically generate pseudo-labels for each other. We face two challenges to make the combination effective: (1) Models' performance fluctuates occasionally during the training process due to the inevitable randomness of the pseudo labels. (2) Thresholds are needed to divide pseudo labels, making the training depend on the accuracy of user intervention. For the first problem, we propose to use wOCC requiring soft labels instead of OCC trained with hard zero/one labels, as soft labels exhibit high consistency throughout different training cycles while hard labels are prone to sudden changes. For the second problem, we repeat the interleaving training module multiple times, during which we propose an adaptive thresholding strategy that can progressively refine a rough threshold to a relatively optimal threshold, which reduces the influence of user interaction. A benefit of employing OCC and WS methods to compose a UVAD method is that we can incorporate the most recent OCC or WS model into our framework. Experiments demonstrate the effectiveness of the proposed UVAD framework. Our code is available at https://github.com/benedictstar/Joint-VAD." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.712, + 0.741, + 0.741 + ], + "angle": 0, + "content": "Keywords: Video anomaly detection \\(\\cdot\\) One-class classification \\(\\cdot\\) Weakly-supervised learning \\(\\cdot\\) Unsupervised learning" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.761, + 0.377, + 0.777 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.788, + 0.789, + 0.819 + ], + "angle": 0, + "content": "Video Anomaly Detection (VAD) is a task that identifies abnormal events in a video, where the abnormal event could be a fire alarm, a flaw in an industrial" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.826, + 0.694, + 0.84 + ], + "angle": 0, + "content": "* Hongmin Cai is the corresponding author. Email: hmcai@scut.edu.cn" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Nie et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "product, or a traffic accident, etc. Most previous VAD methods fall in two categories: One-Class Classification (OCC) methods [12,13,18,19,33,39] and Weakly-Supervised (WS) approaches [21,32,34,38,43,49]. Both kinds of approaches require human-annotated labels for training1. Supervised VAD methods have two prominent problems. First, abnormal events are sparse and difficult to recognize, making the annotation of the training dataset a highly labor-intensive task. Secondly, the categories of abnormal events are unbounded, while human can only collect a limited scope of them, yielding the risk that the trained VAD methods cannot handle unobserved abnormal events." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.797, + 0.42 + ], + "angle": 0, + "content": "Considering the above problems, this paper studies Unsupervised VAD (UVAD) that does not rely on labels. Despite the importance of UVAD, few methods have been developed except [47] and its subsequent works [2, 37] to the best of our knowledge. To achieve unsupervised learning, the method of [47] adversarially trains a generator and a discriminator, where the generator is an autoencoder-based reconstruction model and the discriminator is a fully-connected (FC) binary classifier. While the proposed network architecture is sophisticated, the adopted autoencoder and FC networks are relatively weak which limit the performance of the approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.788, + 0.542 + ], + "angle": 0, + "content": "In this paper, we propose a new UVAD framework. Our key idea is that the techniques of OCC and WS approaches are rapidly developed, while a UVAD method is usually implemented by training two VAD models alternately generating pseudo labels for each other. This motivates us to propose a UVAD framework that directly interleaves the training of a pair of OCC and WS models. This allows us to incorporate the recent advances in the two hot research fields to implement a UVAD method. We have also attempted to combine two OCC or two WS models, but found two homogeneous models are less effective." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.785 + ], + "angle": 0, + "content": "While combining OCC and WS methods, we encounter two problems preventing it from effective. First, we observe the OCC or WS model's performance fluctuates occasionally during the training process, which affects the final accuracy they can achieve. This is partly because the pseudo labels generated by each other are changed frequently, making the training not stable. To solve this problem, we extend OCC to weighted OCC (wOCC), and propose a wOCC-WS interleaving training module. Previous OCC model is trained on normal data, while our wOCC model is trained with soft labels in the range of [0,1]. The soft labels are more consistent across adjacent training cycles, making the performance of wOCC more stable than OCC under changing pseudo labels, thus suppressing the fluctuation effect. Second, we still have to set a threshold to partition normal and abnormal pseudo labels for the WS model, making our method susceptible to a user-provided hyperparameter. To reduce the influence of user interaction, we repeat the interleaving training module multiple times, during which we propose an adaptive threshold mechanism that finds an optimal threshold given just a rough initial value from the user." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.785, + 0.614, + 0.8 + ], + "angle": 0, + "content": "To conclude, the contributions of this work include:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "1 OCC models are trained on normal data only, which requires human labor to exclude abnormal events in preparing the training dataset." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.784, + 0.175 + ], + "angle": 0, + "content": "- We propose a new UVAD framework that interleaves the training of a pair of OCC and WS models." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.178, + 0.785, + 0.221 + ], + "angle": 0, + "content": "- We improve OCC to wOCC which stabilizes the training of the UVAD framework, and also propose an adaptive thresholding strategy to alleviate the influence of user interaction." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.223, + 0.784, + 0.266 + ], + "angle": 0, + "content": "- Our UVAD framework is flexible to employ different pairs of OCC and WS models. Experiments on three OCC and two WS models demonstrate the effectiveness of our method." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.785, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.293, + 0.388, + 0.309 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.786, + 0.461 + ], + "angle": 0, + "content": "Unsupervised VAD (UVAD). Unsupervised VAD requires identifying anomalies from data containing both normal and abnormal samples without any annotation. This is a challenging new task that is almost unexplored in the literature. Although significant progress has been made in OCC and WS VAD, directly applying them independently to address UVAD does not yield good results. Zaheer et al. [47] firstly raise this task and propose to solve UVAD by generative cooperative learning (GCL). Differently, our proposed method modifies both the OCC and WS VAD models and combines the modified two models together to learn from unmarked training data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.787, + 0.689 + ], + "angle": 0, + "content": "One-Class Classification VAD (OCC). OCC-based VAD approaches only have access to normal data and try to model normal data to identify behaviors that are significantly different from normal behaviors as anomalies. Early works use hand-crafted features to help detect anomalies [5,23,40,51]. With the rapid development of deep learning, recent methods turn to normal representations extracted by using a deep neural network [26, 30, 36, 41]. Some methods identify normal patterns by using the reconstruction/prediction model to reconstruct the representations [6,12,14,18,19,24,25,27,31,33,45,46,52]. These models may lead to well-reconstructed anomalies, thereby limiting the performance of the OCC-based methods. Other OCC-based methods turn to identify normal representations by using proxy tasks [4,11,20,29,39]. What's more, recently, Hirschorn and Avidan [13] propose to build a multivariate Gaussian distribution for normal data and detect instances deviating from this distribution as anomalies. However, the above methods treat all samples equally, ignoring the potential differences in importance between normal samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Weakly-Supervised VAD (WS). In weakly-supervised VAD, video-level annotations are available in the training stage. Sultani et al [32] first propose to use the video-level labels and solve WS VAD based on multi-instance learning (MIL) framework [3, 7]. Since then, many works [21, 28, 38, 43, 49, 50] have viewed the WS VAD task as a MIL problem. Tian et al. [34] develop to extend MIL to Top- \\(k\\) MIL method by training with a robust temporal feature magnitude (RTFM) loss function. The methods mentioned above directly train in the supervision of video-level labels and the coarse-grained supervision limits the accuracy of WS models. Recently, two-stage self-training methods [8, 17, 44, 53] adopt a two-stage pipeline to use more fine-grained labels to supervise the networks more" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Nie et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.784, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.457 + ], + "angle": 0, + "content": "Fig. 1: Pipeline of our proposed interleaved framework. The wOCC-WS interleaving training module comprises a weighted OCC (wOCC) model and a WS model, using the output of each model as the pseudo-labels to supervise the other model. Further, the training module is repeated multiple times with the adaptive threshold until the stopping criterion we set. To start the first training module, we randomly sample \\( w_{X} \\) to train the wOCC model and assign a large value as the threshold for the WS model. Once a training module converges, the last WS model of it supplies \\( w_{X} \\) to start the next training module and adaptive thresholding takes the outputs of all previous wOCC models into consideration for updating the threshold." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.787, + 0.518 + ], + "angle": 0, + "content": "strictly. The WS model in our framework shares a similar idea with [8] and is tailored to use finer-grained snippet-level labels." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.541, + 0.331, + 0.557 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.573, + 0.342, + 0.587 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.787, + 0.674 + ], + "angle": 0, + "content": "Our proposed UVAD framework interleaves the training of a pair of OCC and WS models. In theory, any OCC or WS methods can be incorporated into our framework. Without loss of generality, we take STG-NF [13] (a recent OCC model) and RTFM [34] (a recent WS model) as examples to introduce our method in the following sections. We test many other OCC/WS models in Section 4.5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Figure 1 shows an overview of our framework. The basic buildingblock of our method is the wOCC-WS interleaving training module. In the module, the outputs of each model are used as pseudo labels for defining the loss function training the other model. After one module converges, we re-initialize another module to run the interleaving training again with a difference that the threshold of the WS model is updated by an adaptive threshold adjustment mechanism based on wOCC models trained in previous modules. Since wOCC has more stable behaviors than WS, we always initialize the wOCC model to start the training of a module. For the first module, we use a random initialization based on Beta distribution sampling. For other modules, we use the results of WS of the preceding module for the initialization." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.239, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.603, + 0.163 + ], + "angle": 0, + "content": "3.2 wOCC-WS Interleaving Training Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.261 + ], + "angle": 0, + "content": "We first introduce the wOCC-WS interleaving training module which comprises a wOCC model and a WS model. At the very beginning, we initialize the wOCC model to start the interleaving training. Then, pseudo labels created using wOCC are used to train the WS model. Next, pseudo labels created by the WS model are used to improve the wOCC model. The whole training does not use human annotations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.262, + 0.789, + 0.444 + ], + "angle": 0, + "content": "wOCC Model (STG-NF). Typically, OCC models are trained only on normal data. Directly using OCC in our UVAD framework may make the training fluctuate, i.e., the performance of the trained models sometimes gets better, but sometimes worse, finally converging to an inferior point. We use 0 to indicate normal samples, and 1 abnormal samples, where 0, 1 are hard binary labels. During training, we need to identify 0-labelled samples to train the OCC model. The hard partitioned labels may be suddenly changed from 0 to 1 or from 1 to 0, yielding the frequent change of the normal samples used to train the OCC model which is the source of the unstable training. To this end, we propose wOCC which relies on soft labels instead of hard binary labels. Soft labels are in the range of [0, 1]. For example, the soft label 0.7 may be changed to 0.6 but not 0 suddenly. This consistency reduces the fluctuation of the interleaving training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.788, + 0.52 + ], + "angle": 0, + "content": "Taking STG-NF [13] (an OCC model) as an example, we introduce how to improve it to wOCC. STG-NF extracts a set of objects \\(\\mathcal{X} = \\{X_1,\\dots ,X_i,\\dots \\}\\) from the training videos, where \\(X_{i}\\) is a sequence of poses of a person. Each object \\(X_{i}\\) is the basic building block processed by STG-NF. Originally, STG-NF minimizes the negative log-likelihood across the normal data:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.532, + 0.785, + 0.549 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {o c c} = - \\log \\left(p _ {Z} \\left(f _ {S T G - N F} \\left(X _ {i} ^ {+}\\right)\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.558, + 0.788, + 0.651 + ], + "angle": 0, + "content": "where \\( X_{i}^{+} \\in \\mathcal{X}^{+} \\) is sampled from normal data, \\( Z = f_{STG-NF}(X_{i}^{+}) \\) is the feature extracted from \\( X_{i}^{+} \\) by the STG-NF, \\( p_{Z} \\) is the established distribution of the normal data. We upgrade STG-NF by introducing the soft labels \\( w_{X} = \\{w_{X_{1}}, \\dots, w_{X_{i}}, \\dots\\} \\) as the weight to supervise its training on both normal and abnormal data. Formally, for the wOCC model improved from STG-NF, we minimize the following weighted negative log-likelihood across all the data:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.663, + 0.785, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {w o c c} = - \\left(1 - w _ {X _ {i}}\\right) \\log \\left(p _ {Z} \\left(f _ {S T G - N F} \\left(X _ {i}\\right)\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.749 + ], + "angle": 0, + "content": "where \\( X_{i} \\in \\mathcal{X} \\) is sampled from the whole training dataset rather than the normal part and \\( w_{X_i} \\in [0,1] \\) is the weight of the object \\( X_{i} \\) indicating the anomaly degree of the object which serves as the soft label. The object with smaller \\( w_{X_i} \\) means more normal which is better captured by the learned distribution of wOCC." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Pseudo Labels \\( l_{S} \\) from wOCC. Now we use the trained wOCC model to generate pseudo labels \\( l_{S} = \\{l_{S_{1}},\\dots ,l_{S_{i}},\\dots \\} \\) for training the WS model. \"Snippet\" is the basic element processed by WS models. We first split the training videos into snippets, obtaining a set of snippets \\( S = \\{S_1,\\dots ,S_i,\\dots \\} \\). Our aim is to compute the label \\( l_{S_i}\\in \\{0,1\\} \\) for each snippet \\( S_{i} \\). Let \\( x_{i} \\) be the anomaly score of the object \\( X_{i} \\) computed by the trained wOCC model. The anomaly score \\( \\hat{s}_i \\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "Y. Nie et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "of a video snippet \\( S_{i} \\) is computed by averaging all the anomaly scores of objects in the snippet. To generate pseudo label \\( l_{S_i} \\), one way is to check whether \\( \\hat{s}_i \\) is larger than a threshold \\( (l_{S_i} = 1) \\) or not \\( (l_{S_i} = 0) \\). However, as the training goes, the anomaly scores of snippets may vary in a large range, making the absolute threshold invalid. Instead, we sort the snippets according to their anomaly scores in descending order, and get the index of the snippet \\( S_{i} \\) by \\( \\mathrm{Rank}(\\hat{s}_i) \\) in the sorted list. Finally, the label \\( l_{S_i} \\) of snippet \\( S_{i} \\) is computed by" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.263, + 0.786, + 0.302 + ], + "angle": 0, + "content": "\\[\nl _ {S _ {i}} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\quad \\operatorname {R a n k} \\left(\\hat {s} _ {i}\\right) < T _ {w s}, \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.309, + 0.762, + 0.324 + ], + "angle": 0, + "content": "where \\( T_{ws} \\) is the relative threshold not affected by concrete anomaly scores." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.325, + 0.789, + 0.491 + ], + "angle": 0, + "content": "WS Model (RTFM). With snippet-level pseudo annotations \\( l_{S} \\) (obtained in Eq. 3), we now define positive and negative bags and use them to train a WS model. Previously, approaches [32,34] usually treat snippets of the same video as a bag based on video-level labels. Differently, thanks to the snippet-level labels provided by the wOCC model, we can compose positive and negative bags more flexibly. In our method, a positive bag \\( B^{+} = \\{S_{i}^{+}\\}_{i=1}^{C} \\) is composed of randomly selected \\( C = 16 \\) abnormal snippets. Similarly, a negative bag \\( B^{-} = \\{S_{j}^{-}\\}_{j=1}^{C} \\) is composed of randomly selected \\( C \\) normal snippets. Taking RTFM [34] as an example, the WS method employs multiple instance learning, requiring the average Top-\\( k \\) maximum magnitudes of features in positive bags exceeds that in negative bags:" + }, + { + "type": "equation", + "bbox": [ + 0.326, + 0.5, + 0.786, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {w s} = \\max (0, m - d _ {R T F M} (B ^ {+}) + d _ {R T F M} (B ^ {-})) + \\tag {4} \\\\ \\mathcal {L} _ {B C E} \\left(B ^ {+}\\right) + \\mathcal {L} _ {B C E} \\left(B ^ {-}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.545, + 0.788, + 0.59 + ], + "angle": 0, + "content": "where \\(m\\) is the margin, and \\(d_{RTFM}(B)\\) (where \\(B\\) can be \\(B^{+}\\) or \\(B^{-}\\)) returns the average of the Top-\\(k\\) maximum magnitudes of features. Finally, \\(\\mathcal{L}_{BCE}\\) is the binary cross-entropy-based classification loss:" + }, + { + "type": "equation", + "bbox": [ + 0.277, + 0.601, + 0.786, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {B C E} (B) = - Y \\log \\left(f _ {R T F M} (B)\\right) + (1 - Y) \\log \\left(1 - f _ {R T F M} (B)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.788, + 0.683 + ], + "angle": 0, + "content": "where \\( Y \\) is the label of the bag \\( B \\) and \\( f_{RTFM}(B) \\) returns the average of the Top- \\( k \\) maximum anomaly scores. Note that \\( f_{RTFM}(B) \\) is different from \\( d_{RTFM}(B) \\), where \\( f_{RTFM}(B) \\) returns the anomaly scores while \\( d_{RTFM}(B) \\) returns the magnitude of the features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.685, + 0.788, + 0.747 + ], + "angle": 0, + "content": "Pseudo Labels \\( w_{X} \\) from WS. wOCC model needs soft labels \\( w_{X} \\). With the trained WS model, we compute the anomaly score \\( \\hat{x}_i \\) of object \\( X_i \\) by simply setting it as the anomaly score of the snippet containing the object \\( X_i \\). Then, the soft label \\( w_{X_i} \\) is directly set as:" + }, + { + "type": "equation", + "bbox": [ + 0.468, + 0.757, + 0.786, + 0.772 + ], + "angle": 0, + "content": "\\[\nw _ {X _ {i}} = \\hat {x} _ {i}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.726, + 0.795 + ], + "angle": 0, + "content": "\\(\\hat{x}_i\\) is already in the range of \\([0,1]\\), so we do not need to normalize \\(w_{X_i}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Initializing the Interleaving Training Module. As discussed above, in our proposed interleaving training module, wOCC training relies on pseudo labels generated by the WS model, and conversely, WS training requires pseudo" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.283 + ], + "angle": 0, + "content": "labels derived from the wOCC. This egg and chicken problem needs to be solved at the very beginning. Our solution is to provide the wOCC model with randomly generated soft labels to start the interleaving training. Usually, the training dataset contains much more normal samples than abnormal ones, which means most soft labels are around 0 while a small portion of them are near 1. Therefore, we randomly sample the weight \\(w_{X}\\) from the Beta distribution, i.e., \\(w_{X} \\sim \\mathrm{Beta}(\\alpha, \\beta)\\), where \\(\\alpha = 1\\) and \\(\\beta = 5\\). The sampled weights \\(w_{X}\\) are mostly around 0, and a small portion of them is close to 1, satisfying the prior about the distribution of normal and abnormal data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.284, + 0.785, + 0.465 + ], + "angle": 0, + "content": "Convergence Analysis. We have discussed the mechanism and the way to start the interleaving training module above. Then, in the following, we empirically analyze its convergence, which is validated by the experiment in Figure 6. At the very beginning, we randomly initialize the weight \\( w_{X} \\) to train our wOCC model. Although the random initialization is not precise, the wOCC model can still learn normal patterns from the training data, as we assume there is more normal data than abnormal data in the training dataset. Therefore, the wOCC model can produce relatively reliable pseudo labels for the WS model. This is very important as the WS model has a high requirement on the quality of the pseudo labels. In turn, with the trained WS model, we can provide better pseudo supervision to train a stronger wOCC model than the random initialization. In this way, we achieve mutual improvement between two models until convergence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.785, + 0.556 + ], + "angle": 0, + "content": "The above analysis implies the reason why we do not choose to interleave the training of two OCC models or two WS models. The OCC model can provide reliable initialization, while the WS model explicitly enlarges the gap between normal and abnormal samples. Two OCC models fail to utilize the abnormal data, while for two WS models it is not known how to provide a reliable initialization to start the training." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.584, + 0.673, + 0.599 + ], + "angle": 0, + "content": "3.3 Repeating Procedure with Adaptive Thresholding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.784, + 0.734 + ], + "angle": 0, + "content": "In the above, the wOCC and WS models need to provide pseudo labels to supervise the other. As the wOCC model relies on soft labels \\(w_{X}\\) as defined in Eq. 6, its pseudo labels can be directly provided by the WS model without thresholding. However, as defined in Eq. 3, we need to specify a threshold for the WS model since it requires binary hard partition labels to construct positive and negative bags. To alleviate the influence of human interaction, we propose a repeating procedure of the interleaving training module which can adaptively refine a user-specified rough threshold to a relatively optimal one." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Starting and running the next interleaving training module. As shown in Figure 1, we run several interleaving training modules one after another. Recall that we provide \\( w_{X} \\) by random initialization to start the first interleaving training module as discussed in Section 3.2. To start the next module, we use the final WS model in the last module to provide the labels \\( w_{X} \\). For the wOCC and WS models in a new interleaving training module, we train them from scratch, not inheriting parameters from the last module." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "Y. Nie et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "Adaptive Thresholding. The threshold is fixed in each module. Between two modules, we employ an adaptive thresholding mechanism to refine the threshold. As the repeating procedure runs, we obtain a series of thresholds \\(\\{T_{ws}^{1}, T_{ws}^{2}, \\dots, T_{ws}^{i}, \\dots\\}\\) corresponding to the interleaving training modules. The key idea of adaptive thresholding is to ensure:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.236, + 0.786, + 0.254 + ], + "angle": 0, + "content": "\\[\nT _ {w s} ^ {1} \\geq T _ {w s} ^ {2} \\geq \\dots \\geq T _ {w s} ^ {i} \\geq \\dots . \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.266, + 0.785, + 0.327 + ], + "angle": 0, + "content": "Starting from a sufficiently large (larger than the optimal threshold) value of \\( T_{ws}^{1} \\), there must be one threshold \\( T_{ws}^{*} \\in \\{T_{ws}^{1}, T_{ws}^{2}, \\dots, T_{ws}^{i}, \\dots\\} \\) that is closest to the optimal threshold. The problem is how to enforce the monotonically decreasing of the threshold and how to stop the repeating procedure at the optimal point." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.785, + 0.417 + ], + "angle": 0, + "content": "First, we initialize \\( T_{ws}^{1} = R\\% \\times N \\), where \\( R \\) is a user-specific hyper-parameter, e.g. \\( R = 30 \\), and \\( N \\) is the total number of snippets in the training dataset. This setting makes \\( T_{ws}^{1} \\) a large enough threshold, since generally the ratio of abnormal data in the training dataset is less than \\( 30\\% \\). We test different \\( R \\) (e.g., \\( 35\\% \\)) in the experimental section and find that it does not influence our final results much, demonstrating the robustness of our method to the user interaction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.571 + ], + "angle": 0, + "content": "Then, to ensure Eq. 7, our adaptive thresholding method determines the threshold used in the current module based on wOCC models trained in all previous modules. Each training module produces multiple wOCC models, obtaining one after each loop between wOCC and WS. Let \\( M_{i} \\) be the number of wOCC models trained from module 1 to module \\( i \\). When computing \\( T_{ws}^{i+1} \\) for module \\( i+1 \\) with \\( i \\geq 1 \\), we take all the \\( M_{i} \\) wOCC models into consideration. Specifically, for each wOCC, e.g., the \\( j^{th} \\) one where \\( j \\in [1, M_{i}] \\), we use it to identify \\( R\\% \\) of snippets in the training dataset that have high anomaly scores, and denote the resulting set of snippets as \\( A_{j} \\). We then compute intersection between all sets and count the number of the resulting intersection as the threshold:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.583, + 0.785, + 0.601 + ], + "angle": 0, + "content": "\\[\nT _ {w s} ^ {i + 1} = \\operatorname {N u m} \\left(A _ {1} \\cap A _ {2} \\cap \\dots \\cap A _ {M _ {i}}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.718 + ], + "angle": 0, + "content": "where \\(\\mathrm{Num}(\\cdot)\\) counts the number of elements in the intersection. The operation of computing intersection ensures that the obtained threshold is monotonically decreased. In essence, the threshold computed in this way means the number of snippets viewed as abnormal by all wOCC models. At the early stages, the number of wOCC models is small, therefore the size of the intersection is large, yielding large thresholds. As more wOCC models involved, it is more difficult to form a consensus, thus producing smaller thresholds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To elucidate the effect of the adaptive thresholding, we visualize how WS models are gradually improved as the repeating procedure runs in Figure 2. In the figure, we show two examples. Taking the one on the left as an example, it shows a video with an abnormal event in the middle. At the beginning when \\( T_{ws} = 1643 \\), the WS model wrongly identifies the abnormal event as normal. As the repeating procedure runs, \\( T_{ws} \\) becomes smaller and smaller, which is finally reduced to 834 after repeating 5 times, and the WS model begins to identify the anomaly event progressively." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.144, + 0.497, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.144, + 0.696, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.319, + 0.785, + 0.348 + ], + "angle": 0, + "content": "Fig. 2: As repeating procedure goes, \\( T_{ws} \\) monotonically decreases, improving the WS model gradually. Left: example from ShanghaiTech. Right: example from UBnormal." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.526 + ], + "angle": 0, + "content": "Stopping Criterion. As our goal is to find the threshold \\( T_{ws}^{*} \\) close to the optimal threshold, we need to stop the repeating procedure at the right time. Experimentally, we find that \\( T_{ws} \\) drops fast at the first few interleaving training modules, then the rate of change of \\( T_{ws} \\) becomes smaller which indicates that the wOCCs achieve a consensus about anomalies. The size of the consensus set is probably the actual number of abnormal snippets in the dataset. Based on this observation, once change rate of \\( T_{ws} \\) between two consecutive training modules is less than \\( Q\\% \\) of the first change rate at the very beginning, we stop the repeating procedure. Please see the high correlation between the change rate of \\( T_{ws} \\) and the accuracy of the WS model in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.788, + 0.664 + ], + "angle": 0, + "content": "Training Time Analysis. Though we need to repeat the interleaving training module multiple times, our training is as fast as the original wOCC or WS method. First, the repeating procedure usually stops with fewer than 6 modules thanks to the stopping criterion. Furthermore, we train wOCC (or WS) for just one epoch, and then exchange to train WS (or wOCC) for another epoch. If we have 6 modules, each running an average of 5 loops of interleaving training, the wOCC and WS models are trained for just 30 epochs each. The original wOCC or WS models are also trained for the similar number of epochs in their default experimental settings. The detailed experiments are in Section 4.5." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.682, + 0.341, + 0.696 + ], + "angle": 0, + "content": "3.4 Inference" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.705, + 0.787, + 0.737 + ], + "angle": 0, + "content": "At the inference stage, both wOCC and WS models can be used to detect anomalies. We provide results of both models when compared with previous methods." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.757, + 0.378, + 0.775 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.787, + 0.533, + 0.802 + ], + "angle": 0, + "content": "4.1 Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "ShanghaiTech. The ShanghaiTech dataset [18] contains 437 videos and was created as a benchmark for OCC with only normal videos available in the training" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "Y. Nie et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.22 + ], + "angle": 0, + "content": "set. Zhong et al. [53] reorganized the dataset to enable the training of WS systems. The new split contains 63 abnormal videos and 175 normal videos, while there are 44 anomalous and 155 normal videos in the new testing split. In our unsupervised setting, we follow the split for WS approaches but do not provide video-level annotations in the training stage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.784, + 0.297 + ], + "angle": 0, + "content": "UBnormal. The UBnormal dataset [1] is a synthetic open-set benchmark, containing 268 training videos, 64 validation videos and 211 test videos. Normal and abnormal videos are mixed in the three subsets. The dataset is challenging because of disjoint sets of anomaly types in training and testing sets. We follow its original organization but train our proposed UVAD without any annotations." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.299, + 0.784, + 0.328 + ], + "angle": 0, + "content": "Evaluation Metrics. Following prior VAD arts [13, 34], we use the frame-level area under ROC curve (AUC) to measure VAD's accuracy for both datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.353, + 0.54, + 0.369 + ], + "angle": 0, + "content": "4.2 wOCC and WS models Employed" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.785, + 0.438 + ], + "angle": 0, + "content": "As stated, our method is a flexible UVAD framework that can use different wOCC and WS models. We test improving three OCC models to wOCC, including the AE model used in [47], Jigsaw [39], and STG-NF [13]. The wOCC losses for \\(AE\\) and Jigsaw are put into supplemental material." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.441, + 0.784, + 0.467 + ], + "angle": 0, + "content": "(1) AE. GCL [47] proposes to use an AutoEncoder that reconstructs the features extracted from videos as their OCC model. We use it here too." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.471, + 0.784, + 0.529 + ], + "angle": 0, + "content": "(2) Jigsaw. The OCC model Jigsaw [39] addresses VAD by solving a pretext task: spatio-temporal jigsaw puzzles. It divides the spatio-temporal space of a video into smaller cubes, then shuffles the positions of the cubes, and finally tries to restore the original positions of the cubes." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.531, + 0.784, + 0.56 + ], + "angle": 0, + "content": "(3) STG-NF. STG-NF [13] extracts people's pose sequences, and extends Glow [16] to build multivariate Gaussian distribution of normal action sequences." + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.562, + 0.784, + 0.575 + ], + "angle": 0, + "content": "For WS, we test the method proposed in Sultani et al. [32] and RTFM [34]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.578, + 0.784, + 0.62 + ], + "angle": 0, + "content": "(1) Sultani et al. [32]. Sultani et al. [32] propose the first MIL VAD model that maximizes the separability between a positive bag containing snippets of an abnormal video and a negative bag with snippets of a normal video." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.623, + 0.784, + 0.652 + ], + "angle": 0, + "content": "(2) RTFM. RTFM [34] extends Sultani et al. [32] to compare the Top-\\(k\\) largest-magnitude snippets between positive and negative bags." + }, + { + "type": "list", + "bbox": [ + 0.215, + 0.441, + 0.784, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.677, + 0.455, + 0.692 + ], + "angle": 0, + "content": "4.3 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.783, + 0.807 + ], + "angle": 0, + "content": "We implement our method in PyTorch. For STG-NF and RTFM (implementation details of other wOCC/WS models are put in supplemental material), they are optimized by Adam optimizer with \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.999\\), and a weight decay of 0.0005. The batch size of STG-NF is 256, and that of RTFM is 32. The learning rate is \\(5e - 4\\) and \\(1e - 3\\) for STG-NF and RTFM respectively. For RTFM, the margin \\(m\\) is set to 100. During interleaving training, we train a model for one epoch and then exchange to train the other model for another epoch." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Our method has two hyperparameters shared by all types of wOCC or WS models. One is \\( R\\% \\) which determines the initial threshold \\( T_{ws}^{1} \\). By default, we set" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.243 + ], + "angle": 0, + "content": "Table 1: Comparison with previous approaches. For wOCC, we use STG-NF [13]. For WS, we use RTFM [34]. We test using C3D [35] and I3D [15] to extract features for RTFM. Our \\(w_{OCC}\\) gives the AUC of the wOCC model. Our \\(WS\\) gives the AUC of the WS model. The methods split by the dashed line are trained as a whole in our framework. * means we reimplement the method by taking I3D as the feature. ** means our method degenerated to the supervised OCC or WS method. *** means the results are reproduced by us using the official code." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.254, + 0.787, + 0.387 + ], + "angle": 0, + "content": "
UnsupervisedOne-Class ClassificationWeakly Supervised
MethodFeaturesSTech AUC %UB AUC %MethodSTech AUC %UB AUC %MethodFeaturesSTech AUC %UB AUC %
AEAllDataI3D60.51-MemAE [12]71.20-GCN [53]C3D76.44-
Frame Prediction [18]73.40-GCN [53]TSN84.44-
STG-NFAllData [13]-80.2970.48Markovitz et al. [22]76.1052.00Zhang et al. [50]C3D82.50-
HF2-VAD [19]76.20-Sultani et al.*** [32]I3D84.5354.12
GCL* [47]I3D76.14-CT-D2GAN [9]77.70-AR-Net [38]I3D85.38-
CAC [42]79.30-CLAWS [48]C3D89.67-
OurwOCC-81.5071.67SSMTL [10]82.40-MIST [8]I3D94.83-
OurwSC3D85.4359.05Georgescu et al. [11]82.7059.30Li et al. [17]I3D96.08-
SSMTL++ [4]83.8062.10RTFM*** [34]C3D91.5162.30
OurwOCC-82.5774.76Jigsaw [39]84.3056.40RTFM*** [34]I3D96.1066.83
STG-NF [13]85.9071.80S3R [43]I3D97.48-
OurwSI3D88.1863.10OurwOCC (STG-NF)86.3772.81OurwS (RTFM)I3D96.3367.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.401, + 0.782, + 0.442 + ], + "angle": 0, + "content": "Table 2: Ablation study on wOCC im- Table 3: Ablation study on using dif proved from different OCC models, always ferent WS models in our method, always with RTFM as the WS model. with STG-NF as the wOCC model." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.455, + 0.767, + 0.501 + ], + "angle": 0, + "content": "
OCC ModelOurwOCCOurWS(RTFM)GCLClassifierWS ModelOurwOCC(STG-NF)OurWS
AE70.9978.9076.14Sultani et al. [32]81.9277.41
Jigsaw [39]81.2385.35-RTFM [13]82.5788.18
STG-NF [13]82.5788.18-
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.788, + 0.592 + ], + "angle": 0, + "content": "it to \\(15\\%\\) for the ShanghaiTech dataset which contains a small ratio of abnormal data. The UBnormal dataset contains relatively more abnormal data and we set \\(R\\%\\) to \\(30\\%\\). The other parameter is \\(Q\\%\\) used to stop the repeating procedure. For both datasets, we set \\(Q\\%\\) to \\(10\\%\\). Analysis of \\(Q\\%\\) is put into Supp." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.616, + 0.585, + 0.632 + ], + "angle": 0, + "content": "4.4 Comparison with Previous Approaches" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.734 + ], + "angle": 0, + "content": "From left to right, Table 1 shows the comparison between our and previous approaches under Unsupervised, One-Class Classification and Weakly-Supervised. Since the unsupervised methods are scarce, besides GCL [47], we compare with \\(\\mathrm{AE}_{\\text{AllData}}\\) which means training the AE model on the whole training dataset containing both normal and abnormal data. We also compare with STG-NF\\(_{\\text{AllData}}\\). For our method, we adopt STG-NF [13] as wOCC, and RTFM [34] as WS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.799, + 0.841 + ], + "angle": 0, + "content": "First, Our \\( wOCC \\) (trained along with RTFM (I3D)) outperforms STG-NFAllData (see the table on the left). This validates the benefit of using our method to distinguish the normal from abnormal data in the training dataset. On ShanghaiTech, our proposed wOCC improves STG-NF from \\( 80.29\\% \\) to \\( 82.57\\% \\). In contrast, the improvement is from \\( 70.48\\% \\) to \\( 74.76\\% \\) on UBnormal. The improvement on UBnormal is larger as there is more abnormal data in UBnormal which causes STG-NFAllData to perform worse." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "Y. Nie et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.346, + 0.145, + 0.656, + 0.16 + ], + "angle": 0, + "content": "Table 4: Comparison with fixed thresholding." + }, + { + "type": "table", + "bbox": [ + 0.241, + 0.171, + 0.759, + 0.232 + ], + "angle": 0, + "content": "
Weighted OCCWS with Adaptive ThresholdRTFM (AUC %)STG-NF (AUC %)
XX82.0680.52
✓X83.4881.78
X✓85.8681.94
✓✓88.1882.57
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.26, + 0.787, + 0.35 + ], + "angle": 0, + "content": "Second, please compare our method with GCL. Our \\( W_S \\) achieves an AUC of 88.18%, while that of GCL is 76.14% (this score is reported by the classifier of GCL, thus we use our WS model for the comparison). This significant improvement is partly due to the reason that we use a stronger OCC model, i.e., STG-NF, than GCL. If using the same OCC model, our method outperforms GCL by 2.76%, as shown in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.787, + 0.457 + ], + "angle": 0, + "content": "Our UVAD method can be degenerated into a supervised OCC method. For example, we can perform our interleaving training on the normal data only. The obtained wOCC model in this setting is denoted by \\(\\mathrm{Our}_{wOCC}^{**}\\) which outperforms existing supervised OCC approaches (see the table in the middle). This validates the effectiveness of the proposed wOCC with weighted importance, as even among the normal data there is data that is more normal than other data, and the wOCC model can better identify their difference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.787, + 0.548 + ], + "angle": 0, + "content": "To degenerate our method into a supervised WS method, we allow our WS model to know video-level labels. Specifically, snippets of normal videos are treated as normal. Since snippets of abnormal videos can be either normal or abnormal, they are processed by our adaptive thresholding approach. As seen in Table 1 on the right, Our\\(^{**}_{WS}\\) (RTFM) outperforms the baseline RTFM [34]. On UBnormal, our method performs best among all the compared WS methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.572, + 0.525, + 0.587 + ], + "angle": 0, + "content": "4.5 Ablation Study and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.785, + 0.628 + ], + "angle": 0, + "content": "Next, we perform ablation studies on our method. Without extra explanations, experiments are conducted by taking STG-NF as wOCC and RTFM as WS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.787, + 0.718 + ], + "angle": 0, + "content": "Comparison between OCC over wOCC. In Figure 3, we compare the performance fluctuation of using wOCC or OCC (STG-NF) in our framework in the first two interleaving training modules' training. As shown, directly using OCC leads to performance fluctuation of both OCC and WS models and converges to an unsatisfactory point. In comparison, our proposed wOCC stabilizes the training and improves the performance of both models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.787, + 0.809 + ], + "angle": 0, + "content": "Using Different OCC/WS Models. In Table 2, we conduct experiments of upgrading different types of OCC models to wOCC in our method. We test AE [47], Jigsaw [39], and STG-NF [13], while always using RTFM [34] as the WS model. As seen, the effectiveness of our method is essentially affected by wOCC model. Similarly in Table 3, the same wOCC model is used but with different WS models. Overall, a better wOCC or WS model yields better results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Comparison with Fixed Thresholding. In Table 4, the first row shows results of using fixed thresholding for both OCC and WS models. The second and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.495, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.144, + 0.639, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.145, + 0.781, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.276, + 0.782, + 0.346 + ], + "angle": 0, + "content": "Fig. 3: Performance fluctuation in the first Fig. 4: Left: ShanghaiTech. Right: UBnor and second modules. Red line: AUC of mal. Top: \\( T_{ws} \\) in each interleaving trainwOCC (or WS) when interleaving wOCC ing module. Bottom: The AUC of the WS and WS. Blue line: AUC of OCC (or WS) model in each module. Stars in top figures: when interleaving OCC and WS. position where achieves stopping criterion." + }, + { + "type": "table_caption", + "bbox": [ + 0.36, + 0.361, + 0.642, + 0.376 + ], + "angle": 0, + "content": "Table 5: Training time on ShanghaiTech." + }, + { + "type": "table", + "bbox": [ + 0.278, + 0.388, + 0.722, + 0.461 + ], + "angle": 0, + "content": "
ModelEpoch Each StepTraining StepsTraining EpochsTraining TimeSTG-NF AUC %RTFM AUC %
Our UVAD1172×1×172h28m82.5788.18
2122×2×122h42m82.5387.95
3112×3×113h31m82.2786.36
STG-NF [13]--810m85.90-
RTFM [34]--502h32m-97.21
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.785, + 0.534 + ], + "angle": 0, + "content": "third rows use fixed thresholding for either OCC or WS models. The fourth row shows our method with wOCC and adaptive thresholding. As seen, our method outperforms all the other variants." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.787, + 0.643 + ], + "angle": 0, + "content": "Effectiveness of the Stopping Criterion. In Figure 4, we show that with the current stopping criterion, our method can stop when it achieves the best AUC (i.e., the best VAD accuracy). As the repeating procedure runs, \\( T_{ws} \\) drops and AUC increases. \\( T_{ws} \\) drops very fast at the very beginning, and then the drop rate becomes slower. Simultaneously, AUC first increases and then decreases. We stop our method when the change rate of \\( T_{ws} \\) is about \\( 10\\% \\) of its original change rate, which finds the best stopping positions indicated by the stars in top figures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.748 + ], + "angle": 0, + "content": "Robustness to \\( R\\% \\). Our method is robust to the parameter \\( R\\% \\), as demonstrated in Figure 5. On the top, we show that the threshold \\( T_{ws} \\) converges to similar values after 6 interleaving training modules when using different \\( R\\% \\). At the bottom, we show that different \\( R\\% \\) yield similar AUC on both training datasets. The experiments demonstrate that our method is not influenced much by this parameter for each dataset. For different datasets, we need to set different \\( R\\% \\) that roughly match with the ratio of anomaly data in the datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Training Loss Curve. In Figure 6, we show training loss curves of the wOCC (STG-NF) model in our framework. As seen, the loss drops smoothly within each interleaving training module. Since we re-train the wOCC model from scratch in the next module, the loss suddenly increases at the beginning of the next module, but the peak magnitude of the loss is lower than that of the previous module. Please check the zoomed-in curve showing the loss curve of the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Nie et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.77, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.301, + 0.772, + 0.358 + ], + "angle": 0, + "content": "Fig.5: Top: different \\( R\\% \\) converge to Fig.6: Bottom: training loss curve of the similar \\( T_{ws} \\) (on ShanghaiTech). Bottom: wOCC model during the whole repeating different \\( R\\% \\) yield similar AUC (Left: on procedure. Top: zoomed-in training loss ShanghaiTech, Right: on UBnormal). curve in the third module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.785, + 0.437 + ], + "angle": 0, + "content": "wOCC model in the third module. Even though the model is trained alternately with another model, its loss can still decrease smoothly, not affected by the other model. The WS model's training loss curve is put into the supplemental material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.605 + ], + "angle": 0, + "content": "Training Time. As mentioned in Section 3.3, although we need to train the wOCC and WS models alternately many times in multiple modules, our method is fast in training. That is because, we just train a model for one epoch and then exchange to train the other model for another epoch, which form a training step of our method. Our interleaving training converges very fast. As shown in Table 5, there are only 17 training steps during the whole training, yielding a total of \\(34 = 2(\\mathrm{wOCC}\\) and \\(\\mathrm{WS})\\times 1(\\mathrm{epoch})\\times 17\\) (all training steps) epochs. It takes around 2.5 hours to train our method on ShanghaiTech. We have also trained the wOCC or WS model for 2 or 3 epochs in each training step, but obtained worse results, probably because training with more epochs would enhance the supervision of wrong pseudo labels in early training modules." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.635, + 0.36, + 0.65 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this paper, we interleave the One-Classification and Weakly-Supervised models with adaptive thresholding to tackle the unsupervised video anomaly detection problem without human annotation. We face the challenges of performance fluctuation and the demanding of accurate threshold when designing the framework. To alleviate performance fluctuation, We propose the wOCC model which requires soft labels but not binary labels. To obtain a relative optimal threshold, we repeat our proposed interleaving training module, during which we propose a monotonically threshold decreasing mechanism to refine a rough threshold progressively. Extensive experiments demonstrate the effectiveness of our method. Remarkably, our method can be upgraded with the most recent development in OCC and WS VAD fields." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.178, + 0.787, + 0.254 + ], + "angle": 0, + "content": "This work was partly supported by the Natural Science Foundation of China (62072191, U21A20520, 62325204), the National Key Research and Development Program of China (2022YFE0112200), the Key-Area Research and Development Program of Guangzhou City (202206030009), and the Guangdong Basic and Applied Basic Research Fund (2023A1515030002, 2024A1515011995)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.279, + 0.323, + 0.294 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.367 + ], + "angle": 0, + "content": "1. Acsintoae, A., Florescu, A., Georgescu, M., Mare, T., Sumedrea, P., Ionescu, R.T., Khan, F.S., Shah, M.: Unbnormal: New benchmark for supervised open-set video anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.369, + 0.785, + 0.422 + ], + "angle": 0, + "content": "2. Al-lahham, A., Tastan, N., Zaheer, M.Z., Nandakumar, K.: A coarse-to-fine pseudolabeling (c2fpl) framework for unsupervised video anomaly detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6793-6802 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.424, + 0.785, + 0.464 + ], + "angle": 0, + "content": "3. Andrews, S., Tsochantaridis, I., Hofmann, T.: Support vector machines for multiple-instance learning. Advances in neural information processing systems 15 (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.465, + 0.785, + 0.52 + ], + "angle": 0, + "content": "4. Barbalau, A., Ionescu, R.T., Georgescu, M.I., Dueholm, J., Ramachandra, B., Nasrollahi, K., Khan, F.S., Moeslund, T.B., Shah, M.: Ssmtl++: Revisiting self-supervised multi-task learning for video anomaly detection. Computer Vision and Image Understanding 229, 103656 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.521, + 0.785, + 0.562 + ], + "angle": 0, + "content": "5. Basharat, A., Gritai, A., Shah, M.: Learning object motion patterns for anomaly detection and improved object detection. In: 2008 IEEE conference on computer vision and pattern recognition. pp. 1-8. IEEE (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.563, + 0.785, + 0.604 + ], + "angle": 0, + "content": "6. Burlina, P., Joshi, N., Wang, I., et al.: Where's wally now? deep generative and discriminative embeddings for novelty detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11507-11516 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.605, + 0.785, + 0.633 + ], + "angle": 0, + "content": "7. Dietterich, T.G., Lathrop, R.H., Lozano-Pérez, T.: Solving the multiple instance problem with axis-parallel rectangles. Artificial intelligence 89(1-2), 31-71 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.785, + 0.673 + ], + "angle": 0, + "content": "8. Feng, J.C., Hong, F.T., Zheng, W.S.: Mist: Multiple instance self-training framework for video anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14009-14018 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.674, + 0.785, + 0.729 + ], + "angle": 0, + "content": "9. Feng, X., Song, D., Chen, Y., Chen, Z., Ni, J., Chen, H.: Convolutional transformer based dual discriminator generative adversarial networks for video anomaly detection. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 5546-5554 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.73, + 0.785, + 0.785 + ], + "angle": 0, + "content": "10. Georgescu, M.I., Barbalau, A., Ionescu, R.T., Khan, F.S., Popescu, M., Shah, M.: Anomaly detection in video via self-supervised and multi-task learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12742-12752 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "11. Georgescu, M.I., Ionescu, R.T., Khan, F.S., Popescu, M., Shah, M.: A background-agnostic framework with adversarial training for abnormal event detection in video. IEEE transactions on pattern analysis and machine intelligence 44(9), 4505-4523 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.311, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Nie et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "12. Gong, D., Liu, L., Le, V., Saha, B., Mansour, M.R., Venkatesh, S., Hengel, A.v.d.: Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1705-1714 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.786, + 0.247 + ], + "angle": 0, + "content": "13. Hirschorn, O., Avidan, S.: Normalizing flows for human pose anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13545-13554 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.248, + 0.787, + 0.302 + ], + "angle": 0, + "content": "14. Ionescu, R.T., Khan, F.S., Georgescu, M.I., Shao, L.: Object-centric auto-encoders and dummy anomalies for abnormal event detection in video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7842-7851 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.303, + 0.787, + 0.346 + ], + "angle": 0, + "content": "15. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.787, + 0.374 + ], + "angle": 0, + "content": "16. Kingma, D.P., Dhariwal, P.: Glow: Generative flow with invertible 1x1 convolutions. Advances in neural information processing systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.787, + 0.416 + ], + "angle": 0, + "content": "17. Li, S., Liu, F., Jiao, L.: Self-training multi-sequence learning with transformer for weakly supervised video anomaly detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1395-1403 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.787, + 0.459 + ], + "angle": 0, + "content": "18. Liu, W., Luo, W., Lian, D., Gao, S.: Future frame prediction for anomaly detection-- a new baseline. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6536-6545 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.787, + 0.516 + ], + "angle": 0, + "content": "19. Liu, Z., Nie, Y., Long, C., Zhang, Q., Li, G.: A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13588-13597 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.517, + 0.787, + 0.558 + ], + "angle": 0, + "content": "20. Lorre, G., Rabarisoa, J., Orcesi, A., Ainouz, S., Canu, S.: Temporal contrastive pretraining for video action recognition. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 662-670 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.787, + 0.613 + ], + "angle": 0, + "content": "21. Lv, H., Yue, Z., Sun, Q., Luo, B., Cui, Z., Zhang, H.: Unbiased multiple instance learning for weakly supervised video anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8022-8031 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.614, + 0.787, + 0.657 + ], + "angle": 0, + "content": "22. Markovitz, A., Sharir, G., Friedman, I., Zelnik-Manor, L., Avidan, S.: Graph embedded pose clustering for anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10539-10547 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.787, + 0.699 + ], + "angle": 0, + "content": "23. Medioni, G., Cohen, I., Brémond, F., Hongeng, S., Nevatia, R.: Event detection and analysis from video streams. IEEE Transactions on pattern analysis and machine intelligence 23(8), 873-889 (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.787, + 0.755 + ], + "angle": 0, + "content": "24. Morais, R., Le, V., Tran, T., Saha, B., Mansour, M., Venkatesh, S.: Learning regularity in skeleton trajectories for anomaly detection in videos. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11996-12004 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "25. Nguyen, T.N., Meunier, J.: Anomaly detection in video sequence with appearance-motion correspondence. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1273-1283 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "26. Pang, G., Yan, C., Shen, C., Hengel, A.v.d., Bai, X.: Self-trained deep ordinal regression for end-to-end video anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12173-12182 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "27. Park, H., Noh, J., Ham, B.: Learning memory-guided normality for anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14372-14381 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.23 + ], + "angle": 0, + "content": "28. Sapkota, H., Yu, Q.: Bayesian nonparametric submodular video partition for robust anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3212-3221 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.271 + ], + "angle": 0, + "content": "29. Shi, C., Sun, C., Wu, Y., Jia, Y.: Video anomaly detection via sequentially learning multiple pretext tasks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10330-10340 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.325 + ], + "angle": 0, + "content": "30. Smeureanu, S., Ionescu, R.T., Popescu, M., Alexe, B.: Deep appearance features for abnormal behavior detection in video. In: Image Analysis and Processing-ICIAP 2017: 19th International Conference, Catania, Italy, September 11-15, 2017, Proceedings, Part II 19. pp. 779-789. Springer (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.366 + ], + "angle": 0, + "content": "31. Sohrab, F., Raitoharju, J., Gabbouj, M., Iosifidis, A.: Subspace support vector data description. In: 2018 24th International Conference on Pattern Recognition (ICPR). pp. 722-727. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.785, + 0.407 + ], + "angle": 0, + "content": "32. Sultani, W., Chen, C., Shah, M.: Real-world anomaly detection in surveillance videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6479-6488 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.785, + 0.447 + ], + "angle": 0, + "content": "33. Sun, S., Gong, X.: Hierarchical semantic contrast for scene-aware video anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22846-22856 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.502 + ], + "angle": 0, + "content": "34. Tian, Y., Pang, G., Chen, Y., Singh, R., Verjans, J.W., Carneiro, G.: Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4975-4986 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.542 + ], + "angle": 0, + "content": "35. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.583 + ], + "angle": 0, + "content": "36. Tudor Ionescu, R., Smeureanu, S., Alexe, B., Popescu, M.: Unmasking the abnormal events in video. In: Proceedings of the IEEE international conference on computer vision. pp. 2895-2903 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.624 + ], + "angle": 0, + "content": "37. Tur, A.O., Dall'Asen, N., Bryan, C., Ricci, E.: Exploring diffusion models for unsupervised video anomaly detection. In: 2023 IEEE International Conference on Image Processing (ICIP). pp. 2540-2544. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.785, + 0.664 + ], + "angle": 0, + "content": "38. Wan, B., Fang, Y., Xia, X., Mei, J.: Weakly supervised video anomaly detection via center-guided discriminative learning. In: 2020 IEEE international conference on multimedia and expo (ICME). pp. 1-6. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.704 + ], + "angle": 0, + "content": "39. Wang, G., Wang, Y., Qin, J., Zhang, D., Bao, X., Huang, D.: Video anomaly detection by solving decoupled spatio-temporal jigsaw puzzles. In: European Conference on Computer Vision. pp. 494-511. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.759 + ], + "angle": 0, + "content": "40. Wang, J., Song, Y., Leung, T., Rosenberg, C., Wang, J., Philbin, J., Chen, B., Wu, Y.: Learning fine-grained image similarity with deep ranking. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1386-1393 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "41. Wang, J., Cherian, A.: Gods: Generalized one-class discriminative subspaces for anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8201-8211 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "42. Wang, Z., Zou, Y., Zhang, Z.: Cluster attention contrast for video anomaly detection. In: Proceedings of the 28th ACM international conference on multimedia. pp. 2463-2471 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Nie et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "43. Wu, J.C., Hsieh, H.Y., Chen, D.J., Fuh, C.S., Liu, T.L.: Self-supervised sparse representation for video anomaly detection. In: European Conference on Computer Vision. pp. 729-745. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.245 + ], + "angle": 0, + "content": "44. Wu, P., Liu, J., Shi, Y., Sun, Y., Shao, F., Wu, Z., Yang, Z.: Not only look, but also listen: Learning multimodal violence detection under weak supervision. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16. pp. 322-339. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "45. Yan, C., Zhang, S., Liu, Y., Pang, G., Wang, W.: Feature prediction diffusion model for video anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5527-5537 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.342 + ], + "angle": 0, + "content": "46. Yang, Z., Liu, J., Wu, Z., Wu, P., Liu, X.: Video event restoration based on keyframes for video anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14592-14601 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.396 + ], + "angle": 0, + "content": "47. Zaheer, M.Z., Mahmood, A., Khan, M.H., Segu, M., Yu, F., Lee, S.I.: Generative cooperative learning for unsupervised video anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14744-14754 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.453 + ], + "angle": 0, + "content": "48. Zaheer, M.Z., Mahmood, A., Astrid, M., Lee, S.I.: Claws: Clustering assisted weakly supervised learning with normalcy suppression for anomalous event detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 358-376. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.508 + ], + "angle": 0, + "content": "49. Zhang, C., Li, G., Qi, Y., Wang, S., Qing, L., Huang, Q., Yang, M.H.: Exploiting completeness and uncertainty of pseudo labels for weakly supervised video anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16271-16280 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "50. Zhang, J., Qing, L., Miao, J.: Temporal convolutional network with complementary inner bag loss for weakly supervised anomaly detection. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 4030-4034. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "51. Zhang, T., Lu, H., Li, S.Z.: Learning semantic scene models by object classification and trajectory clustering. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1940-1947. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.632 + ], + "angle": 0, + "content": "52. Zhao, Y., Deng, B., Shen, C., Liu, Y., Lu, H., Hua, X.S.: Spatio-temporal autoencoder for video anomaly detection. In: Proceedings of the 25th ACM international conference on Multimedia. pp. 1933-1941 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.688 + ], + "angle": 0, + "content": "53. Zhong, J.X., Li, N., Kong, W., Liu, S., Li, T.H., Li, G.: Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1237-1246 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.688 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection/0ff966a8-e3c5-432a-8beb-e808413549e5_origin.pdf b/2024/Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection/0ff966a8-e3c5-432a-8beb-e808413549e5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4e5e208106bf0ae337bb71063161357e8b3c0823 --- /dev/null +++ b/2024/Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection/0ff966a8-e3c5-432a-8beb-e808413549e5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e12676700056009c03cbad03e8e98f9c7dc7d2eb13ce39ca13cad7b70a91d18 +size 1164165 diff --git a/2024/Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection/full.md b/2024/Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..140b8ba577daa972875efa34cd3da3507fd13014 --- /dev/null +++ b/2024/Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection/full.md @@ -0,0 +1,313 @@ +# Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection + +Yongwei Nie $^{1}$ , Hao Huang $^{1}$ , Chengjiang Long $^{2}$ , Qing Zhang $^{3}$ , Pradipta + +Maji $^{4}$ , and Hongmin Cai $^{1}$ + +$^{1}$ South China University of Technology, China + +Meta Reality Labs, USA + +3 Sun Yat-sen University, China + +$^{4}$ Indian Statistical Institute, India + +Abstract. Video Anomaly Detection (VAD) has been extensively studied under the settings of One-Class Classification (OCC) and Weakly-Supervised learning (WS), which however both require laborious human-annotated normal/abnormal labels. In this paper, we study Unsupervised VAD (UVAD) that does not depend on any label by combining OCC and WS into a unified training framework. Specifically, we extend OCC to weighted OCC (wOCC) and propose a wOCC-WS interleaving training module, where the two models automatically generate pseudo-labels for each other. We face two challenges to make the combination effective: (1) Models' performance fluctuates occasionally during the training process due to the inevitable randomness of the pseudo labels. (2) Thresholds are needed to divide pseudo labels, making the training depend on the accuracy of user intervention. For the first problem, we propose to use wOCC requiring soft labels instead of OCC trained with hard zero/one labels, as soft labels exhibit high consistency throughout different training cycles while hard labels are prone to sudden changes. For the second problem, we repeat the interleaving training module multiple times, during which we propose an adaptive thresholding strategy that can progressively refine a rough threshold to a relatively optimal threshold, which reduces the influence of user interaction. A benefit of employing OCC and WS methods to compose a UVAD method is that we can incorporate the most recent OCC or WS model into our framework. Experiments demonstrate the effectiveness of the proposed UVAD framework. Our code is available at https://github.com/benedictstar/Joint-VAD. + +Keywords: Video anomaly detection $\cdot$ One-class classification $\cdot$ Weakly-supervised learning $\cdot$ Unsupervised learning + +# 1 Introduction + +Video Anomaly Detection (VAD) is a task that identifies abnormal events in a video, where the abnormal event could be a fire alarm, a flaw in an industrial + +product, or a traffic accident, etc. Most previous VAD methods fall in two categories: One-Class Classification (OCC) methods [12,13,18,19,33,39] and Weakly-Supervised (WS) approaches [21,32,34,38,43,49]. Both kinds of approaches require human-annotated labels for training1. Supervised VAD methods have two prominent problems. First, abnormal events are sparse and difficult to recognize, making the annotation of the training dataset a highly labor-intensive task. Secondly, the categories of abnormal events are unbounded, while human can only collect a limited scope of them, yielding the risk that the trained VAD methods cannot handle unobserved abnormal events. + +Considering the above problems, this paper studies Unsupervised VAD (UVAD) that does not rely on labels. Despite the importance of UVAD, few methods have been developed except [47] and its subsequent works [2, 37] to the best of our knowledge. To achieve unsupervised learning, the method of [47] adversarially trains a generator and a discriminator, where the generator is an autoencoder-based reconstruction model and the discriminator is a fully-connected (FC) binary classifier. While the proposed network architecture is sophisticated, the adopted autoencoder and FC networks are relatively weak which limit the performance of the approach. + +In this paper, we propose a new UVAD framework. Our key idea is that the techniques of OCC and WS approaches are rapidly developed, while a UVAD method is usually implemented by training two VAD models alternately generating pseudo labels for each other. This motivates us to propose a UVAD framework that directly interleaves the training of a pair of OCC and WS models. This allows us to incorporate the recent advances in the two hot research fields to implement a UVAD method. We have also attempted to combine two OCC or two WS models, but found two homogeneous models are less effective. + +While combining OCC and WS methods, we encounter two problems preventing it from effective. First, we observe the OCC or WS model's performance fluctuates occasionally during the training process, which affects the final accuracy they can achieve. This is partly because the pseudo labels generated by each other are changed frequently, making the training not stable. To solve this problem, we extend OCC to weighted OCC (wOCC), and propose a wOCC-WS interleaving training module. Previous OCC model is trained on normal data, while our wOCC model is trained with soft labels in the range of [0,1]. The soft labels are more consistent across adjacent training cycles, making the performance of wOCC more stable than OCC under changing pseudo labels, thus suppressing the fluctuation effect. Second, we still have to set a threshold to partition normal and abnormal pseudo labels for the WS model, making our method susceptible to a user-provided hyperparameter. To reduce the influence of user interaction, we repeat the interleaving training module multiple times, during which we propose an adaptive threshold mechanism that finds an optimal threshold given just a rough initial value from the user. + +To conclude, the contributions of this work include: + +- We propose a new UVAD framework that interleaves the training of a pair of OCC and WS models. +- We improve OCC to wOCC which stabilizes the training of the UVAD framework, and also propose an adaptive thresholding strategy to alleviate the influence of user interaction. +- Our UVAD framework is flexible to employ different pairs of OCC and WS models. Experiments on three OCC and two WS models demonstrate the effectiveness of our method. + +# 2 Related Work + +Unsupervised VAD (UVAD). Unsupervised VAD requires identifying anomalies from data containing both normal and abnormal samples without any annotation. This is a challenging new task that is almost unexplored in the literature. Although significant progress has been made in OCC and WS VAD, directly applying them independently to address UVAD does not yield good results. Zaheer et al. [47] firstly raise this task and propose to solve UVAD by generative cooperative learning (GCL). Differently, our proposed method modifies both the OCC and WS VAD models and combines the modified two models together to learn from unmarked training data. + +One-Class Classification VAD (OCC). OCC-based VAD approaches only have access to normal data and try to model normal data to identify behaviors that are significantly different from normal behaviors as anomalies. Early works use hand-crafted features to help detect anomalies [5,23,40,51]. With the rapid development of deep learning, recent methods turn to normal representations extracted by using a deep neural network [26, 30, 36, 41]. Some methods identify normal patterns by using the reconstruction/prediction model to reconstruct the representations [6,12,14,18,19,24,25,27,31,33,45,46,52]. These models may lead to well-reconstructed anomalies, thereby limiting the performance of the OCC-based methods. Other OCC-based methods turn to identify normal representations by using proxy tasks [4,11,20,29,39]. What's more, recently, Hirschorn and Avidan [13] propose to build a multivariate Gaussian distribution for normal data and detect instances deviating from this distribution as anomalies. However, the above methods treat all samples equally, ignoring the potential differences in importance between normal samples. + +Weakly-Supervised VAD (WS). In weakly-supervised VAD, video-level annotations are available in the training stage. Sultani et al [32] first propose to use the video-level labels and solve WS VAD based on multi-instance learning (MIL) framework [3, 7]. Since then, many works [21, 28, 38, 43, 49, 50] have viewed the WS VAD task as a MIL problem. Tian et al. [34] develop to extend MIL to Top- $k$ MIL method by training with a robust temporal feature magnitude (RTFM) loss function. The methods mentioned above directly train in the supervision of video-level labels and the coarse-grained supervision limits the accuracy of WS models. Recently, two-stage self-training methods [8, 17, 44, 53] adopt a two-stage pipeline to use more fine-grained labels to supervise the networks more + +![](images/fc004a50af94976ab7d38dd19d88eb911212310d85aa5cd2ff8131f8cc6bb500.jpg) +Fig. 1: Pipeline of our proposed interleaved framework. The wOCC-WS interleaving training module comprises a weighted OCC (wOCC) model and a WS model, using the output of each model as the pseudo-labels to supervise the other model. Further, the training module is repeated multiple times with the adaptive threshold until the stopping criterion we set. To start the first training module, we randomly sample $w_{X}$ to train the wOCC model and assign a large value as the threshold for the WS model. Once a training module converges, the last WS model of it supplies $w_{X}$ to start the next training module and adaptive thresholding takes the outputs of all previous wOCC models into consideration for updating the threshold. + +strictly. The WS model in our framework shares a similar idea with [8] and is tailored to use finer-grained snippet-level labels. + +# 3 Method + +# 3.1 Overview + +Our proposed UVAD framework interleaves the training of a pair of OCC and WS models. In theory, any OCC or WS methods can be incorporated into our framework. Without loss of generality, we take STG-NF [13] (a recent OCC model) and RTFM [34] (a recent WS model) as examples to introduce our method in the following sections. We test many other OCC/WS models in Section 4.5. + +Figure 1 shows an overview of our framework. The basic buildingblock of our method is the wOCC-WS interleaving training module. In the module, the outputs of each model are used as pseudo labels for defining the loss function training the other model. After one module converges, we re-initialize another module to run the interleaving training again with a difference that the threshold of the WS model is updated by an adaptive threshold adjustment mechanism based on wOCC models trained in previous modules. Since wOCC has more stable behaviors than WS, we always initialize the wOCC model to start the training of a module. For the first module, we use a random initialization based on Beta distribution sampling. For other modules, we use the results of WS of the preceding module for the initialization. + +# 3.2 wOCC-WS Interleaving Training Module + +We first introduce the wOCC-WS interleaving training module which comprises a wOCC model and a WS model. At the very beginning, we initialize the wOCC model to start the interleaving training. Then, pseudo labels created using wOCC are used to train the WS model. Next, pseudo labels created by the WS model are used to improve the wOCC model. The whole training does not use human annotations. + +wOCC Model (STG-NF). Typically, OCC models are trained only on normal data. Directly using OCC in our UVAD framework may make the training fluctuate, i.e., the performance of the trained models sometimes gets better, but sometimes worse, finally converging to an inferior point. We use 0 to indicate normal samples, and 1 abnormal samples, where 0, 1 are hard binary labels. During training, we need to identify 0-labelled samples to train the OCC model. The hard partitioned labels may be suddenly changed from 0 to 1 or from 1 to 0, yielding the frequent change of the normal samples used to train the OCC model which is the source of the unstable training. To this end, we propose wOCC which relies on soft labels instead of hard binary labels. Soft labels are in the range of [0, 1]. For example, the soft label 0.7 may be changed to 0.6 but not 0 suddenly. This consistency reduces the fluctuation of the interleaving training. + +Taking STG-NF [13] (an OCC model) as an example, we introduce how to improve it to wOCC. STG-NF extracts a set of objects $\mathcal{X} = \{X_1,\dots ,X_i,\dots \}$ from the training videos, where $X_{i}$ is a sequence of poses of a person. Each object $X_{i}$ is the basic building block processed by STG-NF. Originally, STG-NF minimizes the negative log-likelihood across the normal data: + +$$ +\mathcal {L} _ {o c c} = - \log \left(p _ {Z} \left(f _ {S T G - N F} \left(X _ {i} ^ {+}\right)\right)\right), \tag {1} +$$ + +where $X_{i}^{+} \in \mathcal{X}^{+}$ is sampled from normal data, $Z = f_{STG-NF}(X_{i}^{+})$ is the feature extracted from $X_{i}^{+}$ by the STG-NF, $p_{Z}$ is the established distribution of the normal data. We upgrade STG-NF by introducing the soft labels $w_{X} = \{w_{X_{1}}, \dots, w_{X_{i}}, \dots\}$ as the weight to supervise its training on both normal and abnormal data. Formally, for the wOCC model improved from STG-NF, we minimize the following weighted negative log-likelihood across all the data: + +$$ +\mathcal {L} _ {w o c c} = - \left(1 - w _ {X _ {i}}\right) \log \left(p _ {Z} \left(f _ {S T G - N F} \left(X _ {i}\right)\right)\right), \tag {2} +$$ + +where $X_{i} \in \mathcal{X}$ is sampled from the whole training dataset rather than the normal part and $w_{X_i} \in [0,1]$ is the weight of the object $X_{i}$ indicating the anomaly degree of the object which serves as the soft label. The object with smaller $w_{X_i}$ means more normal which is better captured by the learned distribution of wOCC. + +Pseudo Labels $l_{S}$ from wOCC. Now we use the trained wOCC model to generate pseudo labels $l_{S} = \{l_{S_{1}},\dots ,l_{S_{i}},\dots \}$ for training the WS model. "Snippet" is the basic element processed by WS models. We first split the training videos into snippets, obtaining a set of snippets $S = \{S_1,\dots ,S_i,\dots \}$ . Our aim is to compute the label $l_{S_i}\in \{0,1\}$ for each snippet $S_{i}$ . Let $x_{i}$ be the anomaly score of the object $X_{i}$ computed by the trained wOCC model. The anomaly score $\hat{s}_i$ + +of a video snippet $S_{i}$ is computed by averaging all the anomaly scores of objects in the snippet. To generate pseudo label $l_{S_i}$ , one way is to check whether $\hat{s}_i$ is larger than a threshold $(l_{S_i} = 1)$ or not $(l_{S_i} = 0)$ . However, as the training goes, the anomaly scores of snippets may vary in a large range, making the absolute threshold invalid. Instead, we sort the snippets according to their anomaly scores in descending order, and get the index of the snippet $S_{i}$ by $\mathrm{Rank}(\hat{s}_i)$ in the sorted list. Finally, the label $l_{S_i}$ of snippet $S_{i}$ is computed by + +$$ +l _ {S _ {i}} = \left\{ \begin{array}{l l} 1, & \text {i f} \quad \operatorname {R a n k} \left(\hat {s} _ {i}\right) < T _ {w s}, \\ 0, & \text {o t h e r w i s e}, \end{array} \right. \tag {3} +$$ + +where $T_{ws}$ is the relative threshold not affected by concrete anomaly scores. + +WS Model (RTFM). With snippet-level pseudo annotations $l_{S}$ (obtained in Eq. 3), we now define positive and negative bags and use them to train a WS model. Previously, approaches [32,34] usually treat snippets of the same video as a bag based on video-level labels. Differently, thanks to the snippet-level labels provided by the wOCC model, we can compose positive and negative bags more flexibly. In our method, a positive bag $B^{+} = \{S_{i}^{+}\}_{i=1}^{C}$ is composed of randomly selected $C = 16$ abnormal snippets. Similarly, a negative bag $B^{-} = \{S_{j}^{-}\}_{j=1}^{C}$ is composed of randomly selected $C$ normal snippets. Taking RTFM [34] as an example, the WS method employs multiple instance learning, requiring the average Top- $k$ maximum magnitudes of features in positive bags exceeds that in negative bags: + +$$ +\begin{array}{l} \mathcal {L} _ {w s} = \max (0, m - d _ {R T F M} (B ^ {+}) + d _ {R T F M} (B ^ {-})) + \tag {4} \\ \mathcal {L} _ {B C E} \left(B ^ {+}\right) + \mathcal {L} _ {B C E} \left(B ^ {-}\right), \\ \end{array} +$$ + +where $m$ is the margin, and $d_{RTFM}(B)$ (where $B$ can be $B^{+}$ or $B^{-}$ ) returns the average of the Top- $k$ maximum magnitudes of features. Finally, $\mathcal{L}_{BCE}$ is the binary cross-entropy-based classification loss: + +$$ +\mathcal {L} _ {B C E} (B) = - Y \log \left(f _ {R T F M} (B)\right) + (1 - Y) \log \left(1 - f _ {R T F M} (B)\right), \tag {5} +$$ + +where $Y$ is the label of the bag $B$ and $f_{RTFM}(B)$ returns the average of the Top- $k$ maximum anomaly scores. Note that $f_{RTFM}(B)$ is different from $d_{RTFM}(B)$ , where $f_{RTFM}(B)$ returns the anomaly scores while $d_{RTFM}(B)$ returns the magnitude of the features. + +Pseudo Labels $w_{X}$ from WS. wOCC model needs soft labels $w_{X}$ . With the trained WS model, we compute the anomaly score $\hat{x}_i$ of object $X_i$ by simply setting it as the anomaly score of the snippet containing the object $X_i$ . Then, the soft label $w_{X_i}$ is directly set as: + +$$ +w _ {X _ {i}} = \hat {x} _ {i}. \tag {6} +$$ + +$\hat{x}_i$ is already in the range of $[0,1]$ , so we do not need to normalize $w_{X_i}$ . + +Initializing the Interleaving Training Module. As discussed above, in our proposed interleaving training module, wOCC training relies on pseudo labels generated by the WS model, and conversely, WS training requires pseudo + +labels derived from the wOCC. This egg and chicken problem needs to be solved at the very beginning. Our solution is to provide the wOCC model with randomly generated soft labels to start the interleaving training. Usually, the training dataset contains much more normal samples than abnormal ones, which means most soft labels are around 0 while a small portion of them are near 1. Therefore, we randomly sample the weight $w_{X}$ from the Beta distribution, i.e., $w_{X} \sim \mathrm{Beta}(\alpha, \beta)$ , where $\alpha = 1$ and $\beta = 5$ . The sampled weights $w_{X}$ are mostly around 0, and a small portion of them is close to 1, satisfying the prior about the distribution of normal and abnormal data. + +Convergence Analysis. We have discussed the mechanism and the way to start the interleaving training module above. Then, in the following, we empirically analyze its convergence, which is validated by the experiment in Figure 6. At the very beginning, we randomly initialize the weight $w_{X}$ to train our wOCC model. Although the random initialization is not precise, the wOCC model can still learn normal patterns from the training data, as we assume there is more normal data than abnormal data in the training dataset. Therefore, the wOCC model can produce relatively reliable pseudo labels for the WS model. This is very important as the WS model has a high requirement on the quality of the pseudo labels. In turn, with the trained WS model, we can provide better pseudo supervision to train a stronger wOCC model than the random initialization. In this way, we achieve mutual improvement between two models until convergence. + +The above analysis implies the reason why we do not choose to interleave the training of two OCC models or two WS models. The OCC model can provide reliable initialization, while the WS model explicitly enlarges the gap between normal and abnormal samples. Two OCC models fail to utilize the abnormal data, while for two WS models it is not known how to provide a reliable initialization to start the training. + +# 3.3 Repeating Procedure with Adaptive Thresholding + +In the above, the wOCC and WS models need to provide pseudo labels to supervise the other. As the wOCC model relies on soft labels $w_{X}$ as defined in Eq. 6, its pseudo labels can be directly provided by the WS model without thresholding. However, as defined in Eq. 3, we need to specify a threshold for the WS model since it requires binary hard partition labels to construct positive and negative bags. To alleviate the influence of human interaction, we propose a repeating procedure of the interleaving training module which can adaptively refine a user-specified rough threshold to a relatively optimal one. + +Starting and running the next interleaving training module. As shown in Figure 1, we run several interleaving training modules one after another. Recall that we provide $w_{X}$ by random initialization to start the first interleaving training module as discussed in Section 3.2. To start the next module, we use the final WS model in the last module to provide the labels $w_{X}$ . For the wOCC and WS models in a new interleaving training module, we train them from scratch, not inheriting parameters from the last module. + +Adaptive Thresholding. The threshold is fixed in each module. Between two modules, we employ an adaptive thresholding mechanism to refine the threshold. As the repeating procedure runs, we obtain a series of thresholds $\{T_{ws}^{1}, T_{ws}^{2}, \dots, T_{ws}^{i}, \dots\}$ corresponding to the interleaving training modules. The key idea of adaptive thresholding is to ensure: + +$$ +T _ {w s} ^ {1} \geq T _ {w s} ^ {2} \geq \dots \geq T _ {w s} ^ {i} \geq \dots . \tag {7} +$$ + +Starting from a sufficiently large (larger than the optimal threshold) value of $T_{ws}^{1}$ , there must be one threshold $T_{ws}^{*} \in \{T_{ws}^{1}, T_{ws}^{2}, \dots, T_{ws}^{i}, \dots\}$ that is closest to the optimal threshold. The problem is how to enforce the monotonically decreasing of the threshold and how to stop the repeating procedure at the optimal point. + +First, we initialize $T_{ws}^{1} = R\% \times N$ , where $R$ is a user-specific hyper-parameter, e.g. $R = 30$ , and $N$ is the total number of snippets in the training dataset. This setting makes $T_{ws}^{1}$ a large enough threshold, since generally the ratio of abnormal data in the training dataset is less than $30\%$ . We test different $R$ (e.g., $35\%$ ) in the experimental section and find that it does not influence our final results much, demonstrating the robustness of our method to the user interaction. + +Then, to ensure Eq. 7, our adaptive thresholding method determines the threshold used in the current module based on wOCC models trained in all previous modules. Each training module produces multiple wOCC models, obtaining one after each loop between wOCC and WS. Let $M_{i}$ be the number of wOCC models trained from module 1 to module $i$ . When computing $T_{ws}^{i+1}$ for module $i+1$ with $i \geq 1$ , we take all the $M_{i}$ wOCC models into consideration. Specifically, for each wOCC, e.g., the $j^{th}$ one where $j \in [1, M_{i}]$ , we use it to identify $R\%$ of snippets in the training dataset that have high anomaly scores, and denote the resulting set of snippets as $A_{j}$ . We then compute intersection between all sets and count the number of the resulting intersection as the threshold: + +$$ +T _ {w s} ^ {i + 1} = \operatorname {N u m} \left(A _ {1} \cap A _ {2} \cap \dots \cap A _ {M _ {i}}\right), \tag {8} +$$ + +where $\mathrm{Num}(\cdot)$ counts the number of elements in the intersection. The operation of computing intersection ensures that the obtained threshold is monotonically decreased. In essence, the threshold computed in this way means the number of snippets viewed as abnormal by all wOCC models. At the early stages, the number of wOCC models is small, therefore the size of the intersection is large, yielding large thresholds. As more wOCC models involved, it is more difficult to form a consensus, thus producing smaller thresholds. + +To elucidate the effect of the adaptive thresholding, we visualize how WS models are gradually improved as the repeating procedure runs in Figure 2. In the figure, we show two examples. Taking the one on the left as an example, it shows a video with an abnormal event in the middle. At the beginning when $T_{ws} = 1643$ , the WS model wrongly identifies the abnormal event as normal. As the repeating procedure runs, $T_{ws}$ becomes smaller and smaller, which is finally reduced to 834 after repeating 5 times, and the WS model begins to identify the anomaly event progressively. + +![](images/860735d1d9744c90ccc3a68a5c408ca563845b9f3d80e1d6e2fbe130fac10d2f.jpg) +Fig. 2: As repeating procedure goes, $T_{ws}$ monotonically decreases, improving the WS model gradually. Left: example from ShanghaiTech. Right: example from UBnormal. + +![](images/662a2edb42455b1676dcec073ef668028956d895e7fcafe2b037cf985cbf2f0d.jpg) + +Stopping Criterion. As our goal is to find the threshold $T_{ws}^{*}$ close to the optimal threshold, we need to stop the repeating procedure at the right time. Experimentally, we find that $T_{ws}$ drops fast at the first few interleaving training modules, then the rate of change of $T_{ws}$ becomes smaller which indicates that the wOCCs achieve a consensus about anomalies. The size of the consensus set is probably the actual number of abnormal snippets in the dataset. Based on this observation, once change rate of $T_{ws}$ between two consecutive training modules is less than $Q\%$ of the first change rate at the very beginning, we stop the repeating procedure. Please see the high correlation between the change rate of $T_{ws}$ and the accuracy of the WS model in Figure 4. + +Training Time Analysis. Though we need to repeat the interleaving training module multiple times, our training is as fast as the original wOCC or WS method. First, the repeating procedure usually stops with fewer than 6 modules thanks to the stopping criterion. Furthermore, we train wOCC (or WS) for just one epoch, and then exchange to train WS (or wOCC) for another epoch. If we have 6 modules, each running an average of 5 loops of interleaving training, the wOCC and WS models are trained for just 30 epochs each. The original wOCC or WS models are also trained for the similar number of epochs in their default experimental settings. The detailed experiments are in Section 4.5. + +# 3.4 Inference + +At the inference stage, both wOCC and WS models can be used to detect anomalies. We provide results of both models when compared with previous methods. + +# 4 Experiments + +# 4.1 Datasets and Evaluation Metrics + +ShanghaiTech. The ShanghaiTech dataset [18] contains 437 videos and was created as a benchmark for OCC with only normal videos available in the training + +set. Zhong et al. [53] reorganized the dataset to enable the training of WS systems. The new split contains 63 abnormal videos and 175 normal videos, while there are 44 anomalous and 155 normal videos in the new testing split. In our unsupervised setting, we follow the split for WS approaches but do not provide video-level annotations in the training stage. + +UBnormal. The UBnormal dataset [1] is a synthetic open-set benchmark, containing 268 training videos, 64 validation videos and 211 test videos. Normal and abnormal videos are mixed in the three subsets. The dataset is challenging because of disjoint sets of anomaly types in training and testing sets. We follow its original organization but train our proposed UVAD without any annotations. + +Evaluation Metrics. Following prior VAD arts [13, 34], we use the frame-level area under ROC curve (AUC) to measure VAD's accuracy for both datasets. + +# 4.2 wOCC and WS models Employed + +As stated, our method is a flexible UVAD framework that can use different wOCC and WS models. We test improving three OCC models to wOCC, including the AE model used in [47], Jigsaw [39], and STG-NF [13]. The wOCC losses for $AE$ and Jigsaw are put into supplemental material. + +(1) AE. GCL [47] proposes to use an AutoEncoder that reconstructs the features extracted from videos as their OCC model. We use it here too. +(2) Jigsaw. The OCC model Jigsaw [39] addresses VAD by solving a pretext task: spatio-temporal jigsaw puzzles. It divides the spatio-temporal space of a video into smaller cubes, then shuffles the positions of the cubes, and finally tries to restore the original positions of the cubes. +(3) STG-NF. STG-NF [13] extracts people's pose sequences, and extends Glow [16] to build multivariate Gaussian distribution of normal action sequences. +For WS, we test the method proposed in Sultani et al. [32] and RTFM [34]. +(1) Sultani et al. [32]. Sultani et al. [32] propose the first MIL VAD model that maximizes the separability between a positive bag containing snippets of an abnormal video and a negative bag with snippets of a normal video. +(2) RTFM. RTFM [34] extends Sultani et al. [32] to compare the Top- $k$ largest-magnitude snippets between positive and negative bags. + +# 4.3 Implementation Details + +We implement our method in PyTorch. For STG-NF and RTFM (implementation details of other wOCC/WS models are put in supplemental material), they are optimized by Adam optimizer with $\beta_{1} = 0.9$ , $\beta_{2} = 0.999$ , and a weight decay of 0.0005. The batch size of STG-NF is 256, and that of RTFM is 32. The learning rate is $5e - 4$ and $1e - 3$ for STG-NF and RTFM respectively. For RTFM, the margin $m$ is set to 100. During interleaving training, we train a model for one epoch and then exchange to train the other model for another epoch. + +Our method has two hyperparameters shared by all types of wOCC or WS models. One is $R\%$ which determines the initial threshold $T_{ws}^{1}$ . By default, we set + +Table 1: Comparison with previous approaches. For wOCC, we use STG-NF [13]. For WS, we use RTFM [34]. We test using C3D [35] and I3D [15] to extract features for RTFM. Our $w_{OCC}$ gives the AUC of the wOCC model. Our $WS$ gives the AUC of the WS model. The methods split by the dashed line are trained as a whole in our framework. * means we reimplement the method by taking I3D as the feature. ** means our method degenerated to the supervised OCC or WS method. *** means the results are reproduced by us using the official code. + +
UnsupervisedOne-Class ClassificationWeakly Supervised
MethodFeaturesSTech AUC %UB AUC %MethodSTech AUC %UB AUC %MethodFeaturesSTech AUC %UB AUC %
AEAllDataI3D60.51-MemAE [12]71.20-GCN [53]C3D76.44-
Frame Prediction [18]73.40-GCN [53]TSN84.44-
STG-NFAllData [13]-80.2970.48Markovitz et al. [22]76.1052.00Zhang et al. [50]C3D82.50-
HF2-VAD [19]76.20-Sultani et al.*** [32]I3D84.5354.12
GCL* [47]I3D76.14-CT-D2GAN [9]77.70-AR-Net [38]I3D85.38-
CAC [42]79.30-CLAWS [48]C3D89.67-
OurwOCC-81.5071.67SSMTL [10]82.40-MIST [8]I3D94.83-
OurwSC3D85.4359.05Georgescu et al. [11]82.7059.30Li et al. [17]I3D96.08-
SSMTL++ [4]83.8062.10RTFM*** [34]C3D91.5162.30
OurwOCC-82.5774.76Jigsaw [39]84.3056.40RTFM*** [34]I3D96.1066.83
STG-NF [13]85.9071.80S3R [43]I3D97.48-
OurwSI3D88.1863.10OurwOCC (STG-NF)86.3772.81OurwS (RTFM)I3D96.3367.42
+ +Table 2: Ablation study on wOCC im- Table 3: Ablation study on using dif proved from different OCC models, always ferent WS models in our method, always with RTFM as the WS model. with STG-NF as the wOCC model. + +
OCC ModelOurwOCCOurWS(RTFM)GCLClassifierWS ModelOurwOCC(STG-NF)OurWS
AE70.9978.9076.14Sultani et al. [32]81.9277.41
Jigsaw [39]81.2385.35-RTFM [13]82.5788.18
STG-NF [13]82.5788.18-
+ +it to $15\%$ for the ShanghaiTech dataset which contains a small ratio of abnormal data. The UBnormal dataset contains relatively more abnormal data and we set $R\%$ to $30\%$ . The other parameter is $Q\%$ used to stop the repeating procedure. For both datasets, we set $Q\%$ to $10\%$ . Analysis of $Q\%$ is put into Supp. + +# 4.4 Comparison with Previous Approaches + +From left to right, Table 1 shows the comparison between our and previous approaches under Unsupervised, One-Class Classification and Weakly-Supervised. Since the unsupervised methods are scarce, besides GCL [47], we compare with $\mathrm{AE}_{\text{AllData}}$ which means training the AE model on the whole training dataset containing both normal and abnormal data. We also compare with STG-NF $_{\text{AllData}}$ . For our method, we adopt STG-NF [13] as wOCC, and RTFM [34] as WS. + +First, Our $wOCC$ (trained along with RTFM (I3D)) outperforms STG-NFAllData (see the table on the left). This validates the benefit of using our method to distinguish the normal from abnormal data in the training dataset. On ShanghaiTech, our proposed wOCC improves STG-NF from $80.29\%$ to $82.57\%$ . In contrast, the improvement is from $70.48\%$ to $74.76\%$ on UBnormal. The improvement on UBnormal is larger as there is more abnormal data in UBnormal which causes STG-NFAllData to perform worse. + +Table 4: Comparison with fixed thresholding. + +
Weighted OCCWS with Adaptive ThresholdRTFM (AUC %)STG-NF (AUC %)
XX82.0680.52
✓X83.4881.78
X✓85.8681.94
✓✓88.1882.57
+ +Second, please compare our method with GCL. Our $W_S$ achieves an AUC of 88.18%, while that of GCL is 76.14% (this score is reported by the classifier of GCL, thus we use our WS model for the comparison). This significant improvement is partly due to the reason that we use a stronger OCC model, i.e., STG-NF, than GCL. If using the same OCC model, our method outperforms GCL by 2.76%, as shown in Table 2. + +Our UVAD method can be degenerated into a supervised OCC method. For example, we can perform our interleaving training on the normal data only. The obtained wOCC model in this setting is denoted by $\mathrm{Our}_{wOCC}^{**}$ which outperforms existing supervised OCC approaches (see the table in the middle). This validates the effectiveness of the proposed wOCC with weighted importance, as even among the normal data there is data that is more normal than other data, and the wOCC model can better identify their difference. + +To degenerate our method into a supervised WS method, we allow our WS model to know video-level labels. Specifically, snippets of normal videos are treated as normal. Since snippets of abnormal videos can be either normal or abnormal, they are processed by our adaptive thresholding approach. As seen in Table 1 on the right, Our $^{**}_{WS}$ (RTFM) outperforms the baseline RTFM [34]. On UBnormal, our method performs best among all the compared WS methods. + +# 4.5 Ablation Study and Discussions + +Next, we perform ablation studies on our method. Without extra explanations, experiments are conducted by taking STG-NF as wOCC and RTFM as WS. + +Comparison between OCC over wOCC. In Figure 3, we compare the performance fluctuation of using wOCC or OCC (STG-NF) in our framework in the first two interleaving training modules' training. As shown, directly using OCC leads to performance fluctuation of both OCC and WS models and converges to an unsatisfactory point. In comparison, our proposed wOCC stabilizes the training and improves the performance of both models. + +Using Different OCC/WS Models. In Table 2, we conduct experiments of upgrading different types of OCC models to wOCC in our method. We test AE [47], Jigsaw [39], and STG-NF [13], while always using RTFM [34] as the WS model. As seen, the effectiveness of our method is essentially affected by wOCC model. Similarly in Table 3, the same wOCC model is used but with different WS models. Overall, a better wOCC or WS model yields better results. + +Comparison with Fixed Thresholding. In Table 4, the first row shows results of using fixed thresholding for both OCC and WS models. The second and + +![](images/e49694ac09a0b54bf3227bc03b60835fad9bbad271e57af2fb697a0a57382436.jpg) +Fig. 3: Performance fluctuation in the first Fig. 4: Left: ShanghaiTech. Right: UBnor and second modules. Red line: AUC of mal. Top: $T_{ws}$ in each interleaving trainwOCC (or WS) when interleaving wOCC ing module. Bottom: The AUC of the WS and WS. Blue line: AUC of OCC (or WS) model in each module. Stars in top figures: when interleaving OCC and WS. position where achieves stopping criterion. + +![](images/87c312a702868bd884db20a804a0401d1327f3b87656714014a95d732e2ce808.jpg) + +![](images/62b28728f4ba29b746f6044d05fe6177e72bd51e49af8b1fbee6991a62dbe5f9.jpg) + +Table 5: Training time on ShanghaiTech. + +
ModelEpoch Each StepTraining StepsTraining EpochsTraining TimeSTG-NF AUC %RTFM AUC %
Our UVAD1172×1×172h28m82.5788.18
2122×2×122h42m82.5387.95
3112×3×113h31m82.2786.36
STG-NF [13]--810m85.90-
RTFM [34]--502h32m-97.21
+ +third rows use fixed thresholding for either OCC or WS models. The fourth row shows our method with wOCC and adaptive thresholding. As seen, our method outperforms all the other variants. + +Effectiveness of the Stopping Criterion. In Figure 4, we show that with the current stopping criterion, our method can stop when it achieves the best AUC (i.e., the best VAD accuracy). As the repeating procedure runs, $T_{ws}$ drops and AUC increases. $T_{ws}$ drops very fast at the very beginning, and then the drop rate becomes slower. Simultaneously, AUC first increases and then decreases. We stop our method when the change rate of $T_{ws}$ is about $10\%$ of its original change rate, which finds the best stopping positions indicated by the stars in top figures. + +Robustness to $R\%$ . Our method is robust to the parameter $R\%$ , as demonstrated in Figure 5. On the top, we show that the threshold $T_{ws}$ converges to similar values after 6 interleaving training modules when using different $R\%$ . At the bottom, we show that different $R\%$ yield similar AUC on both training datasets. The experiments demonstrate that our method is not influenced much by this parameter for each dataset. For different datasets, we need to set different $R\%$ that roughly match with the ratio of anomaly data in the datasets. + +Training Loss Curve. In Figure 6, we show training loss curves of the wOCC (STG-NF) model in our framework. As seen, the loss drops smoothly within each interleaving training module. Since we re-train the wOCC model from scratch in the next module, the loss suddenly increases at the beginning of the next module, but the peak magnitude of the loss is lower than that of the previous module. Please check the zoomed-in curve showing the loss curve of the + +![](images/894b7525f4f435893aebefee2473d7936a0170ca0ab8d3170d020cf0945d1248.jpg) +Fig.5: Top: different $R\%$ converge to Fig.6: Bottom: training loss curve of the similar $T_{ws}$ (on ShanghaiTech). Bottom: wOCC model during the whole repeating different $R\%$ yield similar AUC (Left: on procedure. Top: zoomed-in training loss ShanghaiTech, Right: on UBnormal). curve in the third module. + +wOCC model in the third module. Even though the model is trained alternately with another model, its loss can still decrease smoothly, not affected by the other model. The WS model's training loss curve is put into the supplemental material. + +Training Time. As mentioned in Section 3.3, although we need to train the wOCC and WS models alternately many times in multiple modules, our method is fast in training. That is because, we just train a model for one epoch and then exchange to train the other model for another epoch, which form a training step of our method. Our interleaving training converges very fast. As shown in Table 5, there are only 17 training steps during the whole training, yielding a total of $34 = 2(\mathrm{wOCC}$ and $\mathrm{WS})\times 1(\mathrm{epoch})\times 17$ (all training steps) epochs. It takes around 2.5 hours to train our method on ShanghaiTech. We have also trained the wOCC or WS model for 2 or 3 epochs in each training step, but obtained worse results, probably because training with more epochs would enhance the supervision of wrong pseudo labels in early training modules. + +# 5 Conclusion + +In this paper, we interleave the One-Classification and Weakly-Supervised models with adaptive thresholding to tackle the unsupervised video anomaly detection problem without human annotation. We face the challenges of performance fluctuation and the demanding of accurate threshold when designing the framework. To alleviate performance fluctuation, We propose the wOCC model which requires soft labels but not binary labels. To obtain a relative optimal threshold, we repeat our proposed interleaving training module, during which we propose a monotonically threshold decreasing mechanism to refine a rough threshold progressively. Extensive experiments demonstrate the effectiveness of our method. Remarkably, our method can be upgraded with the most recent development in OCC and WS VAD fields. + +# Acknowledgements + +This work was partly supported by the Natural Science Foundation of China (62072191, U21A20520, 62325204), the National Key Research and Development Program of China (2022YFE0112200), the Key-Area Research and Development Program of Guangzhou City (202206030009), and the Guangdong Basic and Applied Basic Research Fund (2023A1515030002, 2024A1515011995). + +# References + +1. Acsintoae, A., Florescu, A., Georgescu, M., Mare, T., Sumedrea, P., Ionescu, R.T., Khan, F.S., Shah, M.: Unbnormal: New benchmark for supervised open-set video anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2022) +2. Al-lahham, A., Tastan, N., Zaheer, M.Z., Nandakumar, K.: A coarse-to-fine pseudolabeling (c2fpl) framework for unsupervised video anomaly detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6793-6802 (2024) +3. Andrews, S., Tsochantaridis, I., Hofmann, T.: Support vector machines for multiple-instance learning. Advances in neural information processing systems 15 (2002) +4. Barbalau, A., Ionescu, R.T., Georgescu, M.I., Dueholm, J., Ramachandra, B., Nasrollahi, K., Khan, F.S., Moeslund, T.B., Shah, M.: Ssmtl++: Revisiting self-supervised multi-task learning for video anomaly detection. Computer Vision and Image Understanding 229, 103656 (2023) +5. Basharat, A., Gritai, A., Shah, M.: Learning object motion patterns for anomaly detection and improved object detection. In: 2008 IEEE conference on computer vision and pattern recognition. pp. 1-8. IEEE (2008) +6. Burlina, P., Joshi, N., Wang, I., et al.: Where's wally now? deep generative and discriminative embeddings for novelty detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11507-11516 (2019) +7. Dietterich, T.G., Lathrop, R.H., Lozano-Pérez, T.: Solving the multiple instance problem with axis-parallel rectangles. Artificial intelligence 89(1-2), 31-71 (1997) +8. Feng, J.C., Hong, F.T., Zheng, W.S.: Mist: Multiple instance self-training framework for video anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14009-14018 (2021) +9. Feng, X., Song, D., Chen, Y., Chen, Z., Ni, J., Chen, H.: Convolutional transformer based dual discriminator generative adversarial networks for video anomaly detection. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 5546-5554 (2021) +10. Georgescu, M.I., Barbalau, A., Ionescu, R.T., Khan, F.S., Popescu, M., Shah, M.: Anomaly detection in video via self-supervised and multi-task learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12742-12752 (2021) +11. Georgescu, M.I., Ionescu, R.T., Khan, F.S., Popescu, M., Shah, M.: A background-agnostic framework with adversarial training for abnormal event detection in video. IEEE transactions on pattern analysis and machine intelligence 44(9), 4505-4523 (2021) + +12. Gong, D., Liu, L., Le, V., Saha, B., Mansour, M.R., Venkatesh, S., Hengel, A.v.d.: Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1705-1714 (2019) +13. Hirschorn, O., Avidan, S.: Normalizing flows for human pose anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13545-13554 (2023) +14. Ionescu, R.T., Khan, F.S., Georgescu, M.I., Shao, L.: Object-centric auto-encoders and dummy anomalies for abnormal event detection in video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7842-7851 (2019) +15. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017) +16. Kingma, D.P., Dhariwal, P.: Glow: Generative flow with invertible 1x1 convolutions. Advances in neural information processing systems 31 (2018) +17. Li, S., Liu, F., Jiao, L.: Self-training multi-sequence learning with transformer for weakly supervised video anomaly detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1395-1403 (2022) +18. Liu, W., Luo, W., Lian, D., Gao, S.: Future frame prediction for anomaly detection-- a new baseline. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6536-6545 (2018) +19. Liu, Z., Nie, Y., Long, C., Zhang, Q., Li, G.: A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13588-13597 (2021) +20. Lorre, G., Rabarisoa, J., Orcesi, A., Ainouz, S., Canu, S.: Temporal contrastive pretraining for video action recognition. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 662-670 (2020) +21. Lv, H., Yue, Z., Sun, Q., Luo, B., Cui, Z., Zhang, H.: Unbiased multiple instance learning for weakly supervised video anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8022-8031 (June 2023) +22. Markovitz, A., Sharir, G., Friedman, I., Zelnik-Manor, L., Avidan, S.: Graph embedded pose clustering for anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10539-10547 (2020) +23. Medioni, G., Cohen, I., Brémond, F., Hongeng, S., Nevatia, R.: Event detection and analysis from video streams. IEEE Transactions on pattern analysis and machine intelligence 23(8), 873-889 (2001) +24. Morais, R., Le, V., Tran, T., Saha, B., Mansour, M., Venkatesh, S.: Learning regularity in skeleton trajectories for anomaly detection in videos. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11996-12004 (2019) +25. Nguyen, T.N., Meunier, J.: Anomaly detection in video sequence with appearance-motion correspondence. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1273-1283 (2019) +26. Pang, G., Yan, C., Shen, C., Hengel, A.v.d., Bai, X.: Self-trained deep ordinal regression for end-to-end video anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12173-12182 (2020) + +27. Park, H., Noh, J., Ham, B.: Learning memory-guided normality for anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14372-14381 (2020) +28. Sapkota, H., Yu, Q.: Bayesian nonparametric submodular video partition for robust anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3212-3221 (2022) +29. Shi, C., Sun, C., Wu, Y., Jia, Y.: Video anomaly detection via sequentially learning multiple pretext tasks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10330-10340 (2023) +30. Smeureanu, S., Ionescu, R.T., Popescu, M., Alexe, B.: Deep appearance features for abnormal behavior detection in video. In: Image Analysis and Processing-ICIAP 2017: 19th International Conference, Catania, Italy, September 11-15, 2017, Proceedings, Part II 19. pp. 779-789. Springer (2017) +31. Sohrab, F., Raitoharju, J., Gabbouj, M., Iosifidis, A.: Subspace support vector data description. In: 2018 24th International Conference on Pattern Recognition (ICPR). pp. 722-727. IEEE (2018) +32. Sultani, W., Chen, C., Shah, M.: Real-world anomaly detection in surveillance videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6479-6488 (2018) +33. Sun, S., Gong, X.: Hierarchical semantic contrast for scene-aware video anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22846-22856 (June 2023) +34. Tian, Y., Pang, G., Chen, Y., Singh, R., Verjans, J.W., Carneiro, G.: Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4975-4986 (2021) +35. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015) +36. Tudor Ionescu, R., Smeureanu, S., Alexe, B., Popescu, M.: Unmasking the abnormal events in video. In: Proceedings of the IEEE international conference on computer vision. pp. 2895-2903 (2017) +37. Tur, A.O., Dall'Asen, N., Bryan, C., Ricci, E.: Exploring diffusion models for unsupervised video anomaly detection. In: 2023 IEEE International Conference on Image Processing (ICIP). pp. 2540-2544. IEEE (2023) +38. Wan, B., Fang, Y., Xia, X., Mei, J.: Weakly supervised video anomaly detection via center-guided discriminative learning. In: 2020 IEEE international conference on multimedia and expo (ICME). pp. 1-6. IEEE (2020) +39. Wang, G., Wang, Y., Qin, J., Zhang, D., Bao, X., Huang, D.: Video anomaly detection by solving decoupled spatio-temporal jigsaw puzzles. In: European Conference on Computer Vision. pp. 494-511. Springer (2022) +40. Wang, J., Song, Y., Leung, T., Rosenberg, C., Wang, J., Philbin, J., Chen, B., Wu, Y.: Learning fine-grained image similarity with deep ranking. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1386-1393 (2014) +41. Wang, J., Cherian, A.: Gods: Generalized one-class discriminative subspaces for anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8201-8211 (2019) +42. Wang, Z., Zou, Y., Zhang, Z.: Cluster attention contrast for video anomaly detection. In: Proceedings of the 28th ACM international conference on multimedia. pp. 2463-2471 (2020) + +43. Wu, J.C., Hsieh, H.Y., Chen, D.J., Fuh, C.S., Liu, T.L.: Self-supervised sparse representation for video anomaly detection. In: European Conference on Computer Vision. pp. 729-745. Springer (2022) +44. Wu, P., Liu, J., Shi, Y., Sun, Y., Shao, F., Wu, Z., Yang, Z.: Not only look, but also listen: Learning multimodal violence detection under weak supervision. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16. pp. 322-339. Springer (2020) +45. Yan, C., Zhang, S., Liu, Y., Pang, G., Wang, W.: Feature prediction diffusion model for video anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5527-5537 (October 2023) +46. Yang, Z., Liu, J., Wu, Z., Wu, P., Liu, X.: Video event restoration based on keyframes for video anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14592-14601 (June 2023) +47. Zaheer, M.Z., Mahmood, A., Khan, M.H., Segu, M., Yu, F., Lee, S.I.: Generative cooperative learning for unsupervised video anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14744-14754 (2022) +48. Zaheer, M.Z., Mahmood, A., Astrid, M., Lee, S.I.: Claws: Clustering assisted weakly supervised learning with normalcy suppression for anomalous event detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 358-376. Springer (2020) +49. Zhang, C., Li, G., Qi, Y., Wang, S., Qing, L., Huang, Q., Yang, M.H.: Exploiting completeness and uncertainty of pseudo labels for weakly supervised video anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16271-16280 (June 2023) +50. Zhang, J., Qing, L., Miao, J.: Temporal convolutional network with complementary inner bag loss for weakly supervised anomaly detection. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 4030-4034. IEEE (2019) +51. Zhang, T., Lu, H., Li, S.Z.: Learning semantic scene models by object classification and trajectory clustering. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1940-1947. IEEE (2009) +52. Zhao, Y., Deng, B., Shen, C., Liu, Y., Lu, H., Hua, X.S.: Spatio-temporal autoencoder for video anomaly detection. In: Proceedings of the 25th ACM international conference on Multimedia. pp. 1933-1941 (2017) +53. Zhong, J.X., Li, N., Kong, W., Liu, S., Li, T.H., Li, G.: Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1237-1246 (2019) \ No newline at end of file diff --git a/2024/Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection/images.zip b/2024/Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..879b37f4f0bedb3ee63e3811ac87f40c5b0fc2d3 --- /dev/null +++ b/2024/Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8584fad688284a86d4ffb204aaf96cf0022eb25f67e908764dd65ee96166dad1 +size 338719 diff --git a/2024/Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection/layout.json b/2024/Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..183c8ea49e92615ebbf9025ef861f1ae04c83187 --- /dev/null +++ b/2024/Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection/layout.json @@ -0,0 +1,10637 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 111, + 471, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 111, + 471, + 166 + ], + "spans": [ + { + "bbox": [ + 143, + 111, + 471, + 166 + ], + "type": "text", + "content": "Interleaving One-Class and Weakly-Supervised Models with Adaptive Thresholding for Unsupervised Video Anomaly Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 137, + 185, + 478, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 185, + 478, + 200 + ], + "spans": [ + { + "bbox": [ + 137, + 185, + 478, + 200 + ], + "type": "text", + "content": "Yongwei Nie" + }, + { + "bbox": [ + 137, + 185, + 478, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 185, + 478, + 200 + ], + "type": "text", + "content": ", Hao Huang" + }, + { + "bbox": [ + 137, + 185, + 478, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 185, + 478, + 200 + ], + "type": "text", + "content": ", Chengjiang Long" + }, + { + "bbox": [ + 137, + 185, + 478, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 137, + 185, + 478, + 200 + ], + "type": "text", + "content": ", Qing Zhang" + }, + { + "bbox": [ + 137, + 185, + 478, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 137, + 185, + 478, + 200 + ], + "type": "text", + "content": ", Pradipta" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 239, + 201, + 375, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 201, + 375, + 217 + ], + "spans": [ + { + "bbox": [ + 239, + 201, + 375, + 217 + ], + "type": "text", + "content": "Maji" + }, + { + "bbox": [ + 239, + 201, + 375, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 239, + 201, + 375, + 217 + ], + "type": "text", + "content": ", and Hongmin Cai" + }, + { + "bbox": [ + 239, + 201, + 375, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 209, + 224, + 403, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 224, + 403, + 236 + ], + "spans": [ + { + "bbox": [ + 209, + 224, + 403, + 236 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 209, + 224, + 403, + 236 + ], + "type": "text", + "content": " South China University of Technology, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 251, + 236, + 362, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 236, + 362, + 247 + ], + "spans": [ + { + "bbox": [ + 251, + 236, + 362, + 247 + ], + "type": "text", + "content": "Meta Reality Labs, USA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 241, + 247, + 373, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 247, + 373, + 258 + ], + "spans": [ + { + "bbox": [ + 241, + 247, + 373, + 258 + ], + "type": "text", + "content": "3 Sun Yat-sen University, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 235, + 258, + 379, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 258, + 379, + 269 + ], + "spans": [ + { + "bbox": [ + 235, + 258, + 379, + 269 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 235, + 258, + 379, + 269 + ], + "type": "text", + "content": " Indian Statistical Institute, India" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 290, + 455, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 290, + 455, + 555 + ], + "spans": [ + { + "bbox": [ + 159, + 290, + 455, + 555 + ], + "type": "text", + "content": "Abstract. Video Anomaly Detection (VAD) has been extensively studied under the settings of One-Class Classification (OCC) and Weakly-Supervised learning (WS), which however both require laborious human-annotated normal/abnormal labels. In this paper, we study Unsupervised VAD (UVAD) that does not depend on any label by combining OCC and WS into a unified training framework. Specifically, we extend OCC to weighted OCC (wOCC) and propose a wOCC-WS interleaving training module, where the two models automatically generate pseudo-labels for each other. We face two challenges to make the combination effective: (1) Models' performance fluctuates occasionally during the training process due to the inevitable randomness of the pseudo labels. (2) Thresholds are needed to divide pseudo labels, making the training depend on the accuracy of user intervention. For the first problem, we propose to use wOCC requiring soft labels instead of OCC trained with hard zero/one labels, as soft labels exhibit high consistency throughout different training cycles while hard labels are prone to sudden changes. For the second problem, we repeat the interleaving training module multiple times, during which we propose an adaptive thresholding strategy that can progressively refine a rough threshold to a relatively optimal threshold, which reduces the influence of user interaction. A benefit of employing OCC and WS methods to compose a UVAD method is that we can incorporate the most recent OCC or WS model into our framework. Experiments demonstrate the effectiveness of the proposed UVAD framework. Our code is available at https://github.com/benedictstar/Joint-VAD." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 563, + 453, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 563, + 453, + 586 + ], + "spans": [ + { + "bbox": [ + 159, + 563, + 453, + 586 + ], + "type": "text", + "content": "Keywords: Video anomaly detection " + }, + { + "bbox": [ + 159, + 563, + 453, + 586 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 563, + 453, + 586 + ], + "type": "text", + "content": " One-class classification " + }, + { + "bbox": [ + 159, + 563, + 453, + 586 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 563, + 453, + 586 + ], + "type": "text", + "content": " Weakly-supervised learning " + }, + { + "bbox": [ + 159, + 563, + 453, + 586 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 563, + 453, + 586 + ], + "type": "text", + "content": " Unsupervised learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 602, + 230, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 602, + 230, + 615 + ], + "spans": [ + { + "bbox": [ + 132, + 602, + 230, + 615 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "type": "text", + "content": "Video Anomaly Detection (VAD) is a task that identifies abnormal events in a video, where the abnormal event could be a fire alarm, a flaw in an industrial" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 654, + 424, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 654, + 424, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 654, + 424, + 665 + ], + "type": "text", + "content": "* Hongmin Cai is the corresponding author. Email: hmcai@scut.edu.cn" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "product, or a traffic accident, etc. Most previous VAD methods fall in two categories: One-Class Classification (OCC) methods [12,13,18,19,33,39] and Weakly-Supervised (WS) approaches [21,32,34,38,43,49]. Both kinds of approaches require human-annotated labels for training1. Supervised VAD methods have two prominent problems. First, abnormal events are sparse and difficult to recognize, making the annotation of the training dataset a highly labor-intensive task. Secondly, the categories of abnormal events are unbounded, while human can only collect a limited scope of them, yielding the risk that the trained VAD methods cannot handle unobserved abnormal events." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 487, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 487, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 487, + 332 + ], + "type": "text", + "content": "Considering the above problems, this paper studies Unsupervised VAD (UVAD) that does not rely on labels. Despite the importance of UVAD, few methods have been developed except [47] and its subsequent works [2, 37] to the best of our knowledge. To achieve unsupervised learning, the method of [47] adversarially trains a generator and a discriminator, where the generator is an autoencoder-based reconstruction model and the discriminator is a fully-connected (FC) binary classifier. While the proposed network architecture is sophisticated, the adopted autoencoder and FC networks are relatively weak which limit the performance of the approach." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 333, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 482, + 429 + ], + "type": "text", + "content": "In this paper, we propose a new UVAD framework. Our key idea is that the techniques of OCC and WS approaches are rapidly developed, while a UVAD method is usually implemented by training two VAD models alternately generating pseudo labels for each other. This motivates us to propose a UVAD framework that directly interleaves the training of a pair of OCC and WS models. This allows us to incorporate the recent advances in the two hot research fields to implement a UVAD method. We have also attempted to combine two OCC or two WS models, but found two homogeneous models are less effective." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 429, + 482, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 621 + ], + "type": "text", + "content": "While combining OCC and WS methods, we encounter two problems preventing it from effective. First, we observe the OCC or WS model's performance fluctuates occasionally during the training process, which affects the final accuracy they can achieve. This is partly because the pseudo labels generated by each other are changed frequently, making the training not stable. To solve this problem, we extend OCC to weighted OCC (wOCC), and propose a wOCC-WS interleaving training module. Previous OCC model is trained on normal data, while our wOCC model is trained with soft labels in the range of [0,1]. The soft labels are more consistent across adjacent training cycles, making the performance of wOCC more stable than OCC under changing pseudo labels, thus suppressing the fluctuation effect. Second, we still have to set a threshold to partition normal and abnormal pseudo labels for the WS model, making our method susceptible to a user-provided hyperparameter. To reduce the influence of user interaction, we repeat the interleaving training module multiple times, during which we propose an adaptive threshold mechanism that finds an optimal threshold given just a rough initial value from the user." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 621, + 375, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 621, + 375, + 633 + ], + "spans": [ + { + "bbox": [ + 146, + 621, + 375, + 633 + ], + "type": "text", + "content": "To conclude, the contributions of this work include:" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Nie et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": "1 OCC models are trained on normal data only, which requires human labor to exclude abnormal events in preparing the training dataset." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 480, + 210 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 479, + 138 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 479, + 138 + ], + "type": "text", + "content": "- We propose a new UVAD framework that interleaves the training of a pair of OCC and WS models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 175 + ], + "type": "text", + "content": "- We improve OCC to wOCC which stabilizes the training of the UVAD framework, and also propose an adaptive thresholding strategy to alleviate the influence of user interaction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 479, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 479, + 210 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 479, + 210 + ], + "type": "text", + "content": "- Our UVAD framework is flexible to employ different pairs of OCC and WS models. Experiments on three OCC and two WS models demonstrate the effectiveness of our method." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 232, + 237, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 237, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 237, + 244 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 258, + 481, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 481, + 365 + ], + "type": "text", + "content": "Unsupervised VAD (UVAD). Unsupervised VAD requires identifying anomalies from data containing both normal and abnormal samples without any annotation. This is a challenging new task that is almost unexplored in the literature. Although significant progress has been made in OCC and WS VAD, directly applying them independently to address UVAD does not yield good results. Zaheer et al. [47] firstly raise this task and propose to solve UVAD by generative cooperative learning (GCL). Differently, our proposed method modifies both the OCC and WS VAD models and combines the modified two models together to learn from unmarked training data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 366, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 481, + 545 + ], + "type": "text", + "content": "One-Class Classification VAD (OCC). OCC-based VAD approaches only have access to normal data and try to model normal data to identify behaviors that are significantly different from normal behaviors as anomalies. Early works use hand-crafted features to help detect anomalies [5,23,40,51]. With the rapid development of deep learning, recent methods turn to normal representations extracted by using a deep neural network [26, 30, 36, 41]. Some methods identify normal patterns by using the reconstruction/prediction model to reconstruct the representations [6,12,14,18,19,24,25,27,31,33,45,46,52]. These models may lead to well-reconstructed anomalies, thereby limiting the performance of the OCC-based methods. Other OCC-based methods turn to identify normal representations by using proxy tasks [4,11,20,29,39]. What's more, recently, Hirschorn and Avidan [13] propose to build a multivariate Gaussian distribution for normal data and detect instances deviating from this distribution as anomalies. However, the above methods treat all samples equally, ignoring the potential differences in importance between normal samples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "content": "Weakly-Supervised VAD (WS). In weakly-supervised VAD, video-level annotations are available in the training stage. Sultani et al [32] first propose to use the video-level labels and solve WS VAD based on multi-instance learning (MIL) framework [3, 7]. Since then, many works [21, 28, 38, 43, 49, 50] have viewed the WS VAD task as a MIL problem. Tian et al. [34] develop to extend MIL to Top- " + }, + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "content": " MIL method by training with a robust temporal feature magnitude (RTFM) loss function. The methods mentioned above directly train in the supervision of video-level labels and the coarse-grained supervision limits the accuracy of WS models. Recently, two-stage self-training methods [8, 17, 44, 53] adopt a two-stage pipeline to use more fine-grained labels to supervise the networks more" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "text", + "content": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 479, + 251 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 479, + 251 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 479, + 251 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 479, + 251 + ], + "type": "image", + "image_path": "fc004a50af94976ab7d38dd19d88eb911212310d85aa5cd2ff8131f8cc6bb500.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "text", + "content": "Fig. 1: Pipeline of our proposed interleaved framework. The wOCC-WS interleaving training module comprises a weighted OCC (wOCC) model and a WS model, using the output of each model as the pseudo-labels to supervise the other model. Further, the training module is repeated multiple times with the adaptive threshold until the stopping criterion we set. To start the first training module, we randomly sample " + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "inline_equation", + "content": "w_{X}" + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "text", + "content": " to train the wOCC model and assign a large value as the threshold for the WS model. Once a training module converges, the last WS model of it supplies " + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "inline_equation", + "content": "w_{X}" + }, + { + "bbox": [ + 130, + 262, + 482, + 361 + ], + "type": "text", + "content": " to start the next training module and adaptive thresholding takes the outputs of all previous wOCC models into consideration for updating the threshold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 385, + 481, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 481, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 481, + 410 + ], + "type": "text", + "content": "strictly. The WS model in our framework shares a similar idea with [8] and is tailored to use finer-grained snippet-level labels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 428, + 202, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 428, + 202, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 202, + 441 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 453, + 209, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 209, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 209, + 464 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 473, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 481, + 533 + ], + "type": "text", + "content": "Our proposed UVAD framework interleaves the training of a pair of OCC and WS models. In theory, any OCC or WS methods can be incorporated into our framework. Without loss of generality, we take STG-NF [13] (a recent OCC model) and RTFM [34] (a recent WS model) as examples to introduce our method in the following sections. We test many other OCC/WS models in Section 4.5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "Figure 1 shows an overview of our framework. The basic buildingblock of our method is the wOCC-WS interleaving training module. In the module, the outputs of each model are used as pseudo labels for defining the loss function training the other model. After one module converges, we re-initialize another module to run the interleaving training again with a difference that the threshold of the WS model is updated by an adaptive threshold adjustment mechanism based on wOCC models trained in previous modules. Since wOCC has more stable behaviors than WS, we always initialize the wOCC model to start the training of a module. For the first module, we use a random initialization based on Beta distribution sampling. For other modules, we use the results of WS of the preceding module for the initialization." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Nie et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 369, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 369, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 369, + 129 + ], + "type": "text", + "content": "3.2 wOCC-WS Interleaving Training Module" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 206 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 206 + ], + "type": "text", + "content": "We first introduce the wOCC-WS interleaving training module which comprises a wOCC model and a WS model. At the very beginning, we initialize the wOCC model to start the interleaving training. Then, pseudo labels created using wOCC are used to train the WS model. Next, pseudo labels created by the WS model are used to improve the wOCC model. The whole training does not use human annotations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 207, + 482, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 207, + 482, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 482, + 351 + ], + "type": "text", + "content": "wOCC Model (STG-NF). Typically, OCC models are trained only on normal data. Directly using OCC in our UVAD framework may make the training fluctuate, i.e., the performance of the trained models sometimes gets better, but sometimes worse, finally converging to an inferior point. We use 0 to indicate normal samples, and 1 abnormal samples, where 0, 1 are hard binary labels. During training, we need to identify 0-labelled samples to train the OCC model. The hard partitioned labels may be suddenly changed from 0 to 1 or from 1 to 0, yielding the frequent change of the normal samples used to train the OCC model which is the source of the unstable training. To this end, we propose wOCC which relies on soft labels instead of hard binary labels. Soft labels are in the range of [0, 1]. For example, the soft label 0.7 may be changed to 0.6 but not 0 suddenly. This consistency reduces the fluctuation of the interleaving training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 351, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 411 + ], + "type": "text", + "content": "Taking STG-NF [13] (an OCC model) as an example, we introduce how to improve it to wOCC. STG-NF extracts a set of objects " + }, + { + "bbox": [ + 130, + 351, + 482, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{X_1,\\dots ,X_i,\\dots \\}" + }, + { + "bbox": [ + 130, + 351, + 482, + 411 + ], + "type": "text", + "content": " from the training videos, where " + }, + { + "bbox": [ + 130, + 351, + 482, + 411 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 130, + 351, + 482, + 411 + ], + "type": "text", + "content": " is a sequence of poses of a person. Each object " + }, + { + "bbox": [ + 130, + 351, + 482, + 411 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 130, + 351, + 482, + 411 + ], + "type": "text", + "content": " is the basic building block processed by STG-NF. Originally, STG-NF minimizes the negative log-likelihood across the normal data:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 233, + 421, + 480, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 421, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 233, + 421, + 480, + 434 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {o c c} = - \\log \\left(p _ {Z} \\left(f _ {S T G - N F} \\left(X _ {i} ^ {+}\\right)\\right)\\right), \\tag {1}", + "image_path": "d9bc49191ca09523639aa91372f7b2e8dda3a11c0b9ff334d3070ea720195af9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 441, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 441, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 131, + 441, + 482, + 515 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 441, + 482, + 515 + ], + "type": "inline_equation", + "content": "X_{i}^{+} \\in \\mathcal{X}^{+}" + }, + { + "bbox": [ + 131, + 441, + 482, + 515 + ], + "type": "text", + "content": " is sampled from normal data, " + }, + { + "bbox": [ + 131, + 441, + 482, + 515 + ], + "type": "inline_equation", + "content": "Z = f_{STG-NF}(X_{i}^{+})" + }, + { + "bbox": [ + 131, + 441, + 482, + 515 + ], + "type": "text", + "content": " is the feature extracted from " + }, + { + "bbox": [ + 131, + 441, + 482, + 515 + ], + "type": "inline_equation", + "content": "X_{i}^{+}" + }, + { + "bbox": [ + 131, + 441, + 482, + 515 + ], + "type": "text", + "content": " by the STG-NF, " + }, + { + "bbox": [ + 131, + 441, + 482, + 515 + ], + "type": "inline_equation", + "content": "p_{Z}" + }, + { + "bbox": [ + 131, + 441, + 482, + 515 + ], + "type": "text", + "content": " is the established distribution of the normal data. We upgrade STG-NF by introducing the soft labels " + }, + { + "bbox": [ + 131, + 441, + 482, + 515 + ], + "type": "inline_equation", + "content": "w_{X} = \\{w_{X_{1}}, \\dots, w_{X_{i}}, \\dots\\}" + }, + { + "bbox": [ + 131, + 441, + 482, + 515 + ], + "type": "text", + "content": " as the weight to supervise its training on both normal and abnormal data. Formally, for the wOCC model improved from STG-NF, we minimize the following weighted negative log-likelihood across all the data:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 211, + 525, + 480, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 525, + 480, + 537 + ], + "spans": [ + { + "bbox": [ + 211, + 525, + 480, + 537 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {w o c c} = - \\left(1 - w _ {X _ {i}}\\right) \\log \\left(p _ {Z} \\left(f _ {S T G - N F} \\left(X _ {i}\\right)\\right)\\right), \\tag {2}", + "image_path": "d57d3821b8e2cc4b56da655ba3341c3eeea3179d44d0389c162e0a0b0479c79d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 545, + 482, + 593 + ], + "type": "inline_equation", + "content": "X_{i} \\in \\mathcal{X}" + }, + { + "bbox": [ + 130, + 545, + 482, + 593 + ], + "type": "text", + "content": " is sampled from the whole training dataset rather than the normal part and " + }, + { + "bbox": [ + 130, + 545, + 482, + 593 + ], + "type": "inline_equation", + "content": "w_{X_i} \\in [0,1]" + }, + { + "bbox": [ + 130, + 545, + 482, + 593 + ], + "type": "text", + "content": " is the weight of the object " + }, + { + "bbox": [ + 130, + 545, + 482, + 593 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 130, + 545, + 482, + 593 + ], + "type": "text", + "content": " indicating the anomaly degree of the object which serves as the soft label. The object with smaller " + }, + { + "bbox": [ + 130, + 545, + 482, + 593 + ], + "type": "inline_equation", + "content": "w_{X_i}" + }, + { + "bbox": [ + 130, + 545, + 482, + 593 + ], + "type": "text", + "content": " means more normal which is better captured by the learned distribution of wOCC." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Pseudo Labels " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "l_{S}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " from wOCC. Now we use the trained wOCC model to generate pseudo labels " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "l_{S} = \\{l_{S_{1}},\\dots ,l_{S_{i}},\\dots \\}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " for training the WS model. \"Snippet\" is the basic element processed by WS models. We first split the training videos into snippets, obtaining a set of snippets " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "S = \\{S_1,\\dots ,S_i,\\dots \\}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". Our aim is to compute the label " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "l_{S_i}\\in \\{0,1\\}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " for each snippet " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " be the anomaly score of the object " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " computed by the trained wOCC model. The anomaly score " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{s}_i" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 90, + 448, + 102 + ], + "type": "text", + "content": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "of a video snippet " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " is computed by averaging all the anomaly scores of objects in the snippet. To generate pseudo label " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "l_{S_i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": ", one way is to check whether " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\hat{s}_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " is larger than a threshold " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "(l_{S_i} = 1)" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " or not " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "(l_{S_i} = 0)" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": ". However, as the training goes, the anomaly scores of snippets may vary in a large range, making the absolute threshold invalid. Instead, we sort the snippets according to their anomaly scores in descending order, and get the index of the snippet " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathrm{Rank}(\\hat{s}_i)" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " in the sorted list. Finally, the label " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "l_{S_i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " of snippet " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " is computed by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 231, + 208, + 481, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 208, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 231, + 208, + 481, + 239 + ], + "type": "interline_equation", + "content": "l _ {S _ {i}} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\quad \\operatorname {R a n k} \\left(\\hat {s} _ {i}\\right) < T _ {w s}, \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {3}", + "image_path": "b7bd835542f32aee443c4408e396b5a58f1e7e123a42e7f7982e118b5a1d2e22.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 244, + 466, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 244, + 466, + 256 + ], + "spans": [ + { + "bbox": [ + 131, + 244, + 466, + 256 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 244, + 466, + 256 + ], + "type": "inline_equation", + "content": "T_{ws}" + }, + { + "bbox": [ + 131, + 244, + 466, + 256 + ], + "type": "text", + "content": " is the relative threshold not affected by concrete anomaly scores." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 257, + 482, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 388 + ], + "type": "text", + "content": "WS Model (RTFM). With snippet-level pseudo annotations " + }, + { + "bbox": [ + 130, + 257, + 482, + 388 + ], + "type": "inline_equation", + "content": "l_{S}" + }, + { + "bbox": [ + 130, + 257, + 482, + 388 + ], + "type": "text", + "content": " (obtained in Eq. 3), we now define positive and negative bags and use them to train a WS model. Previously, approaches [32,34] usually treat snippets of the same video as a bag based on video-level labels. Differently, thanks to the snippet-level labels provided by the wOCC model, we can compose positive and negative bags more flexibly. In our method, a positive bag " + }, + { + "bbox": [ + 130, + 257, + 482, + 388 + ], + "type": "inline_equation", + "content": "B^{+} = \\{S_{i}^{+}\\}_{i=1}^{C}" + }, + { + "bbox": [ + 130, + 257, + 482, + 388 + ], + "type": "text", + "content": " is composed of randomly selected " + }, + { + "bbox": [ + 130, + 257, + 482, + 388 + ], + "type": "inline_equation", + "content": "C = 16" + }, + { + "bbox": [ + 130, + 257, + 482, + 388 + ], + "type": "text", + "content": " abnormal snippets. Similarly, a negative bag " + }, + { + "bbox": [ + 130, + 257, + 482, + 388 + ], + "type": "inline_equation", + "content": "B^{-} = \\{S_{j}^{-}\\}_{j=1}^{C}" + }, + { + "bbox": [ + 130, + 257, + 482, + 388 + ], + "type": "text", + "content": " is composed of randomly selected " + }, + { + "bbox": [ + 130, + 257, + 482, + 388 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 257, + 482, + 388 + ], + "type": "text", + "content": " normal snippets. Taking RTFM [34] as an example, the WS method employs multiple instance learning, requiring the average Top-" + }, + { + "bbox": [ + 130, + 257, + 482, + 388 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 257, + 482, + 388 + ], + "type": "text", + "content": " maximum magnitudes of features in positive bags exceeds that in negative bags:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 199, + 396, + 481, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 396, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 199, + 396, + 481, + 423 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {w s} = \\max (0, m - d _ {R T F M} (B ^ {+}) + d _ {R T F M} (B ^ {-})) + \\tag {4} \\\\ \\mathcal {L} _ {B C E} \\left(B ^ {+}\\right) + \\mathcal {L} _ {B C E} \\left(B ^ {-}\\right), \\\\ \\end{array}", + "image_path": "9e347508da4f111e9b7ffb7999dfd1b873daae058e479f1e2bac991553e36e22.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "text", + "content": " is the margin, and " + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "inline_equation", + "content": "d_{RTFM}(B)" + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "text", + "content": " can be " + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "inline_equation", + "content": "B^{+}" + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "inline_equation", + "content": "B^{-}" + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "text", + "content": ") returns the average of the Top-" + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "text", + "content": " maximum magnitudes of features. Finally, " + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BCE}" + }, + { + "bbox": [ + 130, + 431, + 482, + 467 + ], + "type": "text", + "content": " is the binary cross-entropy-based classification loss:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 169, + 475, + 481, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 475, + 481, + 488 + ], + "spans": [ + { + "bbox": [ + 169, + 475, + 481, + 488 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {B C E} (B) = - Y \\log \\left(f _ {R T F M} (B)\\right) + (1 - Y) \\log \\left(1 - f _ {R T F M} (B)\\right), \\tag {5}", + "image_path": "f3aacbd848c69f6c57f61133ad4098577247d5eca6ef0a3ee9d5f56c88ee802c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 494, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 494, + 482, + 540 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 494, + 482, + 540 + ], + "type": "text", + "content": " is the label of the bag " + }, + { + "bbox": [ + 130, + 494, + 482, + 540 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 494, + 482, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 494, + 482, + 540 + ], + "type": "inline_equation", + "content": "f_{RTFM}(B)" + }, + { + "bbox": [ + 130, + 494, + 482, + 540 + ], + "type": "text", + "content": " returns the average of the Top- " + }, + { + "bbox": [ + 130, + 494, + 482, + 540 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 494, + 482, + 540 + ], + "type": "text", + "content": " maximum anomaly scores. Note that " + }, + { + "bbox": [ + 130, + 494, + 482, + 540 + ], + "type": "inline_equation", + "content": "f_{RTFM}(B)" + }, + { + "bbox": [ + 130, + 494, + 482, + 540 + ], + "type": "text", + "content": " is different from " + }, + { + "bbox": [ + 130, + 494, + 482, + 540 + ], + "type": "inline_equation", + "content": "d_{RTFM}(B)" + }, + { + "bbox": [ + 130, + 494, + 482, + 540 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 494, + 482, + 540 + ], + "type": "inline_equation", + "content": "f_{RTFM}(B)" + }, + { + "bbox": [ + 130, + 494, + 482, + 540 + ], + "type": "text", + "content": " returns the anomaly scores while " + }, + { + "bbox": [ + 130, + 494, + 482, + 540 + ], + "type": "inline_equation", + "content": "d_{RTFM}(B)" + }, + { + "bbox": [ + 130, + 494, + 482, + 540 + ], + "type": "text", + "content": " returns the magnitude of the features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 542, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 542, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 542, + 482, + 591 + ], + "type": "text", + "content": "Pseudo Labels " + }, + { + "bbox": [ + 130, + 542, + 482, + 591 + ], + "type": "inline_equation", + "content": "w_{X}" + }, + { + "bbox": [ + 130, + 542, + 482, + 591 + ], + "type": "text", + "content": " from WS. wOCC model needs soft labels " + }, + { + "bbox": [ + 130, + 542, + 482, + 591 + ], + "type": "inline_equation", + "content": "w_{X}" + }, + { + "bbox": [ + 130, + 542, + 482, + 591 + ], + "type": "text", + "content": ". With the trained WS model, we compute the anomaly score " + }, + { + "bbox": [ + 130, + 542, + 482, + 591 + ], + "type": "inline_equation", + "content": "\\hat{x}_i" + }, + { + "bbox": [ + 130, + 542, + 482, + 591 + ], + "type": "text", + "content": " of object " + }, + { + "bbox": [ + 130, + 542, + 482, + 591 + ], + "type": "inline_equation", + "content": "X_i" + }, + { + "bbox": [ + 130, + 542, + 482, + 591 + ], + "type": "text", + "content": " by simply setting it as the anomaly score of the snippet containing the object " + }, + { + "bbox": [ + 130, + 542, + 482, + 591 + ], + "type": "inline_equation", + "content": "X_i" + }, + { + "bbox": [ + 130, + 542, + 482, + 591 + ], + "type": "text", + "content": ". Then, the soft label " + }, + { + "bbox": [ + 130, + 542, + 482, + 591 + ], + "type": "inline_equation", + "content": "w_{X_i}" + }, + { + "bbox": [ + 130, + 542, + 482, + 591 + ], + "type": "text", + "content": " is directly set as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 286, + 599, + 481, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 599, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 286, + 599, + 481, + 611 + ], + "type": "interline_equation", + "content": "w _ {X _ {i}} = \\hat {x} _ {i}. \\tag {6}", + "image_path": "9954cf28463808f7f7a27a75f7b524787136f5b2460a8f5c16a4ea6ccf07fab3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 617, + 444, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 444, + 629 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 444, + 629 + ], + "type": "inline_equation", + "content": "\\hat{x}_i" + }, + { + "bbox": [ + 131, + 617, + 444, + 629 + ], + "type": "text", + "content": " is already in the range of " + }, + { + "bbox": [ + 131, + 617, + 444, + 629 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 131, + 617, + 444, + 629 + ], + "type": "text", + "content": ", so we do not need to normalize " + }, + { + "bbox": [ + 131, + 617, + 444, + 629 + ], + "type": "inline_equation", + "content": "w_{X_i}" + }, + { + "bbox": [ + 131, + 617, + 444, + 629 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Initializing the Interleaving Training Module. As discussed above, in our proposed interleaving training module, wOCC training relies on pseudo labels generated by the WS model, and conversely, WS training requires pseudo" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Nie et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": "labels derived from the wOCC. This egg and chicken problem needs to be solved at the very beginning. Our solution is to provide the wOCC model with randomly generated soft labels to start the interleaving training. Usually, the training dataset contains much more normal samples than abnormal ones, which means most soft labels are around 0 while a small portion of them are near 1. Therefore, we randomly sample the weight " + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "inline_equation", + "content": "w_{X}" + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": " from the Beta distribution, i.e., " + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "inline_equation", + "content": "w_{X} \\sim \\mathrm{Beta}(\\alpha, \\beta)" + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "inline_equation", + "content": "\\beta = 5" + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": ". The sampled weights " + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "inline_equation", + "content": "w_{X}" + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": " are mostly around 0, and a small portion of them is close to 1, satisfying the prior about the distribution of normal and abnormal data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 480, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 480, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 480, + 368 + ], + "type": "text", + "content": "Convergence Analysis. We have discussed the mechanism and the way to start the interleaving training module above. Then, in the following, we empirically analyze its convergence, which is validated by the experiment in Figure 6. At the very beginning, we randomly initialize the weight " + }, + { + "bbox": [ + 130, + 224, + 480, + 368 + ], + "type": "inline_equation", + "content": "w_{X}" + }, + { + "bbox": [ + 130, + 224, + 480, + 368 + ], + "type": "text", + "content": " to train our wOCC model. Although the random initialization is not precise, the wOCC model can still learn normal patterns from the training data, as we assume there is more normal data than abnormal data in the training dataset. Therefore, the wOCC model can produce relatively reliable pseudo labels for the WS model. This is very important as the WS model has a high requirement on the quality of the pseudo labels. In turn, with the trained WS model, we can provide better pseudo supervision to train a stronger wOCC model than the random initialization. In this way, we achieve mutual improvement between two models until convergence." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 369, + 480, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 480, + 440 + ], + "type": "text", + "content": "The above analysis implies the reason why we do not choose to interleave the training of two OCC models or two WS models. The OCC model can provide reliable initialization, while the WS model explicitly enlarges the gap between normal and abnormal samples. Two OCC models fail to utilize the abnormal data, while for two WS models it is not known how to provide a reliable initialization to start the training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 462, + 411, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 462, + 411, + 474 + ], + "spans": [ + { + "bbox": [ + 131, + 462, + 411, + 474 + ], + "type": "text", + "content": "3.3 Repeating Procedure with Adaptive Thresholding" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 485, + 479, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 479, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 479, + 581 + ], + "type": "text", + "content": "In the above, the wOCC and WS models need to provide pseudo labels to supervise the other. As the wOCC model relies on soft labels " + }, + { + "bbox": [ + 130, + 485, + 479, + 581 + ], + "type": "inline_equation", + "content": "w_{X}" + }, + { + "bbox": [ + 130, + 485, + 479, + 581 + ], + "type": "text", + "content": " as defined in Eq. 6, its pseudo labels can be directly provided by the WS model without thresholding. However, as defined in Eq. 3, we need to specify a threshold for the WS model since it requires binary hard partition labels to construct positive and negative bags. To alleviate the influence of human interaction, we propose a repeating procedure of the interleaving training module which can adaptively refine a user-specified rough threshold to a relatively optimal one." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "type": "text", + "content": "Starting and running the next interleaving training module. As shown in Figure 1, we run several interleaving training modules one after another. Recall that we provide " + }, + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "type": "inline_equation", + "content": "w_{X}" + }, + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "type": "text", + "content": " by random initialization to start the first interleaving training module as discussed in Section 3.2. To start the next module, we use the final WS model in the last module to provide the labels " + }, + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "type": "inline_equation", + "content": "w_{X}" + }, + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "type": "text", + "content": ". For the wOCC and WS models in a new interleaving training module, we train them from scratch, not inheriting parameters from the last module." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "text", + "content": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 177 + ], + "type": "text", + "content": "Adaptive Thresholding. The threshold is fixed in each module. Between two modules, we employ an adaptive thresholding mechanism to refine the threshold. As the repeating procedure runs, we obtain a series of thresholds " + }, + { + "bbox": [ + 132, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\{T_{ws}^{1}, T_{ws}^{2}, \\dots, T_{ws}^{i}, \\dots\\}" + }, + { + "bbox": [ + 132, + 116, + 482, + 177 + ], + "type": "text", + "content": " corresponding to the interleaving training modules. The key idea of adaptive thresholding is to ensure:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 241, + 186, + 481, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 186, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 241, + 186, + 481, + 201 + ], + "type": "interline_equation", + "content": "T _ {w s} ^ {1} \\geq T _ {w s} ^ {2} \\geq \\dots \\geq T _ {w s} ^ {i} \\geq \\dots . \\tag {7}", + "image_path": "671954d860234a92d01cc82198542ea7b794c732086215c8ada92970e1878322.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 210, + 480, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 210, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 210, + 480, + 258 + ], + "type": "text", + "content": "Starting from a sufficiently large (larger than the optimal threshold) value of " + }, + { + "bbox": [ + 130, + 210, + 480, + 258 + ], + "type": "inline_equation", + "content": "T_{ws}^{1}" + }, + { + "bbox": [ + 130, + 210, + 480, + 258 + ], + "type": "text", + "content": ", there must be one threshold " + }, + { + "bbox": [ + 130, + 210, + 480, + 258 + ], + "type": "inline_equation", + "content": "T_{ws}^{*} \\in \\{T_{ws}^{1}, T_{ws}^{2}, \\dots, T_{ws}^{i}, \\dots\\}" + }, + { + "bbox": [ + 130, + 210, + 480, + 258 + ], + "type": "text", + "content": " that is closest to the optimal threshold. The problem is how to enforce the monotonically decreasing of the threshold and how to stop the repeating procedure at the optimal point." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 258, + 480, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 480, + 330 + ], + "type": "text", + "content": "First, we initialize " + }, + { + "bbox": [ + 130, + 258, + 480, + 330 + ], + "type": "inline_equation", + "content": "T_{ws}^{1} = R\\% \\times N" + }, + { + "bbox": [ + 130, + 258, + 480, + 330 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 258, + 480, + 330 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 258, + 480, + 330 + ], + "type": "text", + "content": " is a user-specific hyper-parameter, e.g. " + }, + { + "bbox": [ + 130, + 258, + 480, + 330 + ], + "type": "inline_equation", + "content": "R = 30" + }, + { + "bbox": [ + 130, + 258, + 480, + 330 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 258, + 480, + 330 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 258, + 480, + 330 + ], + "type": "text", + "content": " is the total number of snippets in the training dataset. This setting makes " + }, + { + "bbox": [ + 130, + 258, + 480, + 330 + ], + "type": "inline_equation", + "content": "T_{ws}^{1}" + }, + { + "bbox": [ + 130, + 258, + 480, + 330 + ], + "type": "text", + "content": " a large enough threshold, since generally the ratio of abnormal data in the training dataset is less than " + }, + { + "bbox": [ + 130, + 258, + 480, + 330 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 258, + 480, + 330 + ], + "type": "text", + "content": ". We test different " + }, + { + "bbox": [ + 130, + 258, + 480, + 330 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 258, + 480, + 330 + ], + "type": "text", + "content": " (e.g., " + }, + { + "bbox": [ + 130, + 258, + 480, + 330 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 130, + 258, + 480, + 330 + ], + "type": "text", + "content": ") in the experimental section and find that it does not influence our final results much, demonstrating the robustness of our method to the user interaction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "content": "Then, to ensure Eq. 7, our adaptive thresholding method determines the threshold used in the current module based on wOCC models trained in all previous modules. Each training module produces multiple wOCC models, obtaining one after each loop between wOCC and WS. Let " + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "content": " be the number of wOCC models trained from module 1 to module " + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "content": ". When computing " + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "inline_equation", + "content": "T_{ws}^{i+1}" + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "content": " for module " + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "inline_equation", + "content": "i+1" + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "inline_equation", + "content": "i \\geq 1" + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "content": ", we take all the " + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "content": " wOCC models into consideration. Specifically, for each wOCC, e.g., the " + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "content": " one where " + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "inline_equation", + "content": "j \\in [1, M_{i}]" + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "content": ", we use it to identify " + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "inline_equation", + "content": "R\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "content": " of snippets in the training dataset that have high anomaly scores, and denote the resulting set of snippets as " + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "inline_equation", + "content": "A_{j}" + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "content": ". We then compute intersection between all sets and count the number of the resulting intersection as the threshold:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 230, + 461, + 480, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 461, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 230, + 461, + 480, + 475 + ], + "type": "interline_equation", + "content": "T _ {w s} ^ {i + 1} = \\operatorname {N u m} \\left(A _ {1} \\cap A _ {2} \\cap \\dots \\cap A _ {M _ {i}}\\right), \\tag {8}", + "image_path": "e636f522d5f1095959efefb8874131a4d278e3986a653c7233a071efedc7432a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\mathrm{Num}(\\cdot)" + }, + { + "bbox": [ + 130, + 485, + 482, + 568 + ], + "type": "text", + "content": " counts the number of elements in the intersection. The operation of computing intersection ensures that the obtained threshold is monotonically decreased. In essence, the threshold computed in this way means the number of snippets viewed as abnormal by all wOCC models. At the early stages, the number of wOCC models is small, therefore the size of the intersection is large, yielding large thresholds. As more wOCC models involved, it is more difficult to form a consensus, thus producing smaller thresholds." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "To elucidate the effect of the adaptive thresholding, we visualize how WS models are gradually improved as the repeating procedure runs in Figure 2. In the figure, we show two examples. Taking the one on the left as an example, it shows a video with an abnormal event in the middle. At the beginning when " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "T_{ws} = 1643" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", the WS model wrongly identifies the abnormal event as normal. As the repeating procedure runs, " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "T_{ws}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " becomes smaller and smaller, which is finally reduced to 834 after repeating 5 times, and the WS model begins to identify the anomaly event progressively." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Nie et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 187, + 114, + 304, + 243 + ], + "blocks": [ + { + "bbox": [ + 187, + 114, + 304, + 243 + ], + "lines": [ + { + "bbox": [ + 187, + 114, + 304, + 243 + ], + "spans": [ + { + "bbox": [ + 187, + 114, + 304, + 243 + ], + "type": "image", + "image_path": "860735d1d9744c90ccc3a68a5c408ca563845b9f3d80e1d6e2fbe130fac10d2f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 252, + 480, + 275 + ], + "lines": [ + { + "bbox": [ + 131, + 252, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 131, + 252, + 480, + 275 + ], + "type": "text", + "content": "Fig. 2: As repeating procedure goes, " + }, + { + "bbox": [ + 131, + 252, + 480, + 275 + ], + "type": "inline_equation", + "content": "T_{ws}" + }, + { + "bbox": [ + 131, + 252, + 480, + 275 + ], + "type": "text", + "content": " monotonically decreases, improving the WS model gradually. Left: example from ShanghaiTech. Right: example from UBnormal." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 425, + 243 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 425, + 243 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 425, + 243 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 425, + 243 + ], + "type": "image", + "image_path": "662a2edb42455b1676dcec073ef668028956d895e7fcafe2b037cf985cbf2f0d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "text", + "content": "Stopping Criterion. As our goal is to find the threshold " + }, + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "inline_equation", + "content": "T_{ws}^{*}" + }, + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "text", + "content": " close to the optimal threshold, we need to stop the repeating procedure at the right time. Experimentally, we find that " + }, + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "inline_equation", + "content": "T_{ws}" + }, + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "text", + "content": " drops fast at the first few interleaving training modules, then the rate of change of " + }, + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "inline_equation", + "content": "T_{ws}" + }, + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "text", + "content": " becomes smaller which indicates that the wOCCs achieve a consensus about anomalies. The size of the consensus set is probably the actual number of abnormal snippets in the dataset. Based on this observation, once change rate of " + }, + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "inline_equation", + "content": "T_{ws}" + }, + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "text", + "content": " between two consecutive training modules is less than " + }, + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "inline_equation", + "content": "Q\\%" + }, + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "text", + "content": " of the first change rate at the very beginning, we stop the repeating procedure. Please see the high correlation between the change rate of " + }, + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "inline_equation", + "content": "T_{ws}" + }, + { + "bbox": [ + 130, + 297, + 482, + 416 + ], + "type": "text", + "content": " and the accuracy of the WS model in Figure 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 416, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 482, + 525 + ], + "type": "text", + "content": "Training Time Analysis. Though we need to repeat the interleaving training module multiple times, our training is as fast as the original wOCC or WS method. First, the repeating procedure usually stops with fewer than 6 modules thanks to the stopping criterion. Furthermore, we train wOCC (or WS) for just one epoch, and then exchange to train WS (or wOCC) for another epoch. If we have 6 modules, each running an average of 5 loops of interleaving training, the wOCC and WS models are trained for just 30 epochs each. The original wOCC or WS models are also trained for the similar number of epochs in their default experimental settings. The detailed experiments are in Section 4.5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 540, + 208, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 540, + 208, + 551 + ], + "spans": [ + { + "bbox": [ + 131, + 540, + 208, + 551 + ], + "type": "text", + "content": "3.4 Inference" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 558, + 481, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 558, + 481, + 583 + ], + "spans": [ + { + "bbox": [ + 131, + 558, + 481, + 583 + ], + "type": "text", + "content": "At the inference stage, both wOCC and WS models can be used to detect anomalies. We provide results of both models when compared with previous methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 599, + 231, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 599, + 231, + 613 + ], + "spans": [ + { + "bbox": [ + 131, + 599, + 231, + 613 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 623, + 326, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 326, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 326, + 635 + ], + "type": "text", + "content": "4.1 Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "ShanghaiTech. The ShanghaiTech dataset [18] contains 437 videos and was created as a benchmark for OCC with only normal videos available in the training" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 90, + 447, + 102 + ], + "type": "text", + "content": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "type": "text", + "content": "set. Zhong et al. [53] reorganized the dataset to enable the training of WS systems. The new split contains 63 abnormal videos and 175 normal videos, while there are 44 anomalous and 155 normal videos in the new testing split. In our unsupervised setting, we follow the split for WS approaches but do not provide video-level annotations in the training stage." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 479, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 479, + 235 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 479, + 235 + ], + "type": "text", + "content": "UBnormal. The UBnormal dataset [1] is a synthetic open-set benchmark, containing 268 training videos, 64 validation videos and 211 test videos. Normal and abnormal videos are mixed in the three subsets. The dataset is challenging because of disjoint sets of anomaly types in training and testing sets. We follow its original organization but train our proposed UVAD without any annotations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 236, + 479, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 236, + 479, + 259 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 479, + 259 + ], + "type": "text", + "content": "Evaluation Metrics. Following prior VAD arts [13, 34], we use the frame-level area under ROC curve (AUC) to measure VAD's accuracy for both datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 279, + 330, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 330, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 330, + 292 + ], + "type": "text", + "content": "4.2 wOCC and WS models Employed" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 300, + 480, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 480, + 346 + ], + "type": "text", + "content": "As stated, our method is a flexible UVAD framework that can use different wOCC and WS models. We test improving three OCC models to wOCC, including the AE model used in [47], Jigsaw [39], and STG-NF [13]. The wOCC losses for " + }, + { + "bbox": [ + 130, + 300, + 480, + 346 + ], + "type": "inline_equation", + "content": "AE" + }, + { + "bbox": [ + 130, + 300, + 480, + 346 + ], + "type": "text", + "content": " and Jigsaw are put into supplemental material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 349, + 479, + 516 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 131, + 349, + 479, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 349, + 479, + 369 + ], + "spans": [ + { + "bbox": [ + 131, + 349, + 479, + 369 + ], + "type": "text", + "content": "(1) AE. GCL [47] proposes to use an AutoEncoder that reconstructs the features extracted from videos as their OCC model. We use it here too." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 373, + 479, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 373, + 479, + 418 + ], + "spans": [ + { + "bbox": [ + 131, + 373, + 479, + 418 + ], + "type": "text", + "content": "(2) Jigsaw. The OCC model Jigsaw [39] addresses VAD by solving a pretext task: spatio-temporal jigsaw puzzles. It divides the spatio-temporal space of a video into smaller cubes, then shuffles the positions of the cubes, and finally tries to restore the original positions of the cubes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 420, + 479, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 420, + 479, + 443 + ], + "spans": [ + { + "bbox": [ + 131, + 420, + 479, + 443 + ], + "type": "text", + "content": "(3) STG-NF. STG-NF [13] extracts people's pose sequences, and extends Glow [16] to build multivariate Gaussian distribution of normal action sequences." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 145, + 445, + 479, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 445, + 479, + 455 + ], + "spans": [ + { + "bbox": [ + 145, + 445, + 479, + 455 + ], + "type": "text", + "content": "For WS, we test the method proposed in Sultani et al. [32] and RTFM [34]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 457, + 479, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 457, + 479, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 457, + 479, + 491 + ], + "type": "text", + "content": "(1) Sultani et al. [32]. Sultani et al. [32] propose the first MIL VAD model that maximizes the separability between a positive bag containing snippets of an abnormal video and a negative bag with snippets of a normal video." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 493, + 479, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 493, + 479, + 516 + ], + "spans": [ + { + "bbox": [ + 131, + 493, + 479, + 516 + ], + "type": "text", + "content": "(2) RTFM. RTFM [34] extends Sultani et al. [32] to compare the Top-" + }, + { + "bbox": [ + 131, + 493, + 479, + 516 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 493, + 479, + 516 + ], + "type": "text", + "content": " largest-magnitude snippets between positive and negative bags." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 536, + 278, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 278, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 278, + 548 + ], + "type": "text", + "content": "4.3 Implementation Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 557, + 479, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 479, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 479, + 639 + ], + "type": "text", + "content": "We implement our method in PyTorch. For STG-NF and RTFM (implementation details of other wOCC/WS models are put in supplemental material), they are optimized by Adam optimizer with " + }, + { + "bbox": [ + 130, + 557, + 479, + 639 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 130, + 557, + 479, + 639 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 557, + 479, + 639 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 130, + 557, + 479, + 639 + ], + "type": "text", + "content": ", and a weight decay of 0.0005. The batch size of STG-NF is 256, and that of RTFM is 32. The learning rate is " + }, + { + "bbox": [ + 130, + 557, + 479, + 639 + ], + "type": "inline_equation", + "content": "5e - 4" + }, + { + "bbox": [ + 130, + 557, + 479, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 557, + 479, + 639 + ], + "type": "inline_equation", + "content": "1e - 3" + }, + { + "bbox": [ + 130, + 557, + 479, + 639 + ], + "type": "text", + "content": " for STG-NF and RTFM respectively. For RTFM, the margin " + }, + { + "bbox": [ + 130, + 557, + 479, + 639 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 557, + 479, + 639 + ], + "type": "text", + "content": " is set to 100. During interleaving training, we train a model for one epoch and then exchange to train the other model for another epoch." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "Our method has two hyperparameters shared by all types of wOCC or WS models. One is " + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "R\\%" + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": " which determines the initial threshold " + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "T_{ws}^{1}" + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": ". By default, we set" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Nie et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 201, + 481, + 306 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "text", + "content": "Table 1: Comparison with previous approaches. For wOCC, we use STG-NF [13]. For WS, we use RTFM [34]. We test using C3D [35] and I3D [15] to extract features for RTFM. Our " + }, + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "inline_equation", + "content": "w_{OCC}" + }, + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "text", + "content": " gives the AUC of the wOCC model. Our " + }, + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "inline_equation", + "content": "WS" + }, + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "text", + "content": " gives the AUC of the WS model. The methods split by the dashed line are trained as a whole in our framework. * means we reimplement the method by taking I3D as the feature. ** means our method degenerated to the supervised OCC or WS method. *** means the results are reproduced by us using the official code." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 201, + 481, + 306 + ], + "lines": [ + { + "bbox": [ + 138, + 201, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 481, + 306 + ], + "type": "table", + "html": "
UnsupervisedOne-Class ClassificationWeakly Supervised
MethodFeaturesSTech AUC %UB AUC %MethodSTech AUC %UB AUC %MethodFeaturesSTech AUC %UB AUC %
AEAllDataI3D60.51-MemAE [12]71.20-GCN [53]C3D76.44-
Frame Prediction [18]73.40-GCN [53]TSN84.44-
STG-NFAllData [13]-80.2970.48Markovitz et al. [22]76.1052.00Zhang et al. [50]C3D82.50-
HF2-VAD [19]76.20-Sultani et al.*** [32]I3D84.5354.12
GCL* [47]I3D76.14-CT-D2GAN [9]77.70-AR-Net [38]I3D85.38-
CAC [42]79.30-CLAWS [48]C3D89.67-
OurwOCC-81.5071.67SSMTL [10]82.40-MIST [8]I3D94.83-
OurwSC3D85.4359.05Georgescu et al. [11]82.7059.30Li et al. [17]I3D96.08-
SSMTL++ [4]83.8062.10RTFM*** [34]C3D91.5162.30
OurwOCC-82.5774.76Jigsaw [39]84.3056.40RTFM*** [34]I3D96.1066.83
STG-NF [13]85.9071.80S3R [43]I3D97.48-
OurwSI3D88.1863.10OurwOCC (STG-NF)86.3772.81OurwS (RTFM)I3D96.3367.42
", + "image_path": "73fe892bf66a361199e8f434b1eecaf3fad7858fb132253057e8c607845c95a5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 137, + 360, + 469, + 396 + ], + "blocks": [ + { + "bbox": [ + 130, + 317, + 478, + 350 + ], + "lines": [ + { + "bbox": [ + 130, + 317, + 478, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 478, + 350 + ], + "type": "text", + "content": "Table 2: Ablation study on wOCC im- Table 3: Ablation study on using dif proved from different OCC models, always ferent WS models in our method, always with RTFM as the WS model. with STG-NF as the wOCC model." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 360, + 469, + 396 + ], + "lines": [ + { + "bbox": [ + 137, + 360, + 469, + 396 + ], + "spans": [ + { + "bbox": [ + 137, + 360, + 469, + 396 + ], + "type": "table", + "html": "
OCC ModelOurwOCCOurWS(RTFM)GCLClassifierWS ModelOurwOCC(STG-NF)OurWS
AE70.9978.9076.14Sultani et al. [32]81.9277.41
Jigsaw [39]81.2385.35-RTFM [13]82.5788.18
STG-NF [13]82.5788.18-
", + "image_path": "50e40612f53917ec5fef2fbd678112f2589477766aaea2a58d175ea207bce192.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 420, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 482, + 468 + ], + "type": "text", + "content": "it to " + }, + { + "bbox": [ + 130, + 420, + 482, + 468 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 130, + 420, + 482, + 468 + ], + "type": "text", + "content": " for the ShanghaiTech dataset which contains a small ratio of abnormal data. The UBnormal dataset contains relatively more abnormal data and we set " + }, + { + "bbox": [ + 130, + 420, + 482, + 468 + ], + "type": "inline_equation", + "content": "R\\%" + }, + { + "bbox": [ + 130, + 420, + 482, + 468 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 420, + 482, + 468 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 420, + 482, + 468 + ], + "type": "text", + "content": ". The other parameter is " + }, + { + "bbox": [ + 130, + 420, + 482, + 468 + ], + "type": "inline_equation", + "content": "Q\\%" + }, + { + "bbox": [ + 130, + 420, + 482, + 468 + ], + "type": "text", + "content": " used to stop the repeating procedure. For both datasets, we set " + }, + { + "bbox": [ + 130, + 420, + 482, + 468 + ], + "type": "inline_equation", + "content": "Q\\%" + }, + { + "bbox": [ + 130, + 420, + 482, + 468 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 420, + 482, + 468 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 420, + 482, + 468 + ], + "type": "text", + "content": ". Analysis of " + }, + { + "bbox": [ + 130, + 420, + 482, + 468 + ], + "type": "inline_equation", + "content": "Q\\%" + }, + { + "bbox": [ + 130, + 420, + 482, + 468 + ], + "type": "text", + "content": " is put into Supp." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 487, + 358, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 487, + 358, + 500 + ], + "spans": [ + { + "bbox": [ + 131, + 487, + 358, + 500 + ], + "type": "text", + "content": "4.4 Comparison with Previous Approaches" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "type": "text", + "content": "From left to right, Table 1 shows the comparison between our and previous approaches under Unsupervised, One-Class Classification and Weakly-Supervised. Since the unsupervised methods are scarce, besides GCL [47], we compare with " + }, + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{AE}_{\\text{AllData}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "type": "text", + "content": " which means training the AE model on the whole training dataset containing both normal and abnormal data. We also compare with STG-NF" + }, + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "type": "inline_equation", + "content": "_{\\text{AllData}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "type": "text", + "content": ". For our method, we adopt STG-NF [13] as wOCC, and RTFM [34] as WS." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 488, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 488, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 488, + 666 + ], + "type": "text", + "content": "First, Our " + }, + { + "bbox": [ + 130, + 582, + 488, + 666 + ], + "type": "inline_equation", + "content": "wOCC" + }, + { + "bbox": [ + 130, + 582, + 488, + 666 + ], + "type": "text", + "content": " (trained along with RTFM (I3D)) outperforms STG-NFAllData (see the table on the left). This validates the benefit of using our method to distinguish the normal from abnormal data in the training dataset. On ShanghaiTech, our proposed wOCC improves STG-NF from " + }, + { + "bbox": [ + 130, + 582, + 488, + 666 + ], + "type": "inline_equation", + "content": "80.29\\%" + }, + { + "bbox": [ + 130, + 582, + 488, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 582, + 488, + 666 + ], + "type": "inline_equation", + "content": "82.57\\%" + }, + { + "bbox": [ + 130, + 582, + 488, + 666 + ], + "type": "text", + "content": ". In contrast, the improvement is from " + }, + { + "bbox": [ + 130, + 582, + 488, + 666 + ], + "type": "inline_equation", + "content": "70.48\\%" + }, + { + "bbox": [ + 130, + 582, + 488, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 582, + 488, + 666 + ], + "type": "inline_equation", + "content": "74.76\\%" + }, + { + "bbox": [ + 130, + 582, + 488, + 666 + ], + "type": "text", + "content": " on UBnormal. The improvement on UBnormal is larger as there is more abnormal data in UBnormal which causes STG-NFAllData to perform worse." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "text", + "content": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 147, + 135, + 464, + 183 + ], + "blocks": [ + { + "bbox": [ + 211, + 114, + 401, + 126 + ], + "lines": [ + { + "bbox": [ + 211, + 114, + 401, + 126 + ], + "spans": [ + { + "bbox": [ + 211, + 114, + 401, + 126 + ], + "type": "text", + "content": "Table 4: Comparison with fixed thresholding." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 147, + 135, + 464, + 183 + ], + "lines": [ + { + "bbox": [ + 147, + 135, + 464, + 183 + ], + "spans": [ + { + "bbox": [ + 147, + 135, + 464, + 183 + ], + "type": "table", + "html": "
Weighted OCCWS with Adaptive ThresholdRTFM (AUC %)STG-NF (AUC %)
XX82.0680.52
✓X83.4881.78
X✓85.8681.94
✓✓88.1882.57
", + "image_path": "e32d1ffcb223522326a9df540ca20ced88ae56082bd13795e972d20db48f3de8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 481, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 481, + 277 + ], + "type": "text", + "content": "Second, please compare our method with GCL. Our " + }, + { + "bbox": [ + 130, + 205, + 481, + 277 + ], + "type": "inline_equation", + "content": "W_S" + }, + { + "bbox": [ + 130, + 205, + 481, + 277 + ], + "type": "text", + "content": " achieves an AUC of 88.18%, while that of GCL is 76.14% (this score is reported by the classifier of GCL, thus we use our WS model for the comparison). This significant improvement is partly due to the reason that we use a stronger OCC model, i.e., STG-NF, than GCL. If using the same OCC model, our method outperforms GCL by 2.76%, as shown in Table 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 277, + 481, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 481, + 361 + ], + "type": "text", + "content": "Our UVAD method can be degenerated into a supervised OCC method. For example, we can perform our interleaving training on the normal data only. The obtained wOCC model in this setting is denoted by " + }, + { + "bbox": [ + 130, + 277, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\mathrm{Our}_{wOCC}^{**}" + }, + { + "bbox": [ + 130, + 277, + 481, + 361 + ], + "type": "text", + "content": " which outperforms existing supervised OCC approaches (see the table in the middle). This validates the effectiveness of the proposed wOCC with weighted importance, as even among the normal data there is data that is more normal than other data, and the wOCC model can better identify their difference." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 361, + 481, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 481, + 434 + ], + "type": "text", + "content": "To degenerate our method into a supervised WS method, we allow our WS model to know video-level labels. Specifically, snippets of normal videos are treated as normal. Since snippets of abnormal videos can be either normal or abnormal, they are processed by our adaptive thresholding approach. As seen in Table 1 on the right, Our" + }, + { + "bbox": [ + 130, + 361, + 481, + 434 + ], + "type": "inline_equation", + "content": "^{**}_{WS}" + }, + { + "bbox": [ + 130, + 361, + 481, + 434 + ], + "type": "text", + "content": " (RTFM) outperforms the baseline RTFM [34]. On UBnormal, our method performs best among all the compared WS methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 453, + 321, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 321, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 321, + 464 + ], + "type": "text", + "content": "4.5 Ablation Study and Discussions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 473, + 480, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 480, + 497 + ], + "type": "text", + "content": "Next, we perform ablation studies on our method. Without extra explanations, experiments are conducted by taking STG-NF as wOCC and RTFM as WS." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 481, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 568 + ], + "type": "text", + "content": "Comparison between OCC over wOCC. In Figure 3, we compare the performance fluctuation of using wOCC or OCC (STG-NF) in our framework in the first two interleaving training modules' training. As shown, directly using OCC leads to performance fluctuation of both OCC and WS models and converges to an unsatisfactory point. In comparison, our proposed wOCC stabilizes the training and improves the performance of both models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 569, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 481, + 640 + ], + "type": "text", + "content": "Using Different OCC/WS Models. In Table 2, we conduct experiments of upgrading different types of OCC models to wOCC in our method. We test AE [47], Jigsaw [39], and STG-NF [13], while always using RTFM [34] as the WS model. As seen, the effectiveness of our method is essentially affected by wOCC model. Similarly in Table 3, the same wOCC model is used but with different WS models. Overall, a better wOCC or WS model yields better results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Comparison with Fixed Thresholding. In Table 4, the first row shows results of using fixed thresholding for both OCC and WS models. The second and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Nie et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 302, + 208 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 302, + 208 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 302, + 208 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 302, + 208 + ], + "type": "image", + "image_path": "e49694ac09a0b54bf3227bc03b60835fad9bbad271e57af2fb697a0a57382436.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 218, + 478, + 274 + ], + "lines": [ + { + "bbox": [ + 130, + 218, + 478, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 478, + 274 + ], + "type": "text", + "content": "Fig. 3: Performance fluctuation in the first Fig. 4: Left: ShanghaiTech. Right: UBnor and second modules. Red line: AUC of mal. Top: " + }, + { + "bbox": [ + 130, + 218, + 478, + 274 + ], + "type": "inline_equation", + "content": "T_{ws}" + }, + { + "bbox": [ + 130, + 218, + 478, + 274 + ], + "type": "text", + "content": " in each interleaving trainwOCC (or WS) when interleaving wOCC ing module. Bottom: The AUC of the WS and WS. Blue line: AUC of OCC (or WS) model in each module. Stars in top figures: when interleaving OCC and WS. position where achieves stopping criterion." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 391, + 205 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 391, + 205 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 391, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 391, + 205 + ], + "type": "image", + "image_path": "87c312a702868bd884db20a804a0401d1327f3b87656714014a95d732e2ce808.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 394, + 114, + 477, + 205 + ], + "blocks": [ + { + "bbox": [ + 394, + 114, + 477, + 205 + ], + "lines": [ + { + "bbox": [ + 394, + 114, + 477, + 205 + ], + "spans": [ + { + "bbox": [ + 394, + 114, + 477, + 205 + ], + "type": "image", + "image_path": "62b28728f4ba29b746f6044d05fe6177e72bd51e49af8b1fbee6991a62dbe5f9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 170, + 307, + 441, + 365 + ], + "blocks": [ + { + "bbox": [ + 220, + 285, + 392, + 297 + ], + "lines": [ + { + "bbox": [ + 220, + 285, + 392, + 297 + ], + "spans": [ + { + "bbox": [ + 220, + 285, + 392, + 297 + ], + "type": "text", + "content": "Table 5: Training time on ShanghaiTech." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 170, + 307, + 441, + 365 + ], + "lines": [ + { + "bbox": [ + 170, + 307, + 441, + 365 + ], + "spans": [ + { + "bbox": [ + 170, + 307, + 441, + 365 + ], + "type": "table", + "html": "
ModelEpoch Each StepTraining StepsTraining EpochsTraining TimeSTG-NF AUC %RTFM AUC %
Our UVAD1172×1×172h28m82.5788.18
2122×2×122h42m82.5387.95
3112×3×113h31m82.2786.36
STG-NF [13]--810m85.90-
RTFM [34]--502h32m-97.21
", + "image_path": "6efe5da69c35f42f6ec420a733397b6431cbf42fd893bb7dd14fb45d1841e923.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 388, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 480, + 422 + ], + "type": "text", + "content": "third rows use fixed thresholding for either OCC or WS models. The fourth row shows our method with wOCC and adaptive thresholding. As seen, our method outperforms all the other variants." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 425, + 481, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 481, + 509 + ], + "type": "text", + "content": "Effectiveness of the Stopping Criterion. In Figure 4, we show that with the current stopping criterion, our method can stop when it achieves the best AUC (i.e., the best VAD accuracy). As the repeating procedure runs, " + }, + { + "bbox": [ + 130, + 425, + 481, + 509 + ], + "type": "inline_equation", + "content": "T_{ws}" + }, + { + "bbox": [ + 130, + 425, + 481, + 509 + ], + "type": "text", + "content": " drops and AUC increases. " + }, + { + "bbox": [ + 130, + 425, + 481, + 509 + ], + "type": "inline_equation", + "content": "T_{ws}" + }, + { + "bbox": [ + 130, + 425, + 481, + 509 + ], + "type": "text", + "content": " drops very fast at the very beginning, and then the drop rate becomes slower. Simultaneously, AUC first increases and then decreases. We stop our method when the change rate of " + }, + { + "bbox": [ + 130, + 425, + 481, + 509 + ], + "type": "inline_equation", + "content": "T_{ws}" + }, + { + "bbox": [ + 130, + 425, + 481, + 509 + ], + "type": "text", + "content": " is about " + }, + { + "bbox": [ + 130, + 425, + 481, + 509 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 425, + 481, + 509 + ], + "type": "text", + "content": " of its original change rate, which finds the best stopping positions indicated by the stars in top figures." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 509, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 592 + ], + "type": "text", + "content": "Robustness to " + }, + { + "bbox": [ + 130, + 509, + 482, + 592 + ], + "type": "inline_equation", + "content": "R\\%" + }, + { + "bbox": [ + 130, + 509, + 482, + 592 + ], + "type": "text", + "content": ". Our method is robust to the parameter " + }, + { + "bbox": [ + 130, + 509, + 482, + 592 + ], + "type": "inline_equation", + "content": "R\\%" + }, + { + "bbox": [ + 130, + 509, + 482, + 592 + ], + "type": "text", + "content": ", as demonstrated in Figure 5. On the top, we show that the threshold " + }, + { + "bbox": [ + 130, + 509, + 482, + 592 + ], + "type": "inline_equation", + "content": "T_{ws}" + }, + { + "bbox": [ + 130, + 509, + 482, + 592 + ], + "type": "text", + "content": " converges to similar values after 6 interleaving training modules when using different " + }, + { + "bbox": [ + 130, + 509, + 482, + 592 + ], + "type": "inline_equation", + "content": "R\\%" + }, + { + "bbox": [ + 130, + 509, + 482, + 592 + ], + "type": "text", + "content": ". At the bottom, we show that different " + }, + { + "bbox": [ + 130, + 509, + 482, + 592 + ], + "type": "inline_equation", + "content": "R\\%" + }, + { + "bbox": [ + 130, + 509, + 482, + 592 + ], + "type": "text", + "content": " yield similar AUC on both training datasets. The experiments demonstrate that our method is not influenced much by this parameter for each dataset. For different datasets, we need to set different " + }, + { + "bbox": [ + 130, + 509, + 482, + 592 + ], + "type": "inline_equation", + "content": "R\\%" + }, + { + "bbox": [ + 130, + 509, + 482, + 592 + ], + "type": "text", + "content": " that roughly match with the ratio of anomaly data in the datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Training Loss Curve. In Figure 6, we show training loss curves of the wOCC (STG-NF) model in our framework. As seen, the loss drops smoothly within each interleaving training module. Since we re-train the wOCC model from scratch in the next module, the loss suddenly increases at the beginning of the next module, but the peak magnitude of the loss is lower than that of the previous module. Please check the zoomed-in curve showing the loss curve of the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "text", + "content": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 471, + 232 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 471, + 232 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 471, + 232 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 471, + 232 + ], + "type": "image", + "image_path": "894b7525f4f435893aebefee2473d7936a0170ca0ab8d3170d020cf0945d1248.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 238, + 472, + 283 + ], + "lines": [ + { + "bbox": [ + 130, + 238, + 472, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 472, + 283 + ], + "type": "text", + "content": "Fig.5: Top: different " + }, + { + "bbox": [ + 130, + 238, + 472, + 283 + ], + "type": "inline_equation", + "content": "R\\%" + }, + { + "bbox": [ + 130, + 238, + 472, + 283 + ], + "type": "text", + "content": " converge to Fig.6: Bottom: training loss curve of the similar " + }, + { + "bbox": [ + 130, + 238, + 472, + 283 + ], + "type": "inline_equation", + "content": "T_{ws}" + }, + { + "bbox": [ + 130, + 238, + 472, + 283 + ], + "type": "text", + "content": " (on ShanghaiTech). Bottom: wOCC model during the whole repeating different " + }, + { + "bbox": [ + 130, + 238, + 472, + 283 + ], + "type": "inline_equation", + "content": "R\\%" + }, + { + "bbox": [ + 130, + 238, + 472, + 283 + ], + "type": "text", + "content": " yield similar AUC (Left: on procedure. Top: zoomed-in training loss ShanghaiTech, Right: on UBnormal). curve in the third module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 309, + 480, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 480, + 346 + ], + "type": "text", + "content": "wOCC model in the third module. Even though the model is trained alternately with another model, its loss can still decrease smoothly, not affected by the other model. The WS model's training loss curve is put into the supplemental material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 346, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 479 + ], + "type": "text", + "content": "Training Time. As mentioned in Section 3.3, although we need to train the wOCC and WS models alternately many times in multiple modules, our method is fast in training. That is because, we just train a model for one epoch and then exchange to train the other model for another epoch, which form a training step of our method. Our interleaving training converges very fast. As shown in Table 5, there are only 17 training steps during the whole training, yielding a total of " + }, + { + "bbox": [ + 130, + 346, + 482, + 479 + ], + "type": "inline_equation", + "content": "34 = 2(\\mathrm{wOCC}" + }, + { + "bbox": [ + 130, + 346, + 482, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 346, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\mathrm{WS})\\times 1(\\mathrm{epoch})\\times 17" + }, + { + "bbox": [ + 130, + 346, + 482, + 479 + ], + "type": "text", + "content": " (all training steps) epochs. It takes around 2.5 hours to train our method on ShanghaiTech. We have also trained the wOCC or WS model for 2 or 3 epochs in each training step, but obtained worse results, probably because training with more epochs would enhance the supervision of wrong pseudo labels in early training modules." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 502, + 220, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 220, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 220, + 514 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": "In this paper, we interleave the One-Classification and Weakly-Supervised models with adaptive thresholding to tackle the unsupervised video anomaly detection problem without human annotation. We face the challenges of performance fluctuation and the demanding of accurate threshold when designing the framework. To alleviate performance fluctuation, We propose the wOCC model which requires soft labels but not binary labels. To obtain a relative optimal threshold, we repeat our proposed interleaving training module, during which we propose a monotonically threshold decreasing mechanism to refine a rough threshold progressively. Extensive experiments demonstrate the effectiveness of our method. Remarkably, our method can be upgraded with the most recent development in OCC and WS VAD fields." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Nie et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 481, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 201 + ], + "type": "text", + "content": "This work was partly supported by the Natural Science Foundation of China (62072191, U21A20520, 62325204), the National Key Research and Development Program of China (2022YFE0112200), the Key-Area Research and Development Program of Guangzhou City (202206030009), and the Guangdong Basic and Applied Basic Research Fund (2023A1515030002, 2024A1515011995)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 246, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 246, + 480, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 290 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 290 + ], + "type": "text", + "content": "1. Acsintoae, A., Florescu, A., Georgescu, M., Mare, T., Sumedrea, P., Ionescu, R.T., Khan, F.S., Shah, M.: Unbnormal: New benchmark for supervised open-set video anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 292, + 480, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 292, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 292, + 480, + 334 + ], + "type": "text", + "content": "2. Al-lahham, A., Tastan, N., Zaheer, M.Z., Nandakumar, K.: A coarse-to-fine pseudolabeling (c2fpl) framework for unsupervised video anomaly detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6793-6802 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 335, + 480, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 480, + 367 + ], + "type": "text", + "content": "3. Andrews, S., Tsochantaridis, I., Hofmann, T.: Support vector machines for multiple-instance learning. Advances in neural information processing systems 15 (2002)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 368, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 368, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 368, + 480, + 411 + ], + "type": "text", + "content": "4. Barbalau, A., Ionescu, R.T., Georgescu, M.I., Dueholm, J., Ramachandra, B., Nasrollahi, K., Khan, F.S., Moeslund, T.B., Shah, M.: Ssmtl++: Revisiting self-supervised multi-task learning for video anomaly detection. Computer Vision and Image Understanding 229, 103656 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 412, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 480, + 445 + ], + "type": "text", + "content": "5. Basharat, A., Gritai, A., Shah, M.: Learning object motion patterns for anomaly detection and improved object detection. In: 2008 IEEE conference on computer vision and pattern recognition. pp. 1-8. IEEE (2008)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 445, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 480, + 478 + ], + "type": "text", + "content": "6. Burlina, P., Joshi, N., Wang, I., et al.: Where's wally now? deep generative and discriminative embeddings for novelty detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11507-11516 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 501 + ], + "type": "text", + "content": "7. Dietterich, T.G., Lathrop, R.H., Lozano-Pérez, T.: Solving the multiple instance problem with axis-parallel rectangles. Artificial intelligence 89(1-2), 31-71 (1997)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 533 + ], + "type": "text", + "content": "8. Feng, J.C., Hong, F.T., Zheng, W.S.: Mist: Multiple instance self-training framework for video anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14009-14018 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 533, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 577 + ], + "type": "text", + "content": "9. Feng, X., Song, D., Chen, Y., Chen, Z., Ni, J., Chen, H.: Convolutional transformer based dual discriminator generative adversarial networks for video anomaly detection. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 5546-5554 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 578, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 578, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 578, + 480, + 621 + ], + "type": "text", + "content": "10. Georgescu, M.I., Barbalau, A., Ionescu, R.T., Khan, F.S., Popescu, M., Shah, M.: Anomaly detection in video via self-supervised and multi-task learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12742-12752 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "type": "text", + "content": "11. Georgescu, M.I., Ionescu, R.T., Khan, F.S., Popescu, M., Shah, M.: A background-agnostic framework with adversarial training for abnormal event detection in video. IEEE transactions on pattern analysis and machine intelligence 44(9), 4505-4523 (2021)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "text", + "content": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "12. Gong, D., Liu, L., Le, V., Saha, B., Mansour, M.R., Venkatesh, S., Hengel, A.v.d.: Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1705-1714 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "type": "text", + "content": "13. Hirschorn, O., Avidan, S.: Normalizing flows for human pose anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13545-13554 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 196, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 481, + 239 + ], + "type": "text", + "content": "14. Ionescu, R.T., Khan, F.S., Georgescu, M.I., Shao, L.: Object-centric auto-encoders and dummy anomalies for abnormal event detection in video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7842-7851 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 239, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 481, + 274 + ], + "type": "text", + "content": "15. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 274, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 296 + ], + "type": "text", + "content": "16. Kingma, D.P., Dhariwal, P.: Glow: Generative flow with invertible 1x1 convolutions. Advances in neural information processing systems 31 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "text", + "content": "17. Li, S., Liu, F., Jiao, L.: Self-training multi-sequence learning with transformer for weakly supervised video anomaly detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1395-1403 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "type": "text", + "content": "18. Liu, W., Luo, W., Lian, D., Gao, S.: Future frame prediction for anomaly detection-- a new baseline. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6536-6545 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 363, + 481, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 481, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 481, + 408 + ], + "type": "text", + "content": "19. Liu, Z., Nie, Y., Long, C., Zhang, Q., Li, G.: A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13588-13597 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 409, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 481, + 441 + ], + "type": "text", + "content": "20. Lorre, G., Rabarisoa, J., Orcesi, A., Ainouz, S., Canu, S.: Temporal contrastive pretraining for video action recognition. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 662-670 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 442, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 485 + ], + "type": "text", + "content": "21. Lv, H., Yue, Z., Sun, Q., Luo, B., Cui, Z., Zhang, H.: Unbiased multiple instance learning for weakly supervised video anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8022-8031 (June 2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 486, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 481, + 520 + ], + "type": "text", + "content": "22. Markovitz, A., Sharir, G., Friedman, I., Zelnik-Manor, L., Avidan, S.: Graph embedded pose clustering for anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10539-10547 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "text", + "content": "23. Medioni, G., Cohen, I., Brémond, F., Hongeng, S., Nevatia, R.: Event detection and analysis from video streams. IEEE Transactions on pattern analysis and machine intelligence 23(8), 873-889 (2001)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 554, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 597 + ], + "type": "text", + "content": "24. Morais, R., Le, V., Tran, T., Saha, B., Mansour, M., Venkatesh, S.: Learning regularity in skeleton trajectories for anomaly detection in videos. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11996-12004 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "text", + "content": "25. Nguyen, T.N., Meunier, J.: Anomaly detection in video sequence with appearance-motion correspondence. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1273-1283 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "26. Pang, G., Yan, C., Shen, C., Hengel, A.v.d., Bai, X.: Self-trained deep ordinal regression for end-to-end video anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12173-12182 (2020)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Nie et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 149 + ], + "type": "text", + "content": "27. Park, H., Noh, J., Ham, B.: Learning memory-guided normality for anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14372-14381 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "text", + "content": "28. Sapkota, H., Yu, Q.: Bayesian nonparametric submodular video partition for robust anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3212-3221 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "type": "text", + "content": "29. Shi, C., Sun, C., Wu, Y., Jia, Y.: Video anomaly detection via sequentially learning multiple pretext tasks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10330-10340 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 257 + ], + "type": "text", + "content": "30. Smeureanu, S., Ionescu, R.T., Popescu, M., Alexe, B.: Deep appearance features for abnormal behavior detection in video. In: Image Analysis and Processing-ICIAP 2017: 19th International Conference, Catania, Italy, September 11-15, 2017, Proceedings, Part II 19. pp. 779-789. Springer (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "text", + "content": "31. Sohrab, F., Raitoharju, J., Gabbouj, M., Iosifidis, A.: Subspace support vector data description. In: 2018 24th International Conference on Pattern Recognition (ICPR). pp. 722-727. IEEE (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 289, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 322 + ], + "type": "text", + "content": "32. Sultani, W., Chen, C., Shah, M.: Real-world anomaly detection in surveillance videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6479-6488 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 322, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 480, + 354 + ], + "type": "text", + "content": "33. Sun, S., Gong, X.: Hierarchical semantic contrast for scene-aware video anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22846-22856 (June 2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 354, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 397 + ], + "type": "text", + "content": "34. Tian, Y., Pang, G., Chen, Y., Singh, R., Verjans, J.W., Carneiro, G.: Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4975-4986 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 429 + ], + "type": "text", + "content": "35. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 429, + 480, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 461 + ], + "type": "text", + "content": "36. Tudor Ionescu, R., Smeureanu, S., Alexe, B., Popescu, M.: Unmasking the abnormal events in video. In: Proceedings of the IEEE international conference on computer vision. pp. 2895-2903 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 461, + 480, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 494 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 494 + ], + "type": "text", + "content": "37. Tur, A.O., Dall'Asen, N., Bryan, C., Ricci, E.: Exploring diffusion models for unsupervised video anomaly detection. In: 2023 IEEE International Conference on Image Processing (ICIP). pp. 2540-2544. IEEE (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 494, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 480, + 525 + ], + "type": "text", + "content": "38. Wan, B., Fang, Y., Xia, X., Mei, J.: Weakly supervised video anomaly detection via center-guided discriminative learning. In: 2020 IEEE international conference on multimedia and expo (ICME). pp. 1-6. IEEE (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "text", + "content": "39. Wang, G., Wang, Y., Qin, J., Zhang, D., Bao, X., Huang, D.: Video anomaly detection by solving decoupled spatio-temporal jigsaw puzzles. In: European Conference on Computer Vision. pp. 494-511. Springer (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 557, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 601 + ], + "type": "text", + "content": "40. Wang, J., Song, Y., Leung, T., Rosenberg, C., Wang, J., Philbin, J., Chen, B., Wu, Y.: Learning fine-grained image similarity with deep ranking. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1386-1393 (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "text", + "content": "41. Wang, J., Cherian, A.: Gods: Generalized one-class discriminative subspaces for anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8201-8211 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "42. Wang, Z., Zou, Y., Zhang, Z.: Cluster attention contrast for video anomaly detection. In: Proceedings of the 28th ACM international conference on multimedia. pp. 2463-2471 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "text", + "content": "Interleaving OCC and WS Models with Adaptive Thresholding for UVAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 544 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "43. Wu, J.C., Hsieh, H.Y., Chen, D.J., Fuh, C.S., Liu, T.L.: Self-supervised sparse representation for video anomaly detection. In: European Conference on Computer Vision. pp. 729-745. Springer (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 194 + ], + "type": "text", + "content": "44. Wu, P., Liu, J., Shi, Y., Sun, Y., Shao, F., Wu, Z., Yang, Z.: Not only look, but also listen: Learning multimodal violence detection under weak supervision. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16. pp. 322-339. Springer (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "45. Yan, C., Zhang, S., Liu, Y., Pang, G., Wang, W.: Feature prediction diffusion model for video anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5527-5537 (October 2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "text", + "content": "46. Yang, Z., Liu, J., Wu, Z., Wu, P., Liu, X.: Video event restoration based on keyframes for video anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14592-14601 (June 2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 480, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 313 + ], + "type": "text", + "content": "47. Zaheer, M.Z., Mahmood, A., Khan, M.H., Segu, M., Yu, F., Lee, S.I.: Generative cooperative learning for unsupervised video anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14744-14754 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 315, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 358 + ], + "type": "text", + "content": "48. Zaheer, M.Z., Mahmood, A., Astrid, M., Lee, S.I.: Claws: Clustering assisted weakly supervised learning with normalcy suppression for anomalous event detection. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 358-376. Springer (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "type": "text", + "content": "49. Zhang, C., Li, G., Qi, Y., Wang, S., Qing, L., Huang, Q., Yang, M.H.: Exploiting completeness and uncertainty of pseudo labels for weakly supervised video anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16271-16280 (June 2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "50. Zhang, J., Qing, L., Miao, J.: Temporal convolutional network with complementary inner bag loss for weakly supervised anomaly detection. In: 2019 IEEE International Conference on Image Processing (ICIP). pp. 4030-4034. IEEE (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "51. Zhang, T., Lu, H., Li, S.Z.: Learning semantic scene models by object classification and trajectory clustering. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1940-1947. IEEE (2009)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "52. Zhao, Y., Deng, B., Shen, C., Liu, Y., Lu, H., Hua, X.S.: Spatio-temporal autoencoder for video anomaly detection. In: Proceedings of the 25th ACM international conference on Multimedia. pp. 1933-1941 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 501, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 544 + ], + "type": "text", + "content": "53. Zhong, J.X., Li, N., Kong, W., Liu, S., Li, T.H., Li, G.: Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1237-1246 (2019)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Nie et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/InternVideo2_ Scaling Foundation Models for Multimodal Video Understanding/3f170c5b-f8dd-444c-a833-330298501085_content_list.json b/2024/InternVideo2_ Scaling Foundation Models for Multimodal Video Understanding/3f170c5b-f8dd-444c-a833-330298501085_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6f8dae5a1ff12a2ca0a366c6d9973e4243a80747 --- /dev/null +++ b/2024/InternVideo2_ Scaling Foundation Models for Multimodal Video Understanding/3f170c5b-f8dd-444c-a833-330298501085_content_list.json @@ -0,0 +1,1922 @@ +[ + { + "type": "text", + "text": "InternVideo2: Scaling Foundation Models for Multimodal Video Understanding", + "text_level": 1, + "bbox": [ + 245, + 140, + 756, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yi Wang $^{1}$ , Kunchang Li $^{3,1}$ , Xinhao Li $^{2,1}$ , Jiashuo Yu $^{3}$ , Yinan He $^{1}$ , Guo Chen $^{2,1}$ , Baoqi Pei $^{1}$ , Rongkun Zheng $^{1}$ , Zun Wang $^{1}$ , Yansong Shi $^{1}$ , Tianxiang Jiang $^{1}$ , Songze Li $^{1}$ , Jilan Xu $^{1}$ , Hongjie Zhang $^{1}$ , Yifei Huang $^{1}$ , Yu Qiao $^{\\dagger}$ , Yali Wang $^{\\dagger 3,1}$ , and Limin Wang $^{\\dagger 2,1}$", + "bbox": [ + 240, + 210, + 761, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 OpenGVLab, Shanghai AI Laboratory, Shanghai, China \n2 Nanjing University, Nanjing, China", + "bbox": [ + 305, + 284, + 696, + 313 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Shenzhen Institutes of Advanced Technology, CAS, Shenzhen, China {wangyi,qiaoyu,wangyali,wanglimin}@pjlab.org.cn https://github.com/OpenGVLab/InternVideo/tree/main/InternVideo?", + "bbox": [ + 263, + 313, + 738, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e510ad049ec3410b35cba384edaeefda48dbe3849da62de1fbc1169937efa682.jpg", + "image_caption": [ + "Fig. 1: InternVideo2 yields strong transferable visual and visual-linguistic representations for action recognition, video-text understanding, to video-centric dialogue. It also exhibits capability of long-form video understanding and procedure-aware reasoning." + ], + "image_footnote": [], + "bbox": [ + 217, + 392, + 777, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We introduce InternVideo2, a new family of video foundation models (ViFM) that achieve the state-of-the-art results in video recognition, video-text tasks, and video-centric dialogue. Our core design is a progressive training approach that unifies the masked video modeling, crossmodal contrastive learning, and next token prediction, scaling up the video encoder size to 6B parameters. At the data level, we prioritize spatiotemporal consistency by semantically segmenting videos and generating video-audio-speech captions. This improves the alignment between video and text. Through extensive experiments, we validate our designs and demonstrate superior performance on over 60 video and audio tasks. Notably, our model outperforms others on various video-related dialogue and long video understanding benchmarks, highlighting its ability to reason and comprehend longer contexts.", + "bbox": [ + 259, + 611, + 743, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Video Foundation Model $\\cdot$ Progressive Training $\\cdot$ MLLM.", + "bbox": [ + 259, + 805, + 723, + 819 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. †Corresponding authors.", + "bbox": [ + 230, + 825, + 540, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 375, + 161 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning transferrable spatiotemporal representations is a critical research area in computer vision, holding diverse applications across domains such as video search [31], game control [7], robotic learning [26], self-driving [103], and scientific studies [84]. Recently, the advancement of Large Language Models (LLMs) [6, 67, 86, 87] and their multimodal variations (MLLMs) [33, 56, 68, 84] have had a profound impact on vision research and other disciplines. Embedding videos effectively into these large models and harnessing their capabilities to enhance video understanding performance has emerged as pivotal tasks [48, 61].", + "bbox": [ + 212, + 175, + 787, + 296 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous research has identified several effective learning schemes for video representations, including reconstructing videos with masked inputs [28,37,85,90], aligning videos with languages [51,52,96,98], and predicting the next token using videos [1,49,83]. These approaches have turned out to be complementary and can be unified through a progressive training scheme. Notably, methods such as InternVideo [94], UMT [51], and VideoPrism [109] have utilized a two-stage training approach involving masked reconstruction and multimodal contrastive learning, leading to improved performance in downstream tasks. Following this line, we aim to further extend this progressive training scheme by incorporating video-based next token prediction and scaling the entire training process, including models and data, to build a new family of video foundation models.", + "bbox": [ + 212, + 296, + 787, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proposed video foundation model, coined as InternVideo2, is built through a progressive training scheme. The learning involves three stages: (1) capturing spatiotemporal structure via unmasked reconstruction, (2) aligning with semantics from other modalities, and (3) enhancing its open-ended dialogue power through next token prediction. In the initial stage, the model learns to reconstruct the unmasked video tokens, allowing the video encoder to develop basic spatiotemporal perception capability. To estimate the existing tokens, vision encoders (InternViT [17] and VideoMAE-g [90]) trained differently are employed as proxies. In the next stage of crossmodal learning, the architecture is expanded to include audio and text encoders. This not only improves the alignment between videos and text but also endows InternVideo2 the ability to handle video-audio tasks. By incorporating these additional modalities, the model's understanding of videos is enriched and aligned with their semantics. Finally, in the next-token prediction stage, a video-centric dialogue system and the corresponding instruction-finetuning dataset are built to further tune the InternVideo2. By connecting InternVideo2 to LLMs, the video encoder is further updated through next-token prediction training, enhancing its ability for open-ended tasks such as VQA and video description.", + "bbox": [ + 212, + 462, + 787, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For the training of InternVideo2, we emphasize the spatiotemporal consistency and labeling quality in the data. We build a large-scale multimodal video-centric dataset consisting of 402M data entries, which includes 2M videos, 50M video-text pairs (from WebVid [3] and InternVid [93]), 50M video-audio-speech-text pairs (InternVid2), and 300M image-text pairs. Specifically, for InternVid2, we segment videos into clips semantically and focus on recalibrating the clip descriptions using three modalities: audio, video, and speech. We first generate captions for", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "these three modalities separately. Then individual captions are fused together to create a more comprehensive description, which will improve the model's ability to comprehend and interpret the video accurately.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We evaluate InternVideo2 across a wide range of video-related tasks. These tasks span from basic spatiotemporal perception, such as action recognition, to high-level reasoning tasks, such as long video or procedure-aware question-answering (QA), as given in Fig. 1. The results (in Sec. 5) demonstrate that InternVideo2 achieves the state-of-the-art performance on multiple tasks and is able to analyze and reason over sequences of actions. This top performance signifies its capability to effectively capture and understand video content. These empirical findings validate that InternVideo2 could serves as a general video encoder for future exploration in video understanding. In summary, our contributions to video understanding are as follows.", + "bbox": [ + 212, + 191, + 787, + 343 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- This paper introduces InternVideo2, a competitive family of video foundation models that leverages masked reconstruction, crossmodal contrastive learning, and next token prediction to make model perceptive, semantic, and capable of reasoning in video understanding.", + "- InternVideo2 achieves the state-of-the-art performance for more than 60 video / audio tasks. Our model demonstrates superior performance in video-related dialogue and long video understanding, highlighting its potential in modeling high-level world knowledge.", + "- We provide an enhanced dataset to train InternVideo2. This includes the validation and incorporation of audio data during training, as well as the improved captioning method. These improvements result in significant enhancements in model performance and generalization ability." + ], + "bbox": [ + 215, + 349, + 784, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 551, + 387, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Foundation Models. Studies on learning video foundation models becomes increasingly crucial considering its wide applications [?, 51, 52, 85, 89, 90, 92, 94, 96, 98, 109]. Typical methods in building video foundation models (ViFM) include video-text contrastive learning [52, 94, 96], masked video modeling [30, 85, 90, 94], and next token prediction [1, 80, 83]. Specifically, All-in-one [89] utilized a single backbone with unified multiple pretraining objectives. On the other hand, UMT [51] combined masked modeling with video-text contrastive learning, demonstrating strong performance in both action recognition and videolanguage tasks. Another approach is mPLUG-2 [95], which introduced a new design for modulating different modalities. It shared a common module across modalities to enhance relations while incorporating modality-specific modules for discrimination. In addition to video-text pretraining, researchers have also explored the use of audio information in videos to improve performance. MERLOT Reserve [104] learned video representations using a large-scale dataset of videospeech-transcript pairs. VALOR [14] employed independent encoders for video, audio, and text and trains a joint visual-audio-text representation. VAST [15] constructed an audio-visual-speech dataset and develops a multimodal backbone", + "bbox": [ + 212, + 582, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "InternVideo2", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/97027eeccf7e96697083cc75455f70ec1857ca205740c8b248bd93fb09a5b82e.jpg", + "image_caption": [ + "Fig. 2: Framework of InternVideo2. It consists of three consecutive training phases: unmasked video token reconstruction, multimodal contrastive learning, and next token prediction. In stage 1, the video encoder is trained from scratch, while in stages 2 and 3, it is initialized from the version used in the previous stage." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 782, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "that excels in video-audio-related tasks. VideoPrism [109] combined video-text contrastive learning and video token reconstruction on a combination of public and proprietary videos, achieving leading results across various video tasks.", + "bbox": [ + 212, + 324, + 782, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multimodal Large Language Models. With advances in large language models (LLMs) [6, 24, 76], their multimodal versions (MLLMs) is becoming popular as it can handle open-world tasks. Seminal works like Flamingo [1] showed outstanding zero/few-shots performances over a range of multimodal tasks [36, 63, 72, 97]. Public MLLMs [33, 56, 113] such as LLaVA [56] and InstructBLIP [21] proposed to use visual instruction-tuning data to improve the visual dialogue ability. Some video-centric MLLMs have been proposed, such as VideoChat [48], VideoChatGPT [61] and Valley [60], by using instruction data to connect video encoders to LLMs for open-world video understanding.", + "bbox": [ + 212, + 369, + 787, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 530, + 330, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We learn InternVideo2 in three stages, illustrated in Fig. 2. The stages include spatiotemporal token reconstruction, video-audio-speech-language contrastive learning, and connecting to a large language model (LLM) for joint training.", + "bbox": [ + 212, + 561, + 785, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video Encoder. The video encoder used in InternVideo2 follows the Vision Transformer (ViT) [25] and includes additional projection layers for distillation. Inspired by previous works [17, 102], we introduce attention pooling to the ViT. For input videos, we sparsely sample 8 frames [91] and perform a $14 \\times 14$ ( $h \\times w$ ) spatial downsampling. These spatiotemporal tokens are then concatenated with a class token and combined with 3D position embeddings. The hyperparameters used for ViT-1B and -6B are the same as CoCa-1B [102] and InternViT-6B [17], respectively. The details of ViT-6B architecture are given in Supp.", + "bbox": [ + 212, + 625, + 787, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Stage1: Reconstructing Unmasked Video Tokens", + "text_level": 1, + "bbox": [ + 212, + 768, + 661, + 784 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We exploit two expert models to guide the video encoder to conduct token-level reconstruction for unmasked areas. Specifically, we adopt InternVL-6B [17] and VideoMAEv2-g [90] to transfer unmasked knowledge via simple projection layers.", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When training, we input the full videos into different teachers and mask out $80\\%$ of the tokens frame by frame, under the semantic guidance of the multimodal model InternVL and motion-aware model VideoMAEv2. We only align the unmasked tokens, by minimizing their mean squared error (MSE) between student and teachers. The learning objective is to reconstruct the remaining tokens as:", + "bbox": [ + 214, + 146, + 787, + 223 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\frac {1}{Z} \\sum \\left(\\alpha_ {1} | f ^ {V} (\\mathbf {V} _ {p}) - h (\\mathbf {V} _ {p}) | ^ {2} + \\alpha_ {2} | f ^ {V} (\\mathbf {V} _ {p}) - g (\\mathbf {V} _ {p}) | ^ {2}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 290, + 234, + 787, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f^V$ , $h$ , and $g$ are our video encoder, InternViT-6B [17], and ViT-g of VideoMAEv2, respectively. $p$ stands for the token index and $f(\\mathbf{V}_p)$ is the corresponding token extracted by InternVideo2 for input video $\\mathbf{V}$ . $Z$ is the normalization factor. $\\alpha_{1}$ and $\\alpha_{2}$ balance the influence between the employed models.", + "bbox": [ + 214, + 281, + 787, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In implementation, we align the last 6 layers of InternVL, last 4 layers of VideoMAEv2, and the final output token of InternVL. Different losses are simply summed for optimization. After pretraining, we drop those projection layers and only use the basic encoder. Compared with only using the multimodal model in UMT and VideoPrism, our strategy makes the vision encoder multimodal-friendly as well as enhances its temporal sensitivity for action modeling.", + "bbox": [ + 214, + 343, + 787, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Stage 2: Aligning Video to Audio-Speech-Text", + "text_level": 1, + "bbox": [ + 215, + 459, + 643, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We exploit the correspondence between video and audio, speech, and text to encourage InternVideo2 to learn more semantics. In practice, InternVideo2 has a huge video encoder, and its employed audio and text encoders are relatively lightweight. The used audio encoder is a 12-layer transformer initialized with BEATs [13] (90M). It takes in audio features, which are 64-dimensional log Mel filterbank spectrograms using a 25ms Hamming window, transformed from 10-second-long clips, padding with zeros. For the text and speech encoders, we initialize the text encoder and multimodal decoder using BERT-Large [24]. Specifically, we utilize the initial 19 layers of BERT-Large as the text encoder, with the subsequent 5 layers equipped with cross-attention layers serving as the multimodal decoder.", + "bbox": [ + 214, + 487, + 787, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For pretraining objectives, we establish alignment across different modalities via text, including video, audio, image, and speech. We employ crossmodal contrastive and matching losses with masked language reconstruction loss as:", + "bbox": [ + 214, + 654, + 787, + 699 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {C O N}} + \\mathcal {L} _ {\\mathrm {M A C}} + \\mathcal {L} _ {\\mathrm {M L M}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 714, + 785, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The employed $\\mathcal{L}_{\\mathrm{MAC}}$ and $\\mathcal{L}_{\\mathrm{MLM}}$ are standard loss from [19]. Specifically, the crossmodal contrastive learning is given as:", + "bbox": [ + 214, + 741, + 787, + 772 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C O N}} = \\sum_ {M, T _ {M ^ {\\prime}}} \\mathcal {L} _ {\\mathrm {C O N}} (M, T _ {M ^ {\\prime}}) = - \\sum_ {M, T _ {M ^ {\\prime}}} \\left(\\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (\\sin \\left(f _ {i} ^ {M} , f _ {i} ^ {T _ {M ^ {\\prime}}}\\right) / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (\\sin \\left(f _ {i} ^ {M} , f _ {j} ^ {T _ {M ^ {\\prime}}}\\right) / \\tau)} + \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (\\sin \\left(f _ {i} ^ {T _ {M ^ {\\prime}}} , f _ {i} ^ {M}\\right) / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (\\sin \\left(f _ {i} ^ {T _ {M ^ {\\prime}}} , f _ {j} ^ {M}\\right) / \\tau)}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 787, + 782, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f^V$ and $f^T$ denote the learned video and text embeddings, respectively.", + "bbox": [ + 220, + 824, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "InternVideo2", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$M$ and $T_{M^{\\prime}}$ indicates the modality of input signals and the text descriptions describing, respectively. $\\sin (\\cdot)$ computes the cosine similarity between two features. $\\tau$ is the learnable temperature.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the matching part, it is given as:", + "bbox": [ + 238, + 191, + 508, + 207 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M A C}} = - y \\log f _ {p} (\\mathbf {V}, \\mathbf {T}) - (1 - y) \\log \\left(1 - f _ {p} (\\mathbf {V}, \\mathbf {T})\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 215, + 785, + 232 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $f_{p}(\\mathbf{V},\\mathbf{T})$ computes the matching likelihood between $\\mathbf{V}$ and $\\mathbf{T}$ . $y$ denotes whether the given video and text are paired $(y = 1)$ or not $(y = 0)$ .", + "bbox": [ + 212, + 238, + 785, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The employed masked language modeling loss is:", + "bbox": [ + 238, + 270, + 593, + 284 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M L M}} = - \\log f _ {p} ^ {T} \\left(\\mathbf {T} _ {j} \\mid \\mathbf {T} _ {< j}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 292, + 785, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $f_{p}^{T}(\\mathbf{T}_{j}|\\mathbf{T}_{< j})$ computes the likelihood of the $j_{\\mathrm{th}}$ text token based on the previous ones. Here $\\mathbf{T}$ refers to video captions.", + "bbox": [ + 212, + 316, + 785, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To improve the training efficiency, we employ the masked learning strategy, aligning unmasked video tokens to tokens from other modalities first, then using full video tokens reconstruction shortly. Specifically, it consists of two steps as follows:", + "bbox": [ + 212, + 349, + 785, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Aligning Masked Visual-Language-Audio. We freeze the audio encoder and focus on aligning visual, audio, and text features. For pre-training, we use a comprehensive set of image, video, and audio-video data. The combinations of modalities used are represented as $\\{M,T_{M^{\\prime}}\\} \\in \\{\\{I,T_I\\},\\{V,T_V\\},\\{V,T_{VAS}\\},\\{VA,T_{VAS}\\}\\}$ where each pair denotes the concatenated features from the respective modalities.", + "bbox": [ + 212, + 420, + 787, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unmasked Visual-Audio-Language Post-Pretraining. We freeze the vision encoder to jointly align audio, visual, and text features. Post-pretraining is conducted using a smaller subset of image and video data (25M samples), along with the full set of audio (0.5M samples) and audio-video data (50M samples). Since the parameters of the largest ViT-6B model are frozen, we do not use masking strategies in this phase to ensure consistency with the inference process and to minimize any performance degradation in downstream tasks. The modality combinations used here are $\\{M, T_{M'}\\} \\in \\{\\{I, T_I\\}, \\{V, T_V\\}, \\{A, T_A\\}, \\{V, T_{VAS}\\}, \\{VA, T_{VA}\\}\\}$ .", + "bbox": [ + 212, + 508, + 787, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Stage3: Predicting Next Token with Video-Centric Inputs", + "text_level": 1, + "bbox": [ + 212, + 648, + 740, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further enrich the semantics embedded in InternVideo2 and improve its support for video-centric dialogue, we tune it by connecting it to a LLM with QFormer design [46, 47]. We employ the progressive learning scheme in [49] by using InternVideo2 as the video encoder and train a video blip for communicating with open-sourced LLM [40, 111]. In training, the video encoder will be updated. Detailed training receipts are the same with VideoChat2 [49].", + "bbox": [ + 212, + 670, + 787, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Multimodal Video Data", + "text_level": 1, + "bbox": [ + 214, + 781, + 485, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We list our training data in Tab. 1. Among the datasets used, KMash and InternVid2 are newly built, whereas the rest are public available.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b16688687ffe4a779a953f1a9b3df0795a5b6a4162d79a50437a6ff15ac849ce.jpg", + "table_caption": [ + "Table 1: Summary of datasets used in InternVideo2 pretraining process." + ], + "table_footnote": [], + "table_body": "
Pretraining Stage DatasetDomain# of clipsAnnotation
Stage 1KMashWeb Video2M-
Stage 2 (img-text)LAION, etcWeb Image300MAlt-text / Generated Caps
Stage 2 (vid-text)WebVid2MWeb Video250kAlt-text
WebVid10MWeb Video9.7MAlt-text
InternVidYoutube Video40MGenerated Caption
InternVid2Youtube Video50MGenerated Caption
Stage 3LLaVA, etcWeb Image/Video2.1MConversation, QA
", + "bbox": [ + 272, + 170, + 723, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a42554fc99b0bb401df147969de40c3b2d0a527c4be924025f4127a4531bac0c.jpg", + "image_caption": [ + "Fig. 3: The framework of our video multimodal annotation system, called VidCap, consists of four main components: video, audio, and speech captioners, along with a LLM for integrating captions from these modalities." + ], + "image_footnote": [], + "bbox": [ + 246, + 292, + 759, + 449 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Video-only Data for Masked Autoencoders", + "text_level": 1, + "bbox": [ + 214, + 530, + 614, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We curate a new video set without labels named $K$ -Mash from action recognition datasets [10, 35, 38, 65, 108], as detailed in Supp. It encompasses a wide range of video types, including first- and third-person perspectives, with both short and long duration, and featuring various settings. Further, we give $K$ - $\\text{Mash}_{2M}$ with additionally sourced and selected 844K videos from YouTube for diversity.", + "bbox": [ + 212, + 554, + 787, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Videos with Audio-Speech Modalities", + "text_level": 1, + "bbox": [ + 214, + 651, + 571, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We build a multimodal video dataset, coined as InternVid2, with video-audiospeech information and their descriptions for strengthening video perception via other modalities. It consists of 100M videos along with their VAS captions. In InternVid2, we collect videos from several sources (detailed in the supp), segment them into clips, and automatically annotate them based on their unimodal or crossmodal inputs. We highlight the importance of temporal segmentation in clip generation and our video multimodal annotation system. We find refining them leads to notable downstream improvements.", + "bbox": [ + 212, + 674, + 787, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Temporal Consistency Matters. We employ a temporal boundary detection model AutoShot [114] to segment videos into clips instead of SceneDet filter from", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "InternVideo2", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "FFmpeg. It predicts boundaries based on temporal semantic variations rather than pixel differences, and is able to generate semantically complete cuts without mixing extra frames with inconsistent context.", + "bbox": [ + 215, + 146, + 785, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Video Multimodal Annotation. We design a video multimodal annotation system VidCap to give proper unimodal and crossmodal descriptions for textualizing videos from different perceptions. It automatically captions visual, audio, and speech of InternVid2, then it corrects them and fuses them for cross-modal captions via LLM. The system frame is given in Fig. 3. VidCap has independent video, audio, speech captioner, and a LLM for caption refinement and fusion. For video, speech, and caption postprocessing, we employ existing methods as the video captioning pipeline in [93], WhisperV2-large model [75], and Vicuna-1.5 [111]. For audio, we craft a audio captioner upon VideoChat [48], as we find no open-sourced one. It extracts audio features from inputs by Beats [13]. We learn it by only tuning its Qformer using a combination of the large-scale audio-text corpus WavCaps [64] dataset. Details are given in the supplementary material.", + "bbox": [ + 215, + 209, + 787, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 417, + 375, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our evaluation of InternVideo2, we assess the models from three learning stages. The evaluation covers a wide range of tasks, including video recognition, video retrieval, question-answering, and more. It includes various scenarios such as zero-shot learning, finetuning, and linear probing.", + "bbox": [ + 215, + 452, + 785, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For InternVideo2 trained in stage 1, 2, and 3, we denote them with InternVideo2 $_{s1}$ , InternVideo2 $_{s2}$ , and InternVideo2 $_{s3}$ , respectively. We also learn a CLIP-style InternVideo2 indicated by InternVideo2 $_{clip}$ . It is post-pretrained from InternVideo2 $_{s2}$ by only preserving video and text encoders and contrastive loss.", + "bbox": [ + 215, + 513, + 803, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Each training stage of InternVideo2-6B uses different configurations and resources. In the first stage, we employ 256 NVIDIA A100 GPUs and train the model for 18 days. The second stage also utilizes 256 A100 GPUs and spans a training period of 14 days. Finally, in the third stage, we use 64 A100 GPUs and train the model for 3 days. We introduce DeepSpeed and FlashAttention [22] for training and inference. More implementation details, experiment results, and ablations are given in the supp.", + "bbox": [ + 215, + 574, + 785, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 Video Classification", + "text_level": 1, + "bbox": [ + 215, + 705, + 426, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Action Recognition. We test InternVideo2 on Kinetics (i.e., K400, 600 and 700 [8-10]), Moments in Time V1 (MiT) [65], Something-Something V2 (SSv2) [35], UCF [79], HMDB [43], Charades [32], ActivityNet [38] (ANet) and HACS [108]. We evaluate in four settings: (a) end-to-end finetuning the whole backbone; (b) attentive probing is similar to linear pooling, but extra trains the attention pooling layer [102]. (c) linear probing which freezes the backbone and only trains the task head; and (d) zero-shot.", + "bbox": [ + 215, + 733, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/51d24fadcfbca8953b2ca4bfff5099583a09486dcd91746c2fc466d9682abf8c.jpg", + "table_caption": [ + "Table 2: End-to-end finetuning action recognition results (top-1 accuracy) on Kinetics, SomethingSomething, and Moments in Time. ${}^{ \\dagger }$ denotes that the result is achieved with different resolutions or frame rates." + ], + "table_footnote": [], + "table_body": "
MethodTraining DataSettingK400K600K700Sth-Sthv2MiTANetHACS
CoVeR [106]IV-3B16 × 44887.187.979.870.846.1--
Hiera-H [78]V-0.25M16 × 22487.888.881.176.5---
CoCa-g [102]I-3B16 × 57688.989.482.7-49.0--
MTV-H [90]IV-370M32 × 28089.990.383.4----
VideoMAEv2-g [90]V-1.35M64 × 26690.089.9-77.0†---
V-JEPA-H [4]V-2M16 × 224---77.0---
MVD-H [92]IV-1.25M16 × 224---77.3---
UniFormerV2-L [50]IV-401M64 × 33690.090.182.773.0†47.8†94.795.4
InternVideo [94]V-12Mensemble91.191.384.077.2---
InternVideo2s1-1BIV-1.1M8 × 22491.391.485.077.150.8--
InternVideo2s1-1BIV-1.1M16 × 22491.691.685.477.150.9--
InternVideo2s1-6BIV-2M8 × 22491.991.785.777.551.0--
InternVideo2s1-6BIV-2M16 × 22492.191.985.977.451.295.997.0
", + "bbox": [ + 222, + 198, + 779, + 338 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/69f826745fff0c5f46fa3b2a86d1d363cb9cf989f403f1b606057a336e52960a.jpg", + "table_caption": [ + "Table 3: Attentive probing recognition results (top-1 accuracy) on Kinetics-400/600/700, Moments in Time and Something-Something V2." + ], + "table_footnote": [], + "table_body": "
MethodTraining DataSettingK400K600K700MiTSSV2
UMT-L [51]IV-25M-82.8--40.354.5
VideoMAEv2-g [90]V-1.35M-82.1--35.056.1
V-JEPA-H [4]V-2M16 × 38481.9---72.2
DINOv2-g [69]I-142M16 × 22483.4---50.0
VideoPrism-g [109]V-619M16 × 28887.2--45.568.5
ViT-e [23]I-4B128 × 22486.5--43.6-
ViT-22B [23]I-4B128 × 22488.0--44.9-
CoCa-g [102]I-3B16 × 57688.088.581.147.4-
InternVideo2s2-1BIV-25.5M16 × 22487.988.079.546.367.3
InternVideo2s2-6BIV-400M16 × 22488.889.181.047.867.7
", + "bbox": [ + 250, + 393, + 748, + 520 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "End-to-end Finetuning. Tab. 2 shows InternVideo2-6B obtains new state-of-the-art (SOTA) results on Kinetics $(92.1\\% / 91.9\\% / 85.9\\%)$ on K400/600/700, respectively), SthSthv2, MiT, ANet, and HACS with only 16 frames, while the previous SOTAs require larger resolution (224 vs. 576) or model ensemble. As for MiT in Table 2, InternVideo2-6B exceeds the previous SOTA, CoCa-g, by a significant margin of $2.2\\%$ $(51.2\\%$ vs. $49.0\\%)$ . Regarding the temporal-related actions in Table 2, our InternVideo2-6B also surpasses MVD [92] on SSv2 $(77.5\\%$ vs. $77.3\\%)$ . Moreover, our InternVideo2-6B showcases top performance on untrimmed video analysis, as indicated in Table 2, with $95.9\\%$ on ActivityNet and $97.0\\%$ on HACS. These results affirm our model's superior capability for robustly identifying complex actions across varied scenes. Note \"I\" and \"V\" denotes images and videos, respectively. \"IV-3B\" means the total number of the used images and videos is 3B, while \"I-3B\" means using 3B images.", + "bbox": [ + 215, + 549, + 787, + 744 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Attentive Probing. As in Tab. 3, InternVideo2-6B not only outperforms ViT-22B [23] and CoCa-g [102] in scene-focused datasets but also surpasses or matches the performance of the latest video foundation model [4,109], on datasets emphasizing temporal dynamics (SthSthV2). This underscores our model's exceptional ability to understand and interpret both spatial and temporal information effectively.", + "bbox": [ + 215, + 763, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "InternVideo2", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/933163192bf7208953121149e82c7cdd2def738fedebbbacc60388a61cc6b80c.jpg", + "table_caption": [ + "Table 4: Linear probing action recognition results (top-1 accuracy) on Kinetics-400, Something-Something V2, UCF-101 and HMDB-51." + ], + "table_footnote": [], + "table_body": "
MethodSettingK400SSV2UCF-101HMDB-51
VideoMAEv2-H [90]12 × 22425.8-56.434.1
TVTSv2-H [105]12 × 22473.1-91.865.7
OpenCLIP-G [20]8 × 22478.335.890.7-
DINOv2-g [69]8 × 22478.438.391.2-
InternVideo2s1-1B16 × 22481.646.396.071.6
InternVideo2s1-6B16 × 22482.047.896.371.8
InternVideo2s2-6B16 × 22484.256.797.380.7
", + "bbox": [ + 272, + 184, + 725, + 289 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/7865090ea8a5eace3424ab253185d7e55fba08d7604ac173f1c4b6166e9e9e33.jpg", + "table_caption": [ + "Table 5: Zero-shot action recognition on UCF, HMDB, MiTv1, SSv2-MC, and Charades." + ], + "table_footnote": [], + "table_body": "
Method#FTraining DataUCFHMDBMiTSSv2-MCCharades
CLIP [74]12I-400M68.943.2-29.6-
TVTSv2 [105]12V-8.5M78.052.1-48.4-
VideoCoCa-g [98]16V-145M----25.8
VideoPrism-g [109]16V-619M----32.4
InternVideo2clip-1B8IV-25.5M88.853.931.661.532.9
InternVideo2clip-6B8IV-400M89.556.732.963.534.6
", + "bbox": [ + 282, + 330, + 710, + 422 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/578eceb125de174e688ffda972410f77e3214d1d9b1c262bac3f2f0f3133acc7.jpg", + "table_caption": [ + "Table 6: Zero-shot action recognition results on Kinetics." + ], + "table_footnote": [], + "table_body": "
Method#FK400K600K700
top-1AVGtop-1AVGtop-1AVG
CLIP [74]858.470.155.167.246.158.4
EVA-CLIP-L [81]1-65.0-64.9-59.1
EVA-CLIP-E [81]1-69.8-69.3-63.4
ViCLIP-L [93]864.875.762.273.554.366.4
VideoCoCa-g [98]1672.081.3-
InternVL-6B [17]869.179.468.978.860.671.5
EVA-CLIP-18B [82]1679.4-79.4-72.2
VideoPrism-g [109]1676.485.4-
InternVideo2clip-1B873.182.472.881.864.975.2
InternVideo2clip-6B872.782.271.781.264.275.2
", + "bbox": [ + 334, + 464, + 661, + 609 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Linear Probing. In Tab. 4, InternVideo2-1B significantly outperforms the previous SOTA, DINOv2-g [69], by notable margins: $+3.2\\%$ on K400, $+8.0\\%$ on SthSthV2, and $+4.8\\%$ on UCF-101. As we scale the model, an upward trend in results is observed, underscoring the benefits of model enhancement. Notably, the integration of multimodal pretraining (stage 2) yields further rise in results. We suppose stage 2 enhances feature discrimination.", + "bbox": [ + 212, + 638, + 787, + 729 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Zero-shot. Table 5 and 6 show InternVideo2 gets $72.7\\%$ / $71.7\\%$ / $64.2\\%$ on K400/600/700, respectively, outperforming others but VideoPrism on K400 $(76.4\\%)$ . On UCF [79], HMDB [43], MiT [65], SSv2-MC, and Charades, InternVideo2 gives an cutting edges over others. The clear gap between VideoPrism and InternVideo2 on K400 may signify the importance of pretraining corpus in VideoPrism (311M videos with text and 36.1M of them are manually labeled) for", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/abbdbb43710db59cfca86f5da030c82cd6d5de05329820a1240a33ccecddaa39.jpg", + "table_caption": [ + "Table 7: Finetuned temporal action localization results on THUMOS14 [39], ActivityNet [42], HACS Segment [108] and FineAction [57]. We report average mAP. \"Flow\" denotes the ensembling I3D flow feature. * denotes the result is achieved with Flow." + ], + "table_footnote": [], + "table_body": "
BackboneTHUMOS14HACSActivityNetFineAction
I3D [10] + Flow66.8-35.6-
R(2+1)D [88]55.6-36.6-
InternVideo71.6*41.339.017.6
VideoMAEv2-g [90]69.5--18.2
InternVideo2s1-1B69.842.440.427.2
InternVideo2s1-6B72.043.341.227.7
", + "bbox": [ + 308, + 198, + 694, + 299 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/31010162db6b7737e17fbe83aabc599eeb4c96fc353f8aa81a33373f1c3a2081.jpg", + "table_caption": [ + "Table 8: Video instance segmentation performance (mAP) on YouTube-VIS19 [99]." + ], + "table_footnote": [], + "table_body": "
MethodBackbone#ParamsYouTubeVIS19
Mask2FormerSwin-L (image) [58]219M60.3
Mask2FormerInternViT (image)6B63.4
Mask2FormerInternVideo2s16B64.2
", + "bbox": [ + 330, + 345, + 666, + 412 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "K400 in zero-shot. Note that on the datasets of Kinetics, UCF101 and HMDB51, which have a distribution closer to the pre-training dataset used in stage1, the performance of Intervvideo2-6B is slightly inferior to that of Intervvideo2-1B. We suppose this is caused by Intervvideo2-6B uses a more abundant pretraining dataset in stage2, leading to the forgetting of pretraining data in stage1.", + "bbox": [ + 212, + 436, + 787, + 513 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Temporal Action Localization. We evaluate models on four temporal action localization (TAL) datasets: THUMOS14 [39], ActivityNet [42], HACS Segment [108] and FineAction [57] in a feature-based manner with finetuning. We employ output of the 7-th layer from InternVideo2 for inputs as the corresponding features without fusing anything else. ActionFormer [107] is used as the detection head. We report mean Average Precision(mAP) under multiple tIoU as in [12,54,100,107]. In Table 7, InternVideo2-6B gets the highest mAP among all comparisons in all datasets, while InternVideo2-1B nearly surpass other methods except in THUMOS14. We find InternVideo2-6B almost consistently improves mAP with a notable margin from InternVideo2-1B except in FineAction. We suppose scaling model capacity without data refinement cannot nontrivially improve fine-grained discrimination abilities of models. Scaling detailed annotations in training may address this issue.", + "bbox": [ + 212, + 532, + 787, + 729 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Video Instance Segmentation. We evaluate on the Video Instance Segmentation (VIS) dataset Youtube-VIS 2019 [99]. Built upon Mask2Former [18], we employ the video encoder of InternVideo2 as backbone with ViT-adapter [16] for features. We also try InternViT [17] for comparisons. In Table 8, InternVideo2 gets the highest mAP among all. This validates its effectiveness in relatively fine-grained spatiotemporal perception.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "InternVideo2", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ad714725eade35976b9de49d57b75bdb2f84e48b84b3e5e9f6a5a06eb3401b04.jpg", + "table_caption": [ + "Table 9: Results of zero-shot video retrieval in both text-to-video (T2V) and video-to-text (V2T) on MSR-VTT, LSMDC, DiDeMo, MSVD, ANet, and VATEX." + ], + "table_footnote": [], + "table_body": "
MethodMSR-VTTLSMDCDiDeMoMSVDANetVATEX
T2VV2TT2VV2TT2VV2TT2VV2TT2VV2TT2VV2T
CLIP [74]30.424.213.911.912.718.740.557.29.113.2--
CLIP4Clip [59]32.0-15.1---38.5-----
ViCLIP [93]42.441.320.116.918.427.949.175.115.124.0--
InternVideo-L [94]40.739.617.613.231.533.543.467.630.731.449.569.5
UMT-L [51]40.737.124.921.948.649.949.074.541.939.4--
VideoCoCa-g [98]34.464.7------34.533.053.273.6
VideoPrism-g [109]39.771.0------52.750.362.577.1
InternVideo2s2-1B51.950.932.027.357.054.358.183.360.454.870.485.4
InternVideo2s2-6B55.953.733.830.157.957.159.383.163.256.571.585.3
", + "bbox": [ + 250, + 184, + 748, + 314 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/66b611211e7b4e683818d834b302778ac42cce2c90214a650e42c74233d57182.jpg", + "table_caption": [ + "Table 10: Results of finetuning video retrieval in both text-to-video (T2V) and video-to-text (V2T) on MSR-VTT, LSMDC, DiDeMo, MSVD, ANet, and VATEX." + ], + "table_footnote": [], + "table_body": "
MethodMSR-VTTLSMDCDiDeMoMSVDANetVATEX
T2VV2TT2VV2TT2VV2TT2VV2TT2VV2TT2VV2T
CLIP [74]38.238.722.522.632.233.9--26.126.9--
CLIP4Clip [59]45.645.924.323.843.043.645.248.440.341.6--
ViCLIP [93]52.551.833.032.549.450.2--49.848.1--
UMT-L [51]58.858.643.041.470.465.758.282.466.864.472.086.0
InternVideo2s2-6B62.860.246.446.774.271.961.485.274.169.775.589.3
", + "bbox": [ + 250, + 367, + 751, + 458 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Video-Audio-Language Tasks", + "text_level": 1, + "bbox": [ + 215, + 484, + 501, + 500 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We evaluate InternVideo2 on video retrieval, captioning, and multi-choice question-answering (QA). The former two tasks are conducted by matching video representation and the candidate text ones using the text encoder in stage 2. The latter is tested by the VideoLLM learned in stage 3. We also test audio tasks.", + "bbox": [ + 212, + 508, + 787, + 570 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Video Retrieval. We evaluate the video retrieval on six popular benchmarks [2, 11,38,77,97], as shown in Tab. 9 and 10. In eval, eight frames from the input videos are uniformly sampled. We report $R@1$ scores for both text-to-video (t2v) and video-to-text (v2t) tasks in Tab. 9 and 10. R@5 and R@10 are given in Supp.", + "bbox": [ + 212, + 590, + 787, + 651 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 9 and 10 demonstrate that InternVideo2 outperforms other state-of-the-arts with a notable margin in both t2v and v2t of all used datasets no matter in zero-shot or finetuned settings, except for the v2t of MSR-VTT, where VideoPrism gives the best result. This shows the video-language semantic alignment of transferrity of InternVideo2.", + "bbox": [ + 212, + 652, + 787, + 727 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Video Temporal Grounding. We evaluate InternVideo2 on two video temporal grounding (VTG) datasets: Qvhighlight [44], and Charade-STA [32]. The eval setting and used features are the same as in TAL. We use CG-DETR [66] as the grounding head. We report R1@0.3, R1@0.5, R1@0.7, and mAP for moment retrieval as in [44, 53, 66]. Highlight Detection is evaluated in terms of \"Very Good\" mAP and HiT@1. In Table 11, InternVideo2-1B and InternVideo2-6B bring", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/0b296c671514b6aad45ae2861063012a8c520d802e5499dd445f13f9d6560297.jpg", + "table_caption": [ + "Table 11: Finetuned temporal grounding on QVHighlight [44] and Charade-STA [32]." + ], + "table_footnote": [], + "table_body": "
FeatureR1@0.5R1@0.7mAPmAPHiT@1
CLIP [74]64.9748.6542.9639.8364.19
CLIP+SlowFast [29]65.4348.3842.8640.3366.21
InternVideo2s2-1B70.0054.4547.0242.3669.74
InternVideo2s2-6B71.4256.4549.2442.9072.00
", + "bbox": [ + 218, + 170, + 506, + 229 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/eb477f6d3c56a7f96fd364cfb2aaca038116080b695c66975e0199a08ae6dcb9.jpg", + "table_caption": [ + "(a) QVHighlight" + ], + "table_footnote": [], + "table_body": "
FeatureR1@0.3R1@0.5R1@0.7mIoU
CLIP [74]65.6252.7730.1645.85
CLIP+SlowFast [29]70.4358.4436.3450.13
InternVideo2s2-1B78.4168.3645.0357.12
InternVideo2s2-6B79.7070.0348.9558.79
", + "bbox": [ + 511, + 170, + 771, + 229 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/363f69c0fcbd50140a67b14b7472d1fce4b3a7368464bd8fd17f1b39cd74cfea.jpg", + "table_caption": [ + "(b) Charade-STA", + "Table 12: Audio retrieval results on AudioCaps [41], Clothov1, and Clothov2 [27]. We report text-to-audio R@1 accuracy in zero-shot and finetuning settings." + ], + "table_footnote": [], + "table_body": "
MethodZero-shotFinetuning
AudioCapsClothoV1ClothoV2AudioCapsClothoV1ClothoV2
VIP-ANT [110]27.7---
VAST [15]---52.025.126.9
LanguageBind [112]-12.112.1---
InternVideo2s2-6B37.117.417.455.225.327.2
", + "bbox": [ + 250, + 310, + 751, + 390 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/433c74bab5f66fbb9511acc44f311a1c2c335cc185298d9a2b0b18cc665914db.jpg", + "table_caption": [ + "Table 13: Results of AudioQA on ClothoAQA [55] and Audio-MusicAVQA (AMAVQA) [5], and audio classification on the ESC-50 [71]. Both in the finetuning setting." + ], + "table_footnote": [], + "table_body": "
BackboneClothQAAMAVQA
AquaNet [55]14.7865.59
MWAFM [45]22.2467.54
InternVideo2s230.1480.51
", + "bbox": [ + 251, + 444, + 464, + 503 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/9f703ac0410f15919ef3e645344499a233514c3ec7410f241e1c037cb1e6c4e3.jpg", + "table_caption": [ + "(a) ClothoAQA and AMAVQA" + ], + "table_footnote": [], + "table_body": "
MethodTop-1 Acc
AST [34]95.60
BEATS [13]98.10
InternVideo2s298.60
", + "bbox": [ + 568, + 444, + 723, + 503 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "(b) ESC-50", + "bbox": [ + 611, + 503, + 678, + 515 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "gradual performance improvements compared to CLIP [73] and CLIP [73] + Slow-fast [29]. This suggests that a larger spatiotemporal model is more beneficial to short-term video semantic alignment capabilities.", + "bbox": [ + 212, + 556, + 785, + 603 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Audio-related Tasks. We evaluate InternVideo2's audio and text encoders on audio tasks, including audio-text retrieval on AudioCaps [41], Clothov1, and Clothov2 [27]; audioQA on ClothoAQA [55] and Audio-MusicAVQA [5]; and audio classification on the ESC-50 [71]. As shown in Tab. 12, 13a, and 13b, our model achieves state-of-the-art performance on all downstream tasks. Considering the limited size of the used audio and text encoders, these audio-related results show crossmodal contrastive learning's benefits are mutual to the used modalities. Audio and the corresponding text models also gain from this learning.", + "bbox": [ + 212, + 625, + 787, + 748 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Video-centric Dialogue and its Applications", + "text_level": 1, + "bbox": [ + 212, + 768, + 620, + 786 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Tab. 14 shows results on MVBench [49], Egoschema [62], and Perception Test [70], by equipping VideoChat2 [49] with our InternVideo2 encoder. VideoChat2-InternVideo2 demonstrates a significant performance advantage over other systems", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "InternVideo2", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/162bba4ff3ed9b2c446b84af73b40cc12924fb6be35b3237a02d2bf654d0ed70.jpg", + "table_caption": [ + "Table 14: Results of Chat-centric Evaluation on Multiple Choice Video-QA on MVBench [49], Egoschema and Perception Test." + ], + "table_footnote": [], + "table_body": "
ModelViEncoderLLMMVBenchEgoschemaPerception Test
GPT-4V [68]-GPT-443.5--
Gemini 1.0 Pro [84]--37.755.751.1
Gemini 1.0 Ultra [84]---61.554.7
Gemini 1.5 Pro [84]---72.2-
LLaVA-Next-VideoCLIP-LVicuna-7B46.543.948.8
VideoLLaMA2CLIP-L-336Mistral-7B54.651.751.4
VideoLLaMA2CLIP-L-336Mistral-8*7B53.953.352.2
VideoChat2UMT-LVicuna-7B51.1--
VideoChat2InternVideo2s3-1BMistral-7B60.355.853.0
", + "bbox": [ + 222, + 184, + 779, + 305 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "in MVBench. In Egoschema and Perception Test evaluations, it stands out as the top performer among open-source models. When contrasted with proprietary commercial models like Gemini, VideoChat2-InternVideo2 exhibits results that are slightly lower but still comparable in the Perception Test. Egoschema requires longer context harnessing, and we plan to further explore leveraging LLM for long multimodal sequence modeling. Considering these benchmarks not only involve perception but also reasoning, these results suggest InternVideo2 does embed knowledge to partially model the world. It also validates the importance of learning a transferrable video representation for current video-related MLLM. We also give several qualitative evaluations with popular GPT-4V and GeminiPro (in supp).", + "bbox": [ + 212, + 333, + 789, + 501 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 215, + 522, + 509, + 539 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have introduced a new family of video foundation models called InternVideo2, which achieves the state-of-the-art performance across various video and audio tasks. In InternVideo2, we combine masked video modeling, video-audio-text contrastive learning, and next token prediction into a unified framework. Additionally, we create a new video-text dataset that incorporates video-audio-speech fused captions as descriptions. The dataset contains temporally segmented clips with high semantic coherence. These designs in InternVideo2 contribute to enhancing video understanding in both perception and reasoning tasks. Notably, InternVideo2 excels in video-related dialogue and long video understanding, demonstrating its effectiveness in capturing high-level semantics.", + "bbox": [ + 212, + 553, + 789, + 705 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations and Discussions. Despite its achievements, InternVideo2 does not introduce specific novel architectural design. Instead, it leverages the existing learning techniques for scaling video foundation models while focusing on improving data processing to enhance its spatiotemporal perception, semantic alignment, and basic knowledge embedding. Similarly to previous studies [51, 101], InternVideo2 still grapples with limitations stemming from fixed input resolutions, sampling rates, and highly compressed tokens, which restrict its ability to express rich video information and capture fine-grained details.", + "bbox": [ + 212, + 719, + 789, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is partially supported by the National Key R&D Program of China (No. 2022ZD0160102), National Natural Science Foundation of China (No. 62076119, No. 61921006)), and the Science and Technology Commission of Shanghai Municipality under Grant No. 23QD1400800 and No. 23YF1461900.", + "bbox": [ + 212, + 175, + 787, + 236 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 258, + 321, + 273 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: a visual language model for few-shot learning. ArXiv abs/2204.14198 (2022)", + "2. Anne Hendricks, L., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.: Localizing moments in video with natural language. In: ICCV. pp. 5803-5812 (2017)", + "3. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: ICCV (2021)", + "4. Bardes, A., Garrido, Q., Ponce, J., Chen, X., Rabbat, M., LeCun, Y., Assran, M., Ballas, N.: V-JEPA: Latent video prediction for visual representation learning (2024), https://openreview.net/forum?id=WFYbBOEOtv", + "5. Behera, S.R., Injeti, K.M., Patibandla, J.S.K., Pokala, P.K., Pailla, B.R.: Aquallm: Audio question answering data generation using large language models. arXiv preprint arXiv:2312.17343 (2023)", + "6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS (2020)", + "7. Bruce, J., Dennis, M., Edwards, A., Parker-Holder, J., Shi, Y., Hughes, E., Lai, M., Mavalankar, A., Steigerwald, R., Apps, C., et al.: Genie: Generative interactive environments. arXiv preprint arXiv:2402.15391 (2024)", + "8. Carreira, J., Noland, E., Banki-Horvath, A., Hillier, C., Zisserman, A.: A short note about kinetics-600. ArXiv abs/1808.01340 (2018)", + "9. Carreira, J., Noland, E., Hillier, C., Zisserman, A.: A short note on the kinetics-700 human action dataset. ArXiv abs/1907.06987 (2019)", + "10. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: CVPR (2017)", + "11. Chen, D.L., Dolan, W.B.: Collecting highly parallel data for paraphrase evaluation. In: Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies-Volume 1. pp. 190-200. Association for Computational Linguistics (2011)", + "12. Chen, G., Zheng, Y.D., Wang, L., Lu, T.: Dscan: improving temporal action detection via dual context aggregation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 248-257 (2022)", + "13. Chen, S., Wu, Y., Wang, C., Liu, S., Tompkins, D., Chen, Z., Che, W., Yu, X., Wei, F.: BEATs: Audio pre-training with acoustic tokenizers. In: Proceedings of the 40th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 202, pp. 5178-5193. PMLR (23-29 Jul 2023)" + ], + "bbox": [ + 225, + 287, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "InternVideo2", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Chen, S., He, X., Guo, L., Zhu, X., Wang, W., Tang, J., Liu, J.: Valor: Vision-audio-language omni-perception pretraining model and dataset. arXiv preprint arXiv:2304.08345 (2023)", + "15. Chen, S., Li, H., Wang, Q., Zhao, Z., Sun, M., Zhu, X., Liu, J.: Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. Advances in Neural Information Processing Systems 36 (2024)", + "16. Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534 (2022)", + "17. Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Zhong, M., Zhang, Q., Zhu, X., Lu, L., Li, B., Luo, P., Lu, T., Qiao, Y., Dai, J.: Intervl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238 (2023)", + "18. Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: NeurIPS (2021)", + "19. Cheng, F., Wang, X., Lei, J., Crandall, D.J., Bansal, M., Bertasius, G.: Vindlu: A recipe for effective video-and-language pretraining. ArXiv abs/2212.05051 (2022)", + "20. Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: CVPR. pp. 2818-2829 (2023)", + "21. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B.A., Fung, P., Hoi, S.C.H.: Instructblip: Towards general-purpose vision-language models with instruction tuning. In: NeurIPS (2023)", + "22. Dao, T., Fu, D., Ermon, S., Rudra, A., Ré, C.: Flashattention: Fast and memory-efficient exact attention with io-awareness. NeurIPS 35, 16344-16359 (2022)", + "23. Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A., Caron, M., Geirhos, R., Alabdulmohsin, I.M., Jenatton, R., Beyer, L., Tschannen, M., Arnab, A., Wang, X., Riquelme, C., Minderer, M., Puigcerver, J., Evci, U., Kumar, M., van Steenkiste, S., Elsayed, G.F., Mahendran, A., Yu, F., Oliver, A., Huot, F., Bastings, J., Collier, M., Gritsenko, A.A., Birodkar, V., Vasconcelos, C.N., Tay, Y., Mensink, T., Kolesnikov, A., Paveti'c, F., Tran, D., Kipf, T., Luvci'c, M., Zhai, X., Keysers, D., Harmsen, J., Houlsby, N.: Scaling vision transformers to 22 billion parameters. In: ICML (2023)", + "24. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. ArXiv abs/1810.04805 (2018)", + "25. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv abs/2010.11929 (2020)", + "26. Driess, D., Xia, F., Sajjadi, M.S.M., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q.H., Yu, T., Huang, W., Chebotar, Y., Sermanet, P., Duckworth, D., Levine, S., Vanhoucke, V., Hausman, K., Toussaint, M., Greff, K., Zeng, A., Mordatch, I., Florence, P.R.: Palm-e: An embodied multimodal language model. In: ICML (2023)", + "27. Drossos, K., Lipping, S., Virtanen, T.: Cloth: An audio captioning dataset. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 736-740. IEEE (2020)", + "28. Feichtenhofer, C., Fan, H., Li, Y., He, K.: Masked autoencoders as spatiotemporal learners. NeurIPS (2022)" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: ICCV (2019)", + "30. Fu, T.J., Li, L., Gan, Z., Lin, K., Wang, W.Y., Wang, L., Liu, Z.: Violet: End-to-end video-language transformers with masked visual-token modeling. arXiv preprint arXiv:2111.12681 (2021)", + "31. Gabeur, V., Sun, C., Alahari, K., Schmid, C.: Multi-modal transformer for video retrieval. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 214-229. Springer (2020)", + "32. Gao, J., Sun, C., Yang, Z., Nevatia, R.: Tall: Temporal activity localization via language query. In: ICCV (2017)", + "33. Gong, T., Lyu, C., Zhang, S., Wang, Y., Zheng, M., Zhao, Q., Liu, K., Zhang, W., Luo, P., Chen, K.: Multimodal-gpt: A vision and language model for dialogue with humans. ArXiv abs/2305.04790 (2023)", + "34. Gong, Y., Chung, Y.A., Glass, J.: Ast: Audio spectrogram transformer. arXiv preprint arXiv:2104.01778 (2021)", + "35. Goyal, R., Kahou, S.E., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fründ, I., Yianilos, P., Mueller-Freitag, M., Hoppe, F., Thurau, C., Bax, I., Memisevic, R.: The \"something something\" video database for learning and evaluating visual common sense. In: ICCV (2017)", + "36. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: CVPR (2017)", + "37. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022)", + "38. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: CVPR (2015)", + "39. Idrees, H., Zamir, A.R., Jiang, Y.G., Gorban, A., Laptev, I., Sukthankar, R., Shah, M.: The thumos challenge on action recognition for videos \"in the wild\". Computer Vision and Image Understanding (2017)", + "40. Jiang, A.Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D.S., Casas, D.d.l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al.: Mistral 7b. arXiv preprint arXiv:2310.06825 (2023)", + "41. Kim, C.D., Kim, B., Lee, H., Kim, G.: Audiocaps: Generating captions for audios in the wild. In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 119-132 (2019)", + "42. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Carlos Niebles, J.: Dense-captioning events in videos. In: ICCV (2017)", + "43. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: 2011 International conference on computer vision. pp. 2556-2563. IEEE (2011)", + "44. Lei, J., Berg, T.L., Bansal, M.: Qvhighlights: Detecting moments and highlights in videos via natural language queries (2021)", + "45. Li, G., Xu, Y., Hu, D.: Multi-scale attention for audio question answering. arXiv preprint arXiv:2305.17993 (2023)", + "46. Li, J., Li, D., Savarese, S., Hoi, S.C.H.: Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In: ICML (2022)", + "47. Li, J., Li, D., Xiong, C., Hoi, S.C.H.: Blip: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In: ICML (2022)" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "InternVideo2", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355 (2023)", + "49. Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Liu, Y., Wang, Z., Xu, J., Chen, G., Luo, P., et al.: Mvbench: A comprehensive multi-modal video understanding benchmark. arXiv preprint arXiv:2311.17005 (2023)", + "50. Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Wang, L., Qiao, Y.: Uniformerv2: Spatiotemporal learning by arming image vits with video uniformer. arXiv preprint arXiv:2211.09552 (2022)", + "51. Li, K., Wang, Y., Li, Y., Wang, Y., He, Y., Wang, L., Qiao, Y.: Unmasked teacher: Towards training-efficient video foundation models. arXiv preprint arXiv:2303.16058 (2023)", + "52. Li, T., Wang, L.: Learning spatiotemporal features via video and text pair discrimination. CoRR abs/2001.05691 (2020), https://arxiv.org/abs/2001.05691", + "53. Lin, K.Q., Zhang, P., Chen, J., Pramanick, S., Gao, D., Wang, A.J., Yan, R., Shou, M.Z.: Univtg: Towards unified video-language temporal grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2794-2804 (2023)", + "54. Lin, T., Liu, X., Li, X., Ding, E., Wen, S.: Bmn: Boundary-matching network for temporal action proposal generation (2019)", + "55. Lipping, S., Sudarsanam, P., Drossos, K., Virtanen, T.: Clotho-aqa: A crowdsourced dataset for audio question answering. In: 2022 30th European Signal Processing Conference (EUSIPCO). pp. 1140-1144. IEEE (2022)", + "56. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023)", + "57. Liu, Y., Wang, L., Wang, Y., Ma, X., Qiao, Y.: Fineaction: A fine-grained video dataset for temporal action localization. IEEE Transactions on Image Processing 31, 6937-6950 (2022)", + "58. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)", + "59. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing (2022)", + "60. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M.H., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability. ArXiv abs/2306.07207 (2023)", + "61. Maaz, M., Rasheed, H.A., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. ArXiv abs/2306.05424 (2023)", + "62. Mangalam, K., Akshulakov, R., Malik, J.: Egoschema: A diagnostic benchmark for very long-form video language understanding. In: Oh, A., Neumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) NeurIPS. pp. 46212-46244 (2023)", + "63. Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. In: CVPR (2019)", + "64. Mei, X., Meng, C., Liu, H., Kong, Q., Ko, T., Zhao, C., Plumbley, M.D., Zou, Y., Wang, W.: Wavcaps: A chatgpt-assisted weakly-labelled audio captioning dataset for audio-language multimodal research (2023)", + "65. Monfort, M., Zhou, B., Bargal, S.A., Andonian, A., Yan, T., Ramakrishnan, K., Brown, L.M., Fan, Q., Gutfreund, D., Vondrick, C., Oliva, A.: Moments in time dataset: One million videos for event understanding. TPAMI (2020)" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "66. Moon, W., Hyun, S., Lee, S., Heo, J.P.: Correlation-guided query-dependency calibration in video representation learning for temporal grounding. arXiv preprint arXiv:2311.08835 (2023)", + "67. OpenAI: Gpt-4 technical report. ArXiv abs/2303.08774 (2023)", + "68. OpenAI: Gpt-4v(ison) system card. https://api_semanticscholar.org/ CorpusID:263218031 (2023)", + "69. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)", + "70. Patraucean, V., Smaira, L., Gupta, A., Continente, A.R., Markeeva, L., Dylan, Banarse, Malinowski, M., Yang, Y., Doersch, C., Matejovicova, T., Sulsky, Y., Antoine, Miech, Koppula, S., Fréchette, A., Klimczak, H., Koster, R., Zhang, J., Stephanie, Winkler, Aytar, Y., Osindero, S., Damen, D., Zisserman, A., Carreira, J.: Perception test: A diagnostic benchmark for multimodal models. In: NeurIPS (2023)", + "71. Piczak, K.J.: Esc: Dataset for environmental sound classification. In: Proceedings of the 23rd ACM international conference on Multimedia. pp. 1015-1018 (2015)", + "72. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: ICCV (2015)", + "73. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)", + "74. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763. PMLR (2021)", + "75. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision. In: International Conference on Machine Learning. pp. 28492-28518. PMLR (2023)", + "76. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR (2020)", + "77. Rohrbach, A., Rohrbach, M., Tandon, N., Schiele, B.: A dataset for movie description. In: CVPR. pp. 3202-3212 (2015)", + "78. Ryali, C., Hu, Y.T., Bolya, D., Wei, C., Fan, H., Huang, P.Y., Aggarwal, V., Chowdhury, A., Poursaeed, O., Hoffman, J., Malik, J., Li, Y., Feichtenhofer, C.: Hiera: A hierarchical vision transformer without the bells-and-whistles (2023)", + "79. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)", + "80. Sun, Q., Cui, Y., Zhang, X., Zhang, F., Yu, Q., Luo, Z., Wang, Y., Rao, Y., Liu, J., Huang, T., et al.: Generative multimodal models are in-context learners. arXiv preprint arXiv:2312.13286 (2023)", + "81. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023)", + "82. Sun, Q., Wang, J., Yu, Q., Cui, Y., Zhang, F., Zhang, X., Wang, X.: Eva-clip-18b: Scaling clip to 18 billion parameters. arXiv preprint arXiv:2402.04252 (2024)", + "83. Sun, Q., Yu, Q., Cui, Y., Zhang, F., Zhang, X., Wang, Y., Gao, H., Liu, J., Huang, T., Wang, X.: Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222 (2023)" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "InternVideo2", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "84. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)", + "85. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. NeurIPS 35, 10078-10093 (2022)", + "86. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models. ArXiv abs/2302.13971 (2023)", + "87. Touvron, H., Martin, L., Stone, K.R., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D.M., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A.S., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I.M., Korenev, A.V., Koura, P.S., Lachaux, M.A., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. ArXiv abs/2307.09288 (2023)", + "88. Tran, D., xiu Wang, H., Torresani, L., Ray, J., LeCun, Y., Paluri, M.: A closer look at spatiotemporal convolutions for action recognition. In: CVPR (2018)", + "89. Wang, J., Ge, Y., Yan, R., Ge, Y., Lin, K.Q., Tsutsui, S., Lin, X., Cai, G., Wu, J., Shan, Y., et al.: All in one: Exploring unified video-language pre-training. In: CVPR. pp. 6598-6608 (2023)", + "90. Wang, L., Huang, B., Zhao, Z., Tong, Z., He, Y., Wang, Y., Wang, Y., Qiao, Y.: Videomae v2: Scaling video masked autoencoders with dual masking. In: CVPR (2023)", + "91. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Gool, L.V.: Temporal segment networks: Towards good practices for deep action recognition. In: ECCV (2016)", + "92. Wang, R., Chen, D., Wu, Z., Chen, Y., Dai, X., Liu, M., Yuan, L., Jiang, Y.G.: Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In: CVPR. pp. 6312-6322 (2023)", + "93. Wang, Y., He, Y., Li, Y., Li, K., Yu, J., Ma, X., Li, X., Chen, G., Chen, X., Wang, Y., et al.: Intervid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942 (2023)", + "94. Wang, Y., Li, K., Li, Y., He, Y., Huang, B., Zhao, Z., Zhang, H., Xu, J., Liu, Y., Wang, Z., Xing, S., Chen, G., Pan, J., Yu, J., Wang, Y., Wang, L., Qiao, Y.: Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191 (2022)", + "95. Xu, H., Ye, Q., Yan, M., Shi, Y., Ye, J., Xu, Y., Li, C., Bi, B., Qian, Q., Wang, W., et al.: mplug-2: A modularized multi-modal foundation model across text, image and video. arXiv preprint arXiv:2302.00402 (2023)", + "96. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettle-moyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084 (2021)", + "97. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: CVPR. pp. 5288-5296 (2016)" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 19 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "98. Yan, S., Zhu, T., Wang, Z., Cao, Y., Zhang, M., Ghosh, S., Wu, Y., Yu, J.: Video-text modeling with zero-shot transfer from contrastive captioners. ArXiv abs/2212.04979 (2022)", + "99. Yang, L., Fan, Y., Xu, N.: Video instance segmentation. In: ICCV (2019)", + "100. Yang, M., Chen, G., Zheng, Y.D., Lu, T., Wang, L.: Basictad: an astounding rgb-only baseline for temporal action detection. Computer Vision and Image Understanding p. 103692 (2023)", + "101. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., Jiang, C., Li, C., Xu, Y., Chen, H., Tian, J., Qi, Q., Zhang, J., Huang, F.: mplug-owl: Modularization empowers large language models with multimodality (2023)", + "102. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 (2022)", + "103. Zablocki, E., Ben-Younes, H., Pérez, P., Cord, M.: Explainability of deep vision-based autonomous driving systems: Review and challenges. International Journal of Computer Vision 130(10), 2425-2452 (2022)", + "104. Zellers, R., Lu, J., Lu, X., Yu, Y., Zhao, Y., Salehi, M., Kusupati, A., Hessel, J., Farhadi, A., Choi, Y.: Merlot reserve: Neural script knowledge through vision and language and sound. In: CVPR. pp. 16375-16387 (2022)", + "105. Zeng, Z., Ge, Y., Tong, Z., Liu, X., Xia, S.T., Shan, Y.: Tvtsv2: Learning out-of-the-box spatiotemporal visual representations at scale. arXiv preprint arXiv:2305.14173 (2023)", + "106. Zhang, B., Yu, J., Fifty, C., Han, W., Dai, A.M., Pang, R., Sha, F.: Cotraining transformer with videos and images improves action recognition. ArXiv abs/2112.07175 (2021)", + "107. Zhang, C.L., Wu, J.Z., Li, Y.: Actionformer: Localizing moments of actions with transformers. ArXiv abs/2202.07925 (2022)", + "108. Zhao, H., Torralba, A., Torresani, L., Yan, Z.: Hacs: Human action clips and segments dataset for recognition and temporal localization. In: ICCV (2019)", + "109. Zhao, L., Gundavarapu, N.B., Yuan, L., Zhou, H., Yan, S., Sun, J.J., Friedman, L., Qian, R., Weyand, T., Zhao, Y., Hornung, R., Schroff, F., Yang, M.H., Ross, D.A., Wang, H., Adam, H., Sirotenko, M., Liu, T., Gong, B.: Videoprism: A foundational visual encoder for video understanding (2024)", + "110. Zhao, Y., Hessel, J., Yu, Y., Lu, X., Zellers, R., Choi, Y.: Connecting the dots between audio and text without parallel data through visual knowledge transfer. arXiv preprint arXiv:2112.08995 (2021)", + "111. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685 (2023)", + "112. Zhu, B., Lin, B., Ning, M., Yan, Y., Cui, J., Wang, H., Pang, Y., Jiang, W., Zhang, J., Li, Z., et al.: Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852 (2023)", + "113. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. ArXiv abs/2304.10592 (2023)", + "114. Zhu, W., Huang, Y., Xie, X., Liu, W., Deng, J., Zhang, D., Wang, Z., Liu, J.: Autoshot: A short video dataset and state-of-the-art shot boundary detection. In: CVPR. pp. 2237-2246 (2023)" + ], + "bbox": [ + 217, + 146, + 785, + 825 + ], + "page_idx": 20 + }, + { + "type": "header", + "text": "InternVideo2", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 20 + }, + { + "type": "page_number", + "text": "21", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 20 + } +] \ No newline at end of file diff --git a/2024/InternVideo2_ Scaling Foundation Models for Multimodal Video Understanding/3f170c5b-f8dd-444c-a833-330298501085_model.json b/2024/InternVideo2_ Scaling Foundation Models for Multimodal Video Understanding/3f170c5b-f8dd-444c-a833-330298501085_model.json new file mode 100644 index 0000000000000000000000000000000000000000..01a44c2a1bc2adc8e14728be48dbfa502423c40b --- /dev/null +++ b/2024/InternVideo2_ Scaling Foundation Models for Multimodal Video Understanding/3f170c5b-f8dd-444c-a833-330298501085_model.json @@ -0,0 +1,3223 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.246, + 0.141, + 0.758, + 0.187 + ], + "angle": 0, + "content": "InternVideo2: Scaling Foundation Models for Multimodal Video Understanding" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.212, + 0.763, + 0.275 + ], + "angle": 0, + "content": "Yi Wang\\(^{1}\\), Kunchang Li\\(^{3,1}\\), Xinhao Li\\(^{2,1}\\), Jiashuo Yu\\(^{3}\\), Yinan He\\(^{1}\\), Guo Chen\\(^{2,1}\\), Baoqi Pei\\(^{1}\\), Rongkun Zheng\\(^{1}\\), Zun Wang\\(^{1}\\), Yansong Shi\\(^{1}\\), Tianxiang Jiang\\(^{1}\\), Songze Li\\(^{1}\\), Jilan Xu\\(^{1}\\), Hongjie Zhang\\(^{1}\\), Yifei Huang\\(^{1}\\), Yu Qiao\\(^{\\dagger}\\), Yali Wang\\(^{\\dagger 3,1}\\), and Limin Wang\\(^{\\dagger 2,1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.285, + 0.697, + 0.314 + ], + "angle": 0, + "content": "1 OpenGVLab, Shanghai AI Laboratory, Shanghai, China \n2 Nanjing University, Nanjing, China" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.314, + 0.739, + 0.356 + ], + "angle": 0, + "content": "3 Shenzhen Institutes of Advanced Technology, CAS, Shenzhen, China {wangyi,qiaoyu,wangyali,wanglimin}@pjlab.org.cn https://github.com/OpenGVLab/InternVideo/tree/main/InternVideo?" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.393, + 0.778, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.519, + 0.788, + 0.562 + ], + "angle": 0, + "content": "Fig. 1: InternVideo2 yields strong transferable visual and visual-linguistic representations for action recognition, video-text understanding, to video-centric dialogue. It also exhibits capability of long-form video understanding and procedure-aware reasoning." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.612, + 0.744, + 0.793 + ], + "angle": 0, + "content": "Abstract. We introduce InternVideo2, a new family of video foundation models (ViFM) that achieve the state-of-the-art results in video recognition, video-text tasks, and video-centric dialogue. Our core design is a progressive training approach that unifies the masked video modeling, crossmodal contrastive learning, and next token prediction, scaling up the video encoder size to 6B parameters. At the data level, we prioritize spatiotemporal consistency by semantically segmenting videos and generating video-audio-speech captions. This improves the alignment between video and text. Through extensive experiments, we validate our designs and demonstrate superior performance on over 60 video and audio tasks. Notably, our model outperforms others on various video-related dialogue and long video understanding benchmarks, highlighting its ability to reason and comprehend longer contexts." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.806, + 0.725, + 0.82 + ], + "angle": 0, + "content": "Keywords: Video Foundation Model \\(\\cdot\\) Progressive Training \\(\\cdot\\) MLLM." + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.541, + 0.84 + ], + "angle": 0, + "content": "*Equal contribution. †Corresponding authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.377, + 0.162 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.297 + ], + "angle": 0, + "content": "Learning transferrable spatiotemporal representations is a critical research area in computer vision, holding diverse applications across domains such as video search [31], game control [7], robotic learning [26], self-driving [103], and scientific studies [84]. Recently, the advancement of Large Language Models (LLMs) [6, 67, 86, 87] and their multimodal variations (MLLMs) [33, 56, 68, 84] have had a profound impact on vision research and other disciplines. Embedding videos effectively into these large models and harnessing their capabilities to enhance video understanding performance has emerged as pivotal tasks [48, 61]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.788, + 0.462 + ], + "angle": 0, + "content": "Previous research has identified several effective learning schemes for video representations, including reconstructing videos with masked inputs [28,37,85,90], aligning videos with languages [51,52,96,98], and predicting the next token using videos [1,49,83]. These approaches have turned out to be complementary and can be unified through a progressive training scheme. Notably, methods such as InternVideo [94], UMT [51], and VideoPrism [109] have utilized a two-stage training approach involving masked reconstruction and multimodal contrastive learning, leading to improved performance in downstream tasks. Following this line, we aim to further extend this progressive training scheme by incorporating video-based next token prediction and scaling the entire training process, including models and data, to build a new family of video foundation models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.734 + ], + "angle": 0, + "content": "The proposed video foundation model, coined as InternVideo2, is built through a progressive training scheme. The learning involves three stages: (1) capturing spatiotemporal structure via unmasked reconstruction, (2) aligning with semantics from other modalities, and (3) enhancing its open-ended dialogue power through next token prediction. In the initial stage, the model learns to reconstruct the unmasked video tokens, allowing the video encoder to develop basic spatiotemporal perception capability. To estimate the existing tokens, vision encoders (InternViT [17] and VideoMAE-g [90]) trained differently are employed as proxies. In the next stage of crossmodal learning, the architecture is expanded to include audio and text encoders. This not only improves the alignment between videos and text but also endows InternVideo2 the ability to handle video-audio tasks. By incorporating these additional modalities, the model's understanding of videos is enriched and aligned with their semantics. Finally, in the next-token prediction stage, a video-centric dialogue system and the corresponding instruction-finetuning dataset are built to further tune the InternVideo2. By connecting InternVideo2 to LLMs, the video encoder is further updated through next-token prediction training, enhancing its ability for open-ended tasks such as VQA and video description." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "For the training of InternVideo2, we emphasize the spatiotemporal consistency and labeling quality in the data. We build a large-scale multimodal video-centric dataset consisting of 402M data entries, which includes 2M videos, 50M video-text pairs (from WebVid [3] and InternVid [93]), 50M video-audio-speech-text pairs (InternVid2), and 300M image-text pairs. Specifically, for InternVid2, we segment videos into clips semantically and focus on recalibrating the clip descriptions using three modalities: audio, video, and speech. We first generate captions for" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InternVideo2" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "these three modalities separately. Then individual captions are fused together to create a more comprehensive description, which will improve the model's ability to comprehend and interpret the video accurately." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.788, + 0.344 + ], + "angle": 0, + "content": "We evaluate InternVideo2 across a wide range of video-related tasks. These tasks span from basic spatiotemporal perception, such as action recognition, to high-level reasoning tasks, such as long video or procedure-aware question-answering (QA), as given in Fig. 1. The results (in Sec. 5) demonstrate that InternVideo2 achieves the state-of-the-art performance on multiple tasks and is able to analyze and reason over sequences of actions. This top performance signifies its capability to effectively capture and understand video content. These empirical findings validate that InternVideo2 could serves as a general video encoder for future exploration in video understanding. In summary, our contributions to video understanding are as follows." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.351, + 0.785, + 0.41 + ], + "angle": 0, + "content": "- This paper introduces InternVideo2, a competitive family of video foundation models that leverages masked reconstruction, crossmodal contrastive learning, and next token prediction to make model perceptive, semantic, and capable of reasoning in video understanding." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.47 + ], + "angle": 0, + "content": "- InternVideo2 achieves the state-of-the-art performance for more than 60 video / audio tasks. Our model demonstrates superior performance in video-related dialogue and long video understanding, highlighting its potential in modeling high-level world knowledge." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.471, + 0.785, + 0.531 + ], + "angle": 0, + "content": "- We provide an enhanced dataset to train InternVideo2. This includes the validation and incorporation of audio data during training, as well as the improved captioning method. These improvements result in significant enhancements in model performance and generalization ability." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.351, + 0.785, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.553, + 0.388, + 0.569 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Video Foundation Models. Studies on learning video foundation models becomes increasingly crucial considering its wide applications [?, 51, 52, 85, 89, 90, 92, 94, 96, 98, 109]. Typical methods in building video foundation models (ViFM) include video-text contrastive learning [52, 94, 96], masked video modeling [30, 85, 90, 94], and next token prediction [1, 80, 83]. Specifically, All-in-one [89] utilized a single backbone with unified multiple pretraining objectives. On the other hand, UMT [51] combined masked modeling with video-text contrastive learning, demonstrating strong performance in both action recognition and videolanguage tasks. Another approach is mPLUG-2 [95], which introduced a new design for modulating different modalities. It shared a common module across modalities to enhance relations while incorporating modality-specific modules for discrimination. In addition to video-text pretraining, researchers have also explored the use of audio information in videos to improve performance. MERLOT Reserve [104] learned video representations using a large-scale dataset of videospeech-transcript pairs. VALOR [14] employed independent encoders for video, audio, and text and trains a joint visual-audio-text representation. VAST [15] constructed an audio-visual-speech dataset and develops a multimodal backbone" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.145, + 0.784, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.239, + 0.788, + 0.296 + ], + "angle": 0, + "content": "Fig. 2: Framework of InternVideo2. It consists of three consecutive training phases: unmasked video token reconstruction, multimodal contrastive learning, and next token prediction. In stage 1, the video encoder is trained from scratch, while in stages 2 and 3, it is initialized from the version used in the previous stage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.325, + 0.784, + 0.37 + ], + "angle": 0, + "content": "that excels in video-audio-related tasks. VideoPrism [109] combined video-text contrastive learning and video token reconstruction on a combination of public and proprietary videos, achieving leading results across various video tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.507 + ], + "angle": 0, + "content": "Multimodal Large Language Models. With advances in large language models (LLMs) [6, 24, 76], their multimodal versions (MLLMs) is becoming popular as it can handle open-world tasks. Seminal works like Flamingo [1] showed outstanding zero/few-shots performances over a range of multimodal tasks [36, 63, 72, 97]. Public MLLMs [33, 56, 113] such as LLaVA [56] and InstructBLIP [21] proposed to use visual instruction-tuning data to improve the visual dialogue ability. Some video-centric MLLMs have been proposed, such as VideoChat [48], VideoChatGPT [61] and Valley [60], by using instruction data to connect video encoders to LLMs for open-world video understanding." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.531, + 0.331, + 0.547 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.787, + 0.61 + ], + "angle": 0, + "content": "We learn InternVideo2 in three stages, illustrated in Fig. 2. The stages include spatiotemporal token reconstruction, video-audio-speech-language contrastive learning, and connecting to a large language model (LLM) for joint training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.788, + 0.747 + ], + "angle": 0, + "content": "Video Encoder. The video encoder used in InternVideo2 follows the Vision Transformer (ViT) [25] and includes additional projection layers for distillation. Inspired by previous works [17, 102], we introduce attention pooling to the ViT. For input videos, we sparsely sample 8 frames [91] and perform a \\(14 \\times 14\\) (\\(h \\times w\\)) spatial downsampling. These spatiotemporal tokens are then concatenated with a class token and combined with 3D position embeddings. The hyperparameters used for ViT-1B and -6B are the same as CoCa-1B [102] and InternViT-6B [17], respectively. The details of ViT-6B architecture are given in Supp." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.769, + 0.663, + 0.785 + ], + "angle": 0, + "content": "3.1 Stage1: Reconstructing Unmasked Video Tokens" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "We exploit two expert models to guide the video encoder to conduct token-level reconstruction for unmasked areas. Specifically, we adopt InternVL-6B [17] and VideoMAEv2-g [90] to transfer unmasked knowledge via simple projection layers." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InternVideo2" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "When training, we input the full videos into different teachers and mask out \\(80\\%\\) of the tokens frame by frame, under the semantic guidance of the multimodal model InternVL and motion-aware model VideoMAEv2. We only align the unmasked tokens, by minimizing their mean squared error (MSE) between student and teachers. The learning objective is to reconstruct the remaining tokens as:" + }, + { + "type": "equation", + "bbox": [ + 0.292, + 0.235, + 0.788, + 0.264 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\frac {1}{Z} \\sum \\left(\\alpha_ {1} | f ^ {V} (\\mathbf {V} _ {p}) - h (\\mathbf {V} _ {p}) | ^ {2} + \\alpha_ {2} | f ^ {V} (\\mathbf {V} _ {p}) - g (\\mathbf {V} _ {p}) | ^ {2}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.282, + 0.789, + 0.342 + ], + "angle": 0, + "content": "where \\( f^V \\), \\( h \\), and \\( g \\) are our video encoder, InternViT-6B [17], and ViT-g of VideoMAEv2, respectively. \\( p \\) stands for the token index and \\( f(\\mathbf{V}_p) \\) is the corresponding token extracted by InternVideo2 for input video \\( \\mathbf{V} \\). \\( Z \\) is the normalization factor. \\( \\alpha_{1} \\) and \\( \\alpha_{2} \\) balance the influence between the employed models." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.344, + 0.789, + 0.435 + ], + "angle": 0, + "content": "In implementation, we align the last 6 layers of InternVL, last 4 layers of VideoMAEv2, and the final output token of InternVL. Different losses are simply summed for optimization. After pretraining, we drop those projection layers and only use the basic encoder. Compared with only using the multimodal model in UMT and VideoPrism, our strategy makes the vision encoder multimodal-friendly as well as enhances its temporal sensitivity for action modeling." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.46, + 0.644, + 0.476 + ], + "angle": 0, + "content": "3.2 Stage 2: Aligning Video to Audio-Speech-Text" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.488, + 0.789, + 0.653 + ], + "angle": 0, + "content": "We exploit the correspondence between video and audio, speech, and text to encourage InternVideo2 to learn more semantics. In practice, InternVideo2 has a huge video encoder, and its employed audio and text encoders are relatively lightweight. The used audio encoder is a 12-layer transformer initialized with BEATs [13] (90M). It takes in audio features, which are 64-dimensional log Mel filterbank spectrograms using a 25ms Hamming window, transformed from 10-second-long clips, padding with zeros. For the text and speech encoders, we initialize the text encoder and multimodal decoder using BERT-Large [24]. Specifically, we utilize the initial 19 layers of BERT-Large as the text encoder, with the subsequent 5 layers equipped with cross-attention layers serving as the multimodal decoder." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.655, + 0.788, + 0.7 + ], + "angle": 0, + "content": "For pretraining objectives, we establish alignment across different modalities via text, including video, audio, image, and speech. We employ crossmodal contrastive and matching losses with masked language reconstruction loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.715, + 0.787, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {C O N}} + \\mathcal {L} _ {\\mathrm {M A C}} + \\mathcal {L} _ {\\mathrm {M L M}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.742, + 0.788, + 0.773 + ], + "angle": 0, + "content": "The employed \\(\\mathcal{L}_{\\mathrm{MAC}}\\) and \\(\\mathcal{L}_{\\mathrm{MLM}}\\) are standard loss from [19]. Specifically, the crossmodal contrastive learning is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.788, + 0.784, + 0.824 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C O N}} = \\sum_ {M, T _ {M ^ {\\prime}}} \\mathcal {L} _ {\\mathrm {C O N}} (M, T _ {M ^ {\\prime}}) = - \\sum_ {M, T _ {M ^ {\\prime}}} \\left(\\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (\\sin \\left(f _ {i} ^ {M} , f _ {i} ^ {T _ {M ^ {\\prime}}}\\right) / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (\\sin \\left(f _ {i} ^ {M} , f _ {j} ^ {T _ {M ^ {\\prime}}}\\right) / \\tau)} + \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (\\sin \\left(f _ {i} ^ {T _ {M ^ {\\prime}}} , f _ {i} ^ {M}\\right) / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (\\sin \\left(f _ {i} ^ {T _ {M ^ {\\prime}}} , f _ {j} ^ {M}\\right) / \\tau)}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.825, + 0.788, + 0.842 + ], + "angle": 0, + "content": "where \\( f^V \\) and \\( f^T \\) denote the learned video and text embeddings, respectively." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.192 + ], + "angle": 0, + "content": "\\(M\\) and \\(T_{M^{\\prime}}\\) indicates the modality of input signals and the text descriptions describing, respectively. \\(\\sin (\\cdot)\\) computes the cosine similarity between two features. \\(\\tau\\) is the learnable temperature." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.193, + 0.509, + 0.208 + ], + "angle": 0, + "content": "For the matching part, it is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.217, + 0.786, + 0.233 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M A C}} = - y \\log f _ {p} (\\mathbf {V}, \\mathbf {T}) - (1 - y) \\log \\left(1 - f _ {p} (\\mathbf {V}, \\mathbf {T})\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.786, + 0.269 + ], + "angle": 0, + "content": "where \\( f_{p}(\\mathbf{V},\\mathbf{T}) \\) computes the matching likelihood between \\( \\mathbf{V} \\) and \\( \\mathbf{T} \\). \\( y \\) denotes whether the given video and text are paired \\( (y = 1) \\) or not \\( (y = 0) \\)." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.271, + 0.594, + 0.285 + ], + "angle": 0, + "content": "The employed masked language modeling loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.293, + 0.787, + 0.311 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M L M}} = - \\log f _ {p} ^ {T} \\left(\\mathbf {T} _ {j} \\mid \\mathbf {T} _ {< j}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.318, + 0.786, + 0.349 + ], + "angle": 0, + "content": "where \\( f_{p}^{T}(\\mathbf{T}_{j}|\\mathbf{T}_{< j}) \\) computes the likelihood of the \\( j_{\\mathrm{th}} \\) text token based on the previous ones. Here \\( \\mathbf{T} \\) refers to video captions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.787, + 0.408 + ], + "angle": 0, + "content": "To improve the training efficiency, we employ the masked learning strategy, aligning unmasked video tokens to tokens from other modalities first, then using full video tokens reconstruction shortly. Specifically, it consists of two steps as follows:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.788, + 0.497 + ], + "angle": 0, + "content": "Aligning Masked Visual-Language-Audio. We freeze the audio encoder and focus on aligning visual, audio, and text features. For pre-training, we use a comprehensive set of image, video, and audio-video data. The combinations of modalities used are represented as \\(\\{M,T_{M^{\\prime}}\\} \\in \\{\\{I,T_I\\},\\{V,T_V\\},\\{V,T_{VAS}\\},\\{VA,T_{VAS}\\}\\}\\) where each pair denotes the concatenated features from the respective modalities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Unmasked Visual-Audio-Language Post-Pretraining. We freeze the vision encoder to jointly align audio, visual, and text features. Post-pretraining is conducted using a smaller subset of image and video data (25M samples), along with the full set of audio (0.5M samples) and audio-video data (50M samples). Since the parameters of the largest ViT-6B model are frozen, we do not use masking strategies in this phase to ensure consistency with the inference process and to minimize any performance degradation in downstream tasks. The modality combinations used here are \\(\\{M, T_{M'}\\} \\in \\{\\{I, T_I\\}, \\{V, T_V\\}, \\{A, T_A\\}, \\{V, T_{VAS}\\}, \\{VA, T_{VA}\\}\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.649, + 0.741, + 0.665 + ], + "angle": 0, + "content": "3.3 Stage3: Predicting Next Token with Video-Centric Inputs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.762 + ], + "angle": 0, + "content": "To further enrich the semantics embedded in InternVideo2 and improve its support for video-centric dialogue, we tune it by connecting it to a LLM with QFormer design [46, 47]. We employ the progressive learning scheme in [49] by using InternVideo2 as the video encoder and train a video blip for communicating with open-sourced LLM [40, 111]. In training, the video encoder will be updated. Detailed training receipts are the same with VideoChat2 [49]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.782, + 0.486, + 0.798 + ], + "angle": 0, + "content": "4 Multimodal Video Data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We list our training data in Tab. 1. Among the datasets used, KMash and InternVid2 are newly built, whereas the rest are public available." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InternVideo2" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "table_caption", + "bbox": [ + 0.258, + 0.145, + 0.744, + 0.16 + ], + "angle": 0, + "content": "Table 1: Summary of datasets used in InternVideo2 pretraining process." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.171, + 0.725, + 0.282 + ], + "angle": 0, + "content": "
Pretraining Stage DatasetDomain# of clipsAnnotation
Stage 1KMashWeb Video2M-
Stage 2 (img-text)LAION, etcWeb Image300MAlt-text / Generated Caps
Stage 2 (vid-text)WebVid2MWeb Video250kAlt-text
WebVid10MWeb Video9.7MAlt-text
InternVidYoutube Video40MGenerated Caption
InternVid2Youtube Video50MGenerated Caption
Stage 3LLaVA, etcWeb Image/Video2.1MConversation, QA
" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.293, + 0.761, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.46, + 0.788, + 0.503 + ], + "angle": 0, + "content": "Fig. 3: The framework of our video multimodal annotation system, called VidCap, consists of four main components: video, audio, and speech captioners, along with a LLM for integrating captions from these modalities." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.531, + 0.616, + 0.546 + ], + "angle": 0, + "content": "4.1 Video-only Data for Masked Autoencoders" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.788, + 0.631 + ], + "angle": 0, + "content": "We curate a new video set without labels named \\( K \\)-Mash from action recognition datasets [10, 35, 38, 65, 108], as detailed in Supp. It encompasses a wide range of video types, including first- and third-person perspectives, with both short and long duration, and featuring various settings. Further, we give \\( K \\)-\\( \\text{Mash}_{2M} \\) with additionally sourced and selected 844K videos from YouTube for diversity." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.652, + 0.573, + 0.667 + ], + "angle": 0, + "content": "4.2 Videos with Audio-Speech Modalities" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.797 + ], + "angle": 0, + "content": "We build a multimodal video dataset, coined as InternVid2, with video-audiospeech information and their descriptions for strengthening video perception via other modalities. It consists of 100M videos along with their VAS captions. In InternVid2, we collect videos from several sources (detailed in the supp), segment them into clips, and automatically annotate them based on their unimodal or crossmodal inputs. We highlight the importance of temporal segmentation in clip generation and our video multimodal annotation system. We find refining them leads to notable downstream improvements." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Temporal Consistency Matters. We employ a temporal boundary detection model AutoShot [114] to segment videos into clips instead of SceneDet filter from" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.193 + ], + "angle": 0, + "content": "FFmpeg. It predicts boundaries based on temporal semantic variations rather than pixel differences, and is able to generate semantically complete cuts without mixing extra frames with inconsistent context." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.21, + 0.789, + 0.392 + ], + "angle": 0, + "content": "Video Multimodal Annotation. We design a video multimodal annotation system VidCap to give proper unimodal and crossmodal descriptions for textualizing videos from different perceptions. It automatically captions visual, audio, and speech of InternVid2, then it corrects them and fuses them for cross-modal captions via LLM. The system frame is given in Fig. 3. VidCap has independent video, audio, speech captioner, and a LLM for caption refinement and fusion. For video, speech, and caption postprocessing, we employ existing methods as the video captioning pipeline in [93], WhisperV2-large model [75], and Vicuna-1.5 [111]. For audio, we craft a audio captioner upon VideoChat [48], as we find no open-sourced one. It extracts audio features from inputs by Beats [13]. We learn it by only tuning its Qformer using a combination of the large-scale audio-text corpus WavCaps [64] dataset. Details are given in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.418, + 0.377, + 0.436 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.453, + 0.787, + 0.514 + ], + "angle": 0, + "content": "In our evaluation of InternVideo2, we assess the models from three learning stages. The evaluation covers a wide range of tasks, including video recognition, video retrieval, question-answering, and more. It includes various scenarios such as zero-shot learning, finetuning, and linear probing." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.514, + 0.804, + 0.574 + ], + "angle": 0, + "content": "For InternVideo2 trained in stage 1, 2, and 3, we denote them with InternVideo2\\(_{s1}\\), InternVideo2\\(_{s2}\\), and InternVideo2\\(_{s3}\\), respectively. We also learn a CLIP-style InternVideo2 indicated by InternVideo2\\(_{clip}\\). It is post-pretrained from InternVideo2\\(_{s2}\\) by only preserving video and text encoders and contrastive loss." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.575, + 0.787, + 0.681 + ], + "angle": 0, + "content": "Each training stage of InternVideo2-6B uses different configurations and resources. In the first stage, we employ 256 NVIDIA A100 GPUs and train the model for 18 days. The second stage also utilizes 256 A100 GPUs and spans a training period of 14 days. Finally, in the third stage, we use 64 A100 GPUs and train the model for 3 days. We introduce DeepSpeed and FlashAttention [22] for training and inference. More implementation details, experiment results, and ablations are given in the supp." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.706, + 0.427, + 0.721 + ], + "angle": 0, + "content": "5.1 Video Classification" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.734, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Action Recognition. We test InternVideo2 on Kinetics (i.e., K400, 600 and 700 [8-10]), Moments in Time V1 (MiT) [65], Something-Something V2 (SSv2) [35], UCF [79], HMDB [43], Charades [32], ActivityNet [38] (ANet) and HACS [108]. We evaluate in four settings: (a) end-to-end finetuning the whole backbone; (b) attentive probing is similar to linear pooling, but extra trains the attention pooling layer [102]. (c) linear probing which freezes the backbone and only trains the task head; and (d) zero-shot." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InternVideo2" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.787, + 0.186 + ], + "angle": 0, + "content": "Table 2: End-to-end finetuning action recognition results (top-1 accuracy) on Kinetics, SomethingSomething, and Moments in Time. \\( {}^{ \\dagger } \\) denotes that the result is achieved with different resolutions or frame rates." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.199, + 0.78, + 0.339 + ], + "angle": 0, + "content": "
MethodTraining DataSettingK400K600K700Sth-Sthv2MiTANetHACS
CoVeR [106]IV-3B16 × 44887.187.979.870.846.1--
Hiera-H [78]V-0.25M16 × 22487.888.881.176.5---
CoCa-g [102]I-3B16 × 57688.989.482.7-49.0--
MTV-H [90]IV-370M32 × 28089.990.383.4----
VideoMAEv2-g [90]V-1.35M64 × 26690.089.9-77.0†---
V-JEPA-H [4]V-2M16 × 224---77.0---
MVD-H [92]IV-1.25M16 × 224---77.3---
UniFormerV2-L [50]IV-401M64 × 33690.090.182.773.0†47.8†94.795.4
InternVideo [94]V-12Mensemble91.191.384.077.2---
InternVideo2s1-1BIV-1.1M8 × 22491.391.485.077.150.8--
InternVideo2s1-1BIV-1.1M16 × 22491.691.685.477.150.9--
InternVideo2s1-6BIV-2M8 × 22491.991.785.777.551.0--
InternVideo2s1-6BIV-2M16 × 22492.191.985.977.451.295.997.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.355, + 0.787, + 0.384 + ], + "angle": 0, + "content": "Table 3: Attentive probing recognition results (top-1 accuracy) on Kinetics-400/600/700, Moments in Time and Something-Something V2." + }, + { + "type": "table", + "bbox": [ + 0.25, + 0.395, + 0.749, + 0.521 + ], + "angle": 0, + "content": "
MethodTraining DataSettingK400K600K700MiTSSV2
UMT-L [51]IV-25M-82.8--40.354.5
VideoMAEv2-g [90]V-1.35M-82.1--35.056.1
V-JEPA-H [4]V-2M16 × 38481.9---72.2
DINOv2-g [69]I-142M16 × 22483.4---50.0
VideoPrism-g [109]V-619M16 × 28887.2--45.568.5
ViT-e [23]I-4B128 × 22486.5--43.6-
ViT-22B [23]I-4B128 × 22488.0--44.9-
CoCa-g [102]I-3B16 × 57688.088.581.147.4-
InternVideo2s2-1BIV-25.5M16 × 22487.988.079.546.367.3
InternVideo2s2-6BIV-400M16 × 22488.889.181.047.867.7
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.55, + 0.788, + 0.746 + ], + "angle": 0, + "content": "End-to-end Finetuning. Tab. 2 shows InternVideo2-6B obtains new state-of-the-art (SOTA) results on Kinetics \\((92.1\\% / 91.9\\% / 85.9\\%)\\) on K400/600/700, respectively), SthSthv2, MiT, ANet, and HACS with only 16 frames, while the previous SOTAs require larger resolution (224 vs. 576) or model ensemble. As for MiT in Table 2, InternVideo2-6B exceeds the previous SOTA, CoCa-g, by a significant margin of \\(2.2\\%\\) \\((51.2\\%\\) vs. \\(49.0\\%)\\). Regarding the temporal-related actions in Table 2, our InternVideo2-6B also surpasses MVD [92] on SSv2 \\((77.5\\%\\) vs. \\(77.3\\%)\\). Moreover, our InternVideo2-6B showcases top performance on untrimmed video analysis, as indicated in Table 2, with \\(95.9\\%\\) on ActivityNet and \\(97.0\\%\\) on HACS. These results affirm our model's superior capability for robustly identifying complex actions across varied scenes. Note \"I\" and \"V\" denotes images and videos, respectively. \"IV-3B\" means the total number of the used images and videos is 3B, while \"I-3B\" means using 3B images." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Attentive Probing. As in Tab. 3, InternVideo2-6B not only outperforms ViT-22B [23] and CoCa-g [102] in scene-focused datasets but also surpasses or matches the performance of the latest video foundation model [4,109], on datasets emphasizing temporal dynamics (SthSthV2). This underscores our model's exceptional ability to understand and interpret both spatial and temporal information effectively." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.173 + ], + "angle": 0, + "content": "Table 4: Linear probing action recognition results (top-1 accuracy) on Kinetics-400, Something-Something V2, UCF-101 and HMDB-51." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.185, + 0.726, + 0.29 + ], + "angle": 0, + "content": "
MethodSettingK400SSV2UCF-101HMDB-51
VideoMAEv2-H [90]12 × 22425.8-56.434.1
TVTSv2-H [105]12 × 22473.1-91.865.7
OpenCLIP-G [20]8 × 22478.335.890.7-
DINOv2-g [69]8 × 22478.438.391.2-
InternVideo2s1-1B16 × 22481.646.396.071.6
InternVideo2s1-6B16 × 22482.047.896.371.8
InternVideo2s2-6B16 × 22484.256.797.380.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.304, + 0.784, + 0.318 + ], + "angle": 0, + "content": "Table 5: Zero-shot action recognition on UCF, HMDB, MiTv1, SSv2-MC, and Charades." + }, + { + "type": "table", + "bbox": [ + 0.283, + 0.331, + 0.712, + 0.424 + ], + "angle": 0, + "content": "
Method#FTraining DataUCFHMDBMiTSSv2-MCCharades
CLIP [74]12I-400M68.943.2-29.6-
TVTSv2 [105]12V-8.5M78.052.1-48.4-
VideoCoCa-g [98]16V-145M----25.8
VideoPrism-g [109]16V-619M----32.4
InternVideo2clip-1B8IV-25.5M88.853.931.661.532.9
InternVideo2clip-6B8IV-400M89.556.732.963.534.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.307, + 0.439, + 0.694, + 0.453 + ], + "angle": 0, + "content": "Table 6: Zero-shot action recognition results on Kinetics." + }, + { + "type": "table", + "bbox": [ + 0.336, + 0.465, + 0.663, + 0.61 + ], + "angle": 0, + "content": "
Method#FK400K600K700
top-1AVGtop-1AVGtop-1AVG
CLIP [74]858.470.155.167.246.158.4
EVA-CLIP-L [81]1-65.0-64.9-59.1
EVA-CLIP-E [81]1-69.8-69.3-63.4
ViCLIP-L [93]864.875.762.273.554.366.4
VideoCoCa-g [98]1672.081.3-
InternVL-6B [17]869.179.468.978.860.671.5
EVA-CLIP-18B [82]1679.4-79.4-72.2
VideoPrism-g [109]1676.485.4-
InternVideo2clip-1B873.182.472.881.864.975.2
InternVideo2clip-6B872.782.271.781.264.275.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.788, + 0.731 + ], + "angle": 0, + "content": "Linear Probing. In Tab. 4, InternVideo2-1B significantly outperforms the previous SOTA, DINOv2-g [69], by notable margins: \\(+3.2\\%\\) on K400, \\(+8.0\\%\\) on SthSthV2, and \\(+4.8\\%\\) on UCF-101. As we scale the model, an upward trend in results is observed, underscoring the benefits of model enhancement. Notably, the integration of multimodal pretraining (stage 2) yields further rise in results. We suppose stage 2 enhances feature discrimination." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Zero-shot. Table 5 and 6 show InternVideo2 gets \\(72.7\\%\\) / \\(71.7\\%\\) / \\(64.2\\%\\) on K400/600/700, respectively, outperforming others but VideoPrism on K400 \\((76.4\\%)\\). On UCF [79], HMDB [43], MiT [65], SSv2-MC, and Charades, InternVideo2 gives an cutting edges over others. The clear gap between VideoPrism and InternVideo2 on K400 may signify the importance of pretraining corpus in VideoPrism (311M videos with text and 36.1M of them are manually labeled) for" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InternVideo2" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.144, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 7: Finetuned temporal action localization results on THUMOS14 [39], ActivityNet [42], HACS Segment [108] and FineAction [57]. We report average mAP. \"Flow\" denotes the ensembling I3D flow feature. * denotes the result is achieved with Flow." + }, + { + "type": "table", + "bbox": [ + 0.309, + 0.199, + 0.696, + 0.3 + ], + "angle": 0, + "content": "
BackboneTHUMOS14HACSActivityNetFineAction
I3D [10] + Flow66.8-35.6-
R(2+1)D [88]55.6-36.6-
InternVideo71.6*41.339.017.6
VideoMAEv2-g [90]69.5--18.2
InternVideo2s1-1B69.842.440.427.2
InternVideo2s1-6B72.043.341.227.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.221, + 0.311, + 0.779, + 0.326 + ], + "angle": 0, + "content": "Table 8: Video instance segmentation performance (mAP) on YouTube-VIS19 [99]." + }, + { + "type": "table", + "bbox": [ + 0.331, + 0.347, + 0.668, + 0.414 + ], + "angle": 0, + "content": "
MethodBackbone#ParamsYouTubeVIS19
Mask2FormerSwin-L (image) [58]219M60.3
Mask2FormerInternViT (image)6B63.4
Mask2FormerInternVideo2s16B64.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.514 + ], + "angle": 0, + "content": "K400 in zero-shot. Note that on the datasets of Kinetics, UCF101 and HMDB51, which have a distribution closer to the pre-training dataset used in stage1, the performance of Intervvideo2-6B is slightly inferior to that of Intervvideo2-1B. We suppose this is caused by Intervvideo2-6B uses a more abundant pretraining dataset in stage2, leading to the forgetting of pretraining data in stage1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.789, + 0.73 + ], + "angle": 0, + "content": "Temporal Action Localization. We evaluate models on four temporal action localization (TAL) datasets: THUMOS14 [39], ActivityNet [42], HACS Segment [108] and FineAction [57] in a feature-based manner with finetuning. We employ output of the 7-th layer from InternVideo2 for inputs as the corresponding features without fusing anything else. ActionFormer [107] is used as the detection head. We report mean Average Precision(mAP) under multiple tIoU as in [12,54,100,107]. In Table 7, InternVideo2-6B gets the highest mAP among all comparisons in all datasets, while InternVideo2-1B nearly surpass other methods except in THUMOS14. We find InternVideo2-6B almost consistently improves mAP with a notable margin from InternVideo2-1B except in FineAction. We suppose scaling model capacity without data refinement cannot nontrivially improve fine-grained discrimination abilities of models. Scaling detailed annotations in training may address this issue." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Video Instance Segmentation. We evaluate on the Video Instance Segmentation (VIS) dataset Youtube-VIS 2019 [99]. Built upon Mask2Former [18], we employ the video encoder of InternVideo2 as backbone with ViT-adapter [16] for features. We also try InternViT [17] for comparisons. In Table 8, InternVideo2 gets the highest mAP among all. This validates its effectiveness in relatively fine-grained spatiotemporal perception." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 9: Results of zero-shot video retrieval in both text-to-video (T2V) and video-to-text (V2T) on MSR-VTT, LSMDC, DiDeMo, MSVD, ANet, and VATEX." + }, + { + "type": "table", + "bbox": [ + 0.25, + 0.185, + 0.749, + 0.315 + ], + "angle": 0, + "content": "
MethodMSR-VTTLSMDCDiDeMoMSVDANetVATEX
T2VV2TT2VV2TT2VV2TT2VV2TT2VV2TT2VV2T
CLIP [74]30.424.213.911.912.718.740.557.29.113.2--
CLIP4Clip [59]32.0-15.1---38.5-----
ViCLIP [93]42.441.320.116.918.427.949.175.115.124.0--
InternVideo-L [94]40.739.617.613.231.533.543.467.630.731.449.569.5
UMT-L [51]40.737.124.921.948.649.949.074.541.939.4--
VideoCoCa-g [98]34.464.7------34.533.053.273.6
VideoPrism-g [109]39.771.0------52.750.362.577.1
InternVideo2s2-1B51.950.932.027.357.054.358.183.360.454.870.485.4
InternVideo2s2-6B55.953.733.830.157.957.159.383.163.256.571.585.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.327, + 0.788, + 0.356 + ], + "angle": 0, + "content": "Table 10: Results of finetuning video retrieval in both text-to-video (T2V) and video-to-text (V2T) on MSR-VTT, LSMDC, DiDeMo, MSVD, ANet, and VATEX." + }, + { + "type": "table", + "bbox": [ + 0.25, + 0.368, + 0.753, + 0.459 + ], + "angle": 0, + "content": "
MethodMSR-VTTLSMDCDiDeMoMSVDANetVATEX
T2VV2TT2VV2TT2VV2TT2VV2TT2VV2TT2VV2T
CLIP [74]38.238.722.522.632.233.9--26.126.9--
CLIP4Clip [59]45.645.924.323.843.043.645.248.440.341.6--
ViCLIP [93]52.551.833.032.549.450.2--49.848.1--
UMT-L [51]58.858.643.041.470.465.758.282.466.864.472.086.0
InternVideo2s2-6B62.860.246.446.774.271.961.485.274.169.775.589.3
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.485, + 0.503, + 0.501 + ], + "angle": 0, + "content": "5.2 Video-Audio-Language Tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.571 + ], + "angle": 0, + "content": "We evaluate InternVideo2 on video retrieval, captioning, and multi-choice question-answering (QA). The former two tasks are conducted by matching video representation and the candidate text ones using the text encoder in stage 2. The latter is tested by the VideoLLM learned in stage 3. We also test audio tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.652 + ], + "angle": 0, + "content": "Video Retrieval. We evaluate the video retrieval on six popular benchmarks [2, 11,38,77,97], as shown in Tab. 9 and 10. In eval, eight frames from the input videos are uniformly sampled. We report \\( R@1 \\) scores for both text-to-video (t2v) and video-to-text (v2t) tasks in Tab. 9 and 10. R@5 and R@10 are given in Supp." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.653, + 0.788, + 0.728 + ], + "angle": 0, + "content": "Table 9 and 10 demonstrate that InternVideo2 outperforms other state-of-the-arts with a notable margin in both t2v and v2t of all used datasets no matter in zero-shot or finetuned settings, except for the v2t of MSR-VTT, where VideoPrism gives the best result. This shows the video-language semantic alignment of transferrity of InternVideo2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Video Temporal Grounding. We evaluate InternVideo2 on two video temporal grounding (VTG) datasets: Qvhighlight [44], and Charade-STA [32]. The eval setting and used features are the same as in TAL. We use CG-DETR [66] as the grounding head. We report R1@0.3, R1@0.5, R1@0.7, and mAP for moment retrieval as in [44, 53, 66]. Highlight Detection is evaluated in terms of \"Very Good\" mAP and HiT@1. In Table 11, InternVideo2-1B and InternVideo2-6B bring" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InternVideo2" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.144, + 0.784, + 0.16 + ], + "angle": 0, + "content": "Table 11: Finetuned temporal grounding on QVHighlight [44] and Charade-STA [32]." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.171, + 0.508, + 0.231 + ], + "angle": 0, + "content": "
FeatureR1@0.5R1@0.7mAPmAPHiT@1
CLIP [74]64.9748.6542.9639.8364.19
CLIP+SlowFast [29]65.4348.3842.8640.3366.21
InternVideo2s2-1B70.0054.4547.0242.3669.74
InternVideo2s2-6B71.4256.4549.2442.9072.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.319, + 0.233, + 0.405, + 0.245 + ], + "angle": 0, + "content": "(a) QVHighlight" + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.171, + 0.772, + 0.231 + ], + "angle": 0, + "content": "
FeatureR1@0.3R1@0.5R1@0.7mIoU
CLIP [74]65.6252.7730.1645.85
CLIP+SlowFast [29]70.4358.4436.3450.13
InternVideo2s2-1B78.4168.3645.0357.12
InternVideo2s2-6B79.7070.0348.9558.79
" + }, + { + "type": "table_caption", + "bbox": [ + 0.595, + 0.233, + 0.688, + 0.242 + ], + "angle": 0, + "content": "(b) Charade-STA" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.27, + 0.785, + 0.299 + ], + "angle": 0, + "content": "Table 12: Audio retrieval results on AudioCaps [41], Clothov1, and Clothov2 [27]. We report text-to-audio R@1 accuracy in zero-shot and finetuning settings." + }, + { + "type": "table", + "bbox": [ + 0.251, + 0.311, + 0.753, + 0.391 + ], + "angle": 0, + "content": "
MethodZero-shotFinetuning
AudioCapsClothoV1ClothoV2AudioCapsClothoV1ClothoV2
VIP-ANT [110]27.7---
VAST [15]---52.025.126.9
LanguageBind [112]-12.112.1---
InternVideo2s2-6B37.117.417.455.225.327.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.404, + 0.787, + 0.434 + ], + "angle": 0, + "content": "Table 13: Results of AudioQA on ClothoAQA [55] and Audio-MusicAVQA (AMAVQA) [5], and audio classification on the ESC-50 [71]. Both in the finetuning setting." + }, + { + "type": "table", + "bbox": [ + 0.253, + 0.445, + 0.465, + 0.504 + ], + "angle": 0, + "content": "
BackboneClothQAAMAVQA
AquaNet [55]14.7865.59
MWAFM [45]22.2467.54
InternVideo2s230.1480.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.272, + 0.505, + 0.446, + 0.517 + ], + "angle": 0, + "content": "(a) ClothoAQA and AMAVQA" + }, + { + "type": "table", + "bbox": [ + 0.569, + 0.445, + 0.724, + 0.504 + ], + "angle": 0, + "content": "
MethodTop-1 Acc
AST [34]95.60
BEATS [13]98.10
InternVideo2s298.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.612, + 0.505, + 0.679, + 0.516 + ], + "angle": 0, + "content": "(b) ESC-50" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.558, + 0.787, + 0.604 + ], + "angle": 0, + "content": "gradual performance improvements compared to CLIP [73] and CLIP [73] + Slow-fast [29]. This suggests that a larger spatiotemporal model is more beneficial to short-term video semantic alignment capabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Audio-related Tasks. We evaluate InternVideo2's audio and text encoders on audio tasks, including audio-text retrieval on AudioCaps [41], Clothov1, and Clothov2 [27]; audioQA on ClothoAQA [55] and Audio-MusicAVQA [5]; and audio classification on the ESC-50 [71]. As shown in Tab. 12, 13a, and 13b, our model achieves state-of-the-art performance on all downstream tasks. Considering the limited size of the used audio and text encoders, these audio-related results show crossmodal contrastive learning's benefits are mutual to the used modalities. Audio and the corresponding text models also gain from this learning." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.77, + 0.622, + 0.787 + ], + "angle": 0, + "content": "5.3 Video-centric Dialogue and its Applications" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Tab. 14 shows results on MVBench [49], Egoschema [62], and Perception Test [70], by equipping VideoChat2 [49] with our InternVideo2 encoder. VideoChat2-InternVideo2 demonstrates a significant performance advantage over other systems" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 14: Results of Chat-centric Evaluation on Multiple Choice Video-QA on MVBench [49], Egoschema and Perception Test." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.185, + 0.781, + 0.306 + ], + "angle": 0, + "content": "
ModelViEncoderLLMMVBenchEgoschemaPerception Test
GPT-4V [68]-GPT-443.5--
Gemini 1.0 Pro [84]--37.755.751.1
Gemini 1.0 Ultra [84]---61.554.7
Gemini 1.5 Pro [84]---72.2-
LLaVA-Next-VideoCLIP-LVicuna-7B46.543.948.8
VideoLLaMA2CLIP-L-336Mistral-7B54.651.751.4
VideoLLaMA2CLIP-L-336Mistral-8*7B53.953.352.2
VideoChat2UMT-LVicuna-7B51.1--
VideoChat2InternVideo2s3-1BMistral-7B60.355.853.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.334, + 0.79, + 0.502 + ], + "angle": 0, + "content": "in MVBench. In Egoschema and Perception Test evaluations, it stands out as the top performer among open-source models. When contrasted with proprietary commercial models like Gemini, VideoChat2-InternVideo2 exhibits results that are slightly lower but still comparable in the Perception Test. Egoschema requires longer context harnessing, and we plan to further explore leveraging LLM for long multimodal sequence modeling. Considering these benchmarks not only involve perception but also reasoning, these results suggest InternVideo2 does embed knowledge to partially model the world. It also validates the importance of learning a transferrable video representation for current video-related MLLM. We also give several qualitative evaluations with popular GPT-4V and GeminiPro (in supp)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.523, + 0.51, + 0.54 + ], + "angle": 0, + "content": "6 Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.79, + 0.706 + ], + "angle": 0, + "content": "We have introduced a new family of video foundation models called InternVideo2, which achieves the state-of-the-art performance across various video and audio tasks. In InternVideo2, we combine masked video modeling, video-audio-text contrastive learning, and next token prediction into a unified framework. Additionally, we create a new video-text dataset that incorporates video-audio-speech fused captions as descriptions. The dataset contains temporally segmented clips with high semantic coherence. These designs in InternVideo2 contribute to enhancing video understanding in both perception and reasoning tasks. Notably, InternVideo2 excels in video-related dialogue and long video understanding, demonstrating its effectiveness in capturing high-level semantics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Limitations and Discussions. Despite its achievements, InternVideo2 does not introduce specific novel architectural design. Instead, it leverages the existing learning techniques for scaling video foundation models while focusing on improving data processing to enhance its spatiotemporal perception, semantic alignment, and basic knowledge embedding. Similarly to previous studies [51, 101], InternVideo2 still grapples with limitations stemming from fixed input resolutions, sampling rates, and highly compressed tokens, which restrict its ability to express rich video information and capture fine-grained details." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InternVideo2" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.237 + ], + "angle": 0, + "content": "This work is partially supported by the National Key R&D Program of China (No. 2022ZD0160102), National Natural Science Foundation of China (No. 62076119, No. 61921006)), and the Science and Technology Commission of Shanghai Municipality under Grant No. 23QD1400800 and No. 23YF1461900." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.259, + 0.323, + 0.275 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.288, + 0.787, + 0.373 + ], + "angle": 0, + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: a visual language model for few-shot learning. ArXiv abs/2204.14198 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.373, + 0.787, + 0.414 + ], + "angle": 0, + "content": "2. Anne Hendricks, L., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.: Localizing moments in video with natural language. In: ICCV. pp. 5803-5812 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.414, + 0.787, + 0.441 + ], + "angle": 0, + "content": "3. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.441, + 0.787, + 0.482 + ], + "angle": 0, + "content": "4. Bardes, A., Garrido, Q., Ponce, J., Chen, X., Rabbat, M., LeCun, Y., Assran, M., Ballas, N.: V-JEPA: Latent video prediction for visual representation learning (2024), https://openreview.net/forum?id=WFYbBOEOtv" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.482, + 0.787, + 0.524 + ], + "angle": 0, + "content": "5. Behera, S.R., Injeti, K.M., Patibandla, J.S.K., Pokala, P.K., Pailla, B.R.: Aquallm: Audio question answering data generation using large language models. arXiv preprint arXiv:2312.17343 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.524, + 0.787, + 0.564 + ], + "angle": 0, + "content": "6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.565, + 0.787, + 0.606 + ], + "angle": 0, + "content": "7. Bruce, J., Dennis, M., Edwards, A., Parker-Holder, J., Shi, Y., Hughes, E., Lai, M., Mavalankar, A., Steigerwald, R., Apps, C., et al.: Genie: Generative interactive environments. arXiv preprint arXiv:2402.15391 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.606, + 0.787, + 0.634 + ], + "angle": 0, + "content": "8. Carreira, J., Noland, E., Banki-Horvath, A., Hillier, C., Zisserman, A.: A short note about kinetics-600. ArXiv abs/1808.01340 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.634, + 0.787, + 0.662 + ], + "angle": 0, + "content": "9. Carreira, J., Noland, E., Hillier, C., Zisserman, A.: A short note on the kinetics-700 human action dataset. ArXiv abs/1907.06987 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.662, + 0.787, + 0.689 + ], + "angle": 0, + "content": "10. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.787, + 0.744 + ], + "angle": 0, + "content": "11. Chen, D.L., Dolan, W.B.: Collecting highly parallel data for paraphrase evaluation. In: Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies-Volume 1. pp. 190-200. Association for Computational Linguistics (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.787, + 0.785 + ], + "angle": 0, + "content": "12. Chen, G., Zheng, Y.D., Wang, L., Lu, T.: Dscan: improving temporal action detection via dual context aggregation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 248-257 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "13. Chen, S., Wu, Y., Wang, C., Liu, S., Tompkins, D., Chen, Z., Che, W., Yu, X., Wei, F.: BEATs: Audio pre-training with acoustic tokenizers. In: Proceedings of the 40th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 202, pp. 5178-5193. PMLR (23-29 Jul 2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.288, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "14. Chen, S., He, X., Guo, L., Zhu, X., Wang, W., Tang, J., Liu, J.: Valor: Vision-audio-language omni-perception pretraining model and dataset. arXiv preprint arXiv:2304.08345 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.19, + 0.787, + 0.231 + ], + "angle": 0, + "content": "15. Chen, S., Li, H., Wang, Q., Zhao, Z., Sun, M., Zhu, X., Liu, J.: Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.231, + 0.787, + 0.259 + ], + "angle": 0, + "content": "16. Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.259, + 0.787, + 0.314 + ], + "angle": 0, + "content": "17. Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Zhong, M., Zhang, Q., Zhu, X., Lu, L., Li, B., Luo, P., Lu, T., Qiao, Y., Dai, J.: Intervl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.314, + 0.787, + 0.342 + ], + "angle": 0, + "content": "18. Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.342, + 0.787, + 0.383 + ], + "angle": 0, + "content": "19. Cheng, F., Wang, X., Lei, J., Crandall, D.J., Bansal, M., Bertasius, G.: Vindlu: A recipe for effective video-and-language pretraining. ArXiv abs/2212.05051 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.383, + 0.787, + 0.425 + ], + "angle": 0, + "content": "20. Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: CVPR. pp. 2818-2829 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.425, + 0.787, + 0.467 + ], + "angle": 0, + "content": "21. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B.A., Fung, P., Hoi, S.C.H.: Instructblip: Towards general-purpose vision-language models with instruction tuning. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.467, + 0.787, + 0.494 + ], + "angle": 0, + "content": "22. Dao, T., Fu, D., Ermon, S., Rudra, A., Ré, C.: Flashattention: Fast and memory-efficient exact attention with io-awareness. NeurIPS 35, 16344-16359 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.494, + 0.787, + 0.605 + ], + "angle": 0, + "content": "23. Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A., Caron, M., Geirhos, R., Alabdulmohsin, I.M., Jenatton, R., Beyer, L., Tschannen, M., Arnab, A., Wang, X., Riquelme, C., Minderer, M., Puigcerver, J., Evci, U., Kumar, M., van Steenkiste, S., Elsayed, G.F., Mahendran, A., Yu, F., Oliver, A., Huot, F., Bastings, J., Collier, M., Gritsenko, A.A., Birodkar, V., Vasconcelos, C.N., Tay, Y., Mensink, T., Kolesnikov, A., Paveti'c, F., Tran, D., Kipf, T., Luvci'c, M., Zhai, X., Keysers, D., Harmsen, J., Houlsby, N.: Scaling vision transformers to 22 billion parameters. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.605, + 0.787, + 0.646 + ], + "angle": 0, + "content": "24. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. ArXiv abs/1810.04805 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.646, + 0.787, + 0.702 + ], + "angle": 0, + "content": "25. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv abs/2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.787, + 0.77 + ], + "angle": 0, + "content": "26. Driess, D., Xia, F., Sajjadi, M.S.M., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q.H., Yu, T., Huang, W., Chebotar, Y., Sermanet, P., Duckworth, D., Levine, S., Vanhoucke, V., Hausman, K., Toussaint, M., Greff, K., Zeng, A., Mordatch, I., Florence, P.R.: Palm-e: An embodied multimodal language model. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.787, + 0.813 + ], + "angle": 0, + "content": "27. Drossos, K., Lipping, S., Virtanen, T.: Cloth: An audio captioning dataset. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 736-740. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "28. Feichtenhofer, C., Fan, H., Li, Y., He, K.: Masked autoencoders as spatiotemporal learners. NeurIPS (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InternVideo2" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.175 + ], + "angle": 0, + "content": "29. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.175, + 0.787, + 0.217 + ], + "angle": 0, + "content": "30. Fu, T.J., Li, L., Gan, Z., Lin, K., Wang, W.Y., Wang, L., Liu, Z.: Violet: End-to-end video-language transformers with masked visual-token modeling. arXiv preprint arXiv:2111.12681 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.217, + 0.787, + 0.257 + ], + "angle": 0, + "content": "31. Gabeur, V., Sun, C., Alahari, K., Schmid, C.: Multi-modal transformer for video retrieval. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 214-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.257, + 0.786, + 0.284 + ], + "angle": 0, + "content": "32. Gao, J., Sun, C., Yang, Z., Nevatia, R.: Tall: Temporal activity localization via language query. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.284, + 0.787, + 0.324 + ], + "angle": 0, + "content": "33. Gong, T., Lyu, C., Zhang, S., Wang, Y., Zheng, M., Zhao, Q., Liu, K., Zhang, W., Luo, P., Chen, K.: Multimodal-gpt: A vision and language model for dialogue with humans. ArXiv abs/2305.04790 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.325, + 0.787, + 0.351 + ], + "angle": 0, + "content": "34. Gong, Y., Chung, Y.A., Glass, J.: Ast: Audio spectrogram transformer. arXiv preprint arXiv:2104.01778 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.352, + 0.787, + 0.405 + ], + "angle": 0, + "content": "35. Goyal, R., Kahou, S.E., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fründ, I., Yianilos, P., Mueller-Freitag, M., Hoppe, F., Thurau, C., Bax, I., Memisevic, R.: The \"something something\" video database for learning and evaluating visual common sense. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.787, + 0.446 + ], + "angle": 0, + "content": "36. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.447, + 0.787, + 0.473 + ], + "angle": 0, + "content": "37. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.474, + 0.787, + 0.5 + ], + "angle": 0, + "content": "38. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.787, + 0.542 + ], + "angle": 0, + "content": "39. Idrees, H., Zamir, A.R., Jiang, Y.G., Gorban, A., Laptev, I., Sukthankar, R., Shah, M.: The thumos challenge on action recognition for videos \"in the wild\". Computer Vision and Image Understanding (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.787, + 0.582 + ], + "angle": 0, + "content": "40. Jiang, A.Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D.S., Casas, D.d.l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al.: Mistral 7b. arXiv preprint arXiv:2310.06825 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.583, + 0.787, + 0.637 + ], + "angle": 0, + "content": "41. Kim, C.D., Kim, B., Lee, H., Kim, G.: Audiocaps: Generating captions for audios in the wild. In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 119-132 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.637, + 0.787, + 0.664 + ], + "angle": 0, + "content": "42. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Carlos Niebles, J.: Dense-captioning events in videos. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.664, + 0.787, + 0.704 + ], + "angle": 0, + "content": "43. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: 2011 International conference on computer vision. pp. 2556-2563. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.705, + 0.787, + 0.732 + ], + "angle": 0, + "content": "44. Lei, J., Berg, T.L., Bansal, M.: Qvhighlights: Detecting moments and highlights in videos via natural language queries (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.787, + 0.758 + ], + "angle": 0, + "content": "45. Li, G., Xu, Y., Hu, D.: Multi-scale attention for audio question answering. arXiv preprint arXiv:2305.17993 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.787, + 0.799 + ], + "angle": 0, + "content": "46. Li, J., Li, D., Savarese, S., Hoi, S.C.H.: Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In: ICML (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "47. Li, J., Li, D., Xiong, C., Hoi, S.C.H.: Blip: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In: ICML (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "48. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.19, + 0.787, + 0.231 + ], + "angle": 0, + "content": "49. Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Liu, Y., Wang, Z., Xu, J., Chen, G., Luo, P., et al.: Mvbench: A comprehensive multi-modal video understanding benchmark. arXiv preprint arXiv:2311.17005 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.231, + 0.787, + 0.272 + ], + "angle": 0, + "content": "50. Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Wang, L., Qiao, Y.: Uniformerv2: Spatiotemporal learning by arming image vits with video uniformer. arXiv preprint arXiv:2211.09552 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "51. Li, K., Wang, Y., Li, Y., Wang, Y., He, Y., Wang, L., Qiao, Y.: Unmasked teacher: Towards training-efficient video foundation models. arXiv preprint arXiv:2303.16058 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.315, + 0.787, + 0.342 + ], + "angle": 0, + "content": "52. Li, T., Wang, L.: Learning spatiotemporal features via video and text pair discrimination. CoRR abs/2001.05691 (2020), https://arxiv.org/abs/2001.05691" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.787, + 0.396 + ], + "angle": 0, + "content": "53. Lin, K.Q., Zhang, P., Chen, J., Pramanick, S., Gao, D., Wang, A.J., Yan, R., Shou, M.Z.: Univtg: Towards unified video-language temporal grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2794-2804 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.787, + 0.424 + ], + "angle": 0, + "content": "54. Lin, T., Liu, X., Li, X., Ding, E., Wen, S.: Bmn: Boundary-matching network for temporal action proposal generation (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.425, + 0.787, + 0.467 + ], + "angle": 0, + "content": "55. Lipping, S., Sudarsanam, P., Drossos, K., Virtanen, T.: Clotho-aqa: A crowdsourced dataset for audio question answering. In: 2022 30th European Signal Processing Conference (EUSIPCO). pp. 1140-1144. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.467, + 0.695, + 0.481 + ], + "angle": 0, + "content": "56. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "57. Liu, Y., Wang, L., Wang, Y., Ma, X., Qiao, Y.: Fineaction: A fine-grained video dataset for temporal action localization. IEEE Transactions on Image Processing 31, 6937-6950 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.522, + 0.787, + 0.563 + ], + "angle": 0, + "content": "58. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.564, + 0.787, + 0.605 + ], + "angle": 0, + "content": "59. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "60. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M.H., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability. ArXiv abs/2306.07207 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "61. Maaz, M., Rasheed, H.A., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. ArXiv abs/2306.05424 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.787, + 0.73 + ], + "angle": 0, + "content": "62. Mangalam, K., Akshulakov, R., Malik, J.: Egoschema: A diagnostic benchmark for very long-form video language understanding. In: Oh, A., Neumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) NeurIPS. pp. 46212-46244 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.787, + 0.757 + ], + "angle": 0, + "content": "63. Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "64. Mei, X., Meng, C., Liu, H., Kong, Q., Ko, T., Zhao, C., Plumbley, M.D., Zou, Y., Wang, W.: Wavcaps: A chatgpt-assisted weakly-labelled audio captioning dataset for audio-language multimodal research (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "65. Monfort, M., Zhou, B., Bargal, S.A., Andonian, A., Yan, T., Ramakrishnan, K., Brown, L.M., Fan, Q., Gutfreund, D., Vondrick, C., Oliva, A.: Moments in time dataset: One million videos for event understanding. TPAMI (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InternVideo2" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "66. Moon, W., Hyun, S., Lee, S., Heo, J.P.: Correlation-guided query-dependency calibration in video representation learning for temporal grounding. arXiv preprint arXiv:2311.08835 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.191, + 0.682, + 0.205 + ], + "angle": 0, + "content": "67. OpenAI: Gpt-4 technical report. ArXiv abs/2303.08774 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.206, + 0.787, + 0.233 + ], + "angle": 0, + "content": "68. OpenAI: Gpt-4v(ison) system card. https://api_semanticscholar.org/ CorpusID:263218031 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.234, + 0.787, + 0.275 + ], + "angle": 0, + "content": "69. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.277, + 0.787, + 0.359 + ], + "angle": 0, + "content": "70. Patraucean, V., Smaira, L., Gupta, A., Continente, A.R., Markeeva, L., Dylan, Banarse, Malinowski, M., Yang, Y., Doersch, C., Matejovicova, T., Sulsky, Y., Antoine, Miech, Koppula, S., Fréchette, A., Klimczak, H., Koster, R., Zhang, J., Stephanie, Winkler, Aytar, Y., Osindero, S., Damen, D., Zisserman, A., Carreira, J.: Perception test: A diagnostic benchmark for multimodal models. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.36, + 0.786, + 0.388 + ], + "angle": 0, + "content": "71. Piczak, K.J.: Esc: Dataset for environmental sound classification. In: Proceedings of the 23rd ACM international conference on Multimedia. pp. 1015-1018 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.389, + 0.787, + 0.43 + ], + "angle": 0, + "content": "72. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: ICCV (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.431, + 0.787, + 0.472 + ], + "angle": 0, + "content": "73. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "74. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.515, + 0.787, + 0.557 + ], + "angle": 0, + "content": "75. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision. In: International Conference on Machine Learning. pp. 28492-28518. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.558, + 0.787, + 0.599 + ], + "angle": 0, + "content": "76. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.601, + 0.787, + 0.628 + ], + "angle": 0, + "content": "77. Rohrbach, A., Rohrbach, M., Tandon, N., Schiele, B.: A dataset for movie description. In: CVPR. pp. 3202-3212 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.629, + 0.787, + 0.67 + ], + "angle": 0, + "content": "78. Ryali, C., Hu, Y.T., Bolya, D., Wei, C., Fan, H., Huang, P.Y., Aggarwal, V., Chowdhury, A., Poursaeed, O., Hoffman, J., Malik, J., Li, Y., Feichtenhofer, C.: Hiera: A hierarchical vision transformer without the bells-and-whistles (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.671, + 0.787, + 0.699 + ], + "angle": 0, + "content": "79. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.787, + 0.741 + ], + "angle": 0, + "content": "80. Sun, Q., Cui, Y., Zhang, X., Zhang, F., Yu, Q., Luo, Z., Wang, Y., Rao, Y., Liu, J., Huang, T., et al.: Generative multimodal models are in-context learners. arXiv preprint arXiv:2312.13286 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.742, + 0.787, + 0.77 + ], + "angle": 0, + "content": "81. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.787, + 0.798 + ], + "angle": 0, + "content": "82. Sun, Q., Wang, J., Yu, Q., Cui, Y., Zhang, F., Zhang, X., Wang, X.: Eva-clip-18b: Scaling clip to 18 billion parameters. arXiv preprint arXiv:2402.04252 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "83. Sun, Q., Yu, Q., Cui, Y., Zhang, F., Zhang, X., Wang, Y., Gao, H., Liu, J., Huang, T., Wang, X.: Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "84. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.189, + 0.787, + 0.229 + ], + "angle": 0, + "content": "85. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. NeurIPS 35, 10078-10093 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.229, + 0.787, + 0.284 + ], + "angle": 0, + "content": "86. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models. ArXiv abs/2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.284, + 0.787, + 0.449 + ], + "angle": 0, + "content": "87. Touvron, H., Martin, L., Stone, K.R., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D.M., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A.S., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I.M., Korenev, A.V., Koura, P.S., Lachaux, M.A., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. ArXiv abs/2307.09288 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.449, + 0.787, + 0.475 + ], + "angle": 0, + "content": "88. Tran, D., xiu Wang, H., Torresani, L., Ray, J., LeCun, Y., Paluri, M.: A closer look at spatiotemporal convolutions for action recognition. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.475, + 0.787, + 0.516 + ], + "angle": 0, + "content": "89. Wang, J., Ge, Y., Yan, R., Ge, Y., Lin, K.Q., Tsutsui, S., Lin, X., Cai, G., Wu, J., Shan, Y., et al.: All in one: Exploring unified video-language pre-training. In: CVPR. pp. 6598-6608 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.516, + 0.787, + 0.556 + ], + "angle": 0, + "content": "90. Wang, L., Huang, B., Zhao, Z., Tong, Z., He, Y., Wang, Y., Wang, Y., Qiao, Y.: Videomae v2: Scaling video masked autoencoders with dual masking. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.556, + 0.787, + 0.597 + ], + "angle": 0, + "content": "91. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Gool, L.V.: Temporal segment networks: Towards good practices for deep action recognition. In: ECCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.597, + 0.787, + 0.638 + ], + "angle": 0, + "content": "92. Wang, R., Chen, D., Wu, Z., Chen, Y., Dai, X., Liu, M., Yuan, L., Jiang, Y.G.: Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In: CVPR. pp. 6312-6322 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.638, + 0.787, + 0.678 + ], + "angle": 0, + "content": "93. Wang, Y., He, Y., Li, Y., Li, K., Yu, J., Ma, X., Li, X., Chen, G., Chen, X., Wang, Y., et al.: Intervid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.678, + 0.787, + 0.733 + ], + "angle": 0, + "content": "94. Wang, Y., Li, K., Li, Y., He, Y., Huang, B., Zhao, Z., Zhang, H., Xu, J., Liu, Y., Wang, Z., Xing, S., Chen, G., Pan, J., Yu, J., Wang, Y., Wang, L., Qiao, Y.: Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.733, + 0.787, + 0.772 + ], + "angle": 0, + "content": "95. Xu, H., Ye, Q., Yan, M., Shi, Y., Ye, J., Xu, Y., Li, C., Bi, B., Qian, Q., Wang, W., et al.: mplug-2: A modularized multi-modal foundation model across text, image and video. arXiv preprint arXiv:2302.00402 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "96. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettle-moyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "97. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: CVPR. pp. 5288-5296 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "InternVideo2" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "21" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "98. Yan, S., Zhu, T., Wang, Z., Cao, Y., Zhang, M., Ghosh, S., Wu, Y., Yu, J.: Video-text modeling with zero-shot transfer from contrastive captioners. ArXiv abs/2212.04979 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.19, + 0.739, + 0.203 + ], + "angle": 0, + "content": "99. Yang, L., Fan, Y., Xu, N.: Video instance segmentation. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.204, + 0.786, + 0.245 + ], + "angle": 0, + "content": "100. Yang, M., Chen, G., Zheng, Y.D., Lu, T., Wang, L.: Basictad: an astounding rgb-only baseline for temporal action detection. Computer Vision and Image Understanding p. 103692 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.787, + 0.299 + ], + "angle": 0, + "content": "101. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., Jiang, C., Li, C., Xu, Y., Chen, H., Tian, J., Qi, Q., Zhang, J., Huang, F.: mplug-owl: Modularization empowers large language models with multimodality (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "102. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.343, + 0.787, + 0.383 + ], + "angle": 0, + "content": "103. Zablocki, E., Ben-Younes, H., Pérez, P., Cord, M.: Explainability of deep vision-based autonomous driving systems: Review and challenges. International Journal of Computer Vision 130(10), 2425-2452 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.384, + 0.787, + 0.425 + ], + "angle": 0, + "content": "104. Zellers, R., Lu, J., Lu, X., Yu, Y., Zhao, Y., Salehi, M., Kusupati, A., Hessel, J., Farhadi, A., Choi, Y.: Merlot reserve: Neural script knowledge through vision and language and sound. In: CVPR. pp. 16375-16387 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.426, + 0.787, + 0.466 + ], + "angle": 0, + "content": "105. Zeng, Z., Ge, Y., Tong, Z., Liu, X., Xia, S.T., Shan, Y.: Tvtsv2: Learning out-of-the-box spatiotemporal visual representations at scale. arXiv preprint arXiv:2305.14173 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.467, + 0.787, + 0.508 + ], + "angle": 0, + "content": "106. Zhang, B., Yu, J., Fifty, C., Han, W., Dai, A.M., Pang, R., Sha, F.: Cotraining transformer with videos and images improves action recognition. ArXiv abs/2112.07175 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.508, + 0.787, + 0.535 + ], + "angle": 0, + "content": "107. Zhang, C.L., Wu, J.Z., Li, Y.: Actionformer: Localizing moments of actions with transformers. ArXiv abs/2202.07925 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.536, + 0.787, + 0.563 + ], + "angle": 0, + "content": "108. Zhao, H., Torralba, A., Torresani, L., Yan, Z.: Hacs: Human action clips and segments dataset for recognition and temporal localization. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.565, + 0.787, + 0.619 + ], + "angle": 0, + "content": "109. Zhao, L., Gundavarapu, N.B., Yuan, L., Zhou, H., Yan, S., Sun, J.J., Friedman, L., Qian, R., Weyand, T., Zhao, Y., Hornung, R., Schroff, F., Yang, M.H., Ross, D.A., Wang, H., Adam, H., Sirotenko, M., Liu, T., Gong, B.: Videoprism: A foundational visual encoder for video understanding (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.62, + 0.787, + 0.66 + ], + "angle": 0, + "content": "110. Zhao, Y., Hessel, J., Yu, Y., Lu, X., Zellers, R., Choi, Y.: Connecting the dots between audio and text without parallel data through visual knowledge transfer. arXiv preprint arXiv:2112.08995 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.661, + 0.787, + 0.702 + ], + "angle": 0, + "content": "111. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.702, + 0.787, + 0.743 + ], + "angle": 0, + "content": "112. Zhu, B., Lin, B., Ning, M., Yan, Y., Cui, J., Wang, H., Pang, Y., Jiang, W., Zhang, J., Li, Z., et al.: Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.744, + 0.787, + 0.784 + ], + "angle": 0, + "content": "113. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. ArXiv abs/2304.10592 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.785, + 0.787, + 0.827 + ], + "angle": 0, + "content": "114. Zhu, W., Huang, Y., Xie, X., Liu, W., Deng, J., Zhang, D., Wang, Z., Liu, J.: Autoshot: A short video dataset and state-of-the-art shot boundary detection. In: CVPR. pp. 2237-2246 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.827 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/InternVideo2_ Scaling Foundation Models for Multimodal Video Understanding/3f170c5b-f8dd-444c-a833-330298501085_origin.pdf b/2024/InternVideo2_ Scaling Foundation Models for Multimodal Video Understanding/3f170c5b-f8dd-444c-a833-330298501085_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..83ddf75781e790d3db8cf940692a006a1bfdb452 --- /dev/null +++ b/2024/InternVideo2_ Scaling Foundation Models for Multimodal Video Understanding/3f170c5b-f8dd-444c-a833-330298501085_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef4e323038749a563fb025860e69d9388d04cc45ee0399837edf56c912786644 +size 12392211 diff --git a/2024/InternVideo2_ Scaling Foundation Models for Multimodal Video Understanding/full.md b/2024/InternVideo2_ Scaling Foundation Models for Multimodal Video Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ef158ee74e39faf01e9a0d79ecb3f1b252ac3de4 --- /dev/null +++ b/2024/InternVideo2_ Scaling Foundation Models for Multimodal Video Understanding/full.md @@ -0,0 +1,377 @@ +# InternVideo2: Scaling Foundation Models for Multimodal Video Understanding + +Yi Wang $^{1}$ , Kunchang Li $^{3,1}$ , Xinhao Li $^{2,1}$ , Jiashuo Yu $^{3}$ , Yinan He $^{1}$ , Guo Chen $^{2,1}$ , Baoqi Pei $^{1}$ , Rongkun Zheng $^{1}$ , Zun Wang $^{1}$ , Yansong Shi $^{1}$ , Tianxiang Jiang $^{1}$ , Songze Li $^{1}$ , Jilan Xu $^{1}$ , Hongjie Zhang $^{1}$ , Yifei Huang $^{1}$ , Yu Qiao $^{\dagger}$ , Yali Wang $^{\dagger 3,1}$ , and Limin Wang $^{\dagger 2,1}$ + +1 OpenGVLab, Shanghai AI Laboratory, Shanghai, China +2 Nanjing University, Nanjing, China + +3 Shenzhen Institutes of Advanced Technology, CAS, Shenzhen, China {wangyi,qiaoyu,wangyali,wanglimin}@pjlab.org.cn https://github.com/OpenGVLab/InternVideo/tree/main/InternVideo? + +![](images/e510ad049ec3410b35cba384edaeefda48dbe3849da62de1fbc1169937efa682.jpg) +Fig. 1: InternVideo2 yields strong transferable visual and visual-linguistic representations for action recognition, video-text understanding, to video-centric dialogue. It also exhibits capability of long-form video understanding and procedure-aware reasoning. + +Abstract. We introduce InternVideo2, a new family of video foundation models (ViFM) that achieve the state-of-the-art results in video recognition, video-text tasks, and video-centric dialogue. Our core design is a progressive training approach that unifies the masked video modeling, crossmodal contrastive learning, and next token prediction, scaling up the video encoder size to 6B parameters. At the data level, we prioritize spatiotemporal consistency by semantically segmenting videos and generating video-audio-speech captions. This improves the alignment between video and text. Through extensive experiments, we validate our designs and demonstrate superior performance on over 60 video and audio tasks. Notably, our model outperforms others on various video-related dialogue and long video understanding benchmarks, highlighting its ability to reason and comprehend longer contexts. + +Keywords: Video Foundation Model $\cdot$ Progressive Training $\cdot$ MLLM. + +# 1 Introduction + +Learning transferrable spatiotemporal representations is a critical research area in computer vision, holding diverse applications across domains such as video search [31], game control [7], robotic learning [26], self-driving [103], and scientific studies [84]. Recently, the advancement of Large Language Models (LLMs) [6, 67, 86, 87] and their multimodal variations (MLLMs) [33, 56, 68, 84] have had a profound impact on vision research and other disciplines. Embedding videos effectively into these large models and harnessing their capabilities to enhance video understanding performance has emerged as pivotal tasks [48, 61]. + +Previous research has identified several effective learning schemes for video representations, including reconstructing videos with masked inputs [28,37,85,90], aligning videos with languages [51,52,96,98], and predicting the next token using videos [1,49,83]. These approaches have turned out to be complementary and can be unified through a progressive training scheme. Notably, methods such as InternVideo [94], UMT [51], and VideoPrism [109] have utilized a two-stage training approach involving masked reconstruction and multimodal contrastive learning, leading to improved performance in downstream tasks. Following this line, we aim to further extend this progressive training scheme by incorporating video-based next token prediction and scaling the entire training process, including models and data, to build a new family of video foundation models. + +The proposed video foundation model, coined as InternVideo2, is built through a progressive training scheme. The learning involves three stages: (1) capturing spatiotemporal structure via unmasked reconstruction, (2) aligning with semantics from other modalities, and (3) enhancing its open-ended dialogue power through next token prediction. In the initial stage, the model learns to reconstruct the unmasked video tokens, allowing the video encoder to develop basic spatiotemporal perception capability. To estimate the existing tokens, vision encoders (InternViT [17] and VideoMAE-g [90]) trained differently are employed as proxies. In the next stage of crossmodal learning, the architecture is expanded to include audio and text encoders. This not only improves the alignment between videos and text but also endows InternVideo2 the ability to handle video-audio tasks. By incorporating these additional modalities, the model's understanding of videos is enriched and aligned with their semantics. Finally, in the next-token prediction stage, a video-centric dialogue system and the corresponding instruction-finetuning dataset are built to further tune the InternVideo2. By connecting InternVideo2 to LLMs, the video encoder is further updated through next-token prediction training, enhancing its ability for open-ended tasks such as VQA and video description. + +For the training of InternVideo2, we emphasize the spatiotemporal consistency and labeling quality in the data. We build a large-scale multimodal video-centric dataset consisting of 402M data entries, which includes 2M videos, 50M video-text pairs (from WebVid [3] and InternVid [93]), 50M video-audio-speech-text pairs (InternVid2), and 300M image-text pairs. Specifically, for InternVid2, we segment videos into clips semantically and focus on recalibrating the clip descriptions using three modalities: audio, video, and speech. We first generate captions for + +these three modalities separately. Then individual captions are fused together to create a more comprehensive description, which will improve the model's ability to comprehend and interpret the video accurately. + +We evaluate InternVideo2 across a wide range of video-related tasks. These tasks span from basic spatiotemporal perception, such as action recognition, to high-level reasoning tasks, such as long video or procedure-aware question-answering (QA), as given in Fig. 1. The results (in Sec. 5) demonstrate that InternVideo2 achieves the state-of-the-art performance on multiple tasks and is able to analyze and reason over sequences of actions. This top performance signifies its capability to effectively capture and understand video content. These empirical findings validate that InternVideo2 could serves as a general video encoder for future exploration in video understanding. In summary, our contributions to video understanding are as follows. + +- This paper introduces InternVideo2, a competitive family of video foundation models that leverages masked reconstruction, crossmodal contrastive learning, and next token prediction to make model perceptive, semantic, and capable of reasoning in video understanding. +- InternVideo2 achieves the state-of-the-art performance for more than 60 video / audio tasks. Our model demonstrates superior performance in video-related dialogue and long video understanding, highlighting its potential in modeling high-level world knowledge. +- We provide an enhanced dataset to train InternVideo2. This includes the validation and incorporation of audio data during training, as well as the improved captioning method. These improvements result in significant enhancements in model performance and generalization ability. + +# 2 Related Work + +Video Foundation Models. Studies on learning video foundation models becomes increasingly crucial considering its wide applications [?, 51, 52, 85, 89, 90, 92, 94, 96, 98, 109]. Typical methods in building video foundation models (ViFM) include video-text contrastive learning [52, 94, 96], masked video modeling [30, 85, 90, 94], and next token prediction [1, 80, 83]. Specifically, All-in-one [89] utilized a single backbone with unified multiple pretraining objectives. On the other hand, UMT [51] combined masked modeling with video-text contrastive learning, demonstrating strong performance in both action recognition and videolanguage tasks. Another approach is mPLUG-2 [95], which introduced a new design for modulating different modalities. It shared a common module across modalities to enhance relations while incorporating modality-specific modules for discrimination. In addition to video-text pretraining, researchers have also explored the use of audio information in videos to improve performance. MERLOT Reserve [104] learned video representations using a large-scale dataset of videospeech-transcript pairs. VALOR [14] employed independent encoders for video, audio, and text and trains a joint visual-audio-text representation. VAST [15] constructed an audio-visual-speech dataset and develops a multimodal backbone + +![](images/97027eeccf7e96697083cc75455f70ec1857ca205740c8b248bd93fb09a5b82e.jpg) +Fig. 2: Framework of InternVideo2. It consists of three consecutive training phases: unmasked video token reconstruction, multimodal contrastive learning, and next token prediction. In stage 1, the video encoder is trained from scratch, while in stages 2 and 3, it is initialized from the version used in the previous stage. + +that excels in video-audio-related tasks. VideoPrism [109] combined video-text contrastive learning and video token reconstruction on a combination of public and proprietary videos, achieving leading results across various video tasks. + +Multimodal Large Language Models. With advances in large language models (LLMs) [6, 24, 76], their multimodal versions (MLLMs) is becoming popular as it can handle open-world tasks. Seminal works like Flamingo [1] showed outstanding zero/few-shots performances over a range of multimodal tasks [36, 63, 72, 97]. Public MLLMs [33, 56, 113] such as LLaVA [56] and InstructBLIP [21] proposed to use visual instruction-tuning data to improve the visual dialogue ability. Some video-centric MLLMs have been proposed, such as VideoChat [48], VideoChatGPT [61] and Valley [60], by using instruction data to connect video encoders to LLMs for open-world video understanding. + +# 3 Method + +We learn InternVideo2 in three stages, illustrated in Fig. 2. The stages include spatiotemporal token reconstruction, video-audio-speech-language contrastive learning, and connecting to a large language model (LLM) for joint training. + +Video Encoder. The video encoder used in InternVideo2 follows the Vision Transformer (ViT) [25] and includes additional projection layers for distillation. Inspired by previous works [17, 102], we introduce attention pooling to the ViT. For input videos, we sparsely sample 8 frames [91] and perform a $14 \times 14$ ( $h \times w$ ) spatial downsampling. These spatiotemporal tokens are then concatenated with a class token and combined with 3D position embeddings. The hyperparameters used for ViT-1B and -6B are the same as CoCa-1B [102] and InternViT-6B [17], respectively. The details of ViT-6B architecture are given in Supp. + +# 3.1 Stage1: Reconstructing Unmasked Video Tokens + +We exploit two expert models to guide the video encoder to conduct token-level reconstruction for unmasked areas. Specifically, we adopt InternVL-6B [17] and VideoMAEv2-g [90] to transfer unmasked knowledge via simple projection layers. + +When training, we input the full videos into different teachers and mask out $80\%$ of the tokens frame by frame, under the semantic guidance of the multimodal model InternVL and motion-aware model VideoMAEv2. We only align the unmasked tokens, by minimizing their mean squared error (MSE) between student and teachers. The learning objective is to reconstruct the remaining tokens as: + +$$ +\mathcal {L} = \frac {1}{Z} \sum \left(\alpha_ {1} | f ^ {V} (\mathbf {V} _ {p}) - h (\mathbf {V} _ {p}) | ^ {2} + \alpha_ {2} | f ^ {V} (\mathbf {V} _ {p}) - g (\mathbf {V} _ {p}) | ^ {2}\right), \tag {1} +$$ + +where $f^V$ , $h$ , and $g$ are our video encoder, InternViT-6B [17], and ViT-g of VideoMAEv2, respectively. $p$ stands for the token index and $f(\mathbf{V}_p)$ is the corresponding token extracted by InternVideo2 for input video $\mathbf{V}$ . $Z$ is the normalization factor. $\alpha_{1}$ and $\alpha_{2}$ balance the influence between the employed models. + +In implementation, we align the last 6 layers of InternVL, last 4 layers of VideoMAEv2, and the final output token of InternVL. Different losses are simply summed for optimization. After pretraining, we drop those projection layers and only use the basic encoder. Compared with only using the multimodal model in UMT and VideoPrism, our strategy makes the vision encoder multimodal-friendly as well as enhances its temporal sensitivity for action modeling. + +# 3.2 Stage 2: Aligning Video to Audio-Speech-Text + +We exploit the correspondence between video and audio, speech, and text to encourage InternVideo2 to learn more semantics. In practice, InternVideo2 has a huge video encoder, and its employed audio and text encoders are relatively lightweight. The used audio encoder is a 12-layer transformer initialized with BEATs [13] (90M). It takes in audio features, which are 64-dimensional log Mel filterbank spectrograms using a 25ms Hamming window, transformed from 10-second-long clips, padding with zeros. For the text and speech encoders, we initialize the text encoder and multimodal decoder using BERT-Large [24]. Specifically, we utilize the initial 19 layers of BERT-Large as the text encoder, with the subsequent 5 layers equipped with cross-attention layers serving as the multimodal decoder. + +For pretraining objectives, we establish alignment across different modalities via text, including video, audio, image, and speech. We employ crossmodal contrastive and matching losses with masked language reconstruction loss as: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {C O N}} + \mathcal {L} _ {\mathrm {M A C}} + \mathcal {L} _ {\mathrm {M L M}}, \tag {2} +$$ + +The employed $\mathcal{L}_{\mathrm{MAC}}$ and $\mathcal{L}_{\mathrm{MLM}}$ are standard loss from [19]. Specifically, the crossmodal contrastive learning is given as: + +$$ +\mathcal {L} _ {\mathrm {C O N}} = \sum_ {M, T _ {M ^ {\prime}}} \mathcal {L} _ {\mathrm {C O N}} (M, T _ {M ^ {\prime}}) = - \sum_ {M, T _ {M ^ {\prime}}} \left(\sum_ {i = 1} ^ {N} \log \frac {\exp (\sin \left(f _ {i} ^ {M} , f _ {i} ^ {T _ {M ^ {\prime}}}\right) / \tau)}{\sum_ {j = 1} ^ {N} \exp (\sin \left(f _ {i} ^ {M} , f _ {j} ^ {T _ {M ^ {\prime}}}\right) / \tau)} + \sum_ {i = 1} ^ {N} \log \frac {\exp (\sin \left(f _ {i} ^ {T _ {M ^ {\prime}}} , f _ {i} ^ {M}\right) / \tau)}{\sum_ {j = 1} ^ {N} \exp (\sin \left(f _ {i} ^ {T _ {M ^ {\prime}}} , f _ {j} ^ {M}\right) / \tau)}\right), \tag {3} +$$ + +where $f^V$ and $f^T$ denote the learned video and text embeddings, respectively. + +$M$ and $T_{M^{\prime}}$ indicates the modality of input signals and the text descriptions describing, respectively. $\sin (\cdot)$ computes the cosine similarity between two features. $\tau$ is the learnable temperature. + +For the matching part, it is given as: + +$$ +\mathcal {L} _ {\mathrm {M A C}} = - y \log f _ {p} (\mathbf {V}, \mathbf {T}) - (1 - y) \log \left(1 - f _ {p} (\mathbf {V}, \mathbf {T})\right), \tag {4} +$$ + +where $f_{p}(\mathbf{V},\mathbf{T})$ computes the matching likelihood between $\mathbf{V}$ and $\mathbf{T}$ . $y$ denotes whether the given video and text are paired $(y = 1)$ or not $(y = 0)$ . + +The employed masked language modeling loss is: + +$$ +\mathcal {L} _ {\mathrm {M L M}} = - \log f _ {p} ^ {T} \left(\mathbf {T} _ {j} \mid \mathbf {T} _ {< j}\right), \tag {5} +$$ + +where $f_{p}^{T}(\mathbf{T}_{j}|\mathbf{T}_{< j})$ computes the likelihood of the $j_{\mathrm{th}}$ text token based on the previous ones. Here $\mathbf{T}$ refers to video captions. + +To improve the training efficiency, we employ the masked learning strategy, aligning unmasked video tokens to tokens from other modalities first, then using full video tokens reconstruction shortly. Specifically, it consists of two steps as follows: + +Aligning Masked Visual-Language-Audio. We freeze the audio encoder and focus on aligning visual, audio, and text features. For pre-training, we use a comprehensive set of image, video, and audio-video data. The combinations of modalities used are represented as $\{M,T_{M^{\prime}}\} \in \{\{I,T_I\},\{V,T_V\},\{V,T_{VAS}\},\{VA,T_{VAS}\}\}$ where each pair denotes the concatenated features from the respective modalities. + +Unmasked Visual-Audio-Language Post-Pretraining. We freeze the vision encoder to jointly align audio, visual, and text features. Post-pretraining is conducted using a smaller subset of image and video data (25M samples), along with the full set of audio (0.5M samples) and audio-video data (50M samples). Since the parameters of the largest ViT-6B model are frozen, we do not use masking strategies in this phase to ensure consistency with the inference process and to minimize any performance degradation in downstream tasks. The modality combinations used here are $\{M, T_{M'}\} \in \{\{I, T_I\}, \{V, T_V\}, \{A, T_A\}, \{V, T_{VAS}\}, \{VA, T_{VA}\}\}$ . + +# 3.3 Stage3: Predicting Next Token with Video-Centric Inputs + +To further enrich the semantics embedded in InternVideo2 and improve its support for video-centric dialogue, we tune it by connecting it to a LLM with QFormer design [46, 47]. We employ the progressive learning scheme in [49] by using InternVideo2 as the video encoder and train a video blip for communicating with open-sourced LLM [40, 111]. In training, the video encoder will be updated. Detailed training receipts are the same with VideoChat2 [49]. + +# 4 Multimodal Video Data + +We list our training data in Tab. 1. Among the datasets used, KMash and InternVid2 are newly built, whereas the rest are public available. + +Table 1: Summary of datasets used in InternVideo2 pretraining process. + +
Pretraining Stage DatasetDomain# of clipsAnnotation
Stage 1KMashWeb Video2M-
Stage 2 (img-text)LAION, etcWeb Image300MAlt-text / Generated Caps
Stage 2 (vid-text)WebVid2MWeb Video250kAlt-text
WebVid10MWeb Video9.7MAlt-text
InternVidYoutube Video40MGenerated Caption
InternVid2Youtube Video50MGenerated Caption
Stage 3LLaVA, etcWeb Image/Video2.1MConversation, QA
+ +![](images/a42554fc99b0bb401df147969de40c3b2d0a527c4be924025f4127a4531bac0c.jpg) +Fig. 3: The framework of our video multimodal annotation system, called VidCap, consists of four main components: video, audio, and speech captioners, along with a LLM for integrating captions from these modalities. + +# 4.1 Video-only Data for Masked Autoencoders + +We curate a new video set without labels named $K$ -Mash from action recognition datasets [10, 35, 38, 65, 108], as detailed in Supp. It encompasses a wide range of video types, including first- and third-person perspectives, with both short and long duration, and featuring various settings. Further, we give $K$ - $\text{Mash}_{2M}$ with additionally sourced and selected 844K videos from YouTube for diversity. + +# 4.2 Videos with Audio-Speech Modalities + +We build a multimodal video dataset, coined as InternVid2, with video-audiospeech information and their descriptions for strengthening video perception via other modalities. It consists of 100M videos along with their VAS captions. In InternVid2, we collect videos from several sources (detailed in the supp), segment them into clips, and automatically annotate them based on their unimodal or crossmodal inputs. We highlight the importance of temporal segmentation in clip generation and our video multimodal annotation system. We find refining them leads to notable downstream improvements. + +Temporal Consistency Matters. We employ a temporal boundary detection model AutoShot [114] to segment videos into clips instead of SceneDet filter from + +FFmpeg. It predicts boundaries based on temporal semantic variations rather than pixel differences, and is able to generate semantically complete cuts without mixing extra frames with inconsistent context. + +Video Multimodal Annotation. We design a video multimodal annotation system VidCap to give proper unimodal and crossmodal descriptions for textualizing videos from different perceptions. It automatically captions visual, audio, and speech of InternVid2, then it corrects them and fuses them for cross-modal captions via LLM. The system frame is given in Fig. 3. VidCap has independent video, audio, speech captioner, and a LLM for caption refinement and fusion. For video, speech, and caption postprocessing, we employ existing methods as the video captioning pipeline in [93], WhisperV2-large model [75], and Vicuna-1.5 [111]. For audio, we craft a audio captioner upon VideoChat [48], as we find no open-sourced one. It extracts audio features from inputs by Beats [13]. We learn it by only tuning its Qformer using a combination of the large-scale audio-text corpus WavCaps [64] dataset. Details are given in the supplementary material. + +# 5 Experiments + +In our evaluation of InternVideo2, we assess the models from three learning stages. The evaluation covers a wide range of tasks, including video recognition, video retrieval, question-answering, and more. It includes various scenarios such as zero-shot learning, finetuning, and linear probing. + +For InternVideo2 trained in stage 1, 2, and 3, we denote them with InternVideo2 $_{s1}$ , InternVideo2 $_{s2}$ , and InternVideo2 $_{s3}$ , respectively. We also learn a CLIP-style InternVideo2 indicated by InternVideo2 $_{clip}$ . It is post-pretrained from InternVideo2 $_{s2}$ by only preserving video and text encoders and contrastive loss. + +Each training stage of InternVideo2-6B uses different configurations and resources. In the first stage, we employ 256 NVIDIA A100 GPUs and train the model for 18 days. The second stage also utilizes 256 A100 GPUs and spans a training period of 14 days. Finally, in the third stage, we use 64 A100 GPUs and train the model for 3 days. We introduce DeepSpeed and FlashAttention [22] for training and inference. More implementation details, experiment results, and ablations are given in the supp. + +# 5.1 Video Classification + +Action Recognition. We test InternVideo2 on Kinetics (i.e., K400, 600 and 700 [8-10]), Moments in Time V1 (MiT) [65], Something-Something V2 (SSv2) [35], UCF [79], HMDB [43], Charades [32], ActivityNet [38] (ANet) and HACS [108]. We evaluate in four settings: (a) end-to-end finetuning the whole backbone; (b) attentive probing is similar to linear pooling, but extra trains the attention pooling layer [102]. (c) linear probing which freezes the backbone and only trains the task head; and (d) zero-shot. + +Table 2: End-to-end finetuning action recognition results (top-1 accuracy) on Kinetics, SomethingSomething, and Moments in Time. ${}^{ \dagger }$ denotes that the result is achieved with different resolutions or frame rates. + +
MethodTraining DataSettingK400K600K700Sth-Sthv2MiTANetHACS
CoVeR [106]IV-3B16 × 44887.187.979.870.846.1--
Hiera-H [78]V-0.25M16 × 22487.888.881.176.5---
CoCa-g [102]I-3B16 × 57688.989.482.7-49.0--
MTV-H [90]IV-370M32 × 28089.990.383.4----
VideoMAEv2-g [90]V-1.35M64 × 26690.089.9-77.0†---
V-JEPA-H [4]V-2M16 × 224---77.0---
MVD-H [92]IV-1.25M16 × 224---77.3---
UniFormerV2-L [50]IV-401M64 × 33690.090.182.773.0†47.8†94.795.4
InternVideo [94]V-12Mensemble91.191.384.077.2---
InternVideo2s1-1BIV-1.1M8 × 22491.391.485.077.150.8--
InternVideo2s1-1BIV-1.1M16 × 22491.691.685.477.150.9--
InternVideo2s1-6BIV-2M8 × 22491.991.785.777.551.0--
InternVideo2s1-6BIV-2M16 × 22492.191.985.977.451.295.997.0
+ +Table 3: Attentive probing recognition results (top-1 accuracy) on Kinetics-400/600/700, Moments in Time and Something-Something V2. + +
MethodTraining DataSettingK400K600K700MiTSSV2
UMT-L [51]IV-25M-82.8--40.354.5
VideoMAEv2-g [90]V-1.35M-82.1--35.056.1
V-JEPA-H [4]V-2M16 × 38481.9---72.2
DINOv2-g [69]I-142M16 × 22483.4---50.0
VideoPrism-g [109]V-619M16 × 28887.2--45.568.5
ViT-e [23]I-4B128 × 22486.5--43.6-
ViT-22B [23]I-4B128 × 22488.0--44.9-
CoCa-g [102]I-3B16 × 57688.088.581.147.4-
InternVideo2s2-1BIV-25.5M16 × 22487.988.079.546.367.3
InternVideo2s2-6BIV-400M16 × 22488.889.181.047.867.7
+ +End-to-end Finetuning. Tab. 2 shows InternVideo2-6B obtains new state-of-the-art (SOTA) results on Kinetics $(92.1\% / 91.9\% / 85.9\%)$ on K400/600/700, respectively), SthSthv2, MiT, ANet, and HACS with only 16 frames, while the previous SOTAs require larger resolution (224 vs. 576) or model ensemble. As for MiT in Table 2, InternVideo2-6B exceeds the previous SOTA, CoCa-g, by a significant margin of $2.2\%$ $(51.2\%$ vs. $49.0\%)$ . Regarding the temporal-related actions in Table 2, our InternVideo2-6B also surpasses MVD [92] on SSv2 $(77.5\%$ vs. $77.3\%)$ . Moreover, our InternVideo2-6B showcases top performance on untrimmed video analysis, as indicated in Table 2, with $95.9\%$ on ActivityNet and $97.0\%$ on HACS. These results affirm our model's superior capability for robustly identifying complex actions across varied scenes. Note "I" and "V" denotes images and videos, respectively. "IV-3B" means the total number of the used images and videos is 3B, while "I-3B" means using 3B images. + +Attentive Probing. As in Tab. 3, InternVideo2-6B not only outperforms ViT-22B [23] and CoCa-g [102] in scene-focused datasets but also surpasses or matches the performance of the latest video foundation model [4,109], on datasets emphasizing temporal dynamics (SthSthV2). This underscores our model's exceptional ability to understand and interpret both spatial and temporal information effectively. + +Table 4: Linear probing action recognition results (top-1 accuracy) on Kinetics-400, Something-Something V2, UCF-101 and HMDB-51. + +
MethodSettingK400SSV2UCF-101HMDB-51
VideoMAEv2-H [90]12 × 22425.8-56.434.1
TVTSv2-H [105]12 × 22473.1-91.865.7
OpenCLIP-G [20]8 × 22478.335.890.7-
DINOv2-g [69]8 × 22478.438.391.2-
InternVideo2s1-1B16 × 22481.646.396.071.6
InternVideo2s1-6B16 × 22482.047.896.371.8
InternVideo2s2-6B16 × 22484.256.797.380.7
+ +Table 5: Zero-shot action recognition on UCF, HMDB, MiTv1, SSv2-MC, and Charades. + +
Method#FTraining DataUCFHMDBMiTSSv2-MCCharades
CLIP [74]12I-400M68.943.2-29.6-
TVTSv2 [105]12V-8.5M78.052.1-48.4-
VideoCoCa-g [98]16V-145M----25.8
VideoPrism-g [109]16V-619M----32.4
InternVideo2clip-1B8IV-25.5M88.853.931.661.532.9
InternVideo2clip-6B8IV-400M89.556.732.963.534.6
+ +Table 6: Zero-shot action recognition results on Kinetics. + +
Method#FK400K600K700
top-1AVGtop-1AVGtop-1AVG
CLIP [74]858.470.155.167.246.158.4
EVA-CLIP-L [81]1-65.0-64.9-59.1
EVA-CLIP-E [81]1-69.8-69.3-63.4
ViCLIP-L [93]864.875.762.273.554.366.4
VideoCoCa-g [98]1672.081.3-
InternVL-6B [17]869.179.468.978.860.671.5
EVA-CLIP-18B [82]1679.4-79.4-72.2
VideoPrism-g [109]1676.485.4-
InternVideo2clip-1B873.182.472.881.864.975.2
InternVideo2clip-6B872.782.271.781.264.275.2
+ +Linear Probing. In Tab. 4, InternVideo2-1B significantly outperforms the previous SOTA, DINOv2-g [69], by notable margins: $+3.2\%$ on K400, $+8.0\%$ on SthSthV2, and $+4.8\%$ on UCF-101. As we scale the model, an upward trend in results is observed, underscoring the benefits of model enhancement. Notably, the integration of multimodal pretraining (stage 2) yields further rise in results. We suppose stage 2 enhances feature discrimination. + +Zero-shot. Table 5 and 6 show InternVideo2 gets $72.7\%$ / $71.7\%$ / $64.2\%$ on K400/600/700, respectively, outperforming others but VideoPrism on K400 $(76.4\%)$ . On UCF [79], HMDB [43], MiT [65], SSv2-MC, and Charades, InternVideo2 gives an cutting edges over others. The clear gap between VideoPrism and InternVideo2 on K400 may signify the importance of pretraining corpus in VideoPrism (311M videos with text and 36.1M of them are manually labeled) for + +Table 7: Finetuned temporal action localization results on THUMOS14 [39], ActivityNet [42], HACS Segment [108] and FineAction [57]. We report average mAP. "Flow" denotes the ensembling I3D flow feature. * denotes the result is achieved with Flow. + +
BackboneTHUMOS14HACSActivityNetFineAction
I3D [10] + Flow66.8-35.6-
R(2+1)D [88]55.6-36.6-
InternVideo71.6*41.339.017.6
VideoMAEv2-g [90]69.5--18.2
InternVideo2s1-1B69.842.440.427.2
InternVideo2s1-6B72.043.341.227.7
+ +Table 8: Video instance segmentation performance (mAP) on YouTube-VIS19 [99]. + +
MethodBackbone#ParamsYouTubeVIS19
Mask2FormerSwin-L (image) [58]219M60.3
Mask2FormerInternViT (image)6B63.4
Mask2FormerInternVideo2s16B64.2
+ +K400 in zero-shot. Note that on the datasets of Kinetics, UCF101 and HMDB51, which have a distribution closer to the pre-training dataset used in stage1, the performance of Intervvideo2-6B is slightly inferior to that of Intervvideo2-1B. We suppose this is caused by Intervvideo2-6B uses a more abundant pretraining dataset in stage2, leading to the forgetting of pretraining data in stage1. + +Temporal Action Localization. We evaluate models on four temporal action localization (TAL) datasets: THUMOS14 [39], ActivityNet [42], HACS Segment [108] and FineAction [57] in a feature-based manner with finetuning. We employ output of the 7-th layer from InternVideo2 for inputs as the corresponding features without fusing anything else. ActionFormer [107] is used as the detection head. We report mean Average Precision(mAP) under multiple tIoU as in [12,54,100,107]. In Table 7, InternVideo2-6B gets the highest mAP among all comparisons in all datasets, while InternVideo2-1B nearly surpass other methods except in THUMOS14. We find InternVideo2-6B almost consistently improves mAP with a notable margin from InternVideo2-1B except in FineAction. We suppose scaling model capacity without data refinement cannot nontrivially improve fine-grained discrimination abilities of models. Scaling detailed annotations in training may address this issue. + +Video Instance Segmentation. We evaluate on the Video Instance Segmentation (VIS) dataset Youtube-VIS 2019 [99]. Built upon Mask2Former [18], we employ the video encoder of InternVideo2 as backbone with ViT-adapter [16] for features. We also try InternViT [17] for comparisons. In Table 8, InternVideo2 gets the highest mAP among all. This validates its effectiveness in relatively fine-grained spatiotemporal perception. + +Table 9: Results of zero-shot video retrieval in both text-to-video (T2V) and video-to-text (V2T) on MSR-VTT, LSMDC, DiDeMo, MSVD, ANet, and VATEX. + +
MethodMSR-VTTLSMDCDiDeMoMSVDANetVATEX
T2VV2TT2VV2TT2VV2TT2VV2TT2VV2TT2VV2T
CLIP [74]30.424.213.911.912.718.740.557.29.113.2--
CLIP4Clip [59]32.0-15.1---38.5-----
ViCLIP [93]42.441.320.116.918.427.949.175.115.124.0--
InternVideo-L [94]40.739.617.613.231.533.543.467.630.731.449.569.5
UMT-L [51]40.737.124.921.948.649.949.074.541.939.4--
VideoCoCa-g [98]34.464.7------34.533.053.273.6
VideoPrism-g [109]39.771.0------52.750.362.577.1
InternVideo2s2-1B51.950.932.027.357.054.358.183.360.454.870.485.4
InternVideo2s2-6B55.953.733.830.157.957.159.383.163.256.571.585.3
+ +Table 10: Results of finetuning video retrieval in both text-to-video (T2V) and video-to-text (V2T) on MSR-VTT, LSMDC, DiDeMo, MSVD, ANet, and VATEX. + +
MethodMSR-VTTLSMDCDiDeMoMSVDANetVATEX
T2VV2TT2VV2TT2VV2TT2VV2TT2VV2TT2VV2T
CLIP [74]38.238.722.522.632.233.9--26.126.9--
CLIP4Clip [59]45.645.924.323.843.043.645.248.440.341.6--
ViCLIP [93]52.551.833.032.549.450.2--49.848.1--
UMT-L [51]58.858.643.041.470.465.758.282.466.864.472.086.0
InternVideo2s2-6B62.860.246.446.774.271.961.485.274.169.775.589.3
+ +# 5.2 Video-Audio-Language Tasks + +We evaluate InternVideo2 on video retrieval, captioning, and multi-choice question-answering (QA). The former two tasks are conducted by matching video representation and the candidate text ones using the text encoder in stage 2. The latter is tested by the VideoLLM learned in stage 3. We also test audio tasks. + +Video Retrieval. We evaluate the video retrieval on six popular benchmarks [2, 11,38,77,97], as shown in Tab. 9 and 10. In eval, eight frames from the input videos are uniformly sampled. We report $R@1$ scores for both text-to-video (t2v) and video-to-text (v2t) tasks in Tab. 9 and 10. R@5 and R@10 are given in Supp. + +Table 9 and 10 demonstrate that InternVideo2 outperforms other state-of-the-arts with a notable margin in both t2v and v2t of all used datasets no matter in zero-shot or finetuned settings, except for the v2t of MSR-VTT, where VideoPrism gives the best result. This shows the video-language semantic alignment of transferrity of InternVideo2. + +Video Temporal Grounding. We evaluate InternVideo2 on two video temporal grounding (VTG) datasets: Qvhighlight [44], and Charade-STA [32]. The eval setting and used features are the same as in TAL. We use CG-DETR [66] as the grounding head. We report R1@0.3, R1@0.5, R1@0.7, and mAP for moment retrieval as in [44, 53, 66]. Highlight Detection is evaluated in terms of "Very Good" mAP and HiT@1. In Table 11, InternVideo2-1B and InternVideo2-6B bring + +Table 11: Finetuned temporal grounding on QVHighlight [44] and Charade-STA [32]. + +
FeatureR1@0.5R1@0.7mAPmAPHiT@1
CLIP [74]64.9748.6542.9639.8364.19
CLIP+SlowFast [29]65.4348.3842.8640.3366.21
InternVideo2s2-1B70.0054.4547.0242.3669.74
InternVideo2s2-6B71.4256.4549.2442.9072.00
+ +(a) QVHighlight + +
FeatureR1@0.3R1@0.5R1@0.7mIoU
CLIP [74]65.6252.7730.1645.85
CLIP+SlowFast [29]70.4358.4436.3450.13
InternVideo2s2-1B78.4168.3645.0357.12
InternVideo2s2-6B79.7070.0348.9558.79
+ +(b) Charade-STA +Table 12: Audio retrieval results on AudioCaps [41], Clothov1, and Clothov2 [27]. We report text-to-audio R@1 accuracy in zero-shot and finetuning settings. + +
MethodZero-shotFinetuning
AudioCapsClothoV1ClothoV2AudioCapsClothoV1ClothoV2
VIP-ANT [110]27.7---
VAST [15]---52.025.126.9
LanguageBind [112]-12.112.1---
InternVideo2s2-6B37.117.417.455.225.327.2
+ +Table 13: Results of AudioQA on ClothoAQA [55] and Audio-MusicAVQA (AMAVQA) [5], and audio classification on the ESC-50 [71]. Both in the finetuning setting. + +
BackboneClothQAAMAVQA
AquaNet [55]14.7865.59
MWAFM [45]22.2467.54
InternVideo2s230.1480.51
+ +(a) ClothoAQA and AMAVQA + +
MethodTop-1 Acc
AST [34]95.60
BEATS [13]98.10
InternVideo2s298.60
+ +(b) ESC-50 + +gradual performance improvements compared to CLIP [73] and CLIP [73] + Slow-fast [29]. This suggests that a larger spatiotemporal model is more beneficial to short-term video semantic alignment capabilities. + +Audio-related Tasks. We evaluate InternVideo2's audio and text encoders on audio tasks, including audio-text retrieval on AudioCaps [41], Clothov1, and Clothov2 [27]; audioQA on ClothoAQA [55] and Audio-MusicAVQA [5]; and audio classification on the ESC-50 [71]. As shown in Tab. 12, 13a, and 13b, our model achieves state-of-the-art performance on all downstream tasks. Considering the limited size of the used audio and text encoders, these audio-related results show crossmodal contrastive learning's benefits are mutual to the used modalities. Audio and the corresponding text models also gain from this learning. + +# 5.3 Video-centric Dialogue and its Applications + +Tab. 14 shows results on MVBench [49], Egoschema [62], and Perception Test [70], by equipping VideoChat2 [49] with our InternVideo2 encoder. VideoChat2-InternVideo2 demonstrates a significant performance advantage over other systems + +Table 14: Results of Chat-centric Evaluation on Multiple Choice Video-QA on MVBench [49], Egoschema and Perception Test. + +
ModelViEncoderLLMMVBenchEgoschemaPerception Test
GPT-4V [68]-GPT-443.5--
Gemini 1.0 Pro [84]--37.755.751.1
Gemini 1.0 Ultra [84]---61.554.7
Gemini 1.5 Pro [84]---72.2-
LLaVA-Next-VideoCLIP-LVicuna-7B46.543.948.8
VideoLLaMA2CLIP-L-336Mistral-7B54.651.751.4
VideoLLaMA2CLIP-L-336Mistral-8*7B53.953.352.2
VideoChat2UMT-LVicuna-7B51.1--
VideoChat2InternVideo2s3-1BMistral-7B60.355.853.0
+ +in MVBench. In Egoschema and Perception Test evaluations, it stands out as the top performer among open-source models. When contrasted with proprietary commercial models like Gemini, VideoChat2-InternVideo2 exhibits results that are slightly lower but still comparable in the Perception Test. Egoschema requires longer context harnessing, and we plan to further explore leveraging LLM for long multimodal sequence modeling. Considering these benchmarks not only involve perception but also reasoning, these results suggest InternVideo2 does embed knowledge to partially model the world. It also validates the importance of learning a transferrable video representation for current video-related MLLM. We also give several qualitative evaluations with popular GPT-4V and GeminiPro (in supp). + +# 6 Conclusion and Discussion + +We have introduced a new family of video foundation models called InternVideo2, which achieves the state-of-the-art performance across various video and audio tasks. In InternVideo2, we combine masked video modeling, video-audio-text contrastive learning, and next token prediction into a unified framework. Additionally, we create a new video-text dataset that incorporates video-audio-speech fused captions as descriptions. The dataset contains temporally segmented clips with high semantic coherence. These designs in InternVideo2 contribute to enhancing video understanding in both perception and reasoning tasks. Notably, InternVideo2 excels in video-related dialogue and long video understanding, demonstrating its effectiveness in capturing high-level semantics. + +Limitations and Discussions. Despite its achievements, InternVideo2 does not introduce specific novel architectural design. Instead, it leverages the existing learning techniques for scaling video foundation models while focusing on improving data processing to enhance its spatiotemporal perception, semantic alignment, and basic knowledge embedding. Similarly to previous studies [51, 101], InternVideo2 still grapples with limitations stemming from fixed input resolutions, sampling rates, and highly compressed tokens, which restrict its ability to express rich video information and capture fine-grained details. + +# Acknowledgements + +This work is partially supported by the National Key R&D Program of China (No. 2022ZD0160102), National Natural Science Foundation of China (No. 62076119, No. 61921006)), and the Science and Technology Commission of Shanghai Municipality under Grant No. 23QD1400800 and No. 23YF1461900. + +# References + +1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: a visual language model for few-shot learning. ArXiv abs/2204.14198 (2022) +2. Anne Hendricks, L., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.: Localizing moments in video with natural language. In: ICCV. pp. 5803-5812 (2017) +3. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: ICCV (2021) +4. Bardes, A., Garrido, Q., Ponce, J., Chen, X., Rabbat, M., LeCun, Y., Assran, M., Ballas, N.: V-JEPA: Latent video prediction for visual representation learning (2024), https://openreview.net/forum?id=WFYbBOEOtv +5. Behera, S.R., Injeti, K.M., Patibandla, J.S.K., Pokala, P.K., Pailla, B.R.: Aquallm: Audio question answering data generation using large language models. arXiv preprint arXiv:2312.17343 (2023) +6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS (2020) +7. Bruce, J., Dennis, M., Edwards, A., Parker-Holder, J., Shi, Y., Hughes, E., Lai, M., Mavalankar, A., Steigerwald, R., Apps, C., et al.: Genie: Generative interactive environments. arXiv preprint arXiv:2402.15391 (2024) +8. Carreira, J., Noland, E., Banki-Horvath, A., Hillier, C., Zisserman, A.: A short note about kinetics-600. ArXiv abs/1808.01340 (2018) +9. Carreira, J., Noland, E., Hillier, C., Zisserman, A.: A short note on the kinetics-700 human action dataset. ArXiv abs/1907.06987 (2019) +10. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: CVPR (2017) +11. Chen, D.L., Dolan, W.B.: Collecting highly parallel data for paraphrase evaluation. In: Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies-Volume 1. pp. 190-200. Association for Computational Linguistics (2011) +12. Chen, G., Zheng, Y.D., Wang, L., Lu, T.: Dscan: improving temporal action detection via dual context aggregation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 248-257 (2022) +13. Chen, S., Wu, Y., Wang, C., Liu, S., Tompkins, D., Chen, Z., Che, W., Yu, X., Wei, F.: BEATs: Audio pre-training with acoustic tokenizers. In: Proceedings of the 40th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 202, pp. 5178-5193. PMLR (23-29 Jul 2023) + +14. Chen, S., He, X., Guo, L., Zhu, X., Wang, W., Tang, J., Liu, J.: Valor: Vision-audio-language omni-perception pretraining model and dataset. arXiv preprint arXiv:2304.08345 (2023) +15. Chen, S., Li, H., Wang, Q., Zhao, Z., Sun, M., Zhu, X., Liu, J.: Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. Advances in Neural Information Processing Systems 36 (2024) +16. Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534 (2022) +17. Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Zhong, M., Zhang, Q., Zhu, X., Lu, L., Li, B., Luo, P., Lu, T., Qiao, Y., Dai, J.: Intervl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238 (2023) +18. Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: NeurIPS (2021) +19. Cheng, F., Wang, X., Lei, J., Crandall, D.J., Bansal, M., Bertasius, G.: Vindlu: A recipe for effective video-and-language pretraining. ArXiv abs/2212.05051 (2022) +20. Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: CVPR. pp. 2818-2829 (2023) +21. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B.A., Fung, P., Hoi, S.C.H.: Instructblip: Towards general-purpose vision-language models with instruction tuning. In: NeurIPS (2023) +22. Dao, T., Fu, D., Ermon, S., Rudra, A., Ré, C.: Flashattention: Fast and memory-efficient exact attention with io-awareness. NeurIPS 35, 16344-16359 (2022) +23. Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A., Caron, M., Geirhos, R., Alabdulmohsin, I.M., Jenatton, R., Beyer, L., Tschannen, M., Arnab, A., Wang, X., Riquelme, C., Minderer, M., Puigcerver, J., Evci, U., Kumar, M., van Steenkiste, S., Elsayed, G.F., Mahendran, A., Yu, F., Oliver, A., Huot, F., Bastings, J., Collier, M., Gritsenko, A.A., Birodkar, V., Vasconcelos, C.N., Tay, Y., Mensink, T., Kolesnikov, A., Paveti'c, F., Tran, D., Kipf, T., Luvci'c, M., Zhai, X., Keysers, D., Harmsen, J., Houlsby, N.: Scaling vision transformers to 22 billion parameters. In: ICML (2023) +24. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. ArXiv abs/1810.04805 (2018) +25. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv abs/2010.11929 (2020) +26. Driess, D., Xia, F., Sajjadi, M.S.M., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q.H., Yu, T., Huang, W., Chebotar, Y., Sermanet, P., Duckworth, D., Levine, S., Vanhoucke, V., Hausman, K., Toussaint, M., Greff, K., Zeng, A., Mordatch, I., Florence, P.R.: Palm-e: An embodied multimodal language model. In: ICML (2023) +27. Drossos, K., Lipping, S., Virtanen, T.: Cloth: An audio captioning dataset. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 736-740. IEEE (2020) +28. Feichtenhofer, C., Fan, H., Li, Y., He, K.: Masked autoencoders as spatiotemporal learners. NeurIPS (2022) + +29. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: ICCV (2019) +30. Fu, T.J., Li, L., Gan, Z., Lin, K., Wang, W.Y., Wang, L., Liu, Z.: Violet: End-to-end video-language transformers with masked visual-token modeling. arXiv preprint arXiv:2111.12681 (2021) +31. Gabeur, V., Sun, C., Alahari, K., Schmid, C.: Multi-modal transformer for video retrieval. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 214-229. Springer (2020) +32. Gao, J., Sun, C., Yang, Z., Nevatia, R.: Tall: Temporal activity localization via language query. In: ICCV (2017) +33. Gong, T., Lyu, C., Zhang, S., Wang, Y., Zheng, M., Zhao, Q., Liu, K., Zhang, W., Luo, P., Chen, K.: Multimodal-gpt: A vision and language model for dialogue with humans. ArXiv abs/2305.04790 (2023) +34. Gong, Y., Chung, Y.A., Glass, J.: Ast: Audio spectrogram transformer. arXiv preprint arXiv:2104.01778 (2021) +35. Goyal, R., Kahou, S.E., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fründ, I., Yianilos, P., Mueller-Freitag, M., Hoppe, F., Thurau, C., Bax, I., Memisevic, R.: The "something something" video database for learning and evaluating visual common sense. In: ICCV (2017) +36. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: CVPR (2017) +37. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022) +38. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: CVPR (2015) +39. Idrees, H., Zamir, A.R., Jiang, Y.G., Gorban, A., Laptev, I., Sukthankar, R., Shah, M.: The thumos challenge on action recognition for videos "in the wild". Computer Vision and Image Understanding (2017) +40. Jiang, A.Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D.S., Casas, D.d.l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al.: Mistral 7b. arXiv preprint arXiv:2310.06825 (2023) +41. Kim, C.D., Kim, B., Lee, H., Kim, G.: Audiocaps: Generating captions for audios in the wild. In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 119-132 (2019) +42. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Carlos Niebles, J.: Dense-captioning events in videos. In: ICCV (2017) +43. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: 2011 International conference on computer vision. pp. 2556-2563. IEEE (2011) +44. Lei, J., Berg, T.L., Bansal, M.: Qvhighlights: Detecting moments and highlights in videos via natural language queries (2021) +45. Li, G., Xu, Y., Hu, D.: Multi-scale attention for audio question answering. arXiv preprint arXiv:2305.17993 (2023) +46. Li, J., Li, D., Savarese, S., Hoi, S.C.H.: Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In: ICML (2022) +47. Li, J., Li, D., Xiong, C., Hoi, S.C.H.: Blip: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In: ICML (2022) + +48. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355 (2023) +49. Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Liu, Y., Wang, Z., Xu, J., Chen, G., Luo, P., et al.: Mvbench: A comprehensive multi-modal video understanding benchmark. arXiv preprint arXiv:2311.17005 (2023) +50. Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Wang, L., Qiao, Y.: Uniformerv2: Spatiotemporal learning by arming image vits with video uniformer. arXiv preprint arXiv:2211.09552 (2022) +51. Li, K., Wang, Y., Li, Y., Wang, Y., He, Y., Wang, L., Qiao, Y.: Unmasked teacher: Towards training-efficient video foundation models. arXiv preprint arXiv:2303.16058 (2023) +52. Li, T., Wang, L.: Learning spatiotemporal features via video and text pair discrimination. CoRR abs/2001.05691 (2020), https://arxiv.org/abs/2001.05691 +53. Lin, K.Q., Zhang, P., Chen, J., Pramanick, S., Gao, D., Wang, A.J., Yan, R., Shou, M.Z.: Univtg: Towards unified video-language temporal grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2794-2804 (2023) +54. Lin, T., Liu, X., Li, X., Ding, E., Wen, S.: Bmn: Boundary-matching network for temporal action proposal generation (2019) +55. Lipping, S., Sudarsanam, P., Drossos, K., Virtanen, T.: Clotho-aqa: A crowdsourced dataset for audio question answering. In: 2022 30th European Signal Processing Conference (EUSIPCO). pp. 1140-1144. IEEE (2022) +56. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023) +57. Liu, Y., Wang, L., Wang, Y., Ma, X., Qiao, Y.: Fineaction: A fine-grained video dataset for temporal action localization. IEEE Transactions on Image Processing 31, 6937-6950 (2022) +58. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021) +59. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing (2022) +60. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M.H., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability. ArXiv abs/2306.07207 (2023) +61. Maaz, M., Rasheed, H.A., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. ArXiv abs/2306.05424 (2023) +62. Mangalam, K., Akshulakov, R., Malik, J.: Egoschema: A diagnostic benchmark for very long-form video language understanding. In: Oh, A., Neumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) NeurIPS. pp. 46212-46244 (2023) +63. Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. In: CVPR (2019) +64. Mei, X., Meng, C., Liu, H., Kong, Q., Ko, T., Zhao, C., Plumbley, M.D., Zou, Y., Wang, W.: Wavcaps: A chatgpt-assisted weakly-labelled audio captioning dataset for audio-language multimodal research (2023) +65. Monfort, M., Zhou, B., Bargal, S.A., Andonian, A., Yan, T., Ramakrishnan, K., Brown, L.M., Fan, Q., Gutfreund, D., Vondrick, C., Oliva, A.: Moments in time dataset: One million videos for event understanding. TPAMI (2020) + +66. Moon, W., Hyun, S., Lee, S., Heo, J.P.: Correlation-guided query-dependency calibration in video representation learning for temporal grounding. arXiv preprint arXiv:2311.08835 (2023) +67. OpenAI: Gpt-4 technical report. ArXiv abs/2303.08774 (2023) +68. OpenAI: Gpt-4v(ison) system card. https://api_semanticscholar.org/ CorpusID:263218031 (2023) +69. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) +70. Patraucean, V., Smaira, L., Gupta, A., Continente, A.R., Markeeva, L., Dylan, Banarse, Malinowski, M., Yang, Y., Doersch, C., Matejovicova, T., Sulsky, Y., Antoine, Miech, Koppula, S., Fréchette, A., Klimczak, H., Koster, R., Zhang, J., Stephanie, Winkler, Aytar, Y., Osindero, S., Damen, D., Zisserman, A., Carreira, J.: Perception test: A diagnostic benchmark for multimodal models. In: NeurIPS (2023) +71. Piczak, K.J.: Esc: Dataset for environmental sound classification. In: Proceedings of the 23rd ACM international conference on Multimedia. pp. 1015-1018 (2015) +72. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: ICCV (2015) +73. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021) +74. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763. PMLR (2021) +75. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision. In: International Conference on Machine Learning. pp. 28492-28518. PMLR (2023) +76. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR (2020) +77. Rohrbach, A., Rohrbach, M., Tandon, N., Schiele, B.: A dataset for movie description. In: CVPR. pp. 3202-3212 (2015) +78. Ryali, C., Hu, Y.T., Bolya, D., Wei, C., Fan, H., Huang, P.Y., Aggarwal, V., Chowdhury, A., Poursaeed, O., Hoffman, J., Malik, J., Li, Y., Feichtenhofer, C.: Hiera: A hierarchical vision transformer without the bells-and-whistles (2023) +79. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012) +80. Sun, Q., Cui, Y., Zhang, X., Zhang, F., Yu, Q., Luo, Z., Wang, Y., Rao, Y., Liu, J., Huang, T., et al.: Generative multimodal models are in-context learners. arXiv preprint arXiv:2312.13286 (2023) +81. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023) +82. Sun, Q., Wang, J., Yu, Q., Cui, Y., Zhang, F., Zhang, X., Wang, X.: Eva-clip-18b: Scaling clip to 18 billion parameters. arXiv preprint arXiv:2402.04252 (2024) +83. Sun, Q., Yu, Q., Cui, Y., Zhang, F., Zhang, X., Wang, Y., Gao, H., Liu, J., Huang, T., Wang, X.: Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222 (2023) + +84. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023) +85. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. NeurIPS 35, 10078-10093 (2022) +86. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models. ArXiv abs/2302.13971 (2023) +87. Touvron, H., Martin, L., Stone, K.R., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D.M., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A.S., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I.M., Korenev, A.V., Koura, P.S., Lachaux, M.A., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. ArXiv abs/2307.09288 (2023) +88. Tran, D., xiu Wang, H., Torresani, L., Ray, J., LeCun, Y., Paluri, M.: A closer look at spatiotemporal convolutions for action recognition. In: CVPR (2018) +89. Wang, J., Ge, Y., Yan, R., Ge, Y., Lin, K.Q., Tsutsui, S., Lin, X., Cai, G., Wu, J., Shan, Y., et al.: All in one: Exploring unified video-language pre-training. In: CVPR. pp. 6598-6608 (2023) +90. Wang, L., Huang, B., Zhao, Z., Tong, Z., He, Y., Wang, Y., Wang, Y., Qiao, Y.: Videomae v2: Scaling video masked autoencoders with dual masking. In: CVPR (2023) +91. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Gool, L.V.: Temporal segment networks: Towards good practices for deep action recognition. In: ECCV (2016) +92. Wang, R., Chen, D., Wu, Z., Chen, Y., Dai, X., Liu, M., Yuan, L., Jiang, Y.G.: Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In: CVPR. pp. 6312-6322 (2023) +93. Wang, Y., He, Y., Li, Y., Li, K., Yu, J., Ma, X., Li, X., Chen, G., Chen, X., Wang, Y., et al.: Intervid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942 (2023) +94. Wang, Y., Li, K., Li, Y., He, Y., Huang, B., Zhao, Z., Zhang, H., Xu, J., Liu, Y., Wang, Z., Xing, S., Chen, G., Pan, J., Yu, J., Wang, Y., Wang, L., Qiao, Y.: Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191 (2022) +95. Xu, H., Ye, Q., Yan, M., Shi, Y., Ye, J., Xu, Y., Li, C., Bi, B., Qian, Q., Wang, W., et al.: mplug-2: A modularized multi-modal foundation model across text, image and video. arXiv preprint arXiv:2302.00402 (2023) +96. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettle-moyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084 (2021) +97. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: CVPR. pp. 5288-5296 (2016) + +98. Yan, S., Zhu, T., Wang, Z., Cao, Y., Zhang, M., Ghosh, S., Wu, Y., Yu, J.: Video-text modeling with zero-shot transfer from contrastive captioners. ArXiv abs/2212.04979 (2022) +99. Yang, L., Fan, Y., Xu, N.: Video instance segmentation. In: ICCV (2019) +100. Yang, M., Chen, G., Zheng, Y.D., Lu, T., Wang, L.: Basictad: an astounding rgb-only baseline for temporal action detection. Computer Vision and Image Understanding p. 103692 (2023) +101. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., Jiang, C., Li, C., Xu, Y., Chen, H., Tian, J., Qi, Q., Zhang, J., Huang, F.: mplug-owl: Modularization empowers large language models with multimodality (2023) +102. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 (2022) +103. Zablocki, E., Ben-Younes, H., Pérez, P., Cord, M.: Explainability of deep vision-based autonomous driving systems: Review and challenges. International Journal of Computer Vision 130(10), 2425-2452 (2022) +104. Zellers, R., Lu, J., Lu, X., Yu, Y., Zhao, Y., Salehi, M., Kusupati, A., Hessel, J., Farhadi, A., Choi, Y.: Merlot reserve: Neural script knowledge through vision and language and sound. In: CVPR. pp. 16375-16387 (2022) +105. Zeng, Z., Ge, Y., Tong, Z., Liu, X., Xia, S.T., Shan, Y.: Tvtsv2: Learning out-of-the-box spatiotemporal visual representations at scale. arXiv preprint arXiv:2305.14173 (2023) +106. Zhang, B., Yu, J., Fifty, C., Han, W., Dai, A.M., Pang, R., Sha, F.: Cotraining transformer with videos and images improves action recognition. ArXiv abs/2112.07175 (2021) +107. Zhang, C.L., Wu, J.Z., Li, Y.: Actionformer: Localizing moments of actions with transformers. ArXiv abs/2202.07925 (2022) +108. Zhao, H., Torralba, A., Torresani, L., Yan, Z.: Hacs: Human action clips and segments dataset for recognition and temporal localization. In: ICCV (2019) +109. Zhao, L., Gundavarapu, N.B., Yuan, L., Zhou, H., Yan, S., Sun, J.J., Friedman, L., Qian, R., Weyand, T., Zhao, Y., Hornung, R., Schroff, F., Yang, M.H., Ross, D.A., Wang, H., Adam, H., Sirotenko, M., Liu, T., Gong, B.: Videoprism: A foundational visual encoder for video understanding (2024) +110. Zhao, Y., Hessel, J., Yu, Y., Lu, X., Zellers, R., Choi, Y.: Connecting the dots between audio and text without parallel data through visual knowledge transfer. arXiv preprint arXiv:2112.08995 (2021) +111. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685 (2023) +112. Zhu, B., Lin, B., Ning, M., Yan, Y., Cui, J., Wang, H., Pang, Y., Jiang, W., Zhang, J., Li, Z., et al.: Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852 (2023) +113. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. ArXiv abs/2304.10592 (2023) +114. Zhu, W., Huang, Y., Xie, X., Liu, W., Deng, J., Zhang, D., Wang, Z., Liu, J.: Autoshot: A short video dataset and state-of-the-art shot boundary detection. In: CVPR. pp. 2237-2246 (2023) \ No newline at end of file diff --git a/2024/InternVideo2_ Scaling Foundation Models for Multimodal Video Understanding/images.zip b/2024/InternVideo2_ Scaling Foundation Models for Multimodal Video Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..80d658bb3f8519dc388cee9db0d6e115673ac6ab --- /dev/null +++ b/2024/InternVideo2_ Scaling Foundation Models for Multimodal Video Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3a4861b4105f9030d80489a9a305b082b798490ed86c534174eeb700e3381b1 +size 605776 diff --git a/2024/InternVideo2_ Scaling Foundation Models for Multimodal Video Understanding/layout.json b/2024/InternVideo2_ Scaling Foundation Models for Multimodal Video Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..038ceecaf94f2661e5024cbe2548e227b8e54002 --- /dev/null +++ b/2024/InternVideo2_ Scaling Foundation Models for Multimodal Video Understanding/layout.json @@ -0,0 +1,11361 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 150, + 111, + 463, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 111, + 463, + 148 + ], + "spans": [ + { + "bbox": [ + 150, + 111, + 463, + 148 + ], + "type": "text", + "content": "InternVideo2: Scaling Foundation Models for Multimodal Video Understanding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "spans": [ + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "text", + "content": "Yi Wang" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "text", + "content": ", Kunchang Li" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "inline_equation", + "content": "^{3,1}" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "text", + "content": ", Xinhao Li" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "inline_equation", + "content": "^{2,1}" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "text", + "content": ", Jiashuo Yu" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "text", + "content": ", Yinan He" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "text", + "content": ", Guo Chen" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "inline_equation", + "content": "^{2,1}" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "text", + "content": ", Baoqi Pei" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "text", + "content": ", Rongkun Zheng" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "text", + "content": ", Zun Wang" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "text", + "content": ", Yansong Shi" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "text", + "content": ", Tianxiang Jiang" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "text", + "content": ", Songze Li" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "text", + "content": ", Jilan Xu" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "text", + "content": ", Hongjie Zhang" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "text", + "content": ", Yifei Huang" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "text", + "content": ", Yu Qiao" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "text", + "content": ", Yali Wang" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "inline_equation", + "content": "^{\\dagger 3,1}" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "text", + "content": ", and Limin Wang" + }, + { + "bbox": [ + 147, + 167, + 466, + 217 + ], + "type": "inline_equation", + "content": "^{\\dagger 2,1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 187, + 225, + 426, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 225, + 426, + 248 + ], + "spans": [ + { + "bbox": [ + 187, + 225, + 426, + 248 + ], + "type": "text", + "content": "1 OpenGVLab, Shanghai AI Laboratory, Shanghai, China \n2 Nanjing University, Nanjing, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 161, + 248, + 452, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 248, + 452, + 281 + ], + "spans": [ + { + "bbox": [ + 161, + 248, + 452, + 281 + ], + "type": "text", + "content": "3 Shenzhen Institutes of Advanced Technology, CAS, Shenzhen, China {wangyi,qiaoyu,wangyali,wanglimin}@pjlab.org.cn https://github.com/OpenGVLab/InternVideo/tree/main/InternVideo?" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 311, + 476, + 401 + ], + "blocks": [ + { + "bbox": [ + 133, + 311, + 476, + 401 + ], + "lines": [ + { + "bbox": [ + 133, + 311, + 476, + 401 + ], + "spans": [ + { + "bbox": [ + 133, + 311, + 476, + 401 + ], + "type": "image", + "image_path": "e510ad049ec3410b35cba384edaeefda48dbe3849da62de1fbc1169937efa682.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 411, + 482, + 445 + ], + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 445 + ], + "type": "text", + "content": "Fig. 1: InternVideo2 yields strong transferable visual and visual-linguistic representations for action recognition, video-text understanding, to video-centric dialogue. It also exhibits capability of long-form video understanding and procedure-aware reasoning." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 484, + 455, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 484, + 455, + 628 + ], + "spans": [ + { + "bbox": [ + 159, + 484, + 455, + 628 + ], + "type": "text", + "content": "Abstract. We introduce InternVideo2, a new family of video foundation models (ViFM) that achieve the state-of-the-art results in video recognition, video-text tasks, and video-centric dialogue. Our core design is a progressive training approach that unifies the masked video modeling, crossmodal contrastive learning, and next token prediction, scaling up the video encoder size to 6B parameters. At the data level, we prioritize spatiotemporal consistency by semantically segmenting videos and generating video-audio-speech captions. This improves the alignment between video and text. Through extensive experiments, we validate our designs and demonstrate superior performance on over 60 video and audio tasks. Notably, our model outperforms others on various video-related dialogue and long video understanding benchmarks, highlighting its ability to reason and comprehend longer contexts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 638, + 443, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 638, + 443, + 649 + ], + "spans": [ + { + "bbox": [ + 159, + 638, + 443, + 649 + ], + "type": "text", + "content": "Keywords: Video Foundation Model " + }, + { + "bbox": [ + 159, + 638, + 443, + 649 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 638, + 443, + 649 + ], + "type": "text", + "content": " Progressive Training " + }, + { + "bbox": [ + 159, + 638, + 443, + 649 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 638, + 443, + 649 + ], + "type": "text", + "content": " MLLM." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 331, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 331, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 331, + 665 + ], + "type": "text", + "content": "*Equal contribution. †Corresponding authors." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 230, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 230, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 230, + 128 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 235 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 235 + ], + "type": "text", + "content": "Learning transferrable spatiotemporal representations is a critical research area in computer vision, holding diverse applications across domains such as video search [31], game control [7], robotic learning [26], self-driving [103], and scientific studies [84]. Recently, the advancement of Large Language Models (LLMs) [6, 67, 86, 87] and their multimodal variations (MLLMs) [33, 56, 68, 84] have had a profound impact on vision research and other disciplines. Embedding videos effectively into these large models and harnessing their capabilities to enhance video understanding performance has emerged as pivotal tasks [48, 61]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 235, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 482, + 365 + ], + "type": "text", + "content": "Previous research has identified several effective learning schemes for video representations, including reconstructing videos with masked inputs [28,37,85,90], aligning videos with languages [51,52,96,98], and predicting the next token using videos [1,49,83]. These approaches have turned out to be complementary and can be unified through a progressive training scheme. Notably, methods such as InternVideo [94], UMT [51], and VideoPrism [109] have utilized a two-stage training approach involving masked reconstruction and multimodal contrastive learning, leading to improved performance in downstream tasks. Following this line, we aim to further extend this progressive training scheme by incorporating video-based next token prediction and scaling the entire training process, including models and data, to build a new family of video foundation models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 366, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 581 + ], + "type": "text", + "content": "The proposed video foundation model, coined as InternVideo2, is built through a progressive training scheme. The learning involves three stages: (1) capturing spatiotemporal structure via unmasked reconstruction, (2) aligning with semantics from other modalities, and (3) enhancing its open-ended dialogue power through next token prediction. In the initial stage, the model learns to reconstruct the unmasked video tokens, allowing the video encoder to develop basic spatiotemporal perception capability. To estimate the existing tokens, vision encoders (InternViT [17] and VideoMAE-g [90]) trained differently are employed as proxies. In the next stage of crossmodal learning, the architecture is expanded to include audio and text encoders. This not only improves the alignment between videos and text but also endows InternVideo2 the ability to handle video-audio tasks. By incorporating these additional modalities, the model's understanding of videos is enriched and aligned with their semantics. Finally, in the next-token prediction stage, a video-centric dialogue system and the corresponding instruction-finetuning dataset are built to further tune the InternVideo2. By connecting InternVideo2 to LLMs, the video encoder is further updated through next-token prediction training, enhancing its ability for open-ended tasks such as VQA and video description." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "For the training of InternVideo2, we emphasize the spatiotemporal consistency and labeling quality in the data. We build a large-scale multimodal video-centric dataset consisting of 402M data entries, which includes 2M videos, 50M video-text pairs (from WebVid [3] and InternVid [93]), 50M video-audio-speech-text pairs (InternVid2), and 300M image-text pairs. Specifically, for InternVid2, we segment videos into clips semantically and focus on recalibrating the clip descriptions using three modalities: audio, video, and speech. We first generate captions for" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "these three modalities separately. Then individual captions are fused together to create a more comprehensive description, which will improve the model's ability to comprehend and interpret the video accurately." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 272 + ], + "type": "text", + "content": "We evaluate InternVideo2 across a wide range of video-related tasks. These tasks span from basic spatiotemporal perception, such as action recognition, to high-level reasoning tasks, such as long video or procedure-aware question-answering (QA), as given in Fig. 1. The results (in Sec. 5) demonstrate that InternVideo2 achieves the state-of-the-art performance on multiple tasks and is able to analyze and reason over sequences of actions. This top performance signifies its capability to effectively capture and understand video content. These empirical findings validate that InternVideo2 could serves as a general video encoder for future exploration in video understanding. In summary, our contributions to video understanding are as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 277, + 480, + 420 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 132, + 277, + 480, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 480, + 324 + ], + "type": "text", + "content": "- This paper introduces InternVideo2, a competitive family of video foundation models that leverages masked reconstruction, crossmodal contrastive learning, and next token prediction to make model perceptive, semantic, and capable of reasoning in video understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 326, + 480, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 372 + ], + "type": "text", + "content": "- InternVideo2 achieves the state-of-the-art performance for more than 60 video / audio tasks. Our model demonstrates superior performance in video-related dialogue and long video understanding, highlighting its potential in modeling high-level world knowledge." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 373, + 480, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 373, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 480, + 420 + ], + "type": "text", + "content": "- We provide an enhanced dataset to train InternVideo2. This includes the validation and incorporation of audio data during training, as well as the improved captioning method. These improvements result in significant enhancements in model performance and generalization ability." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 437, + 237, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 237, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 237, + 450 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "content": "Video Foundation Models. Studies on learning video foundation models becomes increasingly crucial considering its wide applications [?, 51, 52, 85, 89, 90, 92, 94, 96, 98, 109]. Typical methods in building video foundation models (ViFM) include video-text contrastive learning [52, 94, 96], masked video modeling [30, 85, 90, 94], and next token prediction [1, 80, 83]. Specifically, All-in-one [89] utilized a single backbone with unified multiple pretraining objectives. On the other hand, UMT [51] combined masked modeling with video-text contrastive learning, demonstrating strong performance in both action recognition and videolanguage tasks. Another approach is mPLUG-2 [95], which introduced a new design for modulating different modalities. It shared a common module across modalities to enhance relations while incorporating modality-specific modules for discrimination. In addition to video-text pretraining, researchers have also explored the use of audio information in videos to improve performance. MERLOT Reserve [104] learned video representations using a large-scale dataset of videospeech-transcript pairs. VALOR [14] employed independent encoders for video, audio, and text and trains a joint visual-audio-text representation. VAST [15] constructed an audio-visual-speech dataset and develops a multimodal backbone" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "InternVideo2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 479, + 178 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 479, + 178 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 479, + 178 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 479, + 178 + ], + "type": "image", + "image_path": "97027eeccf7e96697083cc75455f70ec1857ca205740c8b248bd93fb09a5b82e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 189, + 482, + 234 + ], + "lines": [ + { + "bbox": [ + 130, + 189, + 482, + 234 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 482, + 234 + ], + "type": "text", + "content": "Fig. 2: Framework of InternVideo2. It consists of three consecutive training phases: unmasked video token reconstruction, multimodal contrastive learning, and next token prediction. In stage 1, the video encoder is trained from scratch, while in stages 2 and 3, it is initialized from the version used in the previous stage." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 257, + 479, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 479, + 293 + ], + "type": "text", + "content": "that excels in video-audio-related tasks. VideoPrism [109] combined video-text contrastive learning and video token reconstruction on a combination of public and proprietary videos, achieving leading results across various video tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 293, + 482, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 401 + ], + "type": "text", + "content": "Multimodal Large Language Models. With advances in large language models (LLMs) [6, 24, 76], their multimodal versions (MLLMs) is becoming popular as it can handle open-world tasks. Seminal works like Flamingo [1] showed outstanding zero/few-shots performances over a range of multimodal tasks [36, 63, 72, 97]. Public MLLMs [33, 56, 113] such as LLaVA [56] and InstructBLIP [21] proposed to use visual instruction-tuning data to improve the visual dialogue ability. Some video-centric MLLMs have been proposed, such as VideoChat [48], VideoChatGPT [61] and Valley [60], by using instruction data to connect video encoders to LLMs for open-world video understanding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 420, + 202, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 202, + 433 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 202, + 433 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 445, + 481, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 481, + 483 + ], + "type": "text", + "content": "We learn InternVideo2 in three stages, illustrated in Fig. 2. The stages include spatiotemporal token reconstruction, video-audio-speech-language contrastive learning, and connecting to a large language model (LLM) for joint training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 495, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 591 + ], + "type": "text", + "content": "Video Encoder. The video encoder used in InternVideo2 follows the Vision Transformer (ViT) [25] and includes additional projection layers for distillation. Inspired by previous works [17, 102], we introduce attention pooling to the ViT. For input videos, we sparsely sample 8 frames [91] and perform a " + }, + { + "bbox": [ + 130, + 495, + 482, + 591 + ], + "type": "inline_equation", + "content": "14 \\times 14" + }, + { + "bbox": [ + 130, + 495, + 482, + 591 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 495, + 482, + 591 + ], + "type": "inline_equation", + "content": "h \\times w" + }, + { + "bbox": [ + 130, + 495, + 482, + 591 + ], + "type": "text", + "content": ") spatial downsampling. These spatiotemporal tokens are then concatenated with a class token and combined with 3D position embeddings. The hyperparameters used for ViT-1B and -6B are the same as CoCa-1B [102] and InternViT-6B [17], respectively. The details of ViT-6B architecture are given in Supp." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 609, + 405, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 405, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 405, + 621 + ], + "type": "text", + "content": "3.1 Stage1: Reconstructing Unmasked Video Tokens" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "We exploit two expert models to guide the video encoder to conduct token-level reconstruction for unmasked areas. Specifically, we adopt InternVL-6B [17] and VideoMAEv2-g [90] to transfer unmasked knowledge via simple projection layers." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "content": "When training, we input the full videos into different teachers and mask out " + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 131, + 116, + 482, + 177 + ], + "type": "text", + "content": " of the tokens frame by frame, under the semantic guidance of the multimodal model InternVL and motion-aware model VideoMAEv2. We only align the unmasked tokens, by minimizing their mean squared error (MSE) between student and teachers. The learning objective is to reconstruct the remaining tokens as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 178, + 186, + 482, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 186, + 482, + 209 + ], + "spans": [ + { + "bbox": [ + 178, + 186, + 482, + 209 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\frac {1}{Z} \\sum \\left(\\alpha_ {1} | f ^ {V} (\\mathbf {V} _ {p}) - h (\\mathbf {V} _ {p}) | ^ {2} + \\alpha_ {2} | f ^ {V} (\\mathbf {V} _ {p}) - g (\\mathbf {V} _ {p}) | ^ {2}\\right), \\tag {1}", + "image_path": "6183d1f66ccd14e5b9f986c73e6ad9759058f9d0d3a832ece9e825ff92fde301.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 223, + 482, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 223, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 131, + 223, + 482, + 270 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 223, + 482, + 270 + ], + "type": "inline_equation", + "content": "f^V" + }, + { + "bbox": [ + 131, + 223, + 482, + 270 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 223, + 482, + 270 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 131, + 223, + 482, + 270 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 223, + 482, + 270 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 131, + 223, + 482, + 270 + ], + "type": "text", + "content": " are our video encoder, InternViT-6B [17], and ViT-g of VideoMAEv2, respectively. " + }, + { + "bbox": [ + 131, + 223, + 482, + 270 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 131, + 223, + 482, + 270 + ], + "type": "text", + "content": " stands for the token index and " + }, + { + "bbox": [ + 131, + 223, + 482, + 270 + ], + "type": "inline_equation", + "content": "f(\\mathbf{V}_p)" + }, + { + "bbox": [ + 131, + 223, + 482, + 270 + ], + "type": "text", + "content": " is the corresponding token extracted by InternVideo2 for input video " + }, + { + "bbox": [ + 131, + 223, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 131, + 223, + 482, + 270 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 223, + 482, + 270 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 131, + 223, + 482, + 270 + ], + "type": "text", + "content": " is the normalization factor. " + }, + { + "bbox": [ + 131, + 223, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 131, + 223, + 482, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 223, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 131, + 223, + 482, + 270 + ], + "type": "text", + "content": " balance the influence between the employed models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 272, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 272, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 131, + 272, + 482, + 344 + ], + "type": "text", + "content": "In implementation, we align the last 6 layers of InternVL, last 4 layers of VideoMAEv2, and the final output token of InternVL. Different losses are simply summed for optimization. After pretraining, we drop those projection layers and only use the basic encoder. Compared with only using the multimodal model in UMT and VideoPrism, our strategy makes the vision encoder multimodal-friendly as well as enhances its temporal sensitivity for action modeling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 364, + 394, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 394, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 394, + 376 + ], + "type": "text", + "content": "3.2 Stage 2: Aligning Video to Audio-Speech-Text" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 386, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 386, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 131, + 386, + 482, + 517 + ], + "type": "text", + "content": "We exploit the correspondence between video and audio, speech, and text to encourage InternVideo2 to learn more semantics. In practice, InternVideo2 has a huge video encoder, and its employed audio and text encoders are relatively lightweight. The used audio encoder is a 12-layer transformer initialized with BEATs [13] (90M). It takes in audio features, which are 64-dimensional log Mel filterbank spectrograms using a 25ms Hamming window, transformed from 10-second-long clips, padding with zeros. For the text and speech encoders, we initialize the text encoder and multimodal decoder using BERT-Large [24]. Specifically, we utilize the initial 19 layers of BERT-Large as the text encoder, with the subsequent 5 layers equipped with cross-attention layers serving as the multimodal decoder." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 518, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 518, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 131, + 518, + 482, + 554 + ], + "type": "text", + "content": "For pretraining objectives, we establish alignment across different modalities via text, including video, audio, image, and speech. We employ crossmodal contrastive and matching losses with masked language reconstruction loss as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 245, + 566, + 481, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 566, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 245, + 566, + 481, + 578 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {C O N}} + \\mathcal {L} _ {\\mathrm {M A C}} + \\mathcal {L} _ {\\mathrm {M L M}}, \\tag {2}", + "image_path": "38d34f8994b64dfc0064e6723533763be5782d45bf6b1de9a1ccde2e6da6c09c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 587, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 587, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 131, + 587, + 482, + 612 + ], + "type": "text", + "content": "The employed " + }, + { + "bbox": [ + 131, + 587, + 482, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MAC}}" + }, + { + "bbox": [ + 131, + 587, + 482, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 587, + 482, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MLM}}" + }, + { + "bbox": [ + 131, + 587, + 482, + 612 + ], + "type": "text", + "content": " are standard loss from [19]. Specifically, the crossmodal contrastive learning is given as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 136, + 624, + 479, + 652 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 624, + 479, + 652 + ], + "spans": [ + { + "bbox": [ + 136, + 624, + 479, + 652 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C O N}} = \\sum_ {M, T _ {M ^ {\\prime}}} \\mathcal {L} _ {\\mathrm {C O N}} (M, T _ {M ^ {\\prime}}) = - \\sum_ {M, T _ {M ^ {\\prime}}} \\left(\\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (\\sin \\left(f _ {i} ^ {M} , f _ {i} ^ {T _ {M ^ {\\prime}}}\\right) / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (\\sin \\left(f _ {i} ^ {M} , f _ {j} ^ {T _ {M ^ {\\prime}}}\\right) / \\tau)} + \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (\\sin \\left(f _ {i} ^ {T _ {M ^ {\\prime}}} , f _ {i} ^ {M}\\right) / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (\\sin \\left(f _ {i} ^ {T _ {M ^ {\\prime}}} , f _ {j} ^ {M}\\right) / \\tau)}\\right), \\tag {3}", + "image_path": "a4b23cd258f2f26b48d56d5bf3004f320570abd59f394c425480a22ee1371b98.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 135, + 653, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 653, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 135, + 653, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 135, + 653, + 482, + 666 + ], + "type": "inline_equation", + "content": "f^V" + }, + { + "bbox": [ + 135, + 653, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 135, + 653, + 482, + 666 + ], + "type": "inline_equation", + "content": "f^T" + }, + { + "bbox": [ + 135, + 653, + 482, + 666 + ], + "type": "text", + "content": " denote the learned video and text embeddings, respectively." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "InternVideo2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "T_{M^{\\prime}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " indicates the modality of input signals and the text descriptions describing, respectively. " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\sin (\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " computes the cosine similarity between two features. " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " is the learnable temperature." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 152, + 311, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 152, + 311, + 164 + ], + "spans": [ + { + "bbox": [ + 146, + 152, + 311, + 164 + ], + "type": "text", + "content": "For the matching part, it is given as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 192, + 171, + 481, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 171, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 192, + 171, + 481, + 184 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M A C}} = - y \\log f _ {p} (\\mathbf {V}, \\mathbf {T}) - (1 - y) \\log \\left(1 - f _ {p} (\\mathbf {V}, \\mathbf {T})\\right), \\tag {4}", + "image_path": "a056a5043098ba7fb0ee8d32cebea2cfc05344b0468f3453407721022b40c140.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 189, + 481, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 481, + 213 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 189, + 481, + 213 + ], + "type": "inline_equation", + "content": "f_{p}(\\mathbf{V},\\mathbf{T})" + }, + { + "bbox": [ + 130, + 189, + 481, + 213 + ], + "type": "text", + "content": " computes the matching likelihood between " + }, + { + "bbox": [ + 130, + 189, + 481, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 130, + 189, + 481, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 189, + 481, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 189, + 481, + 213 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 189, + 481, + 213 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 189, + 481, + 213 + ], + "type": "text", + "content": " denotes whether the given video and text are paired " + }, + { + "bbox": [ + 130, + 189, + 481, + 213 + ], + "type": "inline_equation", + "content": "(y = 1)" + }, + { + "bbox": [ + 130, + 189, + 481, + 213 + ], + "type": "text", + "content": " or not " + }, + { + "bbox": [ + 130, + 189, + 481, + 213 + ], + "type": "inline_equation", + "content": "(y = 0)" + }, + { + "bbox": [ + 130, + 189, + 481, + 213 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 214, + 363, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 214, + 363, + 225 + ], + "spans": [ + { + "bbox": [ + 146, + 214, + 363, + 225 + ], + "type": "text", + "content": "The employed masked language modeling loss is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 247, + 232, + 481, + 246 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 232, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 247, + 232, + 481, + 246 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M L M}} = - \\log f _ {p} ^ {T} \\left(\\mathbf {T} _ {j} \\mid \\mathbf {T} _ {< j}\\right), \\tag {5}", + "image_path": "57bd5257412b7c817644522e97aa284d2dddf19dff308c9396d74ef6dc333bcb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 251, + 481, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 481, + 276 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 481, + 276 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 251, + 481, + 276 + ], + "type": "inline_equation", + "content": "f_{p}^{T}(\\mathbf{T}_{j}|\\mathbf{T}_{< j})" + }, + { + "bbox": [ + 130, + 251, + 481, + 276 + ], + "type": "text", + "content": " computes the likelihood of the " + }, + { + "bbox": [ + 130, + 251, + 481, + 276 + ], + "type": "inline_equation", + "content": "j_{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 251, + 481, + 276 + ], + "type": "text", + "content": " text token based on the previous ones. Here " + }, + { + "bbox": [ + 130, + 251, + 481, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 251, + 481, + 276 + ], + "type": "text", + "content": " refers to video captions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 277, + 481, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 481, + 323 + ], + "type": "text", + "content": "To improve the training efficiency, we employ the masked learning strategy, aligning unmasked video tokens to tokens from other modalities first, then using full video tokens reconstruction shortly. Specifically, it consists of two steps as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 333, + 482, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 482, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 482, + 393 + ], + "type": "text", + "content": "Aligning Masked Visual-Language-Audio. We freeze the audio encoder and focus on aligning visual, audio, and text features. For pre-training, we use a comprehensive set of image, video, and audio-video data. The combinations of modalities used are represented as " + }, + { + "bbox": [ + 130, + 333, + 482, + 393 + ], + "type": "inline_equation", + "content": "\\{M,T_{M^{\\prime}}\\} \\in \\{\\{I,T_I\\},\\{V,T_V\\},\\{V,T_{VAS}\\},\\{VA,T_{VAS}\\}\\}" + }, + { + "bbox": [ + 130, + 333, + 482, + 393 + ], + "type": "text", + "content": " where each pair denotes the concatenated features from the respective modalities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "text", + "content": "Unmasked Visual-Audio-Language Post-Pretraining. We freeze the vision encoder to jointly align audio, visual, and text features. Post-pretraining is conducted using a smaller subset of image and video data (25M samples), along with the full set of audio (0.5M samples) and audio-video data (50M samples). Since the parameters of the largest ViT-6B model are frozen, we do not use masking strategies in this phase to ensure consistency with the inference process and to minimize any performance degradation in downstream tasks. The modality combinations used here are " + }, + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\{M, T_{M'}\\} \\in \\{\\{I, T_I\\}, \\{V, T_V\\}, \\{A, T_A\\}, \\{V, T_{VAS}\\}, \\{VA, T_{VA}\\}\\}" + }, + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 514, + 453, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 453, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 453, + 526 + ], + "type": "text", + "content": "3.3 Stage3: Predicting Next Token with Video-Centric Inputs" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 531, + 482, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 603 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 603 + ], + "type": "text", + "content": "To further enrich the semantics embedded in InternVideo2 and improve its support for video-centric dialogue, we tune it by connecting it to a LLM with QFormer design [46, 47]. We employ the progressive learning scheme in [49] by using InternVideo2 as the video encoder and train a video blip for communicating with open-sourced LLM [40, 111]. In training, the video encoder will be updated. Detailed training receipts are the same with VideoChat2 [49]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 619, + 297, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 619, + 297, + 632 + ], + "spans": [ + { + "bbox": [ + 131, + 619, + 297, + 632 + ], + "type": "text", + "content": "4 Multimodal Video Data" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "We list our training data in Tab. 1. Among the datasets used, KMash and InternVid2 are newly built, whereas the rest are public available." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 135, + 443, + 223 + ], + "blocks": [ + { + "bbox": [ + 157, + 114, + 455, + 126 + ], + "lines": [ + { + "bbox": [ + 157, + 114, + 455, + 126 + ], + "spans": [ + { + "bbox": [ + 157, + 114, + 455, + 126 + ], + "type": "text", + "content": "Table 1: Summary of datasets used in InternVideo2 pretraining process." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 135, + 443, + 223 + ], + "lines": [ + { + "bbox": [ + 167, + 135, + 443, + 223 + ], + "spans": [ + { + "bbox": [ + 167, + 135, + 443, + 223 + ], + "type": "table", + "html": "
Pretraining Stage DatasetDomain# of clipsAnnotation
Stage 1KMashWeb Video2M-
Stage 2 (img-text)LAION, etcWeb Image300MAlt-text / Generated Caps
Stage 2 (vid-text)WebVid2MWeb Video250kAlt-text
WebVid10MWeb Video9.7MAlt-text
InternVidYoutube Video40MGenerated Caption
InternVid2Youtube Video50MGenerated Caption
Stage 3LLaVA, etcWeb Image/Video2.1MConversation, QA
", + "image_path": "b16688687ffe4a779a953f1a9b3df0795a5b6a4162d79a50437a6ff15ac849ce.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 151, + 232, + 465, + 356 + ], + "blocks": [ + { + "bbox": [ + 151, + 232, + 465, + 356 + ], + "lines": [ + { + "bbox": [ + 151, + 232, + 465, + 356 + ], + "spans": [ + { + "bbox": [ + 151, + 232, + 465, + 356 + ], + "type": "image", + "image_path": "a42554fc99b0bb401df147969de40c3b2d0a527c4be924025f4127a4531bac0c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 364, + 482, + 398 + ], + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 398 + ], + "type": "text", + "content": "Fig. 3: The framework of our video multimodal annotation system, called VidCap, consists of four main components: video, audio, and speech captioners, along with a LLM for integrating captions from these modalities." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 420, + 376, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 420, + 376, + 432 + ], + "spans": [ + { + "bbox": [ + 131, + 420, + 376, + 432 + ], + "type": "text", + "content": "4.1 Video-only Data for Masked Autoencoders" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "type": "text", + "content": "We curate a new video set without labels named " + }, + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "type": "text", + "content": "-Mash from action recognition datasets [10, 35, 38, 65, 108], as detailed in Supp. It encompasses a wide range of video types, including first- and third-person perspectives, with both short and long duration, and featuring various settings. Further, we give " + }, + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "type": "text", + "content": "-" + }, + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\text{Mash}_{2M}" + }, + { + "bbox": [ + 130, + 439, + 482, + 499 + ], + "type": "text", + "content": " with additionally sourced and selected 844K videos from YouTube for diversity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 516, + 350, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 516, + 350, + 528 + ], + "spans": [ + { + "bbox": [ + 131, + 516, + 350, + 528 + ], + "type": "text", + "content": "4.2 Videos with Audio-Speech Modalities" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 534, + 482, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 631 + ], + "type": "text", + "content": "We build a multimodal video dataset, coined as InternVid2, with video-audiospeech information and their descriptions for strengthening video perception via other modalities. It consists of 100M videos along with their VAS captions. In InternVid2, we collect videos from several sources (detailed in the supp), segment them into clips, and automatically annotate them based on their unimodal or crossmodal inputs. We highlight the importance of temporal segmentation in clip generation and our video multimodal annotation system. We find refining them leads to notable downstream improvements." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Temporal Consistency Matters. We employ a temporal boundary detection model AutoShot [114] to segment videos into clips instead of SceneDet filter from" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "InternVideo2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 152 + ], + "type": "text", + "content": "FFmpeg. It predicts boundaries based on temporal semantic variations rather than pixel differences, and is able to generate semantically complete cuts without mixing extra frames with inconsistent context." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 166, + 482, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 166, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 166, + 482, + 310 + ], + "type": "text", + "content": "Video Multimodal Annotation. We design a video multimodal annotation system VidCap to give proper unimodal and crossmodal descriptions for textualizing videos from different perceptions. It automatically captions visual, audio, and speech of InternVid2, then it corrects them and fuses them for cross-modal captions via LLM. The system frame is given in Fig. 3. VidCap has independent video, audio, speech captioner, and a LLM for caption refinement and fusion. For video, speech, and caption postprocessing, we employ existing methods as the video captioning pipeline in [93], WhisperV2-large model [75], and Vicuna-1.5 [111]. For audio, we craft a audio captioner upon VideoChat [48], as we find no open-sourced one. It extracts audio features from inputs by Beats [13]. We learn it by only tuning its Qformer using a combination of the large-scale audio-text corpus WavCaps [64] dataset. Details are given in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 331, + 230, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 230, + 345 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 230, + 345 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 358, + 481, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 407 + ], + "type": "text", + "content": "In our evaluation of InternVideo2, we assess the models from three learning stages. The evaluation covers a wide range of tasks, including video recognition, video retrieval, question-answering, and more. It includes various scenarios such as zero-shot learning, finetuning, and linear probing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 407, + 492, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 492, + 454 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 492, + 454 + ], + "type": "text", + "content": "For InternVideo2 trained in stage 1, 2, and 3, we denote them with InternVideo2" + }, + { + "bbox": [ + 132, + 407, + 492, + 454 + ], + "type": "inline_equation", + "content": "_{s1}" + }, + { + "bbox": [ + 132, + 407, + 492, + 454 + ], + "type": "text", + "content": ", InternVideo2" + }, + { + "bbox": [ + 132, + 407, + 492, + 454 + ], + "type": "inline_equation", + "content": "_{s2}" + }, + { + "bbox": [ + 132, + 407, + 492, + 454 + ], + "type": "text", + "content": ", and InternVideo2" + }, + { + "bbox": [ + 132, + 407, + 492, + 454 + ], + "type": "inline_equation", + "content": "_{s3}" + }, + { + "bbox": [ + 132, + 407, + 492, + 454 + ], + "type": "text", + "content": ", respectively. We also learn a CLIP-style InternVideo2 indicated by InternVideo2" + }, + { + "bbox": [ + 132, + 407, + 492, + 454 + ], + "type": "inline_equation", + "content": "_{clip}" + }, + { + "bbox": [ + 132, + 407, + 492, + 454 + ], + "type": "text", + "content": ". It is post-pretrained from InternVideo2" + }, + { + "bbox": [ + 132, + 407, + 492, + 454 + ], + "type": "inline_equation", + "content": "_{s2}" + }, + { + "bbox": [ + 132, + 407, + 492, + 454 + ], + "type": "text", + "content": " by only preserving video and text encoders and contrastive loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 455, + 481, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 455, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 132, + 455, + 481, + 539 + ], + "type": "text", + "content": "Each training stage of InternVideo2-6B uses different configurations and resources. In the first stage, we employ 256 NVIDIA A100 GPUs and train the model for 18 days. The second stage also utilizes 256 A100 GPUs and spans a training period of 14 days. Finally, in the third stage, we use 64 A100 GPUs and train the model for 3 days. We introduce DeepSpeed and FlashAttention [22] for training and inference. More implementation details, experiment results, and ablations are given in the supp." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 559, + 261, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 261, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 261, + 571 + ], + "type": "text", + "content": "5.1 Video Classification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 581, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 482, + 665 + ], + "type": "text", + "content": "Action Recognition. We test InternVideo2 on Kinetics (i.e., K400, 600 and 700 [8-10]), Moments in Time V1 (MiT) [65], Something-Something V2 (SSv2) [35], UCF [79], HMDB [43], Charades [32], ActivityNet [38] (ANet) and HACS [108]. We evaluate in four settings: (a) end-to-end finetuning the whole backbone; (b) attentive probing is similar to linear pooling, but extra trains the attention pooling layer [102]. (c) linear probing which freezes the backbone and only trains the task head; and (d) zero-shot." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 157, + 477, + 268 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 147 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 147 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 147 + ], + "type": "text", + "content": "Table 2: End-to-end finetuning action recognition results (top-1 accuracy) on Kinetics, SomethingSomething, and Moments in Time. " + }, + { + "bbox": [ + 132, + 114, + 481, + 147 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 132, + 114, + 481, + 147 + ], + "type": "text", + "content": " denotes that the result is achieved with different resolutions or frame rates." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 157, + 477, + 268 + ], + "lines": [ + { + "bbox": [ + 136, + 157, + 477, + 268 + ], + "spans": [ + { + "bbox": [ + 136, + 157, + 477, + 268 + ], + "type": "table", + "html": "
MethodTraining DataSettingK400K600K700Sth-Sthv2MiTANetHACS
CoVeR [106]IV-3B16 × 44887.187.979.870.846.1--
Hiera-H [78]V-0.25M16 × 22487.888.881.176.5---
CoCa-g [102]I-3B16 × 57688.989.482.7-49.0--
MTV-H [90]IV-370M32 × 28089.990.383.4----
VideoMAEv2-g [90]V-1.35M64 × 26690.089.9-77.0†---
V-JEPA-H [4]V-2M16 × 224---77.0---
MVD-H [92]IV-1.25M16 × 224---77.3---
UniFormerV2-L [50]IV-401M64 × 33690.090.182.773.0†47.8†94.795.4
InternVideo [94]V-12Mensemble91.191.384.077.2---
InternVideo2s1-1BIV-1.1M8 × 22491.391.485.077.150.8--
InternVideo2s1-1BIV-1.1M16 × 22491.691.685.477.150.9--
InternVideo2s1-6BIV-2M8 × 22491.991.785.777.551.0--
InternVideo2s1-6BIV-2M16 × 22492.191.985.977.451.295.997.0
", + "image_path": "51d24fadcfbca8953b2ca4bfff5099583a09486dcd91746c2fc466d9682abf8c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 153, + 312, + 458, + 412 + ], + "blocks": [ + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "type": "text", + "content": "Table 3: Attentive probing recognition results (top-1 accuracy) on Kinetics-400/600/700, Moments in Time and Something-Something V2." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 312, + 458, + 412 + ], + "lines": [ + { + "bbox": [ + 153, + 312, + 458, + 412 + ], + "spans": [ + { + "bbox": [ + 153, + 312, + 458, + 412 + ], + "type": "table", + "html": "
MethodTraining DataSettingK400K600K700MiTSSV2
UMT-L [51]IV-25M-82.8--40.354.5
VideoMAEv2-g [90]V-1.35M-82.1--35.056.1
V-JEPA-H [4]V-2M16 × 38481.9---72.2
DINOv2-g [69]I-142M16 × 22483.4---50.0
VideoPrism-g [109]V-619M16 × 28887.2--45.568.5
ViT-e [23]I-4B128 × 22486.5--43.6-
ViT-22B [23]I-4B128 × 22488.0--44.9-
CoCa-g [102]I-3B16 × 57688.088.581.147.4-
InternVideo2s2-1BIV-25.5M16 × 22487.988.079.546.367.3
InternVideo2s2-6BIV-400M16 × 22488.889.181.047.867.7
", + "image_path": "69f826745fff0c5f46fa3b2a86d1d363cb9cf989f403f1b606057a336e52960a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 435, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 482, + 590 + ], + "type": "text", + "content": "End-to-end Finetuning. Tab. 2 shows InternVideo2-6B obtains new state-of-the-art (SOTA) results on Kinetics " + }, + { + "bbox": [ + 132, + 435, + 482, + 590 + ], + "type": "inline_equation", + "content": "(92.1\\% / 91.9\\% / 85.9\\%)" + }, + { + "bbox": [ + 132, + 435, + 482, + 590 + ], + "type": "text", + "content": " on K400/600/700, respectively), SthSthv2, MiT, ANet, and HACS with only 16 frames, while the previous SOTAs require larger resolution (224 vs. 576) or model ensemble. As for MiT in Table 2, InternVideo2-6B exceeds the previous SOTA, CoCa-g, by a significant margin of " + }, + { + "bbox": [ + 132, + 435, + 482, + 590 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 132, + 435, + 482, + 590 + ], + "type": "inline_equation", + "content": "(51.2\\%" + }, + { + "bbox": [ + 132, + 435, + 482, + 590 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 132, + 435, + 482, + 590 + ], + "type": "inline_equation", + "content": "49.0\\%)" + }, + { + "bbox": [ + 132, + 435, + 482, + 590 + ], + "type": "text", + "content": ". Regarding the temporal-related actions in Table 2, our InternVideo2-6B also surpasses MVD [92] on SSv2 " + }, + { + "bbox": [ + 132, + 435, + 482, + 590 + ], + "type": "inline_equation", + "content": "(77.5\\%" + }, + { + "bbox": [ + 132, + 435, + 482, + 590 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 132, + 435, + 482, + 590 + ], + "type": "inline_equation", + "content": "77.3\\%)" + }, + { + "bbox": [ + 132, + 435, + 482, + 590 + ], + "type": "text", + "content": ". Moreover, our InternVideo2-6B showcases top performance on untrimmed video analysis, as indicated in Table 2, with " + }, + { + "bbox": [ + 132, + 435, + 482, + 590 + ], + "type": "inline_equation", + "content": "95.9\\%" + }, + { + "bbox": [ + 132, + 435, + 482, + 590 + ], + "type": "text", + "content": " on ActivityNet and " + }, + { + "bbox": [ + 132, + 435, + 482, + 590 + ], + "type": "inline_equation", + "content": "97.0\\%" + }, + { + "bbox": [ + 132, + 435, + 482, + 590 + ], + "type": "text", + "content": " on HACS. These results affirm our model's superior capability for robustly identifying complex actions across varied scenes. Note \"I\" and \"V\" denotes images and videos, respectively. \"IV-3B\" means the total number of the used images and videos is 3B, while \"I-3B\" means using 3B images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 481, + 665 + ], + "type": "text", + "content": "Attentive Probing. As in Tab. 3, InternVideo2-6B not only outperforms ViT-22B [23] and CoCa-g [102] in scene-focused datasets but also surpasses or matches the performance of the latest video foundation model [4,109], on datasets emphasizing temporal dynamics (SthSthV2). This underscores our model's exceptional ability to understand and interpret both spatial and temporal information effectively." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "InternVideo2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 146, + 444, + 229 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 4: Linear probing action recognition results (top-1 accuracy) on Kinetics-400, Something-Something V2, UCF-101 and HMDB-51." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 146, + 444, + 229 + ], + "lines": [ + { + "bbox": [ + 167, + 146, + 444, + 229 + ], + "spans": [ + { + "bbox": [ + 167, + 146, + 444, + 229 + ], + "type": "table", + "html": "
MethodSettingK400SSV2UCF-101HMDB-51
VideoMAEv2-H [90]12 × 22425.8-56.434.1
TVTSv2-H [105]12 × 22473.1-91.865.7
OpenCLIP-G [20]8 × 22478.335.890.7-
DINOv2-g [69]8 × 22478.438.391.2-
InternVideo2s1-1B16 × 22481.646.396.071.6
InternVideo2s1-6B16 × 22482.047.896.371.8
InternVideo2s2-6B16 × 22484.256.797.380.7
", + "image_path": "933163192bf7208953121149e82c7cdd2def738fedebbbacc60388a61cc6b80c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 173, + 262, + 435, + 335 + ], + "blocks": [ + { + "bbox": [ + 132, + 240, + 479, + 251 + ], + "lines": [ + { + "bbox": [ + 132, + 240, + 479, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 479, + 251 + ], + "type": "text", + "content": "Table 5: Zero-shot action recognition on UCF, HMDB, MiTv1, SSv2-MC, and Charades." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 173, + 262, + 435, + 335 + ], + "lines": [ + { + "bbox": [ + 173, + 262, + 435, + 335 + ], + "spans": [ + { + "bbox": [ + 173, + 262, + 435, + 335 + ], + "type": "table", + "html": "
Method#FTraining DataUCFHMDBMiTSSv2-MCCharades
CLIP [74]12I-400M68.943.2-29.6-
TVTSv2 [105]12V-8.5M78.052.1-48.4-
VideoCoCa-g [98]16V-145M----25.8
VideoPrism-g [109]16V-619M----32.4
InternVideo2clip-1B8IV-25.5M88.853.931.661.532.9
InternVideo2clip-6B8IV-400M89.556.732.963.534.6
", + "image_path": "7865090ea8a5eace3424ab253185d7e55fba08d7604ac173f1c4b6166e9e9e33.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 205, + 368, + 405, + 483 + ], + "blocks": [ + { + "bbox": [ + 187, + 347, + 424, + 358 + ], + "lines": [ + { + "bbox": [ + 187, + 347, + 424, + 358 + ], + "spans": [ + { + "bbox": [ + 187, + 347, + 424, + 358 + ], + "type": "text", + "content": "Table 6: Zero-shot action recognition results on Kinetics." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 205, + 368, + 405, + 483 + ], + "lines": [ + { + "bbox": [ + 205, + 368, + 405, + 483 + ], + "spans": [ + { + "bbox": [ + 205, + 368, + 405, + 483 + ], + "type": "table", + "html": "
Method#FK400K600K700
top-1AVGtop-1AVGtop-1AVG
CLIP [74]858.470.155.167.246.158.4
EVA-CLIP-L [81]1-65.0-64.9-59.1
EVA-CLIP-E [81]1-69.8-69.3-63.4
ViCLIP-L [93]864.875.762.273.554.366.4
VideoCoCa-g [98]1672.081.3-
InternVL-6B [17]869.179.468.978.860.671.5
EVA-CLIP-18B [82]1679.4-79.4-72.2
VideoPrism-g [109]1676.485.4-
InternVideo2clip-1B873.182.472.881.864.975.2
InternVideo2clip-6B872.782.271.781.264.275.2
", + "image_path": "578eceb125de174e688ffda972410f77e3214d1d9b1c262bac3f2f0f3133acc7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 506, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 578 + ], + "type": "text", + "content": "Linear Probing. In Tab. 4, InternVideo2-1B significantly outperforms the previous SOTA, DINOv2-g [69], by notable margins: " + }, + { + "bbox": [ + 130, + 506, + 482, + 578 + ], + "type": "inline_equation", + "content": "+3.2\\%" + }, + { + "bbox": [ + 130, + 506, + 482, + 578 + ], + "type": "text", + "content": " on K400, " + }, + { + "bbox": [ + 130, + 506, + 482, + 578 + ], + "type": "inline_equation", + "content": "+8.0\\%" + }, + { + "bbox": [ + 130, + 506, + 482, + 578 + ], + "type": "text", + "content": " on SthSthV2, and " + }, + { + "bbox": [ + 130, + 506, + 482, + 578 + ], + "type": "inline_equation", + "content": "+4.8\\%" + }, + { + "bbox": [ + 130, + 506, + 482, + 578 + ], + "type": "text", + "content": " on UCF-101. As we scale the model, an upward trend in results is observed, underscoring the benefits of model enhancement. Notably, the integration of multimodal pretraining (stage 2) yields further rise in results. We suppose stage 2 enhances feature discrimination." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "Zero-shot. Table 5 and 6 show InternVideo2 gets " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "72.7\\%" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": " / " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "71.7\\%" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": " / " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "64.2\\%" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": " on K400/600/700, respectively, outperforming others but VideoPrism on K400 " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "(76.4\\%)" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": ". On UCF [79], HMDB [43], MiT [65], SSv2-MC, and Charades, InternVideo2 gives an cutting edges over others. The clear gap between VideoPrism and InternVideo2 on K400 may signify the importance of pretraining corpus in VideoPrism (311M videos with text and 36.1M of them are manually labeled) for" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 189, + 157, + 425, + 237 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 7: Finetuned temporal action localization results on THUMOS14 [39], ActivityNet [42], HACS Segment [108] and FineAction [57]. We report average mAP. \"Flow\" denotes the ensembling I3D flow feature. * denotes the result is achieved with Flow." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 189, + 157, + 425, + 237 + ], + "lines": [ + { + "bbox": [ + 189, + 157, + 425, + 237 + ], + "spans": [ + { + "bbox": [ + 189, + 157, + 425, + 237 + ], + "type": "table", + "html": "
BackboneTHUMOS14HACSActivityNetFineAction
I3D [10] + Flow66.8-35.6-
R(2+1)D [88]55.6-36.6-
InternVideo71.6*41.339.017.6
VideoMAEv2-g [90]69.5--18.2
InternVideo2s1-1B69.842.440.427.2
InternVideo2s1-6B72.043.341.227.7
", + "image_path": "abbdbb43710db59cfca86f5da030c82cd6d5de05329820a1240a33ccecddaa39.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 202, + 274, + 408, + 327 + ], + "blocks": [ + { + "bbox": [ + 135, + 246, + 476, + 258 + ], + "lines": [ + { + "bbox": [ + 135, + 246, + 476, + 258 + ], + "spans": [ + { + "bbox": [ + 135, + 246, + 476, + 258 + ], + "type": "text", + "content": "Table 8: Video instance segmentation performance (mAP) on YouTube-VIS19 [99]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 202, + 274, + 408, + 327 + ], + "lines": [ + { + "bbox": [ + 202, + 274, + 408, + 327 + ], + "spans": [ + { + "bbox": [ + 202, + 274, + 408, + 327 + ], + "type": "table", + "html": "
MethodBackbone#ParamsYouTubeVIS19
Mask2FormerSwin-L (image) [58]219M60.3
Mask2FormerInternViT (image)6B63.4
Mask2FormerInternVideo2s16B64.2
", + "image_path": "31010162db6b7737e17fbe83aabc599eeb4c96fc353f8aa81a33373f1c3a2081.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 346, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 407 + ], + "type": "text", + "content": "K400 in zero-shot. Note that on the datasets of Kinetics, UCF101 and HMDB51, which have a distribution closer to the pre-training dataset used in stage1, the performance of Intervvideo2-6B is slightly inferior to that of Intervvideo2-1B. We suppose this is caused by Intervvideo2-6B uses a more abundant pretraining dataset in stage2, leading to the forgetting of pretraining data in stage1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 422, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 578 + ], + "type": "text", + "content": "Temporal Action Localization. We evaluate models on four temporal action localization (TAL) datasets: THUMOS14 [39], ActivityNet [42], HACS Segment [108] and FineAction [57] in a feature-based manner with finetuning. We employ output of the 7-th layer from InternVideo2 for inputs as the corresponding features without fusing anything else. ActionFormer [107] is used as the detection head. We report mean Average Precision(mAP) under multiple tIoU as in [12,54,100,107]. In Table 7, InternVideo2-6B gets the highest mAP among all comparisons in all datasets, while InternVideo2-1B nearly surpass other methods except in THUMOS14. We find InternVideo2-6B almost consistently improves mAP with a notable margin from InternVideo2-1B except in FineAction. We suppose scaling model capacity without data refinement cannot nontrivially improve fine-grained discrimination abilities of models. Scaling detailed annotations in training may address this issue." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Video Instance Segmentation. We evaluate on the Video Instance Segmentation (VIS) dataset Youtube-VIS 2019 [99]. Built upon Mask2Former [18], we employ the video encoder of InternVideo2 as backbone with ViT-adapter [16] for features. We also try InternViT [17] for comparisons. In Table 8, InternVideo2 gets the highest mAP among all. This validates its effectiveness in relatively fine-grained spatiotemporal perception." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "InternVideo2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 153, + 146, + 458, + 249 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 9: Results of zero-shot video retrieval in both text-to-video (T2V) and video-to-text (V2T) on MSR-VTT, LSMDC, DiDeMo, MSVD, ANet, and VATEX." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 146, + 458, + 249 + ], + "lines": [ + { + "bbox": [ + 153, + 146, + 458, + 249 + ], + "spans": [ + { + "bbox": [ + 153, + 146, + 458, + 249 + ], + "type": "table", + "html": "
MethodMSR-VTTLSMDCDiDeMoMSVDANetVATEX
T2VV2TT2VV2TT2VV2TT2VV2TT2VV2TT2VV2T
CLIP [74]30.424.213.911.912.718.740.557.29.113.2--
CLIP4Clip [59]32.0-15.1---38.5-----
ViCLIP [93]42.441.320.116.918.427.949.175.115.124.0--
InternVideo-L [94]40.739.617.613.231.533.543.467.630.731.449.569.5
UMT-L [51]40.737.124.921.948.649.949.074.541.939.4--
VideoCoCa-g [98]34.464.7------34.533.053.273.6
VideoPrism-g [109]39.771.0------52.750.362.577.1
InternVideo2s2-1B51.950.932.027.357.054.358.183.360.454.870.485.4
InternVideo2s2-6B55.953.733.830.157.957.159.383.163.256.571.585.3
", + "image_path": "ad714725eade35976b9de49d57b75bdb2f84e48b84b3e5e9f6a5a06eb3401b04.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 153, + 291, + 460, + 363 + ], + "blocks": [ + { + "bbox": [ + 132, + 258, + 482, + 281 + ], + "lines": [ + { + "bbox": [ + 132, + 258, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 482, + 281 + ], + "type": "text", + "content": "Table 10: Results of finetuning video retrieval in both text-to-video (T2V) and video-to-text (V2T) on MSR-VTT, LSMDC, DiDeMo, MSVD, ANet, and VATEX." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 291, + 460, + 363 + ], + "lines": [ + { + "bbox": [ + 153, + 291, + 460, + 363 + ], + "spans": [ + { + "bbox": [ + 153, + 291, + 460, + 363 + ], + "type": "table", + "html": "
MethodMSR-VTTLSMDCDiDeMoMSVDANetVATEX
T2VV2TT2VV2TT2VV2TT2VV2TT2VV2TT2VV2T
CLIP [74]38.238.722.522.632.233.9--26.126.9--
CLIP4Clip [59]45.645.924.323.843.043.645.248.440.341.6--
ViCLIP [93]52.551.833.032.549.450.2--49.848.1--
UMT-L [51]58.858.643.041.470.465.758.282.466.864.472.086.0
InternVideo2s2-6B62.860.246.446.774.271.961.485.274.169.775.589.3
", + "image_path": "66b611211e7b4e683818d834b302778ac42cce2c90214a650e42c74233d57182.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 384, + 307, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 307, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 307, + 396 + ], + "type": "text", + "content": "5.2 Video-Audio-Language Tasks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 403, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 452 + ], + "type": "text", + "content": "We evaluate InternVideo2 on video retrieval, captioning, and multi-choice question-answering (QA). The former two tasks are conducted by matching video representation and the candidate text ones using the text encoder in stage 2. The latter is tested by the VideoLLM learned in stage 3. We also test audio tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 468, + 482, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 516 + ], + "type": "text", + "content": "Video Retrieval. We evaluate the video retrieval on six popular benchmarks [2, 11,38,77,97], as shown in Tab. 9 and 10. In eval, eight frames from the input videos are uniformly sampled. We report " + }, + { + "bbox": [ + 130, + 468, + 482, + 516 + ], + "type": "inline_equation", + "content": "R@1" + }, + { + "bbox": [ + 130, + 468, + 482, + 516 + ], + "type": "text", + "content": " scores for both text-to-video (t2v) and video-to-text (v2t) tasks in Tab. 9 and 10. R@5 and R@10 are given in Supp." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 517, + 482, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 576 + ], + "type": "text", + "content": "Table 9 and 10 demonstrate that InternVideo2 outperforms other state-of-the-arts with a notable margin in both t2v and v2t of all used datasets no matter in zero-shot or finetuned settings, except for the v2t of MSR-VTT, where VideoPrism gives the best result. This shows the video-language semantic alignment of transferrity of InternVideo2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "Video Temporal Grounding. We evaluate InternVideo2 on two video temporal grounding (VTG) datasets: Qvhighlight [44], and Charade-STA [32]. The eval setting and used features are the same as in TAL. We use CG-DETR [66] as the grounding head. We report R1@0.3, R1@0.5, R1@0.7, and mAP for moment retrieval as in [44, 53, 66]. Highlight Detection is evaluated in terms of \"Very Good\" mAP and HiT@1. In Table 11, InternVideo2-1B and InternVideo2-6B bring" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 135, + 310, + 182 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "type": "text", + "content": "Table 11: Finetuned temporal grounding on QVHighlight [44] and Charade-STA [32]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 135, + 310, + 182 + ], + "lines": [ + { + "bbox": [ + 134, + 135, + 310, + 182 + ], + "spans": [ + { + "bbox": [ + 134, + 135, + 310, + 182 + ], + "type": "table", + "html": "
FeatureR1@0.5R1@0.7mAPmAPHiT@1
CLIP [74]64.9748.6542.9639.8364.19
CLIP+SlowFast [29]65.4348.3842.8640.3366.21
InternVideo2s2-1B70.0054.4547.0242.3669.74
InternVideo2s2-6B71.4256.4549.2442.9072.00
", + "image_path": "0b296c671514b6aad45ae2861063012a8c520d802e5499dd445f13f9d6560297.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 313, + 135, + 472, + 182 + ], + "blocks": [ + { + "bbox": [ + 195, + 184, + 247, + 194 + ], + "lines": [ + { + "bbox": [ + 195, + 184, + 247, + 194 + ], + "spans": [ + { + "bbox": [ + 195, + 184, + 247, + 194 + ], + "type": "text", + "content": "(a) QVHighlight" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 135, + 472, + 182 + ], + "lines": [ + { + "bbox": [ + 313, + 135, + 472, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 135, + 472, + 182 + ], + "type": "table", + "html": "
FeatureR1@0.3R1@0.5R1@0.7mIoU
CLIP [74]65.6252.7730.1645.85
CLIP+SlowFast [29]70.4358.4436.3450.13
InternVideo2s2-1B78.4168.3645.0357.12
InternVideo2s2-6B79.7070.0348.9558.79
", + "image_path": "eb477f6d3c56a7f96fd364cfb2aaca038116080b695c66975e0199a08ae6dcb9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 153, + 246, + 460, + 309 + ], + "blocks": [ + { + "bbox": [ + 364, + 184, + 421, + 191 + ], + "lines": [ + { + "bbox": [ + 364, + 184, + 421, + 191 + ], + "spans": [ + { + "bbox": [ + 364, + 184, + 421, + 191 + ], + "type": "text", + "content": "(b) Charade-STA" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 130, + 213, + 480, + 236 + ], + "lines": [ + { + "bbox": [ + 130, + 213, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 480, + 236 + ], + "type": "text", + "content": "Table 12: Audio retrieval results on AudioCaps [41], Clothov1, and Clothov2 [27]. We report text-to-audio R@1 accuracy in zero-shot and finetuning settings." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 246, + 460, + 309 + ], + "lines": [ + { + "bbox": [ + 153, + 246, + 460, + 309 + ], + "spans": [ + { + "bbox": [ + 153, + 246, + 460, + 309 + ], + "type": "table", + "html": "
MethodZero-shotFinetuning
AudioCapsClothoV1ClothoV2AudioCapsClothoV1ClothoV2
VIP-ANT [110]27.7---
VAST [15]---52.025.126.9
LanguageBind [112]-12.112.1---
InternVideo2s2-6B37.117.417.455.225.327.2
", + "image_path": "363f69c0fcbd50140a67b14b7472d1fce4b3a7368464bd8fd17f1b39cd74cfea.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 154, + 352, + 284, + 399 + ], + "blocks": [ + { + "bbox": [ + 130, + 319, + 481, + 343 + ], + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 343 + ], + "type": "text", + "content": "Table 13: Results of AudioQA on ClothoAQA [55] and Audio-MusicAVQA (AMAVQA) [5], and audio classification on the ESC-50 [71]. Both in the finetuning setting." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 154, + 352, + 284, + 399 + ], + "lines": [ + { + "bbox": [ + 154, + 352, + 284, + 399 + ], + "spans": [ + { + "bbox": [ + 154, + 352, + 284, + 399 + ], + "type": "table", + "html": "
BackboneClothQAAMAVQA
AquaNet [55]14.7865.59
MWAFM [45]22.2467.54
InternVideo2s230.1480.51
", + "image_path": "433c74bab5f66fbb9511acc44f311a1c2c335cc185298d9a2b0b18cc665914db.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 348, + 352, + 443, + 399 + ], + "blocks": [ + { + "bbox": [ + 166, + 399, + 272, + 409 + ], + "lines": [ + { + "bbox": [ + 166, + 399, + 272, + 409 + ], + "spans": [ + { + "bbox": [ + 166, + 399, + 272, + 409 + ], + "type": "text", + "content": "(a) ClothoAQA and AMAVQA" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 348, + 352, + 443, + 399 + ], + "lines": [ + { + "bbox": [ + 348, + 352, + 443, + 399 + ], + "spans": [ + { + "bbox": [ + 348, + 352, + 443, + 399 + ], + "type": "table", + "html": "
MethodTop-1 Acc
AST [34]95.60
BEATS [13]98.10
InternVideo2s298.60
", + "image_path": "9f703ac0410f15919ef3e645344499a233514c3ec7410f241e1c037cb1e6c4e3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 374, + 399, + 415, + 408 + ], + "lines": [ + { + "bbox": [ + 374, + 399, + 415, + 408 + ], + "spans": [ + { + "bbox": [ + 374, + 399, + 415, + 408 + ], + "type": "text", + "content": "(b) ESC-50" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 441, + 481, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 481, + 478 + ], + "type": "text", + "content": "gradual performance improvements compared to CLIP [73] and CLIP [73] + Slow-fast [29]. This suggests that a larger spatiotemporal model is more beneficial to short-term video semantic alignment capabilities." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "text", + "content": "Audio-related Tasks. We evaluate InternVideo2's audio and text encoders on audio tasks, including audio-text retrieval on AudioCaps [41], Clothov1, and Clothov2 [27]; audioQA on ClothoAQA [55] and Audio-MusicAVQA [5]; and audio classification on the ESC-50 [71]. As shown in Tab. 12, 13a, and 13b, our model achieves state-of-the-art performance on all downstream tasks. Considering the limited size of the used audio and text encoders, these audio-related results show crossmodal contrastive learning's benefits are mutual to the used modalities. Audio and the corresponding text models also gain from this learning." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 609, + 380, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 380, + 623 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 380, + 623 + ], + "type": "text", + "content": "5.3 Video-centric Dialogue and its Applications" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Tab. 14 shows results on MVBench [49], Egoschema [62], and Perception Test [70], by equipping VideoChat2 [49] with our InternVideo2 encoder. VideoChat2-InternVideo2 demonstrates a significant performance advantage over other systems" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "InternVideo2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 146, + 477, + 242 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 14: Results of Chat-centric Evaluation on Multiple Choice Video-QA on MVBench [49], Egoschema and Perception Test." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 146, + 477, + 242 + ], + "lines": [ + { + "bbox": [ + 136, + 146, + 477, + 242 + ], + "spans": [ + { + "bbox": [ + 136, + 146, + 477, + 242 + ], + "type": "table", + "html": "
ModelViEncoderLLMMVBenchEgoschemaPerception Test
GPT-4V [68]-GPT-443.5--
Gemini 1.0 Pro [84]--37.755.751.1
Gemini 1.0 Ultra [84]---61.554.7
Gemini 1.5 Pro [84]---72.2-
LLaVA-Next-VideoCLIP-LVicuna-7B46.543.948.8
VideoLLaMA2CLIP-L-336Mistral-7B54.651.751.4
VideoLLaMA2CLIP-L-336Mistral-8*7B53.953.352.2
VideoChat2UMT-LVicuna-7B51.1--
VideoChat2InternVideo2s3-1BMistral-7B60.355.853.0
", + "image_path": "162bba4ff3ed9b2c446b84af73b40cc12924fb6be35b3237a02d2bf654d0ed70.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 264, + 483, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 264, + 483, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 483, + 397 + ], + "type": "text", + "content": "in MVBench. In Egoschema and Perception Test evaluations, it stands out as the top performer among open-source models. When contrasted with proprietary commercial models like Gemini, VideoChat2-InternVideo2 exhibits results that are slightly lower but still comparable in the Perception Test. Egoschema requires longer context harnessing, and we plan to further explore leveraging LLM for long multimodal sequence modeling. Considering these benchmarks not only involve perception but also reasoning, these results suggest InternVideo2 does embed knowledge to partially model the world. It also validates the importance of learning a transferrable video representation for current video-related MLLM. We also give several qualitative evaluations with popular GPT-4V and GeminiPro (in supp)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 414, + 312, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 312, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 312, + 427 + ], + "type": "text", + "content": "6 Conclusion and Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 438, + 483, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 483, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 483, + 559 + ], + "type": "text", + "content": "We have introduced a new family of video foundation models called InternVideo2, which achieves the state-of-the-art performance across various video and audio tasks. In InternVideo2, we combine masked video modeling, video-audio-text contrastive learning, and next token prediction into a unified framework. Additionally, we create a new video-text dataset that incorporates video-audio-speech fused captions as descriptions. The dataset contains temporally segmented clips with high semantic coherence. These designs in InternVideo2 contribute to enhancing video understanding in both perception and reasoning tasks. Notably, InternVideo2 excels in video-related dialogue and long video understanding, demonstrating its effectiveness in capturing high-level semantics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 483, + 666 + ], + "type": "text", + "content": "Limitations and Discussions. Despite its achievements, InternVideo2 does not introduce specific novel architectural design. Instead, it leverages the existing learning techniques for scaling video foundation models while focusing on improving data processing to enhance its spatiotemporal perception, semantic alignment, and basic knowledge embedding. Similarly to previous studies [51, 101], InternVideo2 still grapples with limitations stemming from fixed input resolutions, sampling rates, and highly compressed tokens, which restrict its ability to express rich video information and capture fine-grained details." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 187 + ], + "type": "text", + "content": "This work is partially supported by the National Key R&D Program of China (No. 2022ZD0160102), National Natural Science Foundation of China (No. 62076119, No. 61921006)), and the Science and Technology Commission of Shanghai Municipality under Grant No. 23QD1400800 and No. 23YF1461900." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 228, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 142, + 228, + 481, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 228, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 142, + 228, + 481, + 295 + ], + "type": "text", + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., Ring, R., Rutherford, E., Cabi, S., Han, T., Gong, Z., Samangooei, S., Monteiro, M., Menick, J., Borgeaud, S., Brock, A., Nematzadeh, A., Sharifzadeh, S., Binkowski, M., Barreira, R., Vinyals, O., Zisserman, A., Simonyan, K.: Flamingo: a visual language model for few-shot learning. ArXiv abs/2204.14198 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 142, + 295, + 481, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 295, + 481, + 327 + ], + "spans": [ + { + "bbox": [ + 142, + 295, + 481, + 327 + ], + "type": "text", + "content": "2. Anne Hendricks, L., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.: Localizing moments in video with natural language. In: ICCV. pp. 5803-5812 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 327, + 481, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 327, + 481, + 349 + ], + "spans": [ + { + "bbox": [ + 142, + 327, + 481, + 349 + ], + "type": "text", + "content": "3. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: ICCV (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 349, + 481, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 349, + 481, + 381 + ], + "spans": [ + { + "bbox": [ + 142, + 349, + 481, + 381 + ], + "type": "text", + "content": "4. Bardes, A., Garrido, Q., Ponce, J., Chen, X., Rabbat, M., LeCun, Y., Assran, M., Ballas, N.: V-JEPA: Latent video prediction for visual representation learning (2024), https://openreview.net/forum?id=WFYbBOEOtv" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 381, + 481, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 381, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 142, + 381, + 481, + 415 + ], + "type": "text", + "content": "5. Behera, S.R., Injeti, K.M., Patibandla, J.S.K., Pokala, P.K., Pailla, B.R.: Aquallm: Audio question answering data generation using large language models. arXiv preprint arXiv:2312.17343 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 415, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 415, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 142, + 415, + 481, + 446 + ], + "type": "text", + "content": "6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 447, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 447, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 142, + 447, + 481, + 479 + ], + "type": "text", + "content": "7. Bruce, J., Dennis, M., Edwards, A., Parker-Holder, J., Shi, Y., Hughes, E., Lai, M., Mavalankar, A., Steigerwald, R., Apps, C., et al.: Genie: Generative interactive environments. arXiv preprint arXiv:2402.15391 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 479, + 481, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 479, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 142, + 479, + 481, + 502 + ], + "type": "text", + "content": "8. Carreira, J., Noland, E., Banki-Horvath, A., Hillier, C., Zisserman, A.: A short note about kinetics-600. ArXiv abs/1808.01340 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 502, + 481, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 502, + 481, + 524 + ], + "spans": [ + { + "bbox": [ + 142, + 502, + 481, + 524 + ], + "type": "text", + "content": "9. Carreira, J., Noland, E., Hillier, C., Zisserman, A.: A short note on the kinetics-700 human action dataset. ArXiv abs/1907.06987 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 524, + 481, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 524, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 524, + 481, + 545 + ], + "type": "text", + "content": "10. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: CVPR (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 545, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 481, + 589 + ], + "type": "text", + "content": "11. Chen, D.L., Dolan, W.B.: Collecting highly parallel data for paraphrase evaluation. In: Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies-Volume 1. pp. 190-200. Association for Computational Linguistics (2011)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 481, + 621 + ], + "type": "text", + "content": "12. Chen, G., Zheng, Y.D., Wang, L., Lu, T.: Dscan: improving temporal action detection via dual context aggregation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 248-257 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "text", + "content": "13. Chen, S., Wu, Y., Wang, C., Liu, S., Tompkins, D., Chen, Z., Che, W., Yu, X., Wei, F.: BEATs: Audio pre-training with acoustic tokenizers. In: Proceedings of the 40th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 202, pp. 5178-5193. PMLR (23-29 Jul 2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "InternVideo2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "text", + "content": "14. Chen, S., He, X., Guo, L., Zhu, X., Wang, W., Tang, J., Liu, J.: Valor: Vision-audio-language omni-perception pretraining model and dataset. arXiv preprint arXiv:2304.08345 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 138, + 150, + 481, + 182 + ], + "type": "text", + "content": "15. Chen, S., Li, H., Wang, Q., Zhao, Z., Sun, M., Zhu, X., Liu, J.: Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 182, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 182, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 182, + 481, + 205 + ], + "type": "text", + "content": "16. Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 205, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 481, + 248 + ], + "type": "text", + "content": "17. Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Zhong, M., Zhang, Q., Zhu, X., Lu, L., Li, B., Luo, P., Lu, T., Qiao, Y., Dai, J.: Intervl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 248, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 248, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 248, + 481, + 270 + ], + "type": "text", + "content": "18. Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: NeurIPS (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 270, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 270, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 138, + 270, + 481, + 303 + ], + "type": "text", + "content": "19. Cheng, F., Wang, X., Lei, J., Crandall, D.J., Bansal, M., Bertasius, G.: Vindlu: A recipe for effective video-and-language pretraining. ArXiv abs/2212.05051 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 303, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 303, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 303, + 481, + 336 + ], + "type": "text", + "content": "20. Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: CVPR. pp. 2818-2829 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 336, + 481, + 369 + ], + "type": "text", + "content": "21. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B.A., Fung, P., Hoi, S.C.H.: Instructblip: Towards general-purpose vision-language models with instruction tuning. In: NeurIPS (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 369, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 369, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 138, + 369, + 481, + 391 + ], + "type": "text", + "content": "22. Dao, T., Fu, D., Ermon, S., Rudra, A., Ré, C.: Flashattention: Fast and memory-efficient exact attention with io-awareness. NeurIPS 35, 16344-16359 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 391, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 391, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 391, + 481, + 479 + ], + "type": "text", + "content": "23. Dehghani, M., Djolonga, J., Mustafa, B., Padlewski, P., Heek, J., Gilmer, J., Steiner, A., Caron, M., Geirhos, R., Alabdulmohsin, I.M., Jenatton, R., Beyer, L., Tschannen, M., Arnab, A., Wang, X., Riquelme, C., Minderer, M., Puigcerver, J., Evci, U., Kumar, M., van Steenkiste, S., Elsayed, G.F., Mahendran, A., Yu, F., Oliver, A., Huot, F., Bastings, J., Collier, M., Gritsenko, A.A., Birodkar, V., Vasconcelos, C.N., Tay, Y., Mensink, T., Kolesnikov, A., Paveti'c, F., Tran, D., Kipf, T., Luvci'c, M., Zhai, X., Keysers, D., Harmsen, J., Houlsby, N.: Scaling vision transformers to 22 billion parameters. In: ICML (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 481, + 511 + ], + "type": "text", + "content": "24. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. ArXiv abs/1810.04805 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 511, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 511, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 511, + 481, + 555 + ], + "type": "text", + "content": "25. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv abs/2010.11929 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 555, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 609 + ], + "type": "text", + "content": "26. Driess, D., Xia, F., Sajjadi, M.S.M., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q.H., Yu, T., Huang, W., Chebotar, Y., Sermanet, P., Duckworth, D., Levine, S., Vanhoucke, V., Hausman, K., Toussaint, M., Greff, K., Zeng, A., Mordatch, I., Florence, P.R.: Palm-e: An embodied multimodal language model. In: ICML (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 481, + 643 + ], + "type": "text", + "content": "27. Drossos, K., Lipping, S., Virtanen, T.: Cloth: An audio captioning dataset. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 736-740. IEEE (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "28. Feichtenhofer, C., Fan, H., Li, Y., He, K.: Masked autoencoders as spatiotemporal learners. NeurIPS (2022)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 481, + 138 + ], + "type": "text", + "content": "29. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: ICCV (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 138, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 138, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 138, + 138, + 481, + 171 + ], + "type": "text", + "content": "30. Fu, T.J., Li, L., Gan, Z., Lin, K., Wang, W.Y., Wang, L., Liu, Z.: Violet: End-to-end video-language transformers with masked visual-token modeling. arXiv preprint arXiv:2111.12681 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 171, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 171, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 138, + 171, + 481, + 203 + ], + "type": "text", + "content": "31. Gabeur, V., Sun, C., Alahari, K., Schmid, C.: Multi-modal transformer for video retrieval. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 214-229. Springer (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 203, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 203, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 203, + 481, + 224 + ], + "type": "text", + "content": "32. Gao, J., Sun, C., Yang, Z., Nevatia, R.: Tall: Temporal activity localization via language query. In: ICCV (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 224, + 481, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 481, + 256 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 481, + 256 + ], + "type": "text", + "content": "33. Gong, T., Lyu, C., Zhang, S., Wang, Y., Zheng, M., Zhao, Q., Liu, K., Zhang, W., Luo, P., Chen, K.: Multimodal-gpt: A vision and language model for dialogue with humans. ArXiv abs/2305.04790 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 257, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 257, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 257, + 481, + 277 + ], + "type": "text", + "content": "34. Gong, Y., Chung, Y.A., Glass, J.: Ast: Audio spectrogram transformer. arXiv preprint arXiv:2104.01778 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 278, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 278, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 481, + 320 + ], + "type": "text", + "content": "35. Goyal, R., Kahou, S.E., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fründ, I., Yianilos, P., Mueller-Freitag, M., Hoppe, F., Thurau, C., Bax, I., Memisevic, R.: The \"something something\" video database for learning and evaluating visual common sense. In: ICCV (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 321, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 481, + 353 + ], + "type": "text", + "content": "36. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: CVPR (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 354, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 481, + 374 + ], + "type": "text", + "content": "37. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 375, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 375, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 375, + 481, + 396 + ], + "type": "text", + "content": "38. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: CVPR (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 396, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 481, + 429 + ], + "type": "text", + "content": "39. Idrees, H., Zamir, A.R., Jiang, Y.G., Gorban, A., Laptev, I., Sukthankar, R., Shah, M.: The thumos challenge on action recognition for videos \"in the wild\". Computer Vision and Image Understanding (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 429, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 481, + 460 + ], + "type": "text", + "content": "40. Jiang, A.Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D.S., Casas, D.d.l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al.: Mistral 7b. arXiv preprint arXiv:2310.06825 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 461, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 461, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 138, + 461, + 481, + 504 + ], + "type": "text", + "content": "41. Kim, C.D., Kim, B., Lee, H., Kim, G.: Audiocaps: Generating captions for audios in the wild. In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 119-132 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 504, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 481, + 525 + ], + "type": "text", + "content": "42. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Carlos Niebles, J.: Dense-captioning events in videos. In: ICCV (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 481, + 557 + ], + "type": "text", + "content": "43. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: 2011 International conference on computer vision. pp. 2556-2563. IEEE (2011)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 558, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 558, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 558, + 481, + 579 + ], + "type": "text", + "content": "44. Lei, J., Berg, T.L., Bansal, M.: Qvhighlights: Detecting moments and highlights in videos via natural language queries (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 579, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 481, + 600 + ], + "type": "text", + "content": "45. Li, G., Xu, Y., Hu, D.: Multi-scale attention for audio question answering. arXiv preprint arXiv:2305.17993 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "type": "text", + "content": "46. Li, J., Li, D., Savarese, S., Hoi, S.C.H.: Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In: ICML (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "47. Li, J., Li, D., Xiong, C., Hoi, S.C.H.: Blip: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In: ICML (2022)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "InternVideo2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "text", + "content": "48. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 138, + 150, + 481, + 182 + ], + "type": "text", + "content": "49. Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Liu, Y., Wang, Z., Xu, J., Chen, G., Luo, P., et al.: Mvbench: A comprehensive multi-modal video understanding benchmark. arXiv preprint arXiv:2311.17005 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 182, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 182, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 138, + 182, + 481, + 215 + ], + "type": "text", + "content": "50. Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Wang, L., Qiao, Y.: Uniformerv2: Spatiotemporal learning by arming image vits with video uniformer. arXiv preprint arXiv:2211.09552 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 138, + 216, + 481, + 248 + ], + "type": "text", + "content": "51. Li, K., Wang, Y., Li, Y., Wang, Y., He, Y., Wang, L., Qiao, Y.: Unmasked teacher: Towards training-efficient video foundation models. arXiv preprint arXiv:2303.16058 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 481, + 270 + ], + "type": "text", + "content": "52. Li, T., Wang, L.: Learning spatiotemporal features via video and text pair discrimination. CoRR abs/2001.05691 (2020), https://arxiv.org/abs/2001.05691" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 271, + 481, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 481, + 313 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 481, + 313 + ], + "type": "text", + "content": "53. Lin, K.Q., Zhang, P., Chen, J., Pramanick, S., Gao, D., Wang, A.J., Yan, R., Shou, M.Z.: Univtg: Towards unified video-language temporal grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2794-2804 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 315, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 481, + 335 + ], + "type": "text", + "content": "54. Lin, T., Liu, X., Li, X., Ding, E., Wen, S.: Bmn: Boundary-matching network for temporal action proposal generation (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 336, + 481, + 369 + ], + "type": "text", + "content": "55. Lipping, S., Sudarsanam, P., Drossos, K., Virtanen, T.: Clotho-aqa: A crowdsourced dataset for audio question answering. In: 2022 30th European Signal Processing Conference (EUSIPCO). pp. 1140-1144. IEEE (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 369, + 425, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 369, + 425, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 369, + 425, + 380 + ], + "type": "text", + "content": "56. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 481, + 413 + ], + "type": "text", + "content": "57. Liu, Y., Wang, L., Wang, Y., Ma, X., Qiao, Y.: Fineaction: A fine-grained video dataset for temporal action localization. IEEE Transactions on Image Processing 31, 6937-6950 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 413, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 413, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 413, + 481, + 445 + ], + "type": "text", + "content": "58. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 446, + 481, + 479 + ], + "type": "text", + "content": "59. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 481, + 511 + ], + "type": "text", + "content": "60. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M.H., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability. ArXiv abs/2306.07207 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 481, + 544 + ], + "type": "text", + "content": "61. Maaz, M., Rasheed, H.A., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. ArXiv abs/2306.05424 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 481, + 578 + ], + "type": "text", + "content": "62. Mangalam, K., Akshulakov, R., Malik, J.: Egoschema: A diagnostic benchmark for very long-form video language understanding. In: Oh, A., Neumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) NeurIPS. pp. 46212-46244 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 578, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 599 + ], + "type": "text", + "content": "63. Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. In: CVPR (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "text", + "content": "64. Mei, X., Meng, C., Liu, H., Kong, Q., Ko, T., Zhao, C., Plumbley, M.D., Zou, Y., Wang, W.: Wavcaps: A chatgpt-assisted weakly-labelled audio captioning dataset for audio-language multimodal research (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "65. Monfort, M., Zhou, B., Bargal, S.A., Andonian, A., Yan, T., Ramakrishnan, K., Brown, L.M., Fan, Q., Gutfreund, D., Vondrick, C., Oliva, A.: Moments in time dataset: One million videos for event understanding. TPAMI (2020)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "text", + "content": "66. Moon, W., Hyun, S., Lee, S., Heo, J.P.: Correlation-guided query-dependency calibration in video representation learning for temporal grounding. arXiv preprint arXiv:2311.08835 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 151, + 417, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 151, + 417, + 162 + ], + "spans": [ + { + "bbox": [ + 138, + 151, + 417, + 162 + ], + "type": "text", + "content": "67. OpenAI: Gpt-4 technical report. ArXiv abs/2303.08774 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 163, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 163, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 138, + 163, + 481, + 184 + ], + "type": "text", + "content": "68. OpenAI: Gpt-4v(ison) system card. https://api_semanticscholar.org/ CorpusID:263218031 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 185, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 185, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 185, + 481, + 217 + ], + "type": "text", + "content": "69. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 219, + 481, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 219, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 138, + 219, + 481, + 284 + ], + "type": "text", + "content": "70. Patraucean, V., Smaira, L., Gupta, A., Continente, A.R., Markeeva, L., Dylan, Banarse, Malinowski, M., Yang, Y., Doersch, C., Matejovicova, T., Sulsky, Y., Antoine, Miech, Koppula, S., Fréchette, A., Klimczak, H., Koster, R., Zhang, J., Stephanie, Winkler, Aytar, Y., Osindero, S., Damen, D., Zisserman, A., Carreira, J.: Perception test: A diagnostic benchmark for multimodal models. In: NeurIPS (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 285, + 481, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 481, + 307 + ], + "type": "text", + "content": "71. Piczak, K.J.: Esc: Dataset for environmental sound classification. In: Proceedings of the 23rd ACM international conference on Multimedia. pp. 1015-1018 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 308, + 481, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 308, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 481, + 340 + ], + "type": "text", + "content": "72. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: ICCV (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 341, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 341, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 341, + 481, + 373 + ], + "type": "text", + "content": "73. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 481, + 407 + ], + "type": "text", + "content": "74. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 407, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 481, + 441 + ], + "type": "text", + "content": "75. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision. In: International Conference on Machine Learning. pp. 28492-28518. PMLR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 441, + 481, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 441, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 138, + 441, + 481, + 474 + ], + "type": "text", + "content": "76. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 475, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 475, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 138, + 475, + 481, + 497 + ], + "type": "text", + "content": "77. Rohrbach, A., Rohrbach, M., Tandon, N., Schiele, B.: A dataset for movie description. In: CVPR. pp. 3202-3212 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 498, + 481, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 481, + 530 + ], + "type": "text", + "content": "78. Ryali, C., Hu, Y.T., Bolya, D., Wei, C., Fan, H., Huang, P.Y., Aggarwal, V., Chowdhury, A., Poursaeed, O., Hoffman, J., Malik, J., Li, Y., Feichtenhofer, C.: Hiera: A hierarchical vision transformer without the bells-and-whistles (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 531, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 531, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 531, + 481, + 553 + ], + "type": "text", + "content": "79. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 554, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 481, + 586 + ], + "type": "text", + "content": "80. Sun, Q., Cui, Y., Zhang, X., Zhang, F., Yu, Q., Luo, Z., Wang, Y., Rao, Y., Liu, J., Huang, T., et al.: Generative multimodal models are in-context learners. arXiv preprint arXiv:2312.13286 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 587, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 587, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 587, + 481, + 609 + ], + "type": "text", + "content": "81. Sun, Q., Fang, Y., Wu, L., Wang, X., Cao, Y.: Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 481, + 632 + ], + "type": "text", + "content": "82. Sun, Q., Wang, J., Yu, Q., Cui, Y., Zhang, F., Zhang, X., Wang, X.: Eva-clip-18b: Scaling clip to 18 billion parameters. arXiv preprint arXiv:2402.04252 (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "83. Sun, Q., Yu, Q., Cui, Y., Zhang, F., Zhang, X., Wang, Y., Gao, H., Liu, J., Huang, T., Wang, X.: Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "InternVideo2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "text", + "content": "84. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 149, + 481, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 149, + 481, + 181 + ], + "spans": [ + { + "bbox": [ + 138, + 149, + 481, + 181 + ], + "type": "text", + "content": "85. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. NeurIPS 35, 10078-10093 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 181, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 181, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 181, + 481, + 224 + ], + "type": "text", + "content": "86. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models. ArXiv abs/2302.13971 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 224, + 481, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 481, + 355 + ], + "type": "text", + "content": "87. Touvron, H., Martin, L., Stone, K.R., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D.M., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A.S., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I.M., Korenev, A.V., Koura, P.S., Lachaux, M.A., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. ArXiv abs/2307.09288 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 355, + 481, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 355, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 355, + 481, + 376 + ], + "type": "text", + "content": "88. Tran, D., xiu Wang, H., Torresani, L., Ray, J., LeCun, Y., Paluri, M.: A closer look at spatiotemporal convolutions for action recognition. In: CVPR (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 376, + 481, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 481, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 481, + 408 + ], + "type": "text", + "content": "89. Wang, J., Ge, Y., Yan, R., Ge, Y., Lin, K.Q., Tsutsui, S., Lin, X., Cai, G., Wu, J., Shan, Y., et al.: All in one: Exploring unified video-language pre-training. In: CVPR. pp. 6598-6608 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 408, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 408, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 138, + 408, + 481, + 440 + ], + "type": "text", + "content": "90. Wang, L., Huang, B., Zhao, Z., Tong, Z., He, Y., Wang, Y., Wang, Y., Qiao, Y.: Videomae v2: Scaling video masked autoencoders with dual masking. In: CVPR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 440, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 440, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 138, + 440, + 481, + 472 + ], + "type": "text", + "content": "91. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Gool, L.V.: Temporal segment networks: Towards good practices for deep action recognition. In: ECCV (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 472, + 481, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 481, + 505 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 481, + 505 + ], + "type": "text", + "content": "92. Wang, R., Chen, D., Wu, Z., Chen, Y., Dai, X., Liu, M., Yuan, L., Jiang, Y.G.: Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In: CVPR. pp. 6312-6322 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 505, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 505, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 505, + 481, + 536 + ], + "type": "text", + "content": "93. Wang, Y., He, Y., Li, Y., Li, K., Yu, J., Ma, X., Li, X., Chen, G., Chen, X., Wang, Y., et al.: Intervid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 536, + 481, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 481, + 580 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 481, + 580 + ], + "type": "text", + "content": "94. Wang, Y., Li, K., Li, Y., He, Y., Huang, B., Zhao, Z., Zhang, H., Xu, J., Liu, Y., Wang, Z., Xing, S., Chen, G., Pan, J., Yu, J., Wang, Y., Wang, L., Qiao, Y.: Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 580, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 580, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 580, + 481, + 611 + ], + "type": "text", + "content": "95. Xu, H., Ye, Q., Yan, M., Shi, Y., Ye, J., Xu, Y., Li, C., Bi, B., Qian, Q., Wang, W., et al.: mplug-2: A modularized multi-modal foundation model across text, image and video. arXiv preprint arXiv:2302.00402 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "type": "text", + "content": "96. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettle-moyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "97. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: CVPR. pp. 5288-5296 (2016)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 654 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 136, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 481, + 149 + ], + "type": "text", + "content": "98. Yan, S., Zhu, T., Wang, Z., Cao, Y., Zhang, M., Ghosh, S., Wu, Y., Yu, J.: Video-text modeling with zero-shot transfer from contrastive captioners. ArXiv abs/2212.04979 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 150, + 452, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 150, + 452, + 160 + ], + "spans": [ + { + "bbox": [ + 137, + 150, + 452, + 160 + ], + "type": "text", + "content": "99. Yang, L., Fan, Y., Xu, N.: Video instance segmentation. In: ICCV (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 134, + 161, + 481, + 194 + ], + "type": "text", + "content": "100. Yang, M., Chen, G., Zheng, Y.D., Lu, T., Wang, L.: Basictad: an astounding rgb-only baseline for temporal action detection. Computer Vision and Image Understanding p. 103692 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 481, + 236 + ], + "type": "text", + "content": "101. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., Jiang, C., Li, C., Xu, Y., Chen, H., Tian, J., Qi, Q., Zhang, J., Huang, F.: mplug-owl: Modularization empowers large language models with multimodality (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 481, + 270 + ], + "type": "text", + "content": "102. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 134, + 271, + 481, + 303 + ], + "type": "text", + "content": "103. Zablocki, E., Ben-Younes, H., Pérez, P., Cord, M.: Explainability of deep vision-based autonomous driving systems: Review and challenges. International Journal of Computer Vision 130(10), 2425-2452 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 304, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 304, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 134, + 304, + 481, + 336 + ], + "type": "text", + "content": "104. Zellers, R., Lu, J., Lu, X., Yu, Y., Zhao, Y., Salehi, M., Kusupati, A., Hessel, J., Farhadi, A., Choi, Y.: Merlot reserve: Neural script knowledge through vision and language and sound. In: CVPR. pp. 16375-16387 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 337, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 337, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 134, + 337, + 481, + 369 + ], + "type": "text", + "content": "105. Zeng, Z., Ge, Y., Tong, Z., Liu, X., Xia, S.T., Shan, Y.: Tvtsv2: Learning out-of-the-box spatiotemporal visual representations at scale. arXiv preprint arXiv:2305.14173 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 134, + 369, + 481, + 402 + ], + "type": "text", + "content": "106. Zhang, B., Yu, J., Fifty, C., Han, W., Dai, A.M., Pang, R., Sha, F.: Cotraining transformer with videos and images improves action recognition. ArXiv abs/2112.07175 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 402, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 402, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 134, + 402, + 481, + 423 + ], + "type": "text", + "content": "107. Zhang, C.L., Wu, J.Z., Li, Y.: Actionformer: Localizing moments of actions with transformers. ArXiv abs/2202.07925 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 424, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 424, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 134, + 424, + 481, + 445 + ], + "type": "text", + "content": "108. Zhao, H., Torralba, A., Torresani, L., Yan, Z.: Hacs: Human action clips and segments dataset for recognition and temporal localization. In: ICCV (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 447, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 447, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 134, + 447, + 481, + 490 + ], + "type": "text", + "content": "109. Zhao, L., Gundavarapu, N.B., Yuan, L., Zhou, H., Yan, S., Sun, J.J., Friedman, L., Qian, R., Weyand, T., Zhao, Y., Hornung, R., Schroff, F., Yang, M.H., Ross, D.A., Wang, H., Adam, H., Sirotenko, M., Liu, T., Gong, B.: Videoprism: A foundational visual encoder for video understanding (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 134, + 491, + 481, + 522 + ], + "type": "text", + "content": "110. Zhao, Y., Hessel, J., Yu, Y., Lu, X., Zellers, R., Choi, Y.: Connecting the dots between audio and text without parallel data through visual knowledge transfer. arXiv preprint arXiv:2112.08995 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 134, + 523, + 481, + 555 + ], + "type": "text", + "content": "111. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 555, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 481, + 588 + ], + "type": "text", + "content": "112. Zhu, B., Lin, B., Ning, M., Yan, Y., Cui, J., Wang, H., Pang, Y., Jiang, W., Zhang, J., Li, Z., et al.: Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 589, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 589, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 134, + 589, + 481, + 620 + ], + "type": "text", + "content": "113. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. ArXiv abs/2304.10592 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 621, + 481, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 621, + 481, + 654 + ], + "spans": [ + { + "bbox": [ + 134, + 621, + 481, + 654 + ], + "type": "text", + "content": "114. Zhu, W., Huang, Y., Xie, X., Liu, W., Deng, J., Zhang, D., Wang, Z., Liu, J.: Autoshot: A short video dataset and state-of-the-art shot boundary detection. In: CVPR. pp. 2237-2246 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "InternVideo2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "21" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 20 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Interpretability-Guided Test-Time Adversarial Defense/d8bed3ff-2ae8-4bfb-abf2-7d81aee96721_content_list.json b/2024/Interpretability-Guided Test-Time Adversarial Defense/d8bed3ff-2ae8-4bfb-abf2-7d81aee96721_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..972354a5161c9a296c79de43dc4123ad41920f3e --- /dev/null +++ b/2024/Interpretability-Guided Test-Time Adversarial Defense/d8bed3ff-2ae8-4bfb-abf2-7d81aee96721_content_list.json @@ -0,0 +1,1673 @@ +[ + { + "type": "text", + "text": "Interpretability-Guided Test-Time Adversarial Defense", + "text_level": 1, + "bbox": [ + 236, + 140, + 767, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Akshay Kulkarni $^{1}$ and Tsui-Wei Weng $^{2}$", + "bbox": [ + 352, + 210, + 650, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 CSE, UC San Diego", + "bbox": [ + 423, + 238, + 576, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ HDSI, UC San Diego", + "bbox": [ + 423, + 253, + 578, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{a2kulkarni,lweng}@ucsd.edu", + "bbox": [ + 395, + 268, + 604, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We propose a novel and low-cost test-time adversarial defense by devising interpretability-guided neuron importance ranking methods to identify neurons important to the output classes. Our method is a training-free approach that can significantly improve the robustness-accuracy tradeoff while incurring minimal computational overhead. While being among the most efficient test-time defenses (4× faster), our method is also robust to a wide range of black-box, white-box, and adaptive attacks that break previous test-time defenses. We demonstrate the efficacy of our method for CIFAR10, CIFAR100, and ImageNet-1k on the standard RobustBenchmark (with average gains of 2.6%, 4.9%, and 2.8% respectively). We also show improvements (average 1.5%) over the state-of-the-art test-time defenses even under strong adaptive attacks.", + "bbox": [ + 259, + 316, + 743, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 530, + 375, + 546 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the popularity and success of deep neural networks (DNNs), they have been shown to be brittle against carefully designed small perturbations that are imperceptible to humans [21]. These \"adversarial\" perturbations pose significant risks against deploying DNNs for safety-critical tasks like autonomous driving. To mitigate these risks, there has been great interest in developing defenses, e.g. adversarial training algorithms [40, 42, 44, 52] to obtain a model that is more robust to adversarial samples.", + "bbox": [ + 212, + 561, + 789, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, adversarial training is much more expensive than standard training (e.g. a strong defense like TRADES [52] takes $150 - 200 \\times$ longer than standard training). This inspired training-free defenses known as test-time defenses [3, 26, 30, 39]. The underlying principles in existing test-time defenses are either input purification or model adaptation: input purification [3, 39] modifies the input with an additional perturbation to nullify any adversarial perturbation, whereas model adaptation [14, 26] updates model parameters or adds new parameters at test-time to make the overall model more robust. Despite eliminating the need for training, most existing methods are still quite expensive, $8 - 500 \\times$ more than", + "bbox": [ + 212, + 667, + 789, + 803 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^3$ Our code is available at https://github.com/Trustworthy-ML-Lab/Interpretability-Guided-Defense", + "bbox": [ + 217, + 809, + 789, + 838 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/91dac736033fb7175228124be14c872e907811122a13ec7ccae3e405b11e37cf.jpg", + "image_caption": [ + "Fig. 1: Interpretability-guided masking based on neuron importance ranking for test-time adversarial defense." + ], + "image_footnote": [], + "bbox": [ + 276, + 145, + 727, + 344 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a standard forward pass [16]. Moreover, existing test-time defenses have been broken under strong adaptive attacks [16]. Motivated by these limitations, in this paper, we intend to find an efficient and effective alternative to improve robustness at test-time, so that the defense does not break under adaptive attacks.", + "bbox": [ + 212, + 415, + 787, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To design an effective defense, we leverage recent neuron-level interpretability methods [8, 33], which allow one to interpret the functionalities of individual neurons. In our analysis, we observed that for successful adversarial examples, the majority of activations at an intermediate layer shift from important neurons for the ground-truth class to the important neurons for other classes. Hence, we propose a novel test-time Interpretability-Guided Defense (IG-Defense) enabled by \"neuron importance ranking\" where class-wise importance of each neuron is computed (Fig. 1A). To design effective neuron importance ranking, we repurpose our analysis experiments and existing neuron-level interpretability works to propose two novel neuron importance ranking methods. The key idea of our IG-Defense is to restrict the observed activation shift by masking unimportant neurons (Fig. 1B). We demonstrate the usefulness of our insights by extensively evaluating our IG-Defense on the standard RobustBench benchmark (Fig. 1C). We observe consistent gains for CIFAR10 (+2.6%), CIFAR100 (+4.9%), and ImageNet-1k (+2.8%). We also devise an ensemble of strong adaptive attacks specifically targeting the weaknesses of IG-Defense and observe consistent gains (+1.5%), unlike existing test-time defenses that break under adaptive attacks. Due to our simple yet grounded approach, IG-Defense is among the most efficient test-time defenses (4× faster than the strongest existing defense).", + "bbox": [ + 212, + 477, + 789, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our primary contributions can be summarized as:", + "bbox": [ + 215, + 766, + 576, + 780 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We are the first to propose a neuron-interpretability-guided test-time defense (IG-Defense) utilizing neuron importance ranking (CD-IR, LO-IR) to improve adversarial robustness. IG-Defense is training-free, efficient, and effective.", + "bbox": [ + 223, + 794, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "A. Kulkarni and T. W. Weng", + "bbox": [ + 271, + 114, + 470, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We uncover novel insights into improving adversarial robustness by analyzing adversarial attacks through the lens of neuron-level interpretability in Sec. 3.", + "bbox": [ + 223, + 146, + 785, + 174 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Our proposed IG-Defense consistently improves the robustness on standard CIFAR10, CIFAR100, and ImageNet-1k benchmarks. We also demonstrate improved robustness upto $3.4\\%$ , $3.8\\%$ , and $1.5\\%$ against a wide range of white-box, black-box, and adaptive attacks respectively with the lowest inference time ( $4\\times$ faster) among existing test-time defenses.", + "bbox": [ + 225, + 176, + 785, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 217, + 268, + 387, + 285 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Test-time adversarial robustness. Most of the test-time defenses can be categorized into two approaches: input purification or model adaptation. In input purification, the adversarial input is converted to a safer input by an additional perturbation. For instance, finding a perturbation that either increases top class confidence [3], minimizes contrastive loss [30], or minimizes a self-supervised objective [39]. [24] crafts the perturbation by fooling a real-adversarial discriminator [24] and [49] uses an input passed through a score-based generative model. For model adaptation [14,32,46], the model parameters are either updated at test-time or new parameters are introduced. However, existing test-time defenses are quite expensive (8-500× of single forward pass) and break under adaptive attacks. In contrast, our method is fast and effective (1.5% gain under adaptive attacks with only 2× inference time), and is the first test-time defense to leverage neuron-level interpretability information.", + "bbox": [ + 212, + 296, + 787, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neuron interpretability methods. To understand the functionalities (i.e. concepts) of individual neurons in DNNs, early works like NetDissect [8,9] compare the neuron activation patterns with patterns of predefined concept label masks. MILAN [23] produced natural language descriptions using a supervised image captioning model trained on a large-scale concept-labeled dataset. Recently, CLIP-Dissect [33] leveraged the CLIP paradigm [35] to identify concepts of neurons more efficiently than prior work and without the need of training or densely-concept-labeled datasets. There are other works to identify neuron concepts through clustering [20, 48], via language models [6], or using linear combinations [34]. In this work, we focus on repurposing CLIP-Dissect due to its efficiency and simplicity as well as NetDissect to propose neuron importance ranking methods.", + "bbox": [ + 212, + 493, + 787, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Connections between interpretability and robustness. Prior works attempt to improve robustness via input-dependent interpretability methods like saliency or class activation maps [10, 29]. Wu et al. [45] observed that adversarial attacks lead to changes in attention maps, and proposed a training-time defense with losses to enforce similar attention maps for clean and adversarial samples. Several works [7, 19, 28, 38, 47, 53] use sparsity to improve adversarial robustness. However, they require retraining after or while sparsity constraints are applied. In contrast, we indirectly leverage sparsity without requiring any training. NetDissect [8] investigated the connection between adversarial attacks and interpretability information from their method, but they did not explore how to improve robustness. Hence, we aim to study the unexplored connection of neuron-level interpretability and improving adversarial robustness.", + "bbox": [ + 212, + 659, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Interpretability-Guided Test-Time Adversarial Defense", + "bbox": [ + 366, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Analysis Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 462, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To understand the connection of interpretability of individual neurons with robustness, we aim to design an analysis experiment. One way would be to investigate the changes in neuron activations before and after an adversarial attack. In neuron-labeling works [9,31], they assign an explainable concept to each neuron (or an activation channel in case of CNNs). Specifically, NetDissect [9] qualitatively show that a targeted adversarial attack causes a drop in peak activations for neurons whose concepts are related to the original class, while causing an increase in peak activations for neurons whose concepts are related to the target class of the attack. They also show that the peak activation change is lower for neurons with unrelated concepts. While they only perform the analysis to showcase their interpretability, we are the first to analyze adversarial attacks through neuron-level interpretability with the explicit goal of obtaining insights into improving adversarial robustness.", + "bbox": [ + 212, + 184, + 787, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a more thorough analysis, we observe the average activation change (instead of peak change) and associate the neurons with the category names directly instead of concepts. First, peak activation change may be misleading since it may be caused by specific images while average change takes all images into account for a more well-rounded analysis. Second, our automated evaluation eliminates the need for manual association of concepts with categories [9], which could be difficult, expensive, and noisy, especially with a large number of related categories (e.g. ImageNet classes). We analyze both successful and unsuccessful attacks while [9] only focused on successful attacks.", + "bbox": [ + 212, + 382, + 787, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concretely, we find top- $k$ important neurons related to each task category using neuron importance ranking (discussed in Sec. 4.2). With this, in Fig. 2, we analyze the change in activations corresponding to ground-truth (GT) and non-GT categories before and after an attack. We discuss some preliminaries before delving into further details.", + "bbox": [ + 212, + 518, + 787, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Preliminaries. We denote the dataset as $\\mathcal{D} = \\{(x,y)\\}$ where $x\\in \\mathcal{X}$ is the input image, $y\\in \\mathcal{C}$ is the label, $\\mathcal{C}$ is the label set, and $|\\mathcal{C}|$ is the number of classes. Given a pretrained model $f:\\mathcal{X}\\to \\mathcal{C}$ and its activations at the layer being studied $A:\\mathcal{X}\\rightarrow \\mathbb{R}^{N\\times H\\times W}$ , we can rank the importance of a neuron (out of $N$ neurons/channels) by assessing its contribution to each class. We call this \"neuron importance ranking\" and propose two algorithms for this (Sec. 4.2). Then, we can compute the activations of the top- $k$ important neurons for any class $c_{i}$ as $A_{c_i}(x) = A(x)_{[n_{c_i}]} \\in \\mathbb{R}^{k\\times H\\times W}$ , where $n_{c_i}$ are the indices of the top- $k$ activations for class $c_{i}$ . With access to each $n_{c_i}$ , the remaining indices absent from the top- $k$ of any class are $n_u = [N] \\setminus \\{\\cup_{c_i\\in \\mathcal{C}} n_{c_i}\\}$ , where $[N] = \\{1,2,\\dots,N\\}$ and $\\backslash$ indicates setminus operator. These indices $n_u$ denote unimportant neurons, and corresponding unimportant activations can be computed as $A_u(x) = A(x)_{[n_u]} \\in \\mathbb{R}^{(N - k|\\mathcal{C}|) \\times H \\times W}$ . Next, we define activation change metrics to support our analysis.", + "bbox": [ + 214, + 595, + 787, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Metrics. For an adversarial attack with $\\ell_{\\infty}$ -norm bound of $\\epsilon$ , the perturbation for input $x$ is $\\delta^{*} = \\arg \\max_{\\| \\delta \\|_{\\infty} \\leq \\epsilon} \\ell(f(x + \\delta), y)$ where $\\ell$ is the attack objective. Let $\\hat{y}^{*} = f(x + \\delta^{*})$ be the post-attack prediction. The average activation change", + "bbox": [ + 214, + 794, + 789, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "A. Kulkarni and T. W. Weng", + "bbox": [ + 271, + 114, + 467, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/44563d80def7e9948ace4e50d33af7c4f22143b790342aa4a792eb2d9d1d0ac9.jpg", + "image_caption": [ + "Fig. 2: A. Cat-dog-bird image classifier before a PGD attack. B. After PGD attack, the prediction changes from the ground truth (GT) cat to dog since the activations of neurons important to dog class increase while those important to cat class decrease." + ], + "image_footnote": [], + "bbox": [ + 276, + 143, + 727, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "C. Empirically, successful PGD attacks show a decrease in activations of important GT class neurons while those of post-attack predicted class' important neurons increase.", + "bbox": [ + 215, + 391, + 785, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "D. For unsuccessful PGD attacks, the activations of all neurons reduce even though the prediction remains the same as before the attack.", + "bbox": [ + 215, + 417, + 785, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in the important neurons w.r.t. any class $c_{i}$ is,", + "bbox": [ + 215, + 476, + 550, + 491 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {c _ {i}} = \\underset {(x, y) \\in \\mathcal {D}} {\\mathbb {E}} \\left[ A _ {c _ {i}} \\left(x + \\delta^ {*}\\right) - A _ {c _ {i}} (x) \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 500, + 787, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Replacing $c_{i}$ with GT class $y$ , the average activation change in the important neurons of GT class is $\\varDelta_{y}$ . Similarly, the average activation change in post-attack predicted class neurons is $\\varDelta_{\\hat{y}^*}$ , and that in other non-GT class neurons is $\\varDelta_{\\mathrm{non - GT}} = \\mathbb{E}_{c_i\\neq y,\\hat{y}^*}[\\varDelta_{c_i}]$ . For unsuccessful attacks, where post-attack prediction $\\hat{y}^{*}$ is the same as $y$ , we can compute average activation change in the remaining classes (other than GT) as $\\varDelta_{\\mathrm{rem -cls}}=\\mathbb{E}_{c_i\\neq y}[\\varDelta_{c_i}]$ . Finally, we can also compute the average activation change in unimportant neurons,", + "bbox": [ + 212, + 537, + 787, + 643 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {\\text {u n i m p}} = \\underset {(x, y) \\in \\mathcal {D}} {\\mathbb {E}} \\left[ A _ {u} \\left(x + \\delta^ {*}\\right) - A _ {u} (x) \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 652, + 787, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We empirically compute the above metrics on CIFAR10 with five base models (ResNet18: GAT [40], NuAT [41], OAAT [2], DAJAT [1], and WideResNet-34-10: TRADES-AWP [44]) and present the mean values in Fig. 2C and Fig. 2D. See Appendix for more experimental details and corresponding numerical results.", + "bbox": [ + 212, + 689, + 787, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Observations. In Fig. 2C, we observe $\\varDelta_y<0$ and $\\varDelta_{\\hat{y}^*}>0$ , i.e. successful adversarial attacks boost the activations that are important to the post-attack predicted class $\\hat{y}^*$ while causing a drop for those of the GT class $y$ . Further, from Fig. 2C, the important activations for the remaining classes and unimportant activations are only marginally affected by successful attacks. In Fig. 2D, we observe that unsuccessful attacks cause a drop in the activations of all neurons.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Interpretability-Guided Test-Time Adversarial Defense", + "bbox": [ + 366, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remarks. During both successful and unsuccessful attacks, the activation magnitude of GT class' important neurons drops. Specifically, successful attacks shift the activation magnitude from GT class' important neurons to non-GT class important neurons. Hence, we hypothesize that adversarial robustness can be improved if the activation shift to non-GT class important neurons is restricted.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Further, given the availability of a huge number of adversarially pretrained models [15], we aim to evaluate our hypothesis in a test-time adversarial defense setting, with a low inference overhead and without any training overhead.", + "bbox": [ + 212, + 222, + 785, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Interpretability-Guided Defense", + "text_level": 1, + "bbox": [ + 215, + 291, + 562, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we propose a novel, training-free, test-time Interpretability-Guided Defense (IG-Defense). The core idea of our method is based on interpretability-guided masking, described in Sec. 4.1. In Sec. 4.2, we introduce neuron importance ranking, which is a key enabler in the success of our proposed defense.", + "bbox": [ + 212, + 324, + 787, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Setup. Given a trained model $f: \\mathcal{X} \\to \\mathcal{C}$ , the goal of test-time defense is to improve the robustness of $f$ while retaining its utility (i.e. clean accuracy). We refer to $f$ as the \"base model\" in the rest of the paper.", + "bbox": [ + 212, + 385, + 787, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Interpretability-guided masking", + "text_level": 1, + "bbox": [ + 215, + 454, + 524, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate our hypothesis, we design a simple approach for test-time defense where we mask all the neurons except the important neurons of the correct class. We propose a dual forward pass approach, since access to labels cannot be assumed at test-time. Our approach contains three key steps:", + "bbox": [ + 212, + 479, + 785, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Step 1: Neuron Importance Ranking. First, we compute class-wise neuron importance ranking (Step 1 of Fig. 3). This is computed only once for a given base model $f$ . For each layer, we rank every neuron based on its importance to each class. More details are given in Sec. 4.2.", + "bbox": [ + 225, + 551, + 787, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Given the number of important neurons $k$ (hyperparameter), we compute a binary mask $m \\in \\{0,1\\}^{N \\times C}$ using the indices of important neurons $n_{c_i} \\in \\mathbb{N}^k$ (from the importance ranking in Sec. 4.2) for each class $c_i \\in [C]$ .", + "bbox": [ + 253, + 612, + 787, + 659 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nm \\left[ n _ {c _ {i}}, c _ {i} \\right] = 1 \\forall c _ {i} \\in [ C ], \\quad m \\left[ \\left[ N \\right] \\backslash n _ {c _ {i}}, c _ {i} \\right] = 0 \\forall c _ {i} \\in [ C ] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 672, + 787, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- In the mask $m$ , the top- $k$ important neurons of each class $c_i \\in [C]$ are retained in $m[:, c_i]$ . The remaining neurons unimportant to $c_i$ are masked.", + "bbox": [ + 253, + 696, + 787, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Step 2: Vanilla Forward Pass. In the first forward pass (Step 2 of Fig. 3), we obtain a soft-pseudo-label $\\hat{y} = \\sigma\\left(\\frac{f(x)}{\\tau}\\right) \\in (0,1)^C$ where $\\sigma(\\cdot)$ denotes the softmax function, and $\\tau$ denotes the temperature term which controls the sharpness of the soft-pseudo-label.", + "bbox": [ + 223, + 727, + 787, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Step 3: Masked Forward Pass. In the second forward pass (Step 3 of Fig. 3), we apply a soft-pseudo-label weighted mask i.e. $m\\hat{y} \\in (0,1)^N$ to the activations of the layer being masked. Let $A:\\mathcal{X}\\to \\mathbb{R}^{N\\times H\\times W}$ give", + "bbox": [ + 223, + 794, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "A. Kulkarni and T. W. Weng", + "bbox": [ + 271, + 114, + 468, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/36c4a5e322c35445267218808a94d2a4326a72d99e46df3c2dbd9f1459f16acd.jpg", + "image_caption": [ + "Fig. 3: Step 1. Given a pretrained base model $f$ (e.g. binary classifier here), class-wise neuron importance is computed for a selected layer (Sec. 4.2). A top- $k$ mask $m \\in \\{0,1\\}^{N \\times C}$ is computed to identify top- $k$ neurons important to each class (e.g. $k = 2, N = 5, C = 2$ here). Step 2. During evaluation, a soft-pseudo-label $\\hat{y}$ is computed using the base model $f$ . Step 3. The soft-pseudo-label weighted mask $m\\hat{y} = m\\sigma(\\frac{f(x)}{\\tau}) \\in \\mathbb{R}^N$ is applied to the selected layer to retain only the important neurons of the pseudo-label class." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 266 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the activations at the layer being masked, and $h: \\mathbb{R}^{N \\times H \\times W} \\to \\mathcal{C}$ be the remaining layers after the layer being masked, i.e. $f = h \\circ A$ . Then, the final prediction is given by $\\hat{y}' = h(m\\hat{y} \\cdot A(x))$ .", + "bbox": [ + 240, + 402, + 785, + 448 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The masking re-weights the neurons based on $\\hat{y}$ . However, the temperature term $\\tau$ is set to a small value so that $\\hat{y}$ contains a sharp probability distribution. Thus, the weighted mask corresponds mainly to a single class and retains only the important neurons of that class.", + "- Also note that this enables gradient flow between the two forward passes, to avoid any gradient masking [5]." + ], + "bbox": [ + 254, + 449, + 785, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Intuitively, activations of a deeper layer (e.g. penultimate layer) are masked to minimize any negative impact of the masking on feature extraction, which depends more on the shallower layers [51]. To verify this, we analyze the sensitivity of our method to the layer being masked in Sec. 5.3b.", + "bbox": [ + 212, + 545, + 787, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dependence on soft-pseudo-label. This is an obvious limitation if the soft-pseudo-label is computed directly from the base model $f$ . If it is incorrect, masking will yield no improvement in robustness. Hence, an adaptive attack can be designed to ensure that the soft-pseudo-label is incorrect. For instance, a transfer attack that transfers the perturbation from the base model $f$ would yield an incorrect soft-pseudo-label, and the final robustness would be the same as the base model $f$ . To circumvent this, we propose to use randomized smoothing in the first forward pass (i.e. in Step 2 of Fig. 3).", + "bbox": [ + 212, + 606, + 787, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Randomized smoothing. To obtain the soft-pseudo-label $\\hat{y}$ in a more robust manner, we perform randomized smoothing [13] in the first forward pass. We add a set of $n_s$ Gaussian noises (with zero mean and standard deviation $\\sigma_d$ ) to each input $x$ . For the soft-pseudo-label, the logits of these noisy inputs are averaged and softmax $\\sigma$ is applied. Formally,", + "bbox": [ + 212, + 726, + 787, + 801 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} = \\sigma \\left(\\frac {1}{\\tau} \\underset {v _ {i}} {\\mathbb {E}} [ f (x + v _ {i}) ]\\right); v _ {i} \\sim \\mathcal {N} \\left(0 _ {d}, \\sigma_ {d} ^ {2} I _ {d}\\right) \\forall i \\in [ n _ {s} ] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 810, + 787, + 844 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Interpretability-Guided Test-Time Adversarial Defense", + "bbox": [ + 366, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/517f549af567e09635029a08a80529426a4100b6f1a1ac8f0401ba853b427c4e.jpg", + "image_caption": [ + "Fig. 4: A. Leave-one-Out Importance Ranking (LO-IR) computes class-wise neuron importance as the average change in logits when masking each neuron. B. CLIP-Dissect Importance Ranking (CD-IR) first computes the inner product of class name text embeddings and probing set image embeddings. The importance ranking relies on the similarity between activations of each neuron to the precomputed inner product given the same probing set inputs. C. A higher logit-change $\\Delta_{c_i}^{[j]}$ or similarity $s_{c_i}^{[j]}$ implies that neuron $j$ is more important to class $c_i$ ." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 787, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $0_{d}, I_{d}$ are tensors with the same shape as $x$ but containing all zeros and ones respectively. The number of noises sampled $n_{s}$ and standard deviation $\\sigma_{d}$ are hyperparameters.", + "bbox": [ + 212, + 493, + 787, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Importance ranking methods", + "text_level": 1, + "bbox": [ + 214, + 558, + 503, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, we describe two novel importance ranking methods that can be used in Step 1 of our proposed IG-Defense.", + "bbox": [ + 212, + 580, + 785, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a) Leave-one-Out Importance Ranking (L0-IR). Following the analysis in Sec. 3, given a neuron $j$ , we can compute its importance to a class $c_{i}$ as the average change in the class $c_{i}$ logits, $\\Delta_{c_i}^{[j]}$ , before and after masking out that particular neuron $j$ (Fig. 4A). Intuitively, a higher logit-change for a particular class implies higher dependence of the network on that neuron, i.e. higher importance (Fig. 4C). Formally,", + "bbox": [ + 212, + 611, + 787, + 704 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {c _ {i}} ^ {[ j ]} = \\underset {(x, y) \\in \\mathcal {D} _ {\\text {p r o b e}} | y = c _ {i}} {\\mathbb {E}} \\left[ f _ {c _ {i}} (x) - f _ {c _ {i}} ^ {[ j ]} (x) \\right]; \\forall c _ {i} \\in \\mathcal {C} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 710, + 787, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $f^{[j]}$ is the model where $j^{\\mathrm{th}}$ neuron of the selected layer is masked in the base model $f$ . And the subscript $c_{i}$ indicates the logits of class $c_{i}$ . However, this method could be computationally expensive as average change in logits has to be computed for each neuron. For our proposed method, it is only required once per base model. Thus, it is a feasible option that yields a very good estimate of the importance ranking, which we show in Sec. 5.", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "A. Kulkarni and T. W. Weng", + "bbox": [ + 271, + 114, + 468, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "b) CLIP-Dissect Importance Ranking (CD-IR). CLIP-Dissect [33] uses the multimodal CLIP model [35] to assign concept labels to individual neurons. We extend this idea by replacing concepts with task labels to obtain an importance ranking of neurons for each task label (see Fig. 4B).", + "bbox": [ + 212, + 145, + 782, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Concretely, probing images are passed through the image encoder $E_{I}$ to get a set of image embeddings $\\{A^{[i]} = E_I(x_i)\\}_{i = 1}^M$ (shown in orange in Fig. 4B) while the task label names $\\mathcal{C}$ are passed through the text encoder $E_{T}$ to get the corresponding set of text embeddings $\\{E_T(c_k)\\}_{c_k\\in \\mathcal{C}}$ (shown in green in Fig. 4B). These embeddings i.e. vectors are converted to scalars by taking the inner product of each pair of image-text embeddings, which results in a matrix $P$ with entries $P_{ik} = A^{[i]^\\top}E_T(c_k)$ . Now, for a neuron $j$ , a vector of its activations $q_{j}$ (shown in purple in Fig. 4B) can be obtained where each entry corresponds to each input from the probing dataset. We can compute the similarity score $s_{c_k}^{[j]}$ between each class $c_{k}$ and neuron $j$ using the $k^{\\mathrm{th}}$ column of $P$ and $q_{j}$ . Intuitively, a higher similarity score $s_{c_k}^{[j]}$ implies higher importance of neuron $j$ for class $c_{k}$ (Fig. 4C). We use the same soft-WPMI similarity metric as the original paper [33]. For the probing dataset, we simply use the training dataset for our experiments.", + "bbox": [ + 212, + 207, + 787, + 407 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In practice, CD-IR is computationally less expensive compared to L0-IR (e.g. 2 mins. vs. 28 mins. for 512 neurons). This is because CLIP embeddings can be precomputed and most of the computations can be vectorized easily. As we demonstrate in Sec. 5, although the importance ranking quality is slightly worse than L0-IR (i.e. CD-IR causes a marginal drop in clean accuracy), downstream robustness gains with CD-IR are still significant. Please refer to the Appendix C.3 for details on computational complexity comparisons.", + "bbox": [ + 212, + 409, + 787, + 513 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 537, + 375, + 555 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We extensively evaluate our IG-Defense on the standard RobustBench benchmark [15]. We also follow Croce et al. [16] for properly evaluating test-time defenses and thoroughly test against an ensemble of strong adaptive attacks.", + "bbox": [ + 212, + 566, + 787, + 612 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Experimental setup. We evaluate our methods on standard datasets, CIFAR10, CIFAR100 [27] and ImageNet-1k [36], with ResNet [22] and WideResNet [50] base models. We use an $\\ell_{\\infty}$ threat model with an $\\epsilon$ -bound of 8/255 for CIFAR10/100 and 4/255 for ImageNet-1k following prior works [15]. For evaluation, we use the full CIFAR10/100 test set (10000 samples) and the first 5000 ImageNet-1k validation samples, which is the standard setup as per [15,37]. The hyperparameter $k$ for the top- $k$ neurons in CD-IR or L0-IR is mentioned alongside the method name in the results. For randomized smoothing (RS), we use $n_s = 1$ , $\\sigma_d = \\frac{\\epsilon}{2}$ , and $\\tau = 0.01$ , where $\\epsilon$ is the $\\ell_{\\infty}$ -bound for the attack. We show in Appendix C.2 that the choices of $n_s = 1$ , $\\sigma_d = \\frac{\\epsilon}{2}$ are sufficient for our purpose. Unless otherwise mentioned, RS is used and the penultimate layer output is masked. In the penultimate layer, there are 512 neurons (layer4) in ResNet18, 640 neurons (layer3 or block3) in WideResNet-34-10, and 2048 neurons (layer4) in ResNet50.", + "bbox": [ + 212, + 613, + 787, + 809 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Choice of base models. We experiment with base models from several single-step defenses like FAT [42], GAT [40], NuAT [41], OAAT [2], and multi-step", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Interpretability-Guided Test-Time Adversarial Defense", + "bbox": [ + 366, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/14f7c573807a76a780780ce2c066c5c0236ee456aeb81b7c0ae3e512610b105e.jpg", + "table_caption": [ + "Table 1: Evaluating our IG-Defense (CD-IR, LO-IR) on CIFAR10 with ResNet18 and $\\ell_{\\infty}, \\epsilon = 8/255$ . IW-WC indicates image-wise worst-case, AA indicates AutoAttack." + ], + "table_footnote": [], + "table_body": "
Base Model:NuAT [41] (RN18)OAAT [2] (RN18)DAJAT [1] (RN18)
Test-Time DefenseClean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)
None82.2150.5850.5480.2351.1051.0185.7152.5052.45
CD-IR-50 (Ours)82.4652.0851.58 (+1.04)79.8554.7952.35 (+1.34)85.1154.8153.36 (+0.91)
L0-IR-50 (Ours)82.1752.4152.19 (+1.65)79.9054.9052.30 (+1.29)85.1854.5353.34 (+0.89)
", + "bbox": [ + 217, + 184, + 782, + 258 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/f5fa2c3e92e425c4062639bd4f7cbacb03c4efb9a51db5860d3a1269b95b853d.jpg", + "table_caption": [ + "Table 2: Evaluating our IG-Defense (CD-IR, L0-IR) on CIFAR10 with WRN-34-10 and $\\ell_{\\infty}, \\epsilon = 8/255$ . IW-WC indicates image-wise worst-case, AA indicates AutoAttack." + ], + "table_footnote": [], + "table_body": "
Base Model:NuAT [41] (WRN-34-10)TR-AWP [52] (WRN-34-10)DAJAT [1] (WRN-34-10)
Test-Time DefenseClean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)
None86.3254.7554.7285.3656.1756.1288.7157.8157.72
CD-IR-50 (Ours)86.2857.6956.38 (+1.66)84.9860.2657.46 (+1.34)88.5160.7059.22 (+1.50)
LO-IR-50 (Ours)86.2157.0456.40 (+1.68)85.1759.6157.88 (+1.76)88.6160.0859.41 (+1.69)
", + "bbox": [ + 217, + 311, + 782, + 385 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "defenses like TRADES [52], AWP [44], DAJAT [1], RTB [37], and LAS [25]. We chose these models to ensure a diverse evaluation and showcase the generality of our IG-Defense across base model defense strategies. We also evaluate with the same base model on different architectures as well as on different datasets for a thorough analysis. Please see Appendix B.2 for more details on these defenses. Evaluation. We evaluate our IG-Defense with the base models against a diverse range of attacks. AutoAttack [18] (AA) is an empirically strong attack with a sequence of white-box AutoPGD (APGD) attacks, fast adaptive boundary (FAB) attack [17], and black-box Square attack [4]. For a stronger adaptive attack, we devise an ensemble of attacks. For each image, the adaptive attack is considered successful if any one attack from the ensemble is successful. With this, we get the image-wise worst-case (IW-WC) robust accuracy. To be adaptive for our IG-Defense, the ensemble also includes transfer attacks that specifically target our weakness, i.e. our dependence on pseudo-labels (as discussed in Sec. 4.1).", + "bbox": [ + 212, + 414, + 785, + 627 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Robust accuracy (IW-WC) is computed with 9 attacks. The first three are computed on the defended model: standard white-box AutoAttack, decision-boundary based black-box RayS attack [12], and white-box APGD with Expectation over Transformation (EoT) [5] attack. For the remaining 6 transfer attacks, we compute perturbations on the base model using 6 attacks: APGD with cross-entropy loss, APGD with Carlini-Wagner loss [11], targeted APGD attack with difference of logits ratio loss, standard AutoAttack, RayS attack, and $\\mathrm{APGD + EoT}$ attack. Design choices for IW-WC. While AutoAttack also uses APGD attacks, it returns unperturbed inputs when the attack is unsuccessful. Following [16], the four transfer APGD attacks in IW-WC each return a sample with the highest loss (i.e. closest to the decision boundary). This is a stronger attack than AutoAttack for test-time defenses since samples closer to the decision boundary tend to be misclassified under test-time defenses, as observed by [16]. Expectation over Transformation (EoT) [5] averages gradients over multiple iterations of", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "A. Kulkarni and T. W. Weng", + "bbox": [ + 271, + 114, + 467, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/6b967ed743821a571a10a045aca7740d6fe9a2086c52d0bc084f583f1506da5d.jpg", + "table_caption": [ + "Table 3: Evaluating our IG-Defense (CD-IR, L0-IR) on CIFAR100 with $\\ell_{\\infty}, \\epsilon = 8 / 255$ . IW-WC indicates image-wise worst-case, AA indicates AutoAttack." + ], + "table_footnote": [], + "table_body": "
Base Model:LAS [25] (WRN-34-10)OAAT [2] (RN18)DAJAT [1] (RN18)
Test-Time DefenseClean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)
None64.8930.7430.7162.0227.1527.1265.4527.6727.64
CD-IR-50 (Ours)62.1333.5731.41 (+0.70)61.5629.5129.19 (+2.07)64.9730.2228.27 (+0.63)
LO-IR-50 (Ours)64.8535.6932.22 (+1.51)62.0432.0630.12 (+3.00)65.3732.5528.54 (+0.90)
", + "bbox": [ + 217, + 184, + 782, + 257 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/4f2135ee7db0bfc7f6ceb7a972de2c4b6deed79260e5baf65b571ba1cd45ed49.jpg", + "table_caption": [ + "Table 4: Evaluating our IG-Defense (CD-IR, L0-IR) on ImageNet-1k with ResNet50 and $\\ell_{\\infty}, \\epsilon = 4/255$ . IW-WC indicates image-wise worst-case, AA indicates AutoAttack." + ], + "table_footnote": [], + "table_body": "
Base Model:FAT [42] (RN50)RTB [37] (RN50)
Test-Time DefenseClean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)
None55.6426.2426.2464.1034.6234.62
CD-IR-750 (Ours)54.3027.28 (+1.04)26.80 (+0.56)62.5036.42 (+1.80)35.30 (+0.68)
LO-IR-750 (Ours)55.8829.14 (+2.90)27.08 (+0.84)64.1037.36 (+2.74)35.48 (+0.86)
", + "bbox": [ + 217, + 311, + 782, + 395 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "a particular input to counter the effect of randomness in a defense. Specifically, we use 20 iterations of EoT with APGD (more than $n_s = 1$ ). While [16] also recommend using Backward Pass Differentiable Approximation (BPDA) [5] for test-time defenses with non-differentiable components, our approach already includes BPDA since we use a soft-masking approach instead of hard-masking. Refer to Appendix B.3 for complete implementation details.", + "bbox": [ + 212, + 425, + 787, + 516 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Evaluation on RobustBench", + "text_level": 1, + "bbox": [ + 215, + 540, + 491, + 555 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "a) CIFAR10. In Table 1, we evaluate our methods on CIFAR10 with ResNet18 at $\\epsilon = 8 / 255$ . We observe consistent gains in AutoAttack (avg. $+2.5\\%$ for CD-IR and $+2.55\\%$ for L0-IR) and IW-WC robustness (avg. $+1.1\\%$ for CD-IR and $+1.28\\%$ for L0-IR), with a marginal drop in clean accuracy (avg. $-0.24\\%$ for CD-IR and $-0.3\\%$ for L0-IR). In Table 2, we evaluate on CIFAR10 with the more challenging, large WideResNet-34-10 model at $\\epsilon = 8 / 255$ . We observe consistent gains in IW-WC robustness (avg. $+1.5\\%$ for CD-IR and $+1.7\\%$ for L0-IR), with a marginal drop in clean accuracy (avg. $-0.2\\%$ for CD-IR and $-0.13\\%$ for L0-IR). b) CIFAR100. In Table 3, we evaluate our methods on CIFAR100 with ResNet18 and WideResNet-34-10 at $\\epsilon = 8 / 255$ . Despite the increased number of classes, we get consistently improved worst-case (IW-WC) robust accuracy (avg. $+1.13\\%$ for CD-IR and $+1.8\\%$ for L0-IR). However, the clean accuracy drops for CD-IR (avg. $-0.43\\%$ drop), but only marginally for L0-IR.", + "c) ImageNet-1k. In Table 4, we evaluate our methods on ImageNet-1k with ResNet50 at $\\epsilon = 4 / 255$ . Despite the increased image size and the higher number of classes, we observe consistent gains in worst-case (IW-WC) robustness (avg. $+0.6\\%$ for CD-IR and $+0.85\\%$ for L0-IR) but with a drop in clean accuracy for CD-IR (avg. $-1.47\\%$ similar to CIFAR100). Given the scale of ImageNet-1k, even" + ], + "bbox": [ + 212, + 566, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Interpretability-Guided Test-Time Adversarial Defense", + "bbox": [ + 366, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/8925e1082464aeead1743c17ba5a1526c47c660fd6eba9f13952da1b095924ff.jpg", + "table_caption": [ + "Table 5: Comparison against state-of-the-art test-time defenses with ResNet18 and WideResNet-34-10 on CIFAR10. The number in (green) shows the worst-case robustness gain over the base model. The first row is the base model without any test-time defense. Inference time is a multiple of the time for a single forward pass of the base model." + ], + "table_footnote": [], + "table_body": "
Base Model → Test-Time DefenseOAAT [2] (RN18)DAJAT [1] (RN18)TR-AWP [44] (WRN)Inference Time
Clean Acc.Robust Acc. (IW-WC)Clean Acc.Robust Acc. (IW-WC)Clean Acc.Robust Acc. (IW-WC)
No test-time def.80.2351.0185.7152.4585.3656.121×
Existing Defenses
HD [43]79.8950.68 (-0.33)84.5352.55 (+0.10)84.7855.72 (-0.40)46×
SODEF [26]80.2350.67 (-0.34)84.8652.95 (+0.50)85.2556.05 (-0.07)2×
CAAA [3]80.2351.00 (-0.01)85.7152.45 (+0.00)85.3556.11 (-0.01)8×
Our IG-Defense
CD-IR-5079.8552.35 (+1.34)85.1153.36 (+0.91)84.9857.46 (+1.34)2×
L0-IR-5079.9052.30 (+1.29)85.1853.34 (+0.89)85.1757.88 (+1.76)2×
", + "bbox": [ + 220, + 212, + 787, + 361 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "robustness improvements of $\\sim 0.5\\%$ are quite significant, especially since there is no clean accuracy drop with LO-IR.", + "bbox": [ + 212, + 388, + 784, + 417 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Why does the clean accuracy drop? There are two possible reasons for the drop in clean accuracy. First, due to the randomized smoothing, some samples may get incorrect predictions if the random noise is significant enough to change the class prediction w.r.t. the base model. Second, importance ranking quality and number of neurons masked determine how much clean accuracy can be retained. That is, better quality or lower number of neurons masked would guarantee better clean accuracy retention. However, simply masking lower number of neurons may also reduce the improvements in robustness.", + "bbox": [ + 212, + 417, + 785, + 537 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We see in Table 3, 4 that clean accuracy drop for CD-IR is more for CIFAR100 and ImageNet1k compared to CIFAR10 (Table 1, 2). This is possibly because the ratio of number of neurons to number of classes is decreasing, i.e. more number of neurons will be important to multiple classes. This cannot be captured well by CD-IR, leading to a larger drop in clean accuracy. Note that LO-IR is less susceptible to this problem as the importance ranking of each neuron is computed independently, but it is more expensive. This is why we need to retain more number of important neurons for ImageNet ( $k = 750$ ) than CIFAR ( $k = 50$ ).", + "bbox": [ + 212, + 539, + 785, + 659 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Note that the clean accuracy drops are acceptable since they are significantly smaller than the worst-case robustness gains (e.g. upto $-0.6\\%$ and $+1.7\\%$ respectively for CIFAR10 in Table 1-2). Further, in some cases like ImageNet-1k with LO-IR (Table 4), there are no drops with pure gains in robustness.", + "bbox": [ + 212, + 660, + 785, + 720 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Comparisons with existing test-time defenses", + "text_level": 1, + "bbox": [ + 214, + 739, + 633, + 757 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We compare with three recent state-of-the-art test-time defense methods on CIFAR10 in Table 5. CAAA [3] (input purification) optimizes an \"anti-adversary\" perturbation in a direction opposite to an adversarial attack, by minimizing the loss based on a pseudo-label, to counter the effect of any adversarial perturbation. HD or Hedge Defense [43] (input purification) finds a perturbation that maximizes", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "A. Kulkarni and T. W. Weng", + "bbox": [ + 271, + 114, + 467, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3ac288f046bda527dedb18a14cc180e693373db2dc20a52eeaca6b5fec58b3f3.jpg", + "image_caption": [ + "Fig. 5: Sensitivity analyses for number of retained important neurons $k$ and layer being masked with CIFAR10. For A, B: layer4 with total 512 neurons is used. A lower $k$ and a deeper layer yield better adversarial robustness with minimal loss of clean accuracy." + ], + "image_footnote": [], + "bbox": [ + 222, + 147, + 352, + 220 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/bc7d75d5851e386c8a9880951ca38dcf33b9598a2a7107a8081763c6bf634b2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 148, + 493, + 220 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/417d66132b5d41f9462b058afbd7f54e7b033632bde7a0260337b9717d4cc383.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 148, + 637, + 220 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/74948447b5099fe5443435a8c2c52fd23e911e74e10b22fb195dfee967cba7ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 148, + 779, + 220 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the cross-entropy loss summed over all classes. SODEF [26] (model adaptation) trains a neural ODE block with Lyapunov stability constraints between the pretrained feature extractor and classifier, which reduces the influence of the adversarial perturbation. In contrast, our method uses neither input purification nor model adaptation, since we only mask the activations.", + "bbox": [ + 212, + 306, + 782, + 382 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Table 5, we observe that existing test-time defenses like HD, CAAA, and SODEF do not significantly improve, but rather also reduce the worst-case robust accuracy (IW-WC) compared to the base model. Based on Croce et al. [16], CAAA is the strongest test-time defense among the nine defenses that they evaluated, since its worst-case robust accuracy did not drop below that of the base model. Hence, we only compare with the two strongest input purification defenses, CAAA and HD as well as an efficient model adaptation defense, SODEF from the nine defenses in [16]. Overall, we find that our proposed IG-Defense obtains consistent improvements in image-wise worst-case (IW-WC) robust accuracy, unlike existing test-time defenses. Note that we focus mainly on the worst-case (IW-WC) robust accuracy because gains under weaker attacks would not reflect fundamental gains in robustness [16].", + "bbox": [ + 212, + 383, + 787, + 565 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Analysis", + "text_level": 1, + "bbox": [ + 215, + 585, + 333, + 602 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "a) Ablation study. In Table 6, we study the significance of each proposed component of our approach. First, we start with the baseline (row #1) which uses only one forward pass without masking and without randomized smoothing (RS). Next, we add RS to this baseline (row #2) and observe no gains in robustness and a small drop in clean accuracy. This shows that RS alone cannot improve the robustness. Then, we perform two forward passes without masking and without RS (row #3). Here, the robust accuracy is the same as the baseline indicating that there is no gradient masking with two forward passes. Next, we perform random masking with two forward passes, without and with RS (row #4, #5). Each class is assigned a random set of neurons as the \"important\" neurons to perform the masking. We observe no gains in robustness, validating the significance of our neuron importance ranking. Finally, we show the impact of RS with our CD-IR (row #6, #7) and L0-IR based masking (row #8, #9). Here, we observe consistent gains in robustness, with a small drop in clean accuracy since RS likely changed the prediction w.r.t. the base model for a few samples.", + "bbox": [ + 212, + 609, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Interpretability-Guided Test-Time Adversarial Defense", + "bbox": [ + 366, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "b) Sensitivity analyses. We study the sensitivity of our proposed methods to the number of retained important neurons $k$ in Fig. 5A, B on CIFAR10. Specifically, we test it on ResNet18 layer4 which has total 512 neurons and we vary $k$ from 10 to 500. We observe that the robustness improvement is higher at lower $k$ , although lower than 10 may reduce the clean accuracy due to a high loss", + "bbox": [ + 212, + 146, + 403, + 387 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/e8605505bd2101aacb401971b0fab44204ac6139ec848e1bd144174281df1d22.jpg", + "table_caption": [ + "Table 6: Ablation study of proposed components with TRADES-AWP [44] (WideResNet-34-10) as the base model. RS indicates use of randomized smoothing in the first forward pass. The first row is the base model performance without any test-time defense." + ], + "table_footnote": [], + "table_body": "
Row #No. of forw. passesMasking typeRSClean Acc.Rob. Acc. (AA)Rob. Acc. (IW-WC)
11None✗85.3656.1756.12
21None✓84.9356.1956.14
32None✗85.3556.1556.12
42Random-50✗85.2356.1456.12
52Random-50✓85.3356.2056.15
62CD-IR-50✗85.0959.0056.12
72CD-IR-50✓84.9860.2657.46
82LO-IR-50✗85.3358.5956.12
92LO-IR-50✓85.1759.6157.88
", + "bbox": [ + 413, + 232, + 784, + 369 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "of information. For CD-IR, the clean accuracy can be retained up to $k = 100$ while for L0-IR, it can be retained up to at least $k = 400$ . As explained earlier, it is due to the relatively lower importance ranking quality of CD-IR w.r.t. L0-IR. However, both methods are fairly insensitive to the choice of $k$ .", + "bbox": [ + 212, + 388, + 787, + 446 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We also study the sensitivity to choice of layer being masked in Fig. 5C, D on CIFAR10 with ResNet18. We fix the number of neurons being retained $k$ to 50 and vary the layer being masked from layer1 (shallower) to layer4 (deeper). We observe that masking deeper layers is better, since masking shallower layers affects the feature extraction [51], leading to a large drop in clean accuracy, and consequently a drop in robust accuracy. For all the other experiments, we simply mask the penultimate layer since the clean accuracy drop is the lowest. Please refer to Appendix C for more analysis experiments.", + "bbox": [ + 212, + 450, + 787, + 571 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "c) Inference time comparisons. Since we use $n_s = 1$ in randomized smoothing and perform two forward passes, the inference time of IG-Defense is $2 \\times$ of a single forward pass. The compared test-time defenses HD [43], SODEF [26], and CAAA [3] have computational complexities of $46 \\times$ , $2 \\times$ , and $8 \\times$ respectively. Also, the recent review paper [16] shows that the lowest inference time is $2 \\times$ for existing test-time defenses. Hence, as per Table 5, we achieve the most gains in worst-case robustness with the lowest inference time.", + "bbox": [ + 212, + 571, + 789, + 676 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 709, + 359, + 724 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we proposed a novel, training-free, and effective test-time defense (IG-Defense) guided by interpretability through neuron importance ranking. Our IG-Defense outperformed existing test-time defenses on worst-case adaptive attacks while being among the most efficient test-time defenses to date, establishing the new state-of-the-art. We also demonstrated consistent improvements across standard CIFAR and ImageNet-1k benchmarks.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "A. Kulkarni and T. W. Weng", + "bbox": [ + 271, + 114, + 467, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported in part by National Science Foundation (NSF) awards CNS-1730158, ACI-1540112, ACI1541349, OAC-1826967, OAC-2112167, CNS-2100237, CNS-2120019, the University of California Office of the President, and the University of California San Diego's California Institute for Telecommunications and Information Technology/Qualcomm Institute. Thanks to CENIC for the 100Gbps networks. The authors thank the anonymous reviewers for valuable feedback on the manuscript. T.-W. Weng is supported by National Science Foundation under Grant No. 2107189 and 2313105. T.-W. Weng also thanks the Hellman Fellowship for providing research support.", + "bbox": [ + 212, + 176, + 787, + 313 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 335, + 321, + 351 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Addepalli, S., Jain, S., Babu, R.V.: Efficient and effective augmentation strategy for adversarial training. In: NeurIPS (2022)", + "2. Addepalli, S., Jain, S., Sriramananan, G., Venkatesh Babu, R.: Scaling adversarial training to large perturbation bounds. In: ECCV (2022)", + "3. Alfarra, M., Pérez, J.C., Thabet, A., Bibi, A., Torr, P.H.S., Ghanem, B.: Combating adversaries with anti-adversaries. In: AAAI (2022)", + "4. Andriushchenko, M., Croce, F., Flammarion, N., Hein, M.: Square attack: a query-efficient black-box adversarial attack via random search. In: ECCV (2020)", + "5. Athalye, A., Carlini, N., Wagner, D.: Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In: ICML (2018)", + "6. Bai, N., Iyer, R.A., Oikarinen, T., Weng, T.W.: Describe-and-dissect: Interpreting neurons in vision networks with language models. arXiv preprint arXiv:2403.13771 (2024)", + "7. Bai, Y., Zeng, Y., Jiang, Y., Xia, S.T., Ma, X., Wang, Y.: Improving adversarial robustness via channel-wise activation suppressing. In: ICLR (2021)", + "8. Bau, D., Zhou, B., Khosla, A., Oliva, A., Torralba, A.: Network dissection: Quantifying interpretability of deep visual representations. In: CVPR (2017)", + "9. Bau, D., Zhu, J.Y., Strobelt, H., Lapedriza, A., Zhou, B., Torralba, A.: Understanding the role of individual units in a deep neural network. Proceedings of the National Academy of Sciences (2020)", + "10. Boopathy, A., Liu, S., Zhang, G., Liu, C., Chen, P.Y., Chang, S., Daniel, L.: Proper network interpretability helps adversarial robustness in classification. In: ICML (2020)", + "1.1. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: IEEE Symposium on Security and Privacy (2017)", + "2. Chen, J., Gu, Q.: RayS: A ray searching method for hard-label adversarial attack In: KDD (2020)", + "3. Chen, Y., Oliver, D.S.: Ensemble randomized maximum likelihood method as an iterative ensemble smoother. Mathematical Geosciences 44, 1-26 (2012)", + "4. Chen, Z., Li, Q., Zhang, Z.: Towards robust neural networks via close-loop control In: ICLR (2021)", + "5. Croce, F., Andriushchenko, M., Sehwag, V., Debenedetti, E., Flammarion, N., Chiang, M., Mittal, P., Hein, M.: Robustbench: a standardized adversarial robustness benchmark. In: Datasets and Benchmarks Track, NeurIPS (2021)" + ], + "bbox": [ + 225, + 367, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Interpretability-Guided Test-Time Adversarial Defense", + "bbox": [ + 366, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Croce, F., Gowal, S., Brunner, T., Shelhamer, E., Hein, M., Cemgil, T.: Evaluating the adversarial robustness of adaptive test-time defenses. In: ICML (2022)", + "17. Croce, F., Hein, M.: Minimally distorted adversarial examples with a fast adaptive boundary attack. In: ICML (2020)", + "18. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: ICML (2020)", + "19. Eigen, H., Sadovnik, A.: TopKConv: Increased adversarial robustness through deeper interpretability. In: ICMLA (2021)", + "20. Gerasimou, S., Eniser, H.F., Sen, A., Cakan, A.: Importance-driven deep learning system testing. In: ACM/IEEE ICSE (2020)", + "21. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: ICLR (2015)", + "22. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition In: CVPR (2016)", + "23. Hernandez, E., Schwettmann, S., Bau, D., Bagashvili, T., Torralba, A., Andreas, J.: Natural language descriptions of deep visual features. In: ICLR (2022)", + "24. Hwang, D., Lee, E., Rhee, W.: AID-Purifier: A light auxiliary network for boosting adversarial defense. In: ICPR (2022)", + "25. Jia, X., Zhang, Y., Wu, B., Ma, K., Wang, J., Cao, X.: LAS-AT: adversarial training with learnable attack strategy. In: CVPR (2022)", + "26. Kang, Q., Song, Y., Ding, Q., Tay, W.P.: Stable neural ODE with Lyapunov-stable equilibrium points for defending against adversarial attacks. In: NeurIPS (2021)", + "27. Krizhevsky, A.: Learning multiple layers of features from tiny images. Tech. rep., CIFAR (2009)", + "28. Kundu, S., Nazemi, M., Beerel, P.A., Pedram, M.: A tunable robust pruning framework through dynamic network rewiring of DNNs. In: ASP-DAC (2021)", + "29. Mangla, P., Singh, V., Balasubramanian, V.N.: On saliency maps and adversarial robustness. In: ECML-PKDD (2020)", + "30. Mao, C., Chiquier, M., Wang, H., Yang, J., Vondrick, C.: Adversarial attacks are reversible with natural supervision. In: ICCV (2021)", + "31. Mu, J., Andreas, J.: Compositional explanations of neurons. In: NeurIPS (2020)", + "32. Nayak, G.K., Rawal, R., Chakraborty, A.: DAD: Data-free adversarial defense at test time. In: WACV (2022)", + "33. Oikarinen, T., Weng, T.W.: CLIP-Dissect: Automatic description of neuron representations in deep vision networks. In: ICLR (2023)", + "34. Oikarinen, T., Weng, T.W.: Linear explanations for individual neurons. In: ICML (2024)", + "35. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)", + "36. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) 115(3), 211-252 (2015)", + "37. Salman, H., Ilyas, A., Engstrom, L., Kapoor, A., Madry, A.: Do adversarially robust imagenet models transfer better? In: NeurIPS (2020)", + "38. Sehwag, V., Wang, S., Mittal, P., Jana, S.: Hydra: Pruning adversarially robust neural networks. In: NeurIPS (2020)", + "39. Shi, C., Holtz, C., Mishne, G.: Online adversarial purification based on self-supervised learning. In: ICLR (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "A. Kulkarni and T. W. Weng", + "bbox": [ + 271, + 114, + 468, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "40. Sriramanan, G., Addepalli, S., Baburaj, A., Babu, R.V.: Guided adversarial attack for evaluating and enhancing adversarial defenses. In: NeurIPS (2020)", + "41. Sriramanan, G., Addepalli, S., Baburaj, A., Babu, R.V.: Towards efficient and effective adversarial training. In: NeurIPS (2021)", + "42. Wong, E., Rice, L., Kolter, J.Z.: Fast is better than free: Revisiting adversarial training. In: ICLR (2020)", + "43. Wu, B., Pan, H., Shen, L., Gu, J., Zhao, S., Li, Z., Cai, D., He, X., Liu, W.: Attacking adversarial attacks as a defense. arXiv preprint arXiv:2106.04938 (2021)", + "44. Wu, D., Xia, S.T., Wang, Y.: Adversarial weight perturbation helps robust generalization. In: NeurIPS (2020)", + "45. Wu, S., Sang, J., Xu, K., Zhang, J., Yu, J.: Attention, please! adversarial defense via activation rectification and preservation. ACM Transactions on Multimedia Computing, Communications and Applications 19(4), 1-18 (2023)", + "46. Wu, Y.H., Yuan, C.H., Wu, S.H.: Adversarial robustness via runtime masking and cleansing. In: ICML (2020)", + "47. Xiao, C., Zhong, P., Zheng, C.: Enhancing adversarial defense by k-winners-take-all In: ICLR (2020)", + "48. Xie, X., Li, T., Wang, J., Ma, L., Guo, Q., Yuefei-Xu, F., Liu, Y.: NPC: Neuron path coverage via characterizing decision logic of deep neural networks. ACM Transactions on Software Engineering and Methodology 31(3) (2022)", + "49. Yoon, J., Hwang, S.J., Lee, J.: Adversarial purification with score-based generative models. In: ICML (2021)", + "50. Zagoruyko, S., Komodakis, N.: Wide residual networks. In: BMVC (2016)", + "51. Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: ECCV (2014)", + "52. Zhang, H., Yu, Y., Jiao, J., Xing, E.P., Ghaoui, L.E., Jordan, M.I.: Theoretically principled trade-off between robustness and accuracy. In: ICML (2019)", + "53. Zhao, Q., Wressnegger, C.: Holistic adversarially robust pruning. In: ICLR (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 535 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Interpretability-Guided Test-Time Adversarial Defense", + "bbox": [ + 367, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Interpretability-Guided Test-Time Adversarial Defense/d8bed3ff-2ae8-4bfb-abf2-7d81aee96721_model.json b/2024/Interpretability-Guided Test-Time Adversarial Defense/d8bed3ff-2ae8-4bfb-abf2-7d81aee96721_model.json new file mode 100644 index 0000000000000000000000000000000000000000..45838b42d6a33f22d0c4dd09d8223c73dfd4e9f8 --- /dev/null +++ b/2024/Interpretability-Guided Test-Time Adversarial Defense/d8bed3ff-2ae8-4bfb-abf2-7d81aee96721_model.json @@ -0,0 +1,2313 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.237, + 0.141, + 0.768, + 0.185 + ], + "angle": 0, + "content": "Interpretability-Guided Test-Time Adversarial Defense" + }, + { + "type": "text", + "bbox": [ + 0.354, + 0.212, + 0.651, + 0.229 + ], + "angle": 0, + "content": "Akshay Kulkarni\\(^{1}\\) and Tsui-Wei Weng\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.424, + 0.239, + 0.577, + 0.254 + ], + "angle": 0, + "content": "1 CSE, UC San Diego" + }, + { + "type": "text", + "bbox": [ + 0.424, + 0.254, + 0.58, + 0.269 + ], + "angle": 0, + "content": "\\(^{2}\\) HDSI, UC San Diego" + }, + { + "type": "text", + "bbox": [ + 0.397, + 0.27, + 0.605, + 0.282 + ], + "angle": 0, + "content": "{a2kulkarni,lweng}@ucsd.edu" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.318, + 0.744, + 0.485 + ], + "angle": 0, + "content": "Abstract. We propose a novel and low-cost test-time adversarial defense by devising interpretability-guided neuron importance ranking methods to identify neurons important to the output classes. Our method is a training-free approach that can significantly improve the robustness-accuracy tradeoff while incurring minimal computational overhead. While being among the most efficient test-time defenses (4× faster), our method is also robust to a wide range of black-box, white-box, and adaptive attacks that break previous test-time defenses. We demonstrate the efficacy of our method for CIFAR10, CIFAR100, and ImageNet-1k on the standard RobustBenchmark (with average gains of 2.6%, 4.9%, and 2.8% respectively). We also show improvements (average 1.5%) over the state-of-the-art test-time defenses even under strong adaptive attacks." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.531, + 0.377, + 0.547 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.79, + 0.667 + ], + "angle": 0, + "content": "Despite the popularity and success of deep neural networks (DNNs), they have been shown to be brittle against carefully designed small perturbations that are imperceptible to humans [21]. These \"adversarial\" perturbations pose significant risks against deploying DNNs for safety-critical tasks like autonomous driving. To mitigate these risks, there has been great interest in developing defenses, e.g. adversarial training algorithms [40, 42, 44, 52] to obtain a model that is more robust to adversarial samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.668, + 0.79, + 0.804 + ], + "angle": 0, + "content": "However, adversarial training is much more expensive than standard training (e.g. a strong defense like TRADES [52] takes \\(150 - 200 \\times\\) longer than standard training). This inspired training-free defenses known as test-time defenses [3, 26, 30, 39]. The underlying principles in existing test-time defenses are either input purification or model adaptation: input purification [3, 39] modifies the input with an additional perturbation to nullify any adversarial perturbation, whereas model adaptation [14, 26] updates model parameters or adds new parameters at test-time to make the overall model more robust. Despite eliminating the need for training, most existing methods are still quite expensive, \\(8 - 500 \\times\\) more than" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.81, + 0.79, + 0.839 + ], + "angle": 0, + "content": "\\(^3\\) Our code is available at https://github.com/Trustworthy-ML-Lab/Interpretability-Guided-Defense" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.471, + 0.129 + ], + "angle": 0, + "content": "A. Kulkarni and T. W. Weng" + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.146, + 0.728, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.356, + 0.79, + 0.385 + ], + "angle": 0, + "content": "Fig. 1: Interpretability-guided masking based on neuron importance ranking for test-time adversarial defense." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.789, + 0.476 + ], + "angle": 0, + "content": "a standard forward pass [16]. Moreover, existing test-time defenses have been broken under strong adaptive attacks [16]. Motivated by these limitations, in this paper, we intend to find an efficient and effective alternative to improve robustness at test-time, so that the defense does not break under adaptive attacks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.79, + 0.765 + ], + "angle": 0, + "content": "To design an effective defense, we leverage recent neuron-level interpretability methods [8, 33], which allow one to interpret the functionalities of individual neurons. In our analysis, we observed that for successful adversarial examples, the majority of activations at an intermediate layer shift from important neurons for the ground-truth class to the important neurons for other classes. Hence, we propose a novel test-time Interpretability-Guided Defense (IG-Defense) enabled by \"neuron importance ranking\" where class-wise importance of each neuron is computed (Fig. 1A). To design effective neuron importance ranking, we repurpose our analysis experiments and existing neuron-level interpretability works to propose two novel neuron importance ranking methods. The key idea of our IG-Defense is to restrict the observed activation shift by masking unimportant neurons (Fig. 1B). We demonstrate the usefulness of our insights by extensively evaluating our IG-Defense on the standard RobustBench benchmark (Fig. 1C). We observe consistent gains for CIFAR10 (+2.6%), CIFAR100 (+4.9%), and ImageNet-1k (+2.8%). We also devise an ensemble of strong adaptive attacks specifically targeting the weaknesses of IG-Defense and observe consistent gains (+1.5%), unlike existing test-time defenses that break under adaptive attacks. Due to our simple yet grounded approach, IG-Defense is among the most efficient test-time defenses (4× faster than the strongest existing defense)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.767, + 0.578, + 0.781 + ], + "angle": 0, + "content": "Our primary contributions can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.795, + 0.789, + 0.84 + ], + "angle": 0, + "content": "- We are the first to propose a neuron-interpretability-guided test-time defense (IG-Defense) utilizing neuron importance ranking (CD-IR, LO-IR) to improve adversarial robustness. IG-Defense is training-free, efficient, and effective." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.367, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Interpretability-Guided Test-Time Adversarial Defense" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "- We uncover novel insights into improving adversarial robustness by analyzing adversarial attacks through the lens of neuron-level interpretability in Sec. 3." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.177, + 0.786, + 0.251 + ], + "angle": 0, + "content": "- Our proposed IG-Defense consistently improves the robustness on standard CIFAR10, CIFAR100, and ImageNet-1k benchmarks. We also demonstrate improved robustness upto \\(3.4\\%\\), \\(3.8\\%\\), and \\(1.5\\%\\) against a wide range of white-box, black-box, and adaptive attacks respectively with the lowest inference time (\\(4\\times\\) faster) among existing test-time defenses." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.27, + 0.388, + 0.286 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Test-time adversarial robustness. Most of the test-time defenses can be categorized into two approaches: input purification or model adaptation. In input purification, the adversarial input is converted to a safer input by an additional perturbation. For instance, finding a perturbation that either increases top class confidence [3], minimizes contrastive loss [30], or minimizes a self-supervised objective [39]. [24] crafts the perturbation by fooling a real-adversarial discriminator [24] and [49] uses an input passed through a score-based generative model. For model adaptation [14,32,46], the model parameters are either updated at test-time or new parameters are introduced. However, existing test-time defenses are quite expensive (8-500× of single forward pass) and break under adaptive attacks. In contrast, our method is fast and effective (1.5% gain under adaptive attacks with only 2× inference time), and is the first test-time defense to leverage neuron-level interpretability information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.789, + 0.658 + ], + "angle": 0, + "content": "Neuron interpretability methods. To understand the functionalities (i.e. concepts) of individual neurons in DNNs, early works like NetDissect [8,9] compare the neuron activation patterns with patterns of predefined concept label masks. MILAN [23] produced natural language descriptions using a supervised image captioning model trained on a large-scale concept-labeled dataset. Recently, CLIP-Dissect [33] leveraged the CLIP paradigm [35] to identify concepts of neurons more efficiently than prior work and without the need of training or densely-concept-labeled datasets. There are other works to identify neuron concepts through clustering [20, 48], via language models [6], or using linear combinations [34]. In this work, we focus on repurposing CLIP-Dissect due to its efficiency and simplicity as well as NetDissect to propose neuron importance ranking methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Connections between interpretability and robustness. Prior works attempt to improve robustness via input-dependent interpretability methods like saliency or class activation maps [10, 29]. Wu et al. [45] observed that adversarial attacks lead to changes in attention maps, and proposed a training-time defense with losses to enforce similar attention maps for clean and adversarial samples. Several works [7, 19, 28, 38, 47, 53] use sparsity to improve adversarial robustness. However, they require retraining after or while sparsity constraints are applied. In contrast, we indirectly leverage sparsity without requiring any training. NetDissect [8] investigated the connection between adversarial attacks and interpretability information from their method, but they did not explore how to improve robustness. Hence, we aim to study the unexplored connection of neuron-level interpretability and improving adversarial robustness." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.468, + 0.128 + ], + "angle": 0, + "content": "A. Kulkarni and T. W. Weng" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.463, + 0.163 + ], + "angle": 0, + "content": "3 Analysis Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.185, + 0.788, + 0.381 + ], + "angle": 0, + "content": "To understand the connection of interpretability of individual neurons with robustness, we aim to design an analysis experiment. One way would be to investigate the changes in neuron activations before and after an adversarial attack. In neuron-labeling works [9,31], they assign an explainable concept to each neuron (or an activation channel in case of CNNs). Specifically, NetDissect [9] qualitatively show that a targeted adversarial attack causes a drop in peak activations for neurons whose concepts are related to the original class, while causing an increase in peak activations for neurons whose concepts are related to the target class of the attack. They also show that the peak activation change is lower for neurons with unrelated concepts. While they only perform the analysis to showcase their interpretability, we are the first to analyze adversarial attacks through neuron-level interpretability with the explicit goal of obtaining insights into improving adversarial robustness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.519 + ], + "angle": 0, + "content": "For a more thorough analysis, we observe the average activation change (instead of peak change) and associate the neurons with the category names directly instead of concepts. First, peak activation change may be misleading since it may be caused by specific images while average change takes all images into account for a more well-rounded analysis. Second, our automated evaluation eliminates the need for manual association of concepts with categories [9], which could be difficult, expensive, and noisy, especially with a large number of related categories (e.g. ImageNet classes). We analyze both successful and unsuccessful attacks while [9] only focused on successful attacks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.788, + 0.595 + ], + "angle": 0, + "content": "Concretely, we find top-\\(k\\) important neurons related to each task category using neuron importance ranking (discussed in Sec. 4.2). With this, in Fig. 2, we analyze the change in activations corresponding to ground-truth (GT) and non-GT categories before and after an attack. We discuss some preliminaries before delving into further details." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.597, + 0.789, + 0.793 + ], + "angle": 0, + "content": "Preliminaries. We denote the dataset as \\(\\mathcal{D} = \\{(x,y)\\}\\) where \\(x\\in \\mathcal{X}\\) is the input image, \\(y\\in \\mathcal{C}\\) is the label, \\(\\mathcal{C}\\) is the label set, and \\(|\\mathcal{C}|\\) is the number of classes. Given a pretrained model \\(f:\\mathcal{X}\\to \\mathcal{C}\\) and its activations at the layer being studied \\(A:\\mathcal{X}\\rightarrow \\mathbb{R}^{N\\times H\\times W}\\), we can rank the importance of a neuron (out of \\(N\\) neurons/channels) by assessing its contribution to each class. We call this \"neuron importance ranking\" and propose two algorithms for this (Sec. 4.2). Then, we can compute the activations of the top-\\(k\\) important neurons for any class \\(c_{i}\\) as \\(A_{c_i}(x) = A(x)_{[n_{c_i}]} \\in \\mathbb{R}^{k\\times H\\times W}\\), where \\(n_{c_i}\\) are the indices of the top-\\(k\\) activations for class \\(c_{i}\\). With access to each \\(n_{c_i}\\), the remaining indices absent from the top-\\(k\\) of any class are \\(n_u = [N] \\setminus \\{\\cup_{c_i\\in \\mathcal{C}} n_{c_i}\\}\\), where \\([N] = \\{1,2,\\dots,N\\}\\) and \\(\\backslash\\) indicates setminus operator. These indices \\(n_u\\) denote unimportant neurons, and corresponding unimportant activations can be computed as \\(A_u(x) = A(x)_{[n_u]} \\in \\mathbb{R}^{(N - k|\\mathcal{C}|) \\times H \\times W}\\). Next, we define activation change metrics to support our analysis." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Metrics. For an adversarial attack with \\(\\ell_{\\infty}\\)-norm bound of \\(\\epsilon\\), the perturbation for input \\(x\\) is \\(\\delta^{*} = \\arg \\max_{\\| \\delta \\|_{\\infty} \\leq \\epsilon} \\ell(f(x + \\delta), y)\\) where \\(\\ell\\) is the attack objective. Let \\(\\hat{y}^{*} = f(x + \\delta^{*})\\) be the post-attack prediction. The average activation change" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.367, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Interpretability-Guided Test-Time Adversarial Defense" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.145, + 0.728, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.349, + 0.788, + 0.391 + ], + "angle": 0, + "content": "Fig. 2: A. Cat-dog-bird image classifier before a PGD attack. B. After PGD attack, the prediction changes from the ground truth (GT) cat to dog since the activations of neurons important to dog class increase while those important to cat class decrease." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.392, + 0.786, + 0.418 + ], + "angle": 0, + "content": "C. Empirically, successful PGD attacks show a decrease in activations of important GT class neurons while those of post-attack predicted class' important neurons increase." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.418, + 0.786, + 0.446 + ], + "angle": 0, + "content": "D. For unsuccessful PGD attacks, the activations of all neurons reduce even though the prediction remains the same as before the attack." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.477, + 0.551, + 0.492 + ], + "angle": 0, + "content": "in the important neurons w.r.t. any class \\( c_{i} \\) is," + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.501, + 0.788, + 0.527 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {c _ {i}} = \\underset {(x, y) \\in \\mathcal {D}} {\\mathbb {E}} \\left[ A _ {c _ {i}} \\left(x + \\delta^ {*}\\right) - A _ {c _ {i}} (x) \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.644 + ], + "angle": 0, + "content": "Replacing \\( c_{i} \\) with GT class \\( y \\), the average activation change in the important neurons of GT class is \\( \\varDelta_{y} \\). Similarly, the average activation change in post-attack predicted class neurons is \\( \\varDelta_{\\hat{y}^*} \\), and that in other non-GT class neurons is \\( \\varDelta_{\\mathrm{non - GT}} = \\mathbb{E}_{c_i\\neq y,\\hat{y}^*}[\\varDelta_{c_i}] \\). For unsuccessful attacks, where post-attack prediction \\( \\hat{y}^{*} \\) is the same as \\( y \\), we can compute average activation change in the remaining classes (other than GT) as \\( \\varDelta_{\\mathrm{rem -cls}}=\\mathbb{E}_{c_i\\neq y}[\\varDelta_{c_i}] \\). Finally, we can also compute the average activation change in unimportant neurons," + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.654, + 0.788, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {\\text {u n i m p}} = \\underset {(x, y) \\in \\mathcal {D}} {\\mathbb {E}} \\left[ A _ {u} \\left(x + \\delta^ {*}\\right) - A _ {u} (x) \\right] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.749 + ], + "angle": 0, + "content": "We empirically compute the above metrics on CIFAR10 with five base models (ResNet18: GAT [40], NuAT [41], OAAT [2], DAJAT [1], and WideResNet-34-10: TRADES-AWP [44]) and present the mean values in Fig. 2C and Fig. 2D. See Appendix for more experimental details and corresponding numerical results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Observations. In Fig. 2C, we observe \\(\\varDelta_y<0\\) and \\(\\varDelta_{\\hat{y}^*}>0\\), i.e. successful adversarial attacks boost the activations that are important to the post-attack predicted class \\(\\hat{y}^*\\) while causing a drop for those of the GT class \\(y\\). Further, from Fig. 2C, the important activations for the remaining classes and unimportant activations are only marginally affected by successful attacks. In Fig. 2D, we observe that unsuccessful attacks cause a drop in the activations of all neurons." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.47, + 0.129 + ], + "angle": 0, + "content": "A. Kulkarni and T. W. Weng" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "Remarks. During both successful and unsuccessful attacks, the activation magnitude of GT class' important neurons drops. Specifically, successful attacks shift the activation magnitude from GT class' important neurons to non-GT class important neurons. Hence, we hypothesize that adversarial robustness can be improved if the activation shift to non-GT class important neurons is restricted." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.787, + 0.269 + ], + "angle": 0, + "content": "Further, given the availability of a huge number of adversarially pretrained models [15], we aim to evaluate our hypothesis in a test-time adversarial defense setting, with a low inference overhead and without any training overhead." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.292, + 0.563, + 0.31 + ], + "angle": 0, + "content": "4 Interpretability-Guided Defense" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.385 + ], + "angle": 0, + "content": "In this section, we propose a novel, training-free, test-time Interpretability-Guided Defense (IG-Defense). The core idea of our method is based on interpretability-guided masking, described in Sec. 4.1. In Sec. 4.2, we introduce neuron importance ranking, which is a key enabler in the success of our proposed defense." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.788, + 0.432 + ], + "angle": 0, + "content": "Setup. Given a trained model \\( f: \\mathcal{X} \\to \\mathcal{C} \\), the goal of test-time defense is to improve the robustness of \\( f \\) while retaining its utility (i.e. clean accuracy). We refer to \\( f \\) as the \"base model\" in the rest of the paper." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.455, + 0.525, + 0.47 + ], + "angle": 0, + "content": "4.1 Interpretability-guided masking" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.787, + 0.542 + ], + "angle": 0, + "content": "To validate our hypothesis, we design a simple approach for test-time defense where we mask all the neurons except the important neurons of the correct class. We propose a dual forward pass approach, since access to labels cannot be assumed at test-time. Our approach contains three key steps:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.552, + 0.788, + 0.612 + ], + "angle": 0, + "content": "- Step 1: Neuron Importance Ranking. First, we compute class-wise neuron importance ranking (Step 1 of Fig. 3). This is computed only once for a given base model \\( f \\). For each layer, we rank every neuron based on its importance to each class. More details are given in Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.613, + 0.788, + 0.66 + ], + "angle": 0, + "content": "- Given the number of important neurons \\( k \\) (hyperparameter), we compute a binary mask \\( m \\in \\{0,1\\}^{N \\times C} \\) using the indices of important neurons \\( n_{c_i} \\in \\mathbb{N}^k \\) (from the importance ranking in Sec. 4.2) for each class \\( c_i \\in [C] \\)." + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.673, + 0.788, + 0.691 + ], + "angle": 0, + "content": "\\[\nm \\left[ n _ {c _ {i}}, c _ {i} \\right] = 1 \\forall c _ {i} \\in [ C ], \\quad m \\left[ \\left[ N \\right] \\backslash n _ {c _ {i}}, c _ {i} \\right] = 0 \\forall c _ {i} \\in [ C ] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.697, + 0.788, + 0.728 + ], + "angle": 0, + "content": "- In the mask \\( m \\), the top-\\( k \\) important neurons of each class \\( c_i \\in [C] \\) are retained in \\( m[:, c_i] \\). The remaining neurons unimportant to \\( c_i \\) are masked." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.728, + 0.788, + 0.795 + ], + "angle": 0, + "content": "- Step 2: Vanilla Forward Pass. In the first forward pass (Step 2 of Fig. 3), we obtain a soft-pseudo-label \\(\\hat{y} = \\sigma\\left(\\frac{f(x)}{\\tau}\\right) \\in (0,1)^C\\) where \\(\\sigma(\\cdot)\\) denotes the softmax function, and \\(\\tau\\) denotes the temperature term which controls the sharpness of the soft-pseudo-label." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "- Step 3: Masked Forward Pass. In the second forward pass (Step 3 of Fig. 3), we apply a soft-pseudo-label weighted mask i.e. \\(m\\hat{y} \\in (0,1)^N\\) to the activations of the layer being masked. Let \\(A:\\mathcal{X}\\to \\mathbb{R}^{N\\times H\\times W}\\) give" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.367, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Interpretability-Guided Test-Time Adversarial Defense" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.788, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.278, + 0.79, + 0.377 + ], + "angle": 0, + "content": "Fig. 3: Step 1. Given a pretrained base model \\( f \\) (e.g. binary classifier here), class-wise neuron importance is computed for a selected layer (Sec. 4.2). A top- \\( k \\) mask \\( m \\in \\{0,1\\}^{N \\times C} \\) is computed to identify top- \\( k \\) neurons important to each class (e.g. \\( k = 2, N = 5, C = 2 \\) here). Step 2. During evaluation, a soft-pseudo-label \\( \\hat{y} \\) is computed using the base model \\( f \\). Step 3. The soft-pseudo-label weighted mask \\( m\\hat{y} = m\\sigma(\\frac{f(x)}{\\tau}) \\in \\mathbb{R}^N \\) is applied to the selected layer to retain only the important neurons of the pseudo-label class." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.403, + 0.787, + 0.449 + ], + "angle": 0, + "content": "the activations at the layer being masked, and \\( h: \\mathbb{R}^{N \\times H \\times W} \\to \\mathcal{C} \\) be the remaining layers after the layer being masked, i.e. \\( f = h \\circ A \\). Then, the final prediction is given by \\( \\hat{y}' = h(m\\hat{y} \\cdot A(x)) \\)." + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.45, + 0.787, + 0.508 + ], + "angle": 0, + "content": "- The masking re-weights the neurons based on \\(\\hat{y}\\). However, the temperature term \\(\\tau\\) is set to a small value so that \\(\\hat{y}\\) contains a sharp probability distribution. Thus, the weighted mask corresponds mainly to a single class and retains only the important neurons of that class." + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.509, + 0.787, + 0.539 + ], + "angle": 0, + "content": "- Also note that this enables gradient flow between the two forward passes, to avoid any gradient masking [5]." + }, + { + "type": "list", + "bbox": [ + 0.255, + 0.45, + 0.787, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.788, + 0.606 + ], + "angle": 0, + "content": "Intuitively, activations of a deeper layer (e.g. penultimate layer) are masked to minimize any negative impact of the masking on feature extraction, which depends more on the shallower layers [51]. To verify this, we analyze the sensitivity of our method to the layer being masked in Sec. 5.3b." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.788, + 0.727 + ], + "angle": 0, + "content": "Dependence on soft-pseudo-label. This is an obvious limitation if the soft-pseudo-label is computed directly from the base model \\( f \\). If it is incorrect, masking will yield no improvement in robustness. Hence, an adaptive attack can be designed to ensure that the soft-pseudo-label is incorrect. For instance, a transfer attack that transfers the perturbation from the base model \\( f \\) would yield an incorrect soft-pseudo-label, and the final robustness would be the same as the base model \\( f \\). To circumvent this, we propose to use randomized smoothing in the first forward pass (i.e. in Step 2 of Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.727, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Randomized smoothing. To obtain the soft-pseudo-label \\(\\hat{y}\\) in a more robust manner, we perform randomized smoothing [13] in the first forward pass. We add a set of \\(n_s\\) Gaussian noises (with zero mean and standard deviation \\(\\sigma_d\\)) to each input \\(x\\). For the soft-pseudo-label, the logits of these noisy inputs are averaged and softmax \\(\\sigma\\) is applied. Formally," + }, + { + "type": "equation", + "bbox": [ + 0.309, + 0.811, + 0.788, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\hat {y} = \\sigma \\left(\\frac {1}{\\tau} \\underset {v _ {i}} {\\mathbb {E}} [ f (x + v _ {i}) ]\\right); v _ {i} \\sim \\mathcal {N} \\left(0 _ {d}, \\sigma_ {d} ^ {2} I _ {d}\\right) \\forall i \\in [ n _ {s} ] \\tag {4}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.47, + 0.129 + ], + "angle": 0, + "content": "A. Kulkarni and T. W. Weng" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.145, + 0.788, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.369, + 0.789, + 0.466 + ], + "angle": 0, + "content": "Fig. 4: A. Leave-one-Out Importance Ranking (LO-IR) computes class-wise neuron importance as the average change in logits when masking each neuron. B. CLIP-Dissect Importance Ranking (CD-IR) first computes the inner product of class name text embeddings and probing set image embeddings. The importance ranking relies on the similarity between activations of each neuron to the precomputed inner product given the same probing set inputs. C. A higher logit-change \\(\\Delta_{c_i}^{[j]}\\) or similarity \\(s_{c_i}^{[j]}\\) implies that neuron \\(j\\) is more important to class \\(c_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.789, + 0.54 + ], + "angle": 0, + "content": "where \\(0_{d}, I_{d}\\) are tensors with the same shape as \\(x\\) but containing all zeros and ones respectively. The number of noises sampled \\(n_{s}\\) and standard deviation \\(\\sigma_{d}\\) are hyperparameters." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.559, + 0.504, + 0.575 + ], + "angle": 0, + "content": "4.2 Importance ranking methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.787, + 0.611 + ], + "angle": 0, + "content": "Here, we describe two novel importance ranking methods that can be used in Step 1 of our proposed IG-Defense." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.705 + ], + "angle": 0, + "content": "a) Leave-one-Out Importance Ranking (L0-IR). Following the analysis in Sec. 3, given a neuron \\( j \\), we can compute its importance to a class \\( c_{i} \\) as the average change in the class \\( c_{i} \\) logits, \\( \\Delta_{c_i}^{[j]} \\), before and after masking out that particular neuron \\( j \\) (Fig. 4A). Intuitively, a higher logit-change for a particular class implies higher dependence of the network on that neuron, i.e. higher importance (Fig. 4C). Formally," + }, + { + "type": "equation", + "bbox": [ + 0.324, + 0.711, + 0.788, + 0.741 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {c _ {i}} ^ {[ j ]} = \\underset {(x, y) \\in \\mathcal {D} _ {\\text {p r o b e}} | y = c _ {i}} {\\mathbb {E}} \\left[ f _ {c _ {i}} (x) - f _ {c _ {i}} ^ {[ j ]} (x) \\right]; \\forall c _ {i} \\in \\mathcal {C} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\( f^{[j]} \\) is the model where \\( j^{\\mathrm{th}} \\) neuron of the selected layer is masked in the base model \\( f \\). And the subscript \\( c_{i} \\) indicates the logits of class \\( c_{i} \\). However, this method could be computationally expensive as average change in logits has to be computed for each neuron. For our proposed method, it is only required once per base model. Thus, it is a feasible option that yields a very good estimate of the importance ranking, which we show in Sec. 5." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.367, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Interpretability-Guided Test-Time Adversarial Defense" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.207 + ], + "angle": 0, + "content": "b) CLIP-Dissect Importance Ranking (CD-IR). CLIP-Dissect [33] uses the multimodal CLIP model [35] to assign concept labels to individual neurons. We extend this idea by replacing concepts with task labels to obtain an importance ranking of neurons for each task label (see Fig. 4B)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.789, + 0.409 + ], + "angle": 0, + "content": "Concretely, probing images are passed through the image encoder \\( E_{I} \\) to get a set of image embeddings \\( \\{A^{[i]} = E_I(x_i)\\}_{i = 1}^M \\) (shown in orange in Fig. 4B) while the task label names \\( \\mathcal{C} \\) are passed through the text encoder \\( E_{T} \\) to get the corresponding set of text embeddings \\( \\{E_T(c_k)\\}_{c_k\\in \\mathcal{C}} \\) (shown in green in Fig. 4B). These embeddings i.e. vectors are converted to scalars by taking the inner product of each pair of image-text embeddings, which results in a matrix \\( P \\) with entries \\( P_{ik} = A^{[i]^\\top}E_T(c_k) \\). Now, for a neuron \\( j \\), a vector of its activations \\( q_{j} \\) (shown in purple in Fig. 4B) can be obtained where each entry corresponds to each input from the probing dataset. We can compute the similarity score \\( s_{c_k}^{[j]} \\) between each class \\( c_{k} \\) and neuron \\( j \\) using the \\( k^{\\mathrm{th}} \\) column of \\( P \\) and \\( q_{j} \\). Intuitively, a higher similarity score \\( s_{c_k}^{[j]} \\) implies higher importance of neuron \\( j \\) for class \\( c_{k} \\) (Fig. 4C). We use the same soft-WPMI similarity metric as the original paper [33]. For the probing dataset, we simply use the training dataset for our experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.789, + 0.515 + ], + "angle": 0, + "content": "In practice, CD-IR is computationally less expensive compared to L0-IR (e.g. 2 mins. vs. 28 mins. for 512 neurons). This is because CLIP embeddings can be precomputed and most of the computations can be vectorized easily. As we demonstrate in Sec. 5, although the importance ranking quality is slightly worse than L0-IR (i.e. CD-IR causes a marginal drop in clean accuracy), downstream robustness gains with CD-IR are still significant. Please refer to the Appendix C.3 for details on computational complexity comparisons." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.538, + 0.377, + 0.556 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.789, + 0.613 + ], + "angle": 0, + "content": "We extensively evaluate our IG-Defense on the standard RobustBench benchmark [15]. We also follow Croce et al. [16] for properly evaluating test-time defenses and thoroughly test against an ensemble of strong adaptive attacks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.81 + ], + "angle": 0, + "content": "Experimental setup. We evaluate our methods on standard datasets, CIFAR10, CIFAR100 [27] and ImageNet-1k [36], with ResNet [22] and WideResNet [50] base models. We use an \\(\\ell_{\\infty}\\) threat model with an \\(\\epsilon\\)-bound of 8/255 for CIFAR10/100 and 4/255 for ImageNet-1k following prior works [15]. For evaluation, we use the full CIFAR10/100 test set (10000 samples) and the first 5000 ImageNet-1k validation samples, which is the standard setup as per [15,37]. The hyperparameter \\(k\\) for the top-\\(k\\) neurons in CD-IR or L0-IR is mentioned alongside the method name in the results. For randomized smoothing (RS), we use \\(n_s = 1\\), \\(\\sigma_d = \\frac{\\epsilon}{2}\\), and \\(\\tau = 0.01\\), where \\(\\epsilon\\) is the \\(\\ell_{\\infty}\\)-bound for the attack. We show in Appendix C.2 that the choices of \\(n_s = 1\\), \\(\\sigma_d = \\frac{\\epsilon}{2}\\) are sufficient for our purpose. Unless otherwise mentioned, RS is used and the penultimate layer output is masked. In the penultimate layer, there are 512 neurons (layer4) in ResNet18, 640 neurons (layer3 or block3) in WideResNet-34-10, and 2048 neurons (layer4) in ResNet50." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Choice of base models. We experiment with base models from several single-step defenses like FAT [42], GAT [40], NuAT [41], OAAT [2], and multi-step" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.468, + 0.129 + ], + "angle": 0, + "content": "A. Kulkarni and T. W. Weng" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 1: Evaluating our IG-Defense (CD-IR, LO-IR) on CIFAR10 with ResNet18 and \\(\\ell_{\\infty}, \\epsilon = 8/255\\). IW-WC indicates image-wise worst-case, AA indicates AutoAttack." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.185, + 0.784, + 0.259 + ], + "angle": 0, + "content": "
Base Model:NuAT [41] (RN18)OAAT [2] (RN18)DAJAT [1] (RN18)
Test-Time DefenseClean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)
None82.2150.5850.5480.2351.1051.0185.7152.5052.45
CD-IR-50 (Ours)82.4652.0851.58 (+1.04)79.8554.7952.35 (+1.34)85.1154.8153.36 (+0.91)
L0-IR-50 (Ours)82.1752.4152.19 (+1.65)79.9054.9052.30 (+1.29)85.1854.5353.34 (+0.89)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.273, + 0.785, + 0.302 + ], + "angle": 0, + "content": "Table 2: Evaluating our IG-Defense (CD-IR, L0-IR) on CIFAR10 with WRN-34-10 and \\(\\ell_{\\infty}, \\epsilon = 8/255\\). IW-WC indicates image-wise worst-case, AA indicates AutoAttack." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.313, + 0.784, + 0.386 + ], + "angle": 0, + "content": "
Base Model:NuAT [41] (WRN-34-10)TR-AWP [52] (WRN-34-10)DAJAT [1] (WRN-34-10)
Test-Time DefenseClean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)
None86.3254.7554.7285.3656.1756.1288.7157.8157.72
CD-IR-50 (Ours)86.2857.6956.38 (+1.66)84.9860.2657.46 (+1.34)88.5160.7059.22 (+1.50)
LO-IR-50 (Ours)86.2157.0456.40 (+1.68)85.1759.6157.88 (+1.76)88.6160.0859.41 (+1.69)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.787, + 0.628 + ], + "angle": 0, + "content": "defenses like TRADES [52], AWP [44], DAJAT [1], RTB [37], and LAS [25]. We chose these models to ensure a diverse evaluation and showcase the generality of our IG-Defense across base model defense strategies. We also evaluate with the same base model on different architectures as well as on different datasets for a thorough analysis. Please see Appendix B.2 for more details on these defenses. Evaluation. We evaluate our IG-Defense with the base models against a diverse range of attacks. AutoAttack [18] (AA) is an empirically strong attack with a sequence of white-box AutoPGD (APGD) attacks, fast adaptive boundary (FAB) attack [17], and black-box Square attack [4]. For a stronger adaptive attack, we devise an ensemble of attacks. For each image, the adaptive attack is considered successful if any one attack from the ensemble is successful. With this, we get the image-wise worst-case (IW-WC) robust accuracy. To be adaptive for our IG-Defense, the ensemble also includes transfer attacks that specifically target our weakness, i.e. our dependence on pseudo-labels (as discussed in Sec. 4.1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Robust accuracy (IW-WC) is computed with 9 attacks. The first three are computed on the defended model: standard white-box AutoAttack, decision-boundary based black-box RayS attack [12], and white-box APGD with Expectation over Transformation (EoT) [5] attack. For the remaining 6 transfer attacks, we compute perturbations on the base model using 6 attacks: APGD with cross-entropy loss, APGD with Carlini-Wagner loss [11], targeted APGD attack with difference of logits ratio loss, standard AutoAttack, RayS attack, and \\(\\mathrm{APGD + EoT}\\) attack. Design choices for IW-WC. While AutoAttack also uses APGD attacks, it returns unperturbed inputs when the attack is unsuccessful. Following [16], the four transfer APGD attacks in IW-WC each return a sample with the highest loss (i.e. closest to the decision boundary). This is a stronger attack than AutoAttack for test-time defenses since samples closer to the decision boundary tend to be misclassified under test-time defenses, as observed by [16]. Expectation over Transformation (EoT) [5] averages gradients over multiple iterations of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.367, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Interpretability-Guided Test-Time Adversarial Defense" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.173 + ], + "angle": 0, + "content": "Table 3: Evaluating our IG-Defense (CD-IR, L0-IR) on CIFAR100 with \\(\\ell_{\\infty}, \\epsilon = 8 / 255\\). IW-WC indicates image-wise worst-case, AA indicates AutoAttack." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.185, + 0.784, + 0.258 + ], + "angle": 0, + "content": "
Base Model:LAS [25] (WRN-34-10)OAAT [2] (RN18)DAJAT [1] (RN18)
Test-Time DefenseClean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)
None64.8930.7430.7162.0227.1527.1265.4527.6727.64
CD-IR-50 (Ours)62.1333.5731.41 (+0.70)61.5629.5129.19 (+2.07)64.9730.2228.27 (+0.63)
LO-IR-50 (Ours)64.8535.6932.22 (+1.51)62.0432.0630.12 (+3.00)65.3732.5528.54 (+0.90)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.272, + 0.786, + 0.3 + ], + "angle": 0, + "content": "Table 4: Evaluating our IG-Defense (CD-IR, L0-IR) on ImageNet-1k with ResNet50 and \\(\\ell_{\\infty}, \\epsilon = 4/255\\). IW-WC indicates image-wise worst-case, AA indicates AutoAttack." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.313, + 0.784, + 0.396 + ], + "angle": 0, + "content": "
Base Model:FAT [42] (RN50)RTB [37] (RN50)
Test-Time DefenseClean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)
None55.6426.2426.2464.1034.6234.62
CD-IR-750 (Ours)54.3027.28 (+1.04)26.80 (+0.56)62.5036.42 (+1.80)35.30 (+0.68)
LO-IR-750 (Ours)55.8829.14 (+2.90)27.08 (+0.84)64.1037.36 (+2.74)35.48 (+0.86)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.788, + 0.517 + ], + "angle": 0, + "content": "a particular input to counter the effect of randomness in a defense. Specifically, we use 20 iterations of EoT with APGD (more than \\( n_s = 1 \\)). While [16] also recommend using Backward Pass Differentiable Approximation (BPDA) [5] for test-time defenses with non-differentiable components, our approach already includes BPDA since we use a soft-masking approach instead of hard-masking. Refer to Appendix B.3 for complete implementation details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.541, + 0.492, + 0.556 + ], + "angle": 0, + "content": "5.1 Evaluation on RobustBench" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.789, + 0.763 + ], + "angle": 0, + "content": "a) CIFAR10. In Table 1, we evaluate our methods on CIFAR10 with ResNet18 at \\(\\epsilon = 8 / 255\\). We observe consistent gains in AutoAttack (avg. \\(+2.5\\%\\) for CD-IR and \\(+2.55\\%\\) for L0-IR) and IW-WC robustness (avg. \\(+1.1\\%\\) for CD-IR and \\(+1.28\\%\\) for L0-IR), with a marginal drop in clean accuracy (avg. \\(-0.24\\%\\) for CD-IR and \\(-0.3\\%\\) for L0-IR). In Table 2, we evaluate on CIFAR10 with the more challenging, large WideResNet-34-10 model at \\(\\epsilon = 8 / 255\\). We observe consistent gains in IW-WC robustness (avg. \\(+1.5\\%\\) for CD-IR and \\(+1.7\\%\\) for L0-IR), with a marginal drop in clean accuracy (avg. \\(-0.2\\%\\) for CD-IR and \\(-0.13\\%\\) for L0-IR). b) CIFAR100. In Table 3, we evaluate our methods on CIFAR100 with ResNet18 and WideResNet-34-10 at \\(\\epsilon = 8 / 255\\). Despite the increased number of classes, we get consistently improved worst-case (IW-WC) robust accuracy (avg. \\(+1.13\\%\\) for CD-IR and \\(+1.8\\%\\) for L0-IR). However, the clean accuracy drops for CD-IR (avg. \\(-0.43\\%\\) drop), but only marginally for L0-IR." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.84 + ], + "angle": 0, + "content": "c) ImageNet-1k. In Table 4, we evaluate our methods on ImageNet-1k with ResNet50 at \\(\\epsilon = 4 / 255\\). Despite the increased image size and the higher number of classes, we observe consistent gains in worst-case (IW-WC) robustness (avg. \\(+0.6\\%\\) for CD-IR and \\(+0.85\\%\\) for L0-IR) but with a drop in clean accuracy for CD-IR (avg. \\(-1.47\\%\\) similar to CIFAR100). Given the scale of ImageNet-1k, even" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.567, + 0.789, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.468, + 0.129 + ], + "angle": 0, + "content": "A. Kulkarni and T. W. Weng" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 5: Comparison against state-of-the-art test-time defenses with ResNet18 and WideResNet-34-10 on CIFAR10. The number in (green) shows the worst-case robustness gain over the base model. The first row is the base model without any test-time defense. Inference time is a multiple of the time for a single forward pass of the base model." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.213, + 0.789, + 0.362 + ], + "angle": 0, + "content": "
Base Model → Test-Time DefenseOAAT [2] (RN18)DAJAT [1] (RN18)TR-AWP [44] (WRN)Inference Time
Clean Acc.Robust Acc. (IW-WC)Clean Acc.Robust Acc. (IW-WC)Clean Acc.Robust Acc. (IW-WC)
No test-time def.80.2351.0185.7152.4585.3656.121×
Existing Defenses
HD [43]79.8950.68 (-0.33)84.5352.55 (+0.10)84.7855.72 (-0.40)46×
SODEF [26]80.2350.67 (-0.34)84.8652.95 (+0.50)85.2556.05 (-0.07)2×
CAAA [3]80.2351.00 (-0.01)85.7152.45 (+0.00)85.3556.11 (-0.01)8×
Our IG-Defense
CD-IR-5079.8552.35 (+1.34)85.1153.36 (+0.91)84.9857.46 (+1.34)2×
L0-IR-5079.9052.30 (+1.29)85.1853.34 (+0.89)85.1757.88 (+1.76)2×
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.785, + 0.418 + ], + "angle": 0, + "content": "robustness improvements of \\(\\sim 0.5\\%\\) are quite significant, especially since there is no clean accuracy drop with LO-IR." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.787, + 0.538 + ], + "angle": 0, + "content": "Why does the clean accuracy drop? There are two possible reasons for the drop in clean accuracy. First, due to the randomized smoothing, some samples may get incorrect predictions if the random noise is significant enough to change the class prediction w.r.t. the base model. Second, importance ranking quality and number of neurons masked determine how much clean accuracy can be retained. That is, better quality or lower number of neurons masked would guarantee better clean accuracy retention. However, simply masking lower number of neurons may also reduce the improvements in robustness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.787, + 0.66 + ], + "angle": 0, + "content": "We see in Table 3, 4 that clean accuracy drop for CD-IR is more for CIFAR100 and ImageNet1k compared to CIFAR10 (Table 1, 2). This is possibly because the ratio of number of neurons to number of classes is decreasing, i.e. more number of neurons will be important to multiple classes. This cannot be captured well by CD-IR, leading to a larger drop in clean accuracy. Note that LO-IR is less susceptible to this problem as the importance ranking of each neuron is computed independently, but it is more expensive. This is why we need to retain more number of important neurons for ImageNet (\\( k = 750 \\)) than CIFAR (\\( k = 50 \\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.661, + 0.787, + 0.721 + ], + "angle": 0, + "content": "Note that the clean accuracy drops are acceptable since they are significantly smaller than the worst-case robustness gains (e.g. upto \\(-0.6\\%\\) and \\(+1.7\\%\\) respectively for CIFAR10 in Table 1-2). Further, in some cases like ImageNet-1k with LO-IR (Table 4), there are no drops with pure gains in robustness." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.741, + 0.634, + 0.758 + ], + "angle": 0, + "content": "5.2 Comparisons with existing test-time defenses" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We compare with three recent state-of-the-art test-time defense methods on CIFAR10 in Table 5. CAAA [3] (input purification) optimizes an \"anti-adversary\" perturbation in a direction opposite to an adversarial attack, by minimizing the loss based on a pseudo-label, to counter the effect of any adversarial perturbation. HD or Hedge Defense [43] (input purification) finds a perturbation that maximizes" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.367, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Interpretability-Guided Test-Time Adversarial Defense" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.148, + 0.353, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.149, + 0.495, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.149, + 0.638, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.149, + 0.78, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.237, + 0.788, + 0.28 + ], + "angle": 0, + "content": "Fig. 5: Sensitivity analyses for number of retained important neurons \\( k \\) and layer being masked with CIFAR10. For A, B: layer4 with total 512 neurons is used. A lower \\( k \\) and a deeper layer yield better adversarial robustness with minimal loss of clean accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.784, + 0.383 + ], + "angle": 0, + "content": "the cross-entropy loss summed over all classes. SODEF [26] (model adaptation) trains a neural ODE block with Lyapunov stability constraints between the pretrained feature extractor and classifier, which reduces the influence of the adversarial perturbation. In contrast, our method uses neither input purification nor model adaptation, since we only mask the activations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.789, + 0.566 + ], + "angle": 0, + "content": "In Table 5, we observe that existing test-time defenses like HD, CAAA, and SODEF do not significantly improve, but rather also reduce the worst-case robust accuracy (IW-WC) compared to the base model. Based on Croce et al. [16], CAAA is the strongest test-time defense among the nine defenses that they evaluated, since its worst-case robust accuracy did not drop below that of the base model. Hence, we only compare with the two strongest input purification defenses, CAAA and HD as well as an efficient model adaptation defense, SODEF from the nine defenses in [16]. Overall, we find that our proposed IG-Defense obtains consistent improvements in image-wise worst-case (IW-WC) robust accuracy, unlike existing test-time defenses. Note that we focus mainly on the worst-case (IW-WC) robust accuracy because gains under weaker attacks would not reflect fundamental gains in robustness [16]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.587, + 0.334, + 0.603 + ], + "angle": 0, + "content": "5.3 Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.789, + 0.84 + ], + "angle": 0, + "content": "a) Ablation study. In Table 6, we study the significance of each proposed component of our approach. First, we start with the baseline (row #1) which uses only one forward pass without masking and without randomized smoothing (RS). Next, we add RS to this baseline (row #2) and observe no gains in robustness and a small drop in clean accuracy. This shows that RS alone cannot improve the robustness. Then, we perform two forward passes without masking and without RS (row #3). Here, the robust accuracy is the same as the baseline indicating that there is no gradient masking with two forward passes. Next, we perform random masking with two forward passes, without and with RS (row #4, #5). Each class is assigned a random set of neurons as the \"important\" neurons to perform the masking. We observe no gains in robustness, validating the significance of our neuron importance ranking. Finally, we show the impact of RS with our CD-IR (row #6, #7) and L0-IR based masking (row #8, #9). Here, we observe consistent gains in robustness, with a small drop in clean accuracy since RS likely changed the prediction w.r.t. the base model for a few samples." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.468, + 0.128 + ], + "angle": 0, + "content": "A. Kulkarni and T. W. Weng" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.405, + 0.388 + ], + "angle": 0, + "content": "b) Sensitivity analyses. We study the sensitivity of our proposed methods to the number of retained important neurons \\(k\\) in Fig. 5A, B on CIFAR10. Specifically, we test it on ResNet18 layer4 which has total 512 neurons and we vary \\(k\\) from 10 to 500. We observe that the robustness improvement is higher at lower \\(k\\), although lower than 10 may reduce the clean accuracy due to a high loss" + }, + { + "type": "table_caption", + "bbox": [ + 0.413, + 0.163, + 0.788, + 0.232 + ], + "angle": 0, + "content": "Table 6: Ablation study of proposed components with TRADES-AWP [44] (WideResNet-34-10) as the base model. RS indicates use of randomized smoothing in the first forward pass. The first row is the base model performance without any test-time defense." + }, + { + "type": "table", + "bbox": [ + 0.415, + 0.233, + 0.785, + 0.371 + ], + "angle": 0, + "content": "
Row #No. of forw. passesMasking typeRSClean Acc.Rob. Acc. (AA)Rob. Acc. (IW-WC)
11None✗85.3656.1756.12
21None✓84.9356.1956.14
32None✗85.3556.1556.12
42Random-50✗85.2356.1456.12
52Random-50✓85.3356.2056.15
62CD-IR-50✗85.0959.0056.12
72CD-IR-50✓84.9860.2657.46
82LO-IR-50✗85.3358.5956.12
92LO-IR-50✓85.1759.6157.88
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.789, + 0.448 + ], + "angle": 0, + "content": "of information. For CD-IR, the clean accuracy can be retained up to \\( k = 100 \\) while for L0-IR, it can be retained up to at least \\( k = 400 \\). As explained earlier, it is due to the relatively lower importance ranking quality of CD-IR w.r.t. L0-IR. However, both methods are fairly insensitive to the choice of \\( k \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.451, + 0.789, + 0.572 + ], + "angle": 0, + "content": "We also study the sensitivity to choice of layer being masked in Fig. 5C, D on CIFAR10 with ResNet18. We fix the number of neurons being retained \\( k \\) to 50 and vary the layer being masked from layer1 (shallower) to layer4 (deeper). We observe that masking deeper layers is better, since masking shallower layers affects the feature extraction [51], leading to a large drop in clean accuracy, and consequently a drop in robust accuracy. For all the other experiments, we simply mask the penultimate layer since the clean accuracy drop is the lowest. Please refer to Appendix C for more analysis experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.573, + 0.79, + 0.678 + ], + "angle": 0, + "content": "c) Inference time comparisons. Since we use \\( n_s = 1 \\) in randomized smoothing and perform two forward passes, the inference time of IG-Defense is \\( 2 \\times \\) of a single forward pass. The compared test-time defenses HD [43], SODEF [26], and CAAA [3] have computational complexities of \\( 46 \\times \\), \\( 2 \\times \\), and \\( 8 \\times \\) respectively. Also, the recent review paper [16] shows that the lowest inference time is \\( 2 \\times \\) for existing test-time defenses. Hence, as per Table 5, we achieve the most gains in worst-case robustness with the lowest inference time." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.71, + 0.36, + 0.726 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this work, we proposed a novel, training-free, and effective test-time defense (IG-Defense) guided by interpretability through neuron importance ranking. Our IG-Defense outperformed existing test-time defenses on worst-case adaptive attacks while being among the most efficient test-time defenses to date, establishing the new state-of-the-art. We also demonstrated consistent improvements across standard CIFAR and ImageNet-1k benchmarks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.367, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Interpretability-Guided Test-Time Adversarial Defense" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.314 + ], + "angle": 0, + "content": "This work is supported in part by National Science Foundation (NSF) awards CNS-1730158, ACI-1540112, ACI1541349, OAC-1826967, OAC-2112167, CNS-2100237, CNS-2120019, the University of California Office of the President, and the University of California San Diego's California Institute for Telecommunications and Information Technology/Qualcomm Institute. Thanks to CENIC for the 100Gbps networks. The authors thank the anonymous reviewers for valuable feedback on the manuscript. T.-W. Weng is supported by National Science Foundation under Grant No. 2107189 and 2313105. T.-W. Weng also thanks the Hellman Fellowship for providing research support." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.337, + 0.323, + 0.352 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.368, + 0.785, + 0.396 + ], + "angle": 0, + "content": "1. Addepalli, S., Jain, S., Babu, R.V.: Efficient and effective augmentation strategy for adversarial training. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.397, + 0.785, + 0.424 + ], + "angle": 0, + "content": "2. Addepalli, S., Jain, S., Sriramananan, G., Venkatesh Babu, R.: Scaling adversarial training to large perturbation bounds. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.425, + 0.785, + 0.451 + ], + "angle": 0, + "content": "3. Alfarra, M., Pérez, J.C., Thabet, A., Bibi, A., Torr, P.H.S., Ghanem, B.: Combating adversaries with anti-adversaries. In: AAAI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.452, + 0.787, + 0.479 + ], + "angle": 0, + "content": "4. Andriushchenko, M., Croce, F., Flammarion, N., Hein, M.: Square attack: a query-efficient black-box adversarial attack via random search. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.48, + 0.785, + 0.507 + ], + "angle": 0, + "content": "5. Athalye, A., Carlini, N., Wagner, D.: Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In: ICML (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.508, + 0.785, + 0.548 + ], + "angle": 0, + "content": "6. Bai, N., Iyer, R.A., Oikarinen, T., Weng, T.W.: Describe-and-dissect: Interpreting neurons in vision networks with language models. arXiv preprint arXiv:2403.13771 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.549, + 0.785, + 0.577 + ], + "angle": 0, + "content": "7. Bai, Y., Zeng, Y., Jiang, Y., Xia, S.T., Ma, X., Wang, Y.: Improving adversarial robustness via channel-wise activation suppressing. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.785, + 0.604 + ], + "angle": 0, + "content": "8. Bau, D., Zhou, B., Khosla, A., Oliva, A., Torralba, A.: Network dissection: Quantifying interpretability of deep visual representations. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.605, + 0.785, + 0.646 + ], + "angle": 0, + "content": "9. Bau, D., Zhu, J.Y., Strobelt, H., Lapedriza, A., Zhou, B., Torralba, A.: Understanding the role of individual units in a deep neural network. Proceedings of the National Academy of Sciences (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.687 + ], + "angle": 0, + "content": "10. Boopathy, A., Liu, S., Zhang, G., Liu, C., Chen, P.Y., Chang, S., Daniel, L.: Proper network interpretability helps adversarial robustness in classification. In: ICML (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.785, + 0.715 + ], + "angle": 0, + "content": "1.1. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: IEEE Symposium on Security and Privacy (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.785, + 0.743 + ], + "angle": 0, + "content": "2. Chen, J., Gu, Q.: RayS: A ray searching method for hard-label adversarial attack In: KDD (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "3. Chen, Y., Oliver, D.S.: Ensemble randomized maximum likelihood method as an iterative ensemble smoother. Mathematical Geosciences 44, 1-26 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.798 + ], + "angle": 0, + "content": "4. Chen, Z., Li, Q., Zhang, Z.: Towards robust neural networks via close-loop control In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Croce, F., Andriushchenko, M., Sehwag, V., Debenedetti, E., Flammarion, N., Chiang, M., Mittal, P., Hein, M.: Robustbench: a standardized adversarial robustness benchmark. In: Datasets and Benchmarks Track, NeurIPS (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.368, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.469, + 0.129 + ], + "angle": 0, + "content": "A. Kulkarni and T. W. Weng" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Croce, F., Gowal, S., Brunner, T., Shelhamer, E., Hein, M., Cemgil, T.: Evaluating the adversarial robustness of adaptive test-time defenses. In: ICML (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "17. Croce, F., Hein, M.: Minimally distorted adversarial examples with a fast adaptive boundary attack. In: ICML (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "18. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: ICML (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "19. Eigen, H., Sadovnik, A.: TopKConv: Increased adversarial robustness through deeper interpretability. In: ICMLA (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "20. Gerasimou, S., Eniser, H.F., Sen, A., Cakan, A.: Importance-driven deep learning system testing. In: ACM/IEEE ICSE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "21. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: ICLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "22. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "23. Hernandez, E., Schwettmann, S., Bau, D., Bagashvili, T., Torralba, A., Andreas, J.: Natural language descriptions of deep visual features. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "24. Hwang, D., Lee, E., Rhee, W.: AID-Purifier: A light auxiliary network for boosting adversarial defense. In: ICPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "25. Jia, X., Zhang, Y., Wu, B., Ma, K., Wang, J., Cao, X.: LAS-AT: adversarial training with learnable attack strategy. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "26. Kang, Q., Song, Y., Ding, Q., Tay, W.P.: Stable neural ODE with Lyapunov-stable equilibrium points for defending against adversarial attacks. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "27. Krizhevsky, A.: Learning multiple layers of features from tiny images. Tech. rep., CIFAR (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "28. Kundu, S., Nazemi, M., Beerel, P.A., Pedram, M.: A tunable robust pruning framework through dynamic network rewiring of DNNs. In: ASP-DAC (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "29. Mangla, P., Singh, V., Balasubramanian, V.N.: On saliency maps and adversarial robustness. In: ECML-PKDD (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "30. Mao, C., Chiquier, M., Wang, H., Yang, J., Vondrick, C.: Adversarial attacks are reversible with natural supervision. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.774, + 0.577 + ], + "angle": 0, + "content": "31. Mu, J., Andreas, J.: Compositional explanations of neurons. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.604 + ], + "angle": 0, + "content": "32. Nayak, G.K., Rawal, R., Chakraborty, A.: DAD: Data-free adversarial defense at test time. In: WACV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.632 + ], + "angle": 0, + "content": "33. Oikarinen, T., Weng, T.W.: CLIP-Dissect: Automatic description of neuron representations in deep vision networks. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.66 + ], + "angle": 0, + "content": "34. Oikarinen, T., Weng, T.W.: Linear explanations for individual neurons. In: ICML (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "35. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.757 + ], + "angle": 0, + "content": "36. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) 115(3), 211-252 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.784 + ], + "angle": 0, + "content": "37. Salman, H., Ilyas, A., Engstrom, L., Kapoor, A., Madry, A.: Do adversarially robust imagenet models transfer better? In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "38. Sehwag, V., Wang, S., Mittal, P., Jana, S.: Hydra: Pruning adversarially robust neural networks. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "39. Shi, C., Holtz, C., Mishne, G.: Online adversarial purification based on self-supervised learning. In: ICLR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.368, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Interpretability-Guided Test-Time Adversarial Defense" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "40. Sriramanan, G., Addepalli, S., Baburaj, A., Babu, R.V.: Guided adversarial attack for evaluating and enhancing adversarial defenses. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "41. Sriramanan, G., Addepalli, S., Baburaj, A., Babu, R.V.: Towards efficient and effective adversarial training. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "42. Wong, E., Rice, L., Kolter, J.Z.: Fast is better than free: Revisiting adversarial training. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.259 + ], + "angle": 0, + "content": "43. Wu, B., Pan, H., Shen, L., Gu, J., Zhao, S., Li, Z., Cai, D., He, X., Liu, W.: Attacking adversarial attacks as a defense. arXiv preprint arXiv:2106.04938 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.285 + ], + "angle": 0, + "content": "44. Wu, D., Xia, S.T., Wang, Y.: Adversarial weight perturbation helps robust generalization. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "45. Wu, S., Sang, J., Xu, K., Zhang, J., Yu, J.: Attention, please! adversarial defense via activation rectification and preservation. ACM Transactions on Multimedia Computing, Communications and Applications 19(4), 1-18 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "46. Wu, Y.H., Yuan, C.H., Wu, S.H.: Adversarial robustness via runtime masking and cleansing. In: ICML (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.383 + ], + "angle": 0, + "content": "47. Xiao, C., Zhong, P., Zheng, C.: Enhancing adversarial defense by k-winners-take-all In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "48. Xie, X., Li, T., Wang, J., Ma, L., Guo, Q., Yuefei-Xu, F., Liu, Y.: NPC: Neuron path coverage via characterizing decision logic of deep neural networks. ACM Transactions on Software Engineering and Methodology 31(3) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "49. Yoon, J., Hwang, S.J., Lee, J.: Adversarial purification with score-based generative models. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.733, + 0.466 + ], + "angle": 0, + "content": "50. Zagoruyko, S., Komodakis, N.: Wide residual networks. In: BMVC (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.493 + ], + "angle": 0, + "content": "51. Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: ECCV (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.522 + ], + "angle": 0, + "content": "52. Zhang, H., Yu, Y., Jiao, J., Xing, E.P., Ghaoui, L.E., Jordan, M.I.: Theoretically principled trade-off between robustness and accuracy. In: ICML (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.536 + ], + "angle": 0, + "content": "53. Zhao, Q., Wressnegger, C.: Holistic adversarially robust pruning. In: ICLR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.536 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Interpretability-Guided Test-Time Adversarial Defense/d8bed3ff-2ae8-4bfb-abf2-7d81aee96721_origin.pdf b/2024/Interpretability-Guided Test-Time Adversarial Defense/d8bed3ff-2ae8-4bfb-abf2-7d81aee96721_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2e296ea56a2cd77e1990e6395dc118f4feca4cbf --- /dev/null +++ b/2024/Interpretability-Guided Test-Time Adversarial Defense/d8bed3ff-2ae8-4bfb-abf2-7d81aee96721_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4cf706a9659c74829cb84b2572f382e6d5e2745648d4e29a8c4d64c9ef52cc62 +size 2050453 diff --git a/2024/Interpretability-Guided Test-Time Adversarial Defense/full.md b/2024/Interpretability-Guided Test-Time Adversarial Defense/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0e778ba000513da2c57903e0e8e11027f32acec2 --- /dev/null +++ b/2024/Interpretability-Guided Test-Time Adversarial Defense/full.md @@ -0,0 +1,290 @@ +# Interpretability-Guided Test-Time Adversarial Defense + +Akshay Kulkarni $^{1}$ and Tsui-Wei Weng $^{2}$ + +1 CSE, UC San Diego + +$^{2}$ HDSI, UC San Diego + +{a2kulkarni,lweng}@ucsd.edu + +Abstract. We propose a novel and low-cost test-time adversarial defense by devising interpretability-guided neuron importance ranking methods to identify neurons important to the output classes. Our method is a training-free approach that can significantly improve the robustness-accuracy tradeoff while incurring minimal computational overhead. While being among the most efficient test-time defenses (4× faster), our method is also robust to a wide range of black-box, white-box, and adaptive attacks that break previous test-time defenses. We demonstrate the efficacy of our method for CIFAR10, CIFAR100, and ImageNet-1k on the standard RobustBenchmark (with average gains of 2.6%, 4.9%, and 2.8% respectively). We also show improvements (average 1.5%) over the state-of-the-art test-time defenses even under strong adaptive attacks. + +# 1 Introduction + +Despite the popularity and success of deep neural networks (DNNs), they have been shown to be brittle against carefully designed small perturbations that are imperceptible to humans [21]. These "adversarial" perturbations pose significant risks against deploying DNNs for safety-critical tasks like autonomous driving. To mitigate these risks, there has been great interest in developing defenses, e.g. adversarial training algorithms [40, 42, 44, 52] to obtain a model that is more robust to adversarial samples. + +However, adversarial training is much more expensive than standard training (e.g. a strong defense like TRADES [52] takes $150 - 200 \times$ longer than standard training). This inspired training-free defenses known as test-time defenses [3, 26, 30, 39]. The underlying principles in existing test-time defenses are either input purification or model adaptation: input purification [3, 39] modifies the input with an additional perturbation to nullify any adversarial perturbation, whereas model adaptation [14, 26] updates model parameters or adds new parameters at test-time to make the overall model more robust. Despite eliminating the need for training, most existing methods are still quite expensive, $8 - 500 \times$ more than + +![](images/91dac736033fb7175228124be14c872e907811122a13ec7ccae3e405b11e37cf.jpg) +Fig. 1: Interpretability-guided masking based on neuron importance ranking for test-time adversarial defense. + +a standard forward pass [16]. Moreover, existing test-time defenses have been broken under strong adaptive attacks [16]. Motivated by these limitations, in this paper, we intend to find an efficient and effective alternative to improve robustness at test-time, so that the defense does not break under adaptive attacks. + +To design an effective defense, we leverage recent neuron-level interpretability methods [8, 33], which allow one to interpret the functionalities of individual neurons. In our analysis, we observed that for successful adversarial examples, the majority of activations at an intermediate layer shift from important neurons for the ground-truth class to the important neurons for other classes. Hence, we propose a novel test-time Interpretability-Guided Defense (IG-Defense) enabled by "neuron importance ranking" where class-wise importance of each neuron is computed (Fig. 1A). To design effective neuron importance ranking, we repurpose our analysis experiments and existing neuron-level interpretability works to propose two novel neuron importance ranking methods. The key idea of our IG-Defense is to restrict the observed activation shift by masking unimportant neurons (Fig. 1B). We demonstrate the usefulness of our insights by extensively evaluating our IG-Defense on the standard RobustBench benchmark (Fig. 1C). We observe consistent gains for CIFAR10 (+2.6%), CIFAR100 (+4.9%), and ImageNet-1k (+2.8%). We also devise an ensemble of strong adaptive attacks specifically targeting the weaknesses of IG-Defense and observe consistent gains (+1.5%), unlike existing test-time defenses that break under adaptive attacks. Due to our simple yet grounded approach, IG-Defense is among the most efficient test-time defenses (4× faster than the strongest existing defense). + +Our primary contributions can be summarized as: + +- We are the first to propose a neuron-interpretability-guided test-time defense (IG-Defense) utilizing neuron importance ranking (CD-IR, LO-IR) to improve adversarial robustness. IG-Defense is training-free, efficient, and effective. + +- We uncover novel insights into improving adversarial robustness by analyzing adversarial attacks through the lens of neuron-level interpretability in Sec. 3. + +- Our proposed IG-Defense consistently improves the robustness on standard CIFAR10, CIFAR100, and ImageNet-1k benchmarks. We also demonstrate improved robustness upto $3.4\%$ , $3.8\%$ , and $1.5\%$ against a wide range of white-box, black-box, and adaptive attacks respectively with the lowest inference time ( $4\times$ faster) among existing test-time defenses. + +# 2 Related Work + +Test-time adversarial robustness. Most of the test-time defenses can be categorized into two approaches: input purification or model adaptation. In input purification, the adversarial input is converted to a safer input by an additional perturbation. For instance, finding a perturbation that either increases top class confidence [3], minimizes contrastive loss [30], or minimizes a self-supervised objective [39]. [24] crafts the perturbation by fooling a real-adversarial discriminator [24] and [49] uses an input passed through a score-based generative model. For model adaptation [14,32,46], the model parameters are either updated at test-time or new parameters are introduced. However, existing test-time defenses are quite expensive (8-500× of single forward pass) and break under adaptive attacks. In contrast, our method is fast and effective (1.5% gain under adaptive attacks with only 2× inference time), and is the first test-time defense to leverage neuron-level interpretability information. + +Neuron interpretability methods. To understand the functionalities (i.e. concepts) of individual neurons in DNNs, early works like NetDissect [8,9] compare the neuron activation patterns with patterns of predefined concept label masks. MILAN [23] produced natural language descriptions using a supervised image captioning model trained on a large-scale concept-labeled dataset. Recently, CLIP-Dissect [33] leveraged the CLIP paradigm [35] to identify concepts of neurons more efficiently than prior work and without the need of training or densely-concept-labeled datasets. There are other works to identify neuron concepts through clustering [20, 48], via language models [6], or using linear combinations [34]. In this work, we focus on repurposing CLIP-Dissect due to its efficiency and simplicity as well as NetDissect to propose neuron importance ranking methods. + +Connections between interpretability and robustness. Prior works attempt to improve robustness via input-dependent interpretability methods like saliency or class activation maps [10, 29]. Wu et al. [45] observed that adversarial attacks lead to changes in attention maps, and proposed a training-time defense with losses to enforce similar attention maps for clean and adversarial samples. Several works [7, 19, 28, 38, 47, 53] use sparsity to improve adversarial robustness. However, they require retraining after or while sparsity constraints are applied. In contrast, we indirectly leverage sparsity without requiring any training. NetDissect [8] investigated the connection between adversarial attacks and interpretability information from their method, but they did not explore how to improve robustness. Hence, we aim to study the unexplored connection of neuron-level interpretability and improving adversarial robustness. + +# 3 Analysis Experiments + +To understand the connection of interpretability of individual neurons with robustness, we aim to design an analysis experiment. One way would be to investigate the changes in neuron activations before and after an adversarial attack. In neuron-labeling works [9,31], they assign an explainable concept to each neuron (or an activation channel in case of CNNs). Specifically, NetDissect [9] qualitatively show that a targeted adversarial attack causes a drop in peak activations for neurons whose concepts are related to the original class, while causing an increase in peak activations for neurons whose concepts are related to the target class of the attack. They also show that the peak activation change is lower for neurons with unrelated concepts. While they only perform the analysis to showcase their interpretability, we are the first to analyze adversarial attacks through neuron-level interpretability with the explicit goal of obtaining insights into improving adversarial robustness. + +For a more thorough analysis, we observe the average activation change (instead of peak change) and associate the neurons with the category names directly instead of concepts. First, peak activation change may be misleading since it may be caused by specific images while average change takes all images into account for a more well-rounded analysis. Second, our automated evaluation eliminates the need for manual association of concepts with categories [9], which could be difficult, expensive, and noisy, especially with a large number of related categories (e.g. ImageNet classes). We analyze both successful and unsuccessful attacks while [9] only focused on successful attacks. + +Concretely, we find top- $k$ important neurons related to each task category using neuron importance ranking (discussed in Sec. 4.2). With this, in Fig. 2, we analyze the change in activations corresponding to ground-truth (GT) and non-GT categories before and after an attack. We discuss some preliminaries before delving into further details. + +Preliminaries. We denote the dataset as $\mathcal{D} = \{(x,y)\}$ where $x\in \mathcal{X}$ is the input image, $y\in \mathcal{C}$ is the label, $\mathcal{C}$ is the label set, and $|\mathcal{C}|$ is the number of classes. Given a pretrained model $f:\mathcal{X}\to \mathcal{C}$ and its activations at the layer being studied $A:\mathcal{X}\rightarrow \mathbb{R}^{N\times H\times W}$ , we can rank the importance of a neuron (out of $N$ neurons/channels) by assessing its contribution to each class. We call this "neuron importance ranking" and propose two algorithms for this (Sec. 4.2). Then, we can compute the activations of the top- $k$ important neurons for any class $c_{i}$ as $A_{c_i}(x) = A(x)_{[n_{c_i}]} \in \mathbb{R}^{k\times H\times W}$ , where $n_{c_i}$ are the indices of the top- $k$ activations for class $c_{i}$ . With access to each $n_{c_i}$ , the remaining indices absent from the top- $k$ of any class are $n_u = [N] \setminus \{\cup_{c_i\in \mathcal{C}} n_{c_i}\}$ , where $[N] = \{1,2,\dots,N\}$ and $\backslash$ indicates setminus operator. These indices $n_u$ denote unimportant neurons, and corresponding unimportant activations can be computed as $A_u(x) = A(x)_{[n_u]} \in \mathbb{R}^{(N - k|\mathcal{C}|) \times H \times W}$ . Next, we define activation change metrics to support our analysis. + +Metrics. For an adversarial attack with $\ell_{\infty}$ -norm bound of $\epsilon$ , the perturbation for input $x$ is $\delta^{*} = \arg \max_{\| \delta \|_{\infty} \leq \epsilon} \ell(f(x + \delta), y)$ where $\ell$ is the attack objective. Let $\hat{y}^{*} = f(x + \delta^{*})$ be the post-attack prediction. The average activation change + +![](images/44563d80def7e9948ace4e50d33af7c4f22143b790342aa4a792eb2d9d1d0ac9.jpg) +Fig. 2: A. Cat-dog-bird image classifier before a PGD attack. B. After PGD attack, the prediction changes from the ground truth (GT) cat to dog since the activations of neurons important to dog class increase while those important to cat class decrease. + +C. Empirically, successful PGD attacks show a decrease in activations of important GT class neurons while those of post-attack predicted class' important neurons increase. + +D. For unsuccessful PGD attacks, the activations of all neurons reduce even though the prediction remains the same as before the attack. + +in the important neurons w.r.t. any class $c_{i}$ is, + +$$ +\Delta_ {c _ {i}} = \underset {(x, y) \in \mathcal {D}} {\mathbb {E}} \left[ A _ {c _ {i}} \left(x + \delta^ {*}\right) - A _ {c _ {i}} (x) \right] \tag {1} +$$ + +Replacing $c_{i}$ with GT class $y$ , the average activation change in the important neurons of GT class is $\varDelta_{y}$ . Similarly, the average activation change in post-attack predicted class neurons is $\varDelta_{\hat{y}^*}$ , and that in other non-GT class neurons is $\varDelta_{\mathrm{non - GT}} = \mathbb{E}_{c_i\neq y,\hat{y}^*}[\varDelta_{c_i}]$ . For unsuccessful attacks, where post-attack prediction $\hat{y}^{*}$ is the same as $y$ , we can compute average activation change in the remaining classes (other than GT) as $\varDelta_{\mathrm{rem -cls}}=\mathbb{E}_{c_i\neq y}[\varDelta_{c_i}]$ . Finally, we can also compute the average activation change in unimportant neurons, + +$$ +\Delta_ {\text {u n i m p}} = \underset {(x, y) \in \mathcal {D}} {\mathbb {E}} \left[ A _ {u} \left(x + \delta^ {*}\right) - A _ {u} (x) \right] \tag {2} +$$ + +We empirically compute the above metrics on CIFAR10 with five base models (ResNet18: GAT [40], NuAT [41], OAAT [2], DAJAT [1], and WideResNet-34-10: TRADES-AWP [44]) and present the mean values in Fig. 2C and Fig. 2D. See Appendix for more experimental details and corresponding numerical results. + +Observations. In Fig. 2C, we observe $\varDelta_y<0$ and $\varDelta_{\hat{y}^*}>0$ , i.e. successful adversarial attacks boost the activations that are important to the post-attack predicted class $\hat{y}^*$ while causing a drop for those of the GT class $y$ . Further, from Fig. 2C, the important activations for the remaining classes and unimportant activations are only marginally affected by successful attacks. In Fig. 2D, we observe that unsuccessful attacks cause a drop in the activations of all neurons. + +Remarks. During both successful and unsuccessful attacks, the activation magnitude of GT class' important neurons drops. Specifically, successful attacks shift the activation magnitude from GT class' important neurons to non-GT class important neurons. Hence, we hypothesize that adversarial robustness can be improved if the activation shift to non-GT class important neurons is restricted. + +Further, given the availability of a huge number of adversarially pretrained models [15], we aim to evaluate our hypothesis in a test-time adversarial defense setting, with a low inference overhead and without any training overhead. + +# 4 Interpretability-Guided Defense + +In this section, we propose a novel, training-free, test-time Interpretability-Guided Defense (IG-Defense). The core idea of our method is based on interpretability-guided masking, described in Sec. 4.1. In Sec. 4.2, we introduce neuron importance ranking, which is a key enabler in the success of our proposed defense. + +Setup. Given a trained model $f: \mathcal{X} \to \mathcal{C}$ , the goal of test-time defense is to improve the robustness of $f$ while retaining its utility (i.e. clean accuracy). We refer to $f$ as the "base model" in the rest of the paper. + +# 4.1 Interpretability-guided masking + +To validate our hypothesis, we design a simple approach for test-time defense where we mask all the neurons except the important neurons of the correct class. We propose a dual forward pass approach, since access to labels cannot be assumed at test-time. Our approach contains three key steps: + +- Step 1: Neuron Importance Ranking. First, we compute class-wise neuron importance ranking (Step 1 of Fig. 3). This is computed only once for a given base model $f$ . For each layer, we rank every neuron based on its importance to each class. More details are given in Sec. 4.2. + +- Given the number of important neurons $k$ (hyperparameter), we compute a binary mask $m \in \{0,1\}^{N \times C}$ using the indices of important neurons $n_{c_i} \in \mathbb{N}^k$ (from the importance ranking in Sec. 4.2) for each class $c_i \in [C]$ . + +$$ +m \left[ n _ {c _ {i}}, c _ {i} \right] = 1 \forall c _ {i} \in [ C ], \quad m \left[ \left[ N \right] \backslash n _ {c _ {i}}, c _ {i} \right] = 0 \forall c _ {i} \in [ C ] \tag {3} +$$ + +- In the mask $m$ , the top- $k$ important neurons of each class $c_i \in [C]$ are retained in $m[:, c_i]$ . The remaining neurons unimportant to $c_i$ are masked. + +- Step 2: Vanilla Forward Pass. In the first forward pass (Step 2 of Fig. 3), we obtain a soft-pseudo-label $\hat{y} = \sigma\left(\frac{f(x)}{\tau}\right) \in (0,1)^C$ where $\sigma(\cdot)$ denotes the softmax function, and $\tau$ denotes the temperature term which controls the sharpness of the soft-pseudo-label. + +- Step 3: Masked Forward Pass. In the second forward pass (Step 3 of Fig. 3), we apply a soft-pseudo-label weighted mask i.e. $m\hat{y} \in (0,1)^N$ to the activations of the layer being masked. Let $A:\mathcal{X}\to \mathbb{R}^{N\times H\times W}$ give + +![](images/36c4a5e322c35445267218808a94d2a4326a72d99e46df3c2dbd9f1459f16acd.jpg) +Fig. 3: Step 1. Given a pretrained base model $f$ (e.g. binary classifier here), class-wise neuron importance is computed for a selected layer (Sec. 4.2). A top- $k$ mask $m \in \{0,1\}^{N \times C}$ is computed to identify top- $k$ neurons important to each class (e.g. $k = 2, N = 5, C = 2$ here). Step 2. During evaluation, a soft-pseudo-label $\hat{y}$ is computed using the base model $f$ . Step 3. The soft-pseudo-label weighted mask $m\hat{y} = m\sigma(\frac{f(x)}{\tau}) \in \mathbb{R}^N$ is applied to the selected layer to retain only the important neurons of the pseudo-label class. + +the activations at the layer being masked, and $h: \mathbb{R}^{N \times H \times W} \to \mathcal{C}$ be the remaining layers after the layer being masked, i.e. $f = h \circ A$ . Then, the final prediction is given by $\hat{y}' = h(m\hat{y} \cdot A(x))$ . + +- The masking re-weights the neurons based on $\hat{y}$ . However, the temperature term $\tau$ is set to a small value so that $\hat{y}$ contains a sharp probability distribution. Thus, the weighted mask corresponds mainly to a single class and retains only the important neurons of that class. +- Also note that this enables gradient flow between the two forward passes, to avoid any gradient masking [5]. + +Intuitively, activations of a deeper layer (e.g. penultimate layer) are masked to minimize any negative impact of the masking on feature extraction, which depends more on the shallower layers [51]. To verify this, we analyze the sensitivity of our method to the layer being masked in Sec. 5.3b. + +Dependence on soft-pseudo-label. This is an obvious limitation if the soft-pseudo-label is computed directly from the base model $f$ . If it is incorrect, masking will yield no improvement in robustness. Hence, an adaptive attack can be designed to ensure that the soft-pseudo-label is incorrect. For instance, a transfer attack that transfers the perturbation from the base model $f$ would yield an incorrect soft-pseudo-label, and the final robustness would be the same as the base model $f$ . To circumvent this, we propose to use randomized smoothing in the first forward pass (i.e. in Step 2 of Fig. 3). + +Randomized smoothing. To obtain the soft-pseudo-label $\hat{y}$ in a more robust manner, we perform randomized smoothing [13] in the first forward pass. We add a set of $n_s$ Gaussian noises (with zero mean and standard deviation $\sigma_d$ ) to each input $x$ . For the soft-pseudo-label, the logits of these noisy inputs are averaged and softmax $\sigma$ is applied. Formally, + +$$ +\hat {y} = \sigma \left(\frac {1}{\tau} \underset {v _ {i}} {\mathbb {E}} [ f (x + v _ {i}) ]\right); v _ {i} \sim \mathcal {N} \left(0 _ {d}, \sigma_ {d} ^ {2} I _ {d}\right) \forall i \in [ n _ {s} ] \tag {4} +$$ + +![](images/517f549af567e09635029a08a80529426a4100b6f1a1ac8f0401ba853b427c4e.jpg) +Fig. 4: A. Leave-one-Out Importance Ranking (LO-IR) computes class-wise neuron importance as the average change in logits when masking each neuron. B. CLIP-Dissect Importance Ranking (CD-IR) first computes the inner product of class name text embeddings and probing set image embeddings. The importance ranking relies on the similarity between activations of each neuron to the precomputed inner product given the same probing set inputs. C. A higher logit-change $\Delta_{c_i}^{[j]}$ or similarity $s_{c_i}^{[j]}$ implies that neuron $j$ is more important to class $c_i$ . + +where $0_{d}, I_{d}$ are tensors with the same shape as $x$ but containing all zeros and ones respectively. The number of noises sampled $n_{s}$ and standard deviation $\sigma_{d}$ are hyperparameters. + +# 4.2 Importance ranking methods + +Here, we describe two novel importance ranking methods that can be used in Step 1 of our proposed IG-Defense. + +a) Leave-one-Out Importance Ranking (L0-IR). Following the analysis in Sec. 3, given a neuron $j$ , we can compute its importance to a class $c_{i}$ as the average change in the class $c_{i}$ logits, $\Delta_{c_i}^{[j]}$ , before and after masking out that particular neuron $j$ (Fig. 4A). Intuitively, a higher logit-change for a particular class implies higher dependence of the network on that neuron, i.e. higher importance (Fig. 4C). Formally, + +$$ +\Delta_ {c _ {i}} ^ {[ j ]} = \underset {(x, y) \in \mathcal {D} _ {\text {p r o b e}} | y = c _ {i}} {\mathbb {E}} \left[ f _ {c _ {i}} (x) - f _ {c _ {i}} ^ {[ j ]} (x) \right]; \forall c _ {i} \in \mathcal {C} \tag {5} +$$ + +where $f^{[j]}$ is the model where $j^{\mathrm{th}}$ neuron of the selected layer is masked in the base model $f$ . And the subscript $c_{i}$ indicates the logits of class $c_{i}$ . However, this method could be computationally expensive as average change in logits has to be computed for each neuron. For our proposed method, it is only required once per base model. Thus, it is a feasible option that yields a very good estimate of the importance ranking, which we show in Sec. 5. + +b) CLIP-Dissect Importance Ranking (CD-IR). CLIP-Dissect [33] uses the multimodal CLIP model [35] to assign concept labels to individual neurons. We extend this idea by replacing concepts with task labels to obtain an importance ranking of neurons for each task label (see Fig. 4B). + +Concretely, probing images are passed through the image encoder $E_{I}$ to get a set of image embeddings $\{A^{[i]} = E_I(x_i)\}_{i = 1}^M$ (shown in orange in Fig. 4B) while the task label names $\mathcal{C}$ are passed through the text encoder $E_{T}$ to get the corresponding set of text embeddings $\{E_T(c_k)\}_{c_k\in \mathcal{C}}$ (shown in green in Fig. 4B). These embeddings i.e. vectors are converted to scalars by taking the inner product of each pair of image-text embeddings, which results in a matrix $P$ with entries $P_{ik} = A^{[i]^\top}E_T(c_k)$ . Now, for a neuron $j$ , a vector of its activations $q_{j}$ (shown in purple in Fig. 4B) can be obtained where each entry corresponds to each input from the probing dataset. We can compute the similarity score $s_{c_k}^{[j]}$ between each class $c_{k}$ and neuron $j$ using the $k^{\mathrm{th}}$ column of $P$ and $q_{j}$ . Intuitively, a higher similarity score $s_{c_k}^{[j]}$ implies higher importance of neuron $j$ for class $c_{k}$ (Fig. 4C). We use the same soft-WPMI similarity metric as the original paper [33]. For the probing dataset, we simply use the training dataset for our experiments. + +In practice, CD-IR is computationally less expensive compared to L0-IR (e.g. 2 mins. vs. 28 mins. for 512 neurons). This is because CLIP embeddings can be precomputed and most of the computations can be vectorized easily. As we demonstrate in Sec. 5, although the importance ranking quality is slightly worse than L0-IR (i.e. CD-IR causes a marginal drop in clean accuracy), downstream robustness gains with CD-IR are still significant. Please refer to the Appendix C.3 for details on computational complexity comparisons. + +# 5 Experiments + +We extensively evaluate our IG-Defense on the standard RobustBench benchmark [15]. We also follow Croce et al. [16] for properly evaluating test-time defenses and thoroughly test against an ensemble of strong adaptive attacks. + +Experimental setup. We evaluate our methods on standard datasets, CIFAR10, CIFAR100 [27] and ImageNet-1k [36], with ResNet [22] and WideResNet [50] base models. We use an $\ell_{\infty}$ threat model with an $\epsilon$ -bound of 8/255 for CIFAR10/100 and 4/255 for ImageNet-1k following prior works [15]. For evaluation, we use the full CIFAR10/100 test set (10000 samples) and the first 5000 ImageNet-1k validation samples, which is the standard setup as per [15,37]. The hyperparameter $k$ for the top- $k$ neurons in CD-IR or L0-IR is mentioned alongside the method name in the results. For randomized smoothing (RS), we use $n_s = 1$ , $\sigma_d = \frac{\epsilon}{2}$ , and $\tau = 0.01$ , where $\epsilon$ is the $\ell_{\infty}$ -bound for the attack. We show in Appendix C.2 that the choices of $n_s = 1$ , $\sigma_d = \frac{\epsilon}{2}$ are sufficient for our purpose. Unless otherwise mentioned, RS is used and the penultimate layer output is masked. In the penultimate layer, there are 512 neurons (layer4) in ResNet18, 640 neurons (layer3 or block3) in WideResNet-34-10, and 2048 neurons (layer4) in ResNet50. + +Choice of base models. We experiment with base models from several single-step defenses like FAT [42], GAT [40], NuAT [41], OAAT [2], and multi-step + +Table 1: Evaluating our IG-Defense (CD-IR, LO-IR) on CIFAR10 with ResNet18 and $\ell_{\infty}, \epsilon = 8/255$ . IW-WC indicates image-wise worst-case, AA indicates AutoAttack. + +
Base Model:NuAT [41] (RN18)OAAT [2] (RN18)DAJAT [1] (RN18)
Test-Time DefenseClean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)
None82.2150.5850.5480.2351.1051.0185.7152.5052.45
CD-IR-50 (Ours)82.4652.0851.58 (+1.04)79.8554.7952.35 (+1.34)85.1154.8153.36 (+0.91)
L0-IR-50 (Ours)82.1752.4152.19 (+1.65)79.9054.9052.30 (+1.29)85.1854.5353.34 (+0.89)
+ +Table 2: Evaluating our IG-Defense (CD-IR, L0-IR) on CIFAR10 with WRN-34-10 and $\ell_{\infty}, \epsilon = 8/255$ . IW-WC indicates image-wise worst-case, AA indicates AutoAttack. + +
Base Model:NuAT [41] (WRN-34-10)TR-AWP [52] (WRN-34-10)DAJAT [1] (WRN-34-10)
Test-Time DefenseClean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)
None86.3254.7554.7285.3656.1756.1288.7157.8157.72
CD-IR-50 (Ours)86.2857.6956.38 (+1.66)84.9860.2657.46 (+1.34)88.5160.7059.22 (+1.50)
LO-IR-50 (Ours)86.2157.0456.40 (+1.68)85.1759.6157.88 (+1.76)88.6160.0859.41 (+1.69)
+ +defenses like TRADES [52], AWP [44], DAJAT [1], RTB [37], and LAS [25]. We chose these models to ensure a diverse evaluation and showcase the generality of our IG-Defense across base model defense strategies. We also evaluate with the same base model on different architectures as well as on different datasets for a thorough analysis. Please see Appendix B.2 for more details on these defenses. Evaluation. We evaluate our IG-Defense with the base models against a diverse range of attacks. AutoAttack [18] (AA) is an empirically strong attack with a sequence of white-box AutoPGD (APGD) attacks, fast adaptive boundary (FAB) attack [17], and black-box Square attack [4]. For a stronger adaptive attack, we devise an ensemble of attacks. For each image, the adaptive attack is considered successful if any one attack from the ensemble is successful. With this, we get the image-wise worst-case (IW-WC) robust accuracy. To be adaptive for our IG-Defense, the ensemble also includes transfer attacks that specifically target our weakness, i.e. our dependence on pseudo-labels (as discussed in Sec. 4.1). + +Robust accuracy (IW-WC) is computed with 9 attacks. The first three are computed on the defended model: standard white-box AutoAttack, decision-boundary based black-box RayS attack [12], and white-box APGD with Expectation over Transformation (EoT) [5] attack. For the remaining 6 transfer attacks, we compute perturbations on the base model using 6 attacks: APGD with cross-entropy loss, APGD with Carlini-Wagner loss [11], targeted APGD attack with difference of logits ratio loss, standard AutoAttack, RayS attack, and $\mathrm{APGD + EoT}$ attack. Design choices for IW-WC. While AutoAttack also uses APGD attacks, it returns unperturbed inputs when the attack is unsuccessful. Following [16], the four transfer APGD attacks in IW-WC each return a sample with the highest loss (i.e. closest to the decision boundary). This is a stronger attack than AutoAttack for test-time defenses since samples closer to the decision boundary tend to be misclassified under test-time defenses, as observed by [16]. Expectation over Transformation (EoT) [5] averages gradients over multiple iterations of + +Table 3: Evaluating our IG-Defense (CD-IR, L0-IR) on CIFAR100 with $\ell_{\infty}, \epsilon = 8 / 255$ . IW-WC indicates image-wise worst-case, AA indicates AutoAttack. + +
Base Model:LAS [25] (WRN-34-10)OAAT [2] (RN18)DAJAT [1] (RN18)
Test-Time DefenseClean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)
None64.8930.7430.7162.0227.1527.1265.4527.6727.64
CD-IR-50 (Ours)62.1333.5731.41 (+0.70)61.5629.5129.19 (+2.07)64.9730.2228.27 (+0.63)
LO-IR-50 (Ours)64.8535.6932.22 (+1.51)62.0432.0630.12 (+3.00)65.3732.5528.54 (+0.90)
+ +Table 4: Evaluating our IG-Defense (CD-IR, L0-IR) on ImageNet-1k with ResNet50 and $\ell_{\infty}, \epsilon = 4/255$ . IW-WC indicates image-wise worst-case, AA indicates AutoAttack. + +
Base Model:FAT [42] (RN50)RTB [37] (RN50)
Test-Time DefenseClean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)
None55.6426.2426.2464.1034.6234.62
CD-IR-750 (Ours)54.3027.28 (+1.04)26.80 (+0.56)62.5036.42 (+1.80)35.30 (+0.68)
LO-IR-750 (Ours)55.8829.14 (+2.90)27.08 (+0.84)64.1037.36 (+2.74)35.48 (+0.86)
+ +a particular input to counter the effect of randomness in a defense. Specifically, we use 20 iterations of EoT with APGD (more than $n_s = 1$ ). While [16] also recommend using Backward Pass Differentiable Approximation (BPDA) [5] for test-time defenses with non-differentiable components, our approach already includes BPDA since we use a soft-masking approach instead of hard-masking. Refer to Appendix B.3 for complete implementation details. + +# 5.1 Evaluation on RobustBench + +a) CIFAR10. In Table 1, we evaluate our methods on CIFAR10 with ResNet18 at $\epsilon = 8 / 255$ . We observe consistent gains in AutoAttack (avg. $+2.5\%$ for CD-IR and $+2.55\%$ for L0-IR) and IW-WC robustness (avg. $+1.1\%$ for CD-IR and $+1.28\%$ for L0-IR), with a marginal drop in clean accuracy (avg. $-0.24\%$ for CD-IR and $-0.3\%$ for L0-IR). In Table 2, we evaluate on CIFAR10 with the more challenging, large WideResNet-34-10 model at $\epsilon = 8 / 255$ . We observe consistent gains in IW-WC robustness (avg. $+1.5\%$ for CD-IR and $+1.7\%$ for L0-IR), with a marginal drop in clean accuracy (avg. $-0.2\%$ for CD-IR and $-0.13\%$ for L0-IR). b) CIFAR100. In Table 3, we evaluate our methods on CIFAR100 with ResNet18 and WideResNet-34-10 at $\epsilon = 8 / 255$ . Despite the increased number of classes, we get consistently improved worst-case (IW-WC) robust accuracy (avg. $+1.13\%$ for CD-IR and $+1.8\%$ for L0-IR). However, the clean accuracy drops for CD-IR (avg. $-0.43\%$ drop), but only marginally for L0-IR. +c) ImageNet-1k. In Table 4, we evaluate our methods on ImageNet-1k with ResNet50 at $\epsilon = 4 / 255$ . Despite the increased image size and the higher number of classes, we observe consistent gains in worst-case (IW-WC) robustness (avg. $+0.6\%$ for CD-IR and $+0.85\%$ for L0-IR) but with a drop in clean accuracy for CD-IR (avg. $-1.47\%$ similar to CIFAR100). Given the scale of ImageNet-1k, even + +Table 5: Comparison against state-of-the-art test-time defenses with ResNet18 and WideResNet-34-10 on CIFAR10. The number in (green) shows the worst-case robustness gain over the base model. The first row is the base model without any test-time defense. Inference time is a multiple of the time for a single forward pass of the base model. + +
Base Model → Test-Time DefenseOAAT [2] (RN18)DAJAT [1] (RN18)TR-AWP [44] (WRN)Inference Time
Clean Acc.Robust Acc. (IW-WC)Clean Acc.Robust Acc. (IW-WC)Clean Acc.Robust Acc. (IW-WC)
No test-time def.80.2351.0185.7152.4585.3656.121×
Existing Defenses
HD [43]79.8950.68 (-0.33)84.5352.55 (+0.10)84.7855.72 (-0.40)46×
SODEF [26]80.2350.67 (-0.34)84.8652.95 (+0.50)85.2556.05 (-0.07)2×
CAAA [3]80.2351.00 (-0.01)85.7152.45 (+0.00)85.3556.11 (-0.01)8×
Our IG-Defense
CD-IR-5079.8552.35 (+1.34)85.1153.36 (+0.91)84.9857.46 (+1.34)2×
L0-IR-5079.9052.30 (+1.29)85.1853.34 (+0.89)85.1757.88 (+1.76)2×
+ +robustness improvements of $\sim 0.5\%$ are quite significant, especially since there is no clean accuracy drop with LO-IR. + +Why does the clean accuracy drop? There are two possible reasons for the drop in clean accuracy. First, due to the randomized smoothing, some samples may get incorrect predictions if the random noise is significant enough to change the class prediction w.r.t. the base model. Second, importance ranking quality and number of neurons masked determine how much clean accuracy can be retained. That is, better quality or lower number of neurons masked would guarantee better clean accuracy retention. However, simply masking lower number of neurons may also reduce the improvements in robustness. + +We see in Table 3, 4 that clean accuracy drop for CD-IR is more for CIFAR100 and ImageNet1k compared to CIFAR10 (Table 1, 2). This is possibly because the ratio of number of neurons to number of classes is decreasing, i.e. more number of neurons will be important to multiple classes. This cannot be captured well by CD-IR, leading to a larger drop in clean accuracy. Note that LO-IR is less susceptible to this problem as the importance ranking of each neuron is computed independently, but it is more expensive. This is why we need to retain more number of important neurons for ImageNet ( $k = 750$ ) than CIFAR ( $k = 50$ ). + +Note that the clean accuracy drops are acceptable since they are significantly smaller than the worst-case robustness gains (e.g. upto $-0.6\%$ and $+1.7\%$ respectively for CIFAR10 in Table 1-2). Further, in some cases like ImageNet-1k with LO-IR (Table 4), there are no drops with pure gains in robustness. + +# 5.2 Comparisons with existing test-time defenses + +We compare with three recent state-of-the-art test-time defense methods on CIFAR10 in Table 5. CAAA [3] (input purification) optimizes an "anti-adversary" perturbation in a direction opposite to an adversarial attack, by minimizing the loss based on a pseudo-label, to counter the effect of any adversarial perturbation. HD or Hedge Defense [43] (input purification) finds a perturbation that maximizes + +![](images/3ac288f046bda527dedb18a14cc180e693373db2dc20a52eeaca6b5fec58b3f3.jpg) +Fig. 5: Sensitivity analyses for number of retained important neurons $k$ and layer being masked with CIFAR10. For A, B: layer4 with total 512 neurons is used. A lower $k$ and a deeper layer yield better adversarial robustness with minimal loss of clean accuracy. + +![](images/bc7d75d5851e386c8a9880951ca38dcf33b9598a2a7107a8081763c6bf634b2c.jpg) + +![](images/417d66132b5d41f9462b058afbd7f54e7b033632bde7a0260337b9717d4cc383.jpg) + +![](images/74948447b5099fe5443435a8c2c52fd23e911e74e10b22fb195dfee967cba7ae.jpg) + +the cross-entropy loss summed over all classes. SODEF [26] (model adaptation) trains a neural ODE block with Lyapunov stability constraints between the pretrained feature extractor and classifier, which reduces the influence of the adversarial perturbation. In contrast, our method uses neither input purification nor model adaptation, since we only mask the activations. + +In Table 5, we observe that existing test-time defenses like HD, CAAA, and SODEF do not significantly improve, but rather also reduce the worst-case robust accuracy (IW-WC) compared to the base model. Based on Croce et al. [16], CAAA is the strongest test-time defense among the nine defenses that they evaluated, since its worst-case robust accuracy did not drop below that of the base model. Hence, we only compare with the two strongest input purification defenses, CAAA and HD as well as an efficient model adaptation defense, SODEF from the nine defenses in [16]. Overall, we find that our proposed IG-Defense obtains consistent improvements in image-wise worst-case (IW-WC) robust accuracy, unlike existing test-time defenses. Note that we focus mainly on the worst-case (IW-WC) robust accuracy because gains under weaker attacks would not reflect fundamental gains in robustness [16]. + +# 5.3 Analysis + +a) Ablation study. In Table 6, we study the significance of each proposed component of our approach. First, we start with the baseline (row #1) which uses only one forward pass without masking and without randomized smoothing (RS). Next, we add RS to this baseline (row #2) and observe no gains in robustness and a small drop in clean accuracy. This shows that RS alone cannot improve the robustness. Then, we perform two forward passes without masking and without RS (row #3). Here, the robust accuracy is the same as the baseline indicating that there is no gradient masking with two forward passes. Next, we perform random masking with two forward passes, without and with RS (row #4, #5). Each class is assigned a random set of neurons as the "important" neurons to perform the masking. We observe no gains in robustness, validating the significance of our neuron importance ranking. Finally, we show the impact of RS with our CD-IR (row #6, #7) and L0-IR based masking (row #8, #9). Here, we observe consistent gains in robustness, with a small drop in clean accuracy since RS likely changed the prediction w.r.t. the base model for a few samples. + +b) Sensitivity analyses. We study the sensitivity of our proposed methods to the number of retained important neurons $k$ in Fig. 5A, B on CIFAR10. Specifically, we test it on ResNet18 layer4 which has total 512 neurons and we vary $k$ from 10 to 500. We observe that the robustness improvement is higher at lower $k$ , although lower than 10 may reduce the clean accuracy due to a high loss + +Table 6: Ablation study of proposed components with TRADES-AWP [44] (WideResNet-34-10) as the base model. RS indicates use of randomized smoothing in the first forward pass. The first row is the base model performance without any test-time defense. + +
Row #No. of forw. passesMasking typeRSClean Acc.Rob. Acc. (AA)Rob. Acc. (IW-WC)
11None✗85.3656.1756.12
21None✓84.9356.1956.14
32None✗85.3556.1556.12
42Random-50✗85.2356.1456.12
52Random-50✓85.3356.2056.15
62CD-IR-50✗85.0959.0056.12
72CD-IR-50✓84.9860.2657.46
82LO-IR-50✗85.3358.5956.12
92LO-IR-50✓85.1759.6157.88
+ +of information. For CD-IR, the clean accuracy can be retained up to $k = 100$ while for L0-IR, it can be retained up to at least $k = 400$ . As explained earlier, it is due to the relatively lower importance ranking quality of CD-IR w.r.t. L0-IR. However, both methods are fairly insensitive to the choice of $k$ . + +We also study the sensitivity to choice of layer being masked in Fig. 5C, D on CIFAR10 with ResNet18. We fix the number of neurons being retained $k$ to 50 and vary the layer being masked from layer1 (shallower) to layer4 (deeper). We observe that masking deeper layers is better, since masking shallower layers affects the feature extraction [51], leading to a large drop in clean accuracy, and consequently a drop in robust accuracy. For all the other experiments, we simply mask the penultimate layer since the clean accuracy drop is the lowest. Please refer to Appendix C for more analysis experiments. + +c) Inference time comparisons. Since we use $n_s = 1$ in randomized smoothing and perform two forward passes, the inference time of IG-Defense is $2 \times$ of a single forward pass. The compared test-time defenses HD [43], SODEF [26], and CAAA [3] have computational complexities of $46 \times$ , $2 \times$ , and $8 \times$ respectively. Also, the recent review paper [16] shows that the lowest inference time is $2 \times$ for existing test-time defenses. Hence, as per Table 5, we achieve the most gains in worst-case robustness with the lowest inference time. + +# 6 Conclusion + +In this work, we proposed a novel, training-free, and effective test-time defense (IG-Defense) guided by interpretability through neuron importance ranking. Our IG-Defense outperformed existing test-time defenses on worst-case adaptive attacks while being among the most efficient test-time defenses to date, establishing the new state-of-the-art. We also demonstrated consistent improvements across standard CIFAR and ImageNet-1k benchmarks. + +# Acknowledgements + +This work is supported in part by National Science Foundation (NSF) awards CNS-1730158, ACI-1540112, ACI1541349, OAC-1826967, OAC-2112167, CNS-2100237, CNS-2120019, the University of California Office of the President, and the University of California San Diego's California Institute for Telecommunications and Information Technology/Qualcomm Institute. Thanks to CENIC for the 100Gbps networks. The authors thank the anonymous reviewers for valuable feedback on the manuscript. T.-W. Weng is supported by National Science Foundation under Grant No. 2107189 and 2313105. T.-W. Weng also thanks the Hellman Fellowship for providing research support. + +# References + +1. Addepalli, S., Jain, S., Babu, R.V.: Efficient and effective augmentation strategy for adversarial training. In: NeurIPS (2022) +2. Addepalli, S., Jain, S., Sriramananan, G., Venkatesh Babu, R.: Scaling adversarial training to large perturbation bounds. In: ECCV (2022) +3. Alfarra, M., Pérez, J.C., Thabet, A., Bibi, A., Torr, P.H.S., Ghanem, B.: Combating adversaries with anti-adversaries. In: AAAI (2022) +4. Andriushchenko, M., Croce, F., Flammarion, N., Hein, M.: Square attack: a query-efficient black-box adversarial attack via random search. In: ECCV (2020) +5. Athalye, A., Carlini, N., Wagner, D.: Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In: ICML (2018) +6. Bai, N., Iyer, R.A., Oikarinen, T., Weng, T.W.: Describe-and-dissect: Interpreting neurons in vision networks with language models. arXiv preprint arXiv:2403.13771 (2024) +7. Bai, Y., Zeng, Y., Jiang, Y., Xia, S.T., Ma, X., Wang, Y.: Improving adversarial robustness via channel-wise activation suppressing. In: ICLR (2021) +8. Bau, D., Zhou, B., Khosla, A., Oliva, A., Torralba, A.: Network dissection: Quantifying interpretability of deep visual representations. In: CVPR (2017) +9. Bau, D., Zhu, J.Y., Strobelt, H., Lapedriza, A., Zhou, B., Torralba, A.: Understanding the role of individual units in a deep neural network. Proceedings of the National Academy of Sciences (2020) +10. Boopathy, A., Liu, S., Zhang, G., Liu, C., Chen, P.Y., Chang, S., Daniel, L.: Proper network interpretability helps adversarial robustness in classification. In: ICML (2020) +1.1. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: IEEE Symposium on Security and Privacy (2017) +2. Chen, J., Gu, Q.: RayS: A ray searching method for hard-label adversarial attack In: KDD (2020) +3. Chen, Y., Oliver, D.S.: Ensemble randomized maximum likelihood method as an iterative ensemble smoother. Mathematical Geosciences 44, 1-26 (2012) +4. Chen, Z., Li, Q., Zhang, Z.: Towards robust neural networks via close-loop control In: ICLR (2021) +5. Croce, F., Andriushchenko, M., Sehwag, V., Debenedetti, E., Flammarion, N., Chiang, M., Mittal, P., Hein, M.: Robustbench: a standardized adversarial robustness benchmark. In: Datasets and Benchmarks Track, NeurIPS (2021) + +16. Croce, F., Gowal, S., Brunner, T., Shelhamer, E., Hein, M., Cemgil, T.: Evaluating the adversarial robustness of adaptive test-time defenses. In: ICML (2022) +17. Croce, F., Hein, M.: Minimally distorted adversarial examples with a fast adaptive boundary attack. In: ICML (2020) +18. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: ICML (2020) +19. Eigen, H., Sadovnik, A.: TopKConv: Increased adversarial robustness through deeper interpretability. In: ICMLA (2021) +20. Gerasimou, S., Eniser, H.F., Sen, A., Cakan, A.: Importance-driven deep learning system testing. In: ACM/IEEE ICSE (2020) +21. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: ICLR (2015) +22. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition In: CVPR (2016) +23. Hernandez, E., Schwettmann, S., Bau, D., Bagashvili, T., Torralba, A., Andreas, J.: Natural language descriptions of deep visual features. In: ICLR (2022) +24. Hwang, D., Lee, E., Rhee, W.: AID-Purifier: A light auxiliary network for boosting adversarial defense. In: ICPR (2022) +25. Jia, X., Zhang, Y., Wu, B., Ma, K., Wang, J., Cao, X.: LAS-AT: adversarial training with learnable attack strategy. In: CVPR (2022) +26. Kang, Q., Song, Y., Ding, Q., Tay, W.P.: Stable neural ODE with Lyapunov-stable equilibrium points for defending against adversarial attacks. In: NeurIPS (2021) +27. Krizhevsky, A.: Learning multiple layers of features from tiny images. Tech. rep., CIFAR (2009) +28. Kundu, S., Nazemi, M., Beerel, P.A., Pedram, M.: A tunable robust pruning framework through dynamic network rewiring of DNNs. In: ASP-DAC (2021) +29. Mangla, P., Singh, V., Balasubramanian, V.N.: On saliency maps and adversarial robustness. In: ECML-PKDD (2020) +30. Mao, C., Chiquier, M., Wang, H., Yang, J., Vondrick, C.: Adversarial attacks are reversible with natural supervision. In: ICCV (2021) +31. Mu, J., Andreas, J.: Compositional explanations of neurons. In: NeurIPS (2020) +32. Nayak, G.K., Rawal, R., Chakraborty, A.: DAD: Data-free adversarial defense at test time. In: WACV (2022) +33. Oikarinen, T., Weng, T.W.: CLIP-Dissect: Automatic description of neuron representations in deep vision networks. In: ICLR (2023) +34. Oikarinen, T., Weng, T.W.: Linear explanations for individual neurons. In: ICML (2024) +35. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021) +36. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) 115(3), 211-252 (2015) +37. Salman, H., Ilyas, A., Engstrom, L., Kapoor, A., Madry, A.: Do adversarially robust imagenet models transfer better? In: NeurIPS (2020) +38. Sehwag, V., Wang, S., Mittal, P., Jana, S.: Hydra: Pruning adversarially robust neural networks. In: NeurIPS (2020) +39. Shi, C., Holtz, C., Mishne, G.: Online adversarial purification based on self-supervised learning. In: ICLR (2021) + +40. Sriramanan, G., Addepalli, S., Baburaj, A., Babu, R.V.: Guided adversarial attack for evaluating and enhancing adversarial defenses. In: NeurIPS (2020) +41. Sriramanan, G., Addepalli, S., Baburaj, A., Babu, R.V.: Towards efficient and effective adversarial training. In: NeurIPS (2021) +42. Wong, E., Rice, L., Kolter, J.Z.: Fast is better than free: Revisiting adversarial training. In: ICLR (2020) +43. Wu, B., Pan, H., Shen, L., Gu, J., Zhao, S., Li, Z., Cai, D., He, X., Liu, W.: Attacking adversarial attacks as a defense. arXiv preprint arXiv:2106.04938 (2021) +44. Wu, D., Xia, S.T., Wang, Y.: Adversarial weight perturbation helps robust generalization. In: NeurIPS (2020) +45. Wu, S., Sang, J., Xu, K., Zhang, J., Yu, J.: Attention, please! adversarial defense via activation rectification and preservation. ACM Transactions on Multimedia Computing, Communications and Applications 19(4), 1-18 (2023) +46. Wu, Y.H., Yuan, C.H., Wu, S.H.: Adversarial robustness via runtime masking and cleansing. In: ICML (2020) +47. Xiao, C., Zhong, P., Zheng, C.: Enhancing adversarial defense by k-winners-take-all In: ICLR (2020) +48. Xie, X., Li, T., Wang, J., Ma, L., Guo, Q., Yuefei-Xu, F., Liu, Y.: NPC: Neuron path coverage via characterizing decision logic of deep neural networks. ACM Transactions on Software Engineering and Methodology 31(3) (2022) +49. Yoon, J., Hwang, S.J., Lee, J.: Adversarial purification with score-based generative models. In: ICML (2021) +50. Zagoruyko, S., Komodakis, N.: Wide residual networks. In: BMVC (2016) +51. Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: ECCV (2014) +52. Zhang, H., Yu, Y., Jiao, J., Xing, E.P., Ghaoui, L.E., Jordan, M.I.: Theoretically principled trade-off between robustness and accuracy. In: ICML (2019) +53. Zhao, Q., Wressnegger, C.: Holistic adversarially robust pruning. In: ICLR (2023) \ No newline at end of file diff --git a/2024/Interpretability-Guided Test-Time Adversarial Defense/images.zip b/2024/Interpretability-Guided Test-Time Adversarial Defense/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2102953c392aa8b03e904a100cd3ab1da17315fe --- /dev/null +++ b/2024/Interpretability-Guided Test-Time Adversarial Defense/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3da58c93f5e46036f8084f16ec7e9de8994945a3c5f94840ccfd50a6cae293f6 +size 460161 diff --git a/2024/Interpretability-Guided Test-Time Adversarial Defense/layout.json b/2024/Interpretability-Guided Test-Time Adversarial Defense/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d648dbc79daa55dfa249d41dee7fb3001e8fa376 --- /dev/null +++ b/2024/Interpretability-Guided Test-Time Adversarial Defense/layout.json @@ -0,0 +1,10758 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 145, + 111, + 470, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 111, + 470, + 146 + ], + "spans": [ + { + "bbox": [ + 145, + 111, + 470, + 146 + ], + "type": "text", + "content": "Interpretability-Guided Test-Time Adversarial Defense" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 216, + 167, + 398, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 167, + 398, + 181 + ], + "spans": [ + { + "bbox": [ + 216, + 167, + 398, + 181 + ], + "type": "text", + "content": "Akshay Kulkarni" + }, + { + "bbox": [ + 216, + 167, + 398, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 216, + 167, + 398, + 181 + ], + "type": "text", + "content": " and Tsui-Wei Weng" + }, + { + "bbox": [ + 216, + 167, + 398, + 181 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 259, + 189, + 353, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 189, + 353, + 201 + ], + "spans": [ + { + "bbox": [ + 259, + 189, + 353, + 201 + ], + "type": "text", + "content": "1 CSE, UC San Diego" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 259, + 201, + 354, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 201, + 354, + 213 + ], + "spans": [ + { + "bbox": [ + 259, + 201, + 354, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 259, + 201, + 354, + 213 + ], + "type": "text", + "content": " HDSI, UC San Diego" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 242, + 213, + 370, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 213, + 370, + 223 + ], + "spans": [ + { + "bbox": [ + 242, + 213, + 370, + 223 + ], + "type": "text", + "content": "{a2kulkarni,lweng}@ucsd.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 251, + 455, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 251, + 455, + 384 + ], + "spans": [ + { + "bbox": [ + 159, + 251, + 455, + 384 + ], + "type": "text", + "content": "Abstract. We propose a novel and low-cost test-time adversarial defense by devising interpretability-guided neuron importance ranking methods to identify neurons important to the output classes. Our method is a training-free approach that can significantly improve the robustness-accuracy tradeoff while incurring minimal computational overhead. While being among the most efficient test-time defenses (4× faster), our method is also robust to a wide range of black-box, white-box, and adaptive attacks that break previous test-time defenses. We demonstrate the efficacy of our method for CIFAR10, CIFAR100, and ImageNet-1k on the standard RobustBenchmark (with average gains of 2.6%, 4.9%, and 2.8% respectively). We also show improvements (average 1.5%) over the state-of-the-art test-time defenses even under strong adaptive attacks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 420, + 230, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 230, + 433 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 230, + 433 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 445, + 483, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 483, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 483, + 528 + ], + "type": "text", + "content": "Despite the popularity and success of deep neural networks (DNNs), they have been shown to be brittle against carefully designed small perturbations that are imperceptible to humans [21]. These \"adversarial\" perturbations pose significant risks against deploying DNNs for safety-critical tasks like autonomous driving. To mitigate these risks, there has been great interest in developing defenses, e.g. adversarial training algorithms [40, 42, 44, 52] to obtain a model that is more robust to adversarial samples." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 529, + 483, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 483, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 483, + 636 + ], + "type": "text", + "content": "However, adversarial training is much more expensive than standard training (e.g. a strong defense like TRADES [52] takes " + }, + { + "bbox": [ + 130, + 529, + 483, + 636 + ], + "type": "inline_equation", + "content": "150 - 200 \\times" + }, + { + "bbox": [ + 130, + 529, + 483, + 636 + ], + "type": "text", + "content": " longer than standard training). This inspired training-free defenses known as test-time defenses [3, 26, 30, 39]. The underlying principles in existing test-time defenses are either input purification or model adaptation: input purification [3, 39] modifies the input with an additional perturbation to nullify any adversarial perturbation, whereas model adaptation [14, 26] updates model parameters or adds new parameters at test-time to make the overall model more robust. Despite eliminating the need for training, most existing methods are still quite expensive, " + }, + { + "bbox": [ + 130, + 529, + 483, + 636 + ], + "type": "inline_equation", + "content": "8 - 500 \\times" + }, + { + "bbox": [ + 130, + 529, + 483, + 636 + ], + "type": "text", + "content": " more than" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 641, + 483, + 664 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 641, + 483, + 664 + ], + "spans": [ + { + "bbox": [ + 133, + 641, + 483, + 664 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 133, + 641, + 483, + 664 + ], + "type": "text", + "content": " Our code is available at https://github.com/Trustworthy-ML-Lab/Interpretability-Guided-Defense" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 169, + 115, + 445, + 273 + ], + "blocks": [ + { + "bbox": [ + 169, + 115, + 445, + 273 + ], + "lines": [ + { + "bbox": [ + 169, + 115, + 445, + 273 + ], + "spans": [ + { + "bbox": [ + 169, + 115, + 445, + 273 + ], + "type": "image", + "image_path": "91dac736033fb7175228124be14c872e907811122a13ec7ccae3e405b11e37cf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 281, + 483, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 281, + 483, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 483, + 304 + ], + "type": "text", + "content": "Fig. 1: Interpretability-guided masking based on neuron importance ranking for test-time adversarial defense." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 329, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 376 + ], + "type": "text", + "content": "a standard forward pass [16]. Moreover, existing test-time defenses have been broken under strong adaptive attacks [16]. Motivated by these limitations, in this paper, we intend to find an efficient and effective alternative to improve robustness at test-time, so that the defense does not break under adaptive attacks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 378, + 483, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 483, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 483, + 605 + ], + "type": "text", + "content": "To design an effective defense, we leverage recent neuron-level interpretability methods [8, 33], which allow one to interpret the functionalities of individual neurons. In our analysis, we observed that for successful adversarial examples, the majority of activations at an intermediate layer shift from important neurons for the ground-truth class to the important neurons for other classes. Hence, we propose a novel test-time Interpretability-Guided Defense (IG-Defense) enabled by \"neuron importance ranking\" where class-wise importance of each neuron is computed (Fig. 1A). To design effective neuron importance ranking, we repurpose our analysis experiments and existing neuron-level interpretability works to propose two novel neuron importance ranking methods. The key idea of our IG-Defense is to restrict the observed activation shift by masking unimportant neurons (Fig. 1B). We demonstrate the usefulness of our insights by extensively evaluating our IG-Defense on the standard RobustBench benchmark (Fig. 1C). We observe consistent gains for CIFAR10 (+2.6%), CIFAR100 (+4.9%), and ImageNet-1k (+2.8%). We also devise an ensemble of strong adaptive attacks specifically targeting the weaknesses of IG-Defense and observe consistent gains (+1.5%), unlike existing test-time defenses that break under adaptive attacks. Due to our simple yet grounded approach, IG-Defense is among the most efficient test-time defenses (4× faster than the strongest existing defense)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 607, + 353, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 607, + 353, + 618 + ], + "spans": [ + { + "bbox": [ + 132, + 607, + 353, + 618 + ], + "type": "text", + "content": "Our primary contributions can be summarized as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 629, + 482, + 665 + ], + "type": "text", + "content": "- We are the first to propose a neuron-interpretability-guided test-time defense (IG-Defense) utilizing neuron importance ranking (CD-IR, LO-IR) to improve adversarial robustness. IG-Defense is training-free, efficient, and effective." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 288, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 288, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 288, + 102 + ], + "type": "text", + "content": "A. Kulkarni and T. W. Weng" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 481, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 481, + 138 + ], + "type": "text", + "content": "- We uncover novel insights into improving adversarial robustness by analyzing adversarial attacks through the lens of neuron-level interpretability in Sec. 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 481, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 481, + 198 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 481, + 198 + ], + "type": "text", + "content": "- Our proposed IG-Defense consistently improves the robustness on standard CIFAR10, CIFAR100, and ImageNet-1k benchmarks. We also demonstrate improved robustness upto " + }, + { + "bbox": [ + 138, + 140, + 481, + 198 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 138, + 140, + 481, + 198 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 138, + 140, + 481, + 198 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 138, + 140, + 481, + 198 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 138, + 140, + 481, + 198 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 138, + 140, + 481, + 198 + ], + "type": "text", + "content": " against a wide range of white-box, black-box, and adaptive attacks respectively with the lowest inference time (" + }, + { + "bbox": [ + 138, + 140, + 481, + 198 + ], + "type": "inline_equation", + "content": "4\\times" + }, + { + "bbox": [ + 138, + 140, + 481, + 198 + ], + "type": "text", + "content": " faster) among existing test-time defenses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 213, + 237, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 213, + 237, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 213, + 237, + 226 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 235, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 482, + 390 + ], + "type": "text", + "content": "Test-time adversarial robustness. Most of the test-time defenses can be categorized into two approaches: input purification or model adaptation. In input purification, the adversarial input is converted to a safer input by an additional perturbation. For instance, finding a perturbation that either increases top class confidence [3], minimizes contrastive loss [30], or minimizes a self-supervised objective [39]. [24] crafts the perturbation by fooling a real-adversarial discriminator [24] and [49] uses an input passed through a score-based generative model. For model adaptation [14,32,46], the model parameters are either updated at test-time or new parameters are introduced. However, existing test-time defenses are quite expensive (8-500× of single forward pass) and break under adaptive attacks. In contrast, our method is fast and effective (1.5% gain under adaptive attacks with only 2× inference time), and is the first test-time defense to leverage neuron-level interpretability information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 391, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 521 + ], + "type": "text", + "content": "Neuron interpretability methods. To understand the functionalities (i.e. concepts) of individual neurons in DNNs, early works like NetDissect [8,9] compare the neuron activation patterns with patterns of predefined concept label masks. MILAN [23] produced natural language descriptions using a supervised image captioning model trained on a large-scale concept-labeled dataset. Recently, CLIP-Dissect [33] leveraged the CLIP paradigm [35] to identify concepts of neurons more efficiently than prior work and without the need of training or densely-concept-labeled datasets. There are other works to identify neuron concepts through clustering [20, 48], via language models [6], or using linear combinations [34]. In this work, we focus on repurposing CLIP-Dissect due to its efficiency and simplicity as well as NetDissect to propose neuron importance ranking methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "type": "text", + "content": "Connections between interpretability and robustness. Prior works attempt to improve robustness via input-dependent interpretability methods like saliency or class activation maps [10, 29]. Wu et al. [45] observed that adversarial attacks lead to changes in attention maps, and proposed a training-time defense with losses to enforce similar attention maps for clean and adversarial samples. Several works [7, 19, 28, 38, 47, 53] use sparsity to improve adversarial robustness. However, they require retraining after or while sparsity constraints are applied. In contrast, we indirectly leverage sparsity without requiring any training. NetDissect [8] investigated the connection between adversarial attacks and interpretability information from their method, but they did not explore how to improve robustness. Hence, we aim to study the unexplored connection of neuron-level interpretability and improving adversarial robustness." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 224, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 224, + 91, + 447, + 102 + ], + "type": "text", + "content": "Interpretability-Guided Test-Time Adversarial Defense" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 283, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 283, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 283, + 129 + ], + "type": "text", + "content": "3 Analysis Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 146, + 482, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 146, + 482, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 146, + 482, + 301 + ], + "type": "text", + "content": "To understand the connection of interpretability of individual neurons with robustness, we aim to design an analysis experiment. One way would be to investigate the changes in neuron activations before and after an adversarial attack. In neuron-labeling works [9,31], they assign an explainable concept to each neuron (or an activation channel in case of CNNs). Specifically, NetDissect [9] qualitatively show that a targeted adversarial attack causes a drop in peak activations for neurons whose concepts are related to the original class, while causing an increase in peak activations for neurons whose concepts are related to the target class of the attack. They also show that the peak activation change is lower for neurons with unrelated concepts. While they only perform the analysis to showcase their interpretability, we are the first to analyze adversarial attacks through neuron-level interpretability with the explicit goal of obtaining insights into improving adversarial robustness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 303, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 411 + ], + "type": "text", + "content": "For a more thorough analysis, we observe the average activation change (instead of peak change) and associate the neurons with the category names directly instead of concepts. First, peak activation change may be misleading since it may be caused by specific images while average change takes all images into account for a more well-rounded analysis. Second, our automated evaluation eliminates the need for manual association of concepts with categories [9], which could be difficult, expensive, and noisy, especially with a large number of related categories (e.g. ImageNet classes). We analyze both successful and unsuccessful attacks while [9] only focused on successful attacks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 411, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 471 + ], + "type": "text", + "content": "Concretely, we find top-" + }, + { + "bbox": [ + 130, + 411, + 482, + 471 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 411, + 482, + 471 + ], + "type": "text", + "content": " important neurons related to each task category using neuron importance ranking (discussed in Sec. 4.2). With this, in Fig. 2, we analyze the change in activations corresponding to ground-truth (GT) and non-GT categories before and after an attack. We discuss some preliminaries before delving into further details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": "Preliminaries. We denote the dataset as " + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x,y)\\}" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "inline_equation", + "content": "x\\in \\mathcal{X}" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": " is the input image, " + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "inline_equation", + "content": "y\\in \\mathcal{C}" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": " is the label, " + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": " is the label set, and " + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "inline_equation", + "content": "|\\mathcal{C}|" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": " is the number of classes. Given a pretrained model " + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "inline_equation", + "content": "f:\\mathcal{X}\\to \\mathcal{C}" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": " and its activations at the layer being studied " + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "inline_equation", + "content": "A:\\mathcal{X}\\rightarrow \\mathbb{R}^{N\\times H\\times W}" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": ", we can rank the importance of a neuron (out of " + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": " neurons/channels) by assessing its contribution to each class. We call this \"neuron importance ranking\" and propose two algorithms for this (Sec. 4.2). Then, we can compute the activations of the top-" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": " important neurons for any class " + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "inline_equation", + "content": "A_{c_i}(x) = A(x)_{[n_{c_i}]} \\in \\mathbb{R}^{k\\times H\\times W}" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "inline_equation", + "content": "n_{c_i}" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": " are the indices of the top-" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": " activations for class " + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": ". With access to each " + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "inline_equation", + "content": "n_{c_i}" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": ", the remaining indices absent from the top-" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": " of any class are " + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "inline_equation", + "content": "n_u = [N] \\setminus \\{\\cup_{c_i\\in \\mathcal{C}} n_{c_i}\\}" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "inline_equation", + "content": "[N] = \\{1,2,\\dots,N\\}" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\backslash" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": " indicates setminus operator. These indices " + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "inline_equation", + "content": "n_u" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": " denote unimportant neurons, and corresponding unimportant activations can be computed as " + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "inline_equation", + "content": "A_u(x) = A(x)_{[n_u]} \\in \\mathbb{R}^{(N - k|\\mathcal{C}|) \\times H \\times W}" + }, + { + "bbox": [ + 131, + 472, + 482, + 628 + ], + "type": "text", + "content": ". Next, we define activation change metrics to support our analysis." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 629, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 483, + 666 + ], + "type": "text", + "content": "Metrics. For an adversarial attack with " + }, + { + "bbox": [ + 131, + 629, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 131, + 629, + 483, + 666 + ], + "type": "text", + "content": "-norm bound of " + }, + { + "bbox": [ + 131, + 629, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 131, + 629, + 483, + 666 + ], + "type": "text", + "content": ", the perturbation for input " + }, + { + "bbox": [ + 131, + 629, + 483, + 666 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 629, + 483, + 666 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 131, + 629, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\delta^{*} = \\arg \\max_{\\| \\delta \\|_{\\infty} \\leq \\epsilon} \\ell(f(x + \\delta), y)" + }, + { + "bbox": [ + 131, + 629, + 483, + 666 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 131, + 629, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 131, + 629, + 483, + 666 + ], + "type": "text", + "content": " is the attack objective. Let " + }, + { + "bbox": [ + 131, + 629, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\hat{y}^{*} = f(x + \\delta^{*})" + }, + { + "bbox": [ + 131, + 629, + 483, + 666 + ], + "type": "text", + "content": " be the post-attack prediction. The average activation change" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 286, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 286, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 286, + 101 + ], + "type": "text", + "content": "A. Kulkarni and T. W. Weng" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 169, + 114, + 445, + 266 + ], + "blocks": [ + { + "bbox": [ + 169, + 114, + 445, + 266 + ], + "lines": [ + { + "bbox": [ + 169, + 114, + 445, + 266 + ], + "spans": [ + { + "bbox": [ + 169, + 114, + 445, + 266 + ], + "type": "image", + "image_path": "44563d80def7e9948ace4e50d33af7c4f22143b790342aa4a792eb2d9d1d0ac9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 276, + 482, + 309 + ], + "lines": [ + { + "bbox": [ + 130, + 276, + 482, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 482, + 309 + ], + "type": "text", + "content": "Fig. 2: A. Cat-dog-bird image classifier before a PGD attack. B. After PGD attack, the prediction changes from the ground truth (GT) cat to dog since the activations of neurons important to dog class increase while those important to cat class decrease." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 310, + 481, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 481, + 331 + ], + "type": "text", + "content": "C. Empirically, successful PGD attacks show a decrease in activations of important GT class neurons while those of post-attack predicted class' important neurons increase." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 331, + 481, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 481, + 353 + ], + "type": "text", + "content": "D. For unsuccessful PGD attacks, the activations of all neurons reduce even though the prediction remains the same as before the attack." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 377, + 337, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 377, + 337, + 389 + ], + "spans": [ + { + "bbox": [ + 132, + 377, + 337, + 389 + ], + "type": "text", + "content": "in the important neurons w.r.t. any class " + }, + { + "bbox": [ + 132, + 377, + 337, + 389 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 132, + 377, + 337, + 389 + ], + "type": "text", + "content": " is," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 228, + 396, + 482, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 396, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 228, + 396, + 482, + 417 + ], + "type": "interline_equation", + "content": "\\Delta_ {c _ {i}} = \\underset {(x, y) \\in \\mathcal {D}} {\\mathbb {E}} \\left[ A _ {c _ {i}} \\left(x + \\delta^ {*}\\right) - A _ {c _ {i}} (x) \\right] \\tag {1}", + "image_path": "58cb6b757819c51908298ba5cc6df2ed2ffa85c5a05da16aa85f8f6c83f73135.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "content": "Replacing " + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "content": " with GT class " + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "content": ", the average activation change in the important neurons of GT class is " + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\varDelta_{y}" + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "content": ". Similarly, the average activation change in post-attack predicted class neurons is " + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\varDelta_{\\hat{y}^*}" + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "content": ", and that in other non-GT class neurons is " + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\varDelta_{\\mathrm{non - GT}} = \\mathbb{E}_{c_i\\neq y,\\hat{y}^*}[\\varDelta_{c_i}]" + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "content": ". For unsuccessful attacks, where post-attack prediction " + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\hat{y}^{*}" + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "content": " is the same as " + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "content": ", we can compute average activation change in the remaining classes (other than GT) as " + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\varDelta_{\\mathrm{rem -cls}}=\\mathbb{E}_{c_i\\neq y}[\\varDelta_{c_i}]" + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "content": ". Finally, we can also compute the average activation change in unimportant neurons," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 222, + 517, + 482, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 517, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 222, + 517, + 482, + 538 + ], + "type": "interline_equation", + "content": "\\Delta_ {\\text {u n i m p}} = \\underset {(x, y) \\in \\mathcal {D}} {\\mathbb {E}} \\left[ A _ {u} \\left(x + \\delta^ {*}\\right) - A _ {u} (x) \\right] \\tag {2}", + "image_path": "3870796cffebd25a93f91b7759af6c4e4f4cdb75324bdaa76c653bc30e1df982.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 546, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 593 + ], + "type": "text", + "content": "We empirically compute the above metrics on CIFAR10 with five base models (ResNet18: GAT [40], NuAT [41], OAAT [2], DAJAT [1], and WideResNet-34-10: TRADES-AWP [44]) and present the mean values in Fig. 2C and Fig. 2D. See Appendix for more experimental details and corresponding numerical results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Observations. In Fig. 2C, we observe " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\varDelta_y<0" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\varDelta_{\\hat{y}^*}>0" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": ", i.e. successful adversarial attacks boost the activations that are important to the post-attack predicted class " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\hat{y}^*" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " while causing a drop for those of the GT class " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": ". Further, from Fig. 2C, the important activations for the remaining classes and unimportant activations are only marginally affected by successful attacks. In Fig. 2D, we observe that unsuccessful attacks cause a drop in the activations of all neurons." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 224, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 224, + 91, + 448, + 102 + ], + "type": "text", + "content": "Interpretability-Guided Test-Time Adversarial Defense" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "Remarks. During both successful and unsuccessful attacks, the activation magnitude of GT class' important neurons drops. Specifically, successful attacks shift the activation magnitude from GT class' important neurons to non-GT class important neurons. Hence, we hypothesize that adversarial robustness can be improved if the activation shift to non-GT class important neurons is restricted." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 481, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 481, + 213 + ], + "type": "text", + "content": "Further, given the availability of a huge number of adversarially pretrained models [15], we aim to evaluate our hypothesis in a test-time adversarial defense setting, with a low inference overhead and without any training overhead." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 231, + 344, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 231, + 344, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 231, + 344, + 245 + ], + "type": "text", + "content": "4 Interpretability-Guided Defense" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 257, + 482, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 304 + ], + "type": "text", + "content": "In this section, we propose a novel, training-free, test-time Interpretability-Guided Defense (IG-Defense). The core idea of our method is based on interpretability-guided masking, described in Sec. 4.1. In Sec. 4.2, we introduce neuron importance ranking, which is a key enabler in the success of our proposed defense." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 305, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 482, + 342 + ], + "type": "text", + "content": "Setup. Given a trained model " + }, + { + "bbox": [ + 130, + 305, + 482, + 342 + ], + "type": "inline_equation", + "content": "f: \\mathcal{X} \\to \\mathcal{C}" + }, + { + "bbox": [ + 130, + 305, + 482, + 342 + ], + "type": "text", + "content": ", the goal of test-time defense is to improve the robustness of " + }, + { + "bbox": [ + 130, + 305, + 482, + 342 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 305, + 482, + 342 + ], + "type": "text", + "content": " while retaining its utility (i.e. clean accuracy). We refer to " + }, + { + "bbox": [ + 130, + 305, + 482, + 342 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 305, + 482, + 342 + ], + "type": "text", + "content": " as the \"base model\" in the rest of the paper." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 360, + 321, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 360, + 321, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 360, + 321, + 372 + ], + "type": "text", + "content": "4.1 Interpretability-guided masking" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 380, + 481, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 481, + 429 + ], + "type": "text", + "content": "To validate our hypothesis, we design a simple approach for test-time defense where we mask all the neurons except the important neurons of the correct class. We propose a dual forward pass approach, since access to labels cannot be assumed at test-time. Our approach contains three key steps:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 437, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 482, + 484 + ], + "type": "text", + "content": "- Step 1: Neuron Importance Ranking. First, we compute class-wise neuron importance ranking (Step 1 of Fig. 3). This is computed only once for a given base model " + }, + { + "bbox": [ + 138, + 437, + 482, + 484 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 138, + 437, + 482, + 484 + ], + "type": "text", + "content": ". For each layer, we rank every neuron based on its importance to each class. More details are given in Sec. 4.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 155, + 485, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 485, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 155, + 485, + 482, + 522 + ], + "type": "text", + "content": "- Given the number of important neurons " + }, + { + "bbox": [ + 155, + 485, + 482, + 522 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 155, + 485, + 482, + 522 + ], + "type": "text", + "content": " (hyperparameter), we compute a binary mask " + }, + { + "bbox": [ + 155, + 485, + 482, + 522 + ], + "type": "inline_equation", + "content": "m \\in \\{0,1\\}^{N \\times C}" + }, + { + "bbox": [ + 155, + 485, + 482, + 522 + ], + "type": "text", + "content": " using the indices of important neurons " + }, + { + "bbox": [ + 155, + 485, + 482, + 522 + ], + "type": "inline_equation", + "content": "n_{c_i} \\in \\mathbb{N}^k" + }, + { + "bbox": [ + 155, + 485, + 482, + 522 + ], + "type": "text", + "content": " (from the importance ranking in Sec. 4.2) for each class " + }, + { + "bbox": [ + 155, + 485, + 482, + 522 + ], + "type": "inline_equation", + "content": "c_i \\in [C]" + }, + { + "bbox": [ + 155, + 485, + 482, + 522 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 200, + 533, + 482, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 533, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 200, + 533, + 482, + 547 + ], + "type": "interline_equation", + "content": "m \\left[ n _ {c _ {i}}, c _ {i} \\right] = 1 \\forall c _ {i} \\in [ C ], \\quad m \\left[ \\left[ N \\right] \\backslash n _ {c _ {i}}, c _ {i} \\right] = 0 \\forall c _ {i} \\in [ C ] \\tag {3}", + "image_path": "9fe433c783d30c9d43ccd1686b8c63834de6d87af07064ae7bad4531677eca59.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 155, + 552, + 482, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 552, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 155, + 552, + 482, + 576 + ], + "type": "text", + "content": "- In the mask " + }, + { + "bbox": [ + 155, + 552, + 482, + 576 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 155, + 552, + 482, + 576 + ], + "type": "text", + "content": ", the top-" + }, + { + "bbox": [ + 155, + 552, + 482, + 576 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 155, + 552, + 482, + 576 + ], + "type": "text", + "content": " important neurons of each class " + }, + { + "bbox": [ + 155, + 552, + 482, + 576 + ], + "type": "inline_equation", + "content": "c_i \\in [C]" + }, + { + "bbox": [ + 155, + 552, + 482, + 576 + ], + "type": "text", + "content": " are retained in " + }, + { + "bbox": [ + 155, + 552, + 482, + 576 + ], + "type": "inline_equation", + "content": "m[:, c_i]" + }, + { + "bbox": [ + 155, + 552, + 482, + 576 + ], + "type": "text", + "content": ". The remaining neurons unimportant to " + }, + { + "bbox": [ + 155, + 552, + 482, + 576 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 155, + 552, + 482, + 576 + ], + "type": "text", + "content": " are masked." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 137, + 576, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 576, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 137, + 576, + 482, + 629 + ], + "type": "text", + "content": "- Step 2: Vanilla Forward Pass. In the first forward pass (Step 2 of Fig. 3), we obtain a soft-pseudo-label " + }, + { + "bbox": [ + 137, + 576, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\hat{y} = \\sigma\\left(\\frac{f(x)}{\\tau}\\right) \\in (0,1)^C" + }, + { + "bbox": [ + 137, + 576, + 482, + 629 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 137, + 576, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 137, + 576, + 482, + 629 + ], + "type": "text", + "content": " denotes the softmax function, and " + }, + { + "bbox": [ + 137, + 576, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 137, + 576, + 482, + 629 + ], + "type": "text", + "content": " denotes the temperature term which controls the sharpness of the soft-pseudo-label." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 137, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 137, + 629, + 482, + 666 + ], + "type": "text", + "content": "- Step 3: Masked Forward Pass. In the second forward pass (Step 3 of Fig. 3), we apply a soft-pseudo-label weighted mask i.e. " + }, + { + "bbox": [ + 137, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "m\\hat{y} \\in (0,1)^N" + }, + { + "bbox": [ + 137, + 629, + 482, + 666 + ], + "type": "text", + "content": " to the activations of the layer being masked. Let " + }, + { + "bbox": [ + 137, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "A:\\mathcal{X}\\to \\mathbb{R}^{N\\times H\\times W}" + }, + { + "bbox": [ + 137, + 629, + 482, + 666 + ], + "type": "text", + "content": " give" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 287, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 287, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 287, + 102 + ], + "type": "text", + "content": "A. Kulkarni and T. W. Weng" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 211 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 211 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 211 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 211 + ], + "type": "image", + "image_path": "36c4a5e322c35445267218808a94d2a4326a72d99e46df3c2dbd9f1459f16acd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "lines": [ + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "type": "text", + "content": "Fig. 3: Step 1. Given a pretrained base model " + }, + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "type": "text", + "content": " (e.g. binary classifier here), class-wise neuron importance is computed for a selected layer (Sec. 4.2). A top- " + }, + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "type": "text", + "content": " mask " + }, + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "type": "inline_equation", + "content": "m \\in \\{0,1\\}^{N \\times C}" + }, + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "type": "text", + "content": " is computed to identify top- " + }, + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "type": "text", + "content": " neurons important to each class (e.g. " + }, + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "type": "inline_equation", + "content": "k = 2, N = 5, C = 2" + }, + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "type": "text", + "content": " here). Step 2. During evaluation, a soft-pseudo-label " + }, + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "type": "text", + "content": " is computed using the base model " + }, + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "type": "text", + "content": ". Step 3. The soft-pseudo-label weighted mask " + }, + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "type": "inline_equation", + "content": "m\\hat{y} = m\\sigma(\\frac{f(x)}{\\tau}) \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 130, + 220, + 483, + 298 + ], + "type": "text", + "content": " is applied to the selected layer to retain only the important neurons of the pseudo-label class." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 319, + 481, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 319, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 147, + 319, + 481, + 355 + ], + "type": "text", + "content": "the activations at the layer being masked, and " + }, + { + "bbox": [ + 147, + 319, + 481, + 355 + ], + "type": "inline_equation", + "content": "h: \\mathbb{R}^{N \\times H \\times W} \\to \\mathcal{C}" + }, + { + "bbox": [ + 147, + 319, + 481, + 355 + ], + "type": "text", + "content": " be the remaining layers after the layer being masked, i.e. " + }, + { + "bbox": [ + 147, + 319, + 481, + 355 + ], + "type": "inline_equation", + "content": "f = h \\circ A" + }, + { + "bbox": [ + 147, + 319, + 481, + 355 + ], + "type": "text", + "content": ". Then, the final prediction is given by " + }, + { + "bbox": [ + 147, + 319, + 481, + 355 + ], + "type": "inline_equation", + "content": "\\hat{y}' = h(m\\hat{y} \\cdot A(x))" + }, + { + "bbox": [ + 147, + 319, + 481, + 355 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 156, + 356, + 481, + 426 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 156, + 356, + 481, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 356, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 156, + 356, + 481, + 402 + ], + "type": "text", + "content": "- The masking re-weights the neurons based on " + }, + { + "bbox": [ + 156, + 356, + 481, + 402 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 156, + 356, + 481, + 402 + ], + "type": "text", + "content": ". However, the temperature term " + }, + { + "bbox": [ + 156, + 356, + 481, + 402 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 156, + 356, + 481, + 402 + ], + "type": "text", + "content": " is set to a small value so that " + }, + { + "bbox": [ + 156, + 356, + 481, + 402 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 156, + 356, + 481, + 402 + ], + "type": "text", + "content": " contains a sharp probability distribution. Thus, the weighted mask corresponds mainly to a single class and retains only the important neurons of that class." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 156, + 403, + 481, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 403, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 156, + 403, + 481, + 426 + ], + "type": "text", + "content": "- Also note that this enables gradient flow between the two forward passes, to avoid any gradient masking [5]." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 432, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 482, + 479 + ], + "type": "text", + "content": "Intuitively, activations of a deeper layer (e.g. penultimate layer) are masked to minimize any negative impact of the masking on feature extraction, which depends more on the shallower layers [51]. To verify this, we analyze the sensitivity of our method to the layer being masked in Sec. 5.3b." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 480, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 482, + 575 + ], + "type": "text", + "content": "Dependence on soft-pseudo-label. This is an obvious limitation if the soft-pseudo-label is computed directly from the base model " + }, + { + "bbox": [ + 130, + 480, + 482, + 575 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 480, + 482, + 575 + ], + "type": "text", + "content": ". If it is incorrect, masking will yield no improvement in robustness. Hence, an adaptive attack can be designed to ensure that the soft-pseudo-label is incorrect. For instance, a transfer attack that transfers the perturbation from the base model " + }, + { + "bbox": [ + 130, + 480, + 482, + 575 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 480, + 482, + 575 + ], + "type": "text", + "content": " would yield an incorrect soft-pseudo-label, and the final robustness would be the same as the base model " + }, + { + "bbox": [ + 130, + 480, + 482, + 575 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 480, + 482, + 575 + ], + "type": "text", + "content": ". To circumvent this, we propose to use randomized smoothing in the first forward pass (i.e. in Step 2 of Fig. 3)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 575, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 575, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 575, + 482, + 635 + ], + "type": "text", + "content": "Randomized smoothing. To obtain the soft-pseudo-label " + }, + { + "bbox": [ + 130, + 575, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 130, + 575, + 482, + 635 + ], + "type": "text", + "content": " in a more robust manner, we perform randomized smoothing [13] in the first forward pass. We add a set of " + }, + { + "bbox": [ + 130, + 575, + 482, + 635 + ], + "type": "inline_equation", + "content": "n_s" + }, + { + "bbox": [ + 130, + 575, + 482, + 635 + ], + "type": "text", + "content": " Gaussian noises (with zero mean and standard deviation " + }, + { + "bbox": [ + 130, + 575, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\sigma_d" + }, + { + "bbox": [ + 130, + 575, + 482, + 635 + ], + "type": "text", + "content": ") to each input " + }, + { + "bbox": [ + 130, + 575, + 482, + 635 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 575, + 482, + 635 + ], + "type": "text", + "content": ". For the soft-pseudo-label, the logits of these noisy inputs are averaged and softmax " + }, + { + "bbox": [ + 130, + 575, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 575, + 482, + 635 + ], + "type": "text", + "content": " is applied. Formally," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 189, + 642, + 482, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 642, + 482, + 669 + ], + "spans": [ + { + "bbox": [ + 189, + 642, + 482, + 669 + ], + "type": "interline_equation", + "content": "\\hat {y} = \\sigma \\left(\\frac {1}{\\tau} \\underset {v _ {i}} {\\mathbb {E}} [ f (x + v _ {i}) ]\\right); v _ {i} \\sim \\mathcal {N} \\left(0 _ {d}, \\sigma_ {d} ^ {2} I _ {d}\\right) \\forall i \\in [ n _ {s} ] \\tag {4}", + "image_path": "e049bd81adb33312b4a89a10033c02bc6214d3555bc80d5ca6f1b36bb8c6cf8e.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 224, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 224, + 91, + 448, + 102 + ], + "type": "text", + "content": "Interpretability-Guided Test-Time Adversarial Defense" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 482, + 284 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 284 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 284 + ], + "type": "image", + "image_path": "517f549af567e09635029a08a80529426a4100b6f1a1ac8f0401ba853b427c4e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "type": "text", + "content": "Fig. 4: A. Leave-one-Out Importance Ranking (LO-IR) computes class-wise neuron importance as the average change in logits when masking each neuron. B. CLIP-Dissect Importance Ranking (CD-IR) first computes the inner product of class name text embeddings and probing set image embeddings. The importance ranking relies on the similarity between activations of each neuron to the precomputed inner product given the same probing set inputs. C. A higher logit-change " + }, + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\Delta_{c_i}^{[j]}" + }, + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "type": "text", + "content": " or similarity " + }, + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "type": "inline_equation", + "content": "s_{c_i}^{[j]}" + }, + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "type": "text", + "content": " implies that neuron " + }, + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "type": "text", + "content": " is more important to class " + }, + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 130, + 292, + 482, + 369 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 391, + 482, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 427 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 391, + 482, + 427 + ], + "type": "inline_equation", + "content": "0_{d}, I_{d}" + }, + { + "bbox": [ + 130, + 391, + 482, + 427 + ], + "type": "text", + "content": " are tensors with the same shape as " + }, + { + "bbox": [ + 130, + 391, + 482, + 427 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 391, + 482, + 427 + ], + "type": "text", + "content": " but containing all zeros and ones respectively. The number of noises sampled " + }, + { + "bbox": [ + 130, + 391, + 482, + 427 + ], + "type": "inline_equation", + "content": "n_{s}" + }, + { + "bbox": [ + 130, + 391, + 482, + 427 + ], + "type": "text", + "content": " and standard deviation " + }, + { + "bbox": [ + 130, + 391, + 482, + 427 + ], + "type": "inline_equation", + "content": "\\sigma_{d}" + }, + { + "bbox": [ + 130, + 391, + 482, + 427 + ], + "type": "text", + "content": " are hyperparameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 442, + 308, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 442, + 308, + 455 + ], + "spans": [ + { + "bbox": [ + 131, + 442, + 308, + 455 + ], + "type": "text", + "content": "4.2 Importance ranking methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 460, + 481, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 481, + 483 + ], + "type": "text", + "content": "Here, we describe two novel importance ranking methods that can be used in Step 1 of our proposed IG-Defense." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 484, + 482, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 558 + ], + "type": "text", + "content": "a) Leave-one-Out Importance Ranking (L0-IR). Following the analysis in Sec. 3, given a neuron " + }, + { + "bbox": [ + 130, + 484, + 482, + 558 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 484, + 482, + 558 + ], + "type": "text", + "content": ", we can compute its importance to a class " + }, + { + "bbox": [ + 130, + 484, + 482, + 558 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 484, + 482, + 558 + ], + "type": "text", + "content": " as the average change in the class " + }, + { + "bbox": [ + 130, + 484, + 482, + 558 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 484, + 482, + 558 + ], + "type": "text", + "content": " logits, " + }, + { + "bbox": [ + 130, + 484, + 482, + 558 + ], + "type": "inline_equation", + "content": "\\Delta_{c_i}^{[j]}" + }, + { + "bbox": [ + 130, + 484, + 482, + 558 + ], + "type": "text", + "content": ", before and after masking out that particular neuron " + }, + { + "bbox": [ + 130, + 484, + 482, + 558 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 484, + 482, + 558 + ], + "type": "text", + "content": " (Fig. 4A). Intuitively, a higher logit-change for a particular class implies higher dependence of the network on that neuron, i.e. higher importance (Fig. 4C). Formally," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 198, + 563, + 482, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 563, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 198, + 563, + 482, + 586 + ], + "type": "interline_equation", + "content": "\\Delta_ {c _ {i}} ^ {[ j ]} = \\underset {(x, y) \\in \\mathcal {D} _ {\\text {p r o b e}} | y = c _ {i}} {\\mathbb {E}} \\left[ f _ {c _ {i}} (x) - f _ {c _ {i}} ^ {[ j ]} (x) \\right]; \\forall c _ {i} \\in \\mathcal {C} \\tag {5}", + "image_path": "e1ff81749c7e6e66d9951110767a8196d4f2231cbf747911574d6e49a087a583.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "f^{[j]}" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": " is the model where " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "j^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": " neuron of the selected layer is masked in the base model " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": ". And the subscript " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": " indicates the logits of class " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": ". However, this method could be computationally expensive as average change in logits has to be computed for each neuron. For our proposed method, it is only required once per base model. Thus, it is a feasible option that yields a very good estimate of the importance ranking, which we show in Sec. 5." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 287, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 287, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 287, + 102 + ], + "type": "text", + "content": "A. Kulkarni and T. W. Weng" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": "b) CLIP-Dissect Importance Ranking (CD-IR). CLIP-Dissect [33] uses the multimodal CLIP model [35] to assign concept labels to individual neurons. We extend this idea by replacing concepts with task labels to obtain an importance ranking of neurons for each task label (see Fig. 4B)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "text", + "content": "Concretely, probing images are passed through the image encoder " + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "inline_equation", + "content": "E_{I}" + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "text", + "content": " to get a set of image embeddings " + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "inline_equation", + "content": "\\{A^{[i]} = E_I(x_i)\\}_{i = 1}^M" + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "text", + "content": " (shown in orange in Fig. 4B) while the task label names " + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "text", + "content": " are passed through the text encoder " + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "text", + "content": " to get the corresponding set of text embeddings " + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "inline_equation", + "content": "\\{E_T(c_k)\\}_{c_k\\in \\mathcal{C}}" + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "text", + "content": " (shown in green in Fig. 4B). These embeddings i.e. vectors are converted to scalars by taking the inner product of each pair of image-text embeddings, which results in a matrix " + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "text", + "content": " with entries " + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "inline_equation", + "content": "P_{ik} = A^{[i]^\\top}E_T(c_k)" + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "text", + "content": ". Now, for a neuron " + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "text", + "content": ", a vector of its activations " + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "inline_equation", + "content": "q_{j}" + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "text", + "content": " (shown in purple in Fig. 4B) can be obtained where each entry corresponds to each input from the probing dataset. We can compute the similarity score " + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "inline_equation", + "content": "s_{c_k}^{[j]}" + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "text", + "content": " between each class " + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "inline_equation", + "content": "c_{k}" + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "text", + "content": " and neuron " + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "text", + "content": " using the " + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "inline_equation", + "content": "k^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "text", + "content": " column of " + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "inline_equation", + "content": "q_{j}" + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "text", + "content": ". Intuitively, a higher similarity score " + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "inline_equation", + "content": "s_{c_k}^{[j]}" + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "text", + "content": " implies higher importance of neuron " + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "text", + "content": " for class " + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "inline_equation", + "content": "c_{k}" + }, + { + "bbox": [ + 130, + 164, + 482, + 323 + ], + "type": "text", + "content": " (Fig. 4C). We use the same soft-WPMI similarity metric as the original paper [33]. For the probing dataset, we simply use the training dataset for our experiments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 324, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 407 + ], + "type": "text", + "content": "In practice, CD-IR is computationally less expensive compared to L0-IR (e.g. 2 mins. vs. 28 mins. for 512 neurons). This is because CLIP embeddings can be precomputed and most of the computations can be vectorized easily. As we demonstrate in Sec. 5, although the importance ranking quality is slightly worse than L0-IR (i.e. CD-IR causes a marginal drop in clean accuracy), downstream robustness gains with CD-IR are still significant. Please refer to the Appendix C.3 for details on computational complexity comparisons." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 426, + 230, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 230, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 230, + 440 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 485 + ], + "type": "text", + "content": "We extensively evaluate our IG-Defense on the standard RobustBench benchmark [15]. We also follow Croce et al. [16] for properly evaluating test-time defenses and thoroughly test against an ensemble of strong adaptive attacks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "text", + "content": "Experimental setup. We evaluate our methods on standard datasets, CIFAR10, CIFAR100 [27] and ImageNet-1k [36], with ResNet [22] and WideResNet [50] base models. We use an " + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "text", + "content": " threat model with an " + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "text", + "content": "-bound of 8/255 for CIFAR10/100 and 4/255 for ImageNet-1k following prior works [15]. For evaluation, we use the full CIFAR10/100 test set (10000 samples) and the first 5000 ImageNet-1k validation samples, which is the standard setup as per [15,37]. The hyperparameter " + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "text", + "content": " for the top-" + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "text", + "content": " neurons in CD-IR or L0-IR is mentioned alongside the method name in the results. For randomized smoothing (RS), we use " + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "inline_equation", + "content": "n_s = 1" + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\sigma_d = \\frac{\\epsilon}{2}" + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\tau = 0.01" + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "text", + "content": "-bound for the attack. We show in Appendix C.2 that the choices of " + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "inline_equation", + "content": "n_s = 1" + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\sigma_d = \\frac{\\epsilon}{2}" + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "text", + "content": " are sufficient for our purpose. Unless otherwise mentioned, RS is used and the penultimate layer output is masked. In the penultimate layer, there are 512 neurons (layer4) in ResNet18, 640 neurons (layer3 or block3) in WideResNet-34-10, and 2048 neurons (layer4) in ResNet50." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Choice of base models. We experiment with base models from several single-step defenses like FAT [42], GAT [40], NuAT [41], OAAT [2], and multi-step" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 224, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 224, + 91, + 447, + 102 + ], + "type": "text", + "content": "Interpretability-Guided Test-Time Adversarial Defense" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 146, + 479, + 205 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 1: Evaluating our IG-Defense (CD-IR, LO-IR) on CIFAR10 with ResNet18 and " + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}, \\epsilon = 8/255" + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": ". IW-WC indicates image-wise worst-case, AA indicates AutoAttack." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 146, + 479, + 205 + ], + "lines": [ + { + "bbox": [ + 133, + 146, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 146, + 479, + 205 + ], + "type": "table", + "html": "
Base Model:NuAT [41] (RN18)OAAT [2] (RN18)DAJAT [1] (RN18)
Test-Time DefenseClean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)
None82.2150.5850.5480.2351.1051.0185.7152.5052.45
CD-IR-50 (Ours)82.4652.0851.58 (+1.04)79.8554.7952.35 (+1.34)85.1154.8153.36 (+0.91)
L0-IR-50 (Ours)82.1752.4152.19 (+1.65)79.9054.9052.30 (+1.29)85.1854.5353.34 (+0.89)
", + "image_path": "14f7c573807a76a780780ce2c066c5c0236ee456aeb81b7c0ae3e512610b105e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 247, + 479, + 305 + ], + "blocks": [ + { + "bbox": [ + 132, + 216, + 480, + 239 + ], + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 239 + ], + "type": "text", + "content": "Table 2: Evaluating our IG-Defense (CD-IR, L0-IR) on CIFAR10 with WRN-34-10 and " + }, + { + "bbox": [ + 132, + 216, + 480, + 239 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}, \\epsilon = 8/255" + }, + { + "bbox": [ + 132, + 216, + 480, + 239 + ], + "type": "text", + "content": ". IW-WC indicates image-wise worst-case, AA indicates AutoAttack." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 247, + 479, + 305 + ], + "lines": [ + { + "bbox": [ + 133, + 247, + 479, + 305 + ], + "spans": [ + { + "bbox": [ + 133, + 247, + 479, + 305 + ], + "type": "table", + "html": "
Base Model:NuAT [41] (WRN-34-10)TR-AWP [52] (WRN-34-10)DAJAT [1] (WRN-34-10)
Test-Time DefenseClean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)
None86.3254.7554.7285.3656.1756.1288.7157.8157.72
CD-IR-50 (Ours)86.2857.6956.38 (+1.66)84.9860.2657.46 (+1.34)88.5160.7059.22 (+1.50)
LO-IR-50 (Ours)86.2157.0456.40 (+1.68)85.1759.6157.88 (+1.76)88.6160.0859.41 (+1.69)
", + "image_path": "f5fa2c3e92e425c4062639bd4f7cbacb03c4efb9a51db5860d3a1269b95b853d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 328, + 481, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 481, + 497 + ], + "type": "text", + "content": "defenses like TRADES [52], AWP [44], DAJAT [1], RTB [37], and LAS [25]. We chose these models to ensure a diverse evaluation and showcase the generality of our IG-Defense across base model defense strategies. We also evaluate with the same base model on different architectures as well as on different datasets for a thorough analysis. Please see Appendix B.2 for more details on these defenses. Evaluation. We evaluate our IG-Defense with the base models against a diverse range of attacks. AutoAttack [18] (AA) is an empirically strong attack with a sequence of white-box AutoPGD (APGD) attacks, fast adaptive boundary (FAB) attack [17], and black-box Square attack [4]. For a stronger adaptive attack, we devise an ensemble of attacks. For each image, the adaptive attack is considered successful if any one attack from the ensemble is successful. With this, we get the image-wise worst-case (IW-WC) robust accuracy. To be adaptive for our IG-Defense, the ensemble also includes transfer attacks that specifically target our weakness, i.e. our dependence on pseudo-labels (as discussed in Sec. 4.1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Robust accuracy (IW-WC) is computed with 9 attacks. The first three are computed on the defended model: standard white-box AutoAttack, decision-boundary based black-box RayS attack [12], and white-box APGD with Expectation over Transformation (EoT) [5] attack. For the remaining 6 transfer attacks, we compute perturbations on the base model using 6 attacks: APGD with cross-entropy loss, APGD with Carlini-Wagner loss [11], targeted APGD attack with difference of logits ratio loss, standard AutoAttack, RayS attack, and " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{APGD + EoT}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " attack. Design choices for IW-WC. While AutoAttack also uses APGD attacks, it returns unperturbed inputs when the attack is unsuccessful. Following [16], the four transfer APGD attacks in IW-WC each return a sample with the highest loss (i.e. closest to the decision boundary). This is a stronger attack than AutoAttack for test-time defenses since samples closer to the decision boundary tend to be misclassified under test-time defenses, as observed by [16]. Expectation over Transformation (EoT) [5] averages gradients over multiple iterations of" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 286, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 286, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 286, + 102 + ], + "type": "text", + "content": "A. Kulkarni and T. W. Weng" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 146, + 479, + 204 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 3: Evaluating our IG-Defense (CD-IR, L0-IR) on CIFAR100 with " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}, \\epsilon = 8 / 255" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": ". IW-WC indicates image-wise worst-case, AA indicates AutoAttack." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 146, + 479, + 204 + ], + "lines": [ + { + "bbox": [ + 133, + 146, + 479, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 146, + 479, + 204 + ], + "type": "table", + "html": "
Base Model:LAS [25] (WRN-34-10)OAAT [2] (RN18)DAJAT [1] (RN18)
Test-Time DefenseClean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)
None64.8930.7430.7162.0227.1527.1265.4527.6727.64
CD-IR-50 (Ours)62.1333.5731.41 (+0.70)61.5629.5129.19 (+2.07)64.9730.2228.27 (+0.63)
LO-IR-50 (Ours)64.8535.6932.22 (+1.51)62.0432.0630.12 (+3.00)65.3732.5528.54 (+0.90)
", + "image_path": "6b967ed743821a571a10a045aca7740d6fe9a2086c52d0bc084f583f1506da5d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 247, + 479, + 313 + ], + "blocks": [ + { + "bbox": [ + 132, + 215, + 481, + 237 + ], + "lines": [ + { + "bbox": [ + 132, + 215, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 215, + 481, + 237 + ], + "type": "text", + "content": "Table 4: Evaluating our IG-Defense (CD-IR, L0-IR) on ImageNet-1k with ResNet50 and " + }, + { + "bbox": [ + 132, + 215, + 481, + 237 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}, \\epsilon = 4/255" + }, + { + "bbox": [ + 132, + 215, + 481, + 237 + ], + "type": "text", + "content": ". IW-WC indicates image-wise worst-case, AA indicates AutoAttack." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 247, + 479, + 313 + ], + "lines": [ + { + "bbox": [ + 133, + 247, + 479, + 313 + ], + "spans": [ + { + "bbox": [ + 133, + 247, + 479, + 313 + ], + "type": "table", + "html": "
Base Model:FAT [42] (RN50)RTB [37] (RN50)
Test-Time DefenseClean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)Clean Acc.Robust Acc. (AA)Robust Acc. (IW-WC)
None55.6426.2426.2464.1034.6234.62
CD-IR-750 (Ours)54.3027.28 (+1.04)26.80 (+0.56)62.5036.42 (+1.80)35.30 (+0.68)
LO-IR-750 (Ours)55.8829.14 (+2.90)27.08 (+0.84)64.1037.36 (+2.74)35.48 (+0.86)
", + "image_path": "4f2135ee7db0bfc7f6ceb7a972de2c4b6deed79260e5baf65b571ba1cd45ed49.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 337, + 482, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 482, + 409 + ], + "type": "text", + "content": "a particular input to counter the effect of randomness in a defense. Specifically, we use 20 iterations of EoT with APGD (more than " + }, + { + "bbox": [ + 130, + 337, + 482, + 409 + ], + "type": "inline_equation", + "content": "n_s = 1" + }, + { + "bbox": [ + 130, + 337, + 482, + 409 + ], + "type": "text", + "content": "). While [16] also recommend using Backward Pass Differentiable Approximation (BPDA) [5] for test-time defenses with non-differentiable components, our approach already includes BPDA since we use a soft-masking approach instead of hard-masking. Refer to Appendix B.3 for complete implementation details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 428, + 301, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 428, + 301, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 301, + 440 + ], + "type": "text", + "content": "5.1 Evaluation on RobustBench" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 449, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "text", + "content": "a) CIFAR10. In Table 1, we evaluate our methods on CIFAR10 with ResNet18 at " + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\epsilon = 8 / 255" + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "text", + "content": ". We observe consistent gains in AutoAttack (avg. " + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "inline_equation", + "content": "+2.5\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "text", + "content": " for CD-IR and " + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "inline_equation", + "content": "+2.55\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "text", + "content": " for L0-IR) and IW-WC robustness (avg. " + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "inline_equation", + "content": "+1.1\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "text", + "content": " for CD-IR and " + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "inline_equation", + "content": "+1.28\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "text", + "content": " for L0-IR), with a marginal drop in clean accuracy (avg. " + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "inline_equation", + "content": "-0.24\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "text", + "content": " for CD-IR and " + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "inline_equation", + "content": "-0.3\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "text", + "content": " for L0-IR). In Table 2, we evaluate on CIFAR10 with the more challenging, large WideResNet-34-10 model at " + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\epsilon = 8 / 255" + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "text", + "content": ". We observe consistent gains in IW-WC robustness (avg. " + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "inline_equation", + "content": "+1.5\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "text", + "content": " for CD-IR and " + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "inline_equation", + "content": "+1.7\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "text", + "content": " for L0-IR), with a marginal drop in clean accuracy (avg. " + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "inline_equation", + "content": "-0.2\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "text", + "content": " for CD-IR and " + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "inline_equation", + "content": "-0.13\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "text", + "content": " for L0-IR). b) CIFAR100. In Table 3, we evaluate our methods on CIFAR100 with ResNet18 and WideResNet-34-10 at " + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\epsilon = 8 / 255" + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "text", + "content": ". Despite the increased number of classes, we get consistently improved worst-case (IW-WC) robust accuracy (avg. " + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "inline_equation", + "content": "+1.13\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "text", + "content": " for CD-IR and " + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "inline_equation", + "content": "+1.8\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "text", + "content": " for L0-IR). However, the clean accuracy drops for CD-IR (avg. " + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "inline_equation", + "content": "-0.43\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 604 + ], + "type": "text", + "content": " drop), but only marginally for L0-IR." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "c) ImageNet-1k. In Table 4, we evaluate our methods on ImageNet-1k with ResNet50 at " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon = 4 / 255" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": ". Despite the increased image size and the higher number of classes, we observe consistent gains in worst-case (IW-WC) robustness (avg. " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "+0.6\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " for CD-IR and " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "+0.85\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " for L0-IR) but with a drop in clean accuracy for CD-IR (avg. " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "-1.47\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " similar to CIFAR100). Given the scale of ImageNet-1k, even" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 224, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 224, + 91, + 447, + 102 + ], + "type": "text", + "content": "Interpretability-Guided Test-Time Adversarial Defense" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 168, + 482, + 286 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 5: Comparison against state-of-the-art test-time defenses with ResNet18 and WideResNet-34-10 on CIFAR10. The number in (green) shows the worst-case robustness gain over the base model. The first row is the base model without any test-time defense. Inference time is a multiple of the time for a single forward pass of the base model." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 168, + 482, + 286 + ], + "lines": [ + { + "bbox": [ + 135, + 168, + 482, + 286 + ], + "spans": [ + { + "bbox": [ + 135, + 168, + 482, + 286 + ], + "type": "table", + "html": "
Base Model → Test-Time DefenseOAAT [2] (RN18)DAJAT [1] (RN18)TR-AWP [44] (WRN)Inference Time
Clean Acc.Robust Acc. (IW-WC)Clean Acc.Robust Acc. (IW-WC)Clean Acc.Robust Acc. (IW-WC)
No test-time def.80.2351.0185.7152.4585.3656.121×
Existing Defenses
HD [43]79.8950.68 (-0.33)84.5352.55 (+0.10)84.7855.72 (-0.40)46×
SODEF [26]80.2350.67 (-0.34)84.8652.95 (+0.50)85.2556.05 (-0.07)2×
CAAA [3]80.2351.00 (-0.01)85.7152.45 (+0.00)85.3556.11 (-0.01)8×
Our IG-Defense
CD-IR-5079.8552.35 (+1.34)85.1153.36 (+0.91)84.9857.46 (+1.34)2×
L0-IR-5079.9052.30 (+1.29)85.1853.34 (+0.89)85.1757.88 (+1.76)2×
", + "image_path": "8925e1082464aeead1743c17ba5a1526c47c660fd6eba9f13952da1b095924ff.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 480, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 480, + 331 + ], + "type": "text", + "content": "robustness improvements of " + }, + { + "bbox": [ + 130, + 308, + 480, + 331 + ], + "type": "inline_equation", + "content": "\\sim 0.5\\%" + }, + { + "bbox": [ + 130, + 308, + 480, + 331 + ], + "type": "text", + "content": " are quite significant, especially since there is no clean accuracy drop with LO-IR." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 331, + 481, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 481, + 426 + ], + "type": "text", + "content": "Why does the clean accuracy drop? There are two possible reasons for the drop in clean accuracy. First, due to the randomized smoothing, some samples may get incorrect predictions if the random noise is significant enough to change the class prediction w.r.t. the base model. Second, importance ranking quality and number of neurons masked determine how much clean accuracy can be retained. That is, better quality or lower number of neurons masked would guarantee better clean accuracy retention. However, simply masking lower number of neurons may also reduce the improvements in robustness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 427, + 481, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 481, + 522 + ], + "type": "text", + "content": "We see in Table 3, 4 that clean accuracy drop for CD-IR is more for CIFAR100 and ImageNet1k compared to CIFAR10 (Table 1, 2). This is possibly because the ratio of number of neurons to number of classes is decreasing, i.e. more number of neurons will be important to multiple classes. This cannot be captured well by CD-IR, leading to a larger drop in clean accuracy. Note that LO-IR is less susceptible to this problem as the importance ranking of each neuron is computed independently, but it is more expensive. This is why we need to retain more number of important neurons for ImageNet (" + }, + { + "bbox": [ + 130, + 427, + 481, + 522 + ], + "type": "inline_equation", + "content": "k = 750" + }, + { + "bbox": [ + 130, + 427, + 481, + 522 + ], + "type": "text", + "content": ") than CIFAR (" + }, + { + "bbox": [ + 130, + 427, + 481, + 522 + ], + "type": "inline_equation", + "content": "k = 50" + }, + { + "bbox": [ + 130, + 427, + 481, + 522 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 523, + 481, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 523, + 481, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 523, + 481, + 571 + ], + "type": "text", + "content": "Note that the clean accuracy drops are acceptable since they are significantly smaller than the worst-case robustness gains (e.g. upto " + }, + { + "bbox": [ + 130, + 523, + 481, + 571 + ], + "type": "inline_equation", + "content": "-0.6\\%" + }, + { + "bbox": [ + 130, + 523, + 481, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 523, + 481, + 571 + ], + "type": "inline_equation", + "content": "+1.7\\%" + }, + { + "bbox": [ + 130, + 523, + 481, + 571 + ], + "type": "text", + "content": " respectively for CIFAR10 in Table 1-2). Further, in some cases like ImageNet-1k with LO-IR (Table 4), there are no drops with pure gains in robustness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 586, + 388, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 388, + 600 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 388, + 600 + ], + "type": "text", + "content": "5.2 Comparisons with existing test-time defenses" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "We compare with three recent state-of-the-art test-time defense methods on CIFAR10 in Table 5. CAAA [3] (input purification) optimizes an \"anti-adversary\" perturbation in a direction opposite to an adversarial attack, by minimizing the loss based on a pseudo-label, to counter the effect of any adversarial perturbation. HD or Hedge Defense [43] (input purification) finds a perturbation that maximizes" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 286, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 286, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 286, + 102 + ], + "type": "text", + "content": "A. Kulkarni and T. W. Weng" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 117, + 216, + 175 + ], + "blocks": [ + { + "bbox": [ + 136, + 117, + 216, + 175 + ], + "lines": [ + { + "bbox": [ + 136, + 117, + 216, + 175 + ], + "spans": [ + { + "bbox": [ + 136, + 117, + 216, + 175 + ], + "type": "image", + "image_path": "3ac288f046bda527dedb18a14cc180e693373db2dc20a52eeaca6b5fec58b3f3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 187, + 482, + 221 + ], + "lines": [ + { + "bbox": [ + 130, + 187, + 482, + 221 + ], + "spans": [ + { + "bbox": [ + 130, + 187, + 482, + 221 + ], + "type": "text", + "content": "Fig. 5: Sensitivity analyses for number of retained important neurons " + }, + { + "bbox": [ + 130, + 187, + 482, + 221 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 187, + 482, + 221 + ], + "type": "text", + "content": " and layer being masked with CIFAR10. For A, B: layer4 with total 512 neurons is used. A lower " + }, + { + "bbox": [ + 130, + 187, + 482, + 221 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 187, + 482, + 221 + ], + "type": "text", + "content": " and a deeper layer yield better adversarial robustness with minimal loss of clean accuracy." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 118, + 302, + 175 + ], + "blocks": [ + { + "bbox": [ + 223, + 118, + 302, + 175 + ], + "lines": [ + { + "bbox": [ + 223, + 118, + 302, + 175 + ], + "spans": [ + { + "bbox": [ + 223, + 118, + 302, + 175 + ], + "type": "image", + "image_path": "bc7d75d5851e386c8a9880951ca38dcf33b9598a2a7107a8081763c6bf634b2c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 306, + 118, + 390, + 175 + ], + "blocks": [ + { + "bbox": [ + 306, + 118, + 390, + 175 + ], + "lines": [ + { + "bbox": [ + 306, + 118, + 390, + 175 + ], + "spans": [ + { + "bbox": [ + 306, + 118, + 390, + 175 + ], + "type": "image", + "image_path": "417d66132b5d41f9462b058afbd7f54e7b033632bde7a0260337b9717d4cc383.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 392, + 118, + 477, + 175 + ], + "blocks": [ + { + "bbox": [ + 392, + 118, + 477, + 175 + ], + "lines": [ + { + "bbox": [ + 392, + 118, + 477, + 175 + ], + "spans": [ + { + "bbox": [ + 392, + 118, + 477, + 175 + ], + "type": "image", + "image_path": "74948447b5099fe5443435a8c2c52fd23e911e74e10b22fb195dfee967cba7ae.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 243, + 479, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 479, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 479, + 303 + ], + "type": "text", + "content": "the cross-entropy loss summed over all classes. SODEF [26] (model adaptation) trains a neural ODE block with Lyapunov stability constraints between the pretrained feature extractor and classifier, which reduces the influence of the adversarial perturbation. In contrast, our method uses neither input purification nor model adaptation, since we only mask the activations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 304, + 482, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 448 + ], + "type": "text", + "content": "In Table 5, we observe that existing test-time defenses like HD, CAAA, and SODEF do not significantly improve, but rather also reduce the worst-case robust accuracy (IW-WC) compared to the base model. Based on Croce et al. [16], CAAA is the strongest test-time defense among the nine defenses that they evaluated, since its worst-case robust accuracy did not drop below that of the base model. Hence, we only compare with the two strongest input purification defenses, CAAA and HD as well as an efficient model adaptation defense, SODEF from the nine defenses in [16]. Overall, we find that our proposed IG-Defense obtains consistent improvements in image-wise worst-case (IW-WC) robust accuracy, unlike existing test-time defenses. Note that we focus mainly on the worst-case (IW-WC) robust accuracy because gains under weaker attacks would not reflect fundamental gains in robustness [16]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 464, + 204, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 204, + 477 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 204, + 477 + ], + "type": "text", + "content": "5.3 Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 483, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 665 + ], + "type": "text", + "content": "a) Ablation study. In Table 6, we study the significance of each proposed component of our approach. First, we start with the baseline (row #1) which uses only one forward pass without masking and without randomized smoothing (RS). Next, we add RS to this baseline (row #2) and observe no gains in robustness and a small drop in clean accuracy. This shows that RS alone cannot improve the robustness. Then, we perform two forward passes without masking and without RS (row #3). Here, the robust accuracy is the same as the baseline indicating that there is no gradient masking with two forward passes. Next, we perform random masking with two forward passes, without and with RS (row #4, #5). Each class is assigned a random set of neurons as the \"important\" neurons to perform the masking. We observe no gains in robustness, validating the significance of our neuron importance ranking. Finally, we show the impact of RS with our CD-IR (row #6, #7) and L0-IR based masking (row #8, #9). Here, we observe consistent gains in robustness, with a small drop in clean accuracy since RS likely changed the prediction w.r.t. the base model for a few samples." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 224, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 224, + 91, + 448, + 102 + ], + "type": "text", + "content": "Interpretability-Guided Test-Time Adversarial Defense" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 247, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 247, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 247, + 307 + ], + "type": "text", + "content": "b) Sensitivity analyses. We study the sensitivity of our proposed methods to the number of retained important neurons " + }, + { + "bbox": [ + 130, + 116, + 247, + 307 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 247, + 307 + ], + "type": "text", + "content": " in Fig. 5A, B on CIFAR10. Specifically, we test it on ResNet18 layer4 which has total 512 neurons and we vary " + }, + { + "bbox": [ + 130, + 116, + 247, + 307 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 247, + 307 + ], + "type": "text", + "content": " from 10 to 500. We observe that the robustness improvement is higher at lower " + }, + { + "bbox": [ + 130, + 116, + 247, + 307 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 247, + 307 + ], + "type": "text", + "content": ", although lower than 10 may reduce the clean accuracy due to a high loss" + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 253, + 184, + 480, + 293 + ], + "blocks": [ + { + "bbox": [ + 252, + 129, + 482, + 183 + ], + "lines": [ + { + "bbox": [ + 252, + 129, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 252, + 129, + 482, + 183 + ], + "type": "text", + "content": "Table 6: Ablation study of proposed components with TRADES-AWP [44] (WideResNet-34-10) as the base model. RS indicates use of randomized smoothing in the first forward pass. The first row is the base model performance without any test-time defense." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 253, + 184, + 480, + 293 + ], + "lines": [ + { + "bbox": [ + 253, + 184, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 253, + 184, + 480, + 293 + ], + "type": "table", + "html": "
Row #No. of forw. passesMasking typeRSClean Acc.Rob. Acc. (AA)Rob. Acc. (IW-WC)
11None✗85.3656.1756.12
21None✓84.9356.1956.14
32None✗85.3556.1556.12
42Random-50✗85.2356.1456.12
52Random-50✓85.3356.2056.15
62CD-IR-50✗85.0959.0056.12
72CD-IR-50✓84.9860.2657.46
82LO-IR-50✗85.3358.5956.12
92LO-IR-50✓85.1759.6157.88
", + "image_path": "e8605505bd2101aacb401971b0fab44204ac6139ec848e1bd144174281df1d22.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 308, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 354 + ], + "type": "text", + "content": "of information. For CD-IR, the clean accuracy can be retained up to " + }, + { + "bbox": [ + 130, + 308, + 482, + 354 + ], + "type": "inline_equation", + "content": "k = 100" + }, + { + "bbox": [ + 130, + 308, + 482, + 354 + ], + "type": "text", + "content": " while for L0-IR, it can be retained up to at least " + }, + { + "bbox": [ + 130, + 308, + 482, + 354 + ], + "type": "inline_equation", + "content": "k = 400" + }, + { + "bbox": [ + 130, + 308, + 482, + 354 + ], + "type": "text", + "content": ". As explained earlier, it is due to the relatively lower importance ranking quality of CD-IR w.r.t. L0-IR. However, both methods are fairly insensitive to the choice of " + }, + { + "bbox": [ + 130, + 308, + 482, + 354 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 308, + 482, + 354 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "type": "text", + "content": "We also study the sensitivity to choice of layer being masked in Fig. 5C, D on CIFAR10 with ResNet18. We fix the number of neurons being retained " + }, + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "type": "text", + "content": " to 50 and vary the layer being masked from layer1 (shallower) to layer4 (deeper). We observe that masking deeper layers is better, since masking shallower layers affects the feature extraction [51], leading to a large drop in clean accuracy, and consequently a drop in robust accuracy. For all the other experiments, we simply mask the penultimate layer since the clean accuracy drop is the lowest. Please refer to Appendix C for more analysis experiments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 453, + 483, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 483, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 483, + 536 + ], + "type": "text", + "content": "c) Inference time comparisons. Since we use " + }, + { + "bbox": [ + 130, + 453, + 483, + 536 + ], + "type": "inline_equation", + "content": "n_s = 1" + }, + { + "bbox": [ + 130, + 453, + 483, + 536 + ], + "type": "text", + "content": " in randomized smoothing and perform two forward passes, the inference time of IG-Defense is " + }, + { + "bbox": [ + 130, + 453, + 483, + 536 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 130, + 453, + 483, + 536 + ], + "type": "text", + "content": " of a single forward pass. The compared test-time defenses HD [43], SODEF [26], and CAAA [3] have computational complexities of " + }, + { + "bbox": [ + 130, + 453, + 483, + 536 + ], + "type": "inline_equation", + "content": "46 \\times" + }, + { + "bbox": [ + 130, + 453, + 483, + 536 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 453, + 483, + 536 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 130, + 453, + 483, + 536 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 453, + 483, + 536 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 130, + 453, + 483, + 536 + ], + "type": "text", + "content": " respectively. Also, the recent review paper [16] shows that the lowest inference time is " + }, + { + "bbox": [ + 130, + 453, + 483, + 536 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 130, + 453, + 483, + 536 + ], + "type": "text", + "content": " for existing test-time defenses. Hence, as per Table 5, we achieve the most gains in worst-case robustness with the lowest inference time." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 562, + 220, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 220, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 220, + 574 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "In this work, we proposed a novel, training-free, and effective test-time defense (IG-Defense) guided by interpretability through neuron importance ranking. Our IG-Defense outperformed existing test-time defenses on worst-case adaptive attacks while being among the most efficient test-time defenses to date, establishing the new state-of-the-art. We also demonstrated consistent improvements across standard CIFAR and ImageNet-1k benchmarks." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 286, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 286, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 286, + 101 + ], + "type": "text", + "content": "A. Kulkarni and T. W. Weng" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": "This work is supported in part by National Science Foundation (NSF) awards CNS-1730158, ACI-1540112, ACI1541349, OAC-1826967, OAC-2112167, CNS-2100237, CNS-2120019, the University of California Office of the President, and the University of California San Diego's California Institute for Telecommunications and Information Technology/Qualcomm Institute. Thanks to CENIC for the 100Gbps networks. The authors thank the anonymous reviewers for valuable feedback on the manuscript. T.-W. Weng is supported by National Science Foundation under Grant No. 2107189 and 2313105. T.-W. Weng also thanks the Hellman Fellowship for providing research support." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 266, + 197, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 266, + 197, + 278 + ], + "spans": [ + { + "bbox": [ + 133, + 266, + 197, + 278 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 291, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 291, + 480, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 291, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 138, + 291, + 480, + 313 + ], + "type": "text", + "content": "1. Addepalli, S., Jain, S., Babu, R.V.: Efficient and effective augmentation strategy for adversarial training. In: NeurIPS (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 314, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 314, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 138, + 314, + 480, + 335 + ], + "type": "text", + "content": "2. Addepalli, S., Jain, S., Sriramananan, G., Venkatesh Babu, R.: Scaling adversarial training to large perturbation bounds. In: ECCV (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 336, + 480, + 357 + ], + "type": "text", + "content": "3. Alfarra, M., Pérez, J.C., Thabet, A., Bibi, A., Torr, P.H.S., Ghanem, B.: Combating adversaries with anti-adversaries. In: AAAI (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 357, + 481, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 481, + 379 + ], + "type": "text", + "content": "4. Andriushchenko, M., Croce, F., Flammarion, N., Hein, M.: Square attack: a query-efficient black-box adversarial attack via random search. In: ECCV (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 380, + 480, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 401 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 401 + ], + "type": "text", + "content": "5. Athalye, A., Carlini, N., Wagner, D.: Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In: ICML (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 402, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 402, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 402, + 480, + 434 + ], + "type": "text", + "content": "6. Bai, N., Iyer, R.A., Oikarinen, T., Weng, T.W.: Describe-and-dissect: Interpreting neurons in vision networks with language models. arXiv preprint arXiv:2403.13771 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 434, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 480, + 456 + ], + "type": "text", + "content": "7. Bai, Y., Zeng, Y., Jiang, Y., Xia, S.T., Ma, X., Wang, Y.: Improving adversarial robustness via channel-wise activation suppressing. In: ICLR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 457, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 480, + 478 + ], + "type": "text", + "content": "8. Bau, D., Zhou, B., Khosla, A., Oliva, A., Torralba, A.: Network dissection: Quantifying interpretability of deep visual representations. In: CVPR (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "type": "text", + "content": "9. Bau, D., Zhu, J.Y., Strobelt, H., Lapedriza, A., Zhou, B., Torralba, A.: Understanding the role of individual units in a deep neural network. Proceedings of the National Academy of Sciences (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "text", + "content": "10. Boopathy, A., Liu, S., Zhang, G., Liu, C., Chen, P.Y., Chang, S., Daniel, L.: Proper network interpretability helps adversarial robustness in classification. In: ICML (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 544, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 566 + ], + "type": "text", + "content": "1.1. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: IEEE Symposium on Security and Privacy (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "text", + "content": "2. Chen, J., Gu, Q.: RayS: A ray searching method for hard-label adversarial attack In: KDD (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "type": "text", + "content": "3. Chen, Y., Oliver, D.S.: Ensemble randomized maximum likelihood method as an iterative ensemble smoother. Mathematical Geosciences 44, 1-26 (2012)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "text", + "content": "4. Chen, Z., Li, Q., Zhang, Z.: Towards robust neural networks via close-loop control In: ICLR (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "5. Croce, F., Andriushchenko, M., Sehwag, V., Debenedetti, E., Flammarion, N., Chiang, M., Mittal, P., Hein, M.: Robustbench: a standardized adversarial robustness benchmark. In: Datasets and Benchmarks Track, NeurIPS (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 224, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 224, + 91, + 447, + 102 + ], + "type": "text", + "content": "Interpretability-Guided Test-Time Adversarial Defense" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "16. Croce, F., Gowal, S., Brunner, T., Shelhamer, E., Hein, M., Cemgil, T.: Evaluating the adversarial robustness of adaptive test-time defenses. In: ICML (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "17. Croce, F., Hein, M.: Minimally distorted adversarial examples with a fast adaptive boundary attack. In: ICML (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "18. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: ICML (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "19. Eigen, H., Sadovnik, A.: TopKConv: Increased adversarial robustness through deeper interpretability. In: ICMLA (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "20. Gerasimou, S., Eniser, H.F., Sen, A., Cakan, A.: Importance-driven deep learning system testing. In: ACM/IEEE ICSE (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "21. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: ICLR (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "22. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition In: CVPR (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "text", + "content": "23. Hernandez, E., Schwettmann, S., Bau, D., Bagashvili, T., Torralba, A., Andreas, J.: Natural language descriptions of deep visual features. In: ICLR (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "24. Hwang, D., Lee, E., Rhee, W.: AID-Purifier: A light auxiliary network for boosting adversarial defense. In: ICPR (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "25. Jia, X., Zhang, Y., Wu, B., Ma, K., Wang, J., Cao, X.: LAS-AT: adversarial training with learnable attack strategy. In: CVPR (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "26. Kang, Q., Song, Y., Ding, Q., Tay, W.P.: Stable neural ODE with Lyapunov-stable equilibrium points for defending against adversarial attacks. In: NeurIPS (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "27. Krizhevsky, A.: Learning multiple layers of features from tiny images. Tech. rep., CIFAR (2009)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "28. Kundu, S., Nazemi, M., Beerel, P.A., Pedram, M.: A tunable robust pruning framework through dynamic network rewiring of DNNs. In: ASP-DAC (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "29. Mangla, P., Singh, V., Balasubramanian, V.N.: On saliency maps and adversarial robustness. In: ECML-PKDD (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "text", + "content": "30. Mao, C., Chiquier, M., Wang, H., Yang, J., Vondrick, C.: Adversarial attacks are reversible with natural supervision. In: ICCV (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 447, + 473, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 473, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 473, + 456 + ], + "type": "text", + "content": "31. Mu, J., Andreas, J.: Compositional explanations of neurons. In: NeurIPS (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "type": "text", + "content": "32. Nayak, G.K., Rawal, R., Chakraborty, A.: DAD: Data-free adversarial defense at test time. In: WACV (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "text", + "content": "33. Oikarinen, T., Weng, T.W.: CLIP-Dissect: Automatic description of neuron representations in deep vision networks. In: ICLR (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "text", + "content": "34. Oikarinen, T., Weng, T.W.: Linear explanations for individual neurons. In: ICML (2024)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "text", + "content": "35. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "type": "text", + "content": "36. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) 115(3), 211-252 (2015)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 600, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 620 + ], + "type": "text", + "content": "37. Salman, H., Ilyas, A., Engstrom, L., Kapoor, A., Madry, A.: Do adversarially robust imagenet models transfer better? In: NeurIPS (2020)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "text", + "content": "38. Sehwag, V., Wang, S., Mittal, P., Jana, S.: Hydra: Pruning adversarially robust neural networks. In: NeurIPS (2020)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "39. Shi, C., Holtz, C., Mishne, G.: Online adversarial purification based on self-supervised learning. In: ICLR (2021)" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 287, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 287, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 287, + 102 + ], + "type": "text", + "content": "A. Kulkarni and T. W. Weng" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 424 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "40. Sriramanan, G., Addepalli, S., Baburaj, A., Babu, R.V.: Guided adversarial attack for evaluating and enhancing adversarial defenses. In: NeurIPS (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "text", + "content": "41. Sriramanan, G., Addepalli, S., Baburaj, A., Babu, R.V.: Towards efficient and effective adversarial training. In: NeurIPS (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "text", + "content": "42. Wong, E., Rice, L., Kolter, J.Z.: Fast is better than free: Revisiting adversarial training. In: ICLR (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "43. Wu, B., Pan, H., Shen, L., Gu, J., Zhao, S., Li, Z., Cai, D., He, X., Liu, W.: Attacking adversarial attacks as a defense. arXiv preprint arXiv:2106.04938 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 225 + ], + "type": "text", + "content": "44. Wu, D., Xia, S.T., Wang, Y.: Adversarial weight perturbation helps robust generalization. In: NeurIPS (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "45. Wu, S., Sang, J., Xu, K., Zhang, J., Yu, J.: Attention, please! adversarial defense via activation rectification and preservation. ACM Transactions on Multimedia Computing, Communications and Applications 19(4), 1-18 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "46. Wu, Y.H., Yuan, C.H., Wu, S.H.: Adversarial robustness via runtime masking and cleansing. In: ICML (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "text", + "content": "47. Xiao, C., Zhong, P., Zheng, C.: Enhancing adversarial defense by k-winners-take-all In: ICLR (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "48. Xie, X., Li, T., Wang, J., Ma, L., Guo, Q., Yuefei-Xu, F., Liu, Y.: NPC: Neuron path coverage via characterizing decision logic of deep neural networks. ACM Transactions on Software Engineering and Methodology 31(3) (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "49. Yoon, J., Hwang, S.J., Lee, J.: Adversarial purification with score-based generative models. In: ICML (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 358, + 448, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 448, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 448, + 369 + ], + "type": "text", + "content": "50. Zagoruyko, S., Komodakis, N.: Wide residual networks. In: BMVC (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 369, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 390 + ], + "type": "text", + "content": "51. Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: ECCV (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "type": "text", + "content": "52. Zhang, H., Yu, Y., Jiao, J., Xing, E.P., Ghaoui, L.E., Jordan, M.I.: Theoretically principled trade-off between robustness and accuracy. In: ICML (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 413, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 424 + ], + "type": "text", + "content": "53. Zhao, Q., Wressnegger, C.: Holistic adversarially robust pruning. In: ICLR (2023)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "text", + "content": "Interpretability-Guided Test-Time Adversarial Defense" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Intrinsic Single-Image HDR Reconstruction/e55f56f5-9c00-4140-aab7-15323b235acd_content_list.json b/2024/Intrinsic Single-Image HDR Reconstruction/e55f56f5-9c00-4140-aab7-15323b235acd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..69b13dfae4a5385448ba00b4b32096317a3c7cf6 --- /dev/null +++ b/2024/Intrinsic Single-Image HDR Reconstruction/e55f56f5-9c00-4140-aab7-15323b235acd_content_list.json @@ -0,0 +1,2170 @@ +[ + { + "type": "text", + "text": "Intrinsic Single-Image HDR Reconstruction", + "text_level": 1, + "bbox": [ + 251, + 141, + 751, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sebastian Dille\\* , Chris Careaga\\* , and Yaigiz Aksoy", + "bbox": [ + 292, + 189, + 710, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Simon Fraser University Burnaby, BC V5A 1S6, Canada sdille@sfu.ca", + "bbox": [ + 393, + 217, + 609, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b94aba13b10ed44f5cd7842782273070a4e0b5f3b0f4da4bd2bfc82730171a53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 232, + 273, + 380, + 352 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b88693e2b9c97e4dfc39d2328f535fc12116cd4a23d641c13b60fecb3d2bf424.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 357, + 359, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bd1578a4ae706bc6d0c677422a09c43cf1b0444bfd5cc81ff7cbee8b3f31f08d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 273, + 500, + 395 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9d3fd8a0ef2fa1af403dd8aa883eebc509b42915a2b141639b77fce4939408f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 273, + 617, + 395 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c081a015757e0cb57756c84b3da7b4f32262419542c417dab40e292d7a1c8982.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 273, + 771, + 352 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/76c62823aae19e48d7a890e5542c343b2e24f385e03a5a5f49b965d2f4a81444.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 357, + 751, + 395 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e000213ad39461b9b479c23ae02d2b0756e7b816cb53c004ea627184c328d587.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 397, + 364, + 468 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/15c9daa8008848db1321d6a6d791ed07a2279d495f9d797b75ff39f211f88b79.jpg", + "image_caption": [ + "Fig. 1: In this work, we propose a physically-inspired HDR reconstruction pipeline by conducting reconstruction individually on intrinsic components. Our final reconstruction, as a result, is effective in recovering the high-luminance details and the lost colors." + ], + "image_footnote": [], + "bbox": [ + 230, + 469, + 364, + 497 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e4b7e6972ca5de426dd1bec7af395c005798d28b324822a919b5174eb36214c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 397, + 500, + 468 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/87b751f9e5da6c0d1fcda4f0f9eb51785f4015a147b659f8352c1794c9e07103.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 469, + 500, + 497 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7f8ea3dcfb895c1bd0381bd828404406f3daa497d63797c8d1b67261dc255126.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 397, + 635, + 468 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/24ba8bd9a7bfadc70539f9bfe5fcbfeef396f180752655a2f1112d615f475773.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 469, + 635, + 497 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/135f1be94acde2799ce37e6399e2feb683f080196b5c53342568a8f1ae22d11a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 397, + 769, + 468 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ccda3e24fb4683a5a0dd6c7fdb70494104b6b770a7d6bd6b68e48904339b34e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 469, + 769, + 497 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The low dynamic range (LDR) of common cameras fails to capture the rich contrast in natural scenes, resulting in loss of color and details in saturated pixels. Reconstructing the high dynamic range (HDR) of luminance present in the scene from single LDR photographs is an important task with many applications in computational photography and realistic display of images. The HDR reconstruction task aims to infer the lost details using the context present in the scene, requiring neural networks to understand high-level geometric and illumination cues. This makes it challenging for data-driven algorithms to generate accurate and high-resolution results. In this work, we introduce a physically-inspired remodeling of the HDR reconstruction problem in the intrinsic domain. The intrinsic model allows us to train separate networks to extend the dynamic range in the shading domain and to recover lost color details in the albedo domain. We show that dividing the problem into two simpler sub-tasks improves performance in a wide variety of photographs.", + "bbox": [ + 259, + 571, + 740, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: HDR Reconstruction $\\cdot$ Inverse Tonemapping $\\cdot$ Computational Photography", + "bbox": [ + 259, + 792, + 740, + 820 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$(^{*})$ denotes equal contribution.", + "bbox": [ + 232, + 825, + 439, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 374, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Real-world scenes contain a wider range of luminance than most consumer-grade cameras can record or store. As a result, everyday photographs can only represent a low dynamic range (LDR) version of the scene that lacks details and color in very dark or bright regions. High dynamic range (HDR) imaging aims to capture and store the full contrast present in the scene. HDR images allow realistic display of photographs using high-contrast displays as well as many computational photography applications such as tone mapping and illumination-aware image editing. However, HDR imaging typically requires multi-exposure image stacks which limits their applications to static scenes.", + "bbox": [ + 212, + 175, + 787, + 310 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To extend the HDR-enabled applications to everyday photographs, there has been a growing interest in methods to estimate HDR images from only LDR photographs called single-image HDR reconstruction, or LDR-to-HDR conversion. Single-image HDR reconstruction aims to extend the dynamic range of LDR photographs by recovering the details lost during capture. The missing information in LDR images is typically characterized by saturated pixels in bright regions, lacking textural details and color. This makes HDR reconstruction a high-level problem, requiring networks to utilize geometric and context-dependent cues in the scene to infer the lost information. Formulating this reconstruction problem using standard color spaces such as RGB or Luv requires the networks to statistically model the complex relationship between luminance, surface colors, and geometry, which limits their performance.", + "bbox": [ + 212, + 311, + 787, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present a physically-inspired remodeling of the HDR reconstruction problem in the intrinsic domain. Intrinsic image decomposition is a mid-level vision problem that separates the effect of the illumination in the scene from the reflectance of surfaces [3], representing the input as a combination of a shading and an albedo layer. The shading of an image is highly correlated with the scene geometry and illumination, while the albedo contains the color information and the high-frequency textures. The saturated pixels we aim to recover create different degradations in the shading and albedo layers. The clipping of high values is observed in the shading, while the loss of color information causes desaturation in the albedo. We propose an HDR reconstruction pipeline that individually addresses the loss of information in the LDR image in the two intrinsic components. Our HDR shading reconstruction network extends the dynamic range, taking advantage of the high correlation between shading and the scene geometry. Our LDR albedo reconstruction network, on the other hand, recovers the lost color information in saturated regions. A final refinement stage combines the two components to generate our full HDR reconstruction. Figure 1 shows different stages of our pipeline.", + "bbox": [ + 212, + 492, + 787, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our formulation, we take advantage of the intrinsic image formation model by dividing the recovery of lost details and colors into two separate tasks which are more convenient for networks to model. We show that our physically inspired approach can generate high-fidelity details with rich colors in complex in-the-wild scenes. We evaluate our method qualitatively and quantitatively to demonstrate the state-of-the-art performance of our formulation in a wide variety of scenes.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Dille et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 143, + 382, + 160 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Common imaging sensors for photography are inherently limited in the portion of the scene's dynamic range they can capture in a single shot. This results in saturated pixels in bright regions, representing a significant challenge in the capture and realistic display of real-world scenes.", + "bbox": [ + 212, + 185, + 782, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-exposure fusion methods can reconstruct an HDR image by combining a stack of LDR images taken at different exposures [9,26,40]. Their formulation successfully addresses contrast range extension and reversal of the quantization and nonlinearity of LDR images. The multi-exposure capture, however, assumes a static camera and scene, requiring a challenging alignment process and limiting their application scenario. To enable HDR-related applications in everyday photography, single-image HDR reconstruction approaches aim to recover the lost information directly from an LDR image. Earlier approaches apply specially designed filters [18,34] or median-cut [1,8] to fill in overexposed regions, using different heuristics to scale the intensity of a selected area. Similarly, Bayesian [41] and graph-based formulations [39] have been introduced for chroma restoration. Due to the high-level nature of the reconstruction problem, these low-level approaches struggle to generate realistic details in real-world scenarios.", + "bbox": [ + 212, + 247, + 784, + 443 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In their concurrent publications, Eilertsen et al. [10] and Endo et al. [11] present the first data-driven approaches to the single-image HDR reconstruction problem. They train their networks for direct conversion to HDR from the LDR input, addressing quantization, nonlinearity, and saturation in the LDR image in a single step. Many works in recent literature adapt this unified inverse model [24] and develop tailored solutions for underexposed [6,14] or overexposed [36] areas or losses for color recovery [30]. The direct estimation of HDR outputs, however, creates the challenge of an unconstrained color distribution that hinders training efficiency. Several methods [5, 11, 17, 19, 20, 40] instead adopt the concept of exposure stacks and predict a set of over- and underexposed images from the input LDR image. With each prediction still in the LDR domain, the HDR image can later be created with multi-exposure fusion [9]. Other methods [38,42] employ generative approaches, either generating separate LDR detail layers [42] or full HDR images directly [38], with image-specific HDR reconstruction via GAN inversion. Liu et al. [24] approach HDR reconstruction by inverting the in-camera image formation process, modeling the dequantization, nonlinearization, and saturated pixel reconstruction in three separate steps. They show that this decomposition of the task simplifies the modeling of each individual sub-problem.", + "bbox": [ + 212, + 445, + 785, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we adopt the modular formulation of Liu et al. [24] and focus on recovering the lost details and color information in the saturated pixels. Our method takes inspiration from prior works that formulate computational photography problems in the intrinsic image domain [4, 29]. Rather than defining a direct mapping from LDR to HDR, we divide this task into two sub-problems: shading reconstruction that extends the dynamic range, and reflectance reconstruction that recovers the color information. Our physical deconstruction of the problem significantly improves the recovery of details and color information.", + "bbox": [ + 212, + 718, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Intrinsic Single-Image HDR Reconstruction", + "bbox": [ + 441, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d25827af201bb1f1214a21f93d29b01689f538979860505e0ca505e8b85648b7.jpg", + "image_caption": [ + "Fig. 2: The in-camera pipeline modeling LDR image formation is shown at the top. Liu et al. [24] propose the inverse pipeline shown on the bottom, addressing each type of information loss individually in three steps. In this work, we use the pre-trained networks of Liu et al. [24] for the dequantization and linearization tasks (shown in pink) and focus on developing an intrinsic model for the dynamic range reconstruction problem highlighted in teal." + ], + "image_footnote": [], + "bbox": [ + 215, + 146, + 782, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Background and Motivation", + "text_level": 1, + "bbox": [ + 215, + 372, + 524, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The focus of our work is the formulation of the single-image LDR-to-HDR reconstruction problem in the intrinsic domain. We begin our discussion by first introducing the LDR imaging pipeline and the intrinsic model of image formation. We then draw connections between the two models and motivate our intrinsic HDR formulation.", + "bbox": [ + 212, + 407, + 787, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 LDR image formation and HDR reconstruction", + "text_level": 1, + "bbox": [ + 214, + 508, + 653, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The acquisition of standard RGB photographs involves a series of operations in the camera-internal image processing pipeline, or ISP, that result in different forms of information loss. The initial acquisition of the image is in a raw format, where the luminance received by the sensor from a real-world scene is scaled and stored in 16- or 32-bit data structures. The scale of the captured raw values is determined by the analog camera settings such as the shutter speed, the aperture size, and the ISO. Given the natural high dynamic range of real-world scenes, the raw values are clipped at a maximum due to saturation at the sensor level. As shown in Figure 2, the remainder of the ISP applies nonlinear color and gamma corrections before finally decreasing the bit rate to 8 bits per channel for storage.", + "bbox": [ + 212, + 537, + 787, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The LDR image formation process causes three different types of corruption of the HDR information: quantization, non-linearization, and clipping. Single-image HDR reconstruction methods, as a result, have to address all three of these challenges. While many methods in the literature address all three simultaneously within a single architecture, Liu et al. [24] design their pipeline to address each type of information loss with a separate dedicated network. They propose a dequantization network to first increase the bit depth of the input LDR image and then estimate a CRF curve for linearization. Both tasks are convenient for a network to learn, taking advantage of the ease of dataset generation for dequantization and a simple parametric curve estimation for linearization.", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Dille et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/454f9ce86cdedfe2e9162afe62b8df247d96ff76fb29929baade7e7384a17aaf.jpg", + "image_caption": [ + "Fig. 3: We show the different effects of clipping on the albedo and the shading for a tonemapped HDR image and its LDR counterpart. When clipping the image, the LDR shading loses information compared to HDR. We scale the intensity of both shadings to show the effect. At the same time, the LDR albedo becomes saturated." + ], + "image_footnote": [], + "bbox": [ + 217, + 148, + 787, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "They then reconstruct the values lost to clipping with a third network in what they call the hallucination stage. This last stage of the inverse pipeline addresses the most significant challenge in single-image HDR reconstruction, namely the recovery of the highlights and the color information in clipped regions. Diverging from Liu et al.'s terminology, we refer to this stage as dynamic range extension. We adopt Liu et al.'s inverse pipeline and use their pre-trained networks for dequantization and linearization. In the following, we focus on the dynamic range extension problem and present our solution using an intrinsic model of image formation.", + "bbox": [ + 212, + 406, + 787, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Intrinsic image decomposition", + "text_level": 1, + "bbox": [ + 215, + 564, + 509, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intrinsic image decomposition is a fundamental mid-level computer vision problem that represents a core step in inverse rendering pipelines [2, 12]. It aims to separate the image into illumination and reflectance components following the physical image formation model. Adopting the commonly used Lambertian shading assumption with a single-channel shading representation [3], the intrinsic equation can be written as:", + "bbox": [ + 212, + 589, + 787, + 679 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {r g b} = A _ {r g b} \\cdot S, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 691, + 785, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I$ represents the linear image reflecting relative luminance, $A$ represents albedo, i.e. the surface reflectance containing the color information, and $S$ represents shading reflecting the effect of illumination in the scene.", + "bbox": [ + 212, + 719, + 784, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Shading and albedo represent two orthogonal characteristics of a scene, one containing the effect of light and its interaction with the scene geometry and the other representing material properties. As a result, the two intrinsic components have very different characteristics as demonstrated in Figure 3. Shading is highly correlated with surface orientations, appearing smooth over continuous surfaces", + "bbox": [ + 212, + 765, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Intrinsic Single-Image HDR Reconstruction", + "bbox": [ + 441, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and having a high gradient in depth discontinuities and shadow boundaries. Albedo, on the other hand, contains the high-frequency textures present on the materials in the scene while having a sparse set of values per image as the variation in brightness due to geometry is isolated in the shading layer. In the single-channel shading representation, albedo also contains any variation caused by colorful illumination [3].", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Infrinsics in LDR and HDR", + "text_level": 1, + "bbox": [ + 215, + 273, + 493, + 289 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The bounded representation of LDR images limits the range of luminance values they can contain. HDR reconstruction aims to extend this range of luminance by exploiting statistical properties of real-world photographs derived from high-level scene properties such as geometry and appearance.", + "bbox": [ + 212, + 314, + 785, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A brief look at the intrinsic components of HDR images, as shown in Figure 3, reveals the different characteristics of the shading and albedo layers. The distribution of reflectance values in real-world scenes typically has a narrow distribution, making the albedo a variable that can be represented linearly between [0, 1]. The high dynamic range of the HDR image, on the other hand, is reflected in the shading layer that similarly has an unbounded representation.", + "bbox": [ + 212, + 378, + 787, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As a result, the limited dynamic range of LDR images has different effects on the shading and albedo layers. As Figure 3 demonstrates, the clipping of HDR values results in a similar clipping in shading, showing the relationship between the dynamic range and illumination that has been previously discussed in the literature [10]. The loss of colors induced by clipping, on the other hand, appears in the albedo layer as the dullification of colors in over-exposed regions.", + "bbox": [ + 212, + 472, + 787, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on this observation, we formulate the HDR reconstruction problem separately in shading and albedo. In our setup, the network tasked with converting LDR shading into HDR shading models the high dynamic range of the illumination in the scene, taking advantage of the high correlation between shading and high-level scene properties such as geometry. We task another network with converting the LDR albedo into HDR albedo, training it to recover the lost color information in clipped regions. We detail our full pipeline in Section 4.", + "bbox": [ + 212, + 565, + 787, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our approach effectively separates the two challenges in dynamic range reconstruction, reconstruction of high luminance regions and recovery of the color information, into two separate tasks. Formulating the reconstruction of high luminance in the shading domain results in a more convenient way for the network to statistically model the problem. This can be attributed to the high correlation between shading and geometry, which is often violated in luminance due to textured surfaces as highlighted in Figure 5. Formulating the color recovery problem in the albedo domain simplifies the task similarly, allowing the exploitation of the sparsity of reflectance values in real-world scenes. Our albedo LDR-to-HDR network, as a result, acts comparable to a simple inpainting model, filling in the lost color details based on the scene content.", + "bbox": [ + 212, + 674, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Dille et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/da5ceac200d5241d3e3a7849ce0f20133336571ed2dc04095054c3cff64baef7.jpg", + "image_caption": [ + "Fig. 4: Our HDR reconstruction pipeline starts with the intrinsic decomposition of the image using an off-the-shelf method [3]. The LDR reconstruction is then done individually on shading and albedo layers, which are then combined in a final refinement stage. The input LDR image is provided to each network shown in blue." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 787, + 247 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Dynamic Range Extension via Infrinics", + "text_level": 1, + "bbox": [ + 215, + 342, + 640, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this work, we propose a physically inspired system for dynamic range extension in the single-image HDR reconstruction pipeline. The input to our pipeline is the dequantized and linearized LDR photograph $I_{L}$ . We first compute the intrinsic components of $I_{L}$ , the LDR shading $S_{L}$ , and albedo $A_{L}$ , using the in-the-wild intrinsic decomposition method by Careaga and Aksoy [3]. We then estimate an HDR shading and an HDR albedo from the corresponding LDR components using dedicated networks. The estimated HDR intrinsic components are then combined and refined by a subsequent HDR reconstruction network yielding our final result. We present an overview of our pipeline in Figure 4. In this section, we present our formulation and training details for each step.", + "bbox": [ + 212, + 377, + 787, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 HDR shading reconstruction", + "text_level": 1, + "bbox": [ + 215, + 554, + 500, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The shading layer in the intrinsic model contains the effect of illumination in the scene. The high dynamic range of illumination in the scene, shadows, and specular objects cause a long-tailed distribution of values in the shading map. An LDR input with clipped values similarly produces a clipped LDR shading layer after intrinsic decomposition. We train our HDR shading reconstruction network to extend the range of shading values beyond this clipping.", + "bbox": [ + 212, + 582, + 787, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Several works in the literature define the dynamic range extension task in the luminance domain, targeting a similar goal of isolating the effect of illumination. In their work, Eilertsen et al. [10] motivate this choice from a perceptual standpoint, citing earlier work that explores the effect of illumination versus surface colors in human perception [13]. The luminance, however, is computed as a linear combination of RGB channels and, as a result, still contains information on the reflectance. Shading, on the other hand, directly reflects the illumination and is accordingly highly correlated with the scene geometry as Figure 5 demonstrates. Defining the range extension task in shading instead of luminance benefits the network to model the clipped shading values by easily utilizing the geometry-related cues in the input.", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Intrinsic Single-Image HDR Reconstruction", + "bbox": [ + 441, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4cda54777a5382be486b73abf80ab6cb1963dccf4ca8196cd197edaf12e7c947.jpg", + "image_caption": [ + "Fig. 5: We show the difference between the shading representation and the L-channel from LUV. The shading correlates well with the geometry, while the L-channel includes texture information as shown in the green insets." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 349, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/65935f5a1892b83702eebc1741c8029c66906577e9aa840887a5ee7c134b1732.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 145, + 403, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8ce2a5991430227ba6f7e98ee4ed108a01c24c99464e0f68a683e1d4fb99db10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 143, + 537, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a941d2858a6c5af0c8e6af13b33c8346c21d14a8fc8cf16ccbd16728fd1e6a2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 145, + 591, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9d7215e652dd67b12075570a4f312a9697ed8effebabb29875416c0282d03e3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 593, + 143, + 725, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6667fea6c75d0a69fc8f5f9eb02570232249df2c186dea6d2790e3cb3f2c8ec6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 148, + 781, + 218 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Following Careaga and Aksoy [3], we represent shading in the inverse domain that fits the long-tailed shading in a representation bounded in $[0,1]$ :", + "bbox": [ + 215, + 303, + 782, + 333 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nD _ {L} = \\frac {1}{S _ {L} + 1}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 447, + 343, + 784, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The motivation behind this inverse representation is similar to the use of the log space in HDR reconstruction: It allows the definition of balanced loss functions on variables that could take very large or small values otherwise. The advantage of the inverse domain over the log space lies in its bounded nature that allows using the variable as input without clipping or normalization as well as the use of bounded activation functions such as sigmoid.", + "bbox": [ + 212, + 382, + 782, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our HDR shading reconstruction network receives the LDR image $I_{L}$ and the estimated LDR inverse shading $D_{L}$ as a concatenated $h\\times w\\times 4$ input. Its output is defined as the HDR inverse shading $D_{H}$ . We use multiple loss functions to train our network. The first one is the standard mean-squared error loss defined in the inverse space:", + "bbox": [ + 212, + 473, + 784, + 547 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m s e} ^ {D} = M S E \\left(D _ {H} ^ {*}, D _ {H}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 547, + 784, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $D_H^*$ and $D_H$ denote the ground-truth and estimated HDR inverse shading, respectively. We also use the multi-scale gradient loss [21] as an edge-aware smoothness loss:", + "bbox": [ + 212, + 570, + 782, + 614 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m s g} ^ {D} = M S G \\left(D _ {H} ^ {*}, D _ {H}\\right) = \\frac {1}{N} \\sum_ {m} S A D \\left(\\nabla D _ {H} ^ {* m}, \\nabla D _ {H} ^ {m}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 299, + 623, + 784, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\nabla$ denotes the spatial gradient and $D_H^m$ denotes the estimated inverse HDR shading at scale $m$ .", + "bbox": [ + 212, + 666, + 782, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although the output of our HDR shading reconstruction network is defined only as the HDR inverse shading, we can use the ground-truth HDR image $I_H^*$ and the intrinsic model in Eq. 1 to compute an implied HDR albedo $\\hat{A}_H$ [3]:", + "bbox": [ + 212, + 696, + 782, + 744 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {A} _ {H} = \\frac {I _ {H} ^ {*}}{S _ {H}}, \\quad S _ {H} = \\frac {1 - D _ {H}}{D _ {H}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 753, + 784, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We define the MSE and multi-scale gradient loss also for the implied albedo:", + "bbox": [ + 215, + 792, + 766, + 808 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m s e} ^ {\\hat {A}} = M S E \\left(A _ {H} ^ {*}, \\hat {A} _ {H}\\right), \\quad \\mathcal {L} _ {m s g} ^ {\\hat {A}} = M S G \\left(A _ {H} ^ {*}, \\hat {A} _ {H}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 302, + 816, + 784, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Dille et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and backpropagate these losses to our network through Eq. 5.", + "bbox": [ + 215, + 146, + 661, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our HDR shading reconstruction network only estimates the HDR shading and the HDR albedo component will be estimated with a separate network defined in the next section. Estimating the HDR intrinsic components separately results in possible reconstruction errors when these components are combined to produce an estimation for the HDR image. While we will refine the HDR reconstruction in our final stage with a dedicated network, we define these losses on the implied albedo to promote a better fit within the intrinsic model.", + "bbox": [ + 212, + 162, + 787, + 267 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our final loss for the HDR shading reconstruction network is defined as:", + "bbox": [ + 238, + 268, + 761, + 282 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {D} = \\mathcal {L} _ {m s e} ^ {D} + \\gamma \\cdot \\mathcal {L} _ {m s g} ^ {D} + \\mathcal {L} _ {m s e} ^ {\\hat {A}} + \\gamma \\cdot \\mathcal {L} _ {m s g} ^ {\\hat {A}}, \\quad \\gamma = 0. 1. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 297, + 785, + 316 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 HDR albedo reconstruction", + "text_level": 1, + "bbox": [ + 215, + 340, + 491, + 354 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The clipping in the LDR image formation process results in the loss of color information in the saturated pixels of the LDR albedo map. Our HDR albedo reconstruction network is tasked with recovering the corrupted colors.", + "bbox": [ + 212, + 369, + 784, + 412 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While the LDR shading map has saturated values, signaling the regions to be reconstructed to our HDR shading reconstruction networks, the LDR albedo map does not carry the same signal as albedo maps are not saturated. To denote the regions that require color reconstruction, we provide the network with a soft guidance mask $\\alpha$ defined as:", + "bbox": [ + 212, + 415, + 785, + 491 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha = \\frac {\\operatorname {m a x} \\left(0 , I _ {L} - \\lambda\\right)}{1 - \\lambda}, \\quad \\lambda = 0. 8, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 503, + 785, + 532 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "similar to the soft mask defined by Liu et al. [24]. It should be noted that, unlike Liu et al., we use this mask as an input to our network rather than a blending mask.", + "bbox": [ + 212, + 544, + 784, + 588 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our HDR albedo reconstruction network receives the LDR image $I_{L}$ , the estimated LDR albedo $A_{L}$ , and the soft attention mask $\\alpha$ , as a concatenated $h \\times w \\times 7$ input. Similar to our loss formulation for shading, we define the MSE and multi-scale gradient losses:", + "bbox": [ + 212, + 590, + 787, + 651 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m s e} ^ {A} = M S E \\left(A _ {H} ^ {*}, A _ {H}\\right), \\quad \\mathcal {L} _ {m s g} ^ {A} = M S G \\left(A _ {H} ^ {*}, A _ {H}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 664, + 785, + 681 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "as well as losses on implied inverse shading $\\hat{D}_H = A_H / (I_H^* +A_H)$ using the ground-truth HDR image $I_H^*$ and the intrinsic model in Eq. 1:", + "bbox": [ + 212, + 696, + 787, + 728 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m s e} ^ {\\hat {D}} = M S E \\left(D _ {H} ^ {*}, \\hat {D} _ {H}\\right), \\quad \\mathcal {L} _ {m s g} ^ {\\hat {D}} = M S G \\left(D _ {H} ^ {*}, \\hat {D} _ {H}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 741, + 785, + 766 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our final loss for the HDR albedo reconstruction network is defined as:", + "bbox": [ + 215, + 777, + 728, + 791 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {A} = \\mathcal {L} _ {m s e} ^ {A} + \\gamma \\cdot \\mathcal {L} _ {m s g} ^ {A} + \\mathcal {L} _ {m s e} ^ {\\hat {D}} + \\gamma \\cdot \\mathcal {L} _ {m s g} ^ {\\hat {D}}, \\quad \\gamma = 0. 1. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 804, + 785, + 823 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Intrinsic Single-Image HDR Reconstruction", + "bbox": [ + 441, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/dbd484098b7a32c5717aef01807b4552627f56e3301460a4ebdf0080ba1e563e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 356, + 217 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/cab2a7a40e3d3a507422dfbb14e2927db788ee7be893af66affbde34970cd006.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 143, + 498, + 217 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d6330f3c3bfb31c65b468dd1b1addbc4c7f03714ee69b839b1c160f3b2976cb9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 143, + 642, + 218 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/57958da0bbdd1dbe01365c0e27ae0e1837a0d8aed855e9a8266b81fb93602d8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 143, + 785, + 218 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/4c26592612025a030cba6e49fad6a7ac56336667f4e21fddc431a8951c0da82f.jpg", + "image_caption": [ + "Fig. 6: We show the intermediate results of our pipeline from an LDR input compared with its ground truth. The direct HDR reconstruction from the estimated HDR intrinsics can lead to small artifacts which are reduced by our refinement network." + ], + "image_footnote": [], + "bbox": [ + 217, + 219, + 356, + 292 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/ba67c0a195c4844acbab5e03a8915b4057f2b4e7b4fce570b4694b226d9f1ef0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 219, + 500, + 292 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/80a8eba905890e4235a4c08a8031cf9ff05200f43df6f7f51064e7e46cb14133.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 219, + 643, + 292 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a2b55dbcb2b36a2f4f25a1feaabd562dbfed2b9b16328b9d69715667f2e74e02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 219, + 784, + 292 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Refinement", + "text_level": 1, + "bbox": [ + 215, + 376, + 356, + 388 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The outputs of our HDR reconstruction networks for shading and albedo can be combined using the intrinsic model to give us an initial estimation for the HDR image:", + "bbox": [ + 212, + 402, + 784, + 446 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} _ {H} = A _ {H} * S _ {H}, \\quad S _ {H} = \\frac {1 - D _ {H}}{D _ {H}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 446, + 784, + 477 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "However, as the two intrinsic components are estimated separately, the perfect reconstruction of the HDR image is not guaranteed as shown in Figure 6. In the final step of our pipeline, we train a refinement network that takes the intrinsic HDR reconstruction results as input and generates our final HDR reconstruction result.", + "bbox": [ + 212, + 481, + 784, + 556 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "This network has a similar goal to the hallucination network used in Liu et al. [24]: Reconstructing the final HDR image by extending dynamic range. However, as it receives the intrinsic reconstruction of the HDR image readily as input, its task is significantly easier. Unlike a network trained for direct LDR-to-HDR conversion, it does not need to reason about the high-level cues in the image which is performed by our HDR shading reconstruction network. Similarly, it does not need to recover the corrupted colors in saturated regions, which is performed by our HDR albedo reconstruction network. It is simply tasked with combining the rich input information it receives to reconstruct the final HDR image, which allows it to generate more accurate and sharper results.", + "bbox": [ + 212, + 556, + 785, + 708 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Following the inverse representation we adopt for shading, we use the inverse representation for the HDR images as a bounded representation that fits the long-tailed distribution in HDR images between $[0,1]$ :", + "bbox": [ + 212, + 709, + 784, + 753 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nJ _ {H} = \\frac {1}{I _ {H} + 1}, \\quad \\hat {J} _ {H} = \\frac {1}{\\hat {I} _ {H} + 1}, \\quad J _ {H} ^ {*} = \\frac {1}{I _ {H} ^ {*} + 1}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 765, + 784, + 799 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $J_{H}$ , and $J_{H}^{*}$ denote the estimated and ground-truth inverse HDR images and $\\hat{J}_H$ is the inferred inverse HDR image following Eq. 12.", + "bbox": [ + 212, + 806, + 784, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Dille et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our refinement network receives the LDR image $I_{L}$ , the inferred inverse HDR image $\\hat{J}_{H}$ , the estimated HDR inverse shading $D_{L}$ , and the estimated HDR albedo $A_{L}$ as a concatenated $h\\times w\\times 10$ input. We use the MSE and multi-scale gradient losses to train our refinement network:", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m s e} ^ {J} = M S E \\left(J _ {H} ^ {*}, J _ {H}\\right), \\quad \\mathcal {L} _ {m s g} ^ {J} = M S G \\left(J _ {H} ^ {*}, J _ {H}\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 214, + 785, + 233 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "with the final loss defined as $\\mathcal{L}^J = \\mathcal{L}_{mse}^J +\\gamma \\cdot \\mathcal{L}_{msg}^J$ $\\gamma = 0.1$", + "bbox": [ + 214, + 239, + 653, + 258 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Network architecture and training", + "text_level": 1, + "bbox": [ + 214, + 273, + 547, + 291 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We utilize the same architecture for our 3 networks. We use EfficientNet [37] as the backbone with the decoder from [33], replacing the output activation with a sigmoid. We use RAdam [23] as the optimizer with a learning rate of $1e^{-4}$ and CosineAnnealing [25] scheduling.", + "bbox": [ + 212, + 296, + 787, + 357 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To train our intrinsic formulation, we use the provided synthetic intrinsic ground truth from Hypersim [35] and construct ground truth intrinsics from the raw images of the MultiIllum dataset [31] as described in [3]. We follow the preprocessing steps from Liu et al. [24] to create clipped, linearized LDR images with randomly sampled exposure values $t \\in [-3..3]$ from the ground truth data which we then decompose into our LDR input intrinsics using [3].", + "bbox": [ + 212, + 357, + 787, + 446 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Our refinement network in contrast does not require supervision in intrinsic space. We can thus extend our training set by real images from MultiRAW [22], LSMI [16], and the HDR-Real training subset from Liu et al. [24]. For this stage, we follow the ISP model in [24] to convert the ground truth images into JPEG files, use their pre-trained networks to retrieve linearized LDR images, and again retrieve the intrinsic decomposition via [3]. We provide a more detailed description of our training procedure in the supplementary material.", + "bbox": [ + 212, + 448, + 789, + 554 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 573, + 375, + 590 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we present a thorough evaluation of our proposed method using multiple standard benchmarks. First, we discuss the datasets, metrics and prior work used in our experimental setup, then we present qualitative and quantitative comparisons to state-of-the-art approaches. We present a more detailed evaluation and our ablation studies in the supplementary material.", + "bbox": [ + 212, + 602, + 787, + 679 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Quantitative evaluation", + "text_level": 1, + "bbox": [ + 214, + 696, + 455, + 713 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To quantitatively evaluate our approach against existing methods we follow the evaluation protocol proposed by Hanji et al. [15]. For a given ground truth HDR image, we follow the preprocessing of Liu et al. [24] and map the values into the range $[1 - 1000]$ . We determine the best scale to align the input reconstruction with the ground truth with least-squares, using only values between the $10^{th}$ and $90^{th}$ percentiles to avoid outliers. We then perform CRF correction to match the reconstruction to the ground truth by fitting a per-channel third-degree polynomial in log-RGB space.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Intrinsic Single-Image HDR Reconstruction", + "bbox": [ + 441, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7e9d7790db12538b3d2dc62a6ababd3228c4ad91d42f3f755a6b8d0cc26b3c01.jpg", + "table_caption": [ + "Table 1: Quantitative results against state-of-the-art using the HDR-VDP3 metric. Our method achieves state-of-the-art performance on the difficult high-definition SIHDR benchmark dataset. We are state-of-the-art or competitive on all other datasets." + ], + "table_footnote": [], + "table_body": "
MethodSIHDRHDR-EYEHDR-REALHDR-SYNTHRAISE
Ours8.968.967.578.218.89
DrTMO [11]8.278.727.578.278.64
HDR-CNN [10]8.398.616.708.038.47
ExpandNet [30]8.678.496.977.698.27
Single-HDR [24]8.798.967.478.518.95
Mask-HDR [36]8.258.607.108.148.44
HDRUNet [6]8.828.465.717.618.04
Multi-Exp Gen. [19]8.648.877.348.168.66
Lightweight [14]8.268.207.037.317.68
", + "bbox": [ + 246, + 198, + 750, + 354 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Metrics We follow the guidance of Hanji et al. [15] when choosing metrics to evaluate our approach. In their study, they find that certain metrics better correlate with results from qualitative experiments. We focus our evaluation on the popular HDR-VDP3 [28] metric as it is purpose-built for HDR evaluation and models the subjective evaluation of a human observer. We specifically use the \"quality\" task of HDR-VDP3. Furthermore, we evaluate multiple recommended SDR metrics using the PU21 [27] encoding and provide those results in the supplementary material.", + "bbox": [ + 212, + 382, + 787, + 505 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Datasets We evaluate our method on five publicly available datasets. We use the recent high-resolution SI-HDR Benchmark [15] as well as commonly used real-world datasets HDR-Eye [32], HDR-Real [24], HDR-Synth [24], and RAISE [7].", + "bbox": [ + 212, + 518, + 785, + 564 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We compare our method against a collection of baselines that share their implementation publicly. Table 1 shows the results of our quantitative evaluation. Each value represents the average HDR-VDP3 score in Just-Objectionable-Differences (JOD) units. We show a significant improvement in quality in the SI-HDR Benchmark [15] while achieving state-of-the-art performance on 3 of the other 4 datasets. As our qualitative evaluation in the next section and our results presented in the supplementary material demonstrate, the favorable performance of our method comes from our physically-motivated intrinsic formulation, allowing us to reliably reconstruct the high luminance details and recover lost colors. We provide additional metrics and analysis in the supplementary material.", + "bbox": [ + 212, + 565, + 787, + 715 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Qualitative evaluation", + "text_level": 1, + "bbox": [ + 215, + 738, + 444, + 753 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We present qualitative comparisons in Figures 7 and 8, showing the differences in both detail and color reconstruction. On SI-HDR [15], prior works struggle with predicting accurate details around edges where over- and under-exposed pixels meet. This becomes notable around the leaves in the [roof] scene with visible halos and color shifts. On the floor in the [book] scene, most competitors fail to", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Dille et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d01c1224f68a779baf37ec4aa9aefb287a92309ed3912809d25dba97200c9c7c.jpg", + "image_caption": [ + "Fig. 7: We show results from our method on SI-HDR [15] to the other baselines. We refer to Sec. 5.2 and the supplementary material for an in-depth discussion." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 787, + 633 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "reconstruct the homogenous surface and hallucinate patterns and color blobs. In both cases, our method seamlessly recovers missing details in over-exposed areas without artifacts. For the woman in Fig. 8, HDRUnet [6] and SingleHDR [24] both exhibit strong artifacts as bright spots in the hand and the face, similar to the faces of the [bust] and the [kids]. Our method in contrast can recover the correct color information due to our modeling of the albedo. The results showcase the advantage of the intrinsic formulation for creating visually pleasing results. We provide an extensive qualitative comparison in the supplementary material.", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Intrinsic Single-Image HDR Reconstruction", + "bbox": [ + 441, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a68e7598df6978107bd24a7a1d30819d8bd016f1170eafecd21a198f94eb29cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 143, + 357, + 218 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/38fb0e9c467400789472431a76624ce98c4613efe22db6a41a2a1a36f15856c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 143, + 500, + 218 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3984111da2aaebed448b163c5124bbe12dcf26dfe0bb3541574840daa12bf78c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 143, + 642, + 218 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2c509ad15a4ff5f7381532de693d67aa2c8469f4f192f9742ae0265e1cb40d1b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 143, + 784, + 218 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0334b1b1b9672037db7b8297e0f5a4a399a2fa3ec12f709179c15690ed2793d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 218, + 357, + 306 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b5bfc82b7b334beb34c57d3eff9acc483cbfe807d5b92ff22956e588b6743fa1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 218, + 500, + 306 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1325199239a42b17dcfe76f72a3e448e6cab1ac0826877099534e9db2d8388db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 218, + 642, + 306 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/02e55108d25f0081f4cdb6623ad11595b03086f5bdfce38a9462e415872e9f63.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 218, + 784, + 306 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/7caa7c6a2d016d5c843dfeb26973b03eff6203f70ab278fc85b81ef5db4cef0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 306, + 330, + 402 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b2213ab8689a9b81e00cf4590084b6a4f69c4f7126d8ad1b7d71b3d3fbb81b70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 306, + 444, + 402 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1df629466244146be371967528928d0f1a4795ab7a3d87ce078320badb88d704.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 306, + 557, + 402 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/dadb336317ebfdc6cdf802009298c99771540beb391081d93fc6de81870e388f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 306, + 671, + 402 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0af6ef37a4a6283a521df2d5ac7eaf9f1618b94cc0401e39de2944b8285df794.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 306, + 784, + 402 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/fe1864a602c8beb04e24af78df2114ffd2d6fe9cd4225a415f4fdc4863c77e3c.jpg", + "image_caption": [ + "Fig.8: We compare results from our method with a focus on color reconstruction against HDRUnet [6] and SingleHDR [24]. We refer to Sec. 5.2 and the supplementary material for an in-depth discussion." + ], + "image_footnote": [], + "bbox": [ + 217, + 404, + 330, + 479 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4b7367fc33f5f7d7825fb0255bc5036ade08a3040a5139f3ab7a02a4f71e8b77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 404, + 444, + 479 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/547c2ab0a0540460c4752b12eee0d8e241bf7e8523ebe20cb0d537112080e4e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 404, + 557, + 479 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/601e42538306657d9deab810c93630f380a60ccd79072501a95aa50d632f199f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 404, + 671, + 479 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/63693be8cf6195708200005ad11e9d54dba18466f152c7e74118c3ff36e304fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 404, + 784, + 479 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 564, + 359, + 580 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we remodel the HDR reconstruction problem in the intrinsic domain and propose a multi-stage pipeline to recover lost details. The luminance range extension can be conveniently modeled in the shading space which allows the network to better model the correlation between illumination and geometry. Similarly, neural networks can easily take advantage of albedo sparsity to recover lost color information with high fidelity. Through qualitative and quantitative evaluations, we show that our physically-inspired modeling of the HDR reconstruction problem achieves state-of-the-art performance, improving upon prior work in terms of both detail recovery and color reproduction in a wide variety of in-the-wild photographs.", + "bbox": [ + 212, + 598, + 787, + 750 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 775, + 403, + 792 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We acknowledge the support of the Natural Sciences and Engineering Research Council of Canada (NSERC), [RGPIN-2020-05375].", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Dille et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Banterle, F., Ledda, P., Debattista, K., Chalmers, A.: Inverse tone mapping. In: Proc.GRAPHITE (2006)", + "2. Barrow, H., Tenenbaum, J., Hanson, A., Riseman, E.: Recovering intrinsic scene characteristics. Comput. vis. syst 2(3-26), 2 (1978)", + "3. Careaga, C., Aksoy, Y.: Intrinsic image decomposition via ordinal shading. ACM Trans. Graph. 43(1) (2023)", + "4. Careaga, C., Miangoleh, S.M.H., Aksoy, Y.: Intrinsic harmonization for illumination-aware compositing. In: Proc. SIGGRAPH Asia (2023)", + "5. Chen, S.K., Yen, H.L., Liu, Y.L., Chen, M.H., Hu, H.N., Peng, W.H., Lin, Y.Y.: CEVR: Learning continuous exposure value representations for single-image HDR reconstruction. In: Proc. ICCV (2023)", + "6. Chen, X., Liu, Y., Zhang, Z., Qiao, Y., Dong, C.: HDRUnet: Single image HDR reconstruction with denoising and dequantization. In: Proc. CVPR (2021)", + "7. Dang-Nguyen, D.T., Pasquini, C., Conotter, V., Boato, G.: RAISE: a raw images dataset for digital image forensics. In: Proc. MMSys (2015)", + "8. Debevec, P.: A median cut algorithm for light probe sampling. In: ACM SIGGRAPH 2005 Posters. ACM (2005)", + "9. Debevec, P.E., Malik, J.: Recovering high dynamic range radiance maps from photographs. In: ACM Trans. Graph. (1997)", + "10. Eilertsen, G., Kronander, J., Denes, G., Mantiuk, R.K., Unger, J.: HDR image reconstruction from a single exposure using deep CNNs. ACM Trans. Graph. 36(6) (2017)", + "1. Endo, Y., Kanamori, Y., Mitani, J.: Deep reverse tone mapping. ACM Trans. Graph. 36(6) (2017)", + "2. Garces, E., Rodriguez-Pardo, C., Casas, D., Lopez-Moreno, J.: A survey on intrinsic images: Delving deep into lambert and beyond. Int. J. Comput. Vision (2022)", + "3. Gilchrist, A., Jacobsen, A.: Perception of lightness and illumination in a world of one reflectance. Perception 13, 5-19 (02 1984)", + "4. Guo, C., Xiuhua, J.: LHDR: HDR reconstruction for legacy content using a lightweight DNN. In: Proc. ACCV (2022)", + "5. Hanji, P., Mantiuk, R., Eilertsen, G., Hajisharif, S., Unger, J.: Comparison of single image HDR reconstruction methods — the caveats of quality assessment. In: ACM Trans. Graph. (2022)", + "6. Kim, D., Kim, J., Nam, S., Lee, D., Lee, Y., Kang, N., Lee, H.E., Yoo, B., Han, J.J., Kim, S.J.: Large scale multi-illuminant (LSMI) dataset for developing white balance algorithm under mixed illumination. In: Proc. CVPR (2021)", + "7. Kim, J., Lee, S., Kang, S.J.: End-to-end differentiable learning to HDR image synthesis for multi-exposure images. In: Proc. AAAI (2021)", + "8. Kovaleski, R.P., Oliveira, M.M.: High-quality reverse tone mapping for a wide range of exposures. In: Proc. SIBGRAPI (2014)", + "9. Le, P.H., Le, Q., Nguyen, R., Hua, B.S.: Single-image HDR reconstruction by multi-exposure generation. In: Proc. WACV (2023)", + "20. Lee, S., An, G.H., Kang, S.J.: Deep recursive HDRI: Inverse tone mapping using generative adversarial networks. In: Proc. ECCV (2018)", + "21. Li, Z., Snavely, N.: Learning intrinsic image decomposition from watching the world. In: Proc. CVPR (2018)", + "22. Li, Z., Lu, M., Zhang, X., Feng, X., Asif, M.S., Ma, Z.: Efficient visual computing with camera raw snapshots. IEEE Trans. Pattern Anal. Mach. Intell. (2024)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Intrinsic Single-Image HDR Reconstruction", + "bbox": [ + 441, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "23. Liu, L., Jiang, H., He, P., Chen, W., Liu, X., Gao, J., Han, J.: On the variance of the adaptive learning rate and beyond. In: Proc. ICLR (2020)", + "24. Liu, Y.L., Lai, W.S., Chen, Y.S., Kao, Y.L., Yang, M.H., Chuang, Y.Y., Huang, J.B.: Single-image HDR reconstruction by learning to reverse the camera pipeline. In: Proc. CVPR (2020)", + "25. Loshchilov, I., Hutter, F.: SGDR: Stochastic gradient descent with warm restarts. In: Proc. ICLR (2017)", + "26. Mann, S., Picard, R.: On being \"undigital\" with digital cameras: Extending dynamic range by combining differently exposed pictures. In: Proc. IS&T Annual Conference (1995)", + "27. Mantiuk, R.K., Azimi, M.: Pu21: A novel perceptually uniform encoding for adapting existing quality metrics for hdr. In: Proc. PCS (2021)", + "28. Mantiuk, R.K., Hammou, D., Hanji, P.: HDR-VDP-3: A multi-metric for predicting image differences, quality and contrast distortions in high dynamic range and regular content. arXiv preprint arXiv:2304.13625 (2023)", + "29. Maralan, S.S., Careaga, C., Aksoy, Y.: Computational flash photography through intrinsics. In: Proc. CVPR (2023)", + "30. Marnerides, D., Bashford-Rogers, T., Hatchett, J., Debattista, K.: Expandnet: A deep convolutional neural network for high dynamic range expansion from low dynamic range content. Comput. Graph. Forum 37(2) (2018)", + "31. Murmann, L., Gharbi, M., Aittala, M., Durand, F.: A multi-illumination dataset of indoor object appearance. In: Proc. ICCV (2019)", + "32. Nemoto, H., Korshunov, P., Hanhart, P., Ebrahimi, T.: Visual attention in LDR and HDR images. In: Proc. VPQM (2015)", + "33. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Trans. Pattern Anal. Mach. Intell. 44(3) (2022)", + "34. Rempel, A.G., Trentacoste, M., Seetzen, H., Young, H.D., Heidrich, W., Whitehead, L., Ward, G.: LDR2HDR: on-the-fly reverse tone mapping of legacy video and photographs. ACM Trans. Graph. 26(3) (2007)", + "35. Roberts, M., Ramapuram, J., Ranjan, A., Kumar, A., Bautista, M.A., Paczan, N., Webb, R., Susskind, J.M.: Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In: Proc. ICCV (2021)", + "36. Santos, M.S., Ren, T.I., Kalantari, N.K.: Single image HDR reconstruction using a CNN with masked features and perceptual loss. ACM Trans. Graph. 39(4) (2020)", + "37. Tan, M., Le, Q.: EfficientNet: Rethinking model scaling for convolutional neural networks. In: Proc. ICML (2019)", + "38. Wang, C., Serrano, A., Pan, X., Chen, B., Myszkowski, K., Seidel, H.P., Theobalt, C., Leimkuhler, T.: Glowgan: Unsupervised learning of HDR images from LDR images in the wild. In: Proc. ICCV (2023)", + "39. Xu, D., Doutre, C., Nasiopoulos, P.: Correction of clipped pixels in color images. IEEE Trans. Vis. Comput. Graph. 17(3) (2010)", + "40. Zhang, N., Ye, Y., Zhao, Y., Wang, R.: Revisiting the stack-based inverse tone mapping. In: Proc. CVPR (2023)", + "41. Zhang, X., Brainard, D.H.: Estimation of saturated pixel values in digital color imaging. J. Optical Soc. Am. A 21(12) (2004)", + "42. Zhang, Y., Aydin, T.: Deep HDR estimation with generative detail reconstruction. Comput. Graph. Gorum 40(2), 179-190 (2021)" + ], + "bbox": [ + 215, + 147, + 785, + 811 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Dille et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/Intrinsic Single-Image HDR Reconstruction/e55f56f5-9c00-4140-aab7-15323b235acd_model.json b/2024/Intrinsic Single-Image HDR Reconstruction/e55f56f5-9c00-4140-aab7-15323b235acd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9980cc503a0fb7e4703abe27655a449d57c7e9c9 --- /dev/null +++ b/2024/Intrinsic Single-Image HDR Reconstruction/e55f56f5-9c00-4140-aab7-15323b235acd_model.json @@ -0,0 +1,2564 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.252, + 0.142, + 0.753, + 0.164 + ], + "angle": 0, + "content": "Intrinsic Single-Image HDR Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.19, + 0.711, + 0.207 + ], + "angle": 0, + "content": "Sebastian Dille\\* , Chris Careaga\\* , and Yaigiz Aksoy" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.218, + 0.61, + 0.258 + ], + "angle": 0, + "content": "Simon Fraser University Burnaby, BC V5A 1S6, Canada sdille@sfu.ca" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.274, + 0.382, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.358, + 0.361, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.275, + 0.5, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.275, + 0.619, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.275, + 0.772, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.358, + 0.753, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.398, + 0.366, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.47, + 0.366, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.398, + 0.501, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.47, + 0.501, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.398, + 0.636, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.47, + 0.636, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.398, + 0.771, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.47, + 0.771, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.509, + 0.788, + 0.552 + ], + "angle": 0, + "content": "Fig. 1: In this work, we propose a physically-inspired HDR reconstruction pipeline by conducting reconstruction individually on intrinsic components. Our final reconstruction, as a result, is effective in recovering the high-luminance details and the lost colors." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.572, + 0.741, + 0.78 + ], + "angle": 0, + "content": "Abstract. The low dynamic range (LDR) of common cameras fails to capture the rich contrast in natural scenes, resulting in loss of color and details in saturated pixels. Reconstructing the high dynamic range (HDR) of luminance present in the scene from single LDR photographs is an important task with many applications in computational photography and realistic display of images. The HDR reconstruction task aims to infer the lost details using the context present in the scene, requiring neural networks to understand high-level geometric and illumination cues. This makes it challenging for data-driven algorithms to generate accurate and high-resolution results. In this work, we introduce a physically-inspired remodeling of the HDR reconstruction problem in the intrinsic domain. The intrinsic model allows us to train separate networks to extend the dynamic range in the shading domain and to recover lost color details in the albedo domain. We show that dividing the problem into two simpler sub-tasks improves performance in a wide variety of photographs." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.793, + 0.741, + 0.821 + ], + "angle": 0, + "content": "Keywords: HDR Reconstruction \\(\\cdot\\) Inverse Tonemapping \\(\\cdot\\) Computational Photography" + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.826, + 0.441, + 0.841 + ], + "angle": 0, + "content": "\\((^{*})\\) denotes equal contribution." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "S. Dille et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.375, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.311 + ], + "angle": 0, + "content": "Real-world scenes contain a wider range of luminance than most consumer-grade cameras can record or store. As a result, everyday photographs can only represent a low dynamic range (LDR) version of the scene that lacks details and color in very dark or bright regions. High dynamic range (HDR) imaging aims to capture and store the full contrast present in the scene. HDR images allow realistic display of photographs using high-contrast displays as well as many computational photography applications such as tone mapping and illumination-aware image editing. However, HDR imaging typically requires multi-exposure image stacks which limits their applications to static scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.788, + 0.493 + ], + "angle": 0, + "content": "To extend the HDR-enabled applications to everyday photographs, there has been a growing interest in methods to estimate HDR images from only LDR photographs called single-image HDR reconstruction, or LDR-to-HDR conversion. Single-image HDR reconstruction aims to extend the dynamic range of LDR photographs by recovering the details lost during capture. The missing information in LDR images is typically characterized by saturated pixels in bright regions, lacking textural details and color. This makes HDR reconstruction a high-level problem, requiring networks to utilize geometric and context-dependent cues in the scene to infer the lost information. Formulating this reconstruction problem using standard color spaces such as RGB or Luv requires the networks to statistically model the complex relationship between luminance, surface colors, and geometry, which limits their performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.75 + ], + "angle": 0, + "content": "In this work, we present a physically-inspired remodeling of the HDR reconstruction problem in the intrinsic domain. Intrinsic image decomposition is a mid-level vision problem that separates the effect of the illumination in the scene from the reflectance of surfaces [3], representing the input as a combination of a shading and an albedo layer. The shading of an image is highly correlated with the scene geometry and illumination, while the albedo contains the color information and the high-frequency textures. The saturated pixels we aim to recover create different degradations in the shading and albedo layers. The clipping of high values is observed in the shading, while the loss of color information causes desaturation in the albedo. We propose an HDR reconstruction pipeline that individually addresses the loss of information in the LDR image in the two intrinsic components. Our HDR shading reconstruction network extends the dynamic range, taking advantage of the high correlation between shading and the scene geometry. Our LDR albedo reconstruction network, on the other hand, recovers the lost color information in saturated regions. A final refinement stage combines the two components to generate our full HDR reconstruction. Figure 1 shows different stages of our pipeline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In our formulation, we take advantage of the intrinsic image formation model by dividing the recovery of lost details and colors into two separate tasks which are more convenient for networks to model. We show that our physically inspired approach can generate high-fidelity details with rich colors in complex in-the-wild scenes. We evaluate our method qualitatively and quantitatively to demonstrate the state-of-the-art performance of our formulation in a wide variety of scenes." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.442, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Intrinsic Single-Image HDR Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.383, + 0.161 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.186, + 0.784, + 0.246 + ], + "angle": 0, + "content": "Common imaging sensors for photography are inherently limited in the portion of the scene's dynamic range they can capture in a single shot. This results in saturated pixels in bright regions, representing a significant challenge in the capture and realistic display of real-world scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.248, + 0.785, + 0.444 + ], + "angle": 0, + "content": "Multi-exposure fusion methods can reconstruct an HDR image by combining a stack of LDR images taken at different exposures [9,26,40]. Their formulation successfully addresses contrast range extension and reversal of the quantization and nonlinearity of LDR images. The multi-exposure capture, however, assumes a static camera and scene, requiring a challenging alignment process and limiting their application scenario. To enable HDR-related applications in everyday photography, single-image HDR reconstruction approaches aim to recover the lost information directly from an LDR image. Earlier approaches apply specially designed filters [18,34] or median-cut [1,8] to fill in overexposed regions, using different heuristics to scale the intensity of a selected area. Similarly, Bayesian [41] and graph-based formulations [39] have been introduced for chroma restoration. Due to the high-level nature of the reconstruction problem, these low-level approaches struggle to generate realistic details in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.787, + 0.718 + ], + "angle": 0, + "content": "In their concurrent publications, Eilertsen et al. [10] and Endo et al. [11] present the first data-driven approaches to the single-image HDR reconstruction problem. They train their networks for direct conversion to HDR from the LDR input, addressing quantization, nonlinearity, and saturation in the LDR image in a single step. Many works in recent literature adapt this unified inverse model [24] and develop tailored solutions for underexposed [6,14] or overexposed [36] areas or losses for color recovery [30]. The direct estimation of HDR outputs, however, creates the challenge of an unconstrained color distribution that hinders training efficiency. Several methods [5, 11, 17, 19, 20, 40] instead adopt the concept of exposure stacks and predict a set of over- and underexposed images from the input LDR image. With each prediction still in the LDR domain, the HDR image can later be created with multi-exposure fusion [9]. Other methods [38,42] employ generative approaches, either generating separate LDR detail layers [42] or full HDR images directly [38], with image-specific HDR reconstruction via GAN inversion. Liu et al. [24] approach HDR reconstruction by inverting the in-camera image formation process, modeling the dequantization, nonlinearization, and saturated pixel reconstruction in three separate steps. They show that this decomposition of the task simplifies the modeling of each individual sub-problem." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this work, we adopt the modular formulation of Liu et al. [24] and focus on recovering the lost details and color information in the saturated pixels. Our method takes inspiration from prior works that formulate computational photography problems in the intrinsic image domain [4, 29]. Rather than defining a direct mapping from LDR to HDR, we divide this task into two sub-problems: shading reconstruction that extends the dynamic range, and reflectance reconstruction that recovers the color information. Our physical deconstruction of the problem significantly improves the recovery of details and color information." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "S. Dille et al." + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.147, + 0.784, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.345 + ], + "angle": 0, + "content": "Fig. 2: The in-camera pipeline modeling LDR image formation is shown at the top. Liu et al. [24] propose the inverse pipeline shown on the bottom, addressing each type of information loss individually in three steps. In this work, we use the pre-trained networks of Liu et al. [24] for the dequantization and linearization tasks (shown in pink) and focus on developing an intrinsic model for the dynamic range reconstruction problem highlighted in teal." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.373, + 0.525, + 0.391 + ], + "angle": 0, + "content": "3 Background and Motivation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.789, + 0.485 + ], + "angle": 0, + "content": "The focus of our work is the formulation of the single-image LDR-to-HDR reconstruction problem in the intrinsic domain. We begin our discussion by first introducing the LDR imaging pipeline and the intrinsic model of image formation. We then draw connections between the two models and motivate our intrinsic HDR formulation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.51, + 0.654, + 0.525 + ], + "angle": 0, + "content": "3.1 LDR image formation and HDR reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.689 + ], + "angle": 0, + "content": "The acquisition of standard RGB photographs involves a series of operations in the camera-internal image processing pipeline, or ISP, that result in different forms of information loss. The initial acquisition of the image is in a raw format, where the luminance received by the sensor from a real-world scene is scaled and stored in 16- or 32-bit data structures. The scale of the captured raw values is determined by the analog camera settings such as the shutter speed, the aperture size, and the ISO. Given the natural high dynamic range of real-world scenes, the raw values are clipped at a maximum due to saturation at the sensor level. As shown in Figure 2, the remainder of the ISP applies nonlinear color and gamma corrections before finally decreasing the bit rate to 8 bits per channel for storage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.789, + 0.841 + ], + "angle": 0, + "content": "The LDR image formation process causes three different types of corruption of the HDR information: quantization, non-linearization, and clipping. Single-image HDR reconstruction methods, as a result, have to address all three of these challenges. While many methods in the literature address all three simultaneously within a single architecture, Liu et al. [24] design their pipeline to address each type of information loss with a separate dedicated network. They propose a dequantization network to first increase the bit depth of the input LDR image and then estimate a CRF curve for linearization. Both tasks are convenient for a network to learn, taking advantage of the ease of dataset generation for dequantization and a simple parametric curve estimation for linearization." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.442, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Intrinsic Single-Image HDR Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.149, + 0.789, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.321, + 0.788, + 0.378 + ], + "angle": 0, + "content": "Fig. 3: We show the different effects of clipping on the albedo and the shading for a tonemapped HDR image and its LDR counterpart. When clipping the image, the LDR shading loses information compared to HDR. We scale the intensity of both shadings to show the effect. At the same time, the LDR albedo becomes saturated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.789, + 0.542 + ], + "angle": 0, + "content": "They then reconstruct the values lost to clipping with a third network in what they call the hallucination stage. This last stage of the inverse pipeline addresses the most significant challenge in single-image HDR reconstruction, namely the recovery of the highlights and the color information in clipped regions. Diverging from Liu et al.'s terminology, we refer to this stage as dynamic range extension. We adopt Liu et al.'s inverse pipeline and use their pre-trained networks for dequantization and linearization. In the following, we focus on the dynamic range extension problem and present our solution using an intrinsic model of image formation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.565, + 0.51, + 0.58 + ], + "angle": 0, + "content": "3.2 Intrinsic image decomposition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.788, + 0.68 + ], + "angle": 0, + "content": "Intrinsic image decomposition is a fundamental mid-level computer vision problem that represents a core step in inverse rendering pipelines [2, 12]. It aims to separate the image into illumination and reflectance components following the physical image formation model. Adopting the commonly used Lambertian shading assumption with a single-channel shading representation [3], the intrinsic equation can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.445, + 0.692, + 0.787, + 0.71 + ], + "angle": 0, + "content": "\\[\nI _ {r g b} = A _ {r g b} \\cdot S, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.785, + 0.764 + ], + "angle": 0, + "content": "where \\( I \\) represents the linear image reflecting relative luminance, \\( A \\) represents albedo, i.e. the surface reflectance containing the color information, and \\( S \\) represents shading reflecting the effect of illumination in the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Shading and albedo represent two orthogonal characteristics of a scene, one containing the effect of light and its interaction with the scene geometry and the other representing material properties. As a result, the two intrinsic components have very different characteristics as demonstrated in Figure 3. Shading is highly correlated with surface orientations, appearing smooth over continuous surfaces" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "S. Dille et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "and having a high gradient in depth discontinuities and shadow boundaries. Albedo, on the other hand, contains the high-frequency textures present on the materials in the scene while having a sparse set of values per image as the variation in brightness due to geometry is isolated in the shading layer. In the single-channel shading representation, albedo also contains any variation caused by colorful illumination [3]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.275, + 0.494, + 0.29 + ], + "angle": 0, + "content": "3.3 Infrinsics in LDR and HDR" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.787, + 0.376 + ], + "angle": 0, + "content": "The bounded representation of LDR images limits the range of luminance values they can contain. HDR reconstruction aims to extend this range of luminance by exploiting statistical properties of real-world photographs derived from high-level scene properties such as geometry and appearance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.47 + ], + "angle": 0, + "content": "A brief look at the intrinsic components of HDR images, as shown in Figure 3, reveals the different characteristics of the shading and albedo layers. The distribution of reflectance values in real-world scenes typically has a narrow distribution, making the albedo a variable that can be represented linearly between [0, 1]. The high dynamic range of the HDR image, on the other hand, is reflected in the shading layer that similarly has an unbounded representation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.563 + ], + "angle": 0, + "content": "As a result, the limited dynamic range of LDR images has different effects on the shading and albedo layers. As Figure 3 demonstrates, the clipping of HDR values results in a similar clipping in shading, showing the relationship between the dynamic range and illumination that has been previously discussed in the literature [10]. The loss of colors induced by clipping, on the other hand, appears in the albedo layer as the dullification of colors in over-exposed regions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.673 + ], + "angle": 0, + "content": "Based on this observation, we formulate the HDR reconstruction problem separately in shading and albedo. In our setup, the network tasked with converting LDR shading into HDR shading models the high dynamic range of the illumination in the scene, taking advantage of the high correlation between shading and high-level scene properties such as geometry. We task another network with converting the LDR albedo into HDR albedo, training it to recover the lost color information in clipped regions. We detail our full pipeline in Section 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Our approach effectively separates the two challenges in dynamic range reconstruction, reconstruction of high luminance regions and recovery of the color information, into two separate tasks. Formulating the reconstruction of high luminance in the shading domain results in a more convenient way for the network to statistically model the problem. This can be attributed to the high correlation between shading and geometry, which is often violated in luminance due to textured surfaces as highlighted in Figure 5. Formulating the color recovery problem in the albedo domain simplifies the task similarly, allowing the exploitation of the sparsity of reflectance values in real-world scenes. Our albedo LDR-to-HDR network, as a result, acts comparable to a simple inpainting model, filling in the lost color details based on the scene content." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.442, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Intrinsic Single-Image HDR Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.788, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.257, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Fig. 4: Our HDR reconstruction pipeline starts with the intrinsic decomposition of the image using an off-the-shelf method [3]. The LDR reconstruction is then done individually on shading and albedo layers, which are then combined in a final refinement stage. The input LDR image is provided to each network shown in blue." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.343, + 0.642, + 0.361 + ], + "angle": 0, + "content": "4 Dynamic Range Extension via Infrinics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.789, + 0.53 + ], + "angle": 0, + "content": "In this work, we propose a physically inspired system for dynamic range extension in the single-image HDR reconstruction pipeline. The input to our pipeline is the dequantized and linearized LDR photograph \\( I_{L} \\). We first compute the intrinsic components of \\( I_{L} \\), the LDR shading \\( S_{L} \\), and albedo \\( A_{L} \\), using the in-the-wild intrinsic decomposition method by Careaga and Aksoy [3]. We then estimate an HDR shading and an HDR albedo from the corresponding LDR components using dedicated networks. The estimated HDR intrinsic components are then combined and refined by a subsequent HDR reconstruction network yielding our final result. We present an overview of our pipeline in Figure 4. In this section, we present our formulation and training details for each step." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.555, + 0.5, + 0.57 + ], + "angle": 0, + "content": "4.1 HDR shading reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.789, + 0.674 + ], + "angle": 0, + "content": "The shading layer in the intrinsic model contains the effect of illumination in the scene. The high dynamic range of illumination in the scene, shadows, and specular objects cause a long-tailed distribution of values in the shading map. An LDR input with clipped values similarly produces a clipped LDR shading layer after intrinsic decomposition. We train our HDR shading reconstruction network to extend the range of shading values beyond this clipping." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Several works in the literature define the dynamic range extension task in the luminance domain, targeting a similar goal of isolating the effect of illumination. In their work, Eilertsen et al. [10] motivate this choice from a perceptual standpoint, citing earlier work that explores the effect of illumination versus surface colors in human perception [13]. The luminance, however, is computed as a linear combination of RGB channels and, as a result, still contains information on the reflectance. Shading, on the other hand, directly reflects the illumination and is accordingly highly correlated with the scene geometry as Figure 5 demonstrates. Defining the range extension task in shading instead of luminance benefits the network to model the clipped shading values by easily utilizing the geometry-related cues in the input." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "S. Dille et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.35, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.146, + 0.404, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.145, + 0.538, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.146, + 0.593, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.145, + 0.726, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.15, + 0.782, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.233, + 0.785, + 0.275 + ], + "angle": 0, + "content": "Fig. 5: We show the difference between the shading representation and the L-channel from LUV. The shading correlates well with the geometry, while the L-channel includes texture information as shown in the green insets." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.304, + 0.784, + 0.334 + ], + "angle": 0, + "content": "Following Careaga and Aksoy [3], we represent shading in the inverse domain that fits the long-tailed shading in a representation bounded in \\([0,1]\\):" + }, + { + "type": "equation", + "bbox": [ + 0.449, + 0.344, + 0.785, + 0.375 + ], + "angle": 0, + "content": "\\[\nD _ {L} = \\frac {1}{S _ {L} + 1}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.784, + 0.472 + ], + "angle": 0, + "content": "The motivation behind this inverse representation is similar to the use of the log space in HDR reconstruction: It allows the definition of balanced loss functions on variables that could take very large or small values otherwise. The advantage of the inverse domain over the log space lies in its bounded nature that allows using the variable as input without clipping or normalization as well as the use of bounded activation functions such as sigmoid." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.785, + 0.548 + ], + "angle": 0, + "content": "Our HDR shading reconstruction network receives the LDR image \\( I_{L} \\) and the estimated LDR inverse shading \\( D_{L} \\) as a concatenated \\( h\\times w\\times 4 \\) input. Its output is defined as the HDR inverse shading \\( D_{H} \\). We use multiple loss functions to train our network. The first one is the standard mean-squared error loss defined in the inverse space:" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.548, + 0.785, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m s e} ^ {D} = M S E \\left(D _ {H} ^ {*}, D _ {H}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.571, + 0.784, + 0.616 + ], + "angle": 0, + "content": "where \\( D_H^* \\) and \\( D_H \\) denote the ground-truth and estimated HDR inverse shading, respectively. We also use the multi-scale gradient loss [21] as an edge-aware smoothness loss:" + }, + { + "type": "equation", + "bbox": [ + 0.3, + 0.624, + 0.785, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m s g} ^ {D} = M S G \\left(D _ {H} ^ {*}, D _ {H}\\right) = \\frac {1}{N} \\sum_ {m} S A D \\left(\\nabla D _ {H} ^ {* m}, \\nabla D _ {H} ^ {m}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.784, + 0.697 + ], + "angle": 0, + "content": "where \\(\\nabla\\) denotes the spatial gradient and \\(D_H^m\\) denotes the estimated inverse HDR shading at scale \\(m\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.784, + 0.745 + ], + "angle": 0, + "content": "Although the output of our HDR shading reconstruction network is defined only as the HDR inverse shading, we can use the ground-truth HDR image \\( I_H^* \\) and the intrinsic model in Eq. 1 to compute an implied HDR albedo \\( \\hat{A}_H \\) [3]:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.755, + 0.785, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\hat {A} _ {H} = \\frac {I _ {H} ^ {*}}{S _ {H}}, \\quad S _ {H} = \\frac {1 - D _ {H}}{D _ {H}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.794, + 0.767, + 0.809 + ], + "angle": 0, + "content": "We define the MSE and multi-scale gradient loss also for the implied albedo:" + }, + { + "type": "equation", + "bbox": [ + 0.303, + 0.818, + 0.785, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m s e} ^ {\\hat {A}} = M S E \\left(A _ {H} ^ {*}, \\hat {A} _ {H}\\right), \\quad \\mathcal {L} _ {m s g} ^ {\\hat {A}} = M S G \\left(A _ {H} ^ {*}, \\hat {A} _ {H}\\right), \\tag {6}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.442, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Intrinsic Single-Image HDR Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.662, + 0.162 + ], + "angle": 0, + "content": "and backpropagate these losses to our network through Eq. 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.163, + 0.788, + 0.268 + ], + "angle": 0, + "content": "Our HDR shading reconstruction network only estimates the HDR shading and the HDR albedo component will be estimated with a separate network defined in the next section. Estimating the HDR intrinsic components separately results in possible reconstruction errors when these components are combined to produce an estimation for the HDR image. While we will refine the HDR reconstruction in our final stage with a dedicated network, we define these losses on the implied albedo to promote a better fit within the intrinsic model." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.269, + 0.762, + 0.284 + ], + "angle": 0, + "content": "Our final loss for the HDR shading reconstruction network is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.298, + 0.786, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {D} = \\mathcal {L} _ {m s e} ^ {D} + \\gamma \\cdot \\mathcal {L} _ {m s g} ^ {D} + \\mathcal {L} _ {m s e} ^ {\\hat {A}} + \\gamma \\cdot \\mathcal {L} _ {m s g} ^ {\\hat {A}}, \\quad \\gamma = 0. 1. \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.341, + 0.492, + 0.356 + ], + "angle": 0, + "content": "4.2 HDR albedo reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.785, + 0.414 + ], + "angle": 0, + "content": "The clipping in the LDR image formation process results in the loss of color information in the saturated pixels of the LDR albedo map. Our HDR albedo reconstruction network is tasked with recovering the corrupted colors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.787, + 0.492 + ], + "angle": 0, + "content": "While the LDR shading map has saturated values, signaling the regions to be reconstructed to our HDR shading reconstruction networks, the LDR albedo map does not carry the same signal as albedo maps are not saturated. To denote the regions that require color reconstruction, we provide the network with a soft guidance mask \\(\\alpha\\) defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.504, + 0.786, + 0.533 + ], + "angle": 0, + "content": "\\[\n\\alpha = \\frac {\\operatorname {m a x} \\left(0 , I _ {L} - \\lambda\\right)}{1 - \\lambda}, \\quad \\lambda = 0. 8, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.545, + 0.785, + 0.589 + ], + "angle": 0, + "content": "similar to the soft mask defined by Liu et al. [24]. It should be noted that, unlike Liu et al., we use this mask as an input to our network rather than a blending mask." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.652 + ], + "angle": 0, + "content": "Our HDR albedo reconstruction network receives the LDR image \\( I_{L} \\), the estimated LDR albedo \\( A_{L} \\), and the soft attention mask \\( \\alpha \\), as a concatenated \\( h \\times w \\times 7 \\) input. Similar to our loss formulation for shading, we define the MSE and multi-scale gradient losses:" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.665, + 0.786, + 0.683 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m s e} ^ {A} = M S E \\left(A _ {H} ^ {*}, A _ {H}\\right), \\quad \\mathcal {L} _ {m s g} ^ {A} = M S G \\left(A _ {H} ^ {*}, A _ {H}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.788, + 0.729 + ], + "angle": 0, + "content": "as well as losses on implied inverse shading \\(\\hat{D}_H = A_H / (I_H^* +A_H)\\) using the ground-truth HDR image \\(I_H^*\\) and the intrinsic model in Eq. 1:" + }, + { + "type": "equation", + "bbox": [ + 0.306, + 0.742, + 0.786, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m s e} ^ {\\hat {D}} = M S E \\left(D _ {H} ^ {*}, \\hat {D} _ {H}\\right), \\quad \\mathcal {L} _ {m s g} ^ {\\hat {D}} = M S G \\left(D _ {H} ^ {*}, \\hat {D} _ {H}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.778, + 0.729, + 0.792 + ], + "angle": 0, + "content": "Our final loss for the HDR albedo reconstruction network is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.805, + 0.786, + 0.824 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {A} = \\mathcal {L} _ {m s e} ^ {A} + \\gamma \\cdot \\mathcal {L} _ {m s g} ^ {A} + \\mathcal {L} _ {m s e} ^ {\\hat {D}} + \\gamma \\cdot \\mathcal {L} _ {m s g} ^ {\\hat {D}}, \\quad \\gamma = 0. 1. \\tag {11}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "S. Dille et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.357, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.144, + 0.499, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.144, + 0.643, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.144, + 0.786, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.22, + 0.357, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.22, + 0.5, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.22, + 0.644, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.22, + 0.785, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.304, + 0.785, + 0.346 + ], + "angle": 0, + "content": "Fig. 6: We show the intermediate results of our pipeline from an LDR input compared with its ground truth. The direct HDR reconstruction from the estimated HDR intrinsics can lead to small artifacts which are reduced by our refinement network." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.377, + 0.357, + 0.39 + ], + "angle": 0, + "content": "4.3 Refinement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.785, + 0.447 + ], + "angle": 0, + "content": "The outputs of our HDR reconstruction networks for shading and albedo can be combined using the intrinsic model to give us an initial estimation for the HDR image:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.447, + 0.785, + 0.478 + ], + "angle": 0, + "content": "\\[\n\\hat {I} _ {H} = A _ {H} * S _ {H}, \\quad S _ {H} = \\frac {1 - D _ {H}}{D _ {H}}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.785, + 0.557 + ], + "angle": 0, + "content": "However, as the two intrinsic components are estimated separately, the perfect reconstruction of the HDR image is not guaranteed as shown in Figure 6. In the final step of our pipeline, we train a refinement network that takes the intrinsic HDR reconstruction results as input and generates our final HDR reconstruction result." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.558, + 0.787, + 0.709 + ], + "angle": 0, + "content": "This network has a similar goal to the hallucination network used in Liu et al. [24]: Reconstructing the final HDR image by extending dynamic range. However, as it receives the intrinsic reconstruction of the HDR image readily as input, its task is significantly easier. Unlike a network trained for direct LDR-to-HDR conversion, it does not need to reason about the high-level cues in the image which is performed by our HDR shading reconstruction network. Similarly, it does not need to recover the corrupted colors in saturated regions, which is performed by our HDR albedo reconstruction network. It is simply tasked with combining the rich input information it receives to reconstruct the final HDR image, which allows it to generate more accurate and sharper results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.71, + 0.785, + 0.755 + ], + "angle": 0, + "content": "Following the inverse representation we adopt for shading, we use the inverse representation for the HDR images as a bounded representation that fits the long-tailed distribution in HDR images between \\([0,1]\\):" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.766, + 0.785, + 0.8 + ], + "angle": 0, + "content": "\\[\nJ _ {H} = \\frac {1}{I _ {H} + 1}, \\quad \\hat {J} _ {H} = \\frac {1}{\\hat {I} _ {H} + 1}, \\quad J _ {H} ^ {*} = \\frac {1}{I _ {H} ^ {*} + 1}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.808, + 0.785, + 0.84 + ], + "angle": 0, + "content": "where \\(J_{H}\\), and \\(J_{H}^{*}\\) denote the estimated and ground-truth inverse HDR images and \\(\\hat{J}_H\\) is the inferred inverse HDR image following Eq. 12." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.442, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Intrinsic Single-Image HDR Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "Our refinement network receives the LDR image \\( I_{L} \\), the inferred inverse HDR image \\( \\hat{J}_{H} \\), the estimated HDR inverse shading \\( D_{L} \\), and the estimated HDR albedo \\( A_{L} \\) as a concatenated \\( h\\times w\\times 10 \\) input. We use the MSE and multi-scale gradient losses to train our refinement network:" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.215, + 0.786, + 0.234 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m s e} ^ {J} = M S E \\left(J _ {H} ^ {*}, J _ {H}\\right), \\quad \\mathcal {L} _ {m s g} ^ {J} = M S G \\left(J _ {H} ^ {*}, J _ {H}\\right), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.24, + 0.654, + 0.259 + ], + "angle": 0, + "content": "with the final loss defined as \\(\\mathcal{L}^J = \\mathcal{L}_{mse}^J +\\gamma \\cdot \\mathcal{L}_{msg}^J\\) \\(\\gamma = 0.1\\)" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.275, + 0.548, + 0.292 + ], + "angle": 0, + "content": "4.4 Network architecture and training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.789, + 0.358 + ], + "angle": 0, + "content": "We utilize the same architecture for our 3 networks. We use EfficientNet [37] as the backbone with the decoder from [33], replacing the output activation with a sigmoid. We use RAdam [23] as the optimizer with a learning rate of \\(1e^{-4}\\) and CosineAnnealing [25] scheduling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.789, + 0.448 + ], + "angle": 0, + "content": "To train our intrinsic formulation, we use the provided synthetic intrinsic ground truth from Hypersim [35] and construct ground truth intrinsics from the raw images of the MultiIllum dataset [31] as described in [3]. We follow the preprocessing steps from Liu et al. [24] to create clipped, linearized LDR images with randomly sampled exposure values \\( t \\in [-3..3] \\) from the ground truth data which we then decompose into our LDR input intrinsics using [3]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.449, + 0.79, + 0.555 + ], + "angle": 0, + "content": "Our refinement network in contrast does not require supervision in intrinsic space. We can thus extend our training set by real images from MultiRAW [22], LSMI [16], and the HDR-Real training subset from Liu et al. [24]. For this stage, we follow the ISP model in [24] to convert the ground truth images into JPEG files, use their pre-trained networks to retrieve linearized LDR images, and again retrieve the intrinsic decomposition via [3]. We provide a more detailed description of our training procedure in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.574, + 0.377, + 0.592 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.603, + 0.789, + 0.68 + ], + "angle": 0, + "content": "In this section, we present a thorough evaluation of our proposed method using multiple standard benchmarks. First, we discuss the datasets, metrics and prior work used in our experimental setup, then we present qualitative and quantitative comparisons to state-of-the-art approaches. We present a more detailed evaluation and our ablation studies in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.698, + 0.457, + 0.714 + ], + "angle": 0, + "content": "5.1 Quantitative evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "To quantitatively evaluate our approach against existing methods we follow the evaluation protocol proposed by Hanji et al. [15]. For a given ground truth HDR image, we follow the preprocessing of Liu et al. [24] and map the values into the range \\([1 - 1000]\\). We determine the best scale to align the input reconstruction with the ground truth with least-squares, using only values between the \\(10^{th}\\) and \\(90^{th}\\) percentiles to avoid outliers. We then perform CRF correction to match the reconstruction to the ground truth by fitting a per-channel third-degree polynomial in log-RGB space." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "S. Dille et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Quantitative results against state-of-the-art using the HDR-VDP3 metric. Our method achieves state-of-the-art performance on the difficult high-definition SIHDR benchmark dataset. We are state-of-the-art or competitive on all other datasets." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.199, + 0.751, + 0.355 + ], + "angle": 0, + "content": "
MethodSIHDRHDR-EYEHDR-REALHDR-SYNTHRAISE
Ours8.968.967.578.218.89
DrTMO [11]8.278.727.578.278.64
HDR-CNN [10]8.398.616.708.038.47
ExpandNet [30]8.678.496.977.698.27
Single-HDR [24]8.798.967.478.518.95
Mask-HDR [36]8.258.607.108.148.44
HDRUNet [6]8.828.465.717.618.04
Multi-Exp Gen. [19]8.648.877.348.168.66
Lightweight [14]8.268.207.037.317.68
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.506 + ], + "angle": 0, + "content": "Metrics We follow the guidance of Hanji et al. [15] when choosing metrics to evaluate our approach. In their study, they find that certain metrics better correlate with results from qualitative experiments. We focus our evaluation on the popular HDR-VDP3 [28] metric as it is purpose-built for HDR evaluation and models the subjective evaluation of a human observer. We specifically use the \"quality\" task of HDR-VDP3. Furthermore, we evaluate multiple recommended SDR metrics using the PU21 [27] encoding and provide those results in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.787, + 0.565 + ], + "angle": 0, + "content": "Datasets We evaluate our method on five publicly available datasets. We use the recent high-resolution SI-HDR Benchmark [15] as well as commonly used real-world datasets HDR-Eye [32], HDR-Real [24], HDR-Synth [24], and RAISE [7]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.716 + ], + "angle": 0, + "content": "We compare our method against a collection of baselines that share their implementation publicly. Table 1 shows the results of our quantitative evaluation. Each value represents the average HDR-VDP3 score in Just-Objectionable-Differences (JOD) units. We show a significant improvement in quality in the SI-HDR Benchmark [15] while achieving state-of-the-art performance on 3 of the other 4 datasets. As our qualitative evaluation in the next section and our results presented in the supplementary material demonstrate, the favorable performance of our method comes from our physically-motivated intrinsic formulation, allowing us to reliably reconstruct the high luminance details and recover lost colors. We provide additional metrics and analysis in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.739, + 0.445, + 0.755 + ], + "angle": 0, + "content": "5.2 Qualitative evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.841 + ], + "angle": 0, + "content": "We present qualitative comparisons in Figures 7 and 8, showing the differences in both detail and color reconstruction. On SI-HDR [15], prior works struggle with predicting accurate details around edges where over- and under-exposed pixels meet. This becomes notable around the leaves in the [roof] scene with visible halos and color shifts. On the floor in the [book] scene, most competitors fail to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.442, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Intrinsic Single-Image HDR Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.143, + 0.788, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.642, + 0.788, + 0.673 + ], + "angle": 0, + "content": "Fig. 7: We show results from our method on SI-HDR [15] to the other baselines. We refer to Sec. 5.2 and the supplementary material for an in-depth discussion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.842 + ], + "angle": 0, + "content": "reconstruct the homogenous surface and hallucinate patterns and color blobs. In both cases, our method seamlessly recovers missing details in over-exposed areas without artifacts. For the woman in Fig. 8, HDRUnet [6] and SingleHDR [24] both exhibit strong artifacts as bright spots in the hand and the face, similar to the faces of the [bust] and the [kids]. Our method in contrast can recover the correct color information due to our modeling of the albedo. The results showcase the advantage of the intrinsic formulation for creating visually pleasing results. We provide an extensive qualitative comparison in the supplementary material." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "S. Dille et al." + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.359, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.145, + 0.501, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.145, + 0.643, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.145, + 0.785, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.219, + 0.359, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.219, + 0.501, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.219, + 0.643, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.219, + 0.785, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.308, + 0.331, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.308, + 0.445, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.308, + 0.558, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.308, + 0.672, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.308, + 0.785, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.405, + 0.331, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.405, + 0.445, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.405, + 0.558, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.405, + 0.672, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.405, + 0.785, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.537 + ], + "angle": 0, + "content": "Fig.8: We compare results from our method with a focus on color reconstruction against HDRUnet [6] and SingleHDR [24]. We refer to Sec. 5.2 and the supplementary material for an in-depth discussion." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.565, + 0.36, + 0.581 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.751 + ], + "angle": 0, + "content": "In this work, we remodel the HDR reconstruction problem in the intrinsic domain and propose a multi-stage pipeline to recover lost details. The luminance range extension can be conveniently modeled in the shading space which allows the network to better model the correlation between illumination and geometry. Similarly, neural networks can easily take advantage of albedo sparsity to recover lost color information with high fidelity. Through qualitative and quantitative evaluations, we show that our physically-inspired modeling of the HDR reconstruction problem achieves state-of-the-art performance, improving upon prior work in terms of both detail recovery and color reproduction in a wide variety of in-the-wild photographs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.776, + 0.405, + 0.794 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We acknowledge the support of the Natural Sciences and Engineering Research Council of Canada (NSERC), [RGPIN-2020-05375]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.442, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Intrinsic Single-Image HDR Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Banterle, F., Ledda, P., Debattista, K., Chalmers, A.: Inverse tone mapping. In: Proc.GRAPHITE (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.232 + ], + "angle": 0, + "content": "2. Barrow, H., Tenenbaum, J., Hanson, A., Riseman, E.: Recovering intrinsic scene characteristics. Comput. vis. syst 2(3-26), 2 (1978)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.785, + 0.26 + ], + "angle": 0, + "content": "3. Careaga, C., Aksoy, Y.: Intrinsic image decomposition via ordinal shading. ACM Trans. Graph. 43(1) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.287 + ], + "angle": 0, + "content": "4. Careaga, C., Miangoleh, S.M.H., Aksoy, Y.: Intrinsic harmonization for illumination-aware compositing. In: Proc. SIGGRAPH Asia (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.328 + ], + "angle": 0, + "content": "5. Chen, S.K., Yen, H.L., Liu, Y.L., Chen, M.H., Hu, H.N., Peng, W.H., Lin, Y.Y.: CEVR: Learning continuous exposure value representations for single-image HDR reconstruction. In: Proc. ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.785, + 0.356 + ], + "angle": 0, + "content": "6. Chen, X., Liu, Y., Zhang, Z., Qiao, Y., Dong, C.: HDRUnet: Single image HDR reconstruction with denoising and dequantization. In: Proc. CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.358, + 0.785, + 0.384 + ], + "angle": 0, + "content": "7. Dang-Nguyen, D.T., Pasquini, C., Conotter, V., Boato, G.: RAISE: a raw images dataset for digital image forensics. In: Proc. MMSys (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.785, + 0.412 + ], + "angle": 0, + "content": "8. Debevec, P.: A median cut algorithm for light probe sampling. In: ACM SIGGRAPH 2005 Posters. ACM (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.413, + 0.785, + 0.439 + ], + "angle": 0, + "content": "9. Debevec, P.E., Malik, J.: Recovering high dynamic range radiance maps from photographs. In: ACM Trans. Graph. (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.785, + 0.481 + ], + "angle": 0, + "content": "10. Eilertsen, G., Kronander, J., Denes, G., Mantiuk, R.K., Unger, J.: HDR image reconstruction from a single exposure using deep CNNs. ACM Trans. Graph. 36(6) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.785, + 0.509 + ], + "angle": 0, + "content": "1. Endo, Y., Kanamori, Y., Mitani, J.: Deep reverse tone mapping. ACM Trans. Graph. 36(6) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.51, + 0.785, + 0.536 + ], + "angle": 0, + "content": "2. Garces, E., Rodriguez-Pardo, C., Casas, D., Lopez-Moreno, J.: A survey on intrinsic images: Delving deep into lambert and beyond. Int. J. Comput. Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.785, + 0.564 + ], + "angle": 0, + "content": "3. Gilchrist, A., Jacobsen, A.: Perception of lightness and illumination in a world of one reflectance. Perception 13, 5-19 (02 1984)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.592 + ], + "angle": 0, + "content": "4. Guo, C., Xiuhua, J.: LHDR: HDR reconstruction for legacy content using a lightweight DNN. In: Proc. ACCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.593, + 0.785, + 0.632 + ], + "angle": 0, + "content": "5. Hanji, P., Mantiuk, R., Eilertsen, G., Hajisharif, S., Unger, J.: Comparison of single image HDR reconstruction methods — the caveats of quality assessment. In: ACM Trans. Graph. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "6. Kim, D., Kim, J., Nam, S., Lee, D., Lee, Y., Kang, N., Lee, H.E., Yoo, B., Han, J.J., Kim, S.J.: Large scale multi-illuminant (LSMI) dataset for developing white balance algorithm under mixed illumination. In: Proc. CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "7. Kim, J., Lee, S., Kang, S.J.: End-to-end differentiable learning to HDR image synthesis for multi-exposure images. In: Proc. AAAI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.785, + 0.73 + ], + "angle": 0, + "content": "8. Kovaleski, R.P., Oliveira, M.M.: High-quality reverse tone mapping for a wide range of exposures. In: Proc. SIBGRAPI (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.785, + 0.757 + ], + "angle": 0, + "content": "9. Le, P.H., Le, Q., Nguyen, R., Hua, B.S.: Single-image HDR reconstruction by multi-exposure generation. In: Proc. WACV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "20. Lee, S., An, G.H., Kang, S.J.: Deep recursive HDRI: Inverse tone mapping using generative adversarial networks. In: Proc. ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.812 + ], + "angle": 0, + "content": "21. Li, Z., Snavely, N.: Learning intrinsic image decomposition from watching the world. In: Proc. CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "22. Li, Z., Lu, M., Zhang, X., Feng, X., Asif, M.S., Ma, Z.: Efficient visual computing with camera raw snapshots. IEEE Trans. Pattern Anal. Mach. Intell. (2024)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "S. Dille et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.786, + 0.175 + ], + "angle": 0, + "content": "23. Liu, L., Jiang, H., He, P., Chen, W., Liu, X., Gao, J., Han, J.: On the variance of the adaptive learning rate and beyond. In: Proc. ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.786, + 0.217 + ], + "angle": 0, + "content": "24. Liu, Y.L., Lai, W.S., Chen, Y.S., Kao, Y.L., Yang, M.H., Chuang, Y.Y., Huang, J.B.: Single-image HDR reconstruction by learning to reverse the camera pipeline. In: Proc. CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.786, + 0.244 + ], + "angle": 0, + "content": "25. Loshchilov, I., Hutter, F.: SGDR: Stochastic gradient descent with warm restarts. In: Proc. ICLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.786, + 0.285 + ], + "angle": 0, + "content": "26. Mann, S., Picard, R.: On being \"undigital\" with digital cameras: Extending dynamic range by combining differently exposed pictures. In: Proc. IS&T Annual Conference (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.786, + 0.314 + ], + "angle": 0, + "content": "27. Mantiuk, R.K., Azimi, M.: Pu21: A novel perceptually uniform encoding for adapting existing quality metrics for hdr. In: Proc. PCS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.355 + ], + "angle": 0, + "content": "28. Mantiuk, R.K., Hammou, D., Hanji, P.: HDR-VDP-3: A multi-metric for predicting image differences, quality and contrast distortions in high dynamic range and regular content. arXiv preprint arXiv:2304.13625 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.786, + 0.383 + ], + "angle": 0, + "content": "29. Maralan, S.S., Careaga, C., Aksoy, Y.: Computational flash photography through intrinsics. In: Proc. CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.786, + 0.424 + ], + "angle": 0, + "content": "30. Marnerides, D., Bashford-Rogers, T., Hatchett, J., Debattista, K.: Expandnet: A deep convolutional neural network for high dynamic range expansion from low dynamic range content. Comput. Graph. Forum 37(2) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.786, + 0.452 + ], + "angle": 0, + "content": "31. Murmann, L., Gharbi, M., Aittala, M., Durand, F.: A multi-illumination dataset of indoor object appearance. In: Proc. ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.479 + ], + "angle": 0, + "content": "32. Nemoto, H., Korshunov, P., Hanhart, P., Ebrahimi, T.: Visual attention in LDR and HDR images. In: Proc. VPQM (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.786, + 0.521 + ], + "angle": 0, + "content": "33. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Trans. Pattern Anal. Mach. Intell. 44(3) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.786, + 0.563 + ], + "angle": 0, + "content": "34. Rempel, A.G., Trentacoste, M., Seetzen, H., Young, H.D., Heidrich, W., Whitehead, L., Ward, G.: LDR2HDR: on-the-fly reverse tone mapping of legacy video and photographs. ACM Trans. Graph. 26(3) (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.786, + 0.604 + ], + "angle": 0, + "content": "35. Roberts, M., Ramapuram, J., Ranjan, A., Kumar, A., Bautista, M.A., Paczan, N., Webb, R., Susskind, J.M.: Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In: Proc. ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.786, + 0.632 + ], + "angle": 0, + "content": "36. Santos, M.S., Ren, T.I., Kalantari, N.K.: Single image HDR reconstruction using a CNN with masked features and perceptual loss. ACM Trans. Graph. 39(4) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.786, + 0.659 + ], + "angle": 0, + "content": "37. Tan, M., Le, Q.: EfficientNet: Rethinking model scaling for convolutional neural networks. In: Proc. ICML (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.786, + 0.701 + ], + "angle": 0, + "content": "38. Wang, C., Serrano, A., Pan, X., Chen, B., Myszkowski, K., Seidel, H.P., Theobalt, C., Leimkuhler, T.: Glowgan: Unsupervised learning of HDR images from LDR images in the wild. In: Proc. ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.786, + 0.729 + ], + "angle": 0, + "content": "39. Xu, D., Doutre, C., Nasiopoulos, P.: Correction of clipped pixels in color images. IEEE Trans. Vis. Comput. Graph. 17(3) (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.786, + 0.756 + ], + "angle": 0, + "content": "40. Zhang, N., Ye, Y., Zhao, Y., Wang, R.: Revisiting the stack-based inverse tone mapping. In: Proc. CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.786, + 0.784 + ], + "angle": 0, + "content": "41. Zhang, X., Brainard, D.H.: Estimation of saturated pixel values in digital color imaging. J. Optical Soc. Am. A 21(12) (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.786, + 0.812 + ], + "angle": 0, + "content": "42. Zhang, Y., Aydin, T.: Deep HDR estimation with generative detail reconstruction. Comput. Graph. Gorum 40(2), 179-190 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.786, + 0.812 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Intrinsic Single-Image HDR Reconstruction/e55f56f5-9c00-4140-aab7-15323b235acd_origin.pdf b/2024/Intrinsic Single-Image HDR Reconstruction/e55f56f5-9c00-4140-aab7-15323b235acd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ba642b68341e6aef99320455164c82b9a296a653 --- /dev/null +++ b/2024/Intrinsic Single-Image HDR Reconstruction/e55f56f5-9c00-4140-aab7-15323b235acd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70043e8b986f21d60efe6d29b73c23369990194aaf8f22cd0d81e0242480603b +size 15432628 diff --git a/2024/Intrinsic Single-Image HDR Reconstruction/full.md b/2024/Intrinsic Single-Image HDR Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dc71efe38f256b029025b323cb289e90612931de --- /dev/null +++ b/2024/Intrinsic Single-Image HDR Reconstruction/full.md @@ -0,0 +1,373 @@ +# Intrinsic Single-Image HDR Reconstruction + +Sebastian Dille\* , Chris Careaga\* , and Yaigiz Aksoy + +Simon Fraser University Burnaby, BC V5A 1S6, Canada sdille@sfu.ca + +![](images/b94aba13b10ed44f5cd7842782273070a4e0b5f3b0f4da4bd2bfc82730171a53.jpg) + +![](images/b88693e2b9c97e4dfc39d2328f535fc12116cd4a23d641c13b60fecb3d2bf424.jpg) + +![](images/bd1578a4ae706bc6d0c677422a09c43cf1b0444bfd5cc81ff7cbee8b3f31f08d.jpg) + +![](images/9d3fd8a0ef2fa1af403dd8aa883eebc509b42915a2b141639b77fce4939408f3.jpg) + +![](images/c081a015757e0cb57756c84b3da7b4f32262419542c417dab40e292d7a1c8982.jpg) + +![](images/76c62823aae19e48d7a890e5542c343b2e24f385e03a5a5f49b965d2f4a81444.jpg) + +![](images/e000213ad39461b9b479c23ae02d2b0756e7b816cb53c004ea627184c328d587.jpg) + +![](images/15c9daa8008848db1321d6a6d791ed07a2279d495f9d797b75ff39f211f88b79.jpg) +Fig. 1: In this work, we propose a physically-inspired HDR reconstruction pipeline by conducting reconstruction individually on intrinsic components. Our final reconstruction, as a result, is effective in recovering the high-luminance details and the lost colors. + +![](images/e4b7e6972ca5de426dd1bec7af395c005798d28b324822a919b5174eb36214c9.jpg) + +![](images/87b751f9e5da6c0d1fcda4f0f9eb51785f4015a147b659f8352c1794c9e07103.jpg) + +![](images/7f8ea3dcfb895c1bd0381bd828404406f3daa497d63797c8d1b67261dc255126.jpg) + +![](images/24ba8bd9a7bfadc70539f9bfe5fcbfeef396f180752655a2f1112d615f475773.jpg) + +![](images/135f1be94acde2799ce37e6399e2feb683f080196b5c53342568a8f1ae22d11a.jpg) + +![](images/ccda3e24fb4683a5a0dd6c7fdb70494104b6b770a7d6bd6b68e48904339b34e9.jpg) + +Abstract. The low dynamic range (LDR) of common cameras fails to capture the rich contrast in natural scenes, resulting in loss of color and details in saturated pixels. Reconstructing the high dynamic range (HDR) of luminance present in the scene from single LDR photographs is an important task with many applications in computational photography and realistic display of images. The HDR reconstruction task aims to infer the lost details using the context present in the scene, requiring neural networks to understand high-level geometric and illumination cues. This makes it challenging for data-driven algorithms to generate accurate and high-resolution results. In this work, we introduce a physically-inspired remodeling of the HDR reconstruction problem in the intrinsic domain. The intrinsic model allows us to train separate networks to extend the dynamic range in the shading domain and to recover lost color details in the albedo domain. We show that dividing the problem into two simpler sub-tasks improves performance in a wide variety of photographs. + +Keywords: HDR Reconstruction $\cdot$ Inverse Tonemapping $\cdot$ Computational Photography + +# 1 Introduction + +Real-world scenes contain a wider range of luminance than most consumer-grade cameras can record or store. As a result, everyday photographs can only represent a low dynamic range (LDR) version of the scene that lacks details and color in very dark or bright regions. High dynamic range (HDR) imaging aims to capture and store the full contrast present in the scene. HDR images allow realistic display of photographs using high-contrast displays as well as many computational photography applications such as tone mapping and illumination-aware image editing. However, HDR imaging typically requires multi-exposure image stacks which limits their applications to static scenes. + +To extend the HDR-enabled applications to everyday photographs, there has been a growing interest in methods to estimate HDR images from only LDR photographs called single-image HDR reconstruction, or LDR-to-HDR conversion. Single-image HDR reconstruction aims to extend the dynamic range of LDR photographs by recovering the details lost during capture. The missing information in LDR images is typically characterized by saturated pixels in bright regions, lacking textural details and color. This makes HDR reconstruction a high-level problem, requiring networks to utilize geometric and context-dependent cues in the scene to infer the lost information. Formulating this reconstruction problem using standard color spaces such as RGB or Luv requires the networks to statistically model the complex relationship between luminance, surface colors, and geometry, which limits their performance. + +In this work, we present a physically-inspired remodeling of the HDR reconstruction problem in the intrinsic domain. Intrinsic image decomposition is a mid-level vision problem that separates the effect of the illumination in the scene from the reflectance of surfaces [3], representing the input as a combination of a shading and an albedo layer. The shading of an image is highly correlated with the scene geometry and illumination, while the albedo contains the color information and the high-frequency textures. The saturated pixels we aim to recover create different degradations in the shading and albedo layers. The clipping of high values is observed in the shading, while the loss of color information causes desaturation in the albedo. We propose an HDR reconstruction pipeline that individually addresses the loss of information in the LDR image in the two intrinsic components. Our HDR shading reconstruction network extends the dynamic range, taking advantage of the high correlation between shading and the scene geometry. Our LDR albedo reconstruction network, on the other hand, recovers the lost color information in saturated regions. A final refinement stage combines the two components to generate our full HDR reconstruction. Figure 1 shows different stages of our pipeline. + +In our formulation, we take advantage of the intrinsic image formation model by dividing the recovery of lost details and colors into two separate tasks which are more convenient for networks to model. We show that our physically inspired approach can generate high-fidelity details with rich colors in complex in-the-wild scenes. We evaluate our method qualitatively and quantitatively to demonstrate the state-of-the-art performance of our formulation in a wide variety of scenes. + +# 2 Related work + +Common imaging sensors for photography are inherently limited in the portion of the scene's dynamic range they can capture in a single shot. This results in saturated pixels in bright regions, representing a significant challenge in the capture and realistic display of real-world scenes. + +Multi-exposure fusion methods can reconstruct an HDR image by combining a stack of LDR images taken at different exposures [9,26,40]. Their formulation successfully addresses contrast range extension and reversal of the quantization and nonlinearity of LDR images. The multi-exposure capture, however, assumes a static camera and scene, requiring a challenging alignment process and limiting their application scenario. To enable HDR-related applications in everyday photography, single-image HDR reconstruction approaches aim to recover the lost information directly from an LDR image. Earlier approaches apply specially designed filters [18,34] or median-cut [1,8] to fill in overexposed regions, using different heuristics to scale the intensity of a selected area. Similarly, Bayesian [41] and graph-based formulations [39] have been introduced for chroma restoration. Due to the high-level nature of the reconstruction problem, these low-level approaches struggle to generate realistic details in real-world scenarios. + +In their concurrent publications, Eilertsen et al. [10] and Endo et al. [11] present the first data-driven approaches to the single-image HDR reconstruction problem. They train their networks for direct conversion to HDR from the LDR input, addressing quantization, nonlinearity, and saturation in the LDR image in a single step. Many works in recent literature adapt this unified inverse model [24] and develop tailored solutions for underexposed [6,14] or overexposed [36] areas or losses for color recovery [30]. The direct estimation of HDR outputs, however, creates the challenge of an unconstrained color distribution that hinders training efficiency. Several methods [5, 11, 17, 19, 20, 40] instead adopt the concept of exposure stacks and predict a set of over- and underexposed images from the input LDR image. With each prediction still in the LDR domain, the HDR image can later be created with multi-exposure fusion [9]. Other methods [38,42] employ generative approaches, either generating separate LDR detail layers [42] or full HDR images directly [38], with image-specific HDR reconstruction via GAN inversion. Liu et al. [24] approach HDR reconstruction by inverting the in-camera image formation process, modeling the dequantization, nonlinearization, and saturated pixel reconstruction in three separate steps. They show that this decomposition of the task simplifies the modeling of each individual sub-problem. + +In this work, we adopt the modular formulation of Liu et al. [24] and focus on recovering the lost details and color information in the saturated pixels. Our method takes inspiration from prior works that formulate computational photography problems in the intrinsic image domain [4, 29]. Rather than defining a direct mapping from LDR to HDR, we divide this task into two sub-problems: shading reconstruction that extends the dynamic range, and reflectance reconstruction that recovers the color information. Our physical deconstruction of the problem significantly improves the recovery of details and color information. + +![](images/d25827af201bb1f1214a21f93d29b01689f538979860505e0ca505e8b85648b7.jpg) +Fig. 2: The in-camera pipeline modeling LDR image formation is shown at the top. Liu et al. [24] propose the inverse pipeline shown on the bottom, addressing each type of information loss individually in three steps. In this work, we use the pre-trained networks of Liu et al. [24] for the dequantization and linearization tasks (shown in pink) and focus on developing an intrinsic model for the dynamic range reconstruction problem highlighted in teal. + +# 3 Background and Motivation + +The focus of our work is the formulation of the single-image LDR-to-HDR reconstruction problem in the intrinsic domain. We begin our discussion by first introducing the LDR imaging pipeline and the intrinsic model of image formation. We then draw connections between the two models and motivate our intrinsic HDR formulation. + +# 3.1 LDR image formation and HDR reconstruction + +The acquisition of standard RGB photographs involves a series of operations in the camera-internal image processing pipeline, or ISP, that result in different forms of information loss. The initial acquisition of the image is in a raw format, where the luminance received by the sensor from a real-world scene is scaled and stored in 16- or 32-bit data structures. The scale of the captured raw values is determined by the analog camera settings such as the shutter speed, the aperture size, and the ISO. Given the natural high dynamic range of real-world scenes, the raw values are clipped at a maximum due to saturation at the sensor level. As shown in Figure 2, the remainder of the ISP applies nonlinear color and gamma corrections before finally decreasing the bit rate to 8 bits per channel for storage. + +The LDR image formation process causes three different types of corruption of the HDR information: quantization, non-linearization, and clipping. Single-image HDR reconstruction methods, as a result, have to address all three of these challenges. While many methods in the literature address all three simultaneously within a single architecture, Liu et al. [24] design their pipeline to address each type of information loss with a separate dedicated network. They propose a dequantization network to first increase the bit depth of the input LDR image and then estimate a CRF curve for linearization. Both tasks are convenient for a network to learn, taking advantage of the ease of dataset generation for dequantization and a simple parametric curve estimation for linearization. + +![](images/454f9ce86cdedfe2e9162afe62b8df247d96ff76fb29929baade7e7384a17aaf.jpg) +Fig. 3: We show the different effects of clipping on the albedo and the shading for a tonemapped HDR image and its LDR counterpart. When clipping the image, the LDR shading loses information compared to HDR. We scale the intensity of both shadings to show the effect. At the same time, the LDR albedo becomes saturated. + +They then reconstruct the values lost to clipping with a third network in what they call the hallucination stage. This last stage of the inverse pipeline addresses the most significant challenge in single-image HDR reconstruction, namely the recovery of the highlights and the color information in clipped regions. Diverging from Liu et al.'s terminology, we refer to this stage as dynamic range extension. We adopt Liu et al.'s inverse pipeline and use their pre-trained networks for dequantization and linearization. In the following, we focus on the dynamic range extension problem and present our solution using an intrinsic model of image formation. + +# 3.2 Intrinsic image decomposition + +Intrinsic image decomposition is a fundamental mid-level computer vision problem that represents a core step in inverse rendering pipelines [2, 12]. It aims to separate the image into illumination and reflectance components following the physical image formation model. Adopting the commonly used Lambertian shading assumption with a single-channel shading representation [3], the intrinsic equation can be written as: + +$$ +I _ {r g b} = A _ {r g b} \cdot S, \tag {1} +$$ + +where $I$ represents the linear image reflecting relative luminance, $A$ represents albedo, i.e. the surface reflectance containing the color information, and $S$ represents shading reflecting the effect of illumination in the scene. + +Shading and albedo represent two orthogonal characteristics of a scene, one containing the effect of light and its interaction with the scene geometry and the other representing material properties. As a result, the two intrinsic components have very different characteristics as demonstrated in Figure 3. Shading is highly correlated with surface orientations, appearing smooth over continuous surfaces + +and having a high gradient in depth discontinuities and shadow boundaries. Albedo, on the other hand, contains the high-frequency textures present on the materials in the scene while having a sparse set of values per image as the variation in brightness due to geometry is isolated in the shading layer. In the single-channel shading representation, albedo also contains any variation caused by colorful illumination [3]. + +# 3.3 Infrinsics in LDR and HDR + +The bounded representation of LDR images limits the range of luminance values they can contain. HDR reconstruction aims to extend this range of luminance by exploiting statistical properties of real-world photographs derived from high-level scene properties such as geometry and appearance. + +A brief look at the intrinsic components of HDR images, as shown in Figure 3, reveals the different characteristics of the shading and albedo layers. The distribution of reflectance values in real-world scenes typically has a narrow distribution, making the albedo a variable that can be represented linearly between [0, 1]. The high dynamic range of the HDR image, on the other hand, is reflected in the shading layer that similarly has an unbounded representation. + +As a result, the limited dynamic range of LDR images has different effects on the shading and albedo layers. As Figure 3 demonstrates, the clipping of HDR values results in a similar clipping in shading, showing the relationship between the dynamic range and illumination that has been previously discussed in the literature [10]. The loss of colors induced by clipping, on the other hand, appears in the albedo layer as the dullification of colors in over-exposed regions. + +Based on this observation, we formulate the HDR reconstruction problem separately in shading and albedo. In our setup, the network tasked with converting LDR shading into HDR shading models the high dynamic range of the illumination in the scene, taking advantage of the high correlation between shading and high-level scene properties such as geometry. We task another network with converting the LDR albedo into HDR albedo, training it to recover the lost color information in clipped regions. We detail our full pipeline in Section 4. + +Our approach effectively separates the two challenges in dynamic range reconstruction, reconstruction of high luminance regions and recovery of the color information, into two separate tasks. Formulating the reconstruction of high luminance in the shading domain results in a more convenient way for the network to statistically model the problem. This can be attributed to the high correlation between shading and geometry, which is often violated in luminance due to textured surfaces as highlighted in Figure 5. Formulating the color recovery problem in the albedo domain simplifies the task similarly, allowing the exploitation of the sparsity of reflectance values in real-world scenes. Our albedo LDR-to-HDR network, as a result, acts comparable to a simple inpainting model, filling in the lost color details based on the scene content. + +![](images/da5ceac200d5241d3e3a7849ce0f20133336571ed2dc04095054c3cff64baef7.jpg) +Fig. 4: Our HDR reconstruction pipeline starts with the intrinsic decomposition of the image using an off-the-shelf method [3]. The LDR reconstruction is then done individually on shading and albedo layers, which are then combined in a final refinement stage. The input LDR image is provided to each network shown in blue. + +# 4 Dynamic Range Extension via Infrinics + +In this work, we propose a physically inspired system for dynamic range extension in the single-image HDR reconstruction pipeline. The input to our pipeline is the dequantized and linearized LDR photograph $I_{L}$ . We first compute the intrinsic components of $I_{L}$ , the LDR shading $S_{L}$ , and albedo $A_{L}$ , using the in-the-wild intrinsic decomposition method by Careaga and Aksoy [3]. We then estimate an HDR shading and an HDR albedo from the corresponding LDR components using dedicated networks. The estimated HDR intrinsic components are then combined and refined by a subsequent HDR reconstruction network yielding our final result. We present an overview of our pipeline in Figure 4. In this section, we present our formulation and training details for each step. + +# 4.1 HDR shading reconstruction + +The shading layer in the intrinsic model contains the effect of illumination in the scene. The high dynamic range of illumination in the scene, shadows, and specular objects cause a long-tailed distribution of values in the shading map. An LDR input with clipped values similarly produces a clipped LDR shading layer after intrinsic decomposition. We train our HDR shading reconstruction network to extend the range of shading values beyond this clipping. + +Several works in the literature define the dynamic range extension task in the luminance domain, targeting a similar goal of isolating the effect of illumination. In their work, Eilertsen et al. [10] motivate this choice from a perceptual standpoint, citing earlier work that explores the effect of illumination versus surface colors in human perception [13]. The luminance, however, is computed as a linear combination of RGB channels and, as a result, still contains information on the reflectance. Shading, on the other hand, directly reflects the illumination and is accordingly highly correlated with the scene geometry as Figure 5 demonstrates. Defining the range extension task in shading instead of luminance benefits the network to model the clipped shading values by easily utilizing the geometry-related cues in the input. + +![](images/4cda54777a5382be486b73abf80ab6cb1963dccf4ca8196cd197edaf12e7c947.jpg) +Fig. 5: We show the difference between the shading representation and the L-channel from LUV. The shading correlates well with the geometry, while the L-channel includes texture information as shown in the green insets. + +![](images/65935f5a1892b83702eebc1741c8029c66906577e9aa840887a5ee7c134b1732.jpg) + +![](images/8ce2a5991430227ba6f7e98ee4ed108a01c24c99464e0f68a683e1d4fb99db10.jpg) + +![](images/a941d2858a6c5af0c8e6af13b33c8346c21d14a8fc8cf16ccbd16728fd1e6a2c.jpg) + +![](images/9d7215e652dd67b12075570a4f312a9697ed8effebabb29875416c0282d03e3f.jpg) + +![](images/6667fea6c75d0a69fc8f5f9eb02570232249df2c186dea6d2790e3cb3f2c8ec6.jpg) + +Following Careaga and Aksoy [3], we represent shading in the inverse domain that fits the long-tailed shading in a representation bounded in $[0,1]$ : + +$$ +D _ {L} = \frac {1}{S _ {L} + 1}. \tag {2} +$$ + +The motivation behind this inverse representation is similar to the use of the log space in HDR reconstruction: It allows the definition of balanced loss functions on variables that could take very large or small values otherwise. The advantage of the inverse domain over the log space lies in its bounded nature that allows using the variable as input without clipping or normalization as well as the use of bounded activation functions such as sigmoid. + +Our HDR shading reconstruction network receives the LDR image $I_{L}$ and the estimated LDR inverse shading $D_{L}$ as a concatenated $h\times w\times 4$ input. Its output is defined as the HDR inverse shading $D_{H}$ . We use multiple loss functions to train our network. The first one is the standard mean-squared error loss defined in the inverse space: + +$$ +\mathcal {L} _ {m s e} ^ {D} = M S E \left(D _ {H} ^ {*}, D _ {H}\right), \tag {3} +$$ + +where $D_H^*$ and $D_H$ denote the ground-truth and estimated HDR inverse shading, respectively. We also use the multi-scale gradient loss [21] as an edge-aware smoothness loss: + +$$ +\mathcal {L} _ {m s g} ^ {D} = M S G \left(D _ {H} ^ {*}, D _ {H}\right) = \frac {1}{N} \sum_ {m} S A D \left(\nabla D _ {H} ^ {* m}, \nabla D _ {H} ^ {m}\right), \tag {4} +$$ + +where $\nabla$ denotes the spatial gradient and $D_H^m$ denotes the estimated inverse HDR shading at scale $m$ . + +Although the output of our HDR shading reconstruction network is defined only as the HDR inverse shading, we can use the ground-truth HDR image $I_H^*$ and the intrinsic model in Eq. 1 to compute an implied HDR albedo $\hat{A}_H$ [3]: + +$$ +\hat {A} _ {H} = \frac {I _ {H} ^ {*}}{S _ {H}}, \quad S _ {H} = \frac {1 - D _ {H}}{D _ {H}}. \tag {5} +$$ + +We define the MSE and multi-scale gradient loss also for the implied albedo: + +$$ +\mathcal {L} _ {m s e} ^ {\hat {A}} = M S E \left(A _ {H} ^ {*}, \hat {A} _ {H}\right), \quad \mathcal {L} _ {m s g} ^ {\hat {A}} = M S G \left(A _ {H} ^ {*}, \hat {A} _ {H}\right), \tag {6} +$$ + +and backpropagate these losses to our network through Eq. 5. + +Our HDR shading reconstruction network only estimates the HDR shading and the HDR albedo component will be estimated with a separate network defined in the next section. Estimating the HDR intrinsic components separately results in possible reconstruction errors when these components are combined to produce an estimation for the HDR image. While we will refine the HDR reconstruction in our final stage with a dedicated network, we define these losses on the implied albedo to promote a better fit within the intrinsic model. + +Our final loss for the HDR shading reconstruction network is defined as: + +$$ +\mathcal {L} ^ {D} = \mathcal {L} _ {m s e} ^ {D} + \gamma \cdot \mathcal {L} _ {m s g} ^ {D} + \mathcal {L} _ {m s e} ^ {\hat {A}} + \gamma \cdot \mathcal {L} _ {m s g} ^ {\hat {A}}, \quad \gamma = 0. 1. \tag {7} +$$ + +# 4.2 HDR albedo reconstruction + +The clipping in the LDR image formation process results in the loss of color information in the saturated pixels of the LDR albedo map. Our HDR albedo reconstruction network is tasked with recovering the corrupted colors. + +While the LDR shading map has saturated values, signaling the regions to be reconstructed to our HDR shading reconstruction networks, the LDR albedo map does not carry the same signal as albedo maps are not saturated. To denote the regions that require color reconstruction, we provide the network with a soft guidance mask $\alpha$ defined as: + +$$ +\alpha = \frac {\operatorname {m a x} \left(0 , I _ {L} - \lambda\right)}{1 - \lambda}, \quad \lambda = 0. 8, \tag {8} +$$ + +similar to the soft mask defined by Liu et al. [24]. It should be noted that, unlike Liu et al., we use this mask as an input to our network rather than a blending mask. + +Our HDR albedo reconstruction network receives the LDR image $I_{L}$ , the estimated LDR albedo $A_{L}$ , and the soft attention mask $\alpha$ , as a concatenated $h \times w \times 7$ input. Similar to our loss formulation for shading, we define the MSE and multi-scale gradient losses: + +$$ +\mathcal {L} _ {m s e} ^ {A} = M S E \left(A _ {H} ^ {*}, A _ {H}\right), \quad \mathcal {L} _ {m s g} ^ {A} = M S G \left(A _ {H} ^ {*}, A _ {H}\right), \tag {9} +$$ + +as well as losses on implied inverse shading $\hat{D}_H = A_H / (I_H^* +A_H)$ using the ground-truth HDR image $I_H^*$ and the intrinsic model in Eq. 1: + +$$ +\mathcal {L} _ {m s e} ^ {\hat {D}} = M S E \left(D _ {H} ^ {*}, \hat {D} _ {H}\right), \quad \mathcal {L} _ {m s g} ^ {\hat {D}} = M S G \left(D _ {H} ^ {*}, \hat {D} _ {H}\right). \tag {10} +$$ + +Our final loss for the HDR albedo reconstruction network is defined as: + +$$ +\mathcal {L} ^ {A} = \mathcal {L} _ {m s e} ^ {A} + \gamma \cdot \mathcal {L} _ {m s g} ^ {A} + \mathcal {L} _ {m s e} ^ {\hat {D}} + \gamma \cdot \mathcal {L} _ {m s g} ^ {\hat {D}}, \quad \gamma = 0. 1. \tag {11} +$$ + +![](images/dbd484098b7a32c5717aef01807b4552627f56e3301460a4ebdf0080ba1e563e.jpg) + +![](images/cab2a7a40e3d3a507422dfbb14e2927db788ee7be893af66affbde34970cd006.jpg) + +![](images/d6330f3c3bfb31c65b468dd1b1addbc4c7f03714ee69b839b1c160f3b2976cb9.jpg) + +![](images/57958da0bbdd1dbe01365c0e27ae0e1837a0d8aed855e9a8266b81fb93602d8f.jpg) + +![](images/4c26592612025a030cba6e49fad6a7ac56336667f4e21fddc431a8951c0da82f.jpg) +Fig. 6: We show the intermediate results of our pipeline from an LDR input compared with its ground truth. The direct HDR reconstruction from the estimated HDR intrinsics can lead to small artifacts which are reduced by our refinement network. + +![](images/ba67c0a195c4844acbab5e03a8915b4057f2b4e7b4fce570b4694b226d9f1ef0.jpg) + +![](images/80a8eba905890e4235a4c08a8031cf9ff05200f43df6f7f51064e7e46cb14133.jpg) + +![](images/a2b55dbcb2b36a2f4f25a1feaabd562dbfed2b9b16328b9d69715667f2e74e02.jpg) + +# 4.3 Refinement + +The outputs of our HDR reconstruction networks for shading and albedo can be combined using the intrinsic model to give us an initial estimation for the HDR image: + +$$ +\hat {I} _ {H} = A _ {H} * S _ {H}, \quad S _ {H} = \frac {1 - D _ {H}}{D _ {H}}. \tag {12} +$$ + +However, as the two intrinsic components are estimated separately, the perfect reconstruction of the HDR image is not guaranteed as shown in Figure 6. In the final step of our pipeline, we train a refinement network that takes the intrinsic HDR reconstruction results as input and generates our final HDR reconstruction result. + +This network has a similar goal to the hallucination network used in Liu et al. [24]: Reconstructing the final HDR image by extending dynamic range. However, as it receives the intrinsic reconstruction of the HDR image readily as input, its task is significantly easier. Unlike a network trained for direct LDR-to-HDR conversion, it does not need to reason about the high-level cues in the image which is performed by our HDR shading reconstruction network. Similarly, it does not need to recover the corrupted colors in saturated regions, which is performed by our HDR albedo reconstruction network. It is simply tasked with combining the rich input information it receives to reconstruct the final HDR image, which allows it to generate more accurate and sharper results. + +Following the inverse representation we adopt for shading, we use the inverse representation for the HDR images as a bounded representation that fits the long-tailed distribution in HDR images between $[0,1]$ : + +$$ +J _ {H} = \frac {1}{I _ {H} + 1}, \quad \hat {J} _ {H} = \frac {1}{\hat {I} _ {H} + 1}, \quad J _ {H} ^ {*} = \frac {1}{I _ {H} ^ {*} + 1}, \tag {13} +$$ + +where $J_{H}$ , and $J_{H}^{*}$ denote the estimated and ground-truth inverse HDR images and $\hat{J}_H$ is the inferred inverse HDR image following Eq. 12. + +Our refinement network receives the LDR image $I_{L}$ , the inferred inverse HDR image $\hat{J}_{H}$ , the estimated HDR inverse shading $D_{L}$ , and the estimated HDR albedo $A_{L}$ as a concatenated $h\times w\times 10$ input. We use the MSE and multi-scale gradient losses to train our refinement network: + +$$ +\mathcal {L} _ {m s e} ^ {J} = M S E \left(J _ {H} ^ {*}, J _ {H}\right), \quad \mathcal {L} _ {m s g} ^ {J} = M S G \left(J _ {H} ^ {*}, J _ {H}\right), \tag {14} +$$ + +with the final loss defined as $\mathcal{L}^J = \mathcal{L}_{mse}^J +\gamma \cdot \mathcal{L}_{msg}^J$ $\gamma = 0.1$ + +# 4.4 Network architecture and training + +We utilize the same architecture for our 3 networks. We use EfficientNet [37] as the backbone with the decoder from [33], replacing the output activation with a sigmoid. We use RAdam [23] as the optimizer with a learning rate of $1e^{-4}$ and CosineAnnealing [25] scheduling. + +To train our intrinsic formulation, we use the provided synthetic intrinsic ground truth from Hypersim [35] and construct ground truth intrinsics from the raw images of the MultiIllum dataset [31] as described in [3]. We follow the preprocessing steps from Liu et al. [24] to create clipped, linearized LDR images with randomly sampled exposure values $t \in [-3..3]$ from the ground truth data which we then decompose into our LDR input intrinsics using [3]. + +Our refinement network in contrast does not require supervision in intrinsic space. We can thus extend our training set by real images from MultiRAW [22], LSMI [16], and the HDR-Real training subset from Liu et al. [24]. For this stage, we follow the ISP model in [24] to convert the ground truth images into JPEG files, use their pre-trained networks to retrieve linearized LDR images, and again retrieve the intrinsic decomposition via [3]. We provide a more detailed description of our training procedure in the supplementary material. + +# 5 Experiments + +In this section, we present a thorough evaluation of our proposed method using multiple standard benchmarks. First, we discuss the datasets, metrics and prior work used in our experimental setup, then we present qualitative and quantitative comparisons to state-of-the-art approaches. We present a more detailed evaluation and our ablation studies in the supplementary material. + +# 5.1 Quantitative evaluation + +To quantitatively evaluate our approach against existing methods we follow the evaluation protocol proposed by Hanji et al. [15]. For a given ground truth HDR image, we follow the preprocessing of Liu et al. [24] and map the values into the range $[1 - 1000]$ . We determine the best scale to align the input reconstruction with the ground truth with least-squares, using only values between the $10^{th}$ and $90^{th}$ percentiles to avoid outliers. We then perform CRF correction to match the reconstruction to the ground truth by fitting a per-channel third-degree polynomial in log-RGB space. + +Table 1: Quantitative results against state-of-the-art using the HDR-VDP3 metric. Our method achieves state-of-the-art performance on the difficult high-definition SIHDR benchmark dataset. We are state-of-the-art or competitive on all other datasets. + +
MethodSIHDRHDR-EYEHDR-REALHDR-SYNTHRAISE
Ours8.968.967.578.218.89
DrTMO [11]8.278.727.578.278.64
HDR-CNN [10]8.398.616.708.038.47
ExpandNet [30]8.678.496.977.698.27
Single-HDR [24]8.798.967.478.518.95
Mask-HDR [36]8.258.607.108.148.44
HDRUNet [6]8.828.465.717.618.04
Multi-Exp Gen. [19]8.648.877.348.168.66
Lightweight [14]8.268.207.037.317.68
+ +Metrics We follow the guidance of Hanji et al. [15] when choosing metrics to evaluate our approach. In their study, they find that certain metrics better correlate with results from qualitative experiments. We focus our evaluation on the popular HDR-VDP3 [28] metric as it is purpose-built for HDR evaluation and models the subjective evaluation of a human observer. We specifically use the "quality" task of HDR-VDP3. Furthermore, we evaluate multiple recommended SDR metrics using the PU21 [27] encoding and provide those results in the supplementary material. + +Datasets We evaluate our method on five publicly available datasets. We use the recent high-resolution SI-HDR Benchmark [15] as well as commonly used real-world datasets HDR-Eye [32], HDR-Real [24], HDR-Synth [24], and RAISE [7]. + +We compare our method against a collection of baselines that share their implementation publicly. Table 1 shows the results of our quantitative evaluation. Each value represents the average HDR-VDP3 score in Just-Objectionable-Differences (JOD) units. We show a significant improvement in quality in the SI-HDR Benchmark [15] while achieving state-of-the-art performance on 3 of the other 4 datasets. As our qualitative evaluation in the next section and our results presented in the supplementary material demonstrate, the favorable performance of our method comes from our physically-motivated intrinsic formulation, allowing us to reliably reconstruct the high luminance details and recover lost colors. We provide additional metrics and analysis in the supplementary material. + +# 5.2 Qualitative evaluation + +We present qualitative comparisons in Figures 7 and 8, showing the differences in both detail and color reconstruction. On SI-HDR [15], prior works struggle with predicting accurate details around edges where over- and under-exposed pixels meet. This becomes notable around the leaves in the [roof] scene with visible halos and color shifts. On the floor in the [book] scene, most competitors fail to + +![](images/d01c1224f68a779baf37ec4aa9aefb287a92309ed3912809d25dba97200c9c7c.jpg) +Fig. 7: We show results from our method on SI-HDR [15] to the other baselines. We refer to Sec. 5.2 and the supplementary material for an in-depth discussion. + +reconstruct the homogenous surface and hallucinate patterns and color blobs. In both cases, our method seamlessly recovers missing details in over-exposed areas without artifacts. For the woman in Fig. 8, HDRUnet [6] and SingleHDR [24] both exhibit strong artifacts as bright spots in the hand and the face, similar to the faces of the [bust] and the [kids]. Our method in contrast can recover the correct color information due to our modeling of the albedo. The results showcase the advantage of the intrinsic formulation for creating visually pleasing results. We provide an extensive qualitative comparison in the supplementary material. + +![](images/a68e7598df6978107bd24a7a1d30819d8bd016f1170eafecd21a198f94eb29cf.jpg) + +![](images/38fb0e9c467400789472431a76624ce98c4613efe22db6a41a2a1a36f15856c7.jpg) + +![](images/3984111da2aaebed448b163c5124bbe12dcf26dfe0bb3541574840daa12bf78c.jpg) + +![](images/2c509ad15a4ff5f7381532de693d67aa2c8469f4f192f9742ae0265e1cb40d1b.jpg) + +![](images/0334b1b1b9672037db7b8297e0f5a4a399a2fa3ec12f709179c15690ed2793d4.jpg) + +![](images/b5bfc82b7b334beb34c57d3eff9acc483cbfe807d5b92ff22956e588b6743fa1.jpg) + +![](images/1325199239a42b17dcfe76f72a3e448e6cab1ac0826877099534e9db2d8388db.jpg) + +![](images/02e55108d25f0081f4cdb6623ad11595b03086f5bdfce38a9462e415872e9f63.jpg) + +![](images/7caa7c6a2d016d5c843dfeb26973b03eff6203f70ab278fc85b81ef5db4cef0f.jpg) + +![](images/b2213ab8689a9b81e00cf4590084b6a4f69c4f7126d8ad1b7d71b3d3fbb81b70.jpg) + +![](images/1df629466244146be371967528928d0f1a4795ab7a3d87ce078320badb88d704.jpg) + +![](images/dadb336317ebfdc6cdf802009298c99771540beb391081d93fc6de81870e388f.jpg) + +![](images/0af6ef37a4a6283a521df2d5ac7eaf9f1618b94cc0401e39de2944b8285df794.jpg) + +![](images/fe1864a602c8beb04e24af78df2114ffd2d6fe9cd4225a415f4fdc4863c77e3c.jpg) +Fig.8: We compare results from our method with a focus on color reconstruction against HDRUnet [6] and SingleHDR [24]. We refer to Sec. 5.2 and the supplementary material for an in-depth discussion. + +![](images/4b7367fc33f5f7d7825fb0255bc5036ade08a3040a5139f3ab7a02a4f71e8b77.jpg) + +![](images/547c2ab0a0540460c4752b12eee0d8e241bf7e8523ebe20cb0d537112080e4e8.jpg) + +![](images/601e42538306657d9deab810c93630f380a60ccd79072501a95aa50d632f199f.jpg) + +![](images/63693be8cf6195708200005ad11e9d54dba18466f152c7e74118c3ff36e304fe.jpg) + +# 6 Conclusion + +In this work, we remodel the HDR reconstruction problem in the intrinsic domain and propose a multi-stage pipeline to recover lost details. The luminance range extension can be conveniently modeled in the shading space which allows the network to better model the correlation between illumination and geometry. Similarly, neural networks can easily take advantage of albedo sparsity to recover lost color information with high fidelity. Through qualitative and quantitative evaluations, we show that our physically-inspired modeling of the HDR reconstruction problem achieves state-of-the-art performance, improving upon prior work in terms of both detail recovery and color reproduction in a wide variety of in-the-wild photographs. + +# Acknowledgements + +We acknowledge the support of the Natural Sciences and Engineering Research Council of Canada (NSERC), [RGPIN-2020-05375]. + +# References + +1. Banterle, F., Ledda, P., Debattista, K., Chalmers, A.: Inverse tone mapping. In: Proc.GRAPHITE (2006) +2. Barrow, H., Tenenbaum, J., Hanson, A., Riseman, E.: Recovering intrinsic scene characteristics. Comput. vis. syst 2(3-26), 2 (1978) +3. Careaga, C., Aksoy, Y.: Intrinsic image decomposition via ordinal shading. ACM Trans. Graph. 43(1) (2023) +4. Careaga, C., Miangoleh, S.M.H., Aksoy, Y.: Intrinsic harmonization for illumination-aware compositing. In: Proc. SIGGRAPH Asia (2023) +5. Chen, S.K., Yen, H.L., Liu, Y.L., Chen, M.H., Hu, H.N., Peng, W.H., Lin, Y.Y.: CEVR: Learning continuous exposure value representations for single-image HDR reconstruction. In: Proc. ICCV (2023) +6. Chen, X., Liu, Y., Zhang, Z., Qiao, Y., Dong, C.: HDRUnet: Single image HDR reconstruction with denoising and dequantization. In: Proc. CVPR (2021) +7. Dang-Nguyen, D.T., Pasquini, C., Conotter, V., Boato, G.: RAISE: a raw images dataset for digital image forensics. In: Proc. MMSys (2015) +8. Debevec, P.: A median cut algorithm for light probe sampling. In: ACM SIGGRAPH 2005 Posters. ACM (2005) +9. Debevec, P.E., Malik, J.: Recovering high dynamic range radiance maps from photographs. In: ACM Trans. Graph. (1997) +10. Eilertsen, G., Kronander, J., Denes, G., Mantiuk, R.K., Unger, J.: HDR image reconstruction from a single exposure using deep CNNs. ACM Trans. Graph. 36(6) (2017) +1. Endo, Y., Kanamori, Y., Mitani, J.: Deep reverse tone mapping. ACM Trans. Graph. 36(6) (2017) +2. Garces, E., Rodriguez-Pardo, C., Casas, D., Lopez-Moreno, J.: A survey on intrinsic images: Delving deep into lambert and beyond. Int. J. Comput. Vision (2022) +3. Gilchrist, A., Jacobsen, A.: Perception of lightness and illumination in a world of one reflectance. Perception 13, 5-19 (02 1984) +4. Guo, C., Xiuhua, J.: LHDR: HDR reconstruction for legacy content using a lightweight DNN. In: Proc. ACCV (2022) +5. Hanji, P., Mantiuk, R., Eilertsen, G., Hajisharif, S., Unger, J.: Comparison of single image HDR reconstruction methods — the caveats of quality assessment. In: ACM Trans. Graph. (2022) +6. Kim, D., Kim, J., Nam, S., Lee, D., Lee, Y., Kang, N., Lee, H.E., Yoo, B., Han, J.J., Kim, S.J.: Large scale multi-illuminant (LSMI) dataset for developing white balance algorithm under mixed illumination. In: Proc. CVPR (2021) +7. Kim, J., Lee, S., Kang, S.J.: End-to-end differentiable learning to HDR image synthesis for multi-exposure images. In: Proc. AAAI (2021) +8. Kovaleski, R.P., Oliveira, M.M.: High-quality reverse tone mapping for a wide range of exposures. In: Proc. SIBGRAPI (2014) +9. Le, P.H., Le, Q., Nguyen, R., Hua, B.S.: Single-image HDR reconstruction by multi-exposure generation. In: Proc. WACV (2023) +20. Lee, S., An, G.H., Kang, S.J.: Deep recursive HDRI: Inverse tone mapping using generative adversarial networks. In: Proc. ECCV (2018) +21. Li, Z., Snavely, N.: Learning intrinsic image decomposition from watching the world. In: Proc. CVPR (2018) +22. Li, Z., Lu, M., Zhang, X., Feng, X., Asif, M.S., Ma, Z.: Efficient visual computing with camera raw snapshots. IEEE Trans. Pattern Anal. Mach. Intell. (2024) + +23. Liu, L., Jiang, H., He, P., Chen, W., Liu, X., Gao, J., Han, J.: On the variance of the adaptive learning rate and beyond. In: Proc. ICLR (2020) +24. Liu, Y.L., Lai, W.S., Chen, Y.S., Kao, Y.L., Yang, M.H., Chuang, Y.Y., Huang, J.B.: Single-image HDR reconstruction by learning to reverse the camera pipeline. In: Proc. CVPR (2020) +25. Loshchilov, I., Hutter, F.: SGDR: Stochastic gradient descent with warm restarts. In: Proc. ICLR (2017) +26. Mann, S., Picard, R.: On being "undigital" with digital cameras: Extending dynamic range by combining differently exposed pictures. In: Proc. IS&T Annual Conference (1995) +27. Mantiuk, R.K., Azimi, M.: Pu21: A novel perceptually uniform encoding for adapting existing quality metrics for hdr. In: Proc. PCS (2021) +28. Mantiuk, R.K., Hammou, D., Hanji, P.: HDR-VDP-3: A multi-metric for predicting image differences, quality and contrast distortions in high dynamic range and regular content. arXiv preprint arXiv:2304.13625 (2023) +29. Maralan, S.S., Careaga, C., Aksoy, Y.: Computational flash photography through intrinsics. In: Proc. CVPR (2023) +30. Marnerides, D., Bashford-Rogers, T., Hatchett, J., Debattista, K.: Expandnet: A deep convolutional neural network for high dynamic range expansion from low dynamic range content. Comput. Graph. Forum 37(2) (2018) +31. Murmann, L., Gharbi, M., Aittala, M., Durand, F.: A multi-illumination dataset of indoor object appearance. In: Proc. ICCV (2019) +32. Nemoto, H., Korshunov, P., Hanhart, P., Ebrahimi, T.: Visual attention in LDR and HDR images. In: Proc. VPQM (2015) +33. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Trans. Pattern Anal. Mach. Intell. 44(3) (2022) +34. Rempel, A.G., Trentacoste, M., Seetzen, H., Young, H.D., Heidrich, W., Whitehead, L., Ward, G.: LDR2HDR: on-the-fly reverse tone mapping of legacy video and photographs. ACM Trans. Graph. 26(3) (2007) +35. Roberts, M., Ramapuram, J., Ranjan, A., Kumar, A., Bautista, M.A., Paczan, N., Webb, R., Susskind, J.M.: Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In: Proc. ICCV (2021) +36. Santos, M.S., Ren, T.I., Kalantari, N.K.: Single image HDR reconstruction using a CNN with masked features and perceptual loss. ACM Trans. Graph. 39(4) (2020) +37. Tan, M., Le, Q.: EfficientNet: Rethinking model scaling for convolutional neural networks. In: Proc. ICML (2019) +38. Wang, C., Serrano, A., Pan, X., Chen, B., Myszkowski, K., Seidel, H.P., Theobalt, C., Leimkuhler, T.: Glowgan: Unsupervised learning of HDR images from LDR images in the wild. In: Proc. ICCV (2023) +39. Xu, D., Doutre, C., Nasiopoulos, P.: Correction of clipped pixels in color images. IEEE Trans. Vis. Comput. Graph. 17(3) (2010) +40. Zhang, N., Ye, Y., Zhao, Y., Wang, R.: Revisiting the stack-based inverse tone mapping. In: Proc. CVPR (2023) +41. Zhang, X., Brainard, D.H.: Estimation of saturated pixel values in digital color imaging. J. Optical Soc. Am. A 21(12) (2004) +42. Zhang, Y., Aydin, T.: Deep HDR estimation with generative detail reconstruction. Comput. Graph. Gorum 40(2), 179-190 (2021) \ No newline at end of file diff --git a/2024/Intrinsic Single-Image HDR Reconstruction/images.zip b/2024/Intrinsic Single-Image HDR Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..18e945c565ad36c3a49e5ff5a3af1af4f54255fd --- /dev/null +++ b/2024/Intrinsic Single-Image HDR Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:827e21088f151c70dce9af303bcaa8cd306675778876af1e7f98aa9c90a28f5c +size 920193 diff --git a/2024/Intrinsic Single-Image HDR Reconstruction/layout.json b/2024/Intrinsic Single-Image HDR Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ca317760692587d1effbabd290dfd8aff9beb5af --- /dev/null +++ b/2024/Intrinsic Single-Image HDR Reconstruction/layout.json @@ -0,0 +1,9231 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 154, + 112, + 460, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 112, + 460, + 129 + ], + "spans": [ + { + "bbox": [ + 154, + 112, + 460, + 129 + ], + "type": "text", + "content": "Intrinsic Single-Image HDR Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 179, + 150, + 435, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 150, + 435, + 163 + ], + "spans": [ + { + "bbox": [ + 179, + 150, + 435, + 163 + ], + "type": "text", + "content": "Sebastian Dille\\* , Chris Careaga\\* , and Yaigiz Aksoy" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 241, + 172, + 373, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 172, + 373, + 204 + ], + "spans": [ + { + "bbox": [ + 241, + 172, + 373, + 204 + ], + "type": "text", + "content": "Simon Fraser University Burnaby, BC V5A 1S6, Canada sdille@sfu.ca" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 142, + 217, + 233, + 279 + ], + "blocks": [ + { + "bbox": [ + 142, + 217, + 233, + 279 + ], + "lines": [ + { + "bbox": [ + 142, + 217, + 233, + 279 + ], + "spans": [ + { + "bbox": [ + 142, + 217, + 233, + 279 + ], + "type": "image", + "image_path": "b94aba13b10ed44f5cd7842782273070a4e0b5f3b0f4da4bd2bfc82730171a53.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 156, + 283, + 220, + 314 + ], + "blocks": [ + { + "bbox": [ + 156, + 283, + 220, + 314 + ], + "lines": [ + { + "bbox": [ + 156, + 283, + 220, + 314 + ], + "spans": [ + { + "bbox": [ + 156, + 283, + 220, + 314 + ], + "type": "image", + "image_path": "b88693e2b9c97e4dfc39d2328f535fc12116cd4a23d641c13b60fecb3d2bf424.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 235, + 217, + 306, + 313 + ], + "blocks": [ + { + "bbox": [ + 235, + 217, + 306, + 313 + ], + "lines": [ + { + "bbox": [ + 235, + 217, + 306, + 313 + ], + "spans": [ + { + "bbox": [ + 235, + 217, + 306, + 313 + ], + "type": "image", + "image_path": "bd1578a4ae706bc6d0c677422a09c43cf1b0444bfd5cc81ff7cbee8b3f31f08d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 217, + 378, + 313 + ], + "blocks": [ + { + "bbox": [ + 307, + 217, + 378, + 313 + ], + "lines": [ + { + "bbox": [ + 307, + 217, + 378, + 313 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 378, + 313 + ], + "type": "image", + "image_path": "9d3fd8a0ef2fa1af403dd8aa883eebc509b42915a2b141639b77fce4939408f3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 380, + 217, + 472, + 279 + ], + "blocks": [ + { + "bbox": [ + 380, + 217, + 472, + 279 + ], + "lines": [ + { + "bbox": [ + 380, + 217, + 472, + 279 + ], + "spans": [ + { + "bbox": [ + 380, + 217, + 472, + 279 + ], + "type": "image", + "image_path": "c081a015757e0cb57756c84b3da7b4f32262419542c417dab40e292d7a1c8982.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 396, + 283, + 460, + 313 + ], + "blocks": [ + { + "bbox": [ + 396, + 283, + 460, + 313 + ], + "lines": [ + { + "bbox": [ + 396, + 283, + 460, + 313 + ], + "spans": [ + { + "bbox": [ + 396, + 283, + 460, + 313 + ], + "type": "image", + "image_path": "76c62823aae19e48d7a890e5542c343b2e24f385e03a5a5f49b965d2f4a81444.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 141, + 315, + 223, + 371 + ], + "blocks": [ + { + "bbox": [ + 141, + 315, + 223, + 371 + ], + "lines": [ + { + "bbox": [ + 141, + 315, + 223, + 371 + ], + "spans": [ + { + "bbox": [ + 141, + 315, + 223, + 371 + ], + "type": "image", + "image_path": "e000213ad39461b9b479c23ae02d2b0756e7b816cb53c004ea627184c328d587.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 141, + 372, + 223, + 394 + ], + "blocks": [ + { + "bbox": [ + 141, + 372, + 223, + 394 + ], + "lines": [ + { + "bbox": [ + 141, + 372, + 223, + 394 + ], + "spans": [ + { + "bbox": [ + 141, + 372, + 223, + 394 + ], + "type": "image", + "image_path": "15c9daa8008848db1321d6a6d791ed07a2279d495f9d797b75ff39f211f88b79.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 403, + 482, + 437 + ], + "lines": [ + { + "bbox": [ + 131, + 403, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 131, + 403, + 482, + 437 + ], + "type": "text", + "content": "Fig. 1: In this work, we propose a physically-inspired HDR reconstruction pipeline by conducting reconstruction individually on intrinsic components. Our final reconstruction, as a result, is effective in recovering the high-luminance details and the lost colors." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 225, + 315, + 306, + 371 + ], + "blocks": [ + { + "bbox": [ + 225, + 315, + 306, + 371 + ], + "lines": [ + { + "bbox": [ + 225, + 315, + 306, + 371 + ], + "spans": [ + { + "bbox": [ + 225, + 315, + 306, + 371 + ], + "type": "image", + "image_path": "e4b7e6972ca5de426dd1bec7af395c005798d28b324822a919b5174eb36214c9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 225, + 372, + 306, + 394 + ], + "blocks": [ + { + "bbox": [ + 225, + 372, + 306, + 394 + ], + "lines": [ + { + "bbox": [ + 225, + 372, + 306, + 394 + ], + "spans": [ + { + "bbox": [ + 225, + 372, + 306, + 394 + ], + "type": "image", + "image_path": "87b751f9e5da6c0d1fcda4f0f9eb51785f4015a147b659f8352c1794c9e07103.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 315, + 389, + 371 + ], + "blocks": [ + { + "bbox": [ + 307, + 315, + 389, + 371 + ], + "lines": [ + { + "bbox": [ + 307, + 315, + 389, + 371 + ], + "spans": [ + { + "bbox": [ + 307, + 315, + 389, + 371 + ], + "type": "image", + "image_path": "7f8ea3dcfb895c1bd0381bd828404406f3daa497d63797c8d1b67261dc255126.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 372, + 389, + 394 + ], + "blocks": [ + { + "bbox": [ + 307, + 372, + 389, + 394 + ], + "lines": [ + { + "bbox": [ + 307, + 372, + 389, + 394 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 389, + 394 + ], + "type": "image", + "image_path": "24ba8bd9a7bfadc70539f9bfe5fcbfeef396f180752655a2f1112d615f475773.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 390, + 315, + 471, + 371 + ], + "blocks": [ + { + "bbox": [ + 390, + 315, + 471, + 371 + ], + "lines": [ + { + "bbox": [ + 390, + 315, + 471, + 371 + ], + "spans": [ + { + "bbox": [ + 390, + 315, + 471, + 371 + ], + "type": "image", + "image_path": "135f1be94acde2799ce37e6399e2feb683f080196b5c53342568a8f1ae22d11a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 390, + 372, + 471, + 394 + ], + "blocks": [ + { + "bbox": [ + 390, + 372, + 471, + 394 + ], + "lines": [ + { + "bbox": [ + 390, + 372, + 471, + 394 + ], + "spans": [ + { + "bbox": [ + 390, + 372, + 471, + 394 + ], + "type": "image", + "image_path": "ccda3e24fb4683a5a0dd6c7fdb70494104b6b770a7d6bd6b68e48904339b34e9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 159, + 453, + 453, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 453, + 453, + 617 + ], + "spans": [ + { + "bbox": [ + 159, + 453, + 453, + 617 + ], + "type": "text", + "content": "Abstract. The low dynamic range (LDR) of common cameras fails to capture the rich contrast in natural scenes, resulting in loss of color and details in saturated pixels. Reconstructing the high dynamic range (HDR) of luminance present in the scene from single LDR photographs is an important task with many applications in computational photography and realistic display of images. The HDR reconstruction task aims to infer the lost details using the context present in the scene, requiring neural networks to understand high-level geometric and illumination cues. This makes it challenging for data-driven algorithms to generate accurate and high-resolution results. In this work, we introduce a physically-inspired remodeling of the HDR reconstruction problem in the intrinsic domain. The intrinsic model allows us to train separate networks to extend the dynamic range in the shading domain and to recover lost color details in the albedo domain. We show that dividing the problem into two simpler sub-tasks improves performance in a wide variety of photographs." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 159, + 628, + 453, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 628, + 453, + 650 + ], + "spans": [ + { + "bbox": [ + 159, + 628, + 453, + 650 + ], + "type": "text", + "content": "Keywords: HDR Reconstruction " + }, + { + "bbox": [ + 159, + 628, + 453, + 650 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 628, + 453, + 650 + ], + "type": "text", + "content": " Inverse Tonemapping " + }, + { + "bbox": [ + 159, + 628, + 453, + 650 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 628, + 453, + 650 + ], + "type": "text", + "content": " Computational Photography" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 654, + 269, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 654, + 269, + 666 + ], + "spans": [ + { + "bbox": [ + 142, + 654, + 269, + 666 + ], + "type": "inline_equation", + "content": "(^{*})" + }, + { + "bbox": [ + 142, + 654, + 269, + 666 + ], + "type": "text", + "content": " denotes equal contribution." + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 246 + ], + "type": "text", + "content": "Real-world scenes contain a wider range of luminance than most consumer-grade cameras can record or store. As a result, everyday photographs can only represent a low dynamic range (LDR) version of the scene that lacks details and color in very dark or bright regions. High dynamic range (HDR) imaging aims to capture and store the full contrast present in the scene. HDR images allow realistic display of photographs using high-contrast displays as well as many computational photography applications such as tone mapping and illumination-aware image editing. However, HDR imaging typically requires multi-exposure image stacks which limits their applications to static scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "type": "text", + "content": "To extend the HDR-enabled applications to everyday photographs, there has been a growing interest in methods to estimate HDR images from only LDR photographs called single-image HDR reconstruction, or LDR-to-HDR conversion. Single-image HDR reconstruction aims to extend the dynamic range of LDR photographs by recovering the details lost during capture. The missing information in LDR images is typically characterized by saturated pixels in bright regions, lacking textural details and color. This makes HDR reconstruction a high-level problem, requiring networks to utilize geometric and context-dependent cues in the scene to infer the lost information. Formulating this reconstruction problem using standard color spaces such as RGB or Luv requires the networks to statistically model the complex relationship between luminance, surface colors, and geometry, which limits their performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 390, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 594 + ], + "type": "text", + "content": "In this work, we present a physically-inspired remodeling of the HDR reconstruction problem in the intrinsic domain. Intrinsic image decomposition is a mid-level vision problem that separates the effect of the illumination in the scene from the reflectance of surfaces [3], representing the input as a combination of a shading and an albedo layer. The shading of an image is highly correlated with the scene geometry and illumination, while the albedo contains the color information and the high-frequency textures. The saturated pixels we aim to recover create different degradations in the shading and albedo layers. The clipping of high values is observed in the shading, while the loss of color information causes desaturation in the albedo. We propose an HDR reconstruction pipeline that individually addresses the loss of information in the LDR image in the two intrinsic components. Our HDR shading reconstruction network extends the dynamic range, taking advantage of the high correlation between shading and the scene geometry. Our LDR albedo reconstruction network, on the other hand, recovers the lost color information in saturated regions. A final refinement stage combines the two components to generate our full HDR reconstruction. Figure 1 shows different stages of our pipeline." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "In our formulation, we take advantage of the intrinsic image formation model by dividing the recovery of lost details and colors into two separate tasks which are more convenient for networks to model. We show that our physically inspired approach can generate high-fidelity details with rich colors in complex in-the-wild scenes. We evaluate our method qualitatively and quantitatively to demonstrate the state-of-the-art performance of our formulation in a wide variety of scenes." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "S. Dille et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 147, + 479, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 147, + 479, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 147, + 479, + 194 + ], + "type": "text", + "content": "Common imaging sensors for photography are inherently limited in the portion of the scene's dynamic range they can capture in a single shot. This results in saturated pixels in bright regions, representing a significant challenge in the capture and realistic display of real-world scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 196, + 480, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 196, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 196, + 480, + 351 + ], + "type": "text", + "content": "Multi-exposure fusion methods can reconstruct an HDR image by combining a stack of LDR images taken at different exposures [9,26,40]. Their formulation successfully addresses contrast range extension and reversal of the quantization and nonlinearity of LDR images. The multi-exposure capture, however, assumes a static camera and scene, requiring a challenging alignment process and limiting their application scenario. To enable HDR-related applications in everyday photography, single-image HDR reconstruction approaches aim to recover the lost information directly from an LDR image. Earlier approaches apply specially designed filters [18,34] or median-cut [1,8] to fill in overexposed regions, using different heuristics to scale the intensity of a selected area. Similarly, Bayesian [41] and graph-based formulations [39] have been introduced for chroma restoration. Due to the high-level nature of the reconstruction problem, these low-level approaches struggle to generate realistic details in real-world scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 353, + 481, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 481, + 568 + ], + "type": "text", + "content": "In their concurrent publications, Eilertsen et al. [10] and Endo et al. [11] present the first data-driven approaches to the single-image HDR reconstruction problem. They train their networks for direct conversion to HDR from the LDR input, addressing quantization, nonlinearity, and saturation in the LDR image in a single step. Many works in recent literature adapt this unified inverse model [24] and develop tailored solutions for underexposed [6,14] or overexposed [36] areas or losses for color recovery [30]. The direct estimation of HDR outputs, however, creates the challenge of an unconstrained color distribution that hinders training efficiency. Several methods [5, 11, 17, 19, 20, 40] instead adopt the concept of exposure stacks and predict a set of over- and underexposed images from the input LDR image. With each prediction still in the LDR domain, the HDR image can later be created with multi-exposure fusion [9]. Other methods [38,42] employ generative approaches, either generating separate LDR detail layers [42] or full HDR images directly [38], with image-specific HDR reconstruction via GAN inversion. Liu et al. [24] approach HDR reconstruction by inverting the in-camera image formation process, modeling the dequantization, nonlinearization, and saturated pixel reconstruction in three separate steps. They show that this decomposition of the task simplifies the modeling of each individual sub-problem." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 665 + ], + "type": "text", + "content": "In this work, we adopt the modular formulation of Liu et al. [24] and focus on recovering the lost details and color information in the saturated pixels. Our method takes inspiration from prior works that formulate computational photography problems in the intrinsic image domain [4, 29]. Rather than defining a direct mapping from LDR to HDR, we divide this task into two sub-problems: shading reconstruction that extends the dynamic range, and reflectance reconstruction that recovers the color information. Our physical deconstruction of the problem significantly improves the recovery of details and color information." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "text", + "content": "Intrinsic Single-Image HDR Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 116, + 479, + 196 + ], + "blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 196 + ], + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 196 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 196 + ], + "type": "image", + "image_path": "d25827af201bb1f1214a21f93d29b01689f538979860505e0ca505e8b85648b7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 205, + 482, + 273 + ], + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 273 + ], + "type": "text", + "content": "Fig. 2: The in-camera pipeline modeling LDR image formation is shown at the top. Liu et al. [24] propose the inverse pipeline shown on the bottom, addressing each type of information loss individually in three steps. In this work, we use the pre-trained networks of Liu et al. [24] for the dequantization and linearization tasks (shown in pink) and focus on developing an intrinsic model for the dynamic range reconstruction problem highlighted in teal." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 295, + 321, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 295, + 321, + 309 + ], + "spans": [ + { + "bbox": [ + 132, + 295, + 321, + 309 + ], + "type": "text", + "content": "3 Background and Motivation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "text", + "content": "The focus of our work is the formulation of the single-image LDR-to-HDR reconstruction problem in the intrinsic domain. We begin our discussion by first introducing the LDR imaging pipeline and the intrinsic model of image formation. We then draw connections between the two models and motivate our intrinsic HDR formulation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 403, + 400, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 403, + 400, + 415 + ], + "spans": [ + { + "bbox": [ + 131, + 403, + 400, + 415 + ], + "type": "text", + "content": "3.1 LDR image formation and HDR reconstruction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": "The acquisition of standard RGB photographs involves a series of operations in the camera-internal image processing pipeline, or ISP, that result in different forms of information loss. The initial acquisition of the image is in a raw format, where the luminance received by the sensor from a real-world scene is scaled and stored in 16- or 32-bit data structures. The scale of the captured raw values is determined by the analog camera settings such as the shutter speed, the aperture size, and the ISO. Given the natural high dynamic range of real-world scenes, the raw values are clipped at a maximum due to saturation at the sensor level. As shown in Figure 2, the remainder of the ISP applies nonlinear color and gamma corrections before finally decreasing the bit rate to 8 bits per channel for storage." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "The LDR image formation process causes three different types of corruption of the HDR information: quantization, non-linearization, and clipping. Single-image HDR reconstruction methods, as a result, have to address all three of these challenges. While many methods in the literature address all three simultaneously within a single architecture, Liu et al. [24] design their pipeline to address each type of information loss with a separate dedicated network. They propose a dequantization network to first increase the bit depth of the input LDR image and then estimate a CRF curve for linearization. Both tasks are convenient for a network to learn, taking advantage of the ease of dataset generation for dequantization and a simple parametric curve estimation for linearization." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "S. Dille et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 118, + 482, + 240 + ], + "blocks": [ + { + "bbox": [ + 133, + 118, + 482, + 240 + ], + "lines": [ + { + "bbox": [ + 133, + 118, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 133, + 118, + 482, + 240 + ], + "type": "image", + "image_path": "454f9ce86cdedfe2e9162afe62b8df247d96ff76fb29929baade7e7384a17aaf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 254, + 482, + 299 + ], + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 299 + ], + "type": "text", + "content": "Fig. 3: We show the different effects of clipping on the albedo and the shading for a tonemapped HDR image and its LDR counterpart. When clipping the image, the LDR shading loses information compared to HDR. We scale the intensity of both shadings to show the effect. At the same time, the LDR albedo becomes saturated." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 322, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 429 + ], + "type": "text", + "content": "They then reconstruct the values lost to clipping with a third network in what they call the hallucination stage. This last stage of the inverse pipeline addresses the most significant challenge in single-image HDR reconstruction, namely the recovery of the highlights and the color information in clipped regions. Diverging from Liu et al.'s terminology, we refer to this stage as dynamic range extension. We adopt Liu et al.'s inverse pipeline and use their pre-trained networks for dequantization and linearization. In the following, we focus on the dynamic range extension problem and present our solution using an intrinsic model of image formation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 447, + 312, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 312, + 459 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 312, + 459 + ], + "type": "text", + "content": "3.2 Intrinsic image decomposition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 467, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 482, + 538 + ], + "type": "text", + "content": "Intrinsic image decomposition is a fundamental mid-level computer vision problem that represents a core step in inverse rendering pipelines [2, 12]. It aims to separate the image into illumination and reflectance components following the physical image formation model. Adopting the commonly used Lambertian shading assumption with a single-channel shading representation [3], the intrinsic equation can be written as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 272, + 548, + 481, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 548, + 481, + 562 + ], + "spans": [ + { + "bbox": [ + 272, + 548, + 481, + 562 + ], + "type": "interline_equation", + "content": "I _ {r g b} = A _ {r g b} \\cdot S, \\tag {1}", + "image_path": "d5fd359b399bd34e253df5d780057e643f32ab03c59d7a3e012ada75e0026462.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 480, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 480, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 570, + 480, + 605 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 570, + 480, + 605 + ], + "type": "text", + "content": " represents the linear image reflecting relative luminance, " + }, + { + "bbox": [ + 130, + 570, + 480, + 605 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 570, + 480, + 605 + ], + "type": "text", + "content": " represents albedo, i.e. the surface reflectance containing the color information, and " + }, + { + "bbox": [ + 130, + 570, + 480, + 605 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 570, + 480, + 605 + ], + "type": "text", + "content": " represents shading reflecting the effect of illumination in the scene." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": "Shading and albedo represent two orthogonal characteristics of a scene, one containing the effect of light and its interaction with the scene geometry and the other representing material properties. As a result, the two intrinsic components have very different characteristics as demonstrated in Figure 3. Shading is highly correlated with surface orientations, appearing smooth over continuous surfaces" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "text", + "content": "Intrinsic Single-Image HDR Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "and having a high gradient in depth discontinuities and shadow boundaries. Albedo, on the other hand, contains the high-frequency textures present on the materials in the scene while having a sparse set of values per image as the variation in brightness due to geometry is isolated in the shading layer. In the single-channel shading representation, albedo also contains any variation caused by colorful illumination [3]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 217, + 302, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 302, + 229 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 302, + 229 + ], + "type": "text", + "content": "3.3 Infrinsics in LDR and HDR" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 249, + 481, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 481, + 297 + ], + "type": "text", + "content": "The bounded representation of LDR images limits the range of luminance values they can contain. HDR reconstruction aims to extend this range of luminance by exploiting statistical properties of real-world photographs derived from high-level scene properties such as geometry and appearance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 482, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 372 + ], + "type": "text", + "content": "A brief look at the intrinsic components of HDR images, as shown in Figure 3, reveals the different characteristics of the shading and albedo layers. The distribution of reflectance values in real-world scenes typically has a narrow distribution, making the albedo a variable that can be represented linearly between [0, 1]. The high dynamic range of the HDR image, on the other hand, is reflected in the shading layer that similarly has an unbounded representation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 374, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 445 + ], + "type": "text", + "content": "As a result, the limited dynamic range of LDR images has different effects on the shading and albedo layers. As Figure 3 demonstrates, the clipping of HDR values results in a similar clipping in shading, showing the relationship between the dynamic range and illumination that has been previously discussed in the literature [10]. The loss of colors induced by clipping, on the other hand, appears in the albedo layer as the dullification of colors in over-exposed regions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 448, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 533 + ], + "type": "text", + "content": "Based on this observation, we formulate the HDR reconstruction problem separately in shading and albedo. In our setup, the network tasked with converting LDR shading into HDR shading models the high dynamic range of the illumination in the scene, taking advantage of the high correlation between shading and high-level scene properties such as geometry. We task another network with converting the LDR albedo into HDR albedo, training it to recover the lost color information in clipped regions. We detail our full pipeline in Section 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 534, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 665 + ], + "type": "text", + "content": "Our approach effectively separates the two challenges in dynamic range reconstruction, reconstruction of high luminance regions and recovery of the color information, into two separate tasks. Formulating the reconstruction of high luminance in the shading domain results in a more convenient way for the network to statistically model the problem. This can be attributed to the high correlation between shading and geometry, which is often violated in luminance due to textured surfaces as highlighted in Figure 5. Formulating the color recovery problem in the albedo domain simplifies the task similarly, allowing the exploitation of the sparsity of reflectance values in real-world scenes. Our albedo LDR-to-HDR network, as a result, acts comparable to a simple inpainting model, filling in the lost color details based on the scene content." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "S. Dille et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 482, + 196 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 196 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 196 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 196 + ], + "type": "image", + "image_path": "da5ceac200d5241d3e3a7849ce0f20133336571ed2dc04095054c3cff64baef7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 203, + 482, + 248 + ], + "lines": [ + { + "bbox": [ + 130, + 203, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 203, + 482, + 248 + ], + "type": "text", + "content": "Fig. 4: Our HDR reconstruction pipeline starts with the intrinsic decomposition of the image using an off-the-shelf method [3]. The LDR reconstruction is then done individually on shading and albedo layers, which are then combined in a final refinement stage. The input LDR image is provided to each network shown in blue." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 271, + 392, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 392, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 392, + 285 + ], + "type": "text", + "content": "4 Dynamic Range Extension via Infrinics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 299, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 482, + 419 + ], + "type": "text", + "content": "In this work, we propose a physically inspired system for dynamic range extension in the single-image HDR reconstruction pipeline. The input to our pipeline is the dequantized and linearized LDR photograph " + }, + { + "bbox": [ + 130, + 299, + 482, + 419 + ], + "type": "inline_equation", + "content": "I_{L}" + }, + { + "bbox": [ + 130, + 299, + 482, + 419 + ], + "type": "text", + "content": ". We first compute the intrinsic components of " + }, + { + "bbox": [ + 130, + 299, + 482, + 419 + ], + "type": "inline_equation", + "content": "I_{L}" + }, + { + "bbox": [ + 130, + 299, + 482, + 419 + ], + "type": "text", + "content": ", the LDR shading " + }, + { + "bbox": [ + 130, + 299, + 482, + 419 + ], + "type": "inline_equation", + "content": "S_{L}" + }, + { + "bbox": [ + 130, + 299, + 482, + 419 + ], + "type": "text", + "content": ", and albedo " + }, + { + "bbox": [ + 130, + 299, + 482, + 419 + ], + "type": "inline_equation", + "content": "A_{L}" + }, + { + "bbox": [ + 130, + 299, + 482, + 419 + ], + "type": "text", + "content": ", using the in-the-wild intrinsic decomposition method by Careaga and Aksoy [3]. We then estimate an HDR shading and an HDR albedo from the corresponding LDR components using dedicated networks. The estimated HDR intrinsic components are then combined and refined by a subsequent HDR reconstruction network yielding our final result. We present an overview of our pipeline in Figure 4. In this section, we present our formulation and training details for each step." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 439, + 306, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 306, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 306, + 451 + ], + "type": "text", + "content": "4.1 HDR shading reconstruction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "type": "text", + "content": "The shading layer in the intrinsic model contains the effect of illumination in the scene. The high dynamic range of illumination in the scene, shadows, and specular objects cause a long-tailed distribution of values in the shading map. An LDR input with clipped values similarly produces a clipped LDR shading layer after intrinsic decomposition. We train our HDR shading reconstruction network to extend the range of shading values beyond this clipping." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "Several works in the literature define the dynamic range extension task in the luminance domain, targeting a similar goal of isolating the effect of illumination. In their work, Eilertsen et al. [10] motivate this choice from a perceptual standpoint, citing earlier work that explores the effect of illumination versus surface colors in human perception [13]. The luminance, however, is computed as a linear combination of RGB channels and, as a result, still contains information on the reflectance. Shading, on the other hand, directly reflects the illumination and is accordingly highly correlated with the scene geometry as Figure 5 demonstrates. Defining the range extension task in shading instead of luminance benefits the network to model the clipped shading values by easily utilizing the geometry-related cues in the input." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "text", + "content": "Intrinsic Single-Image HDR Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 214, + 176 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 214, + 176 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 214, + 176 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 214, + 176 + ], + "type": "image", + "image_path": "4cda54777a5382be486b73abf80ab6cb1963dccf4ca8196cd197edaf12e7c947.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 184, + 480, + 217 + ], + "lines": [ + { + "bbox": [ + 132, + 184, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 480, + 217 + ], + "type": "text", + "content": "Fig. 5: We show the difference between the shading representation and the L-channel from LUV. The shading correlates well with the geometry, while the L-channel includes texture information as shown in the green insets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 214, + 115, + 247, + 175 + ], + "blocks": [ + { + "bbox": [ + 214, + 115, + 247, + 175 + ], + "lines": [ + { + "bbox": [ + 214, + 115, + 247, + 175 + ], + "spans": [ + { + "bbox": [ + 214, + 115, + 247, + 175 + ], + "type": "image", + "image_path": "65935f5a1892b83702eebc1741c8029c66906577e9aa840887a5ee7c134b1732.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 248, + 114, + 329, + 175 + ], + "blocks": [ + { + "bbox": [ + 248, + 114, + 329, + 175 + ], + "lines": [ + { + "bbox": [ + 248, + 114, + 329, + 175 + ], + "spans": [ + { + "bbox": [ + 248, + 114, + 329, + 175 + ], + "type": "image", + "image_path": "8ce2a5991430227ba6f7e98ee4ed108a01c24c99464e0f68a683e1d4fb99db10.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 329, + 115, + 362, + 175 + ], + "blocks": [ + { + "bbox": [ + 329, + 115, + 362, + 175 + ], + "lines": [ + { + "bbox": [ + 329, + 115, + 362, + 175 + ], + "spans": [ + { + "bbox": [ + 329, + 115, + 362, + 175 + ], + "type": "image", + "image_path": "a941d2858a6c5af0c8e6af13b33c8346c21d14a8fc8cf16ccbd16728fd1e6a2c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 363, + 114, + 444, + 175 + ], + "blocks": [ + { + "bbox": [ + 363, + 114, + 444, + 175 + ], + "lines": [ + { + "bbox": [ + 363, + 114, + 444, + 175 + ], + "spans": [ + { + "bbox": [ + 363, + 114, + 444, + 175 + ], + "type": "image", + "image_path": "9d7215e652dd67b12075570a4f312a9697ed8effebabb29875416c0282d03e3f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 444, + 118, + 478, + 173 + ], + "blocks": [ + { + "bbox": [ + 444, + 118, + 478, + 173 + ], + "lines": [ + { + "bbox": [ + 444, + 118, + 478, + 173 + ], + "spans": [ + { + "bbox": [ + 444, + 118, + 478, + 173 + ], + "type": "image", + "image_path": "6667fea6c75d0a69fc8f5f9eb02570232249df2c186dea6d2790e3cb3f2c8ec6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 240, + 479, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 479, + 264 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 479, + 264 + ], + "type": "text", + "content": "Following Careaga and Aksoy [3], we represent shading in the inverse domain that fits the long-tailed shading in a representation bounded in " + }, + { + "bbox": [ + 132, + 240, + 479, + 264 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 132, + 240, + 479, + 264 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 274, + 272, + 480, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 272, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 274, + 272, + 480, + 297 + ], + "type": "interline_equation", + "content": "D _ {L} = \\frac {1}{S _ {L} + 1}. \\tag {2}", + "image_path": "f1ade5fd666eda21d69f5c018fb767b6378ca7abdd89a11f330b0614bc773d0c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 303, + 479, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 479, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 479, + 373 + ], + "type": "text", + "content": "The motivation behind this inverse representation is similar to the use of the log space in HDR reconstruction: It allows the definition of balanced loss functions on variables that could take very large or small values otherwise. The advantage of the inverse domain over the log space lies in its bounded nature that allows using the variable as input without clipping or normalization as well as the use of bounded activation functions such as sigmoid." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 375, + 480, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 480, + 434 + ], + "type": "text", + "content": "Our HDR shading reconstruction network receives the LDR image " + }, + { + "bbox": [ + 130, + 375, + 480, + 434 + ], + "type": "inline_equation", + "content": "I_{L}" + }, + { + "bbox": [ + 130, + 375, + 480, + 434 + ], + "type": "text", + "content": " and the estimated LDR inverse shading " + }, + { + "bbox": [ + 130, + 375, + 480, + 434 + ], + "type": "inline_equation", + "content": "D_{L}" + }, + { + "bbox": [ + 130, + 375, + 480, + 434 + ], + "type": "text", + "content": " as a concatenated " + }, + { + "bbox": [ + 130, + 375, + 480, + 434 + ], + "type": "inline_equation", + "content": "h\\times w\\times 4" + }, + { + "bbox": [ + 130, + 375, + 480, + 434 + ], + "type": "text", + "content": " input. Its output is defined as the HDR inverse shading " + }, + { + "bbox": [ + 130, + 375, + 480, + 434 + ], + "type": "inline_equation", + "content": "D_{H}" + }, + { + "bbox": [ + 130, + 375, + 480, + 434 + ], + "type": "text", + "content": ". We use multiple loss functions to train our network. The first one is the standard mean-squared error loss defined in the inverse space:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 249, + 434, + 480, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 434, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 249, + 434, + 480, + 449 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m s e} ^ {D} = M S E \\left(D _ {H} ^ {*}, D _ {H}\\right), \\tag {3}", + "image_path": "50c2d45f2ea36c1c65c18d92cc163e2a531393de5a00128af9cc66702cb24fdc.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 452, + 479, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 452, + 479, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 479, + 487 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 452, + 479, + 487 + ], + "type": "inline_equation", + "content": "D_H^*" + }, + { + "bbox": [ + 130, + 452, + 479, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 452, + 479, + 487 + ], + "type": "inline_equation", + "content": "D_H" + }, + { + "bbox": [ + 130, + 452, + 479, + 487 + ], + "type": "text", + "content": " denote the ground-truth and estimated HDR inverse shading, respectively. We also use the multi-scale gradient loss [21] as an edge-aware smoothness loss:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 183, + 494, + 480, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 494, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 183, + 494, + 480, + 521 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m s g} ^ {D} = M S G \\left(D _ {H} ^ {*}, D _ {H}\\right) = \\frac {1}{N} \\sum_ {m} S A D \\left(\\nabla D _ {H} ^ {* m}, \\nabla D _ {H} ^ {m}\\right), \\tag {4}", + "image_path": "84feb630935a40c007c398ef1039c6a0c29ac4e55e692afda96637959bcc6864.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 528, + 479, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 479, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 479, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 528, + 479, + 552 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 130, + 528, + 479, + 552 + ], + "type": "text", + "content": " denotes the spatial gradient and " + }, + { + "bbox": [ + 130, + 528, + 479, + 552 + ], + "type": "inline_equation", + "content": "D_H^m" + }, + { + "bbox": [ + 130, + 528, + 479, + 552 + ], + "type": "text", + "content": " denotes the estimated inverse HDR shading at scale " + }, + { + "bbox": [ + 130, + 528, + 479, + 552 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 528, + 479, + 552 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 552, + 479, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 479, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 479, + 590 + ], + "type": "text", + "content": "Although the output of our HDR shading reconstruction network is defined only as the HDR inverse shading, we can use the ground-truth HDR image " + }, + { + "bbox": [ + 130, + 552, + 479, + 590 + ], + "type": "inline_equation", + "content": "I_H^*" + }, + { + "bbox": [ + 130, + 552, + 479, + 590 + ], + "type": "text", + "content": " and the intrinsic model in Eq. 1 to compute an implied HDR albedo " + }, + { + "bbox": [ + 130, + 552, + 479, + 590 + ], + "type": "inline_equation", + "content": "\\hat{A}_H" + }, + { + "bbox": [ + 130, + 552, + 479, + 590 + ], + "type": "text", + "content": " [3]:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 242, + 597, + 480, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 597, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 242, + 597, + 480, + 622 + ], + "type": "interline_equation", + "content": "\\hat {A} _ {H} = \\frac {I _ {H} ^ {*}}{S _ {H}}, \\quad S _ {H} = \\frac {1 - D _ {H}}{D _ {H}}. \\tag {5}", + "image_path": "3380ffa75a574ce620e77f382c502af612230978ff676fc832bb8beef882cfd1.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 628, + 469, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 628, + 469, + 640 + ], + "spans": [ + { + "bbox": [ + 132, + 628, + 469, + 640 + ], + "type": "text", + "content": "We define the MSE and multi-scale gradient loss also for the implied albedo:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 185, + 647, + 480, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 647, + 480, + 669 + ], + "spans": [ + { + "bbox": [ + 185, + 647, + 480, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m s e} ^ {\\hat {A}} = M S E \\left(A _ {H} ^ {*}, \\hat {A} _ {H}\\right), \\quad \\mathcal {L} _ {m s g} ^ {\\hat {A}} = M S G \\left(A _ {H} ^ {*}, \\hat {A} _ {H}\\right), \\tag {6}", + "image_path": "8fa247d46792c2e979e97cee53bffcda19a8bd77eaad999c695e3c06b7cacac1.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "S. Dille et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 405, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 405, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 405, + 128 + ], + "type": "text", + "content": "and backpropagate these losses to our network through Eq. 5." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 129, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 129, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 129, + 482, + 212 + ], + "type": "text", + "content": "Our HDR shading reconstruction network only estimates the HDR shading and the HDR albedo component will be estimated with a separate network defined in the next section. Estimating the HDR intrinsic components separately results in possible reconstruction errors when these components are combined to produce an estimation for the HDR image. While we will refine the HDR reconstruction in our final stage with a dedicated network, we define these losses on the implied albedo to promote a better fit within the intrinsic model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 213, + 466, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 213, + 466, + 224 + ], + "spans": [ + { + "bbox": [ + 146, + 213, + 466, + 224 + ], + "type": "text", + "content": "Our final loss for the HDR shading reconstruction network is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 194, + 236, + 481, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 236, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 194, + 236, + 481, + 251 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {D} = \\mathcal {L} _ {m s e} ^ {D} + \\gamma \\cdot \\mathcal {L} _ {m s g} ^ {D} + \\mathcal {L} _ {m s e} ^ {\\hat {A}} + \\gamma \\cdot \\mathcal {L} _ {m s g} ^ {\\hat {A}}, \\quad \\gamma = 0. 1. \\tag {7}", + "image_path": "a751c76de8cb445b8388e7d90dea19349e22a5c71c9261461f75a8f0287c3a2f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 270, + 301, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 301, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 301, + 281 + ], + "type": "text", + "content": "4.2 HDR albedo reconstruction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 480, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 480, + 327 + ], + "type": "text", + "content": "The clipping in the LDR image formation process results in the loss of color information in the saturated pixels of the LDR albedo map. Our HDR albedo reconstruction network is tasked with recovering the corrupted colors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 329, + 481, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 481, + 389 + ], + "type": "text", + "content": "While the LDR shading map has saturated values, signaling the regions to be reconstructed to our HDR shading reconstruction networks, the LDR albedo map does not carry the same signal as albedo maps are not saturated. To denote the regions that require color reconstruction, we provide the network with a soft guidance mask " + }, + { + "bbox": [ + 130, + 329, + 481, + 389 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 329, + 481, + 389 + ], + "type": "text", + "content": " defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 239, + 399, + 481, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 399, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 239, + 399, + 481, + 422 + ], + "type": "interline_equation", + "content": "\\alpha = \\frac {\\operatorname {m a x} \\left(0 , I _ {L} - \\lambda\\right)}{1 - \\lambda}, \\quad \\lambda = 0. 8, \\tag {8}", + "image_path": "8e17772b3c93759874f9f074bf3010eed8178030f2312557d1cff973b4d8025a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 431, + 480, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 480, + 466 + ], + "type": "text", + "content": "similar to the soft mask defined by Liu et al. [24]. It should be noted that, unlike Liu et al., we use this mask as an input to our network rather than a blending mask." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 468, + 482, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 516 + ], + "type": "text", + "content": "Our HDR albedo reconstruction network receives the LDR image " + }, + { + "bbox": [ + 130, + 468, + 482, + 516 + ], + "type": "inline_equation", + "content": "I_{L}" + }, + { + "bbox": [ + 130, + 468, + 482, + 516 + ], + "type": "text", + "content": ", the estimated LDR albedo " + }, + { + "bbox": [ + 130, + 468, + 482, + 516 + ], + "type": "inline_equation", + "content": "A_{L}" + }, + { + "bbox": [ + 130, + 468, + 482, + 516 + ], + "type": "text", + "content": ", and the soft attention mask " + }, + { + "bbox": [ + 130, + 468, + 482, + 516 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 468, + 482, + 516 + ], + "type": "text", + "content": ", as a concatenated " + }, + { + "bbox": [ + 130, + 468, + 482, + 516 + ], + "type": "inline_equation", + "content": "h \\times w \\times 7" + }, + { + "bbox": [ + 130, + 468, + 482, + 516 + ], + "type": "text", + "content": " input. Similar to our loss formulation for shading, we define the MSE and multi-scale gradient losses:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 192, + 526, + 481, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 526, + 481, + 540 + ], + "spans": [ + { + "bbox": [ + 192, + 526, + 481, + 540 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m s e} ^ {A} = M S E \\left(A _ {H} ^ {*}, A _ {H}\\right), \\quad \\mathcal {L} _ {m s g} ^ {A} = M S G \\left(A _ {H} ^ {*}, A _ {H}\\right), \\tag {9}", + "image_path": "ef7389202410cef96a37983f8170a2ab7d298d57460747a2e596060b0b6732a4.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 552, + 482, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 577 + ], + "type": "text", + "content": "as well as losses on implied inverse shading " + }, + { + "bbox": [ + 130, + 552, + 482, + 577 + ], + "type": "inline_equation", + "content": "\\hat{D}_H = A_H / (I_H^* +A_H)" + }, + { + "bbox": [ + 130, + 552, + 482, + 577 + ], + "type": "text", + "content": " using the ground-truth HDR image " + }, + { + "bbox": [ + 130, + 552, + 482, + 577 + ], + "type": "inline_equation", + "content": "I_H^*" + }, + { + "bbox": [ + 130, + 552, + 482, + 577 + ], + "type": "text", + "content": " and the intrinsic model in Eq. 1:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 187, + 587, + 481, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 587, + 481, + 607 + ], + "spans": [ + { + "bbox": [ + 187, + 587, + 481, + 607 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m s e} ^ {\\hat {D}} = M S E \\left(D _ {H} ^ {*}, \\hat {D} _ {H}\\right), \\quad \\mathcal {L} _ {m s g} ^ {\\hat {D}} = M S G \\left(D _ {H} ^ {*}, \\hat {D} _ {H}\\right). \\tag {10}", + "image_path": "c9d6fcbcab49da5b94a96938af2613e6317f8912fd822f418577aa06713e153a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 616, + 446, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 446, + 627 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 446, + 627 + ], + "type": "text", + "content": "Our final loss for the HDR albedo reconstruction network is defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 194, + 637, + 481, + 652 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 637, + 481, + 652 + ], + "spans": [ + { + "bbox": [ + 194, + 637, + 481, + 652 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {A} = \\mathcal {L} _ {m s e} ^ {A} + \\gamma \\cdot \\mathcal {L} _ {m s g} ^ {A} + \\mathcal {L} _ {m s e} ^ {\\hat {D}} + \\gamma \\cdot \\mathcal {L} _ {m s g} ^ {\\hat {D}}, \\quad \\gamma = 0. 1. \\tag {11}", + "image_path": "ccca93836d2cd0e1b7d8f2b0fa24d9ba819f8779e53be0faf2bd0c571499df1e.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "text", + "content": "Intrinsic Single-Image HDR Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 218, + 172 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 218, + 172 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 218, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 218, + 172 + ], + "type": "image", + "image_path": "dbd484098b7a32c5717aef01807b4552627f56e3301460a4ebdf0080ba1e563e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 220, + 114, + 305, + 172 + ], + "blocks": [ + { + "bbox": [ + 220, + 114, + 305, + 172 + ], + "lines": [ + { + "bbox": [ + 220, + 114, + 305, + 172 + ], + "spans": [ + { + "bbox": [ + 220, + 114, + 305, + 172 + ], + "type": "image", + "image_path": "cab2a7a40e3d3a507422dfbb14e2927db788ee7be893af66affbde34970cd006.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 308, + 114, + 393, + 173 + ], + "blocks": [ + { + "bbox": [ + 308, + 114, + 393, + 173 + ], + "lines": [ + { + "bbox": [ + 308, + 114, + 393, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 114, + 393, + 173 + ], + "type": "image", + "image_path": "d6330f3c3bfb31c65b468dd1b1addbc4c7f03714ee69b839b1c160f3b2976cb9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 395, + 114, + 481, + 173 + ], + "blocks": [ + { + "bbox": [ + 395, + 114, + 481, + 173 + ], + "lines": [ + { + "bbox": [ + 395, + 114, + 481, + 173 + ], + "spans": [ + { + "bbox": [ + 395, + 114, + 481, + 173 + ], + "type": "image", + "image_path": "57958da0bbdd1dbe01365c0e27ae0e1837a0d8aed855e9a8266b81fb93602d8f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 174, + 218, + 232 + ], + "blocks": [ + { + "bbox": [ + 133, + 174, + 218, + 232 + ], + "lines": [ + { + "bbox": [ + 133, + 174, + 218, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 174, + 218, + 232 + ], + "type": "image", + "image_path": "4c26592612025a030cba6e49fad6a7ac56336667f4e21fddc431a8951c0da82f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 240, + 480, + 274 + ], + "lines": [ + { + "bbox": [ + 130, + 240, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 480, + 274 + ], + "type": "text", + "content": "Fig. 6: We show the intermediate results of our pipeline from an LDR input compared with its ground truth. The direct HDR reconstruction from the estimated HDR intrinsics can lead to small artifacts which are reduced by our refinement network." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 220, + 174, + 306, + 232 + ], + "blocks": [ + { + "bbox": [ + 220, + 174, + 306, + 232 + ], + "lines": [ + { + "bbox": [ + 220, + 174, + 306, + 232 + ], + "spans": [ + { + "bbox": [ + 220, + 174, + 306, + 232 + ], + "type": "image", + "image_path": "ba67c0a195c4844acbab5e03a8915b4057f2b4e7b4fce570b4694b226d9f1ef0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 174, + 394, + 232 + ], + "blocks": [ + { + "bbox": [ + 309, + 174, + 394, + 232 + ], + "lines": [ + { + "bbox": [ + 309, + 174, + 394, + 232 + ], + "spans": [ + { + "bbox": [ + 309, + 174, + 394, + 232 + ], + "type": "image", + "image_path": "80a8eba905890e4235a4c08a8031cf9ff05200f43df6f7f51064e7e46cb14133.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 395, + 174, + 480, + 232 + ], + "blocks": [ + { + "bbox": [ + 395, + 174, + 480, + 232 + ], + "lines": [ + { + "bbox": [ + 395, + 174, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 395, + 174, + 480, + 232 + ], + "type": "image", + "image_path": "a2b55dbcb2b36a2f4f25a1feaabd562dbfed2b9b16328b9d69715667f2e74e02.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 298, + 218, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 298, + 218, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 298, + 218, + 308 + ], + "type": "text", + "content": "4.3 Refinement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 319, + 480, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 480, + 354 + ], + "type": "text", + "content": "The outputs of our HDR reconstruction networks for shading and albedo can be combined using the intrinsic model to give us an initial estimation for the HDR image:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 233, + 354, + 480, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 354, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 233, + 354, + 480, + 378 + ], + "type": "interline_equation", + "content": "\\hat {I} _ {H} = A _ {H} * S _ {H}, \\quad S _ {H} = \\frac {1 - D _ {H}}{D _ {H}}. \\tag {12}", + "image_path": "09c246e27e45644b7f557f43dea2175283d8677524db310ca0f02287fdb96e38.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 381, + 480, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 480, + 441 + ], + "type": "text", + "content": "However, as the two intrinsic components are estimated separately, the perfect reconstruction of the HDR image is not guaranteed as shown in Figure 6. In the final step of our pipeline, we train a refinement network that takes the intrinsic HDR reconstruction results as input and generates our final HDR reconstruction result." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 441, + 481, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 481, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 481, + 561 + ], + "type": "text", + "content": "This network has a similar goal to the hallucination network used in Liu et al. [24]: Reconstructing the final HDR image by extending dynamic range. However, as it receives the intrinsic reconstruction of the HDR image readily as input, its task is significantly easier. Unlike a network trained for direct LDR-to-HDR conversion, it does not need to reason about the high-level cues in the image which is performed by our HDR shading reconstruction network. Similarly, it does not need to recover the corrupted colors in saturated regions, which is performed by our HDR albedo reconstruction network. It is simply tasked with combining the rich input information it receives to reconstruct the final HDR image, which allows it to generate more accurate and sharper results." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 562, + 480, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 562, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 562, + 480, + 597 + ], + "type": "text", + "content": "Following the inverse representation we adopt for shading, we use the inverse representation for the HDR images as a bounded representation that fits the long-tailed distribution in HDR images between " + }, + { + "bbox": [ + 130, + 562, + 480, + 597 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 130, + 562, + 480, + 597 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 201, + 606, + 480, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 606, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 201, + 606, + 480, + 633 + ], + "type": "interline_equation", + "content": "J _ {H} = \\frac {1}{I _ {H} + 1}, \\quad \\hat {J} _ {H} = \\frac {1}{\\hat {I} _ {H} + 1}, \\quad J _ {H} ^ {*} = \\frac {1}{I _ {H} ^ {*} + 1}, \\tag {13}", + "image_path": "c4613f284b018a1faf12a3497dae503e2c52bea6b39323a73cf7162ae3631afe.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 639, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 639, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 639, + 480, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 639, + 480, + 665 + ], + "type": "inline_equation", + "content": "J_{H}" + }, + { + "bbox": [ + 130, + 639, + 480, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 639, + 480, + 665 + ], + "type": "inline_equation", + "content": "J_{H}^{*}" + }, + { + "bbox": [ + 130, + 639, + 480, + 665 + ], + "type": "text", + "content": " denote the estimated and ground-truth inverse HDR images and " + }, + { + "bbox": [ + 130, + 639, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\hat{J}_H" + }, + { + "bbox": [ + 130, + 639, + 480, + 665 + ], + "type": "text", + "content": " is the inferred inverse HDR image following Eq. 12." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "S. Dille et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "Our refinement network receives the LDR image " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "I_{L}" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": ", the inferred inverse HDR image " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\hat{J}_{H}" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": ", the estimated HDR inverse shading " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "D_{L}" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": ", and the estimated HDR albedo " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "A_{L}" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " as a concatenated " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "h\\times w\\times 10" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " input. We use the MSE and multi-scale gradient losses to train our refinement network:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 195, + 170, + 481, + 185 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 170, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 195, + 170, + 481, + 185 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m s e} ^ {J} = M S E \\left(J _ {H} ^ {*}, J _ {H}\\right), \\quad \\mathcal {L} _ {m s g} ^ {J} = M S G \\left(J _ {H} ^ {*}, J _ {H}\\right), \\tag {14}", + "image_path": "96951cb478fe7da1b2e3a0f2e80beaaf4c265ec9be4acf543cac982b41ed4eba.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 190, + 400, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 190, + 400, + 205 + ], + "spans": [ + { + "bbox": [ + 131, + 190, + 400, + 205 + ], + "type": "text", + "content": "with the final loss defined as " + }, + { + "bbox": [ + 131, + 190, + 400, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^J = \\mathcal{L}_{mse}^J +\\gamma \\cdot \\mathcal{L}_{msg}^J" + }, + { + "bbox": [ + 131, + 190, + 400, + 205 + ], + "type": "inline_equation", + "content": "\\gamma = 0.1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 217, + 335, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 217, + 335, + 231 + ], + "spans": [ + { + "bbox": [ + 131, + 217, + 335, + 231 + ], + "type": "text", + "content": "4.4 Network architecture and training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 235, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 482, + 283 + ], + "type": "text", + "content": "We utilize the same architecture for our 3 networks. We use EfficientNet [37] as the backbone with the decoder from [33], replacing the output activation with a sigmoid. We use RAdam [23] as the optimizer with a learning rate of " + }, + { + "bbox": [ + 130, + 235, + 482, + 283 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 130, + 235, + 482, + 283 + ], + "type": "text", + "content": " and CosineAnnealing [25] scheduling." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 283, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 354 + ], + "type": "text", + "content": "To train our intrinsic formulation, we use the provided synthetic intrinsic ground truth from Hypersim [35] and construct ground truth intrinsics from the raw images of the MultiIllum dataset [31] as described in [3]. We follow the preprocessing steps from Liu et al. [24] to create clipped, linearized LDR images with randomly sampled exposure values " + }, + { + "bbox": [ + 130, + 283, + 482, + 354 + ], + "type": "inline_equation", + "content": "t \\in [-3..3]" + }, + { + "bbox": [ + 130, + 283, + 482, + 354 + ], + "type": "text", + "content": " from the ground truth data which we then decompose into our LDR input intrinsics using [3]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 355, + 483, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 355, + 483, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 483, + 439 + ], + "type": "text", + "content": "Our refinement network in contrast does not require supervision in intrinsic space. We can thus extend our training set by real images from MultiRAW [22], LSMI [16], and the HDR-Real training subset from Liu et al. [24]. For this stage, we follow the ISP model in [24] to convert the ground truth images into JPEG files, use their pre-trained networks to retrieve linearized LDR images, and again retrieve the intrinsic decomposition via [3]. We provide a more detailed description of our training procedure in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 454, + 230, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 454, + 230, + 468 + ], + "spans": [ + { + "bbox": [ + 131, + 454, + 230, + 468 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 477, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 477, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 477, + 482, + 538 + ], + "type": "text", + "content": "In this section, we present a thorough evaluation of our proposed method using multiple standard benchmarks. First, we discuss the datasets, metrics and prior work used in our experimental setup, then we present qualitative and quantitative comparisons to state-of-the-art approaches. We present a more detailed evaluation and our ablation studies in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 552, + 279, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 279, + 565 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 279, + 565 + ], + "type": "text", + "content": "5.1 Quantitative evaluation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "To quantitatively evaluate our approach against existing methods we follow the evaluation protocol proposed by Hanji et al. [15]. For a given ground truth HDR image, we follow the preprocessing of Liu et al. [24] and map the values into the range " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "[1 - 1000]" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". We determine the best scale to align the input reconstruction with the ground truth with least-squares, using only values between the " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "10^{th}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "90^{th}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " percentiles to avoid outliers. We then perform CRF correction to match the reconstruction to the ground truth by fitting a per-channel third-degree polynomial in log-RGB space." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "text", + "content": "Intrinsic Single-Image HDR Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 157, + 459, + 281 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Quantitative results against state-of-the-art using the HDR-VDP3 metric. Our method achieves state-of-the-art performance on the difficult high-definition SIHDR benchmark dataset. We are state-of-the-art or competitive on all other datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 157, + 459, + 281 + ], + "lines": [ + { + "bbox": [ + 151, + 157, + 459, + 281 + ], + "spans": [ + { + "bbox": [ + 151, + 157, + 459, + 281 + ], + "type": "table", + "html": "
MethodSIHDRHDR-EYEHDR-REALHDR-SYNTHRAISE
Ours8.968.967.578.218.89
DrTMO [11]8.278.727.578.278.64
HDR-CNN [10]8.398.616.708.038.47
ExpandNet [30]8.678.496.977.698.27
Single-HDR [24]8.798.967.478.518.95
Mask-HDR [36]8.258.607.108.148.44
HDRUNet [6]8.828.465.717.618.04
Multi-Exp Gen. [19]8.648.877.348.168.66
Lightweight [14]8.268.207.037.317.68
", + "image_path": "7e9d7790db12538b3d2dc62a6ababd3228c4ad91d42f3f755a6b8d0cc26b3c01.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 303, + 482, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 400 + ], + "type": "text", + "content": "Metrics We follow the guidance of Hanji et al. [15] when choosing metrics to evaluate our approach. In their study, they find that certain metrics better correlate with results from qualitative experiments. We focus our evaluation on the popular HDR-VDP3 [28] metric as it is purpose-built for HDR evaluation and models the subjective evaluation of a human observer. We specifically use the \"quality\" task of HDR-VDP3. Furthermore, we evaluate multiple recommended SDR metrics using the PU21 [27] encoding and provide those results in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 411, + 481, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 481, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 481, + 447 + ], + "type": "text", + "content": "Datasets We evaluate our method on five publicly available datasets. We use the recent high-resolution SI-HDR Benchmark [15] as well as commonly used real-world datasets HDR-Eye [32], HDR-Real [24], HDR-Synth [24], and RAISE [7]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 448, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 567 + ], + "type": "text", + "content": "We compare our method against a collection of baselines that share their implementation publicly. Table 1 shows the results of our quantitative evaluation. Each value represents the average HDR-VDP3 score in Just-Objectionable-Differences (JOD) units. We show a significant improvement in quality in the SI-HDR Benchmark [15] while achieving state-of-the-art performance on 3 of the other 4 datasets. As our qualitative evaluation in the next section and our results presented in the supplementary material demonstrate, the favorable performance of our method comes from our physically-motivated intrinsic formulation, allowing us to reliably reconstruct the high luminance details and recover lost colors. We provide additional metrics and analysis in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 585, + 272, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 585, + 272, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 585, + 272, + 597 + ], + "type": "text", + "content": "5.2 Qualitative evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "We present qualitative comparisons in Figures 7 and 8, showing the differences in both detail and color reconstruction. On SI-HDR [15], prior works struggle with predicting accurate details around edges where over- and under-exposed pixels meet. This becomes notable around the leaves in the [roof] scene with visible halos and color shifts. On the floor in the [book] scene, most competitors fail to" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "S. Dille et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 482, + 502 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 482, + 502 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 482, + 502 + ], + "type": "image", + "image_path": "d01c1224f68a779baf37ec4aa9aefb287a92309ed3912809d25dba97200c9c7c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 508, + 482, + 533 + ], + "lines": [ + { + "bbox": [ + 132, + 508, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 482, + 533 + ], + "type": "text", + "content": "Fig. 7: We show results from our method on SI-HDR [15] to the other baselines. We refer to Sec. 5.2 and the supplementary material for an in-depth discussion." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "reconstruct the homogenous surface and hallucinate patterns and color blobs. In both cases, our method seamlessly recovers missing details in over-exposed areas without artifacts. For the woman in Fig. 8, HDRUnet [6] and SingleHDR [24] both exhibit strong artifacts as bright spots in the hand and the face, similar to the faces of the [bust] and the [kids]. Our method in contrast can recover the correct color information due to our modeling of the albedo. The results showcase the advantage of the intrinsic formulation for creating visually pleasing results. We provide an extensive qualitative comparison in the supplementary material." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "text", + "content": "Intrinsic Single-Image HDR Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 219, + 173 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 219, + 173 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 219, + 173 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 219, + 173 + ], + "type": "image", + "image_path": "a68e7598df6978107bd24a7a1d30819d8bd016f1170eafecd21a198f94eb29cf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 220, + 114, + 306, + 173 + ], + "blocks": [ + { + "bbox": [ + 220, + 114, + 306, + 173 + ], + "lines": [ + { + "bbox": [ + 220, + 114, + 306, + 173 + ], + "spans": [ + { + "bbox": [ + 220, + 114, + 306, + 173 + ], + "type": "image", + "image_path": "38fb0e9c467400789472431a76624ce98c4613efe22db6a41a2a1a36f15856c7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 393, + 173 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 393, + 173 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 393, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 393, + 173 + ], + "type": "image", + "image_path": "3984111da2aaebed448b163c5124bbe12dcf26dfe0bb3541574840daa12bf78c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 394, + 114, + 480, + 173 + ], + "blocks": [ + { + "bbox": [ + 394, + 114, + 480, + 173 + ], + "lines": [ + { + "bbox": [ + 394, + 114, + 480, + 173 + ], + "spans": [ + { + "bbox": [ + 394, + 114, + 480, + 173 + ], + "type": "image", + "image_path": "2c509ad15a4ff5f7381532de693d67aa2c8469f4f192f9742ae0265e1cb40d1b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 173, + 219, + 243 + ], + "blocks": [ + { + "bbox": [ + 133, + 173, + 219, + 243 + ], + "lines": [ + { + "bbox": [ + 133, + 173, + 219, + 243 + ], + "spans": [ + { + "bbox": [ + 133, + 173, + 219, + 243 + ], + "type": "image", + "image_path": "0334b1b1b9672037db7b8297e0f5a4a399a2fa3ec12f709179c15690ed2793d4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 220, + 173, + 306, + 243 + ], + "blocks": [ + { + "bbox": [ + 220, + 173, + 306, + 243 + ], + "lines": [ + { + "bbox": [ + 220, + 173, + 306, + 243 + ], + "spans": [ + { + "bbox": [ + 220, + 173, + 306, + 243 + ], + "type": "image", + "image_path": "b5bfc82b7b334beb34c57d3eff9acc483cbfe807d5b92ff22956e588b6743fa1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 173, + 393, + 243 + ], + "blocks": [ + { + "bbox": [ + 307, + 173, + 393, + 243 + ], + "lines": [ + { + "bbox": [ + 307, + 173, + 393, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 393, + 243 + ], + "type": "image", + "image_path": "1325199239a42b17dcfe76f72a3e448e6cab1ac0826877099534e9db2d8388db.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 394, + 173, + 480, + 243 + ], + "blocks": [ + { + "bbox": [ + 394, + 173, + 480, + 243 + ], + "lines": [ + { + "bbox": [ + 394, + 173, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 394, + 173, + 480, + 243 + ], + "type": "image", + "image_path": "02e55108d25f0081f4cdb6623ad11595b03086f5bdfce38a9462e415872e9f63.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 133, + 243, + 202, + 319 + ], + "blocks": [ + { + "bbox": [ + 133, + 243, + 202, + 319 + ], + "lines": [ + { + "bbox": [ + 133, + 243, + 202, + 319 + ], + "spans": [ + { + "bbox": [ + 133, + 243, + 202, + 319 + ], + "type": "image", + "image_path": "7caa7c6a2d016d5c843dfeb26973b03eff6203f70ab278fc85b81ef5db4cef0f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 203, + 243, + 272, + 319 + ], + "blocks": [ + { + "bbox": [ + 203, + 243, + 272, + 319 + ], + "lines": [ + { + "bbox": [ + 203, + 243, + 272, + 319 + ], + "spans": [ + { + "bbox": [ + 203, + 243, + 272, + 319 + ], + "type": "image", + "image_path": "b2213ab8689a9b81e00cf4590084b6a4f69c4f7126d8ad1b7d71b3d3fbb81b70.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 272, + 243, + 341, + 319 + ], + "blocks": [ + { + "bbox": [ + 272, + 243, + 341, + 319 + ], + "lines": [ + { + "bbox": [ + 272, + 243, + 341, + 319 + ], + "spans": [ + { + "bbox": [ + 272, + 243, + 341, + 319 + ], + "type": "image", + "image_path": "1df629466244146be371967528928d0f1a4795ab7a3d87ce078320badb88d704.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 342, + 243, + 411, + 319 + ], + "blocks": [ + { + "bbox": [ + 342, + 243, + 411, + 319 + ], + "lines": [ + { + "bbox": [ + 342, + 243, + 411, + 319 + ], + "spans": [ + { + "bbox": [ + 342, + 243, + 411, + 319 + ], + "type": "image", + "image_path": "dadb336317ebfdc6cdf802009298c99771540beb391081d93fc6de81870e388f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 411, + 243, + 480, + 319 + ], + "blocks": [ + { + "bbox": [ + 411, + 243, + 480, + 319 + ], + "lines": [ + { + "bbox": [ + 411, + 243, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 411, + 243, + 480, + 319 + ], + "type": "image", + "image_path": "0af6ef37a4a6283a521df2d5ac7eaf9f1618b94cc0401e39de2944b8285df794.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 133, + 320, + 202, + 380 + ], + "blocks": [ + { + "bbox": [ + 133, + 320, + 202, + 380 + ], + "lines": [ + { + "bbox": [ + 133, + 320, + 202, + 380 + ], + "spans": [ + { + "bbox": [ + 133, + 320, + 202, + 380 + ], + "type": "image", + "image_path": "fe1864a602c8beb04e24af78df2114ffd2d6fe9cd4225a415f4fdc4863c77e3c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 391, + 482, + 425 + ], + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 425 + ], + "type": "text", + "content": "Fig.8: We compare results from our method with a focus on color reconstruction against HDRUnet [6] and SingleHDR [24]. We refer to Sec. 5.2 and the supplementary material for an in-depth discussion." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 203, + 320, + 272, + 380 + ], + "blocks": [ + { + "bbox": [ + 203, + 320, + 272, + 380 + ], + "lines": [ + { + "bbox": [ + 203, + 320, + 272, + 380 + ], + "spans": [ + { + "bbox": [ + 203, + 320, + 272, + 380 + ], + "type": "image", + "image_path": "4b7367fc33f5f7d7825fb0255bc5036ade08a3040a5139f3ab7a02a4f71e8b77.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 272, + 320, + 341, + 380 + ], + "blocks": [ + { + "bbox": [ + 272, + 320, + 341, + 380 + ], + "lines": [ + { + "bbox": [ + 272, + 320, + 341, + 380 + ], + "spans": [ + { + "bbox": [ + 272, + 320, + 341, + 380 + ], + "type": "image", + "image_path": "547c2ab0a0540460c4752b12eee0d8e241bf7e8523ebe20cb0d537112080e4e8.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 342, + 320, + 411, + 380 + ], + "blocks": [ + { + "bbox": [ + 342, + 320, + 411, + 380 + ], + "lines": [ + { + "bbox": [ + 342, + 320, + 411, + 380 + ], + "spans": [ + { + "bbox": [ + 342, + 320, + 411, + 380 + ], + "type": "image", + "image_path": "601e42538306657d9deab810c93630f380a60ccd79072501a95aa50d632f199f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 411, + 320, + 480, + 380 + ], + "blocks": [ + { + "bbox": [ + 411, + 320, + 480, + 380 + ], + "lines": [ + { + "bbox": [ + 411, + 320, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 411, + 320, + 480, + 380 + ], + "type": "image", + "image_path": "63693be8cf6195708200005ad11e9d54dba18466f152c7e74118c3ff36e304fe.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 447, + 220, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 220, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 220, + 460 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 474, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 594 + ], + "type": "text", + "content": "In this work, we remodel the HDR reconstruction problem in the intrinsic domain and propose a multi-stage pipeline to recover lost details. The luminance range extension can be conveniently modeled in the shading space which allows the network to better model the correlation between illumination and geometry. Similarly, neural networks can easily take advantage of albedo sparsity to recover lost color information with high fidelity. Through qualitative and quantitative evaluations, we show that our physically-inspired modeling of the HDR reconstruction problem achieves state-of-the-art performance, improving upon prior work in terms of both detail recovery and color reproduction in a wide variety of in-the-wild photographs." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 614, + 247, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 614, + 247, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 614, + 247, + 628 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "We acknowledge the support of the Natural Sciences and Engineering Research Council of Canada (NSERC), [RGPIN-2020-05375]." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "S. Dille et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. Banterle, F., Ledda, P., Debattista, K., Chalmers, A.: Inverse tone mapping. In: Proc.GRAPHITE (2006)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "text", + "content": "2. Barrow, H., Tenenbaum, J., Hanson, A., Riseman, E.: Recovering intrinsic scene characteristics. Comput. vis. syst 2(3-26), 2 (1978)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 184, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 480, + 205 + ], + "type": "text", + "content": "3. Careaga, C., Aksoy, Y.: Intrinsic image decomposition via ordinal shading. ACM Trans. Graph. 43(1) (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 206, + 480, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 227 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 227 + ], + "type": "text", + "content": "4. Careaga, C., Miangoleh, S.M.H., Aksoy, Y.: Intrinsic harmonization for illumination-aware compositing. In: Proc. SIGGRAPH Asia (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "type": "text", + "content": "5. Chen, S.K., Yen, H.L., Liu, Y.L., Chen, M.H., Hu, H.N., Peng, W.H., Lin, Y.Y.: CEVR: Learning continuous exposure value representations for single-image HDR reconstruction. In: Proc. ICCV (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 261, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 480, + 281 + ], + "type": "text", + "content": "6. Chen, X., Liu, Y., Zhang, Z., Qiao, Y., Dong, C.: HDRUnet: Single image HDR reconstruction with denoising and dequantization. In: Proc. CVPR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 283, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 283, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 283, + 480, + 304 + ], + "type": "text", + "content": "7. Dang-Nguyen, D.T., Pasquini, C., Conotter, V., Boato, G.: RAISE: a raw images dataset for digital image forensics. In: Proc. MMSys (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 304, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 326 + ], + "type": "text", + "content": "8. Debevec, P.: A median cut algorithm for light probe sampling. In: ACM SIGGRAPH 2005 Posters. ACM (2005)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 327, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 480, + 347 + ], + "type": "text", + "content": "9. Debevec, P.E., Malik, J.: Recovering high dynamic range radiance maps from photographs. In: ACM Trans. Graph. (1997)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "type": "text", + "content": "10. Eilertsen, G., Kronander, J., Denes, G., Mantiuk, R.K., Unger, J.: HDR image reconstruction from a single exposure using deep CNNs. ACM Trans. Graph. 36(6) (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 381, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 480, + 403 + ], + "type": "text", + "content": "1. Endo, Y., Kanamori, Y., Mitani, J.: Deep reverse tone mapping. ACM Trans. Graph. 36(6) (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "type": "text", + "content": "2. Garces, E., Rodriguez-Pardo, C., Casas, D., Lopez-Moreno, J.: A survey on intrinsic images: Delving deep into lambert and beyond. Int. J. Comput. Vision (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 425, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 446 + ], + "type": "text", + "content": "3. Gilchrist, A., Jacobsen, A.: Perception of lightness and illumination in a world of one reflectance. Perception 13, 5-19 (02 1984)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "text", + "content": "4. Guo, C., Xiuhua, J.: LHDR: HDR reconstruction for legacy content using a lightweight DNN. In: Proc. ACCV (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 469, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 469, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 138, + 469, + 480, + 500 + ], + "type": "text", + "content": "5. Hanji, P., Mantiuk, R., Eilertsen, G., Hajisharif, S., Unger, J.: Comparison of single image HDR reconstruction methods — the caveats of quality assessment. In: ACM Trans. Graph. (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "type": "text", + "content": "6. Kim, D., Kim, J., Nam, S., Lee, D., Lee, Y., Kang, N., Lee, H.E., Yoo, B., Han, J.J., Kim, S.J.: Large scale multi-illuminant (LSMI) dataset for developing white balance algorithm under mixed illumination. In: Proc. CVPR (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "type": "text", + "content": "7. Kim, J., Lee, S., Kang, S.J.: End-to-end differentiable learning to HDR image synthesis for multi-exposure images. In: Proc. AAAI (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 556, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 480, + 578 + ], + "type": "text", + "content": "8. Kovaleski, R.P., Oliveira, M.M.: High-quality reverse tone mapping for a wide range of exposures. In: Proc. SIBGRAPI (2014)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "text", + "content": "9. Le, P.H., Le, Q., Nguyen, R., Hua, B.S.: Single-image HDR reconstruction by multi-exposure generation. In: Proc. WACV (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "text", + "content": "20. Lee, S., An, G.H., Kang, S.J.: Deep recursive HDRI: Inverse tone mapping using generative adversarial networks. In: Proc. ECCV (2018)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "text", + "content": "21. Li, Z., Snavely, N.: Learning intrinsic image decomposition from watching the world. In: Proc. CVPR (2018)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "22. Li, Z., Lu, M., Zhang, X., Feng, X., Asif, M.S., Ma, Z.: Efficient visual computing with camera raw snapshots. IEEE Trans. Pattern Anal. Mach. Intell. (2024)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "text", + "content": "Intrinsic Single-Image HDR Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 643 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 481, + 138 + ], + "type": "text", + "content": "23. Liu, L., Jiang, H., He, P., Chen, W., Liu, X., Gao, J., Han, J.: On the variance of the adaptive learning rate and beyond. In: Proc. ICLR (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 171 + ], + "type": "text", + "content": "24. Liu, Y.L., Lai, W.S., Chen, Y.S., Kao, Y.L., Yang, M.H., Chuang, Y.Y., Huang, J.B.: Single-image HDR reconstruction by learning to reverse the camera pipeline. In: Proc. CVPR (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 193 + ], + "type": "text", + "content": "25. Loshchilov, I., Hutter, F.: SGDR: Stochastic gradient descent with warm restarts. In: Proc. ICLR (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 481, + 225 + ], + "type": "text", + "content": "26. Mann, S., Picard, R.: On being \"undigital\" with digital cameras: Extending dynamic range by combining differently exposed pictures. In: Proc. IS&T Annual Conference (1995)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "text", + "content": "27. Mantiuk, R.K., Azimi, M.: Pu21: A novel perceptually uniform encoding for adapting existing quality metrics for hdr. In: Proc. PCS (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "28. Mantiuk, R.K., Hammou, D., Hanji, P.: HDR-VDP-3: A multi-metric for predicting image differences, quality and contrast distortions in high dynamic range and regular content. arXiv preprint arXiv:2304.13625 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "text", + "content": "29. Maralan, S.S., Careaga, C., Aksoy, Y.: Computational flash photography through intrinsics. In: Proc. CVPR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "30. Marnerides, D., Bashford-Rogers, T., Hatchett, J., Debattista, K.: Expandnet: A deep convolutional neural network for high dynamic range expansion from low dynamic range content. Comput. Graph. Forum 37(2) (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 337, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 481, + 357 + ], + "type": "text", + "content": "31. Murmann, L., Gharbi, M., Aittala, M., Durand, F.: A multi-illumination dataset of indoor object appearance. In: Proc. ICCV (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "text", + "content": "32. Nemoto, H., Korshunov, P., Hanhart, P., Ebrahimi, T.: Visual attention in LDR and HDR images. In: Proc. VPQM (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 481, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 412 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 412 + ], + "type": "text", + "content": "33. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Trans. Pattern Anal. Mach. Intell. 44(3) (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "text", + "content": "34. Rempel, A.G., Trentacoste, M., Seetzen, H., Young, H.D., Heidrich, W., Whitehead, L., Ward, G.: LDR2HDR: on-the-fly reverse tone mapping of legacy video and photographs. ACM Trans. Graph. 26(3) (2007)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 447, + 481, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 478 + ], + "type": "text", + "content": "35. Roberts, M., Ramapuram, J., Ranjan, A., Kumar, A., Bautista, M.A., Paczan, N., Webb, R., Susskind, J.M.: Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In: Proc. ICCV (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 500 + ], + "type": "text", + "content": "36. Santos, M.S., Ren, T.I., Kalantari, N.K.: Single image HDR reconstruction using a CNN with masked features and perceptual loss. ACM Trans. Graph. 39(4) (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 502, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 481, + 521 + ], + "type": "text", + "content": "37. Tan, M., Le, Q.: EfficientNet: Rethinking model scaling for convolutional neural networks. In: Proc. ICML (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "text", + "content": "38. Wang, C., Serrano, A., Pan, X., Chen, B., Myszkowski, K., Seidel, H.P., Theobalt, C., Leimkuhler, T.: Glowgan: Unsupervised learning of HDR images from LDR images in the wild. In: Proc. ICCV (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 555, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 577 + ], + "type": "text", + "content": "39. Xu, D., Doutre, C., Nasiopoulos, P.: Correction of clipped pixels in color images. IEEE Trans. Vis. Comput. Graph. 17(3) (2010)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 578, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 598 + ], + "type": "text", + "content": "40. Zhang, N., Ye, Y., Zhao, Y., Wang, R.: Revisiting the stack-based inverse tone mapping. In: Proc. CVPR (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 600, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 620 + ], + "type": "text", + "content": "41. Zhang, X., Brainard, D.H.: Estimation of saturated pixel values in digital color imaging. J. Optical Soc. Am. A 21(12) (2004)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "text", + "content": "42. Zhang, Y., Aydin, T.: Deep HDR estimation with generative detail reconstruction. Comput. Graph. Gorum 40(2), 179-190 (2021)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "S. Dille et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/IntrinsicAnything_ Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination/b4b4420f-76ed-4ddf-baa7-e217742f08de_content_list.json b/2024/IntrinsicAnything_ Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination/b4b4420f-76ed-4ddf-baa7-e217742f08de_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e1220d2a9125e64336895134346ee0646edeec39 --- /dev/null +++ b/2024/IntrinsicAnything_ Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination/b4b4420f-76ed-4ddf-baa7-e217742f08de_content_list.json @@ -0,0 +1,2665 @@ +[ + { + "type": "text", + "text": "Intrinsic Anything: Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination", + "text_level": 1, + "bbox": [ + 250, + 140, + 754, + 207 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xi Chen $^{1}$ , Sida Peng $^{1}$ , Dongchen Yang $^{1}$ , Yuan Liu $^{2}$ , Bowen Pan $^{3}$ , Chengfei Lv $^{3}$ , and Xiaowei Zhou $^{1\\dagger}$", + "bbox": [ + 313, + 234, + 689, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 State Key Lab of CAD&CG, Zhejiang University", + "bbox": [ + 328, + 276, + 669, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 The University of Hong Kong", + "bbox": [ + 393, + 290, + 607, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Tao Technology Department, Alibaba Group", + "bbox": [ + 344, + 305, + 656, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. This paper aims to recover object materials from posed images captured under an unknown static lighting condition. Recent methods solve this task by optimizing material parameters through differentiable physically based rendering. However, due to the coupling between object geometry, materials, and environment lighting, there is inherent ambiguity during the inverse rendering process, preventing previous methods from obtaining accurate results. To overcome this ill-posed problem, our key idea is to learn the material prior with a generative model for regularizing the optimization process. We observe that the general rendering equation can be split into diffuse and specular shading terms, and thus formulate the material prior as diffusion models of albedo and specular. Thanks to this design, our model can be trained using the existing abundant 3D object data, and naturally acts as a versatile tool to resolve the ambiguity when recovering material representations from RGB images. In addition, we develop a coarse-to-fine training strategy that leverages estimated materials to guide diffusion models to satisfy multi-view consistent constraints, leading to more stable and accurate results. Extensive experiments on real-world and synthetic datasets demonstrate that our approach achieves state-of-the-art performance on material recovery. The code will be available at https://zju3dv.github.io/IntrinsicAnything/.", + "bbox": [ + 261, + 343, + 743, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 655, + 375, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recovering the object's geometry, material, and lighting from captured images, also known as inverse rendering, is a long-standing task in computer vision and graphics. These physical attributes of 3D objects are essential for many applications, such as VR/AR, movie production, and video games. Due to the inherent complexity of interaction between real-world objects and environment lighting, inverse rendering remains an ill-posed problem. Previous works overcome this problem using sophisticated capture systems [16,19], or co-located flashlight and camera in a dark environment [5,44,77]. However, these methods require special hardware equipment or constrained environment, limiting their applications.", + "bbox": [ + 212, + 681, + 787, + 819 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 232, + 825, + 392, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d6a8324d8e7504e543a22a6d863ec9b12a76bbadfddc5b5d244daf61b216b48.jpg", + "image_caption": [ + "Fig. 1: Two types of ambiguities in inverse rendering. (a) Ambiguity between diffuse shading and albedo. For example, the Xbox is lit by a yellow light, and the decomposed albedo from NVdiffrecMC [20] tends to be yellow. (b) Ambiguity between shadow and albedo. For example, the porcelain toy is with self-occlusion, and TensoIR [24] bakes the shadow into the recovered albedo. Our method well handles the two types of ambiguities." + ], + "image_footnote": [], + "bbox": [ + 233, + 141, + 787, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on differentiable physically based rendering and neural scene representations, recent methods [20, 24, 76, 78, 80, 81] are able to object materials from images captured under natural lighting. They exploit neural networks to represent object materials and geometry, and combine these object attributes with the learnable lighting to synthesize images, which are compared with captured images to optimize the model paramaters. To achieve better performance of inverse rendering, these methods have attempted to improve object representations [24, 78], lighting representations [20, 76], and training strategies [80, 81]. However, we argue that depending solely on the rendering loss is insufficient for accomplishing accurate decomposition, due to the inherent ambiguity in the inverse rendering process.", + "bbox": [ + 212, + 409, + 787, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We observed that there are two categories of ambiguities, as shown in Fig. 1. First, there is ambiguity between diffuse shading and albedo. Intuitively, when a white object is lit by a yellow light, the decomposed albedo appears to be yellow. Second, when an object is with self-occlusion and illuminated under a complex lighting condition, the hard-cast shadow can be easily baked into the albedo. Due to these ambiguities, the optimization process of previous methods tend to gets trapped in local optimal, leading to the artifacts on the recovered material.", + "bbox": [ + 212, + 579, + 787, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we seek to use data-driven priors for ambiguity-free inverse rendering. Our key idea is learning the distribution of albedo and specular shading with generative models, which is then used to regularize the optimization process. Inspired by the recent progress of generative models [49], we exploit the conditional diffusion model for modeling the material prior. Our choice of material prior for albedo and specular shading is motivated by the fact that the general rendering equation can be separated into diffuse and specular terms, according to the Disney BRDF model, which thus has two advantages. First, we can construct training data on albedo and specular using a large amount of 3D objects [17] that contain various BRDF models. Second, our designed prior can", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Chen. et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "work as a versatile tool to resolve the ambiguity in the inverse rendering process of many material representations in theory.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A challenge of applying diffusion models to regularize the optimization process is that such generative model struggles to perform multi-view consistent constraints. To address this problem, we propose a coarse-to-fine optimization scheme. Our approach first leverages the diffusion prior to recover the coarse object material and environment lighting. Then, the coarse albedo and specular shading are used to guide the diffusion model to produce constraints with better multi-view consistency, leading to more accurate material and lighting recovery.", + "bbox": [ + 212, + 176, + 785, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For training our model, we create a dataset consisting of RGB, albedo, and specular images based on the Objverse dataset [17]. To validate the effective of the proposed approach, we conduct extensive experiments on synthetic and real-world data. Our method achieves state-of-the-art performance on material recovery and relighting.", + "bbox": [ + 212, + 282, + 785, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 377, + 387, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning priors for inverse rendering. Earlier approaches for inverse rendering usually require controlled set-ups [16,30] and strong domain-specific assumptions to reconstruct scene properties from images [74]. With the advance of deep learning, the intrinsic properties of images can be directly learned from large-scale synthetic datasets [32, 33, 54]. Shi et al. [54] leverages ShapeNet [9] and introduces a dataset with image intrinsics to decompose the image into base color and shading. Li et al. [31] introduces a large-scale scene dataset and lifts the model's capability from object level to scene level. Subsequent works [32, 35, 53] have attempted to further recover the BRDF and geometry of a scene. However, obtaining massive high-quality real-world images with ground truth intrinsic decomposition is a significant challenge, which limits models' generalization to real-world data. To mitigate this issue, unsupervised methods [38, 64, 72, 72, 73] and weakly supervised methods [71] seek specially designed training scheme to reduce dependency on labeled data. Recently, [83] proposed a large-scale photorealistic indoor scene dataset InteriorVerse for single-view inverse rendering. Kocsis et al. [28] further uses InteriorVerse [83] to train a diffusion model for single-view material recovery in indoor scenes, boosting the generalization to real-world images. Recent methods [39, 46, 62] also explore diffusion models for texture-related tasks. unidream [39] and Richdreamer [46] leverage diffusion model on albedo generation for reconfigurable 3D mesh generation. ControlMat [62] focuses on capturing tileable SVBRDF from a single image with relatively simple shapes (e.g. a plane) and lighting. Although the above learning-based methods achieve promising progress in inverse rendering, their single-view inference pipeline can't guarantee multi-view consistency. In contrast, our method combines the monocular diffusion models with the multiview physically based rendering and thus achieves accurate material and illumination estimation with multiview consistency.", + "bbox": [ + 212, + 406, + 787, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inverse rendering with neural representation. Inspired by the implicit representations of NeRF [42], various works [7,24,37,78,80] extend the neural representation", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Intrinsic Anything", + "bbox": [ + 611, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tions to factorize a scene into material, geometry, and lighting. NeRV [56] leverages a neural visibility field to reduce the computational cost and models indirect lighting with known environment lighting conditions. InvRender [81] and NeILF series [68,76] further model indirect illumination with a neural representation at unknown light conditions. NvdiffrecMC [20] and NeFII [65] incorporate Monte Carlo ray tracing to recover high-frequency details of lighting. [57] adopts a coarse-to-fine manner to retrieve the materials and illumination of a scene. More recently, GS-IR [34] and GIR [55] leverage Gaussian Splatting [27] to model indirect lighting for inverse Rendering. Furthermore, [66, 69, 82] attempt to scale up the neural representations for solving inverse rendering from object to indoor scene. However, due to the inherited ill-posed feature of the inverse rendering problem, these methods still suffer from disentangling the materials and illumination properties from images. In contrast, by learning a strong material and shading prior, we effectively tackled the ill-posed feature and received a significant improvement in decomposing the image into BRDFs and illuminations.", + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion models. Recently, diffusion models have demonstrated unprecedented performance on generation tasks, such as 2D generation [15, 21, 49, 52] and 3D generation [3, 11, 12, 36, 45, 59], indicating the impressive capability of learning the distribution of target data. Inspired this, some recent methods [14, 26] have attempted to exploit the diffusion prior for ill-posed inverse problems, including image denoising [67, 84], image completion [23, 51], and single-view reconstruction [18, 58]. Diffusion posterior illumination [40] learns the distribution of environment illumination with a diffusion model for material recovery. In contrast, our method learns the material prior for supervision and design it as conditional diffusion models of albedo and specular shading, enabling the direct supervision on material estimation.", + "bbox": [ + 212, + 387, + 787, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Background", + "text_level": 1, + "bbox": [ + 215, + 575, + 369, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Physically-Based Rendering", + "text_level": 1, + "bbox": [ + 215, + 608, + 493, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We adopt the physically based rendering [25] to recover the object material and lighting from images. For a 3D point $\\mathbf{x}$ , its outgoing light $L_{o}$ at direction $\\omega_{o}$ is defined as:", + "bbox": [ + 212, + 633, + 787, + 675 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {o} \\left(\\omega_ {o}; \\mathbf {x}\\right) = \\int_ {\\Omega} L _ {i} \\left(\\omega_ {i}\\right) f _ {r} \\left(\\mathbf {x}; \\omega_ {o}, \\omega_ {i}\\right) \\left(\\omega_ {i} \\cdot \\mathbf {n}\\right) d \\omega_ {i}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 676, + 785, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L_{i}(\\omega_{i})$ is the incident lighting along the direction $\\omega_{i}$ , $\\mathbf{n}$ is surface normal at $\\mathbf{x}$ . In practice, we define the lighting $L_{i}$ as an estimable longitude-latitude environment map. We use the simplified Disney BRDF model [8] for $f_{r}$ :", + "bbox": [ + 212, + 710, + 787, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {r} (\\mathbf {x}; \\omega_ {o}, \\omega_ {i}) = \\frac {\\mathbf {k} _ {d}}{\\pi} + \\frac {D \\cdot G \\cdot F}{4 (\\mathbf {n} \\cdot \\omega_ {\\mathbf {o}}) (\\mathbf {n} \\cdot \\omega_ {i})}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 768, + 785, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D, G$ , and $F$ are the normal distribution, geometry attenuation, and Fresnel effect, respectively. The albedo $\\mathbf{k}_d$ and roughness $\\rho$ parameters are set as", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Chen. et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "learnable, and other parameters are set as constant. The detailed formulation can be found in supplementary material.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The RGB image is rendered by evaluating Eq. (1) with the environment lighting, albedo image, roughness image, and scene mesh. We solve the integral in Eq. (1) using Monte-Carlo sampling with Multiple Importance Sampling [61].", + "bbox": [ + 212, + 176, + 784, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Diffusion Model and Diffusion Posterior Sampling", + "text_level": 1, + "bbox": [ + 214, + 243, + 671, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The diffusion model is a probabilistic generative model that generates images from pure Gaussian noise through a progressive de-noising process [21]. Each step of the denoising process from $\\mathbf{x}_t$ to $\\mathbf{x}_{t - 1}$ is defined as:", + "bbox": [ + 212, + 267, + 782, + 313 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha (t)}} \\left(\\mathbf {x} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\hat {\\alpha} _ {t}}} \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t)\\right) + \\sigma_ {t} \\mathbf {z}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 323, + 784, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}$ is the learnable noise predictor, $\\alpha_{t}$ and $\\sigma_{t}$ are predefined constant values, and $\\mathbf{z}$ is a random value sampled from a standard Gaussian distribution. The noise predictor is typically implemented as a UNet [50].", + "bbox": [ + 212, + 366, + 782, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The diffusion posterior sampling (DPS) [13] is a technique for solving inverse problems (e.g. image deblur) using the diffusion model. Specifically, given an observation $\\mathbf{y}$ derived from $\\mathbf{x}_0$ which defined as:", + "bbox": [ + 212, + 411, + 782, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} = \\mathcal {A} \\left(\\mathbf {x} _ {0}\\right) + \\mathbf {n} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 447, + 468, + 784, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{A}$ is the measurement operator, $\\mathbf{n}$ is the measurement noise. The goal is to estimate $\\mathbf{x}_0$ from observation $\\mathbf{y}$ using posterior distribution $p(\\mathbf{x}_0|\\mathbf{y})$ . To sample from $p(\\mathbf{x}_0|\\mathbf{y})$ with a pre-trained distribution $p(\\mathbf{x}_0)$ , the DPS sampling approximates the posterior sampling functions by adding a guiding term to the denoising score $\\epsilon_{\\theta}(\\mathbf{x}_t,t)$ in Eq. (3):", + "bbox": [ + 212, + 493, + 784, + 569 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\epsilon} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right) = \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right) + \\zeta_ {t} \\nabla_ {\\mathbf {x} _ {t}} \\left| \\left| \\mathcal {A} \\left(\\mathbf {x} _ {t} ^ {0}\\right) - \\mathbf {y} \\right| \\right| _ {2} ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 579, + 784, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_t^0$ is the posterior estimation of $\\mathbf{x}_0$ at timestamp $t$ and $\\zeta_t$ is the step size. This formula assumes a known $\\mathcal{A}$ and $\\mathbf{n}$ follows a Gaussian distribution of $\\mathcal{N}(\\mathbf{0},\\sigma^2\\mathbf{I})$ . We refer to DPS [13] for more details. In Section 4, we will demonstrate how to model the material priors using the diffusion model and use DPS sampling to generate multi-view consistent material samples.", + "bbox": [ + 212, + 604, + 784, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 214, + 703, + 330, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method aims at recovering material and lighting parameters from multi-view images under a single unknown illumination. We define object material as albedo and roughness without considering the metallic attribute and use a lat-long environment map for lighting. Due to the ill-posed nature of inverse rendering [20, 81], ambiguities exist in the optimization process, leading to inaccurate material estimation. We propose using data-driven priors to constrain the optimization to eliminate these ambiguities.", + "bbox": [ + 212, + 733, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Intrinsic Anything", + "bbox": [ + 611, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4243a67edf082b2aa8a26b4b13b784a6f06ed88b91202f3d585c247a089d92e2.jpg", + "image_caption": [ + "Fig. 2: Overview of our pipeline. (a) Based on physically based rendering, our model combines geometry, lighting, roughness, and albedo into RGB and specular images, and optimizes the lighting and materials in a two-stage manner. In the first stage, our model is supervised by images and diffusion priors to output coarse albedo and roughness. Subsequently, the coarse materials are used to guide diffusion models to provide more multi-view consistent constraints. (b) The guided sampling first calculates the L2 loss between the guidance and one-step denoised signals, and then adds the gradient of the L1 loss to the output of the noise predictor." + ], + "image_footnote": [], + "bbox": [ + 223, + 157, + 560, + 321 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b28fce0c0667b4d0ae45e1416d86fe95e7c04dfb24c132185a06ce7e608ee31f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 157, + 779, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inspired by the recent success of generative models [49] on image generation, we design two material priors modeled by the conditional diffusion model (Sec. 4.1). To recover object materials and lighting, we first employ monosdf [75] to reconstruct a 3D mesh from input images. The materials are defined as learnable texture maps in UV space from the UV-mapped mesh using Blender [6], and lighting as estimizable longitude-latitude environment map. Then we apply our coarse-to-fine optimization process (Sec. 4.2) to recover accurate materials and lighting, while the geometry and the prior models are fixed. An overview of our optimization process is shown in Fig 2.", + "bbox": [ + 212, + 478, + 787, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Albedo and Specular Prior Models", + "text_level": 1, + "bbox": [ + 215, + 657, + 550, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Design of the material priors. To better decompose materials and lighting, we propose to use data-driven priors to constrain the optimization process. Ideally, we can directly apply prior distributions of albedo and roughness to narrow down the solution space and effectively alleviate ambiguities. However, roughness, being a concealed characteristic of a material, is difficult to observe and is also hard to acquire on most large-scale synthetic datasets [17]. Because these objects have a variety of material parameters that do not necessarily have a valid roughness.", + "bbox": [ + 212, + 686, + 787, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fortunately, image rendering using the Disney BRDF model [8] is composed of specular shading and diffuse shading regardless of the choice of material parameters, as shown in Eq.(2). Thus, the rendering equation can be rewritten to:", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Chen. et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {o} (\\omega_ {o}; \\mathbf {x}) = \\frac {\\mathbf {k} _ {d}}{\\pi} \\int_ {\\Omega} L _ {i} (\\omega_ {i}) (\\omega_ {i} \\cdot \\mathbf {n}) d \\omega_ {i} + \\int_ {\\Omega} L _ {i} (\\omega_ {i}) \\frac {D \\cdot G \\cdot F}{4 (\\mathbf {n} \\cdot \\omega_ {\\mathbf {o}}) (\\mathbf {n} \\cdot \\omega_ {i})} (\\omega_ {i} \\cdot \\mathbf {n}) d \\omega_ {i} \\\\ = S _ {\\text {d i f f}} (\\mathbf {x}, L _ {i}, k _ {d}) + S _ {\\text {s p e c}} \\left(\\omega_ {o}; \\mathbf {x}, \\rho , L _ {i}\\right). \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 172, + 784, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motivated by Eq. (6), we propose to apply priors on specular shading which exists in all physically based renderings, instead of directly using priors on roughness. Specular shading is closely linked to roughness and has the advantage of being more observable. Moreover, based on any synthetic object, we can create the ground-truth specular shading for training the prior model.", + "bbox": [ + 214, + 253, + 782, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Modeling the material priors. Previous methods [31, 35, 54, 70, 72, 73] generally model the prior of object material using a deterministic prediction network, which attempts to regress the material from a single image. In contrast, we leverage generative models to capture the distribution of albedo and specular, considering that there could be multiple solutions for the underlying object material given captured images [4].", + "bbox": [ + 214, + 351, + 782, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given an RGB image $I$ , we can model the conditional distribution of target albedo $p_{\\theta}(\\mathbf{k}_d|I)$ and specular shading $p_{\\theta}(S_{\\mathrm{spec}}|I)$ by modifying the noise predictor to accept $I$ as its conditioning. Denote the modified noise predictor as $\\epsilon_{\\theta}(\\mathbf{x}_t,t;I)$ . Specifically, we first use the CLIP image encoder [47] to extract a feature vector from the conditioning image. Then, to inject the conditioning signal into the UNet, we utilize a transformer network [60] to perform the cross-attention between the image feature vector and the intermediate feature maps of the UNet. More details can be found in the supplementary.", + "bbox": [ + 214, + 444, + 784, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The diffusion model can model complex distributions in a step-wise denoising manner, which is crucial for albedo and specular shading recovery due to the complexity of lighting and material in real-world data. Moreover, as a generative model, the diffusion model facilitates a guided sampling procedure [13, 22, 41], which permits us to steer the diffusion model toward producing samples that align with multi-view image observations.", + "bbox": [ + 214, + 566, + 784, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Material Prior for Inverse Rendering", + "text_level": 1, + "bbox": [ + 215, + 686, + 568, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on the albedo and specular shading prior models, we design a two-stage optimization process for inverse rendering, which effectively alleviates the ambiguities. An overview of our method is presented in Fig. 2. We first use the conditional diffusion model to regularize the inverse rendering process, which outputs an initialization of lighting, albedo, and roughness. Then, the coarse material is used to guide the sampling of the diffusion model, enabling it to produce more deterministic results and better stabilize the optimization process.", + "bbox": [ + 214, + 719, + 782, + 825 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Intrinsic Anything", + "bbox": [ + 611, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/59d7a59039138aac4981f62a589e84b9dcbf8120f252738062a6b2aa48027378.jpg", + "image_caption": [ + "Input View" + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 282, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac7f5d1df27f44e3d3328745738a092d83775572510de0e6c2dcaa742371af6d.jpg", + "image_caption": [ + "Reference" + ], + "image_footnote": [], + "bbox": [ + 290, + 147, + 336, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f4e8fb00a4f30a28f0048186c15868b78f96fad854cfb572201d657f755855bb.jpg", + "image_caption": [ + "Fig. 3: Effect of the guided sampling. We visualize the samples generated from the albedo prior model of different viewpoints. Without the guided sampling, the materials are inconsistent across multi-views and do not align with the material decomposition from the observed lighting." + ], + "image_footnote": [], + "bbox": [ + 351, + 150, + 398, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/89fc32e2c9da09fa1c50c9e9e90f1eca124e78f1f8965a2c9da563099c557f36.jpg", + "image_caption": [ + "w/o guidance" + ], + "image_footnote": [], + "bbox": [ + 405, + 151, + 452, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/582351d10cf2eadc149d0d83865d5ed324a2e39e342029464234aca8308f872f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 152, + 506, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/55281b6ce5962942555c63cf7f320e7478a94cae361ce5a4f84e2e2d57fe2883.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 150, + 558, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8874c50f029fad77faa398a7bbda971d2cd32d06d605771a323fb763567ca7f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 150, + 612, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1e95969c6d881cb64bbf42cb823259e8cad71c83fbcfb90cdddd20292da13c75.jpg", + "image_caption": [ + "w/guidance" + ], + "image_footnote": [], + "bbox": [ + 619, + 151, + 669, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3a04067962f9c1422c0fc14161d1c0fd7c60639c54165224b9df153397a38505.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 151, + 723, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7adbb3cda7d53baeb6ec15651f8a2126c9ea3ee3a22b7c28fee18d3fdf959016.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 150, + 776, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training with diffusion model. Given input multi-view images of the target object, we utilize the conditional diffusion model to predict their albedo and specular shading components, which are used to supervise the model parameters. Due to the probabilistic nature of the diffusion model, the generated albedo and specular images are inconsistent across camera views in terms of intensity. To overcome this problem, we apply shift and scale invariant loss [48] to compute the difference between the predicted and generated material.", + "bbox": [ + 212, + 297, + 784, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the first stage, the training loss function is defined as:", + "bbox": [ + 238, + 402, + 650, + 417 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o a r s e}} = \\lambda_ {1} \\| \\hat {I} - I \\| + \\lambda_ {2} \\| \\operatorname {S S I} \\left(\\hat {\\mathbf {k}} _ {d}, \\mathbf {k} _ {d} ^ {\\mathrm {s}}\\right) - \\mathbf {k} _ {d} ^ {\\mathrm {s}} \\| \\tag {7} \\\\ + \\lambda_ {3} | | \\hat {S} _ {\\text {s p e c}} - S _ {\\text {s p e c}} ^ {\\mathrm {s}} | | + \\lambda_ {4} \\mathcal {L} _ {\\text {s m o o t h}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 429, + 784, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathrm{SSI}(a,b)$ aligns $a$ to the scale and mean of $b$ . $\\hat{I}$ , $\\hat{k}_d$ , $\\hat{S}_{\\mathrm{spec}}$ are the image, albedo, and specular shading to be optimized respectively, $\\mathbf{k}_d^{\\mathrm{s}}$ and $S_{\\mathrm{spec}}^{\\mathrm{s}}$ are samples from our diffusion models $p_\\theta (\\mathbf{k}_d|I)$ and $p_{\\theta}(S_{\\mathrm{spec}}|I)$ respectively. We also use a smoothness constraint $\\mathcal{L}_{\\mathrm{smooth}}$ on the optimized materials. $\\lambda_1$ , $\\lambda_2$ , $\\lambda_3$ , $\\lambda_4$ are hyper-parameters. More details can be found in the supplementary material.", + "bbox": [ + 212, + 481, + 787, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training with guided diffusion samples. After the training in the first stage, we obtain an initialization of the object material and the environment lighting. To better guide the optimization process, we aimed to generate samples from the diffusion prior models that are consistent with multiple viewpoints, aligned with observed lighting conditions, and still follow the distribution of priors. To this end, we leverage the Diffusion Posterior Sampling (DPS) [13] to guide the sampling process, and use the estimated albedo $\\hat{\\mathbf{k}}_d$ and specular shading $\\hat{S}_{\\mathrm{spec}}$ as guiding signals to re-generate samples. As shown in Fig. 3, samples after guidance are multi-view consistent and fit observations under the current lighting condition. Specifically, we use the score function of DPS in Eq. (5) with $\\mathcal{A}$ as the identity operator to re-generate albedo $\\tilde{\\mathbf{k}}_d$ and specular $\\tilde{S}_{\\mathrm{spec}}$ :", + "bbox": [ + 212, + 571, + 787, + 739 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\epsilon} _ {\\theta} \\left(\\tilde {\\mathbf {x}} _ {t}, t; I\\right) = \\epsilon_ {\\theta} \\left(\\tilde {\\mathbf {x}} _ {t}, t; I\\right) + \\gamma \\bigtriangledown_ {\\tilde {\\mathbf {x}} _ {t}} \\left| \\left| \\tilde {\\mathbf {x}} _ {t} ^ {0} - \\hat {\\mathbf {x}} \\right| \\right| _ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 750, + 784, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{x}}_t$ and $\\tilde{\\mathbf{x}}_t^0$ is the noisy latents and the posterior estimation of $\\hat{\\mathbf{k}}_d$ or $\\tilde{S}_{\\mathrm{spec}}$ on step $t$ , respectively. $\\hat{\\mathbf{x}}$ is $\\hat{\\mathbf{k}}_d$ or $\\hat{S}_{\\mathrm{spec}}$ . We choose the step size following [13] as $\\zeta_i = \\gamma / ||\\tilde{\\mathbf{x}}_t^0 - \\hat{\\mathbf{x}}||$ , where $\\gamma$ is a pre-defined scalar. Finally, we use the re-generated multiview-consistent material $\\tilde{\\mathbf{k}}_d$ or $\\tilde{S}_{\\mathrm{spec}}$ from the DPS to supervise the lighting", + "bbox": [ + 212, + 776, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Chen. et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and materials using Eq. (7). Experiments show that the second stage gives an accurate decomposition of the object material and environment lighting.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "An alternative way to supervise the inverse rendering with diffusion models is the Score Distillation Sampling (SDS) [45], which achieves consistency by finding a mode in the distribution through optimization of the material field. However, our experiments find that the SDS guidance usually produces blurry results. A possible reason is that the SDS guidance relies on noise prediction that has a high variance, which leads to unstable convergence as discussed in [63]. In constraint, our pipeline uses guided image samples that are consistent with low variance.", + "bbox": [ + 212, + 176, + 787, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 304, + 375, + 321 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Implementation details", + "text_level": 1, + "bbox": [ + 214, + 335, + 454, + 349 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Prior Model Training. We create albedo and specular training data based on Objaverse dataset [17]. We randomly select 350k objects, excluding those with pure black or white albedos, and gather 180 HDR environment maps from [1] for rendering. Each object is rendered under 10 random poses with one random HDR environment map. We use a diffusion model fine-tuned from stable-diffusion [49] for the image variation task [2] as a starting point and continue to train 100k iterations for the albedo prior and 350K iterations for the specular shading prior. The training process for the two prior models takes around 10 days on 8 NVIDIA A100 (80GB) GPU.", + "bbox": [ + 212, + 359, + 787, + 494 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Inverse Rendering. The resolutions for texture maps and environment are $2048 \\times 2048$ and $512 \\times 256$ , respectively. We use Nvdiffrast [29] for rendering images with 144 samples per pixel. We train the first stage of our method for 20k iterations and the second stage for 15k iterations. The setting of hyperparameters is presented in the supplement materials. The inverse rendering for each scene takes around 1.5 hours on a single NVIDIA RTX A6000 GPU.", + "bbox": [ + 212, + 508, + 787, + 602 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.2 Evulation", + "text_level": 1, + "bbox": [ + 214, + 619, + 344, + 635 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset. We collect 4 CAD models with a variety of object shapes and materials. Each object is rendered with 120 posed RGBA images using Blender Cycles [6]. We use 100 images for training and 20 for testing. For evaluation, we also render 20 novel test views under novel lights along with their ground truth BRDFs.", + "bbox": [ + 212, + 643, + 785, + 705 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines. We compare our methods with the following two types of baseline methods. (1) Optimization-based methods [24, 43, 76, 81, 81]. InvRender [81] and TensoIR [24] both models indirect illumination with MLPs and uses Spherical Gaussian(SG) as environment light. InvRender [81] uses MLP representations while TensoIR is inspired by tensorial representation [10]. NeILF++ [76] models an incident light field that includes both environemnt light and indirect light with an MLP constrained by the outgoing radiance field. NvDiffRec [43], and NvDiffrecMC [20] extracted explicit topology from neural SDF and directly optimized", + "bbox": [ + 212, + 719, + 787, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Intrinsic Anything", + "bbox": [ + 611, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/f5106305c0d541e6ca0dd44c2671893fbac35746c5b054dbf34c02a3e72dfc56.jpg", + "table_caption": [ + "Table 1: Quantitative results on the synthetic dataset. \"NvDiffrec*\" and \"NvDiffrecMC*\" mean these methods are with our reconstructed mesh. \"Ours w/o guided\" means our method without the guided sampling strategy. TensorIR and NvDiffrec have strong capabilities of fitting observed images, but they struggle to recover correct object materials. Replacing our diffusion priors with other priors [38, 71] leads to the degradation of the performance on material recovery." + ], + "table_footnote": [], + "table_body": "
MethodAlbedoAligned AlbedoView SynthesisRelightingRoughness
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓MSE↓
InvRender [81]17.81810.92470.064523.75920.94910.055927.30510.96730.038624.36320.96210.05150.0420
TensoIR [24]20.29720.95270.063124.93680.95640.061433.12340.98030.026421.34730.95430.05140.0456
NvDiffrec [43]16.54670.92400.070621.53040.95070.057432.57780.98240.019021.99390.95500.07340.0777
NvDiffrec* [43]14.52040.90210.083520.22800.92040.065931.03360.97840.021821.24520.93560.06380.0841
NvDiffrecMC [20]13.04950.88150.106821.67700.94060.069930.10830.96740.044022.53370.94860.07970.1046
NvDiffrecMC* [20]11.13530.87220.117619.83250.93050.080528.16340.95600.044021.03800.93320.08220.1411
NEILF++* [76]13.83360.91720.091518.45430.93660.083129.35270.97750.027620.48780.96290.05900.0855
Yi et al. [71]13.20290.89340.083318.84280.91330.073126.50910.96690.039321.1560.95620.05070.1524
USI3D [38]11.67130.88720.105118.92960.93440.077226.96530.96660.048020.27210.95870.06120.1241
Ours24.67660.95620.050326.88290.95320.037830.24050.97410.032828.97210.97570.02890.0105
", + "bbox": [ + 225, + 239, + 784, + 375 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d9a8da1910ec10810d4118932714860b9647890c877d26017fd6c8dbe89b9c86.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparison in terms of relighting on the synthetic dataset. Zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 217, + 388, + 500, + 547 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/7af24bb2ef373a8441084b1cdf00e41b80fc31325de940550238d7a004acc8c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 388, + 784, + 546 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "the 3D mesh with its PBR materials. Their illumination is expressed as an HDR environment map, which can be optimized by NviDiffRec's [43] split sum environment lighting and NviDiffRec's Monte Carlo ray tracing. (2) Data-driven methods [38, 71]. We keep object and lighting models as the same and just replace diffusion priors with single-view intrinsic image decomposition methods for comparison. USI3D [38] models the decomposition task as domain transfer from the image domain to the albedo or shading domain using generative models. We adopt only the predicted albedo to supervise the optimization because their rendering model is not physically based. Yi. el [71] propose a network to predict and remove specular shading and subsequently recover the albedo. We use the predicted specular shading and albedo as the priors.", + "bbox": [ + 212, + 613, + 787, + 781 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation protocol. We analyze our method and the baselines in the following way. On the 4 synthetic datasets, we compare their reconstruction of albedo, roughness, and novel view synthesis quantitatively with the ground truths. Given", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Chen. et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c0af9d936d35d0b0a761b2df22657ff88a8a34d1af29b93168410777dbe62975.jpg", + "image_caption": [ + "Fig. 5: Quantitative comparison with baselines on the synthetic dataset. Our method estimates more accurate albedo and roughness than other methods. Nvdiffrec [43] and NvdiffrecMC [20] tend to bake the lighting into the albedo, due to the ambiguity between the diffuse shading and albedo. InvRender [81] and TensoIR [24] struggle to decouple the roughness from the albedo and environment lighting." + ], + "image_footnote": [], + "bbox": [ + 218, + 147, + 779, + 375 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/76407c55ede19a6e8c32707f3c943b1d6d29f5e4e3562b7c0e4eb052fc326e2c.jpg", + "image_caption": [ + "Fig.6: Qualitative results on DTU. Our method recovers reasonable materials and relighting on real datasets. Baseline models can not disentangle illumination and materials which results in the baked-in effect on materials estimation." + ], + "image_footnote": [], + "bbox": [ + 215, + 465, + 787, + 743 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "IntrinsicAnything", + "bbox": [ + 611, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "GT", + "bbox": [ + 233, + 157, + 254, + 169 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b8fab22c6ae49b02c1fb1b66808fd826077a329c35f4ce7e485ff5e8eedb74af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 148, + 331, + 186 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/95031cf721cb4b9032ed88964cc0a3d204cf15d86ef154e9160bcc258482c4a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 148, + 392, + 186 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/41e69272b99b23cc41bd26c8e753793d23db5ec795ad43fd1567ea02774ea5fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 148, + 457, + 186 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9d258d7ff7e46c6fcb826403db9ef4e571e00eb2d3fc21c8cd1899aedafc9927.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 460, + 148, + 519, + 186 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2ef8a966bc475e3fe44ba16a17af541baa8a77e51f8f4e3e520365d98ec784ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 148, + 575, + 190 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/710fc2037e5667e2322b0ac0d01ca380a333cbf4cc6131a7b39ef51caab513aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 584, + 147, + 640, + 191 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1ca721e7b50404ea76ed802a5ad335983db7dfff6080775cf86fe451b301adbb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 147, + 702, + 191 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7a2e6a16ce625920ea6c051b73b4feba74d0b9e88c173d614bc1028a246f7ceb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 147, + 767, + 191 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "USIBD", + "bbox": [ + 232, + 208, + 263, + 217 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/da9d0c7a36aaf9d34a4fa74a7e1e7dc75f66202a2f533501f84be66300b5e51f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 199, + 331, + 234 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/07f1738bd7ce200dcaf7b8cd70fe1ccc26be10f239c15616bf8d269b2e84c092.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 196, + 393, + 234 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/047963e72477ca14189977740918482cf11d1014f70b3b72bd95392716b0fc2e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 196, + 457, + 234 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/49fe7d179dbadb95ff6ca4f85fc4622226ed4c97760aa1e248cb16c0906f96f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 460, + 196, + 521, + 234 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9513a7232b5154aaa08b2b03c1c107c4f53151ae4134941396070890558afa9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 195, + 576, + 236 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/23984463a674afed8486ccd9019b495eb67fcf9cecd799f6e5f3c463c24c80bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 584, + 195, + 640, + 238 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/52a1eae2f21686a1686694923c30c7f649b64cd024b98c5f181bb95eae8a855b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 195, + 702, + 238 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/693d1e8184130e652d34038f9bb80e33e56d2aeb40976ed81b8cab21c0a7a6f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 195, + 767, + 238 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Yiet al.", + "bbox": [ + 232, + 256, + 264, + 265 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b913d8ae2acf5bef63e607a28263e9a3930cc7b56bb9a5638e1bed34af38aad6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 244, + 331, + 282 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3041555f2617046202c212c116a35c3d9921038aec4e9a7190258ae68e4a19ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 244, + 393, + 282 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/775ba24bfeb822776b193b1cf2a334d4432efacf4691bf0193a4645c6e4b4262.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 244, + 457, + 282 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7cd7426dcde5d7b3367cc327c76c4c1433254fcd2e44df40cf15a9c0724b56c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 460, + 244, + 521, + 282 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/626931ceb320f989b572f9f0c449de62a4dde28b9dd47b4f17b0e338e752ee4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 244, + 576, + 286 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f6dd90eca91aeac57f3967ed25d55247f4ac12177b4fcc3f354b0bd545ac6a98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 584, + 244, + 640, + 286 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a24d78836285131efce261c026a02a37893f03bde2cb45a0421ba54c25f8000d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 244, + 702, + 286 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/00aee554799ab1f0c4284411e3c42f6bf4d6ce013034bfb9b4200f17c4658a16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 244, + 767, + 287 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ours", + "bbox": [ + 232, + 305, + 256, + 311 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3f88d5f2729a6b63b1baadda0dbb0970e0827e3dc9152203b2615a9c4f08ab5d.jpg", + "image_caption": [ + "Rendering" + ], + "image_footnote": [], + "bbox": [ + 269, + 292, + 331, + 330 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3e504af9128ea8907af5d2fc127594a14d22842eba016d93e0869344f00c80c2.jpg", + "image_caption": [ + "Albedo" + ], + "image_footnote": [], + "bbox": [ + 334, + 292, + 392, + 330 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/eac82589400dbabcc697961e438f8136b99c7cca4b39e4e79eec39d7f34b0a24.jpg", + "image_caption": [ + "roughness" + ], + "image_footnote": [], + "bbox": [ + 395, + 292, + 457, + 330 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/69323977353543fdf250021d5195e067e30b57900190b3a2a4761eabf1e486af.jpg", + "image_caption": [ + "Relight" + ], + "image_footnote": [], + "bbox": [ + 460, + 292, + 519, + 330 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b02d56606838c0339dbb6a89db943a41b9f0a9b1c36eb429de7347b5b1e874d4.jpg", + "image_caption": [ + "Rendering" + ], + "image_footnote": [], + "bbox": [ + 524, + 292, + 576, + 330 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/258c5e2848d08c7b91f0cb92251508f67cdaf55ed9d6f996a7fded522108607e.jpg", + "image_caption": [ + "Albedo", + "Fig. 7: Comparison with other data-driven priors." + ], + "image_footnote": [], + "bbox": [ + 584, + 291, + 640, + 334 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e3f24526805a6d9d2971bc1b5566587094528c4095050b714a737cb9812ff02e.jpg", + "image_caption": [ + "roughness" + ], + "image_footnote": [], + "bbox": [ + 650, + 292, + 702, + 334 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/92aaf15911c5b766be51c4f1f8b55011651a7548f6822f2f6a7997cefc093fd8.jpg", + "image_caption": [ + "Relight" + ], + "image_footnote": [], + "bbox": [ + 714, + 291, + 767, + 334 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "the environment map from our novel light dataset, we also make a comparison on the relighted images. Because Invrender [81] and TensoIR [24] parameterize environment maps with SGs, we first optimize an SG with 128 lobes through Stochastic Gradient Decent optimization. For NvDiffRec [43], NvDiffRecMC [20], and NeILF++ [76], we render the relighted image with Blender [6] using our extracted topology and material texture maps with SPP=4096 and 4 bounces of lights. Due to the inevitable scale ambiguity between lighting and albedo, previous methods [80] align the predicted albedo to the ground truth before relighting and evaluation. Each channel is separately aligned to the ground truth. However, this would cause the incorrectly baked shading in the albedo to be removed during alignment. To this end, we align the average RGB of prediction to ground truth for a more faithful evaluation. Because geometry affects the inverse rendering performance greatly, to make sure the topology for NvDiffRec and NvdiffRecMC is well initialized, we make another set of experiments for NvDiffRec and NvDiffRecMC with MonoSDF [75] reconstructed geometry given and as a fixed topology along the training and evaluation process.", + "bbox": [ + 212, + 405, + 787, + 648 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Metrics. We use Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) [79] as metrics to evaluate the image quality of the aligned albedo, relighting images, and novel view synthesis. Roughness is evaluated using Mean Squared Error(MSE).", + "bbox": [ + 212, + 659, + 787, + 722 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison with optimization-based methods. Tab. 1 quantitatively compares our method with recent inverse rendering methods on the synthetic dataset. We significantly outperform baseline methods in terms of recovered albedo, roughness, and relighting, demonstrating the effectiveness of the proposed material prior. Although TensoIR [24] and NvDiffRec [43] show good results on view synthesis under the original environment lighting, it does not accurately estimate the underlying materials, which will be further shown in the qualitative results.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Chen. et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Tab. 1 also shows that we further improve the performance of material recovery by guiding the diffusion model with estimated albedo and specular.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Fig. 4 and Fig. 5 present the qualitative results on estimated materials and relighting. Fig. 6 demonstrates the performance of our method on the real-world data. More results on real data are presented in the supplementary material. NvDiffrec and $\\mathrm{NeILF}++$ do not disentangle the albedo and light well and show strong light color baked in albedo, as visualized in Fig. 5 and Fig. 6. With our diffusion priors learned from the large-scale dataset, our method can disentangle materials and lighting in a reasonable way.", + "bbox": [ + 212, + 178, + 787, + 285 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Comparison with other data-driven priors. Tab. 1 presents the comparison with data-driven priors [38,71]. Although these methods are able to prevent shadows from being baked into albedo, they still struggle to separate shading from the albedo of objects under complex lighting conditions, as shown in qualitative results in Fig. 7. The incorrect priors will mislead the optimization process and cause large errors in the estimated materials. In contrast, our prior models have a better capacity to decouple shading from materials. We present more qualitative comparisons of the real-world data in the supplementary material.", + "bbox": [ + 212, + 306, + 787, + 429 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation Study. We conduct ablation studies on our diffusion priors and the guided diffusion for fine-level inverse rendering. Quantitative and qualitative results are shown in Tab. 2 and Fig. 8, respectively.", + "bbox": [ + 212, + 450, + 785, + 496 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "\"w/o spec.\" indicates training without the specular shading prior model. This would result in recovering incorrect lighting and roughness, as the diffuse shading and specular shading cannot be correctly separated by relying solely on the albedo prior. As depicted in Fig. 8, the \"hat\" 's roughness prediction is lower and the \"ship\" 's is higher in \"w/o Spec.\" than ground truth, which resulting in inaccurate relighting outcomes. In \"w/o albedo\", the inverse rendering is conducted without albedo prior. The recovered albedo has baked-in shading, as shown in Fig. 8 where the two objects baked shading into the albedo causing incorrect relighting. This indicates that both priors are crucial for accurate lighting and material decomposition.", + "bbox": [ + 212, + 498, + 787, + 648 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "\"w/o guided\" is the result without the guided diffusion for fine-level optimization. Due to the lack of multi-view consistency of diffusion prior, the optimized materials are noisy and have baked-in shading, as shown in Fig. 8. This hurts the relighting quality, as shown in qualitative results in Tab. 2. Therefore, the training process using the guided diffusion model is essential for a smooth and accurate material recovery.", + "bbox": [ + 212, + 650, + 787, + 742 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Modeling indirect illumination and metallic materials. Following the setting of InvRender [81], we extend our method to handle indirect lighting. Inspired by [66], we precompute shading of indirect lighting for each image by tracing the trained SDF and Color field. Additionally, we also include metallic texture as an optimization target. More details and results are in the supplementary materials.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "IntrinsicAnything", + "bbox": [ + 611, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/4b3294f45f372594ad44b37ee13cb337814feb8a6988aac9027b72f07c3b844c.jpg", + "table_caption": [ + "Table 2: Quantitative results of ablation studies." + ], + "table_footnote": [], + "table_body": "
MethodAlbedoAligned AlbedoView SynthesisRelightingRoughness
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓MSE↓
Ours24.67660.95620.050326.88290.95320.037830.24050.97410.032828.97210.97570.02890.0105
Ours w/o spec17.21260.93160.056725.73890.94910.044429.42450.97280.035226.47430.97290.03630.0758
Ours w/o albedo22.71110.94340.057124.34620.93810.048830.59980.97510.026127.28510.97430.02580.0118
Ours w/o guided23.35070.93030.065224.61120.92080.064129.79370.96880.033927.73900.96830.03610.0168
", + "bbox": [ + 225, + 171, + 784, + 242 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4e2f6a49b27b75e2d0ba3041614004a29e2c9704a4e87de8edc5a20b801c5fa1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 257, + 269, + 304 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/25a88a5f0eb08ead0fd7edb23330aff44a9b804605966f0339935a86d23cb549.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 257, + 325, + 303 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3cec1a1aef923e4a068988e341c3470cb2b1e6ab14dd89ffec25161c82331186.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 257, + 380, + 303 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c8a1b0f687370e4be06e9ea7ae937a40c1cd59d26b4e4510fdfc03846207cdef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 257, + 437, + 303 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b6107ce4f5f0e5e13a3f491a254e45d4ccb6a70183fad1726d1a6ef6f0dfa56e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 462, + 258, + 491, + 303 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/7ea3a3e3d0aa644b6f741ebd573cfc9565bd9952d9959e5d3c3cc6ebfc7050d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 263, + 553, + 304 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c26d46edb610b0b2954646e5c289b3dfdfca7905201a8b362e6985e719892b0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 263, + 609, + 304 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/aefe093ef6679008f19ba59badf06c6235d427673a26903784b80ff1bfb38f78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 263, + 665, + 304 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9f68d9eafaf428538eabbb25f98ffe10964132917cdff57efc5d77154745ac31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 263, + 720, + 304 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1d61b8e5fb060787664cacd372d3975d3efc06b25271e67f0a638688eff0e491.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 732, + 263, + 776, + 304 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/73ada74eb334f49b112510d04f61116daa78470a8d6ac6ae310c9b1d2ffeab9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 311, + 264, + 349 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f48c1cc2dfc8a6843acd3dc9eccff5a1f6882a97256fa02b82d3718eca54dc24.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 271, + 309, + 295, + 352 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/06bf811784271d37b1a34275a8a562e2d83cb2142b2c2b417e27d0213d6e2948.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 325, + 309, + 354, + 352 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c7adee1f6384d4b70c4782b24c05289ac931c9a674693c04651df01b51a9a484.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 309, + 436, + 352 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3adb476411b22a08f20ad7f2ddaf450bef30e4d66a2d8fc0df31bfc4d8f570c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 462, + 310, + 491, + 352 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/fabafb0025ca893ea1384173a872c607d23d825096d4775b35af6db2355bad72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 314, + 553, + 354 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/76a35658085605f4313308b50306a9abf1ee447b483357681222cb08f40988fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 565, + 314, + 609, + 354 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/56a1e30bc10532ca715af2e67d59499996d57383d3fb95cf631c74c41c60db92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 314, + 666, + 354 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/618dec07efae921a338f8ecbbcab433f288a23eda4ddef48b85f41c254aef527.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 314, + 720, + 354 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/909d8b23c4930fd696037d3d2f4555817b254e4dccd3afcd4a24ff1edd51e659.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 732, + 314, + 776, + 354 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3fcb9ecaedfb493805aeffb5ec70eaaae0c3617fc4d009c62b278e91844a60c1.jpg", + "image_caption": [ + "Fig. 8: Qualitative results of ablation studies." + ], + "image_footnote": [], + "bbox": [ + 220, + 361, + 263, + 393 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0892ab3d09089931ac3e4428c47116672d126fdb3cf2c7a506d322463f463f48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 359, + 295, + 402 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3380a9e4b7b000d9f9eb272c161b76944277df36d39dab31892c37e62ccc62c2.jpg", + "image_caption": [ + "w/o albedo" + ], + "image_footnote": [], + "bbox": [ + 297, + 359, + 325, + 402 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b82aca426132d6298e93b0f2712841af00f84b8f2285af2e0a71d7cf389a94c0.jpg", + "image_caption": [ + "w/o guided" + ], + "image_footnote": [], + "bbox": [ + 349, + 359, + 380, + 402 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5abebf74672bc6605707ceb6b10c8d7070b202fffd58e4c37eeb6160b02fd46f.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 406, + 359, + 437, + 402 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/cc67f98a679e89001bc2a2b2c67d0d357cc7502ff8c14ec20ed75c034c14d56b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 464, + 359, + 493, + 402 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/56c701a677140b4acac71d9e22760cb60db94f0906aee7e66018d04f90c474f9.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 509, + 364, + 553, + 404 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/50a8b47b1f7b87b0d44efa0299ce009b88857a42bac072c27dea10f8224c5626.jpg", + "image_caption": [ + "w/o spec." + ], + "image_footnote": [], + "bbox": [ + 565, + 364, + 609, + 404 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1efcb4c7cf35bde05878b4689928963d98aa1c9169cc62e1e0762893ba207c47.jpg", + "image_caption": [ + "w/o albedo" + ], + "image_footnote": [], + "bbox": [ + 620, + 364, + 663, + 404 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0020161009a2a31d93a022d013ba381d607023960e176ce612154053fc346d1a.jpg", + "image_caption": [ + "w/o guided" + ], + "image_footnote": [], + "bbox": [ + 676, + 364, + 720, + 404 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/12dee739272c986f386caf2bb96213d05c96399265f096bc7d88032b13776e03.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 732, + 364, + 776, + 404 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 476, + 357, + 491 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We proposed to learn the material prior for inverse rendering under an unknown static lighting condition. To this end, the material prior is designed as conditional diffusion models of albedo and specular shading components, according to the general rendering equation. In addition, a two-stage training scheme is developed for robustly regularizing the inverse rendering process with conditional diffusion models, which uses coarse materials from the first stage training to guide diffusion models to provide multi-view consistent constraints. The proposed model significantly outperforms baseline methods on real-world and synthetic data.", + "bbox": [ + 212, + 507, + 787, + 628 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. Although we achieved impressive results on material recovery, our approach still has some limitations. First, we do not consider the inverse rendering of transparent objects. A solution is revising the geometry representation to a neural field of transparency and jointly optimizing it with object materials and environment lighting. Second, the performance of our model is limited by the accuracy of the reconstructed geometry from MonoSDF. It would be interesting to exploit diffusion models to learn the geometry prior.", + "bbox": [ + 212, + 643, + 787, + 750 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments. The authors would like to acknowledge the support from NSFC (No. 62172364), the Alibaba Innovation Research Program of Alibaba Group and the Information Technology Center and State Key Lab of CAD&CG, Zhejiang University.", + "bbox": [ + 212, + 763, + 787, + 825 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Chen. et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Poly haven: The public 3d asset library. https://polyhaven.com/", + "2. Stable diffusion image variations. https://huggingface.co/spaces/lambdalabs/stable-diffusion-image-variations", + "3. Anciukevicius, T., Xu, Z., Fisher, M., Henderson, P., Bilen, H., Mitra, N.J., Guerrero, P.: Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. In: CVPR (2023)", + "4. Barron, J.T., Malik, J.: Shape, illumination, and reflectance from shading. TPAMI 37(8), 1670-1687 (2014)", + "5. Bi, S., Lombardi, S., Saito, S., Simon, T., Wei, S.E., McPhail, K., Ramamoorthi, R., Sheikh, Y., Saragih, J.: Deep relightable appearance models for animatable faces. ACM Trans. Graph. (Proc. SIGGRAPH) 40(4) (2021)", + "6. Blender: Blender, https://www.blender.org/", + "7. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.P.: Nerd: Neural reflectance decomposition from image collections. In: IEEE International Conference on Computer Vision (ICCV) (2021)", + "8. Burley, B., Studios, W.D.A.: Physically-based shading at disney. In: ACM Trans. Graph. vol. 2012, pp. 1-7. vol. 2012 (2012)", + "9. Chang, A.X., Funkhouser, T.A., Guibas, L.J., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: Shapenet: An information-rich 3d model repository. CoRR abs/1512.03012 (2015), http://arxiv.org/abs/1512.03012", + "0. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: ECCV (2022)", + "1. Chen, H., Gu, J., Chen, A., Tian, W., Tu, Z., Liu, L., Su, H.: Single-stage diffusion nef: A unified approach to 3d generation and reconstruction. In: ICCV (2023)", + "2. Cheng, Y.C., Lee, H.Y., Tulyakov, S., Schwing, A.G., Gui, L.Y.: Sdfusion: Multi-modal 3d shape completion, reconstruction, and generation. In: CVPR (2023)", + "3. Chung, H., Kim, J., Mccann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. In: ICLR (2023), https://openreview.net/forum?id=OnD9zGAGTOK", + "4. Chung, H., Sim, B., Ye, J.C.: Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In: CVPR. pp. 12413-12422 (2022)", + "5. Croitoru, F.A., Hondru, V., Ionescu, R.T., Shah, M.: Diffusion models in vision: A survey. PAMI (2023)", + "6. Debevec, P., Hawkins, T., Tchou, C., Duiker, H.P., Sarokin, W., Sagar, M.: Acquiring the reflectance field of a human face. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 145-156 (2000)", + "7. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objverse: A universe of annotated 3d objects. In: CVPR. pp. 13142-13153 (2023)", + "8. Deng, C., Jiang, C., Qi, C.R., Yan, X., Zhou, Y., Guibas, L., Anguelov, D., et al.: Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In: CVPR. pp. 20637-20647 (2023)", + "9. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escalano, S., Pandey, R., Dourgarian, J., Tang, D., Tkach, A., Kowdle, A., Cooper, E., Dou, M., Fanello, S., Fyffe, G., Rhemann, C., Taylor, J., Debevec, P., Izadi, S.: The relightables: Volumetric performance capture of humans with" + ], + "bbox": [ + 225, + 176, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "IntrinsicAnything", + "bbox": [ + 611, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "realistic relighting. ACM Trans. Graph. 38(6) (nov 2019). https://doi.org/10.1145/3355089.3356571, https://doi.org/10.1145/3355089.3356571", + "20. Hasselgren, J., Hofmann, N., Munkberg, J.: Shape, light & material decomposition from images using monte carlo rendering and denoising. In: NeurIPS (2022)", + "21. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)", + "22. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)", + "23. Horita, D., Yang, J., Chen, D., Koyama, Y., Aizawa, K., Sebe, N.: A structure-guided diffusion model for large-hole image completion. In: BMVC (2023)", + "24. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoir: Tensorial inverse rendering. In: CVPR (2023)", + "25. Kajiya, J.T.: The rendering equation. In: Proceedings of the 13th Annual Conference on Computer Graphics and Interactive Techniques. p. 143-150. SIGGRAPH '86, Association for Computing Machinery, New York, NY, USA (1986). https://doi.org/10.1145/15922.15902, https://doi.org/10.1145/15922.15902", + "26. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. arXiv preprint arXiv:2201.11793 (2022)", + "27. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/", + "28. Kocsis, P., Sitzmann, V., Nießner, M.: Intrinsic image diffusion for single-view material estimation. In: arxiv (2023)", + "29. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Trans. Graph. 39(6) (2020)", + "30. Levoy, M., Pulli, K., Curless, B., Rusinkiewicz, S., Koller, D., Pereira, L., Ginzton, M., Anderson, S.E., Davis, J., Ginsberg, J., Shade, J., Fulk, D.: The digital michelangelo project: 3d scanning of large statues. Proceedings of the 27th annual conference on Computer graphics and interactive techniques (2000), https://apisemantic scholar.org/CorpusID:1546062", + "31. Li, Z., Snavely, N.: Cginintrinsics: Better intrinsic image decomposition through physically-based rendering. In: ECCV (2018)", + "32. Li, Z., Xu, Z., Ramamoorthi, R., Sunkavalli, K., Chandraker, M.: Learning to reconstruct shape and spatially-varying reflectance from a single image. ACM Trans. Graph. 37(6), 1-11 (2018)", + "33. Li, Z., Yu, T.W., Sang, S., Wang, S., Song, M., Liu, Y., Yeh, Y.Y., Zhu, R., Gundavarapu, N., Shi, J., et al.: Openrooms: An open framework for photorealistic indoor scene datasets. In: CVPR. pp. 7190-7199 (2021)", + "34. Liang, Z., Zhang, Q., Feng, Y., Shan, Y., Jia, K.: Gs-ir: 3d gaussian splatting for inverse rendering. Computer Graphics Forum (2023)", + "35. Lichy, D., Wu, J., Sengupta, S., Jacobs, D.W.: Shape and material capture at home. In: CVPR (2021)", + "36. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. In: ICLR (2023)", + "37. Liu, Y., Wang, P., Lin, C., Long, X., Wang, J., Liu, L., Komura, T., Wang, W.: Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In: ACM Trans. Graph. (2023)", + "38. Liu, Y., Li, Y., You, S., Lu, F.: Unsupervised learning for intrinsic image decomposition from a single image. In: CVPR (2020)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Chen. et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Liu, Z., Li, Y., Lin, Y., Yu, X., Peng, S., Cao, Y.P., Qi, X., Huang, X., Liang, D., Ouyang, W.: Unidream: Unifying diffusion priors for relightable text-to-3d generation (2023)", + "40. Lyu, L., Tewari, A., Habermann, M., Saito, S., Zollhöfer, M., Leimkühler, T., Theobalt, C.: Diffusion posterior illumination for ambiguity-aware inverse rendering. ACM Trans. Graph. 42(6) (2023)", + "41. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022)", + "42. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "43. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: CVPR (2022)", + "44. Nam, G., Lee, J.H., Gutierrez, D., Kim, M.H.: Practical svbrdf acquisition of 3d objects with unstructured flash photography. ACM Trans. Graph. (Proc. SIGGRAPH Asia) 37(6), 267:1-12 (2018). https://doi.org/10.1145/3272127.3275017, http://dx.doi.org/10.1145/3272127.3275017", + "45. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: ICLR (2023)", + "46. Qiu, L., Chen, G., Gu, X., Zuo, Q., Xu, M., Wu, Y., Yuan, W., Dong, Z., Bo, L., Han, X.: Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In: CVPR. pp. 9914–9925 (2024)", + "47. Ramanishka, V., Das, A., Zhang, J., Saenko, K.: Learning transferable visual models from natural language. In: CVPR. pp. 4397-4406 (2018)", + "48. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. PAMI 44(3) (2022)", + "49. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)", + "50. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI. pp. 234-241. Springer (2015)", + "51. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)", + "52. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)", + "53. Sang, S., Chandraker, M.: Single-shot neural relighting and svbrdf estimation. In: ECCV (2020)", + "54. Shi, J., Dong, Y., Su, H., Yu, S.X.: Learning non-lambertian object intrinsics across shapenet categories. In: CVPR (July 2017)", + "55. Shi, Y., Wu, Y., Wu, C., Liu, X., Zhao, C., Feng, H., Liu, J., Zhang, L., Zhang, J., Zhou, B., Ding, E., Wang, J.: Gir: 3d gaussian inverse rendering for relightable scene factorization. Arxiv (2023)", + "56. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. CoRR abs/2012.03927 (2020), https://arxiv.org/abs/2012.03927", + "57. Sun, C., Cai, G., Li, Z., Yan, K., Zhang, C., Marshall, C., Huang, J.B., Zhao, S., Dong, Z.: Neural-pbir reconstruction of shape, material, and illumination (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "IntrinsicAnything", + "bbox": [ + 611, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "58. Tang, J., Wang, T., Zhang, B., Zhang, T., Yi, R., Ma, L., Chen, D.: Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In: ICCV. pp. 22819-22829 (October 2023)", + "59. Tang, S., Zhang, F., Chen, J., Wang, P., Furukawa, Y.: Mvdiffusion: Enabling holistic multi-view image generation with correspondence-aware diffusion. In: NeurIPS (2023)", + "60. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) NIPS. vol. 30. Curran Associates, Inc. (2017)", + "61. Veach, E., Guibas, L.J.: Optimally combining sampling techniques for monte carlo rendering. In: Proceedings of the 22nd annual conference on Computer graphics and interactive techniques. pp. 419-428 (1995)", + "62. Vecchio, G., Martin, R., Roullier, A., Kaiser, A., Rouffet, R., Deschaintre, V., Boubekeur, T.: Controlmat: Controlled generative approach to material capture. arXiv preprint arXiv:2309.01700 (2023)", + "63. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In: NeurIPS (2023)", + "64. Wimbauer, F., Wu, S., Rupprecht, C.: De-rendering 3d objects in the wild (2022)", + "65. Wu, H., Hu, Z., Li, L., Zhang, Y., Fan, C., Yu, X.: Nefii: Inverse rendering for reflectance decomposition with near-field indirect illumination. In: CVPR. pp. 4295-4304 (June 2023)", + "66. Wu, L., Zhu, R., Yaldiz, M.B., Zhu, Y., Cai, H., Matai, J., Porikli, F., Li, T.M., Chandraker, M., Ramamoorthi, R.: Factorized inverse path tracing for efficient and accurate material-lighting estimation. In: ICCV. pp. 3848-3858 (2023)", + "67. Xie, Y., Yuan, M., Dong, B., Li, Q.: Diffusion model for generative image denoising. arXiv preprint arXiv:2302.02398 (2023)", + "68. Yao, Y., Zhang, J., Liu, J., Qu, Y., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf: Neural incident light field for physically-based material estimation. In: ECCV (2022)", + "69. Ye, W., Chen, S., Bao, C., Bao, H., Pollefeys, M., Cui, Z., Zhang, G.: IntrinsicNeRF: Learning Intrinsic Neural Radiance Fields for Editable Novel View Synthesis. In: ICCV (2023)", + "70. Yi, R., Tan, P., Lin, S.: Leveraging multi-view image sets for unsupervised intrinsic image decomposition and highlight separation. In: AAAI. vol. 34, pp. 12685-12692 (2020)", + "71. Yi, R., Zhu, C., Xu, K.: Weakly-supervised single-view image relighting. In: CVPR. pp. 8402-8411 (June 2023)", + "72. Yu, Y., Meka, A., Elgharib, M., Seidel, H.P., Theobalt, C., Smith, W.: Self-supervised outdoor scene relighting. In: ECCV (2020)", + "73. Yu, Y., Smith, W.A.: Inverserendernet: Learning single image inverse rendering. In: CVPR. pp. 3155-3164 (2019)", + "74. Yu, Y., Malik, J.: Recovering photometric properties of architectural scenes from photographs. In: Proceedings of the 25th Annual Conference on Computer Graphics and Interactive Techniques. p. 207-217. SIGGRAPH '98, Association for Computing Machinery, New York, NY, USA (1998). https://doi.org/10.1145/280814.280874, https://doi.org/10.1145/280814.280874", + "75. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. NeurIPS (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Chen. et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "76. Zhang, J., Yao, Y., Li, S., Liu, J., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neil++: Inter-reflectable light fields for geometry and material estimation. ICCV (2023)", + "77. Zhang, K., Luan, F., Li, Z., Snavely, N.: Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In: CVPR (2022)", + "78. Zhang, K., Luan, F., Wang, Q., Bala, K., Snavely, N.: PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In: CVPR (2021)", + "79. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)", + "80. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph. 40(6) (dec 2021)", + "81. Zhang, Y., Sun, J., He, X., Fu, H., Jia, R., Zhou, X.: Modeling indirect illumination for inverse rendering. In: CVPR (2022)", + "82. Zhu, J., Huo, Y., Ye, Q., Luan, F., Li, J., Xi, D., Wang, L., Tang, R., Hua, W., Bao, H., et al.: I2-sdf: Intrinsic indoor scene reconstruction and editing via raytracing in neural sdfs. In: CVPR. pp. 12489-12498 (2023)", + "83. Zhu, J., Luan, F., Huo, Y., Lin, Z., Zhong, Z., Xi, D., Wang, R., Bao, H., Zheng, J., Tang, R.: Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In: SIGGRAPH Asia 2022 Conference Papers. ACM (2022), https://doi.org/10.1145/3550469.3555407", + "84. Zhu, Y., Zhang, K., Liang, J., Cao, J., Wen, B., Timofte, R., Van Gool, L.: Denoising diffusion models for plug-and-play image restoration. In: CVPR. pp. 1219-1229 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 494 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "IntrinsicAnything", + "bbox": [ + 611, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/IntrinsicAnything_ Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination/b4b4420f-76ed-4ddf-baa7-e217742f08de_model.json b/2024/IntrinsicAnything_ Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination/b4b4420f-76ed-4ddf-baa7-e217742f08de_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4a16ff455382567133a479ac33b7466983e38df4 --- /dev/null +++ b/2024/IntrinsicAnything_ Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination/b4b4420f-76ed-4ddf-baa7-e217742f08de_model.json @@ -0,0 +1,3626 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.25, + 0.141, + 0.756, + 0.208 + ], + "angle": 0, + "content": "Intrinsic Anything: Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination" + }, + { + "type": "text", + "bbox": [ + 0.314, + 0.235, + 0.69, + 0.266 + ], + "angle": 0, + "content": "Xi Chen\\(^{1}\\), Sida Peng\\(^{1}\\), Dongchen Yang\\(^{1}\\), Yuan Liu\\(^{2}\\), Bowen Pan\\(^{3}\\), Chengfei Lv\\(^{3}\\), and Xiaowei Zhou\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.277, + 0.671, + 0.291 + ], + "angle": 0, + "content": "1 State Key Lab of CAD&CG, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.395, + 0.291, + 0.609, + 0.306 + ], + "angle": 0, + "content": "2 The University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.345, + 0.306, + 0.658, + 0.32 + ], + "angle": 0, + "content": "\\(^{3}\\) Tao Technology Department, Alibaba Group" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.344, + 0.744, + 0.636 + ], + "angle": 0, + "content": "Abstract. This paper aims to recover object materials from posed images captured under an unknown static lighting condition. Recent methods solve this task by optimizing material parameters through differentiable physically based rendering. However, due to the coupling between object geometry, materials, and environment lighting, there is inherent ambiguity during the inverse rendering process, preventing previous methods from obtaining accurate results. To overcome this ill-posed problem, our key idea is to learn the material prior with a generative model for regularizing the optimization process. We observe that the general rendering equation can be split into diffuse and specular shading terms, and thus formulate the material prior as diffusion models of albedo and specular. Thanks to this design, our model can be trained using the existing abundant 3D object data, and naturally acts as a versatile tool to resolve the ambiguity when recovering material representations from RGB images. In addition, we develop a coarse-to-fine training strategy that leverages estimated materials to guide diffusion models to satisfy multi-view consistent constraints, leading to more stable and accurate results. Extensive experiments on real-world and synthetic datasets demonstrate that our approach achieves state-of-the-art performance on material recovery. The code will be available at https://zju3dv.github.io/IntrinsicAnything/." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.656, + 0.377, + 0.672 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.683, + 0.788, + 0.82 + ], + "angle": 0, + "content": "Recovering the object's geometry, material, and lighting from captured images, also known as inverse rendering, is a long-standing task in computer vision and graphics. These physical attributes of 3D objects are essential for many applications, such as VR/AR, movie production, and video games. Due to the inherent complexity of interaction between real-world objects and environment lighting, inverse rendering remains an ill-posed problem. Previous works overcome this problem using sophisticated capture systems [16,19], or co-located flashlight and camera in a dark environment [5,44,77]. However, these methods require special hardware equipment or constrained environment, limiting their applications." + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.826, + 0.393, + 0.84 + ], + "angle": 0, + "content": "† Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "Chen. et al." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.142, + 0.788, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.288, + 0.789, + 0.373 + ], + "angle": 0, + "content": "Fig. 1: Two types of ambiguities in inverse rendering. (a) Ambiguity between diffuse shading and albedo. For example, the Xbox is lit by a yellow light, and the decomposed albedo from NVdiffrecMC [20] tends to be yellow. (b) Ambiguity between shadow and albedo. For example, the porcelain toy is with self-occlusion, and TensoIR [24] bakes the shadow into the recovered albedo. Our method well handles the two types of ambiguities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.789, + 0.576 + ], + "angle": 0, + "content": "Based on differentiable physically based rendering and neural scene representations, recent methods [20, 24, 76, 78, 80, 81] are able to object materials from images captured under natural lighting. They exploit neural networks to represent object materials and geometry, and combine these object attributes with the learnable lighting to synthesize images, which are compared with captured images to optimize the model paramaters. To achieve better performance of inverse rendering, these methods have attempted to improve object representations [24, 78], lighting representations [20, 76], and training strategies [80, 81]. However, we argue that depending solely on the rendering loss is insufficient for accomplishing accurate decomposition, due to the inherent ambiguity in the inverse rendering process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.789, + 0.687 + ], + "angle": 0, + "content": "We observed that there are two categories of ambiguities, as shown in Fig. 1. First, there is ambiguity between diffuse shading and albedo. Intuitively, when a white object is lit by a yellow light, the decomposed albedo appears to be yellow. Second, when an object is with self-occlusion and illuminated under a complex lighting condition, the hard-cast shadow can be easily baked into the albedo. Due to these ambiguities, the optimization process of previous methods tend to gets trapped in local optimal, leading to the artifacts on the recovered material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In this paper, we seek to use data-driven priors for ambiguity-free inverse rendering. Our key idea is learning the distribution of albedo and specular shading with generative models, which is then used to regularize the optimization process. Inspired by the recent progress of generative models [49], we exploit the conditional diffusion model for modeling the material prior. Our choice of material prior for albedo and specular shading is motivated by the fact that the general rendering equation can be separated into diffuse and specular terms, according to the Disney BRDF model, which thus has two advantages. First, we can construct training data on albedo and specular using a large amount of 3D objects [17] that contain various BRDF models. Second, our designed prior can" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.612, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Intrinsic Anything" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "work as a versatile tool to resolve the ambiguity in the inverse rendering process of many material representations in theory." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.283 + ], + "angle": 0, + "content": "A challenge of applying diffusion models to regularize the optimization process is that such generative model struggles to perform multi-view consistent constraints. To address this problem, we propose a coarse-to-fine optimization scheme. Our approach first leverages the diffusion prior to recover the coarse object material and environment lighting. Then, the coarse albedo and specular shading are used to guide the diffusion model to produce constraints with better multi-view consistency, leading to more accurate material and lighting recovery." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.787, + 0.359 + ], + "angle": 0, + "content": "For training our model, we create a dataset consisting of RGB, albedo, and specular images based on the Objverse dataset [17]. To validate the effective of the proposed approach, we conduct extensive experiments on synthetic and real-world data. Our method achieves state-of-the-art performance on material recovery and relighting." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.378, + 0.388, + 0.394 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.788, + 0.801 + ], + "angle": 0, + "content": "Learning priors for inverse rendering. Earlier approaches for inverse rendering usually require controlled set-ups [16,30] and strong domain-specific assumptions to reconstruct scene properties from images [74]. With the advance of deep learning, the intrinsic properties of images can be directly learned from large-scale synthetic datasets [32, 33, 54]. Shi et al. [54] leverages ShapeNet [9] and introduces a dataset with image intrinsics to decompose the image into base color and shading. Li et al. [31] introduces a large-scale scene dataset and lifts the model's capability from object level to scene level. Subsequent works [32, 35, 53] have attempted to further recover the BRDF and geometry of a scene. However, obtaining massive high-quality real-world images with ground truth intrinsic decomposition is a significant challenge, which limits models' generalization to real-world data. To mitigate this issue, unsupervised methods [38, 64, 72, 72, 73] and weakly supervised methods [71] seek specially designed training scheme to reduce dependency on labeled data. Recently, [83] proposed a large-scale photorealistic indoor scene dataset InteriorVerse for single-view inverse rendering. Kocsis et al. [28] further uses InteriorVerse [83] to train a diffusion model for single-view material recovery in indoor scenes, boosting the generalization to real-world images. Recent methods [39, 46, 62] also explore diffusion models for texture-related tasks. unidream [39] and Richdreamer [46] leverage diffusion model on albedo generation for reconfigurable 3D mesh generation. ControlMat [62] focuses on capturing tileable SVBRDF from a single image with relatively simple shapes (e.g. a plane) and lighting. Although the above learning-based methods achieve promising progress in inverse rendering, their single-view inference pipeline can't guarantee multi-view consistency. In contrast, our method combines the monocular diffusion models with the multiview physically based rendering and thus achieves accurate material and illumination estimation with multiview consistency." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Inverse rendering with neural representation. Inspired by the implicit representations of NeRF [42], various works [7,24,37,78,80] extend the neural representation" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Chen. et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": "tions to factorize a scene into material, geometry, and lighting. NeRV [56] leverages a neural visibility field to reduce the computational cost and models indirect lighting with known environment lighting conditions. InvRender [81] and NeILF series [68,76] further model indirect illumination with a neural representation at unknown light conditions. NvdiffrecMC [20] and NeFII [65] incorporate Monte Carlo ray tracing to recover high-frequency details of lighting. [57] adopts a coarse-to-fine manner to retrieve the materials and illumination of a scene. More recently, GS-IR [34] and GIR [55] leverage Gaussian Splatting [27] to model indirect lighting for inverse Rendering. Furthermore, [66, 69, 82] attempt to scale up the neural representations for solving inverse rendering from object to indoor scene. However, due to the inherited ill-posed feature of the inverse rendering problem, these methods still suffer from disentangling the materials and illumination properties from images. In contrast, by learning a strong material and shading prior, we effectively tackled the ill-posed feature and received a significant improvement in decomposing the image into BRDFs and illuminations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.789, + 0.555 + ], + "angle": 0, + "content": "Diffusion models. Recently, diffusion models have demonstrated unprecedented performance on generation tasks, such as 2D generation [15, 21, 49, 52] and 3D generation [3, 11, 12, 36, 45, 59], indicating the impressive capability of learning the distribution of target data. Inspired this, some recent methods [14, 26] have attempted to exploit the diffusion prior for ill-posed inverse problems, including image denoising [67, 84], image completion [23, 51], and single-view reconstruction [18, 58]. Diffusion posterior illumination [40] learns the distribution of environment illumination with a diffusion model for material recovery. In contrast, our method learns the material prior for supervision and design it as conditional diffusion models of albedo and specular shading, enabling the direct supervision on material estimation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.577, + 0.37, + 0.594 + ], + "angle": 0, + "content": "3 Background" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.609, + 0.494, + 0.625 + ], + "angle": 0, + "content": "3.1 Physically-Based Rendering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.634, + 0.788, + 0.676 + ], + "angle": 0, + "content": "We adopt the physically based rendering [25] to recover the object material and lighting from images. For a 3D point \\(\\mathbf{x}\\), its outgoing light \\(L_{o}\\) at direction \\(\\omega_{o}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.677, + 0.787, + 0.707 + ], + "angle": 0, + "content": "\\[\nL _ {o} \\left(\\omega_ {o}; \\mathbf {x}\\right) = \\int_ {\\Omega} L _ {i} \\left(\\omega_ {i}\\right) f _ {r} \\left(\\mathbf {x}; \\omega_ {o}, \\omega_ {i}\\right) \\left(\\omega_ {i} \\cdot \\mathbf {n}\\right) d \\omega_ {i}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.788, + 0.759 + ], + "angle": 0, + "content": "where \\( L_{i}(\\omega_{i}) \\) is the incident lighting along the direction \\( \\omega_{i} \\), \\( \\mathbf{n} \\) is surface normal at \\( \\mathbf{x} \\). In practice, we define the lighting \\( L_{i} \\) as an estimable longitude-latitude environment map. We use the simplified Disney BRDF model [8] for \\( f_{r} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.769, + 0.787, + 0.801 + ], + "angle": 0, + "content": "\\[\nf _ {r} (\\mathbf {x}; \\omega_ {o}, \\omega_ {i}) = \\frac {\\mathbf {k} _ {d}}{\\pi} + \\frac {D \\cdot G \\cdot F}{4 (\\mathbf {n} \\cdot \\omega_ {\\mathbf {o}}) (\\mathbf {n} \\cdot \\omega_ {i})}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "where \\(D, G\\), and \\(F\\) are the normal distribution, geometry attenuation, and Fresnel effect, respectively. The albedo \\(\\mathbf{k}_d\\) and roughness \\(\\rho\\) parameters are set as" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.612, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Intrinsic Anything" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "learnable, and other parameters are set as constant. The detailed formulation can be found in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.224 + ], + "angle": 0, + "content": "The RGB image is rendered by evaluating Eq. (1) with the environment lighting, albedo image, roughness image, and scene mesh. We solve the integral in Eq. (1) using Monte-Carlo sampling with Multiple Importance Sampling [61]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.244, + 0.673, + 0.26 + ], + "angle": 0, + "content": "3.2 Diffusion Model and Diffusion Posterior Sampling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.784, + 0.314 + ], + "angle": 0, + "content": "The diffusion model is a probabilistic generative model that generates images from pure Gaussian noise through a progressive de-noising process [21]. Each step of the denoising process from \\(\\mathbf{x}_t\\) to \\(\\mathbf{x}_{t - 1}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.324, + 0.785, + 0.358 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha (t)}} \\left(\\mathbf {x} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\hat {\\alpha} _ {t}}} \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t)\\right) + \\sigma_ {t} \\mathbf {z}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.784, + 0.412 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}\\) is the learnable noise predictor, \\(\\alpha_{t}\\) and \\(\\sigma_{t}\\) are predefined constant values, and \\(\\mathbf{z}\\) is a random value sampled from a standard Gaussian distribution. The noise predictor is typically implemented as a UNet [50]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.784, + 0.457 + ], + "angle": 0, + "content": "The diffusion posterior sampling (DPS) [13] is a technique for solving inverse problems (e.g. image deblur) using the diffusion model. Specifically, given an observation \\(\\mathbf{y}\\) derived from \\(\\mathbf{x}_0\\) which defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.449, + 0.469, + 0.785, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} = \\mathcal {A} \\left(\\mathbf {x} _ {0}\\right) + \\mathbf {n} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.785, + 0.57 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}\\) is the measurement operator, \\(\\mathbf{n}\\) is the measurement noise. The goal is to estimate \\(\\mathbf{x}_0\\) from observation \\(\\mathbf{y}\\) using posterior distribution \\(p(\\mathbf{x}_0|\\mathbf{y})\\). To sample from \\(p(\\mathbf{x}_0|\\mathbf{y})\\) with a pre-trained distribution \\(p(\\mathbf{x}_0)\\), the DPS sampling approximates the posterior sampling functions by adding a guiding term to the denoising score \\(\\epsilon_{\\theta}(\\mathbf{x}_t,t)\\) in Eq. (3):" + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.58, + 0.785, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\epsilon} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right) = \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right) + \\zeta_ {t} \\nabla_ {\\mathbf {x} _ {t}} \\left| \\left| \\mathcal {A} \\left(\\mathbf {x} _ {t} ^ {0}\\right) - \\mathbf {y} \\right| \\right| _ {2} ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.606, + 0.785, + 0.682 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_t^0\\) is the posterior estimation of \\(\\mathbf{x}_0\\) at timestamp \\(t\\) and \\(\\zeta_t\\) is the step size. This formula assumes a known \\(\\mathcal{A}\\) and \\(\\mathbf{n}\\) follows a Gaussian distribution of \\(\\mathcal{N}(\\mathbf{0},\\sigma^2\\mathbf{I})\\). We refer to DPS [13] for more details. In Section 4, we will demonstrate how to model the material priors using the diffusion model and use DPS sampling to generate multi-view consistent material samples." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.704, + 0.331, + 0.72 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Our method aims at recovering material and lighting parameters from multi-view images under a single unknown illumination. We define object material as albedo and roughness without considering the metallic attribute and use a lat-long environment map for lighting. Due to the ill-posed nature of inverse rendering [20, 81], ambiguities exist in the optimization process, leading to inaccurate material estimation. We propose using data-driven priors to constrain the optimization to eliminate these ambiguities." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Chen. et al." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.158, + 0.561, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.158, + 0.78, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.335, + 0.788, + 0.447 + ], + "angle": 0, + "content": "Fig. 2: Overview of our pipeline. (a) Based on physically based rendering, our model combines geometry, lighting, roughness, and albedo into RGB and specular images, and optimizes the lighting and materials in a two-stage manner. In the first stage, our model is supervised by images and diffusion priors to output coarse albedo and roughness. Subsequently, the coarse materials are used to guide diffusion models to provide more multi-view consistent constraints. (b) The guided sampling first calculates the L2 loss between the guidance and one-step denoised signals, and then adds the gradient of the L1 loss to the output of the noise predictor." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.789, + 0.618 + ], + "angle": 0, + "content": "Inspired by the recent success of generative models [49] on image generation, we design two material priors modeled by the conditional diffusion model (Sec. 4.1). To recover object materials and lighting, we first employ monosdf [75] to reconstruct a 3D mesh from input images. The materials are defined as learnable texture maps in UV space from the UV-mapped mesh using Blender [6], and lighting as estimizable longitude-latitude environment map. Then we apply our coarse-to-fine optimization process (Sec. 4.2) to recover accurate materials and lighting, while the geometry and the prior models are fixed. An overview of our optimization process is shown in Fig 2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.658, + 0.551, + 0.674 + ], + "angle": 0, + "content": "4.1 Albedo and Specular Prior Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Design of the material priors. To better decompose materials and lighting, we propose to use data-driven priors to constrain the optimization process. Ideally, we can directly apply prior distributions of albedo and roughness to narrow down the solution space and effectively alleviate ambiguities. However, roughness, being a concealed characteristic of a material, is difficult to observe and is also hard to acquire on most large-scale synthetic datasets [17]. Because these objects have a variety of material parameters that do not necessarily have a valid roughness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Fortunately, image rendering using the Disney BRDF model [8] is composed of specular shading and diffuse shading regardless of the choice of material parameters, as shown in Eq.(2). Thus, the rendering equation can be rewritten to:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.612, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Intrinsic Anything" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.174, + 0.785, + 0.241 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {o} (\\omega_ {o}; \\mathbf {x}) = \\frac {\\mathbf {k} _ {d}}{\\pi} \\int_ {\\Omega} L _ {i} (\\omega_ {i}) (\\omega_ {i} \\cdot \\mathbf {n}) d \\omega_ {i} + \\int_ {\\Omega} L _ {i} (\\omega_ {i}) \\frac {D \\cdot G \\cdot F}{4 (\\mathbf {n} \\cdot \\omega_ {\\mathbf {o}}) (\\mathbf {n} \\cdot \\omega_ {i})} (\\omega_ {i} \\cdot \\mathbf {n}) d \\omega_ {i} \\\\ = S _ {\\text {d i f f}} (\\mathbf {x}, L _ {i}, k _ {d}) + S _ {\\text {s p e c}} \\left(\\omega_ {o}; \\mathbf {x}, \\rho , L _ {i}\\right). \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.255, + 0.784, + 0.331 + ], + "angle": 0, + "content": "Motivated by Eq. (6), we propose to apply priors on specular shading which exists in all physically based renderings, instead of directly using priors on roughness. Specular shading is closely linked to roughness and has the advantage of being more observable. Moreover, based on any synthetic object, we can create the ground-truth specular shading for training the prior model." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.352, + 0.784, + 0.443 + ], + "angle": 0, + "content": "Modeling the material priors. Previous methods [31, 35, 54, 70, 72, 73] generally model the prior of object material using a deterministic prediction network, which attempts to regress the material from a single image. In contrast, we leverage generative models to capture the distribution of albedo and specular, considering that there could be multiple solutions for the underlying object material given captured images [4]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.445, + 0.785, + 0.566 + ], + "angle": 0, + "content": "Given an RGB image \\(I\\), we can model the conditional distribution of target albedo \\(p_{\\theta}(\\mathbf{k}_d|I)\\) and specular shading \\(p_{\\theta}(S_{\\mathrm{spec}}|I)\\) by modifying the noise predictor to accept \\(I\\) as its conditioning. Denote the modified noise predictor as \\(\\epsilon_{\\theta}(\\mathbf{x}_t,t;I)\\). Specifically, we first use the CLIP image encoder [47] to extract a feature vector from the conditioning image. Then, to inject the conditioning signal into the UNet, we utilize a transformer network [60] to perform the cross-attention between the image feature vector and the intermediate feature maps of the UNet. More details can be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.567, + 0.785, + 0.658 + ], + "angle": 0, + "content": "The diffusion model can model complex distributions in a step-wise denoising manner, which is crucial for albedo and specular shading recovery due to the complexity of lighting and material in real-world data. Moreover, as a generative model, the diffusion model facilitates a guided sampling procedure [13, 22, 41], which permits us to steer the diffusion model toward producing samples that align with multi-view image observations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.687, + 0.569, + 0.703 + ], + "angle": 0, + "content": "4.2 Material Prior for Inverse Rendering" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.72, + 0.784, + 0.826 + ], + "angle": 0, + "content": "Based on the albedo and specular shading prior models, we design a two-stage optimization process for inverse rendering, which effectively alleviates the ambiguities. An overview of our method is presented in Fig. 2. We first use the conditional diffusion model to regularize the inverse rendering process, which outputs an initialization of lighting, albedo, and roughness. Then, the coarse material is used to guide the sampling of the diffusion model, enabling it to produce more deterministic results and better stabilize the optimization process." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "Chen. et al." + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.284, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.195, + 0.266, + 0.202 + ], + "angle": 0, + "content": "Input View" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.148, + 0.338, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.195, + 0.327, + 0.201 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.151, + 0.399, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.152, + 0.454, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.153, + 0.508, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.43, + 0.195, + 0.475, + 0.202 + ], + "angle": 0, + "content": "w/o guidance" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.151, + 0.56, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.151, + 0.614, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.152, + 0.67, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.152, + 0.724, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.194, + 0.691, + 0.202 + ], + "angle": 0, + "content": "w/guidance" + }, + { + "type": "image", + "bbox": [ + 0.732, + 0.151, + 0.777, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.213, + 0.787, + 0.269 + ], + "angle": 0, + "content": "Fig. 3: Effect of the guided sampling. We visualize the samples generated from the albedo prior model of different viewpoints. Without the guided sampling, the materials are inconsistent across multi-views and do not align with the material decomposition from the observed lighting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.785, + 0.404 + ], + "angle": 0, + "content": "Training with diffusion model. Given input multi-view images of the target object, we utilize the conditional diffusion model to predict their albedo and specular shading components, which are used to supervise the model parameters. Due to the probabilistic nature of the diffusion model, the generated albedo and specular images are inconsistent across camera views in terms of intensity. To overcome this problem, we apply shift and scale invariant loss [48] to compute the difference between the predicted and generated material." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.404, + 0.651, + 0.419 + ], + "angle": 0, + "content": "In the first stage, the training loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.43, + 0.785, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o a r s e}} = \\lambda_ {1} \\| \\hat {I} - I \\| + \\lambda_ {2} \\| \\operatorname {S S I} \\left(\\hat {\\mathbf {k}} _ {d}, \\mathbf {k} _ {d} ^ {\\mathrm {s}}\\right) - \\mathbf {k} _ {d} ^ {\\mathrm {s}} \\| \\tag {7} \\\\ + \\lambda_ {3} | | \\hat {S} _ {\\text {s p e c}} - S _ {\\text {s p e c}} ^ {\\mathrm {s}} | | + \\lambda_ {4} \\mathcal {L} _ {\\text {s m o o t h}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.788, + 0.56 + ], + "angle": 0, + "content": "where \\(\\mathrm{SSI}(a,b)\\) aligns \\(a\\) to the scale and mean of \\(b\\). \\(\\hat{I}\\), \\(\\hat{k}_d\\), \\(\\hat{S}_{\\mathrm{spec}}\\) are the image, albedo, and specular shading to be optimized respectively, \\(\\mathbf{k}_d^{\\mathrm{s}}\\) and \\(S_{\\mathrm{spec}}^{\\mathrm{s}}\\) are samples from our diffusion models \\(p_\\theta (\\mathbf{k}_d|I)\\) and \\(p_{\\theta}(S_{\\mathrm{spec}}|I)\\) respectively. We also use a smoothness constraint \\(\\mathcal{L}_{\\mathrm{smooth}}\\) on the optimized materials. \\(\\lambda_1\\), \\(\\lambda_2\\), \\(\\lambda_3\\), \\(\\lambda_4\\) are hyper-parameters. More details can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.573, + 0.788, + 0.74 + ], + "angle": 0, + "content": "Training with guided diffusion samples. After the training in the first stage, we obtain an initialization of the object material and the environment lighting. To better guide the optimization process, we aimed to generate samples from the diffusion prior models that are consistent with multiple viewpoints, aligned with observed lighting conditions, and still follow the distribution of priors. To this end, we leverage the Diffusion Posterior Sampling (DPS) [13] to guide the sampling process, and use the estimated albedo \\(\\hat{\\mathbf{k}}_d\\) and specular shading \\(\\hat{S}_{\\mathrm{spec}}\\) as guiding signals to re-generate samples. As shown in Fig. 3, samples after guidance are multi-view consistent and fit observations under the current lighting condition. Specifically, we use the score function of DPS in Eq. (5) with \\(\\mathcal{A}\\) as the identity operator to re-generate albedo \\(\\tilde{\\mathbf{k}}_d\\) and specular \\(\\tilde{S}_{\\mathrm{spec}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.75, + 0.785, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\epsilon} _ {\\theta} \\left(\\tilde {\\mathbf {x}} _ {t}, t; I\\right) = \\epsilon_ {\\theta} \\left(\\tilde {\\mathbf {x}} _ {t}, t; I\\right) + \\gamma \\bigtriangledown_ {\\tilde {\\mathbf {x}} _ {t}} \\left| \\left| \\tilde {\\mathbf {x}} _ {t} ^ {0} - \\hat {\\mathbf {x}} \\right| \\right| _ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.777, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{x}}_t\\) and \\(\\tilde{\\mathbf{x}}_t^0\\) is the noisy latents and the posterior estimation of \\(\\hat{\\mathbf{k}}_d\\) or \\(\\tilde{S}_{\\mathrm{spec}}\\) on step \\(t\\), respectively. \\(\\hat{\\mathbf{x}}\\) is \\(\\hat{\\mathbf{k}}_d\\) or \\(\\hat{S}_{\\mathrm{spec}}\\). We choose the step size following [13] as \\(\\zeta_i = \\gamma / ||\\tilde{\\mathbf{x}}_t^0 - \\hat{\\mathbf{x}}||\\), where \\(\\gamma\\) is a pre-defined scalar. Finally, we use the re-generated multiview-consistent material \\(\\tilde{\\mathbf{k}}_d\\) or \\(\\tilde{S}_{\\mathrm{spec}}\\) from the DPS to supervise the lighting" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.612, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Intrinsic Anything" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "and materials using Eq. (7). Experiments show that the second stage gives an accurate decomposition of the object material and environment lighting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.284 + ], + "angle": 0, + "content": "An alternative way to supervise the inverse rendering with diffusion models is the Score Distillation Sampling (SDS) [45], which achieves consistency by finding a mode in the distribution through optimization of the material field. However, our experiments find that the SDS guidance usually produces blurry results. A possible reason is that the SDS guidance relies on noise prediction that has a high variance, which leads to unstable convergence as discussed in [63]. In constraint, our pipeline uses guided image samples that are consistent with low variance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.305, + 0.376, + 0.323 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.336, + 0.455, + 0.351 + ], + "angle": 0, + "content": "5.1 Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.788, + 0.496 + ], + "angle": 0, + "content": "Prior Model Training. We create albedo and specular training data based on Objaverse dataset [17]. We randomly select 350k objects, excluding those with pure black or white albedos, and gather 180 HDR environment maps from [1] for rendering. Each object is rendered under 10 random poses with one random HDR environment map. We use a diffusion model fine-tuned from stable-diffusion [49] for the image variation task [2] as a starting point and continue to train 100k iterations for the albedo prior and 350K iterations for the specular shading prior. The training process for the two prior models takes around 10 days on 8 NVIDIA A100 (80GB) GPU." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.603 + ], + "angle": 0, + "content": "Inverse Rendering. The resolutions for texture maps and environment are \\(2048 \\times 2048\\) and \\(512 \\times 256\\), respectively. We use Nvdiffrast [29] for rendering images with 144 samples per pixel. We train the first stage of our method for 20k iterations and the second stage for 15k iterations. The setting of hyperparameters is presented in the supplement materials. The inverse rendering for each scene takes around 1.5 hours on a single NVIDIA RTX A6000 GPU." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.621, + 0.345, + 0.636 + ], + "angle": 0, + "content": "5.2 Evulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.787, + 0.707 + ], + "angle": 0, + "content": "Dataset. We collect 4 CAD models with a variety of object shapes and materials. Each object is rendered with 120 posed RGBA images using Blender Cycles [6]. We use 100 images for training and 20 for testing. For evaluation, we also render 20 novel test views under novel lights along with their ground truth BRDFs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Baselines. We compare our methods with the following two types of baseline methods. (1) Optimization-based methods [24, 43, 76, 81, 81]. InvRender [81] and TensoIR [24] both models indirect illumination with MLPs and uses Spherical Gaussian(SG) as environment light. InvRender [81] uses MLP representations while TensoIR is inspired by tensorial representation [10]. NeILF++ [76] models an incident light field that includes both environemnt light and indirect light with an MLP constrained by the outgoing radiance field. NvDiffRec [43], and NvDiffrecMC [20] extracted explicit topology from neural SDF and directly optimized" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Chen. et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 1: Quantitative results on the synthetic dataset. \"NvDiffrec*\" and \"NvDiffrecMC*\" mean these methods are with our reconstructed mesh. \"Ours w/o guided\" means our method without the guided sampling strategy. TensorIR and NvDiffrec have strong capabilities of fitting observed images, but they struggle to recover correct object materials. Replacing our diffusion priors with other priors [38, 71] leads to the degradation of the performance on material recovery." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.241, + 0.785, + 0.376 + ], + "angle": 0, + "content": "
MethodAlbedoAligned AlbedoView SynthesisRelightingRoughness
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓MSE↓
InvRender [81]17.81810.92470.064523.75920.94910.055927.30510.96730.038624.36320.96210.05150.0420
TensoIR [24]20.29720.95270.063124.93680.95640.061433.12340.98030.026421.34730.95430.05140.0456
NvDiffrec [43]16.54670.92400.070621.53040.95070.057432.57780.98240.019021.99390.95500.07340.0777
NvDiffrec* [43]14.52040.90210.083520.22800.92040.065931.03360.97840.021821.24520.93560.06380.0841
NvDiffrecMC [20]13.04950.88150.106821.67700.94060.069930.10830.96740.044022.53370.94860.07970.1046
NvDiffrecMC* [20]11.13530.87220.117619.83250.93050.080528.16340.95600.044021.03800.93320.08220.1411
NEILF++* [76]13.83360.91720.091518.45430.93660.083129.35270.97750.027620.48780.96290.05900.0855
Yi et al. [71]13.20290.89340.083318.84280.91330.073126.50910.96690.039321.1560.95620.05070.1524
USI3D [38]11.67130.88720.105118.92960.93440.077226.96530.96660.048020.27210.95870.06120.1241
Ours24.67660.95620.050326.88290.95320.037830.24050.97410.032828.97210.97570.02890.0105
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.39, + 0.501, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.39, + 0.785, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.558, + 0.785, + 0.586 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparison in terms of relighting on the synthetic dataset. Zoom in for details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.782 + ], + "angle": 0, + "content": "the 3D mesh with its PBR materials. Their illumination is expressed as an HDR environment map, which can be optimized by NviDiffRec's [43] split sum environment lighting and NviDiffRec's Monte Carlo ray tracing. (2) Data-driven methods [38, 71]. We keep object and lighting models as the same and just replace diffusion priors with single-view intrinsic image decomposition methods for comparison. USI3D [38] models the decomposition task as domain transfer from the image domain to the albedo or shading domain using generative models. We adopt only the predicted albedo to supervise the optimization because their rendering model is not physically based. Yi. el [71] propose a network to predict and remove specular shading and subsequently recover the albedo. We use the predicted specular shading and albedo as the priors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Evaluation protocol. We analyze our method and the baselines in the following way. On the 4 synthetic datasets, we compare their reconstruction of albedo, roughness, and novel view synthesis quantitatively with the ground truths. Given" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.612, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "IntrinsicAnything" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.148, + 0.78, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.385, + 0.788, + 0.456 + ], + "angle": 0, + "content": "Fig. 5: Quantitative comparison with baselines on the synthetic dataset. Our method estimates more accurate albedo and roughness than other methods. Nvdiffrec [43] and NvdiffrecMC [20] tend to bake the lighting into the albedo, due to the ambiguity between the diffuse shading and albedo. InvRender [81] and TensoIR [24] struggle to decouple the roughness from the albedo and environment lighting." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.467, + 0.789, + 0.744 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.754, + 0.788, + 0.798 + ], + "angle": 0, + "content": "Fig.6: Qualitative results on DTU. Our method recovers reasonable materials and relighting on real datasets. Baseline models can not disentangle illumination and materials which results in the baked-in effect on materials estimation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "Chen. et al." + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.159, + 0.256, + 0.17 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.149, + 0.332, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.15, + 0.393, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.15, + 0.458, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.15, + 0.521, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.149, + 0.576, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.148, + 0.641, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.148, + 0.703, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.148, + 0.768, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.209, + 0.264, + 0.218 + ], + "angle": 0, + "content": "USIBD" + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.2, + 0.332, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.198, + 0.394, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.198, + 0.458, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.198, + 0.522, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.196, + 0.577, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.586, + 0.196, + 0.641, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.196, + 0.703, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.196, + 0.768, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.257, + 0.265, + 0.266 + ], + "angle": 0, + "content": "Yiet al." + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.246, + 0.332, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.246, + 0.394, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.246, + 0.458, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.246, + 0.522, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.245, + 0.577, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.586, + 0.245, + 0.641, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.245, + 0.703, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.245, + 0.768, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.306, + 0.257, + 0.313 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.294, + 0.332, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.27, + 0.338, + 0.319, + 0.347 + ], + "angle": 0, + "content": "Rendering" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.294, + 0.393, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.338, + 0.38, + 0.347 + ], + "angle": 0, + "content": "Albedo" + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.294, + 0.458, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.338, + 0.448, + 0.347 + ], + "angle": 0, + "content": "roughness" + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.294, + 0.52, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.472, + 0.338, + 0.505, + 0.347 + ], + "angle": 0, + "content": "Relight" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.293, + 0.577, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.338, + 0.574, + 0.347 + ], + "angle": 0, + "content": "Rendering" + }, + { + "type": "image", + "bbox": [ + 0.586, + 0.292, + 0.641, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.338, + 0.631, + 0.347 + ], + "angle": 0, + "content": "Albedo" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.293, + 0.704, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.65, + 0.338, + 0.7, + 0.347 + ], + "angle": 0, + "content": "roughness" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.292, + 0.768, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.723, + 0.338, + 0.756, + 0.347 + ], + "angle": 0, + "content": "Relight" + }, + { + "type": "image_caption", + "bbox": [ + 0.311, + 0.363, + 0.691, + 0.377 + ], + "angle": 0, + "content": "Fig. 7: Comparison with other data-driven priors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.788, + 0.649 + ], + "angle": 0, + "content": "the environment map from our novel light dataset, we also make a comparison on the relighted images. Because Invrender [81] and TensoIR [24] parameterize environment maps with SGs, we first optimize an SG with 128 lobes through Stochastic Gradient Decent optimization. For NvDiffRec [43], NvDiffRecMC [20], and NeILF++ [76], we render the relighted image with Blender [6] using our extracted topology and material texture maps with SPP=4096 and 4 bounces of lights. Due to the inevitable scale ambiguity between lighting and albedo, previous methods [80] align the predicted albedo to the ground truth before relighting and evaluation. Each channel is separately aligned to the ground truth. However, this would cause the incorrectly baked shading in the albedo to be removed during alignment. To this end, we align the average RGB of prediction to ground truth for a more faithful evaluation. Because geometry affects the inverse rendering performance greatly, to make sure the topology for NvDiffRec and NvdiffRecMC is well initialized, we make another set of experiments for NvDiffRec and NvDiffRecMC with MonoSDF [75] reconstructed geometry given and as a fixed topology along the training and evaluation process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.723 + ], + "angle": 0, + "content": "Metrics. We use Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) [79] as metrics to evaluate the image quality of the aligned albedo, relighting images, and novel view synthesis. Roughness is evaluated using Mean Squared Error(MSE)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Comparison with optimization-based methods. Tab. 1 quantitatively compares our method with recent inverse rendering methods on the synthetic dataset. We significantly outperform baseline methods in terms of recovered albedo, roughness, and relighting, demonstrating the effectiveness of the proposed material prior. Although TensoIR [24] and NvDiffRec [43] show good results on view synthesis under the original environment lighting, it does not accurately estimate the underlying materials, which will be further shown in the qualitative results." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.612, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "IntrinsicAnything" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "Tab. 1 also shows that we further improve the performance of material recovery by guiding the diffusion model with estimated albedo and specular." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.286 + ], + "angle": 0, + "content": "Fig. 4 and Fig. 5 present the qualitative results on estimated materials and relighting. Fig. 6 demonstrates the performance of our method on the real-world data. More results on real data are presented in the supplementary material. NvDiffrec and \\(\\mathrm{NeILF}++\\) do not disentangle the albedo and light well and show strong light color baked in albedo, as visualized in Fig. 5 and Fig. 6. With our diffusion priors learned from the large-scale dataset, our method can disentangle materials and lighting in a reasonable way." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.788, + 0.43 + ], + "angle": 0, + "content": "Comparison with other data-driven priors. Tab. 1 presents the comparison with data-driven priors [38,71]. Although these methods are able to prevent shadows from being baked into albedo, they still struggle to separate shading from the albedo of objects under complex lighting conditions, as shown in qualitative results in Fig. 7. The incorrect priors will mislead the optimization process and cause large errors in the estimated materials. In contrast, our prior models have a better capacity to decouple shading from materials. We present more qualitative comparisons of the real-world data in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.452, + 0.787, + 0.497 + ], + "angle": 0, + "content": "Ablation Study. We conduct ablation studies on our diffusion priors and the guided diffusion for fine-level inverse rendering. Quantitative and qualitative results are shown in Tab. 2 and Fig. 8, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.499, + 0.788, + 0.65 + ], + "angle": 0, + "content": "\"w/o spec.\" indicates training without the specular shading prior model. This would result in recovering incorrect lighting and roughness, as the diffuse shading and specular shading cannot be correctly separated by relying solely on the albedo prior. As depicted in Fig. 8, the \"hat\" 's roughness prediction is lower and the \"ship\" 's is higher in \"w/o Spec.\" than ground truth, which resulting in inaccurate relighting outcomes. In \"w/o albedo\", the inverse rendering is conducted without albedo prior. The recovered albedo has baked-in shading, as shown in Fig. 8 where the two objects baked shading into the albedo causing incorrect relighting. This indicates that both priors are crucial for accurate lighting and material decomposition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.743 + ], + "angle": 0, + "content": "\"w/o guided\" is the result without the guided diffusion for fine-level optimization. Due to the lack of multi-view consistency of diffusion prior, the optimized materials are noisy and have baked-in shading, as shown in Fig. 8. This hurts the relighting quality, as shown in qualitative results in Tab. 2. Therefore, the training process using the guided diffusion model is essential for a smooth and accurate material recovery." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Modeling indirect illumination and metallic materials. Following the setting of InvRender [81], we extend our method to handle indirect lighting. Inspired by [66], we precompute shading of indirect lighting for each image by tracing the trained SDF and Color field. Additionally, we also include metallic texture as an optimization target. More details and results are in the supplementary materials." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "Chen. et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.315, + 0.145, + 0.687, + 0.16 + ], + "angle": 0, + "content": "Table 2: Quantitative results of ablation studies." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.172, + 0.785, + 0.243 + ], + "angle": 0, + "content": "
MethodAlbedoAligned AlbedoView SynthesisRelightingRoughness
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓MSE↓
Ours24.67660.95620.050326.88290.95320.037830.24050.97410.032828.97210.97570.02890.0105
Ours w/o spec17.21260.93160.056725.73890.94910.044429.42450.97280.035226.47430.97290.03630.0758
Ours w/o albedo22.71110.94340.057124.34620.93810.048830.59980.97510.026127.28510.97430.02580.0118
Ours w/o guided23.35070.93030.065224.61120.92080.064129.79370.96880.033927.73900.96830.03610.0168
" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.258, + 0.271, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.258, + 0.326, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.258, + 0.382, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.407, + 0.258, + 0.438, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.464, + 0.259, + 0.493, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.264, + 0.554, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.264, + 0.611, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.264, + 0.666, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.264, + 0.722, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.733, + 0.264, + 0.777, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.312, + 0.266, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.31, + 0.297, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.31, + 0.355, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.407, + 0.31, + 0.437, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.464, + 0.311, + 0.492, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.315, + 0.554, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.315, + 0.611, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.315, + 0.667, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.315, + 0.722, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.733, + 0.315, + 0.777, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.362, + 0.264, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.36, + 0.297, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.36, + 0.326, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.36, + 0.382, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.41, + 0.386, + 0.418 + ], + "angle": 0, + "content": "w/o albedo" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.361, + 0.438, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.403, + 0.411, + 0.449, + 0.418 + ], + "angle": 0, + "content": "w/o guided" + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.361, + 0.494, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.411, + 0.484, + 0.418 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.365, + 0.554, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.411, + 0.535, + 0.418 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.365, + 0.61, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.411, + 0.606, + 0.418 + ], + "angle": 0, + "content": "w/o spec." + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.365, + 0.665, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.411, + 0.663, + 0.418 + ], + "angle": 0, + "content": "w/o albedo" + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.365, + 0.722, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.411, + 0.728, + 0.418 + ], + "angle": 0, + "content": "w/o guided" + }, + { + "type": "image", + "bbox": [ + 0.733, + 0.365, + 0.777, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.411, + 0.762, + 0.418 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.434, + 0.676, + 0.448 + ], + "angle": 0, + "content": "Fig. 8: Qualitative results of ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.477, + 0.359, + 0.492 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.63 + ], + "angle": 0, + "content": "We proposed to learn the material prior for inverse rendering under an unknown static lighting condition. To this end, the material prior is designed as conditional diffusion models of albedo and specular shading components, according to the general rendering equation. In addition, a two-stage training scheme is developed for robustly regularizing the inverse rendering process with conditional diffusion models, which uses coarse materials from the first stage training to guide diffusion models to provide multi-view consistent constraints. The proposed model significantly outperforms baseline methods on real-world and synthetic data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.751 + ], + "angle": 0, + "content": "Limitations. Although we achieved impressive results on material recovery, our approach still has some limitations. First, we do not consider the inverse rendering of transparent objects. A solution is revising the geometry representation to a neural field of transparency and jointly optimizing it with object materials and environment lighting. Second, the performance of our model is limited by the accuracy of the reconstructed geometry from MonoSDF. It would be interesting to exploit diffusion models to learn the geometry prior." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.827 + ], + "angle": 0, + "content": "Acknowledgments. The authors would like to acknowledge the support from NSFC (No. 62172364), the Alibaba Innovation Research Program of Alibaba Group and the Information Technology Center and State Key Lab of CAD&CG, Zhejiang University." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.612, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "IntrinsicAnything" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.684, + 0.19 + ], + "angle": 0, + "content": "1. Poly haven: The public 3d asset library. https://polyhaven.com/" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.191, + 0.786, + 0.218 + ], + "angle": 0, + "content": "2. Stable diffusion image variations. https://huggingface.co/spaces/lambdalabs/stable-diffusion-image-variations" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.786, + 0.259 + ], + "angle": 0, + "content": "3. Anciukevicius, T., Xu, Z., Fisher, M., Henderson, P., Bilen, H., Mitra, N.J., Guerrero, P.: Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.786, + 0.286 + ], + "angle": 0, + "content": "4. Barron, J.T., Malik, J.: Shape, illumination, and reflectance from shading. TPAMI 37(8), 1670-1687 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.786, + 0.329 + ], + "angle": 0, + "content": "5. Bi, S., Lombardi, S., Saito, S., Simon, T., Wei, S.E., McPhail, K., Ramamoorthi, R., Sheikh, Y., Saragih, J.: Deep relightable appearance models for animatable faces. ACM Trans. Graph. (Proc. SIGGRAPH) 40(4) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.553, + 0.342 + ], + "angle": 0, + "content": "6. Blender: Blender, https://www.blender.org/" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.786, + 0.383 + ], + "angle": 0, + "content": "7. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.P.: Nerd: Neural reflectance decomposition from image collections. In: IEEE International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.384, + 0.786, + 0.412 + ], + "angle": 0, + "content": "8. Burley, B., Studios, W.D.A.: Physically-based shading at disney. In: ACM Trans. Graph. vol. 2012, pp. 1-7. vol. 2012 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.786, + 0.467 + ], + "angle": 0, + "content": "9. Chang, A.X., Funkhouser, T.A., Guibas, L.J., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: Shapenet: An information-rich 3d model repository. CoRR abs/1512.03012 (2015), http://arxiv.org/abs/1512.03012" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.786, + 0.494 + ], + "angle": 0, + "content": "0. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.495, + 0.786, + 0.522 + ], + "angle": 0, + "content": "1. Chen, H., Gu, J., Chen, A., Tian, W., Tu, Z., Liu, L., Su, H.: Single-stage diffusion nef: A unified approach to 3d generation and reconstruction. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.786, + 0.55 + ], + "angle": 0, + "content": "2. Cheng, Y.C., Lee, H.Y., Tulyakov, S., Schwing, A.G., Gui, L.Y.: Sdfusion: Multi-modal 3d shape completion, reconstruction, and generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.786, + 0.59 + ], + "angle": 0, + "content": "3. Chung, H., Kim, J., Mccann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. In: ICLR (2023), https://openreview.net/forum?id=OnD9zGAGTOK" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.591, + 0.786, + 0.633 + ], + "angle": 0, + "content": "4. Chung, H., Sim, B., Ye, J.C.: Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In: CVPR. pp. 12413-12422 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.786, + 0.66 + ], + "angle": 0, + "content": "5. Croitoru, F.A., Hondru, V., Ionescu, R.T., Shah, M.: Diffusion models in vision: A survey. PAMI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.786, + 0.702 + ], + "angle": 0, + "content": "6. Debevec, P., Hawkins, T., Tchou, C., Duiker, H.P., Sarokin, W., Sagar, M.: Acquiring the reflectance field of a human face. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 145-156 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.786, + 0.743 + ], + "angle": 0, + "content": "7. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objverse: A universe of annotated 3d objects. In: CVPR. pp. 13142-13153 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.786, + 0.785 + ], + "angle": 0, + "content": "8. Deng, C., Jiang, C., Qi, C.R., Yan, X., Zhou, Y., Guibas, L., Anguelov, D., et al.: Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In: CVPR. pp. 20637-20647 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.786, + 0.84 + ], + "angle": 0, + "content": "9. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escalano, S., Pandey, R., Dourgarian, J., Tang, D., Tkach, A., Kowdle, A., Cooper, E., Dou, M., Fanello, S., Fyffe, G., Rhemann, C., Taylor, J., Debevec, P., Izadi, S.: The relightables: Volumetric performance capture of humans with" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Chen. et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.148, + 0.784, + 0.175 + ], + "angle": 0, + "content": "realistic relighting. ACM Trans. Graph. 38(6) (nov 2019). https://doi.org/10.1145/3355089.3356571, https://doi.org/10.1145/3355089.3356571" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "20. Hasselgren, J., Hofmann, N., Munkberg, J.: Shape, light & material decomposition from images using monte carlo rendering and denoising. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "21. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.258 + ], + "angle": 0, + "content": "22. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.286 + ], + "angle": 0, + "content": "23. Horita, D., Yang, J., Chen, D., Koyama, Y., Aizawa, K., Sebe, N.: A structure-guided diffusion model for large-hole image completion. In: BMVC (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "24. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoir: Tensorial inverse rendering. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.37 + ], + "angle": 0, + "content": "25. Kajiya, J.T.: The rendering equation. In: Proceedings of the 13th Annual Conference on Computer Graphics and Interactive Techniques. p. 143-150. SIGGRAPH '86, Association for Computing Machinery, New York, NY, USA (1986). https://doi.org/10.1145/15922.15902, https://doi.org/10.1145/15922.15902" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "26. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. arXiv preprint arXiv:2201.11793 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "27. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "28. Kocsis, P., Sitzmann, V., Nießner, M.: Intrinsic image diffusion for single-view material estimation. In: arxiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "29. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Trans. Graph. 39(6) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.785, + 0.577 + ], + "angle": 0, + "content": "30. Levoy, M., Pulli, K., Curless, B., Rusinkiewicz, S., Koller, D., Pereira, L., Ginzton, M., Anderson, S.E., Davis, J., Ginsberg, J., Shade, J., Fulk, D.: The digital michelangelo project: 3d scanning of large statues. Proceedings of the 27th annual conference on Computer graphics and interactive techniques (2000), https://apisemantic scholar.org/CorpusID:1546062" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "31. Li, Z., Snavely, N.: Cginintrinsics: Better intrinsic image decomposition through physically-based rendering. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "32. Li, Z., Xu, Z., Ramamoorthi, R., Sunkavalli, K., Chandraker, M.: Learning to reconstruct shape and spatially-varying reflectance from a single image. ACM Trans. Graph. 37(6), 1-11 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.687 + ], + "angle": 0, + "content": "33. Li, Z., Yu, T.W., Sang, S., Wang, S., Song, M., Liu, Y., Yeh, Y.Y., Zhu, R., Gundavarapu, N., Shi, J., et al.: Openrooms: An open framework for photorealistic indoor scene datasets. In: CVPR. pp. 7190-7199 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.785, + 0.716 + ], + "angle": 0, + "content": "34. Liang, Z., Zhang, Q., Feng, Y., Shan, Y., Jia, K.: Gs-ir: 3d gaussian splatting for inverse rendering. Computer Graphics Forum (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.785, + 0.743 + ], + "angle": 0, + "content": "35. Lichy, D., Wu, J., Sengupta, S., Jacobs, D.W.: Shape and material capture at home. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "36. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "37. Liu, Y., Wang, P., Lin, C., Long, X., Wang, J., Liu, L., Komura, T., Wang, W.: Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In: ACM Trans. Graph. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "38. Liu, Y., Li, Y., You, S., Lu, F.: Unsupervised learning for intrinsic image decomposition from a single image. In: CVPR (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.612, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "IntrinsicAnything" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "39. Liu, Z., Li, Y., Lin, Y., Yu, X., Peng, S., Cao, Y.P., Qi, X., Huang, X., Liang, D., Ouyang, W.: Unidream: Unifying diffusion priors for relightable text-to-3d generation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "40. Lyu, L., Tewari, A., Habermann, M., Saito, S., Zollhöfer, M., Leimkühler, T., Theobalt, C.: Diffusion posterior illumination for ambiguity-aware inverse rendering. ACM Trans. Graph. 42(6) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.787, + 0.259 + ], + "angle": 0, + "content": "41. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.787, + 0.3 + ], + "angle": 0, + "content": "42. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "43. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.787, + 0.398 + ], + "angle": 0, + "content": "44. Nam, G., Lee, J.H., Gutierrez, D., Kim, M.H.: Practical svbrdf acquisition of 3d objects with unstructured flash photography. ACM Trans. Graph. (Proc. SIGGRAPH Asia) 37(6), 267:1-12 (2018). https://doi.org/10.1145/3272127.3275017, http://dx.doi.org/10.1145/3272127.3275017" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.787, + 0.424 + ], + "angle": 0, + "content": "45. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.424, + 0.787, + 0.467 + ], + "angle": 0, + "content": "46. Qiu, L., Chen, G., Gu, X., Zuo, Q., Xu, M., Wu, Y., Yuan, W., Dong, Z., Bo, L., Han, X.: Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In: CVPR. pp. 9914–9925 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.787, + 0.494 + ], + "angle": 0, + "content": "47. Ramanishka, V., Das, A., Zhang, J., Saenko, K.: Learning transferable visual models from natural language. In: CVPR. pp. 4397-4406 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.494, + 0.787, + 0.536 + ], + "angle": 0, + "content": "48. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. PAMI 44(3) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.536, + 0.787, + 0.563 + ], + "angle": 0, + "content": "49. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.787, + 0.592 + ], + "angle": 0, + "content": "50. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.592, + 0.787, + 0.633 + ], + "angle": 0, + "content": "51. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.633, + 0.787, + 0.675 + ], + "angle": 0, + "content": "52. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.675, + 0.787, + 0.702 + ], + "angle": 0, + "content": "53. Sang, S., Chandraker, M.: Single-shot neural relighting and svbrdf estimation. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.702, + 0.787, + 0.73 + ], + "angle": 0, + "content": "54. Shi, J., Dong, Y., Su, H., Yu, S.X.: Learning non-lambertian object intrinsics across shapenet categories. In: CVPR (July 2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.73, + 0.787, + 0.771 + ], + "angle": 0, + "content": "55. Shi, Y., Wu, Y., Wu, C., Liu, X., Zhao, C., Feng, H., Liu, J., Zhang, L., Zhang, J., Zhou, B., Ding, E., Wang, J.: Gir: 3d gaussian inverse rendering for relightable scene factorization. Arxiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.771, + 0.787, + 0.813 + ], + "angle": 0, + "content": "56. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. CoRR abs/2012.03927 (2020), https://arxiv.org/abs/2012.03927" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "57. Sun, C., Cai, G., Li, Z., Yan, K., Zhang, C., Marshall, C., Huang, J.B., Zhao, S., Dong, Z.: Neural-pbir reconstruction of shape, material, and illumination (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Chen. et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "58. Tang, J., Wang, T., Zhang, B., Zhang, T., Yi, R., Ma, L., Chen, D.: Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In: ICCV. pp. 22819-22829 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "59. Tang, S., Zhang, F., Chen, J., Wang, P., Furukawa, Y.: Mvdiffusion: Enabling holistic multi-view image generation with correspondence-aware diffusion. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.286 + ], + "angle": 0, + "content": "60. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) NIPS. vol. 30. Curran Associates, Inc. (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.328 + ], + "angle": 0, + "content": "61. Veach, E., Guibas, L.J.: Optimally combining sampling techniques for monte carlo rendering. In: Proceedings of the 22nd annual conference on Computer graphics and interactive techniques. pp. 419-428 (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "62. Vecchio, G., Martin, R., Roullier, A., Kaiser, A., Rouffet, R., Deschaintre, V., Boubekeur, T.: Controlmat: Controlled generative approach to material capture. arXiv preprint arXiv:2309.01700 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "63. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.425 + ], + "angle": 0, + "content": "64. Wimbauer, F., Wu, S., Rupprecht, C.: De-rendering 3d objects in the wild (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.467 + ], + "angle": 0, + "content": "65. Wu, H., Hu, Z., Li, L., Zhang, Y., Fan, C., Yu, X.: Nefii: Inverse rendering for reflectance decomposition with near-field indirect illumination. In: CVPR. pp. 4295-4304 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.509 + ], + "angle": 0, + "content": "66. Wu, L., Zhu, R., Yaldiz, M.B., Zhu, Y., Cai, H., Matai, J., Porikli, F., Li, T.M., Chandraker, M., Ramamoorthi, R.: Factorized inverse path tracing for efficient and accurate material-lighting estimation. In: ICCV. pp. 3848-3858 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "67. Xie, Y., Yuan, M., Dong, B., Li, Q.: Diffusion model for generative image denoising. arXiv preprint arXiv:2302.02398 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "68. Yao, Y., Zhang, J., Liu, J., Qu, Y., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf: Neural incident light field for physically-based material estimation. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.785, + 0.618 + ], + "angle": 0, + "content": "69. Ye, W., Chen, S., Bao, C., Bao, H., Pollefeys, M., Cui, Z., Zhang, G.: IntrinsicNeRF: Learning Intrinsic Neural Radiance Fields for Editable Novel View Synthesis. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.66 + ], + "angle": 0, + "content": "70. Yi, R., Tan, P., Lin, S.: Leveraging multi-view image sets for unsupervised intrinsic image decomposition and highlight separation. In: AAAI. vol. 34, pp. 12685-12692 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.688 + ], + "angle": 0, + "content": "71. Yi, R., Zhu, C., Xu, K.: Weakly-supervised single-view image relighting. In: CVPR. pp. 8402-8411 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.785, + 0.716 + ], + "angle": 0, + "content": "72. Yu, Y., Meka, A., Elgharib, M., Seidel, H.P., Theobalt, C., Smith, W.: Self-supervised outdoor scene relighting. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.743 + ], + "angle": 0, + "content": "73. Yu, Y., Smith, W.A.: Inverserendernet: Learning single image inverse rendering. In: CVPR. pp. 3155-3164 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.812 + ], + "angle": 0, + "content": "74. Yu, Y., Malik, J.: Recovering photometric properties of architectural scenes from photographs. In: Proceedings of the 25th Annual Conference on Computer Graphics and Interactive Techniques. p. 207-217. SIGGRAPH '98, Association for Computing Machinery, New York, NY, USA (1998). https://doi.org/10.1145/280814.280874, https://doi.org/10.1145/280814.280874" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.785, + 0.841 + ], + "angle": 0, + "content": "75. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. NeurIPS (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.612, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "IntrinsicAnything" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "76. Zhang, J., Yao, Y., Li, S., Liu, J., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neil++: Inter-reflectable light fields for geometry and material estimation. ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.217 + ], + "angle": 0, + "content": "77. Zhang, K., Luan, F., Li, Z., Snavely, N.: Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.786, + 0.258 + ], + "angle": 0, + "content": "78. Zhang, K., Luan, F., Wang, Q., Bala, K., Snavely, N.: PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.286 + ], + "angle": 0, + "content": "79. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "80. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph. 40(6) (dec 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.355 + ], + "angle": 0, + "content": "81. Zhang, Y., Sun, J., He, X., Fu, H., Jia, R., Zhou, X.: Modeling indirect illumination for inverse rendering. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.397 + ], + "angle": 0, + "content": "82. Zhu, J., Huo, Y., Ye, Q., Luan, F., Li, J., Xi, D., Wang, L., Tang, R., Hua, W., Bao, H., et al.: I2-sdf: Intrinsic indoor scene reconstruction and editing via raytracing in neural sdfs. In: CVPR. pp. 12489-12498 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.452 + ], + "angle": 0, + "content": "83. Zhu, J., Luan, F., Huo, Y., Lin, Z., Zhong, Z., Xi, D., Wang, R., Bao, H., Zheng, J., Tang, R.: Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In: SIGGRAPH Asia 2022 Conference Papers. ACM (2022), https://doi.org/10.1145/3550469.3555407" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.495 + ], + "angle": 0, + "content": "84. Zhu, Y., Zhang, K., Liang, J., Cao, J., Wen, B., Timofte, R., Van Gool, L.: Denoising diffusion models for plug-and-play image restoration. In: CVPR. pp. 1219-1229 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.495 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/IntrinsicAnything_ Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination/b4b4420f-76ed-4ddf-baa7-e217742f08de_origin.pdf b/2024/IntrinsicAnything_ Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination/b4b4420f-76ed-4ddf-baa7-e217742f08de_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..eb26468597c8e667cdc314ea4ea071f72ed5941f --- /dev/null +++ b/2024/IntrinsicAnything_ Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination/b4b4420f-76ed-4ddf-baa7-e217742f08de_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:263786b8055a8044144a8016fd6c2ef1a527790bd10f7cbc5bdef404c8ab438f +size 8815194 diff --git a/2024/IntrinsicAnything_ Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination/full.md b/2024/IntrinsicAnything_ Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination/full.md new file mode 100644 index 0000000000000000000000000000000000000000..39f4b38f4713eaaac50fd320849fa31d461a558c --- /dev/null +++ b/2024/IntrinsicAnything_ Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination/full.md @@ -0,0 +1,475 @@ +# Intrinsic Anything: Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination + +Xi Chen $^{1}$ , Sida Peng $^{1}$ , Dongchen Yang $^{1}$ , Yuan Liu $^{2}$ , Bowen Pan $^{3}$ , Chengfei Lv $^{3}$ , and Xiaowei Zhou $^{1\dagger}$ + +1 State Key Lab of CAD&CG, Zhejiang University + +2 The University of Hong Kong + +$^{3}$ Tao Technology Department, Alibaba Group + +Abstract. This paper aims to recover object materials from posed images captured under an unknown static lighting condition. Recent methods solve this task by optimizing material parameters through differentiable physically based rendering. However, due to the coupling between object geometry, materials, and environment lighting, there is inherent ambiguity during the inverse rendering process, preventing previous methods from obtaining accurate results. To overcome this ill-posed problem, our key idea is to learn the material prior with a generative model for regularizing the optimization process. We observe that the general rendering equation can be split into diffuse and specular shading terms, and thus formulate the material prior as diffusion models of albedo and specular. Thanks to this design, our model can be trained using the existing abundant 3D object data, and naturally acts as a versatile tool to resolve the ambiguity when recovering material representations from RGB images. In addition, we develop a coarse-to-fine training strategy that leverages estimated materials to guide diffusion models to satisfy multi-view consistent constraints, leading to more stable and accurate results. Extensive experiments on real-world and synthetic datasets demonstrate that our approach achieves state-of-the-art performance on material recovery. The code will be available at https://zju3dv.github.io/IntrinsicAnything/. + +# 1 Introduction + +Recovering the object's geometry, material, and lighting from captured images, also known as inverse rendering, is a long-standing task in computer vision and graphics. These physical attributes of 3D objects are essential for many applications, such as VR/AR, movie production, and video games. Due to the inherent complexity of interaction between real-world objects and environment lighting, inverse rendering remains an ill-posed problem. Previous works overcome this problem using sophisticated capture systems [16,19], or co-located flashlight and camera in a dark environment [5,44,77]. However, these methods require special hardware equipment or constrained environment, limiting their applications. + +![](images/6d6a8324d8e7504e543a22a6d863ec9b12a76bbadfddc5b5d244daf61b216b48.jpg) +Fig. 1: Two types of ambiguities in inverse rendering. (a) Ambiguity between diffuse shading and albedo. For example, the Xbox is lit by a yellow light, and the decomposed albedo from NVdiffrecMC [20] tends to be yellow. (b) Ambiguity between shadow and albedo. For example, the porcelain toy is with self-occlusion, and TensoIR [24] bakes the shadow into the recovered albedo. Our method well handles the two types of ambiguities. + +Based on differentiable physically based rendering and neural scene representations, recent methods [20, 24, 76, 78, 80, 81] are able to object materials from images captured under natural lighting. They exploit neural networks to represent object materials and geometry, and combine these object attributes with the learnable lighting to synthesize images, which are compared with captured images to optimize the model paramaters. To achieve better performance of inverse rendering, these methods have attempted to improve object representations [24, 78], lighting representations [20, 76], and training strategies [80, 81]. However, we argue that depending solely on the rendering loss is insufficient for accomplishing accurate decomposition, due to the inherent ambiguity in the inverse rendering process. + +We observed that there are two categories of ambiguities, as shown in Fig. 1. First, there is ambiguity between diffuse shading and albedo. Intuitively, when a white object is lit by a yellow light, the decomposed albedo appears to be yellow. Second, when an object is with self-occlusion and illuminated under a complex lighting condition, the hard-cast shadow can be easily baked into the albedo. Due to these ambiguities, the optimization process of previous methods tend to gets trapped in local optimal, leading to the artifacts on the recovered material. + +In this paper, we seek to use data-driven priors for ambiguity-free inverse rendering. Our key idea is learning the distribution of albedo and specular shading with generative models, which is then used to regularize the optimization process. Inspired by the recent progress of generative models [49], we exploit the conditional diffusion model for modeling the material prior. Our choice of material prior for albedo and specular shading is motivated by the fact that the general rendering equation can be separated into diffuse and specular terms, according to the Disney BRDF model, which thus has two advantages. First, we can construct training data on albedo and specular using a large amount of 3D objects [17] that contain various BRDF models. Second, our designed prior can + +work as a versatile tool to resolve the ambiguity in the inverse rendering process of many material representations in theory. + +A challenge of applying diffusion models to regularize the optimization process is that such generative model struggles to perform multi-view consistent constraints. To address this problem, we propose a coarse-to-fine optimization scheme. Our approach first leverages the diffusion prior to recover the coarse object material and environment lighting. Then, the coarse albedo and specular shading are used to guide the diffusion model to produce constraints with better multi-view consistency, leading to more accurate material and lighting recovery. + +For training our model, we create a dataset consisting of RGB, albedo, and specular images based on the Objverse dataset [17]. To validate the effective of the proposed approach, we conduct extensive experiments on synthetic and real-world data. Our method achieves state-of-the-art performance on material recovery and relighting. + +# 2 Related Work + +Learning priors for inverse rendering. Earlier approaches for inverse rendering usually require controlled set-ups [16,30] and strong domain-specific assumptions to reconstruct scene properties from images [74]. With the advance of deep learning, the intrinsic properties of images can be directly learned from large-scale synthetic datasets [32, 33, 54]. Shi et al. [54] leverages ShapeNet [9] and introduces a dataset with image intrinsics to decompose the image into base color and shading. Li et al. [31] introduces a large-scale scene dataset and lifts the model's capability from object level to scene level. Subsequent works [32, 35, 53] have attempted to further recover the BRDF and geometry of a scene. However, obtaining massive high-quality real-world images with ground truth intrinsic decomposition is a significant challenge, which limits models' generalization to real-world data. To mitigate this issue, unsupervised methods [38, 64, 72, 72, 73] and weakly supervised methods [71] seek specially designed training scheme to reduce dependency on labeled data. Recently, [83] proposed a large-scale photorealistic indoor scene dataset InteriorVerse for single-view inverse rendering. Kocsis et al. [28] further uses InteriorVerse [83] to train a diffusion model for single-view material recovery in indoor scenes, boosting the generalization to real-world images. Recent methods [39, 46, 62] also explore diffusion models for texture-related tasks. unidream [39] and Richdreamer [46] leverage diffusion model on albedo generation for reconfigurable 3D mesh generation. ControlMat [62] focuses on capturing tileable SVBRDF from a single image with relatively simple shapes (e.g. a plane) and lighting. Although the above learning-based methods achieve promising progress in inverse rendering, their single-view inference pipeline can't guarantee multi-view consistency. In contrast, our method combines the monocular diffusion models with the multiview physically based rendering and thus achieves accurate material and illumination estimation with multiview consistency. + +Inverse rendering with neural representation. Inspired by the implicit representations of NeRF [42], various works [7,24,37,78,80] extend the neural representation + +tions to factorize a scene into material, geometry, and lighting. NeRV [56] leverages a neural visibility field to reduce the computational cost and models indirect lighting with known environment lighting conditions. InvRender [81] and NeILF series [68,76] further model indirect illumination with a neural representation at unknown light conditions. NvdiffrecMC [20] and NeFII [65] incorporate Monte Carlo ray tracing to recover high-frequency details of lighting. [57] adopts a coarse-to-fine manner to retrieve the materials and illumination of a scene. More recently, GS-IR [34] and GIR [55] leverage Gaussian Splatting [27] to model indirect lighting for inverse Rendering. Furthermore, [66, 69, 82] attempt to scale up the neural representations for solving inverse rendering from object to indoor scene. However, due to the inherited ill-posed feature of the inverse rendering problem, these methods still suffer from disentangling the materials and illumination properties from images. In contrast, by learning a strong material and shading prior, we effectively tackled the ill-posed feature and received a significant improvement in decomposing the image into BRDFs and illuminations. + +Diffusion models. Recently, diffusion models have demonstrated unprecedented performance on generation tasks, such as 2D generation [15, 21, 49, 52] and 3D generation [3, 11, 12, 36, 45, 59], indicating the impressive capability of learning the distribution of target data. Inspired this, some recent methods [14, 26] have attempted to exploit the diffusion prior for ill-posed inverse problems, including image denoising [67, 84], image completion [23, 51], and single-view reconstruction [18, 58]. Diffusion posterior illumination [40] learns the distribution of environment illumination with a diffusion model for material recovery. In contrast, our method learns the material prior for supervision and design it as conditional diffusion models of albedo and specular shading, enabling the direct supervision on material estimation. + +# 3 Background + +# 3.1 Physically-Based Rendering + +We adopt the physically based rendering [25] to recover the object material and lighting from images. For a 3D point $\mathbf{x}$ , its outgoing light $L_{o}$ at direction $\omega_{o}$ is defined as: + +$$ +L _ {o} \left(\omega_ {o}; \mathbf {x}\right) = \int_ {\Omega} L _ {i} \left(\omega_ {i}\right) f _ {r} \left(\mathbf {x}; \omega_ {o}, \omega_ {i}\right) \left(\omega_ {i} \cdot \mathbf {n}\right) d \omega_ {i}, \tag {1} +$$ + +where $L_{i}(\omega_{i})$ is the incident lighting along the direction $\omega_{i}$ , $\mathbf{n}$ is surface normal at $\mathbf{x}$ . In practice, we define the lighting $L_{i}$ as an estimable longitude-latitude environment map. We use the simplified Disney BRDF model [8] for $f_{r}$ : + +$$ +f _ {r} (\mathbf {x}; \omega_ {o}, \omega_ {i}) = \frac {\mathbf {k} _ {d}}{\pi} + \frac {D \cdot G \cdot F}{4 (\mathbf {n} \cdot \omega_ {\mathbf {o}}) (\mathbf {n} \cdot \omega_ {i})}, \tag {2} +$$ + +where $D, G$ , and $F$ are the normal distribution, geometry attenuation, and Fresnel effect, respectively. The albedo $\mathbf{k}_d$ and roughness $\rho$ parameters are set as + +learnable, and other parameters are set as constant. The detailed formulation can be found in supplementary material. + +The RGB image is rendered by evaluating Eq. (1) with the environment lighting, albedo image, roughness image, and scene mesh. We solve the integral in Eq. (1) using Monte-Carlo sampling with Multiple Importance Sampling [61]. + +# 3.2 Diffusion Model and Diffusion Posterior Sampling + +The diffusion model is a probabilistic generative model that generates images from pure Gaussian noise through a progressive de-noising process [21]. Each step of the denoising process from $\mathbf{x}_t$ to $\mathbf{x}_{t - 1}$ is defined as: + +$$ +\mathbf {x} _ {t - 1} = \frac {1}{\sqrt {\alpha (t)}} \left(\mathbf {x} _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \hat {\alpha} _ {t}}} \epsilon_ {\theta} (\mathbf {x} _ {t}, t)\right) + \sigma_ {t} \mathbf {z}, \tag {3} +$$ + +where $\epsilon_{\theta}$ is the learnable noise predictor, $\alpha_{t}$ and $\sigma_{t}$ are predefined constant values, and $\mathbf{z}$ is a random value sampled from a standard Gaussian distribution. The noise predictor is typically implemented as a UNet [50]. + +The diffusion posterior sampling (DPS) [13] is a technique for solving inverse problems (e.g. image deblur) using the diffusion model. Specifically, given an observation $\mathbf{y}$ derived from $\mathbf{x}_0$ which defined as: + +$$ +\mathbf {y} = \mathcal {A} \left(\mathbf {x} _ {0}\right) + \mathbf {n} \tag {4} +$$ + +where $\mathcal{A}$ is the measurement operator, $\mathbf{n}$ is the measurement noise. The goal is to estimate $\mathbf{x}_0$ from observation $\mathbf{y}$ using posterior distribution $p(\mathbf{x}_0|\mathbf{y})$ . To sample from $p(\mathbf{x}_0|\mathbf{y})$ with a pre-trained distribution $p(\mathbf{x}_0)$ , the DPS sampling approximates the posterior sampling functions by adding a guiding term to the denoising score $\epsilon_{\theta}(\mathbf{x}_t,t)$ in Eq. (3): + +$$ +\tilde {\epsilon} _ {\theta} \left(\mathbf {x} _ {t}, t\right) = \epsilon_ {\theta} \left(\mathbf {x} _ {t}, t\right) + \zeta_ {t} \nabla_ {\mathbf {x} _ {t}} \left| \left| \mathcal {A} \left(\mathbf {x} _ {t} ^ {0}\right) - \mathbf {y} \right| \right| _ {2} ^ {2}, \tag {5} +$$ + +where $\mathbf{x}_t^0$ is the posterior estimation of $\mathbf{x}_0$ at timestamp $t$ and $\zeta_t$ is the step size. This formula assumes a known $\mathcal{A}$ and $\mathbf{n}$ follows a Gaussian distribution of $\mathcal{N}(\mathbf{0},\sigma^2\mathbf{I})$ . We refer to DPS [13] for more details. In Section 4, we will demonstrate how to model the material priors using the diffusion model and use DPS sampling to generate multi-view consistent material samples. + +# 4 Method + +Our method aims at recovering material and lighting parameters from multi-view images under a single unknown illumination. We define object material as albedo and roughness without considering the metallic attribute and use a lat-long environment map for lighting. Due to the ill-posed nature of inverse rendering [20, 81], ambiguities exist in the optimization process, leading to inaccurate material estimation. We propose using data-driven priors to constrain the optimization to eliminate these ambiguities. + +![](images/4243a67edf082b2aa8a26b4b13b784a6f06ed88b91202f3d585c247a089d92e2.jpg) +Fig. 2: Overview of our pipeline. (a) Based on physically based rendering, our model combines geometry, lighting, roughness, and albedo into RGB and specular images, and optimizes the lighting and materials in a two-stage manner. In the first stage, our model is supervised by images and diffusion priors to output coarse albedo and roughness. Subsequently, the coarse materials are used to guide diffusion models to provide more multi-view consistent constraints. (b) The guided sampling first calculates the L2 loss between the guidance and one-step denoised signals, and then adds the gradient of the L1 loss to the output of the noise predictor. + +![](images/b28fce0c0667b4d0ae45e1416d86fe95e7c04dfb24c132185a06ce7e608ee31f.jpg) + +Inspired by the recent success of generative models [49] on image generation, we design two material priors modeled by the conditional diffusion model (Sec. 4.1). To recover object materials and lighting, we first employ monosdf [75] to reconstruct a 3D mesh from input images. The materials are defined as learnable texture maps in UV space from the UV-mapped mesh using Blender [6], and lighting as estimizable longitude-latitude environment map. Then we apply our coarse-to-fine optimization process (Sec. 4.2) to recover accurate materials and lighting, while the geometry and the prior models are fixed. An overview of our optimization process is shown in Fig 2. + +# 4.1 Albedo and Specular Prior Models + +Design of the material priors. To better decompose materials and lighting, we propose to use data-driven priors to constrain the optimization process. Ideally, we can directly apply prior distributions of albedo and roughness to narrow down the solution space and effectively alleviate ambiguities. However, roughness, being a concealed characteristic of a material, is difficult to observe and is also hard to acquire on most large-scale synthetic datasets [17]. Because these objects have a variety of material parameters that do not necessarily have a valid roughness. + +Fortunately, image rendering using the Disney BRDF model [8] is composed of specular shading and diffuse shading regardless of the choice of material parameters, as shown in Eq.(2). Thus, the rendering equation can be rewritten to: + +$$ +\begin{array}{l} L _ {o} (\omega_ {o}; \mathbf {x}) = \frac {\mathbf {k} _ {d}}{\pi} \int_ {\Omega} L _ {i} (\omega_ {i}) (\omega_ {i} \cdot \mathbf {n}) d \omega_ {i} + \int_ {\Omega} L _ {i} (\omega_ {i}) \frac {D \cdot G \cdot F}{4 (\mathbf {n} \cdot \omega_ {\mathbf {o}}) (\mathbf {n} \cdot \omega_ {i})} (\omega_ {i} \cdot \mathbf {n}) d \omega_ {i} \\ = S _ {\text {d i f f}} (\mathbf {x}, L _ {i}, k _ {d}) + S _ {\text {s p e c}} \left(\omega_ {o}; \mathbf {x}, \rho , L _ {i}\right). \tag {6} \\ \end{array} +$$ + +Motivated by Eq. (6), we propose to apply priors on specular shading which exists in all physically based renderings, instead of directly using priors on roughness. Specular shading is closely linked to roughness and has the advantage of being more observable. Moreover, based on any synthetic object, we can create the ground-truth specular shading for training the prior model. + +Modeling the material priors. Previous methods [31, 35, 54, 70, 72, 73] generally model the prior of object material using a deterministic prediction network, which attempts to regress the material from a single image. In contrast, we leverage generative models to capture the distribution of albedo and specular, considering that there could be multiple solutions for the underlying object material given captured images [4]. + +Given an RGB image $I$ , we can model the conditional distribution of target albedo $p_{\theta}(\mathbf{k}_d|I)$ and specular shading $p_{\theta}(S_{\mathrm{spec}}|I)$ by modifying the noise predictor to accept $I$ as its conditioning. Denote the modified noise predictor as $\epsilon_{\theta}(\mathbf{x}_t,t;I)$ . Specifically, we first use the CLIP image encoder [47] to extract a feature vector from the conditioning image. Then, to inject the conditioning signal into the UNet, we utilize a transformer network [60] to perform the cross-attention between the image feature vector and the intermediate feature maps of the UNet. More details can be found in the supplementary. + +The diffusion model can model complex distributions in a step-wise denoising manner, which is crucial for albedo and specular shading recovery due to the complexity of lighting and material in real-world data. Moreover, as a generative model, the diffusion model facilitates a guided sampling procedure [13, 22, 41], which permits us to steer the diffusion model toward producing samples that align with multi-view image observations. + +# 4.2 Material Prior for Inverse Rendering + +Based on the albedo and specular shading prior models, we design a two-stage optimization process for inverse rendering, which effectively alleviates the ambiguities. An overview of our method is presented in Fig. 2. We first use the conditional diffusion model to regularize the inverse rendering process, which outputs an initialization of lighting, albedo, and roughness. Then, the coarse material is used to guide the sampling of the diffusion model, enabling it to produce more deterministic results and better stabilize the optimization process. + +![](images/59d7a59039138aac4981f62a589e84b9dcbf8120f252738062a6b2aa48027378.jpg) +Input View + +![](images/ac7f5d1df27f44e3d3328745738a092d83775572510de0e6c2dcaa742371af6d.jpg) +Reference + +![](images/f4e8fb00a4f30a28f0048186c15868b78f96fad854cfb572201d657f755855bb.jpg) +Fig. 3: Effect of the guided sampling. We visualize the samples generated from the albedo prior model of different viewpoints. Without the guided sampling, the materials are inconsistent across multi-views and do not align with the material decomposition from the observed lighting. + +![](images/89fc32e2c9da09fa1c50c9e9e90f1eca124e78f1f8965a2c9da563099c557f36.jpg) +w/o guidance + +![](images/582351d10cf2eadc149d0d83865d5ed324a2e39e342029464234aca8308f872f.jpg) + +![](images/55281b6ce5962942555c63cf7f320e7478a94cae361ce5a4f84e2e2d57fe2883.jpg) + +![](images/8874c50f029fad77faa398a7bbda971d2cd32d06d605771a323fb763567ca7f0.jpg) + +![](images/1e95969c6d881cb64bbf42cb823259e8cad71c83fbcfb90cdddd20292da13c75.jpg) +w/guidance + +![](images/3a04067962f9c1422c0fc14161d1c0fd7c60639c54165224b9df153397a38505.jpg) + +![](images/7adbb3cda7d53baeb6ec15651f8a2126c9ea3ee3a22b7c28fee18d3fdf959016.jpg) + +Training with diffusion model. Given input multi-view images of the target object, we utilize the conditional diffusion model to predict their albedo and specular shading components, which are used to supervise the model parameters. Due to the probabilistic nature of the diffusion model, the generated albedo and specular images are inconsistent across camera views in terms of intensity. To overcome this problem, we apply shift and scale invariant loss [48] to compute the difference between the predicted and generated material. + +In the first stage, the training loss function is defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {c o a r s e}} = \lambda_ {1} \| \hat {I} - I \| + \lambda_ {2} \| \operatorname {S S I} \left(\hat {\mathbf {k}} _ {d}, \mathbf {k} _ {d} ^ {\mathrm {s}}\right) - \mathbf {k} _ {d} ^ {\mathrm {s}} \| \tag {7} \\ + \lambda_ {3} | | \hat {S} _ {\text {s p e c}} - S _ {\text {s p e c}} ^ {\mathrm {s}} | | + \lambda_ {4} \mathcal {L} _ {\text {s m o o t h}}, \\ \end{array} +$$ + +where $\mathrm{SSI}(a,b)$ aligns $a$ to the scale and mean of $b$ . $\hat{I}$ , $\hat{k}_d$ , $\hat{S}_{\mathrm{spec}}$ are the image, albedo, and specular shading to be optimized respectively, $\mathbf{k}_d^{\mathrm{s}}$ and $S_{\mathrm{spec}}^{\mathrm{s}}$ are samples from our diffusion models $p_\theta (\mathbf{k}_d|I)$ and $p_{\theta}(S_{\mathrm{spec}}|I)$ respectively. We also use a smoothness constraint $\mathcal{L}_{\mathrm{smooth}}$ on the optimized materials. $\lambda_1$ , $\lambda_2$ , $\lambda_3$ , $\lambda_4$ are hyper-parameters. More details can be found in the supplementary material. + +Training with guided diffusion samples. After the training in the first stage, we obtain an initialization of the object material and the environment lighting. To better guide the optimization process, we aimed to generate samples from the diffusion prior models that are consistent with multiple viewpoints, aligned with observed lighting conditions, and still follow the distribution of priors. To this end, we leverage the Diffusion Posterior Sampling (DPS) [13] to guide the sampling process, and use the estimated albedo $\hat{\mathbf{k}}_d$ and specular shading $\hat{S}_{\mathrm{spec}}$ as guiding signals to re-generate samples. As shown in Fig. 3, samples after guidance are multi-view consistent and fit observations under the current lighting condition. Specifically, we use the score function of DPS in Eq. (5) with $\mathcal{A}$ as the identity operator to re-generate albedo $\tilde{\mathbf{k}}_d$ and specular $\tilde{S}_{\mathrm{spec}}$ : + +$$ +\tilde {\epsilon} _ {\theta} \left(\tilde {\mathbf {x}} _ {t}, t; I\right) = \epsilon_ {\theta} \left(\tilde {\mathbf {x}} _ {t}, t; I\right) + \gamma \bigtriangledown_ {\tilde {\mathbf {x}} _ {t}} \left| \left| \tilde {\mathbf {x}} _ {t} ^ {0} - \hat {\mathbf {x}} \right| \right| _ {2}, \tag {8} +$$ + +where $\tilde{\mathbf{x}}_t$ and $\tilde{\mathbf{x}}_t^0$ is the noisy latents and the posterior estimation of $\hat{\mathbf{k}}_d$ or $\tilde{S}_{\mathrm{spec}}$ on step $t$ , respectively. $\hat{\mathbf{x}}$ is $\hat{\mathbf{k}}_d$ or $\hat{S}_{\mathrm{spec}}$ . We choose the step size following [13] as $\zeta_i = \gamma / ||\tilde{\mathbf{x}}_t^0 - \hat{\mathbf{x}}||$ , where $\gamma$ is a pre-defined scalar. Finally, we use the re-generated multiview-consistent material $\tilde{\mathbf{k}}_d$ or $\tilde{S}_{\mathrm{spec}}$ from the DPS to supervise the lighting + +and materials using Eq. (7). Experiments show that the second stage gives an accurate decomposition of the object material and environment lighting. + +An alternative way to supervise the inverse rendering with diffusion models is the Score Distillation Sampling (SDS) [45], which achieves consistency by finding a mode in the distribution through optimization of the material field. However, our experiments find that the SDS guidance usually produces blurry results. A possible reason is that the SDS guidance relies on noise prediction that has a high variance, which leads to unstable convergence as discussed in [63]. In constraint, our pipeline uses guided image samples that are consistent with low variance. + +# 5 Experiments + +# 5.1 Implementation details + +Prior Model Training. We create albedo and specular training data based on Objaverse dataset [17]. We randomly select 350k objects, excluding those with pure black or white albedos, and gather 180 HDR environment maps from [1] for rendering. Each object is rendered under 10 random poses with one random HDR environment map. We use a diffusion model fine-tuned from stable-diffusion [49] for the image variation task [2] as a starting point and continue to train 100k iterations for the albedo prior and 350K iterations for the specular shading prior. The training process for the two prior models takes around 10 days on 8 NVIDIA A100 (80GB) GPU. + +Inverse Rendering. The resolutions for texture maps and environment are $2048 \times 2048$ and $512 \times 256$ , respectively. We use Nvdiffrast [29] for rendering images with 144 samples per pixel. We train the first stage of our method for 20k iterations and the second stage for 15k iterations. The setting of hyperparameters is presented in the supplement materials. The inverse rendering for each scene takes around 1.5 hours on a single NVIDIA RTX A6000 GPU. + +# 5.2 Evulation + +Dataset. We collect 4 CAD models with a variety of object shapes and materials. Each object is rendered with 120 posed RGBA images using Blender Cycles [6]. We use 100 images for training and 20 for testing. For evaluation, we also render 20 novel test views under novel lights along with their ground truth BRDFs. + +Baselines. We compare our methods with the following two types of baseline methods. (1) Optimization-based methods [24, 43, 76, 81, 81]. InvRender [81] and TensoIR [24] both models indirect illumination with MLPs and uses Spherical Gaussian(SG) as environment light. InvRender [81] uses MLP representations while TensoIR is inspired by tensorial representation [10]. NeILF++ [76] models an incident light field that includes both environemnt light and indirect light with an MLP constrained by the outgoing radiance field. NvDiffRec [43], and NvDiffrecMC [20] extracted explicit topology from neural SDF and directly optimized + +Table 1: Quantitative results on the synthetic dataset. "NvDiffrec*" and "NvDiffrecMC*" mean these methods are with our reconstructed mesh. "Ours w/o guided" means our method without the guided sampling strategy. TensorIR and NvDiffrec have strong capabilities of fitting observed images, but they struggle to recover correct object materials. Replacing our diffusion priors with other priors [38, 71] leads to the degradation of the performance on material recovery. + +
MethodAlbedoAligned AlbedoView SynthesisRelightingRoughness
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓MSE↓
InvRender [81]17.81810.92470.064523.75920.94910.055927.30510.96730.038624.36320.96210.05150.0420
TensoIR [24]20.29720.95270.063124.93680.95640.061433.12340.98030.026421.34730.95430.05140.0456
NvDiffrec [43]16.54670.92400.070621.53040.95070.057432.57780.98240.019021.99390.95500.07340.0777
NvDiffrec* [43]14.52040.90210.083520.22800.92040.065931.03360.97840.021821.24520.93560.06380.0841
NvDiffrecMC [20]13.04950.88150.106821.67700.94060.069930.10830.96740.044022.53370.94860.07970.1046
NvDiffrecMC* [20]11.13530.87220.117619.83250.93050.080528.16340.95600.044021.03800.93320.08220.1411
NEILF++* [76]13.83360.91720.091518.45430.93660.083129.35270.97750.027620.48780.96290.05900.0855
Yi et al. [71]13.20290.89340.083318.84280.91330.073126.50910.96690.039321.1560.95620.05070.1524
USI3D [38]11.67130.88720.105118.92960.93440.077226.96530.96660.048020.27210.95870.06120.1241
Ours24.67660.95620.050326.88290.95320.037830.24050.97410.032828.97210.97570.02890.0105
+ +![](images/d9a8da1910ec10810d4118932714860b9647890c877d26017fd6c8dbe89b9c86.jpg) +Fig. 4: Qualitative comparison in terms of relighting on the synthetic dataset. Zoom in for details. + +![](images/7af24bb2ef373a8441084b1cdf00e41b80fc31325de940550238d7a004acc8c1.jpg) + +the 3D mesh with its PBR materials. Their illumination is expressed as an HDR environment map, which can be optimized by NviDiffRec's [43] split sum environment lighting and NviDiffRec's Monte Carlo ray tracing. (2) Data-driven methods [38, 71]. We keep object and lighting models as the same and just replace diffusion priors with single-view intrinsic image decomposition methods for comparison. USI3D [38] models the decomposition task as domain transfer from the image domain to the albedo or shading domain using generative models. We adopt only the predicted albedo to supervise the optimization because their rendering model is not physically based. Yi. el [71] propose a network to predict and remove specular shading and subsequently recover the albedo. We use the predicted specular shading and albedo as the priors. + +Evaluation protocol. We analyze our method and the baselines in the following way. On the 4 synthetic datasets, we compare their reconstruction of albedo, roughness, and novel view synthesis quantitatively with the ground truths. Given + +![](images/c0af9d936d35d0b0a761b2df22657ff88a8a34d1af29b93168410777dbe62975.jpg) +Fig. 5: Quantitative comparison with baselines on the synthetic dataset. Our method estimates more accurate albedo and roughness than other methods. Nvdiffrec [43] and NvdiffrecMC [20] tend to bake the lighting into the albedo, due to the ambiguity between the diffuse shading and albedo. InvRender [81] and TensoIR [24] struggle to decouple the roughness from the albedo and environment lighting. + +![](images/76407c55ede19a6e8c32707f3c943b1d6d29f5e4e3562b7c0e4eb052fc326e2c.jpg) +Fig.6: Qualitative results on DTU. Our method recovers reasonable materials and relighting on real datasets. Baseline models can not disentangle illumination and materials which results in the baked-in effect on materials estimation. + +GT + +![](images/b8fab22c6ae49b02c1fb1b66808fd826077a329c35f4ce7e485ff5e8eedb74af.jpg) + +![](images/95031cf721cb4b9032ed88964cc0a3d204cf15d86ef154e9160bcc258482c4a4.jpg) + +![](images/41e69272b99b23cc41bd26c8e753793d23db5ec795ad43fd1567ea02774ea5fe.jpg) + +![](images/9d258d7ff7e46c6fcb826403db9ef4e571e00eb2d3fc21c8cd1899aedafc9927.jpg) + +![](images/2ef8a966bc475e3fe44ba16a17af541baa8a77e51f8f4e3e520365d98ec784ba.jpg) + +![](images/710fc2037e5667e2322b0ac0d01ca380a333cbf4cc6131a7b39ef51caab513aa.jpg) + +![](images/1ca721e7b50404ea76ed802a5ad335983db7dfff6080775cf86fe451b301adbb.jpg) + +![](images/7a2e6a16ce625920ea6c051b73b4feba74d0b9e88c173d614bc1028a246f7ceb.jpg) + +USIBD + +![](images/da9d0c7a36aaf9d34a4fa74a7e1e7dc75f66202a2f533501f84be66300b5e51f.jpg) + +![](images/07f1738bd7ce200dcaf7b8cd70fe1ccc26be10f239c15616bf8d269b2e84c092.jpg) + +![](images/047963e72477ca14189977740918482cf11d1014f70b3b72bd95392716b0fc2e.jpg) + +![](images/49fe7d179dbadb95ff6ca4f85fc4622226ed4c97760aa1e248cb16c0906f96f6.jpg) + +![](images/9513a7232b5154aaa08b2b03c1c107c4f53151ae4134941396070890558afa9f.jpg) + +![](images/23984463a674afed8486ccd9019b495eb67fcf9cecd799f6e5f3c463c24c80bc.jpg) + +![](images/52a1eae2f21686a1686694923c30c7f649b64cd024b98c5f181bb95eae8a855b.jpg) + +![](images/693d1e8184130e652d34038f9bb80e33e56d2aeb40976ed81b8cab21c0a7a6f6.jpg) + +Yiet al. + +![](images/b913d8ae2acf5bef63e607a28263e9a3930cc7b56bb9a5638e1bed34af38aad6.jpg) + +![](images/3041555f2617046202c212c116a35c3d9921038aec4e9a7190258ae68e4a19ba.jpg) + +![](images/775ba24bfeb822776b193b1cf2a334d4432efacf4691bf0193a4645c6e4b4262.jpg) + +![](images/7cd7426dcde5d7b3367cc327c76c4c1433254fcd2e44df40cf15a9c0724b56c3.jpg) + +![](images/626931ceb320f989b572f9f0c449de62a4dde28b9dd47b4f17b0e338e752ee4b.jpg) + +![](images/f6dd90eca91aeac57f3967ed25d55247f4ac12177b4fcc3f354b0bd545ac6a98.jpg) + +![](images/a24d78836285131efce261c026a02a37893f03bde2cb45a0421ba54c25f8000d.jpg) + +![](images/00aee554799ab1f0c4284411e3c42f6bf4d6ce013034bfb9b4200f17c4658a16.jpg) + +Ours + +![](images/3f88d5f2729a6b63b1baadda0dbb0970e0827e3dc9152203b2615a9c4f08ab5d.jpg) +Rendering + +![](images/3e504af9128ea8907af5d2fc127594a14d22842eba016d93e0869344f00c80c2.jpg) +Albedo + +![](images/eac82589400dbabcc697961e438f8136b99c7cca4b39e4e79eec39d7f34b0a24.jpg) +roughness + +![](images/69323977353543fdf250021d5195e067e30b57900190b3a2a4761eabf1e486af.jpg) +Relight + +![](images/b02d56606838c0339dbb6a89db943a41b9f0a9b1c36eb429de7347b5b1e874d4.jpg) +Rendering + +![](images/258c5e2848d08c7b91f0cb92251508f67cdaf55ed9d6f996a7fded522108607e.jpg) +Albedo +Fig. 7: Comparison with other data-driven priors. + +![](images/e3f24526805a6d9d2971bc1b5566587094528c4095050b714a737cb9812ff02e.jpg) +roughness + +![](images/92aaf15911c5b766be51c4f1f8b55011651a7548f6822f2f6a7997cefc093fd8.jpg) +Relight + +the environment map from our novel light dataset, we also make a comparison on the relighted images. Because Invrender [81] and TensoIR [24] parameterize environment maps with SGs, we first optimize an SG with 128 lobes through Stochastic Gradient Decent optimization. For NvDiffRec [43], NvDiffRecMC [20], and NeILF++ [76], we render the relighted image with Blender [6] using our extracted topology and material texture maps with SPP=4096 and 4 bounces of lights. Due to the inevitable scale ambiguity between lighting and albedo, previous methods [80] align the predicted albedo to the ground truth before relighting and evaluation. Each channel is separately aligned to the ground truth. However, this would cause the incorrectly baked shading in the albedo to be removed during alignment. To this end, we align the average RGB of prediction to ground truth for a more faithful evaluation. Because geometry affects the inverse rendering performance greatly, to make sure the topology for NvDiffRec and NvdiffRecMC is well initialized, we make another set of experiments for NvDiffRec and NvDiffRecMC with MonoSDF [75] reconstructed geometry given and as a fixed topology along the training and evaluation process. + +Metrics. We use Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) [79] as metrics to evaluate the image quality of the aligned albedo, relighting images, and novel view synthesis. Roughness is evaluated using Mean Squared Error(MSE). + +Comparison with optimization-based methods. Tab. 1 quantitatively compares our method with recent inverse rendering methods on the synthetic dataset. We significantly outperform baseline methods in terms of recovered albedo, roughness, and relighting, demonstrating the effectiveness of the proposed material prior. Although TensoIR [24] and NvDiffRec [43] show good results on view synthesis under the original environment lighting, it does not accurately estimate the underlying materials, which will be further shown in the qualitative results. + +Tab. 1 also shows that we further improve the performance of material recovery by guiding the diffusion model with estimated albedo and specular. + +Fig. 4 and Fig. 5 present the qualitative results on estimated materials and relighting. Fig. 6 demonstrates the performance of our method on the real-world data. More results on real data are presented in the supplementary material. NvDiffrec and $\mathrm{NeILF}++$ do not disentangle the albedo and light well and show strong light color baked in albedo, as visualized in Fig. 5 and Fig. 6. With our diffusion priors learned from the large-scale dataset, our method can disentangle materials and lighting in a reasonable way. + +Comparison with other data-driven priors. Tab. 1 presents the comparison with data-driven priors [38,71]. Although these methods are able to prevent shadows from being baked into albedo, they still struggle to separate shading from the albedo of objects under complex lighting conditions, as shown in qualitative results in Fig. 7. The incorrect priors will mislead the optimization process and cause large errors in the estimated materials. In contrast, our prior models have a better capacity to decouple shading from materials. We present more qualitative comparisons of the real-world data in the supplementary material. + +Ablation Study. We conduct ablation studies on our diffusion priors and the guided diffusion for fine-level inverse rendering. Quantitative and qualitative results are shown in Tab. 2 and Fig. 8, respectively. + +"w/o spec." indicates training without the specular shading prior model. This would result in recovering incorrect lighting and roughness, as the diffuse shading and specular shading cannot be correctly separated by relying solely on the albedo prior. As depicted in Fig. 8, the "hat" 's roughness prediction is lower and the "ship" 's is higher in "w/o Spec." than ground truth, which resulting in inaccurate relighting outcomes. In "w/o albedo", the inverse rendering is conducted without albedo prior. The recovered albedo has baked-in shading, as shown in Fig. 8 where the two objects baked shading into the albedo causing incorrect relighting. This indicates that both priors are crucial for accurate lighting and material decomposition. + +"w/o guided" is the result without the guided diffusion for fine-level optimization. Due to the lack of multi-view consistency of diffusion prior, the optimized materials are noisy and have baked-in shading, as shown in Fig. 8. This hurts the relighting quality, as shown in qualitative results in Tab. 2. Therefore, the training process using the guided diffusion model is essential for a smooth and accurate material recovery. + +Modeling indirect illumination and metallic materials. Following the setting of InvRender [81], we extend our method to handle indirect lighting. Inspired by [66], we precompute shading of indirect lighting for each image by tracing the trained SDF and Color field. Additionally, we also include metallic texture as an optimization target. More details and results are in the supplementary materials. + +Table 2: Quantitative results of ablation studies. + +
MethodAlbedoAligned AlbedoView SynthesisRelightingRoughness
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓MSE↓
Ours24.67660.95620.050326.88290.95320.037830.24050.97410.032828.97210.97570.02890.0105
Ours w/o spec17.21260.93160.056725.73890.94910.044429.42450.97280.035226.47430.97290.03630.0758
Ours w/o albedo22.71110.94340.057124.34620.93810.048830.59980.97510.026127.28510.97430.02580.0118
Ours w/o guided23.35070.93030.065224.61120.92080.064129.79370.96880.033927.73900.96830.03610.0168
+ +![](images/4e2f6a49b27b75e2d0ba3041614004a29e2c9704a4e87de8edc5a20b801c5fa1.jpg) + +![](images/25a88a5f0eb08ead0fd7edb23330aff44a9b804605966f0339935a86d23cb549.jpg) + +![](images/3cec1a1aef923e4a068988e341c3470cb2b1e6ab14dd89ffec25161c82331186.jpg) + +![](images/c8a1b0f687370e4be06e9ea7ae937a40c1cd59d26b4e4510fdfc03846207cdef.jpg) + +![](images/b6107ce4f5f0e5e13a3f491a254e45d4ccb6a70183fad1726d1a6ef6f0dfa56e.jpg) + +![](images/7ea3a3e3d0aa644b6f741ebd573cfc9565bd9952d9959e5d3c3cc6ebfc7050d0.jpg) + +![](images/c26d46edb610b0b2954646e5c289b3dfdfca7905201a8b362e6985e719892b0a.jpg) + +![](images/aefe093ef6679008f19ba59badf06c6235d427673a26903784b80ff1bfb38f78.jpg) + +![](images/9f68d9eafaf428538eabbb25f98ffe10964132917cdff57efc5d77154745ac31.jpg) + +![](images/1d61b8e5fb060787664cacd372d3975d3efc06b25271e67f0a638688eff0e491.jpg) + +![](images/73ada74eb334f49b112510d04f61116daa78470a8d6ac6ae310c9b1d2ffeab9e.jpg) + +![](images/f48c1cc2dfc8a6843acd3dc9eccff5a1f6882a97256fa02b82d3718eca54dc24.jpg) + +![](images/06bf811784271d37b1a34275a8a562e2d83cb2142b2c2b417e27d0213d6e2948.jpg) + +![](images/c7adee1f6384d4b70c4782b24c05289ac931c9a674693c04651df01b51a9a484.jpg) + +![](images/3adb476411b22a08f20ad7f2ddaf450bef30e4d66a2d8fc0df31bfc4d8f570c1.jpg) + +![](images/fabafb0025ca893ea1384173a872c607d23d825096d4775b35af6db2355bad72.jpg) + +![](images/76a35658085605f4313308b50306a9abf1ee447b483357681222cb08f40988fb.jpg) + +![](images/56a1e30bc10532ca715af2e67d59499996d57383d3fb95cf631c74c41c60db92.jpg) + +![](images/618dec07efae921a338f8ecbbcab433f288a23eda4ddef48b85f41c254aef527.jpg) + +![](images/909d8b23c4930fd696037d3d2f4555817b254e4dccd3afcd4a24ff1edd51e659.jpg) + +![](images/3fcb9ecaedfb493805aeffb5ec70eaaae0c3617fc4d009c62b278e91844a60c1.jpg) +Fig. 8: Qualitative results of ablation studies. + +![](images/0892ab3d09089931ac3e4428c47116672d126fdb3cf2c7a506d322463f463f48.jpg) + +![](images/3380a9e4b7b000d9f9eb272c161b76944277df36d39dab31892c37e62ccc62c2.jpg) +w/o albedo + +![](images/b82aca426132d6298e93b0f2712841af00f84b8f2285af2e0a71d7cf389a94c0.jpg) +w/o guided + +![](images/5abebf74672bc6605707ceb6b10c8d7070b202fffd58e4c37eeb6160b02fd46f.jpg) +Ours + +![](images/cc67f98a679e89001bc2a2b2c67d0d357cc7502ff8c14ec20ed75c034c14d56b.jpg) + +![](images/56c701a677140b4acac71d9e22760cb60db94f0906aee7e66018d04f90c474f9.jpg) +GT + +![](images/50a8b47b1f7b87b0d44efa0299ce009b88857a42bac072c27dea10f8224c5626.jpg) +w/o spec. + +![](images/1efcb4c7cf35bde05878b4689928963d98aa1c9169cc62e1e0762893ba207c47.jpg) +w/o albedo + +![](images/0020161009a2a31d93a022d013ba381d607023960e176ce612154053fc346d1a.jpg) +w/o guided + +![](images/12dee739272c986f386caf2bb96213d05c96399265f096bc7d88032b13776e03.jpg) +Ours + +# 6 Conclusion + +We proposed to learn the material prior for inverse rendering under an unknown static lighting condition. To this end, the material prior is designed as conditional diffusion models of albedo and specular shading components, according to the general rendering equation. In addition, a two-stage training scheme is developed for robustly regularizing the inverse rendering process with conditional diffusion models, which uses coarse materials from the first stage training to guide diffusion models to provide multi-view consistent constraints. The proposed model significantly outperforms baseline methods on real-world and synthetic data. + +Limitations. Although we achieved impressive results on material recovery, our approach still has some limitations. First, we do not consider the inverse rendering of transparent objects. A solution is revising the geometry representation to a neural field of transparency and jointly optimizing it with object materials and environment lighting. Second, the performance of our model is limited by the accuracy of the reconstructed geometry from MonoSDF. It would be interesting to exploit diffusion models to learn the geometry prior. + +Acknowledgments. The authors would like to acknowledge the support from NSFC (No. 62172364), the Alibaba Innovation Research Program of Alibaba Group and the Information Technology Center and State Key Lab of CAD&CG, Zhejiang University. + +# References + +1. Poly haven: The public 3d asset library. https://polyhaven.com/ +2. Stable diffusion image variations. https://huggingface.co/spaces/lambdalabs/stable-diffusion-image-variations +3. Anciukevicius, T., Xu, Z., Fisher, M., Henderson, P., Bilen, H., Mitra, N.J., Guerrero, P.: Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. In: CVPR (2023) +4. Barron, J.T., Malik, J.: Shape, illumination, and reflectance from shading. TPAMI 37(8), 1670-1687 (2014) +5. Bi, S., Lombardi, S., Saito, S., Simon, T., Wei, S.E., McPhail, K., Ramamoorthi, R., Sheikh, Y., Saragih, J.: Deep relightable appearance models for animatable faces. ACM Trans. Graph. (Proc. SIGGRAPH) 40(4) (2021) +6. Blender: Blender, https://www.blender.org/ +7. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.P.: Nerd: Neural reflectance decomposition from image collections. In: IEEE International Conference on Computer Vision (ICCV) (2021) +8. Burley, B., Studios, W.D.A.: Physically-based shading at disney. In: ACM Trans. Graph. vol. 2012, pp. 1-7. vol. 2012 (2012) +9. Chang, A.X., Funkhouser, T.A., Guibas, L.J., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: Shapenet: An information-rich 3d model repository. CoRR abs/1512.03012 (2015), http://arxiv.org/abs/1512.03012 +0. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: ECCV (2022) +1. Chen, H., Gu, J., Chen, A., Tian, W., Tu, Z., Liu, L., Su, H.: Single-stage diffusion nef: A unified approach to 3d generation and reconstruction. In: ICCV (2023) +2. Cheng, Y.C., Lee, H.Y., Tulyakov, S., Schwing, A.G., Gui, L.Y.: Sdfusion: Multi-modal 3d shape completion, reconstruction, and generation. In: CVPR (2023) +3. Chung, H., Kim, J., Mccann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. In: ICLR (2023), https://openreview.net/forum?id=OnD9zGAGTOK +4. Chung, H., Sim, B., Ye, J.C.: Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In: CVPR. pp. 12413-12422 (2022) +5. Croitoru, F.A., Hondru, V., Ionescu, R.T., Shah, M.: Diffusion models in vision: A survey. PAMI (2023) +6. Debevec, P., Hawkins, T., Tchou, C., Duiker, H.P., Sarokin, W., Sagar, M.: Acquiring the reflectance field of a human face. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 145-156 (2000) +7. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objverse: A universe of annotated 3d objects. In: CVPR. pp. 13142-13153 (2023) +8. Deng, C., Jiang, C., Qi, C.R., Yan, X., Zhou, Y., Guibas, L., Anguelov, D., et al.: Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In: CVPR. pp. 20637-20647 (2023) +9. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escalano, S., Pandey, R., Dourgarian, J., Tang, D., Tkach, A., Kowdle, A., Cooper, E., Dou, M., Fanello, S., Fyffe, G., Rhemann, C., Taylor, J., Debevec, P., Izadi, S.: The relightables: Volumetric performance capture of humans with + +realistic relighting. ACM Trans. Graph. 38(6) (nov 2019). https://doi.org/10.1145/3355089.3356571, https://doi.org/10.1145/3355089.3356571 +20. Hasselgren, J., Hofmann, N., Munkberg, J.: Shape, light & material decomposition from images using monte carlo rendering and denoising. In: NeurIPS (2022) +21. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020) +22. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) +23. Horita, D., Yang, J., Chen, D., Koyama, Y., Aizawa, K., Sebe, N.: A structure-guided diffusion model for large-hole image completion. In: BMVC (2023) +24. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoir: Tensorial inverse rendering. In: CVPR (2023) +25. Kajiya, J.T.: The rendering equation. In: Proceedings of the 13th Annual Conference on Computer Graphics and Interactive Techniques. p. 143-150. SIGGRAPH '86, Association for Computing Machinery, New York, NY, USA (1986). https://doi.org/10.1145/15922.15902, https://doi.org/10.1145/15922.15902 +26. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. arXiv preprint arXiv:2201.11793 (2022) +27. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/ +28. Kocsis, P., Sitzmann, V., Nießner, M.: Intrinsic image diffusion for single-view material estimation. In: arxiv (2023) +29. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Trans. Graph. 39(6) (2020) +30. Levoy, M., Pulli, K., Curless, B., Rusinkiewicz, S., Koller, D., Pereira, L., Ginzton, M., Anderson, S.E., Davis, J., Ginsberg, J., Shade, J., Fulk, D.: The digital michelangelo project: 3d scanning of large statues. Proceedings of the 27th annual conference on Computer graphics and interactive techniques (2000), https://apisemantic scholar.org/CorpusID:1546062 +31. Li, Z., Snavely, N.: Cginintrinsics: Better intrinsic image decomposition through physically-based rendering. In: ECCV (2018) +32. Li, Z., Xu, Z., Ramamoorthi, R., Sunkavalli, K., Chandraker, M.: Learning to reconstruct shape and spatially-varying reflectance from a single image. ACM Trans. Graph. 37(6), 1-11 (2018) +33. Li, Z., Yu, T.W., Sang, S., Wang, S., Song, M., Liu, Y., Yeh, Y.Y., Zhu, R., Gundavarapu, N., Shi, J., et al.: Openrooms: An open framework for photorealistic indoor scene datasets. In: CVPR. pp. 7190-7199 (2021) +34. Liang, Z., Zhang, Q., Feng, Y., Shan, Y., Jia, K.: Gs-ir: 3d gaussian splatting for inverse rendering. Computer Graphics Forum (2023) +35. Lichy, D., Wu, J., Sengupta, S., Jacobs, D.W.: Shape and material capture at home. In: CVPR (2021) +36. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. In: ICLR (2023) +37. Liu, Y., Wang, P., Lin, C., Long, X., Wang, J., Liu, L., Komura, T., Wang, W.: Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In: ACM Trans. Graph. (2023) +38. Liu, Y., Li, Y., You, S., Lu, F.: Unsupervised learning for intrinsic image decomposition from a single image. In: CVPR (2020) + +39. Liu, Z., Li, Y., Lin, Y., Yu, X., Peng, S., Cao, Y.P., Qi, X., Huang, X., Liang, D., Ouyang, W.: Unidream: Unifying diffusion priors for relightable text-to-3d generation (2023) +40. Lyu, L., Tewari, A., Habermann, M., Saito, S., Zollhöfer, M., Leimkühler, T., Theobalt, C.: Diffusion posterior illumination for ambiguity-aware inverse rendering. ACM Trans. Graph. 42(6) (2023) +41. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022) +42. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +43. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: CVPR (2022) +44. Nam, G., Lee, J.H., Gutierrez, D., Kim, M.H.: Practical svbrdf acquisition of 3d objects with unstructured flash photography. ACM Trans. Graph. (Proc. SIGGRAPH Asia) 37(6), 267:1-12 (2018). https://doi.org/10.1145/3272127.3275017, http://dx.doi.org/10.1145/3272127.3275017 +45. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: ICLR (2023) +46. Qiu, L., Chen, G., Gu, X., Zuo, Q., Xu, M., Wu, Y., Yuan, W., Dong, Z., Bo, L., Han, X.: Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In: CVPR. pp. 9914–9925 (2024) +47. Ramanishka, V., Das, A., Zhang, J., Saenko, K.: Learning transferable visual models from natural language. In: CVPR. pp. 4397-4406 (2018) +48. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. PAMI 44(3) (2022) +49. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) +50. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI. pp. 234-241. Springer (2015) +51. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022) +52. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022) +53. Sang, S., Chandraker, M.: Single-shot neural relighting and svbrdf estimation. In: ECCV (2020) +54. Shi, J., Dong, Y., Su, H., Yu, S.X.: Learning non-lambertian object intrinsics across shapenet categories. In: CVPR (July 2017) +55. Shi, Y., Wu, Y., Wu, C., Liu, X., Zhao, C., Feng, H., Liu, J., Zhang, L., Zhang, J., Zhou, B., Ding, E., Wang, J.: Gir: 3d gaussian inverse rendering for relightable scene factorization. Arxiv (2023) +56. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. CoRR abs/2012.03927 (2020), https://arxiv.org/abs/2012.03927 +57. Sun, C., Cai, G., Li, Z., Yan, K., Zhang, C., Marshall, C., Huang, J.B., Zhao, S., Dong, Z.: Neural-pbir reconstruction of shape, material, and illumination (2023) + +58. Tang, J., Wang, T., Zhang, B., Zhang, T., Yi, R., Ma, L., Chen, D.: Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In: ICCV. pp. 22819-22829 (October 2023) +59. Tang, S., Zhang, F., Chen, J., Wang, P., Furukawa, Y.: Mvdiffusion: Enabling holistic multi-view image generation with correspondence-aware diffusion. In: NeurIPS (2023) +60. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) NIPS. vol. 30. Curran Associates, Inc. (2017) +61. Veach, E., Guibas, L.J.: Optimally combining sampling techniques for monte carlo rendering. In: Proceedings of the 22nd annual conference on Computer graphics and interactive techniques. pp. 419-428 (1995) +62. Vecchio, G., Martin, R., Roullier, A., Kaiser, A., Rouffet, R., Deschaintre, V., Boubekeur, T.: Controlmat: Controlled generative approach to material capture. arXiv preprint arXiv:2309.01700 (2023) +63. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In: NeurIPS (2023) +64. Wimbauer, F., Wu, S., Rupprecht, C.: De-rendering 3d objects in the wild (2022) +65. Wu, H., Hu, Z., Li, L., Zhang, Y., Fan, C., Yu, X.: Nefii: Inverse rendering for reflectance decomposition with near-field indirect illumination. In: CVPR. pp. 4295-4304 (June 2023) +66. Wu, L., Zhu, R., Yaldiz, M.B., Zhu, Y., Cai, H., Matai, J., Porikli, F., Li, T.M., Chandraker, M., Ramamoorthi, R.: Factorized inverse path tracing for efficient and accurate material-lighting estimation. In: ICCV. pp. 3848-3858 (2023) +67. Xie, Y., Yuan, M., Dong, B., Li, Q.: Diffusion model for generative image denoising. arXiv preprint arXiv:2302.02398 (2023) +68. Yao, Y., Zhang, J., Liu, J., Qu, Y., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf: Neural incident light field for physically-based material estimation. In: ECCV (2022) +69. Ye, W., Chen, S., Bao, C., Bao, H., Pollefeys, M., Cui, Z., Zhang, G.: IntrinsicNeRF: Learning Intrinsic Neural Radiance Fields for Editable Novel View Synthesis. In: ICCV (2023) +70. Yi, R., Tan, P., Lin, S.: Leveraging multi-view image sets for unsupervised intrinsic image decomposition and highlight separation. In: AAAI. vol. 34, pp. 12685-12692 (2020) +71. Yi, R., Zhu, C., Xu, K.: Weakly-supervised single-view image relighting. In: CVPR. pp. 8402-8411 (June 2023) +72. Yu, Y., Meka, A., Elgharib, M., Seidel, H.P., Theobalt, C., Smith, W.: Self-supervised outdoor scene relighting. In: ECCV (2020) +73. Yu, Y., Smith, W.A.: Inverserendernet: Learning single image inverse rendering. In: CVPR. pp. 3155-3164 (2019) +74. Yu, Y., Malik, J.: Recovering photometric properties of architectural scenes from photographs. In: Proceedings of the 25th Annual Conference on Computer Graphics and Interactive Techniques. p. 207-217. SIGGRAPH '98, Association for Computing Machinery, New York, NY, USA (1998). https://doi.org/10.1145/280814.280874, https://doi.org/10.1145/280814.280874 +75. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. NeurIPS (2022) + +76. Zhang, J., Yao, Y., Li, S., Liu, J., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neil++: Inter-reflectable light fields for geometry and material estimation. ICCV (2023) +77. Zhang, K., Luan, F., Li, Z., Snavely, N.: Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In: CVPR (2022) +78. Zhang, K., Luan, F., Wang, Q., Bala, K., Snavely, N.: PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In: CVPR (2021) +79. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) +80. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph. 40(6) (dec 2021) +81. Zhang, Y., Sun, J., He, X., Fu, H., Jia, R., Zhou, X.: Modeling indirect illumination for inverse rendering. In: CVPR (2022) +82. Zhu, J., Huo, Y., Ye, Q., Luan, F., Li, J., Xi, D., Wang, L., Tang, R., Hua, W., Bao, H., et al.: I2-sdf: Intrinsic indoor scene reconstruction and editing via raytracing in neural sdfs. In: CVPR. pp. 12489-12498 (2023) +83. Zhu, J., Luan, F., Huo, Y., Lin, Z., Zhong, Z., Xi, D., Wang, R., Bao, H., Zheng, J., Tang, R.: Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In: SIGGRAPH Asia 2022 Conference Papers. ACM (2022), https://doi.org/10.1145/3550469.3555407 +84. Zhu, Y., Zhang, K., Liang, J., Cao, J., Wen, B., Timofte, R., Van Gool, L.: Denoising diffusion models for plug-and-play image restoration. In: CVPR. pp. 1219-1229 (2023) \ No newline at end of file diff --git a/2024/IntrinsicAnything_ Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination/images.zip b/2024/IntrinsicAnything_ Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..525949d7fa78a16d74e611d9b3d719094f097d9c --- /dev/null +++ b/2024/IntrinsicAnything_ Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74603f269cf0025df04fb4105dc389ce65677ada8e2ceb66dd0a8f3eb3fbf6ff +size 644713 diff --git a/2024/IntrinsicAnything_ Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination/layout.json b/2024/IntrinsicAnything_ Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f0b83ec3c233085d6530cc1050956b578a79b738 --- /dev/null +++ b/2024/IntrinsicAnything_ Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination/layout.json @@ -0,0 +1,13532 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 111, + 462, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 111, + 462, + 164 + ], + "spans": [ + { + "bbox": [ + 153, + 111, + 462, + 164 + ], + "type": "text", + "content": "Intrinsic Anything: Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 192, + 186, + 422, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 186, + 422, + 210 + ], + "spans": [ + { + "bbox": [ + 192, + 186, + 422, + 210 + ], + "type": "text", + "content": "Xi Chen" + }, + { + "bbox": [ + 192, + 186, + 422, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 192, + 186, + 422, + 210 + ], + "type": "text", + "content": ", Sida Peng" + }, + { + "bbox": [ + 192, + 186, + 422, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 192, + 186, + 422, + 210 + ], + "type": "text", + "content": ", Dongchen Yang" + }, + { + "bbox": [ + 192, + 186, + 422, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 192, + 186, + 422, + 210 + ], + "type": "text", + "content": ", Yuan Liu" + }, + { + "bbox": [ + 192, + 186, + 422, + 210 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 192, + 186, + 422, + 210 + ], + "type": "text", + "content": ", Bowen Pan" + }, + { + "bbox": [ + 192, + 186, + 422, + 210 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 192, + 186, + 422, + 210 + ], + "type": "text", + "content": ", Chengfei Lv" + }, + { + "bbox": [ + 192, + 186, + 422, + 210 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 192, + 186, + 422, + 210 + ], + "type": "text", + "content": ", and Xiaowei Zhou" + }, + { + "bbox": [ + 192, + 186, + 422, + 210 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 201, + 219, + 410, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 219, + 410, + 230 + ], + "spans": [ + { + "bbox": [ + 201, + 219, + 410, + 230 + ], + "type": "text", + "content": "1 State Key Lab of CAD&CG, Zhejiang University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 241, + 230, + 372, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 230, + 372, + 242 + ], + "spans": [ + { + "bbox": [ + 241, + 230, + 372, + 242 + ], + "type": "text", + "content": "2 The University of Hong Kong" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 211, + 242, + 402, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 242, + 402, + 253 + ], + "spans": [ + { + "bbox": [ + 211, + 242, + 402, + 253 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 211, + 242, + 402, + 253 + ], + "type": "text", + "content": " Tao Technology Department, Alibaba Group" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 272, + 455, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 272, + 455, + 503 + ], + "spans": [ + { + "bbox": [ + 160, + 272, + 455, + 503 + ], + "type": "text", + "content": "Abstract. This paper aims to recover object materials from posed images captured under an unknown static lighting condition. Recent methods solve this task by optimizing material parameters through differentiable physically based rendering. However, due to the coupling between object geometry, materials, and environment lighting, there is inherent ambiguity during the inverse rendering process, preventing previous methods from obtaining accurate results. To overcome this ill-posed problem, our key idea is to learn the material prior with a generative model for regularizing the optimization process. We observe that the general rendering equation can be split into diffuse and specular shading terms, and thus formulate the material prior as diffusion models of albedo and specular. Thanks to this design, our model can be trained using the existing abundant 3D object data, and naturally acts as a versatile tool to resolve the ambiguity when recovering material representations from RGB images. In addition, we develop a coarse-to-fine training strategy that leverages estimated materials to guide diffusion models to satisfy multi-view consistent constraints, leading to more stable and accurate results. Extensive experiments on real-world and synthetic datasets demonstrate that our approach achieves state-of-the-art performance on material recovery. The code will be available at https://zju3dv.github.io/IntrinsicAnything/." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 519, + 230, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 519, + 230, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 230, + 532 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 540, + 482, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 482, + 649 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 482, + 649 + ], + "type": "text", + "content": "Recovering the object's geometry, material, and lighting from captured images, also known as inverse rendering, is a long-standing task in computer vision and graphics. These physical attributes of 3D objects are essential for many applications, such as VR/AR, movie production, and video games. Due to the inherent complexity of interaction between real-world objects and environment lighting, inverse rendering remains an ill-posed problem. Previous works overcome this problem using sophisticated capture systems [16,19], or co-located flashlight and camera in a dark environment [5,44,77]. However, these methods require special hardware equipment or constrained environment, limiting their applications." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 654, + 240, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 654, + 240, + 665 + ], + "spans": [ + { + "bbox": [ + 142, + 654, + 240, + 665 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 112, + 482, + 220 + ], + "blocks": [ + { + "bbox": [ + 143, + 112, + 482, + 220 + ], + "lines": [ + { + "bbox": [ + 143, + 112, + 482, + 220 + ], + "spans": [ + { + "bbox": [ + 143, + 112, + 482, + 220 + ], + "type": "image", + "image_path": "6d6a8324d8e7504e543a22a6d863ec9b12a76bbadfddc5b5d244daf61b216b48.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 228, + 482, + 295 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 295 + ], + "type": "text", + "content": "Fig. 1: Two types of ambiguities in inverse rendering. (a) Ambiguity between diffuse shading and albedo. For example, the Xbox is lit by a yellow light, and the decomposed albedo from NVdiffrecMC [20] tends to be yellow. (b) Ambiguity between shadow and albedo. For example, the porcelain toy is with self-occlusion, and TensoIR [24] bakes the shadow into the recovered albedo. Our method well handles the two types of ambiguities." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 324, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 456 + ], + "type": "text", + "content": "Based on differentiable physically based rendering and neural scene representations, recent methods [20, 24, 76, 78, 80, 81] are able to object materials from images captured under natural lighting. They exploit neural networks to represent object materials and geometry, and combine these object attributes with the learnable lighting to synthesize images, which are compared with captured images to optimize the model paramaters. To achieve better performance of inverse rendering, these methods have attempted to improve object representations [24, 78], lighting representations [20, 76], and training strategies [80, 81]. However, we argue that depending solely on the rendering loss is insufficient for accomplishing accurate decomposition, due to the inherent ambiguity in the inverse rendering process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 459, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 482, + 544 + ], + "type": "text", + "content": "We observed that there are two categories of ambiguities, as shown in Fig. 1. First, there is ambiguity between diffuse shading and albedo. Intuitively, when a white object is lit by a yellow light, the decomposed albedo appears to be yellow. Second, when an object is with self-occlusion and illuminated under a complex lighting condition, the hard-cast shadow can be easily baked into the albedo. Due to these ambiguities, the optimization process of previous methods tend to gets trapped in local optimal, leading to the artifacts on the recovered material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we seek to use data-driven priors for ambiguity-free inverse rendering. Our key idea is learning the distribution of albedo and specular shading with generative models, which is then used to regularize the optimization process. Inspired by the recent progress of generative models [49], we exploit the conditional diffusion model for modeling the material prior. Our choice of material prior for albedo and specular shading is motivated by the fact that the general rendering equation can be separated into diffuse and specular terms, according to the Disney BRDF model, which thus has two advantages. First, we can construct training data on albedo and specular using a large amount of 3D objects [17] that contain various BRDF models. Second, our designed prior can" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "Chen. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "work as a versatile tool to resolve the ambiguity in the inverse rendering process of many material representations in theory." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 224 + ], + "type": "text", + "content": "A challenge of applying diffusion models to regularize the optimization process is that such generative model struggles to perform multi-view consistent constraints. To address this problem, we propose a coarse-to-fine optimization scheme. Our approach first leverages the diffusion prior to recover the coarse object material and environment lighting. Then, the coarse albedo and specular shading are used to guide the diffusion model to produce constraints with better multi-view consistency, leading to more accurate material and lighting recovery." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 481, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 481, + 284 + ], + "type": "text", + "content": "For training our model, we create a dataset consisting of RGB, albedo, and specular images based on the Objverse dataset [17]. To validate the effective of the proposed approach, we conduct extensive experiments on synthetic and real-world data. Our method achieves state-of-the-art performance on material recovery and relighting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 299, + 237, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 299, + 237, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 299, + 237, + 312 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 322, + 482, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 634 + ], + "type": "text", + "content": "Learning priors for inverse rendering. Earlier approaches for inverse rendering usually require controlled set-ups [16,30] and strong domain-specific assumptions to reconstruct scene properties from images [74]. With the advance of deep learning, the intrinsic properties of images can be directly learned from large-scale synthetic datasets [32, 33, 54]. Shi et al. [54] leverages ShapeNet [9] and introduces a dataset with image intrinsics to decompose the image into base color and shading. Li et al. [31] introduces a large-scale scene dataset and lifts the model's capability from object level to scene level. Subsequent works [32, 35, 53] have attempted to further recover the BRDF and geometry of a scene. However, obtaining massive high-quality real-world images with ground truth intrinsic decomposition is a significant challenge, which limits models' generalization to real-world data. To mitigate this issue, unsupervised methods [38, 64, 72, 72, 73] and weakly supervised methods [71] seek specially designed training scheme to reduce dependency on labeled data. Recently, [83] proposed a large-scale photorealistic indoor scene dataset InteriorVerse for single-view inverse rendering. Kocsis et al. [28] further uses InteriorVerse [83] to train a diffusion model for single-view material recovery in indoor scenes, boosting the generalization to real-world images. Recent methods [39, 46, 62] also explore diffusion models for texture-related tasks. unidream [39] and Richdreamer [46] leverage diffusion model on albedo generation for reconfigurable 3D mesh generation. ControlMat [62] focuses on capturing tileable SVBRDF from a single image with relatively simple shapes (e.g. a plane) and lighting. Although the above learning-based methods achieve promising progress in inverse rendering, their single-view inference pipeline can't guarantee multi-view consistency. In contrast, our method combines the monocular diffusion models with the multiview physically based rendering and thus achieves accurate material and illumination estimation with multiview consistency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Inverse rendering with neural representation. Inspired by the implicit representations of NeRF [42], various works [7,24,37,78,80] extend the neural representation" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "type": "text", + "content": "Intrinsic Anything" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "tions to factorize a scene into material, geometry, and lighting. NeRV [56] leverages a neural visibility field to reduce the computational cost and models indirect lighting with known environment lighting conditions. InvRender [81] and NeILF series [68,76] further model indirect illumination with a neural representation at unknown light conditions. NvdiffrecMC [20] and NeFII [65] incorporate Monte Carlo ray tracing to recover high-frequency details of lighting. [57] adopts a coarse-to-fine manner to retrieve the materials and illumination of a scene. More recently, GS-IR [34] and GIR [55] leverage Gaussian Splatting [27] to model indirect lighting for inverse Rendering. Furthermore, [66, 69, 82] attempt to scale up the neural representations for solving inverse rendering from object to indoor scene. However, due to the inherited ill-posed feature of the inverse rendering problem, these methods still suffer from disentangling the materials and illumination properties from images. In contrast, by learning a strong material and shading prior, we effectively tackled the ill-posed feature and received a significant improvement in decomposing the image into BRDFs and illuminations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 307, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 482, + 439 + ], + "type": "text", + "content": "Diffusion models. Recently, diffusion models have demonstrated unprecedented performance on generation tasks, such as 2D generation [15, 21, 49, 52] and 3D generation [3, 11, 12, 36, 45, 59], indicating the impressive capability of learning the distribution of target data. Inspired this, some recent methods [14, 26] have attempted to exploit the diffusion prior for ill-posed inverse problems, including image denoising [67, 84], image completion [23, 51], and single-view reconstruction [18, 58]. Diffusion posterior illumination [40] learns the distribution of environment illumination with a diffusion model for material recovery. In contrast, our method learns the material prior for supervision and design it as conditional diffusion models of albedo and specular shading, enabling the direct supervision on material estimation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 456, + 226, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 226, + 470 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 226, + 470 + ], + "type": "text", + "content": "3 Background" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 482, + 302, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 302, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 302, + 495 + ], + "type": "text", + "content": "3.1 Physically-Based Rendering" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 502, + 482, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 535 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 535 + ], + "type": "text", + "content": "We adopt the physically based rendering [25] to recover the object material and lighting from images. For a 3D point " + }, + { + "bbox": [ + 130, + 502, + 482, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 502, + 482, + 535 + ], + "type": "text", + "content": ", its outgoing light " + }, + { + "bbox": [ + 130, + 502, + 482, + 535 + ], + "type": "inline_equation", + "content": "L_{o}" + }, + { + "bbox": [ + 130, + 502, + 482, + 535 + ], + "type": "text", + "content": " at direction " + }, + { + "bbox": [ + 130, + 502, + 482, + 535 + ], + "type": "inline_equation", + "content": "\\omega_{o}" + }, + { + "bbox": [ + 130, + 502, + 482, + 535 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 209, + 536, + 481, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 536, + 481, + 559 + ], + "spans": [ + { + "bbox": [ + 209, + 536, + 481, + 559 + ], + "type": "interline_equation", + "content": "L _ {o} \\left(\\omega_ {o}; \\mathbf {x}\\right) = \\int_ {\\Omega} L _ {i} \\left(\\omega_ {i}\\right) f _ {r} \\left(\\mathbf {x}; \\omega_ {o}, \\omega_ {i}\\right) \\left(\\omega_ {i} \\cdot \\mathbf {n}\\right) d \\omega_ {i}, \\tag {1}", + "image_path": "fbd69c37ee8b0440e87f5e4306a90f129f7e401fec2de7de89ec70ffbe52ec49.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 563, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 563, + 482, + 601 + ], + "type": "inline_equation", + "content": "L_{i}(\\omega_{i})" + }, + { + "bbox": [ + 130, + 563, + 482, + 601 + ], + "type": "text", + "content": " is the incident lighting along the direction " + }, + { + "bbox": [ + 130, + 563, + 482, + 601 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 130, + 563, + 482, + 601 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 563, + 482, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 563, + 482, + 601 + ], + "type": "text", + "content": " is surface normal at " + }, + { + "bbox": [ + 130, + 563, + 482, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 563, + 482, + 601 + ], + "type": "text", + "content": ". In practice, we define the lighting " + }, + { + "bbox": [ + 130, + 563, + 482, + 601 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 130, + 563, + 482, + 601 + ], + "type": "text", + "content": " as an estimable longitude-latitude environment map. We use the simplified Disney BRDF model [8] for " + }, + { + "bbox": [ + 130, + 563, + 482, + 601 + ], + "type": "inline_equation", + "content": "f_{r}" + }, + { + "bbox": [ + 130, + 563, + 482, + 601 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 224, + 609, + 481, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 609, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 224, + 609, + 481, + 634 + ], + "type": "interline_equation", + "content": "f _ {r} (\\mathbf {x}; \\omega_ {o}, \\omega_ {i}) = \\frac {\\mathbf {k} _ {d}}{\\pi} + \\frac {D \\cdot G \\cdot F}{4 (\\mathbf {n} \\cdot \\omega_ {\\mathbf {o}}) (\\mathbf {n} \\cdot \\omega_ {i})}, \\tag {2}", + "image_path": "628b0e28dce1e0cd249734770ac9c7c392f2d3fb1a2d220d1feaf095c94f4bd4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "D, G" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " are the normal distribution, geometry attenuation, and Fresnel effect, respectively. The albedo " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_d" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " and roughness " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " parameters are set as" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Chen. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "learnable, and other parameters are set as constant. The detailed formulation can be found in supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 177 + ], + "type": "text", + "content": "The RGB image is rendered by evaluating Eq. (1) with the environment lighting, albedo image, roughness image, and scene mesh. We solve the integral in Eq. (1) using Monte-Carlo sampling with Multiple Importance Sampling [61]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 193, + 411, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 193, + 411, + 205 + ], + "spans": [ + { + "bbox": [ + 131, + 193, + 411, + 205 + ], + "type": "text", + "content": "3.2 Diffusion Model and Diffusion Posterior Sampling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 212, + 479, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 479, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 479, + 248 + ], + "type": "text", + "content": "The diffusion model is a probabilistic generative model that generates images from pure Gaussian noise through a progressive de-noising process [21]. Each step of the denoising process from " + }, + { + "bbox": [ + 130, + 212, + 479, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 130, + 212, + 479, + 248 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 212, + 479, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t - 1}" + }, + { + "bbox": [ + 130, + 212, + 479, + 248 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 205, + 256, + 480, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 256, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 205, + 256, + 480, + 283 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha (t)}} \\left(\\mathbf {x} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\hat {\\alpha} _ {t}}} \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t)\\right) + \\sigma_ {t} \\mathbf {z}, \\tag {3}", + "image_path": "868a0b0dca186fd99b141c04510e19c436e7660b09d53f7959b7a03c68d117ab.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 290, + 479, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 479, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 479, + 326 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 290, + 479, + 326 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 290, + 479, + 326 + ], + "type": "text", + "content": " is the learnable noise predictor, " + }, + { + "bbox": [ + 130, + 290, + 479, + 326 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 130, + 290, + 479, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 290, + 479, + 326 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 130, + 290, + 479, + 326 + ], + "type": "text", + "content": " are predefined constant values, and " + }, + { + "bbox": [ + 130, + 290, + 479, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 290, + 479, + 326 + ], + "type": "text", + "content": " is a random value sampled from a standard Gaussian distribution. The noise predictor is typically implemented as a UNet [50]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 326, + 479, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 479, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 479, + 361 + ], + "type": "text", + "content": "The diffusion posterior sampling (DPS) [13] is a technique for solving inverse problems (e.g. image deblur) using the diffusion model. Specifically, given an observation " + }, + { + "bbox": [ + 130, + 326, + 479, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 326, + 479, + 361 + ], + "type": "text", + "content": " derived from " + }, + { + "bbox": [ + 130, + 326, + 479, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 326, + 479, + 361 + ], + "type": "text", + "content": " which defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 274, + 371, + 480, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 371, + 480, + 383 + ], + "spans": [ + { + "bbox": [ + 274, + 371, + 480, + 383 + ], + "type": "interline_equation", + "content": "\\mathbf {y} = \\mathcal {A} \\left(\\mathbf {x} _ {0}\\right) + \\mathbf {n} \\tag {4}", + "image_path": "883f2efd88d3fb263caf53dd0eb30073c1bedc56ce08b8a0d02d6b3140be32f8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 391, + 480, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 480, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 480, + 451 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 391, + 480, + 451 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 391, + 480, + 451 + ], + "type": "text", + "content": " is the measurement operator, " + }, + { + "bbox": [ + 130, + 391, + 480, + 451 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 391, + 480, + 451 + ], + "type": "text", + "content": " is the measurement noise. The goal is to estimate " + }, + { + "bbox": [ + 130, + 391, + 480, + 451 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 391, + 480, + 451 + ], + "type": "text", + "content": " from observation " + }, + { + "bbox": [ + 130, + 391, + 480, + 451 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 391, + 480, + 451 + ], + "type": "text", + "content": " using posterior distribution " + }, + { + "bbox": [ + 130, + 391, + 480, + 451 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0|\\mathbf{y})" + }, + { + "bbox": [ + 130, + 391, + 480, + 451 + ], + "type": "text", + "content": ". To sample from " + }, + { + "bbox": [ + 130, + 391, + 480, + 451 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0|\\mathbf{y})" + }, + { + "bbox": [ + 130, + 391, + 480, + 451 + ], + "type": "text", + "content": " with a pre-trained distribution " + }, + { + "bbox": [ + 130, + 391, + 480, + 451 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0)" + }, + { + "bbox": [ + 130, + 391, + 480, + 451 + ], + "type": "text", + "content": ", the DPS sampling approximates the posterior sampling functions by adding a guiding term to the denoising score " + }, + { + "bbox": [ + 130, + 391, + 480, + 451 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t,t)" + }, + { + "bbox": [ + 130, + 391, + 480, + 451 + ], + "type": "text", + "content": " in Eq. (3):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 214, + 459, + 480, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 459, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 214, + 459, + 480, + 472 + ], + "type": "interline_equation", + "content": "\\tilde {\\epsilon} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right) = \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right) + \\zeta_ {t} \\nabla_ {\\mathbf {x} _ {t}} \\left| \\left| \\mathcal {A} \\left(\\mathbf {x} _ {t} ^ {0}\\right) - \\mathbf {y} \\right| \\right| _ {2} ^ {2}, \\tag {5}", + "image_path": "2d025003ef5f5c73dc2084f951e2c72edf78acc35bd1e3d86770753c365d569c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 479, + 480, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 480, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 480, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 479, + 480, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^0" + }, + { + "bbox": [ + 130, + 479, + 480, + 540 + ], + "type": "text", + "content": " is the posterior estimation of " + }, + { + "bbox": [ + 130, + 479, + 480, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 479, + 480, + 540 + ], + "type": "text", + "content": " at timestamp " + }, + { + "bbox": [ + 130, + 479, + 480, + 540 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 479, + 480, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 479, + 480, + 540 + ], + "type": "inline_equation", + "content": "\\zeta_t" + }, + { + "bbox": [ + 130, + 479, + 480, + 540 + ], + "type": "text", + "content": " is the step size. This formula assumes a known " + }, + { + "bbox": [ + 130, + 479, + 480, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 479, + 480, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 479, + 480, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 479, + 480, + 540 + ], + "type": "text", + "content": " follows a Gaussian distribution of " + }, + { + "bbox": [ + 130, + 479, + 480, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\sigma^2\\mathbf{I})" + }, + { + "bbox": [ + 130, + 479, + 480, + 540 + ], + "type": "text", + "content": ". We refer to DPS [13] for more details. In Section 4, we will demonstrate how to model the material priors using the diffusion model and use DPS sampling to generate multi-view consistent material samples." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 557, + 202, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 557, + 202, + 570 + ], + "spans": [ + { + "bbox": [ + 131, + 557, + 202, + 570 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": "Our method aims at recovering material and lighting parameters from multi-view images under a single unknown illumination. We define object material as albedo and roughness without considering the metallic attribute and use a lat-long environment map for lighting. Due to the ill-posed nature of inverse rendering [20, 81], ambiguities exist in the optimization process, leading to inaccurate material estimation. We propose using data-driven priors to constrain the optimization to eliminate these ambiguities." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "type": "text", + "content": "Intrinsic Anything" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 125, + 343, + 255 + ], + "blocks": [ + { + "bbox": [ + 137, + 125, + 343, + 255 + ], + "lines": [ + { + "bbox": [ + 137, + 125, + 343, + 255 + ], + "spans": [ + { + "bbox": [ + 137, + 125, + 343, + 255 + ], + "type": "image", + "image_path": "4243a67edf082b2aa8a26b4b13b784a6f06ed88b91202f3d585c247a089d92e2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 265, + 482, + 354 + ], + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 354 + ], + "type": "text", + "content": "Fig. 2: Overview of our pipeline. (a) Based on physically based rendering, our model combines geometry, lighting, roughness, and albedo into RGB and specular images, and optimizes the lighting and materials in a two-stage manner. In the first stage, our model is supervised by images and diffusion priors to output coarse albedo and roughness. Subsequently, the coarse materials are used to guide diffusion models to provide more multi-view consistent constraints. (b) The guided sampling first calculates the L2 loss between the guidance and one-step denoised signals, and then adds the gradient of the L1 loss to the output of the noise predictor." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 347, + 125, + 477, + 255 + ], + "blocks": [ + { + "bbox": [ + 347, + 125, + 477, + 255 + ], + "lines": [ + { + "bbox": [ + 347, + 125, + 477, + 255 + ], + "spans": [ + { + "bbox": [ + 347, + 125, + 477, + 255 + ], + "type": "image", + "image_path": "b28fce0c0667b4d0ae45e1416d86fe95e7c04dfb24c132185a06ce7e608ee31f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 379, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 489 + ], + "type": "text", + "content": "Inspired by the recent success of generative models [49] on image generation, we design two material priors modeled by the conditional diffusion model (Sec. 4.1). To recover object materials and lighting, we first employ monosdf [75] to reconstruct a 3D mesh from input images. The materials are defined as learnable texture maps in UV space from the UV-mapped mesh using Blender [6], and lighting as estimizable longitude-latitude environment map. Then we apply our coarse-to-fine optimization process (Sec. 4.2) to recover accurate materials and lighting, while the geometry and the prior models are fixed. An overview of our optimization process is shown in Fig 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 521, + 337, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 337, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 337, + 533 + ], + "type": "text", + "content": "4.1 Albedo and Specular Prior Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 544, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 628 + ], + "type": "text", + "content": "Design of the material priors. To better decompose materials and lighting, we propose to use data-driven priors to constrain the optimization process. Ideally, we can directly apply prior distributions of albedo and roughness to narrow down the solution space and effectively alleviate ambiguities. However, roughness, being a concealed characteristic of a material, is difficult to observe and is also hard to acquire on most large-scale synthetic datasets [17]. Because these objects have a variety of material parameters that do not necessarily have a valid roughness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Fortunately, image rendering using the Disney BRDF model [8] is composed of specular shading and diffuse shading regardless of the choice of material parameters, as shown in Eq.(2). Thus, the rendering equation can be rewritten to:" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Chen. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 144, + 137, + 480, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 137, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 144, + 137, + 480, + 190 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {o} (\\omega_ {o}; \\mathbf {x}) = \\frac {\\mathbf {k} _ {d}}{\\pi} \\int_ {\\Omega} L _ {i} (\\omega_ {i}) (\\omega_ {i} \\cdot \\mathbf {n}) d \\omega_ {i} + \\int_ {\\Omega} L _ {i} (\\omega_ {i}) \\frac {D \\cdot G \\cdot F}{4 (\\mathbf {n} \\cdot \\omega_ {\\mathbf {o}}) (\\mathbf {n} \\cdot \\omega_ {i})} (\\omega_ {i} \\cdot \\mathbf {n}) d \\omega_ {i} \\\\ = S _ {\\text {d i f f}} (\\mathbf {x}, L _ {i}, k _ {d}) + S _ {\\text {s p e c}} \\left(\\omega_ {o}; \\mathbf {x}, \\rho , L _ {i}\\right). \\tag {6} \\\\ \\end{array}", + "image_path": "255832bbdbbbfc94c401118a7d0509bebb51f3f93f791e4b02c2d9de8c08e818.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 201, + 479, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 201, + 479, + 262 + ], + "spans": [ + { + "bbox": [ + 131, + 201, + 479, + 262 + ], + "type": "text", + "content": "Motivated by Eq. (6), we propose to apply priors on specular shading which exists in all physically based renderings, instead of directly using priors on roughness. Specular shading is closely linked to roughness and has the advantage of being more observable. Moreover, based on any synthetic object, we can create the ground-truth specular shading for training the prior model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 278, + 479, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 278, + 479, + 350 + ], + "spans": [ + { + "bbox": [ + 131, + 278, + 479, + 350 + ], + "type": "text", + "content": "Modeling the material priors. Previous methods [31, 35, 54, 70, 72, 73] generally model the prior of object material using a deterministic prediction network, which attempts to regress the material from a single image. In contrast, we leverage generative models to capture the distribution of albedo and specular, considering that there could be multiple solutions for the underlying object material given captured images [4]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 352, + 480, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 352, + 480, + 448 + ], + "spans": [ + { + "bbox": [ + 131, + 352, + 480, + 448 + ], + "type": "text", + "content": "Given an RGB image " + }, + { + "bbox": [ + 131, + 352, + 480, + 448 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 131, + 352, + 480, + 448 + ], + "type": "text", + "content": ", we can model the conditional distribution of target albedo " + }, + { + "bbox": [ + 131, + 352, + 480, + 448 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{k}_d|I)" + }, + { + "bbox": [ + 131, + 352, + 480, + 448 + ], + "type": "text", + "content": " and specular shading " + }, + { + "bbox": [ + 131, + 352, + 480, + 448 + ], + "type": "inline_equation", + "content": "p_{\\theta}(S_{\\mathrm{spec}}|I)" + }, + { + "bbox": [ + 131, + 352, + 480, + 448 + ], + "type": "text", + "content": " by modifying the noise predictor to accept " + }, + { + "bbox": [ + 131, + 352, + 480, + 448 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 131, + 352, + 480, + 448 + ], + "type": "text", + "content": " as its conditioning. Denote the modified noise predictor as " + }, + { + "bbox": [ + 131, + 352, + 480, + 448 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t,t;I)" + }, + { + "bbox": [ + 131, + 352, + 480, + 448 + ], + "type": "text", + "content": ". Specifically, we first use the CLIP image encoder [47] to extract a feature vector from the conditioning image. Then, to inject the conditioning signal into the UNet, we utilize a transformer network [60] to perform the cross-attention between the image feature vector and the intermediate feature maps of the UNet. More details can be found in the supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 449, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 449, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 480, + 521 + ], + "type": "text", + "content": "The diffusion model can model complex distributions in a step-wise denoising manner, which is crucial for albedo and specular shading recovery due to the complexity of lighting and material in real-world data. Moreover, as a generative model, the diffusion model facilitates a guided sampling procedure [13, 22, 41], which permits us to steer the diffusion model toward producing samples that align with multi-view image observations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 544, + 348, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 348, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 348, + 556 + ], + "type": "text", + "content": "4.2 Material Prior for Inverse Rendering" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 570, + 479, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 570, + 479, + 654 + ], + "spans": [ + { + "bbox": [ + 131, + 570, + 479, + 654 + ], + "type": "text", + "content": "Based on the albedo and specular shading prior models, we design a two-stage optimization process for inverse rendering, which effectively alleviates the ambiguities. An overview of our method is presented in Fig. 2. We first use the conditional diffusion model to regularize the inverse rendering process, which outputs an initialization of lighting, albedo, and roughness. Then, the coarse material is used to guide the sampling of the diffusion model, enabling it to produce more deterministic results and better stabilize the optimization process." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "type": "text", + "content": "Intrinsic Anything" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 173, + 153 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 173, + 153 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 173, + 153 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 173, + 153 + ], + "type": "image", + "image_path": "59d7a59039138aac4981f62a589e84b9dcbf8120f252738062a6b2aa48027378.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 154, + 162, + 159 + ], + "lines": [ + { + "bbox": [ + 140, + 154, + 162, + 159 + ], + "spans": [ + { + "bbox": [ + 140, + 154, + 162, + 159 + ], + "type": "text", + "content": "Input View" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 178, + 117, + 206, + 153 + ], + "blocks": [ + { + "bbox": [ + 178, + 117, + 206, + 153 + ], + "lines": [ + { + "bbox": [ + 178, + 117, + 206, + 153 + ], + "spans": [ + { + "bbox": [ + 178, + 117, + 206, + 153 + ], + "type": "image", + "image_path": "ac7f5d1df27f44e3d3328745738a092d83775572510de0e6c2dcaa742371af6d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 154, + 200, + 159 + ], + "lines": [ + { + "bbox": [ + 182, + 154, + 200, + 159 + ], + "spans": [ + { + "bbox": [ + 182, + 154, + 200, + 159 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 215, + 119, + 244, + 152 + ], + "blocks": [ + { + "bbox": [ + 215, + 119, + 244, + 152 + ], + "lines": [ + { + "bbox": [ + 215, + 119, + 244, + 152 + ], + "spans": [ + { + "bbox": [ + 215, + 119, + 244, + 152 + ], + "type": "image", + "image_path": "f4e8fb00a4f30a28f0048186c15868b78f96fad854cfb572201d657f755855bb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 168, + 481, + 213 + ], + "lines": [ + { + "bbox": [ + 130, + 168, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 168, + 481, + 213 + ], + "type": "text", + "content": "Fig. 3: Effect of the guided sampling. We visualize the samples generated from the albedo prior model of different viewpoints. Without the guided sampling, the materials are inconsistent across multi-views and do not align with the material decomposition from the observed lighting." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 248, + 120, + 277, + 152 + ], + "blocks": [ + { + "bbox": [ + 248, + 120, + 277, + 152 + ], + "lines": [ + { + "bbox": [ + 248, + 120, + 277, + 152 + ], + "spans": [ + { + "bbox": [ + 248, + 120, + 277, + 152 + ], + "type": "image", + "image_path": "89fc32e2c9da09fa1c50c9e9e90f1eca124e78f1f8965a2c9da563099c557f36.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 154, + 290, + 159 + ], + "lines": [ + { + "bbox": [ + 263, + 154, + 290, + 159 + ], + "spans": [ + { + "bbox": [ + 263, + 154, + 290, + 159 + ], + "type": "text", + "content": "w/o guidance" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 280, + 121, + 310, + 152 + ], + "blocks": [ + { + "bbox": [ + 280, + 121, + 310, + 152 + ], + "lines": [ + { + "bbox": [ + 280, + 121, + 310, + 152 + ], + "spans": [ + { + "bbox": [ + 280, + 121, + 310, + 152 + ], + "type": "image", + "image_path": "582351d10cf2eadc149d0d83865d5ed324a2e39e342029464234aca8308f872f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 119, + 342, + 152 + ], + "blocks": [ + { + "bbox": [ + 316, + 119, + 342, + 152 + ], + "lines": [ + { + "bbox": [ + 316, + 119, + 342, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 342, + 152 + ], + "type": "image", + "image_path": "55281b6ce5962942555c63cf7f320e7478a94cae361ce5a4f84e2e2d57fe2883.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 345, + 119, + 375, + 152 + ], + "blocks": [ + { + "bbox": [ + 345, + 119, + 375, + 152 + ], + "lines": [ + { + "bbox": [ + 345, + 119, + 375, + 152 + ], + "spans": [ + { + "bbox": [ + 345, + 119, + 375, + 152 + ], + "type": "image", + "image_path": "8874c50f029fad77faa398a7bbda971d2cd32d06d605771a323fb763567ca7f0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 379, + 120, + 410, + 152 + ], + "blocks": [ + { + "bbox": [ + 379, + 120, + 410, + 152 + ], + "lines": [ + { + "bbox": [ + 379, + 120, + 410, + 152 + ], + "spans": [ + { + "bbox": [ + 379, + 120, + 410, + 152 + ], + "type": "image", + "image_path": "1e95969c6d881cb64bbf42cb823259e8cad71c83fbcfb90cdddd20292da13c75.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 153, + 422, + 159 + ], + "lines": [ + { + "bbox": [ + 398, + 153, + 422, + 159 + ], + "spans": [ + { + "bbox": [ + 398, + 153, + 422, + 159 + ], + "type": "text", + "content": "w/guidance" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 413, + 120, + 443, + 152 + ], + "blocks": [ + { + "bbox": [ + 413, + 120, + 443, + 152 + ], + "lines": [ + { + "bbox": [ + 413, + 120, + 443, + 152 + ], + "spans": [ + { + "bbox": [ + 413, + 120, + 443, + 152 + ], + "type": "image", + "image_path": "3a04067962f9c1422c0fc14161d1c0fd7c60639c54165224b9df153397a38505.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 447, + 119, + 475, + 152 + ], + "blocks": [ + { + "bbox": [ + 447, + 119, + 475, + 152 + ], + "lines": [ + { + "bbox": [ + 447, + 119, + 475, + 152 + ], + "spans": [ + { + "bbox": [ + 447, + 119, + 475, + 152 + ], + "type": "image", + "image_path": "7adbb3cda7d53baeb6ec15651f8a2126c9ea3ee3a22b7c28fee18d3fdf959016.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 236, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 480, + 319 + ], + "type": "text", + "content": "Training with diffusion model. Given input multi-view images of the target object, we utilize the conditional diffusion model to predict their albedo and specular shading components, which are used to supervise the model parameters. Due to the probabilistic nature of the diffusion model, the generated albedo and specular images are inconsistent across camera views in terms of intensity. To overcome this problem, we apply shift and scale invariant loss [48] to compute the difference between the predicted and generated material." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 146, + 319, + 398, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 319, + 398, + 331 + ], + "spans": [ + { + "bbox": [ + 146, + 319, + 398, + 331 + ], + "type": "text", + "content": "In the first stage, the training loss function is defined as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 210, + 340, + 480, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 340, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 210, + 340, + 480, + 372 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {c o a r s e}} = \\lambda_ {1} \\| \\hat {I} - I \\| + \\lambda_ {2} \\| \\operatorname {S S I} \\left(\\hat {\\mathbf {k}} _ {d}, \\mathbf {k} _ {d} ^ {\\mathrm {s}}\\right) - \\mathbf {k} _ {d} ^ {\\mathrm {s}} \\| \\tag {7} \\\\ + \\lambda_ {3} | | \\hat {S} _ {\\text {s p e c}} - S _ {\\text {s p e c}} ^ {\\mathrm {s}} | | + \\lambda_ {4} \\mathcal {L} _ {\\text {s m o o t h}}, \\\\ \\end{array}", + "image_path": "220954ab9107e491ab74109f0d43484e6007ad46b86c2dbf1ef4563746269c73.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathrm{SSI}(a,b)" + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "text", + "content": " aligns " + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "text", + "content": " to the scale and mean of " + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\hat{k}_d" + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\hat{S}_{\\mathrm{spec}}" + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "text", + "content": " are the image, albedo, and specular shading to be optimized respectively, " + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_d^{\\mathrm{s}}" + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{spec}}^{\\mathrm{s}}" + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "text", + "content": " are samples from our diffusion models " + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "inline_equation", + "content": "p_\\theta (\\mathbf{k}_d|I)" + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "inline_equation", + "content": "p_{\\theta}(S_{\\mathrm{spec}}|I)" + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "text", + "content": " respectively. We also use a smoothness constraint " + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{smooth}}" + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "text", + "content": " on the optimized materials. " + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\lambda_1" + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\lambda_2" + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\lambda_3" + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\lambda_4" + }, + { + "bbox": [ + 130, + 381, + 482, + 443 + ], + "type": "text", + "content": " are hyper-parameters. More details can be found in the supplementary material." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 453, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 586 + ], + "type": "text", + "content": "Training with guided diffusion samples. After the training in the first stage, we obtain an initialization of the object material and the environment lighting. To better guide the optimization process, we aimed to generate samples from the diffusion prior models that are consistent with multiple viewpoints, aligned with observed lighting conditions, and still follow the distribution of priors. To this end, we leverage the Diffusion Posterior Sampling (DPS) [13] to guide the sampling process, and use the estimated albedo " + }, + { + "bbox": [ + 130, + 453, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{k}}_d" + }, + { + "bbox": [ + 130, + 453, + 482, + 586 + ], + "type": "text", + "content": " and specular shading " + }, + { + "bbox": [ + 130, + 453, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\hat{S}_{\\mathrm{spec}}" + }, + { + "bbox": [ + 130, + 453, + 482, + 586 + ], + "type": "text", + "content": " as guiding signals to re-generate samples. As shown in Fig. 3, samples after guidance are multi-view consistent and fit observations under the current lighting condition. Specifically, we use the score function of DPS in Eq. (5) with " + }, + { + "bbox": [ + 130, + 453, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 453, + 482, + 586 + ], + "type": "text", + "content": " as the identity operator to re-generate albedo " + }, + { + "bbox": [ + 130, + 453, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{k}}_d" + }, + { + "bbox": [ + 130, + 453, + 482, + 586 + ], + "type": "text", + "content": " and specular " + }, + { + "bbox": [ + 130, + 453, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\tilde{S}_{\\mathrm{spec}}" + }, + { + "bbox": [ + 130, + 453, + 482, + 586 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 210, + 594, + 480, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 594, + 480, + 608 + ], + "spans": [ + { + "bbox": [ + 210, + 594, + 480, + 608 + ], + "type": "interline_equation", + "content": "\\tilde {\\epsilon} _ {\\theta} \\left(\\tilde {\\mathbf {x}} _ {t}, t; I\\right) = \\epsilon_ {\\theta} \\left(\\tilde {\\mathbf {x}} _ {t}, t; I\\right) + \\gamma \\bigtriangledown_ {\\tilde {\\mathbf {x}} _ {t}} \\left| \\left| \\tilde {\\mathbf {x}} _ {t} ^ {0} - \\hat {\\mathbf {x}} \\right| \\right| _ {2}, \\tag {8}", + "image_path": "64959387eba5d1fe62fab58ad8aaf1727c40c121c71c7f9233648dde17153318.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_t" + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_t^0" + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "text", + "content": " is the noisy latents and the posterior estimation of " + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{k}}_d" + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\tilde{S}_{\\mathrm{spec}}" + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "text", + "content": " on step " + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{k}}_d" + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{S}_{\\mathrm{spec}}" + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "text", + "content": ". We choose the step size following [13] as " + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\zeta_i = \\gamma / ||\\tilde{\\mathbf{x}}_t^0 - \\hat{\\mathbf{x}}||" + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "text", + "content": " is a pre-defined scalar. Finally, we use the re-generated multiview-consistent material " + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{k}}_d" + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\tilde{S}_{\\mathrm{spec}}" + }, + { + "bbox": [ + 130, + 615, + 482, + 667 + ], + "type": "text", + "content": " from the DPS to supervise the lighting" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "Chen. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "and materials using Eq. (7). Experiments show that the second stage gives an accurate decomposition of the object material and environment lighting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": "An alternative way to supervise the inverse rendering with diffusion models is the Score Distillation Sampling (SDS) [45], which achieves consistency by finding a mode in the distribution through optimization of the material field. However, our experiments find that the SDS guidance usually produces blurry results. A possible reason is that the SDS guidance relies on noise prediction that has a high variance, which leads to unstable convergence as discussed in [63]. In constraint, our pipeline uses guided image samples that are consistent with low variance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 241, + 230, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 241, + 230, + 255 + ], + "spans": [ + { + "bbox": [ + 131, + 241, + 230, + 255 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 266, + 278, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 266, + 278, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 266, + 278, + 277 + ], + "type": "text", + "content": "5.1 Implementation details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 285, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 392 + ], + "type": "text", + "content": "Prior Model Training. We create albedo and specular training data based on Objaverse dataset [17]. We randomly select 350k objects, excluding those with pure black or white albedos, and gather 180 HDR environment maps from [1] for rendering. Each object is rendered under 10 random poses with one random HDR environment map. We use a diffusion model fine-tuned from stable-diffusion [49] for the image variation task [2] as a starting point and continue to train 100k iterations for the albedo prior and 350K iterations for the specular shading prior. The training process for the two prior models takes around 10 days on 8 NVIDIA A100 (80GB) GPU." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 403, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 477 + ], + "type": "text", + "content": "Inverse Rendering. The resolutions for texture maps and environment are " + }, + { + "bbox": [ + 130, + 403, + 482, + 477 + ], + "type": "inline_equation", + "content": "2048 \\times 2048" + }, + { + "bbox": [ + 130, + 403, + 482, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 403, + 482, + 477 + ], + "type": "inline_equation", + "content": "512 \\times 256" + }, + { + "bbox": [ + 130, + 403, + 482, + 477 + ], + "type": "text", + "content": ", respectively. We use Nvdiffrast [29] for rendering images with 144 samples per pixel. We train the first stage of our method for 20k iterations and the second stage for 15k iterations. The setting of hyperparameters is presented in the supplement materials. The inverse rendering for each scene takes around 1.5 hours on a single NVIDIA RTX A6000 GPU." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 491, + 211, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 211, + 503 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 211, + 503 + ], + "type": "text", + "content": "5.2 Evulation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 481, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 559 + ], + "type": "text", + "content": "Dataset. We collect 4 CAD models with a variety of object shapes and materials. Each object is rendered with 120 posed RGBA images using Blender Cycles [6]. We use 100 images for training and 20 for testing. For evaluation, we also render 20 novel test views under novel lights along with their ground truth BRDFs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": "Baselines. We compare our methods with the following two types of baseline methods. (1) Optimization-based methods [24, 43, 76, 81, 81]. InvRender [81] and TensoIR [24] both models indirect illumination with MLPs and uses Spherical Gaussian(SG) as environment light. InvRender [81] uses MLP representations while TensoIR is inspired by tensorial representation [10]. NeILF++ [76] models an incident light field that includes both environemnt light and indirect light with an MLP constrained by the outgoing radiance field. NvDiffRec [43], and NvDiffrecMC [20] extracted explicit topology from neural SDF and directly optimized" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "type": "text", + "content": "Intrinsic Anything" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 190, + 480, + 297 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 1: Quantitative results on the synthetic dataset. \"NvDiffrec*\" and \"NvDiffrecMC*\" mean these methods are with our reconstructed mesh. \"Ours w/o guided\" means our method without the guided sampling strategy. TensorIR and NvDiffrec have strong capabilities of fitting observed images, but they struggle to recover correct object materials. Replacing our diffusion priors with other priors [38, 71] leads to the degradation of the performance on material recovery." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 190, + 480, + 297 + ], + "lines": [ + { + "bbox": [ + 138, + 190, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 190, + 480, + 297 + ], + "type": "table", + "html": "
MethodAlbedoAligned AlbedoView SynthesisRelightingRoughness
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓MSE↓
InvRender [81]17.81810.92470.064523.75920.94910.055927.30510.96730.038624.36320.96210.05150.0420
TensoIR [24]20.29720.95270.063124.93680.95640.061433.12340.98030.026421.34730.95430.05140.0456
NvDiffrec [43]16.54670.92400.070621.53040.95070.057432.57780.98240.019021.99390.95500.07340.0777
NvDiffrec* [43]14.52040.90210.083520.22800.92040.065931.03360.97840.021821.24520.93560.06380.0841
NvDiffrecMC [20]13.04950.88150.106821.67700.94060.069930.10830.96740.044022.53370.94860.07970.1046
NvDiffrecMC* [20]11.13530.87220.117619.83250.93050.080528.16340.95600.044021.03800.93320.08220.1411
NEILF++* [76]13.83360.91720.091518.45430.93660.083129.35270.97750.027620.48780.96290.05900.0855
Yi et al. [71]13.20290.89340.083318.84280.91330.073126.50910.96690.039321.1560.95620.05070.1524
USI3D [38]11.67130.88720.105118.92960.93440.077226.96530.96660.048020.27210.95870.06120.1241
Ours24.67660.95620.050326.88290.95320.037830.24050.97410.032828.97210.97570.02890.0105
", + "image_path": "f5106305c0d541e6ca0dd44c2671893fbac35746c5b054dbf34c02a3e72dfc56.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 308, + 306, + 434 + ], + "blocks": [ + { + "bbox": [ + 133, + 308, + 306, + 434 + ], + "lines": [ + { + "bbox": [ + 133, + 308, + 306, + 434 + ], + "spans": [ + { + "bbox": [ + 133, + 308, + 306, + 434 + ], + "type": "image", + "image_path": "d9a8da1910ec10810d4118932714860b9647890c877d26017fd6c8dbe89b9c86.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 441, + 480, + 464 + ], + "lines": [ + { + "bbox": [ + 130, + 441, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 480, + 464 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparison in terms of relighting on the synthetic dataset. Zoom in for details." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 308, + 480, + 433 + ], + "blocks": [ + { + "bbox": [ + 306, + 308, + 480, + 433 + ], + "lines": [ + { + "bbox": [ + 306, + 308, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 480, + 433 + ], + "type": "image", + "image_path": "7af24bb2ef373a8441084b1cdf00e41b80fc31325de940550238d7a004acc8c1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 619 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 619 + ], + "type": "text", + "content": "the 3D mesh with its PBR materials. Their illumination is expressed as an HDR environment map, which can be optimized by NviDiffRec's [43] split sum environment lighting and NviDiffRec's Monte Carlo ray tracing. (2) Data-driven methods [38, 71]. We keep object and lighting models as the same and just replace diffusion priors with single-view intrinsic image decomposition methods for comparison. USI3D [38] models the decomposition task as domain transfer from the image domain to the albedo or shading domain using generative models. We adopt only the predicted albedo to supervise the optimization because their rendering model is not physically based. Yi. el [71] propose a network to predict and remove specular shading and subsequently recover the albedo. We use the predicted specular shading and albedo as the priors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Evaluation protocol. We analyze our method and the baselines in the following way. On the 4 synthetic datasets, we compare their reconstruction of albedo, roughness, and novel view synthesis quantitatively with the ground truths. Given" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Chen. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 117, + 477, + 297 + ], + "blocks": [ + { + "bbox": [ + 134, + 117, + 477, + 297 + ], + "lines": [ + { + "bbox": [ + 134, + 117, + 477, + 297 + ], + "spans": [ + { + "bbox": [ + 134, + 117, + 477, + 297 + ], + "type": "image", + "image_path": "c0af9d936d35d0b0a761b2df22657ff88a8a34d1af29b93168410777dbe62975.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 304, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 361 + ], + "type": "text", + "content": "Fig. 5: Quantitative comparison with baselines on the synthetic dataset. Our method estimates more accurate albedo and roughness than other methods. Nvdiffrec [43] and NvdiffrecMC [20] tend to bake the lighting into the albedo, due to the ambiguity between the diffuse shading and albedo. InvRender [81] and TensoIR [24] struggle to decouple the roughness from the albedo and environment lighting." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 132, + 369, + 482, + 589 + ], + "blocks": [ + { + "bbox": [ + 132, + 369, + 482, + 589 + ], + "lines": [ + { + "bbox": [ + 132, + 369, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 482, + 589 + ], + "type": "image", + "image_path": "76407c55ede19a6e8c32707f3c943b1d6d29f5e4e3562b7c0e4eb052fc326e2c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 597, + 482, + 632 + ], + "lines": [ + { + "bbox": [ + 130, + 597, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 482, + 632 + ], + "type": "text", + "content": "Fig.6: Qualitative results on DTU. Our method recovers reasonable materials and relighting on real datasets. Baseline models can not disentangle illumination and materials which results in the baked-in effect on materials estimation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "type": "text", + "content": "IntrinsicAnything" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 143, + 125, + 156, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 125, + 156, + 134 + ], + "spans": [ + { + "bbox": [ + 143, + 125, + 156, + 134 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 165, + 118, + 203, + 148 + ], + "blocks": [ + { + "bbox": [ + 165, + 118, + 203, + 148 + ], + "lines": [ + { + "bbox": [ + 165, + 118, + 203, + 148 + ], + "spans": [ + { + "bbox": [ + 165, + 118, + 203, + 148 + ], + "type": "image", + "image_path": "b8fab22c6ae49b02c1fb1b66808fd826077a329c35f4ce7e485ff5e8eedb74af.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 205, + 118, + 240, + 148 + ], + "blocks": [ + { + "bbox": [ + 205, + 118, + 240, + 148 + ], + "lines": [ + { + "bbox": [ + 205, + 118, + 240, + 148 + ], + "spans": [ + { + "bbox": [ + 205, + 118, + 240, + 148 + ], + "type": "image", + "image_path": "95031cf721cb4b9032ed88964cc0a3d204cf15d86ef154e9160bcc258482c4a4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 242, + 118, + 280, + 148 + ], + "blocks": [ + { + "bbox": [ + 242, + 118, + 280, + 148 + ], + "lines": [ + { + "bbox": [ + 242, + 118, + 280, + 148 + ], + "spans": [ + { + "bbox": [ + 242, + 118, + 280, + 148 + ], + "type": "image", + "image_path": "41e69272b99b23cc41bd26c8e753793d23db5ec795ad43fd1567ea02774ea5fe.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 282, + 118, + 318, + 148 + ], + "blocks": [ + { + "bbox": [ + 282, + 118, + 318, + 148 + ], + "lines": [ + { + "bbox": [ + 282, + 118, + 318, + 148 + ], + "spans": [ + { + "bbox": [ + 282, + 118, + 318, + 148 + ], + "type": "image", + "image_path": "9d258d7ff7e46c6fcb826403db9ef4e571e00eb2d3fc21c8cd1899aedafc9927.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 321, + 118, + 352, + 151 + ], + "blocks": [ + { + "bbox": [ + 321, + 118, + 352, + 151 + ], + "lines": [ + { + "bbox": [ + 321, + 118, + 352, + 151 + ], + "spans": [ + { + "bbox": [ + 321, + 118, + 352, + 151 + ], + "type": "image", + "image_path": "2ef8a966bc475e3fe44ba16a17af541baa8a77e51f8f4e3e520365d98ec784ba.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 358, + 117, + 392, + 152 + ], + "blocks": [ + { + "bbox": [ + 358, + 117, + 392, + 152 + ], + "lines": [ + { + "bbox": [ + 358, + 117, + 392, + 152 + ], + "spans": [ + { + "bbox": [ + 358, + 117, + 392, + 152 + ], + "type": "image", + "image_path": "710fc2037e5667e2322b0ac0d01ca380a333cbf4cc6131a7b39ef51caab513aa.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 398, + 117, + 430, + 152 + ], + "blocks": [ + { + "bbox": [ + 398, + 117, + 430, + 152 + ], + "lines": [ + { + "bbox": [ + 398, + 117, + 430, + 152 + ], + "spans": [ + { + "bbox": [ + 398, + 117, + 430, + 152 + ], + "type": "image", + "image_path": "1ca721e7b50404ea76ed802a5ad335983db7dfff6080775cf86fe451b301adbb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 437, + 117, + 470, + 152 + ], + "blocks": [ + { + "bbox": [ + 437, + 117, + 470, + 152 + ], + "lines": [ + { + "bbox": [ + 437, + 117, + 470, + 152 + ], + "spans": [ + { + "bbox": [ + 437, + 117, + 470, + 152 + ], + "type": "image", + "image_path": "7a2e6a16ce625920ea6c051b73b4feba74d0b9e88c173d614bc1028a246f7ceb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 165, + 161, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 165, + 161, + 172 + ], + "spans": [ + { + "bbox": [ + 142, + 165, + 161, + 172 + ], + "type": "text", + "content": "USIBD" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 165, + 158, + 203, + 186 + ], + "blocks": [ + { + "bbox": [ + 165, + 158, + 203, + 186 + ], + "lines": [ + { + "bbox": [ + 165, + 158, + 203, + 186 + ], + "spans": [ + { + "bbox": [ + 165, + 158, + 203, + 186 + ], + "type": "image", + "image_path": "da9d0c7a36aaf9d34a4fa74a7e1e7dc75f66202a2f533501f84be66300b5e51f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 205, + 156, + 241, + 186 + ], + "blocks": [ + { + "bbox": [ + 205, + 156, + 241, + 186 + ], + "lines": [ + { + "bbox": [ + 205, + 156, + 241, + 186 + ], + "spans": [ + { + "bbox": [ + 205, + 156, + 241, + 186 + ], + "type": "image", + "image_path": "07f1738bd7ce200dcaf7b8cd70fe1ccc26be10f239c15616bf8d269b2e84c092.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 242, + 156, + 280, + 186 + ], + "blocks": [ + { + "bbox": [ + 242, + 156, + 280, + 186 + ], + "lines": [ + { + "bbox": [ + 242, + 156, + 280, + 186 + ], + "spans": [ + { + "bbox": [ + 242, + 156, + 280, + 186 + ], + "type": "image", + "image_path": "047963e72477ca14189977740918482cf11d1014f70b3b72bd95392716b0fc2e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 282, + 156, + 319, + 186 + ], + "blocks": [ + { + "bbox": [ + 282, + 156, + 319, + 186 + ], + "lines": [ + { + "bbox": [ + 282, + 156, + 319, + 186 + ], + "spans": [ + { + "bbox": [ + 282, + 156, + 319, + 186 + ], + "type": "image", + "image_path": "49fe7d179dbadb95ff6ca4f85fc4622226ed4c97760aa1e248cb16c0906f96f6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 321, + 155, + 353, + 187 + ], + "blocks": [ + { + "bbox": [ + 321, + 155, + 353, + 187 + ], + "lines": [ + { + "bbox": [ + 321, + 155, + 353, + 187 + ], + "spans": [ + { + "bbox": [ + 321, + 155, + 353, + 187 + ], + "type": "image", + "image_path": "9513a7232b5154aaa08b2b03c1c107c4f53151ae4134941396070890558afa9f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 358, + 155, + 392, + 189 + ], + "blocks": [ + { + "bbox": [ + 358, + 155, + 392, + 189 + ], + "lines": [ + { + "bbox": [ + 358, + 155, + 392, + 189 + ], + "spans": [ + { + "bbox": [ + 358, + 155, + 392, + 189 + ], + "type": "image", + "image_path": "23984463a674afed8486ccd9019b495eb67fcf9cecd799f6e5f3c463c24c80bc.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 398, + 155, + 430, + 189 + ], + "blocks": [ + { + "bbox": [ + 398, + 155, + 430, + 189 + ], + "lines": [ + { + "bbox": [ + 398, + 155, + 430, + 189 + ], + "spans": [ + { + "bbox": [ + 398, + 155, + 430, + 189 + ], + "type": "image", + "image_path": "52a1eae2f21686a1686694923c30c7f649b64cd024b98c5f181bb95eae8a855b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 437, + 155, + 470, + 189 + ], + "blocks": [ + { + "bbox": [ + 437, + 155, + 470, + 189 + ], + "lines": [ + { + "bbox": [ + 437, + 155, + 470, + 189 + ], + "spans": [ + { + "bbox": [ + 437, + 155, + 470, + 189 + ], + "type": "image", + "image_path": "693d1e8184130e652d34038f9bb80e33e56d2aeb40976ed81b8cab21c0a7a6f6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 142, + 203, + 162, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 203, + 162, + 210 + ], + "spans": [ + { + "bbox": [ + 142, + 203, + 162, + 210 + ], + "type": "text", + "content": "Yiet al." + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 165, + 194, + 203, + 224 + ], + "blocks": [ + { + "bbox": [ + 165, + 194, + 203, + 224 + ], + "lines": [ + { + "bbox": [ + 165, + 194, + 203, + 224 + ], + "spans": [ + { + "bbox": [ + 165, + 194, + 203, + 224 + ], + "type": "image", + "image_path": "b913d8ae2acf5bef63e607a28263e9a3930cc7b56bb9a5638e1bed34af38aad6.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 205, + 194, + 241, + 224 + ], + "blocks": [ + { + "bbox": [ + 205, + 194, + 241, + 224 + ], + "lines": [ + { + "bbox": [ + 205, + 194, + 241, + 224 + ], + "spans": [ + { + "bbox": [ + 205, + 194, + 241, + 224 + ], + "type": "image", + "image_path": "3041555f2617046202c212c116a35c3d9921038aec4e9a7190258ae68e4a19ba.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 242, + 194, + 280, + 224 + ], + "blocks": [ + { + "bbox": [ + 242, + 194, + 280, + 224 + ], + "lines": [ + { + "bbox": [ + 242, + 194, + 280, + 224 + ], + "spans": [ + { + "bbox": [ + 242, + 194, + 280, + 224 + ], + "type": "image", + "image_path": "775ba24bfeb822776b193b1cf2a334d4432efacf4691bf0193a4645c6e4b4262.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 282, + 194, + 319, + 224 + ], + "blocks": [ + { + "bbox": [ + 282, + 194, + 319, + 224 + ], + "lines": [ + { + "bbox": [ + 282, + 194, + 319, + 224 + ], + "spans": [ + { + "bbox": [ + 282, + 194, + 319, + 224 + ], + "type": "image", + "image_path": "7cd7426dcde5d7b3367cc327c76c4c1433254fcd2e44df40cf15a9c0724b56c3.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 321, + 194, + 353, + 227 + ], + "blocks": [ + { + "bbox": [ + 321, + 194, + 353, + 227 + ], + "lines": [ + { + "bbox": [ + 321, + 194, + 353, + 227 + ], + "spans": [ + { + "bbox": [ + 321, + 194, + 353, + 227 + ], + "type": "image", + "image_path": "626931ceb320f989b572f9f0c449de62a4dde28b9dd47b4f17b0e338e752ee4b.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 358, + 194, + 392, + 227 + ], + "blocks": [ + { + "bbox": [ + 358, + 194, + 392, + 227 + ], + "lines": [ + { + "bbox": [ + 358, + 194, + 392, + 227 + ], + "spans": [ + { + "bbox": [ + 358, + 194, + 392, + 227 + ], + "type": "image", + "image_path": "f6dd90eca91aeac57f3967ed25d55247f4ac12177b4fcc3f354b0bd545ac6a98.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 398, + 194, + 430, + 227 + ], + "blocks": [ + { + "bbox": [ + 398, + 194, + 430, + 227 + ], + "lines": [ + { + "bbox": [ + 398, + 194, + 430, + 227 + ], + "spans": [ + { + "bbox": [ + 398, + 194, + 430, + 227 + ], + "type": "image", + "image_path": "a24d78836285131efce261c026a02a37893f03bde2cb45a0421ba54c25f8000d.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 437, + 194, + 470, + 228 + ], + "blocks": [ + { + "bbox": [ + 437, + 194, + 470, + 228 + ], + "lines": [ + { + "bbox": [ + 437, + 194, + 470, + 228 + ], + "spans": [ + { + "bbox": [ + 437, + 194, + 470, + 228 + ], + "type": "image", + "image_path": "00aee554799ab1f0c4284411e3c42f6bf4d6ce013034bfb9b4200f17c4658a16.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 142, + 242, + 157, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 242, + 157, + 247 + ], + "spans": [ + { + "bbox": [ + 142, + 242, + 157, + 247 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 165, + 232, + 203, + 262 + ], + "blocks": [ + { + "bbox": [ + 165, + 232, + 203, + 262 + ], + "lines": [ + { + "bbox": [ + 165, + 232, + 203, + 262 + ], + "spans": [ + { + "bbox": [ + 165, + 232, + 203, + 262 + ], + "type": "image", + "image_path": "3f88d5f2729a6b63b1baadda0dbb0970e0827e3dc9152203b2615a9c4f08ab5d.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 267, + 195, + 274 + ], + "lines": [ + { + "bbox": [ + 165, + 267, + 195, + 274 + ], + "spans": [ + { + "bbox": [ + 165, + 267, + 195, + 274 + ], + "type": "text", + "content": "Rendering" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 205, + 232, + 240, + 262 + ], + "blocks": [ + { + "bbox": [ + 205, + 232, + 240, + 262 + ], + "lines": [ + { + "bbox": [ + 205, + 232, + 240, + 262 + ], + "spans": [ + { + "bbox": [ + 205, + 232, + 240, + 262 + ], + "type": "image", + "image_path": "3e504af9128ea8907af5d2fc127594a14d22842eba016d93e0869344f00c80c2.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 267, + 232, + 274 + ], + "lines": [ + { + "bbox": [ + 211, + 267, + 232, + 274 + ], + "spans": [ + { + "bbox": [ + 211, + 267, + 232, + 274 + ], + "type": "text", + "content": "Albedo" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 242, + 232, + 280, + 262 + ], + "blocks": [ + { + "bbox": [ + 242, + 232, + 280, + 262 + ], + "lines": [ + { + "bbox": [ + 242, + 232, + 280, + 262 + ], + "spans": [ + { + "bbox": [ + 242, + 232, + 280, + 262 + ], + "type": "image", + "image_path": "eac82589400dbabcc697961e438f8136b99c7cca4b39e4e79eec39d7f34b0a24.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 267, + 274, + 274 + ], + "lines": [ + { + "bbox": [ + 244, + 267, + 274, + 274 + ], + "spans": [ + { + "bbox": [ + 244, + 267, + 274, + 274 + ], + "type": "text", + "content": "roughness" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 282, + 232, + 318, + 262 + ], + "blocks": [ + { + "bbox": [ + 282, + 232, + 318, + 262 + ], + "lines": [ + { + "bbox": [ + 282, + 232, + 318, + 262 + ], + "spans": [ + { + "bbox": [ + 282, + 232, + 318, + 262 + ], + "type": "image", + "image_path": "69323977353543fdf250021d5195e067e30b57900190b3a2a4761eabf1e486af.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 267, + 309, + 274 + ], + "lines": [ + { + "bbox": [ + 288, + 267, + 309, + 274 + ], + "spans": [ + { + "bbox": [ + 288, + 267, + 309, + 274 + ], + "type": "text", + "content": "Relight" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 321, + 232, + 353, + 262 + ], + "blocks": [ + { + "bbox": [ + 321, + 232, + 353, + 262 + ], + "lines": [ + { + "bbox": [ + 321, + 232, + 353, + 262 + ], + "spans": [ + { + "bbox": [ + 321, + 232, + 353, + 262 + ], + "type": "image", + "image_path": "b02d56606838c0339dbb6a89db943a41b9f0a9b1c36eb429de7347b5b1e874d4.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 267, + 351, + 274 + ], + "lines": [ + { + "bbox": [ + 321, + 267, + 351, + 274 + ], + "spans": [ + { + "bbox": [ + 321, + 267, + 351, + 274 + ], + "type": "text", + "content": "Rendering" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 358, + 231, + 392, + 265 + ], + "blocks": [ + { + "bbox": [ + 358, + 231, + 392, + 265 + ], + "lines": [ + { + "bbox": [ + 358, + 231, + 392, + 265 + ], + "spans": [ + { + "bbox": [ + 358, + 231, + 392, + 265 + ], + "type": "image", + "image_path": "258c5e2848d08c7b91f0cb92251508f67cdaf55ed9d6f996a7fded522108607e.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 267, + 386, + 274 + ], + "lines": [ + { + "bbox": [ + 365, + 267, + 386, + 274 + ], + "spans": [ + { + "bbox": [ + 365, + 267, + 386, + 274 + ], + "type": "text", + "content": "Albedo" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 190, + 287, + 422, + 298 + ], + "lines": [ + { + "bbox": [ + 190, + 287, + 422, + 298 + ], + "spans": [ + { + "bbox": [ + 190, + 287, + 422, + 298 + ], + "type": "text", + "content": "Fig. 7: Comparison with other data-driven priors." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 398, + 232, + 430, + 265 + ], + "blocks": [ + { + "bbox": [ + 398, + 232, + 430, + 265 + ], + "lines": [ + { + "bbox": [ + 398, + 232, + 430, + 265 + ], + "spans": [ + { + "bbox": [ + 398, + 232, + 430, + 265 + ], + "type": "image", + "image_path": "e3f24526805a6d9d2971bc1b5566587094528c4095050b714a737cb9812ff02e.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 267, + 428, + 274 + ], + "lines": [ + { + "bbox": [ + 397, + 267, + 428, + 274 + ], + "spans": [ + { + "bbox": [ + 397, + 267, + 428, + 274 + ], + "type": "text", + "content": "roughness" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 437, + 231, + 470, + 265 + ], + "blocks": [ + { + "bbox": [ + 437, + 231, + 470, + 265 + ], + "lines": [ + { + "bbox": [ + 437, + 231, + 470, + 265 + ], + "spans": [ + { + "bbox": [ + 437, + 231, + 470, + 265 + ], + "type": "image", + "image_path": "92aaf15911c5b766be51c4f1f8b55011651a7548f6822f2f6a7997cefc093fd8.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 442, + 267, + 462, + 274 + ], + "lines": [ + { + "bbox": [ + 442, + 267, + 462, + 274 + ], + "spans": [ + { + "bbox": [ + 442, + 267, + 462, + 274 + ], + "type": "text", + "content": "Relight" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "bbox": [ + 130, + 321, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 514 + ], + "type": "text", + "content": "the environment map from our novel light dataset, we also make a comparison on the relighted images. Because Invrender [81] and TensoIR [24] parameterize environment maps with SGs, we first optimize an SG with 128 lobes through Stochastic Gradient Decent optimization. For NvDiffRec [43], NvDiffRecMC [20], and NeILF++ [76], we render the relighted image with Blender [6] using our extracted topology and material texture maps with SPP=4096 and 4 bounces of lights. Due to the inevitable scale ambiguity between lighting and albedo, previous methods [80] align the predicted albedo to the ground truth before relighting and evaluation. Each channel is separately aligned to the ground truth. However, this would cause the incorrectly baked shading in the albedo to be removed during alignment. To this end, we align the average RGB of prediction to ground truth for a more faithful evaluation. Because geometry affects the inverse rendering performance greatly, to make sure the topology for NvDiffRec and NvdiffRecMC is well initialized, we make another set of experiments for NvDiffRec and NvDiffRecMC with MonoSDF [75] reconstructed geometry given and as a fixed topology along the training and evaluation process." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 130, + 522, + 482, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 572 + ], + "type": "text", + "content": "Metrics. We use Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) [79] as metrics to evaluate the image quality of the aligned albedo, relighting images, and novel view synthesis. Roughness is evaluated using Mean Squared Error(MSE)." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Comparison with optimization-based methods. Tab. 1 quantitatively compares our method with recent inverse rendering methods on the synthetic dataset. We significantly outperform baseline methods in terms of recovered albedo, roughness, and relighting, demonstrating the effectiveness of the proposed material prior. Although TensoIR [24] and NvDiffRec [43] show good results on view synthesis under the original environment lighting, it does not accurately estimate the underlying materials, which will be further shown in the qualitative results." + } + ] + } + ], + "index": 49 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "Chen. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "Tab. 1 also shows that we further improve the performance of material recovery by guiding the diffusion model with estimated albedo and specular." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 226 + ], + "type": "text", + "content": "Fig. 4 and Fig. 5 present the qualitative results on estimated materials and relighting. Fig. 6 demonstrates the performance of our method on the real-world data. More results on real data are presented in the supplementary material. NvDiffrec and " + }, + { + "bbox": [ + 130, + 141, + 482, + 226 + ], + "type": "inline_equation", + "content": "\\mathrm{NeILF}++" + }, + { + "bbox": [ + 130, + 141, + 482, + 226 + ], + "type": "text", + "content": " do not disentangle the albedo and light well and show strong light color baked in albedo, as visualized in Fig. 5 and Fig. 6. With our diffusion priors learned from the large-scale dataset, our method can disentangle materials and lighting in a reasonable way." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 340 + ], + "type": "text", + "content": "Comparison with other data-driven priors. Tab. 1 presents the comparison with data-driven priors [38,71]. Although these methods are able to prevent shadows from being baked into albedo, they still struggle to separate shading from the albedo of objects under complex lighting conditions, as shown in qualitative results in Fig. 7. The incorrect priors will mislead the optimization process and cause large errors in the estimated materials. In contrast, our prior models have a better capacity to decouple shading from materials. We present more qualitative comparisons of the real-world data in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 357, + 481, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 481, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 481, + 393 + ], + "type": "text", + "content": "Ablation Study. We conduct ablation studies on our diffusion priors and the guided diffusion for fine-level inverse rendering. Quantitative and qualitative results are shown in Tab. 2 and Fig. 8, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 395, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 395, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 395, + 482, + 514 + ], + "type": "text", + "content": "\"w/o spec.\" indicates training without the specular shading prior model. This would result in recovering incorrect lighting and roughness, as the diffuse shading and specular shading cannot be correctly separated by relying solely on the albedo prior. As depicted in Fig. 8, the \"hat\" 's roughness prediction is lower and the \"ship\" 's is higher in \"w/o Spec.\" than ground truth, which resulting in inaccurate relighting outcomes. In \"w/o albedo\", the inverse rendering is conducted without albedo prior. The recovered albedo has baked-in shading, as shown in Fig. 8 where the two objects baked shading into the albedo causing incorrect relighting. This indicates that both priors are crucial for accurate lighting and material decomposition." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 515, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 588 + ], + "type": "text", + "content": "\"w/o guided\" is the result without the guided diffusion for fine-level optimization. Due to the lack of multi-view consistency of diffusion prior, the optimized materials are noisy and have baked-in shading, as shown in Fig. 8. This hurts the relighting quality, as shown in qualitative results in Tab. 2. Therefore, the training process using the guided diffusion model is essential for a smooth and accurate material recovery." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Modeling indirect illumination and metallic materials. Following the setting of InvRender [81], we extend our method to handle indirect lighting. Inspired by [66], we precompute shading of indirect lighting for each image by tracing the trained SDF and Color field. Additionally, we also include metallic texture as an optimization target. More details and results are in the supplementary materials." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "type": "text", + "content": "IntrinsicAnything" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 136, + 480, + 192 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 420, + 126 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 420, + 126 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 420, + 126 + ], + "type": "text", + "content": "Table 2: Quantitative results of ablation studies." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 136, + 480, + 192 + ], + "lines": [ + { + "bbox": [ + 138, + 136, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 138, + 136, + 480, + 192 + ], + "type": "table", + "html": "
MethodAlbedoAligned AlbedoView SynthesisRelightingRoughness
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓MSE↓
Ours24.67660.95620.050326.88290.95320.037830.24050.97410.032828.97210.97570.02890.0105
Ours w/o spec17.21260.93160.056725.73890.94910.044429.42450.97280.035226.47430.97290.03630.0758
Ours w/o albedo22.71110.94340.057124.34620.93810.048830.59980.97510.026127.28510.97430.02580.0118
Ours w/o guided23.35070.93030.065224.61120.92080.064129.79370.96880.033927.73900.96830.03610.0168
", + "image_path": "4b3294f45f372594ad44b37ee13cb337814feb8a6988aac9027b72f07c3b844c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 204, + 165, + 241 + ], + "blocks": [ + { + "bbox": [ + 134, + 204, + 165, + 241 + ], + "lines": [ + { + "bbox": [ + 134, + 204, + 165, + 241 + ], + "spans": [ + { + "bbox": [ + 134, + 204, + 165, + 241 + ], + "type": "image", + "image_path": "4e2f6a49b27b75e2d0ba3041614004a29e2c9704a4e87de8edc5a20b801c5fa1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 179, + 204, + 199, + 240 + ], + "blocks": [ + { + "bbox": [ + 179, + 204, + 199, + 240 + ], + "lines": [ + { + "bbox": [ + 179, + 204, + 199, + 240 + ], + "spans": [ + { + "bbox": [ + 179, + 204, + 199, + 240 + ], + "type": "image", + "image_path": "25a88a5f0eb08ead0fd7edb23330aff44a9b804605966f0339935a86d23cb549.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 214, + 204, + 233, + 240 + ], + "blocks": [ + { + "bbox": [ + 214, + 204, + 233, + 240 + ], + "lines": [ + { + "bbox": [ + 214, + 204, + 233, + 240 + ], + "spans": [ + { + "bbox": [ + 214, + 204, + 233, + 240 + ], + "type": "image", + "image_path": "3cec1a1aef923e4a068988e341c3470cb2b1e6ab14dd89ffec25161c82331186.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 249, + 204, + 268, + 240 + ], + "blocks": [ + { + "bbox": [ + 249, + 204, + 268, + 240 + ], + "lines": [ + { + "bbox": [ + 249, + 204, + 268, + 240 + ], + "spans": [ + { + "bbox": [ + 249, + 204, + 268, + 240 + ], + "type": "image", + "image_path": "c8a1b0f687370e4be06e9ea7ae937a40c1cd59d26b4e4510fdfc03846207cdef.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 283, + 205, + 301, + 240 + ], + "blocks": [ + { + "bbox": [ + 283, + 205, + 301, + 240 + ], + "lines": [ + { + "bbox": [ + 283, + 205, + 301, + 240 + ], + "spans": [ + { + "bbox": [ + 283, + 205, + 301, + 240 + ], + "type": "image", + "image_path": "b6107ce4f5f0e5e13a3f491a254e45d4ccb6a70183fad1726d1a6ef6f0dfa56e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 311, + 209, + 339, + 241 + ], + "blocks": [ + { + "bbox": [ + 311, + 209, + 339, + 241 + ], + "lines": [ + { + "bbox": [ + 311, + 209, + 339, + 241 + ], + "spans": [ + { + "bbox": [ + 311, + 209, + 339, + 241 + ], + "type": "image", + "image_path": "7ea3a3e3d0aa644b6f741ebd573cfc9565bd9952d9959e5d3c3cc6ebfc7050d0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 345, + 209, + 373, + 241 + ], + "blocks": [ + { + "bbox": [ + 345, + 209, + 373, + 241 + ], + "lines": [ + { + "bbox": [ + 345, + 209, + 373, + 241 + ], + "spans": [ + { + "bbox": [ + 345, + 209, + 373, + 241 + ], + "type": "image", + "image_path": "c26d46edb610b0b2954646e5c289b3dfdfca7905201a8b362e6985e719892b0a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 381, + 209, + 407, + 241 + ], + "blocks": [ + { + "bbox": [ + 381, + 209, + 407, + 241 + ], + "lines": [ + { + "bbox": [ + 381, + 209, + 407, + 241 + ], + "spans": [ + { + "bbox": [ + 381, + 209, + 407, + 241 + ], + "type": "image", + "image_path": "aefe093ef6679008f19ba59badf06c6235d427673a26903784b80ff1bfb38f78.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 414, + 209, + 441, + 241 + ], + "blocks": [ + { + "bbox": [ + 414, + 209, + 441, + 241 + ], + "lines": [ + { + "bbox": [ + 414, + 209, + 441, + 241 + ], + "spans": [ + { + "bbox": [ + 414, + 209, + 441, + 241 + ], + "type": "image", + "image_path": "9f68d9eafaf428538eabbb25f98ffe10964132917cdff57efc5d77154745ac31.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 448, + 209, + 475, + 241 + ], + "blocks": [ + { + "bbox": [ + 448, + 209, + 475, + 241 + ], + "lines": [ + { + "bbox": [ + 448, + 209, + 475, + 241 + ], + "spans": [ + { + "bbox": [ + 448, + 209, + 475, + 241 + ], + "type": "image", + "image_path": "1d61b8e5fb060787664cacd372d3975d3efc06b25271e67f0a638688eff0e491.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 135, + 247, + 162, + 277 + ], + "blocks": [ + { + "bbox": [ + 135, + 247, + 162, + 277 + ], + "lines": [ + { + "bbox": [ + 135, + 247, + 162, + 277 + ], + "spans": [ + { + "bbox": [ + 135, + 247, + 162, + 277 + ], + "type": "image", + "image_path": "73ada74eb334f49b112510d04f61116daa78470a8d6ac6ae310c9b1d2ffeab9e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 166, + 245, + 181, + 279 + ], + "blocks": [ + { + "bbox": [ + 166, + 245, + 181, + 279 + ], + "lines": [ + { + "bbox": [ + 166, + 245, + 181, + 279 + ], + "spans": [ + { + "bbox": [ + 166, + 245, + 181, + 279 + ], + "type": "image", + "image_path": "f48c1cc2dfc8a6843acd3dc9eccff5a1f6882a97256fa02b82d3718eca54dc24.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 199, + 245, + 217, + 279 + ], + "blocks": [ + { + "bbox": [ + 199, + 245, + 217, + 279 + ], + "lines": [ + { + "bbox": [ + 199, + 245, + 217, + 279 + ], + "spans": [ + { + "bbox": [ + 199, + 245, + 217, + 279 + ], + "type": "image", + "image_path": "06bf811784271d37b1a34275a8a562e2d83cb2142b2c2b417e27d0213d6e2948.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 249, + 245, + 267, + 279 + ], + "blocks": [ + { + "bbox": [ + 249, + 245, + 267, + 279 + ], + "lines": [ + { + "bbox": [ + 249, + 245, + 267, + 279 + ], + "spans": [ + { + "bbox": [ + 249, + 245, + 267, + 279 + ], + "type": "image", + "image_path": "c7adee1f6384d4b70c4782b24c05289ac931c9a674693c04651df01b51a9a484.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 283, + 246, + 301, + 279 + ], + "blocks": [ + { + "bbox": [ + 283, + 246, + 301, + 279 + ], + "lines": [ + { + "bbox": [ + 283, + 246, + 301, + 279 + ], + "spans": [ + { + "bbox": [ + 283, + 246, + 301, + 279 + ], + "type": "image", + "image_path": "3adb476411b22a08f20ad7f2ddaf450bef30e4d66a2d8fc0df31bfc4d8f570c1.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 312, + 249, + 339, + 281 + ], + "blocks": [ + { + "bbox": [ + 312, + 249, + 339, + 281 + ], + "lines": [ + { + "bbox": [ + 312, + 249, + 339, + 281 + ], + "spans": [ + { + "bbox": [ + 312, + 249, + 339, + 281 + ], + "type": "image", + "image_path": "fabafb0025ca893ea1384173a872c607d23d825096d4775b35af6db2355bad72.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 346, + 249, + 373, + 281 + ], + "blocks": [ + { + "bbox": [ + 346, + 249, + 373, + 281 + ], + "lines": [ + { + "bbox": [ + 346, + 249, + 373, + 281 + ], + "spans": [ + { + "bbox": [ + 346, + 249, + 373, + 281 + ], + "type": "image", + "image_path": "76a35658085605f4313308b50306a9abf1ee447b483357681222cb08f40988fb.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 381, + 249, + 408, + 281 + ], + "blocks": [ + { + "bbox": [ + 381, + 249, + 408, + 281 + ], + "lines": [ + { + "bbox": [ + 381, + 249, + 408, + 281 + ], + "spans": [ + { + "bbox": [ + 381, + 249, + 408, + 281 + ], + "type": "image", + "image_path": "56a1e30bc10532ca715af2e67d59499996d57383d3fb95cf631c74c41c60db92.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 414, + 249, + 441, + 281 + ], + "blocks": [ + { + "bbox": [ + 414, + 249, + 441, + 281 + ], + "lines": [ + { + "bbox": [ + 414, + 249, + 441, + 281 + ], + "spans": [ + { + "bbox": [ + 414, + 249, + 441, + 281 + ], + "type": "image", + "image_path": "618dec07efae921a338f8ecbbcab433f288a23eda4ddef48b85f41c254aef527.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 448, + 249, + 475, + 281 + ], + "blocks": [ + { + "bbox": [ + 448, + 249, + 475, + 281 + ], + "lines": [ + { + "bbox": [ + 448, + 249, + 475, + 281 + ], + "spans": [ + { + "bbox": [ + 448, + 249, + 475, + 281 + ], + "type": "image", + "image_path": "909d8b23c4930fd696037d3d2f4555817b254e4dccd3afcd4a24ff1edd51e659.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 135, + 286, + 161, + 312 + ], + "blocks": [ + { + "bbox": [ + 135, + 286, + 161, + 312 + ], + "lines": [ + { + "bbox": [ + 135, + 286, + 161, + 312 + ], + "spans": [ + { + "bbox": [ + 135, + 286, + 161, + 312 + ], + "type": "image", + "image_path": "3fcb9ecaedfb493805aeffb5ec70eaaae0c3617fc4d009c62b278e91844a60c1.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 343, + 413, + 354 + ], + "lines": [ + { + "bbox": [ + 198, + 343, + 413, + 354 + ], + "spans": [ + { + "bbox": [ + 198, + 343, + 413, + 354 + ], + "type": "text", + "content": "Fig. 8: Qualitative results of ablation studies." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 165, + 285, + 181, + 319 + ], + "blocks": [ + { + "bbox": [ + 165, + 285, + 181, + 319 + ], + "lines": [ + { + "bbox": [ + 165, + 285, + 181, + 319 + ], + "spans": [ + { + "bbox": [ + 165, + 285, + 181, + 319 + ], + "type": "image", + "image_path": "0892ab3d09089931ac3e4428c47116672d126fdb3cf2c7a506d322463f463f48.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 182, + 285, + 199, + 319 + ], + "blocks": [ + { + "bbox": [ + 182, + 285, + 199, + 319 + ], + "lines": [ + { + "bbox": [ + 182, + 285, + 199, + 319 + ], + "spans": [ + { + "bbox": [ + 182, + 285, + 199, + 319 + ], + "type": "image", + "image_path": "3380a9e4b7b000d9f9eb272c161b76944277df36d39dab31892c37e62ccc62c2.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 324, + 236, + 331 + ], + "lines": [ + { + "bbox": [ + 208, + 324, + 236, + 331 + ], + "spans": [ + { + "bbox": [ + 208, + 324, + 236, + 331 + ], + "type": "text", + "content": "w/o albedo" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 214, + 285, + 233, + 319 + ], + "blocks": [ + { + "bbox": [ + 214, + 285, + 233, + 319 + ], + "lines": [ + { + "bbox": [ + 214, + 285, + 233, + 319 + ], + "spans": [ + { + "bbox": [ + 214, + 285, + 233, + 319 + ], + "type": "image", + "image_path": "b82aca426132d6298e93b0f2712841af00f84b8f2285af2e0a71d7cf389a94c0.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 325, + 274, + 331 + ], + "lines": [ + { + "bbox": [ + 246, + 325, + 274, + 331 + ], + "spans": [ + { + "bbox": [ + 246, + 325, + 274, + 331 + ], + "type": "text", + "content": "w/o guided" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 249, + 285, + 268, + 319 + ], + "blocks": [ + { + "bbox": [ + 249, + 285, + 268, + 319 + ], + "lines": [ + { + "bbox": [ + 249, + 285, + 268, + 319 + ], + "spans": [ + { + "bbox": [ + 249, + 285, + 268, + 319 + ], + "type": "image", + "image_path": "5abebf74672bc6605707ceb6b10c8d7070b202fffd58e4c37eeb6160b02fd46f.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 325, + 296, + 331 + ], + "lines": [ + { + "bbox": [ + 284, + 325, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 284, + 325, + 296, + 331 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 284, + 285, + 302, + 319 + ], + "blocks": [ + { + "bbox": [ + 284, + 285, + 302, + 319 + ], + "lines": [ + { + "bbox": [ + 284, + 285, + 302, + 319 + ], + "spans": [ + { + "bbox": [ + 284, + 285, + 302, + 319 + ], + "type": "image", + "image_path": "cc67f98a679e89001bc2a2b2c67d0d357cc7502ff8c14ec20ed75c034c14d56b.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 312, + 289, + 339, + 320 + ], + "blocks": [ + { + "bbox": [ + 312, + 289, + 339, + 320 + ], + "lines": [ + { + "bbox": [ + 312, + 289, + 339, + 320 + ], + "spans": [ + { + "bbox": [ + 312, + 289, + 339, + 320 + ], + "type": "image", + "image_path": "56c701a677140b4acac71d9e22760cb60db94f0906aee7e66018d04f90c474f9.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 325, + 327, + 331 + ], + "lines": [ + { + "bbox": [ + 321, + 325, + 327, + 331 + ], + "spans": [ + { + "bbox": [ + 321, + 325, + 327, + 331 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 346, + 289, + 373, + 320 + ], + "blocks": [ + { + "bbox": [ + 346, + 289, + 373, + 320 + ], + "lines": [ + { + "bbox": [ + 346, + 289, + 373, + 320 + ], + "spans": [ + { + "bbox": [ + 346, + 289, + 373, + 320 + ], + "type": "image", + "image_path": "50a8b47b1f7b87b0d44efa0299ce009b88857a42bac072c27dea10f8224c5626.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 325, + 370, + 331 + ], + "lines": [ + { + "bbox": [ + 347, + 325, + 370, + 331 + ], + "spans": [ + { + "bbox": [ + 347, + 325, + 370, + 331 + ], + "type": "text", + "content": "w/o spec." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 380, + 289, + 406, + 320 + ], + "blocks": [ + { + "bbox": [ + 380, + 289, + 406, + 320 + ], + "lines": [ + { + "bbox": [ + 380, + 289, + 406, + 320 + ], + "spans": [ + { + "bbox": [ + 380, + 289, + 406, + 320 + ], + "type": "image", + "image_path": "1efcb4c7cf35bde05878b4689928963d98aa1c9169cc62e1e0762893ba207c47.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 325, + 405, + 331 + ], + "lines": [ + { + "bbox": [ + 378, + 325, + 405, + 331 + ], + "spans": [ + { + "bbox": [ + 378, + 325, + 405, + 331 + ], + "type": "text", + "content": "w/o albedo" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 414, + 289, + 441, + 320 + ], + "blocks": [ + { + "bbox": [ + 414, + 289, + 441, + 320 + ], + "lines": [ + { + "bbox": [ + 414, + 289, + 441, + 320 + ], + "spans": [ + { + "bbox": [ + 414, + 289, + 441, + 320 + ], + "type": "image", + "image_path": "0020161009a2a31d93a022d013ba381d607023960e176ce612154053fc346d1a.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 325, + 445, + 331 + ], + "lines": [ + { + "bbox": [ + 416, + 325, + 445, + 331 + ], + "spans": [ + { + "bbox": [ + 416, + 325, + 445, + 331 + ], + "type": "text", + "content": "w/o guided" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 448, + 289, + 475, + 320 + ], + "blocks": [ + { + "bbox": [ + 448, + 289, + 475, + 320 + ], + "lines": [ + { + "bbox": [ + 448, + 289, + 475, + 320 + ], + "spans": [ + { + "bbox": [ + 448, + 289, + 475, + 320 + ], + "type": "image", + "image_path": "12dee739272c986f386caf2bb96213d05c96399265f096bc7d88032b13776e03.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 325, + 466, + 331 + ], + "lines": [ + { + "bbox": [ + 454, + 325, + 466, + 331 + ], + "spans": [ + { + "bbox": [ + 454, + 325, + 466, + 331 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "bbox": [ + 132, + 377, + 219, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 377, + 219, + 389 + ], + "spans": [ + { + "bbox": [ + 132, + 377, + 219, + 389 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "text", + "content": "We proposed to learn the material prior for inverse rendering under an unknown static lighting condition. To this end, the material prior is designed as conditional diffusion models of albedo and specular shading components, according to the general rendering equation. In addition, a two-stage training scheme is developed for robustly regularizing the inverse rendering process with conditional diffusion models, which uses coarse materials from the first stage training to guide diffusion models to provide multi-view consistent constraints. The proposed model significantly outperforms baseline methods on real-world and synthetic data." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 594 + ], + "type": "text", + "content": "Limitations. Although we achieved impressive results on material recovery, our approach still has some limitations. First, we do not consider the inverse rendering of transparent objects. A solution is revising the geometry representation to a neural field of transparency and jointly optimizing it with object materials and environment lighting. Second, the performance of our model is limited by the accuracy of the reconstructed geometry from MonoSDF. It would be interesting to exploit diffusion models to learn the geometry prior." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 130, + 605, + 482, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 654 + ], + "type": "text", + "content": "Acknowledgments. The authors would like to acknowledge the support from NSFC (No. 62172364), the Alibaba Innovation Research Program of Alibaba Group and the Information Technology Center and State Key Lab of CAD&CG, Zhejiang University." + } + ] + } + ], + "index": 47 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "Chen. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 140, + 418, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 418, + 150 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 418, + 150 + ], + "type": "text", + "content": "1. Poly haven: The public 3d asset library. https://polyhaven.com/" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 151, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 151, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 151, + 481, + 172 + ], + "type": "text", + "content": "2. Stable diffusion image variations. https://huggingface.co/spaces/lambdalabs/stable-diffusion-image-variations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 173, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 481, + 205 + ], + "type": "text", + "content": "3. Anciukevicius, T., Xu, Z., Fisher, M., Henderson, P., Bilen, H., Mitra, N.J., Guerrero, P.: Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. In: CVPR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 481, + 226 + ], + "type": "text", + "content": "4. Barron, J.T., Malik, J.: Shape, illumination, and reflectance from shading. TPAMI 37(8), 1670-1687 (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 228, + 481, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 481, + 260 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 481, + 260 + ], + "type": "text", + "content": "5. Bi, S., Lombardi, S., Saito, S., Simon, T., Wei, S.E., McPhail, K., Ramamoorthi, R., Sheikh, Y., Saragih, J.: Deep relightable appearance models for animatable faces. ACM Trans. Graph. (Proc. SIGGRAPH) 40(4) (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 261, + 338, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 338, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 338, + 270 + ], + "type": "text", + "content": "6. Blender: Blender, https://www.blender.org/" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 481, + 303 + ], + "type": "text", + "content": "7. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.P.: Nerd: Neural reflectance decomposition from image collections. In: IEEE International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 304, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 481, + 326 + ], + "type": "text", + "content": "8. Burley, B., Studios, W.D.A.: Physically-based shading at disney. In: ACM Trans. Graph. vol. 2012, pp. 1-7. vol. 2012 (2012)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 326, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 481, + 369 + ], + "type": "text", + "content": "9. Chang, A.X., Funkhouser, T.A., Guibas, L.J., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: Shapenet: An information-rich 3d model repository. CoRR abs/1512.03012 (2015), http://arxiv.org/abs/1512.03012" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 370, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 481, + 391 + ], + "type": "text", + "content": "0. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: ECCV (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 481, + 413 + ], + "type": "text", + "content": "1. Chen, H., Gu, J., Chen, A., Tian, W., Tu, Z., Liu, L., Su, H.: Single-stage diffusion nef: A unified approach to 3d generation and reconstruction. In: ICCV (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 414, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 481, + 435 + ], + "type": "text", + "content": "2. Cheng, Y.C., Lee, H.Y., Tulyakov, S., Schwing, A.G., Gui, L.Y.: Sdfusion: Multi-modal 3d shape completion, reconstruction, and generation. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 436, + 481, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 481, + 467 + ], + "type": "text", + "content": "3. Chung, H., Kim, J., Mccann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. In: ICLR (2023), https://openreview.net/forum?id=OnD9zGAGTOK" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 481, + 501 + ], + "type": "text", + "content": "4. Chung, H., Sim, B., Ye, J.C.: Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In: CVPR. pp. 12413-12422 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 502, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 481, + 522 + ], + "type": "text", + "content": "5. Croitoru, F.A., Hondru, V., Ionescu, R.T., Shah, M.: Diffusion models in vision: A survey. PAMI (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 481, + 555 + ], + "type": "text", + "content": "6. Debevec, P., Hawkins, T., Tchou, C., Duiker, H.P., Sarokin, W., Sagar, M.: Acquiring the reflectance field of a human face. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 145-156 (2000)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 556, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 481, + 588 + ], + "type": "text", + "content": "7. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objverse: A universe of annotated 3d objects. In: CVPR. pp. 13142-13153 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 481, + 621 + ], + "type": "text", + "content": "8. Deng, C., Jiang, C., Qi, C.R., Yan, X., Zhou, Y., Guibas, L., Anguelov, D., et al.: Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In: CVPR. pp. 20637-20647 (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "type": "text", + "content": "9. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escalano, S., Pandey, R., Dourgarian, J., Tang, D., Tkach, A., Kowdle, A., Cooper, E., Dou, M., Fanello, S., Fyffe, G., Rhemann, C., Taylor, J., Debevec, P., Izadi, S.: The relightables: Volumetric performance capture of humans with" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "type": "text", + "content": "IntrinsicAnything" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 149, + 117, + 479, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 117, + 479, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 117, + 479, + 138 + ], + "type": "text", + "content": "realistic relighting. ACM Trans. Graph. 38(6) (nov 2019). https://doi.org/10.1145/3355089.3356571, https://doi.org/10.1145/3355089.3356571" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "20. Hasselgren, J., Hofmann, N., Munkberg, J.: Shape, light & material decomposition from images using monte carlo rendering and denoising. In: NeurIPS (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "21. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 204 + ], + "type": "text", + "content": "22. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "23. Horita, D., Yang, J., Chen, D., Koyama, Y., Aizawa, K., Sebe, N.: A structure-guided diffusion model for large-hole image completion. In: BMVC (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "24. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoir: Tensorial inverse rendering. In: CVPR (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 249, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 293 + ], + "type": "text", + "content": "25. Kajiya, J.T.: The rendering equation. In: Proceedings of the 13th Annual Conference on Computer Graphics and Interactive Techniques. p. 143-150. SIGGRAPH '86, Association for Computing Machinery, New York, NY, USA (1986). https://doi.org/10.1145/15922.15902, https://doi.org/10.1145/15922.15902" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "26. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. arXiv preprint arXiv:2201.11793 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "27. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "28. Kocsis, P., Sitzmann, V., Nießner, M.: Intrinsic image diffusion for single-view material estimation. In: arxiv (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "29. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Trans. Graph. 39(6) (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 402, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 456 + ], + "type": "text", + "content": "30. Levoy, M., Pulli, K., Curless, B., Rusinkiewicz, S., Koller, D., Pereira, L., Ginzton, M., Anderson, S.E., Davis, J., Ginsberg, J., Shade, J., Fulk, D.: The digital michelangelo project: 3d scanning of large statues. Proceedings of the 27th annual conference on Computer graphics and interactive techniques (2000), https://apisemantic scholar.org/CorpusID:1546062" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "text", + "content": "31. Li, Z., Snavely, N.: Cginintrinsics: Better intrinsic image decomposition through physically-based rendering. In: ECCV (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "text", + "content": "32. Li, Z., Xu, Z., Ramamoorthi, R., Sunkavalli, K., Chandraker, M.: Learning to reconstruct shape and spatially-varying reflectance from a single image. ACM Trans. Graph. 37(6), 1-11 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "33. Li, Z., Yu, T.W., Sang, S., Wang, S., Song, M., Liu, Y., Yeh, Y.Y., Zhu, R., Gundavarapu, N., Shi, J., et al.: Openrooms: An open framework for photorealistic indoor scene datasets. In: CVPR. pp. 7190-7199 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 544, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 567 + ], + "type": "text", + "content": "34. Liang, Z., Zhang, Q., Feng, Y., Shan, Y., Jia, K.: Gs-ir: 3d gaussian splatting for inverse rendering. Computer Graphics Forum (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "35. Lichy, D., Wu, J., Sengupta, S., Jacobs, D.W.: Shape and material capture at home. In: CVPR (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "text", + "content": "36. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. In: ICLR (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "37. Liu, Y., Wang, P., Lin, C., Long, X., Wang, J., Liu, L., Komura, T., Wang, W.: Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In: ACM Trans. Graph. (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "38. Liu, Y., Li, Y., You, S., Lu, F.: Unsupervised learning for intrinsic image decomposition from a single image. In: CVPR (2020)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Chen. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "39. Liu, Z., Li, Y., Lin, Y., Yu, X., Peng, S., Cao, Y.P., Qi, X., Huang, X., Liang, D., Ouyang, W.: Unidream: Unifying diffusion priors for relightable text-to-3d generation (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "text", + "content": "40. Lyu, L., Tewari, A., Habermann, M., Saito, S., Zollhöfer, M., Leimkühler, T., Theobalt, C.: Diffusion posterior illumination for ambiguity-aware inverse rendering. ACM Trans. Graph. 42(6) (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 481, + 205 + ], + "type": "text", + "content": "41. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: ICLR (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 481, + 237 + ], + "type": "text", + "content": "42. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 481, + 270 + ], + "type": "text", + "content": "43. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: CVPR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 270, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 481, + 315 + ], + "type": "text", + "content": "44. Nam, G., Lee, J.H., Gutierrez, D., Kim, M.H.: Practical svbrdf acquisition of 3d objects with unstructured flash photography. ACM Trans. Graph. (Proc. SIGGRAPH Asia) 37(6), 267:1-12 (2018). https://doi.org/10.1145/3272127.3275017, http://dx.doi.org/10.1145/3272127.3275017" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 315, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 481, + 335 + ], + "type": "text", + "content": "45. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: ICLR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 335, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 481, + 369 + ], + "type": "text", + "content": "46. Qiu, L., Chen, G., Gu, X., Zuo, Q., Xu, M., Wu, Y., Yuan, W., Dong, Z., Bo, L., Han, X.: Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In: CVPR. pp. 9914–9925 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 369, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 481, + 391 + ], + "type": "text", + "content": "47. Ramanishka, V., Das, A., Zhang, J., Saenko, K.: Learning transferable visual models from natural language. In: CVPR. pp. 4397-4406 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 391, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 481, + 424 + ], + "type": "text", + "content": "48. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. PAMI 44(3) (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 424, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 481, + 445 + ], + "type": "text", + "content": "49. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 446, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 481, + 468 + ], + "type": "text", + "content": "50. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 481, + 501 + ], + "type": "text", + "content": "51. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 501, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 481, + 534 + ], + "type": "text", + "content": "52. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 534, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 481, + 555 + ], + "type": "text", + "content": "53. Sang, S., Chandraker, M.: Single-shot neural relighting and svbrdf estimation. In: ECCV (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 555, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 481, + 578 + ], + "type": "text", + "content": "54. Shi, J., Dong, Y., Su, H., Yu, S.X.: Learning non-lambertian object intrinsics across shapenet categories. In: CVPR (July 2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 481, + 610 + ], + "type": "text", + "content": "55. Shi, Y., Wu, Y., Wu, C., Liu, X., Zhao, C., Feng, H., Liu, J., Zhang, L., Zhang, J., Zhou, B., Ding, E., Wang, J.: Gir: 3d gaussian inverse rendering for relightable scene factorization. Arxiv (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 610, + 481, + 643 + ], + "type": "text", + "content": "56. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. CoRR abs/2012.03927 (2020), https://arxiv.org/abs/2012.03927" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 481, + 666 + ], + "type": "text", + "content": "57. Sun, C., Cai, G., Li, Z., Yan, K., Zhang, C., Marshall, C., Huang, J.B., Zhao, S., Dong, Z.: Neural-pbir reconstruction of shape, material, and illumination (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "type": "text", + "content": "IntrinsicAnything" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "58. Tang, J., Wang, T., Zhang, B., Zhang, T., Yi, R., Ma, L., Chen, D.: Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In: ICCV. pp. 22819-22829 (October 2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "59. Tang, S., Zhang, F., Chen, J., Wang, P., Furukawa, Y.: Mvdiffusion: Enabling holistic multi-view image generation with correspondence-aware diffusion. In: NeurIPS (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 226 + ], + "type": "text", + "content": "60. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) NIPS. vol. 30. Curran Associates, Inc. (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "type": "text", + "content": "61. Veach, E., Guibas, L.J.: Optimally combining sampling techniques for monte carlo rendering. In: Proceedings of the 22nd annual conference on Computer graphics and interactive techniques. pp. 419-428 (1995)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "62. Vecchio, G., Martin, R., Roullier, A., Kaiser, A., Rouffet, R., Deschaintre, V., Boubekeur, T.: Controlmat: Controlled generative approach to material capture. arXiv preprint arXiv:2309.01700 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "63. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In: NeurIPS (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 325, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 336 + ], + "type": "text", + "content": "64. Wimbauer, F., Wu, S., Rupprecht, C.: De-rendering 3d objects in the wild (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "65. Wu, H., Hu, Z., Li, L., Zhang, Y., Fan, C., Yu, X.: Nefii: Inverse rendering for reflectance decomposition with near-field indirect illumination. In: CVPR. pp. 4295-4304 (June 2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 403 + ], + "type": "text", + "content": "66. Wu, L., Zhu, R., Yaldiz, M.B., Zhu, Y., Cai, H., Matai, J., Porikli, F., Li, T.M., Chandraker, M., Ramamoorthi, R.: Factorized inverse path tracing for efficient and accurate material-lighting estimation. In: ICCV. pp. 3848-3858 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "67. Xie, Y., Yuan, M., Dong, B., Li, Q.: Diffusion model for generative image denoising. arXiv preprint arXiv:2302.02398 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "68. Yao, Y., Zhang, J., Liu, J., Qu, Y., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf: Neural incident light field for physically-based material estimation. In: ECCV (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 456, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 480, + 489 + ], + "type": "text", + "content": "69. Ye, W., Chen, S., Bao, C., Bao, H., Pollefeys, M., Cui, Z., Zhang, G.: IntrinsicNeRF: Learning Intrinsic Neural Radiance Fields for Editable Novel View Synthesis. In: ICCV (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "text", + "content": "70. Yi, R., Tan, P., Lin, S.: Leveraging multi-view image sets for unsupervised intrinsic image decomposition and highlight separation. In: AAAI. vol. 34, pp. 12685-12692 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "text", + "content": "71. Yi, R., Zhu, C., Xu, K.: Weakly-supervised single-view image relighting. In: CVPR. pp. 8402-8411 (June 2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 544, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 567 + ], + "type": "text", + "content": "72. Yu, Y., Meka, A., Elgharib, M., Seidel, H.P., Theobalt, C., Smith, W.: Self-supervised outdoor scene relighting. In: ECCV (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "73. Yu, Y., Smith, W.A.: Inverserendernet: Learning single image inverse rendering. In: CVPR. pp. 3155-3164 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 588, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 643 + ], + "type": "text", + "content": "74. Yu, Y., Malik, J.: Recovering photometric properties of architectural scenes from photographs. In: Proceedings of the 25th Annual Conference on Computer Graphics and Interactive Techniques. p. 207-217. SIGGRAPH '98, Association for Computing Machinery, New York, NY, USA (1998). https://doi.org/10.1145/280814.280874, https://doi.org/10.1145/280814.280874" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "text", + "content": "75. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. NeurIPS (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Chen. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 392 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "76. Zhang, J., Yao, Y., Li, S., Liu, J., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neil++: Inter-reflectable light fields for geometry and material estimation. ICCV (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "text", + "content": "77. Zhang, K., Luan, F., Li, Z., Snavely, N.: Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In: CVPR (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 204 + ], + "type": "text", + "content": "78. Zhang, K., Luan, F., Wang, Q., Bala, K., Snavely, N.: PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In: CVPR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "text", + "content": "79. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "80. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph. 40(6) (dec 2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "81. Zhang, Y., Sun, J., He, X., Fu, H., Jia, R., Zhou, X.: Modeling indirect illumination for inverse rendering. In: CVPR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "text", + "content": "82. Zhu, J., Huo, Y., Ye, Q., Luan, F., Li, J., Xi, D., Wang, L., Tang, R., Hua, W., Bao, H., et al.: I2-sdf: Intrinsic indoor scene reconstruction and editing via raytracing in neural sdfs. In: CVPR. pp. 12489-12498 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "type": "text", + "content": "83. Zhu, J., Luan, F., Huo, Y., Lin, Z., Zhong, Z., Xi, D., Wang, R., Bao, H., Zheng, J., Tang, R.: Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In: SIGGRAPH Asia 2022 Conference Papers. ACM (2022), https://doi.org/10.1145/3550469.3555407" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "text", + "content": "84. Zhu, Y., Zhang, K., Liang, J., Cao, J., Wen, B., Timofte, R., Van Gool, L.: Denoising diffusion models for plug-and-play image restoration. In: CVPR. pp. 1219-1229 (2023)" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 374, + 91, + 447, + 102 + ], + "type": "text", + "content": "IntrinsicAnything" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks/61df44b9-0f88-4bc8-892b-514f717f8da8_content_list.json b/2024/Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks/61df44b9-0f88-4bc8-892b-514f717f8da8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f5bf912f954303b815b115fa8fc6d5992eacdd28 --- /dev/null +++ b/2024/Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks/61df44b9-0f88-4bc8-892b-514f717f8da8_content_list.json @@ -0,0 +1,1854 @@ +[ + { + "type": "text", + "text": "Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tingyu $\\mathrm{Qu}^{1}$ , Tinne Tuytelaars $^{2}$ , and Marie-Francine Moens $^{1}$", + "bbox": [ + 264, + 234, + 736, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Computer Science, KU Leuven", + "bbox": [ + 338, + 261, + 663, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Electrical Engineering, KU Leuven {tingyu.qu; tinne.tuytelaars; sien.moens}@kuleuven.be", + "bbox": [ + 295, + 277, + 705, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Mainstream parameter-efficient fine-tuning (PEFT) methods, such as LoRA or Adapter, project a model's hidden states to a lower dimension, allowing pre-trained models to adapt to new data through this low-rank bottleneck. However, PEFT tasks involving multiple modalities, like vision-language (VL) tasks, require not only adaptation to new data but also learning the relationship between different modalities. Targeting at VL PEFT tasks, we propose a family of operations, called routing functions, to enhance VL alignment in the low-rank bottlenecks. These feature routing functions adopt linear operations and do not introduce new trainable parameters. In-depth analyses are conducted to study their behavior. In various VL PEFT settings, the routing functions significantly improve performance of the original PEFT methods, achieving over $20\\%$ improvement on VQAv2 (RoBERTaLarge + ViT-L/16) and $30\\%$ on COCO Captioning (GPT2-medium + ViT-L/16). Also when fine-tuning a pretrained multimodal model such as CLIP-BART, we observe smaller but consistent improvements across a range of VL PEFT tasks. Our code is available at https://github.com/tingyu215/Routing_VLPEFT.", + "bbox": [ + 261, + 339, + 743, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Vision and Language $\\cdot$ Parameter-Efficient Fine-Tuning $\\cdot$ Low-Rank Approximation", + "bbox": [ + 261, + 589, + 740, + 617 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 642, + 375, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When fine-tuning pre-trained foundation models for downstream tasks, studies show that a relatively high accuracy can be achieved by acting on a much smaller dimension compared to the original dimension of the pre-trained model's feature space. The minimum dimension necessary to approximate the optimization problem of the downstream task is termed the intrinsic dimension [1, 21].", + "bbox": [ + 212, + 672, + 785, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The existence of such intrinsic dimension explains the success of parameter-efficient fine-tuning (PEFT) methods with low-rank bottlenecks for model fine-tuning. Two of the most widely applied techniques are Adapter [9] and LoRA [10]. We demonstrate the use of Adapter and LoRA within a Transformer block in Fig. 1. Despite being added at different positions within the Transformer block, Adapter and LoRA exhibit some key characteristics in their architectural", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "design. Specifically, they first map the hidden states $x_{H}$ to a lower dimension $r$ through a down-projection mapping $W_{down}$ . The down-projected states are then projected back to the original dimension $d$ via an up-projection mapping $W_{up}$ . In the PEFT context, only the parameters of the colored modules in Fig. 1 are updated, while the remaining parameters are unchanged. Since the intermediate dimension $r$ is much smaller than the original dimension $d$ , these methods significantly reduce the number of parameters that need to be updated.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In uni-modal tasks, where $x_{H}$ contains information from a single modality, $W_{down}$ effectively compresses $x_{H}$ into a feature space of lower dimension. This compression usually suffices provided that $r$ is larger than the intrinsic dimension of the downstream task [10]. However, PEFT methods with low-rank bottlenecks, such as Adapter and LoRA, are currently not designed for down-stream tasks involving another modality, such as vision-language (VL) tasks. Even for a pre-trained multimodal foundation model with $x_{H}$ containing information from both modalities, conceptually, directly compressing $x_{H}$ with $W_{down}$ poses challenges, as it is unclear how to effectively balance the two modalities and to enforce alignment between the modalities with this simple linear mapping. Therefore, in this paper we investigate whether it is possible to better route the features through the low-rank bottleneck with some extra guidance. An intuitive ap", + "bbox": [ + 212, + 252, + 488, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "proach is standard cross-attention. However, it introduces additional parameters through query/key/value projections. Moreover, softmax-based operations not only increase time complexity but are limited to dealing with features of length $>1$ . This is incompatible with the PEFT setting, where the goal is to minimize the number of trainable parameters and computational complexity, while maintaining versatility across different types of tasks (including those that use only the [CLS] features). In this work, we introduce feature routing functions in VL PEFT with low-rank bottlenecks. They rely on linear operations and do not introduce extra trainable parameters.", + "bbox": [ + 212, + 646, + 787, + 780 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions include:", + "bbox": [ + 238, + 782, + 522, + 796 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We identify the problem of feature learning in the low-rank bottlenecks of standard PEFT methods considering VL downstream tasks.", + "bbox": [ + 223, + 809, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aa0083a8b1b11a6e9def3f4d3f2507a2625b734bde455f26536c5c1e98bf09e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 295, + 787, + 398 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/17050b95a1c5c37324615644bcb41b4a7c635de76ed152ddfac9697bfd2efdf1.jpg", + "image_caption": [ + "Fig.1: Illustration of LoRA and Adapter used in a Transformer block. Only the green modules are updated in PEFT. We identify the low-rank bottleneck with the orange rectangle. The Adapters are sequentially added to the Transformer block.3" + ], + "image_footnote": [], + "bbox": [ + 501, + 401, + 787, + 520 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "T. Qu et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a family of parameterized operations in the low-rank bottleneck, termed routing functions, that improve the performance of VL tasks.", + "- We conduct extensive experiments on different architectures and tasks, and provide insights on using routing functions in VL PEFT tasks." + ], + "bbox": [ + 223, + 146, + 782, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 231, + 395, + 248 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adapter-based PEFT To reduce the tuning costs of pre-trained language models (LMs), the Adapter layer was first introduced in [9]. It first down-projects the model's hidden state, then up-projects it back to its original dimension, incorporating a nonlinear function after the down-projection. Freezing the pretrained weights of the LM and tuning them with sequentially added Adapter layers achieves performance comparable to full fine-tuning. Since then, various Adapter variants have been proposed to advance the field of PEFT. The focus of developing Adapter variants has been on improving efficiency [18,28,34] and exploring alternative architectures [8,14,24]. Despite being surpassed by its variants on various benchmarks, the original sequential Adapter is still widely used in practice and offers stable results across tasks.", + "bbox": [ + 212, + 265, + 784, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LoRA-based PEFT LoRA [10] adopts the same \"down-projection then up-projection\" approach as the Adapter, with no nonlinear function in the low-rank bottleneck. LoRA is added to the query and value mappings in the attention modules of the pre-trained LM. Important follow-up works include dynamic rank adjustment [38, 41] and quantization adaptation [5, 40]. Moreover, VeRA [17] and NOLA [16] further minimize the number of parameters by leveraging frozen random matrices. LoRA-based PEFT methods have become standard for tuning large language models and large multimodal foundation models.", + "bbox": [ + 212, + 431, + 784, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Other PEFT methods Beyond Adapter and LoRA, prompt-based tuning methods [13, 19, 22] have shown strong performance for tuning both language and vision models. LST [36] learns a compact side network during tuning. BitFit [2] only adjusts the bias terms during fine-tuning, and achieves comparable performance as other PEFT methods. Methods [8, 29, 39, 42] also explore combining PEFT methods to achieve superior performance.", + "bbox": [ + 212, + 551, + 784, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-Language PEFT VL-Adapter [37] is the pioneering work in the field of Vision-Language (VL) PEFT. It empirically studies the use of different PEFT methods on four different VL tasks in a multitask learning setting. VL-Adapter uses CLIP-BART, a similar model architecture as VL-T5 [33]. CLIP-BART uses a projection layer to a connect CLIP [31] vision encoder to BART [20] or T5 [33]. CLIP-BART is tuned using a combination of prompt tuning and other PEFT methods. We conduct extensive experiments using CLIP-BART and VL-Adapter as presented in Sec. 5. As an extension to VL-Adapter, VL-PET [11] combines an Adapter with a rather complicated granularity-controlled mechanism. I-Tuning [27] explores alternative model architectures for VL PEFT tasks, and proposes a parallel Adapter with cross-attention to integrate visual information into a pre-trained LM. Since our work is the first that introduces routing functions into VL PEFT, we limit the comparison of our method to", + "bbox": [ + 212, + 643, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Routing Functions For VL PEFT", + "bbox": [ + 506, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/883fbcf05965f4d4351d5f214883d81dbff7f4aa50fcdc28177442dfdb82da61.jpg", + "image_caption": [ + "Fig. 2: Illustration of our method. Left: Conventional PEFT methods with low-rank bottleneck first map the hidden states $x_{H}$ from a high dimension $d$ to a lower dimension $r$ via down-projection map $W_{down}$ . Then, it maps the hidden states back to the original dimension $d$ via $W_{up}$ . Middle: Our method utilizes the same architecture, but updates the features via a routing function in the low-rank bottleneck. Specifically in a VL task, for the features $x_{R}$ that we want to align to, we use $W_{down}$ to down-project them as $W_{down}x_{R}$ . Then the routing function routes $W_{down}x_{R}$ and $W_{down}x_{H}$ in the low-rank bottleneck. Right: Different routing functions. In Adapter, routing functions are added before nonlinear activation functions." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 380, + 320 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dba17a9499953a423af3f9a7fe68656dcec9b02526d8b98119082da6729b5c66.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 143, + 598, + 320 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e8600ef2c84b203c49cba6c89753629603e094503b30a84aa31dfb12caa91e9a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 143, + 787, + 320 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VL-Adapter and I-Tuning, leaving the exploration of more complicated PEFT methods like VL-PET to future work.", + "bbox": [ + 212, + 486, + 787, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 540, + 380, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2 (Left), standard PEFT modules with low-rank bottleneck, i.e. LoRA/Adapter, learn a down-projection mapping $W_{down}$ followed by an up-projection mapping $W_{up}$ , with no routing in between. These modules are added in each Transformer block of the pre-trained foundation model (Fig. 1). As presented in Fig. 2 (Middle), our method adds a routing function in the low-rank bottleneck. This routing function is designed to align two sets of features, in particular, the extra visual features $x_{R}$ to be integrated in the VL tasks and the hidden states $x_{H}$ of each Transformer block of the foundation model.", + "bbox": [ + 212, + 571, + 787, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The design of a routing function follows two 'parameter-efficient' principles:", + "bbox": [ + 238, + 691, + 784, + 708 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A linear operation is used to reduce computation overload.", + "- No extra trainable parameters are introduced." + ], + "bbox": [ + 223, + 715, + 666, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we explore four types of routing functions. Given $x_{H}$ and $x_{R}^{4}$ , after the $W_{down}$ mapping, we obtain $x_{t} = W_{down}x_{H}$ and $x_{v} = W_{down}x_{R}$ . Here $x_{t}$ is of shape $(L_{t},r)$ and $x_{v}$ is of shape $(L_{v},r)$ , where $L_{t} / L_{v}$ and $r$ are the feature length (i.e. the number of tokens) and the bottleneck dimension or rank,", + "bbox": [ + 212, + 755, + 787, + 816 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "T. Qu et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "Assuming $x_{R}$ to have the same dimension $d$ as $x_{H}$ . See Sec. 4 and Sec. 5 for details.", + "bbox": [ + 217, + 824, + 784, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "respectively. $^5$ In practice, we always use $L_v = 1$ , i.e. we use a single token / global feature vector to represent the visual content. In case of multiple tokens, we use average pooling to reduce it to a single feature vector. $^6$ The proposed routing functions operate as:", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Element-wise multiplication: $x_{t} \\circ x_{v}'$ , where $x_{v}'$ denotes $x_{v}$ . reshape_as( $L_{t}, r$ )7", + "- Element-wise addition: $x_{t} + x_{v}^{\\prime}$ , where $x_{v}^{\\prime}$ denotes $x_{v}$ .reshape_as(L_t, r)", + "- Matrix multiplication 1: $x_{t}(x_{v})^{\\top}x_{v}$", + "- Matrix multiplication 2: $x_{t}x_{v}^{\\prime \\prime}$ , where $x_{v}^{\\prime \\prime}$ denotes $x_{v}$ .reshape_as(r,r)" + ], + "bbox": [ + 225, + 215, + 784, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitively, $x_{t} \\circ x_{v}^{\\prime}$ , reinforces or attenuates individual elements of $x_{t}$ , using the corresponding element of $x_{v}$ as scaling factor. The second routing function, $x_{t} + x_{v}^{\\prime}$ replaces $x_{t}$ with the sum of $x_{v}^{\\prime}$ and $x_{t}$ , thus moving $x_{t}$ closer to $x_{v}$ . $x_{t}(x_{v})^{\\top}x_{v}$ projects $x_{t}$ onto $x_{v}$ , i.e. only keeping the component of $x_{t}$ that is aligned with $x_{v}$ . Finally, $x_{t}x_{v}^{\\prime\\prime}$ replaces $x_{t}$ with a rescaled version of $x_{v}$ , with a different scale factor for each row given by the sum over all elements of the corresponding row of $x_{t}$ . In all these cases, it should be kept in mind that both $x_{v}$ and $x_{t}$ are derived from $x_{H}$ and $x_{R}$ via the down-projection $W_{down}$ . Parameters of $W_{down}$ can be learned so as to select particular dimensions of the original $x_{H}$ and $x_{R}$ , on which to apply the above operations before up-projecting them again with $W_{up}$ .", + "bbox": [ + 212, + 287, + 784, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The features $x_{R}$ contain information from the modalities to which we aim to align $x_{H}$ . The choice of $x_{R}$ can be different for different architectures, for instance, [CLS] features from a ViT in case of an encoder-only or decoder-only LM. We further explain the motivation of the choices of $x_{R}$ for different types of architectures in Sec. 4.1 and Sec. 5.1.", + "bbox": [ + 212, + 455, + 784, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Experiments with En(De)coder-Only Language Models", + "text_level": 1, + "bbox": [ + 214, + 554, + 782, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To explore the value of routing functions, in this section, we start from encoder-only and decoder-only language models (LMs) and incorporate visual information from pre-trained vision encoders for vision-language (VL) parameter-efficient fine-tuning (PEFT) tasks. The section begins with an overview of our experimental design and implementation details, and concludes with a discussion on the comprehensive experimental results obtained.", + "bbox": [ + 212, + 587, + 782, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Experimental Design", + "text_level": 1, + "bbox": [ + 214, + 700, + 434, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we study VL PEFT tasks in their cleanest form. Specifically, (1) No additional mapping networks are adopted, so the feature routing functions form the only module that contributes to the final performance; (2) Only the parameters of the PFET modules in LMs are updated, where we learn the VL alignment; (3)", + "bbox": [ + 212, + 726, + 782, + 787 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Routing Functions For VL PEFT", + "bbox": [ + 506, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "Here we omit the batch dimension for simplicity.", + "bbox": [ + 217, + 796, + 557, + 811 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "6 Only when comparing to cross-attention in Sec. 4.3, we do not use pooled features.", + "bbox": [ + 220, + 811, + 782, + 824 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "e.g. $x_{v}$ . reshape_as( $L_{t}, r$ ): $x_{v}' = \\mathbb{1}^{\\mathsf{T}}x_{v}$ , where $\\mathbb{1}$ is a matrix of ones of shape $(L_{v}, L_{t})$ .", + "bbox": [ + 222, + 824, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f16742f4e8d68edc424b1d0026d006f4863c7e9c61553f2d167ef69bc0189826.jpg", + "image_caption": [ + "Fig.3: Pipeline for VQA and Image Captioning. Only the added classifier (in VQA) and the PEFT modules of the RoBERTa encoder and GPT decoder are tuned. Note that here $x_{H}$ are the text inputs with the visual [CLS] features prepended." + ], + "image_footnote": [], + "bbox": [ + 218, + 142, + 401, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/139a592d9cee0ec8b3811ec9d38446f4adf950a5aabdf741c852d3e6e00147a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 142, + 624, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/af2c687005cf2f604c5af07550a389bb4546c179772493a455da5d749d226dc7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 142, + 785, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "No multimodal tasks (e.g. contrastive language-image pre-training in CLIP [31]) are included in the pre-training phases of the backbone models. We only learn the multimodal embedding space during VL PEFT. These principles guarantee that VL alignment originates strictly from the PEFT modules.", + "bbox": [ + 212, + 376, + 782, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To comprehensively study the role of routing functions in VL tasks, we consider both discriminative and generative VL tasks. We select visual question answering (VQA) and Image Captioning, two of the most widely studied VL tasks, as the representative task for discriminative VL tasks and generative VL tasks, respectively.", + "bbox": [ + 212, + 436, + 784, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present the pipelines for VQA and Image Captioning in Fig. 3. We use ViT [6] as our vision backbone. As for the LM backbone, we use the encoder-only RoBERTa [25] and the decoder-only GPT2 [32] for VQA and Image Captioning, respectively. The [CLS] feature from the ViT is prepended to the text inputs and acts as $x_{R}$ , enabling the learning of VL alignment through PEFT, with or without the integration of our routing functions. Prepending visual features to text inputs is actually the predominant way of learning VL alignment in PEFT [11, 37] without routing functions. The ViT and LMs adopt the same dimensions for their feature spaces, so we do not need additional mappings to match the dimensions.", + "bbox": [ + 212, + 513, + 787, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 686, + 455, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adopt ViT-B/16 pre-trained on ImageNet-21k as our vision backbone, which we keep frozen during training. In LoRA, we set scaling factor $\\alpha = r$ in all setups. For VQAv2, we use RoBERTa $_{\\text{base}}$ as our LM backbone, with a two-layer multi-layer perceptron (MLP) classifier on top following [37]. We train our model using the AdamW optimizer [26] with learning rate=3e-4 and batch size of 512 for 20 epochs, where we warm up the first $5\\%$ of the steps. For the captioning task, COCO Cap., we use GPT2-base as our LM backbone. We train our model for 20", + "bbox": [ + 212, + 710, + 787, + 816 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "T. Qu et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "8 In Appendix B, we report results obtained with larger backbones, e.g. RoBERTalarge.", + "bbox": [ + 217, + 824, + 784, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "epochs using the AdamW optimizer with batch size of 40 and learning rate $= 3\\mathrm{e} - 5$ where we warm up the first 5000 steps. For more details, see Appendix A.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Results", + "text_level": 1, + "bbox": [ + 215, + 196, + 323, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall performance on VQA We first evaluate our routing functions on the discriminative task VQAv2 with RoBERTabase and ViT-B/16, see Tab. 1.", + "bbox": [ + 212, + 219, + 782, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Without the addition of routing functions, models with LoRA and Adapter reach a similar level of performance. All four types of routing functions significantly improve the performance of the original PEFT method in terms of accuracy of the answer. The dimension of the bottleneck does not have a strong impact on the final performance. Sometimes we can even obtain better performance with lower dimensions. It is a sign that the intrinsic dimension of the VL task is not", + "bbox": [ + 212, + 250, + 486, + 445 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a0c9b21fb5aeb52d7b93f889656a8b2184fd5122a4c14b7caf22d7e1c1bc0b5c.jpg", + "table_caption": [ + "Table 1: Accuracy and improvement over the baseline (in blue) obtained by different types of routing functions on VQAv2. The best results and second best results are in bold and underlined, respectively." + ], + "table_footnote": [], + "table_body": "
Routingr = 64r = 128
FunctionsLoRAAdapterLoRAAdapter
None44.1544.1644.4544.28
xt○xt'53.51+9.3652.78+8.6253.86+10.4153.01+8.73
xt+xt'52.60+8.4553.94+9.7852.88+8.4353.95+9.67
xt(xt)Txt53.88+9.7354.48+10.3253.09+8.6455.06+10.78
xtxt''54.21+10.0654.96+10.8051.88+7.4354.38+10.10
", + "bbox": [ + 501, + 364, + 782, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "high, just as observed in uni-modal tasks. Moreover, we observe that the routing functions involving matrix multiplications (i.e. $x_{t}(x_{v})^{\\top}x_{v} \\& x_{t}x_{v}^{\\prime \\prime}$ ) perform quite well in most scenarios for both LoRA and Adapter, while the routing functions involving element-wise operations (i.e. $x_{t} \\circ x_{v}^{\\prime} \\& x_{t} + x_{v}^{\\prime}$ ) are just slightly behind.", + "bbox": [ + 212, + 446, + 784, + 508 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/67bcb20023d60fedbc0063e5f6309306ab792321adeda9e298c7e1b203b6af90.jpg", + "table_caption": [ + "Table 2: Results in terms of accu- Table 3: Results in terms of accuracy with racy with different routing functions on different routing functions on VQAv2 sorted VQAv2 sorted by answer type. The \"all\" by question type. The \"all\" column reports column reports overall accuracy. overall accuracy." + ], + "table_footnote": [], + "table_body": "
Routing FunctionsallLoRA, r = 128Adapter, r = 128
yes/no numberotherallyes/no numberother
None44.4568.6731.2929.2044.2868.3831.19
xt·ox'53.8672.2935.3644.4853.0171.7233.34
xt+xt'52.8872.0334.2043.0053.9572.4335.38
xt(xt)Tx53.0970.6335.4044.1955.0674.0933.59
xtx''51.8869.9233.2942.8254.3873.0033.94
", + "bbox": [ + 218, + 601, + 478, + 660 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0182d84ccf813c27181fee24a511aa132aa464e1f4dee6b7b371d06ecba3af94.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Routing \nFunctionsLoRA, r = 128Adapter, r = 128
allwhere colortimepersonallwhere colortimeperson
None44.4522.6440.0014.2945.1744.2822.9340.1513.7044.94
xt × xv'53.8635.7051.6923.0762.7453.0135.5952.6320.7362.28
xt + xv'52.8836.1249.3321.2262.9553.9537.0852.0821.6663.76
xt(x)Tx53.0936.3352.8622.2961.7455.0636.9253.6624.4463.85
xtxv''51.8834.1951.6022.6361.1454.3835.6452.7721.8562.50
", + "bbox": [ + 485, + 601, + 782, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To dive deeper into the behaviors of different types of routing functions, we evaluate the performance on different answer types as presented in Tab. 2. The largest improvements are obtained for the \"other\" category for all four routing functions. It can be seen that different routing functions excel for different answer types. For example, even when the two element-wise operations reach lower overall accuracy, they lead to higher or comparable accuracy on number-type answers. As for the routing functions involving matrix multiplications, better performance on other-type answers is obtained. For instance, with relatively lower overall accuracy reached in LoRA, matrix multiplications reach comparable performance on other-type answers. No clear pattern can be observed for yes/no", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Routing Functions For VL PEFT", + "bbox": [ + 506, + 113, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/890546d45b1d11b6ac407a02b92e5a6badc067e390e06b227a9f4c5a37dca6fd.jpg", + "table_caption": [ + "Table 4: Results of different types of routing functions on COCO Cap. The Avg. score is the averaged result from BLEU-4, METEOR, ROUGE-L, CIDEr, and SPICE. The best and second best results are in bold and underlined, respectively." + ], + "table_footnote": [], + "table_body": "
Routing FunctionsLoRA, r=64LoRA, r=128
BLEU-4METEORROUGE-LCIDErSPICEAvg.BLEU-4METEORROUGE-LCIDErSPICEAvg.
None18.317.436.855.811.828.0221.020.943.470.314.734.06
xt×ov'26.1+7.823.7+6.348.0+11.888.7+32.917.3+5.540.88+12.8625.4+4.423.7+2.848.2+4.887.9+17.617.7+3.040.58+6.52
xt+xov'22.2+3.920.3+2.941.0+4.273.5+17.714.8+3.034.36+6.3425.4+4.423.1+2.247.1+4.787.1+16.817.0+2.339.94+5.88
xt(xv)Txv24.8+6.522.6+5.245.4+8.684.9+29.116.8+5.038.90+10.8824.9+3.922.7+1.845.4+2.083.8+13.516.9+2.238.74+4.68
xtx''23.9+5.621.9+4.543.9+7.180.5+14.716.2+4.437.28+9.2625.5+4.522.9+2.046.0+2.685.2+14.917.2+2.539.36+5.30
Routing FunctionsAdapter, r=64Adapter, r=128
BLEU-4METEORROUGE-LCIDErSPICEAvg.BLEU-4METEORROUGE-LCIDErSPICEAvg.
None15.918.537.061.614.129.4216.118.436.261.513.929.22
xt×ov'24.6+8.723.1+4.646.4+9.484.5+22.917.2+3.139.16+9.7423.6+7.522.6+4.244.8+8.682.8+21.317.1+3.238.18+8.96
xt+xov'21.0+5.121.5+3.042.4+5.475.0+14.416.0+1.935.18+5.7619.3+3.220.8+2.440.9+4.775.2+13.816.2+2.334.48+5.26
xt(xv)Txv26.1+10.223.2+4.746.9+9.985.4+23.817.3+3.239.78+10.3624.1+8.022.4+4.044.2+8.082.7+21.216.9+3.038.06+8.84
xtx''26.6+10.723.0+4.546.8+9.885.8+24.217.2+3.139.88+10.4626.0+9.923.5+5.146.8+10.686.3+24.817.8+3.940.08+10.86
", + "bbox": [ + 217, + 196, + 784, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "types of answers, which could be explained by the large variety in content of these questions.", + "bbox": [ + 212, + 354, + 785, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 3 shows our method's performance across various question types. Four distinct question types test the method's reasoning capabilities concerning fine-grained visual information: spatial information (\"where\", e.g. Where is the surfboard?), color information (\"color\", e.g. What color is the bike?), time information (\"time\", e.g. What time is on the clock?), and person information (\"person\", e.g. What is the man holding?).9 Adding routing functions significantly improves the performance of LoRA and Adapter. Notably, different routing functions seem to capture different aspects of the visual information. $x_{t} + x_{v}^{\\prime}$ shows great performance while reasoning on spatial information (where) and person information (person). It is assumed that the addition operator reinforces feature values of content both present in the question and image, which might be beneficial for spatial and person questions. Similarly, $x_{t}x_{v}^{\\prime\\prime}$ performs exceptionally well for time information (time). With $x_{t} \\circ x_{v}^{\\prime}$ , we obtain relatively better performance on the color information (color). $x_{t}(x_{v})^{\\top}x_{v}$ presents consistent strong abilities while aligning spatial information (where) and color information (color).", + "bbox": [ + 212, + 386, + 787, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Appendix B, we show an additional ablation study on replacing visual features $(x_{R})$ with random noise or all-ones features. We also compare inference time of LoRA/Adapter with LoRA/Adapter + routing functions, where we observe adding routing functions to PEFT methods does not reduce efficiency.", + "bbox": [ + 212, + 614, + 787, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall performance on image captioning Evaluation of image captioning relies on the overall assessment of the generated captions. We report multiple text generation metrics and their average scores in Tab. 4. The addition of routing functions again leads to significant performance gains for all metrics in all scenarios. As was the case with VQAv2, different patterns are observed. Changing the intermediate dimension $r$ from 128 to 64 seems to have a larger impact on the model performance with conventional LoRA, where without the routing functions the CIDEr score drops from 70.3 to 55.8. In contrast, models augmented with routing functions remain robust and per", + "bbox": [ + 212, + 676, + 787, + 811 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "T. Qu et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "9 The comprehensive list of question types is included in Appendix A.", + "bbox": [ + 217, + 824, + 686, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "formant across setups, despite the changes in intermediate dimension. $x_{t} \\circ x_{v}^{\\prime}$ and $x_{t}x_{v}^{\\prime \\prime}$ perform well, possibly due to their alignment power. We also report the performance of Adapter with larger backbones, namely GPT2-medium and ViT-L/16, in Appendix B. Consistent and significant improvements in performance are observed, showing the efficacy of our routing functions on the task.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Comparison to cross-attention", + "text_level": 1, + "bbox": [ + 214, + 238, + 485, + 252 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Here we compare the results of the integration of routing function $x_{t}(x_{v})^{\\top}x_{v}$ with these of integrating cross-attention between the textual and the visual representation in image captioning. Indeed, in principle, a cross-attention function can also be considered as a kind of routing function with nonlinear operations.", + "bbox": [ + 212, + 253, + 486, + 387 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "I-Tuning [27] implements an Adapter with cross-attention for image captioning. To have a fair comparison, we", + "bbox": [ + 212, + 388, + 500, + 435 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "follow the settings of I-Tuning for both LoRA and Adapter. Specifically, (1) instead of using the visual features, we pretend the \"\" token to the input text. (2) We use the last hidden state from ViT as key and value for cross-attention and as $x_{R}$ for routing functions. This change is needed, since the [CLS] feature is of length 1, which does not work for cross-attention due to the Softmax function.", + "bbox": [ + 212, + 435, + 785, + 508 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the case of LoRA, we also implement a variant with separate $W_{down}$ mappings for $x_{R}$ and $x_{H}$ . This brings LoRA with routing function closer to LoRA with cross-attention. Without separate mapping refers to the setting used in previous experiments, where $x_{R}$ and $x_{H}$ share the same $W_{down}$ . As shown in Tab. 5, despite of having less parameters, $\\mathrm{LoRA} + x_t(x_v)^{\\top}x_v$ (w/ separate mapping) achieves significantly better performance than LoRA with cross-attention. The setup of $\\mathrm{LoRA} + x_t(x_v)^{\\top}x_v$ (w/o separate mapping) outperforms LoRA with cross-attention by a large margin using a similar amount of parameters.", + "bbox": [ + 212, + 511, + 787, + 632 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For Adapter, the cross-attention module has the same intermediate dimension as the $W_{down}$ in Adapter with routing function. Naturally, the separate query/ key/value mappings make that the Adapter with cross-attention has more than double the number of parameters compared to the Adapter equipped with the routing function $(x_{t}(x_{v})^{\\top}x_{v})$ . Both achieve comparable performance (Tab. 5).", + "bbox": [ + 212, + 633, + 787, + 708 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To conclude, Tab. 5 shows that integrating $x_{t}(x_{v})^{\\top}x_{v}$ in LORA and Adapter yields better performance given the same amount of parameters, or is similar in performance while using substantially less parameters. For VL tasks like image captioning, integrating a routing function to guide feature representations in low-rank bottlenecks has superior abilities compared to cross-attention.", + "bbox": [ + 212, + 709, + 787, + 785 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 5: Comparing routing functions with cross-attention on COCO Cap. †: parallel Adapter as used in [27]. Separate Map.: separate mapping for query/ key/value (in cross-attn.) or features $x_{R} / x_{H}$ (in $x_{t}(x_{v})^{\\top}x_{v}$ ). Here $r = 128$ . Avg. indicates the same average scores as in Tab. 4. Full evaluation in Appendix B.", + "bbox": [ + 496, + 266, + 787, + 376 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/4580eae2750b78f10347aeb2f6042918f34239ab5ac180d3d19ad3360cf307a1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PEFTSeparateMap.AlignmentParam.BLEU-4CIDErAvg.
LoRA✓Cross-attn.4.786M28.792.243.02
LoRA✓xt(xt)Txtxt3.932M30.7+2.099.4+7.245.68+2.66
LoRAXxt(xt)Txtxt4.746M30.0+1.399.0+6.845.38+2.36
Adapter†✓Cross-attn.4.732M30.799.845.70
Adapter†Xxt(xt)Txtxt1.830M30.8+0.198.8-1.045.48-0.22
", + "bbox": [ + 501, + 377, + 784, + 429 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Routing Functions For VL PEFT", + "bbox": [ + 506, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "10 Appendix A shows detailed architectures and explains why routing functions outperform cross-attention. We down-project to $r/4$ and $r$ for LoRA w/ & w/o separate mapping, respectively, to have the same level of parameter count.", + "bbox": [ + 214, + 796, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Incorporating nonlinear activation functions into routing functions We have proposed linear routing functions for efficiency reasons, but they do not need to be exclusively linear. Taking inspirations from the linear attention variants [4,15,30], where the linear operation is often equipped with nonlinear activation functions to ensure non-negativity for", + "bbox": [ + 212, + 146, + 486, + 296 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/6f8e15cacab643450ff81922e76895cbfc4e0e7232aa13815a87618dac799148.jpg", + "table_caption": [ + "Table 6: Experiments on adding nonlinear activation functions in routing functions. We use $r = 64$ here." + ], + "table_footnote": [], + "table_body": "
PEFT MethodRouting FunctionsVQA AccuracyCOCO Cap.
BLEU-4CIDErAvg.
LoRAxt(xv)Txv53.8824.884.938.90
LoRAReLU(xt)ReLU(xv)Txv53.4525.386.139.84
Adapterxt(xv)Txv54.4826.185.439.78
AdapterReLU(xt)ReLU(xv)Txv54.6325.084.739.02
", + "bbox": [ + 501, + 203, + 784, + 263 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "similarity scores between query and key, we explore whether a similar operation affects the representation power of the routing functions in the low-rank bottleneck. For routing function $x_{t}(x_{v})^{\\top}x_{v}$ , we use the nonlinear activation function $ReLU(\\cdot)$ as in previous work [30], and transform it to $ReLU(x_{t})ReLU(x_{v})^{\\top}x_{v}$ . The results are presented in Tab. 6. With or without $ReLU(\\cdot)$ , our routing function achieves similar performance on VQA and image captioning.", + "bbox": [ + 212, + 297, + 782, + 388 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Qualitative Analysis We first qualitatively compare the VL alignment pattern of LoRA-tuned GPT2 with or without a routing function on image captioning. An example is presented in Fig. 4. With the addition of routing function $x_{t}(x_{v})^{\\top}x_{v}$ , the model pays more attention to the nouns (e.g. man, a big screen) and the verbs (e.g. watching), which are essential in learning the VL alignment. At the same time, it pays less attention to meaningless words like \"something\". This is in line with the finding that the routing function stresses components of $x_{t}$ that are aligned in the image.", + "bbox": [ + 212, + 392, + 787, + 513 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/197843bbade820f1df61436b07e3b53b5a7ea586049e2db61d95ece54853f6fe.jpg", + "image_caption": [ + "Fig. 4: Example of average attention weights from the last layer of GPT2. We use the final checkpoint trained on COCO Cap. with $r = 64$ . IMAGE: visual [CLS] feature." + ], + "image_footnote": [], + "bbox": [ + 225, + 546, + 718, + 642 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Next we qualitatively examine some examples from the inference phase for VQA and Image Captioning. We present examples in Fig. 5 where conventional LoRA fails. For VQA, previous experiments in Tab. 3 indicate that our routing functions work particularly well on color related questions. As also shown in Fig. 5, conventional LoRA fails to align the color red to the lollipop. The model tuned with conventional LoRA also fails to correctly answer other types of questions as presented in the figure. Without routing functions, captions generated from the LM tuned with conventional LoRA often make mistakes in compre", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "T. Qu et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/845e1ec6070b8ec8a5d1018e8d89fc7f390c5db4aef07b82a89cf8399345cf3c.jpg", + "image_caption": [ + "Fig. 5: Qualitative examples from VQA and Image Captioning tasks. We present cases where the model trained with conventional LoRA fails and results obtained from a model trained with LoRA with a routing function. We use $r = 64$ ." + ], + "image_footnote": [], + "bbox": [ + 250, + 146, + 465, + 333 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0053a476f1e008c39268be92e8b5647fc7c36b2e00fb2160454f9738e25778ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 146, + 754, + 333 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "hending the visual information. Occasionally, we observe a sequence of blank spaces as presented in the first example using conventional LoRA.", + "bbox": [ + 212, + 410, + 784, + 443 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5 Experiments with Encoder-Decoder Language Models", + "text_level": 1, + "bbox": [ + 214, + 462, + 772, + 479 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we study the VL PEFT tasks in a more complicated setting, where we consider an encoder-decoder LM with an additional mapping network dedicated for the vision encoder. The model is trained in a multi-task learning setting, where four different VL tasks are jointly learned, as in [3,37]. Despite the increased complexity of optimizing the network in this scenario, we show that integrating routing functions into the low-rank bottlenecks yields performance enhancements. Moreover, consistent patterns are observed for different scenarios, which shed lights on future research to improve VL PEFT tasks.", + "bbox": [ + 212, + 491, + 787, + 612 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Experimental Design", + "text_level": 1, + "bbox": [ + 214, + 632, + 434, + 648 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Following [37], we use the CLIP-BART architecture to study the effect of our routing functions on the VL PEFT tasks in a multitask learning setting. CLIP-BART is a widely used architecture for VL tasks. We present the simplified model architecture of CLIP-BART with Adapters and the integrated routing functions in Fig. 6. The original CLIP-BART uses a trainable visual projection module to map the visual features extracted by CLIP to the embedding space of BART [20]. CLIP-BART uses task-specific prompts as the textual input: For example, in Fig. 6, \"\" is added as prompt to the text for VQA. The mapped visual features are then prepped to the textual input. Together, they serve as the input to BART. CLIP-BART is trained and evaluated in a multitask setting, that is, VQAv2 [7] and GQA [12] for visual question answering, NLVR $^2$ [35] using natural language for visual reasoning, and COCO Cap. [23] for image captioning.", + "bbox": [ + 212, + 655, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Routing Functions For VL PEFT", + "bbox": [ + 506, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "CLIP-BART unifies the four tasks into one text generation problem following [3], where four different task prompts distinguish the task types. CLIP-BART can be considered as the representative of the current state-of-the-art on VL PEFT tasks.", + "bbox": [ + 212, + 146, + 486, + 251 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To integrate our routing functions into the model, two types of inputs $x_{R}$ can be considered, namely the visual features after the visual projection, and the last encoder hidden state as presented in Fig. 6. The visual features only contain visual information. Because the last encoder hidden state contains both textual and visual information, it can only be added to the decoder. We experiment with both LoRA and Adapter as the PEFT method for fine-tuning CLIP-BART. In this multitask learning setting, one can use (1) a single LoRA/Adapter shared across tasks or (2) multiple LoRAs/Adapters without parameter", + "bbox": [ + 212, + 252, + 486, + 508 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "sharing, where each task has its own LoRA/Adapter. We start by exploring the impact of different routing functions on the single LoRA/Adapter setting. Then, based on the observations obtained in the setting with single Adapter, we explore ways of using the routing functions in the multiple Adapters setting.", + "bbox": [ + 212, + 508, + 785, + 570 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a5e03c18dc9aeea9b48c2643a0f2cf2d65354fe1e58ae3fafc24bb29369af54e.jpg", + "image_caption": [ + "Fig. 6: Architecture of CLIP-BART with Adapter. The standard Adapter architecture is adopted (see Fig. 1). Only the modules in green are trainable. To apply routing functions, we use the pooled last encoder hidden state as feature $x_{R}$ , which is added to the Adapters in the decoder." + ], + "image_footnote": [], + "bbox": [ + 501, + 160, + 787, + 364 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 590, + 455, + 606 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Following [11, 37], we use CLIP-ResNet101 as our vision encoder [31]. We use $\\mathrm{BART}_{\\mathrm{base}}$ as our encoder-decoder LM as in [37]. Following [37], we use $r = 96$ for all experiments. The training pipeline in [37] is used for a fair comparison. Specifically, (1) we use the same seed for all experiments. (2) Due to limited resources, we do not perform an extensive hyper-parameter search as in [37]. (3) We train the model with AdamW optimizer and linear decay for 20 epochs. We warm-up the learning rate from 0 to the highest in the first 2 epochs. (4) The batch size is set to 500. For more details, we refer to Appendix A.", + "bbox": [ + 212, + 613, + 787, + 737 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Results", + "text_level": 1, + "bbox": [ + 214, + 755, + 323, + 768 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Experiments with a single Adapter/LoRA We first report results of experiments on using single Adapter/LoRA shared across tasks in Tab. 7. In this case, the Adapter/LoRA is trained with data from four different tasks, and naturally we can only use one type of routing function throughout the experiments.", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "T. Qu et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e5243c7b6e0bd09ff3030241d721065cd986a0e299e3155fdd09529bb6279ce2.jpg", + "table_caption": [ + "Table 7: Experiments with different types of routing functions using CLIP-BART with single LoRA/Adapter. Following [11] $(\\dagger)^{11}$ , we report the average CIDER score for COCO Cap., accuracy for the other three tasks. We also report the average accuracy from VQA, GQA and NLVR2. We bold and underline the best and second best results." + ], + "table_footnote": [], + "table_body": "
Routing FunctionsSingle LoRASingle Adapter
VQAGQANLVR2Avg.COCOCap.VQAGQANLVR2Avg.
None†65.1553.6672.5863.80115.0165.7654.1673.1964.37
xt○x′v65.6853.9673.4264.35113.9465.9254.3474.2364.83
xt+x′v65.1453.7373.5164.13114.9665.8954.1873.9064.66
xt(xv)Txv64.9453.5673.6064.03117.8065.8453.6574.3164.03
xtx''v64.8453.1372.9863.65119.2665.8353.6173.2764.24
", + "bbox": [ + 220, + 212, + 782, + 308 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Tab. 7 shows that the integration of the routing functions in the single LoRa and single Adapter architecture results in improved performance in nearly all scenarios. But just like the experiments with encoder-only and decoder-only LMs in Sec. 4.3, different routing functions show different patterns. The routing functions involving matrix multiplication, that is, $x_{t}(x_{v})^{\\top}x_{v}$ and $x_{t}x_{v}^{\\prime \\prime}$ , perform exceptionally well in the COCO captioning task, while the other two routing functions involving element-wise operations, i.e. $x_{t}\\circ x_{v}^{\\prime}$ and $x_{t} + x_{v}^{\\prime}$ , show more potential on QA tasks (VQA/GQA). As for the NLVR $^{2}$ task, $x_{t}(x_{v})^{\\top}x_{v}$ and $x_{t}\\circ x_{v}^{\\prime}$ present stronger abilities to learn VL relationships. Note that achieving relatively low performance on one task does not necessarily mean that the corresponding routing function fails. For example, $x_{t}\\circ x_{v}^{\\prime}$ does not work well for captioning, but achieves great performance on the other three tasks. This somehow contrasts with the results obtained with encoder-only and decoder-only LMs in Sec. 4.3, where $x_{t}\\circ x_{v}^{\\prime}$ leads to one of the best performance on COCO Cap. The complexity of the multitask learning setup, where different loss functions are jointly optimized, might here be a confounding factor.", + "bbox": [ + 212, + 339, + 787, + 582 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Experiments with multiple Adapters We further explore the behavior of the routing functions in models with multiple Adapters, where we train a separate Adapter per task. As also observed in [37], using multiple Adapters increases the complexity of the model, and yields worse performance compared to using a single Adapter. As presented in Tab. 8, using the four routing functions with multiple Adapters leads to comparable or better performance as compared to multiple Adapters with no routing functions. Moreover, in line with what we discovered in the experiments with models with a single Adapter, different routing functions shine on different types of tasks.", + "bbox": [ + 212, + 583, + 787, + 718 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this set-up, we also experiment with combinations of routing functions. The four combinations integrate the routing functions that work best for each individual task (e.g. $x_{t}x_{v}^{\\prime}$ for COCO Cap.). The results in Tab. 8 evidence that in this multitask setting combinations of routing functions yield an improved performance compared to using no routing functions. GQA requires more fine", + "bbox": [ + 212, + 720, + 787, + 797 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Routing Functions For VL PEFT", + "bbox": [ + 506, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_footnote", + "text": "11 $\\dagger$ : 3 seeds. Appendix C shows that our improvements are significant using paired t-test with $p = 0.05$ .", + "bbox": [ + 214, + 810, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/4149d018561ae8aeea60c49b3d66bbf1def6fddb7948724a178c4d7e3173a6c0.jpg", + "table_caption": [ + "Table 8: Experiments with different combinations of routing functions using CLIP-BART with multiple Adapters. Following [37] $(\\ddagger)$ , we report CIDEr score for COCO Cap., accuracy for the other three tasks. We also report the average accuracy from VQA, GQA and NLVR2. We bold and underline the best and second best results." + ], + "table_footnote": [], + "table_body": "
Routing FunctionsRouting FunctionsPerformance
VQAGQANLVR2COCO Cap.VQAGQANLVR2Avg.COCO Cap.
None†XXXX65.454.069.863.07114.3
xt○x′v✓✓✓✓65.6153.5170.7863.30113.72
xt+x′v✓✓✓✓65.1953.8470.9963.34115.64
xt(xv)Txv✓✓✓✓64.9352.1170.2562.43116.20
xtx′v✓✓✓✓65.3252.6770.5862.86117.92
comb1xt○x′vxt○x′vxt(xv)Txvxtx′v66.3754.2169.2763.28116.94
comb2xt+x′vxt+x′vxt(xv)Txvxtx′v65.1453.3770.8863.13117.58
comb3xt○x′vxt○x′vxt○x′vxtx′v66.3453.7770.0263.38117.73
comb4xt+x′vxt+x′vxt○x′vxtx′v65.4752.9971.0163.16118.34
", + "bbox": [ + 218, + 212, + 782, + 359 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "grained image representation. For efficiency reason, we only adopt $x_{v}$ as the down-projection from the pooled last encoder hidden state. Nevertheless, comb1 can recover from this shortcoming. More detailed image representations as used in the cross-attention experiments of Tab. 5 might still enhance the results.", + "bbox": [ + 212, + 390, + 784, + 450 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Overall, our research has focused on introducing routing functions in low-rank feature approximations when fine-tuning models for downstream (in our case VL) tasks based on pre-trained foundation models. As a pioneering work on the topic, for obtaining results we did not exhaustively search for all possible choices of hyperparameters. Instead, we have focused on revealing the problem of low-rank bottlenecks and on providing a successful solution. We leave an exhaustive grid search for optimal hyperparameters as future work.", + "bbox": [ + 212, + 450, + 785, + 556 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 580, + 359, + 595 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have introduced feature routing functions to help guide the feature learning in low-rank bottlenecks in architectures that are typical for parameter-efficient fine-tuning such as LoRA and Adapter. We have shown their validity in vision-language tasks, where we have focused on linear functions that do not require any extra parameters. We have demonstrated performance gains in single-task and multi-task settings in combination with a variety of foundation models.", + "bbox": [ + 212, + 612, + 784, + 703 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Routing functions open many avenues for future research. Given the importance of foundation models that are trained in a self-supervised way on large data, there is the need to efficiently fine-tune them for downstream tasks. When computing low-rank approximations of the features, our work shows that the extra guidance of the routing functions is important. We see many possibilities in experimenting with other types of functions in other tasks. For instance, linear and other functions are also suited to model constraints that reflect external knowledge or expectations on the behavior of a model. Future work will show whether they are useful to guide the feature routing in low-rank bottlenecks.", + "bbox": [ + 212, + 704, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "T. Qu et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is funded by the Flanders AI Research Program, European Research Council Advanced Grant H2020-ERC-2017-ADG 788506, and the China Scholarship Council.", + "bbox": [ + 215, + 175, + 784, + 220 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 243, + 321, + 258 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aghajanyan, A., Gupta, S., Zettlemoyer, L.: Intrinsic dimensionality explains the effectiveness of language model fine-tuning. In: Zong, C., Xia, F., Li, W., Navigli, R. (eds.) Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 7319-7328. Association for Computational Linguistics, Online (Aug 2021). https://doi.org/10.18653/v1/2021.acl-long.568, https://aclanthology.org/2021.acl-long.568", + "2. Ben Zaken, E., Goldberg, Y., Ravfogel, S.: BitFit: Simple parameter-efficient fin-tuning for transformer-based masked language-models. In: Muresan, S., Nakov, P., Villavicencio, A. (eds.) Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). pp. 1-9. Association for Computational Linguistics, Dublin, Ireland (May 2022). https://doi.org/10.18653/v1/2022.acl-short.1, https://aclanthology.org/2022.acl-short.1", + "3. Cho, J., Lei, J., Tan, H., Bansal, M.: Unifying vision-and-language tasks via text generation. In: ICML (2021)", + "4. Choromanski, K.M., Likhosherstov, V., Dohan, D., Song, X., Gane, A., Sarlos, T., Hawkins, P., Davis, J.Q., Mohiuddin, A., Kaiser, L., Belanger, D.B., Colwell, L.J., Weller, A.: Rethinking attention with performers. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=Ua6zuk0WRH", + "5. Dettmers, T., Pagnoni, A., Holtzman, A., Zettlemoyer, L.: Qlora: Efficient finetuning of quantized llms. In: Oh, A., Neumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 10088-10115. Curran Associates, Inc. (2023), https://proceedings.neurips.cc/paper_files/paper/2023/file/1feb87871436031bdc0f2beaa62a049b-Paper-Conference.pdf", + "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=YicbFdNTTy", + "7. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2017)", + "8. He, J., Zhou, C., Ma, X., Berg-Kirkpatrick, T., Neubig, G.: Towards a unified view of parameter-efficient transfer learning. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=ORDcd5Axok", + "9. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for NLP. In: Chaudhuri, K., Salakhutdinov, R. (eds.) Proceedings of the 36th International" + ], + "bbox": [ + 223, + 273, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Routing Functions For VL PEFT", + "bbox": [ + 506, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 97, pp. 2790-2799. PMLR (09-15 Jun 2019), https://proceedings.mlr.press/v97/houlsby19a.html", + "10. Hu, E.J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=nZeVKeeFYf9", + "11. Hu, Z.Y., Li, Y., Lyu, M.R., Wang, L.: Vl-pet: Vision-and-language parameter-efficient tuning via granularity control. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3010-3020 (October 2023)", + "12. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)", + "13. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision (ECCV) (2022)", + "14. Karimi Mahabadi, R., Ruder, S., Dehghani, M., Henderson, J.: Parameter-efficient multi-task fine-tuning for transformers via shared hypernetworks. In: Zong, C., Xia, F., Li, W., Navigli, R. (eds.) Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 565-576. Association for Computational Linguistics, Online (Aug 2021). https://doi.org/10.18653/v1/2021.acl-long.47, https://aclanthology.org/2021.acl-long.47", + "15. Katharopoulos, A., Vyas, A., Pappas, N., Fleuret, F.: Transformers are RNNs: Fast autoregressive transformers with linear attention. In: III, H.D., Singh, A. (eds.) Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 5156-5165. PMLR (13-18 Jul 2020), https://proceedings.mlr.press/v119/katharopoulos20a.html", + "16. Koohpayegani, S.A., L, N.K., Nooralinejad, P., Kolouri, S., Pirsiavash, H.: NOLA: Networks as linear combination of low rank random basis. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=TjfXcDgvzk", + "17. Kopiczko, D.J., Blankevoort, T., Asano, Y.M.: VeRA: Vector-based random matrix adaptation. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=NjNfLdxr3A", + "18. Lei, T., Bai, J., Brahma, S., Ainslie, J., Lee, K., Zhou, Y., Du, N., Zhao, V.Y., Wu, Y., Li, B., Zhang, Y., Chang, M.W.: Conditional adapters: Parameter-efficient transfer learning with fast inference. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=IyYyKovOAj", + "19. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. In: Moens, M.F., Huang, X., Specia, L., Yih, S.W.t. (eds.) Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 3045-3059. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic (Nov 2021). https://doi.org/10.18653/v1/2021.emnlp-main.243, https://aclanthology.org/2021.emnlp-main.243", + "20. Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., Stoyanov, V., Zettlemoyer, L.: BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In: Jurafsky, D., Chai, J., Schluter, N., Tetreault, J. (eds.) Proceedings of the 58th Annual Meeting" + ], + "bbox": [ + 217, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "T. Qu et al.", + "bbox": [ + 271, + 114, + 351, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the Association for Computational Linguistics. pp. 7871-7880. Association for Computational Linguistics, Online (Jul 2020). https://doi.org/10.18653/v1/2020.acl-main.703, https://aclanthology.org/2020.acl-main.703", + "21. Li, C., Farkhoor, H., Liu, R., Yosinski, J.: Measuring the intrinsic dimension of objective landscapes. In: International Conference on Learning Representations (2018), https://openreview.net/forum?id=ryup8-WCW", + "22. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. In: Zong, C., Xia, F., Li, W., Navigli, R. (eds.) Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 4582-4597. Association for Computational Linguistics, Online (Aug 2021). https://doi.org/10.18653/v1/2021.acl-long.353, https://aclanthology.org/2021.acl-long.353", + "23. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) Computer Vision - ECCV 2014. pp. 740-755. Springer International Publishing, Cham (2014)", + "24. Lin, Z., Madotto, A., Fung, P.: Exploring versatile generative language model via parameter-efficient transfer learning. In: Cohn, T., He, Y., Liu, Y. (eds.) Findings of the Association for Computational Linguistics: EMNLP 2020. pp. 441-459. Association for Computational Linguistics, Online (Nov 2020). https://doi.org/10.18653/v1/2020-findings-emnlp.41, https://aclanthology.org/2020-findings-emnlp.41", + "25. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized BERT pretraining approach. CoRR abs/1907.11692 (2019), http://arxiv.org/abs/1907.11692", + "26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019), https://openreview.net/forum?id=Bkg6RiCqY7", + "27. Luo, Z., Hu, Z., Xi, Y., Zhang, R., Ma, J.: I-tuning: Tuning frozen language models with image for lightweight image captioning. In: ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5 (2023). https://doi.org/10.1109/ICASSP49357.2023.10096424", + "28. mahabadi, R.K., Henderson, J., Ruder, S.: Compacter: Efficient low-rank hypercomplex adapter layers. In: Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems (2021), https://openreview.net/forum?id=bqGK5PyI6-N", + "29. Mao, Y., Mathias, L., Hou, R., Almahairi, A., Ma, H., Han, J., Yih, S., Khabsa, M.: UniPELT: A unified framework for parameter-efficient language model tuning. In: Muresan, S., Nakov, P., Villavicencio, A. (eds.) Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 6253-6264. Association for Computational Linguistics, Dublin, Ireland (May 2022). https://doi.org/10.18653/v1/2022.acl-long.433, https://aclanthology.org/2022.acl-long.433", + "30. Qin, Z., Sun, W., Deng, H., Li, D., Wei, Y., Lv, B., Yan, J., Kong, L., Zhong, Y.: cosformer: Rethinking softmax in attention. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=B18CQrx2Up4", + "31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.)" + ], + "bbox": [ + 215, + 147, + 785, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Routing Functions For VL PEFT", + "bbox": [ + 506, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 8748-8763. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/radford21a.html", + "32. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I.: Language models are unsupervised multitask learners. In: Preprint (2019)", + "33. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research 21(140), 1-67 (2020), http://jmlr.org/papers/v21/20-074.html", + "34. Rücklé, A., Geigle, G., Glockner, M., Beck, T., Pfeiffer, J., Reimers, N., Gurevych, I.: AdapterDrop: On the efficiency of adapters in transformers. In: Moens, M.F., Huang, X., Specia, L., Yih, S.W.t. (eds.) Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 7930-7946. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic (Nov 2021). https://doi.org/10.18653/v1/2021.emnlp-main.626, https://aclanthology.org/2021.emnlp-main.626", + "35. Suhr, A., Zhou, S., Zhang, A., Zhang, I., Bai, H., Artzi, Y.: A corpus for reasoning about natural language grounded in photographs. In: Korhonen, A., Traum, D., Marquez, L. (eds.) Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. pp. 6418-6428. Association for Computational Linguistics, Florence, Italy (Jul 2019). https://doi.org/10.18653/v1/P19-1644, https://aclanthology.org/P19-1644", + "36. Sung, Y.L., Cho, J., Bansal, M.: LST: Ladder side-tuning for parameter and memory efficient transfer learning. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=isPnnaTZaP5", + "37. Sung, Y.L., Cho, J., Bansal, M.: Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5227-5237 (June 2022)", + "38. Valipour, M., Rezagholizadeh, M., Kobyzev, I., Ghodsi, A.: DyLoRA: Parameter-efficient tuning of pre-trained models using dynamic search-free low-rank adaptation. In: Vlachos, A., Augenstein, I. (eds.) Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics. pp. 3274-3287. Association for Computational Linguistics, Dubrovnik, Croatia (May 2023). https://doi.org/10.18653/v1/2023.eacl-main.239, https://aclanthology.org/2023.eacl-main.239", + "39. Wang, Y., Agarwal, S., Mukherjee, S., Liu, X., Gao, J., Awadallah, A.H., Gao, J.: AdaMix: Mixture-of-adaptations for parameter-efficient model tuning. In: Goldberg, Y., Kozareva, Z., Zhang, Y. (eds.) Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. pp. 5744-5760. Association for Computational Linguistics, Abu Dhabi, United Arab Emirates (Dec 2022). https://doi.org/10.18653/v1/2022.emnlp-main.388, https://aclanthology.org/2022.emnlp-main.388", + "40. Xu, Y., Xie, L., Gu, X., Chen, X., Chang, H., Zhang, H., Chen, Z., ZHANG, X., Tian, Q.: QA-loRA: Quantization-aware low-rank adaptation of large language models. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=WvFoJccpo8", + "41. Zhang, Q., Chen, M., Bukharin, A., He, P., Cheng, Y., Chen, W., Zhao, T.: Adaptive budget allocation for parameter-efficient fine-tuning. In: The Eleventh Inter-" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "T. Qu et al.", + "bbox": [ + 271, + 114, + 351, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "national Conference on Learning Representations (2023), https://openreview.net/forum?id=1q62uWRJjiY", + "42. Zhou, H., Wan, X., Vulic, I., Korhonen, A.: Autopeft: Automatic configuration search for parameter-efficient fine-tuning. Transactions of the Association for Computational Linguistics 12 (2024)" + ], + "bbox": [ + 217, + 147, + 784, + 217 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Routing Functions For VL PEFT", + "bbox": [ + 506, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks/61df44b9-0f88-4bc8-892b-514f717f8da8_model.json b/2024/Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks/61df44b9-0f88-4bc8-892b-514f717f8da8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cd55b5a3ec0e7a91c9e1714885225463f7fd5b8f --- /dev/null +++ b/2024/Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks/61df44b9-0f88-4bc8-892b-514f717f8da8_model.json @@ -0,0 +1,2471 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.22, + 0.141, + 0.784, + 0.207 + ], + "angle": 0, + "content": "Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks" + }, + { + "type": "text", + "bbox": [ + 0.266, + 0.235, + 0.737, + 0.251 + ], + "angle": 0, + "content": "Tingyu \\(\\mathrm{Qu}^{1}\\), Tinne Tuytelaars\\(^{2}\\), and Marie-Francine Moens\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.339, + 0.262, + 0.664, + 0.277 + ], + "angle": 0, + "content": "\\(^{1}\\) Department of Computer Science, KU Leuven" + }, + { + "type": "text", + "bbox": [ + 0.296, + 0.278, + 0.706, + 0.305 + ], + "angle": 0, + "content": "\\(^{2}\\) Department of Electrical Engineering, KU Leuven {tingyu.qu; tinne.tuytelaars; sien.moens}@kuleuven.be" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.34, + 0.744, + 0.577 + ], + "angle": 0, + "content": "Abstract. Mainstream parameter-efficient fine-tuning (PEFT) methods, such as LoRA or Adapter, project a model's hidden states to a lower dimension, allowing pre-trained models to adapt to new data through this low-rank bottleneck. However, PEFT tasks involving multiple modalities, like vision-language (VL) tasks, require not only adaptation to new data but also learning the relationship between different modalities. Targeting at VL PEFT tasks, we propose a family of operations, called routing functions, to enhance VL alignment in the low-rank bottlenecks. These feature routing functions adopt linear operations and do not introduce new trainable parameters. In-depth analyses are conducted to study their behavior. In various VL PEFT settings, the routing functions significantly improve performance of the original PEFT methods, achieving over \\(20\\%\\) improvement on VQAv2 (RoBERTaLarge + ViT-L/16) and \\(30\\%\\) on COCO Captioning (GPT2-medium + ViT-L/16). Also when fine-tuning a pretrained multimodal model such as CLIP-BART, we observe smaller but consistent improvements across a range of VL PEFT tasks. Our code is available at https://github.com/tingyu215/Routing_VLPEFT." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.59, + 0.741, + 0.618 + ], + "angle": 0, + "content": "Keywords: Vision and Language \\(\\cdot\\) Parameter-Efficient Fine-Tuning \\(\\cdot\\) Low-Rank Approximation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.643, + 0.377, + 0.659 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.75 + ], + "angle": 0, + "content": "When fine-tuning pre-trained foundation models for downstream tasks, studies show that a relatively high accuracy can be achieved by acting on a much smaller dimension compared to the original dimension of the pre-trained model's feature space. The minimum dimension necessary to approximate the optimization problem of the downstream task is termed the intrinsic dimension [1, 21]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The existence of such intrinsic dimension explains the success of parameter-efficient fine-tuning (PEFT) methods with low-rank bottlenecks for model fine-tuning. Two of the most widely applied techniques are Adapter [9] and LoRA [10]. We demonstrate the use of Adapter and LoRA within a Transformer block in Fig. 1. Despite being added at different positions within the Transformer block, Adapter and LoRA exhibit some key characteristics in their architectural" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "T. Qu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "design. Specifically, they first map the hidden states \\( x_{H} \\) to a lower dimension \\( r \\) through a down-projection mapping \\( W_{down} \\). The down-projected states are then projected back to the original dimension \\( d \\) via an up-projection mapping \\( W_{up} \\). In the PEFT context, only the parameters of the colored modules in Fig. 1 are updated, while the remaining parameters are unchanged. Since the intermediate dimension \\( r \\) is much smaller than the original dimension \\( d \\), these methods significantly reduce the number of parameters that need to be updated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.489, + 0.645 + ], + "angle": 0, + "content": "In uni-modal tasks, where \\( x_{H} \\) contains information from a single modality, \\( W_{down} \\) effectively compresses \\( x_{H} \\) into a feature space of lower dimension. This compression usually suffices provided that \\( r \\) is larger than the intrinsic dimension of the downstream task [10]. However, PEFT methods with low-rank bottlenecks, such as Adapter and LoRA, are currently not designed for down-stream tasks involving another modality, such as vision-language (VL) tasks. Even for a pre-trained multimodal foundation model with \\( x_{H} \\) containing information from both modalities, conceptually, directly compressing \\( x_{H} \\) with \\( W_{down} \\) poses challenges, as it is unclear how to effectively balance the two modalities and to enforce alignment between the modalities with this simple linear mapping. Therefore, in this paper we investigate whether it is possible to better route the features through the low-rank bottleneck with some extra guidance. An intuitive ap" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.647, + 0.788, + 0.781 + ], + "angle": 0, + "content": "proach is standard cross-attention. However, it introduces additional parameters through query/key/value projections. Moreover, softmax-based operations not only increase time complexity but are limited to dealing with features of length \\(>1\\). This is incompatible with the PEFT setting, where the goal is to minimize the number of trainable parameters and computational complexity, while maintaining versatility across different types of tasks (including those that use only the [CLS] features). In this work, we introduce feature routing functions in VL PEFT with low-rank bottlenecks. They rely on linear operations and do not introduce extra trainable parameters." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.783, + 0.524, + 0.797 + ], + "angle": 0, + "content": "In summary, our contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.81, + 0.789, + 0.84 + ], + "angle": 0, + "content": "- We identify the problem of feature learning in the low-rank bottlenecks of standard PEFT methods considering VL downstream tasks." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.296, + 0.789, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.402, + 0.789, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.542, + 0.788, + 0.626 + ], + "angle": 0, + "content": "Fig.1: Illustration of LoRA and Adapter used in a Transformer block. Only the green modules are updated in PEFT. We identify the low-rank bottleneck with the orange rectangle. The Adapters are sequentially added to the Transformer block.3" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.507, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Routing Functions For VL PEFT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.147, + 0.783, + 0.175 + ], + "angle": 0, + "content": "- We introduce a family of parameterized operations in the low-rank bottleneck, termed routing functions, that improve the performance of VL tasks." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.178, + 0.784, + 0.206 + ], + "angle": 0, + "content": "- We conduct extensive experiments on different architectures and tasks, and provide insights on using routing functions in VL PEFT tasks." + }, + { + "type": "list", + "bbox": [ + 0.224, + 0.147, + 0.784, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.232, + 0.396, + 0.249 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.266, + 0.785, + 0.431 + ], + "angle": 0, + "content": "Adapter-based PEFT To reduce the tuning costs of pre-trained language models (LMs), the Adapter layer was first introduced in [9]. It first down-projects the model's hidden state, then up-projects it back to its original dimension, incorporating a nonlinear function after the down-projection. Freezing the pretrained weights of the LM and tuning them with sequentially added Adapter layers achieves performance comparable to full fine-tuning. Since then, various Adapter variants have been proposed to advance the field of PEFT. The focus of developing Adapter variants has been on improving efficiency [18,28,34] and exploring alternative architectures [8,14,24]. Despite being surpassed by its variants on various benchmarks, the original sequential Adapter is still widely used in practice and offers stable results across tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.785, + 0.551 + ], + "angle": 0, + "content": "LoRA-based PEFT LoRA [10] adopts the same \"down-projection then up-projection\" approach as the Adapter, with no nonlinear function in the low-rank bottleneck. LoRA is added to the query and value mappings in the attention modules of the pre-trained LM. Important follow-up works include dynamic rank adjustment [38, 41] and quantization adaptation [5, 40]. Moreover, VeRA [17] and NOLA [16] further minimize the number of parameters by leveraging frozen random matrices. LoRA-based PEFT methods have become standard for tuning large language models and large multimodal foundation models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.785, + 0.641 + ], + "angle": 0, + "content": "Other PEFT methods Beyond Adapter and LoRA, prompt-based tuning methods [13, 19, 22] have shown strong performance for tuning both language and vision models. LST [36] learns a compact side network during tuning. BitFit [2] only adjusts the bias terms during fine-tuning, and achieves comparable performance as other PEFT methods. Methods [8, 29, 39, 42] also explore combining PEFT methods to achieve superior performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Vision-Language PEFT VL-Adapter [37] is the pioneering work in the field of Vision-Language (VL) PEFT. It empirically studies the use of different PEFT methods on four different VL tasks in a multitask learning setting. VL-Adapter uses CLIP-BART, a similar model architecture as VL-T5 [33]. CLIP-BART uses a projection layer to a connect CLIP [31] vision encoder to BART [20] or T5 [33]. CLIP-BART is tuned using a combination of prompt tuning and other PEFT methods. We conduct extensive experiments using CLIP-BART and VL-Adapter as presented in Sec. 5. As an extension to VL-Adapter, VL-PET [11] combines an Adapter with a rather complicated granularity-controlled mechanism. I-Tuning [27] explores alternative model architectures for VL PEFT tasks, and proposes a parallel Adapter with cross-attention to integrate visual information into a pre-trained LM. Since our work is the first that introduces routing functions into VL PEFT, we limit the comparison of our method to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "T. Qu et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.382, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.145, + 0.599, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.145, + 0.788, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.333, + 0.788, + 0.457 + ], + "angle": 0, + "content": "Fig. 2: Illustration of our method. Left: Conventional PEFT methods with low-rank bottleneck first map the hidden states \\( x_{H} \\) from a high dimension \\( d \\) to a lower dimension \\( r \\) via down-projection map \\( W_{down} \\). Then, it maps the hidden states back to the original dimension \\( d \\) via \\( W_{up} \\). Middle: Our method utilizes the same architecture, but updates the features via a routing function in the low-rank bottleneck. Specifically in a VL task, for the features \\( x_{R} \\) that we want to align to, we use \\( W_{down} \\) to down-project them as \\( W_{down}x_{R} \\). Then the routing function routes \\( W_{down}x_{R} \\) and \\( W_{down}x_{H} \\) in the low-rank bottleneck. Right: Different routing functions. In Adapter, routing functions are added before nonlinear activation functions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.788, + 0.518 + ], + "angle": 0, + "content": "VL-Adapter and I-Tuning, leaving the exploration of more complicated PEFT methods like VL-PET to future work." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.541, + 0.381, + 0.559 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.788, + 0.692 + ], + "angle": 0, + "content": "As illustrated in Fig. 2 (Left), standard PEFT modules with low-rank bottleneck, i.e. LoRA/Adapter, learn a down-projection mapping \\( W_{down} \\) followed by an up-projection mapping \\( W_{up} \\), with no routing in between. These modules are added in each Transformer block of the pre-trained foundation model (Fig. 1). As presented in Fig. 2 (Middle), our method adds a routing function in the low-rank bottleneck. This routing function is designed to align two sets of features, in particular, the extra visual features \\( x_{R} \\) to be integrated in the VL tasks and the hidden states \\( x_{H} \\) of each Transformer block of the foundation model." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.693, + 0.785, + 0.709 + ], + "angle": 0, + "content": "The design of a routing function follows two 'parameter-efficient' principles:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.717, + 0.667, + 0.731 + ], + "angle": 0, + "content": "- A linear operation is used to reduce computation overload." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.732, + 0.576, + 0.746 + ], + "angle": 0, + "content": "- No extra trainable parameters are introduced." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.717, + 0.667, + 0.746 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.756, + 0.788, + 0.817 + ], + "angle": 0, + "content": "In this paper, we explore four types of routing functions. Given \\( x_{H} \\) and \\( x_{R}^{4} \\), after the \\( W_{down} \\) mapping, we obtain \\( x_{t} = W_{down}x_{H} \\) and \\( x_{v} = W_{down}x_{R} \\). Here \\( x_{t} \\) is of shape \\( (L_{t},r) \\) and \\( x_{v} \\) is of shape \\( (L_{v},r) \\), where \\( L_{t} / L_{v} \\) and \\( r \\) are the feature length (i.e. the number of tokens) and the bottleneck dimension or rank," + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Assuming \\( x_{R} \\) to have the same dimension \\( d \\) as \\( x_{H} \\). See Sec. 4 and Sec. 5 for details." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.507, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Routing Functions For VL PEFT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "respectively.\\(^5\\) In practice, we always use \\( L_v = 1 \\), i.e. we use a single token / global feature vector to represent the visual content. In case of multiple tokens, we use average pooling to reduce it to a single feature vector.\\(^6\\) The proposed routing functions operate as:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.217, + 0.785, + 0.234 + ], + "angle": 0, + "content": "- Element-wise multiplication: \\( x_{t} \\circ x_{v}' \\), where \\( x_{v}' \\) denotes \\( x_{v} \\). reshape_as(\\( L_{t}, r \\))7" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.233, + 0.758, + 0.249 + ], + "angle": 0, + "content": "- Element-wise addition: \\( x_{t} + x_{v}^{\\prime} \\), where \\( x_{v}^{\\prime} \\) denotes \\( x_{v} \\).reshape_as(L_t, r)" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.248, + 0.497, + 0.264 + ], + "angle": 0, + "content": "- Matrix multiplication 1: \\( x_{t}(x_{v})^{\\top}x_{v} \\)" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.264, + 0.737, + 0.281 + ], + "angle": 0, + "content": "- Matrix multiplication 2: \\( x_{t}x_{v}^{\\prime \\prime} \\), where \\( x_{v}^{\\prime \\prime} \\) denotes \\( x_{v} \\).reshape_as(r,r)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.217, + 0.785, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.289, + 0.785, + 0.456 + ], + "angle": 0, + "content": "Intuitively, \\( x_{t} \\circ x_{v}^{\\prime} \\), reinforces or attenuates individual elements of \\( x_{t} \\), using the corresponding element of \\( x_{v} \\) as scaling factor. The second routing function, \\( x_{t} + x_{v}^{\\prime} \\) replaces \\( x_{t} \\) with the sum of \\( x_{v}^{\\prime} \\) and \\( x_{t} \\), thus moving \\( x_{t} \\) closer to \\( x_{v} \\). \\( x_{t}(x_{v})^{\\top}x_{v} \\) projects \\( x_{t} \\) onto \\( x_{v} \\), i.e. only keeping the component of \\( x_{t} \\) that is aligned with \\( x_{v} \\). Finally, \\( x_{t}x_{v}^{\\prime\\prime} \\) replaces \\( x_{t} \\) with a rescaled version of \\( x_{v} \\), with a different scale factor for each row given by the sum over all elements of the corresponding row of \\( x_{t} \\). In all these cases, it should be kept in mind that both \\( x_{v} \\) and \\( x_{t} \\) are derived from \\( x_{H} \\) and \\( x_{R} \\) via the down-projection \\( W_{down} \\). Parameters of \\( W_{down} \\) can be learned so as to select particular dimensions of the original \\( x_{H} \\) and \\( x_{R} \\), on which to apply the above operations before up-projecting them again with \\( W_{up} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.785, + 0.53 + ], + "angle": 0, + "content": "The features \\( x_{R} \\) contain information from the modalities to which we aim to align \\( x_{H} \\). The choice of \\( x_{R} \\) can be different for different architectures, for instance, [CLS] features from a ViT in case of an encoder-only or decoder-only LM. We further explain the motivation of the choices of \\( x_{R} \\) for different types of architectures in Sec. 4.1 and Sec. 5.1." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.555, + 0.784, + 0.573 + ], + "angle": 0, + "content": "4 Experiments with En(De)coder-Only Language Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.588, + 0.784, + 0.678 + ], + "angle": 0, + "content": "To explore the value of routing functions, in this section, we start from encoder-only and decoder-only language models (LMs) and incorporate visual information from pre-trained vision encoders for vision-language (VL) parameter-efficient fine-tuning (PEFT) tasks. The section begins with an overview of our experimental design and implementation details, and concludes with a discussion on the comprehensive experimental results obtained." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.701, + 0.436, + 0.717 + ], + "angle": 0, + "content": "4.1 Experimental Design" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.727, + 0.784, + 0.788 + ], + "angle": 0, + "content": "Here, we study VL PEFT tasks in their cleanest form. Specifically, (1) No additional mapping networks are adopted, so the feature routing functions form the only module that contributes to the final performance; (2) Only the parameters of the PFET modules in LMs are updated, where we learn the VL alignment; (3)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.797, + 0.558, + 0.812 + ], + "angle": 0, + "content": "Here we omit the batch dimension for simplicity." + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.812, + 0.783, + 0.825 + ], + "angle": 0, + "content": "6 Only when comparing to cross-attention in Sec. 4.3, we do not use pooled features." + }, + { + "type": "page_footnote", + "bbox": [ + 0.223, + 0.825, + 0.785, + 0.841 + ], + "angle": 0, + "content": "e.g. \\( x_{v} \\). reshape_as(\\( L_{t}, r \\)): \\( x_{v}' = \\mathbb{1}^{\\mathsf{T}}x_{v} \\), where \\( \\mathbb{1} \\) is a matrix of ones of shape \\( (L_{v}, L_{t}) \\)." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.797, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "T. Qu et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.143, + 0.402, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.143, + 0.625, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.143, + 0.787, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.305, + 0.788, + 0.349 + ], + "angle": 0, + "content": "Fig.3: Pipeline for VQA and Image Captioning. Only the added classifier (in VQA) and the PEFT modules of the RoBERTa encoder and GPT decoder are tuned. Note that here \\( x_{H} \\) are the text inputs with the visual [CLS] features prepended." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.784, + 0.437 + ], + "angle": 0, + "content": "No multimodal tasks (e.g. contrastive language-image pre-training in CLIP [31]) are included in the pre-training phases of the backbone models. We only learn the multimodal embedding space during VL PEFT. These principles guarantee that VL alignment originates strictly from the PEFT modules." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.785, + 0.514 + ], + "angle": 0, + "content": "To comprehensively study the role of routing functions in VL tasks, we consider both discriminative and generative VL tasks. We select visual question answering (VQA) and Image Captioning, two of the most widely studied VL tasks, as the representative task for discriminative VL tasks and generative VL tasks, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.789, + 0.664 + ], + "angle": 0, + "content": "We present the pipelines for VQA and Image Captioning in Fig. 3. We use ViT [6] as our vision backbone. As for the LM backbone, we use the encoder-only RoBERTa [25] and the decoder-only GPT2 [32] for VQA and Image Captioning, respectively. The [CLS] feature from the ViT is prepended to the text inputs and acts as \\( x_{R} \\), enabling the learning of VL alignment through PEFT, with or without the integration of our routing functions. Prepending visual features to text inputs is actually the predominant way of learning VL alignment in PEFT [11, 37] without routing functions. The ViT and LMs adopt the same dimensions for their feature spaces, so we do not need additional mappings to match the dimensions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.687, + 0.457, + 0.701 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.711, + 0.788, + 0.817 + ], + "angle": 0, + "content": "We adopt ViT-B/16 pre-trained on ImageNet-21k as our vision backbone, which we keep frozen during training. In LoRA, we set scaling factor \\(\\alpha = r\\) in all setups. For VQAv2, we use RoBERTa\\(_{\\text{base}}\\) as our LM backbone, with a two-layer multi-layer perceptron (MLP) classifier on top following [37]. We train our model using the AdamW optimizer [26] with learning rate=3e-4 and batch size of 512 for 20 epochs, where we warm up the first \\(5\\%\\) of the steps. For the captioning task, COCO Cap., we use GPT2-base as our LM backbone. We train our model for 20" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.785, + 0.841 + ], + "angle": 0, + "content": "8 In Appendix B, we report results obtained with larger backbones, e.g. RoBERTalarge." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.507, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Routing Functions For VL PEFT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "epochs using the AdamW optimizer with batch size of 40 and learning rate \\(= 3\\mathrm{e} - 5\\) where we warm up the first 5000 steps. For more details, see Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.198, + 0.325, + 0.212 + ], + "angle": 0, + "content": "4.3 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.22, + 0.784, + 0.25 + ], + "angle": 0, + "content": "Overall performance on VQA We first evaluate our routing functions on the discriminative task VQAv2 with RoBERTabase and ViT-B/16, see Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.251, + 0.487, + 0.446 + ], + "angle": 0, + "content": "Without the addition of routing functions, models with LoRA and Adapter reach a similar level of performance. All four types of routing functions significantly improve the performance of the original PEFT method in terms of accuracy of the answer. The dimension of the bottleneck does not have a strong impact on the final performance. Sometimes we can even obtain better performance with lower dimensions. It is a sign that the intrinsic dimension of the VL task is not" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.295, + 0.787, + 0.364 + ], + "angle": 0, + "content": "Table 1: Accuracy and improvement over the baseline (in blue) obtained by different types of routing functions on VQAv2. The best results and second best results are in bold and underlined, respectively." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.365, + 0.783, + 0.436 + ], + "angle": 0, + "content": "
Routingr = 64r = 128
FunctionsLoRAAdapterLoRAAdapter
None44.1544.1644.4544.28
xt○xt'53.51+9.3652.78+8.6253.86+10.4153.01+8.73
xt+xt'52.60+8.4553.94+9.7852.88+8.4353.95+9.67
xt(xt)Txt53.88+9.7354.48+10.3253.09+8.6455.06+10.78
xtxt''54.21+10.0654.96+10.8051.88+7.4354.38+10.10
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.785, + 0.509 + ], + "angle": 0, + "content": "high, just as observed in uni-modal tasks. Moreover, we observe that the routing functions involving matrix multiplications (i.e. \\( x_{t}(x_{v})^{\\top}x_{v} \\& x_{t}x_{v}^{\\prime \\prime} \\)) perform quite well in most scenarios for both LoRA and Adapter, while the routing functions involving element-wise operations (i.e. \\( x_{t} \\circ x_{v}^{\\prime} \\& x_{t} + x_{v}^{\\prime} \\)) are just slightly behind." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.534, + 0.788, + 0.59 + ], + "angle": 0, + "content": "Table 2: Results in terms of accu- Table 3: Results in terms of accuracy with racy with different routing functions on different routing functions on VQAv2 sorted VQAv2 sorted by answer type. The \"all\" by question type. The \"all\" column reports column reports overall accuracy. overall accuracy." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.602, + 0.479, + 0.661 + ], + "angle": 0, + "content": "
Routing FunctionsallLoRA, r = 128Adapter, r = 128
yes/no numberotherallyes/no numberother
None44.4568.6731.2929.2044.2868.3831.19
xt·ox'53.8672.2935.3644.4853.0171.7233.34
xt+xt'52.8872.0334.2043.0053.9572.4335.38
xt(xt)Tx53.0970.6335.4044.1955.0674.0933.59
xtx''51.8869.9233.2942.8254.3873.0033.94
" + }, + { + "type": "table", + "bbox": [ + 0.486, + 0.602, + 0.784, + 0.661 + ], + "angle": 0, + "content": "
Routing \nFunctionsLoRA, r = 128Adapter, r = 128
allwhere colortimepersonallwhere colortimeperson
None44.4522.6440.0014.2945.1744.2822.9340.1513.7044.94
xt × xv'53.8635.7051.6923.0762.7453.0135.5952.6320.7362.28
xt + xv'52.8836.1249.3321.2262.9553.9537.0852.0821.6663.76
xt(x)Tx53.0936.3352.8622.2961.7455.0636.9253.6624.4463.85
xtxv''51.8834.1951.6022.6361.1454.3835.6452.7721.8562.50
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To dive deeper into the behaviors of different types of routing functions, we evaluate the performance on different answer types as presented in Tab. 2. The largest improvements are obtained for the \"other\" category for all four routing functions. It can be seen that different routing functions excel for different answer types. For example, even when the two element-wise operations reach lower overall accuracy, they lead to higher or comparable accuracy on number-type answers. As for the routing functions involving matrix multiplications, better performance on other-type answers is obtained. For instance, with relatively lower overall accuracy reached in LoRA, matrix multiplications reach comparable performance on other-type answers. No clear pattern can be observed for yes/no" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "T. Qu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 4: Results of different types of routing functions on COCO Cap. The Avg. score is the averaged result from BLEU-4, METEOR, ROUGE-L, CIDEr, and SPICE. The best and second best results are in bold and underlined, respectively." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.198, + 0.785, + 0.325 + ], + "angle": 0, + "content": "
Routing FunctionsLoRA, r=64LoRA, r=128
BLEU-4METEORROUGE-LCIDErSPICEAvg.BLEU-4METEORROUGE-LCIDErSPICEAvg.
None18.317.436.855.811.828.0221.020.943.470.314.734.06
xt×ov'26.1+7.823.7+6.348.0+11.888.7+32.917.3+5.540.88+12.8625.4+4.423.7+2.848.2+4.887.9+17.617.7+3.040.58+6.52
xt+xov'22.2+3.920.3+2.941.0+4.273.5+17.714.8+3.034.36+6.3425.4+4.423.1+2.247.1+4.787.1+16.817.0+2.339.94+5.88
xt(xv)Txv24.8+6.522.6+5.245.4+8.684.9+29.116.8+5.038.90+10.8824.9+3.922.7+1.845.4+2.083.8+13.516.9+2.238.74+4.68
xtx''23.9+5.621.9+4.543.9+7.180.5+14.716.2+4.437.28+9.2625.5+4.522.9+2.046.0+2.685.2+14.917.2+2.539.36+5.30
Routing FunctionsAdapter, r=64Adapter, r=128
BLEU-4METEORROUGE-LCIDErSPICEAvg.BLEU-4METEORROUGE-LCIDErSPICEAvg.
None15.918.537.061.614.129.4216.118.436.261.513.929.22
xt×ov'24.6+8.723.1+4.646.4+9.484.5+22.917.2+3.139.16+9.7423.6+7.522.6+4.244.8+8.682.8+21.317.1+3.238.18+8.96
xt+xov'21.0+5.121.5+3.042.4+5.475.0+14.416.0+1.935.18+5.7619.3+3.220.8+2.440.9+4.775.2+13.816.2+2.334.48+5.26
xt(xv)Txv26.1+10.223.2+4.746.9+9.985.4+23.817.3+3.239.78+10.3624.1+8.022.4+4.044.2+8.082.7+21.216.9+3.038.06+8.84
xtx''26.6+10.723.0+4.546.8+9.885.8+24.217.2+3.139.88+10.4626.0+9.923.5+5.146.8+10.686.3+24.817.8+3.940.08+10.86
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.787, + 0.385 + ], + "angle": 0, + "content": "types of answers, which could be explained by the large variety in content of these questions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.789, + 0.614 + ], + "angle": 0, + "content": "Tab. 3 shows our method's performance across various question types. Four distinct question types test the method's reasoning capabilities concerning fine-grained visual information: spatial information (\"where\", e.g. Where is the surfboard?), color information (\"color\", e.g. What color is the bike?), time information (\"time\", e.g. What time is on the clock?), and person information (\"person\", e.g. What is the man holding?).9 Adding routing functions significantly improves the performance of LoRA and Adapter. Notably, different routing functions seem to capture different aspects of the visual information. \\( x_{t} + x_{v}^{\\prime} \\) shows great performance while reasoning on spatial information (where) and person information (person). It is assumed that the addition operator reinforces feature values of content both present in the question and image, which might be beneficial for spatial and person questions. Similarly, \\( x_{t}x_{v}^{\\prime\\prime} \\) performs exceptionally well for time information (time). With \\( x_{t} \\circ x_{v}^{\\prime} \\), we obtain relatively better performance on the color information (color). \\( x_{t}(x_{v})^{\\top}x_{v} \\) presents consistent strong abilities while aligning spatial information (where) and color information (color)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.789, + 0.675 + ], + "angle": 0, + "content": "In Appendix B, we show an additional ablation study on replacing visual features \\((x_{R})\\) with random noise or all-ones features. We also compare inference time of LoRA/Adapter with LoRA/Adapter + routing functions, where we observe adding routing functions to PEFT methods does not reduce efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.677, + 0.789, + 0.813 + ], + "angle": 0, + "content": "Overall performance on image captioning Evaluation of image captioning relies on the overall assessment of the generated captions. We report multiple text generation metrics and their average scores in Tab. 4. The addition of routing functions again leads to significant performance gains for all metrics in all scenarios. As was the case with VQAv2, different patterns are observed. Changing the intermediate dimension \\( r \\) from 128 to 64 seems to have a larger impact on the model performance with conventional LoRA, where without the routing functions the CIDEr score drops from 70.3 to 55.8. In contrast, models augmented with routing functions remain robust and per" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.687, + 0.84 + ], + "angle": 0, + "content": "9 The comprehensive list of question types is included in Appendix A." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.507, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Routing Functions For VL PEFT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "formant across setups, despite the changes in intermediate dimension. \\( x_{t} \\circ x_{v}^{\\prime} \\) and \\( x_{t}x_{v}^{\\prime \\prime} \\) perform well, possibly due to their alignment power. We also report the performance of Adapter with larger backbones, namely GPT2-medium and ViT-L/16, in Appendix B. Consistent and significant improvements in performance are observed, showing the efficacy of our routing functions on the task." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.239, + 0.486, + 0.253 + ], + "angle": 0, + "content": "Comparison to cross-attention" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.254, + 0.487, + 0.388 + ], + "angle": 0, + "content": "Here we compare the results of the integration of routing function \\( x_{t}(x_{v})^{\\top}x_{v} \\) with these of integrating cross-attention between the textual and the visual representation in image captioning. Indeed, in principle, a cross-attention function can also be considered as a kind of routing function with nonlinear operations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.501, + 0.436 + ], + "angle": 0, + "content": "I-Tuning [27] implements an Adapter with cross-attention for image captioning. To have a fair comparison, we" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.787, + 0.51 + ], + "angle": 0, + "content": "follow the settings of I-Tuning for both LoRA and Adapter. Specifically, (1) instead of using the visual features, we pretend the \"\" token to the input text. (2) We use the last hidden state from ViT as key and value for cross-attention and as \\( x_{R} \\) for routing functions. This change is needed, since the [CLS] feature is of length 1, which does not work for cross-attention due to the Softmax function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.788, + 0.633 + ], + "angle": 0, + "content": "In the case of LoRA, we also implement a variant with separate \\( W_{down} \\) mappings for \\( x_{R} \\) and \\( x_{H} \\). This brings LoRA with routing function closer to LoRA with cross-attention. Without separate mapping refers to the setting used in previous experiments, where \\( x_{R} \\) and \\( x_{H} \\) share the same \\( W_{down} \\). As shown in Tab. 5, despite of having less parameters, \\( \\mathrm{LoRA} + x_t(x_v)^{\\top}x_v \\) (w/ separate mapping) achieves significantly better performance than LoRA with cross-attention. The setup of \\( \\mathrm{LoRA} + x_t(x_v)^{\\top}x_v \\) (w/o separate mapping) outperforms LoRA with cross-attention by a large margin using a similar amount of parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.634, + 0.788, + 0.709 + ], + "angle": 0, + "content": "For Adapter, the cross-attention module has the same intermediate dimension as the \\( W_{down} \\) in Adapter with routing function. Naturally, the separate query/ key/value mappings make that the Adapter with cross-attention has more than double the number of parameters compared to the Adapter equipped with the routing function \\( (x_{t}(x_{v})^{\\top}x_{v}) \\). Both achieve comparable performance (Tab. 5)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.71, + 0.788, + 0.786 + ], + "angle": 0, + "content": "To conclude, Tab. 5 shows that integrating \\( x_{t}(x_{v})^{\\top}x_{v} \\) in LORA and Adapter yields better performance given the same amount of parameters, or is similar in performance while using substantially less parameters. For VL tasks like image captioning, integrating a routing function to guide feature representations in low-rank bottlenecks has superior abilities compared to cross-attention." + }, + { + "type": "page_footnote", + "bbox": [ + 0.215, + 0.797, + 0.788, + 0.842 + ], + "angle": 0, + "content": "10 Appendix A shows detailed architectures and explains why routing functions outperform cross-attention. We down-project to \\( r/4 \\) and \\( r \\) for LoRA w/ & w/o separate mapping, respectively, to have the same level of parameter count." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.267, + 0.788, + 0.377 + ], + "angle": 0, + "content": "Table 5: Comparing routing functions with cross-attention on COCO Cap. †: parallel Adapter as used in [27]. Separate Map.: separate mapping for query/ key/value (in cross-attn.) or features \\( x_{R} / x_{H} \\) (in \\( x_{t}(x_{v})^{\\top}x_{v} \\)). Here \\( r = 128 \\). Avg. indicates the same average scores as in Tab. 4. Full evaluation in Appendix B." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.378, + 0.785, + 0.43 + ], + "angle": 0, + "content": "
PEFTSeparateMap.AlignmentParam.BLEU-4CIDErAvg.
LoRA✓Cross-attn.4.786M28.792.243.02
LoRA✓xt(xt)Txtxt3.932M30.7+2.099.4+7.245.68+2.66
LoRAXxt(xt)Txtxt4.746M30.0+1.399.0+6.845.38+2.36
Adapter†✓Cross-attn.4.732M30.799.845.70
Adapter†Xxt(xt)Txtxt1.830M30.8+0.198.8-1.045.48-0.22
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "T. Qu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.488, + 0.297 + ], + "angle": 0, + "content": "Incorporating nonlinear activation functions into routing functions We have proposed linear routing functions for efficiency reasons, but they do not need to be exclusively linear. Taking inspirations from the linear attention variants [4,15,30], where the linear operation is often equipped with nonlinear activation functions to ensure non-negativity for" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.163, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 6: Experiments on adding nonlinear activation functions in routing functions. We use \\( r = 64 \\) here." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.204, + 0.785, + 0.264 + ], + "angle": 0, + "content": "
PEFT MethodRouting FunctionsVQA AccuracyCOCO Cap.
BLEU-4CIDErAvg.
LoRAxt(xv)Txv53.8824.884.938.90
LoRAReLU(xt)ReLU(xv)Txv53.4525.386.139.84
Adapterxt(xv)Txv54.4826.185.439.78
AdapterReLU(xt)ReLU(xv)Txv54.6325.084.739.02
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.784, + 0.389 + ], + "angle": 0, + "content": "similarity scores between query and key, we explore whether a similar operation affects the representation power of the routing functions in the low-rank bottleneck. For routing function \\( x_{t}(x_{v})^{\\top}x_{v} \\), we use the nonlinear activation function \\( ReLU(\\cdot) \\) as in previous work [30], and transform it to \\( ReLU(x_{t})ReLU(x_{v})^{\\top}x_{v} \\). The results are presented in Tab. 6. With or without \\( ReLU(\\cdot) \\), our routing function achieves similar performance on VQA and image captioning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.788, + 0.515 + ], + "angle": 0, + "content": "Qualitative Analysis We first qualitatively compare the VL alignment pattern of LoRA-tuned GPT2 with or without a routing function on image captioning. An example is presented in Fig. 4. With the addition of routing function \\( x_{t}(x_{v})^{\\top}x_{v} \\), the model pays more attention to the nouns (e.g. man, a big screen) and the verbs (e.g. watching), which are essential in learning the VL alignment. At the same time, it pays less attention to meaningless words like \"something\". This is in line with the finding that the routing function stresses components of \\( x_{t} \\) that are aligned in the image." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.547, + 0.72, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.652, + 0.788, + 0.682 + ], + "angle": 0, + "content": "Fig. 4: Example of average attention weights from the last layer of GPT2. We use the final checkpoint trained on COCO Cap. with \\( r = 64 \\). IMAGE: visual [CLS] feature." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Next we qualitatively examine some examples from the inference phase for VQA and Image Captioning. We present examples in Fig. 5 where conventional LoRA fails. For VQA, previous experiments in Tab. 3 indicate that our routing functions work particularly well on color related questions. As also shown in Fig. 5, conventional LoRA fails to align the color red to the lollipop. The model tuned with conventional LoRA also fails to correctly answer other types of questions as presented in the figure. Without routing functions, captions generated from the LM tuned with conventional LoRA often make mistakes in compre" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.507, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Routing Functions For VL PEFT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.147, + 0.466, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.479, + 0.147, + 0.756, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Fig. 5: Qualitative examples from VQA and Image Captioning tasks. We present cases where the model trained with conventional LoRA fails and results obtained from a model trained with LoRA with a routing function. We use \\( r = 64 \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.785, + 0.444 + ], + "angle": 0, + "content": "hending the visual information. Occasionally, we observe a sequence of blank spaces as presented in the first example using conventional LoRA." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.463, + 0.774, + 0.481 + ], + "angle": 0, + "content": "5 Experiments with Encoder-Decoder Language Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.613 + ], + "angle": 0, + "content": "In this section, we study the VL PEFT tasks in a more complicated setting, where we consider an encoder-decoder LM with an additional mapping network dedicated for the vision encoder. The model is trained in a multi-task learning setting, where four different VL tasks are jointly learned, as in [3,37]. Despite the increased complexity of optimizing the network in this scenario, we show that integrating routing functions into the low-rank bottlenecks yields performance enhancements. Moreover, consistent patterns are observed for different scenarios, which shed lights on future research to improve VL PEFT tasks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.633, + 0.436, + 0.649 + ], + "angle": 0, + "content": "5.1 Experimental Design" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Following [37], we use the CLIP-BART architecture to study the effect of our routing functions on the VL PEFT tasks in a multitask learning setting. CLIP-BART is a widely used architecture for VL tasks. We present the simplified model architecture of CLIP-BART with Adapters and the integrated routing functions in Fig. 6. The original CLIP-BART uses a trainable visual projection module to map the visual features extracted by CLIP to the embedding space of BART [20]. CLIP-BART uses task-specific prompts as the textual input: For example, in Fig. 6, \"\" is added as prompt to the text for VQA. The mapped visual features are then prepped to the textual input. Together, they serve as the input to BART. CLIP-BART is trained and evaluated in a multitask setting, that is, VQAv2 [7] and GQA [12] for visual question answering, NLVR\\(^2\\) [35] using natural language for visual reasoning, and COCO Cap. [23] for image captioning." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "T. Qu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.488, + 0.252 + ], + "angle": 0, + "content": "CLIP-BART unifies the four tasks into one text generation problem following [3], where four different task prompts distinguish the task types. CLIP-BART can be considered as the representative of the current state-of-the-art on VL PEFT tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.487, + 0.509 + ], + "angle": 0, + "content": "To integrate our routing functions into the model, two types of inputs \\( x_{R} \\) can be considered, namely the visual features after the visual projection, and the last encoder hidden state as presented in Fig. 6. The visual features only contain visual information. Because the last encoder hidden state contains both textual and visual information, it can only be added to the decoder. We experiment with both LoRA and Adapter as the PEFT method for fine-tuning CLIP-BART. In this multitask learning setting, one can use (1) a single LoRA/Adapter shared across tasks or (2) multiple LoRAs/Adapters without parameter" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.787, + 0.571 + ], + "angle": 0, + "content": "sharing, where each task has its own LoRA/Adapter. We start by exploring the impact of different routing functions on the single LoRA/Adapter setting. Then, based on the observations obtained in the setting with single Adapter, we explore ways of using the routing functions in the multiple Adapters setting." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.161, + 0.789, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.388, + 0.788, + 0.487 + ], + "angle": 0, + "content": "Fig. 6: Architecture of CLIP-BART with Adapter. The standard Adapter architecture is adopted (see Fig. 1). Only the modules in green are trainable. To apply routing functions, we use the pooled last encoder hidden state as feature \\( x_{R} \\), which is added to the Adapters in the decoder." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.591, + 0.457, + 0.607 + ], + "angle": 0, + "content": "5.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.738 + ], + "angle": 0, + "content": "Following [11, 37], we use CLIP-ResNet101 as our vision encoder [31]. We use \\(\\mathrm{BART}_{\\mathrm{base}}\\) as our encoder-decoder LM as in [37]. Following [37], we use \\(r = 96\\) for all experiments. The training pipeline in [37] is used for a fair comparison. Specifically, (1) we use the same seed for all experiments. (2) Due to limited resources, we do not perform an extensive hyper-parameter search as in [37]. (3) We train the model with AdamW optimizer and linear decay for 20 epochs. We warm-up the learning rate from 0 to the highest in the first 2 epochs. (4) The batch size is set to 500. For more details, we refer to Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.756, + 0.324, + 0.77 + ], + "angle": 0, + "content": "5.3 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Experiments with a single Adapter/LoRA We first report results of experiments on using single Adapter/LoRA shared across tasks in Tab. 7. In this case, the Adapter/LoRA is trained with data from four different tasks, and naturally we can only use one type of routing function throughout the experiments." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.507, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Routing Functions For VL PEFT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 7: Experiments with different types of routing functions using CLIP-BART with single LoRA/Adapter. Following [11] \\((\\dagger)^{11}\\), we report the average CIDER score for COCO Cap., accuracy for the other three tasks. We also report the average accuracy from VQA, GQA and NLVR2. We bold and underline the best and second best results." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.213, + 0.784, + 0.309 + ], + "angle": 0, + "content": "
Routing FunctionsSingle LoRASingle Adapter
VQAGQANLVR2Avg.COCOCap.VQAGQANLVR2Avg.
None†65.1553.6672.5863.80115.0165.7654.1673.1964.37
xt○x′v65.6853.9673.4264.35113.9465.9254.3474.2364.83
xt+x′v65.1453.7373.5164.13114.9665.8954.1873.9064.66
xt(xv)Txv64.9453.5673.6064.03117.8065.8453.6574.3164.03
xtx''v64.8453.1372.9863.65119.2665.8353.6173.2764.24
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.788, + 0.583 + ], + "angle": 0, + "content": "Tab. 7 shows that the integration of the routing functions in the single LoRa and single Adapter architecture results in improved performance in nearly all scenarios. But just like the experiments with encoder-only and decoder-only LMs in Sec. 4.3, different routing functions show different patterns. The routing functions involving matrix multiplication, that is, \\( x_{t}(x_{v})^{\\top}x_{v} \\) and \\( x_{t}x_{v}^{\\prime \\prime} \\), perform exceptionally well in the COCO captioning task, while the other two routing functions involving element-wise operations, i.e. \\( x_{t}\\circ x_{v}^{\\prime} \\) and \\( x_{t} + x_{v}^{\\prime} \\), show more potential on QA tasks (VQA/GQA). As for the NLVR\\(^{2}\\) task, \\( x_{t}(x_{v})^{\\top}x_{v} \\) and \\( x_{t}\\circ x_{v}^{\\prime} \\) present stronger abilities to learn VL relationships. Note that achieving relatively low performance on one task does not necessarily mean that the corresponding routing function fails. For example, \\( x_{t}\\circ x_{v}^{\\prime} \\) does not work well for captioning, but achieves great performance on the other three tasks. This somehow contrasts with the results obtained with encoder-only and decoder-only LMs in Sec. 4.3, where \\( x_{t}\\circ x_{v}^{\\prime} \\) leads to one of the best performance on COCO Cap. The complexity of the multitask learning setup, where different loss functions are jointly optimized, might here be a confounding factor." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Experiments with multiple Adapters We further explore the behavior of the routing functions in models with multiple Adapters, where we train a separate Adapter per task. As also observed in [37], using multiple Adapters increases the complexity of the model, and yields worse performance compared to using a single Adapter. As presented in Tab. 8, using the four routing functions with multiple Adapters leads to comparable or better performance as compared to multiple Adapters with no routing functions. Moreover, in line with what we discovered in the experiments with models with a single Adapter, different routing functions shine on different types of tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.721, + 0.788, + 0.798 + ], + "angle": 0, + "content": "In this set-up, we also experiment with combinations of routing functions. The four combinations integrate the routing functions that work best for each individual task (e.g. \\( x_{t}x_{v}^{\\prime} \\) for COCO Cap.). The results in Tab. 8 evidence that in this multitask setting combinations of routing functions yield an improved performance compared to using no routing functions. GQA requires more fine" + }, + { + "type": "page_footnote", + "bbox": [ + 0.215, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "11 \\(\\dagger\\): 3 seeds. Appendix C shows that our improvements are significant using paired t-test with \\(p = 0.05\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.129 + ], + "angle": 0, + "content": "T. Qu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 8: Experiments with different combinations of routing functions using CLIP-BART with multiple Adapters. Following [37] \\((\\ddagger)\\), we report CIDEr score for COCO Cap., accuracy for the other three tasks. We also report the average accuracy from VQA, GQA and NLVR2. We bold and underline the best and second best results." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.213, + 0.784, + 0.361 + ], + "angle": 0, + "content": "
Routing FunctionsRouting FunctionsPerformance
VQAGQANLVR2COCO Cap.VQAGQANLVR2Avg.COCO Cap.
None†XXXX65.454.069.863.07114.3
xt○x′v✓✓✓✓65.6153.5170.7863.30113.72
xt+x′v✓✓✓✓65.1953.8470.9963.34115.64
xt(xv)Txv✓✓✓✓64.9352.1170.2562.43116.20
xtx′v✓✓✓✓65.3252.6770.5862.86117.92
comb1xt○x′vxt○x′vxt(xv)Txvxtx′v66.3754.2169.2763.28116.94
comb2xt+x′vxt+x′vxt(xv)Txvxtx′v65.1453.3770.8863.13117.58
comb3xt○x′vxt○x′vxt○x′vxtx′v66.3453.7770.0263.38117.73
comb4xt+x′vxt+x′vxt○x′vxtx′v65.4752.9971.0163.16118.34
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.785, + 0.451 + ], + "angle": 0, + "content": "grained image representation. For efficiency reason, we only adopt \\( x_{v} \\) as the down-projection from the pooled last encoder hidden state. Nevertheless, comb1 can recover from this shortcoming. More detailed image representations as used in the cross-attention experiments of Tab. 5 might still enhance the results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.452, + 0.787, + 0.557 + ], + "angle": 0, + "content": "Overall, our research has focused on introducing routing functions in low-rank feature approximations when fine-tuning models for downstream (in our case VL) tasks based on pre-trained foundation models. As a pioneering work on the topic, for obtaining results we did not exhaustively search for all possible choices of hyperparameters. Instead, we have focused on revealing the problem of low-rank bottlenecks and on providing a successful solution. We leave an exhaustive grid search for optimal hyperparameters as future work." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.581, + 0.36, + 0.597 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.785, + 0.704 + ], + "angle": 0, + "content": "We have introduced feature routing functions to help guide the feature learning in low-rank bottlenecks in architectures that are typical for parameter-efficient fine-tuning such as LoRA and Adapter. We have shown their validity in vision-language tasks, where we have focused on linear functions that do not require any extra parameters. We have demonstrated performance gains in single-task and multi-task settings in combination with a variety of foundation models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Routing functions open many avenues for future research. Given the importance of foundation models that are trained in a self-supervised way on large data, there is the need to efficiently fine-tune them for downstream tasks. When computing low-rank approximations of the features, our work shows that the extra guidance of the routing functions is important. We see many possibilities in experimenting with other types of functions in other tasks. For instance, linear and other functions are also suited to model constraints that reflect external knowledge or expectations on the behavior of a model. Future work will show whether they are useful to guide the feature routing in low-rank bottlenecks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.507, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Routing Functions For VL PEFT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.221 + ], + "angle": 0, + "content": "This work is funded by the Flanders AI Research Program, European Research Council Advanced Grant H2020-ERC-2017-ADG 788506, and the China Scholarship Council." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.244, + 0.323, + 0.26 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.224, + 0.274, + 0.787, + 0.371 + ], + "angle": 0, + "content": "1. Aghajanyan, A., Gupta, S., Zettlemoyer, L.: Intrinsic dimensionality explains the effectiveness of language model fine-tuning. In: Zong, C., Xia, F., Li, W., Navigli, R. (eds.) Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 7319-7328. Association for Computational Linguistics, Online (Aug 2021). https://doi.org/10.18653/v1/2021.acl-long.568, https://aclanthology.org/2021.acl-long.568" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.372, + 0.787, + 0.455 + ], + "angle": 0, + "content": "2. Ben Zaken, E., Goldberg, Y., Ravfogel, S.: BitFit: Simple parameter-efficient fin-tuning for transformer-based masked language-models. In: Muresan, S., Nakov, P., Villavicencio, A. (eds.) Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). pp. 1-9. Association for Computational Linguistics, Dublin, Ireland (May 2022). https://doi.org/10.18653/v1/2022.acl-short.1, https://aclanthology.org/2022.acl-short.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.456, + 0.787, + 0.481 + ], + "angle": 0, + "content": "3. Cho, J., Lei, J., Tan, H., Bansal, M.: Unifying vision-and-language tasks via text generation. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.483, + 0.787, + 0.549 + ], + "angle": 0, + "content": "4. Choromanski, K.M., Likhosherstov, V., Dohan, D., Song, X., Gane, A., Sarlos, T., Hawkins, P., Davis, J.Q., Mohiuddin, A., Kaiser, L., Belanger, D.B., Colwell, L.J., Weller, A.: Rethinking attention with performers. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=Ua6zuk0WRH" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.787, + 0.633 + ], + "angle": 0, + "content": "5. Dettmers, T., Pagnoni, A., Holtzman, A., Zettlemoyer, L.: Qlora: Efficient finetuning of quantized llms. In: Oh, A., Neumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 10088-10115. Curran Associates, Inc. (2023), https://proceedings.neurips.cc/paper_files/paper/2023/file/1feb87871436031bdc0f2beaa62a049b-Paper-Conference.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.787, + 0.702 + ], + "angle": 0, + "content": "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=YicbFdNTTy" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.787, + 0.757 + ], + "angle": 0, + "content": "7. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "8. He, J., Zhou, C., Ma, X., Berg-Kirkpatrick, T., Neubig, G.: Towards a unified view of parameter-efficient transfer learning. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=ORDcd5Axok" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "9. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for NLP. In: Chaudhuri, K., Salakhutdinov, R. (eds.) Proceedings of the 36th International" + }, + { + "type": "list", + "bbox": [ + 0.224, + 0.274, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.129 + ], + "angle": 0, + "content": "T. Qu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 97, pp. 2790-2799. PMLR (09-15 Jun 2019), https://proceedings.mlr.press/v97/houlsby19a.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.243 + ], + "angle": 0, + "content": "10. Hu, E.J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=nZeVKeeFYf9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "11. Hu, Z.Y., Li, Y., Lyu, M.R., Wang, L.: Vl-pet: Vision-and-language parameter-efficient tuning via granularity control. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3010-3020 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.288, + 0.785, + 0.327 + ], + "angle": 0, + "content": "12. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.329, + 0.785, + 0.355 + ], + "angle": 0, + "content": "13. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.357, + 0.786, + 0.466 + ], + "angle": 0, + "content": "14. Karimi Mahabadi, R., Ruder, S., Dehghani, M., Henderson, J.: Parameter-efficient multi-task fine-tuning for transformers via shared hypernetworks. In: Zong, C., Xia, F., Li, W., Navigli, R. (eds.) Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 565-576. Association for Computational Linguistics, Online (Aug 2021). https://doi.org/10.18653/v1/2021.acl-long.47, https://aclanthology.org/2021.acl-long.47" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.467, + 0.786, + 0.535 + ], + "angle": 0, + "content": "15. Katharopoulos, A., Vyas, A., Pappas, N., Fleuret, F.: Transformers are RNNs: Fast autoregressive transformers with linear attention. In: III, H.D., Singh, A. (eds.) Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 5156-5165. PMLR (13-18 Jul 2020), https://proceedings.mlr.press/v119/katharopoulos20a.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.537, + 0.786, + 0.59 + ], + "angle": 0, + "content": "16. Koohpayegani, S.A., L, N.K., Nooralinejad, P., Kolouri, S., Pirsiavash, H.: NOLA: Networks as linear combination of low rank random basis. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=TjfXcDgvzk" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.592, + 0.786, + 0.632 + ], + "angle": 0, + "content": "17. Kopiczko, D.J., Blankevoort, T., Asano, Y.M.: VeRA: Vector-based random matrix adaptation. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=NjNfLdxr3A" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.634, + 0.786, + 0.701 + ], + "angle": 0, + "content": "18. Lei, T., Bai, J., Brahma, S., Ainslie, J., Lee, K., Zhou, Y., Du, N., Zhao, V.Y., Wu, Y., Li, B., Zhang, Y., Chang, M.W.: Conditional adapters: Parameter-efficient transfer learning with fast inference. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=IyYyKovOAj" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.703, + 0.786, + 0.785 + ], + "angle": 0, + "content": "19. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. In: Moens, M.F., Huang, X., Specia, L., Yih, S.W.t. (eds.) Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 3045-3059. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic (Nov 2021). https://doi.org/10.18653/v1/2021.emnlp-main.243, https://aclanthology.org/2021.emnlp-main.243" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.786, + 0.786, + 0.84 + ], + "angle": 0, + "content": "20. Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., Stoyanov, V., Zettlemoyer, L.: BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In: Jurafsky, D., Chai, J., Schluter, N., Tetreault, J. (eds.) Proceedings of the 58th Annual Meeting" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.507, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Routing Functions For VL PEFT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.148, + 0.786, + 0.189 + ], + "angle": 0, + "content": "of the Association for Computational Linguistics. pp. 7871-7880. Association for Computational Linguistics, Online (Jul 2020). https://doi.org/10.18653/v1/2020.acl-main.703, https://aclanthology.org/2020.acl-main.703" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "21. Li, C., Farkhoor, H., Liu, R., Yosinski, J.: Measuring the intrinsic dimension of objective landscapes. In: International Conference on Learning Representations (2018), https://openreview.net/forum?id=ryup8-WCW" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.786, + 0.327 + ], + "angle": 0, + "content": "22. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. In: Zong, C., Xia, F., Li, W., Navigli, R. (eds.) Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 4582-4597. Association for Computational Linguistics, Online (Aug 2021). https://doi.org/10.18653/v1/2021.acl-long.353, https://aclanthology.org/2021.acl-long.353" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.384 + ], + "angle": 0, + "content": "23. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) Computer Vision - ECCV 2014. pp. 740-755. Springer International Publishing, Cham (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.466 + ], + "angle": 0, + "content": "24. Lin, Z., Madotto, A., Fung, P.: Exploring versatile generative language model via parameter-efficient transfer learning. In: Cohn, T., He, Y., Liu, Y. (eds.) Findings of the Association for Computational Linguistics: EMNLP 2020. pp. 441-459. Association for Computational Linguistics, Online (Nov 2020). https://doi.org/10.18653/v1/2020-findings-emnlp.41, https://aclanthology.org/2020-findings-emnlp.41" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.508 + ], + "angle": 0, + "content": "25. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized BERT pretraining approach. CoRR abs/1907.11692 (2019), http://arxiv.org/abs/1907.11692" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019), https://openreview.net/forum?id=Bkg6RiCqY7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.605 + ], + "angle": 0, + "content": "27. Luo, Z., Hu, Z., Xi, Y., Zhang, R., Ma, J.: I-tuning: Tuning frozen language models with image for lightweight image captioning. In: ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5 (2023). https://doi.org/10.1109/ICASSP49357.2023.10096424" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.66 + ], + "angle": 0, + "content": "28. mahabadi, R.K., Henderson, J., Ruder, S.: Compacter: Efficient low-rank hypercomplex adapter layers. In: Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems (2021), https://openreview.net/forum?id=bqGK5PyI6-N" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.787, + 0.757 + ], + "angle": 0, + "content": "29. Mao, Y., Mathias, L., Hou, R., Almahairi, A., Ma, H., Han, J., Yih, S., Khabsa, M.: UniPELT: A unified framework for parameter-efficient language model tuning. In: Muresan, S., Nakov, P., Villavicencio, A. (eds.) Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 6253-6264. Association for Computational Linguistics, Dublin, Ireland (May 2022). https://doi.org/10.18653/v1/2022.acl-long.433, https://aclanthology.org/2022.acl-long.433" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "30. Qin, Z., Sun, W., Deng, H., Li, D., Wei, Y., Lv, B., Yan, J., Kong, L., Zhong, Y.: cosformer: Rethinking softmax in attention. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=B18CQrx2Up4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.129 + ], + "angle": 0, + "content": "T. Qu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 8748-8763. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/radford21a.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.191, + 0.787, + 0.219 + ], + "angle": 0, + "content": "32. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I.: Language models are unsupervised multitask learners. In: Preprint (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.221, + 0.786, + 0.276 + ], + "angle": 0, + "content": "33. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research 21(140), 1-67 (2020), http://jmlr.org/papers/v21/20-074.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.278, + 0.787, + 0.374 + ], + "angle": 0, + "content": "34. Rücklé, A., Geigle, G., Glockner, M., Beck, T., Pfeiffer, J., Reimers, N., Gurevych, I.: AdapterDrop: On the efficiency of adapters in transformers. In: Moens, M.F., Huang, X., Specia, L., Yih, S.W.t. (eds.) Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 7930-7946. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic (Nov 2021). https://doi.org/10.18653/v1/2021.emnlp-main.626, https://aclanthology.org/2021.emnlp-main.626" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.787, + 0.458 + ], + "angle": 0, + "content": "35. Suhr, A., Zhou, S., Zhang, A., Zhang, I., Bai, H., Artzi, Y.: A corpus for reasoning about natural language grounded in photographs. In: Korhonen, A., Traum, D., Marquez, L. (eds.) Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. pp. 6418-6428. Association for Computational Linguistics, Florence, Italy (Jul 2019). https://doi.org/10.18653/v1/P19-1644, https://aclanthology.org/P19-1644" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.786, + 0.514 + ], + "angle": 0, + "content": "36. Sung, Y.L., Cho, J., Bansal, M.: LST: Ladder side-tuning for parameter and memory efficient transfer learning. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=isPnnaTZaP5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.517, + 0.787, + 0.559 + ], + "angle": 0, + "content": "37. Sung, Y.L., Cho, J., Bansal, M.: Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5227-5237 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.56, + 0.787, + 0.656 + ], + "angle": 0, + "content": "38. Valipour, M., Rezagholizadeh, M., Kobyzev, I., Ghodsi, A.: DyLoRA: Parameter-efficient tuning of pre-trained models using dynamic search-free low-rank adaptation. In: Vlachos, A., Augenstein, I. (eds.) Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics. pp. 3274-3287. Association for Computational Linguistics, Dubrovnik, Croatia (May 2023). https://doi.org/10.18653/v1/2023.eacl-main.239, https://aclanthology.org/2023.eacl-main.239" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.787, + 0.755 + ], + "angle": 0, + "content": "39. Wang, Y., Agarwal, S., Mukherjee, S., Liu, X., Gao, J., Awadallah, A.H., Gao, J.: AdaMix: Mixture-of-adaptations for parameter-efficient model tuning. In: Goldberg, Y., Kozareva, Z., Zhang, Y. (eds.) Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. pp. 5744-5760. Association for Computational Linguistics, Abu Dhabi, United Arab Emirates (Dec 2022). https://doi.org/10.18653/v1/2022.emnlp-main.388, https://aclanthology.org/2022.emnlp-main.388" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.756, + 0.787, + 0.811 + ], + "angle": 0, + "content": "40. Xu, Y., Xie, L., Gu, X., Chen, X., Chang, H., Zhang, H., Chen, Z., ZHANG, X., Tian, Q.: QA-loRA: Quantization-aware low-rank adaptation of large language models. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=WvFoJccpo8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "41. Zhang, Q., Chen, M., Bukharin, A., He, P., Cheng, Y., Chen, W., Zhao, T.: Adaptive budget allocation for parameter-efficient fine-tuning. In: The Eleventh Inter-" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.507, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Routing Functions For VL PEFT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.148, + 0.783, + 0.175 + ], + "angle": 0, + "content": "national Conference on Learning Representations (2023), https://openreview.net/forum?id=1q62uWRJjiY" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "42. Zhou, H., Wan, X., Vulic, I., Korhonen, A.: Autopeft: Automatic configuration search for parameter-efficient fine-tuning. Transactions of the Association for Computational Linguistics 12 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.218 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks/61df44b9-0f88-4bc8-892b-514f717f8da8_origin.pdf b/2024/Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks/61df44b9-0f88-4bc8-892b-514f717f8da8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..39cb3e409e579d075dde26b89469074673bd85a5 --- /dev/null +++ b/2024/Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks/61df44b9-0f88-4bc8-892b-514f717f8da8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ccbaa10bc00280e06570f656e65f29efca01791eb98c16904e7293319cf11b83 +size 1008416 diff --git a/2024/Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks/full.md b/2024/Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..850be7521fe431744c5018924b47585112a83a8a --- /dev/null +++ b/2024/Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks/full.md @@ -0,0 +1,284 @@ +# Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks + +Tingyu $\mathrm{Qu}^{1}$ , Tinne Tuytelaars $^{2}$ , and Marie-Francine Moens $^{1}$ + +$^{1}$ Department of Computer Science, KU Leuven + +$^{2}$ Department of Electrical Engineering, KU Leuven {tingyu.qu; tinne.tuytelaars; sien.moens}@kuleuven.be + +Abstract. Mainstream parameter-efficient fine-tuning (PEFT) methods, such as LoRA or Adapter, project a model's hidden states to a lower dimension, allowing pre-trained models to adapt to new data through this low-rank bottleneck. However, PEFT tasks involving multiple modalities, like vision-language (VL) tasks, require not only adaptation to new data but also learning the relationship between different modalities. Targeting at VL PEFT tasks, we propose a family of operations, called routing functions, to enhance VL alignment in the low-rank bottlenecks. These feature routing functions adopt linear operations and do not introduce new trainable parameters. In-depth analyses are conducted to study their behavior. In various VL PEFT settings, the routing functions significantly improve performance of the original PEFT methods, achieving over $20\%$ improvement on VQAv2 (RoBERTaLarge + ViT-L/16) and $30\%$ on COCO Captioning (GPT2-medium + ViT-L/16). Also when fine-tuning a pretrained multimodal model such as CLIP-BART, we observe smaller but consistent improvements across a range of VL PEFT tasks. Our code is available at https://github.com/tingyu215/Routing_VLPEFT. + +Keywords: Vision and Language $\cdot$ Parameter-Efficient Fine-Tuning $\cdot$ Low-Rank Approximation + +# 1 Introduction + +When fine-tuning pre-trained foundation models for downstream tasks, studies show that a relatively high accuracy can be achieved by acting on a much smaller dimension compared to the original dimension of the pre-trained model's feature space. The minimum dimension necessary to approximate the optimization problem of the downstream task is termed the intrinsic dimension [1, 21]. + +The existence of such intrinsic dimension explains the success of parameter-efficient fine-tuning (PEFT) methods with low-rank bottlenecks for model fine-tuning. Two of the most widely applied techniques are Adapter [9] and LoRA [10]. We demonstrate the use of Adapter and LoRA within a Transformer block in Fig. 1. Despite being added at different positions within the Transformer block, Adapter and LoRA exhibit some key characteristics in their architectural + +design. Specifically, they first map the hidden states $x_{H}$ to a lower dimension $r$ through a down-projection mapping $W_{down}$ . The down-projected states are then projected back to the original dimension $d$ via an up-projection mapping $W_{up}$ . In the PEFT context, only the parameters of the colored modules in Fig. 1 are updated, while the remaining parameters are unchanged. Since the intermediate dimension $r$ is much smaller than the original dimension $d$ , these methods significantly reduce the number of parameters that need to be updated. + +In uni-modal tasks, where $x_{H}$ contains information from a single modality, $W_{down}$ effectively compresses $x_{H}$ into a feature space of lower dimension. This compression usually suffices provided that $r$ is larger than the intrinsic dimension of the downstream task [10]. However, PEFT methods with low-rank bottlenecks, such as Adapter and LoRA, are currently not designed for down-stream tasks involving another modality, such as vision-language (VL) tasks. Even for a pre-trained multimodal foundation model with $x_{H}$ containing information from both modalities, conceptually, directly compressing $x_{H}$ with $W_{down}$ poses challenges, as it is unclear how to effectively balance the two modalities and to enforce alignment between the modalities with this simple linear mapping. Therefore, in this paper we investigate whether it is possible to better route the features through the low-rank bottleneck with some extra guidance. An intuitive ap + +proach is standard cross-attention. However, it introduces additional parameters through query/key/value projections. Moreover, softmax-based operations not only increase time complexity but are limited to dealing with features of length $>1$ . This is incompatible with the PEFT setting, where the goal is to minimize the number of trainable parameters and computational complexity, while maintaining versatility across different types of tasks (including those that use only the [CLS] features). In this work, we introduce feature routing functions in VL PEFT with low-rank bottlenecks. They rely on linear operations and do not introduce extra trainable parameters. + +In summary, our contributions include: + +- We identify the problem of feature learning in the low-rank bottlenecks of standard PEFT methods considering VL downstream tasks. + +![](images/aa0083a8b1b11a6e9def3f4d3f2507a2625b734bde455f26536c5c1e98bf09e0.jpg) + +![](images/17050b95a1c5c37324615644bcb41b4a7c635de76ed152ddfac9697bfd2efdf1.jpg) +Fig.1: Illustration of LoRA and Adapter used in a Transformer block. Only the green modules are updated in PEFT. We identify the low-rank bottleneck with the orange rectangle. The Adapters are sequentially added to the Transformer block.3 + +- We introduce a family of parameterized operations in the low-rank bottleneck, termed routing functions, that improve the performance of VL tasks. +- We conduct extensive experiments on different architectures and tasks, and provide insights on using routing functions in VL PEFT tasks. + +# 2 Related Works + +Adapter-based PEFT To reduce the tuning costs of pre-trained language models (LMs), the Adapter layer was first introduced in [9]. It first down-projects the model's hidden state, then up-projects it back to its original dimension, incorporating a nonlinear function after the down-projection. Freezing the pretrained weights of the LM and tuning them with sequentially added Adapter layers achieves performance comparable to full fine-tuning. Since then, various Adapter variants have been proposed to advance the field of PEFT. The focus of developing Adapter variants has been on improving efficiency [18,28,34] and exploring alternative architectures [8,14,24]. Despite being surpassed by its variants on various benchmarks, the original sequential Adapter is still widely used in practice and offers stable results across tasks. + +LoRA-based PEFT LoRA [10] adopts the same "down-projection then up-projection" approach as the Adapter, with no nonlinear function in the low-rank bottleneck. LoRA is added to the query and value mappings in the attention modules of the pre-trained LM. Important follow-up works include dynamic rank adjustment [38, 41] and quantization adaptation [5, 40]. Moreover, VeRA [17] and NOLA [16] further minimize the number of parameters by leveraging frozen random matrices. LoRA-based PEFT methods have become standard for tuning large language models and large multimodal foundation models. + +Other PEFT methods Beyond Adapter and LoRA, prompt-based tuning methods [13, 19, 22] have shown strong performance for tuning both language and vision models. LST [36] learns a compact side network during tuning. BitFit [2] only adjusts the bias terms during fine-tuning, and achieves comparable performance as other PEFT methods. Methods [8, 29, 39, 42] also explore combining PEFT methods to achieve superior performance. + +Vision-Language PEFT VL-Adapter [37] is the pioneering work in the field of Vision-Language (VL) PEFT. It empirically studies the use of different PEFT methods on four different VL tasks in a multitask learning setting. VL-Adapter uses CLIP-BART, a similar model architecture as VL-T5 [33]. CLIP-BART uses a projection layer to a connect CLIP [31] vision encoder to BART [20] or T5 [33]. CLIP-BART is tuned using a combination of prompt tuning and other PEFT methods. We conduct extensive experiments using CLIP-BART and VL-Adapter as presented in Sec. 5. As an extension to VL-Adapter, VL-PET [11] combines an Adapter with a rather complicated granularity-controlled mechanism. I-Tuning [27] explores alternative model architectures for VL PEFT tasks, and proposes a parallel Adapter with cross-attention to integrate visual information into a pre-trained LM. Since our work is the first that introduces routing functions into VL PEFT, we limit the comparison of our method to + +![](images/883fbcf05965f4d4351d5f214883d81dbff7f4aa50fcdc28177442dfdb82da61.jpg) +Fig. 2: Illustration of our method. Left: Conventional PEFT methods with low-rank bottleneck first map the hidden states $x_{H}$ from a high dimension $d$ to a lower dimension $r$ via down-projection map $W_{down}$ . Then, it maps the hidden states back to the original dimension $d$ via $W_{up}$ . Middle: Our method utilizes the same architecture, but updates the features via a routing function in the low-rank bottleneck. Specifically in a VL task, for the features $x_{R}$ that we want to align to, we use $W_{down}$ to down-project them as $W_{down}x_{R}$ . Then the routing function routes $W_{down}x_{R}$ and $W_{down}x_{H}$ in the low-rank bottleneck. Right: Different routing functions. In Adapter, routing functions are added before nonlinear activation functions. + +![](images/dba17a9499953a423af3f9a7fe68656dcec9b02526d8b98119082da6729b5c66.jpg) + +![](images/e8600ef2c84b203c49cba6c89753629603e094503b30a84aa31dfb12caa91e9a.jpg) + +VL-Adapter and I-Tuning, leaving the exploration of more complicated PEFT methods like VL-PET to future work. + +# 3 Methodology + +As illustrated in Fig. 2 (Left), standard PEFT modules with low-rank bottleneck, i.e. LoRA/Adapter, learn a down-projection mapping $W_{down}$ followed by an up-projection mapping $W_{up}$ , with no routing in between. These modules are added in each Transformer block of the pre-trained foundation model (Fig. 1). As presented in Fig. 2 (Middle), our method adds a routing function in the low-rank bottleneck. This routing function is designed to align two sets of features, in particular, the extra visual features $x_{R}$ to be integrated in the VL tasks and the hidden states $x_{H}$ of each Transformer block of the foundation model. + +The design of a routing function follows two 'parameter-efficient' principles: + +- A linear operation is used to reduce computation overload. +- No extra trainable parameters are introduced. + +In this paper, we explore four types of routing functions. Given $x_{H}$ and $x_{R}^{4}$ , after the $W_{down}$ mapping, we obtain $x_{t} = W_{down}x_{H}$ and $x_{v} = W_{down}x_{R}$ . Here $x_{t}$ is of shape $(L_{t},r)$ and $x_{v}$ is of shape $(L_{v},r)$ , where $L_{t} / L_{v}$ and $r$ are the feature length (i.e. the number of tokens) and the bottleneck dimension or rank, + +respectively. $^5$ In practice, we always use $L_v = 1$ , i.e. we use a single token / global feature vector to represent the visual content. In case of multiple tokens, we use average pooling to reduce it to a single feature vector. $^6$ The proposed routing functions operate as: + +- Element-wise multiplication: $x_{t} \circ x_{v}'$ , where $x_{v}'$ denotes $x_{v}$ . reshape_as( $L_{t}, r$ )7 +- Element-wise addition: $x_{t} + x_{v}^{\prime}$ , where $x_{v}^{\prime}$ denotes $x_{v}$ .reshape_as(L_t, r) +- Matrix multiplication 1: $x_{t}(x_{v})^{\top}x_{v}$ +- Matrix multiplication 2: $x_{t}x_{v}^{\prime \prime}$ , where $x_{v}^{\prime \prime}$ denotes $x_{v}$ .reshape_as(r,r) + +Intuitively, $x_{t} \circ x_{v}^{\prime}$ , reinforces or attenuates individual elements of $x_{t}$ , using the corresponding element of $x_{v}$ as scaling factor. The second routing function, $x_{t} + x_{v}^{\prime}$ replaces $x_{t}$ with the sum of $x_{v}^{\prime}$ and $x_{t}$ , thus moving $x_{t}$ closer to $x_{v}$ . $x_{t}(x_{v})^{\top}x_{v}$ projects $x_{t}$ onto $x_{v}$ , i.e. only keeping the component of $x_{t}$ that is aligned with $x_{v}$ . Finally, $x_{t}x_{v}^{\prime\prime}$ replaces $x_{t}$ with a rescaled version of $x_{v}$ , with a different scale factor for each row given by the sum over all elements of the corresponding row of $x_{t}$ . In all these cases, it should be kept in mind that both $x_{v}$ and $x_{t}$ are derived from $x_{H}$ and $x_{R}$ via the down-projection $W_{down}$ . Parameters of $W_{down}$ can be learned so as to select particular dimensions of the original $x_{H}$ and $x_{R}$ , on which to apply the above operations before up-projecting them again with $W_{up}$ . + +The features $x_{R}$ contain information from the modalities to which we aim to align $x_{H}$ . The choice of $x_{R}$ can be different for different architectures, for instance, [CLS] features from a ViT in case of an encoder-only or decoder-only LM. We further explain the motivation of the choices of $x_{R}$ for different types of architectures in Sec. 4.1 and Sec. 5.1. + +# 4 Experiments with En(De)coder-Only Language Models + +To explore the value of routing functions, in this section, we start from encoder-only and decoder-only language models (LMs) and incorporate visual information from pre-trained vision encoders for vision-language (VL) parameter-efficient fine-tuning (PEFT) tasks. The section begins with an overview of our experimental design and implementation details, and concludes with a discussion on the comprehensive experimental results obtained. + +# 4.1 Experimental Design + +Here, we study VL PEFT tasks in their cleanest form. Specifically, (1) No additional mapping networks are adopted, so the feature routing functions form the only module that contributes to the final performance; (2) Only the parameters of the PFET modules in LMs are updated, where we learn the VL alignment; (3) + +![](images/f16742f4e8d68edc424b1d0026d006f4863c7e9c61553f2d167ef69bc0189826.jpg) +Fig.3: Pipeline for VQA and Image Captioning. Only the added classifier (in VQA) and the PEFT modules of the RoBERTa encoder and GPT decoder are tuned. Note that here $x_{H}$ are the text inputs with the visual [CLS] features prepended. + +![](images/139a592d9cee0ec8b3811ec9d38446f4adf950a5aabdf741c852d3e6e00147a9.jpg) + +![](images/af2c687005cf2f604c5af07550a389bb4546c179772493a455da5d749d226dc7.jpg) + +No multimodal tasks (e.g. contrastive language-image pre-training in CLIP [31]) are included in the pre-training phases of the backbone models. We only learn the multimodal embedding space during VL PEFT. These principles guarantee that VL alignment originates strictly from the PEFT modules. + +To comprehensively study the role of routing functions in VL tasks, we consider both discriminative and generative VL tasks. We select visual question answering (VQA) and Image Captioning, two of the most widely studied VL tasks, as the representative task for discriminative VL tasks and generative VL tasks, respectively. + +We present the pipelines for VQA and Image Captioning in Fig. 3. We use ViT [6] as our vision backbone. As for the LM backbone, we use the encoder-only RoBERTa [25] and the decoder-only GPT2 [32] for VQA and Image Captioning, respectively. The [CLS] feature from the ViT is prepended to the text inputs and acts as $x_{R}$ , enabling the learning of VL alignment through PEFT, with or without the integration of our routing functions. Prepending visual features to text inputs is actually the predominant way of learning VL alignment in PEFT [11, 37] without routing functions. The ViT and LMs adopt the same dimensions for their feature spaces, so we do not need additional mappings to match the dimensions. + +# 4.2 Implementation Details + +We adopt ViT-B/16 pre-trained on ImageNet-21k as our vision backbone, which we keep frozen during training. In LoRA, we set scaling factor $\alpha = r$ in all setups. For VQAv2, we use RoBERTa $_{\text{base}}$ as our LM backbone, with a two-layer multi-layer perceptron (MLP) classifier on top following [37]. We train our model using the AdamW optimizer [26] with learning rate=3e-4 and batch size of 512 for 20 epochs, where we warm up the first $5\%$ of the steps. For the captioning task, COCO Cap., we use GPT2-base as our LM backbone. We train our model for 20 + +epochs using the AdamW optimizer with batch size of 40 and learning rate $= 3\mathrm{e} - 5$ where we warm up the first 5000 steps. For more details, see Appendix A. + +# 4.3 Results + +Overall performance on VQA We first evaluate our routing functions on the discriminative task VQAv2 with RoBERTabase and ViT-B/16, see Tab. 1. + +Without the addition of routing functions, models with LoRA and Adapter reach a similar level of performance. All four types of routing functions significantly improve the performance of the original PEFT method in terms of accuracy of the answer. The dimension of the bottleneck does not have a strong impact on the final performance. Sometimes we can even obtain better performance with lower dimensions. It is a sign that the intrinsic dimension of the VL task is not + +Table 1: Accuracy and improvement over the baseline (in blue) obtained by different types of routing functions on VQAv2. The best results and second best results are in bold and underlined, respectively. + +
Routingr = 64r = 128
FunctionsLoRAAdapterLoRAAdapter
None44.1544.1644.4544.28
xt○xt'53.51+9.3652.78+8.6253.86+10.4153.01+8.73
xt+xt'52.60+8.4553.94+9.7852.88+8.4353.95+9.67
xt(xt)Txt53.88+9.7354.48+10.3253.09+8.6455.06+10.78
xtxt''54.21+10.0654.96+10.8051.88+7.4354.38+10.10
+ +high, just as observed in uni-modal tasks. Moreover, we observe that the routing functions involving matrix multiplications (i.e. $x_{t}(x_{v})^{\top}x_{v} \& x_{t}x_{v}^{\prime \prime}$ ) perform quite well in most scenarios for both LoRA and Adapter, while the routing functions involving element-wise operations (i.e. $x_{t} \circ x_{v}^{\prime} \& x_{t} + x_{v}^{\prime}$ ) are just slightly behind. + +Table 2: Results in terms of accu- Table 3: Results in terms of accuracy with racy with different routing functions on different routing functions on VQAv2 sorted VQAv2 sorted by answer type. The "all" by question type. The "all" column reports column reports overall accuracy. overall accuracy. + +
Routing FunctionsallLoRA, r = 128Adapter, r = 128
yes/no numberotherallyes/no numberother
None44.4568.6731.2929.2044.2868.3831.19
xt·ox'53.8672.2935.3644.4853.0171.7233.34
xt+xt'52.8872.0334.2043.0053.9572.4335.38
xt(xt)Tx53.0970.6335.4044.1955.0674.0933.59
xtx''51.8869.9233.2942.8254.3873.0033.94
+ +
Routing +FunctionsLoRA, r = 128Adapter, r = 128
allwhere colortimepersonallwhere colortimeperson
None44.4522.6440.0014.2945.1744.2822.9340.1513.7044.94
xt × xv'53.8635.7051.6923.0762.7453.0135.5952.6320.7362.28
xt + xv'52.8836.1249.3321.2262.9553.9537.0852.0821.6663.76
xt(x)Tx53.0936.3352.8622.2961.7455.0636.9253.6624.4463.85
xtxv''51.8834.1951.6022.6361.1454.3835.6452.7721.8562.50
+ +To dive deeper into the behaviors of different types of routing functions, we evaluate the performance on different answer types as presented in Tab. 2. The largest improvements are obtained for the "other" category for all four routing functions. It can be seen that different routing functions excel for different answer types. For example, even when the two element-wise operations reach lower overall accuracy, they lead to higher or comparable accuracy on number-type answers. As for the routing functions involving matrix multiplications, better performance on other-type answers is obtained. For instance, with relatively lower overall accuracy reached in LoRA, matrix multiplications reach comparable performance on other-type answers. No clear pattern can be observed for yes/no + +Table 4: Results of different types of routing functions on COCO Cap. The Avg. score is the averaged result from BLEU-4, METEOR, ROUGE-L, CIDEr, and SPICE. The best and second best results are in bold and underlined, respectively. + +
Routing FunctionsLoRA, r=64LoRA, r=128
BLEU-4METEORROUGE-LCIDErSPICEAvg.BLEU-4METEORROUGE-LCIDErSPICEAvg.
None18.317.436.855.811.828.0221.020.943.470.314.734.06
xt×ov'26.1+7.823.7+6.348.0+11.888.7+32.917.3+5.540.88+12.8625.4+4.423.7+2.848.2+4.887.9+17.617.7+3.040.58+6.52
xt+xov'22.2+3.920.3+2.941.0+4.273.5+17.714.8+3.034.36+6.3425.4+4.423.1+2.247.1+4.787.1+16.817.0+2.339.94+5.88
xt(xv)Txv24.8+6.522.6+5.245.4+8.684.9+29.116.8+5.038.90+10.8824.9+3.922.7+1.845.4+2.083.8+13.516.9+2.238.74+4.68
xtx''23.9+5.621.9+4.543.9+7.180.5+14.716.2+4.437.28+9.2625.5+4.522.9+2.046.0+2.685.2+14.917.2+2.539.36+5.30
Routing FunctionsAdapter, r=64Adapter, r=128
BLEU-4METEORROUGE-LCIDErSPICEAvg.BLEU-4METEORROUGE-LCIDErSPICEAvg.
None15.918.537.061.614.129.4216.118.436.261.513.929.22
xt×ov'24.6+8.723.1+4.646.4+9.484.5+22.917.2+3.139.16+9.7423.6+7.522.6+4.244.8+8.682.8+21.317.1+3.238.18+8.96
xt+xov'21.0+5.121.5+3.042.4+5.475.0+14.416.0+1.935.18+5.7619.3+3.220.8+2.440.9+4.775.2+13.816.2+2.334.48+5.26
xt(xv)Txv26.1+10.223.2+4.746.9+9.985.4+23.817.3+3.239.78+10.3624.1+8.022.4+4.044.2+8.082.7+21.216.9+3.038.06+8.84
xtx''26.6+10.723.0+4.546.8+9.885.8+24.217.2+3.139.88+10.4626.0+9.923.5+5.146.8+10.686.3+24.817.8+3.940.08+10.86
+ +types of answers, which could be explained by the large variety in content of these questions. + +Tab. 3 shows our method's performance across various question types. Four distinct question types test the method's reasoning capabilities concerning fine-grained visual information: spatial information ("where", e.g. Where is the surfboard?), color information ("color", e.g. What color is the bike?), time information ("time", e.g. What time is on the clock?), and person information ("person", e.g. What is the man holding?).9 Adding routing functions significantly improves the performance of LoRA and Adapter. Notably, different routing functions seem to capture different aspects of the visual information. $x_{t} + x_{v}^{\prime}$ shows great performance while reasoning on spatial information (where) and person information (person). It is assumed that the addition operator reinforces feature values of content both present in the question and image, which might be beneficial for spatial and person questions. Similarly, $x_{t}x_{v}^{\prime\prime}$ performs exceptionally well for time information (time). With $x_{t} \circ x_{v}^{\prime}$ , we obtain relatively better performance on the color information (color). $x_{t}(x_{v})^{\top}x_{v}$ presents consistent strong abilities while aligning spatial information (where) and color information (color). + +In Appendix B, we show an additional ablation study on replacing visual features $(x_{R})$ with random noise or all-ones features. We also compare inference time of LoRA/Adapter with LoRA/Adapter + routing functions, where we observe adding routing functions to PEFT methods does not reduce efficiency. + +Overall performance on image captioning Evaluation of image captioning relies on the overall assessment of the generated captions. We report multiple text generation metrics and their average scores in Tab. 4. The addition of routing functions again leads to significant performance gains for all metrics in all scenarios. As was the case with VQAv2, different patterns are observed. Changing the intermediate dimension $r$ from 128 to 64 seems to have a larger impact on the model performance with conventional LoRA, where without the routing functions the CIDEr score drops from 70.3 to 55.8. In contrast, models augmented with routing functions remain robust and per + +formant across setups, despite the changes in intermediate dimension. $x_{t} \circ x_{v}^{\prime}$ and $x_{t}x_{v}^{\prime \prime}$ perform well, possibly due to their alignment power. We also report the performance of Adapter with larger backbones, namely GPT2-medium and ViT-L/16, in Appendix B. Consistent and significant improvements in performance are observed, showing the efficacy of our routing functions on the task. + +# Comparison to cross-attention + +Here we compare the results of the integration of routing function $x_{t}(x_{v})^{\top}x_{v}$ with these of integrating cross-attention between the textual and the visual representation in image captioning. Indeed, in principle, a cross-attention function can also be considered as a kind of routing function with nonlinear operations. + +I-Tuning [27] implements an Adapter with cross-attention for image captioning. To have a fair comparison, we + +follow the settings of I-Tuning for both LoRA and Adapter. Specifically, (1) instead of using the visual features, we pretend the "" token to the input text. (2) We use the last hidden state from ViT as key and value for cross-attention and as $x_{R}$ for routing functions. This change is needed, since the [CLS] feature is of length 1, which does not work for cross-attention due to the Softmax function. + +In the case of LoRA, we also implement a variant with separate $W_{down}$ mappings for $x_{R}$ and $x_{H}$ . This brings LoRA with routing function closer to LoRA with cross-attention. Without separate mapping refers to the setting used in previous experiments, where $x_{R}$ and $x_{H}$ share the same $W_{down}$ . As shown in Tab. 5, despite of having less parameters, $\mathrm{LoRA} + x_t(x_v)^{\top}x_v$ (w/ separate mapping) achieves significantly better performance than LoRA with cross-attention. The setup of $\mathrm{LoRA} + x_t(x_v)^{\top}x_v$ (w/o separate mapping) outperforms LoRA with cross-attention by a large margin using a similar amount of parameters. + +For Adapter, the cross-attention module has the same intermediate dimension as the $W_{down}$ in Adapter with routing function. Naturally, the separate query/ key/value mappings make that the Adapter with cross-attention has more than double the number of parameters compared to the Adapter equipped with the routing function $(x_{t}(x_{v})^{\top}x_{v})$ . Both achieve comparable performance (Tab. 5). + +To conclude, Tab. 5 shows that integrating $x_{t}(x_{v})^{\top}x_{v}$ in LORA and Adapter yields better performance given the same amount of parameters, or is similar in performance while using substantially less parameters. For VL tasks like image captioning, integrating a routing function to guide feature representations in low-rank bottlenecks has superior abilities compared to cross-attention. + +Table 5: Comparing routing functions with cross-attention on COCO Cap. †: parallel Adapter as used in [27]. Separate Map.: separate mapping for query/ key/value (in cross-attn.) or features $x_{R} / x_{H}$ (in $x_{t}(x_{v})^{\top}x_{v}$ ). Here $r = 128$ . Avg. indicates the same average scores as in Tab. 4. Full evaluation in Appendix B. + +
PEFTSeparateMap.AlignmentParam.BLEU-4CIDErAvg.
LoRA✓Cross-attn.4.786M28.792.243.02
LoRA✓xt(xt)Txtxt3.932M30.7+2.099.4+7.245.68+2.66
LoRAXxt(xt)Txtxt4.746M30.0+1.399.0+6.845.38+2.36
Adapter†✓Cross-attn.4.732M30.799.845.70
Adapter†Xxt(xt)Txtxt1.830M30.8+0.198.8-1.045.48-0.22
+ +Incorporating nonlinear activation functions into routing functions We have proposed linear routing functions for efficiency reasons, but they do not need to be exclusively linear. Taking inspirations from the linear attention variants [4,15,30], where the linear operation is often equipped with nonlinear activation functions to ensure non-negativity for + +Table 6: Experiments on adding nonlinear activation functions in routing functions. We use $r = 64$ here. + +
PEFT MethodRouting FunctionsVQA AccuracyCOCO Cap.
BLEU-4CIDErAvg.
LoRAxt(xv)Txv53.8824.884.938.90
LoRAReLU(xt)ReLU(xv)Txv53.4525.386.139.84
Adapterxt(xv)Txv54.4826.185.439.78
AdapterReLU(xt)ReLU(xv)Txv54.6325.084.739.02
+ +similarity scores between query and key, we explore whether a similar operation affects the representation power of the routing functions in the low-rank bottleneck. For routing function $x_{t}(x_{v})^{\top}x_{v}$ , we use the nonlinear activation function $ReLU(\cdot)$ as in previous work [30], and transform it to $ReLU(x_{t})ReLU(x_{v})^{\top}x_{v}$ . The results are presented in Tab. 6. With or without $ReLU(\cdot)$ , our routing function achieves similar performance on VQA and image captioning. + +Qualitative Analysis We first qualitatively compare the VL alignment pattern of LoRA-tuned GPT2 with or without a routing function on image captioning. An example is presented in Fig. 4. With the addition of routing function $x_{t}(x_{v})^{\top}x_{v}$ , the model pays more attention to the nouns (e.g. man, a big screen) and the verbs (e.g. watching), which are essential in learning the VL alignment. At the same time, it pays less attention to meaningless words like "something". This is in line with the finding that the routing function stresses components of $x_{t}$ that are aligned in the image. + +![](images/197843bbade820f1df61436b07e3b53b5a7ea586049e2db61d95ece54853f6fe.jpg) +Fig. 4: Example of average attention weights from the last layer of GPT2. We use the final checkpoint trained on COCO Cap. with $r = 64$ . IMAGE: visual [CLS] feature. + +Next we qualitatively examine some examples from the inference phase for VQA and Image Captioning. We present examples in Fig. 5 where conventional LoRA fails. For VQA, previous experiments in Tab. 3 indicate that our routing functions work particularly well on color related questions. As also shown in Fig. 5, conventional LoRA fails to align the color red to the lollipop. The model tuned with conventional LoRA also fails to correctly answer other types of questions as presented in the figure. Without routing functions, captions generated from the LM tuned with conventional LoRA often make mistakes in compre + +![](images/845e1ec6070b8ec8a5d1018e8d89fc7f390c5db4aef07b82a89cf8399345cf3c.jpg) +Fig. 5: Qualitative examples from VQA and Image Captioning tasks. We present cases where the model trained with conventional LoRA fails and results obtained from a model trained with LoRA with a routing function. We use $r = 64$ . + +![](images/0053a476f1e008c39268be92e8b5647fc7c36b2e00fb2160454f9738e25778ea.jpg) + +hending the visual information. Occasionally, we observe a sequence of blank spaces as presented in the first example using conventional LoRA. + +# 5 Experiments with Encoder-Decoder Language Models + +In this section, we study the VL PEFT tasks in a more complicated setting, where we consider an encoder-decoder LM with an additional mapping network dedicated for the vision encoder. The model is trained in a multi-task learning setting, where four different VL tasks are jointly learned, as in [3,37]. Despite the increased complexity of optimizing the network in this scenario, we show that integrating routing functions into the low-rank bottlenecks yields performance enhancements. Moreover, consistent patterns are observed for different scenarios, which shed lights on future research to improve VL PEFT tasks. + +# 5.1 Experimental Design + +Following [37], we use the CLIP-BART architecture to study the effect of our routing functions on the VL PEFT tasks in a multitask learning setting. CLIP-BART is a widely used architecture for VL tasks. We present the simplified model architecture of CLIP-BART with Adapters and the integrated routing functions in Fig. 6. The original CLIP-BART uses a trainable visual projection module to map the visual features extracted by CLIP to the embedding space of BART [20]. CLIP-BART uses task-specific prompts as the textual input: For example, in Fig. 6, "" is added as prompt to the text for VQA. The mapped visual features are then prepped to the textual input. Together, they serve as the input to BART. CLIP-BART is trained and evaluated in a multitask setting, that is, VQAv2 [7] and GQA [12] for visual question answering, NLVR $^2$ [35] using natural language for visual reasoning, and COCO Cap. [23] for image captioning. + +CLIP-BART unifies the four tasks into one text generation problem following [3], where four different task prompts distinguish the task types. CLIP-BART can be considered as the representative of the current state-of-the-art on VL PEFT tasks. + +To integrate our routing functions into the model, two types of inputs $x_{R}$ can be considered, namely the visual features after the visual projection, and the last encoder hidden state as presented in Fig. 6. The visual features only contain visual information. Because the last encoder hidden state contains both textual and visual information, it can only be added to the decoder. We experiment with both LoRA and Adapter as the PEFT method for fine-tuning CLIP-BART. In this multitask learning setting, one can use (1) a single LoRA/Adapter shared across tasks or (2) multiple LoRAs/Adapters without parameter + +sharing, where each task has its own LoRA/Adapter. We start by exploring the impact of different routing functions on the single LoRA/Adapter setting. Then, based on the observations obtained in the setting with single Adapter, we explore ways of using the routing functions in the multiple Adapters setting. + +![](images/a5e03c18dc9aeea9b48c2643a0f2cf2d65354fe1e58ae3fafc24bb29369af54e.jpg) +Fig. 6: Architecture of CLIP-BART with Adapter. The standard Adapter architecture is adopted (see Fig. 1). Only the modules in green are trainable. To apply routing functions, we use the pooled last encoder hidden state as feature $x_{R}$ , which is added to the Adapters in the decoder. + +# 5.2 Implementation Details + +Following [11, 37], we use CLIP-ResNet101 as our vision encoder [31]. We use $\mathrm{BART}_{\mathrm{base}}$ as our encoder-decoder LM as in [37]. Following [37], we use $r = 96$ for all experiments. The training pipeline in [37] is used for a fair comparison. Specifically, (1) we use the same seed for all experiments. (2) Due to limited resources, we do not perform an extensive hyper-parameter search as in [37]. (3) We train the model with AdamW optimizer and linear decay for 20 epochs. We warm-up the learning rate from 0 to the highest in the first 2 epochs. (4) The batch size is set to 500. For more details, we refer to Appendix A. + +# 5.3 Results + +Experiments with a single Adapter/LoRA We first report results of experiments on using single Adapter/LoRA shared across tasks in Tab. 7. In this case, the Adapter/LoRA is trained with data from four different tasks, and naturally we can only use one type of routing function throughout the experiments. + +Table 7: Experiments with different types of routing functions using CLIP-BART with single LoRA/Adapter. Following [11] $(\dagger)^{11}$ , we report the average CIDER score for COCO Cap., accuracy for the other three tasks. We also report the average accuracy from VQA, GQA and NLVR2. We bold and underline the best and second best results. + +
Routing FunctionsSingle LoRASingle Adapter
VQAGQANLVR2Avg.COCOCap.VQAGQANLVR2Avg.
None†65.1553.6672.5863.80115.0165.7654.1673.1964.37
xt○x′v65.6853.9673.4264.35113.9465.9254.3474.2364.83
xt+x′v65.1453.7373.5164.13114.9665.8954.1873.9064.66
xt(xv)Txv64.9453.5673.6064.03117.8065.8453.6574.3164.03
xtx''v64.8453.1372.9863.65119.2665.8353.6173.2764.24
+ +Tab. 7 shows that the integration of the routing functions in the single LoRa and single Adapter architecture results in improved performance in nearly all scenarios. But just like the experiments with encoder-only and decoder-only LMs in Sec. 4.3, different routing functions show different patterns. The routing functions involving matrix multiplication, that is, $x_{t}(x_{v})^{\top}x_{v}$ and $x_{t}x_{v}^{\prime \prime}$ , perform exceptionally well in the COCO captioning task, while the other two routing functions involving element-wise operations, i.e. $x_{t}\circ x_{v}^{\prime}$ and $x_{t} + x_{v}^{\prime}$ , show more potential on QA tasks (VQA/GQA). As for the NLVR $^{2}$ task, $x_{t}(x_{v})^{\top}x_{v}$ and $x_{t}\circ x_{v}^{\prime}$ present stronger abilities to learn VL relationships. Note that achieving relatively low performance on one task does not necessarily mean that the corresponding routing function fails. For example, $x_{t}\circ x_{v}^{\prime}$ does not work well for captioning, but achieves great performance on the other three tasks. This somehow contrasts with the results obtained with encoder-only and decoder-only LMs in Sec. 4.3, where $x_{t}\circ x_{v}^{\prime}$ leads to one of the best performance on COCO Cap. The complexity of the multitask learning setup, where different loss functions are jointly optimized, might here be a confounding factor. + +Experiments with multiple Adapters We further explore the behavior of the routing functions in models with multiple Adapters, where we train a separate Adapter per task. As also observed in [37], using multiple Adapters increases the complexity of the model, and yields worse performance compared to using a single Adapter. As presented in Tab. 8, using the four routing functions with multiple Adapters leads to comparable or better performance as compared to multiple Adapters with no routing functions. Moreover, in line with what we discovered in the experiments with models with a single Adapter, different routing functions shine on different types of tasks. + +In this set-up, we also experiment with combinations of routing functions. The four combinations integrate the routing functions that work best for each individual task (e.g. $x_{t}x_{v}^{\prime}$ for COCO Cap.). The results in Tab. 8 evidence that in this multitask setting combinations of routing functions yield an improved performance compared to using no routing functions. GQA requires more fine + +Table 8: Experiments with different combinations of routing functions using CLIP-BART with multiple Adapters. Following [37] $(\ddagger)$ , we report CIDEr score for COCO Cap., accuracy for the other three tasks. We also report the average accuracy from VQA, GQA and NLVR2. We bold and underline the best and second best results. + +
Routing FunctionsRouting FunctionsPerformance
VQAGQANLVR2COCO Cap.VQAGQANLVR2Avg.COCO Cap.
None†XXXX65.454.069.863.07114.3
xt○x′v✓✓✓✓65.6153.5170.7863.30113.72
xt+x′v✓✓✓✓65.1953.8470.9963.34115.64
xt(xv)Txv✓✓✓✓64.9352.1170.2562.43116.20
xtx′v✓✓✓✓65.3252.6770.5862.86117.92
comb1xt○x′vxt○x′vxt(xv)Txvxtx′v66.3754.2169.2763.28116.94
comb2xt+x′vxt+x′vxt(xv)Txvxtx′v65.1453.3770.8863.13117.58
comb3xt○x′vxt○x′vxt○x′vxtx′v66.3453.7770.0263.38117.73
comb4xt+x′vxt+x′vxt○x′vxtx′v65.4752.9971.0163.16118.34
+ +grained image representation. For efficiency reason, we only adopt $x_{v}$ as the down-projection from the pooled last encoder hidden state. Nevertheless, comb1 can recover from this shortcoming. More detailed image representations as used in the cross-attention experiments of Tab. 5 might still enhance the results. + +Overall, our research has focused on introducing routing functions in low-rank feature approximations when fine-tuning models for downstream (in our case VL) tasks based on pre-trained foundation models. As a pioneering work on the topic, for obtaining results we did not exhaustively search for all possible choices of hyperparameters. Instead, we have focused on revealing the problem of low-rank bottlenecks and on providing a successful solution. We leave an exhaustive grid search for optimal hyperparameters as future work. + +# 6 Conclusion + +We have introduced feature routing functions to help guide the feature learning in low-rank bottlenecks in architectures that are typical for parameter-efficient fine-tuning such as LoRA and Adapter. We have shown their validity in vision-language tasks, where we have focused on linear functions that do not require any extra parameters. We have demonstrated performance gains in single-task and multi-task settings in combination with a variety of foundation models. + +Routing functions open many avenues for future research. Given the importance of foundation models that are trained in a self-supervised way on large data, there is the need to efficiently fine-tune them for downstream tasks. When computing low-rank approximations of the features, our work shows that the extra guidance of the routing functions is important. We see many possibilities in experimenting with other types of functions in other tasks. For instance, linear and other functions are also suited to model constraints that reflect external knowledge or expectations on the behavior of a model. Future work will show whether they are useful to guide the feature routing in low-rank bottlenecks. + +# Acknowledgements + +This work is funded by the Flanders AI Research Program, European Research Council Advanced Grant H2020-ERC-2017-ADG 788506, and the China Scholarship Council. + +# References + +1. Aghajanyan, A., Gupta, S., Zettlemoyer, L.: Intrinsic dimensionality explains the effectiveness of language model fine-tuning. In: Zong, C., Xia, F., Li, W., Navigli, R. (eds.) Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 7319-7328. Association for Computational Linguistics, Online (Aug 2021). https://doi.org/10.18653/v1/2021.acl-long.568, https://aclanthology.org/2021.acl-long.568 +2. Ben Zaken, E., Goldberg, Y., Ravfogel, S.: BitFit: Simple parameter-efficient fin-tuning for transformer-based masked language-models. In: Muresan, S., Nakov, P., Villavicencio, A. (eds.) Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). pp. 1-9. Association for Computational Linguistics, Dublin, Ireland (May 2022). https://doi.org/10.18653/v1/2022.acl-short.1, https://aclanthology.org/2022.acl-short.1 +3. Cho, J., Lei, J., Tan, H., Bansal, M.: Unifying vision-and-language tasks via text generation. In: ICML (2021) +4. Choromanski, K.M., Likhosherstov, V., Dohan, D., Song, X., Gane, A., Sarlos, T., Hawkins, P., Davis, J.Q., Mohiuddin, A., Kaiser, L., Belanger, D.B., Colwell, L.J., Weller, A.: Rethinking attention with performers. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=Ua6zuk0WRH +5. Dettmers, T., Pagnoni, A., Holtzman, A., Zettlemoyer, L.: Qlora: Efficient finetuning of quantized llms. In: Oh, A., Neumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 10088-10115. Curran Associates, Inc. (2023), https://proceedings.neurips.cc/paper_files/paper/2023/file/1feb87871436031bdc0f2beaa62a049b-Paper-Conference.pdf +6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=YicbFdNTTy +7. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2017) +8. He, J., Zhou, C., Ma, X., Berg-Kirkpatrick, T., Neubig, G.: Towards a unified view of parameter-efficient transfer learning. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=ORDcd5Axok +9. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for NLP. In: Chaudhuri, K., Salakhutdinov, R. (eds.) Proceedings of the 36th International + +Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 97, pp. 2790-2799. PMLR (09-15 Jun 2019), https://proceedings.mlr.press/v97/houlsby19a.html +10. Hu, E.J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=nZeVKeeFYf9 +11. Hu, Z.Y., Li, Y., Lyu, M.R., Wang, L.: Vl-pet: Vision-and-language parameter-efficient tuning via granularity control. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3010-3020 (October 2023) +12. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019) +13. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision (ECCV) (2022) +14. Karimi Mahabadi, R., Ruder, S., Dehghani, M., Henderson, J.: Parameter-efficient multi-task fine-tuning for transformers via shared hypernetworks. In: Zong, C., Xia, F., Li, W., Navigli, R. (eds.) Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 565-576. Association for Computational Linguistics, Online (Aug 2021). https://doi.org/10.18653/v1/2021.acl-long.47, https://aclanthology.org/2021.acl-long.47 +15. Katharopoulos, A., Vyas, A., Pappas, N., Fleuret, F.: Transformers are RNNs: Fast autoregressive transformers with linear attention. In: III, H.D., Singh, A. (eds.) Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 5156-5165. PMLR (13-18 Jul 2020), https://proceedings.mlr.press/v119/katharopoulos20a.html +16. Koohpayegani, S.A., L, N.K., Nooralinejad, P., Kolouri, S., Pirsiavash, H.: NOLA: Networks as linear combination of low rank random basis. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=TjfXcDgvzk +17. Kopiczko, D.J., Blankevoort, T., Asano, Y.M.: VeRA: Vector-based random matrix adaptation. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=NjNfLdxr3A +18. Lei, T., Bai, J., Brahma, S., Ainslie, J., Lee, K., Zhou, Y., Du, N., Zhao, V.Y., Wu, Y., Li, B., Zhang, Y., Chang, M.W.: Conditional adapters: Parameter-efficient transfer learning with fast inference. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=IyYyKovOAj +19. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. In: Moens, M.F., Huang, X., Specia, L., Yih, S.W.t. (eds.) Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 3045-3059. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic (Nov 2021). https://doi.org/10.18653/v1/2021.emnlp-main.243, https://aclanthology.org/2021.emnlp-main.243 +20. Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., Stoyanov, V., Zettlemoyer, L.: BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In: Jurafsky, D., Chai, J., Schluter, N., Tetreault, J. (eds.) Proceedings of the 58th Annual Meeting + +of the Association for Computational Linguistics. pp. 7871-7880. Association for Computational Linguistics, Online (Jul 2020). https://doi.org/10.18653/v1/2020.acl-main.703, https://aclanthology.org/2020.acl-main.703 +21. Li, C., Farkhoor, H., Liu, R., Yosinski, J.: Measuring the intrinsic dimension of objective landscapes. In: International Conference on Learning Representations (2018), https://openreview.net/forum?id=ryup8-WCW +22. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. In: Zong, C., Xia, F., Li, W., Navigli, R. (eds.) Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 4582-4597. Association for Computational Linguistics, Online (Aug 2021). https://doi.org/10.18653/v1/2021.acl-long.353, https://aclanthology.org/2021.acl-long.353 +23. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) Computer Vision - ECCV 2014. pp. 740-755. Springer International Publishing, Cham (2014) +24. Lin, Z., Madotto, A., Fung, P.: Exploring versatile generative language model via parameter-efficient transfer learning. In: Cohn, T., He, Y., Liu, Y. (eds.) Findings of the Association for Computational Linguistics: EMNLP 2020. pp. 441-459. Association for Computational Linguistics, Online (Nov 2020). https://doi.org/10.18653/v1/2020-findings-emnlp.41, https://aclanthology.org/2020-findings-emnlp.41 +25. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized BERT pretraining approach. CoRR abs/1907.11692 (2019), http://arxiv.org/abs/1907.11692 +26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019), https://openreview.net/forum?id=Bkg6RiCqY7 +27. Luo, Z., Hu, Z., Xi, Y., Zhang, R., Ma, J.: I-tuning: Tuning frozen language models with image for lightweight image captioning. In: ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5 (2023). https://doi.org/10.1109/ICASSP49357.2023.10096424 +28. mahabadi, R.K., Henderson, J., Ruder, S.: Compacter: Efficient low-rank hypercomplex adapter layers. In: Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems (2021), https://openreview.net/forum?id=bqGK5PyI6-N +29. Mao, Y., Mathias, L., Hou, R., Almahairi, A., Ma, H., Han, J., Yih, S., Khabsa, M.: UniPELT: A unified framework for parameter-efficient language model tuning. In: Muresan, S., Nakov, P., Villavicencio, A. (eds.) Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 6253-6264. Association for Computational Linguistics, Dublin, Ireland (May 2022). https://doi.org/10.18653/v1/2022.acl-long.433, https://aclanthology.org/2022.acl-long.433 +30. Qin, Z., Sun, W., Deng, H., Li, D., Wei, Y., Lv, B., Yan, J., Kong, L., Zhong, Y.: cosformer: Rethinking softmax in attention. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=B18CQrx2Up4 +31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) + +Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 8748-8763. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/radford21a.html +32. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I.: Language models are unsupervised multitask learners. In: Preprint (2019) +33. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research 21(140), 1-67 (2020), http://jmlr.org/papers/v21/20-074.html +34. Rücklé, A., Geigle, G., Glockner, M., Beck, T., Pfeiffer, J., Reimers, N., Gurevych, I.: AdapterDrop: On the efficiency of adapters in transformers. In: Moens, M.F., Huang, X., Specia, L., Yih, S.W.t. (eds.) Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 7930-7946. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic (Nov 2021). https://doi.org/10.18653/v1/2021.emnlp-main.626, https://aclanthology.org/2021.emnlp-main.626 +35. Suhr, A., Zhou, S., Zhang, A., Zhang, I., Bai, H., Artzi, Y.: A corpus for reasoning about natural language grounded in photographs. In: Korhonen, A., Traum, D., Marquez, L. (eds.) Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. pp. 6418-6428. Association for Computational Linguistics, Florence, Italy (Jul 2019). https://doi.org/10.18653/v1/P19-1644, https://aclanthology.org/P19-1644 +36. Sung, Y.L., Cho, J., Bansal, M.: LST: Ladder side-tuning for parameter and memory efficient transfer learning. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=isPnnaTZaP5 +37. Sung, Y.L., Cho, J., Bansal, M.: Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5227-5237 (June 2022) +38. Valipour, M., Rezagholizadeh, M., Kobyzev, I., Ghodsi, A.: DyLoRA: Parameter-efficient tuning of pre-trained models using dynamic search-free low-rank adaptation. In: Vlachos, A., Augenstein, I. (eds.) Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics. pp. 3274-3287. Association for Computational Linguistics, Dubrovnik, Croatia (May 2023). https://doi.org/10.18653/v1/2023.eacl-main.239, https://aclanthology.org/2023.eacl-main.239 +39. Wang, Y., Agarwal, S., Mukherjee, S., Liu, X., Gao, J., Awadallah, A.H., Gao, J.: AdaMix: Mixture-of-adaptations for parameter-efficient model tuning. In: Goldberg, Y., Kozareva, Z., Zhang, Y. (eds.) Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. pp. 5744-5760. Association for Computational Linguistics, Abu Dhabi, United Arab Emirates (Dec 2022). https://doi.org/10.18653/v1/2022.emnlp-main.388, https://aclanthology.org/2022.emnlp-main.388 +40. Xu, Y., Xie, L., Gu, X., Chen, X., Chang, H., Zhang, H., Chen, Z., ZHANG, X., Tian, Q.: QA-loRA: Quantization-aware low-rank adaptation of large language models. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=WvFoJccpo8 +41. Zhang, Q., Chen, M., Bukharin, A., He, P., Cheng, Y., Chen, W., Zhao, T.: Adaptive budget allocation for parameter-efficient fine-tuning. In: The Eleventh Inter- + +national Conference on Learning Representations (2023), https://openreview.net/forum?id=1q62uWRJjiY +42. Zhou, H., Wan, X., Vulic, I., Korhonen, A.: Autopeft: Automatic configuration search for parameter-efficient fine-tuning. Transactions of the Association for Computational Linguistics 12 (2024) \ No newline at end of file diff --git a/2024/Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks/images.zip b/2024/Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8a2fbeb305036b5404883cd96b7fca08a7914584 --- /dev/null +++ b/2024/Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80ecefb1235d44ec64e964db54ae1ca45ab181150c7375ac83a7d6a18a761a99 +size 546855 diff --git a/2024/Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks/layout.json b/2024/Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..86dca21c9376961fe5177ffea939fee5fe31a233 --- /dev/null +++ b/2024/Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks/layout.json @@ -0,0 +1,10850 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 163 + ], + "type": "text", + "content": "Introducing Routing Functions to Vision-Language Parameter-Efficient Fine-Tuning with Low-Rank Bottlenecks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 162, + 186, + 451, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 186, + 451, + 198 + ], + "spans": [ + { + "bbox": [ + 162, + 186, + 451, + 198 + ], + "type": "text", + "content": "Tingyu " + }, + { + "bbox": [ + 162, + 186, + 451, + 198 + ], + "type": "inline_equation", + "content": "\\mathrm{Qu}^{1}" + }, + { + "bbox": [ + 162, + 186, + 451, + 198 + ], + "type": "text", + "content": ", Tinne Tuytelaars" + }, + { + "bbox": [ + 162, + 186, + 451, + 198 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 162, + 186, + 451, + 198 + ], + "type": "text", + "content": ", and Marie-Francine Moens" + }, + { + "bbox": [ + 162, + 186, + 451, + 198 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 207, + 207, + 406, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 207, + 406, + 219 + ], + "spans": [ + { + "bbox": [ + 207, + 207, + 406, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 207, + 207, + 406, + 219 + ], + "type": "text", + "content": " Department of Computer Science, KU Leuven" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 181, + 220, + 432, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 220, + 432, + 241 + ], + "spans": [ + { + "bbox": [ + 181, + 220, + 432, + 241 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 181, + 220, + 432, + 241 + ], + "type": "text", + "content": " Department of Electrical Engineering, KU Leuven {tingyu.qu; tinne.tuytelaars; sien.moens}@kuleuven.be" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 269, + 455, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 269, + 455, + 456 + ], + "spans": [ + { + "bbox": [ + 160, + 269, + 455, + 456 + ], + "type": "text", + "content": "Abstract. Mainstream parameter-efficient fine-tuning (PEFT) methods, such as LoRA or Adapter, project a model's hidden states to a lower dimension, allowing pre-trained models to adapt to new data through this low-rank bottleneck. However, PEFT tasks involving multiple modalities, like vision-language (VL) tasks, require not only adaptation to new data but also learning the relationship between different modalities. Targeting at VL PEFT tasks, we propose a family of operations, called routing functions, to enhance VL alignment in the low-rank bottlenecks. These feature routing functions adopt linear operations and do not introduce new trainable parameters. In-depth analyses are conducted to study their behavior. In various VL PEFT settings, the routing functions significantly improve performance of the original PEFT methods, achieving over " + }, + { + "bbox": [ + 160, + 269, + 455, + 456 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 160, + 269, + 455, + 456 + ], + "type": "text", + "content": " improvement on VQAv2 (RoBERTaLarge + ViT-L/16) and " + }, + { + "bbox": [ + 160, + 269, + 455, + 456 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 160, + 269, + 455, + 456 + ], + "type": "text", + "content": " on COCO Captioning (GPT2-medium + ViT-L/16). Also when fine-tuning a pretrained multimodal model such as CLIP-BART, we observe smaller but consistent improvements across a range of VL PEFT tasks. Our code is available at https://github.com/tingyu215/Routing_VLPEFT." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 467, + 453, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 467, + 453, + 489 + ], + "spans": [ + { + "bbox": [ + 160, + 467, + 453, + 489 + ], + "type": "text", + "content": "Keywords: Vision and Language " + }, + { + "bbox": [ + 160, + 467, + 453, + 489 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 467, + 453, + 489 + ], + "type": "text", + "content": " Parameter-Efficient Fine-Tuning " + }, + { + "bbox": [ + 160, + 467, + 453, + 489 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 467, + 453, + 489 + ], + "type": "text", + "content": " Low-Rank Approximation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 509, + 230, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 230, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 230, + 521 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 481, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 594 + ], + "type": "text", + "content": "When fine-tuning pre-trained foundation models for downstream tasks, studies show that a relatively high accuracy can be achieved by acting on a much smaller dimension compared to the original dimension of the pre-trained model's feature space. The minimum dimension necessary to approximate the optimization problem of the downstream task is termed the intrinsic dimension [1, 21]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "The existence of such intrinsic dimension explains the success of parameter-efficient fine-tuning (PEFT) methods with low-rank bottlenecks for model fine-tuning. Two of the most widely applied techniques are Adapter [9] and LoRA [10]. We demonstrate the use of Adapter and LoRA within a Transformer block in Fig. 1. Despite being added at different positions within the Transformer block, Adapter and LoRA exhibit some key characteristics in their architectural" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "design. Specifically, they first map the hidden states " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " to a lower dimension " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " through a down-projection mapping " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "W_{down}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ". The down-projected states are then projected back to the original dimension " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " via an up-projection mapping " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "W_{up}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ". In the PEFT context, only the parameters of the colored modules in Fig. 1 are updated, while the remaining parameters are unchanged. Since the intermediate dimension " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " is much smaller than the original dimension " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ", these methods significantly reduce the number of parameters that need to be updated." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 299, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 299, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 299, + 510 + ], + "type": "text", + "content": "In uni-modal tasks, where " + }, + { + "bbox": [ + 130, + 200, + 299, + 510 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 130, + 200, + 299, + 510 + ], + "type": "text", + "content": " contains information from a single modality, " + }, + { + "bbox": [ + 130, + 200, + 299, + 510 + ], + "type": "inline_equation", + "content": "W_{down}" + }, + { + "bbox": [ + 130, + 200, + 299, + 510 + ], + "type": "text", + "content": " effectively compresses " + }, + { + "bbox": [ + 130, + 200, + 299, + 510 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 130, + 200, + 299, + 510 + ], + "type": "text", + "content": " into a feature space of lower dimension. This compression usually suffices provided that " + }, + { + "bbox": [ + 130, + 200, + 299, + 510 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 200, + 299, + 510 + ], + "type": "text", + "content": " is larger than the intrinsic dimension of the downstream task [10]. However, PEFT methods with low-rank bottlenecks, such as Adapter and LoRA, are currently not designed for down-stream tasks involving another modality, such as vision-language (VL) tasks. Even for a pre-trained multimodal foundation model with " + }, + { + "bbox": [ + 130, + 200, + 299, + 510 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 130, + 200, + 299, + 510 + ], + "type": "text", + "content": " containing information from both modalities, conceptually, directly compressing " + }, + { + "bbox": [ + 130, + 200, + 299, + 510 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 130, + 200, + 299, + 510 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 200, + 299, + 510 + ], + "type": "inline_equation", + "content": "W_{down}" + }, + { + "bbox": [ + 130, + 200, + 299, + 510 + ], + "type": "text", + "content": " poses challenges, as it is unclear how to effectively balance the two modalities and to enforce alignment between the modalities with this simple linear mapping. Therefore, in this paper we investigate whether it is possible to better route the features through the low-rank bottleneck with some extra guidance. An intuitive ap" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 512, + 482, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 482, + 618 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 482, + 618 + ], + "type": "text", + "content": "proach is standard cross-attention. However, it introduces additional parameters through query/key/value projections. Moreover, softmax-based operations not only increase time complexity but are limited to dealing with features of length " + }, + { + "bbox": [ + 130, + 512, + 482, + 618 + ], + "type": "inline_equation", + "content": ">1" + }, + { + "bbox": [ + 130, + 512, + 482, + 618 + ], + "type": "text", + "content": ". This is incompatible with the PEFT setting, where the goal is to minimize the number of trainable parameters and computational complexity, while maintaining versatility across different types of tasks (including those that use only the [CLS] features). In this work, we introduce feature routing functions in VL PEFT with low-rank bottlenecks. They rely on linear operations and do not introduce extra trainable parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 620, + 320, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 620, + 320, + 631 + ], + "spans": [ + { + "bbox": [ + 146, + 620, + 320, + 631 + ], + "type": "text", + "content": "In summary, our contributions include:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 641, + 482, + 665 + ], + "type": "text", + "content": "- We identify the problem of feature learning in the low-rank bottlenecks of standard PEFT methods considering VL downstream tasks." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 234, + 482, + 316 + ], + "blocks": [ + { + "bbox": [ + 306, + 234, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 306, + 234, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 234, + 482, + 316 + ], + "type": "image", + "image_path": "aa0083a8b1b11a6e9def3f4d3f2507a2625b734bde455f26536c5c1e98bf09e0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 318, + 482, + 412 + ], + "blocks": [ + { + "bbox": [ + 307, + 318, + 482, + 412 + ], + "lines": [ + { + "bbox": [ + 307, + 318, + 482, + 412 + ], + "spans": [ + { + "bbox": [ + 307, + 318, + 482, + 412 + ], + "type": "image", + "image_path": "17050b95a1c5c37324615644bcb41b4a7c635de76ed152ddfac9697bfd2efdf1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 429, + 482, + 495 + ], + "lines": [ + { + "bbox": [ + 304, + 429, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 482, + 495 + ], + "type": "text", + "content": "Fig.1: Illustration of LoRA and Adapter used in a Transformer block. Only the green modules are updated in PEFT. We identify the low-rank bottleneck with the orange rectangle. The Adapters are sequentially added to the Transformer block.3" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "T. Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 163 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 479, + 138 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 479, + 138 + ], + "type": "text", + "content": "- We introduce a family of parameterized operations in the low-rank bottleneck, termed routing functions, that improve the performance of VL tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 479, + 163 + ], + "type": "text", + "content": "- We conduct extensive experiments on different architectures and tasks, and provide insights on using routing functions in VL PEFT tasks." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 183, + 242, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 242, + 197 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 242, + 197 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 210, + 480, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 210, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 210, + 480, + 341 + ], + "type": "text", + "content": "Adapter-based PEFT To reduce the tuning costs of pre-trained language models (LMs), the Adapter layer was first introduced in [9]. It first down-projects the model's hidden state, then up-projects it back to its original dimension, incorporating a nonlinear function after the down-projection. Freezing the pretrained weights of the LM and tuning them with sequentially added Adapter layers achieves performance comparable to full fine-tuning. Since then, various Adapter variants have been proposed to advance the field of PEFT. The focus of developing Adapter variants has been on improving efficiency [18,28,34] and exploring alternative architectures [8,14,24]. Despite being surpassed by its variants on various benchmarks, the original sequential Adapter is still widely used in practice and offers stable results across tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 342, + 480, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 480, + 436 + ], + "type": "text", + "content": "LoRA-based PEFT LoRA [10] adopts the same \"down-projection then up-projection\" approach as the Adapter, with no nonlinear function in the low-rank bottleneck. LoRA is added to the query and value mappings in the attention modules of the pre-trained LM. Important follow-up works include dynamic rank adjustment [38, 41] and quantization adaptation [5, 40]. Moreover, VeRA [17] and NOLA [16] further minimize the number of parameters by leveraging frozen random matrices. LoRA-based PEFT methods have become standard for tuning large language models and large multimodal foundation models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 437, + 480, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 480, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 480, + 507 + ], + "type": "text", + "content": "Other PEFT methods Beyond Adapter and LoRA, prompt-based tuning methods [13, 19, 22] have shown strong performance for tuning both language and vision models. LST [36] learns a compact side network during tuning. BitFit [2] only adjusts the bias terms during fine-tuning, and achieves comparable performance as other PEFT methods. Methods [8, 29, 39, 42] also explore combining PEFT methods to achieve superior performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 665 + ], + "type": "text", + "content": "Vision-Language PEFT VL-Adapter [37] is the pioneering work in the field of Vision-Language (VL) PEFT. It empirically studies the use of different PEFT methods on four different VL tasks in a multitask learning setting. VL-Adapter uses CLIP-BART, a similar model architecture as VL-T5 [33]. CLIP-BART uses a projection layer to a connect CLIP [31] vision encoder to BART [20] or T5 [33]. CLIP-BART is tuned using a combination of prompt tuning and other PEFT methods. We conduct extensive experiments using CLIP-BART and VL-Adapter as presented in Sec. 5. As an extension to VL-Adapter, VL-PET [11] combines an Adapter with a rather complicated granularity-controlled mechanism. I-Tuning [27] explores alternative model architectures for VL PEFT tasks, and proposes a parallel Adapter with cross-attention to integrate visual information into a pre-trained LM. Since our work is the first that introduces routing functions into VL PEFT, we limit the comparison of our method to" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "type": "text", + "content": "Routing Functions For VL PEFT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 233, + 254 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 233, + 254 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 233, + 254 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 233, + 254 + ], + "type": "image", + "image_path": "883fbcf05965f4d4351d5f214883d81dbff7f4aa50fcdc28177442dfdb82da61.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "text", + "content": "Fig. 2: Illustration of our method. Left: Conventional PEFT methods with low-rank bottleneck first map the hidden states " + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "text", + "content": " from a high dimension " + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "text", + "content": " to a lower dimension " + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "text", + "content": " via down-projection map " + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "inline_equation", + "content": "W_{down}" + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "text", + "content": ". Then, it maps the hidden states back to the original dimension " + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "inline_equation", + "content": "W_{up}" + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "text", + "content": ". Middle: Our method utilizes the same architecture, but updates the features via a routing function in the low-rank bottleneck. Specifically in a VL task, for the features " + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "inline_equation", + "content": "x_{R}" + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "text", + "content": " that we want to align to, we use " + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "inline_equation", + "content": "W_{down}" + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "text", + "content": " to down-project them as " + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "inline_equation", + "content": "W_{down}x_{R}" + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "text", + "content": ". Then the routing function routes " + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "inline_equation", + "content": "W_{down}x_{R}" + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "inline_equation", + "content": "W_{down}x_{H}" + }, + { + "bbox": [ + 130, + 263, + 482, + 361 + ], + "type": "text", + "content": " in the low-rank bottleneck. Right: Different routing functions. In Adapter, routing functions are added before nonlinear activation functions." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 242, + 114, + 366, + 254 + ], + "blocks": [ + { + "bbox": [ + 242, + 114, + 366, + 254 + ], + "lines": [ + { + "bbox": [ + 242, + 114, + 366, + 254 + ], + "spans": [ + { + "bbox": [ + 242, + 114, + 366, + 254 + ], + "type": "image", + "image_path": "dba17a9499953a423af3f9a7fe68656dcec9b02526d8b98119082da6729b5c66.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 370, + 114, + 482, + 254 + ], + "blocks": [ + { + "bbox": [ + 370, + 114, + 482, + 254 + ], + "lines": [ + { + "bbox": [ + 370, + 114, + 482, + 254 + ], + "spans": [ + { + "bbox": [ + 370, + 114, + 482, + 254 + ], + "type": "image", + "image_path": "e8600ef2c84b203c49cba6c89753629603e094503b30a84aa31dfb12caa91e9a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 385, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 410 + ], + "type": "text", + "content": "VL-Adapter and I-Tuning, leaving the exploration of more complicated PEFT methods like VL-PET to future work." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 428, + 233, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 428, + 233, + 442 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 233, + 442 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "text", + "content": "As illustrated in Fig. 2 (Left), standard PEFT modules with low-rank bottleneck, i.e. LoRA/Adapter, learn a down-projection mapping " + }, + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "inline_equation", + "content": "W_{down}" + }, + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "text", + "content": " followed by an up-projection mapping " + }, + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "inline_equation", + "content": "W_{up}" + }, + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "text", + "content": ", with no routing in between. These modules are added in each Transformer block of the pre-trained foundation model (Fig. 1). As presented in Fig. 2 (Middle), our method adds a routing function in the low-rank bottleneck. This routing function is designed to align two sets of features, in particular, the extra visual features " + }, + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "inline_equation", + "content": "x_{R}" + }, + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "text", + "content": " to be integrated in the VL tasks and the hidden states " + }, + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "text", + "content": " of each Transformer block of the foundation model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 146, + 548, + 480, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 548, + 480, + 561 + ], + "spans": [ + { + "bbox": [ + 146, + 548, + 480, + 561 + ], + "type": "text", + "content": "The design of a routing function follows two 'parameter-efficient' principles:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 567, + 408, + 590 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 137, + 567, + 408, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 567, + 408, + 578 + ], + "spans": [ + { + "bbox": [ + 137, + 567, + 408, + 578 + ], + "type": "text", + "content": "- A linear operation is used to reduce computation overload." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 579, + 352, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 579, + 352, + 590 + ], + "spans": [ + { + "bbox": [ + 137, + 579, + 352, + 590 + ], + "type": "text", + "content": "- No extra trainable parameters are introduced." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": "In this paper, we explore four types of routing functions. Given " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "x_{R}^{4}" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": ", after the " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "W_{down}" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": " mapping, we obtain " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "x_{t} = W_{down}x_{H}" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "x_{v} = W_{down}x_{R}" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": " is of shape " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "(L_{t},r)" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": " is of shape " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "(L_{v},r)" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "L_{t} / L_{v}" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": " are the feature length (i.e. the number of tokens) and the bottleneck dimension or rank," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "T. Qu et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 480, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 480, + 665 + ], + "type": "text", + "content": "Assuming " + }, + { + "bbox": [ + 133, + 653, + 480, + 665 + ], + "type": "inline_equation", + "content": "x_{R}" + }, + { + "bbox": [ + 133, + 653, + 480, + 665 + ], + "type": "text", + "content": " to have the same dimension " + }, + { + "bbox": [ + 133, + 653, + 480, + 665 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 133, + 653, + 480, + 665 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 133, + 653, + 480, + 665 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 133, + 653, + 480, + 665 + ], + "type": "text", + "content": ". See Sec. 4 and Sec. 5 for details." + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "respectively." + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "^5" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " In practice, we always use " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "L_v = 1" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": ", i.e. we use a single token / global feature vector to represent the visual content. In case of multiple tokens, we use average pooling to reduce it to a single feature vector." + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "^6" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " The proposed routing functions operate as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 171, + 480, + 222 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 171, + 480, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 171, + 480, + 185 + ], + "spans": [ + { + "bbox": [ + 138, + 171, + 480, + 185 + ], + "type": "text", + "content": "- Element-wise multiplication: " + }, + { + "bbox": [ + 138, + 171, + 480, + 185 + ], + "type": "inline_equation", + "content": "x_{t} \\circ x_{v}'" + }, + { + "bbox": [ + 138, + 171, + 480, + 185 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 138, + 171, + 480, + 185 + ], + "type": "inline_equation", + "content": "x_{v}'" + }, + { + "bbox": [ + 138, + 171, + 480, + 185 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 138, + 171, + 480, + 185 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 138, + 171, + 480, + 185 + ], + "type": "text", + "content": ". reshape_as(" + }, + { + "bbox": [ + 138, + 171, + 480, + 185 + ], + "type": "inline_equation", + "content": "L_{t}, r" + }, + { + "bbox": [ + 138, + 171, + 480, + 185 + ], + "type": "text", + "content": ")7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 184, + 463, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 463, + 197 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 463, + 197 + ], + "type": "text", + "content": "- Element-wise addition: " + }, + { + "bbox": [ + 138, + 184, + 463, + 197 + ], + "type": "inline_equation", + "content": "x_{t} + x_{v}^{\\prime}" + }, + { + "bbox": [ + 138, + 184, + 463, + 197 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 138, + 184, + 463, + 197 + ], + "type": "inline_equation", + "content": "x_{v}^{\\prime}" + }, + { + "bbox": [ + 138, + 184, + 463, + 197 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 138, + 184, + 463, + 197 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 138, + 184, + 463, + 197 + ], + "type": "text", + "content": ".reshape_as(L_t, r)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 196, + 304, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 196, + 304, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 196, + 304, + 209 + ], + "type": "text", + "content": "- Matrix multiplication 1: " + }, + { + "bbox": [ + 138, + 196, + 304, + 209 + ], + "type": "inline_equation", + "content": "x_{t}(x_{v})^{\\top}x_{v}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 209, + 451, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 451, + 222 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 451, + 222 + ], + "type": "text", + "content": "- Matrix multiplication 2: " + }, + { + "bbox": [ + 138, + 209, + 451, + 222 + ], + "type": "inline_equation", + "content": "x_{t}x_{v}^{\\prime \\prime}" + }, + { + "bbox": [ + 138, + 209, + 451, + 222 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 138, + 209, + 451, + 222 + ], + "type": "inline_equation", + "content": "x_{v}^{\\prime \\prime}" + }, + { + "bbox": [ + 138, + 209, + 451, + 222 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 138, + 209, + 451, + 222 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 138, + 209, + 451, + 222 + ], + "type": "text", + "content": ".reshape_as(r,r)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": "Intuitively, " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{t} \\circ x_{v}^{\\prime}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": ", reinforces or attenuates individual elements of " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": ", using the corresponding element of " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": " as scaling factor. The second routing function, " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{t} + x_{v}^{\\prime}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": " replaces " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": " with the sum of " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{v}^{\\prime}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": ", thus moving " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": " closer to " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{t}(x_{v})^{\\top}x_{v}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": " projects " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": " onto " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": ", i.e. only keeping the component of " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": " that is aligned with " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{t}x_{v}^{\\prime\\prime}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": " replaces " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": " with a rescaled version of " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": ", with a different scale factor for each row given by the sum over all elements of the corresponding row of " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": ". In all these cases, it should be kept in mind that both " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": " are derived from " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{R}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": " via the down-projection " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "W_{down}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": ". Parameters of " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "W_{down}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": " can be learned so as to select particular dimensions of the original " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "x_{R}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": ", on which to apply the above operations before up-projecting them again with " + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "inline_equation", + "content": "W_{up}" + }, + { + "bbox": [ + 130, + 228, + 480, + 361 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 361, + 480, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 480, + 419 + ], + "type": "text", + "content": "The features " + }, + { + "bbox": [ + 130, + 361, + 480, + 419 + ], + "type": "inline_equation", + "content": "x_{R}" + }, + { + "bbox": [ + 130, + 361, + 480, + 419 + ], + "type": "text", + "content": " contain information from the modalities to which we aim to align " + }, + { + "bbox": [ + 130, + 361, + 480, + 419 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 130, + 361, + 480, + 419 + ], + "type": "text", + "content": ". The choice of " + }, + { + "bbox": [ + 130, + 361, + 480, + 419 + ], + "type": "inline_equation", + "content": "x_{R}" + }, + { + "bbox": [ + 130, + 361, + 480, + 419 + ], + "type": "text", + "content": " can be different for different architectures, for instance, [CLS] features from a ViT in case of an encoder-only or decoder-only LM. We further explain the motivation of the choices of " + }, + { + "bbox": [ + 130, + 361, + 480, + 419 + ], + "type": "inline_equation", + "content": "x_{R}" + }, + { + "bbox": [ + 130, + 361, + 480, + 419 + ], + "type": "text", + "content": " for different types of architectures in Sec. 4.1 and Sec. 5.1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 439, + 479, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 439, + 479, + 453 + ], + "spans": [ + { + "bbox": [ + 131, + 439, + 479, + 453 + ], + "type": "text", + "content": "4 Experiments with En(De)coder-Only Language Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 465, + 479, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 465, + 479, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 465, + 479, + 536 + ], + "type": "text", + "content": "To explore the value of routing functions, in this section, we start from encoder-only and decoder-only language models (LMs) and incorporate visual information from pre-trained vision encoders for vision-language (VL) parameter-efficient fine-tuning (PEFT) tasks. The section begins with an overview of our experimental design and implementation details, and concludes with a discussion on the comprehensive experimental results obtained." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 555, + 266, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 555, + 266, + 567 + ], + "spans": [ + { + "bbox": [ + 131, + 555, + 266, + 567 + ], + "type": "text", + "content": "4.1 Experimental Design" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 575, + 479, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 575, + 479, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 575, + 479, + 624 + ], + "type": "text", + "content": "Here, we study VL PEFT tasks in their cleanest form. Specifically, (1) No additional mapping networks are adopted, so the feature routing functions form the only module that contributes to the final performance; (2) Only the parameters of the PFET modules in LMs are updated, where we learn the VL alignment; (3)" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 310, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 310, + 91, + 448, + 102 + ], + "type": "text", + "content": "Routing Functions For VL PEFT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 631, + 341, + 643 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 631, + 341, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 631, + 341, + 643 + ], + "type": "text", + "content": "Here we omit the batch dimension for simplicity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 135, + 643, + 479, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 643, + 479, + 653 + ], + "spans": [ + { + "bbox": [ + 135, + 643, + 479, + 653 + ], + "type": "text", + "content": "6 Only when comparing to cross-attention in Sec. 4.3, we do not use pooled features." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 136, + 653, + 480, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 653, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 136, + 653, + 480, + 666 + ], + "type": "text", + "content": "e.g. " + }, + { + "bbox": [ + 136, + 653, + 480, + 666 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 136, + 653, + 480, + 666 + ], + "type": "text", + "content": ". reshape_as(" + }, + { + "bbox": [ + 136, + 653, + 480, + 666 + ], + "type": "inline_equation", + "content": "L_{t}, r" + }, + { + "bbox": [ + 136, + 653, + 480, + 666 + ], + "type": "text", + "content": "): " + }, + { + "bbox": [ + 136, + 653, + 480, + 666 + ], + "type": "inline_equation", + "content": "x_{v}' = \\mathbb{1}^{\\mathsf{T}}x_{v}" + }, + { + "bbox": [ + 136, + 653, + 480, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 136, + 653, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{1}" + }, + { + "bbox": [ + 136, + 653, + 480, + 666 + ], + "type": "text", + "content": " is a matrix of ones of shape " + }, + { + "bbox": [ + 136, + 653, + 480, + 666 + ], + "type": "inline_equation", + "content": "(L_{v}, L_{t})" + }, + { + "bbox": [ + 136, + 653, + 480, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 113, + 246, + 232 + ], + "blocks": [ + { + "bbox": [ + 134, + 113, + 246, + 232 + ], + "lines": [ + { + "bbox": [ + 134, + 113, + 246, + 232 + ], + "spans": [ + { + "bbox": [ + 134, + 113, + 246, + 232 + ], + "type": "image", + "image_path": "f16742f4e8d68edc424b1d0026d006f4863c7e9c61553f2d167ef69bc0189826.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 241, + 482, + 276 + ], + "lines": [ + { + "bbox": [ + 130, + 241, + 482, + 276 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 482, + 276 + ], + "type": "text", + "content": "Fig.3: Pipeline for VQA and Image Captioning. Only the added classifier (in VQA) and the PEFT modules of the RoBERTa encoder and GPT decoder are tuned. Note that here " + }, + { + "bbox": [ + 130, + 241, + 482, + 276 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 130, + 241, + 482, + 276 + ], + "type": "text", + "content": " are the text inputs with the visual [CLS] features prepended." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 253, + 113, + 382, + 232 + ], + "blocks": [ + { + "bbox": [ + 253, + 113, + 382, + 232 + ], + "lines": [ + { + "bbox": [ + 253, + 113, + 382, + 232 + ], + "spans": [ + { + "bbox": [ + 253, + 113, + 382, + 232 + ], + "type": "image", + "image_path": "139a592d9cee0ec8b3811ec9d38446f4adf950a5aabdf741c852d3e6e00147a9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 389, + 113, + 481, + 232 + ], + "blocks": [ + { + "bbox": [ + 389, + 113, + 481, + 232 + ], + "lines": [ + { + "bbox": [ + 389, + 113, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 389, + 113, + 481, + 232 + ], + "type": "image", + "image_path": "af2c687005cf2f604c5af07550a389bb4546c179772493a455da5d749d226dc7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 298, + 479, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 479, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 479, + 346 + ], + "type": "text", + "content": "No multimodal tasks (e.g. contrastive language-image pre-training in CLIP [31]) are included in the pre-training phases of the backbone models. We only learn the multimodal embedding space during VL PEFT. These principles guarantee that VL alignment originates strictly from the PEFT modules." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 346, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 480, + 407 + ], + "type": "text", + "content": "To comprehensively study the role of routing functions in VL tasks, we consider both discriminative and generative VL tasks. We select visual question answering (VQA) and Image Captioning, two of the most widely studied VL tasks, as the representative task for discriminative VL tasks and generative VL tasks, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 407, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 525 + ], + "type": "text", + "content": "We present the pipelines for VQA and Image Captioning in Fig. 3. We use ViT [6] as our vision backbone. As for the LM backbone, we use the encoder-only RoBERTa [25] and the decoder-only GPT2 [32] for VQA and Image Captioning, respectively. The [CLS] feature from the ViT is prepended to the text inputs and acts as " + }, + { + "bbox": [ + 130, + 407, + 482, + 525 + ], + "type": "inline_equation", + "content": "x_{R}" + }, + { + "bbox": [ + 130, + 407, + 482, + 525 + ], + "type": "text", + "content": ", enabling the learning of VL alignment through PEFT, with or without the integration of our routing functions. Prepending visual features to text inputs is actually the predominant way of learning VL alignment in PEFT [11, 37] without routing functions. The ViT and LMs adopt the same dimensions for their feature spaces, so we do not need additional mappings to match the dimensions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 544, + 279, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 279, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 279, + 555 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "content": "We adopt ViT-B/16 pre-trained on ImageNet-21k as our vision backbone, which we keep frozen during training. In LoRA, we set scaling factor " + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\alpha = r" + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "content": " in all setups. For VQAv2, we use RoBERTa" + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "inline_equation", + "content": "_{\\text{base}}" + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "content": " as our LM backbone, with a two-layer multi-layer perceptron (MLP) classifier on top following [37]. We train our model using the AdamW optimizer [26] with learning rate=3e-4 and batch size of 512 for 20 epochs, where we warm up the first " + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "content": " of the steps. For the captioning task, COCO Cap., we use GPT2-base as our LM backbone. We train our model for 20" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "T. Qu et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 480, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 480, + 666 + ], + "type": "text", + "content": "8 In Appendix B, we report results obtained with larger backbones, e.g. RoBERTalarge." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "epochs using the AdamW optimizer with batch size of 40 and learning rate " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "= 3\\mathrm{e} - 5" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " where we warm up the first 5000 steps. For more details, see Appendix A." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 156, + 198, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 156, + 198, + 167 + ], + "spans": [ + { + "bbox": [ + 132, + 156, + 198, + 167 + ], + "type": "text", + "content": "4.3 Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 174, + 479, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 174, + 479, + 198 + ], + "spans": [ + { + "bbox": [ + 130, + 174, + 479, + 198 + ], + "type": "text", + "content": "Overall performance on VQA We first evaluate our routing functions on the discriminative task VQAv2 with RoBERTabase and ViT-B/16, see Tab. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 198, + 298, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 198, + 298, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 298, + 353 + ], + "type": "text", + "content": "Without the addition of routing functions, models with LoRA and Adapter reach a similar level of performance. All four types of routing functions significantly improve the performance of the original PEFT method in terms of accuracy of the answer. The dimension of the bottleneck does not have a strong impact on the final performance. Sometimes we can even obtain better performance with lower dimensions. It is a sign that the intrinsic dimension of the VL task is not" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 289, + 479, + 345 + ], + "blocks": [ + { + "bbox": [ + 304, + 233, + 481, + 288 + ], + "lines": [ + { + "bbox": [ + 304, + 233, + 481, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 481, + 288 + ], + "type": "text", + "content": "Table 1: Accuracy and improvement over the baseline (in blue) obtained by different types of routing functions on VQAv2. The best results and second best results are in bold and underlined, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 289, + 479, + 345 + ], + "lines": [ + { + "bbox": [ + 307, + 289, + 479, + 345 + ], + "spans": [ + { + "bbox": [ + 307, + 289, + 479, + 345 + ], + "type": "table", + "html": "
Routingr = 64r = 128
FunctionsLoRAAdapterLoRAAdapter
None44.1544.1644.4544.28
xt○xt'53.51+9.3652.78+8.6253.86+10.4153.01+8.73
xt+xt'52.60+8.4553.94+9.7852.88+8.4353.95+9.67
xt(xt)Txt53.88+9.7354.48+10.3253.09+8.6455.06+10.78
xtxt''54.21+10.0654.96+10.8051.88+7.4354.38+10.10
", + "image_path": "a0c9b21fb5aeb52d7b93f889656a8b2184fd5122a4c14b7caf22d7e1c1bc0b5c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "type": "text", + "content": "high, just as observed in uni-modal tasks. Moreover, we observe that the routing functions involving matrix multiplications (i.e. " + }, + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "type": "inline_equation", + "content": "x_{t}(x_{v})^{\\top}x_{v} \\& x_{t}x_{v}^{\\prime \\prime}" + }, + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "type": "text", + "content": ") perform quite well in most scenarios for both LoRA and Adapter, while the routing functions involving element-wise operations (i.e. " + }, + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "type": "inline_equation", + "content": "x_{t} \\circ x_{v}^{\\prime} \\& x_{t} + x_{v}^{\\prime}" + }, + { + "bbox": [ + 130, + 354, + 480, + 403 + ], + "type": "text", + "content": ") are just slightly behind." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 134, + 476, + 293, + 523 + ], + "blocks": [ + { + "bbox": [ + 130, + 422, + 482, + 467 + ], + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 467 + ], + "type": "text", + "content": "Table 2: Results in terms of accu- Table 3: Results in terms of accuracy with racy with different routing functions on different routing functions on VQAv2 sorted VQAv2 sorted by answer type. The \"all\" by question type. The \"all\" column reports column reports overall accuracy. overall accuracy." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 476, + 293, + 523 + ], + "lines": [ + { + "bbox": [ + 134, + 476, + 293, + 523 + ], + "spans": [ + { + "bbox": [ + 134, + 476, + 293, + 523 + ], + "type": "table", + "html": "
Routing FunctionsallLoRA, r = 128Adapter, r = 128
yes/no numberotherallyes/no numberother
None44.4568.6731.2929.2044.2868.3831.19
xt·ox'53.8672.2935.3644.4853.0171.7233.34
xt+xt'52.8872.0334.2043.0053.9572.4335.38
xt(xt)Tx53.0970.6335.4044.1955.0674.0933.59
xtx''51.8869.9233.2942.8254.3873.0033.94
", + "image_path": "67bcb20023d60fedbc0063e5f6309306ab792321adeda9e298c7e1b203b6af90.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 297, + 476, + 479, + 523 + ], + "blocks": [ + { + "bbox": [ + 297, + 476, + 479, + 523 + ], + "lines": [ + { + "bbox": [ + 297, + 476, + 479, + 523 + ], + "spans": [ + { + "bbox": [ + 297, + 476, + 479, + 523 + ], + "type": "table", + "html": "
Routing \nFunctionsLoRA, r = 128Adapter, r = 128
allwhere colortimepersonallwhere colortimeperson
None44.4522.6440.0014.2945.1744.2822.9340.1513.7044.94
xt × xv'53.8635.7051.6923.0762.7453.0135.5952.6320.7362.28
xt + xv'52.8836.1249.3321.2262.9553.9537.0852.0821.6663.76
xt(x)Tx53.0936.3352.8622.2961.7455.0636.9253.6624.4463.85
xtxv''51.8834.1951.6022.6361.1454.3835.6452.7721.8562.50
", + "image_path": "0182d84ccf813c27181fee24a511aa132aa464e1f4dee6b7b371d06ecba3af94.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "To dive deeper into the behaviors of different types of routing functions, we evaluate the performance on different answer types as presented in Tab. 2. The largest improvements are obtained for the \"other\" category for all four routing functions. It can be seen that different routing functions excel for different answer types. For example, even when the two element-wise operations reach lower overall accuracy, they lead to higher or comparable accuracy on number-type answers. As for the routing functions involving matrix multiplications, better performance on other-type answers is obtained. For instance, with relatively lower overall accuracy reached in LoRA, matrix multiplications reach comparable performance on other-type answers. No clear pattern can be observed for yes/no" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 310, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 310, + 90, + 448, + 102 + ], + "type": "text", + "content": "Routing Functions For VL PEFT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 156, + 480, + 257 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 4: Results of different types of routing functions on COCO Cap. The Avg. score is the averaged result from BLEU-4, METEOR, ROUGE-L, CIDEr, and SPICE. The best and second best results are in bold and underlined, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 156, + 480, + 257 + ], + "lines": [ + { + "bbox": [ + 133, + 156, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 133, + 156, + 480, + 257 + ], + "type": "table", + "html": "
Routing FunctionsLoRA, r=64LoRA, r=128
BLEU-4METEORROUGE-LCIDErSPICEAvg.BLEU-4METEORROUGE-LCIDErSPICEAvg.
None18.317.436.855.811.828.0221.020.943.470.314.734.06
xt×ov'26.1+7.823.7+6.348.0+11.888.7+32.917.3+5.540.88+12.8625.4+4.423.7+2.848.2+4.887.9+17.617.7+3.040.58+6.52
xt+xov'22.2+3.920.3+2.941.0+4.273.5+17.714.8+3.034.36+6.3425.4+4.423.1+2.247.1+4.787.1+16.817.0+2.339.94+5.88
xt(xv)Txv24.8+6.522.6+5.245.4+8.684.9+29.116.8+5.038.90+10.8824.9+3.922.7+1.845.4+2.083.8+13.516.9+2.238.74+4.68
xtx''23.9+5.621.9+4.543.9+7.180.5+14.716.2+4.437.28+9.2625.5+4.522.9+2.046.0+2.685.2+14.917.2+2.539.36+5.30
Routing FunctionsAdapter, r=64Adapter, r=128
BLEU-4METEORROUGE-LCIDErSPICEAvg.BLEU-4METEORROUGE-LCIDErSPICEAvg.
None15.918.537.061.614.129.4216.118.436.261.513.929.22
xt×ov'24.6+8.723.1+4.646.4+9.484.5+22.917.2+3.139.16+9.7423.6+7.522.6+4.244.8+8.682.8+21.317.1+3.238.18+8.96
xt+xov'21.0+5.121.5+3.042.4+5.475.0+14.416.0+1.935.18+5.7619.3+3.220.8+2.440.9+4.775.2+13.816.2+2.334.48+5.26
xt(xv)Txv26.1+10.223.2+4.746.9+9.985.4+23.817.3+3.239.78+10.3624.1+8.022.4+4.044.2+8.082.7+21.216.9+3.038.06+8.84
xtx''26.6+10.723.0+4.546.8+9.885.8+24.217.2+3.139.88+10.4626.0+9.923.5+5.146.8+10.686.3+24.817.8+3.940.08+10.86
", + "image_path": "890546d45b1d11b6ac407a02b92e5a6badc067e390e06b227a9f4c5a37dca6fd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 281, + 481, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 481, + 304 + ], + "type": "text", + "content": "types of answers, which could be explained by the large variety in content of these questions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 306, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 486 + ], + "type": "text", + "content": "Tab. 3 shows our method's performance across various question types. Four distinct question types test the method's reasoning capabilities concerning fine-grained visual information: spatial information (\"where\", e.g. Where is the surfboard?), color information (\"color\", e.g. What color is the bike?), time information (\"time\", e.g. What time is on the clock?), and person information (\"person\", e.g. What is the man holding?).9 Adding routing functions significantly improves the performance of LoRA and Adapter. Notably, different routing functions seem to capture different aspects of the visual information. " + }, + { + "bbox": [ + 130, + 306, + 482, + 486 + ], + "type": "inline_equation", + "content": "x_{t} + x_{v}^{\\prime}" + }, + { + "bbox": [ + 130, + 306, + 482, + 486 + ], + "type": "text", + "content": " shows great performance while reasoning on spatial information (where) and person information (person). It is assumed that the addition operator reinforces feature values of content both present in the question and image, which might be beneficial for spatial and person questions. Similarly, " + }, + { + "bbox": [ + 130, + 306, + 482, + 486 + ], + "type": "inline_equation", + "content": "x_{t}x_{v}^{\\prime\\prime}" + }, + { + "bbox": [ + 130, + 306, + 482, + 486 + ], + "type": "text", + "content": " performs exceptionally well for time information (time). With " + }, + { + "bbox": [ + 130, + 306, + 482, + 486 + ], + "type": "inline_equation", + "content": "x_{t} \\circ x_{v}^{\\prime}" + }, + { + "bbox": [ + 130, + 306, + 482, + 486 + ], + "type": "text", + "content": ", we obtain relatively better performance on the color information (color). " + }, + { + "bbox": [ + 130, + 306, + 482, + 486 + ], + "type": "inline_equation", + "content": "x_{t}(x_{v})^{\\top}x_{v}" + }, + { + "bbox": [ + 130, + 306, + 482, + 486 + ], + "type": "text", + "content": " presents consistent strong abilities while aligning spatial information (where) and color information (color)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 487, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 534 + ], + "type": "text", + "content": "In Appendix B, we show an additional ablation study on replacing visual features " + }, + { + "bbox": [ + 130, + 487, + 482, + 534 + ], + "type": "inline_equation", + "content": "(x_{R})" + }, + { + "bbox": [ + 130, + 487, + 482, + 534 + ], + "type": "text", + "content": " with random noise or all-ones features. We also compare inference time of LoRA/Adapter with LoRA/Adapter + routing functions, where we observe adding routing functions to PEFT methods does not reduce efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 536, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 643 + ], + "type": "text", + "content": "Overall performance on image captioning Evaluation of image captioning relies on the overall assessment of the generated captions. We report multiple text generation metrics and their average scores in Tab. 4. The addition of routing functions again leads to significant performance gains for all metrics in all scenarios. As was the case with VQAv2, different patterns are observed. Changing the intermediate dimension " + }, + { + "bbox": [ + 130, + 536, + 482, + 643 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 536, + 482, + 643 + ], + "type": "text", + "content": " from 128 to 64 seems to have a larger impact on the model performance with conventional LoRA, where without the routing functions the CIDEr score drops from 70.3 to 55.8. In contrast, models augmented with routing functions remain robust and per" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "T. Qu et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 420, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 420, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 420, + 665 + ], + "type": "text", + "content": "9 The comprehensive list of question types is included in Appendix A." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "formant across setups, despite the changes in intermediate dimension. " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "x_{t} \\circ x_{v}^{\\prime}" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "x_{t}x_{v}^{\\prime \\prime}" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " perform well, possibly due to their alignment power. We also report the performance of Adapter with larger backbones, namely GPT2-medium and ViT-L/16, in Appendix B. Consistent and significant improvements in performance are observed, showing the efficacy of our routing functions on the task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 189, + 297, + 200 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 189, + 297, + 200 + ], + "spans": [ + { + "bbox": [ + 131, + 189, + 297, + 200 + ], + "type": "text", + "content": "Comparison to cross-attention" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 201, + 298, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 298, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 298, + 307 + ], + "type": "text", + "content": "Here we compare the results of the integration of routing function " + }, + { + "bbox": [ + 130, + 201, + 298, + 307 + ], + "type": "inline_equation", + "content": "x_{t}(x_{v})^{\\top}x_{v}" + }, + { + "bbox": [ + 130, + 201, + 298, + 307 + ], + "type": "text", + "content": " with these of integrating cross-attention between the textual and the visual representation in image captioning. Indeed, in principle, a cross-attention function can also be considered as a kind of routing function with nonlinear operations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 308, + 306, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 306, + 345 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 306, + 345 + ], + "type": "text", + "content": "I-Tuning [27] implements an Adapter with cross-attention for image captioning. To have a fair comparison, we" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 345, + 481, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 481, + 403 + ], + "type": "text", + "content": "follow the settings of I-Tuning for both LoRA and Adapter. Specifically, (1) instead of using the visual features, we pretend the \"\" token to the input text. (2) We use the last hidden state from ViT as key and value for cross-attention and as " + }, + { + "bbox": [ + 130, + 345, + 481, + 403 + ], + "type": "inline_equation", + "content": "x_{R}" + }, + { + "bbox": [ + 130, + 345, + 481, + 403 + ], + "type": "text", + "content": " for routing functions. This change is needed, since the [CLS] feature is of length 1, which does not work for cross-attention due to the Softmax function." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "content": "In the case of LoRA, we also implement a variant with separate " + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "inline_equation", + "content": "W_{down}" + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "content": " mappings for " + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "inline_equation", + "content": "x_{R}" + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "content": ". This brings LoRA with routing function closer to LoRA with cross-attention. Without separate mapping refers to the setting used in previous experiments, where " + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "inline_equation", + "content": "x_{R}" + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "content": " share the same " + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "inline_equation", + "content": "W_{down}" + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "content": ". As shown in Tab. 5, despite of having less parameters, " + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\mathrm{LoRA} + x_t(x_v)^{\\top}x_v" + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "content": " (w/ separate mapping) achieves significantly better performance than LoRA with cross-attention. The setup of " + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "inline_equation", + "content": "\\mathrm{LoRA} + x_t(x_v)^{\\top}x_v" + }, + { + "bbox": [ + 130, + 405, + 482, + 501 + ], + "type": "text", + "content": " (w/o separate mapping) outperforms LoRA with cross-attention by a large margin using a similar amount of parameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 502, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 561 + ], + "type": "text", + "content": "For Adapter, the cross-attention module has the same intermediate dimension as the " + }, + { + "bbox": [ + 130, + 502, + 482, + 561 + ], + "type": "inline_equation", + "content": "W_{down}" + }, + { + "bbox": [ + 130, + 502, + 482, + 561 + ], + "type": "text", + "content": " in Adapter with routing function. Naturally, the separate query/ key/value mappings make that the Adapter with cross-attention has more than double the number of parameters compared to the Adapter equipped with the routing function " + }, + { + "bbox": [ + 130, + 502, + 482, + 561 + ], + "type": "inline_equation", + "content": "(x_{t}(x_{v})^{\\top}x_{v})" + }, + { + "bbox": [ + 130, + 502, + 482, + 561 + ], + "type": "text", + "content": ". Both achieve comparable performance (Tab. 5)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 562, + 482, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 562, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 562, + 482, + 622 + ], + "type": "text", + "content": "To conclude, Tab. 5 shows that integrating " + }, + { + "bbox": [ + 130, + 562, + 482, + 622 + ], + "type": "inline_equation", + "content": "x_{t}(x_{v})^{\\top}x_{v}" + }, + { + "bbox": [ + 130, + 562, + 482, + 622 + ], + "type": "text", + "content": " in LORA and Adapter yields better performance given the same amount of parameters, or is similar in performance while using substantially less parameters. For VL tasks like image captioning, integrating a routing function to guide feature representations in low-rank bottlenecks has superior abilities compared to cross-attention." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 211, + 482, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 211, + 482, + 298 + ], + "spans": [ + { + "bbox": [ + 304, + 211, + 482, + 298 + ], + "type": "text", + "content": "Table 5: Comparing routing functions with cross-attention on COCO Cap. †: parallel Adapter as used in [27]. Separate Map.: separate mapping for query/ key/value (in cross-attn.) or features " + }, + { + "bbox": [ + 304, + 211, + 482, + 298 + ], + "type": "inline_equation", + "content": "x_{R} / x_{H}" + }, + { + "bbox": [ + 304, + 211, + 482, + 298 + ], + "type": "text", + "content": " (in " + }, + { + "bbox": [ + 304, + 211, + 482, + 298 + ], + "type": "inline_equation", + "content": "x_{t}(x_{v})^{\\top}x_{v}" + }, + { + "bbox": [ + 304, + 211, + 482, + 298 + ], + "type": "text", + "content": "). Here " + }, + { + "bbox": [ + 304, + 211, + 482, + 298 + ], + "type": "inline_equation", + "content": "r = 128" + }, + { + "bbox": [ + 304, + 211, + 482, + 298 + ], + "type": "text", + "content": ". Avg. indicates the same average scores as in Tab. 4. Full evaluation in Appendix B." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 307, + 299, + 480, + 340 + ], + "blocks": [ + { + "bbox": [ + 307, + 299, + 480, + 340 + ], + "lines": [ + { + "bbox": [ + 307, + 299, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 480, + 340 + ], + "type": "table", + "html": "
PEFTSeparateMap.AlignmentParam.BLEU-4CIDErAvg.
LoRA✓Cross-attn.4.786M28.792.243.02
LoRA✓xt(xt)Txtxt3.932M30.7+2.099.4+7.245.68+2.66
LoRAXxt(xt)Txtxt4.746M30.0+1.399.0+6.845.38+2.36
Adapter†✓Cross-attn.4.732M30.799.845.70
Adapter†Xxt(xt)Txtxt1.830M30.8+0.198.8-1.045.48-0.22
", + "image_path": "4580eae2750b78f10347aeb2f6042918f34239ab5ac180d3d19ad3360cf307a1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 310, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 310, + 91, + 448, + 102 + ], + "type": "text", + "content": "Routing Functions For VL PEFT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 131, + 631, + 482, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 631, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 631, + 482, + 666 + ], + "type": "text", + "content": "10 Appendix A shows detailed architectures and explains why routing functions outperform cross-attention. We down-project to " + }, + { + "bbox": [ + 131, + 631, + 482, + 666 + ], + "type": "inline_equation", + "content": "r/4" + }, + { + "bbox": [ + 131, + 631, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 631, + 482, + 666 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 131, + 631, + 482, + 666 + ], + "type": "text", + "content": " for LoRA w/ & w/o separate mapping, respectively, to have the same level of parameter count." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 298, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 298, + 235 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 298, + 235 + ], + "type": "text", + "content": "Incorporating nonlinear activation functions into routing functions We have proposed linear routing functions for efficiency reasons, but they do not need to be exclusively linear. Taking inspirations from the linear attention variants [4,15,30], where the linear operation is often equipped with nonlinear activation functions to ensure non-negativity for" + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 307, + 161, + 480, + 209 + ], + "blocks": [ + { + "bbox": [ + 304, + 129, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 304, + 129, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 304, + 129, + 482, + 160 + ], + "type": "text", + "content": "Table 6: Experiments on adding nonlinear activation functions in routing functions. We use " + }, + { + "bbox": [ + 304, + 129, + 482, + 160 + ], + "type": "inline_equation", + "content": "r = 64" + }, + { + "bbox": [ + 304, + 129, + 482, + 160 + ], + "type": "text", + "content": " here." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 161, + 480, + 209 + ], + "lines": [ + { + "bbox": [ + 307, + 161, + 480, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 161, + 480, + 209 + ], + "type": "table", + "html": "
PEFT MethodRouting FunctionsVQA AccuracyCOCO Cap.
BLEU-4CIDErAvg.
LoRAxt(xv)Txv53.8824.884.938.90
LoRAReLU(xt)ReLU(xv)Txv53.4525.386.139.84
Adapterxt(xv)Txv54.4826.185.439.78
AdapterReLU(xt)ReLU(xv)Txv54.6325.084.739.02
", + "image_path": "6f8e15cacab643450ff81922e76895cbfc4e0e7232aa13815a87618dac799148.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 236, + 479, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 479, + 308 + ], + "type": "text", + "content": "similarity scores between query and key, we explore whether a similar operation affects the representation power of the routing functions in the low-rank bottleneck. For routing function " + }, + { + "bbox": [ + 130, + 236, + 479, + 308 + ], + "type": "inline_equation", + "content": "x_{t}(x_{v})^{\\top}x_{v}" + }, + { + "bbox": [ + 130, + 236, + 479, + 308 + ], + "type": "text", + "content": ", we use the nonlinear activation function " + }, + { + "bbox": [ + 130, + 236, + 479, + 308 + ], + "type": "inline_equation", + "content": "ReLU(\\cdot)" + }, + { + "bbox": [ + 130, + 236, + 479, + 308 + ], + "type": "text", + "content": " as in previous work [30], and transform it to " + }, + { + "bbox": [ + 130, + 236, + 479, + 308 + ], + "type": "inline_equation", + "content": "ReLU(x_{t})ReLU(x_{v})^{\\top}x_{v}" + }, + { + "bbox": [ + 130, + 236, + 479, + 308 + ], + "type": "text", + "content": ". The results are presented in Tab. 6. With or without " + }, + { + "bbox": [ + 130, + 236, + 479, + 308 + ], + "type": "inline_equation", + "content": "ReLU(\\cdot)" + }, + { + "bbox": [ + 130, + 236, + 479, + 308 + ], + "type": "text", + "content": ", our routing function achieves similar performance on VQA and image captioning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 311, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 407 + ], + "type": "text", + "content": "Qualitative Analysis We first qualitatively compare the VL alignment pattern of LoRA-tuned GPT2 with or without a routing function on image captioning. An example is presented in Fig. 4. With the addition of routing function " + }, + { + "bbox": [ + 130, + 311, + 482, + 407 + ], + "type": "inline_equation", + "content": "x_{t}(x_{v})^{\\top}x_{v}" + }, + { + "bbox": [ + 130, + 311, + 482, + 407 + ], + "type": "text", + "content": ", the model pays more attention to the nouns (e.g. man, a big screen) and the verbs (e.g. watching), which are essential in learning the VL alignment. At the same time, it pays less attention to meaningless words like \"something\". This is in line with the finding that the routing function stresses components of " + }, + { + "bbox": [ + 130, + 311, + 482, + 407 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 311, + 482, + 407 + ], + "type": "text", + "content": " that are aligned in the image." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 138, + 433, + 440, + 509 + ], + "blocks": [ + { + "bbox": [ + 138, + 433, + 440, + 509 + ], + "lines": [ + { + "bbox": [ + 138, + 433, + 440, + 509 + ], + "spans": [ + { + "bbox": [ + 138, + 433, + 440, + 509 + ], + "type": "image", + "image_path": "197843bbade820f1df61436b07e3b53b5a7ea586049e2db61d95ece54853f6fe.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 516, + 482, + 540 + ], + "lines": [ + { + "bbox": [ + 130, + 516, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 482, + 540 + ], + "type": "text", + "content": "Fig. 4: Example of average attention weights from the last layer of GPT2. We use the final checkpoint trained on COCO Cap. with " + }, + { + "bbox": [ + 130, + 516, + 482, + 540 + ], + "type": "inline_equation", + "content": "r = 64" + }, + { + "bbox": [ + 130, + 516, + 482, + 540 + ], + "type": "text", + "content": ". IMAGE: visual [CLS] feature." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Next we qualitatively examine some examples from the inference phase for VQA and Image Captioning. We present examples in Fig. 5 where conventional LoRA fails. For VQA, previous experiments in Tab. 3 indicate that our routing functions work particularly well on color related questions. As also shown in Fig. 5, conventional LoRA fails to align the color red to the lollipop. The model tuned with conventional LoRA also fails to correctly answer other types of questions as presented in the figure. Without routing functions, captions generated from the LM tuned with conventional LoRA often make mistakes in compre" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "T. Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 116, + 285, + 264 + ], + "blocks": [ + { + "bbox": [ + 153, + 116, + 285, + 264 + ], + "lines": [ + { + "bbox": [ + 153, + 116, + 285, + 264 + ], + "spans": [ + { + "bbox": [ + 153, + 116, + 285, + 264 + ], + "type": "image", + "image_path": "845e1ec6070b8ec8a5d1018e8d89fc7f390c5db4aef07b82a89cf8399345cf3c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "type": "text", + "content": "Fig. 5: Qualitative examples from VQA and Image Captioning tasks. We present cases where the model trained with conventional LoRA fails and results obtained from a model trained with LoRA with a routing function. We use " + }, + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "type": "inline_equation", + "content": "r = 64" + }, + { + "bbox": [ + 130, + 270, + 482, + 304 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 293, + 116, + 462, + 264 + ], + "blocks": [ + { + "bbox": [ + 293, + 116, + 462, + 264 + ], + "lines": [ + { + "bbox": [ + 293, + 116, + 462, + 264 + ], + "spans": [ + { + "bbox": [ + 293, + 116, + 462, + 264 + ], + "type": "image", + "image_path": "0053a476f1e008c39268be92e8b5647fc7c36b2e00fb2160454f9738e25778ea.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 325, + 480, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 480, + 351 + ], + "type": "text", + "content": "hending the visual information. Occasionally, we observe a sequence of blank spaces as presented in the first example using conventional LoRA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 366, + 473, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 366, + 473, + 380 + ], + "spans": [ + { + "bbox": [ + 131, + 366, + 473, + 380 + ], + "type": "text", + "content": "5 Experiments with Encoder-Decoder Language Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 389, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 485 + ], + "type": "text", + "content": "In this section, we study the VL PEFT tasks in a more complicated setting, where we consider an encoder-decoder LM with an additional mapping network dedicated for the vision encoder. The model is trained in a multi-task learning setting, where four different VL tasks are jointly learned, as in [3,37]. Despite the increased complexity of optimizing the network in this scenario, we show that integrating routing functions into the low-rank bottlenecks yields performance enhancements. Moreover, consistent patterns are observed for different scenarios, which shed lights on future research to improve VL PEFT tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 501, + 266, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 501, + 266, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 501, + 266, + 514 + ], + "type": "text", + "content": "5.1 Experimental Design" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 519, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 482, + 665 + ], + "type": "text", + "content": "Following [37], we use the CLIP-BART architecture to study the effect of our routing functions on the VL PEFT tasks in a multitask learning setting. CLIP-BART is a widely used architecture for VL tasks. We present the simplified model architecture of CLIP-BART with Adapters and the integrated routing functions in Fig. 6. The original CLIP-BART uses a trainable visual projection module to map the visual features extracted by CLIP to the embedding space of BART [20]. CLIP-BART uses task-specific prompts as the textual input: For example, in Fig. 6, \"\" is added as prompt to the text for VQA. The mapped visual features are then prepped to the textual input. Together, they serve as the input to BART. CLIP-BART is trained and evaluated in a multitask setting, that is, VQAv2 [7] and GQA [12] for visual question answering, NLVR" + }, + { + "bbox": [ + 130, + 519, + 482, + 665 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 519, + 482, + 665 + ], + "type": "text", + "content": " [35] using natural language for visual reasoning, and COCO Cap. [23] for image captioning." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "type": "text", + "content": "Routing Functions For VL PEFT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 298, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 298, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 298, + 199 + ], + "type": "text", + "content": "CLIP-BART unifies the four tasks into one text generation problem following [3], where four different task prompts distinguish the task types. CLIP-BART can be considered as the representative of the current state-of-the-art on VL PEFT tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 298, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 298, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 298, + 403 + ], + "type": "text", + "content": "To integrate our routing functions into the model, two types of inputs " + }, + { + "bbox": [ + 130, + 200, + 298, + 403 + ], + "type": "inline_equation", + "content": "x_{R}" + }, + { + "bbox": [ + 130, + 200, + 298, + 403 + ], + "type": "text", + "content": " can be considered, namely the visual features after the visual projection, and the last encoder hidden state as presented in Fig. 6. The visual features only contain visual information. Because the last encoder hidden state contains both textual and visual information, it can only be added to the decoder. We experiment with both LoRA and Adapter as the PEFT method for fine-tuning CLIP-BART. In this multitask learning setting, one can use (1) a single LoRA/Adapter shared across tasks or (2) multiple LoRAs/Adapters without parameter" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 403, + 481, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 481, + 452 + ], + "type": "text", + "content": "sharing, where each task has its own LoRA/Adapter. We start by exploring the impact of different routing functions on the single LoRA/Adapter setting. Then, based on the observations obtained in the setting with single Adapter, we explore ways of using the routing functions in the multiple Adapters setting." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 127, + 482, + 289 + ], + "blocks": [ + { + "bbox": [ + 307, + 127, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 307, + 127, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 307, + 127, + 482, + 289 + ], + "type": "image", + "image_path": "a5e03c18dc9aeea9b48c2643a0f2cf2d65354fe1e58ae3fafc24bb29369af54e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 307, + 482, + 385 + ], + "lines": [ + { + "bbox": [ + 304, + 307, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 482, + 385 + ], + "type": "text", + "content": "Fig. 6: Architecture of CLIP-BART with Adapter. The standard Adapter architecture is adopted (see Fig. 1). Only the modules in green are trainable. To apply routing functions, we use the pooled last encoder hidden state as feature " + }, + { + "bbox": [ + 304, + 307, + 482, + 385 + ], + "type": "inline_equation", + "content": "x_{R}" + }, + { + "bbox": [ + 304, + 307, + 482, + 385 + ], + "type": "text", + "content": ", which is added to the Adapters in the decoder." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 468, + 279, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 468, + 279, + 480 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 279, + 480 + ], + "type": "text", + "content": "5.2 Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 486, + 482, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 584 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 584 + ], + "type": "text", + "content": "Following [11, 37], we use CLIP-ResNet101 as our vision encoder [31]. We use " + }, + { + "bbox": [ + 130, + 486, + 482, + 584 + ], + "type": "inline_equation", + "content": "\\mathrm{BART}_{\\mathrm{base}}" + }, + { + "bbox": [ + 130, + 486, + 482, + 584 + ], + "type": "text", + "content": " as our encoder-decoder LM as in [37]. Following [37], we use " + }, + { + "bbox": [ + 130, + 486, + 482, + 584 + ], + "type": "inline_equation", + "content": "r = 96" + }, + { + "bbox": [ + 130, + 486, + 482, + 584 + ], + "type": "text", + "content": " for all experiments. The training pipeline in [37] is used for a fair comparison. Specifically, (1) we use the same seed for all experiments. (2) Due to limited resources, we do not perform an extensive hyper-parameter search as in [37]. (3) We train the model with AdamW optimizer and linear decay for 20 epochs. We warm-up the learning rate from 0 to the highest in the first 2 epochs. (4) The batch size is set to 500. For more details, we refer to Appendix A." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 598, + 198, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 598, + 198, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 598, + 198, + 609 + ], + "type": "text", + "content": "5.3 Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Experiments with a single Adapter/LoRA We first report results of experiments on using single Adapter/LoRA shared across tasks in Tab. 7. In this case, the Adapter/LoRA is trained with data from four different tasks, and naturally we can only use one type of routing function throughout the experiments." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "T. Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 168, + 479, + 244 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 7: Experiments with different types of routing functions using CLIP-BART with single LoRA/Adapter. Following [11] " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "(\\dagger)^{11}" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": ", we report the average CIDER score for COCO Cap., accuracy for the other three tasks. We also report the average accuracy from VQA, GQA and NLVR2. We bold and underline the best and second best results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 168, + 479, + 244 + ], + "lines": [ + { + "bbox": [ + 135, + 168, + 479, + 244 + ], + "spans": [ + { + "bbox": [ + 135, + 168, + 479, + 244 + ], + "type": "table", + "html": "
Routing FunctionsSingle LoRASingle Adapter
VQAGQANLVR2Avg.COCOCap.VQAGQANLVR2Avg.
None†65.1553.6672.5863.80115.0165.7654.1673.1964.37
xt○x′v65.6853.9673.4264.35113.9465.9254.3474.2364.83
xt+x′v65.1453.7373.5164.13114.9665.8954.1873.9064.66
xt(xv)Txv64.9453.5673.6064.03117.8065.8453.6574.3164.03
xtx''v64.8453.1372.9863.65119.2665.8353.6173.2764.24
", + "image_path": "e5243c7b6e0bd09ff3030241d721065cd986a0e299e3155fdd09529bb6279ce2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 269, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 482, + 461 + ], + "type": "text", + "content": "Tab. 7 shows that the integration of the routing functions in the single LoRa and single Adapter architecture results in improved performance in nearly all scenarios. But just like the experiments with encoder-only and decoder-only LMs in Sec. 4.3, different routing functions show different patterns. The routing functions involving matrix multiplication, that is, " + }, + { + "bbox": [ + 130, + 269, + 482, + 461 + ], + "type": "inline_equation", + "content": "x_{t}(x_{v})^{\\top}x_{v}" + }, + { + "bbox": [ + 130, + 269, + 482, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 269, + 482, + 461 + ], + "type": "inline_equation", + "content": "x_{t}x_{v}^{\\prime \\prime}" + }, + { + "bbox": [ + 130, + 269, + 482, + 461 + ], + "type": "text", + "content": ", perform exceptionally well in the COCO captioning task, while the other two routing functions involving element-wise operations, i.e. " + }, + { + "bbox": [ + 130, + 269, + 482, + 461 + ], + "type": "inline_equation", + "content": "x_{t}\\circ x_{v}^{\\prime}" + }, + { + "bbox": [ + 130, + 269, + 482, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 269, + 482, + 461 + ], + "type": "inline_equation", + "content": "x_{t} + x_{v}^{\\prime}" + }, + { + "bbox": [ + 130, + 269, + 482, + 461 + ], + "type": "text", + "content": ", show more potential on QA tasks (VQA/GQA). As for the NLVR" + }, + { + "bbox": [ + 130, + 269, + 482, + 461 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 130, + 269, + 482, + 461 + ], + "type": "text", + "content": " task, " + }, + { + "bbox": [ + 130, + 269, + 482, + 461 + ], + "type": "inline_equation", + "content": "x_{t}(x_{v})^{\\top}x_{v}" + }, + { + "bbox": [ + 130, + 269, + 482, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 269, + 482, + 461 + ], + "type": "inline_equation", + "content": "x_{t}\\circ x_{v}^{\\prime}" + }, + { + "bbox": [ + 130, + 269, + 482, + 461 + ], + "type": "text", + "content": " present stronger abilities to learn VL relationships. Note that achieving relatively low performance on one task does not necessarily mean that the corresponding routing function fails. For example, " + }, + { + "bbox": [ + 130, + 269, + 482, + 461 + ], + "type": "inline_equation", + "content": "x_{t}\\circ x_{v}^{\\prime}" + }, + { + "bbox": [ + 130, + 269, + 482, + 461 + ], + "type": "text", + "content": " does not work well for captioning, but achieves great performance on the other three tasks. This somehow contrasts with the results obtained with encoder-only and decoder-only LMs in Sec. 4.3, where " + }, + { + "bbox": [ + 130, + 269, + 482, + 461 + ], + "type": "inline_equation", + "content": "x_{t}\\circ x_{v}^{\\prime}" + }, + { + "bbox": [ + 130, + 269, + 482, + 461 + ], + "type": "text", + "content": " leads to one of the best performance on COCO Cap. The complexity of the multitask learning setup, where different loss functions are jointly optimized, might here be a confounding factor." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 462, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 569 + ], + "type": "text", + "content": "Experiments with multiple Adapters We further explore the behavior of the routing functions in models with multiple Adapters, where we train a separate Adapter per task. As also observed in [37], using multiple Adapters increases the complexity of the model, and yields worse performance compared to using a single Adapter. As presented in Tab. 8, using the four routing functions with multiple Adapters leads to comparable or better performance as compared to multiple Adapters with no routing functions. Moreover, in line with what we discovered in the experiments with models with a single Adapter, different routing functions shine on different types of tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 571, + 482, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 571, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 571, + 482, + 632 + ], + "type": "text", + "content": "In this set-up, we also experiment with combinations of routing functions. The four combinations integrate the routing functions that work best for each individual task (e.g. " + }, + { + "bbox": [ + 130, + 571, + 482, + 632 + ], + "type": "inline_equation", + "content": "x_{t}x_{v}^{\\prime}" + }, + { + "bbox": [ + 130, + 571, + 482, + 632 + ], + "type": "text", + "content": " for COCO Cap.). The results in Tab. 8 evidence that in this multitask setting combinations of routing functions yield an improved performance compared to using no routing functions. GQA requires more fine" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "type": "text", + "content": "Routing Functions For VL PEFT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 131, + 642, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 642, + 481, + 665 + ], + "type": "text", + "content": "11 " + }, + { + "bbox": [ + 131, + 642, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 131, + 642, + 481, + 665 + ], + "type": "text", + "content": ": 3 seeds. Appendix C shows that our improvements are significant using paired t-test with " + }, + { + "bbox": [ + 131, + 642, + 481, + 665 + ], + "type": "inline_equation", + "content": "p = 0.05" + }, + { + "bbox": [ + 131, + 642, + 481, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 168, + 479, + 285 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 8: Experiments with different combinations of routing functions using CLIP-BART with multiple Adapters. Following [37] " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "(\\ddagger)" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": ", we report CIDEr score for COCO Cap., accuracy for the other three tasks. We also report the average accuracy from VQA, GQA and NLVR2. We bold and underline the best and second best results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 168, + 479, + 285 + ], + "lines": [ + { + "bbox": [ + 134, + 168, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 134, + 168, + 479, + 285 + ], + "type": "table", + "html": "
Routing FunctionsRouting FunctionsPerformance
VQAGQANLVR2COCO Cap.VQAGQANLVR2Avg.COCO Cap.
None†XXXX65.454.069.863.07114.3
xt○x′v✓✓✓✓65.6153.5170.7863.30113.72
xt+x′v✓✓✓✓65.1953.8470.9963.34115.64
xt(xv)Txv✓✓✓✓64.9352.1170.2562.43116.20
xtx′v✓✓✓✓65.3252.6770.5862.86117.92
comb1xt○x′vxt○x′vxt(xv)Txvxtx′v66.3754.2169.2763.28116.94
comb2xt+x′vxt+x′vxt(xv)Txvxtx′v65.1453.3770.8863.13117.58
comb3xt○x′vxt○x′vxt○x′vxtx′v66.3453.7770.0263.38117.73
comb4xt+x′vxt+x′vxt○x′vxtx′v65.4752.9971.0163.16118.34
", + "image_path": "4149d018561ae8aeea60c49b3d66bbf1def6fddb7948724a178c4d7e3173a6c0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 309, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 480, + 357 + ], + "type": "text", + "content": "grained image representation. For efficiency reason, we only adopt " + }, + { + "bbox": [ + 130, + 309, + 480, + 357 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 130, + 309, + 480, + 357 + ], + "type": "text", + "content": " as the down-projection from the pooled last encoder hidden state. Nevertheless, comb1 can recover from this shortcoming. More detailed image representations as used in the cross-attention experiments of Tab. 5 might still enhance the results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 357, + 481, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 481, + 441 + ], + "type": "text", + "content": "Overall, our research has focused on introducing routing functions in low-rank feature approximations when fine-tuning models for downstream (in our case VL) tasks based on pre-trained foundation models. As a pioneering work on the topic, for obtaining results we did not exhaustively search for all possible choices of hyperparameters. Instead, we have focused on revealing the problem of low-rank bottlenecks and on providing a successful solution. We leave an exhaustive grid search for optimal hyperparameters as future work." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 460, + 220, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 220, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 220, + 472 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 485, + 480, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 480, + 557 + ], + "type": "text", + "content": "We have introduced feature routing functions to help guide the feature learning in low-rank bottlenecks in architectures that are typical for parameter-efficient fine-tuning such as LoRA and Adapter. We have shown their validity in vision-language tasks, where we have focused on linear functions that do not require any extra parameters. We have demonstrated performance gains in single-task and multi-task settings in combination with a variety of foundation models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": "Routing functions open many avenues for future research. Given the importance of foundation models that are trained in a self-supervised way on large data, there is the need to efficiently fine-tune them for downstream tasks. When computing low-rank approximations of the features, our work shows that the extra guidance of the routing functions is important. We see many possibilities in experimenting with other types of functions in other tasks. For instance, linear and other functions are also suited to model constraints that reflect external knowledge or expectations on the behavior of a model. Future work will show whether they are useful to guide the feature routing in low-rank bottlenecks." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "T. Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 175 + ], + "type": "text", + "content": "This work is funded by the Flanders AI Research Program, European Research Council Advanced Grant H2020-ERC-2017-ADG 788506, and the China Scholarship Council." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 217, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 137, + 217, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 217, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 137, + 217, + 481, + 293 + ], + "type": "text", + "content": "1. Aghajanyan, A., Gupta, S., Zettlemoyer, L.: Intrinsic dimensionality explains the effectiveness of language model fine-tuning. In: Zong, C., Xia, F., Li, W., Navigli, R. (eds.) Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 7319-7328. Association for Computational Linguistics, Online (Aug 2021). https://doi.org/10.18653/v1/2021.acl-long.568, https://aclanthology.org/2021.acl-long.568" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 294, + 481, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 294, + 481, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 294, + 481, + 360 + ], + "type": "text", + "content": "2. Ben Zaken, E., Goldberg, Y., Ravfogel, S.: BitFit: Simple parameter-efficient fin-tuning for transformer-based masked language-models. In: Muresan, S., Nakov, P., Villavicencio, A. (eds.) Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). pp. 1-9. Association for Computational Linguistics, Dublin, Ireland (May 2022). https://doi.org/10.18653/v1/2022.acl-short.1, https://aclanthology.org/2022.acl-short.1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 361, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 481, + 380 + ], + "type": "text", + "content": "3. Cho, J., Lei, J., Tan, H., Bansal, M.: Unifying vision-and-language tasks via text generation. In: ICML (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 382, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 382, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 382, + 481, + 434 + ], + "type": "text", + "content": "4. Choromanski, K.M., Likhosherstov, V., Dohan, D., Song, X., Gane, A., Sarlos, T., Hawkins, P., Davis, J.Q., Mohiuddin, A., Kaiser, L., Belanger, D.B., Colwell, L.J., Weller, A.: Rethinking attention with performers. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=Ua6zuk0WRH" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 436, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 481, + 501 + ], + "type": "text", + "content": "5. Dettmers, T., Pagnoni, A., Holtzman, A., Zettlemoyer, L.: Qlora: Efficient finetuning of quantized llms. In: Oh, A., Neumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 10088-10115. Curran Associates, Inc. (2023), https://proceedings.neurips.cc/paper_files/paper/2023/file/1feb87871436031bdc0f2beaa62a049b-Paper-Conference.pdf" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 502, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 481, + 555 + ], + "type": "text", + "content": "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=YicbFdNTTy" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 556, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 481, + 599 + ], + "type": "text", + "content": "7. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "text", + "content": "8. He, J., Zhou, C., Ma, X., Berg-Kirkpatrick, T., Neubig, G.: Towards a unified view of parameter-efficient transfer learning. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=ORDcd5Axok" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "9. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for NLP. In: Chaudhuri, K., Salakhutdinov, R. (eds.) Proceedings of the 36th International" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "type": "text", + "content": "Routing Functions For VL PEFT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 149, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 149, + 117, + 480, + 149 + ], + "type": "text", + "content": "Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 97, pp. 2790-2799. PMLR (09-15 Jun 2019), https://proceedings.mlr.press/v97/houlsby19a.html" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 192 + ], + "type": "text", + "content": "10. Hu, E.J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=nZeVKeeFYf9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "text", + "content": "11. Hu, Z.Y., Li, Y., Lyu, M.R., Wang, L.: Vl-pet: Vision-and-language parameter-efficient tuning via granularity control. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3010-3020 (October 2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 228, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 228, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 134, + 228, + 480, + 258 + ], + "type": "text", + "content": "12. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 260, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 260, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 134, + 260, + 480, + 281 + ], + "type": "text", + "content": "13. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 282, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 282, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 134, + 282, + 481, + 369 + ], + "type": "text", + "content": "14. Karimi Mahabadi, R., Ruder, S., Dehghani, M., Henderson, J.: Parameter-efficient multi-task fine-tuning for transformers via shared hypernetworks. In: Zong, C., Xia, F., Li, W., Navigli, R. (eds.) Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 565-576. Association for Computational Linguistics, Online (Aug 2021). https://doi.org/10.18653/v1/2021.acl-long.47, https://aclanthology.org/2021.acl-long.47" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 369, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 369, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 134, + 369, + 481, + 423 + ], + "type": "text", + "content": "15. Katharopoulos, A., Vyas, A., Pappas, N., Fleuret, F.: Transformers are RNNs: Fast autoregressive transformers with linear attention. In: III, H.D., Singh, A. (eds.) Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 5156-5165. PMLR (13-18 Jul 2020), https://proceedings.mlr.press/v119/katharopoulos20a.html" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 425, + 481, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 425, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 134, + 425, + 481, + 467 + ], + "type": "text", + "content": "16. Koohpayegani, S.A., L, N.K., Nooralinejad, P., Kolouri, S., Pirsiavash, H.: NOLA: Networks as linear combination of low rank random basis. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=TjfXcDgvzk" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 468, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 468, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 134, + 468, + 481, + 500 + ], + "type": "text", + "content": "17. Kopiczko, D.J., Blankevoort, T., Asano, Y.M.: VeRA: Vector-based random matrix adaptation. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=NjNfLdxr3A" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 502, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 502, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 134, + 502, + 481, + 555 + ], + "type": "text", + "content": "18. Lei, T., Bai, J., Brahma, S., Ainslie, J., Lee, K., Zhou, Y., Du, N., Zhao, V.Y., Wu, Y., Li, B., Zhang, Y., Chang, M.W.: Conditional adapters: Parameter-efficient transfer learning with fast inference. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=IyYyKovOAj" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 556, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 556, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 556, + 481, + 621 + ], + "type": "text", + "content": "19. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. In: Moens, M.F., Huang, X., Specia, L., Yih, S.W.t. (eds.) Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 3045-3059. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic (Nov 2021). https://doi.org/10.18653/v1/2021.emnlp-main.243, https://aclanthology.org/2021.emnlp-main.243" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "type": "text", + "content": "20. Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., Stoyanov, V., Zettlemoyer, L.: BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In: Jurafsky, D., Chai, J., Schluter, N., Tetreault, J. (eds.) Proceedings of the 58th Annual Meeting" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "text", + "content": "T. Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 149, + 117, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 117, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 149, + 117, + 481, + 149 + ], + "type": "text", + "content": "of the Association for Computational Linguistics. pp. 7871-7880. Association for Computational Linguistics, Online (Jul 2020). https://doi.org/10.18653/v1/2020.acl-main.703, https://aclanthology.org/2020.acl-main.703" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "21. Li, C., Farkhoor, H., Liu, R., Yosinski, J.: Measuring the intrinsic dimension of objective landscapes. In: International Conference on Learning Representations (2018), https://openreview.net/forum?id=ryup8-WCW" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 481, + 258 + ], + "type": "text", + "content": "22. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. In: Zong, C., Xia, F., Li, W., Navigli, R. (eds.) Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). pp. 4582-4597. Association for Computational Linguistics, Online (Aug 2021). https://doi.org/10.18653/v1/2021.acl-long.353, https://aclanthology.org/2021.acl-long.353" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 259, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 304 + ], + "type": "text", + "content": "23. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) Computer Vision - ECCV 2014. pp. 740-755. Springer International Publishing, Cham (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 304, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 369 + ], + "type": "text", + "content": "24. Lin, Z., Madotto, A., Fung, P.: Exploring versatile generative language model via parameter-efficient transfer learning. In: Cohn, T., He, Y., Liu, Y. (eds.) Findings of the Association for Computational Linguistics: EMNLP 2020. pp. 441-459. Association for Computational Linguistics, Online (Nov 2020). https://doi.org/10.18653/v1/2020-findings-emnlp.41, https://aclanthology.org/2020-findings-emnlp.41" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "25. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized BERT pretraining approach. CoRR abs/1907.11692 (2019), http://arxiv.org/abs/1907.11692" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019), https://openreview.net/forum?id=Bkg6RiCqY7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "text", + "content": "27. Luo, Z., Hu, Z., Xi, Y., Zhang, R., Ma, J.: I-tuning: Tuning frozen language models with image for lightweight image captioning. In: ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5 (2023). https://doi.org/10.1109/ICASSP49357.2023.10096424" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 479, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 522 + ], + "type": "text", + "content": "28. mahabadi, R.K., Henderson, J., Ruder, S.: Compacter: Efficient low-rank hypercomplex adapter layers. In: Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems (2021), https://openreview.net/forum?id=bqGK5PyI6-N" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 522, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 599 + ], + "type": "text", + "content": "29. Mao, Y., Mathias, L., Hou, R., Almahairi, A., Ma, H., Han, J., Yih, S., Khabsa, M.: UniPELT: A unified framework for parameter-efficient language model tuning. In: Muresan, S., Nakov, P., Villavicencio, A. (eds.) Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 6253-6264. Association for Computational Linguistics, Dublin, Ireland (May 2022). https://doi.org/10.18653/v1/2022.acl-long.433, https://aclanthology.org/2022.acl-long.433" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "30. Qin, Z., Sun, W., Deng, H., Li, D., Wei, Y., Lv, B., Yan, J., Kong, L., Zhong, Y.: cosformer: Rethinking softmax in attention. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=B18CQrx2Up4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 310, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 310, + 91, + 447, + 101 + ], + "type": "text", + "content": "Routing Functions For VL PEFT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 149 + ], + "type": "text", + "content": "Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 8748-8763. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/radford21a.html" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 151, + 481, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 151, + 481, + 173 + ], + "spans": [ + { + "bbox": [ + 133, + 151, + 481, + 173 + ], + "type": "text", + "content": "32. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I.: Language models are unsupervised multitask learners. In: Preprint (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 175, + 481, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 175, + 481, + 218 + ], + "spans": [ + { + "bbox": [ + 132, + 175, + 481, + 218 + ], + "type": "text", + "content": "33. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research 21(140), 1-67 (2020), http://jmlr.org/papers/v21/20-074.html" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 220, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 220, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 481, + 296 + ], + "type": "text", + "content": "34. Rücklé, A., Geigle, G., Glockner, M., Beck, T., Pfeiffer, J., Reimers, N., Gurevych, I.: AdapterDrop: On the efficiency of adapters in transformers. In: Moens, M.F., Huang, X., Specia, L., Yih, S.W.t. (eds.) Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 7930-7946. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic (Nov 2021). https://doi.org/10.18653/v1/2021.emnlp-main.626, https://aclanthology.org/2021.emnlp-main.626" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 297, + 481, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 362 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 362 + ], + "type": "text", + "content": "35. Suhr, A., Zhou, S., Zhang, A., Zhang, I., Bai, H., Artzi, Y.: A corpus for reasoning about natural language grounded in photographs. In: Korhonen, A., Traum, D., Marquez, L. (eds.) Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. pp. 6418-6428. Association for Computational Linguistics, Florence, Italy (Jul 2019). https://doi.org/10.18653/v1/P19-1644, https://aclanthology.org/P19-1644" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 364, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 481, + 407 + ], + "type": "text", + "content": "36. Sung, Y.L., Cho, J., Bansal, M.: LST: Ladder side-tuning for parameter and memory efficient transfer learning. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=isPnnaTZaP5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 409, + 481, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 481, + 442 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 481, + 442 + ], + "type": "text", + "content": "37. Sung, Y.L., Cho, J., Bansal, M.: Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5227-5237 (June 2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 443, + 481, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 443, + 481, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 443, + 481, + 519 + ], + "type": "text", + "content": "38. Valipour, M., Rezagholizadeh, M., Kobyzev, I., Ghodsi, A.: DyLoRA: Parameter-efficient tuning of pre-trained models using dynamic search-free low-rank adaptation. In: Vlachos, A., Augenstein, I. (eds.) Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics. pp. 3274-3287. Association for Computational Linguistics, Dubrovnik, Croatia (May 2023). https://doi.org/10.18653/v1/2023.eacl-main.239, https://aclanthology.org/2023.eacl-main.239" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 521, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 597 + ], + "type": "text", + "content": "39. Wang, Y., Agarwal, S., Mukherjee, S., Liu, X., Gao, J., Awadallah, A.H., Gao, J.: AdaMix: Mixture-of-adaptations for parameter-efficient model tuning. In: Goldberg, Y., Kozareva, Z., Zhang, Y. (eds.) Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. pp. 5744-5760. Association for Computational Linguistics, Abu Dhabi, United Arab Emirates (Dec 2022). https://doi.org/10.18653/v1/2022.emnlp-main.388, https://aclanthology.org/2022.emnlp-main.388" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 598, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 481, + 642 + ], + "type": "text", + "content": "40. Xu, Y., Xie, L., Gu, X., Chen, X., Chang, H., Zhang, H., Chen, Z., ZHANG, X., Tian, Q.: QA-loRA: Quantization-aware low-rank adaptation of large language models. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=WvFoJccpo8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "41. Zhang, Q., Chen, M., Bukharin, A., He, P., Cheng, Y., Chen, W., Zhao, T.: Adaptive budget allocation for parameter-efficient fine-tuning. In: The Eleventh Inter-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "text", + "content": "T. Qu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 172 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 149, + 117, + 479, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 117, + 479, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 117, + 479, + 138 + ], + "type": "text", + "content": "national Conference on Learning Representations (2023), https://openreview.net/forum?id=1q62uWRJjiY" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "type": "text", + "content": "42. Zhou, H., Wan, X., Vulic, I., Korhonen, A.: Autopeft: Automatic configuration search for parameter-efficient fine-tuning. Transactions of the Association for Computational Linguistics 12 (2024)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "type": "text", + "content": "Routing Functions For VL PEFT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Invertible Neural Warp for NeRF/9222ab52-5e8c-4d58-9015-e57bdb13f7a5_content_list.json b/2024/Invertible Neural Warp for NeRF/9222ab52-5e8c-4d58-9015-e57bdb13f7a5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..84e03c537863e394a523ca3987cc142397d3b0c6 --- /dev/null +++ b/2024/Invertible Neural Warp for NeRF/9222ab52-5e8c-4d58-9015-e57bdb13f7a5_content_list.json @@ -0,0 +1,1697 @@ +[ + { + "type": "text", + "text": "Invertible Neural Warp for NeRF", + "text_level": 1, + "bbox": [ + 308, + 141, + 692, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shin-Fang Chng, Ravi Garg, Hemanth Saratchandran, and Simon Lucey", + "bbox": [ + 238, + 189, + 761, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adelaide University \nAustralian Institute for Machine Learning \nshinfang.chng@adelaide.edu.au \nhttps://sfchng.github.io/ineurowarping-github.io/", + "bbox": [ + 308, + 217, + 692, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. This paper tackles the simultaneous optimization of pose and Neural Radiance Fields (NeRF). Departing from the conventional practice of using explicit global representations for camera pose, we propose a novel overparameterized representation that models camera poses as learnable rigid warp functions. We establish that modeling the rigid warps must be tightly coupled with constraints and regularization imposed. Specifically, we highlight the critical importance of enforcing invertibility when learning rigid warp functions via neural network and propose the use of an Invertible Neural Network (INN) coupled with a geometry-informed constraint for this purpose. We present results on synthetic and real-world datasets, and demonstrate that our approach outperforms existing baselines in terms of pose estimation and high-fidelity reconstruction due to enhanced optimization convergence.", + "bbox": [ + 261, + 306, + 743, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Neural Radiance Fields $\\cdot$ Joint scene reconstruction and pose estimation $\\cdot$ Implicit Neural Representations", + "bbox": [ + 261, + 500, + 740, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 551, + 375, + 569 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "NeRF [25] has recently emerged as a compelling approach for synthesizing photorealistic images from novel views. NeRF employs a multi-layer perceptron (MLP) to model a volumetric representation of a 3D scene. It operates by minimizing the photometric loss, which is the discrepancy between rendered images and actual images. NeRF's ability to reconstruct high-fidelity signals, coupled with its memory efficiency, has propelled its adoption across a wide array of applications [13,18,31,43,44,46-48], demonstrating its significant impact and versatility.", + "bbox": [ + 212, + 583, + 785, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One of the primary challenges with NeRF is the requirement for precisely known camera poses for each captured image. To address this challenge, several approaches such as BARF [23], NeRFmm [40], and GARF [10] have been developed. These methods facilitate the simultaneous optimization of the NeRF and the camera poses, using a compact, six-dimensional vector to represent the camera poses efficiently. However, this compact parameterization, while prevalent in contemporary structure from motion (SfM) literature [32, 34], has been shown to struggle with poor basin of convergence when solved simultaneously with a NeRF [9]. Drawing wisdom from machine learning, where overparameterization has been recognized as a catalyst for enhanced optimization convergence", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8b32bd0a12042b8d9fae560350aa5178f27c48774b4a6761d8e6b5d1b70d6279.jpg", + "image_caption": [ + "(a) Groundtruth" + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 336, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/26f48b667a3667385eafecf1a2bf9ae4d239d497f973548c9616f6bb599ef006.jpg", + "image_caption": [ + "(b) BARF [23]" + ], + "image_footnote": [], + "bbox": [ + 341, + 143, + 444, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7052df06016205a2920c2f73d2037a89586beab5590f896a588aa17ce54b2b87.jpg", + "image_caption": [ + "(c) Naive", + "Fig. 1: We investigate how overparameterizing rigid warps of rays with an MLP benefits the joint optimization task of camera pose and NeRF. This example estimates the warps that align the color-coded patches in Fig. 1a while solving for the neural field. Unlike \"BARF\" and \"naive\" MLP pose overparameterization methods fail catastrophically, enforcing invertibility, either implicitly (Fig. 1d) or explicitly (Fig. 1e) significantly improves warp estimation, see Sec. 4.1 for details of each method. We establish that invertibility is a crucial for MLP-based rigid warp representation." + ], + "image_footnote": [], + "bbox": [ + 447, + 143, + 550, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4f7cbcd4c74cd780f0453a70eff3015258797719adcd6e8ec880f99e7851d94c.jpg", + "image_caption": [ + "(d) Implicit-Invertible MLP" + ], + "image_footnote": [], + "bbox": [ + 553, + 143, + 656, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ac5acfcb88a1cf34ab4ae73180940a053909240fa69adbfaf47ce5e20443b197.jpg", + "image_caption": [ + "(e) Explicit-Invertible INN" + ], + "image_footnote": [], + "bbox": [ + 661, + 143, + 764, + 263 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in modern deep neural networks [22, 26, 35], this paper explores the potential of pose overparameterization for simultaneous pose and neural field estimation.", + "bbox": [ + 212, + 405, + 785, + 434 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach: In traditional NeRF setups, accurately known extrinsic camera pose, comprising of a global rotation and translation for each image, are used to explicitly map pixel coordinates and the camera center to determine the viewing rays in a global world coordinate system [25]. Following the warping operation, the colors and volume densities along each ray in the world coordinate space is manipulated individually through a photometric loss function. In this paper, we explore scenarios where camera poses are not known. Specifically, we propose using a neural network to model the rigid warp function of ray. While it may seem counterintuitive to replace a succinct pose function with a more complex MLP, we argue that the enhanced convergence properties of such overparameterization [1, 26, 27] - in conjunction with the right constraint and prior - outweigh the increased functional flexibility.", + "bbox": [ + 212, + 436, + 787, + 618 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, we highlight the critical role of enforcing invertibility when learning rigid warps using an MLP. To achieve an approximate bijective solution, one remedy is to use an auxiliary network to represent the backward warp; however this will introduce computational overhead. To this end, we propose explicitly modeling inversions in the neural network architectures, formally learning an Invertible Neural Network (INN). Our results demonstrate that opting for an architecture that is explicitly invertible is more effective for jointly optimizing both the pose and radiance field, outperforming existing strong baselines [9,23]. Notably, our INN-based approach achieves an improvement of over $50\\%$ in pose accuracy when compared to the standard SE3 parameterization [23].", + "bbox": [ + 212, + 619, + 787, + 771 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Shin-Fang et al.", + "bbox": [ + 271, + 114, + 379, + 128 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1 Our use of invertibility strictly adheres to the well-established mathematical definition. Let $f$ be a function whose domain is $\\mathcal{X}$ and codomain is $\\mathcal{Y}$ . $f$ is invertible iff there exists a function $g$ from $\\mathcal{Y}$ to $\\mathcal{X}$ such that $g(f(x)) = x \\forall x \\in \\mathcal{X}$ and $f(g(y)) = y \\forall y \\in \\mathcal{Y}$ [14]. We use bijective and invertible interchangeably throughout our paper.", + "bbox": [ + 217, + 782, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 143, + 395, + 160 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Joint NeRF and pose estimation", + "text_level": 1, + "bbox": [ + 215, + 186, + 529, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite NeRF demonstrating compelling results in novel view synthesis, NeRF requires accurate camera poses. The differentiable nature of volume rendering used in NeRF facilitates the backpropagation through the scene representation to update the camera poses. NeRFmm [40] demonstrates the possibility of optimizing camera poses within the NeRF framework. BARF [23] introduces a coarse-to-fine positional encoding scheduling to improve the joint optimization of NeRF and camera poses, and remains a widely adopted method. GARF [10] and SiNeRF [42] advocate for leveraging the smoothness inherent in non-traditional activations to mitigate the noisy gradients due to high frequencies in positional embeddings. In contrast, NoPe-NeRF [5] uses monocular depth prior as a geometry prior to constrain the relative poses. SPARF [37] and SCNeRF [16] demonstrate that using keypoint matches or dense correspondence can constrain the relative pose estimates with ray-to-ray correspondence losses. Park et al. [30] proposes a preconditioning strategy to enhance camera pose optimization. DBARF [8] proposes using low-frequency feature maps to address the joint optimization problem for generalizable NeRF. Bian et al. [4] proposes a pose residual field which learn the pose corrections to refine the initial camera pose for neural surface reconstruction. Closest to our approach is a very recent work L2G by Chen et al. [9], which tackles the camera pose representation using an overparameterization strategy. However, we achieve overparameterization in different manner. Unlike L2G which learn an MLP to predict rigid $SE(3)$ transformations, we propose using an MLP to model the rigid warp function between the pixel and the ray space. Our work argues that while overparameterization can be achieved in different manner, it is tightly coupled with the regularization and constraints imposed. In our case, invertibility of warps becomes an essential constraint.", + "bbox": [ + 215, + 220, + 787, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Overparameterization", + "text_level": 1, + "bbox": [ + 215, + 618, + 441, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overparameterization in deep learning involves employing models with a substantially greater number of parameters than the quantity of training data. Recent research [22, 26, 35] demonstrates that neural networks, when overparameterized, can effectively generalize to new, unseen data, noting that an increase in parameters often correlates with a decrease in test error. Moreover, [26] revealed that this ability to generalize does not necessarily require explicit regularization, suggesting that the optimization process of overparameterized neural networks inherently prefers solutions that are more likely to generalize well. Additionally, [1, 2, 27, 29] illustrates that overparameterized networks are capable of consistently finding a global minimum through gradient-based optimization methods. These insights highlight the remarkable ability of overparameterized neural networks to deliver accurate predictions on unseen data, emphasizing their robustness and efficacy in generalizing beyond the training dataset.", + "bbox": [ + 215, + 652, + 787, + 849 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Invertible Neural Warp for NeRF", + "bbox": [ + 506, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3 Invertible Neural Networks for Deformation Fields", + "text_level": 1, + "bbox": [ + 215, + 146, + 679, + 161 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent advancements in Invertible Neural Networks (INNs) [3,7,12] have broadened their use in the field of 3D deformation. These networks are particularly useful in modeling homermorphic deformation, where the mapping between any frames are bijective and continuous. Because of this capability, they are now being used in various areas, particularly modeling deformation in spatial [17,33,45] and temporal [6,20,28,39].", + "bbox": [ + 212, + 170, + 787, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 284, + 380, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We define the mathematical notations for the camera operations and the joint camera pose estimation in Sec. 3.1. Further, we outline our approach in Sec. 3.2.", + "bbox": [ + 214, + 316, + 785, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Bundle-Adjust NeRF Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 358, + 553, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Camera pose We consider a set of $T$ input images as $\\{\\mathcal{I}_t\\}_{t=1}^T$ taken by camera with intrinsic matrix $K \\in \\mathbb{R}^{3 \\times 3}$ . For each camera $t$ corresponding to these images, BARF-style approaches [23] define its camera-to-world (C2W) as $P = (\\mathbf{R}_t, \\mathbf{t}_t) \\in SE(3)$ , where $\\mathbf{R}_t \\in SO(3)$ and $\\mathbf{t}_t \\in \\mathbb{R}^3$ , respectively.", + "bbox": [ + 214, + 382, + 785, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Camera projection For any vector $\\mathbf{x} \\in \\mathbb{R}^l$ of dimension $l$ , we define its homogeneous representation $\\bar{\\mathbf{x}} \\in \\mathbb{R}^{l+1}$ as $\\bar{\\mathbf{x}} = [\\mathbf{x}^T, 1]$ . We define $\\pi$ as the camera projection operator, which maps a 3D point in the camera coordinate frame denoted as $\\mathbf{x}^{(C)} \\in \\mathbb{R}^3$ to a corresponding 2D pixel coordinate $\\mathbf{u} \\in \\mathbb{R}^2$ . $\\pi^{-1}$ denotes the camera backprojection that maps a pixel $\\mathbf{u}$ coordinate and depth $z$ to a 3D point in the camera coordinate as $\\mathbf{x}^{(C)}$ , for e.g., $\\pi(\\mathbf{x}^{(C)}) \\cong K\\mathbf{x}^{(C)}$ and $\\pi^{-1}(\\mathbf{u}, z) = zK^{-1}\\bar{\\mathbf{u}}$ . We use ${}^{(C)}$ and ${}^{(W)}$ to denote that it is defined within the camera and world coordinate system respectively.", + "bbox": [ + 214, + 452, + 787, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NeRF NeRF represents the volumetric field of a 3D scene as $f(\\gamma(\\mathbf{x}), \\gamma(\\mathbf{d})) \\to (\\mathbf{c}, \\sigma)$ , which maps a 3D location $\\mathbf{x} \\in \\mathbb{R}^3$ and a viewing direction $\\mathbf{d}$ to a RGB color $\\mathbf{c} \\in \\mathbb{R}^3$ and volume density $\\sigma \\in \\mathbb{R}$ . $\\gamma: \\mathbb{R}^3 \\to \\mathbb{R}^{3+6L}$ is the positional embedding function with $L$ frequency bases [25]. This function is parameterized using an MLP as $f_{\\Theta_{rgb}}$ . Given $T$ input images $\\{\\mathcal{I}_t\\}_{t=1}^T$ with corresponding camera poses $\\{P_t\\}_{t=1}^T$ , NeRF is optimized by minimizing photometric loss $\\mathcal{L}_{rgb}$ between synthesized images $\\hat{\\mathcal{I}}$ and original image $\\mathcal{I}$ as", + "bbox": [ + 214, + 583, + 787, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {\\Theta} _ {r g b}} \\sum_ {t = 1} ^ {T} \\sum_ {\\mathbf {u} \\in \\mathbb {R} ^ {2}} \\| \\hat {\\mathcal {I}} (\\mathbf {u}, P _ {t}; \\boldsymbol {\\Theta} _ {r g b}) - \\mathcal {I} _ {i} (\\mathbf {u}) \\| _ {2} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 702, + 785, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Volume rendering For simplicity, let's start by assuming the rendering operation of NeRF operates in the camera coordinate system. We will generalize this later. Each pixel coordinate determines a viewing direction $\\mathbf{d}$ in the camera coordinate system, whose origin is the camera center of projection $\\mathbf{o}^{(C)}$ . We can define a 3D point along the camera ray associated with $\\mathbf{u}$ sampled at depth $z_{i}$", + "bbox": [ + 214, + 763, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Shin-Fang et al.", + "bbox": [ + 271, + 114, + 379, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as $\\mathbf{r}^{(C)}(z) = \\mathbf{o}^{(C)} + z_{i,u}K^{-1}\\bar{\\mathbf{u}}$ . To render the colour of $\\hat{\\mathcal{L}}_{i,u}$ at pixel coordinate $\\mathbf{u}$ , we sample $M$ discrete depth values along the ray between a near bound $z_{n}$ and a far bound $z_{f}$ . For each sampled value, we query the NeRF $f_{\\Theta_{rgb}}$ to obtain the corresponding radiance fields. The output from NeRF is then aggregated to render the RGB colour as", + "bbox": [ + 212, + 143, + 782, + 220 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {I}} (\\mathbf {u}) = \\int_ {z _ {n}} ^ {z _ {f}} T (\\mathbf {u}, z) \\sigma (\\mathbf {r} (z)) \\mathbf {c} (\\mathbf {r} (z)) \\delta z, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 228, + 785, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T(\\mathbf{u},z) = \\exp (-\\int_{z_n}^z\\sigma (\\mathbf{r}(z)))\\delta z'$ denotes the accumulated transmittance value along the ray. We refer readers to [21, 25] for more details of volume rendering operation. In practice, Eq. (2) is approximated using $M$ points sampled along the ray at depth $\\{z_i\\}_{i = 1}^M$ , which produces radiance field outputs as $\\{\\mathbf{y}_i\\}_{i = 1}^M$ . Denoting the ray compositing function in Eq. (2) as $g(.)\\in \\mathbb{R}^{4M}\\to \\mathbb{R}^3$ , we can rewrite $\\mathcal{I}(\\mathbf{u}) = g(\\{\\mathbf{y}_i\\}_{i = 1}^M)$ . Finally, given a camera pose $P$ , we can then transform the ray $\\mathbf{r}^{(C)}$ to the world coordinate to obtain $\\mathbf{r}^{(W)}$ through a 3D rigid transformation $\\mathcal{T}$ . The rendered image is then obtained as", + "bbox": [ + 212, + 272, + 784, + 396 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {I}} (\\mathbf {u}, \\mathbf {p}) = g \\left(\\left\\{f \\left(\\mathcal {T} \\left(\\mathbf {r} ^ {(C)} (z), P\\right); \\boldsymbol {\\Theta} _ {r g b}\\right) \\right\\} _ {i = 1} ^ {M}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 406, + 785, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Joint optimization of pose and NeRF Prior works [10, 23, 40, 42] demonstrate that it is feasible to optimize both camera pose and NeRF by minimizing $\\mathcal{L}_{rgb}$ in Eq. (1). This is achieved by considering $P$ as estimizable parameters, see Sec. 3.1 for its parameterization. Consequently, ray $\\mathbf{r}$ is now dependent on the camera pose. Mathematically, this joint optimization can be rewritten as", + "bbox": [ + 212, + 462, + 782, + 537 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {P, \\boldsymbol {\\Theta} _ {r g b}} \\sum_ {t = 1} ^ {T} \\sum_ {\\mathbf {u} \\in \\mathbb {R} ^ {2}} \\| \\hat {\\mathcal {I}} (\\underbrace {\\mathcal {T} \\left(\\mathbf {r} ^ {(C)} (z , \\bar {\\mathbf {u}}) ; P\\right)} _ {G (.)}; \\boldsymbol {\\Theta} _ {r g b}) - \\mathcal {I} _ {i} (\\mathbf {u}) \\| _ {2} ^ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 547, + 785, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Invertible Neural Warp for Ray Transform", + "text_level": 1, + "bbox": [ + 214, + 609, + 612, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose to overparameterize $P$ using an Invertible Neural Network (INN). There exist two options for parameterization: (i) use a separate INN for each camera $P_{t}$ ; (ii) use a single INN that is shared across all frames, coupled with a learnable code that is unique to frame $t$ . Drawing inspiration from the dynamic NeRF method used for representing deformation fields [6,31,38,39] and also considering parameter efficiency, we have chosen to pursue the latter strategy for our proposed pose overparameterization - a single, globally shared neural $\\Theta_{\\mathcal{W}}$ network across all frames, coupled with an estimizable per-frame latent code $\\varPhi_t\\in\\mathbb{R}^D$ , see supp. (Sec. C) for the comparison of using multiple-INNs versus single-NN. Consequently, we can rewrite $G(.)$ in Eq. (6) as $h(\\mathbf{r}^{(C)};\\Theta_{\\mathcal{W}},\\varPhi_t)$ where $h(.):\\mathbb{R}^{3 + D}\\to \\mathbb{R}^3$ . Fig. 2 presents our approach. In our approach, we", + "bbox": [ + 212, + 635, + 782, + 801 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Invertible Neural Warp for NeRF", + "bbox": [ + 506, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2 This can be succinctly written as $\\mathbf{r}^{(C)}(z) = z_{i,u}\\mathbf{d}$ as $\\mathbf{o}^{(C)}$ is $[0,0,0]^T$ in camera coordinate space.", + "bbox": [ + 217, + 809, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f35034cff4e28a497495b3c340d53d7ccaf3b638b130c164c7fcc9106c443139.jpg", + "image_caption": [ + "Fig. 2: An overview of our INN-based approach, illustrated using two views $\\mathcal{I}_1$ and $\\mathcal{I}_2$ . INN which is denoted as $h_{\\Theta_W}$ takes the pixel locations in the camera coordinate system $\\mathbf{x}_{i,t}^{(C)}$ , along with the frame-dependent latent code $\\varPhi_t$ , and output its corresponding location in the world coordinate system as $\\mathbf{x}_{i,t}^{(W)}$ , see Sec. 3.2 for full details." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 522, + 354 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a7753746da043b6f500c0a3eb653e59ff03a2aafdeee3d0f7323c78f9fa768ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 142, + 785, + 352 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "model each pixel in the camera coordinate system $\\mathbf{x}_{i,t}^{(C)}$ as an individual ray. Our proposed INN is designed for transforming these rays from camera coordinate to the world coordinate. Specifically, our proposed INN takes in the pixel coordinates $\\mathbf{x}_{i,t}^{(C)}$ and camera center $\\mathbf{o}_t^{(C)}$ , both defined in the camera coordinates $t$ , coupled with the frame-dependent latent code and outputs their corresponding equivalent $\\mathbf{x}_{i,t}^{(W)}$ and camera center $\\mathbf{o}_t^{(W)}$ in the world coordinates.", + "bbox": [ + 212, + 441, + 787, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Rigidity prior In our formulation, each pixel is represented as an individual ray within the camera coordinate system. This inherently relaxes the rigidity constraint. As a result, the output from the INN does not necessarily conform to a global rigid motion. Given known camera-world correspondences $(\\mathbf{x}_{i,t}^{(C)},\\mathbf{x}_{i,t}^{(W)})$ we can solve a closed-form rigid registration problem to determine a global pose, which can be integrated into our optimization problem as a rigidity prior $\\mathcal{L}_{rigid}$", + "bbox": [ + 212, + 547, + 787, + 642 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {T ^ {*}} \\sum_ {i = 1} ^ {L} \\| \\mathbf {x} _ {i, t} ^ {(C)} - T ^ {*} \\circ \\mathbf {x} _ {i, t} ^ {(W)} \\| _ {2} ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 648, + 785, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Final optimization problem We solve our final optimization problem as", + "bbox": [ + 215, + 698, + 759, + 714 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {\\Phi} _ {t}, \\boldsymbol {\\Theta} _ {\\mathcal {W}}, \\boldsymbol {\\Theta} _ {r g b}} \\sum_ {t = 1} ^ {T} \\sum_ {\\mathbf {u} \\in \\mathbb {R} ^ {2}} \\| \\hat {\\mathcal {I}} (h (\\mathbf {r} ^ {(c)}; \\boldsymbol {\\Theta} _ {\\mathcal {W}}, \\boldsymbol {\\Phi} _ {t}); \\boldsymbol {\\Theta} _ {r g b}) - \\mathcal {I} _ {i} (\\mathbf {u}) \\| _ {2} ^ {2} + \\lambda \\mathcal {L} _ {r i g i d}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 261, + 723, + 785, + 765 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Advantages of INN for Overparameterizing Rigid Ray Warps", + "text_level": 1, + "bbox": [ + 215, + 785, + 766, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BARF-based approach parameterizes camera pose $P$ of each frame using a $SE(3)$ (see Sec. 3.1), which guarantees that $\\mathcal{T}$ is a bijective mapping. Therefore, when", + "bbox": [ + 215, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Shin-Fang et al.", + "bbox": [ + 271, + 114, + 379, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "overparameterizing camera poses, it is crucial that the neural network adheres to the bijection property because this one-to-one correspondence ensures that there is a unique output in the world space for every point in the camera space. As we will demonstrate in Sec. 4.2, simply applying a rudimentary strategy (denoted as Naive) when overparameterizing the rigid warps of ray (camera-world) with an MLP often does not achieve convergence, see Fig. 1c and Tab. 1. Consequently, to attain this bijective property using an MLP, it is necessary to introduce an auxiliary network to model the backward warps (Implicit-Invertible MLP). While effective, it presents a significant drawback: it results in a twofold rise in the computational complexity due to the existence of the backward network to enforce the self-consistency. To mitigate this substantial increase in computational demands inherent in the modified MLP approach, we propose the use of INNs for parameterize these bijections. INN implements the bijective mappings by composing affine transformations into several blocks. Within each block, the input coordinates are divided into two segments; the first part remains constant and is used to parameterize the transformation that is applied to the second part [6]. Besides their inherent invertibility, INN also offers the advantage of homeomorphic property, which potentially facilitate a more flexible optimization trajectory that is less susceptible to a suboptimal minimum trajectory, see Sec. 4.5", + "bbox": [ + 212, + 146, + 787, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 454, + 375, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1Baselines", + "text_level": 1, + "bbox": [ + 215, + 484, + 339, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our approach with two representative methods in pose-NeRF joint optimization: the standard global $SE3$ -approach BARF [23], and the overparameterized representation L2G [9]. For all experiments, we use the original implementations including their default settings for coarse-to-fine scheduling, architecture and hyperparameters, see supp. (Sec. A) for more details. Additionally, in our 2D planar experiments, we include a comparison with another two variants called the Naive (MLP) and Implicit-Invertible MLP to execute ray transform. This comparison is specifically designed to highlight the significance of invertibility when employing MLPs for executing ray transformation.", + "bbox": [ + 212, + 507, + 787, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Local-to-global (L2G) [9] L2G uses an MLP to predict rigid SE3 transformation for each ray. These predicted transformation parameters are then used to analytically estimate the transformed coordinates in the world space.", + "bbox": [ + 212, + 651, + 787, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Naive This is the simplest version of our baseline that uses one primary network, denoted as $h_{fwd}:(\\mathbf{x}^{(C)},\\varPhi_t)\\to \\mathbf{x}^{(W)}$ , to learn the forward mapping which takes in the coordinates from the camera space (coupled with a per-frame latent code), and outputs the corresponding coordinates in the world space.", + "bbox": [ + 212, + 704, + 787, + 766 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implicit-Invertible MLP We use two networks $h_{fwd}$ and $h_{bwd}$ to enforce approximate invertibility. Alongside primary network $h_{fwd}$ , we use a secondary network $h_{bwd} : (\\mathbf{x}^{(W)}, \\varPhi_t) \\to \\hat{\\mathbf{x}}^{(C)}$ to invert the outputs from the primary network $h_{fwd}$ . To minimize deviations from bijections, we introduce a regularization term $\\mathcal{L}_{\\text{implicit}}$ as $\\| \\mathbf{x}^{(C)} - \\hat{\\mathbf{x}}^{(C)} \\|_2^2$ into the optimization problem Eq. (6).", + "bbox": [ + 214, + 772, + 787, + 849 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Invertible Neural Warp for NeRF", + "bbox": [ + 506, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Explicit-Invertible INN (Ours) We explicitly model inversions in the neural network architecture by formally learning an INN. We have chosen to utilize architecture proposed by NDR-INN [6], see supp. (Sec. A.1) for the architecture details that we use for all our experiments.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 2D Planar Neural Image Alignment", + "text_level": 1, + "bbox": [ + 214, + 229, + 555, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Following BARF [9, 10, 23], we learn a 2D neural image field, for creating a homography-based panoramic image from $N$ patches cropped from the original image, each generated with random homography perturbations. Specifically, we learn a 2D coordinate network $f(\\Theta_{rgb})$ to render the stitched image. Each pixel in the $N$ training patch is warped using the estimated homography $H$ to create the rendered image. We choose the \"cat\" image from ImageNet [11]. We initialize patch warps as identity and fix the gauge freedom by anchoring the first warp to align the neural image to the original image [9, 10, 23]. We randomly generate 20 different homography instances, with scale-noise parameter 0.1 and 0.2 for homography and translation, respectively. We solve Eq. (5) for the homography using a Direct Linear Transform (DLT) solver3.", + "bbox": [ + 212, + 255, + 787, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experiment settings. We evaluate our proposed method against BARF [23], and our three overparameterized network variants: Naive, Implicit- and Explicit-Invertible MLPs, as detailed in Sec. 4.1. For both naive and Implicit-Invertible MLPs, we utilized a Leaky-ReLU MLP with five 256-dimensional hidden units, and a 16-dimensional latent code to represent the frame-dependent embeddings $\\phi_t$ . We also follow the default coarse-to-fine scheduling established by BARF. In the case of Implicit-Invertible MLP, we use the same architecture both $h_{fwd}$ and $h_{bwd}$ . We use the Adam optimizer [19] to optimize for both the network weights $\\Theta_{rgb}$ and $\\Theta_{\\mathcal{W}}$ . We set the learning rate for both $\\Theta_{rgb}$ and $\\Theta_{\\mathcal{W}}$ at $1\\times 10^{-3}$ , with both decaying exponentially to $1\\times 10^{-4}$ and $1\\times 10^{-5}$ , respectively. We set the weighting term for $\\mathcal{L}_{rigid}$ for both overparameterized MLPs to $1\\times 10^{2}$ and the consistency term $\\mathcal{L}_{implicit}$ for Implicit-Invertible MLP to $1\\times 10^{1}$ .", + "bbox": [ + 212, + 431, + 787, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness to noise perturbations Fig. 3 analyzes the robustness of BARF and our approach under different noise perturbations across 20 different homography instances. Each run began with the initialization of the homography using the \"groundtruth\", and noise perturbations were gradually introduced, ranging from 0 to 0.3 to the translation component of the homography. Fig. 3 indicates that our representation exhibits a higher tolerance to noise compared to BARF.", + "bbox": [ + 212, + 623, + 787, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. Tab. 1 summarizes the statistical results from 20 runs, where we report the warp error and patch reconstruction error. We quantify the warp error in terms of corner error, defined as the L2 distance between the groundtruth corner position and estimated corner position, and PSNR as the metric to assess the reconstruction quality. We used 5-pixel threshold to define the success", + "bbox": [ + 212, + 739, + 787, + 816 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Shin-Fang et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "3 https://github.com/kornia/kornia", + "bbox": [ + 217, + 823, + 482, + 839 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b17d1eae930c5bc8390a47b0275281b6897058455224ce79820848965ed1f60.jpg", + "image_caption": [ + "Fig. 3: Basin convergence analysis of our approach versus BARF in a 2D planar experiment. On the left, we show the success rate of 20 runs, where we initialize the homography using groundtruth, and gradually introduced noise perturbations to the translation component. The noise scale is varied from 0 to 0.30. We used 5-pixel threshold to determine success convergence. Notably, our approach (blue) demonstrates higher noise tolerance compared to BARF (red). On the right, we show a qualitative comparison when both methods are perturbed with the highest magnitude of noise." + ], + "image_footnote": [], + "bbox": [ + 220, + 146, + 450, + 267 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/aa5ec814e80ca259bc42627f6d513d55dd5f2185e8655068112a276be203ce23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 146, + 787, + 256 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/e80186b5f90fa0a07d0ece99b402fe9a0423299ba5974b9eb697375d6c194378.jpg", + "table_caption": [ + "Table 1: Statistical result for 20 homography runs, with scale noise of 0.1 for homography and 0.2 for translation. The warp error is quantified in terms of corner error, and the patch reconstruction error in measured in PSNR. We provide mean and std dev for the evaluation. We used 5-pixel threshold to determine success convergence." + ], + "table_footnote": [], + "table_body": "
Corner error (px) ↓Patch PSNR ↑Success rate ↑\n(Upper bound:1.00)
MeanStd. dev.MeanStd. dev.
BARF [23]29.6328.1828.944.380.30
Naive85.5930.3125.862.070.00
Implicit-Invertible13.9222.9333.703.930.65
Explicit-Invertible (INN)4.706.4734.712.370.75
", + "bbox": [ + 220, + 431, + 782, + 526 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "convergence. Fig. 1 presents a qualitative result for a homography instance. By enforcing approximate invertibility, the Implicit-Invertible MLP demonstrates a significantly higher rate of successful convergence compared to the naive version of MLP, with success rates improved by $65\\%$ . We further show that by explicitly injecting invertibility into the architecture (Ours), the success rate is increased to $75\\%$ . This result reinforces that invertibility is crucial when learning rigid warp functions via overparameterization, and using an architecture that guarantees bijective property is effective in ensuring the pose converge to an optimal solution during joint optimization in practice. Henceforth, we will focus exclusively on using our INN-based approach for subsequent results. For completeness, we also compare with another setup, where each frame is parameterized with a neural network, see supp. (Sec. B.1). Interestingly, apart from the parameter efficiency, we find that using a single global neural network for all frames is sufficient to converge to a good pose solutions. We hypothesized that the difference may be attributed to the benefits of gradient sharing in the shared neural network setup. This result is aligned with the findings by Bian et al. [4]. Consequently, we adhere to the design where we use one single INN shared across all the frames, coupled with a frame-specific latent code for the rest of our experiments.", + "bbox": [ + 212, + 568, + 787, + 843 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Invertible Neural Warp for NeRF", + "bbox": [ + 508, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a2ebeb77bc58d43a7af206c0b50a570e50866057e1703f7bb6553d1cb219cd1e.jpg", + "image_caption": [ + "(a) BARF [23]" + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 403, + 271 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3c8165a25bfd71dcc17d991a2e70cc63ad0b0ad3a17c0365743616f4b35d214f.jpg", + "image_caption": [ + "(b) L2G [9]", + "Fig. 4: Qualitative analysis of reconstruction error on leaves (top) and trex (bottom). We present the average image reconstruction error through insets. Our approach presents the lowest misalignment error, as indicated darker areas in the error map." + ], + "image_footnote": [], + "bbox": [ + 406, + 143, + 589, + 271 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/df35eb51c9f26f45c51f65301be34f41a44855fe4a9a341ed753f1c041b842e6.jpg", + "image_caption": [ + "(c) INN(Ours)" + ], + "image_footnote": [], + "bbox": [ + 591, + 143, + 777, + 271 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Neural Radiance Fields (NeRF)", + "text_level": 1, + "bbox": [ + 215, + 353, + 524, + 369 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we compare our proposed representation with BARF [23] and L2G [9]. We assume known intrinsic for all methods. We perform our experiments on both the LLFF [24] (Sec. 4.4), DTU [15] (Sec. 4.6) as well as Blender datasets in supp. (Sec. B2). We solve Eq. (5) for the global rigid SE3 transformation using Umayama algorithm $^4$ .", + "bbox": [ + 212, + 376, + 784, + 452 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation metrics. For pose estimation, we report the accuracy of the poses after globally aligning the optimized poses to the groundtruth [9,10,23,37]. We assess view synthesis using PSNR, SSIM and LPIPS. A standard procedure in view synthesis evaluation involves performing test-time photometric optimization on the trained models. This additional step is intended to factor out the pose errors, which may otherwise compromise the quality of the synthesized views [9,10,23,36,37]. This process is akin to a pose refinement method which minimises the photometric error on the synthesized image while keeping the trained NeRF model fixed. However, it is important to recognize that this pose correction may not accurately represent the initial accuracy of the methods in terms of pose estimation for view synthesis. Therefore, we opt to report the view synthesis quality both before and after the pose refinement step. On DTU, we extend our evaluation to include comparisons of rendered depth with ground-truth depth using mean depth absolute error, as well as reconstruction accuracy using Chamfer distance. For the Chamfer evaluation, we utilize the optimized poses estimated from all methods and employ a neural surface reconstruction algorithm called Voxurf [41] for geometry reconstruction. Further details on pose alignment and metrics computations in the supp. (Sec. A).", + "bbox": [ + 212, + 459, + 787, + 731 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 Forward Facing Scenes: LLFF", + "text_level": 1, + "bbox": [ + 215, + 751, + 506, + 766 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Experiment settings. The standard LLFF benchmark dataset [24] consists of eight real-world, forward-facing scenes captured using hand-held cameras.", + "bbox": [ + 212, + 773, + 784, + 803 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Shin-Fang et al.", + "bbox": [ + 271, + 114, + 379, + 128 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "4 https://github.com/naver/roma", + "bbox": [ + 217, + 809, + 460, + 824 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "For our proposed method, we evaluate the estimated global poses Eq. (5).", + "bbox": [ + 220, + 824, + 723, + 839 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/82c5cb05e2135fb6811afa097f5c6aceee4868daac85ae1644fb719cacd185a1.jpg", + "table_caption": [ + "Table 2: Evaluation of the LLFF dataset [24] using initial identity poses. We compute results for BARF [23] and L2G [9] using their codebases. The results are averaged over all eight scenes, see supp. (Sec. B) for a breakdown of results per scene." + ], + "table_footnote": [], + "table_body": "
Pose accuracyNovel view synthesis
Rotation (°)Translation (×100)Before test-timeAfter test-time
PSNRSSIMLPIPSPSNRSSIMLPIPS
BARF [23]0.900.4017.000.410.3223.820.720.24
L2G [9]0.480.3017.990.460.2624.350.750.21
INN(Ours)0.310.2419.310.520.2524.280.740.22
", + "bbox": [ + 220, + 186, + 782, + 282 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Following [10, 23, 37], we initialize all camera poses with the identity transformations for all the methods. We employ the same training and testing split as in BARF [23]. We use the evaluation metrics described in Sec. 4.3.", + "bbox": [ + 212, + 311, + 782, + 357 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation details. We train all methods for 200k iterations and randomly sample 2048 pixel rays at each optimization step [9, 10, 23]. We train without hierarchical sampling. We set the learning for $\\Theta_{rgb}$ to be $1\\times 10^{-3}$ decaying to $3\\times 10^{-4}$ , and $5\\times 10^{-4}$ for $\\Theta_{\\mathcal{W}}$ decaying to $1\\times 10^{-6}$ . We also follow the default coarse-to-fine scheduling by BARF [23], see supp. for full details.", + "bbox": [ + 212, + 367, + 784, + 444 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results. Our approach achieves a substantial reduction of rotation errors (70% vs. BARF and 35% vs. L2G) and translation errors (50% vs. BARF and 20% vs. L2G). Additionally, the superior performance of both our method and L2G over BARF further highlight the merits of overparameterization for simultaneous pose and neural field estimation. This significant improvement in camera pose accuracy directly enhances the performance of view synthesis, as evident in Tab. 2. In particular, instances such as trex and leaves observe a significant improvement. Fig. 4 illustrates the absolute error between the original image and the rendered image on these two instances. Our approach presents the lowest misalignment error, as indicated darker areas in the error map. It is important to note, however, while L2G and our approach appear comparable after test-time optimization, this refinement procedure has mitigated the pose estimation noise to improve PSNR, as discussed in Sec. 4.3. For more qualitative results and ablation studies, please refer to supp. (Sec. B).", + "bbox": [ + 212, + 454, + 787, + 667 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.5 Homeomorphism perspective: A qualitative analysis", + "text_level": 1, + "bbox": [ + 215, + 676, + 689, + 694 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We present a qualitative analysis on a single-view pose estimation that sheds light on the empirical effectiveness of our approach compared to the L2G method [9]. Our method leverages the concept of an INN, which predict homeomorphisms—continuous, invertible transformations not limited to the rigid motions of the $SE(3)$ group. Unlike the L2G method, which constrains pose estimation within the rigid bounds of $SE(3)$ , our INN-based approach embraces a broader spectrum of transformations. This grants the optimization process a higher degree of flexibility, allowing for a diverse range of optimization paths and facilitating a smoother trajectory towards the solution. To empirically validate our", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Invertible Neural Warp for NeRF", + "bbox": [ + 506, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/de9791282530dc7f2aa2d00955f63115a77d431ac92212169c6d571e7c0c7744.jpg", + "image_caption": [ + "iter=0", + "Fig. 5: Qualitative analysis of intermediate rendered image compared to superimposed Groundtruth (lighter visualization) when using L2G [9] and our approach on single-view pose estimation. Notably, when using our approach for pose estimation, we observe noticeable deformation in the rendered scene depicted in the bottom row (zoom in for better view). These deformations indicate that at each iteration, the INN predicts general homeomorphisms that are not rigid transformation. Thus yielding a flexible optimization trajectory that does not land in a suboptimal minimum trajectory." + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 364, + 287 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d36029cc3d3a8d6d36357dcc155cad1b619265218fd889bb819da70e0a464f18.jpg", + "image_caption": [ + "iter=40" + ], + "image_footnote": [], + "bbox": [ + 366, + 143, + 504, + 287 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3888edd3b5a9d0af273d3b0f90dabbb8c7db78aedc858fe07cf7fdac739c7fca.jpg", + "image_caption": [ + "iter=80" + ], + "image_footnote": [], + "bbox": [ + 504, + 143, + 643, + 287 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/54efe1a573ac49bca4d09c0b97f515f08a3fcf6d8eb19eb754df9eff17e9c4e2.jpg", + "image_caption": [ + "iter=500" + ], + "image_footnote": [], + "bbox": [ + 643, + 143, + 784, + 287 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/2c5d33e8bacd25fb1e184331467d45dbf3707563ee550f19c44d379ac9cd497f.jpg", + "table_caption": [ + "Table 3: Absolute pose accuracy evaluation of the DTU dataset [15] using initial noisy poses that corresponds to an average rotation and translation error of $15^{\\circ}$ and 70 respectively. The upper section shows rotation errors in degrees, while the lower section displays translation errors $(\\times 100)$ . red box denotes the best result." + ], + "table_footnote": [], + "table_body": "
MethodsMetricScan IDsInitial rotation err: 15°, translation err (×100) : 70
243740556365698397105106110114118mean
BARFrotation1.041.880.552.850.430.783.880.542.203.5611.812.011.212.602.52
L2G0.903.070.5710.810.740.941.550.590.642.8513.155.922.9012.474.08
INN(Ours)0.210.480.731.640.400.622.750.550.540.303.892.911.060.251.17
BARFtranslation3.004.082.0911.631.432.7812.802.136.1212.4025.885.524.085.047.07
L2G2.408.802.1332.742.783.485.082.312.375.3436.2014.018.6437.1611.67
INN(Ours)0.501.010.975.240.960.785.300.671.190.439.9611.523.780.643.07
", + "bbox": [ + 217, + 468, + 782, + 565 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "hypothesis, we conducted experiments using a trained NeRF model to estimate the camera pose relative to a 3D scene, aiming to minimize the photometric error between NeRF-rendered and actual observed images. Despite starting from the same initial pose (off by $20^{\\circ}$ ), and employing a random sampling of 2048 rays per iteration for all methods, our INN approach outperformed the L2G method. The L2G method often converged to suboptimal poses, whereas the INN method achieved accurate pose estimation, as evidenced in Fig. 5 where the NeRF-rendered images is well-aligned with the groundtruth. This success can be attributed to the INN's ability to predict a general homeomorphism, a transformation more general than a rigid transformation, which offers a significant advantage in navigating the optimization landscape more effectively and avoiding suboptimal local minima.", + "bbox": [ + 212, + 595, + 787, + 777 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 $360^{\\circ}$ Scenes: DTU", + "text_level": 1, + "bbox": [ + 215, + 785, + 413, + 799 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Experimental settings. We evaluated on 14 test scenes from DTU [4]. Following [9,23,37], we synthetically perturb the ground-truth camera poses with", + "bbox": [ + 212, + 809, + 784, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Shin-Fang et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/6f6960e28f524f0823314e0840d712c0a852834faa37144879c3c58305d58756.jpg", + "table_caption": [ + "Table 4: Geometry evaluation of the DTU dataset [15], using initial noisy poses that corresponds to an average rotation and translation error of $15^{\\circ}$ and 70 respectively. The upper section presents mean absolute depth, while the lower section displays reconstruction error measured by Chamfer distance. red box denotes the best result." + ], + "table_footnote": [], + "table_body": "
Scan IDs | Initial rotation err: 15°, translation err (×100) : 70
PoseMetric243740556365698397105106110114118mean
BARF0.110.150.110.120.140.100.260.170.170.460.530.170.100.170.20
L2Gdepth ↓0.080.240.120.380.170.120.110.180.100.180.240.170.120.840.22
INN(Ours)(abs)0.060.140.130.120.150.080.180.170.110.110.130.300.080.060.13
BARF5.474.074.397.644.835.286.435.748.067.217.758.298.245.476.35
L2GChamfer ↓4.379.084.416.678.035.923.806.006.305.597.437.358.298.166.53
INN(Ours)1.562.255.555.684.584.275.515.695.093.738.157.857.351.144.89
", + "bbox": [ + 215, + 202, + 784, + 305 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "$15\\%$ of additive Gaussian noise, which corresponds to an average rotation and translation error of $15^{\\circ}$ and 70, respectively. For a fair comparison, we used the same initialization for all methods. We refer the readers to supp. (Sec B.5) for the results with Colmap initialization.", + "bbox": [ + 212, + 335, + 782, + 397 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Implementation details. As BARF [23] and L2G [9] have not tested on DTU datasets, given that the DTU dataset encompasses $360^{\\circ}$ scenes which is similar to the original Blender dataset, we adopted the original hyperparameters used by the author for training BARF and L2G on the blender dataset. Following Bian et al. [4], we multiply the output of their local warp network by a small factor which is $\\alpha = 0.01$ for L2G [9]. For our approach, we set the learning rate for $\\Theta_{rgb}$ to be $1\\times 10^{-3}$ decaying to $1\\times 10^{-4}$ , and $\\Theta_{\\mathcal{W}}$ to start from $5\\times 10^{-4}$ and decaying to $1\\times 10^{-8}$ . We use the default coarse-to-fine scheduling by BARF [23].", + "bbox": [ + 212, + 406, + 784, + 529 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results. As demonstrated in Tab. 3, we outperform all baselines by a considerable margin across majority of the sequences in pose accuracy. Overall, our approach achieves approximately a $50\\%$ improvement in rotation and a $60\\%$ improvement in translation over BARF. When compared to L2G, our method shows a $70\\%$ increase in accuracy for both rotation and translation. Additionally, our approach also consistently surpasses both baselines in geometry evaluation, as evidenced in the depth and reconstruction error in Tab. 4 and qualitative results in Fig. 6. For the quantitative results for novel view synthesis and additional qualitative results, we refer the readers to see supp. (Sec B).", + "bbox": [ + 212, + 540, + 787, + 676 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 700, + 359, + 715 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this paper, we examine the benefits of overparameterizing poses via a MLP in the joint optimization task of camera pose and NeRF. We establish that invertibility is a crucial property. We further show that using an Invertible Neural Network, inherently equipped with a guaranteed bijection property, significantly improves the convergence in pose optimization compared to existing representative methods.", + "bbox": [ + 212, + 731, + 787, + 821 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Invertible Neural Warp for NeRF", + "bbox": [ + 506, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9ad49b1e37e52ad4388448c6cba6d2879970f095265cf737a94eb56f5e264925.jpg", + "image_caption": [ + "Fig. 6: We showcase the reconstruction results for scan24, scan55, scan105 and scan106. All meshes are generated using a neural surface reconstruction algorithm called Voxurf [41] using the poses estimated by different approaches." + ], + "image_footnote": [], + "bbox": [ + 218, + 220, + 787, + 703 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Shin-Fang et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement We thank Chee-Kheng (CK) Chng for insightful discussions and technical feedback.", + "bbox": [ + 215, + 161, + 787, + 191 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 214, + 321, + 229 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Allen-Zhu, Z., Li, Y., Liang, Y.: Learning and generalization in overparameterized neural networks, going beyond two layers. Advances in neural information processing systems 32 (2019) 2, 3", + "2. Arora, S., Du, S., Hu, W., Li, Z., Wang, R.: Fine-grained analysis of optimization and generalization for overparameterized two-layer neural networks. In: International Conference on Machine Learning. pp. 322-332. PMLR (2019) 3", + "3. Behrmann, J., Grathwohl, W., Chen, R.T., Duvenaud, D., Jacobsen, J.H.: Invertible residual networks. In: International conference on machine learning. pp. 573-582. PMLR (2019) 4", + "4. Bian, J.W., Bian, W., Prisacariu, V.A., Torr, P.: Porf: Pose residual field for accurate neural surface reconstruction. arXiv preprint arXiv:2310.07449 (2023) 3, 9, 12, 13", + "5. Bian, W., Wang, Z., Li, K., Bian, J.W., Prisacariu, V.A.: Nope-nerf: Optimising neural radiance field with no pose prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4160-4169 (2023) 3", + "6. Cai, H., Feng, W., Feng, X., Wang, Y., Zhang, J.: Neural surface reconstruction of dynamic scenes with monocular rgb-d camera. Advances in Neural Information Processing Systems 35, 967-981 (2022) 4, 5, 7, 8", + "7. Chen, R.T., Rubanova, Y., Bettencourt, J., Duvenaud, D.K.: Neural ordinary differential equations. Advances in neural information processing systems 31 (2018) 4", + "8. Chen, Y., Lee, G.H.: Dbarf: Deep bundle-adjusting generalizable neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24-34 (2023) 3", + "9. Chen, Y., Chen, X., Wang, X., Zhang, Q., Guo, Y., Shan, Y., Wang, F.: Local-to-global registration for bundle-adjusting neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8264-8273 (2023) 1, 2, 3, 7, 8, 10, 11, 12, 13, 14", + "0. Chng, S.F., Ramasinghe, S., Sherrah, J., Lucey, S.: Gaussian activated neural radiance fields for high fidelity reconstruction and pose estimation. In: European Conference on Computer Vision. pp. 264-280. Springer (2022) 1, 3, 5, 8, 10, 11", + "1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) 8", + "2. Dinh, L., Sohl-Dickstein, J., Bengio, S.: Density estimation using real nvp. arXiv preprint arXiv:1605.08803 (2016) 4", + "3. Guo, M., Fathi, A., Wu, J., Funkhouser, T.: Object-centric neural scene rendering. arXiv preprint arXiv:2012.08503 (2020) 1", + "4. Jeffreys, H., Jeffreys, B.: Methods of mathematical physics. Cambridge university press (1999) 2", + "5. Jensen, R., Dahl, A., Vogiatzis, G., Tola, E., Aanaes, H.: Large scale multi-view stereopsis evaluation. In: 2014 IEEE Conference on Computer Vision and Pattern Recognition. pp. 406-413. IEEE (2014) 10, 12, 13" + ], + "bbox": [ + 225, + 244, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Invertible Neural Warp for NeRF", + "bbox": [ + 506, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Jeong, Y., Ahn, S., Choy, C., Anandkumar, A., Cho, M., Park, J.: Self-calibrating neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5846-5854 (2021) 3", + "17. Jiang, C., Huang, J., Tagliasacchi, A., Guibas, L.J.: Shapeflow: Learnable deformation flows among 3d shapes. Advances in Neural Information Processing Systems 33, 9745-9757 (2020) 4", + "18. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: Lerf: Language embedded radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19729-19739 (2023) 1", + "19. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 8", + "20. Lei, J., Daniilidis, K.: Cadex: Learning canonical deformation coordinate space for dynamic surface representation via neural homeomorphism. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6624-6634 (2022) 4", + "21. Levoy, M.: Efficient ray tracing of volume data. ACM Transactions on Graphics (TOG) 9(3), 245-261 (1990) 5", + "22. Li, Y., Liang, Y.: Learning overparameterized neural networks via stochastic gradient descent on structured data. Advances in neural information processing systems 31 (2018) 2, 3", + "23. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: Barf: Bundle-adjusting neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5741-5751 (2021) 1, 2, 3, 4, 5, 7, 8, 9, 10, 11, 12, 13, 14", + "24. Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG) 38(4), 1-14 (2019) 10, 11", + "25. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) 1, 2, 4, 5", + "26. Neyshabur, B., Tomioka, R., Srebro, N.: Norm-based capacity control in neural networks. In: Conference on learning theory. pp. 1376-1401. PMLR (2015) 2, 3", + "27. Nguyen, Q.N., Mondelli, M.: Global convergence of deep networks with one wide layer followed by pyramidal topology. Advances in Neural Information Processing Systems 33, 11961-11972 (2020) 2, 3", + "28. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Occupancy flow: 4d reconstruction by learning particle dynamics. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5379-5389 (2019) 4", + "29. Oymak, S., Soltanolkotabi, M.: Toward moderate overparameterization: Global convergence guarantees for training shallow neural networks. IEEE Journal on Selected Areas in Information Theory 1(1), 84-105 (2020) 3", + "30. Park, K., Henzler, P., Mildenhall, B., Barron, J.T., Martin-Brualla, R.: Camp: Camera preconditioning for neural radiance fields. ACM Transactions on Graphics (TOG) 42(6), 1-11 (2023) 3", + "31. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021) 1, 5", + "32. Parra, A., Chng, S.F., Chin, T.J., Eriksson, A., Reid, I.: Rotation coordinate descent for fast globally optimal rotation averaging. In: Proceedings of the" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Shin-Fang et al.", + "bbox": [ + 271, + 114, + 379, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4298-4307 (2021) 1", + "33. Paschalidou, D., Katharopoulos, A., Geiger, A., Fidler, S.: Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3204-3215 (2021) 4", + "34. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 1", + "35. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014) 2, 3", + "36. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., et al.: Nerfstudio: A modular framework for neural radiance field development. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-12 (2023) 10", + "37. Truong, P., Rakotosaona, M.J., Manhardt, F., Tombari, F.: Sparf: Neural radiance fields from sparse and noisy poses. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4190-4200 (2023) 3, 10, 11, 12", + "38. Wang, C., MacDonald, L.E., Jeni, L.A., Lucey, S.: Flow supervision for deformable nef. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21128-21137 (2023) 5", + "39. Wang, Q., Chang, Y.Y., Cai, R., Li, Z., Hariharan, B., Holynski, A., Snavely, N.: Tracking everything everywhere all at once. arXiv preprint arXiv:2306.05422 (2023) 4, 5", + "40. Wang, Z., Wu, S., Xie, W., Chen, M., Prisacariu, V.A.: Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064 (2021) 1, 3, 5", + "41. Wu, T., Wang, J., Pan, X., Xu, X., Theobalt, C., Liu, Z., Lin, D.: Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv preprint arXiv:2208.12697 (2022) 10, 14", + "42. Xia, Y., Tang, H., Timofte, R., Van Gool, L.: Sinerf: Sinusoidal neural radiance fields for joint pose estimation and scene reconstruction. arXiv preprint arXiv:2210.04553 (2022) 3, 5", + "43. Xu, C., Wu, B., Hou, J., Tsai, S., Li, R., Wang, J., Zhan, W., He, Z., Vajda, P., Keutzer, K., et al.: Nerf-det: Learning geometry-aware volumetric representation for multi-view 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23320-23330 (2023) 1", + "44. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5438-5448 (2022) 1", + "45. Yang, G., Belongie, S., Hariharan, B., Koltun, V.: Geometry processing with neural fields. Advances in Neural Information Processing Systems 34, 22483-22497 (2021) 4", + "46. Zhan, H., Zheng, J., Xu, Y., Reid, I., Rezatofighi, H.: Activermap: Radiance field for active mapping and planning. arXiv preprint arXiv:2211.12656 (2022) 1", + "47. Zhao, F., Yang, W., Zhang, J., Lin, P., Zhang, Y., Yu, J., Xu, L.: Humannerf: Efficiently generated human radiance field from sparse inputs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7743-7753 (2022) 1" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Invertible Neural Warp for NeRF", + "bbox": [ + 506, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "ref_text", + "text": "48. Zhu, Z., Peng, S., Larsson, V., Xu, W., Bao, H., Cui, Z., Oswald, M.R., Pollefeys, M.: Nice-slam: Neural implicit scalable encoding for slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12786-12796 (2022) 1", + "bbox": [ + 217, + 146, + 787, + 204 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Shin-Fang et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Invertible Neural Warp for NeRF/9222ab52-5e8c-4d58-9015-e57bdb13f7a5_model.json b/2024/Invertible Neural Warp for NeRF/9222ab52-5e8c-4d58-9015-e57bdb13f7a5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2b5f6fc479ab64311e1a772e47ea1b3b7657ac90 --- /dev/null +++ b/2024/Invertible Neural Warp for NeRF/9222ab52-5e8c-4d58-9015-e57bdb13f7a5_model.json @@ -0,0 +1,2348 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.309, + 0.142, + 0.694, + 0.163 + ], + "angle": 0, + "content": "Invertible Neural Warp for NeRF" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.19, + 0.763, + 0.207 + ], + "angle": 0, + "content": "Shin-Fang Chng, Ravi Garg, Hemanth Saratchandran, and Simon Lucey" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.218, + 0.693, + 0.274 + ], + "angle": 0, + "content": "Adelaide University \nAustralian Institute for Machine Learning \nshinfang.chng@adelaide.edu.au \nhttps://sfchng.github.io/ineurowarping-github.io/" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.308, + 0.744, + 0.489 + ], + "angle": 0, + "content": "Abstract. This paper tackles the simultaneous optimization of pose and Neural Radiance Fields (NeRF). Departing from the conventional practice of using explicit global representations for camera pose, we propose a novel overparameterized representation that models camera poses as learnable rigid warp functions. We establish that modeling the rigid warps must be tightly coupled with constraints and regularization imposed. Specifically, we highlight the critical importance of enforcing invertibility when learning rigid warp functions via neural network and propose the use of an Invertible Neural Network (INN) coupled with a geometry-informed constraint for this purpose. We present results on synthetic and real-world datasets, and demonstrate that our approach outperforms existing baselines in terms of pose estimation and high-fidelity reconstruction due to enhanced optimization convergence." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.501, + 0.741, + 0.529 + ], + "angle": 0, + "content": "Keywords: Neural Radiance Fields \\(\\cdot\\) Joint scene reconstruction and pose estimation \\(\\cdot\\) Implicit Neural Representations" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.553, + 0.377, + 0.57 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.787, + 0.689 + ], + "angle": 0, + "content": "NeRF [25] has recently emerged as a compelling approach for synthesizing photorealistic images from novel views. NeRF employs a multi-layer perceptron (MLP) to model a volumetric representation of a 3D scene. It operates by minimizing the photometric loss, which is the discrepancy between rendered images and actual images. NeRF's ability to reconstruct high-fidelity signals, coupled with its memory efficiency, has propelled its adoption across a wide array of applications [13,18,31,43,44,46-48], demonstrating its significant impact and versatility." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "One of the primary challenges with NeRF is the requirement for precisely known camera poses for each captured image. To address this challenge, several approaches such as BARF [23], NeRFmm [40], and GARF [10] have been developed. These methods facilitate the simultaneous optimization of the NeRF and the camera poses, using a compact, six-dimensional vector to represent the camera poses efficiently. However, this compact parameterization, while prevalent in contemporary structure from motion (SfM) literature [32, 34], has been shown to struggle with poor basin of convergence when solved simultaneously with a NeRF [9]. Drawing wisdom from machine learning, where overparameterization has been recognized as a catalyst for enhanced optimization convergence" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.38, + 0.129 + ], + "angle": 0, + "content": "Shin-Fang et al." + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.144, + 0.338, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.266, + 0.335, + 0.278 + ], + "angle": 0, + "content": "(a) Groundtruth" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.144, + 0.445, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.266, + 0.436, + 0.278 + ], + "angle": 0, + "content": "(b) BARF [23]" + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.144, + 0.551, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.472, + 0.266, + 0.528, + 0.278 + ], + "angle": 0, + "content": "(c) Naive" + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.144, + 0.658, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.266, + 0.652, + 0.287 + ], + "angle": 0, + "content": "(d) Implicit-Invertible MLP" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.144, + 0.765, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.266, + 0.756, + 0.287 + ], + "angle": 0, + "content": "(e) Explicit-Invertible INN" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.288, + 0.787, + 0.385 + ], + "angle": 0, + "content": "Fig. 1: We investigate how overparameterizing rigid warps of rays with an MLP benefits the joint optimization task of camera pose and NeRF. This example estimates the warps that align the color-coded patches in Fig. 1a while solving for the neural field. Unlike \"BARF\" and \"naive\" MLP pose overparameterization methods fail catastrophically, enforcing invertibility, either implicitly (Fig. 1d) or explicitly (Fig. 1e) significantly improves warp estimation, see Sec. 4.1 for details of each method. We establish that invertibility is a crucial for MLP-based rigid warp representation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.787, + 0.435 + ], + "angle": 0, + "content": "in modern deep neural networks [22, 26, 35], this paper explores the potential of pose overparameterization for simultaneous pose and neural field estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.437, + 0.788, + 0.619 + ], + "angle": 0, + "content": "Our approach: In traditional NeRF setups, accurately known extrinsic camera pose, comprising of a global rotation and translation for each image, are used to explicitly map pixel coordinates and the camera center to determine the viewing rays in a global world coordinate system [25]. Following the warping operation, the colors and volume densities along each ray in the world coordinate space is manipulated individually through a photometric loss function. In this paper, we explore scenarios where camera poses are not known. Specifically, we propose using a neural network to model the rigid warp function of ray. While it may seem counterintuitive to replace a succinct pose function with a more complex MLP, we argue that the enhanced convergence properties of such overparameterization [1, 26, 27] - in conjunction with the right constraint and prior - outweigh the increased functional flexibility." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.788, + 0.772 + ], + "angle": 0, + "content": "Additionally, we highlight the critical role of enforcing invertibility when learning rigid warps using an MLP. To achieve an approximate bijective solution, one remedy is to use an auxiliary network to represent the backward warp; however this will introduce computational overhead. To this end, we propose explicitly modeling inversions in the neural network architectures, formally learning an Invertible Neural Network (INN). Our results demonstrate that opting for an architecture that is explicitly invertible is more effective for jointly optimizing both the pose and radiance field, outperforming existing strong baselines [9,23]. Notably, our INN-based approach achieves an improvement of over \\(50\\%\\) in pose accuracy when compared to the standard SE3 parameterization [23]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.784, + 0.789, + 0.842 + ], + "angle": 0, + "content": "1 Our use of invertibility strictly adheres to the well-established mathematical definition. Let \\( f \\) be a function whose domain is \\( \\mathcal{X} \\) and codomain is \\( \\mathcal{Y} \\). \\( f \\) is invertible iff there exists a function \\( g \\) from \\( \\mathcal{Y} \\) to \\( \\mathcal{X} \\) such that \\( g(f(x)) = x \\forall x \\in \\mathcal{X} \\) and \\( f(g(y)) = y \\forall y \\in \\mathcal{Y} \\) [14]. We use bijective and invertible interchangeably throughout our paper." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.508, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Invertible Neural Warp for NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.396, + 0.161 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.187, + 0.53, + 0.202 + ], + "angle": 0, + "content": "2.1 Joint NeRF and pose estimation" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.222, + 0.788, + 0.6 + ], + "angle": 0, + "content": "Despite NeRF demonstrating compelling results in novel view synthesis, NeRF requires accurate camera poses. The differentiable nature of volume rendering used in NeRF facilitates the backpropagation through the scene representation to update the camera poses. NeRFmm [40] demonstrates the possibility of optimizing camera poses within the NeRF framework. BARF [23] introduces a coarse-to-fine positional encoding scheduling to improve the joint optimization of NeRF and camera poses, and remains a widely adopted method. GARF [10] and SiNeRF [42] advocate for leveraging the smoothness inherent in non-traditional activations to mitigate the noisy gradients due to high frequencies in positional embeddings. In contrast, NoPe-NeRF [5] uses monocular depth prior as a geometry prior to constrain the relative poses. SPARF [37] and SCNeRF [16] demonstrate that using keypoint matches or dense correspondence can constrain the relative pose estimates with ray-to-ray correspondence losses. Park et al. [30] proposes a preconditioning strategy to enhance camera pose optimization. DBARF [8] proposes using low-frequency feature maps to address the joint optimization problem for generalizable NeRF. Bian et al. [4] proposes a pose residual field which learn the pose corrections to refine the initial camera pose for neural surface reconstruction. Closest to our approach is a very recent work L2G by Chen et al. [9], which tackles the camera pose representation using an overparameterization strategy. However, we achieve overparameterization in different manner. Unlike L2G which learn an MLP to predict rigid \\(SE(3)\\) transformations, we propose using an MLP to model the rigid warp function between the pixel and the ray space. Our work argues that while overparameterization can be achieved in different manner, it is tightly coupled with the regularization and constraints imposed. In our case, invertibility of warps becomes an essential constraint." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.619, + 0.442, + 0.634 + ], + "angle": 0, + "content": "2.2 Overparameterization" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.654, + 0.788, + 0.851 + ], + "angle": 0, + "content": "Overparameterization in deep learning involves employing models with a substantially greater number of parameters than the quantity of training data. Recent research [22, 26, 35] demonstrates that neural networks, when overparameterized, can effectively generalize to new, unseen data, noting that an increase in parameters often correlates with a decrease in test error. Moreover, [26] revealed that this ability to generalize does not necessarily require explicit regularization, suggesting that the optimization process of overparameterized neural networks inherently prefers solutions that are more likely to generalize well. Additionally, [1, 2, 27, 29] illustrates that overparameterized networks are capable of consistently finding a global minimum through gradient-based optimization methods. These insights highlight the remarkable ability of overparameterized neural networks to deliver accurate predictions on unseen data, emphasizing their robustness and efficacy in generalizing beyond the training dataset." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.38, + 0.129 + ], + "angle": 0, + "content": "Shin-Fang et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.681, + 0.162 + ], + "angle": 0, + "content": "2.3 Invertible Neural Networks for Deformation Fields" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.264 + ], + "angle": 0, + "content": "Recent advancements in Invertible Neural Networks (INNs) [3,7,12] have broadened their use in the field of 3D deformation. These networks are particularly useful in modeling homermorphic deformation, where the mapping between any frames are bijective and continuous. Because of this capability, they are now being used in various areas, particularly modeling deformation in spatial [17,33,45] and temporal [6,20,28,39]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.285, + 0.381, + 0.303 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.317, + 0.787, + 0.348 + ], + "angle": 0, + "content": "We define the mathematical notations for the camera operations and the joint camera pose estimation in Sec. 3.1. Further, we outline our approach in Sec. 3.2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.359, + 0.554, + 0.374 + ], + "angle": 0, + "content": "3.1 Bundle-Adjust NeRF Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.383, + 0.787, + 0.445 + ], + "angle": 0, + "content": "Camera pose We consider a set of \\(T\\) input images as \\(\\{\\mathcal{I}_t\\}_{t=1}^T\\) taken by camera with intrinsic matrix \\(K \\in \\mathbb{R}^{3 \\times 3}\\). For each camera \\(t\\) corresponding to these images, BARF-style approaches [23] define its camera-to-world (C2W) as \\(P = (\\mathbf{R}_t, \\mathbf{t}_t) \\in SE(3)\\), where \\(\\mathbf{R}_t \\in SO(3)\\) and \\(\\mathbf{t}_t \\in \\mathbb{R}^3\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.453, + 0.788, + 0.575 + ], + "angle": 0, + "content": "Camera projection For any vector \\(\\mathbf{x} \\in \\mathbb{R}^l\\) of dimension \\(l\\), we define its homogeneous representation \\(\\bar{\\mathbf{x}} \\in \\mathbb{R}^{l+1}\\) as \\(\\bar{\\mathbf{x}} = [\\mathbf{x}^T, 1]\\). We define \\(\\pi\\) as the camera projection operator, which maps a 3D point in the camera coordinate frame denoted as \\(\\mathbf{x}^{(C)} \\in \\mathbb{R}^3\\) to a corresponding 2D pixel coordinate \\(\\mathbf{u} \\in \\mathbb{R}^2\\). \\(\\pi^{-1}\\) denotes the camera backprojection that maps a pixel \\(\\mathbf{u}\\) coordinate and depth \\(z\\) to a 3D point in the camera coordinate as \\(\\mathbf{x}^{(C)}\\), for e.g., \\(\\pi(\\mathbf{x}^{(C)}) \\cong K\\mathbf{x}^{(C)}\\) and \\(\\pi^{-1}(\\mathbf{u}, z) = zK^{-1}\\bar{\\mathbf{u}}\\). We use \\({}^{(C)}\\) and \\({}^{(W)}\\) to denote that it is defined within the camera and world coordinate system respectively." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.584, + 0.788, + 0.692 + ], + "angle": 0, + "content": "NeRF NeRF represents the volumetric field of a 3D scene as \\( f(\\gamma(\\mathbf{x}), \\gamma(\\mathbf{d})) \\to (\\mathbf{c}, \\sigma) \\), which maps a 3D location \\( \\mathbf{x} \\in \\mathbb{R}^3 \\) and a viewing direction \\( \\mathbf{d} \\) to a RGB color \\( \\mathbf{c} \\in \\mathbb{R}^3 \\) and volume density \\( \\sigma \\in \\mathbb{R} \\). \\( \\gamma: \\mathbb{R}^3 \\to \\mathbb{R}^{3+6L} \\) is the positional embedding function with \\( L \\) frequency bases [25]. This function is parameterized using an MLP as \\( f_{\\Theta_{rgb}} \\). Given \\( T \\) input images \\( \\{\\mathcal{I}_t\\}_{t=1}^T \\) with corresponding camera poses \\( \\{P_t\\}_{t=1}^T \\), NeRF is optimized by minimizing photometric loss \\( \\mathcal{L}_{rgb} \\) between synthesized images \\( \\hat{\\mathcal{I}} \\) and original image \\( \\mathcal{I} \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.356, + 0.703, + 0.787, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {\\Theta} _ {r g b}} \\sum_ {t = 1} ^ {T} \\sum_ {\\mathbf {u} \\in \\mathbb {R} ^ {2}} \\| \\hat {\\mathcal {I}} (\\mathbf {u}, P _ {t}; \\boldsymbol {\\Theta} _ {r g b}) - \\mathcal {I} _ {i} (\\mathbf {u}) \\| _ {2} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Volume rendering For simplicity, let's start by assuming the rendering operation of NeRF operates in the camera coordinate system. We will generalize this later. Each pixel coordinate determines a viewing direction \\(\\mathbf{d}\\) in the camera coordinate system, whose origin is the camera center of projection \\(\\mathbf{o}^{(C)}\\). We can define a 3D point along the camera ray associated with \\(\\mathbf{u}\\) sampled at depth \\(z_{i}\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.508, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Invertible Neural Warp for NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.784, + 0.221 + ], + "angle": 0, + "content": "as \\(\\mathbf{r}^{(C)}(z) = \\mathbf{o}^{(C)} + z_{i,u}K^{-1}\\bar{\\mathbf{u}}\\). To render the colour of \\(\\hat{\\mathcal{L}}_{i,u}\\) at pixel coordinate \\(\\mathbf{u}\\), we sample \\(M\\) discrete depth values along the ray between a near bound \\(z_{n}\\) and a far bound \\(z_{f}\\). For each sampled value, we query the NeRF \\(f_{\\Theta_{rgb}}\\) to obtain the corresponding radiance fields. The output from NeRF is then aggregated to render the RGB colour as" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.229, + 0.786, + 0.264 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {I}} (\\mathbf {u}) = \\int_ {z _ {n}} ^ {z _ {f}} T (\\mathbf {u}, z) \\sigma (\\mathbf {r} (z)) \\mathbf {c} (\\mathbf {r} (z)) \\delta z, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.273, + 0.785, + 0.397 + ], + "angle": 0, + "content": "where \\( T(\\mathbf{u},z) = \\exp (-\\int_{z_n}^z\\sigma (\\mathbf{r}(z)))\\delta z' \\) denotes the accumulated transmittance value along the ray. We refer readers to [21, 25] for more details of volume rendering operation. In practice, Eq. (2) is approximated using \\( M \\) points sampled along the ray at depth \\( \\{z_i\\}_{i = 1}^M \\), which produces radiance field outputs as \\( \\{\\mathbf{y}_i\\}_{i = 1}^M \\). Denoting the ray compositing function in Eq. (2) as \\( g(.)\\in \\mathbb{R}^{4M}\\to \\mathbb{R}^3 \\), we can rewrite \\( \\mathcal{I}(\\mathbf{u}) = g(\\{\\mathbf{y}_i\\}_{i = 1}^M) \\). Finally, given a camera pose \\( P \\), we can then transform the ray \\( \\mathbf{r}^{(C)} \\) to the world coordinate to obtain \\( \\mathbf{r}^{(W)} \\) through a 3D rigid transformation \\( \\mathcal{T} \\). The rendered image is then obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.407, + 0.786, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {I}} (\\mathbf {u}, \\mathbf {p}) = g \\left(\\left\\{f \\left(\\mathcal {T} \\left(\\mathbf {r} ^ {(C)} (z), P\\right); \\boldsymbol {\\Theta} _ {r g b}\\right) \\right\\} _ {i = 1} ^ {M}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.784, + 0.539 + ], + "angle": 0, + "content": "Joint optimization of pose and NeRF Prior works [10, 23, 40, 42] demonstrate that it is feasible to optimize both camera pose and NeRF by minimizing \\(\\mathcal{L}_{rgb}\\) in Eq. (1). This is achieved by considering \\(P\\) as estimizable parameters, see Sec. 3.1 for its parameterization. Consequently, ray \\(\\mathbf{r}\\) is now dependent on the camera pose. Mathematically, this joint optimization can be rewritten as" + }, + { + "type": "equation", + "bbox": [ + 0.312, + 0.549, + 0.786, + 0.599 + ], + "angle": 0, + "content": "\\[\n\\min _ {P, \\boldsymbol {\\Theta} _ {r g b}} \\sum_ {t = 1} ^ {T} \\sum_ {\\mathbf {u} \\in \\mathbb {R} ^ {2}} \\| \\hat {\\mathcal {I}} (\\underbrace {\\mathcal {T} \\left(\\mathbf {r} ^ {(C)} (z , \\bar {\\mathbf {u}}) ; P\\right)} _ {G (.)}; \\boldsymbol {\\Theta} _ {r g b}) - \\mathcal {I} _ {i} (\\mathbf {u}) \\| _ {2} ^ {2}. \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.611, + 0.614, + 0.627 + ], + "angle": 0, + "content": "3.2 Invertible Neural Warp for Ray Transform" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.784, + 0.803 + ], + "angle": 0, + "content": "We propose to overparameterize \\(P\\) using an Invertible Neural Network (INN). There exist two options for parameterization: (i) use a separate INN for each camera \\(P_{t}\\); (ii) use a single INN that is shared across all frames, coupled with a learnable code that is unique to frame \\(t\\). Drawing inspiration from the dynamic NeRF method used for representing deformation fields [6,31,38,39] and also considering parameter efficiency, we have chosen to pursue the latter strategy for our proposed pose overparameterization - a single, globally shared neural \\(\\Theta_{\\mathcal{W}}\\) network across all frames, coupled with an estimizable per-frame latent code \\(\\varPhi_t\\in\\mathbb{R}^D\\), see supp. (Sec. C) for the comparison of using multiple-INNs versus single-NN. Consequently, we can rewrite \\(G(.)\\) in Eq. (6) as \\(h(\\mathbf{r}^{(C)};\\Theta_{\\mathcal{W}},\\varPhi_t)\\) where \\(h(.):\\mathbb{R}^{3 + D}\\to \\mathbb{R}^3\\). Fig. 2 presents our approach. In our approach, we" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "2 This can be succinctly written as \\(\\mathbf{r}^{(C)}(z) = z_{i,u}\\mathbf{d}\\) as \\(\\mathbf{o}^{(C)}\\) is \\([0,0,0]^T\\) in camera coordinate space." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.38, + 0.129 + ], + "angle": 0, + "content": "Shin-Fang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.523, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.143, + 0.787, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.428 + ], + "angle": 0, + "content": "Fig. 2: An overview of our INN-based approach, illustrated using two views \\(\\mathcal{I}_1\\) and \\(\\mathcal{I}_2\\). INN which is denoted as \\(h_{\\Theta_W}\\) takes the pixel locations in the camera coordinate system \\(\\mathbf{x}_{i,t}^{(C)}\\), along with the frame-dependent latent code \\(\\varPhi_t\\), and output its corresponding location in the world coordinate system as \\(\\mathbf{x}_{i,t}^{(W)}\\), see Sec. 3.2 for full details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.542 + ], + "angle": 0, + "content": "model each pixel in the camera coordinate system \\(\\mathbf{x}_{i,t}^{(C)}\\) as an individual ray. Our proposed INN is designed for transforming these rays from camera coordinate to the world coordinate. Specifically, our proposed INN takes in the pixel coordinates \\(\\mathbf{x}_{i,t}^{(C)}\\) and camera center \\(\\mathbf{o}_t^{(C)}\\), both defined in the camera coordinates \\(t\\), coupled with the frame-dependent latent code and outputs their corresponding equivalent \\(\\mathbf{x}_{i,t}^{(W)}\\) and camera center \\(\\mathbf{o}_t^{(W)}\\) in the world coordinates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.643 + ], + "angle": 0, + "content": "Rigidity prior In our formulation, each pixel is represented as an individual ray within the camera coordinate system. This inherently relaxes the rigidity constraint. As a result, the output from the INN does not necessarily conform to a global rigid motion. Given known camera-world correspondences \\((\\mathbf{x}_{i,t}^{(C)},\\mathbf{x}_{i,t}^{(W)})\\) we can solve a closed-form rigid registration problem to determine a global pose, which can be integrated into our optimization problem as a rigidity prior \\(\\mathcal{L}_{rigid}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.65, + 0.787, + 0.692 + ], + "angle": 0, + "content": "\\[\n\\min _ {T ^ {*}} \\sum_ {i = 1} ^ {L} \\| \\mathbf {x} _ {i, t} ^ {(C)} - T ^ {*} \\circ \\mathbf {x} _ {i, t} ^ {(W)} \\| _ {2} ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.699, + 0.761, + 0.715 + ], + "angle": 0, + "content": "Final optimization problem We solve our final optimization problem as" + }, + { + "type": "equation", + "bbox": [ + 0.262, + 0.724, + 0.787, + 0.766 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {\\Phi} _ {t}, \\boldsymbol {\\Theta} _ {\\mathcal {W}}, \\boldsymbol {\\Theta} _ {r g b}} \\sum_ {t = 1} ^ {T} \\sum_ {\\mathbf {u} \\in \\mathbb {R} ^ {2}} \\| \\hat {\\mathcal {I}} (h (\\mathbf {r} ^ {(c)}; \\boldsymbol {\\Theta} _ {\\mathcal {W}}, \\boldsymbol {\\Phi} _ {t}); \\boldsymbol {\\Theta} _ {r g b}) - \\mathcal {I} _ {i} (\\mathbf {u}) \\| _ {2} ^ {2} + \\lambda \\mathcal {L} _ {r i g i d}. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.786, + 0.767, + 0.802 + ], + "angle": 0, + "content": "3.3 Advantages of INN for Overparameterizing Rigid Ray Warps" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "BARF-based approach parameterizes camera pose \\( P \\) of each frame using a \\( SE(3) \\) (see Sec. 3.1), which guarantees that \\( \\mathcal{T} \\) is a bijective mapping. Therefore, when" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.508, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Invertible Neural Warp for NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.435 + ], + "angle": 0, + "content": "overparameterizing camera poses, it is crucial that the neural network adheres to the bijection property because this one-to-one correspondence ensures that there is a unique output in the world space for every point in the camera space. As we will demonstrate in Sec. 4.2, simply applying a rudimentary strategy (denoted as Naive) when overparameterizing the rigid warps of ray (camera-world) with an MLP often does not achieve convergence, see Fig. 1c and Tab. 1. Consequently, to attain this bijective property using an MLP, it is necessary to introduce an auxiliary network to model the backward warps (Implicit-Invertible MLP). While effective, it presents a significant drawback: it results in a twofold rise in the computational complexity due to the existence of the backward network to enforce the self-consistency. To mitigate this substantial increase in computational demands inherent in the modified MLP approach, we propose the use of INNs for parameterize these bijections. INN implements the bijective mappings by composing affine transformations into several blocks. Within each block, the input coordinates are divided into two segments; the first part remains constant and is used to parameterize the transformation that is applied to the second part [6]. Besides their inherent invertibility, INN also offers the advantage of homeomorphic property, which potentially facilitate a more flexible optimization trajectory that is less susceptible to a suboptimal minimum trajectory, see Sec. 4.5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.455, + 0.377, + 0.473 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.485, + 0.341, + 0.499 + ], + "angle": 0, + "content": "4.1Baselines" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.644 + ], + "angle": 0, + "content": "We compare our approach with two representative methods in pose-NeRF joint optimization: the standard global \\(SE3\\)-approach BARF [23], and the overparameterized representation L2G [9]. For all experiments, we use the original implementations including their default settings for coarse-to-fine scheduling, architecture and hyperparameters, see supp. (Sec. A) for more details. Additionally, in our 2D planar experiments, we include a comparison with another two variants called the Naive (MLP) and Implicit-Invertible MLP to execute ray transform. This comparison is specifically designed to highlight the significance of invertibility when employing MLPs for executing ray transformation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.788, + 0.698 + ], + "angle": 0, + "content": "Local-to-global (L2G) [9] L2G uses an MLP to predict rigid SE3 transformation for each ray. These predicted transformation parameters are then used to analytically estimate the transformed coordinates in the world space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.767 + ], + "angle": 0, + "content": "Naive This is the simplest version of our baseline that uses one primary network, denoted as \\( h_{fwd}:(\\mathbf{x}^{(C)},\\varPhi_t)\\to \\mathbf{x}^{(W)} \\), to learn the forward mapping which takes in the coordinates from the camera space (coupled with a per-frame latent code), and outputs the corresponding coordinates in the world space." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.773, + 0.788, + 0.851 + ], + "angle": 0, + "content": "Implicit-Invertible MLP We use two networks \\( h_{fwd} \\) and \\( h_{bwd} \\) to enforce approximate invertibility. Alongside primary network \\( h_{fwd} \\), we use a secondary network \\( h_{bwd} : (\\mathbf{x}^{(W)}, \\varPhi_t) \\to \\hat{\\mathbf{x}}^{(C)} \\) to invert the outputs from the primary network \\( h_{fwd} \\). To minimize deviations from bijections, we introduce a regularization term \\( \\mathcal{L}_{\\text{implicit}} \\) as \\( \\| \\mathbf{x}^{(C)} - \\hat{\\mathbf{x}}^{(C)} \\|_2^2 \\) into the optimization problem Eq. (6)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Shin-Fang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "Explicit-Invertible INN (Ours) We explicitly model inversions in the neural network architecture by formally learning an INN. We have chosen to utilize architecture proposed by NDR-INN [6], see supp. (Sec. A.1) for the architecture details that we use for all our experiments." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.23, + 0.557, + 0.246 + ], + "angle": 0, + "content": "4.2 2D Planar Neural Image Alignment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.256, + 0.789, + 0.424 + ], + "angle": 0, + "content": "Following BARF [9, 10, 23], we learn a 2D neural image field, for creating a homography-based panoramic image from \\(N\\) patches cropped from the original image, each generated with random homography perturbations. Specifically, we learn a 2D coordinate network \\(f(\\Theta_{rgb})\\) to render the stitched image. Each pixel in the \\(N\\) training patch is warped using the estimated homography \\(H\\) to create the rendered image. We choose the \"cat\" image from ImageNet [11]. We initialize patch warps as identity and fix the gauge freedom by anchoring the first warp to align the neural image to the original image [9, 10, 23]. We randomly generate 20 different homography instances, with scale-noise parameter 0.1 and 0.2 for homography and translation, respectively. We solve Eq. (5) for the homography using a Direct Linear Transform (DLT) solver3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.789, + 0.617 + ], + "angle": 0, + "content": "Experiment settings. We evaluate our proposed method against BARF [23], and our three overparameterized network variants: Naive, Implicit- and Explicit-Invertible MLPs, as detailed in Sec. 4.1. For both naive and Implicit-Invertible MLPs, we utilized a Leaky-ReLU MLP with five 256-dimensional hidden units, and a 16-dimensional latent code to represent the frame-dependent embeddings \\(\\phi_t\\). We also follow the default coarse-to-fine scheduling established by BARF. In the case of Implicit-Invertible MLP, we use the same architecture both \\(h_{fwd}\\) and \\(h_{bwd}\\). We use the Adam optimizer [19] to optimize for both the network weights \\(\\Theta_{rgb}\\) and \\(\\Theta_{\\mathcal{W}}\\). We set the learning rate for both \\(\\Theta_{rgb}\\) and \\(\\Theta_{\\mathcal{W}}\\) at \\(1\\times 10^{-3}\\), with both decaying exponentially to \\(1\\times 10^{-4}\\) and \\(1\\times 10^{-5}\\), respectively. We set the weighting term for \\(\\mathcal{L}_{rigid}\\) for both overparameterized MLPs to \\(1\\times 10^{2}\\) and the consistency term \\(\\mathcal{L}_{implicit}\\) for Implicit-Invertible MLP to \\(1\\times 10^{1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.789, + 0.716 + ], + "angle": 0, + "content": "Robustness to noise perturbations Fig. 3 analyzes the robustness of BARF and our approach under different noise perturbations across 20 different homography instances. Each run began with the initialization of the homography using the \"groundtruth\", and noise perturbations were gradually introduced, ranging from 0 to 0.3 to the translation component of the homography. Fig. 3 indicates that our representation exhibits a higher tolerance to noise compared to BARF." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.741, + 0.789, + 0.817 + ], + "angle": 0, + "content": "Results. Tab. 1 summarizes the statistical results from 20 runs, where we report the warp error and patch reconstruction error. We quantify the warp error in terms of corner error, defined as the L2 distance between the groundtruth corner position and estimated corner position, and PSNR as the metric to assess the reconstruction quality. We used 5-pixel threshold to define the success" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.483, + 0.84 + ], + "angle": 0, + "content": "3 https://github.com/kornia/kornia" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.509, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Invertible Neural Warp for NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.147, + 0.452, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.147, + 0.788, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.272, + 0.785, + 0.369 + ], + "angle": 0, + "content": "Fig. 3: Basin convergence analysis of our approach versus BARF in a 2D planar experiment. On the left, we show the success rate of 20 runs, where we initialize the homography using groundtruth, and gradually introduced noise perturbations to the translation component. The noise scale is varied from 0 to 0.30. We used 5-pixel threshold to determine success convergence. Notably, our approach (blue) demonstrates higher noise tolerance compared to BARF (red). On the right, we show a qualitative comparison when both methods are perturbed with the highest magnitude of noise." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.374, + 0.787, + 0.43 + ], + "angle": 0, + "content": "Table 1: Statistical result for 20 homography runs, with scale noise of 0.1 for homography and 0.2 for translation. The warp error is quantified in terms of corner error, and the patch reconstruction error in measured in PSNR. We provide mean and std dev for the evaluation. We used 5-pixel threshold to determine success convergence." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.432, + 0.784, + 0.527 + ], + "angle": 0, + "content": "
Corner error (px) ↓Patch PSNR ↑Success rate ↑\n(Upper bound:1.00)
MeanStd. dev.MeanStd. dev.
BARF [23]29.6328.1828.944.380.30
Naive85.5930.3125.862.070.00
Implicit-Invertible13.9222.9333.703.930.65
Explicit-Invertible (INN)4.706.4734.712.370.75
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.844 + ], + "angle": 0, + "content": "convergence. Fig. 1 presents a qualitative result for a homography instance. By enforcing approximate invertibility, the Implicit-Invertible MLP demonstrates a significantly higher rate of successful convergence compared to the naive version of MLP, with success rates improved by \\(65\\%\\). We further show that by explicitly injecting invertibility into the architecture (Ours), the success rate is increased to \\(75\\%\\). This result reinforces that invertibility is crucial when learning rigid warp functions via overparameterization, and using an architecture that guarantees bijective property is effective in ensuring the pose converge to an optimal solution during joint optimization in practice. Henceforth, we will focus exclusively on using our INN-based approach for subsequent results. For completeness, we also compare with another setup, where each frame is parameterized with a neural network, see supp. (Sec. B.1). Interestingly, apart from the parameter efficiency, we find that using a single global neural network for all frames is sufficient to converge to a good pose solutions. We hypothesized that the difference may be attributed to the benefits of gradient sharing in the shared neural network setup. This result is aligned with the findings by Bian et al. [4]. Consequently, we adhere to the design where we use one single INN shared across all the frames, coupled with a frame-specific latent code for the rest of our experiments." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.38, + 0.129 + ], + "angle": 0, + "content": "Shin-Fang et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.144, + 0.405, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.27, + 0.273, + 0.358, + 0.285 + ], + "angle": 0, + "content": "(a) BARF [23]" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.144, + 0.591, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.273, + 0.534, + 0.285 + ], + "angle": 0, + "content": "(b) L2G [9]" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.144, + 0.778, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.273, + 0.728, + 0.285 + ], + "angle": 0, + "content": "(c) INN(Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.286, + 0.787, + 0.327 + ], + "angle": 0, + "content": "Fig. 4: Qualitative analysis of reconstruction error on leaves (top) and trex (bottom). We present the average image reconstruction error through insets. Our approach presents the lowest misalignment error, as indicated darker areas in the error map." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.354, + 0.525, + 0.37 + ], + "angle": 0, + "content": "4.3 Neural Radiance Fields (NeRF)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.785, + 0.453 + ], + "angle": 0, + "content": "In this section, we compare our proposed representation with BARF [23] and L2G [9]. We assume known intrinsic for all methods. We perform our experiments on both the LLFF [24] (Sec. 4.4), DTU [15] (Sec. 4.6) as well as Blender datasets in supp. (Sec. B2). We solve Eq. (5) for the global rigid SE3 transformation using Umayama algorithm \\(^4\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.789, + 0.732 + ], + "angle": 0, + "content": "Evaluation metrics. For pose estimation, we report the accuracy of the poses after globally aligning the optimized poses to the groundtruth [9,10,23,37]. We assess view synthesis using PSNR, SSIM and LPIPS. A standard procedure in view synthesis evaluation involves performing test-time photometric optimization on the trained models. This additional step is intended to factor out the pose errors, which may otherwise compromise the quality of the synthesized views [9,10,23,36,37]. This process is akin to a pose refinement method which minimises the photometric error on the synthesized image while keeping the trained NeRF model fixed. However, it is important to recognize that this pose correction may not accurately represent the initial accuracy of the methods in terms of pose estimation for view synthesis. Therefore, we opt to report the view synthesis quality both before and after the pose refinement step. On DTU, we extend our evaluation to include comparisons of rendered depth with ground-truth depth using mean depth absolute error, as well as reconstruction accuracy using Chamfer distance. For the Chamfer evaluation, we utilize the optimized poses estimated from all methods and employ a neural surface reconstruction algorithm called Voxurf [41] for geometry reconstruction. Further details on pose alignment and metrics computations in the supp. (Sec. A)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.752, + 0.507, + 0.767 + ], + "angle": 0, + "content": "4.4 Forward Facing Scenes: LLFF" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.774, + 0.785, + 0.804 + ], + "angle": 0, + "content": "Experiment settings. The standard LLFF benchmark dataset [24] consists of eight real-world, forward-facing scenes captured using hand-held cameras." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.81, + 0.461, + 0.825 + ], + "angle": 0, + "content": "4 https://github.com/naver/roma" + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.825, + 0.725, + 0.84 + ], + "angle": 0, + "content": "For our proposed method, we evaluate the estimated global poses Eq. (5)." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.81, + 0.725, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.508, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Invertible Neural Warp for NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.133, + 0.788, + 0.175 + ], + "angle": 0, + "content": "Table 2: Evaluation of the LLFF dataset [24] using initial identity poses. We compute results for BARF [23] and L2G [9] using their codebases. The results are averaged over all eight scenes, see supp. (Sec. B) for a breakdown of results per scene." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.187, + 0.784, + 0.284 + ], + "angle": 0, + "content": "
Pose accuracyNovel view synthesis
Rotation (°)Translation (×100)Before test-timeAfter test-time
PSNRSSIMLPIPSPSNRSSIMLPIPS
BARF [23]0.900.4017.000.410.3223.820.720.24
L2G [9]0.480.3017.990.460.2624.350.750.21
INN(Ours)0.310.2419.310.520.2524.280.740.22
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.784, + 0.358 + ], + "angle": 0, + "content": "Following [10, 23, 37], we initialize all camera poses with the identity transformations for all the methods. We employ the same training and testing split as in BARF [23]. We use the evaluation metrics described in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.785, + 0.445 + ], + "angle": 0, + "content": "Implementation details. We train all methods for 200k iterations and randomly sample 2048 pixel rays at each optimization step [9, 10, 23]. We train without hierarchical sampling. We set the learning for \\(\\Theta_{rgb}\\) to be \\(1\\times 10^{-3}\\) decaying to \\(3\\times 10^{-4}\\), and \\(5\\times 10^{-4}\\) for \\(\\Theta_{\\mathcal{W}}\\) decaying to \\(1\\times 10^{-6}\\). We also follow the default coarse-to-fine scheduling by BARF [23], see supp. for full details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.788, + 0.669 + ], + "angle": 0, + "content": "Results. Our approach achieves a substantial reduction of rotation errors (70% vs. BARF and 35% vs. L2G) and translation errors (50% vs. BARF and 20% vs. L2G). Additionally, the superior performance of both our method and L2G over BARF further highlight the merits of overparameterization for simultaneous pose and neural field estimation. This significant improvement in camera pose accuracy directly enhances the performance of view synthesis, as evident in Tab. 2. In particular, instances such as trex and leaves observe a significant improvement. Fig. 4 illustrates the absolute error between the original image and the rendered image on these two instances. Our approach presents the lowest misalignment error, as indicated darker areas in the error map. It is important to note, however, while L2G and our approach appear comparable after test-time optimization, this refinement procedure has mitigated the pose estimation noise to improve PSNR, as discussed in Sec. 4.3. For more qualitative results and ablation studies, please refer to supp. (Sec. B)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.678, + 0.691, + 0.695 + ], + "angle": 0, + "content": "4.5 Homeomorphism perspective: A qualitative analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We present a qualitative analysis on a single-view pose estimation that sheds light on the empirical effectiveness of our approach compared to the L2G method [9]. Our method leverages the concept of an INN, which predict homeomorphisms—continuous, invertible transformations not limited to the rigid motions of the \\(SE(3)\\) group. Unlike the L2G method, which constrains pose estimation within the rigid bounds of \\(SE(3)\\), our INN-based approach embraces a broader spectrum of transformations. This grants the optimization process a higher degree of flexibility, allowing for a diverse range of optimization paths and facilitating a smoother trajectory towards the solution. To empirically validate our" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Shin-Fang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.366, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.289, + 0.316, + 0.299 + ], + "angle": 0, + "content": "iter=0" + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.144, + 0.505, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.422, + 0.289, + 0.467, + 0.299 + ], + "angle": 0, + "content": "iter=40" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.144, + 0.645, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.289, + 0.606, + 0.299 + ], + "angle": 0, + "content": "iter=80" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.144, + 0.785, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.289, + 0.748, + 0.298 + ], + "angle": 0, + "content": "iter=500" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.312, + 0.784, + 0.409 + ], + "angle": 0, + "content": "Fig. 5: Qualitative analysis of intermediate rendered image compared to superimposed Groundtruth (lighter visualization) when using L2G [9] and our approach on single-view pose estimation. Notably, when using our approach for pose estimation, we observe noticeable deformation in the rendered scene depicted in the bottom row (zoom in for better view). These deformations indicate that at each iteration, the INN predicts general homeomorphisms that are not rigid transformation. Thus yielding a flexible optimization trajectory that does not land in a suboptimal minimum trajectory." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.411, + 0.785, + 0.467 + ], + "angle": 0, + "content": "Table 3: Absolute pose accuracy evaluation of the DTU dataset [15] using initial noisy poses that corresponds to an average rotation and translation error of \\(15^{\\circ}\\) and 70 respectively. The upper section shows rotation errors in degrees, while the lower section displays translation errors \\((\\times 100)\\). red box denotes the best result." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.469, + 0.784, + 0.566 + ], + "angle": 0, + "content": "
MethodsMetricScan IDsInitial rotation err: 15°, translation err (×100) : 70
243740556365698397105106110114118mean
BARFrotation1.041.880.552.850.430.783.880.542.203.5611.812.011.212.602.52
L2G0.903.070.5710.810.740.941.550.590.642.8513.155.922.9012.474.08
INN(Ours)0.210.480.731.640.400.622.750.550.540.303.892.911.060.251.17
BARFtranslation3.004.082.0911.631.432.7812.802.136.1212.4025.885.524.085.047.07
L2G2.408.802.1332.742.783.485.082.312.375.3436.2014.018.6437.1611.67
INN(Ours)0.501.010.975.240.960.785.300.671.190.439.9611.523.780.643.07
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.779 + ], + "angle": 0, + "content": "hypothesis, we conducted experiments using a trained NeRF model to estimate the camera pose relative to a 3D scene, aiming to minimize the photometric error between NeRF-rendered and actual observed images. Despite starting from the same initial pose (off by \\(20^{\\circ}\\)), and employing a random sampling of 2048 rays per iteration for all methods, our INN approach outperformed the L2G method. The L2G method often converged to suboptimal poses, whereas the INN method achieved accurate pose estimation, as evidenced in Fig. 5 where the NeRF-rendered images is well-aligned with the groundtruth. This success can be attributed to the INN's ability to predict a general homeomorphism, a transformation more general than a rigid transformation, which offers a significant advantage in navigating the optimization landscape more effectively and avoiding suboptimal local minima." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.786, + 0.414, + 0.8 + ], + "angle": 0, + "content": "4.6 \\(360^{\\circ}\\) Scenes: DTU" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.842 + ], + "angle": 0, + "content": "Experimental settings. We evaluated on 14 test scenes from DTU [4]. Following [9,23,37], we synthetically perturb the ground-truth camera poses with" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.508, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Invertible Neural Warp for NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 4: Geometry evaluation of the DTU dataset [15], using initial noisy poses that corresponds to an average rotation and translation error of \\(15^{\\circ}\\) and 70 respectively. The upper section presents mean absolute depth, while the lower section displays reconstruction error measured by Chamfer distance. red box denotes the best result." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.306 + ], + "angle": 0, + "content": "
Scan IDs | Initial rotation err: 15°, translation err (×100) : 70
PoseMetric243740556365698397105106110114118mean
BARF0.110.150.110.120.140.100.260.170.170.460.530.170.100.170.20
L2Gdepth ↓0.080.240.120.380.170.120.110.180.100.180.240.170.120.840.22
INN(Ours)(abs)0.060.140.130.120.150.080.180.170.110.110.130.300.080.060.13
BARF5.474.074.397.644.835.286.435.748.067.217.758.298.245.476.35
L2GChamfer ↓4.379.084.416.678.035.923.806.006.305.597.437.358.298.166.53
INN(Ours)1.562.255.555.684.584.275.515.695.093.738.157.857.351.144.89
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.784, + 0.398 + ], + "angle": 0, + "content": "\\(15\\%\\) of additive Gaussian noise, which corresponds to an average rotation and translation error of \\(15^{\\circ}\\) and 70, respectively. For a fair comparison, we used the same initialization for all methods. We refer the readers to supp. (Sec B.5) for the results with Colmap initialization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.785, + 0.53 + ], + "angle": 0, + "content": "Implementation details. As BARF [23] and L2G [9] have not tested on DTU datasets, given that the DTU dataset encompasses \\(360^{\\circ}\\) scenes which is similar to the original Blender dataset, we adopted the original hyperparameters used by the author for training BARF and L2G on the blender dataset. Following Bian et al. [4], we multiply the output of their local warp network by a small factor which is \\(\\alpha = 0.01\\) for L2G [9]. For our approach, we set the learning rate for \\(\\Theta_{rgb}\\) to be \\(1\\times 10^{-3}\\) decaying to \\(1\\times 10^{-4}\\), and \\(\\Theta_{\\mathcal{W}}\\) to start from \\(5\\times 10^{-4}\\) and decaying to \\(1\\times 10^{-8}\\). We use the default coarse-to-fine scheduling by BARF [23]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.541, + 0.788, + 0.678 + ], + "angle": 0, + "content": "Results. As demonstrated in Tab. 3, we outperform all baselines by a considerable margin across majority of the sequences in pose accuracy. Overall, our approach achieves approximately a \\(50\\%\\) improvement in rotation and a \\(60\\%\\) improvement in translation over BARF. When compared to L2G, our method shows a \\(70\\%\\) increase in accuracy for both rotation and translation. Additionally, our approach also consistently surpasses both baselines in geometry evaluation, as evidenced in the depth and reconstruction error in Tab. 4 and qualitative results in Fig. 6. For the quantitative results for novel view synthesis and additional qualitative results, we refer the readers to see supp. (Sec B)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.701, + 0.36, + 0.716 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.732, + 0.788, + 0.822 + ], + "angle": 0, + "content": "In this paper, we examine the benefits of overparameterizing poses via a MLP in the joint optimization task of camera pose and NeRF. We establish that invertibility is a crucial property. We further show that using an Invertible Neural Network, inherently equipped with a guaranteed bijection property, significantly improves the convergence in pose optimization compared to existing representative methods." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Shin-Fang et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.221, + 0.789, + 0.704 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.713, + 0.789, + 0.756 + ], + "angle": 0, + "content": "Fig. 6: We showcase the reconstruction results for scan24, scan55, scan105 and scan106. All meshes are generated using a neural surface reconstruction algorithm called Voxurf [41] using the poses estimated by different approaches." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.508, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Invertible Neural Warp for NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.162, + 0.788, + 0.192 + ], + "angle": 0, + "content": "Acknowledgement We thank Chee-Kheng (CK) Chng for insightful discussions and technical feedback." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.215, + 0.323, + 0.231 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.246, + 0.785, + 0.288 + ], + "angle": 0, + "content": "1. Allen-Zhu, Z., Li, Y., Liang, Y.: Learning and generalization in overparameterized neural networks, going beyond two layers. Advances in neural information processing systems 32 (2019) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.329 + ], + "angle": 0, + "content": "2. Arora, S., Du, S., Hu, W., Li, Z., Wang, R.: Fine-grained analysis of optimization and generalization for overparameterized two-layer neural networks. In: International Conference on Machine Learning. pp. 322-332. PMLR (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.785, + 0.37 + ], + "angle": 0, + "content": "3. Behrmann, J., Grathwohl, W., Chen, R.T., Duvenaud, D., Jacobsen, J.H.: Invertible residual networks. In: International conference on machine learning. pp. 573-582. PMLR (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "4. Bian, J.W., Bian, W., Prisacariu, V.A., Torr, P.: Porf: Pose residual field for accurate neural surface reconstruction. arXiv preprint arXiv:2310.07449 (2023) 3, 9, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.453 + ], + "angle": 0, + "content": "5. Bian, W., Wang, Z., Li, K., Bian, J.W., Prisacariu, V.A.: Nope-nerf: Optimising neural radiance field with no pose prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4160-4169 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.455, + 0.785, + 0.495 + ], + "angle": 0, + "content": "6. Cai, H., Feng, W., Feng, X., Wang, Y., Zhang, J.: Neural surface reconstruction of dynamic scenes with monocular rgb-d camera. Advances in Neural Information Processing Systems 35, 967-981 (2022) 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.785, + 0.535 + ], + "angle": 0, + "content": "7. Chen, R.T., Rubanova, Y., Bettencourt, J., Duvenaud, D.K.: Neural ordinary differential equations. Advances in neural information processing systems 31 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.536, + 0.785, + 0.578 + ], + "angle": 0, + "content": "8. Chen, Y., Lee, G.H.: Dbarf: Deep bundle-adjusting generalizable neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24-34 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.579, + 0.785, + 0.633 + ], + "angle": 0, + "content": "9. Chen, Y., Chen, X., Wang, X., Zhang, Q., Guo, Y., Shan, Y., Wang, F.: Local-to-global registration for bundle-adjusting neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8264-8273 (2023) 1, 2, 3, 7, 8, 10, 11, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.785, + 0.675 + ], + "angle": 0, + "content": "0. Chng, S.F., Ramasinghe, S., Sherrah, J., Lucey, S.: Gaussian activated neural radiance fields for high fidelity reconstruction and pose estimation. In: European Conference on Computer Vision. pp. 264-280. Springer (2022) 1, 3, 5, 8, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.716 + ], + "angle": 0, + "content": "1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.744 + ], + "angle": 0, + "content": "2. Dinh, L., Sohl-Dickstein, J., Bengio, S.: Density estimation using real nvp. arXiv preprint arXiv:1605.08803 (2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.772 + ], + "angle": 0, + "content": "3. Guo, M., Fathi, A., Wu, J., Funkhouser, T.: Object-centric neural scene rendering. arXiv preprint arXiv:2012.08503 (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "4. Jeffreys, H., Jeffreys, B.: Methods of mathematical physics. Cambridge university press (1999) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Jensen, R., Dahl, A., Vogiatzis, G., Tola, E., Aanaes, H.: Large scale multi-view stereopsis evaluation. In: 2014 IEEE Conference on Computer Vision and Pattern Recognition. pp. 406-413. IEEE (2014) 10, 12, 13" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.246, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.38, + 0.129 + ], + "angle": 0, + "content": "Shin-Fang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "16. Jeong, Y., Ahn, S., Choy, C., Anandkumar, A., Cho, M., Park, J.: Self-calibrating neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5846-5854 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.231 + ], + "angle": 0, + "content": "17. Jiang, C., Huang, J., Tagliasacchi, A., Guibas, L.J.: Shapeflow: Learnable deformation flows among 3d shapes. Advances in Neural Information Processing Systems 33, 9745-9757 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.273 + ], + "angle": 0, + "content": "18. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: Lerf: Language embedded radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19729-19739 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "19. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.356 + ], + "angle": 0, + "content": "20. Lei, J., Daniilidis, K.: Cadex: Learning canonical deformation coordinate space for dynamic surface representation via neural homeomorphism. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6624-6634 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.384 + ], + "angle": 0, + "content": "21. Levoy, M.: Efficient ray tracing of volume data. ACM Transactions on Graphics (TOG) 9(3), 245-261 (1990) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.385, + 0.787, + 0.424 + ], + "angle": 0, + "content": "22. Li, Y., Liang, Y.: Learning overparameterized neural networks via stochastic gradient descent on structured data. Advances in neural information processing systems 31 (2018) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.467 + ], + "angle": 0, + "content": "23. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: Barf: Bundle-adjusting neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5741-5751 (2021) 1, 2, 3, 4, 5, 7, 8, 9, 10, 11, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.787, + 0.522 + ], + "angle": 0, + "content": "24. Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG) 38(4), 1-14 (2019) 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.787, + 0.563 + ], + "angle": 0, + "content": "25. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.787, + 0.592 + ], + "angle": 0, + "content": "26. Neyshabur, B., Tomioka, R., Srebro, N.: Norm-based capacity control in neural networks. In: Conference on learning theory. pp. 1376-1401. PMLR (2015) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.633 + ], + "angle": 0, + "content": "27. Nguyen, Q.N., Mondelli, M.: Global convergence of deep networks with one wide layer followed by pyramidal topology. Advances in Neural Information Processing Systems 33, 11961-11972 (2020) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.787, + 0.674 + ], + "angle": 0, + "content": "28. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Occupancy flow: 4d reconstruction by learning particle dynamics. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5379-5389 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.716 + ], + "angle": 0, + "content": "29. Oymak, S., Soltanolkotabi, M.: Toward moderate overparameterization: Global convergence guarantees for training shallow neural networks. IEEE Journal on Selected Areas in Information Theory 1(1), 84-105 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.787, + 0.758 + ], + "angle": 0, + "content": "30. Park, K., Henzler, P., Mildenhall, B., Barron, J.T., Martin-Brualla, R.: Camp: Camera preconditioning for neural radiance fields. ACM Transactions on Graphics (TOG) 42(6), 1-11 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.811 + ], + "angle": 0, + "content": "31. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021) 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.787, + 0.84 + ], + "angle": 0, + "content": "32. Parra, A., Chng, S.F., Chin, T.J., Eriksson, A., Reid, I.: Rotation coordinate descent for fast globally optimal rotation averaging. In: Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.508, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Invertible Neural Warp for NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4298-4307 (2021) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.787, + 0.232 + ], + "angle": 0, + "content": "33. Paschalidou, D., Katharopoulos, A., Geiger, A., Fidler, S.: Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3204-3215 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.786, + 0.261 + ], + "angle": 0, + "content": "34. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.787, + 0.303 + ], + "angle": 0, + "content": "35. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.787, + 0.359 + ], + "angle": 0, + "content": "36. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., et al.: Nerfstudio: A modular framework for neural radiance field development. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-12 (2023) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.36, + 0.787, + 0.402 + ], + "angle": 0, + "content": "37. Truong, P., Rakotosaona, M.J., Manhardt, F., Tombari, F.: Sparf: Neural radiance fields from sparse and noisy poses. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4190-4200 (2023) 3, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.787, + 0.445 + ], + "angle": 0, + "content": "38. Wang, C., MacDonald, L.E., Jeni, L.A., Lucey, S.: Flow supervision for deformable nef. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21128-21137 (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.787, + 0.487 + ], + "angle": 0, + "content": "39. Wang, Q., Chang, Y.Y., Cai, R., Li, Z., Hariharan, B., Holynski, A., Snavely, N.: Tracking everything everywhere all at once. arXiv preprint arXiv:2306.05422 (2023) 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.528 + ], + "angle": 0, + "content": "40. Wang, Z., Wu, S., Xie, W., Chen, M., Prisacariu, V.A.: Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064 (2021) 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.787, + 0.572 + ], + "angle": 0, + "content": "41. Wu, T., Wang, J., Pan, X., Xu, X., Theobalt, C., Liu, Z., Lin, D.: Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv preprint arXiv:2208.12697 (2022) 10, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.787, + 0.614 + ], + "angle": 0, + "content": "42. Xia, Y., Tang, H., Timofte, R., Van Gool, L.: Sinerf: Sinusoidal neural radiance fields for joint pose estimation and scene reconstruction. arXiv preprint arXiv:2210.04553 (2022) 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.787, + 0.671 + ], + "angle": 0, + "content": "43. Xu, C., Wu, B., Hou, J., Tsai, S., Li, R., Wang, J., Zhan, W., He, Z., Vajda, P., Keutzer, K., et al.: Nerf-det: Learning geometry-aware volumetric representation for multi-view 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23320-23330 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.671, + 0.787, + 0.713 + ], + "angle": 0, + "content": "44. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5438-5448 (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.787, + 0.755 + ], + "angle": 0, + "content": "45. Yang, G., Belongie, S., Hariharan, B., Koltun, V.: Geometry processing with neural fields. Advances in Neural Information Processing Systems 34, 22483-22497 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.756, + 0.787, + 0.785 + ], + "angle": 0, + "content": "46. Zhan, H., Zheng, J., Xu, Y., Reid, I., Rezatofighi, H.: Activermap: Radiance field for active mapping and planning. arXiv preprint arXiv:2211.12656 (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "47. Zhao, F., Yang, W., Zhang, J., Lin, P., Zhang, Y., Yu, J., Xu, L.: Humannerf: Efficiently generated human radiance field from sparse inputs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7743-7753 (2022) 1" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Shin-Fang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.788, + 0.205 + ], + "angle": 0, + "content": "48. Zhu, Z., Peng, S., Larsson, V., Xu, W., Bao, H., Cui, Z., Oswald, M.R., Pollefeys, M.: Nice-slam: Neural implicit scalable encoding for slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12786-12796 (2022) 1" + } + ] +] \ No newline at end of file diff --git a/2024/Invertible Neural Warp for NeRF/9222ab52-5e8c-4d58-9015-e57bdb13f7a5_origin.pdf b/2024/Invertible Neural Warp for NeRF/9222ab52-5e8c-4d58-9015-e57bdb13f7a5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d2ac894abcede6fe2feacb59da19c3ecb96a1c17 --- /dev/null +++ b/2024/Invertible Neural Warp for NeRF/9222ab52-5e8c-4d58-9015-e57bdb13f7a5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c65175c9aac5b6a848074cd2c1320e24e527bf9a0c518fe5ddd505ab3c9b7674 +size 6285556 diff --git a/2024/Invertible Neural Warp for NeRF/full.md b/2024/Invertible Neural Warp for NeRF/full.md new file mode 100644 index 0000000000000000000000000000000000000000..847966913823cd88bd01d9cd60b6014136e7488a --- /dev/null +++ b/2024/Invertible Neural Warp for NeRF/full.md @@ -0,0 +1,286 @@ +# Invertible Neural Warp for NeRF + +Shin-Fang Chng, Ravi Garg, Hemanth Saratchandran, and Simon Lucey + +Adelaide University +Australian Institute for Machine Learning +shinfang.chng@adelaide.edu.au +https://sfchng.github.io/ineurowarping-github.io/ + +Abstract. This paper tackles the simultaneous optimization of pose and Neural Radiance Fields (NeRF). Departing from the conventional practice of using explicit global representations for camera pose, we propose a novel overparameterized representation that models camera poses as learnable rigid warp functions. We establish that modeling the rigid warps must be tightly coupled with constraints and regularization imposed. Specifically, we highlight the critical importance of enforcing invertibility when learning rigid warp functions via neural network and propose the use of an Invertible Neural Network (INN) coupled with a geometry-informed constraint for this purpose. We present results on synthetic and real-world datasets, and demonstrate that our approach outperforms existing baselines in terms of pose estimation and high-fidelity reconstruction due to enhanced optimization convergence. + +Keywords: Neural Radiance Fields $\cdot$ Joint scene reconstruction and pose estimation $\cdot$ Implicit Neural Representations + +# 1 Introduction + +NeRF [25] has recently emerged as a compelling approach for synthesizing photorealistic images from novel views. NeRF employs a multi-layer perceptron (MLP) to model a volumetric representation of a 3D scene. It operates by minimizing the photometric loss, which is the discrepancy between rendered images and actual images. NeRF's ability to reconstruct high-fidelity signals, coupled with its memory efficiency, has propelled its adoption across a wide array of applications [13,18,31,43,44,46-48], demonstrating its significant impact and versatility. + +One of the primary challenges with NeRF is the requirement for precisely known camera poses for each captured image. To address this challenge, several approaches such as BARF [23], NeRFmm [40], and GARF [10] have been developed. These methods facilitate the simultaneous optimization of the NeRF and the camera poses, using a compact, six-dimensional vector to represent the camera poses efficiently. However, this compact parameterization, while prevalent in contemporary structure from motion (SfM) literature [32, 34], has been shown to struggle with poor basin of convergence when solved simultaneously with a NeRF [9]. Drawing wisdom from machine learning, where overparameterization has been recognized as a catalyst for enhanced optimization convergence + +![](images/8b32bd0a12042b8d9fae560350aa5178f27c48774b4a6761d8e6b5d1b70d6279.jpg) +(a) Groundtruth + +![](images/26f48b667a3667385eafecf1a2bf9ae4d239d497f973548c9616f6bb599ef006.jpg) +(b) BARF [23] + +![](images/7052df06016205a2920c2f73d2037a89586beab5590f896a588aa17ce54b2b87.jpg) +(c) Naive +Fig. 1: We investigate how overparameterizing rigid warps of rays with an MLP benefits the joint optimization task of camera pose and NeRF. This example estimates the warps that align the color-coded patches in Fig. 1a while solving for the neural field. Unlike "BARF" and "naive" MLP pose overparameterization methods fail catastrophically, enforcing invertibility, either implicitly (Fig. 1d) or explicitly (Fig. 1e) significantly improves warp estimation, see Sec. 4.1 for details of each method. We establish that invertibility is a crucial for MLP-based rigid warp representation. + +![](images/4f7cbcd4c74cd780f0453a70eff3015258797719adcd6e8ec880f99e7851d94c.jpg) +(d) Implicit-Invertible MLP + +![](images/ac5acfcb88a1cf34ab4ae73180940a053909240fa69adbfaf47ce5e20443b197.jpg) +(e) Explicit-Invertible INN + +in modern deep neural networks [22, 26, 35], this paper explores the potential of pose overparameterization for simultaneous pose and neural field estimation. + +Our approach: In traditional NeRF setups, accurately known extrinsic camera pose, comprising of a global rotation and translation for each image, are used to explicitly map pixel coordinates and the camera center to determine the viewing rays in a global world coordinate system [25]. Following the warping operation, the colors and volume densities along each ray in the world coordinate space is manipulated individually through a photometric loss function. In this paper, we explore scenarios where camera poses are not known. Specifically, we propose using a neural network to model the rigid warp function of ray. While it may seem counterintuitive to replace a succinct pose function with a more complex MLP, we argue that the enhanced convergence properties of such overparameterization [1, 26, 27] - in conjunction with the right constraint and prior - outweigh the increased functional flexibility. + +Additionally, we highlight the critical role of enforcing invertibility when learning rigid warps using an MLP. To achieve an approximate bijective solution, one remedy is to use an auxiliary network to represent the backward warp; however this will introduce computational overhead. To this end, we propose explicitly modeling inversions in the neural network architectures, formally learning an Invertible Neural Network (INN). Our results demonstrate that opting for an architecture that is explicitly invertible is more effective for jointly optimizing both the pose and radiance field, outperforming existing strong baselines [9,23]. Notably, our INN-based approach achieves an improvement of over $50\%$ in pose accuracy when compared to the standard SE3 parameterization [23]. + +# 2 Related Works + +# 2.1 Joint NeRF and pose estimation + +Despite NeRF demonstrating compelling results in novel view synthesis, NeRF requires accurate camera poses. The differentiable nature of volume rendering used in NeRF facilitates the backpropagation through the scene representation to update the camera poses. NeRFmm [40] demonstrates the possibility of optimizing camera poses within the NeRF framework. BARF [23] introduces a coarse-to-fine positional encoding scheduling to improve the joint optimization of NeRF and camera poses, and remains a widely adopted method. GARF [10] and SiNeRF [42] advocate for leveraging the smoothness inherent in non-traditional activations to mitigate the noisy gradients due to high frequencies in positional embeddings. In contrast, NoPe-NeRF [5] uses monocular depth prior as a geometry prior to constrain the relative poses. SPARF [37] and SCNeRF [16] demonstrate that using keypoint matches or dense correspondence can constrain the relative pose estimates with ray-to-ray correspondence losses. Park et al. [30] proposes a preconditioning strategy to enhance camera pose optimization. DBARF [8] proposes using low-frequency feature maps to address the joint optimization problem for generalizable NeRF. Bian et al. [4] proposes a pose residual field which learn the pose corrections to refine the initial camera pose for neural surface reconstruction. Closest to our approach is a very recent work L2G by Chen et al. [9], which tackles the camera pose representation using an overparameterization strategy. However, we achieve overparameterization in different manner. Unlike L2G which learn an MLP to predict rigid $SE(3)$ transformations, we propose using an MLP to model the rigid warp function between the pixel and the ray space. Our work argues that while overparameterization can be achieved in different manner, it is tightly coupled with the regularization and constraints imposed. In our case, invertibility of warps becomes an essential constraint. + +# 2.2 Overparameterization + +Overparameterization in deep learning involves employing models with a substantially greater number of parameters than the quantity of training data. Recent research [22, 26, 35] demonstrates that neural networks, when overparameterized, can effectively generalize to new, unseen data, noting that an increase in parameters often correlates with a decrease in test error. Moreover, [26] revealed that this ability to generalize does not necessarily require explicit regularization, suggesting that the optimization process of overparameterized neural networks inherently prefers solutions that are more likely to generalize well. Additionally, [1, 2, 27, 29] illustrates that overparameterized networks are capable of consistently finding a global minimum through gradient-based optimization methods. These insights highlight the remarkable ability of overparameterized neural networks to deliver accurate predictions on unseen data, emphasizing their robustness and efficacy in generalizing beyond the training dataset. + +# 2.3 Invertible Neural Networks for Deformation Fields + +Recent advancements in Invertible Neural Networks (INNs) [3,7,12] have broadened their use in the field of 3D deformation. These networks are particularly useful in modeling homermorphic deformation, where the mapping between any frames are bijective and continuous. Because of this capability, they are now being used in various areas, particularly modeling deformation in spatial [17,33,45] and temporal [6,20,28,39]. + +# 3 Methodology + +We define the mathematical notations for the camera operations and the joint camera pose estimation in Sec. 3.1. Further, we outline our approach in Sec. 3.2. + +# 3.1 Bundle-Adjust NeRF Preliminaries + +Camera pose We consider a set of $T$ input images as $\{\mathcal{I}_t\}_{t=1}^T$ taken by camera with intrinsic matrix $K \in \mathbb{R}^{3 \times 3}$ . For each camera $t$ corresponding to these images, BARF-style approaches [23] define its camera-to-world (C2W) as $P = (\mathbf{R}_t, \mathbf{t}_t) \in SE(3)$ , where $\mathbf{R}_t \in SO(3)$ and $\mathbf{t}_t \in \mathbb{R}^3$ , respectively. + +Camera projection For any vector $\mathbf{x} \in \mathbb{R}^l$ of dimension $l$ , we define its homogeneous representation $\bar{\mathbf{x}} \in \mathbb{R}^{l+1}$ as $\bar{\mathbf{x}} = [\mathbf{x}^T, 1]$ . We define $\pi$ as the camera projection operator, which maps a 3D point in the camera coordinate frame denoted as $\mathbf{x}^{(C)} \in \mathbb{R}^3$ to a corresponding 2D pixel coordinate $\mathbf{u} \in \mathbb{R}^2$ . $\pi^{-1}$ denotes the camera backprojection that maps a pixel $\mathbf{u}$ coordinate and depth $z$ to a 3D point in the camera coordinate as $\mathbf{x}^{(C)}$ , for e.g., $\pi(\mathbf{x}^{(C)}) \cong K\mathbf{x}^{(C)}$ and $\pi^{-1}(\mathbf{u}, z) = zK^{-1}\bar{\mathbf{u}}$ . We use ${}^{(C)}$ and ${}^{(W)}$ to denote that it is defined within the camera and world coordinate system respectively. + +NeRF NeRF represents the volumetric field of a 3D scene as $f(\gamma(\mathbf{x}), \gamma(\mathbf{d})) \to (\mathbf{c}, \sigma)$ , which maps a 3D location $\mathbf{x} \in \mathbb{R}^3$ and a viewing direction $\mathbf{d}$ to a RGB color $\mathbf{c} \in \mathbb{R}^3$ and volume density $\sigma \in \mathbb{R}$ . $\gamma: \mathbb{R}^3 \to \mathbb{R}^{3+6L}$ is the positional embedding function with $L$ frequency bases [25]. This function is parameterized using an MLP as $f_{\Theta_{rgb}}$ . Given $T$ input images $\{\mathcal{I}_t\}_{t=1}^T$ with corresponding camera poses $\{P_t\}_{t=1}^T$ , NeRF is optimized by minimizing photometric loss $\mathcal{L}_{rgb}$ between synthesized images $\hat{\mathcal{I}}$ and original image $\mathcal{I}$ as + +$$ +\min _ {\boldsymbol {\Theta} _ {r g b}} \sum_ {t = 1} ^ {T} \sum_ {\mathbf {u} \in \mathbb {R} ^ {2}} \| \hat {\mathcal {I}} (\mathbf {u}, P _ {t}; \boldsymbol {\Theta} _ {r g b}) - \mathcal {I} _ {i} (\mathbf {u}) \| _ {2} ^ {2}, \tag {1} +$$ + +Volume rendering For simplicity, let's start by assuming the rendering operation of NeRF operates in the camera coordinate system. We will generalize this later. Each pixel coordinate determines a viewing direction $\mathbf{d}$ in the camera coordinate system, whose origin is the camera center of projection $\mathbf{o}^{(C)}$ . We can define a 3D point along the camera ray associated with $\mathbf{u}$ sampled at depth $z_{i}$ + +as $\mathbf{r}^{(C)}(z) = \mathbf{o}^{(C)} + z_{i,u}K^{-1}\bar{\mathbf{u}}$ . To render the colour of $\hat{\mathcal{L}}_{i,u}$ at pixel coordinate $\mathbf{u}$ , we sample $M$ discrete depth values along the ray between a near bound $z_{n}$ and a far bound $z_{f}$ . For each sampled value, we query the NeRF $f_{\Theta_{rgb}}$ to obtain the corresponding radiance fields. The output from NeRF is then aggregated to render the RGB colour as + +$$ +\hat {\mathcal {I}} (\mathbf {u}) = \int_ {z _ {n}} ^ {z _ {f}} T (\mathbf {u}, z) \sigma (\mathbf {r} (z)) \mathbf {c} (\mathbf {r} (z)) \delta z, \tag {2} +$$ + +where $T(\mathbf{u},z) = \exp (-\int_{z_n}^z\sigma (\mathbf{r}(z)))\delta z'$ denotes the accumulated transmittance value along the ray. We refer readers to [21, 25] for more details of volume rendering operation. In practice, Eq. (2) is approximated using $M$ points sampled along the ray at depth $\{z_i\}_{i = 1}^M$ , which produces radiance field outputs as $\{\mathbf{y}_i\}_{i = 1}^M$ . Denoting the ray compositing function in Eq. (2) as $g(.)\in \mathbb{R}^{4M}\to \mathbb{R}^3$ , we can rewrite $\mathcal{I}(\mathbf{u}) = g(\{\mathbf{y}_i\}_{i = 1}^M)$ . Finally, given a camera pose $P$ , we can then transform the ray $\mathbf{r}^{(C)}$ to the world coordinate to obtain $\mathbf{r}^{(W)}$ through a 3D rigid transformation $\mathcal{T}$ . The rendered image is then obtained as + +$$ +\hat {\mathcal {I}} (\mathbf {u}, \mathbf {p}) = g \left(\left\{f \left(\mathcal {T} \left(\mathbf {r} ^ {(C)} (z), P\right); \boldsymbol {\Theta} _ {r g b}\right) \right\} _ {i = 1} ^ {M}\right). \tag {3} +$$ + +Joint optimization of pose and NeRF Prior works [10, 23, 40, 42] demonstrate that it is feasible to optimize both camera pose and NeRF by minimizing $\mathcal{L}_{rgb}$ in Eq. (1). This is achieved by considering $P$ as estimizable parameters, see Sec. 3.1 for its parameterization. Consequently, ray $\mathbf{r}$ is now dependent on the camera pose. Mathematically, this joint optimization can be rewritten as + +$$ +\min _ {P, \boldsymbol {\Theta} _ {r g b}} \sum_ {t = 1} ^ {T} \sum_ {\mathbf {u} \in \mathbb {R} ^ {2}} \| \hat {\mathcal {I}} (\underbrace {\mathcal {T} \left(\mathbf {r} ^ {(C)} (z , \bar {\mathbf {u}}) ; P\right)} _ {G (.)}; \boldsymbol {\Theta} _ {r g b}) - \mathcal {I} _ {i} (\mathbf {u}) \| _ {2} ^ {2}. \tag {4} +$$ + +# 3.2 Invertible Neural Warp for Ray Transform + +We propose to overparameterize $P$ using an Invertible Neural Network (INN). There exist two options for parameterization: (i) use a separate INN for each camera $P_{t}$ ; (ii) use a single INN that is shared across all frames, coupled with a learnable code that is unique to frame $t$ . Drawing inspiration from the dynamic NeRF method used for representing deformation fields [6,31,38,39] and also considering parameter efficiency, we have chosen to pursue the latter strategy for our proposed pose overparameterization - a single, globally shared neural $\Theta_{\mathcal{W}}$ network across all frames, coupled with an estimizable per-frame latent code $\varPhi_t\in\mathbb{R}^D$ , see supp. (Sec. C) for the comparison of using multiple-INNs versus single-NN. Consequently, we can rewrite $G(.)$ in Eq. (6) as $h(\mathbf{r}^{(C)};\Theta_{\mathcal{W}},\varPhi_t)$ where $h(.):\mathbb{R}^{3 + D}\to \mathbb{R}^3$ . Fig. 2 presents our approach. In our approach, we + +![](images/f35034cff4e28a497495b3c340d53d7ccaf3b638b130c164c7fcc9106c443139.jpg) +Fig. 2: An overview of our INN-based approach, illustrated using two views $\mathcal{I}_1$ and $\mathcal{I}_2$ . INN which is denoted as $h_{\Theta_W}$ takes the pixel locations in the camera coordinate system $\mathbf{x}_{i,t}^{(C)}$ , along with the frame-dependent latent code $\varPhi_t$ , and output its corresponding location in the world coordinate system as $\mathbf{x}_{i,t}^{(W)}$ , see Sec. 3.2 for full details. + +![](images/a7753746da043b6f500c0a3eb653e59ff03a2aafdeee3d0f7323c78f9fa768ad.jpg) + +model each pixel in the camera coordinate system $\mathbf{x}_{i,t}^{(C)}$ as an individual ray. Our proposed INN is designed for transforming these rays from camera coordinate to the world coordinate. Specifically, our proposed INN takes in the pixel coordinates $\mathbf{x}_{i,t}^{(C)}$ and camera center $\mathbf{o}_t^{(C)}$ , both defined in the camera coordinates $t$ , coupled with the frame-dependent latent code and outputs their corresponding equivalent $\mathbf{x}_{i,t}^{(W)}$ and camera center $\mathbf{o}_t^{(W)}$ in the world coordinates. + +Rigidity prior In our formulation, each pixel is represented as an individual ray within the camera coordinate system. This inherently relaxes the rigidity constraint. As a result, the output from the INN does not necessarily conform to a global rigid motion. Given known camera-world correspondences $(\mathbf{x}_{i,t}^{(C)},\mathbf{x}_{i,t}^{(W)})$ we can solve a closed-form rigid registration problem to determine a global pose, which can be integrated into our optimization problem as a rigidity prior $\mathcal{L}_{rigid}$ + +$$ +\min _ {T ^ {*}} \sum_ {i = 1} ^ {L} \| \mathbf {x} _ {i, t} ^ {(C)} - T ^ {*} \circ \mathbf {x} _ {i, t} ^ {(W)} \| _ {2} ^ {2}. \tag {5} +$$ + +Final optimization problem We solve our final optimization problem as + +$$ +\min _ {\boldsymbol {\Phi} _ {t}, \boldsymbol {\Theta} _ {\mathcal {W}}, \boldsymbol {\Theta} _ {r g b}} \sum_ {t = 1} ^ {T} \sum_ {\mathbf {u} \in \mathbb {R} ^ {2}} \| \hat {\mathcal {I}} (h (\mathbf {r} ^ {(c)}; \boldsymbol {\Theta} _ {\mathcal {W}}, \boldsymbol {\Phi} _ {t}); \boldsymbol {\Theta} _ {r g b}) - \mathcal {I} _ {i} (\mathbf {u}) \| _ {2} ^ {2} + \lambda \mathcal {L} _ {r i g i d}. \tag {6} +$$ + +# 3.3 Advantages of INN for Overparameterizing Rigid Ray Warps + +BARF-based approach parameterizes camera pose $P$ of each frame using a $SE(3)$ (see Sec. 3.1), which guarantees that $\mathcal{T}$ is a bijective mapping. Therefore, when + +overparameterizing camera poses, it is crucial that the neural network adheres to the bijection property because this one-to-one correspondence ensures that there is a unique output in the world space for every point in the camera space. As we will demonstrate in Sec. 4.2, simply applying a rudimentary strategy (denoted as Naive) when overparameterizing the rigid warps of ray (camera-world) with an MLP often does not achieve convergence, see Fig. 1c and Tab. 1. Consequently, to attain this bijective property using an MLP, it is necessary to introduce an auxiliary network to model the backward warps (Implicit-Invertible MLP). While effective, it presents a significant drawback: it results in a twofold rise in the computational complexity due to the existence of the backward network to enforce the self-consistency. To mitigate this substantial increase in computational demands inherent in the modified MLP approach, we propose the use of INNs for parameterize these bijections. INN implements the bijective mappings by composing affine transformations into several blocks. Within each block, the input coordinates are divided into two segments; the first part remains constant and is used to parameterize the transformation that is applied to the second part [6]. Besides their inherent invertibility, INN also offers the advantage of homeomorphic property, which potentially facilitate a more flexible optimization trajectory that is less susceptible to a suboptimal minimum trajectory, see Sec. 4.5 + +# 4 Experiments + +# 4.1Baselines + +We compare our approach with two representative methods in pose-NeRF joint optimization: the standard global $SE3$ -approach BARF [23], and the overparameterized representation L2G [9]. For all experiments, we use the original implementations including their default settings for coarse-to-fine scheduling, architecture and hyperparameters, see supp. (Sec. A) for more details. Additionally, in our 2D planar experiments, we include a comparison with another two variants called the Naive (MLP) and Implicit-Invertible MLP to execute ray transform. This comparison is specifically designed to highlight the significance of invertibility when employing MLPs for executing ray transformation. + +Local-to-global (L2G) [9] L2G uses an MLP to predict rigid SE3 transformation for each ray. These predicted transformation parameters are then used to analytically estimate the transformed coordinates in the world space. + +Naive This is the simplest version of our baseline that uses one primary network, denoted as $h_{fwd}:(\mathbf{x}^{(C)},\varPhi_t)\to \mathbf{x}^{(W)}$ , to learn the forward mapping which takes in the coordinates from the camera space (coupled with a per-frame latent code), and outputs the corresponding coordinates in the world space. + +Implicit-Invertible MLP We use two networks $h_{fwd}$ and $h_{bwd}$ to enforce approximate invertibility. Alongside primary network $h_{fwd}$ , we use a secondary network $h_{bwd} : (\mathbf{x}^{(W)}, \varPhi_t) \to \hat{\mathbf{x}}^{(C)}$ to invert the outputs from the primary network $h_{fwd}$ . To minimize deviations from bijections, we introduce a regularization term $\mathcal{L}_{\text{implicit}}$ as $\| \mathbf{x}^{(C)} - \hat{\mathbf{x}}^{(C)} \|_2^2$ into the optimization problem Eq. (6). + +Explicit-Invertible INN (Ours) We explicitly model inversions in the neural network architecture by formally learning an INN. We have chosen to utilize architecture proposed by NDR-INN [6], see supp. (Sec. A.1) for the architecture details that we use for all our experiments. + +# 4.2 2D Planar Neural Image Alignment + +Following BARF [9, 10, 23], we learn a 2D neural image field, for creating a homography-based panoramic image from $N$ patches cropped from the original image, each generated with random homography perturbations. Specifically, we learn a 2D coordinate network $f(\Theta_{rgb})$ to render the stitched image. Each pixel in the $N$ training patch is warped using the estimated homography $H$ to create the rendered image. We choose the "cat" image from ImageNet [11]. We initialize patch warps as identity and fix the gauge freedom by anchoring the first warp to align the neural image to the original image [9, 10, 23]. We randomly generate 20 different homography instances, with scale-noise parameter 0.1 and 0.2 for homography and translation, respectively. We solve Eq. (5) for the homography using a Direct Linear Transform (DLT) solver3. + +Experiment settings. We evaluate our proposed method against BARF [23], and our three overparameterized network variants: Naive, Implicit- and Explicit-Invertible MLPs, as detailed in Sec. 4.1. For both naive and Implicit-Invertible MLPs, we utilized a Leaky-ReLU MLP with five 256-dimensional hidden units, and a 16-dimensional latent code to represent the frame-dependent embeddings $\phi_t$ . We also follow the default coarse-to-fine scheduling established by BARF. In the case of Implicit-Invertible MLP, we use the same architecture both $h_{fwd}$ and $h_{bwd}$ . We use the Adam optimizer [19] to optimize for both the network weights $\Theta_{rgb}$ and $\Theta_{\mathcal{W}}$ . We set the learning rate for both $\Theta_{rgb}$ and $\Theta_{\mathcal{W}}$ at $1\times 10^{-3}$ , with both decaying exponentially to $1\times 10^{-4}$ and $1\times 10^{-5}$ , respectively. We set the weighting term for $\mathcal{L}_{rigid}$ for both overparameterized MLPs to $1\times 10^{2}$ and the consistency term $\mathcal{L}_{implicit}$ for Implicit-Invertible MLP to $1\times 10^{1}$ . + +Robustness to noise perturbations Fig. 3 analyzes the robustness of BARF and our approach under different noise perturbations across 20 different homography instances. Each run began with the initialization of the homography using the "groundtruth", and noise perturbations were gradually introduced, ranging from 0 to 0.3 to the translation component of the homography. Fig. 3 indicates that our representation exhibits a higher tolerance to noise compared to BARF. + +Results. Tab. 1 summarizes the statistical results from 20 runs, where we report the warp error and patch reconstruction error. We quantify the warp error in terms of corner error, defined as the L2 distance between the groundtruth corner position and estimated corner position, and PSNR as the metric to assess the reconstruction quality. We used 5-pixel threshold to define the success + +![](images/4b17d1eae930c5bc8390a47b0275281b6897058455224ce79820848965ed1f60.jpg) +Fig. 3: Basin convergence analysis of our approach versus BARF in a 2D planar experiment. On the left, we show the success rate of 20 runs, where we initialize the homography using groundtruth, and gradually introduced noise perturbations to the translation component. The noise scale is varied from 0 to 0.30. We used 5-pixel threshold to determine success convergence. Notably, our approach (blue) demonstrates higher noise tolerance compared to BARF (red). On the right, we show a qualitative comparison when both methods are perturbed with the highest magnitude of noise. + +![](images/aa5ec814e80ca259bc42627f6d513d55dd5f2185e8655068112a276be203ce23.jpg) + +Table 1: Statistical result for 20 homography runs, with scale noise of 0.1 for homography and 0.2 for translation. The warp error is quantified in terms of corner error, and the patch reconstruction error in measured in PSNR. We provide mean and std dev for the evaluation. We used 5-pixel threshold to determine success convergence. + +
Corner error (px) ↓Patch PSNR ↑Success rate ↑ +(Upper bound:1.00)
MeanStd. dev.MeanStd. dev.
BARF [23]29.6328.1828.944.380.30
Naive85.5930.3125.862.070.00
Implicit-Invertible13.9222.9333.703.930.65
Explicit-Invertible (INN)4.706.4734.712.370.75
+ +convergence. Fig. 1 presents a qualitative result for a homography instance. By enforcing approximate invertibility, the Implicit-Invertible MLP demonstrates a significantly higher rate of successful convergence compared to the naive version of MLP, with success rates improved by $65\%$ . We further show that by explicitly injecting invertibility into the architecture (Ours), the success rate is increased to $75\%$ . This result reinforces that invertibility is crucial when learning rigid warp functions via overparameterization, and using an architecture that guarantees bijective property is effective in ensuring the pose converge to an optimal solution during joint optimization in practice. Henceforth, we will focus exclusively on using our INN-based approach for subsequent results. For completeness, we also compare with another setup, where each frame is parameterized with a neural network, see supp. (Sec. B.1). Interestingly, apart from the parameter efficiency, we find that using a single global neural network for all frames is sufficient to converge to a good pose solutions. We hypothesized that the difference may be attributed to the benefits of gradient sharing in the shared neural network setup. This result is aligned with the findings by Bian et al. [4]. Consequently, we adhere to the design where we use one single INN shared across all the frames, coupled with a frame-specific latent code for the rest of our experiments. + +![](images/a2ebeb77bc58d43a7af206c0b50a570e50866057e1703f7bb6553d1cb219cd1e.jpg) +(a) BARF [23] + +![](images/3c8165a25bfd71dcc17d991a2e70cc63ad0b0ad3a17c0365743616f4b35d214f.jpg) +(b) L2G [9] +Fig. 4: Qualitative analysis of reconstruction error on leaves (top) and trex (bottom). We present the average image reconstruction error through insets. Our approach presents the lowest misalignment error, as indicated darker areas in the error map. + +![](images/df35eb51c9f26f45c51f65301be34f41a44855fe4a9a341ed753f1c041b842e6.jpg) +(c) INN(Ours) + +# 4.3 Neural Radiance Fields (NeRF) + +In this section, we compare our proposed representation with BARF [23] and L2G [9]. We assume known intrinsic for all methods. We perform our experiments on both the LLFF [24] (Sec. 4.4), DTU [15] (Sec. 4.6) as well as Blender datasets in supp. (Sec. B2). We solve Eq. (5) for the global rigid SE3 transformation using Umayama algorithm $^4$ . + +Evaluation metrics. For pose estimation, we report the accuracy of the poses after globally aligning the optimized poses to the groundtruth [9,10,23,37]. We assess view synthesis using PSNR, SSIM and LPIPS. A standard procedure in view synthesis evaluation involves performing test-time photometric optimization on the trained models. This additional step is intended to factor out the pose errors, which may otherwise compromise the quality of the synthesized views [9,10,23,36,37]. This process is akin to a pose refinement method which minimises the photometric error on the synthesized image while keeping the trained NeRF model fixed. However, it is important to recognize that this pose correction may not accurately represent the initial accuracy of the methods in terms of pose estimation for view synthesis. Therefore, we opt to report the view synthesis quality both before and after the pose refinement step. On DTU, we extend our evaluation to include comparisons of rendered depth with ground-truth depth using mean depth absolute error, as well as reconstruction accuracy using Chamfer distance. For the Chamfer evaluation, we utilize the optimized poses estimated from all methods and employ a neural surface reconstruction algorithm called Voxurf [41] for geometry reconstruction. Further details on pose alignment and metrics computations in the supp. (Sec. A). + +# 4.4 Forward Facing Scenes: LLFF + +Experiment settings. The standard LLFF benchmark dataset [24] consists of eight real-world, forward-facing scenes captured using hand-held cameras. + +Table 2: Evaluation of the LLFF dataset [24] using initial identity poses. We compute results for BARF [23] and L2G [9] using their codebases. The results are averaged over all eight scenes, see supp. (Sec. B) for a breakdown of results per scene. + +
Pose accuracyNovel view synthesis
Rotation (°)Translation (×100)Before test-timeAfter test-time
PSNRSSIMLPIPSPSNRSSIMLPIPS
BARF [23]0.900.4017.000.410.3223.820.720.24
L2G [9]0.480.3017.990.460.2624.350.750.21
INN(Ours)0.310.2419.310.520.2524.280.740.22
+ +Following [10, 23, 37], we initialize all camera poses with the identity transformations for all the methods. We employ the same training and testing split as in BARF [23]. We use the evaluation metrics described in Sec. 4.3. + +Implementation details. We train all methods for 200k iterations and randomly sample 2048 pixel rays at each optimization step [9, 10, 23]. We train without hierarchical sampling. We set the learning for $\Theta_{rgb}$ to be $1\times 10^{-3}$ decaying to $3\times 10^{-4}$ , and $5\times 10^{-4}$ for $\Theta_{\mathcal{W}}$ decaying to $1\times 10^{-6}$ . We also follow the default coarse-to-fine scheduling by BARF [23], see supp. for full details. + +Results. Our approach achieves a substantial reduction of rotation errors (70% vs. BARF and 35% vs. L2G) and translation errors (50% vs. BARF and 20% vs. L2G). Additionally, the superior performance of both our method and L2G over BARF further highlight the merits of overparameterization for simultaneous pose and neural field estimation. This significant improvement in camera pose accuracy directly enhances the performance of view synthesis, as evident in Tab. 2. In particular, instances such as trex and leaves observe a significant improvement. Fig. 4 illustrates the absolute error between the original image and the rendered image on these two instances. Our approach presents the lowest misalignment error, as indicated darker areas in the error map. It is important to note, however, while L2G and our approach appear comparable after test-time optimization, this refinement procedure has mitigated the pose estimation noise to improve PSNR, as discussed in Sec. 4.3. For more qualitative results and ablation studies, please refer to supp. (Sec. B). + +# 4.5 Homeomorphism perspective: A qualitative analysis + +We present a qualitative analysis on a single-view pose estimation that sheds light on the empirical effectiveness of our approach compared to the L2G method [9]. Our method leverages the concept of an INN, which predict homeomorphisms—continuous, invertible transformations not limited to the rigid motions of the $SE(3)$ group. Unlike the L2G method, which constrains pose estimation within the rigid bounds of $SE(3)$ , our INN-based approach embraces a broader spectrum of transformations. This grants the optimization process a higher degree of flexibility, allowing for a diverse range of optimization paths and facilitating a smoother trajectory towards the solution. To empirically validate our + +![](images/de9791282530dc7f2aa2d00955f63115a77d431ac92212169c6d571e7c0c7744.jpg) +iter=0 +Fig. 5: Qualitative analysis of intermediate rendered image compared to superimposed Groundtruth (lighter visualization) when using L2G [9] and our approach on single-view pose estimation. Notably, when using our approach for pose estimation, we observe noticeable deformation in the rendered scene depicted in the bottom row (zoom in for better view). These deformations indicate that at each iteration, the INN predicts general homeomorphisms that are not rigid transformation. Thus yielding a flexible optimization trajectory that does not land in a suboptimal minimum trajectory. + +![](images/d36029cc3d3a8d6d36357dcc155cad1b619265218fd889bb819da70e0a464f18.jpg) +iter=40 + +![](images/3888edd3b5a9d0af273d3b0f90dabbb8c7db78aedc858fe07cf7fdac739c7fca.jpg) +iter=80 + +![](images/54efe1a573ac49bca4d09c0b97f515f08a3fcf6d8eb19eb754df9eff17e9c4e2.jpg) +iter=500 + +Table 3: Absolute pose accuracy evaluation of the DTU dataset [15] using initial noisy poses that corresponds to an average rotation and translation error of $15^{\circ}$ and 70 respectively. The upper section shows rotation errors in degrees, while the lower section displays translation errors $(\times 100)$ . red box denotes the best result. + +
MethodsMetricScan IDsInitial rotation err: 15°, translation err (×100) : 70
243740556365698397105106110114118mean
BARFrotation1.041.880.552.850.430.783.880.542.203.5611.812.011.212.602.52
L2G0.903.070.5710.810.740.941.550.590.642.8513.155.922.9012.474.08
INN(Ours)0.210.480.731.640.400.622.750.550.540.303.892.911.060.251.17
BARFtranslation3.004.082.0911.631.432.7812.802.136.1212.4025.885.524.085.047.07
L2G2.408.802.1332.742.783.485.082.312.375.3436.2014.018.6437.1611.67
INN(Ours)0.501.010.975.240.960.785.300.671.190.439.9611.523.780.643.07
+ +hypothesis, we conducted experiments using a trained NeRF model to estimate the camera pose relative to a 3D scene, aiming to minimize the photometric error between NeRF-rendered and actual observed images. Despite starting from the same initial pose (off by $20^{\circ}$ ), and employing a random sampling of 2048 rays per iteration for all methods, our INN approach outperformed the L2G method. The L2G method often converged to suboptimal poses, whereas the INN method achieved accurate pose estimation, as evidenced in Fig. 5 where the NeRF-rendered images is well-aligned with the groundtruth. This success can be attributed to the INN's ability to predict a general homeomorphism, a transformation more general than a rigid transformation, which offers a significant advantage in navigating the optimization landscape more effectively and avoiding suboptimal local minima. + +# 4.6 $360^{\circ}$ Scenes: DTU + +Experimental settings. We evaluated on 14 test scenes from DTU [4]. Following [9,23,37], we synthetically perturb the ground-truth camera poses with + +Table 4: Geometry evaluation of the DTU dataset [15], using initial noisy poses that corresponds to an average rotation and translation error of $15^{\circ}$ and 70 respectively. The upper section presents mean absolute depth, while the lower section displays reconstruction error measured by Chamfer distance. red box denotes the best result. + +
Scan IDs | Initial rotation err: 15°, translation err (×100) : 70
PoseMetric243740556365698397105106110114118mean
BARF0.110.150.110.120.140.100.260.170.170.460.530.170.100.170.20
L2Gdepth ↓0.080.240.120.380.170.120.110.180.100.180.240.170.120.840.22
INN(Ours)(abs)0.060.140.130.120.150.080.180.170.110.110.130.300.080.060.13
BARF5.474.074.397.644.835.286.435.748.067.217.758.298.245.476.35
L2GChamfer ↓4.379.084.416.678.035.923.806.006.305.597.437.358.298.166.53
INN(Ours)1.562.255.555.684.584.275.515.695.093.738.157.857.351.144.89
+ +$15\%$ of additive Gaussian noise, which corresponds to an average rotation and translation error of $15^{\circ}$ and 70, respectively. For a fair comparison, we used the same initialization for all methods. We refer the readers to supp. (Sec B.5) for the results with Colmap initialization. + +Implementation details. As BARF [23] and L2G [9] have not tested on DTU datasets, given that the DTU dataset encompasses $360^{\circ}$ scenes which is similar to the original Blender dataset, we adopted the original hyperparameters used by the author for training BARF and L2G on the blender dataset. Following Bian et al. [4], we multiply the output of their local warp network by a small factor which is $\alpha = 0.01$ for L2G [9]. For our approach, we set the learning rate for $\Theta_{rgb}$ to be $1\times 10^{-3}$ decaying to $1\times 10^{-4}$ , and $\Theta_{\mathcal{W}}$ to start from $5\times 10^{-4}$ and decaying to $1\times 10^{-8}$ . We use the default coarse-to-fine scheduling by BARF [23]. + +Results. As demonstrated in Tab. 3, we outperform all baselines by a considerable margin across majority of the sequences in pose accuracy. Overall, our approach achieves approximately a $50\%$ improvement in rotation and a $60\%$ improvement in translation over BARF. When compared to L2G, our method shows a $70\%$ increase in accuracy for both rotation and translation. Additionally, our approach also consistently surpasses both baselines in geometry evaluation, as evidenced in the depth and reconstruction error in Tab. 4 and qualitative results in Fig. 6. For the quantitative results for novel view synthesis and additional qualitative results, we refer the readers to see supp. (Sec B). + +# 5 Conclusion + +In this paper, we examine the benefits of overparameterizing poses via a MLP in the joint optimization task of camera pose and NeRF. We establish that invertibility is a crucial property. We further show that using an Invertible Neural Network, inherently equipped with a guaranteed bijection property, significantly improves the convergence in pose optimization compared to existing representative methods. + +![](images/9ad49b1e37e52ad4388448c6cba6d2879970f095265cf737a94eb56f5e264925.jpg) +Fig. 6: We showcase the reconstruction results for scan24, scan55, scan105 and scan106. All meshes are generated using a neural surface reconstruction algorithm called Voxurf [41] using the poses estimated by different approaches. + +Acknowledgement We thank Chee-Kheng (CK) Chng for insightful discussions and technical feedback. + +# References + +1. Allen-Zhu, Z., Li, Y., Liang, Y.: Learning and generalization in overparameterized neural networks, going beyond two layers. Advances in neural information processing systems 32 (2019) 2, 3 +2. Arora, S., Du, S., Hu, W., Li, Z., Wang, R.: Fine-grained analysis of optimization and generalization for overparameterized two-layer neural networks. In: International Conference on Machine Learning. pp. 322-332. PMLR (2019) 3 +3. Behrmann, J., Grathwohl, W., Chen, R.T., Duvenaud, D., Jacobsen, J.H.: Invertible residual networks. In: International conference on machine learning. pp. 573-582. PMLR (2019) 4 +4. Bian, J.W., Bian, W., Prisacariu, V.A., Torr, P.: Porf: Pose residual field for accurate neural surface reconstruction. arXiv preprint arXiv:2310.07449 (2023) 3, 9, 12, 13 +5. Bian, W., Wang, Z., Li, K., Bian, J.W., Prisacariu, V.A.: Nope-nerf: Optimising neural radiance field with no pose prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4160-4169 (2023) 3 +6. Cai, H., Feng, W., Feng, X., Wang, Y., Zhang, J.: Neural surface reconstruction of dynamic scenes with monocular rgb-d camera. Advances in Neural Information Processing Systems 35, 967-981 (2022) 4, 5, 7, 8 +7. Chen, R.T., Rubanova, Y., Bettencourt, J., Duvenaud, D.K.: Neural ordinary differential equations. Advances in neural information processing systems 31 (2018) 4 +8. Chen, Y., Lee, G.H.: Dbarf: Deep bundle-adjusting generalizable neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24-34 (2023) 3 +9. Chen, Y., Chen, X., Wang, X., Zhang, Q., Guo, Y., Shan, Y., Wang, F.: Local-to-global registration for bundle-adjusting neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8264-8273 (2023) 1, 2, 3, 7, 8, 10, 11, 12, 13, 14 +0. Chng, S.F., Ramasinghe, S., Sherrah, J., Lucey, S.: Gaussian activated neural radiance fields for high fidelity reconstruction and pose estimation. In: European Conference on Computer Vision. pp. 264-280. Springer (2022) 1, 3, 5, 8, 10, 11 +1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) 8 +2. Dinh, L., Sohl-Dickstein, J., Bengio, S.: Density estimation using real nvp. arXiv preprint arXiv:1605.08803 (2016) 4 +3. Guo, M., Fathi, A., Wu, J., Funkhouser, T.: Object-centric neural scene rendering. arXiv preprint arXiv:2012.08503 (2020) 1 +4. Jeffreys, H., Jeffreys, B.: Methods of mathematical physics. Cambridge university press (1999) 2 +5. Jensen, R., Dahl, A., Vogiatzis, G., Tola, E., Aanaes, H.: Large scale multi-view stereopsis evaluation. In: 2014 IEEE Conference on Computer Vision and Pattern Recognition. pp. 406-413. IEEE (2014) 10, 12, 13 + +16. Jeong, Y., Ahn, S., Choy, C., Anandkumar, A., Cho, M., Park, J.: Self-calibrating neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5846-5854 (2021) 3 +17. Jiang, C., Huang, J., Tagliasacchi, A., Guibas, L.J.: Shapeflow: Learnable deformation flows among 3d shapes. Advances in Neural Information Processing Systems 33, 9745-9757 (2020) 4 +18. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: Lerf: Language embedded radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19729-19739 (2023) 1 +19. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 8 +20. Lei, J., Daniilidis, K.: Cadex: Learning canonical deformation coordinate space for dynamic surface representation via neural homeomorphism. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6624-6634 (2022) 4 +21. Levoy, M.: Efficient ray tracing of volume data. ACM Transactions on Graphics (TOG) 9(3), 245-261 (1990) 5 +22. Li, Y., Liang, Y.: Learning overparameterized neural networks via stochastic gradient descent on structured data. Advances in neural information processing systems 31 (2018) 2, 3 +23. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: Barf: Bundle-adjusting neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5741-5751 (2021) 1, 2, 3, 4, 5, 7, 8, 9, 10, 11, 12, 13, 14 +24. Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG) 38(4), 1-14 (2019) 10, 11 +25. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) 1, 2, 4, 5 +26. Neyshabur, B., Tomioka, R., Srebro, N.: Norm-based capacity control in neural networks. In: Conference on learning theory. pp. 1376-1401. PMLR (2015) 2, 3 +27. Nguyen, Q.N., Mondelli, M.: Global convergence of deep networks with one wide layer followed by pyramidal topology. Advances in Neural Information Processing Systems 33, 11961-11972 (2020) 2, 3 +28. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Occupancy flow: 4d reconstruction by learning particle dynamics. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5379-5389 (2019) 4 +29. Oymak, S., Soltanolkotabi, M.: Toward moderate overparameterization: Global convergence guarantees for training shallow neural networks. IEEE Journal on Selected Areas in Information Theory 1(1), 84-105 (2020) 3 +30. Park, K., Henzler, P., Mildenhall, B., Barron, J.T., Martin-Brualla, R.: Camp: Camera preconditioning for neural radiance fields. ACM Transactions on Graphics (TOG) 42(6), 1-11 (2023) 3 +31. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021) 1, 5 +32. Parra, A., Chng, S.F., Chin, T.J., Eriksson, A., Reid, I.: Rotation coordinate descent for fast globally optimal rotation averaging. In: Proceedings of the + +IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4298-4307 (2021) 1 +33. Paschalidou, D., Katharopoulos, A., Geiger, A., Fidler, S.: Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3204-3215 (2021) 4 +34. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 1 +35. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014) 2, 3 +36. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., et al.: Nerfstudio: A modular framework for neural radiance field development. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-12 (2023) 10 +37. Truong, P., Rakotosaona, M.J., Manhardt, F., Tombari, F.: Sparf: Neural radiance fields from sparse and noisy poses. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4190-4200 (2023) 3, 10, 11, 12 +38. Wang, C., MacDonald, L.E., Jeni, L.A., Lucey, S.: Flow supervision for deformable nef. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21128-21137 (2023) 5 +39. Wang, Q., Chang, Y.Y., Cai, R., Li, Z., Hariharan, B., Holynski, A., Snavely, N.: Tracking everything everywhere all at once. arXiv preprint arXiv:2306.05422 (2023) 4, 5 +40. Wang, Z., Wu, S., Xie, W., Chen, M., Prisacariu, V.A.: Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064 (2021) 1, 3, 5 +41. Wu, T., Wang, J., Pan, X., Xu, X., Theobalt, C., Liu, Z., Lin, D.: Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv preprint arXiv:2208.12697 (2022) 10, 14 +42. Xia, Y., Tang, H., Timofte, R., Van Gool, L.: Sinerf: Sinusoidal neural radiance fields for joint pose estimation and scene reconstruction. arXiv preprint arXiv:2210.04553 (2022) 3, 5 +43. Xu, C., Wu, B., Hou, J., Tsai, S., Li, R., Wang, J., Zhan, W., He, Z., Vajda, P., Keutzer, K., et al.: Nerf-det: Learning geometry-aware volumetric representation for multi-view 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23320-23330 (2023) 1 +44. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5438-5448 (2022) 1 +45. Yang, G., Belongie, S., Hariharan, B., Koltun, V.: Geometry processing with neural fields. Advances in Neural Information Processing Systems 34, 22483-22497 (2021) 4 +46. Zhan, H., Zheng, J., Xu, Y., Reid, I., Rezatofighi, H.: Activermap: Radiance field for active mapping and planning. arXiv preprint arXiv:2211.12656 (2022) 1 +47. Zhao, F., Yang, W., Zhang, J., Lin, P., Zhang, Y., Yu, J., Xu, L.: Humannerf: Efficiently generated human radiance field from sparse inputs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7743-7753 (2022) 1 + +48. Zhu, Z., Peng, S., Larsson, V., Xu, W., Bao, H., Cui, Z., Oswald, M.R., Pollefeys, M.: Nice-slam: Neural implicit scalable encoding for slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12786-12796 (2022) 1 \ No newline at end of file diff --git a/2024/Invertible Neural Warp for NeRF/images.zip b/2024/Invertible Neural Warp for NeRF/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..39e958be11b418627bba32d7e3bf6a2504de7ae2 --- /dev/null +++ b/2024/Invertible Neural Warp for NeRF/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19bee814615149d8d5ad42cd245aeffa840779332ee14e2f9f66be933988395c +size 656515 diff --git a/2024/Invertible Neural Warp for NeRF/layout.json b/2024/Invertible Neural Warp for NeRF/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c37b6efb46d5506e3bf1a2334fb171c8fbc80427 --- /dev/null +++ b/2024/Invertible Neural Warp for NeRF/layout.json @@ -0,0 +1,10415 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 189, + 112, + 424, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 112, + 424, + 129 + ], + "spans": [ + { + "bbox": [ + 189, + 112, + 424, + 129 + ], + "type": "text", + "content": "Invertible Neural Warp for NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 150, + 466, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 150, + 466, + 163 + ], + "spans": [ + { + "bbox": [ + 146, + 150, + 466, + 163 + ], + "type": "text", + "content": "Shin-Fang Chng, Ravi Garg, Hemanth Saratchandran, and Simon Lucey" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 189, + 172, + 424, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 172, + 424, + 217 + ], + "spans": [ + { + "bbox": [ + 189, + 172, + 424, + 217 + ], + "type": "text", + "content": "Adelaide University \nAustralian Institute for Machine Learning \nshinfang.chng@adelaide.edu.au \nhttps://sfchng.github.io/ineurowarping-github.io/" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 243, + 455, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 243, + 455, + 387 + ], + "spans": [ + { + "bbox": [ + 160, + 243, + 455, + 387 + ], + "type": "text", + "content": "Abstract. This paper tackles the simultaneous optimization of pose and Neural Radiance Fields (NeRF). Departing from the conventional practice of using explicit global representations for camera pose, we propose a novel overparameterized representation that models camera poses as learnable rigid warp functions. We establish that modeling the rigid warps must be tightly coupled with constraints and regularization imposed. Specifically, we highlight the critical importance of enforcing invertibility when learning rigid warp functions via neural network and propose the use of an Invertible Neural Network (INN) coupled with a geometry-informed constraint for this purpose. We present results on synthetic and real-world datasets, and demonstrate that our approach outperforms existing baselines in terms of pose estimation and high-fidelity reconstruction due to enhanced optimization convergence." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 396, + 453, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 396, + 453, + 418 + ], + "spans": [ + { + "bbox": [ + 160, + 396, + 453, + 418 + ], + "type": "text", + "content": "Keywords: Neural Radiance Fields " + }, + { + "bbox": [ + 160, + 396, + 453, + 418 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 396, + 453, + 418 + ], + "type": "text", + "content": " Joint scene reconstruction and pose estimation " + }, + { + "bbox": [ + 160, + 396, + 453, + 418 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 396, + 453, + 418 + ], + "type": "text", + "content": " Implicit Neural Representations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 437, + 230, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 230, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 230, + 451 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 481, + 545 + ], + "type": "text", + "content": "NeRF [25] has recently emerged as a compelling approach for synthesizing photorealistic images from novel views. NeRF employs a multi-layer perceptron (MLP) to model a volumetric representation of a 3D scene. It operates by minimizing the photometric loss, which is the discrepancy between rendered images and actual images. NeRF's ability to reconstruct high-fidelity signals, coupled with its memory efficiency, has propelled its adoption across a wide array of applications [13,18,31,43,44,46-48], demonstrating its significant impact and versatility." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "One of the primary challenges with NeRF is the requirement for precisely known camera poses for each captured image. To address this challenge, several approaches such as BARF [23], NeRFmm [40], and GARF [10] have been developed. These methods facilitate the simultaneous optimization of the NeRF and the camera poses, using a compact, six-dimensional vector to represent the camera poses efficiently. However, this compact parameterization, while prevalent in contemporary structure from motion (SfM) literature [32, 34], has been shown to struggle with poor basin of convergence when solved simultaneously with a NeRF [9]. Drawing wisdom from machine learning, where overparameterization has been recognized as a catalyst for enhanced optimization convergence" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 206, + 209 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 206, + 209 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 206, + 209 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 206, + 209 + ], + "type": "image", + "image_path": "8b32bd0a12042b8d9fae560350aa5178f27c48774b4a6761d8e6b5d1b70d6279.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 210, + 205, + 220 + ], + "lines": [ + { + "bbox": [ + 145, + 210, + 205, + 220 + ], + "spans": [ + { + "bbox": [ + 145, + 210, + 205, + 220 + ], + "type": "text", + "content": "(a) Groundtruth" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 209, + 114, + 272, + 209 + ], + "blocks": [ + { + "bbox": [ + 209, + 114, + 272, + 209 + ], + "lines": [ + { + "bbox": [ + 209, + 114, + 272, + 209 + ], + "spans": [ + { + "bbox": [ + 209, + 114, + 272, + 209 + ], + "type": "image", + "image_path": "26f48b667a3667385eafecf1a2bf9ae4d239d497f973548c9616f6bb599ef006.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 210, + 266, + 220 + ], + "lines": [ + { + "bbox": [ + 214, + 210, + 266, + 220 + ], + "spans": [ + { + "bbox": [ + 214, + 210, + 266, + 220 + ], + "type": "text", + "content": "(b) BARF [23]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 274, + 114, + 337, + 209 + ], + "blocks": [ + { + "bbox": [ + 274, + 114, + 337, + 209 + ], + "lines": [ + { + "bbox": [ + 274, + 114, + 337, + 209 + ], + "spans": [ + { + "bbox": [ + 274, + 114, + 337, + 209 + ], + "type": "image", + "image_path": "7052df06016205a2920c2f73d2037a89586beab5590f896a588aa17ce54b2b87.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 210, + 323, + 220 + ], + "lines": [ + { + "bbox": [ + 288, + 210, + 323, + 220 + ], + "spans": [ + { + "bbox": [ + 288, + 210, + 323, + 220 + ], + "type": "text", + "content": "(c) Naive" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 228, + 481, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 481, + 304 + ], + "type": "text", + "content": "Fig. 1: We investigate how overparameterizing rigid warps of rays with an MLP benefits the joint optimization task of camera pose and NeRF. This example estimates the warps that align the color-coded patches in Fig. 1a while solving for the neural field. Unlike \"BARF\" and \"naive\" MLP pose overparameterization methods fail catastrophically, enforcing invertibility, either implicitly (Fig. 1d) or explicitly (Fig. 1e) significantly improves warp estimation, see Sec. 4.1 for details of each method. We establish that invertibility is a crucial for MLP-based rigid warp representation." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 339, + 114, + 402, + 209 + ], + "blocks": [ + { + "bbox": [ + 339, + 114, + 402, + 209 + ], + "lines": [ + { + "bbox": [ + 339, + 114, + 402, + 209 + ], + "spans": [ + { + "bbox": [ + 339, + 114, + 402, + 209 + ], + "type": "image", + "image_path": "4f7cbcd4c74cd780f0453a70eff3015258797719adcd6e8ec880f99e7851d94c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 210, + 399, + 227 + ], + "lines": [ + { + "bbox": [ + 345, + 210, + 399, + 227 + ], + "spans": [ + { + "bbox": [ + 345, + 210, + 399, + 227 + ], + "type": "text", + "content": "(d) Implicit-Invertible MLP" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 405, + 114, + 468, + 209 + ], + "blocks": [ + { + "bbox": [ + 405, + 114, + 468, + 209 + ], + "lines": [ + { + "bbox": [ + 405, + 114, + 468, + 209 + ], + "spans": [ + { + "bbox": [ + 405, + 114, + 468, + 209 + ], + "type": "image", + "image_path": "ac5acfcb88a1cf34ab4ae73180940a053909240fa69adbfaf47ce5e20443b197.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 210, + 462, + 227 + ], + "lines": [ + { + "bbox": [ + 410, + 210, + 462, + 227 + ], + "spans": [ + { + "bbox": [ + 410, + 210, + 462, + 227 + ], + "type": "text", + "content": "(e) Explicit-Invertible INN" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 321, + 481, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 481, + 344 + ], + "type": "text", + "content": "in modern deep neural networks [22, 26, 35], this paper explores the potential of pose overparameterization for simultaneous pose and neural field estimation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 346, + 482, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 490 + ], + "type": "text", + "content": "Our approach: In traditional NeRF setups, accurately known extrinsic camera pose, comprising of a global rotation and translation for each image, are used to explicitly map pixel coordinates and the camera center to determine the viewing rays in a global world coordinate system [25]. Following the warping operation, the colors and volume densities along each ray in the world coordinate space is manipulated individually through a photometric loss function. In this paper, we explore scenarios where camera poses are not known. Specifically, we propose using a neural network to model the rigid warp function of ray. While it may seem counterintuitive to replace a succinct pose function with a more complex MLP, we argue that the enhanced convergence properties of such overparameterization [1, 26, 27] - in conjunction with the right constraint and prior - outweigh the increased functional flexibility." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 491, + 482, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 611 + ], + "type": "text", + "content": "Additionally, we highlight the critical role of enforcing invertibility when learning rigid warps using an MLP. To achieve an approximate bijective solution, one remedy is to use an auxiliary network to represent the backward warp; however this will introduce computational overhead. To this end, we propose explicitly modeling inversions in the neural network architectures, formally learning an Invertible Neural Network (INN). Our results demonstrate that opting for an architecture that is explicitly invertible is more effective for jointly optimizing both the pose and radiance field, outperforming existing strong baselines [9,23]. Notably, our INN-based approach achieves an improvement of over " + }, + { + "bbox": [ + 130, + 491, + 482, + 611 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 491, + 482, + 611 + ], + "type": "text", + "content": " in pose accuracy when compared to the standard SE3 parameterization [23]." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "text", + "content": "Shin-Fang et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "text", + "content": "1 Our use of invertibility strictly adheres to the well-established mathematical definition. Let " + }, + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "text", + "content": " be a function whose domain is " + }, + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "text", + "content": " and codomain is " + }, + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "text", + "content": " is invertible iff there exists a function " + }, + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "inline_equation", + "content": "g(f(x)) = x \\forall x \\in \\mathcal{X}" + }, + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "inline_equation", + "content": "f(g(y)) = y \\forall y \\in \\mathcal{Y}" + }, + { + "bbox": [ + 133, + 620, + 482, + 666 + ], + "type": "text", + "content": " [14]. We use bijective and invertible interchangeably throughout our paper." + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 242, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 242, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 242, + 127 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 148, + 324, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 148, + 324, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 148, + 324, + 159 + ], + "type": "text", + "content": "2.1 Joint NeRF and pose estimation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 175, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 175, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 175, + 482, + 475 + ], + "type": "text", + "content": "Despite NeRF demonstrating compelling results in novel view synthesis, NeRF requires accurate camera poses. The differentiable nature of volume rendering used in NeRF facilitates the backpropagation through the scene representation to update the camera poses. NeRFmm [40] demonstrates the possibility of optimizing camera poses within the NeRF framework. BARF [23] introduces a coarse-to-fine positional encoding scheduling to improve the joint optimization of NeRF and camera poses, and remains a widely adopted method. GARF [10] and SiNeRF [42] advocate for leveraging the smoothness inherent in non-traditional activations to mitigate the noisy gradients due to high frequencies in positional embeddings. In contrast, NoPe-NeRF [5] uses monocular depth prior as a geometry prior to constrain the relative poses. SPARF [37] and SCNeRF [16] demonstrate that using keypoint matches or dense correspondence can constrain the relative pose estimates with ray-to-ray correspondence losses. Park et al. [30] proposes a preconditioning strategy to enhance camera pose optimization. DBARF [8] proposes using low-frequency feature maps to address the joint optimization problem for generalizable NeRF. Bian et al. [4] proposes a pose residual field which learn the pose corrections to refine the initial camera pose for neural surface reconstruction. Closest to our approach is a very recent work L2G by Chen et al. [9], which tackles the camera pose representation using an overparameterization strategy. However, we achieve overparameterization in different manner. Unlike L2G which learn an MLP to predict rigid " + }, + { + "bbox": [ + 132, + 175, + 482, + 475 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 132, + 175, + 482, + 475 + ], + "type": "text", + "content": " transformations, we propose using an MLP to model the rigid warp function between the pixel and the ray space. Our work argues that while overparameterization can be achieved in different manner, it is tightly coupled with the regularization and constraints imposed. In our case, invertibility of warps becomes an essential constraint." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 490, + 270, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 270, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 270, + 502 + ], + "type": "text", + "content": "2.2 Overparameterization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 517, + 482, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 482, + 673 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 482, + 673 + ], + "type": "text", + "content": "Overparameterization in deep learning involves employing models with a substantially greater number of parameters than the quantity of training data. Recent research [22, 26, 35] demonstrates that neural networks, when overparameterized, can effectively generalize to new, unseen data, noting that an increase in parameters often correlates with a decrease in test error. Moreover, [26] revealed that this ability to generalize does not necessarily require explicit regularization, suggesting that the optimization process of overparameterized neural networks inherently prefers solutions that are more likely to generalize well. Additionally, [1, 2, 27, 29] illustrates that overparameterized networks are capable of consistently finding a global minimum through gradient-based optimization methods. These insights highlight the remarkable ability of overparameterized neural networks to deliver accurate predictions on unseen data, emphasizing their robustness and efficacy in generalizing beyond the training dataset." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "type": "text", + "content": "Invertible Neural Warp for NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 416, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 416, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 416, + 128 + ], + "type": "text", + "content": "2.3 Invertible Neural Networks for Deformation Fields" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 209 + ], + "type": "text", + "content": "Recent advancements in Invertible Neural Networks (INNs) [3,7,12] have broadened their use in the field of 3D deformation. These networks are particularly useful in modeling homermorphic deformation, where the mapping between any frames are bijective and continuous. Because of this capability, they are now being used in various areas, particularly modeling deformation in spatial [17,33,45] and temporal [6,20,28,39]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 225, + 233, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 233, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 233, + 239 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 251, + 481, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 251, + 481, + 275 + ], + "spans": [ + { + "bbox": [ + 131, + 251, + 481, + 275 + ], + "type": "text", + "content": "We define the mathematical notations for the camera operations and the joint camera pose estimation in Sec. 3.1. Further, we outline our approach in Sec. 3.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 284, + 339, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 284, + 339, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 284, + 339, + 296 + ], + "type": "text", + "content": "3.1 Bundle-Adjust NeRF Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 303, + 481, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 303, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 131, + 303, + 481, + 352 + ], + "type": "text", + "content": "Camera pose We consider a set of " + }, + { + "bbox": [ + 131, + 303, + 481, + 352 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 303, + 481, + 352 + ], + "type": "text", + "content": " input images as " + }, + { + "bbox": [ + 131, + 303, + 481, + 352 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 131, + 303, + 481, + 352 + ], + "type": "text", + "content": " taken by camera with intrinsic matrix " + }, + { + "bbox": [ + 131, + 303, + 481, + 352 + ], + "type": "inline_equation", + "content": "K \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 131, + 303, + 481, + 352 + ], + "type": "text", + "content": ". For each camera " + }, + { + "bbox": [ + 131, + 303, + 481, + 352 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 303, + 481, + 352 + ], + "type": "text", + "content": " corresponding to these images, BARF-style approaches [23] define its camera-to-world (C2W) as " + }, + { + "bbox": [ + 131, + 303, + 481, + 352 + ], + "type": "inline_equation", + "content": "P = (\\mathbf{R}_t, \\mathbf{t}_t) \\in SE(3)" + }, + { + "bbox": [ + 131, + 303, + 481, + 352 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 303, + 481, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_t \\in SO(3)" + }, + { + "bbox": [ + 131, + 303, + 481, + 352 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 303, + 481, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_t \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 131, + 303, + 481, + 352 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "text", + "content": "Camera projection For any vector " + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^l" + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "text", + "content": " of dimension " + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "text", + "content": ", we define its homogeneous representation " + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}} \\in \\mathbb{R}^{l+1}" + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}} = [\\mathbf{x}^T, 1]" + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "text", + "content": ". We define " + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "text", + "content": " as the camera projection operator, which maps a 3D point in the camera coordinate frame denoted as " + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{(C)} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "text", + "content": " to a corresponding 2D pixel coordinate " + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\mathbf{u} \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\pi^{-1}" + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "text", + "content": " denotes the camera backprojection that maps a pixel " + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "text", + "content": " coordinate and depth " + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "text", + "content": " to a 3D point in the camera coordinate as " + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{(C)}" + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "text", + "content": ", for e.g., " + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\pi(\\mathbf{x}^{(C)}) \\cong K\\mathbf{x}^{(C)}" + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\pi^{-1}(\\mathbf{u}, z) = zK^{-1}\\bar{\\mathbf{u}}" + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "inline_equation", + "content": "{}^{(C)}" + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "inline_equation", + "content": "{}^{(W)}" + }, + { + "bbox": [ + 131, + 358, + 482, + 455 + ], + "type": "text", + "content": " to denote that it is defined within the camera and world coordinate system respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "text", + "content": "NeRF NeRF represents the volumetric field of a 3D scene as " + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "inline_equation", + "content": "f(\\gamma(\\mathbf{x}), \\gamma(\\mathbf{d})) \\to (\\mathbf{c}, \\sigma)" + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "text", + "content": ", which maps a 3D location " + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "text", + "content": " and a viewing direction " + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "text", + "content": " to a RGB color " + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "text", + "content": " and volume density " + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\sigma \\in \\mathbb{R}" + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\gamma: \\mathbb{R}^3 \\to \\mathbb{R}^{3+6L}" + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "text", + "content": " is the positional embedding function with " + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "text", + "content": " frequency bases [25]. This function is parameterized using an MLP as " + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "inline_equation", + "content": "f_{\\Theta_{rgb}}" + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "text", + "content": ". Given " + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "text", + "content": " input images " + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "text", + "content": " with corresponding camera poses " + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\{P_t\\}_{t=1}^T" + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "text", + "content": ", NeRF is optimized by minimizing photometric loss " + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rgb}" + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "text", + "content": " between synthesized images " + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}" + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "text", + "content": " and original image " + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 131, + 462, + 482, + 548 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 217, + 556, + 481, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 556, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 217, + 556, + 481, + 590 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {\\Theta} _ {r g b}} \\sum_ {t = 1} ^ {T} \\sum_ {\\mathbf {u} \\in \\mathbb {R} ^ {2}} \\| \\hat {\\mathcal {I}} (\\mathbf {u}, P _ {t}; \\boldsymbol {\\Theta} _ {r g b}) - \\mathcal {I} _ {i} (\\mathbf {u}) \\| _ {2} ^ {2}, \\tag {1}", + "image_path": "f1a2efcf55d9f4212b0ecfb2e4e6f15128758a7891f53de7fcffd68bf8be3e6f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "text", + "content": "Volume rendering For simplicity, let's start by assuming the rendering operation of NeRF operates in the camera coordinate system. We will generalize this later. Each pixel coordinate determines a viewing direction " + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "text", + "content": " in the camera coordinate system, whose origin is the camera center of projection " + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{o}^{(C)}" + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "text", + "content": ". We can define a 3D point along the camera ray associated with " + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "text", + "content": " sampled at depth " + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "z_{i}" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "text", + "content": "Shin-Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "text", + "content": "as " + }, + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{r}^{(C)}(z) = \\mathbf{o}^{(C)} + z_{i,u}K^{-1}\\bar{\\mathbf{u}}" + }, + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "text", + "content": ". To render the colour of " + }, + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{L}}_{i,u}" + }, + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "text", + "content": " at pixel coordinate " + }, + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "text", + "content": ", we sample " + }, + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "text", + "content": " discrete depth values along the ray between a near bound " + }, + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "inline_equation", + "content": "z_{n}" + }, + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "text", + "content": " and a far bound " + }, + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "inline_equation", + "content": "z_{f}" + }, + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "text", + "content": ". For each sampled value, we query the NeRF " + }, + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "inline_equation", + "content": "f_{\\Theta_{rgb}}" + }, + { + "bbox": [ + 130, + 114, + 479, + 175 + ], + "type": "text", + "content": " to obtain the corresponding radiance fields. The output from NeRF is then aggregated to render the RGB colour as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 224, + 181, + 481, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 181, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 224, + 181, + 481, + 209 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {I}} (\\mathbf {u}) = \\int_ {z _ {n}} ^ {z _ {f}} T (\\mathbf {u}, z) \\sigma (\\mathbf {r} (z)) \\mathbf {c} (\\mathbf {r} (z)) \\delta z, \\tag {2}", + "image_path": "ed969548715c86cd97256772edcae97dbb295f3558963dbbd8646ed14a6a6d05.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "inline_equation", + "content": "T(\\mathbf{u},z) = \\exp (-\\int_{z_n}^z\\sigma (\\mathbf{r}(z)))\\delta z'" + }, + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "text", + "content": " denotes the accumulated transmittance value along the ray. We refer readers to [21, 25] for more details of volume rendering operation. In practice, Eq. (2) is approximated using " + }, + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "text", + "content": " points sampled along the ray at depth " + }, + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "inline_equation", + "content": "\\{z_i\\}_{i = 1}^M" + }, + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "text", + "content": ", which produces radiance field outputs as " + }, + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{y}_i\\}_{i = 1}^M" + }, + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "text", + "content": ". Denoting the ray compositing function in Eq. (2) as " + }, + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "inline_equation", + "content": "g(.)\\in \\mathbb{R}^{4M}\\to \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "text", + "content": ", we can rewrite " + }, + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(\\mathbf{u}) = g(\\{\\mathbf{y}_i\\}_{i = 1}^M)" + }, + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "text", + "content": ". Finally, given a camera pose " + }, + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "text", + "content": ", we can then transform the ray " + }, + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "inline_equation", + "content": "\\mathbf{r}^{(C)}" + }, + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "text", + "content": " to the world coordinate to obtain " + }, + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "inline_equation", + "content": "\\mathbf{r}^{(W)}" + }, + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "text", + "content": " through a 3D rigid transformation " + }, + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 216, + 480, + 314 + ], + "type": "text", + "content": ". The rendered image is then obtained as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 206, + 322, + 481, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 322, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 206, + 322, + 481, + 352 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {I}} (\\mathbf {u}, \\mathbf {p}) = g \\left(\\left\\{f \\left(\\mathcal {T} \\left(\\mathbf {r} ^ {(C)} (z), P\\right); \\boldsymbol {\\Theta} _ {r g b}\\right) \\right\\} _ {i = 1} ^ {M}\\right). \\tag {3}", + "image_path": "ddf5ff9ebe159cd3e9689c5cc114e6a25b1372678d10866c96063a6d12448c78.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 366, + 479, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 479, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 479, + 426 + ], + "type": "text", + "content": "Joint optimization of pose and NeRF Prior works [10, 23, 40, 42] demonstrate that it is feasible to optimize both camera pose and NeRF by minimizing " + }, + { + "bbox": [ + 130, + 366, + 479, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rgb}" + }, + { + "bbox": [ + 130, + 366, + 479, + 426 + ], + "type": "text", + "content": " in Eq. (1). This is achieved by considering " + }, + { + "bbox": [ + 130, + 366, + 479, + 426 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 366, + 479, + 426 + ], + "type": "text", + "content": " as estimizable parameters, see Sec. 3.1 for its parameterization. Consequently, ray " + }, + { + "bbox": [ + 130, + 366, + 479, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 366, + 479, + 426 + ], + "type": "text", + "content": " is now dependent on the camera pose. Mathematically, this joint optimization can be rewritten as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 190, + 434, + 481, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 434, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 190, + 434, + 481, + 474 + ], + "type": "interline_equation", + "content": "\\min _ {P, \\boldsymbol {\\Theta} _ {r g b}} \\sum_ {t = 1} ^ {T} \\sum_ {\\mathbf {u} \\in \\mathbb {R} ^ {2}} \\| \\hat {\\mathcal {I}} (\\underbrace {\\mathcal {T} \\left(\\mathbf {r} ^ {(C)} (z , \\bar {\\mathbf {u}}) ; P\\right)} _ {G (.)}; \\boldsymbol {\\Theta} _ {r g b}) - \\mathcal {I} _ {i} (\\mathbf {u}) \\| _ {2} ^ {2}. \\tag {4}", + "image_path": "f109f9bbea2089e50ccb22edcb8f9cf39459b3b293add0813a9bc99dc4339888.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 483, + 375, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 483, + 375, + 496 + ], + "spans": [ + { + "bbox": [ + 131, + 483, + 375, + 496 + ], + "type": "text", + "content": "3.2 Invertible Neural Warp for Ray Transform" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 503, + 479, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 479, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 479, + 635 + ], + "type": "text", + "content": "We propose to overparameterize " + }, + { + "bbox": [ + 130, + 503, + 479, + 635 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 503, + 479, + 635 + ], + "type": "text", + "content": " using an Invertible Neural Network (INN). There exist two options for parameterization: (i) use a separate INN for each camera " + }, + { + "bbox": [ + 130, + 503, + 479, + 635 + ], + "type": "inline_equation", + "content": "P_{t}" + }, + { + "bbox": [ + 130, + 503, + 479, + 635 + ], + "type": "text", + "content": "; (ii) use a single INN that is shared across all frames, coupled with a learnable code that is unique to frame " + }, + { + "bbox": [ + 130, + 503, + 479, + 635 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 503, + 479, + 635 + ], + "type": "text", + "content": ". Drawing inspiration from the dynamic NeRF method used for representing deformation fields [6,31,38,39] and also considering parameter efficiency, we have chosen to pursue the latter strategy for our proposed pose overparameterization - a single, globally shared neural " + }, + { + "bbox": [ + 130, + 503, + 479, + 635 + ], + "type": "inline_equation", + "content": "\\Theta_{\\mathcal{W}}" + }, + { + "bbox": [ + 130, + 503, + 479, + 635 + ], + "type": "text", + "content": " network across all frames, coupled with an estimizable per-frame latent code " + }, + { + "bbox": [ + 130, + 503, + 479, + 635 + ], + "type": "inline_equation", + "content": "\\varPhi_t\\in\\mathbb{R}^D" + }, + { + "bbox": [ + 130, + 503, + 479, + 635 + ], + "type": "text", + "content": ", see supp. (Sec. C) for the comparison of using multiple-INNs versus single-NN. Consequently, we can rewrite " + }, + { + "bbox": [ + 130, + 503, + 479, + 635 + ], + "type": "inline_equation", + "content": "G(.)" + }, + { + "bbox": [ + 130, + 503, + 479, + 635 + ], + "type": "text", + "content": " in Eq. (6) as " + }, + { + "bbox": [ + 130, + 503, + 479, + 635 + ], + "type": "inline_equation", + "content": "h(\\mathbf{r}^{(C)};\\Theta_{\\mathcal{W}},\\varPhi_t)" + }, + { + "bbox": [ + 130, + 503, + 479, + 635 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 503, + 479, + 635 + ], + "type": "inline_equation", + "content": "h(.):\\mathbb{R}^{3 + D}\\to \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 503, + 479, + 635 + ], + "type": "text", + "content": ". Fig. 2 presents our approach. In our approach, we" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "type": "text", + "content": "Invertible Neural Warp for NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 641, + 479, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 641, + 479, + 665 + ], + "type": "text", + "content": "2 This can be succinctly written as " + }, + { + "bbox": [ + 133, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{r}^{(C)}(z) = z_{i,u}\\mathbf{d}" + }, + { + "bbox": [ + 133, + 641, + 479, + 665 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 133, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{o}^{(C)}" + }, + { + "bbox": [ + 133, + 641, + 479, + 665 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 133, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "[0,0,0]^T" + }, + { + "bbox": [ + 133, + 641, + 479, + 665 + ], + "type": "text", + "content": " in camera coordinate space." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 320, + 281 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 320, + 281 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 320, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 320, + 281 + ], + "type": "image", + "image_path": "f35034cff4e28a497495b3c340d53d7ccaf3b638b130c164c7fcc9106c443139.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 289, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 338 + ], + "type": "text", + "content": "Fig. 2: An overview of our INN-based approach, illustrated using two views " + }, + { + "bbox": [ + 130, + 289, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_1" + }, + { + "bbox": [ + 130, + 289, + 482, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 289, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_2" + }, + { + "bbox": [ + 130, + 289, + 482, + 338 + ], + "type": "text", + "content": ". INN which is denoted as " + }, + { + "bbox": [ + 130, + 289, + 482, + 338 + ], + "type": "inline_equation", + "content": "h_{\\Theta_W}" + }, + { + "bbox": [ + 130, + 289, + 482, + 338 + ], + "type": "text", + "content": " takes the pixel locations in the camera coordinate system " + }, + { + "bbox": [ + 130, + 289, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,t}^{(C)}" + }, + { + "bbox": [ + 130, + 289, + 482, + 338 + ], + "type": "text", + "content": ", along with the frame-dependent latent code " + }, + { + "bbox": [ + 130, + 289, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\varPhi_t" + }, + { + "bbox": [ + 130, + 289, + 482, + 338 + ], + "type": "text", + "content": ", and output its corresponding location in the world coordinate system as " + }, + { + "bbox": [ + 130, + 289, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,t}^{(W)}" + }, + { + "bbox": [ + 130, + 289, + 482, + 338 + ], + "type": "text", + "content": ", see Sec. 3.2 for full details." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 320, + 113, + 481, + 279 + ], + "blocks": [ + { + "bbox": [ + 320, + 113, + 481, + 279 + ], + "lines": [ + { + "bbox": [ + 320, + 113, + 481, + 279 + ], + "spans": [ + { + "bbox": [ + 320, + 113, + 481, + 279 + ], + "type": "image", + "image_path": "a7753746da043b6f500c0a3eb653e59ff03a2aafdeee3d0f7323c78f9fa768ad.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 350, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 429 + ], + "type": "text", + "content": "model each pixel in the camera coordinate system " + }, + { + "bbox": [ + 130, + 350, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,t}^{(C)}" + }, + { + "bbox": [ + 130, + 350, + 482, + 429 + ], + "type": "text", + "content": " as an individual ray. Our proposed INN is designed for transforming these rays from camera coordinate to the world coordinate. Specifically, our proposed INN takes in the pixel coordinates " + }, + { + "bbox": [ + 130, + 350, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,t}^{(C)}" + }, + { + "bbox": [ + 130, + 350, + 482, + 429 + ], + "type": "text", + "content": " and camera center " + }, + { + "bbox": [ + 130, + 350, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_t^{(C)}" + }, + { + "bbox": [ + 130, + 350, + 482, + 429 + ], + "type": "text", + "content": ", both defined in the camera coordinates " + }, + { + "bbox": [ + 130, + 350, + 482, + 429 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 350, + 482, + 429 + ], + "type": "text", + "content": ", coupled with the frame-dependent latent code and outputs their corresponding equivalent " + }, + { + "bbox": [ + 130, + 350, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,t}^{(W)}" + }, + { + "bbox": [ + 130, + 350, + 482, + 429 + ], + "type": "text", + "content": " and camera center " + }, + { + "bbox": [ + 130, + 350, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_t^{(W)}" + }, + { + "bbox": [ + 130, + 350, + 482, + 429 + ], + "type": "text", + "content": " in the world coordinates." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "text", + "content": "Rigidity prior In our formulation, each pixel is represented as an individual ray within the camera coordinate system. This inherently relaxes the rigidity constraint. As a result, the output from the INN does not necessarily conform to a global rigid motion. Given known camera-world correspondences " + }, + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_{i,t}^{(C)},\\mathbf{x}_{i,t}^{(W)})" + }, + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "text", + "content": " we can solve a closed-form rigid registration problem to determine a global pose, which can be integrated into our optimization problem as a rigidity prior " + }, + { + "bbox": [ + 130, + 434, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rigid}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 241, + 514, + 481, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 514, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 241, + 514, + 481, + 548 + ], + "type": "interline_equation", + "content": "\\min _ {T ^ {*}} \\sum_ {i = 1} ^ {L} \\| \\mathbf {x} _ {i, t} ^ {(C)} - T ^ {*} \\circ \\mathbf {x} _ {i, t} ^ {(W)} \\| _ {2} ^ {2}. \\tag {5}", + "image_path": "6432a78f9884e7d69a77a39a4cae407010d224592777860881fbb17f337f088f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 553, + 465, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 553, + 465, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 553, + 465, + 566 + ], + "type": "text", + "content": "Final optimization problem We solve our final optimization problem as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 160, + 573, + 481, + 606 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 573, + 481, + 606 + ], + "spans": [ + { + "bbox": [ + 160, + 573, + 481, + 606 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {\\Phi} _ {t}, \\boldsymbol {\\Theta} _ {\\mathcal {W}}, \\boldsymbol {\\Theta} _ {r g b}} \\sum_ {t = 1} ^ {T} \\sum_ {\\mathbf {u} \\in \\mathbb {R} ^ {2}} \\| \\hat {\\mathcal {I}} (h (\\mathbf {r} ^ {(c)}; \\boldsymbol {\\Theta} _ {\\mathcal {W}}, \\boldsymbol {\\Phi} _ {t}); \\boldsymbol {\\Theta} _ {r g b}) - \\mathcal {I} _ {i} (\\mathbf {u}) \\| _ {2} ^ {2} + \\lambda \\mathcal {L} _ {r i g i d}. \\tag {6}", + "image_path": "f18d106743da8aa6d11b1f81644086b5bddcf439ce94df47e61fd8412f76965e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 622, + 469, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 469, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 469, + 635 + ], + "type": "text", + "content": "3.3 Advantages of INN for Overparameterizing Rigid Ray Warps" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": "BARF-based approach parameterizes camera pose " + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": " of each frame using a " + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": " (see Sec. 3.1), which guarantees that " + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": " is a bijective mapping. Therefore, when" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "text", + "content": "Shin-Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": "overparameterizing camera poses, it is crucial that the neural network adheres to the bijection property because this one-to-one correspondence ensures that there is a unique output in the world space for every point in the camera space. As we will demonstrate in Sec. 4.2, simply applying a rudimentary strategy (denoted as Naive) when overparameterizing the rigid warps of ray (camera-world) with an MLP often does not achieve convergence, see Fig. 1c and Tab. 1. Consequently, to attain this bijective property using an MLP, it is necessary to introduce an auxiliary network to model the backward warps (Implicit-Invertible MLP). While effective, it presents a significant drawback: it results in a twofold rise in the computational complexity due to the existence of the backward network to enforce the self-consistency. To mitigate this substantial increase in computational demands inherent in the modified MLP approach, we propose the use of INNs for parameterize these bijections. INN implements the bijective mappings by composing affine transformations into several blocks. Within each block, the input coordinates are divided into two segments; the first part remains constant and is used to parameterize the transformation that is applied to the second part [6]. Besides their inherent invertibility, INN also offers the advantage of homeomorphic property, which potentially facilitate a more flexible optimization trajectory that is less susceptible to a suboptimal minimum trajectory, see Sec. 4.5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 360, + 230, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 360, + 230, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 360, + 230, + 374 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 384, + 208, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 208, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 208, + 395 + ], + "type": "text", + "content": "4.1Baselines" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 402, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 510 + ], + "type": "text", + "content": "We compare our approach with two representative methods in pose-NeRF joint optimization: the standard global " + }, + { + "bbox": [ + 130, + 402, + 482, + 510 + ], + "type": "inline_equation", + "content": "SE3" + }, + { + "bbox": [ + 130, + 402, + 482, + 510 + ], + "type": "text", + "content": "-approach BARF [23], and the overparameterized representation L2G [9]. For all experiments, we use the original implementations including their default settings for coarse-to-fine scheduling, architecture and hyperparameters, see supp. (Sec. A) for more details. Additionally, in our 2D planar experiments, we include a comparison with another two variants called the Naive (MLP) and Implicit-Invertible MLP to execute ray transform. This comparison is specifically designed to highlight the significance of invertibility when employing MLPs for executing ray transformation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "type": "text", + "content": "Local-to-global (L2G) [9] L2G uses an MLP to predict rigid SE3 transformation for each ray. These predicted transformation parameters are then used to analytically estimate the transformed coordinates in the world space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 607 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 607 + ], + "type": "text", + "content": "Naive This is the simplest version of our baseline that uses one primary network, denoted as " + }, + { + "bbox": [ + 130, + 558, + 482, + 607 + ], + "type": "inline_equation", + "content": "h_{fwd}:(\\mathbf{x}^{(C)},\\varPhi_t)\\to \\mathbf{x}^{(W)}" + }, + { + "bbox": [ + 130, + 558, + 482, + 607 + ], + "type": "text", + "content": ", to learn the forward mapping which takes in the coordinates from the camera space (coupled with a per-frame latent code), and outputs the corresponding coordinates in the world space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 612, + 482, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 612, + 482, + 673 + ], + "spans": [ + { + "bbox": [ + 131, + 612, + 482, + 673 + ], + "type": "text", + "content": "Implicit-Invertible MLP We use two networks " + }, + { + "bbox": [ + 131, + 612, + 482, + 673 + ], + "type": "inline_equation", + "content": "h_{fwd}" + }, + { + "bbox": [ + 131, + 612, + 482, + 673 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 612, + 482, + 673 + ], + "type": "inline_equation", + "content": "h_{bwd}" + }, + { + "bbox": [ + 131, + 612, + 482, + 673 + ], + "type": "text", + "content": " to enforce approximate invertibility. Alongside primary network " + }, + { + "bbox": [ + 131, + 612, + 482, + 673 + ], + "type": "inline_equation", + "content": "h_{fwd}" + }, + { + "bbox": [ + 131, + 612, + 482, + 673 + ], + "type": "text", + "content": ", we use a secondary network " + }, + { + "bbox": [ + 131, + 612, + 482, + 673 + ], + "type": "inline_equation", + "content": "h_{bwd} : (\\mathbf{x}^{(W)}, \\varPhi_t) \\to \\hat{\\mathbf{x}}^{(C)}" + }, + { + "bbox": [ + 131, + 612, + 482, + 673 + ], + "type": "text", + "content": " to invert the outputs from the primary network " + }, + { + "bbox": [ + 131, + 612, + 482, + 673 + ], + "type": "inline_equation", + "content": "h_{fwd}" + }, + { + "bbox": [ + 131, + 612, + 482, + 673 + ], + "type": "text", + "content": ". To minimize deviations from bijections, we introduce a regularization term " + }, + { + "bbox": [ + 131, + 612, + 482, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{implicit}}" + }, + { + "bbox": [ + 131, + 612, + 482, + 673 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 131, + 612, + 482, + 673 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{x}^{(C)} - \\hat{\\mathbf{x}}^{(C)} \\|_2^2" + }, + { + "bbox": [ + 131, + 612, + 482, + 673 + ], + "type": "text", + "content": " into the optimization problem Eq. (6)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "type": "text", + "content": "Invertible Neural Warp for NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "Explicit-Invertible INN (Ours) We explicitly model inversions in the neural network architecture by formally learning an INN. We have chosen to utilize architecture proposed by NDR-INN [6], see supp. (Sec. A.1) for the architecture details that we use for all our experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 182, + 340, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 182, + 340, + 194 + ], + "spans": [ + { + "bbox": [ + 131, + 182, + 340, + 194 + ], + "type": "text", + "content": "4.2 2D Planar Neural Image Alignment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 202, + 482, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 202, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 482, + 335 + ], + "type": "text", + "content": "Following BARF [9, 10, 23], we learn a 2D neural image field, for creating a homography-based panoramic image from " + }, + { + "bbox": [ + 130, + 202, + 482, + 335 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 202, + 482, + 335 + ], + "type": "text", + "content": " patches cropped from the original image, each generated with random homography perturbations. Specifically, we learn a 2D coordinate network " + }, + { + "bbox": [ + 130, + 202, + 482, + 335 + ], + "type": "inline_equation", + "content": "f(\\Theta_{rgb})" + }, + { + "bbox": [ + 130, + 202, + 482, + 335 + ], + "type": "text", + "content": " to render the stitched image. Each pixel in the " + }, + { + "bbox": [ + 130, + 202, + 482, + 335 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 202, + 482, + 335 + ], + "type": "text", + "content": " training patch is warped using the estimated homography " + }, + { + "bbox": [ + 130, + 202, + 482, + 335 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 202, + 482, + 335 + ], + "type": "text", + "content": " to create the rendered image. We choose the \"cat\" image from ImageNet [11]. We initialize patch warps as identity and fix the gauge freedom by anchoring the first warp to align the neural image to the original image [9, 10, 23]. We randomly generate 20 different homography instances, with scale-noise parameter 0.1 and 0.2 for homography and translation, respectively. We solve Eq. (5) for the homography using a Direct Linear Transform (DLT) solver3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "text", + "content": "Experiment settings. We evaluate our proposed method against BARF [23], and our three overparameterized network variants: Naive, Implicit- and Explicit-Invertible MLPs, as detailed in Sec. 4.1. For both naive and Implicit-Invertible MLPs, we utilized a Leaky-ReLU MLP with five 256-dimensional hidden units, and a 16-dimensional latent code to represent the frame-dependent embeddings " + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "inline_equation", + "content": "\\phi_t" + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "text", + "content": ". We also follow the default coarse-to-fine scheduling established by BARF. In the case of Implicit-Invertible MLP, we use the same architecture both " + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "inline_equation", + "content": "h_{fwd}" + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "inline_equation", + "content": "h_{bwd}" + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "text", + "content": ". We use the Adam optimizer [19] to optimize for both the network weights " + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "inline_equation", + "content": "\\Theta_{rgb}" + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "inline_equation", + "content": "\\Theta_{\\mathcal{W}}" + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "text", + "content": ". We set the learning rate for both " + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "inline_equation", + "content": "\\Theta_{rgb}" + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "inline_equation", + "content": "\\Theta_{\\mathcal{W}}" + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "inline_equation", + "content": "1\\times 10^{-3}" + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "text", + "content": ", with both decaying exponentially to " + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "inline_equation", + "content": "1\\times 10^{-4}" + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "inline_equation", + "content": "1\\times 10^{-5}" + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "text", + "content": ", respectively. We set the weighting term for " + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rigid}" + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "text", + "content": " for both overparameterized MLPs to " + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "inline_equation", + "content": "1\\times 10^{2}" + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "text", + "content": " and the consistency term " + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{implicit}" + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "text", + "content": " for Implicit-Invertible MLP to " + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "inline_equation", + "content": "1\\times 10^{1}" + }, + { + "bbox": [ + 130, + 342, + 482, + 488 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 494, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 567 + ], + "type": "text", + "content": "Robustness to noise perturbations Fig. 3 analyzes the robustness of BARF and our approach under different noise perturbations across 20 different homography instances. Each run began with the initialization of the homography using the \"groundtruth\", and noise perturbations were gradually introduced, ranging from 0 to 0.3 to the translation component of the homography. Fig. 3 indicates that our representation exhibits a higher tolerance to noise compared to BARF." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": "Results. Tab. 1 summarizes the statistical results from 20 runs, where we report the warp error and patch reconstruction error. We quantify the warp error in terms of corner error, defined as the L2 distance between the groundtruth corner position and estimated corner position, and PSNR as the metric to assess the reconstruction quality. We used 5-pixel threshold to define the success" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Shin-Fang et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 652, + 295, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 295, + 665 + ], + "type": "text", + "content": "3 https://github.com/kornia/kornia" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 116, + 276, + 212 + ], + "blocks": [ + { + "bbox": [ + 135, + 116, + 276, + 212 + ], + "lines": [ + { + "bbox": [ + 135, + 116, + 276, + 212 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 276, + 212 + ], + "type": "image", + "image_path": "4b17d1eae930c5bc8390a47b0275281b6897058455224ce79820848965ed1f60.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 215, + 480, + 292 + ], + "lines": [ + { + "bbox": [ + 130, + 215, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 480, + 292 + ], + "type": "text", + "content": "Fig. 3: Basin convergence analysis of our approach versus BARF in a 2D planar experiment. On the left, we show the success rate of 20 runs, where we initialize the homography using groundtruth, and gradually introduced noise perturbations to the translation component. The noise scale is varied from 0 to 0.30. We used 5-pixel threshold to determine success convergence. Notably, our approach (blue) demonstrates higher noise tolerance compared to BARF (red). On the right, we show a qualitative comparison when both methods are perturbed with the highest magnitude of noise." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 277, + 116, + 482, + 203 + ], + "blocks": [ + { + "bbox": [ + 277, + 116, + 482, + 203 + ], + "lines": [ + { + "bbox": [ + 277, + 116, + 482, + 203 + ], + "spans": [ + { + "bbox": [ + 277, + 116, + 482, + 203 + ], + "type": "image", + "image_path": "aa5ec814e80ca259bc42627f6d513d55dd5f2185e8655068112a276be203ce23.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 342, + 479, + 417 + ], + "blocks": [ + { + "bbox": [ + 130, + 296, + 481, + 340 + ], + "lines": [ + { + "bbox": [ + 130, + 296, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 481, + 340 + ], + "type": "text", + "content": "Table 1: Statistical result for 20 homography runs, with scale noise of 0.1 for homography and 0.2 for translation. The warp error is quantified in terms of corner error, and the patch reconstruction error in measured in PSNR. We provide mean and std dev for the evaluation. We used 5-pixel threshold to determine success convergence." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 342, + 479, + 417 + ], + "lines": [ + { + "bbox": [ + 135, + 342, + 479, + 417 + ], + "spans": [ + { + "bbox": [ + 135, + 342, + 479, + 417 + ], + "type": "table", + "html": "
Corner error (px) ↓Patch PSNR ↑Success rate ↑\n(Upper bound:1.00)
MeanStd. dev.MeanStd. dev.
BARF [23]29.6328.1828.944.380.30
Naive85.5930.3125.862.070.00
Implicit-Invertible13.9222.9333.703.930.65
Explicit-Invertible (INN)4.706.4734.712.370.75
", + "image_path": "e80186b5f90fa0a07d0ece99b402fe9a0423299ba5974b9eb697375d6c194378.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 450, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 668 + ], + "type": "text", + "content": "convergence. Fig. 1 presents a qualitative result for a homography instance. By enforcing approximate invertibility, the Implicit-Invertible MLP demonstrates a significantly higher rate of successful convergence compared to the naive version of MLP, with success rates improved by " + }, + { + "bbox": [ + 130, + 450, + 482, + 668 + ], + "type": "inline_equation", + "content": "65\\%" + }, + { + "bbox": [ + 130, + 450, + 482, + 668 + ], + "type": "text", + "content": ". We further show that by explicitly injecting invertibility into the architecture (Ours), the success rate is increased to " + }, + { + "bbox": [ + 130, + 450, + 482, + 668 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 130, + 450, + 482, + 668 + ], + "type": "text", + "content": ". This result reinforces that invertibility is crucial when learning rigid warp functions via overparameterization, and using an architecture that guarantees bijective property is effective in ensuring the pose converge to an optimal solution during joint optimization in practice. Henceforth, we will focus exclusively on using our INN-based approach for subsequent results. For completeness, we also compare with another setup, where each frame is parameterized with a neural network, see supp. (Sec. B.1). Interestingly, apart from the parameter efficiency, we find that using a single global neural network for all frames is sufficient to converge to a good pose solutions. We hypothesized that the difference may be attributed to the benefits of gradient sharing in the shared neural network setup. This result is aligned with the findings by Bian et al. [4]. Consequently, we adhere to the design where we use one single INN shared across all the frames, coupled with a frame-specific latent code for the rest of our experiments." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 311, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 311, + 91, + 447, + 102 + ], + "type": "text", + "content": "Invertible Neural Warp for NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 247, + 215 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 247, + 215 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 247, + 215 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 247, + 215 + ], + "type": "image", + "image_path": "a2ebeb77bc58d43a7af206c0b50a570e50866057e1703f7bb6553d1cb219cd1e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 216, + 219, + 225 + ], + "lines": [ + { + "bbox": [ + 165, + 216, + 219, + 225 + ], + "spans": [ + { + "bbox": [ + 165, + 216, + 219, + 225 + ], + "type": "text", + "content": "(a) BARF [23]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 249, + 114, + 361, + 215 + ], + "blocks": [ + { + "bbox": [ + 249, + 114, + 361, + 215 + ], + "lines": [ + { + "bbox": [ + 249, + 114, + 361, + 215 + ], + "spans": [ + { + "bbox": [ + 249, + 114, + 361, + 215 + ], + "type": "image", + "image_path": "3c8165a25bfd71dcc17d991a2e70cc63ad0b0ad3a17c0365743616f4b35d214f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 216, + 326, + 225 + ], + "lines": [ + { + "bbox": [ + 284, + 216, + 326, + 225 + ], + "spans": [ + { + "bbox": [ + 284, + 216, + 326, + 225 + ], + "type": "text", + "content": "(b) L2G [9]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 226, + 481, + 258 + ], + "lines": [ + { + "bbox": [ + 131, + 226, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 131, + 226, + 481, + 258 + ], + "type": "text", + "content": "Fig. 4: Qualitative analysis of reconstruction error on leaves (top) and trex (bottom). We present the average image reconstruction error through insets. Our approach presents the lowest misalignment error, as indicated darker areas in the error map." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 362, + 114, + 476, + 215 + ], + "blocks": [ + { + "bbox": [ + 362, + 114, + 476, + 215 + ], + "lines": [ + { + "bbox": [ + 362, + 114, + 476, + 215 + ], + "spans": [ + { + "bbox": [ + 362, + 114, + 476, + 215 + ], + "type": "image", + "image_path": "df35eb51c9f26f45c51f65301be34f41a44855fe4a9a341ed753f1c041b842e6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 393, + 216, + 445, + 225 + ], + "lines": [ + { + "bbox": [ + 393, + 216, + 445, + 225 + ], + "spans": [ + { + "bbox": [ + 393, + 216, + 445, + 225 + ], + "type": "text", + "content": "(c) INN(Ours)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 280, + 321, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 280, + 321, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 280, + 321, + 293 + ], + "type": "text", + "content": "4.3 Neural Radiance Fields (NeRF)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 298, + 480, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 480, + 358 + ], + "type": "text", + "content": "In this section, we compare our proposed representation with BARF [23] and L2G [9]. We assume known intrinsic for all methods. We perform our experiments on both the LLFF [24] (Sec. 4.4), DTU [15] (Sec. 4.6) as well as Blender datasets in supp. (Sec. B2). We solve Eq. (5) for the global rigid SE3 transformation using Umayama algorithm " + }, + { + "bbox": [ + 130, + 298, + 480, + 358 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 130, + 298, + 480, + 358 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 364, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 579 + ], + "type": "text", + "content": "Evaluation metrics. For pose estimation, we report the accuracy of the poses after globally aligning the optimized poses to the groundtruth [9,10,23,37]. We assess view synthesis using PSNR, SSIM and LPIPS. A standard procedure in view synthesis evaluation involves performing test-time photometric optimization on the trained models. This additional step is intended to factor out the pose errors, which may otherwise compromise the quality of the synthesized views [9,10,23,36,37]. This process is akin to a pose refinement method which minimises the photometric error on the synthesized image while keeping the trained NeRF model fixed. However, it is important to recognize that this pose correction may not accurately represent the initial accuracy of the methods in terms of pose estimation for view synthesis. Therefore, we opt to report the view synthesis quality both before and after the pose refinement step. On DTU, we extend our evaluation to include comparisons of rendered depth with ground-truth depth using mean depth absolute error, as well as reconstruction accuracy using Chamfer distance. For the Chamfer evaluation, we utilize the optimized poses estimated from all methods and employ a neural surface reconstruction algorithm called Voxurf [41] for geometry reconstruction. Further details on pose alignment and metrics computations in the supp. (Sec. A)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 595, + 310, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 595, + 310, + 607 + ], + "spans": [ + { + "bbox": [ + 132, + 595, + 310, + 607 + ], + "type": "text", + "content": "4.4 Forward Facing Scenes: LLFF" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 613, + 480, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 613, + 480, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 613, + 480, + 636 + ], + "type": "text", + "content": "Experiment settings. The standard LLFF benchmark dataset [24] consists of eight real-world, forward-facing scenes captured using hand-held cameras." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "text", + "content": "Shin-Fang et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 641, + 282, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 641, + 282, + 653 + ], + "spans": [ + { + "bbox": [ + 133, + 641, + 282, + 653 + ], + "type": "text", + "content": "4 https://github.com/naver/roma" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 135, + 653, + 443, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 653, + 443, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 653, + 443, + 665 + ], + "type": "text", + "content": "For our proposed method, we evaluate the estimated global poses Eq. (5)." + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 148, + 479, + 224 + ], + "blocks": [ + { + "bbox": [ + 132, + 105, + 482, + 138 + ], + "lines": [ + { + "bbox": [ + 132, + 105, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 105, + 482, + 138 + ], + "type": "text", + "content": "Table 2: Evaluation of the LLFF dataset [24] using initial identity poses. We compute results for BARF [23] and L2G [9] using their codebases. The results are averaged over all eight scenes, see supp. (Sec. B) for a breakdown of results per scene." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 148, + 479, + 224 + ], + "lines": [ + { + "bbox": [ + 135, + 148, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 135, + 148, + 479, + 224 + ], + "type": "table", + "html": "
Pose accuracyNovel view synthesis
Rotation (°)Translation (×100)Before test-timeAfter test-time
PSNRSSIMLPIPSPSNRSSIMLPIPS
BARF [23]0.900.4017.000.410.3223.820.720.24
L2G [9]0.480.3017.990.460.2624.350.750.21
INN(Ours)0.310.2419.310.520.2524.280.740.22
", + "image_path": "82c5cb05e2135fb6811afa097f5c6aceee4868daac85ae1644fb719cacd185a1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 479, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 479, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 479, + 283 + ], + "type": "text", + "content": "Following [10, 23, 37], we initialize all camera poses with the identity transformations for all the methods. We employ the same training and testing split as in BARF [23]. We use the evaluation metrics described in Sec. 4.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 291, + 480, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 480, + 352 + ], + "type": "text", + "content": "Implementation details. We train all methods for 200k iterations and randomly sample 2048 pixel rays at each optimization step [9, 10, 23]. We train without hierarchical sampling. We set the learning for " + }, + { + "bbox": [ + 130, + 291, + 480, + 352 + ], + "type": "inline_equation", + "content": "\\Theta_{rgb}" + }, + { + "bbox": [ + 130, + 291, + 480, + 352 + ], + "type": "text", + "content": " to be " + }, + { + "bbox": [ + 130, + 291, + 480, + 352 + ], + "type": "inline_equation", + "content": "1\\times 10^{-3}" + }, + { + "bbox": [ + 130, + 291, + 480, + 352 + ], + "type": "text", + "content": " decaying to " + }, + { + "bbox": [ + 130, + 291, + 480, + 352 + ], + "type": "inline_equation", + "content": "3\\times 10^{-4}" + }, + { + "bbox": [ + 130, + 291, + 480, + 352 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 291, + 480, + 352 + ], + "type": "inline_equation", + "content": "5\\times 10^{-4}" + }, + { + "bbox": [ + 130, + 291, + 480, + 352 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 291, + 480, + 352 + ], + "type": "inline_equation", + "content": "\\Theta_{\\mathcal{W}}" + }, + { + "bbox": [ + 130, + 291, + 480, + 352 + ], + "type": "text", + "content": " decaying to " + }, + { + "bbox": [ + 130, + 291, + 480, + 352 + ], + "type": "inline_equation", + "content": "1\\times 10^{-6}" + }, + { + "bbox": [ + 130, + 291, + 480, + 352 + ], + "type": "text", + "content": ". We also follow the default coarse-to-fine scheduling by BARF [23], see supp. for full details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 360, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 529 + ], + "type": "text", + "content": "Results. Our approach achieves a substantial reduction of rotation errors (70% vs. BARF and 35% vs. L2G) and translation errors (50% vs. BARF and 20% vs. L2G). Additionally, the superior performance of both our method and L2G over BARF further highlight the merits of overparameterization for simultaneous pose and neural field estimation. This significant improvement in camera pose accuracy directly enhances the performance of view synthesis, as evident in Tab. 2. In particular, instances such as trex and leaves observe a significant improvement. Fig. 4 illustrates the absolute error between the original image and the rendered image on these two instances. Our approach presents the lowest misalignment error, as indicated darker areas in the error map. It is important to note, however, while L2G and our approach appear comparable after test-time optimization, this refinement procedure has mitigated the pose estimation noise to improve PSNR, as discussed in Sec. 4.3. For more qualitative results and ablation studies, please refer to supp. (Sec. B)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 536, + 422, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 422, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 422, + 550 + ], + "type": "text", + "content": "4.5 Homeomorphism perspective: A qualitative analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "We present a qualitative analysis on a single-view pose estimation that sheds light on the empirical effectiveness of our approach compared to the L2G method [9]. Our method leverages the concept of an INN, which predict homeomorphisms—continuous, invertible transformations not limited to the rigid motions of the " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " group. Unlike the L2G method, which constrains pose estimation within the rigid bounds of " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ", our INN-based approach embraces a broader spectrum of transformations. This grants the optimization process a higher degree of flexibility, allowing for a diverse range of optimization paths and facilitating a smoother trajectory towards the solution. To empirically validate our" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 310, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 310, + 91, + 447, + 101 + ], + "type": "text", + "content": "Invertible Neural Warp for NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 223, + 228 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 223, + 228 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 223, + 228 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 223, + 228 + ], + "type": "image", + "image_path": "de9791282530dc7f2aa2d00955f63115a77d431ac92212169c6d571e7c0c7744.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 228, + 193, + 236 + ], + "lines": [ + { + "bbox": [ + 170, + 228, + 193, + 236 + ], + "spans": [ + { + "bbox": [ + 170, + 228, + 193, + 236 + ], + "type": "text", + "content": "iter=0" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 247, + 479, + 323 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 479, + 323 + ], + "type": "text", + "content": "Fig. 5: Qualitative analysis of intermediate rendered image compared to superimposed Groundtruth (lighter visualization) when using L2G [9] and our approach on single-view pose estimation. Notably, when using our approach for pose estimation, we observe noticeable deformation in the rendered scene depicted in the bottom row (zoom in for better view). These deformations indicate that at each iteration, the INN predicts general homeomorphisms that are not rigid transformation. Thus yielding a flexible optimization trajectory that does not land in a suboptimal minimum trajectory." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 224, + 114, + 309, + 228 + ], + "blocks": [ + { + "bbox": [ + 224, + 114, + 309, + 228 + ], + "lines": [ + { + "bbox": [ + 224, + 114, + 309, + 228 + ], + "spans": [ + { + "bbox": [ + 224, + 114, + 309, + 228 + ], + "type": "image", + "image_path": "d36029cc3d3a8d6d36357dcc155cad1b619265218fd889bb819da70e0a464f18.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 228, + 285, + 236 + ], + "lines": [ + { + "bbox": [ + 258, + 228, + 285, + 236 + ], + "spans": [ + { + "bbox": [ + 258, + 228, + 285, + 236 + ], + "type": "text", + "content": "iter=40" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 394, + 228 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 394, + 228 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 394, + 228 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 394, + 228 + ], + "type": "image", + "image_path": "3888edd3b5a9d0af273d3b0f90dabbb8c7db78aedc858fe07cf7fdac739c7fca.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 228, + 370, + 236 + ], + "lines": [ + { + "bbox": [ + 343, + 228, + 370, + 236 + ], + "spans": [ + { + "bbox": [ + 343, + 228, + 370, + 236 + ], + "type": "text", + "content": "iter=80" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 394, + 114, + 480, + 228 + ], + "blocks": [ + { + "bbox": [ + 394, + 114, + 480, + 228 + ], + "lines": [ + { + "bbox": [ + 394, + 114, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 394, + 114, + 480, + 228 + ], + "type": "image", + "image_path": "54efe1a573ac49bca4d09c0b97f515f08a3fcf6d8eb19eb754df9eff17e9c4e2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 228, + 457, + 236 + ], + "lines": [ + { + "bbox": [ + 425, + 228, + 457, + 236 + ], + "spans": [ + { + "bbox": [ + 425, + 228, + 457, + 236 + ], + "type": "text", + "content": "iter=500" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 133, + 371, + 479, + 448 + ], + "blocks": [ + { + "bbox": [ + 130, + 325, + 480, + 369 + ], + "lines": [ + { + "bbox": [ + 130, + 325, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 480, + 369 + ], + "type": "text", + "content": "Table 3: Absolute pose accuracy evaluation of the DTU dataset [15] using initial noisy poses that corresponds to an average rotation and translation error of " + }, + { + "bbox": [ + 130, + 325, + 480, + 369 + ], + "type": "inline_equation", + "content": "15^{\\circ}" + }, + { + "bbox": [ + 130, + 325, + 480, + 369 + ], + "type": "text", + "content": " and 70 respectively. The upper section shows rotation errors in degrees, while the lower section displays translation errors " + }, + { + "bbox": [ + 130, + 325, + 480, + 369 + ], + "type": "inline_equation", + "content": "(\\times 100)" + }, + { + "bbox": [ + 130, + 325, + 480, + 369 + ], + "type": "text", + "content": ". red box denotes the best result." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 371, + 479, + 448 + ], + "lines": [ + { + "bbox": [ + 133, + 371, + 479, + 448 + ], + "spans": [ + { + "bbox": [ + 133, + 371, + 479, + 448 + ], + "type": "table", + "html": "
MethodsMetricScan IDsInitial rotation err: 15°, translation err (×100) : 70
243740556365698397105106110114118mean
BARFrotation1.041.880.552.850.430.783.880.542.203.5611.812.011.212.602.52
L2G0.903.070.5710.810.740.941.550.590.642.8513.155.922.9012.474.08
INN(Ours)0.210.480.731.640.400.622.750.550.540.303.892.911.060.251.17
BARFtranslation3.004.082.0911.631.432.7812.802.136.1212.4025.885.524.085.047.07
L2G2.408.802.1332.742.783.485.082.312.375.3436.2014.018.6437.1611.67
INN(Ours)0.501.010.975.240.960.785.300.671.190.439.9611.523.780.643.07
", + "image_path": "2c5d33e8bacd25fb1e184331467d45dbf3707563ee550f19c44d379ac9cd497f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 472, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 616 + ], + "type": "text", + "content": "hypothesis, we conducted experiments using a trained NeRF model to estimate the camera pose relative to a 3D scene, aiming to minimize the photometric error between NeRF-rendered and actual observed images. Despite starting from the same initial pose (off by " + }, + { + "bbox": [ + 130, + 472, + 482, + 616 + ], + "type": "inline_equation", + "content": "20^{\\circ}" + }, + { + "bbox": [ + 130, + 472, + 482, + 616 + ], + "type": "text", + "content": "), and employing a random sampling of 2048 rays per iteration for all methods, our INN approach outperformed the L2G method. The L2G method often converged to suboptimal poses, whereas the INN method achieved accurate pose estimation, as evidenced in Fig. 5 where the NeRF-rendered images is well-aligned with the groundtruth. This success can be attributed to the INN's ability to predict a general homeomorphism, a transformation more general than a rigid transformation, which offers a significant advantage in navigating the optimization landscape more effectively and avoiding suboptimal local minima." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 622, + 253, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 253, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 253, + 633 + ], + "type": "text", + "content": "4.6 " + }, + { + "bbox": [ + 132, + 622, + 253, + 633 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 132, + 622, + 253, + 633 + ], + "type": "text", + "content": " Scenes: DTU" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 666 + ], + "type": "text", + "content": "Experimental settings. We evaluated on 14 test scenes from DTU [4]. Following [9,23,37], we synthetically perturb the ground-truth camera poses with" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Shin-Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 160, + 480, + 242 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 4: Geometry evaluation of the DTU dataset [15], using initial noisy poses that corresponds to an average rotation and translation error of " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "15^{\\circ}" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": " and 70 respectively. The upper section presents mean absolute depth, while the lower section displays reconstruction error measured by Chamfer distance. red box denotes the best result." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 160, + 480, + 242 + ], + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 242 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 242 + ], + "type": "table", + "html": "
Scan IDs | Initial rotation err: 15°, translation err (×100) : 70
PoseMetric243740556365698397105106110114118mean
BARF0.110.150.110.120.140.100.260.170.170.460.530.170.100.170.20
L2Gdepth ↓0.080.240.120.380.170.120.110.180.100.180.240.170.120.840.22
INN(Ours)(abs)0.060.140.130.120.150.080.180.170.110.110.130.300.080.060.13
BARF5.474.074.397.644.835.286.435.748.067.217.758.298.245.476.35
L2GChamfer ↓4.379.084.416.678.035.923.806.006.305.597.437.358.298.166.53
INN(Ours)1.562.255.555.684.584.275.515.695.093.738.157.857.351.144.89
", + "image_path": "6f6960e28f524f0823314e0840d712c0a852834faa37144879c3c58305d58756.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 266, + 479, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 479, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 479, + 315 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 130, + 266, + 479, + 315 + ], + "type": "text", + "content": " of additive Gaussian noise, which corresponds to an average rotation and translation error of " + }, + { + "bbox": [ + 130, + 266, + 479, + 315 + ], + "type": "inline_equation", + "content": "15^{\\circ}" + }, + { + "bbox": [ + 130, + 266, + 479, + 315 + ], + "type": "text", + "content": " and 70, respectively. For a fair comparison, we used the same initialization for all methods. We refer the readers to supp. (Sec B.5) for the results with Colmap initialization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 322, + 480, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 480, + 419 + ], + "type": "text", + "content": "Implementation details. As BARF [23] and L2G [9] have not tested on DTU datasets, given that the DTU dataset encompasses " + }, + { + "bbox": [ + 130, + 322, + 480, + 419 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 322, + 480, + 419 + ], + "type": "text", + "content": " scenes which is similar to the original Blender dataset, we adopted the original hyperparameters used by the author for training BARF and L2G on the blender dataset. Following Bian et al. [4], we multiply the output of their local warp network by a small factor which is " + }, + { + "bbox": [ + 130, + 322, + 480, + 419 + ], + "type": "inline_equation", + "content": "\\alpha = 0.01" + }, + { + "bbox": [ + 130, + 322, + 480, + 419 + ], + "type": "text", + "content": " for L2G [9]. For our approach, we set the learning rate for " + }, + { + "bbox": [ + 130, + 322, + 480, + 419 + ], + "type": "inline_equation", + "content": "\\Theta_{rgb}" + }, + { + "bbox": [ + 130, + 322, + 480, + 419 + ], + "type": "text", + "content": " to be " + }, + { + "bbox": [ + 130, + 322, + 480, + 419 + ], + "type": "inline_equation", + "content": "1\\times 10^{-3}" + }, + { + "bbox": [ + 130, + 322, + 480, + 419 + ], + "type": "text", + "content": " decaying to " + }, + { + "bbox": [ + 130, + 322, + 480, + 419 + ], + "type": "inline_equation", + "content": "1\\times 10^{-4}" + }, + { + "bbox": [ + 130, + 322, + 480, + 419 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 322, + 480, + 419 + ], + "type": "inline_equation", + "content": "\\Theta_{\\mathcal{W}}" + }, + { + "bbox": [ + 130, + 322, + 480, + 419 + ], + "type": "text", + "content": " to start from " + }, + { + "bbox": [ + 130, + 322, + 480, + 419 + ], + "type": "inline_equation", + "content": "5\\times 10^{-4}" + }, + { + "bbox": [ + 130, + 322, + 480, + 419 + ], + "type": "text", + "content": " and decaying to " + }, + { + "bbox": [ + 130, + 322, + 480, + 419 + ], + "type": "inline_equation", + "content": "1\\times 10^{-8}" + }, + { + "bbox": [ + 130, + 322, + 480, + 419 + ], + "type": "text", + "content": ". We use the default coarse-to-fine scheduling by BARF [23]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 428, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 428, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 428, + 482, + 536 + ], + "type": "text", + "content": "Results. As demonstrated in Tab. 3, we outperform all baselines by a considerable margin across majority of the sequences in pose accuracy. Overall, our approach achieves approximately a " + }, + { + "bbox": [ + 130, + 428, + 482, + 536 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 428, + 482, + 536 + ], + "type": "text", + "content": " improvement in rotation and a " + }, + { + "bbox": [ + 130, + 428, + 482, + 536 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 428, + 482, + 536 + ], + "type": "text", + "content": " improvement in translation over BARF. When compared to L2G, our method shows a " + }, + { + "bbox": [ + 130, + 428, + 482, + 536 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 428, + 482, + 536 + ], + "type": "text", + "content": " increase in accuracy for both rotation and translation. Additionally, our approach also consistently surpasses both baselines in geometry evaluation, as evidenced in the depth and reconstruction error in Tab. 4 and qualitative results in Fig. 6. For the quantitative results for novel view synthesis and additional qualitative results, we refer the readers to see supp. (Sec B)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 555, + 220, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 220, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 220, + 567 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 579, + 482, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 482, + 651 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 482, + 651 + ], + "type": "text", + "content": "In this paper, we examine the benefits of overparameterizing poses via a MLP in the joint optimization task of camera pose and NeRF. We establish that invertibility is a crucial property. We further show that using an Invertible Neural Network, inherently equipped with a guaranteed bijection property, significantly improves the convergence in pose optimization compared to existing representative methods." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 310, + 91, + 447, + 102 + ], + "type": "text", + "content": "Invertible Neural Warp for NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 175, + 482, + 557 + ], + "blocks": [ + { + "bbox": [ + 134, + 175, + 482, + 557 + ], + "lines": [ + { + "bbox": [ + 134, + 175, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 134, + 175, + 482, + 557 + ], + "type": "image", + "image_path": "9ad49b1e37e52ad4388448c6cba6d2879970f095265cf737a94eb56f5e264925.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 564, + 482, + 598 + ], + "lines": [ + { + "bbox": [ + 132, + 564, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 564, + 482, + 598 + ], + "type": "text", + "content": "Fig. 6: We showcase the reconstruction results for scan24, scan55, scan105 and scan106. All meshes are generated using a neural surface reconstruction algorithm called Voxurf [41] using the poses estimated by different approaches." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Shin-Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 128, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 128, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 128, + 482, + 152 + ], + "type": "text", + "content": "Acknowledgement We thank Chee-Kheng (CK) Chng for insightful discussions and technical feedback." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 170, + 197, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 170, + 197, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 170, + 197, + 182 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 194, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 194, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 480, + 228 + ], + "type": "text", + "content": "1. Allen-Zhu, Z., Li, Y., Liang, Y.: Learning and generalization in overparameterized neural networks, going beyond two layers. Advances in neural information processing systems 32 (2019) 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 228, + 480, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 260 + ], + "type": "text", + "content": "2. Arora, S., Du, S., Hu, W., Li, Z., Wang, R.: Fine-grained analysis of optimization and generalization for overparameterized two-layer neural networks. In: International Conference on Machine Learning. pp. 322-332. PMLR (2019) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "type": "text", + "content": "3. Behrmann, J., Grathwohl, W., Chen, R.T., Duvenaud, D., Jacobsen, J.H.: Invertible residual networks. In: International conference on machine learning. pp. 573-582. PMLR (2019) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "type": "text", + "content": "4. Bian, J.W., Bian, W., Prisacariu, V.A., Torr, P.: Porf: Pose residual field for accurate neural surface reconstruction. arXiv preprint arXiv:2310.07449 (2023) 3, 9, 12, 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "type": "text", + "content": "5. Bian, W., Wang, Z., Li, K., Bian, J.W., Prisacariu, V.A.: Nope-nerf: Optimising neural radiance field with no pose prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4160-4169 (2023) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 360, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 360, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 360, + 480, + 392 + ], + "type": "text", + "content": "6. Cai, H., Feng, W., Feng, X., Wang, Y., Zhang, J.: Neural surface reconstruction of dynamic scenes with monocular rgb-d camera. Advances in Neural Information Processing Systems 35, 967-981 (2022) 4, 5, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 423 + ], + "type": "text", + "content": "7. Chen, R.T., Rubanova, Y., Bettencourt, J., Duvenaud, D.K.: Neural ordinary differential equations. Advances in neural information processing systems 31 (2018) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 424, + 480, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 424, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 424, + 480, + 457 + ], + "type": "text", + "content": "8. Chen, Y., Lee, G.H.: Dbarf: Deep bundle-adjusting generalizable neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24-34 (2023) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 458, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 480, + 501 + ], + "type": "text", + "content": "9. Chen, Y., Chen, X., Wang, X., Zhang, Q., Guo, Y., Shan, Y., Wang, F.: Local-to-global registration for bundle-adjusting neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8264-8273 (2023) 1, 2, 3, 7, 8, 10, 11, 12, 13, 14" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 502, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 534 + ], + "type": "text", + "content": "0. Chng, S.F., Ramasinghe, S., Sherrah, J., Lucey, S.: Gaussian activated neural radiance fields for high fidelity reconstruction and pose estimation. In: European Conference on Computer Vision. pp. 264-280. Springer (2022) 1, 3, 5, 8, 10, 11" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 534, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 567 + ], + "type": "text", + "content": "1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "text", + "content": "2. Dinh, L., Sohl-Dickstein, J., Bengio, S.: Density estimation using real nvp. arXiv preprint arXiv:1605.08803 (2016) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 611 + ], + "type": "text", + "content": "3. Guo, M., Fathi, A., Wu, J., Funkhouser, T.: Object-centric neural scene rendering. arXiv preprint arXiv:2012.08503 (2020) 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "text", + "content": "4. Jeffreys, H., Jeffreys, B.: Methods of mathematical physics. Cambridge university press (1999) 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "5. Jensen, R., Dahl, A., Vogiatzis, G., Tola, E., Aanaes, H.: Large scale multi-view stereopsis evaluation. In: 2014 IEEE Conference on Computer Vision and Pattern Recognition. pp. 406-413. IEEE (2014) 10, 12, 13" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 310, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 310, + 91, + 447, + 101 + ], + "type": "text", + "content": "Invertible Neural Warp for NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "16. Jeong, Y., Ahn, S., Choy, C., Anandkumar, A., Cho, M., Park, J.: Self-calibrating neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5846-5854 (2021) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "text", + "content": "17. Jiang, C., Huang, J., Tagliasacchi, A., Guibas, L.J.: Shapeflow: Learnable deformation flows among 3d shapes. Advances in Neural Information Processing Systems 33, 9745-9757 (2020) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "text", + "content": "18. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: Lerf: Language embedded radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19729-19739 (2023) 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "19. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 281 + ], + "type": "text", + "content": "20. Lei, J., Daniilidis, K.: Cadex: Learning canonical deformation coordinate space for dynamic surface representation via neural homeomorphism. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6624-6634 (2022) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "text", + "content": "21. Levoy, M.: Efficient ray tracing of volume data. ACM Transactions on Graphics (TOG) 9(3), 245-261 (1990) 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "22. Li, Y., Liang, Y.: Learning overparameterized neural networks via stochastic gradient descent on structured data. Advances in neural information processing systems 31 (2018) 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "23. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: Barf: Bundle-adjusting neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5741-5751 (2021) 1, 2, 3, 4, 5, 7, 8, 9, 10, 11, 12, 13, 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 370, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 413 + ], + "type": "text", + "content": "24. Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG) 38(4), 1-14 (2019) 10, 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "text", + "content": "25. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) 1, 2, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "text", + "content": "26. Neyshabur, B., Tomioka, R., Srebro, N.: Norm-based capacity control in neural networks. In: Conference on learning theory. pp. 1376-1401. PMLR (2015) 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "27. Nguyen, Q.N., Mondelli, M.: Global convergence of deep networks with one wide layer followed by pyramidal topology. Advances in Neural Information Processing Systems 33, 11961-11972 (2020) 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 502, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 481, + 533 + ], + "type": "text", + "content": "28. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Occupancy flow: 4d reconstruction by learning particle dynamics. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5379-5389 (2019) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "text", + "content": "29. Oymak, S., Soltanolkotabi, M.: Toward moderate overparameterization: Global convergence guarantees for training shallow neural networks. IEEE Journal on Selected Areas in Information Theory 1(1), 84-105 (2020) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "text", + "content": "30. Park, K., Henzler, P., Mildenhall, B., Barron, J.T., Martin-Brualla, R.: Camp: Camera preconditioning for neural radiance fields. ACM Transactions on Graphics (TOG) 42(6), 1-11 (2023) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 642 + ], + "type": "text", + "content": "31. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021) 1, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "32. Parra, A., Chng, S.F., Chin, T.J., Eriksson, A., Reid, I.: Rotation coordinate descent for fast globally optimal rotation averaging. In: Proceedings of the" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "text", + "content": "Shin-Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4298-4307 (2021) 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 183 + ], + "type": "text", + "content": "33. Paschalidou, D., Katharopoulos, A., Geiger, A., Fidler, S.: Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3204-3215 (2021) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "type": "text", + "content": "34. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 481, + 239 + ], + "type": "text", + "content": "35. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research 15(1), 1929-1958 (2014) 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 284 + ], + "type": "text", + "content": "36. Tancik, M., Weber, E., Ng, E., Li, R., Yi, B., Wang, T., Kristoffersen, A., Austin, J., Salahi, K., Ahuja, A., et al.: Nerfstudio: A modular framework for neural radiance field development. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-12 (2023) 10" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 318 + ], + "type": "text", + "content": "37. Truong, P., Rakotosaona, M.J., Manhardt, F., Tombari, F.: Sparf: Neural radiance fields from sparse and noisy poses. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4190-4200 (2023) 3, 10, 11, 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "text", + "content": "38. Wang, C., MacDonald, L.E., Jeni, L.A., Lucey, S.: Flow supervision for deformable nef. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21128-21137 (2023) 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "type": "text", + "content": "39. Wang, Q., Chang, Y.Y., Cai, R., Li, Z., Hariharan, B., Holynski, A., Snavely, N.: Tracking everything everywhere all at once. arXiv preprint arXiv:2306.05422 (2023) 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "40. Wang, Z., Wu, S., Xie, W., Chen, M., Prisacariu, V.A.: Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064 (2021) 1, 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "type": "text", + "content": "41. Wu, T., Wang, J., Pan, X., Xu, X., Theobalt, C., Liu, Z., Lin, D.: Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv preprint arXiv:2208.12697 (2022) 10, 14" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "type": "text", + "content": "42. Xia, Y., Tang, H., Timofte, R., Van Gool, L.: Sinerf: Sinusoidal neural radiance fields for joint pose estimation and scene reconstruction. arXiv preprint arXiv:2210.04553 (2022) 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 487, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 531 + ], + "type": "text", + "content": "43. Xu, C., Wu, B., Hou, J., Tsai, S., Li, R., Wang, J., Zhan, W., He, Z., Vajda, P., Keutzer, K., et al.: Nerf-det: Learning geometry-aware volumetric representation for multi-view 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23320-23330 (2023) 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 531, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 481, + 564 + ], + "type": "text", + "content": "44. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5438-5448 (2022) 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "type": "text", + "content": "45. Yang, G., Belongie, S., Hariharan, B., Koltun, V.: Geometry processing with neural fields. Advances in Neural Information Processing Systems 34, 22483-22497 (2021) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 598, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 481, + 621 + ], + "type": "text", + "content": "46. Zhan, H., Zheng, J., Xu, Y., Reid, I., Rezatofighi, H.: Activermap: Radiance field for active mapping and planning. arXiv preprint arXiv:2211.12656 (2022) 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "47. Zhao, F., Yang, W., Zhang, J., Lin, P., Zhang, Y., Yu, J., Xu, L.: Humannerf: Efficiently generated human radiance field from sparse inputs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7743-7753 (2022) 1" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 310, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 310, + 91, + 447, + 101 + ], + "type": "text", + "content": "Invertible Neural Warp for NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 482, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 482, + 162 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 482, + 162 + ], + "type": "text", + "content": "48. Zhu, Z., Peng, S., Larsson, V., Xu, W., Bao, H., Cui, Z., Oswald, M.R., Pollefeys, M.: Nice-slam: Neural implicit scalable encoding for slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12786-12796 (2022) 1" + } + ] + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Shin-Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Investigating Style Similarity in Diffusion Models/e6cae220-dbdc-4ecf-90b6-28f1684a0527_content_list.json b/2024/Investigating Style Similarity in Diffusion Models/e6cae220-dbdc-4ecf-90b6-28f1684a0527_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ba31f51957739f0c3270c8c2c7174f02423a52cd --- /dev/null +++ b/2024/Investigating Style Similarity in Diffusion Models/e6cae220-dbdc-4ecf-90b6-28f1684a0527_content_list.json @@ -0,0 +1,1750 @@ +[ + { + "type": "text", + "text": "Investigating Style Similarity in Diffusion Models", + "text_level": 1, + "bbox": [ + 218, + 141, + 782, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Growthami Somepalli\\*1, Anubhav Gupta\\*1, Kamal Gupta\\*1, Shramay Palta\\*1, Micah Goldblum\\*3, Jonas Geiping\\*2, Abhinav Shrivastava\\*1, and Tom Goldstein\\*1", + "bbox": [ + 215, + 189, + 782, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Maryland, College Park", + "bbox": [ + 341, + 231, + 658, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 ELLIS Institute, MPI for Intelligent Systems", + "bbox": [ + 341, + 247, + 656, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Columbia University", + "bbox": [ + 424, + 260, + 576, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 https://somepago.github.io/csd", + "bbox": [ + 375, + 273, + 625, + 287 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/33ccc3a732cd6c20c46d042c462473ee361023e66ad35bd3cfacabe4ce8c2a3c.jpg", + "image_caption": [ + "Real paintings" + ], + "image_footnote": [], + "bbox": [ + 222, + 340, + 328, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/97a368689dba65163f359a2ddd324c8f8366910869cc620896f9acb975979c3e.jpg", + "image_caption": [ + "MidJourney" + ], + "image_footnote": [], + "bbox": [ + 334, + 340, + 413, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/58b9eb52ab258da014fbd42d2055b40dff2cf6ac31cce68fd30143b1ed26c538.jpg", + "image_caption": [ + "2.1" + ], + "image_footnote": [], + "bbox": [ + 416, + 340, + 470, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1d858238181bbb2985a5bd8d48cb438624b1201bb35627e8d17040bc094cc814.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 472, + 340, + 500, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cd738bf053c4382f2f80e478bdbe15cabffb239c63dcf22560a357ad1ed31068.jpg", + "image_caption": [ + "1.4" + ], + "image_footnote": [], + "bbox": [ + 501, + 330, + 588, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/39763cba641acaee54e33b5e59d0525927c4d63ae7acd1e59378a95926d2bbd8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 330, + 663, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/67adbfc7450fdcb80e7c411bd2eebad2543abefc10dbd3ef913e8d7c999351cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 330, + 723, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/291721b4c89ab6c5a699ee8eb49f34467111c717a3fe898255e9760867c157a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 330, + 782, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/87d5451b45700ba592927bd1134d0bf7a0cc9530c5f6bcf85cc98371b354d413.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 387, + 330, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/779265e8f59b906ad8956b13068863ba4e4e0414a34dc96abaa465bb44c5ed61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 387, + 413, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8c19ca291a8fe7289f6c4e1fc3cdcec79a4559d92144cdb71796b4e3ab6b4799.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 387, + 498, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/54b65c24568733d08bf36729c68d5f90abd786d69ca35bbc2b409770a39bf80c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 387, + 609, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9ef29bb886a569ef72659c0c3b6f742ac04159e40350b6605e009cdb8270d01d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 387, + 720, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c3811ab33825d3a5417f1c3c934e3a2c27223fd66be2625135dd0a39af80bc48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 387, + 779, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f90092ba02fc4179b6e64c9c5d8d9da4eb9a67c6864641c13e4a124b24e2c59f.jpg", + "image_caption": [ + "Fig. 1: Original artwork of 6 popular artists and the images generated in the style of these artists by three popular text-to-image generative models. The numbers displayed below each image indicate the similarity of the generated image with the artist's style using the proposed method. A high similarity score suggests a strong presence of the artist's style elements in the generated image. Based on our analyses, we postulate that three artists on the right were removed (or unlearned) from SD 2.1 while they were present in MidJourney and SD 1.4. Please refer to Section 2 for more details." + ], + "image_footnote": [], + "bbox": [ + 222, + 435, + 330, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b8b2038b49eab9cba87f72bad42d84fd78c45ce931333a3253a4eaca435862a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 435, + 413, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f6e27c024edf7205ff7698356d0c36f38defa1885e25c7f5616aaffdc91817b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 435, + 498, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/96abf52c86408e3929f388ef765e9007bdad54267970aeee97e50cbfc87415af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 435, + 609, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2e29ad371079ba05a0e8ae8b2b2e4f623cd447a0ac9c3840b1a5fd4bb9b9cdaa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 435, + 694, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eaceb94b33dbc00f9f421cea387eee0934b0a4c3c5d6fc85d27a3afa1d5ed335.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 435, + 779, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Generative models are now widely used by graphic designers and artists. Prior works have shown that these models remember and often replicate content from their training data during generation. Hence as their proliferation increases, it has become important to perform a database search to determine whether the properties of the image are attributable to specific training data, every time before a generated image is used for professional purposes. Existing tools for this purpose focus on retrieving images of similar semantic content. Meanwhile, many artists are concerned with style replication in text-to-image models. We present a framework for understanding and extracting style descriptors from images. Our framework comprises a new dataset curated using the", + "bbox": [ + 259, + 648, + 740, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. Correspondence: growthami@umd.edu.", + "bbox": [ + 232, + 825, + 617, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "insight that style is a subjective property of an image that captures complex yet meaningful interactions of factors including but not limited to colors, textures, shapes, etc. We also propose a method to extract style descriptors that can be used to attribute style of a generated image to the images used in the training dataset of a text-to-image model. We showcase promising results in various style retrieval tasks. We also quantitatively and qualitatively analyze style attribution and matching in the Stable Diffusion model.", + "bbox": [ + 259, + 147, + 743, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: Image Style $\\cdot$ Style Similarity $\\cdot$ Generative models", + "bbox": [ + 261, + 272, + 684, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 316, + 375, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion-based image generators like Stable Diffusion [49], DALL-E [47] and many others [1, 7, 39, 43] learn artistic styles from massive captioned image datasets that are scraped from across the web [54]. Before a generated image is used for commercial purposes, it is wise to understand its relationship to the training data and the origins of its design elements and style attributes. Discovering and attributing these generated images, typically done with image similarity search, is hence becoming increasingly important. Such dataset attribution serves two purposes. It enables users of generated images to understand potential conflicts, associations, and social connotations that their image may evoke. It also enables artists to assess whether and how generative models are using elements of their work.", + "bbox": [ + 212, + 353, + 787, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite a long history of research [60], recovering style from an image is a challenging and open problem in Computer Vision. Many existing retrieval methods [8, 45, 46] for large training datasets focus primarily on matching semantic content between a pair of images. Understanding the origin of the style present in a generated image, however, is much less well understood. To address this gap, we propose a self-supervised objective for learning style descriptors from images. Standard augmentation-based SSL pipelines (e.g. SimCLR and variants) learn feature vectors that are invariant to a set of augmentations. Typically, these augmentations preserve semantic content and treat style as a nuisance variable. In contrast, we choose augmentations that preserve stylistic attributes (colors, textures, or shapes) while minimizing content. Unfortunately, SSL is not enough, as style is inherently subjective, and therefore a good style extractor should be aligned with human perceptions and definitions of style. For this reason, we curate a style attribution dataset, ContraStyles, in which images are associated with the artist that created them.", + "bbox": [ + 212, + 521, + 787, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By training with both SSL and supervised objectives, we create a high-performance model for representing style. Our model, CSD, outperforms other large-scale pre-trained models and prior style retrieval methods on standard datasets. Using CSD, we examine the extent of style replication in the popular open-source text-to-image generative model Stable Diffusion [49], and consider different factors that impact the rate of style replication.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "G. Somepalli et al.", + "bbox": [ + 271, + 114, + 395, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize our contributions, we (1) propose a style attribution dataset ContraStyles, associating images with their styles, (2) introduce a multi-label contrastive learning scheme to extract style descriptors from images and show the efficacy of the scheme by zero-shot evaluation on public domain datasets such as WikiArt and DomainNet (3) We perform a style attribution case study for one of the most popular text-to-image generative models, Stable Diffusion, and propose indicators of how likely an artist's style is to be replicated. Code, artifacts, and dataset links are available at https://somepago.github.io/csd.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Motivation", + "text_level": 1, + "bbox": [ + 214, + 289, + 359, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present a case study that shows how style features can be used to interrogate a generative model, and provide utility to either artists or users. We consider the task of analyzing a model's ability to emulate an artist's style, and of attributing the style of an image to an artist. We begin by curating a list of 96 artists, primarily sourced from the WikiArt database, supplemented by a few contemporary artists who are notably popular within the Stable Diffusion community5. For each artist, we compute a prototype vector by averaging the embeddings of their paintings using our proposed feature extractor, CSD ViT-L. Next, we generate an image for each artist using Stable Diffusion 2.1 with a prompt in the format A painting in the style of . We compute the dot product similarity between each generated image's embedding and the artist's prototype. This process was repeated multiple times for each artist, and we plot mean results in Fig. 2. We refer to this quantity as the General Style Similarity (GSS) score for an artist, as it measures how similar a generated image is to a typical image from that artist while using our style representation model. We also plot an analogous style similarity score, but using \"content-constrained\" prompts. For instance, one prompt template is A painting of a woman doing style of where X is the name of the artist and Y is some setting like reading a book or holding a baby etc. See Sec. 7 for all templates.", + "bbox": [ + 212, + 318, + 787, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each point in Fig. 2 represents an artist. Notice that GSS scores are highly correlated with content-constrained scores, indicating that our feature vectors represent style more than semantic content. Our findings reveal that SD 2.1 is much more capable of emulating some artists than others. Artists like Leonid Afremov, Georges Seurat exhibit high style similarity scores, and visual inspection of generated images confirms that indeed their style is emulated by the model (Fig. 1 - Original artwork vs SD 2.1). On the other end of the spectrum, artists such as Ruan Jia and Greg Rutkowski showed low similarity scores, and likewise the generated images bear little resemblance to the artists' work. Interestingly, after completing this study, we discovered that Greg Rutkowski's work was excluded from the training data for the Stable Diffusion 2.1 model, as reported by [11].", + "bbox": [ + 212, + 606, + 787, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This demonstrates that the Style Similarity score can be used by artists to quantify how well a model emulates their style, or it can be used by users to", + "bbox": [ + 212, + 787, + 787, + 818 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Investigating Style Similarity in Diffusion Models", + "bbox": [ + 401, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "5 https://supagruen.github.io/StableDiffusion-CheatSheet/", + "bbox": [ + 217, + 824, + 661, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ascertain whether a generated image contains stylistic elements associated with a particular artist. After a thorough inspection of the generations from 96 artists, we hypothesize that a single-image Style Similarity score below 0.5 indicates the absence of the artist's style, whereas a score above 0.8 strongly indicates its presence.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Figure 1, we show original artworks for 6 artists, and generations from MidJourney [39], Stable Diffusion 2.1 and Stable Diffusion 1.4 [49] for each of these artists. The 3 artists on the left side have high GSS while the ones on the right side have low GSS. Below each generated image, we display the similarity against the artist's prototype vector. We see high image similarity scores in the Midjourney generations and qualitatively these images look stylistically similar to artists' original artworks. We also see the interesting cases of Greg Rutkowski, Ruan Jia, and Amano whose style is captured by Stable Diffusion 1.4, while being notably absent in Stable Diffusion 2.1. This finding is in line with reports suggesting that some of these artists were removed from the training data", + "bbox": [ + 212, + 223, + 488, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of Stable Diffusion 2.1 [11]. Based on this analysis, we postulate that Ruan Jia, Wadim Kashin, Anton Fadeev, Justin Gerard, and Amano were also either excluded from the training data or post-hoc unlearned/removed from Stable Diffusion 2.1.", + "bbox": [ + 212, + 555, + 787, + 617 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d1f6e6683b8888ef099762a2ff03471432b73f50a7d1ebc04caf8b92ae5584d5.jpg", + "image_caption": [ + "Fig. 2: Style similarity of Stable Diffusion 2.1 generated images against the artist's prototypical representation. On the X-axis, the similarities are depicted when the prompt is not constrained, while the Y-axis represents similarity when the prompt is constrained to generate an image of a \"woman\" in the artist's style." + ], + "image_footnote": [], + "bbox": [ + 506, + 257, + 774, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 What is style?", + "text_level": 1, + "bbox": [ + 215, + 646, + 393, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The precise definition of \"style\" remains in contention, but many named artistic styles (e.g., cubism, impressionism, etc...) are often associated with certain artists. We leverage this social construct, and define style simply as the collection of global characteristics of an image that are identified with an artist or artistic movement. These characteristics encompass various elements such as color usage, brushstroke techniques, composition, and perspective.", + "bbox": [ + 212, + 686, + 787, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Related work. Early computer vision algorithms attempted to model style using low-level visual features like color histograms, texture patterns, edge detection, and shape descriptors. Other computational techniques involve rule-based systems, such as the presence of specific compositional elements, the use of spe", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "G. Somepalli et al.", + "bbox": [ + 271, + 114, + 395, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cific color palettes, or the presence of certain brushstroke patterns to identify specific style characteristics [19, 20, 24, 25, 29, 32, 35, 37, 51, 55, 59, 65, 67].", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Modern studies have focused on the task of transferring style from one image to another [13, 17, 22, 34, 41, 62, 68]. Some works have also concentrated on style classification [2, 4, 10, 15, 27, 28, 30, 33, 38, 48, 53]. A limited number of studies address in-the-wild style quantification, matching, and retrieval [14, 23, 36, 50, 66]. In their seminal work, Gatys et al. [17] introduced Gram Matrices as style descriptors and utilized an optimization loop to transfer style. Another approach proposed by Luan et al. [34] includes a photorealism regularization term to prevent distortions in the reconstructed image. Zhang et al. [68] formulated style transfer using Markov random fields. Beyond Gram-based style representation, Chu et al. [10] explored various other types of correlations and demonstrated performance variations.", + "bbox": [ + 212, + 176, + 785, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In a recent work by Lee et al [31], two separate neural network modules were used - one for image style and another for image content - to facilitate image style retrieval. In the most recent related research, Wang et al. [63] developed an attribution model trained on synthetic style pairs, designed to identify stylistically similar images. In contrast to this approach, our method leverages real image pairs, curated automatically through their caption annotations. Despite our training dataset being approx. $16\\%$ the size of training data used in [63], we demonstrate that our model significantly outperforms this method on many zero-shot style matching tasks in the later sections.", + "bbox": [ + 212, + 343, + 787, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Creating a dataset for style attribution", + "text_level": 1, + "bbox": [ + 214, + 502, + 629, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While many large web datasets now exist, we need one that contains wide variations in artistic styles, and also labels that be used for downstream style retrieval evaluation. Some large-scale datasets specifically designed to handle such a challenge, like BAM [64], are not available in the public domain and others like WikiArt [52] are not large enough to train a good style feature extractor. In the following section, we propose a way to curate a large style dataset out of the LAION [54] Aesthetics $6+$ dataset.", + "bbox": [ + 212, + 534, + 787, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ContraStyles: A dataset for style distillation. We curate our own dataset as a subset of LAION [54]. We start off with the 12M image-text pairs with predicted aesthetics scores of 6 or higher. Note that this dataset is extremely unbalanced, with some popular artists appearing much more frequently than others. Also, a large number of images are duplicated within the dataset which is a major issue for the text-to-image models trained on this data [57,58]. Furthermore, the image captions within the data are often noisy and are often missing a lot of information. We address these challenges and propose a new subset of LAION-Aesthetics consisting of 511,921 images, and 3840 style tags, where each image can have one or more tags. We use this dataset for training our models.", + "bbox": [ + 212, + 643, + 787, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We begin with a bank of styles collated in previous work for image understanding with the CLIP Interrogator [44]. This bank of styles was curated based on typical user prompts for Stable Diffusion. We combine the bank of artists,", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Investigating Style Similarity in Diffusion Models", + "bbox": [ + 401, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "medium, and movement references, to a shortlist of 5600 tags. We then search for these tags in the 12M LAION-Aesthetics captions and shortlist the images that have at least one of the tags present. We further filter out the tags which have over 100,000 hits in the dataset since human inspection found that they refer to common phrases like 'picture' or 'photograph' that do not invoke a distinct style. After discarding images with an unavailable URL, we are left with about 1 million images and 3840 tags. We further deduplicate the images using SSCD [45] with a threshold of 0.8 and merge the tags of images that are near copies of each other. As a by-product, the dedduplication also helps with the missing tags in the images, since we can simply merge the text labels of duplicate images. After dedduplication, we are left with 511,921 images.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5 Contrastive Style Descriptors (CSD)", + "text_level": 1, + "bbox": [ + 215, + 349, + 604, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Self-Supervised Learning. Many successful approaches to SSL [61] use a contrastive [9] approach, where two views (or augmentations) of the same image in the dataset are sampled and passed to a deep network to get their respective image descriptors. The network is trained to maximize the similarity of two views of the same image and minimize agreement with other images in the batch. Standard choices for augmentations include color jitter, blurring, grayscale, etc., as these alter the image's visual properties while preserving content. While these are good augmentations for object recognition tasks, they train the network to ignore image attributes associated with style.", + "bbox": [ + 212, + 393, + 787, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our approach relies on a training pipeline with two parts. First, we use contrastive SSL, but with a set of augmentations that are curated to preserve style. Second, we align our model with human perceptions of style by training on our labelled ContraStyles dataset described in Section 4.", + "bbox": [ + 212, + 532, + 787, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proposed Approach. We seek a model for extracting image descriptors that contain concise and effective style information. To be a useful, the model should be invariant to semantic content and capable of disentangling multiple styles.", + "bbox": [ + 212, + 608, + 785, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given a dataset of $N$ labeled images $\\{\\mathbf{x}_i, l_i\\}_{i=1}^N$ , where each image can have one or more labels from a set of $L$ labels, we define the label vector of the $i^{\\text{th}}$ image as $l_i = (c_1, c_2, \\ldots, c_L)$ , where each $c_k \\in \\{0, 1\\}$ . As mentioned in the previous section, our multi-label dataset consists of $N = 511,921$ images and $L = 3,840$ style tags. We consider a mini-batch of $B$ images. Each of the images are passed to a Vision Transformer (ViT) [12] backbone, and then projected to a $d$ -dimensional vector. We consider two variants of ViT (ViT-B and ViT-L).", + "bbox": [ + 212, + 655, + 787, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our style descriptors $f_{\\mathrm{ViT}}(\\mathbf{x}_i) \\in \\mathbb{R}^d$ are then used to create a matrix of pairwise cosine similarity scores $s_{i,j} = \\cos (f_{\\mathrm{ViT}}(\\mathbf{x}_i), f_{\\mathrm{ViT}}(\\mathbf{x}_j))$ . In order to compute our multi-label contrastive loss (MCL), we also compute the groundtruth similarity scores as $\\hat{s}_{i,j} = \\mathbb{1}\\left(l_i^{\\mathrm{T}}l_j\\right)$ , where $\\mathbb{1}$ is the indicator function that returns 1 if any of the labels of the images $i, j$ match. Our final loss term reduces to:", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "G. Somepalli et al.", + "bbox": [ + 271, + 114, + 395, + 128 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M C L}} = - \\hat {s} _ {i, j} \\log \\frac {\\exp \\left(s _ {i , j}\\right) / \\tau)}{\\sum_ {k \\neq j} \\exp \\left(s _ {i , k}\\right) / \\tau}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 171, + 784, + 214 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\tau$ is the temperature fixed during the training.", + "bbox": [ + 215, + 226, + 598, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since our supervised dataset is modest in size, we add a self-supervised objective. We sample two \"views\" (augmentations) of each image in a batch and add a contrastive SSL term. Standard SSL training routines (e.g., MoCo, SimCLR, BYOL etc.) choose augmentations so that each pair of views has the same semantic content, but different style content. These augmentations typically include Resize, Horizontal Flips, Color Jitter, Grayscale, Gaussian Blur, and Solarization [6, 18]. For our purposes, we depart from standard methods by excluding photometric augmentations (Gaussian Blur, Color Jitter), as they alter the style of the image. We keep the following spatial augmentations - Horizontal Flips, Vertical Flips, Resize and Rotation as they keep style intact.", + "bbox": [ + 212, + 241, + 784, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The overall loss function is a simple combination of the multilabel contrastive loss and self-supervised loss $\\mathcal{L} = \\mathcal{L}_{\\mathrm{MCL}} + \\lambda \\mathcal{L}_{\\mathrm{SSL}}$ . During inference, we use the final layer embedding and the dot product to compute style similarity between any two images. In our experiments, we found that initializing weights to CLIP [46] ViT-B and ViT-L improves performance.", + "bbox": [ + 212, + 392, + 785, + 468 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6 Results", + "text_level": 1, + "bbox": [ + 215, + 494, + 323, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training details. We present the results for two variants of our model CSD ViT-B and CSD ViT-L version. Both the models are initialized with respective CLIP variant checkpoints and are finetuned for 80k iterations on the ContraStyles dataset on 4 A4000/A5000 GPUs. We use an SGD optimizer with momentum 0.9 and learning rate of 0.003 for the projection layer and $1e - 4$ for the backbone. Our mini-batch size per GPU is 16. We use $\\lambda = 0.2$ and $\\tau = 0.1$ for the final model. The training takes about 8 hours for the base model and around 16 hours for the large model. See the Appendix for more details and ablations.", + "bbox": [ + 212, + 527, + 784, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Task. We perform zero-shot evaluation across multiple datasets on a style-retrieval task. Following [3, 26], we split each dataset into two parts: Database and Query. Given a query image at test time, we evaluate whether we can find the ground-truth style in its nearest neighbors from the database.", + "bbox": [ + 212, + 669, + 785, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We compare our model against a recent style attribution model GDA [63] which is trained via fine-tuning on paired synthetic style data, and VGG [17, 56] Gram Matrices which are often used for neural style transfer applications. Further we compare with CLIP [46] models supervised with freeform text captions, and with other self-supervised models such as DINO [8], MoCo [21], SSCD [45]. We use the embeddings from the last layer for each of these models except for VGG where we use the Gram Matrix [17] of the last", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Investigating Style Similarity in Diffusion Models", + "bbox": [ + 401, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "layer. We skipped evaluations of [14, 23, 50] since both pre-trained models and training data are not available.", + "bbox": [ + 212, + 145, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Metrics. We do nearest neighbour searches for $k \\in [1, 10, 100]$ , and report Recall@k, mAP@k. We use the standard definitions of these metrics from the retrieval literature. Like [5], we define positive recall as the existence of a correct label in top-N matches and no recall when none of the top-N matches share a label with the query. Similarly, mAP is defined as average over precision at each rank in N, and then averaged over all queries.", + "bbox": [ + 212, + 181, + 785, + 273 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5e901bfdbad898448f5223b66f99aa351615d7cb8f3b70d513da01f726767580.jpg", + "table_caption": [ + "Table 1: mAP and Recall metrics on DomainNet and WikiArt datasets. Our model consistently performs the best in all cases except one, against both self-supervised and style attribution baselines." + ], + "table_footnote": [], + "table_body": "
DomainNet (mAP@k)WikiArt (mAP@k)DomainNet (Recall@k)WikiArt (Recall@k)
Method110100110100110110100
VGG Gram [16]---25.919.411.4--25.952.780.4
DINO ViT-B/16 [8]69.468.266.244.033.418.969.493.744.069.488.1
DINO ViT-B/8 [8]72.270.969.346.935.920.472.293.846.971.088.9
SSCD RN-50 [45]67.665.962.036.026.514.867.695.036.062.185.4
MOCO ViT-B/16 [21]71.971.169.644.033.218.872.094.044.069.088.0
CLIP ViT-B/16 [46]73.773.071.352.242.026.073.794.552.278.393.5
GDA CLIP ViT-B [63]62.961.659.325.621.014.162.992.325.656.683.8
GDA DINO ViT-B [63]69.568.166.145.534.619.769.593.445.575.889.0
GDA ViT-B [63]67.165.664.242.632.218.267.193.642.667.687.1
CSD ViT-B (Ours)78.377.576.056.246.128.778.394.356.280.393.6
CLIP ViT-L [46]74.073.572.259.448.831.574.094.859.482.995.1
CSD ViT-L (Ours)78.377.876.564.5653.8235.6578.394.564.5685.7395.58
", + "bbox": [ + 220, + 354, + 782, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation Datasets. DomainNet [42] consists of an almost equal number of images from six different domains: Clipart, Infograph, Painting, Quickdraw, Real, and Sketch. Upon examination, we observed a strong stylistic resemblance between the Quickdraw and Sketch domains, leading us to exclude Quickdraw from our analysis. The dataset's content information was utilized to categorize the images into two main clusters of content classes. This clustering was achieved through the application of word2vec [40]. The images within the smaller cluster were designated as part of the Query set (20,000 images), and images in the bigger cluster to Database (206768 images). And the second dataset we evaluate all the models is, WikiArt [52]. It consists of 80096 fine art images spread across 1119 artists and 27 genres. We randomly split the dataset into 64090 Database and 16006 Query images. We use the artist as a proxy for the style since there is large visual variation within each genre for them to be considered as independent styles. Under this setting, WikiArt is a challenging retrieval task as the chance probability of successful match is just .09% while it is 20% for DomainNet.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "G. Somepalli et al.", + "bbox": [ + 271, + 114, + 395, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.1 Analyses and observations", + "text_level": 1, + "bbox": [ + 215, + 146, + 478, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In Table 1, we report the metrics for all the baselines considered and the proposed CSD method using $k$ nearest neighbors on 2 datasets - DomainNet and WikiArt. Note that while style and content are better separated in the case of DomainNet, WikiArt consists of more fine-grained styles and has more practical use cases for style retrieval. Loosely speaking, mAP@ $k$ determines what percentage of the nearest neighbors that are correct predictions, while the recall determines what percentage of queries has a correct match in the top- $k$ neighbors. Our model CSD consistently outperformed all the pretrained feature extractors as well as the recent attribution method GDA [63] on both WikiArt and DomainNet evaluations. Note that all models are evaluated in a zero-shot setting. We see the most gains in the WikiArt dataset which is more challenging with chance probability of only $0.09\\%$ . When we look at mAP@1, which is same as top-1 accuracy, our base model outperforms the next best model by $5\\%$ points on WikiArt and $4.6\\%$ points on DomainNet. Our large model out-performs the closest large competitor by similar margins. Given the complexity of the task, these improvements are non-trivial.", + "bbox": [ + 212, + 185, + 787, + 426 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We attribute the improvements to a couple of factors, (1) The multi-label style contrastive loss on our curated ContraStyles dataset is quite helpful in teaching the model right styles (2) We hypothesize that these SSL models become invariant to styles because the way they were trained, but we are careful to not strip that away in our SSL loss component by carefully curating non-photometric augmentations in training.", + "bbox": [ + 212, + 429, + 787, + 521 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Error Analysis. Even though our model outperforms the previous baselines, our top-1 accuracy for the WikiArt style matching task is still at 64.56. We tried to understand if there is a pattern to these errors. For example, our model is consistently getting confused between impressionist painters claude monet, gustave loiseau, and alfred sisley, all of whom painted many landscapes. They depicted natural scenes, including countryside views, rivers, gardens, and coastal vistas. Another example is pablo picasso and georges braque, who are both cubist painters. Given the impracticality of analyzing all 1,119 artists in the dataset, we opted for a macroscopic examination by categorizing errors at the art movement level. This approach is visualized in the heatmap presented in Fig. 3. In the heatmap, we see most of the errors concentrated along the diagonal, indicating that while the model often correctly identifies the art movement, it struggles to pinpoint the exact artist. There are instances of off-diagonal errors where the model incorrectly identifies both the artist and their art movement. For example, many Post Impressionism and Realism paintings are assigned to Impressionism artists. Upon closer examination, it becomes apparent that they closely align in terms of historical timeline and geographical origin, both being from Europe. This analysis indicates the nuanced nature of style detection and retrieval in Computer Vision. It suggests that the upper limit for accuracy in this task might be considerably lower than $100\\%$ , even for a typical human evaluator, due to the inherent subtleties and complexities involved.", + "bbox": [ + 212, + 523, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Investigating Style Similarity in Diffusion Models", + "bbox": [ + 401, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e9748336a4e095dc167b6706602aea38f9aa6307127c1b62f7a51cf675d7a6c1.jpg", + "image_caption": [ + "Fig. 3: Confusion Matrix of errors in WikiArt: Art movements are predicted correctly. Errors occur in cases where movements share the same historical timeline and/or are derived from the same earlier movement." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 584, + 292 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/13e130720ef7a42129e74a87bf69aa76475dc661dd963502fcd7ba0801414e5f.jpg", + "image_caption": [ + "Fig. 4: Human study on Style Retrieval: Turns out untrained humans are worse than many feature extractors on matching images from the same artist." + ], + "image_footnote": [], + "bbox": [ + 602, + 176, + 785, + 265 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6.2 Human Study", + "text_level": 1, + "bbox": [ + 215, + 392, + 377, + 409 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To understand how our models compare to untrained humans, we conducted a small survey on style matching on 30 humans (excluding the authors). Following the convention in other papers [14,28,50,62] and this paper, we assume, 2 images from same artist can be considered stylistic matches. For each query image, we gave 4 answer images out of which only one is from the same artist and hence is the right answer, so chance accuracy is $25\\%$ . We used the Artcive dataset introduced in [63] to create this survey and we collected 3 responses per item to break any ties. We present the results in Fig 4. Most interestingly, untrained humans are worse than many feature extractors at this task. SSCD is the only feature extractor that did worse than humans. Our model, CSD outperforms all the baselines on this dataset as well. This underpins the difficulty of style matching and also highlights the superior performance of our feature extractor.", + "bbox": [ + 212, + 426, + 787, + 609 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "7 Studying style in the wild: Analysis of Stable Diffusion", + "text_level": 1, + "bbox": [ + 215, + 640, + 782, + 659 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In the previous section, we have quantitatively shown that our model Contrastive Style Descriptors outperforms many baselines on style matching task. Now we try to address the question, Can we do style matching on Stable Diffusion generated images? To answer this question, we first curated multiple synthetic image collections using Stable Diffusion v 2.1 [49] and then compared them against the \"ground truth\" style matches on ContraStyles dataset.", + "bbox": [ + 212, + 681, + 787, + 773 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Creating synthetic style dataset. The first challenge in curating synthetic images through prompts is the choice of prompts to be used for the generation. There have been no in-depth quantitative studies of the effect of prompts on generation styles. For this analysis, we chose 3 types of prompts.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "G. Somepalli et al.", + "bbox": [ + 271, + 114, + 395, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/68ce7432468125cc5da930f877257e5af71c2bf4989f7206201497189c99246f.jpg", + "table_caption": [ + "Table 2: mAP and Recall of SD 2.1 generated synthetic datasets based on Simple prompts and User-generated prompts" + ], + "table_footnote": [], + "table_body": "
Simple prompts (mAP@k)User-generated (mAP@k)Simple prompts (Recall@k)User-generated (Recall@k)
Method110100110100110100110100
GDA - DINO11.610.27.64.454.594.2411.628.152.834.4525.2267.18
CSD-ViTB17.5316.5612.685.855.965.5817.5338.6561.855.8529.2674.2
CLIP ViT-L/1422.320.416.16.15.75.122.344.566.26.126.071.7
CSD (Ours)24.523.318.55.75.95.624.547.267.55.726.571.8
", + "bbox": [ + 227, + 169, + 782, + 272 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. User-generated prompts: We used a Stable Diffusion Prompts $^6$ dataset of 80,000 prompts filtered from Lexica.art. We used the test split and then filtered the prompts to make sure at least one of the keywords from the list we curated in Section 4 is present. We then sampled 4000 prompts from this subset for query split generation.", + "2. Simple prompts: We randomly sampled 400 artists which appeared most frequently in user-generated prompts we analysed. We format the prompt as A painting in the style of , and we generate 10 images per prompt by varying the initialization seed.", + "3. Content-constrained prompts: We wanted to understand if we can detect style when we constrain the model to generate a particular subject/human in the style of an artist. For this, we used the prompt A painting of a woman in the style of or A painting of a woman reading in the style of etc., a total of 5 variations per subject repeated two times. We experimented with subjects, woman,dog and house in this study. We provide the exact templates in the appendix." + ], + "bbox": [ + 215, + 297, + 792, + 540 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/0ed13df6e1a94a1526dfc56c76c1038a27e0261e870a3265f2d342906a09a140.jpg", + "table_caption": [ + "Table 3: mAP and Recall of SD 2.1 generated synthetic datasets based on Content-constrained prompts" + ], + "table_footnote": [], + "table_body": "
Dog (mAP@k)House (mAP@k)Woman (mAP@k)Dog (Recall@k)House (Recall@k)Woman (Recall@k)
Method110100110100110100110100110100110100
GDA-DINO2.2821.63.93.62.72.22.322.288.6828.733.912.3322.210.128.9
CSD-ViT-B4.54.313.614.64.293.877.557.836.424.514.3634.884.615.03397.5520.142.46
CLIP ViT-L/142.32.21.94.54.23.67.47.16.22.29.829.94.513.835.37.419.041.6
CSD (Ours)4.94.84.26.46.25.410.810.18.64.914.534.56.417.840.610.823.444.2
", + "bbox": [ + 218, + 614, + 782, + 714 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We generate 4000 images for each prompt setting using Stable Diffusion 2.1. There is only one style keyword in simple and content-constrained prompts, which we also use as a ground truth label for matching tasks. However, user-generated prompts can have multiple style labels within the caption, and we consider all of them as ground-truth labels.", + "bbox": [ + 212, + 741, + 787, + 816 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Investigating Style Similarity in Diffusion Models", + "bbox": [ + 401, + 114, + 732, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_footnote", + "text": "$^{6}$ https://huggingface.co/datasets/Gustavosta/Stable-Diffusion-Prompts", + "bbox": [ + 217, + 824, + 754, + 840 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c7ab3dff347505d675fe37a2fbf2c41836ec36b8fd0730ccbafb668bcd65387c.jpg", + "image_caption": [ + "Fig. 5: Nearest \"style\" neighbors. For each generated image (referred to as SD Gen), we show the top 5 style neighbors in CSD using our feature extractor. The green and red box around the image indicates whether or not the artist's name used to generate the SD image was present in the caption of the nearest neighbor." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 496, + 281 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3fb9c16ec1ce4d7f28c2b7738a291cfa042fe75fda090ae99cd0d4f3452a9f62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 143, + 785, + 281 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/46379fab2ec7ba3ad4f5608dc6fe95862b66fe6a4499acfe4df62fdc8e24a554.jpg", + "image_caption": [ + "Fig. 6: Top row: Images generated by Stable Diffusion. Middle and Bottom rows: Top matches retrieved by CLIP vs CSD (ours) respectively. CLIP is consistently biased towards image content, for instance retrieving image of a dog in the Column 1, 3, 4, or image of mother and baby in Column 7 or 8. Our method emphasize less on the content but more on the image styles. Please refer to the Appendix for the prompts." + ], + "image_footnote": [], + "bbox": [ + 222, + 368, + 782, + 526 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Style retrieval on generated images. In Tab. 2, we show the retrieval results for Simple and User-generated prompts. We also compare our results with the second-best performing model in the previous section, CLIP ViT-L, and a recent style attribution model GDA [63]. We observe that our method outperforms CLIP on Simple prompt dataset. For User-generated prompts, the performance metrics are closer to CLIP model, but it's important to note that these prompts are inherently more complex. This complexity results in a different label distribution in the query set for the two types of prompts we examine, leading to varied metric ranges in each case. Additionally, our method consistently outperforms both baselines in content-constrained scenarios, as evidenced in Table 3. This indicates the robustness and effectiveness of our approach in dealing with a variety of prompt complexities and content specifications. We refer the reader to Appendix to understand a few caveats of this quantitative study.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "G. Somepalli et al.", + "bbox": [ + 271, + 114, + 395, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative Results. In Fig. 5, we showcase a selection of Stable Diffusion-generated images alongside their top 5 corresponding matches from ContraStyles, as determined by the CSD ViT-L feature extractor. The left section of the figure displays images generated from User-generated prompts, while the right section includes images created from Simple prompts. To aid in visual analysis, matches that share a label with the query image are highlighted in green. We can clearly see that the query image and the matches share multiple stylistic elements, such as color palettes and certain artistic features like motifs or textures. We observed that in generations based on user-generated prompts, perceivable style copying typically occurs only when the prompts are shorter and contain elements that are characteristic of the artist's typical content.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Figure 6, we present several content-constrained prompt generations and their top-1 matches based on the CLIP ViT-L/14 model versus our CSD model. We observe that the CSD model accurately matches the correct artists to queries even when there is no shared content, only style. This is evident in columns 1, 3, 4, 7, and 8, where our model, CSD matches the correct style elements despite the subjects in the images being quite different. In contrast, the CLIP model still prioritizes content, often leading to mismatches in style.", + "bbox": [ + 212, + 313, + 787, + 417 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/daa2e23316851fe8e753543045eb3e3d5615abf083f6395f1de25fee444511ff.jpg", + "image_caption": [ + "Fig. 7: Does the diffusion model prefer some styles over others? When a prompt contains two style tags, we find that SD 2.1 strongly favors the style that it can best reproduce, we suspect because of a prevalence of the style in training data. In each block, the General Style Similarity(GSS) of the left side artist (red color) is less than the right side artist (blue color). (Ref Fig. 2). The generated image is more biased towards the artist with high GSS score." + ], + "image_footnote": [], + "bbox": [ + 246, + 440, + 754, + 547 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Does the model prioritize some artists over others in the prompt? So far in the study we concentrated on the impact of including an artist's name in a style transfer prompt. In this section, we present preliminary findings in scenarios where prompts feature two artists. This is inspired by real-world user prompts from Stable Diffusion and Midjourney, where prompts often include multiple artists. We used the prompt in the style A painting in the style of $\\langle \\mathbf{X} \\rangle$ and $\\langle \\mathbf{Y} \\rangle$ , where $\\mathbf{X}$ and $\\mathbf{Y}$ represent different artists. For this study, we selected five artists with varying General Style Similarity (GSS) scores (referenced in Sec. 2). The artists, ranked by descending GSS scores, are Carne Griffiths, Roy Lichtenstein, Gustav Klimt, Pablo Picasso, and Ivan Bilibin. Note that most of the chosen artists have distinct styles that significantly differ from one another.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Investigating Style Similarity in Diffusion Models", + "bbox": [ + 401, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We chose SD-XL Turbo for this analysis because it is trained on deduplicated data, reducing bias towards frequently featured artists in the train set.", + "bbox": [ + 215, + 145, + 784, + 176 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The results for each pair of artists are presented in Fig. 7. Interestingly, even without specific instructions to generate a female subject, most outputs depicted women, reflecting the common subject matter of the artists studied. We also calculated the style similarity scores for each generated image, comparing them to the prototypical styles of the artists in the prompts. In most cases, the style of the artist with the higher GSS score dominated the generated image. To test for potential bias towards the artist positioned first in the prompt (X), we conducted two trials with reversed artist positions. The results were generally consistent, with the dominant style remaining unchanged. However, in the case of Pablo Picasso and Gustav Klimt, this pattern did not hold; the model favored Picasso's cubist style over Klimt's nouveau style, possibly due to the small difference in their GSS scores. While this is not an extensive study, a consistent trend emerged: styles of artists with higher GSS scores, like Leonid Afremov and Carne Griffiths, predominantly influenced the combined style. We leave the comprehensive study on this topic to future work.", + "bbox": [ + 215, + 178, + 787, + 402 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "8 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 424, + 357, + 441 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This study proposes a framework for learning style descriptors from both labeled and unlabeled data. After building a bespoke dataset, ContraStyles, we train a model that achieves state-of-the-art performance on a range of style matching tasks, including DomainNet, WikiArt, and ContraStyles. Then, we show the substantive practical utility of this model through an investigative analysis on the extent of style copying in popular text-to-image generative models. Here, we show the model is capable of determining the factors that contribute to the frequency of style replication. Through cross-referencing of images with style copies and their original prompts, we have discover that the degree of style copying is increasing with prompt complexity. Such complex prompts lead to greater style copying compared to simple one-line prompts. This finding sheds light on the interplay between textual prompts and style transfer, suggesting that prompt design can influence the level of style copying in generative models. Finally, note that the definition of style used in this work is strictly based on artist attribution. We chose this definition because it can be operationalized and used in dataset construction. This definition is certainly not a golden truth, and we look forward to future studies using alternative, or extended definitions.", + "bbox": [ + 215, + 455, + 787, + 710 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 733, + 401, + 752 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was made possible by the ONR MURI program and the AFOSR MURI program. Commercial support was provided by Capital One Bank, the Amazon Research Award program, and Open Philanthropy. Further support was provided by the National Science Foundation (IIS-2212182), and by the NSF TRAILS Institute (2229885).", + "bbox": [ + 215, + 763, + 784, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "G. Somepalli et al.", + "bbox": [ + 271, + 114, + 395, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Adobe: Firefly (2023), https://www.adobe.com/sensei/generative-ai/firefly.html", + "2. Agarwal, S., Karnick, H., Pant, N., Patel, U.: Genre and style based painting classification. In: 2015 IEEE Winter Conference on Applications of Computer Vision. pp. 588-594. IEEE (2015)", + "3. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5297-5307 (2016)", + "4. Bai, Z., Nakashima, Y., Garcia, N.: Explain me the painting: Multi-topic knowledgeable art description generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5422-5432 (2021)", + "5. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1728-1738 (2021)", + "6. Balestriero, R., Ibrahim, M., Sobal, V., Morcos, A., Shekhar, S., Goldstein, T., Bordes, F., Bardes, A., Mialon, G., Tian, Y., et al.: A cookbook of self-supervised learning. arXiv preprint arXiv:2304.12210 (2023)", + "7. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf 2(3), 8 (2023)", + "8. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021)", + "9. Chopra, S., Hadsell, R., LeCun, Y.: Learning a similarity metric discriminatively, with application to face verification. In: 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05). vol. 1, pp. 539-546. IEEE (2005)", + "0. Chu, W.T., Wu, Y.L.: Image style classification based on learnt deep correlation features. IEEE Transactions on Multimedia 20(9), 2491-2502 (2018)", + "1. Decrypt: Greg rutkowski removed from stable diffusion but brought back by ai artists (March 2024), https://decrypt.co/150575/greg-rutkowski-removed-from-stable-diffusion-but-brought-back-by-ai-artists", + "2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "3. Dumoulin, V., Shlens, J., Kudlur, M.: A learned representation for artistic style. arXiv preprint arXiv:1610.07629 (2016)", + "4. Gairola, S., Shah, R., Narayanan, P.: Unsupervised image style embeddings for retrieval and recognition tasks. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3281-3289 (2020)", + "5. Garcia, N., Vogiatzis, G.: How to read paintings: semantic art understanding with multi-modal retrieval. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018)", + "6. Gatys, L.A., Ecker, A.S., Bethge, M.: A neural algorithm of artistic style. arXiv preprint arXiv:1508.06576 (2015)", + "7. Gatys, L.A., Ecker, A.S., Bethge, M.: Image style transfer using convolutional neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2414-2423 (2016)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Investigating Style Similarity in Diffusion Models", + "bbox": [ + 401, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Geiping, J., Goldblum, M., Somepalli, G., Shwartz-Ziv, R., Goldstein, T., Wilson, A.G.: How much data are augmentations worth? an investigation into scaling laws, invariance, and implicit regularization. arXiv preprint arXiv:2210.06441 (2022)", + "19. Gibson, J.J.: The senses considered as perceptual systems. (1966)", + "20. Graham, D.J., Hughes, J.M., Leder, H., Rockmore, D.N.: Statistics, vision, and the analysis of artistic style. Wiley Interdisciplinary Reviews: Computational Statistics 4(2), 115-123 (2012)", + "21. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)", + "22. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501-1510 (2017)", + "23. Huang, X., Liu, M.Y., Belongie, S., Kautz, J.: Multimodal unsupervised image-to-image translation. In: Proceedings of the European conference on computer vision (ECCV). pp. 172-189 (2018)", + "24. Hughes, J.M., Graham, D.J., Jacobsen, C.R., Rockmore, D.N.: Comparing higher-order spatial statistics and perceptual judgements in the stylometric analysis of art. In: 2011 19th European Signal Processing Conference. pp. 1244-1248. IEEE (2011)", + "25. Hughes, J.M., Graham, D.J., Rockmore, D.N.: Stylometrics of artwork: uses and limitations. In: Computer Vision and Image Analysis of Art. vol. 7531, pp. 91-105. SPIE (2010)", + "26. Jiang, Q.Y., He, Y., Li, G., Lin, J., Li, L., Li, W.J.: Svd: A large-scale short video dataset for near-duplicate video retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5281-5289 (2019)", + "27. Joshi, A., Agrawal, A., Nair, S.: Art style classification with self-trained ensemble of autoencoding transformations. arXiv preprint arXiv:2012.03377 (2020)", + "28. Karayev, S., Trentacoste, M., Han, H., Agarwala, A., Darrell, T., Hertzmann, A., Winnemoeller, H.: Recognizing image style. arXiv preprint arXiv:1311.3715 (2013)", + "29. Lawrence-Lightfoot, S., Davis, J.H.: The art and science of portraiture. John Wiley & Sons (2002)", + "30. Lecoutre, A., Negrevergne, B., Yger, F.: Recognizing art style automatically in painting with deep learning. In: Asian conference on machine learning. pp. 327-342. PMLR (2017)", + "31. Lee, S., Kim, D., Han, B.: Cosmo: Content-style modulation for image retrieval with text feedback. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 802-812 (June 2021)", + "32. Li, J., Yao, L., Hendriks, E., Wang, J.Z.: Rhythmic brushstrokes distinguish van gogh from his contemporaries: findings via automated brushstroke extraction. IEEE transactions on pattern analysis and machine intelligence 34(6), 1159-1176 (2011)", + "33. Lu, X., Lin, Z., Shen, X., Mech, R., Wang, J.Z.: Deep multi-patch aggregation network for image style, aesthetics, and quality estimation. In: Proceedings of the IEEE international conference on computer vision. pp. 990-998 (2015)", + "34. Luan, F., Paris, S., Shechtman, E., Bala, K.: Deep photo style transfer. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4990-4998 (2017)", + "35. Lun, Z., Kalogerakis, E., Sheffer, A.: Elements of style: learning perceptual shape style similarity. ACM Transactions on graphics (TOG) 34(4), 1-14 (2015)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "G. Somepalli et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "36. Matsuo, S., Yanai, K.: Cnn-based style vector for style image retrieval. In: Proceedings of the 2016 ACM on International Conference on Multimedia Retrieval. pp. 309-312 (2016)", + "37. Matthews, R.A., Merriam, T.V.: Distinguishing literary styles using neural networks. In: Handbook of neural computation, pp. G8-1. CRC Press (2020)", + "38. Menis-Mastromichalakis, O., Sofou, N., Stamou, G.: Deep ensemble art style recognition. In: 2020 International Joint Conference on Neural Networks (IJCNN). pp. 1-8. IEEE (2020)", + "39. Midjourney: Midjourney (nd), https://www.midjourney.com/home", + "40. Mikolov, T., Sutskever, I., Chen, K., Corrado, G.S., Dean, J.: Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems 26 (2013)", + "41. Park, T., Zhu, J.Y., Wang, O., Lu, J., Shechtman, E., Efros, A., Zhang, R.: Swapping autoencoder for deep image manipulation. Advances in Neural Information Processing Systems 33, 7198-7211 (2020)", + "42. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1406-1415 (2019)", + "43. Pernias, P., Rampas, D., Richter, M.L., Pal, C., Aubreville, M.: Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In: The Twelfth International Conference on Learning Representations (2023)", + "44. pharmapsychotic: Clip interrogator. https://github.com/pharmapsychotic/clip-interrogator (2023)", + "45. Pizzi, E., Roy, S.D., Ravindra, S.N., Goyal, P., Douze, M.: A self-supervised descriptor for image copy detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14532-14542 (2022)", + "46. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "47. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)", + "48. Rodriguez, C.S., Lech, M., Pirogova, E.: Classification of style in fine-art paintings using transfer learning and weighted image patches. In: 2018 12th International Conference on Signal Processing and Communication Systems (ICSPCS). pp. 1-7. IEEE (2018)", + "49. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022)", + "50. Ruta, D., Motiian, S., Faieta, B., Lin, Z., Jin, H., Filipkowski, A., Gilbert, A., Collomosse, J.: Aladin: all layer adaptive instance normalization for fine-grained style similarity. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11926-11935 (2021)", + "51. Sablatnig, R., Kammerer, P., Zolda, E.: Hierarchical classification of paintings using face-and brush stroke models. In: Proceedings. Fourteenth International Conference on Pattern Recognition (Cat. No. 98EX170). vol. 1, pp. 172-174. IEEE (1998)", + "52. Saleh, B., Elgammal, A.: Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855 (2015)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Investigating Style Similarity in Diffusion Models", + "bbox": [ + 401, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "53. Sandoval, C., Pirogova, E., Lech, M.: Two-stage deep learning approach to the classification of fine-art paintings. IEEE Access 7, 41770-41781 (2019)", + "54. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402 (2022)", + "55. Silva, J.M., Pratas, D., Antunes, R., Matos, S., Pinho, A.J.: Automatic analysis of artistic paintings using information-based measures. Pattern Recognition 114, 107864 (2021)", + "56. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)", + "57. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Diffusion art or digital forgery? investigating data replication in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)", + "58. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Understanding and mitigating copying in diffusion models. Advances in Neural Information Processing Systems 36, 47783-47803 (2023)", + "59. Srinivasa Desikan, B., Shimao, H., Miton, H.: Wikiartvectors: style and color representations of artworks for cultural analysis via information theoretic measures. Entropy 24(9), 1175 (2022)", + "60. Tenenbaum, J., Freeman, W.: Separating style and content. Advances in neural information processing systems 9 (1996)", + "61. Walmer, M., Suri, S., Gupta, K., Shrivastava, A.: Teaching matters: Investigating the role of supervision in vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)", + "62. Wang, J., Yang, H., Fu, J., Yamasaki, T., Guo, B.: Fine-grained image style transfer with visual transformers. In: Proceedings of the Asian Conference on Computer Vision. pp. 841-857 (2022)", + "63. Wang, S.Y., Efros, A.A., Zhu, J.Y., Zhang, R.: Evaluating data attribution for text-to-image models. arXiv preprint arXiv:2306.09345 (2023)", + "64. Wilber, M.J., Fang, C., Jin, H., Hertzmann, A., Collomosse, J., Belongie, S.: Bam! the behave artistic media dataset for recognition beyond photography. In: Proceedings of the IEEE international conference on computer vision. pp. 1202-1211 (2017)", + "65. Willats, J., Durand, F.: Defining pictorial style: Lessons from linguistics and computer graphics. Axiomathes 15(3), 319-351 (2005)", + "66. Wynen, D., Schmid, C., Mairal, J.: Unsupervised learning of artistic styles with archetypal style analysis. Advances in Neural Information Processing Systems 31 (2018)", + "67. Yao, L., Li, J., Wang, J.Z.: Characterizing elegance of curves computationally for distinguishing morisseau paintings and the imitations. In: 2009 16th IEEE International Conference on Image Processing (ICIP). pp. 73-76. IEEE (2009)", + "68. Zhang, W., Cao, C., Chen, S., Liu, J., Tang, X.: Style transfer via image component analysis. IEEE Transactions on multimedia 15(7), 1594-1601 (2013)" + ], + "bbox": [ + 215, + 147, + 784, + 756 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "G. Somepalli et al.", + "bbox": [ + 271, + 114, + 395, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Investigating Style Similarity in Diffusion Models/e6cae220-dbdc-4ecf-90b6-28f1684a0527_model.json b/2024/Investigating Style Similarity in Diffusion Models/e6cae220-dbdc-4ecf-90b6-28f1684a0527_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0eef2e4d609886286e9c324c4cdeaff2a9b0e169 --- /dev/null +++ b/2024/Investigating Style Similarity in Diffusion Models/e6cae220-dbdc-4ecf-90b6-28f1684a0527_model.json @@ -0,0 +1,2535 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.219, + 0.142, + 0.784, + 0.164 + ], + "angle": 0, + "content": "Investigating Style Similarity in Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.19, + 0.784, + 0.222 + ], + "angle": 0, + "content": "Growthami Somepalli\\*1, Anubhav Gupta\\*1, Kamal Gupta\\*1, Shramay Palta\\*1, Micah Goldblum\\*3, Jonas Geiping\\*2, Abhinav Shrivastava\\*1, and Tom Goldstein\\*1" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.232, + 0.66, + 0.248 + ], + "angle": 0, + "content": "1 University of Maryland, College Park" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.248, + 0.658, + 0.261 + ], + "angle": 0, + "content": "2 ELLIS Institute, MPI for Intelligent Systems" + }, + { + "type": "text", + "bbox": [ + 0.425, + 0.261, + 0.578, + 0.275 + ], + "angle": 0, + "content": "3 Columbia University" + }, + { + "type": "text", + "bbox": [ + 0.377, + 0.275, + 0.626, + 0.289 + ], + "angle": 0, + "content": "4 https://somepago.github.io/csd" + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.332, + 0.302, + 0.34 + ], + "angle": 0, + "content": "Real paintings" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.341, + 0.33, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.332, + 0.385, + 0.339 + ], + "angle": 0, + "content": "MidJourney" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.342, + 0.414, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.332, + 0.454, + 0.339 + ], + "angle": 0, + "content": "2.1" + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.342, + 0.472, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.473, + 0.332, + 0.502, + 0.339 + ], + "angle": 0, + "content": "1.4" + }, + { + "type": "image", + "bbox": [ + 0.473, + 0.342, + 0.501, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.332, + 0.589, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.332, + 0.664, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.332, + 0.724, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.332, + 0.783, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.388, + 0.331, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.388, + 0.414, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.388, + 0.499, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.388, + 0.611, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.388, + 0.722, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.388, + 0.78, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.436, + 0.331, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.436, + 0.414, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.436, + 0.499, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.436, + 0.611, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.436, + 0.696, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.436, + 0.78, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.493, + 0.788, + 0.59 + ], + "angle": 0, + "content": "Fig. 1: Original artwork of 6 popular artists and the images generated in the style of these artists by three popular text-to-image generative models. The numbers displayed below each image indicate the similarity of the generated image with the artist's style using the proposed method. A high similarity score suggests a strong presence of the artist's style elements in the generated image. Based on our analyses, we postulate that three artists on the right were removed (or unlearned) from SD 2.1 while they were present in MidJourney and SD 1.4. Please refer to Section 2 for more details." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.649, + 0.741, + 0.802 + ], + "angle": 0, + "content": "Abstract. Generative models are now widely used by graphic designers and artists. Prior works have shown that these models remember and often replicate content from their training data during generation. Hence as their proliferation increases, it has become important to perform a database search to determine whether the properties of the image are attributable to specific training data, every time before a generated image is used for professional purposes. Existing tools for this purpose focus on retrieving images of similar semantic content. Meanwhile, many artists are concerned with style replication in text-to-image models. We present a framework for understanding and extracting style descriptors from images. Our framework comprises a new dataset curated using the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.826, + 0.618, + 0.84 + ], + "angle": 0, + "content": "*Equal contribution. Correspondence: growthami@umd.edu." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.129 + ], + "angle": 0, + "content": "G. Somepalli et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.259 + ], + "angle": 0, + "content": "insight that style is a subjective property of an image that captures complex yet meaningful interactions of factors including but not limited to colors, textures, shapes, etc. We also propose a method to extract style descriptors that can be used to attribute style of a generated image to the images used in the training dataset of a text-to-image model. We showcase promising results in various style retrieval tasks. We also quantitatively and qualitatively analyze style attribution and matching in the Stable Diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.273, + 0.686, + 0.288 + ], + "angle": 0, + "content": "Keywords: Image Style \\(\\cdot\\) Style Similarity \\(\\cdot\\) Generative models" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.318, + 0.376, + 0.334 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.52 + ], + "angle": 0, + "content": "Diffusion-based image generators like Stable Diffusion [49], DALL-E [47] and many others [1, 7, 39, 43] learn artistic styles from massive captioned image datasets that are scraped from across the web [54]. Before a generated image is used for commercial purposes, it is wise to understand its relationship to the training data and the origins of its design elements and style attributes. Discovering and attributing these generated images, typically done with image similarity search, is hence becoming increasingly important. Such dataset attribution serves two purposes. It enables users of generated images to understand potential conflicts, associations, and social connotations that their image may evoke. It also enables artists to assess whether and how generative models are using elements of their work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.748 + ], + "angle": 0, + "content": "Despite a long history of research [60], recovering style from an image is a challenging and open problem in Computer Vision. Many existing retrieval methods [8, 45, 46] for large training datasets focus primarily on matching semantic content between a pair of images. Understanding the origin of the style present in a generated image, however, is much less well understood. To address this gap, we propose a self-supervised objective for learning style descriptors from images. Standard augmentation-based SSL pipelines (e.g. SimCLR and variants) learn feature vectors that are invariant to a set of augmentations. Typically, these augmentations preserve semantic content and treat style as a nuisance variable. In contrast, we choose augmentations that preserve stylistic attributes (colors, textures, or shapes) while minimizing content. Unfortunately, SSL is not enough, as style is inherently subjective, and therefore a good style extractor should be aligned with human perceptions and definitions of style. For this reason, we curate a style attribution dataset, ContraStyles, in which images are associated with the artist that created them." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "By training with both SSL and supervised objectives, we create a high-performance model for representing style. Our model, CSD, outperforms other large-scale pre-trained models and prior style retrieval methods on standard datasets. Using CSD, we examine the extent of style replication in the popular open-source text-to-image generative model Stable Diffusion [49], and consider different factors that impact the rate of style replication." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.402, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Investigating Style Similarity in Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.27 + ], + "angle": 0, + "content": "To summarize our contributions, we (1) propose a style attribution dataset ContraStyles, associating images with their styles, (2) introduce a multi-label contrastive learning scheme to extract style descriptors from images and show the efficacy of the scheme by zero-shot evaluation on public domain datasets such as WikiArt and DomainNet (3) We perform a style attribution case study for one of the most popular text-to-image generative models, Stable Diffusion, and propose indicators of how likely an artist's style is to be replicated. Code, artifacts, and dataset links are available at https://somepago.github.io/csd." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.29, + 0.36, + 0.306 + ], + "angle": 0, + "content": "2 Motivation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.319, + 0.789, + 0.606 + ], + "angle": 0, + "content": "We present a case study that shows how style features can be used to interrogate a generative model, and provide utility to either artists or users. We consider the task of analyzing a model's ability to emulate an artist's style, and of attributing the style of an image to an artist. We begin by curating a list of 96 artists, primarily sourced from the WikiArt database, supplemented by a few contemporary artists who are notably popular within the Stable Diffusion community5. For each artist, we compute a prototype vector by averaging the embeddings of their paintings using our proposed feature extractor, CSD ViT-L. Next, we generate an image for each artist using Stable Diffusion 2.1 with a prompt in the format A painting in the style of . We compute the dot product similarity between each generated image's embedding and the artist's prototype. This process was repeated multiple times for each artist, and we plot mean results in Fig. 2. We refer to this quantity as the General Style Similarity (GSS) score for an artist, as it measures how similar a generated image is to a typical image from that artist while using our style representation model. We also plot an analogous style similarity score, but using \"content-constrained\" prompts. For instance, one prompt template is A painting of a woman doing style of where X is the name of the artist and Y is some setting like reading a book or holding a baby etc. See Sec. 7 for all templates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.789, + 0.788 + ], + "angle": 0, + "content": "Each point in Fig. 2 represents an artist. Notice that GSS scores are highly correlated with content-constrained scores, indicating that our feature vectors represent style more than semantic content. Our findings reveal that SD 2.1 is much more capable of emulating some artists than others. Artists like Leonid Afremov, Georges Seurat exhibit high style similarity scores, and visual inspection of generated images confirms that indeed their style is emulated by the model (Fig. 1 - Original artwork vs SD 2.1). On the other end of the spectrum, artists such as Ruan Jia and Greg Rutkowski showed low similarity scores, and likewise the generated images bear little resemblance to the artists' work. Interestingly, after completing this study, we discovered that Greg Rutkowski's work was excluded from the training data for the Stable Diffusion 2.1 model, as reported by [11]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.788, + 0.789, + 0.819 + ], + "angle": 0, + "content": "This demonstrates that the Style Similarity score can be used by artists to quantify how well a model emulates their style, or it can be used by users to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.663, + 0.84 + ], + "angle": 0, + "content": "5 https://supagruen.github.io/StableDiffusion-CheatSheet/" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.129 + ], + "angle": 0, + "content": "G. Somepalli et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "ascertain whether a generated image contains stylistic elements associated with a particular artist. After a thorough inspection of the generations from 96 artists, we hypothesize that a single-image Style Similarity score below 0.5 indicates the absence of the artist's style, whereas a score above 0.8 strongly indicates its presence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.49, + 0.556 + ], + "angle": 0, + "content": "In Figure 1, we show original artworks for 6 artists, and generations from MidJourney [39], Stable Diffusion 2.1 and Stable Diffusion 1.4 [49] for each of these artists. The 3 artists on the left side have high GSS while the ones on the right side have low GSS. Below each generated image, we display the similarity against the artist's prototype vector. We see high image similarity scores in the Midjourney generations and qualitatively these images look stylistically similar to artists' original artworks. We also see the interesting cases of Greg Rutkowski, Ruan Jia, and Amano whose style is captured by Stable Diffusion 1.4, while being notably absent in Stable Diffusion 2.1. This finding is in line with reports suggesting that some of these artists were removed from the training data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.788, + 0.618 + ], + "angle": 0, + "content": "of Stable Diffusion 2.1 [11]. Based on this analysis, we postulate that Ruan Jia, Wadim Kashin, Anton Fadeev, Justin Gerard, and Amano were also either excluded from the training data or post-hoc unlearned/removed from Stable Diffusion 2.1." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.258, + 0.776, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.443, + 0.788, + 0.554 + ], + "angle": 0, + "content": "Fig. 2: Style similarity of Stable Diffusion 2.1 generated images against the artist's prototypical representation. On the X-axis, the similarities are depicted when the prompt is not constrained, while the Y-axis represents similarity when the prompt is constrained to generate an image of a \"woman\" in the artist's style." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.647, + 0.394, + 0.667 + ], + "angle": 0, + "content": "3 What is style?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.779 + ], + "angle": 0, + "content": "The precise definition of \"style\" remains in contention, but many named artistic styles (e.g., cubism, impressionism, etc...) are often associated with certain artists. We leverage this social construct, and define style simply as the collection of global characteristics of an image that are identified with an artist or artistic movement. These characteristics encompass various elements such as color usage, brushstroke techniques, composition, and perspective." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Related work. Early computer vision algorithms attempted to model style using low-level visual features like color histograms, texture patterns, edge detection, and shape descriptors. Other computational techniques involve rule-based systems, such as the presence of specific compositional elements, the use of spe" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.403, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Investigating Style Similarity in Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "cific color palettes, or the presence of certain brushstroke patterns to identify specific style characteristics [19, 20, 24, 25, 29, 32, 35, 37, 51, 55, 59, 65, 67]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.344 + ], + "angle": 0, + "content": "Modern studies have focused on the task of transferring style from one image to another [13, 17, 22, 34, 41, 62, 68]. Some works have also concentrated on style classification [2, 4, 10, 15, 27, 28, 30, 33, 38, 48, 53]. A limited number of studies address in-the-wild style quantification, matching, and retrieval [14, 23, 36, 50, 66]. In their seminal work, Gatys et al. [17] introduced Gram Matrices as style descriptors and utilized an optimization loop to transfer style. Another approach proposed by Luan et al. [34] includes a photorealism regularization term to prevent distortions in the reconstructed image. Zhang et al. [68] formulated style transfer using Markov random fields. Beyond Gram-based style representation, Chu et al. [10] explored various other types of correlations and demonstrated performance variations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.48 + ], + "angle": 0, + "content": "In a recent work by Lee et al [31], two separate neural network modules were used - one for image style and another for image content - to facilitate image style retrieval. In the most recent related research, Wang et al. [63] developed an attribution model trained on synthetic style pairs, designed to identify stylistically similar images. In contrast to this approach, our method leverages real image pairs, curated automatically through their caption annotations. Despite our training dataset being approx. \\(16\\%\\) the size of training data used in [63], we demonstrate that our model significantly outperforms this method on many zero-shot style matching tasks in the later sections." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.503, + 0.63, + 0.521 + ], + "angle": 0, + "content": "4 Creating a dataset for style attribution" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.642 + ], + "angle": 0, + "content": "While many large web datasets now exist, we need one that contains wide variations in artistic styles, and also labels that be used for downstream style retrieval evaluation. Some large-scale datasets specifically designed to handle such a challenge, like BAM [64], are not available in the public domain and others like WikiArt [52] are not large enough to train a good style feature extractor. In the following section, we propose a way to curate a large style dataset out of the LAION [54] Aesthetics \\(6+\\) dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.795 + ], + "angle": 0, + "content": "ContraStyles: A dataset for style distillation. We curate our own dataset as a subset of LAION [54]. We start off with the 12M image-text pairs with predicted aesthetics scores of 6 or higher. Note that this dataset is extremely unbalanced, with some popular artists appearing much more frequently than others. Also, a large number of images are duplicated within the dataset which is a major issue for the text-to-image models trained on this data [57,58]. Furthermore, the image captions within the data are often noisy and are often missing a lot of information. We address these challenges and propose a new subset of LAION-Aesthetics consisting of 511,921 images, and 3840 style tags, where each image can have one or more tags. We use this dataset for training our models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We begin with a bank of styles collated in previous work for image understanding with the CLIP Interrogator [44]. This bank of styles was curated based on typical user prompts for Stable Diffusion. We combine the bank of artists," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.129 + ], + "angle": 0, + "content": "G. Somepalli et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "medium, and movement references, to a shortlist of 5600 tags. We then search for these tags in the 12M LAION-Aesthetics captions and shortlist the images that have at least one of the tags present. We further filter out the tags which have over 100,000 hits in the dataset since human inspection found that they refer to common phrases like 'picture' or 'photograph' that do not invoke a distinct style. After discarding images with an unavailable URL, we are left with about 1 million images and 3840 tags. We further deduplicate the images using SSCD [45] with a threshold of 0.8 and merge the tags of images that are near copies of each other. As a by-product, the dedduplication also helps with the missing tags in the images, since we can simply merge the text labels of duplicate images. After dedduplication, we are left with 511,921 images." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.35, + 0.606, + 0.368 + ], + "angle": 0, + "content": "5 Contrastive Style Descriptors (CSD)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.788, + 0.53 + ], + "angle": 0, + "content": "Self-Supervised Learning. Many successful approaches to SSL [61] use a contrastive [9] approach, where two views (or augmentations) of the same image in the dataset are sampled and passed to a deep network to get their respective image descriptors. The network is trained to maximize the similarity of two views of the same image and minimize agreement with other images in the batch. Standard choices for augmentations include color jitter, blurring, grayscale, etc., as these alter the image's visual properties while preserving content. While these are good augmentations for object recognition tasks, they train the network to ignore image attributes associated with style." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.788, + 0.593 + ], + "angle": 0, + "content": "Our approach relies on a training pipeline with two parts. First, we use contrastive SSL, but with a set of augmentations that are curated to preserve style. Second, we align our model with human perceptions of style by training on our labelled ContraStyles dataset described in Section 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.787, + 0.655 + ], + "angle": 0, + "content": "Proposed Approach. We seek a model for extracting image descriptors that contain concise and effective style information. To be a useful, the model should be invariant to semantic content and capable of disentangling multiple styles." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Given a dataset of \\(N\\) labeled images \\(\\{\\mathbf{x}_i, l_i\\}_{i=1}^N\\), where each image can have one or more labels from a set of \\(L\\) labels, we define the label vector of the \\(i^{\\text{th}}\\) image as \\(l_i = (c_1, c_2, \\ldots, c_L)\\), where each \\(c_k \\in \\{0, 1\\}\\). As mentioned in the previous section, our multi-label dataset consists of \\(N = 511,921\\) images and \\(L = 3,840\\) style tags. We consider a mini-batch of \\(B\\) images. Each of the images are passed to a Vision Transformer (ViT) [12] backbone, and then projected to a \\(d\\)-dimensional vector. We consider two variants of ViT (ViT-B and ViT-L)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Our style descriptors \\( f_{\\mathrm{ViT}}(\\mathbf{x}_i) \\in \\mathbb{R}^d \\) are then used to create a matrix of pairwise cosine similarity scores \\( s_{i,j} = \\cos (f_{\\mathrm{ViT}}(\\mathbf{x}_i), f_{\\mathrm{ViT}}(\\mathbf{x}_j)) \\). In order to compute our multi-label contrastive loss (MCL), we also compute the groundtruth similarity scores as \\( \\hat{s}_{i,j} = \\mathbb{1}\\left(l_i^{\\mathrm{T}}l_j\\right) \\), where \\( \\mathbb{1} \\) is the indicator function that returns 1 if any of the labels of the images \\( i, j \\) match. Our final loss term reduces to:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.403, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Investigating Style Similarity in Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.172, + 0.785, + 0.215 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M C L}} = - \\hat {s} _ {i, j} \\log \\frac {\\exp \\left(s _ {i , j}\\right) / \\tau)}{\\sum_ {k \\neq j} \\exp \\left(s _ {i , k}\\right) / \\tau}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.227, + 0.599, + 0.241 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the temperature fixed during the training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.242, + 0.785, + 0.392 + ], + "angle": 0, + "content": "Since our supervised dataset is modest in size, we add a self-supervised objective. We sample two \"views\" (augmentations) of each image in a batch and add a contrastive SSL term. Standard SSL training routines (e.g., MoCo, SimCLR, BYOL etc.) choose augmentations so that each pair of views has the same semantic content, but different style content. These augmentations typically include Resize, Horizontal Flips, Color Jitter, Grayscale, Gaussian Blur, and Solarization [6, 18]. For our purposes, we depart from standard methods by excluding photometric augmentations (Gaussian Blur, Color Jitter), as they alter the style of the image. We keep the following spatial augmentations - Horizontal Flips, Vertical Flips, Resize and Rotation as they keep style intact." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.787, + 0.469 + ], + "angle": 0, + "content": "The overall loss function is a simple combination of the multilabel contrastive loss and self-supervised loss \\(\\mathcal{L} = \\mathcal{L}_{\\mathrm{MCL}} + \\lambda \\mathcal{L}_{\\mathrm{SSL}}\\). During inference, we use the final layer embedding and the dot product to compute style similarity between any two images. In our experiments, we found that initializing weights to CLIP [46] ViT-B and ViT-L improves performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.495, + 0.325, + 0.51 + ], + "angle": 0, + "content": "6 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.785, + 0.663 + ], + "angle": 0, + "content": "Training details. We present the results for two variants of our model CSD ViT-B and CSD ViT-L version. Both the models are initialized with respective CLIP variant checkpoints and are finetuned for 80k iterations on the ContraStyles dataset on 4 A4000/A5000 GPUs. We use an SGD optimizer with momentum 0.9 and learning rate of 0.003 for the projection layer and \\(1e - 4\\) for the backbone. Our mini-batch size per GPU is 16. We use \\(\\lambda = 0.2\\) and \\(\\tau = 0.1\\) for the final model. The training takes about 8 hours for the base model and around 16 hours for the large model. See the Appendix for more details and ablations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.67, + 0.787, + 0.731 + ], + "angle": 0, + "content": "Task. We perform zero-shot evaluation across multiple datasets on a style-retrieval task. Following [3, 26], we split each dataset into two parts: Database and Query. Given a query image at test time, we evaluate whether we can find the ground-truth style in its nearest neighbors from the database." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Baselines. We compare our model against a recent style attribution model GDA [63] which is trained via fine-tuning on paired synthetic style data, and VGG [17, 56] Gram Matrices which are often used for neural style transfer applications. Further we compare with CLIP [46] models supervised with freeform text captions, and with other self-supervised models such as DINO [8], MoCo [21], SSCD [45]. We use the embeddings from the last layer for each of these models except for VGG where we use the Gram Matrix [17] of the last" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.129 + ], + "angle": 0, + "content": "G. Somepalli et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.178 + ], + "angle": 0, + "content": "layer. We skipped evaluations of [14, 23, 50] since both pre-trained models and training data are not available." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.182, + 0.787, + 0.274 + ], + "angle": 0, + "content": "Metrics. We do nearest neighbour searches for \\( k \\in [1, 10, 100] \\), and report Recall@k, mAP@k. We use the standard definitions of these metrics from the retrieval literature. Like [5], we define positive recall as the existence of a correct label in top-N matches and no recall when none of the top-N matches share a label with the query. Similarly, mAP is defined as average over precision at each rank in N, and then averaged over all queries." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.301, + 0.787, + 0.343 + ], + "angle": 0, + "content": "Table 1: mAP and Recall metrics on DomainNet and WikiArt datasets. Our model consistently performs the best in all cases except one, against both self-supervised and style attribution baselines." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.356, + 0.784, + 0.58 + ], + "angle": 0, + "content": "
DomainNet (mAP@k)WikiArt (mAP@k)DomainNet (Recall@k)WikiArt (Recall@k)
Method110100110100110110100
VGG Gram [16]---25.919.411.4--25.952.780.4
DINO ViT-B/16 [8]69.468.266.244.033.418.969.493.744.069.488.1
DINO ViT-B/8 [8]72.270.969.346.935.920.472.293.846.971.088.9
SSCD RN-50 [45]67.665.962.036.026.514.867.695.036.062.185.4
MOCO ViT-B/16 [21]71.971.169.644.033.218.872.094.044.069.088.0
CLIP ViT-B/16 [46]73.773.071.352.242.026.073.794.552.278.393.5
GDA CLIP ViT-B [63]62.961.659.325.621.014.162.992.325.656.683.8
GDA DINO ViT-B [63]69.568.166.145.534.619.769.593.445.575.889.0
GDA ViT-B [63]67.165.664.242.632.218.267.193.642.667.687.1
CSD ViT-B (Ours)78.377.576.056.246.128.778.394.356.280.393.6
CLIP ViT-L [46]74.073.572.259.448.831.574.094.859.482.995.1
CSD ViT-L (Ours)78.377.876.564.5653.8235.6578.394.564.5685.7395.58
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Evaluation Datasets. DomainNet [42] consists of an almost equal number of images from six different domains: Clipart, Infograph, Painting, Quickdraw, Real, and Sketch. Upon examination, we observed a strong stylistic resemblance between the Quickdraw and Sketch domains, leading us to exclude Quickdraw from our analysis. The dataset's content information was utilized to categorize the images into two main clusters of content classes. This clustering was achieved through the application of word2vec [40]. The images within the smaller cluster were designated as part of the Query set (20,000 images), and images in the bigger cluster to Database (206768 images). And the second dataset we evaluate all the models is, WikiArt [52]. It consists of 80096 fine art images spread across 1119 artists and 27 genres. We randomly split the dataset into 64090 Database and 16006 Query images. We use the artist as a proxy for the style since there is large visual variation within each genre for them to be considered as independent styles. Under this setting, WikiArt is a challenging retrieval task as the chance probability of successful match is just .09% while it is 20% for DomainNet." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.403, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Investigating Style Similarity in Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.48, + 0.162 + ], + "angle": 0, + "content": "6.1 Analyses and observations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.186, + 0.788, + 0.428 + ], + "angle": 0, + "content": "In Table 1, we report the metrics for all the baselines considered and the proposed CSD method using \\( k \\) nearest neighbors on 2 datasets - DomainNet and WikiArt. Note that while style and content are better separated in the case of DomainNet, WikiArt consists of more fine-grained styles and has more practical use cases for style retrieval. Loosely speaking, mAP@\\( k \\) determines what percentage of the nearest neighbors that are correct predictions, while the recall determines what percentage of queries has a correct match in the top-\\( k \\) neighbors. Our model CSD consistently outperformed all the pretrained feature extractors as well as the recent attribution method GDA [63] on both WikiArt and DomainNet evaluations. Note that all models are evaluated in a zero-shot setting. We see the most gains in the WikiArt dataset which is more challenging with chance probability of only \\( 0.09\\% \\). When we look at mAP@1, which is same as top-1 accuracy, our base model outperforms the next best model by \\( 5\\% \\) points on WikiArt and \\( 4.6\\% \\) points on DomainNet. Our large model out-performs the closest large competitor by similar margins. Given the complexity of the task, these improvements are non-trivial." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.522 + ], + "angle": 0, + "content": "We attribute the improvements to a couple of factors, (1) The multi-label style contrastive loss on our curated ContraStyles dataset is quite helpful in teaching the model right styles (2) We hypothesize that these SSL models become invariant to styles because the way they were trained, but we are careful to not strip that away in our SSL loss component by carefully curating non-photometric augmentations in training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Error Analysis. Even though our model outperforms the previous baselines, our top-1 accuracy for the WikiArt style matching task is still at 64.56. We tried to understand if there is a pattern to these errors. For example, our model is consistently getting confused between impressionist painters claude monet, gustave loiseau, and alfred sisley, all of whom painted many landscapes. They depicted natural scenes, including countryside views, rivers, gardens, and coastal vistas. Another example is pablo picasso and georges braque, who are both cubist painters. Given the impracticality of analyzing all 1,119 artists in the dataset, we opted for a macroscopic examination by categorizing errors at the art movement level. This approach is visualized in the heatmap presented in Fig. 3. In the heatmap, we see most of the errors concentrated along the diagonal, indicating that while the model often correctly identifies the art movement, it struggles to pinpoint the exact artist. There are instances of off-diagonal errors where the model incorrectly identifies both the artist and their art movement. For example, many Post Impressionism and Realism paintings are assigned to Impressionism artists. Upon closer examination, it becomes apparent that they closely align in terms of historical timeline and geographical origin, both being from Europe. This analysis indicates the nuanced nature of style detection and retrieval in Computer Vision. It suggests that the upper limit for accuracy in this task might be considerably lower than \\(100\\%\\), even for a typical human evaluator, due to the inherent subtleties and complexities involved." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.129 + ], + "angle": 0, + "content": "G. Somepalli et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.144, + 0.585, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.304, + 0.596, + 0.361 + ], + "angle": 0, + "content": "Fig. 3: Confusion Matrix of errors in WikiArt: Art movements are predicted correctly. Errors occur in cases where movements share the same historical timeline and/or are derived from the same earlier movement." + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.177, + 0.787, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.277, + 0.788, + 0.361 + ], + "angle": 0, + "content": "Fig. 4: Human study on Style Retrieval: Turns out untrained humans are worse than many feature extractors on matching images from the same artist." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.393, + 0.379, + 0.41 + ], + "angle": 0, + "content": "6.2 Human Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.789, + 0.61 + ], + "angle": 0, + "content": "To understand how our models compare to untrained humans, we conducted a small survey on style matching on 30 humans (excluding the authors). Following the convention in other papers [14,28,50,62] and this paper, we assume, 2 images from same artist can be considered stylistic matches. For each query image, we gave 4 answer images out of which only one is from the same artist and hence is the right answer, so chance accuracy is \\(25\\%\\). We used the Artcive dataset introduced in [63] to create this survey and we collected 3 responses per item to break any ties. We present the results in Fig 4. Most interestingly, untrained humans are worse than many feature extractors at this task. SSCD is the only feature extractor that did worse than humans. Our model, CSD outperforms all the baselines on this dataset as well. This underpins the difficulty of style matching and also highlights the superior performance of our feature extractor." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.641, + 0.784, + 0.66 + ], + "angle": 0, + "content": "7 Studying style in the wild: Analysis of Stable Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.682, + 0.788, + 0.775 + ], + "angle": 0, + "content": "In the previous section, we have quantitatively shown that our model Contrastive Style Descriptors outperforms many baselines on style matching task. Now we try to address the question, Can we do style matching on Stable Diffusion generated images? To answer this question, we first curated multiple synthetic image collections using Stable Diffusion v 2.1 [49] and then compared them against the \"ground truth\" style matches on ContraStyles dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Creating synthetic style dataset. The first challenge in curating synthetic images through prompts is the choice of prompts to be used for the generation. There have been no in-depth quantitative studies of the effect of prompts on generation styles. For this analysis, we chose 3 types of prompts." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.402, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "Investigating Style Similarity in Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.13, + 0.788, + 0.158 + ], + "angle": 0, + "content": "Table 2: mAP and Recall of SD 2.1 generated synthetic datasets based on Simple prompts and User-generated prompts" + }, + { + "type": "table", + "bbox": [ + 0.228, + 0.17, + 0.784, + 0.273 + ], + "angle": 0, + "content": "
Simple prompts (mAP@k)User-generated (mAP@k)Simple prompts (Recall@k)User-generated (Recall@k)
Method110100110100110100110100
GDA - DINO11.610.27.64.454.594.2411.628.152.834.4525.2267.18
CSD-ViTB17.5316.5612.685.855.965.5817.5338.6561.855.8529.2674.2
CLIP ViT-L/1422.320.416.16.15.75.122.344.566.26.126.071.7
CSD (Ours)24.523.318.55.75.95.624.547.267.55.726.571.8
" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.298, + 0.787, + 0.374 + ], + "angle": 0, + "content": "1. User-generated prompts: We used a Stable Diffusion Prompts\\(^6\\) dataset of 80,000 prompts filtered from Lexica.art. We used the test split and then filtered the prompts to make sure at least one of the keywords from the list we curated in Section 4 is present. We then sampled 4000 prompts from this subset for query split generation." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.374, + 0.787, + 0.435 + ], + "angle": 0, + "content": "2. Simple prompts: We randomly sampled 400 artists which appeared most frequently in user-generated prompts we analysed. We format the prompt as A painting in the style of , and we generate 10 images per prompt by varying the initialization seed." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.435, + 0.794, + 0.541 + ], + "angle": 0, + "content": "3. Content-constrained prompts: We wanted to understand if we can detect style when we constrain the model to generate a particular subject/human in the style of an artist. For this, we used the prompt A painting of a woman in the style of or A painting of a woman reading in the style of etc., a total of 5 variations per subject repeated two times. We experimented with subjects, woman,dog and house in this study. We provide the exact templates in the appendix." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.298, + 0.794, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.574, + 0.788, + 0.603 + ], + "angle": 0, + "content": "Table 3: mAP and Recall of SD 2.1 generated synthetic datasets based on Content-constrained prompts" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.615, + 0.784, + 0.715 + ], + "angle": 0, + "content": "
Dog (mAP@k)House (mAP@k)Woman (mAP@k)Dog (Recall@k)House (Recall@k)Woman (Recall@k)
Method110100110100110100110100110100110100
GDA-DINO2.2821.63.93.62.72.22.322.288.6828.733.912.3322.210.128.9
CSD-ViT-B4.54.313.614.64.293.877.557.836.424.514.3634.884.615.03397.5520.142.46
CLIP ViT-L/142.32.21.94.54.23.67.47.16.22.29.829.94.513.835.37.419.041.6
CSD (Ours)4.94.84.26.46.25.410.810.18.64.914.534.56.417.840.610.823.444.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.742, + 0.788, + 0.818 + ], + "angle": 0, + "content": "We generate 4000 images for each prompt setting using Stable Diffusion 2.1. There is only one style keyword in simple and content-constrained prompts, which we also use as a ground truth label for matching tasks. However, user-generated prompts can have multiple style labels within the caption, and we consider all of them as ground-truth labels." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.755, + 0.841 + ], + "angle": 0, + "content": "\\(^{6}\\) https://huggingface.co/datasets/Gustavosta/Stable-Diffusion-Prompts" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.129 + ], + "angle": 0, + "content": "G. Somepalli et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.497, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.144, + 0.787, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.292, + 0.788, + 0.351 + ], + "angle": 0, + "content": "Fig. 5: Nearest \"style\" neighbors. For each generated image (referred to as SD Gen), we show the top 5 style neighbors in CSD using our feature extractor. The green and red box around the image indicates whether or not the artist's name used to generate the SD image was present in the caption of the nearest neighbor." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.369, + 0.784, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.609 + ], + "angle": 0, + "content": "Fig. 6: Top row: Images generated by Stable Diffusion. Middle and Bottom rows: Top matches retrieved by CLIP vs CSD (ours) respectively. CLIP is consistently biased towards image content, for instance retrieving image of a dog in the Column 1, 3, 4, or image of mother and baby in Column 7 or 8. Our method emphasize less on the content but more on the image styles. Please refer to the Appendix for the prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Style retrieval on generated images. In Tab. 2, we show the retrieval results for Simple and User-generated prompts. We also compare our results with the second-best performing model in the previous section, CLIP ViT-L, and a recent style attribution model GDA [63]. We observe that our method outperforms CLIP on Simple prompt dataset. For User-generated prompts, the performance metrics are closer to CLIP model, but it's important to note that these prompts are inherently more complex. This complexity results in a different label distribution in the query set for the two types of prompts we examine, leading to varied metric ranges in each case. Additionally, our method consistently outperforms both baselines in content-constrained scenarios, as evidenced in Table 3. This indicates the robustness and effectiveness of our approach in dealing with a variety of prompt complexities and content specifications. We refer the reader to Appendix to understand a few caveats of this quantitative study." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.403, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Investigating Style Similarity in Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "Qualitative Results. In Fig. 5, we showcase a selection of Stable Diffusion-generated images alongside their top 5 corresponding matches from ContraStyles, as determined by the CSD ViT-L feature extractor. The left section of the figure displays images generated from User-generated prompts, while the right section includes images created from Simple prompts. To aid in visual analysis, matches that share a label with the query image are highlighted in green. We can clearly see that the query image and the matches share multiple stylistic elements, such as color palettes and certain artistic features like motifs or textures. We observed that in generations based on user-generated prompts, perceivable style copying typically occurs only when the prompts are shorter and contain elements that are characteristic of the artist's typical content." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.419 + ], + "angle": 0, + "content": "In Figure 6, we present several content-constrained prompt generations and their top-1 matches based on the CLIP ViT-L/14 model versus our CSD model. We observe that the CSD model accurately matches the correct artists to queries even when there is no shared content, only style. This is evident in columns 1, 3, 4, 7, and 8, where our model, CSD matches the correct style elements despite the subjects in the images being quite different. In contrast, the CLIP model still prioritizes content, often leading to mismatches in style." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.441, + 0.756, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.563, + 0.788, + 0.646 + ], + "angle": 0, + "content": "Fig. 7: Does the diffusion model prefer some styles over others? When a prompt contains two style tags, we find that SD 2.1 strongly favors the style that it can best reproduce, we suspect because of a prevalence of the style in training data. In each block, the General Style Similarity(GSS) of the left side artist (red color) is less than the right side artist (blue color). (Ref Fig. 2). The generated image is more biased towards the artist with high GSS score." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Does the model prioritize some artists over others in the prompt? So far in the study we concentrated on the impact of including an artist's name in a style transfer prompt. In this section, we present preliminary findings in scenarios where prompts feature two artists. This is inspired by real-world user prompts from Stable Diffusion and Midjourney, where prompts often include multiple artists. We used the prompt in the style A painting in the style of \\( \\langle \\mathbf{X} \\rangle \\) and \\( \\langle \\mathbf{Y} \\rangle \\), where \\( \\mathbf{X} \\) and \\( \\mathbf{Y} \\) represent different artists. For this study, we selected five artists with varying General Style Similarity (GSS) scores (referenced in Sec. 2). The artists, ranked by descending GSS scores, are Carne Griffiths, Roy Lichtenstein, Gustav Klimt, Pablo Picasso, and Ivan Bilibin. Note that most of the chosen artists have distinct styles that significantly differ from one another." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.129 + ], + "angle": 0, + "content": "G. Somepalli et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.146, + 0.785, + 0.177 + ], + "angle": 0, + "content": "We chose SD-XL Turbo for this analysis because it is trained on deduplicated data, reducing bias towards frequently featured artists in the train set." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.788, + 0.404 + ], + "angle": 0, + "content": "The results for each pair of artists are presented in Fig. 7. Interestingly, even without specific instructions to generate a female subject, most outputs depicted women, reflecting the common subject matter of the artists studied. We also calculated the style similarity scores for each generated image, comparing them to the prototypical styles of the artists in the prompts. In most cases, the style of the artist with the higher GSS score dominated the generated image. To test for potential bias towards the artist positioned first in the prompt (X), we conducted two trials with reversed artist positions. The results were generally consistent, with the dominant style remaining unchanged. However, in the case of Pablo Picasso and Gustav Klimt, this pattern did not hold; the model favored Picasso's cubist style over Klimt's nouveau style, possibly due to the small difference in their GSS scores. While this is not an extensive study, a consistent trend emerged: styles of artists with higher GSS scores, like Leonid Afremov and Carne Griffiths, predominantly influenced the combined style. We leave the comprehensive study on this topic to future work." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.425, + 0.358, + 0.442 + ], + "angle": 0, + "content": "8 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.456, + 0.788, + 0.712 + ], + "angle": 0, + "content": "This study proposes a framework for learning style descriptors from both labeled and unlabeled data. After building a bespoke dataset, ContraStyles, we train a model that achieves state-of-the-art performance on a range of style matching tasks, including DomainNet, WikiArt, and ContraStyles. Then, we show the substantive practical utility of this model through an investigative analysis on the extent of style copying in popular text-to-image generative models. Here, we show the model is capable of determining the factors that contribute to the frequency of style replication. Through cross-referencing of images with style copies and their original prompts, we have discover that the degree of style copying is increasing with prompt complexity. Such complex prompts lead to greater style copying compared to simple one-line prompts. This finding sheds light on the interplay between textual prompts and style transfer, suggesting that prompt design can influence the level of style copying in generative models. Finally, note that the definition of style used in this work is strictly based on artist attribution. We chose this definition because it can be operationalized and used in dataset construction. This definition is certainly not a golden truth, and we look forward to future studies using alternative, or extended definitions." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.734, + 0.402, + 0.753 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "This work was made possible by the ONR MURI program and the AFOSR MURI program. Commercial support was provided by Capital One Bank, the Amazon Research Award program, and Open Philanthropy. Further support was provided by the National Science Foundation (IIS-2212182), and by the NSF TRAILS Institute (2229885)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.403, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Investigating Style Similarity in Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.2 + ], + "angle": 0, + "content": "1. Adobe: Firefly (2023), https://www.adobe.com/sensei/generative-ai/firefly.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.201, + 0.785, + 0.241 + ], + "angle": 0, + "content": "2. Agarwal, S., Karnick, H., Pant, N., Patel, U.: Genre and style based painting classification. In: 2015 IEEE Winter Conference on Applications of Computer Vision. pp. 588-594. IEEE (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.242, + 0.785, + 0.282 + ], + "angle": 0, + "content": "3. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5297-5307 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.785, + 0.323 + ], + "angle": 0, + "content": "4. Bai, Z., Nakashima, Y., Garcia, N.: Explain me the painting: Multi-topic knowledgeable art description generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5422-5432 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.785, + 0.364 + ], + "angle": 0, + "content": "5. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1728-1738 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.785, + 0.405 + ], + "angle": 0, + "content": "6. Balestriero, R., Ibrahim, M., Sobal, V., Morcos, A., Shekhar, S., Goldstein, T., Bordes, F., Bardes, A., Mialon, G., Tian, Y., et al.: A cookbook of self-supervised learning. arXiv preprint arXiv:2304.12210 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.785, + 0.445 + ], + "angle": 0, + "content": "7. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf 2(3), 8 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.785, + 0.487 + ], + "angle": 0, + "content": "8. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.487, + 0.785, + 0.541 + ], + "angle": 0, + "content": "9. Chopra, S., Hadsell, R., LeCun, Y.: Learning a similarity metric discriminatively, with application to face verification. In: 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05). vol. 1, pp. 539-546. IEEE (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.785, + 0.568 + ], + "angle": 0, + "content": "0. Chu, W.T., Wu, Y.L.: Image style classification based on learnt deep correlation features. IEEE Transactions on Multimedia 20(9), 2491-2502 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.568, + 0.785, + 0.608 + ], + "angle": 0, + "content": "1. Decrypt: Greg rutkowski removed from stable diffusion but brought back by ai artists (March 2024), https://decrypt.co/150575/greg-rutkowski-removed-from-stable-diffusion-but-brought-back-by-ai-artists" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.609, + 0.785, + 0.663 + ], + "angle": 0, + "content": "2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.664, + 0.785, + 0.691 + ], + "angle": 0, + "content": "3. Dumoulin, V., Shlens, J., Kudlur, M.: A learned representation for artistic style. arXiv preprint arXiv:1610.07629 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.785, + 0.732 + ], + "angle": 0, + "content": "4. Gairola, S., Shah, R., Narayanan, P.: Unsupervised image style embeddings for retrieval and recognition tasks. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3281-3289 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.785, + 0.772 + ], + "angle": 0, + "content": "5. Garcia, N., Vogiatzis, G.: How to read paintings: semantic art understanding with multi-modal retrieval. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "6. Gatys, L.A., Ecker, A.S., Bethge, M.: A neural algorithm of artistic style. arXiv preprint arXiv:1508.06576 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Gatys, L.A., Ecker, A.S., Bethge, M.: Image style transfer using convolutional neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2414-2423 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.128 + ], + "angle": 0, + "content": "G. Somepalli et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "18. Geiping, J., Goldblum, M., Somepalli, G., Shwartz-Ziv, R., Goldstein, T., Wilson, A.G.: How much data are augmentations worth? an investigation into scaling laws, invariance, and implicit regularization. arXiv preprint arXiv:2210.06441 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.68, + 0.203 + ], + "angle": 0, + "content": "19. Gibson, J.J.: The senses considered as perceptual systems. (1966)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.245 + ], + "angle": 0, + "content": "20. Graham, D.J., Hughes, J.M., Leder, H., Rockmore, D.N.: Statistics, vision, and the analysis of artistic style. Wiley Interdisciplinary Reviews: Computational Statistics 4(2), 115-123 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "21. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "22. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501-1510 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "23. Huang, X., Liu, M.Y., Belongie, S., Kautz, J.: Multimodal unsupervised image-to-image translation. In: Proceedings of the European conference on computer vision (ECCV). pp. 172-189 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.424 + ], + "angle": 0, + "content": "24. Hughes, J.M., Graham, D.J., Jacobsen, C.R., Rockmore, D.N.: Comparing higher-order spatial statistics and perceptual judgements in the stylometric analysis of art. In: 2011 19th European Signal Processing Conference. pp. 1244-1248. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "25. Hughes, J.M., Graham, D.J., Rockmore, D.N.: Stylometrics of artwork: uses and limitations. In: Computer Vision and Image Analysis of Art. vol. 7531, pp. 91-105. SPIE (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "26. Jiang, Q.Y., He, Y., Li, G., Lin, J., Li, L., Li, W.J.: Svd: A large-scale short video dataset for near-duplicate video retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5281-5289 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "27. Joshi, A., Agrawal, A., Nair, S.: Art style classification with self-trained ensemble of autoencoding transformations. arXiv preprint arXiv:2012.03377 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "28. Karayev, S., Trentacoste, M., Han, H., Agarwala, A., Darrell, T., Hertzmann, A., Winnemoeller, H.: Recognizing image style. arXiv preprint arXiv:1311.3715 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.59 + ], + "angle": 0, + "content": "29. Lawrence-Lightfoot, S., Davis, J.H.: The art and science of portraiture. John Wiley & Sons (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.632 + ], + "angle": 0, + "content": "30. Lecoutre, A., Negrevergne, B., Yger, F.: Recognizing art style automatically in painting with deep learning. In: Asian conference on machine learning. pp. 327-342. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "31. Lee, S., Kim, D., Han, B.: Cosmo: Content-style modulation for image retrieval with text feedback. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 802-812 (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.729 + ], + "angle": 0, + "content": "32. Li, J., Yao, L., Hendriks, E., Wang, J.Z.: Rhythmic brushstrokes distinguish van gogh from his contemporaries: findings via automated brushstroke extraction. IEEE transactions on pattern analysis and machine intelligence 34(6), 1159-1176 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.785, + 0.771 + ], + "angle": 0, + "content": "33. Lu, X., Lin, Z., Shen, X., Mech, R., Wang, J.Z.: Deep multi-patch aggregation network for image style, aesthetics, and quality estimation. In: Proceedings of the IEEE international conference on computer vision. pp. 990-998 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "34. Luan, F., Paris, S., Shechtman, E., Bala, K.: Deep photo style transfer. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4990-4998 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "35. Lun, Z., Kalogerakis, E., Sheffer, A.: Elements of style: learning perceptual shape style similarity. ACM Transactions on graphics (TOG) 34(4), 1-14 (2015)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.403, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Investigating Style Similarity in Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "36. Matsuo, S., Yanai, K.: Cnn-based style vector for style image retrieval. In: Proceedings of the 2016 ACM on International Conference on Multimedia Retrieval. pp. 309-312 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "37. Matthews, R.A., Merriam, T.V.: Distinguishing literary styles using neural networks. In: Handbook of neural computation, pp. G8-1. CRC Press (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.786, + 0.259 + ], + "angle": 0, + "content": "38. Menis-Mastromichalakis, O., Sofou, N., Stamou, G.: Deep ensemble art style recognition. In: 2020 International Joint Conference on Neural Networks (IJCNN). pp. 1-8. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.69, + 0.272 + ], + "angle": 0, + "content": "39. Midjourney: Midjourney (nd), https://www.midjourney.com/home" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "40. Mikolov, T., Sutskever, I., Chen, K., Corrado, G.S., Dean, J.: Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems 26 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.355 + ], + "angle": 0, + "content": "41. Park, T., Zhu, J.Y., Wang, O., Lu, J., Shechtman, E., Efros, A., Zhang, R.: Swapping autoencoder for deep image manipulation. Advances in Neural Information Processing Systems 33, 7198-7211 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.397 + ], + "angle": 0, + "content": "42. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1406-1415 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "43. Pernias, P., Rampas, D., Richter, M.L., Pal, C., Aubreville, M.: Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In: The Twelfth International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.466 + ], + "angle": 0, + "content": "44. pharmapsychotic: Clip interrogator. https://github.com/pharmapsychotic/clip-interrogator (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.508 + ], + "angle": 0, + "content": "45. Pizzi, E., Roy, S.D., Ravindra, S.N., Goyal, P., Douze, M.: A self-supervised descriptor for image copy detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14532-14542 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.563 + ], + "angle": 0, + "content": "46. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.787, + 0.605 + ], + "angle": 0, + "content": "47. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.66 + ], + "angle": 0, + "content": "48. Rodriguez, C.S., Lech, M., Pirogova, E.: Classification of style in fine-art paintings using transfer learning and weighted image patches. In: 2018 12th International Conference on Signal Processing and Communication Systems (ICSPCS). pp. 1-7. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.787, + 0.702 + ], + "angle": 0, + "content": "49. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.787, + 0.757 + ], + "angle": 0, + "content": "50. Ruta, D., Motiian, S., Faieta, B., Lin, Z., Jin, H., Filipkowski, A., Gilbert, A., Collomosse, J.: Aladin: all layer adaptive instance normalization for fine-grained style similarity. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11926-11935 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.813 + ], + "angle": 0, + "content": "51. Sablatnig, R., Kammerer, P., Zolda, E.: Hierarchical classification of paintings using face-and brush stroke models. In: Proceedings. Fourteenth International Conference on Pattern Recognition (Cat. No. 98EX170). vol. 1, pp. 172-174. IEEE (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.787, + 0.84 + ], + "angle": 0, + "content": "52. Saleh, B., Elgammal, A.: Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855 (2015)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.129 + ], + "angle": 0, + "content": "G. Somepalli et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "53. Sandoval, C., Pirogova, E., Lech, M.: Two-stage deep learning approach to the classification of fine-art paintings. IEEE Access 7, 41770-41781 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.23 + ], + "angle": 0, + "content": "54. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "55. Silva, J.M., Pratas, D., Antunes, R., Matos, S., Pinho, A.J.: Automatic analysis of artistic paintings using information-based measures. Pattern Recognition 114, 107864 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "56. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.342 + ], + "angle": 0, + "content": "57. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Diffusion art or digital forgery? investigating data replication in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "58. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Understanding and mitigating copying in diffusion models. Advances in Neural Information Processing Systems 36, 47783-47803 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "59. Srinivasa Desikan, B., Shimao, H., Miton, H.: Wikiartvectors: style and color representations of artworks for cultural analysis via information theoretic measures. Entropy 24(9), 1175 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "60. Tenenbaum, J., Freeman, W.: Separating style and content. Advances in neural information processing systems 9 (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "61. Walmer, M., Suri, S., Gupta, K., Shrivastava, A.: Teaching matters: Investigating the role of supervision in vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.535 + ], + "angle": 0, + "content": "62. Wang, J., Yang, H., Fu, J., Yamasaki, T., Guo, B.: Fine-grained image style transfer with visual transformers. In: Proceedings of the Asian Conference on Computer Vision. pp. 841-857 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.563 + ], + "angle": 0, + "content": "63. Wang, S.Y., Efros, A.A., Zhu, J.Y., Zhang, R.: Evaluating data attribution for text-to-image models. arXiv preprint arXiv:2306.09345 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.618 + ], + "angle": 0, + "content": "64. Wilber, M.J., Fang, C., Jin, H., Hertzmann, A., Collomosse, J., Belongie, S.: Bam! the behave artistic media dataset for recognition beyond photography. In: Proceedings of the IEEE international conference on computer vision. pp. 1202-1211 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.646 + ], + "angle": 0, + "content": "65. Willats, J., Durand, F.: Defining pictorial style: Lessons from linguistics and computer graphics. Axiomathes 15(3), 319-351 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.687 + ], + "angle": 0, + "content": "66. Wynen, D., Schmid, C., Mairal, J.: Unsupervised learning of artistic styles with archetypal style analysis. Advances in Neural Information Processing Systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.785, + 0.73 + ], + "angle": 0, + "content": "67. Yao, L., Li, J., Wang, J.Z.: Characterizing elegance of curves computationally for distinguishing morisseau paintings and the imitations. In: 2009 16th IEEE International Conference on Image Processing (ICIP). pp. 73-76. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.757 + ], + "angle": 0, + "content": "68. Zhang, W., Cao, C., Chen, S., Liu, J., Tang, X.: Style transfer via image component analysis. IEEE Transactions on multimedia 15(7), 1594-1601 (2013)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.757 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Investigating Style Similarity in Diffusion Models/e6cae220-dbdc-4ecf-90b6-28f1684a0527_origin.pdf b/2024/Investigating Style Similarity in Diffusion Models/e6cae220-dbdc-4ecf-90b6-28f1684a0527_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..71e7a0b1d15d4ddbcd6ebabd412a58423b437089 --- /dev/null +++ b/2024/Investigating Style Similarity in Diffusion Models/e6cae220-dbdc-4ecf-90b6-28f1684a0527_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a9a02254b362f75b87537bf948168ed86363c52888cdf0c57bfd762caaaedf9 +size 14286526 diff --git a/2024/Investigating Style Similarity in Diffusion Models/full.md b/2024/Investigating Style Similarity in Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..55c0c2bb52bad5c093f9ecd9b0998e1606a8a3f1 --- /dev/null +++ b/2024/Investigating Style Similarity in Diffusion Models/full.md @@ -0,0 +1,294 @@ +# Investigating Style Similarity in Diffusion Models + +Growthami Somepalli\*1, Anubhav Gupta\*1, Kamal Gupta\*1, Shramay Palta\*1, Micah Goldblum\*3, Jonas Geiping\*2, Abhinav Shrivastava\*1, and Tom Goldstein\*1 + +1 University of Maryland, College Park + +2 ELLIS Institute, MPI for Intelligent Systems + +3 Columbia University + +4 https://somepago.github.io/csd + +![](images/33ccc3a732cd6c20c46d042c462473ee361023e66ad35bd3cfacabe4ce8c2a3c.jpg) +Real paintings + +![](images/97a368689dba65163f359a2ddd324c8f8366910869cc620896f9acb975979c3e.jpg) +MidJourney + +![](images/58b9eb52ab258da014fbd42d2055b40dff2cf6ac31cce68fd30143b1ed26c538.jpg) +2.1 + +![](images/1d858238181bbb2985a5bd8d48cb438624b1201bb35627e8d17040bc094cc814.jpg) + +![](images/cd738bf053c4382f2f80e478bdbe15cabffb239c63dcf22560a357ad1ed31068.jpg) +1.4 + +![](images/39763cba641acaee54e33b5e59d0525927c4d63ae7acd1e59378a95926d2bbd8.jpg) + +![](images/67adbfc7450fdcb80e7c411bd2eebad2543abefc10dbd3ef913e8d7c999351cc.jpg) + +![](images/291721b4c89ab6c5a699ee8eb49f34467111c717a3fe898255e9760867c157a0.jpg) + +![](images/87d5451b45700ba592927bd1134d0bf7a0cc9530c5f6bcf85cc98371b354d413.jpg) + +![](images/779265e8f59b906ad8956b13068863ba4e4e0414a34dc96abaa465bb44c5ed61.jpg) + +![](images/8c19ca291a8fe7289f6c4e1fc3cdcec79a4559d92144cdb71796b4e3ab6b4799.jpg) + +![](images/54b65c24568733d08bf36729c68d5f90abd786d69ca35bbc2b409770a39bf80c.jpg) + +![](images/9ef29bb886a569ef72659c0c3b6f742ac04159e40350b6605e009cdb8270d01d.jpg) + +![](images/c3811ab33825d3a5417f1c3c934e3a2c27223fd66be2625135dd0a39af80bc48.jpg) + +![](images/f90092ba02fc4179b6e64c9c5d8d9da4eb9a67c6864641c13e4a124b24e2c59f.jpg) +Fig. 1: Original artwork of 6 popular artists and the images generated in the style of these artists by three popular text-to-image generative models. The numbers displayed below each image indicate the similarity of the generated image with the artist's style using the proposed method. A high similarity score suggests a strong presence of the artist's style elements in the generated image. Based on our analyses, we postulate that three artists on the right were removed (or unlearned) from SD 2.1 while they were present in MidJourney and SD 1.4. Please refer to Section 2 for more details. + +![](images/b8b2038b49eab9cba87f72bad42d84fd78c45ce931333a3253a4eaca435862a7.jpg) + +![](images/f6e27c024edf7205ff7698356d0c36f38defa1885e25c7f5616aaffdc91817b4.jpg) + +![](images/96abf52c86408e3929f388ef765e9007bdad54267970aeee97e50cbfc87415af.jpg) + +![](images/2e29ad371079ba05a0e8ae8b2b2e4f623cd447a0ac9c3840b1a5fd4bb9b9cdaa.jpg) + +![](images/eaceb94b33dbc00f9f421cea387eee0934b0a4c3c5d6fc85d27a3afa1d5ed335.jpg) + +Abstract. Generative models are now widely used by graphic designers and artists. Prior works have shown that these models remember and often replicate content from their training data during generation. Hence as their proliferation increases, it has become important to perform a database search to determine whether the properties of the image are attributable to specific training data, every time before a generated image is used for professional purposes. Existing tools for this purpose focus on retrieving images of similar semantic content. Meanwhile, many artists are concerned with style replication in text-to-image models. We present a framework for understanding and extracting style descriptors from images. Our framework comprises a new dataset curated using the + +insight that style is a subjective property of an image that captures complex yet meaningful interactions of factors including but not limited to colors, textures, shapes, etc. We also propose a method to extract style descriptors that can be used to attribute style of a generated image to the images used in the training dataset of a text-to-image model. We showcase promising results in various style retrieval tasks. We also quantitatively and qualitatively analyze style attribution and matching in the Stable Diffusion model. + +Keywords: Image Style $\cdot$ Style Similarity $\cdot$ Generative models + +# 1 Introduction + +Diffusion-based image generators like Stable Diffusion [49], DALL-E [47] and many others [1, 7, 39, 43] learn artistic styles from massive captioned image datasets that are scraped from across the web [54]. Before a generated image is used for commercial purposes, it is wise to understand its relationship to the training data and the origins of its design elements and style attributes. Discovering and attributing these generated images, typically done with image similarity search, is hence becoming increasingly important. Such dataset attribution serves two purposes. It enables users of generated images to understand potential conflicts, associations, and social connotations that their image may evoke. It also enables artists to assess whether and how generative models are using elements of their work. + +Despite a long history of research [60], recovering style from an image is a challenging and open problem in Computer Vision. Many existing retrieval methods [8, 45, 46] for large training datasets focus primarily on matching semantic content between a pair of images. Understanding the origin of the style present in a generated image, however, is much less well understood. To address this gap, we propose a self-supervised objective for learning style descriptors from images. Standard augmentation-based SSL pipelines (e.g. SimCLR and variants) learn feature vectors that are invariant to a set of augmentations. Typically, these augmentations preserve semantic content and treat style as a nuisance variable. In contrast, we choose augmentations that preserve stylistic attributes (colors, textures, or shapes) while minimizing content. Unfortunately, SSL is not enough, as style is inherently subjective, and therefore a good style extractor should be aligned with human perceptions and definitions of style. For this reason, we curate a style attribution dataset, ContraStyles, in which images are associated with the artist that created them. + +By training with both SSL and supervised objectives, we create a high-performance model for representing style. Our model, CSD, outperforms other large-scale pre-trained models and prior style retrieval methods on standard datasets. Using CSD, we examine the extent of style replication in the popular open-source text-to-image generative model Stable Diffusion [49], and consider different factors that impact the rate of style replication. + +To summarize our contributions, we (1) propose a style attribution dataset ContraStyles, associating images with their styles, (2) introduce a multi-label contrastive learning scheme to extract style descriptors from images and show the efficacy of the scheme by zero-shot evaluation on public domain datasets such as WikiArt and DomainNet (3) We perform a style attribution case study for one of the most popular text-to-image generative models, Stable Diffusion, and propose indicators of how likely an artist's style is to be replicated. Code, artifacts, and dataset links are available at https://somepago.github.io/csd. + +# 2 Motivation + +We present a case study that shows how style features can be used to interrogate a generative model, and provide utility to either artists or users. We consider the task of analyzing a model's ability to emulate an artist's style, and of attributing the style of an image to an artist. We begin by curating a list of 96 artists, primarily sourced from the WikiArt database, supplemented by a few contemporary artists who are notably popular within the Stable Diffusion community5. For each artist, we compute a prototype vector by averaging the embeddings of their paintings using our proposed feature extractor, CSD ViT-L. Next, we generate an image for each artist using Stable Diffusion 2.1 with a prompt in the format A painting in the style of . We compute the dot product similarity between each generated image's embedding and the artist's prototype. This process was repeated multiple times for each artist, and we plot mean results in Fig. 2. We refer to this quantity as the General Style Similarity (GSS) score for an artist, as it measures how similar a generated image is to a typical image from that artist while using our style representation model. We also plot an analogous style similarity score, but using "content-constrained" prompts. For instance, one prompt template is A painting of a woman doing style of where X is the name of the artist and Y is some setting like reading a book or holding a baby etc. See Sec. 7 for all templates. + +Each point in Fig. 2 represents an artist. Notice that GSS scores are highly correlated with content-constrained scores, indicating that our feature vectors represent style more than semantic content. Our findings reveal that SD 2.1 is much more capable of emulating some artists than others. Artists like Leonid Afremov, Georges Seurat exhibit high style similarity scores, and visual inspection of generated images confirms that indeed their style is emulated by the model (Fig. 1 - Original artwork vs SD 2.1). On the other end of the spectrum, artists such as Ruan Jia and Greg Rutkowski showed low similarity scores, and likewise the generated images bear little resemblance to the artists' work. Interestingly, after completing this study, we discovered that Greg Rutkowski's work was excluded from the training data for the Stable Diffusion 2.1 model, as reported by [11]. + +This demonstrates that the Style Similarity score can be used by artists to quantify how well a model emulates their style, or it can be used by users to + +ascertain whether a generated image contains stylistic elements associated with a particular artist. After a thorough inspection of the generations from 96 artists, we hypothesize that a single-image Style Similarity score below 0.5 indicates the absence of the artist's style, whereas a score above 0.8 strongly indicates its presence. + +In Figure 1, we show original artworks for 6 artists, and generations from MidJourney [39], Stable Diffusion 2.1 and Stable Diffusion 1.4 [49] for each of these artists. The 3 artists on the left side have high GSS while the ones on the right side have low GSS. Below each generated image, we display the similarity against the artist's prototype vector. We see high image similarity scores in the Midjourney generations and qualitatively these images look stylistically similar to artists' original artworks. We also see the interesting cases of Greg Rutkowski, Ruan Jia, and Amano whose style is captured by Stable Diffusion 1.4, while being notably absent in Stable Diffusion 2.1. This finding is in line with reports suggesting that some of these artists were removed from the training data + +of Stable Diffusion 2.1 [11]. Based on this analysis, we postulate that Ruan Jia, Wadim Kashin, Anton Fadeev, Justin Gerard, and Amano were also either excluded from the training data or post-hoc unlearned/removed from Stable Diffusion 2.1. + +![](images/d1f6e6683b8888ef099762a2ff03471432b73f50a7d1ebc04caf8b92ae5584d5.jpg) +Fig. 2: Style similarity of Stable Diffusion 2.1 generated images against the artist's prototypical representation. On the X-axis, the similarities are depicted when the prompt is not constrained, while the Y-axis represents similarity when the prompt is constrained to generate an image of a "woman" in the artist's style. + +# 3 What is style? + +The precise definition of "style" remains in contention, but many named artistic styles (e.g., cubism, impressionism, etc...) are often associated with certain artists. We leverage this social construct, and define style simply as the collection of global characteristics of an image that are identified with an artist or artistic movement. These characteristics encompass various elements such as color usage, brushstroke techniques, composition, and perspective. + +Related work. Early computer vision algorithms attempted to model style using low-level visual features like color histograms, texture patterns, edge detection, and shape descriptors. Other computational techniques involve rule-based systems, such as the presence of specific compositional elements, the use of spe + +cific color palettes, or the presence of certain brushstroke patterns to identify specific style characteristics [19, 20, 24, 25, 29, 32, 35, 37, 51, 55, 59, 65, 67]. + +Modern studies have focused on the task of transferring style from one image to another [13, 17, 22, 34, 41, 62, 68]. Some works have also concentrated on style classification [2, 4, 10, 15, 27, 28, 30, 33, 38, 48, 53]. A limited number of studies address in-the-wild style quantification, matching, and retrieval [14, 23, 36, 50, 66]. In their seminal work, Gatys et al. [17] introduced Gram Matrices as style descriptors and utilized an optimization loop to transfer style. Another approach proposed by Luan et al. [34] includes a photorealism regularization term to prevent distortions in the reconstructed image. Zhang et al. [68] formulated style transfer using Markov random fields. Beyond Gram-based style representation, Chu et al. [10] explored various other types of correlations and demonstrated performance variations. + +In a recent work by Lee et al [31], two separate neural network modules were used - one for image style and another for image content - to facilitate image style retrieval. In the most recent related research, Wang et al. [63] developed an attribution model trained on synthetic style pairs, designed to identify stylistically similar images. In contrast to this approach, our method leverages real image pairs, curated automatically through their caption annotations. Despite our training dataset being approx. $16\%$ the size of training data used in [63], we demonstrate that our model significantly outperforms this method on many zero-shot style matching tasks in the later sections. + +# 4 Creating a dataset for style attribution + +While many large web datasets now exist, we need one that contains wide variations in artistic styles, and also labels that be used for downstream style retrieval evaluation. Some large-scale datasets specifically designed to handle such a challenge, like BAM [64], are not available in the public domain and others like WikiArt [52] are not large enough to train a good style feature extractor. In the following section, we propose a way to curate a large style dataset out of the LAION [54] Aesthetics $6+$ dataset. + +ContraStyles: A dataset for style distillation. We curate our own dataset as a subset of LAION [54]. We start off with the 12M image-text pairs with predicted aesthetics scores of 6 or higher. Note that this dataset is extremely unbalanced, with some popular artists appearing much more frequently than others. Also, a large number of images are duplicated within the dataset which is a major issue for the text-to-image models trained on this data [57,58]. Furthermore, the image captions within the data are often noisy and are often missing a lot of information. We address these challenges and propose a new subset of LAION-Aesthetics consisting of 511,921 images, and 3840 style tags, where each image can have one or more tags. We use this dataset for training our models. + +We begin with a bank of styles collated in previous work for image understanding with the CLIP Interrogator [44]. This bank of styles was curated based on typical user prompts for Stable Diffusion. We combine the bank of artists, + +medium, and movement references, to a shortlist of 5600 tags. We then search for these tags in the 12M LAION-Aesthetics captions and shortlist the images that have at least one of the tags present. We further filter out the tags which have over 100,000 hits in the dataset since human inspection found that they refer to common phrases like 'picture' or 'photograph' that do not invoke a distinct style. After discarding images with an unavailable URL, we are left with about 1 million images and 3840 tags. We further deduplicate the images using SSCD [45] with a threshold of 0.8 and merge the tags of images that are near copies of each other. As a by-product, the dedduplication also helps with the missing tags in the images, since we can simply merge the text labels of duplicate images. After dedduplication, we are left with 511,921 images. + +# 5 Contrastive Style Descriptors (CSD) + +Self-Supervised Learning. Many successful approaches to SSL [61] use a contrastive [9] approach, where two views (or augmentations) of the same image in the dataset are sampled and passed to a deep network to get their respective image descriptors. The network is trained to maximize the similarity of two views of the same image and minimize agreement with other images in the batch. Standard choices for augmentations include color jitter, blurring, grayscale, etc., as these alter the image's visual properties while preserving content. While these are good augmentations for object recognition tasks, they train the network to ignore image attributes associated with style. + +Our approach relies on a training pipeline with two parts. First, we use contrastive SSL, but with a set of augmentations that are curated to preserve style. Second, we align our model with human perceptions of style by training on our labelled ContraStyles dataset described in Section 4. + +Proposed Approach. We seek a model for extracting image descriptors that contain concise and effective style information. To be a useful, the model should be invariant to semantic content and capable of disentangling multiple styles. + +Given a dataset of $N$ labeled images $\{\mathbf{x}_i, l_i\}_{i=1}^N$ , where each image can have one or more labels from a set of $L$ labels, we define the label vector of the $i^{\text{th}}$ image as $l_i = (c_1, c_2, \ldots, c_L)$ , where each $c_k \in \{0, 1\}$ . As mentioned in the previous section, our multi-label dataset consists of $N = 511,921$ images and $L = 3,840$ style tags. We consider a mini-batch of $B$ images. Each of the images are passed to a Vision Transformer (ViT) [12] backbone, and then projected to a $d$ -dimensional vector. We consider two variants of ViT (ViT-B and ViT-L). + +Our style descriptors $f_{\mathrm{ViT}}(\mathbf{x}_i) \in \mathbb{R}^d$ are then used to create a matrix of pairwise cosine similarity scores $s_{i,j} = \cos (f_{\mathrm{ViT}}(\mathbf{x}_i), f_{\mathrm{ViT}}(\mathbf{x}_j))$ . In order to compute our multi-label contrastive loss (MCL), we also compute the groundtruth similarity scores as $\hat{s}_{i,j} = \mathbb{1}\left(l_i^{\mathrm{T}}l_j\right)$ , where $\mathbb{1}$ is the indicator function that returns 1 if any of the labels of the images $i, j$ match. Our final loss term reduces to: + +$$ +\mathcal {L} _ {\mathrm {M C L}} = - \hat {s} _ {i, j} \log \frac {\exp \left(s _ {i , j}\right) / \tau)}{\sum_ {k \neq j} \exp \left(s _ {i , k}\right) / \tau}), \tag {1} +$$ + +where $\tau$ is the temperature fixed during the training. + +Since our supervised dataset is modest in size, we add a self-supervised objective. We sample two "views" (augmentations) of each image in a batch and add a contrastive SSL term. Standard SSL training routines (e.g., MoCo, SimCLR, BYOL etc.) choose augmentations so that each pair of views has the same semantic content, but different style content. These augmentations typically include Resize, Horizontal Flips, Color Jitter, Grayscale, Gaussian Blur, and Solarization [6, 18]. For our purposes, we depart from standard methods by excluding photometric augmentations (Gaussian Blur, Color Jitter), as they alter the style of the image. We keep the following spatial augmentations - Horizontal Flips, Vertical Flips, Resize and Rotation as they keep style intact. + +The overall loss function is a simple combination of the multilabel contrastive loss and self-supervised loss $\mathcal{L} = \mathcal{L}_{\mathrm{MCL}} + \lambda \mathcal{L}_{\mathrm{SSL}}$ . During inference, we use the final layer embedding and the dot product to compute style similarity between any two images. In our experiments, we found that initializing weights to CLIP [46] ViT-B and ViT-L improves performance. + +# 6 Results + +Training details. We present the results for two variants of our model CSD ViT-B and CSD ViT-L version. Both the models are initialized with respective CLIP variant checkpoints and are finetuned for 80k iterations on the ContraStyles dataset on 4 A4000/A5000 GPUs. We use an SGD optimizer with momentum 0.9 and learning rate of 0.003 for the projection layer and $1e - 4$ for the backbone. Our mini-batch size per GPU is 16. We use $\lambda = 0.2$ and $\tau = 0.1$ for the final model. The training takes about 8 hours for the base model and around 16 hours for the large model. See the Appendix for more details and ablations. + +Task. We perform zero-shot evaluation across multiple datasets on a style-retrieval task. Following [3, 26], we split each dataset into two parts: Database and Query. Given a query image at test time, we evaluate whether we can find the ground-truth style in its nearest neighbors from the database. + +Baselines. We compare our model against a recent style attribution model GDA [63] which is trained via fine-tuning on paired synthetic style data, and VGG [17, 56] Gram Matrices which are often used for neural style transfer applications. Further we compare with CLIP [46] models supervised with freeform text captions, and with other self-supervised models such as DINO [8], MoCo [21], SSCD [45]. We use the embeddings from the last layer for each of these models except for VGG where we use the Gram Matrix [17] of the last + +layer. We skipped evaluations of [14, 23, 50] since both pre-trained models and training data are not available. + +Metrics. We do nearest neighbour searches for $k \in [1, 10, 100]$ , and report Recall@k, mAP@k. We use the standard definitions of these metrics from the retrieval literature. Like [5], we define positive recall as the existence of a correct label in top-N matches and no recall when none of the top-N matches share a label with the query. Similarly, mAP is defined as average over precision at each rank in N, and then averaged over all queries. + +Table 1: mAP and Recall metrics on DomainNet and WikiArt datasets. Our model consistently performs the best in all cases except one, against both self-supervised and style attribution baselines. + +
DomainNet (mAP@k)WikiArt (mAP@k)DomainNet (Recall@k)WikiArt (Recall@k)
Method110100110100110110100
VGG Gram [16]---25.919.411.4--25.952.780.4
DINO ViT-B/16 [8]69.468.266.244.033.418.969.493.744.069.488.1
DINO ViT-B/8 [8]72.270.969.346.935.920.472.293.846.971.088.9
SSCD RN-50 [45]67.665.962.036.026.514.867.695.036.062.185.4
MOCO ViT-B/16 [21]71.971.169.644.033.218.872.094.044.069.088.0
CLIP ViT-B/16 [46]73.773.071.352.242.026.073.794.552.278.393.5
GDA CLIP ViT-B [63]62.961.659.325.621.014.162.992.325.656.683.8
GDA DINO ViT-B [63]69.568.166.145.534.619.769.593.445.575.889.0
GDA ViT-B [63]67.165.664.242.632.218.267.193.642.667.687.1
CSD ViT-B (Ours)78.377.576.056.246.128.778.394.356.280.393.6
CLIP ViT-L [46]74.073.572.259.448.831.574.094.859.482.995.1
CSD ViT-L (Ours)78.377.876.564.5653.8235.6578.394.564.5685.7395.58
+ +Evaluation Datasets. DomainNet [42] consists of an almost equal number of images from six different domains: Clipart, Infograph, Painting, Quickdraw, Real, and Sketch. Upon examination, we observed a strong stylistic resemblance between the Quickdraw and Sketch domains, leading us to exclude Quickdraw from our analysis. The dataset's content information was utilized to categorize the images into two main clusters of content classes. This clustering was achieved through the application of word2vec [40]. The images within the smaller cluster were designated as part of the Query set (20,000 images), and images in the bigger cluster to Database (206768 images). And the second dataset we evaluate all the models is, WikiArt [52]. It consists of 80096 fine art images spread across 1119 artists and 27 genres. We randomly split the dataset into 64090 Database and 16006 Query images. We use the artist as a proxy for the style since there is large visual variation within each genre for them to be considered as independent styles. Under this setting, WikiArt is a challenging retrieval task as the chance probability of successful match is just .09% while it is 20% for DomainNet. + +# 6.1 Analyses and observations + +In Table 1, we report the metrics for all the baselines considered and the proposed CSD method using $k$ nearest neighbors on 2 datasets - DomainNet and WikiArt. Note that while style and content are better separated in the case of DomainNet, WikiArt consists of more fine-grained styles and has more practical use cases for style retrieval. Loosely speaking, mAP@ $k$ determines what percentage of the nearest neighbors that are correct predictions, while the recall determines what percentage of queries has a correct match in the top- $k$ neighbors. Our model CSD consistently outperformed all the pretrained feature extractors as well as the recent attribution method GDA [63] on both WikiArt and DomainNet evaluations. Note that all models are evaluated in a zero-shot setting. We see the most gains in the WikiArt dataset which is more challenging with chance probability of only $0.09\%$ . When we look at mAP@1, which is same as top-1 accuracy, our base model outperforms the next best model by $5\%$ points on WikiArt and $4.6\%$ points on DomainNet. Our large model out-performs the closest large competitor by similar margins. Given the complexity of the task, these improvements are non-trivial. + +We attribute the improvements to a couple of factors, (1) The multi-label style contrastive loss on our curated ContraStyles dataset is quite helpful in teaching the model right styles (2) We hypothesize that these SSL models become invariant to styles because the way they were trained, but we are careful to not strip that away in our SSL loss component by carefully curating non-photometric augmentations in training. + +Error Analysis. Even though our model outperforms the previous baselines, our top-1 accuracy for the WikiArt style matching task is still at 64.56. We tried to understand if there is a pattern to these errors. For example, our model is consistently getting confused between impressionist painters claude monet, gustave loiseau, and alfred sisley, all of whom painted many landscapes. They depicted natural scenes, including countryside views, rivers, gardens, and coastal vistas. Another example is pablo picasso and georges braque, who are both cubist painters. Given the impracticality of analyzing all 1,119 artists in the dataset, we opted for a macroscopic examination by categorizing errors at the art movement level. This approach is visualized in the heatmap presented in Fig. 3. In the heatmap, we see most of the errors concentrated along the diagonal, indicating that while the model often correctly identifies the art movement, it struggles to pinpoint the exact artist. There are instances of off-diagonal errors where the model incorrectly identifies both the artist and their art movement. For example, many Post Impressionism and Realism paintings are assigned to Impressionism artists. Upon closer examination, it becomes apparent that they closely align in terms of historical timeline and geographical origin, both being from Europe. This analysis indicates the nuanced nature of style detection and retrieval in Computer Vision. It suggests that the upper limit for accuracy in this task might be considerably lower than $100\%$ , even for a typical human evaluator, due to the inherent subtleties and complexities involved. + +![](images/e9748336a4e095dc167b6706602aea38f9aa6307127c1b62f7a51cf675d7a6c1.jpg) +Fig. 3: Confusion Matrix of errors in WikiArt: Art movements are predicted correctly. Errors occur in cases where movements share the same historical timeline and/or are derived from the same earlier movement. + +![](images/13e130720ef7a42129e74a87bf69aa76475dc661dd963502fcd7ba0801414e5f.jpg) +Fig. 4: Human study on Style Retrieval: Turns out untrained humans are worse than many feature extractors on matching images from the same artist. + +# 6.2 Human Study + +To understand how our models compare to untrained humans, we conducted a small survey on style matching on 30 humans (excluding the authors). Following the convention in other papers [14,28,50,62] and this paper, we assume, 2 images from same artist can be considered stylistic matches. For each query image, we gave 4 answer images out of which only one is from the same artist and hence is the right answer, so chance accuracy is $25\%$ . We used the Artcive dataset introduced in [63] to create this survey and we collected 3 responses per item to break any ties. We present the results in Fig 4. Most interestingly, untrained humans are worse than many feature extractors at this task. SSCD is the only feature extractor that did worse than humans. Our model, CSD outperforms all the baselines on this dataset as well. This underpins the difficulty of style matching and also highlights the superior performance of our feature extractor. + +# 7 Studying style in the wild: Analysis of Stable Diffusion + +In the previous section, we have quantitatively shown that our model Contrastive Style Descriptors outperforms many baselines on style matching task. Now we try to address the question, Can we do style matching on Stable Diffusion generated images? To answer this question, we first curated multiple synthetic image collections using Stable Diffusion v 2.1 [49] and then compared them against the "ground truth" style matches on ContraStyles dataset. + +Creating synthetic style dataset. The first challenge in curating synthetic images through prompts is the choice of prompts to be used for the generation. There have been no in-depth quantitative studies of the effect of prompts on generation styles. For this analysis, we chose 3 types of prompts. + +Table 2: mAP and Recall of SD 2.1 generated synthetic datasets based on Simple prompts and User-generated prompts + +
Simple prompts (mAP@k)User-generated (mAP@k)Simple prompts (Recall@k)User-generated (Recall@k)
Method110100110100110100110100
GDA - DINO11.610.27.64.454.594.2411.628.152.834.4525.2267.18
CSD-ViTB17.5316.5612.685.855.965.5817.5338.6561.855.8529.2674.2
CLIP ViT-L/1422.320.416.16.15.75.122.344.566.26.126.071.7
CSD (Ours)24.523.318.55.75.95.624.547.267.55.726.571.8
+ +1. User-generated prompts: We used a Stable Diffusion Prompts $^6$ dataset of 80,000 prompts filtered from Lexica.art. We used the test split and then filtered the prompts to make sure at least one of the keywords from the list we curated in Section 4 is present. We then sampled 4000 prompts from this subset for query split generation. +2. Simple prompts: We randomly sampled 400 artists which appeared most frequently in user-generated prompts we analysed. We format the prompt as A painting in the style of , and we generate 10 images per prompt by varying the initialization seed. +3. Content-constrained prompts: We wanted to understand if we can detect style when we constrain the model to generate a particular subject/human in the style of an artist. For this, we used the prompt A painting of a woman in the style of or A painting of a woman reading in the style of etc., a total of 5 variations per subject repeated two times. We experimented with subjects, woman,dog and house in this study. We provide the exact templates in the appendix. + +Table 3: mAP and Recall of SD 2.1 generated synthetic datasets based on Content-constrained prompts + +
Dog (mAP@k)House (mAP@k)Woman (mAP@k)Dog (Recall@k)House (Recall@k)Woman (Recall@k)
Method110100110100110100110100110100110100
GDA-DINO2.2821.63.93.62.72.22.322.288.6828.733.912.3322.210.128.9
CSD-ViT-B4.54.313.614.64.293.877.557.836.424.514.3634.884.615.03397.5520.142.46
CLIP ViT-L/142.32.21.94.54.23.67.47.16.22.29.829.94.513.835.37.419.041.6
CSD (Ours)4.94.84.26.46.25.410.810.18.64.914.534.56.417.840.610.823.444.2
+ +We generate 4000 images for each prompt setting using Stable Diffusion 2.1. There is only one style keyword in simple and content-constrained prompts, which we also use as a ground truth label for matching tasks. However, user-generated prompts can have multiple style labels within the caption, and we consider all of them as ground-truth labels. + +![](images/c7ab3dff347505d675fe37a2fbf2c41836ec36b8fd0730ccbafb668bcd65387c.jpg) +Fig. 5: Nearest "style" neighbors. For each generated image (referred to as SD Gen), we show the top 5 style neighbors in CSD using our feature extractor. The green and red box around the image indicates whether or not the artist's name used to generate the SD image was present in the caption of the nearest neighbor. + +![](images/3fb9c16ec1ce4d7f28c2b7738a291cfa042fe75fda090ae99cd0d4f3452a9f62.jpg) + +![](images/46379fab2ec7ba3ad4f5608dc6fe95862b66fe6a4499acfe4df62fdc8e24a554.jpg) +Fig. 6: Top row: Images generated by Stable Diffusion. Middle and Bottom rows: Top matches retrieved by CLIP vs CSD (ours) respectively. CLIP is consistently biased towards image content, for instance retrieving image of a dog in the Column 1, 3, 4, or image of mother and baby in Column 7 or 8. Our method emphasize less on the content but more on the image styles. Please refer to the Appendix for the prompts. + +Style retrieval on generated images. In Tab. 2, we show the retrieval results for Simple and User-generated prompts. We also compare our results with the second-best performing model in the previous section, CLIP ViT-L, and a recent style attribution model GDA [63]. We observe that our method outperforms CLIP on Simple prompt dataset. For User-generated prompts, the performance metrics are closer to CLIP model, but it's important to note that these prompts are inherently more complex. This complexity results in a different label distribution in the query set for the two types of prompts we examine, leading to varied metric ranges in each case. Additionally, our method consistently outperforms both baselines in content-constrained scenarios, as evidenced in Table 3. This indicates the robustness and effectiveness of our approach in dealing with a variety of prompt complexities and content specifications. We refer the reader to Appendix to understand a few caveats of this quantitative study. + +Qualitative Results. In Fig. 5, we showcase a selection of Stable Diffusion-generated images alongside their top 5 corresponding matches from ContraStyles, as determined by the CSD ViT-L feature extractor. The left section of the figure displays images generated from User-generated prompts, while the right section includes images created from Simple prompts. To aid in visual analysis, matches that share a label with the query image are highlighted in green. We can clearly see that the query image and the matches share multiple stylistic elements, such as color palettes and certain artistic features like motifs or textures. We observed that in generations based on user-generated prompts, perceivable style copying typically occurs only when the prompts are shorter and contain elements that are characteristic of the artist's typical content. + +In Figure 6, we present several content-constrained prompt generations and their top-1 matches based on the CLIP ViT-L/14 model versus our CSD model. We observe that the CSD model accurately matches the correct artists to queries even when there is no shared content, only style. This is evident in columns 1, 3, 4, 7, and 8, where our model, CSD matches the correct style elements despite the subjects in the images being quite different. In contrast, the CLIP model still prioritizes content, often leading to mismatches in style. + +![](images/daa2e23316851fe8e753543045eb3e3d5615abf083f6395f1de25fee444511ff.jpg) +Fig. 7: Does the diffusion model prefer some styles over others? When a prompt contains two style tags, we find that SD 2.1 strongly favors the style that it can best reproduce, we suspect because of a prevalence of the style in training data. In each block, the General Style Similarity(GSS) of the left side artist (red color) is less than the right side artist (blue color). (Ref Fig. 2). The generated image is more biased towards the artist with high GSS score. + +Does the model prioritize some artists over others in the prompt? So far in the study we concentrated on the impact of including an artist's name in a style transfer prompt. In this section, we present preliminary findings in scenarios where prompts feature two artists. This is inspired by real-world user prompts from Stable Diffusion and Midjourney, where prompts often include multiple artists. We used the prompt in the style A painting in the style of $\langle \mathbf{X} \rangle$ and $\langle \mathbf{Y} \rangle$ , where $\mathbf{X}$ and $\mathbf{Y}$ represent different artists. For this study, we selected five artists with varying General Style Similarity (GSS) scores (referenced in Sec. 2). The artists, ranked by descending GSS scores, are Carne Griffiths, Roy Lichtenstein, Gustav Klimt, Pablo Picasso, and Ivan Bilibin. Note that most of the chosen artists have distinct styles that significantly differ from one another. + +We chose SD-XL Turbo for this analysis because it is trained on deduplicated data, reducing bias towards frequently featured artists in the train set. + +The results for each pair of artists are presented in Fig. 7. Interestingly, even without specific instructions to generate a female subject, most outputs depicted women, reflecting the common subject matter of the artists studied. We also calculated the style similarity scores for each generated image, comparing them to the prototypical styles of the artists in the prompts. In most cases, the style of the artist with the higher GSS score dominated the generated image. To test for potential bias towards the artist positioned first in the prompt (X), we conducted two trials with reversed artist positions. The results were generally consistent, with the dominant style remaining unchanged. However, in the case of Pablo Picasso and Gustav Klimt, this pattern did not hold; the model favored Picasso's cubist style over Klimt's nouveau style, possibly due to the small difference in their GSS scores. While this is not an extensive study, a consistent trend emerged: styles of artists with higher GSS scores, like Leonid Afremov and Carne Griffiths, predominantly influenced the combined style. We leave the comprehensive study on this topic to future work. + +# 8 Conclusion + +This study proposes a framework for learning style descriptors from both labeled and unlabeled data. After building a bespoke dataset, ContraStyles, we train a model that achieves state-of-the-art performance on a range of style matching tasks, including DomainNet, WikiArt, and ContraStyles. Then, we show the substantive practical utility of this model through an investigative analysis on the extent of style copying in popular text-to-image generative models. Here, we show the model is capable of determining the factors that contribute to the frequency of style replication. Through cross-referencing of images with style copies and their original prompts, we have discover that the degree of style copying is increasing with prompt complexity. Such complex prompts lead to greater style copying compared to simple one-line prompts. This finding sheds light on the interplay between textual prompts and style transfer, suggesting that prompt design can influence the level of style copying in generative models. Finally, note that the definition of style used in this work is strictly based on artist attribution. We chose this definition because it can be operationalized and used in dataset construction. This definition is certainly not a golden truth, and we look forward to future studies using alternative, or extended definitions. + +# Acknowledgements + +This work was made possible by the ONR MURI program and the AFOSR MURI program. Commercial support was provided by Capital One Bank, the Amazon Research Award program, and Open Philanthropy. Further support was provided by the National Science Foundation (IIS-2212182), and by the NSF TRAILS Institute (2229885). + +# References + +1. Adobe: Firefly (2023), https://www.adobe.com/sensei/generative-ai/firefly.html +2. Agarwal, S., Karnick, H., Pant, N., Patel, U.: Genre and style based painting classification. In: 2015 IEEE Winter Conference on Applications of Computer Vision. pp. 588-594. IEEE (2015) +3. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5297-5307 (2016) +4. Bai, Z., Nakashima, Y., Garcia, N.: Explain me the painting: Multi-topic knowledgeable art description generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5422-5432 (2021) +5. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1728-1738 (2021) +6. Balestriero, R., Ibrahim, M., Sobal, V., Morcos, A., Shekhar, S., Goldstein, T., Bordes, F., Bardes, A., Mialon, G., Tian, Y., et al.: A cookbook of self-supervised learning. arXiv preprint arXiv:2304.12210 (2023) +7. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf 2(3), 8 (2023) +8. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021) +9. Chopra, S., Hadsell, R., LeCun, Y.: Learning a similarity metric discriminatively, with application to face verification. In: 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05). vol. 1, pp. 539-546. IEEE (2005) +0. Chu, W.T., Wu, Y.L.: Image style classification based on learnt deep correlation features. IEEE Transactions on Multimedia 20(9), 2491-2502 (2018) +1. Decrypt: Greg rutkowski removed from stable diffusion but brought back by ai artists (March 2024), https://decrypt.co/150575/greg-rutkowski-removed-from-stable-diffusion-but-brought-back-by-ai-artists +2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +3. Dumoulin, V., Shlens, J., Kudlur, M.: A learned representation for artistic style. arXiv preprint arXiv:1610.07629 (2016) +4. Gairola, S., Shah, R., Narayanan, P.: Unsupervised image style embeddings for retrieval and recognition tasks. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3281-3289 (2020) +5. Garcia, N., Vogiatzis, G.: How to read paintings: semantic art understanding with multi-modal retrieval. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018) +6. Gatys, L.A., Ecker, A.S., Bethge, M.: A neural algorithm of artistic style. arXiv preprint arXiv:1508.06576 (2015) +7. Gatys, L.A., Ecker, A.S., Bethge, M.: Image style transfer using convolutional neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2414-2423 (2016) + +18. Geiping, J., Goldblum, M., Somepalli, G., Shwartz-Ziv, R., Goldstein, T., Wilson, A.G.: How much data are augmentations worth? an investigation into scaling laws, invariance, and implicit regularization. arXiv preprint arXiv:2210.06441 (2022) +19. Gibson, J.J.: The senses considered as perceptual systems. (1966) +20. Graham, D.J., Hughes, J.M., Leder, H., Rockmore, D.N.: Statistics, vision, and the analysis of artistic style. Wiley Interdisciplinary Reviews: Computational Statistics 4(2), 115-123 (2012) +21. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) +22. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501-1510 (2017) +23. Huang, X., Liu, M.Y., Belongie, S., Kautz, J.: Multimodal unsupervised image-to-image translation. In: Proceedings of the European conference on computer vision (ECCV). pp. 172-189 (2018) +24. Hughes, J.M., Graham, D.J., Jacobsen, C.R., Rockmore, D.N.: Comparing higher-order spatial statistics and perceptual judgements in the stylometric analysis of art. In: 2011 19th European Signal Processing Conference. pp. 1244-1248. IEEE (2011) +25. Hughes, J.M., Graham, D.J., Rockmore, D.N.: Stylometrics of artwork: uses and limitations. In: Computer Vision and Image Analysis of Art. vol. 7531, pp. 91-105. SPIE (2010) +26. Jiang, Q.Y., He, Y., Li, G., Lin, J., Li, L., Li, W.J.: Svd: A large-scale short video dataset for near-duplicate video retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5281-5289 (2019) +27. Joshi, A., Agrawal, A., Nair, S.: Art style classification with self-trained ensemble of autoencoding transformations. arXiv preprint arXiv:2012.03377 (2020) +28. Karayev, S., Trentacoste, M., Han, H., Agarwala, A., Darrell, T., Hertzmann, A., Winnemoeller, H.: Recognizing image style. arXiv preprint arXiv:1311.3715 (2013) +29. Lawrence-Lightfoot, S., Davis, J.H.: The art and science of portraiture. John Wiley & Sons (2002) +30. Lecoutre, A., Negrevergne, B., Yger, F.: Recognizing art style automatically in painting with deep learning. In: Asian conference on machine learning. pp. 327-342. PMLR (2017) +31. Lee, S., Kim, D., Han, B.: Cosmo: Content-style modulation for image retrieval with text feedback. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 802-812 (June 2021) +32. Li, J., Yao, L., Hendriks, E., Wang, J.Z.: Rhythmic brushstrokes distinguish van gogh from his contemporaries: findings via automated brushstroke extraction. IEEE transactions on pattern analysis and machine intelligence 34(6), 1159-1176 (2011) +33. Lu, X., Lin, Z., Shen, X., Mech, R., Wang, J.Z.: Deep multi-patch aggregation network for image style, aesthetics, and quality estimation. In: Proceedings of the IEEE international conference on computer vision. pp. 990-998 (2015) +34. Luan, F., Paris, S., Shechtman, E., Bala, K.: Deep photo style transfer. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4990-4998 (2017) +35. Lun, Z., Kalogerakis, E., Sheffer, A.: Elements of style: learning perceptual shape style similarity. ACM Transactions on graphics (TOG) 34(4), 1-14 (2015) + +36. Matsuo, S., Yanai, K.: Cnn-based style vector for style image retrieval. In: Proceedings of the 2016 ACM on International Conference on Multimedia Retrieval. pp. 309-312 (2016) +37. Matthews, R.A., Merriam, T.V.: Distinguishing literary styles using neural networks. In: Handbook of neural computation, pp. G8-1. CRC Press (2020) +38. Menis-Mastromichalakis, O., Sofou, N., Stamou, G.: Deep ensemble art style recognition. In: 2020 International Joint Conference on Neural Networks (IJCNN). pp. 1-8. IEEE (2020) +39. Midjourney: Midjourney (nd), https://www.midjourney.com/home +40. Mikolov, T., Sutskever, I., Chen, K., Corrado, G.S., Dean, J.: Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems 26 (2013) +41. Park, T., Zhu, J.Y., Wang, O., Lu, J., Shechtman, E., Efros, A., Zhang, R.: Swapping autoencoder for deep image manipulation. Advances in Neural Information Processing Systems 33, 7198-7211 (2020) +42. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1406-1415 (2019) +43. Pernias, P., Rampas, D., Richter, M.L., Pal, C., Aubreville, M.: Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In: The Twelfth International Conference on Learning Representations (2023) +44. pharmapsychotic: Clip interrogator. https://github.com/pharmapsychotic/clip-interrogator (2023) +45. Pizzi, E., Roy, S.D., Ravindra, S.N., Goyal, P., Douze, M.: A self-supervised descriptor for image copy detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14532-14542 (2022) +46. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +47. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021) +48. Rodriguez, C.S., Lech, M., Pirogova, E.: Classification of style in fine-art paintings using transfer learning and weighted image patches. In: 2018 12th International Conference on Signal Processing and Communication Systems (ICSPCS). pp. 1-7. IEEE (2018) +49. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022) +50. Ruta, D., Motiian, S., Faieta, B., Lin, Z., Jin, H., Filipkowski, A., Gilbert, A., Collomosse, J.: Aladin: all layer adaptive instance normalization for fine-grained style similarity. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11926-11935 (2021) +51. Sablatnig, R., Kammerer, P., Zolda, E.: Hierarchical classification of paintings using face-and brush stroke models. In: Proceedings. Fourteenth International Conference on Pattern Recognition (Cat. No. 98EX170). vol. 1, pp. 172-174. IEEE (1998) +52. Saleh, B., Elgammal, A.: Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855 (2015) + +53. Sandoval, C., Pirogova, E., Lech, M.: Two-stage deep learning approach to the classification of fine-art paintings. IEEE Access 7, 41770-41781 (2019) +54. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402 (2022) +55. Silva, J.M., Pratas, D., Antunes, R., Matos, S., Pinho, A.J.: Automatic analysis of artistic paintings using information-based measures. Pattern Recognition 114, 107864 (2021) +56. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) +57. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Diffusion art or digital forgery? investigating data replication in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) +58. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Understanding and mitigating copying in diffusion models. Advances in Neural Information Processing Systems 36, 47783-47803 (2023) +59. Srinivasa Desikan, B., Shimao, H., Miton, H.: Wikiartvectors: style and color representations of artworks for cultural analysis via information theoretic measures. Entropy 24(9), 1175 (2022) +60. Tenenbaum, J., Freeman, W.: Separating style and content. Advances in neural information processing systems 9 (1996) +61. Walmer, M., Suri, S., Gupta, K., Shrivastava, A.: Teaching matters: Investigating the role of supervision in vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) +62. Wang, J., Yang, H., Fu, J., Yamasaki, T., Guo, B.: Fine-grained image style transfer with visual transformers. In: Proceedings of the Asian Conference on Computer Vision. pp. 841-857 (2022) +63. Wang, S.Y., Efros, A.A., Zhu, J.Y., Zhang, R.: Evaluating data attribution for text-to-image models. arXiv preprint arXiv:2306.09345 (2023) +64. Wilber, M.J., Fang, C., Jin, H., Hertzmann, A., Collomosse, J., Belongie, S.: Bam! the behave artistic media dataset for recognition beyond photography. In: Proceedings of the IEEE international conference on computer vision. pp. 1202-1211 (2017) +65. Willats, J., Durand, F.: Defining pictorial style: Lessons from linguistics and computer graphics. Axiomathes 15(3), 319-351 (2005) +66. Wynen, D., Schmid, C., Mairal, J.: Unsupervised learning of artistic styles with archetypal style analysis. Advances in Neural Information Processing Systems 31 (2018) +67. Yao, L., Li, J., Wang, J.Z.: Characterizing elegance of curves computationally for distinguishing morisseau paintings and the imitations. In: 2009 16th IEEE International Conference on Image Processing (ICIP). pp. 73-76. IEEE (2009) +68. Zhang, W., Cao, C., Chen, S., Liu, J., Tang, X.: Style transfer via image component analysis. IEEE Transactions on multimedia 15(7), 1594-1601 (2013) \ No newline at end of file diff --git a/2024/Investigating Style Similarity in Diffusion Models/images.zip b/2024/Investigating Style Similarity in Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1760c3b11484d64ea938e023a361b99ad2129a48 --- /dev/null +++ b/2024/Investigating Style Similarity in Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f15ec52366789a76ff751b90864a83c769242e7185bbc97d82c59f0f3798c28 +size 571201 diff --git a/2024/Investigating Style Similarity in Diffusion Models/layout.json b/2024/Investigating Style Similarity in Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fc782307e3e0b08040f4b52ebdc155dc137c0cf9 --- /dev/null +++ b/2024/Investigating Style Similarity in Diffusion Models/layout.json @@ -0,0 +1,8734 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 112, + 479, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 112, + 479, + 129 + ], + "spans": [ + { + "bbox": [ + 134, + 112, + 479, + 129 + ], + "type": "text", + "content": "Investigating Style Similarity in Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 150, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 479, + 175 + ], + "type": "text", + "content": "Growthami Somepalli\\*1, Anubhav Gupta\\*1, Kamal Gupta\\*1, Shramay Palta\\*1, Micah Goldblum\\*3, Jonas Geiping\\*2, Abhinav Shrivastava\\*1, and Tom Goldstein\\*1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 209, + 183, + 403, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 183, + 403, + 196 + ], + "spans": [ + { + "bbox": [ + 209, + 183, + 403, + 196 + ], + "type": "text", + "content": "1 University of Maryland, College Park" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 209, + 196, + 402, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 196, + 402, + 206 + ], + "spans": [ + { + "bbox": [ + 209, + 196, + 402, + 206 + ], + "type": "text", + "content": "2 ELLIS Institute, MPI for Intelligent Systems" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 260, + 206, + 353, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 206, + 353, + 217 + ], + "spans": [ + { + "bbox": [ + 260, + 206, + 353, + 217 + ], + "type": "text", + "content": "3 Columbia University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 230, + 217, + 383, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 217, + 383, + 228 + ], + "spans": [ + { + "bbox": [ + 230, + 217, + 383, + 228 + ], + "type": "text", + "content": "4 https://somepago.github.io/csd" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 136, + 270, + 201, + 304 + ], + "blocks": [ + { + "bbox": [ + 148, + 262, + 184, + 269 + ], + "lines": [ + { + "bbox": [ + 148, + 262, + 184, + 269 + ], + "spans": [ + { + "bbox": [ + 148, + 262, + 184, + 269 + ], + "type": "text", + "content": "Real paintings" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 136, + 270, + 201, + 304 + ], + "lines": [ + { + "bbox": [ + 136, + 270, + 201, + 304 + ], + "spans": [ + { + "bbox": [ + 136, + 270, + 201, + 304 + ], + "type": "image", + "image_path": "33ccc3a732cd6c20c46d042c462473ee361023e66ad35bd3cfacabe4ce8c2a3c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 205, + 270, + 253, + 304 + ], + "blocks": [ + { + "bbox": [ + 204, + 262, + 235, + 268 + ], + "lines": [ + { + "bbox": [ + 204, + 262, + 235, + 268 + ], + "spans": [ + { + "bbox": [ + 204, + 262, + 235, + 268 + ], + "type": "text", + "content": "MidJourney" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 205, + 270, + 253, + 304 + ], + "lines": [ + { + "bbox": [ + 205, + 270, + 253, + 304 + ], + "spans": [ + { + "bbox": [ + 205, + 270, + 253, + 304 + ], + "type": "image", + "image_path": "97a368689dba65163f359a2ddd324c8f8366910869cc620896f9acb975979c3e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 255, + 270, + 288, + 304 + ], + "blocks": [ + { + "bbox": [ + 253, + 262, + 277, + 268 + ], + "lines": [ + { + "bbox": [ + 253, + 262, + 277, + 268 + ], + "spans": [ + { + "bbox": [ + 253, + 262, + 277, + 268 + ], + "type": "text", + "content": "2.1" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 255, + 270, + 288, + 304 + ], + "lines": [ + { + "bbox": [ + 255, + 270, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 255, + 270, + 288, + 304 + ], + "type": "image", + "image_path": "58b9eb52ab258da014fbd42d2055b40dff2cf6ac31cce68fd30143b1ed26c538.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 289, + 270, + 306, + 304 + ], + "blocks": [ + { + "bbox": [ + 289, + 270, + 306, + 304 + ], + "lines": [ + { + "bbox": [ + 289, + 270, + 306, + 304 + ], + "spans": [ + { + "bbox": [ + 289, + 270, + 306, + 304 + ], + "type": "image", + "image_path": "1d858238181bbb2985a5bd8d48cb438624b1201bb35627e8d17040bc094cc814.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 262, + 360, + 304 + ], + "blocks": [ + { + "bbox": [ + 289, + 262, + 307, + 268 + ], + "lines": [ + { + "bbox": [ + 289, + 262, + 307, + 268 + ], + "spans": [ + { + "bbox": [ + 289, + 262, + 307, + 268 + ], + "type": "text", + "content": "1.4" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 262, + 360, + 304 + ], + "lines": [ + { + "bbox": [ + 307, + 262, + 360, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 360, + 304 + ], + "type": "image", + "image_path": "cd738bf053c4382f2f80e478bdbe15cabffb239c63dcf22560a357ad1ed31068.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 361, + 262, + 406, + 304 + ], + "blocks": [ + { + "bbox": [ + 361, + 262, + 406, + 304 + ], + "lines": [ + { + "bbox": [ + 361, + 262, + 406, + 304 + ], + "spans": [ + { + "bbox": [ + 361, + 262, + 406, + 304 + ], + "type": "image", + "image_path": "39763cba641acaee54e33b5e59d0525927c4d63ae7acd1e59378a95926d2bbd8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 406, + 262, + 443, + 304 + ], + "blocks": [ + { + "bbox": [ + 406, + 262, + 443, + 304 + ], + "lines": [ + { + "bbox": [ + 406, + 262, + 443, + 304 + ], + "spans": [ + { + "bbox": [ + 406, + 262, + 443, + 304 + ], + "type": "image", + "image_path": "67adbfc7450fdcb80e7c411bd2eebad2543abefc10dbd3ef913e8d7c999351cc.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 443, + 262, + 479, + 304 + ], + "blocks": [ + { + "bbox": [ + 443, + 262, + 479, + 304 + ], + "lines": [ + { + "bbox": [ + 443, + 262, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 443, + 262, + 479, + 304 + ], + "type": "image", + "image_path": "291721b4c89ab6c5a699ee8eb49f34467111c717a3fe898255e9760867c157a0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 136, + 307, + 202, + 342 + ], + "blocks": [ + { + "bbox": [ + 136, + 307, + 202, + 342 + ], + "lines": [ + { + "bbox": [ + 136, + 307, + 202, + 342 + ], + "spans": [ + { + "bbox": [ + 136, + 307, + 202, + 342 + ], + "type": "image", + "image_path": "87d5451b45700ba592927bd1134d0bf7a0cc9530c5f6bcf85cc98371b354d413.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 204, + 307, + 253, + 342 + ], + "blocks": [ + { + "bbox": [ + 204, + 307, + 253, + 342 + ], + "lines": [ + { + "bbox": [ + 204, + 307, + 253, + 342 + ], + "spans": [ + { + "bbox": [ + 204, + 307, + 253, + 342 + ], + "type": "image", + "image_path": "779265e8f59b906ad8956b13068863ba4e4e0414a34dc96abaa465bb44c5ed61.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 253, + 307, + 305, + 342 + ], + "blocks": [ + { + "bbox": [ + 253, + 307, + 305, + 342 + ], + "lines": [ + { + "bbox": [ + 253, + 307, + 305, + 342 + ], + "spans": [ + { + "bbox": [ + 253, + 307, + 305, + 342 + ], + "type": "image", + "image_path": "8c19ca291a8fe7289f6c4e1fc3cdcec79a4559d92144cdb71796b4e3ab6b4799.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 307, + 307, + 373, + 342 + ], + "blocks": [ + { + "bbox": [ + 307, + 307, + 373, + 342 + ], + "lines": [ + { + "bbox": [ + 307, + 307, + 373, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 307, + 373, + 342 + ], + "type": "image", + "image_path": "54b65c24568733d08bf36729c68d5f90abd786d69ca35bbc2b409770a39bf80c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 374, + 307, + 441, + 342 + ], + "blocks": [ + { + "bbox": [ + 374, + 307, + 441, + 342 + ], + "lines": [ + { + "bbox": [ + 374, + 307, + 441, + 342 + ], + "spans": [ + { + "bbox": [ + 374, + 307, + 441, + 342 + ], + "type": "image", + "image_path": "9ef29bb886a569ef72659c0c3b6f742ac04159e40350b6605e009cdb8270d01d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 443, + 307, + 477, + 342 + ], + "blocks": [ + { + "bbox": [ + 443, + 307, + 477, + 342 + ], + "lines": [ + { + "bbox": [ + 443, + 307, + 477, + 342 + ], + "spans": [ + { + "bbox": [ + 443, + 307, + 477, + 342 + ], + "type": "image", + "image_path": "c3811ab33825d3a5417f1c3c934e3a2c27223fd66be2625135dd0a39af80bc48.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 136, + 345, + 202, + 380 + ], + "blocks": [ + { + "bbox": [ + 136, + 345, + 202, + 380 + ], + "lines": [ + { + "bbox": [ + 136, + 345, + 202, + 380 + ], + "spans": [ + { + "bbox": [ + 136, + 345, + 202, + 380 + ], + "type": "image", + "image_path": "f90092ba02fc4179b6e64c9c5d8d9da4eb9a67c6864641c13e4a124b24e2c59f.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 390, + 482, + 467 + ], + "lines": [ + { + "bbox": [ + 131, + 390, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 131, + 390, + 482, + 467 + ], + "type": "text", + "content": "Fig. 1: Original artwork of 6 popular artists and the images generated in the style of these artists by three popular text-to-image generative models. The numbers displayed below each image indicate the similarity of the generated image with the artist's style using the proposed method. A high similarity score suggests a strong presence of the artist's style elements in the generated image. Based on our analyses, we postulate that three artists on the right were removed (or unlearned) from SD 2.1 while they were present in MidJourney and SD 1.4. Please refer to Section 2 for more details." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 204, + 345, + 253, + 380 + ], + "blocks": [ + { + "bbox": [ + 204, + 345, + 253, + 380 + ], + "lines": [ + { + "bbox": [ + 204, + 345, + 253, + 380 + ], + "spans": [ + { + "bbox": [ + 204, + 345, + 253, + 380 + ], + "type": "image", + "image_path": "b8b2038b49eab9cba87f72bad42d84fd78c45ce931333a3253a4eaca435862a7.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 253, + 345, + 305, + 380 + ], + "blocks": [ + { + "bbox": [ + 253, + 345, + 305, + 380 + ], + "lines": [ + { + "bbox": [ + 253, + 345, + 305, + 380 + ], + "spans": [ + { + "bbox": [ + 253, + 345, + 305, + 380 + ], + "type": "image", + "image_path": "f6e27c024edf7205ff7698356d0c36f38defa1885e25c7f5616aaffdc91817b4.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 307, + 345, + 373, + 380 + ], + "blocks": [ + { + "bbox": [ + 307, + 345, + 373, + 380 + ], + "lines": [ + { + "bbox": [ + 307, + 345, + 373, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 345, + 373, + 380 + ], + "type": "image", + "image_path": "96abf52c86408e3929f388ef765e9007bdad54267970aeee97e50cbfc87415af.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 374, + 345, + 425, + 380 + ], + "blocks": [ + { + "bbox": [ + 374, + 345, + 425, + 380 + ], + "lines": [ + { + "bbox": [ + 374, + 345, + 425, + 380 + ], + "spans": [ + { + "bbox": [ + 374, + 345, + 425, + 380 + ], + "type": "image", + "image_path": "2e29ad371079ba05a0e8ae8b2b2e4f623cd447a0ac9c3840b1a5fd4bb9b9cdaa.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 428, + 345, + 477, + 380 + ], + "blocks": [ + { + "bbox": [ + 428, + 345, + 477, + 380 + ], + "lines": [ + { + "bbox": [ + 428, + 345, + 477, + 380 + ], + "spans": [ + { + "bbox": [ + 428, + 345, + 477, + 380 + ], + "type": "image", + "image_path": "eaceb94b33dbc00f9f421cea387eee0934b0a4c3c5d6fc85d27a3afa1d5ed335.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 159, + 514, + 453, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 514, + 453, + 635 + ], + "spans": [ + { + "bbox": [ + 159, + 514, + 453, + 635 + ], + "type": "text", + "content": "Abstract. Generative models are now widely used by graphic designers and artists. Prior works have shown that these models remember and often replicate content from their training data during generation. Hence as their proliferation increases, it has become important to perform a database search to determine whether the properties of the image are attributable to specific training data, every time before a generated image is used for professional purposes. Existing tools for this purpose focus on retrieving images of similar semantic content. Meanwhile, many artists are concerned with style replication in text-to-image models. We present a framework for understanding and extracting style descriptors from images. Our framework comprises a new dataset curated using the" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 654, + 378, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 654, + 378, + 665 + ], + "spans": [ + { + "bbox": [ + 142, + 654, + 378, + 665 + ], + "type": "text", + "content": "*Equal contribution. Correspondence: growthami@umd.edu." + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 205 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 205 + ], + "type": "text", + "content": "insight that style is a subjective property of an image that captures complex yet meaningful interactions of factors including but not limited to colors, textures, shapes, etc. We also propose a method to extract style descriptors that can be used to attribute style of a generated image to the images used in the training dataset of a text-to-image model. We showcase promising results in various style retrieval tasks. We also quantitatively and qualitatively analyze style attribution and matching in the Stable Diffusion model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 216, + 419, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 216, + 419, + 228 + ], + "spans": [ + { + "bbox": [ + 160, + 216, + 419, + 228 + ], + "type": "text", + "content": "Keywords: Image Style " + }, + { + "bbox": [ + 160, + 216, + 419, + 228 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 216, + 419, + 228 + ], + "type": "text", + "content": " Style Similarity " + }, + { + "bbox": [ + 160, + 216, + 419, + 228 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 216, + 419, + 228 + ], + "type": "text", + "content": " Generative models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 251, + 230, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 230, + 264 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 230, + 264 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 280, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 411 + ], + "type": "text", + "content": "Diffusion-based image generators like Stable Diffusion [49], DALL-E [47] and many others [1, 7, 39, 43] learn artistic styles from massive captioned image datasets that are scraped from across the web [54]. Before a generated image is used for commercial purposes, it is wise to understand its relationship to the training data and the origins of its design elements and style attributes. Discovering and attributing these generated images, typically done with image similarity search, is hence becoming increasingly important. Such dataset attribution serves two purposes. It enables users of generated images to understand potential conflicts, associations, and social connotations that their image may evoke. It also enables artists to assess whether and how generative models are using elements of their work." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 413, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 592 + ], + "type": "text", + "content": "Despite a long history of research [60], recovering style from an image is a challenging and open problem in Computer Vision. Many existing retrieval methods [8, 45, 46] for large training datasets focus primarily on matching semantic content between a pair of images. Understanding the origin of the style present in a generated image, however, is much less well understood. To address this gap, we propose a self-supervised objective for learning style descriptors from images. Standard augmentation-based SSL pipelines (e.g. SimCLR and variants) learn feature vectors that are invariant to a set of augmentations. Typically, these augmentations preserve semantic content and treat style as a nuisance variable. In contrast, we choose augmentations that preserve stylistic attributes (colors, textures, or shapes) while minimizing content. Unfortunately, SSL is not enough, as style is inherently subjective, and therefore a good style extractor should be aligned with human perceptions and definitions of style. For this reason, we curate a style attribution dataset, ContraStyles, in which images are associated with the artist that created them." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "By training with both SSL and supervised objectives, we create a high-performance model for representing style. Our model, CSD, outperforms other large-scale pre-trained models and prior style retrieval methods on standard datasets. Using CSD, we examine the extent of style replication in the popular open-source text-to-image generative model Stable Diffusion [49], and consider different factors that impact the rate of style replication." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "text", + "content": "G. Somepalli et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "To summarize our contributions, we (1) propose a style attribution dataset ContraStyles, associating images with their styles, (2) introduce a multi-label contrastive learning scheme to extract style descriptors from images and show the efficacy of the scheme by zero-shot evaluation on public domain datasets such as WikiArt and DomainNet (3) We perform a style attribution case study for one of the most popular text-to-image generative models, Stable Diffusion, and propose indicators of how likely an artist's style is to be replicated. Code, artifacts, and dataset links are available at https://somepago.github.io/csd." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 229, + 220, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 229, + 220, + 242 + ], + "spans": [ + { + "bbox": [ + 131, + 229, + 220, + 242 + ], + "type": "text", + "content": "2 Motivation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 252, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 252, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 482, + 479 + ], + "type": "text", + "content": "We present a case study that shows how style features can be used to interrogate a generative model, and provide utility to either artists or users. We consider the task of analyzing a model's ability to emulate an artist's style, and of attributing the style of an image to an artist. We begin by curating a list of 96 artists, primarily sourced from the WikiArt database, supplemented by a few contemporary artists who are notably popular within the Stable Diffusion community5. For each artist, we compute a prototype vector by averaging the embeddings of their paintings using our proposed feature extractor, CSD ViT-L. Next, we generate an image for each artist using Stable Diffusion 2.1 with a prompt in the format A painting in the style of . We compute the dot product similarity between each generated image's embedding and the artist's prototype. This process was repeated multiple times for each artist, and we plot mean results in Fig. 2. We refer to this quantity as the General Style Similarity (GSS) score for an artist, as it measures how similar a generated image is to a typical image from that artist while using our style representation model. We also plot an analogous style similarity score, but using \"content-constrained\" prompts. For instance, one prompt template is A painting of a woman doing style of where X is the name of the artist and Y is some setting like reading a book or holding a baby etc. See Sec. 7 for all templates." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 480, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 482, + 624 + ], + "type": "text", + "content": "Each point in Fig. 2 represents an artist. Notice that GSS scores are highly correlated with content-constrained scores, indicating that our feature vectors represent style more than semantic content. Our findings reveal that SD 2.1 is much more capable of emulating some artists than others. Artists like Leonid Afremov, Georges Seurat exhibit high style similarity scores, and visual inspection of generated images confirms that indeed their style is emulated by the model (Fig. 1 - Original artwork vs SD 2.1). On the other end of the spectrum, artists such as Ruan Jia and Greg Rutkowski showed low similarity scores, and likewise the generated images bear little resemblance to the artists' work. Interestingly, after completing this study, we discovered that Greg Rutkowski's work was excluded from the training data for the Stable Diffusion 2.1 model, as reported by [11]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "type": "text", + "content": "This demonstrates that the Style Similarity score can be used by artists to quantify how well a model emulates their style, or it can be used by users to" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "type": "text", + "content": "Investigating Style Similarity in Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 405, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 405, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 405, + 665 + ], + "type": "text", + "content": "5 https://supagruen.github.io/StableDiffusion-CheatSheet/" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "ascertain whether a generated image contains stylistic elements associated with a particular artist. After a thorough inspection of the generations from 96 artists, we hypothesize that a single-image Style Similarity score below 0.5 indicates the absence of the artist's style, whereas a score above 0.8 strongly indicates its presence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 177, + 299, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 299, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 299, + 440 + ], + "type": "text", + "content": "In Figure 1, we show original artworks for 6 artists, and generations from MidJourney [39], Stable Diffusion 2.1 and Stable Diffusion 1.4 [49] for each of these artists. The 3 artists on the left side have high GSS while the ones on the right side have low GSS. Below each generated image, we display the similarity against the artist's prototype vector. We see high image similarity scores in the Midjourney generations and qualitatively these images look stylistically similar to artists' original artworks. We also see the interesting cases of Greg Rutkowski, Ruan Jia, and Amano whose style is captured by Stable Diffusion 1.4, while being notably absent in Stable Diffusion 2.1. This finding is in line with reports suggesting that some of these artists were removed from the training data" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 440, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 489 + ], + "type": "text", + "content": "of Stable Diffusion 2.1 [11]. Based on this analysis, we postulate that Ruan Jia, Wadim Kashin, Anton Fadeev, Justin Gerard, and Amano were also either excluded from the training data or post-hoc unlearned/removed from Stable Diffusion 2.1." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 204, + 474, + 340 + ], + "blocks": [ + { + "bbox": [ + 310, + 204, + 474, + 340 + ], + "lines": [ + { + "bbox": [ + 310, + 204, + 474, + 340 + ], + "spans": [ + { + "bbox": [ + 310, + 204, + 474, + 340 + ], + "type": "image", + "image_path": "d1f6e6683b8888ef099762a2ff03471432b73f50a7d1ebc04caf8b92ae5584d5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 350, + 482, + 438 + ], + "lines": [ + { + "bbox": [ + 304, + 350, + 482, + 438 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 482, + 438 + ], + "type": "text", + "content": "Fig. 2: Style similarity of Stable Diffusion 2.1 generated images against the artist's prototypical representation. On the X-axis, the similarities are depicted when the prompt is not constrained, while the Y-axis represents similarity when the prompt is constrained to generate an image of a \"woman\" in the artist's style." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 512, + 241, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 241, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 241, + 528 + ], + "type": "text", + "content": "3 What is style?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "text", + "content": "The precise definition of \"style\" remains in contention, but many named artistic styles (e.g., cubism, impressionism, etc...) are often associated with certain artists. We leverage this social construct, and define style simply as the collection of global characteristics of an image that are identified with an artist or artistic movement. These characteristics encompass various elements such as color usage, brushstroke techniques, composition, and perspective." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Related work. Early computer vision algorithms attempted to model style using low-level visual features like color histograms, texture patterns, edge detection, and shape descriptors. Other computational techniques involve rule-based systems, such as the presence of specific compositional elements, the use of spe" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "text", + "content": "G. Somepalli et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "cific color palettes, or the presence of certain brushstroke patterns to identify specific style characteristics [19, 20, 24, 25, 29, 32, 35, 37, 51, 55, 59, 65, 67]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 272 + ], + "type": "text", + "content": "Modern studies have focused on the task of transferring style from one image to another [13, 17, 22, 34, 41, 62, 68]. Some works have also concentrated on style classification [2, 4, 10, 15, 27, 28, 30, 33, 38, 48, 53]. A limited number of studies address in-the-wild style quantification, matching, and retrieval [14, 23, 36, 50, 66]. In their seminal work, Gatys et al. [17] introduced Gram Matrices as style descriptors and utilized an optimization loop to transfer style. Another approach proposed by Luan et al. [34] includes a photorealism regularization term to prevent distortions in the reconstructed image. Zhang et al. [68] formulated style transfer using Markov random fields. Beyond Gram-based style representation, Chu et al. [10] explored various other types of correlations and demonstrated performance variations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 272, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 380 + ], + "type": "text", + "content": "In a recent work by Lee et al [31], two separate neural network modules were used - one for image style and another for image content - to facilitate image style retrieval. In the most recent related research, Wang et al. [63] developed an attribution model trained on synthetic style pairs, designed to identify stylistically similar images. In contrast to this approach, our method leverages real image pairs, curated automatically through their caption annotations. Despite our training dataset being approx. " + }, + { + "bbox": [ + 130, + 272, + 482, + 380 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 130, + 272, + 482, + 380 + ], + "type": "text", + "content": " the size of training data used in [63], we demonstrate that our model significantly outperforms this method on many zero-shot style matching tasks in the later sections." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 398, + 385, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 398, + 385, + 412 + ], + "spans": [ + { + "bbox": [ + 131, + 398, + 385, + 412 + ], + "type": "text", + "content": "4 Creating a dataset for style attribution" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 423, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 508 + ], + "type": "text", + "content": "While many large web datasets now exist, we need one that contains wide variations in artistic styles, and also labels that be used for downstream style retrieval evaluation. Some large-scale datasets specifically designed to handle such a challenge, like BAM [64], are not available in the public domain and others like WikiArt [52] are not large enough to train a good style feature extractor. In the following section, we propose a way to curate a large style dataset out of the LAION [54] Aesthetics " + }, + { + "bbox": [ + 130, + 423, + 482, + 508 + ], + "type": "inline_equation", + "content": "6+" + }, + { + "bbox": [ + 130, + 423, + 482, + 508 + ], + "type": "text", + "content": " dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": "ContraStyles: A dataset for style distillation. We curate our own dataset as a subset of LAION [54]. We start off with the 12M image-text pairs with predicted aesthetics scores of 6 or higher. Note that this dataset is extremely unbalanced, with some popular artists appearing much more frequently than others. Also, a large number of images are duplicated within the dataset which is a major issue for the text-to-image models trained on this data [57,58]. Furthermore, the image captions within the data are often noisy and are often missing a lot of information. We address these challenges and propose a new subset of LAION-Aesthetics consisting of 511,921 images, and 3840 style tags, where each image can have one or more tags. We use this dataset for training our models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "We begin with a bank of styles collated in previous work for image understanding with the CLIP Interrogator [44]. This bank of styles was curated based on typical user prompts for Stable Diffusion. We combine the bank of artists," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "text", + "content": "Investigating Style Similarity in Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "medium, and movement references, to a shortlist of 5600 tags. We then search for these tags in the 12M LAION-Aesthetics captions and shortlist the images that have at least one of the tags present. We further filter out the tags which have over 100,000 hits in the dataset since human inspection found that they refer to common phrases like 'picture' or 'photograph' that do not invoke a distinct style. After discarding images with an unavailable URL, we are left with about 1 million images and 3840 tags. We further deduplicate the images using SSCD [45] with a threshold of 0.8 and merge the tags of images that are near copies of each other. As a by-product, the dedduplication also helps with the missing tags in the images, since we can simply merge the text labels of duplicate images. After dedduplication, we are left with 511,921 images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 277, + 370, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 370, + 291 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 370, + 291 + ], + "type": "text", + "content": "5 Contrastive Style Descriptors (CSD)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 312, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 419 + ], + "type": "text", + "content": "Self-Supervised Learning. Many successful approaches to SSL [61] use a contrastive [9] approach, where two views (or augmentations) of the same image in the dataset are sampled and passed to a deep network to get their respective image descriptors. The network is trained to maximize the similarity of two views of the same image and minimize agreement with other images in the batch. Standard choices for augmentations include color jitter, blurring, grayscale, etc., as these alter the image's visual properties while preserving content. While these are good augmentations for object recognition tasks, they train the network to ignore image attributes associated with style." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 422, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 469 + ], + "type": "text", + "content": "Our approach relies on a training pipeline with two parts. First, we use contrastive SSL, but with a set of augmentations that are curated to preserve style. Second, we align our model with human perceptions of style by training on our labelled ContraStyles dataset described in Section 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 482, + 481, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 481, + 518 + ], + "type": "text", + "content": "Proposed Approach. We seek a model for extracting image descriptors that contain concise and effective style information. To be a useful, the model should be invariant to semantic content and capable of disentangling multiple styles." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "text", + "content": "Given a dataset of " + }, + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "text", + "content": " labeled images " + }, + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i, l_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "text", + "content": ", where each image can have one or more labels from a set of " + }, + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "text", + "content": " labels, we define the label vector of the " + }, + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "text", + "content": " image as " + }, + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "inline_equation", + "content": "l_i = (c_1, c_2, \\ldots, c_L)" + }, + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "inline_equation", + "content": "c_k \\in \\{0, 1\\}" + }, + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "text", + "content": ". As mentioned in the previous section, our multi-label dataset consists of " + }, + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "inline_equation", + "content": "N = 511,921" + }, + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "text", + "content": " images and " + }, + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "inline_equation", + "content": "L = 3,840" + }, + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "text", + "content": " style tags. We consider a mini-batch of " + }, + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "text", + "content": " images. Each of the images are passed to a Vision Transformer (ViT) [12] backbone, and then projected to a " + }, + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 519, + 482, + 604 + ], + "type": "text", + "content": "-dimensional vector. We consider two variants of ViT (ViT-B and ViT-L)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Our style descriptors " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{ViT}}(\\mathbf{x}_i) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " are then used to create a matrix of pairwise cosine similarity scores " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "s_{i,j} = \\cos (f_{\\mathrm{ViT}}(\\mathbf{x}_i), f_{\\mathrm{ViT}}(\\mathbf{x}_j))" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": ". In order to compute our multi-label contrastive loss (MCL), we also compute the groundtruth similarity scores as " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\hat{s}_{i,j} = \\mathbb{1}\\left(l_i^{\\mathrm{T}}l_j\\right)" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbb{1}" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " is the indicator function that returns 1 if any of the labels of the images " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "i, j" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " match. Our final loss term reduces to:" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "text", + "content": "G. Somepalli et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 231, + 136, + 480, + 170 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 136, + 480, + 170 + ], + "spans": [ + { + "bbox": [ + 231, + 136, + 480, + 170 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M C L}} = - \\hat {s} _ {i, j} \\log \\frac {\\exp \\left(s _ {i , j}\\right) / \\tau)}{\\sum_ {k \\neq j} \\exp \\left(s _ {i , k}\\right) / \\tau}), \\tag {1}", + "image_path": "85c246830737a963b58d834e90ef0258164448a2cce926942c9c9d3f9f865672.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 179, + 366, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 179, + 366, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 179, + 366, + 190 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 179, + 366, + 190 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 132, + 179, + 366, + 190 + ], + "type": "text", + "content": " is the temperature fixed during the training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 191, + 480, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 191, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 480, + 310 + ], + "type": "text", + "content": "Since our supervised dataset is modest in size, we add a self-supervised objective. We sample two \"views\" (augmentations) of each image in a batch and add a contrastive SSL term. Standard SSL training routines (e.g., MoCo, SimCLR, BYOL etc.) choose augmentations so that each pair of views has the same semantic content, but different style content. These augmentations typically include Resize, Horizontal Flips, Color Jitter, Grayscale, Gaussian Blur, and Solarization [6, 18]. For our purposes, we depart from standard methods by excluding photometric augmentations (Gaussian Blur, Color Jitter), as they alter the style of the image. We keep the following spatial augmentations - Horizontal Flips, Vertical Flips, Resize and Rotation as they keep style intact." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 311, + 481, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 481, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 481, + 371 + ], + "type": "text", + "content": "The overall loss function is a simple combination of the multilabel contrastive loss and self-supervised loss " + }, + { + "bbox": [ + 130, + 311, + 481, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{\\mathrm{MCL}} + \\lambda \\mathcal{L}_{\\mathrm{SSL}}" + }, + { + "bbox": [ + 130, + 311, + 481, + 371 + ], + "type": "text", + "content": ". During inference, we use the final layer embedding and the dot product to compute style similarity between any two images. In our experiments, we found that initializing weights to CLIP [46] ViT-B and ViT-L improves performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 392, + 198, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 198, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 198, + 403 + ], + "type": "text", + "content": "6 Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 418, + 480, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 480, + 525 + ], + "type": "text", + "content": "Training details. We present the results for two variants of our model CSD ViT-B and CSD ViT-L version. Both the models are initialized with respective CLIP variant checkpoints and are finetuned for 80k iterations on the ContraStyles dataset on 4 A4000/A5000 GPUs. We use an SGD optimizer with momentum 0.9 and learning rate of 0.003 for the projection layer and " + }, + { + "bbox": [ + 130, + 418, + 480, + 525 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 130, + 418, + 480, + 525 + ], + "type": "text", + "content": " for the backbone. Our mini-batch size per GPU is 16. We use " + }, + { + "bbox": [ + 130, + 418, + 480, + 525 + ], + "type": "inline_equation", + "content": "\\lambda = 0.2" + }, + { + "bbox": [ + 130, + 418, + 480, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 418, + 480, + 525 + ], + "type": "inline_equation", + "content": "\\tau = 0.1" + }, + { + "bbox": [ + 130, + 418, + 480, + 525 + ], + "type": "text", + "content": " for the final model. The training takes about 8 hours for the base model and around 16 hours for the large model. See the Appendix for more details and ablations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 530, + 481, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 530, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 530, + 481, + 578 + ], + "type": "text", + "content": "Task. We perform zero-shot evaluation across multiple datasets on a style-retrieval task. Following [3, 26], we split each dataset into two parts: Database and Query. Given a query image at test time, we evaluate whether we can find the ground-truth style in its nearest neighbors from the database." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Baselines. We compare our model against a recent style attribution model GDA [63] which is trained via fine-tuning on paired synthetic style data, and VGG [17, 56] Gram Matrices which are often used for neural style transfer applications. Further we compare with CLIP [46] models supervised with freeform text captions, and with other self-supervised models such as DINO [8], MoCo [21], SSCD [45]. We use the embeddings from the last layer for each of these models except for VGG where we use the Gram Matrix [17] of the last" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "type": "text", + "content": "Investigating Style Similarity in Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": "layer. We skipped evaluations of [14, 23, 50] since both pre-trained models and training data are not available." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 144, + 481, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 144, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 144, + 481, + 217 + ], + "type": "text", + "content": "Metrics. We do nearest neighbour searches for " + }, + { + "bbox": [ + 130, + 144, + 481, + 217 + ], + "type": "inline_equation", + "content": "k \\in [1, 10, 100]" + }, + { + "bbox": [ + 130, + 144, + 481, + 217 + ], + "type": "text", + "content": ", and report Recall@k, mAP@k. We use the standard definitions of these metrics from the retrieval literature. Like [5], we define positive recall as the existence of a correct label in top-N matches and no recall when none of the top-N matches share a label with the query. Similarly, mAP is defined as average over precision at each rank in N, and then averaged over all queries." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 281, + 479, + 459 + ], + "blocks": [ + { + "bbox": [ + 130, + 238, + 481, + 271 + ], + "lines": [ + { + "bbox": [ + 130, + 238, + 481, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 481, + 271 + ], + "type": "text", + "content": "Table 1: mAP and Recall metrics on DomainNet and WikiArt datasets. Our model consistently performs the best in all cases except one, against both self-supervised and style attribution baselines." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 281, + 479, + 459 + ], + "lines": [ + { + "bbox": [ + 135, + 281, + 479, + 459 + ], + "spans": [ + { + "bbox": [ + 135, + 281, + 479, + 459 + ], + "type": "table", + "html": "
DomainNet (mAP@k)WikiArt (mAP@k)DomainNet (Recall@k)WikiArt (Recall@k)
Method110100110100110110100
VGG Gram [16]---25.919.411.4--25.952.780.4
DINO ViT-B/16 [8]69.468.266.244.033.418.969.493.744.069.488.1
DINO ViT-B/8 [8]72.270.969.346.935.920.472.293.846.971.088.9
SSCD RN-50 [45]67.665.962.036.026.514.867.695.036.062.185.4
MOCO ViT-B/16 [21]71.971.169.644.033.218.872.094.044.069.088.0
CLIP ViT-B/16 [46]73.773.071.352.242.026.073.794.552.278.393.5
GDA CLIP ViT-B [63]62.961.659.325.621.014.162.992.325.656.683.8
GDA DINO ViT-B [63]69.568.166.145.534.619.769.593.445.575.889.0
GDA ViT-B [63]67.165.664.242.632.218.267.193.642.667.687.1
CSD ViT-B (Ours)78.377.576.056.246.128.778.394.356.280.393.6
CLIP ViT-L [46]74.073.572.259.448.831.574.094.859.482.995.1
CSD ViT-L (Ours)78.377.876.564.5653.8235.6578.394.564.5685.7395.58
", + "image_path": "5e901bfdbad898448f5223b66f99aa351615d7cb8f3b70d513da01f726767580.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Evaluation Datasets. DomainNet [42] consists of an almost equal number of images from six different domains: Clipart, Infograph, Painting, Quickdraw, Real, and Sketch. Upon examination, we observed a strong stylistic resemblance between the Quickdraw and Sketch domains, leading us to exclude Quickdraw from our analysis. The dataset's content information was utilized to categorize the images into two main clusters of content classes. This clustering was achieved through the application of word2vec [40]. The images within the smaller cluster were designated as part of the Query set (20,000 images), and images in the bigger cluster to Database (206768 images). And the second dataset we evaluate all the models is, WikiArt [52]. It consists of 80096 fine art images spread across 1119 artists and 27 genres. We randomly split the dataset into 64090 Database and 16006 Query images. We use the artist as a proxy for the style since there is large visual variation within each genre for them to be considered as independent styles. Under this setting, WikiArt is a challenging retrieval task as the chance probability of successful match is just .09% while it is 20% for DomainNet." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "text", + "content": "G. Somepalli et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 293, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 293, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 293, + 128 + ], + "type": "text", + "content": "6.1 Analyses and observations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 147, + 482, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 147, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 147, + 482, + 338 + ], + "type": "text", + "content": "In Table 1, we report the metrics for all the baselines considered and the proposed CSD method using " + }, + { + "bbox": [ + 130, + 147, + 482, + 338 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 147, + 482, + 338 + ], + "type": "text", + "content": " nearest neighbors on 2 datasets - DomainNet and WikiArt. Note that while style and content are better separated in the case of DomainNet, WikiArt consists of more fine-grained styles and has more practical use cases for style retrieval. Loosely speaking, mAP@" + }, + { + "bbox": [ + 130, + 147, + 482, + 338 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 147, + 482, + 338 + ], + "type": "text", + "content": " determines what percentage of the nearest neighbors that are correct predictions, while the recall determines what percentage of queries has a correct match in the top-" + }, + { + "bbox": [ + 130, + 147, + 482, + 338 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 147, + 482, + 338 + ], + "type": "text", + "content": " neighbors. Our model CSD consistently outperformed all the pretrained feature extractors as well as the recent attribution method GDA [63] on both WikiArt and DomainNet evaluations. Note that all models are evaluated in a zero-shot setting. We see the most gains in the WikiArt dataset which is more challenging with chance probability of only " + }, + { + "bbox": [ + 130, + 147, + 482, + 338 + ], + "type": "inline_equation", + "content": "0.09\\%" + }, + { + "bbox": [ + 130, + 147, + 482, + 338 + ], + "type": "text", + "content": ". When we look at mAP@1, which is same as top-1 accuracy, our base model outperforms the next best model by " + }, + { + "bbox": [ + 130, + 147, + 482, + 338 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 147, + 482, + 338 + ], + "type": "text", + "content": " points on WikiArt and " + }, + { + "bbox": [ + 130, + 147, + 482, + 338 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 130, + 147, + 482, + 338 + ], + "type": "text", + "content": " points on DomainNet. Our large model out-performs the closest large competitor by similar margins. Given the complexity of the task, these improvements are non-trivial." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 340, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 413 + ], + "type": "text", + "content": "We attribute the improvements to a couple of factors, (1) The multi-label style contrastive loss on our curated ContraStyles dataset is quite helpful in teaching the model right styles (2) We hypothesize that these SSL models become invariant to styles because the way they were trained, but we are careful to not strip that away in our SSL loss component by carefully curating non-photometric augmentations in training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 415, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 666 + ], + "type": "text", + "content": "Error Analysis. Even though our model outperforms the previous baselines, our top-1 accuracy for the WikiArt style matching task is still at 64.56. We tried to understand if there is a pattern to these errors. For example, our model is consistently getting confused between impressionist painters claude monet, gustave loiseau, and alfred sisley, all of whom painted many landscapes. They depicted natural scenes, including countryside views, rivers, gardens, and coastal vistas. Another example is pablo picasso and georges braque, who are both cubist painters. Given the impracticality of analyzing all 1,119 artists in the dataset, we opted for a macroscopic examination by categorizing errors at the art movement level. This approach is visualized in the heatmap presented in Fig. 3. In the heatmap, we see most of the errors concentrated along the diagonal, indicating that while the model often correctly identifies the art movement, it struggles to pinpoint the exact artist. There are instances of off-diagonal errors where the model incorrectly identifies both the artist and their art movement. For example, many Post Impressionism and Realism paintings are assigned to Impressionism artists. Upon closer examination, it becomes apparent that they closely align in terms of historical timeline and geographical origin, both being from Europe. This analysis indicates the nuanced nature of style detection and retrieval in Computer Vision. It suggests that the upper limit for accuracy in this task might be considerably lower than " + }, + { + "bbox": [ + 130, + 415, + 482, + 666 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 415, + 482, + 666 + ], + "type": "text", + "content": ", even for a typical human evaluator, due to the inherent subtleties and complexities involved." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "text", + "content": "Investigating Style Similarity in Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 358, + 232 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 358, + 232 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 358, + 232 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 358, + 232 + ], + "type": "image", + "image_path": "e9748336a4e095dc167b6706602aea38f9aa6307127c1b62f7a51cf675d7a6c1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 240, + 364, + 285 + ], + "lines": [ + { + "bbox": [ + 131, + 240, + 364, + 285 + ], + "spans": [ + { + "bbox": [ + 131, + 240, + 364, + 285 + ], + "type": "text", + "content": "Fig. 3: Confusion Matrix of errors in WikiArt: Art movements are predicted correctly. Errors occur in cases where movements share the same historical timeline and/or are derived from the same earlier movement." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 369, + 140, + 481, + 210 + ], + "blocks": [ + { + "bbox": [ + 369, + 140, + 481, + 210 + ], + "lines": [ + { + "bbox": [ + 369, + 140, + 481, + 210 + ], + "spans": [ + { + "bbox": [ + 369, + 140, + 481, + 210 + ], + "type": "image", + "image_path": "13e130720ef7a42129e74a87bf69aa76475dc661dd963502fcd7ba0801414e5f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 219, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 366, + 219, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 366, + 219, + 482, + 285 + ], + "type": "text", + "content": "Fig. 4: Human study on Style Retrieval: Turns out untrained humans are worse than many feature extractors on matching images from the same artist." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 311, + 231, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 231, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 231, + 324 + ], + "type": "text", + "content": "6.2 Human Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 338, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 483 + ], + "type": "text", + "content": "To understand how our models compare to untrained humans, we conducted a small survey on style matching on 30 humans (excluding the authors). Following the convention in other papers [14,28,50,62] and this paper, we assume, 2 images from same artist can be considered stylistic matches. For each query image, we gave 4 answer images out of which only one is from the same artist and hence is the right answer, so chance accuracy is " + }, + { + "bbox": [ + 130, + 338, + 482, + 483 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 130, + 338, + 482, + 483 + ], + "type": "text", + "content": ". We used the Artcive dataset introduced in [63] to create this survey and we collected 3 responses per item to break any ties. We present the results in Fig 4. Most interestingly, untrained humans are worse than many feature extractors at this task. SSCD is the only feature extractor that did worse than humans. Our model, CSD outperforms all the baselines on this dataset as well. This underpins the difficulty of style matching and also highlights the superior performance of our feature extractor." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 507, + 479, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 507, + 479, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 507, + 479, + 522 + ], + "type": "text", + "content": "7 Studying style in the wild: Analysis of Stable Diffusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 540, + 482, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 482, + 613 + ], + "type": "text", + "content": "In the previous section, we have quantitatively shown that our model Contrastive Style Descriptors outperforms many baselines on style matching task. Now we try to address the question, Can we do style matching on Stable Diffusion generated images? To answer this question, we first curated multiple synthetic image collections using Stable Diffusion v 2.1 [49] and then compared them against the \"ground truth\" style matches on ContraStyles dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Creating synthetic style dataset. The first challenge in curating synthetic images through prompts is the choice of prompts to be used for the generation. There have been no in-depth quantitative studies of the effect of prompts on generation styles. For this analysis, we chose 3 types of prompts." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "text", + "content": "G. Somepalli et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 139, + 134, + 479, + 216 + ], + "blocks": [ + { + "bbox": [ + 132, + 102, + 482, + 125 + ], + "lines": [ + { + "bbox": [ + 132, + 102, + 482, + 125 + ], + "spans": [ + { + "bbox": [ + 132, + 102, + 482, + 125 + ], + "type": "text", + "content": "Table 2: mAP and Recall of SD 2.1 generated synthetic datasets based on Simple prompts and User-generated prompts" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 139, + 134, + 479, + 216 + ], + "lines": [ + { + "bbox": [ + 139, + 134, + 479, + 216 + ], + "spans": [ + { + "bbox": [ + 139, + 134, + 479, + 216 + ], + "type": "table", + "html": "
Simple prompts (mAP@k)User-generated (mAP@k)Simple prompts (Recall@k)User-generated (Recall@k)
Method110100110100110100110100
GDA - DINO11.610.27.64.454.594.2411.628.152.834.4525.2267.18
CSD-ViTB17.5316.5612.685.855.965.5817.5338.6561.855.8529.2674.2
CLIP ViT-L/1422.320.416.16.15.75.122.344.566.26.126.071.7
CSD (Ours)24.523.318.55.75.95.624.547.267.55.726.571.8
", + "image_path": "68ce7432468125cc5da930f877257e5af71c2bf4989f7206201497189c99246f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 236, + 485, + 428 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 133, + 236, + 481, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 481, + 296 + ], + "type": "text", + "content": "1. User-generated prompts: We used a Stable Diffusion Prompts" + }, + { + "bbox": [ + 133, + 236, + 481, + 296 + ], + "type": "inline_equation", + "content": "^6" + }, + { + "bbox": [ + 133, + 236, + 481, + 296 + ], + "type": "text", + "content": " dataset of 80,000 prompts filtered from Lexica.art. We used the test split and then filtered the prompts to make sure at least one of the keywords from the list we curated in Section 4 is present. We then sampled 4000 prompts from this subset for query split generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 296, + 481, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 296, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 481, + 344 + ], + "type": "text", + "content": "2. Simple prompts: We randomly sampled 400 artists which appeared most frequently in user-generated prompts we analysed. We format the prompt as A painting in the style of , and we generate 10 images per prompt by varying the initialization seed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 344, + 485, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 344, + 485, + 428 + ], + "spans": [ + { + "bbox": [ + 133, + 344, + 485, + 428 + ], + "type": "text", + "content": "3. Content-constrained prompts: We wanted to understand if we can detect style when we constrain the model to generate a particular subject/human in the style of an artist. For this, we used the prompt A painting of a woman in the style of or A painting of a woman reading in the style of etc., a total of 5 variations per subject repeated two times. We experimented with subjects, woman,dog and house in this study. We provide the exact templates in the appendix." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 134, + 487, + 479, + 566 + ], + "blocks": [ + { + "bbox": [ + 132, + 454, + 482, + 477 + ], + "lines": [ + { + "bbox": [ + 132, + 454, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 132, + 454, + 482, + 477 + ], + "type": "text", + "content": "Table 3: mAP and Recall of SD 2.1 generated synthetic datasets based on Content-constrained prompts" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 487, + 479, + 566 + ], + "lines": [ + { + "bbox": [ + 134, + 487, + 479, + 566 + ], + "spans": [ + { + "bbox": [ + 134, + 487, + 479, + 566 + ], + "type": "table", + "html": "
Dog (mAP@k)House (mAP@k)Woman (mAP@k)Dog (Recall@k)House (Recall@k)Woman (Recall@k)
Method110100110100110100110100110100110100
GDA-DINO2.2821.63.93.62.72.22.322.288.6828.733.912.3322.210.128.9
CSD-ViT-B4.54.313.614.64.293.877.557.836.424.514.3634.884.615.03397.5520.142.46
CLIP ViT-L/142.32.21.94.54.23.67.47.16.22.29.829.94.513.835.37.419.041.6
CSD (Ours)4.94.84.26.46.25.410.810.18.64.914.534.56.417.840.610.823.444.2
", + "image_path": "0ed13df6e1a94a1526dfc56c76c1038a27e0261e870a3265f2d342906a09a140.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 587, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 587, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 587, + 482, + 647 + ], + "type": "text", + "content": "We generate 4000 images for each prompt setting using Stable Diffusion 2.1. There is only one style keyword in simple and content-constrained prompts, which we also use as a ground truth label for matching tasks. However, user-generated prompts can have multiple style labels within the caption, and we consider all of them as ground-truth labels." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 448, + 101 + ], + "type": "text", + "content": "Investigating Style Similarity in Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 462, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 462, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 462, + 666 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 133, + 653, + 462, + 666 + ], + "type": "text", + "content": " https://huggingface.co/datasets/Gustavosta/Stable-Diffusion-Prompts" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 304, + 223 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 304, + 223 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 304, + 223 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 304, + 223 + ], + "type": "image", + "image_path": "c7ab3dff347505d675fe37a2fbf2c41836ec36b8fd0730ccbafb668bcd65387c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 231, + 482, + 277 + ], + "lines": [ + { + "bbox": [ + 130, + 231, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 482, + 277 + ], + "type": "text", + "content": "Fig. 5: Nearest \"style\" neighbors. For each generated image (referred to as SD Gen), we show the top 5 style neighbors in CSD using our feature extractor. The green and red box around the image indicates whether or not the artist's name used to generate the SD image was present in the caption of the nearest neighbor." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 481, + 223 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 481, + 223 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 481, + 223 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 481, + 223 + ], + "type": "image", + "image_path": "3fb9c16ec1ce4d7f28c2b7738a291cfa042fe75fda090ae99cd0d4f3452a9f62.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 292, + 479, + 417 + ], + "blocks": [ + { + "bbox": [ + 136, + 292, + 479, + 417 + ], + "lines": [ + { + "bbox": [ + 136, + 292, + 479, + 417 + ], + "spans": [ + { + "bbox": [ + 136, + 292, + 479, + 417 + ], + "type": "image", + "image_path": "46379fab2ec7ba3ad4f5608dc6fe95862b66fe6a4499acfe4df62fdc8e24a554.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 426, + 482, + 482 + ], + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 482 + ], + "type": "text", + "content": "Fig. 6: Top row: Images generated by Stable Diffusion. Middle and Bottom rows: Top matches retrieved by CLIP vs CSD (ours) respectively. CLIP is consistently biased towards image content, for instance retrieving image of a dog in the Column 1, 3, 4, or image of mother and baby in Column 7 or 8. Our method emphasize less on the content but more on the image styles. Please refer to the Appendix for the prompts." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Style retrieval on generated images. In Tab. 2, we show the retrieval results for Simple and User-generated prompts. We also compare our results with the second-best performing model in the previous section, CLIP ViT-L, and a recent style attribution model GDA [63]. We observe that our method outperforms CLIP on Simple prompt dataset. For User-generated prompts, the performance metrics are closer to CLIP model, but it's important to note that these prompts are inherently more complex. This complexity results in a different label distribution in the query set for the two types of prompts we examine, leading to varied metric ranges in each case. Additionally, our method consistently outperforms both baselines in content-constrained scenarios, as evidenced in Table 3. This indicates the robustness and effectiveness of our approach in dealing with a variety of prompt complexities and content specifications. We refer the reader to Appendix to understand a few caveats of this quantitative study." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "text", + "content": "G. Somepalli et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "Qualitative Results. In Fig. 5, we showcase a selection of Stable Diffusion-generated images alongside their top 5 corresponding matches from ContraStyles, as determined by the CSD ViT-L feature extractor. The left section of the figure displays images generated from User-generated prompts, while the right section includes images created from Simple prompts. To aid in visual analysis, matches that share a label with the query image are highlighted in green. We can clearly see that the query image and the matches share multiple stylistic elements, such as color palettes and certain artistic features like motifs or textures. We observed that in generations based on user-generated prompts, perceivable style copying typically occurs only when the prompts are shorter and contain elements that are characteristic of the artist's typical content." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 331 + ], + "type": "text", + "content": "In Figure 6, we present several content-constrained prompt generations and their top-1 matches based on the CLIP ViT-L/14 model versus our CSD model. We observe that the CSD model accurately matches the correct artists to queries even when there is no shared content, only style. This is evident in columns 1, 3, 4, 7, and 8, where our model, CSD matches the correct style elements despite the subjects in the images being quite different. In contrast, the CLIP model still prioritizes content, often leading to mismatches in style." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 151, + 349, + 462, + 434 + ], + "blocks": [ + { + "bbox": [ + 151, + 349, + 462, + 434 + ], + "lines": [ + { + "bbox": [ + 151, + 349, + 462, + 434 + ], + "spans": [ + { + "bbox": [ + 151, + 349, + 462, + 434 + ], + "type": "image", + "image_path": "daa2e23316851fe8e753543045eb3e3d5615abf083f6395f1de25fee444511ff.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 445, + 482, + 511 + ], + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 511 + ], + "type": "text", + "content": "Fig. 7: Does the diffusion model prefer some styles over others? When a prompt contains two style tags, we find that SD 2.1 strongly favors the style that it can best reproduce, we suspect because of a prevalence of the style in training data. In each block, the General Style Similarity(GSS) of the left side artist (red color) is less than the right side artist (blue color). (Ref Fig. 2). The generated image is more biased towards the artist with high GSS score." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Does the model prioritize some artists over others in the prompt? So far in the study we concentrated on the impact of including an artist's name in a style transfer prompt. In this section, we present preliminary findings in scenarios where prompts feature two artists. This is inspired by real-world user prompts from Stable Diffusion and Midjourney, where prompts often include multiple artists. We used the prompt in the style A painting in the style of " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\langle \\mathbf{X} \\rangle" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\langle \\mathbf{Y} \\rangle" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " represent different artists. For this study, we selected five artists with varying General Style Similarity (GSS) scores (referenced in Sec. 2). The artists, ranked by descending GSS scores, are Carne Griffiths, Roy Lichtenstein, Gustav Klimt, Pablo Picasso, and Ivan Bilibin. Note that most of the chosen artists have distinct styles that significantly differ from one another." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "type": "text", + "content": "Investigating Style Similarity in Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 480, + 140 + ], + "type": "text", + "content": "We chose SD-XL Turbo for this analysis because it is trained on deduplicated data, reducing bias towards frequently featured artists in the train set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 482, + 319 + ], + "type": "text", + "content": "The results for each pair of artists are presented in Fig. 7. Interestingly, even without specific instructions to generate a female subject, most outputs depicted women, reflecting the common subject matter of the artists studied. We also calculated the style similarity scores for each generated image, comparing them to the prototypical styles of the artists in the prompts. In most cases, the style of the artist with the higher GSS score dominated the generated image. To test for potential bias towards the artist positioned first in the prompt (X), we conducted two trials with reversed artist positions. The results were generally consistent, with the dominant style remaining unchanged. However, in the case of Pablo Picasso and Gustav Klimt, this pattern did not hold; the model favored Picasso's cubist style over Klimt's nouveau style, possibly due to the small difference in their GSS scores. While this is not an extensive study, a consistent trend emerged: styles of artists with higher GSS scores, like Leonid Afremov and Carne Griffiths, predominantly influenced the combined style. We leave the comprehensive study on this topic to future work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 336, + 219, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 219, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 219, + 350 + ], + "type": "text", + "content": "8 Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 361, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 482, + 563 + ], + "type": "text", + "content": "This study proposes a framework for learning style descriptors from both labeled and unlabeled data. After building a bespoke dataset, ContraStyles, we train a model that achieves state-of-the-art performance on a range of style matching tasks, including DomainNet, WikiArt, and ContraStyles. Then, we show the substantive practical utility of this model through an investigative analysis on the extent of style copying in popular text-to-image generative models. Here, we show the model is capable of determining the factors that contribute to the frequency of style replication. Through cross-referencing of images with style copies and their original prompts, we have discover that the degree of style copying is increasing with prompt complexity. Such complex prompts lead to greater style copying compared to simple one-line prompts. This finding sheds light on the interplay between textual prompts and style transfer, suggesting that prompt design can influence the level of style copying in generative models. Finally, note that the definition of style used in this work is strictly based on artist attribution. We chose this definition because it can be operationalized and used in dataset construction. This definition is certainly not a golden truth, and we look forward to future studies using alternative, or extended definitions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 581, + 246, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 246, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 246, + 596 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 480, + 665 + ], + "type": "text", + "content": "This work was made possible by the ONR MURI program and the AFOSR MURI program. Commercial support was provided by Capital One Bank, the Amazon Research Award program, and Open Philanthropy. Further support was provided by the National Science Foundation (IIS-2212182), and by the NSF TRAILS Institute (2229885)." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "text", + "content": "G. Somepalli et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "text", + "content": "1. Adobe: Firefly (2023), https://www.adobe.com/sensei/generative-ai/firefly.html" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 159, + 480, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 159, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 138, + 159, + 480, + 190 + ], + "type": "text", + "content": "2. Agarwal, S., Karnick, H., Pant, N., Patel, U.: Genre and style based painting classification. In: 2015 IEEE Winter Conference on Applications of Computer Vision. pp. 588-594. IEEE (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 191, + 480, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 191, + 480, + 223 + ], + "spans": [ + { + "bbox": [ + 138, + 191, + 480, + 223 + ], + "type": "text", + "content": "3. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5297-5307 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "type": "text", + "content": "4. Bai, Z., Nakashima, Y., Garcia, N.: Explain me the painting: Multi-topic knowledgeable art description generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5422-5432 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 256, + 480, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 480, + 288 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 480, + 288 + ], + "type": "text", + "content": "5. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1728-1738 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 289, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 480, + 320 + ], + "type": "text", + "content": "6. Balestriero, R., Ibrahim, M., Sobal, V., Morcos, A., Shekhar, S., Goldstein, T., Bordes, F., Bardes, A., Mialon, G., Tian, Y., et al.: A cookbook of self-supervised learning. arXiv preprint arXiv:2304.12210 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 320, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 480, + 352 + ], + "type": "text", + "content": "7. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf 2(3), 8 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 353, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 480, + 385 + ], + "type": "text", + "content": "8. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 385, + 480, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 385, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 138, + 385, + 480, + 428 + ], + "type": "text", + "content": "9. Chopra, S., Hadsell, R., LeCun, Y.: Learning a similarity metric discriminatively, with application to face verification. In: 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05). vol. 1, pp. 539-546. IEEE (2005)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 429, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 480, + 449 + ], + "type": "text", + "content": "0. Chu, W.T., Wu, Y.L.: Image style classification based on learnt deep correlation features. IEEE Transactions on Multimedia 20(9), 2491-2502 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 449, + 480, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 480, + 481 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 480, + 481 + ], + "type": "text", + "content": "1. Decrypt: Greg rutkowski removed from stable diffusion but brought back by ai artists (March 2024), https://decrypt.co/150575/greg-rutkowski-removed-from-stable-diffusion-but-brought-back-by-ai-artists" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 482, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 482, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 482, + 480, + 525 + ], + "type": "text", + "content": "2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 525, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 547 + ], + "type": "text", + "content": "3. Dumoulin, V., Shlens, J., Kudlur, M.: A learned representation for artistic style. arXiv preprint arXiv:1610.07629 (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 547, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 480, + 579 + ], + "type": "text", + "content": "4. Gairola, S., Shah, R., Narayanan, P.: Unsupervised image style embeddings for retrieval and recognition tasks. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3281-3289 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 579, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 480, + 611 + ], + "type": "text", + "content": "5. Garcia, N., Vogiatzis, G.: How to read paintings: semantic art understanding with multi-modal retrieval. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0-0 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "text", + "content": "6. Gatys, L.A., Ecker, A.S., Bethge, M.: A neural algorithm of artistic style. arXiv preprint arXiv:1508.06576 (2015)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "7. Gatys, L.A., Ecker, A.S., Bethge, M.: Image style transfer using convolutional neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2414-2423 (2016)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "text", + "content": "Investigating Style Similarity in Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "18. Geiping, J., Goldblum, M., Somepalli, G., Shwartz-Ziv, R., Goldstein, T., Wilson, A.G.: How much data are augmentations worth? an investigation into scaling laws, invariance, and implicit regularization. arXiv preprint arXiv:2210.06441 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 416, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 416, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 416, + 160 + ], + "type": "text", + "content": "19. Gibson, J.J.: The senses considered as perceptual systems. (1966)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 194 + ], + "type": "text", + "content": "20. Graham, D.J., Hughes, J.M., Leder, H., Rockmore, D.N.: Statistics, vision, and the analysis of artistic style. Wiley Interdisciplinary Reviews: Computational Statistics 4(2), 115-123 (2012)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "21. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "22. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501-1510 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "23. Huang, X., Liu, M.Y., Belongie, S., Kautz, J.: Multimodal unsupervised image-to-image translation. In: Proceedings of the European conference on computer vision (ECCV). pp. 172-189 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "text", + "content": "24. Hughes, J.M., Graham, D.J., Jacobsen, C.R., Rockmore, D.N.: Comparing higher-order spatial statistics and perceptual judgements in the stylometric analysis of art. In: 2011 19th European Signal Processing Conference. pp. 1244-1248. IEEE (2011)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "25. Hughes, J.M., Graham, D.J., Rockmore, D.N.: Stylometrics of artwork: uses and limitations. In: Computer Vision and Image Analysis of Art. vol. 7531, pp. 91-105. SPIE (2010)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "26. Jiang, Q.Y., He, Y., Li, G., Lin, J., Li, L., Li, W.J.: Svd: A large-scale short video dataset for near-duplicate video retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5281-5289 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "27. Joshi, A., Agrawal, A., Nair, S.: Art style classification with self-trained ensemble of autoencoding transformations. arXiv preprint arXiv:2012.03377 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "text", + "content": "28. Karayev, S., Trentacoste, M., Han, H., Agarwala, A., Darrell, T., Hertzmann, A., Winnemoeller, H.: Recognizing image style. arXiv preprint arXiv:1311.3715 (2013)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 446, + 480, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 467 + ], + "type": "text", + "content": "29. Lawrence-Lightfoot, S., Davis, J.H.: The art and science of portraiture. John Wiley & Sons (2002)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "30. Lecoutre, A., Negrevergne, B., Yger, F.: Recognizing art style automatically in painting with deep learning. In: Asian conference on machine learning. pp. 327-342. PMLR (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "31. Lee, S., Kim, D., Han, B.: Cosmo: Content-style modulation for image retrieval with text feedback. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 802-812 (June 2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 577 + ], + "type": "text", + "content": "32. Li, J., Yao, L., Hendriks, E., Wang, J.Z.: Rhythmic brushstrokes distinguish van gogh from his contemporaries: findings via automated brushstroke extraction. IEEE transactions on pattern analysis and machine intelligence 34(6), 1159-1176 (2011)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 577, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 480, + 610 + ], + "type": "text", + "content": "33. Lu, X., Lin, Z., Shen, X., Mech, R., Wang, J.Z.: Deep multi-patch aggregation network for image style, aesthetics, and quality estimation. In: Proceedings of the IEEE international conference on computer vision. pp. 990-998 (2015)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "34. Luan, F., Paris, S., Shechtman, E., Bala, K.: Deep photo style transfer. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4990-4998 (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "35. Lun, Z., Kalogerakis, E., Sheffer, A.: Elements of style: learning perceptual shape style similarity. ACM Transactions on graphics (TOG) 34(4), 1-14 (2015)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "G. Somepalli et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "36. Matsuo, S., Yanai, K.: Cnn-based style vector for style image retrieval. In: Proceedings of the 2016 ACM on International Conference on Multimedia Retrieval. pp. 309-312 (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "text", + "content": "37. Matthews, R.A., Merriam, T.V.: Distinguishing literary styles using neural networks. In: Handbook of neural computation, pp. G8-1. CRC Press (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "38. Menis-Mastromichalakis, O., Sofou, N., Stamou, G.: Deep ensemble art style recognition. In: 2020 International Joint Conference on Neural Networks (IJCNN). pp. 1-8. IEEE (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 422, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 422, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 422, + 215 + ], + "type": "text", + "content": "39. Midjourney: Midjourney (nd), https://www.midjourney.com/home" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "40. Mikolov, T., Sutskever, I., Chen, K., Corrado, G.S., Dean, J.: Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems 26 (2013)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "41. Park, T., Zhu, J.Y., Wang, O., Lu, J., Shechtman, E., Efros, A., Zhang, R.: Swapping autoencoder for deep image manipulation. Advances in Neural Information Processing Systems 33, 7198-7211 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "type": "text", + "content": "42. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1406-1415 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "43. Pernias, P., Rampas, D., Richter, M.L., Pal, C., Aubreville, M.: Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In: The Twelfth International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "text", + "content": "44. pharmapsychotic: Clip interrogator. https://github.com/pharmapsychotic/clip-interrogator (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "45. Pizzi, E., Roy, S.D., Ravindra, S.N., Goyal, P., Douze, M.: A self-supervised descriptor for image copy detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14532-14542 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "type": "text", + "content": "46. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "type": "text", + "content": "47. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 522 + ], + "type": "text", + "content": "48. Rodriguez, C.S., Lech, M., Pirogova, E.: Classification of style in fine-art paintings using transfer learning and weighted image patches. In: 2018 12th International Conference on Signal Processing and Communication Systems (ICSPCS). pp. 1-7. IEEE (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "text", + "content": "49. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 599 + ], + "type": "text", + "content": "50. Ruta, D., Motiian, S., Faieta, B., Lin, Z., Jin, H., Filipkowski, A., Gilbert, A., Collomosse, J.: Aladin: all layer adaptive instance normalization for fine-grained style similarity. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11926-11935 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "text", + "content": "51. Sablatnig, R., Kammerer, P., Zolda, E.: Hierarchical classification of paintings using face-and brush stroke models. In: Proceedings. Fourteenth International Conference on Pattern Recognition (Cat. No. 98EX170). vol. 1, pp. 172-174. IEEE (1998)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "text", + "content": "52. Saleh, B., Elgammal, A.: Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855 (2015)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "text", + "content": "Investigating Style Similarity in Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 599 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "53. Sandoval, C., Pirogova, E., Lech, M.: Two-stage deep learning approach to the classification of fine-art paintings. IEEE Access 7, 41770-41781 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 182 + ], + "type": "text", + "content": "54. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "55. Silva, J.M., Pratas, D., Antunes, R., Matos, S., Pinho, A.J.: Automatic analysis of artistic paintings using information-based measures. Pattern Recognition 114, 107864 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "56. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 270 + ], + "type": "text", + "content": "57. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Diffusion art or digital forgery? investigating data replication in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "text", + "content": "58. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Understanding and mitigating copying in diffusion models. Advances in Neural Information Processing Systems 36, 47783-47803 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "59. Srinivasa Desikan, B., Shimao, H., Miton, H.: Wikiartvectors: style and color representations of artworks for cultural analysis via information theoretic measures. Entropy 24(9), 1175 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "60. Tenenbaum, J., Freeman, W.: Separating style and content. Advances in neural information processing systems 9 (1996)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "text", + "content": "61. Walmer, M., Suri, S., Gupta, K., Shrivastava, A.: Teaching matters: Investigating the role of supervision in vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "type": "text", + "content": "62. Wang, J., Yang, H., Fu, J., Yamasaki, T., Guo, B.: Fine-grained image style transfer with visual transformers. In: Proceedings of the Asian Conference on Computer Vision. pp. 841-857 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "text", + "content": "63. Wang, S.Y., Efros, A.A., Zhu, J.Y., Zhang, R.: Evaluating data attribution for text-to-image models. arXiv preprint arXiv:2306.09345 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 446, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 489 + ], + "type": "text", + "content": "64. Wilber, M.J., Fang, C., Jin, H., Hertzmann, A., Collomosse, J., Belongie, S.: Bam! the behave artistic media dataset for recognition beyond photography. In: Proceedings of the IEEE international conference on computer vision. pp. 1202-1211 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "type": "text", + "content": "65. Willats, J., Durand, F.: Defining pictorial style: Lessons from linguistics and computer graphics. Axiomathes 15(3), 319-351 (2005)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "66. Wynen, D., Schmid, C., Mairal, J.: Unsupervised learning of artistic styles with archetypal style analysis. Advances in Neural Information Processing Systems 31 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "type": "text", + "content": "67. Yao, L., Li, J., Wang, J.Z.: Characterizing elegance of curves computationally for distinguishing morisseau paintings and the imitations. In: 2009 16th IEEE International Conference on Image Processing (ICIP). pp. 73-76. IEEE (2009)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "text", + "content": "68. Zhang, W., Cao, C., Chen, S., Liu, J., Tang, X.: Style transfer via image component analysis. IEEE Transactions on multimedia 15(7), 1594-1601 (2013)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "text", + "content": "G. Somepalli et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Is Retain Set All You Need in Machine Unlearning_ Restoring Performance of Unlearned Models with Out-Of-Distribution Images/ef16a8f0-48f4-4778-ba31-599cb1e6436c_content_list.json b/2024/Is Retain Set All You Need in Machine Unlearning_ Restoring Performance of Unlearned Models with Out-Of-Distribution Images/ef16a8f0-48f4-4778-ba31-599cb1e6436c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..29c144eecc51086df9f144bc5d757bd0a4d20633 --- /dev/null +++ b/2024/Is Retain Set All You Need in Machine Unlearning_ Restoring Performance of Unlearned Models with Out-Of-Distribution Images/ef16a8f0-48f4-4778-ba31-599cb1e6436c_content_list.json @@ -0,0 +1,1679 @@ +[ + { + "type": "text", + "text": "Is Retain Set All You Need in Machine Unlearning? Restoring Performance of Unlearned Models with Out-Of-Distribution Images", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jacopo Bonato1,2, Marco Cotogni2*, and Luigi Sabetta1,2", + "bbox": [ + 272, + 233, + 730, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Leonardo Labs, Rome, Italy,", + "bbox": [ + 375, + 261, + 602, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 These authors contributed equally", + "bbox": [ + 377, + 277, + 622, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jacopo.bonato,luigi.sabetta}.ext@leonardo.com,cotogni@proton.me", + "bbox": [ + 243, + 291, + 754, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this paper, we introduce Selective-distillation for Class and Architecture-agnostic unleaRning (SCAR), a novel approximate unlearning method. SCAR efficiently eliminates specific information while preserving the model's test accuracy without using a retain set, which is a key component in state-of-the-art approximate unlearning algorithms. Our approach utilizes a modified Mahalanobis distance to guide the unlearning of the feature vectors of the instances to be forgotten, aligning them to the nearest wrong class distribution. Moreover, we propose a distillation-trick mechanism that distills the knowledge of the original model into the unlearning model with out-of-distribution images for retaining the original model's test performance without using any retain set. Importantly, we propose a self-forget version of SCAR that unlearns without having access to the forget set. We experimentally verified the effectiveness of our method, on three public datasets, comparing it with state-of-the-art methods. Our method obtains performance higher than methods that operate without the retain set and comparable w.r.t the best methods that rely on the retain set.", + "bbox": [ + 259, + 342, + 743, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Machine Unlearning $\\cdot$ Retain-Free $\\cdot$ Data Removal", + "bbox": [ + 261, + 592, + 686, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 633, + 375, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the past decades, deep learning algorithms have grown significantly in applications like image classification, detection, and segmentation. For optimal performance, these models learn from big training datasets and update their weights to capture intrinsic patterns. Despite their effectiveness, many ethical concerns such as potential biases affecting certain demographics [12] and privacy like unauthorized use of sensitive or personal data [38,48] have been raised. Additionally, methods like model inversion [5,16,29,59,63,66,67], and membership inference attacks [49, 52, 55, 64] raised many concerns about the potential revelation of training data characteristics or can reveal particular data utilization during training. These vulnerabilities highlight the critical necessity for", + "bbox": [ + 212, + 664, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work done while at Leonardo Labs", + "bbox": [ + 217, + 824, + 470, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "implementing safeguards to secure sensitive information within machine learning models. This imperative is fundamental in light of the privacy-preserving regulations that have been put into effect ( e.g. the European Union's General Data Protection Regulation [39] and the California Consumer Privacy Act [45]).", + "bbox": [ + 212, + 146, + 782, + 207 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Machine Unlearning [41, 44, 51, 62] has emerged as a solution to these privacy and ethical challenges, aiming to remove specific information content from algorithms without degrading their performance. Unlearning algorithms can be classified into 3 main groups: model intrinsic, data-driven, and model agnostic. The characteristic of model-intrinsic methods is to change the deep neural network (DNN) architecture or add specific weights to induce the unlearning. The data-driven group is characterized by the usage of data to induce forgetting, and it can be roughly divided into methods that store optimization information during training (e.g. gradient updates) and methods that involve data manipulation. Similarly to this latter group, model-agnostic solutions do not require access to specific optimization-related information and operate directly on the unaltered trained model. Nonetheless, a common practice among these algorithms is to maintain or restore the model performance using the residual portion of the training dataset, known as retain set. Using the retain set to recalibrate models post-unlearning represents a viable strategy, however, it also introduces practical challenges. In scenarios where privacy concerns dictate, only the forget set —images to forget— might be accessible, leaving no retain set for performance restoration. This poses a significant hurdle for models designed to undergo a performance restoration phase after unlearning, potentially compromising their utility. Moreover, when large-scale datasets such as ImageNet or ImageNet-21K are involved and the forget set constitutes a minor fraction, the bulk of the data is the retain set. This disproportionality can significantly inflate the time required to restore the model to its original efficacy. Furthermore, there are even more intricate situations where also the forget set is unavailable, such as when the unlearning task entails forgetting an entire class with access only to its ID.", + "bbox": [ + 212, + 207, + 787, + 585 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper presents a novel, model-agnostic unlearning algorithm, Selective-distillation for Class and Architecture-agnostic unleaRning (SCAR), which leverages metric learning and knowledge distillation [26] to efficiently remove targeted information and maintain model accuracy, notably without relying on a retain set. Specifically, we use the Mahalanobis distance to shift the feature vectors of the instances to forget toward the closest distribution of samples of other classes. This distance includes important information regarding the original dataset distribution of samples resulting in an efficient metric learning-based unlearning strategy. Concurrently, distilling the knowledge of the original model into the unlearning model using out-of-distribution (OOD) images, allows SCAR to maintain the performance of the original model on the test set. The main contributions of this work include: (i) SCAR, a novel model-agnostic unlearning algorithm that achieves competitive unlearning performance without using retain data. (ii) A unique self-forget mechanism in class removal scenarios that operates without direct access to the forget set. (iii) Comprehensive analyses demonstrating SCAR's efficacy across different datasets and architectures in", + "bbox": [ + 212, + 598, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Bonato et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "class-removal and homogeneous-removal scenarios $(iv)$ Experimental evidence showing SCAR's comparable or superior performance respectively to traditional unlearning methods or state-of-the-art (sota) techniques that do not utilize a retain set. Code is available at https://github.com/jbonato1/SCAR", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 214, + 232, + 397, + 248 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Machine Unlearning algorithms can be classified into three main categories: data-driven, model-intrinsic, and model-agnostic unlearning algorithms [44].", + "bbox": [ + 212, + 265, + 782, + 295 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data-driven. The unlearning algorithms in this category employ a data-driven approach to remove knowledge about specific image instances. For example, methods that involve data manipulation, such as transformations or attacks, fall within this group of algorithms [50, 53]. However, this category is not limited to such models: it encompasses a broader range of approaches. In [19, 61] the authors propose similar methods for reverting the gradient updates for the instances of the classes to be removed. Despite their effectiveness, these methods require storing gradient updates for all the batches in the training set. Similarly, in [42], the authors propose two gradient-descent-based approaches for the effective unlearning of convex models. In the first, Gaussian noise is used to perturb the gradient updates, while the second also incorporates reservoir sampling. In [60], the authors modeled the impact of each training point on the model's training process and then compensated for the impact of removing data from the model's knowledge by re-weighting the gradient updates. In [14], the authors employed a strategy of recursively assigning an incorrect class to samples designated for forgetting until a membership inference attack no longer recognizes those samples as part of the training dataset. In [3], the authors segmented the dataset into n parts and trained n models on each segment; during inference, predictions from these models are aggregated. Only the model containing those samples is retrained when a removal request is received. Moreover, novel retain-set-free algorithms such as Random Labels [23] and Negative Gradient [17] have been introduced. In the former approach, random labels are assigned to samples designated for forgetting to calculate the cross-entropy, while the latter employs the negative gradient obtained from these samples to optimize the loss function.", + "bbox": [ + 212, + 296, + 787, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Model-intrinsic. Model-intrinsic approaches to machine unlearning consist of algorithms that depend on a specific architecture. In these cases, the unlearning process is facilitated by sophisticated mechanisms tailored to function within that particular architecture. Depending on the task to be performed several architectures can be involved. For instance in [17, 18] the authors proposed two different unlearning approaches for CNNs. In the former, the authors \"scrubbed\" from the weights of the model the knowledge about instances of images through the computation of Fisher's information matrix. Similarly in the latter, the authors propose a CNN-based unlearning algorithm, inspired by the neural tangent kernel, that models the weights dynamics and can remove from the information about forget data. In [35], the authors proposed a two-phase algorithm for the approximate unlearning of CNNs. During the first phase, called ERM, the model", + "bbox": [ + 212, + 659, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "SCAR", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/508aedf6204bab9f065a87c3edff878b7e2b074de67306ab54a77c6a48f2ed85.jpg", + "image_caption": [ + "A" + ], + "image_footnote": [], + "bbox": [ + 259, + 166, + 521, + 333 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {J S} (\\Phi_ {\\theta} (x _ {j}) | | \\Phi_ {\\theta} ^ {U} (x _ {j}))\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 151, + 696, + 166 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a007c1a9022ea8fed443e9a79f0cbc3c8fb150f9a7f6601305fe512c9e6874a3.jpg", + "image_caption": [ + "B" + ], + "image_footnote": [], + "bbox": [ + 537, + 180, + 635, + 229 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/812df6d6f2395bc5d161cd48572da11d0e04f976d1fb544e91e6ffdc11a2c91c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 180, + 740, + 229 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {M} (\\Phi_ {\\psi} ^ {U} (x _ {j}), Q _ {j} ^ {*})\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 241, + 681, + 255 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/97b2e9e4f74a8e5ddf4ecdafe5c1c2768a918b60e1bb6be0098d8b1ebcff378d.jpg", + "image_caption": [ + "C", + "Fig. 1: SCAR overview. A) SCAR scheme B) distillation-trick through J.S. divergence of the logits of $\\varPhi_{\\theta}(x_j)$ and $\\varPhi_{\\theta}^{U}(x_{j})$ . C) Representation of the feature vector distributions of samples for 3 classes before (left) and after (right) the unlearning process. Ellipses denote the $95\\%$ confidence intervals of the distribution of samples." + ], + "image_footnote": [], + "bbox": [ + 540, + 256, + 740, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is trained on all the data while a binary mask is learned for each layer of the CNN for each class of the dataset. Then during the second phase, called KTP, the knowledge about the classes of the forget set is removed using the previously learned masks. In [20], the authors propose a certified algorithm for removing knowledge about classes on linear classifiers. This method applied a Newton step on the model's weights to significantly reduce the influence of the deleted data point, with the residual error decreasing quadratically as the training set size increases. Then, for certifying the data removal and to prevent the extraction of information from the minimal residual the training loss is perturbed. Recently, additional unlearning approaches have been developed for dealing with specific models on different learning paradigms such as Multimodal learning [7, 46] and Adversarial Learning [37, 57].", + "bbox": [ + 212, + 429, + 787, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model-agnostic. This group of algorithms includes universally applicable methods across different architectures. Their characteristic is the ability to intervene in the feature/activation space to selectively erase data associated with the forget set. In [32], the authors used knowledge distillation to differentiate the logits of the forget set from those of the retain set and erase the knowledge of these examples. A similar strategy was proposed in [30], where distillation was combined with contrastive labeling to achieve efficient unlearning. In [4] the authors proposed two methods for unlearning based on the kinematics of the decision boundary. The methods change the labels of the examples of the forget set to the closest wrong class and then fine-tune the model with the reassigned samples. Similarly, In [9] the authors proposed a metric learning-based approach that associates samples from the forget set to the closest centroid in the feature space. Through this procedure, the model preserves the feature space organization. In [28] the authors proposed a model sparsification via weight pruning for unlearning data samples. In [56], the authors proposed to learn a matrix", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Bonato et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "noise associated with the classes to be forgotten to remove their influence on the model's weights followed by a repair step used to recover knowledge about retain samples.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our approach stands out as a model-agnostic approximate unlearning algorithm and partially aligns with Boundary Unlearning [4] or DUCK [9] that manipulates respectively forget samples labels or forget samples feature vectors to induce unlearning. However, while these methods adjust samples towards incorrect directions based on the nearest wrong class or centroid in feature space, our solution focuses on using the distributions of features to efficiently reorganize the positions of forget samples in the feature space. This enables us to utilize a more precise and efficient mechanism, grounded in the Mahalanobis distance. Nonetheless, SCAR does not rely on any images from the retain set to restore performance during the unlearning process whereas model-agnostic methods like DUCK [9], Boundary Unlearning [4], and SCRUB [32] partially or fully depend on it. SCAR, thanks to an efficient knowledge preservation method based on distillation and OOD data, can operate without accessing training data.", + "bbox": [ + 212, + 191, + 787, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methods", + "text_level": 1, + "bbox": [ + 215, + 411, + 338, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 444, + 372, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consider a dataset $\\{x_i, y_i\\}_{i=1}^N$ consisting of N images $x_i$ each associated with a label $y_i \\in \\{0, \\dots, K\\}$ where $K$ is the number of classes in the dataset. A model $\\varPhi_\\theta$ is trained on the training subset of the dataset $\\mathcal{D}$ to predict the corresponding label for a given image, that is, $\\hat{y}_i = \\varPhi_\\theta(x_i)$ . This model is then evaluated on the test subset $\\mathcal{D}^t$ . The objective of a machine unlearning algorithm is to derive a model, denoted as $\\varPhi_\\theta^U$ , from which information about a specific subset of data, known as the forget set $\\mathcal{D}_f$ , is removed from the original model's parameters $\\theta$ . Concurrently, the algorithm aims to preserve the model's performance on the retain set $\\mathcal{D}_r$ , such that it is comparable to that of a model trained exclusively on the retain set. In this paper, we consider two scenarios: Class-Removal (CR) and Homogeneous Removal (HR).", + "bbox": [ + 212, + 468, + 787, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the CR scenario, the objective is to remove from the model's weights the information about the instances of a class $C$ among the $K$ available. The two subsets $\\mathcal{D}$ and $\\mathcal{D}^t$ are split in forget $(\\mathcal{D}_f, \\mathcal{D}_f^t)$ and retain $(\\mathcal{D}_r, \\mathcal{D}_r^t)$ sets. The forget set contains all the images, from the class to be removed $\\mathcal{D}_f = \\{x_i^f, y_i^f\\}_{i=1}^{N^f}$ with $y_i^f \\in \\{C\\}$ while the retain set contains all the remaining images $\\mathcal{D}_r = \\{x_i^r, y_i^r\\}_{i=1}^{N^r}$ , with $y_i^r \\in \\{0, \\dots, K-1\\}$ ; the same definition stands also for $\\mathcal{D}^t$ . The desiderata in this scenario is to maintain the original accuracy on the $\\mathcal{D}_r^t$ while minimizing the one on the $\\mathcal{D}_f^t$ . In the HR scenario, the training dataset $\\mathcal{D}$ is divided into retain $\\mathcal{D}_r$ and forget $\\mathcal{D}_f$ sets, while the test set $\\mathcal{D}^t$ remains unchanged. In this case, the images constituting $\\mathcal{D}_f$ are chosen from all the classes. Here, the goal is to eliminate the model's knowledge of specific training instances in $\\mathcal{D}_f$ without compromising its performance on $\\mathcal{D}_r$ . Following the application of the unlearning algorithm, the model should no longer differentiate", + "bbox": [ + 212, + 636, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "SCAR", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "between data from $\\mathcal{D}_f$ and $\\mathcal{D}^t$ or in other words it interprets $\\mathcal{D}_f$ samples as never-seen-data. This indistinguishability has also to be reflected in the $\\mathcal{D}_f$ accuracy which has to be approximately similar to $\\mathcal{D}^t$ accuracy. This second scenario is more challenging compared to CR since $\\mathcal{D}_r$ and $\\mathcal{D}_f$ include images from the same classes and the model should maintain its ability to generalize across all classes.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 SCAR", + "text_level": 1, + "bbox": [ + 215, + 258, + 316, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In SCAR, the unlearning process is based on metric learning and involves directly modifying the feature vector projection of the sample to be forgotten within the feature space. Additionally, we have devised a regularization technique that enables the DNN to preserve the original model information through distillation without access to the retain dataset. By leveraging an OOD dataset, SCAR enables the transfer of knowledge from the original model to the unlearning model, thereby ensuring that classification performance on test samples remains consistent throughout the unlearning process. Initially, we present the unlearning strategy of SCAR, followed by a deeper exploration of the distillation with OOD data and the Metric Learning mechanism. Then we introduce a variant of SCAR, called SCAR self-forget, that works without accessing the forget set.", + "bbox": [ + 212, + 282, + 787, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unlearning strategy. The unlearning strategy is composed of 2 synergistic mechanisms, Metric Learning and Distillation-Trick (Fig. 1 A). In the following sections each DNN $\\varPhi_{\\theta}$ will be represented as a sequence of a backbone $\\varPhi_{\\psi}$ and a final fully-connected layer $\\varPhi_{\\pi}$ where $\\theta, \\psi, \\pi$ are the corresponding weight.", + "bbox": [ + 212, + 453, + 787, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During the training phase of the original model $\\varPhi_{\\theta}$ , for each class $i^{th}$ the distribution of feature vectors $Q_{i} = \\{\\varPhi_{\\psi}(x_{j})\\}_{j}$ is prototyped [10, 22, 47] by its mean $\\mu_{i}$ and covariance matrix $\\hat{S}_Q$ (where $\\varPhi_{\\psi}(x_j)$ represent a feature vector), and accurately stored preventing any data leakage. Hence, given a forget sample $(x_{j},y_{j} = k)$ belonging to the $k^{th}$ class, SCAR selects the closest distribution $Q_{j}^{*}$ among the set of distribution $Q_{i}$ with $i\\in \\{0,\\dots,K\\}$ and $i\\neq k$ to the feature vector $\\varPhi_{\\psi}^{U}(x_{j})$ .", + "bbox": [ + 214, + 513, + 787, + 621 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {j} ^ {*} = \\underset {Q _ {i}} {\\operatorname {a r g m i n}} d _ {M} \\left(\\Phi_ {\\psi} ^ {U} \\left(x _ {j}\\right), Q _ {i}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 622, + 785, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hence, during the unlearning procedure, the algorithm, through the Metric Learning mechanism, minimizes the Mahalanobis distance between the forget samples feature vector $\\varPhi_{\\psi}^{U}(x_{j})$ and their corresponding selected distribution $Q_{j}^{*}$ . The overall forget loss is written in eq. 2:", + "bbox": [ + 214, + 654, + 787, + 715 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M} = \\frac {1}{N _ {f , \\text {b a t c h}}} \\sum_ {j = 0} ^ {N _ {f, \\text {b a t c h}} - 1} d _ {M} \\left(\\Phi_ {\\psi} ^ {U} \\left(x _ {j}\\right), Q _ {j} ^ {*}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 726, + 785, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $N_{f,\\mathrm{batch}}$ is the batch size of forget samples.", + "bbox": [ + 214, + 779, + 575, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since in our proposed scenario, the $\\mathcal{D}_r$ dataset is no longer accessible, the algorithm makes use of the Distillation-Trick mechanism to retain past knowledge. We used the Jensen-Shannon divergence between the original model $\\varPhi_{\\theta}$", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Bonato et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/039f2d5e6cab050938a51226004522e67a253b762229ec9546be3223b50a1b15.jpg", + "image_caption": [ + "Fig.2: A-B) TSNE plots of feature vectors for the first 20 classes in CIFAR100 (A, $\\mathcal{D}$ ) and the surrogate dataset subset Imagenet (B, $\\mathcal{D}^{\\mathrm{sur}}$ )" + ], + "image_footnote": [], + "bbox": [ + 267, + 143, + 495, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6641b2536b84e30313f6f89d685f791ede64b203dd34e4f08083a134ceb5d615.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 143, + 727, + 215 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and the unlearning model $\\varPhi_{\\theta}^{U}$ (see par. Distillation-Trick, eq. 5)", + "bbox": [ + 214, + 282, + 679, + 299 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {T D} = \\frac {1}{N _ {r , \\text {b a t c h}}} \\sum_ {j = 0} ^ {N _ {r, \\text {b a t c h}} - 1} d _ {J S} \\left(\\Phi_ {\\theta} ^ {U} \\left(x _ {j}\\right) \\| \\Phi_ {\\theta} \\left(x _ {j}\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 306, + 784, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $N_{r,\\mathrm{batch}}$ is the batch size of the surrogate retain samples. The overall loss combines the forget and retain contribution in eq. 2 and 3.", + "bbox": [ + 214, + 356, + 782, + 386 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {M} + \\lambda_ {2} \\mathcal {L} _ {T D} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 393, + 784, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ are hyperparameters. Instead of selecting a priori the number of unlearning epochs, SCAR computes the $\\mathcal{A}_f$ at each epoch and it stops when $\\mathcal{A}_r$ is lower or equal to a threshold $\\epsilon$ . In CR $\\epsilon = 0$ because the objective is to remove completely the knowledge of the class to remove. In HR $\\epsilon = \\mathcal{A}^t$ because the forget test accuracy has to be close to the accuracy of the original model on the test set. When this condition is not reached the unlearning process stops when it reaches the maximum number of epochs allowed (epochs_max).", + "bbox": [ + 212, + 417, + 782, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Distillation-trick. Numerous instances have been documented in literature wherein DNNs exhibit high confidence predictions [21], even when presented with data significantly distant from their training set. Examples include their performance on fooling images [43], OOD samples [25], or within medical diagnostic tasks [34]. For instance, the t-SNE projection of feature vectors from a DNN, both from the original training dataset and an external OOD dataset, exhibited clustering at identical positions within the feature space (Fig. 2 A-B). In our proposed algorithm, we leverage this characteristic behavior of DNNs on OOD datasets as a form of regularization during the unlearning process, which we term \"distillation-trick\". This approach involves distilling the knowledge from the original frozen model (referred to as the teacher model) into the unlearning model using an external dataset acting as a surrogate dataset, denoted as $\\mathcal{D}^{\\mathrm{sur}}$ [13] (Fig. 1 B). The distillation-trick leverages the Jensen-Shannon divergence between $\\varPhi_{\\theta}^{U}(x_{j})$ and $\\varPhi_{\\theta}(x_{j})$ to retain the knowledge of the original model, even as the unlearning model undergoes training on $\\mathcal{D}^{\\mathrm{sur}}$ (omitting the dependency on $x_{j}$ for brevity):", + "bbox": [ + 214, + 527, + 787, + 770 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nd _ {J S} \\left(\\Phi_ {\\theta} ^ {U} \\| \\Phi_ {\\theta} / T\\right) = \\frac {1}{2} D _ {K L} \\left(\\Phi_ {\\theta} ^ {U} \\| \\Phi_ {\\theta} / T\\right) + \\frac {1}{2} D _ {K L} \\left(\\Phi_ {\\theta} / T \\| \\Phi_ {\\theta} ^ {U}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 281, + 775, + 784, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{T}$ is a scaling temperature. To validate this, we trained $\\varPhi_{\\theta}$ using distillation-trick on a surrogate dataset. Our experiments demonstrate that the accuracy of", + "bbox": [ + 214, + 809, + 789, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "SCAR", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\varPhi_{\\theta}$ on $\\mathcal{D}_t$ remains consistent across epochs and comparable to that of $\\varPhi_{\\theta}^{U}$ (refer to Sec. 1 of the Supp. Materials).", + "bbox": [ + 212, + 145, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Metric learning. In the context of unlearning problems, metric learning serves as a pivotal concept aimed at modifying the feature vectors of forget samples, thereby altering their class predictions. The prevalent strategies involve shifting the feature vectors of forget samples towards the nearest incorrect class centroids [9] or directly shifting the logits towards the nearest incorrect class [4]. In our algorithm, we propose the application of metric learning by adjusting the feature vectors of forget samples toward the nearest distribution of feature vectors of a different class (Fig. 1 C). Thanks to the Mahalanobis distance, which gauges the distance between the forget set samples and the whole distributions of feature vectors, SCAR can integrate more comprehensive information about the distributions through the covariance matrix. This approach contrasts with simpler methods that represent distributions solely using centroids (see Sec.4.4 for a comparison with other distances based on centroids). Therefore, given the set of distribution of feature vectors for each class $Q = \\{Q_{i}\\}_{i = 0}^{K}$ the Mahalanobis distance is defined in eq. 6.", + "bbox": [ + 212, + 180, + 787, + 407 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nd _ {M} (\\varPhi_ {\\psi} ^ {U} (x _ {j}), Q _ {i}) = \\sqrt {(\\varPhi_ {\\psi} ^ {U} (x _ {j}) - \\mu_ {Q}) ^ {T} \\hat {S} _ {Q} ^ {- 1} (\\varPhi_ {\\psi} ^ {U} (x _ {j}) - \\mu_ {Q})} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 303, + 419, + 785, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\varPhi_{\\psi}^{U}(x_{j})$ is the feature vector of $x_{j}$ (i.e. the output of the backbone of $\\varPhi^U$ ), $\\mu_Q$ and $\\hat{S}_Q$ are the mean and the normalized covariance matrix obtained from the distribution $Q_{i}$ . Each class's covariance matrix $S_{Q}$ will exhibit varying levels of scaling and variances across different dimensions. Consequently, the Mahalanobis distances of features from distinct classes will be subjected to differing scaling factors. Therefore, to standardize the covariance matrices per class effectively, we conduct correlation matrix normalization on all the covariance matrices (eq. 7):", + "bbox": [ + 212, + 453, + 787, + 580 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {S} _ {Q} (i, j) = \\frac {S _ {Q} (i , j)}{\\sigma_ {Q} (i) \\sigma_ {Q} (j)}, \\quad \\sigma_ {Q} (j) = \\sqrt {S _ {Q} (j , j)}, \\quad \\sigma_ {Q} (i) = \\sqrt {S _ {Q} (i , i)}. \\qquad (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 264, + 590, + 785, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since the number of samples per class in public datasets ( $\\approx 500$ for CIFAR100 and TinyImagenet) is generally lower than the number of dimensions in the feature space (e.g. $\\mathrm{N} = 512$ for resnet18), we use a covariance shrinkage method to get a full-rank matrix [6], $S_{Q}^{s} = S_{Q} + \\gamma_{0}V_{0}\\mathbb{I} + \\gamma_{1}V_{1}(1 - \\mathbb{I})$ where $V_{0}$ is the average diagonal variance, $V_{2}$ is the average off-diagonal covariance of $S_{Q}$ and $\\mathbb{I}$ is the identity matrix. Furthermore, we added Tukey's normalization [58] to the feature vectors $\\varPhi_{\\psi}^{U}(x)^{\\delta}$ to obtain approximately Gaussian features, where $\\delta$ is a hyperparameter to decide the degree of transformation of the distribution.", + "bbox": [ + 212, + 633, + 787, + 756 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 SCAR Self-forget", + "text_level": 1, + "bbox": [ + 215, + 784, + 406, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, we present an adaptation of SCAR tailored for a more demanding CR scenario, wherein access to the forget set is restricted, and only the class ID", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Bonato et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "slated for removal is provided. We term this procedure SCAR self-forget (Fig. 3). In this version, both the forgetting mechanism and the distillation-trick rely on the surrogate dataset. Given $K_{\\mathrm{ret}}$ , the set of retained classes, and $K_{\\mathrm{fgt}}$ , the set of forget classes, SCAR initially employs the original model $\\varPhi_{\\theta}$ to classify $\\mathcal{D}^{\\mathrm{sur}}$ .", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9bff3dda2b4e3e380f9995e6a43b22677bd66a6be4360ab0533c1aae20a20145.jpg", + "image_caption": [ + "Fig. 3: Scheme of SCAR self-forget in CR. The surrogate dataset $\\mathcal{D}^{\\mathrm{sur}}$ supplies during the unlearning procedure both the surrogates $\\mathcal{D}_r^{\\mathrm{sur}}$ and $\\mathcal{D}_f^{\\mathrm{sur}}$ ." + ], + "image_footnote": [], + "bbox": [ + 218, + 228, + 506, + 364 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Subsequently, SCAR splits the samples in $\\mathcal{D}^{\\mathrm{sur}}$ into $\\mathcal{D}_r^{\\mathrm{sur}}$ and $\\mathcal{D}_f^{\\mathrm{sur}}$ if respectively the predicted class $\\hat{y}$ belongs to $K_{\\mathrm{ret}}$ , and in $\\mathcal{D}_f^{\\mathrm{sur}}$ if $\\hat{y}$ is within $K_{\\mathrm{fgt}}$ . Thus, $\\mathcal{D}_f^{\\mathrm{sur}}$ serves as a surrogate for $\\mathcal{D}_f$ , and SCAR is optimized using the loss function in Eq. 4. This configuration poses an extreme challenge as it lacks both retain and forget data during the unlearning process. SCAR self-forget marks the first attempt to address the unlearning problem within the Class-Removal (CR) scenario without necessitating training or forget data, nor the retention of specific in", + "bbox": [ + 526, + 207, + 787, + 448 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "formation during training such as gradient updates.", + "bbox": [ + 214, + 448, + 589, + 464 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 496, + 460, + 513 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluated SCAR against a range of sota methods and baselines on three datasets CIFAR10 [31], CIFAR100 [31], and TinyImageNet [33]. CIFAR-10 and CIFAR-100 are comprised of 60,000 images of $32 \\times 32$ pixels each (50,000 images for train and 10000 for test), with CIFAR-10 divided into 10 classes, and CIFAR-100 categorized into 100 classes. TinyImageNet consists of 110,000 images of $64 \\times 64$ pixels each, divided into 200 classes (100,000 for train and 10,000 for test). The considered sota methods include DUCK [9], Boundary Shrink [4], Boundary Expanding [4], SCRUB [32], L1-Sparse [28], Bad Teacher [8], ERM-KTP [35], Negative Gradient [17], and Random Labels [23], alongside traditional approaches such as Retraining and Fine-Tuning. While the latter two are computationally demanding as they necessitate access to the $\\mathcal{D}_r$ , they serve as critical benchmarks. Retraining is by definition the upper bound for unlearning, removing the $\\mathcal{D}_f$ from the training process entirely. Fine-Tuning, conversely, adjusts the pre-trained model weights using only the $\\mathcal{D}_r$ . Importantly, changing the number of epochs can substantially affect the finetuning performance: few epochs result in an efficient unlearning whereas if the model is finetuned for many epochs it is substantially retrained. For this reason, the original model is finetuned for a sufficient number of epochs ( $<$ epochs of retrain) that guarantees to unlearn $\\mathcal{D}_f$ (as outlined in [4,35]). Unfortunately, for this reason, Fine-Tuning is extremely time-consuming and inefficient compared to other sota methods [4,9,35]. All", + "bbox": [ + 212, + 537, + 787, + 843 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "SCAR", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/cc54087ead269cb56ed726a66859a8c100115e2ed4883a7d9fd88d74aaa85085.jpg", + "table_caption": [ + "Table 1: Comparison between SCAR performance and different surrogate dataset in CR scenario. The metrics are reported as mean ± std over ten runs." + ], + "table_footnote": [], + "table_body": "
Surrogate dataset DsurSizeCIFAR100TinyImagenet
ArAftAUSp-valArAftAUSp-val
Random Images2*10372.20(01.90)02.60(02.17)0.922(0.027)<10-360.69(01.80)01.20(01.40)0.912(0.022)<10-3
Imagenet subset10472.93(01.78)02.00(02.00)0.935(0.025)<10-362.99(01.39)00.60(01.37)0.940(0.019)<10-3
COCO subset10471.61(01.92)00.84(00.43)0.933(0.023)<10-362.04(01.41)01.00(01.41)0.927(0.019)<10-3
Imagenet distilled10472.13(02.28)02.51(01.62)0.923(0.023)<10-362.36(01.56)01.03(01.41)0.930(0.020)<10-3
Gaussian Noise10416.62(08.64)00.70(00.67)0.388(0.086)<10-317.13(06.08)01.20(01.93)0.482(0.061)<10-3
", + "bbox": [ + 215, + 184, + 785, + 258 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "methods utilized resnet18 [24] as the neural architecture. The details about the hyperparameters are reported in the Supp. Material, Sec. 2, \"Reproducibility\".", + "bbox": [ + 212, + 284, + 782, + 314 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We established two experimental frameworks, CR and HR, based on the scenarios outlined in Sec. 3.1. For CR, we applied a constant seed, dividing each dataset into $\\mathcal{D}_r$ and $\\mathcal{D}_f$ with distinct class distributions. The $\\mathcal{D}_f$ included instances from a single class, while the $\\mathcal{D}_r$ comprised the remaining classes. This setup was replicated across ten splits to calculate average metrics and standard deviations. In the HR scenario, we selected 10 different seeds for dataset partitioning, ensuring a 90:10 split between $\\mathcal{D}_r$ and $\\mathcal{D}_f$ , with class overlaps but unique instances. We reported the mean and standard deviation of the results across these seeds. Further details are available in the Sec. 2 of the Supp. Material.", + "bbox": [ + 212, + 315, + 784, + 450 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For the CR scenario, the metrics considered are the accuracies on retain $(\\mathcal{A}_r^t)$ and forget $(\\mathcal{A}_f^t)$ test sets, to be maximized and minimized, respectively, and the Adaptive Unlearning Score (AUS) [9]. This final score facilitates the comparison among various methods by encapsulating forget and retain test set performances into a single metric. It captures a delicate balance between two essential objectives: preserving high test accuracy while effectively tackling the unlearning task (see Supp. Material Sec. 2). For the HR scenario, the metrics included are accuracy on $\\mathcal{D}_f$ ( $\\mathcal{A}_f$ ), overall test set accuracy $(\\mathcal{A}^t)$ , AUS, and the success of a membership inference attack (MIA) [9, 49, 54], with the optimal MIA value being $\\mathcal{F}_1 = 0.5$ . Specifically the number of forget and test samples are balanced and $\\mathcal{F}_1 = 0.5$ (chance level) means the impossibility to distinguish the membership of the forget data. The Sec. 2 of the Supp. Material provides details about hyperparameters for MIA tuning.", + "bbox": [ + 212, + 450, + 785, + 647 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Impact of different Datasets as $\\mathcal{D}^{\\mathrm{sur}}$", + "text_level": 1, + "bbox": [ + 215, + 666, + 558, + 681 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A critical role in SCAR is played by the distillation-trick mechanism. It allows SCAR to maintain acquired information about the retain set even without accessing it during the unlearning. The distillation-trick works using an OOD dataset, and investigating how different datasets can or cannot affect the efficiency of SCAR is of utmost importance. For this reason, following [2, 15, 27], we selected 4 different OOD external public datasets characterized by different distributions of pixels, number of images, resolution, and number of classes (a random subset of $10\\mathrm{K}$ images of Imagenet1K [11], a random subset of $10\\mathrm{K}$ images of COCO [36], a set of $2\\mathrm{K}$ public natural images [1], and a set of $10\\mathrm{K}$ images distilled with [65] from Imagenet1K). We verified through the Kolmogorov-Smirnov", + "bbox": [ + 212, + 688, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Bonato et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/62afdec70e98e387f2a532b351c728aaa91e5081a0a513370bb51b56ed3fea10.jpg", + "table_caption": [ + "Table 2: Performance comparison between SCAR and sota methods in CR scenario for CIFAR10, CIFAR100, and TinyImagenet datasets. The metrics are reported as mean ± std over ten runs. Results marked with $\\dagger$ are taken from [9]. Columns “ $\\mathcal{D}_r$ free” and “ $\\mathcal{D}_f$ free” indicate whether the method was (✓) or not (✗), Retain-Free or Forget-free, respectively. Green and red colors highlight respectively the best results between methods based on retain set and between retain free methods." + ], + "table_footnote": [], + "table_body": "
DrfreeDffreeCIFAR10CIFAR100TinyImagenet
AicAfjAUSAicAfjAUSAicAfjAUS
Original--88.64(00.63)88.34(00.62)0.531(0.005)77.55(00.11)77.50(02.80)0.563(0.009)68.26(00.08)64.60(15.64)0.607(0.058)
Retrained†--88.05(01.28)0.000(0.000)0.994(0.014)77.97(00.42)0.000(0.000)1.004(0.022)67.67(01.00)0.000(0.000)0.993(0.010)
Fine Tuning‡--87.93(01.14)0.000(0.000)0.993(0.013)77.31(02.17)0.000(0.000)0.998(0.022)67.89(00.24)0.000(0.000)0.994(0.010)
DUCK†XX85.53(01.37)0.000(0.000)0.969(0.015)71.57(02.08)0.010(0.176)0.931(0.026)61.29(01.13)0.020(0.063)0.927(0.013)
Boundary S.†XX83.81(02.29)13.64(03.57)0.840(0.543)55.07(11.85)03.54(02.26)0.749(0.116)55.98(03.33)04.25(02.12)0.840(0.036)
Boundary E.†XX82.36(02.39)13.34(03.21)0.827(0.032)55.18(11.98)03.50(02.32)0.750(0.117)55.92(03.45)02.60(02.50)0.853(0.040)
SCRUBXX87.88(01.18)0.000(0.000)0.992(0.013)77.29(00.25)02.00(05.35)0.977(0.051)68.15(00.25)01.20(03.79)0.986(0.037)
L1-SparseXX85.21(09.41)04.88(02.18)0.862(0.023)62.63(07.56)01.22(05.53)0.879(0.018)63.76(01.29)04.81(01.54)0.901(0.011)
ERM-KTP†XX80.31(-)00.00(-)0.917(-)58.96(-)00.00(-)0.814(-)49.45(-)00.00(-)0.811(-)
Bad TeacherXX88.89(00.87)02.37(05.86)0.982(0.057)77.19(00.19)06.00(10.09)0.940(0.089)64.09(00.78)11.20(09.04)0.862(0.070)
Neg.Grad.†✓X76.27(03.27)00.56(00.12)0.871(0.033)62.84(06.13)00.50(05.50)0.849(0.061)60.09(02.58)0.60(01.35)0.911(0.028)
Rand.Lab.†✓X65.46(07.59)00.83(00.44)0.762(0.076)55.31(07.06)00.40(00.70)0.774(0.070)43.29(10.10)01.20(01.03)0.740(10.100)
SCAR✓X87.71(01.62)00.97(00.24)0.981(0.017)72.93(01.78)02.00(02.00)0.935(0.025)62.99(01.39)0.60(01.37)0.940(0.019)
SCAR Self-Forget✓✓87.18(01.76)01.20(00.43)0.974(0.019)71.09(02.69)00.70(00.95)0.929(0.028)60.79(02.15)00.80(01.40)0.917(0.025)
", + "bbox": [ + 215, + 232, + 787, + 363 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(KS) test [40] that the distributions of pixels of these surrogate datasets are different from the datasets used for benchmarking SCAR (Tab. 1 p-val columns). This is fundamental to assess the generality of SCAR and it does not rely on data leakage between datasets; this is especially true for tinyImagenet and the random subset ofImagenet1K where we observed the KS p-val is $< 0.001$ . Moreover, we also used a surrogate dataset made of Gaussian noise images where the semantic information was absent. We performed our experiments in the CR scenario using CIFAR100 as $\\mathcal{D}$ . Both accuracies and AUS scores were statistically comparable across datasets (Tab. 1) even if theImagenet1K subset gives the overall highest results. Importantly, when in the surrogate dataset semantic information is removed (Gaussian Noise) SCAR performance is poor, suggesting how the surrogate dataset has to contain some spatial feature and patterns to trigger the DNN filters. Hence, the presence of semantic information content in images plays a pivotal role in the distillation-trick mechanism, but, at the same time, the differences in terms of kind of information (e.g. classes and distribution of pixels) do not affect this technique. For the rest of the paper results are presented using as a surrogate the OOD random subset ofImagenet1K.", + "bbox": [ + 212, + 393, + 787, + 650 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Comparison with sota methods", + "text_level": 1, + "bbox": [ + 215, + 676, + 519, + 690 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We evaluated SCAR, and SCAR Self-Forget, within the CR scenario across three datasets (Tab. 2). Our findings indicate that our method surpasses Retain-Free approaches like Negative Gradient and Random Label in performance. Furthermore, SCAR's performance is on par with methods that leverage both forget and retain sets. Notably, SCAR's performance is lower than Fine-Tuning and within 1 std. from SCRUB in CIFAR100 and TinyImagenet, though these methods depend heavily on the $\\mathcal{D}_r$ requiring several finetuning steps on it. Importantly, these outcomes are also achieved with SCAR Self-Forget, which remarkably does not utilize any data from the original training set.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "SCAR", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/c6ee8929e6aa8a2f73ac659b0f43296a7b38fc4d1c66f53a77516ca6dfa86021.jpg", + "table_caption": [ + "Table 3: Performance comparison between SCAR and sota methods in HR scenario for CIFAR10, CIFAR100, and TinyImagenet datasets. The metrics are reported as mean ± std over ten runs. Results marked with $\\dagger$ are taken from [9]. Columns \"D_r free\" indicate whether the method was (✓) or not (X), Retain-Free. Green and red colors highlight respectively the best results between methods based on retain set and between retain free methods." + ], + "table_footnote": [], + "table_body": "
DfreeCIFAR10CIFAR100TinyImagenet
A1AfF1AUSA1AfF1AUSA1AfF1AUS
Original-88.54(00.25)99.49(00.08)54.62(00.80)0.900(0.004)77.23(00.53)99.63(00.10)61.80(00.59)0.814(0.006)68.08(00.39)84.83(00.44)54.06(01.48)0.854(0.007)
Retrained†-84.13(00.59)84.56(00.72)50.22(00.80)0.950(0.011)77.27(01.03)76.87(00.95)49.45(00.40)0.993(0.017)63.45(02.34)63.24(02.58)49.56(00.47)0.950(0.040)
Fine Tuning†-85.70(00.48)88.66(00.44)50.06(01.30)0.942(0.008)72.06(00.52)74.97(00.72)49.68(00.43)0.918(0.009)67.03(00.45)70.52(00.65)49.86(00.56)0.937(0.011)
DUCK†X85.96(00.64)86.05(00.55)50.26(00.49)0.972(0.011)74.74(00.74)75.51(02.01)50.66(00.64)0.965(0.022)62.01(00.79)64.43(00.76)49.87(00.73)0.916(0.014)
SCRBX78.02(07.80)82.53(10.80)50.59(00.89)0.854(0.132)73.33(00.85)92.46(01.50)56.45(01.48)0.804(0.014)68.52(00.40)83.14(00.53)53.89(00.96)0.875(0.008)
L1-SparseX93.05(00.53)99.06(00.65)52.21(00.95)0.927(0.011)52.26(01.65)66.89(02.34)53.46(00.94)0.690(0.024)63.76(01.19)84.30(01.04)53.29(00.73)0.794(0.032)
Bad TeacherX86.66(00.44)92.91(00.57)50.26(00.87)0.922(0.008)73.57(00.61)86.57(00.57)54.44(01.81)0.851(0.006)67.41(00.49)80.28(00.81)54.83(01.44)0.876(0.006)
Neg. Grad.†✓79.35(00.85)87.11(00.86)51.74(00.54)0.841(0.012)60.83(00.77)76.77(00.57)50.98(00.47)0.718(0.009)54.83(00.81)67.27(00.59)49.78(00.33)0.770(0.011)
Rand. Lab.†✓77.28(01.37)87.07(01.20)51.45(00.79)0.807(0.018)60.51(00.82)77.18(00.50)50.55(00.49)0.711(0.009)54.34(00.85)67.65(00.45)49.84(00.64)0.760(0.011)
SCAR✓86.14(00.47)88.40(00.31)50.07(00.59)0.953(0.008)73.23(00.74)75.63(00.54)49.77(00.28)0.934(0.011)61.35(00.69)66.51(00.44)49.85(00.19)0.886(0.011)
", + "bbox": [ + 217, + 233, + 785, + 327 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/29bc71f939f68ccb0143babfa36e45ba9ea58ae0bfd7fe5a05bfacf90982db3d.jpg", + "table_caption": [ + "Table 4: Results of the ablation study on CIFAR100 in CR and HR scenarios. The metrics for CR and HR are reported as mean ± std over ten runs." + ], + "table_footnote": [], + "table_body": "
LTDLMCIFAR100TinyImagenet
ArAfAUSArAfAUS
CRXX77.55(00.11)77.50(02.80)0.563(0.009)68.26(00.08)64.60(15.64)0.607(0.058)
✓X72.09(02.72)40.00(13.36)0.675(0.067)63.14(01.77)29.80(12.09)0.730(0.069)
X✓66.90(05.03)02.60(02.27)0.871(0.053)49.72(02.54)00.80(01.40)0.807(0.028)
✓✓72.93(01.78)02.00(02.00)0.935(0.025)62.99(01.39)00.60(01.37)0.940(0.019)
AtAfAUSAtAfAUS
HRXX77.39(00.46)99.64(00.09)0.817(0.005)68.08(00.39)84.83(00.44)0.854(0.007)
✓X76.53(00.41)99.18(00.14)0.807(0.005)64.84(00.60)80.51(00.51)0.835(0.009)
X✓18.83(01.06)19.69(01.19)0.409(0.013)05.88(02.94)05.51(02.63)0.376(0.033)
✓✓73.23(00.74)75.63(00.54)0.934(0.011)61.35(00.69)66.51(00.44)0.886(0.011)
", + "bbox": [ + 259, + 373, + 743, + 494 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We also extended our analysis to the HR scenario (Tab. 3). Importantly, the Self-Forget variant of our method is not applicable in HR, as the $\\mathcal{D}_f$ is essential to identify the specific instance to forget. In this more demanding context, SCAR surpasses all the performance of Retain-Free approaches. Between the methods that use the $\\mathcal{D}_r$ only DUCK in CIFAR100 (within 1 std.) and TinyImagenet and Fine-Tuning in TinyImagenet achieve superior results.", + "bbox": [ + 212, + 522, + 787, + 612 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Notably, our method outperforms the Retain-Free approaches in both scenarios and demonstrates commendable performance relative to methods that rely on the $\\mathcal{D}_r$ . This distinction highlights the effectiveness and versatility of our approach, making it a viable alternative for scenarios where the $\\mathcal{D}_r$ is entirely unavailable. We also confirmed our result using the COCO dataset as a surrogate dataset (Sec. 1 of the Supp. Material).", + "bbox": [ + 212, + 613, + 787, + 704 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 724, + 390, + 739 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conducted an ablation study to evaluate the impact of different loss components $\\mathcal{L}_{LD}$ and $\\mathcal{L}_M$ of our method on the unlearning performance across CR and HR scenarios. The results of this study are presented in Tab. 4. In CR, considering only $\\mathcal{L}_{LD}$ leads to similar retain set accuracy but with high $\\mathcal{D}_f$ accuracy. In fact, the stopping criteria is hardly fulfilled and the method stops at the maximum number of epochs allowed. Conversely, considering only $\\mathcal{L}_M$ the", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Bonato et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/83f6cd007b2b5a574d42d2053216e1ea9681be5fd4b02fa386deabc8fb015779.jpg", + "table_caption": [ + "Table 5: Results of SCAR changing the measure used in $\\mathcal{L}_M$ on the CIFAR100. The metrics for CR and HR are reported as mean $\\pm$ std over ten runs." + ], + "table_footnote": [], + "table_body": "
CIFAR100 CRCIFAR100 HR
ArrAftAUSArrAftAUS
Cosine Similarity72.74(02.19)02.00(02.36)0.933(0.031)73.88(00.46)77.05(00.54)0.934(0.008)
L2 Distance72.45(02.23)02.70(02.54)0.924(0.032)69.58(00.82)76.93(00.55)0.857(0.011)
Mahalanobis72.93(01.78)02.00(02.00)0.935(0.025)73.23(00.74)75.63(00.54)0.934(0.011)
", + "bbox": [ + 256, + 178, + 741, + 238 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "methods can push down the accuracy on the $\\mathcal{D}_f$ , but at the cost of a low retain set accuracy. In the HR scenario, we observed similar but more pronounced results. The presence of $\\mathcal{L}_{LD}$ alone is insufficient for effectively erasing knowledge about forget samples, as evidenced by a high forget accuracy. Employing only $\\mathcal{L}_{LD}$ results in a steep decrease in both accuracies. This result suggests how the unlearning strategy without a method for retaining knowledge can't be used to unlearn. This effect is more evident in HR because images from the same class are distributed across both the $\\mathcal{D}_r$ and $\\mathcal{D}_f$ , meaning that the $\\mathcal{L}_{LD}$ compromises the general knowledge of the classes.", + "bbox": [ + 212, + 267, + 787, + 404 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/863d2b9093ed822a2889d3f27a4101a611ed54d4561b7aa209158e3ed304aff7.jpg", + "image_caption": [ + "Fig. 4: Values of final AUS for SCAR and SCAR self-forget as a function of the number of samples available in $\\mathcal{D}^{\\mathrm{sur}}$ (subset of Imagenet1K). AUS is reported as mean $\\pm$ std over ten runs" + ], + "image_footnote": [], + "bbox": [ + 240, + 404, + 480, + 525 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/a7a14f3b6e2ed6ea4a704b327f4478145d32ec11e480fee12e3c8923a8b1fbfb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CIFAR100
AfAfAUS
Original (AliCNN)70.08(00.15)69.70(14.83)0.545(0.048)
Unlearned65.35(02.37)02.90(03.38)0.926(0.036)
Original (Resnet18)77.55(00.11)77.50(02.80)0.563(0.009)
Unlearned72.93(01.78)02.00(02.00)0.935(0.025)
Original (Resnet34)80.13(00.10)80.30(10.02)0.569(0.031)
Unlearned75.67(01.74)01.90(00.92)0.937(0.019)
Original (Resnet50)80.64(00.10)80.80(09.50)0.570(0.030)
Unlearned75.99(01.77)00.80(02.36)0.942(0.028)
Original (ViT-B16)85.51(01.45)83.30(06.24)0.546(0.023)
Unlearned81.99(02.59)02.04(03.10)0.945(0.031)
", + "bbox": [ + 529, + 409, + 756, + 523 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 6: Results of SCAR on CIFAR100 dataset in the CR scenario changing the model to unlearn. The metrics are reported as mean ± std over ten runs before and after the unlearning.", + "bbox": [ + 511, + 534, + 782, + 604 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Analysis on the effect of different measures in the metric learning mechanism of SCAR", + "text_level": 1, + "bbox": [ + 214, + 628, + 740, + 659 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We report in Tab. 5 and Supp. Tab. 2 the results of the SCAR using different distance measures for the $\\mathcal{L}_M$ component across CIFAR100 and TinyImagenet datasets in both the CR and HR scenarios. Using the Mahalanobis distance yields a perfect balance, achieving the highest AUS scores in both CR and HR. Cosine Similarity distance achieves compatible results (even if constantly lower) in CR and HR scenarios but at the cost of being $(22 \\pm 5\\%)$ slower than the Mahalanobis. This result suggests how the information about the distributions of samples embedded into the Mahalanobis distance is fundamental to guide the optimization efficiently. In contrast, the $L_2$ distance resulted in lower performance in both CR and HR scenarios suggesting how this metric hampers the unlearning process.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "SCAR", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Analysis of the impact of the dimension of $\\mathcal{D}^{\\mathrm{sur}}$", + "text_level": 1, + "bbox": [ + 215, + 146, + 653, + 161 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Fig. 4 illustrates the relationship between the number of images in the surrogate dataset $D_{sur}$ (subset of ImageNet1K) and the AUS for both SCAR and its self-forget variant, in the CR scenario. SCAR maintains a relatively stable AUS score as the number of images increases, suggesting that the method's unlearning ability does not significantly depend on the size of the surrogate dataset. The self-forget variant of SCAR shows a plateau in AUS score when at least 6K surrogate images are used which suggests how the dimension of $\\mathcal{D}^{\\mathrm{sur}}$ can affect the self-forget mechanism. Importantly, the self-forget mechanism that can access 6K surrogate samples or more exhibits performance comparable with the method that can access forget data. This pattern reflects the effectiveness of SCAR in handling the unlearning process, with only minor improvements seen with the addition of more surrogate data when forget data are not available.", + "bbox": [ + 212, + 169, + 787, + 351 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.6 Architectural-Agnostic Results", + "text_level": 1, + "bbox": [ + 215, + 372, + 517, + 387 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "For additional insight, we examined the performance of SCAR (Tab. 6) across various DNN architectures within the CR scenario. We investigated the effect of the unlearning with SCAR on a basic CNN network (AllCNN), ResNet architectures (18, 34, and 50), and a Vision Transformer (ViT-B16). Our method consistently demonstrates strong performance across all tested architectures. Notably, the results achieved with the ViT confirm the model-agnostic nature of our approach beyond the CNN architectures. It is worth mentioning that the efficacy of our method is not contingent on the initial accuracy of the model, which underscores its adaptability and broad applicability.", + "bbox": [ + 212, + 395, + 787, + 532 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 551, + 369, + 568 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduced SCAR a novel method for approximate unlearning, designed to effectively erase information about the forget-set data while maintaining the original model performance. Our experimental results indicate that SCAR outperforms the methods that do not require access to the retain set and achieves performance levels comparable to, or even surpassing, those of existing approximate unlearning methods that depend on the retain set, across both CR and HR scenarios. Furthermore, we have developed a self-forget variation of SCAR capable of delivering impressive results in CR without accessing both the retain and forget sets. Future work should include a mathematical exploration of the algorithms' certifiability and the development of alternatives to the distillation-trick, enabling the retention of past knowledge without the need for a surrogate dataset, which can limit the application of SCAR to other domain.", + "bbox": [ + 212, + 583, + 785, + 763 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Despite these challenges, we firmly believe that our algorithm, that do not require access to any retain set, can significantly impact society: for example in the performance restoration of models poisoned by corrupted data or in recognition systems (such as facial recognition), by enabling individuals to exercise their right to have their data removed from the model's knowledge.", + "bbox": [ + 212, + 765, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Bonato et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Image captioning dataset, random images. https://www.kaggle.com/datasets/shamsaddin97/image-captioning-dataset-random-images/data, accessed: 2024-02-14 10", + "2. Bibas, K., Feder, M., Hassner, T.: Single layer predictive normalized maximum likelihood for out-of-distribution detection. Advances in Neural Information Processing Systems 34, 1179-1191 (2021) 10", + "3. Bourtoule, L., Chandrasekaran, V., Choquette-Choo, C.A., Jia, H., Travers, A., Zhang, B., Lie, D., Papernot, N.: Machine unlearning. In: 2021 IEEE Symposium on Security and Privacy (SP). pp. 141-159. IEEE (2021) 3", + "4. Chen, M., Gao, W., Liu, G., Peng, K., Wang, C.: Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7766-7775 (June 2023) 4, 5, 8, 9", + "5. Chen, S., Kahla, M., Jia, R., Qi, G.J.: Knowledge-enriched distributional model inversion attacks pp. 16158-16167 (2021). https://doi.org/10.1109/ICCV48922.2021.015871", + "6. Chen, Y., Wiesel, A., Eldar, Y.C., Hero, A.O.: Shrinkage algorithms for mmse covariance estimation. IEEE Transactions on Signal Processing 58, 5016-5029 (2009), https://api-semanticscholar.org/CorpusID:12616083 8", + "7. Cheng, J., Amiri, H.: Multimodal machine unlearning. arXiv preprint arXiv:2311.12047 (2023) 4", + "8. Chandawat, V.S., Tarun, A.K., Mandal, M., Kankanhalli, M.: Can bad teaching induce forgetting? unlearning in deep networks using an incompetent teacher. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 7210-7217 (2023) 9", + "9. Cotogni, M., Bonato, J., Sabetta, L., Pelosiin, F., Nicolosi, A.: Duck: Distance-based unlearning via centroid kinematics. arXiv preprint arXiv:2312.02052 (2023) 4, 5, 8, 9, 10, 11, 12", + "0. De Lange, M., Tuytelaars, T.: Continual prototype evolution: Learning online from non-stationary data streams (2021). https://doi.org/10.1109/ICCV48922.2021.008146", + "1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) 10", + "2. Du, M., Yang, F., Zou, N., Hu, X.: Fairness in deep learning: A computational perspective. IEEE Intelligent Systems 36(4), 25-34 (2021). https://doi.org/10.1109/MIS.2020.3000681 1", + "3. Fang, G., Bao, Y., Song, J., Wang, X., Xie, D., Shen, C., Song, M.: Mosaicking to distill: Knowledge distillation from out-of-domain data. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 11920-11932. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/63dc7ed1010d3c3b8269faf0ba7491d4-Paper.pdf 7", + "4. Felps, D.L., Schwickerath, A.D., Williams, J.D., Vuong, T.N., Briggs, A., Hunt, M., Sakmar, E., Saranchak, D.D., Shumaker, T.: Class clown: Data redaction in machine unlearning at enterprise scale. arXiv preprint arXiv:2012.04699 (2020) 3", + "5. Fort, S., Ren, J., Lakshminarayanan, B.: Exploring the limits of out-of-distribution detection. Advances in Neural Information Processing Systems 34, 7068-7081 (2021) 10" + ], + "bbox": [ + 225, + 172, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "SCAR", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Fredrikson, M., Jha, S., Ristenpart, T.: Model inversion attacks that exploit confidence information and basic countermeasures. Proceedings of the 22nd ACM SIGSAC Conference on Computer and Communications Security (2015), https://apisemantic scholar.org/CorpusID:207229839 1", + "17. Golatkar, A., Achille, A., Soatto, S.: Eternal sunshine of the spotless net: Selective forgetting in deep networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9304-9312 (2020) 3, 9", + "18. Golatkar, A., Achille, A., Soatto, S.: Forgetting outside the box: Scrubbing deep networks of information accessible from input-output observations. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIX 16. pp. 383-398. Springer (2020) 3", + "19. Graves, L., Nagisetty, V., Ganesh, V.: Amnesiac machine learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 11516-11524 (2021) 3", + "20. Guo, C., Goldstein, T., Hannun, A., Van Der Maaten, L.: Certified data removal from machine learning models. In: Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 3832-3842. PMLR (2020) 4", + "21. Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: International conference on machine learning. pp. 1321-1330. PMLR (2017) 7", + "22. Han, X., Dai, Y., Gao, T., Lin, Y., Liu, Z., Li, P., Sun, M., Zhou, J.: Continual relation learning via episodic memory activation and reconsolidation (2020). https://doi.org/10.18653/v1/2020.acl-main.573, https://aclanthology.org/2020.acl-main.573 6", + "23. Hayase, T., Yasutomi, S., Katoh, T.: Selective forgetting of deep networks at a finer level than samples. arXiv preprint arXiv:2012.11849 (2020) 3, 9", + "24. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 10", + "25. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: International Conference on Learning Representations (2017), https://openreview.net/forum?id=Hkg4TI9x1 7", + "26. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) 2", + "27. Hsu, Y.C., Shen, Y., Jin, H., Kira, Z.: Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10951-10960 (2020) 10", + "28. Jia, J., Liu, J., Ram, P., Yao, Y., Liu, G., Liu, Y., Sharma, P., Liu, S.: Model sparsity can simplify machine unlearning. In: Annual Conference on Neural Information Processing Systems (2023) 4, 9", + "29. Kahla, M., Chen, S., Just, H.A., Jia, R.: Label-only model inversion attacks via boundary repulsion. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15025-15033 (2022). https://doi.org/10.1109/CVPR52688.2022.014621", + "30. Kim, J., Woo, S.S.: Efficient two-stage model retraining for machine unlearning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4361-4369 (2022) 4", + "31. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) 9" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Bonato et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Kurmanji, M., Triantafillou, P., Hayes, J., Triantafillou, E.: Towards unbounded machine unlearning. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=OveBaTtUAT 4, 5, 9", + "33. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015) 9", + "34. Leibig, C., Allken, V., Ayhan, M.S., Berens, P., Wahl, S.: Leveraging uncertainty information from deep neural networks for disease detection. Scientific reports 7(1), 17816 (2017) 7", + "35. Lin, S., Zhang, X., Chen, C., Chen, X., Susilo, W.: Erm-ktp: Knowledge-level machine unlearning via knowledge transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20147-20155 (2023) 3, 9", + "36. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) 10", + "37. Liu, J., Xue, M., Lou, J., Zhang, X., Xiong, L., Qin, Z.: Mutter: Machine unlearning on adversarially trained models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4892-4902 (October 2023) 4", + "38. Liu, X., Xie, L., Wang, Y., Zou, J., Xiong, J., Ying, Z., Vasilakos, A.V.: Privacy and security issues in deep learning: A survey. IEEE Access 9, 4566-4593 (2021), https://api-semanticscholar.org/CorpusID:231423554 1", + "39. Magdziarczyk, M.: Right to be forgotten in light of regulation (eu) 2016/679 of the european parliament and of the council of 27 april 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing directive 95/46/ec. In: 6th International Multidisciplinary Scientific Conference on Social Sciences and Art Sgem 2019. pp. 177-184 (2019) 2", + "40. Massey Jr, F.J.: The kolmogorov-smirnov test for goodness of fit. Journal of the American statistical Association 46(253), 68-78 (1951) 11", + "41. Mercuri, S., Khraishi, R., Okhrati, R., Batra, D., Hamill, C., Ghasempour, T., Nowlan, A.: An introduction to machine unlearning. arXiv preprint arXiv:2209.00939 (2022) 2", + "42. Neel, S., Roth, A., Sharifi-Malvajerdi, S.: Descent-to-delete: Gradient-based methods for machine unlearning. In: Algorithmic Learning Theory. pp. 931-962. PMLR (2021) 3", + "43. Nguyen, A., Yosinski, J., Clune, J.: Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 427-436 (2015) 7", + "44. Nguyen, T.T., Huynh, T.T., Nguyen, P.L., Liew, A.W.C., Yin, H., Nguyen, Q.V.H.: A survey of machine unlearning. arXiv preprint arXiv:2209.02299 (2022) 2, 3", + "45. Pardau, S.L.: The california consumer privacy act: Towards a european-style privacy regime in the united states. J. Tech. L. & Pol'y 23, 68 (2018) 2", + "46. Poppi, S., Poppi, T., Cocchi, F., Cornia, M., Baraldi, L., Cucchiara, R.: Removing NSFW Concepts from Vision-and-Language Models for Text-to-Image Retrieval and Generation. arXiv preprint arXiv:2311.16254 (2023) 4", + "47. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning (2017). https://doi.org/10.1109/CVPR.2017.587 6" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "SCAR", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Rigaki, M., Garcia, S.: A survey of privacy attacks in machine learning. ACM Computing Surveys 56(4), 1-34 (Nov 2023). https://doi.org/10.1145/3624010, http://dx.doi.org/10.1145/3624010 1", + "49. Salem, A., Zhang, Y., Humbert, M., Berrang, P., Fritz, M., Backes, M.: Ml-leaks: Model and data independent membership inference attacks and defenses on machine learning models. In: Proceedings of the 26th Annual Network and Distributed System Security Symposium (NDSS) (2019) 1, 10", + "50. Setlur, A., Eysenbach, B., Smith, V., Levine, S.: Adversarial unlearning: Reducing confidence along adversarial directions. Advances in Neural Information Processing Systems 35, 18556-18570 (2022) 3", + "51. Shaik, T.B., Tao, X., Xie, H., Li, L., Zhu, X., Li, Q.: Exploring the landscape of machine unlearning: A comprehensive survey and taxonomy. CoRR (2023) 2", + "52. Shokri, R., Stronati, M., Song, C., Shmatikov, V.: Membership inference attacks against machine learning models. 2017 IEEE Symposium on Security and Privacy (SP) pp. 3-18 (2017) 1", + "53. Sommer, D.M., Song, L., Wagh, S., Mittal, P.: Towards probabilistic verification of machine unlearning. arXiv preprint arXiv:2003.04247 (2020) 3", + "54. Song, L., Mittal, P.: Systematic evaluation of privacy risks of machine learning models (2020) 10", + "55. Song, L., Shokri, R., Mittal, P.: Privacy risks of securing machine learning models against adversarial examples. Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security (2019), https://api-semanticscholar.org/CorpusID:165163934 1", + "56. Tarun, A.K., Chandawat, V.S., Mandal, M., Kankanhalli, M.: Fast yet effective machine unlearning. IEEE Transactions on Neural Networks and Learning Systems (2023) 4", + "57. Tiwary, P., Guha, A., Panda, S., et al.: Adapt then unlearn: Exploiting parameter space semantics for unlearning in generative adversarial networks. arXiv preprint arXiv:2309.14054 (2023) 4", + "58. Tukey, J.W.: Exploratory Data Analysis. Addison-Wesley (1977) 8", + "59. Wang, K.C., FU, Y., Li, K., Khisti, A., Zemel, R., Makhzani, A.: Variational model inversion attacks. Advances in Neural Information Processing Systems 34, 9706-9719 (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/50a074e6a8da4662ae0a29edde722179-Paper.pdf 1", + "60. Wu, G., Hashemi, M., Srinivasa, C.: Puma: Performance unchanged model augmentation for training data removal. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 8675-8682 (2022) 3", + "61. Wu, Y., Dobriban, E., Davidson, S.: Deltagrad: Rapid retraining of machine learning models. In: International Conference on Machine Learning. pp. 10355-10366. PMLR (2020) 3", + "62. Xu, J., Wu, Z., Wang, C., Jia, X.: Machine unlearning: Solutions and challenges. arXiv preprint arXiv:2308.07061 (2023) 2", + "63. Yang, Z., Zhang, J., Chang, E.C., Liang, Z.: Neural network inversion in adversarial setting via background knowledge alignment. Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security p. 225-240 (2019). https://doi.org/10.1145/3319535.3354261, https://doi.org/10.1145/3319535.3354261 1", + "64. Yeom, S., Giacomelli, I., Fredrikson, M., Jha, S.: Privacy risk in machine learning: Analyzing the connection to overfitting. 2018 IEEE 31st Computer Security Foundations Symposium (CSF) pp. 268-282 (2017), https://api-semanticscholar.org/CorpusID:2656445 1" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "J. Bonato et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "65. Yin, Z., Xing, E., Shen, Z.: Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. Advances in Neural Information Processing Systems 36 (2024) 10", + "66. Zhang, Y., Jia, R., Pei, H., Wang, W., Li, B., Song, D.: The secret revealer: Generative model-inversion attacks against deep neural networks. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 250-258 (jun 2020). https://doi.org/10.1109/CVPR42600.2020.00033, https://doi.ieeecomputersociety.org/10.1109/CVPR42600.2020.00033 1", + "67. Zhao, X., Zhang, W., Xiao, X., Lim, B.: Exploiting explanations for model inversion attacks. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 662-672 (oct 2021). https://doi.org/10.1109/ICCV48922.2021.00072, https://doi.ieeeccomputersociety.org/10.1109/ICCV48922.2021.00072 1" + ], + "bbox": [ + 212, + 146, + 787, + 314 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "SCAR", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Is Retain Set All You Need in Machine Unlearning_ Restoring Performance of Unlearned Models with Out-Of-Distribution Images/ef16a8f0-48f4-4778-ba31-599cb1e6436c_model.json b/2024/Is Retain Set All You Need in Machine Unlearning_ Restoring Performance of Unlearned Models with Out-Of-Distribution Images/ef16a8f0-48f4-4778-ba31-599cb1e6436c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..013dc7c2c7a45570ce4b7a07e880eedf128af7eb --- /dev/null +++ b/2024/Is Retain Set All You Need in Machine Unlearning_ Restoring Performance of Unlearned Models with Out-Of-Distribution Images/ef16a8f0-48f4-4778-ba31-599cb1e6436c_model.json @@ -0,0 +1,2427 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.22, + 0.141, + 0.784, + 0.21 + ], + "angle": 0, + "content": "Is Retain Set All You Need in Machine Unlearning? Restoring Performance of Unlearned Models with Out-Of-Distribution Images" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.234, + 0.731, + 0.252 + ], + "angle": 0, + "content": "Jacopo Bonato1,2, Marco Cotogni2*, and Luigi Sabetta1,2" + }, + { + "type": "text", + "bbox": [ + 0.376, + 0.262, + 0.603, + 0.277 + ], + "angle": 0, + "content": "1 Leonardo Labs, Rome, Italy," + }, + { + "type": "text", + "bbox": [ + 0.379, + 0.278, + 0.623, + 0.291 + ], + "angle": 0, + "content": "2 These authors contributed equally" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.292, + 0.756, + 0.305 + ], + "angle": 0, + "content": "{jacopo.bonato,luigi.sabetta}.ext@leonardo.com,cotogni@proton.me" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.343, + 0.744, + 0.581 + ], + "angle": 0, + "content": "Abstract. In this paper, we introduce Selective-distillation for Class and Architecture-agnostic unleaRning (SCAR), a novel approximate unlearning method. SCAR efficiently eliminates specific information while preserving the model's test accuracy without using a retain set, which is a key component in state-of-the-art approximate unlearning algorithms. Our approach utilizes a modified Mahalanobis distance to guide the unlearning of the feature vectors of the instances to be forgotten, aligning them to the nearest wrong class distribution. Moreover, we propose a distillation-trick mechanism that distills the knowledge of the original model into the unlearning model with out-of-distribution images for retaining the original model's test performance without using any retain set. Importantly, we propose a self-forget version of SCAR that unlearns without having access to the forget set. We experimentally verified the effectiveness of our method, on three public datasets, comparing it with state-of-the-art methods. Our method obtains performance higher than methods that operate without the retain set and comparable w.r.t the best methods that rely on the retain set." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.593, + 0.687, + 0.607 + ], + "angle": 0, + "content": "Keywords: Machine Unlearning \\(\\cdot\\) Retain-Free \\(\\cdot\\) Data Removal" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.634, + 0.376, + 0.649 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.665, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Over the past decades, deep learning algorithms have grown significantly in applications like image classification, detection, and segmentation. For optimal performance, these models learn from big training datasets and update their weights to capture intrinsic patterns. Despite their effectiveness, many ethical concerns such as potential biases affecting certain demographics [12] and privacy like unauthorized use of sensitive or personal data [38,48] have been raised. Additionally, methods like model inversion [5,16,29,59,63,66,67], and membership inference attacks [49, 52, 55, 64] raised many concerns about the potential revelation of training data characteristics or can reveal particular data utilization during training. These vulnerabilities highlight the critical necessity for" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.471, + 0.839 + ], + "angle": 0, + "content": "* Work done while at Leonardo Labs" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "J. Bonato et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.208 + ], + "angle": 0, + "content": "implementing safeguards to secure sensitive information within machine learning models. This imperative is fundamental in light of the privacy-preserving regulations that have been put into effect ( e.g. the European Union's General Data Protection Regulation [39] and the California Consumer Privacy Act [45])." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.788, + 0.586 + ], + "angle": 0, + "content": "Machine Unlearning [41, 44, 51, 62] has emerged as a solution to these privacy and ethical challenges, aiming to remove specific information content from algorithms without degrading their performance. Unlearning algorithms can be classified into 3 main groups: model intrinsic, data-driven, and model agnostic. The characteristic of model-intrinsic methods is to change the deep neural network (DNN) architecture or add specific weights to induce the unlearning. The data-driven group is characterized by the usage of data to induce forgetting, and it can be roughly divided into methods that store optimization information during training (e.g. gradient updates) and methods that involve data manipulation. Similarly to this latter group, model-agnostic solutions do not require access to specific optimization-related information and operate directly on the unaltered trained model. Nonetheless, a common practice among these algorithms is to maintain or restore the model performance using the residual portion of the training dataset, known as retain set. Using the retain set to recalibrate models post-unlearning represents a viable strategy, however, it also introduces practical challenges. In scenarios where privacy concerns dictate, only the forget set —images to forget— might be accessible, leaving no retain set for performance restoration. This poses a significant hurdle for models designed to undergo a performance restoration phase after unlearning, potentially compromising their utility. Moreover, when large-scale datasets such as ImageNet or ImageNet-21K are involved and the forget set constitutes a minor fraction, the bulk of the data is the retain set. This disproportionality can significantly inflate the time required to restore the model to its original efficacy. Furthermore, there are even more intricate situations where also the forget set is unavailable, such as when the unlearning task entails forgetting an entire class with access only to its ID." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.84 + ], + "angle": 0, + "content": "This paper presents a novel, model-agnostic unlearning algorithm, Selective-distillation for Class and Architecture-agnostic unleaRning (SCAR), which leverages metric learning and knowledge distillation [26] to efficiently remove targeted information and maintain model accuracy, notably without relying on a retain set. Specifically, we use the Mahalanobis distance to shift the feature vectors of the instances to forget toward the closest distribution of samples of other classes. This distance includes important information regarding the original dataset distribution of samples resulting in an efficient metric learning-based unlearning strategy. Concurrently, distilling the knowledge of the original model into the unlearning model using out-of-distribution (OOD) images, allows SCAR to maintain the performance of the original model on the test set. The main contributions of this work include: (i) SCAR, a novel model-agnostic unlearning algorithm that achieves competitive unlearning performance without using retain data. (ii) A unique self-forget mechanism in class removal scenarios that operates without direct access to the forget set. (iii) Comprehensive analyses demonstrating SCAR's efficacy across different datasets and architectures in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SCAR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "class-removal and homogeneous-removal scenarios \\((iv)\\) Experimental evidence showing SCAR's comparable or superior performance respectively to traditional unlearning methods or state-of-the-art (sota) techniques that do not utilize a retain set. Code is available at https://github.com/jbonato1/SCAR" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.233, + 0.398, + 0.249 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.266, + 0.784, + 0.296 + ], + "angle": 0, + "content": "Machine Unlearning algorithms can be classified into three main categories: data-driven, model-intrinsic, and model-agnostic unlearning algorithms [44]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.789, + 0.659 + ], + "angle": 0, + "content": "Data-driven. The unlearning algorithms in this category employ a data-driven approach to remove knowledge about specific image instances. For example, methods that involve data manipulation, such as transformations or attacks, fall within this group of algorithms [50, 53]. However, this category is not limited to such models: it encompasses a broader range of approaches. In [19, 61] the authors propose similar methods for reverting the gradient updates for the instances of the classes to be removed. Despite their effectiveness, these methods require storing gradient updates for all the batches in the training set. Similarly, in [42], the authors propose two gradient-descent-based approaches for the effective unlearning of convex models. In the first, Gaussian noise is used to perturb the gradient updates, while the second also incorporates reservoir sampling. In [60], the authors modeled the impact of each training point on the model's training process and then compensated for the impact of removing data from the model's knowledge by re-weighting the gradient updates. In [14], the authors employed a strategy of recursively assigning an incorrect class to samples designated for forgetting until a membership inference attack no longer recognizes those samples as part of the training dataset. In [3], the authors segmented the dataset into n parts and trained n models on each segment; during inference, predictions from these models are aggregated. Only the model containing those samples is retrained when a removal request is received. Moreover, novel retain-set-free algorithms such as Random Labels [23] and Negative Gradient [17] have been introduced. In the former approach, random labels are assigned to samples designated for forgetting to calculate the cross-entropy, while the latter employs the negative gradient obtained from these samples to optimize the loss function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Model-intrinsic. Model-intrinsic approaches to machine unlearning consist of algorithms that depend on a specific architecture. In these cases, the unlearning process is facilitated by sophisticated mechanisms tailored to function within that particular architecture. Depending on the task to be performed several architectures can be involved. For instance in [17, 18] the authors proposed two different unlearning approaches for CNNs. In the former, the authors \"scrubbed\" from the weights of the model the knowledge about instances of images through the computation of Fisher's information matrix. Similarly in the latter, the authors propose a CNN-based unlearning algorithm, inspired by the neural tangent kernel, that models the weights dynamics and can remove from the information about forget data. In [35], the authors proposed a two-phase algorithm for the approximate unlearning of CNNs. During the first phase, called ERM, the model" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "J. Bonato et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.147, + 0.279, + 0.158 + ], + "angle": 0, + "content": "A" + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.167, + 0.522, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.147, + 0.56, + 0.157 + ], + "angle": 0, + "content": "B" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.152, + 0.697, + 0.167 + ], + "angle": 0, + "content": "\\[\nd _ {J S} (\\Phi_ {\\theta} (x _ {j}) | | \\Phi_ {\\theta} ^ {U} (x _ {j}))\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.181, + 0.637, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.181, + 0.741, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.239, + 0.561, + 0.25 + ], + "angle": 0, + "content": "C" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.242, + 0.683, + 0.256 + ], + "angle": 0, + "content": "\\[\nd _ {M} (\\Phi_ {\\psi} ^ {U} (x _ {j}), Q _ {j} ^ {*})\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.257, + 0.741, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.339, + 0.788, + 0.396 + ], + "angle": 0, + "content": "Fig. 1: SCAR overview. A) SCAR scheme B) distillation-trick through J.S. divergence of the logits of \\(\\varPhi_{\\theta}(x_j)\\) and \\(\\varPhi_{\\theta}^{U}(x_{j})\\) . C) Representation of the feature vector distributions of samples for 3 classes before (left) and after (right) the unlearning process. Ellipses denote the \\(95\\%\\) confidence intervals of the distribution of samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.789, + 0.612 + ], + "angle": 0, + "content": "is trained on all the data while a binary mask is learned for each layer of the CNN for each class of the dataset. Then during the second phase, called KTP, the knowledge about the classes of the forget set is removed using the previously learned masks. In [20], the authors propose a certified algorithm for removing knowledge about classes on linear classifiers. This method applied a Newton step on the model's weights to significantly reduce the influence of the deleted data point, with the residual error decreasing quadratically as the training set size increases. Then, for certifying the data removal and to prevent the extraction of information from the minimal residual the training loss is perturbed. Recently, additional unlearning approaches have been developed for dealing with specific models on different learning paradigms such as Multimodal learning [7, 46] and Adversarial Learning [37, 57]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Model-agnostic. This group of algorithms includes universally applicable methods across different architectures. Their characteristic is the ability to intervene in the feature/activation space to selectively erase data associated with the forget set. In [32], the authors used knowledge distillation to differentiate the logits of the forget set from those of the retain set and erase the knowledge of these examples. A similar strategy was proposed in [30], where distillation was combined with contrastive labeling to achieve efficient unlearning. In [4] the authors proposed two methods for unlearning based on the kinematics of the decision boundary. The methods change the labels of the examples of the forget set to the closest wrong class and then fine-tune the model with the reassigned samples. Similarly, In [9] the authors proposed a metric learning-based approach that associates samples from the forget set to the closest centroid in the feature space. Through this procedure, the model preserves the feature space organization. In [28] the authors proposed a model sparsification via weight pruning for unlearning data samples. In [56], the authors proposed to learn a matrix" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SCAR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "noise associated with the classes to be forgotten to remove their influence on the model's weights followed by a repair step used to recover knowledge about retain samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.39 + ], + "angle": 0, + "content": "Our approach stands out as a model-agnostic approximate unlearning algorithm and partially aligns with Boundary Unlearning [4] or DUCK [9] that manipulates respectively forget samples labels or forget samples feature vectors to induce unlearning. However, while these methods adjust samples towards incorrect directions based on the nearest wrong class or centroid in feature space, our solution focuses on using the distributions of features to efficiently reorganize the positions of forget samples in the feature space. This enables us to utilize a more precise and efficient mechanism, grounded in the Mahalanobis distance. Nonetheless, SCAR does not rely on any images from the retain set to restore performance during the unlearning process whereas model-agnostic methods like DUCK [9], Boundary Unlearning [4], and SCRUB [32] partially or fully depend on it. SCAR, thanks to an efficient knowledge preservation method based on distillation and OOD data, can operate without accessing training data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.412, + 0.339, + 0.428 + ], + "angle": 0, + "content": "3 Methods" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.445, + 0.373, + 0.459 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.788, + 0.636 + ], + "angle": 0, + "content": "Consider a dataset \\(\\{x_i, y_i\\}_{i=1}^N\\) consisting of N images \\(x_i\\) each associated with a label \\(y_i \\in \\{0, \\dots, K\\}\\) where \\(K\\) is the number of classes in the dataset. A model \\(\\varPhi_\\theta\\) is trained on the training subset of the dataset \\(\\mathcal{D}\\) to predict the corresponding label for a given image, that is, \\(\\hat{y}_i = \\varPhi_\\theta(x_i)\\). This model is then evaluated on the test subset \\(\\mathcal{D}^t\\). The objective of a machine unlearning algorithm is to derive a model, denoted as \\(\\varPhi_\\theta^U\\), from which information about a specific subset of data, known as the forget set \\(\\mathcal{D}_f\\), is removed from the original model's parameters \\(\\theta\\). Concurrently, the algorithm aims to preserve the model's performance on the retain set \\(\\mathcal{D}_r\\), such that it is comparable to that of a model trained exclusively on the retain set. In this paper, we consider two scenarios: Class-Removal (CR) and Homogeneous Removal (HR)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In the CR scenario, the objective is to remove from the model's weights the information about the instances of a class \\( C \\) among the \\( K \\) available. The two subsets \\( \\mathcal{D} \\) and \\( \\mathcal{D}^t \\) are split in forget \\( (\\mathcal{D}_f, \\mathcal{D}_f^t) \\) and retain \\( (\\mathcal{D}_r, \\mathcal{D}_r^t) \\) sets. The forget set contains all the images, from the class to be removed \\( \\mathcal{D}_f = \\{x_i^f, y_i^f\\}_{i=1}^{N^f} \\) with \\( y_i^f \\in \\{C\\} \\) while the retain set contains all the remaining images \\( \\mathcal{D}_r = \\{x_i^r, y_i^r\\}_{i=1}^{N^r} \\), with \\( y_i^r \\in \\{0, \\dots, K-1\\} \\); the same definition stands also for \\( \\mathcal{D}^t \\). The desiderata in this scenario is to maintain the original accuracy on the \\( \\mathcal{D}_r^t \\) while minimizing the one on the \\( \\mathcal{D}_f^t \\). In the HR scenario, the training dataset \\( \\mathcal{D} \\) is divided into retain \\( \\mathcal{D}_r \\) and forget \\( \\mathcal{D}_f \\) sets, while the test set \\( \\mathcal{D}^t \\) remains unchanged. In this case, the images constituting \\( \\mathcal{D}_f \\) are chosen from all the classes. Here, the goal is to eliminate the model's knowledge of specific training instances in \\( \\mathcal{D}_f \\) without compromising its performance on \\( \\mathcal{D}_r \\). Following the application of the unlearning algorithm, the model should no longer differentiate" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "J. Bonato et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "between data from \\(\\mathcal{D}_f\\) and \\(\\mathcal{D}^t\\) or in other words it interprets \\(\\mathcal{D}_f\\) samples as never-seen-data. This indistinguishability has also to be reflected in the \\(\\mathcal{D}_f\\) accuracy which has to be approximately similar to \\(\\mathcal{D}^t\\) accuracy. This second scenario is more challenging compared to CR since \\(\\mathcal{D}_r\\) and \\(\\mathcal{D}_f\\) include images from the same classes and the model should maintain its ability to generalize across all classes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.259, + 0.317, + 0.273 + ], + "angle": 0, + "content": "3.2 SCAR" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.284, + 0.789, + 0.451 + ], + "angle": 0, + "content": "In SCAR, the unlearning process is based on metric learning and involves directly modifying the feature vector projection of the sample to be forgotten within the feature space. Additionally, we have devised a regularization technique that enables the DNN to preserve the original model information through distillation without access to the retain dataset. By leveraging an OOD dataset, SCAR enables the transfer of knowledge from the original model to the unlearning model, thereby ensuring that classification performance on test samples remains consistent throughout the unlearning process. Initially, we present the unlearning strategy of SCAR, followed by a deeper exploration of the distillation with OOD data and the Metric Learning mechanism. Then we introduce a variant of SCAR, called SCAR self-forget, that works without accessing the forget set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.789, + 0.514 + ], + "angle": 0, + "content": "Unlearning strategy. The unlearning strategy is composed of 2 synergistic mechanisms, Metric Learning and Distillation-Trick (Fig. 1 A). In the following sections each DNN \\(\\varPhi_{\\theta}\\) will be represented as a sequence of a backbone \\(\\varPhi_{\\psi}\\) and a final fully-connected layer \\(\\varPhi_{\\pi}\\) where \\(\\theta, \\psi, \\pi\\) are the corresponding weight." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.515, + 0.789, + 0.622 + ], + "angle": 0, + "content": "During the training phase of the original model \\(\\varPhi_{\\theta}\\), for each class \\(i^{th}\\) the distribution of feature vectors \\(Q_{i} = \\{\\varPhi_{\\psi}(x_{j})\\}_{j}\\) is prototyped [10, 22, 47] by its mean \\(\\mu_{i}\\) and covariance matrix \\(\\hat{S}_Q\\) (where \\(\\varPhi_{\\psi}(x_j)\\) represent a feature vector), and accurately stored preventing any data leakage. Hence, given a forget sample \\((x_{j},y_{j} = k)\\) belonging to the \\(k^{th}\\) class, SCAR selects the closest distribution \\(Q_{j}^{*}\\) among the set of distribution \\(Q_{i}\\) with \\(i\\in \\{0,\\dots,K\\}\\) and \\(i\\neq k\\) to the feature vector \\(\\varPhi_{\\psi}^{U}(x_{j})\\)." + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.623, + 0.787, + 0.649 + ], + "angle": 0, + "content": "\\[\nQ _ {j} ^ {*} = \\underset {Q _ {i}} {\\operatorname {a r g m i n}} d _ {M} \\left(\\Phi_ {\\psi} ^ {U} \\left(x _ {j}\\right), Q _ {i}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.655, + 0.788, + 0.716 + ], + "angle": 0, + "content": "Hence, during the unlearning procedure, the algorithm, through the Metric Learning mechanism, minimizes the Mahalanobis distance between the forget samples feature vector \\(\\varPhi_{\\psi}^{U}(x_{j})\\) and their corresponding selected distribution \\(Q_{j}^{*}\\). The overall forget loss is written in eq. 2:" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.727, + 0.787, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M} = \\frac {1}{N _ {f , \\text {b a t c h}}} \\sum_ {j = 0} ^ {N _ {f, \\text {b a t c h}} - 1} d _ {M} \\left(\\Phi_ {\\psi} ^ {U} \\left(x _ {j}\\right), Q _ {j} ^ {*}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.576, + 0.795 + ], + "angle": 0, + "content": "where \\(N_{f,\\mathrm{batch}}\\) is the batch size of forget samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Since in our proposed scenario, the \\(\\mathcal{D}_r\\) dataset is no longer accessible, the algorithm makes use of the Distillation-Trick mechanism to retain past knowledge. We used the Jensen-Shannon divergence between the original model \\(\\varPhi_{\\theta}\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SCAR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.144, + 0.496, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.144, + 0.728, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.228, + 0.784, + 0.258 + ], + "angle": 0, + "content": "Fig.2: A-B) TSNE plots of feature vectors for the first 20 classes in CIFAR100 (A, \\(\\mathcal{D}\\)) and the surrogate dataset subset Imagenet (B, \\(\\mathcal{D}^{\\mathrm{sur}}\\))" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.283, + 0.68, + 0.3 + ], + "angle": 0, + "content": "and the unlearning model \\(\\varPhi_{\\theta}^{U}\\) (see par. Distillation-Trick, eq. 5)" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.307, + 0.785, + 0.351 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {T D} = \\frac {1}{N _ {r , \\text {b a t c h}}} \\sum_ {j = 0} ^ {N _ {r, \\text {b a t c h}} - 1} d _ {J S} \\left(\\Phi_ {\\theta} ^ {U} \\left(x _ {j}\\right) \\| \\Phi_ {\\theta} \\left(x _ {j}\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.357, + 0.784, + 0.387 + ], + "angle": 0, + "content": "where \\( N_{r,\\mathrm{batch}} \\) is the batch size of the surrogate retain samples. The overall loss combines the forget and retain contribution in eq. 2 and 3." + }, + { + "type": "equation", + "bbox": [ + 0.426, + 0.395, + 0.785, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {M} + \\lambda_ {2} \\mathcal {L} _ {T D} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.784, + 0.525 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are hyperparameters. Instead of selecting a priori the number of unlearning epochs, SCAR computes the \\(\\mathcal{A}_f\\) at each epoch and it stops when \\(\\mathcal{A}_r\\) is lower or equal to a threshold \\(\\epsilon\\). In CR \\(\\epsilon = 0\\) because the objective is to remove completely the knowledge of the class to remove. In HR \\(\\epsilon = \\mathcal{A}^t\\) because the forget test accuracy has to be close to the accuracy of the original model on the test set. When this condition is not reached the unlearning process stops when it reaches the maximum number of epochs allowed (epochs_max)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.528, + 0.788, + 0.771 + ], + "angle": 0, + "content": "Distillation-trick. Numerous instances have been documented in literature wherein DNNs exhibit high confidence predictions [21], even when presented with data significantly distant from their training set. Examples include their performance on fooling images [43], OOD samples [25], or within medical diagnostic tasks [34]. For instance, the t-SNE projection of feature vectors from a DNN, both from the original training dataset and an external OOD dataset, exhibited clustering at identical positions within the feature space (Fig. 2 A-B). In our proposed algorithm, we leverage this characteristic behavior of DNNs on OOD datasets as a form of regularization during the unlearning process, which we term \"distillation-trick\". This approach involves distilling the knowledge from the original frozen model (referred to as the teacher model) into the unlearning model using an external dataset acting as a surrogate dataset, denoted as \\(\\mathcal{D}^{\\mathrm{sur}}\\) [13] (Fig. 1 B). The distillation-trick leverages the Jensen-Shannon divergence between \\(\\varPhi_{\\theta}^{U}(x_{j})\\) and \\(\\varPhi_{\\theta}(x_{j})\\) to retain the knowledge of the original model, even as the unlearning model undergoes training on \\(\\mathcal{D}^{\\mathrm{sur}}\\) (omitting the dependency on \\(x_{j}\\) for brevity):" + }, + { + "type": "equation", + "bbox": [ + 0.282, + 0.776, + 0.785, + 0.805 + ], + "angle": 0, + "content": "\\[\nd _ {J S} \\left(\\Phi_ {\\theta} ^ {U} \\| \\Phi_ {\\theta} / T\\right) = \\frac {1}{2} D _ {K L} \\left(\\Phi_ {\\theta} ^ {U} \\| \\Phi_ {\\theta} / T\\right) + \\frac {1}{2} D _ {K L} \\left(\\Phi_ {\\theta} / T \\| \\Phi_ {\\theta} ^ {U}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.79, + 0.84 + ], + "angle": 0, + "content": "where \\( \\mathrm{T} \\) is a scaling temperature. To validate this, we trained \\( \\varPhi_{\\theta} \\) using distillation-trick on a surrogate dataset. Our experiments demonstrate that the accuracy of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "J. Bonato et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.178 + ], + "angle": 0, + "content": "\\(\\varPhi_{\\theta}\\) on \\(\\mathcal{D}_t\\) remains consistent across epochs and comparable to that of \\(\\varPhi_{\\theta}^{U}\\) (refer to Sec. 1 of the Supp. Materials)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.409 + ], + "angle": 0, + "content": "Metric learning. In the context of unlearning problems, metric learning serves as a pivotal concept aimed at modifying the feature vectors of forget samples, thereby altering their class predictions. The prevalent strategies involve shifting the feature vectors of forget samples towards the nearest incorrect class centroids [9] or directly shifting the logits towards the nearest incorrect class [4]. In our algorithm, we propose the application of metric learning by adjusting the feature vectors of forget samples toward the nearest distribution of feature vectors of a different class (Fig. 1 C). Thanks to the Mahalanobis distance, which gauges the distance between the forget set samples and the whole distributions of feature vectors, SCAR can integrate more comprehensive information about the distributions through the covariance matrix. This approach contrasts with simpler methods that represent distributions solely using centroids (see Sec.4.4 for a comparison with other distances based on centroids). Therefore, given the set of distribution of feature vectors for each class \\( Q = \\{Q_{i}\\}_{i = 0}^{K} \\) the Mahalanobis distance is defined in eq. 6." + }, + { + "type": "equation", + "bbox": [ + 0.304, + 0.42, + 0.786, + 0.445 + ], + "angle": 0, + "content": "\\[\nd _ {M} (\\varPhi_ {\\psi} ^ {U} (x _ {j}), Q _ {i}) = \\sqrt {(\\varPhi_ {\\psi} ^ {U} (x _ {j}) - \\mu_ {Q}) ^ {T} \\hat {S} _ {Q} ^ {- 1} (\\varPhi_ {\\psi} ^ {U} (x _ {j}) - \\mu_ {Q})} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.788, + 0.581 + ], + "angle": 0, + "content": "where \\(\\varPhi_{\\psi}^{U}(x_{j})\\) is the feature vector of \\(x_{j}\\) (i.e. the output of the backbone of \\(\\varPhi^U\\)), \\(\\mu_Q\\) and \\(\\hat{S}_Q\\) are the mean and the normalized covariance matrix obtained from the distribution \\(Q_{i}\\). Each class's covariance matrix \\(S_{Q}\\) will exhibit varying levels of scaling and variances across different dimensions. Consequently, the Mahalanobis distances of features from distinct classes will be subjected to differing scaling factors. Therefore, to standardize the covariance matrices per class effectively, we conduct correlation matrix normalization on all the covariance matrices (eq. 7):" + }, + { + "type": "equation", + "bbox": [ + 0.265, + 0.592, + 0.787, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\hat {S} _ {Q} (i, j) = \\frac {S _ {Q} (i , j)}{\\sigma_ {Q} (i) \\sigma_ {Q} (j)}, \\quad \\sigma_ {Q} (j) = \\sqrt {S _ {Q} (j , j)}, \\quad \\sigma_ {Q} (i) = \\sqrt {S _ {Q} (i , i)}. \\qquad (7)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.634, + 0.788, + 0.757 + ], + "angle": 0, + "content": "Since the number of samples per class in public datasets (\\(\\approx 500\\) for CIFAR100 and TinyImagenet) is generally lower than the number of dimensions in the feature space (e.g. \\(\\mathrm{N} = 512\\) for resnet18), we use a covariance shrinkage method to get a full-rank matrix [6], \\(S_{Q}^{s} = S_{Q} + \\gamma_{0}V_{0}\\mathbb{I} + \\gamma_{1}V_{1}(1 - \\mathbb{I})\\) where \\(V_{0}\\) is the average diagonal variance, \\(V_{2}\\) is the average off-diagonal covariance of \\(S_{Q}\\) and \\(\\mathbb{I}\\) is the identity matrix. Furthermore, we added Tukey's normalization [58] to the feature vectors \\(\\varPhi_{\\psi}^{U}(x)^{\\delta}\\) to obtain approximately Gaussian features, where \\(\\delta\\) is a hyperparameter to decide the degree of transformation of the distribution." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.785, + 0.407, + 0.801 + ], + "angle": 0, + "content": "3.3 SCAR Self-forget" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Additionally, we present an adaptation of SCAR tailored for a more demanding CR scenario, wherein access to the forget set is restricted, and only the class ID" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SCAR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "slated for removal is provided. We term this procedure SCAR self-forget (Fig. 3). In this version, both the forgetting mechanism and the distillation-trick rely on the surrogate dataset. Given \\( K_{\\mathrm{ret}} \\), the set of retained classes, and \\( K_{\\mathrm{fgt}} \\), the set of forget classes, SCAR initially employs the original model \\( \\varPhi_{\\theta} \\) to classify \\( \\mathcal{D}^{\\mathrm{sur}} \\)." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.229, + 0.508, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.373, + 0.518, + 0.43 + ], + "angle": 0, + "content": "Fig. 3: Scheme of SCAR self-forget in CR. The surrogate dataset \\(\\mathcal{D}^{\\mathrm{sur}}\\) supplies during the unlearning procedure both the surrogates \\(\\mathcal{D}_r^{\\mathrm{sur}}\\) and \\(\\mathcal{D}_f^{\\mathrm{sur}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.208, + 0.789, + 0.449 + ], + "angle": 0, + "content": "Subsequently, SCAR splits the samples in \\(\\mathcal{D}^{\\mathrm{sur}}\\) into \\(\\mathcal{D}_r^{\\mathrm{sur}}\\) and \\(\\mathcal{D}_f^{\\mathrm{sur}}\\) if respectively the predicted class \\(\\hat{y}\\) belongs to \\(K_{\\mathrm{ret}}\\), and in \\(\\mathcal{D}_f^{\\mathrm{sur}}\\) if \\(\\hat{y}\\) is within \\(K_{\\mathrm{fgt}}\\). Thus, \\(\\mathcal{D}_f^{\\mathrm{sur}}\\) serves as a surrogate for \\(\\mathcal{D}_f\\), and SCAR is optimized using the loss function in Eq. 4. This configuration poses an extreme challenge as it lacks both retain and forget data during the unlearning process. SCAR self-forget marks the first attempt to address the unlearning problem within the Class-Removal (CR) scenario without necessitating training or forget data, nor the retention of specific in" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.449, + 0.59, + 0.465 + ], + "angle": 0, + "content": "formation during training such as gradient updates." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.497, + 0.462, + 0.515 + ], + "angle": 0, + "content": "4 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.789, + 0.844 + ], + "angle": 0, + "content": "We evaluated SCAR against a range of sota methods and baselines on three datasets CIFAR10 [31], CIFAR100 [31], and TinyImageNet [33]. CIFAR-10 and CIFAR-100 are comprised of 60,000 images of \\(32 \\times 32\\) pixels each (50,000 images for train and 10000 for test), with CIFAR-10 divided into 10 classes, and CIFAR-100 categorized into 100 classes. TinyImageNet consists of 110,000 images of \\(64 \\times 64\\) pixels each, divided into 200 classes (100,000 for train and 10,000 for test). The considered sota methods include DUCK [9], Boundary Shrink [4], Boundary Expanding [4], SCRUB [32], L1-Sparse [28], Bad Teacher [8], ERM-KTP [35], Negative Gradient [17], and Random Labels [23], alongside traditional approaches such as Retraining and Fine-Tuning. While the latter two are computationally demanding as they necessitate access to the \\(\\mathcal{D}_r\\), they serve as critical benchmarks. Retraining is by definition the upper bound for unlearning, removing the \\(\\mathcal{D}_f\\) from the training process entirely. Fine-Tuning, conversely, adjusts the pre-trained model weights using only the \\(\\mathcal{D}_r\\). Importantly, changing the number of epochs can substantially affect the finetuning performance: few epochs result in an efficient unlearning whereas if the model is finetuned for many epochs it is substantially retrained. For this reason, the original model is finetuned for a sufficient number of epochs (\\(<\\) epochs of retrain) that guarantees to unlearn \\(\\mathcal{D}_f\\) (as outlined in [4,35]). Unfortunately, for this reason, Fine-Tuning is extremely time-consuming and inefficient compared to other sota methods [4,9,35]. All" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "J. Bonato et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 1: Comparison between SCAR performance and different surrogate dataset in CR scenario. The metrics are reported as mean ± std over ten runs." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.185, + 0.787, + 0.259 + ], + "angle": 0, + "content": "
Surrogate dataset DsurSizeCIFAR100TinyImagenet
ArAftAUSp-valArAftAUSp-val
Random Images2*10372.20(01.90)02.60(02.17)0.922(0.027)<10-360.69(01.80)01.20(01.40)0.912(0.022)<10-3
Imagenet subset10472.93(01.78)02.00(02.00)0.935(0.025)<10-362.99(01.39)00.60(01.37)0.940(0.019)<10-3
COCO subset10471.61(01.92)00.84(00.43)0.933(0.023)<10-362.04(01.41)01.00(01.41)0.927(0.019)<10-3
Imagenet distilled10472.13(02.28)02.51(01.62)0.923(0.023)<10-362.36(01.56)01.03(01.41)0.930(0.020)<10-3
Gaussian Noise10416.62(08.64)00.70(00.67)0.388(0.086)<10-317.13(06.08)01.20(01.93)0.482(0.061)<10-3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.285, + 0.784, + 0.315 + ], + "angle": 0, + "content": "methods utilized resnet18 [24] as the neural architecture. The details about the hyperparameters are reported in the Supp. Material, Sec. 2, \"Reproducibility\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.316, + 0.785, + 0.451 + ], + "angle": 0, + "content": "We established two experimental frameworks, CR and HR, based on the scenarios outlined in Sec. 3.1. For CR, we applied a constant seed, dividing each dataset into \\( \\mathcal{D}_r \\) and \\( \\mathcal{D}_f \\) with distinct class distributions. The \\( \\mathcal{D}_f \\) included instances from a single class, while the \\( \\mathcal{D}_r \\) comprised the remaining classes. This setup was replicated across ten splits to calculate average metrics and standard deviations. In the HR scenario, we selected 10 different seeds for dataset partitioning, ensuring a 90:10 split between \\( \\mathcal{D}_r \\) and \\( \\mathcal{D}_f \\), with class overlaps but unique instances. We reported the mean and standard deviation of the results across these seeds. Further details are available in the Sec. 2 of the Supp. Material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.452, + 0.787, + 0.648 + ], + "angle": 0, + "content": "For the CR scenario, the metrics considered are the accuracies on retain \\((\\mathcal{A}_r^t)\\) and forget \\((\\mathcal{A}_f^t)\\) test sets, to be maximized and minimized, respectively, and the Adaptive Unlearning Score (AUS) [9]. This final score facilitates the comparison among various methods by encapsulating forget and retain test set performances into a single metric. It captures a delicate balance between two essential objectives: preserving high test accuracy while effectively tackling the unlearning task (see Supp. Material Sec. 2). For the HR scenario, the metrics included are accuracy on \\(\\mathcal{D}_f\\) (\\(\\mathcal{A}_f\\)), overall test set accuracy \\((\\mathcal{A}^t)\\), AUS, and the success of a membership inference attack (MIA) [9, 49, 54], with the optimal MIA value being \\(\\mathcal{F}_1 = 0.5\\). Specifically the number of forget and test samples are balanced and \\(\\mathcal{F}_1 = 0.5\\) (chance level) means the impossibility to distinguish the membership of the forget data. The Sec. 2 of the Supp. Material provides details about hyperparameters for MIA tuning." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.667, + 0.559, + 0.683 + ], + "angle": 0, + "content": "4.1 Impact of different Datasets as \\(\\mathcal{D}^{\\mathrm{sur}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.841 + ], + "angle": 0, + "content": "A critical role in SCAR is played by the distillation-trick mechanism. It allows SCAR to maintain acquired information about the retain set even without accessing it during the unlearning. The distillation-trick works using an OOD dataset, and investigating how different datasets can or cannot affect the efficiency of SCAR is of utmost importance. For this reason, following [2, 15, 27], we selected 4 different OOD external public datasets characterized by different distributions of pixels, number of images, resolution, and number of classes (a random subset of \\(10\\mathrm{K}\\) images of Imagenet1K [11], a random subset of \\(10\\mathrm{K}\\) images of COCO [36], a set of \\(2\\mathrm{K}\\) public natural images [1], and a set of \\(10\\mathrm{K}\\) images distilled with [65] from Imagenet1K). We verified through the Kolmogorov-Smirnov" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SCAR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 2: Performance comparison between SCAR and sota methods in CR scenario for CIFAR10, CIFAR100, and TinyImagenet datasets. The metrics are reported as mean ± std over ten runs. Results marked with \\(\\dagger\\) are taken from [9]. Columns “\\(\\mathcal{D}_r\\) free” and “\\(\\mathcal{D}_f\\) free” indicate whether the method was (✓) or not (✗), Retain-Free or Forget-free, respectively. Green and red colors highlight respectively the best results between methods based on retain set and between retain free methods." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.233, + 0.789, + 0.364 + ], + "angle": 0, + "content": "
DrfreeDffreeCIFAR10CIFAR100TinyImagenet
AicAfjAUSAicAfjAUSAicAfjAUS
Original--88.64(00.63)88.34(00.62)0.531(0.005)77.55(00.11)77.50(02.80)0.563(0.009)68.26(00.08)64.60(15.64)0.607(0.058)
Retrained†--88.05(01.28)0.000(0.000)0.994(0.014)77.97(00.42)0.000(0.000)1.004(0.022)67.67(01.00)0.000(0.000)0.993(0.010)
Fine Tuning‡--87.93(01.14)0.000(0.000)0.993(0.013)77.31(02.17)0.000(0.000)0.998(0.022)67.89(00.24)0.000(0.000)0.994(0.010)
DUCK†XX85.53(01.37)0.000(0.000)0.969(0.015)71.57(02.08)0.010(0.176)0.931(0.026)61.29(01.13)0.020(0.063)0.927(0.013)
Boundary S.†XX83.81(02.29)13.64(03.57)0.840(0.543)55.07(11.85)03.54(02.26)0.749(0.116)55.98(03.33)04.25(02.12)0.840(0.036)
Boundary E.†XX82.36(02.39)13.34(03.21)0.827(0.032)55.18(11.98)03.50(02.32)0.750(0.117)55.92(03.45)02.60(02.50)0.853(0.040)
SCRUBXX87.88(01.18)0.000(0.000)0.992(0.013)77.29(00.25)02.00(05.35)0.977(0.051)68.15(00.25)01.20(03.79)0.986(0.037)
L1-SparseXX85.21(09.41)04.88(02.18)0.862(0.023)62.63(07.56)01.22(05.53)0.879(0.018)63.76(01.29)04.81(01.54)0.901(0.011)
ERM-KTP†XX80.31(-)00.00(-)0.917(-)58.96(-)00.00(-)0.814(-)49.45(-)00.00(-)0.811(-)
Bad TeacherXX88.89(00.87)02.37(05.86)0.982(0.057)77.19(00.19)06.00(10.09)0.940(0.089)64.09(00.78)11.20(09.04)0.862(0.070)
Neg.Grad.†✓X76.27(03.27)00.56(00.12)0.871(0.033)62.84(06.13)00.50(05.50)0.849(0.061)60.09(02.58)0.60(01.35)0.911(0.028)
Rand.Lab.†✓X65.46(07.59)00.83(00.44)0.762(0.076)55.31(07.06)00.40(00.70)0.774(0.070)43.29(10.10)01.20(01.03)0.740(10.100)
SCAR✓X87.71(01.62)00.97(00.24)0.981(0.017)72.93(01.78)02.00(02.00)0.935(0.025)62.99(01.39)0.60(01.37)0.940(0.019)
SCAR Self-Forget✓✓87.18(01.76)01.20(00.43)0.974(0.019)71.09(02.69)00.70(00.95)0.929(0.028)60.79(02.15)00.80(01.40)0.917(0.025)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.789, + 0.651 + ], + "angle": 0, + "content": "(KS) test [40] that the distributions of pixels of these surrogate datasets are different from the datasets used for benchmarking SCAR (Tab. 1 p-val columns). This is fundamental to assess the generality of SCAR and it does not rely on data leakage between datasets; this is especially true for tinyImagenet and the random subset ofImagenet1K where we observed the KS p-val is \\(< 0.001\\). Moreover, we also used a surrogate dataset made of Gaussian noise images where the semantic information was absent. We performed our experiments in the CR scenario using CIFAR100 as \\(\\mathcal{D}\\). Both accuracies and AUS scores were statistically comparable across datasets (Tab. 1) even if theImagenet1K subset gives the overall highest results. Importantly, when in the surrogate dataset semantic information is removed (Gaussian Noise) SCAR performance is poor, suggesting how the surrogate dataset has to contain some spatial feature and patterns to trigger the DNN filters. Hence, the presence of semantic information content in images plays a pivotal role in the distillation-trick mechanism, but, at the same time, the differences in terms of kind of information (e.g. classes and distribution of pixels) do not affect this technique. For the rest of the paper results are presented using as a surrogate the OOD random subset ofImagenet1K." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.677, + 0.52, + 0.691 + ], + "angle": 0, + "content": "4.2 Comparison with sota methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We evaluated SCAR, and SCAR Self-Forget, within the CR scenario across three datasets (Tab. 2). Our findings indicate that our method surpasses Retain-Free approaches like Negative Gradient and Random Label in performance. Furthermore, SCAR's performance is on par with methods that leverage both forget and retain sets. Notably, SCAR's performance is lower than Fine-Tuning and within 1 std. from SCRUB in CIFAR100 and TinyImagenet, though these methods depend heavily on the \\(\\mathcal{D}_r\\) requiring several finetuning steps on it. Importantly, these outcomes are also achieved with SCAR Self-Forget, which remarkably does not utilize any data from the original training set." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "J. Bonato et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 3: Performance comparison between SCAR and sota methods in HR scenario for CIFAR10, CIFAR100, and TinyImagenet datasets. The metrics are reported as mean ± std over ten runs. Results marked with \\(\\dagger\\) are taken from [9]. Columns \"D_r free\" indicate whether the method was (✓) or not (X), Retain-Free. Green and red colors highlight respectively the best results between methods based on retain set and between retain free methods." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.234, + 0.787, + 0.328 + ], + "angle": 0, + "content": "
DfreeCIFAR10CIFAR100TinyImagenet
A1AfF1AUSA1AfF1AUSA1AfF1AUS
Original-88.54(00.25)99.49(00.08)54.62(00.80)0.900(0.004)77.23(00.53)99.63(00.10)61.80(00.59)0.814(0.006)68.08(00.39)84.83(00.44)54.06(01.48)0.854(0.007)
Retrained†-84.13(00.59)84.56(00.72)50.22(00.80)0.950(0.011)77.27(01.03)76.87(00.95)49.45(00.40)0.993(0.017)63.45(02.34)63.24(02.58)49.56(00.47)0.950(0.040)
Fine Tuning†-85.70(00.48)88.66(00.44)50.06(01.30)0.942(0.008)72.06(00.52)74.97(00.72)49.68(00.43)0.918(0.009)67.03(00.45)70.52(00.65)49.86(00.56)0.937(0.011)
DUCK†X85.96(00.64)86.05(00.55)50.26(00.49)0.972(0.011)74.74(00.74)75.51(02.01)50.66(00.64)0.965(0.022)62.01(00.79)64.43(00.76)49.87(00.73)0.916(0.014)
SCRBX78.02(07.80)82.53(10.80)50.59(00.89)0.854(0.132)73.33(00.85)92.46(01.50)56.45(01.48)0.804(0.014)68.52(00.40)83.14(00.53)53.89(00.96)0.875(0.008)
L1-SparseX93.05(00.53)99.06(00.65)52.21(00.95)0.927(0.011)52.26(01.65)66.89(02.34)53.46(00.94)0.690(0.024)63.76(01.19)84.30(01.04)53.29(00.73)0.794(0.032)
Bad TeacherX86.66(00.44)92.91(00.57)50.26(00.87)0.922(0.008)73.57(00.61)86.57(00.57)54.44(01.81)0.851(0.006)67.41(00.49)80.28(00.81)54.83(01.44)0.876(0.006)
Neg. Grad.†✓79.35(00.85)87.11(00.86)51.74(00.54)0.841(0.012)60.83(00.77)76.77(00.57)50.98(00.47)0.718(0.009)54.83(00.81)67.27(00.59)49.78(00.33)0.770(0.011)
Rand. Lab.†✓77.28(01.37)87.07(01.20)51.45(00.79)0.807(0.018)60.51(00.82)77.18(00.50)50.55(00.49)0.711(0.009)54.34(00.85)67.65(00.45)49.84(00.64)0.760(0.011)
SCAR✓86.14(00.47)88.40(00.31)50.07(00.59)0.953(0.008)73.23(00.74)75.63(00.54)49.77(00.28)0.934(0.011)61.35(00.69)66.51(00.44)49.85(00.19)0.886(0.011)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.34, + 0.786, + 0.369 + ], + "angle": 0, + "content": "Table 4: Results of the ablation study on CIFAR100 in CR and HR scenarios. The metrics for CR and HR are reported as mean ± std over ten runs." + }, + { + "type": "table", + "bbox": [ + 0.26, + 0.374, + 0.744, + 0.496 + ], + "angle": 0, + "content": "
LTDLMCIFAR100TinyImagenet
ArAfAUSArAfAUS
CRXX77.55(00.11)77.50(02.80)0.563(0.009)68.26(00.08)64.60(15.64)0.607(0.058)
✓X72.09(02.72)40.00(13.36)0.675(0.067)63.14(01.77)29.80(12.09)0.730(0.069)
X✓66.90(05.03)02.60(02.27)0.871(0.053)49.72(02.54)00.80(01.40)0.807(0.028)
✓✓72.93(01.78)02.00(02.00)0.935(0.025)62.99(01.39)00.60(01.37)0.940(0.019)
AtAfAUSAtAfAUS
HRXX77.39(00.46)99.64(00.09)0.817(0.005)68.08(00.39)84.83(00.44)0.854(0.007)
✓X76.53(00.41)99.18(00.14)0.807(0.005)64.84(00.60)80.51(00.51)0.835(0.009)
X✓18.83(01.06)19.69(01.19)0.409(0.013)05.88(02.94)05.51(02.63)0.376(0.033)
✓✓73.23(00.74)75.63(00.54)0.934(0.011)61.35(00.69)66.51(00.44)0.886(0.011)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.613 + ], + "angle": 0, + "content": "We also extended our analysis to the HR scenario (Tab. 3). Importantly, the Self-Forget variant of our method is not applicable in HR, as the \\(\\mathcal{D}_f\\) is essential to identify the specific instance to forget. In this more demanding context, SCAR surpasses all the performance of Retain-Free approaches. Between the methods that use the \\(\\mathcal{D}_r\\) only DUCK in CIFAR100 (within 1 std.) and TinyImagenet and Fine-Tuning in TinyImagenet achieve superior results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.705 + ], + "angle": 0, + "content": "Notably, our method outperforms the Retain-Free approaches in both scenarios and demonstrates commendable performance relative to methods that rely on the \\(\\mathcal{D}_r\\). This distinction highlights the effectiveness and versatility of our approach, making it a viable alternative for scenarios where the \\(\\mathcal{D}_r\\) is entirely unavailable. We also confirmed our result using the COCO dataset as a surrogate dataset (Sec. 1 of the Supp. Material)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.726, + 0.391, + 0.741 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We conducted an ablation study to evaluate the impact of different loss components \\(\\mathcal{L}_{LD}\\) and \\(\\mathcal{L}_M\\) of our method on the unlearning performance across CR and HR scenarios. The results of this study are presented in Tab. 4. In CR, considering only \\(\\mathcal{L}_{LD}\\) leads to similar retain set accuracy but with high \\(\\mathcal{D}_f\\) accuracy. In fact, the stopping criteria is hardly fulfilled and the method stops at the maximum number of epochs allowed. Conversely, considering only \\(\\mathcal{L}_M\\) the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SCAR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.173 + ], + "angle": 0, + "content": "Table 5: Results of SCAR changing the measure used in \\(\\mathcal{L}_M\\) on the CIFAR100. The metrics for CR and HR are reported as mean \\(\\pm\\) std over ten runs." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.179, + 0.742, + 0.239 + ], + "angle": 0, + "content": "
CIFAR100 CRCIFAR100 HR
ArrAftAUSArrAftAUS
Cosine Similarity72.74(02.19)02.00(02.36)0.933(0.031)73.88(00.46)77.05(00.54)0.934(0.008)
L2 Distance72.45(02.23)02.70(02.54)0.924(0.032)69.58(00.82)76.93(00.55)0.857(0.011)
Mahalanobis72.93(01.78)02.00(02.00)0.935(0.025)73.23(00.74)75.63(00.54)0.934(0.011)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.405 + ], + "angle": 0, + "content": "methods can push down the accuracy on the \\(\\mathcal{D}_f\\), but at the cost of a low retain set accuracy. In the HR scenario, we observed similar but more pronounced results. The presence of \\(\\mathcal{L}_{LD}\\) alone is insufficient for effectively erasing knowledge about forget samples, as evidenced by a high forget accuracy. Employing only \\(\\mathcal{L}_{LD}\\) results in a steep decrease in both accuracies. This result suggests how the unlearning strategy without a method for retaining knowledge can't be used to unlearn. This effect is more evident in HR because images from the same class are distributed across both the \\(\\mathcal{D}_r\\) and \\(\\mathcal{D}_f\\), meaning that the \\(\\mathcal{L}_{LD}\\) compromises the general knowledge of the classes." + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.405, + 0.482, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.535, + 0.489, + 0.605 + ], + "angle": 0, + "content": "Fig. 4: Values of final AUS for SCAR and SCAR self-forget as a function of the number of samples available in \\(\\mathcal{D}^{\\mathrm{sur}}\\) (subset of Imagenet1K). AUS is reported as mean \\(\\pm\\) std over ten runs" + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.41, + 0.758, + 0.525 + ], + "angle": 0, + "content": "
CIFAR100
AfAfAUS
Original (AliCNN)70.08(00.15)69.70(14.83)0.545(0.048)
Unlearned65.35(02.37)02.90(03.38)0.926(0.036)
Original (Resnet18)77.55(00.11)77.50(02.80)0.563(0.009)
Unlearned72.93(01.78)02.00(02.00)0.935(0.025)
Original (Resnet34)80.13(00.10)80.30(10.02)0.569(0.031)
Unlearned75.67(01.74)01.90(00.92)0.937(0.019)
Original (Resnet50)80.64(00.10)80.80(09.50)0.570(0.030)
Unlearned75.99(01.77)00.80(02.36)0.942(0.028)
Original (ViT-B16)85.51(01.45)83.30(06.24)0.546(0.023)
Unlearned81.99(02.59)02.04(03.10)0.945(0.031)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.535, + 0.784, + 0.605 + ], + "angle": 0, + "content": "Table 6: Results of SCAR on CIFAR100 dataset in the CR scenario changing the model to unlearn. The metrics are reported as mean ± std over ten runs before and after the unlearning." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.63, + 0.741, + 0.66 + ], + "angle": 0, + "content": "4.4 Analysis on the effect of different measures in the metric learning mechanism of SCAR" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We report in Tab. 5 and Supp. Tab. 2 the results of the SCAR using different distance measures for the \\(\\mathcal{L}_M\\) component across CIFAR100 and TinyImagenet datasets in both the CR and HR scenarios. Using the Mahalanobis distance yields a perfect balance, achieving the highest AUS scores in both CR and HR. Cosine Similarity distance achieves compatible results (even if constantly lower) in CR and HR scenarios but at the cost of being \\((22 \\pm 5\\%)\\) slower than the Mahalanobis. This result suggests how the information about the distributions of samples embedded into the Mahalanobis distance is fundamental to guide the optimization efficiently. In contrast, the \\(L_2\\) distance resulted in lower performance in both CR and HR scenarios suggesting how this metric hampers the unlearning process." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "J. Bonato et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.655, + 0.162 + ], + "angle": 0, + "content": "4.5 Analysis of the impact of the dimension of \\(\\mathcal{D}^{\\mathrm{sur}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.352 + ], + "angle": 0, + "content": "Fig. 4 illustrates the relationship between the number of images in the surrogate dataset \\( D_{sur} \\) (subset of ImageNet1K) and the AUS for both SCAR and its self-forget variant, in the CR scenario. SCAR maintains a relatively stable AUS score as the number of images increases, suggesting that the method's unlearning ability does not significantly depend on the size of the surrogate dataset. The self-forget variant of SCAR shows a plateau in AUS score when at least 6K surrogate images are used which suggests how the dimension of \\( \\mathcal{D}^{\\mathrm{sur}} \\) can affect the self-forget mechanism. Importantly, the self-forget mechanism that can access 6K surrogate samples or more exhibits performance comparable with the method that can access forget data. This pattern reflects the effectiveness of SCAR in handling the unlearning process, with only minor improvements seen with the addition of more surrogate data when forget data are not available." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.373, + 0.518, + 0.388 + ], + "angle": 0, + "content": "4.6 Architectural-Agnostic Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.788, + 0.533 + ], + "angle": 0, + "content": "For additional insight, we examined the performance of SCAR (Tab. 6) across various DNN architectures within the CR scenario. We investigated the effect of the unlearning with SCAR on a basic CNN network (AllCNN), ResNet architectures (18, 34, and 50), and a Vision Transformer (ViT-B16). Our method consistently demonstrates strong performance across all tested architectures. Notably, the results achieved with the ViT confirm the model-agnostic nature of our approach beyond the CNN architectures. It is worth mentioning that the efficacy of our method is not contingent on the initial accuracy of the model, which underscores its adaptability and broad applicability." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.553, + 0.37, + 0.569 + ], + "angle": 0, + "content": "5 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.787, + 0.765 + ], + "angle": 0, + "content": "In this paper, we introduced SCAR a novel method for approximate unlearning, designed to effectively erase information about the forget-set data while maintaining the original model performance. Our experimental results indicate that SCAR outperforms the methods that do not require access to the retain set and achieves performance levels comparable to, or even surpassing, those of existing approximate unlearning methods that depend on the retain set, across both CR and HR scenarios. Furthermore, we have developed a self-forget variation of SCAR capable of delivering impressive results in CR without accessing both the retain and forget sets. Future work should include a mathematical exploration of the algorithms' certifiability and the development of alternatives to the distillation-trick, enabling the retention of past knowledge without the need for a surrogate dataset, which can limit the application of SCAR to other domain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Despite these challenges, we firmly believe that our algorithm, that do not require access to any retain set, can significantly impact society: for example in the performance restoration of models poisoned by corrupted data or in recognition systems (such as facial recognition), by enabling individuals to exercise their right to have their data removed from the model's knowledge." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SCAR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.786, + 0.214 + ], + "angle": 0, + "content": "1. Image captioning dataset, random images. https://www.kaggle.com/datasets/shamsaddin97/image-captioning-dataset-random-images/data, accessed: 2024-02-14 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.215, + 0.786, + 0.255 + ], + "angle": 0, + "content": "2. Bibas, K., Feder, M., Hassner, T.: Single layer predictive normalized maximum likelihood for out-of-distribution detection. Advances in Neural Information Processing Systems 34, 1179-1191 (2021) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.256, + 0.786, + 0.296 + ], + "angle": 0, + "content": "3. Bourtoule, L., Chandrasekaran, V., Choquette-Choo, C.A., Jia, H., Travers, A., Zhang, B., Lie, D., Papernot, N.: Machine unlearning. In: 2021 IEEE Symposium on Security and Privacy (SP). pp. 141-159. IEEE (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.297, + 0.786, + 0.35 + ], + "angle": 0, + "content": "4. Chen, M., Gao, W., Liu, G., Peng, K., Wang, C.: Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7766-7775 (June 2023) 4, 5, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.786, + 0.39 + ], + "angle": 0, + "content": "5. Chen, S., Kahla, M., Jia, R., Qi, G.J.: Knowledge-enriched distributional model inversion attacks pp. 16158-16167 (2021). https://doi.org/10.1109/ICCV48922.2021.015871" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.786, + 0.432 + ], + "angle": 0, + "content": "6. Chen, Y., Wiesel, A., Eldar, Y.C., Hero, A.O.: Shrinkage algorithms for mmse covariance estimation. IEEE Transactions on Signal Processing 58, 5016-5029 (2009), https://api-semanticscholar.org/CorpusID:12616083 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.786, + 0.459 + ], + "angle": 0, + "content": "7. Cheng, J., Amiri, H.: Multimodal machine unlearning. arXiv preprint arXiv:2311.12047 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.459, + 0.786, + 0.513 + ], + "angle": 0, + "content": "8. Chandawat, V.S., Tarun, A.K., Mandal, M., Kankanhalli, M.: Can bad teaching induce forgetting? unlearning in deep networks using an incompetent teacher. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 7210-7217 (2023) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.514, + 0.786, + 0.554 + ], + "angle": 0, + "content": "9. Cotogni, M., Bonato, J., Sabetta, L., Pelosiin, F., Nicolosi, A.: Duck: Distance-based unlearning via centroid kinematics. arXiv preprint arXiv:2312.02052 (2023) 4, 5, 8, 9, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.555, + 0.786, + 0.593 + ], + "angle": 0, + "content": "0. De Lange, M., Tuytelaars, T.: Continual prototype evolution: Learning online from non-stationary data streams (2021). https://doi.org/10.1109/ICCV48922.2021.008146" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.595, + 0.786, + 0.636 + ], + "angle": 0, + "content": "1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.636, + 0.786, + 0.675 + ], + "angle": 0, + "content": "2. Du, M., Yang, F., Zou, N., Hu, X.: Fairness in deep learning: A computational perspective. IEEE Intelligent Systems 36(4), 25-34 (2021). https://doi.org/10.1109/MIS.2020.3000681 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.676, + 0.786, + 0.758 + ], + "angle": 0, + "content": "3. Fang, G., Bao, Y., Song, J., Wang, X., Xie, D., Shen, C., Song, M.: Mosaicking to distill: Knowledge distillation from out-of-domain data. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 11920-11932. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/63dc7ed1010d3c3b8269faf0ba7491d4-Paper.pdf 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.786, + 0.799 + ], + "angle": 0, + "content": "4. Felps, D.L., Schwickerath, A.D., Williams, J.D., Vuong, T.N., Briggs, A., Hunt, M., Sakmar, E., Saranchak, D.D., Shumaker, T.: Class clown: Data redaction in machine unlearning at enterprise scale. arXiv preprint arXiv:2012.04699 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "5. Fort, S., Ren, J., Lakshminarayanan, B.: Exploring the limits of out-of-distribution detection. Advances in Neural Information Processing Systems 34, 7068-7081 (2021) 10" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "J. Bonato et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "16. Fredrikson, M., Jha, S., Ristenpart, T.: Model inversion attacks that exploit confidence information and basic countermeasures. Proceedings of the 22nd ACM SIGSAC Conference on Computer and Communications Security (2015), https://apisemantic scholar.org/CorpusID:207229839 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.244 + ], + "angle": 0, + "content": "17. Golatkar, A., Achille, A., Soatto, S.: Eternal sunshine of the spotless net: Selective forgetting in deep networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9304-9312 (2020) 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.244, + 0.787, + 0.299 + ], + "angle": 0, + "content": "18. Golatkar, A., Achille, A., Soatto, S.: Forgetting outside the box: Scrubbing deep networks of information accessible from input-output observations. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIX 16. pp. 383-398. Springer (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.299, + 0.787, + 0.337 + ], + "angle": 0, + "content": "19. Graves, L., Nagisetty, V., Ganesh, V.: Amnesiac machine learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 11516-11524 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.787, + 0.393 + ], + "angle": 0, + "content": "20. Guo, C., Goldstein, T., Hannun, A., Van Der Maaten, L.: Certified data removal from machine learning models. In: Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 3832-3842. PMLR (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.787, + 0.434 + ], + "angle": 0, + "content": "21. Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: International conference on machine learning. pp. 1321-1330. PMLR (2017) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.787, + 0.489 + ], + "angle": 0, + "content": "22. Han, X., Dai, Y., Gao, T., Lin, Y., Liu, Z., Li, P., Sun, M., Zhou, J.: Continual relation learning via episodic memory activation and reconsolidation (2020). https://doi.org/10.18653/v1/2020.acl-main.573, https://aclanthology.org/2020.acl-main.573 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.489, + 0.787, + 0.516 + ], + "angle": 0, + "content": "23. Hayase, T., Yasutomi, S., Katoh, T.: Selective forgetting of deep networks at a finer level than samples. arXiv preprint arXiv:2012.11849 (2020) 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.787, + 0.556 + ], + "angle": 0, + "content": "24. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.787, + 0.598 + ], + "angle": 0, + "content": "25. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: International Conference on Learning Representations (2017), https://openreview.net/forum?id=Hkg4TI9x1 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.787, + 0.623 + ], + "angle": 0, + "content": "26. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.787, + 0.678 + ], + "angle": 0, + "content": "27. Hsu, Y.C., Shen, Y., Jin, H., Kira, Z.: Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10951-10960 (2020) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.787, + 0.718 + ], + "angle": 0, + "content": "28. Jia, J., Liu, J., Ram, P., Yao, Y., Liu, G., Liu, Y., Sharma, P., Liu, S.: Model sparsity can simplify machine unlearning. In: Annual Conference on Neural Information Processing Systems (2023) 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.772 + ], + "angle": 0, + "content": "29. Kahla, M., Chen, S., Just, H.A., Jia, R.: Label-only model inversion attacks via boundary repulsion. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15025-15033 (2022). https://doi.org/10.1109/CVPR52688.2022.014621" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "30. Kim, J., Woo, S.S.: Efficient two-stage model retraining for machine unlearning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4361-4369 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "31. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) 9" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SCAR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.202 + ], + "angle": 0, + "content": "32. Kurmanji, M., Triantafillou, P., Hayes, J., Triantafillou, E.: Towards unbounded machine unlearning. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=OveBaTtUAT 4, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.233 + ], + "angle": 0, + "content": "33. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.234, + 0.786, + 0.274 + ], + "angle": 0, + "content": "34. Leibig, C., Allken, V., Ayhan, M.S., Berens, P., Wahl, S.: Leveraging uncertainty information from deep neural networks for disease detection. Scientific reports 7(1), 17816 (2017) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.276, + 0.787, + 0.33 + ], + "angle": 0, + "content": "35. Lin, S., Zhang, X., Chen, C., Chen, X., Susilo, W.: Erm-ktp: Knowledge-level machine unlearning via knowledge transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20147-20155 (2023) 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.332, + 0.786, + 0.387 + ], + "angle": 0, + "content": "36. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.389, + 0.787, + 0.431 + ], + "angle": 0, + "content": "37. Liu, J., Xue, M., Lou, J., Zhang, X., Xiong, L., Qin, Z.: Mutter: Machine unlearning on adversarially trained models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4892-4902 (October 2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.432, + 0.787, + 0.472 + ], + "angle": 0, + "content": "38. Liu, X., Xie, L., Wang, Y., Zou, J., Xiong, J., Ying, Z., Vasilakos, A.V.: Privacy and security issues in deep learning: A survey. IEEE Access 9, 4566-4593 (2021), https://api-semanticscholar.org/CorpusID:231423554 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.474, + 0.787, + 0.555 + ], + "angle": 0, + "content": "39. Magdziarczyk, M.: Right to be forgotten in light of regulation (eu) 2016/679 of the european parliament and of the council of 27 april 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing directive 95/46/ec. In: 6th International Multidisciplinary Scientific Conference on Social Sciences and Art Sgem 2019. pp. 177-184 (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.557, + 0.787, + 0.584 + ], + "angle": 0, + "content": "40. Massey Jr, F.J.: The kolmogorov-smirnov test for goodness of fit. Journal of the American statistical Association 46(253), 68-78 (1951) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.586, + 0.787, + 0.626 + ], + "angle": 0, + "content": "41. Mercuri, S., Khraishi, R., Okhrati, R., Batra, D., Hamill, C., Ghasempour, T., Nowlan, A.: An introduction to machine unlearning. arXiv preprint arXiv:2209.00939 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.628, + 0.787, + 0.669 + ], + "angle": 0, + "content": "42. Neel, S., Roth, A., Sharifi-Malvajerdi, S.: Descent-to-delete: Gradient-based methods for machine unlearning. In: Algorithmic Learning Theory. pp. 931-962. PMLR (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.671, + 0.787, + 0.712 + ], + "angle": 0, + "content": "43. Nguyen, A., Yosinski, J., Clune, J.: Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 427-436 (2015) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.714, + 0.787, + 0.741 + ], + "angle": 0, + "content": "44. Nguyen, T.T., Huynh, T.T., Nguyen, P.L., Liew, A.W.C., Yin, H., Nguyen, Q.V.H.: A survey of machine unlearning. arXiv preprint arXiv:2209.02299 (2022) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.742, + 0.787, + 0.769 + ], + "angle": 0, + "content": "45. Pardau, S.L.: The california consumer privacy act: Towards a european-style privacy regime in the united states. J. Tech. L. & Pol'y 23, 68 (2018) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.771, + 0.787, + 0.811 + ], + "angle": 0, + "content": "46. Poppi, S., Poppi, T., Cocchi, F., Cornia, M., Baraldi, L., Cucchiara, R.: Removing NSFW Concepts from Vision-and-Language Models for Text-to-Image Retrieval and Generation. arXiv preprint arXiv:2311.16254 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "47. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning (2017). https://doi.org/10.1109/CVPR.2017.587 6" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "J. Bonato et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "48. Rigaki, M., Garcia, S.: A survey of privacy attacks in machine learning. ACM Computing Surveys 56(4), 1-34 (Nov 2023). https://doi.org/10.1145/3624010, http://dx.doi.org/10.1145/3624010 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.244 + ], + "angle": 0, + "content": "49. Salem, A., Zhang, Y., Humbert, M., Berrang, P., Fritz, M., Backes, M.: Ml-leaks: Model and data independent membership inference attacks and defenses on machine learning models. In: Proceedings of the 26th Annual Network and Distributed System Security Symposium (NDSS) (2019) 1, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.244, + 0.787, + 0.284 + ], + "angle": 0, + "content": "50. Setlur, A., Eysenbach, B., Smith, V., Levine, S.: Adversarial unlearning: Reducing confidence along adversarial directions. Advances in Neural Information Processing Systems 35, 18556-18570 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.787, + 0.311 + ], + "angle": 0, + "content": "51. Shaik, T.B., Tao, X., Xie, H., Li, L., Zhu, X., Li, Q.: Exploring the landscape of machine unlearning: A comprehensive survey and taxonomy. CoRR (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.787, + 0.352 + ], + "angle": 0, + "content": "52. Shokri, R., Stronati, M., Song, C., Shmatikov, V.: Membership inference attacks against machine learning models. 2017 IEEE Symposium on Security and Privacy (SP) pp. 3-18 (2017) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.787, + 0.379 + ], + "angle": 0, + "content": "53. Sommer, D.M., Song, L., Wagh, S., Mittal, P.: Towards probabilistic verification of machine unlearning. arXiv preprint arXiv:2003.04247 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.405 + ], + "angle": 0, + "content": "54. Song, L., Mittal, P.: Systematic evaluation of privacy risks of machine learning models (2020) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.405, + 0.787, + 0.459 + ], + "angle": 0, + "content": "55. Song, L., Shokri, R., Mittal, P.: Privacy risks of securing machine learning models against adversarial examples. Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security (2019), https://api-semanticscholar.org/CorpusID:165163934 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.787, + 0.5 + ], + "angle": 0, + "content": "56. Tarun, A.K., Chandawat, V.S., Mandal, M., Kankanhalli, M.: Fast yet effective machine unlearning. IEEE Transactions on Neural Networks and Learning Systems (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.541 + ], + "angle": 0, + "content": "57. Tiwary, P., Guha, A., Panda, S., et al.: Adapt then unlearn: Exploiting parameter space semantics for unlearning in generative adversarial networks. arXiv preprint arXiv:2309.14054 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.541, + 0.69, + 0.554 + ], + "angle": 0, + "content": "58. Tukey, J.W.: Exploratory Data Analysis. Addison-Wesley (1977) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.554, + 0.787, + 0.609 + ], + "angle": 0, + "content": "59. Wang, K.C., FU, Y., Li, K., Khisti, A., Zemel, R., Makhzani, A.: Variational model inversion attacks. Advances in Neural Information Processing Systems 34, 9706-9719 (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/50a074e6a8da4662ae0a29edde722179-Paper.pdf 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.787, + 0.65 + ], + "angle": 0, + "content": "60. Wu, G., Hashemi, M., Srinivasa, C.: Puma: Performance unchanged model augmentation for training data removal. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 8675-8682 (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.787, + 0.69 + ], + "angle": 0, + "content": "61. Wu, Y., Dobriban, E., Davidson, S.: Deltagrad: Rapid retraining of machine learning models. In: International Conference on Machine Learning. pp. 10355-10366. PMLR (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.69, + 0.787, + 0.718 + ], + "angle": 0, + "content": "62. Xu, J., Wu, Z., Wang, C., Jia, X.: Machine unlearning: Solutions and challenges. arXiv preprint arXiv:2308.07061 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.784 + ], + "angle": 0, + "content": "63. Yang, Z., Zhang, J., Chang, E.C., Liang, Z.: Neural network inversion in adversarial setting via background knowledge alignment. Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security p. 225-240 (2019). https://doi.org/10.1145/3319535.3354261, https://doi.org/10.1145/3319535.3354261 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "64. Yeom, S., Giacomelli, I., Fredrikson, M., Jha, S.: Privacy risk in machine learning: Analyzing the connection to overfitting. 2018 IEEE 31st Computer Security Foundations Symposium (CSF) pp. 268-282 (2017), https://api-semanticscholar.org/CorpusID:2656445 1" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "SCAR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "65. Yin, Z., Xing, E., Shen, Z.: Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. Advances in Neural Information Processing Systems 36 (2024) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.258 + ], + "angle": 0, + "content": "66. Zhang, Y., Jia, R., Pei, H., Wang, W., Li, B., Song, D.: The secret revealer: Generative model-inversion attacks against deep neural networks. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 250-258 (jun 2020). https://doi.org/10.1109/CVPR42600.2020.00033, https://doi.ieeecomputersociety.org/10.1109/CVPR42600.2020.00033 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.315 + ], + "angle": 0, + "content": "67. Zhao, X., Zhang, W., Xiao, X., Lim, B.: Exploiting explanations for model inversion attacks. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 662-672 (oct 2021). https://doi.org/10.1109/ICCV48922.2021.00072, https://doi.ieeeccomputersociety.org/10.1109/ICCV48922.2021.00072 1" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.315 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Is Retain Set All You Need in Machine Unlearning_ Restoring Performance of Unlearned Models with Out-Of-Distribution Images/ef16a8f0-48f4-4778-ba31-599cb1e6436c_origin.pdf b/2024/Is Retain Set All You Need in Machine Unlearning_ Restoring Performance of Unlearned Models with Out-Of-Distribution Images/ef16a8f0-48f4-4778-ba31-599cb1e6436c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..74dea8856c75948e4ed2b0a151478d921a92d80b --- /dev/null +++ b/2024/Is Retain Set All You Need in Machine Unlearning_ Restoring Performance of Unlearned Models with Out-Of-Distribution Images/ef16a8f0-48f4-4778-ba31-599cb1e6436c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6904b4a190fbc39c5b27913a30e30ce3969da68a83d023a7a3e447613d607607 +size 3762064 diff --git a/2024/Is Retain Set All You Need in Machine Unlearning_ Restoring Performance of Unlearned Models with Out-Of-Distribution Images/full.md b/2024/Is Retain Set All You Need in Machine Unlearning_ Restoring Performance of Unlearned Models with Out-Of-Distribution Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ef819eb0dc6b1d163a7758b2b521ccf722176401 --- /dev/null +++ b/2024/Is Retain Set All You Need in Machine Unlearning_ Restoring Performance of Unlearned Models with Out-Of-Distribution Images/full.md @@ -0,0 +1,298 @@ +# Is Retain Set All You Need in Machine Unlearning? Restoring Performance of Unlearned Models with Out-Of-Distribution Images + +Jacopo Bonato1,2, Marco Cotogni2*, and Luigi Sabetta1,2 + +1 Leonardo Labs, Rome, Italy, + +2 These authors contributed equally + +{jacopo.bonato,luigi.sabetta}.ext@leonardo.com,cotogni@proton.me + +Abstract. In this paper, we introduce Selective-distillation for Class and Architecture-agnostic unleaRning (SCAR), a novel approximate unlearning method. SCAR efficiently eliminates specific information while preserving the model's test accuracy without using a retain set, which is a key component in state-of-the-art approximate unlearning algorithms. Our approach utilizes a modified Mahalanobis distance to guide the unlearning of the feature vectors of the instances to be forgotten, aligning them to the nearest wrong class distribution. Moreover, we propose a distillation-trick mechanism that distills the knowledge of the original model into the unlearning model with out-of-distribution images for retaining the original model's test performance without using any retain set. Importantly, we propose a self-forget version of SCAR that unlearns without having access to the forget set. We experimentally verified the effectiveness of our method, on three public datasets, comparing it with state-of-the-art methods. Our method obtains performance higher than methods that operate without the retain set and comparable w.r.t the best methods that rely on the retain set. + +Keywords: Machine Unlearning $\cdot$ Retain-Free $\cdot$ Data Removal + +# 1 Introduction + +Over the past decades, deep learning algorithms have grown significantly in applications like image classification, detection, and segmentation. For optimal performance, these models learn from big training datasets and update their weights to capture intrinsic patterns. Despite their effectiveness, many ethical concerns such as potential biases affecting certain demographics [12] and privacy like unauthorized use of sensitive or personal data [38,48] have been raised. Additionally, methods like model inversion [5,16,29,59,63,66,67], and membership inference attacks [49, 52, 55, 64] raised many concerns about the potential revelation of training data characteristics or can reveal particular data utilization during training. These vulnerabilities highlight the critical necessity for + +implementing safeguards to secure sensitive information within machine learning models. This imperative is fundamental in light of the privacy-preserving regulations that have been put into effect ( e.g. the European Union's General Data Protection Regulation [39] and the California Consumer Privacy Act [45]). + +Machine Unlearning [41, 44, 51, 62] has emerged as a solution to these privacy and ethical challenges, aiming to remove specific information content from algorithms without degrading their performance. Unlearning algorithms can be classified into 3 main groups: model intrinsic, data-driven, and model agnostic. The characteristic of model-intrinsic methods is to change the deep neural network (DNN) architecture or add specific weights to induce the unlearning. The data-driven group is characterized by the usage of data to induce forgetting, and it can be roughly divided into methods that store optimization information during training (e.g. gradient updates) and methods that involve data manipulation. Similarly to this latter group, model-agnostic solutions do not require access to specific optimization-related information and operate directly on the unaltered trained model. Nonetheless, a common practice among these algorithms is to maintain or restore the model performance using the residual portion of the training dataset, known as retain set. Using the retain set to recalibrate models post-unlearning represents a viable strategy, however, it also introduces practical challenges. In scenarios where privacy concerns dictate, only the forget set —images to forget— might be accessible, leaving no retain set for performance restoration. This poses a significant hurdle for models designed to undergo a performance restoration phase after unlearning, potentially compromising their utility. Moreover, when large-scale datasets such as ImageNet or ImageNet-21K are involved and the forget set constitutes a minor fraction, the bulk of the data is the retain set. This disproportionality can significantly inflate the time required to restore the model to its original efficacy. Furthermore, there are even more intricate situations where also the forget set is unavailable, such as when the unlearning task entails forgetting an entire class with access only to its ID. + +This paper presents a novel, model-agnostic unlearning algorithm, Selective-distillation for Class and Architecture-agnostic unleaRning (SCAR), which leverages metric learning and knowledge distillation [26] to efficiently remove targeted information and maintain model accuracy, notably without relying on a retain set. Specifically, we use the Mahalanobis distance to shift the feature vectors of the instances to forget toward the closest distribution of samples of other classes. This distance includes important information regarding the original dataset distribution of samples resulting in an efficient metric learning-based unlearning strategy. Concurrently, distilling the knowledge of the original model into the unlearning model using out-of-distribution (OOD) images, allows SCAR to maintain the performance of the original model on the test set. The main contributions of this work include: (i) SCAR, a novel model-agnostic unlearning algorithm that achieves competitive unlearning performance without using retain data. (ii) A unique self-forget mechanism in class removal scenarios that operates without direct access to the forget set. (iii) Comprehensive analyses demonstrating SCAR's efficacy across different datasets and architectures in + +class-removal and homogeneous-removal scenarios $(iv)$ Experimental evidence showing SCAR's comparable or superior performance respectively to traditional unlearning methods or state-of-the-art (sota) techniques that do not utilize a retain set. Code is available at https://github.com/jbonato1/SCAR + +# 2 Related Works + +Machine Unlearning algorithms can be classified into three main categories: data-driven, model-intrinsic, and model-agnostic unlearning algorithms [44]. + +Data-driven. The unlearning algorithms in this category employ a data-driven approach to remove knowledge about specific image instances. For example, methods that involve data manipulation, such as transformations or attacks, fall within this group of algorithms [50, 53]. However, this category is not limited to such models: it encompasses a broader range of approaches. In [19, 61] the authors propose similar methods for reverting the gradient updates for the instances of the classes to be removed. Despite their effectiveness, these methods require storing gradient updates for all the batches in the training set. Similarly, in [42], the authors propose two gradient-descent-based approaches for the effective unlearning of convex models. In the first, Gaussian noise is used to perturb the gradient updates, while the second also incorporates reservoir sampling. In [60], the authors modeled the impact of each training point on the model's training process and then compensated for the impact of removing data from the model's knowledge by re-weighting the gradient updates. In [14], the authors employed a strategy of recursively assigning an incorrect class to samples designated for forgetting until a membership inference attack no longer recognizes those samples as part of the training dataset. In [3], the authors segmented the dataset into n parts and trained n models on each segment; during inference, predictions from these models are aggregated. Only the model containing those samples is retrained when a removal request is received. Moreover, novel retain-set-free algorithms such as Random Labels [23] and Negative Gradient [17] have been introduced. In the former approach, random labels are assigned to samples designated for forgetting to calculate the cross-entropy, while the latter employs the negative gradient obtained from these samples to optimize the loss function. + +Model-intrinsic. Model-intrinsic approaches to machine unlearning consist of algorithms that depend on a specific architecture. In these cases, the unlearning process is facilitated by sophisticated mechanisms tailored to function within that particular architecture. Depending on the task to be performed several architectures can be involved. For instance in [17, 18] the authors proposed two different unlearning approaches for CNNs. In the former, the authors "scrubbed" from the weights of the model the knowledge about instances of images through the computation of Fisher's information matrix. Similarly in the latter, the authors propose a CNN-based unlearning algorithm, inspired by the neural tangent kernel, that models the weights dynamics and can remove from the information about forget data. In [35], the authors proposed a two-phase algorithm for the approximate unlearning of CNNs. During the first phase, called ERM, the model + +![](images/508aedf6204bab9f065a87c3edff878b7e2b074de67306ab54a77c6a48f2ed85.jpg) +A + +$$ +d _ {J S} (\Phi_ {\theta} (x _ {j}) | | \Phi_ {\theta} ^ {U} (x _ {j})) +$$ + +![](images/a007c1a9022ea8fed443e9a79f0cbc3c8fb150f9a7f6601305fe512c9e6874a3.jpg) +B + +![](images/812df6d6f2395bc5d161cd48572da11d0e04f976d1fb544e91e6ffdc11a2c91c.jpg) + +$$ +d _ {M} (\Phi_ {\psi} ^ {U} (x _ {j}), Q _ {j} ^ {*}) +$$ + +![](images/97b2e9e4f74a8e5ddf4ecdafe5c1c2768a918b60e1bb6be0098d8b1ebcff378d.jpg) +C +Fig. 1: SCAR overview. A) SCAR scheme B) distillation-trick through J.S. divergence of the logits of $\varPhi_{\theta}(x_j)$ and $\varPhi_{\theta}^{U}(x_{j})$ . C) Representation of the feature vector distributions of samples for 3 classes before (left) and after (right) the unlearning process. Ellipses denote the $95\%$ confidence intervals of the distribution of samples. + +is trained on all the data while a binary mask is learned for each layer of the CNN for each class of the dataset. Then during the second phase, called KTP, the knowledge about the classes of the forget set is removed using the previously learned masks. In [20], the authors propose a certified algorithm for removing knowledge about classes on linear classifiers. This method applied a Newton step on the model's weights to significantly reduce the influence of the deleted data point, with the residual error decreasing quadratically as the training set size increases. Then, for certifying the data removal and to prevent the extraction of information from the minimal residual the training loss is perturbed. Recently, additional unlearning approaches have been developed for dealing with specific models on different learning paradigms such as Multimodal learning [7, 46] and Adversarial Learning [37, 57]. + +Model-agnostic. This group of algorithms includes universally applicable methods across different architectures. Their characteristic is the ability to intervene in the feature/activation space to selectively erase data associated with the forget set. In [32], the authors used knowledge distillation to differentiate the logits of the forget set from those of the retain set and erase the knowledge of these examples. A similar strategy was proposed in [30], where distillation was combined with contrastive labeling to achieve efficient unlearning. In [4] the authors proposed two methods for unlearning based on the kinematics of the decision boundary. The methods change the labels of the examples of the forget set to the closest wrong class and then fine-tune the model with the reassigned samples. Similarly, In [9] the authors proposed a metric learning-based approach that associates samples from the forget set to the closest centroid in the feature space. Through this procedure, the model preserves the feature space organization. In [28] the authors proposed a model sparsification via weight pruning for unlearning data samples. In [56], the authors proposed to learn a matrix + +noise associated with the classes to be forgotten to remove their influence on the model's weights followed by a repair step used to recover knowledge about retain samples. + +Our approach stands out as a model-agnostic approximate unlearning algorithm and partially aligns with Boundary Unlearning [4] or DUCK [9] that manipulates respectively forget samples labels or forget samples feature vectors to induce unlearning. However, while these methods adjust samples towards incorrect directions based on the nearest wrong class or centroid in feature space, our solution focuses on using the distributions of features to efficiently reorganize the positions of forget samples in the feature space. This enables us to utilize a more precise and efficient mechanism, grounded in the Mahalanobis distance. Nonetheless, SCAR does not rely on any images from the retain set to restore performance during the unlearning process whereas model-agnostic methods like DUCK [9], Boundary Unlearning [4], and SCRUB [32] partially or fully depend on it. SCAR, thanks to an efficient knowledge preservation method based on distillation and OOD data, can operate without accessing training data. + +# 3 Methods + +# 3.1 Preliminaries + +Consider a dataset $\{x_i, y_i\}_{i=1}^N$ consisting of N images $x_i$ each associated with a label $y_i \in \{0, \dots, K\}$ where $K$ is the number of classes in the dataset. A model $\varPhi_\theta$ is trained on the training subset of the dataset $\mathcal{D}$ to predict the corresponding label for a given image, that is, $\hat{y}_i = \varPhi_\theta(x_i)$ . This model is then evaluated on the test subset $\mathcal{D}^t$ . The objective of a machine unlearning algorithm is to derive a model, denoted as $\varPhi_\theta^U$ , from which information about a specific subset of data, known as the forget set $\mathcal{D}_f$ , is removed from the original model's parameters $\theta$ . Concurrently, the algorithm aims to preserve the model's performance on the retain set $\mathcal{D}_r$ , such that it is comparable to that of a model trained exclusively on the retain set. In this paper, we consider two scenarios: Class-Removal (CR) and Homogeneous Removal (HR). + +In the CR scenario, the objective is to remove from the model's weights the information about the instances of a class $C$ among the $K$ available. The two subsets $\mathcal{D}$ and $\mathcal{D}^t$ are split in forget $(\mathcal{D}_f, \mathcal{D}_f^t)$ and retain $(\mathcal{D}_r, \mathcal{D}_r^t)$ sets. The forget set contains all the images, from the class to be removed $\mathcal{D}_f = \{x_i^f, y_i^f\}_{i=1}^{N^f}$ with $y_i^f \in \{C\}$ while the retain set contains all the remaining images $\mathcal{D}_r = \{x_i^r, y_i^r\}_{i=1}^{N^r}$ , with $y_i^r \in \{0, \dots, K-1\}$ ; the same definition stands also for $\mathcal{D}^t$ . The desiderata in this scenario is to maintain the original accuracy on the $\mathcal{D}_r^t$ while minimizing the one on the $\mathcal{D}_f^t$ . In the HR scenario, the training dataset $\mathcal{D}$ is divided into retain $\mathcal{D}_r$ and forget $\mathcal{D}_f$ sets, while the test set $\mathcal{D}^t$ remains unchanged. In this case, the images constituting $\mathcal{D}_f$ are chosen from all the classes. Here, the goal is to eliminate the model's knowledge of specific training instances in $\mathcal{D}_f$ without compromising its performance on $\mathcal{D}_r$ . Following the application of the unlearning algorithm, the model should no longer differentiate + +between data from $\mathcal{D}_f$ and $\mathcal{D}^t$ or in other words it interprets $\mathcal{D}_f$ samples as never-seen-data. This indistinguishability has also to be reflected in the $\mathcal{D}_f$ accuracy which has to be approximately similar to $\mathcal{D}^t$ accuracy. This second scenario is more challenging compared to CR since $\mathcal{D}_r$ and $\mathcal{D}_f$ include images from the same classes and the model should maintain its ability to generalize across all classes. + +# 3.2 SCAR + +In SCAR, the unlearning process is based on metric learning and involves directly modifying the feature vector projection of the sample to be forgotten within the feature space. Additionally, we have devised a regularization technique that enables the DNN to preserve the original model information through distillation without access to the retain dataset. By leveraging an OOD dataset, SCAR enables the transfer of knowledge from the original model to the unlearning model, thereby ensuring that classification performance on test samples remains consistent throughout the unlearning process. Initially, we present the unlearning strategy of SCAR, followed by a deeper exploration of the distillation with OOD data and the Metric Learning mechanism. Then we introduce a variant of SCAR, called SCAR self-forget, that works without accessing the forget set. + +Unlearning strategy. The unlearning strategy is composed of 2 synergistic mechanisms, Metric Learning and Distillation-Trick (Fig. 1 A). In the following sections each DNN $\varPhi_{\theta}$ will be represented as a sequence of a backbone $\varPhi_{\psi}$ and a final fully-connected layer $\varPhi_{\pi}$ where $\theta, \psi, \pi$ are the corresponding weight. + +During the training phase of the original model $\varPhi_{\theta}$ , for each class $i^{th}$ the distribution of feature vectors $Q_{i} = \{\varPhi_{\psi}(x_{j})\}_{j}$ is prototyped [10, 22, 47] by its mean $\mu_{i}$ and covariance matrix $\hat{S}_Q$ (where $\varPhi_{\psi}(x_j)$ represent a feature vector), and accurately stored preventing any data leakage. Hence, given a forget sample $(x_{j},y_{j} = k)$ belonging to the $k^{th}$ class, SCAR selects the closest distribution $Q_{j}^{*}$ among the set of distribution $Q_{i}$ with $i\in \{0,\dots,K\}$ and $i\neq k$ to the feature vector $\varPhi_{\psi}^{U}(x_{j})$ . + +$$ +Q _ {j} ^ {*} = \underset {Q _ {i}} {\operatorname {a r g m i n}} d _ {M} \left(\Phi_ {\psi} ^ {U} \left(x _ {j}\right), Q _ {i}\right) \tag {1} +$$ + +Hence, during the unlearning procedure, the algorithm, through the Metric Learning mechanism, minimizes the Mahalanobis distance between the forget samples feature vector $\varPhi_{\psi}^{U}(x_{j})$ and their corresponding selected distribution $Q_{j}^{*}$ . The overall forget loss is written in eq. 2: + +$$ +\mathcal {L} _ {M} = \frac {1}{N _ {f , \text {b a t c h}}} \sum_ {j = 0} ^ {N _ {f, \text {b a t c h}} - 1} d _ {M} \left(\Phi_ {\psi} ^ {U} \left(x _ {j}\right), Q _ {j} ^ {*}\right). \tag {2} +$$ + +where $N_{f,\mathrm{batch}}$ is the batch size of forget samples. + +Since in our proposed scenario, the $\mathcal{D}_r$ dataset is no longer accessible, the algorithm makes use of the Distillation-Trick mechanism to retain past knowledge. We used the Jensen-Shannon divergence between the original model $\varPhi_{\theta}$ + +![](images/039f2d5e6cab050938a51226004522e67a253b762229ec9546be3223b50a1b15.jpg) +Fig.2: A-B) TSNE plots of feature vectors for the first 20 classes in CIFAR100 (A, $\mathcal{D}$ ) and the surrogate dataset subset Imagenet (B, $\mathcal{D}^{\mathrm{sur}}$ ) + +![](images/6641b2536b84e30313f6f89d685f791ede64b203dd34e4f08083a134ceb5d615.jpg) + +and the unlearning model $\varPhi_{\theta}^{U}$ (see par. Distillation-Trick, eq. 5) + +$$ +\mathcal {L} _ {T D} = \frac {1}{N _ {r , \text {b a t c h}}} \sum_ {j = 0} ^ {N _ {r, \text {b a t c h}} - 1} d _ {J S} \left(\Phi_ {\theta} ^ {U} \left(x _ {j}\right) \| \Phi_ {\theta} \left(x _ {j}\right)\right) \tag {3} +$$ + +where $N_{r,\mathrm{batch}}$ is the batch size of the surrogate retain samples. The overall loss combines the forget and retain contribution in eq. 2 and 3. + +$$ +\mathcal {L} = \lambda_ {1} \mathcal {L} _ {M} + \lambda_ {2} \mathcal {L} _ {T D} \tag {4} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are hyperparameters. Instead of selecting a priori the number of unlearning epochs, SCAR computes the $\mathcal{A}_f$ at each epoch and it stops when $\mathcal{A}_r$ is lower or equal to a threshold $\epsilon$ . In CR $\epsilon = 0$ because the objective is to remove completely the knowledge of the class to remove. In HR $\epsilon = \mathcal{A}^t$ because the forget test accuracy has to be close to the accuracy of the original model on the test set. When this condition is not reached the unlearning process stops when it reaches the maximum number of epochs allowed (epochs_max). + +Distillation-trick. Numerous instances have been documented in literature wherein DNNs exhibit high confidence predictions [21], even when presented with data significantly distant from their training set. Examples include their performance on fooling images [43], OOD samples [25], or within medical diagnostic tasks [34]. For instance, the t-SNE projection of feature vectors from a DNN, both from the original training dataset and an external OOD dataset, exhibited clustering at identical positions within the feature space (Fig. 2 A-B). In our proposed algorithm, we leverage this characteristic behavior of DNNs on OOD datasets as a form of regularization during the unlearning process, which we term "distillation-trick". This approach involves distilling the knowledge from the original frozen model (referred to as the teacher model) into the unlearning model using an external dataset acting as a surrogate dataset, denoted as $\mathcal{D}^{\mathrm{sur}}$ [13] (Fig. 1 B). The distillation-trick leverages the Jensen-Shannon divergence between $\varPhi_{\theta}^{U}(x_{j})$ and $\varPhi_{\theta}(x_{j})$ to retain the knowledge of the original model, even as the unlearning model undergoes training on $\mathcal{D}^{\mathrm{sur}}$ (omitting the dependency on $x_{j}$ for brevity): + +$$ +d _ {J S} \left(\Phi_ {\theta} ^ {U} \| \Phi_ {\theta} / T\right) = \frac {1}{2} D _ {K L} \left(\Phi_ {\theta} ^ {U} \| \Phi_ {\theta} / T\right) + \frac {1}{2} D _ {K L} \left(\Phi_ {\theta} / T \| \Phi_ {\theta} ^ {U}\right) \tag {5} +$$ + +where $\mathrm{T}$ is a scaling temperature. To validate this, we trained $\varPhi_{\theta}$ using distillation-trick on a surrogate dataset. Our experiments demonstrate that the accuracy of + +$\varPhi_{\theta}$ on $\mathcal{D}_t$ remains consistent across epochs and comparable to that of $\varPhi_{\theta}^{U}$ (refer to Sec. 1 of the Supp. Materials). + +Metric learning. In the context of unlearning problems, metric learning serves as a pivotal concept aimed at modifying the feature vectors of forget samples, thereby altering their class predictions. The prevalent strategies involve shifting the feature vectors of forget samples towards the nearest incorrect class centroids [9] or directly shifting the logits towards the nearest incorrect class [4]. In our algorithm, we propose the application of metric learning by adjusting the feature vectors of forget samples toward the nearest distribution of feature vectors of a different class (Fig. 1 C). Thanks to the Mahalanobis distance, which gauges the distance between the forget set samples and the whole distributions of feature vectors, SCAR can integrate more comprehensive information about the distributions through the covariance matrix. This approach contrasts with simpler methods that represent distributions solely using centroids (see Sec.4.4 for a comparison with other distances based on centroids). Therefore, given the set of distribution of feature vectors for each class $Q = \{Q_{i}\}_{i = 0}^{K}$ the Mahalanobis distance is defined in eq. 6. + +$$ +d _ {M} (\varPhi_ {\psi} ^ {U} (x _ {j}), Q _ {i}) = \sqrt {(\varPhi_ {\psi} ^ {U} (x _ {j}) - \mu_ {Q}) ^ {T} \hat {S} _ {Q} ^ {- 1} (\varPhi_ {\psi} ^ {U} (x _ {j}) - \mu_ {Q})} \tag {6} +$$ + +where $\varPhi_{\psi}^{U}(x_{j})$ is the feature vector of $x_{j}$ (i.e. the output of the backbone of $\varPhi^U$ ), $\mu_Q$ and $\hat{S}_Q$ are the mean and the normalized covariance matrix obtained from the distribution $Q_{i}$ . Each class's covariance matrix $S_{Q}$ will exhibit varying levels of scaling and variances across different dimensions. Consequently, the Mahalanobis distances of features from distinct classes will be subjected to differing scaling factors. Therefore, to standardize the covariance matrices per class effectively, we conduct correlation matrix normalization on all the covariance matrices (eq. 7): + +$$ +\hat {S} _ {Q} (i, j) = \frac {S _ {Q} (i , j)}{\sigma_ {Q} (i) \sigma_ {Q} (j)}, \quad \sigma_ {Q} (j) = \sqrt {S _ {Q} (j , j)}, \quad \sigma_ {Q} (i) = \sqrt {S _ {Q} (i , i)}. \qquad (7) +$$ + +Since the number of samples per class in public datasets ( $\approx 500$ for CIFAR100 and TinyImagenet) is generally lower than the number of dimensions in the feature space (e.g. $\mathrm{N} = 512$ for resnet18), we use a covariance shrinkage method to get a full-rank matrix [6], $S_{Q}^{s} = S_{Q} + \gamma_{0}V_{0}\mathbb{I} + \gamma_{1}V_{1}(1 - \mathbb{I})$ where $V_{0}$ is the average diagonal variance, $V_{2}$ is the average off-diagonal covariance of $S_{Q}$ and $\mathbb{I}$ is the identity matrix. Furthermore, we added Tukey's normalization [58] to the feature vectors $\varPhi_{\psi}^{U}(x)^{\delta}$ to obtain approximately Gaussian features, where $\delta$ is a hyperparameter to decide the degree of transformation of the distribution. + +# 3.3 SCAR Self-forget + +Additionally, we present an adaptation of SCAR tailored for a more demanding CR scenario, wherein access to the forget set is restricted, and only the class ID + +slated for removal is provided. We term this procedure SCAR self-forget (Fig. 3). In this version, both the forgetting mechanism and the distillation-trick rely on the surrogate dataset. Given $K_{\mathrm{ret}}$ , the set of retained classes, and $K_{\mathrm{fgt}}$ , the set of forget classes, SCAR initially employs the original model $\varPhi_{\theta}$ to classify $\mathcal{D}^{\mathrm{sur}}$ . + +![](images/9bff3dda2b4e3e380f9995e6a43b22677bd66a6be4360ab0533c1aae20a20145.jpg) +Fig. 3: Scheme of SCAR self-forget in CR. The surrogate dataset $\mathcal{D}^{\mathrm{sur}}$ supplies during the unlearning procedure both the surrogates $\mathcal{D}_r^{\mathrm{sur}}$ and $\mathcal{D}_f^{\mathrm{sur}}$ . + +Subsequently, SCAR splits the samples in $\mathcal{D}^{\mathrm{sur}}$ into $\mathcal{D}_r^{\mathrm{sur}}$ and $\mathcal{D}_f^{\mathrm{sur}}$ if respectively the predicted class $\hat{y}$ belongs to $K_{\mathrm{ret}}$ , and in $\mathcal{D}_f^{\mathrm{sur}}$ if $\hat{y}$ is within $K_{\mathrm{fgt}}$ . Thus, $\mathcal{D}_f^{\mathrm{sur}}$ serves as a surrogate for $\mathcal{D}_f$ , and SCAR is optimized using the loss function in Eq. 4. This configuration poses an extreme challenge as it lacks both retain and forget data during the unlearning process. SCAR self-forget marks the first attempt to address the unlearning problem within the Class-Removal (CR) scenario without necessitating training or forget data, nor the retention of specific in + +formation during training such as gradient updates. + +# 4 Experimental Results + +We evaluated SCAR against a range of sota methods and baselines on three datasets CIFAR10 [31], CIFAR100 [31], and TinyImageNet [33]. CIFAR-10 and CIFAR-100 are comprised of 60,000 images of $32 \times 32$ pixels each (50,000 images for train and 10000 for test), with CIFAR-10 divided into 10 classes, and CIFAR-100 categorized into 100 classes. TinyImageNet consists of 110,000 images of $64 \times 64$ pixels each, divided into 200 classes (100,000 for train and 10,000 for test). The considered sota methods include DUCK [9], Boundary Shrink [4], Boundary Expanding [4], SCRUB [32], L1-Sparse [28], Bad Teacher [8], ERM-KTP [35], Negative Gradient [17], and Random Labels [23], alongside traditional approaches such as Retraining and Fine-Tuning. While the latter two are computationally demanding as they necessitate access to the $\mathcal{D}_r$ , they serve as critical benchmarks. Retraining is by definition the upper bound for unlearning, removing the $\mathcal{D}_f$ from the training process entirely. Fine-Tuning, conversely, adjusts the pre-trained model weights using only the $\mathcal{D}_r$ . Importantly, changing the number of epochs can substantially affect the finetuning performance: few epochs result in an efficient unlearning whereas if the model is finetuned for many epochs it is substantially retrained. For this reason, the original model is finetuned for a sufficient number of epochs ( $<$ epochs of retrain) that guarantees to unlearn $\mathcal{D}_f$ (as outlined in [4,35]). Unfortunately, for this reason, Fine-Tuning is extremely time-consuming and inefficient compared to other sota methods [4,9,35]. All + +Table 1: Comparison between SCAR performance and different surrogate dataset in CR scenario. The metrics are reported as mean ± std over ten runs. + +
Surrogate dataset DsurSizeCIFAR100TinyImagenet
ArAftAUSp-valArAftAUSp-val
Random Images2*10372.20(01.90)02.60(02.17)0.922(0.027)<10-360.69(01.80)01.20(01.40)0.912(0.022)<10-3
Imagenet subset10472.93(01.78)02.00(02.00)0.935(0.025)<10-362.99(01.39)00.60(01.37)0.940(0.019)<10-3
COCO subset10471.61(01.92)00.84(00.43)0.933(0.023)<10-362.04(01.41)01.00(01.41)0.927(0.019)<10-3
Imagenet distilled10472.13(02.28)02.51(01.62)0.923(0.023)<10-362.36(01.56)01.03(01.41)0.930(0.020)<10-3
Gaussian Noise10416.62(08.64)00.70(00.67)0.388(0.086)<10-317.13(06.08)01.20(01.93)0.482(0.061)<10-3
+ +methods utilized resnet18 [24] as the neural architecture. The details about the hyperparameters are reported in the Supp. Material, Sec. 2, "Reproducibility". + +We established two experimental frameworks, CR and HR, based on the scenarios outlined in Sec. 3.1. For CR, we applied a constant seed, dividing each dataset into $\mathcal{D}_r$ and $\mathcal{D}_f$ with distinct class distributions. The $\mathcal{D}_f$ included instances from a single class, while the $\mathcal{D}_r$ comprised the remaining classes. This setup was replicated across ten splits to calculate average metrics and standard deviations. In the HR scenario, we selected 10 different seeds for dataset partitioning, ensuring a 90:10 split between $\mathcal{D}_r$ and $\mathcal{D}_f$ , with class overlaps but unique instances. We reported the mean and standard deviation of the results across these seeds. Further details are available in the Sec. 2 of the Supp. Material. + +For the CR scenario, the metrics considered are the accuracies on retain $(\mathcal{A}_r^t)$ and forget $(\mathcal{A}_f^t)$ test sets, to be maximized and minimized, respectively, and the Adaptive Unlearning Score (AUS) [9]. This final score facilitates the comparison among various methods by encapsulating forget and retain test set performances into a single metric. It captures a delicate balance between two essential objectives: preserving high test accuracy while effectively tackling the unlearning task (see Supp. Material Sec. 2). For the HR scenario, the metrics included are accuracy on $\mathcal{D}_f$ ( $\mathcal{A}_f$ ), overall test set accuracy $(\mathcal{A}^t)$ , AUS, and the success of a membership inference attack (MIA) [9, 49, 54], with the optimal MIA value being $\mathcal{F}_1 = 0.5$ . Specifically the number of forget and test samples are balanced and $\mathcal{F}_1 = 0.5$ (chance level) means the impossibility to distinguish the membership of the forget data. The Sec. 2 of the Supp. Material provides details about hyperparameters for MIA tuning. + +# 4.1 Impact of different Datasets as $\mathcal{D}^{\mathrm{sur}}$ + +A critical role in SCAR is played by the distillation-trick mechanism. It allows SCAR to maintain acquired information about the retain set even without accessing it during the unlearning. The distillation-trick works using an OOD dataset, and investigating how different datasets can or cannot affect the efficiency of SCAR is of utmost importance. For this reason, following [2, 15, 27], we selected 4 different OOD external public datasets characterized by different distributions of pixels, number of images, resolution, and number of classes (a random subset of $10\mathrm{K}$ images of Imagenet1K [11], a random subset of $10\mathrm{K}$ images of COCO [36], a set of $2\mathrm{K}$ public natural images [1], and a set of $10\mathrm{K}$ images distilled with [65] from Imagenet1K). We verified through the Kolmogorov-Smirnov + +Table 2: Performance comparison between SCAR and sota methods in CR scenario for CIFAR10, CIFAR100, and TinyImagenet datasets. The metrics are reported as mean ± std over ten runs. Results marked with $\dagger$ are taken from [9]. Columns “ $\mathcal{D}_r$ free” and “ $\mathcal{D}_f$ free” indicate whether the method was (✓) or not (✗), Retain-Free or Forget-free, respectively. Green and red colors highlight respectively the best results between methods based on retain set and between retain free methods. + +
DrfreeDffreeCIFAR10CIFAR100TinyImagenet
AicAfjAUSAicAfjAUSAicAfjAUS
Original--88.64(00.63)88.34(00.62)0.531(0.005)77.55(00.11)77.50(02.80)0.563(0.009)68.26(00.08)64.60(15.64)0.607(0.058)
Retrained†--88.05(01.28)0.000(0.000)0.994(0.014)77.97(00.42)0.000(0.000)1.004(0.022)67.67(01.00)0.000(0.000)0.993(0.010)
Fine Tuning‡--87.93(01.14)0.000(0.000)0.993(0.013)77.31(02.17)0.000(0.000)0.998(0.022)67.89(00.24)0.000(0.000)0.994(0.010)
DUCK†XX85.53(01.37)0.000(0.000)0.969(0.015)71.57(02.08)0.010(0.176)0.931(0.026)61.29(01.13)0.020(0.063)0.927(0.013)
Boundary S.†XX83.81(02.29)13.64(03.57)0.840(0.543)55.07(11.85)03.54(02.26)0.749(0.116)55.98(03.33)04.25(02.12)0.840(0.036)
Boundary E.†XX82.36(02.39)13.34(03.21)0.827(0.032)55.18(11.98)03.50(02.32)0.750(0.117)55.92(03.45)02.60(02.50)0.853(0.040)
SCRUBXX87.88(01.18)0.000(0.000)0.992(0.013)77.29(00.25)02.00(05.35)0.977(0.051)68.15(00.25)01.20(03.79)0.986(0.037)
L1-SparseXX85.21(09.41)04.88(02.18)0.862(0.023)62.63(07.56)01.22(05.53)0.879(0.018)63.76(01.29)04.81(01.54)0.901(0.011)
ERM-KTP†XX80.31(-)00.00(-)0.917(-)58.96(-)00.00(-)0.814(-)49.45(-)00.00(-)0.811(-)
Bad TeacherXX88.89(00.87)02.37(05.86)0.982(0.057)77.19(00.19)06.00(10.09)0.940(0.089)64.09(00.78)11.20(09.04)0.862(0.070)
Neg.Grad.†✓X76.27(03.27)00.56(00.12)0.871(0.033)62.84(06.13)00.50(05.50)0.849(0.061)60.09(02.58)0.60(01.35)0.911(0.028)
Rand.Lab.†✓X65.46(07.59)00.83(00.44)0.762(0.076)55.31(07.06)00.40(00.70)0.774(0.070)43.29(10.10)01.20(01.03)0.740(10.100)
SCAR✓X87.71(01.62)00.97(00.24)0.981(0.017)72.93(01.78)02.00(02.00)0.935(0.025)62.99(01.39)0.60(01.37)0.940(0.019)
SCAR Self-Forget✓✓87.18(01.76)01.20(00.43)0.974(0.019)71.09(02.69)00.70(00.95)0.929(0.028)60.79(02.15)00.80(01.40)0.917(0.025)
+ +(KS) test [40] that the distributions of pixels of these surrogate datasets are different from the datasets used for benchmarking SCAR (Tab. 1 p-val columns). This is fundamental to assess the generality of SCAR and it does not rely on data leakage between datasets; this is especially true for tinyImagenet and the random subset ofImagenet1K where we observed the KS p-val is $< 0.001$ . Moreover, we also used a surrogate dataset made of Gaussian noise images where the semantic information was absent. We performed our experiments in the CR scenario using CIFAR100 as $\mathcal{D}$ . Both accuracies and AUS scores were statistically comparable across datasets (Tab. 1) even if theImagenet1K subset gives the overall highest results. Importantly, when in the surrogate dataset semantic information is removed (Gaussian Noise) SCAR performance is poor, suggesting how the surrogate dataset has to contain some spatial feature and patterns to trigger the DNN filters. Hence, the presence of semantic information content in images plays a pivotal role in the distillation-trick mechanism, but, at the same time, the differences in terms of kind of information (e.g. classes and distribution of pixels) do not affect this technique. For the rest of the paper results are presented using as a surrogate the OOD random subset ofImagenet1K. + +# 4.2 Comparison with sota methods + +We evaluated SCAR, and SCAR Self-Forget, within the CR scenario across three datasets (Tab. 2). Our findings indicate that our method surpasses Retain-Free approaches like Negative Gradient and Random Label in performance. Furthermore, SCAR's performance is on par with methods that leverage both forget and retain sets. Notably, SCAR's performance is lower than Fine-Tuning and within 1 std. from SCRUB in CIFAR100 and TinyImagenet, though these methods depend heavily on the $\mathcal{D}_r$ requiring several finetuning steps on it. Importantly, these outcomes are also achieved with SCAR Self-Forget, which remarkably does not utilize any data from the original training set. + +Table 3: Performance comparison between SCAR and sota methods in HR scenario for CIFAR10, CIFAR100, and TinyImagenet datasets. The metrics are reported as mean ± std over ten runs. Results marked with $\dagger$ are taken from [9]. Columns "D_r free" indicate whether the method was (✓) or not (X), Retain-Free. Green and red colors highlight respectively the best results between methods based on retain set and between retain free methods. + +
DfreeCIFAR10CIFAR100TinyImagenet
A1AfF1AUSA1AfF1AUSA1AfF1AUS
Original-88.54(00.25)99.49(00.08)54.62(00.80)0.900(0.004)77.23(00.53)99.63(00.10)61.80(00.59)0.814(0.006)68.08(00.39)84.83(00.44)54.06(01.48)0.854(0.007)
Retrained†-84.13(00.59)84.56(00.72)50.22(00.80)0.950(0.011)77.27(01.03)76.87(00.95)49.45(00.40)0.993(0.017)63.45(02.34)63.24(02.58)49.56(00.47)0.950(0.040)
Fine Tuning†-85.70(00.48)88.66(00.44)50.06(01.30)0.942(0.008)72.06(00.52)74.97(00.72)49.68(00.43)0.918(0.009)67.03(00.45)70.52(00.65)49.86(00.56)0.937(0.011)
DUCK†X85.96(00.64)86.05(00.55)50.26(00.49)0.972(0.011)74.74(00.74)75.51(02.01)50.66(00.64)0.965(0.022)62.01(00.79)64.43(00.76)49.87(00.73)0.916(0.014)
SCRBX78.02(07.80)82.53(10.80)50.59(00.89)0.854(0.132)73.33(00.85)92.46(01.50)56.45(01.48)0.804(0.014)68.52(00.40)83.14(00.53)53.89(00.96)0.875(0.008)
L1-SparseX93.05(00.53)99.06(00.65)52.21(00.95)0.927(0.011)52.26(01.65)66.89(02.34)53.46(00.94)0.690(0.024)63.76(01.19)84.30(01.04)53.29(00.73)0.794(0.032)
Bad TeacherX86.66(00.44)92.91(00.57)50.26(00.87)0.922(0.008)73.57(00.61)86.57(00.57)54.44(01.81)0.851(0.006)67.41(00.49)80.28(00.81)54.83(01.44)0.876(0.006)
Neg. Grad.†✓79.35(00.85)87.11(00.86)51.74(00.54)0.841(0.012)60.83(00.77)76.77(00.57)50.98(00.47)0.718(0.009)54.83(00.81)67.27(00.59)49.78(00.33)0.770(0.011)
Rand. Lab.†✓77.28(01.37)87.07(01.20)51.45(00.79)0.807(0.018)60.51(00.82)77.18(00.50)50.55(00.49)0.711(0.009)54.34(00.85)67.65(00.45)49.84(00.64)0.760(0.011)
SCAR✓86.14(00.47)88.40(00.31)50.07(00.59)0.953(0.008)73.23(00.74)75.63(00.54)49.77(00.28)0.934(0.011)61.35(00.69)66.51(00.44)49.85(00.19)0.886(0.011)
+ +Table 4: Results of the ablation study on CIFAR100 in CR and HR scenarios. The metrics for CR and HR are reported as mean ± std over ten runs. + +
LTDLMCIFAR100TinyImagenet
ArAfAUSArAfAUS
CRXX77.55(00.11)77.50(02.80)0.563(0.009)68.26(00.08)64.60(15.64)0.607(0.058)
✓X72.09(02.72)40.00(13.36)0.675(0.067)63.14(01.77)29.80(12.09)0.730(0.069)
X✓66.90(05.03)02.60(02.27)0.871(0.053)49.72(02.54)00.80(01.40)0.807(0.028)
✓✓72.93(01.78)02.00(02.00)0.935(0.025)62.99(01.39)00.60(01.37)0.940(0.019)
AtAfAUSAtAfAUS
HRXX77.39(00.46)99.64(00.09)0.817(0.005)68.08(00.39)84.83(00.44)0.854(0.007)
✓X76.53(00.41)99.18(00.14)0.807(0.005)64.84(00.60)80.51(00.51)0.835(0.009)
X✓18.83(01.06)19.69(01.19)0.409(0.013)05.88(02.94)05.51(02.63)0.376(0.033)
✓✓73.23(00.74)75.63(00.54)0.934(0.011)61.35(00.69)66.51(00.44)0.886(0.011)
+ +We also extended our analysis to the HR scenario (Tab. 3). Importantly, the Self-Forget variant of our method is not applicable in HR, as the $\mathcal{D}_f$ is essential to identify the specific instance to forget. In this more demanding context, SCAR surpasses all the performance of Retain-Free approaches. Between the methods that use the $\mathcal{D}_r$ only DUCK in CIFAR100 (within 1 std.) and TinyImagenet and Fine-Tuning in TinyImagenet achieve superior results. + +Notably, our method outperforms the Retain-Free approaches in both scenarios and demonstrates commendable performance relative to methods that rely on the $\mathcal{D}_r$ . This distinction highlights the effectiveness and versatility of our approach, making it a viable alternative for scenarios where the $\mathcal{D}_r$ is entirely unavailable. We also confirmed our result using the COCO dataset as a surrogate dataset (Sec. 1 of the Supp. Material). + +# 4.3 Ablation Study + +We conducted an ablation study to evaluate the impact of different loss components $\mathcal{L}_{LD}$ and $\mathcal{L}_M$ of our method on the unlearning performance across CR and HR scenarios. The results of this study are presented in Tab. 4. In CR, considering only $\mathcal{L}_{LD}$ leads to similar retain set accuracy but with high $\mathcal{D}_f$ accuracy. In fact, the stopping criteria is hardly fulfilled and the method stops at the maximum number of epochs allowed. Conversely, considering only $\mathcal{L}_M$ the + +Table 5: Results of SCAR changing the measure used in $\mathcal{L}_M$ on the CIFAR100. The metrics for CR and HR are reported as mean $\pm$ std over ten runs. + +
CIFAR100 CRCIFAR100 HR
ArrAftAUSArrAftAUS
Cosine Similarity72.74(02.19)02.00(02.36)0.933(0.031)73.88(00.46)77.05(00.54)0.934(0.008)
L2 Distance72.45(02.23)02.70(02.54)0.924(0.032)69.58(00.82)76.93(00.55)0.857(0.011)
Mahalanobis72.93(01.78)02.00(02.00)0.935(0.025)73.23(00.74)75.63(00.54)0.934(0.011)
+ +methods can push down the accuracy on the $\mathcal{D}_f$ , but at the cost of a low retain set accuracy. In the HR scenario, we observed similar but more pronounced results. The presence of $\mathcal{L}_{LD}$ alone is insufficient for effectively erasing knowledge about forget samples, as evidenced by a high forget accuracy. Employing only $\mathcal{L}_{LD}$ results in a steep decrease in both accuracies. This result suggests how the unlearning strategy without a method for retaining knowledge can't be used to unlearn. This effect is more evident in HR because images from the same class are distributed across both the $\mathcal{D}_r$ and $\mathcal{D}_f$ , meaning that the $\mathcal{L}_{LD}$ compromises the general knowledge of the classes. + +![](images/863d2b9093ed822a2889d3f27a4101a611ed54d4561b7aa209158e3ed304aff7.jpg) +Fig. 4: Values of final AUS for SCAR and SCAR self-forget as a function of the number of samples available in $\mathcal{D}^{\mathrm{sur}}$ (subset of Imagenet1K). AUS is reported as mean $\pm$ std over ten runs + +
CIFAR100
AfAfAUS
Original (AliCNN)70.08(00.15)69.70(14.83)0.545(0.048)
Unlearned65.35(02.37)02.90(03.38)0.926(0.036)
Original (Resnet18)77.55(00.11)77.50(02.80)0.563(0.009)
Unlearned72.93(01.78)02.00(02.00)0.935(0.025)
Original (Resnet34)80.13(00.10)80.30(10.02)0.569(0.031)
Unlearned75.67(01.74)01.90(00.92)0.937(0.019)
Original (Resnet50)80.64(00.10)80.80(09.50)0.570(0.030)
Unlearned75.99(01.77)00.80(02.36)0.942(0.028)
Original (ViT-B16)85.51(01.45)83.30(06.24)0.546(0.023)
Unlearned81.99(02.59)02.04(03.10)0.945(0.031)
+ +Table 6: Results of SCAR on CIFAR100 dataset in the CR scenario changing the model to unlearn. The metrics are reported as mean ± std over ten runs before and after the unlearning. + +# 4.4 Analysis on the effect of different measures in the metric learning mechanism of SCAR + +We report in Tab. 5 and Supp. Tab. 2 the results of the SCAR using different distance measures for the $\mathcal{L}_M$ component across CIFAR100 and TinyImagenet datasets in both the CR and HR scenarios. Using the Mahalanobis distance yields a perfect balance, achieving the highest AUS scores in both CR and HR. Cosine Similarity distance achieves compatible results (even if constantly lower) in CR and HR scenarios but at the cost of being $(22 \pm 5\%)$ slower than the Mahalanobis. This result suggests how the information about the distributions of samples embedded into the Mahalanobis distance is fundamental to guide the optimization efficiently. In contrast, the $L_2$ distance resulted in lower performance in both CR and HR scenarios suggesting how this metric hampers the unlearning process. + +# 4.5 Analysis of the impact of the dimension of $\mathcal{D}^{\mathrm{sur}}$ + +Fig. 4 illustrates the relationship between the number of images in the surrogate dataset $D_{sur}$ (subset of ImageNet1K) and the AUS for both SCAR and its self-forget variant, in the CR scenario. SCAR maintains a relatively stable AUS score as the number of images increases, suggesting that the method's unlearning ability does not significantly depend on the size of the surrogate dataset. The self-forget variant of SCAR shows a plateau in AUS score when at least 6K surrogate images are used which suggests how the dimension of $\mathcal{D}^{\mathrm{sur}}$ can affect the self-forget mechanism. Importantly, the self-forget mechanism that can access 6K surrogate samples or more exhibits performance comparable with the method that can access forget data. This pattern reflects the effectiveness of SCAR in handling the unlearning process, with only minor improvements seen with the addition of more surrogate data when forget data are not available. + +# 4.6 Architectural-Agnostic Results + +For additional insight, we examined the performance of SCAR (Tab. 6) across various DNN architectures within the CR scenario. We investigated the effect of the unlearning with SCAR on a basic CNN network (AllCNN), ResNet architectures (18, 34, and 50), and a Vision Transformer (ViT-B16). Our method consistently demonstrates strong performance across all tested architectures. Notably, the results achieved with the ViT confirm the model-agnostic nature of our approach beyond the CNN architectures. It is worth mentioning that the efficacy of our method is not contingent on the initial accuracy of the model, which underscores its adaptability and broad applicability. + +# 5 Conclusions + +In this paper, we introduced SCAR a novel method for approximate unlearning, designed to effectively erase information about the forget-set data while maintaining the original model performance. Our experimental results indicate that SCAR outperforms the methods that do not require access to the retain set and achieves performance levels comparable to, or even surpassing, those of existing approximate unlearning methods that depend on the retain set, across both CR and HR scenarios. Furthermore, we have developed a self-forget variation of SCAR capable of delivering impressive results in CR without accessing both the retain and forget sets. Future work should include a mathematical exploration of the algorithms' certifiability and the development of alternatives to the distillation-trick, enabling the retention of past knowledge without the need for a surrogate dataset, which can limit the application of SCAR to other domain. + +Despite these challenges, we firmly believe that our algorithm, that do not require access to any retain set, can significantly impact society: for example in the performance restoration of models poisoned by corrupted data or in recognition systems (such as facial recognition), by enabling individuals to exercise their right to have their data removed from the model's knowledge. + +# References + +1. Image captioning dataset, random images. https://www.kaggle.com/datasets/shamsaddin97/image-captioning-dataset-random-images/data, accessed: 2024-02-14 10 +2. Bibas, K., Feder, M., Hassner, T.: Single layer predictive normalized maximum likelihood for out-of-distribution detection. Advances in Neural Information Processing Systems 34, 1179-1191 (2021) 10 +3. Bourtoule, L., Chandrasekaran, V., Choquette-Choo, C.A., Jia, H., Travers, A., Zhang, B., Lie, D., Papernot, N.: Machine unlearning. In: 2021 IEEE Symposium on Security and Privacy (SP). pp. 141-159. IEEE (2021) 3 +4. Chen, M., Gao, W., Liu, G., Peng, K., Wang, C.: Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7766-7775 (June 2023) 4, 5, 8, 9 +5. Chen, S., Kahla, M., Jia, R., Qi, G.J.: Knowledge-enriched distributional model inversion attacks pp. 16158-16167 (2021). https://doi.org/10.1109/ICCV48922.2021.015871 +6. Chen, Y., Wiesel, A., Eldar, Y.C., Hero, A.O.: Shrinkage algorithms for mmse covariance estimation. IEEE Transactions on Signal Processing 58, 5016-5029 (2009), https://api-semanticscholar.org/CorpusID:12616083 8 +7. Cheng, J., Amiri, H.: Multimodal machine unlearning. arXiv preprint arXiv:2311.12047 (2023) 4 +8. Chandawat, V.S., Tarun, A.K., Mandal, M., Kankanhalli, M.: Can bad teaching induce forgetting? unlearning in deep networks using an incompetent teacher. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 7210-7217 (2023) 9 +9. Cotogni, M., Bonato, J., Sabetta, L., Pelosiin, F., Nicolosi, A.: Duck: Distance-based unlearning via centroid kinematics. arXiv preprint arXiv:2312.02052 (2023) 4, 5, 8, 9, 10, 11, 12 +0. De Lange, M., Tuytelaars, T.: Continual prototype evolution: Learning online from non-stationary data streams (2021). https://doi.org/10.1109/ICCV48922.2021.008146 +1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) 10 +2. Du, M., Yang, F., Zou, N., Hu, X.: Fairness in deep learning: A computational perspective. IEEE Intelligent Systems 36(4), 25-34 (2021). https://doi.org/10.1109/MIS.2020.3000681 1 +3. Fang, G., Bao, Y., Song, J., Wang, X., Xie, D., Shen, C., Song, M.: Mosaicking to distill: Knowledge distillation from out-of-domain data. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 11920-11932. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/63dc7ed1010d3c3b8269faf0ba7491d4-Paper.pdf 7 +4. Felps, D.L., Schwickerath, A.D., Williams, J.D., Vuong, T.N., Briggs, A., Hunt, M., Sakmar, E., Saranchak, D.D., Shumaker, T.: Class clown: Data redaction in machine unlearning at enterprise scale. arXiv preprint arXiv:2012.04699 (2020) 3 +5. Fort, S., Ren, J., Lakshminarayanan, B.: Exploring the limits of out-of-distribution detection. Advances in Neural Information Processing Systems 34, 7068-7081 (2021) 10 + +16. Fredrikson, M., Jha, S., Ristenpart, T.: Model inversion attacks that exploit confidence information and basic countermeasures. Proceedings of the 22nd ACM SIGSAC Conference on Computer and Communications Security (2015), https://apisemantic scholar.org/CorpusID:207229839 1 +17. Golatkar, A., Achille, A., Soatto, S.: Eternal sunshine of the spotless net: Selective forgetting in deep networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9304-9312 (2020) 3, 9 +18. Golatkar, A., Achille, A., Soatto, S.: Forgetting outside the box: Scrubbing deep networks of information accessible from input-output observations. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIX 16. pp. 383-398. Springer (2020) 3 +19. Graves, L., Nagisetty, V., Ganesh, V.: Amnesiac machine learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 11516-11524 (2021) 3 +20. Guo, C., Goldstein, T., Hannun, A., Van Der Maaten, L.: Certified data removal from machine learning models. In: Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 3832-3842. PMLR (2020) 4 +21. Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: International conference on machine learning. pp. 1321-1330. PMLR (2017) 7 +22. Han, X., Dai, Y., Gao, T., Lin, Y., Liu, Z., Li, P., Sun, M., Zhou, J.: Continual relation learning via episodic memory activation and reconsolidation (2020). https://doi.org/10.18653/v1/2020.acl-main.573, https://aclanthology.org/2020.acl-main.573 6 +23. Hayase, T., Yasutomi, S., Katoh, T.: Selective forgetting of deep networks at a finer level than samples. arXiv preprint arXiv:2012.11849 (2020) 3, 9 +24. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 10 +25. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: International Conference on Learning Representations (2017), https://openreview.net/forum?id=Hkg4TI9x1 7 +26. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) 2 +27. Hsu, Y.C., Shen, Y., Jin, H., Kira, Z.: Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10951-10960 (2020) 10 +28. Jia, J., Liu, J., Ram, P., Yao, Y., Liu, G., Liu, Y., Sharma, P., Liu, S.: Model sparsity can simplify machine unlearning. In: Annual Conference on Neural Information Processing Systems (2023) 4, 9 +29. Kahla, M., Chen, S., Just, H.A., Jia, R.: Label-only model inversion attacks via boundary repulsion. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15025-15033 (2022). https://doi.org/10.1109/CVPR52688.2022.014621 +30. Kim, J., Woo, S.S.: Efficient two-stage model retraining for machine unlearning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4361-4369 (2022) 4 +31. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) 9 + +32. Kurmanji, M., Triantafillou, P., Hayes, J., Triantafillou, E.: Towards unbounded machine unlearning. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=OveBaTtUAT 4, 5, 9 +33. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015) 9 +34. Leibig, C., Allken, V., Ayhan, M.S., Berens, P., Wahl, S.: Leveraging uncertainty information from deep neural networks for disease detection. Scientific reports 7(1), 17816 (2017) 7 +35. Lin, S., Zhang, X., Chen, C., Chen, X., Susilo, W.: Erm-ktp: Knowledge-level machine unlearning via knowledge transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20147-20155 (2023) 3, 9 +36. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) 10 +37. Liu, J., Xue, M., Lou, J., Zhang, X., Xiong, L., Qin, Z.: Mutter: Machine unlearning on adversarially trained models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4892-4902 (October 2023) 4 +38. Liu, X., Xie, L., Wang, Y., Zou, J., Xiong, J., Ying, Z., Vasilakos, A.V.: Privacy and security issues in deep learning: A survey. IEEE Access 9, 4566-4593 (2021), https://api-semanticscholar.org/CorpusID:231423554 1 +39. Magdziarczyk, M.: Right to be forgotten in light of regulation (eu) 2016/679 of the european parliament and of the council of 27 april 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing directive 95/46/ec. In: 6th International Multidisciplinary Scientific Conference on Social Sciences and Art Sgem 2019. pp. 177-184 (2019) 2 +40. Massey Jr, F.J.: The kolmogorov-smirnov test for goodness of fit. Journal of the American statistical Association 46(253), 68-78 (1951) 11 +41. Mercuri, S., Khraishi, R., Okhrati, R., Batra, D., Hamill, C., Ghasempour, T., Nowlan, A.: An introduction to machine unlearning. arXiv preprint arXiv:2209.00939 (2022) 2 +42. Neel, S., Roth, A., Sharifi-Malvajerdi, S.: Descent-to-delete: Gradient-based methods for machine unlearning. In: Algorithmic Learning Theory. pp. 931-962. PMLR (2021) 3 +43. Nguyen, A., Yosinski, J., Clune, J.: Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 427-436 (2015) 7 +44. Nguyen, T.T., Huynh, T.T., Nguyen, P.L., Liew, A.W.C., Yin, H., Nguyen, Q.V.H.: A survey of machine unlearning. arXiv preprint arXiv:2209.02299 (2022) 2, 3 +45. Pardau, S.L.: The california consumer privacy act: Towards a european-style privacy regime in the united states. J. Tech. L. & Pol'y 23, 68 (2018) 2 +46. Poppi, S., Poppi, T., Cocchi, F., Cornia, M., Baraldi, L., Cucchiara, R.: Removing NSFW Concepts from Vision-and-Language Models for Text-to-Image Retrieval and Generation. arXiv preprint arXiv:2311.16254 (2023) 4 +47. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning (2017). https://doi.org/10.1109/CVPR.2017.587 6 + +48. Rigaki, M., Garcia, S.: A survey of privacy attacks in machine learning. ACM Computing Surveys 56(4), 1-34 (Nov 2023). https://doi.org/10.1145/3624010, http://dx.doi.org/10.1145/3624010 1 +49. Salem, A., Zhang, Y., Humbert, M., Berrang, P., Fritz, M., Backes, M.: Ml-leaks: Model and data independent membership inference attacks and defenses on machine learning models. In: Proceedings of the 26th Annual Network and Distributed System Security Symposium (NDSS) (2019) 1, 10 +50. Setlur, A., Eysenbach, B., Smith, V., Levine, S.: Adversarial unlearning: Reducing confidence along adversarial directions. Advances in Neural Information Processing Systems 35, 18556-18570 (2022) 3 +51. Shaik, T.B., Tao, X., Xie, H., Li, L., Zhu, X., Li, Q.: Exploring the landscape of machine unlearning: A comprehensive survey and taxonomy. CoRR (2023) 2 +52. Shokri, R., Stronati, M., Song, C., Shmatikov, V.: Membership inference attacks against machine learning models. 2017 IEEE Symposium on Security and Privacy (SP) pp. 3-18 (2017) 1 +53. Sommer, D.M., Song, L., Wagh, S., Mittal, P.: Towards probabilistic verification of machine unlearning. arXiv preprint arXiv:2003.04247 (2020) 3 +54. Song, L., Mittal, P.: Systematic evaluation of privacy risks of machine learning models (2020) 10 +55. Song, L., Shokri, R., Mittal, P.: Privacy risks of securing machine learning models against adversarial examples. Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security (2019), https://api-semanticscholar.org/CorpusID:165163934 1 +56. Tarun, A.K., Chandawat, V.S., Mandal, M., Kankanhalli, M.: Fast yet effective machine unlearning. IEEE Transactions on Neural Networks and Learning Systems (2023) 4 +57. Tiwary, P., Guha, A., Panda, S., et al.: Adapt then unlearn: Exploiting parameter space semantics for unlearning in generative adversarial networks. arXiv preprint arXiv:2309.14054 (2023) 4 +58. Tukey, J.W.: Exploratory Data Analysis. Addison-Wesley (1977) 8 +59. Wang, K.C., FU, Y., Li, K., Khisti, A., Zemel, R., Makhzani, A.: Variational model inversion attacks. Advances in Neural Information Processing Systems 34, 9706-9719 (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/50a074e6a8da4662ae0a29edde722179-Paper.pdf 1 +60. Wu, G., Hashemi, M., Srinivasa, C.: Puma: Performance unchanged model augmentation for training data removal. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 8675-8682 (2022) 3 +61. Wu, Y., Dobriban, E., Davidson, S.: Deltagrad: Rapid retraining of machine learning models. In: International Conference on Machine Learning. pp. 10355-10366. PMLR (2020) 3 +62. Xu, J., Wu, Z., Wang, C., Jia, X.: Machine unlearning: Solutions and challenges. arXiv preprint arXiv:2308.07061 (2023) 2 +63. Yang, Z., Zhang, J., Chang, E.C., Liang, Z.: Neural network inversion in adversarial setting via background knowledge alignment. Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security p. 225-240 (2019). https://doi.org/10.1145/3319535.3354261, https://doi.org/10.1145/3319535.3354261 1 +64. Yeom, S., Giacomelli, I., Fredrikson, M., Jha, S.: Privacy risk in machine learning: Analyzing the connection to overfitting. 2018 IEEE 31st Computer Security Foundations Symposium (CSF) pp. 268-282 (2017), https://api-semanticscholar.org/CorpusID:2656445 1 + +65. Yin, Z., Xing, E., Shen, Z.: Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. Advances in Neural Information Processing Systems 36 (2024) 10 +66. Zhang, Y., Jia, R., Pei, H., Wang, W., Li, B., Song, D.: The secret revealer: Generative model-inversion attacks against deep neural networks. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 250-258 (jun 2020). https://doi.org/10.1109/CVPR42600.2020.00033, https://doi.ieeecomputersociety.org/10.1109/CVPR42600.2020.00033 1 +67. Zhao, X., Zhang, W., Xiao, X., Lim, B.: Exploiting explanations for model inversion attacks. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 662-672 (oct 2021). https://doi.org/10.1109/ICCV48922.2021.00072, https://doi.ieeeccomputersociety.org/10.1109/ICCV48922.2021.00072 1 \ No newline at end of file diff --git a/2024/Is Retain Set All You Need in Machine Unlearning_ Restoring Performance of Unlearned Models with Out-Of-Distribution Images/images.zip b/2024/Is Retain Set All You Need in Machine Unlearning_ Restoring Performance of Unlearned Models with Out-Of-Distribution Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2749b427570cc948554057a2efc4fe6543b73bb2 --- /dev/null +++ b/2024/Is Retain Set All You Need in Machine Unlearning_ Restoring Performance of Unlearned Models with Out-Of-Distribution Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f639406457270fbaa80630acba7bddda2560e5788c061c1218bf7b80d2c00041 +size 403358 diff --git a/2024/Is Retain Set All You Need in Machine Unlearning_ Restoring Performance of Unlearned Models with Out-Of-Distribution Images/layout.json b/2024/Is Retain Set All You Need in Machine Unlearning_ Restoring Performance of Unlearned Models with Out-Of-Distribution Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d71761d65e207acb8463b546d818cf1f56f84fb8 --- /dev/null +++ b/2024/Is Retain Set All You Need in Machine Unlearning_ Restoring Performance of Unlearned Models with Out-Of-Distribution Images/layout.json @@ -0,0 +1,11004 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 166 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 166 + ], + "type": "text", + "content": "Is Retain Set All You Need in Machine Unlearning? Restoring Performance of Unlearned Models with Out-Of-Distribution Images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 167, + 185, + 447, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 185, + 447, + 199 + ], + "spans": [ + { + "bbox": [ + 167, + 185, + 447, + 199 + ], + "type": "text", + "content": "Jacopo Bonato1,2, Marco Cotogni2*, and Luigi Sabetta1,2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 230, + 207, + 369, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 207, + 369, + 219 + ], + "spans": [ + { + "bbox": [ + 230, + 207, + 369, + 219 + ], + "type": "text", + "content": "1 Leonardo Labs, Rome, Italy," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 231, + 220, + 381, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 220, + 381, + 230 + ], + "spans": [ + { + "bbox": [ + 231, + 220, + 381, + 230 + ], + "type": "text", + "content": "2 These authors contributed equally" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 231, + 462, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 231, + 462, + 241 + ], + "spans": [ + { + "bbox": [ + 149, + 231, + 462, + 241 + ], + "type": "text", + "content": "{jacopo.bonato,luigi.sabetta}.ext@leonardo.com,cotogni@proton.me" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 271, + 455, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 271, + 455, + 460 + ], + "spans": [ + { + "bbox": [ + 159, + 271, + 455, + 460 + ], + "type": "text", + "content": "Abstract. In this paper, we introduce Selective-distillation for Class and Architecture-agnostic unleaRning (SCAR), a novel approximate unlearning method. SCAR efficiently eliminates specific information while preserving the model's test accuracy without using a retain set, which is a key component in state-of-the-art approximate unlearning algorithms. Our approach utilizes a modified Mahalanobis distance to guide the unlearning of the feature vectors of the instances to be forgotten, aligning them to the nearest wrong class distribution. Moreover, we propose a distillation-trick mechanism that distills the knowledge of the original model into the unlearning model with out-of-distribution images for retaining the original model's test performance without using any retain set. Importantly, we propose a self-forget version of SCAR that unlearns without having access to the forget set. We experimentally verified the effectiveness of our method, on three public datasets, comparing it with state-of-the-art methods. Our method obtains performance higher than methods that operate without the retain set and comparable w.r.t the best methods that rely on the retain set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 469, + 420, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 469, + 420, + 480 + ], + "spans": [ + { + "bbox": [ + 160, + 469, + 420, + 480 + ], + "type": "text", + "content": "Keywords: Machine Unlearning " + }, + { + "bbox": [ + 160, + 469, + 420, + 480 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 469, + 420, + 480 + ], + "type": "text", + "content": " Retain-Free " + }, + { + "bbox": [ + 160, + 469, + 420, + 480 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 469, + 420, + 480 + ], + "type": "text", + "content": " Data Removal" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 502, + 230, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 230, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 230, + 514 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "text", + "content": "Over the past decades, deep learning algorithms have grown significantly in applications like image classification, detection, and segmentation. For optimal performance, these models learn from big training datasets and update their weights to capture intrinsic patterns. Despite their effectiveness, many ethical concerns such as potential biases affecting certain demographics [12] and privacy like unauthorized use of sensitive or personal data [38,48] have been raised. Additionally, methods like model inversion [5,16,29,59,63,66,67], and membership inference attacks [49, 52, 55, 64] raised many concerns about the potential revelation of training data characteristics or can reveal particular data utilization during training. These vulnerabilities highlight the critical necessity for" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 288, + 664 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 288, + 664 + ], + "type": "text", + "content": "* Work done while at Leonardo Labs" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "content": "implementing safeguards to secure sensitive information within machine learning models. This imperative is fundamental in light of the privacy-preserving regulations that have been put into effect ( e.g. the European Union's General Data Protection Regulation [39] and the California Consumer Privacy Act [45])." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 482, + 464 + ], + "type": "text", + "content": "Machine Unlearning [41, 44, 51, 62] has emerged as a solution to these privacy and ethical challenges, aiming to remove specific information content from algorithms without degrading their performance. Unlearning algorithms can be classified into 3 main groups: model intrinsic, data-driven, and model agnostic. The characteristic of model-intrinsic methods is to change the deep neural network (DNN) architecture or add specific weights to induce the unlearning. The data-driven group is characterized by the usage of data to induce forgetting, and it can be roughly divided into methods that store optimization information during training (e.g. gradient updates) and methods that involve data manipulation. Similarly to this latter group, model-agnostic solutions do not require access to specific optimization-related information and operate directly on the unaltered trained model. Nonetheless, a common practice among these algorithms is to maintain or restore the model performance using the residual portion of the training dataset, known as retain set. Using the retain set to recalibrate models post-unlearning represents a viable strategy, however, it also introduces practical challenges. In scenarios where privacy concerns dictate, only the forget set —images to forget— might be accessible, leaving no retain set for performance restoration. This poses a significant hurdle for models designed to undergo a performance restoration phase after unlearning, potentially compromising their utility. Moreover, when large-scale datasets such as ImageNet or ImageNet-21K are involved and the forget set constitutes a minor fraction, the bulk of the data is the retain set. This disproportionality can significantly inflate the time required to restore the model to its original efficacy. Furthermore, there are even more intricate situations where also the forget set is unavailable, such as when the unlearning task entails forgetting an entire class with access only to its ID." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 474, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 665 + ], + "type": "text", + "content": "This paper presents a novel, model-agnostic unlearning algorithm, Selective-distillation for Class and Architecture-agnostic unleaRning (SCAR), which leverages metric learning and knowledge distillation [26] to efficiently remove targeted information and maintain model accuracy, notably without relying on a retain set. Specifically, we use the Mahalanobis distance to shift the feature vectors of the instances to forget toward the closest distribution of samples of other classes. This distance includes important information regarding the original dataset distribution of samples resulting in an efficient metric learning-based unlearning strategy. Concurrently, distilling the knowledge of the original model into the unlearning model using out-of-distribution (OOD) images, allows SCAR to maintain the performance of the original model on the test set. The main contributions of this work include: (i) SCAR, a novel model-agnostic unlearning algorithm that achieves competitive unlearning performance without using retain data. (ii) A unique self-forget mechanism in class removal scenarios that operates without direct access to the forget set. (iii) Comprehensive analyses demonstrating SCAR's efficacy across different datasets and architectures in" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "J. Bonato et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "class-removal and homogeneous-removal scenarios " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "(iv)" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " Experimental evidence showing SCAR's comparable or superior performance respectively to traditional unlearning methods or state-of-the-art (sota) techniques that do not utilize a retain set. Code is available at https://github.com/jbonato1/SCAR" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 184, + 243, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 184, + 243, + 197 + ], + "spans": [ + { + "bbox": [ + 131, + 184, + 243, + 197 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 210, + 479, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 210, + 479, + 234 + ], + "spans": [ + { + "bbox": [ + 130, + 210, + 479, + 234 + ], + "type": "text", + "content": "Machine Unlearning algorithms can be classified into three main categories: data-driven, model-intrinsic, and model-agnostic unlearning algorithms [44]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 235, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 482, + 521 + ], + "type": "text", + "content": "Data-driven. The unlearning algorithms in this category employ a data-driven approach to remove knowledge about specific image instances. For example, methods that involve data manipulation, such as transformations or attacks, fall within this group of algorithms [50, 53]. However, this category is not limited to such models: it encompasses a broader range of approaches. In [19, 61] the authors propose similar methods for reverting the gradient updates for the instances of the classes to be removed. Despite their effectiveness, these methods require storing gradient updates for all the batches in the training set. Similarly, in [42], the authors propose two gradient-descent-based approaches for the effective unlearning of convex models. In the first, Gaussian noise is used to perturb the gradient updates, while the second also incorporates reservoir sampling. In [60], the authors modeled the impact of each training point on the model's training process and then compensated for the impact of removing data from the model's knowledge by re-weighting the gradient updates. In [14], the authors employed a strategy of recursively assigning an incorrect class to samples designated for forgetting until a membership inference attack no longer recognizes those samples as part of the training dataset. In [3], the authors segmented the dataset into n parts and trained n models on each segment; during inference, predictions from these models are aggregated. Only the model containing those samples is retrained when a removal request is received. Moreover, novel retain-set-free algorithms such as Random Labels [23] and Negative Gradient [17] have been introduced. In the former approach, random labels are assigned to samples designated for forgetting to calculate the cross-entropy, while the latter employs the negative gradient obtained from these samples to optimize the loss function." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": "Model-intrinsic. Model-intrinsic approaches to machine unlearning consist of algorithms that depend on a specific architecture. In these cases, the unlearning process is facilitated by sophisticated mechanisms tailored to function within that particular architecture. Depending on the task to be performed several architectures can be involved. For instance in [17, 18] the authors proposed two different unlearning approaches for CNNs. In the former, the authors \"scrubbed\" from the weights of the model the knowledge about instances of images through the computation of Fisher's information matrix. Similarly in the latter, the authors propose a CNN-based unlearning algorithm, inspired by the neural tangent kernel, that models the weights dynamics and can remove from the information about forget data. In [35], the authors proposed a two-phase algorithm for the approximate unlearning of CNNs. During the first phase, called ERM, the model" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "SCAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 132, + 319, + 264 + ], + "blocks": [ + { + "bbox": [ + 160, + 116, + 170, + 125 + ], + "lines": [ + { + "bbox": [ + 160, + 116, + 170, + 125 + ], + "spans": [ + { + "bbox": [ + 160, + 116, + 170, + 125 + ], + "type": "text", + "content": "A" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 132, + 319, + 264 + ], + "lines": [ + { + "bbox": [ + 159, + 132, + 319, + 264 + ], + "spans": [ + { + "bbox": [ + 159, + 132, + 319, + 264 + ], + "type": "image", + "image_path": "508aedf6204bab9f065a87c3edff878b7e2b074de67306ab54a77c6a48f2ed85.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 361, + 120, + 426, + 132 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 120, + 426, + 132 + ], + "spans": [ + { + "bbox": [ + 361, + 120, + 426, + 132 + ], + "type": "interline_equation", + "content": "d _ {J S} (\\Phi_ {\\theta} (x _ {j}) | | \\Phi_ {\\theta} ^ {U} (x _ {j}))", + "image_path": "78fe9018593c6bd682bf0a5810e91bc06df3dd1754205e6feefda520f2383535.jpg" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 329, + 143, + 389, + 182 + ], + "blocks": [ + { + "bbox": [ + 334, + 116, + 342, + 124 + ], + "lines": [ + { + "bbox": [ + 334, + 116, + 342, + 124 + ], + "spans": [ + { + "bbox": [ + 334, + 116, + 342, + 124 + ], + "type": "text", + "content": "B" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 329, + 143, + 389, + 182 + ], + "lines": [ + { + "bbox": [ + 329, + 143, + 389, + 182 + ], + "spans": [ + { + "bbox": [ + 329, + 143, + 389, + 182 + ], + "type": "image", + "image_path": "a007c1a9022ea8fed443e9a79f0cbc3c8fb150f9a7f6601305fe512c9e6874a3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 393, + 143, + 453, + 182 + ], + "blocks": [ + { + "bbox": [ + 393, + 143, + 453, + 182 + ], + "lines": [ + { + "bbox": [ + 393, + 143, + 453, + 182 + ], + "spans": [ + { + "bbox": [ + 393, + 143, + 453, + 182 + ], + "type": "image", + "image_path": "812df6d6f2395bc5d161cd48572da11d0e04f976d1fb544e91e6ffdc11a2c91c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 369, + 191, + 417, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 191, + 417, + 202 + ], + "spans": [ + { + "bbox": [ + 369, + 191, + 417, + 202 + ], + "type": "interline_equation", + "content": "d _ {M} (\\Phi_ {\\psi} ^ {U} (x _ {j}), Q _ {j} ^ {*})", + "image_path": "ad6fecd1e08c425fdc21b2ab201c902581cadfe599124275345646b758b5648e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 331, + 203, + 453, + 263 + ], + "blocks": [ + { + "bbox": [ + 335, + 189, + 343, + 198 + ], + "lines": [ + { + "bbox": [ + 335, + 189, + 343, + 198 + ], + "spans": [ + { + "bbox": [ + 335, + 189, + 343, + 198 + ], + "type": "text", + "content": "C" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 203, + 453, + 263 + ], + "lines": [ + { + "bbox": [ + 331, + 203, + 453, + 263 + ], + "spans": [ + { + "bbox": [ + 331, + 203, + 453, + 263 + ], + "type": "image", + "image_path": "97b2e9e4f74a8e5ddf4ecdafe5c1c2768a918b60e1bb6be0098d8b1ebcff378d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 268, + 482, + 313 + ], + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 313 + ], + "type": "text", + "content": "Fig. 1: SCAR overview. A) SCAR scheme B) distillation-trick through J.S. divergence of the logits of " + }, + { + "bbox": [ + 130, + 268, + 482, + 313 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}(x_j)" + }, + { + "bbox": [ + 130, + 268, + 482, + 313 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 268, + 482, + 313 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}^{U}(x_{j})" + }, + { + "bbox": [ + 130, + 268, + 482, + 313 + ], + "type": "text", + "content": " . C) Representation of the feature vector distributions of samples for 3 classes before (left) and after (right) the unlearning process. Ellipses denote the " + }, + { + "bbox": [ + 130, + 268, + 482, + 313 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 268, + 482, + 313 + ], + "type": "text", + "content": " confidence intervals of the distribution of samples." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 340, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 484 + ], + "type": "text", + "content": "is trained on all the data while a binary mask is learned for each layer of the CNN for each class of the dataset. Then during the second phase, called KTP, the knowledge about the classes of the forget set is removed using the previously learned masks. In [20], the authors propose a certified algorithm for removing knowledge about classes on linear classifiers. This method applied a Newton step on the model's weights to significantly reduce the influence of the deleted data point, with the residual error decreasing quadratically as the training set size increases. Then, for certifying the data removal and to prevent the extraction of information from the minimal residual the training loss is perturbed. Recently, additional unlearning approaches have been developed for dealing with specific models on different learning paradigms such as Multimodal learning [7, 46] and Adversarial Learning [37, 57]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Model-agnostic. This group of algorithms includes universally applicable methods across different architectures. Their characteristic is the ability to intervene in the feature/activation space to selectively erase data associated with the forget set. In [32], the authors used knowledge distillation to differentiate the logits of the forget set from those of the retain set and erase the knowledge of these examples. A similar strategy was proposed in [30], where distillation was combined with contrastive labeling to achieve efficient unlearning. In [4] the authors proposed two methods for unlearning based on the kinematics of the decision boundary. The methods change the labels of the examples of the forget set to the closest wrong class and then fine-tune the model with the reassigned samples. Similarly, In [9] the authors proposed a metric learning-based approach that associates samples from the forget set to the closest centroid in the feature space. Through this procedure, the model preserves the feature space organization. In [28] the authors proposed a model sparsification via weight pruning for unlearning data samples. In [56], the authors proposed to learn a matrix" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "J. Bonato et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "noise associated with the classes to be forgotten to remove their influence on the model's weights followed by a repair step used to recover knowledge about retain samples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 308 + ], + "type": "text", + "content": "Our approach stands out as a model-agnostic approximate unlearning algorithm and partially aligns with Boundary Unlearning [4] or DUCK [9] that manipulates respectively forget samples labels or forget samples feature vectors to induce unlearning. However, while these methods adjust samples towards incorrect directions based on the nearest wrong class or centroid in feature space, our solution focuses on using the distributions of features to efficiently reorganize the positions of forget samples in the feature space. This enables us to utilize a more precise and efficient mechanism, grounded in the Mahalanobis distance. Nonetheless, SCAR does not rely on any images from the retain set to restore performance during the unlearning process whereas model-agnostic methods like DUCK [9], Boundary Unlearning [4], and SCRUB [32] partially or fully depend on it. SCAR, thanks to an efficient knowledge preservation method based on distillation and OOD data, can operate without accessing training data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 326, + 207, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 207, + 338 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 207, + 338 + ], + "type": "text", + "content": "3 Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 352, + 228, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 228, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 228, + 363 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "text", + "content": "Consider a dataset " + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\{x_i, y_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "text", + "content": " consisting of N images " + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "text", + "content": " each associated with a label " + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "inline_equation", + "content": "y_i \\in \\{0, \\dots, K\\}" + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "text", + "content": " is the number of classes in the dataset. A model " + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\varPhi_\\theta" + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "text", + "content": " is trained on the training subset of the dataset " + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "text", + "content": " to predict the corresponding label for a given image, that is, " + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\hat{y}_i = \\varPhi_\\theta(x_i)" + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "text", + "content": ". This model is then evaluated on the test subset " + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^t" + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "text", + "content": ". The objective of a machine unlearning algorithm is to derive a model, denoted as " + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\varPhi_\\theta^U" + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "text", + "content": ", from which information about a specific subset of data, known as the forget set " + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "text", + "content": ", is removed from the original model's parameters " + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "text", + "content": ". Concurrently, the algorithm aims to preserve the model's performance on the retain set " + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 371, + 482, + 503 + ], + "type": "text", + "content": ", such that it is comparable to that of a model trained exclusively on the retain set. In this paper, we consider two scenarios: Class-Removal (CR) and Homogeneous Removal (HR)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "content": "In the CR scenario, the objective is to remove from the model's weights the information about the instances of a class " + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "content": " among the " + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "content": " available. The two subsets " + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^t" + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "content": " are split in forget " + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_f, \\mathcal{D}_f^t)" + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "content": " and retain " + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_r, \\mathcal{D}_r^t)" + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "content": " sets. The forget set contains all the images, from the class to be removed " + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f = \\{x_i^f, y_i^f\\}_{i=1}^{N^f}" + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "inline_equation", + "content": "y_i^f \\in \\{C\\}" + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "content": " while the retain set contains all the remaining images " + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r = \\{x_i^r, y_i^r\\}_{i=1}^{N^r}" + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "inline_equation", + "content": "y_i^r \\in \\{0, \\dots, K-1\\}" + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "content": "; the same definition stands also for " + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^t" + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "content": ". The desiderata in this scenario is to maintain the original accuracy on the " + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r^t" + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "content": " while minimizing the one on the " + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f^t" + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "content": ". In the HR scenario, the training dataset " + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "content": " is divided into retain " + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "content": " and forget " + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "content": " sets, while the test set " + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^t" + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "content": " remains unchanged. In this case, the images constituting " + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "content": " are chosen from all the classes. Here, the goal is to eliminate the model's knowledge of specific training instances in " + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "content": " without compromising its performance on " + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 504, + 482, + 666 + ], + "type": "text", + "content": ". Following the application of the unlearning algorithm, the model should no longer differentiate" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "SCAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "between data from " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^t" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": " or in other words it interprets " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": " samples as never-seen-data. This indistinguishability has also to be reflected in the " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": " accuracy which has to be approximately similar to " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^t" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": " accuracy. This second scenario is more challenging compared to CR since " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": " include images from the same classes and the model should maintain its ability to generalize across all classes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 205, + 194, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 194, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 194, + 216 + ], + "type": "text", + "content": "3.2 SCAR" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 357 + ], + "type": "text", + "content": "In SCAR, the unlearning process is based on metric learning and involves directly modifying the feature vector projection of the sample to be forgotten within the feature space. Additionally, we have devised a regularization technique that enables the DNN to preserve the original model information through distillation without access to the retain dataset. By leveraging an OOD dataset, SCAR enables the transfer of knowledge from the original model to the unlearning model, thereby ensuring that classification performance on test samples remains consistent throughout the unlearning process. Initially, we present the unlearning strategy of SCAR, followed by a deeper exploration of the distillation with OOD data and the Metric Learning mechanism. Then we introduce a variant of SCAR, called SCAR self-forget, that works without accessing the forget set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 359, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 482, + 407 + ], + "type": "text", + "content": "Unlearning strategy. The unlearning strategy is composed of 2 synergistic mechanisms, Metric Learning and Distillation-Trick (Fig. 1 A). In the following sections each DNN " + }, + { + "bbox": [ + 130, + 359, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}" + }, + { + "bbox": [ + 130, + 359, + 482, + 407 + ], + "type": "text", + "content": " will be represented as a sequence of a backbone " + }, + { + "bbox": [ + 130, + 359, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\psi}" + }, + { + "bbox": [ + 130, + 359, + 482, + 407 + ], + "type": "text", + "content": " and a final fully-connected layer " + }, + { + "bbox": [ + 130, + 359, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\pi}" + }, + { + "bbox": [ + 130, + 359, + 482, + 407 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 359, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\theta, \\psi, \\pi" + }, + { + "bbox": [ + 130, + 359, + 482, + 407 + ], + "type": "text", + "content": " are the corresponding weight." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "text", + "content": "During the training phase of the original model " + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}" + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "text", + "content": ", for each class " + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "text", + "content": " the distribution of feature vectors " + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "inline_equation", + "content": "Q_{i} = \\{\\varPhi_{\\psi}(x_{j})\\}_{j}" + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "text", + "content": " is prototyped [10, 22, 47] by its mean " + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\mu_{i}" + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "text", + "content": " and covariance matrix " + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\hat{S}_Q" + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\psi}(x_j)" + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "text", + "content": " represent a feature vector), and accurately stored preventing any data leakage. Hence, given a forget sample " + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "inline_equation", + "content": "(x_{j},y_{j} = k)" + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "text", + "content": " belonging to the " + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "text", + "content": " class, SCAR selects the closest distribution " + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "inline_equation", + "content": "Q_{j}^{*}" + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "text", + "content": " among the set of distribution " + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "inline_equation", + "content": "i\\in \\{0,\\dots,K\\}" + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "inline_equation", + "content": "i\\neq k" + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "text", + "content": " to the feature vector " + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\psi}^{U}(x_{j})" + }, + { + "bbox": [ + 131, + 407, + 482, + 492 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 244, + 493, + 481, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 493, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 244, + 493, + 481, + 514 + ], + "type": "interline_equation", + "content": "Q _ {j} ^ {*} = \\underset {Q _ {i}} {\\operatorname {a r g m i n}} d _ {M} \\left(\\Phi_ {\\psi} ^ {U} \\left(x _ {j}\\right), Q _ {i}\\right) \\tag {1}", + "image_path": "4f7789bb949fd5755138468672b6b926dd25922d36091706fdf387a9f3cf03fb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 518, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 518, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 131, + 518, + 482, + 567 + ], + "type": "text", + "content": "Hence, during the unlearning procedure, the algorithm, through the Metric Learning mechanism, minimizes the Mahalanobis distance between the forget samples feature vector " + }, + { + "bbox": [ + 131, + 518, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\psi}^{U}(x_{j})" + }, + { + "bbox": [ + 131, + 518, + 482, + 567 + ], + "type": "text", + "content": " and their corresponding selected distribution " + }, + { + "bbox": [ + 131, + 518, + 482, + 567 + ], + "type": "inline_equation", + "content": "Q_{j}^{*}" + }, + { + "bbox": [ + 131, + 518, + 482, + 567 + ], + "type": "text", + "content": ". The overall forget loss is written in eq. 2:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 216, + 575, + 481, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 575, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 216, + 575, + 481, + 609 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M} = \\frac {1}{N _ {f , \\text {b a t c h}}} \\sum_ {j = 0} ^ {N _ {f, \\text {b a t c h}} - 1} d _ {M} \\left(\\Phi_ {\\psi} ^ {U} \\left(x _ {j}\\right), Q _ {j} ^ {*}\\right). \\tag {2}", + "image_path": "a45936e35267e3a6d3c006356100b48f9a214d9277bf9d4fb4282605247754f7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 617, + 352, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 352, + 629 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 352, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 617, + 352, + 629 + ], + "type": "inline_equation", + "content": "N_{f,\\mathrm{batch}}" + }, + { + "bbox": [ + 131, + 617, + 352, + 629 + ], + "type": "text", + "content": " is the batch size of forget samples." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Since in our proposed scenario, the " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": " dataset is no longer accessible, the algorithm makes use of the Distillation-Trick mechanism to retain past knowledge. We used the Jensen-Shannon divergence between the original model " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "J. Bonato et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 164, + 114, + 303, + 171 + ], + "blocks": [ + { + "bbox": [ + 164, + 114, + 303, + 171 + ], + "lines": [ + { + "bbox": [ + 164, + 114, + 303, + 171 + ], + "spans": [ + { + "bbox": [ + 164, + 114, + 303, + 171 + ], + "type": "image", + "image_path": "039f2d5e6cab050938a51226004522e67a253b762229ec9546be3223b50a1b15.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 180, + 479, + 204 + ], + "lines": [ + { + "bbox": [ + 131, + 180, + 479, + 204 + ], + "spans": [ + { + "bbox": [ + 131, + 180, + 479, + 204 + ], + "type": "text", + "content": "Fig.2: A-B) TSNE plots of feature vectors for the first 20 classes in CIFAR100 (A, " + }, + { + "bbox": [ + 131, + 180, + 479, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 131, + 180, + 479, + 204 + ], + "type": "text", + "content": ") and the surrogate dataset subset Imagenet (B, " + }, + { + "bbox": [ + 131, + 180, + 479, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{sur}}" + }, + { + "bbox": [ + 131, + 180, + 479, + 204 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 114, + 445, + 171 + ], + "blocks": [ + { + "bbox": [ + 306, + 114, + 445, + 171 + ], + "lines": [ + { + "bbox": [ + 306, + 114, + 445, + 171 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 445, + 171 + ], + "type": "image", + "image_path": "6641b2536b84e30313f6f89d685f791ede64b203dd34e4f08083a134ceb5d615.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 224, + 416, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 224, + 416, + 237 + ], + "spans": [ + { + "bbox": [ + 131, + 224, + 416, + 237 + ], + "type": "text", + "content": "and the unlearning model " + }, + { + "bbox": [ + 131, + 224, + 416, + 237 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}^{U}" + }, + { + "bbox": [ + 131, + 224, + 416, + 237 + ], + "type": "text", + "content": " (see par. Distillation-Trick, eq. 5)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 201, + 243, + 480, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 243, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 201, + 243, + 480, + 277 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {T D} = \\frac {1}{N _ {r , \\text {b a t c h}}} \\sum_ {j = 0} ^ {N _ {r, \\text {b a t c h}} - 1} d _ {J S} \\left(\\Phi_ {\\theta} ^ {U} \\left(x _ {j}\\right) \\| \\Phi_ {\\theta} \\left(x _ {j}\\right)\\right) \\tag {3}", + "image_path": "5649a2e2fa3b0d3b82bb67ab5f890610b37cc598e9f38b84c9245745e56e4f8f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 282, + 479, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 282, + 479, + 306 + ], + "spans": [ + { + "bbox": [ + 131, + 282, + 479, + 306 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 282, + 479, + 306 + ], + "type": "inline_equation", + "content": "N_{r,\\mathrm{batch}}" + }, + { + "bbox": [ + 131, + 282, + 479, + 306 + ], + "type": "text", + "content": " is the batch size of the surrogate retain samples. The overall loss combines the forget and retain contribution in eq. 2 and 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 260, + 312, + 480, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 312, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 260, + 312, + 480, + 326 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {M} + \\lambda_ {2} \\mathcal {L} _ {T D} \\tag {4}", + "image_path": "84ac5479de48c47733c3c6e35b23c33aa83f79bd9bf21bc5ca2e984fb5508b8a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 331, + 479, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 479, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 479, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 331, + 479, + 415 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 331, + 479, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 331, + 479, + 415 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 331, + 479, + 415 + ], + "type": "text", + "content": " are hyperparameters. Instead of selecting a priori the number of unlearning epochs, SCAR computes the " + }, + { + "bbox": [ + 130, + 331, + 479, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_f" + }, + { + "bbox": [ + 130, + 331, + 479, + 415 + ], + "type": "text", + "content": " at each epoch and it stops when " + }, + { + "bbox": [ + 130, + 331, + 479, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_r" + }, + { + "bbox": [ + 130, + 331, + 479, + 415 + ], + "type": "text", + "content": " is lower or equal to a threshold " + }, + { + "bbox": [ + 130, + 331, + 479, + 415 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 331, + 479, + 415 + ], + "type": "text", + "content": ". In CR " + }, + { + "bbox": [ + 130, + 331, + 479, + 415 + ], + "type": "inline_equation", + "content": "\\epsilon = 0" + }, + { + "bbox": [ + 130, + 331, + 479, + 415 + ], + "type": "text", + "content": " because the objective is to remove completely the knowledge of the class to remove. In HR " + }, + { + "bbox": [ + 130, + 331, + 479, + 415 + ], + "type": "inline_equation", + "content": "\\epsilon = \\mathcal{A}^t" + }, + { + "bbox": [ + 130, + 331, + 479, + 415 + ], + "type": "text", + "content": " because the forget test accuracy has to be close to the accuracy of the original model on the test set. When this condition is not reached the unlearning process stops when it reaches the maximum number of epochs allowed (epochs_max)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 418, + 482, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 418, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 131, + 418, + 482, + 610 + ], + "type": "text", + "content": "Distillation-trick. Numerous instances have been documented in literature wherein DNNs exhibit high confidence predictions [21], even when presented with data significantly distant from their training set. Examples include their performance on fooling images [43], OOD samples [25], or within medical diagnostic tasks [34]. For instance, the t-SNE projection of feature vectors from a DNN, both from the original training dataset and an external OOD dataset, exhibited clustering at identical positions within the feature space (Fig. 2 A-B). In our proposed algorithm, we leverage this characteristic behavior of DNNs on OOD datasets as a form of regularization during the unlearning process, which we term \"distillation-trick\". This approach involves distilling the knowledge from the original frozen model (referred to as the teacher model) into the unlearning model using an external dataset acting as a surrogate dataset, denoted as " + }, + { + "bbox": [ + 131, + 418, + 482, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{sur}}" + }, + { + "bbox": [ + 131, + 418, + 482, + 610 + ], + "type": "text", + "content": " [13] (Fig. 1 B). The distillation-trick leverages the Jensen-Shannon divergence between " + }, + { + "bbox": [ + 131, + 418, + 482, + 610 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}^{U}(x_{j})" + }, + { + "bbox": [ + 131, + 418, + 482, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 418, + 482, + 610 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}(x_{j})" + }, + { + "bbox": [ + 131, + 418, + 482, + 610 + ], + "type": "text", + "content": " to retain the knowledge of the original model, even as the unlearning model undergoes training on " + }, + { + "bbox": [ + 131, + 418, + 482, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{sur}}" + }, + { + "bbox": [ + 131, + 418, + 482, + 610 + ], + "type": "text", + "content": " (omitting the dependency on " + }, + { + "bbox": [ + 131, + 418, + 482, + 610 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 131, + 418, + 482, + 610 + ], + "type": "text", + "content": " for brevity):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 172, + 614, + 480, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 614, + 480, + 637 + ], + "spans": [ + { + "bbox": [ + 172, + 614, + 480, + 637 + ], + "type": "interline_equation", + "content": "d _ {J S} \\left(\\Phi_ {\\theta} ^ {U} \\| \\Phi_ {\\theta} / T\\right) = \\frac {1}{2} D _ {K L} \\left(\\Phi_ {\\theta} ^ {U} \\| \\Phi_ {\\theta} / T\\right) + \\frac {1}{2} D _ {K L} \\left(\\Phi_ {\\theta} / T \\| \\Phi_ {\\theta} ^ {U}\\right) \\tag {5}", + "image_path": "eea91eff317103addbaf61c36a097d90d17510064e944218bbe870053ecd0736.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 641, + 483, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 483, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 483, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 641, + 483, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{T}" + }, + { + "bbox": [ + 131, + 641, + 483, + 665 + ], + "type": "text", + "content": " is a scaling temperature. To validate this, we trained " + }, + { + "bbox": [ + 131, + 641, + 483, + 665 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}" + }, + { + "bbox": [ + 131, + 641, + 483, + 665 + ], + "type": "text", + "content": " using distillation-trick on a surrogate dataset. Our experiments demonstrate that the accuracy of" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "SCAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": " remains consistent across epochs and comparable to that of " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}^{U}" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": " (refer to Sec. 1 of the Supp. Materials)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 323 + ], + "type": "text", + "content": "Metric learning. In the context of unlearning problems, metric learning serves as a pivotal concept aimed at modifying the feature vectors of forget samples, thereby altering their class predictions. The prevalent strategies involve shifting the feature vectors of forget samples towards the nearest incorrect class centroids [9] or directly shifting the logits towards the nearest incorrect class [4]. In our algorithm, we propose the application of metric learning by adjusting the feature vectors of forget samples toward the nearest distribution of feature vectors of a different class (Fig. 1 C). Thanks to the Mahalanobis distance, which gauges the distance between the forget set samples and the whole distributions of feature vectors, SCAR can integrate more comprehensive information about the distributions through the covariance matrix. This approach contrasts with simpler methods that represent distributions solely using centroids (see Sec.4.4 for a comparison with other distances based on centroids). Therefore, given the set of distribution of feature vectors for each class " + }, + { + "bbox": [ + 130, + 143, + 482, + 323 + ], + "type": "inline_equation", + "content": "Q = \\{Q_{i}\\}_{i = 0}^{K}" + }, + { + "bbox": [ + 130, + 143, + 482, + 323 + ], + "type": "text", + "content": " the Mahalanobis distance is defined in eq. 6." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 186, + 332, + 481, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 332, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 186, + 332, + 481, + 352 + ], + "type": "interline_equation", + "content": "d _ {M} (\\varPhi_ {\\psi} ^ {U} (x _ {j}), Q _ {i}) = \\sqrt {(\\varPhi_ {\\psi} ^ {U} (x _ {j}) - \\mu_ {Q}) ^ {T} \\hat {S} _ {Q} ^ {- 1} (\\varPhi_ {\\psi} ^ {U} (x _ {j}) - \\mu_ {Q})} \\tag {6}", + "image_path": "61761ea3421a771fecdeb91d2c6bc4c2bc494a26a2a577b49970895732cbeec3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 359, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 482, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 359, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\psi}^{U}(x_{j})" + }, + { + "bbox": [ + 130, + 359, + 482, + 460 + ], + "type": "text", + "content": " is the feature vector of " + }, + { + "bbox": [ + 130, + 359, + 482, + 460 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 130, + 359, + 482, + 460 + ], + "type": "text", + "content": " (i.e. the output of the backbone of " + }, + { + "bbox": [ + 130, + 359, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\varPhi^U" + }, + { + "bbox": [ + 130, + 359, + 482, + 460 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 130, + 359, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\mu_Q" + }, + { + "bbox": [ + 130, + 359, + 482, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 359, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\hat{S}_Q" + }, + { + "bbox": [ + 130, + 359, + 482, + 460 + ], + "type": "text", + "content": " are the mean and the normalized covariance matrix obtained from the distribution " + }, + { + "bbox": [ + 130, + 359, + 482, + 460 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 130, + 359, + 482, + 460 + ], + "type": "text", + "content": ". Each class's covariance matrix " + }, + { + "bbox": [ + 130, + 359, + 482, + 460 + ], + "type": "inline_equation", + "content": "S_{Q}" + }, + { + "bbox": [ + 130, + 359, + 482, + 460 + ], + "type": "text", + "content": " will exhibit varying levels of scaling and variances across different dimensions. Consequently, the Mahalanobis distances of features from distinct classes will be subjected to differing scaling factors. Therefore, to standardize the covariance matrices per class effectively, we conduct correlation matrix normalization on all the covariance matrices (eq. 7):" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 162, + 468, + 481, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 468, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 162, + 468, + 481, + 495 + ], + "type": "interline_equation", + "content": "\\hat {S} _ {Q} (i, j) = \\frac {S _ {Q} (i , j)}{\\sigma_ {Q} (i) \\sigma_ {Q} (j)}, \\quad \\sigma_ {Q} (j) = \\sqrt {S _ {Q} (j , j)}, \\quad \\sigma_ {Q} (i) = \\sqrt {S _ {Q} (i , i)}. \\qquad (7)", + "image_path": "b1da28d14e8588b76b4597aeaea140f1e3374f9b6f25639547083755d4147c33.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 502, + 482, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 599 + ], + "type": "text", + "content": "Since the number of samples per class in public datasets (" + }, + { + "bbox": [ + 130, + 502, + 482, + 599 + ], + "type": "inline_equation", + "content": "\\approx 500" + }, + { + "bbox": [ + 130, + 502, + 482, + 599 + ], + "type": "text", + "content": " for CIFAR100 and TinyImagenet) is generally lower than the number of dimensions in the feature space (e.g. " + }, + { + "bbox": [ + 130, + 502, + 482, + 599 + ], + "type": "inline_equation", + "content": "\\mathrm{N} = 512" + }, + { + "bbox": [ + 130, + 502, + 482, + 599 + ], + "type": "text", + "content": " for resnet18), we use a covariance shrinkage method to get a full-rank matrix [6], " + }, + { + "bbox": [ + 130, + 502, + 482, + 599 + ], + "type": "inline_equation", + "content": "S_{Q}^{s} = S_{Q} + \\gamma_{0}V_{0}\\mathbb{I} + \\gamma_{1}V_{1}(1 - \\mathbb{I})" + }, + { + "bbox": [ + 130, + 502, + 482, + 599 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 502, + 482, + 599 + ], + "type": "inline_equation", + "content": "V_{0}" + }, + { + "bbox": [ + 130, + 502, + 482, + 599 + ], + "type": "text", + "content": " is the average diagonal variance, " + }, + { + "bbox": [ + 130, + 502, + 482, + 599 + ], + "type": "inline_equation", + "content": "V_{2}" + }, + { + "bbox": [ + 130, + 502, + 482, + 599 + ], + "type": "text", + "content": " is the average off-diagonal covariance of " + }, + { + "bbox": [ + 130, + 502, + 482, + 599 + ], + "type": "inline_equation", + "content": "S_{Q}" + }, + { + "bbox": [ + 130, + 502, + 482, + 599 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 502, + 482, + 599 + ], + "type": "inline_equation", + "content": "\\mathbb{I}" + }, + { + "bbox": [ + 130, + 502, + 482, + 599 + ], + "type": "text", + "content": " is the identity matrix. Furthermore, we added Tukey's normalization [58] to the feature vectors " + }, + { + "bbox": [ + 130, + 502, + 482, + 599 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\psi}^{U}(x)^{\\delta}" + }, + { + "bbox": [ + 130, + 502, + 482, + 599 + ], + "type": "text", + "content": " to obtain approximately Gaussian features, where " + }, + { + "bbox": [ + 130, + 502, + 482, + 599 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 502, + 482, + 599 + ], + "type": "text", + "content": " is a hyperparameter to decide the degree of transformation of the distribution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 621, + 249, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 249, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 249, + 634 + ], + "type": "text", + "content": "3.3 SCAR Self-forget" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Additionally, we present an adaptation of SCAR tailored for a more demanding CR scenario, wherein access to the forget set is restricted, and only the class ID" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "J. Bonato et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "slated for removal is provided. We term this procedure SCAR self-forget (Fig. 3). In this version, both the forgetting mechanism and the distillation-trick rely on the surrogate dataset. Given " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{ret}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": ", the set of retained classes, and " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{fgt}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": ", the set of forget classes, SCAR initially employs the original model " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\theta}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " to classify " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{sur}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 181, + 310, + 289 + ], + "blocks": [ + { + "bbox": [ + 134, + 181, + 310, + 289 + ], + "lines": [ + { + "bbox": [ + 134, + 181, + 310, + 289 + ], + "spans": [ + { + "bbox": [ + 134, + 181, + 310, + 289 + ], + "type": "image", + "image_path": "9bff3dda2b4e3e380f9995e6a43b22677bd66a6be4360ab0533c1aae20a20145.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 295, + 317, + 340 + ], + "lines": [ + { + "bbox": [ + 131, + 295, + 317, + 340 + ], + "spans": [ + { + "bbox": [ + 131, + 295, + 317, + 340 + ], + "type": "text", + "content": "Fig. 3: Scheme of SCAR self-forget in CR. The surrogate dataset " + }, + { + "bbox": [ + 131, + 295, + 317, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{sur}}" + }, + { + "bbox": [ + 131, + 295, + 317, + 340 + ], + "type": "text", + "content": " supplies during the unlearning procedure both the surrogates " + }, + { + "bbox": [ + 131, + 295, + 317, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r^{\\mathrm{sur}}" + }, + { + "bbox": [ + 131, + 295, + 317, + 340 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 295, + 317, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f^{\\mathrm{sur}}" + }, + { + "bbox": [ + 131, + 295, + 317, + 340 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "text", + "content": "Subsequently, SCAR splits the samples in " + }, + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{sur}}" + }, + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r^{\\mathrm{sur}}" + }, + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f^{\\mathrm{sur}}" + }, + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "text", + "content": " if respectively the predicted class " + }, + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "text", + "content": " belongs to " + }, + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{ret}}" + }, + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "text", + "content": ", and in " + }, + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f^{\\mathrm{sur}}" + }, + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "text", + "content": " is within " + }, + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{fgt}}" + }, + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "text", + "content": ". Thus, " + }, + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f^{\\mathrm{sur}}" + }, + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "text", + "content": " serves as a surrogate for " + }, + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 322, + 164, + 482, + 355 + ], + "type": "text", + "content": ", and SCAR is optimized using the loss function in Eq. 4. This configuration poses an extreme challenge as it lacks both retain and forget data during the unlearning process. SCAR self-forget marks the first attempt to address the unlearning problem within the Class-Removal (CR) scenario without necessitating training or forget data, nor the retention of specific in" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 355, + 361, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 355, + 361, + 368 + ], + "spans": [ + { + "bbox": [ + 131, + 355, + 361, + 368 + ], + "type": "text", + "content": "formation during training such as gradient updates." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 393, + 282, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 393, + 282, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 393, + 282, + 407 + ], + "type": "text", + "content": "4 Experimental Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": "We evaluated SCAR against a range of sota methods and baselines on three datasets CIFAR10 [31], CIFAR100 [31], and TinyImageNet [33]. CIFAR-10 and CIFAR-100 are comprised of 60,000 images of " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " pixels each (50,000 images for train and 10000 for test), with CIFAR-10 divided into 10 classes, and CIFAR-100 categorized into 100 classes. TinyImageNet consists of 110,000 images of " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " pixels each, divided into 200 classes (100,000 for train and 10,000 for test). The considered sota methods include DUCK [9], Boundary Shrink [4], Boundary Expanding [4], SCRUB [32], L1-Sparse [28], Bad Teacher [8], ERM-KTP [35], Negative Gradient [17], and Random Labels [23], alongside traditional approaches such as Retraining and Fine-Tuning. While the latter two are computationally demanding as they necessitate access to the " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ", they serve as critical benchmarks. Retraining is by definition the upper bound for unlearning, removing the " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " from the training process entirely. Fine-Tuning, conversely, adjusts the pre-trained model weights using only the " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ". Importantly, changing the number of epochs can substantially affect the finetuning performance: few epochs result in an efficient unlearning whereas if the model is finetuned for many epochs it is substantially retrained. For this reason, the original model is finetuned for a sufficient number of epochs (" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " epochs of retrain) that guarantees to unlearn " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " (as outlined in [4,35]). Unfortunately, for this reason, Fine-Tuning is extremely time-consuming and inefficient compared to other sota methods [4,9,35]. All" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "SCAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 146, + 481, + 205 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 1: Comparison between SCAR performance and different surrogate dataset in CR scenario. The metrics are reported as mean ± std over ten runs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 146, + 481, + 205 + ], + "lines": [ + { + "bbox": [ + 132, + 146, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 481, + 205 + ], + "type": "table", + "html": "
Surrogate dataset DsurSizeCIFAR100TinyImagenet
ArAftAUSp-valArAftAUSp-val
Random Images2*10372.20(01.90)02.60(02.17)0.922(0.027)<10-360.69(01.80)01.20(01.40)0.912(0.022)<10-3
Imagenet subset10472.93(01.78)02.00(02.00)0.935(0.025)<10-362.99(01.39)00.60(01.37)0.940(0.019)<10-3
COCO subset10471.61(01.92)00.84(00.43)0.933(0.023)<10-362.04(01.41)01.00(01.41)0.927(0.019)<10-3
Imagenet distilled10472.13(02.28)02.51(01.62)0.923(0.023)<10-362.36(01.56)01.03(01.41)0.930(0.020)<10-3
Gaussian Noise10416.62(08.64)00.70(00.67)0.388(0.086)<10-317.13(06.08)01.20(01.93)0.482(0.061)<10-3
", + "image_path": "cc54087ead269cb56ed726a66859a8c100115e2ed4883a7d9fd88d74aaa85085.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 225, + 479, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 479, + 249 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 479, + 249 + ], + "type": "text", + "content": "methods utilized resnet18 [24] as the neural architecture. The details about the hyperparameters are reported in the Supp. Material, Sec. 2, \"Reproducibility\"." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "text", + "content": "We established two experimental frameworks, CR and HR, based on the scenarios outlined in Sec. 3.1. For CR, we applied a constant seed, dividing each dataset into " + }, + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "text", + "content": " with distinct class distributions. The " + }, + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "text", + "content": " included instances from a single class, while the " + }, + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "text", + "content": " comprised the remaining classes. This setup was replicated across ten splits to calculate average metrics and standard deviations. In the HR scenario, we selected 10 different seeds for dataset partitioning, ensuring a 90:10 split between " + }, + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "text", + "content": ", with class overlaps but unique instances. We reported the mean and standard deviation of the results across these seeds. Further details are available in the Sec. 2 of the Supp. Material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 357, + 481, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 481, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 481, + 513 + ], + "type": "text", + "content": "For the CR scenario, the metrics considered are the accuracies on retain " + }, + { + "bbox": [ + 130, + 357, + 481, + 513 + ], + "type": "inline_equation", + "content": "(\\mathcal{A}_r^t)" + }, + { + "bbox": [ + 130, + 357, + 481, + 513 + ], + "type": "text", + "content": " and forget " + }, + { + "bbox": [ + 130, + 357, + 481, + 513 + ], + "type": "inline_equation", + "content": "(\\mathcal{A}_f^t)" + }, + { + "bbox": [ + 130, + 357, + 481, + 513 + ], + "type": "text", + "content": " test sets, to be maximized and minimized, respectively, and the Adaptive Unlearning Score (AUS) [9]. This final score facilitates the comparison among various methods by encapsulating forget and retain test set performances into a single metric. It captures a delicate balance between two essential objectives: preserving high test accuracy while effectively tackling the unlearning task (see Supp. Material Sec. 2). For the HR scenario, the metrics included are accuracy on " + }, + { + "bbox": [ + 130, + 357, + 481, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 357, + 481, + 513 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 357, + 481, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_f" + }, + { + "bbox": [ + 130, + 357, + 481, + 513 + ], + "type": "text", + "content": "), overall test set accuracy " + }, + { + "bbox": [ + 130, + 357, + 481, + 513 + ], + "type": "inline_equation", + "content": "(\\mathcal{A}^t)" + }, + { + "bbox": [ + 130, + 357, + 481, + 513 + ], + "type": "text", + "content": ", AUS, and the success of a membership inference attack (MIA) [9, 49, 54], with the optimal MIA value being " + }, + { + "bbox": [ + 130, + 357, + 481, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_1 = 0.5" + }, + { + "bbox": [ + 130, + 357, + 481, + 513 + ], + "type": "text", + "content": ". Specifically the number of forget and test samples are balanced and " + }, + { + "bbox": [ + 130, + 357, + 481, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_1 = 0.5" + }, + { + "bbox": [ + 130, + 357, + 481, + 513 + ], + "type": "text", + "content": " (chance level) means the impossibility to distinguish the membership of the forget data. The Sec. 2 of the Supp. Material provides details about hyperparameters for MIA tuning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 528, + 342, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 528, + 342, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 528, + 342, + 540 + ], + "type": "text", + "content": "4.1 Impact of different Datasets as " + }, + { + "bbox": [ + 132, + 528, + 342, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{sur}}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": "A critical role in SCAR is played by the distillation-trick mechanism. It allows SCAR to maintain acquired information about the retain set even without accessing it during the unlearning. The distillation-trick works using an OOD dataset, and investigating how different datasets can or cannot affect the efficiency of SCAR is of utmost importance. For this reason, following [2, 15, 27], we selected 4 different OOD external public datasets characterized by different distributions of pixels, number of images, resolution, and number of classes (a random subset of " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "10\\mathrm{K}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " images of Imagenet1K [11], a random subset of " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "10\\mathrm{K}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " images of COCO [36], a set of " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "2\\mathrm{K}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " public natural images [1], and a set of " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "10\\mathrm{K}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " images distilled with [65] from Imagenet1K). We verified through the Kolmogorov-Smirnov" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "J. Bonato et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 184, + 482, + 288 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 2: Performance comparison between SCAR and sota methods in CR scenario for CIFAR10, CIFAR100, and TinyImagenet datasets. The metrics are reported as mean ± std over ten runs. Results marked with " + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": " are taken from [9]. Columns “" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": " free” and “" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": " free” indicate whether the method was (✓) or not (✗), Retain-Free or Forget-free, respectively. Green and red colors highlight respectively the best results between methods based on retain set and between retain free methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 184, + 482, + 288 + ], + "lines": [ + { + "bbox": [ + 132, + 184, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 482, + 288 + ], + "type": "table", + "html": "
DrfreeDffreeCIFAR10CIFAR100TinyImagenet
AicAfjAUSAicAfjAUSAicAfjAUS
Original--88.64(00.63)88.34(00.62)0.531(0.005)77.55(00.11)77.50(02.80)0.563(0.009)68.26(00.08)64.60(15.64)0.607(0.058)
Retrained†--88.05(01.28)0.000(0.000)0.994(0.014)77.97(00.42)0.000(0.000)1.004(0.022)67.67(01.00)0.000(0.000)0.993(0.010)
Fine Tuning‡--87.93(01.14)0.000(0.000)0.993(0.013)77.31(02.17)0.000(0.000)0.998(0.022)67.89(00.24)0.000(0.000)0.994(0.010)
DUCK†XX85.53(01.37)0.000(0.000)0.969(0.015)71.57(02.08)0.010(0.176)0.931(0.026)61.29(01.13)0.020(0.063)0.927(0.013)
Boundary S.†XX83.81(02.29)13.64(03.57)0.840(0.543)55.07(11.85)03.54(02.26)0.749(0.116)55.98(03.33)04.25(02.12)0.840(0.036)
Boundary E.†XX82.36(02.39)13.34(03.21)0.827(0.032)55.18(11.98)03.50(02.32)0.750(0.117)55.92(03.45)02.60(02.50)0.853(0.040)
SCRUBXX87.88(01.18)0.000(0.000)0.992(0.013)77.29(00.25)02.00(05.35)0.977(0.051)68.15(00.25)01.20(03.79)0.986(0.037)
L1-SparseXX85.21(09.41)04.88(02.18)0.862(0.023)62.63(07.56)01.22(05.53)0.879(0.018)63.76(01.29)04.81(01.54)0.901(0.011)
ERM-KTP†XX80.31(-)00.00(-)0.917(-)58.96(-)00.00(-)0.814(-)49.45(-)00.00(-)0.811(-)
Bad TeacherXX88.89(00.87)02.37(05.86)0.982(0.057)77.19(00.19)06.00(10.09)0.940(0.089)64.09(00.78)11.20(09.04)0.862(0.070)
Neg.Grad.†✓X76.27(03.27)00.56(00.12)0.871(0.033)62.84(06.13)00.50(05.50)0.849(0.061)60.09(02.58)0.60(01.35)0.911(0.028)
Rand.Lab.†✓X65.46(07.59)00.83(00.44)0.762(0.076)55.31(07.06)00.40(00.70)0.774(0.070)43.29(10.10)01.20(01.03)0.740(10.100)
SCAR✓X87.71(01.62)00.97(00.24)0.981(0.017)72.93(01.78)02.00(02.00)0.935(0.025)62.99(01.39)0.60(01.37)0.940(0.019)
SCAR Self-Forget✓✓87.18(01.76)01.20(00.43)0.974(0.019)71.09(02.69)00.70(00.95)0.929(0.028)60.79(02.15)00.80(01.40)0.917(0.025)
", + "image_path": "62afdec70e98e387f2a532b351c728aaa91e5081a0a513370bb51b56ed3fea10.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 312, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 515 + ], + "type": "text", + "content": "(KS) test [40] that the distributions of pixels of these surrogate datasets are different from the datasets used for benchmarking SCAR (Tab. 1 p-val columns). This is fundamental to assess the generality of SCAR and it does not rely on data leakage between datasets; this is especially true for tinyImagenet and the random subset ofImagenet1K where we observed the KS p-val is " + }, + { + "bbox": [ + 130, + 312, + 482, + 515 + ], + "type": "inline_equation", + "content": "< 0.001" + }, + { + "bbox": [ + 130, + 312, + 482, + 515 + ], + "type": "text", + "content": ". Moreover, we also used a surrogate dataset made of Gaussian noise images where the semantic information was absent. We performed our experiments in the CR scenario using CIFAR100 as " + }, + { + "bbox": [ + 130, + 312, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 312, + 482, + 515 + ], + "type": "text", + "content": ". Both accuracies and AUS scores were statistically comparable across datasets (Tab. 1) even if theImagenet1K subset gives the overall highest results. Importantly, when in the surrogate dataset semantic information is removed (Gaussian Noise) SCAR performance is poor, suggesting how the surrogate dataset has to contain some spatial feature and patterns to trigger the DNN filters. Hence, the presence of semantic information content in images plays a pivotal role in the distillation-trick mechanism, but, at the same time, the differences in terms of kind of information (e.g. classes and distribution of pixels) do not affect this technique. For the rest of the paper results are presented using as a surrogate the OOD random subset ofImagenet1K." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 536, + 318, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 318, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 318, + 547 + ], + "type": "text", + "content": "4.2 Comparison with sota methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "We evaluated SCAR, and SCAR Self-Forget, within the CR scenario across three datasets (Tab. 2). Our findings indicate that our method surpasses Retain-Free approaches like Negative Gradient and Random Label in performance. Furthermore, SCAR's performance is on par with methods that leverage both forget and retain sets. Notably, SCAR's performance is lower than Fine-Tuning and within 1 std. from SCRUB in CIFAR100 and TinyImagenet, though these methods depend heavily on the " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " requiring several finetuning steps on it. Importantly, these outcomes are also achieved with SCAR Self-Forget, which remarkably does not utilize any data from the original training set." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "SCAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 185, + 481, + 259 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 3: Performance comparison between SCAR and sota methods in HR scenario for CIFAR10, CIFAR100, and TinyImagenet datasets. The metrics are reported as mean ± std over ten runs. Results marked with " + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": " are taken from [9]. Columns \"D_r free\" indicate whether the method was (✓) or not (X), Retain-Free. Green and red colors highlight respectively the best results between methods based on retain set and between retain free methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 185, + 481, + 259 + ], + "lines": [ + { + "bbox": [ + 133, + 185, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 481, + 259 + ], + "type": "table", + "html": "
DfreeCIFAR10CIFAR100TinyImagenet
A1AfF1AUSA1AfF1AUSA1AfF1AUS
Original-88.54(00.25)99.49(00.08)54.62(00.80)0.900(0.004)77.23(00.53)99.63(00.10)61.80(00.59)0.814(0.006)68.08(00.39)84.83(00.44)54.06(01.48)0.854(0.007)
Retrained†-84.13(00.59)84.56(00.72)50.22(00.80)0.950(0.011)77.27(01.03)76.87(00.95)49.45(00.40)0.993(0.017)63.45(02.34)63.24(02.58)49.56(00.47)0.950(0.040)
Fine Tuning†-85.70(00.48)88.66(00.44)50.06(01.30)0.942(0.008)72.06(00.52)74.97(00.72)49.68(00.43)0.918(0.009)67.03(00.45)70.52(00.65)49.86(00.56)0.937(0.011)
DUCK†X85.96(00.64)86.05(00.55)50.26(00.49)0.972(0.011)74.74(00.74)75.51(02.01)50.66(00.64)0.965(0.022)62.01(00.79)64.43(00.76)49.87(00.73)0.916(0.014)
SCRBX78.02(07.80)82.53(10.80)50.59(00.89)0.854(0.132)73.33(00.85)92.46(01.50)56.45(01.48)0.804(0.014)68.52(00.40)83.14(00.53)53.89(00.96)0.875(0.008)
L1-SparseX93.05(00.53)99.06(00.65)52.21(00.95)0.927(0.011)52.26(01.65)66.89(02.34)53.46(00.94)0.690(0.024)63.76(01.19)84.30(01.04)53.29(00.73)0.794(0.032)
Bad TeacherX86.66(00.44)92.91(00.57)50.26(00.87)0.922(0.008)73.57(00.61)86.57(00.57)54.44(01.81)0.851(0.006)67.41(00.49)80.28(00.81)54.83(01.44)0.876(0.006)
Neg. Grad.†✓79.35(00.85)87.11(00.86)51.74(00.54)0.841(0.012)60.83(00.77)76.77(00.57)50.98(00.47)0.718(0.009)54.83(00.81)67.27(00.59)49.78(00.33)0.770(0.011)
Rand. Lab.†✓77.28(01.37)87.07(01.20)51.45(00.79)0.807(0.018)60.51(00.82)77.18(00.50)50.55(00.49)0.711(0.009)54.34(00.85)67.65(00.45)49.84(00.64)0.760(0.011)
SCAR✓86.14(00.47)88.40(00.31)50.07(00.59)0.953(0.008)73.23(00.74)75.63(00.54)49.77(00.28)0.934(0.011)61.35(00.69)66.51(00.44)49.85(00.19)0.886(0.011)
", + "image_path": "c6ee8929e6aa8a2f73ac659b0f43296a7b38fc4d1c66f53a77516ca6dfa86021.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 159, + 296, + 455, + 392 + ], + "blocks": [ + { + "bbox": [ + 130, + 269, + 481, + 292 + ], + "lines": [ + { + "bbox": [ + 130, + 269, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 481, + 292 + ], + "type": "text", + "content": "Table 4: Results of the ablation study on CIFAR100 in CR and HR scenarios. The metrics for CR and HR are reported as mean ± std over ten runs." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 296, + 455, + 392 + ], + "lines": [ + { + "bbox": [ + 159, + 296, + 455, + 392 + ], + "spans": [ + { + "bbox": [ + 159, + 296, + 455, + 392 + ], + "type": "table", + "html": "
LTDLMCIFAR100TinyImagenet
ArAfAUSArAfAUS
CRXX77.55(00.11)77.50(02.80)0.563(0.009)68.26(00.08)64.60(15.64)0.607(0.058)
✓X72.09(02.72)40.00(13.36)0.675(0.067)63.14(01.77)29.80(12.09)0.730(0.069)
X✓66.90(05.03)02.60(02.27)0.871(0.053)49.72(02.54)00.80(01.40)0.807(0.028)
✓✓72.93(01.78)02.00(02.00)0.935(0.025)62.99(01.39)00.60(01.37)0.940(0.019)
AtAfAUSAtAfAUS
HRXX77.39(00.46)99.64(00.09)0.817(0.005)68.08(00.39)84.83(00.44)0.854(0.007)
✓X76.53(00.41)99.18(00.14)0.807(0.005)64.84(00.60)80.51(00.51)0.835(0.009)
X✓18.83(01.06)19.69(01.19)0.409(0.013)05.88(02.94)05.51(02.63)0.376(0.033)
✓✓73.23(00.74)75.63(00.54)0.934(0.011)61.35(00.69)66.51(00.44)0.886(0.011)
", + "image_path": "29bc71f939f68ccb0143babfa36e45ba9ea58ae0bfd7fe5a05bfacf90982db3d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 414, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 485 + ], + "type": "text", + "content": "We also extended our analysis to the HR scenario (Tab. 3). Importantly, the Self-Forget variant of our method is not applicable in HR, as the " + }, + { + "bbox": [ + 130, + 414, + 482, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 414, + 482, + 485 + ], + "type": "text", + "content": " is essential to identify the specific instance to forget. In this more demanding context, SCAR surpasses all the performance of Retain-Free approaches. Between the methods that use the " + }, + { + "bbox": [ + 130, + 414, + 482, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 414, + 482, + 485 + ], + "type": "text", + "content": " only DUCK in CIFAR100 (within 1 std.) and TinyImagenet and Fine-Tuning in TinyImagenet achieve superior results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "text", + "content": "Notably, our method outperforms the Retain-Free approaches in both scenarios and demonstrates commendable performance relative to methods that rely on the " + }, + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "text", + "content": ". This distinction highlights the effectiveness and versatility of our approach, making it a viable alternative for scenarios where the " + }, + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 486, + 482, + 558 + ], + "type": "text", + "content": " is entirely unavailable. We also confirmed our result using the COCO dataset as a surrogate dataset (Sec. 1 of the Supp. Material)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 574, + 239, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 239, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 239, + 586 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We conducted an ablation study to evaluate the impact of different loss components " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LD}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_M" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " of our method on the unlearning performance across CR and HR scenarios. The results of this study are presented in Tab. 4. In CR, considering only " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LD}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " leads to similar retain set accuracy but with high " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " accuracy. In fact, the stopping criteria is hardly fulfilled and the method stops at the maximum number of epochs allowed. Conversely, considering only " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_M" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "J. Bonato et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 157, + 141, + 454, + 189 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 5: Results of SCAR changing the measure used in " + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_M" + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": " on the CIFAR100. The metrics for CR and HR are reported as mean " + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "\\pm" + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": " std over ten runs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 141, + 454, + 189 + ], + "lines": [ + { + "bbox": [ + 157, + 141, + 454, + 189 + ], + "spans": [ + { + "bbox": [ + 157, + 141, + 454, + 189 + ], + "type": "table", + "html": "
CIFAR100 CRCIFAR100 HR
ArrAftAUSArrAftAUS
Cosine Similarity72.74(02.19)02.00(02.36)0.933(0.031)73.88(00.46)77.05(00.54)0.934(0.008)
L2 Distance72.45(02.23)02.70(02.54)0.924(0.032)69.58(00.82)76.93(00.55)0.857(0.011)
Mahalanobis72.93(01.78)02.00(02.00)0.935(0.025)73.23(00.74)75.63(00.54)0.934(0.011)
", + "image_path": "83f6cd007b2b5a574d42d2053216e1ea9681be5fd4b02fa386deabc8fb015779.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 212, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 320 + ], + "type": "text", + "content": "methods can push down the accuracy on the " + }, + { + "bbox": [ + 130, + 212, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 212, + 482, + 320 + ], + "type": "text", + "content": ", but at the cost of a low retain set accuracy. In the HR scenario, we observed similar but more pronounced results. The presence of " + }, + { + "bbox": [ + 130, + 212, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LD}" + }, + { + "bbox": [ + 130, + 212, + 482, + 320 + ], + "type": "text", + "content": " alone is insufficient for effectively erasing knowledge about forget samples, as evidenced by a high forget accuracy. Employing only " + }, + { + "bbox": [ + 130, + 212, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LD}" + }, + { + "bbox": [ + 130, + 212, + 482, + 320 + ], + "type": "text", + "content": " results in a steep decrease in both accuracies. This result suggests how the unlearning strategy without a method for retaining knowledge can't be used to unlearn. This effect is more evident in HR because images from the same class are distributed across both the " + }, + { + "bbox": [ + 130, + 212, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 212, + 482, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 212, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 130, + 212, + 482, + 320 + ], + "type": "text", + "content": ", meaning that the " + }, + { + "bbox": [ + 130, + 212, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LD}" + }, + { + "bbox": [ + 130, + 212, + 482, + 320 + ], + "type": "text", + "content": " compromises the general knowledge of the classes." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 147, + 320, + 294, + 416 + ], + "blocks": [ + { + "bbox": [ + 147, + 320, + 294, + 416 + ], + "lines": [ + { + "bbox": [ + 147, + 320, + 294, + 416 + ], + "spans": [ + { + "bbox": [ + 147, + 320, + 294, + 416 + ], + "type": "image", + "image_path": "863d2b9093ed822a2889d3f27a4101a611ed54d4561b7aa209158e3ed304aff7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 423, + 299, + 479 + ], + "lines": [ + { + "bbox": [ + 146, + 423, + 299, + 479 + ], + "spans": [ + { + "bbox": [ + 146, + 423, + 299, + 479 + ], + "type": "text", + "content": "Fig. 4: Values of final AUS for SCAR and SCAR self-forget as a function of the number of samples available in " + }, + { + "bbox": [ + 146, + 423, + 299, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{sur}}" + }, + { + "bbox": [ + 146, + 423, + 299, + 479 + ], + "type": "text", + "content": " (subset of Imagenet1K). AUS is reported as mean " + }, + { + "bbox": [ + 146, + 423, + 299, + 479 + ], + "type": "inline_equation", + "content": "\\pm" + }, + { + "bbox": [ + 146, + 423, + 299, + 479 + ], + "type": "text", + "content": " std over ten runs" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 324, + 324, + 463, + 415 + ], + "blocks": [ + { + "bbox": [ + 324, + 324, + 463, + 415 + ], + "lines": [ + { + "bbox": [ + 324, + 324, + 463, + 415 + ], + "spans": [ + { + "bbox": [ + 324, + 324, + 463, + 415 + ], + "type": "table", + "html": "
CIFAR100
AfAfAUS
Original (AliCNN)70.08(00.15)69.70(14.83)0.545(0.048)
Unlearned65.35(02.37)02.90(03.38)0.926(0.036)
Original (Resnet18)77.55(00.11)77.50(02.80)0.563(0.009)
Unlearned72.93(01.78)02.00(02.00)0.935(0.025)
Original (Resnet34)80.13(00.10)80.30(10.02)0.569(0.031)
Unlearned75.67(01.74)01.90(00.92)0.937(0.019)
Original (Resnet50)80.64(00.10)80.80(09.50)0.570(0.030)
Unlearned75.99(01.77)00.80(02.36)0.942(0.028)
Original (ViT-B16)85.51(01.45)83.30(06.24)0.546(0.023)
Unlearned81.99(02.59)02.04(03.10)0.945(0.031)
", + "image_path": "a7a14f3b6e2ed6ea4a704b327f4478145d32ec11e480fee12e3c8923a8b1fbfb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 423, + 479, + 479 + ], + "lines": [ + { + "bbox": [ + 313, + 423, + 479, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 479, + 479 + ], + "type": "text", + "content": "Table 6: Results of SCAR on CIFAR100 dataset in the CR scenario changing the model to unlearn. The metrics are reported as mean ± std over ten runs before and after the unlearning." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 131, + 498, + 453, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 453, + 522 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 453, + 522 + ], + "type": "text", + "content": "4.4 Analysis on the effect of different measures in the metric learning mechanism of SCAR" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "We report in Tab. 5 and Supp. Tab. 2 the results of the SCAR using different distance measures for the " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_M" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " component across CIFAR100 and TinyImagenet datasets in both the CR and HR scenarios. Using the Mahalanobis distance yields a perfect balance, achieving the highest AUS scores in both CR and HR. Cosine Similarity distance achieves compatible results (even if constantly lower) in CR and HR scenarios but at the cost of being " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "(22 \\pm 5\\%)" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " slower than the Mahalanobis. This result suggests how the information about the distributions of samples embedded into the Mahalanobis distance is fundamental to guide the optimization efficiently. In contrast, the " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " distance resulted in lower performance in both CR and HR scenarios suggesting how this metric hampers the unlearning process." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "SCAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 400, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 400, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 400, + 128 + ], + "type": "text", + "content": "4.5 Analysis of the impact of the dimension of " + }, + { + "bbox": [ + 132, + 116, + 400, + 128 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{sur}}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 278 + ], + "type": "text", + "content": "Fig. 4 illustrates the relationship between the number of images in the surrogate dataset " + }, + { + "bbox": [ + 130, + 134, + 482, + 278 + ], + "type": "inline_equation", + "content": "D_{sur}" + }, + { + "bbox": [ + 130, + 134, + 482, + 278 + ], + "type": "text", + "content": " (subset of ImageNet1K) and the AUS for both SCAR and its self-forget variant, in the CR scenario. SCAR maintains a relatively stable AUS score as the number of images increases, suggesting that the method's unlearning ability does not significantly depend on the size of the surrogate dataset. The self-forget variant of SCAR shows a plateau in AUS score when at least 6K surrogate images are used which suggests how the dimension of " + }, + { + "bbox": [ + 130, + 134, + 482, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{sur}}" + }, + { + "bbox": [ + 130, + 134, + 482, + 278 + ], + "type": "text", + "content": " can affect the self-forget mechanism. Importantly, the self-forget mechanism that can access 6K surrogate samples or more exhibits performance comparable with the method that can access forget data. This pattern reflects the effectiveness of SCAR in handling the unlearning process, with only minor improvements seen with the addition of more surrogate data when forget data are not available." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 295, + 317, + 307 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 295, + 317, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 295, + 317, + 307 + ], + "type": "text", + "content": "4.6 Architectural-Agnostic Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 313, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 482, + 422 + ], + "type": "text", + "content": "For additional insight, we examined the performance of SCAR (Tab. 6) across various DNN architectures within the CR scenario. We investigated the effect of the unlearning with SCAR on a basic CNN network (AllCNN), ResNet architectures (18, 34, and 50), and a Vision Transformer (ViT-B16). Our method consistently demonstrates strong performance across all tested architectures. Notably, the results achieved with the ViT confirm the model-agnostic nature of our approach beyond the CNN architectures. It is worth mentioning that the efficacy of our method is not contingent on the initial accuracy of the model, which underscores its adaptability and broad applicability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 437, + 226, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 226, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 226, + 450 + ], + "type": "text", + "content": "5 Conclusions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 481, + 605 + ], + "type": "text", + "content": "In this paper, we introduced SCAR a novel method for approximate unlearning, designed to effectively erase information about the forget-set data while maintaining the original model performance. Our experimental results indicate that SCAR outperforms the methods that do not require access to the retain set and achieves performance levels comparable to, or even surpassing, those of existing approximate unlearning methods that depend on the retain set, across both CR and HR scenarios. Furthermore, we have developed a self-forget variation of SCAR capable of delivering impressive results in CR without accessing both the retain and forget sets. Future work should include a mathematical exploration of the algorithms' certifiability and the development of alternatives to the distillation-trick, enabling the retention of past knowledge without the need for a surrogate dataset, which can limit the application of SCAR to other domain." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 606, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 481, + 665 + ], + "type": "text", + "content": "Despite these challenges, we firmly believe that our algorithm, that do not require access to any retain set, can significantly impact society: for example in the performance restoration of models poisoned by corrupted data or in recognition systems (such as facial recognition), by enabling individuals to exercise their right to have their data removed from the model's knowledge." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "J. Bonato et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 137, + 481, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 481, + 169 + ], + "type": "text", + "content": "1. Image captioning dataset, random images. https://www.kaggle.com/datasets/shamsaddin97/image-captioning-dataset-random-images/data, accessed: 2024-02-14 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 170, + 481, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 170, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 138, + 170, + 481, + 201 + ], + "type": "text", + "content": "2. Bibas, K., Feder, M., Hassner, T.: Single layer predictive normalized maximum likelihood for out-of-distribution detection. Advances in Neural Information Processing Systems 34, 1179-1191 (2021) 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 202, + 481, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 202, + 481, + 234 + ], + "spans": [ + { + "bbox": [ + 138, + 202, + 481, + 234 + ], + "type": "text", + "content": "3. Bourtoule, L., Chandrasekaran, V., Choquette-Choo, C.A., Jia, H., Travers, A., Zhang, B., Lie, D., Papernot, N.: Machine unlearning. In: 2021 IEEE Symposium on Security and Privacy (SP). pp. 141-159. IEEE (2021) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 235, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 235, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 235, + 481, + 277 + ], + "type": "text", + "content": "4. Chen, M., Gao, W., Liu, G., Peng, K., Wang, C.: Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7766-7775 (June 2023) 4, 5, 8, 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 277, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 481, + 308 + ], + "type": "text", + "content": "5. Chen, S., Kahla, M., Jia, R., Qi, G.J.: Knowledge-enriched distributional model inversion attacks pp. 16158-16167 (2021). https://doi.org/10.1109/ICCV48922.2021.015871" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 309, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 481, + 342 + ], + "type": "text", + "content": "6. Chen, Y., Wiesel, A., Eldar, Y.C., Hero, A.O.: Shrinkage algorithms for mmse covariance estimation. IEEE Transactions on Signal Processing 58, 5016-5029 (2009), https://api-semanticscholar.org/CorpusID:12616083 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 342, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 481, + 363 + ], + "type": "text", + "content": "7. Cheng, J., Amiri, H.: Multimodal machine unlearning. arXiv preprint arXiv:2311.12047 (2023) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 363, + 481, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 363, + 481, + 406 + ], + "spans": [ + { + "bbox": [ + 138, + 363, + 481, + 406 + ], + "type": "text", + "content": "8. Chandawat, V.S., Tarun, A.K., Mandal, M., Kankanhalli, M.: Can bad teaching induce forgetting? unlearning in deep networks using an incompetent teacher. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 7210-7217 (2023) 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 407, + 481, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 481, + 438 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 481, + 438 + ], + "type": "text", + "content": "9. Cotogni, M., Bonato, J., Sabetta, L., Pelosiin, F., Nicolosi, A.: Duck: Distance-based unlearning via centroid kinematics. arXiv preprint arXiv:2312.02052 (2023) 4, 5, 8, 9, 10, 11, 12" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 439, + 481, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 439, + 481, + 469 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 481, + 469 + ], + "type": "text", + "content": "0. De Lange, M., Tuytelaars, T.: Continual prototype evolution: Learning online from non-stationary data streams (2021). https://doi.org/10.1109/ICCV48922.2021.008146" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 471, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 471, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 471, + 481, + 503 + ], + "type": "text", + "content": "1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 503, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 481, + 534 + ], + "type": "text", + "content": "2. Du, M., Yang, F., Zou, N., Hu, X.: Fairness in deep learning: A computational perspective. IEEE Intelligent Systems 36(4), 25-34 (2021). https://doi.org/10.1109/MIS.2020.3000681 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 535, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 535, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 535, + 481, + 600 + ], + "type": "text", + "content": "3. Fang, G., Bao, Y., Song, J., Wang, X., Xie, D., Shen, C., Song, M.: Mosaicking to distill: Knowledge distillation from out-of-domain data. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 11920-11932. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/63dc7ed1010d3c3b8269faf0ba7491d4-Paper.pdf 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "type": "text", + "content": "4. Felps, D.L., Schwickerath, A.D., Williams, J.D., Vuong, T.N., Briggs, A., Hunt, M., Sakmar, E., Saranchak, D.D., Shumaker, T.: Class clown: Data redaction in machine unlearning at enterprise scale. arXiv preprint arXiv:2012.04699 (2020) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "5. Fort, S., Ren, J., Lakshminarayanan, B.: Exploring the limits of out-of-distribution detection. Advances in Neural Information Processing Systems 34, 7068-7081 (2021) 10" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "SCAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "16. Fredrikson, M., Jha, S., Ristenpart, T.: Model inversion attacks that exploit confidence information and basic countermeasures. Proceedings of the 22nd ACM SIGSAC Conference on Computer and Communications Security (2015), https://apisemantic scholar.org/CorpusID:207229839 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 193 + ], + "type": "text", + "content": "17. Golatkar, A., Achille, A., Soatto, S.: Eternal sunshine of the spotless net: Selective forgetting in deep networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9304-9312 (2020) 3, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 481, + 236 + ], + "type": "text", + "content": "18. Golatkar, A., Achille, A., Soatto, S.: Forgetting outside the box: Scrubbing deep networks of information accessible from input-output observations. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIX 16. pp. 383-398. Springer (2020) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 236, + 481, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 481, + 266 + ], + "type": "text", + "content": "19. Graves, L., Nagisetty, V., Ganesh, V.: Amnesiac machine learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 11516-11524 (2021) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 267, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 481, + 311 + ], + "type": "text", + "content": "20. Guo, C., Goldstein, T., Hannun, A., Van Der Maaten, L.: Certified data removal from machine learning models. In: Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 3832-3842. PMLR (2020) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 311, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 481, + 343 + ], + "type": "text", + "content": "21. Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: International conference on machine learning. pp. 1321-1330. PMLR (2017) 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 343, + 481, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 481, + 387 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 481, + 387 + ], + "type": "text", + "content": "22. Han, X., Dai, Y., Gao, T., Lin, Y., Liu, Z., Li, P., Sun, M., Zhou, J.: Continual relation learning via episodic memory activation and reconsolidation (2020). https://doi.org/10.18653/v1/2020.acl-main.573, https://aclanthology.org/2020.acl-main.573 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 387, + 481, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 481, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 481, + 408 + ], + "type": "text", + "content": "23. Hayase, T., Yasutomi, S., Katoh, T.: Selective forgetting of deep networks at a finer level than samples. arXiv preprint arXiv:2012.11849 (2020) 3, 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 408, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 481, + 440 + ], + "type": "text", + "content": "24. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 10" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 440, + 481, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 481, + 473 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 481, + 473 + ], + "type": "text", + "content": "25. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: International Conference on Learning Representations (2017), https://openreview.net/forum?id=Hkg4TI9x1 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 472, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 493 + ], + "type": "text", + "content": "26. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 493, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 481, + 536 + ], + "type": "text", + "content": "27. Hsu, Y.C., Shen, Y., Jin, H., Kira, Z.: Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10951-10960 (2020) 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "text", + "content": "28. Jia, J., Liu, J., Ram, P., Yao, Y., Liu, G., Liu, Y., Sharma, P., Liu, S.: Model sparsity can simplify machine unlearning. In: Annual Conference on Neural Information Processing Systems (2023) 4, 9" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 568, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 611 + ], + "type": "text", + "content": "29. Kahla, M., Chen, S., Just, H.A., Jia, R.: Label-only model inversion attacks via boundary repulsion. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 15025-15033 (2022). https://doi.org/10.1109/CVPR52688.2022.014621" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "30. Kim, J., Woo, S.S.: Efficient two-stage model retraining for machine unlearning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4361-4369 (2022) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "31. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) 9" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "J. Bonato et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 159 + ], + "type": "text", + "content": "32. Kurmanji, M., Triantafillou, P., Hayes, J., Triantafillou, E.: Towards unbounded machine unlearning. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=OveBaTtUAT 4, 5, 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 184 + ], + "type": "text", + "content": "33. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015) 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 217 + ], + "type": "text", + "content": "34. Leibig, C., Allken, V., Ayhan, M.S., Berens, P., Wahl, S.: Leveraging uncertainty information from deep neural networks for disease detection. Scientific reports 7(1), 17816 (2017) 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 261 + ], + "type": "text", + "content": "35. Lin, S., Zhang, X., Chen, C., Chen, X., Susilo, W.: Erm-ktp: Knowledge-level machine unlearning via knowledge transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20147-20155 (2023) 3, 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 481, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 306 + ], + "type": "text", + "content": "36. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) 10" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "text", + "content": "37. Liu, J., Xue, M., Lou, J., Zhang, X., Xiong, L., Qin, Z.: Mutter: Machine unlearning on adversarially trained models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4892-4902 (October 2023) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 342, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 373 + ], + "type": "text", + "content": "38. Liu, X., Xie, L., Wang, Y., Zou, J., Xiong, J., Ying, Z., Vasilakos, A.V.: Privacy and security issues in deep learning: A survey. IEEE Access 9, 4566-4593 (2021), https://api-semanticscholar.org/CorpusID:231423554 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 375, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 439 + ], + "type": "text", + "content": "39. Magdziarczyk, M.: Right to be forgotten in light of regulation (eu) 2016/679 of the european parliament and of the council of 27 april 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing directive 95/46/ec. In: 6th International Multidisciplinary Scientific Conference on Social Sciences and Art Sgem 2019. pp. 177-184 (2019) 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 441, + 481, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 481, + 462 + ], + "type": "text", + "content": "40. Massey Jr, F.J.: The kolmogorov-smirnov test for goodness of fit. Journal of the American statistical Association 46(253), 68-78 (1951) 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 464, + 481, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 495 + ], + "type": "text", + "content": "41. Mercuri, S., Khraishi, R., Okhrati, R., Batra, D., Hamill, C., Ghasempour, T., Nowlan, A.: An introduction to machine unlearning. arXiv preprint arXiv:2209.00939 (2022) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 497, + 481, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 481, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 481, + 529 + ], + "type": "text", + "content": "42. Neel, S., Roth, A., Sharifi-Malvajerdi, S.: Descent-to-delete: Gradient-based methods for machine unlearning. In: Algorithmic Learning Theory. pp. 931-962. PMLR (2021) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 531, + 481, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 481, + 563 + ], + "type": "text", + "content": "43. Nguyen, A., Yosinski, J., Clune, J.: Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 427-436 (2015) 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 565, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 586 + ], + "type": "text", + "content": "44. Nguyen, T.T., Huynh, T.T., Nguyen, P.L., Liew, A.W.C., Yin, H., Nguyen, Q.V.H.: A survey of machine unlearning. arXiv preprint arXiv:2209.02299 (2022) 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "type": "text", + "content": "45. Pardau, S.L.: The california consumer privacy act: Towards a european-style privacy regime in the united states. J. Tech. L. & Pol'y 23, 68 (2018) 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 610, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 642 + ], + "type": "text", + "content": "46. Poppi, S., Poppi, T., Cocchi, F., Cornia, M., Baraldi, L., Cucchiara, R.: Removing NSFW Concepts from Vision-and-Language Models for Text-to-Image Retrieval and Generation. arXiv preprint arXiv:2311.16254 (2023) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "47. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning (2017). https://doi.org/10.1109/CVPR.2017.587 6" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "SCAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "48. Rigaki, M., Garcia, S.: A survey of privacy attacks in machine learning. ACM Computing Surveys 56(4), 1-34 (Nov 2023). https://doi.org/10.1145/3624010, http://dx.doi.org/10.1145/3624010 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 193 + ], + "type": "text", + "content": "49. Salem, A., Zhang, Y., Humbert, M., Berrang, P., Fritz, M., Backes, M.: Ml-leaks: Model and data independent membership inference attacks and defenses on machine learning models. In: Proceedings of the 26th Annual Network and Distributed System Security Symposium (NDSS) (2019) 1, 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 193, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 481, + 224 + ], + "type": "text", + "content": "50. Setlur, A., Eysenbach, B., Smith, V., Levine, S.: Adversarial unlearning: Reducing confidence along adversarial directions. Advances in Neural Information Processing Systems 35, 18556-18570 (2022) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 224, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 481, + 246 + ], + "type": "text", + "content": "51. Shaik, T.B., Tao, X., Xie, H., Li, L., Zhu, X., Li, Q.: Exploring the landscape of machine unlearning: A comprehensive survey and taxonomy. CoRR (2023) 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "text", + "content": "52. Shokri, R., Stronati, M., Song, C., Shmatikov, V.: Membership inference attacks against machine learning models. 2017 IEEE Symposium on Security and Privacy (SP) pp. 3-18 (2017) 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 278, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 300 + ], + "type": "text", + "content": "53. Sommer, D.M., Song, L., Wagh, S., Mittal, P.: Towards probabilistic verification of machine unlearning. arXiv preprint arXiv:2003.04247 (2020) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 320 + ], + "type": "text", + "content": "54. Song, L., Mittal, P.: Systematic evaluation of privacy risks of machine learning models (2020) 10" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 320, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 481, + 363 + ], + "type": "text", + "content": "55. Song, L., Shokri, R., Mittal, P.: Privacy risks of securing machine learning models against adversarial examples. Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security (2019), https://api-semanticscholar.org/CorpusID:165163934 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 481, + 396 + ], + "type": "text", + "content": "56. Tarun, A.K., Chandawat, V.S., Mandal, M., Kankanhalli, M.: Fast yet effective machine unlearning. IEEE Transactions on Neural Networks and Learning Systems (2023) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 396, + 481, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 428 + ], + "type": "text", + "content": "57. Tiwary, P., Guha, A., Panda, S., et al.: Adapt then unlearn: Exploiting parameter space semantics for unlearning in generative adversarial networks. arXiv preprint arXiv:2309.14054 (2023) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 428, + 422, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 428, + 422, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 422, + 438 + ], + "type": "text", + "content": "58. Tukey, J.W.: Exploratory Data Analysis. Addison-Wesley (1977) 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 438, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 438, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 438, + 481, + 482 + ], + "type": "text", + "content": "59. Wang, K.C., FU, Y., Li, K., Khisti, A., Zemel, R., Makhzani, A.: Variational model inversion attacks. Advances in Neural Information Processing Systems 34, 9706-9719 (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/50a074e6a8da4662ae0a29edde722179-Paper.pdf 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 482, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 481, + 514 + ], + "type": "text", + "content": "60. Wu, G., Hashemi, M., Srinivasa, C.: Puma: Performance unchanged model augmentation for training data removal. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 8675-8682 (2022) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 514, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 481, + 546 + ], + "type": "text", + "content": "61. Wu, Y., Dobriban, E., Davidson, S.: Deltagrad: Rapid retraining of machine learning models. In: International Conference on Machine Learning. pp. 10355-10366. PMLR (2020) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 546, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 481, + 568 + ], + "type": "text", + "content": "62. Xu, J., Wu, Z., Wang, C., Jia, X.: Machine unlearning: Solutions and challenges. arXiv preprint arXiv:2308.07061 (2023) 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 568, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 620 + ], + "type": "text", + "content": "63. Yang, Z., Zhang, J., Chang, E.C., Liang, Z.: Neural network inversion in adversarial setting via background knowledge alignment. Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security p. 225-240 (2019). https://doi.org/10.1145/3319535.3354261, https://doi.org/10.1145/3319535.3354261 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "64. Yeom, S., Giacomelli, I., Fredrikson, M., Jha, S.: Privacy risk in machine learning: Analyzing the connection to overfitting. 2018 IEEE 31st Computer Security Foundations Symposium (CSF) pp. 268-282 (2017), https://api-semanticscholar.org/CorpusID:2656445 1" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "J. Bonato et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 249 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "65. Yin, Z., Xing, E., Shen, Z.: Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. Advances in Neural Information Processing Systems 36 (2024) 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 204 + ], + "type": "text", + "content": "66. Zhang, Y., Jia, R., Pei, H., Wang, W., Li, B., Song, D.: The secret revealer: Generative model-inversion attacks against deep neural networks. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 250-258 (jun 2020). https://doi.org/10.1109/CVPR42600.2020.00033, https://doi.ieeecomputersociety.org/10.1109/CVPR42600.2020.00033 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 482, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 249 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 249 + ], + "type": "text", + "content": "67. Zhao, X., Zhang, W., Xiao, X., Lim, B.: Exploiting explanations for model inversion attacks. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 662-672 (oct 2021). https://doi.org/10.1109/ICCV48922.2021.00072, https://doi.ieeeccomputersociety.org/10.1109/ICCV48922.2021.00072 1" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "SCAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Is user feedback always informative_ Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data/0589fa53-73c5-413a-aac9-f37670a1de05_content_list.json b/2024/Is user feedback always informative_ Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data/0589fa53-73c5-413a-aac9-f37670a1de05_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5f15ae8027ba135540ccf844f2d0d272691f8aa3 --- /dev/null +++ b/2024/Is user feedback always informative_ Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data/0589fa53-73c5-413a-aac9-f37670a1de05_content_list.json @@ -0,0 +1,1790 @@ +[ + { + "type": "text", + "text": "Is user feedback always informative? Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data", + "text_level": 1, + "bbox": [ + 228, + 141, + 774, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junha Song $^{1,2}$ , Tae Soo Kim $^{2}$ , Junha Kim $^{2}$ , Gunhee Nam $^{2}$ , Thijs Kooi $^{2}$ , and Jaegul Choo $^{1}$", + "bbox": [ + 263, + 233, + 738, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 KAIST {sb020518, jchoo}@kaist.ac.kr 2 Lunit Inc. {junha.kim, taesoo.kim, ghnam, tkooi}@lunit.io", + "bbox": [ + 277, + 276, + 723, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. This paper aims to adapt the source model to the target environment, leveraging small user feedback (i.e., labeled target data) readily available in real-world applications. We find that existing semi-supervised domain adaptation (SemiSDA) methods often suffer from poorly improved adaptation performance when directly utilizing such feedback data, as shown in Figure 1. We analyze this phenomenon via a novel concept called Negatively Biased Feedback (NBF), which stems from the observation that user feedback is more likely for data points where the model produces incorrect predictions. To leverage this feedback while avoiding the issue, we propose a scalable adapting approach, Retrieval Latent Defending. This approach helps existing SemiSDA methods to adapt the model with a balanced supervised signal by utilizing latent defending samples throughout the adaptation process. We demonstrate the problem caused by NBF and the efficacy of our approach across various benchmarks, including image classification, semantic segmentation, and a real-world medical imaging application. Our extensive experiments reveal that integrating our approach with multiple state-of-the-art SemiSDA methods leads to significant performance improvements.", + "bbox": [ + 261, + 342, + 743, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Rethinking user-provided feedback $\\cdot$ Semi-supervised & Source-free domain adaptation $\\cdot$ Medical image diagnosis", + "bbox": [ + 259, + 604, + 720, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 662, + 375, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While deep neural networks have demonstrated remarkable performance in the development domain (i.e., source domain) [15,23], they often suffer from performance degradation in the deployed domain (i.e., target domain) due to domain shift [17, 72, 78]. To mitigate this issue, domain adaptation (DA) techniques have been introduced [34, 58, 70]. The most common DA tasks include semi-supervised domain adaptation (SemiSDA) and source-free domain adaptation (SFDA). SemiSDA aims to adapt the model given a small amount of labeled target data along with massive unlabeled target data [6, 58, 66, 99]. SFDA conducts adaptation with only target data without accessing source data considering data privacy or memory constraints in edge devices [34, 67, 92].", + "bbox": [ + 212, + 688, + 787, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2665f627bd6c298bd33a6f9b2a5cd19d930495806a579cc983bb24a3ecd883a.jpg", + "image_caption": [ + "Fig. 1: (a) User feedback. Users can provide feedback while interacting with an ML product, where feedback is likely to be biased towards misclassified samples, which we define as Negatively Biased Feedback (NBF). (b) Adaptation results. We adapt the source model with small user feedback and large unlabeled target data using previous semi-supervised domain adaptation (SemiSDA) algorithms. Compared to random feedback, which is the classical SemiSDA setup where labeled data is a random subset of target data, model adaptation with NBF leads to subpar performance. This paper analyzes this problem and introduces a scalable solution." + ], + "image_footnote": [], + "bbox": [ + 220, + 145, + 496, + 243 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1fa349cee07a879821bb1b0df1354fb785a490a9f81e0ae77032670bdc690d4a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 146, + 774, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite such advances in DA, adapting the model with user feedback still remains an open area for further research, even though practical machine learning (ML) products often allow users to provide feedback in order to further improve the model in the target environment. For example, facial recognition or medical image diagnosis applications enable users to give feedback correcting wrong model predictions, as depicted in Figure 1 (a). Since feedback can be modeled in this case as a small amount of labeled target data, it is anticipated that previous SemiSDA methods assuming the same setup would yield promising results. However, we observe that they show inferior adaptation performance on multiple DA benchmarks when using such user feedback in practice, as shown in the dark-gray bar in Figure 1 (b).", + "bbox": [ + 212, + 368, + 787, + 535 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce a novel concept called Negatively Biased Feedback (NBF) to explain this phenomenon. NBF is based on the observation that user feedback is more likely to be derived from incorrect model predictions. For example, a radiologist might log a misdiagnosed chest X-ray by the model, as its accuracy directly impacts the patient's survival. Interestingly, our observation aligns with findings from cognitive psychology literature [3,57] that proves that humans are more likely to react and provide feedback to negative events (i.e., wrong model predictions). Since such an NBF scenario is feasible, we analyze its unexpected impact on SemiSDA observed above. We identify that a biased distribution of NBF within the overall data distribution leads to sub-optimal adaptation results, particularly compared to Random Feedback (RF). RF represents the classical SemiSDA setup, where labeled data is randomly selected from the target data.", + "bbox": [ + 212, + 536, + 787, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the problem caused by NBF, we present a scalable approach named Retrieval Latent Defending, which can be seamlessly integrated with existing SemiSDA methods. Our approach allows them to adapt the model without a strong dependence on the biasedly distributed labeled data. Specifically, we balance the supervised adapting signal by appending latent defending samples to the mini-batch and help to keep the model's balanced class discriminability throughout adapting iterations. We evaluate the unexpected influence of NBF using various benchmarks, including image classification, semantic segmentation,", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Song et al.", + "bbox": [ + 271, + 114, + 361, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f43b6f978ec654ada808edc9a32d3e1c9a0d4b92b99841e0deeeb68abbd1f86f.jpg", + "image_caption": [ + "Fig. 2: Adaptation with user feedback can be effective in alleviating performance degradation caused by domain shift. However, there are some challenges: (i) user feedback may be a biased sampling of the true target distribution due to the nature of feedback, (ii) the amount of the ground truths (GT) labels obtained through feedback is small, and (iii) only unlabeled target data is typically available, not source data." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 777, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and medical image diagnosis. Building upon these evaluations, we demonstrate that our approach not only complements, but significantly enhances the performance of multiple SemiSDA methods.", + "bbox": [ + 212, + 330, + 782, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The contributions of the paper are as follows:", + "bbox": [ + 238, + 377, + 568, + 392 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "We introduce the novel concept called Negatively Biased Feedback and uncover that it can lead to sub-optimal adaptation performance of existing SemiSDA methods.", + "- We analyze this problem and present a scalable solution, Retrieval Latent Defending, that combines with SemiSDA methods and allows them to avoid the unexpected effect of NBF.", + "- We show that our approach generalizes through diverse DA benchmarks and improves adaptation results of state-of-the-art SemiSDA methods.", + "We publicly release the code on https://github.com/junha1125/RLD-SemiSDA." + ], + "bbox": [ + 230, + 395, + 784, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 553, + 387, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adaptation in the deployment environment. Real-world ML products often encounter performance degradation caused by gaps between the source and target environment [17]. One solution is to adapt the model using unlabeled data observed in the target domain, referred to as unsupervised domain adaptation (UDA) [37, 59, 72]. Works on UDA use both source and target data to improve the target performance by using methods such as domain discrepancy minimization by adversarial training [18, 41, 59, 70, 72, 73, 76], and self-training with pseudo labels [45, 51, 97, 98]. Source-free DA (SFDA) builds on UDA and imposes an additional constraint that the source data can not be accessed during domain adaptation. This has practical implications for addressing data privacy concerns or barriers in data transmission to edge devices [34, 38, 77, 95]. The majority of recent SFDA works rely on strategies like domain clustering [34], nearest neighbors [91-93], and contrastive learning [8, 35, 101]. Nevertheless, SFDA does not consider the availability of small labeled data, which may be available in practical ML systems. Semi-supervised DA (SemiSDA) works mainly demonstrate that permitting small labeled data in the target domain can substantially enhance adaptation performance compared to traditional UDA [58].", + "bbox": [ + 212, + 583, + 787, + 843 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "NBF & RLD", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Their primary strategy is to use domain alignment [20, 33, 58, 94], multi-view consistency [2, 6, 33, 89], and asymmetric co-training [36, 90].", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Active domain adaptation (ActiveDA) [25, 55, 86] envisions a scenario in which the machine selects specific target samples and instructs annotators to label them. The primary objective of ActiveDA is to strategically identify and select the most informative samples for annotation. These chosen samples (i.e., labeled target data) are subsequently utilized to update the source model using SemiSDA methods [33, 58], and the effectiveness of ActiveDA is assessed by evaluating the target performance of the adapted model.", + "bbox": [ + 212, + 181, + 784, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semi-supervised learning (SemiSL) aims to reduce expensive human annotations, and propose methods to train a model from scratch using massive unlabeled data along with limited amounts of labeled data [43,74]. The majority of SemiSL methods depend on consistency regularization [4,5,16,60,66,87], which helps the model to make similar predictions for augmented versions of the same image. Moreover, adaptive thresholding [9,12,24,66,81,88,99] is also popularly utilized to produce reliable pseudo labels from unlabeled data.", + "bbox": [ + 212, + 291, + 784, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SemiSDA and SemiSL setups mimic small labeled datasets by randomly selecting subsets of the target dataset, whereas ActiveDA involves selections instructed by the machine. In contrast, this paper posits that in real-world applications, labeled data is typically acquired through user intervention. Additionally, users often provide feedback on samples misclassified by the model (i.e., negatively biased feedback), a process detailed in the following section.", + "bbox": [ + 212, + 397, + 784, + 488 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/5e746c549ebb8dcd0b0088beb454eb3d772d8024870d203e9273ca7d13ddd968.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
UDASFDAActiveDASemiSDASemiSLOur setup
Adaptation○○○○×○
Source-free×○××-○
Feedback××machine-instructedrandomly selectedrandomly selecteduser-provided
", + "bbox": [ + 218, + 494, + 781, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The table above summarizes the comparison of relevant studies to our setup. In the table, adaptation means fine-tuning the source pre-trained model (as opposed to training from scratch); feedback represents a small number of labeled target samples. Appendix A provides further comparisons with settings like class-imbalanced SemiSDA and test-time adaptation (TTA).", + "bbox": [ + 212, + 535, + 784, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Negatively Biased Feedback", + "text_level": 1, + "bbox": [ + 214, + 626, + 521, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Adaptation with user feedback.", + "text_level": 1, + "bbox": [ + 214, + 655, + 519, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our adaptation setup is illustrated in Figure 2. A model is pre-trained on the source data $D_{s}$ . Next, the model is deployed to the target domain, such as a smartphone or a hospital, where we assume the transfer of $D_{s}$ is prohibited due to data privacy regulations or resource constraints (same setup as SFDA [34]). While users utilize ML products on the target domain, the model provides prediction results for data observed in the target domain $D_{t}$ and occasionally obtains user feedback in the form of annotations $y$ . We represent the target data as $D_{t} = X_{t}^{lb} \\cup X_{t}^{ulb}$ , where $X_{t}^{lb} = \\{(x_{lb}^{n},y_{lb}^{n}):n\\in [1..N_{lb}]\\}$ , $X_{t}^{ulb} = \\{(x_{ulb}^{n}):n\\in [1..N_{ulb}]\\}$ , $x_{lb}$ and $x_{ulb}$ denote labeled and unlabeled data and $N_{lb}$ and $N_{ulb}$ is their number of data. Lastly, the model can utilize $D_{t}$ and SemiSDA algorithms for adaptation during its inactive phase (e.g., when users do not use the product,", + "bbox": [ + 212, + 672, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Song et al.", + "bbox": [ + 271, + 114, + 359, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/df2bd396b0b566c13e4bd4684d90efe02dc750deafbbad1cc6f1a1fae253d0e6.jpg", + "image_caption": [ + "Fig.3: Effect of negatively biased feedback. Our novel observations are that $(a)$ user-provided feedback in practice has a biased distribution in each class cluster (the bottom center sub-figure) which is in contrast to random feedback, $(b)$ Existing SemiSDA methods adapt the model by dominating the labeled data points (the right sub-figures) even though they are biasedly positioned, and $(c)$ NBF prevents the model from having a decision boundary for true class clusters and leads to inferior adaptation performance (the bottom right sub-figure)." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 753, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "like at nighttime) in order to alleviate performance degradation due to domain shift or to personalize the model based on user feedback.", + "bbox": [ + 212, + 459, + 784, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rethinking user-provided feedback. Classical SemiSDA works simply assume that a random subset in target data $D_{t}$ is labeled by users when building $X_{t}^{lb}$ . However, as illustrated in Figure 2 (i), we suggest that users are more likely to provide feedback on misclassified samples by the source model, named negatively biased feedback (NBF). This behavior can be understood from two perspectives: (a) users generally expect their feedback to be used as a basis of model improvement, motivating them to provide NBF, and (b) humans tend to react more strongly to negative experiences, such as receiving incorrect predictions, as observed in psychological studies [3,57]. We note that the NBF assumption holds more strongly for the medical application: it is reasonable to imagine that the user (i.e., radiologist) logs the mistakes of the model while diagnosing a chest X-ray exam because the diagnostic accuracy of the model is directly related to the patient's chances of survival. Furthermore, applications beyond the medical domain can also exhibit NBF. For instance, users in self-driving cars can report errors, such as object detection failures or navigation mistakes, to enhance the car's driving capabilities.", + "bbox": [ + 212, + 494, + 787, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Influence of NBF on SemiSDA", + "text_level": 1, + "bbox": [ + 215, + 753, + 516, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Simulation study. As shown in Figure 3, we conduct a simulation study to understand the effect of NBF on SemiSDA. We first use the blobs dataset [53] and construct the source and target data so that domain shift exists between them (left sub-figures). We pre-train a source model on the source data and", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "NBF & RLD", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "compute the accuracy in the target domain, where the performance drop due to domain shift is observed $(98.5\\% \\rightarrow 76.4\\%)$ . Next, we simulate two types of feedback (i.e., labeled data): random feedback and negatively biased feedback following a previous SemiSDA setup and our setup, respectively. Specifically, NBF is randomly selected among misclassified samples by the source model. We find that random feedback (RF) points are evenly distributed, while NBF points are biasedly positioned across each class cluster (refer to blue points in the dashed circle in the center sub-figures).", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To alleviate the performance drop caused by domain shift, we adapt the model using the target data and a semi-supervised method, Pseudo-labeling [1]. This method iteratively optimizes the model by the cross-entropy loss computed by the ground truth of labeled data and pseudo labels of unlabeled data in a minibatch (pseudo labels are predicted by the current adapting model so they can be changed according to an updating decision boundary. Further comprehension can be achieved by referring Appendix B.). The SemiSDA results are shown in the right sub-figures, where we make two interesting observations: (i) the distribution of labeled data can contribute significantly to a decision boundary of the adapted model (red arrows in the figure), and (ii) the adapted model under NBF has poorly improved performance compared with one under RF $(76.4\\% \\rightarrow 88.1\\%$ with NBF, but $76.4\\% \\rightarrow 91.7\\%$ with RF).", + "bbox": [ + 212, + 267, + 787, + 449 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unexpected influence of NBF. Our intuitive reasoning probably suggests that NBF provides more information than RF by correcting more source model deficiencies, and thus leads to better adaptation performance. However, we empirically show that NBF can result in inferior adaptation performance due to its biased distribution across each class cluster, as illustrated in Figure 3. Surprisingly, we also show that this problem persists, even with other state-of-the-art SemiSDA methods and large datasets for various DA benchmarks, including image classification, semantic segmentation, and medical image diagnosis. Our work highlights the importance of careful design when using user feedback in real-world scenarios and, to the best of our knowledge, is the first study to uncover and analyze this phenomenon.", + "bbox": [ + 212, + 453, + 787, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Approach", + "text_level": 1, + "bbox": [ + 215, + 630, + 346, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Prerequisite: Previous SemiSDA method", + "text_level": 1, + "bbox": [ + 215, + 655, + 599, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Previous SemiSDA and SemiSL works typically construct a mini-batch with labeled data $\\{(x_{lb}^b, y_{lb}^b) : b \\in [1..B]\\}$ , and unlabeled data whose size is $\\mu$ times larger than labeled ones $\\{(x_{ulb}^b) : b \\in [1..\\mu \\cdot B]\\}$ , where $B$ is the mini-batch size for labeled data. To adapt the model iteratively, they compute the cross-entropy loss $\\mathcal{H}(\\cdot, \\cdot)$ with labeled data and the consistency regularization to multi-view of unlabeled data, which are formulated as the following:", + "bbox": [ + 212, + 676, + 787, + 768 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s u p} = \\frac {1}{B} \\sum_ {b = 1} ^ {B} \\mathcal {H} \\left(y _ {l b} ^ {b}, f _ {\\theta} \\left(x _ {l b} ^ {b}\\right)\\right), \\quad \\mathcal {L} _ {u n s u p} = \\frac {1}{\\mu \\cdot B} \\sum_ {b = 1} ^ {\\mu \\cdot B} \\mathcal {H} \\left(\\hat {y} _ {u l b} ^ {b}, f _ {\\theta} \\left(\\Omega \\left(x _ {u l b} ^ {b}\\right)\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 267, + 770, + 785, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $f_{\\theta}(\\cdot)$ is the output probability from the model, $\\hat{y}_{ulb}$ denotes a pseudo label obtained from $f_{\\theta}(\\omega(x_{ulb}))$ , and $\\omega(\\cdot)$ and $\\Omega(\\cdot)$ represent weak and strong image", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Song et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/623147e71e483c912e40ebb7b5a40dd915dc5a23b83f3ae1375eb1e375a43f13.jpg", + "image_caption": [ + "Fig. 4: Even though labeled data $(x_{lb}, y_{lb})$ is biasedly positioned, the model needs to be adapted with balanced class discriminability (i.e., decision boundary). (i) However, previous SemiSDA methods have overlooked this fact and used the labeled data naively by applying a cross-entropy loss, leading to inadequate adaptation performance. (ii) To alleviate this problem, we propose a scalable adapting approach, retrieval latent defending, which allows the model to adjust the balance of a mini-batch on each iteration by using latent defending samples $x_{LD}$ together with labeled data $x_{lb}$ ." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 759, + 362 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "augmentation, respectively. While sharing the core framework, each SemiSDA method employs distinct adapting strategies, especially to enhance the effectiveness of the use of unlabeled data rather than labeled data [6,81,96].", + "bbox": [ + 212, + 474, + 784, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Problem of previous works. Since previous SemiSDA methods have overlooked the unexpected impact of NBF, they often suffer from sub-optimal performance under the NBF assumption (shown in Section 5). To address this problem, we focus on developing a scalable solution that (i) can easily combine with existing DA methods without modifying their core adapting strategies and (ii) can be applied to a wide range of benchmarks, including medical image diagnosis.", + "bbox": [ + 212, + 525, + 787, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Retrieval Latent Defending", + "text_level": 1, + "bbox": [ + 215, + 635, + 486, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on the observations in Figure 3, we illustrate the unintended effect of NBF when using an existing SemiSDA method in Figure 4 (top center), where NBF is likely to exhibit a biased distribution, leading to undesirable adaptation results. To alleviate this issue, we propose Retrieval Latent Defending as depicted in Figure 4 (bottom). $①$ Prior to each epoch, we generate a candidate bank of data points, denoted as $x_{LD}$ . $② \\sim ④$ For each adapting iteration, we balance the mini-batch by retrieving latent defending samples $x_{LD}$ from the bank. $⑤ \\sim ⑥$ The model is then adapted using the reconfigured mini-batch and following the baseline SemiSDA approach. We hypothesize that the latent space progressively created by the $x_{LD}$ candidates throughout the adaptation process (bold dashed circle in Figure 4 (top right)) mitigates the issue caused by NBF, thereby allowing the SemiSDA baseline to achieve robust adaptation against NBF.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "NBF & RLD", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Candidate bank generation. The candidate bank serves as a repository of pseudo labels $\\hat{Y}_t^{ulb}$ for a subset of the target unlabeled data $X_t^{ulb}$ . Before each epoch, we freeze the model and use it to generate pseudo labels $\\hat{Y}_t^{ulb} = \\{(\\hat{y}_{ulb}^n): n \\in [1..N_{ulb}]\\}$ , where $\\hat{y}_{ulb}^n$ is assigned to $x_{ulb}^n$ as the predicted class with the highest softmax probability: $\\hat{y}_{ulb}^n = \\operatorname{argmax}_c [f_\\theta(x_{ulb}^n)]_c$ . We then retain only samples with the top $p\\%$ highest probabilities within each class. This filtering step helps mitigate the inclusion of data with potentially inaccurate pseudo labels, as the model's predictions on $X_t^{ulb}$ might not always be perfect.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Defending sample selection. We select $k$ latent defending samples $x_{LD}$ from the bank at random for each labeled data $(x_{lb}^{b}, y_{lb}^{b})$ . These selected samples share the same pseudo label as the ground-truth label of their associated counterparts (i.e., $\\hat{y}_{LD} = y_{lb}^{b}$ ). By incorporating these defending samples, we balance the data distribution within the current mini-batch. For example, consider $x_{lb}^{1}$ and $x_{lb}^{2}$ in Figure 4 (top right). As these labeled samples are included in the current mini-batch alongside the selected defending samples $x_{LD}^{1}$ and $x_{LD}^{2}$ , we expect to prevent the supervised adapting signal from becoming overly dependent on the labeled samples. We imagine the effect of the defending samples throughout the adaptation process and depict the latent space formed gradually by the $x_{LD}$ candidates as bold dashed circles in Figure 4 (top right).", + "bbox": [ + 212, + 271, + 787, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Consequently, the overall loss consists of the sum of losses in Eq. (1) and a loss from our proposed method as,", + "bbox": [ + 212, + 439, + 785, + 469 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\underbrace {\\mathcal {L} _ {\\text {s u p}} + \\mathcal {L} _ {\\text {u n s u p}}} _ {\\text {b a s e l i n e}} + \\underbrace {\\frac {1}{k \\cdot B} \\sum_ {b = 1} ^ {k \\cdot B} \\mathcal {H} \\left(y _ {L D} ^ {b} , f _ {\\theta} \\left(x _ {L D} ^ {b}\\right)\\right)} _ {\\text {r e t r i e v a l l a t e n t d e f e n d i n g}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 472, + 785, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Importance of our method. Understanding the impact of NBF on adaptation performance is crucial. For example, naively adapting a model for a medical application using radiologist-provided feedback can actually cause performance degradation (shown in Table 5), potentially posing significant risks to patients. We propose a scalable and simple approach to solve the problem caused by NBF, which can not be addressed by existing methods. Given the practicality of the NBF problem and the scalability of our solution, we believe our work holds considerable potential for real-world applications.", + "bbox": [ + 212, + 531, + 787, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 666, + 377, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 Experimental Setups", + "text_level": 1, + "bbox": [ + 214, + 691, + 434, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our approach is simple enough to seamlessly combine with existing SemiSDA algorithms and also be applied to diverse benchmarks. This section describes our experimental setup for natural image classification tasks and a real-world medical application. Details for semantic segmentation experiments are in Appendix D.", + "bbox": [ + 212, + 715, + 785, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baselines. We validate our approach by combining various state-of-the-art algorithms for SemiSDA [58] (e.g., CDAC [33] and AdaMatch [6]) and SemiSL [66,87] (e.g., FlexMatch [96] and FreeMatch [81]). Note that the SemiSL methods have been demonstrated to be strong SemiSDA learners [99], so we can consider them", + "bbox": [ + 214, + 779, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Song et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/913cf31ba67ae89e6447208b285243fc9f78116ff20179b00a14e23bd23f63dc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
methodfeedbackaverage\\( \\mathrm{r} \\rightarrow \\mathrm{c} \\)\\( \\mathrm{r} \\rightarrow \\mathrm{p} \\)\\( \\mathrm{p} \\rightarrow \\mathrm{c} \\)\\( \\mathrm{c} \\rightarrow \\mathrm{s} \\)\\( \\mathrm{s} \\rightarrow \\mathrm{p} \\)\\( \\mathrm{r} \\rightarrow \\mathrm{s} \\)\\( \\mathrm{p} \\rightarrow \\mathrm{r} \\)
RerNetAdaMatch [6]RF67.666.668.568.560.369.258.781.5
NBF\\( {64.5}\\left( {+{3.1}}\\right) \\)64.366.165.656.965.654.278.9
w/ oursNBF\\( \\mathbf{{72.0}\\left( {+{7.5}}\\right) } \\)74.572.773.965.570.064.383.2
ViTAdaMatch [6]RF74.775.376.973.868.076.367.185.5
NBF\\( {73.7}\\left( {+{1.0}}\\right) \\)74.776.274.765.774.066.884.0
w/ oursNBF\\( \\mathbf{{75.9}\\left( {+{2.2}}\\right) } \\)76.977.877.868.576.668.385.1
", + "bbox": [ + 236, + 143, + 767, + 207 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 1: Adaptation results on DomainNet-126. We simulate seven domain-shift scenarios (i.e., source $\\rightarrow$ target). The model is pre-trained on the source domain and then adapted to a training set of the target domain. The results on the test set of the target domain are reported as the top-1 accuracy (\\%). DomainNet-126 [54,58] dataset includes real, painting, sketch, and clip-art domains. In this experiment, we assume that the 378 feedback samples (i.e., 3 labeled data per class) are obtained from users. A state-of-the-art SemiSDA method, AdaMatch [6], is used as a baseline.", + "bbox": [ + 212, + 205, + 787, + 303 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/4afba3dfdf9ade3d9156f63d1563bea8e2273195228ecc8f5c1218f92d06835b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
methodfeedbackaveragea→ca→pa→rc→ac→pc→rp→ap→cp→rr→ar→cr→p
AdaMatch [6]RF70.955.480.475.965.781.574.665.958.778.468.861.584.3
NBF69.3 (+1.6)54.276.675.365.979.375.563.757.475.966.756.884.2
w/oursNBF73.8 (+4.5)62.281.079.768.885.478.667.761.779.569.064.188.2
", + "bbox": [ + 235, + 308, + 764, + 343 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 2: Adaptation results on OfficeHome. OfficeHome [75] dataset includes real, product, art, and clip-art domain. We assume that the 195 feedback samples (i.e., 3 labeled data per class) are obtained. AdaMatch [6] and ResNet-50 [23] are used.", + "bbox": [ + 212, + 343, + 784, + 385 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "as SemiSDA methods. For medical experiments, we use Pseudo-labeling [1] as a baseline since it is easily applicable to medical image adaptation.", + "bbox": [ + 212, + 393, + 785, + 425 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. We utilize natural image datasets containing multiple kinds of domains (e.g., real and painting). The datasets include DomainNet-126 [54,58] with 142k images of 126 classes, and OfficeHome [75] with 15K images of 65 classes.", + "bbox": [ + 212, + 428, + 785, + 472 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To conduct medical experiments, we present a practical medical setting. We adopt the MIMIC-CXR-V2 dataset [27]. It assumes a multi-finding binary classification setup, where multiple radiographic findings, like Pneumonia and Atelectasis, can coexist in a single chest X-ray (CXR) sample. Thus, the model predicts the presence or absence (binary classes) of each individual finding. We simulate domain shift by using Posterior-Anterior (PA)-view data as the source and AP-view data as the target, capturing real-world variations in data acquisition. Typically, patients requiring an AP X-ray are those facing positioning challenges that prevent them from undergoing a PA X-ray. Therefore, this setup can be seen as a scenario where the target environment is the intensive care unit, which hospitalizes critically ill patients.", + "bbox": [ + 212, + 474, + 787, + 638 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Following the recent SemiSDA [94] and SFDA [8] setups, we assume the model is pre-trained in the source domain and deployed in the target domain. Since the datasets above were not initially divided into training and test sets, we performed a random 8:2 split within each domain, designating them respectively for training and testing. The training set is used to adapt the model, while the test set is used to report the top-1 accuracy.", + "bbox": [ + 212, + 640, + 787, + 731 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "User feedback. Feedback given by users is modeled as annotations $\\{y_{lb}^n : n \\in [1..N_{lb}]\\}$ on a small subset of the target's training set $D_t^{train}$ , while the remaining of them are used as unlabeled target data. In our experiments, we take into account two types of feedback: random feedback (RF) and negatively biased feedback (NBF). RF is the same setup of classical SemiSDA and SemiSL, where randomly selected samples from $D_t^{train}$ are used as small labeled set $X_t^{lb}$ . For NBF, we randomly select samples that are incorrectly predicted in $D_t^{train}$ by", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "NBF & RLD", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/170b81269cf22ac510e3f79814deeeb72cf49c9a02aa11f24411ae8d41de50b5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
feed, amount method378 (3 labeled data per class)630 (5 labeled data per class)
RFNBFw/oursRFNBFw/ours
Source model56.5
MME [58]69.568.4 (-1.1)70.8 (+2.4)71.270.1 (-1.1)72.5 (+2.4)
CDAC [33]68.364.6 (-3.7)73.2 (+8.6)71.768.1 (-3.6)74.9 (+6.8)
AdaMatch [6]67.664.5 (-3.1)72.0 (+7.5)70.967.7 (-3.2)74.3 (+6.6)
FixMatch [66]67.663.4 (-4.2)73.2 (+9.8)71.566.1 (-5.4)75.1 (+9.0)
UDA [87]69.264.9 (-4.3)73.4 (+8.5)72.968.8 (-4.1)75.3 (+6.5)
FlexMatch [96]73.371.4 (-1.9)74.7 (+3.3)75.373.9 (-1.4)76.0 (+2.1)
FreeMatch [81]73.872.0 (-1.8)74.8 (+2.8)75.674.4 (-1.2)76.1 (+1.7)
Fully supervised83.6
Source model64.5
MME [58]73.272.7 (-0.5)74.1 (+1.4)74.574.0 (-0.5)75.2 (+1.2)
CDAC [33]74.272.8 (-1.4)75.4 (+2.6)75.474.1 (-1.3)76.2 (+2.1)
AdaMatch [6]74.773.7 (-1.0)75.9 (+2.2)75.975.1 (-0.8)76.7 (+1.6)
FixMatch [66]74.673.0 (-1.6)75.6 (+2.6)75.774.3 (-1.4)76.5 (+2.2)
UDA [87]74.873.3 (-1.5)75.8 (+2.5)75.974.5 (-1.4)76.7 (+2.2)
FlexMatch [96]74.973.9 (-1.0)75.8 (+1.9)76.075.1 (-0.9)76.9 (+1.8)
FreeMatch [81]74.973.9 (-1.0)75.7 (+1.8)76.075.1 (-0.9)76.8 (+1.7)
Fully supervised85.4
", + "bbox": [ + 303, + 143, + 699, + 305 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/85d5b51d0e95f98724905a9eaacf44b18726daeacb236407af0c3cfcfdea2b99.jpg", + "table_caption": [ + "Table 3: Comparisons on DomainNet-126. We evaluate our method by integrating it with SemiSDA and SemiSL methods. The average accuracy of seven domain-shift scenarios in Table 1 is reported. Source model represents the pre-trained model without adaptation. Fully supervised means the model is adapted with fully labeled target data." + ], + "table_footnote": [], + "table_body": "
feed, amount method195 (3 labeled data per class)325 (5 labeled data per class)
RFNBFw/oursRFNBFw/ours
Source model57.6
MME [58]71.270.2 (-1.0)73.4 (+3.2)73.573.1 (-0.4)75.6 (+2.5)
CDAC [33]71.269.0 (-2.2)74.3 (+5.3)73.572.3 (-1.2)75.7 (+3.4)
AdaMatch [6]70.969.3 (-1.6)73.8 (+4.5)73.472.7 (-0.7)75.5 (+2.8)
FixMatch [66]71.468.6 (-2.8)73.7 (+5.1)73.972.2 (-1.7)75.3 (+3.1)
UDA [87]72.269.5 (-2.7)74.1 (+4.6)74.473.0 (-1.4)76.0 (+3.0)
FlexMatch [96]73.772.1 (-1.6)74.7 (+2.6)75.974.9 (-1.0)76.6 (+1.7)
FreeMatch [81]74.072.7 (-1.3)74.8 (+2.1)75.875.0 (-0.8)76.6 (+1.6)
Fully supervised87.4
", + "bbox": [ + 303, + 364, + 699, + 455 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 4: Comparisons on OfficeHome. The average accuracy of twelve domain-shift scenarios in Table 2 is reported. ResNet-50 is used.", + "bbox": [ + 215, + 455, + 784, + 483 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "the source model (i.e., the pre-trained model before adaptation). Note that we focus on the impact of a biased label distribution within the same class, as shown in Figure 3, and thus take the same number of feedback for each class. Further discussion about the imbalance in the number of feedback between classes presented in [31, 49, 83] is provided in Appendix A.2.", + "bbox": [ + 212, + 496, + 785, + 571 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Network architectures. We adopt commonly used networks, ResNet [23] and ViT [15] for natural image tasks and DenseNet [26] for a medical task. We employ ResNet-50 with the last classification layer comprising a weight normalization layer and a bottleneck layer, following previous works [8, 34] and use the ViT-Small (i.e., ViT-S) introduced in [80]. The DenseNet-121 is used, provided in TorchXrayVision [13], like existing medical works [32, 44].", + "bbox": [ + 212, + 574, + 785, + 665 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation details. We implement our framework by extending the publicly available USB [80] repository. Both pre-training and adaptation are conducted with a mini-batch size of 128 and the SGD optimizer. Diverse baselines for SemiSDA and SemiSL are used to compute the losses in Eq. (1). The hyperparameters for each baseline simply follow USB [80] or public code [33,58]. For all experiments, our approach uses the same hyper-parameters of the appended defending samples $k$ and reliable filtering rate $p$ as 3 and 0.4, respectively.", + "bbox": [ + 212, + 667, + 785, + 773 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Main Results", + "text_level": 1, + "bbox": [ + 215, + 789, + 372, + 804 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Natural image classification. Following recent DA works [8, 94], we conduct experiments on seven and twelve domain shift scenarios provided with the", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Song et al.", + "bbox": [ + 271, + 114, + 361, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/1a0a93f6bc6c7b1f36a1f94d5389a287c4d30798c41abd0fe1ed981fc4076aeb.jpg", + "image_caption": [ + "Table 5: Adaptation in a medical application. We use samples with PA-view as the source data and samples with AP-view as the target data in MIMIC-CXR-V2 dataset [27]. NBF-CE represents a scenario when NBF is composed of cases with confident errors. We use DensNet-121 [13, 26] and assume the 20 feedback for the absence and presence per finding." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 803, + 351 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "DomainNet-126 and OfficeHome datasets, respectively. Table 1 and Table 2 show the results, where AdaMatch [6] is used as the baseline. We observe consistent results with Figure 3 even on large natural datasets: when simply applying the baseline under the NBF assumption, the adapted model shows inferior performance for most domain shifts than applying it under RF, e.g., $64.5 < 67.6$ . Combining our approach with the baseline mitigates this issue and achieves a performance increase, e.g., $64.5 \\rightarrow 72.0$ .", + "bbox": [ + 212, + 429, + 787, + 535 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We also use other promising baselines and report the average accuracy of all domain shifts in Table 3 and Table 4 (all results can be found in Appendix F). While both feedback types bring performance improvement from the source model, lower performance is observed with NBF. Our method enables the baselines to not only address this problem but surpass performance under RF. The above results suggest that the biased distribution of labeled samples, which has been overlooked in previous SemiSDA works, is actually problematic, and our retrieval latent defending approach is effective.", + "bbox": [ + 212, + 536, + 787, + 656 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Medical image diagnosis. Table 5 shows the results (bottom) and also depicts the effect of NBF (top center). We report the AUROC [7] for each finding following standard practice for measuring computer-aided-diagnosis model evaluation [32,44]. The baseline SemiSDA method under NBF exhibits inferior performance compared to one under RF, but this issue can be mitigated by combining our approach.", + "bbox": [ + 212, + 660, + 787, + 750 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In addition, we propose an interesting and practical scenario named NBF with more confident errors (NBF-CE). In this scenario, we assume that a radiologist is likely to give feedback when the model makes confidently wrong predictions. Imagine that the model predicts a $1\\%$ likelihood of cancer in a CXR image, but the person actually has cancer. Such failure to detect potential patients early on can significantly reduce the patient's chances of survival, so a radiologist may", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "NBF & RLD", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/6538a72ac99fe2db0591c6563e68ade36a2d6af2e08eeb186558787cc2dc7079.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
labeling typefeed. amountRFNBFNBF w/oursENT [62]ENT w/ours
IAST [46,63]PA, 40 points55.353.0 (-2.3)56.3 (+3.3)53.556.0 (+2.5)
RIPU [85]PA, 40 points57.654.5 (-3.1)58.0 (+3.5)54.657.7 (+3.1)
", + "bbox": [ + 218, + 143, + 781, + 176 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 6: Adaptation on semantic segmentation. The GTA5 [56] $\\rightarrow$ Cityscapes [14] setup is used [72]. The target performance of the source model is $36.6\\mathrm{mIoU}$ .", + "bbox": [ + 215, + 176, + 782, + 205 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "provide feedback to the model. To simulate NBF-CE, we select samples where the source model most confidently predicts a finding to be absent $(\\hat{y} \\approx 0)$ although it is clearly visible in the radiograph $(y = 1)$ , and vice versa, i.e., samples of $\\hat{y} \\approx 1$ but $y = 0$ . Table 5 also shows the results under an NBF-CE scenario, where the model's adaptation performance is further reduced compared with NBF (0.7691 for $\\mathrm{NBF} \\to 0.7639$ for NBF-CE). By combining our method, we observe performance improvements for both NBF variants, e.g., 0.7639 for $\\mathrm{NBF - CE} \\to 0.7875$ with ours. We illustrate the hypothesized impact of our method in Table 5.", + "bbox": [ + 212, + 213, + 787, + 335 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Semantic segmentation. We evaluate the influence of NBF and our approach on a semantic segmentation task. We utilize the most common adaptation benchmark of GTA5 [56] to Cityscapes [14]. The baseline DA algorithms are used as IAST [46, 63] and RIPU [85] in a source-free scenario. We regard Pixel-based Annotation (PA) in which we assume 40 pixels per image like LabOR [63]. Table 6 shows results similar to those we observed in the classification and medical imaging tasks. The baselines under NBF exhibit inferior performance compared to those under RF (54.5 for $\\mathrm{NBF} < 57.6$ for RF), but this issue is addressed by combining our approach with them $(+3.5\\mathrm{mIoU})$ . Although out of our scope (refer to Appendix A.1), we validate one active labeling strategy ENT [62], which assigns highly uncertain (i.e., probably misclassified) pixels as feedback. Consequently, the feedback instructed by ENT is biasedly distributed in a manner similar to NBF. ENT also causes unexpected results (54.6 for $\\mathrm{ENT} < 57.6$ for RF), and our approach alleviates this issue $(+3.1\\mathrm{mIoU})$ .", + "bbox": [ + 212, + 338, + 787, + 551 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 565, + 388, + 580 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Positive vs. Negative feedback. We study the role of feedback on the adaptation results by varying feedback configurations. Let positively-provided feedback (PF) be obtained from samples that the source model correctly predicts, as opposed to negatively-provided feedback (NF). We adjust the ratio of PF:NF while keeping the total number of labeled samples constant, as shown in Figure 5.", + "bbox": [ + 212, + 590, + 785, + 666 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "When using only FreeMatch (gray dot-dashed line), both biased feedback types (i.e., NBF and PBF) result in worse adaptation performance compared to balanced feedback for the baseline, e.g., 72.6 in 378:0 (PBF) $< 73.3$ in 252:126. In contrast, when our method is applied (red line), NBF yields the best performance. PBF and NBF can be respectively regarded as contributing previously known knowledge of the model and new knowledge that complements model deficiencies. Hence, it may be natural that NBF, which actually encodes the model's mistakes, contributes to favorable adaptation results.", + "bbox": [ + 212, + 666, + 787, + 787 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Number of unlabeled samples in a mini-batch. Existing SemiSDA methods [6,81] typically set the ratio $\\mu$ between labeled and unlabeled samples in a", + "bbox": [ + 212, + 790, + 787, + 821 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Song et al.", + "bbox": [ + 271, + 114, + 359, + 128 + ], + "page_idx": 11 + }, + { + "type": "page_footnote", + "text": "If not specified, we use ResNet-50 and report the average accuracy (\\%) of seven domain shift scenarios in Table 1 for ablation studies.", + "bbox": [ + 228, + 830, + 785, + 859 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e86a614287b88fe342b0453b03d79c126febf09d14fab568454cd2058975d420.jpg", + "image_caption": [ + "Fig. 5: NBF leads to higher performance than PBF. We compare different user-feedback configurations when the total number of feedback is 378 (top) and 630 (bottom). Positive and negative feedback refers to feedback from correct and incorrect model predictions, respectively. We run three random seed experiments and describe the average performance and standard deviation in the parenthesis." + ], + "image_footnote": [], + "bbox": [ + 341, + 143, + 661, + 268 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/c42ab6948b203bc7b34e63cae005d44094813cc7110b4b2685df021d71e7a162.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
methodfeed.negatively biased feedback (NBF)
#xulb, #xLD, #xlb112, 0, 16112, 48, 1664, 48, 16
total batch size128176128
FreeMatch [81]36872.074.274.8 (+0.6)
AdaMatch [6]64.571.372.0 (+0.7)
FreeMatch [81]63074.475.576.1 (+0.6)
AdaMatch [6]67.773.474.3 (+0.9)
", + "bbox": [ + 344, + 342, + 656, + 405 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 7: In the mini-batch, diminishing the number of unlabeled samples and adding our defending samples achieves better performance with our approach. We ablate them by changing the ratio $\\mu$ in Section 4.1, while keeping the size of labeled samples.", + "bbox": [ + 215, + 405, + 785, + 446 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "mini-batch to 1:7. However, we observe that adhering to this ratio is not optimal for our approach, as shown in Table 7. Our method shows better performance when the ratio is varied to 1:4, i.e., decreasing unlabeled sample sizes. This finding contradicts observations in several TTA works [28, 48, 67], where adaptation performance tends to increase with larger batch sizes. We speculate that it is beneficial to prioritize more reliable information, which refers to labeled data and our defending samples selected from the filtering-applied bank, during the adapting process. This result may be aligned with previous works for curriculum learning [40, 100] and adaptive thresholding [96].", + "bbox": [ + 212, + 454, + 787, + 590 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Number of labeled data. We measure the impact of feedback size (number of labeled samples) in Figure 6. The results show that the inferior performance on NBF persists even with an increased amount of feedback (gray $\\rightarrow$ black line); however, our approach mitigates it and improves performance (black $\\rightarrow$ red line). We make an interesting observation that the performance gap between black and red lines becomes larger as the number of available feedback decreases. Since obtaining large feedback may be challenging in real-world applications, our method is expected to be more helpful in this practical case.", + "bbox": [ + 212, + 594, + 787, + 715 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Data selecting strategy. We explore various strategies for selecting defending samples to balance the mini-batch, as shown in Table 8(top). The strategies include: in the $x_{LD}$ candidate bank, (i) random selection regardless of the class of the labeled data, (ii) random selection in the same class as the labeled data (i.e., class-aware), (iii) selecting samples close to the cluster center obtained by k means clustering [21] and (iv) selecting samples with embedded features distant from the labeled data where cosine distance is used. While our approach consistently outperforms the baseline regardless of the chosen strategy, we empirically", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "NBF & RLD", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/75da3a9bd27938ce43af20c7a5b7e3e3269010c907386b3157e2e3cfae87737d.jpg", + "image_caption": [ + "Fig. 6: More reliable adaptation with NBF. In addition to Table 3, we conduct experiments with different amounts (1,3,5,10, and 15 labeled data per class) of feedback using FreeMatch [81]. The number of available feedback is likely to be small in practice. In this case, our method achieves large performance improvement, e.g., our method increases the baseline performance by $+4.9$ when one feedback per class is available." + ], + "image_footnote": [], + "bbox": [ + 328, + 143, + 671, + 226 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/b2a4d413d7cbc7d3e15b3ba573cc5957d57b22d3dc3561596179a0f7a44dafa5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
selection strategy\nclass-awarerandom\nxrandom\n✓kmeans\n✓cosine\n✓baseline\n-
FreeMatch [81]Res.74.174.874.674.072.0
FreeMatch [81]ViT.75.075.775.675.173.9
filtering rate0.20.40.60.8baseline only
FreeMatch [81]Res.74.574.874.373.772.0
FreeMatch [81]ViT.75.575.775.975.573.9
", + "bbox": [ + 287, + 299, + 712, + 368 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 8: We ablate a component of our approach with 378 feedback: $x_{LD}$ selection strategy and filtering rate $p$ for bank generation.", + "bbox": [ + 212, + 369, + 785, + 397 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "find that strategy (ii) achieves the best performance. Therefore, we adopt this strategy for our proposed method.", + "bbox": [ + 212, + 406, + 785, + 436 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Further studies, such as extension to a TTA scenario, combining with SFDA methods and different feedback configurations, are presented in Appendix C.", + "bbox": [ + 212, + 436, + 785, + 468 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion & Discussion", + "text_level": 1, + "bbox": [ + 214, + 478, + 491, + 494 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "User feedback can play an integral part in adapting the practical ML product to the target environment. However, we have shown that naive adaptation using existing SemiSDA methods led to undesirable adaptation results. We explained this through the lens of Negatively-Biased Feedback (NBF). In this paper, we uncovered the unexpected results of NBF and presented a scalable solution, Retrieval Latent Defending. This method prevents the mini-batch from becoming overly dependent on labeled samples that may have a biased distribution within the overall target distribution. Under the diverse DA benchmarks, from the simulation study to the medical imaging task, we demonstrated the practical problem caused by NBF and the effectiveness of our approach by combining it with multiple SemiSDA baselines. We hope our efforts will inspire future DA works on leveraging user feedback to improve an ML model in the deployment environment.", + "bbox": [ + 212, + 505, + 787, + 700 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Broader impact. The proposed setup assumes that an ML product obtains feedback as a form of annotations (i.e., labeled data). In some cases, users can provide feedback in different forms, like thumbs up & down and rating of model prediction, or noise feedback whose information is different from the ground truth. Further research considering these points will pave the way for developing safer and more reliable adapting strategies.", + "bbox": [ + 212, + 702, + 787, + 794 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment. We sincerely appreciate the abundant support provided by Lunit Inc., and we would like to thank Donggeun Yoo, Seonwook Park, and Sérgio Pereira for their valuable feedback.", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Song et al.", + "bbox": [ + 271, + 114, + 359, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Arazo, E., Ortego, D., Albert, P., O'Connor, N.E., McGuinness, K.: Pseudolabeling and confirmation bias in deep semi-supervised learning. In: International Joint Conference on Neural Networks (IJCNN) (2020) 6, 9, 11, 23, 24, 25, 26, 29, 30", + "2. Basak, H., Yin, Z.: Semi-supervised domain adaptive medical image segmentation through consistency regularized disentangled contrastive learning. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (2023) 4", + "3. Baumeister, R.F., Bratslavsky, E., Finkenauer, C., Vohs, K.D.: Bad is stronger than good. Review of general psychology (2001) 2, 5", + "4. Berthelot, D., Carlini, N., Cubuk, E.D., Kurakin, A., Sohn, K., Zhang, H., Raffel, C.: Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. In: ICLR (2020) 4", + "5. Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.A.: Mixmatch: A holistic approach to semi-supervised learning. In: NeurIPS (2019) 4", + "6. Berthelot, D., Roelofs, R., Sohn, K., Carlini, N., Kurakin, A.: Adamatch: A unified approach to semi-supervised learning and domain adaptation. In: ICLR (2022) 1, 4, 7, 8, 9, 10, 11, 12, 13, 22, 28, 30, 32, 33, 34", + "7. Bradley, A.P.: The use of the area under the roc curve in the evaluation of machine learning algorithms. Pattern recognition (1997) 11", + "8. Chen, D., Wang, D., Darrell, T., Ebrahimi, S.: Contrastive test-time adaptation. In: CVPR (2022) 3, 9, 10, 24, 25, 27, 28, 30, 35", + "9. Chen, H., Tao, R., Fan, Y., Wang, Y., Wang, J., Schiele, B., Xie, X., Raj, B., Savvides, M.: Softmatch: Addressing the quantity-quality trade-off in semi-supervised learning. In: ICLR (2023) 4", + "10. Chen, W., Lin, L., Yang, S., Xie, D., Pu, S., Zhuang, Y.: Self-supervised noisy label learning for source-free unsupervised domain adaptation. In: IROS (2022) 29", + "11. Chen, X., Zhao, Z., Zhang, Y., Duan, M., Qi, D., Zhao, H.: Focalclick: Towards practical interactive image segmentation. In: CVPR (2022) 25", + "12. Chen, Y., Tan, X., Zhao, B., Chen, Z., Song, R., Liang, J., Lu, X.: Boosting semi-supervised learning by exploiting all unlabeled data. In: CVPR (2023) 4", + "13. Cohen, J.P., Viviano, J.D., Bertin, P., Morrison, P., Torabian, P., Guerrera, M., Lungren, M.P., Chaudhari, A., Brooks, R., Hashir, M., et al.: Torchxrayvision: A library of chest x-ray datasets and models. In: International Conference on Medical Imaging with Deep Learning (2022) 10, 11, 28", + "14. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR (2016) 12, 29", + "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) 1, 10, 32, 33, 35", + "16. Fini, E., Astolfi, P., Alahari, K., Alameda-Pineda, X., Mairal, J., Nabi, M., Ricci, E.: Semi-supervised learning made simple with self-supervised clustering. In: CVPR (2023) 4" + ], + "bbox": [ + 225, + 178, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "NBF & RLD", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Ganin, Y., Lempitsky, V.: Unsupervised domain adaptation by backpropagation. In: ICML (2015) 1, 3", + "18. Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., Marchand, M., Lempitsky, V.: Domain-adversarial training of neural networks. The journal of machine learning research (2016) 3", + "19. Gong, T., Jeong, J., Kim, T., Kim, Y., Shin, J., Lee, S.J.: Robust continual test-time adaptation: Instance-aware bn and prediction-balanced memory. In: NeurIPS (2023) 24", + "20. Harada, S., Bise, R., Araki, K., Yoshizawa, A., Terada, K., Kurata, M., Nakajima, N., Abe, H., Ushiku, T., Uchida, S.: Cluster-guided semi-supervised domain adaptation for imbalanced medical image classification. arXiv preprint arXiv:2303.01283 (2023) 4", + "21. Hartigan, J.A., Wong, M.A.: Algorithm as 136: A k-means clustering algorithm. Journal of the royal statistical society. series c (applied statistics) (1979) 13, 27", + "22. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR (2020) 27", + "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) 1, 9, 10, 32, 33, 35", + "24. Higuchi, Y., Moritz, N., Roux, J.L., Hori, T.: Momentum pseudo-labeling for semi-supervised speech recognition. In: Interspeech (2021) 4", + "25. Huang, D., Li, J., Chen, W., Huang, J., Chai, Z., Li, G.: Divide and adapt: Active domain adaptation via customized learning. In: CVPR (2023) 4, 21, 22, 29", + "26. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: CVPR (2017) 10, 11, 28", + "27. Johnson, A.E., Pollard, T.J., Greenbaum, N.R., Lungren, M.P., Deng, C.y., Peng, Y., Lu, Z., Mark, R.G., Berkowitz, S.J., Horng, S.: Mimic-cxr-jpg, a large publicly available database of labeled chest radiographs. arXiv preprint arXiv:1901.07042 (2019) 9, 11", + "28. Khurana, A., Paul, S., Rai, P., Biswas, S., Aggarwal, G.: Sita: Single image test-time adaptation. arXiv preprint arXiv:2112.02355 (2021) 13", + "29. Kim, J., Hur, Y., Park, S., Yang, E., Hwang, S.J., Shin, J.: Distribution aligning refinery of pseudo-label for imbalanced semi-supervised learning. In: NeurIPS (2020) 22, 23", + "30. Knox, W.B., Stone, P.: Tamer: Training an agent manually via evaluative reinforcement. In: IEEE international conference on development and learning (2008) 25", + "31. Lee, H., Shin, S., Kim, H.: Abc: Auxiliary balanced classifier for class-imbalanced semi-supervised learning. In: NeurIPS (2021) 10, 22", + "32. Lenga, M., Schulz, H., Saalbach, A.: Continual learning for domain adaptation in chest x-ray classification. In: Medical Imaging with Deep Learning (2020) 10, 11", + "33. Li, J., Li, G., Shi, Y., Yu, Y.: Cross-domain adaptive clustering for semi-supervised domain adaptation. In: CVPR (2021) 4, 8, 10, 32, 33, 34", + "34. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: ICML (2020) 1, 3, 4, 10, 26, 27, 35", + "35. Litrico, M., Del Bue, A., Morerio, P.: Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation. In: CVPR (2023) 3, 26, 27, 29, 30, 35", + "36. Liu, X., Xing, F., Shusharina, N., Lim, R., Jay Kuo, C.C., El Fakhri, G., Woo, J.: Act: Semi-supervised domain-adaptive medical image segmentation with asym" + ], + "bbox": [ + 225, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Song et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "metric co-training. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (2022) 4", + "37. Liu, X., Yoo, C., Xing, F., Oh, H., El Fakhri, G., Kang, J.W., Woo, J., et al.: Deep unsupervised domain adaptation: A review of recent advances and perspectives. APSIPA Transactions on Signal and Information Processing (2022) 3", + "38. Liu, Y., Zhang, W., Wang, J.: Source-free domain adaptation for semantic segmentation. In: CVPR (2021) 3", + "39. Liu, Y., Kothari, P., van Delft, B., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? In: NeurIPS (2021) 24, 30", + "40. Liu, Z., Miao, Z., Pan, X., Zhan, X., Lin, D., Yu, S.X., Gong, B.: Open compound domain adaptation. In: CVPR (2020) 13", + "41. Long, M., Zhu, H., Wang, J., Jordan, M.I.: Deep transfer learning with joint adaptation networks. In: ICML (2017) 3", + "42. MacGlashan, J., Ho, M.K., Loftin, R., Peng, B., Wang, G., Roberts, D.L., Taylor, M.E., Littman, M.L.: Interactive learning from policy-dependent human feedback. In: International conference on machine learning (2017) 25", + "43. Madani, A., Moradi, M., Karargyris, A., Syeda-Mahmood, T.: Semi-supervised learning with generative adversarial networks for chest x-ray classification with ability of data domain adaptation. In: International symposium on biomedical imaging (2018) 4", + "44. Mahapatra, D., Korevaar, S., Bozortabar, B., Tennakoon, R.: Unsupervised domain adaptation using feature disentanglement and gcns for medical image classification. In: ECCV (2022) 10, 11", + "45. Mei, K., Zhu, C., Zou, J., Zhang, S.: Instance adaptive self-training for unsupervised domain adaptation. In: ECCV (2020) 3", + "46. Mei, K., Zhu, C., Zou, J., Zhang, S.: Instance adaptive self-training for unsupervised domain adaptation. In: ECCV (2020) 12, 29", + "47. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: ICML (2022) 24, 30", + "48. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: ICLR (2023) 13", + "49. Oh, Y., Kim, D.J., Kweon, I.S.: Daso: Distribution-aware semantics-oriented pseudo-label for imbalanced semi-supervised learning. In: CVPR (2022) 10, 22", + "50. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. In: NeurIPS (2022) 25", + "51. Pan, Y., Yao, T., Li, Y., Wang, Y., Ngo, C.W., Mei, T.: Transferrable prototypical networks for unsupervised domain adaptation. In: CVPR (2019) 3", + "52. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, e.a.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019), https://github.com/pytorch/vision/blob/main/torchvision/models/resnet.py 28", + "53. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournaepau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine learning in Python. Journal of Machine Learning Research (2011) 5, 25, 26", + "54. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: ICCV (2019) 9" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "NBF & RLD", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "55. Prabhu, V., Chandrasekaran, A., Saenko, K., Hoffman, J.: Active domain adaptation via clustering uncertainty-weighted embeddings. In: ICCV (2021) 4, 21, 22, 29", + "56. Richter, S.R., Vineet, V., Roth, S., Koltun, V.: Playing for data: Ground truth from computer games. In: ECCV (2016) 12, 29", + "57. Rozin, P., Royzman, E.B.: Negativity bias, negativity dominance, and contagion. Personality and social psychology review (2001) 2, 5", + "58. Saito, K., Kim, D., Sclaroff, S., Darrell, T., Saenko, K.: Semi-supervised domain adaptation via minimax entropy. In: ICCV (2019) 1, 3, 4, 8, 9, 10, 28, 32, 33, 34", + "59. Saito, K., Watanabe, K., Ushiku, Y., Harada, T.: Maximum classifier discrepancy for unsupervised domain adaptation. In: CVPR (2018) 3", + "60. Sajjadi, M., Javanmardi, M., Tasdizen, T.: Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In: NeurIPS (2016) 4", + "61. Schulman, J., Zoph, B., Kim, C., Hilton, J., Menick, J., Weng, J., Uribe, J.F.C., Fedus, L., Metz, L., Pokorny, M., et al.: Chatgpt: Optimizing language models for dialogue. OpenAI blog (2022) 25", + "62. Shen, Y., Yun, H., Lipton, Z.C., Kronrod, Y., Anandkumar, A.: Deep active learning for named entity recognition. In: ICLR (2017) 12, 22", + "63. Shin, I., Kim, D.J., Cho, J.W., Woo, S., Park, K., Kweon, I.S.: Labor: Labeling only if required for domain adaptive semantic segmentation. In: ICCV (2021) 12, 29", + "64. Sofiuk, K., Petrov, I., Barinova, O., Konushin, A.: f-brs: Rethinking backpropagating refinement for interactive segmentation. In: CVPR (2020) 25", + "65. Sofiuk, K., Petrov, I.A., Konushin, A.: Reviving iterative training with mask guidance for interactive segmentation. In: IEEE International Conference on Image Processing (ICIP) (2022) 25", + "66. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In: NeurIPS (2020) 1, 4, 8, 10, 28, 32, 33", + "67. Song, J., Lee, J., Kweon, I.S., Choi, S.: Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. In: CVPR (2023) 1, 13, 24, 30", + "68. Song, J., Park, K., Shin, I., Woo, S., Zhang, C., Kweon, I.S.: Test-time adaptation in the dynamic world with compound domain knowledge management. IEEE Robotics and Automation Letters (2023) 24", + "69. Stiennon, N., Ouyang, L., Wu, J., Ziegler, D., Lowe, R., Voss, C., Radford, A., Amodei, D., Christiano, P.F.: Learning to summarize with human feedback. In: NeurIPS (2020) 25", + "70. Sun, B., Saenko, K.: Deep coral: Correlation alignment for deep domain adaptation. In: ECCV Workshops (2016) 1, 3", + "71. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 25", + "72. Tsai, Y.H., Hung, W.C., Schulter, S., Sohn, K., Yang, M.H., Chandraker, M.: Learning to adapt structured output space for semantic segmentation. In: CVPR (2018) 1, 3, 12", + "73. Tsai, Y.H., Hung, W.C., Schulter, S., Sohn, K., Yang, M.H., Chandraker, M.: Learning to adapt structured output space for semantic segmentation. In: CVPR (2018) 3", + "74. Van Engelen, J.E., Hoos, H.H.: A survey on semi-supervised learning. Machine learning (2020) 4" + ], + "bbox": [ + 223, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "J. Song et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "75. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: CVPR (2017) 9", + "76. Vu, T.H., Jain, H., Bucher, M., Cord, M., Pérez, P.: Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In: CVPR (2019) 3", + "77. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: ICLR (2021) 3, 24, 30", + "78. Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing (2018) 1", + "79. Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: CVPR (2022) 30", + "80. Wang, Y., Chen, H., Fan, Y., Sun, W., Tao, R., Hou, W., Wang, R., Yang, L., Zhou, Z., Guo, L.Z., et al.: Usb: A unified semi-supervised learning benchmark for classification. In: NeurIPS (2022) 10", + "81. Wang, Y., Chen, H., Heng, Q., Hou, W., Fan, Y., Wu, Z., Wang, J., Savvides, M., Shinozaki, T., Raj, B., et al.: Freematch: Self-adaptive thresholding for semi-supervised learning. In: ICLR (2023) 4, 7, 8, 10, 12, 13, 14, 24, 25, 27, 28, 30, 32, 33, 34", + "82. Warnell, G., Waytowich, N., Lawhern, V., Stone, P.: Deep tamer: Interactive agent shaping in high-dimensional state spaces. In: AAAI (2018) 25", + "83. Wei, C., Sohn, K., Mellina, C., Yuille, A., Yang, F.: Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In: CVPR (2021) 10, 22, 23, 24, 29", + "84. Wirth, C., Akhour, R., Neumann, G., Fünkranz, J., et al.: A survey of preference-based reinforcement learning methods. Journal of Machine Learning Research (2017) 25", + "85. Xie, B., Yuan, L., Li, S., Liu, C.H., Cheng, X.: Towards fewer annotations: Active learning via region impurity and prediction uncertainty for domain adaptive semantic segmentation. In: CVPR (2022) 12, 29", + "86. Xie, M., Li, Y., Wang, Y., Luo, Z., Gan, Z., Sun, Z., Chi, M., Wang, C., Wang, P.: Learning distinctive margin toward active domain adaptation. In: CVPR (2022) 4, 21", + "87. Xie, Q., Dai, Z., Hovy, E., Luong, T., Le, Q.: Unsupervised data augmentation for consistency training. In: NeurIPS (2020) 4, 8, 10, 32, 33, 34", + "88. Xu, Y., Shang, L., Ye, J., Qian, Q., Li, Y.F., Sun, B., Li, H., Jin, R.: Dash: Semi-supervised learning with dynamic thresholding. In: ICML (2021) 4", + "89. Yan, Z., Wu, Y., Li, G., Qin, Y., Han, X., Cui, S.: Multi-level consistency learning for semi-supervised domain adaptation. In: IJCAI (2022) 4", + "90. Yang, L., Wang, Y., Gao, M., Shrivastava, A., Weinberger, K.Q., Chao, W.L., Lim, S.N.: Deep co-training with task decomposition for semi-supervised domain adaptation. In: ICCV (2021) 4", + "91. Yang, S., Jui, S., van de Weijer, J., et al.: Attracting and dispersing: A simple approach for source-free domain adaptation. In: NeurIPS (2022) 3", + "92. Yang, S., Wang, Y., Van De Weijer, J., Herranz, L., Jui, S.: Generalized source-free domain adaptation. In: ICCV (2021) 1, 3", + "93. Yang, S., van de Weijer, J., Herranz, L., Jui, S., et al.: Exploiting the intrinsic neighborhood structure for source-free domain adaptation. In: NeurIPS (2021) 3, 27, 28, 35", + "94. Yu, Y.C., Lin, H.T.: Semi-supervised domain adaptation with source label adaptation. In: CVPR (2023) 4, 9, 10" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "NBF & RLD", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "95. Yu, Z., Li, J., Du, Z., Zhu, L., Shen, H.T.: A comprehensive survey on source-free domain adaptation. arXiv preprint arXiv:2302.11803 (2023) 3", + "96. Zhang, B., Wang, Y., Hou, W., Wu, H., Wang, J., Okumura, M., Shinozaki, T.: Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In: NeurIPS (2021) 7, 8, 10, 13, 32, 33, 34", + "97. Zhang, C., Miech, A., Shen, J., Alayrac, J.B., Luc, P.: Making the most of what you have: Adapting pre-trained visual language models in the low-data regime. arXiv preprint arXiv:2305.02297 (2023) 3", + "98. Zhang, W., Ouyang, W., Li, W., Xu, D.: Collaborative and adversarial network for unsupervised domain adaptation. In: CVPR (2018) 3", + "99. Zhang, Y., Zhang, H., Deng, B., Li, S., Jia, K., Zhang, L.: Semi-supervised models are strong unsupervised domain adaptation learners. arXiv preprint arXiv:2106.00417 (2021) 1, 4, 8", + "100. Zhang, Y., David, P., Gong, B.: Curriculum domain adaptation for semantic segmentation of urban scenes. In: ICCV (2017) 13", + "101. Zhang, Y., Wang, Z., He, W.: Class relationship embedded learning for source-free unsupervised domain adaptation. In: CVPR (2023) 3" + ], + "bbox": [ + 218, + 146, + 785, + 383 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "J. Song et al.", + "bbox": [ + 271, + 114, + 359, + 128 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/Is user feedback always informative_ Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data/0589fa53-73c5-413a-aac9-f37670a1de05_model.json b/2024/Is user feedback always informative_ Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data/0589fa53-73c5-413a-aac9-f37670a1de05_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e0a9fd405111957920b9ef3e84d600e6105d77a0 --- /dev/null +++ b/2024/Is user feedback always informative_ Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data/0589fa53-73c5-413a-aac9-f37670a1de05_model.json @@ -0,0 +1,2891 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.229, + 0.142, + 0.776, + 0.21 + ], + "angle": 0, + "content": "Is user feedback always informative? Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.234, + 0.74, + 0.266 + ], + "angle": 0, + "content": "Junha Song\\(^{1,2}\\), Tae Soo Kim\\(^{2}\\), Junha Kim\\(^{2}\\), Gunhee Nam\\(^{2}\\), Thijs Kooi\\(^{2}\\), and Jaegul Choo\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.279, + 0.277, + 0.724, + 0.308 + ], + "angle": 0, + "content": "1 KAIST {sb020518, jchoo}@kaist.ac.kr 2 Lunit Inc. {junha.kim, taesoo.kim, ghnam, tkooi}@lunit.io" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.343, + 0.744, + 0.593 + ], + "angle": 0, + "content": "Abstract. This paper aims to adapt the source model to the target environment, leveraging small user feedback (i.e., labeled target data) readily available in real-world applications. We find that existing semi-supervised domain adaptation (SemiSDA) methods often suffer from poorly improved adaptation performance when directly utilizing such feedback data, as shown in Figure 1. We analyze this phenomenon via a novel concept called Negatively Biased Feedback (NBF), which stems from the observation that user feedback is more likely for data points where the model produces incorrect predictions. To leverage this feedback while avoiding the issue, we propose a scalable adapting approach, Retrieval Latent Defending. This approach helps existing SemiSDA methods to adapt the model with a balanced supervised signal by utilizing latent defending samples throughout the adaptation process. We demonstrate the problem caused by NBF and the efficacy of our approach across various benchmarks, including image classification, semantic segmentation, and a real-world medical imaging application. Our extensive experiments reveal that integrating our approach with multiple state-of-the-art SemiSDA methods leads to significant performance improvements." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.606, + 0.721, + 0.634 + ], + "angle": 0, + "content": "Keywords: Rethinking user-provided feedback \\(\\cdot\\) Semi-supervised & Source-free domain adaptation \\(\\cdot\\) Medical image diagnosis" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.664, + 0.377, + 0.68 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.844 + ], + "angle": 0, + "content": "While deep neural networks have demonstrated remarkable performance in the development domain (i.e., source domain) [15,23], they often suffer from performance degradation in the deployed domain (i.e., target domain) due to domain shift [17, 72, 78]. To mitigate this issue, domain adaptation (DA) techniques have been introduced [34, 58, 70]. The most common DA tasks include semi-supervised domain adaptation (SemiSDA) and source-free domain adaptation (SFDA). SemiSDA aims to adapt the model given a small amount of labeled target data along with massive unlabeled target data [6, 58, 66, 99]. SFDA conducts adaptation with only target data without accessing source data considering data privacy or memory constraints in edge devices [34, 67, 92]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.129 + ], + "angle": 0, + "content": "J. Song et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.146, + 0.497, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.147, + 0.775, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.248, + 0.788, + 0.36 + ], + "angle": 0, + "content": "Fig. 1: (a) User feedback. Users can provide feedback while interacting with an ML product, where feedback is likely to be biased towards misclassified samples, which we define as Negatively Biased Feedback (NBF). (b) Adaptation results. We adapt the source model with small user feedback and large unlabeled target data using previous semi-supervised domain adaptation (SemiSDA) algorithms. Compared to random feedback, which is the classical SemiSDA setup where labeled data is a random subset of target data, model adaptation with NBF leads to subpar performance. This paper analyzes this problem and introduces a scalable solution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.789, + 0.536 + ], + "angle": 0, + "content": "Despite such advances in DA, adapting the model with user feedback still remains an open area for further research, even though practical machine learning (ML) products often allow users to provide feedback in order to further improve the model in the target environment. For example, facial recognition or medical image diagnosis applications enable users to give feedback correcting wrong model predictions, as depicted in Figure 1 (a). Since feedback can be modeled in this case as a small amount of labeled target data, it is anticipated that previous SemiSDA methods assuming the same setup would yield promising results. However, we observe that they show inferior adaptation performance on multiple DA benchmarks when using such user feedback in practice, as shown in the dark-gray bar in Figure 1 (b)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.789, + 0.718 + ], + "angle": 0, + "content": "We introduce a novel concept called Negatively Biased Feedback (NBF) to explain this phenomenon. NBF is based on the observation that user feedback is more likely to be derived from incorrect model predictions. For example, a radiologist might log a misdiagnosed chest X-ray by the model, as its accuracy directly impacts the patient's survival. Interestingly, our observation aligns with findings from cognitive psychology literature [3,57] that proves that humans are more likely to react and provide feedback to negative events (i.e., wrong model predictions). Since such an NBF scenario is feasible, we analyze its unexpected impact on SemiSDA observed above. We identify that a biased distribution of NBF within the overall data distribution leads to sub-optimal adaptation results, particularly compared to Random Feedback (RF). RF represents the classical SemiSDA setup, where labeled data is randomly selected from the target data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "To address the problem caused by NBF, we present a scalable approach named Retrieval Latent Defending, which can be seamlessly integrated with existing SemiSDA methods. Our approach allows them to adapt the model without a strong dependence on the biasedly distributed labeled data. Specifically, we balance the supervised adapting signal by appending latent defending samples to the mini-batch and help to keep the model's balanced class discriminability throughout adapting iterations. We evaluate the unexpected influence of NBF using various benchmarks, including image classification, semantic segmentation," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "NBF & RLD" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.778, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.248, + 0.787, + 0.318 + ], + "angle": 0, + "content": "Fig. 2: Adaptation with user feedback can be effective in alleviating performance degradation caused by domain shift. However, there are some challenges: (i) user feedback may be a biased sampling of the true target distribution due to the nature of feedback, (ii) the amount of the ground truths (GT) labels obtained through feedback is small, and (iii) only unlabeled target data is typically available, not source data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.784, + 0.375 + ], + "angle": 0, + "content": "and medical image diagnosis. Building upon these evaluations, we demonstrate that our approach not only complements, but significantly enhances the performance of multiple SemiSDA methods." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.378, + 0.57, + 0.393 + ], + "angle": 0, + "content": "The contributions of the paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.396, + 0.785, + 0.439 + ], + "angle": 0, + "content": "We introduce the novel concept called Negatively Biased Feedback and uncover that it can lead to sub-optimal adaptation performance of existing SemiSDA methods." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.441, + 0.785, + 0.485 + ], + "angle": 0, + "content": "- We analyze this problem and present a scalable solution, Retrieval Latent Defending, that combines with SemiSDA methods and allows them to avoid the unexpected effect of NBF." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.486, + 0.785, + 0.514 + ], + "angle": 0, + "content": "- We show that our approach generalizes through diverse DA benchmarks and improves adaptation results of state-of-the-art SemiSDA methods." + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.515, + 0.785, + 0.543 + ], + "angle": 0, + "content": "We publicly release the code on https://github.com/junha1125/RLD-SemiSDA." + }, + { + "type": "list", + "bbox": [ + 0.232, + 0.396, + 0.785, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.554, + 0.388, + 0.57 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Adaptation in the deployment environment. Real-world ML products often encounter performance degradation caused by gaps between the source and target environment [17]. One solution is to adapt the model using unlabeled data observed in the target domain, referred to as unsupervised domain adaptation (UDA) [37, 59, 72]. Works on UDA use both source and target data to improve the target performance by using methods such as domain discrepancy minimization by adversarial training [18, 41, 59, 70, 72, 73, 76], and self-training with pseudo labels [45, 51, 97, 98]. Source-free DA (SFDA) builds on UDA and imposes an additional constraint that the source data can not be accessed during domain adaptation. This has practical implications for addressing data privacy concerns or barriers in data transmission to edge devices [34, 38, 77, 95]. The majority of recent SFDA works rely on strategies like domain clustering [34], nearest neighbors [91-93], and contrastive learning [8, 35, 101]. Nevertheless, SFDA does not consider the availability of small labeled data, which may be available in practical ML systems. Semi-supervised DA (SemiSDA) works mainly demonstrate that permitting small labeled data in the target domain can substantially enhance adaptation performance compared to traditional UDA [58]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.129 + ], + "angle": 0, + "content": "J. Song et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "Their primary strategy is to use domain alignment [20, 33, 58, 94], multi-view consistency [2, 6, 33, 89], and asymmetric co-training [36, 90]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.182, + 0.785, + 0.287 + ], + "angle": 0, + "content": "Active domain adaptation (ActiveDA) [25, 55, 86] envisions a scenario in which the machine selects specific target samples and instructs annotators to label them. The primary objective of ActiveDA is to strategically identify and select the most informative samples for annotation. These chosen samples (i.e., labeled target data) are subsequently utilized to update the source model using SemiSDA methods [33, 58], and the effectiveness of ActiveDA is assessed by evaluating the target performance of the adapted model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.292, + 0.785, + 0.397 + ], + "angle": 0, + "content": "Semi-supervised learning (SemiSL) aims to reduce expensive human annotations, and propose methods to train a model from scratch using massive unlabeled data along with limited amounts of labeled data [43,74]. The majority of SemiSL methods depend on consistency regularization [4,5,16,60,66,87], which helps the model to make similar predictions for augmented versions of the same image. Moreover, adaptive thresholding [9,12,24,66,81,88,99] is also popularly utilized to produce reliable pseudo labels from unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.785, + 0.489 + ], + "angle": 0, + "content": "SemiSDA and SemiSL setups mimic small labeled datasets by randomly selecting subsets of the target dataset, whereas ActiveDA involves selections instructed by the machine. In contrast, this paper posits that in real-world applications, labeled data is typically acquired through user intervention. Additionally, users often provide feedback on samples misclassified by the model (i.e., negatively biased feedback), a process detailed in the following section." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.495, + 0.782, + 0.535 + ], + "angle": 0, + "content": "
UDASFDAActiveDASemiSDASemiSLOur setup
Adaptation○○○○×○
Source-free×○××-○
Feedback××machine-instructedrandomly selectedrandomly selecteduser-provided
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.785, + 0.61 + ], + "angle": 0, + "content": "The table above summarizes the comparison of relevant studies to our setup. In the table, adaptation means fine-tuning the source pre-trained model (as opposed to training from scratch); feedback represents a small number of labeled target samples. Appendix A provides further comparisons with settings like class-imbalanced SemiSDA and test-time adaptation (TTA)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.627, + 0.522, + 0.646 + ], + "angle": 0, + "content": "3 Negatively Biased Feedback" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.656, + 0.52, + 0.671 + ], + "angle": 0, + "content": "3.1 Adaptation with user feedback." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Our adaptation setup is illustrated in Figure 2. A model is pre-trained on the source data \\( D_{s} \\). Next, the model is deployed to the target domain, such as a smartphone or a hospital, where we assume the transfer of \\( D_{s} \\) is prohibited due to data privacy regulations or resource constraints (same setup as SFDA [34]). While users utilize ML products on the target domain, the model provides prediction results for data observed in the target domain \\( D_{t} \\) and occasionally obtains user feedback in the form of annotations \\( y \\). We represent the target data as \\( D_{t} = X_{t}^{lb} \\cup X_{t}^{ulb} \\), where \\( X_{t}^{lb} = \\{(x_{lb}^{n},y_{lb}^{n}):n\\in [1..N_{lb}]\\} \\), \\( X_{t}^{ulb} = \\{(x_{ulb}^{n}):n\\in [1..N_{ulb}]\\} \\), \\( x_{lb} \\) and \\( x_{ulb} \\) denote labeled and unlabeled data and \\( N_{lb} \\) and \\( N_{ulb} \\) is their number of data. Lastly, the model can utilize \\( D_{t} \\) and SemiSDA algorithms for adaptation during its inactive phase (e.g., when users do not use the product," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "NBF & RLD" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.754, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.35, + 0.788, + 0.447 + ], + "angle": 0, + "content": "Fig.3: Effect of negatively biased feedback. Our novel observations are that \\((a)\\) user-provided feedback in practice has a biased distribution in each class cluster (the bottom center sub-figure) which is in contrast to random feedback, \\((b)\\) Existing SemiSDA methods adapt the model by dominating the labeled data points (the right sub-figures) even though they are biasedly positioned, and \\((c)\\) NBF prevents the model from having a decision boundary for true class clusters and leads to inferior adaptation performance (the bottom right sub-figure)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.785, + 0.492 + ], + "angle": 0, + "content": "like at nighttime) in order to alleviate performance degradation due to domain shift or to personalize the model based on user feedback." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.789, + 0.737 + ], + "angle": 0, + "content": "Rethinking user-provided feedback. Classical SemiSDA works simply assume that a random subset in target data \\( D_{t} \\) is labeled by users when building \\( X_{t}^{lb} \\). However, as illustrated in Figure 2 (i), we suggest that users are more likely to provide feedback on misclassified samples by the source model, named negatively biased feedback (NBF). This behavior can be understood from two perspectives: (a) users generally expect their feedback to be used as a basis of model improvement, motivating them to provide NBF, and (b) humans tend to react more strongly to negative experiences, such as receiving incorrect predictions, as observed in psychological studies [3,57]. We note that the NBF assumption holds more strongly for the medical application: it is reasonable to imagine that the user (i.e., radiologist) logs the mistakes of the model while diagnosing a chest X-ray exam because the diagnostic accuracy of the model is directly related to the patient's chances of survival. Furthermore, applications beyond the medical domain can also exhibit NBF. For instance, users in self-driving cars can report errors, such as object detection failures or navigation mistakes, to enhance the car's driving capabilities." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.754, + 0.517, + 0.768 + ], + "angle": 0, + "content": "3.2 Influence of NBF on SemiSDA" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Simulation study. As shown in Figure 3, we conduct a simulation study to understand the effect of NBF on SemiSDA. We first use the blobs dataset [53] and construct the source and target data so that domain shift exists between them (left sub-figures). We pre-train a source model on the source data and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.363, + 0.129 + ], + "angle": 0, + "content": "J. Song et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "compute the accuracy in the target domain, where the performance drop due to domain shift is observed \\((98.5\\% \\rightarrow 76.4\\%)\\). Next, we simulate two types of feedback (i.e., labeled data): random feedback and negatively biased feedback following a previous SemiSDA setup and our setup, respectively. Specifically, NBF is randomly selected among misclassified samples by the source model. We find that random feedback (RF) points are evenly distributed, while NBF points are biasedly positioned across each class cluster (refer to blue points in the dashed circle in the center sub-figures)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.45 + ], + "angle": 0, + "content": "To alleviate the performance drop caused by domain shift, we adapt the model using the target data and a semi-supervised method, Pseudo-labeling [1]. This method iteratively optimizes the model by the cross-entropy loss computed by the ground truth of labeled data and pseudo labels of unlabeled data in a minibatch (pseudo labels are predicted by the current adapting model so they can be changed according to an updating decision boundary. Further comprehension can be achieved by referring Appendix B.). The SemiSDA results are shown in the right sub-figures, where we make two interesting observations: (i) the distribution of labeled data can contribute significantly to a decision boundary of the adapted model (red arrows in the figure), and (ii) the adapted model under NBF has poorly improved performance compared with one under RF \\((76.4\\% \\rightarrow 88.1\\%\\) with NBF, but \\(76.4\\% \\rightarrow 91.7\\%\\) with RF)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.788, + 0.621 + ], + "angle": 0, + "content": "Unexpected influence of NBF. Our intuitive reasoning probably suggests that NBF provides more information than RF by correcting more source model deficiencies, and thus leads to better adaptation performance. However, we empirically show that NBF can result in inferior adaptation performance due to its biased distribution across each class cluster, as illustrated in Figure 3. Surprisingly, we also show that this problem persists, even with other state-of-the-art SemiSDA methods and large datasets for various DA benchmarks, including image classification, semantic segmentation, and medical image diagnosis. Our work highlights the importance of careful design when using user feedback in real-world scenarios and, to the best of our knowledge, is the first study to uncover and analyze this phenomenon." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.631, + 0.348, + 0.648 + ], + "angle": 0, + "content": "4 Approach" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.656, + 0.6, + 0.673 + ], + "angle": 0, + "content": "4.1 Prerequisite: Previous SemiSDA method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.678, + 0.788, + 0.769 + ], + "angle": 0, + "content": "Previous SemiSDA and SemiSL works typically construct a mini-batch with labeled data \\(\\{(x_{lb}^b, y_{lb}^b) : b \\in [1..B]\\}\\), and unlabeled data whose size is \\(\\mu\\) times larger than labeled ones \\(\\{(x_{ulb}^b) : b \\in [1..\\mu \\cdot B]\\}\\), where \\(B\\) is the mini-batch size for labeled data. To adapt the model iteratively, they compute the cross-entropy loss \\(\\mathcal{H}(\\cdot, \\cdot)\\) with labeled data and the consistency regularization to multi-view of unlabeled data, which are formulated as the following:" + }, + { + "type": "equation", + "bbox": [ + 0.268, + 0.771, + 0.787, + 0.807 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s u p} = \\frac {1}{B} \\sum_ {b = 1} ^ {B} \\mathcal {H} \\left(y _ {l b} ^ {b}, f _ {\\theta} \\left(x _ {l b} ^ {b}\\right)\\right), \\quad \\mathcal {L} _ {u n s u p} = \\frac {1}{\\mu \\cdot B} \\sum_ {b = 1} ^ {\\mu \\cdot B} \\mathcal {H} \\left(\\hat {y} _ {u l b} ^ {b}, f _ {\\theta} \\left(\\Omega \\left(x _ {u l b} ^ {b}\\right)\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "where \\( f_{\\theta}(\\cdot) \\) is the output probability from the model, \\( \\hat{y}_{ulb} \\) denotes a pseudo label obtained from \\( f_{\\theta}(\\omega(x_{ulb})) \\), and \\( \\omega(\\cdot) \\) and \\( \\Omega(\\cdot) \\) represent weak and strong image" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "NBF & RLD" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.144, + 0.761, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.366, + 0.788, + 0.465 + ], + "angle": 0, + "content": "Fig. 4: Even though labeled data \\((x_{lb}, y_{lb})\\) is biasedly positioned, the model needs to be adapted with balanced class discriminability (i.e., decision boundary). (i) However, previous SemiSDA methods have overlooked this fact and used the labeled data naively by applying a cross-entropy loss, leading to inadequate adaptation performance. (ii) To alleviate this problem, we propose a scalable adapting approach, retrieval latent defending, which allows the model to adjust the balance of a mini-batch on each iteration by using latent defending samples \\(x_{LD}\\) together with labeled data \\(x_{lb}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.785, + 0.522 + ], + "angle": 0, + "content": "augmentation, respectively. While sharing the core framework, each SemiSDA method employs distinct adapting strategies, especially to enhance the effectiveness of the use of unlabeled data rather than labeled data [6,81,96]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.788, + 0.618 + ], + "angle": 0, + "content": "Problem of previous works. Since previous SemiSDA methods have overlooked the unexpected impact of NBF, they often suffer from sub-optimal performance under the NBF assumption (shown in Section 5). To address this problem, we focus on developing a scalable solution that (i) can easily combine with existing DA methods without modifying their core adapting strategies and (ii) can be applied to a wide range of benchmarks, including medical image diagnosis." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.636, + 0.488, + 0.652 + ], + "angle": 0, + "content": "4.2 Retrieval Latent Defending" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Based on the observations in Figure 3, we illustrate the unintended effect of NBF when using an existing SemiSDA method in Figure 4 (top center), where NBF is likely to exhibit a biased distribution, leading to undesirable adaptation results. To alleviate this issue, we propose Retrieval Latent Defending as depicted in Figure 4 (bottom). \\(①\\) Prior to each epoch, we generate a candidate bank of data points, denoted as \\(x_{LD}\\). \\(② \\sim ④\\) For each adapting iteration, we balance the mini-batch by retrieving latent defending samples \\(x_{LD}\\) from the bank. \\(⑤ \\sim ⑥\\) The model is then adapted using the reconfigured mini-batch and following the baseline SemiSDA approach. We hypothesize that the latent space progressively created by the \\(x_{LD}\\) candidates throughout the adaptation process (bold dashed circle in Figure 4 (top right)) mitigates the issue caused by NBF, thereby allowing the SemiSDA baseline to achieve robust adaptation against NBF." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.363, + 0.129 + ], + "angle": 0, + "content": "J. Song et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "Candidate bank generation. The candidate bank serves as a repository of pseudo labels \\(\\hat{Y}_t^{ulb}\\) for a subset of the target unlabeled data \\(X_t^{ulb}\\). Before each epoch, we freeze the model and use it to generate pseudo labels \\(\\hat{Y}_t^{ulb} = \\{(\\hat{y}_{ulb}^n): n \\in [1..N_{ulb}]\\}\\), where \\(\\hat{y}_{ulb}^n\\) is assigned to \\(x_{ulb}^n\\) as the predicted class with the highest softmax probability: \\(\\hat{y}_{ulb}^n = \\operatorname{argmax}_c [f_\\theta(x_{ulb}^n)]_c\\). We then retain only samples with the top \\(p\\%\\) highest probabilities within each class. This filtering step helps mitigate the inclusion of data with potentially inaccurate pseudo labels, as the model's predictions on \\(X_t^{ulb}\\) might not always be perfect." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.439 + ], + "angle": 0, + "content": "Defending sample selection. We select \\( k \\) latent defending samples \\( x_{LD} \\) from the bank at random for each labeled data \\( (x_{lb}^{b}, y_{lb}^{b}) \\). These selected samples share the same pseudo label as the ground-truth label of their associated counterparts (i.e., \\( \\hat{y}_{LD} = y_{lb}^{b} \\)). By incorporating these defending samples, we balance the data distribution within the current mini-batch. For example, consider \\( x_{lb}^{1} \\) and \\( x_{lb}^{2} \\) in Figure 4 (top right). As these labeled samples are included in the current mini-batch alongside the selected defending samples \\( x_{LD}^{1} \\) and \\( x_{LD}^{2} \\), we expect to prevent the supervised adapting signal from becoming overly dependent on the labeled samples. We imagine the effect of the defending samples throughout the adaptation process and depict the latent space formed gradually by the \\( x_{LD} \\) candidates as bold dashed circles in Figure 4 (top right)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.787, + 0.47 + ], + "angle": 0, + "content": "Consequently, the overall loss consists of the sum of losses in Eq. (1) and a loss from our proposed method as," + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.473, + 0.787, + 0.524 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\underbrace {\\mathcal {L} _ {\\text {s u p}} + \\mathcal {L} _ {\\text {u n s u p}}} _ {\\text {b a s e l i n e}} + \\underbrace {\\frac {1}{k \\cdot B} \\sum_ {b = 1} ^ {k \\cdot B} \\mathcal {H} \\left(y _ {L D} ^ {b} , f _ {\\theta} \\left(x _ {L D} ^ {b}\\right)\\right)} _ {\\text {r e t r i e v a l l a t e n t d e f e n d i n g}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.532, + 0.788, + 0.653 + ], + "angle": 0, + "content": "Importance of our method. Understanding the impact of NBF on adaptation performance is crucial. For example, naively adapting a model for a medical application using radiologist-provided feedback can actually cause performance degradation (shown in Table 5), potentially posing significant risks to patients. We propose a scalable and simple approach to solve the problem caused by NBF, which can not be addressed by existing methods. Given the practicality of the NBF problem and the scalability of our solution, we believe our work holds considerable potential for real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.667, + 0.378, + 0.684 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.692, + 0.436, + 0.708 + ], + "angle": 0, + "content": "5.1 Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.716, + 0.787, + 0.778 + ], + "angle": 0, + "content": "Our approach is simple enough to seamlessly combine with existing SemiSDA algorithms and also be applied to diverse benchmarks. This section describes our experimental setup for natural image classification tasks and a real-world medical application. Details for semantic segmentation experiments are in Appendix D." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Baselines. We validate our approach by combining various state-of-the-art algorithms for SemiSDA [58] (e.g., CDAC [33] and AdaMatch [6]) and SemiSL [66,87] (e.g., FlexMatch [96] and FreeMatch [81]). Note that the SemiSL methods have been demonstrated to be strong SemiSDA learners [99], so we can consider them" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "NBF & RLD" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table", + "bbox": [ + 0.237, + 0.145, + 0.768, + 0.208 + ], + "angle": 0, + "content": "
methodfeedbackaverage\\( \\mathrm{r} \\rightarrow \\mathrm{c} \\)\\( \\mathrm{r} \\rightarrow \\mathrm{p} \\)\\( \\mathrm{p} \\rightarrow \\mathrm{c} \\)\\( \\mathrm{c} \\rightarrow \\mathrm{s} \\)\\( \\mathrm{s} \\rightarrow \\mathrm{p} \\)\\( \\mathrm{r} \\rightarrow \\mathrm{s} \\)\\( \\mathrm{p} \\rightarrow \\mathrm{r} \\)
RerNetAdaMatch [6]RF67.666.668.568.560.369.258.781.5
NBF\\( {64.5}\\left( {+{3.1}}\\right) \\)64.366.165.656.965.654.278.9
w/ oursNBF\\( \\mathbf{{72.0}\\left( {+{7.5}}\\right) } \\)74.572.773.965.570.064.383.2
ViTAdaMatch [6]RF74.775.376.973.868.076.367.185.5
NBF\\( {73.7}\\left( {+{1.0}}\\right) \\)74.776.274.765.774.066.884.0
w/ oursNBF\\( \\mathbf{{75.9}\\left( {+{2.2}}\\right) } \\)76.977.877.868.576.668.385.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.207, + 0.788, + 0.304 + ], + "angle": 0, + "content": "Table 1: Adaptation results on DomainNet-126. We simulate seven domain-shift scenarios (i.e., source \\(\\rightarrow\\) target). The model is pre-trained on the source domain and then adapted to a training set of the target domain. The results on the test set of the target domain are reported as the top-1 accuracy (\\%). DomainNet-126 [54,58] dataset includes real, painting, sketch, and clip-art domains. In this experiment, we assume that the 378 feedback samples (i.e., 3 labeled data per class) are obtained from users. A state-of-the-art SemiSDA method, AdaMatch [6], is used as a baseline." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.309, + 0.766, + 0.344 + ], + "angle": 0, + "content": "
methodfeedbackaveragea→ca→pa→rc→ac→pc→rp→ap→cp→rr→ar→cr→p
AdaMatch [6]RF70.955.480.475.965.781.574.665.958.778.468.861.584.3
NBF69.3 (+1.6)54.276.675.365.979.375.563.757.475.966.756.884.2
w/oursNBF73.8 (+4.5)62.281.079.768.885.478.667.761.779.569.064.188.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.785, + 0.386 + ], + "angle": 0, + "content": "Table 2: Adaptation results on OfficeHome. OfficeHome [75] dataset includes real, product, art, and clip-art domain. We assume that the 195 feedback samples (i.e., 3 labeled data per class) are obtained. AdaMatch [6] and ResNet-50 [23] are used." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.787, + 0.426 + ], + "angle": 0, + "content": "as SemiSDA methods. For medical experiments, we use Pseudo-labeling [1] as a baseline since it is easily applicable to medical image adaptation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.787, + 0.473 + ], + "angle": 0, + "content": "Datasets. We utilize natural image datasets containing multiple kinds of domains (e.g., real and painting). The datasets include DomainNet-126 [54,58] with 142k images of 126 classes, and OfficeHome [75] with 15K images of 65 classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.788, + 0.64 + ], + "angle": 0, + "content": "To conduct medical experiments, we present a practical medical setting. We adopt the MIMIC-CXR-V2 dataset [27]. It assumes a multi-finding binary classification setup, where multiple radiographic findings, like Pneumonia and Atelectasis, can coexist in a single chest X-ray (CXR) sample. Thus, the model predicts the presence or absence (binary classes) of each individual finding. We simulate domain shift by using Posterior-Anterior (PA)-view data as the source and AP-view data as the target, capturing real-world variations in data acquisition. Typically, patients requiring an AP X-ray are those facing positioning challenges that prevent them from undergoing a PA X-ray. Therefore, this setup can be seen as a scenario where the target environment is the intensive care unit, which hospitalizes critically ill patients." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.788, + 0.732 + ], + "angle": 0, + "content": "Following the recent SemiSDA [94] and SFDA [8] setups, we assume the model is pre-trained in the source domain and deployed in the target domain. Since the datasets above were not initially divided into training and test sets, we performed a random 8:2 split within each domain, designating them respectively for training and testing. The training set is used to adapt the model, while the test set is used to report the top-1 accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.842 + ], + "angle": 0, + "content": "User feedback. Feedback given by users is modeled as annotations \\(\\{y_{lb}^n : n \\in [1..N_{lb}]\\}\\) on a small subset of the target's training set \\(D_t^{train}\\), while the remaining of them are used as unlabeled target data. In our experiments, we take into account two types of feedback: random feedback (RF) and negatively biased feedback (NBF). RF is the same setup of classical SemiSDA and SemiSL, where randomly selected samples from \\(D_t^{train}\\) are used as small labeled set \\(X_t^{lb}\\). For NBF, we randomly select samples that are incorrectly predicted in \\(D_t^{train}\\) by" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.129 + ], + "angle": 0, + "content": "J. Song et al." + }, + { + "type": "table", + "bbox": [ + 0.304, + 0.145, + 0.7, + 0.306 + ], + "angle": 0, + "content": "
feed, amount method378 (3 labeled data per class)630 (5 labeled data per class)
RFNBFw/oursRFNBFw/ours
Source model56.5
MME [58]69.568.4 (-1.1)70.8 (+2.4)71.270.1 (-1.1)72.5 (+2.4)
CDAC [33]68.364.6 (-3.7)73.2 (+8.6)71.768.1 (-3.6)74.9 (+6.8)
AdaMatch [6]67.664.5 (-3.1)72.0 (+7.5)70.967.7 (-3.2)74.3 (+6.6)
FixMatch [66]67.663.4 (-4.2)73.2 (+9.8)71.566.1 (-5.4)75.1 (+9.0)
UDA [87]69.264.9 (-4.3)73.4 (+8.5)72.968.8 (-4.1)75.3 (+6.5)
FlexMatch [96]73.371.4 (-1.9)74.7 (+3.3)75.373.9 (-1.4)76.0 (+2.1)
FreeMatch [81]73.872.0 (-1.8)74.8 (+2.8)75.674.4 (-1.2)76.1 (+1.7)
Fully supervised83.6
Source model64.5
MME [58]73.272.7 (-0.5)74.1 (+1.4)74.574.0 (-0.5)75.2 (+1.2)
CDAC [33]74.272.8 (-1.4)75.4 (+2.6)75.474.1 (-1.3)76.2 (+2.1)
AdaMatch [6]74.773.7 (-1.0)75.9 (+2.2)75.975.1 (-0.8)76.7 (+1.6)
FixMatch [66]74.673.0 (-1.6)75.6 (+2.6)75.774.3 (-1.4)76.5 (+2.2)
UDA [87]74.873.3 (-1.5)75.8 (+2.5)75.974.5 (-1.4)76.7 (+2.2)
FlexMatch [96]74.973.9 (-1.0)75.8 (+1.9)76.075.1 (-0.9)76.9 (+1.8)
FreeMatch [81]74.973.9 (-1.0)75.7 (+1.8)76.075.1 (-0.9)76.8 (+1.7)
Fully supervised85.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.306, + 0.787, + 0.36 + ], + "angle": 0, + "content": "Table 3: Comparisons on DomainNet-126. We evaluate our method by integrating it with SemiSDA and SemiSL methods. The average accuracy of seven domain-shift scenarios in Table 1 is reported. Source model represents the pre-trained model without adaptation. Fully supervised means the model is adapted with fully labeled target data." + }, + { + "type": "table", + "bbox": [ + 0.304, + 0.365, + 0.7, + 0.456 + ], + "angle": 0, + "content": "
feed, amount method195 (3 labeled data per class)325 (5 labeled data per class)
RFNBFw/oursRFNBFw/ours
Source model57.6
MME [58]71.270.2 (-1.0)73.4 (+3.2)73.573.1 (-0.4)75.6 (+2.5)
CDAC [33]71.269.0 (-2.2)74.3 (+5.3)73.572.3 (-1.2)75.7 (+3.4)
AdaMatch [6]70.969.3 (-1.6)73.8 (+4.5)73.472.7 (-0.7)75.5 (+2.8)
FixMatch [66]71.468.6 (-2.8)73.7 (+5.1)73.972.2 (-1.7)75.3 (+3.1)
UDA [87]72.269.5 (-2.7)74.1 (+4.6)74.473.0 (-1.4)76.0 (+3.0)
FlexMatch [96]73.772.1 (-1.6)74.7 (+2.6)75.974.9 (-1.0)76.6 (+1.7)
FreeMatch [81]74.072.7 (-1.3)74.8 (+2.1)75.875.0 (-0.8)76.6 (+1.6)
Fully supervised87.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.457, + 0.785, + 0.484 + ], + "angle": 0, + "content": "Table 4: Comparisons on OfficeHome. The average accuracy of twelve domain-shift scenarios in Table 2 is reported. ResNet-50 is used." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.787, + 0.572 + ], + "angle": 0, + "content": "the source model (i.e., the pre-trained model before adaptation). Note that we focus on the impact of a biased label distribution within the same class, as shown in Figure 3, and thus take the same number of feedback for each class. Further discussion about the imbalance in the number of feedback between classes presented in [31, 49, 83] is provided in Appendix A.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.787, + 0.666 + ], + "angle": 0, + "content": "Network architectures. We adopt commonly used networks, ResNet [23] and ViT [15] for natural image tasks and DenseNet [26] for a medical task. We employ ResNet-50 with the last classification layer comprising a weight normalization layer and a bottleneck layer, following previous works [8, 34] and use the ViT-Small (i.e., ViT-S) introduced in [80]. The DenseNet-121 is used, provided in TorchXrayVision [13], like existing medical works [32, 44]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.668, + 0.787, + 0.775 + ], + "angle": 0, + "content": "Implementation details. We implement our framework by extending the publicly available USB [80] repository. Both pre-training and adaptation are conducted with a mini-batch size of 128 and the SGD optimizer. Diverse baselines for SemiSDA and SemiSL are used to compute the losses in Eq. (1). The hyperparameters for each baseline simply follow USB [80] or public code [33,58]. For all experiments, our approach uses the same hyper-parameters of the appended defending samples \\( k \\) and reliable filtering rate \\( p \\) as 3 and 0.4, respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.79, + 0.373, + 0.805 + ], + "angle": 0, + "content": "5.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Natural image classification. Following recent DA works [8, 94], we conduct experiments on seven and twelve domain shift scenarios provided with the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "NBF & RLD" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.805, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.353, + 0.788, + 0.424 + ], + "angle": 0, + "content": "Table 5: Adaptation in a medical application. We use samples with PA-view as the source data and samples with AP-view as the target data in MIMIC-CXR-V2 dataset [27]. NBF-CE represents a scenario when NBF is composed of cases with confident errors. We use DensNet-121 [13, 26] and assume the 20 feedback for the absence and presence per finding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.536 + ], + "angle": 0, + "content": "DomainNet-126 and OfficeHome datasets, respectively. Table 1 and Table 2 show the results, where AdaMatch [6] is used as the baseline. We observe consistent results with Figure 3 even on large natural datasets: when simply applying the baseline under the NBF assumption, the adapted model shows inferior performance for most domain shifts than applying it under RF, e.g., \\(64.5 < 67.6\\). Combining our approach with the baseline mitigates this issue and achieves a performance increase, e.g., \\(64.5 \\rightarrow 72.0\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.657 + ], + "angle": 0, + "content": "We also use other promising baselines and report the average accuracy of all domain shifts in Table 3 and Table 4 (all results can be found in Appendix F). While both feedback types bring performance improvement from the source model, lower performance is observed with NBF. Our method enables the baselines to not only address this problem but surpass performance under RF. The above results suggest that the biased distribution of labeled samples, which has been overlooked in previous SemiSDA works, is actually problematic, and our retrieval latent defending approach is effective." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.661, + 0.788, + 0.751 + ], + "angle": 0, + "content": "Medical image diagnosis. Table 5 shows the results (bottom) and also depicts the effect of NBF (top center). We report the AUROC [7] for each finding following standard practice for measuring computer-aided-diagnosis model evaluation [32,44]. The baseline SemiSDA method under NBF exhibits inferior performance compared to one under RF, but this issue can be mitigated by combining our approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.751, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In addition, we propose an interesting and practical scenario named NBF with more confident errors (NBF-CE). In this scenario, we assume that a radiologist is likely to give feedback when the model makes confidently wrong predictions. Imagine that the model predicts a \\(1\\%\\) likelihood of cancer in a CXR image, but the person actually has cancer. Such failure to detect potential patients early on can significantly reduce the patient's chances of survival, so a radiologist may" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.129 + ], + "angle": 0, + "content": "J. Song et al." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.144, + 0.782, + 0.178 + ], + "angle": 0, + "content": "
labeling typefeed. amountRFNBFNBF w/oursENT [62]ENT w/ours
IAST [46,63]PA, 40 points55.353.0 (-2.3)56.3 (+3.3)53.556.0 (+2.5)
RIPU [85]PA, 40 points57.654.5 (-3.1)58.0 (+3.5)54.657.7 (+3.1)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.178, + 0.784, + 0.207 + ], + "angle": 0, + "content": "Table 6: Adaptation on semantic segmentation. The GTA5 [56] \\(\\rightarrow\\) Cityscapes [14] setup is used [72]. The target performance of the source model is \\(36.6\\mathrm{mIoU}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.214, + 0.788, + 0.336 + ], + "angle": 0, + "content": "provide feedback to the model. To simulate NBF-CE, we select samples where the source model most confidently predicts a finding to be absent \\((\\hat{y} \\approx 0)\\) although it is clearly visible in the radiograph \\((y = 1)\\), and vice versa, i.e., samples of \\(\\hat{y} \\approx 1\\) but \\(y = 0\\). Table 5 also shows the results under an NBF-CE scenario, where the model's adaptation performance is further reduced compared with NBF (0.7691 for \\(\\mathrm{NBF} \\to 0.7639\\) for NBF-CE). By combining our method, we observe performance improvements for both NBF variants, e.g., 0.7639 for \\(\\mathrm{NBF - CE} \\to 0.7875\\) with ours. We illustrate the hypothesized impact of our method in Table 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.789, + 0.552 + ], + "angle": 0, + "content": "Semantic segmentation. We evaluate the influence of NBF and our approach on a semantic segmentation task. We utilize the most common adaptation benchmark of GTA5 [56] to Cityscapes [14]. The baseline DA algorithms are used as IAST [46, 63] and RIPU [85] in a source-free scenario. We regard Pixel-based Annotation (PA) in which we assume 40 pixels per image like LabOR [63]. Table 6 shows results similar to those we observed in the classification and medical imaging tasks. The baselines under NBF exhibit inferior performance compared to those under RF (54.5 for \\(\\mathrm{NBF} < 57.6\\) for RF), but this issue is addressed by combining our approach with them \\((+3.5\\mathrm{mIoU})\\). Although out of our scope (refer to Appendix A.1), we validate one active labeling strategy ENT [62], which assigns highly uncertain (i.e., probably misclassified) pixels as feedback. Consequently, the feedback instructed by ENT is biasedly distributed in a manner similar to NBF. ENT also causes unexpected results (54.6 for \\(\\mathrm{ENT} < 57.6\\) for RF), and our approach alleviates this issue \\((+3.1\\mathrm{mIoU})\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.566, + 0.39, + 0.581 + ], + "angle": 0, + "content": "5.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.591, + 0.787, + 0.667 + ], + "angle": 0, + "content": "Positive vs. Negative feedback. We study the role of feedback on the adaptation results by varying feedback configurations. Let positively-provided feedback (PF) be obtained from samples that the source model correctly predicts, as opposed to negatively-provided feedback (NF). We adjust the ratio of PF:NF while keeping the total number of labeled samples constant, as shown in Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.788, + 0.789 + ], + "angle": 0, + "content": "When using only FreeMatch (gray dot-dashed line), both biased feedback types (i.e., NBF and PBF) result in worse adaptation performance compared to balanced feedback for the baseline, e.g., 72.6 in 378:0 (PBF) \\(< 73.3\\) in 252:126. In contrast, when our method is applied (red line), NBF yields the best performance. PBF and NBF can be respectively regarded as contributing previously known knowledge of the model and new knowledge that complements model deficiencies. Hence, it may be natural that NBF, which actually encodes the model's mistakes, contributes to favorable adaptation results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.791, + 0.788, + 0.823 + ], + "angle": 0, + "content": "Number of unlabeled samples in a mini-batch. Existing SemiSDA methods [6,81] typically set the ratio \\(\\mu\\) between labeled and unlabeled samples in a" + }, + { + "type": "page_footnote", + "bbox": [ + 0.23, + 0.831, + 0.787, + 0.86 + ], + "angle": 0, + "content": "If not specified, we use ResNet-50 and report the average accuracy (\\%) of seven domain shift scenarios in Table 1 for ablation studies." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "NBF & RLD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.145, + 0.663, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.27, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Fig. 5: NBF leads to higher performance than PBF. We compare different user-feedback configurations when the total number of feedback is 378 (top) and 630 (bottom). Positive and negative feedback refers to feedback from correct and incorrect model predictions, respectively. We run three random seed experiments and describe the average performance and standard deviation in the parenthesis." + }, + { + "type": "table", + "bbox": [ + 0.345, + 0.343, + 0.658, + 0.406 + ], + "angle": 0, + "content": "
methodfeed.negatively biased feedback (NBF)
#xulb, #xLD, #xlb112, 0, 16112, 48, 1664, 48, 16
total batch size128176128
FreeMatch [81]36872.074.274.8 (+0.6)
AdaMatch [6]64.571.372.0 (+0.7)
FreeMatch [81]63074.475.576.1 (+0.6)
AdaMatch [6]67.773.474.3 (+0.9)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.406, + 0.787, + 0.448 + ], + "angle": 0, + "content": "Table 7: In the mini-batch, diminishing the number of unlabeled samples and adding our defending samples achieves better performance with our approach. We ablate them by changing the ratio \\(\\mu\\) in Section 4.1, while keeping the size of labeled samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.789, + 0.592 + ], + "angle": 0, + "content": "mini-batch to 1:7. However, we observe that adhering to this ratio is not optimal for our approach, as shown in Table 7. Our method shows better performance when the ratio is varied to 1:4, i.e., decreasing unlabeled sample sizes. This finding contradicts observations in several TTA works [28, 48, 67], where adaptation performance tends to increase with larger batch sizes. We speculate that it is beneficial to prioritize more reliable information, which refers to labeled data and our defending samples selected from the filtering-applied bank, during the adapting process. This result may be aligned with previous works for curriculum learning [40, 100] and adaptive thresholding [96]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.789, + 0.716 + ], + "angle": 0, + "content": "Number of labeled data. We measure the impact of feedback size (number of labeled samples) in Figure 6. The results show that the inferior performance on NBF persists even with an increased amount of feedback (gray \\(\\rightarrow\\) black line); however, our approach mitigates it and improves performance (black \\(\\rightarrow\\) red line). We make an interesting observation that the performance gap between black and red lines becomes larger as the number of available feedback decreases. Since obtaining large feedback may be challenging in real-world applications, our method is expected to be more helpful in this practical case." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Data selecting strategy. We explore various strategies for selecting defending samples to balance the mini-batch, as shown in Table 8(top). The strategies include: in the \\(x_{LD}\\) candidate bank, (i) random selection regardless of the class of the labeled data, (ii) random selection in the same class as the labeled data (i.e., class-aware), (iii) selecting samples close to the cluster center obtained by k means clustering [21] and (iv) selecting samples with embedded features distant from the labeled data where cosine distance is used. While our approach consistently outperforms the baseline regardless of the chosen strategy, we empirically" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.129 + ], + "angle": 0, + "content": "J. Song et al." + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.144, + 0.672, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.228, + 0.788, + 0.298 + ], + "angle": 0, + "content": "Fig. 6: More reliable adaptation with NBF. In addition to Table 3, we conduct experiments with different amounts (1,3,5,10, and 15 labeled data per class) of feedback using FreeMatch [81]. The number of available feedback is likely to be small in practice. In this case, our method achieves large performance improvement, e.g., our method increases the baseline performance by \\(+4.9\\) when one feedback per class is available." + }, + { + "type": "table", + "bbox": [ + 0.289, + 0.3, + 0.714, + 0.369 + ], + "angle": 0, + "content": "
selection strategy\nclass-awarerandom\nxrandom\n✓kmeans\n✓cosine\n✓baseline\n-
FreeMatch [81]Res.74.174.874.674.072.0
FreeMatch [81]ViT.75.075.775.675.173.9
filtering rate0.20.40.60.8baseline only
FreeMatch [81]Res.74.574.874.373.772.0
FreeMatch [81]ViT.75.575.775.975.573.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.371, + 0.786, + 0.398 + ], + "angle": 0, + "content": "Table 8: We ablate a component of our approach with 378 feedback: \\( x_{LD} \\) selection strategy and filtering rate \\( p \\) for bank generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.786, + 0.437 + ], + "angle": 0, + "content": "find that strategy (ii) achieves the best performance. Therefore, we adopt this strategy for our proposed method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.437, + 0.787, + 0.469 + ], + "angle": 0, + "content": "Further studies, such as extension to a TTA scenario, combining with SFDA methods and different feedback configurations, are presented in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.479, + 0.492, + 0.495 + ], + "angle": 0, + "content": "6 Conclusion & Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.701 + ], + "angle": 0, + "content": "User feedback can play an integral part in adapting the practical ML product to the target environment. However, we have shown that naive adaptation using existing SemiSDA methods led to undesirable adaptation results. We explained this through the lens of Negatively-Biased Feedback (NBF). In this paper, we uncovered the unexpected results of NBF and presented a scalable solution, Retrieval Latent Defending. This method prevents the mini-batch from becoming overly dependent on labeled samples that may have a biased distribution within the overall target distribution. Under the diverse DA benchmarks, from the simulation study to the medical imaging task, we demonstrated the practical problem caused by NBF and the effectiveness of our approach by combining it with multiple SemiSDA baselines. We hope our efforts will inspire future DA works on leveraging user feedback to improve an ML model in the deployment environment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Broader impact. The proposed setup assumes that an ML product obtains feedback as a form of annotations (i.e., labeled data). In some cases, users can provide feedback in different forms, like thumbs up & down and rating of model prediction, or noise feedback whose information is different from the ground truth. Further research considering these points will pave the way for developing safer and more reliable adapting strategies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Acknowledgment. We sincerely appreciate the abundant support provided by Lunit Inc., and we would like to thank Donggeun Yoo, Seonwook Park, and Sérgio Pereira for their valuable feedback." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "NBF & RLD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.179, + 0.787, + 0.234 + ], + "angle": 0, + "content": "1. Arazo, E., Ortego, D., Albert, P., O'Connor, N.E., McGuinness, K.: Pseudolabeling and confirmation bias in deep semi-supervised learning. In: International Joint Conference on Neural Networks (IJCNN) (2020) 6, 9, 11, 23, 24, 25, 26, 29, 30" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.236, + 0.787, + 0.292 + ], + "angle": 0, + "content": "2. Basak, H., Yin, Z.: Semi-supervised domain adaptive medical image segmentation through consistency regularized disentangled contrastive learning. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.293, + 0.787, + 0.32 + ], + "angle": 0, + "content": "3. Baumeister, R.F., Bratslavsky, E., Finkenauer, C., Vohs, K.D.: Bad is stronger than good. Review of general psychology (2001) 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.321, + 0.787, + 0.362 + ], + "angle": 0, + "content": "4. Berthelot, D., Carlini, N., Cubuk, E.D., Kurakin, A., Sohn, K., Zhang, H., Raffel, C.: Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. In: ICLR (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.363, + 0.787, + 0.403 + ], + "angle": 0, + "content": "5. Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.A.: Mixmatch: A holistic approach to semi-supervised learning. In: NeurIPS (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.405, + 0.787, + 0.446 + ], + "angle": 0, + "content": "6. Berthelot, D., Roelofs, R., Sohn, K., Carlini, N., Kurakin, A.: Adamatch: A unified approach to semi-supervised learning and domain adaptation. In: ICLR (2022) 1, 4, 7, 8, 9, 10, 11, 12, 13, 22, 28, 30, 32, 33, 34" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.448, + 0.787, + 0.474 + ], + "angle": 0, + "content": "7. Bradley, A.P.: The use of the area under the roc curve in the evaluation of machine learning algorithms. Pattern recognition (1997) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.476, + 0.787, + 0.503 + ], + "angle": 0, + "content": "8. Chen, D., Wang, D., Darrell, T., Ebrahimi, S.: Contrastive test-time adaptation. In: CVPR (2022) 3, 9, 10, 24, 25, 27, 28, 30, 35" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.504, + 0.787, + 0.545 + ], + "angle": 0, + "content": "9. Chen, H., Tao, R., Fan, Y., Wang, Y., Wang, J., Schiele, B., Xie, X., Raj, B., Savvides, M.: Softmatch: Addressing the quantity-quality trade-off in semi-supervised learning. In: ICLR (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.546, + 0.787, + 0.585 + ], + "angle": 0, + "content": "10. Chen, W., Lin, L., Yang, S., Xie, D., Pu, S., Zhuang, Y.: Self-supervised noisy label learning for source-free unsupervised domain adaptation. In: IROS (2022) 29" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.588, + 0.787, + 0.616 + ], + "angle": 0, + "content": "11. Chen, X., Zhao, Z., Zhang, Y., Duan, M., Qi, D., Zhao, H.: Focalclick: Towards practical interactive image segmentation. In: CVPR (2022) 25" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.617, + 0.787, + 0.644 + ], + "angle": 0, + "content": "12. Chen, Y., Tan, X., Zhao, B., Chen, Z., Song, R., Liang, J., Lu, X.: Boosting semi-supervised learning by exploiting all unlabeled data. In: CVPR (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.645, + 0.787, + 0.7 + ], + "angle": 0, + "content": "13. Cohen, J.P., Viviano, J.D., Bertin, P., Morrison, P., Torabian, P., Guerrera, M., Lungren, M.P., Chaudhari, A., Brooks, R., Hashir, M., et al.: Torchxrayvision: A library of chest x-ray datasets and models. In: International Conference on Medical Imaging with Deep Learning (2022) 10, 11, 28" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.701, + 0.787, + 0.742 + ], + "angle": 0, + "content": "14. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR (2016) 12, 29" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.743, + 0.787, + 0.797 + ], + "angle": 0, + "content": "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) 1, 10, 32, 33, 35" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "16. Fini, E., Astolfi, P., Alahari, K., Alameda-Pineda, X., Mairal, J., Nabi, M., Ricci, E.: Semi-supervised learning made simple with self-supervised clustering. In: CVPR (2023) 4" + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.179, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "J. Song et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.148, + 0.784, + 0.175 + ], + "angle": 0, + "content": "17. Ganin, Y., Lempitsky, V.: Unsupervised domain adaptation by backpropagation. In: ICML (2015) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.217 + ], + "angle": 0, + "content": "18. Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., Marchand, M., Lempitsky, V.: Domain-adversarial training of neural networks. The journal of machine learning research (2016) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.218, + 0.785, + 0.258 + ], + "angle": 0, + "content": "19. Gong, T., Jeong, J., Kim, T., Kim, Y., Shin, J., Lee, S.J.: Robust continual test-time adaptation: Instance-aware bn and prediction-balanced memory. In: NeurIPS (2023) 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.259, + 0.785, + 0.313 + ], + "angle": 0, + "content": "20. Harada, S., Bise, R., Araki, K., Yoshizawa, A., Terada, K., Kurata, M., Nakajima, N., Abe, H., Ushiku, T., Uchida, S.: Cluster-guided semi-supervised domain adaptation for imbalanced medical image classification. arXiv preprint arXiv:2303.01283 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "21. Hartigan, J.A., Wong, M.A.: Algorithm as 136: A k-means clustering algorithm. Journal of the royal statistical society. series c (applied statistics) (1979) 13, 27" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "22. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR (2020) 27" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) 1, 9, 10, 32, 33, 35" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "24. Higuchi, Y., Moritz, N., Roux, J.L., Hori, T.: Momentum pseudo-labeling for semi-supervised speech recognition. In: Interspeech (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "25. Huang, D., Li, J., Chen, W., Huang, J., Chai, Z., Li, G.: Divide and adapt: Active domain adaptation via customized learning. In: CVPR (2023) 4, 21, 22, 29" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "26. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: CVPR (2017) 10, 11, 28" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.481, + 0.785, + 0.535 + ], + "angle": 0, + "content": "27. Johnson, A.E., Pollard, T.J., Greenbaum, N.R., Lungren, M.P., Deng, C.y., Peng, Y., Lu, Z., Mark, R.G., Berkowitz, S.J., Horng, S.: Mimic-cxr-jpg, a large publicly available database of labeled chest radiographs. arXiv preprint arXiv:1901.07042 (2019) 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.536, + 0.785, + 0.563 + ], + "angle": 0, + "content": "28. Khurana, A., Paul, S., Rai, P., Biswas, S., Aggarwal, G.: Sita: Single image test-time adaptation. arXiv preprint arXiv:2112.02355 (2021) 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.564, + 0.785, + 0.604 + ], + "angle": 0, + "content": "29. Kim, J., Hur, Y., Park, S., Yang, E., Hwang, S.J., Shin, J.: Distribution aligning refinery of pseudo-label for imbalanced semi-supervised learning. In: NeurIPS (2020) 22, 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.605, + 0.785, + 0.644 + ], + "angle": 0, + "content": "30. Knox, W.B., Stone, P.: Tamer: Training an agent manually via evaluative reinforcement. In: IEEE international conference on development and learning (2008) 25" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.646, + 0.785, + 0.673 + ], + "angle": 0, + "content": "31. Lee, H., Shin, S., Kim, H.: Abc: Auxiliary balanced classifier for class-imbalanced semi-supervised learning. In: NeurIPS (2021) 10, 22" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "32. Lenga, M., Schulz, H., Saalbach, A.: Continual learning for domain adaptation in chest x-ray classification. In: Medical Imaging with Deep Learning (2020) 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.785, + 0.729 + ], + "angle": 0, + "content": "33. Li, J., Li, G., Shi, Y., Yu, Y.: Cross-domain adaptive clustering for semi-supervised domain adaptation. In: CVPR (2021) 4, 8, 10, 32, 33, 34" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.785, + 0.77 + ], + "angle": 0, + "content": "34. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: ICML (2020) 1, 3, 4, 10, 26, 27, 35" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.811 + ], + "angle": 0, + "content": "35. Litrico, M., Del Bue, A., Morerio, P.: Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation. In: CVPR (2023) 3, 26, 27, 29, 30, 35" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.812, + 0.785, + 0.84 + ], + "angle": 0, + "content": "36. Liu, X., Xing, F., Shusharina, N., Lim, R., Jay Kuo, C.C., El Fakhri, G., Woo, J.: Act: Semi-supervised domain-adaptive medical image segmentation with asym" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "NBF & RLD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.251, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "metric co-training. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.786, + 0.218 + ], + "angle": 0, + "content": "37. Liu, X., Yoo, C., Xing, F., Oh, H., El Fakhri, G., Kang, J.W., Woo, J., et al.: Deep unsupervised domain adaptation: A review of recent advances and perspectives. APSIPA Transactions on Signal and Information Processing (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.786, + 0.247 + ], + "angle": 0, + "content": "38. Liu, Y., Zhang, W., Wang, J.: Source-free domain adaptation for semantic segmentation. In: CVPR (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.248, + 0.786, + 0.288 + ], + "angle": 0, + "content": "39. Liu, Y., Kothari, P., van Delft, B., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? In: NeurIPS (2021) 24, 30" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.786, + 0.318 + ], + "angle": 0, + "content": "40. Liu, Z., Miao, Z., Pan, X., Zhan, X., Lin, D., Yu, S.X., Gong, B.: Open compound domain adaptation. In: CVPR (2020) 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.786, + 0.346 + ], + "angle": 0, + "content": "41. Long, M., Zhu, H., Wang, J., Jordan, M.I.: Deep transfer learning with joint adaptation networks. In: ICML (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.346, + 0.786, + 0.389 + ], + "angle": 0, + "content": "42. MacGlashan, J., Ho, M.K., Loftin, R., Peng, B., Wang, G., Roberts, D.L., Taylor, M.E., Littman, M.L.: Interactive learning from policy-dependent human feedback. In: International conference on machine learning (2017) 25" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.389, + 0.786, + 0.445 + ], + "angle": 0, + "content": "43. Madani, A., Moradi, M., Karargyris, A., Syeda-Mahmood, T.: Semi-supervised learning with generative adversarial networks for chest x-ray classification with ability of data domain adaptation. In: International symposium on biomedical imaging (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.445, + 0.786, + 0.487 + ], + "angle": 0, + "content": "44. Mahapatra, D., Korevaar, S., Bozortabar, B., Tennakoon, R.: Unsupervised domain adaptation using feature disentanglement and gcns for medical image classification. In: ECCV (2022) 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.487, + 0.786, + 0.515 + ], + "angle": 0, + "content": "45. Mei, K., Zhu, C., Zou, J., Zhang, S.: Instance adaptive self-training for unsupervised domain adaptation. In: ECCV (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.515, + 0.786, + 0.543 + ], + "angle": 0, + "content": "46. Mei, K., Zhu, C., Zou, J., Zhang, S.: Instance adaptive self-training for unsupervised domain adaptation. In: ECCV (2020) 12, 29" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.544, + 0.786, + 0.572 + ], + "angle": 0, + "content": "47. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: ICML (2022) 24, 30" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.572, + 0.786, + 0.601 + ], + "angle": 0, + "content": "48. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: ICLR (2023) 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.601, + 0.786, + 0.629 + ], + "angle": 0, + "content": "49. Oh, Y., Kim, D.J., Kweon, I.S.: Daso: Distribution-aware semantics-oriented pseudo-label for imbalanced semi-supervised learning. In: CVPR (2022) 10, 22" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.629, + 0.786, + 0.671 + ], + "angle": 0, + "content": "50. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. In: NeurIPS (2022) 25" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.671, + 0.786, + 0.699 + ], + "angle": 0, + "content": "51. Pan, Y., Yao, T., Li, Y., Wang, Y., Ngo, C.W., Mei, T.: Transferrable prototypical networks for unsupervised domain adaptation. In: CVPR (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.7, + 0.786, + 0.756 + ], + "angle": 0, + "content": "52. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, e.a.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019), https://github.com/pytorch/vision/blob/main/torchvision/models/resnet.py 28" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.786, + 0.812 + ], + "angle": 0, + "content": "53. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournaepau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine learning in Python. Journal of Machine Learning Research (2011) 5, 25, 26" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.812, + 0.786, + 0.84 + ], + "angle": 0, + "content": "54. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: ICCV (2019) 9" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "J. Song et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.147, + 0.784, + 0.188 + ], + "angle": 0, + "content": "55. Prabhu, V., Chandrasekaran, A., Saenko, K., Hoffman, J.: Active domain adaptation via clustering uncertainty-weighted embeddings. In: ICCV (2021) 4, 21, 22, 29" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "56. Richter, S.R., Vineet, V., Roth, S., Koltun, V.: Playing for data: Ground truth from computer games. In: ECCV (2016) 12, 29" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.217, + 0.785, + 0.243 + ], + "angle": 0, + "content": "57. Rozin, P., Royzman, E.B.: Negativity bias, negativity dominance, and contagion. Personality and social psychology review (2001) 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.243, + 0.785, + 0.27 + ], + "angle": 0, + "content": "58. Saito, K., Kim, D., Sclaroff, S., Darrell, T., Saenko, K.: Semi-supervised domain adaptation via minimax entropy. In: ICCV (2019) 1, 3, 4, 8, 9, 10, 28, 32, 33, 34" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.271, + 0.785, + 0.297 + ], + "angle": 0, + "content": "59. Saito, K., Watanabe, K., Ushiku, Y., Harada, T.: Maximum classifier discrepancy for unsupervised domain adaptation. In: CVPR (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.297, + 0.785, + 0.337 + ], + "angle": 0, + "content": "60. Sajjadi, M., Javanmardi, M., Tasdizen, T.: Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In: NeurIPS (2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.785, + 0.379 + ], + "angle": 0, + "content": "61. Schulman, J., Zoph, B., Kim, C., Hilton, J., Menick, J., Weng, J., Uribe, J.F.C., Fedus, L., Metz, L., Pokorny, M., et al.: Chatgpt: Optimizing language models for dialogue. OpenAI blog (2022) 25" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.785, + 0.405 + ], + "angle": 0, + "content": "62. Shen, Y., Yun, H., Lipton, Z.C., Kronrod, Y., Anandkumar, A.: Deep active learning for named entity recognition. In: ICLR (2017) 12, 22" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.785, + 0.445 + ], + "angle": 0, + "content": "63. Shin, I., Kim, D.J., Cho, J.W., Woo, S., Park, K., Kweon, I.S.: Labor: Labeling only if required for domain adaptive semantic segmentation. In: ICCV (2021) 12, 29" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.785, + 0.473 + ], + "angle": 0, + "content": "64. Sofiuk, K., Petrov, I., Barinova, O., Konushin, A.: f-brs: Rethinking backpropagating refinement for interactive segmentation. In: CVPR (2020) 25" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "65. Sofiuk, K., Petrov, I.A., Konushin, A.: Reviving iterative training with mask guidance for interactive segmentation. In: IEEE International Conference on Image Processing (ICIP) (2022) 25" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.514, + 0.785, + 0.555 + ], + "angle": 0, + "content": "66. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In: NeurIPS (2020) 1, 4, 8, 10, 28, 32, 33" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.555, + 0.785, + 0.582 + ], + "angle": 0, + "content": "67. Song, J., Lee, J., Kweon, I.S., Choi, S.: Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. In: CVPR (2023) 1, 13, 24, 30" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.785, + 0.623 + ], + "angle": 0, + "content": "68. Song, J., Park, K., Shin, I., Woo, S., Zhang, C., Kweon, I.S.: Test-time adaptation in the dynamic world with compound domain knowledge management. IEEE Robotics and Automation Letters (2023) 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.785, + 0.663 + ], + "angle": 0, + "content": "69. Stiennon, N., Ouyang, L., Wu, J., Ziegler, D., Lowe, R., Voss, C., Radford, A., Amodei, D., Christiano, P.F.: Learning to summarize with human feedback. In: NeurIPS (2020) 25" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.785, + 0.691 + ], + "angle": 0, + "content": "70. Sun, B., Saenko, K.: Deep coral: Correlation alignment for deep domain adaptation. In: ECCV Workshops (2016) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.785, + 0.732 + ], + "angle": 0, + "content": "71. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 25" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.785, + 0.771 + ], + "angle": 0, + "content": "72. Tsai, Y.H., Hung, W.C., Schulter, S., Sohn, K., Yang, M.H., Chandraker, M.: Learning to adapt structured output space for semantic segmentation. In: CVPR (2018) 1, 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "73. Tsai, Y.H., Hung, W.C., Schulter, S., Sohn, K., Yang, M.H., Chandraker, M.: Learning to adapt structured output space for semantic segmentation. In: CVPR (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.812, + 0.785, + 0.84 + ], + "angle": 0, + "content": "74. Van Engelen, J.E., Hoos, H.H.: A survey on semi-supervised learning. Machine learning (2020) 4" + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "NBF & RLD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "75. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: CVPR (2017) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.215 + ], + "angle": 0, + "content": "76. Vu, T.H., Jain, H., Bucher, M., Cord, M., Pérez, P.: Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In: CVPR (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.217, + 0.785, + 0.244 + ], + "angle": 0, + "content": "77. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: ICLR (2021) 3, 24, 30" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.246, + 0.785, + 0.272 + ], + "angle": 0, + "content": "78. Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing (2018) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.785, + 0.299 + ], + "angle": 0, + "content": "79. Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: CVPR (2022) 30" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.301, + 0.785, + 0.34 + ], + "angle": 0, + "content": "80. Wang, Y., Chen, H., Fan, Y., Sun, W., Tao, R., Hou, W., Wang, R., Yang, L., Zhou, Z., Guo, L.Z., et al.: Usb: A unified semi-supervised learning benchmark for classification. In: NeurIPS (2022) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.395 + ], + "angle": 0, + "content": "81. Wang, Y., Chen, H., Heng, Q., Hou, W., Fan, Y., Wu, Z., Wang, J., Savvides, M., Shinozaki, T., Raj, B., et al.: Freematch: Self-adaptive thresholding for semi-supervised learning. In: ICLR (2023) 4, 7, 8, 10, 12, 13, 14, 24, 25, 27, 28, 30, 32, 33, 34" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "82. Warnell, G., Waytowich, N., Lawhern, V., Stone, P.: Deep tamer: Interactive agent shaping in high-dimensional state spaces. In: AAAI (2018) 25" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.465 + ], + "angle": 0, + "content": "83. Wei, C., Sohn, K., Mellina, C., Yuille, A., Yang, F.: Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In: CVPR (2021) 10, 22, 23, 24, 29" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "84. Wirth, C., Akhour, R., Neumann, G., Fünkranz, J., et al.: A survey of preference-based reinforcement learning methods. Journal of Machine Learning Research (2017) 25" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.549 + ], + "angle": 0, + "content": "85. Xie, B., Yuan, L., Li, S., Liu, C.H., Cheng, X.: Towards fewer annotations: Active learning via region impurity and prediction uncertainty for domain adaptive semantic segmentation. In: CVPR (2022) 12, 29" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.55, + 0.785, + 0.589 + ], + "angle": 0, + "content": "86. Xie, M., Li, Y., Wang, Y., Luo, Z., Gan, Z., Sun, Z., Chi, M., Wang, C., Wang, P.: Learning distinctive margin toward active domain adaptation. In: CVPR (2022) 4, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.591, + 0.785, + 0.619 + ], + "angle": 0, + "content": "87. Xie, Q., Dai, Z., Hovy, E., Luong, T., Le, Q.: Unsupervised data augmentation for consistency training. In: NeurIPS (2020) 4, 8, 10, 32, 33, 34" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "88. Xu, Y., Shang, L., Ye, J., Qian, Q., Li, Y.F., Sun, B., Li, H., Jin, R.: Dash: Semi-supervised learning with dynamic thresholding. In: ICML (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.673 + ], + "angle": 0, + "content": "89. Yan, Z., Wu, Y., Li, G., Qin, Y., Han, X., Cui, S.: Multi-level consistency learning for semi-supervised domain adaptation. In: IJCAI (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "90. Yang, L., Wang, Y., Gao, M., Shrivastava, A., Weinberger, K.Q., Chao, W.L., Lim, S.N.: Deep co-training with task decomposition for semi-supervised domain adaptation. In: ICCV (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.743 + ], + "angle": 0, + "content": "91. Yang, S., Jui, S., van de Weijer, J., et al.: Attracting and dispersing: A simple approach for source-free domain adaptation. In: NeurIPS (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.771 + ], + "angle": 0, + "content": "92. Yang, S., Wang, Y., Van De Weijer, J., Herranz, L., Jui, S.: Generalized source-free domain adaptation. In: ICCV (2021) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.811 + ], + "angle": 0, + "content": "93. Yang, S., van de Weijer, J., Herranz, L., Jui, S., et al.: Exploiting the intrinsic neighborhood structure for source-free domain adaptation. In: NeurIPS (2021) 3, 27, 28, 35" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "94. Yu, Y.C., Lin, H.T.: Semi-supervised domain adaptation with source label adaptation. In: CVPR (2023) 4, 9, 10" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.129 + ], + "angle": 0, + "content": "J. Song et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.147, + 0.787, + 0.175 + ], + "angle": 0, + "content": "95. Yu, Z., Li, J., Du, Z., Zhu, L., Shen, H.T.: A comprehensive survey on source-free domain adaptation. arXiv preprint arXiv:2302.11803 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.786, + 0.217 + ], + "angle": 0, + "content": "96. Zhang, B., Wang, Y., Hou, W., Wu, H., Wang, J., Okumura, M., Shinozaki, T.: Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In: NeurIPS (2021) 7, 8, 10, 13, 32, 33, 34" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "97. Zhang, C., Miech, A., Shen, J., Alayrac, J.B., Luc, P.: Making the most of what you have: Adapting pre-trained visual language models in the low-data regime. arXiv preprint arXiv:2305.02297 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.786, + 0.286 + ], + "angle": 0, + "content": "98. Zhang, W., Ouyang, W., Li, W., Xu, D.: Collaborative and adversarial network for unsupervised domain adaptation. In: CVPR (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.786, + 0.327 + ], + "angle": 0, + "content": "99. Zhang, Y., Zhang, H., Deng, B., Li, S., Jia, K., Zhang, L.: Semi-supervised models are strong unsupervised domain adaptation learners. arXiv preprint arXiv:2106.00417 (2021) 1, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "100. Zhang, Y., David, P., Gong, B.: Curriculum domain adaptation for semantic segmentation of urban scenes. In: ICCV (2017) 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.356, + 0.785, + 0.384 + ], + "angle": 0, + "content": "101. Zhang, Y., Wang, Z., He, W.: Class relationship embedded learning for source-free unsupervised domain adaptation. In: CVPR (2023) 3" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.147, + 0.787, + 0.384 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Is user feedback always informative_ Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data/0589fa53-73c5-413a-aac9-f37670a1de05_origin.pdf b/2024/Is user feedback always informative_ Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data/0589fa53-73c5-413a-aac9-f37670a1de05_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..635bdd16f61bcb0026a0f8fecd97722366259394 --- /dev/null +++ b/2024/Is user feedback always informative_ Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data/0589fa53-73c5-413a-aac9-f37670a1de05_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:86e0c99b240b0c5b53094bdd227d185bf1863f43f82db06c339516ae77aa2f35 +size 4598354 diff --git a/2024/Is user feedback always informative_ Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data/full.md b/2024/Is user feedback always informative_ Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8358892752b3d0d95846fa3192d49ec843be6ce3 --- /dev/null +++ b/2024/Is user feedback always informative_ Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data/full.md @@ -0,0 +1,329 @@ +# Is user feedback always informative? Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data + +Junha Song $^{1,2}$ , Tae Soo Kim $^{2}$ , Junha Kim $^{2}$ , Gunhee Nam $^{2}$ , Thijs Kooi $^{2}$ , and Jaegul Choo $^{1}$ + +1 KAIST {sb020518, jchoo}@kaist.ac.kr 2 Lunit Inc. {junha.kim, taesoo.kim, ghnam, tkooi}@lunit.io + +Abstract. This paper aims to adapt the source model to the target environment, leveraging small user feedback (i.e., labeled target data) readily available in real-world applications. We find that existing semi-supervised domain adaptation (SemiSDA) methods often suffer from poorly improved adaptation performance when directly utilizing such feedback data, as shown in Figure 1. We analyze this phenomenon via a novel concept called Negatively Biased Feedback (NBF), which stems from the observation that user feedback is more likely for data points where the model produces incorrect predictions. To leverage this feedback while avoiding the issue, we propose a scalable adapting approach, Retrieval Latent Defending. This approach helps existing SemiSDA methods to adapt the model with a balanced supervised signal by utilizing latent defending samples throughout the adaptation process. We demonstrate the problem caused by NBF and the efficacy of our approach across various benchmarks, including image classification, semantic segmentation, and a real-world medical imaging application. Our extensive experiments reveal that integrating our approach with multiple state-of-the-art SemiSDA methods leads to significant performance improvements. + +Keywords: Rethinking user-provided feedback $\cdot$ Semi-supervised & Source-free domain adaptation $\cdot$ Medical image diagnosis + +# 1 Introduction + +While deep neural networks have demonstrated remarkable performance in the development domain (i.e., source domain) [15,23], they often suffer from performance degradation in the deployed domain (i.e., target domain) due to domain shift [17, 72, 78]. To mitigate this issue, domain adaptation (DA) techniques have been introduced [34, 58, 70]. The most common DA tasks include semi-supervised domain adaptation (SemiSDA) and source-free domain adaptation (SFDA). SemiSDA aims to adapt the model given a small amount of labeled target data along with massive unlabeled target data [6, 58, 66, 99]. SFDA conducts adaptation with only target data without accessing source data considering data privacy or memory constraints in edge devices [34, 67, 92]. + +![](images/a2665f627bd6c298bd33a6f9b2a5cd19d930495806a579cc983bb24a3ecd883a.jpg) +Fig. 1: (a) User feedback. Users can provide feedback while interacting with an ML product, where feedback is likely to be biased towards misclassified samples, which we define as Negatively Biased Feedback (NBF). (b) Adaptation results. We adapt the source model with small user feedback and large unlabeled target data using previous semi-supervised domain adaptation (SemiSDA) algorithms. Compared to random feedback, which is the classical SemiSDA setup where labeled data is a random subset of target data, model adaptation with NBF leads to subpar performance. This paper analyzes this problem and introduces a scalable solution. + +![](images/1fa349cee07a879821bb1b0df1354fb785a490a9f81e0ae77032670bdc690d4a.jpg) + +Despite such advances in DA, adapting the model with user feedback still remains an open area for further research, even though practical machine learning (ML) products often allow users to provide feedback in order to further improve the model in the target environment. For example, facial recognition or medical image diagnosis applications enable users to give feedback correcting wrong model predictions, as depicted in Figure 1 (a). Since feedback can be modeled in this case as a small amount of labeled target data, it is anticipated that previous SemiSDA methods assuming the same setup would yield promising results. However, we observe that they show inferior adaptation performance on multiple DA benchmarks when using such user feedback in practice, as shown in the dark-gray bar in Figure 1 (b). + +We introduce a novel concept called Negatively Biased Feedback (NBF) to explain this phenomenon. NBF is based on the observation that user feedback is more likely to be derived from incorrect model predictions. For example, a radiologist might log a misdiagnosed chest X-ray by the model, as its accuracy directly impacts the patient's survival. Interestingly, our observation aligns with findings from cognitive psychology literature [3,57] that proves that humans are more likely to react and provide feedback to negative events (i.e., wrong model predictions). Since such an NBF scenario is feasible, we analyze its unexpected impact on SemiSDA observed above. We identify that a biased distribution of NBF within the overall data distribution leads to sub-optimal adaptation results, particularly compared to Random Feedback (RF). RF represents the classical SemiSDA setup, where labeled data is randomly selected from the target data. + +To address the problem caused by NBF, we present a scalable approach named Retrieval Latent Defending, which can be seamlessly integrated with existing SemiSDA methods. Our approach allows them to adapt the model without a strong dependence on the biasedly distributed labeled data. Specifically, we balance the supervised adapting signal by appending latent defending samples to the mini-batch and help to keep the model's balanced class discriminability throughout adapting iterations. We evaluate the unexpected influence of NBF using various benchmarks, including image classification, semantic segmentation, + +![](images/f43b6f978ec654ada808edc9a32d3e1c9a0d4b92b99841e0deeeb68abbd1f86f.jpg) +Fig. 2: Adaptation with user feedback can be effective in alleviating performance degradation caused by domain shift. However, there are some challenges: (i) user feedback may be a biased sampling of the true target distribution due to the nature of feedback, (ii) the amount of the ground truths (GT) labels obtained through feedback is small, and (iii) only unlabeled target data is typically available, not source data. + +and medical image diagnosis. Building upon these evaluations, we demonstrate that our approach not only complements, but significantly enhances the performance of multiple SemiSDA methods. + +The contributions of the paper are as follows: + +We introduce the novel concept called Negatively Biased Feedback and uncover that it can lead to sub-optimal adaptation performance of existing SemiSDA methods. +- We analyze this problem and present a scalable solution, Retrieval Latent Defending, that combines with SemiSDA methods and allows them to avoid the unexpected effect of NBF. +- We show that our approach generalizes through diverse DA benchmarks and improves adaptation results of state-of-the-art SemiSDA methods. +We publicly release the code on https://github.com/junha1125/RLD-SemiSDA. + +# 2 Related Work + +Adaptation in the deployment environment. Real-world ML products often encounter performance degradation caused by gaps between the source and target environment [17]. One solution is to adapt the model using unlabeled data observed in the target domain, referred to as unsupervised domain adaptation (UDA) [37, 59, 72]. Works on UDA use both source and target data to improve the target performance by using methods such as domain discrepancy minimization by adversarial training [18, 41, 59, 70, 72, 73, 76], and self-training with pseudo labels [45, 51, 97, 98]. Source-free DA (SFDA) builds on UDA and imposes an additional constraint that the source data can not be accessed during domain adaptation. This has practical implications for addressing data privacy concerns or barriers in data transmission to edge devices [34, 38, 77, 95]. The majority of recent SFDA works rely on strategies like domain clustering [34], nearest neighbors [91-93], and contrastive learning [8, 35, 101]. Nevertheless, SFDA does not consider the availability of small labeled data, which may be available in practical ML systems. Semi-supervised DA (SemiSDA) works mainly demonstrate that permitting small labeled data in the target domain can substantially enhance adaptation performance compared to traditional UDA [58]. + +Their primary strategy is to use domain alignment [20, 33, 58, 94], multi-view consistency [2, 6, 33, 89], and asymmetric co-training [36, 90]. + +Active domain adaptation (ActiveDA) [25, 55, 86] envisions a scenario in which the machine selects specific target samples and instructs annotators to label them. The primary objective of ActiveDA is to strategically identify and select the most informative samples for annotation. These chosen samples (i.e., labeled target data) are subsequently utilized to update the source model using SemiSDA methods [33, 58], and the effectiveness of ActiveDA is assessed by evaluating the target performance of the adapted model. + +Semi-supervised learning (SemiSL) aims to reduce expensive human annotations, and propose methods to train a model from scratch using massive unlabeled data along with limited amounts of labeled data [43,74]. The majority of SemiSL methods depend on consistency regularization [4,5,16,60,66,87], which helps the model to make similar predictions for augmented versions of the same image. Moreover, adaptive thresholding [9,12,24,66,81,88,99] is also popularly utilized to produce reliable pseudo labels from unlabeled data. + +SemiSDA and SemiSL setups mimic small labeled datasets by randomly selecting subsets of the target dataset, whereas ActiveDA involves selections instructed by the machine. In contrast, this paper posits that in real-world applications, labeled data is typically acquired through user intervention. Additionally, users often provide feedback on samples misclassified by the model (i.e., negatively biased feedback), a process detailed in the following section. + +
UDASFDAActiveDASemiSDASemiSLOur setup
Adaptation○○○○×○
Source-free×○××-○
Feedback××machine-instructedrandomly selectedrandomly selecteduser-provided
+ +The table above summarizes the comparison of relevant studies to our setup. In the table, adaptation means fine-tuning the source pre-trained model (as opposed to training from scratch); feedback represents a small number of labeled target samples. Appendix A provides further comparisons with settings like class-imbalanced SemiSDA and test-time adaptation (TTA). + +# 3 Negatively Biased Feedback + +# 3.1 Adaptation with user feedback. + +Our adaptation setup is illustrated in Figure 2. A model is pre-trained on the source data $D_{s}$ . Next, the model is deployed to the target domain, such as a smartphone or a hospital, where we assume the transfer of $D_{s}$ is prohibited due to data privacy regulations or resource constraints (same setup as SFDA [34]). While users utilize ML products on the target domain, the model provides prediction results for data observed in the target domain $D_{t}$ and occasionally obtains user feedback in the form of annotations $y$ . We represent the target data as $D_{t} = X_{t}^{lb} \cup X_{t}^{ulb}$ , where $X_{t}^{lb} = \{(x_{lb}^{n},y_{lb}^{n}):n\in [1..N_{lb}]\}$ , $X_{t}^{ulb} = \{(x_{ulb}^{n}):n\in [1..N_{ulb}]\}$ , $x_{lb}$ and $x_{ulb}$ denote labeled and unlabeled data and $N_{lb}$ and $N_{ulb}$ is their number of data. Lastly, the model can utilize $D_{t}$ and SemiSDA algorithms for adaptation during its inactive phase (e.g., when users do not use the product, + +![](images/df2bd396b0b566c13e4bd4684d90efe02dc750deafbbad1cc6f1a1fae253d0e6.jpg) +Fig.3: Effect of negatively biased feedback. Our novel observations are that $(a)$ user-provided feedback in practice has a biased distribution in each class cluster (the bottom center sub-figure) which is in contrast to random feedback, $(b)$ Existing SemiSDA methods adapt the model by dominating the labeled data points (the right sub-figures) even though they are biasedly positioned, and $(c)$ NBF prevents the model from having a decision boundary for true class clusters and leads to inferior adaptation performance (the bottom right sub-figure). + +like at nighttime) in order to alleviate performance degradation due to domain shift or to personalize the model based on user feedback. + +Rethinking user-provided feedback. Classical SemiSDA works simply assume that a random subset in target data $D_{t}$ is labeled by users when building $X_{t}^{lb}$ . However, as illustrated in Figure 2 (i), we suggest that users are more likely to provide feedback on misclassified samples by the source model, named negatively biased feedback (NBF). This behavior can be understood from two perspectives: (a) users generally expect their feedback to be used as a basis of model improvement, motivating them to provide NBF, and (b) humans tend to react more strongly to negative experiences, such as receiving incorrect predictions, as observed in psychological studies [3,57]. We note that the NBF assumption holds more strongly for the medical application: it is reasonable to imagine that the user (i.e., radiologist) logs the mistakes of the model while diagnosing a chest X-ray exam because the diagnostic accuracy of the model is directly related to the patient's chances of survival. Furthermore, applications beyond the medical domain can also exhibit NBF. For instance, users in self-driving cars can report errors, such as object detection failures or navigation mistakes, to enhance the car's driving capabilities. + +# 3.2 Influence of NBF on SemiSDA + +Simulation study. As shown in Figure 3, we conduct a simulation study to understand the effect of NBF on SemiSDA. We first use the blobs dataset [53] and construct the source and target data so that domain shift exists between them (left sub-figures). We pre-train a source model on the source data and + +compute the accuracy in the target domain, where the performance drop due to domain shift is observed $(98.5\% \rightarrow 76.4\%)$ . Next, we simulate two types of feedback (i.e., labeled data): random feedback and negatively biased feedback following a previous SemiSDA setup and our setup, respectively. Specifically, NBF is randomly selected among misclassified samples by the source model. We find that random feedback (RF) points are evenly distributed, while NBF points are biasedly positioned across each class cluster (refer to blue points in the dashed circle in the center sub-figures). + +To alleviate the performance drop caused by domain shift, we adapt the model using the target data and a semi-supervised method, Pseudo-labeling [1]. This method iteratively optimizes the model by the cross-entropy loss computed by the ground truth of labeled data and pseudo labels of unlabeled data in a minibatch (pseudo labels are predicted by the current adapting model so they can be changed according to an updating decision boundary. Further comprehension can be achieved by referring Appendix B.). The SemiSDA results are shown in the right sub-figures, where we make two interesting observations: (i) the distribution of labeled data can contribute significantly to a decision boundary of the adapted model (red arrows in the figure), and (ii) the adapted model under NBF has poorly improved performance compared with one under RF $(76.4\% \rightarrow 88.1\%$ with NBF, but $76.4\% \rightarrow 91.7\%$ with RF). + +Unexpected influence of NBF. Our intuitive reasoning probably suggests that NBF provides more information than RF by correcting more source model deficiencies, and thus leads to better adaptation performance. However, we empirically show that NBF can result in inferior adaptation performance due to its biased distribution across each class cluster, as illustrated in Figure 3. Surprisingly, we also show that this problem persists, even with other state-of-the-art SemiSDA methods and large datasets for various DA benchmarks, including image classification, semantic segmentation, and medical image diagnosis. Our work highlights the importance of careful design when using user feedback in real-world scenarios and, to the best of our knowledge, is the first study to uncover and analyze this phenomenon. + +# 4 Approach + +# 4.1 Prerequisite: Previous SemiSDA method + +Previous SemiSDA and SemiSL works typically construct a mini-batch with labeled data $\{(x_{lb}^b, y_{lb}^b) : b \in [1..B]\}$ , and unlabeled data whose size is $\mu$ times larger than labeled ones $\{(x_{ulb}^b) : b \in [1..\mu \cdot B]\}$ , where $B$ is the mini-batch size for labeled data. To adapt the model iteratively, they compute the cross-entropy loss $\mathcal{H}(\cdot, \cdot)$ with labeled data and the consistency regularization to multi-view of unlabeled data, which are formulated as the following: + +$$ +\mathcal {L} _ {s u p} = \frac {1}{B} \sum_ {b = 1} ^ {B} \mathcal {H} \left(y _ {l b} ^ {b}, f _ {\theta} \left(x _ {l b} ^ {b}\right)\right), \quad \mathcal {L} _ {u n s u p} = \frac {1}{\mu \cdot B} \sum_ {b = 1} ^ {\mu \cdot B} \mathcal {H} \left(\hat {y} _ {u l b} ^ {b}, f _ {\theta} \left(\Omega \left(x _ {u l b} ^ {b}\right)\right)\right), \tag {1} +$$ + +where $f_{\theta}(\cdot)$ is the output probability from the model, $\hat{y}_{ulb}$ denotes a pseudo label obtained from $f_{\theta}(\omega(x_{ulb}))$ , and $\omega(\cdot)$ and $\Omega(\cdot)$ represent weak and strong image + +![](images/623147e71e483c912e40ebb7b5a40dd915dc5a23b83f3ae1375eb1e375a43f13.jpg) +Fig. 4: Even though labeled data $(x_{lb}, y_{lb})$ is biasedly positioned, the model needs to be adapted with balanced class discriminability (i.e., decision boundary). (i) However, previous SemiSDA methods have overlooked this fact and used the labeled data naively by applying a cross-entropy loss, leading to inadequate adaptation performance. (ii) To alleviate this problem, we propose a scalable adapting approach, retrieval latent defending, which allows the model to adjust the balance of a mini-batch on each iteration by using latent defending samples $x_{LD}$ together with labeled data $x_{lb}$ . + +augmentation, respectively. While sharing the core framework, each SemiSDA method employs distinct adapting strategies, especially to enhance the effectiveness of the use of unlabeled data rather than labeled data [6,81,96]. + +Problem of previous works. Since previous SemiSDA methods have overlooked the unexpected impact of NBF, they often suffer from sub-optimal performance under the NBF assumption (shown in Section 5). To address this problem, we focus on developing a scalable solution that (i) can easily combine with existing DA methods without modifying their core adapting strategies and (ii) can be applied to a wide range of benchmarks, including medical image diagnosis. + +# 4.2 Retrieval Latent Defending + +Based on the observations in Figure 3, we illustrate the unintended effect of NBF when using an existing SemiSDA method in Figure 4 (top center), where NBF is likely to exhibit a biased distribution, leading to undesirable adaptation results. To alleviate this issue, we propose Retrieval Latent Defending as depicted in Figure 4 (bottom). $①$ Prior to each epoch, we generate a candidate bank of data points, denoted as $x_{LD}$ . $② \sim ④$ For each adapting iteration, we balance the mini-batch by retrieving latent defending samples $x_{LD}$ from the bank. $⑤ \sim ⑥$ The model is then adapted using the reconfigured mini-batch and following the baseline SemiSDA approach. We hypothesize that the latent space progressively created by the $x_{LD}$ candidates throughout the adaptation process (bold dashed circle in Figure 4 (top right)) mitigates the issue caused by NBF, thereby allowing the SemiSDA baseline to achieve robust adaptation against NBF. + +Candidate bank generation. The candidate bank serves as a repository of pseudo labels $\hat{Y}_t^{ulb}$ for a subset of the target unlabeled data $X_t^{ulb}$ . Before each epoch, we freeze the model and use it to generate pseudo labels $\hat{Y}_t^{ulb} = \{(\hat{y}_{ulb}^n): n \in [1..N_{ulb}]\}$ , where $\hat{y}_{ulb}^n$ is assigned to $x_{ulb}^n$ as the predicted class with the highest softmax probability: $\hat{y}_{ulb}^n = \operatorname{argmax}_c [f_\theta(x_{ulb}^n)]_c$ . We then retain only samples with the top $p\%$ highest probabilities within each class. This filtering step helps mitigate the inclusion of data with potentially inaccurate pseudo labels, as the model's predictions on $X_t^{ulb}$ might not always be perfect. + +Defending sample selection. We select $k$ latent defending samples $x_{LD}$ from the bank at random for each labeled data $(x_{lb}^{b}, y_{lb}^{b})$ . These selected samples share the same pseudo label as the ground-truth label of their associated counterparts (i.e., $\hat{y}_{LD} = y_{lb}^{b}$ ). By incorporating these defending samples, we balance the data distribution within the current mini-batch. For example, consider $x_{lb}^{1}$ and $x_{lb}^{2}$ in Figure 4 (top right). As these labeled samples are included in the current mini-batch alongside the selected defending samples $x_{LD}^{1}$ and $x_{LD}^{2}$ , we expect to prevent the supervised adapting signal from becoming overly dependent on the labeled samples. We imagine the effect of the defending samples throughout the adaptation process and depict the latent space formed gradually by the $x_{LD}$ candidates as bold dashed circles in Figure 4 (top right). + +Consequently, the overall loss consists of the sum of losses in Eq. (1) and a loss from our proposed method as, + +$$ +\mathcal {L} _ {\text {t o t a l}} = \underbrace {\mathcal {L} _ {\text {s u p}} + \mathcal {L} _ {\text {u n s u p}}} _ {\text {b a s e l i n e}} + \underbrace {\frac {1}{k \cdot B} \sum_ {b = 1} ^ {k \cdot B} \mathcal {H} \left(y _ {L D} ^ {b} , f _ {\theta} \left(x _ {L D} ^ {b}\right)\right)} _ {\text {r e t r i e v a l l a t e n t d e f e n d i n g}}. \tag {2} +$$ + +Importance of our method. Understanding the impact of NBF on adaptation performance is crucial. For example, naively adapting a model for a medical application using radiologist-provided feedback can actually cause performance degradation (shown in Table 5), potentially posing significant risks to patients. We propose a scalable and simple approach to solve the problem caused by NBF, which can not be addressed by existing methods. Given the practicality of the NBF problem and the scalability of our solution, we believe our work holds considerable potential for real-world applications. + +# 5 Experiments + +# 5.1 Experimental Setups + +Our approach is simple enough to seamlessly combine with existing SemiSDA algorithms and also be applied to diverse benchmarks. This section describes our experimental setup for natural image classification tasks and a real-world medical application. Details for semantic segmentation experiments are in Appendix D. + +Baselines. We validate our approach by combining various state-of-the-art algorithms for SemiSDA [58] (e.g., CDAC [33] and AdaMatch [6]) and SemiSL [66,87] (e.g., FlexMatch [96] and FreeMatch [81]). Note that the SemiSL methods have been demonstrated to be strong SemiSDA learners [99], so we can consider them + +
methodfeedbackaverage\( \mathrm{r} \rightarrow \mathrm{c} \)\( \mathrm{r} \rightarrow \mathrm{p} \)\( \mathrm{p} \rightarrow \mathrm{c} \)\( \mathrm{c} \rightarrow \mathrm{s} \)\( \mathrm{s} \rightarrow \mathrm{p} \)\( \mathrm{r} \rightarrow \mathrm{s} \)\( \mathrm{p} \rightarrow \mathrm{r} \)
RerNetAdaMatch [6]RF67.666.668.568.560.369.258.781.5
NBF\( {64.5}\left( {+{3.1}}\right) \)64.366.165.656.965.654.278.9
w/ oursNBF\( \mathbf{{72.0}\left( {+{7.5}}\right) } \)74.572.773.965.570.064.383.2
ViTAdaMatch [6]RF74.775.376.973.868.076.367.185.5
NBF\( {73.7}\left( {+{1.0}}\right) \)74.776.274.765.774.066.884.0
w/ oursNBF\( \mathbf{{75.9}\left( {+{2.2}}\right) } \)76.977.877.868.576.668.385.1
+ +Table 1: Adaptation results on DomainNet-126. We simulate seven domain-shift scenarios (i.e., source $\rightarrow$ target). The model is pre-trained on the source domain and then adapted to a training set of the target domain. The results on the test set of the target domain are reported as the top-1 accuracy (\%). DomainNet-126 [54,58] dataset includes real, painting, sketch, and clip-art domains. In this experiment, we assume that the 378 feedback samples (i.e., 3 labeled data per class) are obtained from users. A state-of-the-art SemiSDA method, AdaMatch [6], is used as a baseline. + +
methodfeedbackaveragea→ca→pa→rc→ac→pc→rp→ap→cp→rr→ar→cr→p
AdaMatch [6]RF70.955.480.475.965.781.574.665.958.778.468.861.584.3
NBF69.3 (+1.6)54.276.675.365.979.375.563.757.475.966.756.884.2
w/oursNBF73.8 (+4.5)62.281.079.768.885.478.667.761.779.569.064.188.2
+ +Table 2: Adaptation results on OfficeHome. OfficeHome [75] dataset includes real, product, art, and clip-art domain. We assume that the 195 feedback samples (i.e., 3 labeled data per class) are obtained. AdaMatch [6] and ResNet-50 [23] are used. + +as SemiSDA methods. For medical experiments, we use Pseudo-labeling [1] as a baseline since it is easily applicable to medical image adaptation. + +Datasets. We utilize natural image datasets containing multiple kinds of domains (e.g., real and painting). The datasets include DomainNet-126 [54,58] with 142k images of 126 classes, and OfficeHome [75] with 15K images of 65 classes. + +To conduct medical experiments, we present a practical medical setting. We adopt the MIMIC-CXR-V2 dataset [27]. It assumes a multi-finding binary classification setup, where multiple radiographic findings, like Pneumonia and Atelectasis, can coexist in a single chest X-ray (CXR) sample. Thus, the model predicts the presence or absence (binary classes) of each individual finding. We simulate domain shift by using Posterior-Anterior (PA)-view data as the source and AP-view data as the target, capturing real-world variations in data acquisition. Typically, patients requiring an AP X-ray are those facing positioning challenges that prevent them from undergoing a PA X-ray. Therefore, this setup can be seen as a scenario where the target environment is the intensive care unit, which hospitalizes critically ill patients. + +Following the recent SemiSDA [94] and SFDA [8] setups, we assume the model is pre-trained in the source domain and deployed in the target domain. Since the datasets above were not initially divided into training and test sets, we performed a random 8:2 split within each domain, designating them respectively for training and testing. The training set is used to adapt the model, while the test set is used to report the top-1 accuracy. + +User feedback. Feedback given by users is modeled as annotations $\{y_{lb}^n : n \in [1..N_{lb}]\}$ on a small subset of the target's training set $D_t^{train}$ , while the remaining of them are used as unlabeled target data. In our experiments, we take into account two types of feedback: random feedback (RF) and negatively biased feedback (NBF). RF is the same setup of classical SemiSDA and SemiSL, where randomly selected samples from $D_t^{train}$ are used as small labeled set $X_t^{lb}$ . For NBF, we randomly select samples that are incorrectly predicted in $D_t^{train}$ by + +
feed, amount method378 (3 labeled data per class)630 (5 labeled data per class)
RFNBFw/oursRFNBFw/ours
Source model56.5
MME [58]69.568.4 (-1.1)70.8 (+2.4)71.270.1 (-1.1)72.5 (+2.4)
CDAC [33]68.364.6 (-3.7)73.2 (+8.6)71.768.1 (-3.6)74.9 (+6.8)
AdaMatch [6]67.664.5 (-3.1)72.0 (+7.5)70.967.7 (-3.2)74.3 (+6.6)
FixMatch [66]67.663.4 (-4.2)73.2 (+9.8)71.566.1 (-5.4)75.1 (+9.0)
UDA [87]69.264.9 (-4.3)73.4 (+8.5)72.968.8 (-4.1)75.3 (+6.5)
FlexMatch [96]73.371.4 (-1.9)74.7 (+3.3)75.373.9 (-1.4)76.0 (+2.1)
FreeMatch [81]73.872.0 (-1.8)74.8 (+2.8)75.674.4 (-1.2)76.1 (+1.7)
Fully supervised83.6
Source model64.5
MME [58]73.272.7 (-0.5)74.1 (+1.4)74.574.0 (-0.5)75.2 (+1.2)
CDAC [33]74.272.8 (-1.4)75.4 (+2.6)75.474.1 (-1.3)76.2 (+2.1)
AdaMatch [6]74.773.7 (-1.0)75.9 (+2.2)75.975.1 (-0.8)76.7 (+1.6)
FixMatch [66]74.673.0 (-1.6)75.6 (+2.6)75.774.3 (-1.4)76.5 (+2.2)
UDA [87]74.873.3 (-1.5)75.8 (+2.5)75.974.5 (-1.4)76.7 (+2.2)
FlexMatch [96]74.973.9 (-1.0)75.8 (+1.9)76.075.1 (-0.9)76.9 (+1.8)
FreeMatch [81]74.973.9 (-1.0)75.7 (+1.8)76.075.1 (-0.9)76.8 (+1.7)
Fully supervised85.4
+ +Table 3: Comparisons on DomainNet-126. We evaluate our method by integrating it with SemiSDA and SemiSL methods. The average accuracy of seven domain-shift scenarios in Table 1 is reported. Source model represents the pre-trained model without adaptation. Fully supervised means the model is adapted with fully labeled target data. + +
feed, amount method195 (3 labeled data per class)325 (5 labeled data per class)
RFNBFw/oursRFNBFw/ours
Source model57.6
MME [58]71.270.2 (-1.0)73.4 (+3.2)73.573.1 (-0.4)75.6 (+2.5)
CDAC [33]71.269.0 (-2.2)74.3 (+5.3)73.572.3 (-1.2)75.7 (+3.4)
AdaMatch [6]70.969.3 (-1.6)73.8 (+4.5)73.472.7 (-0.7)75.5 (+2.8)
FixMatch [66]71.468.6 (-2.8)73.7 (+5.1)73.972.2 (-1.7)75.3 (+3.1)
UDA [87]72.269.5 (-2.7)74.1 (+4.6)74.473.0 (-1.4)76.0 (+3.0)
FlexMatch [96]73.772.1 (-1.6)74.7 (+2.6)75.974.9 (-1.0)76.6 (+1.7)
FreeMatch [81]74.072.7 (-1.3)74.8 (+2.1)75.875.0 (-0.8)76.6 (+1.6)
Fully supervised87.4
+ +Table 4: Comparisons on OfficeHome. The average accuracy of twelve domain-shift scenarios in Table 2 is reported. ResNet-50 is used. + +the source model (i.e., the pre-trained model before adaptation). Note that we focus on the impact of a biased label distribution within the same class, as shown in Figure 3, and thus take the same number of feedback for each class. Further discussion about the imbalance in the number of feedback between classes presented in [31, 49, 83] is provided in Appendix A.2. + +Network architectures. We adopt commonly used networks, ResNet [23] and ViT [15] for natural image tasks and DenseNet [26] for a medical task. We employ ResNet-50 with the last classification layer comprising a weight normalization layer and a bottleneck layer, following previous works [8, 34] and use the ViT-Small (i.e., ViT-S) introduced in [80]. The DenseNet-121 is used, provided in TorchXrayVision [13], like existing medical works [32, 44]. + +Implementation details. We implement our framework by extending the publicly available USB [80] repository. Both pre-training and adaptation are conducted with a mini-batch size of 128 and the SGD optimizer. Diverse baselines for SemiSDA and SemiSL are used to compute the losses in Eq. (1). The hyperparameters for each baseline simply follow USB [80] or public code [33,58]. For all experiments, our approach uses the same hyper-parameters of the appended defending samples $k$ and reliable filtering rate $p$ as 3 and 0.4, respectively. + +# 5.2 Main Results + +Natural image classification. Following recent DA works [8, 94], we conduct experiments on seven and twelve domain shift scenarios provided with the + +![](images/1a0a93f6bc6c7b1f36a1f94d5389a287c4d30798c41abd0fe1ed981fc4076aeb.jpg) +Table 5: Adaptation in a medical application. We use samples with PA-view as the source data and samples with AP-view as the target data in MIMIC-CXR-V2 dataset [27]. NBF-CE represents a scenario when NBF is composed of cases with confident errors. We use DensNet-121 [13, 26] and assume the 20 feedback for the absence and presence per finding. + +DomainNet-126 and OfficeHome datasets, respectively. Table 1 and Table 2 show the results, where AdaMatch [6] is used as the baseline. We observe consistent results with Figure 3 even on large natural datasets: when simply applying the baseline under the NBF assumption, the adapted model shows inferior performance for most domain shifts than applying it under RF, e.g., $64.5 < 67.6$ . Combining our approach with the baseline mitigates this issue and achieves a performance increase, e.g., $64.5 \rightarrow 72.0$ . + +We also use other promising baselines and report the average accuracy of all domain shifts in Table 3 and Table 4 (all results can be found in Appendix F). While both feedback types bring performance improvement from the source model, lower performance is observed with NBF. Our method enables the baselines to not only address this problem but surpass performance under RF. The above results suggest that the biased distribution of labeled samples, which has been overlooked in previous SemiSDA works, is actually problematic, and our retrieval latent defending approach is effective. + +Medical image diagnosis. Table 5 shows the results (bottom) and also depicts the effect of NBF (top center). We report the AUROC [7] for each finding following standard practice for measuring computer-aided-diagnosis model evaluation [32,44]. The baseline SemiSDA method under NBF exhibits inferior performance compared to one under RF, but this issue can be mitigated by combining our approach. + +In addition, we propose an interesting and practical scenario named NBF with more confident errors (NBF-CE). In this scenario, we assume that a radiologist is likely to give feedback when the model makes confidently wrong predictions. Imagine that the model predicts a $1\%$ likelihood of cancer in a CXR image, but the person actually has cancer. Such failure to detect potential patients early on can significantly reduce the patient's chances of survival, so a radiologist may + +
labeling typefeed. amountRFNBFNBF w/oursENT [62]ENT w/ours
IAST [46,63]PA, 40 points55.353.0 (-2.3)56.3 (+3.3)53.556.0 (+2.5)
RIPU [85]PA, 40 points57.654.5 (-3.1)58.0 (+3.5)54.657.7 (+3.1)
+ +Table 6: Adaptation on semantic segmentation. The GTA5 [56] $\rightarrow$ Cityscapes [14] setup is used [72]. The target performance of the source model is $36.6\mathrm{mIoU}$ . + +provide feedback to the model. To simulate NBF-CE, we select samples where the source model most confidently predicts a finding to be absent $(\hat{y} \approx 0)$ although it is clearly visible in the radiograph $(y = 1)$ , and vice versa, i.e., samples of $\hat{y} \approx 1$ but $y = 0$ . Table 5 also shows the results under an NBF-CE scenario, where the model's adaptation performance is further reduced compared with NBF (0.7691 for $\mathrm{NBF} \to 0.7639$ for NBF-CE). By combining our method, we observe performance improvements for both NBF variants, e.g., 0.7639 for $\mathrm{NBF - CE} \to 0.7875$ with ours. We illustrate the hypothesized impact of our method in Table 5. + +Semantic segmentation. We evaluate the influence of NBF and our approach on a semantic segmentation task. We utilize the most common adaptation benchmark of GTA5 [56] to Cityscapes [14]. The baseline DA algorithms are used as IAST [46, 63] and RIPU [85] in a source-free scenario. We regard Pixel-based Annotation (PA) in which we assume 40 pixels per image like LabOR [63]. Table 6 shows results similar to those we observed in the classification and medical imaging tasks. The baselines under NBF exhibit inferior performance compared to those under RF (54.5 for $\mathrm{NBF} < 57.6$ for RF), but this issue is addressed by combining our approach with them $(+3.5\mathrm{mIoU})$ . Although out of our scope (refer to Appendix A.1), we validate one active labeling strategy ENT [62], which assigns highly uncertain (i.e., probably misclassified) pixels as feedback. Consequently, the feedback instructed by ENT is biasedly distributed in a manner similar to NBF. ENT also causes unexpected results (54.6 for $\mathrm{ENT} < 57.6$ for RF), and our approach alleviates this issue $(+3.1\mathrm{mIoU})$ . + +# 5.3 Ablation Study + +Positive vs. Negative feedback. We study the role of feedback on the adaptation results by varying feedback configurations. Let positively-provided feedback (PF) be obtained from samples that the source model correctly predicts, as opposed to negatively-provided feedback (NF). We adjust the ratio of PF:NF while keeping the total number of labeled samples constant, as shown in Figure 5. + +When using only FreeMatch (gray dot-dashed line), both biased feedback types (i.e., NBF and PBF) result in worse adaptation performance compared to balanced feedback for the baseline, e.g., 72.6 in 378:0 (PBF) $< 73.3$ in 252:126. In contrast, when our method is applied (red line), NBF yields the best performance. PBF and NBF can be respectively regarded as contributing previously known knowledge of the model and new knowledge that complements model deficiencies. Hence, it may be natural that NBF, which actually encodes the model's mistakes, contributes to favorable adaptation results. + +Number of unlabeled samples in a mini-batch. Existing SemiSDA methods [6,81] typically set the ratio $\mu$ between labeled and unlabeled samples in a + +![](images/e86a614287b88fe342b0453b03d79c126febf09d14fab568454cd2058975d420.jpg) +Fig. 5: NBF leads to higher performance than PBF. We compare different user-feedback configurations when the total number of feedback is 378 (top) and 630 (bottom). Positive and negative feedback refers to feedback from correct and incorrect model predictions, respectively. We run three random seed experiments and describe the average performance and standard deviation in the parenthesis. + +
methodfeed.negatively biased feedback (NBF)
#xulb, #xLD, #xlb112, 0, 16112, 48, 1664, 48, 16
total batch size128176128
FreeMatch [81]36872.074.274.8 (+0.6)
AdaMatch [6]64.571.372.0 (+0.7)
FreeMatch [81]63074.475.576.1 (+0.6)
AdaMatch [6]67.773.474.3 (+0.9)
+ +Table 7: In the mini-batch, diminishing the number of unlabeled samples and adding our defending samples achieves better performance with our approach. We ablate them by changing the ratio $\mu$ in Section 4.1, while keeping the size of labeled samples. + +mini-batch to 1:7. However, we observe that adhering to this ratio is not optimal for our approach, as shown in Table 7. Our method shows better performance when the ratio is varied to 1:4, i.e., decreasing unlabeled sample sizes. This finding contradicts observations in several TTA works [28, 48, 67], where adaptation performance tends to increase with larger batch sizes. We speculate that it is beneficial to prioritize more reliable information, which refers to labeled data and our defending samples selected from the filtering-applied bank, during the adapting process. This result may be aligned with previous works for curriculum learning [40, 100] and adaptive thresholding [96]. + +Number of labeled data. We measure the impact of feedback size (number of labeled samples) in Figure 6. The results show that the inferior performance on NBF persists even with an increased amount of feedback (gray $\rightarrow$ black line); however, our approach mitigates it and improves performance (black $\rightarrow$ red line). We make an interesting observation that the performance gap between black and red lines becomes larger as the number of available feedback decreases. Since obtaining large feedback may be challenging in real-world applications, our method is expected to be more helpful in this practical case. + +Data selecting strategy. We explore various strategies for selecting defending samples to balance the mini-batch, as shown in Table 8(top). The strategies include: in the $x_{LD}$ candidate bank, (i) random selection regardless of the class of the labeled data, (ii) random selection in the same class as the labeled data (i.e., class-aware), (iii) selecting samples close to the cluster center obtained by k means clustering [21] and (iv) selecting samples with embedded features distant from the labeled data where cosine distance is used. While our approach consistently outperforms the baseline regardless of the chosen strategy, we empirically + +![](images/75da3a9bd27938ce43af20c7a5b7e3e3269010c907386b3157e2e3cfae87737d.jpg) +Fig. 6: More reliable adaptation with NBF. In addition to Table 3, we conduct experiments with different amounts (1,3,5,10, and 15 labeled data per class) of feedback using FreeMatch [81]. The number of available feedback is likely to be small in practice. In this case, our method achieves large performance improvement, e.g., our method increases the baseline performance by $+4.9$ when one feedback per class is available. + +
selection strategy +class-awarerandom +xrandom +✓kmeans +✓cosine +✓baseline +-
FreeMatch [81]Res.74.174.874.674.072.0
FreeMatch [81]ViT.75.075.775.675.173.9
filtering rate0.20.40.60.8baseline only
FreeMatch [81]Res.74.574.874.373.772.0
FreeMatch [81]ViT.75.575.775.975.573.9
+ +Table 8: We ablate a component of our approach with 378 feedback: $x_{LD}$ selection strategy and filtering rate $p$ for bank generation. + +find that strategy (ii) achieves the best performance. Therefore, we adopt this strategy for our proposed method. + +Further studies, such as extension to a TTA scenario, combining with SFDA methods and different feedback configurations, are presented in Appendix C. + +# 6 Conclusion & Discussion + +User feedback can play an integral part in adapting the practical ML product to the target environment. However, we have shown that naive adaptation using existing SemiSDA methods led to undesirable adaptation results. We explained this through the lens of Negatively-Biased Feedback (NBF). In this paper, we uncovered the unexpected results of NBF and presented a scalable solution, Retrieval Latent Defending. This method prevents the mini-batch from becoming overly dependent on labeled samples that may have a biased distribution within the overall target distribution. Under the diverse DA benchmarks, from the simulation study to the medical imaging task, we demonstrated the practical problem caused by NBF and the effectiveness of our approach by combining it with multiple SemiSDA baselines. We hope our efforts will inspire future DA works on leveraging user feedback to improve an ML model in the deployment environment. + +Broader impact. The proposed setup assumes that an ML product obtains feedback as a form of annotations (i.e., labeled data). In some cases, users can provide feedback in different forms, like thumbs up & down and rating of model prediction, or noise feedback whose information is different from the ground truth. Further research considering these points will pave the way for developing safer and more reliable adapting strategies. + +Acknowledgment. We sincerely appreciate the abundant support provided by Lunit Inc., and we would like to thank Donggeun Yoo, Seonwook Park, and Sérgio Pereira for their valuable feedback. + +# References + +1. Arazo, E., Ortego, D., Albert, P., O'Connor, N.E., McGuinness, K.: Pseudolabeling and confirmation bias in deep semi-supervised learning. In: International Joint Conference on Neural Networks (IJCNN) (2020) 6, 9, 11, 23, 24, 25, 26, 29, 30 +2. Basak, H., Yin, Z.: Semi-supervised domain adaptive medical image segmentation through consistency regularized disentangled contrastive learning. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (2023) 4 +3. Baumeister, R.F., Bratslavsky, E., Finkenauer, C., Vohs, K.D.: Bad is stronger than good. Review of general psychology (2001) 2, 5 +4. Berthelot, D., Carlini, N., Cubuk, E.D., Kurakin, A., Sohn, K., Zhang, H., Raffel, C.: Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. In: ICLR (2020) 4 +5. Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.A.: Mixmatch: A holistic approach to semi-supervised learning. In: NeurIPS (2019) 4 +6. Berthelot, D., Roelofs, R., Sohn, K., Carlini, N., Kurakin, A.: Adamatch: A unified approach to semi-supervised learning and domain adaptation. In: ICLR (2022) 1, 4, 7, 8, 9, 10, 11, 12, 13, 22, 28, 30, 32, 33, 34 +7. Bradley, A.P.: The use of the area under the roc curve in the evaluation of machine learning algorithms. Pattern recognition (1997) 11 +8. Chen, D., Wang, D., Darrell, T., Ebrahimi, S.: Contrastive test-time adaptation. In: CVPR (2022) 3, 9, 10, 24, 25, 27, 28, 30, 35 +9. Chen, H., Tao, R., Fan, Y., Wang, Y., Wang, J., Schiele, B., Xie, X., Raj, B., Savvides, M.: Softmatch: Addressing the quantity-quality trade-off in semi-supervised learning. In: ICLR (2023) 4 +10. Chen, W., Lin, L., Yang, S., Xie, D., Pu, S., Zhuang, Y.: Self-supervised noisy label learning for source-free unsupervised domain adaptation. In: IROS (2022) 29 +11. Chen, X., Zhao, Z., Zhang, Y., Duan, M., Qi, D., Zhao, H.: Focalclick: Towards practical interactive image segmentation. In: CVPR (2022) 25 +12. Chen, Y., Tan, X., Zhao, B., Chen, Z., Song, R., Liang, J., Lu, X.: Boosting semi-supervised learning by exploiting all unlabeled data. In: CVPR (2023) 4 +13. Cohen, J.P., Viviano, J.D., Bertin, P., Morrison, P., Torabian, P., Guerrera, M., Lungren, M.P., Chaudhari, A., Brooks, R., Hashir, M., et al.: Torchxrayvision: A library of chest x-ray datasets and models. In: International Conference on Medical Imaging with Deep Learning (2022) 10, 11, 28 +14. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR (2016) 12, 29 +15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) 1, 10, 32, 33, 35 +16. Fini, E., Astolfi, P., Alahari, K., Alameda-Pineda, X., Mairal, J., Nabi, M., Ricci, E.: Semi-supervised learning made simple with self-supervised clustering. In: CVPR (2023) 4 + +17. Ganin, Y., Lempitsky, V.: Unsupervised domain adaptation by backpropagation. In: ICML (2015) 1, 3 +18. Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., Marchand, M., Lempitsky, V.: Domain-adversarial training of neural networks. The journal of machine learning research (2016) 3 +19. Gong, T., Jeong, J., Kim, T., Kim, Y., Shin, J., Lee, S.J.: Robust continual test-time adaptation: Instance-aware bn and prediction-balanced memory. In: NeurIPS (2023) 24 +20. Harada, S., Bise, R., Araki, K., Yoshizawa, A., Terada, K., Kurata, M., Nakajima, N., Abe, H., Ushiku, T., Uchida, S.: Cluster-guided semi-supervised domain adaptation for imbalanced medical image classification. arXiv preprint arXiv:2303.01283 (2023) 4 +21. Hartigan, J.A., Wong, M.A.: Algorithm as 136: A k-means clustering algorithm. Journal of the royal statistical society. series c (applied statistics) (1979) 13, 27 +22. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR (2020) 27 +23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) 1, 9, 10, 32, 33, 35 +24. Higuchi, Y., Moritz, N., Roux, J.L., Hori, T.: Momentum pseudo-labeling for semi-supervised speech recognition. In: Interspeech (2021) 4 +25. Huang, D., Li, J., Chen, W., Huang, J., Chai, Z., Li, G.: Divide and adapt: Active domain adaptation via customized learning. In: CVPR (2023) 4, 21, 22, 29 +26. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: CVPR (2017) 10, 11, 28 +27. Johnson, A.E., Pollard, T.J., Greenbaum, N.R., Lungren, M.P., Deng, C.y., Peng, Y., Lu, Z., Mark, R.G., Berkowitz, S.J., Horng, S.: Mimic-cxr-jpg, a large publicly available database of labeled chest radiographs. arXiv preprint arXiv:1901.07042 (2019) 9, 11 +28. Khurana, A., Paul, S., Rai, P., Biswas, S., Aggarwal, G.: Sita: Single image test-time adaptation. arXiv preprint arXiv:2112.02355 (2021) 13 +29. Kim, J., Hur, Y., Park, S., Yang, E., Hwang, S.J., Shin, J.: Distribution aligning refinery of pseudo-label for imbalanced semi-supervised learning. In: NeurIPS (2020) 22, 23 +30. Knox, W.B., Stone, P.: Tamer: Training an agent manually via evaluative reinforcement. In: IEEE international conference on development and learning (2008) 25 +31. Lee, H., Shin, S., Kim, H.: Abc: Auxiliary balanced classifier for class-imbalanced semi-supervised learning. In: NeurIPS (2021) 10, 22 +32. Lenga, M., Schulz, H., Saalbach, A.: Continual learning for domain adaptation in chest x-ray classification. In: Medical Imaging with Deep Learning (2020) 10, 11 +33. Li, J., Li, G., Shi, Y., Yu, Y.: Cross-domain adaptive clustering for semi-supervised domain adaptation. In: CVPR (2021) 4, 8, 10, 32, 33, 34 +34. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: ICML (2020) 1, 3, 4, 10, 26, 27, 35 +35. Litrico, M., Del Bue, A., Morerio, P.: Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation. In: CVPR (2023) 3, 26, 27, 29, 30, 35 +36. Liu, X., Xing, F., Shusharina, N., Lim, R., Jay Kuo, C.C., El Fakhri, G., Woo, J.: Act: Semi-supervised domain-adaptive medical image segmentation with asym + +metric co-training. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (2022) 4 +37. Liu, X., Yoo, C., Xing, F., Oh, H., El Fakhri, G., Kang, J.W., Woo, J., et al.: Deep unsupervised domain adaptation: A review of recent advances and perspectives. APSIPA Transactions on Signal and Information Processing (2022) 3 +38. Liu, Y., Zhang, W., Wang, J.: Source-free domain adaptation for semantic segmentation. In: CVPR (2021) 3 +39. Liu, Y., Kothari, P., van Delft, B., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? In: NeurIPS (2021) 24, 30 +40. Liu, Z., Miao, Z., Pan, X., Zhan, X., Lin, D., Yu, S.X., Gong, B.: Open compound domain adaptation. In: CVPR (2020) 13 +41. Long, M., Zhu, H., Wang, J., Jordan, M.I.: Deep transfer learning with joint adaptation networks. In: ICML (2017) 3 +42. MacGlashan, J., Ho, M.K., Loftin, R., Peng, B., Wang, G., Roberts, D.L., Taylor, M.E., Littman, M.L.: Interactive learning from policy-dependent human feedback. In: International conference on machine learning (2017) 25 +43. Madani, A., Moradi, M., Karargyris, A., Syeda-Mahmood, T.: Semi-supervised learning with generative adversarial networks for chest x-ray classification with ability of data domain adaptation. In: International symposium on biomedical imaging (2018) 4 +44. Mahapatra, D., Korevaar, S., Bozortabar, B., Tennakoon, R.: Unsupervised domain adaptation using feature disentanglement and gcns for medical image classification. In: ECCV (2022) 10, 11 +45. Mei, K., Zhu, C., Zou, J., Zhang, S.: Instance adaptive self-training for unsupervised domain adaptation. In: ECCV (2020) 3 +46. Mei, K., Zhu, C., Zou, J., Zhang, S.: Instance adaptive self-training for unsupervised domain adaptation. In: ECCV (2020) 12, 29 +47. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: ICML (2022) 24, 30 +48. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: ICLR (2023) 13 +49. Oh, Y., Kim, D.J., Kweon, I.S.: Daso: Distribution-aware semantics-oriented pseudo-label for imbalanced semi-supervised learning. In: CVPR (2022) 10, 22 +50. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. In: NeurIPS (2022) 25 +51. Pan, Y., Yao, T., Li, Y., Wang, Y., Ngo, C.W., Mei, T.: Transferrable prototypical networks for unsupervised domain adaptation. In: CVPR (2019) 3 +52. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, e.a.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019), https://github.com/pytorch/vision/blob/main/torchvision/models/resnet.py 28 +53. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournaepau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine learning in Python. Journal of Machine Learning Research (2011) 5, 25, 26 +54. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: ICCV (2019) 9 + +55. Prabhu, V., Chandrasekaran, A., Saenko, K., Hoffman, J.: Active domain adaptation via clustering uncertainty-weighted embeddings. In: ICCV (2021) 4, 21, 22, 29 +56. Richter, S.R., Vineet, V., Roth, S., Koltun, V.: Playing for data: Ground truth from computer games. In: ECCV (2016) 12, 29 +57. Rozin, P., Royzman, E.B.: Negativity bias, negativity dominance, and contagion. Personality and social psychology review (2001) 2, 5 +58. Saito, K., Kim, D., Sclaroff, S., Darrell, T., Saenko, K.: Semi-supervised domain adaptation via minimax entropy. In: ICCV (2019) 1, 3, 4, 8, 9, 10, 28, 32, 33, 34 +59. Saito, K., Watanabe, K., Ushiku, Y., Harada, T.: Maximum classifier discrepancy for unsupervised domain adaptation. In: CVPR (2018) 3 +60. Sajjadi, M., Javanmardi, M., Tasdizen, T.: Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In: NeurIPS (2016) 4 +61. Schulman, J., Zoph, B., Kim, C., Hilton, J., Menick, J., Weng, J., Uribe, J.F.C., Fedus, L., Metz, L., Pokorny, M., et al.: Chatgpt: Optimizing language models for dialogue. OpenAI blog (2022) 25 +62. Shen, Y., Yun, H., Lipton, Z.C., Kronrod, Y., Anandkumar, A.: Deep active learning for named entity recognition. In: ICLR (2017) 12, 22 +63. Shin, I., Kim, D.J., Cho, J.W., Woo, S., Park, K., Kweon, I.S.: Labor: Labeling only if required for domain adaptive semantic segmentation. In: ICCV (2021) 12, 29 +64. Sofiuk, K., Petrov, I., Barinova, O., Konushin, A.: f-brs: Rethinking backpropagating refinement for interactive segmentation. In: CVPR (2020) 25 +65. Sofiuk, K., Petrov, I.A., Konushin, A.: Reviving iterative training with mask guidance for interactive segmentation. In: IEEE International Conference on Image Processing (ICIP) (2022) 25 +66. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In: NeurIPS (2020) 1, 4, 8, 10, 28, 32, 33 +67. Song, J., Lee, J., Kweon, I.S., Choi, S.: Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. In: CVPR (2023) 1, 13, 24, 30 +68. Song, J., Park, K., Shin, I., Woo, S., Zhang, C., Kweon, I.S.: Test-time adaptation in the dynamic world with compound domain knowledge management. IEEE Robotics and Automation Letters (2023) 24 +69. Stiennon, N., Ouyang, L., Wu, J., Ziegler, D., Lowe, R., Voss, C., Radford, A., Amodei, D., Christiano, P.F.: Learning to summarize with human feedback. In: NeurIPS (2020) 25 +70. Sun, B., Saenko, K.: Deep coral: Correlation alignment for deep domain adaptation. In: ECCV Workshops (2016) 1, 3 +71. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 25 +72. Tsai, Y.H., Hung, W.C., Schulter, S., Sohn, K., Yang, M.H., Chandraker, M.: Learning to adapt structured output space for semantic segmentation. In: CVPR (2018) 1, 3, 12 +73. Tsai, Y.H., Hung, W.C., Schulter, S., Sohn, K., Yang, M.H., Chandraker, M.: Learning to adapt structured output space for semantic segmentation. In: CVPR (2018) 3 +74. Van Engelen, J.E., Hoos, H.H.: A survey on semi-supervised learning. Machine learning (2020) 4 + +75. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: CVPR (2017) 9 +76. Vu, T.H., Jain, H., Bucher, M., Cord, M., Pérez, P.: Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In: CVPR (2019) 3 +77. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: ICLR (2021) 3, 24, 30 +78. Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing (2018) 1 +79. Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: CVPR (2022) 30 +80. Wang, Y., Chen, H., Fan, Y., Sun, W., Tao, R., Hou, W., Wang, R., Yang, L., Zhou, Z., Guo, L.Z., et al.: Usb: A unified semi-supervised learning benchmark for classification. In: NeurIPS (2022) 10 +81. Wang, Y., Chen, H., Heng, Q., Hou, W., Fan, Y., Wu, Z., Wang, J., Savvides, M., Shinozaki, T., Raj, B., et al.: Freematch: Self-adaptive thresholding for semi-supervised learning. In: ICLR (2023) 4, 7, 8, 10, 12, 13, 14, 24, 25, 27, 28, 30, 32, 33, 34 +82. Warnell, G., Waytowich, N., Lawhern, V., Stone, P.: Deep tamer: Interactive agent shaping in high-dimensional state spaces. In: AAAI (2018) 25 +83. Wei, C., Sohn, K., Mellina, C., Yuille, A., Yang, F.: Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In: CVPR (2021) 10, 22, 23, 24, 29 +84. Wirth, C., Akhour, R., Neumann, G., Fünkranz, J., et al.: A survey of preference-based reinforcement learning methods. Journal of Machine Learning Research (2017) 25 +85. Xie, B., Yuan, L., Li, S., Liu, C.H., Cheng, X.: Towards fewer annotations: Active learning via region impurity and prediction uncertainty for domain adaptive semantic segmentation. In: CVPR (2022) 12, 29 +86. Xie, M., Li, Y., Wang, Y., Luo, Z., Gan, Z., Sun, Z., Chi, M., Wang, C., Wang, P.: Learning distinctive margin toward active domain adaptation. In: CVPR (2022) 4, 21 +87. Xie, Q., Dai, Z., Hovy, E., Luong, T., Le, Q.: Unsupervised data augmentation for consistency training. In: NeurIPS (2020) 4, 8, 10, 32, 33, 34 +88. Xu, Y., Shang, L., Ye, J., Qian, Q., Li, Y.F., Sun, B., Li, H., Jin, R.: Dash: Semi-supervised learning with dynamic thresholding. In: ICML (2021) 4 +89. Yan, Z., Wu, Y., Li, G., Qin, Y., Han, X., Cui, S.: Multi-level consistency learning for semi-supervised domain adaptation. In: IJCAI (2022) 4 +90. Yang, L., Wang, Y., Gao, M., Shrivastava, A., Weinberger, K.Q., Chao, W.L., Lim, S.N.: Deep co-training with task decomposition for semi-supervised domain adaptation. In: ICCV (2021) 4 +91. Yang, S., Jui, S., van de Weijer, J., et al.: Attracting and dispersing: A simple approach for source-free domain adaptation. In: NeurIPS (2022) 3 +92. Yang, S., Wang, Y., Van De Weijer, J., Herranz, L., Jui, S.: Generalized source-free domain adaptation. In: ICCV (2021) 1, 3 +93. Yang, S., van de Weijer, J., Herranz, L., Jui, S., et al.: Exploiting the intrinsic neighborhood structure for source-free domain adaptation. In: NeurIPS (2021) 3, 27, 28, 35 +94. Yu, Y.C., Lin, H.T.: Semi-supervised domain adaptation with source label adaptation. In: CVPR (2023) 4, 9, 10 + +95. Yu, Z., Li, J., Du, Z., Zhu, L., Shen, H.T.: A comprehensive survey on source-free domain adaptation. arXiv preprint arXiv:2302.11803 (2023) 3 +96. Zhang, B., Wang, Y., Hou, W., Wu, H., Wang, J., Okumura, M., Shinozaki, T.: Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In: NeurIPS (2021) 7, 8, 10, 13, 32, 33, 34 +97. Zhang, C., Miech, A., Shen, J., Alayrac, J.B., Luc, P.: Making the most of what you have: Adapting pre-trained visual language models in the low-data regime. arXiv preprint arXiv:2305.02297 (2023) 3 +98. Zhang, W., Ouyang, W., Li, W., Xu, D.: Collaborative and adversarial network for unsupervised domain adaptation. In: CVPR (2018) 3 +99. Zhang, Y., Zhang, H., Deng, B., Li, S., Jia, K., Zhang, L.: Semi-supervised models are strong unsupervised domain adaptation learners. arXiv preprint arXiv:2106.00417 (2021) 1, 4, 8 +100. Zhang, Y., David, P., Gong, B.: Curriculum domain adaptation for semantic segmentation of urban scenes. In: ICCV (2017) 13 +101. Zhang, Y., Wang, Z., He, W.: Class relationship embedded learning for source-free unsupervised domain adaptation. In: CVPR (2023) 3 \ No newline at end of file diff --git a/2024/Is user feedback always informative_ Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data/images.zip b/2024/Is user feedback always informative_ Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b20c5916b2762506a47231d0761d4aa8e6118ac6 --- /dev/null +++ b/2024/Is user feedback always informative_ Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7890d26a02db9f9bc5b50a90a48c32ce51a53e6d79bdc5433f2adab2d6bc2e1 +size 630923 diff --git a/2024/Is user feedback always informative_ Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data/layout.json b/2024/Is user feedback always informative_ Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b0afe27b5b5dc0fcc4e2aab88fdd7b2a8dc472b6 --- /dev/null +++ b/2024/Is user feedback always informative_ Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data/layout.json @@ -0,0 +1,10751 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 112, + 474, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 112, + 474, + 166 + ], + "spans": [ + { + "bbox": [ + 140, + 112, + 474, + 166 + ], + "type": "text", + "content": "Is user feedback always informative? Retrieval Latent Defending for Semi-Supervised Domain Adaptation without Source Data" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 161, + 185, + 452, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 185, + 452, + 210 + ], + "spans": [ + { + "bbox": [ + 161, + 185, + 452, + 210 + ], + "type": "text", + "content": "Junha Song" + }, + { + "bbox": [ + 161, + 185, + 452, + 210 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 161, + 185, + 452, + 210 + ], + "type": "text", + "content": ", Tae Soo Kim" + }, + { + "bbox": [ + 161, + 185, + 452, + 210 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 161, + 185, + 452, + 210 + ], + "type": "text", + "content": ", Junha Kim" + }, + { + "bbox": [ + 161, + 185, + 452, + 210 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 161, + 185, + 452, + 210 + ], + "type": "text", + "content": ", Gunhee Nam" + }, + { + "bbox": [ + 161, + 185, + 452, + 210 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 161, + 185, + 452, + 210 + ], + "type": "text", + "content": ", Thijs Kooi" + }, + { + "bbox": [ + 161, + 185, + 452, + 210 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 161, + 185, + 452, + 210 + ], + "type": "text", + "content": ", and Jaegul Choo" + }, + { + "bbox": [ + 161, + 185, + 452, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 170, + 219, + 443, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 219, + 443, + 243 + ], + "spans": [ + { + "bbox": [ + 170, + 219, + 443, + 243 + ], + "type": "text", + "content": "1 KAIST {sb020518, jchoo}@kaist.ac.kr 2 Lunit Inc. {junha.kim, taesoo.kim, ghnam, tkooi}@lunit.io" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 271, + 455, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 271, + 455, + 469 + ], + "spans": [ + { + "bbox": [ + 160, + 271, + 455, + 469 + ], + "type": "text", + "content": "Abstract. This paper aims to adapt the source model to the target environment, leveraging small user feedback (i.e., labeled target data) readily available in real-world applications. We find that existing semi-supervised domain adaptation (SemiSDA) methods often suffer from poorly improved adaptation performance when directly utilizing such feedback data, as shown in Figure 1. We analyze this phenomenon via a novel concept called Negatively Biased Feedback (NBF), which stems from the observation that user feedback is more likely for data points where the model produces incorrect predictions. To leverage this feedback while avoiding the issue, we propose a scalable adapting approach, Retrieval Latent Defending. This approach helps existing SemiSDA methods to adapt the model with a balanced supervised signal by utilizing latent defending samples throughout the adaptation process. We demonstrate the problem caused by NBF and the efficacy of our approach across various benchmarks, including image classification, semantic segmentation, and a real-world medical imaging application. Our extensive experiments reveal that integrating our approach with multiple state-of-the-art SemiSDA methods leads to significant performance improvements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 479, + 441, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 479, + 441, + 502 + ], + "spans": [ + { + "bbox": [ + 159, + 479, + 441, + 502 + ], + "type": "text", + "content": "Keywords: Rethinking user-provided feedback " + }, + { + "bbox": [ + 159, + 479, + 441, + 502 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 479, + 441, + 502 + ], + "type": "text", + "content": " Semi-supervised & Source-free domain adaptation " + }, + { + "bbox": [ + 159, + 479, + 441, + 502 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 479, + 441, + 502 + ], + "type": "text", + "content": " Medical image diagnosis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 525, + 230, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 230, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 230, + 538 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": "While deep neural networks have demonstrated remarkable performance in the development domain (i.e., source domain) [15,23], they often suffer from performance degradation in the deployed domain (i.e., target domain) due to domain shift [17, 72, 78]. To mitigate this issue, domain adaptation (DA) techniques have been introduced [34, 58, 70]. The most common DA tasks include semi-supervised domain adaptation (SemiSDA) and source-free domain adaptation (SFDA). SemiSDA aims to adapt the model given a small amount of labeled target data along with massive unlabeled target data [6, 58, 66, 99]. SFDA conducts adaptation with only target data without accessing source data considering data privacy or memory constraints in edge devices [34, 67, 92]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 115, + 304, + 193 + ], + "blocks": [ + { + "bbox": [ + 135, + 115, + 304, + 193 + ], + "lines": [ + { + "bbox": [ + 135, + 115, + 304, + 193 + ], + "spans": [ + { + "bbox": [ + 135, + 115, + 304, + 193 + ], + "type": "image", + "image_path": "a2665f627bd6c298bd33a6f9b2a5cd19d930495806a579cc983bb24a3ecd883a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 196, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 130, + 196, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 196, + 482, + 285 + ], + "type": "text", + "content": "Fig. 1: (a) User feedback. Users can provide feedback while interacting with an ML product, where feedback is likely to be biased towards misclassified samples, which we define as Negatively Biased Feedback (NBF). (b) Adaptation results. We adapt the source model with small user feedback and large unlabeled target data using previous semi-supervised domain adaptation (SemiSDA) algorithms. Compared to random feedback, which is the classical SemiSDA setup where labeled data is a random subset of target data, model adaptation with NBF leads to subpar performance. This paper analyzes this problem and introduces a scalable solution." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 116, + 474, + 194 + ], + "blocks": [ + { + "bbox": [ + 306, + 116, + 474, + 194 + ], + "lines": [ + { + "bbox": [ + 306, + 116, + 474, + 194 + ], + "spans": [ + { + "bbox": [ + 306, + 116, + 474, + 194 + ], + "type": "image", + "image_path": "1fa349cee07a879821bb1b0df1354fb785a490a9f81e0ae77032670bdc690d4a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 292, + 482, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 424 + ], + "type": "text", + "content": "Despite such advances in DA, adapting the model with user feedback still remains an open area for further research, even though practical machine learning (ML) products often allow users to provide feedback in order to further improve the model in the target environment. For example, facial recognition or medical image diagnosis applications enable users to give feedback correcting wrong model predictions, as depicted in Figure 1 (a). Since feedback can be modeled in this case as a small amount of labeled target data, it is anticipated that previous SemiSDA methods assuming the same setup would yield promising results. However, we observe that they show inferior adaptation performance on multiple DA benchmarks when using such user feedback in practice, as shown in the dark-gray bar in Figure 1 (b)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "type": "text", + "content": "We introduce a novel concept called Negatively Biased Feedback (NBF) to explain this phenomenon. NBF is based on the observation that user feedback is more likely to be derived from incorrect model predictions. For example, a radiologist might log a misdiagnosed chest X-ray by the model, as its accuracy directly impacts the patient's survival. Interestingly, our observation aligns with findings from cognitive psychology literature [3,57] that proves that humans are more likely to react and provide feedback to negative events (i.e., wrong model predictions). Since such an NBF scenario is feasible, we analyze its unexpected impact on SemiSDA observed above. We identify that a biased distribution of NBF within the overall data distribution leads to sub-optimal adaptation results, particularly compared to Random Feedback (RF). RF represents the classical SemiSDA setup, where labeled data is randomly selected from the target data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "To address the problem caused by NBF, we present a scalable approach named Retrieval Latent Defending, which can be seamlessly integrated with existing SemiSDA methods. Our approach allows them to adapt the model without a strong dependence on the biasedly distributed labeled data. Specifically, we balance the supervised adapting signal by appending latent defending samples to the mini-batch and help to keep the model's balanced class discriminability throughout adapting iterations. We evaluate the unexpected influence of NBF using various benchmarks, including image classification, semantic segmentation," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "type": "text", + "content": "J. Song et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 476, + 193 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 476, + 193 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 476, + 193 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 476, + 193 + ], + "type": "image", + "image_path": "f43b6f978ec654ada808edc9a32d3e1c9a0d4b92b99841e0deeeb68abbd1f86f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 196, + 481, + 251 + ], + "lines": [ + { + "bbox": [ + 130, + 196, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 196, + 481, + 251 + ], + "type": "text", + "content": "Fig. 2: Adaptation with user feedback can be effective in alleviating performance degradation caused by domain shift. However, there are some challenges: (i) user feedback may be a biased sampling of the true target distribution due to the nature of feedback, (ii) the amount of the ground truths (GT) labels obtained through feedback is small, and (iii) only unlabeled target data is typically available, not source data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 262, + 479, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 479, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 479, + 297 + ], + "type": "text", + "content": "and medical image diagnosis. Building upon these evaluations, we demonstrate that our approach not only complements, but significantly enhances the performance of multiple SemiSDA methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 299, + 348, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 299, + 348, + 311 + ], + "spans": [ + { + "bbox": [ + 146, + 299, + 348, + 311 + ], + "type": "text", + "content": "The contributions of the paper are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 313, + 480, + 430 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 141, + 313, + 480, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 313, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 141, + 313, + 480, + 347 + ], + "type": "text", + "content": "We introduce the novel concept called Negatively Biased Feedback and uncover that it can lead to sub-optimal adaptation performance of existing SemiSDA methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 349, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 349, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 141, + 349, + 480, + 384 + ], + "type": "text", + "content": "- We analyze this problem and present a scalable solution, Retrieval Latent Defending, that combines with SemiSDA methods and allows them to avoid the unexpected effect of NBF." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 384, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 384, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 141, + 384, + 480, + 407 + ], + "type": "text", + "content": "- We show that our approach generalizes through diverse DA benchmarks and improves adaptation results of state-of-the-art SemiSDA methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 407, + 480, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 407, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 141, + 407, + 480, + 430 + ], + "type": "text", + "content": "We publicly release the code on https://github.com/junha1125/RLD-SemiSDA." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 438, + 237, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 438, + 237, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 438, + 237, + 451 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": "Adaptation in the deployment environment. Real-world ML products often encounter performance degradation caused by gaps between the source and target environment [17]. One solution is to adapt the model using unlabeled data observed in the target domain, referred to as unsupervised domain adaptation (UDA) [37, 59, 72]. Works on UDA use both source and target data to improve the target performance by using methods such as domain discrepancy minimization by adversarial training [18, 41, 59, 70, 72, 73, 76], and self-training with pseudo labels [45, 51, 97, 98]. Source-free DA (SFDA) builds on UDA and imposes an additional constraint that the source data can not be accessed during domain adaptation. This has practical implications for addressing data privacy concerns or barriers in data transmission to edge devices [34, 38, 77, 95]. The majority of recent SFDA works rely on strategies like domain clustering [34], nearest neighbors [91-93], and contrastive learning [8, 35, 101]. Nevertheless, SFDA does not consider the availability of small labeled data, which may be available in practical ML systems. Semi-supervised DA (SemiSDA) works mainly demonstrate that permitting small labeled data in the target domain can substantially enhance adaptation performance compared to traditional UDA [58]." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "NBF & RLD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "Their primary strategy is to use domain alignment [20, 33, 58, 94], multi-view consistency [2, 6, 33, 89], and asymmetric co-training [36, 90]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 144, + 480, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 144, + 480, + 227 + ], + "spans": [ + { + "bbox": [ + 130, + 144, + 480, + 227 + ], + "type": "text", + "content": "Active domain adaptation (ActiveDA) [25, 55, 86] envisions a scenario in which the machine selects specific target samples and instructs annotators to label them. The primary objective of ActiveDA is to strategically identify and select the most informative samples for annotation. These chosen samples (i.e., labeled target data) are subsequently utilized to update the source model using SemiSDA methods [33, 58], and the effectiveness of ActiveDA is assessed by evaluating the target performance of the adapted model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 231, + 480, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 231, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 480, + 314 + ], + "type": "text", + "content": "Semi-supervised learning (SemiSL) aims to reduce expensive human annotations, and propose methods to train a model from scratch using massive unlabeled data along with limited amounts of labeled data [43,74]. The majority of SemiSL methods depend on consistency regularization [4,5,16,60,66,87], which helps the model to make similar predictions for augmented versions of the same image. Moreover, adaptive thresholding [9,12,24,66,81,88,99] is also popularly utilized to produce reliable pseudo labels from unlabeled data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 315, + 480, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 480, + 387 + ], + "type": "text", + "content": "SemiSDA and SemiSL setups mimic small labeled datasets by randomly selecting subsets of the target dataset, whereas ActiveDA involves selections instructed by the machine. In contrast, this paper posits that in real-world applications, labeled data is typically acquired through user intervention. Additionally, users often provide feedback on samples misclassified by the model (i.e., negatively biased feedback), a process detailed in the following section." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 134, + 392, + 478, + 423 + ], + "blocks": [ + { + "bbox": [ + 134, + 392, + 478, + 423 + ], + "lines": [ + { + "bbox": [ + 134, + 392, + 478, + 423 + ], + "spans": [ + { + "bbox": [ + 134, + 392, + 478, + 423 + ], + "type": "table", + "html": "
UDASFDAActiveDASemiSDASemiSLOur setup
Adaptation○○○○×○
Source-free×○××-○
Feedback××machine-instructedrandomly selectedrandomly selecteduser-provided
", + "image_path": "5e746c549ebb8dcd0b0088beb454eb3d772d8024870d203e9273ca7d13ddd968.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 424, + 480, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 480, + 483 + ], + "type": "text", + "content": "The table above summarizes the comparison of relevant studies to our setup. In the table, adaptation means fine-tuning the source pre-trained model (as opposed to training from scratch); feedback represents a small number of labeled target samples. Appendix A provides further comparisons with settings like class-imbalanced SemiSDA and test-time adaptation (TTA)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 496, + 319, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 496, + 319, + 511 + ], + "spans": [ + { + "bbox": [ + 131, + 496, + 319, + 511 + ], + "type": "text", + "content": "3 Negatively Biased Feedback" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 519, + 318, + 531 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 519, + 318, + 531 + ], + "spans": [ + { + "bbox": [ + 131, + 519, + 318, + 531 + ], + "type": "text", + "content": "3.1 Adaptation with user feedback." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": "Our adaptation setup is illustrated in Figure 2. A model is pre-trained on the source data " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "D_{s}" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": ". Next, the model is deployed to the target domain, such as a smartphone or a hospital, where we assume the transfer of " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "D_{s}" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " is prohibited due to data privacy regulations or resource constraints (same setup as SFDA [34]). While users utilize ML products on the target domain, the model provides prediction results for data observed in the target domain " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " and occasionally obtains user feedback in the form of annotations " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": ". We represent the target data as " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "D_{t} = X_{t}^{lb} \\cup X_{t}^{ulb}" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "X_{t}^{lb} = \\{(x_{lb}^{n},y_{lb}^{n}):n\\in [1..N_{lb}]\\}" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "X_{t}^{ulb} = \\{(x_{ulb}^{n}):n\\in [1..N_{ulb}]\\}" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "x_{lb}" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "x_{ulb}" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " denote labeled and unlabeled data and " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "N_{lb}" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "N_{ulb}" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " is their number of data. Lastly, the model can utilize " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " and SemiSDA algorithms for adaptation during its inactive phase (e.g., when users do not use the product," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "text", + "content": "J. Song et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 461, + 275 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 461, + 275 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 461, + 275 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 461, + 275 + ], + "type": "image", + "image_path": "df2bd396b0b566c13e4bd4684d90efe02dc750deafbbad1cc6f1a1fae253d0e6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 277, + 482, + 354 + ], + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 354 + ], + "type": "text", + "content": "Fig.3: Effect of negatively biased feedback. Our novel observations are that " + }, + { + "bbox": [ + 130, + 277, + 482, + 354 + ], + "type": "inline_equation", + "content": "(a)" + }, + { + "bbox": [ + 130, + 277, + 482, + 354 + ], + "type": "text", + "content": " user-provided feedback in practice has a biased distribution in each class cluster (the bottom center sub-figure) which is in contrast to random feedback, " + }, + { + "bbox": [ + 130, + 277, + 482, + 354 + ], + "type": "inline_equation", + "content": "(b)" + }, + { + "bbox": [ + 130, + 277, + 482, + 354 + ], + "type": "text", + "content": " Existing SemiSDA methods adapt the model by dominating the labeled data points (the right sub-figures) even though they are biasedly positioned, and " + }, + { + "bbox": [ + 130, + 277, + 482, + 354 + ], + "type": "inline_equation", + "content": "(c)" + }, + { + "bbox": [ + 130, + 277, + 482, + 354 + ], + "type": "text", + "content": " NBF prevents the model from having a decision boundary for true class clusters and leads to inferior adaptation performance (the bottom right sub-figure)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 364, + 480, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 480, + 389 + ], + "type": "text", + "content": "like at nighttime) in order to alleviate performance degradation due to domain shift or to personalize the model based on user feedback." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 392, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 583 + ], + "type": "text", + "content": "Rethinking user-provided feedback. Classical SemiSDA works simply assume that a random subset in target data " + }, + { + "bbox": [ + 130, + 392, + 482, + 583 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 130, + 392, + 482, + 583 + ], + "type": "text", + "content": " is labeled by users when building " + }, + { + "bbox": [ + 130, + 392, + 482, + 583 + ], + "type": "inline_equation", + "content": "X_{t}^{lb}" + }, + { + "bbox": [ + 130, + 392, + 482, + 583 + ], + "type": "text", + "content": ". However, as illustrated in Figure 2 (i), we suggest that users are more likely to provide feedback on misclassified samples by the source model, named negatively biased feedback (NBF). This behavior can be understood from two perspectives: (a) users generally expect their feedback to be used as a basis of model improvement, motivating them to provide NBF, and (b) humans tend to react more strongly to negative experiences, such as receiving incorrect predictions, as observed in psychological studies [3,57]. We note that the NBF assumption holds more strongly for the medical application: it is reasonable to imagine that the user (i.e., radiologist) logs the mistakes of the model while diagnosing a chest X-ray exam because the diagnostic accuracy of the model is directly related to the patient's chances of survival. Furthermore, applications beyond the medical domain can also exhibit NBF. For instance, users in self-driving cars can report errors, such as object detection failures or navigation mistakes, to enhance the car's driving capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 597, + 316, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 316, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 316, + 608 + ], + "type": "text", + "content": "3.2 Influence of NBF on SemiSDA" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Simulation study. As shown in Figure 3, we conduct a simulation study to understand the effect of NBF on SemiSDA. We first use the blobs dataset [53] and construct the source and target data so that domain shift exists between them (left sub-figures). We pre-train a source model on the source data and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "NBF & RLD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "compute the accuracy in the target domain, where the performance drop due to domain shift is observed " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "(98.5\\% \\rightarrow 76.4\\%)" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": ". Next, we simulate two types of feedback (i.e., labeled data): random feedback and negatively biased feedback following a previous SemiSDA setup and our setup, respectively. Specifically, NBF is randomly selected among misclassified samples by the source model. We find that random feedback (RF) points are evenly distributed, while NBF points are biasedly positioned across each class cluster (refer to blue points in the dashed circle in the center sub-figures)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 482, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 356 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 356 + ], + "type": "text", + "content": "To alleviate the performance drop caused by domain shift, we adapt the model using the target data and a semi-supervised method, Pseudo-labeling [1]. This method iteratively optimizes the model by the cross-entropy loss computed by the ground truth of labeled data and pseudo labels of unlabeled data in a minibatch (pseudo labels are predicted by the current adapting model so they can be changed according to an updating decision boundary. Further comprehension can be achieved by referring Appendix B.). The SemiSDA results are shown in the right sub-figures, where we make two interesting observations: (i) the distribution of labeled data can contribute significantly to a decision boundary of the adapted model (red arrows in the figure), and (ii) the adapted model under NBF has poorly improved performance compared with one under RF " + }, + { + "bbox": [ + 130, + 212, + 482, + 356 + ], + "type": "inline_equation", + "content": "(76.4\\% \\rightarrow 88.1\\%" + }, + { + "bbox": [ + 130, + 212, + 482, + 356 + ], + "type": "text", + "content": " with NBF, but " + }, + { + "bbox": [ + 130, + 212, + 482, + 356 + ], + "type": "inline_equation", + "content": "76.4\\% \\rightarrow 91.7\\%" + }, + { + "bbox": [ + 130, + 212, + 482, + 356 + ], + "type": "text", + "content": " with RF)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 359, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 482, + 491 + ], + "type": "text", + "content": "Unexpected influence of NBF. Our intuitive reasoning probably suggests that NBF provides more information than RF by correcting more source model deficiencies, and thus leads to better adaptation performance. However, we empirically show that NBF can result in inferior adaptation performance due to its biased distribution across each class cluster, as illustrated in Figure 3. Surprisingly, we also show that this problem persists, even with other state-of-the-art SemiSDA methods and large datasets for various DA benchmarks, including image classification, semantic segmentation, and medical image diagnosis. Our work highlights the importance of careful design when using user feedback in real-world scenarios and, to the best of our knowledge, is the first study to uncover and analyze this phenomenon." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 499, + 212, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 499, + 212, + 513 + ], + "spans": [ + { + "bbox": [ + 132, + 499, + 212, + 513 + ], + "type": "text", + "content": "4 Approach" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 519, + 367, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 519, + 367, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 367, + 533 + ], + "type": "text", + "content": "4.1 Prerequisite: Previous SemiSDA method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": "Previous SemiSDA and SemiSL works typically construct a mini-batch with labeled data " + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\{(x_{lb}^b, y_{lb}^b) : b \\in [1..B]\\}" + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": ", and unlabeled data whose size is " + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": " times larger than labeled ones " + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\{(x_{ulb}^b) : b \\in [1..\\mu \\cdot B]\\}" + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": " is the mini-batch size for labeled data. To adapt the model iteratively, they compute the cross-entropy loss " + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 536, + 482, + 609 + ], + "type": "text", + "content": " with labeled data and the consistency regularization to multi-view of unlabeled data, which are formulated as the following:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 164, + 610, + 481, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 610, + 481, + 639 + ], + "spans": [ + { + "bbox": [ + 164, + 610, + 481, + 639 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s u p} = \\frac {1}{B} \\sum_ {b = 1} ^ {B} \\mathcal {H} \\left(y _ {l b} ^ {b}, f _ {\\theta} \\left(x _ {l b} ^ {b}\\right)\\right), \\quad \\mathcal {L} _ {u n s u p} = \\frac {1}{\\mu \\cdot B} \\sum_ {b = 1} ^ {\\mu \\cdot B} \\mathcal {H} \\left(\\hat {y} _ {u l b} ^ {b}, f _ {\\theta} \\left(\\Omega \\left(x _ {u l b} ^ {b}\\right)\\right)\\right), \\tag {1}", + "image_path": "fd695a05cceb7c1032e7bb29159286bf3cf5da0147eedc513337a6a4707e07a2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " is the output probability from the model, " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\hat{y}_{ulb}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " denotes a pseudo label obtained from " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\omega(x_{ulb}))" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\omega(\\cdot)" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\Omega(\\cdot)" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " represent weak and strong image" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Song et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 465, + 287 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 465, + 287 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 465, + 287 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 465, + 287 + ], + "type": "image", + "image_path": "623147e71e483c912e40ebb7b5a40dd915dc5a23b83f3ae1375eb1e375a43f13.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 289, + 482, + 368 + ], + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 368 + ], + "type": "text", + "content": "Fig. 4: Even though labeled data " + }, + { + "bbox": [ + 130, + 289, + 482, + 368 + ], + "type": "inline_equation", + "content": "(x_{lb}, y_{lb})" + }, + { + "bbox": [ + 130, + 289, + 482, + 368 + ], + "type": "text", + "content": " is biasedly positioned, the model needs to be adapted with balanced class discriminability (i.e., decision boundary). (i) However, previous SemiSDA methods have overlooked this fact and used the labeled data naively by applying a cross-entropy loss, leading to inadequate adaptation performance. (ii) To alleviate this problem, we propose a scalable adapting approach, retrieval latent defending, which allows the model to adjust the balance of a mini-batch on each iteration by using latent defending samples " + }, + { + "bbox": [ + 130, + 289, + 482, + 368 + ], + "type": "inline_equation", + "content": "x_{LD}" + }, + { + "bbox": [ + 130, + 289, + 482, + 368 + ], + "type": "text", + "content": " together with labeled data " + }, + { + "bbox": [ + 130, + 289, + 482, + 368 + ], + "type": "inline_equation", + "content": "x_{lb}" + }, + { + "bbox": [ + 130, + 289, + 482, + 368 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 376, + 480, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 480, + 413 + ], + "type": "text", + "content": "augmentation, respectively. While sharing the core framework, each SemiSDA method employs distinct adapting strategies, especially to enhance the effectiveness of the use of unlabeled data rather than labeled data [6,81,96]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 416, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 482, + 489 + ], + "type": "text", + "content": "Problem of previous works. Since previous SemiSDA methods have overlooked the unexpected impact of NBF, they often suffer from sub-optimal performance under the NBF assumption (shown in Section 5). To address this problem, we focus on developing a scalable solution that (i) can easily combine with existing DA methods without modifying their core adapting strategies and (ii) can be applied to a wide range of benchmarks, including medical image diagnosis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 503, + 298, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 298, + 516 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 298, + 516 + ], + "type": "text", + "content": "4.2 Retrieval Latent Defending" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Based on the observations in Figure 3, we illustrate the unintended effect of NBF when using an existing SemiSDA method in Figure 4 (top center), where NBF is likely to exhibit a biased distribution, leading to undesirable adaptation results. To alleviate this issue, we propose Retrieval Latent Defending as depicted in Figure 4 (bottom). " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "①" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " Prior to each epoch, we generate a candidate bank of data points, denoted as " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{LD}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "② \\sim ④" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " For each adapting iteration, we balance the mini-batch by retrieving latent defending samples " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{LD}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " from the bank. " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "⑤ \\sim ⑥" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " The model is then adapted using the reconfigured mini-batch and following the baseline SemiSDA approach. We hypothesize that the latent space progressively created by the " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{LD}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " candidates throughout the adaptation process (bold dashed circle in Figure 4 (top right)) mitigates the issue caused by NBF, thereby allowing the SemiSDA baseline to achieve robust adaptation against NBF." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "NBF & RLD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "Candidate bank generation. The candidate bank serves as a repository of pseudo labels " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\hat{Y}_t^{ulb}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " for a subset of the target unlabeled data " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "X_t^{ulb}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ". Before each epoch, we freeze the model and use it to generate pseudo labels " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\hat{Y}_t^{ulb} = \\{(\\hat{y}_{ulb}^n): n \\in [1..N_{ulb}]\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\hat{y}_{ulb}^n" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " is assigned to " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "x_{ulb}^n" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " as the predicted class with the highest softmax probability: " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\hat{y}_{ulb}^n = \\operatorname{argmax}_c [f_\\theta(x_{ulb}^n)]_c" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ". We then retain only samples with the top " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "p\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " highest probabilities within each class. This filtering step helps mitigate the inclusion of data with potentially inaccurate pseudo labels, as the model's predictions on " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "X_t^{ulb}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " might not always be perfect." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 215, + 482, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 347 + ], + "type": "text", + "content": "Defending sample selection. We select " + }, + { + "bbox": [ + 130, + 215, + 482, + 347 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 215, + 482, + 347 + ], + "type": "text", + "content": " latent defending samples " + }, + { + "bbox": [ + 130, + 215, + 482, + 347 + ], + "type": "inline_equation", + "content": "x_{LD}" + }, + { + "bbox": [ + 130, + 215, + 482, + 347 + ], + "type": "text", + "content": " from the bank at random for each labeled data " + }, + { + "bbox": [ + 130, + 215, + 482, + 347 + ], + "type": "inline_equation", + "content": "(x_{lb}^{b}, y_{lb}^{b})" + }, + { + "bbox": [ + 130, + 215, + 482, + 347 + ], + "type": "text", + "content": ". These selected samples share the same pseudo label as the ground-truth label of their associated counterparts (i.e., " + }, + { + "bbox": [ + 130, + 215, + 482, + 347 + ], + "type": "inline_equation", + "content": "\\hat{y}_{LD} = y_{lb}^{b}" + }, + { + "bbox": [ + 130, + 215, + 482, + 347 + ], + "type": "text", + "content": "). By incorporating these defending samples, we balance the data distribution within the current mini-batch. For example, consider " + }, + { + "bbox": [ + 130, + 215, + 482, + 347 + ], + "type": "inline_equation", + "content": "x_{lb}^{1}" + }, + { + "bbox": [ + 130, + 215, + 482, + 347 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 215, + 482, + 347 + ], + "type": "inline_equation", + "content": "x_{lb}^{2}" + }, + { + "bbox": [ + 130, + 215, + 482, + 347 + ], + "type": "text", + "content": " in Figure 4 (top right). As these labeled samples are included in the current mini-batch alongside the selected defending samples " + }, + { + "bbox": [ + 130, + 215, + 482, + 347 + ], + "type": "inline_equation", + "content": "x_{LD}^{1}" + }, + { + "bbox": [ + 130, + 215, + 482, + 347 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 215, + 482, + 347 + ], + "type": "inline_equation", + "content": "x_{LD}^{2}" + }, + { + "bbox": [ + 130, + 215, + 482, + 347 + ], + "type": "text", + "content": ", we expect to prevent the supervised adapting signal from becoming overly dependent on the labeled samples. We imagine the effect of the defending samples throughout the adaptation process and depict the latent space formed gradually by the " + }, + { + "bbox": [ + 130, + 215, + 482, + 347 + ], + "type": "inline_equation", + "content": "x_{LD}" + }, + { + "bbox": [ + 130, + 215, + 482, + 347 + ], + "type": "text", + "content": " candidates as bold dashed circles in Figure 4 (top right)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 348, + 481, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 481, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 481, + 372 + ], + "type": "text", + "content": "Consequently, the overall loss consists of the sum of losses in Eq. (1) and a loss from our proposed method as," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 208, + 374, + 481, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 374, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 208, + 374, + 481, + 415 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\underbrace {\\mathcal {L} _ {\\text {s u p}} + \\mathcal {L} _ {\\text {u n s u p}}} _ {\\text {b a s e l i n e}} + \\underbrace {\\frac {1}{k \\cdot B} \\sum_ {b = 1} ^ {k \\cdot B} \\mathcal {H} \\left(y _ {L D} ^ {b} , f _ {\\theta} \\left(x _ {L D} ^ {b}\\right)\\right)} _ {\\text {r e t r i e v a l l a t e n t d e f e n d i n g}}. \\tag {2}", + "image_path": "a71243fab22d0375c9415599aec3f33919c1eae23ada8f73265908cfdf5817ec.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 421, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 482, + 517 + ], + "type": "text", + "content": "Importance of our method. Understanding the impact of NBF on adaptation performance is crucial. For example, naively adapting a model for a medical application using radiologist-provided feedback can actually cause performance degradation (shown in Table 5), potentially posing significant risks to patients. We propose a scalable and simple approach to solve the problem caused by NBF, which can not be addressed by existing methods. Given the practicality of the NBF problem and the scalability of our solution, we believe our work holds considerable potential for real-world applications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 528, + 231, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 528, + 231, + 541 + ], + "spans": [ + { + "bbox": [ + 131, + 528, + 231, + 541 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 548, + 266, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 266, + 560 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 266, + 560 + ], + "type": "text", + "content": "5.1 Experimental Setups" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 567, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 481, + 616 + ], + "type": "text", + "content": "Our approach is simple enough to seamlessly combine with existing SemiSDA algorithms and also be applied to diverse benchmarks. This section describes our experimental setup for natural image classification tasks and a real-world medical application. Details for semantic segmentation experiments are in Appendix D." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": "Baselines. We validate our approach by combining various state-of-the-art algorithms for SemiSDA [58] (e.g., CDAC [33] and AdaMatch [6]) and SemiSL [66,87] (e.g., FlexMatch [96] and FreeMatch [81]). Note that the SemiSL methods have been demonstrated to be strong SemiSDA learners [99], so we can consider them" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Song et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 145, + 114, + 470, + 164 + ], + "blocks": [ + { + "bbox": [ + 145, + 114, + 470, + 164 + ], + "lines": [ + { + "bbox": [ + 145, + 114, + 470, + 164 + ], + "spans": [ + { + "bbox": [ + 145, + 114, + 470, + 164 + ], + "type": "table", + "html": "
methodfeedbackaverage\\( \\mathrm{r} \\rightarrow \\mathrm{c} \\)\\( \\mathrm{r} \\rightarrow \\mathrm{p} \\)\\( \\mathrm{p} \\rightarrow \\mathrm{c} \\)\\( \\mathrm{c} \\rightarrow \\mathrm{s} \\)\\( \\mathrm{s} \\rightarrow \\mathrm{p} \\)\\( \\mathrm{r} \\rightarrow \\mathrm{s} \\)\\( \\mathrm{p} \\rightarrow \\mathrm{r} \\)
RerNetAdaMatch [6]RF67.666.668.568.560.369.258.781.5
NBF\\( {64.5}\\left( {+{3.1}}\\right) \\)64.366.165.656.965.654.278.9
w/ oursNBF\\( \\mathbf{{72.0}\\left( {+{7.5}}\\right) } \\)74.572.773.965.570.064.383.2
ViTAdaMatch [6]RF74.775.376.973.868.076.367.185.5
NBF\\( {73.7}\\left( {+{1.0}}\\right) \\)74.776.274.765.774.066.884.0
w/ oursNBF\\( \\mathbf{{75.9}\\left( {+{2.2}}\\right) } \\)76.977.877.868.576.668.385.1
", + "image_path": "913cf31ba67ae89e6447208b285243fc9f78116ff20179b00a14e23bd23f63dc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 163, + 482, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 163, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 163, + 482, + 240 + ], + "type": "text", + "content": "Table 1: Adaptation results on DomainNet-126. We simulate seven domain-shift scenarios (i.e., source " + }, + { + "bbox": [ + 130, + 163, + 482, + 240 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 163, + 482, + 240 + ], + "type": "text", + "content": " target). The model is pre-trained on the source domain and then adapted to a training set of the target domain. The results on the test set of the target domain are reported as the top-1 accuracy (\\%). DomainNet-126 [54,58] dataset includes real, painting, sketch, and clip-art domains. In this experiment, we assume that the 378 feedback samples (i.e., 3 labeled data per class) are obtained from users. A state-of-the-art SemiSDA method, AdaMatch [6], is used as a baseline." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 144, + 244, + 468, + 272 + ], + "blocks": [ + { + "bbox": [ + 144, + 244, + 468, + 272 + ], + "lines": [ + { + "bbox": [ + 144, + 244, + 468, + 272 + ], + "spans": [ + { + "bbox": [ + 144, + 244, + 468, + 272 + ], + "type": "table", + "html": "
methodfeedbackaveragea→ca→pa→rc→ac→pc→rp→ap→cp→rr→ar→cr→p
AdaMatch [6]RF70.955.480.475.965.781.574.665.958.778.468.861.584.3
NBF69.3 (+1.6)54.276.675.365.979.375.563.757.475.966.756.884.2
w/oursNBF73.8 (+4.5)62.281.079.768.885.478.667.761.779.569.064.188.2
", + "image_path": "4afba3dfdf9ade3d9156f63d1563bea8e2273195228ecc8f5c1218f92d06835b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 272, + 480, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 480, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 480, + 305 + ], + "type": "text", + "content": "Table 2: Adaptation results on OfficeHome. OfficeHome [75] dataset includes real, product, art, and clip-art domain. We assume that the 195 feedback samples (i.e., 3 labeled data per class) are obtained. AdaMatch [6] and ResNet-50 [23] are used." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 312, + 481, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 481, + 337 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 481, + 337 + ], + "type": "text", + "content": "as SemiSDA methods. For medical experiments, we use Pseudo-labeling [1] as a baseline since it is easily applicable to medical image adaptation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 339, + 481, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 481, + 374 + ], + "type": "text", + "content": "Datasets. We utilize natural image datasets containing multiple kinds of domains (e.g., real and painting). The datasets include DomainNet-126 [54,58] with 142k images of 126 classes, and OfficeHome [75] with 15K images of 65 classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 376, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 506 + ], + "type": "text", + "content": "To conduct medical experiments, we present a practical medical setting. We adopt the MIMIC-CXR-V2 dataset [27]. It assumes a multi-finding binary classification setup, where multiple radiographic findings, like Pneumonia and Atelectasis, can coexist in a single chest X-ray (CXR) sample. Thus, the model predicts the presence or absence (binary classes) of each individual finding. We simulate domain shift by using Posterior-Anterior (PA)-view data as the source and AP-view data as the target, capturing real-world variations in data acquisition. Typically, patients requiring an AP X-ray are those facing positioning challenges that prevent them from undergoing a PA X-ray. Therefore, this setup can be seen as a scenario where the target environment is the intensive care unit, which hospitalizes critically ill patients." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 507, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 579 + ], + "type": "text", + "content": "Following the recent SemiSDA [94] and SFDA [8] setups, we assume the model is pre-trained in the source domain and deployed in the target domain. Since the datasets above were not initially divided into training and test sets, we performed a random 8:2 split within each domain, designating them respectively for training and testing. The training set is used to adapt the model, while the test set is used to report the top-1 accuracy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "User feedback. Feedback given by users is modeled as annotations " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{y_{lb}^n : n \\in [1..N_{lb}]\\}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " on a small subset of the target's training set " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "D_t^{train}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ", while the remaining of them are used as unlabeled target data. In our experiments, we take into account two types of feedback: random feedback (RF) and negatively biased feedback (NBF). RF is the same setup of classical SemiSDA and SemiSL, where randomly selected samples from " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "D_t^{train}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " are used as small labeled set " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "X_t^{lb}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ". For NBF, we randomly select samples that are incorrectly predicted in " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "D_t^{train}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "NBF & RLD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 186, + 114, + 428, + 242 + ], + "blocks": [ + { + "bbox": [ + 186, + 114, + 428, + 242 + ], + "lines": [ + { + "bbox": [ + 186, + 114, + 428, + 242 + ], + "spans": [ + { + "bbox": [ + 186, + 114, + 428, + 242 + ], + "type": "table", + "html": "
feed, amount method378 (3 labeled data per class)630 (5 labeled data per class)
RFNBFw/oursRFNBFw/ours
Source model56.5
MME [58]69.568.4 (-1.1)70.8 (+2.4)71.270.1 (-1.1)72.5 (+2.4)
CDAC [33]68.364.6 (-3.7)73.2 (+8.6)71.768.1 (-3.6)74.9 (+6.8)
AdaMatch [6]67.664.5 (-3.1)72.0 (+7.5)70.967.7 (-3.2)74.3 (+6.6)
FixMatch [66]67.663.4 (-4.2)73.2 (+9.8)71.566.1 (-5.4)75.1 (+9.0)
UDA [87]69.264.9 (-4.3)73.4 (+8.5)72.968.8 (-4.1)75.3 (+6.5)
FlexMatch [96]73.371.4 (-1.9)74.7 (+3.3)75.373.9 (-1.4)76.0 (+2.1)
FreeMatch [81]73.872.0 (-1.8)74.8 (+2.8)75.674.4 (-1.2)76.1 (+1.7)
Fully supervised83.6
Source model64.5
MME [58]73.272.7 (-0.5)74.1 (+1.4)74.574.0 (-0.5)75.2 (+1.2)
CDAC [33]74.272.8 (-1.4)75.4 (+2.6)75.474.1 (-1.3)76.2 (+2.1)
AdaMatch [6]74.773.7 (-1.0)75.9 (+2.2)75.975.1 (-0.8)76.7 (+1.6)
FixMatch [66]74.673.0 (-1.6)75.6 (+2.6)75.774.3 (-1.4)76.5 (+2.2)
UDA [87]74.873.3 (-1.5)75.8 (+2.5)75.974.5 (-1.4)76.7 (+2.2)
FlexMatch [96]74.973.9 (-1.0)75.8 (+1.9)76.075.1 (-0.9)76.9 (+1.8)
FreeMatch [81]74.973.9 (-1.0)75.7 (+1.8)76.075.1 (-0.9)76.8 (+1.7)
Fully supervised85.4
", + "image_path": "170b81269cf22ac510e3f79814deeeb72cf49c9a02aa11f24411ae8d41de50b5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 186, + 289, + 428, + 361 + ], + "blocks": [ + { + "bbox": [ + 132, + 242, + 481, + 285 + ], + "lines": [ + { + "bbox": [ + 132, + 242, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 242, + 481, + 285 + ], + "type": "text", + "content": "Table 3: Comparisons on DomainNet-126. We evaluate our method by integrating it with SemiSDA and SemiSL methods. The average accuracy of seven domain-shift scenarios in Table 1 is reported. Source model represents the pre-trained model without adaptation. Fully supervised means the model is adapted with fully labeled target data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 186, + 289, + 428, + 361 + ], + "lines": [ + { + "bbox": [ + 186, + 289, + 428, + 361 + ], + "spans": [ + { + "bbox": [ + 186, + 289, + 428, + 361 + ], + "type": "table", + "html": "
feed, amount method195 (3 labeled data per class)325 (5 labeled data per class)
RFNBFw/oursRFNBFw/ours
Source model57.6
MME [58]71.270.2 (-1.0)73.4 (+3.2)73.573.1 (-0.4)75.6 (+2.5)
CDAC [33]71.269.0 (-2.2)74.3 (+5.3)73.572.3 (-1.2)75.7 (+3.4)
AdaMatch [6]70.969.3 (-1.6)73.8 (+4.5)73.472.7 (-0.7)75.5 (+2.8)
FixMatch [66]71.468.6 (-2.8)73.7 (+5.1)73.972.2 (-1.7)75.3 (+3.1)
UDA [87]72.269.5 (-2.7)74.1 (+4.6)74.473.0 (-1.4)76.0 (+3.0)
FlexMatch [96]73.772.1 (-1.6)74.7 (+2.6)75.974.9 (-1.0)76.6 (+1.7)
FreeMatch [81]74.072.7 (-1.3)74.8 (+2.1)75.875.0 (-0.8)76.6 (+1.6)
Fully supervised87.4
", + "image_path": "85d5b51d0e95f98724905a9eaacf44b18726daeacb236407af0c3cfcfdea2b99.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 361, + 480, + 383 + ], + "lines": [ + { + "bbox": [ + 132, + 361, + 480, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 480, + 383 + ], + "type": "text", + "content": "Table 4: Comparisons on OfficeHome. The average accuracy of twelve domain-shift scenarios in Table 2 is reported. ResNet-50 is used." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 393, + 481, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 481, + 453 + ], + "type": "text", + "content": "the source model (i.e., the pre-trained model before adaptation). Note that we focus on the impact of a biased label distribution within the same class, as shown in Figure 3, and thus take the same number of feedback for each class. Further discussion about the imbalance in the number of feedback between classes presented in [31, 49, 83] is provided in Appendix A.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 455, + 481, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 481, + 527 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 481, + 527 + ], + "type": "text", + "content": "Network architectures. We adopt commonly used networks, ResNet [23] and ViT [15] for natural image tasks and DenseNet [26] for a medical task. We employ ResNet-50 with the last classification layer comprising a weight normalization layer and a bottleneck layer, following previous works [8, 34] and use the ViT-Small (i.e., ViT-S) introduced in [80]. The DenseNet-121 is used, provided in TorchXrayVision [13], like existing medical works [32, 44]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 529, + 481, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 481, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 481, + 613 + ], + "type": "text", + "content": "Implementation details. We implement our framework by extending the publicly available USB [80] repository. Both pre-training and adaptation are conducted with a mini-batch size of 128 and the SGD optimizer. Diverse baselines for SemiSDA and SemiSL are used to compute the losses in Eq. (1). The hyperparameters for each baseline simply follow USB [80] or public code [33,58]. For all experiments, our approach uses the same hyper-parameters of the appended defending samples " + }, + { + "bbox": [ + 130, + 529, + 481, + 613 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 529, + 481, + 613 + ], + "type": "text", + "content": " and reliable filtering rate " + }, + { + "bbox": [ + 130, + 529, + 481, + 613 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 529, + 481, + 613 + ], + "type": "text", + "content": " as 3 and 0.4, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 625, + 228, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 625, + 228, + 637 + ], + "spans": [ + { + "bbox": [ + 132, + 625, + 228, + 637 + ], + "type": "text", + "content": "5.2 Main Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Natural image classification. Following recent DA works [8, 94], we conduct experiments on seven and twelve domain shift scenarios provided with the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 102 + ], + "type": "text", + "content": "J. Song et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 492, + 278 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 492, + 278 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 492, + 278 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 492, + 278 + ], + "type": "image", + "image_path": "1a0a93f6bc6c7b1f36a1f94d5389a287c4d30798c41abd0fe1ed981fc4076aeb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 279, + 482, + 335 + ], + "lines": [ + { + "bbox": [ + 130, + 279, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 482, + 335 + ], + "type": "text", + "content": "Table 5: Adaptation in a medical application. We use samples with PA-view as the source data and samples with AP-view as the target data in MIMIC-CXR-V2 dataset [27]. NBF-CE represents a scenario when NBF is composed of cases with confident errors. We use DensNet-121 [13, 26] and assume the 20 feedback for the absence and presence per finding." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 340, + 482, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 424 + ], + "type": "text", + "content": "DomainNet-126 and OfficeHome datasets, respectively. Table 1 and Table 2 show the results, where AdaMatch [6] is used as the baseline. We observe consistent results with Figure 3 even on large natural datasets: when simply applying the baseline under the NBF assumption, the adapted model shows inferior performance for most domain shifts than applying it under RF, e.g., " + }, + { + "bbox": [ + 130, + 340, + 482, + 424 + ], + "type": "inline_equation", + "content": "64.5 < 67.6" + }, + { + "bbox": [ + 130, + 340, + 482, + 424 + ], + "type": "text", + "content": ". Combining our approach with the baseline mitigates this issue and achieves a performance increase, e.g., " + }, + { + "bbox": [ + 130, + 340, + 482, + 424 + ], + "type": "inline_equation", + "content": "64.5 \\rightarrow 72.0" + }, + { + "bbox": [ + 130, + 340, + 482, + 424 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 425, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 520 + ], + "type": "text", + "content": "We also use other promising baselines and report the average accuracy of all domain shifts in Table 3 and Table 4 (all results can be found in Appendix F). While both feedback types bring performance improvement from the source model, lower performance is observed with NBF. Our method enables the baselines to not only address this problem but surpass performance under RF. The above results suggest that the biased distribution of labeled samples, which has been overlooked in previous SemiSDA works, is actually problematic, and our retrieval latent defending approach is effective." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 523, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 523, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 523, + 482, + 594 + ], + "type": "text", + "content": "Medical image diagnosis. Table 5 shows the results (bottom) and also depicts the effect of NBF (top center). We report the AUROC [7] for each finding following standard practice for measuring computer-aided-diagnosis model evaluation [32,44]. The baseline SemiSDA method under NBF exhibits inferior performance compared to one under RF, but this issue can be mitigated by combining our approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In addition, we propose an interesting and practical scenario named NBF with more confident errors (NBF-CE). In this scenario, we assume that a radiologist is likely to give feedback when the model makes confidently wrong predictions. Imagine that the model predicts a " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " likelihood of cancer in a CXR image, but the person actually has cancer. Such failure to detect potential patients early on can significantly reduce the patient's chances of survival, so a radiologist may" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "NBF & RLD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 478, + 140 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 478, + 140 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 478, + 140 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 478, + 140 + ], + "type": "table", + "html": "
labeling typefeed. amountRFNBFNBF w/oursENT [62]ENT w/ours
IAST [46,63]PA, 40 points55.353.0 (-2.3)56.3 (+3.3)53.556.0 (+2.5)
RIPU [85]PA, 40 points57.654.5 (-3.1)58.0 (+3.5)54.657.7 (+3.1)
", + "image_path": "6538a72ac99fe2db0591c6563e68ade36a2d6af2e08eeb186558787cc2dc7079.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 479, + 163 + ], + "lines": [ + { + "bbox": [ + 132, + 140, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 479, + 163 + ], + "type": "text", + "content": "Table 6: Adaptation on semantic segmentation. The GTA5 [56] " + }, + { + "bbox": [ + 132, + 140, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 132, + 140, + 479, + 163 + ], + "type": "text", + "content": " Cityscapes [14] setup is used [72]. The target performance of the source model is " + }, + { + "bbox": [ + 132, + 140, + 479, + 163 + ], + "type": "inline_equation", + "content": "36.6\\mathrm{mIoU}" + }, + { + "bbox": [ + 132, + 140, + 479, + 163 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 169, + 482, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 169, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 169, + 482, + 266 + ], + "type": "text", + "content": "provide feedback to the model. To simulate NBF-CE, we select samples where the source model most confidently predicts a finding to be absent " + }, + { + "bbox": [ + 130, + 169, + 482, + 266 + ], + "type": "inline_equation", + "content": "(\\hat{y} \\approx 0)" + }, + { + "bbox": [ + 130, + 169, + 482, + 266 + ], + "type": "text", + "content": " although it is clearly visible in the radiograph " + }, + { + "bbox": [ + 130, + 169, + 482, + 266 + ], + "type": "inline_equation", + "content": "(y = 1)" + }, + { + "bbox": [ + 130, + 169, + 482, + 266 + ], + "type": "text", + "content": ", and vice versa, i.e., samples of " + }, + { + "bbox": [ + 130, + 169, + 482, + 266 + ], + "type": "inline_equation", + "content": "\\hat{y} \\approx 1" + }, + { + "bbox": [ + 130, + 169, + 482, + 266 + ], + "type": "text", + "content": " but " + }, + { + "bbox": [ + 130, + 169, + 482, + 266 + ], + "type": "inline_equation", + "content": "y = 0" + }, + { + "bbox": [ + 130, + 169, + 482, + 266 + ], + "type": "text", + "content": ". Table 5 also shows the results under an NBF-CE scenario, where the model's adaptation performance is further reduced compared with NBF (0.7691 for " + }, + { + "bbox": [ + 130, + 169, + 482, + 266 + ], + "type": "inline_equation", + "content": "\\mathrm{NBF} \\to 0.7639" + }, + { + "bbox": [ + 130, + 169, + 482, + 266 + ], + "type": "text", + "content": " for NBF-CE). By combining our method, we observe performance improvements for both NBF variants, e.g., 0.7639 for " + }, + { + "bbox": [ + 130, + 169, + 482, + 266 + ], + "type": "inline_equation", + "content": "\\mathrm{NBF - CE} \\to 0.7875" + }, + { + "bbox": [ + 130, + 169, + 482, + 266 + ], + "type": "text", + "content": " with ours. We illustrate the hypothesized impact of our method in Table 5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 268, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 437 + ], + "type": "text", + "content": "Semantic segmentation. We evaluate the influence of NBF and our approach on a semantic segmentation task. We utilize the most common adaptation benchmark of GTA5 [56] to Cityscapes [14]. The baseline DA algorithms are used as IAST [46, 63] and RIPU [85] in a source-free scenario. We regard Pixel-based Annotation (PA) in which we assume 40 pixels per image like LabOR [63]. Table 6 shows results similar to those we observed in the classification and medical imaging tasks. The baselines under NBF exhibit inferior performance compared to those under RF (54.5 for " + }, + { + "bbox": [ + 130, + 268, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathrm{NBF} < 57.6" + }, + { + "bbox": [ + 130, + 268, + 482, + 437 + ], + "type": "text", + "content": " for RF), but this issue is addressed by combining our approach with them " + }, + { + "bbox": [ + 130, + 268, + 482, + 437 + ], + "type": "inline_equation", + "content": "(+3.5\\mathrm{mIoU})" + }, + { + "bbox": [ + 130, + 268, + 482, + 437 + ], + "type": "text", + "content": ". Although out of our scope (refer to Appendix A.1), we validate one active labeling strategy ENT [62], which assigns highly uncertain (i.e., probably misclassified) pixels as feedback. Consequently, the feedback instructed by ENT is biasedly distributed in a manner similar to NBF. ENT also causes unexpected results (54.6 for " + }, + { + "bbox": [ + 130, + 268, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathrm{ENT} < 57.6" + }, + { + "bbox": [ + 130, + 268, + 482, + 437 + ], + "type": "text", + "content": " for RF), and our approach alleviates this issue " + }, + { + "bbox": [ + 130, + 268, + 482, + 437 + ], + "type": "inline_equation", + "content": "(+3.1\\mathrm{mIoU})" + }, + { + "bbox": [ + 130, + 268, + 482, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 448, + 238, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 448, + 238, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 448, + 238, + 460 + ], + "type": "text", + "content": "5.3 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 468, + 481, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 481, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 481, + 528 + ], + "type": "text", + "content": "Positive vs. Negative feedback. We study the role of feedback on the adaptation results by varying feedback configurations. Let positively-provided feedback (PF) be obtained from samples that the source model correctly predicts, as opposed to negatively-provided feedback (NF). We adjust the ratio of PF:NF while keeping the total number of labeled samples constant, as shown in Figure 5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "content": "When using only FreeMatch (gray dot-dashed line), both biased feedback types (i.e., NBF and PBF) result in worse adaptation performance compared to balanced feedback for the baseline, e.g., 72.6 in 378:0 (PBF) " + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "inline_equation", + "content": "< 73.3" + }, + { + "bbox": [ + 130, + 528, + 482, + 624 + ], + "type": "text", + "content": " in 252:126. In contrast, when our method is applied (red line), NBF yields the best performance. PBF and NBF can be respectively regarded as contributing previously known knowledge of the model and new knowledge that complements model deficiencies. Hence, it may be natural that NBF, which actually encodes the model's mistakes, contributes to favorable adaptation results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 626, + 482, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 626, + 482, + 651 + ], + "spans": [ + { + "bbox": [ + 130, + 626, + 482, + 651 + ], + "type": "text", + "content": "Number of unlabeled samples in a mini-batch. Existing SemiSDA methods [6,81] typically set the ratio " + }, + { + "bbox": [ + 130, + 626, + 482, + 651 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 626, + 482, + 651 + ], + "type": "text", + "content": " between labeled and unlabeled samples in a" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "text", + "content": "J. Song et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 140, + 658, + 481, + 681 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 658, + 481, + 681 + ], + "spans": [ + { + "bbox": [ + 140, + 658, + 481, + 681 + ], + "type": "text", + "content": "If not specified, we use ResNet-50 and report the average accuracy (\\%) of seven domain shift scenarios in Table 1 for ablation studies." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 209, + 114, + 405, + 213 + ], + "blocks": [ + { + "bbox": [ + 209, + 114, + 405, + 213 + ], + "lines": [ + { + "bbox": [ + 209, + 114, + 405, + 213 + ], + "spans": [ + { + "bbox": [ + 209, + 114, + 405, + 213 + ], + "type": "image", + "image_path": "e86a614287b88fe342b0453b03d79c126febf09d14fab568454cd2058975d420.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 213, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 132, + 213, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 482, + 269 + ], + "type": "text", + "content": "Fig. 5: NBF leads to higher performance than PBF. We compare different user-feedback configurations when the total number of feedback is 378 (top) and 630 (bottom). Positive and negative feedback refers to feedback from correct and incorrect model predictions, respectively. We run three random seed experiments and describe the average performance and standard deviation in the parenthesis." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 211, + 271, + 402, + 321 + ], + "blocks": [ + { + "bbox": [ + 211, + 271, + 402, + 321 + ], + "lines": [ + { + "bbox": [ + 211, + 271, + 402, + 321 + ], + "spans": [ + { + "bbox": [ + 211, + 271, + 402, + 321 + ], + "type": "table", + "html": "
methodfeed.negatively biased feedback (NBF)
#xulb, #xLD, #xlb112, 0, 16112, 48, 1664, 48, 16
total batch size128176128
FreeMatch [81]36872.074.274.8 (+0.6)
AdaMatch [6]64.571.372.0 (+0.7)
FreeMatch [81]63074.475.576.1 (+0.6)
AdaMatch [6]67.773.474.3 (+0.9)
", + "image_path": "c42ab6948b203bc7b34e63cae005d44094813cc7110b4b2685df021d71e7a162.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 321, + 481, + 354 + ], + "lines": [ + { + "bbox": [ + 132, + 321, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 481, + 354 + ], + "type": "text", + "content": "Table 7: In the mini-batch, diminishing the number of unlabeled samples and adding our defending samples achieves better performance with our approach. We ablate them by changing the ratio " + }, + { + "bbox": [ + 132, + 321, + 481, + 354 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 132, + 321, + 481, + 354 + ], + "type": "text", + "content": " in Section 4.1, while keeping the size of labeled samples." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 360, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 468 + ], + "type": "text", + "content": "mini-batch to 1:7. However, we observe that adhering to this ratio is not optimal for our approach, as shown in Table 7. Our method shows better performance when the ratio is varied to 1:4, i.e., decreasing unlabeled sample sizes. This finding contradicts observations in several TTA works [28, 48, 67], where adaptation performance tends to increase with larger batch sizes. We speculate that it is beneficial to prioritize more reliable information, which refers to labeled data and our defending samples selected from the filtering-applied bank, during the adapting process. This result may be aligned with previous works for curriculum learning [40, 100] and adaptive thresholding [96]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "text", + "content": "Number of labeled data. We measure the impact of feedback size (number of labeled samples) in Figure 6. The results show that the inferior performance on NBF persists even with an increased amount of feedback (gray " + }, + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "text", + "content": " black line); however, our approach mitigates it and improves performance (black " + }, + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 471, + 482, + 567 + ], + "type": "text", + "content": " red line). We make an interesting observation that the performance gap between black and red lines becomes larger as the number of available feedback decreases. Since obtaining large feedback may be challenging in real-world applications, our method is expected to be more helpful in this practical case." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Data selecting strategy. We explore various strategies for selecting defending samples to balance the mini-batch, as shown in Table 8(top). The strategies include: in the " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{LD}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " candidate bank, (i) random selection regardless of the class of the labeled data, (ii) random selection in the same class as the labeled data (i.e., class-aware), (iii) selecting samples close to the cluster center obtained by k means clustering [21] and (iv) selecting samples with embedded features distant from the labeled data where cosine distance is used. While our approach consistently outperforms the baseline regardless of the chosen strategy, we empirically" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "NBF & RLD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 201, + 114, + 411, + 179 + ], + "blocks": [ + { + "bbox": [ + 201, + 114, + 411, + 179 + ], + "lines": [ + { + "bbox": [ + 201, + 114, + 411, + 179 + ], + "spans": [ + { + "bbox": [ + 201, + 114, + 411, + 179 + ], + "type": "image", + "image_path": "75da3a9bd27938ce43af20c7a5b7e3e3269010c907386b3157e2e3cfae87737d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 180, + 482, + 236 + ], + "lines": [ + { + "bbox": [ + 130, + 180, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 180, + 482, + 236 + ], + "type": "text", + "content": "Fig. 6: More reliable adaptation with NBF. In addition to Table 3, we conduct experiments with different amounts (1,3,5,10, and 15 labeled data per class) of feedback using FreeMatch [81]. The number of available feedback is likely to be small in practice. In this case, our method achieves large performance improvement, e.g., our method increases the baseline performance by " + }, + { + "bbox": [ + 130, + 180, + 482, + 236 + ], + "type": "inline_equation", + "content": "+4.9" + }, + { + "bbox": [ + 130, + 180, + 482, + 236 + ], + "type": "text", + "content": " when one feedback per class is available." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 176, + 237, + 436, + 292 + ], + "blocks": [ + { + "bbox": [ + 176, + 237, + 436, + 292 + ], + "lines": [ + { + "bbox": [ + 176, + 237, + 436, + 292 + ], + "spans": [ + { + "bbox": [ + 176, + 237, + 436, + 292 + ], + "type": "table", + "html": "
selection strategy\nclass-awarerandom\nxrandom\n✓kmeans\n✓cosine\n✓baseline\n-
FreeMatch [81]Res.74.174.874.674.072.0
FreeMatch [81]ViT.75.075.775.675.173.9
filtering rate0.20.40.60.8baseline only
FreeMatch [81]Res.74.574.874.373.772.0
FreeMatch [81]ViT.75.575.775.975.573.9
", + "image_path": "b2a4d413d7cbc7d3e15b3ba573cc5957d57b22d3dc3561596179a0f7a44dafa5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 293, + 481, + 315 + ], + "lines": [ + { + "bbox": [ + 130, + 293, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 481, + 315 + ], + "type": "text", + "content": "Table 8: We ablate a component of our approach with 378 feedback: " + }, + { + "bbox": [ + 130, + 293, + 481, + 315 + ], + "type": "inline_equation", + "content": "x_{LD}" + }, + { + "bbox": [ + 130, + 293, + 481, + 315 + ], + "type": "text", + "content": " selection strategy and filtering rate " + }, + { + "bbox": [ + 130, + 293, + 481, + 315 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 293, + 481, + 315 + ], + "type": "text", + "content": " for bank generation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 322, + 481, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 481, + 346 + ], + "type": "text", + "content": "find that strategy (ii) achieves the best performance. Therefore, we adopt this strategy for our proposed method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 346, + 481, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 481, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 481, + 371 + ], + "type": "text", + "content": "Further studies, such as extension to a TTA scenario, combining with SFDA methods and different feedback configurations, are presented in Appendix C." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 379, + 301, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 379, + 301, + 392 + ], + "spans": [ + { + "bbox": [ + 131, + 379, + 301, + 392 + ], + "type": "text", + "content": "6 Conclusion & Discussion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 400, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 555 + ], + "type": "text", + "content": "User feedback can play an integral part in adapting the practical ML product to the target environment. However, we have shown that naive adaptation using existing SemiSDA methods led to undesirable adaptation results. We explained this through the lens of Negatively-Biased Feedback (NBF). In this paper, we uncovered the unexpected results of NBF and presented a scalable solution, Retrieval Latent Defending. This method prevents the mini-batch from becoming overly dependent on labeled samples that may have a biased distribution within the overall target distribution. Under the diverse DA benchmarks, from the simulation study to the medical imaging task, we demonstrated the practical problem caused by NBF and the effectiveness of our approach by combining it with multiple SemiSDA baselines. We hope our efforts will inspire future DA works on leveraging user feedback to improve an ML model in the deployment environment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 482, + 629 + ], + "type": "text", + "content": "Broader impact. The proposed setup assumes that an ML product obtains feedback as a form of annotations (i.e., labeled data). In some cases, users can provide feedback in different forms, like thumbs up & down and rating of model prediction, or noise feedback whose information is different from the ground truth. Further research considering these points will pave the way for developing safer and more reliable adapting strategies." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Acknowledgment. We sincerely appreciate the abundant support provided by Lunit Inc., and we would like to thank Donggeun Yoo, Seonwook Park, and Sérgio Pereira for their valuable feedback." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "text", + "content": "J. Song et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 142, + 141, + 481, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 141, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 142, + 141, + 481, + 185 + ], + "type": "text", + "content": "1. Arazo, E., Ortego, D., Albert, P., O'Connor, N.E., McGuinness, K.: Pseudolabeling and confirmation bias in deep semi-supervised learning. In: International Joint Conference on Neural Networks (IJCNN) (2020) 6, 9, 11, 23, 24, 25, 26, 29, 30" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 142, + 186, + 481, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 186, + 481, + 231 + ], + "spans": [ + { + "bbox": [ + 142, + 186, + 481, + 231 + ], + "type": "text", + "content": "2. Basak, H., Yin, Z.: Semi-supervised domain adaptive medical image segmentation through consistency regularized disentangled contrastive learning. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (2023) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 142, + 232, + 481, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 232, + 481, + 253 + ], + "spans": [ + { + "bbox": [ + 142, + 232, + 481, + 253 + ], + "type": "text", + "content": "3. Baumeister, R.F., Bratslavsky, E., Finkenauer, C., Vohs, K.D.: Bad is stronger than good. Review of general psychology (2001) 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 142, + 254, + 481, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 254, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 142, + 254, + 481, + 286 + ], + "type": "text", + "content": "4. Berthelot, D., Carlini, N., Cubuk, E.D., Kurakin, A., Sohn, K., Zhang, H., Raffel, C.: Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. In: ICLR (2020) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 287, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 287, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 142, + 287, + 481, + 319 + ], + "type": "text", + "content": "5. Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.A.: Mixmatch: A holistic approach to semi-supervised learning. In: NeurIPS (2019) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 320, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 320, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 142, + 320, + 481, + 353 + ], + "type": "text", + "content": "6. Berthelot, D., Roelofs, R., Sohn, K., Carlini, N., Kurakin, A.: Adamatch: A unified approach to semi-supervised learning and domain adaptation. In: ICLR (2022) 1, 4, 7, 8, 9, 10, 11, 12, 13, 22, 28, 30, 32, 33, 34" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 354, + 481, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 354, + 481, + 375 + ], + "spans": [ + { + "bbox": [ + 142, + 354, + 481, + 375 + ], + "type": "text", + "content": "7. Bradley, A.P.: The use of the area under the roc curve in the evaluation of machine learning algorithms. Pattern recognition (1997) 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 376, + 481, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 376, + 481, + 398 + ], + "spans": [ + { + "bbox": [ + 142, + 376, + 481, + 398 + ], + "type": "text", + "content": "8. Chen, D., Wang, D., Darrell, T., Ebrahimi, S.: Contrastive test-time adaptation. In: CVPR (2022) 3, 9, 10, 24, 25, 27, 28, 30, 35" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 399, + 481, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 399, + 481, + 431 + ], + "spans": [ + { + "bbox": [ + 142, + 399, + 481, + 431 + ], + "type": "text", + "content": "9. Chen, H., Tao, R., Fan, Y., Wang, Y., Wang, J., Schiele, B., Xie, X., Raj, B., Savvides, M.: Softmatch: Addressing the quantity-quality trade-off in semi-supervised learning. In: ICLR (2023) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 432, + 481, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 481, + 463 + ], + "type": "text", + "content": "10. Chen, W., Lin, L., Yang, S., Xie, D., Pu, S., Zhuang, Y.: Self-supervised noisy label learning for source-free unsupervised domain adaptation. In: IROS (2022) 29" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 465, + 481, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 465, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 465, + 481, + 487 + ], + "type": "text", + "content": "11. Chen, X., Zhao, Z., Zhang, Y., Duan, M., Qi, D., Zhao, H.: Focalclick: Towards practical interactive image segmentation. In: CVPR (2022) 25" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 488, + 481, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 481, + 510 + ], + "type": "text", + "content": "12. Chen, Y., Tan, X., Zhao, B., Chen, Z., Song, R., Liang, J., Lu, X.: Boosting semi-supervised learning by exploiting all unlabeled data. In: CVPR (2023) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 510, + 481, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 481, + 554 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 481, + 554 + ], + "type": "text", + "content": "13. Cohen, J.P., Viviano, J.D., Bertin, P., Morrison, P., Torabian, P., Guerrera, M., Lungren, M.P., Chaudhari, A., Brooks, R., Hashir, M., et al.: Torchxrayvision: A library of chest x-ray datasets and models. In: International Conference on Medical Imaging with Deep Learning (2022) 10, 11, 28" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "text", + "content": "14. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR (2016) 12, 29" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 588, + 481, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 481, + 631 + ], + "type": "text", + "content": "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) 1, 10, 32, 33, 35" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "16. Fini, E., Astolfi, P., Alahari, K., Alameda-Pineda, X., Mairal, J., Nabi, M., Ricci, E.: Semi-supervised learning made simple with self-supervised clustering. In: CVPR (2023) 4" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "NBF & RLD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 117, + 479, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 117, + 479, + 138 + ], + "spans": [ + { + "bbox": [ + 138, + 117, + 479, + 138 + ], + "type": "text", + "content": "17. Ganin, Y., Lempitsky, V.: Unsupervised domain adaptation by backpropagation. In: ICML (2015) 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 480, + 171 + ], + "type": "text", + "content": "18. Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., Marchand, M., Lempitsky, V.: Domain-adversarial training of neural networks. The journal of machine learning research (2016) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 172, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 172, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 138, + 172, + 480, + 204 + ], + "type": "text", + "content": "19. Gong, T., Jeong, J., Kim, T., Kim, Y., Shin, J., Lee, S.J.: Robust continual test-time adaptation: Instance-aware bn and prediction-balanced memory. In: NeurIPS (2023) 24" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 205, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 480, + 247 + ], + "type": "text", + "content": "20. Harada, S., Bise, R., Araki, K., Yoshizawa, A., Terada, K., Kurata, M., Nakajima, N., Abe, H., Ushiku, T., Uchida, S.: Cluster-guided semi-supervised domain adaptation for imbalanced medical image classification. arXiv preprint arXiv:2303.01283 (2023) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 480, + 270 + ], + "type": "text", + "content": "21. Hartigan, J.A., Wong, M.A.: Algorithm as 136: A k-means clustering algorithm. Journal of the royal statistical society. series c (applied statistics) (1979) 13, 27" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "text", + "content": "22. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR (2020) 27" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 314 + ], + "type": "text", + "content": "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) 1, 9, 10, 32, 33, 35" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 480, + 335 + ], + "type": "text", + "content": "24. Higuchi, Y., Moritz, N., Roux, J.L., Hori, T.: Momentum pseudo-labeling for semi-supervised speech recognition. In: Interspeech (2021) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 336, + 480, + 357 + ], + "type": "text", + "content": "25. Huang, D., Li, J., Chen, W., Huang, J., Chai, Z., Li, G.: Divide and adapt: Active domain adaptation via customized learning. In: CVPR (2023) 4, 21, 22, 29" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 138, + 358, + 480, + 379 + ], + "type": "text", + "content": "26. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: CVPR (2017) 10, 11, 28" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 380, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 423 + ], + "type": "text", + "content": "27. Johnson, A.E., Pollard, T.J., Greenbaum, N.R., Lungren, M.P., Deng, C.y., Peng, Y., Lu, Z., Mark, R.G., Berkowitz, S.J., Horng, S.: Mimic-cxr-jpg, a large publicly available database of labeled chest radiographs. arXiv preprint arXiv:1901.07042 (2019) 9, 11" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 424, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 424, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 424, + 480, + 445 + ], + "type": "text", + "content": "28. Khurana, A., Paul, S., Rai, P., Biswas, S., Aggarwal, G.: Sita: Single image test-time adaptation. arXiv preprint arXiv:2112.02355 (2021) 13" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 446, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 446, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 138, + 446, + 480, + 478 + ], + "type": "text", + "content": "29. Kim, J., Hur, Y., Park, S., Yang, E., Hwang, S.J., Shin, J.: Distribution aligning refinery of pseudo-label for imbalanced semi-supervised learning. In: NeurIPS (2020) 22, 23" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 479, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 510 + ], + "type": "text", + "content": "30. Knox, W.B., Stone, P.: Tamer: Training an agent manually via evaluative reinforcement. In: IEEE international conference on development and learning (2008) 25" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 511, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 511, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 511, + 480, + 533 + ], + "type": "text", + "content": "31. Lee, H., Shin, S., Kim, H.: Abc: Auxiliary balanced classifier for class-imbalanced semi-supervised learning. In: NeurIPS (2021) 10, 22" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "type": "text", + "content": "32. Lenga, M., Schulz, H., Saalbach, A.: Continual learning for domain adaptation in chest x-ray classification. In: Medical Imaging with Deep Learning (2020) 10, 11" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 556, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 480, + 577 + ], + "type": "text", + "content": "33. Li, J., Li, G., Shi, Y., Yu, Y.: Cross-domain adaptive clustering for semi-supervised domain adaptation. In: CVPR (2021) 4, 8, 10, 32, 33, 34" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 578, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 609 + ], + "type": "text", + "content": "34. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: ICML (2020) 1, 3, 4, 10, 26, 27, 35" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 610, + 480, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 642 + ], + "type": "text", + "content": "35. Litrico, M., Del Bue, A., Morerio, P.: Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation. In: CVPR (2023) 3, 26, 27, 29, 30, 35" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "36. Liu, X., Xing, F., Shusharina, N., Lim, R., Jay Kuo, C.C., El Fakhri, G., Woo, J.: Act: Semi-supervised domain-adaptive medical image segmentation with asym" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "J. Song et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 153, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 153, + 116, + 481, + 138 + ], + "type": "text", + "content": "metric co-training. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (2022) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 481, + 172 + ], + "type": "text", + "content": "37. Liu, X., Yoo, C., Xing, F., Oh, H., El Fakhri, G., Kang, J.W., Woo, J., et al.: Deep unsupervised domain adaptation: A review of recent advances and perspectives. APSIPA Transactions on Signal and Information Processing (2022) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 481, + 195 + ], + "type": "text", + "content": "38. Liu, Y., Zhang, W., Wang, J.: Source-free domain adaptation for semantic segmentation. In: CVPR (2021) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 196, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 196, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 138, + 196, + 481, + 228 + ], + "type": "text", + "content": "39. Liu, Y., Kothari, P., van Delft, B., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? In: NeurIPS (2021) 24, 30" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 481, + 251 + ], + "type": "text", + "content": "40. Liu, Z., Miao, Z., Pan, X., Zhan, X., Lin, D., Yu, S.X., Gong, B.: Open compound domain adaptation. In: CVPR (2020) 13" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 251, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 481, + 274 + ], + "type": "text", + "content": "41. Long, M., Zhu, H., Wang, J., Jordan, M.I.: Deep transfer learning with joint adaptation networks. In: ICML (2017) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 274, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 481, + 308 + ], + "type": "text", + "content": "42. MacGlashan, J., Ho, M.K., Loftin, R., Peng, B., Wang, G., Roberts, D.L., Taylor, M.E., Littman, M.L.: Interactive learning from policy-dependent human feedback. In: International conference on machine learning (2017) 25" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 308, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 308, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 481, + 352 + ], + "type": "text", + "content": "43. Madani, A., Moradi, M., Karargyris, A., Syeda-Mahmood, T.: Semi-supervised learning with generative adversarial networks for chest x-ray classification with ability of data domain adaptation. In: International symposium on biomedical imaging (2018) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 352, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 352, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 352, + 481, + 385 + ], + "type": "text", + "content": "44. Mahapatra, D., Korevaar, S., Bozortabar, B., Tennakoon, R.: Unsupervised domain adaptation using feature disentanglement and gcns for medical image classification. In: ECCV (2022) 10, 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 385, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 385, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 385, + 481, + 407 + ], + "type": "text", + "content": "45. Mei, K., Zhu, C., Zou, J., Zhang, S.: Instance adaptive self-training for unsupervised domain adaptation. In: ECCV (2020) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 407, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 481, + 430 + ], + "type": "text", + "content": "46. Mei, K., Zhu, C., Zou, J., Zhang, S.: Instance adaptive self-training for unsupervised domain adaptation. In: ECCV (2020) 12, 29" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 430, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 430, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 430, + 481, + 453 + ], + "type": "text", + "content": "47. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: ICML (2022) 24, 30" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 453, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 453, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 453, + 481, + 475 + ], + "type": "text", + "content": "48. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: ICLR (2023) 13" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 475, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 475, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 475, + 481, + 498 + ], + "type": "text", + "content": "49. Oh, Y., Kim, D.J., Kweon, I.S.: Daso: Distribution-aware semantics-oriented pseudo-label for imbalanced semi-supervised learning. In: CVPR (2022) 10, 22" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 498, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 481, + 531 + ], + "type": "text", + "content": "50. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. In: NeurIPS (2022) 25" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 531, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 531, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 531, + 481, + 553 + ], + "type": "text", + "content": "51. Pan, Y., Yao, T., Li, Y., Wang, Y., Ngo, C.W., Mei, T.: Transferrable prototypical networks for unsupervised domain adaptation. In: CVPR (2019) 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 554, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 481, + 598 + ], + "type": "text", + "content": "52. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, e.a.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019), https://github.com/pytorch/vision/blob/main/torchvision/models/resnet.py 28" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 599, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 481, + 643 + ], + "type": "text", + "content": "53. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournaepau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine learning in Python. Journal of Machine Learning Research (2011) 5, 25, 26" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "54. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: ICCV (2019) 9" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "NBF & RLD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 479, + 148 + ], + "type": "text", + "content": "55. Prabhu, V., Chandrasekaran, A., Saenko, K., Hoffman, J.: Active domain adaptation via clustering uncertainty-weighted embeddings. In: ICCV (2021) 4, 21, 22, 29" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 137, + 149, + 480, + 171 + ], + "type": "text", + "content": "56. Richter, S.R., Vineet, V., Roth, S., Koltun, V.: Playing for data: Ground truth from computer games. In: ECCV (2016) 12, 29" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 171, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 171, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 137, + 171, + 480, + 192 + ], + "type": "text", + "content": "57. Rozin, P., Royzman, E.B.: Negativity bias, negativity dominance, and contagion. Personality and social psychology review (2001) 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 192, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 192, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 138, + 192, + 480, + 213 + ], + "type": "text", + "content": "58. Saito, K., Kim, D., Sclaroff, S., Darrell, T., Saenko, K.: Semi-supervised domain adaptation via minimax entropy. In: ICCV (2019) 1, 3, 4, 8, 9, 10, 28, 32, 33, 34" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 214, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 214, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 138, + 214, + 480, + 235 + ], + "type": "text", + "content": "59. Saito, K., Watanabe, K., Ushiku, Y., Harada, T.: Maximum classifier discrepancy for unsupervised domain adaptation. In: CVPR (2018) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 235, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 235, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 235, + 480, + 266 + ], + "type": "text", + "content": "60. Sajjadi, M., Javanmardi, M., Tasdizen, T.: Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In: NeurIPS (2016) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 267, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 480, + 300 + ], + "type": "text", + "content": "61. Schulman, J., Zoph, B., Kim, C., Hilton, J., Menick, J., Weng, J., Uribe, J.F.C., Fedus, L., Metz, L., Pokorny, M., et al.: Chatgpt: Optimizing language models for dialogue. OpenAI blog (2022) 25" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 300, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 480, + 320 + ], + "type": "text", + "content": "62. Shen, Y., Yun, H., Lipton, Z.C., Kronrod, Y., Anandkumar, A.: Deep active learning for named entity recognition. In: ICLR (2017) 12, 22" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 321, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 480, + 352 + ], + "type": "text", + "content": "63. Shin, I., Kim, D.J., Cho, J.W., Woo, S., Park, K., Kweon, I.S.: Labor: Labeling only if required for domain adaptive semantic segmentation. In: ICCV (2021) 12, 29" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 353, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 480, + 374 + ], + "type": "text", + "content": "64. Sofiuk, K., Petrov, I., Barinova, O., Konushin, A.: f-brs: Rethinking backpropagating refinement for interactive segmentation. In: CVPR (2020) 25" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 480, + 407 + ], + "type": "text", + "content": "65. Sofiuk, K., Petrov, I.A., Konushin, A.: Reviving iterative training with mask guidance for interactive segmentation. In: IEEE International Conference on Image Processing (ICIP) (2022) 25" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 407, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 480, + 439 + ], + "type": "text", + "content": "66. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In: NeurIPS (2020) 1, 4, 8, 10, 28, 32, 33" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 439, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 439, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 480, + 460 + ], + "type": "text", + "content": "67. Song, J., Lee, J., Kweon, I.S., Choi, S.: Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. In: CVPR (2023) 1, 13, 24, 30" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 460, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 493 + ], + "type": "text", + "content": "68. Song, J., Park, K., Shin, I., Woo, S., Zhang, C., Kweon, I.S.: Test-time adaptation in the dynamic world with compound domain knowledge management. IEEE Robotics and Automation Letters (2023) 24" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 493, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 480, + 525 + ], + "type": "text", + "content": "69. Stiennon, N., Ouyang, L., Wu, J., Ziegler, D., Lowe, R., Voss, C., Radford, A., Amodei, D., Christiano, P.F.: Learning to summarize with human feedback. In: NeurIPS (2020) 25" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 525, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 547 + ], + "type": "text", + "content": "70. Sun, B., Saenko, K.: Deep coral: Correlation alignment for deep domain adaptation. In: ECCV Workshops (2016) 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 547, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 480, + 579 + ], + "type": "text", + "content": "71. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 25" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 579, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 480, + 610 + ], + "type": "text", + "content": "72. Tsai, Y.H., Hung, W.C., Schulter, S., Sohn, K., Yang, M.H., Chandraker, M.: Learning to adapt structured output space for semantic segmentation. In: CVPR (2018) 1, 3, 12" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "text", + "content": "73. Tsai, Y.H., Hung, W.C., Schulter, S., Sohn, K., Yang, M.H., Chandraker, M.: Learning to adapt structured output space for semantic segmentation. In: CVPR (2018) 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "74. Van Engelen, J.E., Hoos, H.H.: A survey on semi-supervised learning. Machine learning (2020) 4" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "J. Song et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 138 + ], + "type": "text", + "content": "75. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: CVPR (2017) 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 170 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 170 + ], + "type": "text", + "content": "76. Vu, T.H., Jain, H., Bucher, M., Cord, M., Pérez, P.: Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In: CVPR (2019) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 171, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 171, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 138, + 171, + 480, + 193 + ], + "type": "text", + "content": "77. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: ICLR (2021) 3, 24, 30" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 194, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 480, + 215 + ], + "type": "text", + "content": "78. Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing (2018) 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 217, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 480, + 236 + ], + "type": "text", + "content": "79. Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: CVPR (2022) 30" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 238, + 480, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 480, + 269 + ], + "type": "text", + "content": "80. Wang, Y., Chen, H., Fan, Y., Sun, W., Tao, R., Hou, W., Wang, R., Yang, L., Zhou, Z., Guo, L.Z., et al.: Usb: A unified semi-supervised learning benchmark for classification. In: NeurIPS (2022) 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 271, + 480, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 312 + ], + "type": "text", + "content": "81. Wang, Y., Chen, H., Heng, Q., Hou, W., Fan, Y., Wu, Z., Wang, J., Savvides, M., Shinozaki, T., Raj, B., et al.: Freematch: Self-adaptive thresholding for semi-supervised learning. In: ICLR (2023) 4, 7, 8, 10, 12, 13, 14, 24, 25, 27, 28, 30, 32, 33, 34" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 480, + 335 + ], + "type": "text", + "content": "82. Warnell, G., Waytowich, N., Lawhern, V., Stone, P.: Deep tamer: Interactive agent shaping in high-dimensional state spaces. In: AAAI (2018) 25" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 337, + 480, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 480, + 368 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 480, + 368 + ], + "type": "text", + "content": "83. Wei, C., Sohn, K., Mellina, C., Yuille, A., Yang, F.: Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In: CVPR (2021) 10, 22, 23, 24, 29" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 138, + 369, + 480, + 402 + ], + "type": "text", + "content": "84. Wirth, C., Akhour, R., Neumann, G., Fünkranz, J., et al.: A survey of preference-based reinforcement learning methods. Journal of Machine Learning Research (2017) 25" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 403, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 434 + ], + "type": "text", + "content": "85. Xie, B., Yuan, L., Li, S., Liu, C.H., Cheng, X.: Towards fewer annotations: Active learning via region impurity and prediction uncertainty for domain adaptive semantic segmentation. In: CVPR (2022) 12, 29" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 435, + 480, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 480, + 466 + ], + "type": "text", + "content": "86. Xie, M., Li, Y., Wang, Y., Luo, Z., Gan, Z., Sun, Z., Chi, M., Wang, C., Wang, P.: Learning distinctive margin toward active domain adaptation. In: CVPR (2022) 4, 21" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 490 + ], + "type": "text", + "content": "87. Xie, Q., Dai, Z., Hovy, E., Luong, T., Le, Q.: Unsupervised data augmentation for consistency training. In: NeurIPS (2020) 4, 8, 10, 32, 33, 34" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "type": "text", + "content": "88. Xu, Y., Shang, L., Ye, J., Qian, Q., Li, Y.F., Sun, B., Li, H., Jin, R.: Dash: Semi-supervised learning with dynamic thresholding. In: ICML (2021) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "text", + "content": "89. Yan, Z., Wu, Y., Li, G., Qin, Y., Han, X., Cui, S.: Multi-level consistency learning for semi-supervised domain adaptation. In: IJCAI (2022) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 566 + ], + "type": "text", + "content": "90. Yang, L., Wang, Y., Gao, M., Shrivastava, A., Weinberger, K.Q., Chao, W.L., Lim, S.N.: Deep co-training with task decomposition for semi-supervised domain adaptation. In: ICCV (2021) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "text", + "content": "91. Yang, S., Jui, S., van de Weijer, J., et al.: Attracting and dispersing: A simple approach for source-free domain adaptation. In: NeurIPS (2022) 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "type": "text", + "content": "92. Yang, S., Wang, Y., Van De Weijer, J., Herranz, L., Jui, S.: Generalized source-free domain adaptation. In: ICCV (2021) 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 611, + 480, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 642 + ], + "type": "text", + "content": "93. Yang, S., van de Weijer, J., Herranz, L., Jui, S., et al.: Exploiting the intrinsic neighborhood structure for source-free domain adaptation. In: NeurIPS (2021) 3, 27, 28, 35" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "94. Yu, Y.C., Lin, H.T.: Semi-supervised domain adaptation with source label adaptation. In: CVPR (2023) 4, 9, 10" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "NBF & RLD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 116, + 481, + 304 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 136, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 481, + 138 + ], + "type": "text", + "content": "95. Yu, Z., Li, J., Du, Z., Zhu, L., Shen, H.T.: A comprehensive survey on source-free domain adaptation. arXiv preprint arXiv:2302.11803 (2023) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 481, + 171 + ], + "type": "text", + "content": "96. Zhang, B., Wang, Y., Hou, W., Wu, H., Wang, J., Okumura, M., Shinozaki, T.: Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In: NeurIPS (2021) 7, 8, 10, 13, 32, 33, 34" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 172, + 481, + 205 + ], + "type": "text", + "content": "97. Zhang, C., Miech, A., Shen, J., Alayrac, J.B., Luc, P.: Making the most of what you have: Adapting pre-trained visual language models in the low-data regime. arXiv preprint arXiv:2305.02297 (2023) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 481, + 226 + ], + "type": "text", + "content": "98. Zhang, W., Ouyang, W., Li, W., Xu, D.: Collaborative and adversarial network for unsupervised domain adaptation. In: CVPR (2018) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 481, + 258 + ], + "type": "text", + "content": "99. Zhang, Y., Zhang, H., Deng, B., Li, S., Jia, K., Zhang, L.: Semi-supervised models are strong unsupervised domain adaptation learners. arXiv preprint arXiv:2106.00417 (2021) 1, 4, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 134, + 259, + 480, + 281 + ], + "type": "text", + "content": "100. Zhang, Y., David, P., Gong, B.: Curriculum domain adaptation for semantic segmentation of urban scenes. In: ICCV (2017) 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 281, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 281, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 134, + 281, + 480, + 304 + ], + "type": "text", + "content": "101. Zhang, Y., Wang, Z., He, W.: Class relationship embedded learning for source-free unsupervised domain adaptation. In: CVPR (2023) 3" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "text", + "content": "J. Song et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Isomorphic Pruning for Vision Models/bf292aec-a3cb-46ef-92f7-27df67a69e35_content_list.json b/2024/Isomorphic Pruning for Vision Models/bf292aec-a3cb-46ef-92f7-27df67a69e35_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..04896d85cdfe5b56b525a4d438415643c63dc824 --- /dev/null +++ b/2024/Isomorphic Pruning for Vision Models/bf292aec-a3cb-46ef-92f7-27df67a69e35_content_list.json @@ -0,0 +1,1596 @@ +[ + { + "type": "text", + "text": "Isomorphic Pruning for Vision Models", + "text_level": 1, + "bbox": [ + 281, + 141, + 722, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gongfan Fang $^{1}$ , Xinyin Ma $^{1}$ , Michael Bi Mi $^{2}$ , and Xinchao Wang $^{1*}$", + "bbox": [ + 236, + 188, + 767, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National University of Singapore", + "bbox": [ + 383, + 215, + 617, + 231 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Huawei Technologies Ltd.", + "bbox": [ + 406, + 231, + 594, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gongfan@u.nus.edu，xinchao@nus.edu.sg", + "bbox": [ + 357, + 246, + 645, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Structured pruning reduces the computational overhead of deep neural networks by removing redundant sub-structures. However, assessing the relative importance of different sub-structures remains a significant challenge, particularly in advanced vision models featuring novel mechanisms and architectures like self-attention, depth-wise convolutions, or residual connections. These heterogeneous substructures usually exhibit diverged parameter scales, weight distributions, and computational topology, introducing considerable difficulty to importance comparison. To overcome this, we present Isomorphic Pruning, a simple approach that demonstrates effectiveness across a range of network architectures such as Vision Transformers and CNNs, and delivers competitive performance across different model sizes. Isomorphic Pruning originates from an observation that, when evaluated under a pre-defined importance criterion, heterogeneous sub-structures demonstrate significant divergence in their importance distribution, as opposed to isomorphic structures that present similar importance patterns. This inspires us to perform isolated ranking and comparison on different types of sub-structures for more reliable pruning. Our empirical results on ImageNet-1K demonstrate that Isomorphic Pruning surpasses several pruning baselines dedicatedly designed for Transformers or CNNs. For instance, we improve the accuracy of DeiT-Tiny from $74.52\\%$ to $77.50\\%$ by pruning an off-the-shelf DeiT-Base model. And for ConvNext-Tiny, we enhanced performance from $82.06\\%$ to $82.18\\%$ , while reducing the number of parameters and memory usage. Code is available at https://github.com/VainF/Isomorphic-Pruning.", + "bbox": [ + 259, + 287, + 743, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Network Pruning $\\cdot$ Vision Transformers $\\cdot$ CNNs", + "bbox": [ + 261, + 646, + 669, + 661 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 683, + 375, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep Neural Networks are typically over-parameterized, containing numerous redundant sub-structures. Such redundancy can be effectively eliminated via structured pruning, without significantly sacrificing the network's performance [16, 22, 25]. Pruning usually follows a \"ranking-to-prune\" framework: To facilitate compression while retaining the learned capabilities of networks, a carefully-designed criterion is utilized to assess the importance score of different structures [11, 22, 40], followed by a ranking process to identify and remove those", + "bbox": [ + 212, + 710, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding Author.", + "bbox": [ + 217, + 824, + 388, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5d6fc412fc9fe29a137bf1527c879ab991715f988911296aaef04fe3e8d941bd.jpg", + "image_caption": [ + "Fig. 1: ImageNet Top-1 Accuracy vs. Multiply-Accumulate Operations (MACs) of pruned DeiT [50] and pruned ConvNext [33]. The pruned models, marked as “★” have comparable or better latency yet superior performance compared to scratch training counterparts highlighted by “○”. The MACs are in log scale for better visualization" + ], + "image_footnote": [], + "bbox": [ + 253, + 148, + 743, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "relatively unimportant ones. After pruning, the compressed model is fine-tuned to recover its performance [11, 18].", + "bbox": [ + 212, + 443, + 782, + 473 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, ranking-based structured pruning presupposes an underlying assumption that the parameters from different structures are directly comparable. Classic deep neural networks, such as VGG [46], are built by stacking similar convolutional layers or blocks where the neuron connections share a similar computational topology. Under the circumstances, a straightforward global ranking is expected to reveal their relative importance and remove unimportant parameters [11, 22]. However, with recent innovation in network structure design, advanced vision models like Vision Transformers [10, 31, 50] and ConvNext [33] incorporate more sophisticated designs, such as complicated residual connections [19, 24], attention mechanisms [52] and depth-wise convolutions [6]. These novel structures increase the network's internal complexity and heterogeneity, challenging the reliability of pruning importance criteria. For instance, a vanilla Vision Transformer [10] is composed of MLP, Multi-head Attention, and skip connections. Even in the same Multi-head attention layer, there are two pruning schemes to removing either entire attention heads [37, 53] or the within-head dimensions [58, 68]. These structures vary by their computational topologies and, as will be illustrated in the experiments, tend to exhibit diverged importance distributions. In this case, directly comparing the importance of these heterogeneous might be problematic.", + "bbox": [ + 212, + 474, + 785, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the problem, we propose a simple pruning approach, termed Isomorphic Pruning, to alleviate the incomparability issue arising from heterogeneous sub-structures. The essence of this method lies in decomposing a network into multiple isomorphic groups, as illustrated in Figure 2 (d), based on the computational topology of the sub-structures. Each isomorphic group contains", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "G. Fang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c60658c4438fb36ae5860712318177d6e42191850ac2abd987ea02b23ec8875d.jpg", + "image_caption": [ + "(a) Network" + ], + "image_footnote": [], + "bbox": [ + 256, + 160, + 313, + 231 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1e43295c1daf7b27db9057b7ee834b9531f436bccda89115995b1721c48a6d20.jpg", + "image_caption": [ + "(b) Local Pruning" + ], + "image_footnote": [], + "bbox": [ + 362, + 160, + 449, + 229 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/89d548bb24841bb9486b5878797ae9f1fe8eae9eb5d8700b20c5defcc3a35728.jpg", + "image_caption": [ + "(c) Global Pruning" + ], + "image_footnote": [], + "bbox": [ + 478, + 160, + 566, + 231 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6dc360afa098946100a2127d209eb63a33d41a02f1042c44a1dc61033f2e96b4.jpg", + "image_caption": [ + "(d) Isomorphic Pruning", + "Fig. 2: For a pre-trained network (a), we show three pruning strategies: (b) Local Pruning that compares parameter importance within current layers; (c) Global Pruning that performs global ranking for all parameters; (d) the proposed Isomorphic pruning that groups parameters by the computational topology and applies importance ranking within groups. In each group, the importance distributions are more comparable. Details about the distribution can be found in Fig. 4 of the experiments section." + ], + "image_footnote": [], + "bbox": [ + 593, + 159, + 754, + 232 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "elements with the same structure and tends to present a similar distribution of importance, making it more rational to compare their relative importance.", + "bbox": [ + 212, + 359, + 787, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To validate the effectiveness of our method, we conducted experiments on a variety of visual models, including DeiT [50], ConvNext [33], ResNet [19], and MobileNet-V2 [44]. Pruning results, as shown in Figure 1, on ImageNet-1K [8] demonstrate that Isomorphic Pruning achieves comparable or even superior performance relative to pruning algorithms specifically designed for CNNs or transformers. For instance, by pruning a pre-trained large DeiT model, we enhanced the performance of the DeiT-T model from $74.52\\%$ to $77.50\\%$ , when compared to the scratch training one. Similarly, our algorithm also yields satisfactory results on CNNs. We constructed compact ConvNexts and ResNets via pruning, which has fewer MACs and parameters while surpassing a conventionally trained model. Additionally, we also studied the real-world latency and memory footprint of the compressed models on both GPU and CPU devices. Our findings indicate that models pruned with Isomorphic Pruning achieve significant actual speed-ups and are competitive to carefully designed efficient architectures.", + "bbox": [ + 212, + 390, + 787, + 601 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, Our contribution lies in a practical and general pruning approach, which can handle various vision models with novel mechanisms and architectures. The pruned models show superior performance to several pruning competitors on ImageNet-1K benchmark.", + "bbox": [ + 212, + 601, + 787, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 214, + 685, + 397, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Network Pruning. Deep neural networks are typically over-parameterized and can be pruned without significant performance drop [16, 25]. The exploration of network pruning can be dated back to early works [1, 17, 25] about neural networks. Over the past few years, a large number of works have been conducted on convolutional networks to verify the effectiveness of pruning, which can be roughly categorized into unstructured [16, 29, 35] and structured pruning [5, 11, 18, 22, 56]. Unstructured pruning compresses a pre-trained model by forcing partial weights to be zero, leading sparse networks [29]. To accelerate", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Isomorphic Pruning for Vision Models", + "bbox": [ + 473, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the sparse model, special patterns and devices will be used such as N:M sparsity [12, 42, 48]. Structured pruning provides a hardware-friendly approach by removing sub-structures such as channels [11] or blocks [14] in CNNs, which improves the model efficiency without dedicated accelerators. In recent years, a large number of works adapted the idea of network pruning to vision transformers. It demonstrated that even the vanilla vision transformers can be squeezed for a better efficiency-performance trade-off [2, 4, 45, 47, 58, 60–62, 64, 65, 68, 70]. In the literature, transformer pruning has been studied from several perspectives, such as width reduction [70], depth pruning [61], attention head removal [58], sparse training [4] and architecture transformation [18].", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pruning by Global Ranking. In the context of network pruning, global pruning is recognized as a pivotal topic for exploration. This method entails a comprehensive ranking of all network layers, assigning adaptive pruning ratios to each layer according to the ranks [5,11,22,58]. Different layers within a network often exhibit varying levels of redundancy. To this end, the advantage of global pruning lies in its ability to create a more parameter-efficient, non-uniformly compressed model [58]. However, a key challenge remains in the ranking: how to ensure the consistency and comparability across different layers? A widely-used solution for this issue is normalizing scores on a per-layer basis to even out the importance distribution across the network [30]. Alternatively, some approaches use sparse training to learn an effective target structure [11], which might be costly due to the additional optimization. In this work, we introduce a more straightforward approach termed isomorphic pruning to facilitate a reliable ranking for pruning.", + "bbox": [ + 212, + 321, + 787, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 544, + 330, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The core idea of Isomorphic Pruning is straightforward: we group sub-structures in a network by their network topology and perform importance ranking within groups. This imposes two questions 1) how to identify removable sub-structures in networks and 2) how to compare the topology graph for grouping and pruning.", + "bbox": [ + 212, + 578, + 787, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Sub-structure in Network", + "text_level": 1, + "bbox": [ + 215, + 662, + 475, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Structured Pruning aims to remove sub-structures in networks, such as the dimensions of hidden states or features. To make the formulation clear, we primarily concentrate on stacked fully-connected networks, although the principles established herein can be readily extended to CNNs and transformers. Initially, we consider a single linear layer which is parameterized by a matrix $W \\in \\mathbb{R}^{m \\times n}$ . To articulate the pruning process, we introduce a pruning function $\\mathrm{g}(W,d,k)$ defined as follows:", + "bbox": [ + 212, + 691, + 787, + 796 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {w} = \\mathrm {g} (W, d, k) = \\left\\{ \\begin{array}{l l} \\mathrm {g} (W, 0, k) = W _ {[ k,: ]} \\in \\mathbb {R} ^ {1 \\times n}, & \\text {i f p r u n i n g o u t p u t d i m e n s i o n ,} \\\\ \\mathrm {g} (W, 1, k) = W _ {[:, k ]} \\in \\mathbb {R} ^ {m \\times 1}, & \\text {i f p r u n i n g i n p u t d i m e n s i o n ,} \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 806, + 785, + 844 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "G. Fang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{w}$ refers to the weight column or row slated for removal, acquired via a slicing operation denoted by subscript $[:,k]$ or $[k,:]$ . The pruning function $\\mathbf{g}(W,d,k)$ is defined by three parameters, the parameter matrix $W$ , the axis $d\\in \\{0,1\\}$ indicating the dimension to be compressed, and the index $k$ identifying the $k$ -th rows or columns in $W$ . In the context of a linear layer represented by a 2-D matrix $W\\in \\mathbb{R}^{m\\times n}$ , the axis $d$ can assume values of either 0 or 1, with $k$ having a permissible range of $[0,m]$ and $[0,n]$ , respectively. This formulation encapsulates the pruning of a single layer; to extend it to the entire network, we also need to take the dependency [5,11,30] into account, where parameters from different layers are coupled and should be pruned simultaneously. As depicted in Figure 3, for two adjacent linear layers parameterized by $W_{1}$ and $W_{2}$ , the output dimension of the former is integrally coupled to the input dimension of the latter, thereby constituting a minimal sub-structure eligible for removal. If we treat each pruning operation $\\mathbf{g}(W,d,k)$ as vertices defined by a triplet $(W,d,k)$ and the dependency as edges, then it is natural to model sub-structures as graphs.", + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Graph Modeling of Sub-structures", + "text_level": 1, + "bbox": [ + 214, + 395, + 545, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we delve deeper into the methodology for decomposing a network into several independent and minimal sub-structures in pruning, and model them as graphs. We can begin with the MLP network as illustrated in Figure 3, where the activation functions are omitted. The network comprises a sequence of 2-D parameter matrices $\\{W_1, W_2, \\dots, W_L\\}$ , with $L$ denoting the network depth. Initially, all parameters are not assigned to any sub-structure. We randomly select one parameter $W$ as the seed to find its associated substructures and build the corresponding graph. For instance, selecting parameter $W_1$ and setting the axis $d = 0$ allows us to establish an initial graph $G = (V = \\{(W, 0, k)\\}, E = \\{\\})$ , focusing on the elimination of $k$ -th the output dimension in $W_1$ . Then we iterate through all potential $(W', d', k') \\notin V$ to ascertain any dependency to $(W, d, k) \\in V$ . Dependencies between layers can be ascertained through either manually-crafted patterns or automatic algorithms. In this work, we adhere to the approach of [11] to assess dependency based on two rudimentary rules:", + "bbox": [ + 212, + 420, + 787, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dependency Modeling. For a pair of triplets $(W', d', k')$ , $(W, d, k)$ , a dependency is established if either of the following criteria is met: 1) $W'$ and $W$ are adjacent layers with $d = d'$ and $k = k'$ ; 2) $W' = W$ and $\\mathrm{g}(W', d', k') = \\mathrm{g}(W, d, k)$ . The first criterion addresses adjacent layers as exemplified in Figure 3, while the second pertains to specialized layers, such as normalization layers, which share the same pruning pattern for both outputs and inputs. Thereafter, we update the vertex set $V$ and edge set $E$ with the discovered dependency between $(W', d', k')$ and $(W, d, k)$ accordingly:", + "bbox": [ + 212, + 631, + 787, + 752 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} V = V \\cup \\left\\{\\left(W ^ {\\prime}, d ^ {\\prime}, k ^ {\\prime}\\right) \\right\\}, \\\\ E = E + \\left\\{\\left(\\left(W _ {1} - 1\\right), \\left(W ^ {\\prime}, d ^ {\\prime}, 1\\right)\\right) \\right\\} \\end{array} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 763, + 784, + 787 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE = E \\cup \\left\\{\\left(\\left(W, d, k\\right), \\left(W ^ {\\prime}, d ^ {\\prime}, k ^ {\\prime}\\right)\\right) \\right\\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 782, + 622, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can further repeat the above process until no additional vertex is incorporated into the current graph. The resultant graph $G$ then serves as a topological", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Isomorphic Pruning for Vision Models", + "bbox": [ + 473, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c39cbbbb1c8a68e49ea3f7544e05c8267b8b5b5ca0e7a6687cdf9a9bc326b2cc.jpg", + "image_caption": [ + "Fig. 3: Isomorphic Pruning models sub-structures as graphs, and apply isolated ranking and pruning with isomorphic structures. We highlight three removable substructures in an MLP and show their corresponding graphs below. Substructures 1 & 3 are isomorphic but 2 & 3 are non-isomorphic due to the additional residual connections." + ], + "image_footnote": [], + "bbox": [ + 292, + 143, + 715, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "representation of a single sub-structure associated with the initial seed parameter. To map all substructures onto different graphs, we subsequently embark on a recursive procedure, commencing from those nodes not yet allocated to any sub-structure. Figure 3 furnishes an exemplar of this process, showcasing three distinct substructures labeled as sub-structures 1, 2, and 3.", + "bbox": [ + 212, + 396, + 787, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Ranking with Graph Isomorphism", + "text_level": 1, + "bbox": [ + 214, + 494, + 545, + 510 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the graph-based representation of network sub-structures, we advance to the ranking phase. This stage involves estimating and ranking the significance of each sub-structure to identify redundancies. A pivotal component of this process is an importance criterion, which quantifies the relative significance of various sub-structures. Previous works in the literature presents numerous well-crafted criteria, including magnitude pruning and Taylor expansion. In this work, we focus on the general importance function $I(\\cdot)$ for a single layer, represented as:", + "bbox": [ + 212, + 521, + 787, + 628 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI (W, d, k) = \\left\\{ \\begin{array}{l l} \\| \\mathrm {g} (W, d, k)) \\|, & \\text {i f M a g n i t u d e P r u n i n g} \\\\ \\| \\frac {\\partial \\mathcal {L}}{\\partial \\mathrm {g} (W , d , k)} \\mathrm {g} (W, d, k)) \\| _ {2}, & \\text {i f T a y l o r P r u n i n g} \\\\ \\dots & \\text {O t h e r i m p o r t a n c e C r i t e r i a .} \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 271, + 638, + 785, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In structural pruning, our objective is to remove sub-structures encompassing parameters across multiple layers. Existing approaches [5,11,30] generally extend the single-layer importance to a holistic measure by aggregating scores across all constituents $(W,d,k)$ for the vertexes set, denoted as $I^{*}(\\cdot)$ as follows:", + "bbox": [ + 214, + 700, + 787, + 763 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI ^ {*} (G (V, E)) = \\sum_ {(W, d, k) \\in V} I (W, d, k), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 773, + 785, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "thereby cumulating the importance of all elements within graph $G$ . However, the inherent diversity in parameters and computational topology among different", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "G. Fang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "substructures $G$ introduces potential ambiguities in this metric. Experimental evidence, particularly in the context of a vision transformer as shown in Figure 4, elucidates this issue: the largest substructure encompasses all embedding dimensions, containing a substantial number of parameters, whereas the most minimal sub-structure comprises merely two linear layers. Consequently, such aggregation may fail to accurately reflect the true relative importance of these entities. To address the challenges posed by the heterogeneity of substructures, we delineate a simple approach designed to preserve the nuanced information encapsulated by the aggregated importance, as outlined in Equation 4, whilst alleviating the detrimental repercussions raised by the heterogeneity of substructures. Our principal strategy entails the clustering of sub-structures into various isomorphic groups, predicated on their topological logic and parameter configurations. In this case, applying aggregated importance within groups can be meaningful, since isomorphic sub-structures show one-to-one correspondence among their parameters, and how different parameters participate in computations is entirely identical. To facilitate this, the key step is to examine the isomorphism of two graphs $G_{1}$ and $G_{2}$ . Formally, we adopt the following definition to ascertain graph isomorphism.", + "bbox": [ + 212, + 146, + 787, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Definition 1 (Graph Isomorphism). Two labeled graphs $G$ and $G'$ are isomorphic if there exists a bijection between their vertices so that, all mapped vertices preserve the same labels, and are connected in the same way.", + "bbox": [ + 214, + 425, + 789, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To develop a general indicator Isomorphic $(\\mathrm{g}_1,\\mathrm{g}_2)$ , we need to examine both the label and connectivity in two graphs.", + "bbox": [ + 214, + 479, + 787, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Vertex Label and Connectivity. In our settings, each vertex is labeled by the triplet $(W,d,k)$ , encompassing the parameter matrix $\\mathbf{W}$ , axis $d$ , and index $k$ . We follow a simple principle for labling: Two vertices $(\\mathbf{W},d,k)$ and $(\\mathbf{W}',d',k')$ are labeled identically if and only if the parameter $\\mathbf{W}'$ originates from the same layer type such as Linear or Conv, with the same pruning dimension $d = d'$ . We permit the index $k$ to differ, since within the same parameter matrix, the weight vectors sliced along the $d$ -th dimension are consistently homogeneous and thus can be considered isomorphic. After determining the label for each vertex, we further explore the connectivity in two graphs $G_{1}$ and $G_{2}$ . As illustrated in Figure 3, we show three substructures alongside their respective graphs $G_{1}$ , $G_{2}$ , and $G_{3}$ , where $G_{1}$ and $G_{3}$ incorporate a singular edge, whereas $G_{2}$ encompasses three edges. If we check the edges following the execution order of the computational graphs, it's natural to observe the isomorphism of $G_{1}$ and $G_{3}$ , since they share the same topology. To make this process more formal, we assume the edges in graph $G$ are already sorted according to their execution order in forwarding, and check the isomorphism by", + "bbox": [ + 212, + 530, + 787, + 772 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {I s o m o r p h i c} \\left(G, G ^ {\\prime}\\right)) = \\mathbb {1} \\left\\{\\left| G \\right| = \\left| G ^ {\\prime} \\right| \\wedge \\operatorname {l a b e l} \\left(V _ {i}\\right) = \\operatorname {l a b e l} \\left(V _ {i} ^ {\\prime}\\right); \\forall i \\in \\{1, 2, \\dots , \\left| G ^ {\\prime} \\right| \\} \\right\\},\n$$\n", + "text_format": "latex", + "bbox": [ + 248, + 784, + 785, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbb{1}$ is an indicator function. $|G^{\\prime}|$ denotes the number of edges in the graph and $\\wedge$ signifies the logical AND operation. The first condition ensures the equiva", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Isomorphic Pruning for Vision Models", + "bbox": [ + 473, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Isomorphic Pruning" + ], + "code_body": "1: Input: Network with parameters $\\{W_1, W_2, \\ldots, W_n\\}$ , Pruning ratio $p\\%$ . \n2: Output: Pruned Network $\\{W_1', W_2', \\ldots, W_n'\\}$ \n3: procedure IDENTIFYSUBSTRUCTURES(\\{W1, W2, \\ldots, Wn\\}) \n4: Initialize empty list of sub-structures $\\mathbf{S} = \\{\\}$ \n5: for each layer $W_i$ in $\\{W_1, W_2, \\ldots, W_n\\}$ do \n6: if $W_i$ is not assigned to any substructure then \n7: Initialize a graph $G(V = \\{W_i\\}, E = \\{\\})$ . \n8: Find and add coupled $W_j$ to $G$ until there is no new dependency. \n9: $\\mathbf{S} = \\mathbf{S} \\cup \\{G\\}$ \n10: return $\\mathbf{S}$ \n11: procedure PRUNEBYGROUPEDRANKING(S, p) \n12: for each substructure $G_i$ in $\\{G_1, G_2, \\ldots, G_n\\}$ do \n13: if $G_i$ is not assigned to any group then \n14: Initialize a Group $\\mathbf{R} = \\{G_i\\}$ . \n15: Find and add $G_j$ that satisfies Isomorphism $(G_i, G_j) = 1$ to $\\mathbf{R}$ . \n16: Calculate the importance scores $\\{I^*(G) | G \\in \\mathbf{R}\\}$ . \n17: Remove $p\\%$ substructures in $\\mathbf{R}$ by ranking the importance scores. \n18: return $\\mathbf{S}$ \n19: $\\mathbf{S} \\gets \\text{IDENTIFYSUBSTRUCTURES}(\\{W_1, W_2, \\ldots, W_n\\})$ \n20: $\\mathbf{S} \\gets \\text{PRUNEBYGROUPEDRANKING}(\\mathbf{S}, p)$", + "bbox": [ + 215, + 164, + 754, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lence of the number of vertices in the graphs, while the second condition confirms the matching labels of vertices on either side of the edge pairs $V_{i}$ and $V_{i}^{\\prime}$ .", + "bbox": [ + 212, + 484, + 784, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Using Equation 5, it is easy to cluster substructures into different groups and perform rankings within each group exclusively. Specifically, we examine a collection of substructures $\\{G_1, G_2, \\ldots, G_N\\}$ that share identical graph topologies. We estimate the aggregated significance of each graph independently using Equation 4, resulting in an $N$ -dimensional importance vector $[I^*(G_1), \\ldots, I^*(G_N)]$ . This facilitates a straightforward ranking of the importance vector, followed by the elimination of the $\\mathfrak{p}\\%$ unimportant substructures based on their scores. This ranking method is deemed reliable since all substructures within a group possess the same architectural design, parameter scale, and computational graph. A similar principle can be uniformly applied to other isomorphic groups for pruning, as detailed in Algorithm 1.", + "bbox": [ + 212, + 515, + 785, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 703, + 375, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This section reports our empirical results on the ImageNet-1K dataset [8], which comprises 1,281,167 images for training and 50,000 for validation. We deploy the proposed Isomorphic Pruning to compress several vision models, including ConvNext [33], ResNet [36], MobileNetv2 [44], and Vision Transformers [10, 50]. We report the parameter amount (#Params), MACs, and Top-1 validation accuracy in Table 1 and 2, and report the actual latency and peak memory consumption in Table 3.", + "bbox": [ + 212, + 733, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "G. Fang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Settings for ImageNet-1K", + "text_level": 1, + "bbox": [ + 215, + 146, + 473, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Pruning. For simplicity, we adopt the Taylor-based pruning framework [39], which applies a data-driven Taylor expansion to the loss function. For this purpose, we randomly sampled 100 mini-batches from the ImageNet dataset, each containing 64 images. These images were resized to $256 \\times 256$ pixels and then center-cropped to $224 \\times 224$ pixels. We normalize all images using the mean and standard deviation specific to ImageNet. Notably, while some methods, such as ConvNext such as ConvNext [33] and DeiT [50] employ additional training techniques like label smoothing and knowledge distillation, our approach only leverages a straightforward cross-entropy loss for the Taylor expansion. We accumulate gradients from all the batches for a more reliable importance estimation and apply Isomorphic pruning in a one-shot manner without iterative pruning and fine-tuning [40].", + "bbox": [ + 212, + 171, + 787, + 354 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Finetuning. During the fine-tuning process of pruned models, we adhere to the original training protocol for a fair comparison. While maintaining most of the original settings unchanged, we adjusted the learning rates and batch sizes to make the training suitable for A5000 GPUs, following the linear LR scaling rule proposed in [13]. For example, to compressing a pre-trained DeiT models [50], we adopt the same augmentation strategies for training, including RandAugmentation [7], Mixup [67], CutMix [66], Random Erasing [69], Repeated Augmentation [23], label smoothing [49]. The pruned models are distilled with a pre-trained RegNetY [43, 50] for 300 epochs using the AdamW [34] optimizer, with a learning rate of 0.0005 and a total batch size of 2048. A weight decay of 0.05 and a cosine annealing scheduler is deployed for training. For other models like ConvNext [33], ResNet [19] and MobileNet-v2 [44], the same principle will be applied. More details of hyper-parameters, pruning, and finetuning can be found in the appendix.", + "bbox": [ + 212, + 376, + 787, + 588 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation. We report the optimal classification accuracy achieved on the ImageNet-1K validation set, employing the standard resize-and-crop protocol [19, 33] for evaluation. To further validate the efficiency of models, we profiled the latency and the peak memory consumption of the pruned models on both GPU (a single RTX A5000) and CPU. To guarantee a fair and standardized comparison, we also tested the latency of several baseline models sourced from the Pytorch-Image Models [57] and Torchvision [38], following the same testing procedure. The latency was estimated with a batch size of 256 for GPU and 8 for CPU, averaged on 100 repeats. Appropriate GPU warmup was deployed in our test.", + "bbox": [ + 212, + 609, + 787, + 748 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Pruning Convolutional Neural Networks", + "text_level": 1, + "bbox": [ + 215, + 768, + 596, + 784 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Convolutional Neural Network is a critical focus in the area of network pruning. In this section, we concentrate on pruning three popular ConvNets: ConvNext [33], ResNet [19] and MobileNet-v2 [44].", + "bbox": [ + 212, + 794, + 785, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Isomorphic Pruning for Vision Models", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/184a1b705ee0abc35115746d98b63b430510eda61563aa6a5cdd5123f86ae588.jpg", + "table_caption": [ + "Table 1: Pruning CNNs on ImageNet-1K." + ], + "table_footnote": [], + "table_body": "
Pruning Method#Params (M)MACs (G)Base Acc.Final Acc.Δ Acc.
ConvNext-B† [33]88.5915.39-83.83-
ConvNext-S† [33]50.228.71-83.14-
ConvNext-S (Ours)47.368.4883.8383.17-0.66
ConvNext-T† [33]28.594.47-82.06-
ConvNext-T (Ours)25.324.1983.8382.19-1.64
ResNet-152†60.1911.58-78.31-
AOFP-E2 [9]-4.1377.3777.00-0.37
ABCPruner [28]24.074.3178.3177.12-1.19
EPruner-0.63 [27]21.564.0878.3176.83-1.48
ResNet-152-4.0G (Ours)23.114.0578.3177.84-0.47
ResNet-101†44.557.85-77.38-
IE [39]31.24.7077.3777.35-0.02
FPGM [21]-4.5177.3777.32-0.05
SFP [20]-4.5177.3777.51+0.14
ISTA [59]-4.4776.4075.27-1.13
Res101-4.5G (Ours)29.144.4877.3877.56+0.16
GFP [30]28.023.9078.2978.33+0.04
AOFP [9]-3.8976.6376.40-0.23
Res101-3.8G (Ours)24.873.8577.3877.43+0.05
ResNet-50† [19]25.564.13-76.13-
Taylor [39]14.202.2576.1874.50-1.68
CCP [41]-2.1176.1575.50-0.65
GFP [30]19.422.0476.7976.42-0.37
AutoSlim [63]20.602.0076.1075.60-0.50
DepGraph [11]-1.9976.1575.83-0.32
ResNet50-2G (Ours)15.052.0676.1375.91-0.22
Mobv2†3.500.32-71.88-
Meta [32]-0.1474.7068.20-6.50
GFP [30]-0.1575.7469.16-6.58
DepGraph [11]-0.1571.9068.49-3.41
Mobv2-0.15G (Ours)2.190.1571.8868.91-2.97
", + "bbox": [ + 272, + 160, + 723, + 534 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ConvNext. Our experiments in Table 1 begin with the evaluation on an advanced CNN, ConvNext [33], which achieves impressive performance on several benchmarks. We use the official ConvNext-Base [38] as the pre-trained model for pruning, which achieves an accuracy of $83.83\\%$ on the ImageNet-1K validation set. The pruning of ConvNext models is a relatively underexplored topic. Hence, we establish our baselines using the official variants with fewer parameters, ConvNext-S and ConvNext-T. For ConvNext, we also generalize our method to depth pruning, which removes unimportant layers with the minimal average scores on substructures, so that the pruned models have the same depth as baselines. The pruned models are fine-tuned for 300 epochs following the same training protocol as [33]. During fine-tuning, we observed that the pruned ConvNext with half of the parameters removed, achieves an accuracy of 80.09 in only 20 epochs, which effectively validates the efficiency of pruning. Moreover, as illustrated in Table 1, the final model after 300 full fine-tuning, achieves competitive performance (82.19% vs. 82.06%) to pre-trained models with fewer parameters (25.32M vs. 28.59M) and MACs (4.19G vs. 4.47G).", + "bbox": [ + 212, + 547, + 787, + 789 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ResNet. In line with prior studies [5, 11, 22, 30], we demonstrate the efficacy of our method by structurally pruning pre-trained ResNet MACs. ResNet-", + "bbox": [ + 212, + 808, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "G. Fang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/11bd7db5add35241b1f1b835eab3ca82855034a77d9b9c36f78790a92b7ca135.jpg", + "table_caption": [ + "Table 2: Pruning DeiTs on ImageNet-1K." + ], + "table_footnote": [], + "table_body": "
Method#Params (M) MACs (G) Acc (%)Method#Params (M) MACs (G) Acc (%)
DeiT-B† [50]87.3417.6983.32DeiT-T† [50]5.911.2774.52
DeiT-S† [50]22.444.6481.17DeiT-T-600EP [58]5.911.2775.00
DeiT-S-600EP [58]22.444.6481.80WDPruning [61]13.302.6078.38
VTP - 40% [70]48.0010.0080.70X-Pruner [64]-2.4078.93
WDPruning [61]55.309.9080.76UVC [65]-2.3278.82
UVC [65]-8.0080.57ViT-Slim [70]11.402.3077.94
X-Pruner [64]-8.5081.02ViT-Slim [70]11.402.3077.94
NViT-B [58]34.006.8083.29NViT-T [58]6.901.3076.21
SAViT [68]25.405.3081.66SAViT+KD [68]6.601.3076.95
UP-DeiT-S [62]22.10-81.56UP-DeiT-T [62]5.70-75.79
NViT-S [58]21.004.2082.19DeiT-T (Ours)5.741.2177.50
Upop [45]19.904.1081.10VTC-LFC [55]5.10-71.60
DeiT-S (Ours)20.694.1682.41SViT-E [68]4.210.9970.12
GOHSP [60]14.402.9979.98GOHSP [60]4.000.9170.24
GOHSP [60]11.102.8179.86SAViT [68]4.200.9070.72
ViT-Slim [2]13.502.8079.20CP-ViT [47]-0.7471.24
Upop [45]13.502.8079.60WDPruning [61]3.500.7070.34
ViT-Slim [2]13.502.8079.20X-Pruner [64]-0.6071.10
CP-ViT [47]-2.6679.08UVC [65]-0.5170.60
DeiT-2.6G (Ours)13.072.6281.13DeiT-0.6G (Ours)3.080.6272.60
", + "bbox": [ + 228, + 160, + 776, + 368 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "50/101/152 consists of four residual blocks, each formed by several Bottle-neck structures and residual connections [19]. The proposed Isomorphic Pruning identifies five kinds of distinct isomorphic groups within ResNet-50, four targeting the block structures and one associated with bottleneck structures [19]. It is worthwhile to note that different sub-structures vary in their parameter and MAC counts. For example, residual blocks naturally contain more parameters than the bottleneck structures which are only composed of a few convolutional layers. The conventional global pruning, employing a naive global ranking threshold, faces the potential risks of over-pruning certain layers due to the biased importance distribution. Isomorphic Pruning, however, employs a more stable scheme for pruning. As illustrated in Table 1, we report both the accuracy and the accuracy drop compared to their base models. Isomorphic pruning achieves lossless compression on ResNet-101, under both 4.5G and 3.8G settings.", + "bbox": [ + 212, + 393, + 787, + 592 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "MobileNet-v2. We also conduct experiments on lightweight models, MobileNet-v2, by compressing its MACs from 0.32 G to 0.15 G. In MobileNet-v2, we can also discover 5 isomorphic groups in MobileNet. The pruned model achieves better accuracy compared to DepGraph [11] which requires sparse training.", + "bbox": [ + 212, + 612, + 789, + 676 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Pruning Vision Transformers.", + "text_level": 1, + "bbox": [ + 214, + 694, + 506, + 709 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Vision Transformers (ViT) has profoundly reshaped the paradigm of visual modeling over the past few years. Nevertheless, their inherent over-parameterization often imposes huge overheads to both training and inference. Compared to CNNs that are composed of similar convolution layers or blocks, Vision Transformers are usually built with MLP, Multi-head Attention, and skip connections, which are more challenging for pruning due to the heterogeneous internal structures.", + "bbox": [ + 212, + 718, + 785, + 809 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 2 presents our pruning results on DeiT, a plain vision transformer enhanced with knowledge distillation [50]. We benchmark our approach with", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Isomorphic Pruning for Vision Models", + "bbox": [ + 475, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/fd634b3a6df3020c23ea4e8bde3762b05ba0c4bd98672cfc59cfe361f27d75d0.jpg", + "table_caption": [ + "Table 3: Profiling the pruned models and other off-the-shelf vision models from PyTorch-Image-Models [57] on a single RTX A5000 GPU and CPU. For NViT [58], we report their best speed up in the original paper, tested on a single V100 or RTX 3080." + ], + "table_footnote": [], + "table_body": "
Method#Params ↓ MACs ↓ Peak Mem. ↓ GPU Latency (256) ↓ CPU Latency (8) ↓ Acc. (%)
DeiT-B [50]86.40 M17.60 G2909.78 MB802.97 ms (1.00×)197.73 ms (1.00×)83.32
DeiT-S [50]22.44 M4.64 G1363.47 MB241.81 ms (3.32×)80.64 ms (2.45×)81.17
Swin-T [31]28.29 M4.51 G4257.98 MB375.45 ms (2.14×)203.86 ms (0.97×)81.19
NViT-S [58]21.00 M4.20 G--(2.52×)-82.19
TNT-S [15]23.76 M4.85 G1457.56 MB631.02 ms (1.27×)175.58 ms (1.13×)81.50
CaiT-XS24 [51]26.56 M5.40 G1496.42 MB557.41 ms (1.44×)342.28 ms (0.58×)82.00
CrossViT-S [3]26.86 M5.64 G1546.94 MB315.02 ms (2.55×)115.51 ms (1.71×)81.00
EfficientFormer-L3 [26]31.41 M3.94 G2681.75 MB249.05 ms (3.22×)167.16 ms (1.18×)82.40
PVTv2-B2 [54]25.36 M4.00 G4478.18 MB364.00 ms (2.21×)212.44 ms (0.93×)82.00
DeiT-S (Ours)20.69 M4.16 G1547.97 MB230.84 ms (3.48×)44.10 ms (4.48×)82.41
DeiT-2.6G (Ours)13.07 M2.62 G1203.07 MB189.68 ms (4.23×)56.59 ms (3.49×)81.13
DeiT-T [50]5.91 M1.27 G720.22 MB93.78 ms (8.56×)46.83 ms (4.22×)74.52
NViT-T [58]6.90 M1.30 G--(4.97×)-76.21
DeiT-T (Ours)5.71 M1.20 G837.16 MB93.17 ms (8.62×)47.89 ms (4.13×)77.50
DeiT-0.6G (Ours)3.08 M0.62 G629.28 MB63.23 ms (12.70×)40.98 ms (4.83×)72.60
ConvNext-B [33]88.59 M15.36 G5394.10 MB789.24 ms (1.00×)262.58 ms (1.00×)83.83
ConvNext-S [33]50.22 M8.69 G3958.45 MB508.20 ms (1.57×)188.07 ms (1.40×)83.14
ConvNext-S (Ours)47.36 M8.48 G3788.65 MB535.95 ms (1.59×)179.72 ms (1.46×)83.17
ConvNext-T [33]28.59 M4.46 G3785.37 MB291.02 ms (2.82×)126.17 ms (2.08×)82.06
ConvNext-T (Ours)25.32 M4.19 G3607.06 MB292.77 ms (2.83×)97.82 ms (2.68×)82.19
ResNet152 [19]60.19 M11.58 G3320.24 MB386.65 ms (1.00×)272.21 ms (1.00×)78.31
ResNet152-4G (Ours)23.11 M4.06 G1938.91 MB238.12 ms (1.62×)180.65 ms (1.51×)77.84
ResNet101 [19]44.55 M7.85 G3196.77 MB271.34 ms (1.00×)202.66 ms (1.00×)77.38
ResNet101-4.5G (Ours)29.14 M4.48 G2639.40 MB220.72 ms (1.23×)162.33 ms (1.25×)77.56
ResNet101-3.8G (Ours)24.87 M3.85 G2286.36 MB194.39 ms (1.40×)158.98 ms (1.28×)77.43
ResNet50 [19]25.56 M4.13 G3035.84 MB171.54 ms (1.00×)114.86 ms (1.00×)76.13
ResNet50-2.0G (Ours)15.05 M2.06 G2238.15 MB78.75 ms (1.33×)80.88 ms (1.42×)75.91
Mobv2 [44]3.50 M0.32 G2855.44 MB71.42 ms (1.00×)74.46 ms (1.00×)71.88
Mobv2-0.15G (Ours)2.19 M0.15 G1354.82 MB43.48 ms (1.64×)38.63 ms (1.93×)68.91
", + "bbox": [ + 217, + 198, + 784, + 502 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "a range of innovative pruning baselines specifically tailored for vision transformers [45, 47, 50, 55, 58, 60, 61, 64, 65, 68, 70]. A DeiT model consists of various components, such as MLP and attention layers. Within the attention layers, we can even find two feasible pruning schemes: pruning 1) the number of heads or 2) within-head dimensions. We implement Isomorphic Pruning to compress a pre-trained DeiT-Base model by pruning both heads, head dimensions, and the embedding sizes following [58]. This is achieved through a straightforward first-order Taylor expansion for importance estimation, followed by a grouped ranking in Isomorphic Pruning. The details of pruning vision transformers can be found in the appendix.", + "bbox": [ + 212, + 536, + 787, + 686 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We developed four lightweight vision transformers: DeiT-S, DeiT-2.6G, DeiT-T and DeiT-0.6G. The pruned DeiT-S and DeiT-T models achieve comparable latencies comparable to the standard DeiT-S and DeiT-T with manually designed architecture. The empirical results are reported in Table 2, which shows that our method yields greater performance gains in complex models with extensive heterogeneous internal structures, especially in comparison to CNNs. Specifically, the optimized DeiT-T achieves an accuracy of $77.50\\%$ , with only 1.2G MACs, which is better than the results $(74.52\\%)$ of the pre-trained DeiT-T with uniform width across layers. Detailed configurations, such as base models and pruning ratios, are available in the appendix.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "G. Fang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e3600dd6856369561d35f1101ff00315d23351fc6c524ef5700d3bbb6544dff2.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 243, + 145, + 380, + 306 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9ea2c9c087d773e289c1bc67b427415f2b26b09f77e1206d308f314c1b97a7e9.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 388, + 145, + 527, + 306 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0279396134af7b62b84d0b38091ee89e433508089848e776920f76d687e0c1dc.jpg", + "image_caption": [ + "(c)", + "Fig. 4: (a-b) Top-1 Accuracy and loss of pruning ResNet-50 on ImageNet-1K validation set without finetuning. We report the mean and standard derivation of 10 experiments. IsomorphicP Pruning consistently improves the performance of pruned models. (c) Histogram of Taylor importance score of DeiT sub-structures. Isomorphic structures are highlighted with the same color. Zoom in for more details about each structure. We visualize the threshold of $50\\%$ pruning ratio for naive global pruning (the black dash) and isomorphic pruning (the colored dash)." + ], + "image_footnote": [], + "bbox": [ + 542, + 145, + 753, + 308 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Analysis on the Pruned Models", + "text_level": 1, + "bbox": [ + 215, + 460, + 522, + 476 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "On-device Latency Table 3 measures the actual latency and peak memory consumptions on a single RTX A5000 GPU and CPU. We evaluate all models with a batch size of 256 for GPU and 8 for CPU, respectively. The average latency over 100-round experiments are reported, with a 20-step warmup process before evaluation. Firstly, results show that the pruned DeiT models can achieve comparable performance (82.41% vs. 82.40%) to some advanced transformers like EfficientFormer [26], while preserving a competitive latency. Additionally, we observe that the peak memory consumption is primarily determined by the largest layer width. Therefore, under the same budget, a uniform DeiT may require less peak memory (1363 MB for DeiT-S) for inference. In contrast, a non-uniform architecture may require slightly more memory (1547 MB for the DeiT-S Pruned) under the given budget. But non-uniform pruning can achieve better accuracy, which reveals an important trade-off between accuracy and memory consumption for transformers. In addition, for CNNs like ResNets, the pruned model requires much less memory due to the compressed feature dimensions. Furthermore, we note that the actual speed-up may not be linearly correlated with the MAC reductions. This is particularly evident in ResNet-50, where only a $1.33 \\times$ actual speed-up is achieved despite the removal of half of the MACs. In contrast, compression on Mobv2 is more effective, achieving a $1.64 \\times$ acceleration with a similar compression ratio on MACs.", + "bbox": [ + 212, + 484, + 787, + 787 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Distribution of Importance Scores. To further explore how Isomorphic Pruning affects the pruning behavior, we study the distribution of importance", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Isomorphic Pruning for Vision Models", + "bbox": [ + 475, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "score in a DeiT-Base model [50]. As shown in Figure 4 (c), the importance histogram of different sub-structures is highlighted with different colors, each corresponding to one kind of isomorphic structure. It is natural to find that, the MLP layer and the attention layer have diverged importance distributions due to their different computation process. Besides, when we apply a simple $50\\%$ global pruning on the DeiT model, most parameters in Attentions $\\# 5$ , $\\# 9$ , and $\\# 13$ will be removed, since the importance scores in attentions are centered around 0. In contrast, the embedding number as $\\# 1$ will not be even been pruned in this case, due to the relatively large importance scores. This inevitably leads to an imbalanced pruning of different structures. Notably, the embedding structure, as illustrated in the appendix, contains massive parameters in the DeiT, where biased pruning may lead to very limited acceleration.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Importance Criteria. Note that Isomorphic Pruning can be combined with different pruning criteria like Taylor [39], Hessian [25] and even L1-norm [22]. We quantify the performance gain of Isomorphic Pruning by visualizing the accuracy and loss curve with different pruning ratios in Figure 4 (a) and (b). Specifically, we gradually increase the pruning ratio for each criterion and report the average accuracy and loss without finetuning over 10 repeated runs with random inputs for Taylor expansion [25, 39]. Note that L1 pruning is deterministic and thus presents 0 std. We find the Isomorphic Pruning, marked as solid lines, consistently improves the counterpart baselines highlighted in dashed lines.", + "bbox": [ + 212, + 351, + 787, + 488 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Transfer Learning. We also evaluate the performance of transfer learning based on our pruned models. Table 4 reports the accuracy of transferring different models to CIFAR-10 and CIFAR-100. We resize all images to $224 \\times 224$ and apply the same training recipe for transfer. For the DeiT models, we did not include the supervision from the teacher models. It is observed that our method achieves better transfer accuracy compared to the scratch training model and baseline methods.", + "bbox": [ + 212, + 511, + 545, + 676 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/c6c68c098f78a8a275c90a2aa105c087ba93255d6422ca1cdaadb09b0745b5a4.jpg", + "table_caption": [ + "Table 4: Transfer learning on pruned models." + ], + "table_footnote": [], + "table_body": "
ModelsCIFAR-10CIFAR-100
DeiT-S [39]98.5287.07
NViT-S [58]98.7887.90
DeiT-S (Ours)99.0290.09
DeiT-T93.9385.66
NViT-T [58]98.3185.88
DeiT-S (Ours)98.2887.08
ConvNext-S98.7789.15
ConvNext-S (Ours)98.7889.36
", + "bbox": [ + 558, + 554, + 782, + 661 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 700, + 359, + 717 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce Isomorphic Pruning, a practical approach designed to compress vision networks with novel mechanisms and architectures. This method focuses on mitigating the challenges posed by heterogeneous sub-structures in networks, thereby enhancing the reliability of ranking and pruning processes. Our empirical results on the ImageNet-1k dataset, showcase the effectiveness of Isomorphic Pruning across several vision models like ConvNext, ResNet, MobileNetv2, and Vision Transformers.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "G. Fang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This project is supported by the National Research Foundation Singapore under its Medium Sized Center for Advanced Robotics Technology Innovation, and the Advanced Research and Technology Innovation Centre (ARTIC), the National University of Singapore under Grant (project number: A-0005947-21-00, project reference: ECT-RP2).", + "bbox": [ + 212, + 175, + 787, + 251 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 273, + 321, + 287 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Baum, E., Haussler, D.: What size net gives valid generalization? Advances in neural information processing systems 1 (1988)", + "2. Chavan, A., Shen, Z., Liu, Z., Liu, Z., Cheng, K.T., Xing, E.P.: Vision transformer slimming: Multi-dimension searching in continuous optimization space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4931-4941 (2022)", + "3. Chen, C.F.R., Fan, Q., Panda, R.: Crossvit: Cross-attention multi-scale vision transformer for image classification. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 357-366 (2021)", + "4. Chen, T., Cheng, Y., Gan, Z., Yuan, L., Zhang, L., Wang, Z.: Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems 34, 19974-19988 (2021)", + "5. Chen, T., Liang, L., Ding, T., Zhu, Z., Zharkov, I.: Otov2: Automatic, generic, user-friendly. arXiv preprint arXiv:2303.06862 (2023)", + "6. Chollet, F.: Xception: Deep learning with depthwise separable convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1251-1258 (2017)", + "7. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 702-703 (2020)", + "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "9. Ding, X., Ding, G., Guo, Y., Han, J., Yan, C.: Approximated oracle filter pruning for destructive cnn width optimization. In: International Conference on Machine Learning. pp. 1607-1616. PMLR (2019)", + "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "11. Fang, G., Ma, X., Song, M., Mi, M.B., Wang, X.: Depgraph: Towards any structural pruning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16091-16101 (2023)", + "12. Frantar, E., Alistarh, D.: Sparsegpt: Massive language models can be accurately pruned in one-shot (2023)", + "13. Goyal, P., Dolkar, P., Girshick, R., Noordhuis, P., Wesolowski, L., Kyrola, A., Tulloch, A., Jia, Y., He, K.: Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677 (2017)" + ], + "bbox": [ + 218, + 301, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Isomorphic Pruning for Vision Models", + "bbox": [ + 475, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Guo, J., Ouyang, W., Xu, D.: Multi-dimensional pruning: A unified framework for model compression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1508-1517 (2020)", + "15. Han, K., Xiao, A., Wu, E., Guo, J., Xu, C., Wang, Y.: Transformer in transformer. Advances in Neural Information Processing Systems 34, 15908-15919 (2021)", + "16. Han, S., Mao, H., Dally, W.J.: Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. International Conference on Learning Representations (ICLR) (2016)", + "17. Hassibi, B., Stork, D.G., Wolff, G.J.: Optimal brain surgeon and general network pruning. In: IEEE international conference on neural networks. pp. 293-299. IEEE (1993)", + "18. He, H., Liu, J., Pan, Z., Cai, J., Zhang, J., Tao, D., Zhuang, B.: Pruning self-attention into convolutional layers in single path. arXiv preprint arXiv:2111.11802 (2021)", + "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "20. He, Y., Kang, G., Dong, X., Fu, Y., Yang, Y.: Soft filter pruning for accelerating deep convolutional neural networks. arXiv preprint arXiv:1808.06866 (2018)", + "21. He, Y., Liu, P., Wang, Z., Hu, Z., Yang, Y.: Filter pruning via geometric median for deep convolutional neural networks acceleration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4340-4349 (2019)", + "22. He, Y., Zhang, X., Sun, J.: Channel pruning for accelerating very deep neural networks. In: Proceedings of the IEEE international conference on computer vision. pp. 1389-1397 (2017)", + "23. Hoffer, E., Ben-Nun, T., Hubara, I., Giladi, N., Hoefler, T., Soudry, D.: Augment your batch: Improving generalization through instance repetition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8129-8138 (2020)", + "24. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)", + "25. LeCun, Y., Denker, J., Solla, S.: Optimal brain damage. Advances in neural information processing systems 2 (1989)", + "26. Li, Y., Yuan, G., Wen, Y., Hu, J., Evangelidis, G., Tulyakov, S., Wang, Y., Ren, J.: Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems 35, 12934-12949 (2022)", + "27. Lin, M., Ji, R., Li, S., Wang, Y., Wu, Y., Huang, F., Ye, Q.: Network pruning using adaptive exemplar filters. IEEE Transactions on Neural Networks and Learning Systems 33(12), 7357-7366 (2021)", + "28. Lin, M., Ji, R., Zhang, Y., Zhang, B., Wu, Y., Tian, Y.: Channel pruning via automatic structure search. arXiv preprint arXiv:2001.08565 (2020)", + "29. Liu, B., Wang, M., Foroosh, H., Tappen, M., Pensky, M.: Sparse convolutional neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 806-814 (2015)", + "30. Liu, L., Zhang, S., Kuang, Z., Zhou, A., Xue, J.H., Wang, X., Chen, Y., Yang, W., Liao, Q., Zhang, W.: Group fisher pruning for practical network compression. In: International Conference on Machine Learning. pp. 7021-7032. PMLR (2021)", + "31. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "G. Fang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)", + "32. Liu, Z., Mu, H., Zhang, X., Guo, Z., Yang, X., Cheng, K.T., Sun, J.: Metapruning: Meta learning for automatic neural network channel pruning. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3296-3305 (2019)", + "33. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11976-11986 (2022)", + "34. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "35. Louizos, C., Welling, M., Kingma, D.P.: Learning sparse neural networks through $l\\_0$ regularization. arXiv preprint arXiv:1712.01312 (2017)", + "36. Ma, X., Yuan, G., Lin, S., Li, Z., Sun, H., Wang, Y.: Resnet can be pruned $60 \\times$ : Introducing network purification and unused path removal (p-rm) after weight pruning. In: 2019 IEEE/ACM International Symposium on Nanoscale Architectures (NANOARCH). pp. 1-2. IEEE (2019)", + "37. Ma, X., Fang, G., Wang, X.: Llm-pruner: On the structural pruning of large language models. In: Advances in Neural Information Processing Systems (2023)", + "38. maintainers, T., contributors: Torchvision: Pytorch's computer vision library. GitHub repository (2016)", + "39. Molchanov, P., Mallya, A., Tyree, S., Frosio, I., Kautz, J.: Importance estimation for neural network pruning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11264-11272 (2019)", + "40. Molchanov, P., Tyree, S., Karras, T., Aila, T., Kautz, J.: Pruning convolutional neural networks for resource efficient inference. arXiv preprint arXiv:1611.06440 (2016)", + "41. Peng, H., Wu, J., Chen, S., Huang, J.: Collaborative channel pruning for deep networks. In: International Conference on Machine Learning. pp. 5113-5122. PMLR (2019)", + "42. Pool, J., Sawarkar, A., Rodge, J.: Accelerating inference with sparsity using the nvidia ampere architecture and nvidia tensorsr. NVIDIA Developer Technical Blog, https://developer.nvidia.com/blog/accelerating-inference-with-sparsityusing-ampere-and-tensorr (2021)", + "43. Radosavovic, I., Kosaraju, R.P., Girshick, R., He, K., Dollár, P.: Designing network design spaces. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10428-10436 (2020)", + "44. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: Mobilenetv2: Inverted residuals and linear bottlenecks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4510-4520 (2018)", + "45. Shi, D., Tao, C., Jin, Y., Yang, Z., Yuan, C., Wang, J.: UPop: Unified and progressive pruning for compressing vision-language transformers. In: Proceedings of the 40th International Conference on Machine Learning. vol. 202, pp. 31292-31311. PMLR (2023)", + "46. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)", + "47. Song, Z., Xu, Y., He, Z., Jiang, L., Jing, N., Liang, X.: Cp-vit: Cascade vision transformer pruning via progressive sparsity prediction. arXiv preprint arXiv:2203.04570 (2022)", + "48. Sun, M., Liu, Z., Bair, A., Kolter, J.Z.: A simple and effective pruning approach for large language models. arXiv preprint arXiv:2306.11695 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Isomorphic Pruning for Vision Models", + "bbox": [ + 475, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818-2826 (2016)", + "50. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers & distillation through attention. In: International Conference on Machine Learning. vol. 139, pp. 10347-10357 (July 2021)", + "51. Touvron, H., Cord, M., Sablayrolles, A., Synnaeve, G., Jégou, H.: Going deeper with image transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 32-42 (2021)", + "52. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "53. Voita, E., Talbot, D., Moiseev, F., Sennrich, R., Titov, I.: Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. arXiv preprint arXiv:1905.09418 (2019)", + "54. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media 8(3), 415-424 (2022)", + "55. Wang, Z., Luo, H., Wang, P., Ding, F., Wang, F., Li, H.: Vtc-lfc: Vision transformer compression with low-frequency components. Advances in Neural Information Processing Systems 35, 13974-13988 (2022)", + "56. Wen, W., Wu, C., Wang, Y., Chen, Y., Li, H.: Learning structured sparsity in deep neural networks. Advances in neural information processing systems 29 (2016)", + "57. Wightman, R.: Pytorch-image-models. https://github.com/huggingface/pytorch-image-models", + "58. Yang, H., Yin, H., Shen, M., Molchanov, P., Li, H., Kautz, J.: Global vision transformer pruning with hessian-aware saliency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18547-18557 (2023)", + "59. Ye, J., Lu, X., Lin, Z., Wang, J.Z.: Rethinking the smaller-norm-less-informative assumption in channel pruning of convolution layers. arXiv preprint arXiv:1802.00124 (2018)", + "60. Yin, M., Uzkent, B., Shen, Y., Jin, H., Yuan, B.: Gohsp: A unified framework of graph and optimization-based heterogeneous structured pruning for vision transformer. arXiv preprint arXiv:2301.05345 (2023)", + "61. Yu, F., Huang, K., Wang, M., Cheng, Y., Chu, W., Cui, L.: Width & depth pruning for vision transformers. In: AAAI Conference on Artificial Intelligence (AAAI). vol. 2022 (2022)", + "62. Yu, H., Wu, J.: A unified pruning framework for vision transformers. Science China Information Sciences 66(7), 1-2 (2023)", + "63. Yu, J., Huang, T.: Autoslim: Towards one-shot architecture search for channel numbers. arXiv preprint arXiv:1903.11728 (2019)", + "64. Yu, L., Xiang, W.: X-pruner: explainable pruning for vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24355-24363 (2023)", + "65. Yu, S., Chen, T., Shen, J., Yuan, H., Tan, J., Yang, S., Liu, J., Wang, Z.: Unified visual transformer compression. arXiv preprint arXiv:2203.08243 (2022)", + "66. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6023-6032 (2019)", + "67. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "G. Fang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "68. Zheng, C., Zhang, K., Yang, Z., Tan, W., Xiao, J., Ren, Y., Pu, S., et al.: Savit: Structure-aware vision transformer pruning via collaborative optimization. Advances in Neural Information Processing Systems 35, 9010-9023 (2022)", + "69. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 13001-13008 (2020)", + "70. Zhu, M., Tang, Y., Han, K.: Vision transformer pruning. arXiv preprint arXiv:2104.08500 (2021)" + ], + "bbox": [ + 215, + 146, + 787, + 258 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Isomorphic Pruning for Vision Models", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Isomorphic Pruning for Vision Models/bf292aec-a3cb-46ef-92f7-27df67a69e35_model.json b/2024/Isomorphic Pruning for Vision Models/bf292aec-a3cb-46ef-92f7-27df67a69e35_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dbdf8562e9f06189ecf1951961c30af707e46116 --- /dev/null +++ b/2024/Isomorphic Pruning for Vision Models/bf292aec-a3cb-46ef-92f7-27df67a69e35_model.json @@ -0,0 +1,2427 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.282, + 0.142, + 0.723, + 0.163 + ], + "angle": 0, + "content": "Isomorphic Pruning for Vision Models" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.189, + 0.768, + 0.207 + ], + "angle": 0, + "content": "Gongfan Fang\\(^{1}\\), Xinyin Ma\\(^{1}\\), Michael Bi Mi\\(^{2}\\), and Xinchao Wang\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.384, + 0.217, + 0.619, + 0.232 + ], + "angle": 0, + "content": "\\(^{1}\\) National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.407, + 0.232, + 0.595, + 0.246 + ], + "angle": 0, + "content": "\\(^{2}\\) Huawei Technologies Ltd." + }, + { + "type": "text", + "bbox": [ + 0.358, + 0.247, + 0.647, + 0.261 + ], + "angle": 0, + "content": "gongfan@u.nus.edu，xinchao@nus.edu.sg" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.288, + 0.744, + 0.635 + ], + "angle": 0, + "content": "Abstract. Structured pruning reduces the computational overhead of deep neural networks by removing redundant sub-structures. However, assessing the relative importance of different sub-structures remains a significant challenge, particularly in advanced vision models featuring novel mechanisms and architectures like self-attention, depth-wise convolutions, or residual connections. These heterogeneous substructures usually exhibit diverged parameter scales, weight distributions, and computational topology, introducing considerable difficulty to importance comparison. To overcome this, we present Isomorphic Pruning, a simple approach that demonstrates effectiveness across a range of network architectures such as Vision Transformers and CNNs, and delivers competitive performance across different model sizes. Isomorphic Pruning originates from an observation that, when evaluated under a pre-defined importance criterion, heterogeneous sub-structures demonstrate significant divergence in their importance distribution, as opposed to isomorphic structures that present similar importance patterns. This inspires us to perform isolated ranking and comparison on different types of sub-structures for more reliable pruning. Our empirical results on ImageNet-1K demonstrate that Isomorphic Pruning surpasses several pruning baselines dedicatedly designed for Transformers or CNNs. For instance, we improve the accuracy of DeiT-Tiny from \\(74.52\\%\\) to \\(77.50\\%\\) by pruning an off-the-shelf DeiT-Base model. And for ConvNext-Tiny, we enhanced performance from \\(82.06\\%\\) to \\(82.18\\%\\), while reducing the number of parameters and memory usage. Code is available at https://github.com/VainF/Isomorphic-Pruning." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.647, + 0.671, + 0.662 + ], + "angle": 0, + "content": "Keywords: Network Pruning \\(\\cdot\\) Vision Transformers \\(\\cdot\\) CNNs" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.684, + 0.377, + 0.699 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Deep Neural Networks are typically over-parameterized, containing numerous redundant sub-structures. Such redundancy can be effectively eliminated via structured pruning, without significantly sacrificing the network's performance [16, 22, 25]. Pruning usually follows a \"ranking-to-prune\" framework: To facilitate compression while retaining the learned capabilities of networks, a carefully-designed criterion is utilized to assess the importance score of different structures [11, 22, 40], followed by a ranking process to identify and remove those" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.39, + 0.84 + ], + "angle": 0, + "content": "* Corresponding Author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "G. Fang et al." + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.149, + 0.745, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.353, + 0.788, + 0.41 + ], + "angle": 0, + "content": "Fig. 1: ImageNet Top-1 Accuracy vs. Multiply-Accumulate Operations (MACs) of pruned DeiT [50] and pruned ConvNext [33]. The pruned models, marked as “★” have comparable or better latency yet superior performance compared to scratch training counterparts highlighted by “○”. The MACs are in log scale for better visualization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.784, + 0.474 + ], + "angle": 0, + "content": "relatively unimportant ones. After pruning, the compressed model is fine-tuned to recover its performance [11, 18]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.786, + 0.763 + ], + "angle": 0, + "content": "However, ranking-based structured pruning presupposes an underlying assumption that the parameters from different structures are directly comparable. Classic deep neural networks, such as VGG [46], are built by stacking similar convolutional layers or blocks where the neuron connections share a similar computational topology. Under the circumstances, a straightforward global ranking is expected to reveal their relative importance and remove unimportant parameters [11, 22]. However, with recent innovation in network structure design, advanced vision models like Vision Transformers [10, 31, 50] and ConvNext [33] incorporate more sophisticated designs, such as complicated residual connections [19, 24], attention mechanisms [52] and depth-wise convolutions [6]. These novel structures increase the network's internal complexity and heterogeneity, challenging the reliability of pruning importance criteria. For instance, a vanilla Vision Transformer [10] is composed of MLP, Multi-head Attention, and skip connections. Even in the same Multi-head attention layer, there are two pruning schemes to removing either entire attention heads [37, 53] or the within-head dimensions [58, 68]. These structures vary by their computational topologies and, as will be illustrated in the experiments, tend to exhibit diverged importance distributions. In this case, directly comparing the importance of these heterogeneous might be problematic." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To address the problem, we propose a simple pruning approach, termed Isomorphic Pruning, to alleviate the incomparability issue arising from heterogeneous sub-structures. The essence of this method lies in decomposing a network into multiple isomorphic groups, as illustrated in Figure 2 (d), based on the computational topology of the sub-structures. Each isomorphic group contains" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.475, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Isomorphic Pruning for Vision Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.147, + 0.314, + 0.159 + ], + "angle": 0, + "content": "(a) Network" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.161, + 0.315, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.147, + 0.43, + 0.158 + ], + "angle": 0, + "content": "(b) Local Pruning" + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.161, + 0.45, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.467, + 0.147, + 0.547, + 0.158 + ], + "angle": 0, + "content": "(c) Global Pruning" + }, + { + "type": "image", + "bbox": [ + 0.479, + 0.161, + 0.567, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.608, + 0.147, + 0.709, + 0.159 + ], + "angle": 0, + "content": "(d) Isomorphic Pruning" + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.16, + 0.755, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.245, + 0.788, + 0.33 + ], + "angle": 0, + "content": "Fig. 2: For a pre-trained network (a), we show three pruning strategies: (b) Local Pruning that compares parameter importance within current layers; (c) Global Pruning that performs global ranking for all parameters; (d) the proposed Isomorphic pruning that groups parameters by the computational topology and applies importance ranking within groups. In each group, the importance distributions are more comparable. Details about the distribution can be found in Fig. 4 of the experiments section." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.788, + 0.39 + ], + "angle": 0, + "content": "elements with the same structure and tends to present a similar distribution of importance, making it more rational to compare their relative importance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.788, + 0.602 + ], + "angle": 0, + "content": "To validate the effectiveness of our method, we conducted experiments on a variety of visual models, including DeiT [50], ConvNext [33], ResNet [19], and MobileNet-V2 [44]. Pruning results, as shown in Figure 1, on ImageNet-1K [8] demonstrate that Isomorphic Pruning achieves comparable or even superior performance relative to pruning algorithms specifically designed for CNNs or transformers. For instance, by pruning a pre-trained large DeiT model, we enhanced the performance of the DeiT-T model from \\(74.52\\%\\) to \\(77.50\\%\\), when compared to the scratch training one. Similarly, our algorithm also yields satisfactory results on CNNs. We constructed compact ConvNexts and ResNets via pruning, which has fewer MACs and parameters while surpassing a conventionally trained model. Additionally, we also studied the real-world latency and memory footprint of the compressed models on both GPU and CPU devices. Our findings indicate that models pruned with Isomorphic Pruning achieve significant actual speed-ups and are competitive to carefully designed efficient architectures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.602, + 0.788, + 0.663 + ], + "angle": 0, + "content": "Therefore, Our contribution lies in a practical and general pruning approach, which can handle various vision models with novel mechanisms and architectures. The pruned models show superior performance to several pruning competitors on ImageNet-1K benchmark." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.686, + 0.398, + 0.703 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Network Pruning. Deep neural networks are typically over-parameterized and can be pruned without significant performance drop [16, 25]. The exploration of network pruning can be dated back to early works [1, 17, 25] about neural networks. Over the past few years, a large number of works have been conducted on convolutional networks to verify the effectiveness of pruning, which can be roughly categorized into unstructured [16, 29, 35] and structured pruning [5, 11, 18, 22, 56]. Unstructured pruning compresses a pre-trained model by forcing partial weights to be zero, leading sparse networks [29]. To accelerate" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "G. Fang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "the sparse model, special patterns and devices will be used such as N:M sparsity [12, 42, 48]. Structured pruning provides a hardware-friendly approach by removing sub-structures such as channels [11] or blocks [14] in CNNs, which improves the model efficiency without dedicated accelerators. In recent years, a large number of works adapted the idea of network pruning to vision transformers. It demonstrated that even the vanilla vision transformers can be squeezed for a better efficiency-performance trade-off [2, 4, 45, 47, 58, 60–62, 64, 65, 68, 70]. In the literature, transformer pruning has been studied from several perspectives, such as width reduction [70], depth pruning [61], attention head removal [58], sparse training [4] and architecture transformation [18]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.789, + 0.522 + ], + "angle": 0, + "content": "Pruning by Global Ranking. In the context of network pruning, global pruning is recognized as a pivotal topic for exploration. This method entails a comprehensive ranking of all network layers, assigning adaptive pruning ratios to each layer according to the ranks [5,11,22,58]. Different layers within a network often exhibit varying levels of redundancy. To this end, the advantage of global pruning lies in its ability to create a more parameter-efficient, non-uniformly compressed model [58]. However, a key challenge remains in the ranking: how to ensure the consistency and comparability across different layers? A widely-used solution for this issue is normalizing scores on a per-layer basis to even out the importance distribution across the network [30]. Alternatively, some approaches use sparse training to learn an effective target structure [11], which might be costly due to the additional optimization. In this work, we introduce a more straightforward approach termed isomorphic pruning to facilitate a reliable ranking for pruning." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.545, + 0.331, + 0.56 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.788, + 0.641 + ], + "angle": 0, + "content": "The core idea of Isomorphic Pruning is straightforward: we group sub-structures in a network by their network topology and perform importance ranking within groups. This imposes two questions 1) how to identify removable sub-structures in networks and 2) how to compare the topology graph for grouping and pruning." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.664, + 0.476, + 0.679 + ], + "angle": 0, + "content": "3.1 Sub-structure in Network" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.692, + 0.788, + 0.797 + ], + "angle": 0, + "content": "Structured Pruning aims to remove sub-structures in networks, such as the dimensions of hidden states or features. To make the formulation clear, we primarily concentrate on stacked fully-connected networks, although the principles established herein can be readily extended to CNNs and transformers. Initially, we consider a single linear layer which is parameterized by a matrix \\( W \\in \\mathbb{R}^{m \\times n} \\). To articulate the pruning process, we introduce a pruning function \\( \\mathrm{g}(W,d,k) \\) defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.807, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\hat {w} = \\mathrm {g} (W, d, k) = \\left\\{ \\begin{array}{l l} \\mathrm {g} (W, 0, k) = W _ {[ k,: ]} \\in \\mathbb {R} ^ {1 \\times n}, & \\text {i f p r u n i n g o u t p u t d i m e n s i o n ,} \\\\ \\mathrm {g} (W, 1, k) = W _ {[:, k ]} \\in \\mathbb {R} ^ {m \\times 1}, & \\text {i f p r u n i n g i n p u t d i m e n s i o n ,} \\end{array} \\right. \\tag {1}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.475, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Isomorphic Pruning for Vision Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": "where \\(\\hat{w}\\) refers to the weight column or row slated for removal, acquired via a slicing operation denoted by subscript \\([:,k]\\) or \\([k,:]\\). The pruning function \\(\\mathbf{g}(W,d,k)\\) is defined by three parameters, the parameter matrix \\(W\\), the axis \\(d\\in \\{0,1\\}\\) indicating the dimension to be compressed, and the index \\(k\\) identifying the \\(k\\)-th rows or columns in \\(W\\). In the context of a linear layer represented by a 2-D matrix \\(W\\in \\mathbb{R}^{m\\times n}\\), the axis \\(d\\) can assume values of either 0 or 1, with \\(k\\) having a permissible range of \\([0,m]\\) and \\([0,n]\\), respectively. This formulation encapsulates the pruning of a single layer; to extend it to the entire network, we also need to take the dependency [5,11,30] into account, where parameters from different layers are coupled and should be pruned simultaneously. As depicted in Figure 3, for two adjacent linear layers parameterized by \\(W_{1}\\) and \\(W_{2}\\), the output dimension of the former is integrally coupled to the input dimension of the latter, thereby constituting a minimal sub-structure eligible for removal. If we treat each pruning operation \\(\\mathbf{g}(W,d,k)\\) as vertices defined by a triplet \\((W,d,k)\\) and the dependency as edges, then it is natural to model sub-structures as graphs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.396, + 0.547, + 0.412 + ], + "angle": 0, + "content": "3.2 Graph Modeling of Sub-structures" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.789, + 0.632 + ], + "angle": 0, + "content": "In this section, we delve deeper into the methodology for decomposing a network into several independent and minimal sub-structures in pruning, and model them as graphs. We can begin with the MLP network as illustrated in Figure 3, where the activation functions are omitted. The network comprises a sequence of 2-D parameter matrices \\(\\{W_1, W_2, \\dots, W_L\\}\\), with \\(L\\) denoting the network depth. Initially, all parameters are not assigned to any sub-structure. We randomly select one parameter \\(W\\) as the seed to find its associated substructures and build the corresponding graph. For instance, selecting parameter \\(W_1\\) and setting the axis \\(d = 0\\) allows us to establish an initial graph \\(G = (V = \\{(W, 0, k)\\}, E = \\{\\})\\), focusing on the elimination of \\(k\\)-th the output dimension in \\(W_1\\). Then we iterate through all potential \\((W', d', k') \\notin V\\) to ascertain any dependency to \\((W, d, k) \\in V\\). Dependencies between layers can be ascertained through either manually-crafted patterns or automatic algorithms. In this work, we adhere to the approach of [11] to assess dependency based on two rudimentary rules:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.632, + 0.789, + 0.753 + ], + "angle": 0, + "content": "Dependency Modeling. For a pair of triplets \\((W', d', k')\\), \\((W, d, k)\\), a dependency is established if either of the following criteria is met: 1) \\(W'\\) and \\(W\\) are adjacent layers with \\(d = d'\\) and \\(k = k'\\); 2) \\(W' = W\\) and \\(\\mathrm{g}(W', d', k') = \\mathrm{g}(W, d, k)\\). The first criterion addresses adjacent layers as exemplified in Figure 3, while the second pertains to specialized layers, such as normalization layers, which share the same pruning pattern for both outputs and inputs. Thereafter, we update the vertex set \\(V\\) and edge set \\(E\\) with the discovered dependency between \\((W', d', k')\\) and \\((W, d, k)\\) accordingly:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.765, + 0.785, + 0.788 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} V = V \\cup \\left\\{\\left(W ^ {\\prime}, d ^ {\\prime}, k ^ {\\prime}\\right) \\right\\}, \\\\ E = E + \\left\\{\\left(\\left(W _ {1} - 1\\right), \\left(W ^ {\\prime}, d ^ {\\prime}, 1\\right)\\right) \\right\\} \\end{array} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.784, + 0.623, + 0.799 + ], + "angle": 0, + "content": "\\[\nE = E \\cup \\left\\{\\left(\\left(W, d, k\\right), \\left(W ^ {\\prime}, d ^ {\\prime}, k ^ {\\prime}\\right)\\right) \\right\\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We can further repeat the above process until no additional vertex is incorporated into the current graph. The resultant graph \\( G \\) then serves as a topological" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "G. Fang et al." + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.144, + 0.716, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.31, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Fig. 3: Isomorphic Pruning models sub-structures as graphs, and apply isolated ranking and pruning with isomorphic structures. We highlight three removable substructures in an MLP and show their corresponding graphs below. Substructures 1 & 3 are isomorphic but 2 & 3 are non-isomorphic due to the additional residual connections." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.789, + 0.473 + ], + "angle": 0, + "content": "representation of a single sub-structure associated with the initial seed parameter. To map all substructures onto different graphs, we subsequently embark on a recursive procedure, commencing from those nodes not yet allocated to any sub-structure. Figure 3 furnishes an exemplar of this process, showcasing three distinct substructures labeled as sub-structures 1, 2, and 3." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.495, + 0.546, + 0.511 + ], + "angle": 0, + "content": "3.3 Ranking with Graph Isomorphism" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.629 + ], + "angle": 0, + "content": "With the graph-based representation of network sub-structures, we advance to the ranking phase. This stage involves estimating and ranking the significance of each sub-structure to identify redundancies. A pivotal component of this process is an importance criterion, which quantifies the relative significance of various sub-structures. Previous works in the literature presents numerous well-crafted criteria, including magnitude pruning and Taylor expansion. In this work, we focus on the general importance function \\( I(\\cdot) \\) for a single layer, represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.272, + 0.64, + 0.787, + 0.692 + ], + "angle": 0, + "content": "\\[\nI (W, d, k) = \\left\\{ \\begin{array}{l l} \\| \\mathrm {g} (W, d, k)) \\|, & \\text {i f M a g n i t u d e P r u n i n g} \\\\ \\| \\frac {\\partial \\mathcal {L}}{\\partial \\mathrm {g} (W , d , k)} \\mathrm {g} (W, d, k)) \\| _ {2}, & \\text {i f T a y l o r P r u n i n g} \\\\ \\dots & \\text {O t h e r i m p o r t a n c e C r i t e r i a .} \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.702, + 0.788, + 0.765 + ], + "angle": 0, + "content": "In structural pruning, our objective is to remove sub-structures encompassing parameters across multiple layers. Existing approaches [5,11,30] generally extend the single-layer importance to a holistic measure by aggregating scores across all constituents \\((W,d,k)\\) for the vertexes set, denoted as \\(I^{*}(\\cdot)\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.774, + 0.787, + 0.801 + ], + "angle": 0, + "content": "\\[\nI ^ {*} (G (V, E)) = \\sum_ {(W, d, k) \\in V} I (W, d, k), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "thereby cumulating the importance of all elements within graph \\( G \\). However, the inherent diversity in parameters and computational topology among different" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.475, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Isomorphic Pruning for Vision Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.42 + ], + "angle": 0, + "content": "substructures \\( G \\) introduces potential ambiguities in this metric. Experimental evidence, particularly in the context of a vision transformer as shown in Figure 4, elucidates this issue: the largest substructure encompasses all embedding dimensions, containing a substantial number of parameters, whereas the most minimal sub-structure comprises merely two linear layers. Consequently, such aggregation may fail to accurately reflect the true relative importance of these entities. To address the challenges posed by the heterogeneity of substructures, we delineate a simple approach designed to preserve the nuanced information encapsulated by the aggregated importance, as outlined in Equation 4, whilst alleviating the detrimental repercussions raised by the heterogeneity of substructures. Our principal strategy entails the clustering of sub-structures into various isomorphic groups, predicated on their topological logic and parameter configurations. In this case, applying aggregated importance within groups can be meaningful, since isomorphic sub-structures show one-to-one correspondence among their parameters, and how different parameters participate in computations is entirely identical. To facilitate this, the key step is to examine the isomorphism of two graphs \\( G_{1} \\) and \\( G_{2} \\). Formally, we adopt the following definition to ascertain graph isomorphism." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.426, + 0.79, + 0.473 + ], + "angle": 0, + "content": "Definition 1 (Graph Isomorphism). Two labeled graphs \\( G \\) and \\( G' \\) are isomorphic if there exists a bijection between their vertices so that, all mapped vertices preserve the same labels, and are connected in the same way." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.48, + 0.789, + 0.511 + ], + "angle": 0, + "content": "To develop a general indicator Isomorphic \\((\\mathrm{g}_1,\\mathrm{g}_2)\\), we need to examine both the label and connectivity in two graphs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.788, + 0.773 + ], + "angle": 0, + "content": "Vertex Label and Connectivity. In our settings, each vertex is labeled by the triplet \\((W,d,k)\\), encompassing the parameter matrix \\(\\mathbf{W}\\), axis \\(d\\), and index \\(k\\). We follow a simple principle for labling: Two vertices \\((\\mathbf{W},d,k)\\) and \\((\\mathbf{W}',d',k')\\) are labeled identically if and only if the parameter \\(\\mathbf{W}'\\) originates from the same layer type such as Linear or Conv, with the same pruning dimension \\(d = d'\\). We permit the index \\(k\\) to differ, since within the same parameter matrix, the weight vectors sliced along the \\(d\\)-th dimension are consistently homogeneous and thus can be considered isomorphic. After determining the label for each vertex, we further explore the connectivity in two graphs \\(G_{1}\\) and \\(G_{2}\\). As illustrated in Figure 3, we show three substructures alongside their respective graphs \\(G_{1}\\), \\(G_{2}\\), and \\(G_{3}\\), where \\(G_{1}\\) and \\(G_{3}\\) incorporate a singular edge, whereas \\(G_{2}\\) encompasses three edges. If we check the edges following the execution order of the computational graphs, it's natural to observe the isomorphism of \\(G_{1}\\) and \\(G_{3}\\), since they share the same topology. To make this process more formal, we assume the edges in graph \\(G\\) are already sorted according to their execution order in forwarding, and check the isomorphism by" + }, + { + "type": "equation", + "bbox": [ + 0.249, + 0.785, + 0.786, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\operatorname {I s o m o r p h i c} \\left(G, G ^ {\\prime}\\right)) = \\mathbb {1} \\left\\{\\left| G \\right| = \\left| G ^ {\\prime} \\right| \\wedge \\operatorname {l a b e l} \\left(V _ {i}\\right) = \\operatorname {l a b e l} \\left(V _ {i} ^ {\\prime}\\right); \\forall i \\in \\{1, 2, \\dots , \\left| G ^ {\\prime} \\right| \\} \\right\\},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}\\) is an indicator function. \\(|G^{\\prime}|\\) denotes the number of edges in the graph and \\(\\wedge\\) signifies the logical AND operation. The first condition ensures the equiva" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "G. Fang et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.146, + 0.47, + 0.163 + ], + "angle": 0, + "content": "Algorithm 1 Isomorphic Pruning" + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.165, + 0.755, + 0.453 + ], + "angle": 0, + "content": "1: Input: Network with parameters \\(\\{W_1, W_2, \\ldots, W_n\\}\\), Pruning ratio \\(p\\%\\). \n2: Output: Pruned Network \\(\\{W_1', W_2', \\ldots, W_n'\\}\\) \n3: procedure IDENTIFYSUBSTRUCTURES(\\{W1, W2, \\ldots, Wn\\}) \n4: Initialize empty list of sub-structures \\(\\mathbf{S} = \\{\\}\\) \n5: for each layer \\(W_i\\) in \\(\\{W_1, W_2, \\ldots, W_n\\}\\) do \n6: if \\(W_i\\) is not assigned to any substructure then \n7: Initialize a graph \\(G(V = \\{W_i\\}, E = \\{\\})\\). \n8: Find and add coupled \\(W_j\\) to \\(G\\) until there is no new dependency. \n9: \\(\\mathbf{S} = \\mathbf{S} \\cup \\{G\\}\\) \n10: return \\(\\mathbf{S}\\) \n11: procedure PRUNEBYGROUPEDRANKING(S, p) \n12: for each substructure \\(G_i\\) in \\(\\{G_1, G_2, \\ldots, G_n\\}\\) do \n13: if \\(G_i\\) is not assigned to any group then \n14: Initialize a Group \\(\\mathbf{R} = \\{G_i\\}\\). \n15: Find and add \\(G_j\\) that satisfies Isomorphism \\((G_i, G_j) = 1\\) to \\(\\mathbf{R}\\). \n16: Calculate the importance scores \\(\\{I^*(G) | G \\in \\mathbf{R}\\}\\). \n17: Remove \\(p\\%\\) substructures in \\(\\mathbf{R}\\) by ranking the importance scores. \n18: return \\(\\mathbf{S}\\) \n19: \\(\\mathbf{S} \\gets \\text{IDENTIFYSUBSTRUCTURES}(\\{W_1, W_2, \\ldots, W_n\\})\\) \n20: \\(\\mathbf{S} \\gets \\text{PRUNEBYGROUPEDRANKING}(\\mathbf{S}, p)\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.785, + 0.514 + ], + "angle": 0, + "content": "lence of the number of vertices in the graphs, while the second condition confirms the matching labels of vertices on either side of the edge pairs \\( V_{i} \\) and \\( V_{i}^{\\prime} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.787, + 0.681 + ], + "angle": 0, + "content": "Using Equation 5, it is easy to cluster substructures into different groups and perform rankings within each group exclusively. Specifically, we examine a collection of substructures \\(\\{G_1, G_2, \\ldots, G_N\\}\\) that share identical graph topologies. We estimate the aggregated significance of each graph independently using Equation 4, resulting in an \\(N\\)-dimensional importance vector \\([I^*(G_1), \\ldots, I^*(G_N)]\\). This facilitates a straightforward ranking of the importance vector, followed by the elimination of the \\(\\mathfrak{p}\\%\\) unimportant substructures based on their scores. This ranking method is deemed reliable since all substructures within a group possess the same architectural design, parameter scale, and computational graph. A similar principle can be uniformly applied to other isomorphic groups for pruning, as detailed in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.704, + 0.376, + 0.722 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.84 + ], + "angle": 0, + "content": "This section reports our empirical results on the ImageNet-1K dataset [8], which comprises 1,281,167 images for training and 50,000 for validation. We deploy the proposed Isomorphic Pruning to compress several vision models, including ConvNext [33], ResNet [36], MobileNetv2 [44], and Vision Transformers [10, 50]. We report the parameter amount (#Params), MACs, and Top-1 validation accuracy in Table 1 and 2, and report the actual latency and peak memory consumption in Table 3." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.475, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Isomorphic Pruning for Vision Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.475, + 0.162 + ], + "angle": 0, + "content": "4.1 Settings for ImageNet-1K" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.355 + ], + "angle": 0, + "content": "Pruning. For simplicity, we adopt the Taylor-based pruning framework [39], which applies a data-driven Taylor expansion to the loss function. For this purpose, we randomly sampled 100 mini-batches from the ImageNet dataset, each containing 64 images. These images were resized to \\(256 \\times 256\\) pixels and then center-cropped to \\(224 \\times 224\\) pixels. We normalize all images using the mean and standard deviation specific to ImageNet. Notably, while some methods, such as ConvNext such as ConvNext [33] and DeiT [50] employ additional training techniques like label smoothing and knowledge distillation, our approach only leverages a straightforward cross-entropy loss for the Taylor expansion. We accumulate gradients from all the batches for a more reliable importance estimation and apply Isomorphic pruning in a one-shot manner without iterative pruning and fine-tuning [40]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.789, + 0.589 + ], + "angle": 0, + "content": "Finetuning. During the fine-tuning process of pruned models, we adhere to the original training protocol for a fair comparison. While maintaining most of the original settings unchanged, we adjusted the learning rates and batch sizes to make the training suitable for A5000 GPUs, following the linear LR scaling rule proposed in [13]. For example, to compressing a pre-trained DeiT models [50], we adopt the same augmentation strategies for training, including RandAugmentation [7], Mixup [67], CutMix [66], Random Erasing [69], Repeated Augmentation [23], label smoothing [49]. The pruned models are distilled with a pre-trained RegNetY [43, 50] for 300 epochs using the AdamW [34] optimizer, with a learning rate of 0.0005 and a total batch size of 2048. A weight decay of 0.05 and a cosine annealing scheduler is deployed for training. For other models like ConvNext [33], ResNet [19] and MobileNet-v2 [44], the same principle will be applied. More details of hyper-parameters, pruning, and finetuning can be found in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Evaluation. We report the optimal classification accuracy achieved on the ImageNet-1K validation set, employing the standard resize-and-crop protocol [19, 33] for evaluation. To further validate the efficiency of models, we profiled the latency and the peak memory consumption of the pruned models on both GPU (a single RTX A5000) and CPU. To guarantee a fair and standardized comparison, we also tested the latency of several baseline models sourced from the Pytorch-Image Models [57] and Torchvision [38], following the same testing procedure. The latency was estimated with a batch size of 256 for GPU and 8 for CPU, averaged on 100 repeats. Appropriate GPU warmup was deployed in our test." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.597, + 0.785 + ], + "angle": 0, + "content": "4.2 Pruning Convolutional Neural Networks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.843 + ], + "angle": 0, + "content": "Convolutional Neural Network is a critical focus in the area of network pruning. In this section, we concentrate on pruning three popular ConvNets: ConvNext [33], ResNet [19] and MobileNet-v2 [44]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "G. Fang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.359, + 0.145, + 0.645, + 0.159 + ], + "angle": 0, + "content": "Table 1: Pruning CNNs on ImageNet-1K." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.161, + 0.724, + 0.535 + ], + "angle": 0, + "content": "
Pruning Method#Params (M)MACs (G)Base Acc.Final Acc.Δ Acc.
ConvNext-B† [33]88.5915.39-83.83-
ConvNext-S† [33]50.228.71-83.14-
ConvNext-S (Ours)47.368.4883.8383.17-0.66
ConvNext-T† [33]28.594.47-82.06-
ConvNext-T (Ours)25.324.1983.8382.19-1.64
ResNet-152†60.1911.58-78.31-
AOFP-E2 [9]-4.1377.3777.00-0.37
ABCPruner [28]24.074.3178.3177.12-1.19
EPruner-0.63 [27]21.564.0878.3176.83-1.48
ResNet-152-4.0G (Ours)23.114.0578.3177.84-0.47
ResNet-101†44.557.85-77.38-
IE [39]31.24.7077.3777.35-0.02
FPGM [21]-4.5177.3777.32-0.05
SFP [20]-4.5177.3777.51+0.14
ISTA [59]-4.4776.4075.27-1.13
Res101-4.5G (Ours)29.144.4877.3877.56+0.16
GFP [30]28.023.9078.2978.33+0.04
AOFP [9]-3.8976.6376.40-0.23
Res101-3.8G (Ours)24.873.8577.3877.43+0.05
ResNet-50† [19]25.564.13-76.13-
Taylor [39]14.202.2576.1874.50-1.68
CCP [41]-2.1176.1575.50-0.65
GFP [30]19.422.0476.7976.42-0.37
AutoSlim [63]20.602.0076.1075.60-0.50
DepGraph [11]-1.9976.1575.83-0.32
ResNet50-2G (Ours)15.052.0676.1375.91-0.22
Mobv2†3.500.32-71.88-
Meta [32]-0.1474.7068.20-6.50
GFP [30]-0.1575.7469.16-6.58
DepGraph [11]-0.1571.9068.49-3.41
Mobv2-0.15G (Ours)2.190.1571.8868.91-2.97
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.79 + ], + "angle": 0, + "content": "ConvNext. Our experiments in Table 1 begin with the evaluation on an advanced CNN, ConvNext [33], which achieves impressive performance on several benchmarks. We use the official ConvNext-Base [38] as the pre-trained model for pruning, which achieves an accuracy of \\(83.83\\%\\) on the ImageNet-1K validation set. The pruning of ConvNext models is a relatively underexplored topic. Hence, we establish our baselines using the official variants with fewer parameters, ConvNext-S and ConvNext-T. For ConvNext, we also generalize our method to depth pruning, which removes unimportant layers with the minimal average scores on substructures, so that the pruned models have the same depth as baselines. The pruned models are fine-tuned for 300 epochs following the same training protocol as [33]. During fine-tuning, we observed that the pruned ConvNext with half of the parameters removed, achieves an accuracy of 80.09 in only 20 epochs, which effectively validates the efficiency of pruning. Moreover, as illustrated in Table 1, the final model after 300 full fine-tuning, achieves competitive performance (82.19% vs. 82.06%) to pre-trained models with fewer parameters (25.32M vs. 28.59M) and MACs (4.19G vs. 4.47G)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.809, + 0.787, + 0.84 + ], + "angle": 0, + "content": "ResNet. In line with prior studies [5, 11, 22, 30], we demonstrate the efficacy of our method by structurally pruning pre-trained ResNet MACs. ResNet-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.476, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Isomorphic Pruning for Vision Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.359, + 0.145, + 0.645, + 0.159 + ], + "angle": 0, + "content": "Table 2: Pruning DeiTs on ImageNet-1K." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.161, + 0.777, + 0.369 + ], + "angle": 0, + "content": "
Method#Params (M) MACs (G) Acc (%)Method#Params (M) MACs (G) Acc (%)
DeiT-B† [50]87.3417.6983.32DeiT-T† [50]5.911.2774.52
DeiT-S† [50]22.444.6481.17DeiT-T-600EP [58]5.911.2775.00
DeiT-S-600EP [58]22.444.6481.80WDPruning [61]13.302.6078.38
VTP - 40% [70]48.0010.0080.70X-Pruner [64]-2.4078.93
WDPruning [61]55.309.9080.76UVC [65]-2.3278.82
UVC [65]-8.0080.57ViT-Slim [70]11.402.3077.94
X-Pruner [64]-8.5081.02ViT-Slim [70]11.402.3077.94
NViT-B [58]34.006.8083.29NViT-T [58]6.901.3076.21
SAViT [68]25.405.3081.66SAViT+KD [68]6.601.3076.95
UP-DeiT-S [62]22.10-81.56UP-DeiT-T [62]5.70-75.79
NViT-S [58]21.004.2082.19DeiT-T (Ours)5.741.2177.50
Upop [45]19.904.1081.10VTC-LFC [55]5.10-71.60
DeiT-S (Ours)20.694.1682.41SViT-E [68]4.210.9970.12
GOHSP [60]14.402.9979.98GOHSP [60]4.000.9170.24
GOHSP [60]11.102.8179.86SAViT [68]4.200.9070.72
ViT-Slim [2]13.502.8079.20CP-ViT [47]-0.7471.24
Upop [45]13.502.8079.60WDPruning [61]3.500.7070.34
ViT-Slim [2]13.502.8079.20X-Pruner [64]-0.6071.10
CP-ViT [47]-2.6679.08UVC [65]-0.5170.60
DeiT-2.6G (Ours)13.072.6281.13DeiT-0.6G (Ours)3.080.6272.60
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.788, + 0.593 + ], + "angle": 0, + "content": "50/101/152 consists of four residual blocks, each formed by several Bottle-neck structures and residual connections [19]. The proposed Isomorphic Pruning identifies five kinds of distinct isomorphic groups within ResNet-50, four targeting the block structures and one associated with bottleneck structures [19]. It is worthwhile to note that different sub-structures vary in their parameter and MAC counts. For example, residual blocks naturally contain more parameters than the bottleneck structures which are only composed of a few convolutional layers. The conventional global pruning, employing a naive global ranking threshold, faces the potential risks of over-pruning certain layers due to the biased importance distribution. Isomorphic Pruning, however, employs a more stable scheme for pruning. As illustrated in Table 1, we report both the accuracy and the accuracy drop compared to their base models. Isomorphic pruning achieves lossless compression on ResNet-101, under both 4.5G and 3.8G settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.79, + 0.677 + ], + "angle": 0, + "content": "MobileNet-v2. We also conduct experiments on lightweight models, MobileNet-v2, by compressing its MACs from 0.32 G to 0.15 G. In MobileNet-v2, we can also discover 5 isomorphic groups in MobileNet. The pruned model achieves better accuracy compared to DepGraph [11] which requires sparse training." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.695, + 0.507, + 0.71 + ], + "angle": 0, + "content": "4.3 Pruning Vision Transformers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.787, + 0.81 + ], + "angle": 0, + "content": "Vision Transformers (ViT) has profoundly reshaped the paradigm of visual modeling over the past few years. Nevertheless, their inherent over-parameterization often imposes huge overheads to both training and inference. Compared to CNNs that are composed of similar convolution layers or blocks, Vision Transformers are usually built with MLP, Multi-head Attention, and skip connections, which are more challenging for pruning due to the heterogeneous internal structures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Table 2 presents our pruning results on DeiT, a plain vision transformer enhanced with knowledge distillation [50]. We benchmark our approach with" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "G. Fang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 3: Profiling the pruned models and other off-the-shelf vision models from PyTorch-Image-Models [57] on a single RTX A5000 GPU and CPU. For NViT [58], we report their best speed up in the original paper, tested on a single V100 or RTX 3080." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.199, + 0.785, + 0.503 + ], + "angle": 0, + "content": "
Method#Params ↓ MACs ↓ Peak Mem. ↓ GPU Latency (256) ↓ CPU Latency (8) ↓ Acc. (%)
DeiT-B [50]86.40 M17.60 G2909.78 MB802.97 ms (1.00×)197.73 ms (1.00×)83.32
DeiT-S [50]22.44 M4.64 G1363.47 MB241.81 ms (3.32×)80.64 ms (2.45×)81.17
Swin-T [31]28.29 M4.51 G4257.98 MB375.45 ms (2.14×)203.86 ms (0.97×)81.19
NViT-S [58]21.00 M4.20 G--(2.52×)-82.19
TNT-S [15]23.76 M4.85 G1457.56 MB631.02 ms (1.27×)175.58 ms (1.13×)81.50
CaiT-XS24 [51]26.56 M5.40 G1496.42 MB557.41 ms (1.44×)342.28 ms (0.58×)82.00
CrossViT-S [3]26.86 M5.64 G1546.94 MB315.02 ms (2.55×)115.51 ms (1.71×)81.00
EfficientFormer-L3 [26]31.41 M3.94 G2681.75 MB249.05 ms (3.22×)167.16 ms (1.18×)82.40
PVTv2-B2 [54]25.36 M4.00 G4478.18 MB364.00 ms (2.21×)212.44 ms (0.93×)82.00
DeiT-S (Ours)20.69 M4.16 G1547.97 MB230.84 ms (3.48×)44.10 ms (4.48×)82.41
DeiT-2.6G (Ours)13.07 M2.62 G1203.07 MB189.68 ms (4.23×)56.59 ms (3.49×)81.13
DeiT-T [50]5.91 M1.27 G720.22 MB93.78 ms (8.56×)46.83 ms (4.22×)74.52
NViT-T [58]6.90 M1.30 G--(4.97×)-76.21
DeiT-T (Ours)5.71 M1.20 G837.16 MB93.17 ms (8.62×)47.89 ms (4.13×)77.50
DeiT-0.6G (Ours)3.08 M0.62 G629.28 MB63.23 ms (12.70×)40.98 ms (4.83×)72.60
ConvNext-B [33]88.59 M15.36 G5394.10 MB789.24 ms (1.00×)262.58 ms (1.00×)83.83
ConvNext-S [33]50.22 M8.69 G3958.45 MB508.20 ms (1.57×)188.07 ms (1.40×)83.14
ConvNext-S (Ours)47.36 M8.48 G3788.65 MB535.95 ms (1.59×)179.72 ms (1.46×)83.17
ConvNext-T [33]28.59 M4.46 G3785.37 MB291.02 ms (2.82×)126.17 ms (2.08×)82.06
ConvNext-T (Ours)25.32 M4.19 G3607.06 MB292.77 ms (2.83×)97.82 ms (2.68×)82.19
ResNet152 [19]60.19 M11.58 G3320.24 MB386.65 ms (1.00×)272.21 ms (1.00×)78.31
ResNet152-4G (Ours)23.11 M4.06 G1938.91 MB238.12 ms (1.62×)180.65 ms (1.51×)77.84
ResNet101 [19]44.55 M7.85 G3196.77 MB271.34 ms (1.00×)202.66 ms (1.00×)77.38
ResNet101-4.5G (Ours)29.14 M4.48 G2639.40 MB220.72 ms (1.23×)162.33 ms (1.25×)77.56
ResNet101-3.8G (Ours)24.87 M3.85 G2286.36 MB194.39 ms (1.40×)158.98 ms (1.28×)77.43
ResNet50 [19]25.56 M4.13 G3035.84 MB171.54 ms (1.00×)114.86 ms (1.00×)76.13
ResNet50-2.0G (Ours)15.05 M2.06 G2238.15 MB78.75 ms (1.33×)80.88 ms (1.42×)75.91
Mobv2 [44]3.50 M0.32 G2855.44 MB71.42 ms (1.00×)74.46 ms (1.00×)71.88
Mobv2-0.15G (Ours)2.19 M0.15 G1354.82 MB43.48 ms (1.64×)38.63 ms (1.93×)68.91
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.688 + ], + "angle": 0, + "content": "a range of innovative pruning baselines specifically tailored for vision transformers [45, 47, 50, 55, 58, 60, 61, 64, 65, 68, 70]. A DeiT model consists of various components, such as MLP and attention layers. Within the attention layers, we can even find two feasible pruning schemes: pruning 1) the number of heads or 2) within-head dimensions. We implement Isomorphic Pruning to compress a pre-trained DeiT-Base model by pruning both heads, head dimensions, and the embedding sizes following [58]. This is achieved through a straightforward first-order Taylor expansion for importance estimation, followed by a grouped ranking in Isomorphic Pruning. The details of pruning vision transformers can be found in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We developed four lightweight vision transformers: DeiT-S, DeiT-2.6G, DeiT-T and DeiT-0.6G. The pruned DeiT-S and DeiT-T models achieve comparable latencies comparable to the standard DeiT-S and DeiT-T with manually designed architecture. The empirical results are reported in Table 2, which shows that our method yields greater performance gains in complex models with extensive heterogeneous internal structures, especially in comparison to CNNs. Specifically, the optimized DeiT-T achieves an accuracy of \\(77.50\\%\\), with only 1.2G MACs, which is better than the results \\((74.52\\%)\\) of the pre-trained DeiT-T with uniform width across layers. Detailed configurations, such as base models and pruning ratios, are available in the appendix." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.476, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Isomorphic Pruning for Vision Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.146, + 0.382, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.312, + 0.323, + 0.324 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.146, + 0.528, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.45, + 0.312, + 0.47, + 0.323 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.146, + 0.754, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.312, + 0.658, + 0.324 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.335, + 0.785, + 0.432 + ], + "angle": 0, + "content": "Fig. 4: (a-b) Top-1 Accuracy and loss of pruning ResNet-50 on ImageNet-1K validation set without finetuning. We report the mean and standard derivation of 10 experiments. IsomorphicP Pruning consistently improves the performance of pruned models. (c) Histogram of Taylor importance score of DeiT sub-structures. Isomorphic structures are highlighted with the same color. Zoom in for more details about each structure. We visualize the threshold of \\(50\\%\\) pruning ratio for naive global pruning (the black dash) and isomorphic pruning (the colored dash)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.462, + 0.523, + 0.477 + ], + "angle": 0, + "content": "4.4 Analysis on the Pruned Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.789 + ], + "angle": 0, + "content": "On-device Latency Table 3 measures the actual latency and peak memory consumptions on a single RTX A5000 GPU and CPU. We evaluate all models with a batch size of 256 for GPU and 8 for CPU, respectively. The average latency over 100-round experiments are reported, with a 20-step warmup process before evaluation. Firstly, results show that the pruned DeiT models can achieve comparable performance (82.41% vs. 82.40%) to some advanced transformers like EfficientFormer [26], while preserving a competitive latency. Additionally, we observe that the peak memory consumption is primarily determined by the largest layer width. Therefore, under the same budget, a uniform DeiT may require less peak memory (1363 MB for DeiT-S) for inference. In contrast, a non-uniform architecture may require slightly more memory (1547 MB for the DeiT-S Pruned) under the given budget. But non-uniform pruning can achieve better accuracy, which reveals an important trade-off between accuracy and memory consumption for transformers. In addition, for CNNs like ResNets, the pruned model requires much less memory due to the compressed feature dimensions. Furthermore, we note that the actual speed-up may not be linearly correlated with the MAC reductions. This is particularly evident in ResNet-50, where only a \\(1.33 \\times\\) actual speed-up is achieved despite the removal of half of the MACs. In contrast, compression on Mobv2 is more effective, achieving a \\(1.64 \\times\\) acceleration with a similar compression ratio on MACs." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Distribution of Importance Scores. To further explore how Isomorphic Pruning affects the pruning behavior, we study the distribution of importance" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "G. Fang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "score in a DeiT-Base model [50]. As shown in Figure 4 (c), the importance histogram of different sub-structures is highlighted with different colors, each corresponding to one kind of isomorphic structure. It is natural to find that, the MLP layer and the attention layer have diverged importance distributions due to their different computation process. Besides, when we apply a simple \\(50\\%\\) global pruning on the DeiT model, most parameters in Attentions \\(\\# 5\\), \\(\\# 9\\), and \\(\\# 13\\) will be removed, since the importance scores in attentions are centered around 0. In contrast, the embedding number as \\(\\# 1\\) will not be even been pruned in this case, due to the relatively large importance scores. This inevitably leads to an imbalanced pruning of different structures. Notably, the embedding structure, as illustrated in the appendix, contains massive parameters in the DeiT, where biased pruning may lead to very limited acceleration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.489 + ], + "angle": 0, + "content": "Importance Criteria. Note that Isomorphic Pruning can be combined with different pruning criteria like Taylor [39], Hessian [25] and even L1-norm [22]. We quantify the performance gain of Isomorphic Pruning by visualizing the accuracy and loss curve with different pruning ratios in Figure 4 (a) and (b). Specifically, we gradually increase the pruning ratio for each criterion and report the average accuracy and loss without finetuning over 10 repeated runs with random inputs for Taylor expansion [25, 39]. Note that L1 pruning is deterministic and thus presents 0 std. We find the Isomorphic Pruning, marked as solid lines, consistently improves the counterpart baselines highlighted in dashed lines." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.546, + 0.678 + ], + "angle": 0, + "content": "Transfer Learning. We also evaluate the performance of transfer learning based on our pruned models. Table 4 reports the accuracy of transferring different models to CIFAR-10 and CIFAR-100. We resize all images to \\(224 \\times 224\\) and apply the same training recipe for transfer. For the DeiT models, we did not include the supervision from the teacher models. It is observed that our method achieves better transfer accuracy compared to the scratch training model and baseline methods." + }, + { + "type": "table_caption", + "bbox": [ + 0.555, + 0.527, + 0.787, + 0.553 + ], + "angle": 0, + "content": "Table 4: Transfer learning on pruned models." + }, + { + "type": "table", + "bbox": [ + 0.56, + 0.555, + 0.784, + 0.662 + ], + "angle": 0, + "content": "
ModelsCIFAR-10CIFAR-100
DeiT-S [39]98.5287.07
NViT-S [58]98.7887.90
DeiT-S (Ours)99.0290.09
DeiT-T93.9385.66
NViT-T [58]98.3185.88
DeiT-S (Ours)98.2887.08
ConvNext-S98.7789.15
ConvNext-S (Ours)98.7889.36
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.702, + 0.36, + 0.718 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we introduce Isomorphic Pruning, a practical approach designed to compress vision networks with novel mechanisms and architectures. This method focuses on mitigating the challenges posed by heterogeneous sub-structures in networks, thereby enhancing the reliability of ranking and pruning processes. Our empirical results on the ImageNet-1k dataset, showcase the effectiveness of Isomorphic Pruning across several vision models like ConvNext, ResNet, MobileNetv2, and Vision Transformers." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.476, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Isomorphic Pruning for Vision Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.252 + ], + "angle": 0, + "content": "This project is supported by the National Research Foundation Singapore under its Medium Sized Center for Advanced Robotics Technology Innovation, and the Advanced Research and Technology Innovation Centre (ARTIC), the National University of Singapore under Grant (project number: A-0005947-21-00, project reference: ECT-RP2)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.274, + 0.323, + 0.289 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.303, + 0.785, + 0.331 + ], + "angle": 0, + "content": "1. Baum, E., Haussler, D.: What size net gives valid generalization? Advances in neural information processing systems 1 (1988)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.785, + 0.386 + ], + "angle": 0, + "content": "2. Chavan, A., Shen, Z., Liu, Z., Liu, Z., Cheng, K.T., Xing, E.P.: Vision transformer slimming: Multi-dimension searching in continuous optimization space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4931-4941 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.388, + 0.785, + 0.428 + ], + "angle": 0, + "content": "3. Chen, C.F.R., Fan, Q., Panda, R.: Crossvit: Cross-attention multi-scale vision transformer for image classification. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 357-366 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.429, + 0.785, + 0.469 + ], + "angle": 0, + "content": "4. Chen, T., Cheng, Y., Gan, Z., Yuan, L., Zhang, L., Wang, Z.: Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems 34, 19974-19988 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.47, + 0.785, + 0.496 + ], + "angle": 0, + "content": "5. Chen, T., Liang, L., Ding, T., Zhu, Z., Zharkov, I.: Otov2: Automatic, generic, user-friendly. arXiv preprint arXiv:2303.06862 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.497, + 0.785, + 0.537 + ], + "angle": 0, + "content": "6. Chollet, F.: Xception: Deep learning with depthwise separable convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1251-1258 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.538, + 0.785, + 0.592 + ], + "angle": 0, + "content": "7. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 702-703 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.594, + 0.785, + 0.634 + ], + "angle": 0, + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.635, + 0.785, + 0.675 + ], + "angle": 0, + "content": "9. Ding, X., Ding, G., Guo, Y., Han, J., Yan, C.: Approximated oracle filter pruning for destructive cnn width optimization. In: International Conference on Machine Learning. pp. 1607-1616. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.676, + 0.785, + 0.731 + ], + "angle": 0, + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.731, + 0.785, + 0.772 + ], + "angle": 0, + "content": "11. Fang, G., Ma, X., Song, M., Mi, M.B., Wang, X.: Depgraph: Towards any structural pruning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16091-16101 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.773, + 0.785, + 0.799 + ], + "angle": 0, + "content": "12. Frantar, E., Alistarh, D.: Sparsegpt: Massive language models can be accurately pruned in one-shot (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Goyal, P., Dolkar, P., Girshick, R., Noordhuis, P., Wesolowski, L., Kyrola, A., Tulloch, A., Jia, Y., He, K.: Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.303, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "G. Fang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "14. Guo, J., Ouyang, W., Xu, D.: Multi-dimensional pruning: A unified framework for model compression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1508-1517 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.786, + 0.218 + ], + "angle": 0, + "content": "15. Han, K., Xiao, A., Wu, E., Guo, J., Xu, C., Wang, Y.: Transformer in transformer. Advances in Neural Information Processing Systems 34, 15908-15919 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.786, + 0.259 + ], + "angle": 0, + "content": "16. Han, S., Mao, H., Dally, W.J.: Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. International Conference on Learning Representations (ICLR) (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.3 + ], + "angle": 0, + "content": "17. Hassibi, B., Stork, D.G., Wolff, G.J.: Optimal brain surgeon and general network pruning. In: IEEE international conference on neural networks. pp. 293-299. IEEE (1993)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "18. He, H., Liu, J., Pan, Z., Cai, J., Zhang, J., Tao, D., Zhuang, B.: Pruning self-attention into convolutional layers in single path. arXiv preprint arXiv:2111.11802 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.787, + 0.383 + ], + "angle": 0, + "content": "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.787, + 0.411 + ], + "angle": 0, + "content": "20. He, Y., Kang, G., Dong, X., Fu, Y., Yang, Y.: Soft filter pruning for accelerating deep convolutional neural networks. arXiv preprint arXiv:1808.06866 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.787, + 0.453 + ], + "angle": 0, + "content": "21. He, Y., Liu, P., Wang, Z., Hu, Z., Yang, Y.: Filter pruning via geometric median for deep convolutional neural networks acceleration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4340-4349 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.495 + ], + "angle": 0, + "content": "22. He, Y., Zhang, X., Sun, J.: Channel pruning for accelerating very deep neural networks. In: Proceedings of the IEEE international conference on computer vision. pp. 1389-1397 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.549 + ], + "angle": 0, + "content": "23. Hoffer, E., Ben-Nun, T., Hubara, I., Giladi, N., Hoefler, T., Soudry, D.: Augment your batch: Improving generalization through instance repetition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8129-8138 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.549, + 0.787, + 0.592 + ], + "angle": 0, + "content": "24. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.619 + ], + "angle": 0, + "content": "25. LeCun, Y., Denker, J., Solla, S.: Optimal brain damage. Advances in neural information processing systems 2 (1989)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.787, + 0.66 + ], + "angle": 0, + "content": "26. Li, Y., Yuan, G., Wen, Y., Hu, J., Evangelidis, G., Tulyakov, S., Wang, Y., Ren, J.: Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems 35, 12934-12949 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.787, + 0.702 + ], + "angle": 0, + "content": "27. Lin, M., Ji, R., Li, S., Wang, Y., Wu, Y., Huang, F., Ye, Q.: Network pruning using adaptive exemplar filters. IEEE Transactions on Neural Networks and Learning Systems 33(12), 7357-7366 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.787, + 0.73 + ], + "angle": 0, + "content": "28. Lin, M., Ji, R., Zhang, Y., Zhang, B., Wu, Y., Tian, Y.: Channel pruning via automatic structure search. arXiv preprint arXiv:2001.08565 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.787, + 0.772 + ], + "angle": 0, + "content": "29. Liu, B., Wang, M., Foroosh, H., Tappen, M., Pensky, M.: Sparse convolutional neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 806-814 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "30. Liu, L., Zhang, S., Kuang, Z., Zhou, A., Xue, J.H., Wang, X., Chen, Y., Yang, W., Liao, Q., Zhang, W.: Group fisher pruning for practical network compression. In: International Conference on Machine Learning. pp. 7021-7032. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "31. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.476, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Isomorphic Pruning for Vision Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.787, + 0.217 + ], + "angle": 0, + "content": "32. Liu, Z., Mu, H., Zhang, X., Guo, Z., Yang, X., Cheng, K.T., Sun, J.: Metapruning: Meta learning for automatic neural network channel pruning. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3296-3305 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.786, + 0.259 + ], + "angle": 0, + "content": "33. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11976-11986 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "34. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "35. Louizos, C., Welling, M., Kingma, D.P.: Learning sparse neural networks through \\( l\\_0 \\) regularization. arXiv preprint arXiv:1712.01312 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.37 + ], + "angle": 0, + "content": "36. Ma, X., Yuan, G., Lin, S., Li, Z., Sun, H., Wang, Y.: Resnet can be pruned \\(60 \\times\\): Introducing network purification and unused path removal (p-rm) after weight pruning. In: 2019 IEEE/ACM International Symposium on Nanoscale Architectures (NANOARCH). pp. 1-2. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.786, + 0.397 + ], + "angle": 0, + "content": "37. Ma, X., Fang, G., Wang, X.: Llm-pruner: On the structural pruning of large language models. In: Advances in Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.786, + 0.424 + ], + "angle": 0, + "content": "38. maintainers, T., contributors: Torchvision: Pytorch's computer vision library. GitHub repository (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.786, + 0.467 + ], + "angle": 0, + "content": "39. Molchanov, P., Mallya, A., Tyree, S., Frosio, I., Kautz, J.: Importance estimation for neural network pruning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11264-11272 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.786, + 0.508 + ], + "angle": 0, + "content": "40. Molchanov, P., Tyree, S., Karras, T., Aila, T., Kautz, J.: Pruning convolutional neural networks for resource efficient inference. arXiv preprint arXiv:1611.06440 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.786, + 0.55 + ], + "angle": 0, + "content": "41. Peng, H., Wu, J., Chen, S., Huang, J.: Collaborative channel pruning for deep networks. In: International Conference on Machine Learning. pp. 5113-5122. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.786, + 0.605 + ], + "angle": 0, + "content": "42. Pool, J., Sawarkar, A., Rodge, J.: Accelerating inference with sparsity using the nvidia ampere architecture and nvidia tensorsr. NVIDIA Developer Technical Blog, https://developer.nvidia.com/blog/accelerating-inference-with-sparsityusing-ampere-and-tensorr (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.786, + 0.646 + ], + "angle": 0, + "content": "43. Radosavovic, I., Kosaraju, R.P., Girshick, R., He, K., Dollár, P.: Designing network design spaces. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10428-10436 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.786, + 0.688 + ], + "angle": 0, + "content": "44. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: Mobilenetv2: Inverted residuals and linear bottlenecks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4510-4520 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.786, + 0.743 + ], + "angle": 0, + "content": "45. Shi, D., Tao, C., Jin, Y., Yang, Z., Yuan, C., Wang, J.: UPop: Unified and progressive pruning for compressing vision-language transformers. In: Proceedings of the 40th International Conference on Machine Learning. vol. 202, pp. 31292-31311. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.786, + 0.771 + ], + "angle": 0, + "content": "46. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.786, + 0.813 + ], + "angle": 0, + "content": "47. Song, Z., Xu, Y., He, Z., Jiang, L., Jing, N., Liang, X.: Cp-vit: Cascade vision transformer pruning via progressive sparsity prediction. arXiv preprint arXiv:2203.04570 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.786, + 0.84 + ], + "angle": 0, + "content": "48. Sun, M., Liu, Z., Bair, A., Kolter, J.Z.: A simple and effective pruning approach for large language models. arXiv preprint arXiv:2306.11695 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "G. Fang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "49. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818-2826 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.23 + ], + "angle": 0, + "content": "50. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers & distillation through attention. In: International Conference on Machine Learning. vol. 139, pp. 10347-10357 (July 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.271 + ], + "angle": 0, + "content": "51. Touvron, H., Cord, M., Sablayrolles, A., Synnaeve, G., Jégou, H.: Going deeper with image transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 32-42 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.311 + ], + "angle": 0, + "content": "52. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.352 + ], + "angle": 0, + "content": "53. Voita, E., Talbot, D., Moiseev, F., Sennrich, R., Titov, I.: Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. arXiv preprint arXiv:1905.09418 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.393 + ], + "angle": 0, + "content": "54. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media 8(3), 415-424 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.434 + ], + "angle": 0, + "content": "55. Wang, Z., Luo, H., Wang, P., Ding, F., Wang, F., Li, H.: Vtc-lfc: Vision transformer compression with low-frequency components. Advances in Neural Information Processing Systems 35, 13974-13988 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.462 + ], + "angle": 0, + "content": "56. Wen, W., Wu, C., Wang, Y., Chen, Y., Li, H.: Learning structured sparsity in deep neural networks. Advances in neural information processing systems 29 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.462, + 0.785, + 0.487 + ], + "angle": 0, + "content": "57. Wightman, R.: Pytorch-image-models. https://github.com/huggingface/pytorch-image-models" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.528 + ], + "angle": 0, + "content": "58. Yang, H., Yin, H., Shen, M., Molchanov, P., Li, H., Kautz, J.: Global vision transformer pruning with hessian-aware saliency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18547-18557 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.569 + ], + "angle": 0, + "content": "59. Ye, J., Lu, X., Lin, Z., Wang, J.Z.: Rethinking the smaller-norm-less-informative assumption in channel pruning of convolution layers. arXiv preprint arXiv:1802.00124 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.61 + ], + "angle": 0, + "content": "60. Yin, M., Uzkent, B., Shen, Y., Jin, H., Yuan, B.: Gohsp: A unified framework of graph and optimization-based heterogeneous structured pruning for vision transformer. arXiv preprint arXiv:2301.05345 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.651 + ], + "angle": 0, + "content": "61. Yu, F., Huang, K., Wang, M., Cheng, Y., Chu, W., Cui, L.: Width & depth pruning for vision transformers. In: AAAI Conference on Artificial Intelligence (AAAI). vol. 2022 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.785, + 0.678 + ], + "angle": 0, + "content": "62. Yu, H., Wu, J.: A unified pruning framework for vision transformers. Science China Information Sciences 66(7), 1-2 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.704 + ], + "angle": 0, + "content": "63. Yu, J., Huang, T.: Autoslim: Towards one-shot architecture search for channel numbers. arXiv preprint arXiv:1903.11728 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.785, + 0.745 + ], + "angle": 0, + "content": "64. Yu, L., Xiang, W.: X-pruner: explainable pruning for vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24355-24363 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.772 + ], + "angle": 0, + "content": "65. Yu, S., Chen, T., Shen, J., Yuan, H., Tan, J., Yang, S., Liu, J., Wang, Z.: Unified visual transformer compression. arXiv preprint arXiv:2203.08243 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "66. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6023-6032 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "67. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.475, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Isomorphic Pruning for Vision Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "68. Zheng, C., Zhang, K., Yang, Z., Tan, W., Xiao, J., Ren, Y., Pu, S., et al.: Savit: Structure-aware vision transformer pruning via collaborative optimization. Advances in Neural Information Processing Systems 35, 9010-9023 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "69. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 13001-13008 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.788, + 0.259 + ], + "angle": 0, + "content": "70. Zhu, M., Tang, Y., Han, K.: Vision transformer pruning. arXiv preprint arXiv:2104.08500 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.259 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Isomorphic Pruning for Vision Models/bf292aec-a3cb-46ef-92f7-27df67a69e35_origin.pdf b/2024/Isomorphic Pruning for Vision Models/bf292aec-a3cb-46ef-92f7-27df67a69e35_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6b95daaaf5494b8197476c2839a55d329943690c --- /dev/null +++ b/2024/Isomorphic Pruning for Vision Models/bf292aec-a3cb-46ef-92f7-27df67a69e35_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7dc630453e0770d0843223d33342287f7b71c008806e4b8245650f5ef573af8b +size 1075762 diff --git a/2024/Isomorphic Pruning for Vision Models/full.md b/2024/Isomorphic Pruning for Vision Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..00d9d7f64fe53e3953c83cd1c23092e519fa5894 --- /dev/null +++ b/2024/Isomorphic Pruning for Vision Models/full.md @@ -0,0 +1,302 @@ +# Isomorphic Pruning for Vision Models + +Gongfan Fang $^{1}$ , Xinyin Ma $^{1}$ , Michael Bi Mi $^{2}$ , and Xinchao Wang $^{1*}$ + +$^{1}$ National University of Singapore + +$^{2}$ Huawei Technologies Ltd. + +gongfan@u.nus.edu，xinchao@nus.edu.sg + +Abstract. Structured pruning reduces the computational overhead of deep neural networks by removing redundant sub-structures. However, assessing the relative importance of different sub-structures remains a significant challenge, particularly in advanced vision models featuring novel mechanisms and architectures like self-attention, depth-wise convolutions, or residual connections. These heterogeneous substructures usually exhibit diverged parameter scales, weight distributions, and computational topology, introducing considerable difficulty to importance comparison. To overcome this, we present Isomorphic Pruning, a simple approach that demonstrates effectiveness across a range of network architectures such as Vision Transformers and CNNs, and delivers competitive performance across different model sizes. Isomorphic Pruning originates from an observation that, when evaluated under a pre-defined importance criterion, heterogeneous sub-structures demonstrate significant divergence in their importance distribution, as opposed to isomorphic structures that present similar importance patterns. This inspires us to perform isolated ranking and comparison on different types of sub-structures for more reliable pruning. Our empirical results on ImageNet-1K demonstrate that Isomorphic Pruning surpasses several pruning baselines dedicatedly designed for Transformers or CNNs. For instance, we improve the accuracy of DeiT-Tiny from $74.52\%$ to $77.50\%$ by pruning an off-the-shelf DeiT-Base model. And for ConvNext-Tiny, we enhanced performance from $82.06\%$ to $82.18\%$ , while reducing the number of parameters and memory usage. Code is available at https://github.com/VainF/Isomorphic-Pruning. + +Keywords: Network Pruning $\cdot$ Vision Transformers $\cdot$ CNNs + +# 1 Introduction + +Deep Neural Networks are typically over-parameterized, containing numerous redundant sub-structures. Such redundancy can be effectively eliminated via structured pruning, without significantly sacrificing the network's performance [16, 22, 25]. Pruning usually follows a "ranking-to-prune" framework: To facilitate compression while retaining the learned capabilities of networks, a carefully-designed criterion is utilized to assess the importance score of different structures [11, 22, 40], followed by a ranking process to identify and remove those + +![](images/5d6fc412fc9fe29a137bf1527c879ab991715f988911296aaef04fe3e8d941bd.jpg) +Fig. 1: ImageNet Top-1 Accuracy vs. Multiply-Accumulate Operations (MACs) of pruned DeiT [50] and pruned ConvNext [33]. The pruned models, marked as “★” have comparable or better latency yet superior performance compared to scratch training counterparts highlighted by “○”. The MACs are in log scale for better visualization + +relatively unimportant ones. After pruning, the compressed model is fine-tuned to recover its performance [11, 18]. + +However, ranking-based structured pruning presupposes an underlying assumption that the parameters from different structures are directly comparable. Classic deep neural networks, such as VGG [46], are built by stacking similar convolutional layers or blocks where the neuron connections share a similar computational topology. Under the circumstances, a straightforward global ranking is expected to reveal their relative importance and remove unimportant parameters [11, 22]. However, with recent innovation in network structure design, advanced vision models like Vision Transformers [10, 31, 50] and ConvNext [33] incorporate more sophisticated designs, such as complicated residual connections [19, 24], attention mechanisms [52] and depth-wise convolutions [6]. These novel structures increase the network's internal complexity and heterogeneity, challenging the reliability of pruning importance criteria. For instance, a vanilla Vision Transformer [10] is composed of MLP, Multi-head Attention, and skip connections. Even in the same Multi-head attention layer, there are two pruning schemes to removing either entire attention heads [37, 53] or the within-head dimensions [58, 68]. These structures vary by their computational topologies and, as will be illustrated in the experiments, tend to exhibit diverged importance distributions. In this case, directly comparing the importance of these heterogeneous might be problematic. + +To address the problem, we propose a simple pruning approach, termed Isomorphic Pruning, to alleviate the incomparability issue arising from heterogeneous sub-structures. The essence of this method lies in decomposing a network into multiple isomorphic groups, as illustrated in Figure 2 (d), based on the computational topology of the sub-structures. Each isomorphic group contains + +![](images/c60658c4438fb36ae5860712318177d6e42191850ac2abd987ea02b23ec8875d.jpg) +(a) Network + +![](images/1e43295c1daf7b27db9057b7ee834b9531f436bccda89115995b1721c48a6d20.jpg) +(b) Local Pruning + +![](images/89d548bb24841bb9486b5878797ae9f1fe8eae9eb5d8700b20c5defcc3a35728.jpg) +(c) Global Pruning + +![](images/6dc360afa098946100a2127d209eb63a33d41a02f1042c44a1dc61033f2e96b4.jpg) +(d) Isomorphic Pruning +Fig. 2: For a pre-trained network (a), we show three pruning strategies: (b) Local Pruning that compares parameter importance within current layers; (c) Global Pruning that performs global ranking for all parameters; (d) the proposed Isomorphic pruning that groups parameters by the computational topology and applies importance ranking within groups. In each group, the importance distributions are more comparable. Details about the distribution can be found in Fig. 4 of the experiments section. + +elements with the same structure and tends to present a similar distribution of importance, making it more rational to compare their relative importance. + +To validate the effectiveness of our method, we conducted experiments on a variety of visual models, including DeiT [50], ConvNext [33], ResNet [19], and MobileNet-V2 [44]. Pruning results, as shown in Figure 1, on ImageNet-1K [8] demonstrate that Isomorphic Pruning achieves comparable or even superior performance relative to pruning algorithms specifically designed for CNNs or transformers. For instance, by pruning a pre-trained large DeiT model, we enhanced the performance of the DeiT-T model from $74.52\%$ to $77.50\%$ , when compared to the scratch training one. Similarly, our algorithm also yields satisfactory results on CNNs. We constructed compact ConvNexts and ResNets via pruning, which has fewer MACs and parameters while surpassing a conventionally trained model. Additionally, we also studied the real-world latency and memory footprint of the compressed models on both GPU and CPU devices. Our findings indicate that models pruned with Isomorphic Pruning achieve significant actual speed-ups and are competitive to carefully designed efficient architectures. + +Therefore, Our contribution lies in a practical and general pruning approach, which can handle various vision models with novel mechanisms and architectures. The pruned models show superior performance to several pruning competitors on ImageNet-1K benchmark. + +# 2 Related Works + +Network Pruning. Deep neural networks are typically over-parameterized and can be pruned without significant performance drop [16, 25]. The exploration of network pruning can be dated back to early works [1, 17, 25] about neural networks. Over the past few years, a large number of works have been conducted on convolutional networks to verify the effectiveness of pruning, which can be roughly categorized into unstructured [16, 29, 35] and structured pruning [5, 11, 18, 22, 56]. Unstructured pruning compresses a pre-trained model by forcing partial weights to be zero, leading sparse networks [29]. To accelerate + +the sparse model, special patterns and devices will be used such as N:M sparsity [12, 42, 48]. Structured pruning provides a hardware-friendly approach by removing sub-structures such as channels [11] or blocks [14] in CNNs, which improves the model efficiency without dedicated accelerators. In recent years, a large number of works adapted the idea of network pruning to vision transformers. It demonstrated that even the vanilla vision transformers can be squeezed for a better efficiency-performance trade-off [2, 4, 45, 47, 58, 60–62, 64, 65, 68, 70]. In the literature, transformer pruning has been studied from several perspectives, such as width reduction [70], depth pruning [61], attention head removal [58], sparse training [4] and architecture transformation [18]. + +Pruning by Global Ranking. In the context of network pruning, global pruning is recognized as a pivotal topic for exploration. This method entails a comprehensive ranking of all network layers, assigning adaptive pruning ratios to each layer according to the ranks [5,11,22,58]. Different layers within a network often exhibit varying levels of redundancy. To this end, the advantage of global pruning lies in its ability to create a more parameter-efficient, non-uniformly compressed model [58]. However, a key challenge remains in the ranking: how to ensure the consistency and comparability across different layers? A widely-used solution for this issue is normalizing scores on a per-layer basis to even out the importance distribution across the network [30]. Alternatively, some approaches use sparse training to learn an effective target structure [11], which might be costly due to the additional optimization. In this work, we introduce a more straightforward approach termed isomorphic pruning to facilitate a reliable ranking for pruning. + +# 3 Method + +The core idea of Isomorphic Pruning is straightforward: we group sub-structures in a network by their network topology and perform importance ranking within groups. This imposes two questions 1) how to identify removable sub-structures in networks and 2) how to compare the topology graph for grouping and pruning. + +# 3.1 Sub-structure in Network + +Structured Pruning aims to remove sub-structures in networks, such as the dimensions of hidden states or features. To make the formulation clear, we primarily concentrate on stacked fully-connected networks, although the principles established herein can be readily extended to CNNs and transformers. Initially, we consider a single linear layer which is parameterized by a matrix $W \in \mathbb{R}^{m \times n}$ . To articulate the pruning process, we introduce a pruning function $\mathrm{g}(W,d,k)$ defined as follows: + +$$ +\hat {w} = \mathrm {g} (W, d, k) = \left\{ \begin{array}{l l} \mathrm {g} (W, 0, k) = W _ {[ k,: ]} \in \mathbb {R} ^ {1 \times n}, & \text {i f p r u n i n g o u t p u t d i m e n s i o n ,} \\ \mathrm {g} (W, 1, k) = W _ {[:, k ]} \in \mathbb {R} ^ {m \times 1}, & \text {i f p r u n i n g i n p u t d i m e n s i o n ,} \end{array} \right. \tag {1} +$$ + +where $\hat{w}$ refers to the weight column or row slated for removal, acquired via a slicing operation denoted by subscript $[:,k]$ or $[k,:]$ . The pruning function $\mathbf{g}(W,d,k)$ is defined by three parameters, the parameter matrix $W$ , the axis $d\in \{0,1\}$ indicating the dimension to be compressed, and the index $k$ identifying the $k$ -th rows or columns in $W$ . In the context of a linear layer represented by a 2-D matrix $W\in \mathbb{R}^{m\times n}$ , the axis $d$ can assume values of either 0 or 1, with $k$ having a permissible range of $[0,m]$ and $[0,n]$ , respectively. This formulation encapsulates the pruning of a single layer; to extend it to the entire network, we also need to take the dependency [5,11,30] into account, where parameters from different layers are coupled and should be pruned simultaneously. As depicted in Figure 3, for two adjacent linear layers parameterized by $W_{1}$ and $W_{2}$ , the output dimension of the former is integrally coupled to the input dimension of the latter, thereby constituting a minimal sub-structure eligible for removal. If we treat each pruning operation $\mathbf{g}(W,d,k)$ as vertices defined by a triplet $(W,d,k)$ and the dependency as edges, then it is natural to model sub-structures as graphs. + +# 3.2 Graph Modeling of Sub-structures + +In this section, we delve deeper into the methodology for decomposing a network into several independent and minimal sub-structures in pruning, and model them as graphs. We can begin with the MLP network as illustrated in Figure 3, where the activation functions are omitted. The network comprises a sequence of 2-D parameter matrices $\{W_1, W_2, \dots, W_L\}$ , with $L$ denoting the network depth. Initially, all parameters are not assigned to any sub-structure. We randomly select one parameter $W$ as the seed to find its associated substructures and build the corresponding graph. For instance, selecting parameter $W_1$ and setting the axis $d = 0$ allows us to establish an initial graph $G = (V = \{(W, 0, k)\}, E = \{\})$ , focusing on the elimination of $k$ -th the output dimension in $W_1$ . Then we iterate through all potential $(W', d', k') \notin V$ to ascertain any dependency to $(W, d, k) \in V$ . Dependencies between layers can be ascertained through either manually-crafted patterns or automatic algorithms. In this work, we adhere to the approach of [11] to assess dependency based on two rudimentary rules: + +Dependency Modeling. For a pair of triplets $(W', d', k')$ , $(W, d, k)$ , a dependency is established if either of the following criteria is met: 1) $W'$ and $W$ are adjacent layers with $d = d'$ and $k = k'$ ; 2) $W' = W$ and $\mathrm{g}(W', d', k') = \mathrm{g}(W, d, k)$ . The first criterion addresses adjacent layers as exemplified in Figure 3, while the second pertains to specialized layers, such as normalization layers, which share the same pruning pattern for both outputs and inputs. Thereafter, we update the vertex set $V$ and edge set $E$ with the discovered dependency between $(W', d', k')$ and $(W, d, k)$ accordingly: + +$$ +\begin{array}{l} V = V \cup \left\{\left(W ^ {\prime}, d ^ {\prime}, k ^ {\prime}\right) \right\}, \\ E = E + \left\{\left(\left(W _ {1} - 1\right), \left(W ^ {\prime}, d ^ {\prime}, 1\right)\right) \right\} \end{array} \tag {2} +$$ + +$$ +E = E \cup \left\{\left(\left(W, d, k\right), \left(W ^ {\prime}, d ^ {\prime}, k ^ {\prime}\right)\right) \right\}. +$$ + +We can further repeat the above process until no additional vertex is incorporated into the current graph. The resultant graph $G$ then serves as a topological + +![](images/c39cbbbb1c8a68e49ea3f7544e05c8267b8b5b5ca0e7a6687cdf9a9bc326b2cc.jpg) +Fig. 3: Isomorphic Pruning models sub-structures as graphs, and apply isolated ranking and pruning with isomorphic structures. We highlight three removable substructures in an MLP and show their corresponding graphs below. Substructures 1 & 3 are isomorphic but 2 & 3 are non-isomorphic due to the additional residual connections. + +representation of a single sub-structure associated with the initial seed parameter. To map all substructures onto different graphs, we subsequently embark on a recursive procedure, commencing from those nodes not yet allocated to any sub-structure. Figure 3 furnishes an exemplar of this process, showcasing three distinct substructures labeled as sub-structures 1, 2, and 3. + +# 3.3 Ranking with Graph Isomorphism + +With the graph-based representation of network sub-structures, we advance to the ranking phase. This stage involves estimating and ranking the significance of each sub-structure to identify redundancies. A pivotal component of this process is an importance criterion, which quantifies the relative significance of various sub-structures. Previous works in the literature presents numerous well-crafted criteria, including magnitude pruning and Taylor expansion. In this work, we focus on the general importance function $I(\cdot)$ for a single layer, represented as: + +$$ +I (W, d, k) = \left\{ \begin{array}{l l} \| \mathrm {g} (W, d, k)) \|, & \text {i f M a g n i t u d e P r u n i n g} \\ \| \frac {\partial \mathcal {L}}{\partial \mathrm {g} (W , d , k)} \mathrm {g} (W, d, k)) \| _ {2}, & \text {i f T a y l o r P r u n i n g} \\ \dots & \text {O t h e r i m p o r t a n c e C r i t e r i a .} \end{array} \right. \tag {3} +$$ + +In structural pruning, our objective is to remove sub-structures encompassing parameters across multiple layers. Existing approaches [5,11,30] generally extend the single-layer importance to a holistic measure by aggregating scores across all constituents $(W,d,k)$ for the vertexes set, denoted as $I^{*}(\cdot)$ as follows: + +$$ +I ^ {*} (G (V, E)) = \sum_ {(W, d, k) \in V} I (W, d, k), \tag {4} +$$ + +thereby cumulating the importance of all elements within graph $G$ . However, the inherent diversity in parameters and computational topology among different + +substructures $G$ introduces potential ambiguities in this metric. Experimental evidence, particularly in the context of a vision transformer as shown in Figure 4, elucidates this issue: the largest substructure encompasses all embedding dimensions, containing a substantial number of parameters, whereas the most minimal sub-structure comprises merely two linear layers. Consequently, such aggregation may fail to accurately reflect the true relative importance of these entities. To address the challenges posed by the heterogeneity of substructures, we delineate a simple approach designed to preserve the nuanced information encapsulated by the aggregated importance, as outlined in Equation 4, whilst alleviating the detrimental repercussions raised by the heterogeneity of substructures. Our principal strategy entails the clustering of sub-structures into various isomorphic groups, predicated on their topological logic and parameter configurations. In this case, applying aggregated importance within groups can be meaningful, since isomorphic sub-structures show one-to-one correspondence among their parameters, and how different parameters participate in computations is entirely identical. To facilitate this, the key step is to examine the isomorphism of two graphs $G_{1}$ and $G_{2}$ . Formally, we adopt the following definition to ascertain graph isomorphism. + +Definition 1 (Graph Isomorphism). Two labeled graphs $G$ and $G'$ are isomorphic if there exists a bijection between their vertices so that, all mapped vertices preserve the same labels, and are connected in the same way. + +To develop a general indicator Isomorphic $(\mathrm{g}_1,\mathrm{g}_2)$ , we need to examine both the label and connectivity in two graphs. + +Vertex Label and Connectivity. In our settings, each vertex is labeled by the triplet $(W,d,k)$ , encompassing the parameter matrix $\mathbf{W}$ , axis $d$ , and index $k$ . We follow a simple principle for labling: Two vertices $(\mathbf{W},d,k)$ and $(\mathbf{W}',d',k')$ are labeled identically if and only if the parameter $\mathbf{W}'$ originates from the same layer type such as Linear or Conv, with the same pruning dimension $d = d'$ . We permit the index $k$ to differ, since within the same parameter matrix, the weight vectors sliced along the $d$ -th dimension are consistently homogeneous and thus can be considered isomorphic. After determining the label for each vertex, we further explore the connectivity in two graphs $G_{1}$ and $G_{2}$ . As illustrated in Figure 3, we show three substructures alongside their respective graphs $G_{1}$ , $G_{2}$ , and $G_{3}$ , where $G_{1}$ and $G_{3}$ incorporate a singular edge, whereas $G_{2}$ encompasses three edges. If we check the edges following the execution order of the computational graphs, it's natural to observe the isomorphism of $G_{1}$ and $G_{3}$ , since they share the same topology. To make this process more formal, we assume the edges in graph $G$ are already sorted according to their execution order in forwarding, and check the isomorphism by + +$$ +\operatorname {I s o m o r p h i c} \left(G, G ^ {\prime}\right)) = \mathbb {1} \left\{\left| G \right| = \left| G ^ {\prime} \right| \wedge \operatorname {l a b e l} \left(V _ {i}\right) = \operatorname {l a b e l} \left(V _ {i} ^ {\prime}\right); \forall i \in \{1, 2, \dots , \left| G ^ {\prime} \right| \} \right\}, +$$ + +where $\mathbb{1}$ is an indicator function. $|G^{\prime}|$ denotes the number of edges in the graph and $\wedge$ signifies the logical AND operation. The first condition ensures the equiva + +Algorithm 1 Isomorphic Pruning +1: Input: Network with parameters $\{W_1, W_2, \ldots, W_n\}$ , Pruning ratio $p\%$ . +2: Output: Pruned Network $\{W_1', W_2', \ldots, W_n'\}$ +3: procedure IDENTIFYSUBSTRUCTURES(\{W1, W2, \ldots, Wn\}) +4: Initialize empty list of sub-structures $\mathbf{S} = \{\}$ +5: for each layer $W_i$ in $\{W_1, W_2, \ldots, W_n\}$ do +6: if $W_i$ is not assigned to any substructure then +7: Initialize a graph $G(V = \{W_i\}, E = \{\})$ . +8: Find and add coupled $W_j$ to $G$ until there is no new dependency. +9: $\mathbf{S} = \mathbf{S} \cup \{G\}$ +10: return $\mathbf{S}$ +11: procedure PRUNEBYGROUPEDRANKING(S, p) +12: for each substructure $G_i$ in $\{G_1, G_2, \ldots, G_n\}$ do +13: if $G_i$ is not assigned to any group then +14: Initialize a Group $\mathbf{R} = \{G_i\}$ . +15: Find and add $G_j$ that satisfies Isomorphism $(G_i, G_j) = 1$ to $\mathbf{R}$ . +16: Calculate the importance scores $\{I^*(G) | G \in \mathbf{R}\}$ . +17: Remove $p\%$ substructures in $\mathbf{R}$ by ranking the importance scores. +18: return $\mathbf{S}$ +19: $\mathbf{S} \gets \text{IDENTIFYSUBSTRUCTURES}(\{W_1, W_2, \ldots, W_n\})$ +20: $\mathbf{S} \gets \text{PRUNEBYGROUPEDRANKING}(\mathbf{S}, p)$ + +lence of the number of vertices in the graphs, while the second condition confirms the matching labels of vertices on either side of the edge pairs $V_{i}$ and $V_{i}^{\prime}$ . + +Using Equation 5, it is easy to cluster substructures into different groups and perform rankings within each group exclusively. Specifically, we examine a collection of substructures $\{G_1, G_2, \ldots, G_N\}$ that share identical graph topologies. We estimate the aggregated significance of each graph independently using Equation 4, resulting in an $N$ -dimensional importance vector $[I^*(G_1), \ldots, I^*(G_N)]$ . This facilitates a straightforward ranking of the importance vector, followed by the elimination of the $\mathfrak{p}\%$ unimportant substructures based on their scores. This ranking method is deemed reliable since all substructures within a group possess the same architectural design, parameter scale, and computational graph. A similar principle can be uniformly applied to other isomorphic groups for pruning, as detailed in Algorithm 1. + +# 4 Experiments + +This section reports our empirical results on the ImageNet-1K dataset [8], which comprises 1,281,167 images for training and 50,000 for validation. We deploy the proposed Isomorphic Pruning to compress several vision models, including ConvNext [33], ResNet [36], MobileNetv2 [44], and Vision Transformers [10, 50]. We report the parameter amount (#Params), MACs, and Top-1 validation accuracy in Table 1 and 2, and report the actual latency and peak memory consumption in Table 3. + +# 4.1 Settings for ImageNet-1K + +Pruning. For simplicity, we adopt the Taylor-based pruning framework [39], which applies a data-driven Taylor expansion to the loss function. For this purpose, we randomly sampled 100 mini-batches from the ImageNet dataset, each containing 64 images. These images were resized to $256 \times 256$ pixels and then center-cropped to $224 \times 224$ pixels. We normalize all images using the mean and standard deviation specific to ImageNet. Notably, while some methods, such as ConvNext such as ConvNext [33] and DeiT [50] employ additional training techniques like label smoothing and knowledge distillation, our approach only leverages a straightforward cross-entropy loss for the Taylor expansion. We accumulate gradients from all the batches for a more reliable importance estimation and apply Isomorphic pruning in a one-shot manner without iterative pruning and fine-tuning [40]. + +Finetuning. During the fine-tuning process of pruned models, we adhere to the original training protocol for a fair comparison. While maintaining most of the original settings unchanged, we adjusted the learning rates and batch sizes to make the training suitable for A5000 GPUs, following the linear LR scaling rule proposed in [13]. For example, to compressing a pre-trained DeiT models [50], we adopt the same augmentation strategies for training, including RandAugmentation [7], Mixup [67], CutMix [66], Random Erasing [69], Repeated Augmentation [23], label smoothing [49]. The pruned models are distilled with a pre-trained RegNetY [43, 50] for 300 epochs using the AdamW [34] optimizer, with a learning rate of 0.0005 and a total batch size of 2048. A weight decay of 0.05 and a cosine annealing scheduler is deployed for training. For other models like ConvNext [33], ResNet [19] and MobileNet-v2 [44], the same principle will be applied. More details of hyper-parameters, pruning, and finetuning can be found in the appendix. + +Evaluation. We report the optimal classification accuracy achieved on the ImageNet-1K validation set, employing the standard resize-and-crop protocol [19, 33] for evaluation. To further validate the efficiency of models, we profiled the latency and the peak memory consumption of the pruned models on both GPU (a single RTX A5000) and CPU. To guarantee a fair and standardized comparison, we also tested the latency of several baseline models sourced from the Pytorch-Image Models [57] and Torchvision [38], following the same testing procedure. The latency was estimated with a batch size of 256 for GPU and 8 for CPU, averaged on 100 repeats. Appropriate GPU warmup was deployed in our test. + +# 4.2 Pruning Convolutional Neural Networks + +Convolutional Neural Network is a critical focus in the area of network pruning. In this section, we concentrate on pruning three popular ConvNets: ConvNext [33], ResNet [19] and MobileNet-v2 [44]. + +Table 1: Pruning CNNs on ImageNet-1K. + +
Pruning Method#Params (M)MACs (G)Base Acc.Final Acc.Δ Acc.
ConvNext-B† [33]88.5915.39-83.83-
ConvNext-S† [33]50.228.71-83.14-
ConvNext-S (Ours)47.368.4883.8383.17-0.66
ConvNext-T† [33]28.594.47-82.06-
ConvNext-T (Ours)25.324.1983.8382.19-1.64
ResNet-152†60.1911.58-78.31-
AOFP-E2 [9]-4.1377.3777.00-0.37
ABCPruner [28]24.074.3178.3177.12-1.19
EPruner-0.63 [27]21.564.0878.3176.83-1.48
ResNet-152-4.0G (Ours)23.114.0578.3177.84-0.47
ResNet-101†44.557.85-77.38-
IE [39]31.24.7077.3777.35-0.02
FPGM [21]-4.5177.3777.32-0.05
SFP [20]-4.5177.3777.51+0.14
ISTA [59]-4.4776.4075.27-1.13
Res101-4.5G (Ours)29.144.4877.3877.56+0.16
GFP [30]28.023.9078.2978.33+0.04
AOFP [9]-3.8976.6376.40-0.23
Res101-3.8G (Ours)24.873.8577.3877.43+0.05
ResNet-50† [19]25.564.13-76.13-
Taylor [39]14.202.2576.1874.50-1.68
CCP [41]-2.1176.1575.50-0.65
GFP [30]19.422.0476.7976.42-0.37
AutoSlim [63]20.602.0076.1075.60-0.50
DepGraph [11]-1.9976.1575.83-0.32
ResNet50-2G (Ours)15.052.0676.1375.91-0.22
Mobv2†3.500.32-71.88-
Meta [32]-0.1474.7068.20-6.50
GFP [30]-0.1575.7469.16-6.58
DepGraph [11]-0.1571.9068.49-3.41
Mobv2-0.15G (Ours)2.190.1571.8868.91-2.97
+ +ConvNext. Our experiments in Table 1 begin with the evaluation on an advanced CNN, ConvNext [33], which achieves impressive performance on several benchmarks. We use the official ConvNext-Base [38] as the pre-trained model for pruning, which achieves an accuracy of $83.83\%$ on the ImageNet-1K validation set. The pruning of ConvNext models is a relatively underexplored topic. Hence, we establish our baselines using the official variants with fewer parameters, ConvNext-S and ConvNext-T. For ConvNext, we also generalize our method to depth pruning, which removes unimportant layers with the minimal average scores on substructures, so that the pruned models have the same depth as baselines. The pruned models are fine-tuned for 300 epochs following the same training protocol as [33]. During fine-tuning, we observed that the pruned ConvNext with half of the parameters removed, achieves an accuracy of 80.09 in only 20 epochs, which effectively validates the efficiency of pruning. Moreover, as illustrated in Table 1, the final model after 300 full fine-tuning, achieves competitive performance (82.19% vs. 82.06%) to pre-trained models with fewer parameters (25.32M vs. 28.59M) and MACs (4.19G vs. 4.47G). + +ResNet. In line with prior studies [5, 11, 22, 30], we demonstrate the efficacy of our method by structurally pruning pre-trained ResNet MACs. ResNet- + +Table 2: Pruning DeiTs on ImageNet-1K. + +
Method#Params (M) MACs (G) Acc (%)Method#Params (M) MACs (G) Acc (%)
DeiT-B† [50]87.3417.6983.32DeiT-T† [50]5.911.2774.52
DeiT-S† [50]22.444.6481.17DeiT-T-600EP [58]5.911.2775.00
DeiT-S-600EP [58]22.444.6481.80WDPruning [61]13.302.6078.38
VTP - 40% [70]48.0010.0080.70X-Pruner [64]-2.4078.93
WDPruning [61]55.309.9080.76UVC [65]-2.3278.82
UVC [65]-8.0080.57ViT-Slim [70]11.402.3077.94
X-Pruner [64]-8.5081.02ViT-Slim [70]11.402.3077.94
NViT-B [58]34.006.8083.29NViT-T [58]6.901.3076.21
SAViT [68]25.405.3081.66SAViT+KD [68]6.601.3076.95
UP-DeiT-S [62]22.10-81.56UP-DeiT-T [62]5.70-75.79
NViT-S [58]21.004.2082.19DeiT-T (Ours)5.741.2177.50
Upop [45]19.904.1081.10VTC-LFC [55]5.10-71.60
DeiT-S (Ours)20.694.1682.41SViT-E [68]4.210.9970.12
GOHSP [60]14.402.9979.98GOHSP [60]4.000.9170.24
GOHSP [60]11.102.8179.86SAViT [68]4.200.9070.72
ViT-Slim [2]13.502.8079.20CP-ViT [47]-0.7471.24
Upop [45]13.502.8079.60WDPruning [61]3.500.7070.34
ViT-Slim [2]13.502.8079.20X-Pruner [64]-0.6071.10
CP-ViT [47]-2.6679.08UVC [65]-0.5170.60
DeiT-2.6G (Ours)13.072.6281.13DeiT-0.6G (Ours)3.080.6272.60
+ +50/101/152 consists of four residual blocks, each formed by several Bottle-neck structures and residual connections [19]. The proposed Isomorphic Pruning identifies five kinds of distinct isomorphic groups within ResNet-50, four targeting the block structures and one associated with bottleneck structures [19]. It is worthwhile to note that different sub-structures vary in their parameter and MAC counts. For example, residual blocks naturally contain more parameters than the bottleneck structures which are only composed of a few convolutional layers. The conventional global pruning, employing a naive global ranking threshold, faces the potential risks of over-pruning certain layers due to the biased importance distribution. Isomorphic Pruning, however, employs a more stable scheme for pruning. As illustrated in Table 1, we report both the accuracy and the accuracy drop compared to their base models. Isomorphic pruning achieves lossless compression on ResNet-101, under both 4.5G and 3.8G settings. + +MobileNet-v2. We also conduct experiments on lightweight models, MobileNet-v2, by compressing its MACs from 0.32 G to 0.15 G. In MobileNet-v2, we can also discover 5 isomorphic groups in MobileNet. The pruned model achieves better accuracy compared to DepGraph [11] which requires sparse training. + +# 4.3 Pruning Vision Transformers. + +Vision Transformers (ViT) has profoundly reshaped the paradigm of visual modeling over the past few years. Nevertheless, their inherent over-parameterization often imposes huge overheads to both training and inference. Compared to CNNs that are composed of similar convolution layers or blocks, Vision Transformers are usually built with MLP, Multi-head Attention, and skip connections, which are more challenging for pruning due to the heterogeneous internal structures. + +Table 2 presents our pruning results on DeiT, a plain vision transformer enhanced with knowledge distillation [50]. We benchmark our approach with + +Table 3: Profiling the pruned models and other off-the-shelf vision models from PyTorch-Image-Models [57] on a single RTX A5000 GPU and CPU. For NViT [58], we report their best speed up in the original paper, tested on a single V100 or RTX 3080. + +
Method#Params ↓ MACs ↓ Peak Mem. ↓ GPU Latency (256) ↓ CPU Latency (8) ↓ Acc. (%)
DeiT-B [50]86.40 M17.60 G2909.78 MB802.97 ms (1.00×)197.73 ms (1.00×)83.32
DeiT-S [50]22.44 M4.64 G1363.47 MB241.81 ms (3.32×)80.64 ms (2.45×)81.17
Swin-T [31]28.29 M4.51 G4257.98 MB375.45 ms (2.14×)203.86 ms (0.97×)81.19
NViT-S [58]21.00 M4.20 G--(2.52×)-82.19
TNT-S [15]23.76 M4.85 G1457.56 MB631.02 ms (1.27×)175.58 ms (1.13×)81.50
CaiT-XS24 [51]26.56 M5.40 G1496.42 MB557.41 ms (1.44×)342.28 ms (0.58×)82.00
CrossViT-S [3]26.86 M5.64 G1546.94 MB315.02 ms (2.55×)115.51 ms (1.71×)81.00
EfficientFormer-L3 [26]31.41 M3.94 G2681.75 MB249.05 ms (3.22×)167.16 ms (1.18×)82.40
PVTv2-B2 [54]25.36 M4.00 G4478.18 MB364.00 ms (2.21×)212.44 ms (0.93×)82.00
DeiT-S (Ours)20.69 M4.16 G1547.97 MB230.84 ms (3.48×)44.10 ms (4.48×)82.41
DeiT-2.6G (Ours)13.07 M2.62 G1203.07 MB189.68 ms (4.23×)56.59 ms (3.49×)81.13
DeiT-T [50]5.91 M1.27 G720.22 MB93.78 ms (8.56×)46.83 ms (4.22×)74.52
NViT-T [58]6.90 M1.30 G--(4.97×)-76.21
DeiT-T (Ours)5.71 M1.20 G837.16 MB93.17 ms (8.62×)47.89 ms (4.13×)77.50
DeiT-0.6G (Ours)3.08 M0.62 G629.28 MB63.23 ms (12.70×)40.98 ms (4.83×)72.60
ConvNext-B [33]88.59 M15.36 G5394.10 MB789.24 ms (1.00×)262.58 ms (1.00×)83.83
ConvNext-S [33]50.22 M8.69 G3958.45 MB508.20 ms (1.57×)188.07 ms (1.40×)83.14
ConvNext-S (Ours)47.36 M8.48 G3788.65 MB535.95 ms (1.59×)179.72 ms (1.46×)83.17
ConvNext-T [33]28.59 M4.46 G3785.37 MB291.02 ms (2.82×)126.17 ms (2.08×)82.06
ConvNext-T (Ours)25.32 M4.19 G3607.06 MB292.77 ms (2.83×)97.82 ms (2.68×)82.19
ResNet152 [19]60.19 M11.58 G3320.24 MB386.65 ms (1.00×)272.21 ms (1.00×)78.31
ResNet152-4G (Ours)23.11 M4.06 G1938.91 MB238.12 ms (1.62×)180.65 ms (1.51×)77.84
ResNet101 [19]44.55 M7.85 G3196.77 MB271.34 ms (1.00×)202.66 ms (1.00×)77.38
ResNet101-4.5G (Ours)29.14 M4.48 G2639.40 MB220.72 ms (1.23×)162.33 ms (1.25×)77.56
ResNet101-3.8G (Ours)24.87 M3.85 G2286.36 MB194.39 ms (1.40×)158.98 ms (1.28×)77.43
ResNet50 [19]25.56 M4.13 G3035.84 MB171.54 ms (1.00×)114.86 ms (1.00×)76.13
ResNet50-2.0G (Ours)15.05 M2.06 G2238.15 MB78.75 ms (1.33×)80.88 ms (1.42×)75.91
Mobv2 [44]3.50 M0.32 G2855.44 MB71.42 ms (1.00×)74.46 ms (1.00×)71.88
Mobv2-0.15G (Ours)2.19 M0.15 G1354.82 MB43.48 ms (1.64×)38.63 ms (1.93×)68.91
+ +a range of innovative pruning baselines specifically tailored for vision transformers [45, 47, 50, 55, 58, 60, 61, 64, 65, 68, 70]. A DeiT model consists of various components, such as MLP and attention layers. Within the attention layers, we can even find two feasible pruning schemes: pruning 1) the number of heads or 2) within-head dimensions. We implement Isomorphic Pruning to compress a pre-trained DeiT-Base model by pruning both heads, head dimensions, and the embedding sizes following [58]. This is achieved through a straightforward first-order Taylor expansion for importance estimation, followed by a grouped ranking in Isomorphic Pruning. The details of pruning vision transformers can be found in the appendix. + +We developed four lightweight vision transformers: DeiT-S, DeiT-2.6G, DeiT-T and DeiT-0.6G. The pruned DeiT-S and DeiT-T models achieve comparable latencies comparable to the standard DeiT-S and DeiT-T with manually designed architecture. The empirical results are reported in Table 2, which shows that our method yields greater performance gains in complex models with extensive heterogeneous internal structures, especially in comparison to CNNs. Specifically, the optimized DeiT-T achieves an accuracy of $77.50\%$ , with only 1.2G MACs, which is better than the results $(74.52\%)$ of the pre-trained DeiT-T with uniform width across layers. Detailed configurations, such as base models and pruning ratios, are available in the appendix. + +![](images/e3600dd6856369561d35f1101ff00315d23351fc6c524ef5700d3bbb6544dff2.jpg) +(a) + +![](images/9ea2c9c087d773e289c1bc67b427415f2b26b09f77e1206d308f314c1b97a7e9.jpg) +(b) + +![](images/0279396134af7b62b84d0b38091ee89e433508089848e776920f76d687e0c1dc.jpg) +(c) +Fig. 4: (a-b) Top-1 Accuracy and loss of pruning ResNet-50 on ImageNet-1K validation set without finetuning. We report the mean and standard derivation of 10 experiments. IsomorphicP Pruning consistently improves the performance of pruned models. (c) Histogram of Taylor importance score of DeiT sub-structures. Isomorphic structures are highlighted with the same color. Zoom in for more details about each structure. We visualize the threshold of $50\%$ pruning ratio for naive global pruning (the black dash) and isomorphic pruning (the colored dash). + +# 4.4 Analysis on the Pruned Models + +On-device Latency Table 3 measures the actual latency and peak memory consumptions on a single RTX A5000 GPU and CPU. We evaluate all models with a batch size of 256 for GPU and 8 for CPU, respectively. The average latency over 100-round experiments are reported, with a 20-step warmup process before evaluation. Firstly, results show that the pruned DeiT models can achieve comparable performance (82.41% vs. 82.40%) to some advanced transformers like EfficientFormer [26], while preserving a competitive latency. Additionally, we observe that the peak memory consumption is primarily determined by the largest layer width. Therefore, under the same budget, a uniform DeiT may require less peak memory (1363 MB for DeiT-S) for inference. In contrast, a non-uniform architecture may require slightly more memory (1547 MB for the DeiT-S Pruned) under the given budget. But non-uniform pruning can achieve better accuracy, which reveals an important trade-off between accuracy and memory consumption for transformers. In addition, for CNNs like ResNets, the pruned model requires much less memory due to the compressed feature dimensions. Furthermore, we note that the actual speed-up may not be linearly correlated with the MAC reductions. This is particularly evident in ResNet-50, where only a $1.33 \times$ actual speed-up is achieved despite the removal of half of the MACs. In contrast, compression on Mobv2 is more effective, achieving a $1.64 \times$ acceleration with a similar compression ratio on MACs. + +Distribution of Importance Scores. To further explore how Isomorphic Pruning affects the pruning behavior, we study the distribution of importance + +score in a DeiT-Base model [50]. As shown in Figure 4 (c), the importance histogram of different sub-structures is highlighted with different colors, each corresponding to one kind of isomorphic structure. It is natural to find that, the MLP layer and the attention layer have diverged importance distributions due to their different computation process. Besides, when we apply a simple $50\%$ global pruning on the DeiT model, most parameters in Attentions $\# 5$ , $\# 9$ , and $\# 13$ will be removed, since the importance scores in attentions are centered around 0. In contrast, the embedding number as $\# 1$ will not be even been pruned in this case, due to the relatively large importance scores. This inevitably leads to an imbalanced pruning of different structures. Notably, the embedding structure, as illustrated in the appendix, contains massive parameters in the DeiT, where biased pruning may lead to very limited acceleration. + +Importance Criteria. Note that Isomorphic Pruning can be combined with different pruning criteria like Taylor [39], Hessian [25] and even L1-norm [22]. We quantify the performance gain of Isomorphic Pruning by visualizing the accuracy and loss curve with different pruning ratios in Figure 4 (a) and (b). Specifically, we gradually increase the pruning ratio for each criterion and report the average accuracy and loss without finetuning over 10 repeated runs with random inputs for Taylor expansion [25, 39]. Note that L1 pruning is deterministic and thus presents 0 std. We find the Isomorphic Pruning, marked as solid lines, consistently improves the counterpart baselines highlighted in dashed lines. + +Transfer Learning. We also evaluate the performance of transfer learning based on our pruned models. Table 4 reports the accuracy of transferring different models to CIFAR-10 and CIFAR-100. We resize all images to $224 \times 224$ and apply the same training recipe for transfer. For the DeiT models, we did not include the supervision from the teacher models. It is observed that our method achieves better transfer accuracy compared to the scratch training model and baseline methods. + +Table 4: Transfer learning on pruned models. + +
ModelsCIFAR-10CIFAR-100
DeiT-S [39]98.5287.07
NViT-S [58]98.7887.90
DeiT-S (Ours)99.0290.09
DeiT-T93.9385.66
NViT-T [58]98.3185.88
DeiT-S (Ours)98.2887.08
ConvNext-S98.7789.15
ConvNext-S (Ours)98.7889.36
+ +# 5 Conclusion + +In this paper, we introduce Isomorphic Pruning, a practical approach designed to compress vision networks with novel mechanisms and architectures. This method focuses on mitigating the challenges posed by heterogeneous sub-structures in networks, thereby enhancing the reliability of ranking and pruning processes. Our empirical results on the ImageNet-1k dataset, showcase the effectiveness of Isomorphic Pruning across several vision models like ConvNext, ResNet, MobileNetv2, and Vision Transformers. + +# Acknowledgement + +This project is supported by the National Research Foundation Singapore under its Medium Sized Center for Advanced Robotics Technology Innovation, and the Advanced Research and Technology Innovation Centre (ARTIC), the National University of Singapore under Grant (project number: A-0005947-21-00, project reference: ECT-RP2). + +# References + +1. Baum, E., Haussler, D.: What size net gives valid generalization? Advances in neural information processing systems 1 (1988) +2. Chavan, A., Shen, Z., Liu, Z., Liu, Z., Cheng, K.T., Xing, E.P.: Vision transformer slimming: Multi-dimension searching in continuous optimization space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4931-4941 (2022) +3. Chen, C.F.R., Fan, Q., Panda, R.: Crossvit: Cross-attention multi-scale vision transformer for image classification. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 357-366 (2021) +4. Chen, T., Cheng, Y., Gan, Z., Yuan, L., Zhang, L., Wang, Z.: Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems 34, 19974-19988 (2021) +5. Chen, T., Liang, L., Ding, T., Zhu, Z., Zharkov, I.: Otov2: Automatic, generic, user-friendly. arXiv preprint arXiv:2303.06862 (2023) +6. Chollet, F.: Xception: Deep learning with depthwise separable convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1251-1258 (2017) +7. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 702-703 (2020) +8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +9. Ding, X., Ding, G., Guo, Y., Han, J., Yan, C.: Approximated oracle filter pruning for destructive cnn width optimization. In: International Conference on Machine Learning. pp. 1607-1616. PMLR (2019) +10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +11. Fang, G., Ma, X., Song, M., Mi, M.B., Wang, X.: Depgraph: Towards any structural pruning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16091-16101 (2023) +12. Frantar, E., Alistarh, D.: Sparsegpt: Massive language models can be accurately pruned in one-shot (2023) +13. Goyal, P., Dolkar, P., Girshick, R., Noordhuis, P., Wesolowski, L., Kyrola, A., Tulloch, A., Jia, Y., He, K.: Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677 (2017) + +14. Guo, J., Ouyang, W., Xu, D.: Multi-dimensional pruning: A unified framework for model compression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1508-1517 (2020) +15. Han, K., Xiao, A., Wu, E., Guo, J., Xu, C., Wang, Y.: Transformer in transformer. Advances in Neural Information Processing Systems 34, 15908-15919 (2021) +16. Han, S., Mao, H., Dally, W.J.: Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. International Conference on Learning Representations (ICLR) (2016) +17. Hassibi, B., Stork, D.G., Wolff, G.J.: Optimal brain surgeon and general network pruning. In: IEEE international conference on neural networks. pp. 293-299. IEEE (1993) +18. He, H., Liu, J., Pan, Z., Cai, J., Zhang, J., Tao, D., Zhuang, B.: Pruning self-attention into convolutional layers in single path. arXiv preprint arXiv:2111.11802 (2021) +19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +20. He, Y., Kang, G., Dong, X., Fu, Y., Yang, Y.: Soft filter pruning for accelerating deep convolutional neural networks. arXiv preprint arXiv:1808.06866 (2018) +21. He, Y., Liu, P., Wang, Z., Hu, Z., Yang, Y.: Filter pruning via geometric median for deep convolutional neural networks acceleration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4340-4349 (2019) +22. He, Y., Zhang, X., Sun, J.: Channel pruning for accelerating very deep neural networks. In: Proceedings of the IEEE international conference on computer vision. pp. 1389-1397 (2017) +23. Hoffer, E., Ben-Nun, T., Hubara, I., Giladi, N., Hoefler, T., Soudry, D.: Augment your batch: Improving generalization through instance repetition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8129-8138 (2020) +24. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017) +25. LeCun, Y., Denker, J., Solla, S.: Optimal brain damage. Advances in neural information processing systems 2 (1989) +26. Li, Y., Yuan, G., Wen, Y., Hu, J., Evangelidis, G., Tulyakov, S., Wang, Y., Ren, J.: Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems 35, 12934-12949 (2022) +27. Lin, M., Ji, R., Li, S., Wang, Y., Wu, Y., Huang, F., Ye, Q.: Network pruning using adaptive exemplar filters. IEEE Transactions on Neural Networks and Learning Systems 33(12), 7357-7366 (2021) +28. Lin, M., Ji, R., Zhang, Y., Zhang, B., Wu, Y., Tian, Y.: Channel pruning via automatic structure search. arXiv preprint arXiv:2001.08565 (2020) +29. Liu, B., Wang, M., Foroosh, H., Tappen, M., Pensky, M.: Sparse convolutional neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 806-814 (2015) +30. Liu, L., Zhang, S., Kuang, Z., Zhou, A., Xue, J.H., Wang, X., Chen, Y., Yang, W., Liao, Q., Zhang, W.: Group fisher pruning for practical network compression. In: International Conference on Machine Learning. pp. 7021-7032. PMLR (2021) +31. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings + +of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) +32. Liu, Z., Mu, H., Zhang, X., Guo, Z., Yang, X., Cheng, K.T., Sun, J.: Metapruning: Meta learning for automatic neural network channel pruning. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3296-3305 (2019) +33. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11976-11986 (2022) +34. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +35. Louizos, C., Welling, M., Kingma, D.P.: Learning sparse neural networks through $l\_0$ regularization. arXiv preprint arXiv:1712.01312 (2017) +36. Ma, X., Yuan, G., Lin, S., Li, Z., Sun, H., Wang, Y.: Resnet can be pruned $60 \times$ : Introducing network purification and unused path removal (p-rm) after weight pruning. In: 2019 IEEE/ACM International Symposium on Nanoscale Architectures (NANOARCH). pp. 1-2. IEEE (2019) +37. Ma, X., Fang, G., Wang, X.: Llm-pruner: On the structural pruning of large language models. In: Advances in Neural Information Processing Systems (2023) +38. maintainers, T., contributors: Torchvision: Pytorch's computer vision library. GitHub repository (2016) +39. Molchanov, P., Mallya, A., Tyree, S., Frosio, I., Kautz, J.: Importance estimation for neural network pruning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11264-11272 (2019) +40. Molchanov, P., Tyree, S., Karras, T., Aila, T., Kautz, J.: Pruning convolutional neural networks for resource efficient inference. arXiv preprint arXiv:1611.06440 (2016) +41. Peng, H., Wu, J., Chen, S., Huang, J.: Collaborative channel pruning for deep networks. In: International Conference on Machine Learning. pp. 5113-5122. PMLR (2019) +42. Pool, J., Sawarkar, A., Rodge, J.: Accelerating inference with sparsity using the nvidia ampere architecture and nvidia tensorsr. NVIDIA Developer Technical Blog, https://developer.nvidia.com/blog/accelerating-inference-with-sparsityusing-ampere-and-tensorr (2021) +43. Radosavovic, I., Kosaraju, R.P., Girshick, R., He, K., Dollár, P.: Designing network design spaces. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10428-10436 (2020) +44. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: Mobilenetv2: Inverted residuals and linear bottlenecks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4510-4520 (2018) +45. Shi, D., Tao, C., Jin, Y., Yang, Z., Yuan, C., Wang, J.: UPop: Unified and progressive pruning for compressing vision-language transformers. In: Proceedings of the 40th International Conference on Machine Learning. vol. 202, pp. 31292-31311. PMLR (2023) +46. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) +47. Song, Z., Xu, Y., He, Z., Jiang, L., Jing, N., Liang, X.: Cp-vit: Cascade vision transformer pruning via progressive sparsity prediction. arXiv preprint arXiv:2203.04570 (2022) +48. Sun, M., Liu, Z., Bair, A., Kolter, J.Z.: A simple and effective pruning approach for large language models. arXiv preprint arXiv:2306.11695 (2023) + +49. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818-2826 (2016) +50. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers & distillation through attention. In: International Conference on Machine Learning. vol. 139, pp. 10347-10357 (July 2021) +51. Touvron, H., Cord, M., Sablayrolles, A., Synnaeve, G., Jégou, H.: Going deeper with image transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 32-42 (2021) +52. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +53. Voita, E., Talbot, D., Moiseev, F., Sennrich, R., Titov, I.: Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. arXiv preprint arXiv:1905.09418 (2019) +54. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media 8(3), 415-424 (2022) +55. Wang, Z., Luo, H., Wang, P., Ding, F., Wang, F., Li, H.: Vtc-lfc: Vision transformer compression with low-frequency components. Advances in Neural Information Processing Systems 35, 13974-13988 (2022) +56. Wen, W., Wu, C., Wang, Y., Chen, Y., Li, H.: Learning structured sparsity in deep neural networks. Advances in neural information processing systems 29 (2016) +57. Wightman, R.: Pytorch-image-models. https://github.com/huggingface/pytorch-image-models +58. Yang, H., Yin, H., Shen, M., Molchanov, P., Li, H., Kautz, J.: Global vision transformer pruning with hessian-aware saliency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18547-18557 (2023) +59. Ye, J., Lu, X., Lin, Z., Wang, J.Z.: Rethinking the smaller-norm-less-informative assumption in channel pruning of convolution layers. arXiv preprint arXiv:1802.00124 (2018) +60. Yin, M., Uzkent, B., Shen, Y., Jin, H., Yuan, B.: Gohsp: A unified framework of graph and optimization-based heterogeneous structured pruning for vision transformer. arXiv preprint arXiv:2301.05345 (2023) +61. Yu, F., Huang, K., Wang, M., Cheng, Y., Chu, W., Cui, L.: Width & depth pruning for vision transformers. In: AAAI Conference on Artificial Intelligence (AAAI). vol. 2022 (2022) +62. Yu, H., Wu, J.: A unified pruning framework for vision transformers. Science China Information Sciences 66(7), 1-2 (2023) +63. Yu, J., Huang, T.: Autoslim: Towards one-shot architecture search for channel numbers. arXiv preprint arXiv:1903.11728 (2019) +64. Yu, L., Xiang, W.: X-pruner: explainable pruning for vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24355-24363 (2023) +65. Yu, S., Chen, T., Shen, J., Yuan, H., Tan, J., Yang, S., Liu, J., Wang, Z.: Unified visual transformer compression. arXiv preprint arXiv:2203.08243 (2022) +66. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6023-6032 (2019) +67. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017) + +68. Zheng, C., Zhang, K., Yang, Z., Tan, W., Xiao, J., Ren, Y., Pu, S., et al.: Savit: Structure-aware vision transformer pruning via collaborative optimization. Advances in Neural Information Processing Systems 35, 9010-9023 (2022) +69. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 13001-13008 (2020) +70. Zhu, M., Tang, Y., Han, K.: Vision transformer pruning. arXiv preprint arXiv:2104.08500 (2021) \ No newline at end of file diff --git a/2024/Isomorphic Pruning for Vision Models/images.zip b/2024/Isomorphic Pruning for Vision Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..db8e30309bf18e79109aec18737e5b3776ad2c02 --- /dev/null +++ b/2024/Isomorphic Pruning for Vision Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ecf2c8c95391feaff9a2945dde676509f03fae11b4be8c85fb4d7386dd695056 +size 576614 diff --git a/2024/Isomorphic Pruning for Vision Models/layout.json b/2024/Isomorphic Pruning for Vision Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..727eae34e4776cccfae06ca1bd96de83c47d526b --- /dev/null +++ b/2024/Isomorphic Pruning for Vision Models/layout.json @@ -0,0 +1,10180 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 172, + 112, + 442, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 112, + 442, + 129 + ], + "spans": [ + { + "bbox": [ + 172, + 112, + 442, + 129 + ], + "type": "text", + "content": "Isomorphic Pruning for Vision Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 149, + 470, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 149, + 470, + 163 + ], + "spans": [ + { + "bbox": [ + 145, + 149, + 470, + 163 + ], + "type": "text", + "content": "Gongfan Fang" + }, + { + "bbox": [ + 145, + 149, + 470, + 163 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 149, + 470, + 163 + ], + "type": "text", + "content": ", Xinyin Ma" + }, + { + "bbox": [ + 145, + 149, + 470, + 163 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 149, + 470, + 163 + ], + "type": "text", + "content": ", Michael Bi Mi" + }, + { + "bbox": [ + 145, + 149, + 470, + 163 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 149, + 470, + 163 + ], + "type": "text", + "content": ", and Xinchao Wang" + }, + { + "bbox": [ + 145, + 149, + 470, + 163 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 235, + 171, + 378, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 171, + 378, + 183 + ], + "spans": [ + { + "bbox": [ + 235, + 171, + 378, + 183 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 235, + 171, + 378, + 183 + ], + "type": "text", + "content": " National University of Singapore" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 249, + 183, + 364, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 183, + 364, + 194 + ], + "spans": [ + { + "bbox": [ + 249, + 183, + 364, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 249, + 183, + 364, + 194 + ], + "type": "text", + "content": " Huawei Technologies Ltd." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 219, + 195, + 395, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 195, + 395, + 206 + ], + "spans": [ + { + "bbox": [ + 219, + 195, + 395, + 206 + ], + "type": "text", + "content": "gongfan@u.nus.edu，xinchao@nus.edu.sg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 228, + 455, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 228, + 455, + 502 + ], + "spans": [ + { + "bbox": [ + 159, + 228, + 455, + 502 + ], + "type": "text", + "content": "Abstract. Structured pruning reduces the computational overhead of deep neural networks by removing redundant sub-structures. However, assessing the relative importance of different sub-structures remains a significant challenge, particularly in advanced vision models featuring novel mechanisms and architectures like self-attention, depth-wise convolutions, or residual connections. These heterogeneous substructures usually exhibit diverged parameter scales, weight distributions, and computational topology, introducing considerable difficulty to importance comparison. To overcome this, we present Isomorphic Pruning, a simple approach that demonstrates effectiveness across a range of network architectures such as Vision Transformers and CNNs, and delivers competitive performance across different model sizes. Isomorphic Pruning originates from an observation that, when evaluated under a pre-defined importance criterion, heterogeneous sub-structures demonstrate significant divergence in their importance distribution, as opposed to isomorphic structures that present similar importance patterns. This inspires us to perform isolated ranking and comparison on different types of sub-structures for more reliable pruning. Our empirical results on ImageNet-1K demonstrate that Isomorphic Pruning surpasses several pruning baselines dedicatedly designed for Transformers or CNNs. For instance, we improve the accuracy of DeiT-Tiny from " + }, + { + "bbox": [ + 159, + 228, + 455, + 502 + ], + "type": "inline_equation", + "content": "74.52\\%" + }, + { + "bbox": [ + 159, + 228, + 455, + 502 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 159, + 228, + 455, + 502 + ], + "type": "inline_equation", + "content": "77.50\\%" + }, + { + "bbox": [ + 159, + 228, + 455, + 502 + ], + "type": "text", + "content": " by pruning an off-the-shelf DeiT-Base model. And for ConvNext-Tiny, we enhanced performance from " + }, + { + "bbox": [ + 159, + 228, + 455, + 502 + ], + "type": "inline_equation", + "content": "82.06\\%" + }, + { + "bbox": [ + 159, + 228, + 455, + 502 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 159, + 228, + 455, + 502 + ], + "type": "inline_equation", + "content": "82.18\\%" + }, + { + "bbox": [ + 159, + 228, + 455, + 502 + ], + "type": "text", + "content": ", while reducing the number of parameters and memory usage. Code is available at https://github.com/VainF/Isomorphic-Pruning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 512, + 410, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 512, + 410, + 524 + ], + "spans": [ + { + "bbox": [ + 160, + 512, + 410, + 524 + ], + "type": "text", + "content": "Keywords: Network Pruning " + }, + { + "bbox": [ + 160, + 512, + 410, + 524 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 512, + 410, + 524 + ], + "type": "text", + "content": " Vision Transformers " + }, + { + "bbox": [ + 160, + 512, + 410, + 524 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 512, + 410, + 524 + ], + "type": "text", + "content": " CNNs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 541, + 230, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 541, + 230, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 541, + 230, + 553 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 563, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 648 + ], + "type": "text", + "content": "Deep Neural Networks are typically over-parameterized, containing numerous redundant sub-structures. Such redundancy can be effectively eliminated via structured pruning, without significantly sacrificing the network's performance [16, 22, 25]. Pruning usually follows a \"ranking-to-prune\" framework: To facilitate compression while retaining the learned capabilities of networks, a carefully-designed criterion is utilized to assess the importance score of different structures [11, 22, 40], followed by a ranking process to identify and remove those" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 238, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 238, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 238, + 665 + ], + "type": "text", + "content": "* Corresponding Author." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 155, + 118, + 455, + 266 + ], + "blocks": [ + { + "bbox": [ + 155, + 118, + 455, + 266 + ], + "lines": [ + { + "bbox": [ + 155, + 118, + 455, + 266 + ], + "spans": [ + { + "bbox": [ + 155, + 118, + 455, + 266 + ], + "type": "image", + "image_path": "5d6fc412fc9fe29a137bf1527c879ab991715f988911296aaef04fe3e8d941bd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 279, + 482, + 324 + ], + "lines": [ + { + "bbox": [ + 130, + 279, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 482, + 324 + ], + "type": "text", + "content": "Fig. 1: ImageNet Top-1 Accuracy vs. Multiply-Accumulate Operations (MACs) of pruned DeiT [50] and pruned ConvNext [33]. The pruned models, marked as “★” have comparable or better latency yet superior performance compared to scratch training counterparts highlighted by “○”. The MACs are in log scale for better visualization" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 351, + 479, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 479, + 375 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 479, + 375 + ], + "type": "text", + "content": "relatively unimportant ones. After pruning, the compressed model is fine-tuned to recover its performance [11, 18]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 481, + 604 + ], + "type": "text", + "content": "However, ranking-based structured pruning presupposes an underlying assumption that the parameters from different structures are directly comparable. Classic deep neural networks, such as VGG [46], are built by stacking similar convolutional layers or blocks where the neuron connections share a similar computational topology. Under the circumstances, a straightforward global ranking is expected to reveal their relative importance and remove unimportant parameters [11, 22]. However, with recent innovation in network structure design, advanced vision models like Vision Transformers [10, 31, 50] and ConvNext [33] incorporate more sophisticated designs, such as complicated residual connections [19, 24], attention mechanisms [52] and depth-wise convolutions [6]. These novel structures increase the network's internal complexity and heterogeneity, challenging the reliability of pruning importance criteria. For instance, a vanilla Vision Transformer [10] is composed of MLP, Multi-head Attention, and skip connections. Even in the same Multi-head attention layer, there are two pruning schemes to removing either entire attention heads [37, 53] or the within-head dimensions [58, 68]. These structures vary by their computational topologies and, as will be illustrated in the experiments, tend to exhibit diverged importance distributions. In this case, directly comparing the importance of these heterogeneous might be problematic." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "To address the problem, we propose a simple pruning approach, termed Isomorphic Pruning, to alleviate the incomparability issue arising from heterogeneous sub-structures. The essence of this method lies in decomposing a network into multiple isomorphic groups, as illustrated in Figure 2 (d), based on the computational topology of the sub-structures. Each isomorphic group contains" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "G. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 127, + 192, + 183 + ], + "blocks": [ + { + "bbox": [ + 157, + 116, + 192, + 125 + ], + "lines": [ + { + "bbox": [ + 157, + 116, + 192, + 125 + ], + "spans": [ + { + "bbox": [ + 157, + 116, + 192, + 125 + ], + "type": "text", + "content": "(a) Network" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 157, + 127, + 192, + 183 + ], + "lines": [ + { + "bbox": [ + 157, + 127, + 192, + 183 + ], + "spans": [ + { + "bbox": [ + 157, + 127, + 192, + 183 + ], + "type": "image", + "image_path": "c60658c4438fb36ae5860712318177d6e42191850ac2abd987ea02b23ec8875d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 222, + 127, + 275, + 182 + ], + "blocks": [ + { + "bbox": [ + 215, + 116, + 263, + 125 + ], + "lines": [ + { + "bbox": [ + 215, + 116, + 263, + 125 + ], + "spans": [ + { + "bbox": [ + 215, + 116, + 263, + 125 + ], + "type": "text", + "content": "(b) Local Pruning" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 222, + 127, + 275, + 182 + ], + "lines": [ + { + "bbox": [ + 222, + 127, + 275, + 182 + ], + "spans": [ + { + "bbox": [ + 222, + 127, + 275, + 182 + ], + "type": "image", + "image_path": "1e43295c1daf7b27db9057b7ee834b9531f436bccda89115995b1721c48a6d20.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 293, + 127, + 347, + 183 + ], + "blocks": [ + { + "bbox": [ + 285, + 116, + 334, + 125 + ], + "lines": [ + { + "bbox": [ + 285, + 116, + 334, + 125 + ], + "spans": [ + { + "bbox": [ + 285, + 116, + 334, + 125 + ], + "type": "text", + "content": "(c) Global Pruning" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 293, + 127, + 347, + 183 + ], + "lines": [ + { + "bbox": [ + 293, + 127, + 347, + 183 + ], + "spans": [ + { + "bbox": [ + 293, + 127, + 347, + 183 + ], + "type": "image", + "image_path": "89d548bb24841bb9486b5878797ae9f1fe8eae9eb5d8700b20c5defcc3a35728.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 363, + 126, + 462, + 184 + ], + "blocks": [ + { + "bbox": [ + 372, + 116, + 433, + 125 + ], + "lines": [ + { + "bbox": [ + 372, + 116, + 433, + 125 + ], + "spans": [ + { + "bbox": [ + 372, + 116, + 433, + 125 + ], + "type": "text", + "content": "(d) Isomorphic Pruning" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 363, + 126, + 462, + 184 + ], + "lines": [ + { + "bbox": [ + 363, + 126, + 462, + 184 + ], + "spans": [ + { + "bbox": [ + 363, + 126, + 462, + 184 + ], + "type": "image", + "image_path": "6dc360afa098946100a2127d209eb63a33d41a02f1042c44a1dc61033f2e96b4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 194, + 482, + 261 + ], + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 261 + ], + "type": "text", + "content": "Fig. 2: For a pre-trained network (a), we show three pruning strategies: (b) Local Pruning that compares parameter importance within current layers; (c) Global Pruning that performs global ranking for all parameters; (d) the proposed Isomorphic pruning that groups parameters by the computational topology and applies importance ranking within groups. In each group, the importance distributions are more comparable. Details about the distribution can be found in Fig. 4 of the experiments section." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 285, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 308 + ], + "type": "text", + "content": "elements with the same structure and tends to present a similar distribution of importance, making it more rational to compare their relative importance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 309, + 482, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 482, + 476 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 482, + 476 + ], + "type": "text", + "content": "To validate the effectiveness of our method, we conducted experiments on a variety of visual models, including DeiT [50], ConvNext [33], ResNet [19], and MobileNet-V2 [44]. Pruning results, as shown in Figure 1, on ImageNet-1K [8] demonstrate that Isomorphic Pruning achieves comparable or even superior performance relative to pruning algorithms specifically designed for CNNs or transformers. For instance, by pruning a pre-trained large DeiT model, we enhanced the performance of the DeiT-T model from " + }, + { + "bbox": [ + 130, + 309, + 482, + 476 + ], + "type": "inline_equation", + "content": "74.52\\%" + }, + { + "bbox": [ + 130, + 309, + 482, + 476 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 309, + 482, + 476 + ], + "type": "inline_equation", + "content": "77.50\\%" + }, + { + "bbox": [ + 130, + 309, + 482, + 476 + ], + "type": "text", + "content": ", when compared to the scratch training one. Similarly, our algorithm also yields satisfactory results on CNNs. We constructed compact ConvNexts and ResNets via pruning, which has fewer MACs and parameters while surpassing a conventionally trained model. Additionally, we also studied the real-world latency and memory footprint of the compressed models on both GPU and CPU devices. Our findings indicate that models pruned with Isomorphic Pruning achieve significant actual speed-ups and are competitive to carefully designed efficient architectures." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 476, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 476, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 476, + 482, + 525 + ], + "type": "text", + "content": "Therefore, Our contribution lies in a practical and general pruning approach, which can handle various vision models with novel mechanisms and architectures. The pruned models show superior performance to several pruning competitors on ImageNet-1K benchmark." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 543, + 243, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 543, + 243, + 556 + ], + "spans": [ + { + "bbox": [ + 131, + 543, + 243, + 556 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Network Pruning. Deep neural networks are typically over-parameterized and can be pruned without significant performance drop [16, 25]. The exploration of network pruning can be dated back to early works [1, 17, 25] about neural networks. Over the past few years, a large number of works have been conducted on convolutional networks to verify the effectiveness of pruning, which can be roughly categorized into unstructured [16, 29, 35] and structured pruning [5, 11, 18, 22, 56]. Unstructured pruning compresses a pre-trained model by forcing partial weights to be zero, leading sparse networks [29]. To accelerate" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 448, + 102 + ], + "type": "text", + "content": "Isomorphic Pruning for Vision Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "the sparse model, special patterns and devices will be used such as N:M sparsity [12, 42, 48]. Structured pruning provides a hardware-friendly approach by removing sub-structures such as channels [11] or blocks [14] in CNNs, which improves the model efficiency without dedicated accelerators. In recent years, a large number of works adapted the idea of network pruning to vision transformers. It demonstrated that even the vanilla vision transformers can be squeezed for a better efficiency-performance trade-off [2, 4, 45, 47, 58, 60–62, 64, 65, 68, 70]. In the literature, transformer pruning has been studied from several perspectives, such as width reduction [70], depth pruning [61], attention head removal [58], sparse training [4] and architecture transformation [18]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 255, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 413 + ], + "type": "text", + "content": "Pruning by Global Ranking. In the context of network pruning, global pruning is recognized as a pivotal topic for exploration. This method entails a comprehensive ranking of all network layers, assigning adaptive pruning ratios to each layer according to the ranks [5,11,22,58]. Different layers within a network often exhibit varying levels of redundancy. To this end, the advantage of global pruning lies in its ability to create a more parameter-efficient, non-uniformly compressed model [58]. However, a key challenge remains in the ranking: how to ensure the consistency and comparability across different layers? A widely-used solution for this issue is normalizing scores on a per-layer basis to even out the importance distribution across the network [30]. Alternatively, some approaches use sparse training to learn an effective target structure [11], which might be costly due to the additional optimization. In this work, we introduce a more straightforward approach termed isomorphic pruning to facilitate a reliable ranking for pruning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 431, + 202, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 202, + 443 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 202, + 443 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 458, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 482, + 507 + ], + "type": "text", + "content": "The core idea of Isomorphic Pruning is straightforward: we group sub-structures in a network by their network topology and perform importance ranking within groups. This imposes two questions 1) how to identify removable sub-structures in networks and 2) how to compare the topology graph for grouping and pruning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 525, + 291, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 291, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 291, + 537 + ], + "type": "text", + "content": "3.1 Sub-structure in Network" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 548, + 482, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 631 + ], + "type": "text", + "content": "Structured Pruning aims to remove sub-structures in networks, such as the dimensions of hidden states or features. To make the formulation clear, we primarily concentrate on stacked fully-connected networks, although the principles established herein can be readily extended to CNNs and transformers. Initially, we consider a single linear layer which is parameterized by a matrix " + }, + { + "bbox": [ + 130, + 548, + 482, + 631 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{m \\times n}" + }, + { + "bbox": [ + 130, + 548, + 482, + 631 + ], + "type": "text", + "content": ". To articulate the pruning process, we introduce a pruning function " + }, + { + "bbox": [ + 130, + 548, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\mathrm{g}(W,d,k)" + }, + { + "bbox": [ + 130, + 548, + 482, + 631 + ], + "type": "text", + "content": " defined as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 639, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 639, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 143, + 639, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\hat {w} = \\mathrm {g} (W, d, k) = \\left\\{ \\begin{array}{l l} \\mathrm {g} (W, 0, k) = W _ {[ k,: ]} \\in \\mathbb {R} ^ {1 \\times n}, & \\text {i f p r u n i n g o u t p u t d i m e n s i o n ,} \\\\ \\mathrm {g} (W, 1, k) = W _ {[:, k ]} \\in \\mathbb {R} ^ {m \\times 1}, & \\text {i f p r u n i n g i n p u t d i m e n s i o n ,} \\end{array} \\right. \\tag {1}", + "image_path": "d96639e7c3e784632b1355d48cb987b9d1e74c3a1325bde03bfc6ba88fc8bad8.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "G. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\hat{w}" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " refers to the weight column or row slated for removal, acquired via a slicing operation denoted by subscript " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "[:,k]" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "[k,:]" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": ". The pruning function " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{g}(W,d,k)" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " is defined by three parameters, the parameter matrix " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": ", the axis " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "d\\in \\{0,1\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " indicating the dimension to be compressed, and the index " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " identifying the " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "-th rows or columns in " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": ". In the context of a linear layer represented by a 2-D matrix " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "W\\in \\mathbb{R}^{m\\times n}" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": ", the axis " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " can assume values of either 0 or 1, with " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " having a permissible range of " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "[0,m]" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "[0,n]" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": ", respectively. This formulation encapsulates the pruning of a single layer; to extend it to the entire network, we also need to take the dependency [5,11,30] into account, where parameters from different layers are coupled and should be pruned simultaneously. As depicted in Figure 3, for two adjacent linear layers parameterized by " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "W_{1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": ", the output dimension of the former is integrally coupled to the input dimension of the latter, thereby constituting a minimal sub-structure eligible for removal. If we treat each pruning operation " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{g}(W,d,k)" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " as vertices defined by a triplet " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "(W,d,k)" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": " and the dependency as edges, then it is natural to model sub-structures as graphs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 313, + 334, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 313, + 334, + 326 + ], + "spans": [ + { + "bbox": [ + 131, + 313, + 334, + 326 + ], + "type": "text", + "content": "3.2 Graph Modeling of Sub-structures" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "text", + "content": "In this section, we delve deeper into the methodology for decomposing a network into several independent and minimal sub-structures in pruning, and model them as graphs. We can begin with the MLP network as illustrated in Figure 3, where the activation functions are omitted. The network comprises a sequence of 2-D parameter matrices " + }, + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "inline_equation", + "content": "\\{W_1, W_2, \\dots, W_L\\}" + }, + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "text", + "content": " denoting the network depth. Initially, all parameters are not assigned to any sub-structure. We randomly select one parameter " + }, + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "text", + "content": " as the seed to find its associated substructures and build the corresponding graph. For instance, selecting parameter " + }, + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "inline_equation", + "content": "W_1" + }, + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "text", + "content": " and setting the axis " + }, + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "inline_equation", + "content": "d = 0" + }, + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "text", + "content": " allows us to establish an initial graph " + }, + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "inline_equation", + "content": "G = (V = \\{(W, 0, k)\\}, E = \\{\\})" + }, + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "text", + "content": ", focusing on the elimination of " + }, + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "text", + "content": "-th the output dimension in " + }, + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "inline_equation", + "content": "W_1" + }, + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "text", + "content": ". Then we iterate through all potential " + }, + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "inline_equation", + "content": "(W', d', k') \\notin V" + }, + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "text", + "content": " to ascertain any dependency to " + }, + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "inline_equation", + "content": "(W, d, k) \\in V" + }, + { + "bbox": [ + 130, + 333, + 482, + 500 + ], + "type": "text", + "content": ". Dependencies between layers can be ascertained through either manually-crafted patterns or automatic algorithms. In this work, we adhere to the approach of [11] to assess dependency based on two rudimentary rules:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "text", + "content": "Dependency Modeling. For a pair of triplets " + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "inline_equation", + "content": "(W', d', k')" + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "inline_equation", + "content": "(W, d, k)" + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "text", + "content": ", a dependency is established if either of the following criteria is met: 1) " + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "inline_equation", + "content": "W'" + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "text", + "content": " are adjacent layers with " + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "inline_equation", + "content": "d = d'" + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "inline_equation", + "content": "k = k'" + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "text", + "content": "; 2) " + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "inline_equation", + "content": "W' = W" + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathrm{g}(W', d', k') = \\mathrm{g}(W, d, k)" + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "text", + "content": ". The first criterion addresses adjacent layers as exemplified in Figure 3, while the second pertains to specialized layers, such as normalization layers, which share the same pruning pattern for both outputs and inputs. Thereafter, we update the vertex set " + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "text", + "content": " and edge set " + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "text", + "content": " with the discovered dependency between " + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "inline_equation", + "content": "(W', d', k')" + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "inline_equation", + "content": "(W, d, k)" + }, + { + "bbox": [ + 130, + 500, + 482, + 596 + ], + "type": "text", + "content": " accordingly:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 233, + 605, + 480, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 605, + 480, + 624 + ], + "spans": [ + { + "bbox": [ + 233, + 605, + 480, + 624 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} V = V \\cup \\left\\{\\left(W ^ {\\prime}, d ^ {\\prime}, k ^ {\\prime}\\right) \\right\\}, \\\\ E = E + \\left\\{\\left(\\left(W _ {1} - 1\\right), \\left(W ^ {\\prime}, d ^ {\\prime}, 1\\right)\\right) \\right\\} \\end{array} \\tag {2}", + "image_path": "083fcbb4ed3903c2c20b2a575f42c4bab8e3fca97c0a14aead33f79212d38973.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 233, + 620, + 381, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 620, + 381, + 632 + ], + "spans": [ + { + "bbox": [ + 233, + 620, + 381, + 632 + ], + "type": "interline_equation", + "content": "E = E \\cup \\left\\{\\left(\\left(W, d, k\\right), \\left(W ^ {\\prime}, d ^ {\\prime}, k ^ {\\prime}\\right)\\right) \\right\\}.", + "image_path": "2e848ab20cf49d0b18d80dd900b863f4904a0202074a488b2e36081dfbd673e4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "We can further repeat the above process until no additional vertex is incorporated into the current graph. The resultant graph " + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": " then serves as a topological" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 448, + 102 + ], + "type": "text", + "content": "Isomorphic Pruning for Vision Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 179, + 114, + 438, + 236 + ], + "blocks": [ + { + "bbox": [ + 179, + 114, + 438, + 236 + ], + "lines": [ + { + "bbox": [ + 179, + 114, + 438, + 236 + ], + "spans": [ + { + "bbox": [ + 179, + 114, + 438, + 236 + ], + "type": "image", + "image_path": "c39cbbbb1c8a68e49ea3f7544e05c8267b8b5b5ca0e7a6687cdf9a9bc326b2cc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 245, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 131, + 245, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 131, + 245, + 482, + 291 + ], + "type": "text", + "content": "Fig. 3: Isomorphic Pruning models sub-structures as graphs, and apply isolated ranking and pruning with isomorphic structures. We highlight three removable substructures in an MLP and show their corresponding graphs below. Substructures 1 & 3 are isomorphic but 2 & 3 are non-isomorphic due to the additional residual connections." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 314, + 482, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 374 + ], + "type": "text", + "content": "representation of a single sub-structure associated with the initial seed parameter. To map all substructures onto different graphs, we subsequently embark on a recursive procedure, commencing from those nodes not yet allocated to any sub-structure. Figure 3 furnishes an exemplar of this process, showcasing three distinct substructures labeled as sub-structures 1, 2, and 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 392, + 334, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 392, + 334, + 404 + ], + "spans": [ + { + "bbox": [ + 131, + 392, + 334, + 404 + ], + "type": "text", + "content": "3.3 Ranking with Graph Isomorphism" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 413, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 498 + ], + "type": "text", + "content": "With the graph-based representation of network sub-structures, we advance to the ranking phase. This stage involves estimating and ranking the significance of each sub-structure to identify redundancies. A pivotal component of this process is an importance criterion, which quantifies the relative significance of various sub-structures. Previous works in the literature presents numerous well-crafted criteria, including magnitude pruning and Taylor expansion. In this work, we focus on the general importance function " + }, + { + "bbox": [ + 130, + 413, + 482, + 498 + ], + "type": "inline_equation", + "content": "I(\\cdot)" + }, + { + "bbox": [ + 130, + 413, + 482, + 498 + ], + "type": "text", + "content": " for a single layer, represented as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 166, + 506, + 481, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 506, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 166, + 506, + 481, + 548 + ], + "type": "interline_equation", + "content": "I (W, d, k) = \\left\\{ \\begin{array}{l l} \\| \\mathrm {g} (W, d, k)) \\|, & \\text {i f M a g n i t u d e P r u n i n g} \\\\ \\| \\frac {\\partial \\mathcal {L}}{\\partial \\mathrm {g} (W , d , k)} \\mathrm {g} (W, d, k)) \\| _ {2}, & \\text {i f T a y l o r P r u n i n g} \\\\ \\dots & \\text {O t h e r i m p o r t a n c e C r i t e r i a .} \\end{array} \\right. \\tag {3}", + "image_path": "0a9448224e6752a27108de1b27edf4c21880dc2213baf6c260b8ddc286e3b459.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 555, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 555, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 555, + 482, + 605 + ], + "type": "text", + "content": "In structural pruning, our objective is to remove sub-structures encompassing parameters across multiple layers. Existing approaches [5,11,30] generally extend the single-layer importance to a holistic measure by aggregating scores across all constituents " + }, + { + "bbox": [ + 131, + 555, + 482, + 605 + ], + "type": "inline_equation", + "content": "(W,d,k)" + }, + { + "bbox": [ + 131, + 555, + 482, + 605 + ], + "type": "text", + "content": " for the vertexes set, denoted as " + }, + { + "bbox": [ + 131, + 555, + 482, + 605 + ], + "type": "inline_equation", + "content": "I^{*}(\\cdot)" + }, + { + "bbox": [ + 131, + 555, + 482, + 605 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 222, + 613, + 481, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 613, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 222, + 613, + 481, + 634 + ], + "type": "interline_equation", + "content": "I ^ {*} (G (V, E)) = \\sum_ {(W, d, k) \\in V} I (W, d, k), \\tag {4}", + "image_path": "27b1511c6c64694c92697aa3b937eab9af5854fd0804cacf5d479e9091fc7ff1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "thereby cumulating the importance of all elements within graph " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": ". However, the inherent diversity in parameters and computational topology among different" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "G. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": "substructures " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " introduces potential ambiguities in this metric. Experimental evidence, particularly in the context of a vision transformer as shown in Figure 4, elucidates this issue: the largest substructure encompasses all embedding dimensions, containing a substantial number of parameters, whereas the most minimal sub-structure comprises merely two linear layers. Consequently, such aggregation may fail to accurately reflect the true relative importance of these entities. To address the challenges posed by the heterogeneity of substructures, we delineate a simple approach designed to preserve the nuanced information encapsulated by the aggregated importance, as outlined in Equation 4, whilst alleviating the detrimental repercussions raised by the heterogeneity of substructures. Our principal strategy entails the clustering of sub-structures into various isomorphic groups, predicated on their topological logic and parameter configurations. In this case, applying aggregated importance within groups can be meaningful, since isomorphic sub-structures show one-to-one correspondence among their parameters, and how different parameters participate in computations is entirely identical. To facilitate this, the key step is to examine the isomorphism of two graphs " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "G_{1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "G_{2}" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": ". Formally, we adopt the following definition to ascertain graph isomorphism." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 337, + 483, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 337, + 483, + 374 + ], + "spans": [ + { + "bbox": [ + 131, + 337, + 483, + 374 + ], + "type": "text", + "content": "Definition 1 (Graph Isomorphism). Two labeled graphs " + }, + { + "bbox": [ + 131, + 337, + 483, + 374 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 131, + 337, + 483, + 374 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 337, + 483, + 374 + ], + "type": "inline_equation", + "content": "G'" + }, + { + "bbox": [ + 131, + 337, + 483, + 374 + ], + "type": "text", + "content": " are isomorphic if there exists a bijection between their vertices so that, all mapped vertices preserve the same labels, and are connected in the same way." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 380, + 482, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 380, + 482, + 404 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 482, + 404 + ], + "type": "text", + "content": "To develop a general indicator Isomorphic " + }, + { + "bbox": [ + 131, + 380, + 482, + 404 + ], + "type": "inline_equation", + "content": "(\\mathrm{g}_1,\\mathrm{g}_2)" + }, + { + "bbox": [ + 131, + 380, + 482, + 404 + ], + "type": "text", + "content": ", we need to examine both the label and connectivity in two graphs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": "Vertex Label and Connectivity. In our settings, each vertex is labeled by the triplet " + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "inline_equation", + "content": "(W,d,k)" + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": ", encompassing the parameter matrix " + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": ", axis " + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": ", and index " + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": ". We follow a simple principle for labling: Two vertices " + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "inline_equation", + "content": "(\\mathbf{W},d,k)" + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "inline_equation", + "content": "(\\mathbf{W}',d',k')" + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": " are labeled identically if and only if the parameter " + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{W}'" + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": " originates from the same layer type such as Linear or Conv, with the same pruning dimension " + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "inline_equation", + "content": "d = d'" + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": ". We permit the index " + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": " to differ, since within the same parameter matrix, the weight vectors sliced along the " + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": "-th dimension are consistently homogeneous and thus can be considered isomorphic. After determining the label for each vertex, we further explore the connectivity in two graphs " + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "inline_equation", + "content": "G_{1}" + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "inline_equation", + "content": "G_{2}" + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": ". As illustrated in Figure 3, we show three substructures alongside their respective graphs " + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "inline_equation", + "content": "G_{1}" + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "inline_equation", + "content": "G_{2}" + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "inline_equation", + "content": "G_{3}" + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "inline_equation", + "content": "G_{1}" + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "inline_equation", + "content": "G_{3}" + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": " incorporate a singular edge, whereas " + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "inline_equation", + "content": "G_{2}" + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": " encompasses three edges. If we check the edges following the execution order of the computational graphs, it's natural to observe the isomorphism of " + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "inline_equation", + "content": "G_{1}" + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "inline_equation", + "content": "G_{3}" + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": ", since they share the same topology. To make this process more formal, we assume the edges in graph " + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": " are already sorted according to their execution order in forwarding, and check the isomorphism by" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 621, + 481, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 621, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 152, + 621, + 481, + 633 + ], + "type": "interline_equation", + "content": "\\operatorname {I s o m o r p h i c} \\left(G, G ^ {\\prime}\\right)) = \\mathbb {1} \\left\\{\\left| G \\right| = \\left| G ^ {\\prime} \\right| \\wedge \\operatorname {l a b e l} \\left(V _ {i}\\right) = \\operatorname {l a b e l} \\left(V _ {i} ^ {\\prime}\\right); \\forall i \\in \\{1, 2, \\dots , \\left| G ^ {\\prime} \\right| \\} \\right\\},", + "image_path": "0e9e54fcde4733ded03c49e3a4b2093334a338ebd7077824145408e879f0e632.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbb{1}" + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": " is an indicator function. " + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "|G^{\\prime}|" + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": " denotes the number of edges in the graph and " + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\wedge" + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": " signifies the logical AND operation. The first condition ensures the equiva" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 448, + 102 + ], + "type": "text", + "content": "Isomorphic Pruning for Vision Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 130, + 462, + 358 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 287, + 129 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 287, + 129 + ], + "type": "text", + "content": "Algorithm 1 Isomorphic Pruning" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "lines": [ + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": "1: Input: Network with parameters " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "\\{W_1, W_2, \\ldots, W_n\\}" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": ", Pruning ratio " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "p\\%" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": ". \n2: Output: Pruned Network " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "\\{W_1', W_2', \\ldots, W_n'\\}" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": " \n3: procedure IDENTIFYSUBSTRUCTURES(\\{W1, W2, \\ldots, Wn\\}) \n4: Initialize empty list of sub-structures " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = \\{\\}" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": " \n5: for each layer " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "W_i" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "\\{W_1, W_2, \\ldots, W_n\\}" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": " do \n6: if " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "W_i" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": " is not assigned to any substructure then \n7: Initialize a graph " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "G(V = \\{W_i\\}, E = \\{\\})" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": ". \n8: Find and add coupled " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "W_j" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": " until there is no new dependency. \n9: " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = \\mathbf{S} \\cup \\{G\\}" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": " \n10: return " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": " \n11: procedure PRUNEBYGROUPEDRANKING(S, p) \n12: for each substructure " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "G_i" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "\\{G_1, G_2, \\ldots, G_n\\}" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": " do \n13: if " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "G_i" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": " is not assigned to any group then \n14: Initialize a Group " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "\\mathbf{R} = \\{G_i\\}" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": ". \n15: Find and add " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "G_j" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": " that satisfies Isomorphism " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "(G_i, G_j) = 1" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": ". \n16: Calculate the importance scores " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "\\{I^*(G) | G \\in \\mathbf{R}\\}" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": ". \n17: Remove " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "p\\%" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": " substructures in " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": " by ranking the importance scores. \n18: return " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": " \n19: " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\gets \\text{IDENTIFYSUBSTRUCTURES}(\\{W_1, W_2, \\ldots, W_n\\})" + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "text", + "content": " \n20: " + }, + { + "bbox": [ + 132, + 130, + 462, + 358 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\gets \\text{PRUNEBYGROUPEDRANKING}(\\mathbf{S}, p)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 384, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 480, + 407 + ], + "type": "text", + "content": "lence of the number of vertices in the graphs, while the second condition confirms the matching labels of vertices on either side of the edge pairs " + }, + { + "bbox": [ + 130, + 384, + 480, + 407 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 130, + 384, + 480, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 384, + 480, + 407 + ], + "type": "inline_equation", + "content": "V_{i}^{\\prime}" + }, + { + "bbox": [ + 130, + 384, + 480, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 408, + 481, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 481, + 539 + ], + "type": "text", + "content": "Using Equation 5, it is easy to cluster substructures into different groups and perform rankings within each group exclusively. Specifically, we examine a collection of substructures " + }, + { + "bbox": [ + 130, + 408, + 481, + 539 + ], + "type": "inline_equation", + "content": "\\{G_1, G_2, \\ldots, G_N\\}" + }, + { + "bbox": [ + 130, + 408, + 481, + 539 + ], + "type": "text", + "content": " that share identical graph topologies. We estimate the aggregated significance of each graph independently using Equation 4, resulting in an " + }, + { + "bbox": [ + 130, + 408, + 481, + 539 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 408, + 481, + 539 + ], + "type": "text", + "content": "-dimensional importance vector " + }, + { + "bbox": [ + 130, + 408, + 481, + 539 + ], + "type": "inline_equation", + "content": "[I^*(G_1), \\ldots, I^*(G_N)]" + }, + { + "bbox": [ + 130, + 408, + 481, + 539 + ], + "type": "text", + "content": ". This facilitates a straightforward ranking of the importance vector, followed by the elimination of the " + }, + { + "bbox": [ + 130, + 408, + 481, + 539 + ], + "type": "inline_equation", + "content": "\\mathfrak{p}\\%" + }, + { + "bbox": [ + 130, + 408, + 481, + 539 + ], + "type": "text", + "content": " unimportant substructures based on their scores. This ranking method is deemed reliable since all substructures within a group possess the same architectural design, parameter scale, and computational graph. A similar principle can be uniformly applied to other isomorphic groups for pruning, as detailed in Algorithm 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 557, + 230, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 230, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 230, + 571 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": "This section reports our empirical results on the ImageNet-1K dataset [8], which comprises 1,281,167 images for training and 50,000 for validation. We deploy the proposed Isomorphic Pruning to compress several vision models, including ConvNext [33], ResNet [36], MobileNetv2 [44], and Vision Transformers [10, 50]. We report the parameter amount (#Params), MACs, and Top-1 validation accuracy in Table 1 and 2, and report the actual latency and peak memory consumption in Table 3." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "G. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 290, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 290, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 290, + 128 + ], + "type": "text", + "content": "4.1 Settings for ImageNet-1K" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 281 + ], + "type": "text", + "content": "Pruning. For simplicity, we adopt the Taylor-based pruning framework [39], which applies a data-driven Taylor expansion to the loss function. For this purpose, we randomly sampled 100 mini-batches from the ImageNet dataset, each containing 64 images. These images were resized to " + }, + { + "bbox": [ + 130, + 136, + 482, + 281 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 136, + 482, + 281 + ], + "type": "text", + "content": " pixels and then center-cropped to " + }, + { + "bbox": [ + 130, + 136, + 482, + 281 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 136, + 482, + 281 + ], + "type": "text", + "content": " pixels. We normalize all images using the mean and standard deviation specific to ImageNet. Notably, while some methods, such as ConvNext such as ConvNext [33] and DeiT [50] employ additional training techniques like label smoothing and knowledge distillation, our approach only leverages a straightforward cross-entropy loss for the Taylor expansion. We accumulate gradients from all the batches for a more reliable importance estimation and apply Isomorphic pruning in a one-shot manner without iterative pruning and fine-tuning [40]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 298, + 482, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 466 + ], + "type": "text", + "content": "Finetuning. During the fine-tuning process of pruned models, we adhere to the original training protocol for a fair comparison. While maintaining most of the original settings unchanged, we adjusted the learning rates and batch sizes to make the training suitable for A5000 GPUs, following the linear LR scaling rule proposed in [13]. For example, to compressing a pre-trained DeiT models [50], we adopt the same augmentation strategies for training, including RandAugmentation [7], Mixup [67], CutMix [66], Random Erasing [69], Repeated Augmentation [23], label smoothing [49]. The pruned models are distilled with a pre-trained RegNetY [43, 50] for 300 epochs using the AdamW [34] optimizer, with a learning rate of 0.0005 and a total batch size of 2048. A weight decay of 0.05 and a cosine annealing scheduler is deployed for training. For other models like ConvNext [33], ResNet [19] and MobileNet-v2 [44], the same principle will be applied. More details of hyper-parameters, pruning, and finetuning can be found in the appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 483, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 593 + ], + "type": "text", + "content": "Evaluation. We report the optimal classification accuracy achieved on the ImageNet-1K validation set, employing the standard resize-and-crop protocol [19, 33] for evaluation. To further validate the efficiency of models, we profiled the latency and the peak memory consumption of the pruned models on both GPU (a single RTX A5000) and CPU. To guarantee a fair and standardized comparison, we also tested the latency of several baseline models sourced from the Pytorch-Image Models [57] and Torchvision [38], following the same testing procedure. The latency was estimated with a batch size of 256 for GPU and 8 for CPU, averaged on 100 repeats. Appropriate GPU warmup was deployed in our test." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 609, + 365, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 365, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 365, + 621 + ], + "type": "text", + "content": "4.2 Pruning Convolutional Neural Networks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": "Convolutional Neural Network is a critical focus in the area of network pruning. In this section, we concentrate on pruning three popular ConvNets: ConvNext [33], ResNet [19] and MobileNet-v2 [44]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Isomorphic Pruning for Vision Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 127, + 443, + 423 + ], + "blocks": [ + { + "bbox": [ + 219, + 114, + 394, + 125 + ], + "lines": [ + { + "bbox": [ + 219, + 114, + 394, + 125 + ], + "spans": [ + { + "bbox": [ + 219, + 114, + 394, + 125 + ], + "type": "text", + "content": "Table 1: Pruning CNNs on ImageNet-1K." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 127, + 443, + 423 + ], + "lines": [ + { + "bbox": [ + 167, + 127, + 443, + 423 + ], + "spans": [ + { + "bbox": [ + 167, + 127, + 443, + 423 + ], + "type": "table", + "html": "
Pruning Method#Params (M)MACs (G)Base Acc.Final Acc.Δ Acc.
ConvNext-B† [33]88.5915.39-83.83-
ConvNext-S† [33]50.228.71-83.14-
ConvNext-S (Ours)47.368.4883.8383.17-0.66
ConvNext-T† [33]28.594.47-82.06-
ConvNext-T (Ours)25.324.1983.8382.19-1.64
ResNet-152†60.1911.58-78.31-
AOFP-E2 [9]-4.1377.3777.00-0.37
ABCPruner [28]24.074.3178.3177.12-1.19
EPruner-0.63 [27]21.564.0878.3176.83-1.48
ResNet-152-4.0G (Ours)23.114.0578.3177.84-0.47
ResNet-101†44.557.85-77.38-
IE [39]31.24.7077.3777.35-0.02
FPGM [21]-4.5177.3777.32-0.05
SFP [20]-4.5177.3777.51+0.14
ISTA [59]-4.4776.4075.27-1.13
Res101-4.5G (Ours)29.144.4877.3877.56+0.16
GFP [30]28.023.9078.2978.33+0.04
AOFP [9]-3.8976.6376.40-0.23
Res101-3.8G (Ours)24.873.8577.3877.43+0.05
ResNet-50† [19]25.564.13-76.13-
Taylor [39]14.202.2576.1874.50-1.68
CCP [41]-2.1176.1575.50-0.65
GFP [30]19.422.0476.7976.42-0.37
AutoSlim [63]20.602.0076.1075.60-0.50
DepGraph [11]-1.9976.1575.83-0.32
ResNet50-2G (Ours)15.052.0676.1375.91-0.22
Mobv2†3.500.32-71.88-
Meta [32]-0.1474.7068.20-6.50
GFP [30]-0.1575.7469.16-6.58
DepGraph [11]-0.1571.9068.49-3.41
Mobv2-0.15G (Ours)2.190.1571.8868.91-2.97
", + "image_path": "184a1b705ee0abc35115746d98b63b430510eda61563aa6a5cdd5123f86ae588.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 434, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 625 + ], + "type": "text", + "content": "ConvNext. Our experiments in Table 1 begin with the evaluation on an advanced CNN, ConvNext [33], which achieves impressive performance on several benchmarks. We use the official ConvNext-Base [38] as the pre-trained model for pruning, which achieves an accuracy of " + }, + { + "bbox": [ + 130, + 434, + 482, + 625 + ], + "type": "inline_equation", + "content": "83.83\\%" + }, + { + "bbox": [ + 130, + 434, + 482, + 625 + ], + "type": "text", + "content": " on the ImageNet-1K validation set. The pruning of ConvNext models is a relatively underexplored topic. Hence, we establish our baselines using the official variants with fewer parameters, ConvNext-S and ConvNext-T. For ConvNext, we also generalize our method to depth pruning, which removes unimportant layers with the minimal average scores on substructures, so that the pruned models have the same depth as baselines. The pruned models are fine-tuned for 300 epochs following the same training protocol as [33]. During fine-tuning, we observed that the pruned ConvNext with half of the parameters removed, achieves an accuracy of 80.09 in only 20 epochs, which effectively validates the efficiency of pruning. Moreover, as illustrated in Table 1, the final model after 300 full fine-tuning, achieves competitive performance (82.19% vs. 82.06%) to pre-trained models with fewer parameters (25.32M vs. 28.59M) and MACs (4.19G vs. 4.47G)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "text", + "content": "ResNet. In line with prior studies [5, 11, 22, 30], we demonstrate the efficacy of our method by structurally pruning pre-trained ResNet MACs. ResNet-" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "G. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 127, + 475, + 292 + ], + "blocks": [ + { + "bbox": [ + 219, + 114, + 394, + 125 + ], + "lines": [ + { + "bbox": [ + 219, + 114, + 394, + 125 + ], + "spans": [ + { + "bbox": [ + 219, + 114, + 394, + 125 + ], + "type": "text", + "content": "Table 2: Pruning DeiTs on ImageNet-1K." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 127, + 475, + 292 + ], + "lines": [ + { + "bbox": [ + 140, + 127, + 475, + 292 + ], + "spans": [ + { + "bbox": [ + 140, + 127, + 475, + 292 + ], + "type": "table", + "html": "
Method#Params (M) MACs (G) Acc (%)Method#Params (M) MACs (G) Acc (%)
DeiT-B† [50]87.3417.6983.32DeiT-T† [50]5.911.2774.52
DeiT-S† [50]22.444.6481.17DeiT-T-600EP [58]5.911.2775.00
DeiT-S-600EP [58]22.444.6481.80WDPruning [61]13.302.6078.38
VTP - 40% [70]48.0010.0080.70X-Pruner [64]-2.4078.93
WDPruning [61]55.309.9080.76UVC [65]-2.3278.82
UVC [65]-8.0080.57ViT-Slim [70]11.402.3077.94
X-Pruner [64]-8.5081.02ViT-Slim [70]11.402.3077.94
NViT-B [58]34.006.8083.29NViT-T [58]6.901.3076.21
SAViT [68]25.405.3081.66SAViT+KD [68]6.601.3076.95
UP-DeiT-S [62]22.10-81.56UP-DeiT-T [62]5.70-75.79
NViT-S [58]21.004.2082.19DeiT-T (Ours)5.741.2177.50
Upop [45]19.904.1081.10VTC-LFC [55]5.10-71.60
DeiT-S (Ours)20.694.1682.41SViT-E [68]4.210.9970.12
GOHSP [60]14.402.9979.98GOHSP [60]4.000.9170.24
GOHSP [60]11.102.8179.86SAViT [68]4.200.9070.72
ViT-Slim [2]13.502.8079.20CP-ViT [47]-0.7471.24
Upop [45]13.502.8079.60WDPruning [61]3.500.7070.34
ViT-Slim [2]13.502.8079.20X-Pruner [64]-0.6071.10
CP-ViT [47]-2.6679.08UVC [65]-0.5170.60
DeiT-2.6G (Ours)13.072.6281.13DeiT-0.6G (Ours)3.080.6272.60
", + "image_path": "11bd7db5add35241b1f1b835eab3ca82855034a77d9b9c36f78790a92b7ca135.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 312, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 469 + ], + "type": "text", + "content": "50/101/152 consists of four residual blocks, each formed by several Bottle-neck structures and residual connections [19]. The proposed Isomorphic Pruning identifies five kinds of distinct isomorphic groups within ResNet-50, four targeting the block structures and one associated with bottleneck structures [19]. It is worthwhile to note that different sub-structures vary in their parameter and MAC counts. For example, residual blocks naturally contain more parameters than the bottleneck structures which are only composed of a few convolutional layers. The conventional global pruning, employing a naive global ranking threshold, faces the potential risks of over-pruning certain layers due to the biased importance distribution. Isomorphic Pruning, however, employs a more stable scheme for pruning. As illustrated in Table 1, we report both the accuracy and the accuracy drop compared to their base models. Isomorphic pruning achieves lossless compression on ResNet-101, under both 4.5G and 3.8G settings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 485, + 483, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 483, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 483, + 536 + ], + "type": "text", + "content": "MobileNet-v2. We also conduct experiments on lightweight models, MobileNet-v2, by compressing its MACs from 0.32 G to 0.15 G. In MobileNet-v2, we can also discover 5 isomorphic groups in MobileNet. The pruned model achieves better accuracy compared to DepGraph [11] which requires sparse training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 550, + 310, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 550, + 310, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 550, + 310, + 562 + ], + "type": "text", + "content": "4.3 Pruning Vision Transformers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 569, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 481, + 641 + ], + "type": "text", + "content": "Vision Transformers (ViT) has profoundly reshaped the paradigm of visual modeling over the past few years. Nevertheless, their inherent over-parameterization often imposes huge overheads to both training and inference. Compared to CNNs that are composed of similar convolution layers or blocks, Vision Transformers are usually built with MLP, Multi-head Attention, and skip connections, which are more challenging for pruning due to the heterogeneous internal structures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Table 2 presents our pruning results on DeiT, a plain vision transformer enhanced with knowledge distillation [50]. We benchmark our approach with" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "type": "text", + "content": "Isomorphic Pruning for Vision Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 157, + 480, + 398 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: Profiling the pruned models and other off-the-shelf vision models from PyTorch-Image-Models [57] on a single RTX A5000 GPU and CPU. For NViT [58], we report their best speed up in the original paper, tested on a single V100 or RTX 3080." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 157, + 480, + 398 + ], + "lines": [ + { + "bbox": [ + 133, + 157, + 480, + 398 + ], + "spans": [ + { + "bbox": [ + 133, + 157, + 480, + 398 + ], + "type": "table", + "html": "
Method#Params ↓ MACs ↓ Peak Mem. ↓ GPU Latency (256) ↓ CPU Latency (8) ↓ Acc. (%)
DeiT-B [50]86.40 M17.60 G2909.78 MB802.97 ms (1.00×)197.73 ms (1.00×)83.32
DeiT-S [50]22.44 M4.64 G1363.47 MB241.81 ms (3.32×)80.64 ms (2.45×)81.17
Swin-T [31]28.29 M4.51 G4257.98 MB375.45 ms (2.14×)203.86 ms (0.97×)81.19
NViT-S [58]21.00 M4.20 G--(2.52×)-82.19
TNT-S [15]23.76 M4.85 G1457.56 MB631.02 ms (1.27×)175.58 ms (1.13×)81.50
CaiT-XS24 [51]26.56 M5.40 G1496.42 MB557.41 ms (1.44×)342.28 ms (0.58×)82.00
CrossViT-S [3]26.86 M5.64 G1546.94 MB315.02 ms (2.55×)115.51 ms (1.71×)81.00
EfficientFormer-L3 [26]31.41 M3.94 G2681.75 MB249.05 ms (3.22×)167.16 ms (1.18×)82.40
PVTv2-B2 [54]25.36 M4.00 G4478.18 MB364.00 ms (2.21×)212.44 ms (0.93×)82.00
DeiT-S (Ours)20.69 M4.16 G1547.97 MB230.84 ms (3.48×)44.10 ms (4.48×)82.41
DeiT-2.6G (Ours)13.07 M2.62 G1203.07 MB189.68 ms (4.23×)56.59 ms (3.49×)81.13
DeiT-T [50]5.91 M1.27 G720.22 MB93.78 ms (8.56×)46.83 ms (4.22×)74.52
NViT-T [58]6.90 M1.30 G--(4.97×)-76.21
DeiT-T (Ours)5.71 M1.20 G837.16 MB93.17 ms (8.62×)47.89 ms (4.13×)77.50
DeiT-0.6G (Ours)3.08 M0.62 G629.28 MB63.23 ms (12.70×)40.98 ms (4.83×)72.60
ConvNext-B [33]88.59 M15.36 G5394.10 MB789.24 ms (1.00×)262.58 ms (1.00×)83.83
ConvNext-S [33]50.22 M8.69 G3958.45 MB508.20 ms (1.57×)188.07 ms (1.40×)83.14
ConvNext-S (Ours)47.36 M8.48 G3788.65 MB535.95 ms (1.59×)179.72 ms (1.46×)83.17
ConvNext-T [33]28.59 M4.46 G3785.37 MB291.02 ms (2.82×)126.17 ms (2.08×)82.06
ConvNext-T (Ours)25.32 M4.19 G3607.06 MB292.77 ms (2.83×)97.82 ms (2.68×)82.19
ResNet152 [19]60.19 M11.58 G3320.24 MB386.65 ms (1.00×)272.21 ms (1.00×)78.31
ResNet152-4G (Ours)23.11 M4.06 G1938.91 MB238.12 ms (1.62×)180.65 ms (1.51×)77.84
ResNet101 [19]44.55 M7.85 G3196.77 MB271.34 ms (1.00×)202.66 ms (1.00×)77.38
ResNet101-4.5G (Ours)29.14 M4.48 G2639.40 MB220.72 ms (1.23×)162.33 ms (1.25×)77.56
ResNet101-3.8G (Ours)24.87 M3.85 G2286.36 MB194.39 ms (1.40×)158.98 ms (1.28×)77.43
ResNet50 [19]25.56 M4.13 G3035.84 MB171.54 ms (1.00×)114.86 ms (1.00×)76.13
ResNet50-2.0G (Ours)15.05 M2.06 G2238.15 MB78.75 ms (1.33×)80.88 ms (1.42×)75.91
Mobv2 [44]3.50 M0.32 G2855.44 MB71.42 ms (1.00×)74.46 ms (1.00×)71.88
Mobv2-0.15G (Ours)2.19 M0.15 G1354.82 MB43.48 ms (1.64×)38.63 ms (1.93×)68.91
", + "image_path": "fd634b3a6df3020c23ea4e8bde3762b05ba0c4bd98672cfc59cfe361f27d75d0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": "a range of innovative pruning baselines specifically tailored for vision transformers [45, 47, 50, 55, 58, 60, 61, 64, 65, 68, 70]. A DeiT model consists of various components, such as MLP and attention layers. Within the attention layers, we can even find two feasible pruning schemes: pruning 1) the number of heads or 2) within-head dimensions. We implement Isomorphic Pruning to compress a pre-trained DeiT-Base model by pruning both heads, head dimensions, and the embedding sizes following [58]. This is achieved through a straightforward first-order Taylor expansion for importance estimation, followed by a grouped ranking in Isomorphic Pruning. The details of pruning vision transformers can be found in the appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "We developed four lightweight vision transformers: DeiT-S, DeiT-2.6G, DeiT-T and DeiT-0.6G. The pruned DeiT-S and DeiT-T models achieve comparable latencies comparable to the standard DeiT-S and DeiT-T with manually designed architecture. The empirical results are reported in Table 2, which shows that our method yields greater performance gains in complex models with extensive heterogeneous internal structures, especially in comparison to CNNs. Specifically, the optimized DeiT-T achieves an accuracy of " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "77.50\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ", with only 1.2G MACs, which is better than the results " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "(74.52\\%)" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " of the pre-trained DeiT-T with uniform width across layers. Detailed configurations, such as base models and pruning ratios, are available in the appendix." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "G. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 115, + 233, + 243 + ], + "blocks": [ + { + "bbox": [ + 149, + 115, + 233, + 243 + ], + "lines": [ + { + "bbox": [ + 149, + 115, + 233, + 243 + ], + "spans": [ + { + "bbox": [ + 149, + 115, + 233, + 243 + ], + "type": "image", + "image_path": "e3600dd6856369561d35f1101ff00315d23351fc6c524ef5700d3bbb6544dff2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 247, + 197, + 256 + ], + "lines": [ + { + "bbox": [ + 186, + 247, + 197, + 256 + ], + "spans": [ + { + "bbox": [ + 186, + 247, + 197, + 256 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 238, + 115, + 323, + 243 + ], + "blocks": [ + { + "bbox": [ + 238, + 115, + 323, + 243 + ], + "lines": [ + { + "bbox": [ + 238, + 115, + 323, + 243 + ], + "spans": [ + { + "bbox": [ + 238, + 115, + 323, + 243 + ], + "type": "image", + "image_path": "9ea2c9c087d773e289c1bc67b427415f2b26b09f77e1206d308f314c1b97a7e9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 275, + 247, + 287, + 255 + ], + "lines": [ + { + "bbox": [ + 275, + 247, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 275, + 247, + 287, + 255 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 332, + 115, + 461, + 244 + ], + "blocks": [ + { + "bbox": [ + 332, + 115, + 461, + 244 + ], + "lines": [ + { + "bbox": [ + 332, + 115, + 461, + 244 + ], + "spans": [ + { + "bbox": [ + 332, + 115, + 461, + 244 + ], + "type": "image", + "image_path": "0279396134af7b62b84d0b38091ee89e433508089848e776920f76d687e0c1dc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 247, + 402, + 256 + ], + "lines": [ + { + "bbox": [ + 391, + 247, + 402, + 256 + ], + "spans": [ + { + "bbox": [ + 391, + 247, + 402, + 256 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 265, + 480, + 342 + ], + "lines": [ + { + "bbox": [ + 130, + 265, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 480, + 342 + ], + "type": "text", + "content": "Fig. 4: (a-b) Top-1 Accuracy and loss of pruning ResNet-50 on ImageNet-1K validation set without finetuning. We report the mean and standard derivation of 10 experiments. IsomorphicP Pruning consistently improves the performance of pruned models. (c) Histogram of Taylor importance score of DeiT sub-structures. Isomorphic structures are highlighted with the same color. Zoom in for more details about each structure. We visualize the threshold of " + }, + { + "bbox": [ + 130, + 265, + 480, + 342 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 265, + 480, + 342 + ], + "type": "text", + "content": " pruning ratio for naive global pruning (the black dash) and isomorphic pruning (the colored dash)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 365, + 320, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 320, + 377 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 320, + 377 + ], + "type": "text", + "content": "4.4 Analysis on the Pruned Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 384, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 624 + ], + "type": "text", + "content": "On-device Latency Table 3 measures the actual latency and peak memory consumptions on a single RTX A5000 GPU and CPU. We evaluate all models with a batch size of 256 for GPU and 8 for CPU, respectively. The average latency over 100-round experiments are reported, with a 20-step warmup process before evaluation. Firstly, results show that the pruned DeiT models can achieve comparable performance (82.41% vs. 82.40%) to some advanced transformers like EfficientFormer [26], while preserving a competitive latency. Additionally, we observe that the peak memory consumption is primarily determined by the largest layer width. Therefore, under the same budget, a uniform DeiT may require less peak memory (1363 MB for DeiT-S) for inference. In contrast, a non-uniform architecture may require slightly more memory (1547 MB for the DeiT-S Pruned) under the given budget. But non-uniform pruning can achieve better accuracy, which reveals an important trade-off between accuracy and memory consumption for transformers. In addition, for CNNs like ResNets, the pruned model requires much less memory due to the compressed feature dimensions. Furthermore, we note that the actual speed-up may not be linearly correlated with the MAC reductions. This is particularly evident in ResNet-50, where only a " + }, + { + "bbox": [ + 130, + 384, + 482, + 624 + ], + "type": "inline_equation", + "content": "1.33 \\times" + }, + { + "bbox": [ + 130, + 384, + 482, + 624 + ], + "type": "text", + "content": " actual speed-up is achieved despite the removal of half of the MACs. In contrast, compression on Mobv2 is more effective, achieving a " + }, + { + "bbox": [ + 130, + 384, + 482, + 624 + ], + "type": "inline_equation", + "content": "1.64 \\times" + }, + { + "bbox": [ + 130, + 384, + 482, + 624 + ], + "type": "text", + "content": " acceleration with a similar compression ratio on MACs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Distribution of Importance Scores. To further explore how Isomorphic Pruning affects the pruning behavior, we study the distribution of importance" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "type": "text", + "content": "Isomorphic Pruning for Vision Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "score in a DeiT-Base model [50]. As shown in Figure 4 (c), the importance histogram of different sub-structures is highlighted with different colors, each corresponding to one kind of isomorphic structure. It is natural to find that, the MLP layer and the attention layer have diverged importance distributions due to their different computation process. Besides, when we apply a simple " + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": " global pruning on the DeiT model, most parameters in Attentions " + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "inline_equation", + "content": "\\# 5" + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "inline_equation", + "content": "\\# 9" + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "inline_equation", + "content": "\\# 13" + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": " will be removed, since the importance scores in attentions are centered around 0. In contrast, the embedding number as " + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "inline_equation", + "content": "\\# 1" + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": " will not be even been pruned in this case, due to the relatively large importance scores. This inevitably leads to an imbalanced pruning of different structures. Notably, the embedding structure, as illustrated in the appendix, contains massive parameters in the DeiT, where biased pruning may lead to very limited acceleration." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 278, + 482, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 387 + ], + "type": "text", + "content": "Importance Criteria. Note that Isomorphic Pruning can be combined with different pruning criteria like Taylor [39], Hessian [25] and even L1-norm [22]. We quantify the performance gain of Isomorphic Pruning by visualizing the accuracy and loss curve with different pruning ratios in Figure 4 (a) and (b). Specifically, we gradually increase the pruning ratio for each criterion and report the average accuracy and loss without finetuning over 10 repeated runs with random inputs for Taylor expansion [25, 39]. Note that L1 pruning is deterministic and thus presents 0 std. We find the Isomorphic Pruning, marked as solid lines, consistently improves the counterpart baselines highlighted in dashed lines." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 405, + 334, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 334, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 334, + 536 + ], + "type": "text", + "content": "Transfer Learning. We also evaluate the performance of transfer learning based on our pruned models. Table 4 reports the accuracy of transferring different models to CIFAR-10 and CIFAR-100. We resize all images to " + }, + { + "bbox": [ + 130, + 405, + 334, + 536 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 405, + 334, + 536 + ], + "type": "text", + "content": " and apply the same training recipe for transfer. For the DeiT models, we did not include the supervision from the teacher models. It is observed that our method achieves better transfer accuracy compared to the scratch training model and baseline methods." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 342, + 439, + 479, + 524 + ], + "blocks": [ + { + "bbox": [ + 339, + 417, + 481, + 437 + ], + "lines": [ + { + "bbox": [ + 339, + 417, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 339, + 417, + 481, + 437 + ], + "type": "text", + "content": "Table 4: Transfer learning on pruned models." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 342, + 439, + 479, + 524 + ], + "lines": [ + { + "bbox": [ + 342, + 439, + 479, + 524 + ], + "spans": [ + { + "bbox": [ + 342, + 439, + 479, + 524 + ], + "type": "table", + "html": "
ModelsCIFAR-10CIFAR-100
DeiT-S [39]98.5287.07
NViT-S [58]98.7887.90
DeiT-S (Ours)99.0290.09
DeiT-T93.9385.66
NViT-T [58]98.3185.88
DeiT-S (Ours)98.2887.08
ConvNext-S98.7789.15
ConvNext-S (Ours)98.7889.36
", + "image_path": "c6c68c098f78a8a275c90a2aa105c087ba93255d6422ca1cdaadb09b0745b5a4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 555, + 220, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 220, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 220, + 568 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduce Isomorphic Pruning, a practical approach designed to compress vision networks with novel mechanisms and architectures. This method focuses on mitigating the challenges posed by heterogeneous sub-structures in networks, thereby enhancing the reliability of ranking and pruning processes. Our empirical results on the ImageNet-1k dataset, showcase the effectiveness of Isomorphic Pruning across several vision models like ConvNext, ResNet, MobileNetv2, and Vision Transformers." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "G. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 199 + ], + "type": "text", + "content": "This project is supported by the National Research Foundation Singapore under its Medium Sized Center for Advanced Robotics Technology Innovation, and the Advanced Research and Technology Innovation Centre (ARTIC), the National University of Singapore under Grant (project number: A-0005947-21-00, project reference: ECT-RP2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 217, + 197, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 197, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 197, + 228 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 239, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 239, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 262 + ], + "type": "text", + "content": "1. Baum, E., Haussler, D.: What size net gives valid generalization? Advances in neural information processing systems 1 (1988)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 262, + 480, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 480, + 305 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 480, + 305 + ], + "type": "text", + "content": "2. Chavan, A., Shen, Z., Liu, Z., Liu, Z., Cheng, K.T., Xing, E.P.: Vision transformer slimming: Multi-dimension searching in continuous optimization space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4931-4941 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 307, + 480, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 307, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 138, + 307, + 480, + 338 + ], + "type": "text", + "content": "3. Chen, C.F.R., Fan, Q., Panda, R.: Crossvit: Cross-attention multi-scale vision transformer for image classification. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 357-366 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 339, + 480, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 339, + 480, + 371 + ], + "spans": [ + { + "bbox": [ + 138, + 339, + 480, + 371 + ], + "type": "text", + "content": "4. Chen, T., Cheng, Y., Gan, Z., Yuan, L., Zhang, L., Wang, Z.: Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems 34, 19974-19988 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 372, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 372, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 372, + 480, + 392 + ], + "type": "text", + "content": "5. Chen, T., Liang, L., Ding, T., Zhu, Z., Zharkov, I.: Otov2: Automatic, generic, user-friendly. arXiv preprint arXiv:2303.06862 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 393, + 480, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 393, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 138, + 393, + 480, + 425 + ], + "type": "text", + "content": "6. Chollet, F.: Xception: Deep learning with depthwise separable convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1251-1258 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 426, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 480, + 468 + ], + "type": "text", + "content": "7. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 702-703 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 470, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 470, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 470, + 480, + 502 + ], + "type": "text", + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 502, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 534 + ], + "type": "text", + "content": "9. Ding, X., Ding, G., Guo, Y., Han, J., Yan, C.: Approximated oracle filter pruning for destructive cnn width optimization. In: International Conference on Machine Learning. pp. 1607-1616. PMLR (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 535, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 535, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 134, + 535, + 480, + 578 + ], + "type": "text", + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 578, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 578, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 134, + 578, + 480, + 611 + ], + "type": "text", + "content": "11. Fang, G., Ma, X., Song, M., Mi, M.B., Wang, X.: Depgraph: Towards any structural pruning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16091-16101 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "type": "text", + "content": "12. Frantar, E., Alistarh, D.: Sparsegpt: Massive language models can be accurately pruned in one-shot (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "13. Goyal, P., Dolkar, P., Girshick, R., Noordhuis, P., Wesolowski, L., Kyrola, A., Tulloch, A., Jia, Y., He, K.: Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677 (2017)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "type": "text", + "content": "Isomorphic Pruning for Vision Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "14. Guo, J., Ouyang, W., Xu, D.: Multi-dimensional pruning: A unified framework for model compression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1508-1517 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 172 + ], + "type": "text", + "content": "15. Han, K., Xiao, A., Wu, E., Guo, J., Xu, C., Wang, Y.: Transformer in transformer. Advances in Neural Information Processing Systems 34, 15908-15919 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "text", + "content": "16. Han, S., Mao, H., Dally, W.J.: Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. International Conference on Learning Representations (ICLR) (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "17. Hassibi, B., Stork, D.G., Wolff, G.J.: Optimal brain surgeon and general network pruning. In: IEEE international conference on neural networks. pp. 293-299. IEEE (1993)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "18. He, H., Liu, J., Pan, Z., Cai, J., Zhang, J., Tao, D., Zhuang, B.: Pruning self-attention into convolutional layers in single path. arXiv preprint arXiv:2111.11802 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 303 + ], + "type": "text", + "content": "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 303, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 481, + 325 + ], + "type": "text", + "content": "20. He, Y., Kang, G., Dong, X., Fu, Y., Yang, Y.: Soft filter pruning for accelerating deep convolutional neural networks. arXiv preprint arXiv:1808.06866 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "text", + "content": "21. He, Y., Liu, P., Wang, Z., Hu, Z., Yang, Y.: Filter pruning via geometric median for deep convolutional neural networks acceleration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4340-4349 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "text", + "content": "22. He, Y., Zhang, X., Sun, J.: Channel pruning for accelerating very deep neural networks. In: Proceedings of the IEEE international conference on computer vision. pp. 1389-1397 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 434 + ], + "type": "text", + "content": "23. Hoffer, E., Ben-Nun, T., Hubara, I., Giladi, N., Hoefler, T., Soudry, D.: Augment your batch: Improving generalization through instance repetition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8129-8138 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 434, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 481, + 468 + ], + "type": "text", + "content": "24. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "text", + "content": "25. LeCun, Y., Denker, J., Solla, S.: Optimal brain damage. Advances in neural information processing systems 2 (1989)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "text", + "content": "26. Li, Y., Yuan, G., Wen, Y., Hu, J., Evangelidis, G., Tulyakov, S., Wang, Y., Ren, J.: Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems 35, 12934-12949 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "type": "text", + "content": "27. Lin, M., Ji, R., Li, S., Wang, Y., Wu, Y., Huang, F., Ye, Q.: Network pruning using adaptive exemplar filters. IEEE Transactions on Neural Networks and Learning Systems 33(12), 7357-7366 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "type": "text", + "content": "28. Lin, M., Ji, R., Zhang, Y., Zhang, B., Wu, Y., Tian, Y.: Channel pruning via automatic structure search. arXiv preprint arXiv:2001.08565 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "text", + "content": "29. Liu, B., Wang, M., Foroosh, H., Tappen, M., Pensky, M.: Sparse convolutional neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 806-814 (2015)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "30. Liu, L., Zhang, S., Kuang, Z., Zhou, A., Xue, J.H., Wang, X., Chen, Y., Yang, W., Liao, Q., Zhang, W.: Group fisher pruning for practical network compression. In: International Conference on Machine Learning. pp. 7021-7032. PMLR (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "31. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "G. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "text", + "content": "of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "type": "text", + "content": "32. Liu, Z., Mu, H., Zhang, X., Guo, Z., Yang, X., Cheng, K.T., Sun, J.: Metapruning: Meta learning for automatic neural network channel pruning. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3296-3305 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "text", + "content": "33. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11976-11986 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "34. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "35. Louizos, C., Welling, M., Kingma, D.P.: Learning sparse neural networks through " + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "inline_equation", + "content": "l\\_0" + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": " regularization. arXiv preprint arXiv:1712.01312 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 293 + ], + "type": "text", + "content": "36. Ma, X., Yuan, G., Lin, S., Li, Z., Sun, H., Wang, Y.: Resnet can be pruned " + }, + { + "bbox": [ + 132, + 249, + 481, + 293 + ], + "type": "inline_equation", + "content": "60 \\times" + }, + { + "bbox": [ + 132, + 249, + 481, + 293 + ], + "type": "text", + "content": ": Introducing network purification and unused path removal (p-rm) after weight pruning. In: 2019 IEEE/ACM International Symposium on Nanoscale Architectures (NANOARCH). pp. 1-2. IEEE (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "text", + "content": "37. Ma, X., Fang, G., Wang, X.: Llm-pruner: On the structural pruning of large language models. In: Advances in Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "text", + "content": "38. maintainers, T., contributors: Torchvision: Pytorch's computer vision library. GitHub repository (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "39. Molchanov, P., Mallya, A., Tyree, S., Frosio, I., Kautz, J.: Importance estimation for neural network pruning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11264-11272 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 370, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 402 + ], + "type": "text", + "content": "40. Molchanov, P., Tyree, S., Karras, T., Aila, T., Kautz, J.: Pruning convolutional neural networks for resource efficient inference. arXiv preprint arXiv:1611.06440 (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "41. Peng, H., Wu, J., Chen, S., Huang, J.: Collaborative channel pruning for deep networks. In: International Conference on Machine Learning. pp. 5113-5122. PMLR (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 479 + ], + "type": "text", + "content": "42. Pool, J., Sawarkar, A., Rodge, J.: Accelerating inference with sparsity using the nvidia ampere architecture and nvidia tensorsr. NVIDIA Developer Technical Blog, https://developer.nvidia.com/blog/accelerating-inference-with-sparsityusing-ampere-and-tensorr (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "43. Radosavovic, I., Kosaraju, R.P., Girshick, R., He, K., Dollár, P.: Designing network design spaces. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10428-10436 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "44. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: Mobilenetv2: Inverted residuals and linear bottlenecks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4510-4520 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "text", + "content": "45. Shi, D., Tao, C., Jin, Y., Yang, Z., Yuan, C., Wang, J.: UPop: Unified and progressive pruning for compressing vision-language transformers. In: Proceedings of the 40th International Conference on Machine Learning. vol. 202, pp. 31292-31311. PMLR (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "text", + "content": "46. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "47. Song, Z., Xu, Y., He, Z., Jiang, L., Jing, N., Liang, X.: Cp-vit: Cascade vision transformer pruning via progressive sparsity prediction. arXiv preprint arXiv:2203.04570 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "text", + "content": "48. Sun, M., Liu, Z., Bair, A., Kolter, J.Z.: A simple and effective pruning approach for large language models. arXiv preprint arXiv:2306.11695 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 291, + 91, + 447, + 102 + ], + "type": "text", + "content": "Isomorphic Pruning for Vision Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "49. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818-2826 (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "text", + "content": "50. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jegou, H.: Training data-efficient image transformers & distillation through attention. In: International Conference on Machine Learning. vol. 139, pp. 10347-10357 (July 2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "type": "text", + "content": "51. Touvron, H., Cord, M., Sablayrolles, A., Synnaeve, G., Jégou, H.: Going deeper with image transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 32-42 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "type": "text", + "content": "52. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 246, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 278 + ], + "type": "text", + "content": "53. Voita, E., Talbot, D., Moiseev, F., Sennrich, R., Titov, I.: Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. arXiv preprint arXiv:1905.09418 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 278, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 311 + ], + "type": "text", + "content": "54. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media 8(3), 415-424 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 311, + 480, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 343 + ], + "type": "text", + "content": "55. Wang, Z., Luo, H., Wang, P., Ding, F., Wang, F., Li, H.: Vtc-lfc: Vision transformer compression with low-frequency components. Advances in Neural Information Processing Systems 35, 13974-13988 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 343, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 365 + ], + "type": "text", + "content": "56. Wen, W., Wu, C., Wang, Y., Chen, Y., Li, H.: Learning structured sparsity in deep neural networks. Advances in neural information processing systems 29 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "text", + "content": "57. Wightman, R.: Pytorch-image-models. https://github.com/huggingface/pytorch-image-models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "58. Yang, H., Yin, H., Shen, M., Molchanov, P., Li, H., Kautz, J.: Global vision transformer pruning with hessian-aware saliency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18547-18557 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "type": "text", + "content": "59. Ye, J., Lu, X., Lin, Z., Wang, J.Z.: Rethinking the smaller-norm-less-informative assumption in channel pruning of convolution layers. arXiv preprint arXiv:1802.00124 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "type": "text", + "content": "60. Yin, M., Uzkent, B., Shen, Y., Jin, H., Yuan, B.: Gohsp: A unified framework of graph and optimization-based heterogeneous structured pruning for vision transformer. arXiv preprint arXiv:2301.05345 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 483, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 515 + ], + "type": "text", + "content": "61. Yu, F., Huang, K., Wang, M., Cheng, Y., Chu, W., Cui, L.: Width & depth pruning for vision transformers. In: AAAI Conference on Artificial Intelligence (AAAI). vol. 2022 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 515, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 480, + 536 + ], + "type": "text", + "content": "62. Yu, H., Wu, J.: A unified pruning framework for vision transformers. Science China Information Sciences 66(7), 1-2 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "text", + "content": "63. Yu, J., Huang, T.: Autoslim: Towards one-shot architecture search for channel numbers. arXiv preprint arXiv:1903.11728 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 558, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 480, + 590 + ], + "type": "text", + "content": "64. Yu, L., Xiang, W.: X-pruner: explainable pruning for vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24355-24363 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "text", + "content": "65. Yu, S., Chen, T., Shen, J., Yuan, H., Tan, J., Yang, S., Liu, J., Wang, Z.: Unified visual transformer compression. arXiv preprint arXiv:2203.08243 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "66. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6023-6032 (2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "67. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "G. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 205 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "68. Zheng, C., Zhang, K., Yang, Z., Tan, W., Xiao, J., Ren, Y., Pu, S., et al.: Savit: Structure-aware vision transformer pruning via collaborative optimization. Advances in Neural Information Processing Systems 35, 9010-9023 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "type": "text", + "content": "69. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 13001-13008 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 482, + 205 + ], + "type": "text", + "content": "70. Zhu, M., Tang, Y., Han, K.: Vision transformer pruning. arXiv preprint arXiv:2104.08500 (2021)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Isomorphic Pruning for Vision Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ItTakesTwo_ Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation/69e10020-97d7-4fa9-96f3-57b936f4e871_content_list.json b/2024/ItTakesTwo_ Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation/69e10020-97d7-4fa9-96f3-57b936f4e871_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f986ad1240cc1451fa009fb937380591a1991a9e --- /dev/null +++ b/2024/ItTakesTwo_ Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation/69e10020-97d7-4fa9-96f3-57b936f4e871_content_list.json @@ -0,0 +1,1812 @@ +[ + { + "type": "text", + "text": "ItTakesTwo: Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation", + "text_level": 1, + "bbox": [ + 84, + 48, + 910, + 107 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuyuan Liu\\*1(), Yuanhong Chen\\*1, Hu Wang2,1, Vasileios Belagiannis3, Ian Reid2,1, and Gustavo Carneiro4,1", + "bbox": [ + 104, + 138, + 891, + 180 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Australian Institute for Machine Learning (AIML), University of Adelaide, AU", + "bbox": [ + 96, + 193, + 895, + 213 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), UAE", + "bbox": [ + 123, + 213, + 871, + 230 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 University of Erlangen-Nuremberg, Germany", + "bbox": [ + 263, + 230, + 731, + 248 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Centre for Vision, Speech and Signal Processing (CVSSP), University of Surrey, UK", + "bbox": [ + 75, + 248, + 917, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The costly and time-consuming annotation process to produce large training sets for modelling semantic LiDAR segmentation methods has motivated the development of semi-supervised learning (SSL) methods. However, such SSL approaches often concentrate on employing consistency learning only for individual LiDAR representations. This narrow focus results in limited perturbations that generally fail to enable effective consistency learning. Additionally, these SSL approaches employ contrastive learning based on the sampling from a limited set of positive and negative embedding samples. This paper introduces a novel semisupervised LiDAR semantic segmentation framework called ItTakesTwo (IT2). IT2 is designed to ensure consistent predictions from peer LiDAR representations, thereby improving the perturbation effectiveness in consistency learning. Furthermore, our contrastive learning employs informative samples drawn from a distribution of positive and negative embeddings learned from the entire training set. Results on public benchmarks show that our approach achieves remarkable improvements over the previous state-of-the-art (SOTA) methods in the field. The code is available at: https://github.com/yyliu01/IT2.", + "bbox": [ + 142, + 302, + 857, + 624 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Semi-supervised Learning $\\cdot$ LiDAR Semantic Segmentation", + "bbox": [ + 142, + 640, + 849, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 75, + 687, + 309, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Outdoor LiDAR semantic segmentation is emerging as an important area of research given its application to self-driving vehicles [16, 17]. Although LIDAR data has advantages relative to photometric camera data, it also presents unique challenges, such as data sparsity [66] and varying point density [27, 45]. Most existing works alleviate these issues by transforming the LiDAR point-clouds to other representations, including range-based image [13, 26, 45, 55], voxel grids [15, 21, 66], or fusing them with other modalities [38, 41] and views [58]. Despite their success, those methods rely on supervised learning that requires large labelled training sets, which can be problematic given the costly and time-consuming labelling process [39, 47].", + "bbox": [ + 72, + 723, + 920, + 918 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* denotes equal contribution. (☑) Email: yuyuan.liu@adelaide.edu.au.", + "bbox": [ + 99, + 928, + 820, + 947 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a9315ad125de3c5b132dd037dbbdb275a82c73d44310d0125031ce5043815f67.jpg", + "image_caption": [ + "Fig. 1: The left figure represents the 3D point cloud scan (ground truth of four classes - Car, Pole, Terrain, and Sidewalk) while the right figures denote the prediction results (into the same four classes), provided by our model, from range (top-right) and voxel (bottom-right) representations, where the incorrect predictions are marked in gray. Please notice that: 1) the mapping of the four classes from the 3D point cloud scan to each representation demonstrates the consistent distribution of semantic classes, and 2) the histograms of the predictions for each representation exhibit distinct differences." + ], + "image_footnote": [], + "bbox": [ + 77, + 50, + 922, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These issues have motivated semi-supervised learning (SSL) innovations for LiDAR semantic segmentation [23,27,31], which require a small labelled dataset used in conjunction with a large unlabelled dataset. Consistency Learning, one of the main strategies explored in SSL for image semantic segmentation, penalises inconsistent predictions from perturbed inputs, where these perturbations take various forms, including input augmentation [62, 69], adversarial noise [22, 39], network perturbations [6] and dropout [61]. The incorporation of strong and varied perturbations is crucial for enhancing generalisation [14, 39].", + "bbox": [ + 72, + 430, + 917, + 586 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current research [27,31] applies such consistency learning in SSL for LiDAR points using many strategies, such as laser-beam augmentation [27] or Mean Teacher (MT) [31] framework. Despite their success, these approaches solely rely on a single LiDAR representation (e.g., voxel grid) for training, which could potentially limit the model's ability to generalize effectively due to the restricted perturbations. Such an issue can be mitigated with the use of multiple LiDAR representations for consistency learning that explores the fundamental Clustering assumption5, where the results from different representations naturally indicate the same underlying semantic information from the 3D LiDAR points. For example, as illustrated in Fig. 1, the class Terrain (▲) in the top-right pixels and bottom-right voxels should be \"clustered\" according to the (left) original 3D point data. Thus, the peer LiDAR representations can be regarded as a novel form similar to perturbation, but yet under-explored in the field. Other disadvantages of focusing on only one of the representations when using small labelled datasets are: the range view suffers from the information loss during the projection process, particularly for the tail categories, such as Pole (▲), which has", + "bbox": [ + 72, + 586, + 920, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 77, + 16, + 94, + 30 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 157, + 14, + 280, + 30 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "5 Cluster hypothesis (or assumption) [42]: the data of various forms that behave similarly with respect to information relevance should be clustered together.", + "bbox": [ + 79, + 910, + 917, + 947 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a low mIoU of 52.02; and the voxel grid tends to show lower accuracy in dense areas that are far away from the LiDAR source, such as Sidewalk ( $\\lambda$ ), which has a low mIoU of 69.50. Such limited pseudo label results will inevitably introduce confirmation bias (i.e., the problem of relying on and reinforcing wrong predictions during training), reducing even further the efficacy of consistency learning based on single representation [23, 27, 31].", + "bbox": [ + 77, + 55, + 917, + 175 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another effective technique for semi-supervised LiDAR segmentation is contrastive learning [7,20,23,31], which consists of attracting positive and repelling negative embedding pairs at instance [46] or category [23,31] levels, thereby enhancing the latent space point distribution. However, to generate these pairs, methods typically rely on a limited number of embeddings extracted from randomly selected training samples due to the high computational cost associated with dense contrastive learning tasks [23,31]. Unfortunately, this limited sample size hinders a comprehensive understanding of the distribution across the entire embedding space [5], resulting in suboptimal convergence of LiDAR segmentation in SSL. This limitation becomes particularly pronounced when dealing with embeddings from different models in multiple views, where potential noisy data [44, 51] or model bias [59] from one network can adversely affect others, thereby impeding the optimization process.", + "bbox": [ + 77, + 178, + 917, + 427 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we propose a novel semi-supervised LiDAR semantic segmentation framework, named ItTakesTwo (IT2). IT2 employs consistency learning based on peer-representations perturbation, which enforces the consistent predictions across two different representations. Such perturbation contributes with a more effective SSL generalisation and alleviates issues related to pseudo label qualities from a single representation. Furthermore, we propose cross-distribution contrastive learning, where instead of randomly selecting training samples in the embedding space, we learn informative embedding distributions (also known as prototype) via Gaussian Mixture Models from both representations. Then the embedding samples are drawn from such prototypical distributions to produce representative features that enhance the understanding of the latent space in peer representation scenarios. In addition, instead of applying a single type of data augmentation to both LiDAR representations [27], we empirically show that applying different augmentations to distinct representations provides better generalisation. To summarise, our work makes the following key contributions:", + "bbox": [ + 77, + 430, + 917, + 720 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. A novel SSL LiDAR segmentation framework (IT2) that leverages peer-representation perturbation in consistency learning, encouraging the same predictions for the points with the same semantic meaning in different LiDAR representations;", + "2. A new contrastive learning where informative positive and negative pairs of embeddings are sampled from prototype distributions (modeled via Gaussian Mixture Models) of both representations; and", + "3. An innovative representation-specific data augmentation that enables IT2 achieve better generalisation." + ], + "bbox": [ + 84, + 726, + 917, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our IT2 has state-of-the-art (SOTA) performance on nuScenes [4], SemanticKitti [2] and ScribbleKitti [52], and we also achieve best performance with different", + "bbox": [ + 77, + 908, + 944, + 947 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "ItTakesTwo", + "bbox": [ + 721, + 16, + 837, + 30 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 903, + 16, + 917, + 30 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sampling strategies to form the labelled and unlabelled subsets. Notably, in nuScenes [4] dataset, we outperform the previous SOTA [27] by around $4.6\\%$ in the limited labelled data situations.", + "bbox": [ + 79, + 55, + 917, + 112 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 79, + 144, + 326, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LiDAR Semantic Segmentation is an essential task for autonomous driving [17]. To more effectively extract the feature information from the unordered LiDAR point data, current approaches [45, 66, 67] transform such point data into different representations. Range representation methods [26, 45, 65] transfer the LiDAR points to the spherical coordinates [45] and project them to the image space [26, 65]. Voxel representation methods [15, 28, 66] voxelise the LiDAR points and project them into different sub-types of geometry views to alleviate the sparse data challenges [10, 15], encompassing cartesian [53], spherical [28] and cylindrical [66] coordinate systems. Furthermore, some methods [34, 38] take inputs from a combination of LiDAR representations [48, 67] or incorporate inputs from camera modalities [34, 40], fusing them in the intermediate layer to enhance feature understanding. Although these supervised methods show promising results [58, 65, 66], there has been a longstanding interest in minimising the extensive labeling efforts, with approaches such as scribble annotation [52], box supervision [37], active learning [35, 60], and SSL [23, 27]. In this paper, we focus on SSL to alleviate such annotation costs.", + "bbox": [ + 79, + 188, + 917, + 497 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semi-supervised LiDAR Semantic Segmentation leverages unlabelled data to enhance the model's generalisation beyond the small labelled dataset. The current literature on LiDAR SSL primarily focuses on indoor [7, 12, 32, 60] or object [25, 49] point cloud scenarios, which have uniformly distributed and high density points in the scan. Such scenario is markedly different from the representation of outdoor LiDAR points that pose unique challenges, such as complex structure [27] and low density points [66]. Previous semi-supervised LiDAR segmentation methods [23, 31, 46] mainly rely on contrastive learning to improve the latent space understanding. For instance, GPC [23] utilises confidence-threshold to constrain potentially noisy pseudo labels, and Lim3D [31] uses a memory bank to save negative samples. To increase the variation of the LiDAR data, LaserMix [27] proposes a novel augmentation that mixes laser beams in various spatial positions and holds the consistency learning under the Mean Teacher (MT) architecture [50]. However, these methods use only single LiDAR representations which, as explained in Sec. 1, imposes many limitations, such as limited perturbation effectiveness and poor performance for specific categories. Our framework introduces a novel peer-representation perturbation that employs various LiDAR representations during training, resulting in improved generalisation.", + "bbox": [ + 79, + 502, + 917, + 850 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contrastive Learning [5, 24, 54, 56] enforces latent embeddings to be closer to instances with the same semantic meaning (i.e. positive) and to be far away from instances with different semantic meanings (i.e. negative). This approach has been explored in LiDAR segmentation [7, 12, 23, 31, 46], but the reliance on actual training samples leads to sub-optimal training as only a subset of em", + "bbox": [ + 79, + 850, + 917, + 947 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 79, + 17, + 94, + 29 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 159, + 16, + 280, + 30 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d3f22719c0b6c5a2809055d3d6a1f0aba4dceb103fcdfcc6e5bca31ade8d09ab.jpg", + "image_caption": [ + "Fig. 2: Illustration of our approach with unlabeled data. The point scans in the batch $(\\mathbf{x}_1$ and $\\mathbf{x}_2$ on the left) are transformed into range (and processed by $f_{\\phi_{\\mathrm{FCN}}}(.))$ and voxel (and processed by $f_{\\theta_{\\mathrm{EXR}}}(.))$ representations. Subsequently, the results from $f_{\\phi_{\\mathrm{SEG}}}(f_{\\phi_{\\mathrm{FCN}}}(.))$ and $f_{\\theta_{\\mathrm{SEG}}}(f_{\\theta_{\\mathrm{EXR}}}(.))$ are transformed to the other representation with Eq. (2) and cross-supervised with Eq. (4). Also, the embedding $\\{\\cdot ,\\cdot \\}$ selected by the pseudo labels are utilised to model the GMMs with Eq. (6) and penalised by the informative samples (generated from GMMs) based on $\\ell_{cross}(.)$ in Eq. (8). Please note the arrow indicates that the operation does not require gradient computation." + ], + "image_footnote": [], + "bbox": [ + 79, + 52, + 922, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "bedding pairs [23, 31] can be selected for computing the contrastive loss. This becomes more challenging when working with embeddings obtained from different representations [36, 51, 59], as potentially noisy predictions [44, 59] from each representation can result in confirmation bias during training [1]. Prototypical contrastive learning [5, 18, 29, 30, 33] explores the use of a set of representative embeddings for the contrastive learning via memory bank [18, 54], optimal transport [5], or deep clustering [29, 30]. Nevertheless, such prototypes represent the modes of a more complex underlying distribution of prototypes that has not been explored before for SSL LiDAR segmentation. In our framework, we employ Gaussian Mixture Models (GMMs) to represent class-specific embedding distributions from both LiDAR representations and to generate informative embeddings for all classes, enhancing the effectiveness of the contrastive learning.", + "bbox": [ + 72, + 471, + 917, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 75, + 738, + 317, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our approach is based on a labelled point dataset $\\mathcal{P}_L = \\{(\\mathbf{x}_i,\\mathbf{y}_i)\\}_{i = 1}^{| \\mathcal{P}_L |}$ consisting of $|\\mathcal{P}_L|$ scans and an unlabelled set $\\mathcal{P}_U = \\{(\\mathbf{x}_i\\}_{i = 1}^{| \\mathcal{P}_U|}$ with $|\\mathcal{P}_U|$ scans. Each scan $\\mathbf{x}_i\\in \\mathcal{P}\\subset \\mathbb{R}^{3\\times C\\times N}$ has 3 spatial coordinates, with $C$ denoting the number of features (intensity, elongation, etc.), $N$ representing the number of points, and $\\mathbf{y}_i\\in \\mathcal{V}\\subset \\{0,1\\}^{Y\\times N}$ represents the one-hot label with $Y$ being the number of classes in the labelled set. This point cloud dataset is transformed into the range and voxel representations to build the two peer representation datasets that is required by our method. These transformations are defined by [58] as:", + "bbox": [ + 72, + 786, + 920, + 949 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "ItTakesTwo", + "bbox": [ + 721, + 16, + 837, + 30 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 903, + 16, + 917, + 29 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $\\psi_{p\\rightarrow v}:\\mathcal{P}\\rightarrow \\mathcal{V}$ is a point to voxel projection function that forms $\\mathcal{V}_L = \\{(x_i^v,y_i^v)|(x_i,y_i)\\in \\mathcal{P}_L\\}_{i = 1}^{|\\mathcal{P}_L|}$ and $\\mathcal{V}_U = \\{\\mathbf{x}_i^v |\\mathbf{x}_i\\in \\mathcal{P}_U\\}_{i = 1}^{|\\mathcal{P}_U|}$ , with each volume $\\mathbf{x}_i^v\\in \\mathcal{V}\\subset \\mathbb{R}^{C_v\\times H\\times W\\times L}$ having size $H\\times W\\times L$ and $C_v$ features, and the projected label being $\\mathbf{y}_i^v\\in \\{0,1\\}^{Y\\times H\\times W\\times L}$ in one-hot format.", + "- $\\psi_{p\\rightarrow r}:\\mathcal{P}\\rightarrow \\mathcal{R}$ is a function to project point to the range image6. With this, we build $\\mathcal{R}_L = \\{(\\mathbf{x}_i^r,\\mathbf{y}_i^r)|(x_i,y_i)\\in \\mathcal{P}_L\\}_{i = 1}^{|\\mathcal{P}_L|}$ and $\\mathcal{R}_U = \\{\\mathbf{x}_i^r |\\mathbf{x}_i\\in \\mathcal{P}_U\\}_{i = 1}^{|\\mathcal{P}_U|}$ , with each range image $\\mathbf{x}_i^r\\in \\mathcal{R}\\subset \\mathbb{R}^{C_r\\times U\\times V}$ having size $U\\times V$ with $C_r$ features, and the projected label being $\\mathbf{y}_i^r\\in \\{0,1\\}^{Y\\times U\\times V}$ ." + ], + "bbox": [ + 89, + 55, + 920, + 219 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The inverse of the transformations above are respectively denoted by $\\psi_{v\\rightarrow p}$ and $\\psi_{r\\rightarrow p}$ . Based on the transformed range and voxel representations from $\\mathcal{P}_L$ and $\\mathcal{P}_U$ , we define the labelled set $\\mathcal{D}_L = \\{(\\mathbf{x}_i^v,\\mathbf{y}_i^v),(\\mathbf{x}_i^r,\\mathbf{y}_i^r)|(\\mathbf{x}_i^v,\\mathbf{y}_i^v)\\in \\mathcal{V}_L,(\\mathbf{x}_i^r,\\mathbf{y}_i^r)\\in \\mathcal{R}_L,\\}_{i = 1}^{|\\mathcal{D}_L|}$ and the larger unlabelled set $\\mathcal{D}_U = \\{(\\mathbf{x}_i^v,\\mathbf{x}_i^r)|\\mathbf{x}_i^v\\in \\mathcal{V}_U,\\mathbf{x}_i^r\\in \\mathcal{R}_U\\}_{i = 1}^{|\\mathcal{D}_U|}$ where $|\\mathcal{D}_L|\\leq |\\mathcal{D}_U|$ . These datasets are used to train our IT2 methodology in Sec. 3.1, while Sec. 3.2 describes our Gaussian Mixture Models (GMMs) based prototypical contrastive learning. Lastly, we introduce the details of our augmentation method in Sec. 3.3.", + "bbox": [ + 72, + 227, + 922, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 ItTakesTwo (IT2) Framework", + "text_level": 1, + "bbox": [ + 75, + 406, + 503, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, our approach enforces prediction consistency between two peer representations to explore a potentially more effective perturbation. Our method can use various LiDAR representations, but we follow [27] and employ range and voxel networks for a fair comparison in the benchmark. The range image network consists of an FCN Network (e.g., ResNet [19]) that maps the input to a latent dense feature space with $f_{\\phi_{\\mathbf{FCN}}} : \\mathcal{R} \\to \\mathcal{I}$ , where $\\mathcal{I} \\subset \\mathbb{R}^{U \\times V \\times D_r}$ with $D_r$ denoting feature dimensionality. Then a segmentation head will map $\\mathcal{I}$ to the categorical segmentation map with $f_{\\phi_{\\mathbf{SEG}}} : \\mathcal{I} \\to \\mathbb{R}^{U \\times V \\times Y}$ . Similarly, the voxel network extracts features with $f_{\\theta_{\\mathbf{EXR}}} : \\mathcal{V} \\to \\mathcal{K}$ , where $\\mathcal{K} \\in \\mathbb{R}^{H \\times W \\times L \\times D_v}$ with $D_v$ denoting the feature dimensionality, followed by a segmentation head $f_{\\theta_{\\mathbf{SEG}}} : \\mathcal{K} \\to \\mathbb{R}^{H \\times W \\times L \\times Y}$ . The peer representation consistency training depends on the following categorical distributions for both views:", + "bbox": [ + 72, + 435, + 922, + 671 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {y}} ^ {r} \\left(\\omega^ {r}\\right) = \\operatorname {s o f t m a x} ^ {\\left(\\omega^ {r}\\right)} \\left(f _ {\\phi_ {\\mathrm {S E G}}} \\left(f _ {\\phi_ {\\mathrm {F C N}}} \\left(\\mathbf {x} ^ {r}\\right)\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 273, + 681, + 915, + 713 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {y}} ^ {v} \\left(\\omega^ {v}\\right) = \\operatorname {s o f t m a x} \\left(\\omega^ {v}\\right) \\left(f _ {\\theta_ {\\mathrm {S E G}}} \\left(f _ {\\theta_ {\\mathrm {E X R}}} \\left(\\mathbf {x} ^ {v}\\right)\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 278, + 710, + 716, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{y}}^r$ and $\\hat{\\mathbf{y}}^v$ are the categorical distributions for the range image (in $\\omega^r$ in the $U\\times V$ lattice $\\Omega^r$ ) and voxel (in $\\omega^v$ in the $H\\times W\\times L$ lattice $\\Omega^v$ ) representations, and $\\mathrm{softmax}^{(\\omega)}(.)$ computes the softmax activation at $\\omega$ . Based on these outputs, we perform the following cross-view transformations:", + "bbox": [ + 72, + 744, + 922, + 822 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {y}} ^ {r} \\left(\\omega^ {r}\\right) = \\operatorname {a r g m a x} \\Psi_ {v \\rightarrow r} ^ {\\left(\\omega^ {r}\\right)} \\left(\\hat {\\mathbf {y}} ^ {v}\\right), \\quad \\mathbf {c} ^ {r} \\left(\\omega^ {r}\\right) = \\max \\Psi_ {v \\rightarrow r} ^ {\\left(\\omega^ {r}\\right)} \\left(\\hat {\\mathbf {y}} ^ {v}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 832, + 915, + 866 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {y}} ^ {v} \\left(\\omega^ {v}\\right) = \\operatorname {a r g m a x} \\Psi_ {r \\rightarrow v} ^ {\\left(\\omega^ {v}\\right)} \\left(\\hat {\\mathbf {y}} ^ {r}\\right), \\quad \\mathbf {c} ^ {v} \\left(\\omega^ {v}\\right) = \\max \\Psi_ {r \\rightarrow v} ^ {\\left(\\omega^ {v}\\right)} \\left(\\hat {\\mathbf {y}} ^ {r}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 861, + 772, + 882 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 77, + 16, + 94, + 30 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 157, + 14, + 280, + 30 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "6 Range-to-point projection can lead to information loss [45], which is typically mitigated by post-processing with K-Nearest Neighbors (KNN). In pursuit of efficiency, our IT2 does not use any post-processing during training.", + "bbox": [ + 79, + 892, + 922, + 949 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{y}}^r$ and $\\tilde{\\mathbf{y}}^v$ represent the transformed hard pseudo label maps, $\\mathbf{c}^r$ and $\\mathbf{c}^v$ denote the transformed confidence maps for range and voxel representations, respectively, $\\varPsi_{v\\rightarrow r}=\\psi_{p\\rightarrow r}\\circ\\psi_{v\\rightarrow p},\\varPsi_{r\\rightarrow v}=\\psi_{p\\rightarrow v}\\circ\\psi_{r\\rightarrow p}$ , and the superscript $\\omega$ indicates that we only consider the output at $\\omega$ . We use the cross-view pseudo labels from Eq. (2) to formulate the loss function to train IT2, as follows:", + "bbox": [ + 72, + 56, + 920, + 152 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\mathrm {I T 2}} \\left(\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\phi_ {\\mathrm {F C N}}}, f _ {\\phi_ {\\mathrm {E X R}}}, f _ {\\phi_ {\\mathrm {S E G}}}\\right) = \\ell_ {\\text {r a n g e}} \\left(\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\phi_ {\\mathrm {F C N}}}, f _ {\\phi_ {\\mathrm {S E G}}}\\right) + \\ell_ {\\text {v o x e l}} \\left(\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\theta_ {\\mathrm {E X R}}}, f _ {\\theta_ {\\mathrm {S E G}}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 167, + 915, + 204 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where we formulate the loss for each view as:", + "bbox": [ + 75, + 206, + 559, + 222 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\ell_ {\\mathrm {r a n g e}} (\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\phi_ {\\mathbf {F C N}}}, f _ {\\phi_ {\\mathbf {S E G}}}) = \\frac {1}{| \\mathcal {D} _ {L} | \\times | \\Omega^ {r} |} \\sum_ {i = 1} ^ {| \\mathcal {D} _ {L} |} \\sum_ {\\omega^ {r} \\in \\Omega^ {r}} \\ell (\\mathbf {y} _ {i} ^ {r} (\\omega^ {r}), f _ {\\phi_ {\\mathbf {S E G}}} ^ {(\\omega^ {r})} (f _ {\\phi_ {\\mathbf {F C N}}} (\\mathbf {x} _ {i} ^ {r})) \\\\ + \\frac {1}{\\left| \\mathcal {D} _ {U} \\right| \\times \\left| \\Omega^ {r} \\right|} \\sum_ {i = 1} ^ {\\left| \\mathcal {D} _ {U} \\right|} \\sum_ {\\omega^ {r} \\in \\Omega^ {r}} \\ell \\left(\\tilde {\\mathbf {y}} _ {i} ^ {r} \\left(\\omega^ {r}\\right), f _ {\\phi_ {\\mathrm {S E G}}} ^ {\\left(\\omega^ {r}\\right)} \\left(f _ {\\phi_ {\\mathrm {F C N}}} \\left(\\mathbf {x} _ {i} ^ {r}\\right)\\right), \\right. \\tag {4} \\\\ \\ell_ {\\mathrm {v o x e l}} (\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\theta_ {\\mathbf {E X R}}}, f _ {\\theta_ {\\mathbf {S E G}}}) = \\frac {1}{| \\mathcal {D} _ {L} | \\times | \\Omega^ {v} |} \\sum_ {i = 1} ^ {| \\mathcal {D} _ {L} |} \\sum_ {\\omega^ {v} \\in \\Omega^ {v}} \\ell (\\mathbf {y} _ {i} ^ {v} (\\omega^ {v}), f _ {\\theta_ {\\mathbf {S E G}}} ^ {(\\omega^ {v})} (f _ {\\theta_ {\\mathbf {E X R}}} (\\mathbf {x} _ {i} ^ {v})) \\\\ + \\frac {1}{| \\mathcal {D} _ {U} | \\times | \\Omega^ {v} |} \\sum_ {i = 1} ^ {| \\mathcal {D} _ {U} |} \\sum_ {\\omega^ {v} \\in \\Omega^ {v}} \\ell (\\tilde {\\mathbf {y}} _ {i} ^ {v} (\\omega^ {v}), f _ {\\theta_ {\\mathrm {S E G}}} ^ {(\\omega^ {v})} (f _ {\\theta_ {\\mathrm {E X R}}} (\\mathbf {x} _ {i} ^ {v})), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 235, + 915, + 437 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with the loss function $\\ell(.)$ denoting the combination of cross-entropy loss and Lovasz-softmax [3] loss, following [27, 31, 66].", + "bbox": [ + 75, + 450, + 917, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Cross-distribution Contrastive Learning", + "text_level": 1, + "bbox": [ + 75, + 517, + 629, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We model the cross distribution of the embedding space using a multi-layer perceptron to generate latent embeddings in parallel with the segmentation head. Specifically, we obtain the range image embedding space via $f_{\\phi_{\\mathbf{PRJ}}}^r: \\mathcal{I} \\to \\mathbb{R}^{U \\times V \\times Z}$ with $Z$ denoting the pixel-wise number of dimensions in the projected space. Similarly, the voxel embedding space is obtained via $f_{\\theta_{\\mathbf{PRJ}}}^v: \\mathcal{K} \\to \\mathbb{R}^{H \\times W \\times L \\times Z}$ , where $\\mathcal{I}$ and $\\mathcal{K}$ are intermediate feature spaces introduced in Sec. 3.1. We formulate the observed embedding results with:", + "bbox": [ + 72, + 547, + 917, + 687 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} ^ {r} \\left(\\omega^ {r}\\right) = f _ {\\phi_ {\\mathbf {P R J}}} ^ {\\left(\\omega^ {r}\\right)} \\left(f _ {\\phi_ {\\mathbf {F C N}}} \\left(\\mathbf {x} ^ {r}\\right)\\right), \\quad \\mathbf {z} ^ {v} \\left(\\omega^ {v}\\right) = f _ {\\theta_ {\\mathbf {P R J}}} ^ {\\left(\\omega^ {v}\\right)} \\left(f _ {\\theta_ {\\mathbf {E X R}}} \\left(\\mathbf {x} ^ {v}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 699, + 915, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{z}^r$ and $\\mathbf{z}^v$ represent the latent point-wise embeddings of the range and voxel representations, respectively.", + "bbox": [ + 72, + 734, + 917, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As mentioned in Sec. 2, contrastive learning methods [7, 12, 23, 31, 46] only use a limited subset of actual training samples present in a mini batch, potentially leading to suboptimal convergence with confirmation bias. We propose a novel approach to represent the range and voxel embeddings with class-specific Gaussian Mixture Models (GMMs). Such approach enables the sampling of the most informative embeddings, fostering a comprehensive understanding of the distribution across the feature space of different representations. In practice, we concatenate the embeddings from both range and voxel based on their categories, where such embedding set is defined by $\\mathcal{G}^y = \\{(\\mathbf{z},\\mathbf{c},y)|\\mathbf{z} = \\mathbf{z}_i^r (\\omega^r),\\mathbf{c} =$", + "bbox": [ + 72, + 773, + 920, + 949 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "ItTakesTwo", + "bbox": [ + 721, + 14, + 837, + 30 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 903, + 16, + 917, + 30 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathbf{c}_i^r (\\omega^r),y = \\tilde{\\mathbf{y}}_i^r (\\omega^r)\\} \\bigcup \\{(\\mathbf{z},\\mathbf{c},y)|\\mathbf{z} = \\mathbf{z}_i^v (\\omega^v),\\mathbf{c} = \\mathbf{c}_i^v (\\omega^v),y = \\tilde{\\mathbf{y}}_i^v (\\omega^v)\\}$ for the unlabelled data indexed by $i\\in \\{1,\\dots,|\\mathcal{D}_U|\\}$ , with the related confidence (i.e., $\\mathbf{c}^r$ and $\\mathbf{c}^v$ ) and pseudo labels (i.e., $\\tilde{\\mathbf{y}}^r$ and $\\tilde{\\mathbf{y}}^v$ ) from Eq. (2) at positions $\\omega^r\\in \\Omega^r$ and $\\omega^v\\in \\varOmega^v$ . The class-specific GMMs are defined for both representations by", + "bbox": [ + 72, + 55, + 920, + 134 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {\\Gamma^ {y}} (\\mathbf {z} | y) = \\sum_ {m = 1} ^ {M} \\mathbf {P} _ {\\Gamma^ {y}} (m | y) \\mathbf {P} _ {\\Gamma^ {y}} (\\mathbf {z} | m, y) = \\sum_ {m = 1} ^ {M} \\pi_ {m} ^ {y} \\mathcal {N} (\\mathbf {z}; \\boldsymbol {\\mu} _ {m} ^ {y}, \\boldsymbol {\\Sigma} _ {m} ^ {y}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 141, + 917, + 188 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\pmb{\\pi}_{m}^{y} = \\mathbf{P}_{\\varGamma^{y}}(m|y)$ denotes the prior probability for the $m$ -th GMM component of class $y \\in \\{0,1\\}^Y$ , $\\pmb{\\mu}_{m}^{y} \\in \\mathbb{R}^{Z}$ and $\\pmb{\\Sigma}_{m}^{y} \\in \\mathbb{R}^{Z \\times Z}$ represent the mean and covariance of the corresponding GMM component and we fix $M = 5$ for all experiments. The GMM parameter optimisation in Eq. (6) relies on the EM [11] algorithm using the training set $\\mathcal{G}^{y}$ defined above with the following steps:", + "bbox": [ + 72, + 195, + 917, + 294 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "E step: $\\mathbf{q}_{n,m}^{y} = \\frac{\\pi_{m}^{y}\\mathcal{N}(\\mathbf{z}_{n}\\mid\\boldsymbol{\\mu}_{m}^{y},\\boldsymbol{\\Sigma}_{m}^{y})}{\\sum_{m}^{M}\\pi_{m}^{y}\\mathcal{N}(\\mathbf{z}_{n}\\mid\\boldsymbol{\\mu}_{m}^{y},\\boldsymbol{\\Sigma}_{m}^{y})}$ , for $(\\mathbf{z}_n,\\mathbf{c}_n,y_n)\\in \\mathcal{G}^y$", + "bbox": [ + 147, + 297, + 852, + 344 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "M step: $\\pmb{\\mu}_{m}^{y} = \\frac{1}{|\\mathcal{G}^{y}|}\\sum_{(\\mathbf{z}_{n},\\mathbf{c}_{n},y_{n})\\in \\mathcal{G}^{y}}\\mathbf{c}_{n}\\mathbf{q}_{n,m}^{y_{n}}\\mathbf{z}_{n}$ (7)", + "bbox": [ + 147, + 347, + 915, + 393 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol{\\Sigma}_{m}^{y} = \\frac{1}{|\\mathcal{G}^{y}|}\\sum_{(\\mathbf{z}_{n},\\mathbf{c}_{n},y_{n})\\in \\mathcal{G}^{y}}\\mathbf{c}_{n}\\mathbf{q}_{n,m}^{y_{n}}(\\mathbf{z}_{n} - \\boldsymbol{\\mu}_{m}^{y_{n}})\\bigl(\\mathbf{z}_{n} - \\boldsymbol{\\mu}_{m}^{y_{n}}\\bigr)^{\\intercal},\n$$\n", + "text_format": "latex", + "bbox": [ + 263, + 400, + 842, + 447 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{G}^y$ represents the set of embedding samples that share the same label $y$ . The posterior $\\mathbf{q}_{n,m}^y$ estimates the responsibility of the $n^{th}$ embedding $\\mathbf{z}$ from class $y$ in $\\mathcal{G}^y$ being assigned to the $m$ -th component. The confidence $\\mathbf{c}_n$ represents the potential correctness of the pseudo label from Eq. (2), which is used as a weighting factor to reduce the effect of low-confidence segmentation results [8,9] in the estimation of the GMM parameters. Note that we maintain a uniform weight of $\\pi_{m}^{y} = 1 / M$ for each component [43, 57] to prevent one component from dominating others.", + "bbox": [ + 72, + 453, + 917, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GMM-based Prototypical Contrastive Learning. The class-specific GMMs from Eq. (7) are used to generate a set of highly representative embeddings for the contrastive learning process explained next. These representative class-specific embeddings is denoted by $S^y = \\{\\mathbf{s}|\\mathbf{s}\\sim \\mathbf{P}_{\\Gamma^y}(\\mathbf{z}|y)\\}$ , with $\\mathbf{P}_{\\Gamma^y}(\\mathbf{z}|y)$ being the class-specific GMM defined in Eq. (6). Subsequently, these generated virtual prototypes are utilised to guide the contrastive learning for the incoming embedding features from both range and voxel representations [54], without joining the backpropagation computation graph. We define the set of incoming features with the anchor set $\\mathcal{A} = \\{(\\mathbf{z},y)\\}$ that is formed similarly to how we form the set $\\mathcal{G}^y$ without including the confidence value. The samples in $\\mathcal{A}$ are selectively chosen for backpropagation [54], but the informative samples in $S^y$ contribute to a comprehensive understanding of the cross-distribution for both representations. The contrastive learning relies on the minimisation of the following loss:", + "bbox": [ + 72, + 627, + 917, + 882 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {c r o s s}} \\left(\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\phi_ {\\mathbf {P R J}}}, f _ {\\phi_ {\\mathbf {F C N}}}, f _ {\\theta_ {\\mathbf {P R J}}}, f _ {\\theta_ {\\mathbf {E X R}}} \\right) = \\sum_ {(\\mathbf {z}, y) \\in \\mathcal {A}} \\sum_ {\\mathbf {s} \\in \\mathcal {S} ^ {y}} \\frac {\\exp (\\mathbf {z} \\cdot \\mathbf {s} / \\tau)}{\\exp (\\mathbf {z} \\cdot \\mathbf {s} / \\tau) + \\sum_ {\\mathbf {s} ^ {-} \\in \\bar {\\mathcal {S}} ^ {y}} \\exp (\\mathbf {z} \\cdot \\mathbf {s} ^ {-} / \\tau)}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 79, + 889, + 915, + 947 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 77, + 16, + 94, + 29 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 157, + 14, + 280, + 30 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\bar{S}^y$ is the set of virtual negative embeddings for class $y$ .", + "bbox": [ + 75, + 55, + 736, + 74 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 Representation Specific Data Augmentation", + "text_level": 1, + "bbox": [ + 75, + 105, + 680, + 125 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Following [39, 61], we apply spatial data augmentations to the pseudo-labeling strategy in Eq. (2) to increase the robustness of our approach to complex scenes. After experimenting with classic data augmentation processes (e.g., CutMix [63] and LaserMix [27]), we notice that the application of distinct spatial augmentations to different viewpoints in our IT2 framework results in more effective generalization. We have empirically found the multi-boxes CutMix [63] augmentation for the range image and single-inclination LaserMix [27] augmentation for the voxel representation yield the most robust results. Although this representation-specific data augmentation is not our main contribution, it is shown empirically to produce relevant improvements. Please see details in supplementary Sec. 2.", + "bbox": [ + 72, + 139, + 922, + 336 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Overall Training", + "text_level": 1, + "bbox": [ + 75, + 364, + 346, + 385 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The overall training of our IT2 minimises the following loss function:", + "bbox": [ + 75, + 398, + 811, + 419 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} \\left(\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, \\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\phi_ {\\mathrm {F C N}}}, f _ {\\phi_ {\\mathrm {E X R}}}, f _ {\\phi_ {\\mathrm {S E G}}} f _ {\\phi_ {\\mathrm {P R J}}}, f _ {\\phi_ {\\mathrm {P R J}}}\\right) = \\\\ \\ell_ {\\mathrm {I T 2}} \\left(\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\phi_ {\\mathbf {F C N}}}, f _ {\\phi_ {\\mathbf {E X R}}}, f _ {\\phi_ {\\mathbf {S E G}}}\\right) + \\ell_ {\\mathrm {c r o s s}} \\left(\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\phi_ {\\mathbf {P R J}}}, f _ {\\phi_ {\\mathbf {F C N}}}, f _ {\\theta_ {\\mathbf {P R J}}}, f _ {\\theta_ {\\mathbf {E X R}}}\\right), \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 79, + 435, + 915, + 495 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where the entire training pipeline is performed in an end-to-end manner, where in each iteration, we minimise the loss from $\\ell_{\\mathrm{IT2}}(.)$ and $\\ell_{\\mathrm{cross}}(.)$ simultaneously.", + "bbox": [ + 75, + 497, + 917, + 538 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 75, + 560, + 300, + 582 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 75, + 603, + 389, + 622 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate our approach on three benchmarks, including nuScenes [4], SemanticKITTI [2] and ScribbleKITTI [52]. The nuScenes [4] dataset has 29,130 training scans and 6,019 validation scans. Each LiDAR scan contains 32 beams with $[-30^{\\circ},10^{\\circ}]$ vertical field of views (FOV). The dataset has 23 categories, but following previous works [27,66], we merge them into 16 for training and validation. The SemanticKITTI [2] and its extension ScribbleKITTI [52] are popular laser-based segmentation benchmarks, where the latter one is based on line-scribbles labelling to reduce the annotation effort. The SemanticKITTI [2] dataset has 19 categories for semantic segmentation, where every LiDAR scan contains 64 beams with $[3^{\\circ}, -25^{\\circ}]$ vertical FOV. There are 22 sequences of LiDAR scans in the dataset, where sequences 0 to 9 are used for training and validation. Following [23,27], we use 19,130 scans (sequences 0 to 7 and sequence 9) for training and 4,071 scans from sequence 8 for validation.", + "bbox": [ + 72, + 635, + 920, + 889 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Sampling strategies. Previous works [23, 27, 31] propose different sampling strategies to select the labelled data for the semi-supervised learning setting, including: 1) uniform [27] sampling of the labelled data across all the training", + "bbox": [ + 72, + 889, + 920, + 949 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ItTakesTwo", + "bbox": [ + 719, + 14, + 837, + 30 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 900, + 16, + 917, + 30 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/637f723ca22bc082fe4d2ea7a1ce64b966416418a589cd4373ab8d3b8b6d03a4.jpg", + "table_caption": [ + "Table 1: Uniform sampling results across various benchmarks, where we follow [27] to utilise half feature dimension and lower resolution in Cylinder3D [66] for all the experiments. Best results are highlighted in red." + ], + "table_footnote": [], + "table_body": "
Repr.MethodnuScenes [4]SemanticKITTI [2]ScribbleKITTI [52]
1%10%20%50%1%10%20%50%1%10%20%50%
Rangesup.38.357.562.767.636.252.255.957.233.147.749.952.5
MT [50]42.160.465.469.437.553.156.157.434.249.851.653.3
CBST [68]40.960.564.369.339.953.456.156.935.750.752.754.6
CPS [6]40.760.864.968.036.552.356.357.433.750.052.854.6
LaserMix [27]49.568.270.673.043.458.859.461.438.354.455.658.7
IT2 (Ours)56.571.373.474.051.960.361.762.146.657.157.358.6
Voxelsup.50.965.966.671.245.456.157.858.739.248.052.153.8
MT [50]51.666.067.171.745.457.159.260.041.050.152.853.9
CBST [68]53.066.569.671.648.858.359.459.741.550.653.354.5
CPS [6]52.966.370.072.546.758.759.660.541.451.853.954.8
LaserMix [27]55.369.971.873.250.660.061.962.344.253.755.156.8
IT2 (Ours)57.572.173.674.152.061.462.162.547.956.757.558.3
", + "bbox": [ + 84, + 123, + 917, + 422 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "sequences, 2) partial [23] sampling of the labelled data only from certain sequences, and 3) sampling of significant [31] labelled frames from the sequences. In the experiment, we examine our approach under all these sampling protocols. Metrics. We measure performance with mean intersection over union (mIoU). In each benchmark, we use the sampling strategies above to label $n\\%$ of the training set, with the rest $(100 - n)\\%$ being unlabelled.", + "bbox": [ + 77, + 456, + 917, + 573 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation details. In the range view, we use the FIDNet architecture [65] with ResNet34 [19] backbone. The input range image has size $32 \\times 1920$ for nuScenes [4] and $64 \\times 2048$ for SemanticKITTI [2] and ScirbbleKITTI [52], following [27]. In the voxel representation, we use the Cylinder3D architecture [66], where we fix the intermediate layer at 16 dimensions with smaller resolution as specified in [27] for the uniform sampling experiments to enable a fair comparison. For contrastive learning, we build a 3-layer projector [24,54]. The projector outputs an embedding map with 64 channels, where we apply the temperature factor $\\tau = .1$ in Eq. (8). See supplementary Sec. 3 and Sec. 7 for more details.", + "bbox": [ + 77, + 575, + 917, + 749 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Results on Different Partition Protocols", + "text_level": 1, + "bbox": [ + 79, + 760, + 629, + 780 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Uniform sampling results. We first compare our method using the uniform sampling benchmark [27] in Tab. 1, with the experimental setup from [27] (i.e., architecture, split ID and the resolution of input representations). Our framework outperforms the previous SOTA Lasermix [27] by a large margin for almost all partition protocols. For example, we observe $\\approx 3.4\\%$ and $\\approx 2.0\\%$ improvements on average for range and voxel representations, respectively. Notably, we highlight that our approach yields significant improvements for the $1\\%$ partition protocol in the line-scribble dataset [52], providing $8.28\\%$ improvement in the", + "bbox": [ + 79, + 791, + 917, + 947 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 157, + 14, + 280, + 30 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/275f85492edccb84f1b28bbacda925eb05360cabf20f6551e69be8ccdf89a89b.jpg", + "table_caption": [ + "Table 2: Partial & significant sampling results in the SemanticKITTI [2] benchmark. We compare with previous SOTA based on Cylinder3D [66] architecture, where TTA denotes test time augmentation. $\\dagger$ indicates that we re-implement results based on the supported checkpoints from their GitHub - please see Supplementary Sec. 1 for more details. Best results in each evaluation protocol are highlighted in red." + ], + "table_footnote": [], + "table_body": "
TTAMethodSemanticKITTI(par.)MethodSemanticKITTI(sig.)
5%10%20%40%5%10%20%40%
XGPC† [23]39.945.955.757.1lim3D [31]56.859.660.561.4
IT2 (Ours)43.448.156.560.9IT2 (Ours)60.363.364.064.8
✓GPC† [23]40.246.356.557.6lim3D [31]59.562.263.163.3
IT2 (Ours)43.849.258.862.4IT2 (Ours)61.764.464.765.8
", + "bbox": [ + 125, + 159, + 878, + 295 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "range image and $3.66\\%$ in the voxel representation.", + "bbox": [ + 77, + 331, + 619, + 351 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Partial and significant sampling results. In Tab. 2, we compare our results with GPC [23] and lim3D [31] in partial sampling (par.) and significant sampling (sig.), based on both original performance and test time augmentation (TTA). In the partial sampling protocol7, our approach yields superior results across all the labelled data situations for both with and without TTA. Notably, our results surpass GPC [23] by $4.5\\%$ and $3.6\\%$ in the limited $5\\%$ labelled data condition, demonstrating its good generalisation capability. In the significant sampling results, our method also achieves better performance than lim3D [31] for all the partition protocols. For example, our results are better than the previous SOTA by about $2.2\\%$ and $1.6\\%$ in $10\\%$ and $20\\%$ labelled data with TTA, respectively.", + "bbox": [ + 79, + 352, + 917, + 546 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 79, + 560, + 329, + 580 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Improvements from our contributions. Table 3 presents our ablation study on the nuScenes [4] and ScribbleKITTI [52] datasets across various partition protocols. We investigate the roles of the IT2 architecture, cross-distribution contrastive learning (Ctrs.), and representation-related data augmentation (Aug.). The baseline utilises network perturbation through single representations [6] from [27]. Our IT2 architecture results in improvements of approximately $4.1\\%$ and $2.5\\%$ for the voxel representation in $10\\%$ labelled data for nuScenes [4] and ScribbleKITTI [52], respectively. The Ctrs. further improves by $\\approx 2.4\\%$ and $\\approx 1.1\\%$ for range and voxel representations in both datasets in $10\\%$ labelled data. Such enhancements show the effectiveness of our cross-distribution prototypes and contrastive learning strategy. Furthermore, Aug. results in improvements of $2.2\\%$ and $1.4\\%$ for range and voxel representations in the ScribbleKITTI [52] dataset for $10\\%$ labelled data, demonstrating the better generalisation brought by the representation-specific data augmentation.", + "bbox": [ + 79, + 593, + 917, + 863 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ItTakesTwo", + "bbox": [ + 721, + 14, + 837, + 30 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 893, + 16, + 915, + 30 + ], + "page_idx": 10 + }, + { + "type": "page_footnote", + "text": "The original GPC paper [23] calculates the mIoU differently compared to common practices [27, 31, 66]. The reported performance is based on the re-implementation from the official supported checkpoints of their GitHub, where more details are in Supplementary Sec. 1.", + "bbox": [ + 84, + 874, + 917, + 947 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/c9253c28f32cb00ce96b0da69eb8b1ab73b44b0ddf912c93723b345497593ca8.jpg", + "table_caption": [ + "Table 3: Ablation study in the nuScenes [4] and ScribbleKITTI [52] datasets via uniform sampling. IT2 represents our architecture, Ctrs. denotes our cross-distribution contrastive learning and Aug. is our representation-specific data augmentation. The yellow rows denote our single representation baseline with CPS [6] from [27]." + ], + "table_footnote": [], + "table_body": "
Repr.IT2 Ctrs. Aug.nuScenes [4]ScribbleKITTI [52]
10%20%50%10%20%50%
RangeXXX60.864.968.050.052.854.6
✓XX67.3 (6.5↑)70.8 (5.9↑)72.6 (4.6↑)53.2 (3.2↑)55.0 (2.2↑)56.2 (1.6↑)
✓✓X70.3 (9.5↑)72.6 (7.7↑)73.5 (5.5↑)54.9 (4.9↑)56.1 (3.3↑)57.2 (2.6↑)
✓✓✓71.3 (10.5↑)73.4 (8.5↑)74.0 (6.0↑)57.1 (7.1↑)57.3 (4.5↑)58.6 (4.0↑)
VoxelXXX66.370.072.551.853.954.8
✓XX70.4 (4.1↑)71.9 (1.9↑)73.1 (0.6↑)54.3 (2.5↑)55.6 (2.7↑)56.6 (1.8↑)
✓✓X71.6 (5.3↑)72.9 (2.9↑)73.9 (1.4↑)55.3 (3.5↑)56.5 (3.6↑)57.6 (2.8↑)
✓✓✓72.1 (5.8↑)73.6 (3.6↑)74.1 (1.6↑)56.7 (4.9↑)57.5 (4.6↑)58.3 (3.5↑)
", + "bbox": [ + 84, + 136, + 912, + 343 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d42650b2c6538733dd1a8509370f8d1a5b6f5ab41155658c657037dac6d51bb6.jpg", + "image_caption": [ + "Fig. 3: Ablation study of temperature value $\\tau$ from Eq. (8) in nuScenes [4] dataset under $10\\%$ labelled partition protocol." + ], + "image_footnote": [], + "bbox": [ + 75, + 354, + 491, + 495 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/c0645e36836fea1bb72101f8768681c20c3f2c29c038b62a288b1864e5b2a4fd.jpg", + "table_caption": [ + "Table 4: Contrastive Learning study in nuScenes [4], where ContrasSeg [54] is re-implemented from the SOTA image (pixel-level) semantic segmentation." + ], + "table_footnote": [], + "table_body": "
MethodRepr.nuScenes [4]
10%20%
ContrasSeg [54]range70.372.3
voxel71.272.8
Oursrange71.3 (1.0↑)73.4 (1.1↑)
voxel72.1 (0.9↑)73.6 (0.8↑)
", + "bbox": [ + 501, + 430, + 912, + 549 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Cross-distribution contrastive learning. We investigate the impact of the hyperparameter $\\tau$ from Eq. (8) in Fig. 3 using the $10\\%$ labelled protocol on the nuScenes [4] dataset. Our observations indicate that $\\tau = 0.10$ yields the best results, so we fix this value for all experiments. In Tab. 4, we compare our approach with our re-implementation of the label-guided pixel-wise contrastive learning [54] based on our IT2 architecture with different ratio of labelled data, where our method shows $\\approx 1\\%$ improvement for both representations, demonstrating the effectiveness of our proposed cross-distribution contrastive learning. Representation-specific augmentation. In Fig. 4, we have demonstrated the augmentation results for each representation in the ScribbleKITTI [52] dataset, where 'multi' denotes multi-box [63], and '1-inc' represents single inclination setups [27]. In our framework, one representation's results rely on the other's generalisation. For instance, '(multi) CutMix' reaches $57.1\\%$ mIoU with '(1-inc) LaserMix,' improving by $0.5\\%$ compared to the results with 'LaserMix.' Meanwhile, '(multi) CutMix' combined with '(1-inc) LaserMix' improves by $1.3\\%$ in the range and $0.7\\%$ in voxel representations compared to LaserMix [27], highlighting the improved generalisation benefits from our augmentation.", + "bbox": [ + 72, + 598, + 917, + 928 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation study for other representations. In Tab. 5, we conduct abla", + "bbox": [ + 75, + 928, + 917, + 947 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 157, + 14, + 280, + 30 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/cc11fcdc44998bfbbecc11735861c931911ebd5bc1dd6af6a279196d2897c507.jpg", + "image_caption": [ + "Table 5: Ablation study for the bird's eye view (BEV) experiment in the nuScenes [4], where results are from PolarNet [64]. The baseline relies on single view [6], followed by BEV with range image or voxel grid." + ], + "image_footnote": [], + "bbox": [ + 82, + 53, + 496, + 204 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b039b272353df8f458ace7117515cc0af069724c1fc9fa3c59c8941db585589c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Repr.BEV (PolarNet [64])
10%20%50%
sup.58.563.968.4
w/ bev62.4 (3.9↑)66.7 (2.8↑)70.0 (1.6↑)
w/ range64.8 (6.3↑)67.9 (4.0↑)70.6 (2.2↑)
w/ voxel66.3 (7.8↑)69.1 (5.2↑)71.6 (3.2↑)
", + "bbox": [ + 508, + 144, + 915, + 287 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d6f08f6ff4cb749c3191e7b450ced6b3186db14ad833561a843dea7fe1593797.jpg", + "image_caption": [ + "Fig. 4: Ablation study for the repr.-specific augmentations in ScribbleKITTI [52] with $10\\%$ labelled data, where 'multi' is multi-boxes and '1-inc' is single inclination." + ], + "image_footnote": [], + "bbox": [ + 82, + 323, + 290, + 458 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e9b95a5defc6d47b1d7e7eb8da00a20818242b9f1bce0bef7583071d11a5983a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 326, + 498, + 461 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2e3804f3d5fd133ae4ec4cb6f161a8b076a3c4aa9eddfec5db5d2db024a60314.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 326, + 704, + 458 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/423e5f59caf966f7b1a488cc4f83d10575d1bfad41049cef93bf65630d7f0309.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 326, + 915, + 458 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d20a87dd5527309d4d7128d0c548069c54da799d0b8ab2283d7ce32ced6a0b41.jpg", + "image_caption": [ + "(a) Ground truth" + ], + "image_footnote": [], + "bbox": [ + 82, + 461, + 288, + 648 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/49b013b85c98998b551c2c75e72dc8da8e7e0c74329fa5869ffa3d211ca185d5.jpg", + "image_caption": [ + "(b) MT [50]" + ], + "image_footnote": [], + "bbox": [ + 292, + 463, + 496, + 648 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/947c3dec638db299244e1dd7a020717d624d6d8a2baf0785bcbaee4f36e2f722.jpg", + "image_caption": [ + "(c) LaserMix [27]" + ], + "image_footnote": [], + "bbox": [ + 498, + 463, + 704, + 648 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d7e4b2a21f93eddcef5173424ac2b33b630d2f9333bca2849037336b8bb148c5.jpg", + "image_caption": [ + "(d) Ours", + "Fig. 5: Error maps visualised from LiDAR points (top) and range view (bottom) in 2 examples of the nuScenes [4] dataset. The correct predictions are in green and the mistakes are highlighted in red." + ], + "image_footnote": [], + "bbox": [ + 707, + 463, + 912, + 648 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "tion studies with an additional representation, the Bird's Eye View (BEV), on the nuScenes [4] dataset, where all the results in the table are based on PolarNet [64]. We re-implement the supervised results and single representation supervision [6] in the first and second rows, respectively. We emphasize that employing peer representations can significantly enhance performance compared to the single-representation baseline [6]. For instance, under the $10\\%$ labelled partition protocol, employing BEV (based on PolarNet [64]) with the range image (based on FidNet [65]) and with the voxel grid (based on Cylinder3D [66]) lead to improvements of $2.4\\%$ and $3.9\\%$ in mIoU, respectively.", + "bbox": [ + 72, + 780, + 920, + 959 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ItTakesTwo", + "bbox": [ + 721, + 16, + 837, + 30 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 893, + 16, + 917, + 29 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a7d369ee918029812c70d2e104eba80df1a04ad4510cf4db1bd8a56f9479c05a.jpg", + "image_caption": [ + "(a) w/o Contras." + ], + "image_footnote": [], + "bbox": [ + 82, + 56, + 355, + 193 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5c9d838b19176cfbe93572eaa2d87dad574f734e7ff4c8e08512e61c21729b25.jpg", + "image_caption": [ + "(b) ContrasSeg [54]", + "Fig. 6: T-SNE visualisation for the pixel- (and voxel-) wise latent embedding features in the nuScenes [4] dataset. The features $\\mathcal{V}$ and $\\mathcal{O}$ indicate range and voxel representations and they are colored according to the category label." + ], + "image_footnote": [], + "bbox": [ + 360, + 56, + 636, + 193 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/430e59adc6899344eadebe4b0138047bc33a9cfe5739ffe63f41b9e7bec59766.jpg", + "image_caption": [ + "(c) Ours" + ], + "image_footnote": [], + "bbox": [ + 639, + 56, + 915, + 193 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.4 Visualisation", + "text_level": 1, + "bbox": [ + 75, + 312, + 302, + 330 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Visualisation of prediction error maps. We compare the error maps from predictions in Fig. 5, including Mean Teacher [50] (2nd column), LaserMix [27] (3rd column) and ours (last column). The visualisations show the bird's eye view's output in the first row, followed by the range view results in the second row. Notably, our method's prediction demonstrates a superior qualitative result, exhibiting fewer mistaken predictions (highlighted in red) and more correct predictions (marked in green) for both views.", + "bbox": [ + 72, + 344, + 920, + 479 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "T-SNE visualisation of the latent features. Figure 6 displays the latent embedding distribution from the nuScenes [4] dataset. Fig. 6a shows the results without contrastive learning, Fig. 6b displays the results based on the label-guided contrastive learning [54], and our prototypical contrastive learning is in Fig. 6c. The samples denoted by $\\text{和}$ and $\\odot$ indicate the latent embeddings from range and voxel models, respectively, where different colours represent the classes. Fig. 6a displays an unstructured latent representation due to the absence of contrastive learning, but Fig. 6b successfully clusters class embeddings from the same modalities (e.g., $\\{\\text{和}, \\text{和}\\}$ and $\\{\\text{和}, \\text{和}\\}$ ), yet it fails to capture the semantic relationships between different modalities within the same category. Our method, depicted in Fig. 6c, successfully clusters embeddings from the same class and different modalities $\\{\\text{和}, \\text{和}\\}$ , enabling a more effective distribution of embeddings for subsequent segmentation.", + "bbox": [ + 72, + 481, + 920, + 734 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 75, + 752, + 288, + 772 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this study, we introduced the novel semi-supervised LiDAR semantic segmentation method ItTakesTwo (IT2). IT2 improves generalisation with consistent predictions from peer LiDAR representations, and it also introduces a more effective contrastive learning using informative embeddings sampled from a learned distribution of positives and negatives. Finally, through experimental studies, we identify an efficient way to apply data augmentation to each representation. Our approach achieves best results across various benchmarks under all partition protocols, demonstrating its effectiveness.", + "bbox": [ + 72, + 791, + 917, + 949 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 157, + 14, + 280, + 30 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. The project is supported by the Australian Research Council (ARC) through grant FT190100525.", + "bbox": [ + 75, + 55, + 920, + 95 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 77, + 121, + 237, + 141 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Arazo, E., Ortego, D., Albert, P., O'Connor, N.E., McGuinness, K.: Pseudolabeling and confirmation bias in deep semi-supervised learning. In: 2020 International Joint Conference on Neural Networks (IJCNN). pp. 1-8. IEEE (2020)", + "2. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9297-9307 (2019)", + "3. Berman, M., Triki, A.R., Blaschko, M.B.: The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4413-4421 (2018)", + "4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)", + "5. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems 33, 9912-9924 (2020)", + "6. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2613-2622 (2021)", + "7. Cheng, M., Hui, L., Xie, J., Yang, J.: Sspc-net: Semi-supervised semantic 3d point cloud segmentation network. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 1140-1147 (2021)", + "8. Cheung, Y.m.: A rival penalized em algorithm towards maximizing weighted likelihood for density mixture clustering with automatic model selection. In: Proceedings of the 17th International Conference on Pattern Recognition, 2004. ICPR 2004. vol. 4, pp. 633-636. IEEE (2004)", + "9. Cheung, Y.m.: Maximum weighted likelihood via rival penalized em for density mixture clustering with automatic model selection. IEEE Transactions on Knowledge and Data Engineering 17(6), 750-761 (2005)", + "10. Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d u-net: learning dense volumetric segmentation from sparse annotation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part II 19. pp. 424-432. Springer (2016)", + "11. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological) 39(1), 1-22 (1977)", + "12. Deng, S., Dong, Q., Liu, B., Hu, Z.: Superpoint-guided semi-supervised semantic segmentation of 3d point clouds. In: 2022 International conference on robotics and automation (ICRA). pp. 9214-9220. IEEE (2022)", + "13. Fan, L., Xiong, X., Wang, F., Wang, N., Zhang, Z.: Rangedet: In defense of range view for lidar-based 3d object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2918-2927 (2021)" + ], + "bbox": [ + 79, + 157, + 920, + 949 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "ItTakesTwo", + "bbox": [ + 721, + 16, + 837, + 30 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 893, + 16, + 917, + 30 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. French, G., Aila, T., Laine, S., Mackiewicz, M., Finlayson, G.: Semi-supervised semantic segmentation needs strong, high-dimensional perturbations (2019)", + "15. Graham, B., Engelcke, M., Van Der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9224-9232 (2018)", + "16. Grigorescu, S., Trasnea, B., Cocias, T., Macesanu, G.: A survey of deep learning techniques for autonomous driving. Journal of Field Robotics 37(3), 362-386 (2020)", + "17. Guo, Y., Wang, H., Hu, Q., Liu, H., Liu, L., Bennamoun, M.: Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence 43(12), 4338-4364 (2020)", + "18. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)", + "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "20. Hou, J., Graham, B., Nießner, M., Xie, S.: Exploring data-efficient 3d scene understanding with contrastive scene contexts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15587-15597 (2021)", + "21. Hou, Y., Zhu, X., Ma, Y., Loy, C.C., Li, Y.: Point-to-voxel knowledge distillation for lidar semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8479-8488 (2022)", + "22. Hung, W.C., Tsai, Y.H., Liou, Y.T., Lin, Y.Y., Yang, M.H.: Adversarial learning for semi-supervised semantic segmentation. arXiv preprint arXiv:1802.07934 (2018)", + "23. Jiang, L., Shi, S., Tian, Z., Lai, X., Liu, S., Fu, C.W., Jia, J.: Guided point contrastive learning for semi-supervised point cloud semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6423-6432 (2021)", + "24. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in Neural Information Processing Systems 33, 18661-18673 (2020)", + "25. Kohli, A.P.S., Sitzmann, V., Wetzstein, G.: Semantic implicit neural scene representations with semi-supervised training. In: 2020 International Conference on 3D Vision (3DV). pp. 423-433. IEEE (2020)", + "26. Kong, L., Liu, Y., Chen, R., Ma, Y., Zhu, X., Li, Y., Hou, Y., Qiao, Y., Liu, Z.: Rethinking range view representation for lidar segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 228-240 (2023)", + "27. Kong, L., Ren, J., Pan, L., Liu, Z.: Lasermix for semi-supervised lidar semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21705-21715 (2023)", + "28. Lai, X., Chen, Y., Lu, F., Liu, J., Jia, J.: Spherical transformer for lidar-based 3d recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17545-17555 (2023)", + "29. Le-Khac, P.H., Healy, G., Smeaton, A.F.: Contrastive representation learning: A framework and review. IEEE Access 8, 193907-193934 (2020)", + "30. Li, J., Zhou, P., Xiong, C., Hoi, S.C.: Prototypical contrastive learning of unsupervised representations. arXiv preprint arXiv:2005.04966 (2020)", + "31. Li, L., Shum, H.P., Breckon, T.P.: Less is more: Reducing task and model complexity for 3d point cloud semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9361-9371 (2023)" + ], + "bbox": [ + 77, + 56, + 917, + 947 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 157, + 14, + 280, + 30 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Li, M., Xie, Y., Shen, Y., Ke, B., Qiao, R., Ren, B., Lin, S., Ma, L.: Hybridcr: Weakly-supervised 3d point cloud semantic segmentation via hybrid contrastive regularization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14930-14939 (2022)", + "33. Liang, C., Wang, W., Miao, J., Yang, Y.: Gmmseg: Gaussian mixture based generative semantic segmentation models. Advances in Neural Information Processing Systems 35, 31360-31375 (2022)", + "34. Liang, T., Xie, H., Yu, K., Xia, Z., Lin, Z., Wang, Y., Tang, T., Wang, B., Tang, Z.: Bevfusion: A simple and robust lidar-camera fusion framework. Advances in Neural Information Processing Systems 35, 10421-10434 (2022)", + "35. Liu, M., Zhou, Y., Qi, C.R., Gong, B., Su, H., Anguelov, D.: Less: Label-efficient semantic segmentation for lidar point clouds. In: European Conference on Computer Vision. pp. 70-89. Springer (2022)", + "36. Liu, W., Yue, X., Chen, Y., Denoeux, T.: Trusted multi-view deep learning with opinion aggregation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 7585-7593 (2022)", + "37. Liu, Y., Hu, Q., Lei, Y., Xu, K., Li, J., Guo, Y.: Box2seg: Learning semantics of 3d point clouds with box-level supervision. arXiv preprint arXiv:2201.02963 (2022)", + "38. Liu, Y., Kong, L., Cen, J., Chen, R., Zhang, W., Pan, L., Chen, K., Liu, Z.: Segment any point cloud sequences by distilling vision foundation models. arXiv preprint arXiv:2306.09347 (2023)", + "39. Liu, Y., Tian, Y., Chen, Y., Liu, F., Belagiannis, V., Carneiro, G.: Perturbed and strict mean teachers for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4258-4267 (2022)", + "40. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: IEEE International Conference on Robotics and Automation (ICRA) (2023)", + "41. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2774-2781. IEEE (2023)", + "42. Manning, C.D.: An introduction to information retrieval. Cambridge university press (2009)", + "43. Mena, G., Nejatbakhsh, A., Varol, E., Niles-Weed, J.: Sinkhorn em: an expectation-maximization algorithm based on entropic optimal transport. arXiv preprint arXiv:2006.16548 (2020)", + "44. Miech, A., Alayrac, J.B., Smaira, L., Laptev, I., Sivic, J., Zisserman, A.: End-to-end learning of visual representations from uncurated instructional videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9879-9889 (2020)", + "45. Milioto, A., Vizzo, I., Behley, J., Stachniss, C.: Rangenet++: Fast and accurate lidar semantic segmentation. In: 2019 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 4213-4220. IEEE (2019)", + "46. Nunes, L., Marcuzzi, R., Chen, X., Behley, J., Stachniss, C.: Segcontrast: 3d point cloud feature representation learning through self-supervised segment discrimination. IEEE Robotics and Automation Letters 7(2), 2116-2123 (2022)", + "47. Ouali, Y., Hudelot, C., Tami, M.: Semi-supervised semantic segmentation with cross-consistency training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12674-12684 (2020)" + ], + "bbox": [ + 77, + 56, + 917, + 947 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "ItTakesTwo", + "bbox": [ + 721, + 16, + 837, + 30 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 893, + 16, + 917, + 30 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Reichardt, L., Ebert, N., Wasenmüller, O.: 360deg from a single camera: A few-shot approach for lidar segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1075-1083 (2023)", + "49. Sun, C.Y., Yang, Y.Q., Guo, H.X., Wang, P.S., Tong, X., Liu, Y., Shum, H.Y.: Semi-supervised 3d shape segmentation with multilevel consistency and part substitution. Computational Visual Media 9(2), 229-247 (2023)", + "50. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. arXiv preprint arXiv:1703.01780 (2017)", + "51. Tian, Y., Krishnan, D., Isola, P.: Contrastive multiview coding. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16. pp. 776-794. Springer (2020)", + "52. Unal, O., Dai, D., Van Gool, L.: Scribble-supervised lidar semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2697-2707 (2022)", + "53. Vora, S., Radwan, N., Greff, K., Meyer, H., Genova, K., Sajjadi, M.S., Pot, E., Tagliasacchi, A., Duckworth, D.: Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes. arXiv preprint arXiv:2111.13260 (2021)", + "54. Wang, W., Zhou, T., Yu, F., Dai, J., Konukoglu, E., Van Gool, L.: Exploring cross-image pixel contrast for semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7303-7313 (2021)", + "55. Wu, B., Wan, A., Yue, X., Keutzer, K.: Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In: 2018 IEEE international conference on robotics and automation (ICRA). pp. 1887-1893. IEEE (2018)", + "56. Xie, S., Gu, J., Guo, D., Qi, C.R., Guibas, L., Litany, O.: Pointcontrast: Unsupervised pre-training for 3d point cloud understanding. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 574-591. Springer (2020)", + "57. Xu, J., Hsu, D.J., Maleki, A.: Benefits of over-parameterization with em. Advances in Neural Information Processing Systems 31 (2018)", + "58. Xu, J., Zhang, R., Dou, J., Zhu, Y., Sun, J., Pu, S.: Rpvnet: A deep and efficient range-point-voxel fusion network for lidar point cloud segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16024-16033 (2021)", + "59. Xu, J., Tang, H., Ren, Y., Peng, L., Zhu, X., He, L.: Multi-level feature learning for contrastive multi-view clustering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16051-16060 (2022)", + "60. Xu, Z., Yuan, B., Zhao, S., Zhang, Q., Gao, X.: Hierarchical point-based active learning for semi-supervised point cloud semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18098-18108 (2023)", + "61. Yang, L., Qi, L., Feng, L., Zhang, W., Shi, Y.: Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7236-7246 (2023)", + "62. Yang, L., Zhuo, W., Qi, L., Shi, Y., Gao, Y.: St++: Make self-training work better for semi-supervised semantic segmentation. arXiv preprint arXiv:2106.05095 (2021)", + "63. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6023-6032 (2019)" + ], + "bbox": [ + 79, + 56, + 917, + 947 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 79, + 16, + 104, + 29 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 157, + 14, + 280, + 30 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Zhang, Y., Zhou, Z., David, P., Yue, X., Xi, Z., Gong, B., Foroosh, H.: Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9601-9610 (2020)", + "65. Zhao, Y., Bai, L., Huang, X.: Fidnet: Lidar point cloud semantic segmentation with fully interpolation decoding. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4453-4458. IEEE (2021)", + "66. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9939-9948 (2021)", + "67. Zhuang, Z., Li, R., Jia, K., Wang, Q., Li, Y., Tan, M.: Perception-aware multi-sensor fusion for 3d lidar semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16280-16290 (2021)", + "68. Zou, Y., Yu, Z., Kumar, B., Wang, J.: Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In: Proceedings of the European conference on computer vision (ECCV). pp. 289-305 (2018)", + "69. Zou, Y., Zhang, Z., Zhang, H., Li, C.L., Bian, X., Huang, J.B., Pfister, T.: Pseudoseg: Designing pseudo labels for semantic segmentation. arXiv preprint arXiv:2010.09713 (2020)" + ], + "bbox": [ + 75, + 55, + 917, + 413 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "It Takes Two", + "bbox": [ + 721, + 16, + 837, + 30 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 893, + 16, + 917, + 30 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/ItTakesTwo_ Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation/69e10020-97d7-4fa9-96f3-57b936f4e871_model.json b/2024/ItTakesTwo_ Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation/69e10020-97d7-4fa9-96f3-57b936f4e871_model.json new file mode 100644 index 0000000000000000000000000000000000000000..37bd8e159a5ab53eea51790142c35d8bf7432e56 --- /dev/null +++ b/2024/ItTakesTwo_ Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation/69e10020-97d7-4fa9-96f3-57b936f4e871_model.json @@ -0,0 +1,2669 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.086, + 0.049, + 0.912, + 0.108 + ], + "angle": 0, + "content": "ItTakesTwo: Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.139, + 0.892, + 0.181 + ], + "angle": 0, + "content": "Yuyuan Liu\\*1(), Yuanhong Chen\\*1, Hu Wang2,1, Vasileios Belagiannis3, Ian Reid2,1, and Gustavo Carneiro4,1" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.195, + 0.896, + 0.214 + ], + "angle": 0, + "content": "\\(^{1}\\) Australian Institute for Machine Learning (AIML), University of Adelaide, AU" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.214, + 0.873, + 0.231 + ], + "angle": 0, + "content": "\\(^{2}\\) Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), UAE" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.232, + 0.733, + 0.249 + ], + "angle": 0, + "content": "3 University of Erlangen-Nuremberg, Germany" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.25, + 0.918, + 0.268 + ], + "angle": 0, + "content": "4 Centre for Vision, Speech and Signal Processing (CVSSP), University of Surrey, UK" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.304, + 0.858, + 0.626 + ], + "angle": 0, + "content": "Abstract. The costly and time-consuming annotation process to produce large training sets for modelling semantic LiDAR segmentation methods has motivated the development of semi-supervised learning (SSL) methods. However, such SSL approaches often concentrate on employing consistency learning only for individual LiDAR representations. This narrow focus results in limited perturbations that generally fail to enable effective consistency learning. Additionally, these SSL approaches employ contrastive learning based on the sampling from a limited set of positive and negative embedding samples. This paper introduces a novel semisupervised LiDAR semantic segmentation framework called ItTakesTwo (IT2). IT2 is designed to ensure consistent predictions from peer LiDAR representations, thereby improving the perturbation effectiveness in consistency learning. Furthermore, our contrastive learning employs informative samples drawn from a distribution of positive and negative embeddings learned from the entire training set. Results on public benchmarks show that our approach achieves remarkable improvements over the previous state-of-the-art (SOTA) methods in the field. The code is available at: https://github.com/yyliu01/IT2." + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.642, + 0.851, + 0.661 + ], + "angle": 0, + "content": "Keywords: Semi-supervised Learning \\(\\cdot\\) LiDAR Semantic Segmentation" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.688, + 0.31, + 0.709 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.725, + 0.922, + 0.92 + ], + "angle": 0, + "content": "Outdoor LiDAR semantic segmentation is emerging as an important area of research given its application to self-driving vehicles [16, 17]. Although LIDAR data has advantages relative to photometric camera data, it also presents unique challenges, such as data sparsity [66] and varying point density [27, 45]. Most existing works alleviate these issues by transforming the LiDAR point-clouds to other representations, including range-based image [13, 26, 45, 55], voxel grids [15, 21, 66], or fusing them with other modalities [38, 41] and views [58]. Despite their success, those methods rely on supervised learning that requires large labelled training sets, which can be problematic given the costly and time-consuming labelling process [39, 47]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.929, + 0.821, + 0.949 + ], + "angle": 0, + "content": "* denotes equal contribution. (☑) Email: yuyuan.liu@adelaide.edu.au." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.079, + 0.017, + 0.096, + 0.031 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.282, + 0.032 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.052, + 0.924, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.28, + 0.921, + 0.407 + ], + "angle": 0, + "content": "Fig. 1: The left figure represents the 3D point cloud scan (ground truth of four classes - Car, Pole, Terrain, and Sidewalk) while the right figures denote the prediction results (into the same four classes), provided by our model, from range (top-right) and voxel (bottom-right) representations, where the incorrect predictions are marked in gray. Please notice that: 1) the mapping of the four classes from the 3D point cloud scan to each representation demonstrates the consistent distribution of semantic classes, and 2) the histograms of the predictions for each representation exhibit distinct differences." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.431, + 0.92, + 0.587 + ], + "angle": 0, + "content": "These issues have motivated semi-supervised learning (SSL) innovations for LiDAR semantic segmentation [23,27,31], which require a small labelled dataset used in conjunction with a large unlabelled dataset. Consistency Learning, one of the main strategies explored in SSL for image semantic segmentation, penalises inconsistent predictions from perturbed inputs, where these perturbations take various forms, including input augmentation [62, 69], adversarial noise [22, 39], network perturbations [6] and dropout [61]. The incorporation of strong and varied perturbations is crucial for enhancing generalisation [14, 39]." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.587, + 0.922, + 0.903 + ], + "angle": 0, + "content": "Current research [27,31] applies such consistency learning in SSL for LiDAR points using many strategies, such as laser-beam augmentation [27] or Mean Teacher (MT) [31] framework. Despite their success, these approaches solely rely on a single LiDAR representation (e.g., voxel grid) for training, which could potentially limit the model's ability to generalize effectively due to the restricted perturbations. Such an issue can be mitigated with the use of multiple LiDAR representations for consistency learning that explores the fundamental Clustering assumption5, where the results from different representations naturally indicate the same underlying semantic information from the 3D LiDAR points. For example, as illustrated in Fig. 1, the class Terrain (▲) in the top-right pixels and bottom-right voxels should be \"clustered\" according to the (left) original 3D point data. Thus, the peer LiDAR representations can be regarded as a novel form similar to perturbation, but yet under-explored in the field. Other disadvantages of focusing on only one of the representations when using small labelled datasets are: the range view suffers from the information loss during the projection process, particularly for the tail categories, such as Pole (▲), which has" + }, + { + "type": "page_footnote", + "bbox": [ + 0.082, + 0.912, + 0.918, + 0.949 + ], + "angle": 0, + "content": "5 Cluster hypothesis (or assumption) [42]: the data of various forms that behave similarly with respect to information relevance should be clustered together." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.722, + 0.017, + 0.839, + 0.032 + ], + "angle": 0, + "content": "ItTakesTwo" + }, + { + "type": "page_number", + "bbox": [ + 0.904, + 0.017, + 0.919, + 0.031 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.056, + 0.919, + 0.177 + ], + "angle": 0, + "content": "a low mIoU of 52.02; and the voxel grid tends to show lower accuracy in dense areas that are far away from the LiDAR source, such as Sidewalk (\\(\\lambda\\)), which has a low mIoU of 69.50. Such limited pseudo label results will inevitably introduce confirmation bias (i.e., the problem of relying on and reinforcing wrong predictions during training), reducing even further the efficacy of consistency learning based on single representation [23, 27, 31]." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.179, + 0.92, + 0.429 + ], + "angle": 0, + "content": "Another effective technique for semi-supervised LiDAR segmentation is contrastive learning [7,20,23,31], which consists of attracting positive and repelling negative embedding pairs at instance [46] or category [23,31] levels, thereby enhancing the latent space point distribution. However, to generate these pairs, methods typically rely on a limited number of embeddings extracted from randomly selected training samples due to the high computational cost associated with dense contrastive learning tasks [23,31]. Unfortunately, this limited sample size hinders a comprehensive understanding of the distribution across the entire embedding space [5], resulting in suboptimal convergence of LiDAR segmentation in SSL. This limitation becomes particularly pronounced when dealing with embeddings from different models in multiple views, where potential noisy data [44, 51] or model bias [59] from one network can adversely affect others, thereby impeding the optimization process." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.431, + 0.92, + 0.721 + ], + "angle": 0, + "content": "In this work, we propose a novel semi-supervised LiDAR semantic segmentation framework, named ItTakesTwo (IT2). IT2 employs consistency learning based on peer-representations perturbation, which enforces the consistent predictions across two different representations. Such perturbation contributes with a more effective SSL generalisation and alleviates issues related to pseudo label qualities from a single representation. Furthermore, we propose cross-distribution contrastive learning, where instead of randomly selecting training samples in the embedding space, we learn informative embedding distributions (also known as prototype) via Gaussian Mixture Models from both representations. Then the embedding samples are drawn from such prototypical distributions to produce representative features that enhance the understanding of the latent space in peer representation scenarios. In addition, instead of applying a single type of data augmentation to both LiDAR representations [27], we empirically show that applying different augmentations to distinct representations provides better generalisation. To summarise, our work makes the following key contributions:" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.728, + 0.918, + 0.806 + ], + "angle": 0, + "content": "1. A novel SSL LiDAR segmentation framework (IT2) that leverages peer-representation perturbation in consistency learning, encouraging the same predictions for the points with the same semantic meaning in different LiDAR representations;" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.806, + 0.918, + 0.864 + ], + "angle": 0, + "content": "2. A new contrastive learning where informative positive and negative pairs of embeddings are sampled from prototype distributions (modeled via Gaussian Mixture Models) of both representations; and" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.864, + 0.918, + 0.901 + ], + "angle": 0, + "content": "3. An innovative representation-specific data augmentation that enables IT2 achieve better generalisation." + }, + { + "type": "list", + "bbox": [ + 0.086, + 0.728, + 0.918, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.91, + 0.946, + 0.949 + ], + "angle": 0, + "content": "Our IT2 has state-of-the-art (SOTA) performance on nuScenes [4], SemanticKitti [2] and ScribbleKitti [52], and we also achieve best performance with different" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.018, + 0.095, + 0.03 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.16, + 0.017, + 0.282, + 0.032 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.056, + 0.918, + 0.113 + ], + "angle": 0, + "content": "sampling strategies to form the labelled and unlabelled subsets. Notably, in nuScenes [4] dataset, we outperform the previous SOTA [27] by around \\(4.6\\%\\) in the limited labelled data situations." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.146, + 0.329, + 0.167 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.19, + 0.918, + 0.499 + ], + "angle": 0, + "content": "LiDAR Semantic Segmentation is an essential task for autonomous driving [17]. To more effectively extract the feature information from the unordered LiDAR point data, current approaches [45, 66, 67] transform such point data into different representations. Range representation methods [26, 45, 65] transfer the LiDAR points to the spherical coordinates [45] and project them to the image space [26, 65]. Voxel representation methods [15, 28, 66] voxelise the LiDAR points and project them into different sub-types of geometry views to alleviate the sparse data challenges [10, 15], encompassing cartesian [53], spherical [28] and cylindrical [66] coordinate systems. Furthermore, some methods [34, 38] take inputs from a combination of LiDAR representations [48, 67] or incorporate inputs from camera modalities [34, 40], fusing them in the intermediate layer to enhance feature understanding. Although these supervised methods show promising results [58, 65, 66], there has been a longstanding interest in minimising the extensive labeling efforts, with approaches such as scribble annotation [52], box supervision [37], active learning [35, 60], and SSL [23, 27]. In this paper, we focus on SSL to alleviate such annotation costs." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.503, + 0.918, + 0.851 + ], + "angle": 0, + "content": "Semi-supervised LiDAR Semantic Segmentation leverages unlabelled data to enhance the model's generalisation beyond the small labelled dataset. The current literature on LiDAR SSL primarily focuses on indoor [7, 12, 32, 60] or object [25, 49] point cloud scenarios, which have uniformly distributed and high density points in the scan. Such scenario is markedly different from the representation of outdoor LiDAR points that pose unique challenges, such as complex structure [27] and low density points [66]. Previous semi-supervised LiDAR segmentation methods [23, 31, 46] mainly rely on contrastive learning to improve the latent space understanding. For instance, GPC [23] utilises confidence-threshold to constrain potentially noisy pseudo labels, and Lim3D [31] uses a memory bank to save negative samples. To increase the variation of the LiDAR data, LaserMix [27] proposes a novel augmentation that mixes laser beams in various spatial positions and holds the consistency learning under the Mean Teacher (MT) architecture [50]. However, these methods use only single LiDAR representations which, as explained in Sec. 1, imposes many limitations, such as limited perturbation effectiveness and poor performance for specific categories. Our framework introduces a novel peer-representation perturbation that employs various LiDAR representations during training, resulting in improved generalisation." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.852, + 0.918, + 0.948 + ], + "angle": 0, + "content": "Contrastive Learning [5, 24, 54, 56] enforces latent embeddings to be closer to instances with the same semantic meaning (i.e. positive) and to be far away from instances with different semantic meanings (i.e. negative). This approach has been explored in LiDAR segmentation [7, 12, 23, 31, 46], but the reliance on actual training samples leads to sub-optimal training as only a subset of em" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.722, + 0.017, + 0.839, + 0.031 + ], + "angle": 0, + "content": "ItTakesTwo" + }, + { + "type": "page_number", + "bbox": [ + 0.904, + 0.017, + 0.919, + 0.03 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.053, + 0.924, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.302, + 0.922, + 0.446 + ], + "angle": 0, + "content": "Fig. 2: Illustration of our approach with unlabeled data. The point scans in the batch \\((\\mathbf{x}_1\\) and \\(\\mathbf{x}_2\\) on the left) are transformed into range (and processed by \\(f_{\\phi_{\\mathrm{FCN}}}(.))\\) and voxel (and processed by \\(f_{\\theta_{\\mathrm{EXR}}}(.))\\) representations. Subsequently, the results from \\(f_{\\phi_{\\mathrm{SEG}}}(f_{\\phi_{\\mathrm{FCN}}}(.))\\) and \\(f_{\\theta_{\\mathrm{SEG}}}(f_{\\theta_{\\mathrm{EXR}}}(.))\\) are transformed to the other representation with Eq. (2) and cross-supervised with Eq. (4). Also, the embedding \\(\\{\\cdot ,\\cdot \\}\\) selected by the pseudo labels are utilised to model the GMMs with Eq. (6) and penalised by the informative samples (generated from GMMs) based on \\(\\ell_{cross}(.)\\) in Eq. (8). Please note the arrow indicates that the operation does not require gradient computation." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.472, + 0.92, + 0.707 + ], + "angle": 0, + "content": "bedding pairs [23, 31] can be selected for computing the contrastive loss. This becomes more challenging when working with embeddings obtained from different representations [36, 51, 59], as potentially noisy predictions [44, 59] from each representation can result in confirmation bias during training [1]. Prototypical contrastive learning [5, 18, 29, 30, 33] explores the use of a set of representative embeddings for the contrastive learning via memory bank [18, 54], optimal transport [5], or deep clustering [29, 30]. Nevertheless, such prototypes represent the modes of a more complex underlying distribution of prototypes that has not been explored before for SSL LiDAR segmentation. In our framework, we employ Gaussian Mixture Models (GMMs) to represent class-specific embedding distributions from both LiDAR representations and to generate informative embeddings for all classes, enhancing the effectiveness of the contrastive learning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.739, + 0.319, + 0.762 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.788, + 0.922, + 0.95 + ], + "angle": 0, + "content": "Our approach is based on a labelled point dataset \\(\\mathcal{P}_L = \\{(\\mathbf{x}_i,\\mathbf{y}_i)\\}_{i = 1}^{| \\mathcal{P}_L |}\\) consisting of \\(|\\mathcal{P}_L|\\) scans and an unlabelled set \\(\\mathcal{P}_U = \\{(\\mathbf{x}_i\\}_{i = 1}^{| \\mathcal{P}_U|}\\) with \\(|\\mathcal{P}_U|\\) scans. Each scan \\(\\mathbf{x}_i\\in \\mathcal{P}\\subset \\mathbb{R}^{3\\times C\\times N}\\) has 3 spatial coordinates, with \\(C\\) denoting the number of features (intensity, elongation, etc.), \\(N\\) representing the number of points, and \\(\\mathbf{y}_i\\in \\mathcal{V}\\subset \\{0,1\\}^{Y\\times N}\\) represents the one-hot label with \\(Y\\) being the number of classes in the labelled set. This point cloud dataset is transformed into the range and voxel representations to build the two peer representation datasets that is required by our method. These transformations are defined by [58] as:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.078, + 0.017, + 0.096, + 0.031 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.282, + 0.032 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.056, + 0.921, + 0.137 + ], + "angle": 0, + "content": "- \\(\\psi_{p\\rightarrow v}:\\mathcal{P}\\rightarrow \\mathcal{V}\\) is a point to voxel projection function that forms \\(\\mathcal{V}_L = \\{(x_i^v,y_i^v)|(x_i,y_i)\\in \\mathcal{P}_L\\}_{i = 1}^{|\\mathcal{P}_L|}\\) and \\(\\mathcal{V}_U = \\{\\mathbf{x}_i^v |\\mathbf{x}_i\\in \\mathcal{P}_U\\}_{i = 1}^{|\\mathcal{P}_U|}\\), with each volume \\(\\mathbf{x}_i^v\\in \\mathcal{V}\\subset \\mathbb{R}^{C_v\\times H\\times W\\times L}\\) having size \\(H\\times W\\times L\\) and \\(C_v\\) features, and the projected label being \\(\\mathbf{y}_i^v\\in \\{0,1\\}^{Y\\times H\\times W\\times L}\\) in one-hot format." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.922, + 0.22 + ], + "angle": 0, + "content": "- \\(\\psi_{p\\rightarrow r}:\\mathcal{P}\\rightarrow \\mathcal{R}\\) is a function to project point to the range image6. With this, we build \\(\\mathcal{R}_L = \\{(\\mathbf{x}_i^r,\\mathbf{y}_i^r)|(x_i,y_i)\\in \\mathcal{P}_L\\}_{i = 1}^{|\\mathcal{P}_L|}\\) and \\(\\mathcal{R}_U = \\{\\mathbf{x}_i^r |\\mathbf{x}_i\\in \\mathcal{P}_U\\}_{i = 1}^{|\\mathcal{P}_U|}\\), with each range image \\(\\mathbf{x}_i^r\\in \\mathcal{R}\\subset \\mathbb{R}^{C_r\\times U\\times V}\\) having size \\(U\\times V\\) with \\(C_r\\) features, and the projected label being \\(\\mathbf{y}_i^r\\in \\{0,1\\}^{Y\\times U\\times V}\\)." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.056, + 0.922, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.229, + 0.924, + 0.39 + ], + "angle": 0, + "content": "The inverse of the transformations above are respectively denoted by \\(\\psi_{v\\rightarrow p}\\) and \\(\\psi_{r\\rightarrow p}\\). Based on the transformed range and voxel representations from \\(\\mathcal{P}_L\\) and \\(\\mathcal{P}_U\\), we define the labelled set \\(\\mathcal{D}_L = \\{(\\mathbf{x}_i^v,\\mathbf{y}_i^v),(\\mathbf{x}_i^r,\\mathbf{y}_i^r)|(\\mathbf{x}_i^v,\\mathbf{y}_i^v)\\in \\mathcal{V}_L,(\\mathbf{x}_i^r,\\mathbf{y}_i^r)\\in \\mathcal{R}_L,\\}_{i = 1}^{|\\mathcal{D}_L|}\\) and the larger unlabelled set \\(\\mathcal{D}_U = \\{(\\mathbf{x}_i^v,\\mathbf{x}_i^r)|\\mathbf{x}_i^v\\in \\mathcal{V}_U,\\mathbf{x}_i^r\\in \\mathcal{R}_U\\}_{i = 1}^{|\\mathcal{D}_U|}\\) where \\(|\\mathcal{D}_L|\\leq |\\mathcal{D}_U|\\). These datasets are used to train our IT2 methodology in Sec. 3.1, while Sec. 3.2 describes our Gaussian Mixture Models (GMMs) based prototypical contrastive learning. Lastly, we introduce the details of our augmentation method in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.407, + 0.506, + 0.427 + ], + "angle": 0, + "content": "3.1 ItTakesTwo (IT2) Framework" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.437, + 0.924, + 0.673 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, our approach enforces prediction consistency between two peer representations to explore a potentially more effective perturbation. Our method can use various LiDAR representations, but we follow [27] and employ range and voxel networks for a fair comparison in the benchmark. The range image network consists of an FCN Network (e.g., ResNet [19]) that maps the input to a latent dense feature space with \\( f_{\\phi_{\\mathbf{FCN}}} : \\mathcal{R} \\to \\mathcal{I} \\), where \\( \\mathcal{I} \\subset \\mathbb{R}^{U \\times V \\times D_r} \\) with \\( D_r \\) denoting feature dimensionality. Then a segmentation head will map \\( \\mathcal{I} \\) to the categorical segmentation map with \\( f_{\\phi_{\\mathbf{SEG}}} : \\mathcal{I} \\to \\mathbb{R}^{U \\times V \\times Y} \\). Similarly, the voxel network extracts features with \\( f_{\\theta_{\\mathbf{EXR}}} : \\mathcal{V} \\to \\mathcal{K} \\), where \\( \\mathcal{K} \\in \\mathbb{R}^{H \\times W \\times L \\times D_v} \\) with \\( D_v \\) denoting the feature dimensionality, followed by a segmentation head \\( f_{\\theta_{\\mathbf{SEG}}} : \\mathcal{K} \\to \\mathbb{R}^{H \\times W \\times L \\times Y} \\). The peer representation consistency training depends on the following categorical distributions for both views:" + }, + { + "type": "equation", + "bbox": [ + 0.275, + 0.682, + 0.917, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {y}} ^ {r} \\left(\\omega^ {r}\\right) = \\operatorname {s o f t m a x} ^ {\\left(\\omega^ {r}\\right)} \\left(f _ {\\phi_ {\\mathrm {S E G}}} \\left(f _ {\\phi_ {\\mathrm {F C N}}} \\left(\\mathbf {x} ^ {r}\\right)\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.279, + 0.711, + 0.719, + 0.733 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {y}} ^ {v} \\left(\\omega^ {v}\\right) = \\operatorname {s o f t m a x} \\left(\\omega^ {v}\\right) \\left(f _ {\\theta_ {\\mathrm {S E G}}} \\left(f _ {\\theta_ {\\mathrm {E X R}}} \\left(\\mathbf {x} ^ {v}\\right)\\right)\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.745, + 0.924, + 0.824 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{y}}^r\\) and \\(\\hat{\\mathbf{y}}^v\\) are the categorical distributions for the range image (in \\(\\omega^r\\) in the \\(U\\times V\\) lattice \\(\\Omega^r\\)) and voxel (in \\(\\omega^v\\) in the \\(H\\times W\\times L\\) lattice \\(\\Omega^v\\)) representations, and \\(\\mathrm{softmax}^{(\\omega)}(.)\\) computes the softmax activation at \\(\\omega\\). Based on these outputs, we perform the following cross-view transformations:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.833, + 0.917, + 0.867 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {y}} ^ {r} \\left(\\omega^ {r}\\right) = \\operatorname {a r g m a x} \\Psi_ {v \\rightarrow r} ^ {\\left(\\omega^ {r}\\right)} \\left(\\hat {\\mathbf {y}} ^ {v}\\right), \\quad \\mathbf {c} ^ {r} \\left(\\omega^ {r}\\right) = \\max \\Psi_ {v \\rightarrow r} ^ {\\left(\\omega^ {r}\\right)} \\left(\\hat {\\mathbf {y}} ^ {v}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.224, + 0.862, + 0.773, + 0.884 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {y}} ^ {v} \\left(\\omega^ {v}\\right) = \\operatorname {a r g m a x} \\Psi_ {r \\rightarrow v} ^ {\\left(\\omega^ {v}\\right)} \\left(\\hat {\\mathbf {y}} ^ {r}\\right), \\quad \\mathbf {c} ^ {v} \\left(\\omega^ {v}\\right) = \\max \\Psi_ {r \\rightarrow v} ^ {\\left(\\omega^ {v}\\right)} \\left(\\hat {\\mathbf {y}} ^ {r}\\right),\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.08, + 0.894, + 0.924, + 0.95 + ], + "angle": 0, + "content": "6 Range-to-point projection can lead to information loss [45], which is typically mitigated by post-processing with K-Nearest Neighbors (KNN). In pursuit of efficiency, our IT2 does not use any post-processing during training." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.722, + 0.016, + 0.84, + 0.032 + ], + "angle": 0, + "content": "ItTakesTwo" + }, + { + "type": "page_number", + "bbox": [ + 0.904, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.057, + 0.921, + 0.154 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{y}}^r\\) and \\(\\tilde{\\mathbf{y}}^v\\) represent the transformed hard pseudo label maps, \\(\\mathbf{c}^r\\) and \\(\\mathbf{c}^v\\) denote the transformed confidence maps for range and voxel representations, respectively, \\(\\varPsi_{v\\rightarrow r}=\\psi_{p\\rightarrow r}\\circ\\psi_{v\\rightarrow p},\\varPsi_{r\\rightarrow v}=\\psi_{p\\rightarrow v}\\circ\\psi_{r\\rightarrow p}\\), and the superscript \\(\\omega\\) indicates that we only consider the output at \\(\\omega\\). We use the cross-view pseudo labels from Eq. (2) to formulate the loss function to train IT2, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.168, + 0.917, + 0.206 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\mathrm {I T 2}} \\left(\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\phi_ {\\mathrm {F C N}}}, f _ {\\phi_ {\\mathrm {E X R}}}, f _ {\\phi_ {\\mathrm {S E G}}}\\right) = \\ell_ {\\text {r a n g e}} \\left(\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\phi_ {\\mathrm {F C N}}}, f _ {\\phi_ {\\mathrm {S E G}}}\\right) + \\ell_ {\\text {v o x e l}} \\left(\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\theta_ {\\mathrm {E X R}}}, f _ {\\theta_ {\\mathrm {S E G}}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.207, + 0.56, + 0.224 + ], + "angle": 0, + "content": "where we formulate the loss for each view as:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.236, + 0.917, + 0.439 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\ell_ {\\mathrm {r a n g e}} (\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\phi_ {\\mathbf {F C N}}}, f _ {\\phi_ {\\mathbf {S E G}}}) = \\frac {1}{| \\mathcal {D} _ {L} | \\times | \\Omega^ {r} |} \\sum_ {i = 1} ^ {| \\mathcal {D} _ {L} |} \\sum_ {\\omega^ {r} \\in \\Omega^ {r}} \\ell (\\mathbf {y} _ {i} ^ {r} (\\omega^ {r}), f _ {\\phi_ {\\mathbf {S E G}}} ^ {(\\omega^ {r})} (f _ {\\phi_ {\\mathbf {F C N}}} (\\mathbf {x} _ {i} ^ {r})) \\\\ + \\frac {1}{\\left| \\mathcal {D} _ {U} \\right| \\times \\left| \\Omega^ {r} \\right|} \\sum_ {i = 1} ^ {\\left| \\mathcal {D} _ {U} \\right|} \\sum_ {\\omega^ {r} \\in \\Omega^ {r}} \\ell \\left(\\tilde {\\mathbf {y}} _ {i} ^ {r} \\left(\\omega^ {r}\\right), f _ {\\phi_ {\\mathrm {S E G}}} ^ {\\left(\\omega^ {r}\\right)} \\left(f _ {\\phi_ {\\mathrm {F C N}}} \\left(\\mathbf {x} _ {i} ^ {r}\\right)\\right), \\right. \\tag {4} \\\\ \\ell_ {\\mathrm {v o x e l}} (\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\theta_ {\\mathbf {E X R}}}, f _ {\\theta_ {\\mathbf {S E G}}}) = \\frac {1}{| \\mathcal {D} _ {L} | \\times | \\Omega^ {v} |} \\sum_ {i = 1} ^ {| \\mathcal {D} _ {L} |} \\sum_ {\\omega^ {v} \\in \\Omega^ {v}} \\ell (\\mathbf {y} _ {i} ^ {v} (\\omega^ {v}), f _ {\\theta_ {\\mathbf {S E G}}} ^ {(\\omega^ {v})} (f _ {\\theta_ {\\mathbf {E X R}}} (\\mathbf {x} _ {i} ^ {v})) \\\\ + \\frac {1}{| \\mathcal {D} _ {U} | \\times | \\Omega^ {v} |} \\sum_ {i = 1} ^ {| \\mathcal {D} _ {U} |} \\sum_ {\\omega^ {v} \\in \\Omega^ {v}} \\ell (\\tilde {\\mathbf {y}} _ {i} ^ {v} (\\omega^ {v}), f _ {\\theta_ {\\mathrm {S E G}}} ^ {(\\omega^ {v})} (f _ {\\theta_ {\\mathrm {E X R}}} (\\mathbf {x} _ {i} ^ {v})), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.92, + 0.493 + ], + "angle": 0, + "content": "with the loss function \\(\\ell(.)\\) denoting the combination of cross-entropy loss and Lovasz-softmax [3] loss, following [27, 31, 66]." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.518, + 0.631, + 0.538 + ], + "angle": 0, + "content": "3.2 Cross-distribution Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.548, + 0.92, + 0.689 + ], + "angle": 0, + "content": "We model the cross distribution of the embedding space using a multi-layer perceptron to generate latent embeddings in parallel with the segmentation head. Specifically, we obtain the range image embedding space via \\( f_{\\phi_{\\mathbf{PRJ}}}^r: \\mathcal{I} \\to \\mathbb{R}^{U \\times V \\times Z} \\) with \\( Z \\) denoting the pixel-wise number of dimensions in the projected space. Similarly, the voxel embedding space is obtained via \\( f_{\\theta_{\\mathbf{PRJ}}}^v: \\mathcal{K} \\to \\mathbb{R}^{H \\times W \\times L \\times Z} \\), where \\( \\mathcal{I} \\) and \\( \\mathcal{K} \\) are intermediate feature spaces introduced in Sec. 3.1. We formulate the observed embedding results with:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.7, + 0.917, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} ^ {r} \\left(\\omega^ {r}\\right) = f _ {\\phi_ {\\mathbf {P R J}}} ^ {\\left(\\omega^ {r}\\right)} \\left(f _ {\\phi_ {\\mathbf {F C N}}} \\left(\\mathbf {x} ^ {r}\\right)\\right), \\quad \\mathbf {z} ^ {v} \\left(\\omega^ {v}\\right) = f _ {\\theta_ {\\mathbf {P R J}}} ^ {\\left(\\omega^ {v}\\right)} \\left(f _ {\\theta_ {\\mathbf {E X R}}} \\left(\\mathbf {x} ^ {v}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.736, + 0.919, + 0.773 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}^r\\) and \\(\\mathbf{z}^v\\) represent the latent point-wise embeddings of the range and voxel representations, respectively." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.775, + 0.922, + 0.951 + ], + "angle": 0, + "content": "As mentioned in Sec. 2, contrastive learning methods [7, 12, 23, 31, 46] only use a limited subset of actual training samples present in a mini batch, potentially leading to suboptimal convergence with confirmation bias. We propose a novel approach to represent the range and voxel embeddings with class-specific Gaussian Mixture Models (GMMs). Such approach enables the sampling of the most informative embeddings, fostering a comprehensive understanding of the distribution across the feature space of different representations. In practice, we concatenate the embeddings from both range and voxel based on their categories, where such embedding set is defined by \\(\\mathcal{G}^y = \\{(\\mathbf{z},\\mathbf{c},y)|\\mathbf{z} = \\mathbf{z}_i^r (\\omega^r),\\mathbf{c} =\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.079, + 0.017, + 0.096, + 0.03 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.282, + 0.032 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.056, + 0.921, + 0.136 + ], + "angle": 0, + "content": "\\(\\mathbf{c}_i^r (\\omega^r),y = \\tilde{\\mathbf{y}}_i^r (\\omega^r)\\} \\bigcup \\{(\\mathbf{z},\\mathbf{c},y)|\\mathbf{z} = \\mathbf{z}_i^v (\\omega^v),\\mathbf{c} = \\mathbf{c}_i^v (\\omega^v),y = \\tilde{\\mathbf{y}}_i^v (\\omega^v)\\}\\) for the unlabelled data indexed by \\(i\\in \\{1,\\dots,|\\mathcal{D}_U|\\}\\), with the related confidence (i.e., \\(\\mathbf{c}^r\\) and \\(\\mathbf{c}^v\\)) and pseudo labels (i.e., \\(\\tilde{\\mathbf{y}}^r\\) and \\(\\tilde{\\mathbf{y}}^v\\)) from Eq. (2) at positions \\(\\omega^r\\in \\Omega^r\\) and \\(\\omega^v\\in \\varOmega^v\\). The class-specific GMMs are defined for both representations by" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.143, + 0.92, + 0.19 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {\\Gamma^ {y}} (\\mathbf {z} | y) = \\sum_ {m = 1} ^ {M} \\mathbf {P} _ {\\Gamma^ {y}} (m | y) \\mathbf {P} _ {\\Gamma^ {y}} (\\mathbf {z} | m, y) = \\sum_ {m = 1} ^ {M} \\pi_ {m} ^ {y} \\mathcal {N} (\\mathbf {z}; \\boldsymbol {\\mu} _ {m} ^ {y}, \\boldsymbol {\\Sigma} _ {m} ^ {y}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.196, + 0.92, + 0.295 + ], + "angle": 0, + "content": "where \\(\\pmb{\\pi}_{m}^{y} = \\mathbf{P}_{\\varGamma^{y}}(m|y)\\) denotes the prior probability for the \\(m\\)-th GMM component of class \\(y \\in \\{0,1\\}^Y\\), \\(\\pmb{\\mu}_{m}^{y} \\in \\mathbb{R}^{Z}\\) and \\(\\pmb{\\Sigma}_{m}^{y} \\in \\mathbb{R}^{Z \\times Z}\\) represent the mean and covariance of the corresponding GMM component and we fix \\(M = 5\\) for all experiments. The GMM parameter optimisation in Eq. (6) relies on the EM [11] algorithm using the training set \\(\\mathcal{G}^{y}\\) defined above with the following steps:" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.298, + 0.854, + 0.345 + ], + "angle": 0, + "content": "E step: \\(\\mathbf{q}_{n,m}^{y} = \\frac{\\pi_{m}^{y}\\mathcal{N}(\\mathbf{z}_{n}\\mid\\boldsymbol{\\mu}_{m}^{y},\\boldsymbol{\\Sigma}_{m}^{y})}{\\sum_{m}^{M}\\pi_{m}^{y}\\mathcal{N}(\\mathbf{z}_{n}\\mid\\boldsymbol{\\mu}_{m}^{y},\\boldsymbol{\\Sigma}_{m}^{y})}\\), for \\((\\mathbf{z}_n,\\mathbf{c}_n,y_n)\\in \\mathcal{G}^y\\)" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.349, + 0.917, + 0.395 + ], + "angle": 0, + "content": "M step: \\(\\pmb{\\mu}_{m}^{y} = \\frac{1}{|\\mathcal{G}^{y}|}\\sum_{(\\mathbf{z}_{n},\\mathbf{c}_{n},y_{n})\\in \\mathcal{G}^{y}}\\mathbf{c}_{n}\\mathbf{q}_{n,m}^{y_{n}}\\mathbf{z}_{n}\\) (7)" + }, + { + "type": "equation", + "bbox": [ + 0.266, + 0.401, + 0.844, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol{\\Sigma}_{m}^{y} = \\frac{1}{|\\mathcal{G}^{y}|}\\sum_{(\\mathbf{z}_{n},\\mathbf{c}_{n},y_{n})\\in \\mathcal{G}^{y}}\\mathbf{c}_{n}\\mathbf{q}_{n,m}^{y_{n}}(\\mathbf{z}_{n} - \\boldsymbol{\\mu}_{m}^{y_{n}})\\bigl(\\mathbf{z}_{n} - \\boldsymbol{\\mu}_{m}^{y_{n}}\\bigr)^{\\intercal},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.454, + 0.92, + 0.611 + ], + "angle": 0, + "content": "where \\(\\mathcal{G}^y\\) represents the set of embedding samples that share the same label \\(y\\). The posterior \\(\\mathbf{q}_{n,m}^y\\) estimates the responsibility of the \\(n^{th}\\) embedding \\(\\mathbf{z}\\) from class \\(y\\) in \\(\\mathcal{G}^y\\) being assigned to the \\(m\\)-th component. The confidence \\(\\mathbf{c}_n\\) represents the potential correctness of the pseudo label from Eq. (2), which is used as a weighting factor to reduce the effect of low-confidence segmentation results [8,9] in the estimation of the GMM parameters. Note that we maintain a uniform weight of \\(\\pi_{m}^{y} = 1 / M\\) for each component [43, 57] to prevent one component from dominating others." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.629, + 0.92, + 0.883 + ], + "angle": 0, + "content": "GMM-based Prototypical Contrastive Learning. The class-specific GMMs from Eq. (7) are used to generate a set of highly representative embeddings for the contrastive learning process explained next. These representative class-specific embeddings is denoted by \\( S^y = \\{\\mathbf{s}|\\mathbf{s}\\sim \\mathbf{P}_{\\Gamma^y}(\\mathbf{z}|y)\\} \\), with \\( \\mathbf{P}_{\\Gamma^y}(\\mathbf{z}|y) \\) being the class-specific GMM defined in Eq. (6). Subsequently, these generated virtual prototypes are utilised to guide the contrastive learning for the incoming embedding features from both range and voxel representations [54], without joining the backpropagation computation graph. We define the set of incoming features with the anchor set \\( \\mathcal{A} = \\{(\\mathbf{z},y)\\} \\) that is formed similarly to how we form the set \\( \\mathcal{G}^y \\) without including the confidence value. The samples in \\( \\mathcal{A} \\) are selectively chosen for backpropagation [54], but the informative samples in \\( S^y \\) contribute to a comprehensive understanding of the cross-distribution for both representations. The contrastive learning relies on the minimisation of the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.081, + 0.89, + 0.917, + 0.949 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {c r o s s}} \\left(\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\phi_ {\\mathbf {P R J}}}, f _ {\\phi_ {\\mathbf {F C N}}}, f _ {\\theta_ {\\mathbf {P R J}}}, f _ {\\theta_ {\\mathbf {E X R}}} \\right) = \\sum_ {(\\mathbf {z}, y) \\in \\mathcal {A}} \\sum_ {\\mathbf {s} \\in \\mathcal {S} ^ {y}} \\frac {\\exp (\\mathbf {z} \\cdot \\mathbf {s} / \\tau)}{\\exp (\\mathbf {z} \\cdot \\mathbf {s} / \\tau) + \\sum_ {\\mathbf {s} ^ {-} \\in \\bar {\\mathcal {S}} ^ {y}} \\exp (\\mathbf {z} \\cdot \\mathbf {s} ^ {-} / \\tau)}, \\tag {8}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.721, + 0.016, + 0.84, + 0.032 + ], + "angle": 0, + "content": "ItTakesTwo" + }, + { + "type": "page_number", + "bbox": [ + 0.903, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.056, + 0.738, + 0.076 + ], + "angle": 0, + "content": "where \\(\\bar{S}^y\\) is the set of virtual negative embeddings for class \\(y\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.106, + 0.681, + 0.126 + ], + "angle": 0, + "content": "3.3 Representation Specific Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.14, + 0.923, + 0.337 + ], + "angle": 0, + "content": "Following [39, 61], we apply spatial data augmentations to the pseudo-labeling strategy in Eq. (2) to increase the robustness of our approach to complex scenes. After experimenting with classic data augmentation processes (e.g., CutMix [63] and LaserMix [27]), we notice that the application of distinct spatial augmentations to different viewpoints in our IT2 framework results in more effective generalization. We have empirically found the multi-boxes CutMix [63] augmentation for the range image and single-inclination LaserMix [27] augmentation for the voxel representation yield the most robust results. Although this representation-specific data augmentation is not our main contribution, it is shown empirically to produce relevant improvements. Please see details in supplementary Sec. 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.365, + 0.348, + 0.386 + ], + "angle": 0, + "content": "3.4 Overall Training" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.399, + 0.812, + 0.42 + ], + "angle": 0, + "content": "The overall training of our IT2 minimises the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.081, + 0.436, + 0.917, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} \\left(\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, \\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\phi_ {\\mathrm {F C N}}}, f _ {\\phi_ {\\mathrm {E X R}}}, f _ {\\phi_ {\\mathrm {S E G}}} f _ {\\phi_ {\\mathrm {P R J}}}, f _ {\\phi_ {\\mathrm {P R J}}}\\right) = \\\\ \\ell_ {\\mathrm {I T 2}} \\left(\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\phi_ {\\mathbf {F C N}}}, f _ {\\phi_ {\\mathbf {E X R}}}, f _ {\\phi_ {\\mathbf {S E G}}}\\right) + \\ell_ {\\mathrm {c r o s s}} \\left(\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\phi_ {\\mathbf {P R J}}}, f _ {\\phi_ {\\mathbf {F C N}}}, f _ {\\theta_ {\\mathbf {P R J}}}, f _ {\\theta_ {\\mathbf {E X R}}}\\right), \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.499, + 0.92, + 0.539 + ], + "angle": 0, + "content": "where the entire training pipeline is performed in an end-to-end manner, where in each iteration, we minimise the loss from \\(\\ell_{\\mathrm{IT2}}(.)\\) and \\(\\ell_{\\mathrm{cross}}(.)\\) simultaneously." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.561, + 0.301, + 0.583 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.604, + 0.39, + 0.624 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.637, + 0.921, + 0.89 + ], + "angle": 0, + "content": "We evaluate our approach on three benchmarks, including nuScenes [4], SemanticKITTI [2] and ScribbleKITTI [52]. The nuScenes [4] dataset has 29,130 training scans and 6,019 validation scans. Each LiDAR scan contains 32 beams with \\([-30^{\\circ},10^{\\circ}]\\) vertical field of views (FOV). The dataset has 23 categories, but following previous works [27,66], we merge them into 16 for training and validation. The SemanticKITTI [2] and its extension ScribbleKITTI [52] are popular laser-based segmentation benchmarks, where the latter one is based on line-scribbles labelling to reduce the annotation effort. The SemanticKITTI [2] dataset has 19 categories for semantic segmentation, where every LiDAR scan contains 64 beams with \\([3^{\\circ}, -25^{\\circ}]\\) vertical FOV. There are 22 sequences of LiDAR scans in the dataset, where sequences 0 to 9 are used for training and validation. Following [23,27], we use 19,130 scans (sequences 0 to 7 and sequence 9) for training and 4,071 scans from sequence 8 for validation." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.891, + 0.922, + 0.951 + ], + "angle": 0, + "content": "Sampling strategies. Previous works [23, 27, 31] propose different sampling strategies to select the labelled data for the semi-supervised learning setting, including: 1) uniform [27] sampling of the labelled data across all the training" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.032 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.282, + 0.032 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.079, + 0.053, + 0.92, + 0.108 + ], + "angle": 0, + "content": "Table 1: Uniform sampling results across various benchmarks, where we follow [27] to utilise half feature dimension and lower resolution in Cylinder3D [66] for all the experiments. Best results are highlighted in red." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.124, + 0.92, + 0.424 + ], + "angle": 0, + "content": "
Repr.MethodnuScenes [4]SemanticKITTI [2]ScribbleKITTI [52]
1%10%20%50%1%10%20%50%1%10%20%50%
Rangesup.38.357.562.767.636.252.255.957.233.147.749.952.5
MT [50]42.160.465.469.437.553.156.157.434.249.851.653.3
CBST [68]40.960.564.369.339.953.456.156.935.750.752.754.6
CPS [6]40.760.864.968.036.552.356.357.433.750.052.854.6
LaserMix [27]49.568.270.673.043.458.859.461.438.354.455.658.7
IT2 (Ours)56.571.373.474.051.960.361.762.146.657.157.358.6
Voxelsup.50.965.966.671.245.456.157.858.739.248.052.153.8
MT [50]51.666.067.171.745.457.159.260.041.050.152.853.9
CBST [68]53.066.569.671.648.858.359.459.741.550.653.354.5
CPS [6]52.966.370.072.546.758.759.660.541.451.853.954.8
LaserMix [27]55.369.971.873.250.660.061.962.344.253.755.156.8
IT2 (Ours)57.572.173.674.152.061.462.162.547.956.757.558.3
" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.458, + 0.92, + 0.575 + ], + "angle": 0, + "content": "sequences, 2) partial [23] sampling of the labelled data only from certain sequences, and 3) sampling of significant [31] labelled frames from the sequences. In the experiment, we examine our approach under all these sampling protocols. Metrics. We measure performance with mean intersection over union (mIoU). In each benchmark, we use the sampling strategies above to label \\( n\\% \\) of the training set, with the rest \\( (100 - n)\\% \\) being unlabelled." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.577, + 0.92, + 0.751 + ], + "angle": 0, + "content": "Implementation details. In the range view, we use the FIDNet architecture [65] with ResNet34 [19] backbone. The input range image has size \\( 32 \\times 1920 \\) for nuScenes [4] and \\( 64 \\times 2048 \\) for SemanticKITTI [2] and ScirbbleKITTI [52], following [27]. In the voxel representation, we use the Cylinder3D architecture [66], where we fix the intermediate layer at 16 dimensions with smaller resolution as specified in [27] for the uniform sampling experiments to enable a fair comparison. For contrastive learning, we build a 3-layer projector [24,54]. The projector outputs an embedding map with 64 channels, where we apply the temperature factor \\( \\tau = .1 \\) in Eq. (8). See supplementary Sec. 3 and Sec. 7 for more details." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.762, + 0.631, + 0.781 + ], + "angle": 0, + "content": "4.2 Results on Different Partition Protocols" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.793, + 0.92, + 0.949 + ], + "angle": 0, + "content": "Uniform sampling results. We first compare our method using the uniform sampling benchmark [27] in Tab. 1, with the experimental setup from [27] (i.e., architecture, split ID and the resolution of input representations). Our framework outperforms the previous SOTA Lasermix [27] by a large margin for almost all partition protocols. For example, we observe \\(\\approx 3.4\\%\\) and \\(\\approx 2.0\\%\\) improvements on average for range and voxel representations, respectively. Notably, we highlight that our approach yields significant improvements for the \\(1\\%\\) partition protocol in the line-scribble dataset [52], providing \\(8.28\\%\\) improvement in the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.722, + 0.016, + 0.839, + 0.032 + ], + "angle": 0, + "content": "ItTakesTwo" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.917, + 0.031 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.079, + 0.053, + 0.92, + 0.144 + ], + "angle": 0, + "content": "Table 2: Partial & significant sampling results in the SemanticKITTI [2] benchmark. We compare with previous SOTA based on Cylinder3D [66] architecture, where TTA denotes test time augmentation. \\(\\dagger\\) indicates that we re-implement results based on the supported checkpoints from their GitHub - please see Supplementary Sec. 1 for more details. Best results in each evaluation protocol are highlighted in red." + }, + { + "type": "table", + "bbox": [ + 0.128, + 0.16, + 0.879, + 0.297 + ], + "angle": 0, + "content": "
TTAMethodSemanticKITTI(par.)MethodSemanticKITTI(sig.)
5%10%20%40%5%10%20%40%
XGPC† [23]39.945.955.757.1lim3D [31]56.859.660.561.4
IT2 (Ours)43.448.156.560.9IT2 (Ours)60.363.364.064.8
✓GPC† [23]40.246.356.557.6lim3D [31]59.562.263.163.3
IT2 (Ours)43.849.258.862.4IT2 (Ours)61.764.464.765.8
" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.333, + 0.621, + 0.352 + ], + "angle": 0, + "content": "range image and \\(3.66\\%\\) in the voxel representation." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.354, + 0.919, + 0.547 + ], + "angle": 0, + "content": "Partial and significant sampling results. In Tab. 2, we compare our results with GPC [23] and lim3D [31] in partial sampling (par.) and significant sampling (sig.), based on both original performance and test time augmentation (TTA). In the partial sampling protocol7, our approach yields superior results across all the labelled data situations for both with and without TTA. Notably, our results surpass GPC [23] by \\(4.5\\%\\) and \\(3.6\\%\\) in the limited \\(5\\%\\) labelled data condition, demonstrating its good generalisation capability. In the significant sampling results, our method also achieves better performance than lim3D [31] for all the partition protocols. For example, our results are better than the previous SOTA by about \\(2.2\\%\\) and \\(1.6\\%\\) in \\(10\\%\\) and \\(20\\%\\) labelled data with TTA, respectively." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.562, + 0.331, + 0.582 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.594, + 0.919, + 0.865 + ], + "angle": 0, + "content": "Improvements from our contributions. Table 3 presents our ablation study on the nuScenes [4] and ScribbleKITTI [52] datasets across various partition protocols. We investigate the roles of the IT2 architecture, cross-distribution contrastive learning (Ctrs.), and representation-related data augmentation (Aug.). The baseline utilises network perturbation through single representations [6] from [27]. Our IT2 architecture results in improvements of approximately \\(4.1\\%\\) and \\(2.5\\%\\) for the voxel representation in \\(10\\%\\) labelled data for nuScenes [4] and ScribbleKITTI [52], respectively. The Ctrs. further improves by \\(\\approx 2.4\\%\\) and \\(\\approx 1.1\\%\\) for range and voxel representations in both datasets in \\(10\\%\\) labelled data. Such enhancements show the effectiveness of our cross-distribution prototypes and contrastive learning strategy. Furthermore, Aug. results in improvements of \\(2.2\\%\\) and \\(1.4\\%\\) for range and voxel representations in the ScribbleKITTI [52] dataset for \\(10\\%\\) labelled data, demonstrating the better generalisation brought by the representation-specific data augmentation." + }, + { + "type": "page_footnote", + "bbox": [ + 0.086, + 0.876, + 0.919, + 0.949 + ], + "angle": 0, + "content": "The original GPC paper [23] calculates the mIoU differently compared to common practices [27, 31, 66]. The reported performance is based on the re-implementation from the official supported checkpoints of their GitHub, where more details are in Supplementary Sec. 1." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.031 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.282, + 0.032 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.053, + 0.921, + 0.128 + ], + "angle": 0, + "content": "Table 3: Ablation study in the nuScenes [4] and ScribbleKITTI [52] datasets via uniform sampling. IT2 represents our architecture, Ctrs. denotes our cross-distribution contrastive learning and Aug. is our representation-specific data augmentation. The yellow rows denote our single representation baseline with CPS [6] from [27]." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.137, + 0.913, + 0.344 + ], + "angle": 0, + "content": "
Repr.IT2 Ctrs. Aug.nuScenes [4]ScribbleKITTI [52]
10%20%50%10%20%50%
RangeXXX60.864.968.050.052.854.6
✓XX67.3 (6.5↑)70.8 (5.9↑)72.6 (4.6↑)53.2 (3.2↑)55.0 (2.2↑)56.2 (1.6↑)
✓✓X70.3 (9.5↑)72.6 (7.7↑)73.5 (5.5↑)54.9 (4.9↑)56.1 (3.3↑)57.2 (2.6↑)
✓✓✓71.3 (10.5↑)73.4 (8.5↑)74.0 (6.0↑)57.1 (7.1↑)57.3 (4.5↑)58.6 (4.0↑)
VoxelXXX66.370.072.551.853.954.8
✓XX70.4 (4.1↑)71.9 (1.9↑)73.1 (0.6↑)54.3 (2.5↑)55.6 (2.7↑)56.6 (1.8↑)
✓✓X71.6 (5.3↑)72.9 (2.9↑)73.9 (1.4↑)55.3 (3.5↑)56.5 (3.6↑)57.6 (2.8↑)
✓✓✓72.1 (5.8↑)73.6 (3.6↑)74.1 (1.6↑)56.7 (4.9↑)57.5 (4.6↑)58.3 (3.5↑)
" + }, + { + "type": "image", + "bbox": [ + 0.076, + 0.355, + 0.492, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.501, + 0.495, + 0.556 + ], + "angle": 0, + "content": "Fig. 3: Ablation study of temperature value \\(\\tau\\) from Eq. (8) in nuScenes [4] dataset under \\(10\\%\\) labelled partition protocol." + }, + { + "type": "table_caption", + "bbox": [ + 0.495, + 0.357, + 0.913, + 0.429 + ], + "angle": 0, + "content": "Table 4: Contrastive Learning study in nuScenes [4], where ContrasSeg [54] is re-implemented from the SOTA image (pixel-level) semantic segmentation." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.431, + 0.913, + 0.551 + ], + "angle": 0, + "content": "
MethodRepr.nuScenes [4]
10%20%
ContrasSeg [54]range70.372.3
voxel71.272.8
Oursrange71.3 (1.0↑)73.4 (1.1↑)
voxel72.1 (0.9↑)73.6 (0.8↑)
" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.599, + 0.92, + 0.929 + ], + "angle": 0, + "content": "Cross-distribution contrastive learning. We investigate the impact of the hyperparameter \\(\\tau\\) from Eq. (8) in Fig. 3 using the \\(10\\%\\) labelled protocol on the nuScenes [4] dataset. Our observations indicate that \\(\\tau = 0.10\\) yields the best results, so we fix this value for all experiments. In Tab. 4, we compare our approach with our re-implementation of the label-guided pixel-wise contrastive learning [54] based on our IT2 architecture with different ratio of labelled data, where our method shows \\(\\approx 1\\%\\) improvement for both representations, demonstrating the effectiveness of our proposed cross-distribution contrastive learning. Representation-specific augmentation. In Fig. 4, we have demonstrated the augmentation results for each representation in the ScribbleKITTI [52] dataset, where 'multi' denotes multi-box [63], and '1-inc' represents single inclination setups [27]. In our framework, one representation's results rely on the other's generalisation. For instance, '(multi) CutMix' reaches \\(57.1\\%\\) mIoU with '(1-inc) LaserMix,' improving by \\(0.5\\%\\) compared to the results with 'LaserMix.' Meanwhile, '(multi) CutMix' combined with '(1-inc) LaserMix' improves by \\(1.3\\%\\) in the range and \\(0.7\\%\\) in voxel representations compared to LaserMix [27], highlighting the improved generalisation benefits from our augmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.929, + 0.919, + 0.949 + ], + "angle": 0, + "content": "Ablation study for other representations. In Tab. 5, we conduct abla" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.722, + 0.017, + 0.839, + 0.031 + ], + "angle": 0, + "content": "ItTakesTwo" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.919, + 0.03 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.054, + 0.497, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.221, + 0.5, + 0.292 + ], + "angle": 0, + "content": "Fig. 4: Ablation study for the repr.-specific augmentations in ScribbleKITTI [52] with \\(10\\%\\) labelled data, where 'multi' is multi-boxes and '1-inc' is single inclination." + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.054, + 0.925, + 0.145 + ], + "angle": 0, + "content": "Table 5: Ablation study for the bird's eye view (BEV) experiment in the nuScenes [4], where results are from PolarNet [64]. The baseline relies on single view [6], followed by BEV with range image or voxel grid." + }, + { + "type": "table", + "bbox": [ + 0.509, + 0.146, + 0.917, + 0.289 + ], + "angle": 0, + "content": "
Repr.BEV (PolarNet [64])
10%20%50%
sup.58.563.968.4
w/ bev62.4 (3.9↑)66.7 (2.8↑)70.0 (1.6↑)
w/ range64.8 (6.3↑)67.9 (4.0↑)70.6 (2.2↑)
w/ voxel66.3 (7.8↑)69.1 (5.2↑)71.6 (3.2↑)
" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.324, + 0.291, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.328, + 0.5, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.328, + 0.707, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.328, + 0.916, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.463, + 0.29, + 0.65 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.465, + 0.498, + 0.65 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.465, + 0.707, + 0.65 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.465, + 0.915, + 0.65 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.652, + 0.263, + 0.666 + ], + "angle": 0, + "content": "(a) Ground truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.652, + 0.448, + 0.667 + ], + "angle": 0, + "content": "(b) MT [50]" + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.652, + 0.678, + 0.667 + ], + "angle": 0, + "content": "(c) LaserMix [27]" + }, + { + "type": "image_caption", + "bbox": [ + 0.772, + 0.652, + 0.851, + 0.666 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.682, + 0.92, + 0.737 + ], + "angle": 0, + "content": "Fig. 5: Error maps visualised from LiDAR points (top) and range view (bottom) in 2 examples of the nuScenes [4] dataset. The correct predictions are in green and the mistakes are highlighted in red." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.782, + 0.922, + 0.96 + ], + "angle": 0, + "content": "tion studies with an additional representation, the Bird's Eye View (BEV), on the nuScenes [4] dataset, where all the results in the table are based on PolarNet [64]. We re-implement the supervised results and single representation supervision [6] in the first and second rows, respectively. We emphasize that employing peer representations can significantly enhance performance compared to the single-representation baseline [6]. For instance, under the \\(10\\%\\) labelled partition protocol, employing BEV (based on PolarNet [64]) with the range image (based on FidNet [65]) and with the voxel grid (based on Cylinder3D [66]) lead to improvements of \\(2.4\\%\\) and \\(3.9\\%\\) in mIoU, respectively." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.031 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.282, + 0.032 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.057, + 0.357, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.198, + 0.294, + 0.213 + ], + "angle": 0, + "content": "(a) w/o Contras." + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.057, + 0.637, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.198, + 0.585, + 0.214 + ], + "angle": 0, + "content": "(b) ContrasSeg [54]" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.057, + 0.917, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.199, + 0.817, + 0.213 + ], + "angle": 0, + "content": "(c) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.228, + 0.921, + 0.283 + ], + "angle": 0, + "content": "Fig. 6: T-SNE visualisation for the pixel- (and voxel-) wise latent embedding features in the nuScenes [4] dataset. The features \\(\\mathcal{V}\\) and \\(\\mathcal{O}\\) indicate range and voxel representations and they are colored according to the category label." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.313, + 0.304, + 0.331 + ], + "angle": 0, + "content": "4.4 Visualisation" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.346, + 0.921, + 0.481 + ], + "angle": 0, + "content": "Visualisation of prediction error maps. We compare the error maps from predictions in Fig. 5, including Mean Teacher [50] (2nd column), LaserMix [27] (3rd column) and ours (last column). The visualisations show the bird's eye view's output in the first row, followed by the range view results in the second row. Notably, our method's prediction demonstrates a superior qualitative result, exhibiting fewer mistaken predictions (highlighted in red) and more correct predictions (marked in green) for both views." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.482, + 0.922, + 0.735 + ], + "angle": 0, + "content": "T-SNE visualisation of the latent features. Figure 6 displays the latent embedding distribution from the nuScenes [4] dataset. Fig. 6a shows the results without contrastive learning, Fig. 6b displays the results based on the label-guided contrastive learning [54], and our prototypical contrastive learning is in Fig. 6c. The samples denoted by \\(\\text{和}\\) and \\(\\odot\\) indicate the latent embeddings from range and voxel models, respectively, where different colours represent the classes. Fig. 6a displays an unstructured latent representation due to the absence of contrastive learning, but Fig. 6b successfully clusters class embeddings from the same modalities (e.g., \\(\\{\\text{和}, \\text{和}\\}\\) and \\(\\{\\text{和}, \\text{和}\\}\\)), yet it fails to capture the semantic relationships between different modalities within the same category. Our method, depicted in Fig. 6c, successfully clusters embeddings from the same class and different modalities \\(\\{\\text{和}, \\text{和}\\}\\), enabling a more effective distribution of embeddings for subsequent segmentation." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.753, + 0.289, + 0.773 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.793, + 0.92, + 0.95 + ], + "angle": 0, + "content": "In this study, we introduced the novel semi-supervised LiDAR semantic segmentation method ItTakesTwo (IT2). IT2 improves generalisation with consistent predictions from peer LiDAR representations, and it also introduces a more effective contrastive learning using informative embeddings sampled from a learned distribution of positives and negatives. Finally, through experimental studies, we identify an efficient way to apply data augmentation to each representation. Our approach achieves best results across various benchmarks under all partition protocols, demonstrating its effectiveness." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.722, + 0.017, + 0.839, + 0.032 + ], + "angle": 0, + "content": "ItTakesTwo" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.919, + 0.031 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.056, + 0.921, + 0.096 + ], + "angle": 0, + "content": "Acknowledgements. The project is supported by the Australian Research Council (ARC) through grant FT190100525." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.122, + 0.238, + 0.143 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.158, + 0.918, + 0.211 + ], + "angle": 0, + "content": "1. Arazo, E., Ortego, D., Albert, P., O'Connor, N.E., McGuinness, K.: Pseudolabeling and confirmation bias in deep semi-supervised learning. In: 2020 International Joint Conference on Neural Networks (IJCNN). pp. 1-8. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.211, + 0.92, + 0.282 + ], + "angle": 0, + "content": "2. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9297-9307 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.282, + 0.921, + 0.351 + ], + "angle": 0, + "content": "3. Berman, M., Triki, A.R., Blaschko, M.B.: The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4413-4421 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.352, + 0.921, + 0.423 + ], + "angle": 0, + "content": "4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.423, + 0.92, + 0.475 + ], + "angle": 0, + "content": "5. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems 33, 9912-9924 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.476, + 0.92, + 0.528 + ], + "angle": 0, + "content": "6. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2613-2622 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.528, + 0.92, + 0.58 + ], + "angle": 0, + "content": "7. Cheng, M., Hui, L., Xie, J., Yang, J.: Sspc-net: Semi-supervised semantic 3d point cloud segmentation network. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 1140-1147 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.58, + 0.92, + 0.65 + ], + "angle": 0, + "content": "8. Cheung, Y.m.: A rival penalized em algorithm towards maximizing weighted likelihood for density mixture clustering with automatic model selection. In: Proceedings of the 17th International Conference on Pattern Recognition, 2004. ICPR 2004. vol. 4, pp. 633-636. IEEE (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.651, + 0.92, + 0.702 + ], + "angle": 0, + "content": "9. Cheung, Y.m.: Maximum weighted likelihood via rival penalized em for density mixture clustering with automatic model selection. IEEE Transactions on Knowledge and Data Engineering 17(6), 750-761 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.702, + 0.92, + 0.791 + ], + "angle": 0, + "content": "10. Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d u-net: learning dense volumetric segmentation from sparse annotation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part II 19. pp. 424-432. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.791, + 0.92, + 0.844 + ], + "angle": 0, + "content": "11. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological) 39(1), 1-22 (1977)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.844, + 0.92, + 0.896 + ], + "angle": 0, + "content": "12. Deng, S., Dong, Q., Liu, B., Hu, Z.: Superpoint-guided semi-supervised semantic segmentation of 3d point clouds. In: 2022 International conference on robotics and automation (ICRA). pp. 9214-9220. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.896, + 0.92, + 0.95 + ], + "angle": 0, + "content": "13. Fan, L., Xiong, X., Wang, F., Wang, N., Zhang, Z.: Rangedet: In defense of range view for lidar-based 3d object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2918-2927 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.158, + 0.921, + 0.95 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.106, + 0.031 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.282, + 0.032 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.057, + 0.918, + 0.093 + ], + "angle": 0, + "content": "14. French, G., Aila, T., Laine, S., Mackiewicz, M., Finlayson, G.: Semi-supervised semantic segmentation needs strong, high-dimensional perturbations (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.094, + 0.92, + 0.145 + ], + "angle": 0, + "content": "15. Graham, B., Engelcke, M., Van Der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9224-9232 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.146, + 0.92, + 0.197 + ], + "angle": 0, + "content": "16. Grigorescu, S., Trasnea, B., Cocias, T., Macesanu, G.: A survey of deep learning techniques for autonomous driving. Journal of Field Robotics 37(3), 362-386 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.198, + 0.918, + 0.25 + ], + "angle": 0, + "content": "17. Guo, Y., Wang, H., Hu, Q., Liu, H., Liu, L., Bennamoun, M.: Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence 43(12), 4338-4364 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.251, + 0.918, + 0.303 + ], + "angle": 0, + "content": "18. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.304, + 0.917, + 0.355 + ], + "angle": 0, + "content": "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.356, + 0.918, + 0.407 + ], + "angle": 0, + "content": "20. Hou, J., Graham, B., Nießner, M., Xie, S.: Exploring data-efficient 3d scene understanding with contrastive scene contexts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15587-15597 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.408, + 0.917, + 0.46 + ], + "angle": 0, + "content": "21. Hou, Y., Zhu, X., Ma, Y., Loy, C.C., Li, Y.: Point-to-voxel knowledge distillation for lidar semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8479-8488 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.46, + 0.917, + 0.495 + ], + "angle": 0, + "content": "22. Hung, W.C., Tsai, Y.H., Liou, Y.T., Lin, Y.Y., Yang, M.H.: Adversarial learning for semi-supervised semantic segmentation. arXiv preprint arXiv:1802.07934 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.495, + 0.918, + 0.564 + ], + "angle": 0, + "content": "23. Jiang, L., Shi, S., Tian, Z., Lai, X., Liu, S., Fu, C.W., Jia, J.: Guided point contrastive learning for semi-supervised point cloud semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6423-6432 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.565, + 0.918, + 0.617 + ], + "angle": 0, + "content": "24. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in Neural Information Processing Systems 33, 18661-18673 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.918, + 0.669 + ], + "angle": 0, + "content": "25. Kohli, A.P.S., Sitzmann, V., Wetzstein, G.: Semantic implicit neural scene representations with semi-supervised training. In: 2020 International Conference on 3D Vision (3DV). pp. 423-433. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.67, + 0.919, + 0.722 + ], + "angle": 0, + "content": "26. Kong, L., Liu, Y., Chen, R., Ma, Y., Zhu, X., Li, Y., Hou, Y., Qiao, Y., Liu, Z.: Rethinking range view representation for lidar segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 228-240 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.723, + 0.918, + 0.774 + ], + "angle": 0, + "content": "27. Kong, L., Ren, J., Pan, L., Liu, Z.: Lasermix for semi-supervised lidar semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21705-21715 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.918, + 0.827 + ], + "angle": 0, + "content": "28. Lai, X., Chen, Y., Lu, F., Liu, J., Jia, J.: Spherical transformer for lidar-based 3d recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17545-17555 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.828, + 0.917, + 0.861 + ], + "angle": 0, + "content": "29. Le-Khac, P.H., Healy, G., Smeaton, A.F.: Contrastive representation learning: A framework and review. IEEE Access 8, 193907-193934 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.862, + 0.917, + 0.896 + ], + "angle": 0, + "content": "30. Li, J., Zhou, P., Xiong, C., Hoi, S.C.: Prototypical contrastive learning of unsupervised representations. arXiv preprint arXiv:2005.04966 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.897, + 0.918, + 0.949 + ], + "angle": 0, + "content": "31. Li, L., Shum, H.P., Breckon, T.P.: Less is more: Reducing task and model complexity for 3d point cloud semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9361-9371 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.057, + 0.92, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.722, + 0.017, + 0.839, + 0.032 + ], + "angle": 0, + "content": "ItTakesTwo" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.057, + 0.92, + 0.129 + ], + "angle": 0, + "content": "32. Li, M., Xie, Y., Shen, Y., Ke, B., Qiao, R., Ren, B., Lin, S., Ma, L.: Hybridcr: Weakly-supervised 3d point cloud semantic segmentation via hybrid contrastive regularization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14930-14939 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.131, + 0.918, + 0.182 + ], + "angle": 0, + "content": "33. Liang, C., Wang, W., Miao, J., Yang, Y.: Gmmseg: Gaussian mixture based generative semantic segmentation models. Advances in Neural Information Processing Systems 35, 31360-31375 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.184, + 0.918, + 0.236 + ], + "angle": 0, + "content": "34. Liang, T., Xie, H., Yu, K., Xia, Z., Lin, Z., Wang, Y., Tang, T., Wang, B., Tang, Z.: Bevfusion: A simple and robust lidar-camera fusion framework. Advances in Neural Information Processing Systems 35, 10421-10434 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.237, + 0.918, + 0.289 + ], + "angle": 0, + "content": "35. Liu, M., Zhou, Y., Qi, C.R., Gong, B., Su, H., Anguelov, D.: Less: Label-efficient semantic segmentation for lidar point clouds. In: European Conference on Computer Vision. pp. 70-89. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.918, + 0.342 + ], + "angle": 0, + "content": "36. Liu, W., Yue, X., Chen, Y., Denoeux, T.: Trusted multi-view deep learning with opinion aggregation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 7585-7593 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.918, + 0.379 + ], + "angle": 0, + "content": "37. Liu, Y., Hu, Q., Lei, Y., Xu, K., Li, J., Guo, Y.: Box2seg: Learning semantics of 3d point clouds with box-level supervision. arXiv preprint arXiv:2201.02963 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.38, + 0.918, + 0.432 + ], + "angle": 0, + "content": "38. Liu, Y., Kong, L., Cen, J., Chen, R., Zhang, W., Pan, L., Chen, K., Liu, Z.: Segment any point cloud sequences by distilling vision foundation models. arXiv preprint arXiv:2306.09347 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.918, + 0.503 + ], + "angle": 0, + "content": "39. Liu, Y., Tian, Y., Chen, Y., Liu, F., Belagiannis, V., Carneiro, G.: Perturbed and strict mean teachers for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4258-4267 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.918, + 0.557 + ], + "angle": 0, + "content": "40. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: IEEE International Conference on Robotics and Automation (ICRA) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.558, + 0.918, + 0.628 + ], + "angle": 0, + "content": "41. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2774-2781. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.63, + 0.918, + 0.664 + ], + "angle": 0, + "content": "42. Manning, C.D.: An introduction to information retrieval. Cambridge university press (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.918, + 0.717 + ], + "angle": 0, + "content": "43. Mena, G., Nejatbakhsh, A., Varol, E., Niles-Weed, J.: Sinkhorn em: an expectation-maximization algorithm based on entropic optimal transport. arXiv preprint arXiv:2006.16548 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.918, + 0.789 + ], + "angle": 0, + "content": "44. Miech, A., Alayrac, J.B., Smaira, L., Laptev, I., Sivic, J., Zisserman, A.: End-to-end learning of visual representations from uncurated instructional videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9879-9889 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.918, + 0.842 + ], + "angle": 0, + "content": "45. Milioto, A., Vizzo, I., Behley, J., Stachniss, C.: Rangenet++: Fast and accurate lidar semantic segmentation. In: 2019 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 4213-4220. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.843, + 0.918, + 0.895 + ], + "angle": 0, + "content": "46. Nunes, L., Marcuzzi, R., Chen, X., Behley, J., Stachniss, C.: Segcontrast: 3d point cloud feature representation learning through self-supervised segment discrimination. IEEE Robotics and Automation Letters 7(2), 2116-2123 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.896, + 0.918, + 0.949 + ], + "angle": 0, + "content": "47. Ouali, Y., Hudelot, C., Tami, M.: Semi-supervised semantic segmentation with cross-consistency training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12674-12684 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.057, + 0.92, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.03 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.282, + 0.032 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.058, + 0.918, + 0.111 + ], + "angle": 0, + "content": "48. Reichardt, L., Ebert, N., Wasenmüller, O.: 360deg from a single camera: A few-shot approach for lidar segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1075-1083 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.112, + 0.918, + 0.164 + ], + "angle": 0, + "content": "49. Sun, C.Y., Yang, Y.Q., Guo, H.X., Wang, P.S., Tong, X., Liu, Y., Shum, H.Y.: Semi-supervised 3d shape segmentation with multilevel consistency and part substitution. Computational Visual Media 9(2), 229-247 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.918, + 0.218 + ], + "angle": 0, + "content": "50. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. arXiv preprint arXiv:1703.01780 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.918, + 0.272 + ], + "angle": 0, + "content": "51. Tian, Y., Krishnan, D., Isola, P.: Contrastive multiview coding. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16. pp. 776-794. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.272, + 0.918, + 0.325 + ], + "angle": 0, + "content": "52. Unal, O., Dai, D., Van Gool, L.: Scribble-supervised lidar semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2697-2707 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.325, + 0.918, + 0.378 + ], + "angle": 0, + "content": "53. Vora, S., Radwan, N., Greff, K., Meyer, H., Genova, K., Sajjadi, M.S., Pot, E., Tagliasacchi, A., Duckworth, D.: Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes. arXiv preprint arXiv:2111.13260 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.379, + 0.918, + 0.432 + ], + "angle": 0, + "content": "54. Wang, W., Zhou, T., Yu, F., Dai, J., Konukoglu, E., Van Gool, L.: Exploring cross-image pixel contrast for semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7303-7313 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.432, + 0.918, + 0.502 + ], + "angle": 0, + "content": "55. Wu, B., Wan, A., Yue, X., Keutzer, K.: Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In: 2018 IEEE international conference on robotics and automation (ICRA). pp. 1887-1893. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.918, + 0.576 + ], + "angle": 0, + "content": "56. Xie, S., Gu, J., Guo, D., Qi, C.R., Guibas, L., Litany, O.: Pointcontrast: Unsupervised pre-training for 3d point cloud understanding. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 574-591. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.918, + 0.61 + ], + "angle": 0, + "content": "57. Xu, J., Hsu, D.J., Maleki, A.: Benefits of over-parameterization with em. Advances in Neural Information Processing Systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.611, + 0.918, + 0.682 + ], + "angle": 0, + "content": "58. Xu, J., Zhang, R., Dou, J., Zhu, Y., Sun, J., Pu, S.: Rpvnet: A deep and efficient range-point-voxel fusion network for lidar point cloud segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16024-16033 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.682, + 0.918, + 0.735 + ], + "angle": 0, + "content": "59. Xu, J., Tang, H., Ren, Y., Peng, L., Zhu, X., He, L.: Multi-level feature learning for contrastive multi-view clustering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16051-16060 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.918, + 0.806 + ], + "angle": 0, + "content": "60. Xu, Z., Yuan, B., Zhao, S., Zhang, Q., Gao, X.: Hierarchical point-based active learning for semi-supervised point cloud semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18098-18108 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.806, + 0.918, + 0.859 + ], + "angle": 0, + "content": "61. Yang, L., Qi, L., Feng, L., Zhang, W., Shi, Y.: Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7236-7246 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.918, + 0.896 + ], + "angle": 0, + "content": "62. Yang, L., Zhuo, W., Qi, L., Shi, Y., Gao, Y.: St++: Make self-training work better for semi-supervised semantic segmentation. arXiv preprint arXiv:2106.05095 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.896, + 0.918, + 0.949 + ], + "angle": 0, + "content": "63. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6023-6032 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.058, + 0.918, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.722, + 0.017, + 0.839, + 0.032 + ], + "angle": 0, + "content": "It Takes Two" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.919, + 0.031 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.056, + 0.92, + 0.129 + ], + "angle": 0, + "content": "64. Zhang, Y., Zhou, Z., David, P., Yue, X., Xi, Z., Gong, B., Foroosh, H.: Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9601-9610 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.131, + 0.92, + 0.183 + ], + "angle": 0, + "content": "65. Zhao, Y., Bai, L., Huang, X.: Fidnet: Lidar point cloud semantic segmentation with fully interpolation decoding. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4453-4458. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.184, + 0.92, + 0.254 + ], + "angle": 0, + "content": "66. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9939-9948 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.255, + 0.918, + 0.307 + ], + "angle": 0, + "content": "67. Zhuang, Z., Li, R., Jia, K., Wang, Q., Li, Y., Tan, M.: Perception-aware multi-sensor fusion for 3d lidar semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16280-16290 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.308, + 0.918, + 0.36 + ], + "angle": 0, + "content": "68. Zou, Y., Yu, Z., Kumar, B., Wang, J.: Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In: Proceedings of the European conference on computer vision (ECCV). pp. 289-305 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.361, + 0.918, + 0.414 + ], + "angle": 0, + "content": "69. Zou, Y., Zhang, Z., Zhang, H., Li, C.L., Bian, X., Huang, J.B., Pfister, T.: Pseudoseg: Designing pseudo labels for semantic segmentation. arXiv preprint arXiv:2010.09713 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.056, + 0.92, + 0.414 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/ItTakesTwo_ Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation/69e10020-97d7-4fa9-96f3-57b936f4e871_origin.pdf b/2024/ItTakesTwo_ Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation/69e10020-97d7-4fa9-96f3-57b936f4e871_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..645f99db1cfde26dd1c50e58a6fe56d47836ff9d --- /dev/null +++ b/2024/ItTakesTwo_ Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation/69e10020-97d7-4fa9-96f3-57b936f4e871_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb27b2b56531a6416ef8365cf30a0f9dfc7be5d33ffa44cfd886ce082d7c59dd +size 8147783 diff --git a/2024/ItTakesTwo_ Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation/full.md b/2024/ItTakesTwo_ Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1dab244a9d98e10d1354755817e05adbe403b5a0 --- /dev/null +++ b/2024/ItTakesTwo_ Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation/full.md @@ -0,0 +1,324 @@ +# ItTakesTwo: Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation + +Yuyuan Liu\*1(), Yuanhong Chen\*1, Hu Wang2,1, Vasileios Belagiannis3, Ian Reid2,1, and Gustavo Carneiro4,1 + +$^{1}$ Australian Institute for Machine Learning (AIML), University of Adelaide, AU + +$^{2}$ Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), UAE + +3 University of Erlangen-Nuremberg, Germany + +4 Centre for Vision, Speech and Signal Processing (CVSSP), University of Surrey, UK + +Abstract. The costly and time-consuming annotation process to produce large training sets for modelling semantic LiDAR segmentation methods has motivated the development of semi-supervised learning (SSL) methods. However, such SSL approaches often concentrate on employing consistency learning only for individual LiDAR representations. This narrow focus results in limited perturbations that generally fail to enable effective consistency learning. Additionally, these SSL approaches employ contrastive learning based on the sampling from a limited set of positive and negative embedding samples. This paper introduces a novel semisupervised LiDAR semantic segmentation framework called ItTakesTwo (IT2). IT2 is designed to ensure consistent predictions from peer LiDAR representations, thereby improving the perturbation effectiveness in consistency learning. Furthermore, our contrastive learning employs informative samples drawn from a distribution of positive and negative embeddings learned from the entire training set. Results on public benchmarks show that our approach achieves remarkable improvements over the previous state-of-the-art (SOTA) methods in the field. The code is available at: https://github.com/yyliu01/IT2. + +Keywords: Semi-supervised Learning $\cdot$ LiDAR Semantic Segmentation + +# 1 Introduction + +Outdoor LiDAR semantic segmentation is emerging as an important area of research given its application to self-driving vehicles [16, 17]. Although LIDAR data has advantages relative to photometric camera data, it also presents unique challenges, such as data sparsity [66] and varying point density [27, 45]. Most existing works alleviate these issues by transforming the LiDAR point-clouds to other representations, including range-based image [13, 26, 45, 55], voxel grids [15, 21, 66], or fusing them with other modalities [38, 41] and views [58]. Despite their success, those methods rely on supervised learning that requires large labelled training sets, which can be problematic given the costly and time-consuming labelling process [39, 47]. + +![](images/a9315ad125de3c5b132dd037dbbdb275a82c73d44310d0125031ce5043815f67.jpg) +Fig. 1: The left figure represents the 3D point cloud scan (ground truth of four classes - Car, Pole, Terrain, and Sidewalk) while the right figures denote the prediction results (into the same four classes), provided by our model, from range (top-right) and voxel (bottom-right) representations, where the incorrect predictions are marked in gray. Please notice that: 1) the mapping of the four classes from the 3D point cloud scan to each representation demonstrates the consistent distribution of semantic classes, and 2) the histograms of the predictions for each representation exhibit distinct differences. + +These issues have motivated semi-supervised learning (SSL) innovations for LiDAR semantic segmentation [23,27,31], which require a small labelled dataset used in conjunction with a large unlabelled dataset. Consistency Learning, one of the main strategies explored in SSL for image semantic segmentation, penalises inconsistent predictions from perturbed inputs, where these perturbations take various forms, including input augmentation [62, 69], adversarial noise [22, 39], network perturbations [6] and dropout [61]. The incorporation of strong and varied perturbations is crucial for enhancing generalisation [14, 39]. + +Current research [27,31] applies such consistency learning in SSL for LiDAR points using many strategies, such as laser-beam augmentation [27] or Mean Teacher (MT) [31] framework. Despite their success, these approaches solely rely on a single LiDAR representation (e.g., voxel grid) for training, which could potentially limit the model's ability to generalize effectively due to the restricted perturbations. Such an issue can be mitigated with the use of multiple LiDAR representations for consistency learning that explores the fundamental Clustering assumption5, where the results from different representations naturally indicate the same underlying semantic information from the 3D LiDAR points. For example, as illustrated in Fig. 1, the class Terrain (▲) in the top-right pixels and bottom-right voxels should be "clustered" according to the (left) original 3D point data. Thus, the peer LiDAR representations can be regarded as a novel form similar to perturbation, but yet under-explored in the field. Other disadvantages of focusing on only one of the representations when using small labelled datasets are: the range view suffers from the information loss during the projection process, particularly for the tail categories, such as Pole (▲), which has + +a low mIoU of 52.02; and the voxel grid tends to show lower accuracy in dense areas that are far away from the LiDAR source, such as Sidewalk ( $\lambda$ ), which has a low mIoU of 69.50. Such limited pseudo label results will inevitably introduce confirmation bias (i.e., the problem of relying on and reinforcing wrong predictions during training), reducing even further the efficacy of consistency learning based on single representation [23, 27, 31]. + +Another effective technique for semi-supervised LiDAR segmentation is contrastive learning [7,20,23,31], which consists of attracting positive and repelling negative embedding pairs at instance [46] or category [23,31] levels, thereby enhancing the latent space point distribution. However, to generate these pairs, methods typically rely on a limited number of embeddings extracted from randomly selected training samples due to the high computational cost associated with dense contrastive learning tasks [23,31]. Unfortunately, this limited sample size hinders a comprehensive understanding of the distribution across the entire embedding space [5], resulting in suboptimal convergence of LiDAR segmentation in SSL. This limitation becomes particularly pronounced when dealing with embeddings from different models in multiple views, where potential noisy data [44, 51] or model bias [59] from one network can adversely affect others, thereby impeding the optimization process. + +In this work, we propose a novel semi-supervised LiDAR semantic segmentation framework, named ItTakesTwo (IT2). IT2 employs consistency learning based on peer-representations perturbation, which enforces the consistent predictions across two different representations. Such perturbation contributes with a more effective SSL generalisation and alleviates issues related to pseudo label qualities from a single representation. Furthermore, we propose cross-distribution contrastive learning, where instead of randomly selecting training samples in the embedding space, we learn informative embedding distributions (also known as prototype) via Gaussian Mixture Models from both representations. Then the embedding samples are drawn from such prototypical distributions to produce representative features that enhance the understanding of the latent space in peer representation scenarios. In addition, instead of applying a single type of data augmentation to both LiDAR representations [27], we empirically show that applying different augmentations to distinct representations provides better generalisation. To summarise, our work makes the following key contributions: + +1. A novel SSL LiDAR segmentation framework (IT2) that leverages peer-representation perturbation in consistency learning, encouraging the same predictions for the points with the same semantic meaning in different LiDAR representations; +2. A new contrastive learning where informative positive and negative pairs of embeddings are sampled from prototype distributions (modeled via Gaussian Mixture Models) of both representations; and +3. An innovative representation-specific data augmentation that enables IT2 achieve better generalisation. + +Our IT2 has state-of-the-art (SOTA) performance on nuScenes [4], SemanticKitti [2] and ScribbleKitti [52], and we also achieve best performance with different + +sampling strategies to form the labelled and unlabelled subsets. Notably, in nuScenes [4] dataset, we outperform the previous SOTA [27] by around $4.6\%$ in the limited labelled data situations. + +# 2 Related Work + +LiDAR Semantic Segmentation is an essential task for autonomous driving [17]. To more effectively extract the feature information from the unordered LiDAR point data, current approaches [45, 66, 67] transform such point data into different representations. Range representation methods [26, 45, 65] transfer the LiDAR points to the spherical coordinates [45] and project them to the image space [26, 65]. Voxel representation methods [15, 28, 66] voxelise the LiDAR points and project them into different sub-types of geometry views to alleviate the sparse data challenges [10, 15], encompassing cartesian [53], spherical [28] and cylindrical [66] coordinate systems. Furthermore, some methods [34, 38] take inputs from a combination of LiDAR representations [48, 67] or incorporate inputs from camera modalities [34, 40], fusing them in the intermediate layer to enhance feature understanding. Although these supervised methods show promising results [58, 65, 66], there has been a longstanding interest in minimising the extensive labeling efforts, with approaches such as scribble annotation [52], box supervision [37], active learning [35, 60], and SSL [23, 27]. In this paper, we focus on SSL to alleviate such annotation costs. + +Semi-supervised LiDAR Semantic Segmentation leverages unlabelled data to enhance the model's generalisation beyond the small labelled dataset. The current literature on LiDAR SSL primarily focuses on indoor [7, 12, 32, 60] or object [25, 49] point cloud scenarios, which have uniformly distributed and high density points in the scan. Such scenario is markedly different from the representation of outdoor LiDAR points that pose unique challenges, such as complex structure [27] and low density points [66]. Previous semi-supervised LiDAR segmentation methods [23, 31, 46] mainly rely on contrastive learning to improve the latent space understanding. For instance, GPC [23] utilises confidence-threshold to constrain potentially noisy pseudo labels, and Lim3D [31] uses a memory bank to save negative samples. To increase the variation of the LiDAR data, LaserMix [27] proposes a novel augmentation that mixes laser beams in various spatial positions and holds the consistency learning under the Mean Teacher (MT) architecture [50]. However, these methods use only single LiDAR representations which, as explained in Sec. 1, imposes many limitations, such as limited perturbation effectiveness and poor performance for specific categories. Our framework introduces a novel peer-representation perturbation that employs various LiDAR representations during training, resulting in improved generalisation. + +Contrastive Learning [5, 24, 54, 56] enforces latent embeddings to be closer to instances with the same semantic meaning (i.e. positive) and to be far away from instances with different semantic meanings (i.e. negative). This approach has been explored in LiDAR segmentation [7, 12, 23, 31, 46], but the reliance on actual training samples leads to sub-optimal training as only a subset of em + +![](images/d3f22719c0b6c5a2809055d3d6a1f0aba4dceb103fcdfcc6e5bca31ade8d09ab.jpg) +Fig. 2: Illustration of our approach with unlabeled data. The point scans in the batch $(\mathbf{x}_1$ and $\mathbf{x}_2$ on the left) are transformed into range (and processed by $f_{\phi_{\mathrm{FCN}}}(.))$ and voxel (and processed by $f_{\theta_{\mathrm{EXR}}}(.))$ representations. Subsequently, the results from $f_{\phi_{\mathrm{SEG}}}(f_{\phi_{\mathrm{FCN}}}(.))$ and $f_{\theta_{\mathrm{SEG}}}(f_{\theta_{\mathrm{EXR}}}(.))$ are transformed to the other representation with Eq. (2) and cross-supervised with Eq. (4). Also, the embedding $\{\cdot ,\cdot \}$ selected by the pseudo labels are utilised to model the GMMs with Eq. (6) and penalised by the informative samples (generated from GMMs) based on $\ell_{cross}(.)$ in Eq. (8). Please note the arrow indicates that the operation does not require gradient computation. + +bedding pairs [23, 31] can be selected for computing the contrastive loss. This becomes more challenging when working with embeddings obtained from different representations [36, 51, 59], as potentially noisy predictions [44, 59] from each representation can result in confirmation bias during training [1]. Prototypical contrastive learning [5, 18, 29, 30, 33] explores the use of a set of representative embeddings for the contrastive learning via memory bank [18, 54], optimal transport [5], or deep clustering [29, 30]. Nevertheless, such prototypes represent the modes of a more complex underlying distribution of prototypes that has not been explored before for SSL LiDAR segmentation. In our framework, we employ Gaussian Mixture Models (GMMs) to represent class-specific embedding distributions from both LiDAR representations and to generate informative embeddings for all classes, enhancing the effectiveness of the contrastive learning. + +# 3 Methodology + +Our approach is based on a labelled point dataset $\mathcal{P}_L = \{(\mathbf{x}_i,\mathbf{y}_i)\}_{i = 1}^{| \mathcal{P}_L |}$ consisting of $|\mathcal{P}_L|$ scans and an unlabelled set $\mathcal{P}_U = \{(\mathbf{x}_i\}_{i = 1}^{| \mathcal{P}_U|}$ with $|\mathcal{P}_U|$ scans. Each scan $\mathbf{x}_i\in \mathcal{P}\subset \mathbb{R}^{3\times C\times N}$ has 3 spatial coordinates, with $C$ denoting the number of features (intensity, elongation, etc.), $N$ representing the number of points, and $\mathbf{y}_i\in \mathcal{V}\subset \{0,1\}^{Y\times N}$ represents the one-hot label with $Y$ being the number of classes in the labelled set. This point cloud dataset is transformed into the range and voxel representations to build the two peer representation datasets that is required by our method. These transformations are defined by [58] as: + +- $\psi_{p\rightarrow v}:\mathcal{P}\rightarrow \mathcal{V}$ is a point to voxel projection function that forms $\mathcal{V}_L = \{(x_i^v,y_i^v)|(x_i,y_i)\in \mathcal{P}_L\}_{i = 1}^{|\mathcal{P}_L|}$ and $\mathcal{V}_U = \{\mathbf{x}_i^v |\mathbf{x}_i\in \mathcal{P}_U\}_{i = 1}^{|\mathcal{P}_U|}$ , with each volume $\mathbf{x}_i^v\in \mathcal{V}\subset \mathbb{R}^{C_v\times H\times W\times L}$ having size $H\times W\times L$ and $C_v$ features, and the projected label being $\mathbf{y}_i^v\in \{0,1\}^{Y\times H\times W\times L}$ in one-hot format. +- $\psi_{p\rightarrow r}:\mathcal{P}\rightarrow \mathcal{R}$ is a function to project point to the range image6. With this, we build $\mathcal{R}_L = \{(\mathbf{x}_i^r,\mathbf{y}_i^r)|(x_i,y_i)\in \mathcal{P}_L\}_{i = 1}^{|\mathcal{P}_L|}$ and $\mathcal{R}_U = \{\mathbf{x}_i^r |\mathbf{x}_i\in \mathcal{P}_U\}_{i = 1}^{|\mathcal{P}_U|}$ , with each range image $\mathbf{x}_i^r\in \mathcal{R}\subset \mathbb{R}^{C_r\times U\times V}$ having size $U\times V$ with $C_r$ features, and the projected label being $\mathbf{y}_i^r\in \{0,1\}^{Y\times U\times V}$ . + +The inverse of the transformations above are respectively denoted by $\psi_{v\rightarrow p}$ and $\psi_{r\rightarrow p}$ . Based on the transformed range and voxel representations from $\mathcal{P}_L$ and $\mathcal{P}_U$ , we define the labelled set $\mathcal{D}_L = \{(\mathbf{x}_i^v,\mathbf{y}_i^v),(\mathbf{x}_i^r,\mathbf{y}_i^r)|(\mathbf{x}_i^v,\mathbf{y}_i^v)\in \mathcal{V}_L,(\mathbf{x}_i^r,\mathbf{y}_i^r)\in \mathcal{R}_L,\}_{i = 1}^{|\mathcal{D}_L|}$ and the larger unlabelled set $\mathcal{D}_U = \{(\mathbf{x}_i^v,\mathbf{x}_i^r)|\mathbf{x}_i^v\in \mathcal{V}_U,\mathbf{x}_i^r\in \mathcal{R}_U\}_{i = 1}^{|\mathcal{D}_U|}$ where $|\mathcal{D}_L|\leq |\mathcal{D}_U|$ . These datasets are used to train our IT2 methodology in Sec. 3.1, while Sec. 3.2 describes our Gaussian Mixture Models (GMMs) based prototypical contrastive learning. Lastly, we introduce the details of our augmentation method in Sec. 3.3. + +# 3.1 ItTakesTwo (IT2) Framework + +As illustrated in Fig. 2, our approach enforces prediction consistency between two peer representations to explore a potentially more effective perturbation. Our method can use various LiDAR representations, but we follow [27] and employ range and voxel networks for a fair comparison in the benchmark. The range image network consists of an FCN Network (e.g., ResNet [19]) that maps the input to a latent dense feature space with $f_{\phi_{\mathbf{FCN}}} : \mathcal{R} \to \mathcal{I}$ , where $\mathcal{I} \subset \mathbb{R}^{U \times V \times D_r}$ with $D_r$ denoting feature dimensionality. Then a segmentation head will map $\mathcal{I}$ to the categorical segmentation map with $f_{\phi_{\mathbf{SEG}}} : \mathcal{I} \to \mathbb{R}^{U \times V \times Y}$ . Similarly, the voxel network extracts features with $f_{\theta_{\mathbf{EXR}}} : \mathcal{V} \to \mathcal{K}$ , where $\mathcal{K} \in \mathbb{R}^{H \times W \times L \times D_v}$ with $D_v$ denoting the feature dimensionality, followed by a segmentation head $f_{\theta_{\mathbf{SEG}}} : \mathcal{K} \to \mathbb{R}^{H \times W \times L \times Y}$ . The peer representation consistency training depends on the following categorical distributions for both views: + +$$ +\hat {\mathbf {y}} ^ {r} \left(\omega^ {r}\right) = \operatorname {s o f t m a x} ^ {\left(\omega^ {r}\right)} \left(f _ {\phi_ {\mathrm {S E G}}} \left(f _ {\phi_ {\mathrm {F C N}}} \left(\mathbf {x} ^ {r}\right)\right)\right), \tag {1} +$$ + +$$ +\hat {\mathbf {y}} ^ {v} \left(\omega^ {v}\right) = \operatorname {s o f t m a x} \left(\omega^ {v}\right) \left(f _ {\theta_ {\mathrm {S E G}}} \left(f _ {\theta_ {\mathrm {E X R}}} \left(\mathbf {x} ^ {v}\right)\right)\right), +$$ + +where $\hat{\mathbf{y}}^r$ and $\hat{\mathbf{y}}^v$ are the categorical distributions for the range image (in $\omega^r$ in the $U\times V$ lattice $\Omega^r$ ) and voxel (in $\omega^v$ in the $H\times W\times L$ lattice $\Omega^v$ ) representations, and $\mathrm{softmax}^{(\omega)}(.)$ computes the softmax activation at $\omega$ . Based on these outputs, we perform the following cross-view transformations: + +$$ +\tilde {\mathbf {y}} ^ {r} \left(\omega^ {r}\right) = \operatorname {a r g m a x} \Psi_ {v \rightarrow r} ^ {\left(\omega^ {r}\right)} \left(\hat {\mathbf {y}} ^ {v}\right), \quad \mathbf {c} ^ {r} \left(\omega^ {r}\right) = \max \Psi_ {v \rightarrow r} ^ {\left(\omega^ {r}\right)} \left(\hat {\mathbf {y}} ^ {v}\right), \tag {2} +$$ + +$$ +\tilde {\mathbf {y}} ^ {v} \left(\omega^ {v}\right) = \operatorname {a r g m a x} \Psi_ {r \rightarrow v} ^ {\left(\omega^ {v}\right)} \left(\hat {\mathbf {y}} ^ {r}\right), \quad \mathbf {c} ^ {v} \left(\omega^ {v}\right) = \max \Psi_ {r \rightarrow v} ^ {\left(\omega^ {v}\right)} \left(\hat {\mathbf {y}} ^ {r}\right), +$$ + +where $\tilde{\mathbf{y}}^r$ and $\tilde{\mathbf{y}}^v$ represent the transformed hard pseudo label maps, $\mathbf{c}^r$ and $\mathbf{c}^v$ denote the transformed confidence maps for range and voxel representations, respectively, $\varPsi_{v\rightarrow r}=\psi_{p\rightarrow r}\circ\psi_{v\rightarrow p},\varPsi_{r\rightarrow v}=\psi_{p\rightarrow v}\circ\psi_{r\rightarrow p}$ , and the superscript $\omega$ indicates that we only consider the output at $\omega$ . We use the cross-view pseudo labels from Eq. (2) to formulate the loss function to train IT2, as follows: + +$$ +\ell_ {\mathrm {I T 2}} \left(\mathcal {D} _ {L}, \mathcal {D} _ {U}, f _ {\phi_ {\mathrm {F C N}}}, f _ {\phi_ {\mathrm {E X R}}}, f _ {\phi_ {\mathrm {S E G}}}\right) = \ell_ {\text {r a n g e}} \left(\mathcal {D} _ {L}, \mathcal {D} _ {U}, f _ {\phi_ {\mathrm {F C N}}}, f _ {\phi_ {\mathrm {S E G}}}\right) + \ell_ {\text {v o x e l}} \left(\mathcal {D} _ {L}, \mathcal {D} _ {U}, f _ {\theta_ {\mathrm {E X R}}}, f _ {\theta_ {\mathrm {S E G}}}\right), \tag {3} +$$ + +where we formulate the loss for each view as: + +$$ +\begin{array}{l} \ell_ {\mathrm {r a n g e}} (\mathcal {D} _ {L}, \mathcal {D} _ {U}, f _ {\phi_ {\mathbf {F C N}}}, f _ {\phi_ {\mathbf {S E G}}}) = \frac {1}{| \mathcal {D} _ {L} | \times | \Omega^ {r} |} \sum_ {i = 1} ^ {| \mathcal {D} _ {L} |} \sum_ {\omega^ {r} \in \Omega^ {r}} \ell (\mathbf {y} _ {i} ^ {r} (\omega^ {r}), f _ {\phi_ {\mathbf {S E G}}} ^ {(\omega^ {r})} (f _ {\phi_ {\mathbf {F C N}}} (\mathbf {x} _ {i} ^ {r})) \\ + \frac {1}{\left| \mathcal {D} _ {U} \right| \times \left| \Omega^ {r} \right|} \sum_ {i = 1} ^ {\left| \mathcal {D} _ {U} \right|} \sum_ {\omega^ {r} \in \Omega^ {r}} \ell \left(\tilde {\mathbf {y}} _ {i} ^ {r} \left(\omega^ {r}\right), f _ {\phi_ {\mathrm {S E G}}} ^ {\left(\omega^ {r}\right)} \left(f _ {\phi_ {\mathrm {F C N}}} \left(\mathbf {x} _ {i} ^ {r}\right)\right), \right. \tag {4} \\ \ell_ {\mathrm {v o x e l}} (\mathcal {D} _ {L}, \mathcal {D} _ {U}, f _ {\theta_ {\mathbf {E X R}}}, f _ {\theta_ {\mathbf {S E G}}}) = \frac {1}{| \mathcal {D} _ {L} | \times | \Omega^ {v} |} \sum_ {i = 1} ^ {| \mathcal {D} _ {L} |} \sum_ {\omega^ {v} \in \Omega^ {v}} \ell (\mathbf {y} _ {i} ^ {v} (\omega^ {v}), f _ {\theta_ {\mathbf {S E G}}} ^ {(\omega^ {v})} (f _ {\theta_ {\mathbf {E X R}}} (\mathbf {x} _ {i} ^ {v})) \\ + \frac {1}{| \mathcal {D} _ {U} | \times | \Omega^ {v} |} \sum_ {i = 1} ^ {| \mathcal {D} _ {U} |} \sum_ {\omega^ {v} \in \Omega^ {v}} \ell (\tilde {\mathbf {y}} _ {i} ^ {v} (\omega^ {v}), f _ {\theta_ {\mathrm {S E G}}} ^ {(\omega^ {v})} (f _ {\theta_ {\mathrm {E X R}}} (\mathbf {x} _ {i} ^ {v})), \\ \end{array} +$$ + +with the loss function $\ell(.)$ denoting the combination of cross-entropy loss and Lovasz-softmax [3] loss, following [27, 31, 66]. + +# 3.2 Cross-distribution Contrastive Learning + +We model the cross distribution of the embedding space using a multi-layer perceptron to generate latent embeddings in parallel with the segmentation head. Specifically, we obtain the range image embedding space via $f_{\phi_{\mathbf{PRJ}}}^r: \mathcal{I} \to \mathbb{R}^{U \times V \times Z}$ with $Z$ denoting the pixel-wise number of dimensions in the projected space. Similarly, the voxel embedding space is obtained via $f_{\theta_{\mathbf{PRJ}}}^v: \mathcal{K} \to \mathbb{R}^{H \times W \times L \times Z}$ , where $\mathcal{I}$ and $\mathcal{K}$ are intermediate feature spaces introduced in Sec. 3.1. We formulate the observed embedding results with: + +$$ +\mathbf {z} ^ {r} \left(\omega^ {r}\right) = f _ {\phi_ {\mathbf {P R J}}} ^ {\left(\omega^ {r}\right)} \left(f _ {\phi_ {\mathbf {F C N}}} \left(\mathbf {x} ^ {r}\right)\right), \quad \mathbf {z} ^ {v} \left(\omega^ {v}\right) = f _ {\theta_ {\mathbf {P R J}}} ^ {\left(\omega^ {v}\right)} \left(f _ {\theta_ {\mathbf {E X R}}} \left(\mathbf {x} ^ {v}\right)\right), \tag {5} +$$ + +where $\mathbf{z}^r$ and $\mathbf{z}^v$ represent the latent point-wise embeddings of the range and voxel representations, respectively. + +As mentioned in Sec. 2, contrastive learning methods [7, 12, 23, 31, 46] only use a limited subset of actual training samples present in a mini batch, potentially leading to suboptimal convergence with confirmation bias. We propose a novel approach to represent the range and voxel embeddings with class-specific Gaussian Mixture Models (GMMs). Such approach enables the sampling of the most informative embeddings, fostering a comprehensive understanding of the distribution across the feature space of different representations. In practice, we concatenate the embeddings from both range and voxel based on their categories, where such embedding set is defined by $\mathcal{G}^y = \{(\mathbf{z},\mathbf{c},y)|\mathbf{z} = \mathbf{z}_i^r (\omega^r),\mathbf{c} =$ + +$\mathbf{c}_i^r (\omega^r),y = \tilde{\mathbf{y}}_i^r (\omega^r)\} \bigcup \{(\mathbf{z},\mathbf{c},y)|\mathbf{z} = \mathbf{z}_i^v (\omega^v),\mathbf{c} = \mathbf{c}_i^v (\omega^v),y = \tilde{\mathbf{y}}_i^v (\omega^v)\}$ for the unlabelled data indexed by $i\in \{1,\dots,|\mathcal{D}_U|\}$ , with the related confidence (i.e., $\mathbf{c}^r$ and $\mathbf{c}^v$ ) and pseudo labels (i.e., $\tilde{\mathbf{y}}^r$ and $\tilde{\mathbf{y}}^v$ ) from Eq. (2) at positions $\omega^r\in \Omega^r$ and $\omega^v\in \varOmega^v$ . The class-specific GMMs are defined for both representations by + +$$ +\mathbf {P} _ {\Gamma^ {y}} (\mathbf {z} | y) = \sum_ {m = 1} ^ {M} \mathbf {P} _ {\Gamma^ {y}} (m | y) \mathbf {P} _ {\Gamma^ {y}} (\mathbf {z} | m, y) = \sum_ {m = 1} ^ {M} \pi_ {m} ^ {y} \mathcal {N} (\mathbf {z}; \boldsymbol {\mu} _ {m} ^ {y}, \boldsymbol {\Sigma} _ {m} ^ {y}), \tag {6} +$$ + +where $\pmb{\pi}_{m}^{y} = \mathbf{P}_{\varGamma^{y}}(m|y)$ denotes the prior probability for the $m$ -th GMM component of class $y \in \{0,1\}^Y$ , $\pmb{\mu}_{m}^{y} \in \mathbb{R}^{Z}$ and $\pmb{\Sigma}_{m}^{y} \in \mathbb{R}^{Z \times Z}$ represent the mean and covariance of the corresponding GMM component and we fix $M = 5$ for all experiments. The GMM parameter optimisation in Eq. (6) relies on the EM [11] algorithm using the training set $\mathcal{G}^{y}$ defined above with the following steps: + +E step: $\mathbf{q}_{n,m}^{y} = \frac{\pi_{m}^{y}\mathcal{N}(\mathbf{z}_{n}\mid\boldsymbol{\mu}_{m}^{y},\boldsymbol{\Sigma}_{m}^{y})}{\sum_{m}^{M}\pi_{m}^{y}\mathcal{N}(\mathbf{z}_{n}\mid\boldsymbol{\mu}_{m}^{y},\boldsymbol{\Sigma}_{m}^{y})}$ , for $(\mathbf{z}_n,\mathbf{c}_n,y_n)\in \mathcal{G}^y$ + +M step: $\pmb{\mu}_{m}^{y} = \frac{1}{|\mathcal{G}^{y}|}\sum_{(\mathbf{z}_{n},\mathbf{c}_{n},y_{n})\in \mathcal{G}^{y}}\mathbf{c}_{n}\mathbf{q}_{n,m}^{y_{n}}\mathbf{z}_{n}$ (7) + +$$ +\boldsymbol{\Sigma}_{m}^{y} = \frac{1}{|\mathcal{G}^{y}|}\sum_{(\mathbf{z}_{n},\mathbf{c}_{n},y_{n})\in \mathcal{G}^{y}}\mathbf{c}_{n}\mathbf{q}_{n,m}^{y_{n}}(\mathbf{z}_{n} - \boldsymbol{\mu}_{m}^{y_{n}})\bigl(\mathbf{z}_{n} - \boldsymbol{\mu}_{m}^{y_{n}}\bigr)^{\intercal}, +$$ + +where $\mathcal{G}^y$ represents the set of embedding samples that share the same label $y$ . The posterior $\mathbf{q}_{n,m}^y$ estimates the responsibility of the $n^{th}$ embedding $\mathbf{z}$ from class $y$ in $\mathcal{G}^y$ being assigned to the $m$ -th component. The confidence $\mathbf{c}_n$ represents the potential correctness of the pseudo label from Eq. (2), which is used as a weighting factor to reduce the effect of low-confidence segmentation results [8,9] in the estimation of the GMM parameters. Note that we maintain a uniform weight of $\pi_{m}^{y} = 1 / M$ for each component [43, 57] to prevent one component from dominating others. + +GMM-based Prototypical Contrastive Learning. The class-specific GMMs from Eq. (7) are used to generate a set of highly representative embeddings for the contrastive learning process explained next. These representative class-specific embeddings is denoted by $S^y = \{\mathbf{s}|\mathbf{s}\sim \mathbf{P}_{\Gamma^y}(\mathbf{z}|y)\}$ , with $\mathbf{P}_{\Gamma^y}(\mathbf{z}|y)$ being the class-specific GMM defined in Eq. (6). Subsequently, these generated virtual prototypes are utilised to guide the contrastive learning for the incoming embedding features from both range and voxel representations [54], without joining the backpropagation computation graph. We define the set of incoming features with the anchor set $\mathcal{A} = \{(\mathbf{z},y)\}$ that is formed similarly to how we form the set $\mathcal{G}^y$ without including the confidence value. The samples in $\mathcal{A}$ are selectively chosen for backpropagation [54], but the informative samples in $S^y$ contribute to a comprehensive understanding of the cross-distribution for both representations. The contrastive learning relies on the minimisation of the following loss: + +$$ +\ell_ {\text {c r o s s}} \left(\mathcal {D} _ {L}, \mathcal {D} _ {U}, f _ {\phi_ {\mathbf {P R J}}}, f _ {\phi_ {\mathbf {F C N}}}, f _ {\theta_ {\mathbf {P R J}}}, f _ {\theta_ {\mathbf {E X R}}} \right) = \sum_ {(\mathbf {z}, y) \in \mathcal {A}} \sum_ {\mathbf {s} \in \mathcal {S} ^ {y}} \frac {\exp (\mathbf {z} \cdot \mathbf {s} / \tau)}{\exp (\mathbf {z} \cdot \mathbf {s} / \tau) + \sum_ {\mathbf {s} ^ {-} \in \bar {\mathcal {S}} ^ {y}} \exp (\mathbf {z} \cdot \mathbf {s} ^ {-} / \tau)}, \tag {8} +$$ + +where $\bar{S}^y$ is the set of virtual negative embeddings for class $y$ . + +# 3.3 Representation Specific Data Augmentation + +Following [39, 61], we apply spatial data augmentations to the pseudo-labeling strategy in Eq. (2) to increase the robustness of our approach to complex scenes. After experimenting with classic data augmentation processes (e.g., CutMix [63] and LaserMix [27]), we notice that the application of distinct spatial augmentations to different viewpoints in our IT2 framework results in more effective generalization. We have empirically found the multi-boxes CutMix [63] augmentation for the range image and single-inclination LaserMix [27] augmentation for the voxel representation yield the most robust results. Although this representation-specific data augmentation is not our main contribution, it is shown empirically to produce relevant improvements. Please see details in supplementary Sec. 2. + +# 3.4 Overall Training + +The overall training of our IT2 minimises the following loss function: + +$$ +\begin{array}{l} \mathcal {L} \left(\mathcal {D} _ {L}, \mathcal {D} _ {U}, \mathcal {D} _ {L}, \mathcal {D} _ {U}, f _ {\phi_ {\mathrm {F C N}}}, f _ {\phi_ {\mathrm {E X R}}}, f _ {\phi_ {\mathrm {S E G}}} f _ {\phi_ {\mathrm {P R J}}}, f _ {\phi_ {\mathrm {P R J}}}\right) = \\ \ell_ {\mathrm {I T 2}} \left(\mathcal {D} _ {L}, \mathcal {D} _ {U}, f _ {\phi_ {\mathbf {F C N}}}, f _ {\phi_ {\mathbf {E X R}}}, f _ {\phi_ {\mathbf {S E G}}}\right) + \ell_ {\mathrm {c r o s s}} \left(\mathcal {D} _ {L}, \mathcal {D} _ {U}, f _ {\phi_ {\mathbf {P R J}}}, f _ {\phi_ {\mathbf {F C N}}}, f _ {\theta_ {\mathbf {P R J}}}, f _ {\theta_ {\mathbf {E X R}}}\right), \tag {9} \\ \end{array} +$$ + +where the entire training pipeline is performed in an end-to-end manner, where in each iteration, we minimise the loss from $\ell_{\mathrm{IT2}}(.)$ and $\ell_{\mathrm{cross}}(.)$ simultaneously. + +# 4 Experiment + +# 4.1 Experimental Setup + +We evaluate our approach on three benchmarks, including nuScenes [4], SemanticKITTI [2] and ScribbleKITTI [52]. The nuScenes [4] dataset has 29,130 training scans and 6,019 validation scans. Each LiDAR scan contains 32 beams with $[-30^{\circ},10^{\circ}]$ vertical field of views (FOV). The dataset has 23 categories, but following previous works [27,66], we merge them into 16 for training and validation. The SemanticKITTI [2] and its extension ScribbleKITTI [52] are popular laser-based segmentation benchmarks, where the latter one is based on line-scribbles labelling to reduce the annotation effort. The SemanticKITTI [2] dataset has 19 categories for semantic segmentation, where every LiDAR scan contains 64 beams with $[3^{\circ}, -25^{\circ}]$ vertical FOV. There are 22 sequences of LiDAR scans in the dataset, where sequences 0 to 9 are used for training and validation. Following [23,27], we use 19,130 scans (sequences 0 to 7 and sequence 9) for training and 4,071 scans from sequence 8 for validation. + +Sampling strategies. Previous works [23, 27, 31] propose different sampling strategies to select the labelled data for the semi-supervised learning setting, including: 1) uniform [27] sampling of the labelled data across all the training + +Table 1: Uniform sampling results across various benchmarks, where we follow [27] to utilise half feature dimension and lower resolution in Cylinder3D [66] for all the experiments. Best results are highlighted in red. + +
Repr.MethodnuScenes [4]SemanticKITTI [2]ScribbleKITTI [52]
1%10%20%50%1%10%20%50%1%10%20%50%
Rangesup.38.357.562.767.636.252.255.957.233.147.749.952.5
MT [50]42.160.465.469.437.553.156.157.434.249.851.653.3
CBST [68]40.960.564.369.339.953.456.156.935.750.752.754.6
CPS [6]40.760.864.968.036.552.356.357.433.750.052.854.6
LaserMix [27]49.568.270.673.043.458.859.461.438.354.455.658.7
IT2 (Ours)56.571.373.474.051.960.361.762.146.657.157.358.6
Voxelsup.50.965.966.671.245.456.157.858.739.248.052.153.8
MT [50]51.666.067.171.745.457.159.260.041.050.152.853.9
CBST [68]53.066.569.671.648.858.359.459.741.550.653.354.5
CPS [6]52.966.370.072.546.758.759.660.541.451.853.954.8
LaserMix [27]55.369.971.873.250.660.061.962.344.253.755.156.8
IT2 (Ours)57.572.173.674.152.061.462.162.547.956.757.558.3
+ +sequences, 2) partial [23] sampling of the labelled data only from certain sequences, and 3) sampling of significant [31] labelled frames from the sequences. In the experiment, we examine our approach under all these sampling protocols. Metrics. We measure performance with mean intersection over union (mIoU). In each benchmark, we use the sampling strategies above to label $n\%$ of the training set, with the rest $(100 - n)\%$ being unlabelled. + +Implementation details. In the range view, we use the FIDNet architecture [65] with ResNet34 [19] backbone. The input range image has size $32 \times 1920$ for nuScenes [4] and $64 \times 2048$ for SemanticKITTI [2] and ScirbbleKITTI [52], following [27]. In the voxel representation, we use the Cylinder3D architecture [66], where we fix the intermediate layer at 16 dimensions with smaller resolution as specified in [27] for the uniform sampling experiments to enable a fair comparison. For contrastive learning, we build a 3-layer projector [24,54]. The projector outputs an embedding map with 64 channels, where we apply the temperature factor $\tau = .1$ in Eq. (8). See supplementary Sec. 3 and Sec. 7 for more details. + +# 4.2 Results on Different Partition Protocols + +Uniform sampling results. We first compare our method using the uniform sampling benchmark [27] in Tab. 1, with the experimental setup from [27] (i.e., architecture, split ID and the resolution of input representations). Our framework outperforms the previous SOTA Lasermix [27] by a large margin for almost all partition protocols. For example, we observe $\approx 3.4\%$ and $\approx 2.0\%$ improvements on average for range and voxel representations, respectively. Notably, we highlight that our approach yields significant improvements for the $1\%$ partition protocol in the line-scribble dataset [52], providing $8.28\%$ improvement in the + +Table 2: Partial & significant sampling results in the SemanticKITTI [2] benchmark. We compare with previous SOTA based on Cylinder3D [66] architecture, where TTA denotes test time augmentation. $\dagger$ indicates that we re-implement results based on the supported checkpoints from their GitHub - please see Supplementary Sec. 1 for more details. Best results in each evaluation protocol are highlighted in red. + +
TTAMethodSemanticKITTI(par.)MethodSemanticKITTI(sig.)
5%10%20%40%5%10%20%40%
XGPC† [23]39.945.955.757.1lim3D [31]56.859.660.561.4
IT2 (Ours)43.448.156.560.9IT2 (Ours)60.363.364.064.8
✓GPC† [23]40.246.356.557.6lim3D [31]59.562.263.163.3
IT2 (Ours)43.849.258.862.4IT2 (Ours)61.764.464.765.8
+ +range image and $3.66\%$ in the voxel representation. + +Partial and significant sampling results. In Tab. 2, we compare our results with GPC [23] and lim3D [31] in partial sampling (par.) and significant sampling (sig.), based on both original performance and test time augmentation (TTA). In the partial sampling protocol7, our approach yields superior results across all the labelled data situations for both with and without TTA. Notably, our results surpass GPC [23] by $4.5\%$ and $3.6\%$ in the limited $5\%$ labelled data condition, demonstrating its good generalisation capability. In the significant sampling results, our method also achieves better performance than lim3D [31] for all the partition protocols. For example, our results are better than the previous SOTA by about $2.2\%$ and $1.6\%$ in $10\%$ and $20\%$ labelled data with TTA, respectively. + +# 4.3 Ablation Study + +Improvements from our contributions. Table 3 presents our ablation study on the nuScenes [4] and ScribbleKITTI [52] datasets across various partition protocols. We investigate the roles of the IT2 architecture, cross-distribution contrastive learning (Ctrs.), and representation-related data augmentation (Aug.). The baseline utilises network perturbation through single representations [6] from [27]. Our IT2 architecture results in improvements of approximately $4.1\%$ and $2.5\%$ for the voxel representation in $10\%$ labelled data for nuScenes [4] and ScribbleKITTI [52], respectively. The Ctrs. further improves by $\approx 2.4\%$ and $\approx 1.1\%$ for range and voxel representations in both datasets in $10\%$ labelled data. Such enhancements show the effectiveness of our cross-distribution prototypes and contrastive learning strategy. Furthermore, Aug. results in improvements of $2.2\%$ and $1.4\%$ for range and voxel representations in the ScribbleKITTI [52] dataset for $10\%$ labelled data, demonstrating the better generalisation brought by the representation-specific data augmentation. + +Table 3: Ablation study in the nuScenes [4] and ScribbleKITTI [52] datasets via uniform sampling. IT2 represents our architecture, Ctrs. denotes our cross-distribution contrastive learning and Aug. is our representation-specific data augmentation. The yellow rows denote our single representation baseline with CPS [6] from [27]. + +
Repr.IT2 Ctrs. Aug.nuScenes [4]ScribbleKITTI [52]
10%20%50%10%20%50%
RangeXXX60.864.968.050.052.854.6
✓XX67.3 (6.5↑)70.8 (5.9↑)72.6 (4.6↑)53.2 (3.2↑)55.0 (2.2↑)56.2 (1.6↑)
✓✓X70.3 (9.5↑)72.6 (7.7↑)73.5 (5.5↑)54.9 (4.9↑)56.1 (3.3↑)57.2 (2.6↑)
✓✓✓71.3 (10.5↑)73.4 (8.5↑)74.0 (6.0↑)57.1 (7.1↑)57.3 (4.5↑)58.6 (4.0↑)
VoxelXXX66.370.072.551.853.954.8
✓XX70.4 (4.1↑)71.9 (1.9↑)73.1 (0.6↑)54.3 (2.5↑)55.6 (2.7↑)56.6 (1.8↑)
✓✓X71.6 (5.3↑)72.9 (2.9↑)73.9 (1.4↑)55.3 (3.5↑)56.5 (3.6↑)57.6 (2.8↑)
✓✓✓72.1 (5.8↑)73.6 (3.6↑)74.1 (1.6↑)56.7 (4.9↑)57.5 (4.6↑)58.3 (3.5↑)
+ +![](images/d42650b2c6538733dd1a8509370f8d1a5b6f5ab41155658c657037dac6d51bb6.jpg) +Fig. 3: Ablation study of temperature value $\tau$ from Eq. (8) in nuScenes [4] dataset under $10\%$ labelled partition protocol. + +Table 4: Contrastive Learning study in nuScenes [4], where ContrasSeg [54] is re-implemented from the SOTA image (pixel-level) semantic segmentation. + +
MethodRepr.nuScenes [4]
10%20%
ContrasSeg [54]range70.372.3
voxel71.272.8
Oursrange71.3 (1.0↑)73.4 (1.1↑)
voxel72.1 (0.9↑)73.6 (0.8↑)
+ +Cross-distribution contrastive learning. We investigate the impact of the hyperparameter $\tau$ from Eq. (8) in Fig. 3 using the $10\%$ labelled protocol on the nuScenes [4] dataset. Our observations indicate that $\tau = 0.10$ yields the best results, so we fix this value for all experiments. In Tab. 4, we compare our approach with our re-implementation of the label-guided pixel-wise contrastive learning [54] based on our IT2 architecture with different ratio of labelled data, where our method shows $\approx 1\%$ improvement for both representations, demonstrating the effectiveness of our proposed cross-distribution contrastive learning. Representation-specific augmentation. In Fig. 4, we have demonstrated the augmentation results for each representation in the ScribbleKITTI [52] dataset, where 'multi' denotes multi-box [63], and '1-inc' represents single inclination setups [27]. In our framework, one representation's results rely on the other's generalisation. For instance, '(multi) CutMix' reaches $57.1\%$ mIoU with '(1-inc) LaserMix,' improving by $0.5\%$ compared to the results with 'LaserMix.' Meanwhile, '(multi) CutMix' combined with '(1-inc) LaserMix' improves by $1.3\%$ in the range and $0.7\%$ in voxel representations compared to LaserMix [27], highlighting the improved generalisation benefits from our augmentation. + +Ablation study for other representations. In Tab. 5, we conduct abla + +![](images/cc11fcdc44998bfbbecc11735861c931911ebd5bc1dd6af6a279196d2897c507.jpg) +Table 5: Ablation study for the bird's eye view (BEV) experiment in the nuScenes [4], where results are from PolarNet [64]. The baseline relies on single view [6], followed by BEV with range image or voxel grid. + +
Repr.BEV (PolarNet [64])
10%20%50%
sup.58.563.968.4
w/ bev62.4 (3.9↑)66.7 (2.8↑)70.0 (1.6↑)
w/ range64.8 (6.3↑)67.9 (4.0↑)70.6 (2.2↑)
w/ voxel66.3 (7.8↑)69.1 (5.2↑)71.6 (3.2↑)
+ +![](images/d6f08f6ff4cb749c3191e7b450ced6b3186db14ad833561a843dea7fe1593797.jpg) +Fig. 4: Ablation study for the repr.-specific augmentations in ScribbleKITTI [52] with $10\%$ labelled data, where 'multi' is multi-boxes and '1-inc' is single inclination. + +![](images/e9b95a5defc6d47b1d7e7eb8da00a20818242b9f1bce0bef7583071d11a5983a.jpg) + +![](images/2e3804f3d5fd133ae4ec4cb6f161a8b076a3c4aa9eddfec5db5d2db024a60314.jpg) + +![](images/423e5f59caf966f7b1a488cc4f83d10575d1bfad41049cef93bf65630d7f0309.jpg) + +![](images/d20a87dd5527309d4d7128d0c548069c54da799d0b8ab2283d7ce32ced6a0b41.jpg) +(a) Ground truth + +![](images/49b013b85c98998b551c2c75e72dc8da8e7e0c74329fa5869ffa3d211ca185d5.jpg) +(b) MT [50] + +![](images/947c3dec638db299244e1dd7a020717d624d6d8a2baf0785bcbaee4f36e2f722.jpg) +(c) LaserMix [27] + +![](images/d7e4b2a21f93eddcef5173424ac2b33b630d2f9333bca2849037336b8bb148c5.jpg) +(d) Ours +Fig. 5: Error maps visualised from LiDAR points (top) and range view (bottom) in 2 examples of the nuScenes [4] dataset. The correct predictions are in green and the mistakes are highlighted in red. + +tion studies with an additional representation, the Bird's Eye View (BEV), on the nuScenes [4] dataset, where all the results in the table are based on PolarNet [64]. We re-implement the supervised results and single representation supervision [6] in the first and second rows, respectively. We emphasize that employing peer representations can significantly enhance performance compared to the single-representation baseline [6]. For instance, under the $10\%$ labelled partition protocol, employing BEV (based on PolarNet [64]) with the range image (based on FidNet [65]) and with the voxel grid (based on Cylinder3D [66]) lead to improvements of $2.4\%$ and $3.9\%$ in mIoU, respectively. + +![](images/a7d369ee918029812c70d2e104eba80df1a04ad4510cf4db1bd8a56f9479c05a.jpg) +(a) w/o Contras. + +![](images/5c9d838b19176cfbe93572eaa2d87dad574f734e7ff4c8e08512e61c21729b25.jpg) +(b) ContrasSeg [54] +Fig. 6: T-SNE visualisation for the pixel- (and voxel-) wise latent embedding features in the nuScenes [4] dataset. The features $\mathcal{V}$ and $\mathcal{O}$ indicate range and voxel representations and they are colored according to the category label. + +![](images/430e59adc6899344eadebe4b0138047bc33a9cfe5739ffe63f41b9e7bec59766.jpg) +(c) Ours + +# 4.4 Visualisation + +Visualisation of prediction error maps. We compare the error maps from predictions in Fig. 5, including Mean Teacher [50] (2nd column), LaserMix [27] (3rd column) and ours (last column). The visualisations show the bird's eye view's output in the first row, followed by the range view results in the second row. Notably, our method's prediction demonstrates a superior qualitative result, exhibiting fewer mistaken predictions (highlighted in red) and more correct predictions (marked in green) for both views. + +T-SNE visualisation of the latent features. Figure 6 displays the latent embedding distribution from the nuScenes [4] dataset. Fig. 6a shows the results without contrastive learning, Fig. 6b displays the results based on the label-guided contrastive learning [54], and our prototypical contrastive learning is in Fig. 6c. The samples denoted by $\text{和}$ and $\odot$ indicate the latent embeddings from range and voxel models, respectively, where different colours represent the classes. Fig. 6a displays an unstructured latent representation due to the absence of contrastive learning, but Fig. 6b successfully clusters class embeddings from the same modalities (e.g., $\{\text{和}, \text{和}\}$ and $\{\text{和}, \text{和}\}$ ), yet it fails to capture the semantic relationships between different modalities within the same category. Our method, depicted in Fig. 6c, successfully clusters embeddings from the same class and different modalities $\{\text{和}, \text{和}\}$ , enabling a more effective distribution of embeddings for subsequent segmentation. + +# 5 Conclusion + +In this study, we introduced the novel semi-supervised LiDAR semantic segmentation method ItTakesTwo (IT2). IT2 improves generalisation with consistent predictions from peer LiDAR representations, and it also introduces a more effective contrastive learning using informative embeddings sampled from a learned distribution of positives and negatives. Finally, through experimental studies, we identify an efficient way to apply data augmentation to each representation. Our approach achieves best results across various benchmarks under all partition protocols, demonstrating its effectiveness. + +Acknowledgements. The project is supported by the Australian Research Council (ARC) through grant FT190100525. + +# References + +1. Arazo, E., Ortego, D., Albert, P., O'Connor, N.E., McGuinness, K.: Pseudolabeling and confirmation bias in deep semi-supervised learning. In: 2020 International Joint Conference on Neural Networks (IJCNN). pp. 1-8. IEEE (2020) +2. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9297-9307 (2019) +3. Berman, M., Triki, A.R., Blaschko, M.B.: The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4413-4421 (2018) +4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020) +5. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems 33, 9912-9924 (2020) +6. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2613-2622 (2021) +7. Cheng, M., Hui, L., Xie, J., Yang, J.: Sspc-net: Semi-supervised semantic 3d point cloud segmentation network. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 1140-1147 (2021) +8. Cheung, Y.m.: A rival penalized em algorithm towards maximizing weighted likelihood for density mixture clustering with automatic model selection. In: Proceedings of the 17th International Conference on Pattern Recognition, 2004. ICPR 2004. vol. 4, pp. 633-636. IEEE (2004) +9. Cheung, Y.m.: Maximum weighted likelihood via rival penalized em for density mixture clustering with automatic model selection. IEEE Transactions on Knowledge and Data Engineering 17(6), 750-761 (2005) +10. Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d u-net: learning dense volumetric segmentation from sparse annotation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part II 19. pp. 424-432. Springer (2016) +11. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological) 39(1), 1-22 (1977) +12. Deng, S., Dong, Q., Liu, B., Hu, Z.: Superpoint-guided semi-supervised semantic segmentation of 3d point clouds. In: 2022 International conference on robotics and automation (ICRA). pp. 9214-9220. IEEE (2022) +13. Fan, L., Xiong, X., Wang, F., Wang, N., Zhang, Z.: Rangedet: In defense of range view for lidar-based 3d object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2918-2927 (2021) + +14. French, G., Aila, T., Laine, S., Mackiewicz, M., Finlayson, G.: Semi-supervised semantic segmentation needs strong, high-dimensional perturbations (2019) +15. Graham, B., Engelcke, M., Van Der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9224-9232 (2018) +16. Grigorescu, S., Trasnea, B., Cocias, T., Macesanu, G.: A survey of deep learning techniques for autonomous driving. Journal of Field Robotics 37(3), 362-386 (2020) +17. Guo, Y., Wang, H., Hu, Q., Liu, H., Liu, L., Bennamoun, M.: Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence 43(12), 4338-4364 (2020) +18. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) +19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +20. Hou, J., Graham, B., Nießner, M., Xie, S.: Exploring data-efficient 3d scene understanding with contrastive scene contexts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15587-15597 (2021) +21. Hou, Y., Zhu, X., Ma, Y., Loy, C.C., Li, Y.: Point-to-voxel knowledge distillation for lidar semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8479-8488 (2022) +22. Hung, W.C., Tsai, Y.H., Liou, Y.T., Lin, Y.Y., Yang, M.H.: Adversarial learning for semi-supervised semantic segmentation. arXiv preprint arXiv:1802.07934 (2018) +23. Jiang, L., Shi, S., Tian, Z., Lai, X., Liu, S., Fu, C.W., Jia, J.: Guided point contrastive learning for semi-supervised point cloud semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6423-6432 (2021) +24. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in Neural Information Processing Systems 33, 18661-18673 (2020) +25. Kohli, A.P.S., Sitzmann, V., Wetzstein, G.: Semantic implicit neural scene representations with semi-supervised training. In: 2020 International Conference on 3D Vision (3DV). pp. 423-433. IEEE (2020) +26. Kong, L., Liu, Y., Chen, R., Ma, Y., Zhu, X., Li, Y., Hou, Y., Qiao, Y., Liu, Z.: Rethinking range view representation for lidar segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 228-240 (2023) +27. Kong, L., Ren, J., Pan, L., Liu, Z.: Lasermix for semi-supervised lidar semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21705-21715 (2023) +28. Lai, X., Chen, Y., Lu, F., Liu, J., Jia, J.: Spherical transformer for lidar-based 3d recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17545-17555 (2023) +29. Le-Khac, P.H., Healy, G., Smeaton, A.F.: Contrastive representation learning: A framework and review. IEEE Access 8, 193907-193934 (2020) +30. Li, J., Zhou, P., Xiong, C., Hoi, S.C.: Prototypical contrastive learning of unsupervised representations. arXiv preprint arXiv:2005.04966 (2020) +31. Li, L., Shum, H.P., Breckon, T.P.: Less is more: Reducing task and model complexity for 3d point cloud semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9361-9371 (2023) + +32. Li, M., Xie, Y., Shen, Y., Ke, B., Qiao, R., Ren, B., Lin, S., Ma, L.: Hybridcr: Weakly-supervised 3d point cloud semantic segmentation via hybrid contrastive regularization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14930-14939 (2022) +33. Liang, C., Wang, W., Miao, J., Yang, Y.: Gmmseg: Gaussian mixture based generative semantic segmentation models. Advances in Neural Information Processing Systems 35, 31360-31375 (2022) +34. Liang, T., Xie, H., Yu, K., Xia, Z., Lin, Z., Wang, Y., Tang, T., Wang, B., Tang, Z.: Bevfusion: A simple and robust lidar-camera fusion framework. Advances in Neural Information Processing Systems 35, 10421-10434 (2022) +35. Liu, M., Zhou, Y., Qi, C.R., Gong, B., Su, H., Anguelov, D.: Less: Label-efficient semantic segmentation for lidar point clouds. In: European Conference on Computer Vision. pp. 70-89. Springer (2022) +36. Liu, W., Yue, X., Chen, Y., Denoeux, T.: Trusted multi-view deep learning with opinion aggregation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 7585-7593 (2022) +37. Liu, Y., Hu, Q., Lei, Y., Xu, K., Li, J., Guo, Y.: Box2seg: Learning semantics of 3d point clouds with box-level supervision. arXiv preprint arXiv:2201.02963 (2022) +38. Liu, Y., Kong, L., Cen, J., Chen, R., Zhang, W., Pan, L., Chen, K., Liu, Z.: Segment any point cloud sequences by distilling vision foundation models. arXiv preprint arXiv:2306.09347 (2023) +39. Liu, Y., Tian, Y., Chen, Y., Liu, F., Belagiannis, V., Carneiro, G.: Perturbed and strict mean teachers for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4258-4267 (2022) +40. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: IEEE International Conference on Robotics and Automation (ICRA) (2023) +41. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2774-2781. IEEE (2023) +42. Manning, C.D.: An introduction to information retrieval. Cambridge university press (2009) +43. Mena, G., Nejatbakhsh, A., Varol, E., Niles-Weed, J.: Sinkhorn em: an expectation-maximization algorithm based on entropic optimal transport. arXiv preprint arXiv:2006.16548 (2020) +44. Miech, A., Alayrac, J.B., Smaira, L., Laptev, I., Sivic, J., Zisserman, A.: End-to-end learning of visual representations from uncurated instructional videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9879-9889 (2020) +45. Milioto, A., Vizzo, I., Behley, J., Stachniss, C.: Rangenet++: Fast and accurate lidar semantic segmentation. In: 2019 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 4213-4220. IEEE (2019) +46. Nunes, L., Marcuzzi, R., Chen, X., Behley, J., Stachniss, C.: Segcontrast: 3d point cloud feature representation learning through self-supervised segment discrimination. IEEE Robotics and Automation Letters 7(2), 2116-2123 (2022) +47. Ouali, Y., Hudelot, C., Tami, M.: Semi-supervised semantic segmentation with cross-consistency training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12674-12684 (2020) + +48. Reichardt, L., Ebert, N., Wasenmüller, O.: 360deg from a single camera: A few-shot approach for lidar segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1075-1083 (2023) +49. Sun, C.Y., Yang, Y.Q., Guo, H.X., Wang, P.S., Tong, X., Liu, Y., Shum, H.Y.: Semi-supervised 3d shape segmentation with multilevel consistency and part substitution. Computational Visual Media 9(2), 229-247 (2023) +50. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. arXiv preprint arXiv:1703.01780 (2017) +51. Tian, Y., Krishnan, D., Isola, P.: Contrastive multiview coding. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16. pp. 776-794. Springer (2020) +52. Unal, O., Dai, D., Van Gool, L.: Scribble-supervised lidar semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2697-2707 (2022) +53. Vora, S., Radwan, N., Greff, K., Meyer, H., Genova, K., Sajjadi, M.S., Pot, E., Tagliasacchi, A., Duckworth, D.: Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes. arXiv preprint arXiv:2111.13260 (2021) +54. Wang, W., Zhou, T., Yu, F., Dai, J., Konukoglu, E., Van Gool, L.: Exploring cross-image pixel contrast for semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7303-7313 (2021) +55. Wu, B., Wan, A., Yue, X., Keutzer, K.: Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In: 2018 IEEE international conference on robotics and automation (ICRA). pp. 1887-1893. IEEE (2018) +56. Xie, S., Gu, J., Guo, D., Qi, C.R., Guibas, L., Litany, O.: Pointcontrast: Unsupervised pre-training for 3d point cloud understanding. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 574-591. Springer (2020) +57. Xu, J., Hsu, D.J., Maleki, A.: Benefits of over-parameterization with em. Advances in Neural Information Processing Systems 31 (2018) +58. Xu, J., Zhang, R., Dou, J., Zhu, Y., Sun, J., Pu, S.: Rpvnet: A deep and efficient range-point-voxel fusion network for lidar point cloud segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16024-16033 (2021) +59. Xu, J., Tang, H., Ren, Y., Peng, L., Zhu, X., He, L.: Multi-level feature learning for contrastive multi-view clustering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16051-16060 (2022) +60. Xu, Z., Yuan, B., Zhao, S., Zhang, Q., Gao, X.: Hierarchical point-based active learning for semi-supervised point cloud semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18098-18108 (2023) +61. Yang, L., Qi, L., Feng, L., Zhang, W., Shi, Y.: Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7236-7246 (2023) +62. Yang, L., Zhuo, W., Qi, L., Shi, Y., Gao, Y.: St++: Make self-training work better for semi-supervised semantic segmentation. arXiv preprint arXiv:2106.05095 (2021) +63. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6023-6032 (2019) + +64. Zhang, Y., Zhou, Z., David, P., Yue, X., Xi, Z., Gong, B., Foroosh, H.: Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9601-9610 (2020) +65. Zhao, Y., Bai, L., Huang, X.: Fidnet: Lidar point cloud semantic segmentation with fully interpolation decoding. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4453-4458. IEEE (2021) +66. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9939-9948 (2021) +67. Zhuang, Z., Li, R., Jia, K., Wang, Q., Li, Y., Tan, M.: Perception-aware multi-sensor fusion for 3d lidar semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16280-16290 (2021) +68. Zou, Y., Yu, Z., Kumar, B., Wang, J.: Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In: Proceedings of the European conference on computer vision (ECCV). pp. 289-305 (2018) +69. Zou, Y., Zhang, Z., Zhang, H., Li, C.L., Bian, X., Huang, J.B., Pfister, T.: Pseudoseg: Designing pseudo labels for semantic segmentation. arXiv preprint arXiv:2010.09713 (2020) \ No newline at end of file diff --git a/2024/ItTakesTwo_ Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation/images.zip b/2024/ItTakesTwo_ Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3e856eb91a435747d85f4268d4b90a28f11eb3bd --- /dev/null +++ b/2024/ItTakesTwo_ Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f6b40fa569e90585ca83933065812a5c33e3a659ca11f70994749ba2f1ed3cb +size 777136 diff --git a/2024/ItTakesTwo_ Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation/layout.json b/2024/ItTakesTwo_ Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ef231d395c3ad0a7f8e342e5b9713ca8ad278bfd --- /dev/null +++ b/2024/ItTakesTwo_ Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation/layout.json @@ -0,0 +1,11495 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 35, + 30, + 376, + 66 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 35, + 30, + 376, + 66 + ], + "spans": [ + { + "bbox": [ + 35, + 30, + 376, + 66 + ], + "type": "text", + "content": "ItTakesTwo: Leveraging Peer Representations for Semi-supervised LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 43, + 85, + 368, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 43, + 85, + 368, + 111 + ], + "spans": [ + { + "bbox": [ + 43, + 85, + 368, + 111 + ], + "type": "text", + "content": "Yuyuan Liu\\*1(), Yuanhong Chen\\*1, Hu Wang2,1, Vasileios Belagiannis3, Ian Reid2,1, and Gustavo Carneiro4,1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 40, + 119, + 370, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 119, + 370, + 131 + ], + "spans": [ + { + "bbox": [ + 40, + 119, + 370, + 131 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 40, + 119, + 370, + 131 + ], + "type": "text", + "content": " Australian Institute for Machine Learning (AIML), University of Adelaide, AU" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 131, + 360, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 131, + 360, + 142 + ], + "spans": [ + { + "bbox": [ + 51, + 131, + 360, + 142 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 51, + 131, + 360, + 142 + ], + "type": "text", + "content": " Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), UAE" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 109, + 142, + 302, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 142, + 302, + 153 + ], + "spans": [ + { + "bbox": [ + 109, + 142, + 302, + 153 + ], + "type": "text", + "content": "3 University of Erlangen-Nuremberg, Germany" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 153, + 379, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 153, + 379, + 164 + ], + "spans": [ + { + "bbox": [ + 31, + 153, + 379, + 164 + ], + "type": "text", + "content": "4 Centre for Vision, Speech and Signal Processing (CVSSP), University of Surrey, UK" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 186, + 354, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 186, + 354, + 384 + ], + "spans": [ + { + "bbox": [ + 59, + 186, + 354, + 384 + ], + "type": "text", + "content": "Abstract. The costly and time-consuming annotation process to produce large training sets for modelling semantic LiDAR segmentation methods has motivated the development of semi-supervised learning (SSL) methods. However, such SSL approaches often concentrate on employing consistency learning only for individual LiDAR representations. This narrow focus results in limited perturbations that generally fail to enable effective consistency learning. Additionally, these SSL approaches employ contrastive learning based on the sampling from a limited set of positive and negative embedding samples. This paper introduces a novel semisupervised LiDAR semantic segmentation framework called ItTakesTwo (IT2). IT2 is designed to ensure consistent predictions from peer LiDAR representations, thereby improving the perturbation effectiveness in consistency learning. Furthermore, our contrastive learning employs informative samples drawn from a distribution of positive and negative embeddings learned from the entire training set. Results on public benchmarks show that our approach achieves remarkable improvements over the previous state-of-the-art (SOTA) methods in the field. The code is available at: https://github.com/yyliu01/IT2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 394, + 351, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 394, + 351, + 406 + ], + "spans": [ + { + "bbox": [ + 59, + 394, + 351, + 406 + ], + "type": "text", + "content": "Keywords: Semi-supervised Learning " + }, + { + "bbox": [ + 59, + 394, + 351, + 406 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 59, + 394, + 351, + 406 + ], + "type": "text", + "content": " LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 31, + 423, + 128, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 423, + 128, + 436 + ], + "spans": [ + { + "bbox": [ + 31, + 423, + 128, + 436 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 30, + 445, + 380, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 445, + 380, + 565 + ], + "spans": [ + { + "bbox": [ + 30, + 445, + 380, + 565 + ], + "type": "text", + "content": "Outdoor LiDAR semantic segmentation is emerging as an important area of research given its application to self-driving vehicles [16, 17]. Although LIDAR data has advantages relative to photometric camera data, it also presents unique challenges, such as data sparsity [66] and varying point density [27, 45]. Most existing works alleviate these issues by transforming the LiDAR point-clouds to other representations, including range-based image [13, 26, 45, 55], voxel grids [15, 21, 66], or fusing them with other modalities [38, 41] and views [58]. Despite their success, those methods rely on supervised learning that requires large labelled training sets, which can be problematic given the costly and time-consuming labelling process [39, 47]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 41, + 571, + 339, + 583 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 571, + 339, + 583 + ], + "spans": [ + { + "bbox": [ + 41, + 571, + 339, + 583 + ], + "type": "text", + "content": "* denotes equal contribution. (☑) Email: yuyuan.liu@adelaide.edu.au." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 32, + 31, + 381, + 167 + ], + "blocks": [ + { + "bbox": [ + 32, + 31, + 381, + 167 + ], + "lines": [ + { + "bbox": [ + 32, + 31, + 381, + 167 + ], + "spans": [ + { + "bbox": [ + 32, + 31, + 381, + 167 + ], + "type": "image", + "image_path": "a9315ad125de3c5b132dd037dbbdb275a82c73d44310d0125031ce5043815f67.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 30, + 172, + 380, + 250 + ], + "lines": [ + { + "bbox": [ + 30, + 172, + 380, + 250 + ], + "spans": [ + { + "bbox": [ + 30, + 172, + 380, + 250 + ], + "type": "text", + "content": "Fig. 1: The left figure represents the 3D point cloud scan (ground truth of four classes - Car, Pole, Terrain, and Sidewalk) while the right figures denote the prediction results (into the same four classes), provided by our model, from range (top-right) and voxel (bottom-right) representations, where the incorrect predictions are marked in gray. Please notice that: 1) the mapping of the four classes from the 3D point cloud scan to each representation demonstrates the consistent distribution of semantic classes, and 2) the histograms of the predictions for each representation exhibit distinct differences." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 30, + 265, + 379, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 265, + 379, + 361 + ], + "spans": [ + { + "bbox": [ + 30, + 265, + 379, + 361 + ], + "type": "text", + "content": "These issues have motivated semi-supervised learning (SSL) innovations for LiDAR semantic segmentation [23,27,31], which require a small labelled dataset used in conjunction with a large unlabelled dataset. Consistency Learning, one of the main strategies explored in SSL for image semantic segmentation, penalises inconsistent predictions from perturbed inputs, where these perturbations take various forms, including input augmentation [62, 69], adversarial noise [22, 39], network perturbations [6] and dropout [61]. The incorporation of strong and varied perturbations is crucial for enhancing generalisation [14, 39]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 30, + 361, + 380, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 361, + 380, + 555 + ], + "spans": [ + { + "bbox": [ + 30, + 361, + 380, + 555 + ], + "type": "text", + "content": "Current research [27,31] applies such consistency learning in SSL for LiDAR points using many strategies, such as laser-beam augmentation [27] or Mean Teacher (MT) [31] framework. Despite their success, these approaches solely rely on a single LiDAR representation (e.g., voxel grid) for training, which could potentially limit the model's ability to generalize effectively due to the restricted perturbations. Such an issue can be mitigated with the use of multiple LiDAR representations for consistency learning that explores the fundamental Clustering assumption5, where the results from different representations naturally indicate the same underlying semantic information from the 3D LiDAR points. For example, as illustrated in Fig. 1, the class Terrain (▲) in the top-right pixels and bottom-right voxels should be \"clustered\" according to the (left) original 3D point data. Thus, the peer LiDAR representations can be regarded as a novel form similar to perturbation, but yet under-explored in the field. Other disadvantages of focusing on only one of the representations when using small labelled datasets are: the range view suffers from the information loss during the projection process, particularly for the tail categories, such as Pole (▲), which has" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "spans": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 33, + 560, + 379, + 583 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 560, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 560, + 379, + 583 + ], + "type": "text", + "content": "5 Cluster hypothesis (or assumption) [42]: the data of various forms that behave similarly with respect to information relevance should be clustered together." + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 34, + 379, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 34, + 379, + 108 + ], + "spans": [ + { + "bbox": [ + 32, + 34, + 379, + 108 + ], + "type": "text", + "content": "a low mIoU of 52.02; and the voxel grid tends to show lower accuracy in dense areas that are far away from the LiDAR source, such as Sidewalk (" + }, + { + "bbox": [ + 32, + 34, + 379, + 108 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 32, + 34, + 379, + 108 + ], + "type": "text", + "content": "), which has a low mIoU of 69.50. Such limited pseudo label results will inevitably introduce confirmation bias (i.e., the problem of relying on and reinforcing wrong predictions during training), reducing even further the efficacy of consistency learning based on single representation [23, 27, 31]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 110, + 379, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 110, + 379, + 263 + ], + "spans": [ + { + "bbox": [ + 32, + 110, + 379, + 263 + ], + "type": "text", + "content": "Another effective technique for semi-supervised LiDAR segmentation is contrastive learning [7,20,23,31], which consists of attracting positive and repelling negative embedding pairs at instance [46] or category [23,31] levels, thereby enhancing the latent space point distribution. However, to generate these pairs, methods typically rely on a limited number of embeddings extracted from randomly selected training samples due to the high computational cost associated with dense contrastive learning tasks [23,31]. Unfortunately, this limited sample size hinders a comprehensive understanding of the distribution across the entire embedding space [5], resulting in suboptimal convergence of LiDAR segmentation in SSL. This limitation becomes particularly pronounced when dealing with embeddings from different models in multiple views, where potential noisy data [44, 51] or model bias [59] from one network can adversely affect others, thereby impeding the optimization process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 32, + 265, + 379, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 265, + 379, + 443 + ], + "spans": [ + { + "bbox": [ + 32, + 265, + 379, + 443 + ], + "type": "text", + "content": "In this work, we propose a novel semi-supervised LiDAR semantic segmentation framework, named ItTakesTwo (IT2). IT2 employs consistency learning based on peer-representations perturbation, which enforces the consistent predictions across two different representations. Such perturbation contributes with a more effective SSL generalisation and alleviates issues related to pseudo label qualities from a single representation. Furthermore, we propose cross-distribution contrastive learning, where instead of randomly selecting training samples in the embedding space, we learn informative embedding distributions (also known as prototype) via Gaussian Mixture Models from both representations. Then the embedding samples are drawn from such prototypical distributions to produce representative features that enhance the understanding of the latent space in peer representation scenarios. In addition, instead of applying a single type of data augmentation to both LiDAR representations [27], we empirically show that applying different augmentations to distinct representations provides better generalisation. To summarise, our work makes the following key contributions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 35, + 447, + 379, + 554 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 36, + 447, + 379, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 447, + 379, + 495 + ], + "spans": [ + { + "bbox": [ + 36, + 447, + 379, + 495 + ], + "type": "text", + "content": "1. A novel SSL LiDAR segmentation framework (IT2) that leverages peer-representation perturbation in consistency learning, encouraging the same predictions for the points with the same semantic meaning in different LiDAR representations;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 35, + 495, + 379, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 35, + 495, + 379, + 531 + ], + "spans": [ + { + "bbox": [ + 35, + 495, + 379, + 531 + ], + "type": "text", + "content": "2. A new contrastive learning where informative positive and negative pairs of embeddings are sampled from prototype distributions (modeled via Gaussian Mixture Models) of both representations; and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 35, + 531, + 379, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 35, + 531, + 379, + 554 + ], + "spans": [ + { + "bbox": [ + 35, + 531, + 379, + 554 + ], + "type": "text", + "content": "3. An innovative representation-specific data augmentation that enables IT2 achieve better generalisation." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 32, + 559, + 390, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 559, + 390, + 583 + ], + "spans": [ + { + "bbox": [ + 32, + 559, + 390, + 583 + ], + "type": "text", + "content": "Our IT2 has state-of-the-art (SOTA) performance on nuScenes [4], SemanticKitti [2] and ScribbleKitti [52], and we also achieve best performance with different" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 10, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 10, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 298, + 10, + 346, + 19 + ], + "type": "text", + "content": "ItTakesTwo" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 33, + 34, + 379, + 69 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 34, + 379, + 69 + ], + "spans": [ + { + "bbox": [ + 33, + 34, + 379, + 69 + ], + "type": "text", + "content": "sampling strategies to form the labelled and unlabelled subsets. Notably, in nuScenes [4] dataset, we outperform the previous SOTA [27] by around " + }, + { + "bbox": [ + 33, + 34, + 379, + 69 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 33, + 34, + 379, + 69 + ], + "type": "text", + "content": " in the limited labelled data situations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 89, + 135, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 89, + 135, + 102 + ], + "spans": [ + { + "bbox": [ + 33, + 89, + 135, + 102 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 116, + 379, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 116, + 379, + 306 + ], + "spans": [ + { + "bbox": [ + 33, + 116, + 379, + 306 + ], + "type": "text", + "content": "LiDAR Semantic Segmentation is an essential task for autonomous driving [17]. To more effectively extract the feature information from the unordered LiDAR point data, current approaches [45, 66, 67] transform such point data into different representations. Range representation methods [26, 45, 65] transfer the LiDAR points to the spherical coordinates [45] and project them to the image space [26, 65]. Voxel representation methods [15, 28, 66] voxelise the LiDAR points and project them into different sub-types of geometry views to alleviate the sparse data challenges [10, 15], encompassing cartesian [53], spherical [28] and cylindrical [66] coordinate systems. Furthermore, some methods [34, 38] take inputs from a combination of LiDAR representations [48, 67] or incorporate inputs from camera modalities [34, 40], fusing them in the intermediate layer to enhance feature understanding. Although these supervised methods show promising results [58, 65, 66], there has been a longstanding interest in minimising the extensive labeling efforts, with approaches such as scribble annotation [52], box supervision [37], active learning [35, 60], and SSL [23, 27]. In this paper, we focus on SSL to alleviate such annotation costs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 309, + 379, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 309, + 379, + 523 + ], + "spans": [ + { + "bbox": [ + 33, + 309, + 379, + 523 + ], + "type": "text", + "content": "Semi-supervised LiDAR Semantic Segmentation leverages unlabelled data to enhance the model's generalisation beyond the small labelled dataset. The current literature on LiDAR SSL primarily focuses on indoor [7, 12, 32, 60] or object [25, 49] point cloud scenarios, which have uniformly distributed and high density points in the scan. Such scenario is markedly different from the representation of outdoor LiDAR points that pose unique challenges, such as complex structure [27] and low density points [66]. Previous semi-supervised LiDAR segmentation methods [23, 31, 46] mainly rely on contrastive learning to improve the latent space understanding. For instance, GPC [23] utilises confidence-threshold to constrain potentially noisy pseudo labels, and Lim3D [31] uses a memory bank to save negative samples. To increase the variation of the LiDAR data, LaserMix [27] proposes a novel augmentation that mixes laser beams in various spatial positions and holds the consistency learning under the Mean Teacher (MT) architecture [50]. However, these methods use only single LiDAR representations which, as explained in Sec. 1, imposes many limitations, such as limited perturbation effectiveness and poor performance for specific categories. Our framework introduces a novel peer-representation perturbation that employs various LiDAR representations during training, resulting in improved generalisation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 523, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 523, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 523, + 379, + 583 + ], + "type": "text", + "content": "Contrastive Learning [5, 24, 54, 56] enforces latent embeddings to be closer to instances with the same semantic meaning (i.e. positive) and to be far away from instances with different semantic meanings (i.e. negative). This approach has been explored in LiDAR segmentation [7, 12, 23, 31, 46], but the reliance on actual training samples leads to sub-optimal training as only a subset of em" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 10, + 116, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 10, + 116, + 19 + ], + "spans": [ + { + "bbox": [ + 66, + 10, + 116, + 19 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 33, + 32, + 381, + 178 + ], + "blocks": [ + { + "bbox": [ + 33, + 32, + 381, + 178 + ], + "lines": [ + { + "bbox": [ + 33, + 32, + 381, + 178 + ], + "spans": [ + { + "bbox": [ + 33, + 32, + 381, + 178 + ], + "type": "image", + "image_path": "d3f22719c0b6c5a2809055d3d6a1f0aba4dceb103fcdfcc6e5bca31ade8d09ab.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 31, + 185, + 380, + 274 + ], + "lines": [ + { + "bbox": [ + 31, + 185, + 380, + 274 + ], + "spans": [ + { + "bbox": [ + 31, + 185, + 380, + 274 + ], + "type": "text", + "content": "Fig. 2: Illustration of our approach with unlabeled data. The point scans in the batch " + }, + { + "bbox": [ + 31, + 185, + 380, + 274 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_1" + }, + { + "bbox": [ + 31, + 185, + 380, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 31, + 185, + 380, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_2" + }, + { + "bbox": [ + 31, + 185, + 380, + 274 + ], + "type": "text", + "content": " on the left) are transformed into range (and processed by " + }, + { + "bbox": [ + 31, + 185, + 380, + 274 + ], + "type": "inline_equation", + "content": "f_{\\phi_{\\mathrm{FCN}}}(.))" + }, + { + "bbox": [ + 31, + 185, + 380, + 274 + ], + "type": "text", + "content": " and voxel (and processed by " + }, + { + "bbox": [ + 31, + 185, + 380, + 274 + ], + "type": "inline_equation", + "content": "f_{\\theta_{\\mathrm{EXR}}}(.))" + }, + { + "bbox": [ + 31, + 185, + 380, + 274 + ], + "type": "text", + "content": " representations. Subsequently, the results from " + }, + { + "bbox": [ + 31, + 185, + 380, + 274 + ], + "type": "inline_equation", + "content": "f_{\\phi_{\\mathrm{SEG}}}(f_{\\phi_{\\mathrm{FCN}}}(.))" + }, + { + "bbox": [ + 31, + 185, + 380, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 31, + 185, + 380, + 274 + ], + "type": "inline_equation", + "content": "f_{\\theta_{\\mathrm{SEG}}}(f_{\\theta_{\\mathrm{EXR}}}(.))" + }, + { + "bbox": [ + 31, + 185, + 380, + 274 + ], + "type": "text", + "content": " are transformed to the other representation with Eq. (2) and cross-supervised with Eq. (4). Also, the embedding " + }, + { + "bbox": [ + 31, + 185, + 380, + 274 + ], + "type": "inline_equation", + "content": "\\{\\cdot ,\\cdot \\}" + }, + { + "bbox": [ + 31, + 185, + 380, + 274 + ], + "type": "text", + "content": " selected by the pseudo labels are utilised to model the GMMs with Eq. (6) and penalised by the informative samples (generated from GMMs) based on " + }, + { + "bbox": [ + 31, + 185, + 380, + 274 + ], + "type": "inline_equation", + "content": "\\ell_{cross}(.)" + }, + { + "bbox": [ + 31, + 185, + 380, + 274 + ], + "type": "text", + "content": " in Eq. (8). Please note the arrow indicates that the operation does not require gradient computation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 30, + 290, + 379, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 290, + 379, + 434 + ], + "spans": [ + { + "bbox": [ + 30, + 290, + 379, + 434 + ], + "type": "text", + "content": "bedding pairs [23, 31] can be selected for computing the contrastive loss. This becomes more challenging when working with embeddings obtained from different representations [36, 51, 59], as potentially noisy predictions [44, 59] from each representation can result in confirmation bias during training [1]. Prototypical contrastive learning [5, 18, 29, 30, 33] explores the use of a set of representative embeddings for the contrastive learning via memory bank [18, 54], optimal transport [5], or deep clustering [29, 30]. Nevertheless, such prototypes represent the modes of a more complex underlying distribution of prototypes that has not been explored before for SSL LiDAR segmentation. In our framework, we employ Gaussian Mixture Models (GMMs) to represent class-specific embedding distributions from both LiDAR representations and to generate informative embeddings for all classes, enhancing the effectiveness of the contrastive learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 454, + 131, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 454, + 131, + 468 + ], + "spans": [ + { + "bbox": [ + 31, + 454, + 131, + 468 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 484, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 484, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 484, + 380, + 584 + ], + "type": "text", + "content": "Our approach is based on a labelled point dataset " + }, + { + "bbox": [ + 30, + 484, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_L = \\{(\\mathbf{x}_i,\\mathbf{y}_i)\\}_{i = 1}^{| \\mathcal{P}_L |}" + }, + { + "bbox": [ + 30, + 484, + 380, + 584 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 30, + 484, + 380, + 584 + ], + "type": "inline_equation", + "content": "|\\mathcal{P}_L|" + }, + { + "bbox": [ + 30, + 484, + 380, + 584 + ], + "type": "text", + "content": " scans and an unlabelled set " + }, + { + "bbox": [ + 30, + 484, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_U = \\{(\\mathbf{x}_i\\}_{i = 1}^{| \\mathcal{P}_U|}" + }, + { + "bbox": [ + 30, + 484, + 380, + 584 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 30, + 484, + 380, + 584 + ], + "type": "inline_equation", + "content": "|\\mathcal{P}_U|" + }, + { + "bbox": [ + 30, + 484, + 380, + 584 + ], + "type": "text", + "content": " scans. Each scan " + }, + { + "bbox": [ + 30, + 484, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i\\in \\mathcal{P}\\subset \\mathbb{R}^{3\\times C\\times N}" + }, + { + "bbox": [ + 30, + 484, + 380, + 584 + ], + "type": "text", + "content": " has 3 spatial coordinates, with " + }, + { + "bbox": [ + 30, + 484, + 380, + 584 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 30, + 484, + 380, + 584 + ], + "type": "text", + "content": " denoting the number of features (intensity, elongation, etc.), " + }, + { + "bbox": [ + 30, + 484, + 380, + 584 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 30, + 484, + 380, + 584 + ], + "type": "text", + "content": " representing the number of points, and " + }, + { + "bbox": [ + 30, + 484, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i\\in \\mathcal{V}\\subset \\{0,1\\}^{Y\\times N}" + }, + { + "bbox": [ + 30, + 484, + 380, + 584 + ], + "type": "text", + "content": " represents the one-hot label with " + }, + { + "bbox": [ + 30, + 484, + 380, + 584 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 30, + 484, + 380, + 584 + ], + "type": "text", + "content": " being the number of classes in the labelled set. This point cloud dataset is transformed into the range and voxel representations to build the two peer representation datasets that is required by our method. These transformations are defined by [58] as:" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 10, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 10, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 298, + 10, + 346, + 19 + ], + "type": "text", + "content": "ItTakesTwo" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 373, + 10, + 379, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 10, + 379, + 18 + ], + "spans": [ + { + "bbox": [ + 373, + 10, + 379, + 18 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 37, + 34, + 380, + 135 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 37, + 34, + 380, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 34, + 380, + 84 + ], + "spans": [ + { + "bbox": [ + 37, + 34, + 380, + 84 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 37, + 34, + 380, + 84 + ], + "type": "inline_equation", + "content": "\\psi_{p\\rightarrow v}:\\mathcal{P}\\rightarrow \\mathcal{V}" + }, + { + "bbox": [ + 37, + 34, + 380, + 84 + ], + "type": "text", + "content": " is a point to voxel projection function that forms " + }, + { + "bbox": [ + 37, + 34, + 380, + 84 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_L = \\{(x_i^v,y_i^v)|(x_i,y_i)\\in \\mathcal{P}_L\\}_{i = 1}^{|\\mathcal{P}_L|}" + }, + { + "bbox": [ + 37, + 34, + 380, + 84 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 37, + 34, + 380, + 84 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_U = \\{\\mathbf{x}_i^v |\\mathbf{x}_i\\in \\mathcal{P}_U\\}_{i = 1}^{|\\mathcal{P}_U|}" + }, + { + "bbox": [ + 37, + 34, + 380, + 84 + ], + "type": "text", + "content": ", with each volume " + }, + { + "bbox": [ + 37, + 34, + 380, + 84 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^v\\in \\mathcal{V}\\subset \\mathbb{R}^{C_v\\times H\\times W\\times L}" + }, + { + "bbox": [ + 37, + 34, + 380, + 84 + ], + "type": "text", + "content": " having size " + }, + { + "bbox": [ + 37, + 34, + 380, + 84 + ], + "type": "inline_equation", + "content": "H\\times W\\times L" + }, + { + "bbox": [ + 37, + 34, + 380, + 84 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 37, + 34, + 380, + 84 + ], + "type": "inline_equation", + "content": "C_v" + }, + { + "bbox": [ + 37, + 34, + 380, + 84 + ], + "type": "text", + "content": " features, and the projected label being " + }, + { + "bbox": [ + 37, + 34, + 380, + 84 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^v\\in \\{0,1\\}^{Y\\times H\\times W\\times L}" + }, + { + "bbox": [ + 37, + 34, + 380, + 84 + ], + "type": "text", + "content": " in one-hot format." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 37, + 84, + 380, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 84, + 380, + 135 + ], + "spans": [ + { + "bbox": [ + 37, + 84, + 380, + 135 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 37, + 84, + 380, + 135 + ], + "type": "inline_equation", + "content": "\\psi_{p\\rightarrow r}:\\mathcal{P}\\rightarrow \\mathcal{R}" + }, + { + "bbox": [ + 37, + 84, + 380, + 135 + ], + "type": "text", + "content": " is a function to project point to the range image6. With this, we build " + }, + { + "bbox": [ + 37, + 84, + 380, + 135 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_L = \\{(\\mathbf{x}_i^r,\\mathbf{y}_i^r)|(x_i,y_i)\\in \\mathcal{P}_L\\}_{i = 1}^{|\\mathcal{P}_L|}" + }, + { + "bbox": [ + 37, + 84, + 380, + 135 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 37, + 84, + 380, + 135 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_U = \\{\\mathbf{x}_i^r |\\mathbf{x}_i\\in \\mathcal{P}_U\\}_{i = 1}^{|\\mathcal{P}_U|}" + }, + { + "bbox": [ + 37, + 84, + 380, + 135 + ], + "type": "text", + "content": ", with each range image " + }, + { + "bbox": [ + 37, + 84, + 380, + 135 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^r\\in \\mathcal{R}\\subset \\mathbb{R}^{C_r\\times U\\times V}" + }, + { + "bbox": [ + 37, + 84, + 380, + 135 + ], + "type": "text", + "content": " having size " + }, + { + "bbox": [ + 37, + 84, + 380, + 135 + ], + "type": "inline_equation", + "content": "U\\times V" + }, + { + "bbox": [ + 37, + 84, + 380, + 135 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 37, + 84, + 380, + 135 + ], + "type": "inline_equation", + "content": "C_r" + }, + { + "bbox": [ + 37, + 84, + 380, + 135 + ], + "type": "text", + "content": " features, and the projected label being " + }, + { + "bbox": [ + 37, + 84, + 380, + 135 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^r\\in \\{0,1\\}^{Y\\times U\\times V}" + }, + { + "bbox": [ + 37, + 84, + 380, + 135 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 30, + 140, + 381, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 140, + 381, + 239 + ], + "spans": [ + { + "bbox": [ + 30, + 140, + 381, + 239 + ], + "type": "text", + "content": "The inverse of the transformations above are respectively denoted by " + }, + { + "bbox": [ + 30, + 140, + 381, + 239 + ], + "type": "inline_equation", + "content": "\\psi_{v\\rightarrow p}" + }, + { + "bbox": [ + 30, + 140, + 381, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 140, + 381, + 239 + ], + "type": "inline_equation", + "content": "\\psi_{r\\rightarrow p}" + }, + { + "bbox": [ + 30, + 140, + 381, + 239 + ], + "type": "text", + "content": ". Based on the transformed range and voxel representations from " + }, + { + "bbox": [ + 30, + 140, + 381, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_L" + }, + { + "bbox": [ + 30, + 140, + 381, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 140, + 381, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_U" + }, + { + "bbox": [ + 30, + 140, + 381, + 239 + ], + "type": "text", + "content": ", we define the labelled set " + }, + { + "bbox": [ + 30, + 140, + 381, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_L = \\{(\\mathbf{x}_i^v,\\mathbf{y}_i^v),(\\mathbf{x}_i^r,\\mathbf{y}_i^r)|(\\mathbf{x}_i^v,\\mathbf{y}_i^v)\\in \\mathcal{V}_L,(\\mathbf{x}_i^r,\\mathbf{y}_i^r)\\in \\mathcal{R}_L,\\}_{i = 1}^{|\\mathcal{D}_L|}" + }, + { + "bbox": [ + 30, + 140, + 381, + 239 + ], + "type": "text", + "content": " and the larger unlabelled set " + }, + { + "bbox": [ + 30, + 140, + 381, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_U = \\{(\\mathbf{x}_i^v,\\mathbf{x}_i^r)|\\mathbf{x}_i^v\\in \\mathcal{V}_U,\\mathbf{x}_i^r\\in \\mathcal{R}_U\\}_{i = 1}^{|\\mathcal{D}_U|}" + }, + { + "bbox": [ + 30, + 140, + 381, + 239 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 30, + 140, + 381, + 239 + ], + "type": "inline_equation", + "content": "|\\mathcal{D}_L|\\leq |\\mathcal{D}_U|" + }, + { + "bbox": [ + 30, + 140, + 381, + 239 + ], + "type": "text", + "content": ". These datasets are used to train our IT2 methodology in Sec. 3.1, while Sec. 3.2 describes our Gaussian Mixture Models (GMMs) based prototypical contrastive learning. Lastly, we introduce the details of our augmentation method in Sec. 3.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 31, + 250, + 208, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 250, + 208, + 262 + ], + "spans": [ + { + "bbox": [ + 31, + 250, + 208, + 262 + ], + "type": "text", + "content": "3.1 ItTakesTwo (IT2) Framework" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 268, + 381, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 268, + 381, + 413 + ], + "spans": [ + { + "bbox": [ + 30, + 268, + 381, + 413 + ], + "type": "text", + "content": "As illustrated in Fig. 2, our approach enforces prediction consistency between two peer representations to explore a potentially more effective perturbation. Our method can use various LiDAR representations, but we follow [27] and employ range and voxel networks for a fair comparison in the benchmark. The range image network consists of an FCN Network (e.g., ResNet [19]) that maps the input to a latent dense feature space with " + }, + { + "bbox": [ + 30, + 268, + 381, + 413 + ], + "type": "inline_equation", + "content": "f_{\\phi_{\\mathbf{FCN}}} : \\mathcal{R} \\to \\mathcal{I}" + }, + { + "bbox": [ + 30, + 268, + 381, + 413 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 30, + 268, + 381, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{I} \\subset \\mathbb{R}^{U \\times V \\times D_r}" + }, + { + "bbox": [ + 30, + 268, + 381, + 413 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 30, + 268, + 381, + 413 + ], + "type": "inline_equation", + "content": "D_r" + }, + { + "bbox": [ + 30, + 268, + 381, + 413 + ], + "type": "text", + "content": " denoting feature dimensionality. Then a segmentation head will map " + }, + { + "bbox": [ + 30, + 268, + 381, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 30, + 268, + 381, + 413 + ], + "type": "text", + "content": " to the categorical segmentation map with " + }, + { + "bbox": [ + 30, + 268, + 381, + 413 + ], + "type": "inline_equation", + "content": "f_{\\phi_{\\mathbf{SEG}}} : \\mathcal{I} \\to \\mathbb{R}^{U \\times V \\times Y}" + }, + { + "bbox": [ + 30, + 268, + 381, + 413 + ], + "type": "text", + "content": ". Similarly, the voxel network extracts features with " + }, + { + "bbox": [ + 30, + 268, + 381, + 413 + ], + "type": "inline_equation", + "content": "f_{\\theta_{\\mathbf{EXR}}} : \\mathcal{V} \\to \\mathcal{K}" + }, + { + "bbox": [ + 30, + 268, + 381, + 413 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 30, + 268, + 381, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{K} \\in \\mathbb{R}^{H \\times W \\times L \\times D_v}" + }, + { + "bbox": [ + 30, + 268, + 381, + 413 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 30, + 268, + 381, + 413 + ], + "type": "inline_equation", + "content": "D_v" + }, + { + "bbox": [ + 30, + 268, + 381, + 413 + ], + "type": "text", + "content": " denoting the feature dimensionality, followed by a segmentation head " + }, + { + "bbox": [ + 30, + 268, + 381, + 413 + ], + "type": "inline_equation", + "content": "f_{\\theta_{\\mathbf{SEG}}} : \\mathcal{K} \\to \\mathbb{R}^{H \\times W \\times L \\times Y}" + }, + { + "bbox": [ + 30, + 268, + 381, + 413 + ], + "type": "text", + "content": ". The peer representation consistency training depends on the following categorical distributions for both views:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 113, + 419, + 378, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 419, + 378, + 439 + ], + "spans": [ + { + "bbox": [ + 113, + 419, + 378, + 439 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {y}} ^ {r} \\left(\\omega^ {r}\\right) = \\operatorname {s o f t m a x} ^ {\\left(\\omega^ {r}\\right)} \\left(f _ {\\phi_ {\\mathrm {S E G}}} \\left(f _ {\\phi_ {\\mathrm {F C N}}} \\left(\\mathbf {x} ^ {r}\\right)\\right)\\right), \\tag {1}", + "image_path": "0fa8dd0b489f8837e3537801e4304b842722dec11f06025ecb239b851db9eff5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 437, + 296, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 437, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 115, + 437, + 296, + 450 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {y}} ^ {v} \\left(\\omega^ {v}\\right) = \\operatorname {s o f t m a x} \\left(\\omega^ {v}\\right) \\left(f _ {\\theta_ {\\mathrm {S E G}}} \\left(f _ {\\theta_ {\\mathrm {E X R}}} \\left(\\mathbf {x} ^ {v}\\right)\\right)\\right),", + "image_path": "52f23ee5fbc02eb9cb3a24c279109634552f125214fdb70c8baf97f194385f5b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "spans": [ + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}^r" + }, + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}^v" + }, + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "text", + "content": " are the categorical distributions for the range image (in " + }, + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "inline_equation", + "content": "\\omega^r" + }, + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "inline_equation", + "content": "U\\times V" + }, + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "text", + "content": " lattice " + }, + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "inline_equation", + "content": "\\Omega^r" + }, + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "text", + "content": ") and voxel (in " + }, + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "inline_equation", + "content": "\\omega^v" + }, + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "inline_equation", + "content": "H\\times W\\times L" + }, + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "text", + "content": " lattice " + }, + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "inline_equation", + "content": "\\Omega^v" + }, + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "text", + "content": ") representations, and " + }, + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{softmax}^{(\\omega)}(.)" + }, + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "text", + "content": " computes the softmax activation at " + }, + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 30, + 458, + 381, + 506 + ], + "type": "text", + "content": ". Based on these outputs, we perform the following cross-view transformations:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 90, + 512, + 378, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 512, + 378, + 533 + ], + "spans": [ + { + "bbox": [ + 90, + 512, + 378, + 533 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {y}} ^ {r} \\left(\\omega^ {r}\\right) = \\operatorname {a r g m a x} \\Psi_ {v \\rightarrow r} ^ {\\left(\\omega^ {r}\\right)} \\left(\\hat {\\mathbf {y}} ^ {v}\\right), \\quad \\mathbf {c} ^ {r} \\left(\\omega^ {r}\\right) = \\max \\Psi_ {v \\rightarrow r} ^ {\\left(\\omega^ {r}\\right)} \\left(\\hat {\\mathbf {y}} ^ {v}\\right), \\tag {2}", + "image_path": "a67ddce9e7372ec70cf4c2d276e220a24f351942286620ad6dfa91ec13c13702.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 92, + 530, + 319, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 530, + 319, + 543 + ], + "spans": [ + { + "bbox": [ + 92, + 530, + 319, + 543 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {y}} ^ {v} \\left(\\omega^ {v}\\right) = \\operatorname {a r g m a x} \\Psi_ {r \\rightarrow v} ^ {\\left(\\omega^ {v}\\right)} \\left(\\hat {\\mathbf {y}} ^ {r}\\right), \\quad \\mathbf {c} ^ {v} \\left(\\omega^ {v}\\right) = \\max \\Psi_ {r \\rightarrow v} ^ {\\left(\\omega^ {v}\\right)} \\left(\\hat {\\mathbf {y}} ^ {r}\\right),", + "image_path": "129eb3f5cc375b69e990e40fb5f99ff7001e8378c0fa804d308ceeb9e449c231.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "spans": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 33, + 549, + 381, + 584 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 549, + 381, + 584 + ], + "spans": [ + { + "bbox": [ + 33, + 549, + 381, + 584 + ], + "type": "text", + "content": "6 Range-to-point projection can lead to information loss [45], which is typically mitigated by post-processing with K-Nearest Neighbors (KNN). In pursuit of efficiency, our IT2 does not use any post-processing during training." + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "spans": [ + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{y}}^r" + }, + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{y}}^v" + }, + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "text", + "content": " represent the transformed hard pseudo label maps, " + }, + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^r" + }, + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^v" + }, + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "text", + "content": " denote the transformed confidence maps for range and voxel representations, respectively, " + }, + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "inline_equation", + "content": "\\varPsi_{v\\rightarrow r}=\\psi_{p\\rightarrow r}\\circ\\psi_{v\\rightarrow p},\\varPsi_{r\\rightarrow v}=\\psi_{p\\rightarrow v}\\circ\\psi_{r\\rightarrow p}" + }, + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "text", + "content": ", and the superscript " + }, + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "text", + "content": " indicates that we only consider the output at " + }, + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 30, + 35, + 380, + 94 + ], + "type": "text", + "content": ". We use the cross-view pseudo labels from Eq. (2) to formulate the loss function to train IT2, as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 35, + 103, + 378, + 126 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 35, + 103, + 378, + 126 + ], + "spans": [ + { + "bbox": [ + 35, + 103, + 378, + 126 + ], + "type": "interline_equation", + "content": "\\ell_ {\\mathrm {I T 2}} \\left(\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\phi_ {\\mathrm {F C N}}}, f _ {\\phi_ {\\mathrm {E X R}}}, f _ {\\phi_ {\\mathrm {S E G}}}\\right) = \\ell_ {\\text {r a n g e}} \\left(\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\phi_ {\\mathrm {F C N}}}, f _ {\\phi_ {\\mathrm {S E G}}}\\right) + \\ell_ {\\text {v o x e l}} \\left(\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\theta_ {\\mathrm {E X R}}}, f _ {\\theta_ {\\mathrm {S E G}}}\\right), \\tag {3}", + "image_path": "ba063a2e87aa8243f6d237142d1b62722e3a9ac2373d71963ede6c56fb9f7beb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 31, + 127, + 231, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 127, + 231, + 137 + ], + "spans": [ + { + "bbox": [ + 31, + 127, + 231, + 137 + ], + "type": "text", + "content": "where we formulate the loss for each view as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 44, + 145, + 378, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 44, + 145, + 378, + 269 + ], + "spans": [ + { + "bbox": [ + 44, + 145, + 378, + 269 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\ell_ {\\mathrm {r a n g e}} (\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\phi_ {\\mathbf {F C N}}}, f _ {\\phi_ {\\mathbf {S E G}}}) = \\frac {1}{| \\mathcal {D} _ {L} | \\times | \\Omega^ {r} |} \\sum_ {i = 1} ^ {| \\mathcal {D} _ {L} |} \\sum_ {\\omega^ {r} \\in \\Omega^ {r}} \\ell (\\mathbf {y} _ {i} ^ {r} (\\omega^ {r}), f _ {\\phi_ {\\mathbf {S E G}}} ^ {(\\omega^ {r})} (f _ {\\phi_ {\\mathbf {F C N}}} (\\mathbf {x} _ {i} ^ {r})) \\\\ + \\frac {1}{\\left| \\mathcal {D} _ {U} \\right| \\times \\left| \\Omega^ {r} \\right|} \\sum_ {i = 1} ^ {\\left| \\mathcal {D} _ {U} \\right|} \\sum_ {\\omega^ {r} \\in \\Omega^ {r}} \\ell \\left(\\tilde {\\mathbf {y}} _ {i} ^ {r} \\left(\\omega^ {r}\\right), f _ {\\phi_ {\\mathrm {S E G}}} ^ {\\left(\\omega^ {r}\\right)} \\left(f _ {\\phi_ {\\mathrm {F C N}}} \\left(\\mathbf {x} _ {i} ^ {r}\\right)\\right), \\right. \\tag {4} \\\\ \\ell_ {\\mathrm {v o x e l}} (\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\theta_ {\\mathbf {E X R}}}, f _ {\\theta_ {\\mathbf {S E G}}}) = \\frac {1}{| \\mathcal {D} _ {L} | \\times | \\Omega^ {v} |} \\sum_ {i = 1} ^ {| \\mathcal {D} _ {L} |} \\sum_ {\\omega^ {v} \\in \\Omega^ {v}} \\ell (\\mathbf {y} _ {i} ^ {v} (\\omega^ {v}), f _ {\\theta_ {\\mathbf {S E G}}} ^ {(\\omega^ {v})} (f _ {\\theta_ {\\mathbf {E X R}}} (\\mathbf {x} _ {i} ^ {v})) \\\\ + \\frac {1}{| \\mathcal {D} _ {U} | \\times | \\Omega^ {v} |} \\sum_ {i = 1} ^ {| \\mathcal {D} _ {U} |} \\sum_ {\\omega^ {v} \\in \\Omega^ {v}} \\ell (\\tilde {\\mathbf {y}} _ {i} ^ {v} (\\omega^ {v}), f _ {\\theta_ {\\mathrm {S E G}}} ^ {(\\omega^ {v})} (f _ {\\theta_ {\\mathrm {E X R}}} (\\mathbf {x} _ {i} ^ {v})), \\\\ \\end{array}", + "image_path": "3fea289a202a0423f2e26bd14102d46c5bcbb8c807c99a16c2ce3c5d2e8de151.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 31, + 277, + 379, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 277, + 379, + 303 + ], + "spans": [ + { + "bbox": [ + 31, + 277, + 379, + 303 + ], + "type": "text", + "content": "with the loss function " + }, + { + "bbox": [ + 31, + 277, + 379, + 303 + ], + "type": "inline_equation", + "content": "\\ell(.)" + }, + { + "bbox": [ + 31, + 277, + 379, + 303 + ], + "type": "text", + "content": " denoting the combination of cross-entropy loss and Lovasz-softmax [3] loss, following [27, 31, 66]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 31, + 318, + 260, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 318, + 260, + 330 + ], + "spans": [ + { + "bbox": [ + 31, + 318, + 260, + 330 + ], + "type": "text", + "content": "3.2 Cross-distribution Contrastive Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 337, + 379, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 337, + 379, + 423 + ], + "spans": [ + { + "bbox": [ + 30, + 337, + 379, + 423 + ], + "type": "text", + "content": "We model the cross distribution of the embedding space using a multi-layer perceptron to generate latent embeddings in parallel with the segmentation head. Specifically, we obtain the range image embedding space via " + }, + { + "bbox": [ + 30, + 337, + 379, + 423 + ], + "type": "inline_equation", + "content": "f_{\\phi_{\\mathbf{PRJ}}}^r: \\mathcal{I} \\to \\mathbb{R}^{U \\times V \\times Z}" + }, + { + "bbox": [ + 30, + 337, + 379, + 423 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 30, + 337, + 379, + 423 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 30, + 337, + 379, + 423 + ], + "type": "text", + "content": " denoting the pixel-wise number of dimensions in the projected space. Similarly, the voxel embedding space is obtained via " + }, + { + "bbox": [ + 30, + 337, + 379, + 423 + ], + "type": "inline_equation", + "content": "f_{\\theta_{\\mathbf{PRJ}}}^v: \\mathcal{K} \\to \\mathbb{R}^{H \\times W \\times L \\times Z}" + }, + { + "bbox": [ + 30, + 337, + 379, + 423 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 30, + 337, + 379, + 423 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 30, + 337, + 379, + 423 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 337, + 379, + 423 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 30, + 337, + 379, + 423 + ], + "type": "text", + "content": " are intermediate feature spaces introduced in Sec. 3.1. We formulate the observed embedding results with:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 430, + 378, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 430, + 378, + 447 + ], + "spans": [ + { + "bbox": [ + 76, + 430, + 378, + 447 + ], + "type": "interline_equation", + "content": "\\mathbf {z} ^ {r} \\left(\\omega^ {r}\\right) = f _ {\\phi_ {\\mathbf {P R J}}} ^ {\\left(\\omega^ {r}\\right)} \\left(f _ {\\phi_ {\\mathbf {F C N}}} \\left(\\mathbf {x} ^ {r}\\right)\\right), \\quad \\mathbf {z} ^ {v} \\left(\\omega^ {v}\\right) = f _ {\\theta_ {\\mathbf {P R J}}} ^ {\\left(\\omega^ {v}\\right)} \\left(f _ {\\theta_ {\\mathbf {E X R}}} \\left(\\mathbf {x} ^ {v}\\right)\\right), \\tag {5}", + "image_path": "a36297be71929dd4d3a7ccf1237cff36193b15a06122699bca86677e995c3d22.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 452, + 379, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 452, + 379, + 475 + ], + "spans": [ + { + "bbox": [ + 30, + 452, + 379, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 30, + 452, + 379, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^r" + }, + { + "bbox": [ + 30, + 452, + 379, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 452, + 379, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^v" + }, + { + "bbox": [ + 30, + 452, + 379, + 475 + ], + "type": "text", + "content": " represent the latent point-wise embeddings of the range and voxel representations, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "text", + "content": "As mentioned in Sec. 2, contrastive learning methods [7, 12, 23, 31, 46] only use a limited subset of actual training samples present in a mini batch, potentially leading to suboptimal convergence with confirmation bias. We propose a novel approach to represent the range and voxel embeddings with class-specific Gaussian Mixture Models (GMMs). Such approach enables the sampling of the most informative embeddings, fostering a comprehensive understanding of the distribution across the feature space of different representations. In practice, we concatenate the embeddings from both range and voxel based on their categories, where such embedding set is defined by " + }, + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^y = \\{(\\mathbf{z},\\mathbf{c},y)|\\mathbf{z} = \\mathbf{z}_i^r (\\omega^r),\\mathbf{c} =" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 9, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 9, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 298, + 9, + 346, + 19 + ], + "type": "text", + "content": "ItTakesTwo" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "spans": [ + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i^r (\\omega^r),y = \\tilde{\\mathbf{y}}_i^r (\\omega^r)\\} \\bigcup \\{(\\mathbf{z},\\mathbf{c},y)|\\mathbf{z} = \\mathbf{z}_i^v (\\omega^v),\\mathbf{c} = \\mathbf{c}_i^v (\\omega^v),y = \\tilde{\\mathbf{y}}_i^v (\\omega^v)\\}" + }, + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "text", + "content": " for the unlabelled data indexed by " + }, + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "inline_equation", + "content": "i\\in \\{1,\\dots,|\\mathcal{D}_U|\\}" + }, + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "text", + "content": ", with the related confidence (i.e., " + }, + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^r" + }, + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^v" + }, + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "text", + "content": ") and pseudo labels (i.e., " + }, + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{y}}^r" + }, + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{y}}^v" + }, + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "text", + "content": ") from Eq. (2) at positions " + }, + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "inline_equation", + "content": "\\omega^r\\in \\Omega^r" + }, + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "inline_equation", + "content": "\\omega^v\\in \\varOmega^v" + }, + { + "bbox": [ + 30, + 34, + 380, + 83 + ], + "type": "text", + "content": ". The class-specific GMMs are defined for both representations by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 76, + 87, + 379, + 116 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 87, + 379, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 87, + 379, + 116 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {\\Gamma^ {y}} (\\mathbf {z} | y) = \\sum_ {m = 1} ^ {M} \\mathbf {P} _ {\\Gamma^ {y}} (m | y) \\mathbf {P} _ {\\Gamma^ {y}} (\\mathbf {z} | m, y) = \\sum_ {m = 1} ^ {M} \\pi_ {m} ^ {y} \\mathcal {N} (\\mathbf {z}; \\boldsymbol {\\mu} _ {m} ^ {y}, \\boldsymbol {\\Sigma} _ {m} ^ {y}), \\tag {6}", + "image_path": "f03f1c123769713ce302266e5baf062cebe1814ce43ad87ad0979e431a4c2cfa.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 120, + 379, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 120, + 379, + 181 + ], + "spans": [ + { + "bbox": [ + 30, + 120, + 379, + 181 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 30, + 120, + 379, + 181 + ], + "type": "inline_equation", + "content": "\\pmb{\\pi}_{m}^{y} = \\mathbf{P}_{\\varGamma^{y}}(m|y)" + }, + { + "bbox": [ + 30, + 120, + 379, + 181 + ], + "type": "text", + "content": " denotes the prior probability for the " + }, + { + "bbox": [ + 30, + 120, + 379, + 181 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 30, + 120, + 379, + 181 + ], + "type": "text", + "content": "-th GMM component of class " + }, + { + "bbox": [ + 30, + 120, + 379, + 181 + ], + "type": "inline_equation", + "content": "y \\in \\{0,1\\}^Y" + }, + { + "bbox": [ + 30, + 120, + 379, + 181 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 30, + 120, + 379, + 181 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{m}^{y} \\in \\mathbb{R}^{Z}" + }, + { + "bbox": [ + 30, + 120, + 379, + 181 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 120, + 379, + 181 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_{m}^{y} \\in \\mathbb{R}^{Z \\times Z}" + }, + { + "bbox": [ + 30, + 120, + 379, + 181 + ], + "type": "text", + "content": " represent the mean and covariance of the corresponding GMM component and we fix " + }, + { + "bbox": [ + 30, + 120, + 379, + 181 + ], + "type": "inline_equation", + "content": "M = 5" + }, + { + "bbox": [ + 30, + 120, + 379, + 181 + ], + "type": "text", + "content": " for all experiments. The GMM parameter optimisation in Eq. (6) relies on the EM [11] algorithm using the training set " + }, + { + "bbox": [ + 30, + 120, + 379, + 181 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{y}" + }, + { + "bbox": [ + 30, + 120, + 379, + 181 + ], + "type": "text", + "content": " defined above with the following steps:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 183, + 352, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 183, + 352, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 183, + 352, + 212 + ], + "type": "text", + "content": "E step: " + }, + { + "bbox": [ + 61, + 183, + 352, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{n,m}^{y} = \\frac{\\pi_{m}^{y}\\mathcal{N}(\\mathbf{z}_{n}\\mid\\boldsymbol{\\mu}_{m}^{y},\\boldsymbol{\\Sigma}_{m}^{y})}{\\sum_{m}^{M}\\pi_{m}^{y}\\mathcal{N}(\\mathbf{z}_{n}\\mid\\boldsymbol{\\mu}_{m}^{y},\\boldsymbol{\\Sigma}_{m}^{y})}" + }, + { + "bbox": [ + 61, + 183, + 352, + 212 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 61, + 183, + 352, + 212 + ], + "type": "inline_equation", + "content": "(\\mathbf{z}_n,\\mathbf{c}_n,y_n)\\in \\mathcal{G}^y" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 214, + 378, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 214, + 378, + 242 + ], + "spans": [ + { + "bbox": [ + 61, + 214, + 378, + 242 + ], + "type": "text", + "content": "M step: " + }, + { + "bbox": [ + 61, + 214, + 378, + 242 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{m}^{y} = \\frac{1}{|\\mathcal{G}^{y}|}\\sum_{(\\mathbf{z}_{n},\\mathbf{c}_{n},y_{n})\\in \\mathcal{G}^{y}}\\mathbf{c}_{n}\\mathbf{q}_{n,m}^{y_{n}}\\mathbf{z}_{n}" + }, + { + "bbox": [ + 61, + 214, + 378, + 242 + ], + "type": "text", + "content": " (7)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 109, + 246, + 348, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 246, + 348, + 275 + ], + "spans": [ + { + "bbox": [ + 109, + 246, + 348, + 275 + ], + "type": "interline_equation", + "content": "\\boldsymbol{\\Sigma}_{m}^{y} = \\frac{1}{|\\mathcal{G}^{y}|}\\sum_{(\\mathbf{z}_{n},\\mathbf{c}_{n},y_{n})\\in \\mathcal{G}^{y}}\\mathbf{c}_{n}\\mathbf{q}_{n,m}^{y_{n}}(\\mathbf{z}_{n} - \\boldsymbol{\\mu}_{m}^{y_{n}})\\bigl(\\mathbf{z}_{n} - \\boldsymbol{\\mu}_{m}^{y_{n}}\\bigr)^{\\intercal},", + "image_path": "f5171e3372bd2a345db1370776cb70e753adee13a95daac2a5fa410597c4605d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "spans": [ + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^y" + }, + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "text", + "content": " represents the set of embedding samples that share the same label " + }, + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "text", + "content": ". The posterior " + }, + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{n,m}^y" + }, + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "text", + "content": " estimates the responsibility of the " + }, + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "inline_equation", + "content": "n^{th}" + }, + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "text", + "content": " embedding " + }, + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "text", + "content": " from class " + }, + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^y" + }, + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "text", + "content": " being assigned to the " + }, + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "text", + "content": "-th component. The confidence " + }, + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_n" + }, + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "text", + "content": " represents the potential correctness of the pseudo label from Eq. (2), which is used as a weighting factor to reduce the effect of low-confidence segmentation results [8,9] in the estimation of the GMM parameters. Note that we maintain a uniform weight of " + }, + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "inline_equation", + "content": "\\pi_{m}^{y} = 1 / M" + }, + { + "bbox": [ + 30, + 279, + 379, + 375 + ], + "type": "text", + "content": " for each component [43, 57] to prevent one component from dominating others." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 30, + 386, + 379, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 386, + 379, + 543 + ], + "spans": [ + { + "bbox": [ + 30, + 386, + 379, + 543 + ], + "type": "text", + "content": "GMM-based Prototypical Contrastive Learning. The class-specific GMMs from Eq. (7) are used to generate a set of highly representative embeddings for the contrastive learning process explained next. These representative class-specific embeddings is denoted by " + }, + { + "bbox": [ + 30, + 386, + 379, + 543 + ], + "type": "inline_equation", + "content": "S^y = \\{\\mathbf{s}|\\mathbf{s}\\sim \\mathbf{P}_{\\Gamma^y}(\\mathbf{z}|y)\\}" + }, + { + "bbox": [ + 30, + 386, + 379, + 543 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 30, + 386, + 379, + 543 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\Gamma^y}(\\mathbf{z}|y)" + }, + { + "bbox": [ + 30, + 386, + 379, + 543 + ], + "type": "text", + "content": " being the class-specific GMM defined in Eq. (6). Subsequently, these generated virtual prototypes are utilised to guide the contrastive learning for the incoming embedding features from both range and voxel representations [54], without joining the backpropagation computation graph. We define the set of incoming features with the anchor set " + }, + { + "bbox": [ + 30, + 386, + 379, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{A} = \\{(\\mathbf{z},y)\\}" + }, + { + "bbox": [ + 30, + 386, + 379, + 543 + ], + "type": "text", + "content": " that is formed similarly to how we form the set " + }, + { + "bbox": [ + 30, + 386, + 379, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^y" + }, + { + "bbox": [ + 30, + 386, + 379, + 543 + ], + "type": "text", + "content": " without including the confidence value. The samples in " + }, + { + "bbox": [ + 30, + 386, + 379, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 30, + 386, + 379, + 543 + ], + "type": "text", + "content": " are selectively chosen for backpropagation [54], but the informative samples in " + }, + { + "bbox": [ + 30, + 386, + 379, + 543 + ], + "type": "inline_equation", + "content": "S^y" + }, + { + "bbox": [ + 30, + 386, + 379, + 543 + ], + "type": "text", + "content": " contribute to a comprehensive understanding of the cross-distribution for both representations. The contrastive learning relies on the minimisation of the following loss:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 547, + 378, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 547, + 378, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 547, + 378, + 583 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {c r o s s}} \\left(\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\phi_ {\\mathbf {P R J}}}, f _ {\\phi_ {\\mathbf {F C N}}}, f _ {\\theta_ {\\mathbf {P R J}}}, f _ {\\theta_ {\\mathbf {E X R}}} \\right) = \\sum_ {(\\mathbf {z}, y) \\in \\mathcal {A}} \\sum_ {\\mathbf {s} \\in \\mathcal {S} ^ {y}} \\frac {\\exp (\\mathbf {z} \\cdot \\mathbf {s} / \\tau)}{\\exp (\\mathbf {z} \\cdot \\mathbf {s} / \\tau) + \\sum_ {\\mathbf {s} ^ {-} \\in \\bar {\\mathcal {S}} ^ {y}} \\exp (\\mathbf {z} \\cdot \\mathbf {s} ^ {-} / \\tau)}, \\tag {8}", + "image_path": "ad6f300de7635ecf33fb0e882bc04d580b790d463931da443eac15ed79e7e569.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 31, + 34, + 304, + 46 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 34, + 304, + 46 + ], + "spans": [ + { + "bbox": [ + 31, + 34, + 304, + 46 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 31, + 34, + 304, + 46 + ], + "type": "inline_equation", + "content": "\\bar{S}^y" + }, + { + "bbox": [ + 31, + 34, + 304, + 46 + ], + "type": "text", + "content": " is the set of virtual negative embeddings for class " + }, + { + "bbox": [ + 31, + 34, + 304, + 46 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 31, + 34, + 304, + 46 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 65, + 281, + 77 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 65, + 281, + 77 + ], + "spans": [ + { + "bbox": [ + 31, + 65, + 281, + 77 + ], + "type": "text", + "content": "3.3 Representation Specific Data Augmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 86, + 381, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 86, + 381, + 207 + ], + "spans": [ + { + "bbox": [ + 30, + 86, + 381, + 207 + ], + "type": "text", + "content": "Following [39, 61], we apply spatial data augmentations to the pseudo-labeling strategy in Eq. (2) to increase the robustness of our approach to complex scenes. After experimenting with classic data augmentation processes (e.g., CutMix [63] and LaserMix [27]), we notice that the application of distinct spatial augmentations to different viewpoints in our IT2 framework results in more effective generalization. We have empirically found the multi-boxes CutMix [63] augmentation for the range image and single-inclination LaserMix [27] augmentation for the voxel representation yield the most robust results. Although this representation-specific data augmentation is not our main contribution, it is shown empirically to produce relevant improvements. Please see details in supplementary Sec. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 224, + 143, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 224, + 143, + 237 + ], + "spans": [ + { + "bbox": [ + 31, + 224, + 143, + 237 + ], + "type": "text", + "content": "3.4 Overall Training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 31, + 245, + 335, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 245, + 335, + 258 + ], + "spans": [ + { + "bbox": [ + 31, + 245, + 335, + 258 + ], + "type": "text", + "content": "The overall training of our IT2 minimises the following loss function:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 268, + 378, + 305 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 268, + 378, + 305 + ], + "spans": [ + { + "bbox": [ + 33, + 268, + 378, + 305 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} \\left(\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, \\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\phi_ {\\mathrm {F C N}}}, f _ {\\phi_ {\\mathrm {E X R}}}, f _ {\\phi_ {\\mathrm {S E G}}} f _ {\\phi_ {\\mathrm {P R J}}}, f _ {\\phi_ {\\mathrm {P R J}}}\\right) = \\\\ \\ell_ {\\mathrm {I T 2}} \\left(\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\phi_ {\\mathbf {F C N}}}, f _ {\\phi_ {\\mathbf {E X R}}}, f _ {\\phi_ {\\mathbf {S E G}}}\\right) + \\ell_ {\\mathrm {c r o s s}} \\left(\\mathcal {D} _ {L}, \\mathcal {D} _ {U}, f _ {\\phi_ {\\mathbf {P R J}}}, f _ {\\phi_ {\\mathbf {F C N}}}, f _ {\\theta_ {\\mathbf {P R J}}}, f _ {\\theta_ {\\mathbf {E X R}}}\\right), \\tag {9} \\\\ \\end{array}", + "image_path": "3bbf0254a0c7018bcf4e19d6ed0fbb38da5dedf85da648cbdb1ac67cb16687d5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 31, + 306, + 379, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 306, + 379, + 331 + ], + "spans": [ + { + "bbox": [ + 31, + 306, + 379, + 331 + ], + "type": "text", + "content": "where the entire training pipeline is performed in an end-to-end manner, where in each iteration, we minimise the loss from " + }, + { + "bbox": [ + 31, + 306, + 379, + 331 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{IT2}}(.)" + }, + { + "bbox": [ + 31, + 306, + 379, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 31, + 306, + 379, + 331 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{cross}}(.)" + }, + { + "bbox": [ + 31, + 306, + 379, + 331 + ], + "type": "text", + "content": " simultaneously." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 31, + 345, + 124, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 345, + 124, + 358 + ], + "spans": [ + { + "bbox": [ + 31, + 345, + 124, + 358 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 31, + 371, + 161, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 371, + 161, + 383 + ], + "spans": [ + { + "bbox": [ + 31, + 371, + 161, + 383 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 30, + 391, + 380, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 391, + 380, + 547 + ], + "spans": [ + { + "bbox": [ + 30, + 391, + 380, + 547 + ], + "type": "text", + "content": "We evaluate our approach on three benchmarks, including nuScenes [4], SemanticKITTI [2] and ScribbleKITTI [52]. The nuScenes [4] dataset has 29,130 training scans and 6,019 validation scans. Each LiDAR scan contains 32 beams with " + }, + { + "bbox": [ + 30, + 391, + 380, + 547 + ], + "type": "inline_equation", + "content": "[-30^{\\circ},10^{\\circ}]" + }, + { + "bbox": [ + 30, + 391, + 380, + 547 + ], + "type": "text", + "content": " vertical field of views (FOV). The dataset has 23 categories, but following previous works [27,66], we merge them into 16 for training and validation. The SemanticKITTI [2] and its extension ScribbleKITTI [52] are popular laser-based segmentation benchmarks, where the latter one is based on line-scribbles labelling to reduce the annotation effort. The SemanticKITTI [2] dataset has 19 categories for semantic segmentation, where every LiDAR scan contains 64 beams with " + }, + { + "bbox": [ + 30, + 391, + 380, + 547 + ], + "type": "inline_equation", + "content": "[3^{\\circ}, -25^{\\circ}]" + }, + { + "bbox": [ + 30, + 391, + 380, + 547 + ], + "type": "text", + "content": " vertical FOV. There are 22 sequences of LiDAR scans in the dataset, where sequences 0 to 9 are used for training and validation. Following [23,27], we use 19,130 scans (sequences 0 to 7 and sequence 9) for training and 4,071 scans from sequence 8 for validation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 30, + 547, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 547, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 547, + 380, + 584 + ], + "type": "text", + "content": "Sampling strategies. Previous works [23, 27, 31] propose different sampling strategies to select the labelled data for the semi-supervised learning setting, including: 1) uniform [27] sampling of the labelled data across all the training" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 297, + 9, + 346, + 19 + ], + "type": "text", + "content": "ItTakesTwo" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 35, + 76, + 379, + 260 + ], + "blocks": [ + { + "bbox": [ + 32, + 32, + 379, + 66 + ], + "lines": [ + { + "bbox": [ + 32, + 32, + 379, + 66 + ], + "spans": [ + { + "bbox": [ + 32, + 32, + 379, + 66 + ], + "type": "text", + "content": "Table 1: Uniform sampling results across various benchmarks, where we follow [27] to utilise half feature dimension and lower resolution in Cylinder3D [66] for all the experiments. Best results are highlighted in red." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 35, + 76, + 379, + 260 + ], + "lines": [ + { + "bbox": [ + 35, + 76, + 379, + 260 + ], + "spans": [ + { + "bbox": [ + 35, + 76, + 379, + 260 + ], + "type": "table", + "html": "
Repr.MethodnuScenes [4]SemanticKITTI [2]ScribbleKITTI [52]
1%10%20%50%1%10%20%50%1%10%20%50%
Rangesup.38.357.562.767.636.252.255.957.233.147.749.952.5
MT [50]42.160.465.469.437.553.156.157.434.249.851.653.3
CBST [68]40.960.564.369.339.953.456.156.935.750.752.754.6
CPS [6]40.760.864.968.036.552.356.357.433.750.052.854.6
LaserMix [27]49.568.270.673.043.458.859.461.438.354.455.658.7
IT2 (Ours)56.571.373.474.051.960.361.762.146.657.157.358.6
Voxelsup.50.965.966.671.245.456.157.858.739.248.052.153.8
MT [50]51.666.067.171.745.457.159.260.041.050.152.853.9
CBST [68]53.066.569.671.648.858.359.459.741.550.653.354.5
CPS [6]52.966.370.072.546.758.759.660.541.451.853.954.8
LaserMix [27]55.369.971.873.250.660.061.962.344.253.755.156.8
IT2 (Ours)57.572.173.674.152.061.462.162.547.956.757.558.3
", + "image_path": "637f723ca22bc082fe4d2ea7a1ce64b966416418a589cd4373ab8d3b8b6d03a4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 32, + 281, + 379, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 281, + 379, + 353 + ], + "spans": [ + { + "bbox": [ + 32, + 281, + 379, + 353 + ], + "type": "text", + "content": "sequences, 2) partial [23] sampling of the labelled data only from certain sequences, and 3) sampling of significant [31] labelled frames from the sequences. In the experiment, we examine our approach under all these sampling protocols. Metrics. We measure performance with mean intersection over union (mIoU). In each benchmark, we use the sampling strategies above to label " + }, + { + "bbox": [ + 32, + 281, + 379, + 353 + ], + "type": "inline_equation", + "content": "n\\%" + }, + { + "bbox": [ + 32, + 281, + 379, + 353 + ], + "type": "text", + "content": " of the training set, with the rest " + }, + { + "bbox": [ + 32, + 281, + 379, + 353 + ], + "type": "inline_equation", + "content": "(100 - n)\\%" + }, + { + "bbox": [ + 32, + 281, + 379, + 353 + ], + "type": "text", + "content": " being unlabelled." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 32, + 354, + 379, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 354, + 379, + 461 + ], + "spans": [ + { + "bbox": [ + 32, + 354, + 379, + 461 + ], + "type": "text", + "content": "Implementation details. In the range view, we use the FIDNet architecture [65] with ResNet34 [19] backbone. The input range image has size " + }, + { + "bbox": [ + 32, + 354, + 379, + 461 + ], + "type": "inline_equation", + "content": "32 \\times 1920" + }, + { + "bbox": [ + 32, + 354, + 379, + 461 + ], + "type": "text", + "content": " for nuScenes [4] and " + }, + { + "bbox": [ + 32, + 354, + 379, + 461 + ], + "type": "inline_equation", + "content": "64 \\times 2048" + }, + { + "bbox": [ + 32, + 354, + 379, + 461 + ], + "type": "text", + "content": " for SemanticKITTI [2] and ScirbbleKITTI [52], following [27]. In the voxel representation, we use the Cylinder3D architecture [66], where we fix the intermediate layer at 16 dimensions with smaller resolution as specified in [27] for the uniform sampling experiments to enable a fair comparison. For contrastive learning, we build a 3-layer projector [24,54]. The projector outputs an embedding map with 64 channels, where we apply the temperature factor " + }, + { + "bbox": [ + 32, + 354, + 379, + 461 + ], + "type": "inline_equation", + "content": "\\tau = .1" + }, + { + "bbox": [ + 32, + 354, + 379, + 461 + ], + "type": "text", + "content": " in Eq. (8). See supplementary Sec. 3 and Sec. 7 for more details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 468, + 260, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 468, + 260, + 480 + ], + "spans": [ + { + "bbox": [ + 33, + 468, + 260, + 480 + ], + "type": "text", + "content": "4.2 Results on Different Partition Protocols" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 487, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 487, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 487, + 379, + 583 + ], + "type": "text", + "content": "Uniform sampling results. We first compare our method using the uniform sampling benchmark [27] in Tab. 1, with the experimental setup from [27] (i.e., architecture, split ID and the resolution of input representations). Our framework outperforms the previous SOTA Lasermix [27] by a large margin for almost all partition protocols. For example, we observe " + }, + { + "bbox": [ + 33, + 487, + 379, + 583 + ], + "type": "inline_equation", + "content": "\\approx 3.4\\%" + }, + { + "bbox": [ + 33, + 487, + 379, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 33, + 487, + 379, + 583 + ], + "type": "inline_equation", + "content": "\\approx 2.0\\%" + }, + { + "bbox": [ + 33, + 487, + 379, + 583 + ], + "type": "text", + "content": " improvements on average for range and voxel representations, respectively. Notably, we highlight that our approach yields significant improvements for the " + }, + { + "bbox": [ + 33, + 487, + 379, + 583 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 33, + 487, + 379, + 583 + ], + "type": "text", + "content": " partition protocol in the line-scribble dataset [52], providing " + }, + { + "bbox": [ + 33, + 487, + 379, + 583 + ], + "type": "inline_equation", + "content": "8.28\\%" + }, + { + "bbox": [ + 33, + 487, + 379, + 583 + ], + "type": "text", + "content": " improvement in the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 98, + 363, + 182 + ], + "blocks": [ + { + "bbox": [ + 32, + 32, + 379, + 88 + ], + "lines": [ + { + "bbox": [ + 32, + 32, + 379, + 88 + ], + "spans": [ + { + "bbox": [ + 32, + 32, + 379, + 88 + ], + "type": "text", + "content": "Table 2: Partial & significant sampling results in the SemanticKITTI [2] benchmark. We compare with previous SOTA based on Cylinder3D [66] architecture, where TTA denotes test time augmentation. " + }, + { + "bbox": [ + 32, + 32, + 379, + 88 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 32, + 32, + 379, + 88 + ], + "type": "text", + "content": " indicates that we re-implement results based on the supported checkpoints from their GitHub - please see Supplementary Sec. 1 for more details. Best results in each evaluation protocol are highlighted in red." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 98, + 363, + 182 + ], + "lines": [ + { + "bbox": [ + 52, + 98, + 363, + 182 + ], + "spans": [ + { + "bbox": [ + 52, + 98, + 363, + 182 + ], + "type": "table", + "html": "
TTAMethodSemanticKITTI(par.)MethodSemanticKITTI(sig.)
5%10%20%40%5%10%20%40%
XGPC† [23]39.945.955.757.1lim3D [31]56.859.660.561.4
IT2 (Ours)43.448.156.560.9IT2 (Ours)60.363.364.064.8
✓GPC† [23]40.246.356.557.6lim3D [31]59.562.263.163.3
IT2 (Ours)43.849.258.862.4IT2 (Ours)61.764.464.765.8
", + "image_path": "275f85492edccb84f1b28bbacda925eb05360cabf20f6551e69be8ccdf89a89b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 32, + 204, + 256, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 204, + 256, + 216 + ], + "spans": [ + { + "bbox": [ + 32, + 204, + 256, + 216 + ], + "type": "text", + "content": "range image and " + }, + { + "bbox": [ + 32, + 204, + 256, + 216 + ], + "type": "inline_equation", + "content": "3.66\\%" + }, + { + "bbox": [ + 32, + 204, + 256, + 216 + ], + "type": "text", + "content": " in the voxel representation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 217, + 379, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 217, + 379, + 336 + ], + "spans": [ + { + "bbox": [ + 33, + 217, + 379, + 336 + ], + "type": "text", + "content": "Partial and significant sampling results. In Tab. 2, we compare our results with GPC [23] and lim3D [31] in partial sampling (par.) and significant sampling (sig.), based on both original performance and test time augmentation (TTA). In the partial sampling protocol7, our approach yields superior results across all the labelled data situations for both with and without TTA. Notably, our results surpass GPC [23] by " + }, + { + "bbox": [ + 33, + 217, + 379, + 336 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 33, + 217, + 379, + 336 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 33, + 217, + 379, + 336 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 33, + 217, + 379, + 336 + ], + "type": "text", + "content": " in the limited " + }, + { + "bbox": [ + 33, + 217, + 379, + 336 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 33, + 217, + 379, + 336 + ], + "type": "text", + "content": " labelled data condition, demonstrating its good generalisation capability. In the significant sampling results, our method also achieves better performance than lim3D [31] for all the partition protocols. For example, our results are better than the previous SOTA by about " + }, + { + "bbox": [ + 33, + 217, + 379, + 336 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 33, + 217, + 379, + 336 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 33, + 217, + 379, + 336 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 33, + 217, + 379, + 336 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 33, + 217, + 379, + 336 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 33, + 217, + 379, + 336 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 33, + 217, + 379, + 336 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 33, + 217, + 379, + 336 + ], + "type": "text", + "content": " labelled data with TTA, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 345, + 136, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 345, + 136, + 357 + ], + "spans": [ + { + "bbox": [ + 33, + 345, + 136, + 357 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 365, + 379, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 365, + 379, + 531 + ], + "spans": [ + { + "bbox": [ + 33, + 365, + 379, + 531 + ], + "type": "text", + "content": "Improvements from our contributions. Table 3 presents our ablation study on the nuScenes [4] and ScribbleKITTI [52] datasets across various partition protocols. We investigate the roles of the IT2 architecture, cross-distribution contrastive learning (Ctrs.), and representation-related data augmentation (Aug.). The baseline utilises network perturbation through single representations [6] from [27]. Our IT2 architecture results in improvements of approximately " + }, + { + "bbox": [ + 33, + 365, + 379, + 531 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 33, + 365, + 379, + 531 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 33, + 365, + 379, + 531 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 33, + 365, + 379, + 531 + ], + "type": "text", + "content": " for the voxel representation in " + }, + { + "bbox": [ + 33, + 365, + 379, + 531 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 33, + 365, + 379, + 531 + ], + "type": "text", + "content": " labelled data for nuScenes [4] and ScribbleKITTI [52], respectively. The Ctrs. further improves by " + }, + { + "bbox": [ + 33, + 365, + 379, + 531 + ], + "type": "inline_equation", + "content": "\\approx 2.4\\%" + }, + { + "bbox": [ + 33, + 365, + 379, + 531 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 33, + 365, + 379, + 531 + ], + "type": "inline_equation", + "content": "\\approx 1.1\\%" + }, + { + "bbox": [ + 33, + 365, + 379, + 531 + ], + "type": "text", + "content": " for range and voxel representations in both datasets in " + }, + { + "bbox": [ + 33, + 365, + 379, + 531 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 33, + 365, + 379, + 531 + ], + "type": "text", + "content": " labelled data. Such enhancements show the effectiveness of our cross-distribution prototypes and contrastive learning strategy. Furthermore, Aug. results in improvements of " + }, + { + "bbox": [ + 33, + 365, + 379, + 531 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 33, + 365, + 379, + 531 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 33, + 365, + 379, + 531 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 33, + 365, + 379, + 531 + ], + "type": "text", + "content": " for range and voxel representations in the ScribbleKITTI [52] dataset for " + }, + { + "bbox": [ + 33, + 365, + 379, + 531 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 33, + 365, + 379, + 531 + ], + "type": "text", + "content": " labelled data, demonstrating the better generalisation brought by the representation-specific data augmentation." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 9, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 9, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 298, + 9, + 346, + 19 + ], + "type": "text", + "content": "ItTakesTwo" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 378, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 378, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 378, + 19 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 35, + 538, + 379, + 583 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 35, + 538, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 35, + 538, + 379, + 583 + ], + "type": "text", + "content": "The original GPC paper [23] calculates the mIoU differently compared to common practices [27, 31, 66]. The reported performance is based on the re-implementation from the official supported checkpoints of their GitHub, where more details are in Supplementary Sec. 1." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 35, + 84, + 377, + 211 + ], + "blocks": [ + { + "bbox": [ + 31, + 32, + 380, + 78 + ], + "lines": [ + { + "bbox": [ + 31, + 32, + 380, + 78 + ], + "spans": [ + { + "bbox": [ + 31, + 32, + 380, + 78 + ], + "type": "text", + "content": "Table 3: Ablation study in the nuScenes [4] and ScribbleKITTI [52] datasets via uniform sampling. IT2 represents our architecture, Ctrs. denotes our cross-distribution contrastive learning and Aug. is our representation-specific data augmentation. The yellow rows denote our single representation baseline with CPS [6] from [27]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 35, + 84, + 377, + 211 + ], + "lines": [ + { + "bbox": [ + 35, + 84, + 377, + 211 + ], + "spans": [ + { + "bbox": [ + 35, + 84, + 377, + 211 + ], + "type": "table", + "html": "
Repr.IT2 Ctrs. Aug.nuScenes [4]ScribbleKITTI [52]
10%20%50%10%20%50%
RangeXXX60.864.968.050.052.854.6
✓XX67.3 (6.5↑)70.8 (5.9↑)72.6 (4.6↑)53.2 (3.2↑)55.0 (2.2↑)56.2 (1.6↑)
✓✓X70.3 (9.5↑)72.6 (7.7↑)73.5 (5.5↑)54.9 (4.9↑)56.1 (3.3↑)57.2 (2.6↑)
✓✓✓71.3 (10.5↑)73.4 (8.5↑)74.0 (6.0↑)57.1 (7.1↑)57.3 (4.5↑)58.6 (4.0↑)
VoxelXXX66.370.072.551.853.954.8
✓XX70.4 (4.1↑)71.9 (1.9↑)73.1 (0.6↑)54.3 (2.5↑)55.6 (2.7↑)56.6 (1.8↑)
✓✓X71.6 (5.3↑)72.9 (2.9↑)73.9 (1.4↑)55.3 (3.5↑)56.5 (3.6↑)57.6 (2.8↑)
✓✓✓72.1 (5.8↑)73.6 (3.6↑)74.1 (1.6↑)56.7 (4.9↑)57.5 (4.6↑)58.3 (3.5↑)
", + "image_path": "c9253c28f32cb00ce96b0da69eb8b1ab73b44b0ddf912c93723b345497593ca8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 31, + 218, + 203, + 305 + ], + "blocks": [ + { + "bbox": [ + 31, + 218, + 203, + 305 + ], + "lines": [ + { + "bbox": [ + 31, + 218, + 203, + 305 + ], + "spans": [ + { + "bbox": [ + 31, + 218, + 203, + 305 + ], + "type": "image", + "image_path": "d42650b2c6538733dd1a8509370f8d1a5b6f5ab41155658c657037dac6d51bb6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 31, + 308, + 204, + 341 + ], + "lines": [ + { + "bbox": [ + 31, + 308, + 204, + 341 + ], + "spans": [ + { + "bbox": [ + 31, + 308, + 204, + 341 + ], + "type": "text", + "content": "Fig. 3: Ablation study of temperature value " + }, + { + "bbox": [ + 31, + 308, + 204, + 341 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 31, + 308, + 204, + 341 + ], + "type": "text", + "content": " from Eq. (8) in nuScenes [4] dataset under " + }, + { + "bbox": [ + 31, + 308, + 204, + 341 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 31, + 308, + 204, + 341 + ], + "type": "text", + "content": " labelled partition protocol." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 207, + 265, + 377, + 338 + ], + "blocks": [ + { + "bbox": [ + 204, + 219, + 377, + 263 + ], + "lines": [ + { + "bbox": [ + 204, + 219, + 377, + 263 + ], + "spans": [ + { + "bbox": [ + 204, + 219, + 377, + 263 + ], + "type": "text", + "content": "Table 4: Contrastive Learning study in nuScenes [4], where ContrasSeg [54] is re-implemented from the SOTA image (pixel-level) semantic segmentation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 207, + 265, + 377, + 338 + ], + "lines": [ + { + "bbox": [ + 207, + 265, + 377, + 338 + ], + "spans": [ + { + "bbox": [ + 207, + 265, + 377, + 338 + ], + "type": "table", + "html": "
MethodRepr.nuScenes [4]
10%20%
ContrasSeg [54]range70.372.3
voxel71.272.8
Oursrange71.3 (1.0↑)73.4 (1.1↑)
voxel72.1 (0.9↑)73.6 (0.8↑)
", + "image_path": "c0645e36836fea1bb72101f8768681c20c3f2c29c038b62a288b1864e5b2a4fd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 368, + 379, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 368, + 379, + 571 + ], + "spans": [ + { + "bbox": [ + 30, + 368, + 379, + 571 + ], + "type": "text", + "content": "Cross-distribution contrastive learning. We investigate the impact of the hyperparameter " + }, + { + "bbox": [ + 30, + 368, + 379, + 571 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 30, + 368, + 379, + 571 + ], + "type": "text", + "content": " from Eq. (8) in Fig. 3 using the " + }, + { + "bbox": [ + 30, + 368, + 379, + 571 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 30, + 368, + 379, + 571 + ], + "type": "text", + "content": " labelled protocol on the nuScenes [4] dataset. Our observations indicate that " + }, + { + "bbox": [ + 30, + 368, + 379, + 571 + ], + "type": "inline_equation", + "content": "\\tau = 0.10" + }, + { + "bbox": [ + 30, + 368, + 379, + 571 + ], + "type": "text", + "content": " yields the best results, so we fix this value for all experiments. In Tab. 4, we compare our approach with our re-implementation of the label-guided pixel-wise contrastive learning [54] based on our IT2 architecture with different ratio of labelled data, where our method shows " + }, + { + "bbox": [ + 30, + 368, + 379, + 571 + ], + "type": "inline_equation", + "content": "\\approx 1\\%" + }, + { + "bbox": [ + 30, + 368, + 379, + 571 + ], + "type": "text", + "content": " improvement for both representations, demonstrating the effectiveness of our proposed cross-distribution contrastive learning. Representation-specific augmentation. In Fig. 4, we have demonstrated the augmentation results for each representation in the ScribbleKITTI [52] dataset, where 'multi' denotes multi-box [63], and '1-inc' represents single inclination setups [27]. In our framework, one representation's results rely on the other's generalisation. For instance, '(multi) CutMix' reaches " + }, + { + "bbox": [ + 30, + 368, + 379, + 571 + ], + "type": "inline_equation", + "content": "57.1\\%" + }, + { + "bbox": [ + 30, + 368, + 379, + 571 + ], + "type": "text", + "content": " mIoU with '(1-inc) LaserMix,' improving by " + }, + { + "bbox": [ + 30, + 368, + 379, + 571 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 30, + 368, + 379, + 571 + ], + "type": "text", + "content": " compared to the results with 'LaserMix.' Meanwhile, '(multi) CutMix' combined with '(1-inc) LaserMix' improves by " + }, + { + "bbox": [ + 30, + 368, + 379, + 571 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 30, + 368, + 379, + 571 + ], + "type": "text", + "content": " in the range and " + }, + { + "bbox": [ + 30, + 368, + 379, + 571 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 30, + 368, + 379, + 571 + ], + "type": "text", + "content": " in voxel representations compared to LaserMix [27], highlighting the improved generalisation benefits from our augmentation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 31, + 571, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 571, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 31, + 571, + 379, + 583 + ], + "type": "text", + "content": "Ablation study for other representations. In Tab. 5, we conduct abla" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 34, + 33, + 205, + 126 + ], + "blocks": [ + { + "bbox": [ + 34, + 33, + 205, + 126 + ], + "lines": [ + { + "bbox": [ + 34, + 33, + 205, + 126 + ], + "spans": [ + { + "bbox": [ + 34, + 33, + 205, + 126 + ], + "type": "image", + "image_path": "cc11fcdc44998bfbbecc11735861c931911ebd5bc1dd6af6a279196d2897c507.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 33, + 382, + 89 + ], + "lines": [ + { + "bbox": [ + 206, + 33, + 382, + 89 + ], + "spans": [ + { + "bbox": [ + 206, + 33, + 382, + 89 + ], + "type": "text", + "content": "Table 5: Ablation study for the bird's eye view (BEV) experiment in the nuScenes [4], where results are from PolarNet [64]. The baseline relies on single view [6], followed by BEV with range image or voxel grid." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 210, + 89, + 378, + 177 + ], + "blocks": [ + { + "bbox": [ + 210, + 89, + 378, + 177 + ], + "lines": [ + { + "bbox": [ + 210, + 89, + 378, + 177 + ], + "spans": [ + { + "bbox": [ + 210, + 89, + 378, + 177 + ], + "type": "table", + "html": "
Repr.BEV (PolarNet [64])
10%20%50%
sup.58.563.968.4
w/ bev62.4 (3.9↑)66.7 (2.8↑)70.0 (1.6↑)
w/ range64.8 (6.3↑)67.9 (4.0↑)70.6 (2.2↑)
w/ voxel66.3 (7.8↑)69.1 (5.2↑)71.6 (3.2↑)
", + "image_path": "b039b272353df8f458ace7117515cc0af069724c1fc9fa3c59c8941db585589c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 34, + 199, + 120, + 282 + ], + "blocks": [ + { + "bbox": [ + 31, + 135, + 206, + 179 + ], + "lines": [ + { + "bbox": [ + 31, + 135, + 206, + 179 + ], + "spans": [ + { + "bbox": [ + 31, + 135, + 206, + 179 + ], + "type": "text", + "content": "Fig. 4: Ablation study for the repr.-specific augmentations in ScribbleKITTI [52] with " + }, + { + "bbox": [ + 31, + 135, + 206, + 179 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 31, + 135, + 206, + 179 + ], + "type": "text", + "content": " labelled data, where 'multi' is multi-boxes and '1-inc' is single inclination." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 34, + 199, + 120, + 282 + ], + "lines": [ + { + "bbox": [ + 34, + 199, + 120, + 282 + ], + "spans": [ + { + "bbox": [ + 34, + 199, + 120, + 282 + ], + "type": "image", + "image_path": "d6f08f6ff4cb749c3191e7b450ced6b3186db14ad833561a843dea7fe1593797.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 121, + 201, + 206, + 284 + ], + "blocks": [ + { + "bbox": [ + 121, + 201, + 206, + 284 + ], + "lines": [ + { + "bbox": [ + 121, + 201, + 206, + 284 + ], + "spans": [ + { + "bbox": [ + 121, + 201, + 206, + 284 + ], + "type": "image", + "image_path": "e9b95a5defc6d47b1d7e7eb8da00a20818242b9f1bce0bef7583071d11a5983a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 206, + 201, + 291, + 282 + ], + "blocks": [ + { + "bbox": [ + 206, + 201, + 291, + 282 + ], + "lines": [ + { + "bbox": [ + 206, + 201, + 291, + 282 + ], + "spans": [ + { + "bbox": [ + 206, + 201, + 291, + 282 + ], + "type": "image", + "image_path": "2e3804f3d5fd133ae4ec4cb6f161a8b076a3c4aa9eddfec5db5d2db024a60314.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 293, + 201, + 378, + 282 + ], + "blocks": [ + { + "bbox": [ + 293, + 201, + 378, + 282 + ], + "lines": [ + { + "bbox": [ + 293, + 201, + 378, + 282 + ], + "spans": [ + { + "bbox": [ + 293, + 201, + 378, + 282 + ], + "type": "image", + "image_path": "423e5f59caf966f7b1a488cc4f83d10575d1bfad41049cef93bf65630d7f0309.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 34, + 284, + 119, + 399 + ], + "blocks": [ + { + "bbox": [ + 34, + 284, + 119, + 399 + ], + "lines": [ + { + "bbox": [ + 34, + 284, + 119, + 399 + ], + "spans": [ + { + "bbox": [ + 34, + 284, + 119, + 399 + ], + "type": "image", + "image_path": "d20a87dd5527309d4d7128d0c548069c54da799d0b8ab2283d7ce32ced6a0b41.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 400, + 108, + 409 + ], + "lines": [ + { + "bbox": [ + 46, + 400, + 108, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 108, + 409 + ], + "type": "text", + "content": "(a) Ground truth" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 121, + 285, + 205, + 399 + ], + "blocks": [ + { + "bbox": [ + 121, + 285, + 205, + 399 + ], + "lines": [ + { + "bbox": [ + 121, + 285, + 205, + 399 + ], + "spans": [ + { + "bbox": [ + 121, + 285, + 205, + 399 + ], + "type": "image", + "image_path": "49b013b85c98998b551c2c75e72dc8da8e7e0c74329fa5869ffa3d211ca185d5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 400, + 185, + 410 + ], + "lines": [ + { + "bbox": [ + 141, + 400, + 185, + 410 + ], + "spans": [ + { + "bbox": [ + 141, + 400, + 185, + 410 + ], + "type": "text", + "content": "(b) MT [50]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 206, + 285, + 291, + 399 + ], + "blocks": [ + { + "bbox": [ + 206, + 285, + 291, + 399 + ], + "lines": [ + { + "bbox": [ + 206, + 285, + 291, + 399 + ], + "spans": [ + { + "bbox": [ + 206, + 285, + 291, + 399 + ], + "type": "image", + "image_path": "947c3dec638db299244e1dd7a020717d624d6d8a2baf0785bcbaee4f36e2f722.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 218, + 400, + 280, + 410 + ], + "lines": [ + { + "bbox": [ + 218, + 400, + 280, + 410 + ], + "spans": [ + { + "bbox": [ + 218, + 400, + 280, + 410 + ], + "type": "text", + "content": "(c) LaserMix [27]" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 292, + 285, + 377, + 399 + ], + "blocks": [ + { + "bbox": [ + 292, + 285, + 377, + 399 + ], + "lines": [ + { + "bbox": [ + 292, + 285, + 377, + 399 + ], + "spans": [ + { + "bbox": [ + 292, + 285, + 377, + 399 + ], + "type": "image", + "image_path": "d7e4b2a21f93eddcef5173424ac2b33b630d2f9333bca2849037336b8bb148c5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 400, + 351, + 409 + ], + "lines": [ + { + "bbox": [ + 318, + 400, + 351, + 409 + ], + "spans": [ + { + "bbox": [ + 318, + 400, + 351, + 409 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 30, + 419, + 379, + 453 + ], + "lines": [ + { + "bbox": [ + 30, + 419, + 379, + 453 + ], + "spans": [ + { + "bbox": [ + 30, + 419, + 379, + 453 + ], + "type": "text", + "content": "Fig. 5: Error maps visualised from LiDAR points (top) and range view (bottom) in 2 examples of the nuScenes [4] dataset. The correct predictions are in green and the mistakes are highlighted in red." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 30, + 480, + 380, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 480, + 380, + 590 + ], + "spans": [ + { + "bbox": [ + 30, + 480, + 380, + 590 + ], + "type": "text", + "content": "tion studies with an additional representation, the Bird's Eye View (BEV), on the nuScenes [4] dataset, where all the results in the table are based on PolarNet [64]. We re-implement the supervised results and single representation supervision [6] in the first and second rows, respectively. We emphasize that employing peer representations can significantly enhance performance compared to the single-representation baseline [6]. For instance, under the " + }, + { + "bbox": [ + 30, + 480, + 380, + 590 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 30, + 480, + 380, + 590 + ], + "type": "text", + "content": " labelled partition protocol, employing BEV (based on PolarNet [64]) with the range image (based on FidNet [65]) and with the voxel grid (based on Cylinder3D [66]) lead to improvements of " + }, + { + "bbox": [ + 30, + 480, + 380, + 590 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 30, + 480, + 380, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 480, + 380, + 590 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 30, + 480, + 380, + 590 + ], + "type": "text", + "content": " in mIoU, respectively." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 10, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 10, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 298, + 10, + 346, + 19 + ], + "type": "text", + "content": "ItTakesTwo" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 18 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 18 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 34, + 35, + 147, + 119 + ], + "blocks": [ + { + "bbox": [ + 34, + 35, + 147, + 119 + ], + "lines": [ + { + "bbox": [ + 34, + 35, + 147, + 119 + ], + "spans": [ + { + "bbox": [ + 34, + 35, + 147, + 119 + ], + "type": "image", + "image_path": "a7d369ee918029812c70d2e104eba80df1a04ad4510cf4db1bd8a56f9479c05a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 121, + 121, + 130 + ], + "lines": [ + { + "bbox": [ + 60, + 121, + 121, + 130 + ], + "spans": [ + { + "bbox": [ + 60, + 121, + 121, + 130 + ], + "type": "text", + "content": "(a) w/o Contras." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 149, + 35, + 263, + 119 + ], + "blocks": [ + { + "bbox": [ + 149, + 35, + 263, + 119 + ], + "lines": [ + { + "bbox": [ + 149, + 35, + 263, + 119 + ], + "spans": [ + { + "bbox": [ + 149, + 35, + 263, + 119 + ], + "type": "image", + "image_path": "5c9d838b19176cfbe93572eaa2d87dad574f734e7ff4c8e08512e61c21729b25.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 121, + 241, + 131 + ], + "lines": [ + { + "bbox": [ + 171, + 121, + 241, + 131 + ], + "spans": [ + { + "bbox": [ + 171, + 121, + 241, + 131 + ], + "type": "text", + "content": "(b) ContrasSeg [54]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 30, + 140, + 380, + 174 + ], + "lines": [ + { + "bbox": [ + 30, + 140, + 380, + 174 + ], + "spans": [ + { + "bbox": [ + 30, + 140, + 380, + 174 + ], + "type": "text", + "content": "Fig. 6: T-SNE visualisation for the pixel- (and voxel-) wise latent embedding features in the nuScenes [4] dataset. The features " + }, + { + "bbox": [ + 30, + 140, + 380, + 174 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 30, + 140, + 380, + 174 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 140, + 380, + 174 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 30, + 140, + 380, + 174 + ], + "type": "text", + "content": " indicate range and voxel representations and they are colored according to the category label." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 264, + 35, + 378, + 119 + ], + "blocks": [ + { + "bbox": [ + 264, + 35, + 378, + 119 + ], + "lines": [ + { + "bbox": [ + 264, + 35, + 378, + 119 + ], + "spans": [ + { + "bbox": [ + 264, + 35, + 378, + 119 + ], + "type": "image", + "image_path": "430e59adc6899344eadebe4b0138047bc33a9cfe5739ffe63f41b9e7bec59766.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 122, + 337, + 130 + ], + "lines": [ + { + "bbox": [ + 306, + 122, + 337, + 130 + ], + "spans": [ + { + "bbox": [ + 306, + 122, + 337, + 130 + ], + "type": "text", + "content": "(c) Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 31, + 192, + 125, + 203 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 192, + 125, + 203 + ], + "spans": [ + { + "bbox": [ + 31, + 192, + 125, + 203 + ], + "type": "text", + "content": "4.4 Visualisation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 212, + 380, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 212, + 380, + 295 + ], + "spans": [ + { + "bbox": [ + 30, + 212, + 380, + 295 + ], + "type": "text", + "content": "Visualisation of prediction error maps. We compare the error maps from predictions in Fig. 5, including Mean Teacher [50] (2nd column), LaserMix [27] (3rd column) and ours (last column). The visualisations show the bird's eye view's output in the first row, followed by the range view results in the second row. Notably, our method's prediction demonstrates a superior qualitative result, exhibiting fewer mistaken predictions (highlighted in red) and more correct predictions (marked in green) for both views." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 30, + 296, + 380, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 296, + 380, + 452 + ], + "spans": [ + { + "bbox": [ + 30, + 296, + 380, + 452 + ], + "type": "text", + "content": "T-SNE visualisation of the latent features. Figure 6 displays the latent embedding distribution from the nuScenes [4] dataset. Fig. 6a shows the results without contrastive learning, Fig. 6b displays the results based on the label-guided contrastive learning [54], and our prototypical contrastive learning is in Fig. 6c. The samples denoted by " + }, + { + "bbox": [ + 30, + 296, + 380, + 452 + ], + "type": "inline_equation", + "content": "\\text{和}" + }, + { + "bbox": [ + 30, + 296, + 380, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 296, + 380, + 452 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 30, + 296, + 380, + 452 + ], + "type": "text", + "content": " indicate the latent embeddings from range and voxel models, respectively, where different colours represent the classes. Fig. 6a displays an unstructured latent representation due to the absence of contrastive learning, but Fig. 6b successfully clusters class embeddings from the same modalities (e.g., " + }, + { + "bbox": [ + 30, + 296, + 380, + 452 + ], + "type": "inline_equation", + "content": "\\{\\text{和}, \\text{和}\\}" + }, + { + "bbox": [ + 30, + 296, + 380, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 296, + 380, + 452 + ], + "type": "inline_equation", + "content": "\\{\\text{和}, \\text{和}\\}" + }, + { + "bbox": [ + 30, + 296, + 380, + 452 + ], + "type": "text", + "content": "), yet it fails to capture the semantic relationships between different modalities within the same category. Our method, depicted in Fig. 6c, successfully clusters embeddings from the same class and different modalities " + }, + { + "bbox": [ + 30, + 296, + 380, + 452 + ], + "type": "inline_equation", + "content": "\\{\\text{和}, \\text{和}\\}" + }, + { + "bbox": [ + 30, + 296, + 380, + 452 + ], + "type": "text", + "content": ", enabling a more effective distribution of embeddings for subsequent segmentation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 31, + 463, + 119, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 463, + 119, + 475 + ], + "spans": [ + { + "bbox": [ + 31, + 463, + 119, + 475 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 30, + 487, + 379, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 487, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 487, + 379, + 584 + ], + "type": "text", + "content": "In this study, we introduced the novel semi-supervised LiDAR semantic segmentation method ItTakesTwo (IT2). IT2 improves generalisation with consistent predictions from peer LiDAR representations, and it also introduces a more effective contrastive learning using informative embeddings sampled from a learned distribution of positives and negatives. Finally, through experimental studies, we identify an efficient way to apply data augmentation to each representation. Our approach achieves best results across various benchmarks under all partition protocols, demonstrating its effectiveness." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 31, + 34, + 380, + 59 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 34, + 380, + 59 + ], + "spans": [ + { + "bbox": [ + 31, + 34, + 380, + 59 + ], + "type": "text", + "content": "Acknowledgements. The project is supported by the Australian Research Council (ARC) through grant FT190100525." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 75, + 98, + 87 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 75, + 98, + 87 + ], + "spans": [ + { + "bbox": [ + 32, + 75, + 98, + 87 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 97, + 380, + 584 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 36, + 97, + 379, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 97, + 379, + 129 + ], + "spans": [ + { + "bbox": [ + 36, + 97, + 379, + 129 + ], + "type": "text", + "content": "1. Arazo, E., Ortego, D., Albert, P., O'Connor, N.E., McGuinness, K.: Pseudolabeling and confirmation bias in deep semi-supervised learning. In: 2020 International Joint Conference on Neural Networks (IJCNN). pp. 1-8. IEEE (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 36, + 129, + 379, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 129, + 379, + 173 + ], + "spans": [ + { + "bbox": [ + 36, + 129, + 379, + 173 + ], + "type": "text", + "content": "2. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9297-9307 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 37, + 173, + 380, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 173, + 380, + 215 + ], + "spans": [ + { + "bbox": [ + 37, + 173, + 380, + 215 + ], + "type": "text", + "content": "3. Berman, M., Triki, A.R., Blaschko, M.B.: The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4413-4421 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 37, + 216, + 380, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 216, + 380, + 260 + ], + "spans": [ + { + "bbox": [ + 37, + 216, + 380, + 260 + ], + "type": "text", + "content": "4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 37, + 260, + 379, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 260, + 379, + 292 + ], + "spans": [ + { + "bbox": [ + 37, + 260, + 379, + 292 + ], + "type": "text", + "content": "5. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems 33, 9912-9924 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 37, + 292, + 379, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 292, + 379, + 324 + ], + "spans": [ + { + "bbox": [ + 37, + 292, + 379, + 324 + ], + "type": "text", + "content": "6. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2613-2622 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 37, + 324, + 379, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 324, + 379, + 356 + ], + "spans": [ + { + "bbox": [ + 37, + 324, + 379, + 356 + ], + "type": "text", + "content": "7. Cheng, M., Hui, L., Xie, J., Yang, J.: Sspc-net: Semi-supervised semantic 3d point cloud segmentation network. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 1140-1147 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 37, + 356, + 379, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 356, + 379, + 399 + ], + "spans": [ + { + "bbox": [ + 37, + 356, + 379, + 399 + ], + "type": "text", + "content": "8. Cheung, Y.m.: A rival penalized em algorithm towards maximizing weighted likelihood for density mixture clustering with automatic model selection. In: Proceedings of the 17th International Conference on Pattern Recognition, 2004. ICPR 2004. vol. 4, pp. 633-636. IEEE (2004)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 37, + 400, + 379, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 400, + 379, + 431 + ], + "spans": [ + { + "bbox": [ + 37, + 400, + 379, + 431 + ], + "type": "text", + "content": "9. Cheung, Y.m.: Maximum weighted likelihood via rival penalized em for density mixture clustering with automatic model selection. IEEE Transactions on Knowledge and Data Engineering 17(6), 750-761 (2005)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 431, + 379, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 431, + 379, + 486 + ], + "spans": [ + { + "bbox": [ + 33, + 431, + 379, + 486 + ], + "type": "text", + "content": "10. Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d u-net: learning dense volumetric segmentation from sparse annotation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part II 19. pp. 424-432. Springer (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 486, + 379, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 486, + 379, + 519 + ], + "spans": [ + { + "bbox": [ + 33, + 486, + 379, + 519 + ], + "type": "text", + "content": "11. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological) 39(1), 1-22 (1977)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 519, + 379, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 519, + 379, + 551 + ], + "spans": [ + { + "bbox": [ + 33, + 519, + 379, + 551 + ], + "type": "text", + "content": "12. Deng, S., Dong, Q., Liu, B., Hu, Z.: Superpoint-guided semi-supervised semantic segmentation of 3d point clouds. In: 2022 International conference on robotics and automation (ICRA). pp. 9214-9220. IEEE (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 551, + 379, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 551, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 33, + 551, + 379, + 584 + ], + "type": "text", + "content": "13. Fan, L., Xiong, X., Wang, F., Wang, N., Zhang, Z.: Rangedet: In defense of range view for lidar-based 3d object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2918-2927 (2021)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 10, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 10, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 298, + 10, + 346, + 19 + ], + "type": "text", + "content": "ItTakesTwo" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 35, + 379, + 583 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 57 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 35, + 379, + 57 + ], + "spans": [ + { + "bbox": [ + 33, + 35, + 379, + 57 + ], + "type": "text", + "content": "14. French, G., Aila, T., Laine, S., Mackiewicz, M., Finlayson, G.: Semi-supervised semantic segmentation needs strong, high-dimensional perturbations (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 57, + 379, + 89 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 57, + 379, + 89 + ], + "spans": [ + { + "bbox": [ + 32, + 57, + 379, + 89 + ], + "type": "text", + "content": "15. Graham, B., Engelcke, M., Van Der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9224-9232 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 89, + 379, + 121 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 89, + 379, + 121 + ], + "spans": [ + { + "bbox": [ + 33, + 89, + 379, + 121 + ], + "type": "text", + "content": "16. Grigorescu, S., Trasnea, B., Cocias, T., Macesanu, G.: A survey of deep learning techniques for autonomous driving. Journal of Field Robotics 37(3), 362-386 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 121, + 379, + 153 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 121, + 379, + 153 + ], + "spans": [ + { + "bbox": [ + 33, + 121, + 379, + 153 + ], + "type": "text", + "content": "17. Guo, Y., Wang, H., Hu, Q., Liu, H., Liu, L., Bennamoun, M.: Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence 43(12), 4338-4364 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 154, + 379, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 154, + 379, + 186 + ], + "spans": [ + { + "bbox": [ + 33, + 154, + 379, + 186 + ], + "type": "text", + "content": "18. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 186, + 378, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 186, + 378, + 218 + ], + "spans": [ + { + "bbox": [ + 33, + 186, + 378, + 218 + ], + "type": "text", + "content": "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 218, + 379, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 218, + 379, + 250 + ], + "spans": [ + { + "bbox": [ + 33, + 218, + 379, + 250 + ], + "type": "text", + "content": "20. Hou, J., Graham, B., Nießner, M., Xie, S.: Exploring data-efficient 3d scene understanding with contrastive scene contexts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15587-15597 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 250, + 378, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 250, + 378, + 282 + ], + "spans": [ + { + "bbox": [ + 33, + 250, + 378, + 282 + ], + "type": "text", + "content": "21. Hou, Y., Zhu, X., Ma, Y., Loy, C.C., Li, Y.: Point-to-voxel knowledge distillation for lidar semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8479-8488 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 282, + 378, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 282, + 378, + 304 + ], + "spans": [ + { + "bbox": [ + 33, + 282, + 378, + 304 + ], + "type": "text", + "content": "22. Hung, W.C., Tsai, Y.H., Liou, Y.T., Lin, Y.Y., Yang, M.H.: Adversarial learning for semi-supervised semantic segmentation. arXiv preprint arXiv:1802.07934 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 33, + 304, + 379, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 304, + 379, + 346 + ], + "spans": [ + { + "bbox": [ + 33, + 304, + 379, + 346 + ], + "type": "text", + "content": "23. Jiang, L., Shi, S., Tian, Z., Lai, X., Liu, S., Fu, C.W., Jia, J.: Guided point contrastive learning for semi-supervised point cloud semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6423-6432 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 347, + 379, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 347, + 379, + 379 + ], + "spans": [ + { + "bbox": [ + 33, + 347, + 379, + 379 + ], + "type": "text", + "content": "24. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in Neural Information Processing Systems 33, 18661-18673 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 380, + 379, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 380, + 379, + 411 + ], + "spans": [ + { + "bbox": [ + 33, + 380, + 379, + 411 + ], + "type": "text", + "content": "25. Kohli, A.P.S., Sitzmann, V., Wetzstein, G.: Semantic implicit neural scene representations with semi-supervised training. In: 2020 International Conference on 3D Vision (3DV). pp. 423-433. IEEE (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 412, + 379, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 412, + 379, + 444 + ], + "spans": [ + { + "bbox": [ + 33, + 412, + 379, + 444 + ], + "type": "text", + "content": "26. Kong, L., Liu, Y., Chen, R., Ma, Y., Zhu, X., Li, Y., Hou, Y., Qiao, Y., Liu, Z.: Rethinking range view representation for lidar segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 228-240 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 444, + 379, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 444, + 379, + 476 + ], + "spans": [ + { + "bbox": [ + 33, + 444, + 379, + 476 + ], + "type": "text", + "content": "27. Kong, L., Ren, J., Pan, L., Liu, Z.: Lasermix for semi-supervised lidar semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21705-21715 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 476, + 379, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 476, + 379, + 508 + ], + "spans": [ + { + "bbox": [ + 33, + 476, + 379, + 508 + ], + "type": "text", + "content": "28. Lai, X., Chen, Y., Lu, F., Liu, J., Jia, J.: Spherical transformer for lidar-based 3d recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17545-17555 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 33, + 509, + 378, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 509, + 378, + 529 + ], + "spans": [ + { + "bbox": [ + 33, + 509, + 378, + 529 + ], + "type": "text", + "content": "29. Le-Khac, P.H., Healy, G., Smeaton, A.F.: Contrastive representation learning: A framework and review. IEEE Access 8, 193907-193934 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 33, + 530, + 378, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 530, + 378, + 551 + ], + "spans": [ + { + "bbox": [ + 33, + 530, + 378, + 551 + ], + "type": "text", + "content": "30. Li, J., Zhou, P., Xiong, C., Hoi, S.C.: Prototypical contrastive learning of unsupervised representations. arXiv preprint arXiv:2005.04966 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "type": "text", + "content": "31. Li, L., Shum, H.P., Breckon, T.P.: Less is more: Reducing task and model complexity for 3d point cloud semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9361-9371 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 35, + 379, + 583 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 32, + 35, + 379, + 79 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 35, + 379, + 79 + ], + "spans": [ + { + "bbox": [ + 32, + 35, + 379, + 79 + ], + "type": "text", + "content": "32. Li, M., Xie, Y., Shen, Y., Ke, B., Qiao, R., Ren, B., Lin, S., Ma, L.: Hybridcr: Weakly-supervised 3d point cloud semantic segmentation via hybrid contrastive regularization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14930-14939 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 80, + 379, + 111 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 80, + 379, + 111 + ], + "spans": [ + { + "bbox": [ + 33, + 80, + 379, + 111 + ], + "type": "text", + "content": "33. Liang, C., Wang, W., Miao, J., Yang, Y.: Gmmseg: Gaussian mixture based generative semantic segmentation models. Advances in Neural Information Processing Systems 35, 31360-31375 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 113, + 379, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 113, + 379, + 145 + ], + "spans": [ + { + "bbox": [ + 33, + 113, + 379, + 145 + ], + "type": "text", + "content": "34. Liang, T., Xie, H., Yu, K., Xia, Z., Lin, Z., Wang, Y., Tang, T., Wang, B., Tang, Z.: Bevfusion: A simple and robust lidar-camera fusion framework. Advances in Neural Information Processing Systems 35, 10421-10434 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 145, + 379, + 177 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 145, + 379, + 177 + ], + "spans": [ + { + "bbox": [ + 33, + 145, + 379, + 177 + ], + "type": "text", + "content": "35. Liu, M., Zhou, Y., Qi, C.R., Gong, B., Su, H., Anguelov, D.: Less: Label-efficient semantic segmentation for lidar point clouds. In: European Conference on Computer Vision. pp. 70-89. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 178, + 379, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 178, + 379, + 210 + ], + "spans": [ + { + "bbox": [ + 33, + 178, + 379, + 210 + ], + "type": "text", + "content": "36. Liu, W., Yue, X., Chen, Y., Denoeux, T.: Trusted multi-view deep learning with opinion aggregation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 7585-7593 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 211, + 379, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 211, + 379, + 233 + ], + "spans": [ + { + "bbox": [ + 33, + 211, + 379, + 233 + ], + "type": "text", + "content": "37. Liu, Y., Hu, Q., Lei, Y., Xu, K., Li, J., Guo, Y.: Box2seg: Learning semantics of 3d point clouds with box-level supervision. arXiv preprint arXiv:2201.02963 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 233, + 379, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 233, + 379, + 265 + ], + "spans": [ + { + "bbox": [ + 33, + 233, + 379, + 265 + ], + "type": "text", + "content": "38. Liu, Y., Kong, L., Cen, J., Chen, R., Zhang, W., Pan, L., Chen, K., Liu, Z.: Segment any point cloud sequences by distilling vision foundation models. arXiv preprint arXiv:2306.09347 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 266, + 379, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 266, + 379, + 309 + ], + "spans": [ + { + "bbox": [ + 33, + 266, + 379, + 309 + ], + "type": "text", + "content": "39. Liu, Y., Tian, Y., Chen, Y., Liu, F., Belagiannis, V., Carneiro, G.: Perturbed and strict mean teachers for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4258-4267 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 310, + 379, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 310, + 379, + 342 + ], + "spans": [ + { + "bbox": [ + 33, + 310, + 379, + 342 + ], + "type": "text", + "content": "40. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: IEEE International Conference on Robotics and Automation (ICRA) (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 33, + 343, + 379, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 343, + 379, + 386 + ], + "spans": [ + { + "bbox": [ + 33, + 343, + 379, + 386 + ], + "type": "text", + "content": "41. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2774-2781. IEEE (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 387, + 379, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 387, + 379, + 408 + ], + "spans": [ + { + "bbox": [ + 33, + 387, + 379, + 408 + ], + "type": "text", + "content": "42. Manning, C.D.: An introduction to information retrieval. Cambridge university press (2009)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 408, + 379, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 408, + 379, + 440 + ], + "spans": [ + { + "bbox": [ + 33, + 408, + 379, + 440 + ], + "type": "text", + "content": "43. Mena, G., Nejatbakhsh, A., Varol, E., Niles-Weed, J.: Sinkhorn em: an expectation-maximization algorithm based on entropic optimal transport. arXiv preprint arXiv:2006.16548 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 441, + 379, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 441, + 379, + 485 + ], + "spans": [ + { + "bbox": [ + 33, + 441, + 379, + 485 + ], + "type": "text", + "content": "44. Miech, A., Alayrac, J.B., Smaira, L., Laptev, I., Sivic, J., Zisserman, A.: End-to-end learning of visual representations from uncurated instructional videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9879-9889 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 485, + 379, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 485, + 379, + 517 + ], + "spans": [ + { + "bbox": [ + 33, + 485, + 379, + 517 + ], + "type": "text", + "content": "45. Milioto, A., Vizzo, I., Behley, J., Stachniss, C.: Rangenet++: Fast and accurate lidar semantic segmentation. In: 2019 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 4213-4220. IEEE (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 518, + 379, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 518, + 379, + 550 + ], + "spans": [ + { + "bbox": [ + 33, + 518, + 379, + 550 + ], + "type": "text", + "content": "46. Nunes, L., Marcuzzi, R., Chen, X., Behley, J., Stachniss, C.: Segcontrast: 3d point cloud feature representation learning through self-supervised segment discrimination. IEEE Robotics and Automation Letters 7(2), 2116-2123 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "type": "text", + "content": "47. Ouali, Y., Hudelot, C., Tami, M.: Semi-supervised semantic segmentation with cross-consistency training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12674-12684 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 10, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 10, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 298, + 10, + 346, + 19 + ], + "type": "text", + "content": "ItTakesTwo" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 583 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 68 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 35, + 379, + 68 + ], + "spans": [ + { + "bbox": [ + 33, + 35, + 379, + 68 + ], + "type": "text", + "content": "48. Reichardt, L., Ebert, N., Wasenmüller, O.: 360deg from a single camera: A few-shot approach for lidar segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1075-1083 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 68, + 379, + 100 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 68, + 379, + 100 + ], + "spans": [ + { + "bbox": [ + 33, + 68, + 379, + 100 + ], + "type": "text", + "content": "49. Sun, C.Y., Yang, Y.Q., Guo, H.X., Wang, P.S., Tong, X., Liu, Y., Shum, H.Y.: Semi-supervised 3d shape segmentation with multilevel consistency and part substitution. Computational Visual Media 9(2), 229-247 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 101, + 379, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 101, + 379, + 134 + ], + "spans": [ + { + "bbox": [ + 33, + 101, + 379, + 134 + ], + "type": "text", + "content": "50. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. arXiv preprint arXiv:1703.01780 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 134, + 379, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 134, + 379, + 167 + ], + "spans": [ + { + "bbox": [ + 33, + 134, + 379, + 167 + ], + "type": "text", + "content": "51. Tian, Y., Krishnan, D., Isola, P.: Contrastive multiview coding. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16. pp. 776-794. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 167, + 379, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 167, + 379, + 199 + ], + "spans": [ + { + "bbox": [ + 33, + 167, + 379, + 199 + ], + "type": "text", + "content": "52. Unal, O., Dai, D., Van Gool, L.: Scribble-supervised lidar semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2697-2707 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 199, + 379, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 199, + 379, + 232 + ], + "spans": [ + { + "bbox": [ + 33, + 199, + 379, + 232 + ], + "type": "text", + "content": "53. Vora, S., Radwan, N., Greff, K., Meyer, H., Genova, K., Sajjadi, M.S., Pot, E., Tagliasacchi, A., Duckworth, D.: Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes. arXiv preprint arXiv:2111.13260 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 233, + 379, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 233, + 379, + 265 + ], + "spans": [ + { + "bbox": [ + 33, + 233, + 379, + 265 + ], + "type": "text", + "content": "54. Wang, W., Zhou, T., Yu, F., Dai, J., Konukoglu, E., Van Gool, L.: Exploring cross-image pixel contrast for semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7303-7313 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 265, + 379, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 265, + 379, + 308 + ], + "spans": [ + { + "bbox": [ + 33, + 265, + 379, + 308 + ], + "type": "text", + "content": "55. Wu, B., Wan, A., Yue, X., Keutzer, K.: Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In: 2018 IEEE international conference on robotics and automation (ICRA). pp. 1887-1893. IEEE (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 309, + 379, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 309, + 379, + 354 + ], + "spans": [ + { + "bbox": [ + 33, + 309, + 379, + 354 + ], + "type": "text", + "content": "56. Xie, S., Gu, J., Guo, D., Qi, C.R., Guibas, L., Litany, O.: Pointcontrast: Unsupervised pre-training for 3d point cloud understanding. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 574-591. Springer (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 33, + 353, + 379, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 353, + 379, + 375 + ], + "spans": [ + { + "bbox": [ + 33, + 353, + 379, + 375 + ], + "type": "text", + "content": "57. Xu, J., Hsu, D.J., Maleki, A.: Benefits of over-parameterization with em. Advances in Neural Information Processing Systems 31 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 375, + 379, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 375, + 379, + 419 + ], + "spans": [ + { + "bbox": [ + 33, + 375, + 379, + 419 + ], + "type": "text", + "content": "58. Xu, J., Zhang, R., Dou, J., Zhu, Y., Sun, J., Pu, S.: Rpvnet: A deep and efficient range-point-voxel fusion network for lidar point cloud segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16024-16033 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 419, + 379, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 419, + 379, + 452 + ], + "spans": [ + { + "bbox": [ + 33, + 419, + 379, + 452 + ], + "type": "text", + "content": "59. Xu, J., Tang, H., Ren, Y., Peng, L., Zhu, X., He, L.: Multi-level feature learning for contrastive multi-view clustering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16051-16060 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 452, + 379, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 452, + 379, + 495 + ], + "spans": [ + { + "bbox": [ + 33, + 452, + 379, + 495 + ], + "type": "text", + "content": "60. Xu, Z., Yuan, B., Zhao, S., Zhang, Q., Gao, X.: Hierarchical point-based active learning for semi-supervised point cloud semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18098-18108 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 495, + 379, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 495, + 379, + 528 + ], + "spans": [ + { + "bbox": [ + 33, + 495, + 379, + 528 + ], + "type": "text", + "content": "61. Yang, L., Qi, L., Feng, L., Zhang, W., Shi, Y.: Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7236-7246 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 528, + 379, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 528, + 379, + 551 + ], + "spans": [ + { + "bbox": [ + 33, + 528, + 379, + 551 + ], + "type": "text", + "content": "62. Yang, L., Zhuo, W., Qi, L., Shi, Y., Gao, Y.: St++: Make self-training work better for semi-supervised semantic segmentation. arXiv preprint arXiv:2106.05095 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "type": "text", + "content": "63. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6023-6032 (2019)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 18 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 18 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 116, + 19 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 31, + 34, + 379, + 254 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 31, + 34, + 379, + 79 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 34, + 379, + 79 + ], + "spans": [ + { + "bbox": [ + 31, + 34, + 379, + 79 + ], + "type": "text", + "content": "64. Zhang, Y., Zhou, Z., David, P., Yue, X., Xi, Z., Gong, B., Foroosh, H.: Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9601-9610 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 80, + 379, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 80, + 379, + 112 + ], + "spans": [ + { + "bbox": [ + 31, + 80, + 379, + 112 + ], + "type": "text", + "content": "65. Zhao, Y., Bai, L., Huang, X.: Fidnet: Lidar point cloud semantic segmentation with fully interpolation decoding. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4453-4458. IEEE (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 31, + 113, + 379, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 113, + 379, + 156 + ], + "spans": [ + { + "bbox": [ + 31, + 113, + 379, + 156 + ], + "type": "text", + "content": "66. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9939-9948 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 156, + 379, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 156, + 379, + 188 + ], + "spans": [ + { + "bbox": [ + 31, + 156, + 379, + 188 + ], + "type": "text", + "content": "67. Zhuang, Z., Li, R., Jia, K., Wang, Q., Li, Y., Tan, M.: Perception-aware multi-sensor fusion for 3d lidar semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16280-16290 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 32, + 189, + 379, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 189, + 379, + 221 + ], + "spans": [ + { + "bbox": [ + 32, + 189, + 379, + 221 + ], + "type": "text", + "content": "68. Zou, Y., Yu, Z., Kumar, B., Wang, J.: Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In: Proceedings of the European conference on computer vision (ECCV). pp. 289-305 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 32, + 222, + 379, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 222, + 379, + 254 + ], + "spans": [ + { + "bbox": [ + 32, + 222, + 379, + 254 + ], + "type": "text", + "content": "69. Zou, Y., Zhang, Z., Zhang, H., Li, C.L., Bian, X., Huang, J.B., Pfister, T.: Pseudoseg: Designing pseudo labels for semantic segmentation. arXiv preprint arXiv:2010.09713 (2020)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 10, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 10, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 298, + 10, + 346, + 19 + ], + "type": "text", + "content": "It Takes Two" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models/d774ff0f-2daf-4eeb-815c-c6da2eb3ab17_content_list.json b/2024/Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models/d774ff0f-2daf-4eeb-815c-c6da2eb3ab17_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e9e950680567db39de0e0d04286c85a5c29b7058 --- /dev/null +++ b/2024/Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models/d774ff0f-2daf-4eeb-815c-c6da2eb3ab17_content_list.json @@ -0,0 +1,1790 @@ +[ + { + "type": "text", + "text": "Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models", + "text_level": 1, + "bbox": [ + 220, + 140, + 781, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiao Liu $^{1*}$ , Xiaoliu Guan $^{1*}$ , Yu Wu $^{1\\dagger}$ , and Jiaxu Miao $^{2\\dagger}$", + "bbox": [ + 274, + 233, + 732, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science, Wuhan University", + "bbox": [ + 334, + 261, + 668, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xiaoliu, liuxiaoguan, wuyucs}@whu.edu.cn", + "bbox": [ + 349, + 277, + 651, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Cyber Science and Technology, Sun Yat-sen University", + "bbox": [ + 279, + 290, + 722, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "miaojx@mail.sysu.edu.cn", + "bbox": [ + 410, + 305, + 591, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Diffusion models, known for their tremendous ability to generate novel and high-quality samples, have recently raised concerns due to their data memorization behavior, which poses privacy risks. Recent approaches for memory mitigation either only focused on the text modality problem in cross-modal generation tasks or utilized data augmentation strategies. In this paper, we propose a novel training framework for diffusion models from the perspective of visual modality, which is more generic and fundamental for mitigating memorization. To facilitate \"forgetting\" of stored information in diffusion model parameters, we propose an iterative ensemble training strategy by splitting the data into multiple shards for training multiple models and intermittently aggregating these model parameters. Moreover, practical analysis of losses illustrates that the training loss for easily memorable images tends to be obviously lower. Thus, we propose an anti-gradient control method to exclude the sample with a lower loss value from the current mini-batch to avoid memorizing. Extensive experiments and analysis on four datasets are conducted to illustrate the effectiveness of our method, and results show that our method successfully reduces memory capacity while even improving the performance slightly. Moreover, to save the computing cost, we successfully apply our method to fine-tune the well-trained diffusion models by limited epochs, demonstrating the applicability of our method. Code is available in https://github.com/liuxiao-guan/IET_AGC.", + "bbox": [ + 259, + 356, + 743, + 661 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Diffusion Models $\\cdot$ Model Memorization $\\cdot$ Data Privacy", + "bbox": [ + 261, + 674, + 718, + 689 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 738, + 375, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in diffusion models have significantly transformed the landscape of image generation. Modern diffusion models, such as Stable Diffusion [23],", + "bbox": [ + 212, + 770, + 787, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 218, + 811, + 364, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 222, + 824, + 383, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Midjourney [1], and SORA [2], have the ability to generate realistic images that are hard for humans to distinguish, demonstrating the unparalleled capabilities in producing diverse images. However, recent works [5, 27, 31] suggested that diffusion models are capable of memorizing images from the training set and reproducing them, which has the potential risk of privacy leakage. To address the severe problem, some works [10, 17, 19, 33] proposed to make diffusion models \"forget\" specific concepts such as a portrait of a certain celebrity, or the style of a particular artist. However, these works can only blacklist specific content that users want to conceal, but cannot completely cover the privacy-sensitive information that the model might remember, which still has a risk of privacy leakage.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, some works [7,26,27,31] have proposed to mitigate diffusion memorization without specific content limitations, thus reducing the risk of diffusion models leaking privacy-sensitive training data. Most of them focused on tackling the training data memorization in text-to-image diffusion models, and proposed data augmentation for captions to reduce model memorization since the insufficient diversity in captions easily leads to training data generation. For instance, Somepalli et al. [27] utilized random caption replacement, random token replacement, caption word repetition, etc., to reduce memorization. Another previous approach Daras et al. [7] proposed to use corrupted images for diffusion model training for memorization reduction. Wen et al. [31] introduced methods for detecting memorized prompts through text-conditional predictions and proposed two strategies to mitigate memorization: minimizing during inference or filtering during training.", + "bbox": [ + 212, + 316, + 787, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although these works represented an important step forward in understanding the memorization issue in diffusion models, they either utilized simple data augmentation strategies or only focused on the easily memorable images that are related to specific captions in cross-modal generation tasks. However, diffusion models have been proven capable of generating images from memory without the need for text guidance [4], and previous methods focused on a limited scope of the memorization issue in diffusion models. Differently, in this paper, we propose a novel training framework for diffusion models from the perspective of visual modality, i.e., IET-AGC (Iterative Ensemble Training with Anti-Gradient Control), which not only reduces memorization fundamentally but also provides a more generic approach for both unconditional and text-conditional diffusion models.", + "bbox": [ + 212, + 518, + 787, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "First, we propose an iterative ensemble training (IET) framework to mitigate memorization by parameter aggregation. Training data are stored in parameters of diffusion models due to over-optimization, and the model ensemble strategy aggregates parameters to re-organize model knowledge, which contributes to alleviating the memorization of training data. Thus, we separate the training data into several groups, train diffusion models individually, and ensemble them for memorization reduction. However, aggregating models trained on subsets of data may potentially decrease the performance due to the divergent optimizations on different subsets of these models. Motivated by Federated Learning tech", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "niques [18], we iteratively ensemble the models during training, which reduces memorization by multiple aggregation and maintains generation performance.", + "bbox": [ + 218, + 146, + 784, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on our IET training framework, we further introduce an Anti-Gradient Control module to reduce memorization of training data ulteriorly. We conduct practical analysis on losses during training and find that the training loss for easily memorable images tends to be obviously lower than that for less memorable images. Thus, we propose to exclude the sample with a relatively small loss value from the current mini-batch to avoid memorizing these samples. During the training process, as the diffusion model exhibits varying average loss values across different time steps, we maintain a memory bank to store the average loss of each time step. Subsequently, we discard samples with losses below a certain proportion of the average loss. Since the model has encountered the sample during training, excluding it is unlikely to have a significant impact on the model's performance.", + "bbox": [ + 217, + 178, + 785, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extensive experiments on four datasets show the significance of our IET-AGC framework. Our method significantly reduces the memorized quantity by $87.3\\%$ , $66.4\\%$ , and $85.3\\%$ compared with the default training (DDPM [14]) on CIFAR-10 and CIFAR-100 and AFHQ-DOG, respectively. In addition, considering the high cost of re-training existing well-trained diffusion models, we also propose an efficient way to address the memorization issue of existing pre-trained models by simply fine-tuning using our method for several epochs. Experiments show that using our method to fine-tune for just two epochs can significantly reduce the memorization by $25.22\\%$ in unconditional diffusion models on CIFAR-10. Furthermore, when fine-tuning the text-conditional diffusion model, Stable Diffusion, our approach decreases the memorization score by $42.18\\%$ compared to conventional fine-tuning methods. These results demonstrate that our method performs excellently on both unconditional and text-conditional diffusion models.", + "bbox": [ + 217, + 359, + 785, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 218, + 594, + 385, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Memorization in Generative Models. Several studies have examined the memorization capabilities of the generative model. Generative Adversarial Networks (GANs) [12] have been at the forefront of this research area. As Webster et al. [29] demonstrated when applied to face datasets, GANs can occasionally replicate. Prior study [4] explored an adversarial attack on language models like GPT-2 [21], where individual training examples can be recovered, including personally identifiable information and unique text sequences.", + "bbox": [ + 217, + 628, + 784, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent studies have shifted their attention toward diffusion models. Somepalli et al. [26] found that diffusion models accurately recall and replicate training images, especially noted with models like the Stable Diffusion model [23]. Building upon this discovery, Carlini et al. [5] developed a tailored black-box attack for diffusion models. They generated images and implemented a membership inference attack to assess density. Webster et al. [28] demonstrated a more efficient extraction attack with fewer network evaluations, identified \"template", + "bbox": [ + 217, + 734, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "IET-AGC: Mitigating Memorization in Diffusion Models", + "bbox": [ + 354, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "verbatim,\" and discussed its persistence in newer systems. Recent research has shifted towards exploring the theoretical aspects of memory in diffusion models. [32] discovered that generalization and memorization are mutually exclusive occurrences and further demonstrated that the dichotomy between memorization and generalization can be apparent at the class level. Gu et al. [13] extensively studied how factors like data dimension, model size, time embedding, and class conditions affect the memory capacity of the diffusion model.", + "bbox": [ + 215, + 146, + 785, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Memorization Mitigation. The mitigation measures have primarily been concerned with filtering inputs and deduplication. For example, Stable Diffusion employed well-trained detectors to identify unsuitable generated content. However, these temporary solutions can be easily bypassed [22, 30] and do not effectively prevent or lessen copying behavior on a broad scale. Kumari et al. [17] designed an algorithm to align the image distribution with a specific style, instance, or text prompt they aim to remove, to the distribution related to a core concept. This stopped the model from producing target concepts based on its text condition. However, these approaches are inefficient because they necessitate a list of all concepts to be erased, and have not addressed the key issue of how to reduce the memory capacity of the model. [8, 11] explored the use of differential privacy (DP) [9] to train diffusion models or fine-tune ImageNet pre-trained models. However, their focus was on ensuring the privacy of the training of diffusion models, not on the privacy of the images generated by the diffusion models. Daras et al. [7] introduced a technique for training diffusion models utilizing tainted data. By incorporating additional corruption before applying noise, their methodology prevents the model from overfitting to the training data. But their training requires a considerable amount of time. [27, 31] also suggested a series of recommendations to mitigate copying such as randomly replacing the caption of an image with a random sequence of words, but most of which are limited to text-to-image models. Our work focuses on the nature of memorization in diffusion models, especially for unconditional ones.", + "bbox": [ + 215, + 279, + 785, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 217, + 640, + 328, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we present our methodology aimed at mitigating the memorization in diffusion models, without sacrificing excessive image quality.", + "bbox": [ + 217, + 676, + 784, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Iterative Ensemble Training", + "text_level": 1, + "bbox": [ + 217, + 734, + 491, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training data are stored in parameters of diffusion models due to over-optimization, and the model ensemble strategy aggregates parameters to re-organize model knowledge. Thus, to mitigate the memorization, we propose a method that trains multiple diffusion models on different data shards of a dataset, merges them after a certain period, and then repeats these two stages iteratively.", + "bbox": [ + 217, + 763, + 797, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/25cdf156e98300101500475d8a4f9ee9ca31a8a7ed59a3cac40599a0011aa9cd.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 228, + 147, + 472, + 315 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5b7a0371dc4605e1eb438f22214df4eada10b36a36ffe6fd5e14289c72122c27.jpg", + "image_caption": [ + "(b)", + "Fig. 1: Overview of our IET-AGC method. (a) Iterative Ensemble Training (IET): we divide the dataset $D$ into $K$ different data shards. Each shard $D_{i}$ trains a separate diffusion model $\\theta_{i}$ . After a period of training, these models are merged by averaging and repeating this training strategy. (b) Anti-Gradient Control (AGC): during the training process, we dynamically update and maintain a memory bank of losses at each time step. For loss values smaller than $\\lambda$ times the corresponding memory bank, we exclude these losses to prevent the model from memorizing such images." + ], + "image_footnote": [], + "bbox": [ + 480, + 148, + 774, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training on Different Data Shards. Unlike the training methods of previous diffusion models, which train a single model on the entire dataset once. In this paper, we divide a dataset into an equal number of shards and then train the corresponding diffusion models on each separate part. If the dataset contains class information, we divide it equally along the class dimension, so that each data shard contains all classes, with the same number of instances for each class.", + "bbox": [ + 212, + 469, + 787, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, suppose the dataset $D$ contains $C$ classes, with each class having $N_{c}$ samples, and the total number of samples is $N = \\sum_{c=1}^{C} N_{c}$ . We divide the dataset into $K$ equal parts, with each part containing $\\frac{N}{K}$ samples. So the $i$ th part of the dataset $D_{i}$ is represented as", + "bbox": [ + 212, + 561, + 787, + 623 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {i} = \\bigcup_ {c = 1} ^ {C} \\left\\{\\left(x _ {c, i, k}, y _ {c, i, k}\\right) \\mid k = 1, 2, \\dots , \\frac {N _ {c}}{K} \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 637, + 785, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(x,y)$ represents the sample and its corresponding label. Notably, $C = 1$ indicates that the dataset does not contain class information. Then, each part $i$ trains a separate diffusion model $\\theta_{i}$ scaled by the learning rate $\\eta$ on its own dataset,", + "bbox": [ + 212, + 694, + 787, + 753 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {i} \\leftarrow \\theta_ {i} - \\eta \\nabla \\mathcal {L} \\left(\\theta_ {i}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 761, + 785, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Merging the Multiple Diffusion Models. After training for some time, each part obtains a different diffusion model. We simply average the weights of all", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "IET-AGC: Mitigating Memorization in Diffusion Models", + "bbox": [ + 354, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "models $\\theta_{i}$ to obtain a global model $\\widehat{\\theta}$ as", + "bbox": [ + 215, + 143, + 503, + 160 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\theta} \\leftarrow \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\theta_ {i}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 450, + 175, + 785, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, we repeat the two stages of training on separate shards of the data and merging models, where we use the obtained global model as the initial model for the first stage. As each shard contains only $\\frac{1}{K}$ of the total data, the training time for each model is proportionally reduced, maintaining the overall computational cost nearly constant compared to training a single model on the entire dataset. The only additional computational cost comes from the periodic merging of models, which is minimal.", + "bbox": [ + 214, + 229, + 787, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Loss Analysis", + "text_level": 1, + "bbox": [ + 215, + 361, + 374, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further reduce memorization of training data, we delve deeper into the causes of memorization phenomena, specifically analyzing it through the lens of the loss. We begin by establishing the fundamental notation linked with diffusion models. Diffusion models [14] originate from the non-equilibrium statistical physics [25]. They are essentially straightforward: they operate as image denoisers. During the training process, when given a clean image $x$ , time-step $t$ is sampled from the interval $[0, T]$ , along with a Gaussian noise vector $\\epsilon \\sim N(0, I)$ , resulting in a noised image $x_{t}$ :", + "bbox": [ + 214, + 391, + 787, + 512 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\sqrt {\\alpha_ {t}} x + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 513, + 784, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the scheduled variance $\\alpha_{t}$ varies between 0 and 1, with $\\alpha_0 = 1$ and $\\alpha_{T} = 0$ . The diffusion model then removes the noise to reconstruct the original image $x$ by predicting the noise introduced, achieved through stochastic minimization of the objective function $\\frac{1}{N}\\sum_{i}\\mathbb{E}_{t,\\epsilon}\\mathcal{L}(x_i,t,\\epsilon ;\\theta)$ , where", + "bbox": [ + 214, + 540, + 787, + 602 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(x _ {i}, t, \\epsilon ; \\theta\\right) = \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\alpha_ {t}} x _ {i} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , t\\right) \\| _ {2} ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 616, + 784, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To analyze the correlation between losses and image memorization, we identify images memorized on CIFAR-10 and calculate their loss functions at each time step. Similarly, we sample 256 non-memorized images from the remaining training data and compute their losses at each time step. Fig. 2 shows the comparisons of the losses when the time step is in the interval [0, 600] ( $T = 1000$ ). Memorized images exhibit significantly smaller loss values during this period, indicating that the model tends to reconstruct noise into such images.", + "bbox": [ + 214, + 647, + 787, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Anti-Gradient Control", + "text_level": 1, + "bbox": [ + 215, + 779, + 447, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this subsection, we elaborate on how to utilize the aforementioned loss analysis to devise a training strategy aimed at alleviating the occurrence of memorization.", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9e2da4cfb6dde66cf0dfa6474105892f115d5e2d517b1564f48aa035190e7f61.jpg", + "image_caption": [ + "Fig. 2: Comparison of the losses between memorized and non-memorized images. The solid line represents the averaged losses of memorized and non-memorized images, while the two dashed lines represent the losses of the 15th and 85th percentile data respectively." + ], + "image_footnote": [], + "bbox": [ + 308, + 164, + 676, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Memory Bank. In order to identify images with exceptionally low loss values that are prone to memorization during training, we need to maintain the average losses for each time step. However, computing the average loss at each time step entails substantial computational expenses, as it necessitates evaluating the losses for all images using the model at each time step. Thus, we propose a memory bank to store and update losses during mini-batch training without increasing the time cost. However, the losses generally decrease with the training step growing. Therefore, when calculating the average loss in the memory bank, instead of directly averaging all losses at this time step, the aggregating weights should be higher for losses closer to the current update.", + "bbox": [ + 212, + 448, + 787, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, we initialize an array of length $T$ with zeros, termed the memory bank. The initialization values of the memory bank seem insignificant. After calculating the loss for a mini-batch, we update the memory bank using the Exponential Moving Average (EMA) method based on the loss and the sampled time step, thereby better reflecting the current state of the model:", + "bbox": [ + 212, + 601, + 787, + 676 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nl _ {t} \\leftarrow \\gamma \\cdot l _ {t} + (1 - \\gamma) \\cdot \\mathcal {L} (x, t, \\epsilon ; \\theta), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 691, + 785, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\gamma$ represents the smoothing factor, and $l_{t}$ represents the averaged loss in the memory bank at time step $t$ .", + "bbox": [ + 214, + 720, + 785, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Anti-Gradient Control (AGC). In previous observations, if the model exhibits memorization of a certain sample, the loss value of the model on that sample tends to be abnormally small. Thus, we use the ratio of the training loss of a certain sample to the mean loss in the memory bank at the time step $t$ as", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "IET-AGC: Mitigating Memorization in Diffusion Models", + "bbox": [ + 354, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "a measure to detect memorization:", + "bbox": [ + 215, + 147, + 467, + 160 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nr = \\frac {\\mathcal {L} (x , t , \\epsilon ; \\theta)}{l _ {t}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 171, + 784, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A smaller ratio of $r$ indicates a higher likelihood of the image being memorized. To facilitate this determination, we can establish a configurable threshold denoted as $\\lambda$ . If the loss ratio $r$ falls below this threshold $\\lambda$ , the image is classified as memorized. At this point, we generate a mask that sets the loss value corresponding to this image to zero, i.e. skipping this image in the mini-batch, as shown in the following function,", + "bbox": [ + 215, + 212, + 782, + 303 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (x, t, \\epsilon ; \\theta) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} r < \\lambda \\\\ \\mathcal {L} (x, t, \\epsilon ; \\theta) & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 315, + 784, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since the model has encountered the sample during training, excluding it is unlikely to have a significant impact on the model's performance.", + "bbox": [ + 215, + 364, + 782, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 420, + 374, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 452, + 426, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on CIFAR-10 [16], CIFAR-100 [16], AFHQDOG [6] for unconditional generation, and LAION-10k [27] for text-conditioned generation. CIFAR-10 and CIFAR-100 consist of 50,000 32x32 color images, divided into 10 and 100 classes, respectively. AFHQ-DOG is a subset of the AFHQ dataset with approximately 5,000 512x512 dog images, resized to 64x64 for our experiments. LAION-10k is a subset of LAION [24], comprising 10,000 image-text pairs with each image having a resolution of 256x256 pixels.", + "bbox": [ + 212, + 478, + 784, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation Details of Training. We conduct experiments on training unconditional diffusion models from scratch using the CIFAR-10, CIFAR-100, and AFHQ-DOG datasets. The IET framework divides CIFAR datasets into 10 shards and AFHQ-DOG into 5 shards. Threshold $\\lambda$ is set to 0.5 for CIFAR datasets and 0.714 for AFHQ-DOG. Additionally, we perform experiments on fine-tuning pre-trained diffusion models using the CIFAR-10 dataset, maintaining the same hyperparameters as those used for training from scratch. To demonstrate the effectiveness of our method in text-conditioned diffusion models, we fine-tune Stable Diffusion on LAION-10k. The IET framework divides the LAION-10k dataset into 8 shards, with the threshold $\\lambda$ set to 0.8. The smoothing factor $\\gamma$ is 0.8 for all datasets. Further details are in the supplementary material.", + "bbox": [ + 212, + 585, + 784, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Extracting Memorized Image. We adopt Carlini's detection rule [5] for unconditional generation, considering $\\bar{x}$ as memorized if the $\\ell_2$ distance to its nearest neighbor is significantly lower compared to the $n$ closest neighbors. We modify this rule to:", + "bbox": [ + 212, + 750, + 784, + 810 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {2} (\\bar {x}, x; \\mathbb {S} _ {\\bar {x}} ^ {n}) = \\frac {\\ell_ {2} (\\bar {x} , x)}{\\mathbb {E} _ {y \\in \\mathbb {S} _ {\\bar {x}} ^ {n}} [ \\ell_ {2} (\\bar {x} , y) ]}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 811, + 784, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $n = 50$ in our experiment. A binary classifier is defined as:", + "bbox": [ + 215, + 146, + 683, + 161 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nI s M e m o (\\bar {x}, x; \\mathbb {S} _ {\\bar {x}} ^ {n}, \\delta_ {V}) = 1 _ {\\ell_ {2} (\\bar {x}, x; \\mathbb {S} _ {\\bar {x}} ^ {n}) \\leq \\delta_ {V}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 172, + 785, + 189 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The more images below $\\delta_V$ , the stronger the model's memorization. We generate 65,536 images per model, calculate their $\\ell_2$ distances, and count images below thresholds $\\delta_V$ of 0.4, 0.5, and 0.6 to quantitatively evaluate the model's memorization, denoted as $\\mathrm{MQ}_{0.4}$ , $\\mathrm{MQ}_{0.5}$ and $\\mathrm{MQ}_{0.6}$ .", + "bbox": [ + 212, + 196, + 782, + 257 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We adopt Somepalli's detection rule [27] for text-conditioned generation, quantify memorization using a similarity score derived from the dot product of SSCD features [20] of $\\bar{x}$ and the nearest neighbor $n_0$ :", + "bbox": [ + 212, + 257, + 782, + 301 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\zeta = E (\\bar {x}) ^ {T} \\cdot E (n _ {0}) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 436, + 311, + 785, + 329 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $E(\\cdot)$ is SSCD [20]. The dataset similarity score is then defined as the 95th percentile of this distribution.", + "bbox": [ + 214, + 335, + 782, + 367 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 388, + 439, + 402 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training from Scratch. The experimental results of our method and four competitive methods are shown in Tab. 1. \"Default\" denotes the conventional training approach of DDPM [14]. \"DP-SGD\" denotes the method of Differentially Private Stochastic Gradient Descent [3], which involves clipping and adding noise to the model's gradients to protect privacy, albeit at the cost of some image quality. Carlina et al. [5] found that utilizing DP-SGD could lead to consistent model divergence. To make DP training more stable, we modify the amplitude of the added noise to be the product of the gradient norm and the noise multiplier:", + "bbox": [ + 214, + 411, + 784, + 532 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma = \\left| \\left| \\nabla \\mathcal {L} \\right| \\right| \\times \\tau , \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 449, + 547, + 784, + 561 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\tau$ represents the noise multiplier. \"Adding noise\" denotes a method of directly adding Gaussian noise to the images during training, with a mean of 0 and a variance of 0.1. \"Ambient Diffusion\" [7] protects privacy by training generative models on highly corrupted samples, preventing the model from observing clean training data.", + "bbox": [ + 212, + 568, + 782, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Results in Tab. 1 show that adding noise to the training images or gradients reduces the quality of the generated images. However, it still does not resolve the issue of training image memorization. Despite Ambient Diffusion also reducing memorization, it leads to a significant increase in FID (from 8.81 to 11.7), indicating a notable degradation of image quality. Compared with the default training approach, our method maintains or even slightly improves the generative quality by reducing the FID score. At the same time, our method significantly reduces the diffusion model's memorization of the training data. As shown in Tab. 1, for the $\\mathrm{MQ}_{0.4}$ score, the number of memorized images reduced by $87.3\\%$ , $66.4\\%$ , and $85.3\\%$ compared with the default training on CIFAR-10, CIFAR-100 and AFHQ-DOG, respectively, illustrating the effectiveness of our method.", + "bbox": [ + 214, + 643, + 785, + 809 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Visualization. To provide a more intuitive confirmation of our training method, we visualize the images generated by the model along with their closest", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "IET-AGC: Mitigating Memorization in Diffusion Models", + "bbox": [ + 354, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/dbe66bf6134d44e4d115f7c19e99ba33fc49cca3c4a0e6c6a67342648079ac87.jpg", + "table_caption": [ + "Table 1: Comparisons of unconditional generation on three datasets in terms of memorized quantity denoted as MQ. We also report the FID to evaluate the quality of images produced by the model." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR-10CIFAR-100AFHQ-DOG
\\(MQ_{0.4}\\)\\(MQ_{0.5}\\)\\(MQ_{0.6}\\)↓FID↓\\(MQ_{0.4}\\)\\(MQ_{0.5}\\)\\(MQ_{0.6}\\)↓FID↓\\(MQ_{0.4}\\)\\(MQ_{0.5}\\)\\(MQ_{0.6}\\)↓FID↓
Default11146520308.81429172756209.2912344190533079523.59
Adding Noise197593209194.611791037438386.1811700192952722461.18
DP-SGD [3]148728320012.55--------
Ambient Diffusion [7]2213885111.7--------
IET-AGC141178398.3414476032748.51181154351523722.2
", + "bbox": [ + 215, + 198, + 787, + 310 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/b36517505ffc6e2d207c7f6c6dccfabd05f59f46a1a0f07d04fe5438e9b0b6b0.jpg", + "image_caption": [ + "(a) Default Training" + ], + "image_footnote": [], + "bbox": [ + 271, + 335, + 500, + 512 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d9601d3604a8b0b0d4e62ef6cd6321e245ecafb738a547504cbfe0f84ce83d3e.jpg", + "image_caption": [ + "(b) Our Method", + "Fig. 3: The similar grid image of the default training and our method. Odd-numbered columns represent images from the training set, while even-numbered columns represent images from the generation set that has the smallest $\\ell_2$ distance to the corresponding training set image. The images in the grid are arranged in ascending order of $\\ell_2$ distance, and the selected images from both methods correspond to the same position in terms of $\\ell_2$ distance." + ], + "image_footnote": [], + "bbox": [ + 501, + 335, + 732, + 512 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "counterparts in the training dataset in Fig. 3. There is a similarity between the images from the training set and their corresponding images in the generation set, while our method shows lower similarity compared to the default training method. It is also important to note that our model's FID slightly decreases on all three datasets compared to the default indicating that our method also improves the quality of the generated images.", + "bbox": [ + 212, + 652, + 787, + 743 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Finetuning Unconditional DDPMs. The results are presented in Tab. 2. The IET method shows improvement over the pre-trained in terms of $\\mathrm{MQ}_{0.4}$ and FID, indicating more efficient data forgetting and higher image quality. The AGC method rapidly causes the model to forget the memorized data. However, it also led to a substantial increase in FID. The IET-AGC method not only reduces the FID but also effectively lowers the model's memory rate.", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 2: Fine-tuning results of the pre-trained DDPM on CIFAR-10 dataset for 2 epochs.", + "bbox": [ + 215, + 154, + 444, + 195 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/c8d8dd9132e1a0454f8e405ce0e553ee1ad6294ab99b09eed78f09f3a245e468.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCIFAR-10
\\(MQ_{0.4}\\)\\(MQ_{0.5}\\)\\(MQ_{0.6}\\)↓FID↓
Default11146520308.81
IET5139623678.33
AGC44235131711.8
IET-AGC8340817967.93
", + "bbox": [ + 218, + 208, + 439, + 309 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/5513efa0e034abbec3508d4393e2c7eee925814656bb0f38d7b0206731f46aa8.jpg", + "table_caption": [ + "Table 3: Fine-tuning results of Stable Diffusion model on LAION-10k dataset." + ], + "table_footnote": [], + "table_body": "
MethodSim Score↓Clip Score↑FID↓
Default SD0.6430.518.7
Train TimeMC0.4230.316.6
RC0.5730.616.0
Mitigation [27]CWR0.6130.816.7
Test TimeRT0.5229.518.7
CWR0.5830.118.1
Mitigation [27]GNI0.6230.318.9
Our MethodIET0.4131.316.5
AGC0.5330.618.5
IET-AGC0.3731.316.7
", + "bbox": [ + 447, + 184, + 746, + 329 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e547a1c59dc7c57b8a707d3648a523831aa9a1c75b7f5e1e4e3f5fbe4c254fa9.jpg", + "image_caption": [ + "(a) Distribution of distances to the most similar images in the dataset." + ], + "image_footnote": [], + "bbox": [ + 246, + 359, + 477, + 489 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/315847b1019744e493e278031641b66fbaa1eb26e90a5d000529215d55213d10.jpg", + "image_caption": [ + "(b) Energy distribution. The greater the energy, the more complex the image.", + "Fig. 4: Most skipped images vs least skipped images." + ], + "image_footnote": [], + "bbox": [ + 504, + 359, + 736, + 491 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Finetuning Text-conditional Stable Diffusion. The results are presented in Tab. 3. Somepalli et al. [27] protects privacy by randomizing conditional information during training and inference, thereby reducing the likelihood of the model replicating specific training data. Our method IET-AGC achieves the best overall results with a Similarity Score of 0.37, maintaining the highest Clip Score of 31.3 and a competitive FID of 16.7. This indicates our approach effectively balances memorization and generative quality, outperforming the default and other mitigation methods.", + "bbox": [ + 212, + 556, + 787, + 676 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Analysis of Skipping", + "text_level": 1, + "bbox": [ + 215, + 700, + 431, + 715 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we conduct comparative experiments on the AFHQ-DOG dataset to delve into which types of images are prone to be skipped, as well as the relationship between memorizable images and those that are skipped.", + "bbox": [ + 212, + 726, + 785, + 771 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Frequency of Skipped Images. Throughout the training process, we record the identifiers of skipped images. As shown in Fig. 5, our method does not entail skipping all images. In our approach, $90\\%$ of the images are skipped fewer than", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "IET-AGC: Mitigating Memorization in Diffusion Models", + "bbox": [ + 354, + 113, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/339a94c939876f4af05d39443c70935c383938ece7450c5025c9cd15d8b1e3ff.jpg", + "image_caption": [ + "Fig. 5: Distribution of skipped image counts." + ], + "image_footnote": [], + "bbox": [ + 305, + 170, + 663, + 371 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "648 times (across a total of 2,278 training epochs), indicating that our method can effectively differentiate between different images. This suggests that we are not simply reducing memorization by constraining the model's learning. On the other hand, while our method requires skipping images with exceptionally low loss values, all images still contribute to the model's training.", + "bbox": [ + 212, + 407, + 787, + 484 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Images Most Easily Skipped. We believe these images are more easily skipped for two main reasons. Firstly, data aggregation. We computed the $\\ell_2$ distance between these easily skipped images and all other images in the dataset, as well as between those not easily skipped images and all other images in the dataset. Fig. 4a indicates that the distribution of the skipped images is more clustered. Consistent with the findings of Carlini et al. [5], which suggested that removing duplicate training images effectively reduces memorization capacity, skipping these clustered images can also reduce memorization capacity. Secondly, data simplicity. We performed Fourier transforms on these easily skipped and not easily skipped images to obtain their energy distributions, as shown in Fig. 4b. The easily skipped images have less energy, indicating that they lack finer details. We believe both factors contribute to making these images easier for the model to memorize, thus skipping them has a positive impact on reducing the model's memorization capacity.", + "bbox": [ + 212, + 510, + 787, + 724 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 750, + 390, + 763 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Performance Comparisons of Each Component. To further understand the effectiveness of our approach, we conduct ablation experiments to investigate the individual impacts of different components on CIFAR-10. In addition, in our approach, the dataset is evenly distributed, meaning each data shard is set in", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "an IID (independently and identically distributed) manner. To validate the reasonableness of this dataset configuration, we also establish an experiment where each data shard is set up in a non-independent and identically distributed (non-IID) manner. Similar to [15], we employ the Dirichlet distribution to generate data, thereby establishing such a setting.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results in Tab. 4 show that when the data shards are set in non-IID, Iterative Ensemble Training (IET) successfully reduces the memorization $\\mathrm{(IET_{non - IID})}$ . However, the non-IID splitting reduces the quality of performance since the divergent optimization of different models and aggregating these models affects the performance. Differently, when the data shards are set in IID, $\\mathrm{IET}_{\\mathrm{IID}}$ can reduce both the memorization and FID score. Thus in this paper, we choose to use the IID splitting strategy. Besides, Tab. 4 shows that both AntiGradient Control (AGC) and Iterative En", + "bbox": [ + 212, + 222, + 545, + 417 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "semble Training (IET) reduce memorization effectively and AGC is more effective than the IET on the memorization reduction. Compared to the conventional method, AGC reduces $\\mathrm{MQ}_{0.5}$ by approximately $66\\%$ , while $\\mathrm{IET}_{\\mathrm{IID}}$ reduces it by around $18\\%$ . However, using AGC only shows a slight growth in FID, which affects the quality of the generated images. Interestingly, combining both IET and AGC leads to better FID and memory reduction effects.", + "bbox": [ + 212, + 419, + 787, + 510 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/992b1ba4ca128c5b36c21753a0e531dcae6b9ba6f8e794dfd0266b79b95177f7.jpg", + "table_caption": [ + "Table 4: Performance comparisons of each component." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR-10
\\( MQ_{0.4} \\)\\( MQ_{0.5} \\)\\( MQ_{0.6} \\)↓FID↓
Default11146520308.81
\\( IET_{IID} \\)8938217837.61
\\( IET_{non-IID} \\)69315143912.5
AGC2615497611.36
IET-AGC141178398.34
", + "bbox": [ + 555, + 273, + 789, + 388 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Exploring Parameters Impact on Experimental Results. In this study, we examine how various parameters affect our experimental outcomes. By systematically varying these parameters, we aim to understand how they influence our results and to identify the optimal settings for our experiments. Specifically, we conduct a series of experiments where we change the number of shards, training epochs per aggregation period, skipping threshold $\\lambda$ , and smoothing factor $\\gamma$ of the memory bank. For each variation, we measure the impact on MQ and FID. The default parameters are set with the number of shards $K$ as 10, epochs per shard $E$ as 100, skipping threshold $\\lambda$ as 0.5, and smoothing factor $\\gamma$ as 0.8. Results are reported in Tab. 5.", + "bbox": [ + 212, + 536, + 787, + 686 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Number of Shards $K$ . We investigate the impact of the number of data shards on model performance by setting it to 1, 5, 10, and 15. $K = 1$ means the default training strategy of diffusion models. Results show that the MQ scores of $K = 5, 10, 15$ are all lower than $K = 1$ , indicating that using our IET method can effectively reduce memorization. When $K = 10$ , the MQ score achieves the best performance. However, with the growth of $K$ , the FID score increases. Thus we need to balance the FID and MQ to choose the proper number of shards.", + "bbox": [ + 212, + 688, + 787, + 794 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Training Epochs per Aggregation Period $E$ . We set the total number of training epochs unchanged and conduct experiments by varying the number of epochs per model aggregation period, i.e., the aggregating frequency of IET.", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "IET-AGC: Mitigating Memorization in Diffusion Models", + "bbox": [ + 354, + 113, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results show that both high and low frequencies of aggregation will reduce the performance of the memorization and image quality. Thus, we choose $E = 100$ to achieve the best performance of MQ.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Skipping Threshold $\\lambda$ .", + "text_level": 1, + "bbox": [ + 240, + 191, + 429, + 205 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We evaluate the importance of $\\lambda$ in mitigating the memorization effect by setting the values of $\\lambda$ to 0.4, 0.5, 0.66 and 0.8. A large threshold means skipping more training samples that are easily memorized. Results in Tab. 5 show as $\\lambda$ grows, more memorable training samples are skipped and the memorization phenomenon is further reduced. However, skipping more samples will reduce the model performance, i.e., the generation quality.", + "bbox": [ + 212, + 207, + 431, + 446 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/0fb4fff06b9007ba3f16c34ce04a80d58a4e39b1fb6368d877cdcf040ec9308a.jpg", + "table_caption": [ + "Table 5: Parameters impact on experimental results." + ], + "table_footnote": [], + "table_body": "
ParametersCIFAR-10
\\(MQ_{0.4}\\)\\(MQ_{0.5}\\)\\(MQ_{0.6} \\downarrow\\)FID↓
Number of Shards K111146520308.81
5201297986.37
10141178398.34
152116911439.1
Epoches per Aggregation E102616710189.31
501619411207.76
100141178398.34
2001719313079.01
25024181111410.36
Skipping Threshold λ0.44731015737.72
0.5141178398.34
0.663696088.25
0.834550310.43
Smoothing Factor γ0.52719211938.99
0.8141178398.34
0.9201328868.94
", + "bbox": [ + 442, + 234, + 784, + 445 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Smoothing Factor $\\gamma$", + "text_level": 1, + "bbox": [ + 240, + 449, + 429, + 463 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The selection of the smooth-", + "bbox": [ + 212, + 464, + 431, + 477 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "ing factor $\\gamma$ is also crucial. When we set $\\gamma$ to 0.9, the update of the memory bank becomes overly sluggish, failing to faithfully reflect the current model's loss across different time steps. Conversely, if we set $\\gamma$ to 0.5, the memory bank becomes excessively sensitive, which can lead to instability and fragility during the training process.", + "bbox": [ + 212, + 478, + 787, + 554 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Limitations and Future Work", + "text_level": 1, + "bbox": [ + 214, + 575, + 537, + 592 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Although our strategy has proven effective in reducing the model's memorization capacity, the selection of the threshold $\\lambda$ , which balances between the model's memorization capacity and the quality of the generated images, can only rely on experiments. In future research, we plan to design a quick analysis method for datasets to help determine the appropriate choice of threshold.", + "bbox": [ + 212, + 606, + 787, + 681 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 703, + 359, + 719 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper innovatively proposes a training strategy for diffusion models, which train models on multiple data shards and ignore data with abnormal loss values. This strategy effectively reduces the memory capacity of the model, further strengthening data privacy protection without compromising the quality of the generated images. We firmly believe that this training strategy has a broad application prospect and great development potential in the field of data privacy protection.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 185, + 392, + 204 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by National Natural Science Foundation of China (Grant No. 62306273), National Natural Science Foundation of China (Grant No. 62372341), and the Fundamental Research Funds for the Central Universities (Grant No. 2042024kf0040). The numerical calculations in this paper have been done on the supercomputing system in the Supercomputing Center of Wuhan University.", + "bbox": [ + 212, + 218, + 787, + 309 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 333, + 323, + 349 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Midjourney team (2022), https://www.midjourney.com/home", + "2. Sora team (2024), https://openai.com/sora", + "3. Abadi, M., Chu, A., Goodfellow, I., McMahan, H.B., Mironov, I., Talwar, K., Zhang, L.: Deep learning with differential privacy. In: Proceedings of the 2016 ACM SIGSAC conference on computer and communications security. pp. 308-318 (2016)", + "4. Carlini, N., Tramer, F., Wallace, E., Jagielski, M., Herbert-Voss, A., Lee, K., Roberts, A., Brown, T., Song, D., Erlingsson, U., et al.: Extracting training data from large language models. In: 30th USENIX Security Symposium (USENIX Security 21). pp. 2633-2650 (2021)", + "5. Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramer, F., Balle, B., Ippolito, D., Wallace, E.: Extracting training data from diffusion models. In: 32nd USENIX Security Symposium (USENIX Security 23). pp. 5253-5270 (2023)", + "6. Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8188-8197 (2020)", + "7. Daras, G., Shah, K., Dagan, Y., Gollakota, A., Dimakis, A., Klivans, A.: Ambient diffusion: Learning clean distributions from corrupted data. Advances in Neural Information Processing Systems 36 (2024)", + "8. Dockhorn, T., Cao, T., Vahdat, A., Kreis, K.: Differentially private diffusion models. arXiv preprint arXiv:2210.09929 (2022)", + "9. Dwork, C.: Differential privacy. In: International colloquium on automata, languages, and programming. pp. 1-12. Springer (2006)", + "0. Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5111-5120 (2024)", + "1. Ghalebikesabi, S., Berrada, L., Gowal, S., Ktena, I., Stanforth, R., Hayes, J., De, S., Smith, S.L., Wiles, O., Balle, B.: Differentially private diffusion models generate useful synthetic images. arXiv preprint arXiv:2302.13861 (2023)", + "12. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020)", + "13. Gu, X., Du, C., Pang, T., Li, C., Lin, M., Wang, Y.: On memorization in diffusion models. arXiv preprint arXiv:2310.02664 (2023)" + ], + "bbox": [ + 218, + 367, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "IET-AGC: Mitigating Memorization in Diffusion Models", + "bbox": [ + 354, + 114, + 732, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)", + "15. Hsu, T.M.H., Qi, H., Brown, M.: Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335 (2019)", + "16. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)", + "17. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22691-22702 (2023)", + "18. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017)", + "19. Ni, Z., Wei, L., Li, J., Tang, S., Zhuang, Y., Tian, Q.: Degeneration-tuning: Using scrambled grid shield unwanted concepts from stable diffusion. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 8900-8909 (2023)", + "20. Pizzi, E., Roy, S.D., Ravindra, S.N., Goyal, P., Douze, M.: A self-supervised descriptor for image copy detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14532-14542 (2022)", + "21. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)", + "22. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022)", + "23. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "24. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)", + "25. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)", + "26. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Diffusion art or digital forgery? investigating data replication in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6048-6058 (2023)", + "27. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Understanding and mitigating copying in diffusion models. Advances in Neural Information Processing Systems 36 (2024)", + "28. Webster, R.: A reproducible extraction of training images from diffusion models. arXiv preprint arXiv:2305.08694 (2023)", + "29. Webster, R., Rabin, J., Simon, L., Jurie, F.: This person (probably) exists. identity membership attacks against gan generated faces. arXiv preprint arXiv:2107.06018 (2021)", + "30. Wen, Y., Jain, N., Kirchenbauer, J., Goldblum, M., Geiping, J., Goldstein, T.: Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. Advances in Neural Information Processing Systems 36 (2024)", + "31. Wen, Y., Liu, Y., Chen, C., Lyu, L.: Detecting, explaining, and mitigating memorization in diffusion models. In: The Twelfth International Conference on Learning Representations (2023)" + ], + "bbox": [ + 217, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "X. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Yoon, T., Choi, J.Y., Kwon, S., Ryu, E.K.: Diffusion probabilistic models generalize when they fail to memorize. In: ICML 2023 Workshop on Structured Probabilistic Inference $\\{\\backslash\\&\\}$ Generative Modeling (2023)", + "33. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 217 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "IET-AGC: Mitigating Memorization in Diffusion Models", + "bbox": [ + 354, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models/d774ff0f-2daf-4eeb-815c-c6da2eb3ab17_model.json b/2024/Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models/d774ff0f-2daf-4eeb-815c-c6da2eb3ab17_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dd016181b89e1e550741d36e76176dbe22350c99 --- /dev/null +++ b/2024/Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models/d774ff0f-2daf-4eeb-815c-c6da2eb3ab17_model.json @@ -0,0 +1,2236 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.221, + 0.141, + 0.782, + 0.207 + ], + "angle": 0, + "content": "Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.234, + 0.733, + 0.251 + ], + "angle": 0, + "content": "Xiao Liu\\(^{1*}\\), Xiaoliu Guan\\(^{1*}\\), Yu Wu\\(^{1\\dagger}\\), and Jiaxu Miao\\(^{2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.262, + 0.669, + 0.277 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Computer Science, Wuhan University" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.278, + 0.653, + 0.291 + ], + "angle": 0, + "content": "{xiaoliu, liuxiaoguan, wuyucs}@whu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.291, + 0.723, + 0.305 + ], + "angle": 0, + "content": "\\(^{2}\\) School of Cyber Science and Technology, Sun Yat-sen University" + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.306, + 0.593, + 0.319 + ], + "angle": 0, + "content": "miaojx@mail.sysu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.357, + 0.744, + 0.662 + ], + "angle": 0, + "content": "Abstract. Diffusion models, known for their tremendous ability to generate novel and high-quality samples, have recently raised concerns due to their data memorization behavior, which poses privacy risks. Recent approaches for memory mitigation either only focused on the text modality problem in cross-modal generation tasks or utilized data augmentation strategies. In this paper, we propose a novel training framework for diffusion models from the perspective of visual modality, which is more generic and fundamental for mitigating memorization. To facilitate \"forgetting\" of stored information in diffusion model parameters, we propose an iterative ensemble training strategy by splitting the data into multiple shards for training multiple models and intermittently aggregating these model parameters. Moreover, practical analysis of losses illustrates that the training loss for easily memorable images tends to be obviously lower. Thus, we propose an anti-gradient control method to exclude the sample with a lower loss value from the current mini-batch to avoid memorizing. Extensive experiments and analysis on four datasets are conducted to illustrate the effectiveness of our method, and results show that our method successfully reduces memory capacity while even improving the performance slightly. Moreover, to save the computing cost, we successfully apply our method to fine-tune the well-trained diffusion models by limited epochs, demonstrating the applicability of our method. Code is available in https://github.com/liuxiao-guan/IET_AGC." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.675, + 0.72, + 0.69 + ], + "angle": 0, + "content": "Keywords: Diffusion Models \\(\\cdot\\) Model Memorization \\(\\cdot\\) Data Privacy" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.739, + 0.376, + 0.755 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.771, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Recent advancements in diffusion models have significantly transformed the landscape of image generation. Modern diffusion models, such as Stable Diffusion [23]," + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.812, + 0.365, + 0.825 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.223, + 0.825, + 0.385, + 0.84 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.812, + 0.385, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "Midjourney [1], and SORA [2], have the ability to generate realistic images that are hard for humans to distinguish, demonstrating the unparalleled capabilities in producing diverse images. However, recent works [5, 27, 31] suggested that diffusion models are capable of memorizing images from the training set and reproducing them, which has the potential risk of privacy leakage. To address the severe problem, some works [10, 17, 19, 33] proposed to make diffusion models \"forget\" specific concepts such as a portrait of a certain celebrity, or the style of a particular artist. However, these works can only blacklist specific content that users want to conceal, but cannot completely cover the privacy-sensitive information that the model might remember, which still has a risk of privacy leakage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.318, + 0.788, + 0.515 + ], + "angle": 0, + "content": "Recently, some works [7,26,27,31] have proposed to mitigate diffusion memorization without specific content limitations, thus reducing the risk of diffusion models leaking privacy-sensitive training data. Most of them focused on tackling the training data memorization in text-to-image diffusion models, and proposed data augmentation for captions to reduce model memorization since the insufficient diversity in captions easily leads to training data generation. For instance, Somepalli et al. [27] utilized random caption replacement, random token replacement, caption word repetition, etc., to reduce memorization. Another previous approach Daras et al. [7] proposed to use corrupted images for diffusion model training for memorization reduction. Wen et al. [31] introduced methods for detecting memorized prompts through text-conditional predictions and proposed two strategies to mitigate memorization: minimizing during inference or filtering during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.788, + 0.699 + ], + "angle": 0, + "content": "Although these works represented an important step forward in understanding the memorization issue in diffusion models, they either utilized simple data augmentation strategies or only focused on the easily memorable images that are related to specific captions in cross-modal generation tasks. However, diffusion models have been proven capable of generating images from memory without the need for text guidance [4], and previous methods focused on a limited scope of the memorization issue in diffusion models. Differently, in this paper, we propose a novel training framework for diffusion models from the perspective of visual modality, i.e., IET-AGC (Iterative Ensemble Training with Anti-Gradient Control), which not only reduces memorization fundamentally but also provides a more generic approach for both unconditional and text-conditional diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "First, we propose an iterative ensemble training (IET) framework to mitigate memorization by parameter aggregation. Training data are stored in parameters of diffusion models due to over-optimization, and the model ensemble strategy aggregates parameters to re-organize model knowledge, which contributes to alleviating the memorization of training data. Thus, we separate the training data into several groups, train diffusion models individually, and ensemble them for memorization reduction. However, aggregating models trained on subsets of data may potentially decrease the performance due to the divergent optimizations on different subsets of these models. Motivated by Federated Learning tech" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.356, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "IET-AGC: Mitigating Memorization in Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.147, + 0.785, + 0.177 + ], + "angle": 0, + "content": "niques [18], we iteratively ensemble the models during training, which reduces memorization by multiple aggregation and maintains generation performance." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.179, + 0.786, + 0.358 + ], + "angle": 0, + "content": "Based on our IET training framework, we further introduce an Anti-Gradient Control module to reduce memorization of training data ulteriorly. We conduct practical analysis on losses during training and find that the training loss for easily memorable images tends to be obviously lower than that for less memorable images. Thus, we propose to exclude the sample with a relatively small loss value from the current mini-batch to avoid memorizing these samples. During the training process, as the diffusion model exhibits varying average loss values across different time steps, we maintain a memory bank to store the average loss of each time step. Subsequently, we discard samples with losses below a certain proportion of the average loss. Since the model has encountered the sample during training, excluding it is unlikely to have a significant impact on the model's performance." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.36, + 0.787, + 0.569 + ], + "angle": 0, + "content": "Extensive experiments on four datasets show the significance of our IET-AGC framework. Our method significantly reduces the memorized quantity by \\(87.3\\%\\), \\(66.4\\%\\), and \\(85.3\\%\\) compared with the default training (DDPM [14]) on CIFAR-10 and CIFAR-100 and AFHQ-DOG, respectively. In addition, considering the high cost of re-training existing well-trained diffusion models, we also propose an efficient way to address the memorization issue of existing pre-trained models by simply fine-tuning using our method for several epochs. Experiments show that using our method to fine-tune for just two epochs can significantly reduce the memorization by \\(25.22\\%\\) in unconditional diffusion models on CIFAR-10. Furthermore, when fine-tuning the text-conditional diffusion model, Stable Diffusion, our approach decreases the memorization score by \\(42.18\\%\\) compared to conventional fine-tuning methods. These results demonstrate that our method performs excellently on both unconditional and text-conditional diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.595, + 0.387, + 0.61 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.629, + 0.785, + 0.734 + ], + "angle": 0, + "content": "Memorization in Generative Models. Several studies have examined the memorization capabilities of the generative model. Generative Adversarial Networks (GANs) [12] have been at the forefront of this research area. As Webster et al. [29] demonstrated when applied to face datasets, GANs can occasionally replicate. Prior study [4] explored an adversarial attack on language models like GPT-2 [21], where individual training examples can be recovered, including personally identifiable information and unique text sequences." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.735, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Recent studies have shifted their attention toward diffusion models. Somepalli et al. [26] found that diffusion models accurately recall and replicate training images, especially noted with models like the Stable Diffusion model [23]. Building upon this discovery, Carlini et al. [5] developed a tailored black-box attack for diffusion models. They generated images and implemented a membership inference attack to assess density. Webster et al. [28] demonstrated a more efficient extraction attack with fewer network evaluations, identified \"template" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.253 + ], + "angle": 0, + "content": "verbatim,\" and discussed its persistence in newer systems. Recent research has shifted towards exploring the theoretical aspects of memory in diffusion models. [32] discovered that generalization and memorization are mutually exclusive occurrences and further demonstrated that the dichotomy between memorization and generalization can be apparent at the class level. Gu et al. [13] extensively studied how factors like data dimension, model size, time embedding, and class conditions affect the memory capacity of the diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.28, + 0.787, + 0.612 + ], + "angle": 0, + "content": "Memorization Mitigation. The mitigation measures have primarily been concerned with filtering inputs and deduplication. For example, Stable Diffusion employed well-trained detectors to identify unsuitable generated content. However, these temporary solutions can be easily bypassed [22, 30] and do not effectively prevent or lessen copying behavior on a broad scale. Kumari et al. [17] designed an algorithm to align the image distribution with a specific style, instance, or text prompt they aim to remove, to the distribution related to a core concept. This stopped the model from producing target concepts based on its text condition. However, these approaches are inefficient because they necessitate a list of all concepts to be erased, and have not addressed the key issue of how to reduce the memory capacity of the model. [8, 11] explored the use of differential privacy (DP) [9] to train diffusion models or fine-tune ImageNet pre-trained models. However, their focus was on ensuring the privacy of the training of diffusion models, not on the privacy of the images generated by the diffusion models. Daras et al. [7] introduced a technique for training diffusion models utilizing tainted data. By incorporating additional corruption before applying noise, their methodology prevents the model from overfitting to the training data. But their training requires a considerable amount of time. [27, 31] also suggested a series of recommendations to mitigate copying such as randomly replacing the caption of an image with a random sequence of words, but most of which are limited to text-to-image models. Our work focuses on the nature of memorization in diffusion models, especially for unconditional ones." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.641, + 0.33, + 0.656 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.678, + 0.785, + 0.708 + ], + "angle": 0, + "content": "In this section, we present our methodology aimed at mitigating the memorization in diffusion models, without sacrificing excessive image quality." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.735, + 0.493, + 0.75 + ], + "angle": 0, + "content": "3.1 Iterative Ensemble Training" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.765, + 0.799, + 0.84 + ], + "angle": 0, + "content": "Training data are stored in parameters of diffusion models due to over-optimization, and the model ensemble strategy aggregates parameters to re-organize model knowledge. Thus, to mitigate the memorization, we propose a method that trains multiple diffusion models on different data shards of a dataset, merges them after a certain period, and then repeats these two stages iteratively." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.355, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "IET-AGC: Mitigating Memorization in Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.148, + 0.473, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.328, + 0.357, + 0.342 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.149, + 0.776, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.328, + 0.651, + 0.343 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.455 + ], + "angle": 0, + "content": "Fig. 1: Overview of our IET-AGC method. (a) Iterative Ensemble Training (IET): we divide the dataset \\(D\\) into \\(K\\) different data shards. Each shard \\(D_{i}\\) trains a separate diffusion model \\(\\theta_{i}\\). After a period of training, these models are merged by averaging and repeating this training strategy. (b) Anti-Gradient Control (AGC): during the training process, we dynamically update and maintain a memory bank of losses at each time step. For loss values smaller than \\(\\lambda\\) times the corresponding memory bank, we exclude these losses to prevent the model from memorizing such images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.788, + 0.56 + ], + "angle": 0, + "content": "Training on Different Data Shards. Unlike the training methods of previous diffusion models, which train a single model on the entire dataset once. In this paper, we divide a dataset into an equal number of shards and then train the corresponding diffusion models on each separate part. If the dataset contains class information, we divide it equally along the class dimension, so that each data shard contains all classes, with the same number of instances for each class." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.788, + 0.624 + ], + "angle": 0, + "content": "Specifically, suppose the dataset \\(D\\) contains \\(C\\) classes, with each class having \\(N_{c}\\) samples, and the total number of samples is \\(N = \\sum_{c=1}^{C} N_{c}\\). We divide the dataset into \\(K\\) equal parts, with each part containing \\(\\frac{N}{K}\\) samples. So the \\(i\\)th part of the dataset \\(D_{i}\\) is represented as" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.638, + 0.786, + 0.679 + ], + "angle": 0, + "content": "\\[\nD _ {i} = \\bigcup_ {c = 1} ^ {C} \\left\\{\\left(x _ {c, i, k}, y _ {c, i, k}\\right) \\mid k = 1, 2, \\dots , \\frac {N _ {c}}{K} \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.788, + 0.755 + ], + "angle": 0, + "content": "where \\((x,y)\\) represents the sample and its corresponding label. Notably, \\(C = 1\\) indicates that the dataset does not contain class information. Then, each part \\(i\\) trains a separate diffusion model \\(\\theta_{i}\\) scaled by the learning rate \\(\\eta\\) on its own dataset," + }, + { + "type": "equation", + "bbox": [ + 0.43, + 0.762, + 0.786, + 0.778 + ], + "angle": 0, + "content": "\\[\n\\theta_ {i} \\leftarrow \\theta_ {i} - \\eta \\nabla \\mathcal {L} \\left(\\theta_ {i}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Merging the Multiple Diffusion Models. After training for some time, each part obtains a different diffusion model. We simply average the weights of all" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.145, + 0.504, + 0.161 + ], + "angle": 0, + "content": "models \\(\\theta_{i}\\) to obtain a global model \\(\\widehat{\\theta}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.451, + 0.176, + 0.786, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\theta} \\leftarrow \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\theta_ {i}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.23, + 0.788, + 0.335 + ], + "angle": 0, + "content": "Then, we repeat the two stages of training on separate shards of the data and merging models, where we use the obtained global model as the initial model for the first stage. As each shard contains only \\(\\frac{1}{K}\\) of the total data, the training time for each model is proportionally reduced, maintaining the overall computational cost nearly constant compared to training a single model on the entire dataset. The only additional computational cost comes from the periodic merging of models, which is minimal." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.362, + 0.375, + 0.379 + ], + "angle": 0, + "content": "3.2 Loss Analysis" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.392, + 0.789, + 0.513 + ], + "angle": 0, + "content": "To further reduce memorization of training data, we delve deeper into the causes of memorization phenomena, specifically analyzing it through the lens of the loss. We begin by establishing the fundamental notation linked with diffusion models. Diffusion models [14] originate from the non-equilibrium statistical physics [25]. They are essentially straightforward: they operate as image denoisers. During the training process, when given a clean image \\( x \\), time-step \\( t \\) is sampled from the interval \\( [0, T] \\), along with a Gaussian noise vector \\( \\epsilon \\sim N(0, I) \\), resulting in a noised image \\( x_{t} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.515, + 0.785, + 0.532 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\sqrt {\\alpha_ {t}} x + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.541, + 0.788, + 0.603 + ], + "angle": 0, + "content": "where the scheduled variance \\(\\alpha_{t}\\) varies between 0 and 1, with \\(\\alpha_0 = 1\\) and \\(\\alpha_{T} = 0\\). The diffusion model then removes the noise to reconstruct the original image \\(x\\) by predicting the noise introduced, achieved through stochastic minimization of the objective function \\(\\frac{1}{N}\\sum_{i}\\mathbb{E}_{t,\\epsilon}\\mathcal{L}(x_i,t,\\epsilon ;\\theta)\\), where" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.617, + 0.785, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(x _ {i}, t, \\epsilon ; \\theta\\right) = \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\alpha_ {t}} x _ {i} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , t\\right) \\| _ {2} ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.648, + 0.788, + 0.754 + ], + "angle": 0, + "content": "To analyze the correlation between losses and image memorization, we identify images memorized on CIFAR-10 and calculate their loss functions at each time step. Similarly, we sample 256 non-memorized images from the remaining training data and compute their losses at each time step. Fig. 2 shows the comparisons of the losses when the time step is in the interval [0, 600] (\\(T = 1000\\)). Memorized images exhibit significantly smaller loss values during this period, indicating that the model tends to reconstruct noise into such images." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.78, + 0.449, + 0.794 + ], + "angle": 0, + "content": "3.3 Anti-Gradient Control" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this subsection, we elaborate on how to utilize the aforementioned loss analysis to devise a training strategy aimed at alleviating the occurrence of memorization." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.356, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "IET-AGC: Mitigating Memorization in Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.165, + 0.677, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.378, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Fig. 2: Comparison of the losses between memorized and non-memorized images. The solid line represents the averaged losses of memorized and non-memorized images, while the two dashed lines represent the losses of the 15th and 85th percentile data respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.449, + 0.788, + 0.601 + ], + "angle": 0, + "content": "Memory Bank. In order to identify images with exceptionally low loss values that are prone to memorization during training, we need to maintain the average losses for each time step. However, computing the average loss at each time step entails substantial computational expenses, as it necessitates evaluating the losses for all images using the model at each time step. Thus, we propose a memory bank to store and update losses during mini-batch training without increasing the time cost. However, the losses generally decrease with the training step growing. Therefore, when calculating the average loss in the memory bank, instead of directly averaging all losses at this time step, the aggregating weights should be higher for losses closer to the current update." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.602, + 0.788, + 0.678 + ], + "angle": 0, + "content": "Specifically, we initialize an array of length \\( T \\) with zeros, termed the memory bank. The initialization values of the memory bank seem insignificant. After calculating the loss for a mini-batch, we update the memory bank using the Exponential Moving Average (EMA) method based on the loss and the sampled time step, thereby better reflecting the current state of the model:" + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.692, + 0.786, + 0.709 + ], + "angle": 0, + "content": "\\[\nl _ {t} \\leftarrow \\gamma \\cdot l _ {t} + (1 - \\gamma) \\cdot \\mathcal {L} (x, t, \\epsilon ; \\theta), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.722, + 0.787, + 0.753 + ], + "angle": 0, + "content": "where \\(\\gamma\\) represents the smoothing factor, and \\(l_{t}\\) represents the averaged loss in the memory bank at time step \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Anti-Gradient Control (AGC). In previous observations, if the model exhibits memorization of a certain sample, the loss value of the model on that sample tends to be abnormally small. Thus, we use the ratio of the training loss of a certain sample to the mean loss in the memory bank at the time step \\( t \\) as" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.148, + 0.468, + 0.161 + ], + "angle": 0, + "content": "a measure to detect memorization:" + }, + { + "type": "equation", + "bbox": [ + 0.445, + 0.172, + 0.785, + 0.202 + ], + "angle": 0, + "content": "\\[\nr = \\frac {\\mathcal {L} (x , t , \\epsilon ; \\theta)}{l _ {t}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.213, + 0.784, + 0.304 + ], + "angle": 0, + "content": "A smaller ratio of \\( r \\) indicates a higher likelihood of the image being memorized. To facilitate this determination, we can establish a configurable threshold denoted as \\( \\lambda \\). If the loss ratio \\( r \\) falls below this threshold \\( \\lambda \\), the image is classified as memorized. At this point, we generate a mask that sets the loss value corresponding to this image to zero, i.e. skipping this image in the mini-batch, as shown in the following function," + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.316, + 0.785, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (x, t, \\epsilon ; \\theta) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} r < \\lambda \\\\ \\mathcal {L} (x, t, \\epsilon ; \\theta) & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.366, + 0.784, + 0.396 + ], + "angle": 0, + "content": "Since the model has encountered the sample during training, excluding it is unlikely to have a significant impact on the model's performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.421, + 0.375, + 0.438 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.453, + 0.427, + 0.468 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.785, + 0.585 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on CIFAR-10 [16], CIFAR-100 [16], AFHQDOG [6] for unconditional generation, and LAION-10k [27] for text-conditioned generation. CIFAR-10 and CIFAR-100 consist of 50,000 32x32 color images, divided into 10 and 100 classes, respectively. AFHQ-DOG is a subset of the AFHQ dataset with approximately 5,000 512x512 dog images, resized to 64x64 for our experiments. LAION-10k is a subset of LAION [24], comprising 10,000 image-text pairs with each image having a resolution of 256x256 pixels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.586, + 0.785, + 0.751 + ], + "angle": 0, + "content": "Implementation Details of Training. We conduct experiments on training unconditional diffusion models from scratch using the CIFAR-10, CIFAR-100, and AFHQ-DOG datasets. The IET framework divides CIFAR datasets into 10 shards and AFHQ-DOG into 5 shards. Threshold \\(\\lambda\\) is set to 0.5 for CIFAR datasets and 0.714 for AFHQ-DOG. Additionally, we perform experiments on fine-tuning pre-trained diffusion models using the CIFAR-10 dataset, maintaining the same hyperparameters as those used for training from scratch. To demonstrate the effectiveness of our method in text-conditioned diffusion models, we fine-tune Stable Diffusion on LAION-10k. The IET framework divides the LAION-10k dataset into 8 shards, with the threshold \\(\\lambda\\) set to 0.8. The smoothing factor \\(\\gamma\\) is 0.8 for all datasets. Further details are in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.751, + 0.785, + 0.811 + ], + "angle": 0, + "content": "Extracting Memorized Image. We adopt Carlini's detection rule [5] for unconditional generation, considering \\(\\bar{x}\\) as memorized if the \\(\\ell_2\\) distance to its nearest neighbor is significantly lower compared to the \\(n\\) closest neighbors. We modify this rule to:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.812, + 0.785, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\ell_ {2} (\\bar {x}, x; \\mathbb {S} _ {\\bar {x}} ^ {n}) = \\frac {\\ell_ {2} (\\bar {x} , x)}{\\mathbb {E} _ {y \\in \\mathbb {S} _ {\\bar {x}} ^ {n}} [ \\ell_ {2} (\\bar {x} , y) ]}, \\tag {9}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.355, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "IET-AGC: Mitigating Memorization in Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.684, + 0.162 + ], + "angle": 0, + "content": "where \\( n = 50 \\) in our experiment. A binary classifier is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.173, + 0.786, + 0.19 + ], + "angle": 0, + "content": "\\[\nI s M e m o (\\bar {x}, x; \\mathbb {S} _ {\\bar {x}} ^ {n}, \\delta_ {V}) = 1 _ {\\ell_ {2} (\\bar {x}, x; \\mathbb {S} _ {\\bar {x}} ^ {n}) \\leq \\delta_ {V}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.197, + 0.784, + 0.258 + ], + "angle": 0, + "content": "The more images below \\(\\delta_V\\), the stronger the model's memorization. We generate 65,536 images per model, calculate their \\(\\ell_2\\) distances, and count images below thresholds \\(\\delta_V\\) of 0.4, 0.5, and 0.6 to quantitatively evaluate the model's memorization, denoted as \\(\\mathrm{MQ}_{0.4}\\), \\(\\mathrm{MQ}_{0.5}\\) and \\(\\mathrm{MQ}_{0.6}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.258, + 0.784, + 0.303 + ], + "angle": 0, + "content": "We adopt Somepalli's detection rule [27] for text-conditioned generation, quantify memorization using a similarity score derived from the dot product of SSCD features [20] of \\(\\bar{x}\\) and the nearest neighbor \\(n_0\\):" + }, + { + "type": "equation", + "bbox": [ + 0.437, + 0.313, + 0.786, + 0.33 + ], + "angle": 0, + "content": "\\[\n\\zeta = E (\\bar {x}) ^ {T} \\cdot E (n _ {0}) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.337, + 0.784, + 0.368 + ], + "angle": 0, + "content": "where \\( E(\\cdot) \\) is SSCD [20]. The dataset similarity score is then defined as the 95th percentile of this distribution." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.389, + 0.441, + 0.403 + ], + "angle": 0, + "content": "4.2 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.412, + 0.785, + 0.533 + ], + "angle": 0, + "content": "Training from Scratch. The experimental results of our method and four competitive methods are shown in Tab. 1. \"Default\" denotes the conventional training approach of DDPM [14]. \"DP-SGD\" denotes the method of Differentially Private Stochastic Gradient Descent [3], which involves clipping and adding noise to the model's gradients to protect privacy, albeit at the cost of some image quality. Carlina et al. [5] found that utilizing DP-SGD could lead to consistent model divergence. To make DP training more stable, we modify the amplitude of the added noise to be the product of the gradient norm and the noise multiplier:" + }, + { + "type": "equation", + "bbox": [ + 0.45, + 0.549, + 0.785, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\sigma = \\left| \\left| \\nabla \\mathcal {L} \\right| \\right| \\times \\tau , \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.784, + 0.644 + ], + "angle": 0, + "content": "where \\(\\tau\\) represents the noise multiplier. \"Adding noise\" denotes a method of directly adding Gaussian noise to the images during training, with a mean of 0 and a variance of 0.1. \"Ambient Diffusion\" [7] protects privacy by training generative models on highly corrupted samples, preventing the model from observing clean training data." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.645, + 0.786, + 0.81 + ], + "angle": 0, + "content": "Results in Tab. 1 show that adding noise to the training images or gradients reduces the quality of the generated images. However, it still does not resolve the issue of training image memorization. Despite Ambient Diffusion also reducing memorization, it leads to a significant increase in FID (from 8.81 to 11.7), indicating a notable degradation of image quality. Compared with the default training approach, our method maintains or even slightly improves the generative quality by reducing the FID score. At the same time, our method significantly reduces the diffusion model's memorization of the training data. As shown in Tab. 1, for the \\(\\mathrm{MQ}_{0.4}\\) score, the number of memorized images reduced by \\(87.3\\%\\), \\(66.4\\%\\), and \\(85.3\\%\\) compared with the default training on CIFAR-10, CIFAR-100 and AFHQ-DOG, respectively, illustrating the effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Visualization. To provide a more intuitive confirmation of our training method, we visualize the images generated by the model along with their closest" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Comparisons of unconditional generation on three datasets in terms of memorized quantity denoted as MQ. We also report the FID to evaluate the quality of images produced by the model." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.199, + 0.789, + 0.311 + ], + "angle": 0, + "content": "
MethodCIFAR-10CIFAR-100AFHQ-DOG
\\(MQ_{0.4}\\)\\(MQ_{0.5}\\)\\(MQ_{0.6}\\)↓FID↓\\(MQ_{0.4}\\)\\(MQ_{0.5}\\)\\(MQ_{0.6}\\)↓FID↓\\(MQ_{0.4}\\)\\(MQ_{0.5}\\)\\(MQ_{0.6}\\)↓FID↓
Default11146520308.81429172756209.2912344190533079523.59
Adding Noise197593209194.611791037438386.1811700192952722461.18
DP-SGD [3]148728320012.55--------
Ambient Diffusion [7]2213885111.7--------
IET-AGC141178398.3414476032748.51181154351523722.2
" + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.336, + 0.501, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.327, + 0.515, + 0.447, + 0.527 + ], + "angle": 0, + "content": "(a) Default Training" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.336, + 0.733, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.515, + 0.667, + 0.527 + ], + "angle": 0, + "content": "(b) Our Method" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.54, + 0.788, + 0.625 + ], + "angle": 0, + "content": "Fig. 3: The similar grid image of the default training and our method. Odd-numbered columns represent images from the training set, while even-numbered columns represent images from the generation set that has the smallest \\(\\ell_2\\) distance to the corresponding training set image. The images in the grid are arranged in ascending order of \\(\\ell_2\\) distance, and the selected images from both methods correspond to the same position in terms of \\(\\ell_2\\) distance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.653, + 0.788, + 0.744 + ], + "angle": 0, + "content": "counterparts in the training dataset in Fig. 3. There is a similarity between the images from the training set and their corresponding images in the generation set, while our method shows lower similarity compared to the default training method. It is also important to note that our model's FID slightly decreases on all three datasets compared to the default indicating that our method also improves the quality of the generated images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Finetuning Unconditional DDPMs. The results are presented in Tab. 2. The IET method shows improvement over the pre-trained in terms of \\(\\mathrm{MQ}_{0.4}\\) and FID, indicating more efficient data forgetting and higher image quality. The AGC method rapidly causes the model to forget the memorized data. However, it also led to a substantial increase in FID. The IET-AGC method not only reduces the FID but also effectively lowers the model's memory rate." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.355, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "IET-AGC: Mitigating Memorization in Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.155, + 0.446, + 0.196 + ], + "angle": 0, + "content": "Table 2: Fine-tuning results of the pre-trained DDPM on CIFAR-10 dataset for 2 epochs." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.209, + 0.44, + 0.31 + ], + "angle": 0, + "content": "
MethodCIFAR-10
\\(MQ_{0.4}\\)\\(MQ_{0.5}\\)\\(MQ_{0.6}\\)↓FID↓
Default11146520308.81
IET5139623678.33
AGC44235131711.8
IET-AGC8340817967.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.446, + 0.145, + 0.734, + 0.172 + ], + "angle": 0, + "content": "Table 3: Fine-tuning results of Stable Diffusion model on LAION-10k dataset." + }, + { + "type": "table", + "bbox": [ + 0.448, + 0.185, + 0.747, + 0.33 + ], + "angle": 0, + "content": "
MethodSim Score↓Clip Score↑FID↓
Default SD0.6430.518.7
Train TimeMC0.4230.316.6
RC0.5730.616.0
Mitigation [27]CWR0.6130.816.7
Test TimeRT0.5229.518.7
CWR0.5830.118.1
Mitigation [27]GNI0.6230.318.9
Our MethodIET0.4131.316.5
AGC0.5330.618.5
IET-AGC0.3731.316.7
" + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.36, + 0.478, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.495, + 0.499, + 0.516 + ], + "angle": 0, + "content": "(a) Distribution of distances to the most similar images in the dataset." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.36, + 0.738, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.495, + 0.76, + 0.517 + ], + "angle": 0, + "content": "(b) Energy distribution. The greater the energy, the more complex the image." + }, + { + "type": "image_caption", + "bbox": [ + 0.322, + 0.53, + 0.678, + 0.544 + ], + "angle": 0, + "content": "Fig. 4: Most skipped images vs least skipped images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.788, + 0.678 + ], + "angle": 0, + "content": "Finetuning Text-conditional Stable Diffusion. The results are presented in Tab. 3. Somepalli et al. [27] protects privacy by randomizing conditional information during training and inference, thereby reducing the likelihood of the model replicating specific training data. Our method IET-AGC achieves the best overall results with a Similarity Score of 0.37, maintaining the highest Clip Score of 31.3 and a competitive FID of 16.7. This indicates our approach effectively balances memorization and generative quality, outperforming the default and other mitigation methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.701, + 0.432, + 0.716 + ], + "angle": 0, + "content": "4.3 Analysis of Skipping" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.727, + 0.786, + 0.772 + ], + "angle": 0, + "content": "In this section, we conduct comparative experiments on the AFHQ-DOG dataset to delve into which types of images are prone to be skipped, as well as the relationship between memorizable images and those that are skipped." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Frequency of Skipped Images. Throughout the training process, we record the identifiers of skipped images. As shown in Fig. 5, our method does not entail skipping all images. In our approach, \\(90\\%\\) of the images are skipped fewer than" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.171, + 0.665, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.379, + 0.653, + 0.394 + ], + "angle": 0, + "content": "Fig. 5: Distribution of skipped image counts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.409, + 0.788, + 0.485 + ], + "angle": 0, + "content": "648 times (across a total of 2,278 training epochs), indicating that our method can effectively differentiate between different images. This suggests that we are not simply reducing memorization by constraining the model's learning. On the other hand, while our method requires skipping images with exceptionally low loss values, all images still contribute to the model's training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.788, + 0.725 + ], + "angle": 0, + "content": "Images Most Easily Skipped. We believe these images are more easily skipped for two main reasons. Firstly, data aggregation. We computed the \\(\\ell_2\\) distance between these easily skipped images and all other images in the dataset, as well as between those not easily skipped images and all other images in the dataset. Fig. 4a indicates that the distribution of the skipped images is more clustered. Consistent with the findings of Carlini et al. [5], which suggested that removing duplicate training images effectively reduces memorization capacity, skipping these clustered images can also reduce memorization capacity. Secondly, data simplicity. We performed Fourier transforms on these easily skipped and not easily skipped images to obtain their energy distributions, as shown in Fig. 4b. The easily skipped images have less energy, indicating that they lack finer details. We believe both factors contribute to making these images easier for the model to memorize, thus skipping them has a positive impact on reducing the model's memorization capacity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.75, + 0.391, + 0.765 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Performance Comparisons of Each Component. To further understand the effectiveness of our approach, we conduct ablation experiments to investigate the individual impacts of different components on CIFAR-10. In addition, in our approach, the dataset is evenly distributed, meaning each data shard is set in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.355, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "IET-AGC: Mitigating Memorization in Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "an IID (independently and identically distributed) manner. To validate the reasonableness of this dataset configuration, we also establish an experiment where each data shard is set up in a non-independent and identically distributed (non-IID) manner. Similar to [15], we employ the Dirichlet distribution to generate data, thereby establishing such a setting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.546, + 0.419 + ], + "angle": 0, + "content": "Results in Tab. 4 show that when the data shards are set in non-IID, Iterative Ensemble Training (IET) successfully reduces the memorization \\(\\mathrm{(IET_{non - IID})}\\). However, the non-IID splitting reduces the quality of performance since the divergent optimization of different models and aggregating these models affects the performance. Differently, when the data shards are set in IID, \\(\\mathrm{IET}_{\\mathrm{IID}}\\) can reduce both the memorization and FID score. Thus in this paper, we choose to use the IID splitting strategy. Besides, Tab. 4 shows that both AntiGradient Control (AGC) and Iterative En" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.788, + 0.511 + ], + "angle": 0, + "content": "semble Training (IET) reduce memorization effectively and AGC is more effective than the IET on the memorization reduction. Compared to the conventional method, AGC reduces \\(\\mathrm{MQ}_{0.5}\\) by approximately \\(66\\%\\), while \\(\\mathrm{IET}_{\\mathrm{IID}}\\) reduces it by around \\(18\\%\\). However, using AGC only shows a slight growth in FID, which affects the quality of the generated images. Interestingly, combining both IET and AGC leads to better FID and memory reduction effects." + }, + { + "type": "table_caption", + "bbox": [ + 0.555, + 0.245, + 0.786, + 0.272 + ], + "angle": 0, + "content": "Table 4: Performance comparisons of each component." + }, + { + "type": "table", + "bbox": [ + 0.556, + 0.274, + 0.79, + 0.389 + ], + "angle": 0, + "content": "
MethodCIFAR-10
\\( MQ_{0.4} \\)\\( MQ_{0.5} \\)\\( MQ_{0.6} \\)↓FID↓
Default11146520308.81
\\( IET_{IID} \\)8938217837.61
\\( IET_{non-IID} \\)69315143912.5
AGC2615497611.36
IET-AGC141178398.34
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.687 + ], + "angle": 0, + "content": "Exploring Parameters Impact on Experimental Results. In this study, we examine how various parameters affect our experimental outcomes. By systematically varying these parameters, we aim to understand how they influence our results and to identify the optimal settings for our experiments. Specifically, we conduct a series of experiments where we change the number of shards, training epochs per aggregation period, skipping threshold \\(\\lambda\\), and smoothing factor \\(\\gamma\\) of the memory bank. For each variation, we measure the impact on MQ and FID. The default parameters are set with the number of shards \\(K\\) as 10, epochs per shard \\(E\\) as 100, skipping threshold \\(\\lambda\\) as 0.5, and smoothing factor \\(\\gamma\\) as 0.8. Results are reported in Tab. 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Number of Shards \\( K \\). We investigate the impact of the number of data shards on model performance by setting it to 1, 5, 10, and 15. \\( K = 1 \\) means the default training strategy of diffusion models. Results show that the MQ scores of \\( K = 5, 10, 15 \\) are all lower than \\( K = 1 \\), indicating that using our IET method can effectively reduce memorization. When \\( K = 10 \\), the MQ score achieves the best performance. However, with the growth of \\( K \\), the FID score increases. Thus we need to balance the FID and MQ to choose the proper number of shards." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Training Epochs per Aggregation Period \\( E \\). We set the total number of training epochs unchanged and conduct experiments by varying the number of epochs per model aggregation period, i.e., the aggregating frequency of IET." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.192 + ], + "angle": 0, + "content": "Results show that both high and low frequencies of aggregation will reduce the performance of the memorization and image quality. Thus, we choose \\( E = 100 \\) to achieve the best performance of MQ." + }, + { + "type": "title", + "bbox": [ + 0.241, + 0.193, + 0.431, + 0.207 + ], + "angle": 0, + "content": "Skipping Threshold \\(\\lambda\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.432, + 0.448 + ], + "angle": 0, + "content": "We evaluate the importance of \\(\\lambda\\) in mitigating the memorization effect by setting the values of \\(\\lambda\\) to 0.4, 0.5, 0.66 and 0.8. A large threshold means skipping more training samples that are easily memorized. Results in Tab. 5 show as \\(\\lambda\\) grows, more memorable training samples are skipped and the memorization phenomenon is further reduced. However, skipping more samples will reduce the model performance, i.e., the generation quality." + }, + { + "type": "table_caption", + "bbox": [ + 0.442, + 0.208, + 0.786, + 0.234 + ], + "angle": 0, + "content": "Table 5: Parameters impact on experimental results." + }, + { + "type": "table", + "bbox": [ + 0.443, + 0.236, + 0.785, + 0.446 + ], + "angle": 0, + "content": "
ParametersCIFAR-10
\\(MQ_{0.4}\\)\\(MQ_{0.5}\\)\\(MQ_{0.6} \\downarrow\\)FID↓
Number of Shards K111146520308.81
5201297986.37
10141178398.34
152116911439.1
Epoches per Aggregation E102616710189.31
501619411207.76
100141178398.34
2001719313079.01
25024181111410.36
Skipping Threshold λ0.44731015737.72
0.5141178398.34
0.663696088.25
0.834550310.43
Smoothing Factor γ0.52719211938.99
0.8141178398.34
0.9201328868.94
" + }, + { + "type": "title", + "bbox": [ + 0.241, + 0.45, + 0.431, + 0.464 + ], + "angle": 0, + "content": "Smoothing Factor \\(\\gamma\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.432, + 0.478 + ], + "angle": 0, + "content": "The selection of the smooth-" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.555 + ], + "angle": 0, + "content": "ing factor \\(\\gamma\\) is also crucial. When we set \\(\\gamma\\) to 0.9, the update of the memory bank becomes overly sluggish, failing to faithfully reflect the current model's loss across different time steps. Conversely, if we set \\(\\gamma\\) to 0.5, the memory bank becomes excessively sensitive, which can lead to instability and fragility during the training process." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.576, + 0.538, + 0.593 + ], + "angle": 0, + "content": "5 Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.788, + 0.683 + ], + "angle": 0, + "content": "Although our strategy has proven effective in reducing the model's memorization capacity, the selection of the threshold \\(\\lambda\\), which balances between the model's memorization capacity and the quality of the generated images, can only rely on experiments. In future research, we plan to design a quick analysis method for datasets to help determine the appropriate choice of threshold." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.704, + 0.36, + 0.72 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This paper innovatively proposes a training strategy for diffusion models, which train models on multiple data shards and ignore data with abnormal loss values. This strategy effectively reduces the memory capacity of the model, further strengthening data privacy protection without compromising the quality of the generated images. We firmly believe that this training strategy has a broad application prospect and great development potential in the field of data privacy protection." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.355, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "IET-AGC: Mitigating Memorization in Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.186, + 0.393, + 0.205 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.219, + 0.788, + 0.31 + ], + "angle": 0, + "content": "This work was supported by National Natural Science Foundation of China (Grant No. 62306273), National Natural Science Foundation of China (Grant No. 62372341), and the Fundamental Research Funds for the Central Universities (Grant No. 2042024kf0040). The numerical calculations in this paper have been done on the supercomputing system in the Supercomputing Center of Wuhan University." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.334, + 0.324, + 0.35 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.368, + 0.655, + 0.381 + ], + "angle": 0, + "content": "1. Midjourney team (2022), https://www.midjourney.com/home" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.382, + 0.549, + 0.394 + ], + "angle": 0, + "content": "2. Sora team (2024), https://openai.com/sora" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.396, + 0.786, + 0.45 + ], + "angle": 0, + "content": "3. Abadi, M., Chu, A., Goodfellow, I., McMahan, H.B., Mironov, I., Talwar, K., Zhang, L.: Deep learning with differential privacy. In: Proceedings of the 2016 ACM SIGSAC conference on computer and communications security. pp. 308-318 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.451, + 0.786, + 0.506 + ], + "angle": 0, + "content": "4. Carlini, N., Tramer, F., Wallace, E., Jagielski, M., Herbert-Voss, A., Lee, K., Roberts, A., Brown, T., Song, D., Erlingsson, U., et al.: Extracting training data from large language models. In: 30th USENIX Security Symposium (USENIX Security 21). pp. 2633-2650 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.507, + 0.786, + 0.548 + ], + "angle": 0, + "content": "5. Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramer, F., Balle, B., Ippolito, D., Wallace, E.: Extracting training data from diffusion models. In: 32nd USENIX Security Symposium (USENIX Security 23). pp. 5253-5270 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.549, + 0.786, + 0.589 + ], + "angle": 0, + "content": "6. Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8188-8197 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.59, + 0.786, + 0.631 + ], + "angle": 0, + "content": "7. Daras, G., Shah, K., Dagan, Y., Gollakota, A., Dimakis, A., Klivans, A.: Ambient diffusion: Learning clean distributions from corrupted data. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.632, + 0.786, + 0.659 + ], + "angle": 0, + "content": "8. Dockhorn, T., Cao, T., Vahdat, A., Kreis, K.: Differentially private diffusion models. arXiv preprint arXiv:2210.09929 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.786, + 0.687 + ], + "angle": 0, + "content": "9. Dwork, C.: Differential privacy. In: International colloquium on automata, languages, and programming. pp. 1-12. Springer (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.786, + 0.729 + ], + "angle": 0, + "content": "0. Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5111-5120 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.786, + 0.77 + ], + "angle": 0, + "content": "1. Ghalebikesabi, S., Berrada, L., Gowal, S., Ktena, I., Stanforth, R., Hayes, J., De, S., Smith, S.L., Wiles, O., Balle, B.: Differentially private diffusion models generate useful synthetic images. arXiv preprint arXiv:2302.13861 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.771, + 0.786, + 0.812 + ], + "angle": 0, + "content": "12. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "13. Gu, X., Du, C., Pang, T., Li, C., Lin, M., Wang, Y.: On memorization in diffusion models. arXiv preprint arXiv:2310.02664 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.368, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "X. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "15. Hsu, T.M.H., Qi, H., Brown, M.: Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.206, + 0.785, + 0.233 + ], + "angle": 0, + "content": "16. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.234, + 0.785, + 0.275 + ], + "angle": 0, + "content": "17. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22691-22702 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.277, + 0.785, + 0.317 + ], + "angle": 0, + "content": "18. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.319, + 0.785, + 0.36 + ], + "angle": 0, + "content": "19. Ni, Z., Wei, L., Li, J., Tang, S., Zhuang, Y., Tian, Q.: Degeneration-tuning: Using scrambled grid shield unwanted concepts from stable diffusion. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 8900-8909 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.361, + 0.785, + 0.402 + ], + "angle": 0, + "content": "20. Pizzi, E., Roy, S.D., Ravindra, S.N., Goyal, P., Douze, M.: A self-supervised descriptor for image copy detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14532-14542 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.403, + 0.785, + 0.431 + ], + "angle": 0, + "content": "21. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.432, + 0.785, + 0.459 + ], + "angle": 0, + "content": "22. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.46, + 0.785, + 0.501 + ], + "angle": 0, + "content": "23. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.502, + 0.785, + 0.543 + ], + "angle": 0, + "content": "24. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.545, + 0.785, + 0.587 + ], + "angle": 0, + "content": "25. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.588, + 0.785, + 0.642 + ], + "angle": 0, + "content": "26. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Diffusion art or digital forgery? investigating data replication in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6048-6058 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.643, + 0.785, + 0.685 + ], + "angle": 0, + "content": "27. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Understanding and mitigating copying in diffusion models. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.686, + 0.785, + 0.713 + ], + "angle": 0, + "content": "28. Webster, R.: A reproducible extraction of training images from diffusion models. arXiv preprint arXiv:2305.08694 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.714, + 0.785, + 0.755 + ], + "angle": 0, + "content": "29. Webster, R., Rabin, J., Simon, L., Jurie, F.: This person (probably) exists. identity membership attacks against gan generated faces. arXiv preprint arXiv:2107.06018 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "30. Wen, Y., Jain, N., Kirchenbauer, J., Goldblum, M., Geiping, J., Goldstein, T.: Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "31. Wen, Y., Liu, Y., Chen, C., Lyu, L.: Detecting, explaining, and mitigating memorization in diffusion models. In: The Twelfth International Conference on Learning Representations (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.356, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "IET-AGC: Mitigating Memorization in Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "32. Yoon, T., Choi, J.Y., Kwon, S., Ryu, E.K.: Diffusion probabilistic models generalize when they fail to memorize. In: ICML 2023 Workshop on Structured Probabilistic Inference \\(\\{\\backslash\\&\\}\\) Generative Modeling (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.786, + 0.218 + ], + "angle": 0, + "content": "33. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.218 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models/d774ff0f-2daf-4eeb-815c-c6da2eb3ab17_origin.pdf b/2024/Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models/d774ff0f-2daf-4eeb-815c-c6da2eb3ab17_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e519c5f1666e8fd4aab01b880577c1ca18b498b3 --- /dev/null +++ b/2024/Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models/d774ff0f-2daf-4eeb-815c-c6da2eb3ab17_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd3d27c4fbe22492f71f79d0e730666ed3e3ab8a0f239ace8ed38a2195381972 +size 1715818 diff --git a/2024/Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models/full.md b/2024/Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a22d4738757f384897b0e70d8511e23ee97d3115 --- /dev/null +++ b/2024/Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models/full.md @@ -0,0 +1,306 @@ +# Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models + +Xiao Liu $^{1*}$ , Xiaoliu Guan $^{1*}$ , Yu Wu $^{1\dagger}$ , and Jiaxu Miao $^{2\dagger}$ + +$^{1}$ School of Computer Science, Wuhan University + +{xiaoliu, liuxiaoguan, wuyucs}@whu.edu.cn + +$^{2}$ School of Cyber Science and Technology, Sun Yat-sen University + +miaojx@mail.sysu.edu.cn + +Abstract. Diffusion models, known for their tremendous ability to generate novel and high-quality samples, have recently raised concerns due to their data memorization behavior, which poses privacy risks. Recent approaches for memory mitigation either only focused on the text modality problem in cross-modal generation tasks or utilized data augmentation strategies. In this paper, we propose a novel training framework for diffusion models from the perspective of visual modality, which is more generic and fundamental for mitigating memorization. To facilitate "forgetting" of stored information in diffusion model parameters, we propose an iterative ensemble training strategy by splitting the data into multiple shards for training multiple models and intermittently aggregating these model parameters. Moreover, practical analysis of losses illustrates that the training loss for easily memorable images tends to be obviously lower. Thus, we propose an anti-gradient control method to exclude the sample with a lower loss value from the current mini-batch to avoid memorizing. Extensive experiments and analysis on four datasets are conducted to illustrate the effectiveness of our method, and results show that our method successfully reduces memory capacity while even improving the performance slightly. Moreover, to save the computing cost, we successfully apply our method to fine-tune the well-trained diffusion models by limited epochs, demonstrating the applicability of our method. Code is available in https://github.com/liuxiao-guan/IET_AGC. + +Keywords: Diffusion Models $\cdot$ Model Memorization $\cdot$ Data Privacy + +# 1 Introduction + +Recent advancements in diffusion models have significantly transformed the landscape of image generation. Modern diffusion models, such as Stable Diffusion [23], + +Midjourney [1], and SORA [2], have the ability to generate realistic images that are hard for humans to distinguish, demonstrating the unparalleled capabilities in producing diverse images. However, recent works [5, 27, 31] suggested that diffusion models are capable of memorizing images from the training set and reproducing them, which has the potential risk of privacy leakage. To address the severe problem, some works [10, 17, 19, 33] proposed to make diffusion models "forget" specific concepts such as a portrait of a certain celebrity, or the style of a particular artist. However, these works can only blacklist specific content that users want to conceal, but cannot completely cover the privacy-sensitive information that the model might remember, which still has a risk of privacy leakage. + +Recently, some works [7,26,27,31] have proposed to mitigate diffusion memorization without specific content limitations, thus reducing the risk of diffusion models leaking privacy-sensitive training data. Most of them focused on tackling the training data memorization in text-to-image diffusion models, and proposed data augmentation for captions to reduce model memorization since the insufficient diversity in captions easily leads to training data generation. For instance, Somepalli et al. [27] utilized random caption replacement, random token replacement, caption word repetition, etc., to reduce memorization. Another previous approach Daras et al. [7] proposed to use corrupted images for diffusion model training for memorization reduction. Wen et al. [31] introduced methods for detecting memorized prompts through text-conditional predictions and proposed two strategies to mitigate memorization: minimizing during inference or filtering during training. + +Although these works represented an important step forward in understanding the memorization issue in diffusion models, they either utilized simple data augmentation strategies or only focused on the easily memorable images that are related to specific captions in cross-modal generation tasks. However, diffusion models have been proven capable of generating images from memory without the need for text guidance [4], and previous methods focused on a limited scope of the memorization issue in diffusion models. Differently, in this paper, we propose a novel training framework for diffusion models from the perspective of visual modality, i.e., IET-AGC (Iterative Ensemble Training with Anti-Gradient Control), which not only reduces memorization fundamentally but also provides a more generic approach for both unconditional and text-conditional diffusion models. + +First, we propose an iterative ensemble training (IET) framework to mitigate memorization by parameter aggregation. Training data are stored in parameters of diffusion models due to over-optimization, and the model ensemble strategy aggregates parameters to re-organize model knowledge, which contributes to alleviating the memorization of training data. Thus, we separate the training data into several groups, train diffusion models individually, and ensemble them for memorization reduction. However, aggregating models trained on subsets of data may potentially decrease the performance due to the divergent optimizations on different subsets of these models. Motivated by Federated Learning tech + +niques [18], we iteratively ensemble the models during training, which reduces memorization by multiple aggregation and maintains generation performance. + +Based on our IET training framework, we further introduce an Anti-Gradient Control module to reduce memorization of training data ulteriorly. We conduct practical analysis on losses during training and find that the training loss for easily memorable images tends to be obviously lower than that for less memorable images. Thus, we propose to exclude the sample with a relatively small loss value from the current mini-batch to avoid memorizing these samples. During the training process, as the diffusion model exhibits varying average loss values across different time steps, we maintain a memory bank to store the average loss of each time step. Subsequently, we discard samples with losses below a certain proportion of the average loss. Since the model has encountered the sample during training, excluding it is unlikely to have a significant impact on the model's performance. + +Extensive experiments on four datasets show the significance of our IET-AGC framework. Our method significantly reduces the memorized quantity by $87.3\%$ , $66.4\%$ , and $85.3\%$ compared with the default training (DDPM [14]) on CIFAR-10 and CIFAR-100 and AFHQ-DOG, respectively. In addition, considering the high cost of re-training existing well-trained diffusion models, we also propose an efficient way to address the memorization issue of existing pre-trained models by simply fine-tuning using our method for several epochs. Experiments show that using our method to fine-tune for just two epochs can significantly reduce the memorization by $25.22\%$ in unconditional diffusion models on CIFAR-10. Furthermore, when fine-tuning the text-conditional diffusion model, Stable Diffusion, our approach decreases the memorization score by $42.18\%$ compared to conventional fine-tuning methods. These results demonstrate that our method performs excellently on both unconditional and text-conditional diffusion models. + +# 2 Related Work + +Memorization in Generative Models. Several studies have examined the memorization capabilities of the generative model. Generative Adversarial Networks (GANs) [12] have been at the forefront of this research area. As Webster et al. [29] demonstrated when applied to face datasets, GANs can occasionally replicate. Prior study [4] explored an adversarial attack on language models like GPT-2 [21], where individual training examples can be recovered, including personally identifiable information and unique text sequences. + +Recent studies have shifted their attention toward diffusion models. Somepalli et al. [26] found that diffusion models accurately recall and replicate training images, especially noted with models like the Stable Diffusion model [23]. Building upon this discovery, Carlini et al. [5] developed a tailored black-box attack for diffusion models. They generated images and implemented a membership inference attack to assess density. Webster et al. [28] demonstrated a more efficient extraction attack with fewer network evaluations, identified "template + +verbatim," and discussed its persistence in newer systems. Recent research has shifted towards exploring the theoretical aspects of memory in diffusion models. [32] discovered that generalization and memorization are mutually exclusive occurrences and further demonstrated that the dichotomy between memorization and generalization can be apparent at the class level. Gu et al. [13] extensively studied how factors like data dimension, model size, time embedding, and class conditions affect the memory capacity of the diffusion model. + +Memorization Mitigation. The mitigation measures have primarily been concerned with filtering inputs and deduplication. For example, Stable Diffusion employed well-trained detectors to identify unsuitable generated content. However, these temporary solutions can be easily bypassed [22, 30] and do not effectively prevent or lessen copying behavior on a broad scale. Kumari et al. [17] designed an algorithm to align the image distribution with a specific style, instance, or text prompt they aim to remove, to the distribution related to a core concept. This stopped the model from producing target concepts based on its text condition. However, these approaches are inefficient because they necessitate a list of all concepts to be erased, and have not addressed the key issue of how to reduce the memory capacity of the model. [8, 11] explored the use of differential privacy (DP) [9] to train diffusion models or fine-tune ImageNet pre-trained models. However, their focus was on ensuring the privacy of the training of diffusion models, not on the privacy of the images generated by the diffusion models. Daras et al. [7] introduced a technique for training diffusion models utilizing tainted data. By incorporating additional corruption before applying noise, their methodology prevents the model from overfitting to the training data. But their training requires a considerable amount of time. [27, 31] also suggested a series of recommendations to mitigate copying such as randomly replacing the caption of an image with a random sequence of words, but most of which are limited to text-to-image models. Our work focuses on the nature of memorization in diffusion models, especially for unconditional ones. + +# 3 Method + +In this section, we present our methodology aimed at mitigating the memorization in diffusion models, without sacrificing excessive image quality. + +# 3.1 Iterative Ensemble Training + +Training data are stored in parameters of diffusion models due to over-optimization, and the model ensemble strategy aggregates parameters to re-organize model knowledge. Thus, to mitigate the memorization, we propose a method that trains multiple diffusion models on different data shards of a dataset, merges them after a certain period, and then repeats these two stages iteratively. + +![](images/25cdf156e98300101500475d8a4f9ee9ca31a8a7ed59a3cac40599a0011aa9cd.jpg) +(a) + +![](images/5b7a0371dc4605e1eb438f22214df4eada10b36a36ffe6fd5e14289c72122c27.jpg) +(b) +Fig. 1: Overview of our IET-AGC method. (a) Iterative Ensemble Training (IET): we divide the dataset $D$ into $K$ different data shards. Each shard $D_{i}$ trains a separate diffusion model $\theta_{i}$ . After a period of training, these models are merged by averaging and repeating this training strategy. (b) Anti-Gradient Control (AGC): during the training process, we dynamically update and maintain a memory bank of losses at each time step. For loss values smaller than $\lambda$ times the corresponding memory bank, we exclude these losses to prevent the model from memorizing such images. + +Training on Different Data Shards. Unlike the training methods of previous diffusion models, which train a single model on the entire dataset once. In this paper, we divide a dataset into an equal number of shards and then train the corresponding diffusion models on each separate part. If the dataset contains class information, we divide it equally along the class dimension, so that each data shard contains all classes, with the same number of instances for each class. + +Specifically, suppose the dataset $D$ contains $C$ classes, with each class having $N_{c}$ samples, and the total number of samples is $N = \sum_{c=1}^{C} N_{c}$ . We divide the dataset into $K$ equal parts, with each part containing $\frac{N}{K}$ samples. So the $i$ th part of the dataset $D_{i}$ is represented as + +$$ +D _ {i} = \bigcup_ {c = 1} ^ {C} \left\{\left(x _ {c, i, k}, y _ {c, i, k}\right) \mid k = 1, 2, \dots , \frac {N _ {c}}{K} \right\}, \tag {1} +$$ + +where $(x,y)$ represents the sample and its corresponding label. Notably, $C = 1$ indicates that the dataset does not contain class information. Then, each part $i$ trains a separate diffusion model $\theta_{i}$ scaled by the learning rate $\eta$ on its own dataset, + +$$ +\theta_ {i} \leftarrow \theta_ {i} - \eta \nabla \mathcal {L} \left(\theta_ {i}\right). \tag {2} +$$ + +Merging the Multiple Diffusion Models. After training for some time, each part obtains a different diffusion model. We simply average the weights of all + +models $\theta_{i}$ to obtain a global model $\widehat{\theta}$ as + +$$ +\widehat {\theta} \leftarrow \frac {1}{K} \sum_ {i = 1} ^ {K} \theta_ {i}. \tag {3} +$$ + +Then, we repeat the two stages of training on separate shards of the data and merging models, where we use the obtained global model as the initial model for the first stage. As each shard contains only $\frac{1}{K}$ of the total data, the training time for each model is proportionally reduced, maintaining the overall computational cost nearly constant compared to training a single model on the entire dataset. The only additional computational cost comes from the periodic merging of models, which is minimal. + +# 3.2 Loss Analysis + +To further reduce memorization of training data, we delve deeper into the causes of memorization phenomena, specifically analyzing it through the lens of the loss. We begin by establishing the fundamental notation linked with diffusion models. Diffusion models [14] originate from the non-equilibrium statistical physics [25]. They are essentially straightforward: they operate as image denoisers. During the training process, when given a clean image $x$ , time-step $t$ is sampled from the interval $[0, T]$ , along with a Gaussian noise vector $\epsilon \sim N(0, I)$ , resulting in a noised image $x_{t}$ : + +$$ +x _ {t} = \sqrt {\alpha_ {t}} x + \sqrt {1 - \alpha_ {t}} \epsilon , \tag {4} +$$ + +where the scheduled variance $\alpha_{t}$ varies between 0 and 1, with $\alpha_0 = 1$ and $\alpha_{T} = 0$ . The diffusion model then removes the noise to reconstruct the original image $x$ by predicting the noise introduced, achieved through stochastic minimization of the objective function $\frac{1}{N}\sum_{i}\mathbb{E}_{t,\epsilon}\mathcal{L}(x_i,t,\epsilon ;\theta)$ , where + +$$ +\mathcal {L} \left(x _ {i}, t, \epsilon ; \theta\right) = \| \epsilon - \epsilon_ {\theta} \left(\sqrt {\alpha_ {t}} x _ {i} + \sqrt {1 - \alpha_ {t}} \epsilon , t\right) \| _ {2} ^ {2}. \tag {5} +$$ + +To analyze the correlation between losses and image memorization, we identify images memorized on CIFAR-10 and calculate their loss functions at each time step. Similarly, we sample 256 non-memorized images from the remaining training data and compute their losses at each time step. Fig. 2 shows the comparisons of the losses when the time step is in the interval [0, 600] ( $T = 1000$ ). Memorized images exhibit significantly smaller loss values during this period, indicating that the model tends to reconstruct noise into such images. + +# 3.3 Anti-Gradient Control + +In this subsection, we elaborate on how to utilize the aforementioned loss analysis to devise a training strategy aimed at alleviating the occurrence of memorization. + +![](images/9e2da4cfb6dde66cf0dfa6474105892f115d5e2d517b1564f48aa035190e7f61.jpg) +Fig. 2: Comparison of the losses between memorized and non-memorized images. The solid line represents the averaged losses of memorized and non-memorized images, while the two dashed lines represent the losses of the 15th and 85th percentile data respectively. + +Memory Bank. In order to identify images with exceptionally low loss values that are prone to memorization during training, we need to maintain the average losses for each time step. However, computing the average loss at each time step entails substantial computational expenses, as it necessitates evaluating the losses for all images using the model at each time step. Thus, we propose a memory bank to store and update losses during mini-batch training without increasing the time cost. However, the losses generally decrease with the training step growing. Therefore, when calculating the average loss in the memory bank, instead of directly averaging all losses at this time step, the aggregating weights should be higher for losses closer to the current update. + +Specifically, we initialize an array of length $T$ with zeros, termed the memory bank. The initialization values of the memory bank seem insignificant. After calculating the loss for a mini-batch, we update the memory bank using the Exponential Moving Average (EMA) method based on the loss and the sampled time step, thereby better reflecting the current state of the model: + +$$ +l _ {t} \leftarrow \gamma \cdot l _ {t} + (1 - \gamma) \cdot \mathcal {L} (x, t, \epsilon ; \theta), \tag {6} +$$ + +where $\gamma$ represents the smoothing factor, and $l_{t}$ represents the averaged loss in the memory bank at time step $t$ . + +Anti-Gradient Control (AGC). In previous observations, if the model exhibits memorization of a certain sample, the loss value of the model on that sample tends to be abnormally small. Thus, we use the ratio of the training loss of a certain sample to the mean loss in the memory bank at the time step $t$ as + +a measure to detect memorization: + +$$ +r = \frac {\mathcal {L} (x , t , \epsilon ; \theta)}{l _ {t}}. \tag {7} +$$ + +A smaller ratio of $r$ indicates a higher likelihood of the image being memorized. To facilitate this determination, we can establish a configurable threshold denoted as $\lambda$ . If the loss ratio $r$ falls below this threshold $\lambda$ , the image is classified as memorized. At this point, we generate a mask that sets the loss value corresponding to this image to zero, i.e. skipping this image in the mini-batch, as shown in the following function, + +$$ +\mathcal {L} (x, t, \epsilon ; \theta) = \left\{ \begin{array}{l l} 0 & \text {i f} r < \lambda \\ \mathcal {L} (x, t, \epsilon ; \theta) & \text {o t h e r w i s e .} \end{array} \right. \tag {8} +$$ + +Since the model has encountered the sample during training, excluding it is unlikely to have a significant impact on the model's performance. + +# 4 Experiments + +# 4.1 Experimental Setup + +Datasets. We evaluate our method on CIFAR-10 [16], CIFAR-100 [16], AFHQDOG [6] for unconditional generation, and LAION-10k [27] for text-conditioned generation. CIFAR-10 and CIFAR-100 consist of 50,000 32x32 color images, divided into 10 and 100 classes, respectively. AFHQ-DOG is a subset of the AFHQ dataset with approximately 5,000 512x512 dog images, resized to 64x64 for our experiments. LAION-10k is a subset of LAION [24], comprising 10,000 image-text pairs with each image having a resolution of 256x256 pixels. + +Implementation Details of Training. We conduct experiments on training unconditional diffusion models from scratch using the CIFAR-10, CIFAR-100, and AFHQ-DOG datasets. The IET framework divides CIFAR datasets into 10 shards and AFHQ-DOG into 5 shards. Threshold $\lambda$ is set to 0.5 for CIFAR datasets and 0.714 for AFHQ-DOG. Additionally, we perform experiments on fine-tuning pre-trained diffusion models using the CIFAR-10 dataset, maintaining the same hyperparameters as those used for training from scratch. To demonstrate the effectiveness of our method in text-conditioned diffusion models, we fine-tune Stable Diffusion on LAION-10k. The IET framework divides the LAION-10k dataset into 8 shards, with the threshold $\lambda$ set to 0.8. The smoothing factor $\gamma$ is 0.8 for all datasets. Further details are in the supplementary material. + +Extracting Memorized Image. We adopt Carlini's detection rule [5] for unconditional generation, considering $\bar{x}$ as memorized if the $\ell_2$ distance to its nearest neighbor is significantly lower compared to the $n$ closest neighbors. We modify this rule to: + +$$ +\ell_ {2} (\bar {x}, x; \mathbb {S} _ {\bar {x}} ^ {n}) = \frac {\ell_ {2} (\bar {x} , x)}{\mathbb {E} _ {y \in \mathbb {S} _ {\bar {x}} ^ {n}} [ \ell_ {2} (\bar {x} , y) ]}, \tag {9} +$$ + +where $n = 50$ in our experiment. A binary classifier is defined as: + +$$ +I s M e m o (\bar {x}, x; \mathbb {S} _ {\bar {x}} ^ {n}, \delta_ {V}) = 1 _ {\ell_ {2} (\bar {x}, x; \mathbb {S} _ {\bar {x}} ^ {n}) \leq \delta_ {V}}. \tag {10} +$$ + +The more images below $\delta_V$ , the stronger the model's memorization. We generate 65,536 images per model, calculate their $\ell_2$ distances, and count images below thresholds $\delta_V$ of 0.4, 0.5, and 0.6 to quantitatively evaluate the model's memorization, denoted as $\mathrm{MQ}_{0.4}$ , $\mathrm{MQ}_{0.5}$ and $\mathrm{MQ}_{0.6}$ . + +We adopt Somepalli's detection rule [27] for text-conditioned generation, quantify memorization using a similarity score derived from the dot product of SSCD features [20] of $\bar{x}$ and the nearest neighbor $n_0$ : + +$$ +\zeta = E (\bar {x}) ^ {T} \cdot E (n _ {0}) \tag {11} +$$ + +where $E(\cdot)$ is SSCD [20]. The dataset similarity score is then defined as the 95th percentile of this distribution. + +# 4.2 Experimental Results + +Training from Scratch. The experimental results of our method and four competitive methods are shown in Tab. 1. "Default" denotes the conventional training approach of DDPM [14]. "DP-SGD" denotes the method of Differentially Private Stochastic Gradient Descent [3], which involves clipping and adding noise to the model's gradients to protect privacy, albeit at the cost of some image quality. Carlina et al. [5] found that utilizing DP-SGD could lead to consistent model divergence. To make DP training more stable, we modify the amplitude of the added noise to be the product of the gradient norm and the noise multiplier: + +$$ +\sigma = \left| \left| \nabla \mathcal {L} \right| \right| \times \tau , \tag {12} +$$ + +where $\tau$ represents the noise multiplier. "Adding noise" denotes a method of directly adding Gaussian noise to the images during training, with a mean of 0 and a variance of 0.1. "Ambient Diffusion" [7] protects privacy by training generative models on highly corrupted samples, preventing the model from observing clean training data. + +Results in Tab. 1 show that adding noise to the training images or gradients reduces the quality of the generated images. However, it still does not resolve the issue of training image memorization. Despite Ambient Diffusion also reducing memorization, it leads to a significant increase in FID (from 8.81 to 11.7), indicating a notable degradation of image quality. Compared with the default training approach, our method maintains or even slightly improves the generative quality by reducing the FID score. At the same time, our method significantly reduces the diffusion model's memorization of the training data. As shown in Tab. 1, for the $\mathrm{MQ}_{0.4}$ score, the number of memorized images reduced by $87.3\%$ , $66.4\%$ , and $85.3\%$ compared with the default training on CIFAR-10, CIFAR-100 and AFHQ-DOG, respectively, illustrating the effectiveness of our method. + +Visualization. To provide a more intuitive confirmation of our training method, we visualize the images generated by the model along with their closest + +Table 1: Comparisons of unconditional generation on three datasets in terms of memorized quantity denoted as MQ. We also report the FID to evaluate the quality of images produced by the model. + +
MethodCIFAR-10CIFAR-100AFHQ-DOG
\(MQ_{0.4}\)\(MQ_{0.5}\)\(MQ_{0.6}\)↓FID↓\(MQ_{0.4}\)\(MQ_{0.5}\)\(MQ_{0.6}\)↓FID↓\(MQ_{0.4}\)\(MQ_{0.5}\)\(MQ_{0.6}\)↓FID↓
Default11146520308.81429172756209.2912344190533079523.59
Adding Noise197593209194.611791037438386.1811700192952722461.18
DP-SGD [3]148728320012.55--------
Ambient Diffusion [7]2213885111.7--------
IET-AGC141178398.3414476032748.51181154351523722.2
+ +![](images/b36517505ffc6e2d207c7f6c6dccfabd05f59f46a1a0f07d04fe5438e9b0b6b0.jpg) +(a) Default Training + +![](images/d9601d3604a8b0b0d4e62ef6cd6321e245ecafb738a547504cbfe0f84ce83d3e.jpg) +(b) Our Method +Fig. 3: The similar grid image of the default training and our method. Odd-numbered columns represent images from the training set, while even-numbered columns represent images from the generation set that has the smallest $\ell_2$ distance to the corresponding training set image. The images in the grid are arranged in ascending order of $\ell_2$ distance, and the selected images from both methods correspond to the same position in terms of $\ell_2$ distance. + +counterparts in the training dataset in Fig. 3. There is a similarity between the images from the training set and their corresponding images in the generation set, while our method shows lower similarity compared to the default training method. It is also important to note that our model's FID slightly decreases on all three datasets compared to the default indicating that our method also improves the quality of the generated images. + +Finetuning Unconditional DDPMs. The results are presented in Tab. 2. The IET method shows improvement over the pre-trained in terms of $\mathrm{MQ}_{0.4}$ and FID, indicating more efficient data forgetting and higher image quality. The AGC method rapidly causes the model to forget the memorized data. However, it also led to a substantial increase in FID. The IET-AGC method not only reduces the FID but also effectively lowers the model's memory rate. + +Table 2: Fine-tuning results of the pre-trained DDPM on CIFAR-10 dataset for 2 epochs. + +
MethodCIFAR-10
\(MQ_{0.4}\)\(MQ_{0.5}\)\(MQ_{0.6}\)↓FID↓
Default11146520308.81
IET5139623678.33
AGC44235131711.8
IET-AGC8340817967.93
+ +Table 3: Fine-tuning results of Stable Diffusion model on LAION-10k dataset. + +
MethodSim Score↓Clip Score↑FID↓
Default SD0.6430.518.7
Train TimeMC0.4230.316.6
RC0.5730.616.0
Mitigation [27]CWR0.6130.816.7
Test TimeRT0.5229.518.7
CWR0.5830.118.1
Mitigation [27]GNI0.6230.318.9
Our MethodIET0.4131.316.5
AGC0.5330.618.5
IET-AGC0.3731.316.7
+ +![](images/e547a1c59dc7c57b8a707d3648a523831aa9a1c75b7f5e1e4e3f5fbe4c254fa9.jpg) +(a) Distribution of distances to the most similar images in the dataset. + +![](images/315847b1019744e493e278031641b66fbaa1eb26e90a5d000529215d55213d10.jpg) +(b) Energy distribution. The greater the energy, the more complex the image. +Fig. 4: Most skipped images vs least skipped images. + +Finetuning Text-conditional Stable Diffusion. The results are presented in Tab. 3. Somepalli et al. [27] protects privacy by randomizing conditional information during training and inference, thereby reducing the likelihood of the model replicating specific training data. Our method IET-AGC achieves the best overall results with a Similarity Score of 0.37, maintaining the highest Clip Score of 31.3 and a competitive FID of 16.7. This indicates our approach effectively balances memorization and generative quality, outperforming the default and other mitigation methods. + +# 4.3 Analysis of Skipping + +In this section, we conduct comparative experiments on the AFHQ-DOG dataset to delve into which types of images are prone to be skipped, as well as the relationship between memorizable images and those that are skipped. + +Frequency of Skipped Images. Throughout the training process, we record the identifiers of skipped images. As shown in Fig. 5, our method does not entail skipping all images. In our approach, $90\%$ of the images are skipped fewer than + +![](images/339a94c939876f4af05d39443c70935c383938ece7450c5025c9cd15d8b1e3ff.jpg) +Fig. 5: Distribution of skipped image counts. + +648 times (across a total of 2,278 training epochs), indicating that our method can effectively differentiate between different images. This suggests that we are not simply reducing memorization by constraining the model's learning. On the other hand, while our method requires skipping images with exceptionally low loss values, all images still contribute to the model's training. + +Images Most Easily Skipped. We believe these images are more easily skipped for two main reasons. Firstly, data aggregation. We computed the $\ell_2$ distance between these easily skipped images and all other images in the dataset, as well as between those not easily skipped images and all other images in the dataset. Fig. 4a indicates that the distribution of the skipped images is more clustered. Consistent with the findings of Carlini et al. [5], which suggested that removing duplicate training images effectively reduces memorization capacity, skipping these clustered images can also reduce memorization capacity. Secondly, data simplicity. We performed Fourier transforms on these easily skipped and not easily skipped images to obtain their energy distributions, as shown in Fig. 4b. The easily skipped images have less energy, indicating that they lack finer details. We believe both factors contribute to making these images easier for the model to memorize, thus skipping them has a positive impact on reducing the model's memorization capacity. + +# 4.4 Ablation Study + +Performance Comparisons of Each Component. To further understand the effectiveness of our approach, we conduct ablation experiments to investigate the individual impacts of different components on CIFAR-10. In addition, in our approach, the dataset is evenly distributed, meaning each data shard is set in + +an IID (independently and identically distributed) manner. To validate the reasonableness of this dataset configuration, we also establish an experiment where each data shard is set up in a non-independent and identically distributed (non-IID) manner. Similar to [15], we employ the Dirichlet distribution to generate data, thereby establishing such a setting. + +Results in Tab. 4 show that when the data shards are set in non-IID, Iterative Ensemble Training (IET) successfully reduces the memorization $\mathrm{(IET_{non - IID})}$ . However, the non-IID splitting reduces the quality of performance since the divergent optimization of different models and aggregating these models affects the performance. Differently, when the data shards are set in IID, $\mathrm{IET}_{\mathrm{IID}}$ can reduce both the memorization and FID score. Thus in this paper, we choose to use the IID splitting strategy. Besides, Tab. 4 shows that both AntiGradient Control (AGC) and Iterative En + +semble Training (IET) reduce memorization effectively and AGC is more effective than the IET on the memorization reduction. Compared to the conventional method, AGC reduces $\mathrm{MQ}_{0.5}$ by approximately $66\%$ , while $\mathrm{IET}_{\mathrm{IID}}$ reduces it by around $18\%$ . However, using AGC only shows a slight growth in FID, which affects the quality of the generated images. Interestingly, combining both IET and AGC leads to better FID and memory reduction effects. + +Table 4: Performance comparisons of each component. + +
MethodCIFAR-10
\( MQ_{0.4} \)\( MQ_{0.5} \)\( MQ_{0.6} \)↓FID↓
Default11146520308.81
\( IET_{IID} \)8938217837.61
\( IET_{non-IID} \)69315143912.5
AGC2615497611.36
IET-AGC141178398.34
+ +Exploring Parameters Impact on Experimental Results. In this study, we examine how various parameters affect our experimental outcomes. By systematically varying these parameters, we aim to understand how they influence our results and to identify the optimal settings for our experiments. Specifically, we conduct a series of experiments where we change the number of shards, training epochs per aggregation period, skipping threshold $\lambda$ , and smoothing factor $\gamma$ of the memory bank. For each variation, we measure the impact on MQ and FID. The default parameters are set with the number of shards $K$ as 10, epochs per shard $E$ as 100, skipping threshold $\lambda$ as 0.5, and smoothing factor $\gamma$ as 0.8. Results are reported in Tab. 5. + +Number of Shards $K$ . We investigate the impact of the number of data shards on model performance by setting it to 1, 5, 10, and 15. $K = 1$ means the default training strategy of diffusion models. Results show that the MQ scores of $K = 5, 10, 15$ are all lower than $K = 1$ , indicating that using our IET method can effectively reduce memorization. When $K = 10$ , the MQ score achieves the best performance. However, with the growth of $K$ , the FID score increases. Thus we need to balance the FID and MQ to choose the proper number of shards. + +Training Epochs per Aggregation Period $E$ . We set the total number of training epochs unchanged and conduct experiments by varying the number of epochs per model aggregation period, i.e., the aggregating frequency of IET. + +Results show that both high and low frequencies of aggregation will reduce the performance of the memorization and image quality. Thus, we choose $E = 100$ to achieve the best performance of MQ. + +# Skipping Threshold $\lambda$ . + +We evaluate the importance of $\lambda$ in mitigating the memorization effect by setting the values of $\lambda$ to 0.4, 0.5, 0.66 and 0.8. A large threshold means skipping more training samples that are easily memorized. Results in Tab. 5 show as $\lambda$ grows, more memorable training samples are skipped and the memorization phenomenon is further reduced. However, skipping more samples will reduce the model performance, i.e., the generation quality. + +Table 5: Parameters impact on experimental results. + +
ParametersCIFAR-10
\(MQ_{0.4}\)\(MQ_{0.5}\)\(MQ_{0.6} \downarrow\)FID↓
Number of Shards K111146520308.81
5201297986.37
10141178398.34
152116911439.1
Epoches per Aggregation E102616710189.31
501619411207.76
100141178398.34
2001719313079.01
25024181111410.36
Skipping Threshold λ0.44731015737.72
0.5141178398.34
0.663696088.25
0.834550310.43
Smoothing Factor γ0.52719211938.99
0.8141178398.34
0.9201328868.94
+ +# Smoothing Factor $\gamma$ + +The selection of the smooth- + +ing factor $\gamma$ is also crucial. When we set $\gamma$ to 0.9, the update of the memory bank becomes overly sluggish, failing to faithfully reflect the current model's loss across different time steps. Conversely, if we set $\gamma$ to 0.5, the memory bank becomes excessively sensitive, which can lead to instability and fragility during the training process. + +# 5 Limitations and Future Work + +Although our strategy has proven effective in reducing the model's memorization capacity, the selection of the threshold $\lambda$ , which balances between the model's memorization capacity and the quality of the generated images, can only rely on experiments. In future research, we plan to design a quick analysis method for datasets to help determine the appropriate choice of threshold. + +# 6 Conclusion + +This paper innovatively proposes a training strategy for diffusion models, which train models on multiple data shards and ignore data with abnormal loss values. This strategy effectively reduces the memory capacity of the model, further strengthening data privacy protection without compromising the quality of the generated images. We firmly believe that this training strategy has a broad application prospect and great development potential in the field of data privacy protection. + +# Acknowledgments + +This work was supported by National Natural Science Foundation of China (Grant No. 62306273), National Natural Science Foundation of China (Grant No. 62372341), and the Fundamental Research Funds for the Central Universities (Grant No. 2042024kf0040). The numerical calculations in this paper have been done on the supercomputing system in the Supercomputing Center of Wuhan University. + +# References + +1. Midjourney team (2022), https://www.midjourney.com/home +2. Sora team (2024), https://openai.com/sora +3. Abadi, M., Chu, A., Goodfellow, I., McMahan, H.B., Mironov, I., Talwar, K., Zhang, L.: Deep learning with differential privacy. In: Proceedings of the 2016 ACM SIGSAC conference on computer and communications security. pp. 308-318 (2016) +4. Carlini, N., Tramer, F., Wallace, E., Jagielski, M., Herbert-Voss, A., Lee, K., Roberts, A., Brown, T., Song, D., Erlingsson, U., et al.: Extracting training data from large language models. In: 30th USENIX Security Symposium (USENIX Security 21). pp. 2633-2650 (2021) +5. Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramer, F., Balle, B., Ippolito, D., Wallace, E.: Extracting training data from diffusion models. In: 32nd USENIX Security Symposium (USENIX Security 23). pp. 5253-5270 (2023) +6. Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8188-8197 (2020) +7. Daras, G., Shah, K., Dagan, Y., Gollakota, A., Dimakis, A., Klivans, A.: Ambient diffusion: Learning clean distributions from corrupted data. Advances in Neural Information Processing Systems 36 (2024) +8. Dockhorn, T., Cao, T., Vahdat, A., Kreis, K.: Differentially private diffusion models. arXiv preprint arXiv:2210.09929 (2022) +9. Dwork, C.: Differential privacy. In: International colloquium on automata, languages, and programming. pp. 1-12. Springer (2006) +0. Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5111-5120 (2024) +1. Ghalebikesabi, S., Berrada, L., Gowal, S., Ktena, I., Stanforth, R., Hayes, J., De, S., Smith, S.L., Wiles, O., Balle, B.: Differentially private diffusion models generate useful synthetic images. arXiv preprint arXiv:2302.13861 (2023) +12. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020) +13. Gu, X., Du, C., Pang, T., Li, C., Lin, M., Wang, Y.: On memorization in diffusion models. arXiv preprint arXiv:2310.02664 (2023) + +14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) +15. Hsu, T.M.H., Qi, H., Brown, M.: Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335 (2019) +16. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) +17. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22691-22702 (2023) +18. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017) +19. Ni, Z., Wei, L., Li, J., Tang, S., Zhuang, Y., Tian, Q.: Degeneration-tuning: Using scrambled grid shield unwanted concepts from stable diffusion. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 8900-8909 (2023) +20. Pizzi, E., Roy, S.D., Ravindra, S.N., Goyal, P., Douze, M.: A self-supervised descriptor for image copy detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14532-14542 (2022) +21. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019) +22. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022) +23. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +24. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021) +25. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015) +26. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Diffusion art or digital forgery? investigating data replication in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6048-6058 (2023) +27. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Understanding and mitigating copying in diffusion models. Advances in Neural Information Processing Systems 36 (2024) +28. Webster, R.: A reproducible extraction of training images from diffusion models. arXiv preprint arXiv:2305.08694 (2023) +29. Webster, R., Rabin, J., Simon, L., Jurie, F.: This person (probably) exists. identity membership attacks against gan generated faces. arXiv preprint arXiv:2107.06018 (2021) +30. Wen, Y., Jain, N., Kirchenbauer, J., Goldblum, M., Geiping, J., Goldstein, T.: Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. Advances in Neural Information Processing Systems 36 (2024) +31. Wen, Y., Liu, Y., Chen, C., Lyu, L.: Detecting, explaining, and mitigating memorization in diffusion models. In: The Twelfth International Conference on Learning Representations (2023) + +32. Yoon, T., Choi, J.Y., Kwon, S., Ryu, E.K.: Diffusion probabilistic models generalize when they fail to memorize. In: ICML 2023 Workshop on Structured Probabilistic Inference $\{\backslash\&\}$ Generative Modeling (2023) +33. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591 (2023) \ No newline at end of file diff --git a/2024/Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models/images.zip b/2024/Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4eb4cfa46337849f5e3fa8edf2d65ab26e354ac1 --- /dev/null +++ b/2024/Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29134d7fdbbc63c5a92fd291d11354f0ca631dd69e727a09d95db2c02bed934f +size 383646 diff --git a/2024/Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models/layout.json b/2024/Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9fcf0ecb6acf5ddf32091559552de4b1ad5aece4 --- /dev/null +++ b/2024/Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models/layout.json @@ -0,0 +1,9056 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 111, + 478, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 111, + 478, + 163 + ], + "spans": [ + { + "bbox": [ + 135, + 111, + 478, + 163 + ], + "type": "text", + "content": "Iterative Ensemble Training with Anti-Gradient Control for Mitigating Memorization in Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 168, + 185, + 448, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 185, + 448, + 198 + ], + "spans": [ + { + "bbox": [ + 168, + 185, + 448, + 198 + ], + "type": "text", + "content": "Xiao Liu" + }, + { + "bbox": [ + 168, + 185, + 448, + 198 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 168, + 185, + 448, + 198 + ], + "type": "text", + "content": ", Xiaoliu Guan" + }, + { + "bbox": [ + 168, + 185, + 448, + 198 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 168, + 185, + 448, + 198 + ], + "type": "text", + "content": ", Yu Wu" + }, + { + "bbox": [ + 168, + 185, + 448, + 198 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 168, + 185, + 448, + 198 + ], + "type": "text", + "content": ", and Jiaxu Miao" + }, + { + "bbox": [ + 168, + 185, + 448, + 198 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 205, + 207, + 409, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 207, + 409, + 219 + ], + "spans": [ + { + "bbox": [ + 205, + 207, + 409, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 205, + 207, + 409, + 219 + ], + "type": "text", + "content": " School of Computer Science, Wuhan University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 214, + 220, + 399, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 220, + 399, + 230 + ], + "spans": [ + { + "bbox": [ + 214, + 220, + 399, + 230 + ], + "type": "text", + "content": "{xiaoliu, liuxiaoguan, wuyucs}@whu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 171, + 230, + 442, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 230, + 442, + 241 + ], + "spans": [ + { + "bbox": [ + 171, + 230, + 442, + 241 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 171, + 230, + 442, + 241 + ], + "type": "text", + "content": " School of Cyber Science and Technology, Sun Yat-sen University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 251, + 242, + 362, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 242, + 362, + 252 + ], + "spans": [ + { + "bbox": [ + 251, + 242, + 362, + 252 + ], + "type": "text", + "content": "miaojx@mail.sysu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 282, + 455, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 282, + 455, + 524 + ], + "spans": [ + { + "bbox": [ + 159, + 282, + 455, + 524 + ], + "type": "text", + "content": "Abstract. Diffusion models, known for their tremendous ability to generate novel and high-quality samples, have recently raised concerns due to their data memorization behavior, which poses privacy risks. Recent approaches for memory mitigation either only focused on the text modality problem in cross-modal generation tasks or utilized data augmentation strategies. In this paper, we propose a novel training framework for diffusion models from the perspective of visual modality, which is more generic and fundamental for mitigating memorization. To facilitate \"forgetting\" of stored information in diffusion model parameters, we propose an iterative ensemble training strategy by splitting the data into multiple shards for training multiple models and intermittently aggregating these model parameters. Moreover, practical analysis of losses illustrates that the training loss for easily memorable images tends to be obviously lower. Thus, we propose an anti-gradient control method to exclude the sample with a lower loss value from the current mini-batch to avoid memorizing. Extensive experiments and analysis on four datasets are conducted to illustrate the effectiveness of our method, and results show that our method successfully reduces memory capacity while even improving the performance slightly. Moreover, to save the computing cost, we successfully apply our method to fine-tune the well-trained diffusion models by limited epochs, demonstrating the applicability of our method. Code is available in https://github.com/liuxiao-guan/IET_AGC." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 534, + 440, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 534, + 440, + 546 + ], + "spans": [ + { + "bbox": [ + 160, + 534, + 440, + 546 + ], + "type": "text", + "content": "Keywords: Diffusion Models " + }, + { + "bbox": [ + 160, + 534, + 440, + 546 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 534, + 440, + 546 + ], + "type": "text", + "content": " Model Memorization " + }, + { + "bbox": [ + 160, + 534, + 440, + 546 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 534, + 440, + 546 + ], + "type": "text", + "content": " Data Privacy" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 585, + 230, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 585, + 230, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 585, + 230, + 597 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 610, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 610, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 610, + 482, + 635 + ], + "type": "text", + "content": "Recent advancements in diffusion models have significantly transformed the landscape of image generation. Modern diffusion models, such as Stable Diffusion [23]," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 643, + 223, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 223, + 653 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 223, + 653 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 136, + 653, + 235, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 653, + 235, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 653, + 235, + 665 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "Midjourney [1], and SORA [2], have the ability to generate realistic images that are hard for humans to distinguish, demonstrating the unparalleled capabilities in producing diverse images. However, recent works [5, 27, 31] suggested that diffusion models are capable of memorizing images from the training set and reproducing them, which has the potential risk of privacy leakage. To address the severe problem, some works [10, 17, 19, 33] proposed to make diffusion models \"forget\" specific concepts such as a portrait of a certain celebrity, or the style of a particular artist. However, these works can only blacklist specific content that users want to conceal, but cannot completely cover the privacy-sensitive information that the model might remember, which still has a risk of privacy leakage." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 251, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 407 + ], + "type": "text", + "content": "Recently, some works [7,26,27,31] have proposed to mitigate diffusion memorization without specific content limitations, thus reducing the risk of diffusion models leaking privacy-sensitive training data. Most of them focused on tackling the training data memorization in text-to-image diffusion models, and proposed data augmentation for captions to reduce model memorization since the insufficient diversity in captions easily leads to training data generation. For instance, Somepalli et al. [27] utilized random caption replacement, random token replacement, caption word repetition, etc., to reduce memorization. Another previous approach Daras et al. [7] proposed to use corrupted images for diffusion model training for memorization reduction. Wen et al. [31] introduced methods for detecting memorized prompts through text-conditional predictions and proposed two strategies to mitigate memorization: minimizing during inference or filtering during training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 411, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 553 + ], + "type": "text", + "content": "Although these works represented an important step forward in understanding the memorization issue in diffusion models, they either utilized simple data augmentation strategies or only focused on the easily memorable images that are related to specific captions in cross-modal generation tasks. However, diffusion models have been proven capable of generating images from memory without the need for text guidance [4], and previous methods focused on a limited scope of the memorization issue in diffusion models. Differently, in this paper, we propose a novel training framework for diffusion models from the perspective of visual modality, i.e., IET-AGC (Iterative Ensemble Training with Anti-Gradient Control), which not only reduces memorization fundamentally but also provides a more generic approach for both unconditional and text-conditional diffusion models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "First, we propose an iterative ensemble training (IET) framework to mitigate memorization by parameter aggregation. Training data are stored in parameters of diffusion models due to over-optimization, and the model ensemble strategy aggregates parameters to re-organize model knowledge, which contributes to alleviating the memorization of training data. Thus, we separate the training data into several groups, train diffusion models individually, and ensemble them for memorization reduction. However, aggregating models trained on subsets of data may potentially decrease the performance due to the divergent optimizations on different subsets of these models. Motivated by Federated Learning tech" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 480, + 140 + ], + "type": "text", + "content": "niques [18], we iteratively ensemble the models during training, which reduces memorization by multiple aggregation and maintains generation performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 141, + 481, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 141, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 133, + 141, + 481, + 283 + ], + "type": "text", + "content": "Based on our IET training framework, we further introduce an Anti-Gradient Control module to reduce memorization of training data ulteriorly. We conduct practical analysis on losses during training and find that the training loss for easily memorable images tends to be obviously lower than that for less memorable images. Thus, we propose to exclude the sample with a relatively small loss value from the current mini-batch to avoid memorizing these samples. During the training process, as the diffusion model exhibits varying average loss values across different time steps, we maintain a memory bank to store the average loss of each time step. Subsequently, we discard samples with losses below a certain proportion of the average loss. Since the model has encountered the sample during training, excluding it is unlikely to have a significant impact on the model's performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 285, + 481, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 285, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 133, + 285, + 481, + 450 + ], + "type": "text", + "content": "Extensive experiments on four datasets show the significance of our IET-AGC framework. Our method significantly reduces the memorized quantity by " + }, + { + "bbox": [ + 133, + 285, + 481, + 450 + ], + "type": "inline_equation", + "content": "87.3\\%" + }, + { + "bbox": [ + 133, + 285, + 481, + 450 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 133, + 285, + 481, + 450 + ], + "type": "inline_equation", + "content": "66.4\\%" + }, + { + "bbox": [ + 133, + 285, + 481, + 450 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 133, + 285, + 481, + 450 + ], + "type": "inline_equation", + "content": "85.3\\%" + }, + { + "bbox": [ + 133, + 285, + 481, + 450 + ], + "type": "text", + "content": " compared with the default training (DDPM [14]) on CIFAR-10 and CIFAR-100 and AFHQ-DOG, respectively. In addition, considering the high cost of re-training existing well-trained diffusion models, we also propose an efficient way to address the memorization issue of existing pre-trained models by simply fine-tuning using our method for several epochs. Experiments show that using our method to fine-tune for just two epochs can significantly reduce the memorization by " + }, + { + "bbox": [ + 133, + 285, + 481, + 450 + ], + "type": "inline_equation", + "content": "25.22\\%" + }, + { + "bbox": [ + 133, + 285, + 481, + 450 + ], + "type": "text", + "content": " in unconditional diffusion models on CIFAR-10. Furthermore, when fine-tuning the text-conditional diffusion model, Stable Diffusion, our approach decreases the memorization score by " + }, + { + "bbox": [ + 133, + 285, + 481, + 450 + ], + "type": "inline_equation", + "content": "42.18\\%" + }, + { + "bbox": [ + 133, + 285, + 481, + 450 + ], + "type": "text", + "content": " compared to conventional fine-tuning methods. These results demonstrate that our method performs excellently on both unconditional and text-conditional diffusion models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 471, + 236, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 471, + 236, + 483 + ], + "spans": [ + { + "bbox": [ + 134, + 471, + 236, + 483 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 498, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 498, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 133, + 498, + 480, + 581 + ], + "type": "text", + "content": "Memorization in Generative Models. Several studies have examined the memorization capabilities of the generative model. Generative Adversarial Networks (GANs) [12] have been at the forefront of this research area. As Webster et al. [29] demonstrated when applied to face datasets, GANs can occasionally replicate. Prior study [4] explored an adversarial attack on language models like GPT-2 [21], where individual training examples can be recovered, including personally identifiable information and unique text sequences." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "text", + "content": "Recent studies have shifted their attention toward diffusion models. Somepalli et al. [26] found that diffusion models accurately recall and replicate training images, especially noted with models like the Stable Diffusion model [23]. Building upon this discovery, Carlini et al. [5] developed a tailored black-box attack for diffusion models. They generated images and implemented a membership inference attack to assess density. Webster et al. [28] demonstrated a more efficient extraction attack with fewer network evaluations, identified \"template" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "text", + "content": "IET-AGC: Mitigating Memorization in Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 200 + ], + "type": "text", + "content": "verbatim,\" and discussed its persistence in newer systems. Recent research has shifted towards exploring the theoretical aspects of memory in diffusion models. [32] discovered that generalization and memorization are mutually exclusive occurrences and further demonstrated that the dichotomy between memorization and generalization can be apparent at the class level. Gu et al. [13] extensively studied how factors like data dimension, model size, time embedding, and class conditions affect the memory capacity of the diffusion model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 221, + 481, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 221, + 481, + 484 + ], + "spans": [ + { + "bbox": [ + 132, + 221, + 481, + 484 + ], + "type": "text", + "content": "Memorization Mitigation. The mitigation measures have primarily been concerned with filtering inputs and deduplication. For example, Stable Diffusion employed well-trained detectors to identify unsuitable generated content. However, these temporary solutions can be easily bypassed [22, 30] and do not effectively prevent or lessen copying behavior on a broad scale. Kumari et al. [17] designed an algorithm to align the image distribution with a specific style, instance, or text prompt they aim to remove, to the distribution related to a core concept. This stopped the model from producing target concepts based on its text condition. However, these approaches are inefficient because they necessitate a list of all concepts to be erased, and have not addressed the key issue of how to reduce the memory capacity of the model. [8, 11] explored the use of differential privacy (DP) [9] to train diffusion models or fine-tune ImageNet pre-trained models. However, their focus was on ensuring the privacy of the training of diffusion models, not on the privacy of the images generated by the diffusion models. Daras et al. [7] introduced a technique for training diffusion models utilizing tainted data. By incorporating additional corruption before applying noise, their methodology prevents the model from overfitting to the training data. But their training requires a considerable amount of time. [27, 31] also suggested a series of recommendations to mitigate copying such as randomly replacing the caption of an image with a random sequence of words, but most of which are limited to text-to-image models. Our work focuses on the nature of memorization in diffusion models, especially for unconditional ones." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 507, + 201, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 507, + 201, + 519 + ], + "spans": [ + { + "bbox": [ + 133, + 507, + 201, + 519 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 536, + 480, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 536, + 480, + 560 + ], + "spans": [ + { + "bbox": [ + 133, + 536, + 480, + 560 + ], + "type": "text", + "content": "In this section, we present our methodology aimed at mitigating the memorization in diffusion models, without sacrificing excessive image quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 582, + 301, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 582, + 301, + 594 + ], + "spans": [ + { + "bbox": [ + 133, + 582, + 301, + 594 + ], + "type": "text", + "content": "3.1 Iterative Ensemble Training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 605, + 488, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 605, + 488, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 605, + 488, + 665 + ], + "type": "text", + "content": "Training data are stored in parameters of diffusion models due to over-optimization, and the model ensemble strategy aggregates parameters to re-organize model knowledge. Thus, to mitigate the memorization, we propose a method that trains multiple diffusion models on different data shards of a dataset, merges them after a certain period, and then repeats these two stages iteratively." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 117, + 289, + 250 + ], + "blocks": [ + { + "bbox": [ + 140, + 117, + 289, + 250 + ], + "lines": [ + { + "bbox": [ + 140, + 117, + 289, + 250 + ], + "spans": [ + { + "bbox": [ + 140, + 117, + 289, + 250 + ], + "type": "image", + "image_path": "25cdf156e98300101500475d8a4f9ee9ca31a8a7ed59a3cac40599a0011aa9cd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 259, + 218, + 270 + ], + "lines": [ + { + "bbox": [ + 204, + 259, + 218, + 270 + ], + "spans": [ + { + "bbox": [ + 204, + 259, + 218, + 270 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 294, + 118, + 474, + 245 + ], + "blocks": [ + { + "bbox": [ + 294, + 118, + 474, + 245 + ], + "lines": [ + { + "bbox": [ + 294, + 118, + 474, + 245 + ], + "spans": [ + { + "bbox": [ + 294, + 118, + 474, + 245 + ], + "type": "image", + "image_path": "5b7a0371dc4605e1eb438f22214df4eada10b36a36ffe6fd5e14289c72122c27.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 259, + 398, + 271 + ], + "lines": [ + { + "bbox": [ + 383, + 259, + 398, + 271 + ], + "spans": [ + { + "bbox": [ + 383, + 259, + 398, + 271 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 281, + 482, + 360 + ], + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 360 + ], + "type": "text", + "content": "Fig. 1: Overview of our IET-AGC method. (a) Iterative Ensemble Training (IET): we divide the dataset " + }, + { + "bbox": [ + 130, + 281, + 482, + 360 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 281, + 482, + 360 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 281, + 482, + 360 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 281, + 482, + 360 + ], + "type": "text", + "content": " different data shards. Each shard " + }, + { + "bbox": [ + 130, + 281, + 482, + 360 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 130, + 281, + 482, + 360 + ], + "type": "text", + "content": " trains a separate diffusion model " + }, + { + "bbox": [ + 130, + 281, + 482, + 360 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 130, + 281, + 482, + 360 + ], + "type": "text", + "content": ". After a period of training, these models are merged by averaging and repeating this training strategy. (b) Anti-Gradient Control (AGC): during the training process, we dynamically update and maintain a memory bank of losses at each time step. For loss values smaller than " + }, + { + "bbox": [ + 130, + 281, + 482, + 360 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 281, + 482, + 360 + ], + "type": "text", + "content": " times the corresponding memory bank, we exclude these losses to prevent the model from memorizing such images." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 372, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 482, + 443 + ], + "type": "text", + "content": "Training on Different Data Shards. Unlike the training methods of previous diffusion models, which train a single model on the entire dataset once. In this paper, we divide a dataset into an equal number of shards and then train the corresponding diffusion models on each separate part. If the dataset contains class information, we divide it equally along the class dimension, so that each data shard contains all classes, with the same number of instances for each class." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "type": "text", + "content": "Specifically, suppose the dataset " + }, + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "type": "text", + "content": " classes, with each class having " + }, + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "type": "text", + "content": " samples, and the total number of samples is " + }, + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "type": "inline_equation", + "content": "N = \\sum_{c=1}^{C} N_{c}" + }, + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "type": "text", + "content": ". We divide the dataset into " + }, + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "type": "text", + "content": " equal parts, with each part containing " + }, + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "type": "inline_equation", + "content": "\\frac{N}{K}" + }, + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "type": "text", + "content": " samples. So the " + }, + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "type": "text", + "content": "th part of the dataset " + }, + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "type": "text", + "content": " is represented as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 212, + 505, + 481, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 505, + 481, + 537 + ], + "spans": [ + { + "bbox": [ + 212, + 505, + 481, + 537 + ], + "type": "interline_equation", + "content": "D _ {i} = \\bigcup_ {c = 1} ^ {C} \\left\\{\\left(x _ {c, i, k}, y _ {c, i, k}\\right) \\mid k = 1, 2, \\dots , \\frac {N _ {c}}{K} \\right\\}, \\tag {1}", + "image_path": "ead8436a5b160210663a5d6c151eb9df5c6c50d5c483b08c9bf5a1481c906c7d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "text", + "content": " represents the sample and its corresponding label. Notably, " + }, + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "inline_equation", + "content": "C = 1" + }, + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "text", + "content": " indicates that the dataset does not contain class information. Then, each part " + }, + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "text", + "content": " trains a separate diffusion model " + }, + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "text", + "content": " scaled by the learning rate " + }, + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "text", + "content": " on its own dataset," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 263, + 603, + 481, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 603, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 263, + 603, + 481, + 616 + ], + "type": "interline_equation", + "content": "\\theta_ {i} \\leftarrow \\theta_ {i} - \\eta \\nabla \\mathcal {L} \\left(\\theta_ {i}\\right). \\tag {2}", + "image_path": "4d653d42fa7f80a513fb4d75341af03cc991778120136949b7db71f2494aff5b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Merging the Multiple Diffusion Models. After training for some time, each part obtains a different diffusion model. We simply average the weights of all" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "text", + "content": "IET-AGC: Mitigating Memorization in Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 308, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 308, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 308, + 127 + ], + "type": "text", + "content": "models " + }, + { + "bbox": [ + 132, + 114, + 308, + 127 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 132, + 114, + 308, + 127 + ], + "type": "text", + "content": " to obtain a global model " + }, + { + "bbox": [ + 132, + 114, + 308, + 127 + ], + "type": "inline_equation", + "content": "\\widehat{\\theta}" + }, + { + "bbox": [ + 132, + 114, + 308, + 127 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 276, + 139, + 481, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 139, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 276, + 139, + 481, + 171 + ], + "type": "interline_equation", + "content": "\\widehat {\\theta} \\leftarrow \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\theta_ {i}. \\tag {3}", + "image_path": "264785aa63f9f7762dbec763e42c2bf16a087dc73c894846c57399df019f8ac9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 182, + 482, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 182, + 482, + 265 + ], + "spans": [ + { + "bbox": [ + 131, + 182, + 482, + 265 + ], + "type": "text", + "content": "Then, we repeat the two stages of training on separate shards of the data and merging models, where we use the obtained global model as the initial model for the first stage. As each shard contains only " + }, + { + "bbox": [ + 131, + 182, + 482, + 265 + ], + "type": "inline_equation", + "content": "\\frac{1}{K}" + }, + { + "bbox": [ + 131, + 182, + 482, + 265 + ], + "type": "text", + "content": " of the total data, the training time for each model is proportionally reduced, maintaining the overall computational cost nearly constant compared to training a single model on the entire dataset. The only additional computational cost comes from the periodic merging of models, which is minimal." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 286, + 229, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 286, + 229, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 286, + 229, + 300 + ], + "type": "text", + "content": "3.2 Loss Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 310, + 482, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 310, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 131, + 310, + 482, + 406 + ], + "type": "text", + "content": "To further reduce memorization of training data, we delve deeper into the causes of memorization phenomena, specifically analyzing it through the lens of the loss. We begin by establishing the fundamental notation linked with diffusion models. Diffusion models [14] originate from the non-equilibrium statistical physics [25]. They are essentially straightforward: they operate as image denoisers. During the training process, when given a clean image " + }, + { + "bbox": [ + 131, + 310, + 482, + 406 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 310, + 482, + 406 + ], + "type": "text", + "content": ", time-step " + }, + { + "bbox": [ + 131, + 310, + 482, + 406 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 310, + 482, + 406 + ], + "type": "text", + "content": " is sampled from the interval " + }, + { + "bbox": [ + 131, + 310, + 482, + 406 + ], + "type": "inline_equation", + "content": "[0, T]" + }, + { + "bbox": [ + 131, + 310, + 482, + 406 + ], + "type": "text", + "content": ", along with a Gaussian noise vector " + }, + { + "bbox": [ + 131, + 310, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim N(0, I)" + }, + { + "bbox": [ + 131, + 310, + 482, + 406 + ], + "type": "text", + "content": ", resulting in a noised image " + }, + { + "bbox": [ + 131, + 310, + 482, + 406 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 131, + 310, + 482, + 406 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 256, + 407, + 480, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 407, + 480, + 421 + ], + "spans": [ + { + "bbox": [ + 256, + 407, + 480, + 421 + ], + "type": "interline_equation", + "content": "x _ {t} = \\sqrt {\\alpha_ {t}} x + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\tag {4}", + "image_path": "509b454aca574de3daf583e2cdc275295f6e5af62dc14409dcb34600a65d8160.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 428, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 428, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 131, + 428, + 482, + 477 + ], + "type": "text", + "content": "where the scheduled variance " + }, + { + "bbox": [ + 131, + 428, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 131, + 428, + 482, + 477 + ], + "type": "text", + "content": " varies between 0 and 1, with " + }, + { + "bbox": [ + 131, + 428, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\alpha_0 = 1" + }, + { + "bbox": [ + 131, + 428, + 482, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 428, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\alpha_{T} = 0" + }, + { + "bbox": [ + 131, + 428, + 482, + 477 + ], + "type": "text", + "content": ". The diffusion model then removes the noise to reconstruct the original image " + }, + { + "bbox": [ + 131, + 428, + 482, + 477 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 428, + 482, + 477 + ], + "type": "text", + "content": " by predicting the noise introduced, achieved through stochastic minimization of the objective function " + }, + { + "bbox": [ + 131, + 428, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\frac{1}{N}\\sum_{i}\\mathbb{E}_{t,\\epsilon}\\mathcal{L}(x_i,t,\\epsilon ;\\theta)" + }, + { + "bbox": [ + 131, + 428, + 482, + 477 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 207, + 488, + 480, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 488, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 207, + 488, + 480, + 502 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(x _ {i}, t, \\epsilon ; \\theta\\right) = \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\alpha_ {t}} x _ {i} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , t\\right) \\| _ {2} ^ {2}. \\tag {5}", + "image_path": "ed0f750811ff38b493d1859e5159165364e842a564cb9b41a1d731f3c765317b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 513, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 513, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 513, + 482, + 597 + ], + "type": "text", + "content": "To analyze the correlation between losses and image memorization, we identify images memorized on CIFAR-10 and calculate their loss functions at each time step. Similarly, we sample 256 non-memorized images from the remaining training data and compute their losses at each time step. Fig. 2 shows the comparisons of the losses when the time step is in the interval [0, 600] (" + }, + { + "bbox": [ + 131, + 513, + 482, + 597 + ], + "type": "inline_equation", + "content": "T = 1000" + }, + { + "bbox": [ + 131, + 513, + 482, + 597 + ], + "type": "text", + "content": "). Memorized images exhibit significantly smaller loss values during this period, indicating that the model tends to reconstruct noise into such images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 617, + 274, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 274, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 274, + 628 + ], + "type": "text", + "content": "3.3 Anti-Gradient Control" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "In this subsection, we elaborate on how to utilize the aforementioned loss analysis to devise a training strategy aimed at alleviating the occurrence of memorization." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 189, + 130, + 414, + 293 + ], + "blocks": [ + { + "bbox": [ + 189, + 130, + 414, + 293 + ], + "lines": [ + { + "bbox": [ + 189, + 130, + 414, + 293 + ], + "spans": [ + { + "bbox": [ + 189, + 130, + 414, + 293 + ], + "type": "image", + "image_path": "9e2da4cfb6dde66cf0dfa6474105892f115d5e2d517b1564f48aa035190e7f61.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 299, + 482, + 344 + ], + "lines": [ + { + "bbox": [ + 130, + 299, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 482, + 344 + ], + "type": "text", + "content": "Fig. 2: Comparison of the losses between memorized and non-memorized images. The solid line represents the averaged losses of memorized and non-memorized images, while the two dashed lines represent the losses of the 15th and 85th percentile data respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 355, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 355, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 482, + 475 + ], + "type": "text", + "content": "Memory Bank. In order to identify images with exceptionally low loss values that are prone to memorization during training, we need to maintain the average losses for each time step. However, computing the average loss at each time step entails substantial computational expenses, as it necessitates evaluating the losses for all images using the model at each time step. Thus, we propose a memory bank to store and update losses during mini-batch training without increasing the time cost. However, the losses generally decrease with the training step growing. Therefore, when calculating the average loss in the memory bank, instead of directly averaging all losses at this time step, the aggregating weights should be higher for losses closer to the current update." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 476, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 476, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 476, + 482, + 536 + ], + "type": "text", + "content": "Specifically, we initialize an array of length " + }, + { + "bbox": [ + 130, + 476, + 482, + 536 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 476, + 482, + 536 + ], + "type": "text", + "content": " with zeros, termed the memory bank. The initialization values of the memory bank seem insignificant. After calculating the loss for a mini-batch, we update the memory bank using the Exponential Moving Average (EMA) method based on the loss and the sampled time step, thereby better reflecting the current state of the model:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 234, + 548, + 481, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 548, + 481, + 561 + ], + "spans": [ + { + "bbox": [ + 234, + 548, + 481, + 561 + ], + "type": "interline_equation", + "content": "l _ {t} \\leftarrow \\gamma \\cdot l _ {t} + (1 - \\gamma) \\cdot \\mathcal {L} (x, t, \\epsilon ; \\theta), \\tag {6}", + "image_path": "e3d323fcba65c368e399e8c37b8867e35b22e24262e035c3198b44fc86e51d48.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 571, + 481, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 571, + 481, + 596 + ], + "spans": [ + { + "bbox": [ + 131, + 571, + 481, + 596 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 571, + 481, + 596 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 571, + 481, + 596 + ], + "type": "text", + "content": " represents the smoothing factor, and " + }, + { + "bbox": [ + 131, + 571, + 481, + 596 + ], + "type": "inline_equation", + "content": "l_{t}" + }, + { + "bbox": [ + 131, + 571, + 481, + 596 + ], + "type": "text", + "content": " represents the averaged loss in the memory bank at time step " + }, + { + "bbox": [ + 131, + 571, + 481, + 596 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 571, + 481, + 596 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Anti-Gradient Control (AGC). In previous observations, if the model exhibits memorization of a certain sample, the loss value of the model on that sample tends to be abnormally small. Thus, we use the ratio of the training loss of a certain sample to the mean loss in the memory bank at the time step " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "text", + "content": "IET-AGC: Mitigating Memorization in Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 286, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 286, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 286, + 127 + ], + "type": "text", + "content": "a measure to detect memorization:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 272, + 136, + 480, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 136, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 272, + 136, + 480, + 159 + ], + "type": "interline_equation", + "content": "r = \\frac {\\mathcal {L} (x , t , \\epsilon ; \\theta)}{l _ {t}}. \\tag {7}", + "image_path": "36ffb8fd6522304808f087cc7298a8364458ecb7f8896fa1e32998a76dcb2b0b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 168, + 479, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 168, + 479, + 240 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 479, + 240 + ], + "type": "text", + "content": "A smaller ratio of " + }, + { + "bbox": [ + 132, + 168, + 479, + 240 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 132, + 168, + 479, + 240 + ], + "type": "text", + "content": " indicates a higher likelihood of the image being memorized. To facilitate this determination, we can establish a configurable threshold denoted as " + }, + { + "bbox": [ + 132, + 168, + 479, + 240 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 132, + 168, + 479, + 240 + ], + "type": "text", + "content": ". If the loss ratio " + }, + { + "bbox": [ + 132, + 168, + 479, + 240 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 132, + 168, + 479, + 240 + ], + "type": "text", + "content": " falls below this threshold " + }, + { + "bbox": [ + 132, + 168, + 479, + 240 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 132, + 168, + 479, + 240 + ], + "type": "text", + "content": ", the image is classified as memorized. At this point, we generate a mask that sets the loss value corresponding to this image to zero, i.e. skipping this image in the mini-batch, as shown in the following function," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 223, + 250, + 480, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 250, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 223, + 250, + 480, + 281 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (x, t, \\epsilon ; \\theta) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} r < \\lambda \\\\ \\mathcal {L} (x, t, \\epsilon ; \\theta) & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {8}", + "image_path": "c91b04e86f18b8f89e1305f9e5374e48fca84cd6bdf9466d63eca4e14cfa09b7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 289, + 479, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 479, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 479, + 313 + ], + "type": "text", + "content": "Since the model has encountered the sample during training, excluding it is unlikely to have a significant impact on the model's performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 333, + 229, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 229, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 229, + 346 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 358, + 261, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 261, + 370 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 261, + 370 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 379, + 480, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 480, + 463 + ], + "type": "text", + "content": "Datasets. We evaluate our method on CIFAR-10 [16], CIFAR-100 [16], AFHQDOG [6] for unconditional generation, and LAION-10k [27] for text-conditioned generation. CIFAR-10 and CIFAR-100 consist of 50,000 32x32 color images, divided into 10 and 100 classes, respectively. AFHQ-DOG is a subset of the AFHQ dataset with approximately 5,000 512x512 dog images, resized to 64x64 for our experiments. LAION-10k is a subset of LAION [24], comprising 10,000 image-text pairs with each image having a resolution of 256x256 pixels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 464, + 480, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 480, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 480, + 594 + ], + "type": "text", + "content": "Implementation Details of Training. We conduct experiments on training unconditional diffusion models from scratch using the CIFAR-10, CIFAR-100, and AFHQ-DOG datasets. The IET framework divides CIFAR datasets into 10 shards and AFHQ-DOG into 5 shards. Threshold " + }, + { + "bbox": [ + 130, + 464, + 480, + 594 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 464, + 480, + 594 + ], + "type": "text", + "content": " is set to 0.5 for CIFAR datasets and 0.714 for AFHQ-DOG. Additionally, we perform experiments on fine-tuning pre-trained diffusion models using the CIFAR-10 dataset, maintaining the same hyperparameters as those used for training from scratch. To demonstrate the effectiveness of our method in text-conditioned diffusion models, we fine-tune Stable Diffusion on LAION-10k. The IET framework divides the LAION-10k dataset into 8 shards, with the threshold " + }, + { + "bbox": [ + 130, + 464, + 480, + 594 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 464, + 480, + 594 + ], + "type": "text", + "content": " set to 0.8. The smoothing factor " + }, + { + "bbox": [ + 130, + 464, + 480, + 594 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 464, + 480, + 594 + ], + "type": "text", + "content": " is 0.8 for all datasets. Further details are in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 480, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 642 + ], + "type": "text", + "content": "Extracting Memorized Image. We adopt Carlini's detection rule [5] for unconditional generation, considering " + }, + { + "bbox": [ + 130, + 594, + 480, + 642 + ], + "type": "inline_equation", + "content": "\\bar{x}" + }, + { + "bbox": [ + 130, + 594, + 480, + 642 + ], + "type": "text", + "content": " as memorized if the " + }, + { + "bbox": [ + 130, + 594, + 480, + 642 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 594, + 480, + 642 + ], + "type": "text", + "content": " distance to its nearest neighbor is significantly lower compared to the " + }, + { + "bbox": [ + 130, + 594, + 480, + 642 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 594, + 480, + 642 + ], + "type": "text", + "content": " closest neighbors. We modify this rule to:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 242, + 643, + 480, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 643, + 480, + 669 + ], + "spans": [ + { + "bbox": [ + 242, + 643, + 480, + 669 + ], + "type": "interline_equation", + "content": "\\ell_ {2} (\\bar {x}, x; \\mathbb {S} _ {\\bar {x}} ^ {n}) = \\frac {\\ell_ {2} (\\bar {x} , x)}{\\mathbb {E} _ {y \\in \\mathbb {S} _ {\\bar {x}} ^ {n}} [ \\ell_ {2} (\\bar {x} , y) ]}, \\tag {9}", + "image_path": "1c3c6060edbf40c0832923528fbc3691a1e8f9dd2856f26c74012f52e9853341.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 418, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 418, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 418, + 128 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 116, + 418, + 128 + ], + "type": "inline_equation", + "content": "n = 50" + }, + { + "bbox": [ + 132, + 116, + 418, + 128 + ], + "type": "text", + "content": " in our experiment. A binary classifier is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 224, + 137, + 481, + 150 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 137, + 481, + 150 + ], + "spans": [ + { + "bbox": [ + 224, + 137, + 481, + 150 + ], + "type": "interline_equation", + "content": "I s M e m o (\\bar {x}, x; \\mathbb {S} _ {\\bar {x}} ^ {n}, \\delta_ {V}) = 1 _ {\\ell_ {2} (\\bar {x}, x; \\mathbb {S} _ {\\bar {x}} ^ {n}) \\leq \\delta_ {V}}. \\tag {10}", + "image_path": "0eddb82aaa3eb536d3ac621855cdbe688128d78e64fa68ad42376ed5d1f3d208.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 156, + 479, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 156, + 479, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 156, + 479, + 204 + ], + "type": "text", + "content": "The more images below " + }, + { + "bbox": [ + 130, + 156, + 479, + 204 + ], + "type": "inline_equation", + "content": "\\delta_V" + }, + { + "bbox": [ + 130, + 156, + 479, + 204 + ], + "type": "text", + "content": ", the stronger the model's memorization. We generate 65,536 images per model, calculate their " + }, + { + "bbox": [ + 130, + 156, + 479, + 204 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 156, + 479, + 204 + ], + "type": "text", + "content": " distances, and count images below thresholds " + }, + { + "bbox": [ + 130, + 156, + 479, + 204 + ], + "type": "inline_equation", + "content": "\\delta_V" + }, + { + "bbox": [ + 130, + 156, + 479, + 204 + ], + "type": "text", + "content": " of 0.4, 0.5, and 0.6 to quantitatively evaluate the model's memorization, denoted as " + }, + { + "bbox": [ + 130, + 156, + 479, + 204 + ], + "type": "inline_equation", + "content": "\\mathrm{MQ}_{0.4}" + }, + { + "bbox": [ + 130, + 156, + 479, + 204 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 156, + 479, + 204 + ], + "type": "inline_equation", + "content": "\\mathrm{MQ}_{0.5}" + }, + { + "bbox": [ + 130, + 156, + 479, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 156, + 479, + 204 + ], + "type": "inline_equation", + "content": "\\mathrm{MQ}_{0.6}" + }, + { + "bbox": [ + 130, + 156, + 479, + 204 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 204, + 479, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 204, + 479, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 479, + 239 + ], + "type": "text", + "content": "We adopt Somepalli's detection rule [27] for text-conditioned generation, quantify memorization using a similarity score derived from the dot product of SSCD features [20] of " + }, + { + "bbox": [ + 130, + 204, + 479, + 239 + ], + "type": "inline_equation", + "content": "\\bar{x}" + }, + { + "bbox": [ + 130, + 204, + 479, + 239 + ], + "type": "text", + "content": " and the nearest neighbor " + }, + { + "bbox": [ + 130, + 204, + 479, + 239 + ], + "type": "inline_equation", + "content": "n_0" + }, + { + "bbox": [ + 130, + 204, + 479, + 239 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 267, + 247, + 481, + 261 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 247, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 267, + 247, + 481, + 261 + ], + "type": "interline_equation", + "content": "\\zeta = E (\\bar {x}) ^ {T} \\cdot E (n _ {0}) \\tag {11}", + "image_path": "9ef46107177b2c1b2964e5f89b3d478047bb124ea0b3a55893a7d2d81f1a0c24.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 266, + 479, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 266, + 479, + 291 + ], + "spans": [ + { + "bbox": [ + 131, + 266, + 479, + 291 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 266, + 479, + 291 + ], + "type": "inline_equation", + "content": "E(\\cdot)" + }, + { + "bbox": [ + 131, + 266, + 479, + 291 + ], + "type": "text", + "content": " is SSCD [20]. The dataset similarity score is then defined as the 95th percentile of this distribution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 269, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 269, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 269, + 319 + ], + "type": "text", + "content": "4.2 Experimental Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 326, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 326, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 131, + 326, + 480, + 422 + ], + "type": "text", + "content": "Training from Scratch. The experimental results of our method and four competitive methods are shown in Tab. 1. \"Default\" denotes the conventional training approach of DDPM [14]. \"DP-SGD\" denotes the method of Differentially Private Stochastic Gradient Descent [3], which involves clipping and adding noise to the model's gradients to protect privacy, albeit at the cost of some image quality. Carlina et al. [5] found that utilizing DP-SGD could lead to consistent model divergence. To make DP training more stable, we modify the amplitude of the added noise to be the product of the gradient norm and the noise multiplier:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 275, + 434, + 480, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 434, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 275, + 434, + 480, + 445 + ], + "type": "interline_equation", + "content": "\\sigma = \\left| \\left| \\nabla \\mathcal {L} \\right| \\right| \\times \\tau , \\tag {12}", + "image_path": "f7562d74dbfa8befaa1fc36d1097bd262429aa146df983917bcf4da58a487865.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 450, + 479, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 479, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 479, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 450, + 479, + 510 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 450, + 479, + 510 + ], + "type": "text", + "content": " represents the noise multiplier. \"Adding noise\" denotes a method of directly adding Gaussian noise to the images during training, with a mean of 0 and a variance of 0.1. \"Ambient Diffusion\" [7] protects privacy by training generative models on highly corrupted samples, preventing the model from observing clean training data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 510, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 510, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 131, + 510, + 481, + 641 + ], + "type": "text", + "content": "Results in Tab. 1 show that adding noise to the training images or gradients reduces the quality of the generated images. However, it still does not resolve the issue of training image memorization. Despite Ambient Diffusion also reducing memorization, it leads to a significant increase in FID (from 8.81 to 11.7), indicating a notable degradation of image quality. Compared with the default training approach, our method maintains or even slightly improves the generative quality by reducing the FID score. At the same time, our method significantly reduces the diffusion model's memorization of the training data. As shown in Tab. 1, for the " + }, + { + "bbox": [ + 131, + 510, + 481, + 641 + ], + "type": "inline_equation", + "content": "\\mathrm{MQ}_{0.4}" + }, + { + "bbox": [ + 131, + 510, + 481, + 641 + ], + "type": "text", + "content": " score, the number of memorized images reduced by " + }, + { + "bbox": [ + 131, + 510, + 481, + 641 + ], + "type": "inline_equation", + "content": "87.3\\%" + }, + { + "bbox": [ + 131, + 510, + 481, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 510, + 481, + 641 + ], + "type": "inline_equation", + "content": "66.4\\%" + }, + { + "bbox": [ + 131, + 510, + 481, + 641 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 510, + 481, + 641 + ], + "type": "inline_equation", + "content": "85.3\\%" + }, + { + "bbox": [ + 131, + 510, + 481, + 641 + ], + "type": "text", + "content": " compared with the default training on CIFAR-10, CIFAR-100 and AFHQ-DOG, respectively, illustrating the effectiveness of our method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Visualization. To provide a more intuitive confirmation of our training method, we visualize the images generated by the model along with their closest" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "text", + "content": "IET-AGC: Mitigating Memorization in Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 157, + 482, + 246 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Comparisons of unconditional generation on three datasets in terms of memorized quantity denoted as MQ. We also report the FID to evaluate the quality of images produced by the model." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 157, + 482, + 246 + ], + "lines": [ + { + "bbox": [ + 132, + 157, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 482, + 246 + ], + "type": "table", + "html": "
MethodCIFAR-10CIFAR-100AFHQ-DOG
\\(MQ_{0.4}\\)\\(MQ_{0.5}\\)\\(MQ_{0.6}\\)↓FID↓\\(MQ_{0.4}\\)\\(MQ_{0.5}\\)\\(MQ_{0.6}\\)↓FID↓\\(MQ_{0.4}\\)\\(MQ_{0.5}\\)\\(MQ_{0.6}\\)↓FID↓
Default11146520308.81429172756209.2912344190533079523.59
Adding Noise197593209194.611791037438386.1811700192952722461.18
DP-SGD [3]148728320012.55--------
Ambient Diffusion [7]2213885111.7--------
IET-AGC141178398.3414476032748.51181154351523722.2
", + "image_path": "dbe66bf6134d44e4d115f7c19e99ba33fc49cca3c4a0e6c6a67342648079ac87.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 166, + 266, + 306, + 406 + ], + "blocks": [ + { + "bbox": [ + 166, + 266, + 306, + 406 + ], + "lines": [ + { + "bbox": [ + 166, + 266, + 306, + 406 + ], + "spans": [ + { + "bbox": [ + 166, + 266, + 306, + 406 + ], + "type": "image", + "image_path": "b36517505ffc6e2d207c7f6c6dccfabd05f59f46a1a0f07d04fe5438e9b0b6b0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 407, + 273, + 417 + ], + "lines": [ + { + "bbox": [ + 200, + 407, + 273, + 417 + ], + "spans": [ + { + "bbox": [ + 200, + 407, + 273, + 417 + ], + "type": "text", + "content": "(a) Default Training" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 266, + 448, + 406 + ], + "blocks": [ + { + "bbox": [ + 307, + 266, + 448, + 406 + ], + "lines": [ + { + "bbox": [ + 307, + 266, + 448, + 406 + ], + "spans": [ + { + "bbox": [ + 307, + 266, + 448, + 406 + ], + "type": "image", + "image_path": "d9601d3604a8b0b0d4e62ef6cd6321e245ecafb738a547504cbfe0f84ce83d3e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 407, + 408, + 417 + ], + "lines": [ + { + "bbox": [ + 348, + 407, + 408, + 417 + ], + "spans": [ + { + "bbox": [ + 348, + 407, + 408, + 417 + ], + "type": "text", + "content": "(b) Our Method" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 427, + 482, + 495 + ], + "lines": [ + { + "bbox": [ + 130, + 427, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 482, + 495 + ], + "type": "text", + "content": "Fig. 3: The similar grid image of the default training and our method. Odd-numbered columns represent images from the training set, while even-numbered columns represent images from the generation set that has the smallest " + }, + { + "bbox": [ + 130, + 427, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 427, + 482, + 495 + ], + "type": "text", + "content": " distance to the corresponding training set image. The images in the grid are arranged in ascending order of " + }, + { + "bbox": [ + 130, + 427, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 427, + 482, + 495 + ], + "type": "text", + "content": " distance, and the selected images from both methods correspond to the same position in terms of " + }, + { + "bbox": [ + 130, + 427, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 427, + 482, + 495 + ], + "type": "text", + "content": " distance." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 517, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 589 + ], + "type": "text", + "content": "counterparts in the training dataset in Fig. 3. There is a similarity between the images from the training set and their corresponding images in the generation set, while our method shows lower similarity compared to the default training method. It is also important to note that our model's FID slightly decreases on all three datasets compared to the default indicating that our method also improves the quality of the generated images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "Finetuning Unconditional DDPMs. The results are presented in Tab. 2. The IET method shows improvement over the pre-trained in terms of " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{MQ}_{0.4}" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": " and FID, indicating more efficient data forgetting and higher image quality. The AGC method rapidly causes the model to forget the memorized data. However, it also led to a substantial increase in FID. The IET-AGC method not only reduces the FID but also effectively lowers the model's memory rate." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 122, + 272, + 155 + ], + "lines": [ + { + "bbox": [ + 132, + 122, + 272, + 155 + ], + "spans": [ + { + "bbox": [ + 132, + 122, + 272, + 155 + ], + "type": "text", + "content": "Table 2: Fine-tuning results of the pre-trained DDPM on CIFAR-10 dataset for 2 epochs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 134, + 165, + 269, + 245 + ], + "blocks": [ + { + "bbox": [ + 134, + 165, + 269, + 245 + ], + "lines": [ + { + "bbox": [ + 134, + 165, + 269, + 245 + ], + "spans": [ + { + "bbox": [ + 134, + 165, + 269, + 245 + ], + "type": "table", + "html": "
MethodCIFAR-10
\\(MQ_{0.4}\\)\\(MQ_{0.5}\\)\\(MQ_{0.6}\\)↓FID↓
Default11146520308.81
IET5139623678.33
AGC44235131711.8
IET-AGC8340817967.93
", + "image_path": "c8d8dd9132e1a0454f8e405ce0e553ee1ad6294ab99b09eed78f09f3a245e468.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 274, + 146, + 457, + 261 + ], + "blocks": [ + { + "bbox": [ + 272, + 114, + 449, + 136 + ], + "lines": [ + { + "bbox": [ + 272, + 114, + 449, + 136 + ], + "spans": [ + { + "bbox": [ + 272, + 114, + 449, + 136 + ], + "type": "text", + "content": "Table 3: Fine-tuning results of Stable Diffusion model on LAION-10k dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 274, + 146, + 457, + 261 + ], + "lines": [ + { + "bbox": [ + 274, + 146, + 457, + 261 + ], + "spans": [ + { + "bbox": [ + 274, + 146, + 457, + 261 + ], + "type": "table", + "html": "
MethodSim Score↓Clip Score↑FID↓
Default SD0.6430.518.7
Train TimeMC0.4230.316.6
RC0.5730.616.0
Mitigation [27]CWR0.6130.816.7
Test TimeRT0.5229.518.7
CWR0.5830.118.1
Mitigation [27]GNI0.6230.318.9
Our MethodIET0.4131.316.5
AGC0.5330.618.5
IET-AGC0.3731.316.7
", + "image_path": "5513efa0e034abbec3508d4393e2c7eee925814656bb0f38d7b0206731f46aa8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 151, + 285, + 292, + 388 + ], + "blocks": [ + { + "bbox": [ + 151, + 285, + 292, + 388 + ], + "lines": [ + { + "bbox": [ + 151, + 285, + 292, + 388 + ], + "spans": [ + { + "bbox": [ + 151, + 285, + 292, + 388 + ], + "type": "image", + "image_path": "e547a1c59dc7c57b8a707d3648a523831aa9a1c75b7f5e1e4e3f5fbe4c254fa9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 392, + 305, + 408 + ], + "lines": [ + { + "bbox": [ + 148, + 392, + 305, + 408 + ], + "spans": [ + { + "bbox": [ + 148, + 392, + 305, + 408 + ], + "type": "text", + "content": "(a) Distribution of distances to the most similar images in the dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 285, + 451, + 389 + ], + "blocks": [ + { + "bbox": [ + 309, + 285, + 451, + 389 + ], + "lines": [ + { + "bbox": [ + 309, + 285, + 451, + 389 + ], + "spans": [ + { + "bbox": [ + 309, + 285, + 451, + 389 + ], + "type": "image", + "image_path": "315847b1019744e493e278031641b66fbaa1eb26e90a5d000529215d55213d10.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 392, + 465, + 409 + ], + "lines": [ + { + "bbox": [ + 307, + 392, + 465, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 392, + 465, + 409 + ], + "type": "text", + "content": "(b) Energy distribution. The greater the energy, the more complex the image." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 197, + 419, + 414, + 430 + ], + "lines": [ + { + "bbox": [ + 197, + 419, + 414, + 430 + ], + "spans": [ + { + "bbox": [ + 197, + 419, + 414, + 430 + ], + "type": "text", + "content": "Fig. 4: Most skipped images vs least skipped images." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 441, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 536 + ], + "type": "text", + "content": "Finetuning Text-conditional Stable Diffusion. The results are presented in Tab. 3. Somepalli et al. [27] protects privacy by randomizing conditional information during training and inference, thereby reducing the likelihood of the model replicating specific training data. Our method IET-AGC achieves the best overall results with a Similarity Score of 0.37, maintaining the highest Clip Score of 31.3 and a competitive FID of 16.7. This indicates our approach effectively balances memorization and generative quality, outperforming the default and other mitigation methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 555, + 264, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 264, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 264, + 567 + ], + "type": "text", + "content": "4.3 Analysis of Skipping" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 575, + 481, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 575, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 575, + 481, + 611 + ], + "type": "text", + "content": "In this section, we conduct comparative experiments on the AFHQ-DOG dataset to delve into which types of images are prone to be skipped, as well as the relationship between memorizable images and those that are skipped." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Frequency of Skipped Images. Throughout the training process, we record the identifiers of skipped images. As shown in Fig. 5, our method does not entail skipping all images. In our approach, " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " of the images are skipped fewer than" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 90, + 448, + 102 + ], + "type": "text", + "content": "IET-AGC: Mitigating Memorization in Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 187, + 135, + 406, + 294 + ], + "blocks": [ + { + "bbox": [ + 187, + 135, + 406, + 294 + ], + "lines": [ + { + "bbox": [ + 187, + 135, + 406, + 294 + ], + "spans": [ + { + "bbox": [ + 187, + 135, + 406, + 294 + ], + "type": "image", + "image_path": "339a94c939876f4af05d39443c70935c383938ece7450c5025c9cd15d8b1e3ff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 213, + 300, + 399, + 312 + ], + "lines": [ + { + "bbox": [ + 213, + 300, + 399, + 312 + ], + "spans": [ + { + "bbox": [ + 213, + 300, + 399, + 312 + ], + "type": "text", + "content": "Fig. 5: Distribution of skipped image counts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "text", + "content": "648 times (across a total of 2,278 training epochs), indicating that our method can effectively differentiate between different images. This suggests that we are not simply reducing memorization by constraining the model's learning. On the other hand, while our method requires skipping images with exceptionally low loss values, all images still contribute to the model's training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 404, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 482, + 574 + ], + "type": "text", + "content": "Images Most Easily Skipped. We believe these images are more easily skipped for two main reasons. Firstly, data aggregation. We computed the " + }, + { + "bbox": [ + 130, + 404, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 404, + 482, + 574 + ], + "type": "text", + "content": " distance between these easily skipped images and all other images in the dataset, as well as between those not easily skipped images and all other images in the dataset. Fig. 4a indicates that the distribution of the skipped images is more clustered. Consistent with the findings of Carlini et al. [5], which suggested that removing duplicate training images effectively reduces memorization capacity, skipping these clustered images can also reduce memorization capacity. Secondly, data simplicity. We performed Fourier transforms on these easily skipped and not easily skipped images to obtain their energy distributions, as shown in Fig. 4b. The easily skipped images have less energy, indicating that they lack finer details. We believe both factors contribute to making these images easier for the model to memorize, thus skipping them has a positive impact on reducing the model's memorization capacity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 594, + 239, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 239, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 239, + 605 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Performance Comparisons of Each Component. To further understand the effectiveness of our approach, we conduct ablation experiments to investigate the individual impacts of different components on CIFAR-10. In addition, in our approach, the dataset is evenly distributed, meaning each data shard is set in" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "an IID (independently and identically distributed) manner. To validate the reasonableness of this dataset configuration, we also establish an experiment where each data shard is set up in a non-independent and identically distributed (non-IID) manner. Similar to [15], we employ the Dirichlet distribution to generate data, thereby establishing such a setting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 334, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 334, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 334, + 331 + ], + "type": "text", + "content": "Results in Tab. 4 show that when the data shards are set in non-IID, Iterative Ensemble Training (IET) successfully reduces the memorization " + }, + { + "bbox": [ + 130, + 176, + 334, + 331 + ], + "type": "inline_equation", + "content": "\\mathrm{(IET_{non - IID})}" + }, + { + "bbox": [ + 130, + 176, + 334, + 331 + ], + "type": "text", + "content": ". However, the non-IID splitting reduces the quality of performance since the divergent optimization of different models and aggregating these models affects the performance. Differently, when the data shards are set in IID, " + }, + { + "bbox": [ + 130, + 176, + 334, + 331 + ], + "type": "inline_equation", + "content": "\\mathrm{IET}_{\\mathrm{IID}}" + }, + { + "bbox": [ + 130, + 176, + 334, + 331 + ], + "type": "text", + "content": " can reduce both the memorization and FID score. Thus in this paper, we choose to use the IID splitting strategy. Besides, Tab. 4 shows that both AntiGradient Control (AGC) and Iterative En" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 332, + 482, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 404 + ], + "type": "text", + "content": "semble Training (IET) reduce memorization effectively and AGC is more effective than the IET on the memorization reduction. Compared to the conventional method, AGC reduces " + }, + { + "bbox": [ + 130, + 332, + 482, + 404 + ], + "type": "inline_equation", + "content": "\\mathrm{MQ}_{0.5}" + }, + { + "bbox": [ + 130, + 332, + 482, + 404 + ], + "type": "text", + "content": " by approximately " + }, + { + "bbox": [ + 130, + 332, + 482, + 404 + ], + "type": "inline_equation", + "content": "66\\%" + }, + { + "bbox": [ + 130, + 332, + 482, + 404 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 130, + 332, + 482, + 404 + ], + "type": "inline_equation", + "content": "\\mathrm{IET}_{\\mathrm{IID}}" + }, + { + "bbox": [ + 130, + 332, + 482, + 404 + ], + "type": "text", + "content": " reduces it by around " + }, + { + "bbox": [ + 130, + 332, + 482, + 404 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 130, + 332, + 482, + 404 + ], + "type": "text", + "content": ". However, using AGC only shows a slight growth in FID, which affects the quality of the generated images. Interestingly, combining both IET and AGC leads to better FID and memory reduction effects." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 340, + 217, + 483, + 308 + ], + "blocks": [ + { + "bbox": [ + 339, + 194, + 481, + 215 + ], + "lines": [ + { + "bbox": [ + 339, + 194, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 339, + 194, + 481, + 215 + ], + "type": "text", + "content": "Table 4: Performance comparisons of each component." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 217, + 483, + 308 + ], + "lines": [ + { + "bbox": [ + 340, + 217, + 483, + 308 + ], + "spans": [ + { + "bbox": [ + 340, + 217, + 483, + 308 + ], + "type": "table", + "html": "
MethodCIFAR-10
\\( MQ_{0.4} \\)\\( MQ_{0.5} \\)\\( MQ_{0.6} \\)↓FID↓
Default11146520308.81
\\( IET_{IID} \\)8938217837.61
\\( IET_{non-IID} \\)69315143912.5
AGC2615497611.36
IET-AGC141178398.34
", + "image_path": "992b1ba4ca128c5b36c21753a0e531dcae6b9ba6f8e794dfd0266b79b95177f7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": "Exploring Parameters Impact on Experimental Results. In this study, we examine how various parameters affect our experimental outcomes. By systematically varying these parameters, we aim to understand how they influence our results and to identify the optimal settings for our experiments. Specifically, we conduct a series of experiments where we change the number of shards, training epochs per aggregation period, skipping threshold " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": ", and smoothing factor " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": " of the memory bank. For each variation, we measure the impact on MQ and FID. The default parameters are set with the number of shards " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": " as 10, epochs per shard " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": " as 100, skipping threshold " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": " as 0.5, and smoothing factor " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": " as 0.8. Results are reported in Tab. 5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "content": "Number of Shards " + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "content": ". We investigate the impact of the number of data shards on model performance by setting it to 1, 5, 10, and 15. " + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "inline_equation", + "content": "K = 1" + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "content": " means the default training strategy of diffusion models. Results show that the MQ scores of " + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "inline_equation", + "content": "K = 5, 10, 15" + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "content": " are all lower than " + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "inline_equation", + "content": "K = 1" + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "content": ", indicating that using our IET method can effectively reduce memorization. When " + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "inline_equation", + "content": "K = 10" + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "content": ", the MQ score achieves the best performance. However, with the growth of " + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 545, + 482, + 629 + ], + "type": "text", + "content": ", the FID score increases. Thus we need to balance the FID and MQ to choose the proper number of shards." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Training Epochs per Aggregation Period " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": ". We set the total number of training epochs unchanged and conduct experiments by varying the number of epochs per model aggregation period, i.e., the aggregating frequency of IET." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 90, + 448, + 102 + ], + "type": "text", + "content": "IET-AGC: Mitigating Memorization in Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "Results show that both high and low frequencies of aggregation will reduce the performance of the memorization and image quality. Thus, we choose " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "E = 100" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " to achieve the best performance of MQ." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 152, + 263, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 152, + 263, + 163 + ], + "spans": [ + { + "bbox": [ + 147, + 152, + 263, + 163 + ], + "type": "text", + "content": "Skipping Threshold " + }, + { + "bbox": [ + 147, + 152, + 263, + 163 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 147, + 152, + 263, + 163 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 164, + 264, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 264, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 264, + 354 + ], + "type": "text", + "content": "We evaluate the importance of " + }, + { + "bbox": [ + 130, + 164, + 264, + 354 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 164, + 264, + 354 + ], + "type": "text", + "content": " in mitigating the memorization effect by setting the values of " + }, + { + "bbox": [ + 130, + 164, + 264, + 354 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 164, + 264, + 354 + ], + "type": "text", + "content": " to 0.4, 0.5, 0.66 and 0.8. A large threshold means skipping more training samples that are easily memorized. Results in Tab. 5 show as " + }, + { + "bbox": [ + 130, + 164, + 264, + 354 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 164, + 264, + 354 + ], + "type": "text", + "content": " grows, more memorable training samples are skipped and the memorization phenomenon is further reduced. However, skipping more samples will reduce the model performance, i.e., the generation quality." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 271, + 186, + 480, + 353 + ], + "blocks": [ + { + "bbox": [ + 270, + 164, + 481, + 185 + ], + "lines": [ + { + "bbox": [ + 270, + 164, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 270, + 164, + 481, + 185 + ], + "type": "text", + "content": "Table 5: Parameters impact on experimental results." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 271, + 186, + 480, + 353 + ], + "lines": [ + { + "bbox": [ + 271, + 186, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 271, + 186, + 480, + 353 + ], + "type": "table", + "html": "
ParametersCIFAR-10
\\(MQ_{0.4}\\)\\(MQ_{0.5}\\)\\(MQ_{0.6} \\downarrow\\)FID↓
Number of Shards K111146520308.81
5201297986.37
10141178398.34
152116911439.1
Epoches per Aggregation E102616710189.31
501619411207.76
100141178398.34
2001719313079.01
25024181111410.36
Skipping Threshold λ0.44731015737.72
0.5141178398.34
0.663696088.25
0.834550310.43
Smoothing Factor γ0.52719211938.99
0.8141178398.34
0.9201328868.94
", + "image_path": "0fb4fff06b9007ba3f16c34ce04a80d58a4e39b1fb6368d877cdcf040ec9308a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 147, + 356, + 263, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 356, + 263, + 367 + ], + "spans": [ + { + "bbox": [ + 147, + 356, + 263, + 367 + ], + "type": "text", + "content": "Smoothing Factor " + }, + { + "bbox": [ + 147, + 356, + 263, + 367 + ], + "type": "inline_equation", + "content": "\\gamma" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 368, + 264, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 264, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 264, + 378 + ], + "type": "text", + "content": "The selection of the smooth-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 379, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 439 + ], + "type": "text", + "content": "ing factor " + }, + { + "bbox": [ + 130, + 379, + 482, + 439 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 379, + 482, + 439 + ], + "type": "text", + "content": " is also crucial. When we set " + }, + { + "bbox": [ + 130, + 379, + 482, + 439 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 379, + 482, + 439 + ], + "type": "text", + "content": " to 0.9, the update of the memory bank becomes overly sluggish, failing to faithfully reflect the current model's loss across different time steps. Conversely, if we set " + }, + { + "bbox": [ + 130, + 379, + 482, + 439 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 379, + 482, + 439 + ], + "type": "text", + "content": " to 0.5, the memory bank becomes excessively sensitive, which can lead to instability and fragility during the training process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 456, + 329, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 456, + 329, + 469 + ], + "spans": [ + { + "bbox": [ + 131, + 456, + 329, + 469 + ], + "type": "text", + "content": "5 Limitations and Future Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 480, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 482, + 540 + ], + "type": "text", + "content": "Although our strategy has proven effective in reducing the model's memorization capacity, the selection of the threshold " + }, + { + "bbox": [ + 130, + 480, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 480, + 482, + 540 + ], + "type": "text", + "content": ", which balances between the model's memorization capacity and the quality of the generated images, can only rely on experiments. In future research, we plan to design a quick analysis method for datasets to help determine the appropriate choice of threshold." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 557, + 220, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 557, + 220, + 570 + ], + "spans": [ + { + "bbox": [ + 131, + 557, + 220, + 570 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "This paper innovatively proposes a training strategy for diffusion models, which train models on multiple data shards and ignore data with abnormal loss values. This strategy effectively reduces the memory capacity of the model, further strengthening data privacy protection without compromising the quality of the generated images. We firmly believe that this training strategy has a broad application prospect and great development potential in the field of data privacy protection." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 147, + 240, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 147, + 240, + 162 + ], + "spans": [ + { + "bbox": [ + 133, + 147, + 240, + 162 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 173, + 482, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 173, + 482, + 245 + ], + "spans": [ + { + "bbox": [ + 130, + 173, + 482, + 245 + ], + "type": "text", + "content": "This work was supported by National Natural Science Foundation of China (Grant No. 62306273), National Natural Science Foundation of China (Grant No. 62372341), and the Fundamental Research Funds for the Central Universities (Grant No. 2042024kf0040). The numerical calculations in this paper have been done on the supercomputing system in the Supercomputing Center of Wuhan University." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 264, + 198, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 264, + 198, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 264, + 198, + 277 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 291, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 291, + 400, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 291, + 400, + 301 + ], + "spans": [ + { + "bbox": [ + 138, + 291, + 400, + 301 + ], + "type": "text", + "content": "1. Midjourney team (2022), https://www.midjourney.com/home" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 302, + 335, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 302, + 335, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 302, + 335, + 312 + ], + "type": "text", + "content": "2. Sora team (2024), https://openai.com/sora" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 313, + 481, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 313, + 481, + 356 + ], + "spans": [ + { + "bbox": [ + 138, + 313, + 481, + 356 + ], + "type": "text", + "content": "3. Abadi, M., Chu, A., Goodfellow, I., McMahan, H.B., Mironov, I., Talwar, K., Zhang, L.: Deep learning with differential privacy. In: Proceedings of the 2016 ACM SIGSAC conference on computer and communications security. pp. 308-318 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 357, + 481, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 481, + 400 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 481, + 400 + ], + "type": "text", + "content": "4. Carlini, N., Tramer, F., Wallace, E., Jagielski, M., Herbert-Voss, A., Lee, K., Roberts, A., Brown, T., Song, D., Erlingsson, U., et al.: Extracting training data from large language models. In: 30th USENIX Security Symposium (USENIX Security 21). pp. 2633-2650 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 401, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 401, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 401, + 481, + 434 + ], + "type": "text", + "content": "5. Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramer, F., Balle, B., Ippolito, D., Wallace, E.: Extracting training data from diffusion models. In: 32nd USENIX Security Symposium (USENIX Security 23). pp. 5253-5270 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 434, + 481, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 481, + 466 + ], + "type": "text", + "content": "6. Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8188-8197 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 467, + 481, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 481, + 499 + ], + "type": "text", + "content": "7. Daras, G., Shah, K., Dagan, Y., Gollakota, A., Dimakis, A., Klivans, A.: Ambient diffusion: Learning clean distributions from corrupted data. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 500, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 500, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 500, + 481, + 521 + ], + "type": "text", + "content": "8. Dockhorn, T., Cao, T., Vahdat, A., Kreis, K.: Differentially private diffusion models. arXiv preprint arXiv:2210.09929 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 522, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 481, + 544 + ], + "type": "text", + "content": "9. Dwork, C.: Differential privacy. In: International colloquium on automata, languages, and programming. pp. 1-12. Springer (2006)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 544, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 481, + 577 + ], + "type": "text", + "content": "0. Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5111-5120 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 577, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 481, + 609 + ], + "type": "text", + "content": "1. Ghalebikesabi, S., Berrada, L., Gowal, S., Ktena, I., Stanforth, R., Hayes, J., De, S., Smith, S.L., Wiles, O., Balle, B.: Differentially private diffusion models generate useful synthetic images. arXiv preprint arXiv:2302.13861 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 610, + 481, + 643 + ], + "type": "text", + "content": "12. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "text", + "content": "13. Gu, X., Du, C., Pang, T., Li, C., Lin, M., Wang, Y.: On memorization in diffusion models. arXiv preprint arXiv:2310.02664 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "text", + "content": "IET-AGC: Mitigating Memorization in Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "type": "text", + "content": "15. Hsu, T.M.H., Qi, H., Brown, M.: Measuring the effects of non-identical data distribution for federated visual classification. arXiv preprint arXiv:1909.06335 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 163, + 480, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 163, + 480, + 184 + ], + "spans": [ + { + "bbox": [ + 133, + 163, + 480, + 184 + ], + "type": "text", + "content": "16. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 185, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 185, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 134, + 185, + 480, + 217 + ], + "type": "text", + "content": "17. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22691-22702 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 219, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 219, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 134, + 219, + 480, + 251 + ], + "type": "text", + "content": "18. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273-1282. PMLR (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 252, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 252, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 134, + 252, + 480, + 285 + ], + "type": "text", + "content": "19. Ni, Z., Wei, L., Li, J., Tang, S., Zhuang, Y., Tian, Q.: Degeneration-tuning: Using scrambled grid shield unwanted concepts from stable diffusion. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 8900-8909 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 285, + 480, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 285, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 134, + 285, + 480, + 318 + ], + "type": "text", + "content": "20. Pizzi, E., Roy, S.D., Ravindra, S.N., Goyal, P., Douze, M.: A self-supervised descriptor for image copy detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14532-14542 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 319, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 319, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 134, + 319, + 480, + 341 + ], + "type": "text", + "content": "21. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 342, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 342, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 134, + 342, + 480, + 363 + ], + "type": "text", + "content": "22. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 134, + 364, + 480, + 396 + ], + "type": "text", + "content": "23. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 397, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 397, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 134, + 397, + 480, + 430 + ], + "type": "text", + "content": "24. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 431, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 431, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 134, + 431, + 480, + 464 + ], + "type": "text", + "content": "25. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 465, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 465, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 134, + 465, + 480, + 508 + ], + "type": "text", + "content": "26. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Diffusion art or digital forgery? investigating data replication in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6048-6058 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 509, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 509, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 134, + 509, + 480, + 542 + ], + "type": "text", + "content": "27. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Understanding and mitigating copying in diffusion models. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 543, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 543, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 134, + 543, + 480, + 564 + ], + "type": "text", + "content": "28. Webster, R.: A reproducible extraction of training images from diffusion models. arXiv preprint arXiv:2305.08694 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 565, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 565, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 134, + 565, + 480, + 597 + ], + "type": "text", + "content": "29. Webster, R., Rabin, J., Simon, L., Jurie, F.: This person (probably) exists. identity membership attacks against gan generated faces. arXiv preprint arXiv:2107.06018 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 599, + 480, + 632 + ], + "type": "text", + "content": "30. Wen, Y., Jain, N., Kirchenbauer, J., Goldblum, M., Geiping, J., Goldstein, T.: Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "text", + "content": "31. Wen, Y., Liu, Y., Chen, C., Lyu, L.: Detecting, explaining, and mitigating memorization in diffusion models. In: The Twelfth International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "X. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 172 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "text", + "content": "32. Yoon, T., Choi, J.Y., Kwon, S., Ryu, E.K.: Diffusion probabilistic models generalize when they fail to memorize. In: ICML 2023 Workshop on Structured Probabilistic Inference " + }, + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "inline_equation", + "content": "\\{\\backslash\\&\\}" + }, + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "text", + "content": " Generative Modeling (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "text", + "content": "33. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591 (2023)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "text", + "content": "IET-AGC: Mitigating Memorization in Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/JDT3D_ Addressing the Gaps in LiDAR-Based Tracking-by-Attention/b2b98b8d-db63-4b59-ae29-0da5121d4534_content_list.json b/2024/JDT3D_ Addressing the Gaps in LiDAR-Based Tracking-by-Attention/b2b98b8d-db63-4b59-ae29-0da5121d4534_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9ee9661b50de32162c99e957d4c8d4dc2094633b --- /dev/null +++ b/2024/JDT3D_ Addressing the Gaps in LiDAR-Based Tracking-by-Attention/b2b98b8d-db63-4b59-ae29-0da5121d4534_content_list.json @@ -0,0 +1,1703 @@ +[ + { + "type": "text", + "text": "JDT3D: Addressing the Gaps in LiDAR-Based Tracking-by-Attention", + "text_level": 1, + "bbox": [ + 233, + 140, + 769, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Brian Cheong\\* Jiachen Zhou\\* and Steven Waslander", + "bbox": [ + 287, + 212, + 712, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Toronto \n{brian.cheong, jason.zhou, steven.waslander}@robotics.utias.utoronto.ca", + "bbox": [ + 225, + 239, + 774, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Tracking-by-detection (TBD) methods achieve state-of-the-art performance on 3D tracking benchmarks for autonomous driving. On the other hand, tracking-by-attention (TBA) methods have the potential to outperform TBD methods, particularly for long occlusions and challenging detection settings. This work investigates why TBA methods continue to lag in performance behind TBD methods using a LiDAR-based joint detector and tracker called JDT3D. Based on this analysis, we propose two generalizable methods to bridge the gap between TBD and TBA methods: track sampling augmentation and confidence-based query propagation. JDT3D is trained and evaluated on the nuScenes dataset, achieving 0.574 on the AMOTA metric on the nuScenes test set, outperforming all existing LiDAR-based TBA approaches by over $6\\%$ . Based on our results, we further discuss some potential challenges with the existing TBA model formulation to explain the continued gap in performance with TBD methods. The implementation of JDT3D can be found at the following link: https://github.com/TRAILab/JDT3D.", + "bbox": [ + 261, + 308, + 743, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Multi-Object Tracking $\\cdot$ Autonomous Vehicles $\\cdot$ Computer Vision", + "bbox": [ + 261, + 542, + 743, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 599, + 375, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-object tracking (MOT) is an essential component in the perception system of autonomous vehicles, as it allows autonomous agents to reason and plan appropriately in dynamic environments. The task involves identifying object trajectories, requiring that each object is accurately detected and consistently identified over time within the scene.", + "bbox": [ + 212, + 632, + 787, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The majority of MOT methods adhere to the tracking-by-detection (TBD) paradigm. This paradigm is comprised of two separate processes: generating bounding box predictions without track IDs and associating these predictions to a set of maintained tracks based on some matching criteria. TBD methods currently achieve state-of-the-art results on 3D MOT benchmarks [12, 15, 23].", + "bbox": [ + 212, + 707, + 787, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast to the decoupled approach of TBD, joint detection and tracking (JDT) approaches have attempted to unify the detection and tracking tasks in an", + "bbox": [ + 215, + 784, + 787, + 814 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Denotes equal contribution.", + "bbox": [ + 217, + 824, + 419, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "end-to-end manner [18, 19, 29, 31]. Recent works in autonomous driving perception [7, 17] and computer vision [8] demonstrate that networks trained end-to-end have shown superior performance compared to pipeline-based techniques, despite initially lower performance from earlier works [21, 27]. End-to-end methods can learn richer and more generalizable representations while minimizing the task-specific engineering of individual modules required in a pipeline-based approach. While TBD methods enforce a unidirectional flow of information from the detector to the tracker, the JDT formulation allows both the detector and tracker to exchange useful information to enhance the performance of both tasks.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Integrating detection and tracking into a joint learning task requires a unique approach to formulating the problem and training the model. One such approach is tracking-by-attention (TBA), illustrated in Fig. 1. In TBA, objects are represented as vector embeddings, or \"queries\", that are used to detect the same object over multiple frames. While recent TBA methods [18,19,29,31] have been proposed as alternatives to TBD, they continue to underperform on the MOT task, especially in the LiDAR domain [29]. However, they have shown potential in certain cases, such as significantly fewer ID-switch errors [19]. The unexplored potential of TBA in the LiDAR domain motivates our work in addressing a key question: what is holding back LiDAR-based TBA? In this work, we propose a LiDAR-based TBA tracking method called JDT3D, based on which we explore different aspects that hold back LiDAR-based TBA.", + "bbox": [ + 212, + 286, + 787, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One insight is that LiDAR-based tracking methods can suffer from sparse supervision signals due to point cloud and object sparsity. Inspired by object sampling augmentation [25] for LiDAR-based detection methods, we propose track sampling, a temporally consistent augmentation method that injects consistent objects over multiple LiDAR frames to enrich supervision signals while maintaining temporal consistency.", + "bbox": [ + 212, + 470, + 787, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, existing TBA methods maintain an inconsistency between training and inference regarding their query propagation method. During training, queries that are matched to ground truth samples are propagated to the next frame, but at inference, propagation is based on a confidence threshold. This inconsistency between training and inference could confuse the model to over-trust false positive queries. To address this, we propose a confidence threshold propagation strategy consistent for both training and inference and conduct comprehensive evaluations and discussions on the strategy design.", + "bbox": [ + 212, + 566, + 787, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contributions as follows:", + "bbox": [ + 238, + 691, + 557, + 705 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "JDT3D We present JDT3D, a LiDAR-based tracking-by-attention model that outperforms existing LiDAR-based tracking-by-attention methods on the nuScenes test set by over $6\\%$ on the AMOTA metric.", + "bbox": [ + 212, + 710, + 797, + 756 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Confidence-based query propagation We perform ablations to validate the effectiveness and generalizability of the confidence-based propagation criterion.", + "bbox": [ + 212, + 760, + 787, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Track Sampling Augmentation We propose and validate a novel data augmentation technique for multi-frame LiDAR methods that significantly improves performance and convergence rates during training.", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "B. Cheong et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/901813337f9bd5a1067850121f1a0273eb4f80e96490e461851e5b3a37ae91f1.jpg", + "image_caption": [ + "Fig. 1: Illustration of tracking by attention. Each object is represented by a query. In the diagram, the yellow motorcycle leaves the frame, so it is removed from the set of maintained queries. To detect new objects in the scene, such as the tan bus, proposal queries are appended at each time step, represented by the blank squares." + ], + "image_footnote": [], + "bbox": [ + 290, + 146, + 709, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Analysis Based on our experiments using JDT3D, we provide a detailed analysis to explain the continued gap in performance between LiDAR-based TBD and TBA. We attribute the gap to the reduced detection performance when training on multiple consecutive frames and the handling of queries that results in \"temporal confusion\".", + "bbox": [ + 212, + 364, + 787, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 460, + 395, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-object tracking (MOT) methods either follow the tracking-by-detection paradigm, in which object detection is independently performed on single frames before the tracking module, or they perform detection and tracking jointly in an end-to-end fashion. MOT can also be further categorized into 2D and 3D MOT. The task of 2D MOT involves predicting object tracks in 2D space, while 3D MOT predicts the tracks in 3D space. 2D MOT methods generally take in camera images as sensor input, while 3D MOT methods may take in camera images, LiDAR point clouds, or both. In this work, we will be focusing our analysis on 3D MOT methods.", + "bbox": [ + 212, + 492, + 787, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Tracking-by-detection (TBD)", + "text_level": 1, + "bbox": [ + 215, + 650, + 501, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the tracking-by-detection paradigm, the detection and tracking problems are solved by two distinct modules. First, a single-frame object detector is used to detect objects in a given scene. Next, the detections are associated with a set of maintained tracks by a tracking module and used to update them with current detection information. Unassigned detections are used to initialize new tracks.", + "bbox": [ + 212, + 672, + 787, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While different tracking methods exist among TBD methods, they are all consistently tied to the detector's performance. Several top-performing 3D tracking methods [1,4,15] are strong detectors that have been paired with the simple tracking method of CenterPoint [26]. The formulation of TBD inhibits the detector from taking advantage of temporal object information that would enhance object detection. As well, errors in the detection are guaranteed to be propagated", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "JDT3D", + "bbox": [ + 678, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "downstream to the tracking task. This motivates the exploration of methods that can jointly optimize detection and tracking to achieve MOT.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Joint Detection and Tracking (JDT)", + "text_level": 1, + "bbox": [ + 214, + 195, + 562, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Joint detection and tracking (JDT) is a recent paradigm in both 2D and 3D MOT that involves training a model to perform detection and tracking end-to-end. In contrast to the TBD methods that rely on a dedicated detector to generate detections, JDT methods simultaneously output object detections with associated tracking IDs. Without heuristic matching or hand-crafted rules, SimTrack [16] builds upon CenterPoint [26] by learning a hybrid-time centerness map between two consecutive LiDAR point clouds. AlphaTrack [28] adds a cross-modal fusion scheme to the CenterPoint tracking method [26] and applies an alternating training strategy to balance between the detection branch and appearance affinity. The above two methods perform detection and tracking jointly, but known tracks are not utilized to inform future detections. 3D DetecTrack [9] proposes tracklet-aware 3D detection that reconfigures the initial detection using the latest tracklets' information and employs a spatio-temporal gated GNN for association.", + "bbox": [ + 212, + 217, + 787, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tracking-by-Attention (TBA) A growing trend within JDT is the tracking-by-attention (TBA) approach that emerged from TrackFormer [18] for 2D MOT. TBA uses the transformer architecture [22] and formulates tracking as a frame-to-frame set prediction problem that can be trained in an end-to-end manner.", + "bbox": [ + 212, + 431, + 785, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TrackFormer [18] extends a 2D query-based detector, DETR [3] and its variant Deformable DETR [33], by designing autoregressive track queries that are passed across frames with associated object identities. MOTRv2 [31] uses a pretrainedYOLOX detector to initialize their object queries and uses the detections to anchor the decoded bounding box positions. In addition, track queries during training are propagated based on the prediction confidence rather than whether a prediction was matched to a ground truth object. This naturally creates false positive and false negative cases during training, removing the need for query injection and removal augmentations [18,27]. JDT3D applies confidence-based track query propagation to LiDAR-based TBA, and we perform ablations showing the method's improved performance over the ground truth matching criterion.", + "bbox": [ + 212, + 492, + 787, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MUTR3D [30] and PF-Track [19] demonstrate the transfer of the tracking-by-attention framework to multi-view vision-only 3D MOT. PF-Track [19] also designs past and future reasoning modules to enhance the spatio-temporal representation of queries. MotionTrack [29] investigates the feasibility of extending the TBA paradigm to LiDAR-based and LiDAR-camera fusion-based 3D JDT by building upon TransFusion [1], a query-based 3D object detector. Rather than implicitly learning which queries should be suppressed through the transformer decoder, MotionTrack learns an explicit affinity matrix between tracks and objects in the scene and performs greedy matching to determine their association. This explicit matching echoes the decoupling found in TBD methods and potentially detracts from the benefits of an end-to-end method. While vision-only", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "B. Cheong et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d7d0f7ec76b29fe2ec4998c43430ba14e8da77e45bcfc6d02eedca62b5c8dacc.jpg", + "image_caption": [ + "Fig. 2: JDT3D Architecture. At each time step, a BEV feature map is extracted and used to initialize a set of proposal queries. The proposal queries are concatenated to track queries passed from the previous frame and used to predict objects in the scene. Track queries detect the same unique objects in each time step, while proposal queries detect untracked or new objects." + ], + "image_footnote": [], + "bbox": [ + 246, + 146, + 751, + 300 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TBA methods have shown surprising results on the MOT task, the baseline results set by MotionTrack's LiDAR-only and fusion TBA methods underperform compared to LiDAR-based TBD baselines such as CenterPoint [26] and SimpleTrack [20], prompting further exploration of LiDAR TBA methods. Our work explores different potential hypotheses for the reported performance gap between LiDAR-based TBD and TBA methods and proposes improvements to the existing TBA methods.", + "bbox": [ + 212, + 426, + 787, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3 LiDAR Data Augmentations", + "text_level": 1, + "bbox": [ + 214, + 579, + 498, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As with other machine learning tasks, data augmentation can greatly enhance the performance of LiDAR-based 3D object detection methods. Some common augmentations include random flipping, rotation, translation, and scaling of the point cloud [6]. In addition to these, object sampling is also a widely used augmentation method for LiDAR-based object detectors [1,15,26]. First introduced in SECOND [25], object sampling involves creating a database of ground truth object point cloud scans and bounding boxes and randomly sampling and injecting them into the scene during training to improve the convergence and performance overall. However, this augmentation does not apply directly to training TBA methods, since it would result in sporadic track instances in the scene in each training sequence. Our proposed track sampling augmentation samples consecutive track instances and injects them into a sequence of frames during training. This maintains the temporal nature of the object tracks while still offering the same benefits of faster convergence.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "JDT3D", + "bbox": [ + 678, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 143, + 330, + 160 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To perform our analysis, we developed our own baseline LiDAR-based TBA method since no such implementations were made available as of the writing of this paper. We build on aspects from query-based 3D object detection [1] and vision-only TBA [19] to formulate our method, called JDT3D. JDT3D is summarized in Fig. 2.", + "bbox": [ + 212, + 172, + 787, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 266, + 339, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At each time step, a feature extractor takes in the LiDAR scan of the scene, $\\mathbf{P}$ , and converts it into a BEV feature map representation, $\\mathbf{F}$ , using a standard voxel feature extractor [32]. This feature map is used to predict heatmaps of potential object locations and initialize a set of proposal queries, $\\mathbf{PQ}$ [1].", + "bbox": [ + 212, + 287, + 785, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the first time step, the proposal queries are passed through a transformer decoder layer with the BEV feature map to obtain a set of updated queries, $\\mathbf{Q}^D$ . The past reasoning module refines the updated queries with historical track information [19] before a feedforward network decodes the refined queries, $\\mathbf{Q}^R$ into bounding boxes and class predictions, $\\mathbf{B}^R$ . During training, $\\mathbf{Q}^D$ is decoded into a set of bounding boxes and class predictions, $\\mathbf{B}^D$ , for an auxiliary loss.", + "bbox": [ + 212, + 348, + 785, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To create the track queries for the next frame, the predictions and their corresponding queries are filtered with a confidence threshold. The filtered queries are passed to the future reasoning module [19] which predicts a motion trajectory and projects each query into the next time step.", + "bbox": [ + 212, + 438, + 785, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At every following time step, new proposal queries are initialized with the heatmap as described above and are concatenated with track queries from the previous time step, $\\mathbf{TQ}$ , before being passed to the transformer decoder.", + "bbox": [ + 212, + 500, + 785, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To perform tracking, high-confidence predictions and associated queries are assigned an ID. Track queries are used to detect the same object consistently over multiple frames, while proposal queries are used to detect objects to be added to the track query set. Tracking is achieved implicitly through the direct relationship between each prediction and each query passed into the transformer decoder. Trajectories can be obtained by tracing a given track query back in time.", + "bbox": [ + 212, + 544, + 785, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Query Propagation", + "text_level": 1, + "bbox": [ + 215, + 652, + 419, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During training, only the refined queries corresponding to predictions with a confidence score greater than $\\tau_{pass}$ are passed to the next time step. This method of passing and pruning queries has the advantage of naturally creating instances of false negative and false positive track queries during training [31]. False negative track queries are missed tracks that should be re-detected in the next frame as if they were new objects entering the scene. False positive track queries simulate tracks that have left the scene and should be pruned in the next time step.", + "bbox": [ + 212, + 672, + 787, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This method has the benefit of better matching the inference query propagation behaviour. As well, it creates more instances during training where the decoder must suppress track query predictions and output positive proposal query predictions, ensuring robust handling of track births and deaths.", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "B. Cheong et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cd0239ea36c858749982b1139b5d1298c1d48296c36ee3f0dfde4ee0a09b691b.jpg", + "image_caption": [ + "Fig. 3: An example of track sampling augmentation over three consecutive frames from the nuScenes dataset. The original trajectories and sampled trajectories are shown in blue and orange, respectively. Only a subset of the tracks and the third LiDAR scan are shown for visual clarity. The older boxes are shown with transparency." + ], + "image_footnote": [], + "bbox": [ + 318, + 145, + 498, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3bd3bc5c2b0c7d9f1126b7739e7e1008337f8090d5a86ba3dea183404952ae29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 146, + 684, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Ground Truth Assignment", + "text_level": 1, + "bbox": [ + 215, + 382, + 480, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The ground truth assignment of the decoder predictions, $\\mathbf{B}_t^D$ , and the refined predictions, $\\mathbf{B}_t^R$ , are both done in the same manner using a two-stage assignment process. The first stage involves matching the track queries based on their previously matched ground truth objects [18]. If the previously matched ground truth track is not present in the current time step, such as in the case of a track leaving the scene, then the track query prediction is assigned to no object. False positive track queries were never assigned a ground truth track previously, so they are also assigned to no object.", + "bbox": [ + 212, + 409, + 787, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The second stage of the assignment involves matching the proposal queries to the unmatched ground truths. The unmatched ground truths would include new objects in the scene and missed objects from previous time steps. A Hungarian matcher is used to compute the optimal assignment between the unmatched queries and ground truths based on an association cost [10].", + "bbox": [ + 212, + 531, + 787, + 607 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Track Sampling Augmentation", + "text_level": 1, + "bbox": [ + 215, + 631, + 514, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To improve the convergence rate and performance, we introduce track sampling augmentation during training. Inspired by the object sampling augmentation [25], our track sampling involves adding additional tracks into a training clip from a database of tracks in a temporally consistent manner, as illustrated in Fig. 3. The original object sampling augmentation does not account for temporal consistency, and applying it naively to our multi-frame training scheme would result in disjointed track instances during training.", + "bbox": [ + 212, + 657, + 787, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, the training set is preprocessed to generate a database of ground truth tracks, including bounding boxes, object classes, track IDs, and LiDAR points contained in the bounding boxes. To perform track sample augmentation on a training clip, a set of tracks is randomly sampled, avoiding track IDs that overlap with the existing ground truth objects in the scene. This enforces unique IDs for", + "bbox": [ + 212, + 765, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "JDT3D", + "bbox": [ + 678, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the tracks in the sequence. To inject a track, $L$ consecutive ground-truth boxes are randomly selected from each sampled track. In each training frame, the LiDAR points corresponding to the sampled objects are used to replace regions in the original point cloud. The objects are injected in chronological order over the $L$ frames. A collision check is performed to ensure that there are no collisions between injected objects and the original ground truth objects in the scene. In the case of a collision, the would-be sampled object is pruned from that frame, simulating an occlusion.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The benefits of track sampling mirror those of object sampling [25], with the extra ground truth objects in the scene providing more samples to learn from in a single frame. This results in a stronger training signal and more varied scenes.", + "bbox": [ + 212, + 267, + 787, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Network Training and Losses", + "text_level": 1, + "bbox": [ + 215, + 335, + 503, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To train JDT3D, $L$ LiDAR frames are passed sequentially to the model as a training sample. The loss for a given frame, $\\mathcal{L}_i$ , is computed as the weighted sum of the following losses:", + "bbox": [ + 212, + 361, + 787, + 407 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i} = \\lambda_ {h} \\mathcal {L} _ {h} + \\lambda_ {f} \\mathcal {L} _ {f} + \\sum_ {m = \\{D, R \\}} \\left(\\lambda_ {r e g} ^ {m} \\mathcal {L} _ {r e g} ^ {m} + \\lambda_ {c l s} ^ {m} \\mathcal {L} _ {c l s} ^ {m}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 419, + 785, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The superscripts $D$ and $R$ correspond to predictions output by $\\mathbf{Q}^D$ and $\\mathbf{Q}^R$ , respectively. $\\mathcal{L}_{reg}^{m}$ is the L1 loss over the bounding boxes, and $\\mathcal{L}_{cls}^{m}$ is the focal loss [14] over the classification predictions. $\\mathcal{L}_h$ is the Gaussian Focal Loss on the predicted object heatmaps [11,26], and $\\mathcal{L}_f$ is the L1 loss over the motion trajectory predicted by the future reasoning module [19]. Each loss has an associated scalar weight term, $\\lambda$ . The total loss is computed as the sum of each individual frame loss:", + "bbox": [ + 212, + 464, + 787, + 569 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c l i p}} = \\sum_ {i = 1} ^ {L} \\mathcal {L} _ {i} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 450, + 569, + 785, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 630, + 375, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Dataset and Metrics", + "text_level": 1, + "bbox": [ + 215, + 662, + 431, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "JDT3D is trained and evaluated on the nuScenes dataset [2], a multi-modal autonomous driving dataset containing data from six multi-view cameras, five radars, and a 360-degree LiDAR. The dataset contains a total of 40,157 frames across 1000 scenes, of which 850 scenes are annotated for training (700) and validation (150) at a frame rate of $2\\mathrm{Hz}$ . We train on the seven object classes annotated for the tracking task.", + "bbox": [ + 212, + 688, + 787, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate JDT3D, the average multi-object tracking accuracy (AMOTA) and average multi-object tracking precision (AMOTP) tracking metrics [24] are computed across the seven tracking classes of nuScenes, evaluating the tracking accuracy and tracking position misalignment, respectively. Secondary metrics are", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "B. Cheong et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/04def74aad5ef068e27f800fd22e8d1895f187d20928533edb95b6541e9e37fe.jpg", + "table_caption": [ + "Table 1: A comparison of 3D multi-object tracking methods evaluated on the nuScenes [2] test set. Bold represents the best results among the Transformer-based TBA methods, while Underlined numbers are the best overall." + ], + "table_footnote": [], + "table_body": "
MethodTBAModalityAMOTA ↑AMOTP ↓FP ↓FN ↓IDS ↓
AB3DMOT [24]L0.1511.50115088757309027
Chui et al. [5]L0.5500.7981753333216950
CenterPoint [26]L0.6380.5551861222928760
SimpleTrack [20]L0.6680.5501751423451575
PF-Track [19]✓C0.4341.2521904842758249
DQTrack [13]✓C0.5231.096--1204
MotionTrack-L [29]✓L0.510.99--9705
MotionTrack-LC [29]✓L+C0.550.871--8716
JDT3D (ours)✓L0.5740.8371115229919254
", + "bbox": [ + 230, + 198, + 767, + 345 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "also computed by the nuScenes MOT benchmark [2], such as the number of false positives, false negatives, and identity switches. We also evaluate the detection performance across the tracking classes, with the mean average precision (mAP) and nuScenes Detection Score (NDS) metrics.", + "bbox": [ + 215, + 375, + 785, + 436 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 217, + 459, + 455, + 474 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "VoxelNet [25, 32] is used as the LiDAR backbone to extract the BEV feature map. For each frame, the top 200 predictions from the heatmap are used to initialize the proposal queries. The training clip length $L$ is set to 3 frames. The confidence threshold for track queries $\\tau_{pass}$ is set to 0.4 for training and inference. The Hungarian matching cost for ground truth assignment consists of the sum of the L1 cost on the predicted box and the focal loss [14] on the predicted class confidence. All detections are generated in the car ego-frame, so the object centers of the track queries are updated to their relative positions in the current training frame based on the car's ego-motion between time steps. A pretrained VoxelNet backbone from BEVFusion [15] is used to initialize the feature extractor. To train JDT3D, the model is first trained on the detection task using $L = 1$ for 20 epochs. It is then further trained on the tracking task using $L = 3$ for another 20 epochs. All experiments were trained and evaluated on eight NVIDIA V100 GPUs using random flipping, rotation, translation, and scaling on the training sequence, and track sampling augmentation, unless otherwise specified. JDT3D performs inference at $3.85\\mathrm{Hz}$ on a Tesla V100.", + "bbox": [ + 215, + 486, + 787, + 728 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Main Results", + "text_level": 1, + "bbox": [ + 217, + 752, + 372, + 766 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our main results compare JDT3D with baseline TBD methods and relevant TBA methods evaluated on the nuScenes test set [2], summarized in Tab. 1. JDT3D outperforms the current state-of-the-art LiDAR-based TBA methods, with a relative improvement of $12.5\\%$ on the AMOTA metric and $15.5\\%$ on the", + "bbox": [ + 215, + 779, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "JDT3D", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/debc44cdbd77910d9f471ee093348e47734e1f0f9ae4e8c647cd2553d3b26e96.jpg", + "table_caption": [ + "Table 2: Ablation on the track sampling augmentation, reported on the nuScenes validation set. The mAP metric is computed on the seven tracking categories." + ], + "table_footnote": [], + "table_body": "
Track Sampling RateAMOTA ↑AMOTP ↓IDS ↓mAP ↑
0.00.3141.3773880.225
0.250.6000.8642100.3867
0.500.6220.8162030.400
1.00.6210.7901500.399
", + "bbox": [ + 303, + 184, + 697, + 263 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "AMOTP metric on the test split. Moreover, with LiDAR input only, JDT3D also outperforms the fusion-based MotionTrack-LC [29] with a relative improvement of $4.36\\%$ on AMOTA and $3.79\\%$ on AMOTP on the test set. While JDT3D underperforms compared to baseline TBD methods on most metrics, it achieves the lowest number of ID switches among LiDAR-only methods, indicating a strong track coherence.", + "bbox": [ + 212, + 289, + 782, + 378 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "JDT3D has significantly more false negatives compared to TBD methods, but fewer false positives. The potential sources of false negatives are either due to missed detections or tracks diverging from the ground truths. However, the latter case would result in both high false negatives and false positives. Based on the low false positive results from JDT3D, we conclude that the false negatives are mainly due to missed detections.", + "bbox": [ + 212, + 380, + 784, + 470 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 492, + 401, + 507 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Track Sampling. By comparing the performance of JDT3D on the nuScenes validation set with and without track sampling augmentation, we verify that sampling extra tracks into a training sequence significantly improves the convergence rate and the final performance. The comparison summarized in Tab. 2 shows a $97\\%$ relative improvement on the AMOTA metric from track sampling augmentation compared to no track sampling, with a similar trend across all other metrics, including detection metrics such as the mAP. We also analyzed the effect of varying the number of objects that are sampled by reducing the sampling frequency to 0.5 and 0.25 of the nominal rate. We observe that the AMOTA and mAP metrics plateau after a sampling rate of 0.5, although the AMOTP and IDS metrics see continued improvement with increased track sampling. The benefits of implementing track sampling are expected to generalize to other LiDAR-based TBA methods, but there are currently no available implementations to verify this hypothesis.", + "bbox": [ + 212, + 516, + 787, + 729 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training Clip Length. To validate our choice of clip length, the performances of JDT3D models trained with two to four frame sequences were compared and summarized in Tab. 3. The best performance is achieved when trained over $L = 3$ frames, with a significant jump when moving from $L = 2$ to $L = 3$ . We hypothesize that the training process benefits from the longer-term, temporally consistent information, resulting in improved tracking performance. Moving from", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "B. Cheong et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e848affa980486f41d8d901921bd33e84249aedf941a95b8ae2e0440db64f9f1.jpg", + "table_caption": [ + "Table 3: Ablation on the training sequence length, $L$ , reported on the nuScenes validation set." + ], + "table_footnote": [], + "table_body": "
Clip Length (L)AMOTA ↑AMOTP ↓IDS ↓mAP ↑
20.3971.2944460.214
30.6210.7901500.399
40.5990.7831730.392
", + "bbox": [ + 320, + 184, + 678, + 247 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/473d0e6a7b7837f9d950ccf6200f5cdc45fa7266a8abc43e446fccd8e1aa3445.jpg", + "table_caption": [ + "Table 4: Ablation on the query propagation criteria, reported on the nuScenes [2] validation set. Passing queries with high confidence encourages the re-detection of poor detections and handling of false positives and false negatives." + ], + "table_footnote": [], + "table_body": "
Query PropagationAMOTA↑ AMOTP↓ Recall↑ IDS
Matched to a GT0.4640.9170.51182
High Confidence0.6210.7900.669150
", + "bbox": [ + 305, + 315, + 691, + 364 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "$L = 3$ to $L = 4$ , the performance drops slightly while also significantly increasing the train time and required memory.", + "bbox": [ + 212, + 392, + 782, + 422 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Query Propagation Method. In Tab. 4, we compare different query propagation methods; the first method is passing queries that have been matched to a ground truth [18, 19], and the second is passing based on the $\\tau_{pass}$ threshold for high confidence predictions as described in Sec. 3.2. Results show that using confidence as a filtering criterion during training enhances the overall tracking metrics AMOTA and AMOTP by a large margin and significantly reduces the number of false positives and false negatives while increasing the number of true positives. While the ID switch errors are higher, this increase is less than $2\\%$ relative to the increase in true positives. Thus, we verify that using confidence as the query propagation criterion is more appropriate for the TBA paradigm.", + "bbox": [ + 212, + 443, + 784, + 594 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In addition, we show the generalizability of this query propagation method by performing the same ablation on PF-Track. The results in Tab. 5 show improvements across most metrics, with a similar behaviour of more IDS errors.", + "bbox": [ + 212, + 595, + 782, + 640 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison with TBD Baseline. We compare the performance between our TBA framework and baseline TBD methods such as the CenterPoint greedy-based tracking method [26] and SimpleTrack [20]. To compare our tracking", + "bbox": [ + 212, + 661, + 782, + 708 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/d043737515186980b35145d0e7b4ab5f027d487633915926f2a566f14f23d553.jpg", + "table_caption": [ + "Table 5: An ablation comparing the original ground truth-based propagation and our confidence-based propagation on PF-Track-S [19] evaluated on the nuScenes [2] validation set." + ], + "table_footnote": [], + "table_body": "
Query PropagationAMOTA ↑AMOTP ↓Recall ↑IDS ↓
Matched to a GT0.4081.3430.507166
High Confidence0.4261.3160.558306
", + "bbox": [ + 308, + 787, + 694, + 837 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "JDT3D", + "bbox": [ + 678, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/541a0e67fc43415bfedc611710944d71f37129363778697280bbc4e95100f55c.jpg", + "table_caption": [ + "Table 6: Comparison with \"Tracking by Detection\" on the nuScenes validation set, using the detections generated by JDT3D." + ], + "table_footnote": [], + "table_body": "
Tracking MethodAMOTA ↑AMOTP ↓IDS ↓
CenterPoint [26]0.6080.825491
SimpleTrack [20]0.5730.8775585
JDT3D (Ours)0.6210.790150
", + "bbox": [ + 341, + 184, + 655, + 250 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/8513fbb5476ae0ccc3477c9b0a94ede906d8e38cf4ac907b43ce2481828e6bff.jpg", + "table_caption": [ + "Table 7: Comparison of detection performance between the single-frame pretraining and the three-frame fine-tuning. Evaluated on the seven tracking classes on the nuScenes validation set. The best results are typeset in boldface." + ], + "table_footnote": [], + "table_body": "
JDT3D variantmAP ↑NDS ↑
Single-frame0.6050.549
Three-frame0.3990.502
", + "bbox": [ + 393, + 316, + 602, + 366 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "method to the TBD trackers independent of detection performance, we pass the detections generated by JDT3D to the tracker and evaluate its tracking performance compared to our joint approach. The results in Tab. 6 demonstrate clear advantages of our joint end-to-end approach that offers temporal consistency and data association and reaffirm that the gap in tracking performance lies in the missed detections.", + "bbox": [ + 215, + 393, + 784, + 484 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Detection Performance", + "text_level": 1, + "bbox": [ + 215, + 508, + 450, + 522 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "As mentioned in Sec. 4.2, JDT3D is pre-trained to perform single-frame detection and fine-tuned to perform tracking on sequences of multiple LiDAR frames. While the pre-training step achieves reasonable detection performance, the performance drops significantly after fine-tuning, as seen in Tab. 7. This indicates that the current TBA problem formulation may not be the best alignment of the detection and tracking objectives.", + "bbox": [ + 215, + 534, + 784, + 625 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 Proposal and Track Query Imbalances", + "text_level": 1, + "bbox": [ + 215, + 648, + 575, + 662 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Tab. 8 shows further analysis of the JDT3D outputs. During training, the positively matched proposal queries, or untracked newborn objects, are more confident in the first frame than in subsequent frames. Also, the positively matched track queries, or previously tracked objects, tend to be more confident than the newborn objects. This indicates that the model has a strong bias toward previously tracked objects. This bias could lead to more missed objects and fewer initialized tracks, explaining the high number of false negatives indicated in Tab. 1.", + "bbox": [ + 215, + 672, + 784, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "One hypothesis for this behaviour is that this is caused by the imbalance in training samples assigned to each type of query. Tab. 9 shows the average distribution of true positive predictions of both proposal and track queries in", + "bbox": [ + 215, + 795, + 784, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "B. Cheong et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/89deea2f500553d138f12ca7e963d29b88be34f37eb9cc1874b2cfd1e426fe76.jpg", + "table_caption": [ + "Table 8: The average confidence of positively matched proposal and track predictions over a three-frame training sequence. There are no track queries in the first frame." + ], + "table_footnote": [], + "table_body": "
Avg. ConfidenceFrame 1Frame 2Frame 3
Positively Matched Proposals0.6690.4090.387
Track Predictionsn/a0.7590.760
", + "bbox": [ + 310, + 184, + 691, + 234 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/a7934a657cb4e4410e2b2361fa1524df21db095de0e94605334bfd3b89a4bf07.jpg", + "table_caption": [ + "Table 9: The distribution of positive and negative matches for proposal queries and track queries in the second and third frames of training." + ], + "table_footnote": [], + "table_body": "
PQ PredictionsTQ Predictions
Avg Positive Matches4 (2%)47 (94%)
Avg Negative Matches196 (98%)3 (6%)
", + "bbox": [ + 313, + 287, + 683, + 337 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the second and third frames of training. Note that the proportion of positively matched proposal queries is significantly lower than for track queries. This is somewhat expected since the average track birthrate per frame is 0.7, so most objects in a scene were present in the previous frame. This imbalance in positive training samples between the proposal and track queries could explain the confidence imbalance observed.", + "bbox": [ + 212, + 364, + 784, + 455 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Discussion", + "text_level": 1, + "bbox": [ + 215, + 476, + 354, + 492 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Based on our analysis from Tab. 6 we show that the reason LiDAR-based TBA methods continue to underperform compared to TBD methods is due to poorer detection performance. In addition, we note that there is a significant drop in the detection performance when fine-tuning the model on multiple frames in Tab. 7. In the following sections, we discuss hypotheses to explain the observed results and propose potential improvements to build on JDT3D.", + "bbox": [ + 212, + 507, + 784, + 599 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.1 Temporal Confusion", + "text_level": 1, + "bbox": [ + 215, + 619, + 429, + 633 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our hypothesis for the observed drop in performance in Tab. 7 lies in the fact that the problem formulation of TBA is more complex than query-based detection. In query-based detection, the handling of the queries by the decoder is straightforward: if the query represents an object in the scene, it should output a high-confidence bounding box from the query. On the other hand, in TBA, the handling of queries is conditional: depending on whether the decoder receives a track query that represents the same object, the decoder needs to determine whether the incoming proposal query is a duplicate that should have a low confidence or a unique object that should have a high confidence, as illustrated in Fig. 4. This complex interaction results in \"temporal confusion\" such that the expected model behaviour is different depending on the query inputs. We hypothesize that the reason temporal confusion has not been an issue in vision-based TBA methods [18,19,30] is due to their larger decoder size enabling more", + "bbox": [ + 212, + 643, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "JDT3D", + "bbox": [ + 678, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/996dca4637e4186161c434653079f8c5d4346437748636108c52e20c99fa6d55.jpg", + "image_caption": [ + "Fig. 4: An example of temporal confusion, where the decoder must handle the same proposal query differently based on the presence of track queries. In Case 1, the proposal query should be a positive prediction, while in Case 2, it should be a negative prediction." + ], + "image_footnote": [], + "bbox": [ + 305, + 161, + 697, + 213 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "complex reasoning about the queries. To test this hypothesis, further optimizations would be needed to enable a larger decoder with a LiDAR backbone due to hardware constraints.", + "bbox": [ + 212, + 308, + 784, + 354 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.2 Potential Improvements", + "text_level": 1, + "bbox": [ + 214, + 377, + 460, + 393 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "With these results in mind, we propose two potential directions that can be further explored to bridge the gap between JDT and TBD methods.", + "bbox": [ + 212, + 404, + 784, + 433 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Firstly, based on the ablation performed to evaluate the training clip size, we observed that three-frame training significantly improved performance over two-frame training, although extending to four frames reduced performance slightly. Further exploration of the effect of clip length on tracking performance may yield insights on how to better leverage longer term temporal information.", + "bbox": [ + 212, + 434, + 784, + 508 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Secondly, optimizations to the model size and memory footprint would enable a larger and more complex decoder, that could potentially address the temporal confusion hypothesis.", + "bbox": [ + 212, + 510, + 785, + 555 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 580, + 359, + 595 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper presents JDT3D, a novel LiDAR tracking method that makes significant progress in closing the gap between TBD and TBA methods. JDT3D is trained using track sampling augmentation, a generalizable data augmentation method that injects extra long-term tracks into training samples. We show that track sampling augmentation significantly improves training performance and convergence, and we hope that future LiDAR-based trackers can leverage these findings. As well, we show that the query propagation scheme during training also has a significant impact on the performance, showing that using a confidence threshold to filter low confidence predictions improves tracking performance greatly. Based on our experiments, we hypothesize that the complexity of tracking-by-attention combined with the memory constraints imposed by LiDAR feature extractors limiting the size of transformer decoders is the cause of the drop in detection performance between single and multi-frame variants of JDT3D. We propose further areas of exploration including exploring different training sequence lengths and enabling more decoder layers.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "B. Cheong et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1090-1099 (June 2022)", + "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11618-11628 (2019), https://apisemantic scholar.org/CorpusID: 85517967", + "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European Conference on Computer Vision. pp. 213-229. Springer (2020)", + "4. Chen, Y., Yu, Z., Chen, Y., Lan, S., Anandkumar, A., Jia, J., Alvarez, J.M.: Focalformer3d: Focusing on hard instance for 3d object detection. ArXiv abs/2308.04556 (2023), https://api-semanticscholar.org/CorpusID: 260736073", + "5. Chiu, H.k., Prioletti, A., Li, J., Bohg, J.: Probabilistic 3d multi-object tracking for autonomous driving. arXiv preprint arXiv:2001.05673 (2020)", + "6. Hahner, M., Dai, D., Liniger, A., Gool, L.V.: Quantifying data augmentation for lidar based 3d object detection. CoRR abs/2004.01643 (2020), https://arxiv.org/abs/2004.01643", + "7. Hu, Y., Yang, J., Chen, L., Li, K., Sima, C., Zhu, X., Chai, S., Du, S., Lin, T., Wang, W., et al.: Planning-oriented autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17853-17862 (2023)", + "8. Huang, T.E., Liu, Y., Van Gool, L., Yu, F.: Video task decathlon: Unifying image and video tasks in autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8647-8657 (2023)", + "9. Koh, J., Kim, J., Yoo, J.H., Kim, Y., Kum, D., Choi, J.W.: Joint 3d object detection and tracking using spatio-temporal representation of camera image and lidar point clouds. Proceedings of the AAAI Conference on Artificial Intelligence 36(1), 1210-1218 (Jun 2022). https://doi.org/10.1609/aaai.v36i1.20007, https://ojs.aaai.org/index.php/AAAI/article/view/20007", + "0. Kuhn, H.W.: The hungarian method for the assignment problem. Naval Research Logistics (NRL) 52 (1955), https://api-semanticscholar.org/CorpusID: 9426884", + "1. Law, H., Deng, J.: Cornernet: Detecting objects as paired keypoints. Int. J. Comput. Vis. 128(3), 642-656 (2020). https://doi.org/10.1007/S11263-019-01204-1, https://doi.org/10.1007/s11263-019-01204-1", + "2. Li, X., Xie, T., Liu, D., Gao, J., Dai, K., Jiang, Z., Zhao, L., Wang, K.: Poly-mot: A polyhedral framework for 3d multi-object tracking. ArXiv abs/2307.16675 (2023), https://api.sementicscholar.org/CorpusID:260334560", + "3. Li, Y., Yu, Z., Philion, J., Anandkumar, A., Fidler, S., Jia, J., Alvarez, J.: End-to-end 3d tracking with decoupled queries. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2023)", + "4. Lin, T.Y., Goyal, P., Girshick, R.B., He, K., Dollar, P.: Focal loss for dense object detection. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 2999-3007 (2017), https://api-semanticscholar.org/CorpusID:47252984" + ], + "bbox": [ + 225, + 175, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "JDT3D", + "bbox": [ + 678, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. 2023 IEEE International Conference on Robotics and Automation (ICRA) pp. 2774-2781 (2022), https://apisemantic scholar.org/CorpusID:249097415", + "16. Luo, C., Yang, X., Yuille, A.: Exploring simple 3d multi-object tracking for autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10488-10497 (October 2021)", + "17. Mahmoud, A., Hu, J.S.K., Waslander, S.L.: Dense voxel fusion for 3d object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 663-672 (January 2023)", + "18. Meinhardt, T., Kirillov, A., Leal-Taixe, L., Feichtenhofer, C.: Trackformer: Multiobject tracking with transformers. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8834-8844 (2022). https://doi.org/10.1109/CVPR52688.2022.00864", + "19. Pang, Z., Li, J., Tokmakov, P., Chen, D., Zagoruyko, S., Wang, Y.X.: Standing between past and future: Spatio-temporal modeling for multi-camera 3d multi-object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17928-17938 (June 2023)", + "20. Pang, Z., Li, Z., Wang, N.: SimpleTrack: Understanding and Rethinking 3D Multiobject Tracking. In: Karlinsky, L., Michaeli, T., Nishino, K. (eds.) Computer Vision - ECCV 2022 Workshops. pp. 680-696. Lecture Notes in Computer Science, Springer Nature Switzerland (2021). https://doi.org/10.1007/978-3-031-25056-9_43", + "21. Sun, P., Cao, J., Jiang, Y., Zhang, R., Xie, E., Yuan, Z., Wang, C., Luo, P.: Transtrack: Multiple-object tracking with transformer. arXiv preprint arXiv: 2012.15460 (2020)", + "22. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017), https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf", + "23. Wang, L., Zhang, X.N., Qin, W., Li, X., Yang, L., Li, Z., Zhu, L., Wang, H., Li, J., Liu, H.: Camo-mot: Combined appearance-motion optimization for 3d multi-object tracking with camera-lidar fusion. IEEE Transactions on Intelligent Transportation Systems 24, 11981-11996 (2022), https://api_semanticscholar.org/CorpusID: 252089635", + "24. Weng, X., Wang, J., Held, D., Kitani, K.: 3d multi-object tracking: A baseline and new evaluation metrics. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 10359-10366 (2020). https://doi.org/10.1109/IROS45743.2020.9341164", + "25. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors (Basel, Switzerland) 18 (2018), https://api_semanticscholar.org/CorpusID:52957856", + "26. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11784-11793 (June 2021)", + "27. Zeng, F., Dong, B., Wang, T., Chen, C., Zhang, X., Wei, Y.: Motr: End-to-end multiple-object tracking with transformer. ArXiv abs/2105.03247 (2021), https://api-semanticscholar.org/CorpusID:234096063" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "B. Cheong et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "28. Zeng, Y., Ma, C., Zhu, M., Fan, Z., Yang, X.: Cross-modal 3d object detection and tracking for auto-driving. In: IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2021, Prague, Czech Republic, September 2021. IEEE (2021)", + "29. Zhang, C., Zhang, C., Guo, Y., Chen, L., Happold, M.: Motiontrack: End-to-end transformer-based multi-object tracking with lidar-camera fusion. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) pp. 151-160 (2023), https://apisemanticscholar.org/CorpusID:259286967", + "30. Zhang, T., Chen, X., Wang, Y., Wang, Y., Zhao, H.: Muitr3d: A multi-camera tracking framework via 3d-to-2d queries. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 4536-4545 (2022). https://doi.org/10.1109/CVPRW56347.2022.00500", + "31. Zhang, Y., Wang, T., Zhang, X.: Motrv2: Bootstrapping end-to-end multi-object tracking by pretrained object detectors. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 22056-22065 (2022), https://apisemantic scholar.org/CorpusID:253581846", + "32. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition pp. 4490-4499 (2017), https://apisemantic scholar.org/CorpusID:42427078", + "33. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: deformable transformers for end-to-end object detection. In: 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 479 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "JDT3D", + "bbox": [ + 678, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/JDT3D_ Addressing the Gaps in LiDAR-Based Tracking-by-Attention/b2b98b8d-db63-4b59-ae29-0da5121d4534_model.json b/2024/JDT3D_ Addressing the Gaps in LiDAR-Based Tracking-by-Attention/b2b98b8d-db63-4b59-ae29-0da5121d4534_model.json new file mode 100644 index 0000000000000000000000000000000000000000..57345c8b9e68d9ac121d1fc02b29c933a6003a67 --- /dev/null +++ b/2024/JDT3D_ Addressing the Gaps in LiDAR-Based Tracking-by-Attention/b2b98b8d-db63-4b59-ae29-0da5121d4534_model.json @@ -0,0 +1,2115 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.235, + 0.141, + 0.77, + 0.187 + ], + "angle": 0, + "content": "JDT3D: Addressing the Gaps in LiDAR-Based Tracking-by-Attention" + }, + { + "type": "text", + "bbox": [ + 0.289, + 0.213, + 0.714, + 0.228 + ], + "angle": 0, + "content": "Brian Cheong\\* Jiachen Zhou\\* and Steven Waslander" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.241, + 0.776, + 0.269 + ], + "angle": 0, + "content": "University of Toronto \n{brian.cheong, jason.zhou, steven.waslander}@robotics.utias.utoronto.ca" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.309, + 0.744, + 0.533 + ], + "angle": 0, + "content": "Abstract. Tracking-by-detection (TBD) methods achieve state-of-the-art performance on 3D tracking benchmarks for autonomous driving. On the other hand, tracking-by-attention (TBA) methods have the potential to outperform TBD methods, particularly for long occlusions and challenging detection settings. This work investigates why TBA methods continue to lag in performance behind TBD methods using a LiDAR-based joint detector and tracker called JDT3D. Based on this analysis, we propose two generalizable methods to bridge the gap between TBD and TBA methods: track sampling augmentation and confidence-based query propagation. JDT3D is trained and evaluated on the nuScenes dataset, achieving 0.574 on the AMOTA metric on the nuScenes test set, outperforming all existing LiDAR-based TBA approaches by over \\(6\\%\\). Based on our results, we further discuss some potential challenges with the existing TBA model formulation to explain the continued gap in performance with TBD methods. The implementation of JDT3D can be found at the following link: https://github.com/TRAILab/JDT3D." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.544, + 0.744, + 0.571 + ], + "angle": 0, + "content": "Keywords: Multi-Object Tracking \\(\\cdot\\) Autonomous Vehicles \\(\\cdot\\) Computer Vision" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.6, + 0.377, + 0.615 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.707 + ], + "angle": 0, + "content": "Multi-object tracking (MOT) is an essential component in the perception system of autonomous vehicles, as it allows autonomous agents to reason and plan appropriately in dynamic environments. The task involves identifying object trajectories, requiring that each object is accurately detected and consistently identified over time within the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.708, + 0.788, + 0.784 + ], + "angle": 0, + "content": "The majority of MOT methods adhere to the tracking-by-detection (TBD) paradigm. This paradigm is comprised of two separate processes: generating bounding box predictions without track IDs and associating these predictions to a set of maintained tracks based on some matching criteria. TBD methods currently achieve state-of-the-art results on 3D MOT benchmarks [12, 15, 23]." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.785, + 0.788, + 0.815 + ], + "angle": 0, + "content": "In contrast to the decoupled approach of TBD, joint detection and tracking (JDT) approaches have attempted to unify the detection and tracking tasks in an" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.421, + 0.84 + ], + "angle": 0, + "content": "* Denotes equal contribution." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.382, + 0.129 + ], + "angle": 0, + "content": "B. Cheong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "end-to-end manner [18, 19, 29, 31]. Recent works in autonomous driving perception [7, 17] and computer vision [8] demonstrate that networks trained end-to-end have shown superior performance compared to pipeline-based techniques, despite initially lower performance from earlier works [21, 27]. End-to-end methods can learn richer and more generalizable representations while minimizing the task-specific engineering of individual modules required in a pipeline-based approach. While TBD methods enforce a unidirectional flow of information from the detector to the tracker, the JDT formulation allows both the detector and tracker to exchange useful information to enhance the performance of both tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.287, + 0.789, + 0.468 + ], + "angle": 0, + "content": "Integrating detection and tracking into a joint learning task requires a unique approach to formulating the problem and training the model. One such approach is tracking-by-attention (TBA), illustrated in Fig. 1. In TBA, objects are represented as vector embeddings, or \"queries\", that are used to detect the same object over multiple frames. While recent TBA methods [18,19,29,31] have been proposed as alternatives to TBD, they continue to underperform on the MOT task, especially in the LiDAR domain [29]. However, they have shown potential in certain cases, such as significantly fewer ID-switch errors [19]. The unexplored potential of TBA in the LiDAR domain motivates our work in addressing a key question: what is holding back LiDAR-based TBA? In this work, we propose a LiDAR-based TBA tracking method called JDT3D, based on which we explore different aspects that hold back LiDAR-based TBA." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.789, + 0.563 + ], + "angle": 0, + "content": "One insight is that LiDAR-based tracking methods can suffer from sparse supervision signals due to point cloud and object sparsity. Inspired by object sampling augmentation [25] for LiDAR-based detection methods, we propose track sampling, a temporally consistent augmentation method that injects consistent objects over multiple LiDAR frames to enrich supervision signals while maintaining temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.789, + 0.689 + ], + "angle": 0, + "content": "In addition, existing TBA methods maintain an inconsistency between training and inference regarding their query propagation method. During training, queries that are matched to ground truth samples are propagated to the next frame, but at inference, propagation is based on a confidence threshold. This inconsistency between training and inference could confuse the model to over-trust false positive queries. To address this, we propose a confidence threshold propagation strategy consistent for both training and inference and conduct comprehensive evaluations and discussions on the strategy design." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.692, + 0.558, + 0.707 + ], + "angle": 0, + "content": "We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.711, + 0.799, + 0.757 + ], + "angle": 0, + "content": "JDT3D We present JDT3D, a LiDAR-based tracking-by-attention model that outperforms existing LiDAR-based tracking-by-attention methods on the nuScenes test set by over \\(6\\%\\) on the AMOTA metric." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.761, + 0.789, + 0.792 + ], + "angle": 0, + "content": "Confidence-based query propagation We perform ablations to validate the effectiveness and generalizability of the confidence-based propagation criterion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Track Sampling Augmentation We propose and validate a novel data augmentation technique for multi-frame LiDAR methods that significantly improves performance and convergence rates during training." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "JDT3D" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.147, + 0.71, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.28, + 0.788, + 0.337 + ], + "angle": 0, + "content": "Fig. 1: Illustration of tracking by attention. Each object is represented by a query. In the diagram, the yellow motorcycle leaves the frame, so it is removed from the set of maintained queries. To detect new objects in the scene, such as the tan bus, proposal queries are appended at each time step, represented by the blank squares." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.442 + ], + "angle": 0, + "content": "Analysis Based on our experiments using JDT3D, we provide a detailed analysis to explain the continued gap in performance between LiDAR-based TBD and TBA. We attribute the gap to the reduced detection performance when training on multiple consecutive frames and the handling of queries that results in \"temporal confusion\"." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.462, + 0.396, + 0.479 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.63 + ], + "angle": 0, + "content": "Multi-object tracking (MOT) methods either follow the tracking-by-detection paradigm, in which object detection is independently performed on single frames before the tracking module, or they perform detection and tracking jointly in an end-to-end fashion. MOT can also be further categorized into 2D and 3D MOT. The task of 2D MOT involves predicting object tracks in 2D space, while 3D MOT predicts the tracks in 3D space. 2D MOT methods generally take in camera images as sensor input, while 3D MOT methods may take in camera images, LiDAR point clouds, or both. In this work, we will be focusing our analysis on 3D MOT methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.651, + 0.503, + 0.667 + ], + "angle": 0, + "content": "2.1 Tracking-by-detection (TBD)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.749 + ], + "angle": 0, + "content": "In the tracking-by-detection paradigm, the detection and tracking problems are solved by two distinct modules. First, a single-frame object detector is used to detect objects in a given scene. Next, the detections are associated with a set of maintained tracks by a tracking module and used to update them with current detection information. Unassigned detections are used to initialize new tracks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "While different tracking methods exist among TBD methods, they are all consistently tied to the detector's performance. Several top-performing 3D tracking methods [1,4,15] are strong detectors that have been paired with the simple tracking method of CenterPoint [26]. The formulation of TBD inhibits the detector from taking advantage of temporal object information that would enhance object detection. As well, errors in the detection are guaranteed to be propagated" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.382, + 0.129 + ], + "angle": 0, + "content": "B. Cheong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "downstream to the tracking task. This motivates the exploration of methods that can jointly optimize detection and tracking to achieve MOT." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.196, + 0.563, + 0.212 + ], + "angle": 0, + "content": "2.2 Joint Detection and Tracking (JDT)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.218, + 0.788, + 0.415 + ], + "angle": 0, + "content": "Joint detection and tracking (JDT) is a recent paradigm in both 2D and 3D MOT that involves training a model to perform detection and tracking end-to-end. In contrast to the TBD methods that rely on a dedicated detector to generate detections, JDT methods simultaneously output object detections with associated tracking IDs. Without heuristic matching or hand-crafted rules, SimTrack [16] builds upon CenterPoint [26] by learning a hybrid-time centerness map between two consecutive LiDAR point clouds. AlphaTrack [28] adds a cross-modal fusion scheme to the CenterPoint tracking method [26] and applies an alternating training strategy to balance between the detection branch and appearance affinity. The above two methods perform detection and tracking jointly, but known tracks are not utilized to inform future detections. 3D DetecTrack [9] proposes tracklet-aware 3D detection that reconfigures the initial detection using the latest tracklets' information and employs a spatio-temporal gated GNN for association." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.787, + 0.493 + ], + "angle": 0, + "content": "Tracking-by-Attention (TBA) A growing trend within JDT is the tracking-by-attention (TBA) approach that emerged from TrackFormer [18] for 2D MOT. TBA uses the transformer architecture [22] and formulates tracking as a frame-to-frame set prediction problem that can be trained in an end-to-end manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.672 + ], + "angle": 0, + "content": "TrackFormer [18] extends a 2D query-based detector, DETR [3] and its variant Deformable DETR [33], by designing autoregressive track queries that are passed across frames with associated object identities. MOTRv2 [31] uses a pretrainedYOLOX detector to initialize their object queries and uses the detections to anchor the decoded bounding box positions. In addition, track queries during training are propagated based on the prediction confidence rather than whether a prediction was matched to a ground truth object. This naturally creates false positive and false negative cases during training, removing the need for query injection and removal augmentations [18,27]. JDT3D applies confidence-based track query propagation to LiDAR-based TBA, and we perform ablations showing the method's improved performance over the ground truth matching criterion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.842 + ], + "angle": 0, + "content": "MUTR3D [30] and PF-Track [19] demonstrate the transfer of the tracking-by-attention framework to multi-view vision-only 3D MOT. PF-Track [19] also designs past and future reasoning modules to enhance the spatio-temporal representation of queries. MotionTrack [29] investigates the feasibility of extending the TBA paradigm to LiDAR-based and LiDAR-camera fusion-based 3D JDT by building upon TransFusion [1], a query-based 3D object detector. Rather than implicitly learning which queries should be suppressed through the transformer decoder, MotionTrack learns an explicit affinity matrix between tracks and objects in the scene and performs greedy matching to determine their association. This explicit matching echoes the decoupling found in TBD methods and potentially detracts from the benefits of an end-to-end method. While vision-only" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "JDT3D" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.147, + 0.752, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.318, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Fig. 2: JDT3D Architecture. At each time step, a BEV feature map is extracted and used to initialize a set of proposal queries. The proposal queries are concatenated to track queries passed from the previous frame and used to predict objects in the scene. Track queries detect the same unique objects in each time step, while proposal queries detect untracked or new objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.788, + 0.533 + ], + "angle": 0, + "content": "TBA methods have shown surprising results on the MOT task, the baseline results set by MotionTrack's LiDAR-only and fusion TBA methods underperform compared to LiDAR-based TBD baselines such as CenterPoint [26] and SimpleTrack [20], prompting further exploration of LiDAR TBA methods. Our work explores different potential hypotheses for the reported performance gap between LiDAR-based TBD and TBA methods and proposes improvements to the existing TBA methods." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.58, + 0.499, + 0.594 + ], + "angle": 0, + "content": "2.3 LiDAR Data Augmentations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As with other machine learning tasks, data augmentation can greatly enhance the performance of LiDAR-based 3D object detection methods. Some common augmentations include random flipping, rotation, translation, and scaling of the point cloud [6]. In addition to these, object sampling is also a widely used augmentation method for LiDAR-based object detectors [1,15,26]. First introduced in SECOND [25], object sampling involves creating a database of ground truth object point cloud scans and bounding boxes and randomly sampling and injecting them into the scene during training to improve the convergence and performance overall. However, this augmentation does not apply directly to training TBA methods, since it would result in sporadic track instances in the scene in each training sequence. Our proposed track sampling augmentation samples consecutive track instances and injects them into a sequence of frames during training. This maintains the temporal nature of the object tracks while still offering the same benefits of faster convergence." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.382, + 0.129 + ], + "angle": 0, + "content": "B. Cheong et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.331, + 0.161 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.248 + ], + "angle": 0, + "content": "To perform our analysis, we developed our own baseline LiDAR-based TBA method since no such implementations were made available as of the writing of this paper. We build on aspects from query-based 3D object detection [1] and vision-only TBA [19] to formulate our method, called JDT3D. JDT3D is summarized in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.267, + 0.341, + 0.282 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.288, + 0.787, + 0.349 + ], + "angle": 0, + "content": "At each time step, a feature extractor takes in the LiDAR scan of the scene, \\(\\mathbf{P}\\), and converts it into a BEV feature map representation, \\(\\mathbf{F}\\), using a standard voxel feature extractor [32]. This feature map is used to predict heatmaps of potential object locations and initialize a set of proposal queries, \\(\\mathbf{PQ}\\) [1]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.787, + 0.439 + ], + "angle": 0, + "content": "In the first time step, the proposal queries are passed through a transformer decoder layer with the BEV feature map to obtain a set of updated queries, \\(\\mathbf{Q}^D\\). The past reasoning module refines the updated queries with historical track information [19] before a feedforward network decodes the refined queries, \\(\\mathbf{Q}^R\\) into bounding boxes and class predictions, \\(\\mathbf{B}^R\\). During training, \\(\\mathbf{Q}^D\\) is decoded into a set of bounding boxes and class predictions, \\(\\mathbf{B}^D\\), for an auxiliary loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.787, + 0.5 + ], + "angle": 0, + "content": "To create the track queries for the next frame, the predictions and their corresponding queries are filtered with a confidence threshold. The filtered queries are passed to the future reasoning module [19] which predicts a motion trajectory and projects each query into the next time step." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.787, + 0.544 + ], + "angle": 0, + "content": "At every following time step, new proposal queries are initialized with the heatmap as described above and are concatenated with track queries from the previous time step, \\(\\mathbf{TQ}\\), before being passed to the transformer decoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.545, + 0.787, + 0.636 + ], + "angle": 0, + "content": "To perform tracking, high-confidence predictions and associated queries are assigned an ID. Track queries are used to detect the same object consistently over multiple frames, while proposal queries are used to detect objects to be added to the track query set. Tracking is achieved implicitly through the direct relationship between each prediction and each query passed into the transformer decoder. Trajectories can be obtained by tracing a given track query back in time." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.653, + 0.421, + 0.669 + ], + "angle": 0, + "content": "3.2 Query Propagation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.78 + ], + "angle": 0, + "content": "During training, only the refined queries corresponding to predictions with a confidence score greater than \\(\\tau_{pass}\\) are passed to the next time step. This method of passing and pruning queries has the advantage of naturally creating instances of false negative and false positive track queries during training [31]. False negative track queries are missed tracks that should be re-detected in the next frame as if they were new objects entering the scene. False positive track queries simulate tracks that have left the scene and should be pruned in the next time step." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "This method has the benefit of better matching the inference query propagation behaviour. As well, it creates more instances during training where the decoder must suppress track query predictions and output positive proposal query predictions, ensuring robust handling of track births and deaths." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "JDT3D" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.146, + 0.499, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.147, + 0.685, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.296, + 0.788, + 0.352 + ], + "angle": 0, + "content": "Fig. 3: An example of track sampling augmentation over three consecutive frames from the nuScenes dataset. The original trajectories and sampled trajectories are shown in blue and orange, respectively. Only a subset of the tracks and the third LiDAR scan are shown for visual clarity. The older boxes are shown with transparency." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.383, + 0.482, + 0.399 + ], + "angle": 0, + "content": "3.3 Ground Truth Assignment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.788, + 0.531 + ], + "angle": 0, + "content": "The ground truth assignment of the decoder predictions, \\(\\mathbf{B}_t^D\\), and the refined predictions, \\(\\mathbf{B}_t^R\\), are both done in the same manner using a two-stage assignment process. The first stage involves matching the track queries based on their previously matched ground truth objects [18]. If the previously matched ground truth track is not present in the current time step, such as in the case of a track leaving the scene, then the track query prediction is assigned to no object. False positive track queries were never assigned a ground truth track previously, so they are also assigned to no object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.532, + 0.788, + 0.608 + ], + "angle": 0, + "content": "The second stage of the assignment involves matching the proposal queries to the unmatched ground truths. The unmatched ground truths would include new objects in the scene and missed objects from previous time steps. A Hungarian matcher is used to compute the optimal assignment between the unmatched queries and ground truths based on an association cost [10]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.632, + 0.515, + 0.646 + ], + "angle": 0, + "content": "3.4 Track Sampling Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.764 + ], + "angle": 0, + "content": "To improve the convergence rate and performance, we introduce track sampling augmentation during training. Inspired by the object sampling augmentation [25], our track sampling involves adding additional tracks into a training clip from a database of tracks in a temporally consistent manner, as illustrated in Fig. 3. The original object sampling augmentation does not account for temporal consistency, and applying it naively to our multi-frame training scheme would result in disjointed track instances during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.84 + ], + "angle": 0, + "content": "First, the training set is preprocessed to generate a database of ground truth tracks, including bounding boxes, object classes, track IDs, and LiDAR points contained in the bounding boxes. To perform track sample augmentation on a training clip, a set of tracks is randomly sampled, avoiding track IDs that overlap with the existing ground truth objects in the scene. This enforces unique IDs for" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.382, + 0.129 + ], + "angle": 0, + "content": "B. Cheong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "the tracks in the sequence. To inject a track, \\( L \\) consecutive ground-truth boxes are randomly selected from each sampled track. In each training frame, the LiDAR points corresponding to the sampled objects are used to replace regions in the original point cloud. The objects are injected in chronological order over the \\( L \\) frames. A collision check is performed to ensure that there are no collisions between injected objects and the original ground truth objects in the scene. In the case of a collision, the would-be sampled object is pruned from that frame, simulating an occlusion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.314 + ], + "angle": 0, + "content": "The benefits of track sampling mirror those of object sampling [25], with the extra ground truth objects in the scene providing more samples to learn from in a single frame. This results in a stronger training signal and more varied scenes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.336, + 0.504, + 0.352 + ], + "angle": 0, + "content": "3.5 Network Training and Losses" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.362, + 0.788, + 0.408 + ], + "angle": 0, + "content": "To train JDT3D, \\( L \\) LiDAR frames are passed sequentially to the model as a training sample. The loss for a given frame, \\( \\mathcal{L}_i \\), is computed as the weighted sum of the following losses:" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.42, + 0.786, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i} = \\lambda_ {h} \\mathcal {L} _ {h} + \\lambda_ {f} \\mathcal {L} _ {f} + \\sum_ {m = \\{D, R \\}} \\left(\\lambda_ {r e g} ^ {m} \\mathcal {L} _ {r e g} ^ {m} + \\lambda_ {c l s} ^ {m} \\mathcal {L} _ {c l s} ^ {m}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.57 + ], + "angle": 0, + "content": "The superscripts \\(D\\) and \\(R\\) correspond to predictions output by \\(\\mathbf{Q}^D\\) and \\(\\mathbf{Q}^R\\), respectively. \\(\\mathcal{L}_{reg}^{m}\\) is the L1 loss over the bounding boxes, and \\(\\mathcal{L}_{cls}^{m}\\) is the focal loss [14] over the classification predictions. \\(\\mathcal{L}_h\\) is the Gaussian Focal Loss on the predicted object heatmaps [11,26], and \\(\\mathcal{L}_f\\) is the L1 loss over the motion trajectory predicted by the future reasoning module [19]. Each loss has an associated scalar weight term, \\(\\lambda\\). The total loss is computed as the sum of each individual frame loss:" + }, + { + "type": "equation", + "bbox": [ + 0.452, + 0.57, + 0.786, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c l i p}} = \\sum_ {i = 1} ^ {L} \\mathcal {L} _ {i} \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.631, + 0.376, + 0.649 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.664, + 0.433, + 0.679 + ], + "angle": 0, + "content": "4.1 Dataset and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.779 + ], + "angle": 0, + "content": "JDT3D is trained and evaluated on the nuScenes dataset [2], a multi-modal autonomous driving dataset containing data from six multi-view cameras, five radars, and a 360-degree LiDAR. The dataset contains a total of 40,157 frames across 1000 scenes, of which 850 scenes are annotated for training (700) and validation (150) at a frame rate of \\(2\\mathrm{Hz}\\). We train on the seven object classes annotated for the tracking task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To evaluate JDT3D, the average multi-object tracking accuracy (AMOTA) and average multi-object tracking precision (AMOTP) tracking metrics [24] are computed across the seven tracking classes of nuScenes, evaluating the tracking accuracy and tracking position misalignment, respectively. Secondary metrics are" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.68, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "JDT3D" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.786, + 0.187 + ], + "angle": 0, + "content": "Table 1: A comparison of 3D multi-object tracking methods evaluated on the nuScenes [2] test set. Bold represents the best results among the Transformer-based TBA methods, while Underlined numbers are the best overall." + }, + { + "type": "table", + "bbox": [ + 0.232, + 0.199, + 0.769, + 0.347 + ], + "angle": 0, + "content": "
MethodTBAModalityAMOTA ↑AMOTP ↓FP ↓FN ↓IDS ↓
AB3DMOT [24]L0.1511.50115088757309027
Chui et al. [5]L0.5500.7981753333216950
CenterPoint [26]L0.6380.5551861222928760
SimpleTrack [20]L0.6680.5501751423451575
PF-Track [19]✓C0.4341.2521904842758249
DQTrack [13]✓C0.5231.096--1204
MotionTrack-L [29]✓L0.510.99--9705
MotionTrack-LC [29]✓L+C0.550.871--8716
JDT3D (ours)✓L0.5740.8371115229919254
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.375, + 0.787, + 0.438 + ], + "angle": 0, + "content": "also computed by the nuScenes MOT benchmark [2], such as the number of false positives, false negatives, and identity switches. We also evaluate the detection performance across the tracking classes, with the mean average precision (mAP) and nuScenes Detection Score (NDS) metrics." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.46, + 0.456, + 0.475 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.487, + 0.788, + 0.729 + ], + "angle": 0, + "content": "VoxelNet [25, 32] is used as the LiDAR backbone to extract the BEV feature map. For each frame, the top 200 predictions from the heatmap are used to initialize the proposal queries. The training clip length \\( L \\) is set to 3 frames. The confidence threshold for track queries \\( \\tau_{pass} \\) is set to 0.4 for training and inference. The Hungarian matching cost for ground truth assignment consists of the sum of the L1 cost on the predicted box and the focal loss [14] on the predicted class confidence. All detections are generated in the car ego-frame, so the object centers of the track queries are updated to their relative positions in the current training frame based on the car's ego-motion between time steps. A pretrained VoxelNet backbone from BEVFusion [15] is used to initialize the feature extractor. To train JDT3D, the model is first trained on the detection task using \\( L = 1 \\) for 20 epochs. It is then further trained on the tracking task using \\( L = 3 \\) for another 20 epochs. All experiments were trained and evaluated on eight NVIDIA V100 GPUs using random flipping, rotation, translation, and scaling on the training sequence, and track sampling augmentation, unless otherwise specified. JDT3D performs inference at \\( 3.85\\mathrm{Hz} \\) on a Tesla V100." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.753, + 0.373, + 0.767 + ], + "angle": 0, + "content": "4.3 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Our main results compare JDT3D with baseline TBD methods and relevant TBA methods evaluated on the nuScenes test set [2], summarized in Tab. 1. JDT3D outperforms the current state-of-the-art LiDAR-based TBA methods, with a relative improvement of \\(12.5\\%\\) on the AMOTA metric and \\(15.5\\%\\) on the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.382, + 0.129 + ], + "angle": 0, + "content": "B. Cheong et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 2: Ablation on the track sampling augmentation, reported on the nuScenes validation set. The mAP metric is computed on the seven tracking categories." + }, + { + "type": "table", + "bbox": [ + 0.304, + 0.185, + 0.699, + 0.264 + ], + "angle": 0, + "content": "
Track Sampling RateAMOTA ↑AMOTP ↓IDS ↓mAP ↑
0.00.3141.3773880.225
0.250.6000.8642100.3867
0.500.6220.8162030.400
1.00.6210.7901500.399
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.29, + 0.784, + 0.379 + ], + "angle": 0, + "content": "AMOTP metric on the test split. Moreover, with LiDAR input only, JDT3D also outperforms the fusion-based MotionTrack-LC [29] with a relative improvement of \\(4.36\\%\\) on AMOTA and \\(3.79\\%\\) on AMOTP on the test set. While JDT3D underperforms compared to baseline TBD methods on most metrics, it achieves the lowest number of ID switches among LiDAR-only methods, indicating a strong track coherence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.785, + 0.471 + ], + "angle": 0, + "content": "JDT3D has significantly more false negatives compared to TBD methods, but fewer false positives. The potential sources of false negatives are either due to missed detections or tracks diverging from the ground truths. However, the latter case would result in both high false negatives and false positives. Based on the low false positive results from JDT3D, we conclude that the false negatives are mainly due to missed detections." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.493, + 0.403, + 0.508 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.788, + 0.73 + ], + "angle": 0, + "content": "Track Sampling. By comparing the performance of JDT3D on the nuScenes validation set with and without track sampling augmentation, we verify that sampling extra tracks into a training sequence significantly improves the convergence rate and the final performance. The comparison summarized in Tab. 2 shows a \\(97\\%\\) relative improvement on the AMOTA metric from track sampling augmentation compared to no track sampling, with a similar trend across all other metrics, including detection metrics such as the mAP. We also analyzed the effect of varying the number of objects that are sampled by reducing the sampling frequency to 0.5 and 0.25 of the nominal rate. We observe that the AMOTA and mAP metrics plateau after a sampling rate of 0.5, although the AMOTP and IDS metrics see continued improvement with increased track sampling. The benefits of implementing track sampling are expected to generalize to other LiDAR-based TBA methods, but there are currently no available implementations to verify this hypothesis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Training Clip Length. To validate our choice of clip length, the performances of JDT3D models trained with two to four frame sequences were compared and summarized in Tab. 3. The best performance is achieved when trained over \\( L = 3 \\) frames, with a significant jump when moving from \\( L = 2 \\) to \\( L = 3 \\). We hypothesize that the training process benefits from the longer-term, temporally consistent information, resulting in improved tracking performance. Moving from" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "JDT3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.784, + 0.172 + ], + "angle": 0, + "content": "Table 3: Ablation on the training sequence length, \\( L \\), reported on the nuScenes validation set." + }, + { + "type": "table", + "bbox": [ + 0.321, + 0.185, + 0.679, + 0.248 + ], + "angle": 0, + "content": "
Clip Length (L)AMOTA ↑AMOTP ↓IDS ↓mAP ↑
20.3971.2944460.214
30.6210.7901500.399
40.5990.7831730.392
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.261, + 0.785, + 0.303 + ], + "angle": 0, + "content": "Table 4: Ablation on the query propagation criteria, reported on the nuScenes [2] validation set. Passing queries with high confidence encourages the re-detection of poor detections and handling of false positives and false negatives." + }, + { + "type": "table", + "bbox": [ + 0.307, + 0.316, + 0.692, + 0.365 + ], + "angle": 0, + "content": "
Query PropagationAMOTA↑ AMOTP↓ Recall↑ IDS
Matched to a GT0.4640.9170.51182
High Confidence0.6210.7900.669150
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.784, + 0.423 + ], + "angle": 0, + "content": "\\(L = 3\\) to \\(L = 4\\), the performance drops slightly while also significantly increasing the train time and required memory." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.785, + 0.595 + ], + "angle": 0, + "content": "Query Propagation Method. In Tab. 4, we compare different query propagation methods; the first method is passing queries that have been matched to a ground truth [18, 19], and the second is passing based on the \\(\\tau_{pass}\\) threshold for high confidence predictions as described in Sec. 3.2. Results show that using confidence as a filtering criterion during training enhances the overall tracking metrics AMOTA and AMOTP by a large margin and significantly reduces the number of false positives and false negatives while increasing the number of true positives. While the ID switch errors are higher, this increase is less than \\(2\\%\\) relative to the increase in true positives. Thus, we verify that using confidence as the query propagation criterion is more appropriate for the TBA paradigm." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.784, + 0.641 + ], + "angle": 0, + "content": "In addition, we show the generalizability of this query propagation method by performing the same ablation on PF-Track. The results in Tab. 5 show improvements across most metrics, with a similar behaviour of more IDS errors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.662, + 0.784, + 0.709 + ], + "angle": 0, + "content": "Comparison with TBD Baseline. We compare the performance between our TBA framework and baseline TBD methods such as the CenterPoint greedy-based tracking method [26] and SimpleTrack [20]. To compare our tracking" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.734, + 0.784, + 0.774 + ], + "angle": 0, + "content": "Table 5: An ablation comparing the original ground truth-based propagation and our confidence-based propagation on PF-Track-S [19] evaluated on the nuScenes [2] validation set." + }, + { + "type": "table", + "bbox": [ + 0.31, + 0.788, + 0.695, + 0.838 + ], + "angle": 0, + "content": "
Query PropagationAMOTA ↑AMOTP ↓Recall ↑IDS ↓
Matched to a GT0.4081.3430.507166
High Confidence0.4261.3160.558306
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.382, + 0.129 + ], + "angle": 0, + "content": "B. Cheong et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 6: Comparison with \"Tracking by Detection\" on the nuScenes validation set, using the detections generated by JDT3D." + }, + { + "type": "table", + "bbox": [ + 0.342, + 0.185, + 0.656, + 0.25 + ], + "angle": 0, + "content": "
Tracking MethodAMOTA ↑AMOTP ↓IDS ↓
CenterPoint [26]0.6080.825491
SimpleTrack [20]0.5730.8775585
JDT3D (Ours)0.6210.790150
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.263, + 0.785, + 0.304 + ], + "angle": 0, + "content": "Table 7: Comparison of detection performance between the single-frame pretraining and the three-frame fine-tuning. Evaluated on the seven tracking classes on the nuScenes validation set. The best results are typeset in boldface." + }, + { + "type": "table", + "bbox": [ + 0.395, + 0.317, + 0.603, + 0.367 + ], + "angle": 0, + "content": "
JDT3D variantmAP ↑NDS ↑
Single-frame0.6050.549
Three-frame0.3990.502
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.395, + 0.785, + 0.485 + ], + "angle": 0, + "content": "method to the TBD trackers independent of detection performance, we pass the detections generated by JDT3D to the tracker and evaluate its tracking performance compared to our joint approach. The results in Tab. 6 demonstrate clear advantages of our joint end-to-end approach that offers temporal consistency and data association and reaffirm that the gap in tracking performance lies in the missed detections." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.509, + 0.452, + 0.523 + ], + "angle": 0, + "content": "4.5 Detection Performance" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.535, + 0.785, + 0.625 + ], + "angle": 0, + "content": "As mentioned in Sec. 4.2, JDT3D is pre-trained to perform single-frame detection and fine-tuned to perform tracking on sequences of multiple LiDAR frames. While the pre-training step achieves reasonable detection performance, the performance drops significantly after fine-tuning, as seen in Tab. 7. This indicates that the current TBA problem formulation may not be the best alignment of the detection and tracking objectives." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.649, + 0.576, + 0.664 + ], + "angle": 0, + "content": "4.6 Proposal and Track Query Imbalances" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.674, + 0.785, + 0.794 + ], + "angle": 0, + "content": "Tab. 8 shows further analysis of the JDT3D outputs. During training, the positively matched proposal queries, or untracked newborn objects, are more confident in the first frame than in subsequent frames. Also, the positively matched track queries, or previously tracked objects, tend to be more confident than the newborn objects. This indicates that the model has a strong bias toward previously tracked objects. This bias could lead to more missed objects and fewer initialized tracks, explaining the high number of false negatives indicated in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.796, + 0.785, + 0.84 + ], + "angle": 0, + "content": "One hypothesis for this behaviour is that this is caused by the imbalance in training samples assigned to each type of query. Tab. 9 shows the average distribution of true positive predictions of both proposal and track queries in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "JDT3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 8: The average confidence of positively matched proposal and track predictions over a three-frame training sequence. There are no track queries in the first frame." + }, + { + "type": "table", + "bbox": [ + 0.312, + 0.185, + 0.692, + 0.235 + ], + "angle": 0, + "content": "
Avg. ConfidenceFrame 1Frame 2Frame 3
Positively Matched Proposals0.6690.4090.387
Track Predictionsn/a0.7590.760
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.247, + 0.785, + 0.276 + ], + "angle": 0, + "content": "Table 9: The distribution of positive and negative matches for proposal queries and track queries in the second and third frames of training." + }, + { + "type": "table", + "bbox": [ + 0.315, + 0.288, + 0.684, + 0.338 + ], + "angle": 0, + "content": "
PQ PredictionsTQ Predictions
Avg Positive Matches4 (2%)47 (94%)
Avg Negative Matches196 (98%)3 (6%)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.785, + 0.456 + ], + "angle": 0, + "content": "the second and third frames of training. Note that the proportion of positively matched proposal queries is significantly lower than for track queries. This is somewhat expected since the average track birthrate per frame is 0.7, so most objects in a scene were present in the previous frame. This imbalance in positive training samples between the proposal and track queries could explain the confidence imbalance observed." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.477, + 0.355, + 0.493 + ], + "angle": 0, + "content": "5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.785, + 0.6 + ], + "angle": 0, + "content": "Based on our analysis from Tab. 6 we show that the reason LiDAR-based TBA methods continue to underperform compared to TBD methods is due to poorer detection performance. In addition, we note that there is a significant drop in the detection performance when fine-tuning the model on multiple frames in Tab. 7. In the following sections, we discuss hypotheses to explain the observed results and propose potential improvements to build on JDT3D." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.62, + 0.431, + 0.635 + ], + "angle": 0, + "content": "5.1 Temporal Confusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Our hypothesis for the observed drop in performance in Tab. 7 lies in the fact that the problem formulation of TBA is more complex than query-based detection. In query-based detection, the handling of the queries by the decoder is straightforward: if the query represents an object in the scene, it should output a high-confidence bounding box from the query. On the other hand, in TBA, the handling of queries is conditional: depending on whether the decoder receives a track query that represents the same object, the decoder needs to determine whether the incoming proposal query is a duplicate that should have a low confidence or a unique object that should have a high confidence, as illustrated in Fig. 4. This complex interaction results in \"temporal confusion\" such that the expected model behaviour is different depending on the query inputs. We hypothesize that the reason temporal confusion has not been an issue in vision-based TBA methods [18,19,30] is due to their larger decoder size enabling more" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.382, + 0.129 + ], + "angle": 0, + "content": "B. Cheong et al." + }, + { + "type": "image", + "bbox": [ + 0.307, + 0.162, + 0.699, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.236, + 0.788, + 0.28 + ], + "angle": 0, + "content": "Fig. 4: An example of temporal confusion, where the decoder must handle the same proposal query differently based on the presence of track queries. In Case 1, the proposal query should be a positive prediction, while in Case 2, it should be a negative prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.309, + 0.785, + 0.355 + ], + "angle": 0, + "content": "complex reasoning about the queries. To test this hypothesis, further optimizations would be needed to enable a larger decoder with a LiDAR backbone due to hardware constraints." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.378, + 0.461, + 0.394 + ], + "angle": 0, + "content": "5.2 Potential Improvements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.785, + 0.434 + ], + "angle": 0, + "content": "With these results in mind, we propose two potential directions that can be further explored to bridge the gap between JDT and TBD methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.785, + 0.51 + ], + "angle": 0, + "content": "Firstly, based on the ablation performed to evaluate the training clip size, we observed that three-frame training significantly improved performance over two-frame training, although extending to four frames reduced performance slightly. Further exploration of the effect of clip length on tracking performance may yield insights on how to better leverage longer term temporal information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.787, + 0.556 + ], + "angle": 0, + "content": "Secondly, optimizations to the model size and memory footprint would enable a larger and more complex decoder, that could potentially address the temporal confusion hypothesis." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.581, + 0.36, + 0.597 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This paper presents JDT3D, a novel LiDAR tracking method that makes significant progress in closing the gap between TBD and TBA methods. JDT3D is trained using track sampling augmentation, a generalizable data augmentation method that injects extra long-term tracks into training samples. We show that track sampling augmentation significantly improves training performance and convergence, and we hope that future LiDAR-based trackers can leverage these findings. As well, we show that the query propagation scheme during training also has a significant impact on the performance, showing that using a confidence threshold to filter low confidence predictions improves tracking performance greatly. Based on our experiments, we hypothesize that the complexity of tracking-by-attention combined with the memory constraints imposed by LiDAR feature extractors limiting the size of transformer decoders is the cause of the drop in detection performance between single and multi-frame variants of JDT3D. We propose further areas of exploration including exploring different training sequence lengths and enabling more decoder layers." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "JDT3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.786, + 0.232 + ], + "angle": 0, + "content": "1. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1090-1099 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.786, + 0.3 + ], + "angle": 0, + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11618-11628 (2019), https://apisemantic scholar.org/CorpusID: 85517967" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.301, + 0.786, + 0.342 + ], + "angle": 0, + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European Conference on Computer Vision. pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.786, + 0.396 + ], + "angle": 0, + "content": "4. Chen, Y., Yu, Z., Chen, Y., Lan, S., Anandkumar, A., Jia, J., Alvarez, J.M.: Focalformer3d: Focusing on hard instance for 3d object detection. ArXiv abs/2308.04556 (2023), https://api-semanticscholar.org/CorpusID: 260736073" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.786, + 0.425 + ], + "angle": 0, + "content": "5. Chiu, H.k., Prioletti, A., Li, J., Bohg, J.: Probabilistic 3d multi-object tracking for autonomous driving. arXiv preprint arXiv:2001.05673 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.786, + 0.467 + ], + "angle": 0, + "content": "6. Hahner, M., Dai, D., Liniger, A., Gool, L.V.: Quantifying data augmentation for lidar based 3d object detection. CoRR abs/2004.01643 (2020), https://arxiv.org/abs/2004.01643" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.786, + 0.522 + ], + "angle": 0, + "content": "7. Hu, Y., Yang, J., Chen, L., Li, K., Sima, C., Zhu, X., Chai, S., Du, S., Lin, T., Wang, W., et al.: Planning-oriented autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17853-17862 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.786, + 0.564 + ], + "angle": 0, + "content": "8. Huang, T.E., Liu, Y., Van Gool, L., Yu, F.: Video task decathlon: Unifying image and video tasks in autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8647-8657 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.786, + 0.632 + ], + "angle": 0, + "content": "9. Koh, J., Kim, J., Yoo, J.H., Kim, Y., Kum, D., Choi, J.W.: Joint 3d object detection and tracking using spatio-temporal representation of camera image and lidar point clouds. Proceedings of the AAAI Conference on Artificial Intelligence 36(1), 1210-1218 (Jun 2022). https://doi.org/10.1609/aaai.v36i1.20007, https://ojs.aaai.org/index.php/AAAI/article/view/20007" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.786, + 0.673 + ], + "angle": 0, + "content": "0. Kuhn, H.W.: The hungarian method for the assignment problem. Naval Research Logistics (NRL) 52 (1955), https://api-semanticscholar.org/CorpusID: 9426884" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.786, + 0.715 + ], + "angle": 0, + "content": "1. Law, H., Deng, J.: Cornernet: Detecting objects as paired keypoints. Int. J. Comput. Vis. 128(3), 642-656 (2020). https://doi.org/10.1007/S11263-019-01204-1, https://doi.org/10.1007/s11263-019-01204-1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.786, + 0.757 + ], + "angle": 0, + "content": "2. Li, X., Xie, T., Liu, D., Gao, J., Dai, K., Jiang, Z., Zhao, L., Wang, K.: Poly-mot: A polyhedral framework for 3d multi-object tracking. ArXiv abs/2307.16675 (2023), https://api.sementicscholar.org/CorpusID:260334560" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.786, + 0.799 + ], + "angle": 0, + "content": "3. Li, Y., Yu, Z., Philion, J., Anandkumar, A., Fidler, S., Jia, J., Alvarez, J.: End-to-end 3d tracking with decoupled queries. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "4. Lin, T.Y., Goyal, P., Girshick, R.B., He, K., Dollar, P.: Focal loss for dense object detection. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 2999-3007 (2017), https://api-semanticscholar.org/CorpusID:47252984" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.176, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.382, + 0.129 + ], + "angle": 0, + "content": "B. Cheong et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "15. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. 2023 IEEE International Conference on Robotics and Automation (ICRA) pp. 2774-2781 (2022), https://apisemantic scholar.org/CorpusID:249097415" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.787, + 0.245 + ], + "angle": 0, + "content": "16. Luo, C., Yang, X., Yuille, A.: Exploring simple 3d multi-object tracking for autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10488-10497 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.787, + 0.286 + ], + "angle": 0, + "content": "17. Mahmoud, A., Hu, J.S.K., Waslander, S.L.: Dense voxel fusion for 3d object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 663-672 (January 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.287, + 0.787, + 0.342 + ], + "angle": 0, + "content": "18. Meinhardt, T., Kirillov, A., Leal-Taixe, L., Feichtenhofer, C.: Trackformer: Multiobject tracking with transformers. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8834-8844 (2022). https://doi.org/10.1109/CVPR52688.2022.00864" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.343, + 0.787, + 0.398 + ], + "angle": 0, + "content": "19. Pang, Z., Li, J., Tokmakov, P., Chen, D., Zagoruyko, S., Wang, Y.X.: Standing between past and future: Spatio-temporal modeling for multi-camera 3d multi-object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17928-17938 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.399, + 0.787, + 0.466 + ], + "angle": 0, + "content": "20. Pang, Z., Li, Z., Wang, N.: SimpleTrack: Understanding and Rethinking 3D Multiobject Tracking. In: Karlinsky, L., Michaeli, T., Nishino, K. (eds.) Computer Vision - ECCV 2022 Workshops. pp. 680-696. Lecture Notes in Computer Science, Springer Nature Switzerland (2021). https://doi.org/10.1007/978-3-031-25056-9_43" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.508 + ], + "angle": 0, + "content": "21. Sun, P., Cao, J., Jiang, Y., Zhang, R., Xie, E., Yuan, Z., Wang, C., Luo, P.: Transtrack: Multiple-object tracking with transformer. arXiv preprint arXiv: 2012.15460 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.591 + ], + "angle": 0, + "content": "22. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017), https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.787, + 0.659 + ], + "angle": 0, + "content": "23. Wang, L., Zhang, X.N., Qin, W., Li, X., Yang, L., Li, Z., Zhu, L., Wang, H., Li, J., Liu, H.: Camo-mot: Combined appearance-motion optimization for 3d multi-object tracking with camera-lidar fusion. IEEE Transactions on Intelligent Transportation Systems 24, 11981-11996 (2022), https://api_semanticscholar.org/CorpusID: 252089635" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.787, + 0.715 + ], + "angle": 0, + "content": "24. Weng, X., Wang, J., Held, D., Kitani, K.: 3d multi-object tracking: A baseline and new evaluation metrics. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 10359-10366 (2020). https://doi.org/10.1109/IROS45743.2020.9341164" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.757 + ], + "angle": 0, + "content": "25. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors (Basel, Switzerland) 18 (2018), https://api_semanticscholar.org/CorpusID:52957856" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "26. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11784-11793 (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "27. Zeng, F., Dong, B., Wang, T., Chen, C., Zhang, X., Wei, Y.: Motr: End-to-end multiple-object tracking with transformer. ArXiv abs/2105.03247 (2021), https://api-semanticscholar.org/CorpusID:234096063" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.679, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "JDT3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "28. Zeng, Y., Ma, C., Zhu, M., Fan, Z., Yang, X.: Cross-modal 3d object detection and tracking for auto-driving. In: IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2021, Prague, Czech Republic, September 2021. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.788, + 0.259 + ], + "angle": 0, + "content": "29. Zhang, C., Zhang, C., Guo, Y., Chen, L., Happold, M.: Motiontrack: End-to-end transformer-based multi-object tracking with lidar-camera fusion. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) pp. 151-160 (2023), https://apisemanticscholar.org/CorpusID:259286967" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.26, + 0.788, + 0.313 + ], + "angle": 0, + "content": "30. Zhang, T., Chen, X., Wang, Y., Wang, Y., Zhao, H.: Muitr3d: A multi-camera tracking framework via 3d-to-2d queries. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 4536-4545 (2022). https://doi.org/10.1109/CVPRW56347.2022.00500" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.369 + ], + "angle": 0, + "content": "31. Zhang, Y., Wang, T., Zhang, X.: Motrv2: Bootstrapping end-to-end multi-object tracking by pretrained object detectors. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 22056-22065 (2022), https://apisemantic scholar.org/CorpusID:253581846" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.424 + ], + "angle": 0, + "content": "32. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition pp. 4490-4499 (2017), https://apisemantic scholar.org/CorpusID:42427078" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.481 + ], + "angle": 0, + "content": "33. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: deformable transformers for end-to-end object detection. In: 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.481 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/JDT3D_ Addressing the Gaps in LiDAR-Based Tracking-by-Attention/b2b98b8d-db63-4b59-ae29-0da5121d4534_origin.pdf b/2024/JDT3D_ Addressing the Gaps in LiDAR-Based Tracking-by-Attention/b2b98b8d-db63-4b59-ae29-0da5121d4534_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..98ee11250de453914e28a2310459b4d7a028cca2 --- /dev/null +++ b/2024/JDT3D_ Addressing the Gaps in LiDAR-Based Tracking-by-Attention/b2b98b8d-db63-4b59-ae29-0da5121d4534_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8678063e0349238d42259960829db8c9542288ac548dcb70afe44ac3acc2f8ac +size 2274114 diff --git a/2024/JDT3D_ Addressing the Gaps in LiDAR-Based Tracking-by-Attention/full.md b/2024/JDT3D_ Addressing the Gaps in LiDAR-Based Tracking-by-Attention/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ae8adf9847015741226eef610ccf05d0f4160970 --- /dev/null +++ b/2024/JDT3D_ Addressing the Gaps in LiDAR-Based Tracking-by-Attention/full.md @@ -0,0 +1,276 @@ +# JDT3D: Addressing the Gaps in LiDAR-Based Tracking-by-Attention + +Brian Cheong\* Jiachen Zhou\* and Steven Waslander + +University of Toronto +{brian.cheong, jason.zhou, steven.waslander}@robotics.utias.utoronto.ca + +Abstract. Tracking-by-detection (TBD) methods achieve state-of-the-art performance on 3D tracking benchmarks for autonomous driving. On the other hand, tracking-by-attention (TBA) methods have the potential to outperform TBD methods, particularly for long occlusions and challenging detection settings. This work investigates why TBA methods continue to lag in performance behind TBD methods using a LiDAR-based joint detector and tracker called JDT3D. Based on this analysis, we propose two generalizable methods to bridge the gap between TBD and TBA methods: track sampling augmentation and confidence-based query propagation. JDT3D is trained and evaluated on the nuScenes dataset, achieving 0.574 on the AMOTA metric on the nuScenes test set, outperforming all existing LiDAR-based TBA approaches by over $6\%$ . Based on our results, we further discuss some potential challenges with the existing TBA model formulation to explain the continued gap in performance with TBD methods. The implementation of JDT3D can be found at the following link: https://github.com/TRAILab/JDT3D. + +Keywords: Multi-Object Tracking $\cdot$ Autonomous Vehicles $\cdot$ Computer Vision + +# 1 Introduction + +Multi-object tracking (MOT) is an essential component in the perception system of autonomous vehicles, as it allows autonomous agents to reason and plan appropriately in dynamic environments. The task involves identifying object trajectories, requiring that each object is accurately detected and consistently identified over time within the scene. + +The majority of MOT methods adhere to the tracking-by-detection (TBD) paradigm. This paradigm is comprised of two separate processes: generating bounding box predictions without track IDs and associating these predictions to a set of maintained tracks based on some matching criteria. TBD methods currently achieve state-of-the-art results on 3D MOT benchmarks [12, 15, 23]. + +In contrast to the decoupled approach of TBD, joint detection and tracking (JDT) approaches have attempted to unify the detection and tracking tasks in an + +end-to-end manner [18, 19, 29, 31]. Recent works in autonomous driving perception [7, 17] and computer vision [8] demonstrate that networks trained end-to-end have shown superior performance compared to pipeline-based techniques, despite initially lower performance from earlier works [21, 27]. End-to-end methods can learn richer and more generalizable representations while minimizing the task-specific engineering of individual modules required in a pipeline-based approach. While TBD methods enforce a unidirectional flow of information from the detector to the tracker, the JDT formulation allows both the detector and tracker to exchange useful information to enhance the performance of both tasks. + +Integrating detection and tracking into a joint learning task requires a unique approach to formulating the problem and training the model. One such approach is tracking-by-attention (TBA), illustrated in Fig. 1. In TBA, objects are represented as vector embeddings, or "queries", that are used to detect the same object over multiple frames. While recent TBA methods [18,19,29,31] have been proposed as alternatives to TBD, they continue to underperform on the MOT task, especially in the LiDAR domain [29]. However, they have shown potential in certain cases, such as significantly fewer ID-switch errors [19]. The unexplored potential of TBA in the LiDAR domain motivates our work in addressing a key question: what is holding back LiDAR-based TBA? In this work, we propose a LiDAR-based TBA tracking method called JDT3D, based on which we explore different aspects that hold back LiDAR-based TBA. + +One insight is that LiDAR-based tracking methods can suffer from sparse supervision signals due to point cloud and object sparsity. Inspired by object sampling augmentation [25] for LiDAR-based detection methods, we propose track sampling, a temporally consistent augmentation method that injects consistent objects over multiple LiDAR frames to enrich supervision signals while maintaining temporal consistency. + +In addition, existing TBA methods maintain an inconsistency between training and inference regarding their query propagation method. During training, queries that are matched to ground truth samples are propagated to the next frame, but at inference, propagation is based on a confidence threshold. This inconsistency between training and inference could confuse the model to over-trust false positive queries. To address this, we propose a confidence threshold propagation strategy consistent for both training and inference and conduct comprehensive evaluations and discussions on the strategy design. + +We summarize our contributions as follows: + +JDT3D We present JDT3D, a LiDAR-based tracking-by-attention model that outperforms existing LiDAR-based tracking-by-attention methods on the nuScenes test set by over $6\%$ on the AMOTA metric. + +Confidence-based query propagation We perform ablations to validate the effectiveness and generalizability of the confidence-based propagation criterion. + +Track Sampling Augmentation We propose and validate a novel data augmentation technique for multi-frame LiDAR methods that significantly improves performance and convergence rates during training. + +![](images/901813337f9bd5a1067850121f1a0273eb4f80e96490e461851e5b3a37ae91f1.jpg) +Fig. 1: Illustration of tracking by attention. Each object is represented by a query. In the diagram, the yellow motorcycle leaves the frame, so it is removed from the set of maintained queries. To detect new objects in the scene, such as the tan bus, proposal queries are appended at each time step, represented by the blank squares. + +Analysis Based on our experiments using JDT3D, we provide a detailed analysis to explain the continued gap in performance between LiDAR-based TBD and TBA. We attribute the gap to the reduced detection performance when training on multiple consecutive frames and the handling of queries that results in "temporal confusion". + +# 2 Related Works + +Multi-object tracking (MOT) methods either follow the tracking-by-detection paradigm, in which object detection is independently performed on single frames before the tracking module, or they perform detection and tracking jointly in an end-to-end fashion. MOT can also be further categorized into 2D and 3D MOT. The task of 2D MOT involves predicting object tracks in 2D space, while 3D MOT predicts the tracks in 3D space. 2D MOT methods generally take in camera images as sensor input, while 3D MOT methods may take in camera images, LiDAR point clouds, or both. In this work, we will be focusing our analysis on 3D MOT methods. + +# 2.1 Tracking-by-detection (TBD) + +In the tracking-by-detection paradigm, the detection and tracking problems are solved by two distinct modules. First, a single-frame object detector is used to detect objects in a given scene. Next, the detections are associated with a set of maintained tracks by a tracking module and used to update them with current detection information. Unassigned detections are used to initialize new tracks. + +While different tracking methods exist among TBD methods, they are all consistently tied to the detector's performance. Several top-performing 3D tracking methods [1,4,15] are strong detectors that have been paired with the simple tracking method of CenterPoint [26]. The formulation of TBD inhibits the detector from taking advantage of temporal object information that would enhance object detection. As well, errors in the detection are guaranteed to be propagated + +downstream to the tracking task. This motivates the exploration of methods that can jointly optimize detection and tracking to achieve MOT. + +# 2.2 Joint Detection and Tracking (JDT) + +Joint detection and tracking (JDT) is a recent paradigm in both 2D and 3D MOT that involves training a model to perform detection and tracking end-to-end. In contrast to the TBD methods that rely on a dedicated detector to generate detections, JDT methods simultaneously output object detections with associated tracking IDs. Without heuristic matching or hand-crafted rules, SimTrack [16] builds upon CenterPoint [26] by learning a hybrid-time centerness map between two consecutive LiDAR point clouds. AlphaTrack [28] adds a cross-modal fusion scheme to the CenterPoint tracking method [26] and applies an alternating training strategy to balance between the detection branch and appearance affinity. The above two methods perform detection and tracking jointly, but known tracks are not utilized to inform future detections. 3D DetecTrack [9] proposes tracklet-aware 3D detection that reconfigures the initial detection using the latest tracklets' information and employs a spatio-temporal gated GNN for association. + +Tracking-by-Attention (TBA) A growing trend within JDT is the tracking-by-attention (TBA) approach that emerged from TrackFormer [18] for 2D MOT. TBA uses the transformer architecture [22] and formulates tracking as a frame-to-frame set prediction problem that can be trained in an end-to-end manner. + +TrackFormer [18] extends a 2D query-based detector, DETR [3] and its variant Deformable DETR [33], by designing autoregressive track queries that are passed across frames with associated object identities. MOTRv2 [31] uses a pretrainedYOLOX detector to initialize their object queries and uses the detections to anchor the decoded bounding box positions. In addition, track queries during training are propagated based on the prediction confidence rather than whether a prediction was matched to a ground truth object. This naturally creates false positive and false negative cases during training, removing the need for query injection and removal augmentations [18,27]. JDT3D applies confidence-based track query propagation to LiDAR-based TBA, and we perform ablations showing the method's improved performance over the ground truth matching criterion. + +MUTR3D [30] and PF-Track [19] demonstrate the transfer of the tracking-by-attention framework to multi-view vision-only 3D MOT. PF-Track [19] also designs past and future reasoning modules to enhance the spatio-temporal representation of queries. MotionTrack [29] investigates the feasibility of extending the TBA paradigm to LiDAR-based and LiDAR-camera fusion-based 3D JDT by building upon TransFusion [1], a query-based 3D object detector. Rather than implicitly learning which queries should be suppressed through the transformer decoder, MotionTrack learns an explicit affinity matrix between tracks and objects in the scene and performs greedy matching to determine their association. This explicit matching echoes the decoupling found in TBD methods and potentially detracts from the benefits of an end-to-end method. While vision-only + +![](images/d7d0f7ec76b29fe2ec4998c43430ba14e8da77e45bcfc6d02eedca62b5c8dacc.jpg) +Fig. 2: JDT3D Architecture. At each time step, a BEV feature map is extracted and used to initialize a set of proposal queries. The proposal queries are concatenated to track queries passed from the previous frame and used to predict objects in the scene. Track queries detect the same unique objects in each time step, while proposal queries detect untracked or new objects. + +TBA methods have shown surprising results on the MOT task, the baseline results set by MotionTrack's LiDAR-only and fusion TBA methods underperform compared to LiDAR-based TBD baselines such as CenterPoint [26] and SimpleTrack [20], prompting further exploration of LiDAR TBA methods. Our work explores different potential hypotheses for the reported performance gap between LiDAR-based TBD and TBA methods and proposes improvements to the existing TBA methods. + +# 2.3 LiDAR Data Augmentations + +As with other machine learning tasks, data augmentation can greatly enhance the performance of LiDAR-based 3D object detection methods. Some common augmentations include random flipping, rotation, translation, and scaling of the point cloud [6]. In addition to these, object sampling is also a widely used augmentation method for LiDAR-based object detectors [1,15,26]. First introduced in SECOND [25], object sampling involves creating a database of ground truth object point cloud scans and bounding boxes and randomly sampling and injecting them into the scene during training to improve the convergence and performance overall. However, this augmentation does not apply directly to training TBA methods, since it would result in sporadic track instances in the scene in each training sequence. Our proposed track sampling augmentation samples consecutive track instances and injects them into a sequence of frames during training. This maintains the temporal nature of the object tracks while still offering the same benefits of faster convergence. + +# 3 Method + +To perform our analysis, we developed our own baseline LiDAR-based TBA method since no such implementations were made available as of the writing of this paper. We build on aspects from query-based 3D object detection [1] and vision-only TBA [19] to formulate our method, called JDT3D. JDT3D is summarized in Fig. 2. + +# 3.1 Overview + +At each time step, a feature extractor takes in the LiDAR scan of the scene, $\mathbf{P}$ , and converts it into a BEV feature map representation, $\mathbf{F}$ , using a standard voxel feature extractor [32]. This feature map is used to predict heatmaps of potential object locations and initialize a set of proposal queries, $\mathbf{PQ}$ [1]. + +In the first time step, the proposal queries are passed through a transformer decoder layer with the BEV feature map to obtain a set of updated queries, $\mathbf{Q}^D$ . The past reasoning module refines the updated queries with historical track information [19] before a feedforward network decodes the refined queries, $\mathbf{Q}^R$ into bounding boxes and class predictions, $\mathbf{B}^R$ . During training, $\mathbf{Q}^D$ is decoded into a set of bounding boxes and class predictions, $\mathbf{B}^D$ , for an auxiliary loss. + +To create the track queries for the next frame, the predictions and their corresponding queries are filtered with a confidence threshold. The filtered queries are passed to the future reasoning module [19] which predicts a motion trajectory and projects each query into the next time step. + +At every following time step, new proposal queries are initialized with the heatmap as described above and are concatenated with track queries from the previous time step, $\mathbf{TQ}$ , before being passed to the transformer decoder. + +To perform tracking, high-confidence predictions and associated queries are assigned an ID. Track queries are used to detect the same object consistently over multiple frames, while proposal queries are used to detect objects to be added to the track query set. Tracking is achieved implicitly through the direct relationship between each prediction and each query passed into the transformer decoder. Trajectories can be obtained by tracing a given track query back in time. + +# 3.2 Query Propagation + +During training, only the refined queries corresponding to predictions with a confidence score greater than $\tau_{pass}$ are passed to the next time step. This method of passing and pruning queries has the advantage of naturally creating instances of false negative and false positive track queries during training [31]. False negative track queries are missed tracks that should be re-detected in the next frame as if they were new objects entering the scene. False positive track queries simulate tracks that have left the scene and should be pruned in the next time step. + +This method has the benefit of better matching the inference query propagation behaviour. As well, it creates more instances during training where the decoder must suppress track query predictions and output positive proposal query predictions, ensuring robust handling of track births and deaths. + +![](images/cd0239ea36c858749982b1139b5d1298c1d48296c36ee3f0dfde4ee0a09b691b.jpg) +Fig. 3: An example of track sampling augmentation over three consecutive frames from the nuScenes dataset. The original trajectories and sampled trajectories are shown in blue and orange, respectively. Only a subset of the tracks and the third LiDAR scan are shown for visual clarity. The older boxes are shown with transparency. + +![](images/3bd3bc5c2b0c7d9f1126b7739e7e1008337f8090d5a86ba3dea183404952ae29.jpg) + +# 3.3 Ground Truth Assignment + +The ground truth assignment of the decoder predictions, $\mathbf{B}_t^D$ , and the refined predictions, $\mathbf{B}_t^R$ , are both done in the same manner using a two-stage assignment process. The first stage involves matching the track queries based on their previously matched ground truth objects [18]. If the previously matched ground truth track is not present in the current time step, such as in the case of a track leaving the scene, then the track query prediction is assigned to no object. False positive track queries were never assigned a ground truth track previously, so they are also assigned to no object. + +The second stage of the assignment involves matching the proposal queries to the unmatched ground truths. The unmatched ground truths would include new objects in the scene and missed objects from previous time steps. A Hungarian matcher is used to compute the optimal assignment between the unmatched queries and ground truths based on an association cost [10]. + +# 3.4 Track Sampling Augmentation + +To improve the convergence rate and performance, we introduce track sampling augmentation during training. Inspired by the object sampling augmentation [25], our track sampling involves adding additional tracks into a training clip from a database of tracks in a temporally consistent manner, as illustrated in Fig. 3. The original object sampling augmentation does not account for temporal consistency, and applying it naively to our multi-frame training scheme would result in disjointed track instances during training. + +First, the training set is preprocessed to generate a database of ground truth tracks, including bounding boxes, object classes, track IDs, and LiDAR points contained in the bounding boxes. To perform track sample augmentation on a training clip, a set of tracks is randomly sampled, avoiding track IDs that overlap with the existing ground truth objects in the scene. This enforces unique IDs for + +the tracks in the sequence. To inject a track, $L$ consecutive ground-truth boxes are randomly selected from each sampled track. In each training frame, the LiDAR points corresponding to the sampled objects are used to replace regions in the original point cloud. The objects are injected in chronological order over the $L$ frames. A collision check is performed to ensure that there are no collisions between injected objects and the original ground truth objects in the scene. In the case of a collision, the would-be sampled object is pruned from that frame, simulating an occlusion. + +The benefits of track sampling mirror those of object sampling [25], with the extra ground truth objects in the scene providing more samples to learn from in a single frame. This results in a stronger training signal and more varied scenes. + +# 3.5 Network Training and Losses + +To train JDT3D, $L$ LiDAR frames are passed sequentially to the model as a training sample. The loss for a given frame, $\mathcal{L}_i$ , is computed as the weighted sum of the following losses: + +$$ +\mathcal {L} _ {i} = \lambda_ {h} \mathcal {L} _ {h} + \lambda_ {f} \mathcal {L} _ {f} + \sum_ {m = \{D, R \}} \left(\lambda_ {r e g} ^ {m} \mathcal {L} _ {r e g} ^ {m} + \lambda_ {c l s} ^ {m} \mathcal {L} _ {c l s} ^ {m}\right) \tag {1} +$$ + +The superscripts $D$ and $R$ correspond to predictions output by $\mathbf{Q}^D$ and $\mathbf{Q}^R$ , respectively. $\mathcal{L}_{reg}^{m}$ is the L1 loss over the bounding boxes, and $\mathcal{L}_{cls}^{m}$ is the focal loss [14] over the classification predictions. $\mathcal{L}_h$ is the Gaussian Focal Loss on the predicted object heatmaps [11,26], and $\mathcal{L}_f$ is the L1 loss over the motion trajectory predicted by the future reasoning module [19]. Each loss has an associated scalar weight term, $\lambda$ . The total loss is computed as the sum of each individual frame loss: + +$$ +\mathcal {L} _ {\text {c l i p}} = \sum_ {i = 1} ^ {L} \mathcal {L} _ {i} \tag {2} +$$ + +# 4 Experiments + +# 4.1 Dataset and Metrics + +JDT3D is trained and evaluated on the nuScenes dataset [2], a multi-modal autonomous driving dataset containing data from six multi-view cameras, five radars, and a 360-degree LiDAR. The dataset contains a total of 40,157 frames across 1000 scenes, of which 850 scenes are annotated for training (700) and validation (150) at a frame rate of $2\mathrm{Hz}$ . We train on the seven object classes annotated for the tracking task. + +To evaluate JDT3D, the average multi-object tracking accuracy (AMOTA) and average multi-object tracking precision (AMOTP) tracking metrics [24] are computed across the seven tracking classes of nuScenes, evaluating the tracking accuracy and tracking position misalignment, respectively. Secondary metrics are + +Table 1: A comparison of 3D multi-object tracking methods evaluated on the nuScenes [2] test set. Bold represents the best results among the Transformer-based TBA methods, while Underlined numbers are the best overall. + +
MethodTBAModalityAMOTA ↑AMOTP ↓FP ↓FN ↓IDS ↓
AB3DMOT [24]L0.1511.50115088757309027
Chui et al. [5]L0.5500.7981753333216950
CenterPoint [26]L0.6380.5551861222928760
SimpleTrack [20]L0.6680.5501751423451575
PF-Track [19]✓C0.4341.2521904842758249
DQTrack [13]✓C0.5231.096--1204
MotionTrack-L [29]✓L0.510.99--9705
MotionTrack-LC [29]✓L+C0.550.871--8716
JDT3D (ours)✓L0.5740.8371115229919254
+ +also computed by the nuScenes MOT benchmark [2], such as the number of false positives, false negatives, and identity switches. We also evaluate the detection performance across the tracking classes, with the mean average precision (mAP) and nuScenes Detection Score (NDS) metrics. + +# 4.2 Implementation Details + +VoxelNet [25, 32] is used as the LiDAR backbone to extract the BEV feature map. For each frame, the top 200 predictions from the heatmap are used to initialize the proposal queries. The training clip length $L$ is set to 3 frames. The confidence threshold for track queries $\tau_{pass}$ is set to 0.4 for training and inference. The Hungarian matching cost for ground truth assignment consists of the sum of the L1 cost on the predicted box and the focal loss [14] on the predicted class confidence. All detections are generated in the car ego-frame, so the object centers of the track queries are updated to their relative positions in the current training frame based on the car's ego-motion between time steps. A pretrained VoxelNet backbone from BEVFusion [15] is used to initialize the feature extractor. To train JDT3D, the model is first trained on the detection task using $L = 1$ for 20 epochs. It is then further trained on the tracking task using $L = 3$ for another 20 epochs. All experiments were trained and evaluated on eight NVIDIA V100 GPUs using random flipping, rotation, translation, and scaling on the training sequence, and track sampling augmentation, unless otherwise specified. JDT3D performs inference at $3.85\mathrm{Hz}$ on a Tesla V100. + +# 4.3 Main Results + +Our main results compare JDT3D with baseline TBD methods and relevant TBA methods evaluated on the nuScenes test set [2], summarized in Tab. 1. JDT3D outperforms the current state-of-the-art LiDAR-based TBA methods, with a relative improvement of $12.5\%$ on the AMOTA metric and $15.5\%$ on the + +Table 2: Ablation on the track sampling augmentation, reported on the nuScenes validation set. The mAP metric is computed on the seven tracking categories. + +
Track Sampling RateAMOTA ↑AMOTP ↓IDS ↓mAP ↑
0.00.3141.3773880.225
0.250.6000.8642100.3867
0.500.6220.8162030.400
1.00.6210.7901500.399
+ +AMOTP metric on the test split. Moreover, with LiDAR input only, JDT3D also outperforms the fusion-based MotionTrack-LC [29] with a relative improvement of $4.36\%$ on AMOTA and $3.79\%$ on AMOTP on the test set. While JDT3D underperforms compared to baseline TBD methods on most metrics, it achieves the lowest number of ID switches among LiDAR-only methods, indicating a strong track coherence. + +JDT3D has significantly more false negatives compared to TBD methods, but fewer false positives. The potential sources of false negatives are either due to missed detections or tracks diverging from the ground truths. However, the latter case would result in both high false negatives and false positives. Based on the low false positive results from JDT3D, we conclude that the false negatives are mainly due to missed detections. + +# 4.4 Ablation Studies + +Track Sampling. By comparing the performance of JDT3D on the nuScenes validation set with and without track sampling augmentation, we verify that sampling extra tracks into a training sequence significantly improves the convergence rate and the final performance. The comparison summarized in Tab. 2 shows a $97\%$ relative improvement on the AMOTA metric from track sampling augmentation compared to no track sampling, with a similar trend across all other metrics, including detection metrics such as the mAP. We also analyzed the effect of varying the number of objects that are sampled by reducing the sampling frequency to 0.5 and 0.25 of the nominal rate. We observe that the AMOTA and mAP metrics plateau after a sampling rate of 0.5, although the AMOTP and IDS metrics see continued improvement with increased track sampling. The benefits of implementing track sampling are expected to generalize to other LiDAR-based TBA methods, but there are currently no available implementations to verify this hypothesis. + +Training Clip Length. To validate our choice of clip length, the performances of JDT3D models trained with two to four frame sequences were compared and summarized in Tab. 3. The best performance is achieved when trained over $L = 3$ frames, with a significant jump when moving from $L = 2$ to $L = 3$ . We hypothesize that the training process benefits from the longer-term, temporally consistent information, resulting in improved tracking performance. Moving from + +Table 3: Ablation on the training sequence length, $L$ , reported on the nuScenes validation set. + +
Clip Length (L)AMOTA ↑AMOTP ↓IDS ↓mAP ↑
20.3971.2944460.214
30.6210.7901500.399
40.5990.7831730.392
+ +Table 4: Ablation on the query propagation criteria, reported on the nuScenes [2] validation set. Passing queries with high confidence encourages the re-detection of poor detections and handling of false positives and false negatives. + +
Query PropagationAMOTA↑ AMOTP↓ Recall↑ IDS
Matched to a GT0.4640.9170.51182
High Confidence0.6210.7900.669150
+ +$L = 3$ to $L = 4$ , the performance drops slightly while also significantly increasing the train time and required memory. + +Query Propagation Method. In Tab. 4, we compare different query propagation methods; the first method is passing queries that have been matched to a ground truth [18, 19], and the second is passing based on the $\tau_{pass}$ threshold for high confidence predictions as described in Sec. 3.2. Results show that using confidence as a filtering criterion during training enhances the overall tracking metrics AMOTA and AMOTP by a large margin and significantly reduces the number of false positives and false negatives while increasing the number of true positives. While the ID switch errors are higher, this increase is less than $2\%$ relative to the increase in true positives. Thus, we verify that using confidence as the query propagation criterion is more appropriate for the TBA paradigm. + +In addition, we show the generalizability of this query propagation method by performing the same ablation on PF-Track. The results in Tab. 5 show improvements across most metrics, with a similar behaviour of more IDS errors. + +Comparison with TBD Baseline. We compare the performance between our TBA framework and baseline TBD methods such as the CenterPoint greedy-based tracking method [26] and SimpleTrack [20]. To compare our tracking + +Table 5: An ablation comparing the original ground truth-based propagation and our confidence-based propagation on PF-Track-S [19] evaluated on the nuScenes [2] validation set. + +
Query PropagationAMOTA ↑AMOTP ↓Recall ↑IDS ↓
Matched to a GT0.4081.3430.507166
High Confidence0.4261.3160.558306
+ +Table 6: Comparison with "Tracking by Detection" on the nuScenes validation set, using the detections generated by JDT3D. + +
Tracking MethodAMOTA ↑AMOTP ↓IDS ↓
CenterPoint [26]0.6080.825491
SimpleTrack [20]0.5730.8775585
JDT3D (Ours)0.6210.790150
+ +Table 7: Comparison of detection performance between the single-frame pretraining and the three-frame fine-tuning. Evaluated on the seven tracking classes on the nuScenes validation set. The best results are typeset in boldface. + +
JDT3D variantmAP ↑NDS ↑
Single-frame0.6050.549
Three-frame0.3990.502
+ +method to the TBD trackers independent of detection performance, we pass the detections generated by JDT3D to the tracker and evaluate its tracking performance compared to our joint approach. The results in Tab. 6 demonstrate clear advantages of our joint end-to-end approach that offers temporal consistency and data association and reaffirm that the gap in tracking performance lies in the missed detections. + +# 4.5 Detection Performance + +As mentioned in Sec. 4.2, JDT3D is pre-trained to perform single-frame detection and fine-tuned to perform tracking on sequences of multiple LiDAR frames. While the pre-training step achieves reasonable detection performance, the performance drops significantly after fine-tuning, as seen in Tab. 7. This indicates that the current TBA problem formulation may not be the best alignment of the detection and tracking objectives. + +# 4.6 Proposal and Track Query Imbalances + +Tab. 8 shows further analysis of the JDT3D outputs. During training, the positively matched proposal queries, or untracked newborn objects, are more confident in the first frame than in subsequent frames. Also, the positively matched track queries, or previously tracked objects, tend to be more confident than the newborn objects. This indicates that the model has a strong bias toward previously tracked objects. This bias could lead to more missed objects and fewer initialized tracks, explaining the high number of false negatives indicated in Tab. 1. + +One hypothesis for this behaviour is that this is caused by the imbalance in training samples assigned to each type of query. Tab. 9 shows the average distribution of true positive predictions of both proposal and track queries in + +Table 8: The average confidence of positively matched proposal and track predictions over a three-frame training sequence. There are no track queries in the first frame. + +
Avg. ConfidenceFrame 1Frame 2Frame 3
Positively Matched Proposals0.6690.4090.387
Track Predictionsn/a0.7590.760
+ +Table 9: The distribution of positive and negative matches for proposal queries and track queries in the second and third frames of training. + +
PQ PredictionsTQ Predictions
Avg Positive Matches4 (2%)47 (94%)
Avg Negative Matches196 (98%)3 (6%)
+ +the second and third frames of training. Note that the proportion of positively matched proposal queries is significantly lower than for track queries. This is somewhat expected since the average track birthrate per frame is 0.7, so most objects in a scene were present in the previous frame. This imbalance in positive training samples between the proposal and track queries could explain the confidence imbalance observed. + +# 5 Discussion + +Based on our analysis from Tab. 6 we show that the reason LiDAR-based TBA methods continue to underperform compared to TBD methods is due to poorer detection performance. In addition, we note that there is a significant drop in the detection performance when fine-tuning the model on multiple frames in Tab. 7. In the following sections, we discuss hypotheses to explain the observed results and propose potential improvements to build on JDT3D. + +# 5.1 Temporal Confusion + +Our hypothesis for the observed drop in performance in Tab. 7 lies in the fact that the problem formulation of TBA is more complex than query-based detection. In query-based detection, the handling of the queries by the decoder is straightforward: if the query represents an object in the scene, it should output a high-confidence bounding box from the query. On the other hand, in TBA, the handling of queries is conditional: depending on whether the decoder receives a track query that represents the same object, the decoder needs to determine whether the incoming proposal query is a duplicate that should have a low confidence or a unique object that should have a high confidence, as illustrated in Fig. 4. This complex interaction results in "temporal confusion" such that the expected model behaviour is different depending on the query inputs. We hypothesize that the reason temporal confusion has not been an issue in vision-based TBA methods [18,19,30] is due to their larger decoder size enabling more + +![](images/996dca4637e4186161c434653079f8c5d4346437748636108c52e20c99fa6d55.jpg) +Fig. 4: An example of temporal confusion, where the decoder must handle the same proposal query differently based on the presence of track queries. In Case 1, the proposal query should be a positive prediction, while in Case 2, it should be a negative prediction. + +complex reasoning about the queries. To test this hypothesis, further optimizations would be needed to enable a larger decoder with a LiDAR backbone due to hardware constraints. + +# 5.2 Potential Improvements + +With these results in mind, we propose two potential directions that can be further explored to bridge the gap between JDT and TBD methods. + +Firstly, based on the ablation performed to evaluate the training clip size, we observed that three-frame training significantly improved performance over two-frame training, although extending to four frames reduced performance slightly. Further exploration of the effect of clip length on tracking performance may yield insights on how to better leverage longer term temporal information. + +Secondly, optimizations to the model size and memory footprint would enable a larger and more complex decoder, that could potentially address the temporal confusion hypothesis. + +# 6 Conclusion + +This paper presents JDT3D, a novel LiDAR tracking method that makes significant progress in closing the gap between TBD and TBA methods. JDT3D is trained using track sampling augmentation, a generalizable data augmentation method that injects extra long-term tracks into training samples. We show that track sampling augmentation significantly improves training performance and convergence, and we hope that future LiDAR-based trackers can leverage these findings. As well, we show that the query propagation scheme during training also has a significant impact on the performance, showing that using a confidence threshold to filter low confidence predictions improves tracking performance greatly. Based on our experiments, we hypothesize that the complexity of tracking-by-attention combined with the memory constraints imposed by LiDAR feature extractors limiting the size of transformer decoders is the cause of the drop in detection performance between single and multi-frame variants of JDT3D. We propose further areas of exploration including exploring different training sequence lengths and enabling more decoder layers. + +# References + +1. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1090-1099 (June 2022) +2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11618-11628 (2019), https://apisemantic scholar.org/CorpusID: 85517967 +3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European Conference on Computer Vision. pp. 213-229. Springer (2020) +4. Chen, Y., Yu, Z., Chen, Y., Lan, S., Anandkumar, A., Jia, J., Alvarez, J.M.: Focalformer3d: Focusing on hard instance for 3d object detection. ArXiv abs/2308.04556 (2023), https://api-semanticscholar.org/CorpusID: 260736073 +5. Chiu, H.k., Prioletti, A., Li, J., Bohg, J.: Probabilistic 3d multi-object tracking for autonomous driving. arXiv preprint arXiv:2001.05673 (2020) +6. Hahner, M., Dai, D., Liniger, A., Gool, L.V.: Quantifying data augmentation for lidar based 3d object detection. CoRR abs/2004.01643 (2020), https://arxiv.org/abs/2004.01643 +7. Hu, Y., Yang, J., Chen, L., Li, K., Sima, C., Zhu, X., Chai, S., Du, S., Lin, T., Wang, W., et al.: Planning-oriented autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17853-17862 (2023) +8. Huang, T.E., Liu, Y., Van Gool, L., Yu, F.: Video task decathlon: Unifying image and video tasks in autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8647-8657 (2023) +9. Koh, J., Kim, J., Yoo, J.H., Kim, Y., Kum, D., Choi, J.W.: Joint 3d object detection and tracking using spatio-temporal representation of camera image and lidar point clouds. Proceedings of the AAAI Conference on Artificial Intelligence 36(1), 1210-1218 (Jun 2022). https://doi.org/10.1609/aaai.v36i1.20007, https://ojs.aaai.org/index.php/AAAI/article/view/20007 +0. Kuhn, H.W.: The hungarian method for the assignment problem. Naval Research Logistics (NRL) 52 (1955), https://api-semanticscholar.org/CorpusID: 9426884 +1. Law, H., Deng, J.: Cornernet: Detecting objects as paired keypoints. Int. J. Comput. Vis. 128(3), 642-656 (2020). https://doi.org/10.1007/S11263-019-01204-1, https://doi.org/10.1007/s11263-019-01204-1 +2. Li, X., Xie, T., Liu, D., Gao, J., Dai, K., Jiang, Z., Zhao, L., Wang, K.: Poly-mot: A polyhedral framework for 3d multi-object tracking. ArXiv abs/2307.16675 (2023), https://api.sementicscholar.org/CorpusID:260334560 +3. Li, Y., Yu, Z., Philion, J., Anandkumar, A., Fidler, S., Jia, J., Alvarez, J.: End-to-end 3d tracking with decoupled queries. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2023) +4. Lin, T.Y., Goyal, P., Girshick, R.B., He, K., Dollar, P.: Focal loss for dense object detection. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 2999-3007 (2017), https://api-semanticscholar.org/CorpusID:47252984 + +15. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. 2023 IEEE International Conference on Robotics and Automation (ICRA) pp. 2774-2781 (2022), https://apisemantic scholar.org/CorpusID:249097415 +16. Luo, C., Yang, X., Yuille, A.: Exploring simple 3d multi-object tracking for autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10488-10497 (October 2021) +17. Mahmoud, A., Hu, J.S.K., Waslander, S.L.: Dense voxel fusion for 3d object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 663-672 (January 2023) +18. Meinhardt, T., Kirillov, A., Leal-Taixe, L., Feichtenhofer, C.: Trackformer: Multiobject tracking with transformers. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8834-8844 (2022). https://doi.org/10.1109/CVPR52688.2022.00864 +19. Pang, Z., Li, J., Tokmakov, P., Chen, D., Zagoruyko, S., Wang, Y.X.: Standing between past and future: Spatio-temporal modeling for multi-camera 3d multi-object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17928-17938 (June 2023) +20. Pang, Z., Li, Z., Wang, N.: SimpleTrack: Understanding and Rethinking 3D Multiobject Tracking. In: Karlinsky, L., Michaeli, T., Nishino, K. (eds.) Computer Vision - ECCV 2022 Workshops. pp. 680-696. Lecture Notes in Computer Science, Springer Nature Switzerland (2021). https://doi.org/10.1007/978-3-031-25056-9_43 +21. Sun, P., Cao, J., Jiang, Y., Zhang, R., Xie, E., Yuan, Z., Wang, C., Luo, P.: Transtrack: Multiple-object tracking with transformer. arXiv preprint arXiv: 2012.15460 (2020) +22. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017), https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf +23. Wang, L., Zhang, X.N., Qin, W., Li, X., Yang, L., Li, Z., Zhu, L., Wang, H., Li, J., Liu, H.: Camo-mot: Combined appearance-motion optimization for 3d multi-object tracking with camera-lidar fusion. IEEE Transactions on Intelligent Transportation Systems 24, 11981-11996 (2022), https://api_semanticscholar.org/CorpusID: 252089635 +24. Weng, X., Wang, J., Held, D., Kitani, K.: 3d multi-object tracking: A baseline and new evaluation metrics. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 10359-10366 (2020). https://doi.org/10.1109/IROS45743.2020.9341164 +25. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors (Basel, Switzerland) 18 (2018), https://api_semanticscholar.org/CorpusID:52957856 +26. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11784-11793 (June 2021) +27. Zeng, F., Dong, B., Wang, T., Chen, C., Zhang, X., Wei, Y.: Motr: End-to-end multiple-object tracking with transformer. ArXiv abs/2105.03247 (2021), https://api-semanticscholar.org/CorpusID:234096063 + +28. Zeng, Y., Ma, C., Zhu, M., Fan, Z., Yang, X.: Cross-modal 3d object detection and tracking for auto-driving. In: IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2021, Prague, Czech Republic, September 2021. IEEE (2021) +29. Zhang, C., Zhang, C., Guo, Y., Chen, L., Happold, M.: Motiontrack: End-to-end transformer-based multi-object tracking with lidar-camera fusion. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) pp. 151-160 (2023), https://apisemanticscholar.org/CorpusID:259286967 +30. Zhang, T., Chen, X., Wang, Y., Wang, Y., Zhao, H.: Muitr3d: A multi-camera tracking framework via 3d-to-2d queries. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 4536-4545 (2022). https://doi.org/10.1109/CVPRW56347.2022.00500 +31. Zhang, Y., Wang, T., Zhang, X.: Motrv2: Bootstrapping end-to-end multi-object tracking by pretrained object detectors. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 22056-22065 (2022), https://apisemantic scholar.org/CorpusID:253581846 +32. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition pp. 4490-4499 (2017), https://apisemantic scholar.org/CorpusID:42427078 +33. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: deformable transformers for end-to-end object detection. In: 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 (2021) \ No newline at end of file diff --git a/2024/JDT3D_ Addressing the Gaps in LiDAR-Based Tracking-by-Attention/images.zip b/2024/JDT3D_ Addressing the Gaps in LiDAR-Based Tracking-by-Attention/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..277b8439be81144c72b0bfbbf67acb61f2634337 --- /dev/null +++ b/2024/JDT3D_ Addressing the Gaps in LiDAR-Based Tracking-by-Attention/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b95819c6252349a9d1eb724843c782300dcb138b8cf0f78c54c0f91aa5ad930f +size 300548 diff --git a/2024/JDT3D_ Addressing the Gaps in LiDAR-Based Tracking-by-Attention/layout.json b/2024/JDT3D_ Addressing the Gaps in LiDAR-Based Tracking-by-Attention/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2b4e296b7a594422bdd6712c8e1ce2aa5851c591 --- /dev/null +++ b/2024/JDT3D_ Addressing the Gaps in LiDAR-Based Tracking-by-Attention/layout.json @@ -0,0 +1,7502 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 111, + 471, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 111, + 471, + 148 + ], + "spans": [ + { + "bbox": [ + 143, + 111, + 471, + 148 + ], + "type": "text", + "content": "JDT3D: Addressing the Gaps in LiDAR-Based Tracking-by-Attention" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 176, + 168, + 436, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 168, + 436, + 180 + ], + "spans": [ + { + "bbox": [ + 176, + 168, + 436, + 180 + ], + "type": "text", + "content": "Brian Cheong\\* Jiachen Zhou\\* and Steven Waslander" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 190, + 474, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 190, + 474, + 213 + ], + "spans": [ + { + "bbox": [ + 138, + 190, + 474, + 213 + ], + "type": "text", + "content": "University of Toronto \n{brian.cheong, jason.zhou, steven.waslander}@robotics.utias.utoronto.ca" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 244, + 455, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 244, + 455, + 422 + ], + "spans": [ + { + "bbox": [ + 160, + 244, + 455, + 422 + ], + "type": "text", + "content": "Abstract. Tracking-by-detection (TBD) methods achieve state-of-the-art performance on 3D tracking benchmarks for autonomous driving. On the other hand, tracking-by-attention (TBA) methods have the potential to outperform TBD methods, particularly for long occlusions and challenging detection settings. This work investigates why TBA methods continue to lag in performance behind TBD methods using a LiDAR-based joint detector and tracker called JDT3D. Based on this analysis, we propose two generalizable methods to bridge the gap between TBD and TBA methods: track sampling augmentation and confidence-based query propagation. JDT3D is trained and evaluated on the nuScenes dataset, achieving 0.574 on the AMOTA metric on the nuScenes test set, outperforming all existing LiDAR-based TBA approaches by over " + }, + { + "bbox": [ + 160, + 244, + 455, + 422 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 160, + 244, + 455, + 422 + ], + "type": "text", + "content": ". Based on our results, we further discuss some potential challenges with the existing TBA model formulation to explain the continued gap in performance with TBD methods. The implementation of JDT3D can be found at the following link: https://github.com/TRAILab/JDT3D." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 430, + 455, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 430, + 455, + 452 + ], + "spans": [ + { + "bbox": [ + 160, + 430, + 455, + 452 + ], + "type": "text", + "content": "Keywords: Multi-Object Tracking " + }, + { + "bbox": [ + 160, + 430, + 455, + 452 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 430, + 455, + 452 + ], + "type": "text", + "content": " Autonomous Vehicles " + }, + { + "bbox": [ + 160, + 430, + 455, + 452 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 430, + 455, + 452 + ], + "type": "text", + "content": " Computer Vision" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 475, + 230, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 230, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 230, + 487 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 501, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 559 + ], + "type": "text", + "content": "Multi-object tracking (MOT) is an essential component in the perception system of autonomous vehicles, as it allows autonomous agents to reason and plan appropriately in dynamic environments. The task involves identifying object trajectories, requiring that each object is accurately detected and consistently identified over time within the scene." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 560, + 482, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 560, + 482, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 560, + 482, + 620 + ], + "type": "text", + "content": "The majority of MOT methods adhere to the tracking-by-detection (TBD) paradigm. This paradigm is comprised of two separate processes: generating bounding box predictions without track IDs and associating these predictions to a set of maintained tracks based on some matching criteria. TBD methods currently achieve state-of-the-art results on 3D MOT benchmarks [12, 15, 23]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 621, + 482, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 482, + 645 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 482, + 645 + ], + "type": "text", + "content": "In contrast to the decoupled approach of TBD, joint detection and tracking (JDT) approaches have attempted to unify the detection and tracking tasks in an" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 257, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 257, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 257, + 665 + ], + "type": "text", + "content": "* Denotes equal contribution." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "end-to-end manner [18, 19, 29, 31]. Recent works in autonomous driving perception [7, 17] and computer vision [8] demonstrate that networks trained end-to-end have shown superior performance compared to pipeline-based techniques, despite initially lower performance from earlier works [21, 27]. End-to-end methods can learn richer and more generalizable representations while minimizing the task-specific engineering of individual modules required in a pipeline-based approach. While TBD methods enforce a unidirectional flow of information from the detector to the tracker, the JDT formulation allows both the detector and tracker to exchange useful information to enhance the performance of both tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 227, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 370 + ], + "type": "text", + "content": "Integrating detection and tracking into a joint learning task requires a unique approach to formulating the problem and training the model. One such approach is tracking-by-attention (TBA), illustrated in Fig. 1. In TBA, objects are represented as vector embeddings, or \"queries\", that are used to detect the same object over multiple frames. While recent TBA methods [18,19,29,31] have been proposed as alternatives to TBD, they continue to underperform on the MOT task, especially in the LiDAR domain [29]. However, they have shown potential in certain cases, such as significantly fewer ID-switch errors [19]. The unexplored potential of TBA in the LiDAR domain motivates our work in addressing a key question: what is holding back LiDAR-based TBA? In this work, we propose a LiDAR-based TBA tracking method called JDT3D, based on which we explore different aspects that hold back LiDAR-based TBA." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 373, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 445 + ], + "type": "text", + "content": "One insight is that LiDAR-based tracking methods can suffer from sparse supervision signals due to point cloud and object sparsity. Inspired by object sampling augmentation [25] for LiDAR-based detection methods, we propose track sampling, a temporally consistent augmentation method that injects consistent objects over multiple LiDAR frames to enrich supervision signals while maintaining temporal consistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 449, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 545 + ], + "type": "text", + "content": "In addition, existing TBA methods maintain an inconsistency between training and inference regarding their query propagation method. During training, queries that are matched to ground truth samples are propagated to the next frame, but at inference, propagation is based on a confidence threshold. This inconsistency between training and inference could confuse the model to over-trust false positive queries. To address this, we propose a confidence threshold propagation strategy consistent for both training and inference and conduct comprehensive evaluations and discussions on the strategy design." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 548, + 341, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 548, + 341, + 559 + ], + "spans": [ + { + "bbox": [ + 146, + 548, + 341, + 559 + ], + "type": "text", + "content": "We summarize our contributions as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 563, + 488, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 488, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 488, + 599 + ], + "type": "text", + "content": "JDT3D We present JDT3D, a LiDAR-based tracking-by-attention model that outperforms existing LiDAR-based tracking-by-attention methods on the nuScenes test set by over " + }, + { + "bbox": [ + 130, + 563, + 488, + 599 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 130, + 563, + 488, + 599 + ], + "type": "text", + "content": " on the AMOTA metric." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 602, + 482, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 602, + 482, + 627 + ], + "spans": [ + { + "bbox": [ + 130, + 602, + 482, + 627 + ], + "type": "text", + "content": "Confidence-based query propagation We perform ablations to validate the effectiveness and generalizability of the confidence-based propagation criterion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Track Sampling Augmentation We propose and validate a novel data augmentation technique for multi-frame LiDAR methods that significantly improves performance and convergence rates during training." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "B. Cheong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 178, + 116, + 434, + 211 + ], + "blocks": [ + { + "bbox": [ + 178, + 116, + 434, + 211 + ], + "lines": [ + { + "bbox": [ + 178, + 116, + 434, + 211 + ], + "spans": [ + { + "bbox": [ + 178, + 116, + 434, + 211 + ], + "type": "image", + "image_path": "901813337f9bd5a1067850121f1a0273eb4f80e96490e461851e5b3a37ae91f1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 221, + 482, + 266 + ], + "lines": [ + { + "bbox": [ + 130, + 221, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 482, + 266 + ], + "type": "text", + "content": "Fig. 1: Illustration of tracking by attention. Each object is represented by a query. In the diagram, the yellow motorcycle leaves the frame, so it is removed from the set of maintained queries. To detect new objects in the scene, such as the tan bus, proposal queries are appended at each time step, represented by the blank squares." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 289, + 482, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 350 + ], + "type": "text", + "content": "Analysis Based on our experiments using JDT3D, we provide a detailed analysis to explain the continued gap in performance between LiDAR-based TBD and TBA. We attribute the gap to the reduced detection performance when training on multiple consecutive frames and the handling of queries that results in \"temporal confusion\"." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 365, + 242, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 242, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 242, + 379 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 390, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 498 + ], + "type": "text", + "content": "Multi-object tracking (MOT) methods either follow the tracking-by-detection paradigm, in which object detection is independently performed on single frames before the tracking module, or they perform detection and tracking jointly in an end-to-end fashion. MOT can also be further categorized into 2D and 3D MOT. The task of 2D MOT involves predicting object tracks in 2D space, while 3D MOT predicts the tracks in 3D space. 2D MOT methods generally take in camera images as sensor input, while 3D MOT methods may take in camera images, LiDAR point clouds, or both. In this work, we will be focusing our analysis on 3D MOT methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 515, + 307, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 307, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 307, + 528 + ], + "type": "text", + "content": "2.1 Tracking-by-detection (TBD)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": "In the tracking-by-detection paradigm, the detection and tracking problems are solved by two distinct modules. First, a single-frame object detector is used to detect objects in a given scene. Next, the detections are associated with a set of maintained tracks by a tracking module and used to update them with current detection information. Unassigned detections are used to initialize new tracks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "While different tracking methods exist among TBD methods, they are all consistently tied to the detector's performance. Several top-performing 3D tracking methods [1,4,15] are strong detectors that have been paired with the simple tracking method of CenterPoint [26]. The formulation of TBD inhibits the detector from taking advantage of temporal object information that would enhance object detection. As well, errors in the detection are guaranteed to be propagated" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "text", + "content": "JDT3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "downstream to the tracking task. This motivates the exploration of methods that can jointly optimize detection and tracking to achieve MOT." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 155, + 344, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 155, + 344, + 167 + ], + "spans": [ + { + "bbox": [ + 131, + 155, + 344, + 167 + ], + "type": "text", + "content": "2.2 Joint Detection and Tracking (JDT)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 172, + 482, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 482, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 482, + 328 + ], + "type": "text", + "content": "Joint detection and tracking (JDT) is a recent paradigm in both 2D and 3D MOT that involves training a model to perform detection and tracking end-to-end. In contrast to the TBD methods that rely on a dedicated detector to generate detections, JDT methods simultaneously output object detections with associated tracking IDs. Without heuristic matching or hand-crafted rules, SimTrack [16] builds upon CenterPoint [26] by learning a hybrid-time centerness map between two consecutive LiDAR point clouds. AlphaTrack [28] adds a cross-modal fusion scheme to the CenterPoint tracking method [26] and applies an alternating training strategy to balance between the detection branch and appearance affinity. The above two methods perform detection and tracking jointly, but known tracks are not utilized to inform future detections. 3D DetecTrack [9] proposes tracklet-aware 3D detection that reconfigures the initial detection using the latest tracklets' information and employs a spatio-temporal gated GNN for association." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 481, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 481, + 390 + ], + "type": "text", + "content": "Tracking-by-Attention (TBA) A growing trend within JDT is the tracking-by-attention (TBA) approach that emerged from TrackFormer [18] for 2D MOT. TBA uses the transformer architecture [22] and formulates tracking as a frame-to-frame set prediction problem that can be trained in an end-to-end manner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 390, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 532 + ], + "type": "text", + "content": "TrackFormer [18] extends a 2D query-based detector, DETR [3] and its variant Deformable DETR [33], by designing autoregressive track queries that are passed across frames with associated object identities. MOTRv2 [31] uses a pretrainedYOLOX detector to initialize their object queries and uses the detections to anchor the decoded bounding box positions. In addition, track queries during training are propagated based on the prediction confidence rather than whether a prediction was matched to a ground truth object. This naturally creates false positive and false negative cases during training, removing the need for query injection and removal augmentations [18,27]. JDT3D applies confidence-based track query propagation to LiDAR-based TBA, and we perform ablations showing the method's improved performance over the ground truth matching criterion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "MUTR3D [30] and PF-Track [19] demonstrate the transfer of the tracking-by-attention framework to multi-view vision-only 3D MOT. PF-Track [19] also designs past and future reasoning modules to enhance the spatio-temporal representation of queries. MotionTrack [29] investigates the feasibility of extending the TBA paradigm to LiDAR-based and LiDAR-camera fusion-based 3D JDT by building upon TransFusion [1], a query-based 3D object detector. Rather than implicitly learning which queries should be suppressed through the transformer decoder, MotionTrack learns an explicit affinity matrix between tracks and objects in the scene and performs greedy matching to determine their association. This explicit matching echoes the decoupling found in TBD methods and potentially detracts from the benefits of an end-to-end method. While vision-only" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "B. Cheong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 116, + 460, + 238 + ], + "blocks": [ + { + "bbox": [ + 151, + 116, + 460, + 238 + ], + "lines": [ + { + "bbox": [ + 151, + 116, + 460, + 238 + ], + "spans": [ + { + "bbox": [ + 151, + 116, + 460, + 238 + ], + "type": "image", + "image_path": "d7d0f7ec76b29fe2ec4998c43430ba14e8da77e45bcfc6d02eedca62b5c8dacc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 251, + 482, + 307 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 307 + ], + "type": "text", + "content": "Fig. 2: JDT3D Architecture. At each time step, a BEV feature map is extracted and used to initialize a set of proposal queries. The proposal queries are concatenated to track queries passed from the previous frame and used to predict objects in the scene. Track queries detect the same unique objects in each time step, while proposal queries detect untracked or new objects." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 338, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 422 + ], + "type": "text", + "content": "TBA methods have shown surprising results on the MOT task, the baseline results set by MotionTrack's LiDAR-only and fusion TBA methods underperform compared to LiDAR-based TBD baselines such as CenterPoint [26] and SimpleTrack [20], prompting further exploration of LiDAR TBA methods. Our work explores different potential hypotheses for the reported performance gap between LiDAR-based TBD and TBA methods and proposes improvements to the existing TBA methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 459, + 305, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 459, + 305, + 470 + ], + "spans": [ + { + "bbox": [ + 131, + 459, + 305, + 470 + ], + "type": "text", + "content": "2.3 LiDAR Data Augmentations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "As with other machine learning tasks, data augmentation can greatly enhance the performance of LiDAR-based 3D object detection methods. Some common augmentations include random flipping, rotation, translation, and scaling of the point cloud [6]. In addition to these, object sampling is also a widely used augmentation method for LiDAR-based object detectors [1,15,26]. First introduced in SECOND [25], object sampling involves creating a database of ground truth object point cloud scans and bounding boxes and randomly sampling and injecting them into the scene during training to improve the convergence and performance overall. However, this augmentation does not apply directly to training TBA methods, since it would result in sporadic track instances in the scene in each training sequence. Our proposed track sampling augmentation samples consecutive track instances and injects them into a sequence of frames during training. This maintains the temporal nature of the object tracks while still offering the same benefits of faster convergence." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "text", + "content": "JDT3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 196 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 196 + ], + "type": "text", + "content": "To perform our analysis, we developed our own baseline LiDAR-based TBA method since no such implementations were made available as of the writing of this paper. We build on aspects from query-based 3D object detection [1] and vision-only TBA [19] to formulate our method, called JDT3D. JDT3D is summarized in Fig. 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 211, + 208, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 211, + 208, + 223 + ], + "spans": [ + { + "bbox": [ + 132, + 211, + 208, + 223 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 228, + 481, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 481, + 276 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 481, + 276 + ], + "type": "text", + "content": "At each time step, a feature extractor takes in the LiDAR scan of the scene, " + }, + { + "bbox": [ + 130, + 228, + 481, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 228, + 481, + 276 + ], + "type": "text", + "content": ", and converts it into a BEV feature map representation, " + }, + { + "bbox": [ + 130, + 228, + 481, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 130, + 228, + 481, + 276 + ], + "type": "text", + "content": ", using a standard voxel feature extractor [32]. This feature map is used to predict heatmaps of potential object locations and initialize a set of proposal queries, " + }, + { + "bbox": [ + 130, + 228, + 481, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{PQ}" + }, + { + "bbox": [ + 130, + 228, + 481, + 276 + ], + "type": "text", + "content": " [1]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 276, + 481, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 481, + 347 + ], + "type": "text", + "content": "In the first time step, the proposal queries are passed through a transformer decoder layer with the BEV feature map to obtain a set of updated queries, " + }, + { + "bbox": [ + 130, + 276, + 481, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^D" + }, + { + "bbox": [ + 130, + 276, + 481, + 347 + ], + "type": "text", + "content": ". The past reasoning module refines the updated queries with historical track information [19] before a feedforward network decodes the refined queries, " + }, + { + "bbox": [ + 130, + 276, + 481, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^R" + }, + { + "bbox": [ + 130, + 276, + 481, + 347 + ], + "type": "text", + "content": " into bounding boxes and class predictions, " + }, + { + "bbox": [ + 130, + 276, + 481, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{B}^R" + }, + { + "bbox": [ + 130, + 276, + 481, + 347 + ], + "type": "text", + "content": ". During training, " + }, + { + "bbox": [ + 130, + 276, + 481, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^D" + }, + { + "bbox": [ + 130, + 276, + 481, + 347 + ], + "type": "text", + "content": " is decoded into a set of bounding boxes and class predictions, " + }, + { + "bbox": [ + 130, + 276, + 481, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{B}^D" + }, + { + "bbox": [ + 130, + 276, + 481, + 347 + ], + "type": "text", + "content": ", for an auxiliary loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 347, + 481, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 481, + 396 + ], + "type": "text", + "content": "To create the track queries for the next frame, the predictions and their corresponding queries are filtered with a confidence threshold. The filtered queries are passed to the future reasoning module [19] which predicts a motion trajectory and projects each query into the next time step." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 396, + 481, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 481, + 430 + ], + "type": "text", + "content": "At every following time step, new proposal queries are initialized with the heatmap as described above and are concatenated with track queries from the previous time step, " + }, + { + "bbox": [ + 130, + 396, + 481, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{TQ}" + }, + { + "bbox": [ + 130, + 396, + 481, + 430 + ], + "type": "text", + "content": ", before being passed to the transformer decoder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 431, + 481, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 481, + 503 + ], + "type": "text", + "content": "To perform tracking, high-confidence predictions and associated queries are assigned an ID. Track queries are used to detect the same object consistently over multiple frames, while proposal queries are used to detect objects to be added to the track query set. Tracking is achieved implicitly through the direct relationship between each prediction and each query passed into the transformer decoder. Trajectories can be obtained by tracing a given track query back in time." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 517, + 257, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 257, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 257, + 529 + ], + "type": "text", + "content": "3.2 Query Propagation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 533, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 617 + ], + "type": "text", + "content": "During training, only the refined queries corresponding to predictions with a confidence score greater than " + }, + { + "bbox": [ + 130, + 533, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\tau_{pass}" + }, + { + "bbox": [ + 130, + 533, + 482, + 617 + ], + "type": "text", + "content": " are passed to the next time step. This method of passing and pruning queries has the advantage of naturally creating instances of false negative and false positive track queries during training [31]. False negative track queries are missed tracks that should be re-detected in the next frame as if they were new objects entering the scene. False positive track queries simulate tracks that have left the scene and should be pruned in the next time step." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "This method has the benefit of better matching the inference query propagation behaviour. As well, it creates more instances during training where the decoder must suppress track query predictions and output positive proposal query predictions, ensuring robust handling of track births and deaths." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "B. Cheong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 195, + 115, + 305, + 224 + ], + "blocks": [ + { + "bbox": [ + 195, + 115, + 305, + 224 + ], + "lines": [ + { + "bbox": [ + 195, + 115, + 305, + 224 + ], + "spans": [ + { + "bbox": [ + 195, + 115, + 305, + 224 + ], + "type": "image", + "image_path": "cd0239ea36c858749982b1139b5d1298c1d48296c36ee3f0dfde4ee0a09b691b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 234, + 482, + 278 + ], + "lines": [ + { + "bbox": [ + 130, + 234, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 482, + 278 + ], + "type": "text", + "content": "Fig. 3: An example of track sampling augmentation over three consecutive frames from the nuScenes dataset. The original trajectories and sampled trajectories are shown in blue and orange, respectively. Only a subset of the tracks and the third LiDAR scan are shown for visual clarity. The older boxes are shown with transparency." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 116, + 419, + 224 + ], + "blocks": [ + { + "bbox": [ + 309, + 116, + 419, + 224 + ], + "lines": [ + { + "bbox": [ + 309, + 116, + 419, + 224 + ], + "spans": [ + { + "bbox": [ + 309, + 116, + 419, + 224 + ], + "type": "image", + "image_path": "3bd3bc5c2b0c7d9f1126b7739e7e1008337f8090d5a86ba3dea183404952ae29.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 303, + 294, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 294, + 316 + ], + "type": "text", + "content": "3.3 Ground Truth Assignment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 324, + 482, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 420 + ], + "type": "text", + "content": "The ground truth assignment of the decoder predictions, " + }, + { + "bbox": [ + 130, + 324, + 482, + 420 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_t^D" + }, + { + "bbox": [ + 130, + 324, + 482, + 420 + ], + "type": "text", + "content": ", and the refined predictions, " + }, + { + "bbox": [ + 130, + 324, + 482, + 420 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_t^R" + }, + { + "bbox": [ + 130, + 324, + 482, + 420 + ], + "type": "text", + "content": ", are both done in the same manner using a two-stage assignment process. The first stage involves matching the track queries based on their previously matched ground truth objects [18]. If the previously matched ground truth track is not present in the current time step, such as in the case of a track leaving the scene, then the track query prediction is assigned to no object. False positive track queries were never assigned a ground truth track previously, so they are also assigned to no object." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 421, + 482, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 482, + 481 + ], + "type": "text", + "content": "The second stage of the assignment involves matching the proposal queries to the unmatched ground truths. The unmatched ground truths would include new objects in the scene and missed objects from previous time steps. A Hungarian matcher is used to compute the optimal assignment between the unmatched queries and ground truths based on an association cost [10]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 500, + 315, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 315, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 315, + 511 + ], + "type": "text", + "content": "3.4 Track Sampling Augmentation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": "To improve the convergence rate and performance, we introduce track sampling augmentation during training. Inspired by the object sampling augmentation [25], our track sampling involves adding additional tracks into a training clip from a database of tracks in a temporally consistent manner, as illustrated in Fig. 3. The original object sampling augmentation does not account for temporal consistency, and applying it naively to our multi-frame training scheme would result in disjointed track instances during training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": "First, the training set is preprocessed to generate a database of ground truth tracks, including bounding boxes, object classes, track IDs, and LiDAR points contained in the bounding boxes. To perform track sample augmentation on a training clip, a set of tracks is randomly sampled, avoiding track IDs that overlap with the existing ground truth objects in the scene. This enforces unique IDs for" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "text", + "content": "JDT3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "the tracks in the sequence. To inject a track, " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " consecutive ground-truth boxes are randomly selected from each sampled track. In each training frame, the LiDAR points corresponding to the sampled objects are used to replace regions in the original point cloud. The objects are injected in chronological order over the " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " frames. A collision check is performed to ensure that there are no collisions between injected objects and the original ground truth objects in the scene. In the case of a collision, the would-be sampled object is pruned from that frame, simulating an occlusion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 248 + ], + "type": "text", + "content": "The benefits of track sampling mirror those of object sampling [25], with the extra ground truth objects in the scene providing more samples to learn from in a single frame. This results in a stronger training signal and more varied scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 266, + 308, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 266, + 308, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 266, + 308, + 278 + ], + "type": "text", + "content": "3.5 Network Training and Losses" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 286, + 482, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 286, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 482, + 323 + ], + "type": "text", + "content": "To train JDT3D, " + }, + { + "bbox": [ + 130, + 286, + 482, + 323 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 286, + 482, + 323 + ], + "type": "text", + "content": " LiDAR frames are passed sequentially to the model as a training sample. The loss for a given frame, " + }, + { + "bbox": [ + 130, + 286, + 482, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_i" + }, + { + "bbox": [ + 130, + 286, + 482, + 323 + ], + "type": "text", + "content": ", is computed as the weighted sum of the following losses:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 195, + 332, + 481, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 332, + 481, + 360 + ], + "spans": [ + { + "bbox": [ + 195, + 332, + 481, + 360 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i} = \\lambda_ {h} \\mathcal {L} _ {h} + \\lambda_ {f} \\mathcal {L} _ {f} + \\sum_ {m = \\{D, R \\}} \\left(\\lambda_ {r e g} ^ {m} \\mathcal {L} _ {r e g} ^ {m} + \\lambda_ {c l s} ^ {m} \\mathcal {L} _ {c l s} ^ {m}\\right) \\tag {1}", + "image_path": "f194d65f933d590c714d702efb6207758878a1d062370dafe5c71f37db446cc4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "text", + "content": "The superscripts " + }, + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "text", + "content": " correspond to predictions output by " + }, + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^D" + }, + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^R" + }, + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}^{m}" + }, + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "text", + "content": " is the L1 loss over the bounding boxes, and " + }, + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}^{m}" + }, + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "text", + "content": " is the focal loss [14] over the classification predictions. " + }, + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_h" + }, + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "text", + "content": " is the Gaussian Focal Loss on the predicted object heatmaps [11,26], and " + }, + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "text", + "content": " is the L1 loss over the motion trajectory predicted by the future reasoning module [19]. Each loss has an associated scalar weight term, " + }, + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 368, + 482, + 451 + ], + "type": "text", + "content": ". The total loss is computed as the sum of each individual frame loss:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 276, + 451, + 481, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 451, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 276, + 451, + 481, + 483 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c l i p}} = \\sum_ {i = 1} ^ {L} \\mathcal {L} _ {i} \\tag {2}", + "image_path": "edb1d13c39612ce52386005867cb78637dfceddc199ae27b727d8c5a3613dc67.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 499, + 230, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 499, + 230, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 499, + 230, + 514 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 525, + 264, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 264, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 264, + 537 + ], + "type": "text", + "content": "4.1 Dataset and Metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": "JDT3D is trained and evaluated on the nuScenes dataset [2], a multi-modal autonomous driving dataset containing data from six multi-view cameras, five radars, and a 360-degree LiDAR. The dataset contains a total of 40,157 frames across 1000 scenes, of which 850 scenes are annotated for training (700) and validation (150) at a frame rate of " + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": ". We train on the seven object classes annotated for the tracking task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "To evaluate JDT3D, the average multi-object tracking accuracy (AMOTA) and average multi-object tracking precision (AMOTP) tracking metrics [24] are computed across the seven tracking classes of nuScenes, evaluating the tracking accuracy and tracking position misalignment, respectively. Secondary metrics are" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "B. Cheong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 157, + 470, + 274 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "type": "text", + "content": "Table 1: A comparison of 3D multi-object tracking methods evaluated on the nuScenes [2] test set. Bold represents the best results among the Transformer-based TBA methods, while Underlined numbers are the best overall." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 157, + 470, + 274 + ], + "lines": [ + { + "bbox": [ + 141, + 157, + 470, + 274 + ], + "spans": [ + { + "bbox": [ + 141, + 157, + 470, + 274 + ], + "type": "table", + "html": "
MethodTBAModalityAMOTA ↑AMOTP ↓FP ↓FN ↓IDS ↓
AB3DMOT [24]L0.1511.50115088757309027
Chui et al. [5]L0.5500.7981753333216950
CenterPoint [26]L0.6380.5551861222928760
SimpleTrack [20]L0.6680.5501751423451575
PF-Track [19]✓C0.4341.2521904842758249
DQTrack [13]✓C0.5231.096--1204
MotionTrack-L [29]✓L0.510.99--9705
MotionTrack-LC [29]✓L+C0.550.871--8716
JDT3D (ours)✓L0.5740.8371115229919254
", + "image_path": "04def74aad5ef068e27f800fd22e8d1895f187d20928533edb95b6541e9e37fe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 297, + 481, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 346 + ], + "type": "text", + "content": "also computed by the nuScenes MOT benchmark [2], such as the number of false positives, false negatives, and identity switches. We also evaluate the detection performance across the tracking classes, with the mean average precision (mAP) and nuScenes Detection Score (NDS) metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 364, + 279, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 364, + 279, + 376 + ], + "spans": [ + { + "bbox": [ + 133, + 364, + 279, + 376 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 385, + 482, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 482, + 577 + ], + "type": "text", + "content": "VoxelNet [25, 32] is used as the LiDAR backbone to extract the BEV feature map. For each frame, the top 200 predictions from the heatmap are used to initialize the proposal queries. The training clip length " + }, + { + "bbox": [ + 132, + 385, + 482, + 577 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 132, + 385, + 482, + 577 + ], + "type": "text", + "content": " is set to 3 frames. The confidence threshold for track queries " + }, + { + "bbox": [ + 132, + 385, + 482, + 577 + ], + "type": "inline_equation", + "content": "\\tau_{pass}" + }, + { + "bbox": [ + 132, + 385, + 482, + 577 + ], + "type": "text", + "content": " is set to 0.4 for training and inference. The Hungarian matching cost for ground truth assignment consists of the sum of the L1 cost on the predicted box and the focal loss [14] on the predicted class confidence. All detections are generated in the car ego-frame, so the object centers of the track queries are updated to their relative positions in the current training frame based on the car's ego-motion between time steps. A pretrained VoxelNet backbone from BEVFusion [15] is used to initialize the feature extractor. To train JDT3D, the model is first trained on the detection task using " + }, + { + "bbox": [ + 132, + 385, + 482, + 577 + ], + "type": "inline_equation", + "content": "L = 1" + }, + { + "bbox": [ + 132, + 385, + 482, + 577 + ], + "type": "text", + "content": " for 20 epochs. It is then further trained on the tracking task using " + }, + { + "bbox": [ + 132, + 385, + 482, + 577 + ], + "type": "inline_equation", + "content": "L = 3" + }, + { + "bbox": [ + 132, + 385, + 482, + 577 + ], + "type": "text", + "content": " for another 20 epochs. All experiments were trained and evaluated on eight NVIDIA V100 GPUs using random flipping, rotation, translation, and scaling on the training sequence, and track sampling augmentation, unless otherwise specified. JDT3D performs inference at " + }, + { + "bbox": [ + 132, + 385, + 482, + 577 + ], + "type": "inline_equation", + "content": "3.85\\mathrm{Hz}" + }, + { + "bbox": [ + 132, + 385, + 482, + 577 + ], + "type": "text", + "content": " on a Tesla V100." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 596, + 228, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 596, + 228, + 607 + ], + "spans": [ + { + "bbox": [ + 133, + 596, + 228, + 607 + ], + "type": "text", + "content": "4.3 Main Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 481, + 666 + ], + "type": "text", + "content": "Our main results compare JDT3D with baseline TBD methods and relevant TBA methods evaluated on the nuScenes test set [2], summarized in Tab. 1. JDT3D outperforms the current state-of-the-art LiDAR-based TBA methods, with a relative improvement of " + }, + { + "bbox": [ + 132, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "12.5\\%" + }, + { + "bbox": [ + 132, + 617, + 481, + 666 + ], + "type": "text", + "content": " on the AMOTA metric and " + }, + { + "bbox": [ + 132, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "15.5\\%" + }, + { + "bbox": [ + 132, + 617, + 481, + 666 + ], + "type": "text", + "content": " on the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "JDT3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 186, + 146, + 427, + 209 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: Ablation on the track sampling augmentation, reported on the nuScenes validation set. The mAP metric is computed on the seven tracking categories." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 186, + 146, + 427, + 209 + ], + "lines": [ + { + "bbox": [ + 186, + 146, + 427, + 209 + ], + "spans": [ + { + "bbox": [ + 186, + 146, + 427, + 209 + ], + "type": "table", + "html": "
Track Sampling RateAMOTA ↑AMOTP ↓IDS ↓mAP ↑
0.00.3141.3773880.225
0.250.6000.8642100.3867
0.500.6220.8162030.400
1.00.6210.7901500.399
", + "image_path": "debc44cdbd77910d9f471ee093348e47734e1f0f9ae4e8c647cd2553d3b26e96.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 229, + 479, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 229, + 479, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 229, + 479, + 300 + ], + "type": "text", + "content": "AMOTP metric on the test split. Moreover, with LiDAR input only, JDT3D also outperforms the fusion-based MotionTrack-LC [29] with a relative improvement of " + }, + { + "bbox": [ + 130, + 229, + 479, + 300 + ], + "type": "inline_equation", + "content": "4.36\\%" + }, + { + "bbox": [ + 130, + 229, + 479, + 300 + ], + "type": "text", + "content": " on AMOTA and " + }, + { + "bbox": [ + 130, + 229, + 479, + 300 + ], + "type": "inline_equation", + "content": "3.79\\%" + }, + { + "bbox": [ + 130, + 229, + 479, + 300 + ], + "type": "text", + "content": " on AMOTP on the test set. While JDT3D underperforms compared to baseline TBD methods on most metrics, it achieves the lowest number of ID switches among LiDAR-only methods, indicating a strong track coherence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 301, + 480, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 480, + 373 + ], + "type": "text", + "content": "JDT3D has significantly more false negatives compared to TBD methods, but fewer false positives. The potential sources of false negatives are either due to missed detections or tracks diverging from the ground truths. However, the latter case would result in both high false negatives and false positives. Based on the low false positive results from JDT3D, we conclude that the false negatives are mainly due to missed detections." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 390, + 246, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 390, + 246, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 390, + 246, + 402 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 409, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 578 + ], + "type": "text", + "content": "Track Sampling. By comparing the performance of JDT3D on the nuScenes validation set with and without track sampling augmentation, we verify that sampling extra tracks into a training sequence significantly improves the convergence rate and the final performance. The comparison summarized in Tab. 2 shows a " + }, + { + "bbox": [ + 130, + 409, + 482, + 578 + ], + "type": "inline_equation", + "content": "97\\%" + }, + { + "bbox": [ + 130, + 409, + 482, + 578 + ], + "type": "text", + "content": " relative improvement on the AMOTA metric from track sampling augmentation compared to no track sampling, with a similar trend across all other metrics, including detection metrics such as the mAP. We also analyzed the effect of varying the number of objects that are sampled by reducing the sampling frequency to 0.5 and 0.25 of the nominal rate. We observe that the AMOTA and mAP metrics plateau after a sampling rate of 0.5, although the AMOTP and IDS metrics see continued improvement with increased track sampling. The benefits of implementing track sampling are expected to generalize to other LiDAR-based TBA methods, but there are currently no available implementations to verify this hypothesis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Training Clip Length. To validate our choice of clip length, the performances of JDT3D models trained with two to four frame sequences were compared and summarized in Tab. 3. The best performance is achieved when trained over " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "L = 3" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " frames, with a significant jump when moving from " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "L = 2" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "L = 3" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". We hypothesize that the training process benefits from the longer-term, temporally consistent information, resulting in improved tracking performance. Moving from" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "B. Cheong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 196, + 146, + 415, + 196 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 479, + 136 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 479, + 136 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 479, + 136 + ], + "type": "text", + "content": "Table 3: Ablation on the training sequence length, " + }, + { + "bbox": [ + 130, + 114, + 479, + 136 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 114, + 479, + 136 + ], + "type": "text", + "content": ", reported on the nuScenes validation set." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 196, + 146, + 415, + 196 + ], + "lines": [ + { + "bbox": [ + 196, + 146, + 415, + 196 + ], + "spans": [ + { + "bbox": [ + 196, + 146, + 415, + 196 + ], + "type": "table", + "html": "
Clip Length (L)AMOTA ↑AMOTP ↓IDS ↓mAP ↑
20.3971.2944460.214
30.6210.7901500.399
40.5990.7831730.392
", + "image_path": "e848affa980486f41d8d901921bd33e84249aedf941a95b8ae2e0440db64f9f1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 187, + 250, + 423, + 289 + ], + "blocks": [ + { + "bbox": [ + 130, + 206, + 480, + 239 + ], + "lines": [ + { + "bbox": [ + 130, + 206, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 206, + 480, + 239 + ], + "type": "text", + "content": "Table 4: Ablation on the query propagation criteria, reported on the nuScenes [2] validation set. Passing queries with high confidence encourages the re-detection of poor detections and handling of false positives and false negatives." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 250, + 423, + 289 + ], + "lines": [ + { + "bbox": [ + 187, + 250, + 423, + 289 + ], + "spans": [ + { + "bbox": [ + 187, + 250, + 423, + 289 + ], + "type": "table", + "html": "
Query PropagationAMOTA↑ AMOTP↓ Recall↑ IDS
Matched to a GT0.4640.9170.51182
High Confidence0.6210.7900.669150
", + "image_path": "473d0e6a7b7837f9d950ccf6200f5cdc45fa7266a8abc43e446fccd8e1aa3445.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 311, + 479, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 479, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 479, + 335 + ], + "type": "inline_equation", + "content": "L = 3" + }, + { + "bbox": [ + 130, + 311, + 479, + 335 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 311, + 479, + 335 + ], + "type": "inline_equation", + "content": "L = 4" + }, + { + "bbox": [ + 130, + 311, + 479, + 335 + ], + "type": "text", + "content": ", the performance drops slightly while also significantly increasing the train time and required memory." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 351, + 480, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 480, + 471 + ], + "type": "text", + "content": "Query Propagation Method. In Tab. 4, we compare different query propagation methods; the first method is passing queries that have been matched to a ground truth [18, 19], and the second is passing based on the " + }, + { + "bbox": [ + 130, + 351, + 480, + 471 + ], + "type": "inline_equation", + "content": "\\tau_{pass}" + }, + { + "bbox": [ + 130, + 351, + 480, + 471 + ], + "type": "text", + "content": " threshold for high confidence predictions as described in Sec. 3.2. Results show that using confidence as a filtering criterion during training enhances the overall tracking metrics AMOTA and AMOTP by a large margin and significantly reduces the number of false positives and false negatives while increasing the number of true positives. While the ID switch errors are higher, this increase is less than " + }, + { + "bbox": [ + 130, + 351, + 480, + 471 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 351, + 480, + 471 + ], + "type": "text", + "content": " relative to the increase in true positives. Thus, we verify that using confidence as the query propagation criterion is more appropriate for the TBA paradigm." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 472, + 479, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 479, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 479, + 507 + ], + "type": "text", + "content": "In addition, we show the generalizability of this query propagation method by performing the same ablation on PF-Track. The results in Tab. 5 show improvements across most metrics, with a similar behaviour of more IDS errors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 524, + 479, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 524, + 479, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 524, + 479, + 561 + ], + "type": "text", + "content": "Comparison with TBD Baseline. We compare the performance between our TBA framework and baseline TBD methods such as the CenterPoint greedy-based tracking method [26] and SimpleTrack [20]. To compare our tracking" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 189, + 624, + 425, + 663 + ], + "blocks": [ + { + "bbox": [ + 130, + 581, + 479, + 613 + ], + "lines": [ + { + "bbox": [ + 130, + 581, + 479, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 479, + 613 + ], + "type": "text", + "content": "Table 5: An ablation comparing the original ground truth-based propagation and our confidence-based propagation on PF-Track-S [19] evaluated on the nuScenes [2] validation set." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 189, + 624, + 425, + 663 + ], + "lines": [ + { + "bbox": [ + 189, + 624, + 425, + 663 + ], + "spans": [ + { + "bbox": [ + 189, + 624, + 425, + 663 + ], + "type": "table", + "html": "
Query PropagationAMOTA ↑AMOTP ↓Recall ↑IDS ↓
Matched to a GT0.4081.3430.507166
High Confidence0.4261.3160.558306
", + "image_path": "d043737515186980b35145d0e7b4ab5f027d487633915926f2a566f14f23d553.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "text", + "content": "JDT3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 209, + 146, + 401, + 198 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 6: Comparison with \"Tracking by Detection\" on the nuScenes validation set, using the detections generated by JDT3D." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 209, + 146, + 401, + 198 + ], + "lines": [ + { + "bbox": [ + 209, + 146, + 401, + 198 + ], + "spans": [ + { + "bbox": [ + 209, + 146, + 401, + 198 + ], + "type": "table", + "html": "
Tracking MethodAMOTA ↑AMOTP ↓IDS ↓
CenterPoint [26]0.6080.825491
SimpleTrack [20]0.5730.8775585
JDT3D (Ours)0.6210.790150
", + "image_path": "541a0e67fc43415bfedc611710944d71f37129363778697280bbc4e95100f55c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 241, + 251, + 369, + 290 + ], + "blocks": [ + { + "bbox": [ + 132, + 208, + 480, + 240 + ], + "lines": [ + { + "bbox": [ + 132, + 208, + 480, + 240 + ], + "spans": [ + { + "bbox": [ + 132, + 208, + 480, + 240 + ], + "type": "text", + "content": "Table 7: Comparison of detection performance between the single-frame pretraining and the three-frame fine-tuning. Evaluated on the seven tracking classes on the nuScenes validation set. The best results are typeset in boldface." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 241, + 251, + 369, + 290 + ], + "lines": [ + { + "bbox": [ + 241, + 251, + 369, + 290 + ], + "spans": [ + { + "bbox": [ + 241, + 251, + 369, + 290 + ], + "type": "table", + "html": "
JDT3D variantmAP ↑NDS ↑
Single-frame0.6050.549
Three-frame0.3990.502
", + "image_path": "8513fbb5476ae0ccc3477c9b0a94ede906d8e38cf4ac907b43ce2481828e6bff.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 312, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 480, + 384 + ], + "type": "text", + "content": "method to the TBD trackers independent of detection performance, we pass the detections generated by JDT3D to the tracker and evaluate its tracking performance compared to our joint approach. The results in Tab. 6 demonstrate clear advantages of our joint end-to-end approach that offers temporal consistency and data association and reaffirm that the gap in tracking performance lies in the missed detections." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 403, + 276, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 276, + 414 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 276, + 414 + ], + "type": "text", + "content": "4.5 Detection Performance" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 423, + 480, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 423, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 480, + 495 + ], + "type": "text", + "content": "As mentioned in Sec. 4.2, JDT3D is pre-trained to perform single-frame detection and fine-tuned to perform tracking on sequences of multiple LiDAR frames. While the pre-training step achieves reasonable detection performance, the performance drops significantly after fine-tuning, as seen in Tab. 7. This indicates that the current TBA problem formulation may not be the best alignment of the detection and tracking objectives." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 514, + 352, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 352, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 352, + 525 + ], + "type": "text", + "content": "4.6 Proposal and Track Query Imbalances" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 533, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 628 + ], + "type": "text", + "content": "Tab. 8 shows further analysis of the JDT3D outputs. During training, the positively matched proposal queries, or untracked newborn objects, are more confident in the first frame than in subsequent frames. Also, the positively matched track queries, or previously tracked objects, tend to be more confident than the newborn objects. This indicates that the model has a strong bias toward previously tracked objects. This bias could lead to more missed objects and fewer initialized tracks, explaining the high number of false negatives indicated in Tab. 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 630, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 630, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 630, + 480, + 665 + ], + "type": "text", + "content": "One hypothesis for this behaviour is that this is caused by the imbalance in training samples assigned to each type of query. Tab. 9 shows the average distribution of true positive predictions of both proposal and track queries in" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "B. Cheong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 190, + 146, + 423, + 186 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 8: The average confidence of positively matched proposal and track predictions over a three-frame training sequence. There are no track queries in the first frame." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 190, + 146, + 423, + 186 + ], + "lines": [ + { + "bbox": [ + 190, + 146, + 423, + 186 + ], + "spans": [ + { + "bbox": [ + 190, + 146, + 423, + 186 + ], + "type": "table", + "html": "
Avg. ConfidenceFrame 1Frame 2Frame 3
Positively Matched Proposals0.6690.4090.387
Track Predictionsn/a0.7590.760
", + "image_path": "89deea2f500553d138f12ca7e963d29b88be34f37eb9cc1874b2cfd1e426fe76.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 192, + 228, + 418, + 267 + ], + "blocks": [ + { + "bbox": [ + 132, + 195, + 480, + 218 + ], + "lines": [ + { + "bbox": [ + 132, + 195, + 480, + 218 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 480, + 218 + ], + "type": "text", + "content": "Table 9: The distribution of positive and negative matches for proposal queries and track queries in the second and third frames of training." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 192, + 228, + 418, + 267 + ], + "lines": [ + { + "bbox": [ + 192, + 228, + 418, + 267 + ], + "spans": [ + { + "bbox": [ + 192, + 228, + 418, + 267 + ], + "type": "table", + "html": "
PQ PredictionsTQ Predictions
Avg Positive Matches4 (2%)47 (94%)
Avg Negative Matches196 (98%)3 (6%)
", + "image_path": "a7934a657cb4e4410e2b2361fa1524df21db095de0e94605334bfd3b89a4bf07.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 289, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 480, + 361 + ], + "type": "text", + "content": "the second and third frames of training. Note that the proportion of positively matched proposal queries is significantly lower than for track queries. This is somewhat expected since the average track birthrate per frame is 0.7, so most objects in a scene were present in the previous frame. This imbalance in positive training samples between the proposal and track queries could explain the confidence imbalance observed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 377, + 217, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 377, + 217, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 377, + 217, + 390 + ], + "type": "text", + "content": "5 Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 402, + 480, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 480, + 475 + ], + "type": "text", + "content": "Based on our analysis from Tab. 6 we show that the reason LiDAR-based TBA methods continue to underperform compared to TBD methods is due to poorer detection performance. In addition, we note that there is a significant drop in the detection performance when fine-tuning the model on multiple frames in Tab. 7. In the following sections, we discuss hypotheses to explain the observed results and propose potential improvements to build on JDT3D." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 491, + 263, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 263, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 263, + 502 + ], + "type": "text", + "content": "5.1 Temporal Confusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "type": "text", + "content": "Our hypothesis for the observed drop in performance in Tab. 7 lies in the fact that the problem formulation of TBA is more complex than query-based detection. In query-based detection, the handling of the queries by the decoder is straightforward: if the query represents an object in the scene, it should output a high-confidence bounding box from the query. On the other hand, in TBA, the handling of queries is conditional: depending on whether the decoder receives a track query that represents the same object, the decoder needs to determine whether the incoming proposal query is a duplicate that should have a low confidence or a unique object that should have a high confidence, as illustrated in Fig. 4. This complex interaction results in \"temporal confusion\" such that the expected model behaviour is different depending on the query inputs. We hypothesize that the reason temporal confusion has not been an issue in vision-based TBA methods [18,19,30] is due to their larger decoder size enabling more" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "text", + "content": "JDT3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 187, + 128, + 427, + 169 + ], + "blocks": [ + { + "bbox": [ + 187, + 128, + 427, + 169 + ], + "lines": [ + { + "bbox": [ + 187, + 128, + 427, + 169 + ], + "spans": [ + { + "bbox": [ + 187, + 128, + 427, + 169 + ], + "type": "image", + "image_path": "996dca4637e4186161c434653079f8c5d4346437748636108c52e20c99fa6d55.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 186, + 482, + 221 + ], + "lines": [ + { + "bbox": [ + 130, + 186, + 482, + 221 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 482, + 221 + ], + "type": "text", + "content": "Fig. 4: An example of temporal confusion, where the decoder must handle the same proposal query differently based on the presence of track queries. In Case 1, the proposal query should be a positive prediction, while in Case 2, it should be a negative prediction." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 244, + 480, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 244, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 480, + 281 + ], + "type": "text", + "content": "complex reasoning about the queries. To test this hypothesis, further optimizations would be needed to enable a larger decoder with a LiDAR backbone due to hardware constraints." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 299, + 282, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 299, + 282, + 312 + ], + "spans": [ + { + "bbox": [ + 131, + 299, + 282, + 312 + ], + "type": "text", + "content": "5.2 Potential Improvements" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 320, + 480, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 480, + 343 + ], + "type": "text", + "content": "With these results in mind, we propose two potential directions that can be further explored to bridge the gap between JDT and TBD methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 344, + 480, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 480, + 403 + ], + "type": "text", + "content": "Firstly, based on the ablation performed to evaluate the training clip size, we observed that three-frame training significantly improved performance over two-frame training, although extending to four frames reduced performance slightly. Further exploration of the effect of clip length on tracking performance may yield insights on how to better leverage longer term temporal information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 404, + 481, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 481, + 440 + ], + "type": "text", + "content": "Secondly, optimizations to the model size and memory footprint would enable a larger and more complex decoder, that could potentially address the temporal confusion hypothesis." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 460, + 220, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 460, + 220, + 472 + ], + "spans": [ + { + "bbox": [ + 131, + 460, + 220, + 472 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "This paper presents JDT3D, a novel LiDAR tracking method that makes significant progress in closing the gap between TBD and TBA methods. JDT3D is trained using track sampling augmentation, a generalizable data augmentation method that injects extra long-term tracks into training samples. We show that track sampling augmentation significantly improves training performance and convergence, and we hope that future LiDAR-based trackers can leverage these findings. As well, we show that the query propagation scheme during training also has a significant impact on the performance, showing that using a confidence threshold to filter low confidence predictions improves tracking performance greatly. Based on our experiments, we hypothesize that the complexity of tracking-by-attention combined with the memory constraints imposed by LiDAR feature extractors limiting the size of transformer decoders is the cause of the drop in detection performance between single and multi-frame variants of JDT3D. We propose further areas of exploration including exploring different training sequence lengths and enabling more decoder layers." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "B. Cheong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 139, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 139, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 481, + 183 + ], + "type": "text", + "content": "1. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1090-1099 (June 2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 184, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 481, + 237 + ], + "type": "text", + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 11618-11628 (2019), https://apisemantic scholar.org/CorpusID: 85517967" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 481, + 270 + ], + "type": "text", + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European Conference on Computer Vision. pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 271, + 481, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 481, + 313 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 481, + 313 + ], + "type": "text", + "content": "4. Chen, Y., Yu, Z., Chen, Y., Lan, S., Anandkumar, A., Jia, J., Alvarez, J.M.: Focalformer3d: Focusing on hard instance for 3d object detection. ArXiv abs/2308.04556 (2023), https://api-semanticscholar.org/CorpusID: 260736073" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 315, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 481, + 336 + ], + "type": "text", + "content": "5. Chiu, H.k., Prioletti, A., Li, J., Bohg, J.: Probabilistic 3d multi-object tracking for autonomous driving. arXiv preprint arXiv:2001.05673 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 337, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 481, + 369 + ], + "type": "text", + "content": "6. Hahner, M., Dai, D., Liniger, A., Gool, L.V.: Quantifying data augmentation for lidar based 3d object detection. CoRR abs/2004.01643 (2020), https://arxiv.org/abs/2004.01643" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 370, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 481, + 413 + ], + "type": "text", + "content": "7. Hu, Y., Yang, J., Chen, L., Li, K., Sima, C., Zhu, X., Chai, S., Du, S., Lin, T., Wang, W., et al.: Planning-oriented autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17853-17862 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 414, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 481, + 446 + ], + "type": "text", + "content": "8. Huang, T.E., Liu, Y., Van Gool, L., Yu, F.: Video task decathlon: Unifying image and video tasks in autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8647-8657 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 447, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 481, + 500 + ], + "type": "text", + "content": "9. Koh, J., Kim, J., Yoo, J.H., Kim, Y., Kum, D., Choi, J.W.: Joint 3d object detection and tracking using spatio-temporal representation of camera image and lidar point clouds. Proceedings of the AAAI Conference on Artificial Intelligence 36(1), 1210-1218 (Jun 2022). https://doi.org/10.1609/aaai.v36i1.20007, https://ojs.aaai.org/index.php/AAAI/article/view/20007" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 481, + 533 + ], + "type": "text", + "content": "0. Kuhn, H.W.: The hungarian method for the assignment problem. Naval Research Logistics (NRL) 52 (1955), https://api-semanticscholar.org/CorpusID: 9426884" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 481, + 566 + ], + "type": "text", + "content": "1. Law, H., Deng, J.: Cornernet: Detecting objects as paired keypoints. Int. J. Comput. Vis. 128(3), 642-656 (2020). https://doi.org/10.1007/S11263-019-01204-1, https://doi.org/10.1007/s11263-019-01204-1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 599 + ], + "type": "text", + "content": "2. Li, X., Xie, T., Liu, D., Gao, J., Dai, K., Jiang, Z., Zhao, L., Wang, K.: Poly-mot: A polyhedral framework for 3d multi-object tracking. ArXiv abs/2307.16675 (2023), https://api.sementicscholar.org/CorpusID:260334560" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "text", + "content": "3. Li, Y., Yu, Z., Philion, J., Anandkumar, A., Fidler, S., Jia, J., Alvarez, J.: End-to-end 3d tracking with decoupled queries. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "4. Lin, T.Y., Goyal, P., Girshick, R.B., He, K., Dollar, P.: Focal loss for dense object detection. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 2999-3007 (2017), https://api-semanticscholar.org/CorpusID:47252984" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "text", + "content": "JDT3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "15. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. 2023 IEEE International Conference on Robotics and Automation (ICRA) pp. 2774-2781 (2022), https://apisemantic scholar.org/CorpusID:249097415" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "type": "text", + "content": "16. Luo, C., Yang, X., Yuille, A.: Exploring simple 3d multi-object tracking for autonomous driving. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10488-10497 (October 2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 481, + 226 + ], + "type": "text", + "content": "17. Mahmoud, A., Hu, J.S.K., Waslander, S.L.: Dense voxel fusion for 3d object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 663-672 (January 2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 227, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 227, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 227, + 481, + 270 + ], + "type": "text", + "content": "18. Meinhardt, T., Kirillov, A., Leal-Taixe, L., Feichtenhofer, C.: Trackformer: Multiobject tracking with transformers. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8834-8844 (2022). https://doi.org/10.1109/CVPR52688.2022.00864" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 271, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 271, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 133, + 271, + 481, + 315 + ], + "type": "text", + "content": "19. Pang, Z., Li, J., Tokmakov, P., Chen, D., Zagoruyko, S., Wang, Y.X.: Standing between past and future: Spatio-temporal modeling for multi-camera 3d multi-object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17928-17938 (June 2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 316, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 481, + 369 + ], + "type": "text", + "content": "20. Pang, Z., Li, Z., Wang, N.: SimpleTrack: Understanding and Rethinking 3D Multiobject Tracking. In: Karlinsky, L., Michaeli, T., Nishino, K. (eds.) Computer Vision - ECCV 2022 Workshops. pp. 680-696. Lecture Notes in Computer Science, Springer Nature Switzerland (2021). https://doi.org/10.1007/978-3-031-25056-9_43" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "21. Sun, P., Cao, J., Jiang, Y., Zhang, R., Xie, E., Yuan, Z., Wang, C., Luo, P.: Transtrack: Multiple-object tracking with transformer. arXiv preprint arXiv: 2012.15460 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 403, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 468 + ], + "type": "text", + "content": "22. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017), https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 468, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 521 + ], + "type": "text", + "content": "23. Wang, L., Zhang, X.N., Qin, W., Li, X., Yang, L., Li, Z., Zhu, L., Wang, H., Li, J., Liu, H.: Camo-mot: Combined appearance-motion optimization for 3d multi-object tracking with camera-lidar fusion. IEEE Transactions on Intelligent Transportation Systems 24, 11981-11996 (2022), https://api_semanticscholar.org/CorpusID: 252089635" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 522, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 566 + ], + "type": "text", + "content": "24. Weng, X., Wang, J., Held, D., Kitani, K.: 3d multi-object tracking: A baseline and new evaluation metrics. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 10359-10366 (2020). https://doi.org/10.1109/IROS45743.2020.9341164" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "25. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors (Basel, Switzerland) 18 (2018), https://api_semanticscholar.org/CorpusID:52957856" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "26. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11784-11793 (June 2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "27. Zeng, F., Dong, B., Wang, T., Chen, C., Zhang, X., Wei, Y.: Motr: End-to-end multiple-object tracking with transformer. ArXiv abs/2105.03247 (2021), https://api-semanticscholar.org/CorpusID:234096063" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "B. Cheong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 380 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "28. Zeng, Y., Ma, C., Zhu, M., Fan, Z., Yang, X.: Cross-modal 3d object detection and tracking for auto-driving. In: IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2021, Prague, Czech Republic, September 2021. IEEE (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 482, + 205 + ], + "type": "text", + "content": "29. Zhang, C., Zhang, C., Guo, Y., Chen, L., Happold, M.: Motiontrack: End-to-end transformer-based multi-object tracking with lidar-camera fusion. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) pp. 151-160 (2023), https://apisemanticscholar.org/CorpusID:259286967" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 482, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 247 + ], + "type": "text", + "content": "30. Zhang, T., Chen, X., Wang, Y., Wang, Y., Zhao, H.: Muitr3d: A multi-camera tracking framework via 3d-to-2d queries. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 4536-4545 (2022). https://doi.org/10.1109/CVPRW56347.2022.00500" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 249, + 482, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 292 + ], + "type": "text", + "content": "31. Zhang, Y., Wang, T., Zhang, X.: Motrv2: Bootstrapping end-to-end multi-object tracking by pretrained object detectors. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 22056-22065 (2022), https://apisemantic scholar.org/CorpusID:253581846" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 293, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 335 + ], + "type": "text", + "content": "32. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition pp. 4490-4499 (2017), https://apisemantic scholar.org/CorpusID:42427078" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 335, + 482, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 380 + ], + "type": "text", + "content": "33. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: deformable transformers for end-to-end object detection. In: 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 (2021)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 415, + 91, + 447, + 100 + ], + "type": "text", + "content": "JDT3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography/2ac99c58-8fa1-4a8d-b568-c7870f80ab53_content_list.json b/2024/Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography/2ac99c58-8fa1-4a8d-b568-c7870f80ab53_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6062587aab5c29d2211acb889965343ad4c03d9b --- /dev/null +++ b/2024/Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography/2ac99c58-8fa1-4a8d-b568-c7870f80ab53_content_list.json @@ -0,0 +1,1608 @@ +[ + { + "type": "text", + "text": "Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography", + "text_level": 1, + "bbox": [ + 259, + 140, + 743, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kailai Zhou1, Lijing Cai1, Yibo Wang1, Mengya Zhang1, Bihan Wen3, Qiu Shen†,2, and Xun Cao1,2", + "bbox": [ + 246, + 234, + 754, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 School of Electronic Science and Engineering, Nanjing University, Nanjing, China", + "bbox": [ + 225, + 276, + 774, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Key Laboratory of Optoelectronic Devices and Systems with Extreme", + "bbox": [ + 259, + 291, + 740, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Performances of MOE, Nanjing University, Nanjing, China", + "bbox": [ + 305, + 305, + 696, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Nanyang Technological University, Singapore", + "bbox": [ + 343, + 319, + 658, + 334 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{calayzhou, cailijing, ybwang, 522023230161}@smail.nju.edu.cn,", + "bbox": [ + 263, + 334, + 738, + 348 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "bihan.wen@ntu.edu.sg, {caoxun, shenqiu}@nju.edu.cn", + "bbox": [ + 321, + 349, + 679, + 361 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The integration of miniaturized spectrometers into mobile devices offers new avenues for image quality enhancement and facilitates novel downstream tasks. However, the broader application of spectral sensors in mobile photography is hindered by the inherent complexity of spectral images and the constraints of spectral imaging capabilities. To overcome these challenges, we propose a joint RGB-Spectral decomposition model guided enhancement framework, which consists of two steps: joint decomposition and prior-guided enhancement. Firstly, we leverage the complementarity between RGB and Low-resolution Multi-Spectral Images (Lr-MSI) to predict shading, reflectance, and material semantic priors. Subsequently, these priors are seamlessly integrated into the established HDRNet to promote dynamic range enhancement, color mapping, and grid expert learning, respectively. Additionally, we construct a high-quality Mobile-Spec dataset to support our research, and our experiments validate the effectiveness of Lr-MSI in the tone enhancement task. This work aims to establish a solid foundation for advancing spectral vision in mobile photography. The code is available at https://github.com/CalayZhou/JDM-HDRNet.", + "bbox": [ + 259, + 401, + 743, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Image enhancement $\\cdot$ RGB-Spectral fusion $\\cdot$ Spectral image intrinsic decomposition $\\cdot$ Spectral dataset and application", + "bbox": [ + 259, + 664, + 741, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 720, + 375, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in miniaturized spectrometers [29,37] have facilitated their integration into mobile devices, thereby expanding the scope of possible applications such as medical diagnosis [15], food safety evaluation [10], anti-spoofing face recognition [31] and color analysis [9]. Nevertheless, spectral sensors [2] in", + "bbox": [ + 212, + 753, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 217, + 824, + 385, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ebb7f5c05940a03ff7132713b05d71fa271a380e13d856d5bf3426e999c78ed4.jpg", + "image_caption": [ + "Fig. 1: The Lr-MSI is decomposed into the illumination $L(\\lambda)$ , reflectance $R(\\lambda, x)$ and shading $S(x)$ terms. We then analyze their respective roles in the mobile ISP pipeline." + ], + "image_footnote": [], + "bbox": [ + 250, + 143, + 754, + 229 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mobile photography are mainly employed for illuminant estimation in the Auto White Balance (AWB) operation at present. The spectral information holds potential to enhance the aesthetic appeal and perceptual quality of images. Exploring how to leverage additional spectral information to improve other image enhancement operations in the mobile Image Signal Processing (ISP) pipeline is an area requiring deeper investigation.", + "bbox": [ + 212, + 304, + 782, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We attempt to make a deeper analysis of the role of Lr-MSI in mobile photography by drawing insights from the spectral image intrinsic decomposition. As illustrated in Fig. 1, the spectral image can be decomposed into its constituent terms [5]: illumination curve $L(\\lambda)$ , reflectance $R(\\lambda, x)$ , and shading $S(x)$ . $L(\\lambda)$ captures the spectral characteristics of the light source and has traditionally been applied for illuminant estimation [8]. $R(\\lambda, x)$ embodies the albedo invariant color and texture of the material. The reflectance of spectral images which possesses fine-grained color channels shows promise for material segmentation and color mapping tasks. $S(x)$ represents interaction between the geometry of objects and illumination, which may contribute to the local brightness adjustment of high dynamic scenes. We contend that the potential of the reflectance $R(\\lambda, x)$ and shading $S(x)$ terms has not been thoroughly explored.", + "bbox": [ + 212, + 397, + 784, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we explore the application of low-resolution multi-spectral images in the tone enhancement task within the mobile ISP pipeline, especially for outdoor High Dynamic Range (HDR) scenes. The objective of tone enhancement is to adjust the brightness, contrast, and color characteristics of an image to enhance details in both the highlights and shadows while preserving its natural appearance. The primary challenge is the heterogeneous information fusion involving the integration of extra Lr-MSI with the original RGB workflow. This challenge manifests in two aspects: Firstly, variations in imaging mechanisms lead to the inherent complexity of spectral images, encompassing factors like scene geometry, inter-reflections and intricate artificial illumination [5], making it difficult to directly integrate spectral information into the mobile ISP workflow. Secondly, despite commercial spectral sensors offering over ten spectral channels [2], their spatial resolution is often confined due to the constrained spectral imaging capabilities on mobile devices. This limited spatial resolution remains a problem for the application of Lr-MSI in tone enhancement.", + "bbox": [ + 212, + 580, + 785, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above challenges, we propose a joint RGB-Spectral decomposition model guided image enhancement framework which consists of two", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "K Zhou et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "phases: joint decomposition and prior-guided enhancement. The effectiveness of this framework relies on three assumptions: (1) The near-infrared band in Lr-MSI can serve as an approximation of the shading term [7]. (2) Lr-MSI and RGB images exhibit complementary characteristics in both spatial and spectral resolutions. (3) The increased color channels in Lr-MSI contribute to material segmentation. Firstly, to mitigate limited spectral imaging capabilities, the joint decomposition phase leverages the complementarity between Lr-MSI and RGB images to predict shading, reflectance, and material semantic priors. Secondly, to overcome the inherent complexity of spectral images, the prior-guided enhancement phase seamlessly integrates these priors into the established HDRNet [11] to provide clear mechanistic guidance. Specifically, the shading component is separated from original RGB images to enhance the adaptability in dealing with localized brightness variations. Subsequently, the $R(\\lambda, x)$ of Lr-MSI is employed to guide the prediction of bilateral grids via the spectral perception map, particularly within the reflectance domain. Harnessing the reflectance domain alongside higher-dimensional spectral channels facilitates enhanced learning of color relationships for tone mapping. Lastly, we introduce the mixture of semantic grid experts to specifically adapt to the distinct color characteristics of individual material categories. The Joint DecoMposition framework guided HDRNet (JDM-HDRNet) exhibits superior performance compared to previous enhancement methods in terms of PSNR, SSIM [35] and a color difference metric.", + "bbox": [ + 217, + 146, + 785, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The contributions of this paper are as follows: (1) The pioneering Mobile-Spec dataset is meticulously constructed with high quality aligned RGB and hyperspectral images, which serves as a solid foundation for spectral applications. (2) We propose a joint RGB-Spectral decomposition model guided enhancement framework to address challenges of inherent complexity of spectral images and limited spectral imaging capabilities. The Lr-MSI and RGB images are decomposed into shading, reflectance and material semantic priors, and these priors are employed for promoting dynamic range enhancement, color mapping, and semantic grid expert learning, respectively. (3) Experiments demonstrate the effectiveness of introducing extra Lr-MSI in the tone enhancement task, which provides valuable exploration for the industrial application of spectral sensors.", + "bbox": [ + 217, + 464, + 785, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 217, + 650, + 385, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Potential of Mobile Spectral Sensors", + "text_level": 1, + "bbox": [ + 217, + 680, + 560, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As the cost of miniaturized spectrometers [29,37] decreases and their availability becomes more widespread, the integration of these spectral sensors with mobile devices creates new opportunities. Studies have demonstrated the potential of hyperspectral imaging in medical applications [15,16,23,30] such as hemodynamics monitoring [15] and skin analysis [16,30]. Moreover, smartphones equipped with snapshot hyperspectral sensors can enhance anti-spoofing face recognition [31] by preventing high-quality 3D mask attacks. Other applications include food safety evaluation [10], color analysis [9], and water quality assessment [32]. Despite these advancements, the use of spectral sensors in mobile photography remains", + "bbox": [ + 217, + 703, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "A Joint Decomposition Model Guided Image Enhancement", + "bbox": [ + 338, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "limited. As RGB sensors lack inherent illumination adaptation abilities to maintain color constancy, in recent years, most high-end smartphones are equipped with spectral sensors which aim to improve the accuracy of illuminant estimation [8, 13, 33] in an uncontrolled environment. The spectral sensor has evolved from its origins as an ambient light sensor to the refined capture of spectral signatures, primarily employed in auto white balance because its spatial resolution is often confined to a single pixel [2]. Our objective is to validate the applicability of Lr-MSI to other enhancement tasks within the mobile ISP pipeline, particularly under assumptions where the constraints of spectral imaging are mitigated by dual-camera acquisition in the Mobile-Spec dataset.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Enhancement in the Mobile ISP Pipeline", + "text_level": 1, + "bbox": [ + 214, + 318, + 602, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A standard mobile ISP pipeline [19] comprises both the Bayer processing routine and the subsequent photo-finishing routine. There are several enhancement procedures such as exposure correction [38], white-balance [12], contrast enhancement [4], color manipulation [20], local and global tone-mapping [28]. In this study, we aim to fully exploit the intrinsic decomposition components of spectral images by integrating Lr-MSI into the tone enhancement. This task involves adjusting the tonal range while preserving overall appearance and details, especially for HDR scenes. The commonly used method is the histogram equalization [21] which globally improves the contrast of entire images. However, global operations often fall short when dealing with scenes of high dynamic range. In the deep learning era, color transform-based methods are often employed for real-time processing of high-resolution images in a collaborative global-local manner. According to the color transform functions, these methods can be classified into affine transformation matrices [11, 34], multilayer perceptrons (MLPs) [14], 3D LUTs [36, 39, 41], etc. The pioneering work is the HDRNet [11] which learns the bilateral grid of coefficients at low resolution and performs transformations from input to output at full resolution. 3D LUT [39] replaces the affine transformation matrices with image-adaptive 3-dimensional lookup tables, enabling rapid and robust photo enhancement. We choose the HDRNet as the baseline due to its simple and flexible network design. The key challenge lies in the heterogeneous information fusion involving the integration of extra Lr-MSI with the HDRNet.", + "bbox": [ + 212, + 342, + 787, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Spectral Image Intrinsic Decomposition", + "text_level": 1, + "bbox": [ + 214, + 680, + 589, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To fuse the extra Lr-MSI into the original RGB workflow in HDRNet, we consider the role of spectral information in mobile photography from the perspective of Spectral Intrinsic Image Decomposition (SIID) [5]. The SIID problem aims to separate a spectral image into more elaborate intrinsic properties, such as the spectrum of the illumination, reflectance and shading components. This decomposition is typically performed in the spectral domain, which allows much more flexibility in the algorithm design and leads to a more accurate analysis. Zheng et al. [42] model illumination and reflectance spectra separation of a hyperspectral image into a low-rank matrix factorization problem. Chen et al. [5]", + "bbox": [ + 212, + 704, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "K Zhou et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1b73e1aef31acaf251064c69286f2f99f982ab5b6ce9072879d3fbd1c0cca14e.jpg", + "image_caption": [ + "Fig. 2: (a) The Mobile-Spec dataset comprises RGB images (16-bit input and 8-bit target), hyperspectral images and their corresponding shading, reflectance and material segmentation images. Near-infrared images serve as the guide map to approximate the shading term. (b) Spectral responses of different material categories. (c-d) Visualizations of RGB and hyperspectral data for different categories with t-SNE [27]." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 754, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "resolve a spectral image into independent intrinsic components: illumination, shading, and reflectance, and develop an effective algorithm which first reduces the spatial and spectral resolution with the super-pixel down-sampling, then the SIID problem is optimized under local constant and global sparse constraint. Huang et al. [17] extend the Retinex model to the multispectral domain based on the subspace constraint, which assumes the reflectance and shading vectors both live in a low dimensional subspace along the spectral domain. SIID is an ill-posed and under-constrained problem and previous algorithms are developed under the ideal lab environment, underscoring the challenges of tackling complex outdoor scenes. Fortunately, Cheng et al. [7] find that near-infrared images can function as reliable approximations of shading priors, and Zhang et al. [40] further perform intrinsic decomposition for outdoor scenes using aligned RGB and hyperspectral images. Based on this assumption, our joint decomposition model can be trained in an end-to-end manner, which will be discussed in Section 4.1.", + "bbox": [ + 212, + 373, + 787, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 The Mobile-Spec Dataset", + "text_level": 1, + "bbox": [ + 215, + 609, + 496, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To the best of our knowledge, no existing image enhancement dataset includes both aligned hyperspectral and RGB images. To bridge this data gap, we have constructed the high-quality Mobile-Spec dataset, consisting of 200 image groups. As shown in Fig. 2(a), each group represents an independent scene and includes a 16-bit input image, an 8-bit target image, a hyperspectral image, and the corresponding shading, reflectance, and material segmentation images.", + "bbox": [ + 212, + 643, + 787, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dual Camera System. To address the limited spectral imaging capabilities of sensors on smartphones, the Mobile-Spec is captured using a dual camera system: a high-end commercial smartphone and a Gaiasky-mini2 scanning hyperspectral camera [1]. The Gaiasky-mini2 covers a spectral range from 400 nm to 1000 nm with 176 spectral channels. The 16-bit input images are fused from multiple exposure frames, and the 8-bit target images are generated by a commercial privacy model that integrates the aesthetics of multiple experts.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "A Joint Decomposition Model Guided Image Enhancement", + "bbox": [ + 338, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Matching. The hyperspectral image size is $1057 \\times 960 \\times 176$ , while the 16-bit input and 8-bit target RGB images are $4096 \\times 3072 \\times 3$ . To align RGB images with a wide angle of view onto hyperspectral images, we employ an automatic alignment algorithm that computes SIFT descriptors [26] and estimates homography transformation matrices. Consequently, both RGB and hyperspectral images achieve a consistent spatial resolution of $1057 \\times 960$ .", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Influence of Light Sources. The Mobile-Spec dataset is captured exclusively outdoors to avoid the impact of intricate indoor lighting. Solar radiation spans a broad range of wavelengths, ranging from ultraviolet to visible light and extending into the infrared region. It is reasonable to assume that near-infrared images can approximate shading priors in outdoor settings [7].", + "bbox": [ + 212, + 237, + 787, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset Collection and Filtering. Our dataset spans multiple seasons and includes five categories commonly observed in outdoor environments: sky, building, plant, trunk, and road. Their respective spectral curves are illustrated in Fig. 2(b). Mobile-Spec undergoes rigorous filtering to ensure high quality. Firstly, samples are selected based on dynamic range, determined by an automated algorithm analyzing the difference between maximum and minimum pixel intensities. Secondly, samples exhibiting significant alignment errors are filtered out. Thirdly, samples with unsatisfactory visual quality are excluded, considering factors such as chromatic aberration, sharpness, noise, and artifacts. Lastly, to augment dataset diversity, scenes with high similarity are eliminated.", + "bbox": [ + 212, + 314, + 787, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Shading, Reflectance and Material Segmentation. To obtain targets of $S$ , $R$ and $M$ for the joint decomposition model, shading images are approximated by averaging near-infrared bands from $850~\\mathrm{nm}$ to $1000~\\mathrm{nm}$ in hyperspectral images. Reflectance images can be derived based on the Retinex theory [3, 22]. Ground truths of material segmentation are meticulously labeled by human annotators. More details about the Mobile-Spec are provided in the appendix.", + "bbox": [ + 212, + 465, + 787, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Proposed Method", + "text_level": 1, + "bbox": [ + 215, + 583, + 426, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Joint RGB-Spectral Decomposition Model", + "text_level": 1, + "bbox": [ + 215, + 618, + 612, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To integrate low-resolution spectral images into the RGB tone enhancement workflow, we first analyze the differences between hyperspectral and RGB camera imaging mechanisms. The imaging model for spectral cameras can be represented as follows based on [5,18]:", + "bbox": [ + 212, + 647, + 787, + 708 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI _ {k, x} = \\int_ {4 0 0 \\mathrm {n m}} ^ {1 0 0 0 \\mathrm {n m}} C _ {k} (\\lambda) L (\\lambda) S (x) R (\\lambda , x) d \\lambda , \\mathrm {k} = 1, 2, 3 \\dots \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 299, + 720, + 785, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$I_{k,x}$ represents the pixel intensities at position $x$ of the k-th band, $C_k(\\lambda)$ denotes the camera sensitivity function, $L(\\lambda)$ indicates the illumination curve of the light source, $S(x)$ refers to shading, and $R(\\lambda ,x)$ is the reflectance of scenes. $L(\\lambda)$ is usually utilized for illuminant estimation in the auto white balance procedure, while the role of the shading $S(x)$ and reflectance $R(\\lambda ,x)$ components", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "K Zhou et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/85361946d33ea537467f1f1d69b82e3613ed8b408f18d5c1728367d1d15b797f.jpg", + "image_caption": [ + "Fig. 3: (a) The joint RGB-Spectral decomposition model leverages the complementarity between RGB images and Lr-MSIs to predict shading (S), reflectance (R) and material category (M) priors. (b) The pixel intensity histogram. (c-d) The distribution of Pearson correlation coefficient $\\rho$ in the original RGB space and reflectance space." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 490, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb09ad04b2291750e192938aa2437412d87e89e9ad3e6b4a335f78a6f8750a73.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 143, + 779, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "has been rarely explored. The core difference between the RGB and hyperspectral cameras lies in $C_k(\\lambda)$ , where hyperspectral images have fine-grained color channels due to a higher spectral sampling rate. To simulate the limited spectral imaging capability on smartphones, we downsample the hyperspectral image $(1057 \\times 960 \\times 176)$ to the Lr-MSI $(16 \\times 16 \\times 10)$ . This downsampling is conducted to match the typical 10-channel configuration of commercial smartphone spectral sensors. [2]. We maintain flexibility in setting the spatial resolution to a default of $16 \\times 16$ , which will be discussed in Section 5.2. We posit that the effect of Lr-MSI in the joint decomposition model can be observed in two aspects.", + "bbox": [ + 212, + 351, + 787, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Approximate estimation of shading priors with the near-infrared band. To obtain the shading term, traditional optimization-based methods design various hand-crafted priors to constrain the solution [5,17], which are often inapplicable in complex outdoor scenes. Recently, near-infrared images have offered a simple yet powerful approach for shading smoothness [7], as the spectral curves of different colors gradually flatten out and texture variation is considerably reduced in the near-infrared band (refer to the appendix). Given that the Mobile-Spec dataset is exclusively captured outdoors to avoid the impact of complex indoor light sources, and considering the wide spectrum of solar radiation covering the infrared region, the shading term can be predicted through the joint decomposition of Lr-MSI and RGB images, as they share the same $S(x)$ .", + "bbox": [ + 212, + 489, + 787, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Enhancing material segmentation with the visible band. We employ t-SNE [27] to analyze the complex spectral patterns and relationships between different spectral bands. In Fig. 2(c), clusters representing different categories exhibit denser aggregation in the t-SNE visualization of hyperspectral images compared to that of RGB images. This observation suggests that the increased granularity of spectral channels provides enhanced discriminative capacity among various materials, and Lr-MSI may potentially contribute to the material segmentation task.", + "bbox": [ + 212, + 657, + 787, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 3(a), we employ a joint RGB-Spectral decomposition model to leverage both near-infrared and visible bands in Lr-MSI for shading prediction and material segmentation. Lr-MSI provides high spectral resolution but limited spatial resolution, while the RGB image offers rich spatial information but", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "A Joint Decomposition Model Guided Image Enhancement", + "bbox": [ + 338, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "lower spectral resolution. We transform the prediction of shading images from the regression problem to the classification problem by dividing shading images into eight brightness levels. Consequently, shading prediction and material segmentation can be decomposed in a collaborative manner. The joint RGB-Spectral decomposition model incorporates two independent encoders and decoders, adapted from the basic segmentation model FCN [25]. Specifically, LrMSI is resized to match the spatial resolution of the RGB image. The RGB encoder $E_{rgb}$ and spectral encoder $E_{spec}$ are trained to project the 16-bit input image $I_{rgb}$ and Lr-MSI $I_{msi}$ to the same latent space for representation alignment: $\\mathcal{F}_{rgb} = E_{rgb}(I_{rgb})$ and $\\mathcal{F}_{spec} = E_{spec}(I_{msi})$ . $\\mathcal{F}_{rgb}$ and $\\mathcal{F}_{spec}$ are fused together with the concatenation to share the common and complementary feature representations. Material segmentation $M$ and shading $S$ are independently predicted with the decoder $D_{m}$ and $D_{s}$ based on the fused feature representation. The joint RGB-Spectral decomposition model can be formulated as:", + "bbox": [ + 212, + 146, + 787, + 358 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nM, S = D _ {m, s} \\left(\\operatorname {c o n c a t} \\left(\\mathcal {F} _ {r g b}, \\mathcal {F} _ {\\text {s p e c}}\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 375, + 785, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With the shading component $S$ , the reflectance component $R_{rgb}$ , $R_{msi}$ can subsequently be obtained according to the Retinex theory [3,22].", + "bbox": [ + 212, + 396, + 785, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 JDM-HDRNet", + "text_level": 1, + "bbox": [ + 215, + 450, + 385, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Based on the established HDRNet [11], the JDM-HDRNet comprehensively exploits the $S$ , $R$ and $M$ priors to provide explicit guidance for tone enhancement.", + "bbox": [ + 212, + 474, + 785, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$S$ : Localized Brightness Adaptation. To adjust both highlight and shadow areas during tone enhancement, the HDRNet's bilateral grid depth is set to 8, accommodating eight different luminance levels. However, this design may not effectively handle various localized brightness variations in high dynamic range scenes. We propose that separating the shading component would yield benefits in such contexts. As shown in Fig. 3(b), the pixel histogram vectors between the 16-bit input and 8-bit target exhibits a Pearson correlation coefficient $\\rho$ of 0.66 in the original RGB space, which increases to 0.91 in the reflectance space. A larger $\\rho$ indicates greater similarity in pixel histogram characteristics. Analyzing the statistics of the $\\rho$ distribution in the Mobile-Spec dataset, we observe that the reflectance space (Fig. 3(d)) has a greater proportion of samples with high $\\rho$ than the RGB space (Fig. 3(c)). This finding indicates that separating the shading component from the RGB space to the reflectance space may reduce the difficulty of color mapping learning. In our implementation, the shading component $S$ is converted to the brightness representation $\\hat{S}$ with a localized brightness adaptation module, comprising two layers of convolution and deconvolution. This lightweight module helps mitigate drastic pixel intensity changes in localized high dynamic range areas, thereby enhancing the adaptability. Subsequently, the reflectance images of 16-bit input $R_{rgb} = I_{rgb} / \\hat{S}$ and Lr-MSI $R_{msi} = I_{msi} / \\hat{S}$ are fed into the coefficient prediction part of the bilateral grid.", + "bbox": [ + 212, + 506, + 787, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$R$ : Spectral Perception Self-Attention. After transforming from RGB space to reflectance space, we aim to leverage the reflectance of Lr-MSI $R_{msi}$", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "K Zhou et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c4ecf87738825d0f68a113b3aa327ba099ebcc3aa9ab51e766b089fdd6ac391d.jpg", + "image_caption": [ + "Fig. 4: The overall architecture of the proposed JDM-HDRNet. The $S$ prior is employed to enhance localized dynamic ranges, while $R$ and $M$ priors contribute to the color mapping and grid expert learning, respectively." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 258 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "which has fine-grained spectral channels to enhance the color mapping learning of bilateral grid. It should be noted that $R_{msi}$ has limited spatial resolution. We argue that downsampling in the spatial dimension will not cause significant degradation since bilateral grid coefficient prediction is carried out at a much lower spatial resolution. Hence, we design the Spectral Perception Self-Attention (SPSA) module which utilizes the $R_{msi}$ to guide $R_{rgb}$ for better bilateral coefficient prediction. $R_{msi}$ is resized to match the spatial resolution of $R_{rgb}$ , which is $256 \\times 256$ by default in HDRNet. The $R_{msi}$ and $R_{rgb}$ are processed with a stack of strided convolutional layers to extract the low-level features and generate hierarchical feature maps. These feature maps $\\mathcal{F}_{R_{msi}}$ , $\\mathcal{F}_{R_{rgb}}$ are concatenated along the channel dimension to form the cross-spectral feature embedding $\\mathcal{F}_R$ . Then $1 \\times 1$ convolutions $W_1$ and $3 \\times 3$ depth-wise convolutions $W_3$ are applied to the $\\mathcal{F}_{\\mathcal{R}}$ to generate query $Q$ , key $K$ and value $V$ :", + "bbox": [ + 212, + 340, + 787, + 537 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nQ = W _ {1} ^ {Q} W _ {3} ^ {Q} \\mathcal {F} _ {\\mathcal {R}}, K = W _ {1} ^ {K} W _ {3} ^ {K} \\mathcal {F} _ {\\mathcal {R}}, V = W _ {1} ^ {V} W _ {3} ^ {V} \\mathcal {F} _ {\\mathcal {R}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 553, + 784, + 573 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The query $Q$ and key $K$ are reshaped into $\\hat{Q} \\in \\mathbb{R}^{HW \\times C}$ and $\\hat{K} \\in \\mathbb{R}^{C \\times HW}$ so that their dot-product interaction generates a spectral perception map $A \\in \\mathbb{R}^{C \\times C}$ , which models the mutual information across the different spectral channels. The importance of different channels of the reshaped value $\\hat{V} \\in \\mathbb{R}^{HW \\times C}$ is reweighted by $A$ , and the SPSA module can be formulated as follows:", + "bbox": [ + 212, + 580, + 787, + 657 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nA = \\operatorname {s o f t m a x} (\\sigma \\hat {K} \\cdot \\hat {Q}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 405, + 686, + 785, + 705 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {F}} _ {R _ {m s i}} = \\hat {W} _ {3} ^ {m s i} \\hat {V} \\cdot A + W _ {3} ^ {m s i} \\mathcal {F} _ {R _ {m s i}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 374, + 707, + 785, + 727 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {F}} _ {R _ {r g b}} = \\hat {W} _ {3} ^ {r g b} \\hat {V} \\cdot A + W _ {3} ^ {r g b} \\mathcal {F} _ {R _ {r g b}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 729, + 785, + 750 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The $\\sigma$ is a learnable scaling parameter to control the magnitude of the dot product. The SPSA module serves as the residual learning which fuse the cross-spectral feature $\\hat{V} \\cdot A$ with the original feature $\\mathcal{F}_{R_{msi}}$ , $\\mathcal{F}_{R_{rgb}}$ by the adaptive weighting of $3 \\times 3$ convolution $\\hat{W}_3^{msi}$ , $\\hat{W}_3^{rgb}$ . The fused outputs $\\hat{\\mathcal{F}}_{R_{msi}}$ , $\\hat{\\mathcal{F}}_{R_{rgb}}$ of the SPSA module are fed into next layer in a progressive way.", + "bbox": [ + 212, + 762, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "A Joint Decomposition Model Guided Image Enhancement", + "bbox": [ + 336, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "$M$ : Mixture of Semantic Grid Experts. Common outdoor scenes can be roughly categorized into several primary types (e.g., sky, building, plant). As illustrated in Fig. 2(b), the plant spectral curves typically feature a reflection peak at around $550~\\mathrm{nm}$ , whereas the spectral curve of sky primarily resides within the blue wavelength range. Given these distinct characteristics, a multi-expert design can be introduced to effectively adapt to each material category's color preference. We believe that incorporating the material semantic prior $M$ can offer robust constraints and yield visually pleasing results for tone enhancement. Consequently, we introduce a mixture of semantic grid experts, wherein each category is individually associated with a unique bilateral grid coefficient. For a specific semantic grid expert, the binarized material segmentation of the $i$ -th category $M_{i}$ is transformed into the probability map $\\varPsi_{i}$ with a mapping function comprising two convolutional and deconvolutional layers. Subsequently, a pair of affine transformation parameters $(\\alpha, \\beta) \\in \\mathbb{R}^{HW \\times C}$ is learned based on the probability map $\\varPsi_{i}$ . The $(\\alpha, \\beta)$ is then introduced to the $Q$ and $K$ branch to modulate features conditioned on the probability map as follows:", + "bbox": [ + 212, + 146, + 787, + 388 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {Q} = (1 + \\alpha_ {Q}) \\cdot Q + \\beta_ {Q}, \\hat {K} = (1 + \\alpha_ {K}) \\cdot K + \\beta_ {K} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 404, + 785, + 422 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In each semantic grid expert $\\varPhi_{i}(x,y)$ , the material category $M_{i}$ is introduced in the same way. The bilateral grids of different experts are dynamically fused based on the category-dependent weight coefficient denoted as $w_{i}$ :", + "bbox": [ + 212, + 430, + 787, + 476 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi (x, y) = \\sum_ {i = 1} ^ {N} w _ {i} \\Phi_ {i} (x, y) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 489, + 785, + 530 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Here, $N$ is the number of material categories, and $\\varPhi(x,y)$ denotes the final multi-expert bilateral grid. The full-resolution output of reflectance image $O_R$ is obtained through interpolated affine transformations, which is operated based on the bilateral grid $\\varPhi(x,y)$ and the guidance map $g$ . The $g$ is learned from the $R_{rgb}$ using the original guidance map auxiliary network. With the multiplication of $O_R$ and $\\hat{S}$ , we get the final 8-bit output of JDM-HDRNet.", + "bbox": [ + 212, + 537, + 787, + 630 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 656, + 375, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 691, + 455, + 707 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In the Mobile-Spec dataset, $80\\%$ of the samples are allocated for the training set, while the remaining $20\\%$ are designated for the test set. The JDM-HDRNet is trained with a batch size of 4 and a learning rate of 0.0001. Specifically, the joint decomposition model is first trained with cross-entropy loss for both shading and segmentation predictions, then the JDM-HDRNet is trained with MSE loss in a supervised manner. During the training, input images are cropped to the size of $512 \\times 512$ , and these images undergo further downsampling to $256 \\times 256$ in the low-resolution stream.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "K Zhou et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/b796a268c7ba9e5734b8c2ee3662a49a5837054fff35a742eac93241560b70c9.jpg", + "table_caption": [ + "Table 1: Performance evaluation of HDRNet with shading $S^{*}$ and reflectance prior $R^{*}$ . Ablation studies are conducted on the spectral channel and spatial resolution.", + "(a) Performance evaluation of HDRNet and HDRNet with shading prior (HDRNet $+S^{*}$ )." + ], + "table_footnote": [], + "table_body": "
PSNR↑SSIM↑△E*↓PSNR↑SSIM↑△E*↓
HDRNet27.750.9395.12HDRNet+S*28.680.9574.29
", + "bbox": [ + 245, + 196, + 732, + 223 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/35a1db4d39efdae9ecab3ee86436945c8e0b3af82866a1351f09875a2793a639.jpg", + "table_caption": [ + "(b) Ablation study of reflectance prior $R^{*}$ on (c) Ablation study of reflectance prior $R^{*}$ on channels with different spectral ranges. the spatial resolution, which is ranging from" + ], + "table_footnote": [], + "table_body": "
PSNR↑SSIM↑△E*↓
baseline28.680.9574.29
400-520 nm28.740.9574.00
520-640 nm29.090.9643.82
640-760 nm29.190.9623.92
400-760 nm29.240.9673.72
760-1000 nm29.070.9653.88
400-1000 nm29.680.9683.55
", + "bbox": [ + 251, + 247, + 491, + 351 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7724f7be9e0bea9ad3566d5dfeafc0edb8d5f88831d06091b8e1d151c3dc8130.jpg", + "table_caption": [ + "(c) Ablation study of reflectance prior $R^{*}$ on the spatial resolution, which is ranging from $1\\times 1$ to $256\\times 256$" + ], + "table_footnote": [], + "table_body": "
PSNR↑SSIM↑△E*↓
baseline28.680.9574.29
1 × 129.050.9634.06
4 × 429.210.9633.80
16 × 1629.680.9683.55
64 × 6429.810.9693.47
256 × 25629.780.9673.49
", + "bbox": [ + 511, + 257, + 750, + 349 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/89326dea4468eadad816b5111fc1dd00d713da0339729136718dabe0b65ad50a.jpg", + "table_caption": [ + "Table 2: Performance evaluation of HDRNet with material semantic prior $M^{*}$ .", + "(a) The effectiveness of incorporating extra Lr-MSI in (b) Ablation study of $M^{*}$ on the numthe joint RGB-Spectral decomposition model for the ber of material categories. material segmentation task." + ], + "table_footnote": [], + "table_body": "
buildingplantskytrunkroadmIoU
RGB81.9987.6394.8410.9683.8771.86
+LrMSI87.5391.0195.8731.1489.1178.93
", + "bbox": [ + 233, + 422, + 527, + 464 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7b41ff8eb437a82d884fe25f67871cb9cb2e394723d590d7918b349281bdde59.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
1246
PSNR↑27.7527.9828.4729.11
SSIM↑0.9390.9440.9540.964
△E*↓5.125.354.694.28
", + "bbox": [ + 553, + 412, + 759, + 465 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Three metrics are employed to validate the effectiveness of our method, including the PSNR, SSIM [35], and $\\triangle E^{*}$ . $\\triangle E^{*}$ measures the perceptual difference between two colors in the CIELAB space, a smaller $\\triangle E^{*}$ value indicates better color accuracy or image quality. For the evaluation of the material segmentation, we adopt the standard metric as mean Intersection-over-Union (mIoU).", + "bbox": [ + 212, + 492, + 787, + 569 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Effectiveness of $S^{*}$ , $R^{*}$ , $M^{*}$ Priors", + "text_level": 1, + "bbox": [ + 214, + 589, + 545, + 604 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To avoid the impact of inaccurate decomposition, we begin by systematically integrating the ideal values of these priors (denoted as $S^{*}, R^{*}, M^{*}$ ) individually into the HDRNet baseline for validation.", + "bbox": [ + 212, + 613, + 784, + 657 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Shading prior $S^{*}$ . Tab. 1(a) illustrates that the HDRNet baseline attains a PSNR of 27.75 dB, while the separation of the shading component ( $S^{*}$ ) elevates the PSNR from 27.75 dB to 28.68 dB. Notably, the architecture of HDRNet remains unchanged in this sub-experiment. This finding demonstrates that transforming from RGB into reflectance space of input images with $S^{*}$ is a simple yet effective design for the tone enhancement task, which exhibits enhanced adaptability in diverse high dynamic range scenes.", + "bbox": [ + 212, + 657, + 787, + 763 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Reflectance prior $R^{*}$ . Given the practical limitations of spectral imaging capability on mobile devices, we conduct ablation studies to examine the spectral and spatial resolution configurations of Lr-MSI. We first maintain the spatial resolution at the default setting of $16 \\times 16$ and extract channels of different spectral ranges. The Lr-MSI has ten spectral channels, ranging from $400 \\mathrm{~nm}$ to", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "A Joint Decomposition Model Guided Image Enhancement", + "bbox": [ + 336, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "$1000\\mathrm{nm}$ , with each channel featuring a $60~\\mathrm{nm}$ bandwidth. Notably, $400 - 760\\mathrm{nm}$ encompasses six visible wavelength channels, while $760 - 1000\\mathrm{nm}$ includes four near-infrared channels. As observed in Tab. 1(b), the closer to the wavelength of red channel, the higher PSNR it will achieve. The possible explanation lies in the composition of our Mobile-Spec dataset, where a larger proportion of the area is covered by the sky and plant, while the presence of red objects is comparatively limited. Consequently, the HDRNet baseline tends to prioritize the blue and green wavelengths over the red wavelength during the training. Furthermore, the combination of all ten channels $(400 - 1000\\mathrm{nm})$ achieves the best results of $29.68\\mathrm{dB}$ . In our view, fine-grained spectral information enhances bilateral grids with heightened color perception capabilities and compensates for the imbalanced learning of different color channels. Then we maintain the number of channels at ten, and the spatial resolution is resized from $1\\times 1$ to $256\\times 256$ . As depicted in Tab. 1(c), there is a trend of improved performance in evaluation metrics with higher spatial resolution. Considering the inherent constraint of spatial resolution in Lr-MSI on mobile devices, we contend that the optimal parameter configuration for Lr-MSI may align to the spatial resolution of the bilateral grid $(16\\times 16)$ , which is the 8x downsampling corresponding to the spatial resolution of low-resolution stream input $(256\\times 256)$ . Our experimental results suggest that the $16\\times 16$ configuration yields satisfactory results, beyond which increasing the spatial resolution leads to diminishing marginal utility.", + "bbox": [ + 212, + 145, + 787, + 464 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Material semantic prior $M^{*}$ . Tab. 2(a) demonstrates that incorporating the Lr-MSI branch into the joint RGB-Spectral decomposition model enhances the mIoU of material segmentation from $71.86\\%$ to $78.93\\%$ . Despite Lr-MSI's notably lower spatial resolution compared to the 16-bit, it exhibits superior material identification capability owing to its fine-grained spectral channels. Subsequently, we conduct ablation experiments to evaluate the number of material categories. The category number of the original HDRNet baseline is set to one. In the two-category set, the sky, plant, and trunk are grouped into one category, while building, road, and others form another. For the four-category set, the plant and trunk, as well as building and road, are each combined into single categories, whereas the sky and others are treated as separate, independent categories. By refining the division of material categories, the mixture of semantic grid experts becomes more specialized in enhancing tones for specific materials. Tab. 2(b) demonstrates improved quantitative results across three evaluation metrics with the inclusion of six semantic grid experts.", + "bbox": [ + 212, + 464, + 787, + 691 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Ablation Study of JDM-HDRNet", + "text_level": 1, + "bbox": [ + 215, + 718, + 539, + 734 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Tab. 3(a) presents the ablation study incorporating ideal values $S^{*}$ , $R^{*}$ , $M^{*}$ into HDRNet (denoted as JDM-HDRNet\\*). The collaboration of shading $S$ and reflectance $R$ priors leads to a substantial decrease in $\\triangle E^{*}$ from 5.12 to 3.55 in the CIELAB color space. Moreover, the design of the mixture of semantic grid experts enables specific experts to focus on individual material categories, resulting in a PSNR of 30.14 dB. The above experiments are conducted with ideal priors.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "K Zhou et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/24081b16f55a82d4ea078bfef3472ac581e776cd7040b41c0ce8a4bf57b82cef.jpg", + "table_caption": [ + "Table 3: Ablation study of the proposed JDM-HDRNet on the $S$ , $R$ and $M$ priors.", + "(a) Ablation study on the ideal values of $S^*$ , (b) Ablation study on $S$ , $R$ and $M$ priors predicted from the joint decomposition model." + ], + "table_footnote": [], + "table_body": "
S*R* M*PSNR↑SSIM↑ΔE*↓
HDRNet27.750.9395.12HDRNet27.750.939
JDM-HDRNet*✓28.680.9574.29JDM-HDRNet✓28.590.953
✓29.680.9683.55✓29.570.966
✓30.140.9723.44✓29.830.967
", + "bbox": [ + 236, + 191, + 754, + 258 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1d104033313de3e9c84d5a507236b29f387d3e258d40418506ad83ad97702be8.jpg", + "table_caption": [ + "Table 4: Comparisons of our JDM-HDRNet with previous enhancement methods." + ], + "table_footnote": [], + "table_body": "
DPECSRNet3D LUTCLUTSepLUT-SSepLUT-L4D LUTHDRNetUPEOurs
PSNR↑22.8126.3427.5227.3027.5728.0827.7727.7528.1929.83
SSIM↑0.8060.9230.9260.9380.9330.9440.9350.9390.9460.967
△E*↓11.066.445.394.635.374.264.325.124.793.60
", + "bbox": [ + 217, + 284, + 787, + 339 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Subsequently, the predictions for $S$ , $R$ , $M$ from joint RGB-Spectral decomposition model are integrated with HDRNet (denoted as JDM-HDRNet). Tab. 3(b) shows that the JDM-HDRNet exhibits comparable performance across three metrics in the $S$ set, indicating that the localized brightness adaptation module demonstrates adaptive adjustment capabilities in response to shading image degradation. In addition, we observe that the performance of the reflectance $(R)$ set remains relatively unaffected, achieving an improvement of 0.98 dB. Because both the Lr-MSI and 16-bit RGB images have undergone consistent transformation into the reflectance space using the same $S$ from the joint decomposition model. Regarding the material semantic prior (M), JDM-HDRNet dynamically fuses different grid experts at the bilateral grid level to avoid the inaccurate segmentation of $M$ . Overall, JDM-HDRNet achieves a gain of 2.08 dB compared to the HDRNet baseline, which demonstrates that the low resolution spectral information can be effectively leveraged by the joint decomposition model.", + "bbox": [ + 212, + 369, + 787, + 582 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Comparisons with Previous Methods", + "text_level": 1, + "bbox": [ + 215, + 603, + 568, + 619 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Quantitative Comparisons. We compare JDM-HDRNet with previous color transform-based methods for tone enhancement. In Tab. 4, DPE [6] exhibits inferior performance compared to other methods, attributed to its unpaired setting, leading to output images affected by undesired artifacts introduced by the generative adversarial network. Among the color transform functions, the MLPs based CSRNet [14] exhibits a slightly lower PSNR of 26.34 dB. In contrast, the 3D LUTs based methods, such as 3D LUT [39], CLUT [41], SepLUT-S [36], and 4D LUT [24], demonstrate the similar PSNR ranging from 27.30 dB to 27.77 dB. Notably, SepLUT-L [36] with larger parameters exhibits a superior performance because it simultaneously takes advantage of both 1D and 3D LUTs for enhancement, the 1D LUT adjusts image contrast to achieve a more uniform distribution in an image-adaptive manner. Additionally, UPE [34] estimates an image-to-illumination mapping with constraints of smoothness loss, which is exclusively designed to enhance images under diverse lighting conditions, and it", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "A Joint Decomposition Model Guided Image Enhancement", + "bbox": [ + 338, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/86ef7e3909dbfdd8c134b3bc31ffb95c7233d26e487c8159b17100487647efd9.jpg", + "image_caption": [ + "Fig. 5: Qualitative comparisons on the Mobile-Spec dataset. The error maps displayed at the top-right corner illustrate the differences with ground truth for each image." + ], + "image_footnote": [], + "bbox": [ + 232, + 143, + 772, + 281 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "achieves good performance in handling high dynamic range scenes in the Mobile-Spec dataset. Previous methods face challenges in learning accurate tone enhancement for HDR scenes, while our priors-guided JDM-HDRNet outperforms previous methods by effectively leveraging $S$ , $R$ , $M$ priors predicted from the joint RGB-Spectral decomposition model to provide explicit guidance.", + "bbox": [ + 212, + 348, + 782, + 422 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Qualitative Comparisons. Fig. 5 presents qualitative comparisons of two representative examples. In the first row, JDM-HDRNet exhibits enhanced contrast and a wider dynamic range in the plant area, while other methods appear comparatively dim. In the second row, JDM-HDRNet shows reduced color deviation with the ground truth in the wall area, indicating the beneficial role of spectral information integration in improving photographic realism. In summary, JDM-HDRNet, guided by $S$ , $R$ and $M$ priors, achieves more accurate and pleasing colors for high dynamic range scenes.", + "bbox": [ + 212, + 424, + 784, + 542 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. The spatial resolution of Lr-MSI in this study is fixed at $16 \\times 16$ , a setting not yet achievable on commercial smartphones. Increasing the spatial resolution would improve the accuracy of the joint decomposition model. Further exploration is required to address the constrained spectral imaging capabilities in the context of image enhancement and related applications on mobile devices.", + "bbox": [ + 212, + 544, + 785, + 619 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 642, + 359, + 657 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have investigated the effectiveness of Lr-MSI in the mobile ISP pipeline for enhancing tone enhancement. We first construct the high-quality Mobile-Spec, a pioneering dataset comprising both aligned spectral and RGB images. Then we analyze the role of Lr-MSI from the perspective of spectral image intrinsic decomposition. Shading, reflectance and material semantic priors are derived using the joint decomposition model to address the inherent complexity of spectral images and the constraints of spectral imaging capabilities. Subsequently, JDM-HDRNet is meticulously designed under the explicit guidance of above priors, which exhibits superior performance than the competitive methods. Our dataset and method are expected to attract further research into this area and explore the untapped potential of spectral information on mobile devices.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "K Zhou et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research was supported by National Key Research and Development Program of China (2023YFF0713300), National Natural Science Foundation of China under Grant (62071216), Leading Technology of Jiangsu Basic Research Plan under Grant (BK20192003), and National Research Foundation Singapore Competitive Research Program (award number CRP29-2022-0003). We thank the device support from NJU-ZPTECH University-Enterprise Joint Laboratory for Computational Spectral Imaging.", + "bbox": [ + 212, + 172, + 787, + 279 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 297, + 321, + 314 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Gaiasky-mini hyperspectral imaging camera. https://www.dualix.com.cn/en/Goods/desc/id/123/aid/954.html", + "2. Spectral Sensing -ams.com. https://ams.com/spectral-sensing", + "3. Barrow, H., Tenenbaum, J., Hanson, A., Riseman, E.: Recovering intrinsic scene characteristics. Comput. vis. syst 2(3-26), 2 (1978)", + "4. Cai, J., Gu, S., Zhang, L.: Learning a deep single image contrast enhancer from multi-exposure images. IEEE Transactions on Image Processing 27(4), 2049-2062 (2018)", + "5. Chen, X., Zhu, W., Zhao, Y., Yu, Y., Zhou, Y., Yue, T., Du, S., Cao, X.: Intrinsic decomposition from a single spectral image. Applied optics 56(20), 5676-5684 (2017)", + "6. Chen, Y.S., Wang, Y.C., Kao, M.H., Chuang, Y.Y.: Deep photo enhancer: Unpaired learning for image enhancement from photographs with gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6306-6314 (2018)", + "7. Cheng, Z., Zheng, Y., You, S., Sato, I.: Non-local intrinsic decomposition with near-infrared priors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2521-2530 (2019)", + "8. Erba, I., et al.: Computational color constancy beyond rgb images: Multispectral and temporal extensions (2024)", + "9. Finlayson, G.D., Zhu, Y.: Designing color filters that make cameras more colorimetric. IEEE Transactions on Image Processing 30, 853-867 (2020)", + "10. Fu, X., Ying, Y.: Food safety evaluation based on near infrared spectroscopy and imaging: A review. Critical Reviews in Food Science and Nutrition 56(11), 1913-1924 (2016)", + "11. Gharbi, M., Chen, J., Barron, J.T., Hasinoff, S.W., Durand, F.: Deep bilateral learning for real-time image enhancement. ACM Transactions on Graphics (TOG) 36(4), 1-12 (2017)", + "12. Gijsenij, A., Gevers, T., Van De Weijer, J.: Computational color constancy: Survey and experiments. IEEE transactions on image processing 20(9), 2475-2489 (2011)", + "13. Glatt, O., Ater, Y., Kim, W.S., Werman, S., Berby, O., Zini, Y., Zelinger, S., Lee, S., Choi, H., Soloveichik, E.: Beyond rgb: A real world dataset for multispectral imaging in mobile devices. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4344-4354 (2024)", + "14. He, J., Liu, Y., Qiao, Y., Dong, C.: Conditional sequential modulation for efficient global image retouching. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16. pp. 679-695. Springer (2020)" + ], + "bbox": [ + 218, + 325, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "A Joint Decomposition Model Guided Image Enhancement", + "bbox": [ + 338, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. He, Q., Wang, R.: Hyperspectral imaging enabled by an unmodified smartphone for analyzing skin morphological features and monitoring hemodynamics. Biomedical optics express 11(2), 895-910 (2020)", + "16. He, Q., Wang, R.K.: Analysis of skin morphological features and real-time monitoring using snapshot hyperspectral imaging. Biomedical optics express 10(11), 5625-5638 (2019)", + "17. Huang, Q., Zhu, W., Zhao, Y., Chen, L., Wang, Y., Yue, T., Cao, X.: Multispectral image intrinsic decomposition via subspace constraint. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6430-6439 (2018)", + "18. Jiang, J., Liu, D., Gu, J., Süssstrunk, S.: What is the space of spectral sensitivity functions for digital color cameras? In: 2013 IEEE Workshop on Applications of Computer Vision (WACV). pp. 168-179. IEEE (2013)", + "19. Karaimer, H.C., Brown, M.S.: A software platform for manipulating the camera imaging pipeline. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 429-444. Springer (2016)", + "20. Kim, S.J., Lin, H.T., Lu, Z., Susstrunk, S., Lin, S., Brown, M.S.: A new in-camera imaging model for color computer vision and its application. IEEE Transactions on Pattern Analysis and Machine Intelligence 34(12), 2289-2302 (2012)", + "21. Kim, Y.T.: Contrast enhancement using brightness preserving bi-histogram equalization. IEEE transactions on Consumer Electronics 43(1), 1-8 (1997)", + "22. Land, E.H., McCann, J.J.: Lightness and retinex theory. Josa 61(1), 1-11 (1971)", + "23. Lindholm, V., Raita-Hakola, A.M., Annala, L., Salmivuori, M., Jeskanen, L., Saari, H., Koskenmies, S., Pitkänen, S., Pölönen, I., Isoherranen, K., et al.: Differentiating malignant from benign pigmented or non-pigmented skin tumours—a pilot study on 3d hyperspectral imaging of complex skin surfaces and convolutional neural networks. Journal of clinical medicine 11(7), 1914 (2022)", + "24. Liu, C., Yang, H., Fu, J., Qian, X.: 4d lut: learnable context-aware 4d lookup table for image enhancement. IEEE Transactions on Image Processing 32, 4742-4756 (2023)", + "25. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015)", + "26. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International journal of computer vision 60, 91-110 (2004)", + "27. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008)", + "28. Mantiuk, R., Daly, S., Kerofsky, L.: Display adaptive tone mapping. In: ACM SIGGRAPH 2008 papers, pp. 1-10 (2008)", + "29. McGonigle, A.J., Wilkes, T.C., Pering, T.D., Willmott, J.R., Cook, J.M., Mims III, F.M., Parisi, A.V.: Smartphone spectrometers. Sensors 18(1), 223 (2018)", + "30. Ng, P.C., Chi, Z., Verdie, Y., Lu, J., Plataniotis, K.N.: Hyper-skin: A hyperspectral dataset for reconstructing facial skin-spectra from rgb images. Advances in Neural Information Processing Systems 36 (2024)", + "31. Rao, S., Huang, Y., Cui, K., Li, Y.: Anti-spoofing face recognition using a metasurface-based snapshot hyperspectral image sensor. Optica 9(11), 1253-1259 (2022)", + "32. e Silva, G.M., Garcia, J.A., de Alencar Garitta, J., Cunha, D.G.F., Finkler, N.R., Mendiondo, E.M., Ghiglieno, F.: Smartphone-based spectrometry system as a prescreening assessment of copper and iron for real time control of water pollution. Journal of Environmental Management 323, 116214 (2022)" + ], + "bbox": [ + 217, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "K Zhou et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Thomas, J.B.: Illuminant estimation from uncalibrated multispectral images. In: 2015 Colour and Visual Computing Symposium (CVCS). pp. 1-6. IEEE (2015)", + "34. Wang, R., Zhang, Q., Fu, C.W., Shen, X., Zheng, W.S., Jia, J.: Underexposed photo enhancement using deep illumination estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6849-6857 (2019)", + "35. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)", + "36. Yang, C., Jin, M., Xu, Y., Zhang, R., Chen, Y., Liu, H.: Seplut: Separable image-adaptive lookup tables for real-time image enhancement. In: European Conference on Computer Vision. pp. 201-217. Springer (2022)", + "37. Yang, Z., Albrow-Owen, T., Cai, W., Hasan, T.: Miniaturization of optical spectrometers. Science 371(6528), eabe0722 (2021)", + "38. Yuan, L., Sun, J.: Automatic exposure correction of consumer photographs. In: Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part IV 12. pp. 771-785. Springer (2012)", + "39. Zeng, H., Cai, J., Li, L., Cao, Z., Zhang, L.: Learning image-adaptive 3d lookup tables for high performance photo enhancement in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(4), 2058-2073 (2020)", + "40. Zhang, F., You, S., Li, Y., Fu, Y.: Hsi-guided intrinsic image decomposition for outdoor scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 313-322 (2022)", + "41. Zhang, F., Zeng, H., Zhang, T., Zhang, L.: Clut-net: Learning adaptively compressed representations of 3dluts for lightweight image enhancement. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 6493-6501 (2022)", + "42. Zheng, Y., Sato, I., Sato, Y.: Illumination and reflectance spectra separation of a hyperspectral image meets low-rank matrix factorization. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1779-1787 (2015)" + ], + "bbox": [ + 212, + 146, + 785, + 589 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "A Joint Decomposition Model Guided Image Enhancement", + "bbox": [ + 338, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography/2ac99c58-8fa1-4a8d-b568-c7870f80ab53_model.json b/2024/Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography/2ac99c58-8fa1-4a8d-b568-c7870f80ab53_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cda1548f53f14a0c458c40c8a10ad7561707fd0a --- /dev/null +++ b/2024/Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography/2ac99c58-8fa1-4a8d-b568-c7870f80ab53_model.json @@ -0,0 +1,2126 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.26, + 0.141, + 0.745, + 0.21 + ], + "angle": 0, + "content": "Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.235, + 0.756, + 0.268 + ], + "angle": 0, + "content": "Kailai Zhou1, Lijing Cai1, Yibo Wang1, Mengya Zhang1, Bihan Wen3, Qiu Shen†,2, and Xun Cao1,2" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.277, + 0.776, + 0.292 + ], + "angle": 0, + "content": "1 School of Electronic Science and Engineering, Nanjing University, Nanjing, China" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.292, + 0.741, + 0.306 + ], + "angle": 0, + "content": "\\(^{2}\\) Key Laboratory of Optoelectronic Devices and Systems with Extreme" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.306, + 0.697, + 0.32 + ], + "angle": 0, + "content": "Performances of MOE, Nanjing University, Nanjing, China" + }, + { + "type": "text", + "bbox": [ + 0.344, + 0.32, + 0.659, + 0.335 + ], + "angle": 0, + "content": "3 Nanyang Technological University, Singapore" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.335, + 0.74, + 0.349 + ], + "angle": 0, + "content": "{calayzhou, cailijing, ybwang, 522023230161}@smail.nju.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.35, + 0.68, + 0.362 + ], + "angle": 0, + "content": "bihan.wen@ntu.edu.sg, {caoxun, shenqiu}@nju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.402, + 0.744, + 0.651 + ], + "angle": 0, + "content": "Abstract. The integration of miniaturized spectrometers into mobile devices offers new avenues for image quality enhancement and facilitates novel downstream tasks. However, the broader application of spectral sensors in mobile photography is hindered by the inherent complexity of spectral images and the constraints of spectral imaging capabilities. To overcome these challenges, we propose a joint RGB-Spectral decomposition model guided enhancement framework, which consists of two steps: joint decomposition and prior-guided enhancement. Firstly, we leverage the complementarity between RGB and Low-resolution Multi-Spectral Images (Lr-MSI) to predict shading, reflectance, and material semantic priors. Subsequently, these priors are seamlessly integrated into the established HDRNet to promote dynamic range enhancement, color mapping, and grid expert learning, respectively. Additionally, we construct a high-quality Mobile-Spec dataset to support our research, and our experiments validate the effectiveness of Lr-MSI in the tone enhancement task. This work aims to establish a solid foundation for advancing spectral vision in mobile photography. The code is available at https://github.com/CalayZhou/JDM-HDRNet." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.665, + 0.743, + 0.694 + ], + "angle": 0, + "content": "Keywords: Image enhancement \\(\\cdot\\) RGB-Spectral fusion \\(\\cdot\\) Spectral image intrinsic decomposition \\(\\cdot\\) Spectral dataset and application" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.721, + 0.377, + 0.737 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.754, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Recent advancements in miniaturized spectrometers [29,37] have facilitated their integration into mobile devices, thereby expanding the scope of possible applications such as medical diagnosis [15], food safety evaluation [10], anti-spoofing face recognition [31] and color analysis [9]. Nevertheless, spectral sensors [2] in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.386, + 0.84 + ], + "angle": 0, + "content": "† Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "K Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.145, + 0.756, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.242, + 0.788, + 0.272 + ], + "angle": 0, + "content": "Fig. 1: The Lr-MSI is decomposed into the illumination \\(L(\\lambda)\\), reflectance \\(R(\\lambda, x)\\) and shading \\(S(x)\\) terms. We then analyze their respective roles in the mobile ISP pipeline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.784, + 0.395 + ], + "angle": 0, + "content": "mobile photography are mainly employed for illuminant estimation in the Auto White Balance (AWB) operation at present. The spectral information holds potential to enhance the aesthetic appeal and perceptual quality of images. Exploring how to leverage additional spectral information to improve other image enhancement operations in the mobile Image Signal Processing (ISP) pipeline is an area requiring deeper investigation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.785, + 0.579 + ], + "angle": 0, + "content": "We attempt to make a deeper analysis of the role of Lr-MSI in mobile photography by drawing insights from the spectral image intrinsic decomposition. As illustrated in Fig. 1, the spectral image can be decomposed into its constituent terms [5]: illumination curve \\(L(\\lambda)\\), reflectance \\(R(\\lambda, x)\\), and shading \\(S(x)\\). \\(L(\\lambda)\\) captures the spectral characteristics of the light source and has traditionally been applied for illuminant estimation [8]. \\(R(\\lambda, x)\\) embodies the albedo invariant color and texture of the material. The reflectance of spectral images which possesses fine-grained color channels shows promise for material segmentation and color mapping tasks. \\(S(x)\\) represents interaction between the geometry of objects and illumination, which may contribute to the local brightness adjustment of high dynamic scenes. We contend that the potential of the reflectance \\(R(\\lambda, x)\\) and shading \\(S(x)\\) terms has not been thoroughly explored." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.786, + 0.807 + ], + "angle": 0, + "content": "In this paper, we explore the application of low-resolution multi-spectral images in the tone enhancement task within the mobile ISP pipeline, especially for outdoor High Dynamic Range (HDR) scenes. The objective of tone enhancement is to adjust the brightness, contrast, and color characteristics of an image to enhance details in both the highlights and shadows while preserving its natural appearance. The primary challenge is the heterogeneous information fusion involving the integration of extra Lr-MSI with the original RGB workflow. This challenge manifests in two aspects: Firstly, variations in imaging mechanisms lead to the inherent complexity of spectral images, encompassing factors like scene geometry, inter-reflections and intricate artificial illumination [5], making it difficult to directly integrate spectral information into the mobile ISP workflow. Secondly, despite commercial spectral sensors offering over ten spectral channels [2], their spatial resolution is often confined due to the constrained spectral imaging capabilities on mobile devices. This limited spatial resolution remains a problem for the application of Lr-MSI in tone enhancement." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.786, + 0.84 + ], + "angle": 0, + "content": "To address the above challenges, we propose a joint RGB-Spectral decomposition model guided image enhancement framework which consists of two" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Joint Decomposition Model Guided Image Enhancement" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.786, + 0.463 + ], + "angle": 0, + "content": "phases: joint decomposition and prior-guided enhancement. The effectiveness of this framework relies on three assumptions: (1) The near-infrared band in Lr-MSI can serve as an approximation of the shading term [7]. (2) Lr-MSI and RGB images exhibit complementary characteristics in both spatial and spectral resolutions. (3) The increased color channels in Lr-MSI contribute to material segmentation. Firstly, to mitigate limited spectral imaging capabilities, the joint decomposition phase leverages the complementarity between Lr-MSI and RGB images to predict shading, reflectance, and material semantic priors. Secondly, to overcome the inherent complexity of spectral images, the prior-guided enhancement phase seamlessly integrates these priors into the established HDRNet [11] to provide clear mechanistic guidance. Specifically, the shading component is separated from original RGB images to enhance the adaptability in dealing with localized brightness variations. Subsequently, the \\( R(\\lambda, x) \\) of Lr-MSI is employed to guide the prediction of bilateral grids via the spectral perception map, particularly within the reflectance domain. Harnessing the reflectance domain alongside higher-dimensional spectral channels facilitates enhanced learning of color relationships for tone mapping. Lastly, we introduce the mixture of semantic grid experts to specifically adapt to the distinct color characteristics of individual material categories. The Joint DecoMposition framework guided HDRNet (JDM-HDRNet) exhibits superior performance compared to previous enhancement methods in terms of PSNR, SSIM [35] and a color difference metric." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.465, + 0.786, + 0.629 + ], + "angle": 0, + "content": "The contributions of this paper are as follows: (1) The pioneering Mobile-Spec dataset is meticulously constructed with high quality aligned RGB and hyperspectral images, which serves as a solid foundation for spectral applications. (2) We propose a joint RGB-Spectral decomposition model guided enhancement framework to address challenges of inherent complexity of spectral images and limited spectral imaging capabilities. The Lr-MSI and RGB images are decomposed into shading, reflectance and material semantic priors, and these priors are employed for promoting dynamic range enhancement, color mapping, and semantic grid expert learning, respectively. (3) Experiments demonstrate the effectiveness of introducing extra Lr-MSI in the tone enhancement task, which provides valuable exploration for the industrial application of spectral sensors." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.651, + 0.386, + 0.667 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.681, + 0.562, + 0.696 + ], + "angle": 0, + "content": "2.1 Potential of Mobile Spectral Sensors" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.704, + 0.786, + 0.84 + ], + "angle": 0, + "content": "As the cost of miniaturized spectrometers [29,37] decreases and their availability becomes more widespread, the integration of these spectral sensors with mobile devices creates new opportunities. Studies have demonstrated the potential of hyperspectral imaging in medical applications [15,16,23,30] such as hemodynamics monitoring [15] and skin analysis [16,30]. Moreover, smartphones equipped with snapshot hyperspectral sensors can enhance anti-spoofing face recognition [31] by preventing high-quality 3D mask attacks. Other applications include food safety evaluation [10], color analysis [9], and water quality assessment [32]. Despite these advancements, the use of spectral sensors in mobile photography remains" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "K Zhou et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "limited. As RGB sensors lack inherent illumination adaptation abilities to maintain color constancy, in recent years, most high-end smartphones are equipped with spectral sensors which aim to improve the accuracy of illuminant estimation [8, 13, 33] in an uncontrolled environment. The spectral sensor has evolved from its origins as an ambient light sensor to the refined capture of spectral signatures, primarily employed in auto white balance because its spatial resolution is often confined to a single pixel [2]. Our objective is to validate the applicability of Lr-MSI to other enhancement tasks within the mobile ISP pipeline, particularly under assumptions where the constraints of spectral imaging are mitigated by dual-camera acquisition in the Mobile-Spec dataset." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.319, + 0.603, + 0.334 + ], + "angle": 0, + "content": "2.2 Enhancement in the Mobile ISP Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.789, + 0.661 + ], + "angle": 0, + "content": "A standard mobile ISP pipeline [19] comprises both the Bayer processing routine and the subsequent photo-finishing routine. There are several enhancement procedures such as exposure correction [38], white-balance [12], contrast enhancement [4], color manipulation [20], local and global tone-mapping [28]. In this study, we aim to fully exploit the intrinsic decomposition components of spectral images by integrating Lr-MSI into the tone enhancement. This task involves adjusting the tonal range while preserving overall appearance and details, especially for HDR scenes. The commonly used method is the histogram equalization [21] which globally improves the contrast of entire images. However, global operations often fall short when dealing with scenes of high dynamic range. In the deep learning era, color transform-based methods are often employed for real-time processing of high-resolution images in a collaborative global-local manner. According to the color transform functions, these methods can be classified into affine transformation matrices [11, 34], multilayer perceptrons (MLPs) [14], 3D LUTs [36, 39, 41], etc. The pioneering work is the HDRNet [11] which learns the bilateral grid of coefficients at low resolution and performs transformations from input to output at full resolution. 3D LUT [39] replaces the affine transformation matrices with image-adaptive 3-dimensional lookup tables, enabling rapid and robust photo enhancement. We choose the HDRNet as the baseline due to its simple and flexible network design. The key challenge lies in the heterogeneous information fusion involving the integration of extra Lr-MSI with the HDRNet." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.681, + 0.591, + 0.697 + ], + "angle": 0, + "content": "2.3 Spectral Image Intrinsic Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.843 + ], + "angle": 0, + "content": "To fuse the extra Lr-MSI into the original RGB workflow in HDRNet, we consider the role of spectral information in mobile photography from the perspective of Spectral Intrinsic Image Decomposition (SIID) [5]. The SIID problem aims to separate a spectral image into more elaborate intrinsic properties, such as the spectrum of the illumination, reflectance and shading components. This decomposition is typically performed in the spectral domain, which allows much more flexibility in the algorithm design and leads to a more accurate analysis. Zheng et al. [42] model illumination and reflectance spectra separation of a hyperspectral image into a low-rank matrix factorization problem. Chen et al. [5]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "A Joint Decomposition Model Guided Image Enhancement" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.756, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.273, + 0.785, + 0.343 + ], + "angle": 0, + "content": "Fig. 2: (a) The Mobile-Spec dataset comprises RGB images (16-bit input and 8-bit target), hyperspectral images and their corresponding shading, reflectance and material segmentation images. Near-infrared images serve as the guide map to approximate the shading term. (b) Spectral responses of different material categories. (c-d) Visualizations of RGB and hyperspectral data for different categories with t-SNE [27]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.585 + ], + "angle": 0, + "content": "resolve a spectral image into independent intrinsic components: illumination, shading, and reflectance, and develop an effective algorithm which first reduces the spatial and spectral resolution with the super-pixel down-sampling, then the SIID problem is optimized under local constant and global sparse constraint. Huang et al. [17] extend the Retinex model to the multispectral domain based on the subspace constraint, which assumes the reflectance and shading vectors both live in a low dimensional subspace along the spectral domain. SIID is an ill-posed and under-constrained problem and previous algorithms are developed under the ideal lab environment, underscoring the challenges of tackling complex outdoor scenes. Fortunately, Cheng et al. [7] find that near-infrared images can function as reliable approximations of shading priors, and Zhang et al. [40] further perform intrinsic decomposition for outdoor scenes using aligned RGB and hyperspectral images. Based on this assumption, our joint decomposition model can be trained in an end-to-end manner, which will be discussed in Section 4.1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.61, + 0.498, + 0.627 + ], + "angle": 0, + "content": "3 The Mobile-Spec Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.734 + ], + "angle": 0, + "content": "To the best of our knowledge, no existing image enhancement dataset includes both aligned hyperspectral and RGB images. To bridge this data gap, we have constructed the high-quality Mobile-Spec dataset, consisting of 200 image groups. As shown in Fig. 2(a), each group represents an independent scene and includes a 16-bit input image, an 8-bit target image, a hyperspectral image, and the corresponding shading, reflectance, and material segmentation images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Dual Camera System. To address the limited spectral imaging capabilities of sensors on smartphones, the Mobile-Spec is captured using a dual camera system: a high-end commercial smartphone and a Gaiasky-mini2 scanning hyperspectral camera [1]. The Gaiasky-mini2 covers a spectral range from 400 nm to 1000 nm with 176 spectral channels. The 16-bit input images are fused from multiple exposure frames, and the 8-bit target images are generated by a commercial privacy model that integrates the aesthetics of multiple experts." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "K Zhou et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "Image Matching. The hyperspectral image size is \\(1057 \\times 960 \\times 176\\), while the 16-bit input and 8-bit target RGB images are \\(4096 \\times 3072 \\times 3\\). To align RGB images with a wide angle of view onto hyperspectral images, we employ an automatic alignment algorithm that computes SIFT descriptors [26] and estimates homography transformation matrices. Consequently, both RGB and hyperspectral images achieve a consistent spatial resolution of \\(1057 \\times 960\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Influence of Light Sources. The Mobile-Spec dataset is captured exclusively outdoors to avoid the impact of intricate indoor lighting. Solar radiation spans a broad range of wavelengths, ranging from ultraviolet to visible light and extending into the infrared region. It is reasonable to assume that near-infrared images can approximate shading priors in outdoor settings [7]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.465 + ], + "angle": 0, + "content": "Dataset Collection and Filtering. Our dataset spans multiple seasons and includes five categories commonly observed in outdoor environments: sky, building, plant, trunk, and road. Their respective spectral curves are illustrated in Fig. 2(b). Mobile-Spec undergoes rigorous filtering to ensure high quality. Firstly, samples are selected based on dynamic range, determined by an automated algorithm analyzing the difference between maximum and minimum pixel intensities. Secondly, samples exhibiting significant alignment errors are filtered out. Thirdly, samples with unsatisfactory visual quality are excluded, considering factors such as chromatic aberration, sharpness, noise, and artifacts. Lastly, to augment dataset diversity, scenes with high similarity are eliminated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.789, + 0.557 + ], + "angle": 0, + "content": "Shading, Reflectance and Material Segmentation. To obtain targets of \\( S \\), \\( R \\) and \\( M \\) for the joint decomposition model, shading images are approximated by averaging near-infrared bands from \\( 850~\\mathrm{nm} \\) to \\( 1000~\\mathrm{nm} \\) in hyperspectral images. Reflectance images can be derived based on the Retinex theory [3, 22]. Ground truths of material segmentation are meticulously labeled by human annotators. More details about the Mobile-Spec are provided in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.584, + 0.427, + 0.601 + ], + "angle": 0, + "content": "4 Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.619, + 0.614, + 0.634 + ], + "angle": 0, + "content": "4.1 Joint RGB-Spectral Decomposition Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.648, + 0.788, + 0.709 + ], + "angle": 0, + "content": "To integrate low-resolution spectral images into the RGB tone enhancement workflow, we first analyze the differences between hyperspectral and RGB camera imaging mechanisms. The imaging model for spectral cameras can be represented as follows based on [5,18]:" + }, + { + "type": "equation", + "bbox": [ + 0.3, + 0.722, + 0.787, + 0.757 + ], + "angle": 0, + "content": "\\[\nI _ {k, x} = \\int_ {4 0 0 \\mathrm {n m}} ^ {1 0 0 0 \\mathrm {n m}} C _ {k} (\\lambda) L (\\lambda) S (x) R (\\lambda , x) d \\lambda , \\mathrm {k} = 1, 2, 3 \\dots \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.841 + ], + "angle": 0, + "content": "\\(I_{k,x}\\) represents the pixel intensities at position \\(x\\) of the k-th band, \\(C_k(\\lambda)\\) denotes the camera sensitivity function, \\(L(\\lambda)\\) indicates the illumination curve of the light source, \\(S(x)\\) refers to shading, and \\(R(\\lambda ,x)\\) is the reflectance of scenes. \\(L(\\lambda)\\) is usually utilized for illuminant estimation in the auto white balance procedure, while the role of the shading \\(S(x)\\) and reflectance \\(R(\\lambda ,x)\\) components" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "A Joint Decomposition Model Guided Image Enhancement" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.491, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.144, + 0.78, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.263, + 0.788, + 0.32 + ], + "angle": 0, + "content": "Fig. 3: (a) The joint RGB-Spectral decomposition model leverages the complementarity between RGB images and Lr-MSIs to predict shading (S), reflectance (R) and material category (M) priors. (b) The pixel intensity histogram. (c-d) The distribution of Pearson correlation coefficient \\(\\rho\\) in the original RGB space and reflectance space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.488 + ], + "angle": 0, + "content": "has been rarely explored. The core difference between the RGB and hyperspectral cameras lies in \\( C_k(\\lambda) \\), where hyperspectral images have fine-grained color channels due to a higher spectral sampling rate. To simulate the limited spectral imaging capability on smartphones, we downsample the hyperspectral image \\( (1057 \\times 960 \\times 176) \\) to the Lr-MSI \\( (16 \\times 16 \\times 10) \\). This downsampling is conducted to match the typical 10-channel configuration of commercial smartphone spectral sensors. [2]. We maintain flexibility in setting the spatial resolution to a default of \\( 16 \\times 16 \\), which will be discussed in Section 5.2. We posit that the effect of Lr-MSI in the joint decomposition model can be observed in two aspects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.788, + 0.656 + ], + "angle": 0, + "content": "Approximate estimation of shading priors with the near-infrared band. To obtain the shading term, traditional optimization-based methods design various hand-crafted priors to constrain the solution [5,17], which are often inapplicable in complex outdoor scenes. Recently, near-infrared images have offered a simple yet powerful approach for shading smoothness [7], as the spectral curves of different colors gradually flatten out and texture variation is considerably reduced in the near-infrared band (refer to the appendix). Given that the Mobile-Spec dataset is exclusively captured outdoors to avoid the impact of complex indoor light sources, and considering the wide spectrum of solar radiation covering the infrared region, the shading term can be predicted through the joint decomposition of Lr-MSI and RGB images, as they share the same \\( S(x) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Enhancing material segmentation with the visible band. We employ t-SNE [27] to analyze the complex spectral patterns and relationships between different spectral bands. In Fig. 2(c), clusters representing different categories exhibit denser aggregation in the t-SNE visualization of hyperspectral images compared to that of RGB images. This observation suggests that the increased granularity of spectral channels provides enhanced discriminative capacity among various materials, and Lr-MSI may potentially contribute to the material segmentation task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In Fig. 3(a), we employ a joint RGB-Spectral decomposition model to leverage both near-infrared and visible bands in Lr-MSI for shading prediction and material segmentation. Lr-MSI provides high spectral resolution but limited spatial resolution, while the RGB image offers rich spatial information but" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "K Zhou et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.359 + ], + "angle": 0, + "content": "lower spectral resolution. We transform the prediction of shading images from the regression problem to the classification problem by dividing shading images into eight brightness levels. Consequently, shading prediction and material segmentation can be decomposed in a collaborative manner. The joint RGB-Spectral decomposition model incorporates two independent encoders and decoders, adapted from the basic segmentation model FCN [25]. Specifically, LrMSI is resized to match the spatial resolution of the RGB image. The RGB encoder \\( E_{rgb} \\) and spectral encoder \\( E_{spec} \\) are trained to project the 16-bit input image \\( I_{rgb} \\) and Lr-MSI \\( I_{msi} \\) to the same latent space for representation alignment: \\( \\mathcal{F}_{rgb} = E_{rgb}(I_{rgb}) \\) and \\( \\mathcal{F}_{spec} = E_{spec}(I_{msi}) \\). \\( \\mathcal{F}_{rgb} \\) and \\( \\mathcal{F}_{spec} \\) are fused together with the concatenation to share the common and complementary feature representations. Material segmentation \\( M \\) and shading \\( S \\) are independently predicted with the decoder \\( D_{m} \\) and \\( D_{s} \\) based on the fused feature representation. The joint RGB-Spectral decomposition model can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.375, + 0.786, + 0.391 + ], + "angle": 0, + "content": "\\[\nM, S = D _ {m, s} \\left(\\operatorname {c o n c a t} \\left(\\mathcal {F} _ {r g b}, \\mathcal {F} _ {\\text {s p e c}}\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.787, + 0.428 + ], + "angle": 0, + "content": "With the shading component \\( S \\), the reflectance component \\( R_{rgb} \\), \\( R_{msi} \\) can subsequently be obtained according to the Retinex theory [3,22]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.451, + 0.386, + 0.464 + ], + "angle": 0, + "content": "4.2 JDM-HDRNet" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.787, + 0.506 + ], + "angle": 0, + "content": "Based on the established HDRNet [11], the JDM-HDRNet comprehensively exploits the \\(S\\), \\(R\\) and \\(M\\) priors to provide explicit guidance for tone enhancement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.789, + 0.81 + ], + "angle": 0, + "content": "\\(S\\): Localized Brightness Adaptation. To adjust both highlight and shadow areas during tone enhancement, the HDRNet's bilateral grid depth is set to 8, accommodating eight different luminance levels. However, this design may not effectively handle various localized brightness variations in high dynamic range scenes. We propose that separating the shading component would yield benefits in such contexts. As shown in Fig. 3(b), the pixel histogram vectors between the 16-bit input and 8-bit target exhibits a Pearson correlation coefficient \\(\\rho\\) of 0.66 in the original RGB space, which increases to 0.91 in the reflectance space. A larger \\(\\rho\\) indicates greater similarity in pixel histogram characteristics. Analyzing the statistics of the \\(\\rho\\) distribution in the Mobile-Spec dataset, we observe that the reflectance space (Fig. 3(d)) has a greater proportion of samples with high \\(\\rho\\) than the RGB space (Fig. 3(c)). This finding indicates that separating the shading component from the RGB space to the reflectance space may reduce the difficulty of color mapping learning. In our implementation, the shading component \\(S\\) is converted to the brightness representation \\(\\hat{S}\\) with a localized brightness adaptation module, comprising two layers of convolution and deconvolution. This lightweight module helps mitigate drastic pixel intensity changes in localized high dynamic range areas, thereby enhancing the adaptability. Subsequently, the reflectance images of 16-bit input \\(R_{rgb} = I_{rgb} / \\hat{S}\\) and Lr-MSI \\(R_{msi} = I_{msi} / \\hat{S}\\) are fed into the coefficient prediction part of the bilateral grid." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "\\(R\\): Spectral Perception Self-Attention. After transforming from RGB space to reflectance space, we aim to leverage the reflectance of Lr-MSI \\(R_{msi}\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.338, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "A Joint Decomposition Model Guided Image Enhancement" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.788, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.268, + 0.789, + 0.311 + ], + "angle": 0, + "content": "Fig. 4: The overall architecture of the proposed JDM-HDRNet. The \\(S\\) prior is employed to enhance localized dynamic ranges, while \\(R\\) and \\(M\\) priors contribute to the color mapping and grid expert learning, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.789, + 0.538 + ], + "angle": 0, + "content": "which has fine-grained spectral channels to enhance the color mapping learning of bilateral grid. It should be noted that \\( R_{msi} \\) has limited spatial resolution. We argue that downsampling in the spatial dimension will not cause significant degradation since bilateral grid coefficient prediction is carried out at a much lower spatial resolution. Hence, we design the Spectral Perception Self-Attention (SPSA) module which utilizes the \\( R_{msi} \\) to guide \\( R_{rgb} \\) for better bilateral coefficient prediction. \\( R_{msi} \\) is resized to match the spatial resolution of \\( R_{rgb} \\), which is \\( 256 \\times 256 \\) by default in HDRNet. The \\( R_{msi} \\) and \\( R_{rgb} \\) are processed with a stack of strided convolutional layers to extract the low-level features and generate hierarchical feature maps. These feature maps \\( \\mathcal{F}_{R_{msi}} \\), \\( \\mathcal{F}_{R_{rgb}} \\) are concatenated along the channel dimension to form the cross-spectral feature embedding \\( \\mathcal{F}_R \\). Then \\( 1 \\times 1 \\) convolutions \\( W_1 \\) and \\( 3 \\times 3 \\) depth-wise convolutions \\( W_3 \\) are applied to the \\( \\mathcal{F}_{\\mathcal{R}} \\) to generate query \\( Q \\), key \\( K \\) and value \\( V \\):" + }, + { + "type": "equation", + "bbox": [ + 0.314, + 0.554, + 0.785, + 0.574 + ], + "angle": 0, + "content": "\\[\nQ = W _ {1} ^ {Q} W _ {3} ^ {Q} \\mathcal {F} _ {\\mathcal {R}}, K = W _ {1} ^ {K} W _ {3} ^ {K} \\mathcal {F} _ {\\mathcal {R}}, V = W _ {1} ^ {V} W _ {3} ^ {V} \\mathcal {F} _ {\\mathcal {R}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.788, + 0.659 + ], + "angle": 0, + "content": "The query \\(Q\\) and key \\(K\\) are reshaped into \\(\\hat{Q} \\in \\mathbb{R}^{HW \\times C}\\) and \\(\\hat{K} \\in \\mathbb{R}^{C \\times HW}\\) so that their dot-product interaction generates a spectral perception map \\(A \\in \\mathbb{R}^{C \\times C}\\), which models the mutual information across the different spectral channels. The importance of different channels of the reshaped value \\(\\hat{V} \\in \\mathbb{R}^{HW \\times C}\\) is reweighted by \\(A\\), and the SPSA module can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.406, + 0.687, + 0.786, + 0.706 + ], + "angle": 0, + "content": "\\[\nA = \\operatorname {s o f t m a x} (\\sigma \\hat {K} \\cdot \\hat {Q}) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.708, + 0.787, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {F}} _ {R _ {m s i}} = \\hat {W} _ {3} ^ {m s i} \\hat {V} \\cdot A + W _ {3} ^ {m s i} \\mathcal {F} _ {R _ {m s i}} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.73, + 0.787, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {F}} _ {R _ {r g b}} = \\hat {W} _ {3} ^ {r g b} \\hat {V} \\cdot A + W _ {3} ^ {r g b} \\mathcal {F} _ {R _ {r g b}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.763, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The \\(\\sigma\\) is a learnable scaling parameter to control the magnitude of the dot product. The SPSA module serves as the residual learning which fuse the cross-spectral feature \\(\\hat{V} \\cdot A\\) with the original feature \\(\\mathcal{F}_{R_{msi}}\\), \\(\\mathcal{F}_{R_{rgb}}\\) by the adaptive weighting of \\(3 \\times 3\\) convolution \\(\\hat{W}_3^{msi}\\), \\(\\hat{W}_3^{rgb}\\). The fused outputs \\(\\hat{\\mathcal{F}}_{R_{msi}}\\), \\(\\hat{\\mathcal{F}}_{R_{rgb}}\\) of the SPSA module are fed into next layer in a progressive way." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "K Zhou et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.39 + ], + "angle": 0, + "content": "\\(M\\): Mixture of Semantic Grid Experts. Common outdoor scenes can be roughly categorized into several primary types (e.g., sky, building, plant). As illustrated in Fig. 2(b), the plant spectral curves typically feature a reflection peak at around \\(550~\\mathrm{nm}\\), whereas the spectral curve of sky primarily resides within the blue wavelength range. Given these distinct characteristics, a multi-expert design can be introduced to effectively adapt to each material category's color preference. We believe that incorporating the material semantic prior \\(M\\) can offer robust constraints and yield visually pleasing results for tone enhancement. Consequently, we introduce a mixture of semantic grid experts, wherein each category is individually associated with a unique bilateral grid coefficient. For a specific semantic grid expert, the binarized material segmentation of the \\(i\\)-th category \\(M_{i}\\) is transformed into the probability map \\(\\varPsi_{i}\\) with a mapping function comprising two convolutional and deconvolutional layers. Subsequently, a pair of affine transformation parameters \\((\\alpha, \\beta) \\in \\mathbb{R}^{HW \\times C}\\) is learned based on the probability map \\(\\varPsi_{i}\\). The \\((\\alpha, \\beta)\\) is then introduced to the \\(Q\\) and \\(K\\) branch to modulate features conditioned on the probability map as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.405, + 0.786, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\hat {Q} = (1 + \\alpha_ {Q}) \\cdot Q + \\beta_ {Q}, \\hat {K} = (1 + \\alpha_ {K}) \\cdot K + \\beta_ {K} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.477 + ], + "angle": 0, + "content": "In each semantic grid expert \\(\\varPhi_{i}(x,y)\\), the material category \\(M_{i}\\) is introduced in the same way. The bilateral grids of different experts are dynamically fused based on the category-dependent weight coefficient denoted as \\(w_{i}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.491, + 0.786, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\Phi (x, y) = \\sum_ {i = 1} ^ {N} w _ {i} \\Phi_ {i} (x, y) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Here, \\(N\\) is the number of material categories, and \\(\\varPhi(x,y)\\) denotes the final multi-expert bilateral grid. The full-resolution output of reflectance image \\(O_R\\) is obtained through interpolated affine transformations, which is operated based on the bilateral grid \\(\\varPhi(x,y)\\) and the guidance map \\(g\\). The \\(g\\) is learned from the \\(R_{rgb}\\) using the original guidance map auxiliary network. With the multiplication of \\(O_R\\) and \\(\\hat{S}\\), we get the final 8-bit output of JDM-HDRNet." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.657, + 0.376, + 0.674 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.692, + 0.457, + 0.708 + ], + "angle": 0, + "content": "5.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In the Mobile-Spec dataset, \\(80\\%\\) of the samples are allocated for the training set, while the remaining \\(20\\%\\) are designated for the test set. The JDM-HDRNet is trained with a batch size of 4 and a learning rate of 0.0001. Specifically, the joint decomposition model is first trained with cross-entropy loss for both shading and segmentation predictions, then the JDM-HDRNet is trained with MSE loss in a supervised manner. During the training, input images are cropped to the size of \\(512 \\times 512\\), and these images undergo further downsampling to \\(256 \\times 256\\) in the low-resolution stream." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.338, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Joint Decomposition Model Guided Image Enhancement" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 1: Performance evaluation of HDRNet with shading \\( S^{*} \\) and reflectance prior \\( R^{*} \\). Ablation studies are conducted on the spectral channel and spatial resolution." + }, + { + "type": "table_caption", + "bbox": [ + 0.246, + 0.185, + 0.757, + 0.197 + ], + "angle": 0, + "content": "(a) Performance evaluation of HDRNet and HDRNet with shading prior (HDRNet \\(+S^{*}\\))." + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.197, + 0.733, + 0.224 + ], + "angle": 0, + "content": "
PSNR↑SSIM↑△E*↓PSNR↑SSIM↑△E*↓
HDRNet27.750.9395.12HDRNet+S*28.680.9574.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.243, + 0.226, + 0.76, + 0.248 + ], + "angle": 0, + "content": "(b) Ablation study of reflectance prior \\( R^{*} \\) on (c) Ablation study of reflectance prior \\( R^{*} \\) on channels with different spectral ranges. the spatial resolution, which is ranging from" + }, + { + "type": "table", + "bbox": [ + 0.252, + 0.248, + 0.492, + 0.352 + ], + "angle": 0, + "content": "
PSNR↑SSIM↑△E*↓
baseline28.680.9574.29
400-520 nm28.740.9574.00
520-640 nm29.090.9643.82
640-760 nm29.190.9623.92
400-760 nm29.240.9673.72
760-1000 nm29.070.9653.88
400-1000 nm29.680.9683.55
" + }, + { + "type": "table_caption", + "bbox": [ + 0.504, + 0.228, + 0.761, + 0.257 + ], + "angle": 0, + "content": "(c) Ablation study of reflectance prior \\(R^{*}\\) on the spatial resolution, which is ranging from \\(1\\times 1\\) to \\(256\\times 256\\)" + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.258, + 0.75, + 0.351 + ], + "angle": 0, + "content": "
PSNR↑SSIM↑△E*↓
baseline28.680.9574.29
1 × 129.050.9634.06
4 × 429.210.9633.80
16 × 1629.680.9683.55
64 × 6429.810.9693.47
256 × 25629.780.9673.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.235, + 0.365, + 0.767, + 0.379 + ], + "angle": 0, + "content": "Table 2: Performance evaluation of HDRNet with material semantic prior \\(M^{*}\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.226, + 0.391, + 0.772, + 0.423 + ], + "angle": 0, + "content": "(a) The effectiveness of incorporating extra Lr-MSI in (b) Ablation study of \\( M^{*} \\) on the numthe joint RGB-Spectral decomposition model for the ber of material categories. material segmentation task." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.423, + 0.529, + 0.465 + ], + "angle": 0, + "content": "
buildingplantskytrunkroadmIoU
RGB81.9987.6394.8410.9683.8771.86
+LrMSI87.5391.0195.8731.1489.1178.93
" + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.413, + 0.761, + 0.466 + ], + "angle": 0, + "content": "
1246
PSNR↑27.7527.9828.4729.11
SSIM↑0.9390.9440.9540.964
△E*↓5.125.354.694.28
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.57 + ], + "angle": 0, + "content": "Three metrics are employed to validate the effectiveness of our method, including the PSNR, SSIM [35], and \\(\\triangle E^{*}\\). \\(\\triangle E^{*}\\) measures the perceptual difference between two colors in the CIELAB space, a smaller \\(\\triangle E^{*}\\) value indicates better color accuracy or image quality. For the evaluation of the material segmentation, we adopt the standard metric as mean Intersection-over-Union (mIoU)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.59, + 0.547, + 0.605 + ], + "angle": 0, + "content": "5.2 Effectiveness of \\(S^{*}\\), \\(R^{*}\\), \\(M^{*}\\) Priors" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.785, + 0.658 + ], + "angle": 0, + "content": "To avoid the impact of inaccurate decomposition, we begin by systematically integrating the ideal values of these priors (denoted as \\( S^{*}, R^{*}, M^{*} \\)) individually into the HDRNet baseline for validation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Shading prior \\( S^{*} \\). Tab. 1(a) illustrates that the HDRNet baseline attains a PSNR of 27.75 dB, while the separation of the shading component (\\( S^{*} \\)) elevates the PSNR from 27.75 dB to 28.68 dB. Notably, the architecture of HDRNet remains unchanged in this sub-experiment. This finding demonstrates that transforming from RGB into reflectance space of input images with \\( S^{*} \\) is a simple yet effective design for the tone enhancement task, which exhibits enhanced adaptability in diverse high dynamic range scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Reflectance prior \\( R^{*} \\). Given the practical limitations of spectral imaging capability on mobile devices, we conduct ablation studies to examine the spectral and spatial resolution configurations of Lr-MSI. We first maintain the spatial resolution at the default setting of \\( 16 \\times 16 \\) and extract channels of different spectral ranges. The Lr-MSI has ten spectral channels, ranging from \\( 400 \\mathrm{~nm} \\) to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "K Zhou et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.465 + ], + "angle": 0, + "content": "\\(1000\\mathrm{nm}\\), with each channel featuring a \\(60~\\mathrm{nm}\\) bandwidth. Notably, \\(400 - 760\\mathrm{nm}\\) encompasses six visible wavelength channels, while \\(760 - 1000\\mathrm{nm}\\) includes four near-infrared channels. As observed in Tab. 1(b), the closer to the wavelength of red channel, the higher PSNR it will achieve. The possible explanation lies in the composition of our Mobile-Spec dataset, where a larger proportion of the area is covered by the sky and plant, while the presence of red objects is comparatively limited. Consequently, the HDRNet baseline tends to prioritize the blue and green wavelengths over the red wavelength during the training. Furthermore, the combination of all ten channels \\((400 - 1000\\mathrm{nm})\\) achieves the best results of \\(29.68\\mathrm{dB}\\). In our view, fine-grained spectral information enhances bilateral grids with heightened color perception capabilities and compensates for the imbalanced learning of different color channels. Then we maintain the number of channels at ten, and the spatial resolution is resized from \\(1\\times 1\\) to \\(256\\times 256\\). As depicted in Tab. 1(c), there is a trend of improved performance in evaluation metrics with higher spatial resolution. Considering the inherent constraint of spatial resolution in Lr-MSI on mobile devices, we contend that the optimal parameter configuration for Lr-MSI may align to the spatial resolution of the bilateral grid \\((16\\times 16)\\), which is the 8x downsampling corresponding to the spatial resolution of low-resolution stream input \\((256\\times 256)\\). Our experimental results suggest that the \\(16\\times 16\\) configuration yields satisfactory results, beyond which increasing the spatial resolution leads to diminishing marginal utility." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.693 + ], + "angle": 0, + "content": "Material semantic prior \\( M^{*} \\). Tab. 2(a) demonstrates that incorporating the Lr-MSI branch into the joint RGB-Spectral decomposition model enhances the mIoU of material segmentation from \\( 71.86\\% \\) to \\( 78.93\\% \\). Despite Lr-MSI's notably lower spatial resolution compared to the 16-bit, it exhibits superior material identification capability owing to its fine-grained spectral channels. Subsequently, we conduct ablation experiments to evaluate the number of material categories. The category number of the original HDRNet baseline is set to one. In the two-category set, the sky, plant, and trunk are grouped into one category, while building, road, and others form another. For the four-category set, the plant and trunk, as well as building and road, are each combined into single categories, whereas the sky and others are treated as separate, independent categories. By refining the division of material categories, the mixture of semantic grid experts becomes more specialized in enhancing tones for specific materials. Tab. 2(b) demonstrates improved quantitative results across three evaluation metrics with the inclusion of six semantic grid experts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.719, + 0.54, + 0.735 + ], + "angle": 0, + "content": "5.3 Ablation Study of JDM-HDRNet" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Tab. 3(a) presents the ablation study incorporating ideal values \\( S^{*} \\), \\( R^{*} \\), \\( M^{*} \\) into HDRNet (denoted as JDM-HDRNet\\*). The collaboration of shading \\( S \\) and reflectance \\( R \\) priors leads to a substantial decrease in \\( \\triangle E^{*} \\) from 5.12 to 3.55 in the CIELAB color space. Moreover, the design of the mixture of semantic grid experts enables specific experts to focus on individual material categories, resulting in a PSNR of 30.14 dB. The above experiments are conducted with ideal priors." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Joint Decomposition Model Guided Image Enhancement" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.222, + 0.145, + 0.78, + 0.159 + ], + "angle": 0, + "content": "Table 3: Ablation study of the proposed JDM-HDRNet on the \\( S \\), \\( R \\) and \\( M \\) priors." + }, + { + "type": "table_caption", + "bbox": [ + 0.228, + 0.171, + 0.768, + 0.193 + ], + "angle": 0, + "content": "(a) Ablation study on the ideal values of \\( S^* \\), (b) Ablation study on \\( S \\), \\( R \\) and \\( M \\) priors predicted from the joint decomposition model." + }, + { + "type": "table", + "bbox": [ + 0.237, + 0.193, + 0.755, + 0.259 + ], + "angle": 0, + "content": "
S*R* M*PSNR↑SSIM↑ΔE*↓
HDRNet27.750.9395.12HDRNet27.750.939
JDM-HDRNet*✓28.680.9574.29JDM-HDRNet✓28.590.953
✓29.680.9683.55✓29.570.966
✓30.140.9723.44✓29.830.967
" + }, + { + "type": "table_caption", + "bbox": [ + 0.226, + 0.271, + 0.776, + 0.284 + ], + "angle": 0, + "content": "Table 4: Comparisons of our JDM-HDRNet with previous enhancement methods." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.285, + 0.789, + 0.34 + ], + "angle": 0, + "content": "
DPECSRNet3D LUTCLUTSepLUT-SSepLUT-L4D LUTHDRNetUPEOurs
PSNR↑22.8126.3427.5227.3027.5728.0827.7727.7528.1929.83
SSIM↑0.8060.9230.9260.9380.9330.9440.9350.9390.9460.967
△E*↓11.066.445.394.635.374.264.325.124.793.60
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.789, + 0.583 + ], + "angle": 0, + "content": "Subsequently, the predictions for \\( S \\), \\( R \\), \\( M \\) from joint RGB-Spectral decomposition model are integrated with HDRNet (denoted as JDM-HDRNet). Tab. 3(b) shows that the JDM-HDRNet exhibits comparable performance across three metrics in the \\( S \\) set, indicating that the localized brightness adaptation module demonstrates adaptive adjustment capabilities in response to shading image degradation. In addition, we observe that the performance of the reflectance \\( (R) \\) set remains relatively unaffected, achieving an improvement of 0.98 dB. Because both the Lr-MSI and 16-bit RGB images have undergone consistent transformation into the reflectance space using the same \\( S \\) from the joint decomposition model. Regarding the material semantic prior (M), JDM-HDRNet dynamically fuses different grid experts at the bilateral grid level to avoid the inaccurate segmentation of \\( M \\). Overall, JDM-HDRNet achieves a gain of 2.08 dB compared to the HDRNet baseline, which demonstrates that the low resolution spectral information can be effectively leveraged by the joint decomposition model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.604, + 0.569, + 0.62 + ], + "angle": 0, + "content": "5.4 Comparisons with Previous Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Quantitative Comparisons. We compare JDM-HDRNet with previous color transform-based methods for tone enhancement. In Tab. 4, DPE [6] exhibits inferior performance compared to other methods, attributed to its unpaired setting, leading to output images affected by undesired artifacts introduced by the generative adversarial network. Among the color transform functions, the MLPs based CSRNet [14] exhibits a slightly lower PSNR of 26.34 dB. In contrast, the 3D LUTs based methods, such as 3D LUT [39], CLUT [41], SepLUT-S [36], and 4D LUT [24], demonstrate the similar PSNR ranging from 27.30 dB to 27.77 dB. Notably, SepLUT-L [36] with larger parameters exhibits a superior performance because it simultaneously takes advantage of both 1D and 3D LUTs for enhancement, the 1D LUT adjusts image contrast to achieve a more uniform distribution in an image-adaptive manner. Additionally, UPE [34] estimates an image-to-illumination mapping with constraints of smoothness loss, which is exclusively designed to enhance images under diverse lighting conditions, and it" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "K Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.144, + 0.773, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.291, + 0.788, + 0.32 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparisons on the Mobile-Spec dataset. The error maps displayed at the top-right corner illustrate the differences with ground truth for each image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.784, + 0.424 + ], + "angle": 0, + "content": "achieves good performance in handling high dynamic range scenes in the Mobile-Spec dataset. Previous methods face challenges in learning accurate tone enhancement for HDR scenes, while our priors-guided JDM-HDRNet outperforms previous methods by effectively leveraging \\( S \\), \\( R \\), \\( M \\) priors predicted from the joint RGB-Spectral decomposition model to provide explicit guidance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.785, + 0.544 + ], + "angle": 0, + "content": "Qualitative Comparisons. Fig. 5 presents qualitative comparisons of two representative examples. In the first row, JDM-HDRNet exhibits enhanced contrast and a wider dynamic range in the plant area, while other methods appear comparatively dim. In the second row, JDM-HDRNet shows reduced color deviation with the ground truth in the wall area, indicating the beneficial role of spectral information integration in improving photographic realism. In summary, JDM-HDRNet, guided by \\( S \\), \\( R \\) and \\( M \\) priors, achieves more accurate and pleasing colors for high dynamic range scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.545, + 0.787, + 0.621 + ], + "angle": 0, + "content": "Limitations. The spatial resolution of Lr-MSI in this study is fixed at \\(16 \\times 16\\), a setting not yet achievable on commercial smartphones. Increasing the spatial resolution would improve the accuracy of the joint decomposition model. Further exploration is required to address the constrained spectral imaging capabilities in the context of image enhancement and related applications on mobile devices." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.643, + 0.36, + 0.659 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We have investigated the effectiveness of Lr-MSI in the mobile ISP pipeline for enhancing tone enhancement. We first construct the high-quality Mobile-Spec, a pioneering dataset comprising both aligned spectral and RGB images. Then we analyze the role of Lr-MSI from the perspective of spectral image intrinsic decomposition. Shading, reflectance and material semantic priors are derived using the joint decomposition model to address the inherent complexity of spectral images and the constraints of spectral imaging capabilities. Subsequently, JDM-HDRNet is meticulously designed under the explicit guidance of above priors, which exhibits superior performance than the competitive methods. Our dataset and method are expected to attract further research into this area and explore the untapped potential of spectral information on mobile devices." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Joint Decomposition Model Guided Image Enhancement" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.28 + ], + "angle": 0, + "content": "This research was supported by National Key Research and Development Program of China (2023YFF0713300), National Natural Science Foundation of China under Grant (62071216), Leading Technology of Jiangsu Basic Research Plan under Grant (BK20192003), and National Research Foundation Singapore Competitive Research Program (award number CRP29-2022-0003). We thank the device support from NJU-ZPTECH University-Enterprise Joint Laboratory for Computational Spectral Imaging." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.299, + 0.323, + 0.315 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.326, + 0.785, + 0.353 + ], + "angle": 0, + "content": "1. Gaiasky-mini hyperspectral imaging camera. https://www.dualix.com.cn/en/Goods/desc/id/123/aid/954.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.354, + 0.695, + 0.366 + ], + "angle": 0, + "content": "2. Spectral Sensing -ams.com. https://ams.com/spectral-sensing" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.367, + 0.785, + 0.393 + ], + "angle": 0, + "content": "3. Barrow, H., Tenenbaum, J., Hanson, A., Riseman, E.: Recovering intrinsic scene characteristics. Comput. vis. syst 2(3-26), 2 (1978)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.394, + 0.785, + 0.433 + ], + "angle": 0, + "content": "4. Cai, J., Gu, S., Zhang, L.: Learning a deep single image contrast enhancer from multi-exposure images. IEEE Transactions on Image Processing 27(4), 2049-2062 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.473 + ], + "angle": 0, + "content": "5. Chen, X., Zhu, W., Zhao, Y., Yu, Y., Zhou, Y., Yue, T., Du, S., Cao, X.: Intrinsic decomposition from a single spectral image. Applied optics 56(20), 5676-5684 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.475, + 0.785, + 0.529 + ], + "angle": 0, + "content": "6. Chen, Y.S., Wang, Y.C., Kao, M.H., Chuang, Y.Y.: Deep photo enhancer: Unpaired learning for image enhancement from photographs with gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6306-6314 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.53, + 0.785, + 0.569 + ], + "angle": 0, + "content": "7. Cheng, Z., Zheng, Y., You, S., Sato, I.: Non-local intrinsic decomposition with near-infrared priors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2521-2530 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.57, + 0.785, + 0.596 + ], + "angle": 0, + "content": "8. Erba, I., et al.: Computational color constancy beyond rgb images: Multispectral and temporal extensions (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.597, + 0.785, + 0.623 + ], + "angle": 0, + "content": "9. Finlayson, G.D., Zhu, Y.: Designing color filters that make cameras more colorimetric. IEEE Transactions on Image Processing 30, 853-867 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.624, + 0.785, + 0.663 + ], + "angle": 0, + "content": "10. Fu, X., Ying, Y.: Food safety evaluation based on near infrared spectroscopy and imaging: A review. Critical Reviews in Food Science and Nutrition 56(11), 1913-1924 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.664, + 0.785, + 0.704 + ], + "angle": 0, + "content": "11. Gharbi, M., Chen, J., Barron, J.T., Hasinoff, S.W., Durand, F.: Deep bilateral learning for real-time image enhancement. ACM Transactions on Graphics (TOG) 36(4), 1-12 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.705, + 0.785, + 0.731 + ], + "angle": 0, + "content": "12. Gijsenij, A., Gevers, T., Van De Weijer, J.: Computational color constancy: Survey and experiments. IEEE transactions on image processing 20(9), 2475-2489 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.732, + 0.785, + 0.785 + ], + "angle": 0, + "content": "13. Glatt, O., Ater, Y., Kim, W.S., Werman, S., Berby, O., Zini, Y., Zelinger, S., Lee, S., Choi, H., Soloveichik, E.: Beyond rgb: A real world dataset for multispectral imaging in mobile devices. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4344-4354 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "14. He, J., Liu, Y., Qiao, Y., Dong, C.: Conditional sequential modulation for efficient global image retouching. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16. pp. 679-695. Springer (2020)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.326, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "K Zhou et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "15. He, Q., Wang, R.: Hyperspectral imaging enabled by an unmodified smartphone for analyzing skin morphological features and monitoring hemodynamics. Biomedical optics express 11(2), 895-910 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "16. He, Q., Wang, R.K.: Analysis of skin morphological features and real-time monitoring using snapshot hyperspectral imaging. Biomedical optics express 10(11), 5625-5638 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.271 + ], + "angle": 0, + "content": "17. Huang, Q., Zhu, W., Zhao, Y., Chen, L., Wang, Y., Yue, T., Cao, X.: Multispectral image intrinsic decomposition via subspace constraint. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6430-6439 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.785, + 0.312 + ], + "angle": 0, + "content": "18. Jiang, J., Liu, D., Gu, J., Süssstrunk, S.: What is the space of spectral sensitivity functions for digital color cameras? In: 2013 IEEE Workshop on Applications of Computer Vision (WACV). pp. 168-179. IEEE (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.312, + 0.785, + 0.367 + ], + "angle": 0, + "content": "19. Karaimer, H.C., Brown, M.S.: A software platform for manipulating the camera imaging pipeline. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 429-444. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.367, + 0.785, + 0.407 + ], + "angle": 0, + "content": "20. Kim, S.J., Lin, H.T., Lu, Z., Susstrunk, S., Lin, S., Brown, M.S.: A new in-camera imaging model for color computer vision and its application. IEEE Transactions on Pattern Analysis and Machine Intelligence 34(12), 2289-2302 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.407, + 0.785, + 0.434 + ], + "angle": 0, + "content": "21. Kim, Y.T.: Contrast enhancement using brightness preserving bi-histogram equalization. IEEE transactions on Consumer Electronics 43(1), 1-8 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.434, + 0.785, + 0.447 + ], + "angle": 0, + "content": "22. Land, E.H., McCann, J.J.: Lightness and retinex theory. Josa 61(1), 1-11 (1971)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.447, + 0.785, + 0.515 + ], + "angle": 0, + "content": "23. Lindholm, V., Raita-Hakola, A.M., Annala, L., Salmivuori, M., Jeskanen, L., Saari, H., Koskenmies, S., Pitkänen, S., Pölönen, I., Isoherranen, K., et al.: Differentiating malignant from benign pigmented or non-pigmented skin tumours—a pilot study on 3d hyperspectral imaging of complex skin surfaces and convolutional neural networks. Journal of clinical medicine 11(7), 1914 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.515, + 0.785, + 0.556 + ], + "angle": 0, + "content": "24. Liu, C., Yang, H., Fu, J., Qian, X.: 4d lut: learnable context-aware 4d lookup table for image enhancement. IEEE Transactions on Image Processing 32, 4742-4756 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.556, + 0.785, + 0.598 + ], + "angle": 0, + "content": "25. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.597, + 0.785, + 0.624 + ], + "angle": 0, + "content": "26. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International journal of computer vision 60, 91-110 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.624, + 0.785, + 0.651 + ], + "angle": 0, + "content": "27. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.651, + 0.785, + 0.677 + ], + "angle": 0, + "content": "28. Mantiuk, R., Daly, S., Kerofsky, L.: Display adaptive tone mapping. In: ACM SIGGRAPH 2008 papers, pp. 1-10 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.677, + 0.785, + 0.704 + ], + "angle": 0, + "content": "29. McGonigle, A.J., Wilkes, T.C., Pering, T.D., Willmott, J.R., Cook, J.M., Mims III, F.M., Parisi, A.V.: Smartphone spectrometers. Sensors 18(1), 223 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.704, + 0.785, + 0.745 + ], + "angle": 0, + "content": "30. Ng, P.C., Chi, Z., Verdie, Y., Lu, J., Plataniotis, K.N.: Hyper-skin: A hyperspectral dataset for reconstructing facial skin-spectra from rgb images. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.745, + 0.785, + 0.786 + ], + "angle": 0, + "content": "31. Rao, S., Huang, Y., Cui, K., Li, Y.: Anti-spoofing face recognition using a metasurface-based snapshot hyperspectral image sensor. Optica 9(11), 1253-1259 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. e Silva, G.M., Garcia, J.A., de Alencar Garitta, J., Cunha, D.G.F., Finkler, N.R., Mendiondo, E.M., Ghiglieno, F.: Smartphone-based spectrometry system as a prescreening assessment of copper and iron for real time control of water pollution. Journal of Environmental Management 323, 116214 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.339, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "A Joint Decomposition Model Guided Image Enhancement" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "33. Thomas, J.B.: Illuminant estimation from uncalibrated multispectral images. In: 2015 Colour and Visual Computing Symposium (CVCS). pp. 1-6. IEEE (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.176, + 0.787, + 0.23 + ], + "angle": 0, + "content": "34. Wang, R., Zhang, Q., Fu, C.W., Shen, X., Zheng, W.S., Jia, J.: Underexposed photo enhancement using deep illumination estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6849-6857 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.231, + 0.785, + 0.272 + ], + "angle": 0, + "content": "35. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "36. Yang, C., Jin, M., Xu, Y., Zhang, R., Chen, Y., Liu, H.: Seplut: Separable image-adaptive lookup tables for real-time image enhancement. In: European Conference on Computer Vision. pp. 201-217. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "37. Yang, Z., Albrow-Owen, T., Cai, W., Hasan, T.: Miniaturization of optical spectrometers. Science 371(6528), eabe0722 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.343, + 0.785, + 0.395 + ], + "angle": 0, + "content": "38. Yuan, L., Sun, J.: Automatic exposure correction of consumer photographs. In: Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part IV 12. pp. 771-785. Springer (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.397, + 0.785, + 0.438 + ], + "angle": 0, + "content": "39. Zeng, H., Cai, J., Li, L., Cao, Z., Zhang, L.: Learning image-adaptive 3d lookup tables for high performance photo enhancement in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(4), 2058-2073 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "40. Zhang, F., You, S., Li, Y., Fu, Y.: Hsi-guided intrinsic image decomposition for outdoor scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 313-322 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.481, + 0.785, + 0.535 + ], + "angle": 0, + "content": "41. Zhang, F., Zeng, H., Zhang, T., Zhang, L.: Clut-net: Learning adaptively compressed representations of 3dluts for lightweight image enhancement. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 6493-6501 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.536, + 0.785, + 0.59 + ], + "angle": 0, + "content": "42. Zheng, Y., Sato, I., Sato, Y.: Illumination and reflectance spectra separation of a hyperspectral image meets low-rank matrix factorization. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1779-1787 (2015)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.59 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography/2ac99c58-8fa1-4a8d-b568-c7870f80ab53_origin.pdf b/2024/Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography/2ac99c58-8fa1-4a8d-b568-c7870f80ab53_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b72a462cae5ba894970bc1685409af98e505e8c8 --- /dev/null +++ b/2024/Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography/2ac99c58-8fa1-4a8d-b568-c7870f80ab53_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74d6411bbfb832570f4d8a097b75adff18e51e1f9988f43253b68a8dd128e983 +size 4214778 diff --git a/2024/Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography/full.md b/2024/Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d526c9e3d1a5567ba06d3af89ffc81336590c454 --- /dev/null +++ b/2024/Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography/full.md @@ -0,0 +1,276 @@ +# Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography + +Kailai Zhou1, Lijing Cai1, Yibo Wang1, Mengya Zhang1, Bihan Wen3, Qiu Shen†,2, and Xun Cao1,2 + +1 School of Electronic Science and Engineering, Nanjing University, Nanjing, China + +$^{2}$ Key Laboratory of Optoelectronic Devices and Systems with Extreme + +Performances of MOE, Nanjing University, Nanjing, China + +3 Nanyang Technological University, Singapore + +{calayzhou, cailijing, ybwang, 522023230161}@smail.nju.edu.cn, + +bihan.wen@ntu.edu.sg, {caoxun, shenqiu}@nju.edu.cn + +Abstract. The integration of miniaturized spectrometers into mobile devices offers new avenues for image quality enhancement and facilitates novel downstream tasks. However, the broader application of spectral sensors in mobile photography is hindered by the inherent complexity of spectral images and the constraints of spectral imaging capabilities. To overcome these challenges, we propose a joint RGB-Spectral decomposition model guided enhancement framework, which consists of two steps: joint decomposition and prior-guided enhancement. Firstly, we leverage the complementarity between RGB and Low-resolution Multi-Spectral Images (Lr-MSI) to predict shading, reflectance, and material semantic priors. Subsequently, these priors are seamlessly integrated into the established HDRNet to promote dynamic range enhancement, color mapping, and grid expert learning, respectively. Additionally, we construct a high-quality Mobile-Spec dataset to support our research, and our experiments validate the effectiveness of Lr-MSI in the tone enhancement task. This work aims to establish a solid foundation for advancing spectral vision in mobile photography. The code is available at https://github.com/CalayZhou/JDM-HDRNet. + +Keywords: Image enhancement $\cdot$ RGB-Spectral fusion $\cdot$ Spectral image intrinsic decomposition $\cdot$ Spectral dataset and application + +# 1 Introduction + +Recent advancements in miniaturized spectrometers [29,37] have facilitated their integration into mobile devices, thereby expanding the scope of possible applications such as medical diagnosis [15], food safety evaluation [10], anti-spoofing face recognition [31] and color analysis [9]. Nevertheless, spectral sensors [2] in + +![](images/ebb7f5c05940a03ff7132713b05d71fa271a380e13d856d5bf3426e999c78ed4.jpg) +Fig. 1: The Lr-MSI is decomposed into the illumination $L(\lambda)$ , reflectance $R(\lambda, x)$ and shading $S(x)$ terms. We then analyze their respective roles in the mobile ISP pipeline. + +mobile photography are mainly employed for illuminant estimation in the Auto White Balance (AWB) operation at present. The spectral information holds potential to enhance the aesthetic appeal and perceptual quality of images. Exploring how to leverage additional spectral information to improve other image enhancement operations in the mobile Image Signal Processing (ISP) pipeline is an area requiring deeper investigation. + +We attempt to make a deeper analysis of the role of Lr-MSI in mobile photography by drawing insights from the spectral image intrinsic decomposition. As illustrated in Fig. 1, the spectral image can be decomposed into its constituent terms [5]: illumination curve $L(\lambda)$ , reflectance $R(\lambda, x)$ , and shading $S(x)$ . $L(\lambda)$ captures the spectral characteristics of the light source and has traditionally been applied for illuminant estimation [8]. $R(\lambda, x)$ embodies the albedo invariant color and texture of the material. The reflectance of spectral images which possesses fine-grained color channels shows promise for material segmentation and color mapping tasks. $S(x)$ represents interaction between the geometry of objects and illumination, which may contribute to the local brightness adjustment of high dynamic scenes. We contend that the potential of the reflectance $R(\lambda, x)$ and shading $S(x)$ terms has not been thoroughly explored. + +In this paper, we explore the application of low-resolution multi-spectral images in the tone enhancement task within the mobile ISP pipeline, especially for outdoor High Dynamic Range (HDR) scenes. The objective of tone enhancement is to adjust the brightness, contrast, and color characteristics of an image to enhance details in both the highlights and shadows while preserving its natural appearance. The primary challenge is the heterogeneous information fusion involving the integration of extra Lr-MSI with the original RGB workflow. This challenge manifests in two aspects: Firstly, variations in imaging mechanisms lead to the inherent complexity of spectral images, encompassing factors like scene geometry, inter-reflections and intricate artificial illumination [5], making it difficult to directly integrate spectral information into the mobile ISP workflow. Secondly, despite commercial spectral sensors offering over ten spectral channels [2], their spatial resolution is often confined due to the constrained spectral imaging capabilities on mobile devices. This limited spatial resolution remains a problem for the application of Lr-MSI in tone enhancement. + +To address the above challenges, we propose a joint RGB-Spectral decomposition model guided image enhancement framework which consists of two + +phases: joint decomposition and prior-guided enhancement. The effectiveness of this framework relies on three assumptions: (1) The near-infrared band in Lr-MSI can serve as an approximation of the shading term [7]. (2) Lr-MSI and RGB images exhibit complementary characteristics in both spatial and spectral resolutions. (3) The increased color channels in Lr-MSI contribute to material segmentation. Firstly, to mitigate limited spectral imaging capabilities, the joint decomposition phase leverages the complementarity between Lr-MSI and RGB images to predict shading, reflectance, and material semantic priors. Secondly, to overcome the inherent complexity of spectral images, the prior-guided enhancement phase seamlessly integrates these priors into the established HDRNet [11] to provide clear mechanistic guidance. Specifically, the shading component is separated from original RGB images to enhance the adaptability in dealing with localized brightness variations. Subsequently, the $R(\lambda, x)$ of Lr-MSI is employed to guide the prediction of bilateral grids via the spectral perception map, particularly within the reflectance domain. Harnessing the reflectance domain alongside higher-dimensional spectral channels facilitates enhanced learning of color relationships for tone mapping. Lastly, we introduce the mixture of semantic grid experts to specifically adapt to the distinct color characteristics of individual material categories. The Joint DecoMposition framework guided HDRNet (JDM-HDRNet) exhibits superior performance compared to previous enhancement methods in terms of PSNR, SSIM [35] and a color difference metric. + +The contributions of this paper are as follows: (1) The pioneering Mobile-Spec dataset is meticulously constructed with high quality aligned RGB and hyperspectral images, which serves as a solid foundation for spectral applications. (2) We propose a joint RGB-Spectral decomposition model guided enhancement framework to address challenges of inherent complexity of spectral images and limited spectral imaging capabilities. The Lr-MSI and RGB images are decomposed into shading, reflectance and material semantic priors, and these priors are employed for promoting dynamic range enhancement, color mapping, and semantic grid expert learning, respectively. (3) Experiments demonstrate the effectiveness of introducing extra Lr-MSI in the tone enhancement task, which provides valuable exploration for the industrial application of spectral sensors. + +# 2 Related Work + +# 2.1 Potential of Mobile Spectral Sensors + +As the cost of miniaturized spectrometers [29,37] decreases and their availability becomes more widespread, the integration of these spectral sensors with mobile devices creates new opportunities. Studies have demonstrated the potential of hyperspectral imaging in medical applications [15,16,23,30] such as hemodynamics monitoring [15] and skin analysis [16,30]. Moreover, smartphones equipped with snapshot hyperspectral sensors can enhance anti-spoofing face recognition [31] by preventing high-quality 3D mask attacks. Other applications include food safety evaluation [10], color analysis [9], and water quality assessment [32]. Despite these advancements, the use of spectral sensors in mobile photography remains + +limited. As RGB sensors lack inherent illumination adaptation abilities to maintain color constancy, in recent years, most high-end smartphones are equipped with spectral sensors which aim to improve the accuracy of illuminant estimation [8, 13, 33] in an uncontrolled environment. The spectral sensor has evolved from its origins as an ambient light sensor to the refined capture of spectral signatures, primarily employed in auto white balance because its spatial resolution is often confined to a single pixel [2]. Our objective is to validate the applicability of Lr-MSI to other enhancement tasks within the mobile ISP pipeline, particularly under assumptions where the constraints of spectral imaging are mitigated by dual-camera acquisition in the Mobile-Spec dataset. + +# 2.2 Enhancement in the Mobile ISP Pipeline + +A standard mobile ISP pipeline [19] comprises both the Bayer processing routine and the subsequent photo-finishing routine. There are several enhancement procedures such as exposure correction [38], white-balance [12], contrast enhancement [4], color manipulation [20], local and global tone-mapping [28]. In this study, we aim to fully exploit the intrinsic decomposition components of spectral images by integrating Lr-MSI into the tone enhancement. This task involves adjusting the tonal range while preserving overall appearance and details, especially for HDR scenes. The commonly used method is the histogram equalization [21] which globally improves the contrast of entire images. However, global operations often fall short when dealing with scenes of high dynamic range. In the deep learning era, color transform-based methods are often employed for real-time processing of high-resolution images in a collaborative global-local manner. According to the color transform functions, these methods can be classified into affine transformation matrices [11, 34], multilayer perceptrons (MLPs) [14], 3D LUTs [36, 39, 41], etc. The pioneering work is the HDRNet [11] which learns the bilateral grid of coefficients at low resolution and performs transformations from input to output at full resolution. 3D LUT [39] replaces the affine transformation matrices with image-adaptive 3-dimensional lookup tables, enabling rapid and robust photo enhancement. We choose the HDRNet as the baseline due to its simple and flexible network design. The key challenge lies in the heterogeneous information fusion involving the integration of extra Lr-MSI with the HDRNet. + +# 2.3 Spectral Image Intrinsic Decomposition + +To fuse the extra Lr-MSI into the original RGB workflow in HDRNet, we consider the role of spectral information in mobile photography from the perspective of Spectral Intrinsic Image Decomposition (SIID) [5]. The SIID problem aims to separate a spectral image into more elaborate intrinsic properties, such as the spectrum of the illumination, reflectance and shading components. This decomposition is typically performed in the spectral domain, which allows much more flexibility in the algorithm design and leads to a more accurate analysis. Zheng et al. [42] model illumination and reflectance spectra separation of a hyperspectral image into a low-rank matrix factorization problem. Chen et al. [5] + +![](images/1b73e1aef31acaf251064c69286f2f99f982ab5b6ce9072879d3fbd1c0cca14e.jpg) +Fig. 2: (a) The Mobile-Spec dataset comprises RGB images (16-bit input and 8-bit target), hyperspectral images and their corresponding shading, reflectance and material segmentation images. Near-infrared images serve as the guide map to approximate the shading term. (b) Spectral responses of different material categories. (c-d) Visualizations of RGB and hyperspectral data for different categories with t-SNE [27]. + +resolve a spectral image into independent intrinsic components: illumination, shading, and reflectance, and develop an effective algorithm which first reduces the spatial and spectral resolution with the super-pixel down-sampling, then the SIID problem is optimized under local constant and global sparse constraint. Huang et al. [17] extend the Retinex model to the multispectral domain based on the subspace constraint, which assumes the reflectance and shading vectors both live in a low dimensional subspace along the spectral domain. SIID is an ill-posed and under-constrained problem and previous algorithms are developed under the ideal lab environment, underscoring the challenges of tackling complex outdoor scenes. Fortunately, Cheng et al. [7] find that near-infrared images can function as reliable approximations of shading priors, and Zhang et al. [40] further perform intrinsic decomposition for outdoor scenes using aligned RGB and hyperspectral images. Based on this assumption, our joint decomposition model can be trained in an end-to-end manner, which will be discussed in Section 4.1. + +# 3 The Mobile-Spec Dataset + +To the best of our knowledge, no existing image enhancement dataset includes both aligned hyperspectral and RGB images. To bridge this data gap, we have constructed the high-quality Mobile-Spec dataset, consisting of 200 image groups. As shown in Fig. 2(a), each group represents an independent scene and includes a 16-bit input image, an 8-bit target image, a hyperspectral image, and the corresponding shading, reflectance, and material segmentation images. + +Dual Camera System. To address the limited spectral imaging capabilities of sensors on smartphones, the Mobile-Spec is captured using a dual camera system: a high-end commercial smartphone and a Gaiasky-mini2 scanning hyperspectral camera [1]. The Gaiasky-mini2 covers a spectral range from 400 nm to 1000 nm with 176 spectral channels. The 16-bit input images are fused from multiple exposure frames, and the 8-bit target images are generated by a commercial privacy model that integrates the aesthetics of multiple experts. + +Image Matching. The hyperspectral image size is $1057 \times 960 \times 176$ , while the 16-bit input and 8-bit target RGB images are $4096 \times 3072 \times 3$ . To align RGB images with a wide angle of view onto hyperspectral images, we employ an automatic alignment algorithm that computes SIFT descriptors [26] and estimates homography transformation matrices. Consequently, both RGB and hyperspectral images achieve a consistent spatial resolution of $1057 \times 960$ . + +Influence of Light Sources. The Mobile-Spec dataset is captured exclusively outdoors to avoid the impact of intricate indoor lighting. Solar radiation spans a broad range of wavelengths, ranging from ultraviolet to visible light and extending into the infrared region. It is reasonable to assume that near-infrared images can approximate shading priors in outdoor settings [7]. + +Dataset Collection and Filtering. Our dataset spans multiple seasons and includes five categories commonly observed in outdoor environments: sky, building, plant, trunk, and road. Their respective spectral curves are illustrated in Fig. 2(b). Mobile-Spec undergoes rigorous filtering to ensure high quality. Firstly, samples are selected based on dynamic range, determined by an automated algorithm analyzing the difference between maximum and minimum pixel intensities. Secondly, samples exhibiting significant alignment errors are filtered out. Thirdly, samples with unsatisfactory visual quality are excluded, considering factors such as chromatic aberration, sharpness, noise, and artifacts. Lastly, to augment dataset diversity, scenes with high similarity are eliminated. + +Shading, Reflectance and Material Segmentation. To obtain targets of $S$ , $R$ and $M$ for the joint decomposition model, shading images are approximated by averaging near-infrared bands from $850~\mathrm{nm}$ to $1000~\mathrm{nm}$ in hyperspectral images. Reflectance images can be derived based on the Retinex theory [3, 22]. Ground truths of material segmentation are meticulously labeled by human annotators. More details about the Mobile-Spec are provided in the appendix. + +# 4 Proposed Method + +# 4.1 Joint RGB-Spectral Decomposition Model + +To integrate low-resolution spectral images into the RGB tone enhancement workflow, we first analyze the differences between hyperspectral and RGB camera imaging mechanisms. The imaging model for spectral cameras can be represented as follows based on [5,18]: + +$$ +I _ {k, x} = \int_ {4 0 0 \mathrm {n m}} ^ {1 0 0 0 \mathrm {n m}} C _ {k} (\lambda) L (\lambda) S (x) R (\lambda , x) d \lambda , \mathrm {k} = 1, 2, 3 \dots \tag {1} +$$ + +$I_{k,x}$ represents the pixel intensities at position $x$ of the k-th band, $C_k(\lambda)$ denotes the camera sensitivity function, $L(\lambda)$ indicates the illumination curve of the light source, $S(x)$ refers to shading, and $R(\lambda ,x)$ is the reflectance of scenes. $L(\lambda)$ is usually utilized for illuminant estimation in the auto white balance procedure, while the role of the shading $S(x)$ and reflectance $R(\lambda ,x)$ components + +![](images/85361946d33ea537467f1f1d69b82e3613ed8b408f18d5c1728367d1d15b797f.jpg) +Fig. 3: (a) The joint RGB-Spectral decomposition model leverages the complementarity between RGB images and Lr-MSIs to predict shading (S), reflectance (R) and material category (M) priors. (b) The pixel intensity histogram. (c-d) The distribution of Pearson correlation coefficient $\rho$ in the original RGB space and reflectance space. + +![](images/fb09ad04b2291750e192938aa2437412d87e89e9ad3e6b4a335f78a6f8750a73.jpg) + +has been rarely explored. The core difference between the RGB and hyperspectral cameras lies in $C_k(\lambda)$ , where hyperspectral images have fine-grained color channels due to a higher spectral sampling rate. To simulate the limited spectral imaging capability on smartphones, we downsample the hyperspectral image $(1057 \times 960 \times 176)$ to the Lr-MSI $(16 \times 16 \times 10)$ . This downsampling is conducted to match the typical 10-channel configuration of commercial smartphone spectral sensors. [2]. We maintain flexibility in setting the spatial resolution to a default of $16 \times 16$ , which will be discussed in Section 5.2. We posit that the effect of Lr-MSI in the joint decomposition model can be observed in two aspects. + +Approximate estimation of shading priors with the near-infrared band. To obtain the shading term, traditional optimization-based methods design various hand-crafted priors to constrain the solution [5,17], which are often inapplicable in complex outdoor scenes. Recently, near-infrared images have offered a simple yet powerful approach for shading smoothness [7], as the spectral curves of different colors gradually flatten out and texture variation is considerably reduced in the near-infrared band (refer to the appendix). Given that the Mobile-Spec dataset is exclusively captured outdoors to avoid the impact of complex indoor light sources, and considering the wide spectrum of solar radiation covering the infrared region, the shading term can be predicted through the joint decomposition of Lr-MSI and RGB images, as they share the same $S(x)$ . + +Enhancing material segmentation with the visible band. We employ t-SNE [27] to analyze the complex spectral patterns and relationships between different spectral bands. In Fig. 2(c), clusters representing different categories exhibit denser aggregation in the t-SNE visualization of hyperspectral images compared to that of RGB images. This observation suggests that the increased granularity of spectral channels provides enhanced discriminative capacity among various materials, and Lr-MSI may potentially contribute to the material segmentation task. + +In Fig. 3(a), we employ a joint RGB-Spectral decomposition model to leverage both near-infrared and visible bands in Lr-MSI for shading prediction and material segmentation. Lr-MSI provides high spectral resolution but limited spatial resolution, while the RGB image offers rich spatial information but + +lower spectral resolution. We transform the prediction of shading images from the regression problem to the classification problem by dividing shading images into eight brightness levels. Consequently, shading prediction and material segmentation can be decomposed in a collaborative manner. The joint RGB-Spectral decomposition model incorporates two independent encoders and decoders, adapted from the basic segmentation model FCN [25]. Specifically, LrMSI is resized to match the spatial resolution of the RGB image. The RGB encoder $E_{rgb}$ and spectral encoder $E_{spec}$ are trained to project the 16-bit input image $I_{rgb}$ and Lr-MSI $I_{msi}$ to the same latent space for representation alignment: $\mathcal{F}_{rgb} = E_{rgb}(I_{rgb})$ and $\mathcal{F}_{spec} = E_{spec}(I_{msi})$ . $\mathcal{F}_{rgb}$ and $\mathcal{F}_{spec}$ are fused together with the concatenation to share the common and complementary feature representations. Material segmentation $M$ and shading $S$ are independently predicted with the decoder $D_{m}$ and $D_{s}$ based on the fused feature representation. The joint RGB-Spectral decomposition model can be formulated as: + +$$ +M, S = D _ {m, s} \left(\operatorname {c o n c a t} \left(\mathcal {F} _ {r g b}, \mathcal {F} _ {\text {s p e c}}\right)\right) \tag {2} +$$ + +With the shading component $S$ , the reflectance component $R_{rgb}$ , $R_{msi}$ can subsequently be obtained according to the Retinex theory [3,22]. + +# 4.2 JDM-HDRNet + +Based on the established HDRNet [11], the JDM-HDRNet comprehensively exploits the $S$ , $R$ and $M$ priors to provide explicit guidance for tone enhancement. + +$S$ : Localized Brightness Adaptation. To adjust both highlight and shadow areas during tone enhancement, the HDRNet's bilateral grid depth is set to 8, accommodating eight different luminance levels. However, this design may not effectively handle various localized brightness variations in high dynamic range scenes. We propose that separating the shading component would yield benefits in such contexts. As shown in Fig. 3(b), the pixel histogram vectors between the 16-bit input and 8-bit target exhibits a Pearson correlation coefficient $\rho$ of 0.66 in the original RGB space, which increases to 0.91 in the reflectance space. A larger $\rho$ indicates greater similarity in pixel histogram characteristics. Analyzing the statistics of the $\rho$ distribution in the Mobile-Spec dataset, we observe that the reflectance space (Fig. 3(d)) has a greater proportion of samples with high $\rho$ than the RGB space (Fig. 3(c)). This finding indicates that separating the shading component from the RGB space to the reflectance space may reduce the difficulty of color mapping learning. In our implementation, the shading component $S$ is converted to the brightness representation $\hat{S}$ with a localized brightness adaptation module, comprising two layers of convolution and deconvolution. This lightweight module helps mitigate drastic pixel intensity changes in localized high dynamic range areas, thereby enhancing the adaptability. Subsequently, the reflectance images of 16-bit input $R_{rgb} = I_{rgb} / \hat{S}$ and Lr-MSI $R_{msi} = I_{msi} / \hat{S}$ are fed into the coefficient prediction part of the bilateral grid. + +$R$ : Spectral Perception Self-Attention. After transforming from RGB space to reflectance space, we aim to leverage the reflectance of Lr-MSI $R_{msi}$ + +![](images/c4ecf87738825d0f68a113b3aa327ba099ebcc3aa9ab51e766b089fdd6ac391d.jpg) +Fig. 4: The overall architecture of the proposed JDM-HDRNet. The $S$ prior is employed to enhance localized dynamic ranges, while $R$ and $M$ priors contribute to the color mapping and grid expert learning, respectively. + +which has fine-grained spectral channels to enhance the color mapping learning of bilateral grid. It should be noted that $R_{msi}$ has limited spatial resolution. We argue that downsampling in the spatial dimension will not cause significant degradation since bilateral grid coefficient prediction is carried out at a much lower spatial resolution. Hence, we design the Spectral Perception Self-Attention (SPSA) module which utilizes the $R_{msi}$ to guide $R_{rgb}$ for better bilateral coefficient prediction. $R_{msi}$ is resized to match the spatial resolution of $R_{rgb}$ , which is $256 \times 256$ by default in HDRNet. The $R_{msi}$ and $R_{rgb}$ are processed with a stack of strided convolutional layers to extract the low-level features and generate hierarchical feature maps. These feature maps $\mathcal{F}_{R_{msi}}$ , $\mathcal{F}_{R_{rgb}}$ are concatenated along the channel dimension to form the cross-spectral feature embedding $\mathcal{F}_R$ . Then $1 \times 1$ convolutions $W_1$ and $3 \times 3$ depth-wise convolutions $W_3$ are applied to the $\mathcal{F}_{\mathcal{R}}$ to generate query $Q$ , key $K$ and value $V$ : + +$$ +Q = W _ {1} ^ {Q} W _ {3} ^ {Q} \mathcal {F} _ {\mathcal {R}}, K = W _ {1} ^ {K} W _ {3} ^ {K} \mathcal {F} _ {\mathcal {R}}, V = W _ {1} ^ {V} W _ {3} ^ {V} \mathcal {F} _ {\mathcal {R}} \tag {3} +$$ + +The query $Q$ and key $K$ are reshaped into $\hat{Q} \in \mathbb{R}^{HW \times C}$ and $\hat{K} \in \mathbb{R}^{C \times HW}$ so that their dot-product interaction generates a spectral perception map $A \in \mathbb{R}^{C \times C}$ , which models the mutual information across the different spectral channels. The importance of different channels of the reshaped value $\hat{V} \in \mathbb{R}^{HW \times C}$ is reweighted by $A$ , and the SPSA module can be formulated as follows: + +$$ +A = \operatorname {s o f t m a x} (\sigma \hat {K} \cdot \hat {Q}) \tag {4} +$$ + +$$ +\hat {\mathcal {F}} _ {R _ {m s i}} = \hat {W} _ {3} ^ {m s i} \hat {V} \cdot A + W _ {3} ^ {m s i} \mathcal {F} _ {R _ {m s i}} \tag {5} +$$ + +$$ +\hat {\mathcal {F}} _ {R _ {r g b}} = \hat {W} _ {3} ^ {r g b} \hat {V} \cdot A + W _ {3} ^ {r g b} \mathcal {F} _ {R _ {r g b}} \tag {6} +$$ + +The $\sigma$ is a learnable scaling parameter to control the magnitude of the dot product. The SPSA module serves as the residual learning which fuse the cross-spectral feature $\hat{V} \cdot A$ with the original feature $\mathcal{F}_{R_{msi}}$ , $\mathcal{F}_{R_{rgb}}$ by the adaptive weighting of $3 \times 3$ convolution $\hat{W}_3^{msi}$ , $\hat{W}_3^{rgb}$ . The fused outputs $\hat{\mathcal{F}}_{R_{msi}}$ , $\hat{\mathcal{F}}_{R_{rgb}}$ of the SPSA module are fed into next layer in a progressive way. + +$M$ : Mixture of Semantic Grid Experts. Common outdoor scenes can be roughly categorized into several primary types (e.g., sky, building, plant). As illustrated in Fig. 2(b), the plant spectral curves typically feature a reflection peak at around $550~\mathrm{nm}$ , whereas the spectral curve of sky primarily resides within the blue wavelength range. Given these distinct characteristics, a multi-expert design can be introduced to effectively adapt to each material category's color preference. We believe that incorporating the material semantic prior $M$ can offer robust constraints and yield visually pleasing results for tone enhancement. Consequently, we introduce a mixture of semantic grid experts, wherein each category is individually associated with a unique bilateral grid coefficient. For a specific semantic grid expert, the binarized material segmentation of the $i$ -th category $M_{i}$ is transformed into the probability map $\varPsi_{i}$ with a mapping function comprising two convolutional and deconvolutional layers. Subsequently, a pair of affine transformation parameters $(\alpha, \beta) \in \mathbb{R}^{HW \times C}$ is learned based on the probability map $\varPsi_{i}$ . The $(\alpha, \beta)$ is then introduced to the $Q$ and $K$ branch to modulate features conditioned on the probability map as follows: + +$$ +\hat {Q} = (1 + \alpha_ {Q}) \cdot Q + \beta_ {Q}, \hat {K} = (1 + \alpha_ {K}) \cdot K + \beta_ {K} \tag {7} +$$ + +In each semantic grid expert $\varPhi_{i}(x,y)$ , the material category $M_{i}$ is introduced in the same way. The bilateral grids of different experts are dynamically fused based on the category-dependent weight coefficient denoted as $w_{i}$ : + +$$ +\Phi (x, y) = \sum_ {i = 1} ^ {N} w _ {i} \Phi_ {i} (x, y) \tag {8} +$$ + +Here, $N$ is the number of material categories, and $\varPhi(x,y)$ denotes the final multi-expert bilateral grid. The full-resolution output of reflectance image $O_R$ is obtained through interpolated affine transformations, which is operated based on the bilateral grid $\varPhi(x,y)$ and the guidance map $g$ . The $g$ is learned from the $R_{rgb}$ using the original guidance map auxiliary network. With the multiplication of $O_R$ and $\hat{S}$ , we get the final 8-bit output of JDM-HDRNet. + +# 5 Experiments + +# 5.1 Implementation Details + +In the Mobile-Spec dataset, $80\%$ of the samples are allocated for the training set, while the remaining $20\%$ are designated for the test set. The JDM-HDRNet is trained with a batch size of 4 and a learning rate of 0.0001. Specifically, the joint decomposition model is first trained with cross-entropy loss for both shading and segmentation predictions, then the JDM-HDRNet is trained with MSE loss in a supervised manner. During the training, input images are cropped to the size of $512 \times 512$ , and these images undergo further downsampling to $256 \times 256$ in the low-resolution stream. + +Table 1: Performance evaluation of HDRNet with shading $S^{*}$ and reflectance prior $R^{*}$ . Ablation studies are conducted on the spectral channel and spatial resolution. +(a) Performance evaluation of HDRNet and HDRNet with shading prior (HDRNet $+S^{*}$ ). + +
PSNR↑SSIM↑△E*↓PSNR↑SSIM↑△E*↓
HDRNet27.750.9395.12HDRNet+S*28.680.9574.29
+ +(b) Ablation study of reflectance prior $R^{*}$ on (c) Ablation study of reflectance prior $R^{*}$ on channels with different spectral ranges. the spatial resolution, which is ranging from + +
PSNR↑SSIM↑△E*↓
baseline28.680.9574.29
400-520 nm28.740.9574.00
520-640 nm29.090.9643.82
640-760 nm29.190.9623.92
400-760 nm29.240.9673.72
760-1000 nm29.070.9653.88
400-1000 nm29.680.9683.55
+ +(c) Ablation study of reflectance prior $R^{*}$ on the spatial resolution, which is ranging from $1\times 1$ to $256\times 256$ + +
PSNR↑SSIM↑△E*↓
baseline28.680.9574.29
1 × 129.050.9634.06
4 × 429.210.9633.80
16 × 1629.680.9683.55
64 × 6429.810.9693.47
256 × 25629.780.9673.49
+ +Table 2: Performance evaluation of HDRNet with material semantic prior $M^{*}$ . +(a) The effectiveness of incorporating extra Lr-MSI in (b) Ablation study of $M^{*}$ on the numthe joint RGB-Spectral decomposition model for the ber of material categories. material segmentation task. + +
buildingplantskytrunkroadmIoU
RGB81.9987.6394.8410.9683.8771.86
+LrMSI87.5391.0195.8731.1489.1178.93
+ +
1246
PSNR↑27.7527.9828.4729.11
SSIM↑0.9390.9440.9540.964
△E*↓5.125.354.694.28
+ +Three metrics are employed to validate the effectiveness of our method, including the PSNR, SSIM [35], and $\triangle E^{*}$ . $\triangle E^{*}$ measures the perceptual difference between two colors in the CIELAB space, a smaller $\triangle E^{*}$ value indicates better color accuracy or image quality. For the evaluation of the material segmentation, we adopt the standard metric as mean Intersection-over-Union (mIoU). + +# 5.2 Effectiveness of $S^{*}$ , $R^{*}$ , $M^{*}$ Priors + +To avoid the impact of inaccurate decomposition, we begin by systematically integrating the ideal values of these priors (denoted as $S^{*}, R^{*}, M^{*}$ ) individually into the HDRNet baseline for validation. + +Shading prior $S^{*}$ . Tab. 1(a) illustrates that the HDRNet baseline attains a PSNR of 27.75 dB, while the separation of the shading component ( $S^{*}$ ) elevates the PSNR from 27.75 dB to 28.68 dB. Notably, the architecture of HDRNet remains unchanged in this sub-experiment. This finding demonstrates that transforming from RGB into reflectance space of input images with $S^{*}$ is a simple yet effective design for the tone enhancement task, which exhibits enhanced adaptability in diverse high dynamic range scenes. + +Reflectance prior $R^{*}$ . Given the practical limitations of spectral imaging capability on mobile devices, we conduct ablation studies to examine the spectral and spatial resolution configurations of Lr-MSI. We first maintain the spatial resolution at the default setting of $16 \times 16$ and extract channels of different spectral ranges. The Lr-MSI has ten spectral channels, ranging from $400 \mathrm{~nm}$ to + +$1000\mathrm{nm}$ , with each channel featuring a $60~\mathrm{nm}$ bandwidth. Notably, $400 - 760\mathrm{nm}$ encompasses six visible wavelength channels, while $760 - 1000\mathrm{nm}$ includes four near-infrared channels. As observed in Tab. 1(b), the closer to the wavelength of red channel, the higher PSNR it will achieve. The possible explanation lies in the composition of our Mobile-Spec dataset, where a larger proportion of the area is covered by the sky and plant, while the presence of red objects is comparatively limited. Consequently, the HDRNet baseline tends to prioritize the blue and green wavelengths over the red wavelength during the training. Furthermore, the combination of all ten channels $(400 - 1000\mathrm{nm})$ achieves the best results of $29.68\mathrm{dB}$ . In our view, fine-grained spectral information enhances bilateral grids with heightened color perception capabilities and compensates for the imbalanced learning of different color channels. Then we maintain the number of channels at ten, and the spatial resolution is resized from $1\times 1$ to $256\times 256$ . As depicted in Tab. 1(c), there is a trend of improved performance in evaluation metrics with higher spatial resolution. Considering the inherent constraint of spatial resolution in Lr-MSI on mobile devices, we contend that the optimal parameter configuration for Lr-MSI may align to the spatial resolution of the bilateral grid $(16\times 16)$ , which is the 8x downsampling corresponding to the spatial resolution of low-resolution stream input $(256\times 256)$ . Our experimental results suggest that the $16\times 16$ configuration yields satisfactory results, beyond which increasing the spatial resolution leads to diminishing marginal utility. + +Material semantic prior $M^{*}$ . Tab. 2(a) demonstrates that incorporating the Lr-MSI branch into the joint RGB-Spectral decomposition model enhances the mIoU of material segmentation from $71.86\%$ to $78.93\%$ . Despite Lr-MSI's notably lower spatial resolution compared to the 16-bit, it exhibits superior material identification capability owing to its fine-grained spectral channels. Subsequently, we conduct ablation experiments to evaluate the number of material categories. The category number of the original HDRNet baseline is set to one. In the two-category set, the sky, plant, and trunk are grouped into one category, while building, road, and others form another. For the four-category set, the plant and trunk, as well as building and road, are each combined into single categories, whereas the sky and others are treated as separate, independent categories. By refining the division of material categories, the mixture of semantic grid experts becomes more specialized in enhancing tones for specific materials. Tab. 2(b) demonstrates improved quantitative results across three evaluation metrics with the inclusion of six semantic grid experts. + +# 5.3 Ablation Study of JDM-HDRNet + +Tab. 3(a) presents the ablation study incorporating ideal values $S^{*}$ , $R^{*}$ , $M^{*}$ into HDRNet (denoted as JDM-HDRNet\*). The collaboration of shading $S$ and reflectance $R$ priors leads to a substantial decrease in $\triangle E^{*}$ from 5.12 to 3.55 in the CIELAB color space. Moreover, the design of the mixture of semantic grid experts enables specific experts to focus on individual material categories, resulting in a PSNR of 30.14 dB. The above experiments are conducted with ideal priors. + +Table 3: Ablation study of the proposed JDM-HDRNet on the $S$ , $R$ and $M$ priors. +(a) Ablation study on the ideal values of $S^*$ , (b) Ablation study on $S$ , $R$ and $M$ priors predicted from the joint decomposition model. + +
S*R* M*PSNR↑SSIM↑ΔE*↓
HDRNet27.750.9395.12HDRNet27.750.939
JDM-HDRNet*✓28.680.9574.29JDM-HDRNet✓28.590.953
✓29.680.9683.55✓29.570.966
✓30.140.9723.44✓29.830.967
+ +Table 4: Comparisons of our JDM-HDRNet with previous enhancement methods. + +
DPECSRNet3D LUTCLUTSepLUT-SSepLUT-L4D LUTHDRNetUPEOurs
PSNR↑22.8126.3427.5227.3027.5728.0827.7727.7528.1929.83
SSIM↑0.8060.9230.9260.9380.9330.9440.9350.9390.9460.967
△E*↓11.066.445.394.635.374.264.325.124.793.60
+ +Subsequently, the predictions for $S$ , $R$ , $M$ from joint RGB-Spectral decomposition model are integrated with HDRNet (denoted as JDM-HDRNet). Tab. 3(b) shows that the JDM-HDRNet exhibits comparable performance across three metrics in the $S$ set, indicating that the localized brightness adaptation module demonstrates adaptive adjustment capabilities in response to shading image degradation. In addition, we observe that the performance of the reflectance $(R)$ set remains relatively unaffected, achieving an improvement of 0.98 dB. Because both the Lr-MSI and 16-bit RGB images have undergone consistent transformation into the reflectance space using the same $S$ from the joint decomposition model. Regarding the material semantic prior (M), JDM-HDRNet dynamically fuses different grid experts at the bilateral grid level to avoid the inaccurate segmentation of $M$ . Overall, JDM-HDRNet achieves a gain of 2.08 dB compared to the HDRNet baseline, which demonstrates that the low resolution spectral information can be effectively leveraged by the joint decomposition model. + +# 5.4 Comparisons with Previous Methods + +Quantitative Comparisons. We compare JDM-HDRNet with previous color transform-based methods for tone enhancement. In Tab. 4, DPE [6] exhibits inferior performance compared to other methods, attributed to its unpaired setting, leading to output images affected by undesired artifacts introduced by the generative adversarial network. Among the color transform functions, the MLPs based CSRNet [14] exhibits a slightly lower PSNR of 26.34 dB. In contrast, the 3D LUTs based methods, such as 3D LUT [39], CLUT [41], SepLUT-S [36], and 4D LUT [24], demonstrate the similar PSNR ranging from 27.30 dB to 27.77 dB. Notably, SepLUT-L [36] with larger parameters exhibits a superior performance because it simultaneously takes advantage of both 1D and 3D LUTs for enhancement, the 1D LUT adjusts image contrast to achieve a more uniform distribution in an image-adaptive manner. Additionally, UPE [34] estimates an image-to-illumination mapping with constraints of smoothness loss, which is exclusively designed to enhance images under diverse lighting conditions, and it + +![](images/86ef7e3909dbfdd8c134b3bc31ffb95c7233d26e487c8159b17100487647efd9.jpg) +Fig. 5: Qualitative comparisons on the Mobile-Spec dataset. The error maps displayed at the top-right corner illustrate the differences with ground truth for each image. + +achieves good performance in handling high dynamic range scenes in the Mobile-Spec dataset. Previous methods face challenges in learning accurate tone enhancement for HDR scenes, while our priors-guided JDM-HDRNet outperforms previous methods by effectively leveraging $S$ , $R$ , $M$ priors predicted from the joint RGB-Spectral decomposition model to provide explicit guidance. + +Qualitative Comparisons. Fig. 5 presents qualitative comparisons of two representative examples. In the first row, JDM-HDRNet exhibits enhanced contrast and a wider dynamic range in the plant area, while other methods appear comparatively dim. In the second row, JDM-HDRNet shows reduced color deviation with the ground truth in the wall area, indicating the beneficial role of spectral information integration in improving photographic realism. In summary, JDM-HDRNet, guided by $S$ , $R$ and $M$ priors, achieves more accurate and pleasing colors for high dynamic range scenes. + +Limitations. The spatial resolution of Lr-MSI in this study is fixed at $16 \times 16$ , a setting not yet achievable on commercial smartphones. Increasing the spatial resolution would improve the accuracy of the joint decomposition model. Further exploration is required to address the constrained spectral imaging capabilities in the context of image enhancement and related applications on mobile devices. + +# 6 Conclusion + +We have investigated the effectiveness of Lr-MSI in the mobile ISP pipeline for enhancing tone enhancement. We first construct the high-quality Mobile-Spec, a pioneering dataset comprising both aligned spectral and RGB images. Then we analyze the role of Lr-MSI from the perspective of spectral image intrinsic decomposition. Shading, reflectance and material semantic priors are derived using the joint decomposition model to address the inherent complexity of spectral images and the constraints of spectral imaging capabilities. Subsequently, JDM-HDRNet is meticulously designed under the explicit guidance of above priors, which exhibits superior performance than the competitive methods. Our dataset and method are expected to attract further research into this area and explore the untapped potential of spectral information on mobile devices. + +# Acknowledgments + +This research was supported by National Key Research and Development Program of China (2023YFF0713300), National Natural Science Foundation of China under Grant (62071216), Leading Technology of Jiangsu Basic Research Plan under Grant (BK20192003), and National Research Foundation Singapore Competitive Research Program (award number CRP29-2022-0003). We thank the device support from NJU-ZPTECH University-Enterprise Joint Laboratory for Computational Spectral Imaging. + +# References + +1. Gaiasky-mini hyperspectral imaging camera. https://www.dualix.com.cn/en/Goods/desc/id/123/aid/954.html +2. Spectral Sensing -ams.com. https://ams.com/spectral-sensing +3. Barrow, H., Tenenbaum, J., Hanson, A., Riseman, E.: Recovering intrinsic scene characteristics. Comput. vis. syst 2(3-26), 2 (1978) +4. Cai, J., Gu, S., Zhang, L.: Learning a deep single image contrast enhancer from multi-exposure images. IEEE Transactions on Image Processing 27(4), 2049-2062 (2018) +5. Chen, X., Zhu, W., Zhao, Y., Yu, Y., Zhou, Y., Yue, T., Du, S., Cao, X.: Intrinsic decomposition from a single spectral image. Applied optics 56(20), 5676-5684 (2017) +6. Chen, Y.S., Wang, Y.C., Kao, M.H., Chuang, Y.Y.: Deep photo enhancer: Unpaired learning for image enhancement from photographs with gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6306-6314 (2018) +7. Cheng, Z., Zheng, Y., You, S., Sato, I.: Non-local intrinsic decomposition with near-infrared priors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2521-2530 (2019) +8. Erba, I., et al.: Computational color constancy beyond rgb images: Multispectral and temporal extensions (2024) +9. Finlayson, G.D., Zhu, Y.: Designing color filters that make cameras more colorimetric. IEEE Transactions on Image Processing 30, 853-867 (2020) +10. Fu, X., Ying, Y.: Food safety evaluation based on near infrared spectroscopy and imaging: A review. Critical Reviews in Food Science and Nutrition 56(11), 1913-1924 (2016) +11. Gharbi, M., Chen, J., Barron, J.T., Hasinoff, S.W., Durand, F.: Deep bilateral learning for real-time image enhancement. ACM Transactions on Graphics (TOG) 36(4), 1-12 (2017) +12. Gijsenij, A., Gevers, T., Van De Weijer, J.: Computational color constancy: Survey and experiments. IEEE transactions on image processing 20(9), 2475-2489 (2011) +13. Glatt, O., Ater, Y., Kim, W.S., Werman, S., Berby, O., Zini, Y., Zelinger, S., Lee, S., Choi, H., Soloveichik, E.: Beyond rgb: A real world dataset for multispectral imaging in mobile devices. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4344-4354 (2024) +14. He, J., Liu, Y., Qiao, Y., Dong, C.: Conditional sequential modulation for efficient global image retouching. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16. pp. 679-695. Springer (2020) + +15. He, Q., Wang, R.: Hyperspectral imaging enabled by an unmodified smartphone for analyzing skin morphological features and monitoring hemodynamics. Biomedical optics express 11(2), 895-910 (2020) +16. He, Q., Wang, R.K.: Analysis of skin morphological features and real-time monitoring using snapshot hyperspectral imaging. Biomedical optics express 10(11), 5625-5638 (2019) +17. Huang, Q., Zhu, W., Zhao, Y., Chen, L., Wang, Y., Yue, T., Cao, X.: Multispectral image intrinsic decomposition via subspace constraint. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6430-6439 (2018) +18. Jiang, J., Liu, D., Gu, J., Süssstrunk, S.: What is the space of spectral sensitivity functions for digital color cameras? In: 2013 IEEE Workshop on Applications of Computer Vision (WACV). pp. 168-179. IEEE (2013) +19. Karaimer, H.C., Brown, M.S.: A software platform for manipulating the camera imaging pipeline. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 429-444. Springer (2016) +20. Kim, S.J., Lin, H.T., Lu, Z., Susstrunk, S., Lin, S., Brown, M.S.: A new in-camera imaging model for color computer vision and its application. IEEE Transactions on Pattern Analysis and Machine Intelligence 34(12), 2289-2302 (2012) +21. Kim, Y.T.: Contrast enhancement using brightness preserving bi-histogram equalization. IEEE transactions on Consumer Electronics 43(1), 1-8 (1997) +22. Land, E.H., McCann, J.J.: Lightness and retinex theory. Josa 61(1), 1-11 (1971) +23. Lindholm, V., Raita-Hakola, A.M., Annala, L., Salmivuori, M., Jeskanen, L., Saari, H., Koskenmies, S., Pitkänen, S., Pölönen, I., Isoherranen, K., et al.: Differentiating malignant from benign pigmented or non-pigmented skin tumours—a pilot study on 3d hyperspectral imaging of complex skin surfaces and convolutional neural networks. Journal of clinical medicine 11(7), 1914 (2022) +24. Liu, C., Yang, H., Fu, J., Qian, X.: 4d lut: learnable context-aware 4d lookup table for image enhancement. IEEE Transactions on Image Processing 32, 4742-4756 (2023) +25. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015) +26. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International journal of computer vision 60, 91-110 (2004) +27. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) +28. Mantiuk, R., Daly, S., Kerofsky, L.: Display adaptive tone mapping. In: ACM SIGGRAPH 2008 papers, pp. 1-10 (2008) +29. McGonigle, A.J., Wilkes, T.C., Pering, T.D., Willmott, J.R., Cook, J.M., Mims III, F.M., Parisi, A.V.: Smartphone spectrometers. Sensors 18(1), 223 (2018) +30. Ng, P.C., Chi, Z., Verdie, Y., Lu, J., Plataniotis, K.N.: Hyper-skin: A hyperspectral dataset for reconstructing facial skin-spectra from rgb images. Advances in Neural Information Processing Systems 36 (2024) +31. Rao, S., Huang, Y., Cui, K., Li, Y.: Anti-spoofing face recognition using a metasurface-based snapshot hyperspectral image sensor. Optica 9(11), 1253-1259 (2022) +32. e Silva, G.M., Garcia, J.A., de Alencar Garitta, J., Cunha, D.G.F., Finkler, N.R., Mendiondo, E.M., Ghiglieno, F.: Smartphone-based spectrometry system as a prescreening assessment of copper and iron for real time control of water pollution. Journal of Environmental Management 323, 116214 (2022) + +33. Thomas, J.B.: Illuminant estimation from uncalibrated multispectral images. In: 2015 Colour and Visual Computing Symposium (CVCS). pp. 1-6. IEEE (2015) +34. Wang, R., Zhang, Q., Fu, C.W., Shen, X., Zheng, W.S., Jia, J.: Underexposed photo enhancement using deep illumination estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6849-6857 (2019) +35. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004) +36. Yang, C., Jin, M., Xu, Y., Zhang, R., Chen, Y., Liu, H.: Seplut: Separable image-adaptive lookup tables for real-time image enhancement. In: European Conference on Computer Vision. pp. 201-217. Springer (2022) +37. Yang, Z., Albrow-Owen, T., Cai, W., Hasan, T.: Miniaturization of optical spectrometers. Science 371(6528), eabe0722 (2021) +38. Yuan, L., Sun, J.: Automatic exposure correction of consumer photographs. In: Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part IV 12. pp. 771-785. Springer (2012) +39. Zeng, H., Cai, J., Li, L., Cao, Z., Zhang, L.: Learning image-adaptive 3d lookup tables for high performance photo enhancement in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(4), 2058-2073 (2020) +40. Zhang, F., You, S., Li, Y., Fu, Y.: Hsi-guided intrinsic image decomposition for outdoor scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 313-322 (2022) +41. Zhang, F., Zeng, H., Zhang, T., Zhang, L.: Clut-net: Learning adaptively compressed representations of 3dluts for lightweight image enhancement. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 6493-6501 (2022) +42. Zheng, Y., Sato, I., Sato, Y.: Illumination and reflectance spectra separation of a hyperspectral image meets low-rank matrix factorization. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1779-1787 (2015) \ No newline at end of file diff --git a/2024/Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography/images.zip b/2024/Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..60ac91734313002f36cb43065ae7407f949c6ac0 --- /dev/null +++ b/2024/Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3857e6cd804770fe0ab4af755433736ecb8f2e243beb9ad9b3ca4dff6abcb20 +size 415276 diff --git a/2024/Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography/layout.json b/2024/Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3b064ec502163b3f453d769c085b3596f93d55db --- /dev/null +++ b/2024/Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography/layout.json @@ -0,0 +1,10437 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 111, + 455, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 111, + 455, + 166 + ], + "spans": [ + { + "bbox": [ + 159, + 111, + 455, + 166 + ], + "type": "text", + "content": "Joint RGB-Spectral Decomposition Model Guided Image Enhancement in Mobile Photography" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 151, + 186, + 462, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 186, + 462, + 212 + ], + "spans": [ + { + "bbox": [ + 151, + 186, + 462, + 212 + ], + "type": "text", + "content": "Kailai Zhou1, Lijing Cai1, Yibo Wang1, Mengya Zhang1, Bihan Wen3, Qiu Shen†,2, and Xun Cao1,2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 219, + 474, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 219, + 474, + 231 + ], + "spans": [ + { + "bbox": [ + 138, + 219, + 474, + 231 + ], + "type": "text", + "content": "1 School of Electronic Science and Engineering, Nanjing University, Nanjing, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 231, + 453, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 231, + 453, + 242 + ], + "spans": [ + { + "bbox": [ + 159, + 231, + 453, + 242 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 159, + 231, + 453, + 242 + ], + "type": "text", + "content": " Key Laboratory of Optoelectronic Devices and Systems with Extreme" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 242, + 426, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 242, + 426, + 253 + ], + "spans": [ + { + "bbox": [ + 187, + 242, + 426, + 253 + ], + "type": "text", + "content": "Performances of MOE, Nanjing University, Nanjing, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 210, + 253, + 403, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 253, + 403, + 265 + ], + "spans": [ + { + "bbox": [ + 210, + 253, + 403, + 265 + ], + "type": "text", + "content": "3 Nanyang Technological University, Singapore" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 161, + 265, + 452, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 265, + 452, + 276 + ], + "spans": [ + { + "bbox": [ + 161, + 265, + 452, + 276 + ], + "type": "text", + "content": "{calayzhou, cailijing, ybwang, 522023230161}@smail.nju.edu.cn," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 197, + 277, + 416, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 277, + 416, + 286 + ], + "spans": [ + { + "bbox": [ + 197, + 277, + 416, + 286 + ], + "type": "text", + "content": "bihan.wen@ntu.edu.sg, {caoxun, shenqiu}@nju.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 318, + 455, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 318, + 455, + 515 + ], + "spans": [ + { + "bbox": [ + 159, + 318, + 455, + 515 + ], + "type": "text", + "content": "Abstract. The integration of miniaturized spectrometers into mobile devices offers new avenues for image quality enhancement and facilitates novel downstream tasks. However, the broader application of spectral sensors in mobile photography is hindered by the inherent complexity of spectral images and the constraints of spectral imaging capabilities. To overcome these challenges, we propose a joint RGB-Spectral decomposition model guided enhancement framework, which consists of two steps: joint decomposition and prior-guided enhancement. Firstly, we leverage the complementarity between RGB and Low-resolution Multi-Spectral Images (Lr-MSI) to predict shading, reflectance, and material semantic priors. Subsequently, these priors are seamlessly integrated into the established HDRNet to promote dynamic range enhancement, color mapping, and grid expert learning, respectively. Additionally, we construct a high-quality Mobile-Spec dataset to support our research, and our experiments validate the effectiveness of Lr-MSI in the tone enhancement task. This work aims to establish a solid foundation for advancing spectral vision in mobile photography. The code is available at https://github.com/CalayZhou/JDM-HDRNet." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 159, + 526, + 454, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 526, + 454, + 549 + ], + "spans": [ + { + "bbox": [ + 159, + 526, + 454, + 549 + ], + "type": "text", + "content": "Keywords: Image enhancement " + }, + { + "bbox": [ + 159, + 526, + 454, + 549 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 526, + 454, + 549 + ], + "type": "text", + "content": " RGB-Spectral fusion " + }, + { + "bbox": [ + 159, + 526, + 454, + 549 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 526, + 454, + 549 + ], + "type": "text", + "content": " Spectral image intrinsic decomposition " + }, + { + "bbox": [ + 159, + 526, + 454, + 549 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 526, + 454, + 549 + ], + "type": "text", + "content": " Spectral dataset and application" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 571, + 230, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 571, + 230, + 583 + ], + "spans": [ + { + "bbox": [ + 132, + 571, + 230, + 583 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 482, + 647 + ], + "type": "text", + "content": "Recent advancements in miniaturized spectrometers [29,37] have facilitated their integration into mobile devices, thereby expanding the scope of possible applications such as medical diagnosis [15], food safety evaluation [10], anti-spoofing face recognition [31] and color analysis [9]. Nevertheless, spectral sensors [2] in" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 114, + 462, + 182 + ], + "blocks": [ + { + "bbox": [ + 153, + 114, + 462, + 182 + ], + "lines": [ + { + "bbox": [ + 153, + 114, + 462, + 182 + ], + "spans": [ + { + "bbox": [ + 153, + 114, + 462, + 182 + ], + "type": "image", + "image_path": "ebb7f5c05940a03ff7132713b05d71fa271a380e13d856d5bf3426e999c78ed4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 191, + 482, + 215 + ], + "lines": [ + { + "bbox": [ + 131, + 191, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 131, + 191, + 482, + 215 + ], + "type": "text", + "content": "Fig. 1: The Lr-MSI is decomposed into the illumination " + }, + { + "bbox": [ + 131, + 191, + 482, + 215 + ], + "type": "inline_equation", + "content": "L(\\lambda)" + }, + { + "bbox": [ + 131, + 191, + 482, + 215 + ], + "type": "text", + "content": ", reflectance " + }, + { + "bbox": [ + 131, + 191, + 482, + 215 + ], + "type": "inline_equation", + "content": "R(\\lambda, x)" + }, + { + "bbox": [ + 131, + 191, + 482, + 215 + ], + "type": "text", + "content": " and shading " + }, + { + "bbox": [ + 131, + 191, + 482, + 215 + ], + "type": "inline_equation", + "content": "S(x)" + }, + { + "bbox": [ + 131, + 191, + 482, + 215 + ], + "type": "text", + "content": " terms. We then analyze their respective roles in the mobile ISP pipeline." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 241, + 479, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 479, + 312 + ], + "type": "text", + "content": "mobile photography are mainly employed for illuminant estimation in the Auto White Balance (AWB) operation at present. The spectral information holds potential to enhance the aesthetic appeal and perceptual quality of images. Exploring how to leverage additional spectral information to improve other image enhancement operations in the mobile Image Signal Processing (ISP) pipeline is an area requiring deeper investigation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 315, + 480, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 480, + 458 + ], + "type": "text", + "content": "We attempt to make a deeper analysis of the role of Lr-MSI in mobile photography by drawing insights from the spectral image intrinsic decomposition. As illustrated in Fig. 1, the spectral image can be decomposed into its constituent terms [5]: illumination curve " + }, + { + "bbox": [ + 130, + 315, + 480, + 458 + ], + "type": "inline_equation", + "content": "L(\\lambda)" + }, + { + "bbox": [ + 130, + 315, + 480, + 458 + ], + "type": "text", + "content": ", reflectance " + }, + { + "bbox": [ + 130, + 315, + 480, + 458 + ], + "type": "inline_equation", + "content": "R(\\lambda, x)" + }, + { + "bbox": [ + 130, + 315, + 480, + 458 + ], + "type": "text", + "content": ", and shading " + }, + { + "bbox": [ + 130, + 315, + 480, + 458 + ], + "type": "inline_equation", + "content": "S(x)" + }, + { + "bbox": [ + 130, + 315, + 480, + 458 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 315, + 480, + 458 + ], + "type": "inline_equation", + "content": "L(\\lambda)" + }, + { + "bbox": [ + 130, + 315, + 480, + 458 + ], + "type": "text", + "content": " captures the spectral characteristics of the light source and has traditionally been applied for illuminant estimation [8]. " + }, + { + "bbox": [ + 130, + 315, + 480, + 458 + ], + "type": "inline_equation", + "content": "R(\\lambda, x)" + }, + { + "bbox": [ + 130, + 315, + 480, + 458 + ], + "type": "text", + "content": " embodies the albedo invariant color and texture of the material. The reflectance of spectral images which possesses fine-grained color channels shows promise for material segmentation and color mapping tasks. " + }, + { + "bbox": [ + 130, + 315, + 480, + 458 + ], + "type": "inline_equation", + "content": "S(x)" + }, + { + "bbox": [ + 130, + 315, + 480, + 458 + ], + "type": "text", + "content": " represents interaction between the geometry of objects and illumination, which may contribute to the local brightness adjustment of high dynamic scenes. We contend that the potential of the reflectance " + }, + { + "bbox": [ + 130, + 315, + 480, + 458 + ], + "type": "inline_equation", + "content": "R(\\lambda, x)" + }, + { + "bbox": [ + 130, + 315, + 480, + 458 + ], + "type": "text", + "content": " and shading " + }, + { + "bbox": [ + 130, + 315, + 480, + 458 + ], + "type": "inline_equation", + "content": "S(x)" + }, + { + "bbox": [ + 130, + 315, + 480, + 458 + ], + "type": "text", + "content": " terms has not been thoroughly explored." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 460, + 481, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 481, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 481, + 639 + ], + "type": "text", + "content": "In this paper, we explore the application of low-resolution multi-spectral images in the tone enhancement task within the mobile ISP pipeline, especially for outdoor High Dynamic Range (HDR) scenes. The objective of tone enhancement is to adjust the brightness, contrast, and color characteristics of an image to enhance details in both the highlights and shadows while preserving its natural appearance. The primary challenge is the heterogeneous information fusion involving the integration of extra Lr-MSI with the original RGB workflow. This challenge manifests in two aspects: Firstly, variations in imaging mechanisms lead to the inherent complexity of spectral images, encompassing factors like scene geometry, inter-reflections and intricate artificial illumination [5], making it difficult to directly integrate spectral information into the mobile ISP workflow. Secondly, despite commercial spectral sensors offering over ten spectral channels [2], their spatial resolution is often confined due to the constrained spectral imaging capabilities on mobile devices. This limited spatial resolution remains a problem for the application of Lr-MSI in tone enhancement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "To address the above challenges, we propose a joint RGB-Spectral decomposition model guided image enhancement framework which consists of two" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "K Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 366 + ], + "type": "text", + "content": "phases: joint decomposition and prior-guided enhancement. The effectiveness of this framework relies on three assumptions: (1) The near-infrared band in Lr-MSI can serve as an approximation of the shading term [7]. (2) Lr-MSI and RGB images exhibit complementary characteristics in both spatial and spectral resolutions. (3) The increased color channels in Lr-MSI contribute to material segmentation. Firstly, to mitigate limited spectral imaging capabilities, the joint decomposition phase leverages the complementarity between Lr-MSI and RGB images to predict shading, reflectance, and material semantic priors. Secondly, to overcome the inherent complexity of spectral images, the prior-guided enhancement phase seamlessly integrates these priors into the established HDRNet [11] to provide clear mechanistic guidance. Specifically, the shading component is separated from original RGB images to enhance the adaptability in dealing with localized brightness variations. Subsequently, the " + }, + { + "bbox": [ + 133, + 116, + 481, + 366 + ], + "type": "inline_equation", + "content": "R(\\lambda, x)" + }, + { + "bbox": [ + 133, + 116, + 481, + 366 + ], + "type": "text", + "content": " of Lr-MSI is employed to guide the prediction of bilateral grids via the spectral perception map, particularly within the reflectance domain. Harnessing the reflectance domain alongside higher-dimensional spectral channels facilitates enhanced learning of color relationships for tone mapping. Lastly, we introduce the mixture of semantic grid experts to specifically adapt to the distinct color characteristics of individual material categories. The Joint DecoMposition framework guided HDRNet (JDM-HDRNet) exhibits superior performance compared to previous enhancement methods in terms of PSNR, SSIM [35] and a color difference metric." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 368, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 368, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 133, + 368, + 481, + 498 + ], + "type": "text", + "content": "The contributions of this paper are as follows: (1) The pioneering Mobile-Spec dataset is meticulously constructed with high quality aligned RGB and hyperspectral images, which serves as a solid foundation for spectral applications. (2) We propose a joint RGB-Spectral decomposition model guided enhancement framework to address challenges of inherent complexity of spectral images and limited spectral imaging capabilities. The Lr-MSI and RGB images are decomposed into shading, reflectance and material semantic priors, and these priors are employed for promoting dynamic range enhancement, color mapping, and semantic grid expert learning, respectively. (3) Experiments demonstrate the effectiveness of introducing extra Lr-MSI in the tone enhancement task, which provides valuable exploration for the industrial application of spectral sensors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 515, + 236, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 515, + 236, + 528 + ], + "spans": [ + { + "bbox": [ + 133, + 515, + 236, + 528 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 539, + 343, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 539, + 343, + 551 + ], + "spans": [ + { + "bbox": [ + 133, + 539, + 343, + 551 + ], + "type": "text", + "content": "2.1 Potential of Mobile Spectral Sensors" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 557, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 557, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 557, + 481, + 665 + ], + "type": "text", + "content": "As the cost of miniaturized spectrometers [29,37] decreases and their availability becomes more widespread, the integration of these spectral sensors with mobile devices creates new opportunities. Studies have demonstrated the potential of hyperspectral imaging in medical applications [15,16,23,30] such as hemodynamics monitoring [15] and skin analysis [16,30]. Moreover, smartphones equipped with snapshot hyperspectral sensors can enhance anti-spoofing face recognition [31] by preventing high-quality 3D mask attacks. Other applications include food safety evaluation [10], color analysis [9], and water quality assessment [32]. Despite these advancements, the use of spectral sensors in mobile photography remains" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Joint Decomposition Model Guided Image Enhancement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "limited. As RGB sensors lack inherent illumination adaptation abilities to maintain color constancy, in recent years, most high-end smartphones are equipped with spectral sensors which aim to improve the accuracy of illuminant estimation [8, 13, 33] in an uncontrolled environment. The spectral sensor has evolved from its origins as an ambient light sensor to the refined capture of spectral signatures, primarily employed in auto white balance because its spatial resolution is often confined to a single pixel [2]. Our objective is to validate the applicability of Lr-MSI to other enhancement tasks within the mobile ISP pipeline, particularly under assumptions where the constraints of spectral imaging are mitigated by dual-camera acquisition in the Mobile-Spec dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 252, + 369, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 252, + 369, + 264 + ], + "spans": [ + { + "bbox": [ + 131, + 252, + 369, + 264 + ], + "type": "text", + "content": "2.2 Enhancement in the Mobile ISP Pipeline" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 271, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 523 + ], + "type": "text", + "content": "A standard mobile ISP pipeline [19] comprises both the Bayer processing routine and the subsequent photo-finishing routine. There are several enhancement procedures such as exposure correction [38], white-balance [12], contrast enhancement [4], color manipulation [20], local and global tone-mapping [28]. In this study, we aim to fully exploit the intrinsic decomposition components of spectral images by integrating Lr-MSI into the tone enhancement. This task involves adjusting the tonal range while preserving overall appearance and details, especially for HDR scenes. The commonly used method is the histogram equalization [21] which globally improves the contrast of entire images. However, global operations often fall short when dealing with scenes of high dynamic range. In the deep learning era, color transform-based methods are often employed for real-time processing of high-resolution images in a collaborative global-local manner. According to the color transform functions, these methods can be classified into affine transformation matrices [11, 34], multilayer perceptrons (MLPs) [14], 3D LUTs [36, 39, 41], etc. The pioneering work is the HDRNet [11] which learns the bilateral grid of coefficients at low resolution and performs transformations from input to output at full resolution. 3D LUT [39] replaces the affine transformation matrices with image-adaptive 3-dimensional lookup tables, enabling rapid and robust photo enhancement. We choose the HDRNet as the baseline due to its simple and flexible network design. The key challenge lies in the heterogeneous information fusion involving the integration of extra Lr-MSI with the HDRNet." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 539, + 361, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 539, + 361, + 552 + ], + "spans": [ + { + "bbox": [ + 131, + 539, + 361, + 552 + ], + "type": "text", + "content": "2.3 Spectral Image Intrinsic Decomposition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 667 + ], + "type": "text", + "content": "To fuse the extra Lr-MSI into the original RGB workflow in HDRNet, we consider the role of spectral information in mobile photography from the perspective of Spectral Intrinsic Image Decomposition (SIID) [5]. The SIID problem aims to separate a spectral image into more elaborate intrinsic properties, such as the spectrum of the illumination, reflectance and shading components. This decomposition is typically performed in the spectral domain, which allows much more flexibility in the algorithm design and leads to a more accurate analysis. Zheng et al. [42] model illumination and reflectance spectra separation of a hyperspectral image into a low-rank matrix factorization problem. Chen et al. [5]" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "K Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 462, + 208 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 462, + 208 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 462, + 208 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 462, + 208 + ], + "type": "image", + "image_path": "1b73e1aef31acaf251064c69286f2f99f982ab5b6ce9072879d3fbd1c0cca14e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 216, + 480, + 271 + ], + "lines": [ + { + "bbox": [ + 131, + 216, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 131, + 216, + 480, + 271 + ], + "type": "text", + "content": "Fig. 2: (a) The Mobile-Spec dataset comprises RGB images (16-bit input and 8-bit target), hyperspectral images and their corresponding shading, reflectance and material segmentation images. Near-infrared images serve as the guide map to approximate the shading term. (b) Spectral responses of different material categories. (c-d) Visualizations of RGB and hyperspectral data for different categories with t-SNE [27]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 296, + 482, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 463 + ], + "type": "text", + "content": "resolve a spectral image into independent intrinsic components: illumination, shading, and reflectance, and develop an effective algorithm which first reduces the spatial and spectral resolution with the super-pixel down-sampling, then the SIID problem is optimized under local constant and global sparse constraint. Huang et al. [17] extend the Retinex model to the multispectral domain based on the subspace constraint, which assumes the reflectance and shading vectors both live in a low dimensional subspace along the spectral domain. SIID is an ill-posed and under-constrained problem and previous algorithms are developed under the ideal lab environment, underscoring the challenges of tackling complex outdoor scenes. Fortunately, Cheng et al. [7] find that near-infrared images can function as reliable approximations of shading priors, and Zhang et al. [40] further perform intrinsic decomposition for outdoor scenes using aligned RGB and hyperspectral images. Based on this assumption, our joint decomposition model can be trained in an end-to-end manner, which will be discussed in Section 4.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 483, + 304, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 304, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 304, + 496 + ], + "type": "text", + "content": "3 The Mobile-Spec Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 581 + ], + "type": "text", + "content": "To the best of our knowledge, no existing image enhancement dataset includes both aligned hyperspectral and RGB images. To bridge this data gap, we have constructed the high-quality Mobile-Spec dataset, consisting of 200 image groups. As shown in Fig. 2(a), each group represents an independent scene and includes a 16-bit input image, an 8-bit target image, a hyperspectral image, and the corresponding shading, reflectance, and material segmentation images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Dual Camera System. To address the limited spectral imaging capabilities of sensors on smartphones, the Mobile-Spec is captured using a dual camera system: a high-end commercial smartphone and a Gaiasky-mini2 scanning hyperspectral camera [1]. The Gaiasky-mini2 covers a spectral range from 400 nm to 1000 nm with 176 spectral channels. The 16-bit input images are fused from multiple exposure frames, and the 8-bit target images are generated by a commercial privacy model that integrates the aesthetics of multiple experts." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Joint Decomposition Model Guided Image Enhancement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "Image Matching. The hyperspectral image size is " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "1057 \\times 960 \\times 176" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": ", while the 16-bit input and 8-bit target RGB images are " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "4096 \\times 3072 \\times 3" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": ". To align RGB images with a wide angle of view onto hyperspectral images, we employ an automatic alignment algorithm that computes SIFT descriptors [26] and estimates homography transformation matrices. Consequently, both RGB and hyperspectral images achieve a consistent spatial resolution of " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "1057 \\times 960" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 248 + ], + "type": "text", + "content": "Influence of Light Sources. The Mobile-Spec dataset is captured exclusively outdoors to avoid the impact of intricate indoor lighting. Solar radiation spans a broad range of wavelengths, ranging from ultraviolet to visible light and extending into the infrared region. It is reasonable to assume that near-infrared images can approximate shading priors in outdoor settings [7]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 249, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 368 + ], + "type": "text", + "content": "Dataset Collection and Filtering. Our dataset spans multiple seasons and includes five categories commonly observed in outdoor environments: sky, building, plant, trunk, and road. Their respective spectral curves are illustrated in Fig. 2(b). Mobile-Spec undergoes rigorous filtering to ensure high quality. Firstly, samples are selected based on dynamic range, determined by an automated algorithm analyzing the difference between maximum and minimum pixel intensities. Secondly, samples exhibiting significant alignment errors are filtered out. Thirdly, samples with unsatisfactory visual quality are excluded, considering factors such as chromatic aberration, sharpness, noise, and artifacts. Lastly, to augment dataset diversity, scenes with high similarity are eliminated." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "type": "text", + "content": "Shading, Reflectance and Material Segmentation. To obtain targets of " + }, + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "type": "text", + "content": " for the joint decomposition model, shading images are approximated by averaging near-infrared bands from " + }, + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "type": "inline_equation", + "content": "850~\\mathrm{nm}" + }, + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "type": "inline_equation", + "content": "1000~\\mathrm{nm}" + }, + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "type": "text", + "content": " in hyperspectral images. Reflectance images can be derived based on the Retinex theory [3, 22]. Ground truths of material segmentation are meticulously labeled by human annotators. More details about the Mobile-Spec are provided in the appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 462, + 261, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 462, + 261, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 462, + 261, + 475 + ], + "type": "text", + "content": "4 Proposed Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 490, + 375, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 375, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 375, + 502 + ], + "type": "text", + "content": "4.1 Joint RGB-Spectral Decomposition Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 513, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 513, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 513, + 482, + 561 + ], + "type": "text", + "content": "To integrate low-resolution spectral images into the RGB tone enhancement workflow, we first analyze the differences between hyperspectral and RGB camera imaging mechanisms. The imaging model for spectral cameras can be represented as follows based on [5,18]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 183, + 571, + 481, + 599 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 571, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 183, + 571, + 481, + 599 + ], + "type": "interline_equation", + "content": "I _ {k, x} = \\int_ {4 0 0 \\mathrm {n m}} ^ {1 0 0 0 \\mathrm {n m}} C _ {k} (\\lambda) L (\\lambda) S (x) R (\\lambda , x) d \\lambda , \\mathrm {k} = 1, 2, 3 \\dots \\tag {1}", + "image_path": "3b47089c8789009c9e476b67e429a75937a66b965854d640fca3264adfe0df88.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "I_{k,x}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " represents the pixel intensities at position " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " of the k-th band, " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "C_k(\\lambda)" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " denotes the camera sensitivity function, " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "L(\\lambda)" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " indicates the illumination curve of the light source, " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "S(x)" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " refers to shading, and " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "R(\\lambda ,x)" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " is the reflectance of scenes. " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "L(\\lambda)" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " is usually utilized for illuminant estimation in the auto white balance procedure, while the role of the shading " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "S(x)" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " and reflectance " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "R(\\lambda ,x)" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " components" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "K Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 300, + 198 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 300, + 198 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 300, + 198 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 300, + 198 + ], + "type": "image", + "image_path": "85361946d33ea537467f1f1d69b82e3613ed8b408f18d5c1728367d1d15b797f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 208, + 482, + 253 + ], + "lines": [ + { + "bbox": [ + 130, + 208, + 482, + 253 + ], + "spans": [ + { + "bbox": [ + 130, + 208, + 482, + 253 + ], + "type": "text", + "content": "Fig. 3: (a) The joint RGB-Spectral decomposition model leverages the complementarity between RGB images and Lr-MSIs to predict shading (S), reflectance (R) and material category (M) priors. (b) The pixel intensity histogram. (c-d) The distribution of Pearson correlation coefficient " + }, + { + "bbox": [ + 130, + 208, + 482, + 253 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 208, + 482, + 253 + ], + "type": "text", + "content": " in the original RGB space and reflectance space." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 303, + 114, + 477, + 198 + ], + "blocks": [ + { + "bbox": [ + 303, + 114, + 477, + 198 + ], + "lines": [ + { + "bbox": [ + 303, + 114, + 477, + 198 + ], + "spans": [ + { + "bbox": [ + 303, + 114, + 477, + 198 + ], + "type": "image", + "image_path": "fb09ad04b2291750e192938aa2437412d87e89e9ad3e6b4a335f78a6f8750a73.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "type": "text", + "content": "has been rarely explored. The core difference between the RGB and hyperspectral cameras lies in " + }, + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "type": "inline_equation", + "content": "C_k(\\lambda)" + }, + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "type": "text", + "content": ", where hyperspectral images have fine-grained color channels due to a higher spectral sampling rate. To simulate the limited spectral imaging capability on smartphones, we downsample the hyperspectral image " + }, + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "type": "inline_equation", + "content": "(1057 \\times 960 \\times 176)" + }, + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "type": "text", + "content": " to the Lr-MSI " + }, + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "type": "inline_equation", + "content": "(16 \\times 16 \\times 10)" + }, + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "type": "text", + "content": ". This downsampling is conducted to match the typical 10-channel configuration of commercial smartphone spectral sensors. [2]. We maintain flexibility in setting the spatial resolution to a default of " + }, + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "type": "text", + "content": ", which will be discussed in Section 5.2. We posit that the effect of Lr-MSI in the joint decomposition model can be observed in two aspects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 388, + 482, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 519 + ], + "type": "text", + "content": "Approximate estimation of shading priors with the near-infrared band. To obtain the shading term, traditional optimization-based methods design various hand-crafted priors to constrain the solution [5,17], which are often inapplicable in complex outdoor scenes. Recently, near-infrared images have offered a simple yet powerful approach for shading smoothness [7], as the spectral curves of different colors gradually flatten out and texture variation is considerably reduced in the near-infrared band (refer to the appendix). Given that the Mobile-Spec dataset is exclusively captured outdoors to avoid the impact of complex indoor light sources, and considering the wide spectrum of solar radiation covering the infrared region, the shading term can be predicted through the joint decomposition of Lr-MSI and RGB images, as they share the same " + }, + { + "bbox": [ + 130, + 388, + 482, + 519 + ], + "type": "inline_equation", + "content": "S(x)" + }, + { + "bbox": [ + 130, + 388, + 482, + 519 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "text", + "content": "Enhancing material segmentation with the visible band. We employ t-SNE [27] to analyze the complex spectral patterns and relationships between different spectral bands. In Fig. 2(c), clusters representing different categories exhibit denser aggregation in the t-SNE visualization of hyperspectral images compared to that of RGB images. This observation suggests that the increased granularity of spectral channels provides enhanced discriminative capacity among various materials, and Lr-MSI may potentially contribute to the material segmentation task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In Fig. 3(a), we employ a joint RGB-Spectral decomposition model to leverage both near-infrared and visible bands in Lr-MSI for shading prediction and material segmentation. Lr-MSI provides high spectral resolution but limited spatial resolution, while the RGB image offers rich spatial information but" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Joint Decomposition Model Guided Image Enhancement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": "lower spectral resolution. We transform the prediction of shading images from the regression problem to the classification problem by dividing shading images into eight brightness levels. Consequently, shading prediction and material segmentation can be decomposed in a collaborative manner. The joint RGB-Spectral decomposition model incorporates two independent encoders and decoders, adapted from the basic segmentation model FCN [25]. Specifically, LrMSI is resized to match the spatial resolution of the RGB image. The RGB encoder " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "E_{rgb}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " and spectral encoder " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "E_{spec}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " are trained to project the 16-bit input image " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "I_{rgb}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " and Lr-MSI " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "I_{msi}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " to the same latent space for representation alignment: " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{rgb} = E_{rgb}(I_{rgb})" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{spec} = E_{spec}(I_{msi})" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{rgb}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{spec}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " are fused together with the concatenation to share the common and complementary feature representations. Material segmentation " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " and shading " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " are independently predicted with the decoder " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "D_{m}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "D_{s}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " based on the fused feature representation. The joint RGB-Spectral decomposition model can be formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 233, + 297, + 481, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 297, + 481, + 309 + ], + "spans": [ + { + "bbox": [ + 233, + 297, + 481, + 309 + ], + "type": "interline_equation", + "content": "M, S = D _ {m, s} \\left(\\operatorname {c o n c a t} \\left(\\mathcal {F} _ {r g b}, \\mathcal {F} _ {\\text {s p e c}}\\right)\\right) \\tag {2}", + "image_path": "4b10f6bbdb83f135e261e3a417b9aaf4731c2be0e8738a620ad85a3526c57a9c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 314, + 481, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 481, + 338 + ], + "type": "text", + "content": "With the shading component " + }, + { + "bbox": [ + 130, + 314, + 481, + 338 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 314, + 481, + 338 + ], + "type": "text", + "content": ", the reflectance component " + }, + { + "bbox": [ + 130, + 314, + 481, + 338 + ], + "type": "inline_equation", + "content": "R_{rgb}" + }, + { + "bbox": [ + 130, + 314, + 481, + 338 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 314, + 481, + 338 + ], + "type": "inline_equation", + "content": "R_{msi}" + }, + { + "bbox": [ + 130, + 314, + 481, + 338 + ], + "type": "text", + "content": " can subsequently be obtained according to the Retinex theory [3,22]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 357, + 236, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 236, + 367 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 236, + 367 + ], + "type": "text", + "content": "4.2 JDM-HDRNet" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 376, + 481, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 481, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 481, + 400 + ], + "type": "text", + "content": "Based on the established HDRNet [11], the JDM-HDRNet comprehensively exploits the " + }, + { + "bbox": [ + 130, + 376, + 481, + 400 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 376, + 481, + 400 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 376, + 481, + 400 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 376, + 481, + 400 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 376, + 481, + 400 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 376, + 481, + 400 + ], + "type": "text", + "content": " priors to provide explicit guidance for tone enhancement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 401, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 641 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 401, + 482, + 641 + ], + "type": "text", + "content": ": Localized Brightness Adaptation. To adjust both highlight and shadow areas during tone enhancement, the HDRNet's bilateral grid depth is set to 8, accommodating eight different luminance levels. However, this design may not effectively handle various localized brightness variations in high dynamic range scenes. We propose that separating the shading component would yield benefits in such contexts. As shown in Fig. 3(b), the pixel histogram vectors between the 16-bit input and 8-bit target exhibits a Pearson correlation coefficient " + }, + { + "bbox": [ + 130, + 401, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 401, + 482, + 641 + ], + "type": "text", + "content": " of 0.66 in the original RGB space, which increases to 0.91 in the reflectance space. A larger " + }, + { + "bbox": [ + 130, + 401, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 401, + 482, + 641 + ], + "type": "text", + "content": " indicates greater similarity in pixel histogram characteristics. Analyzing the statistics of the " + }, + { + "bbox": [ + 130, + 401, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 401, + 482, + 641 + ], + "type": "text", + "content": " distribution in the Mobile-Spec dataset, we observe that the reflectance space (Fig. 3(d)) has a greater proportion of samples with high " + }, + { + "bbox": [ + 130, + 401, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 401, + 482, + 641 + ], + "type": "text", + "content": " than the RGB space (Fig. 3(c)). This finding indicates that separating the shading component from the RGB space to the reflectance space may reduce the difficulty of color mapping learning. In our implementation, the shading component " + }, + { + "bbox": [ + 130, + 401, + 482, + 641 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 401, + 482, + 641 + ], + "type": "text", + "content": " is converted to the brightness representation " + }, + { + "bbox": [ + 130, + 401, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 130, + 401, + 482, + 641 + ], + "type": "text", + "content": " with a localized brightness adaptation module, comprising two layers of convolution and deconvolution. This lightweight module helps mitigate drastic pixel intensity changes in localized high dynamic range areas, thereby enhancing the adaptability. Subsequently, the reflectance images of 16-bit input " + }, + { + "bbox": [ + 130, + 401, + 482, + 641 + ], + "type": "inline_equation", + "content": "R_{rgb} = I_{rgb} / \\hat{S}" + }, + { + "bbox": [ + 130, + 401, + 482, + 641 + ], + "type": "text", + "content": " and Lr-MSI " + }, + { + "bbox": [ + 130, + 401, + 482, + 641 + ], + "type": "inline_equation", + "content": "R_{msi} = I_{msi} / \\hat{S}" + }, + { + "bbox": [ + 130, + 401, + 482, + 641 + ], + "type": "text", + "content": " are fed into the coefficient prediction part of the bilateral grid." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": ": Spectral Perception Self-Attention. After transforming from RGB space to reflectance space, we aim to leverage the reflectance of Lr-MSI " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "R_{msi}" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "K Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 205 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 205 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 205 + ], + "type": "image", + "image_path": "c4ecf87738825d0f68a113b3aa327ba099ebcc3aa9ab51e766b089fdd6ac391d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 212, + 482, + 246 + ], + "lines": [ + { + "bbox": [ + 131, + 212, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 131, + 212, + 482, + 246 + ], + "type": "text", + "content": "Fig. 4: The overall architecture of the proposed JDM-HDRNet. The " + }, + { + "bbox": [ + 131, + 212, + 482, + 246 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 131, + 212, + 482, + 246 + ], + "type": "text", + "content": " prior is employed to enhance localized dynamic ranges, while " + }, + { + "bbox": [ + 131, + 212, + 482, + 246 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 131, + 212, + 482, + 246 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 212, + 482, + 246 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 212, + 482, + 246 + ], + "type": "text", + "content": " priors contribute to the color mapping and grid expert learning, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": "which has fine-grained spectral channels to enhance the color mapping learning of bilateral grid. It should be noted that " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "R_{msi}" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": " has limited spatial resolution. We argue that downsampling in the spatial dimension will not cause significant degradation since bilateral grid coefficient prediction is carried out at a much lower spatial resolution. Hence, we design the Spectral Perception Self-Attention (SPSA) module which utilizes the " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "R_{msi}" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": " to guide " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "R_{rgb}" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": " for better bilateral coefficient prediction. " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "R_{msi}" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": " is resized to match the spatial resolution of " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "R_{rgb}" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": ", which is " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": " by default in HDRNet. The " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "R_{msi}" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "R_{rgb}" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": " are processed with a stack of strided convolutional layers to extract the low-level features and generate hierarchical feature maps. These feature maps " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{R_{msi}}" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{R_{rgb}}" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": " are concatenated along the channel dimension to form the cross-spectral feature embedding " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_R" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": ". Then " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": " convolutions " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "W_1" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": " depth-wise convolutions " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "W_3" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": " are applied to the " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathcal{R}}" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": " to generate query " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 270, + 482, + 426 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 192, + 438, + 480, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 438, + 480, + 454 + ], + "spans": [ + { + "bbox": [ + 192, + 438, + 480, + 454 + ], + "type": "interline_equation", + "content": "Q = W _ {1} ^ {Q} W _ {3} ^ {Q} \\mathcal {F} _ {\\mathcal {R}}, K = W _ {1} ^ {K} W _ {3} ^ {K} \\mathcal {F} _ {\\mathcal {R}}, V = W _ {1} ^ {V} W _ {3} ^ {V} \\mathcal {F} _ {\\mathcal {R}} \\tag {3}", + "image_path": "36e3912f873a6017855cd8a2ce1c82fb0a604420051fa9a0c17599bf3b5a00ba.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "text", + "content": "The query " + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "text", + "content": " and key " + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "text", + "content": " are reshaped into " + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\hat{Q} \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\hat{K} \\in \\mathbb{R}^{C \\times HW}" + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "text", + "content": " so that their dot-product interaction generates a spectral perception map " + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{C \\times C}" + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "text", + "content": ", which models the mutual information across the different spectral channels. The importance of different channels of the reshaped value " + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\hat{V} \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "text", + "content": " is reweighted by " + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "text", + "content": ", and the SPSA module can be formulated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 248, + 544, + 481, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 544, + 481, + 559 + ], + "spans": [ + { + "bbox": [ + 248, + 544, + 481, + 559 + ], + "type": "interline_equation", + "content": "A = \\operatorname {s o f t m a x} (\\sigma \\hat {K} \\cdot \\hat {Q}) \\tag {4}", + "image_path": "980fdb2f50f62ce107d4d6480aa0e9aa850b4bb429e15cfa2a170825220dedeb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 229, + 560, + 481, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 560, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 229, + 560, + 481, + 576 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {F}} _ {R _ {m s i}} = \\hat {W} _ {3} ^ {m s i} \\hat {V} \\cdot A + W _ {3} ^ {m s i} \\mathcal {F} _ {R _ {m s i}} \\tag {5}", + "image_path": "358147b9ef55c9e4af16a32d6fb1bb38bbafed2620ed56c2e6620b186e728ee9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 231, + 578, + 481, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 578, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 231, + 578, + 481, + 594 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {F}} _ {R _ {r g b}} = \\hat {W} _ {3} ^ {r g b} \\hat {V} \\cdot A + W _ {3} ^ {r g b} \\mathcal {F} _ {R _ {r g b}} \\tag {6}", + "image_path": "54c3601201d3ee94195190c04f47365aa7accc4ed7ed92d0d3c6e0590e5a69d4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": " is a learnable scaling parameter to control the magnitude of the dot product. The SPSA module serves as the residual learning which fuse the cross-spectral feature " + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{V} \\cdot A" + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": " with the original feature " + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{R_{msi}}" + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{R_{rgb}}" + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": " by the adaptive weighting of " + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": " convolution " + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{W}_3^{msi}" + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{W}_3^{rgb}" + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": ". The fused outputs " + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_{R_{msi}}" + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_{R_{rgb}}" + }, + { + "bbox": [ + 130, + 604, + 482, + 666 + ], + "type": "text", + "content": " of the SPSA module are fed into next layer in a progressive way." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 206, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 206, + 91, + 448, + 102 + ], + "type": "text", + "content": "A Joint Decomposition Model Guided Image Enhancement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": ": Mixture of Semantic Grid Experts. Common outdoor scenes can be roughly categorized into several primary types (e.g., sky, building, plant). As illustrated in Fig. 2(b), the plant spectral curves typically feature a reflection peak at around " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "550~\\mathrm{nm}" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": ", whereas the spectral curve of sky primarily resides within the blue wavelength range. Given these distinct characteristics, a multi-expert design can be introduced to effectively adapt to each material category's color preference. We believe that incorporating the material semantic prior " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " can offer robust constraints and yield visually pleasing results for tone enhancement. Consequently, we introduce a mixture of semantic grid experts, wherein each category is individually associated with a unique bilateral grid coefficient. For a specific semantic grid expert, the binarized material segmentation of the " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": "-th category " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " is transformed into the probability map " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\varPsi_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " with a mapping function comprising two convolutional and deconvolutional layers. Subsequently, a pair of affine transformation parameters " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "(\\alpha, \\beta) \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " is learned based on the probability map " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\varPsi_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "(\\alpha, \\beta)" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " is then introduced to the " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": " branch to modulate features conditioned on the probability map as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 203, + 320, + 481, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 320, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 203, + 320, + 481, + 335 + ], + "type": "interline_equation", + "content": "\\hat {Q} = (1 + \\alpha_ {Q}) \\cdot Q + \\beta_ {Q}, \\hat {K} = (1 + \\alpha_ {K}) \\cdot K + \\beta_ {K} \\tag {7}", + "image_path": "a8531e3f0bf074796cb87d708712632524d30b9e32914d74a399092a1f80053c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 341, + 482, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 377 + ], + "type": "text", + "content": "In each semantic grid expert " + }, + { + "bbox": [ + 130, + 341, + 482, + 377 + ], + "type": "inline_equation", + "content": "\\varPhi_{i}(x,y)" + }, + { + "bbox": [ + 130, + 341, + 482, + 377 + ], + "type": "text", + "content": ", the material category " + }, + { + "bbox": [ + 130, + 341, + 482, + 377 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 130, + 341, + 482, + 377 + ], + "type": "text", + "content": " is introduced in the same way. The bilateral grids of different experts are dynamically fused based on the category-dependent weight coefficient denoted as " + }, + { + "bbox": [ + 130, + 341, + 482, + 377 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 130, + 341, + 482, + 377 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 255, + 388, + 481, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 388, + 481, + 420 + ], + "spans": [ + { + "bbox": [ + 255, + 388, + 481, + 420 + ], + "type": "interline_equation", + "content": "\\Phi (x, y) = \\sum_ {i = 1} ^ {N} w _ {i} \\Phi_ {i} (x, y) \\tag {8}", + "image_path": "7768fcbaf4fc66f76f2bdb6612f8f5c55d1e5acdb36049f5851467d4d0f674f1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": " is the number of material categories, and " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\varPhi(x,y)" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": " denotes the final multi-expert bilateral grid. The full-resolution output of reflectance image " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "O_R" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": " is obtained through interpolated affine transformations, which is operated based on the bilateral grid " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\varPhi(x,y)" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": " and the guidance map " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": " is learned from the " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "R_{rgb}" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": " using the original guidance map auxiliary network. With the multiplication of " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "O_R" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 130, + 426, + 482, + 499 + ], + "type": "text", + "content": ", we get the final 8-bit output of JDM-HDRNet." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 520, + 230, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 230, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 230, + 533 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 548, + 279, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 279, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 279, + 560 + ], + "type": "text", + "content": "5.1 Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "In the Mobile-Spec dataset, " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " of the samples are allocated for the training set, while the remaining " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " are designated for the test set. The JDM-HDRNet is trained with a batch size of 4 and a learning rate of 0.0001. Specifically, the joint decomposition model is first trained with cross-entropy loss for both shading and segmentation predictions, then the JDM-HDRNet is trained with MSE loss in a supervised manner. During the training, input images are cropped to the size of " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ", and these images undergo further downsampling to " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " in the low-resolution stream." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "K Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 150, + 156, + 448, + 177 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 1: Performance evaluation of HDRNet with shading " + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "inline_equation", + "content": "S^{*}" + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": " and reflectance prior " + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "inline_equation", + "content": "R^{*}" + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": ". Ablation studies are conducted on the spectral channel and spatial resolution." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 150, + 146, + 463, + 156 + ], + "lines": [ + { + "bbox": [ + 150, + 146, + 463, + 156 + ], + "spans": [ + { + "bbox": [ + 150, + 146, + 463, + 156 + ], + "type": "text", + "content": "(a) Performance evaluation of HDRNet and HDRNet with shading prior (HDRNet " + }, + { + "bbox": [ + 150, + 146, + 463, + 156 + ], + "type": "inline_equation", + "content": "+S^{*}" + }, + { + "bbox": [ + 150, + 146, + 463, + 156 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 150, + 156, + 448, + 177 + ], + "lines": [ + { + "bbox": [ + 150, + 156, + 448, + 177 + ], + "spans": [ + { + "bbox": [ + 150, + 156, + 448, + 177 + ], + "type": "table", + "html": "
PSNR↑SSIM↑△E*↓PSNR↑SSIM↑△E*↓
HDRNet27.750.9395.12HDRNet+S*28.680.9574.29
", + "image_path": "b796a268c7ba9e5734b8c2ee3662a49a5837054fff35a742eac93241560b70c9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 154, + 196, + 301, + 278 + ], + "blocks": [ + { + "bbox": [ + 148, + 178, + 465, + 196 + ], + "lines": [ + { + "bbox": [ + 148, + 178, + 465, + 196 + ], + "spans": [ + { + "bbox": [ + 148, + 178, + 465, + 196 + ], + "type": "text", + "content": "(b) Ablation study of reflectance prior " + }, + { + "bbox": [ + 148, + 178, + 465, + 196 + ], + "type": "inline_equation", + "content": "R^{*}" + }, + { + "bbox": [ + 148, + 178, + 465, + 196 + ], + "type": "text", + "content": " on (c) Ablation study of reflectance prior " + }, + { + "bbox": [ + 148, + 178, + 465, + 196 + ], + "type": "inline_equation", + "content": "R^{*}" + }, + { + "bbox": [ + 148, + 178, + 465, + 196 + ], + "type": "text", + "content": " on channels with different spectral ranges. the spatial resolution, which is ranging from" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 154, + 196, + 301, + 278 + ], + "lines": [ + { + "bbox": [ + 154, + 196, + 301, + 278 + ], + "spans": [ + { + "bbox": [ + 154, + 196, + 301, + 278 + ], + "type": "table", + "html": "
PSNR↑SSIM↑△E*↓
baseline28.680.9574.29
400-520 nm28.740.9574.00
520-640 nm29.090.9643.82
640-760 nm29.190.9623.92
400-760 nm29.240.9673.72
760-1000 nm29.070.9653.88
400-1000 nm29.680.9683.55
", + "image_path": "35a1db4d39efdae9ecab3ee86436945c8e0b3af82866a1351f09875a2793a639.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 313, + 204, + 459, + 277 + ], + "blocks": [ + { + "bbox": [ + 308, + 180, + 465, + 203 + ], + "lines": [ + { + "bbox": [ + 308, + 180, + 465, + 203 + ], + "spans": [ + { + "bbox": [ + 308, + 180, + 465, + 203 + ], + "type": "text", + "content": "(c) Ablation study of reflectance prior " + }, + { + "bbox": [ + 308, + 180, + 465, + 203 + ], + "type": "inline_equation", + "content": "R^{*}" + }, + { + "bbox": [ + 308, + 180, + 465, + 203 + ], + "type": "text", + "content": " on the spatial resolution, which is ranging from " + }, + { + "bbox": [ + 308, + 180, + 465, + 203 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 308, + 180, + 465, + 203 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 308, + 180, + 465, + 203 + ], + "type": "inline_equation", + "content": "256\\times 256" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 204, + 459, + 277 + ], + "lines": [ + { + "bbox": [ + 313, + 204, + 459, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 459, + 277 + ], + "type": "table", + "html": "
PSNR↑SSIM↑△E*↓
baseline28.680.9574.29
1 × 129.050.9634.06
4 × 429.210.9633.80
16 × 1629.680.9683.55
64 × 6429.810.9693.47
256 × 25629.780.9673.49
", + "image_path": "7724f7be9e0bea9ad3566d5dfeafc0edb8d5f88831d06091b8e1d151c3dc8130.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 143, + 335, + 323, + 368 + ], + "blocks": [ + { + "bbox": [ + 143, + 289, + 469, + 300 + ], + "lines": [ + { + "bbox": [ + 143, + 289, + 469, + 300 + ], + "spans": [ + { + "bbox": [ + 143, + 289, + 469, + 300 + ], + "type": "text", + "content": "Table 2: Performance evaluation of HDRNet with material semantic prior " + }, + { + "bbox": [ + 143, + 289, + 469, + 300 + ], + "type": "inline_equation", + "content": "M^{*}" + }, + { + "bbox": [ + 143, + 289, + 469, + 300 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 309, + 472, + 335 + ], + "lines": [ + { + "bbox": [ + 138, + 309, + 472, + 335 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 472, + 335 + ], + "type": "text", + "content": "(a) The effectiveness of incorporating extra Lr-MSI in (b) Ablation study of " + }, + { + "bbox": [ + 138, + 309, + 472, + 335 + ], + "type": "inline_equation", + "content": "M^{*}" + }, + { + "bbox": [ + 138, + 309, + 472, + 335 + ], + "type": "text", + "content": " on the numthe joint RGB-Spectral decomposition model for the ber of material categories. material segmentation task." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 335, + 323, + 368 + ], + "lines": [ + { + "bbox": [ + 143, + 335, + 323, + 368 + ], + "spans": [ + { + "bbox": [ + 143, + 335, + 323, + 368 + ], + "type": "table", + "html": "
buildingplantskytrunkroadmIoU
RGB81.9987.6394.8410.9683.8771.86
+LrMSI87.5391.0195.8731.1489.1178.93
", + "image_path": "89326dea4468eadad816b5111fc1dd00d713da0339729136718dabe0b65ad50a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 339, + 327, + 465, + 369 + ], + "blocks": [ + { + "bbox": [ + 339, + 327, + 465, + 369 + ], + "lines": [ + { + "bbox": [ + 339, + 327, + 465, + 369 + ], + "spans": [ + { + "bbox": [ + 339, + 327, + 465, + 369 + ], + "type": "table", + "html": "
1246
PSNR↑27.7527.9828.4729.11
SSIM↑0.9390.9440.9540.964
△E*↓5.125.354.694.28
", + "image_path": "7b41ff8eb437a82d884fe25f67871cb9cb2e394723d590d7918b349281bdde59.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 390, + 482, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 451 + ], + "type": "text", + "content": "Three metrics are employed to validate the effectiveness of our method, including the PSNR, SSIM [35], and " + }, + { + "bbox": [ + 130, + 390, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\triangle E^{*}" + }, + { + "bbox": [ + 130, + 390, + 482, + 451 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 390, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\triangle E^{*}" + }, + { + "bbox": [ + 130, + 390, + 482, + 451 + ], + "type": "text", + "content": " measures the perceptual difference between two colors in the CIELAB space, a smaller " + }, + { + "bbox": [ + 130, + 390, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\triangle E^{*}" + }, + { + "bbox": [ + 130, + 390, + 482, + 451 + ], + "type": "text", + "content": " value indicates better color accuracy or image quality. For the evaluation of the material segmentation, we adopt the standard metric as mean Intersection-over-Union (mIoU)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 467, + 334, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 467, + 334, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 467, + 334, + 479 + ], + "type": "text", + "content": "5.2 Effectiveness of " + }, + { + "bbox": [ + 131, + 467, + 334, + 479 + ], + "type": "inline_equation", + "content": "S^{*}" + }, + { + "bbox": [ + 131, + 467, + 334, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 467, + 334, + 479 + ], + "type": "inline_equation", + "content": "R^{*}" + }, + { + "bbox": [ + 131, + 467, + 334, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 467, + 334, + 479 + ], + "type": "inline_equation", + "content": "M^{*}" + }, + { + "bbox": [ + 131, + 467, + 334, + 479 + ], + "type": "text", + "content": " Priors" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 486, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 480, + 521 + ], + "type": "text", + "content": "To avoid the impact of inaccurate decomposition, we begin by systematically integrating the ideal values of these priors (denoted as " + }, + { + "bbox": [ + 130, + 486, + 480, + 521 + ], + "type": "inline_equation", + "content": "S^{*}, R^{*}, M^{*}" + }, + { + "bbox": [ + 130, + 486, + 480, + 521 + ], + "type": "text", + "content": ") individually into the HDRNet baseline for validation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": "Shading prior " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "S^{*}" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": ". Tab. 1(a) illustrates that the HDRNet baseline attains a PSNR of 27.75 dB, while the separation of the shading component (" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "S^{*}" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": ") elevates the PSNR from 27.75 dB to 28.68 dB. Notably, the architecture of HDRNet remains unchanged in this sub-experiment. This finding demonstrates that transforming from RGB into reflectance space of input images with " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "S^{*}" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": " is a simple yet effective design for the tone enhancement task, which exhibits enhanced adaptability in diverse high dynamic range scenes." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Reflectance prior " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "R^{*}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ". Given the practical limitations of spectral imaging capability on mobile devices, we conduct ablation studies to examine the spectral and spatial resolution configurations of Lr-MSI. We first maintain the spatial resolution at the default setting of " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " and extract channels of different spectral ranges. The Lr-MSI has ten spectral channels, ranging from " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "400 \\mathrm{~nm}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " to" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 206, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 206, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Joint Decomposition Model Guided Image Enhancement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "inline_equation", + "content": "1000\\mathrm{nm}" + }, + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "text", + "content": ", with each channel featuring a " + }, + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "inline_equation", + "content": "60~\\mathrm{nm}" + }, + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "text", + "content": " bandwidth. Notably, " + }, + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "inline_equation", + "content": "400 - 760\\mathrm{nm}" + }, + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "text", + "content": " encompasses six visible wavelength channels, while " + }, + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "inline_equation", + "content": "760 - 1000\\mathrm{nm}" + }, + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "text", + "content": " includes four near-infrared channels. As observed in Tab. 1(b), the closer to the wavelength of red channel, the higher PSNR it will achieve. The possible explanation lies in the composition of our Mobile-Spec dataset, where a larger proportion of the area is covered by the sky and plant, while the presence of red objects is comparatively limited. Consequently, the HDRNet baseline tends to prioritize the blue and green wavelengths over the red wavelength during the training. Furthermore, the combination of all ten channels " + }, + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "inline_equation", + "content": "(400 - 1000\\mathrm{nm})" + }, + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "text", + "content": " achieves the best results of " + }, + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "inline_equation", + "content": "29.68\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "text", + "content": ". In our view, fine-grained spectral information enhances bilateral grids with heightened color perception capabilities and compensates for the imbalanced learning of different color channels. Then we maintain the number of channels at ten, and the spatial resolution is resized from " + }, + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "text", + "content": ". As depicted in Tab. 1(c), there is a trend of improved performance in evaluation metrics with higher spatial resolution. Considering the inherent constraint of spatial resolution in Lr-MSI on mobile devices, we contend that the optimal parameter configuration for Lr-MSI may align to the spatial resolution of the bilateral grid " + }, + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "inline_equation", + "content": "(16\\times 16)" + }, + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "text", + "content": ", which is the 8x downsampling corresponding to the spatial resolution of low-resolution stream input " + }, + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "inline_equation", + "content": "(256\\times 256)" + }, + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "text", + "content": ". Our experimental results suggest that the " + }, + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "inline_equation", + "content": "16\\times 16" + }, + { + "bbox": [ + 130, + 115, + 482, + 368 + ], + "type": "text", + "content": " configuration yields satisfactory results, beyond which increasing the spatial resolution leads to diminishing marginal utility." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 368, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 548 + ], + "type": "text", + "content": "Material semantic prior " + }, + { + "bbox": [ + 130, + 368, + 482, + 548 + ], + "type": "inline_equation", + "content": "M^{*}" + }, + { + "bbox": [ + 130, + 368, + 482, + 548 + ], + "type": "text", + "content": ". Tab. 2(a) demonstrates that incorporating the Lr-MSI branch into the joint RGB-Spectral decomposition model enhances the mIoU of material segmentation from " + }, + { + "bbox": [ + 130, + 368, + 482, + 548 + ], + "type": "inline_equation", + "content": "71.86\\%" + }, + { + "bbox": [ + 130, + 368, + 482, + 548 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 368, + 482, + 548 + ], + "type": "inline_equation", + "content": "78.93\\%" + }, + { + "bbox": [ + 130, + 368, + 482, + 548 + ], + "type": "text", + "content": ". Despite Lr-MSI's notably lower spatial resolution compared to the 16-bit, it exhibits superior material identification capability owing to its fine-grained spectral channels. Subsequently, we conduct ablation experiments to evaluate the number of material categories. The category number of the original HDRNet baseline is set to one. In the two-category set, the sky, plant, and trunk are grouped into one category, while building, road, and others form another. For the four-category set, the plant and trunk, as well as building and road, are each combined into single categories, whereas the sky and others are treated as separate, independent categories. By refining the division of material categories, the mixture of semantic grid experts becomes more specialized in enhancing tones for specific materials. Tab. 2(b) demonstrates improved quantitative results across three evaluation metrics with the inclusion of six semantic grid experts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 569, + 330, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 330, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 330, + 582 + ], + "type": "text", + "content": "5.3 Ablation Study of JDM-HDRNet" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Tab. 3(a) presents the ablation study incorporating ideal values " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "S^{*}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "R^{*}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "M^{*}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " into HDRNet (denoted as JDM-HDRNet\\*). The collaboration of shading " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and reflectance " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " priors leads to a substantial decrease in " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\triangle E^{*}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " from 5.12 to 3.55 in the CIELAB color space. Moreover, the design of the mixture of semantic grid experts enables specific experts to focus on individual material categories, resulting in a PSNR of 30.14 dB. The above experiments are conducted with ideal priors." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "K Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 145, + 152, + 462, + 205 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 477, + 125 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 477, + 125 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 477, + 125 + ], + "type": "text", + "content": "Table 3: Ablation study of the proposed JDM-HDRNet on the " + }, + { + "bbox": [ + 135, + 114, + 477, + 125 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 135, + 114, + 477, + 125 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 135, + 114, + 477, + 125 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 135, + 114, + 477, + 125 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 135, + 114, + 477, + 125 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 135, + 114, + 477, + 125 + ], + "type": "text", + "content": " priors." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 139, + 135, + 470, + 152 + ], + "lines": [ + { + "bbox": [ + 139, + 135, + 470, + 152 + ], + "spans": [ + { + "bbox": [ + 139, + 135, + 470, + 152 + ], + "type": "text", + "content": "(a) Ablation study on the ideal values of " + }, + { + "bbox": [ + 139, + 135, + 470, + 152 + ], + "type": "inline_equation", + "content": "S^*" + }, + { + "bbox": [ + 139, + 135, + 470, + 152 + ], + "type": "text", + "content": ", (b) Ablation study on " + }, + { + "bbox": [ + 139, + 135, + 470, + 152 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 139, + 135, + 470, + 152 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 139, + 135, + 470, + 152 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 139, + 135, + 470, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 139, + 135, + 470, + 152 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 139, + 135, + 470, + 152 + ], + "type": "text", + "content": " priors predicted from the joint decomposition model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 145, + 152, + 462, + 205 + ], + "lines": [ + { + "bbox": [ + 145, + 152, + 462, + 205 + ], + "spans": [ + { + "bbox": [ + 145, + 152, + 462, + 205 + ], + "type": "table", + "html": "
S*R* M*PSNR↑SSIM↑ΔE*↓
HDRNet27.750.9395.12HDRNet27.750.939
JDM-HDRNet*✓28.680.9574.29JDM-HDRNet✓28.590.953
✓29.680.9683.55✓29.570.966
✓30.140.9723.44✓29.830.967
", + "image_path": "24081b16f55a82d4ea078bfef3472ac581e776cd7040b41c0ce8a4bf57b82cef.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 133, + 225, + 482, + 269 + ], + "blocks": [ + { + "bbox": [ + 138, + 214, + 474, + 224 + ], + "lines": [ + { + "bbox": [ + 138, + 214, + 474, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 214, + 474, + 224 + ], + "type": "text", + "content": "Table 4: Comparisons of our JDM-HDRNet with previous enhancement methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 225, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 133, + 225, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 133, + 225, + 482, + 269 + ], + "type": "table", + "html": "
DPECSRNet3D LUTCLUTSepLUT-SSepLUT-L4D LUTHDRNetUPEOurs
PSNR↑22.8126.3427.5227.3027.5728.0827.7727.7528.1929.83
SSIM↑0.8060.9230.9260.9380.9330.9440.9350.9390.9460.967
△E*↓11.066.445.394.635.374.264.325.124.793.60
", + "image_path": "1d104033313de3e9c84d5a507236b29f387d3e258d40418506ad83ad97702be8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 461 + ], + "type": "text", + "content": "Subsequently, the predictions for " + }, + { + "bbox": [ + 130, + 293, + 482, + 461 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 293, + 482, + 461 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 293, + 482, + 461 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 293, + 482, + 461 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 293, + 482, + 461 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 293, + 482, + 461 + ], + "type": "text", + "content": " from joint RGB-Spectral decomposition model are integrated with HDRNet (denoted as JDM-HDRNet). Tab. 3(b) shows that the JDM-HDRNet exhibits comparable performance across three metrics in the " + }, + { + "bbox": [ + 130, + 293, + 482, + 461 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 293, + 482, + 461 + ], + "type": "text", + "content": " set, indicating that the localized brightness adaptation module demonstrates adaptive adjustment capabilities in response to shading image degradation. In addition, we observe that the performance of the reflectance " + }, + { + "bbox": [ + 130, + 293, + 482, + 461 + ], + "type": "inline_equation", + "content": "(R)" + }, + { + "bbox": [ + 130, + 293, + 482, + 461 + ], + "type": "text", + "content": " set remains relatively unaffected, achieving an improvement of 0.98 dB. Because both the Lr-MSI and 16-bit RGB images have undergone consistent transformation into the reflectance space using the same " + }, + { + "bbox": [ + 130, + 293, + 482, + 461 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 293, + 482, + 461 + ], + "type": "text", + "content": " from the joint decomposition model. Regarding the material semantic prior (M), JDM-HDRNet dynamically fuses different grid experts at the bilateral grid level to avoid the inaccurate segmentation of " + }, + { + "bbox": [ + 130, + 293, + 482, + 461 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 293, + 482, + 461 + ], + "type": "text", + "content": ". Overall, JDM-HDRNet achieves a gain of 2.08 dB compared to the HDRNet baseline, which demonstrates that the low resolution spectral information can be effectively leveraged by the joint decomposition model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 478, + 348, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 478, + 348, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 478, + 348, + 491 + ], + "type": "text", + "content": "5.4 Comparisons with Previous Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Quantitative Comparisons. We compare JDM-HDRNet with previous color transform-based methods for tone enhancement. In Tab. 4, DPE [6] exhibits inferior performance compared to other methods, attributed to its unpaired setting, leading to output images affected by undesired artifacts introduced by the generative adversarial network. Among the color transform functions, the MLPs based CSRNet [14] exhibits a slightly lower PSNR of 26.34 dB. In contrast, the 3D LUTs based methods, such as 3D LUT [39], CLUT [41], SepLUT-S [36], and 4D LUT [24], demonstrate the similar PSNR ranging from 27.30 dB to 27.77 dB. Notably, SepLUT-L [36] with larger parameters exhibits a superior performance because it simultaneously takes advantage of both 1D and 3D LUTs for enhancement, the 1D LUT adjusts image contrast to achieve a more uniform distribution in an image-adaptive manner. Additionally, UPE [34] estimates an image-to-illumination mapping with constraints of smoothness loss, which is exclusively designed to enhance images under diverse lighting conditions, and it" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Joint Decomposition Model Guided Image Enhancement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 114, + 473, + 223 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 473, + 223 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 473, + 223 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 473, + 223 + ], + "type": "image", + "image_path": "86ef7e3909dbfdd8c134b3bc31ffb95c7233d26e487c8159b17100487647efd9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 230, + 482, + 253 + ], + "lines": [ + { + "bbox": [ + 131, + 230, + 482, + 253 + ], + "spans": [ + { + "bbox": [ + 131, + 230, + 482, + 253 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparisons on the Mobile-Spec dataset. The error maps displayed at the top-right corner illustrate the differences with ground truth for each image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 276, + 479, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 479, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 479, + 335 + ], + "type": "text", + "content": "achieves good performance in handling high dynamic range scenes in the Mobile-Spec dataset. Previous methods face challenges in learning accurate tone enhancement for HDR scenes, while our priors-guided JDM-HDRNet outperforms previous methods by effectively leveraging " + }, + { + "bbox": [ + 130, + 276, + 479, + 335 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 276, + 479, + 335 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 276, + 479, + 335 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 276, + 479, + 335 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 276, + 479, + 335 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 276, + 479, + 335 + ], + "type": "text", + "content": " priors predicted from the joint RGB-Spectral decomposition model to provide explicit guidance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 336, + 480, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 480, + 430 + ], + "type": "text", + "content": "Qualitative Comparisons. Fig. 5 presents qualitative comparisons of two representative examples. In the first row, JDM-HDRNet exhibits enhanced contrast and a wider dynamic range in the plant area, while other methods appear comparatively dim. In the second row, JDM-HDRNet shows reduced color deviation with the ground truth in the wall area, indicating the beneficial role of spectral information integration in improving photographic realism. In summary, JDM-HDRNet, guided by " + }, + { + "bbox": [ + 130, + 336, + 480, + 430 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 336, + 480, + 430 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 336, + 480, + 430 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 336, + 480, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 336, + 480, + 430 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 336, + 480, + 430 + ], + "type": "text", + "content": " priors, achieves more accurate and pleasing colors for high dynamic range scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 431, + 481, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 481, + 491 + ], + "type": "text", + "content": "Limitations. The spatial resolution of Lr-MSI in this study is fixed at " + }, + { + "bbox": [ + 130, + 431, + 481, + 491 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 130, + 431, + 481, + 491 + ], + "type": "text", + "content": ", a setting not yet achievable on commercial smartphones. Increasing the spatial resolution would improve the accuracy of the joint decomposition model. Further exploration is required to address the constrained spectral imaging capabilities in the context of image enhancement and related applications on mobile devices." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 509, + 220, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 220, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 220, + 521 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "We have investigated the effectiveness of Lr-MSI in the mobile ISP pipeline for enhancing tone enhancement. We first construct the high-quality Mobile-Spec, a pioneering dataset comprising both aligned spectral and RGB images. Then we analyze the role of Lr-MSI from the perspective of spectral image intrinsic decomposition. Shading, reflectance and material semantic priors are derived using the joint decomposition model to address the inherent complexity of spectral images and the constraints of spectral imaging capabilities. Subsequently, JDM-HDRNet is meticulously designed under the explicit guidance of above priors, which exhibits superior performance than the competitive methods. Our dataset and method are expected to attract further research into this area and explore the untapped potential of spectral information on mobile devices." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "K Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 221 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 221 + ], + "type": "text", + "content": "This research was supported by National Key Research and Development Program of China (2023YFF0713300), National Natural Science Foundation of China under Grant (62071216), Leading Technology of Jiangsu Basic Research Plan under Grant (BK20192003), and National Research Foundation Singapore Competitive Research Program (award number CRP29-2022-0003). We thank the device support from NJU-ZPTECH University-Enterprise Joint Laboratory for Computational Spectral Imaging." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 236, + 197, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 197, + 249 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 197, + 249 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 258, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 258, + 480, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 258, + 480, + 279 + ], + "spans": [ + { + "bbox": [ + 138, + 258, + 480, + 279 + ], + "type": "text", + "content": "1. Gaiasky-mini hyperspectral imaging camera. https://www.dualix.com.cn/en/Goods/desc/id/123/aid/954.html" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 280, + 425, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 280, + 425, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 280, + 425, + 289 + ], + "type": "text", + "content": "2. Spectral Sensing -ams.com. https://ams.com/spectral-sensing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 290, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 290, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 290, + 480, + 311 + ], + "type": "text", + "content": "3. Barrow, H., Tenenbaum, J., Hanson, A., Riseman, E.: Recovering intrinsic scene characteristics. Comput. vis. syst 2(3-26), 2 (1978)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 312, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 480, + 342 + ], + "type": "text", + "content": "4. Cai, J., Gu, S., Zhang, L.: Learning a deep single image contrast enhancer from multi-exposure images. IEEE Transactions on Image Processing 27(4), 2049-2062 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 343, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 374 + ], + "type": "text", + "content": "5. Chen, X., Zhu, W., Zhao, Y., Yu, Y., Zhou, Y., Yue, T., Du, S., Cao, X.: Intrinsic decomposition from a single spectral image. Applied optics 56(20), 5676-5684 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 376, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 480, + 418 + ], + "type": "text", + "content": "6. Chen, Y.S., Wang, Y.C., Kao, M.H., Chuang, Y.Y.: Deep photo enhancer: Unpaired learning for image enhancement from photographs with gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6306-6314 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 419, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 419, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 138, + 419, + 480, + 450 + ], + "type": "text", + "content": "7. Cheng, Z., Zheng, Y., You, S., Sato, I.: Non-local intrinsic decomposition with near-infrared priors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2521-2530 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 451, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 451, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 138, + 451, + 480, + 472 + ], + "type": "text", + "content": "8. Erba, I., et al.: Computational color constancy beyond rgb images: Multispectral and temporal extensions (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 472, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 480, + 493 + ], + "type": "text", + "content": "9. Finlayson, G.D., Zhu, Y.: Designing color filters that make cameras more colorimetric. IEEE Transactions on Image Processing 30, 853-867 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 494, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 494, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 134, + 494, + 480, + 525 + ], + "type": "text", + "content": "10. Fu, X., Ying, Y.: Food safety evaluation based on near infrared spectroscopy and imaging: A review. Critical Reviews in Food Science and Nutrition 56(11), 1913-1924 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 525, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 525, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 134, + 525, + 480, + 557 + ], + "type": "text", + "content": "11. Gharbi, M., Chen, J., Barron, J.T., Hasinoff, S.W., Durand, F.: Deep bilateral learning for real-time image enhancement. ACM Transactions on Graphics (TOG) 36(4), 1-12 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 558, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 558, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 134, + 558, + 480, + 578 + ], + "type": "text", + "content": "12. Gijsenij, A., Gevers, T., Van De Weijer, J.: Computational color constancy: Survey and experiments. IEEE transactions on image processing 20(9), 2475-2489 (2011)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 579, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 579, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 579, + 480, + 621 + ], + "type": "text", + "content": "13. Glatt, O., Ater, Y., Kim, W.S., Werman, S., Berby, O., Zini, Y., Zelinger, S., Lee, S., Choi, H., Soloveichik, E.: Beyond rgb: A real world dataset for multispectral imaging in mobile devices. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4344-4354 (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "type": "text", + "content": "14. He, J., Liu, Y., Qiao, Y., Dong, C.: Conditional sequential modulation for efficient global image retouching. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16. pp. 679-695. Springer (2020)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Joint Decomposition Model Guided Image Enhancement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "15. He, Q., Wang, R.: Hyperspectral imaging enabled by an unmodified smartphone for analyzing skin morphological features and monitoring hemodynamics. Biomedical optics express 11(2), 895-910 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "text", + "content": "16. He, Q., Wang, R.K.: Analysis of skin morphological features and real-time monitoring using snapshot hyperspectral imaging. Biomedical optics express 10(11), 5625-5638 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "type": "text", + "content": "17. Huang, Q., Zhu, W., Zhao, Y., Chen, L., Wang, Y., Yue, T., Cao, X.: Multispectral image intrinsic decomposition via subspace constraint. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6430-6439 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 480, + 247 + ], + "type": "text", + "content": "18. Jiang, J., Liu, D., Gu, J., Süssstrunk, S.: What is the space of spectral sensitivity functions for digital color cameras? In: 2013 IEEE Workshop on Applications of Computer Vision (WACV). pp. 168-179. IEEE (2013)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 247, + 480, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 247, + 480, + 290 + ], + "spans": [ + { + "bbox": [ + 133, + 247, + 480, + 290 + ], + "type": "text", + "content": "19. Karaimer, H.C., Brown, M.S.: A software platform for manipulating the camera imaging pipeline. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 429-444. Springer (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 290, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 290, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 133, + 290, + 480, + 322 + ], + "type": "text", + "content": "20. Kim, S.J., Lin, H.T., Lu, Z., Susstrunk, S., Lin, S., Brown, M.S.: A new in-camera imaging model for color computer vision and its application. IEEE Transactions on Pattern Analysis and Machine Intelligence 34(12), 2289-2302 (2012)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 322, + 480, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 322, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 133, + 322, + 480, + 343 + ], + "type": "text", + "content": "21. Kim, Y.T.: Contrast enhancement using brightness preserving bi-histogram equalization. IEEE transactions on Consumer Electronics 43(1), 1-8 (1997)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 343, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 343, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 133, + 343, + 480, + 354 + ], + "type": "text", + "content": "22. Land, E.H., McCann, J.J.: Lightness and retinex theory. Josa 61(1), 1-11 (1971)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 354, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 354, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 354, + 480, + 407 + ], + "type": "text", + "content": "23. Lindholm, V., Raita-Hakola, A.M., Annala, L., Salmivuori, M., Jeskanen, L., Saari, H., Koskenmies, S., Pitkänen, S., Pölönen, I., Isoherranen, K., et al.: Differentiating malignant from benign pigmented or non-pigmented skin tumours—a pilot study on 3d hyperspectral imaging of complex skin surfaces and convolutional neural networks. Journal of clinical medicine 11(7), 1914 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 407, + 480, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 407, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 133, + 407, + 480, + 440 + ], + "type": "text", + "content": "24. Liu, C., Yang, H., Fu, J., Qian, X.: 4d lut: learnable context-aware 4d lookup table for image enhancement. IEEE Transactions on Image Processing 32, 4742-4756 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 440, + 480, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 440, + 480, + 473 + ], + "spans": [ + { + "bbox": [ + 133, + 440, + 480, + 473 + ], + "type": "text", + "content": "25. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 472, + 480, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 472, + 480, + 494 + ], + "spans": [ + { + "bbox": [ + 133, + 472, + 480, + 494 + ], + "type": "text", + "content": "26. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International journal of computer vision 60, 91-110 (2004)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 494, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 494, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 133, + 494, + 480, + 515 + ], + "type": "text", + "content": "27. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 515, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 515, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 133, + 515, + 480, + 536 + ], + "type": "text", + "content": "28. Mantiuk, R., Daly, S., Kerofsky, L.: Display adaptive tone mapping. In: ACM SIGGRAPH 2008 papers, pp. 1-10 (2008)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 536, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 536, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 133, + 536, + 480, + 557 + ], + "type": "text", + "content": "29. McGonigle, A.J., Wilkes, T.C., Pering, T.D., Willmott, J.R., Cook, J.M., Mims III, F.M., Parisi, A.V.: Smartphone spectrometers. Sensors 18(1), 223 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 557, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 557, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 133, + 557, + 480, + 590 + ], + "type": "text", + "content": "30. Ng, P.C., Chi, Z., Verdie, Y., Lu, J., Plataniotis, K.N.: Hyper-skin: A hyperspectral dataset for reconstructing facial skin-spectra from rgb images. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 590, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 590, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 133, + 590, + 480, + 622 + ], + "type": "text", + "content": "31. Rao, S., Huang, Y., Cui, K., Li, Y.: Anti-spoofing face recognition using a metasurface-based snapshot hyperspectral image sensor. Optica 9(11), 1253-1259 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 622, + 480, + 665 + ], + "type": "text", + "content": "32. e Silva, G.M., Garcia, J.A., de Alencar Garitta, J., Cunha, D.G.F., Finkler, N.R., Mendiondo, E.M., Ghiglieno, F.: Smartphone-based spectrometry system as a prescreening assessment of copper and iron for real time control of water pollution. Journal of Environmental Management 323, 116214 (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "K Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 467 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "33. Thomas, J.B.: Illuminant estimation from uncalibrated multispectral images. In: 2015 Colour and Visual Computing Symposium (CVCS). pp. 1-6. IEEE (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 481, + 182 + ], + "type": "text", + "content": "34. Wang, R., Zhang, Q., Fu, C.W., Shen, X., Zheng, W.S., Jia, J.: Underexposed photo enhancement using deep illumination estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6849-6857 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "type": "text", + "content": "35. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "36. Yang, C., Jin, M., Xu, Y., Zhang, R., Chen, Y., Liu, H.: Seplut: Separable image-adaptive lookup tables for real-time image enhancement. In: European Conference on Computer Vision. pp. 201-217. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "37. Yang, Z., Albrow-Owen, T., Cai, W., Hasan, T.: Miniaturization of optical spectrometers. Science 371(6528), eabe0722 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 480, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 312 + ], + "type": "text", + "content": "38. Yuan, L., Sun, J.: Automatic exposure correction of consumer photographs. In: Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part IV 12. pp. 771-785. Springer (2012)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 314, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 480, + 346 + ], + "type": "text", + "content": "39. Zeng, H., Cai, J., Li, L., Cao, Z., Zhang, L.: Learning image-adaptive 3d lookup tables for high performance photo enhancement in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(4), 2058-2073 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "40. Zhang, F., You, S., Li, Y., Fu, Y.: Hsi-guided intrinsic image decomposition for outdoor scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 313-322 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 423 + ], + "type": "text", + "content": "41. Zhang, F., Zeng, H., Zhang, T., Zhang, L.: Clut-net: Learning adaptively compressed representations of 3dluts for lightweight image enhancement. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 6493-6501 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 480, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 467 + ], + "type": "text", + "content": "42. Zheng, Y., Sato, I., Sato, Y.: Illumination and reflectance spectra separation of a hyperspectral image meets low-rank matrix factorization. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1779-1787 (2015)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 91, + 447, + 102 + ], + "type": "text", + "content": "A Joint Decomposition Model Guided Image Enhancement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/JointDreamer_ Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation/4667ae70-c38c-469d-a697-9d92865631f2_content_list.json b/2024/JointDreamer_ Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation/4667ae70-c38c-469d-a697-9d92865631f2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d2b74db2bab1ada4f8c757aafd3d2cf1c8595493 --- /dev/null +++ b/2024/JointDreamer_ Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation/4667ae70-c38c-469d-a697-9d92865631f2_content_list.json @@ -0,0 +1,1770 @@ +[ + { + "type": "text", + "text": "JointDreamer: Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation", + "text_level": 1, + "bbox": [ + 228, + 140, + 774, + 207 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chenhan Jiang $^{1*†}$ , Yihan Zeng $^{2*}$ , Tianyang Hu $^{2}$ , Songcun Xu $^{2}$ , Wei Zhang $^{2}$ , Hang Xu $^{2}$ , and Dit-Yan Yeung $^{1}$", + "bbox": [ + 217, + 233, + 784, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The Hong Kong University of Science and Technology", + "bbox": [ + 310, + 276, + 687, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Huawei Noah's Ark Lab", + "bbox": [ + 413, + 292, + 589, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "*Equal contribution. †Corresponding Author: jchcyan@gmail.com", + "bbox": [ + 276, + 305, + 725, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://jointdreamer.github.io", + "bbox": [ + 383, + 320, + 617, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Score Distillation Sampling (SDS) by well-trained 2D diffusion models has shown great promise in text-to-3D generation. However, this paradigm distills view-agnostic 2D image distributions into the rendering distribution of 3D representation for each view independently, overlooking the coherence across views and yielding 3D inconsistency in generations. In this work, we propose Joint Score Distillation (JSD), a new paradigm that ensures coherent 3D generations. Specifically, we model the joint image distribution, which introduces an energy function to capture the coherence among denoised images from the diffusion model. We then derive the joint score distillation on multiple rendered views of the 3D representation, as opposed to a single view in SDS. In addition, we instantiate three universal view-aware models as energy functions, demonstrating compatibility with JSD. Empirically, JSD significantly mitigates the 3D inconsistency problem in SDS, while maintaining text congruence. Moreover, we introduce the Geometry Fading scheme and Classifier-Free Guidance (CFG) Switching strategy to enhance generative details. Our framework, JointDreamer, establishes a new benchmark in text-to-3D generation, achieving outstanding results with an $88.5\\%$ CLIP R-Precision and $27.7\\%$ CLIP Score. These metrics demonstrate exceptional text congruence, as well as remarkable geometric consistency and texture fidelity.", + "bbox": [ + 259, + 367, + 743, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D Vision $\\cdot$ 3D Generation $\\cdot$ Energy Function", + "bbox": [ + 261, + 672, + 656, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 733, + 375, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D content creation is essential for diverse applications, including gaming, robotics simulation, and virtual reality. However, it is labor-intensive, demanding substantial time for skilled designers to create a single 3D asset. Hence, automating 3D creation with text input has attracted considerable attention. Recently, the score distillation sampling (SDS) algorithm pioneered by DreamFusion [37]", + "bbox": [ + 212, + 763, + 792, + 842 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3f3a1bb770864999f37a8549ab0e64c86a81933d0589418a70a502ab4eda53ce.jpg", + "image_caption": [ + "Fig. 1: Text-to-3D generations by JointDreamer from scratch. JointDreamer excels in generating geometrically consistent and high-fidelity 3D assets, adhering to complex textual descriptions that are challenging for previous methods." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 774, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "shows promise in text-to-3D tasks, which lifts image distribution from a well-trained diffusion model [38] into parameterized 3D representation like NeRF [31]. Compared to 3D generative models [2, 19, 34] that struggle with producing arbitrary objects due to limited text-3D training data, SDS-based methods [6, 24, 37, 42] can generate arbitrary 3D assets with diverse text input.", + "bbox": [ + 212, + 638, + 787, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although SDS-based methods benefit from the generalizability of diffusion models, they often encounter a common issue known as Janus artifacts [6, 24, 37, 42]. These artifacts manifest as repeated content from different viewpoints of a 3D generation, yielding a lack of realism and coherence in the rendered views. We investigate the Janus artifacts by visualizing image distributions of the diffusion model [38] from multiple viewpoints, as illustrated in Fig. 2(a). The results reveal the view-agnostic nature and the content inconsistency across views of the diffusion model. Consequently, SDS optimizes each rendered view of 3D", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "C. Jiang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "representation independently and inherits image distribution without coherent multi-view perspective, leading to the geometric inconsistency of 3D generations.", + "bbox": [ + 214, + 146, + 784, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing works [1,22] address the aforementioned challenges within the SDS framework by employing prompt engineering techniques. However, the effectiveness of such methods remains unsatisfactory, as evident from the results depicted in Fig. 2(b). Alternative methods [16,41] propose to finetune view-aware diffusion models using rendered images of 3D datasets [7,8]. Nevertheless, they are prone to overfitting on limited text-3D training data [23], decreasing the text congruence when handling complex text inputs. Based on the above observations, it requires a rethinking of the SDS optimization to enhance the 3D coherence of generations while maintaining generalizability.", + "bbox": [ + 212, + 186, + 787, + 323 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we first present Joint Score Distillation (JSD), which significantly promotes the 3D consistency of generation and inherits generalizability from diffusion models. Specifically, we model the joint image distribution of diffusion model via an energy function measuring coherence across denoised images. It facilitates the extension of the KL-Divergence in SDS from single-view into multi-view. We then derive the joint score distillation function from multi-view KL-Divergence, which ensures inter-view coherence in the optimization process of 3D generation. We show that SDS is a special case of JSD with the energy term omitted, which indicates the absence of coherence constraint across views.", + "bbox": [ + 212, + 330, + 517, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building upon JSD optimization, we present three view-aware models as energy terms to showcase the compatibility of JSD: the Binary Classification Model, the Image-to-Image Translation Model, and the Multi-view Generation Model. Through empirical analysis, it is observed that different view-aware models introduce distinct coherence measurements,", + "bbox": [ + 212, + 613, + 517, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "leading to diverse 3D generations, while all contributing to 3D consistency. Furthermore, to facilitate a more comprehensive comparison with existing text-to-3D generation methods, we introduce JointDreamer, an innovative framework capable of producing geometric-consistent and high-fidelity 3D assets adhering to complex text descriptions. Notably, in addition to incorporating a Multi-view Generation model as an energy term in JSD, we introduce two complementary", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/60635b6e1dfbdb5e4b443c7cb208dff487674074281f21709a1c38f9b1835a20.jpg", + "image_caption": [ + "Fig.2: Illustration of text-conditioned images for different viewpoints, where input texts are augmented with corresponding direction prompts for each view. (a) The original generations from 2D diffusion model [38] are view-agnostic and inconsistent across views. (b) Text prompt tuning [1] has limited improvement in the directional structure of generated images for each view. (c) JSD injects coherence measurement from the proposed binary classifier (refer to Section 4.2), contributing to modified directional structures and semantical consistency across views." + ], + "image_footnote": [], + "bbox": [ + 531, + 335, + 781, + 498 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "JointDreamer", + "bbox": [ + 637, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "techniques, namely the Geometry Fading scheme and the Classifier-Free Guidance (CFG) Switching strategy, to enhance generative details.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We systematically assess the quality of our approach, both qualitatively and quantitatively, compared to existing methods. Qualitative results gallery can be found in Fig. 1. Our JointDreamer consistently produces high-fidelity 3D assets and mitigates Janus artifacts in SDS. It maintains text congruence even when confronted with complex text input, achieving $88.5\\%$ CLIP R-Precision and $27.7\\%$ CLIP Score.", + "bbox": [ + 212, + 176, + 784, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In brief, our contributions are summarized as follows:", + "bbox": [ + 238, + 268, + 624, + 281 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel Joint Score Distillation (JSD) for text-to-3D generation, optimizing multiple views jointly via an energy function to capture inter-view coherence.", + "- We present three view-aware models as energy functions to show compatibility with JSD, all of them mitigate the Janus problem in SDS.", + "- We introduce the text-to-3D framework JointDreamer, incorporating complementary Geometry Fading and CFG Switching techniques. Our JointDreamer achieves geometrically consistent and high-fidelity 3D assets even with complex textual inputs." + ], + "bbox": [ + 225, + 294, + 784, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 455, + 395, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text-to-3D Generation. Existing text-to-3D generation methods can be categorized into two streams: 3D generative models and 2D optimization methods. The former encompasses various deep generative models such as Variational Auto Encoders (VAEs) [12, 13], Generative Adversarial Models (GAN) [4, 9, 10, 33, 35], diffusion models [5, 28, 34] and transformer architectures [2, 19]. These models are efficient in inference but often struggle with generalizability and training stability, attributed to the limited scope and complexity of available 3D datasets. The latter approach centers around the Score Distillation Sampling (SDS) algorithm proposed by [37], which leverages 2D diffusion model priors [38] for optimizing 3D representations. Subsequent advancements have refined this technique by improving the 3D representations [6, 24], the sampling scheduler [17] and loss design [42]. However, the above approaches overlook the geometric consistency problem, facing inherent multi-face Janus issues in SDS. Prior works try to alleviate the Janus issues with prompt tuning [1] yet they achieve limited effect. Very recent work MVDream [41] addresses the problem by fine-tuning a multi-view diffusion model, but it is susceptible to overfitting on scarce 3D training data, compromising semantic consistency in text-to-3D generations. In this work, we address the fundamental flaw of SDS that optimizes each view independently by introducing a joint optimization function that enforces inter-view consistency, essentially solving the Janus issues in SDS while preserving its generalizability.", + "bbox": [ + 212, + 483, + 785, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion-based Novel View Synthesis. As an alternative to 3D generation, novel view synthesis models the challenge as a view-conditioned image-to-image translation task. There have been proposals for pose-conditioned image-to-image", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "C. Jiang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a09b64fcd74f2b86786d190bb558a9ecae1f0da3bd57f94f32b1fcb31dca274a.jpg", + "image_caption": [ + "Fig. 3: Overview of JointDreamer Framework. We introduce an energy function to model the joint distribution for multi-view denoised images from 2D diffusion model, facilitating the Joint Score Distillation (JSD) optimization for text-to-3D generation." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 781, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "diffusion models [43] that generate novel views on synthetic data in 3D. Recently, [26] promoted the generalizability of novel view synthesis by fine-tuning the 2D diffusion model [38] on renderings of 3D dataset, which facilitates images-to-3D tasks with 3D reconstruction or SDS algorithm. To further improve the 3D consistency across generated views and input view, very recent works [27, 29, 40] modify the generation process into multi-view generation and present corresponding architecture designs. Generally, these methods take camera specifications as conditions and enable the viewpoint-aware generations, but they can hardly accurately capture a complete 3D scene consistently and densely. Our method acknowledges the potential of these models to discern relative interview relationships, which we harness to provide inter-view coherence for our JSD, thus serving as universal guidance models.", + "bbox": [ + 212, + 335, + 787, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 537, + 380, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We review the original SDS and address its fundamental limitation: optimizing independently for each single view of 3D representation and distilling with the view-agnostic image distribution of diffusion model. It results in geometric inconsistency issues, dubbed as Multi-Face Janus Problem.", + "bbox": [ + 212, + 566, + 784, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Score Distillation Sampling (SDS). SDS optimization is widely adopted by text-to-3D generation pipelines [6,24,30,37,42]. Given a 3D representation with learnable parameters $\\theta$ and a pre-trained 2D diffusion model with noise prediction network $\\epsilon_{\\Phi}(\\mathbf{x}_t,t,y)$ , SDS optimizes $\\theta$ by minimizing the KL-divergence:", + "bbox": [ + 212, + 626, + 784, + 686 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} D _ {K L} \\left(q _ {t} ^ {\\theta} \\left(\\mathbf {x} _ {t} | c, y\\right) \\mid \\mid p _ {t} \\left(\\mathbf {x} _ {t} | y\\right)\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 699, + 785, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $p_t(\\mathbf{x}_t | y)$ is the image distribution sampled from diffusion model, $q_t^\\theta(\\mathbf{x}_t | c, y)$ is the distribution of rendered image $\\mathbf{x}_t = g(\\theta, c)$ with respect to camera pose $c$ at timestep $t$ of the forward diffusion process, where $g$ is the renderer. To solve Eq. (1), the score distillation function is derived as:", + "bbox": [ + 212, + 723, + 787, + 786 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\theta} L _ {S D S} (\\theta) \\triangleq \\mathrm {E} _ {t, \\mathbf {x}} [ w (t) \\frac {\\sigma_ {t}}{\\alpha_ {t}} \\nabla_ {\\theta} K L \\left(q _ {t} ^ {\\theta} \\left(\\mathbf {x} _ {t} | c, y\\right) \\mid \\mid p _ {t} \\left(\\mathbf {x} _ {t} | y\\right)\\right) ] \\\\ \\triangleq \\mathrm {E} _ {t, \\epsilon_ {\\Phi}} [ w (t) (\\hat {\\epsilon} _ {\\Phi} (\\mathbf {x} _ {t}, t, y) - \\epsilon) \\frac {\\delta g (\\theta , c)}{\\delta \\theta} ], \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 790, + 784, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "JointDreamer", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\epsilon}_{\\Phi} := (1 + s)\\epsilon_{\\Phi}(\\mathbf{x}_t, t, y) - s\\epsilon_{\\Phi}(\\mathbf{x}_t, t, \\emptyset)$ is modification of predicted noise with classifier-free guidance (CFG) $s$ , $w(t)$ is time-dependent weighting function. Multi-Face Janus Problem. To achieve consistency, it is essential that the rendering distributions $q_0^\\theta(\\mathbf{x}_0|c, y)$ adhere to text condition $y$ and image distribution $p_t(\\mathbf{x}_t|y)$ keep consistency across views with different poses. For the image distribution $p_0(\\mathbf{x}_0|y)$ of 2D diffusion model, the pose condition can be injected via input text with the corresponding directional prompt [37, 42]. As illustrated in Fig. 2(a), the pre-trained 2D image distribution, trained on individual images, is view-agnostic and lacks identity consistency across views. Even with a text tuning mechanism [1] specifically designed for multi-view image generation, as shown in Fig. 2(b), the above issues are far from resolved. Since SDS minimizes KL-divergence between the image distribution and rendering distribution independently for each rendered view, it can only inevitably inherit the 3D-awareness deficit of the 2D diffusion model, resulting in inconsistent 3D generation, which is commonly referred to as the Multi-face Janus Problem of SDS.", + "bbox": [ + 212, + 146, + 787, + 373 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 395, + 330, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we introduce JointDreamer, a novel text-to-3D generation framework as illustrated in Fig. 3. We first present the derivation of Joint Score Distillation (JSD) in Sec. 4.1, which extends the single-view optimization in SDS into a multi-view KL-Divergence. Then we integrate universal view-aware models into JSD to show the compatibility of JSD in Sec. 4.2, where we instantiate three kinds of view-aware models to capture inter-view coherence. Finally, we elaborate on the overall framework JointDreamer in Sec. 4.3, where we integrate the multi-view generation model into JSD. We also propose a geometry fading scheme and CFG switching strategy to further enhance generative quality.", + "bbox": [ + 212, + 426, + 787, + 564 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Joint Score Distillation (JSD)", + "text_level": 1, + "bbox": [ + 215, + 584, + 506, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address the multi-face problem arising from SDS, we extend the score distillation from single-view to multi-view settings and promote inter-view coherence across 2D image distribution, and thus derive our JSD optimization function.", + "bbox": [ + 212, + 608, + 785, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Coherence Modeling for Joint Image Distribution. As discussed above, the rendering distributions of the 3D representation should maintain 3D consistency across views $\\tilde{\\mathbf{x}} = \\{\\mathbf{x}^1,\\mathbf{x}^2,\\dots ,\\mathbf{x}^V\\}$ with respect to different poses $\\tilde{\\mathbf{c}} = \\{c^{1},c^{2},\\ldots ,c^{V}\\}$ . However, for 2D pre-trained diffusion models, different views are generated independently. To ensure consistency, we propose modeling the joint image distribution of multiple views, denoted as $p_0(\\tilde{\\mathbf{x}} |\\tilde{\\mathbf{c}},y)$ , within the diffusion model. Following the commonly adopted assumption in energy-based distribution modeling [21, 44, 45], we introduce an energy function that measures the inter-view coherence by $\\mathcal{C}(\\tilde{\\mathbf{x}},\\tilde{\\mathbf{c}}):\\mathbb{R}^{Vd}\\to \\mathbb{R}$ and define:", + "bbox": [ + 215, + 660, + 787, + 797 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {0} (\\tilde {\\mathbf {x}} | \\tilde {\\mathbf {c}}, y) \\propto \\exp (\\mathcal {C} (\\tilde {\\mathbf {x}}, \\tilde {\\mathbf {c}})) \\prod_ {i = 1} ^ {V} p _ {0} (\\mathbf {x} ^ {i} | c ^ {i}, y), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 806, + 785, + 844 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "C. Jiang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where a larger $\\mathcal{C}(\\tilde{\\mathbf{x}},\\tilde{\\mathbf{c}})$ indicates greater coherence among the denoised view images. As a result, the joint image distribution is no longer view-independent. In practice, the joint energy function can be implemented via various view-aware models, as long as they can reflect the coherence across multiple views. In Sec. 4.2, we explore different choices in depth. The modeling of coherence across joint image distributions facilitates our JSD on multi-view, integrating inter-view coherence to ensure consistent 3D representation.", + "bbox": [ + 214, + 146, + 787, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "KL-Divergence on Multiple Views. We extend the single-view KL-divergence in SDS to a multi-view version, based on the joint image distribution:", + "bbox": [ + 214, + 260, + 789, + 289 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\theta} D _ {K L} \\left(q _ {t} ^ {\\theta} (\\tilde {\\mathbf {x}} | \\tilde {\\mathbf {c}}, y) \\| p _ {t} (\\tilde {\\mathbf {x}} | \\tilde {\\mathbf {c}}, y)\\right) \\\\ = \\min _ {\\theta} D _ {K L} \\left(q _ {t} ^ {\\theta} (\\tilde {\\mathbf {x}} | \\tilde {\\mathbf {c}}, y) \\right\\lVert \\exp \\left(\\mathcal {C} \\left(\\tilde {\\mathbf {x}}, \\tilde {\\mathbf {c}}\\right)\\right) \\prod_ {i = 1} ^ {V} p _ {t} \\left(\\mathbf {x} ^ {i} \\mid c ^ {i}, y\\right), \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 300, + 784, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the extra energy term $\\mathcal{C}(\\tilde{\\mathbf{x}},\\tilde{\\mathbf{c}})$ in Eq. (3) accounts for the inter-view coherence. Without this constraint, e.g., $\\mathcal{C}(\\tilde{\\mathbf{x}},\\tilde{\\mathbf{c}})\\equiv 0$ , different rendering views are optimized independently with the 2D diffusion model separately. In this sense, the original SDS can be seen as a special case of JSD.", + "bbox": [ + 214, + 369, + 782, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Joint Score Distillation Function. To correspond to the gradient of the new rule of multi-view KL-divergence in Eq. (4), we derive our score distillation function that is jointly conducted on multiple views as follows:", + "bbox": [ + 214, + 438, + 782, + 483 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\theta} L _ {J S D} (\\theta) \\\\ \\triangleq \\mathrm {E} _ {t, \\epsilon_ {\\Phi}} [ w (t) \\mathbb {E} (\\nabla_ {\\tilde {\\mathbf {x}}} \\log q _ {t} (\\tilde {\\mathbf {x}} _ {t} | \\tilde {\\mathbf {x}} _ {0}) - \\nabla_ {\\tilde {\\mathbf {x}}} \\log p _ {t} (\\tilde {\\mathbf {x}} _ {t} | y)) ] \\tag {5} \\\\ = \\sum_ {i = 1} ^ {V} \\mathrm {E} _ {t, \\epsilon_ {\\Phi} ^ {i}} [ w (t) (\\hat {\\epsilon} _ {\\Phi} (\\mathbf {x} _ {t} ^ {i}, y) - \\frac {\\partial \\mathcal {C} (\\tilde {\\mathbf {x}})}{\\partial \\mathbf {x} _ {t} ^ {i}} - \\epsilon^ {i}) \\frac {\\delta g (\\theta , c ^ {i})}{\\delta \\theta} ], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 494, + 784, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\{\\epsilon^i\\}_{i = 1}^V$ are noises during score matching for different views. The proof can be found in the Appendix. To intuitively compare SDS with JSD, we sample multi-view images from the forward pass of pre-trained diffusion model [38] as shown in Fig. 2(c). For each view $\\mathbf{x}_t$ , we randomly select $\\mathbf{x}_t^\\prime$ from a different view and adopt the binary classification model presented in Sec. 4.2 as the energy function to measure coherence across the two views. The results illustrate that JSD significantly enhances the correspondence with directional prompts and consistency across different views, particularly for the initially biased views such as side and back views. These theoretical and empirical results prove that the multi-face Janus issue in SDS is rooted in the view-agnostic and view-biased 2D image distribution, a challenge effectively mitigated by JSD.", + "bbox": [ + 214, + 579, + 787, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Universal View-Aware Models as Energy Function", + "text_level": 1, + "bbox": [ + 214, + 768, + 674, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "JSD requires an energy function $\\mathcal{C}(\\tilde{\\mathbf{x}},\\tilde{\\mathbf{c}})$ to measure coherence across denoised images, as presented in Eq. (3). This energy function plays a crucial role in assessing the consistency between different views in image distribution. To demonstrate", + "bbox": [ + 214, + 794, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "JointDreamer", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e09f1c9ad9fd4190ae30fb8f1027fa4246413df1e8eb1bcb7d934abc5ec3222e.jpg", + "image_caption": [ + "Fig. 4: Illustration of the binary classification model and qualitative results with JSD. (a) The classification model $M_{\\mathrm{CLS}}$ produces the binary logit to measure the consistency between two input views $x^i$ and $x^j$ . (b) JSD integrated with the classification model effectively alleviates Janus issues compared to SDS." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 395, + 351 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8310201a10e7cba4b295123a8f7c531556b987951d0cd38f7015d09ecab44872.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 143, + 781, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the compatibility of JSD, we employ three different types of models trained for various representative multi-view tasks.", + "bbox": [ + 212, + 441, + 784, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Binary classification model $M_{\\mathrm{CLS}}$ . The binary classification model $M_{\\mathrm{CLS}}$ is designed to classify the content consistency between two input images based on their relative camera pose. To ensure computational efficiency, we introduce a dedicated classification model, as shown in Fig. 4 (a). The training process takes two days with a single A800 GPU on Objaverse dataset [8]. Further details can be found in the appendix. The classification model $M_{\\mathrm{CLS}}$ processes pairs of images $\\mathbf{x}^i$ and $\\mathbf{x}^j$ captured from different viewpoints $c^i$ and $c^j$ . It extracts image features using the DINO-ViT/s16 backbone [3]. These image features are conditioned on the camera feature obtained through MLP layers from the relative camera transformation matrix $\\Delta(c^j, c^i)$ . Finally, the classification head produces the binary score. To incorporate with JSD, we only consider neighboring views in $V$ as image pairs for coherence measurement, which is denoted as:", + "bbox": [ + 212, + 473, + 787, + 654 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {\\mathrm {C L S}} \\left(\\tilde {\\mathbf {x}}, \\tilde {\\mathbf {c}}\\right) = \\sum_ {i, j \\in 1, \\dots , V; i \\neq j} M _ {\\mathrm {C L S}} \\left(\\mathbf {x} _ {t} ^ {i}, \\mathbf {x} _ {t} ^ {j}, \\Delta \\left(c ^ {j}, c ^ {i}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 361, + 664, + 785, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where the higher logit indicates the stronger geometric consistency.", + "bbox": [ + 212, + 703, + 699, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Image-to-image translation model $M_{\\mathrm{I2I}}$ . The image-to-image translation model $M_{\\mathrm{I2I}}$ is tailored for novel view synthesis [26, 27, 29, 40]. We employ the most recent model, Wonder3D [29] as $M_{\\mathrm{I2I}}$ , which is a viewpoint-conditioned image translation model and generates consistent content in the target viewpoint. When integrated with JSD, a random reference view $\\mathbf{x}^{\\mathrm{ref}}$ is selected from the set of 3D rendered images. Then we input the relative camera transformation $\\Delta (c^{i}, c^{\\mathrm{ref}})$ and rendered images to $M_{\\mathrm{I2I}}$ . The measure of consistency is determined by calculating the reconstruction loss between the synthesized new", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "C. Jiang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/05ee31543744ce6424a8ff61b64d3a716f75d5b6b3cc881732c67aa4b1f6f364.jpg", + "image_caption": [ + "Fig. 5: Comparison of text-to-3D generation. See Appendix for more results." + ], + "image_footnote": [], + "bbox": [ + 212, + 141, + 787, + 481 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "image and its corresponding rendered image:", + "bbox": [ + 212, + 532, + 539, + 547 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {\\mathrm {I 2 I}} (\\tilde {\\mathbf {x}}, \\tilde {\\mathbf {c}}) = - \\sum_ {i \\in 1, \\dots , V} | | M _ {\\mathrm {I 2 I}} (\\mathbf {x} _ {t} ^ {\\mathbf {r e f}}, \\Delta (c ^ {i}, c ^ {\\mathbf {r e f}})) - \\mathbf {x} _ {t} ^ {i} | | _ {2} ^ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 558, + 785, + 589 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where a smaller reconstruction loss indicates stronger geometric consistency under the estimation of $M_{\\mathrm{I2I}}$ .", + "bbox": [ + 212, + 602, + 784, + 632 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Multi-view synthesis model $M_{\\mathbf{MVS}}$ . The multi-view synthesis model is designed to generate multiple images conditioned on text prompts and camera poses, wherein we employ very recent work MVDream [41] as $M_{\\mathbf{MVS}}$ . We compute the reconstruction loss for multiple generative views and rendered views:", + "bbox": [ + 212, + 633, + 784, + 694 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {\\mathbf {M V S}} (\\tilde {\\mathbf {x}}, \\tilde {\\mathbf {c}}) = - | | M _ {\\mathbf {M V S}} (y, \\tilde {\\mathbf {c}}) - \\tilde {\\mathbf {x}} | | _ {2} ^ {2} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 710, + 785, + 728 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The smaller reconstruction loss signifies better geometric consistency within $M_{\\mathbf{MVS}}$ . These view-aware models measure the coherence across views according to their own 3D-aware insights. Incorporated with JSD, they provide distinct constraints, resulting in different 3D generations that nonetheless contribute to enhanced geometric consistency. We believe that JSD can be adapted to more universal view-aware models, which enables us to progressively redefine the benchmark of 3D generation with the advancement of multi-view tasks.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "JointDreamer", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/102b4a6ebb3b696522f4e2d89ea052289fe8e8ecb09e08d9c900169d2752a6fa.jpg", + "image_caption": [ + "A pink Spiderman dancing ballet, Marvel character HD, highly detailed 3D model" + ], + "image_footnote": [], + "bbox": [ + 217, + 152, + 356, + 210 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3af7a027f877022bb7afa73cd7789c9983710496e07cac7a9cd7fb3cb47c2aa0.jpg", + "image_caption": [ + "a blue jay standing on a large basket of rainbow macarons" + ], + "image_footnote": [], + "bbox": [ + 359, + 152, + 501, + 209 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/1998689416a7d15587413fd8edaa95a77158f265e79006ff00bce141c91294ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 151, + 638, + 209 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3bf4446cea63ec8c1e91bed4a1f1ea8f725aaf73698439af7c00f5c0f6e68594.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 151, + 779, + 210 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/01eac03f88f6d4290468fb46d98dfca58dd4a52be5656373f7c754c23eb28048.jpg", + "image_caption": [ + "SDS" + ], + "image_footnote": [], + "bbox": [ + 217, + 210, + 356, + 275 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/69e36279d7daf173dbe8bab8cefae7d3f48863bbf6f2e02b3906265b591d5f1f.jpg", + "image_caption": [ + "JSD w/ $C_{\\mathrm{CLS}}$", + "Fig. 6: Ablations on JSD incorporated with different energy function $\\mathcal{C}$ . CLS: Binary Classification model; I2I: Image-to-Image Translation model (Wonder3D [29]); MVS: Multi-View Image Synthesis model (MVDream [41])." + ], + "image_footnote": [], + "bbox": [ + 370, + 218, + 500, + 277 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d72f5478b2d4cf76bd21e429cf0f6c55e92504ce12e7261538d26229a13e2bb3.jpg", + "image_caption": [ + "JSD w/C12I" + ], + "image_footnote": [], + "bbox": [ + 514, + 218, + 638, + 277 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/1fc711d4cbf5a4eef051a0c6dc9d7104b6a863b16842ee8a70fe681190930047.jpg", + "image_caption": [ + "JSD w/C_MVS" + ], + "image_footnote": [], + "bbox": [ + 651, + 218, + 781, + 277 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Framework of JointDreamer", + "text_level": 1, + "bbox": [ + 215, + 368, + 495, + 383 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Building upon the JSD optimization, we propose the overall framework JointDreamer. The optimization is based on neural radiance field (NeRF) [31], adopting Instant-NGP [32] with volume renderer. We adopt the multi-view synthesis model $M_{\\mathrm{MVS}}$ as the energy function to integrate with JSD in JointDreamer. During optimization training, we utilize the common techniques including time-annealing and resolution scaling-up following [41, 42]. Besides, we propose two novel techniques to further enhance the generation quality, including a Geometry Fading scheme and a Classifier-Free Guidance Scale (CFG) Switching strategy. Geometry Fading. We aim to shift attention between geometric structure and texture details during optimization. Specifically, starting from iteration 5K, we reduce the learning rate of the density network of NeRF from $1e - 2$ to $1e - 6$ and set orientation loss to 0. Consequently, it benefits geometric convergence in the early phase of optimization, while allowing for decreased attention on geometry and increased attention on texture enhancement in the later stages.", + "bbox": [ + 212, + 392, + 787, + 603 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "CFG Switching. CFG scheduling strategies have been employed in 2D domain [15,39] to enhance quality. In this work, we propose to modify the CFG scale $s$ during the training for 3D generation. We are motivated by the observation that a large CFG scale can lead to accelerated geometric convergence but may result in under-optimized geometry and distorted texture. Unlike the annealed CFG approach in CLIP-Sculptor [39], our increasing strategy prioritizes texture while maintaining accurate geometry. Specifically, a smaller $s = 30$ is employed in the early stages to preserve shape integrity, which allows for stronger coherence guidance from JSD. After 5K iterations, we increase the $s$ to 50, enhancing texture fidelity and overall quality.", + "bbox": [ + 212, + 604, + 787, + 756 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiment", + "text_level": 1, + "bbox": [ + 215, + 777, + 366, + 795 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we present the text-to-3D generation results of JointDreamer with qualitative and quantitative evaluations, illustrating state-of-the-art per", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "C. Jiang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "a rabbit cutting grass with a lawnmower", + "text_level": 1, + "bbox": [ + 406, + 143, + 617, + 152 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0371e9faf6e2e8e83e1302f14370651347026ce9da59593af65db1532a4204e1.jpg", + "image_caption": [ + "SDS", + "Fig. 7: Ablations on CFG Switching and Geometry Fading techniques in JointDreamer pipeline, demonstrating the effectiveness of quality enhancement." + ], + "image_footnote": [], + "bbox": [ + 217, + 155, + 326, + 234 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/42c9d79ec65651d1664022af306902beb92bac090fbc22f4db7319ca0e7cbc2d.jpg", + "image_caption": [ + "original JSD" + ], + "image_footnote": [], + "bbox": [ + 330, + 152, + 478, + 234 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/cebe4868e9d9921597ed69400d942ae6a6489b2a36e5573d18a8247e7637d85e.jpg", + "image_caption": [ + "+CFG Switching" + ], + "image_footnote": [], + "bbox": [ + 480, + 152, + 630, + 234 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3fdc32514adaa3be98d6b95b0af59b380c4d4cd51a361f0d083cda149af1bd97.jpg", + "image_caption": [ + "+ Geo Fading" + ], + "image_footnote": [], + "bbox": [ + 632, + 152, + 784, + 234 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "performance. We also make further ablation analysis on the proposed JSD. More training and evaluation details can be found in the Appendix.", + "bbox": [ + 212, + 308, + 782, + 338 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Text-to-3D Generation", + "text_level": 1, + "bbox": [ + 214, + 359, + 450, + 373 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/51f9d1580bf1f5e25e8157774556395ba0088d10d44e943c57071816056103c0.jpg", + "table_caption": [ + "Table 1: Quantitative results on textual consistency and user preference, tested on object-centric subset of MS-COCO [25]." + ], + "table_footnote": [], + "table_body": "
MethodCLIP Score ↑R-Precision ↑User Study↑
DreamFusion [37]20.127.718.2
ProlificDreamer [42]25.018.716.2
MVDream [41]20.833.623.5
JointDreamer27.788.542.1
", + "bbox": [ + 223, + 426, + 540, + 492 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/088b748911caa73510d1906497d96864c0a9c19f59804325ae7af353a6d86a43.jpg", + "table_caption": [ + "Table 2: Ablation study on CFG Switching (CFGS) and Geometry Fading (GF)." + ], + "table_footnote": [], + "table_body": "
SDS JSD CFGS GF|CLIP Score↑ FID↓
✓20.0 429.2
✓27.6 360.7
✓ ✓28.2 357.6
✓ ✓ ✓28.8 353.9
", + "bbox": [ + 560, + 426, + 772, + 492 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative Comparisons. We compare with several representative baselines on threestudio [11] project, which modularizes the text-to-3D framework allowing for fair comparisons by ablating individual components. The generative samples are shown in Fig. 5. (i) DreamFusion [37]: Compared to DreamFusion which utilizes traditional SDS optimization, JointDreamer significantly improves the Multi-Face Janus issues in the generations of DreamFusion by introducing an energy term to ensure inter-view coherence. (ii) Magic3D [24]: Magic3D introduces a two-stage generation pipeline, transferring NeRF to DMTet in the second stage to enhance generation quality. We also transfer NeRF to DMTet as JointDreamer-DMTet, showcasing consistent superiority in geometry and texture quality by utilizing JSD optimization instead of SDS. (iii) ProlificDreamer [42]: ProlificDreamer presents variational score distillation (VSD) as a variant of score distillation function. While VSD enhances photorealism in 3D renderings by introducing a LoRA model, the ill-posed association of pose and images during LoRA training deepens the geometric inconsistency of 3D representation, resulting in severe multi-view artifacts in generations. In contrast, JointDreamer with JSD achieves geometric consistency while maintaining high-fidelity texture quality. (iv) MVDream [41]: Compared to direct distillation from a finetuned model as MVDream, JointDreamer employs view-aware models as the coherence constraint in JSD, while still inherit the generalization capabilities of original diffusion models [38]. The results indicate that JointDreamer mitigates the overfitting issues of MVDream, which accurately responds to complex input text and enhances the 3D consistency of generations.", + "bbox": [ + 212, + 493, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "JointDreamer", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative Comparisons. Following [18,37], we evaluate CLIP Score [14], CLIP R-Precision [36] and user preference on the object-centric caption subset of MS-COCO [25] with 153 prompts to measure the text congruence and generative quality. For computational efficiency, we generate each 3D asset using 5K iterations of $64 \\times 64$ rendered images and render 20 images per caption for evaluation. CLIP ViT-B/32 is adopted as the feature extractor for Clip Score and Clip R-Precision. As illustrated by the results in Table 1, JointDreamer can outperform all baselines on CLIP Score and CLIP R-Precision by large margins. Specifically, JointDreamer achieves an improvement of the R-Precision by $60.8\\%$ and $54.9\\%$ over DreamFusion and MVDream, demonstrating its superior corresponding to textual description. Notably, the severe Janus artifacts in ProlificDreamer compromise the quality of rendering with noisy background and semantic distortion, resulting in the lowest R-Precision. We also conduct a user study about shape preference on these prompts in Table 1.", + "bbox": [ + 212, + 146, + 787, + 359 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Ablation Analyses", + "text_level": 1, + "bbox": [ + 215, + 383, + 413, + 398 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablations on Energy Functions. Sec. 4.2 discusses the varying interview", + "bbox": [ + 215, + 407, + 785, + 422 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "coherency measurements provided by view-aware models trained. When incorporated with JSD, these models have distinct impacts on 3D generations, as shown in Fig. 6. The binary classifier effectively corrects inaccurate geometric structures in SDS. However, it cannot introduce additional imaginative elements as a discriminative model, resulting in oversaturated and monotonous textures. In con", + "bbox": [ + 212, + 422, + 514, + 571 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/07988e616f80e452f474179d33e7e988608c31405acf13acb23e5d6287550fd9.jpg", + "table_caption": [ + "Table 3: Quantitative comparison of energy functions, showcasing the effectiveness of JSD in mitigating Janus artifacts with comparable computational efficiency to SDS." + ], + "table_footnote": [], + "table_body": "
MethodsJanus Rate ↓ GPU Memory ↓ Train Time ↓
SDS100%16.1 G50 min.
JSD w/CCLS12.5%22.1 G80 min.
JSD w/C2I31.2%16.0 G119 min.
JSD w/C_MVS6.2%19.4 G54 min.
", + "bbox": [ + 532, + 507, + 782, + 561 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "trast, as generative models, Wonder3D [29] and MVDream [41] employ reconstruction loss to estimate 3D consistency. Hence, they not only guide geometric structural modifications but also influence texture quality. we further conduct a quantitative comparison using 16 complex multi-Janus prompts, as outlined in Table 3. The experimental setup details can be found in the Appendix. The results indicate that our JSD consistently mitigates Janus artifacts across different view-aware models, with only a slight increase in computational requirements compared to SDS. We find that the energy function $C_{12\\mathrm{I}}$ derived from the imagedto-image translation model exhibits poor performance, likely attributed to a mismatch in camera range and inaccurate translated results.", + "bbox": [ + 212, + 571, + 787, + 724 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Analysis on Geometry Fading and CFG Switching Mechanisms. We conduct incremental ablations on our proposed techniques in JointDreamer, including the Geometry Fading scheme and CFG Switching strategy. As shown in Fig. 7, increasing the CFG value enhances texture detail compared to the original JSD, but over-optimizes the geometry, resulting in a bumpier shape. Compared to “+CFG Switching”, the Geometry fading effectively protects shape when larger CFG guidance. We also conduct a quantitative evaluation on a $30\\%$", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "C. Jiang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e157a862a628d12ee2b43482add96ef10c550bad5e80f4812bd18bd605ff4218.jpg", + "image_caption": [ + "Fig. 9: Discuss the impact of prompts and random seeds, demonstrating the robustness of JointDreamer. $/...$ means the same content as the prompt in Fig. 7." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 784, + 292 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "MS-COCO subset in Table 2. JSD demonstrates superior texture quality compared to SDS, as evidenced by the Clip Score (CS) and FID metrics. The two proposed mechanisms further enhance the texture quality.", + "bbox": [ + 212, + 354, + 782, + 400 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Discussions on Training Loss. To make further comparisons with JSD", + "bbox": [ + 214, + 406, + 784, + 421 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "and SDS, we aggregate training losses from multiple prompts on two optimization functions and visualize the training loss curve as illustrated in Fig. 8. We observe that SDS experiences significant fluctuations due to the randomness of single-view optimization. In contrast, JSD can converge gradually and smoothly, demonstrating the introduction of multi", + "bbox": [ + 212, + 422, + 486, + 571 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/bbe2af54d2d91285ab5177d2cb8a553833888e958203ae0ca9e1017c2c00c2c4.jpg", + "image_caption": [ + "Fig.8: Training loss comparisons of JSD and SDS. JSD eliminates randomness fluctuations in SDS convergence." + ], + "image_footnote": [], + "bbox": [ + 500, + 425, + 785, + 512 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "view optimization with inter-view coherence in JSD can reduce the randomness of optimization and contribute to better convergence for 3D representation.", + "bbox": [ + 212, + 573, + 782, + 603 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Discussions on Robustness for Prompts and Random Seed. To show the impact of seed and prompt, we modify seeds and key prompt components in Fig. 7, such as subject, object and verb. Results in Fig. 9 demonstrate the robustness of JointDreamer for different seed. Note that the default seed is 0.", + "bbox": [ + 212, + 609, + 782, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The Effectiveness of Classification Model. The classification model surpasses MVDream in training speed by a factor of 48. Fig.4 (b) provides additional high-quality results obtained using JSD in conjunction with the binary classification model $M_{\\mathrm{CLS}}$ . Furthermore, we conduct an ablation study by replacing $\\frac{\\partial\\mathcal{C}(\\hat{\\mathbf{x}})}{\\partial\\mathbf{x}_i^i}$ in Eq.5 with a randomly generated value between [0,1]. However, it yields shapeless results due to unrelated disturbances in the optimization process. These findings highlight the significance of the proposed classification model in achieving a balance between computational efficiency and generation quality.", + "bbox": [ + 212, + 676, + 784, + 801 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "More Comparison with MVDream. View-aware models can be incorporated into 3D generation by combining them with JSD or using them directly with", + "bbox": [ + 214, + 809, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "JointDreamer", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5449a51b441ed97c9047f6bbe0da35a093cb2df9d2136515ff4b5f87a604bbc0.jpg", + "image_caption": [ + "Fig. 10: Comparison of JointDreamer and SDS+MVDream, illustrating the superiority of JointDreamer in mitigating the Janus problem. SDS+MVDream consistently exhibits semantic missing or Janus issues with different weight options." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 777, + 265 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "SDS. In this discussion, we use MVDream [41] as an example to demonstrate the superiority of JSD. MVDream can generate consistent shapes when calculating the SDS loss directly. However, it may miss components in the complete input text due to its fine-tuning on limited 3D data, as illustrated in Fig. 5. To address this limitation, a straightforward approach is to \"SDS+MVDream\".", + "bbox": [ + 212, + 335, + 782, + 411 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "However, balancing the impact of SDS and MVDream is challenging. As shown in Fig. 10, setting $\\lambda_{\\mathrm{SDS}}$ to 0 degrades to MVDream alone, while setting $\\lambda_{\\mathrm{MVDream}}$ to 0 yields a combination similar to DreamFusion [37]. Achieving a balanced impact between SDS and MVDream through a simple combination is difficult. When $\\lambda_{\\mathrm{MVDream}}$ is large, textual consistency remains constrained, whereas decreasing $\\lambda_{\\mathrm{MVDream}}$ leads to the Janus problem. This is due to gradient misalignment across multiple views in the \"SDS+MVDream\" combination, as SDS lacks multi-view information and cannot derive our objective outlined in Eq. 5. In contrast, JSD based on the joint image distribution provides supervision for text consistency and high-fidelity texture. JSD also promotes inter-view consistency by introducing an energy term with the view-aware model.", + "bbox": [ + 212, + 412, + 787, + 580 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 606, + 359, + 622 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we introduce Joint Score Distillation (JSD) as a new paradigm for text-to-3D generation, which conducts multi-view optimization jointly and accounts for inter-view coherence. We demonstrate that JSD can significantly enhance 3D coherence while maintaining generalizability. With other proposed techniques, our overall framework, JointDreamer, is capable of geometric-consistent and high-fidelity 3D generation adhering to complex text input.", + "bbox": [ + 212, + 642, + 802, + 733 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. While the training time of JointDreamer is comparable to existing SDS pipelines, there is room for improvement in terms of acceleration. Future work will explore alternative 3D representations, such as 3D Gaussian [20]. Additionally, JSD utilizes view-aware models to ensure geometry consistency and mitigate the impact of limited data. However, view-aware models still require 3D data for training, and further efficient 3D data collection or reconstruction from multi-view images is also worth investigating.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "C. Jiang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Armandpour, M., Zheng, H., Sadeghian, A., Sadeghian, A., Zhou, M.: Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. In: ICLR (2024)", + "2. Cao, Z., Hong, F., Wu, T., Pan, L., Liu, Z.: Large-vocabulary 3d diffusion model with transformer. In: ICLR (2024)", + "3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV. pp. 9650-9660 (2021)", + "4. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: CVPR. pp. 16123-16133 (2022)", + "5. Chen, H., Gu, J., Chen, A., Tian, W., Tu, Z., Liu, L., Su, H.: Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction. arXiv preprint arXiv:2304.06714 (2023)", + "6. Chen, R., Chen, Y., Jiao, N., Jia, K.: *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation*. In: ICCV. pp. 22246–22256 (2023)", + "7. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. In: NeurIPS (2024)", + "8. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objverse: A universe of annotated 3d objects. In: CVPR. pp. 13142-13153 (2023)", + "9. Deng, Y., Yang, J., Xiang, J., Tong, X.: Gram: Generative radiance manifolds for 3d-aware image generation. In: CVPR. pp. 10673-10683 (2022)", + "10. Gao, J., Shen, T., Wang, Z., Chen, W., Yin, K., Li, D., Litany, O., Gojcic, Z., Fidler, S.: Get3d: A generative model of high quality 3d textured shapes learned from images. NeurIPS 35, 31841-31854 (2022)", + "11. Guo, Y.C., Liu, Y.T., Shao, R., Laforte, C., Voleti, V., Luo, G., Chen, C.H., Zou, Z.X., Wang, C., Cao, Y.P., Zhang, S.H.: threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio (2023)", + "12. Henderson, P., Ferrari, V.: Learning single-image 3d reconstruction by generative modelling of shape, pose and shading. IJCV 128(4), 835-854 (2020)", + "13. Henderson, P., Tsiminaki, V., Lampert, C.H.: Leveraging 2d data to learn textured 3d mesh generation. In: CVPR. pp. 7498-7507 (2020)", + "14. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: CLIPScore: a reference-free evaluation metric for image captioning. In: EMNLP (2021)", + "15. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)", + "16. Hu, Z., Zhao, M., Zhao, C., Liang, X., Li, L., Zhao, Z., Fan, C., Zhou, X., Yu, X.: Efficientdreamer: High-fidelity and robust 3d creation via orthogonal-view diffusion priors. In: CVPR. pp. 4949-4958 (2024)", + "17. Huang, Y., Wang, J., Shi, Y., Tang, B., Qi, X., Zhang, L.: Dreamtime: An improved optimization strategy for diffusion-guided 3d generation. In: ICLR (2023)", + "18. Jain, A., Mildenhall, B., Barron, J.T., Abbeel, P., Poole, B.: Zero-shot text-guided object generation with dream fields. In: CVPR. pp. 867-876 (2022)" + ], + "bbox": [ + 218, + 176, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "JointDreamer", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Jun, H., Nichol, A.: Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023)", + "20. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4), 1-14 (2023)", + "21. LeCun, Y., Chopra, S., Hadsell, R., Ranzato, M., Huang, F.: A tutorial on energy-based learning. Predicting structured data 1(0) (2006)", + "22. Li, M., Zhou, P., Liu, J.W., Keppo, J., Lin, M., Yan, S., Xu, X.: Instant3d: Instant text-to-3d generation. arXiv preprint arXiv:2311.08403 (2023)", + "23. Li, W., Chen, R., Chen, X., Tan, P.: Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. In: ICLR (2024)", + "24. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: CVPR (2023)", + "25. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "26. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: ICCV. pp. 9298-9309 (2023)", + "27. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. In: ICLR (2024)", + "28. Liu, Z., Feng, Y., Black, M.J., Nowrouzezahrai, D., Paull, L., Liu, W.: Meshdiffusion: Score-based generative 3d mesh modeling. In: ICLR (2023)", + "29. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3d: Single image to 3d using cross-domain diffusion. In: CVPR (2024)", + "30. Luo, W., Hu, T., Zhang, S., Sun, J., Li, Z., Zhang, Z.: Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. NeurIPS 36 (2024)", + "31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "32. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)", + "33. Nguyen-Phuoc, T., Li, C., Theis, L., Richardt, C., Yang, Y.L.: Hologan: Unsupervised learning of 3d representations from natural images. In: ICCV. pp. 7588-7597 (2019)", + "34. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751 (2022)", + "35. Niemeyer, M., Geiger, A.: Giraffe: Representing scenes as compositional generative neural feature fields. In: CVPR. pp. 11453-11464 (2021)", + "36. Park, D.H., Azadi, S., Liu, X., Darrell, T., Rohrbach, A.: Benchmark for compositional text-to-image synthesis. In: NeurIPS Datasets and Benchmarks Track (2021)", + "37. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: ICLR (2023)", + "38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + ], + "bbox": [ + 217, + 147, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "C. Jiang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Sanghi, A., Fu, R., Liu, V., Willis, K.D., Shayani, H., Khasahmadi, A.H., Sridhar, S., Ritchie, D.: Clip-sculptor: Zero-shot generation of high-fidelity and diverse shapes from natural language. In: CVPR. pp. 18339-18348 (2023)", + "40. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110 (2023)", + "41. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. In: ICLR (2024)", + "42. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In: NeurIPS (2024)", + "43. Watson, D., Chan, W., Martin-Brualla, R., Ho, J., Tagliasacchi, A., Norouzi, M.: Novel view synthesis with diffusion models. In: ICLR (2023)", + "44. Weese, J., Kaus, M., Lorenz, C., Lobregt, S., Truyen, R., Pekar, V.: Shape constrained deformable models for 3d medical image segmentation. In: Information Processing in Medical Imaging: 17th International Conference, IPMI 2001 Davis, CA, USA, June 18-22, 2001 Proceedings 17. pp. 380-387. Springer (2001)", + "45. Zhao, W., Yan, L., Zhang, Y.: Geometric-constrained multi-view image matching method based on semi-global optimization. Geo-spatial information science 21(2), 115-126 (2018)" + ], + "bbox": [ + 212, + 146, + 787, + 424 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "JointDreamer", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/JointDreamer_ Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation/4667ae70-c38c-469d-a697-9d92865631f2_model.json b/2024/JointDreamer_ Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation/4667ae70-c38c-469d-a697-9d92865631f2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0921fd113a3b0b480f70a7bb742da41b8db6cb0b --- /dev/null +++ b/2024/JointDreamer_ Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation/4667ae70-c38c-469d-a697-9d92865631f2_model.json @@ -0,0 +1,2412 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.23, + 0.141, + 0.776, + 0.208 + ], + "angle": 0, + "content": "JointDreamer: Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.234, + 0.785, + 0.267 + ], + "angle": 0, + "content": "Chenhan Jiang\\(^{1*†}\\), Yihan Zeng\\(^{2*}\\), Tianyang Hu\\(^{2}\\), Songcun Xu\\(^{2}\\), Wei Zhang\\(^{2}\\), Hang Xu\\(^{2}\\), and Dit-Yan Yeung\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.277, + 0.689, + 0.292 + ], + "angle": 0, + "content": "1 The Hong Kong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.414, + 0.293, + 0.59, + 0.305 + ], + "angle": 0, + "content": "2 Huawei Noah's Ark Lab" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.306, + 0.726, + 0.32 + ], + "angle": 0, + "content": "*Equal contribution. †Corresponding Author: jchcyan@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.384, + 0.321, + 0.619, + 0.334 + ], + "angle": 0, + "content": "https://jointdreamer.github.io" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.368, + 0.744, + 0.66 + ], + "angle": 0, + "content": "Abstract. Score Distillation Sampling (SDS) by well-trained 2D diffusion models has shown great promise in text-to-3D generation. However, this paradigm distills view-agnostic 2D image distributions into the rendering distribution of 3D representation for each view independently, overlooking the coherence across views and yielding 3D inconsistency in generations. In this work, we propose Joint Score Distillation (JSD), a new paradigm that ensures coherent 3D generations. Specifically, we model the joint image distribution, which introduces an energy function to capture the coherence among denoised images from the diffusion model. We then derive the joint score distillation on multiple rendered views of the 3D representation, as opposed to a single view in SDS. In addition, we instantiate three universal view-aware models as energy functions, demonstrating compatibility with JSD. Empirically, JSD significantly mitigates the 3D inconsistency problem in SDS, while maintaining text congruence. Moreover, we introduce the Geometry Fading scheme and Classifier-Free Guidance (CFG) Switching strategy to enhance generative details. Our framework, JointDreamer, establishes a new benchmark in text-to-3D generation, achieving outstanding results with an \\(88.5\\%\\) CLIP R-Precision and \\(27.7\\%\\) CLIP Score. These metrics demonstrate exceptional text congruence, as well as remarkable geometric consistency and texture fidelity." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.673, + 0.658, + 0.688 + ], + "angle": 0, + "content": "Keywords: 3D Vision \\(\\cdot\\) 3D Generation \\(\\cdot\\) Energy Function" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.734, + 0.377, + 0.75 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.794, + 0.843 + ], + "angle": 0, + "content": "3D content creation is essential for diverse applications, including gaming, robotics simulation, and virtual reality. However, it is labor-intensive, demanding substantial time for skilled designers to create a single 3D asset. Hence, automating 3D creation with text input has attracted considerable attention. Recently, the score distillation sampling (SDS) algorithm pioneered by DreamFusion [37]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "C. Jiang et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.775, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.578, + 0.788, + 0.621 + ], + "angle": 0, + "content": "Fig. 1: Text-to-3D generations by JointDreamer from scratch. JointDreamer excels in generating geometrically consistent and high-fidelity 3D assets, adhering to complex textual descriptions that are challenging for previous methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.639, + 0.788, + 0.716 + ], + "angle": 0, + "content": "shows promise in text-to-3D tasks, which lifts image distribution from a well-trained diffusion model [38] into parameterized 3D representation like NeRF [31]. Compared to 3D generative models [2, 19, 34] that struggle with producing arbitrary objects due to limited text-3D training data, SDS-based methods [6, 24, 37, 42] can generate arbitrary 3D assets with diverse text input." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Although SDS-based methods benefit from the generalizability of diffusion models, they often encounter a common issue known as Janus artifacts [6, 24, 37, 42]. These artifacts manifest as repeated content from different viewpoints of a 3D generation, yielding a lack of realism and coherence in the rendered views. We investigate the Janus artifacts by visualizing image distributions of the diffusion model [38] from multiple viewpoints, as illustrated in Fig. 2(a). The results reveal the view-agnostic nature and the content inconsistency across views of the diffusion model. Consequently, SDS optimizes each rendered view of 3D" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "JointDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "representation independently and inherits image distribution without coherent multi-view perspective, leading to the geometric inconsistency of 3D generations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.187, + 0.788, + 0.324 + ], + "angle": 0, + "content": "Existing works [1,22] address the aforementioned challenges within the SDS framework by employing prompt engineering techniques. However, the effectiveness of such methods remains unsatisfactory, as evident from the results depicted in Fig. 2(b). Alternative methods [16,41] propose to finetune view-aware diffusion models using rendered images of 3D datasets [7,8]. Nevertheless, they are prone to overfitting on limited text-3D training data [23], decreasing the text congruence when handling complex text inputs. Based on the above observations, it requires a rethinking of the SDS optimization to enhance the 3D coherence of generations while maintaining generalizability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.518, + 0.604 + ], + "angle": 0, + "content": "In this work, we first present Joint Score Distillation (JSD), which significantly promotes the 3D consistency of generation and inherits generalizability from diffusion models. Specifically, we model the joint image distribution of diffusion model via an energy function measuring coherence across denoised images. It facilitates the extension of the KL-Divergence in SDS from single-view into multi-view. We then derive the joint score distillation function from multi-view KL-Divergence, which ensures inter-view coherence in the optimization process of 3D generation. We show that SDS is a special case of JSD with the energy term omitted, which indicates the absence of coherence constraint across views." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.518, + 0.748 + ], + "angle": 0, + "content": "Building upon JSD optimization, we present three view-aware models as energy terms to showcase the compatibility of JSD: the Binary Classification Model, the Image-to-Image Translation Model, and the Multi-view Generation Model. Through empirical analysis, it is observed that different view-aware models introduce distinct coherence measurements," + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "leading to diverse 3D generations, while all contributing to 3D consistency. Furthermore, to facilitate a more comprehensive comparison with existing text-to-3D generation methods, we introduce JointDreamer, an innovative framework capable of producing geometric-consistent and high-fidelity 3D assets adhering to complex text descriptions. Notably, in addition to incorporating a Multi-view Generation model as an energy term in JSD, we introduce two complementary" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.336, + 0.782, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.506, + 0.789, + 0.729 + ], + "angle": 0, + "content": "Fig.2: Illustration of text-conditioned images for different viewpoints, where input texts are augmented with corresponding direction prompts for each view. (a) The original generations from 2D diffusion model [38] are view-agnostic and inconsistent across views. (b) Text prompt tuning [1] has limited improvement in the directional structure of generated images for each view. (c) JSD injects coherence measurement from the proposed binary classifier (refer to Section 4.2), contributing to modified directional structures and semantical consistency across views." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "C. Jiang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "techniques, namely the Geometry Fading scheme and the Classifier-Free Guidance (CFG) Switching strategy, to enhance generative details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.266 + ], + "angle": 0, + "content": "We systematically assess the quality of our approach, both qualitatively and quantitatively, compared to existing methods. Qualitative results gallery can be found in Fig. 1. Our JointDreamer consistently produces high-fidelity 3D assets and mitigates Janus artifacts in SDS. It maintains text congruence even when confronted with complex text input, achieving \\(88.5\\%\\) CLIP R-Precision and \\(27.7\\%\\) CLIP Score." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.269, + 0.625, + 0.282 + ], + "angle": 0, + "content": "In brief, our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.295, + 0.784, + 0.338 + ], + "angle": 0, + "content": "- We introduce a novel Joint Score Distillation (JSD) for text-to-3D generation, optimizing multiple views jointly via an energy function to capture inter-view coherence." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.341, + 0.784, + 0.37 + ], + "angle": 0, + "content": "- We present three view-aware models as energy functions to show compatibility with JSD, all of them mitigate the Janus problem in SDS." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.431 + ], + "angle": 0, + "content": "- We introduce the text-to-3D framework JointDreamer, incorporating complementary Geometry Fading and CFG Switching techniques. Our JointDreamer achieves geometrically consistent and high-fidelity 3D assets even with complex textual inputs." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.295, + 0.785, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.456, + 0.396, + 0.472 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.484, + 0.787, + 0.787 + ], + "angle": 0, + "content": "Text-to-3D Generation. Existing text-to-3D generation methods can be categorized into two streams: 3D generative models and 2D optimization methods. The former encompasses various deep generative models such as Variational Auto Encoders (VAEs) [12, 13], Generative Adversarial Models (GAN) [4, 9, 10, 33, 35], diffusion models [5, 28, 34] and transformer architectures [2, 19]. These models are efficient in inference but often struggle with generalizability and training stability, attributed to the limited scope and complexity of available 3D datasets. The latter approach centers around the Score Distillation Sampling (SDS) algorithm proposed by [37], which leverages 2D diffusion model priors [38] for optimizing 3D representations. Subsequent advancements have refined this technique by improving the 3D representations [6, 24], the sampling scheduler [17] and loss design [42]. However, the above approaches overlook the geometric consistency problem, facing inherent multi-face Janus issues in SDS. Prior works try to alleviate the Janus issues with prompt tuning [1] yet they achieve limited effect. Very recent work MVDream [41] addresses the problem by fine-tuning a multi-view diffusion model, but it is susceptible to overfitting on scarce 3D training data, compromising semantic consistency in text-to-3D generations. In this work, we address the fundamental flaw of SDS that optimizes each view independently by introducing a joint optimization function that enforces inter-view consistency, essentially solving the Janus issues in SDS while preserving its generalizability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Diffusion-based Novel View Synthesis. As an alternative to 3D generation, novel view synthesis models the challenge as a view-conditioned image-to-image translation task. There have been proposals for pose-conditioned image-to-image" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "JointDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.782, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.279, + 0.785, + 0.322 + ], + "angle": 0, + "content": "Fig. 3: Overview of JointDreamer Framework. We introduce an energy function to model the joint distribution for multi-view denoised images from 2D diffusion model, facilitating the Joint Score Distillation (JSD) optimization for text-to-3D generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.788, + 0.519 + ], + "angle": 0, + "content": "diffusion models [43] that generate novel views on synthetic data in 3D. Recently, [26] promoted the generalizability of novel view synthesis by fine-tuning the 2D diffusion model [38] on renderings of 3D dataset, which facilitates images-to-3D tasks with 3D reconstruction or SDS algorithm. To further improve the 3D consistency across generated views and input view, very recent works [27, 29, 40] modify the generation process into multi-view generation and present corresponding architecture designs. Generally, these methods take camera specifications as conditions and enable the viewpoint-aware generations, but they can hardly accurately capture a complete 3D scene consistently and densely. Our method acknowledges the potential of these models to discern relative interview relationships, which we harness to provide inter-view coherence for our JSD, thus serving as universal guidance models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.538, + 0.382, + 0.554 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.785, + 0.626 + ], + "angle": 0, + "content": "We review the original SDS and address its fundamental limitation: optimizing independently for each single view of 3D representation and distilling with the view-agnostic image distribution of diffusion model. It results in geometric inconsistency issues, dubbed as Multi-Face Janus Problem." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.785, + 0.688 + ], + "angle": 0, + "content": "Score Distillation Sampling (SDS). SDS optimization is widely adopted by text-to-3D generation pipelines [6,24,30,37,42]. Given a 3D representation with learnable parameters \\(\\theta\\) and a pre-trained 2D diffusion model with noise prediction network \\(\\epsilon_{\\Phi}(\\mathbf{x}_t,t,y)\\), SDS optimizes \\(\\theta\\) by minimizing the KL-divergence:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.7, + 0.787, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} D _ {K L} \\left(q _ {t} ^ {\\theta} \\left(\\mathbf {x} _ {t} | c, y\\right) \\mid \\mid p _ {t} \\left(\\mathbf {x} _ {t} | y\\right)\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.724, + 0.788, + 0.787 + ], + "angle": 0, + "content": "Here, \\( p_t(\\mathbf{x}_t | y) \\) is the image distribution sampled from diffusion model, \\( q_t^\\theta(\\mathbf{x}_t | c, y) \\) is the distribution of rendered image \\( \\mathbf{x}_t = g(\\theta, c) \\) with respect to camera pose \\( c \\) at timestep \\( t \\) of the forward diffusion process, where \\( g \\) is the renderer. To solve Eq. (1), the score distillation function is derived as:" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.791, + 0.785, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\theta} L _ {S D S} (\\theta) \\triangleq \\mathrm {E} _ {t, \\mathbf {x}} [ w (t) \\frac {\\sigma_ {t}}{\\alpha_ {t}} \\nabla_ {\\theta} K L \\left(q _ {t} ^ {\\theta} \\left(\\mathbf {x} _ {t} | c, y\\right) \\mid \\mid p _ {t} \\left(\\mathbf {x} _ {t} | y\\right)\\right) ] \\\\ \\triangleq \\mathrm {E} _ {t, \\epsilon_ {\\Phi}} [ w (t) (\\hat {\\epsilon} _ {\\Phi} (\\mathbf {x} _ {t}, t, y) - \\epsilon) \\frac {\\delta g (\\theta , c)}{\\delta \\theta} ], \\tag {2} \\\\ \\end{array}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "C. Jiang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.374 + ], + "angle": 0, + "content": "where \\(\\hat{\\epsilon}_{\\Phi} := (1 + s)\\epsilon_{\\Phi}(\\mathbf{x}_t, t, y) - s\\epsilon_{\\Phi}(\\mathbf{x}_t, t, \\emptyset)\\) is modification of predicted noise with classifier-free guidance (CFG) \\(s\\), \\(w(t)\\) is time-dependent weighting function. Multi-Face Janus Problem. To achieve consistency, it is essential that the rendering distributions \\(q_0^\\theta(\\mathbf{x}_0|c, y)\\) adhere to text condition \\(y\\) and image distribution \\(p_t(\\mathbf{x}_t|y)\\) keep consistency across views with different poses. For the image distribution \\(p_0(\\mathbf{x}_0|y)\\) of 2D diffusion model, the pose condition can be injected via input text with the corresponding directional prompt [37, 42]. As illustrated in Fig. 2(a), the pre-trained 2D image distribution, trained on individual images, is view-agnostic and lacks identity consistency across views. Even with a text tuning mechanism [1] specifically designed for multi-view image generation, as shown in Fig. 2(b), the above issues are far from resolved. Since SDS minimizes KL-divergence between the image distribution and rendering distribution independently for each rendered view, it can only inevitably inherit the 3D-awareness deficit of the 2D diffusion model, resulting in inconsistent 3D generation, which is commonly referred to as the Multi-face Janus Problem of SDS." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.396, + 0.331, + 0.412 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.788, + 0.565 + ], + "angle": 0, + "content": "In this section, we introduce JointDreamer, a novel text-to-3D generation framework as illustrated in Fig. 3. We first present the derivation of Joint Score Distillation (JSD) in Sec. 4.1, which extends the single-view optimization in SDS into a multi-view KL-Divergence. Then we integrate universal view-aware models into JSD to show the compatibility of JSD in Sec. 4.2, where we instantiate three kinds of view-aware models to capture inter-view coherence. Finally, we elaborate on the overall framework JointDreamer in Sec. 4.3, where we integrate the multi-view generation model into JSD. We also propose a geometry fading scheme and CFG switching strategy to further enhance generative quality." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.585, + 0.507, + 0.601 + ], + "angle": 0, + "content": "4.1 Joint Score Distillation (JSD)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.787, + 0.656 + ], + "angle": 0, + "content": "To address the multi-face problem arising from SDS, we extend the score distillation from single-view to multi-view settings and promote inter-view coherence across 2D image distribution, and thus derive our JSD optimization function." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.661, + 0.788, + 0.798 + ], + "angle": 0, + "content": "Coherence Modeling for Joint Image Distribution. As discussed above, the rendering distributions of the 3D representation should maintain 3D consistency across views \\(\\tilde{\\mathbf{x}} = \\{\\mathbf{x}^1,\\mathbf{x}^2,\\dots ,\\mathbf{x}^V\\}\\) with respect to different poses \\(\\tilde{\\mathbf{c}} = \\{c^{1},c^{2},\\ldots ,c^{V}\\}\\). However, for 2D pre-trained diffusion models, different views are generated independently. To ensure consistency, we propose modeling the joint image distribution of multiple views, denoted as \\(p_0(\\tilde{\\mathbf{x}} |\\tilde{\\mathbf{c}},y)\\), within the diffusion model. Following the commonly adopted assumption in energy-based distribution modeling [21, 44, 45], we introduce an energy function that measures the inter-view coherence by \\(\\mathcal{C}(\\tilde{\\mathbf{x}},\\tilde{\\mathbf{c}}):\\mathbb{R}^{Vd}\\to \\mathbb{R}\\) and define:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.808, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\np _ {0} (\\tilde {\\mathbf {x}} | \\tilde {\\mathbf {c}}, y) \\propto \\exp (\\mathcal {C} (\\tilde {\\mathbf {x}}, \\tilde {\\mathbf {c}})) \\prod_ {i = 1} ^ {V} p _ {0} (\\mathbf {x} ^ {i} | c ^ {i}, y), \\tag {3}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "JointDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "where a larger \\(\\mathcal{C}(\\tilde{\\mathbf{x}},\\tilde{\\mathbf{c}})\\) indicates greater coherence among the denoised view images. As a result, the joint image distribution is no longer view-independent. In practice, the joint energy function can be implemented via various view-aware models, as long as they can reflect the coherence across multiple views. In Sec. 4.2, we explore different choices in depth. The modeling of coherence across joint image distributions facilitates our JSD on multi-view, integrating inter-view coherence to ensure consistent 3D representation." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.261, + 0.79, + 0.29 + ], + "angle": 0, + "content": "KL-Divergence on Multiple Views. We extend the single-view KL-divergence in SDS to a multi-view version, based on the joint image distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.301, + 0.785, + 0.361 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\theta} D _ {K L} \\left(q _ {t} ^ {\\theta} (\\tilde {\\mathbf {x}} | \\tilde {\\mathbf {c}}, y) \\| p _ {t} (\\tilde {\\mathbf {x}} | \\tilde {\\mathbf {c}}, y)\\right) \\\\ = \\min _ {\\theta} D _ {K L} \\left(q _ {t} ^ {\\theta} (\\tilde {\\mathbf {x}} | \\tilde {\\mathbf {c}}, y) \\right\\lVert \\exp \\left(\\mathcal {C} \\left(\\tilde {\\mathbf {x}}, \\tilde {\\mathbf {c}}\\right)\\right) \\prod_ {i = 1} ^ {V} p _ {t} \\left(\\mathbf {x} ^ {i} \\mid c ^ {i}, y\\right), \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.371, + 0.784, + 0.432 + ], + "angle": 0, + "content": "where the extra energy term \\(\\mathcal{C}(\\tilde{\\mathbf{x}},\\tilde{\\mathbf{c}})\\) in Eq. (3) accounts for the inter-view coherence. Without this constraint, e.g., \\(\\mathcal{C}(\\tilde{\\mathbf{x}},\\tilde{\\mathbf{c}})\\equiv 0\\), different rendering views are optimized independently with the 2D diffusion model separately. In this sense, the original SDS can be seen as a special case of JSD." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.439, + 0.784, + 0.484 + ], + "angle": 0, + "content": "Joint Score Distillation Function. To correspond to the gradient of the new rule of multi-view KL-divergence in Eq. (4), we derive our score distillation function that is jointly conducted on multiple views as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.337, + 0.495, + 0.785, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\theta} L _ {J S D} (\\theta) \\\\ \\triangleq \\mathrm {E} _ {t, \\epsilon_ {\\Phi}} [ w (t) \\mathbb {E} (\\nabla_ {\\tilde {\\mathbf {x}}} \\log q _ {t} (\\tilde {\\mathbf {x}} _ {t} | \\tilde {\\mathbf {x}} _ {0}) - \\nabla_ {\\tilde {\\mathbf {x}}} \\log p _ {t} (\\tilde {\\mathbf {x}} _ {t} | y)) ] \\tag {5} \\\\ = \\sum_ {i = 1} ^ {V} \\mathrm {E} _ {t, \\epsilon_ {\\Phi} ^ {i}} [ w (t) (\\hat {\\epsilon} _ {\\Phi} (\\mathbf {x} _ {t} ^ {i}, y) - \\frac {\\partial \\mathcal {C} (\\tilde {\\mathbf {x}})}{\\partial \\mathbf {x} _ {t} ^ {i}} - \\epsilon^ {i}) \\frac {\\delta g (\\theta , c ^ {i})}{\\delta \\theta} ], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.58, + 0.788, + 0.747 + ], + "angle": 0, + "content": "where \\(\\{\\epsilon^i\\}_{i = 1}^V\\) are noises during score matching for different views. The proof can be found in the Appendix. To intuitively compare SDS with JSD, we sample multi-view images from the forward pass of pre-trained diffusion model [38] as shown in Fig. 2(c). For each view \\(\\mathbf{x}_t\\), we randomly select \\(\\mathbf{x}_t^\\prime\\) from a different view and adopt the binary classification model presented in Sec. 4.2 as the energy function to measure coherence across the two views. The results illustrate that JSD significantly enhances the correspondence with directional prompts and consistency across different views, particularly for the initially biased views such as side and back views. These theoretical and empirical results prove that the multi-face Janus issue in SDS is rooted in the view-agnostic and view-biased 2D image distribution, a challenge effectively mitigated by JSD." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.77, + 0.676, + 0.785 + ], + "angle": 0, + "content": "4.2 Universal View-Aware Models as Energy Function" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "JSD requires an energy function \\(\\mathcal{C}(\\tilde{\\mathbf{x}},\\tilde{\\mathbf{c}})\\) to measure coherence across denoised images, as presented in Eq. (3). This energy function plays a crucial role in assessing the consistency between different views in image distribution. To demonstrate" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "C. Jiang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.396, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.144, + 0.782, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Fig. 4: Illustration of the binary classification model and qualitative results with JSD. (a) The classification model \\( M_{\\mathrm{CLS}} \\) produces the binary logit to measure the consistency between two input views \\( x^i \\) and \\( x^j \\). (b) JSD integrated with the classification model effectively alleviates Janus issues compared to SDS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.785, + 0.472 + ], + "angle": 0, + "content": "the compatibility of JSD, we employ three different types of models trained for various representative multi-view tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.789, + 0.655 + ], + "angle": 0, + "content": "Binary classification model \\( M_{\\mathrm{CLS}} \\). The binary classification model \\( M_{\\mathrm{CLS}} \\) is designed to classify the content consistency between two input images based on their relative camera pose. To ensure computational efficiency, we introduce a dedicated classification model, as shown in Fig. 4 (a). The training process takes two days with a single A800 GPU on Objaverse dataset [8]. Further details can be found in the appendix. The classification model \\( M_{\\mathrm{CLS}} \\) processes pairs of images \\( \\mathbf{x}^i \\) and \\( \\mathbf{x}^j \\) captured from different viewpoints \\( c^i \\) and \\( c^j \\). It extracts image features using the DINO-ViT/s16 backbone [3]. These image features are conditioned on the camera feature obtained through MLP layers from the relative camera transformation matrix \\( \\Delta(c^j, c^i) \\). Finally, the classification head produces the binary score. To incorporate with JSD, we only consider neighboring views in \\( V \\) as image pairs for coherence measurement, which is denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.362, + 0.665, + 0.787, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {\\mathrm {C L S}} \\left(\\tilde {\\mathbf {x}}, \\tilde {\\mathbf {c}}\\right) = \\sum_ {i, j \\in 1, \\dots , V; i \\neq j} M _ {\\mathrm {C L S}} \\left(\\mathbf {x} _ {t} ^ {i}, \\mathbf {x} _ {t} ^ {j}, \\Delta \\left(c ^ {j}, c ^ {i}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.7, + 0.719 + ], + "angle": 0, + "content": "where the higher logit indicates the stronger geometric consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Image-to-image translation model \\( M_{\\mathrm{I2I}} \\). The image-to-image translation model \\( M_{\\mathrm{I2I}} \\) is tailored for novel view synthesis [26, 27, 29, 40]. We employ the most recent model, Wonder3D [29] as \\( M_{\\mathrm{I2I}} \\), which is a viewpoint-conditioned image translation model and generates consistent content in the target viewpoint. When integrated with JSD, a random reference view \\( \\mathbf{x}^{\\mathrm{ref}} \\) is selected from the set of 3D rendered images. Then we input the relative camera transformation \\( \\Delta (c^{i}, c^{\\mathrm{ref}}) \\) and rendered images to \\( M_{\\mathrm{I2I}} \\). The measure of consistency is determined by calculating the reconstruction loss between the synthesized new" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "JointDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.142, + 0.788, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.485, + 0.777, + 0.5 + ], + "angle": 0, + "content": "Fig. 5: Comparison of text-to-3D generation. See Appendix for more results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.54, + 0.548 + ], + "angle": 0, + "content": "image and its corresponding rendered image:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.559, + 0.787, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {\\mathrm {I 2 I}} (\\tilde {\\mathbf {x}}, \\tilde {\\mathbf {c}}) = - \\sum_ {i \\in 1, \\dots , V} | | M _ {\\mathrm {I 2 I}} (\\mathbf {x} _ {t} ^ {\\mathbf {r e f}}, \\Delta (c ^ {i}, c ^ {\\mathbf {r e f}})) - \\mathbf {x} _ {t} ^ {i} | | _ {2} ^ {2}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.603, + 0.785, + 0.633 + ], + "angle": 0, + "content": "where a smaller reconstruction loss indicates stronger geometric consistency under the estimation of \\( M_{\\mathrm{I2I}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.634, + 0.785, + 0.695 + ], + "angle": 0, + "content": "Multi-view synthesis model \\( M_{\\mathbf{MVS}} \\). The multi-view synthesis model is designed to generate multiple images conditioned on text prompts and camera poses, wherein we employ very recent work MVDream [41] as \\( M_{\\mathbf{MVS}} \\). We compute the reconstruction loss for multiple generative views and rendered views:" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.711, + 0.787, + 0.729 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {\\mathbf {M V S}} (\\tilde {\\mathbf {x}}, \\tilde {\\mathbf {c}}) = - | | M _ {\\mathbf {M V S}} (y, \\tilde {\\mathbf {c}}) - \\tilde {\\mathbf {x}} | | _ {2} ^ {2} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The smaller reconstruction loss signifies better geometric consistency within \\( M_{\\mathbf{MVS}} \\). These view-aware models measure the coherence across views according to their own 3D-aware insights. Incorporated with JSD, they provide distinct constraints, resulting in different 3D generations that nonetheless contribute to enhanced geometric consistency. We believe that JSD can be adapted to more universal view-aware models, which enables us to progressively redefine the benchmark of 3D generation with the advancement of multi-view tasks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "C. Jiang et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.144, + 0.642, + 0.153 + ], + "angle": 0, + "content": "a blue jay standing on a large basket of rainbow macarons" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.153, + 0.357, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.153, + 0.503, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.152, + 0.64, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.152, + 0.781, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.211, + 0.357, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.27, + 0.278, + 0.291, + 0.285 + ], + "angle": 0, + "content": "SDS" + }, + { + "type": "image_caption", + "bbox": [ + 0.357, + 0.211, + 0.704, + 0.22 + ], + "angle": 0, + "content": "A pink Spiderman dancing ballet, Marvel character HD, highly detailed 3D model" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.219, + 0.501, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.412, + 0.279, + 0.458, + 0.287 + ], + "angle": 0, + "content": "JSD w/ \\(C_{\\mathrm{CLS}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.219, + 0.64, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.279, + 0.6, + 0.287 + ], + "angle": 0, + "content": "JSD w/C12I" + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.219, + 0.782, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.279, + 0.734, + 0.287 + ], + "angle": 0, + "content": "JSD w/C_MVS" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.297, + 0.787, + 0.34 + ], + "angle": 0, + "content": "Fig. 6: Ablations on JSD incorporated with different energy function \\(\\mathcal{C}\\). CLS: Binary Classification model; I2I: Image-to-Image Translation model (Wonder3D [29]); MVS: Multi-View Image Synthesis model (MVDream [41])." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.369, + 0.496, + 0.384 + ], + "angle": 0, + "content": "4.3 Framework of JointDreamer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.789, + 0.604 + ], + "angle": 0, + "content": "Building upon the JSD optimization, we propose the overall framework JointDreamer. The optimization is based on neural radiance field (NeRF) [31], adopting Instant-NGP [32] with volume renderer. We adopt the multi-view synthesis model \\( M_{\\mathrm{MVS}} \\) as the energy function to integrate with JSD in JointDreamer. During optimization training, we utilize the common techniques including time-annealing and resolution scaling-up following [41, 42]. Besides, we propose two novel techniques to further enhance the generation quality, including a Geometry Fading scheme and a Classifier-Free Guidance Scale (CFG) Switching strategy. Geometry Fading. We aim to shift attention between geometric structure and texture details during optimization. Specifically, starting from iteration 5K, we reduce the learning rate of the density network of NeRF from \\( 1e - 2 \\) to \\( 1e - 6 \\) and set orientation loss to 0. Consequently, it benefits geometric convergence in the early phase of optimization, while allowing for decreased attention on geometry and increased attention on texture enhancement in the later stages." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.789, + 0.757 + ], + "angle": 0, + "content": "CFG Switching. CFG scheduling strategies have been employed in 2D domain [15,39] to enhance quality. In this work, we propose to modify the CFG scale \\(s\\) during the training for 3D generation. We are motivated by the observation that a large CFG scale can lead to accelerated geometric convergence but may result in under-optimized geometry and distorted texture. Unlike the annealed CFG approach in CLIP-Sculptor [39], our increasing strategy prioritizes texture while maintaining accurate geometry. Specifically, a smaller \\(s = 30\\) is employed in the early stages to preserve shape integrity, which allows for stronger coherence guidance from JSD. After 5K iterations, we increase the \\(s\\) to 50, enhancing texture fidelity and overall quality." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.779, + 0.367, + 0.796 + ], + "angle": 0, + "content": "5 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In this section, we present the text-to-3D generation results of JointDreamer with qualitative and quantitative evaluations, illustrating state-of-the-art per" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "JointDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "title", + "bbox": [ + 0.407, + 0.144, + 0.619, + 0.154 + ], + "angle": 0, + "content": "a rabbit cutting grass with a lawnmower" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.156, + 0.328, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.263, + 0.236, + 0.282, + 0.244 + ], + "angle": 0, + "content": "SDS" + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.153, + 0.479, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.236, + 0.422, + 0.244 + ], + "angle": 0, + "content": "original JSD" + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.153, + 0.631, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.236, + 0.589, + 0.245 + ], + "angle": 0, + "content": "+CFG Switching" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.153, + 0.785, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.236, + 0.746, + 0.245 + ], + "angle": 0, + "content": "+ Geo Fading" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.251, + 0.785, + 0.278 + ], + "angle": 0, + "content": "Fig. 7: Ablations on CFG Switching and Geometry Fading techniques in JointDreamer pipeline, demonstrating the effectiveness of quality enhancement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.309, + 0.784, + 0.339 + ], + "angle": 0, + "content": "performance. We also make further ablation analysis on the proposed JSD. More training and evaluation details can be found in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.361, + 0.451, + 0.374 + ], + "angle": 0, + "content": "5.1 Text-to-3D Generation" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.385, + 0.547, + 0.427 + ], + "angle": 0, + "content": "Table 1: Quantitative results on textual consistency and user preference, tested on object-centric subset of MS-COCO [25]." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.428, + 0.541, + 0.493 + ], + "angle": 0, + "content": "
MethodCLIP Score ↑R-Precision ↑User Study↑
DreamFusion [37]20.127.718.2
ProlificDreamer [42]25.018.716.2
MVDream [41]20.833.623.5
JointDreamer27.788.542.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.55, + 0.385, + 0.781, + 0.427 + ], + "angle": 0, + "content": "Table 2: Ablation study on CFG Switching (CFGS) and Geometry Fading (GF)." + }, + { + "type": "table", + "bbox": [ + 0.561, + 0.428, + 0.774, + 0.493 + ], + "angle": 0, + "content": "
SDS JSD CFGS GF|CLIP Score↑ FID↓
✓20.0 429.2
✓27.6 360.7
✓ ✓28.2 357.6
✓ ✓ ✓28.8 353.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Qualitative Comparisons. We compare with several representative baselines on threestudio [11] project, which modularizes the text-to-3D framework allowing for fair comparisons by ablating individual components. The generative samples are shown in Fig. 5. (i) DreamFusion [37]: Compared to DreamFusion which utilizes traditional SDS optimization, JointDreamer significantly improves the Multi-Face Janus issues in the generations of DreamFusion by introducing an energy term to ensure inter-view coherence. (ii) Magic3D [24]: Magic3D introduces a two-stage generation pipeline, transferring NeRF to DMTet in the second stage to enhance generation quality. We also transfer NeRF to DMTet as JointDreamer-DMTet, showcasing consistent superiority in geometry and texture quality by utilizing JSD optimization instead of SDS. (iii) ProlificDreamer [42]: ProlificDreamer presents variational score distillation (VSD) as a variant of score distillation function. While VSD enhances photorealism in 3D renderings by introducing a LoRA model, the ill-posed association of pose and images during LoRA training deepens the geometric inconsistency of 3D representation, resulting in severe multi-view artifacts in generations. In contrast, JointDreamer with JSD achieves geometric consistency while maintaining high-fidelity texture quality. (iv) MVDream [41]: Compared to direct distillation from a finetuned model as MVDream, JointDreamer employs view-aware models as the coherence constraint in JSD, while still inherit the generalization capabilities of original diffusion models [38]. The results indicate that JointDreamer mitigates the overfitting issues of MVDream, which accurately responds to complex input text and enhances the 3D consistency of generations." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "C. Jiang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.36 + ], + "angle": 0, + "content": "Quantitative Comparisons. Following [18,37], we evaluate CLIP Score [14], CLIP R-Precision [36] and user preference on the object-centric caption subset of MS-COCO [25] with 153 prompts to measure the text congruence and generative quality. For computational efficiency, we generate each 3D asset using 5K iterations of \\(64 \\times 64\\) rendered images and render 20 images per caption for evaluation. CLIP ViT-B/32 is adopted as the feature extractor for Clip Score and Clip R-Precision. As illustrated by the results in Table 1, JointDreamer can outperform all baselines on CLIP Score and CLIP R-Precision by large margins. Specifically, JointDreamer achieves an improvement of the R-Precision by \\(60.8\\%\\) and \\(54.9\\%\\) over DreamFusion and MVDream, demonstrating its superior corresponding to textual description. Notably, the severe Janus artifacts in ProlificDreamer compromise the quality of rendering with noisy background and semantic distortion, resulting in the lowest R-Precision. We also conduct a user study about shape preference on these prompts in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.384, + 0.414, + 0.399 + ], + "angle": 0, + "content": "5.2 Ablation Analyses" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.408, + 0.787, + 0.423 + ], + "angle": 0, + "content": "Ablations on Energy Functions. Sec. 4.2 discusses the varying interview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.423, + 0.516, + 0.572 + ], + "angle": 0, + "content": "coherency measurements provided by view-aware models trained. When incorporated with JSD, these models have distinct impacts on 3D generations, as shown in Fig. 6. The binary classifier effectively corrects inaccurate geometric structures in SDS. However, it cannot introduce additional imaginative elements as a discriminative model, resulting in oversaturated and monotonous textures. In con" + }, + { + "type": "table_caption", + "bbox": [ + 0.526, + 0.438, + 0.788, + 0.507 + ], + "angle": 0, + "content": "Table 3: Quantitative comparison of energy functions, showcasing the effectiveness of JSD in mitigating Janus artifacts with comparable computational efficiency to SDS." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.508, + 0.784, + 0.563 + ], + "angle": 0, + "content": "
MethodsJanus Rate ↓ GPU Memory ↓ Train Time ↓
SDS100%16.1 G50 min.
JSD w/CCLS12.5%22.1 G80 min.
JSD w/C2I31.2%16.0 G119 min.
JSD w/C_MVS6.2%19.4 G54 min.
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.573, + 0.788, + 0.725 + ], + "angle": 0, + "content": "trast, as generative models, Wonder3D [29] and MVDream [41] employ reconstruction loss to estimate 3D consistency. Hence, they not only guide geometric structural modifications but also influence texture quality. we further conduct a quantitative comparison using 16 complex multi-Janus prompts, as outlined in Table 3. The experimental setup details can be found in the Appendix. The results indicate that our JSD consistently mitigates Janus artifacts across different view-aware models, with only a slight increase in computational requirements compared to SDS. We find that the energy function \\(C_{12\\mathrm{I}}\\) derived from the imagedto-image translation model exhibits poor performance, likely attributed to a mismatch in camera range and inaccurate translated results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Analysis on Geometry Fading and CFG Switching Mechanisms. We conduct incremental ablations on our proposed techniques in JointDreamer, including the Geometry Fading scheme and CFG Switching strategy. As shown in Fig. 7, increasing the CFG value enhances texture detail compared to the original JSD, but over-optimizes the geometry, resulting in a bumpier shape. Compared to “+CFG Switching”, the Geometry fading effectively protects shape when larger CFG guidance. We also conduct a quantitative evaluation on a \\(30\\%\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "JointDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.785, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.3, + 0.785, + 0.329 + ], + "angle": 0, + "content": "Fig. 9: Discuss the impact of prompts and random seeds, demonstrating the robustness of JointDreamer. \\(/...\\) means the same content as the prompt in Fig. 7." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.784, + 0.401 + ], + "angle": 0, + "content": "MS-COCO subset in Table 2. JSD demonstrates superior texture quality compared to SDS, as evidenced by the Clip Score (CS) and FID metrics. The two proposed mechanisms further enhance the texture quality." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.407, + 0.785, + 0.422 + ], + "angle": 0, + "content": "Discussions on Training Loss. To make further comparisons with JSD" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.423, + 0.487, + 0.572 + ], + "angle": 0, + "content": "and SDS, we aggregate training losses from multiple prompts on two optimization functions and visualize the training loss curve as illustrated in Fig. 8. We observe that SDS experiences significant fluctuations due to the randomness of single-view optimization. In contrast, JSD can converge gradually and smoothly, demonstrating the introduction of multi" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.426, + 0.787, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.52, + 0.788, + 0.562 + ], + "angle": 0, + "content": "Fig.8: Training loss comparisons of JSD and SDS. JSD eliminates randomness fluctuations in SDS convergence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.784, + 0.604 + ], + "angle": 0, + "content": "view optimization with inter-view coherence in JSD can reduce the randomness of optimization and contribute to better convergence for 3D representation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.784, + 0.673 + ], + "angle": 0, + "content": "Discussions on Robustness for Prompts and Random Seed. To show the impact of seed and prompt, we modify seeds and key prompt components in Fig. 7, such as subject, object and verb. Results in Fig. 9 demonstrate the robustness of JointDreamer for different seed. Note that the default seed is 0." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.678, + 0.785, + 0.803 + ], + "angle": 0, + "content": "The Effectiveness of Classification Model. The classification model surpasses MVDream in training speed by a factor of 48. Fig.4 (b) provides additional high-quality results obtained using JSD in conjunction with the binary classification model \\(M_{\\mathrm{CLS}}\\). Furthermore, we conduct an ablation study by replacing \\(\\frac{\\partial\\mathcal{C}(\\hat{\\mathbf{x}})}{\\partial\\mathbf{x}_i^i}\\) in Eq.5 with a randomly generated value between [0,1]. However, it yields shapeless results due to unrelated disturbances in the optimization process. These findings highlight the significance of the proposed classification model in achieving a balance between computational efficiency and generation quality." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "More Comparison with MVDream. View-aware models can be incorporated into 3D generation by combining them with JSD or using them directly with" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "C. Jiang et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.778, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.27, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Fig. 10: Comparison of JointDreamer and SDS+MVDream, illustrating the superiority of JointDreamer in mitigating the Janus problem. SDS+MVDream consistently exhibits semantic missing or Janus issues with different weight options." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.784, + 0.412 + ], + "angle": 0, + "content": "SDS. In this discussion, we use MVDream [41] as an example to demonstrate the superiority of JSD. MVDream can generate consistent shapes when calculating the SDS loss directly. However, it may miss components in the complete input text due to its fine-tuning on limited 3D data, as illustrated in Fig. 5. To address this limitation, a straightforward approach is to \"SDS+MVDream\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.788, + 0.581 + ], + "angle": 0, + "content": "However, balancing the impact of SDS and MVDream is challenging. As shown in Fig. 10, setting \\(\\lambda_{\\mathrm{SDS}}\\) to 0 degrades to MVDream alone, while setting \\(\\lambda_{\\mathrm{MVDream}}\\) to 0 yields a combination similar to DreamFusion [37]. Achieving a balanced impact between SDS and MVDream through a simple combination is difficult. When \\(\\lambda_{\\mathrm{MVDream}}\\) is large, textual consistency remains constrained, whereas decreasing \\(\\lambda_{\\mathrm{MVDream}}\\) leads to the Janus problem. This is due to gradient misalignment across multiple views in the \"SDS+MVDream\" combination, as SDS lacks multi-view information and cannot derive our objective outlined in Eq. 5. In contrast, JSD based on the joint image distribution provides supervision for text consistency and high-fidelity texture. JSD also promotes inter-view consistency by introducing an energy term with the view-aware model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.607, + 0.36, + 0.623 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.803, + 0.734 + ], + "angle": 0, + "content": "In this work, we introduce Joint Score Distillation (JSD) as a new paradigm for text-to-3D generation, which conducts multi-view optimization jointly and accounts for inter-view coherence. We demonstrate that JSD can significantly enhance 3D coherence while maintaining generalizability. With other proposed techniques, our overall framework, JointDreamer, is capable of geometric-consistent and high-fidelity 3D generation adhering to complex text input." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Limitations. While the training time of JointDreamer is comparable to existing SDS pipelines, there is room for improvement in terms of acceleration. Future work will explore alternative 3D representations, such as 3D Gaussian [20]. Additionally, JSD utilizes view-aware models to ensure geometry consistency and mitigate the impact of limited data. However, view-aware models still require 3D data for training, and further efficient 3D data collection or reconstruction from multi-view images is also worth investigating." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "JointDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Armandpour, M., Zheng, H., Sadeghian, A., Sadeghian, A., Zhou, M.: Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.245 + ], + "angle": 0, + "content": "2. Cao, Z., Hong, F., Wu, T., Pan, L., Liu, Z.: Large-vocabulary 3d diffusion model with transformer. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.785, + 0.286 + ], + "angle": 0, + "content": "3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV. pp. 9650-9660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.328 + ], + "angle": 0, + "content": "4. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: CVPR. pp. 16123-16133 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.785, + 0.37 + ], + "angle": 0, + "content": "5. Chen, H., Gu, J., Chen, A., Tian, W., Tu, Z., Liu, L., Su, H.: Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction. arXiv preprint arXiv:2304.06714 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "6. Chen, R., Chen, Y., Jiao, N., Jia, K.: *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation*. In: ICCV. pp. 22246–22256 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.453 + ], + "angle": 0, + "content": "7. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. In: NeurIPS (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.454, + 0.787, + 0.495 + ], + "angle": 0, + "content": "8. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objverse: A universe of annotated 3d objects. In: CVPR. pp. 13142-13153 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.785, + 0.522 + ], + "angle": 0, + "content": "9. Deng, Y., Yang, J., Xiang, J., Tong, X.: Gram: Generative radiance manifolds for 3d-aware image generation. In: CVPR. pp. 10673-10683 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "10. Gao, J., Shen, T., Wang, Z., Chen, W., Yin, K., Li, D., Litany, O., Gojcic, Z., Fidler, S.: Get3d: A generative model of high quality 3d textured shapes learned from images. NeurIPS 35, 31841-31854 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.565, + 0.785, + 0.618 + ], + "angle": 0, + "content": "11. Guo, Y.C., Liu, Y.T., Shao, R., Laforte, C., Voleti, V., Luo, G., Chen, C.H., Zou, Z.X., Wang, C., Cao, Y.P., Zhang, S.H.: threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "12. Henderson, P., Ferrari, V.: Learning single-image 3d reconstruction by generative modelling of shape, pose and shading. IJCV 128(4), 835-854 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "13. Henderson, P., Tsiminaki, V., Lampert, C.H.: Leveraging 2d data to learn textured 3d mesh generation. In: CVPR. pp. 7498-7507 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "14. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: CLIPScore: a reference-free evaluation metric for image captioning. In: EMNLP (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.703, + 0.785, + 0.743 + ], + "angle": 0, + "content": "15. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "16. Hu, Z., Zhao, M., Zhao, C., Liang, X., Li, L., Zhao, Z., Fan, C., Zhou, X., Yu, X.: Efficientdreamer: High-fidelity and robust 3d creation via orthogonal-view diffusion priors. In: CVPR. pp. 4949-4958 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "17. Huang, Y., Wang, J., Shi, Y., Tang, B., Qi, X., Zhang, L.: Dreamtime: An improved optimization strategy for diffusion-guided 3d generation. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "18. Jain, A., Mildenhall, B., Barron, J.T., Abbeel, P., Poole, B.: Zero-shot text-guided object generation with dream fields. In: CVPR. pp. 867-876 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.177, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "C. Jiang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "19. Jun, H., Nichol, A.: Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.217 + ], + "angle": 0, + "content": "20. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4), 1-14 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.785, + 0.243 + ], + "angle": 0, + "content": "21. LeCun, Y., Chopra, S., Hadsell, R., Ranzato, M., Huang, F.: A tutorial on energy-based learning. Predicting structured data 1(0) (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.785, + 0.27 + ], + "angle": 0, + "content": "22. Li, M., Zhou, P., Liu, J.W., Keppo, J., Lin, M., Yan, S., Xu, X.: Instant3d: Instant text-to-3d generation. arXiv preprint arXiv:2311.08403 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.27, + 0.785, + 0.297 + ], + "angle": 0, + "content": "23. Li, W., Chen, R., Chen, X., Tan, P.: Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.297, + 0.785, + 0.337 + ], + "angle": 0, + "content": "24. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.337, + 0.785, + 0.393 + ], + "angle": 0, + "content": "25. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.393, + 0.785, + 0.42 + ], + "angle": 0, + "content": "26. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: ICCV. pp. 9298-9309 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.42, + 0.785, + 0.447 + ], + "angle": 0, + "content": "27. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.447, + 0.785, + 0.473 + ], + "angle": 0, + "content": "28. Liu, Z., Feng, Y., Black, M.J., Nowrouzezahrai, D., Paull, L., Liu, W.: Meshdiffusion: Score-based generative 3d mesh modeling. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "29. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3d: Single image to 3d using cross-domain diffusion. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.514, + 0.785, + 0.554 + ], + "angle": 0, + "content": "30. Luo, W., Hu, T., Zhang, S., Sun, J., Li, Z., Zhang, Z.: Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. NeurIPS 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.554, + 0.785, + 0.596 + ], + "angle": 0, + "content": "31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.596, + 0.785, + 0.636 + ], + "angle": 0, + "content": "32. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.636, + 0.785, + 0.678 + ], + "angle": 0, + "content": "33. Nguyen-Phuoc, T., Li, C., Theis, L., Richardt, C., Yang, Y.L.: Hologan: Unsupervised learning of 3d representations from natural images. In: ICCV. pp. 7588-7597 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "34. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.718, + 0.785, + 0.746 + ], + "angle": 0, + "content": "35. Niemeyer, M., Geiger, A.: Giraffe: Representing scenes as compositional generative neural feature fields. In: CVPR. pp. 11453-11464 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.746, + 0.785, + 0.786 + ], + "angle": 0, + "content": "36. Park, D.H., Azadi, S., Liu, X., Darrell, T., Rohrbach, A.: Benchmark for compositional text-to-image synthesis. In: NeurIPS Datasets and Benchmarks Track (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "37. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.813, + 0.785, + 0.841 + ], + "angle": 0, + "content": "38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "JointDreamer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "39. Sanghi, A., Fu, R., Liu, V., Willis, K.D., Shayani, H., Khasahmadi, A.H., Sridhar, S., Ritchie, D.: Clip-sculptor: Zero-shot generation of high-fidelity and diverse shapes from natural language. In: CVPR. pp. 18339-18348 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "40. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.259 + ], + "angle": 0, + "content": "41. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.3 + ], + "angle": 0, + "content": "42. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In: NeurIPS (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.328 + ], + "angle": 0, + "content": "43. Watson, D., Chan, W., Martin-Brualla, R., Ho, J., Tagliasacchi, A., Norouzi, M.: Novel view synthesis with diffusion models. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.384 + ], + "angle": 0, + "content": "44. Weese, J., Kaus, M., Lorenz, C., Lobregt, S., Truyen, R., Pekar, V.: Shape constrained deformable models for 3d medical image segmentation. In: Information Processing in Medical Imaging: 17th International Conference, IPMI 2001 Davis, CA, USA, June 18-22, 2001 Proceedings 17. pp. 380-387. Springer (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.425 + ], + "angle": 0, + "content": "45. Zhao, W., Yan, L., Zhang, Y.: Geometric-constrained multi-view image matching method based on semi-global optimization. Geo-spatial information science 21(2), 115-126 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.425 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/JointDreamer_ Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation/4667ae70-c38c-469d-a697-9d92865631f2_origin.pdf b/2024/JointDreamer_ Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation/4667ae70-c38c-469d-a697-9d92865631f2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ecb88e424a680f8d007b3c795443e884f74b0245 --- /dev/null +++ b/2024/JointDreamer_ Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation/4667ae70-c38c-469d-a697-9d92865631f2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce724227956e18e8912c48e3d2c7de3359223b1e84fc25bd53963930127a463b +size 9374026 diff --git a/2024/JointDreamer_ Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation/full.md b/2024/JointDreamer_ Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e835c542cf88e953199e1f752a8115a3024b0aa3 --- /dev/null +++ b/2024/JointDreamer_ Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation/full.md @@ -0,0 +1,311 @@ +# JointDreamer: Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation + +Chenhan Jiang $^{1*†}$ , Yihan Zeng $^{2*}$ , Tianyang Hu $^{2}$ , Songcun Xu $^{2}$ , Wei Zhang $^{2}$ , Hang Xu $^{2}$ , and Dit-Yan Yeung $^{1}$ + +1 The Hong Kong University of Science and Technology + +2 Huawei Noah's Ark Lab + +*Equal contribution. †Corresponding Author: jchcyan@gmail.com + +https://jointdreamer.github.io + +Abstract. Score Distillation Sampling (SDS) by well-trained 2D diffusion models has shown great promise in text-to-3D generation. However, this paradigm distills view-agnostic 2D image distributions into the rendering distribution of 3D representation for each view independently, overlooking the coherence across views and yielding 3D inconsistency in generations. In this work, we propose Joint Score Distillation (JSD), a new paradigm that ensures coherent 3D generations. Specifically, we model the joint image distribution, which introduces an energy function to capture the coherence among denoised images from the diffusion model. We then derive the joint score distillation on multiple rendered views of the 3D representation, as opposed to a single view in SDS. In addition, we instantiate three universal view-aware models as energy functions, demonstrating compatibility with JSD. Empirically, JSD significantly mitigates the 3D inconsistency problem in SDS, while maintaining text congruence. Moreover, we introduce the Geometry Fading scheme and Classifier-Free Guidance (CFG) Switching strategy to enhance generative details. Our framework, JointDreamer, establishes a new benchmark in text-to-3D generation, achieving outstanding results with an $88.5\%$ CLIP R-Precision and $27.7\%$ CLIP Score. These metrics demonstrate exceptional text congruence, as well as remarkable geometric consistency and texture fidelity. + +Keywords: 3D Vision $\cdot$ 3D Generation $\cdot$ Energy Function + +# 1 Introduction + +3D content creation is essential for diverse applications, including gaming, robotics simulation, and virtual reality. However, it is labor-intensive, demanding substantial time for skilled designers to create a single 3D asset. Hence, automating 3D creation with text input has attracted considerable attention. Recently, the score distillation sampling (SDS) algorithm pioneered by DreamFusion [37] + +![](images/3f3a1bb770864999f37a8549ab0e64c86a81933d0589418a70a502ab4eda53ce.jpg) +Fig. 1: Text-to-3D generations by JointDreamer from scratch. JointDreamer excels in generating geometrically consistent and high-fidelity 3D assets, adhering to complex textual descriptions that are challenging for previous methods. + +shows promise in text-to-3D tasks, which lifts image distribution from a well-trained diffusion model [38] into parameterized 3D representation like NeRF [31]. Compared to 3D generative models [2, 19, 34] that struggle with producing arbitrary objects due to limited text-3D training data, SDS-based methods [6, 24, 37, 42] can generate arbitrary 3D assets with diverse text input. + +Although SDS-based methods benefit from the generalizability of diffusion models, they often encounter a common issue known as Janus artifacts [6, 24, 37, 42]. These artifacts manifest as repeated content from different viewpoints of a 3D generation, yielding a lack of realism and coherence in the rendered views. We investigate the Janus artifacts by visualizing image distributions of the diffusion model [38] from multiple viewpoints, as illustrated in Fig. 2(a). The results reveal the view-agnostic nature and the content inconsistency across views of the diffusion model. Consequently, SDS optimizes each rendered view of 3D + +representation independently and inherits image distribution without coherent multi-view perspective, leading to the geometric inconsistency of 3D generations. + +Existing works [1,22] address the aforementioned challenges within the SDS framework by employing prompt engineering techniques. However, the effectiveness of such methods remains unsatisfactory, as evident from the results depicted in Fig. 2(b). Alternative methods [16,41] propose to finetune view-aware diffusion models using rendered images of 3D datasets [7,8]. Nevertheless, they are prone to overfitting on limited text-3D training data [23], decreasing the text congruence when handling complex text inputs. Based on the above observations, it requires a rethinking of the SDS optimization to enhance the 3D coherence of generations while maintaining generalizability. + +In this work, we first present Joint Score Distillation (JSD), which significantly promotes the 3D consistency of generation and inherits generalizability from diffusion models. Specifically, we model the joint image distribution of diffusion model via an energy function measuring coherence across denoised images. It facilitates the extension of the KL-Divergence in SDS from single-view into multi-view. We then derive the joint score distillation function from multi-view KL-Divergence, which ensures inter-view coherence in the optimization process of 3D generation. We show that SDS is a special case of JSD with the energy term omitted, which indicates the absence of coherence constraint across views. + +Building upon JSD optimization, we present three view-aware models as energy terms to showcase the compatibility of JSD: the Binary Classification Model, the Image-to-Image Translation Model, and the Multi-view Generation Model. Through empirical analysis, it is observed that different view-aware models introduce distinct coherence measurements, + +leading to diverse 3D generations, while all contributing to 3D consistency. Furthermore, to facilitate a more comprehensive comparison with existing text-to-3D generation methods, we introduce JointDreamer, an innovative framework capable of producing geometric-consistent and high-fidelity 3D assets adhering to complex text descriptions. Notably, in addition to incorporating a Multi-view Generation model as an energy term in JSD, we introduce two complementary + +![](images/60635b6e1dfbdb5e4b443c7cb208dff487674074281f21709a1c38f9b1835a20.jpg) +Fig.2: Illustration of text-conditioned images for different viewpoints, where input texts are augmented with corresponding direction prompts for each view. (a) The original generations from 2D diffusion model [38] are view-agnostic and inconsistent across views. (b) Text prompt tuning [1] has limited improvement in the directional structure of generated images for each view. (c) JSD injects coherence measurement from the proposed binary classifier (refer to Section 4.2), contributing to modified directional structures and semantical consistency across views. + +techniques, namely the Geometry Fading scheme and the Classifier-Free Guidance (CFG) Switching strategy, to enhance generative details. + +We systematically assess the quality of our approach, both qualitatively and quantitatively, compared to existing methods. Qualitative results gallery can be found in Fig. 1. Our JointDreamer consistently produces high-fidelity 3D assets and mitigates Janus artifacts in SDS. It maintains text congruence even when confronted with complex text input, achieving $88.5\%$ CLIP R-Precision and $27.7\%$ CLIP Score. + +In brief, our contributions are summarized as follows: + +- We introduce a novel Joint Score Distillation (JSD) for text-to-3D generation, optimizing multiple views jointly via an energy function to capture inter-view coherence. +- We present three view-aware models as energy functions to show compatibility with JSD, all of them mitigate the Janus problem in SDS. +- We introduce the text-to-3D framework JointDreamer, incorporating complementary Geometry Fading and CFG Switching techniques. Our JointDreamer achieves geometrically consistent and high-fidelity 3D assets even with complex textual inputs. + +# 2 Related Works + +Text-to-3D Generation. Existing text-to-3D generation methods can be categorized into two streams: 3D generative models and 2D optimization methods. The former encompasses various deep generative models such as Variational Auto Encoders (VAEs) [12, 13], Generative Adversarial Models (GAN) [4, 9, 10, 33, 35], diffusion models [5, 28, 34] and transformer architectures [2, 19]. These models are efficient in inference but often struggle with generalizability and training stability, attributed to the limited scope and complexity of available 3D datasets. The latter approach centers around the Score Distillation Sampling (SDS) algorithm proposed by [37], which leverages 2D diffusion model priors [38] for optimizing 3D representations. Subsequent advancements have refined this technique by improving the 3D representations [6, 24], the sampling scheduler [17] and loss design [42]. However, the above approaches overlook the geometric consistency problem, facing inherent multi-face Janus issues in SDS. Prior works try to alleviate the Janus issues with prompt tuning [1] yet they achieve limited effect. Very recent work MVDream [41] addresses the problem by fine-tuning a multi-view diffusion model, but it is susceptible to overfitting on scarce 3D training data, compromising semantic consistency in text-to-3D generations. In this work, we address the fundamental flaw of SDS that optimizes each view independently by introducing a joint optimization function that enforces inter-view consistency, essentially solving the Janus issues in SDS while preserving its generalizability. + +Diffusion-based Novel View Synthesis. As an alternative to 3D generation, novel view synthesis models the challenge as a view-conditioned image-to-image translation task. There have been proposals for pose-conditioned image-to-image + +![](images/a09b64fcd74f2b86786d190bb558a9ecae1f0da3bd57f94f32b1fcb31dca274a.jpg) +Fig. 3: Overview of JointDreamer Framework. We introduce an energy function to model the joint distribution for multi-view denoised images from 2D diffusion model, facilitating the Joint Score Distillation (JSD) optimization for text-to-3D generation. + +diffusion models [43] that generate novel views on synthetic data in 3D. Recently, [26] promoted the generalizability of novel view synthesis by fine-tuning the 2D diffusion model [38] on renderings of 3D dataset, which facilitates images-to-3D tasks with 3D reconstruction or SDS algorithm. To further improve the 3D consistency across generated views and input view, very recent works [27, 29, 40] modify the generation process into multi-view generation and present corresponding architecture designs. Generally, these methods take camera specifications as conditions and enable the viewpoint-aware generations, but they can hardly accurately capture a complete 3D scene consistently and densely. Our method acknowledges the potential of these models to discern relative interview relationships, which we harness to provide inter-view coherence for our JSD, thus serving as universal guidance models. + +# 3 Preliminaries + +We review the original SDS and address its fundamental limitation: optimizing independently for each single view of 3D representation and distilling with the view-agnostic image distribution of diffusion model. It results in geometric inconsistency issues, dubbed as Multi-Face Janus Problem. + +Score Distillation Sampling (SDS). SDS optimization is widely adopted by text-to-3D generation pipelines [6,24,30,37,42]. Given a 3D representation with learnable parameters $\theta$ and a pre-trained 2D diffusion model with noise prediction network $\epsilon_{\Phi}(\mathbf{x}_t,t,y)$ , SDS optimizes $\theta$ by minimizing the KL-divergence: + +$$ +\min _ {\theta} D _ {K L} \left(q _ {t} ^ {\theta} \left(\mathbf {x} _ {t} | c, y\right) \mid \mid p _ {t} \left(\mathbf {x} _ {t} | y\right)\right). \tag {1} +$$ + +Here, $p_t(\mathbf{x}_t | y)$ is the image distribution sampled from diffusion model, $q_t^\theta(\mathbf{x}_t | c, y)$ is the distribution of rendered image $\mathbf{x}_t = g(\theta, c)$ with respect to camera pose $c$ at timestep $t$ of the forward diffusion process, where $g$ is the renderer. To solve Eq. (1), the score distillation function is derived as: + +$$ +\begin{array}{l} \nabla_ {\theta} L _ {S D S} (\theta) \triangleq \mathrm {E} _ {t, \mathbf {x}} [ w (t) \frac {\sigma_ {t}}{\alpha_ {t}} \nabla_ {\theta} K L \left(q _ {t} ^ {\theta} \left(\mathbf {x} _ {t} | c, y\right) \mid \mid p _ {t} \left(\mathbf {x} _ {t} | y\right)\right) ] \\ \triangleq \mathrm {E} _ {t, \epsilon_ {\Phi}} [ w (t) (\hat {\epsilon} _ {\Phi} (\mathbf {x} _ {t}, t, y) - \epsilon) \frac {\delta g (\theta , c)}{\delta \theta} ], \tag {2} \\ \end{array} +$$ + +where $\hat{\epsilon}_{\Phi} := (1 + s)\epsilon_{\Phi}(\mathbf{x}_t, t, y) - s\epsilon_{\Phi}(\mathbf{x}_t, t, \emptyset)$ is modification of predicted noise with classifier-free guidance (CFG) $s$ , $w(t)$ is time-dependent weighting function. Multi-Face Janus Problem. To achieve consistency, it is essential that the rendering distributions $q_0^\theta(\mathbf{x}_0|c, y)$ adhere to text condition $y$ and image distribution $p_t(\mathbf{x}_t|y)$ keep consistency across views with different poses. For the image distribution $p_0(\mathbf{x}_0|y)$ of 2D diffusion model, the pose condition can be injected via input text with the corresponding directional prompt [37, 42]. As illustrated in Fig. 2(a), the pre-trained 2D image distribution, trained on individual images, is view-agnostic and lacks identity consistency across views. Even with a text tuning mechanism [1] specifically designed for multi-view image generation, as shown in Fig. 2(b), the above issues are far from resolved. Since SDS minimizes KL-divergence between the image distribution and rendering distribution independently for each rendered view, it can only inevitably inherit the 3D-awareness deficit of the 2D diffusion model, resulting in inconsistent 3D generation, which is commonly referred to as the Multi-face Janus Problem of SDS. + +# 4 Method + +In this section, we introduce JointDreamer, a novel text-to-3D generation framework as illustrated in Fig. 3. We first present the derivation of Joint Score Distillation (JSD) in Sec. 4.1, which extends the single-view optimization in SDS into a multi-view KL-Divergence. Then we integrate universal view-aware models into JSD to show the compatibility of JSD in Sec. 4.2, where we instantiate three kinds of view-aware models to capture inter-view coherence. Finally, we elaborate on the overall framework JointDreamer in Sec. 4.3, where we integrate the multi-view generation model into JSD. We also propose a geometry fading scheme and CFG switching strategy to further enhance generative quality. + +# 4.1 Joint Score Distillation (JSD) + +To address the multi-face problem arising from SDS, we extend the score distillation from single-view to multi-view settings and promote inter-view coherence across 2D image distribution, and thus derive our JSD optimization function. + +Coherence Modeling for Joint Image Distribution. As discussed above, the rendering distributions of the 3D representation should maintain 3D consistency across views $\tilde{\mathbf{x}} = \{\mathbf{x}^1,\mathbf{x}^2,\dots ,\mathbf{x}^V\}$ with respect to different poses $\tilde{\mathbf{c}} = \{c^{1},c^{2},\ldots ,c^{V}\}$ . However, for 2D pre-trained diffusion models, different views are generated independently. To ensure consistency, we propose modeling the joint image distribution of multiple views, denoted as $p_0(\tilde{\mathbf{x}} |\tilde{\mathbf{c}},y)$ , within the diffusion model. Following the commonly adopted assumption in energy-based distribution modeling [21, 44, 45], we introduce an energy function that measures the inter-view coherence by $\mathcal{C}(\tilde{\mathbf{x}},\tilde{\mathbf{c}}):\mathbb{R}^{Vd}\to \mathbb{R}$ and define: + +$$ +p _ {0} (\tilde {\mathbf {x}} | \tilde {\mathbf {c}}, y) \propto \exp (\mathcal {C} (\tilde {\mathbf {x}}, \tilde {\mathbf {c}})) \prod_ {i = 1} ^ {V} p _ {0} (\mathbf {x} ^ {i} | c ^ {i}, y), \tag {3} +$$ + +where a larger $\mathcal{C}(\tilde{\mathbf{x}},\tilde{\mathbf{c}})$ indicates greater coherence among the denoised view images. As a result, the joint image distribution is no longer view-independent. In practice, the joint energy function can be implemented via various view-aware models, as long as they can reflect the coherence across multiple views. In Sec. 4.2, we explore different choices in depth. The modeling of coherence across joint image distributions facilitates our JSD on multi-view, integrating inter-view coherence to ensure consistent 3D representation. + +KL-Divergence on Multiple Views. We extend the single-view KL-divergence in SDS to a multi-view version, based on the joint image distribution: + +$$ +\begin{array}{l} \min _ {\theta} D _ {K L} \left(q _ {t} ^ {\theta} (\tilde {\mathbf {x}} | \tilde {\mathbf {c}}, y) \| p _ {t} (\tilde {\mathbf {x}} | \tilde {\mathbf {c}}, y)\right) \\ = \min _ {\theta} D _ {K L} \left(q _ {t} ^ {\theta} (\tilde {\mathbf {x}} | \tilde {\mathbf {c}}, y) \right\lVert \exp \left(\mathcal {C} \left(\tilde {\mathbf {x}}, \tilde {\mathbf {c}}\right)\right) \prod_ {i = 1} ^ {V} p _ {t} \left(\mathbf {x} ^ {i} \mid c ^ {i}, y\right), \tag {4} \\ \end{array} +$$ + +where the extra energy term $\mathcal{C}(\tilde{\mathbf{x}},\tilde{\mathbf{c}})$ in Eq. (3) accounts for the inter-view coherence. Without this constraint, e.g., $\mathcal{C}(\tilde{\mathbf{x}},\tilde{\mathbf{c}})\equiv 0$ , different rendering views are optimized independently with the 2D diffusion model separately. In this sense, the original SDS can be seen as a special case of JSD. + +Joint Score Distillation Function. To correspond to the gradient of the new rule of multi-view KL-divergence in Eq. (4), we derive our score distillation function that is jointly conducted on multiple views as follows: + +$$ +\begin{array}{l} \nabla_ {\theta} L _ {J S D} (\theta) \\ \triangleq \mathrm {E} _ {t, \epsilon_ {\Phi}} [ w (t) \mathbb {E} (\nabla_ {\tilde {\mathbf {x}}} \log q _ {t} (\tilde {\mathbf {x}} _ {t} | \tilde {\mathbf {x}} _ {0}) - \nabla_ {\tilde {\mathbf {x}}} \log p _ {t} (\tilde {\mathbf {x}} _ {t} | y)) ] \tag {5} \\ = \sum_ {i = 1} ^ {V} \mathrm {E} _ {t, \epsilon_ {\Phi} ^ {i}} [ w (t) (\hat {\epsilon} _ {\Phi} (\mathbf {x} _ {t} ^ {i}, y) - \frac {\partial \mathcal {C} (\tilde {\mathbf {x}})}{\partial \mathbf {x} _ {t} ^ {i}} - \epsilon^ {i}) \frac {\delta g (\theta , c ^ {i})}{\delta \theta} ], \\ \end{array} +$$ + +where $\{\epsilon^i\}_{i = 1}^V$ are noises during score matching for different views. The proof can be found in the Appendix. To intuitively compare SDS with JSD, we sample multi-view images from the forward pass of pre-trained diffusion model [38] as shown in Fig. 2(c). For each view $\mathbf{x}_t$ , we randomly select $\mathbf{x}_t^\prime$ from a different view and adopt the binary classification model presented in Sec. 4.2 as the energy function to measure coherence across the two views. The results illustrate that JSD significantly enhances the correspondence with directional prompts and consistency across different views, particularly for the initially biased views such as side and back views. These theoretical and empirical results prove that the multi-face Janus issue in SDS is rooted in the view-agnostic and view-biased 2D image distribution, a challenge effectively mitigated by JSD. + +# 4.2 Universal View-Aware Models as Energy Function + +JSD requires an energy function $\mathcal{C}(\tilde{\mathbf{x}},\tilde{\mathbf{c}})$ to measure coherence across denoised images, as presented in Eq. (3). This energy function plays a crucial role in assessing the consistency between different views in image distribution. To demonstrate + +![](images/e09f1c9ad9fd4190ae30fb8f1027fa4246413df1e8eb1bcb7d934abc5ec3222e.jpg) +Fig. 4: Illustration of the binary classification model and qualitative results with JSD. (a) The classification model $M_{\mathrm{CLS}}$ produces the binary logit to measure the consistency between two input views $x^i$ and $x^j$ . (b) JSD integrated with the classification model effectively alleviates Janus issues compared to SDS. + +![](images/8310201a10e7cba4b295123a8f7c531556b987951d0cd38f7015d09ecab44872.jpg) + +the compatibility of JSD, we employ three different types of models trained for various representative multi-view tasks. + +Binary classification model $M_{\mathrm{CLS}}$ . The binary classification model $M_{\mathrm{CLS}}$ is designed to classify the content consistency between two input images based on their relative camera pose. To ensure computational efficiency, we introduce a dedicated classification model, as shown in Fig. 4 (a). The training process takes two days with a single A800 GPU on Objaverse dataset [8]. Further details can be found in the appendix. The classification model $M_{\mathrm{CLS}}$ processes pairs of images $\mathbf{x}^i$ and $\mathbf{x}^j$ captured from different viewpoints $c^i$ and $c^j$ . It extracts image features using the DINO-ViT/s16 backbone [3]. These image features are conditioned on the camera feature obtained through MLP layers from the relative camera transformation matrix $\Delta(c^j, c^i)$ . Finally, the classification head produces the binary score. To incorporate with JSD, we only consider neighboring views in $V$ as image pairs for coherence measurement, which is denoted as: + +$$ +\mathcal {C} _ {\mathrm {C L S}} \left(\tilde {\mathbf {x}}, \tilde {\mathbf {c}}\right) = \sum_ {i, j \in 1, \dots , V; i \neq j} M _ {\mathrm {C L S}} \left(\mathbf {x} _ {t} ^ {i}, \mathbf {x} _ {t} ^ {j}, \Delta \left(c ^ {j}, c ^ {i}\right)\right), \tag {6} +$$ + +where the higher logit indicates the stronger geometric consistency. + +Image-to-image translation model $M_{\mathrm{I2I}}$ . The image-to-image translation model $M_{\mathrm{I2I}}$ is tailored for novel view synthesis [26, 27, 29, 40]. We employ the most recent model, Wonder3D [29] as $M_{\mathrm{I2I}}$ , which is a viewpoint-conditioned image translation model and generates consistent content in the target viewpoint. When integrated with JSD, a random reference view $\mathbf{x}^{\mathrm{ref}}$ is selected from the set of 3D rendered images. Then we input the relative camera transformation $\Delta (c^{i}, c^{\mathrm{ref}})$ and rendered images to $M_{\mathrm{I2I}}$ . The measure of consistency is determined by calculating the reconstruction loss between the synthesized new + +![](images/05ee31543744ce6424a8ff61b64d3a716f75d5b6b3cc881732c67aa4b1f6f364.jpg) +Fig. 5: Comparison of text-to-3D generation. See Appendix for more results. + +image and its corresponding rendered image: + +$$ +\mathcal {C} _ {\mathrm {I 2 I}} (\tilde {\mathbf {x}}, \tilde {\mathbf {c}}) = - \sum_ {i \in 1, \dots , V} | | M _ {\mathrm {I 2 I}} (\mathbf {x} _ {t} ^ {\mathbf {r e f}}, \Delta (c ^ {i}, c ^ {\mathbf {r e f}})) - \mathbf {x} _ {t} ^ {i} | | _ {2} ^ {2}, \tag {7} +$$ + +where a smaller reconstruction loss indicates stronger geometric consistency under the estimation of $M_{\mathrm{I2I}}$ . + +Multi-view synthesis model $M_{\mathbf{MVS}}$ . The multi-view synthesis model is designed to generate multiple images conditioned on text prompts and camera poses, wherein we employ very recent work MVDream [41] as $M_{\mathbf{MVS}}$ . We compute the reconstruction loss for multiple generative views and rendered views: + +$$ +\mathcal {C} _ {\mathbf {M V S}} (\tilde {\mathbf {x}}, \tilde {\mathbf {c}}) = - | | M _ {\mathbf {M V S}} (y, \tilde {\mathbf {c}}) - \tilde {\mathbf {x}} | | _ {2} ^ {2} \tag {8} +$$ + +The smaller reconstruction loss signifies better geometric consistency within $M_{\mathbf{MVS}}$ . These view-aware models measure the coherence across views according to their own 3D-aware insights. Incorporated with JSD, they provide distinct constraints, resulting in different 3D generations that nonetheless contribute to enhanced geometric consistency. We believe that JSD can be adapted to more universal view-aware models, which enables us to progressively redefine the benchmark of 3D generation with the advancement of multi-view tasks. + +![](images/102b4a6ebb3b696522f4e2d89ea052289fe8e8ecb09e08d9c900169d2752a6fa.jpg) +A pink Spiderman dancing ballet, Marvel character HD, highly detailed 3D model + +![](images/3af7a027f877022bb7afa73cd7789c9983710496e07cac7a9cd7fb3cb47c2aa0.jpg) +a blue jay standing on a large basket of rainbow macarons + +![](images/1998689416a7d15587413fd8edaa95a77158f265e79006ff00bce141c91294ce.jpg) + +![](images/3bf4446cea63ec8c1e91bed4a1f1ea8f725aaf73698439af7c00f5c0f6e68594.jpg) + +![](images/01eac03f88f6d4290468fb46d98dfca58dd4a52be5656373f7c754c23eb28048.jpg) +SDS + +![](images/69e36279d7daf173dbe8bab8cefae7d3f48863bbf6f2e02b3906265b591d5f1f.jpg) +JSD w/ $C_{\mathrm{CLS}}$ +Fig. 6: Ablations on JSD incorporated with different energy function $\mathcal{C}$ . CLS: Binary Classification model; I2I: Image-to-Image Translation model (Wonder3D [29]); MVS: Multi-View Image Synthesis model (MVDream [41]). + +![](images/d72f5478b2d4cf76bd21e429cf0f6c55e92504ce12e7261538d26229a13e2bb3.jpg) +JSD w/C12I + +![](images/1fc711d4cbf5a4eef051a0c6dc9d7104b6a863b16842ee8a70fe681190930047.jpg) +JSD w/C_MVS + +# 4.3 Framework of JointDreamer + +Building upon the JSD optimization, we propose the overall framework JointDreamer. The optimization is based on neural radiance field (NeRF) [31], adopting Instant-NGP [32] with volume renderer. We adopt the multi-view synthesis model $M_{\mathrm{MVS}}$ as the energy function to integrate with JSD in JointDreamer. During optimization training, we utilize the common techniques including time-annealing and resolution scaling-up following [41, 42]. Besides, we propose two novel techniques to further enhance the generation quality, including a Geometry Fading scheme and a Classifier-Free Guidance Scale (CFG) Switching strategy. Geometry Fading. We aim to shift attention between geometric structure and texture details during optimization. Specifically, starting from iteration 5K, we reduce the learning rate of the density network of NeRF from $1e - 2$ to $1e - 6$ and set orientation loss to 0. Consequently, it benefits geometric convergence in the early phase of optimization, while allowing for decreased attention on geometry and increased attention on texture enhancement in the later stages. + +CFG Switching. CFG scheduling strategies have been employed in 2D domain [15,39] to enhance quality. In this work, we propose to modify the CFG scale $s$ during the training for 3D generation. We are motivated by the observation that a large CFG scale can lead to accelerated geometric convergence but may result in under-optimized geometry and distorted texture. Unlike the annealed CFG approach in CLIP-Sculptor [39], our increasing strategy prioritizes texture while maintaining accurate geometry. Specifically, a smaller $s = 30$ is employed in the early stages to preserve shape integrity, which allows for stronger coherence guidance from JSD. After 5K iterations, we increase the $s$ to 50, enhancing texture fidelity and overall quality. + +# 5 Experiment + +In this section, we present the text-to-3D generation results of JointDreamer with qualitative and quantitative evaluations, illustrating state-of-the-art per + +# a rabbit cutting grass with a lawnmower + +![](images/0371e9faf6e2e8e83e1302f14370651347026ce9da59593af65db1532a4204e1.jpg) +SDS +Fig. 7: Ablations on CFG Switching and Geometry Fading techniques in JointDreamer pipeline, demonstrating the effectiveness of quality enhancement. + +![](images/42c9d79ec65651d1664022af306902beb92bac090fbc22f4db7319ca0e7cbc2d.jpg) +original JSD + +![](images/cebe4868e9d9921597ed69400d942ae6a6489b2a36e5573d18a8247e7637d85e.jpg) ++CFG Switching + +![](images/3fdc32514adaa3be98d6b95b0af59b380c4d4cd51a361f0d083cda149af1bd97.jpg) ++ Geo Fading + +performance. We also make further ablation analysis on the proposed JSD. More training and evaluation details can be found in the Appendix. + +# 5.1 Text-to-3D Generation + +Table 1: Quantitative results on textual consistency and user preference, tested on object-centric subset of MS-COCO [25]. + +
MethodCLIP Score ↑R-Precision ↑User Study↑
DreamFusion [37]20.127.718.2
ProlificDreamer [42]25.018.716.2
MVDream [41]20.833.623.5
JointDreamer27.788.542.1
+ +Table 2: Ablation study on CFG Switching (CFGS) and Geometry Fading (GF). + +
SDS JSD CFGS GF|CLIP Score↑ FID↓
✓20.0 429.2
✓27.6 360.7
✓ ✓28.2 357.6
✓ ✓ ✓28.8 353.9
+ +Qualitative Comparisons. We compare with several representative baselines on threestudio [11] project, which modularizes the text-to-3D framework allowing for fair comparisons by ablating individual components. The generative samples are shown in Fig. 5. (i) DreamFusion [37]: Compared to DreamFusion which utilizes traditional SDS optimization, JointDreamer significantly improves the Multi-Face Janus issues in the generations of DreamFusion by introducing an energy term to ensure inter-view coherence. (ii) Magic3D [24]: Magic3D introduces a two-stage generation pipeline, transferring NeRF to DMTet in the second stage to enhance generation quality. We also transfer NeRF to DMTet as JointDreamer-DMTet, showcasing consistent superiority in geometry and texture quality by utilizing JSD optimization instead of SDS. (iii) ProlificDreamer [42]: ProlificDreamer presents variational score distillation (VSD) as a variant of score distillation function. While VSD enhances photorealism in 3D renderings by introducing a LoRA model, the ill-posed association of pose and images during LoRA training deepens the geometric inconsistency of 3D representation, resulting in severe multi-view artifacts in generations. In contrast, JointDreamer with JSD achieves geometric consistency while maintaining high-fidelity texture quality. (iv) MVDream [41]: Compared to direct distillation from a finetuned model as MVDream, JointDreamer employs view-aware models as the coherence constraint in JSD, while still inherit the generalization capabilities of original diffusion models [38]. The results indicate that JointDreamer mitigates the overfitting issues of MVDream, which accurately responds to complex input text and enhances the 3D consistency of generations. + +Quantitative Comparisons. Following [18,37], we evaluate CLIP Score [14], CLIP R-Precision [36] and user preference on the object-centric caption subset of MS-COCO [25] with 153 prompts to measure the text congruence and generative quality. For computational efficiency, we generate each 3D asset using 5K iterations of $64 \times 64$ rendered images and render 20 images per caption for evaluation. CLIP ViT-B/32 is adopted as the feature extractor for Clip Score and Clip R-Precision. As illustrated by the results in Table 1, JointDreamer can outperform all baselines on CLIP Score and CLIP R-Precision by large margins. Specifically, JointDreamer achieves an improvement of the R-Precision by $60.8\%$ and $54.9\%$ over DreamFusion and MVDream, demonstrating its superior corresponding to textual description. Notably, the severe Janus artifacts in ProlificDreamer compromise the quality of rendering with noisy background and semantic distortion, resulting in the lowest R-Precision. We also conduct a user study about shape preference on these prompts in Table 1. + +# 5.2 Ablation Analyses + +Ablations on Energy Functions. Sec. 4.2 discusses the varying interview + +coherency measurements provided by view-aware models trained. When incorporated with JSD, these models have distinct impacts on 3D generations, as shown in Fig. 6. The binary classifier effectively corrects inaccurate geometric structures in SDS. However, it cannot introduce additional imaginative elements as a discriminative model, resulting in oversaturated and monotonous textures. In con + +Table 3: Quantitative comparison of energy functions, showcasing the effectiveness of JSD in mitigating Janus artifacts with comparable computational efficiency to SDS. + +
MethodsJanus Rate ↓ GPU Memory ↓ Train Time ↓
SDS100%16.1 G50 min.
JSD w/CCLS12.5%22.1 G80 min.
JSD w/C2I31.2%16.0 G119 min.
JSD w/C_MVS6.2%19.4 G54 min.
+ +trast, as generative models, Wonder3D [29] and MVDream [41] employ reconstruction loss to estimate 3D consistency. Hence, they not only guide geometric structural modifications but also influence texture quality. we further conduct a quantitative comparison using 16 complex multi-Janus prompts, as outlined in Table 3. The experimental setup details can be found in the Appendix. The results indicate that our JSD consistently mitigates Janus artifacts across different view-aware models, with only a slight increase in computational requirements compared to SDS. We find that the energy function $C_{12\mathrm{I}}$ derived from the imagedto-image translation model exhibits poor performance, likely attributed to a mismatch in camera range and inaccurate translated results. + +Analysis on Geometry Fading and CFG Switching Mechanisms. We conduct incremental ablations on our proposed techniques in JointDreamer, including the Geometry Fading scheme and CFG Switching strategy. As shown in Fig. 7, increasing the CFG value enhances texture detail compared to the original JSD, but over-optimizes the geometry, resulting in a bumpier shape. Compared to “+CFG Switching”, the Geometry fading effectively protects shape when larger CFG guidance. We also conduct a quantitative evaluation on a $30\%$ + +![](images/e157a862a628d12ee2b43482add96ef10c550bad5e80f4812bd18bd605ff4218.jpg) +Fig. 9: Discuss the impact of prompts and random seeds, demonstrating the robustness of JointDreamer. $/...$ means the same content as the prompt in Fig. 7. + +MS-COCO subset in Table 2. JSD demonstrates superior texture quality compared to SDS, as evidenced by the Clip Score (CS) and FID metrics. The two proposed mechanisms further enhance the texture quality. + +Discussions on Training Loss. To make further comparisons with JSD + +and SDS, we aggregate training losses from multiple prompts on two optimization functions and visualize the training loss curve as illustrated in Fig. 8. We observe that SDS experiences significant fluctuations due to the randomness of single-view optimization. In contrast, JSD can converge gradually and smoothly, demonstrating the introduction of multi + +![](images/bbe2af54d2d91285ab5177d2cb8a553833888e958203ae0ca9e1017c2c00c2c4.jpg) +Fig.8: Training loss comparisons of JSD and SDS. JSD eliminates randomness fluctuations in SDS convergence. + +view optimization with inter-view coherence in JSD can reduce the randomness of optimization and contribute to better convergence for 3D representation. + +Discussions on Robustness for Prompts and Random Seed. To show the impact of seed and prompt, we modify seeds and key prompt components in Fig. 7, such as subject, object and verb. Results in Fig. 9 demonstrate the robustness of JointDreamer for different seed. Note that the default seed is 0. + +The Effectiveness of Classification Model. The classification model surpasses MVDream in training speed by a factor of 48. Fig.4 (b) provides additional high-quality results obtained using JSD in conjunction with the binary classification model $M_{\mathrm{CLS}}$ . Furthermore, we conduct an ablation study by replacing $\frac{\partial\mathcal{C}(\hat{\mathbf{x}})}{\partial\mathbf{x}_i^i}$ in Eq.5 with a randomly generated value between [0,1]. However, it yields shapeless results due to unrelated disturbances in the optimization process. These findings highlight the significance of the proposed classification model in achieving a balance between computational efficiency and generation quality. + +More Comparison with MVDream. View-aware models can be incorporated into 3D generation by combining them with JSD or using them directly with + +![](images/5449a51b441ed97c9047f6bbe0da35a093cb2df9d2136515ff4b5f87a604bbc0.jpg) +Fig. 10: Comparison of JointDreamer and SDS+MVDream, illustrating the superiority of JointDreamer in mitigating the Janus problem. SDS+MVDream consistently exhibits semantic missing or Janus issues with different weight options. + +SDS. In this discussion, we use MVDream [41] as an example to demonstrate the superiority of JSD. MVDream can generate consistent shapes when calculating the SDS loss directly. However, it may miss components in the complete input text due to its fine-tuning on limited 3D data, as illustrated in Fig. 5. To address this limitation, a straightforward approach is to "SDS+MVDream". + +However, balancing the impact of SDS and MVDream is challenging. As shown in Fig. 10, setting $\lambda_{\mathrm{SDS}}$ to 0 degrades to MVDream alone, while setting $\lambda_{\mathrm{MVDream}}$ to 0 yields a combination similar to DreamFusion [37]. Achieving a balanced impact between SDS and MVDream through a simple combination is difficult. When $\lambda_{\mathrm{MVDream}}$ is large, textual consistency remains constrained, whereas decreasing $\lambda_{\mathrm{MVDream}}$ leads to the Janus problem. This is due to gradient misalignment across multiple views in the "SDS+MVDream" combination, as SDS lacks multi-view information and cannot derive our objective outlined in Eq. 5. In contrast, JSD based on the joint image distribution provides supervision for text consistency and high-fidelity texture. JSD also promotes inter-view consistency by introducing an energy term with the view-aware model. + +# 6 Conclusion + +In this work, we introduce Joint Score Distillation (JSD) as a new paradigm for text-to-3D generation, which conducts multi-view optimization jointly and accounts for inter-view coherence. We demonstrate that JSD can significantly enhance 3D coherence while maintaining generalizability. With other proposed techniques, our overall framework, JointDreamer, is capable of geometric-consistent and high-fidelity 3D generation adhering to complex text input. + +Limitations. While the training time of JointDreamer is comparable to existing SDS pipelines, there is room for improvement in terms of acceleration. Future work will explore alternative 3D representations, such as 3D Gaussian [20]. Additionally, JSD utilizes view-aware models to ensure geometry consistency and mitigate the impact of limited data. However, view-aware models still require 3D data for training, and further efficient 3D data collection or reconstruction from multi-view images is also worth investigating. + +# References + +1. Armandpour, M., Zheng, H., Sadeghian, A., Sadeghian, A., Zhou, M.: Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. In: ICLR (2024) +2. Cao, Z., Hong, F., Wu, T., Pan, L., Liu, Z.: Large-vocabulary 3d diffusion model with transformer. In: ICLR (2024) +3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV. pp. 9650-9660 (2021) +4. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: CVPR. pp. 16123-16133 (2022) +5. Chen, H., Gu, J., Chen, A., Tian, W., Tu, Z., Liu, L., Su, H.: Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction. arXiv preprint arXiv:2304.06714 (2023) +6. Chen, R., Chen, Y., Jiao, N., Jia, K.: *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation*. In: ICCV. pp. 22246–22256 (2023) +7. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. In: NeurIPS (2024) +8. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objverse: A universe of annotated 3d objects. In: CVPR. pp. 13142-13153 (2023) +9. Deng, Y., Yang, J., Xiang, J., Tong, X.: Gram: Generative radiance manifolds for 3d-aware image generation. In: CVPR. pp. 10673-10683 (2022) +10. Gao, J., Shen, T., Wang, Z., Chen, W., Yin, K., Li, D., Litany, O., Gojcic, Z., Fidler, S.: Get3d: A generative model of high quality 3d textured shapes learned from images. NeurIPS 35, 31841-31854 (2022) +11. Guo, Y.C., Liu, Y.T., Shao, R., Laforte, C., Voleti, V., Luo, G., Chen, C.H., Zou, Z.X., Wang, C., Cao, Y.P., Zhang, S.H.: threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio (2023) +12. Henderson, P., Ferrari, V.: Learning single-image 3d reconstruction by generative modelling of shape, pose and shading. IJCV 128(4), 835-854 (2020) +13. Henderson, P., Tsiminaki, V., Lampert, C.H.: Leveraging 2d data to learn textured 3d mesh generation. In: CVPR. pp. 7498-7507 (2020) +14. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: CLIPScore: a reference-free evaluation metric for image captioning. In: EMNLP (2021) +15. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022) +16. Hu, Z., Zhao, M., Zhao, C., Liang, X., Li, L., Zhao, Z., Fan, C., Zhou, X., Yu, X.: Efficientdreamer: High-fidelity and robust 3d creation via orthogonal-view diffusion priors. In: CVPR. pp. 4949-4958 (2024) +17. Huang, Y., Wang, J., Shi, Y., Tang, B., Qi, X., Zhang, L.: Dreamtime: An improved optimization strategy for diffusion-guided 3d generation. In: ICLR (2023) +18. Jain, A., Mildenhall, B., Barron, J.T., Abbeel, P., Poole, B.: Zero-shot text-guided object generation with dream fields. In: CVPR. pp. 867-876 (2022) + +19. Jun, H., Nichol, A.: Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023) +20. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4), 1-14 (2023) +21. LeCun, Y., Chopra, S., Hadsell, R., Ranzato, M., Huang, F.: A tutorial on energy-based learning. Predicting structured data 1(0) (2006) +22. Li, M., Zhou, P., Liu, J.W., Keppo, J., Lin, M., Yan, S., Xu, X.: Instant3d: Instant text-to-3d generation. arXiv preprint arXiv:2311.08403 (2023) +23. Li, W., Chen, R., Chen, X., Tan, P.: Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. In: ICLR (2024) +24. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: CVPR (2023) +25. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +26. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: ICCV. pp. 9298-9309 (2023) +27. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. In: ICLR (2024) +28. Liu, Z., Feng, Y., Black, M.J., Nowrouzezahrai, D., Paull, L., Liu, W.: Meshdiffusion: Score-based generative 3d mesh modeling. In: ICLR (2023) +29. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3d: Single image to 3d using cross-domain diffusion. In: CVPR (2024) +30. Luo, W., Hu, T., Zhang, S., Sun, J., Li, Z., Zhang, Z.: Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. NeurIPS 36 (2024) +31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +32. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022) +33. Nguyen-Phuoc, T., Li, C., Theis, L., Richardt, C., Yang, Y.L.: Hologan: Unsupervised learning of 3d representations from natural images. In: ICCV. pp. 7588-7597 (2019) +34. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751 (2022) +35. Niemeyer, M., Geiger, A.: Giraffe: Representing scenes as compositional generative neural feature fields. In: CVPR. pp. 11453-11464 (2021) +36. Park, D.H., Azadi, S., Liu, X., Darrell, T., Rohrbach, A.: Benchmark for compositional text-to-image synthesis. In: NeurIPS Datasets and Benchmarks Track (2021) +37. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: ICLR (2023) +38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022) + +39. Sanghi, A., Fu, R., Liu, V., Willis, K.D., Shayani, H., Khasahmadi, A.H., Sridhar, S., Ritchie, D.: Clip-sculptor: Zero-shot generation of high-fidelity and diverse shapes from natural language. In: CVPR. pp. 18339-18348 (2023) +40. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110 (2023) +41. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. In: ICLR (2024) +42. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In: NeurIPS (2024) +43. Watson, D., Chan, W., Martin-Brualla, R., Ho, J., Tagliasacchi, A., Norouzi, M.: Novel view synthesis with diffusion models. In: ICLR (2023) +44. Weese, J., Kaus, M., Lorenz, C., Lobregt, S., Truyen, R., Pekar, V.: Shape constrained deformable models for 3d medical image segmentation. In: Information Processing in Medical Imaging: 17th International Conference, IPMI 2001 Davis, CA, USA, June 18-22, 2001 Proceedings 17. pp. 380-387. Springer (2001) +45. Zhao, W., Yan, L., Zhang, Y.: Geometric-constrained multi-view image matching method based on semi-global optimization. Geo-spatial information science 21(2), 115-126 (2018) \ No newline at end of file diff --git a/2024/JointDreamer_ Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation/images.zip b/2024/JointDreamer_ Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bd304733e7f3bd6464c7165af47b25df8bea816c --- /dev/null +++ b/2024/JointDreamer_ Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:862f234299f9d127fe2f7af0f7a631dd321fba69a2eb45c39b135801e6119fea +size 782634 diff --git a/2024/JointDreamer_ Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation/layout.json b/2024/JointDreamer_ Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..79db22ff487188663558051cd237ba04971a2cd7 --- /dev/null +++ b/2024/JointDreamer_ Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation/layout.json @@ -0,0 +1,9173 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 474, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 474, + 164 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 474, + 164 + ], + "type": "text", + "content": "JointDreamer: Ensuring Geometry Consistency and Text Congruence in Text-to-3D Generation via Joint Score Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 185, + 480, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 185, + 480, + 211 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 480, + 211 + ], + "type": "text", + "content": "Chenhan Jiang" + }, + { + "bbox": [ + 133, + 185, + 480, + 211 + ], + "type": "inline_equation", + "content": "^{1*†}" + }, + { + "bbox": [ + 133, + 185, + 480, + 211 + ], + "type": "text", + "content": ", Yihan Zeng" + }, + { + "bbox": [ + 133, + 185, + 480, + 211 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 133, + 185, + 480, + 211 + ], + "type": "text", + "content": ", Tianyang Hu" + }, + { + "bbox": [ + 133, + 185, + 480, + 211 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 185, + 480, + 211 + ], + "type": "text", + "content": ", Songcun Xu" + }, + { + "bbox": [ + 133, + 185, + 480, + 211 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 185, + 480, + 211 + ], + "type": "text", + "content": ", Wei Zhang" + }, + { + "bbox": [ + 133, + 185, + 480, + 211 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 185, + 480, + 211 + ], + "type": "text", + "content": ", Hang Xu" + }, + { + "bbox": [ + 133, + 185, + 480, + 211 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 185, + 480, + 211 + ], + "type": "text", + "content": ", and Dit-Yan Yeung" + }, + { + "bbox": [ + 133, + 185, + 480, + 211 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 190, + 219, + 421, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 219, + 421, + 231 + ], + "spans": [ + { + "bbox": [ + 190, + 219, + 421, + 231 + ], + "type": "text", + "content": "1 The Hong Kong University of Science and Technology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 253, + 232, + 361, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 232, + 361, + 241 + ], + "spans": [ + { + "bbox": [ + 253, + 232, + 361, + 241 + ], + "type": "text", + "content": "2 Huawei Noah's Ark Lab" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 169, + 242, + 444, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 242, + 444, + 253 + ], + "spans": [ + { + "bbox": [ + 169, + 242, + 444, + 253 + ], + "type": "text", + "content": "*Equal contribution. †Corresponding Author: jchcyan@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 235, + 254, + 378, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 254, + 378, + 264 + ], + "spans": [ + { + "bbox": [ + 235, + 254, + 378, + 264 + ], + "type": "text", + "content": "https://jointdreamer.github.io" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 291, + 455, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 291, + 455, + 522 + ], + "spans": [ + { + "bbox": [ + 159, + 291, + 455, + 522 + ], + "type": "text", + "content": "Abstract. Score Distillation Sampling (SDS) by well-trained 2D diffusion models has shown great promise in text-to-3D generation. However, this paradigm distills view-agnostic 2D image distributions into the rendering distribution of 3D representation for each view independently, overlooking the coherence across views and yielding 3D inconsistency in generations. In this work, we propose Joint Score Distillation (JSD), a new paradigm that ensures coherent 3D generations. Specifically, we model the joint image distribution, which introduces an energy function to capture the coherence among denoised images from the diffusion model. We then derive the joint score distillation on multiple rendered views of the 3D representation, as opposed to a single view in SDS. In addition, we instantiate three universal view-aware models as energy functions, demonstrating compatibility with JSD. Empirically, JSD significantly mitigates the 3D inconsistency problem in SDS, while maintaining text congruence. Moreover, we introduce the Geometry Fading scheme and Classifier-Free Guidance (CFG) Switching strategy to enhance generative details. Our framework, JointDreamer, establishes a new benchmark in text-to-3D generation, achieving outstanding results with an " + }, + { + "bbox": [ + 159, + 291, + 455, + 522 + ], + "type": "inline_equation", + "content": "88.5\\%" + }, + { + "bbox": [ + 159, + 291, + 455, + 522 + ], + "type": "text", + "content": " CLIP R-Precision and " + }, + { + "bbox": [ + 159, + 291, + 455, + 522 + ], + "type": "inline_equation", + "content": "27.7\\%" + }, + { + "bbox": [ + 159, + 291, + 455, + 522 + ], + "type": "text", + "content": " CLIP Score. These metrics demonstrate exceptional text congruence, as well as remarkable geometric consistency and texture fidelity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 533, + 402, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 533, + 402, + 544 + ], + "spans": [ + { + "bbox": [ + 160, + 533, + 402, + 544 + ], + "type": "text", + "content": "Keywords: 3D Vision " + }, + { + "bbox": [ + 160, + 533, + 402, + 544 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 533, + 402, + 544 + ], + "type": "text", + "content": " 3D Generation " + }, + { + "bbox": [ + 160, + 533, + 402, + 544 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 533, + 402, + 544 + ], + "type": "text", + "content": " Energy Function" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 485, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 485, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 485, + 667 + ], + "type": "text", + "content": "3D content creation is essential for diverse applications, including gaming, robotics simulation, and virtual reality. However, it is labor-intensive, demanding substantial time for skilled designers to create a single 3D asset. Hence, automating 3D creation with text input has attracted considerable attention. Recently, the score distillation sampling (SDS) algorithm pioneered by DreamFusion [37]" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 474, + 454 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 474, + 454 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 474, + 454 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 474, + 454 + ], + "type": "image", + "image_path": "3f3a1bb770864999f37a8549ab0e64c86a81933d0589418a70a502ab4eda53ce.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 457, + 482, + 491 + ], + "lines": [ + { + "bbox": [ + 132, + 457, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 482, + 491 + ], + "type": "text", + "content": "Fig. 1: Text-to-3D generations by JointDreamer from scratch. JointDreamer excels in generating geometrically consistent and high-fidelity 3D assets, adhering to complex textual descriptions that are challenging for previous methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 506, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 567 + ], + "type": "text", + "content": "shows promise in text-to-3D tasks, which lifts image distribution from a well-trained diffusion model [38] into parameterized 3D representation like NeRF [31]. Compared to 3D generative models [2, 19, 34] that struggle with producing arbitrary objects due to limited text-3D training data, SDS-based methods [6, 24, 37, 42] can generate arbitrary 3D assets with diverse text input." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Although SDS-based methods benefit from the generalizability of diffusion models, they often encounter a common issue known as Janus artifacts [6, 24, 37, 42]. These artifacts manifest as repeated content from different viewpoints of a 3D generation, yielding a lack of realism and coherence in the rendered views. We investigate the Janus artifacts by visualizing image distributions of the diffusion model [38] from multiple viewpoints, as illustrated in Fig. 2(a). The results reveal the view-agnostic nature and the content inconsistency across views of the diffusion model. Consequently, SDS optimizes each rendered view of 3D" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "C. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 480, + 140 + ], + "type": "text", + "content": "representation independently and inherits image distribution without coherent multi-view perspective, leading to the geometric inconsistency of 3D generations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 148, + 482, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 148, + 482, + 256 + ], + "spans": [ + { + "bbox": [ + 130, + 148, + 482, + 256 + ], + "type": "text", + "content": "Existing works [1,22] address the aforementioned challenges within the SDS framework by employing prompt engineering techniques. However, the effectiveness of such methods remains unsatisfactory, as evident from the results depicted in Fig. 2(b). Alternative methods [16,41] propose to finetune view-aware diffusion models using rendered images of 3D datasets [7,8]. Nevertheless, they are prone to overfitting on limited text-3D training data [23], decreasing the text congruence when handling complex text inputs. Based on the above observations, it requires a rethinking of the SDS optimization to enhance the 3D coherence of generations while maintaining generalizability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 262, + 317, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 317, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 317, + 478 + ], + "type": "text", + "content": "In this work, we first present Joint Score Distillation (JSD), which significantly promotes the 3D consistency of generation and inherits generalizability from diffusion models. Specifically, we model the joint image distribution of diffusion model via an energy function measuring coherence across denoised images. It facilitates the extension of the KL-Divergence in SDS from single-view into multi-view. We then derive the joint score distillation function from multi-view KL-Divergence, which ensures inter-view coherence in the optimization process of 3D generation. We show that SDS is a special case of JSD with the energy term omitted, which indicates the absence of coherence constraint across views." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 486, + 317, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 317, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 317, + 592 + ], + "type": "text", + "content": "Building upon JSD optimization, we present three view-aware models as energy terms to showcase the compatibility of JSD: the Binary Classification Model, the Image-to-Image Translation Model, and the Multi-view Generation Model. Through empirical analysis, it is observed that different view-aware models introduce distinct coherence measurements," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "leading to diverse 3D generations, while all contributing to 3D consistency. Furthermore, to facilitate a more comprehensive comparison with existing text-to-3D generation methods, we introduce JointDreamer, an innovative framework capable of producing geometric-consistent and high-fidelity 3D assets adhering to complex text descriptions. Notably, in addition to incorporating a Multi-view Generation model as an energy term in JSD, we introduce two complementary" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 325, + 266, + 478, + 395 + ], + "blocks": [ + { + "bbox": [ + 325, + 266, + 478, + 395 + ], + "lines": [ + { + "bbox": [ + 325, + 266, + 478, + 395 + ], + "spans": [ + { + "bbox": [ + 325, + 266, + 478, + 395 + ], + "type": "image", + "image_path": "60635b6e1dfbdb5e4b443c7cb208dff487674074281f21709a1c38f9b1835a20.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 400, + 482, + 577 + ], + "lines": [ + { + "bbox": [ + 321, + 400, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 321, + 400, + 482, + 577 + ], + "type": "text", + "content": "Fig.2: Illustration of text-conditioned images for different viewpoints, where input texts are augmented with corresponding direction prompts for each view. (a) The original generations from 2D diffusion model [38] are view-agnostic and inconsistent across views. (b) Text prompt tuning [1] has limited improvement in the directional structure of generated images for each view. (c) JSD injects coherence measurement from the proposed binary classifier (refer to Section 4.2), contributing to modified directional structures and semantical consistency across views." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 101 + ], + "type": "text", + "content": "JointDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "techniques, namely the Geometry Fading scheme and the Classifier-Free Guidance (CFG) Switching strategy, to enhance generative details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 210 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 210 + ], + "type": "text", + "content": "We systematically assess the quality of our approach, both qualitatively and quantitatively, compared to existing methods. Qualitative results gallery can be found in Fig. 1. Our JointDreamer consistently produces high-fidelity 3D assets and mitigates Janus artifacts in SDS. It maintains text congruence even when confronted with complex text input, achieving " + }, + { + "bbox": [ + 130, + 140, + 480, + 210 + ], + "type": "inline_equation", + "content": "88.5\\%" + }, + { + "bbox": [ + 130, + 140, + 480, + 210 + ], + "type": "text", + "content": " CLIP R-Precision and " + }, + { + "bbox": [ + 130, + 140, + 480, + 210 + ], + "type": "inline_equation", + "content": "27.7\\%" + }, + { + "bbox": [ + 130, + 140, + 480, + 210 + ], + "type": "text", + "content": " CLIP Score." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 213, + 382, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 213, + 382, + 223 + ], + "spans": [ + { + "bbox": [ + 146, + 213, + 382, + 223 + ], + "type": "text", + "content": "In brief, our contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 233, + 480, + 341 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 233, + 479, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 233, + 479, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 233, + 479, + 267 + ], + "type": "text", + "content": "- We introduce a novel Joint Score Distillation (JSD) for text-to-3D generation, optimizing multiple views jointly via an energy function to capture inter-view coherence." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 270, + 479, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 270, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 270, + 479, + 293 + ], + "type": "text", + "content": "- We present three view-aware models as energy functions to show compatibility with JSD, all of them mitigate the Janus problem in SDS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 293, + 480, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 341 + ], + "type": "text", + "content": "- We introduce the text-to-3D framework JointDreamer, incorporating complementary Geometry Fading and CFG Switching techniques. Our JointDreamer achieves geometrically consistent and high-fidelity 3D assets even with complex textual inputs." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 361, + 242, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 242, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 242, + 373 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 383, + 481, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 481, + 623 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 481, + 623 + ], + "type": "text", + "content": "Text-to-3D Generation. Existing text-to-3D generation methods can be categorized into two streams: 3D generative models and 2D optimization methods. The former encompasses various deep generative models such as Variational Auto Encoders (VAEs) [12, 13], Generative Adversarial Models (GAN) [4, 9, 10, 33, 35], diffusion models [5, 28, 34] and transformer architectures [2, 19]. These models are efficient in inference but often struggle with generalizability and training stability, attributed to the limited scope and complexity of available 3D datasets. The latter approach centers around the Score Distillation Sampling (SDS) algorithm proposed by [37], which leverages 2D diffusion model priors [38] for optimizing 3D representations. Subsequent advancements have refined this technique by improving the 3D representations [6, 24], the sampling scheduler [17] and loss design [42]. However, the above approaches overlook the geometric consistency problem, facing inherent multi-face Janus issues in SDS. Prior works try to alleviate the Janus issues with prompt tuning [1] yet they achieve limited effect. Very recent work MVDream [41] addresses the problem by fine-tuning a multi-view diffusion model, but it is susceptible to overfitting on scarce 3D training data, compromising semantic consistency in text-to-3D generations. In this work, we address the fundamental flaw of SDS that optimizes each view independently by introducing a joint optimization function that enforces inter-view consistency, essentially solving the Janus issues in SDS while preserving its generalizability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "Diffusion-based Novel View Synthesis. As an alternative to 3D generation, novel view synthesis models the challenge as a view-conditioned image-to-image translation task. There have been proposals for pose-conditioned image-to-image" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "C. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 478, + 216 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 478, + 216 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 478, + 216 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 478, + 216 + ], + "type": "image", + "image_path": "a09b64fcd74f2b86786d190bb558a9ecae1f0da3bd57f94f32b1fcb31dca274a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 220, + 480, + 255 + ], + "lines": [ + { + "bbox": [ + 130, + 220, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 480, + 255 + ], + "type": "text", + "content": "Fig. 3: Overview of JointDreamer Framework. We introduce an energy function to model the joint distribution for multi-view denoised images from 2D diffusion model, facilitating the Joint Score Distillation (JSD) optimization for text-to-3D generation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 411 + ], + "type": "text", + "content": "diffusion models [43] that generate novel views on synthetic data in 3D. Recently, [26] promoted the generalizability of novel view synthesis by fine-tuning the 2D diffusion model [38] on renderings of 3D dataset, which facilitates images-to-3D tasks with 3D reconstruction or SDS algorithm. To further improve the 3D consistency across generated views and input view, very recent works [27, 29, 40] modify the generation process into multi-view generation and present corresponding architecture designs. Generally, these methods take camera specifications as conditions and enable the viewpoint-aware generations, but they can hardly accurately capture a complete 3D scene consistently and densely. Our method acknowledges the potential of these models to discern relative interview relationships, which we harness to provide inter-view coherence for our JSD, thus serving as universal guidance models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 426, + 233, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 233, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 233, + 438 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 480, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 480, + 495 + ], + "type": "text", + "content": "We review the original SDS and address its fundamental limitation: optimizing independently for each single view of 3D representation and distilling with the view-agnostic image distribution of diffusion model. It results in geometric inconsistency issues, dubbed as Multi-Face Janus Problem." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 496, + 480, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 480, + 544 + ], + "type": "text", + "content": "Score Distillation Sampling (SDS). SDS optimization is widely adopted by text-to-3D generation pipelines [6,24,30,37,42]. Given a 3D representation with learnable parameters " + }, + { + "bbox": [ + 130, + 496, + 480, + 544 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 496, + 480, + 544 + ], + "type": "text", + "content": " and a pre-trained 2D diffusion model with noise prediction network " + }, + { + "bbox": [ + 130, + 496, + 480, + 544 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\Phi}(\\mathbf{x}_t,t,y)" + }, + { + "bbox": [ + 130, + 496, + 480, + 544 + ], + "type": "text", + "content": ", SDS optimizes " + }, + { + "bbox": [ + 130, + 496, + 480, + 544 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 496, + 480, + 544 + ], + "type": "text", + "content": " by minimizing the KL-divergence:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 242, + 554, + 481, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 554, + 481, + 571 + ], + "spans": [ + { + "bbox": [ + 242, + 554, + 481, + 571 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} D _ {K L} \\left(q _ {t} ^ {\\theta} \\left(\\mathbf {x} _ {t} | c, y\\right) \\mid \\mid p _ {t} \\left(\\mathbf {x} _ {t} | y\\right)\\right). \\tag {1}", + "image_path": "5e3a1c1517b671026fa964e3b48129050e68895113e5c3691dcbcd0c6456879d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 573, + 482, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 573, + 482, + 623 + ], + "spans": [ + { + "bbox": [ + 130, + 573, + 482, + 623 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 573, + 482, + 623 + ], + "type": "inline_equation", + "content": "p_t(\\mathbf{x}_t | y)" + }, + { + "bbox": [ + 130, + 573, + 482, + 623 + ], + "type": "text", + "content": " is the image distribution sampled from diffusion model, " + }, + { + "bbox": [ + 130, + 573, + 482, + 623 + ], + "type": "inline_equation", + "content": "q_t^\\theta(\\mathbf{x}_t | c, y)" + }, + { + "bbox": [ + 130, + 573, + 482, + 623 + ], + "type": "text", + "content": " is the distribution of rendered image " + }, + { + "bbox": [ + 130, + 573, + 482, + 623 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t = g(\\theta, c)" + }, + { + "bbox": [ + 130, + 573, + 482, + 623 + ], + "type": "text", + "content": " with respect to camera pose " + }, + { + "bbox": [ + 130, + 573, + 482, + 623 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 573, + 482, + 623 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 130, + 573, + 482, + 623 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 573, + 482, + 623 + ], + "type": "text", + "content": " of the forward diffusion process, where " + }, + { + "bbox": [ + 130, + 573, + 482, + 623 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 573, + 482, + 623 + ], + "type": "text", + "content": " is the renderer. To solve Eq. (1), the score distillation function is derived as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 194, + 626, + 480, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 626, + 480, + 668 + ], + "spans": [ + { + "bbox": [ + 194, + 626, + 480, + 668 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\theta} L _ {S D S} (\\theta) \\triangleq \\mathrm {E} _ {t, \\mathbf {x}} [ w (t) \\frac {\\sigma_ {t}}{\\alpha_ {t}} \\nabla_ {\\theta} K L \\left(q _ {t} ^ {\\theta} \\left(\\mathbf {x} _ {t} | c, y\\right) \\mid \\mid p _ {t} \\left(\\mathbf {x} _ {t} | y\\right)\\right) ] \\\\ \\triangleq \\mathrm {E} _ {t, \\epsilon_ {\\Phi}} [ w (t) (\\hat {\\epsilon} _ {\\Phi} (\\mathbf {x} _ {t}, t, y) - \\epsilon) \\frac {\\delta g (\\theta , c)}{\\delta \\theta} ], \\tag {2} \\\\ \\end{array}", + "image_path": "084d1731851a5e02adf3b924a581ba1d458cc84d3498b6119bc62221c33fdf6b.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "JointDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\Phi} := (1 + s)\\epsilon_{\\Phi}(\\mathbf{x}_t, t, y) - s\\epsilon_{\\Phi}(\\mathbf{x}_t, t, \\emptyset)" + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "content": " is modification of predicted noise with classifier-free guidance (CFG) " + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "content": " is time-dependent weighting function. Multi-Face Janus Problem. To achieve consistency, it is essential that the rendering distributions " + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "inline_equation", + "content": "q_0^\\theta(\\mathbf{x}_0|c, y)" + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "content": " adhere to text condition " + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "content": " and image distribution " + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "inline_equation", + "content": "p_t(\\mathbf{x}_t|y)" + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "content": " keep consistency across views with different poses. For the image distribution " + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "inline_equation", + "content": "p_0(\\mathbf{x}_0|y)" + }, + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "content": " of 2D diffusion model, the pose condition can be injected via input text with the corresponding directional prompt [37, 42]. As illustrated in Fig. 2(a), the pre-trained 2D image distribution, trained on individual images, is view-agnostic and lacks identity consistency across views. Even with a text tuning mechanism [1] specifically designed for multi-view image generation, as shown in Fig. 2(b), the above issues are far from resolved. Since SDS minimizes KL-divergence between the image distribution and rendering distribution independently for each rendered view, it can only inevitably inherit the 3D-awareness deficit of the 2D diffusion model, resulting in inconsistent 3D generation, which is commonly referred to as the Multi-face Janus Problem of SDS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 313, + 202, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 313, + 202, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 313, + 202, + 326 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 338, + 482, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 447 + ], + "type": "text", + "content": "In this section, we introduce JointDreamer, a novel text-to-3D generation framework as illustrated in Fig. 3. We first present the derivation of Joint Score Distillation (JSD) in Sec. 4.1, which extends the single-view optimization in SDS into a multi-view KL-Divergence. Then we integrate universal view-aware models into JSD to show the compatibility of JSD in Sec. 4.2, where we instantiate three kinds of view-aware models to capture inter-view coherence. Finally, we elaborate on the overall framework JointDreamer in Sec. 4.3, where we integrate the multi-view generation model into JSD. We also propose a geometry fading scheme and CFG switching strategy to further enhance generative quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 463, + 310, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 463, + 310, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 463, + 310, + 475 + ], + "type": "text", + "content": "4.1 Joint Score Distillation (JSD)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 482, + 481, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 481, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 481, + 519 + ], + "type": "text", + "content": "To address the multi-face problem arising from SDS, we extend the score distillation from single-view to multi-view settings and promote inter-view coherence across 2D image distribution, and thus derive our JSD optimization function." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 523, + 482, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 482, + 632 + ], + "type": "text", + "content": "Coherence Modeling for Joint Image Distribution. As discussed above, the rendering distributions of the 3D representation should maintain 3D consistency across views " + }, + { + "bbox": [ + 132, + 523, + 482, + 632 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}} = \\{\\mathbf{x}^1,\\mathbf{x}^2,\\dots ,\\mathbf{x}^V\\}" + }, + { + "bbox": [ + 132, + 523, + 482, + 632 + ], + "type": "text", + "content": " with respect to different poses " + }, + { + "bbox": [ + 132, + 523, + 482, + 632 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{c}} = \\{c^{1},c^{2},\\ldots ,c^{V}\\}" + }, + { + "bbox": [ + 132, + 523, + 482, + 632 + ], + "type": "text", + "content": ". However, for 2D pre-trained diffusion models, different views are generated independently. To ensure consistency, we propose modeling the joint image distribution of multiple views, denoted as " + }, + { + "bbox": [ + 132, + 523, + 482, + 632 + ], + "type": "inline_equation", + "content": "p_0(\\tilde{\\mathbf{x}} |\\tilde{\\mathbf{c}},y)" + }, + { + "bbox": [ + 132, + 523, + 482, + 632 + ], + "type": "text", + "content": ", within the diffusion model. Following the commonly adopted assumption in energy-based distribution modeling [21, 44, 45], we introduce an energy function that measures the inter-view coherence by " + }, + { + "bbox": [ + 132, + 523, + 482, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\tilde{\\mathbf{x}},\\tilde{\\mathbf{c}}):\\mathbb{R}^{Vd}\\to \\mathbb{R}" + }, + { + "bbox": [ + 132, + 523, + 482, + 632 + ], + "type": "text", + "content": " and define:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 225, + 639, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 639, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 225, + 639, + 481, + 669 + ], + "type": "interline_equation", + "content": "p _ {0} (\\tilde {\\mathbf {x}} | \\tilde {\\mathbf {c}}, y) \\propto \\exp (\\mathcal {C} (\\tilde {\\mathbf {x}}, \\tilde {\\mathbf {c}})) \\prod_ {i = 1} ^ {V} p _ {0} (\\mathbf {x} ^ {i} | c ^ {i}, y), \\tag {3}", + "image_path": "68122f04bba7a9e0ab1dfc961ff5b79100474a384e8f2a03909dcd3d6416692c.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "C. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 200 + ], + "type": "text", + "content": "where a larger " + }, + { + "bbox": [ + 131, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\tilde{\\mathbf{x}},\\tilde{\\mathbf{c}})" + }, + { + "bbox": [ + 131, + 116, + 482, + 200 + ], + "type": "text", + "content": " indicates greater coherence among the denoised view images. As a result, the joint image distribution is no longer view-independent. In practice, the joint energy function can be implemented via various view-aware models, as long as they can reflect the coherence across multiple views. In Sec. 4.2, we explore different choices in depth. The modeling of coherence across joint image distributions facilitates our JSD on multi-view, integrating inter-view coherence to ensure consistent 3D representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 206, + 483, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 206, + 483, + 229 + ], + "spans": [ + { + "bbox": [ + 131, + 206, + 483, + 229 + ], + "type": "text", + "content": "KL-Divergence on Multiple Views. We extend the single-view KL-divergence in SDS to a multi-view version, based on the joint image distribution:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 206, + 238, + 480, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 238, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 206, + 238, + 480, + 285 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\theta} D _ {K L} \\left(q _ {t} ^ {\\theta} (\\tilde {\\mathbf {x}} | \\tilde {\\mathbf {c}}, y) \\| p _ {t} (\\tilde {\\mathbf {x}} | \\tilde {\\mathbf {c}}, y)\\right) \\\\ = \\min _ {\\theta} D _ {K L} \\left(q _ {t} ^ {\\theta} (\\tilde {\\mathbf {x}} | \\tilde {\\mathbf {c}}, y) \\right\\lVert \\exp \\left(\\mathcal {C} \\left(\\tilde {\\mathbf {x}}, \\tilde {\\mathbf {c}}\\right)\\right) \\prod_ {i = 1} ^ {V} p _ {t} \\left(\\mathbf {x} ^ {i} \\mid c ^ {i}, y\\right), \\tag {4} \\\\ \\end{array}", + "image_path": "e4f25bb86c9b71d0028332adb0aa4fd204bc20593b3f9f389ffa89b4c5e45110.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 293, + 479, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 293, + 479, + 342 + ], + "spans": [ + { + "bbox": [ + 131, + 293, + 479, + 342 + ], + "type": "text", + "content": "where the extra energy term " + }, + { + "bbox": [ + 131, + 293, + 479, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\tilde{\\mathbf{x}},\\tilde{\\mathbf{c}})" + }, + { + "bbox": [ + 131, + 293, + 479, + 342 + ], + "type": "text", + "content": " in Eq. (3) accounts for the inter-view coherence. Without this constraint, e.g., " + }, + { + "bbox": [ + 131, + 293, + 479, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\tilde{\\mathbf{x}},\\tilde{\\mathbf{c}})\\equiv 0" + }, + { + "bbox": [ + 131, + 293, + 479, + 342 + ], + "type": "text", + "content": ", different rendering views are optimized independently with the 2D diffusion model separately. In this sense, the original SDS can be seen as a special case of JSD." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 347, + 479, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 347, + 479, + 383 + ], + "spans": [ + { + "bbox": [ + 131, + 347, + 479, + 383 + ], + "type": "text", + "content": "Joint Score Distillation Function. To correspond to the gradient of the new rule of multi-view KL-divergence in Eq. (4), we derive our score distillation function that is jointly conducted on multiple views as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 206, + 392, + 480, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 392, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 206, + 392, + 480, + 449 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\theta} L _ {J S D} (\\theta) \\\\ \\triangleq \\mathrm {E} _ {t, \\epsilon_ {\\Phi}} [ w (t) \\mathbb {E} (\\nabla_ {\\tilde {\\mathbf {x}}} \\log q _ {t} (\\tilde {\\mathbf {x}} _ {t} | \\tilde {\\mathbf {x}} _ {0}) - \\nabla_ {\\tilde {\\mathbf {x}}} \\log p _ {t} (\\tilde {\\mathbf {x}} _ {t} | y)) ] \\tag {5} \\\\ = \\sum_ {i = 1} ^ {V} \\mathrm {E} _ {t, \\epsilon_ {\\Phi} ^ {i}} [ w (t) (\\hat {\\epsilon} _ {\\Phi} (\\mathbf {x} _ {t} ^ {i}, y) - \\frac {\\partial \\mathcal {C} (\\tilde {\\mathbf {x}})}{\\partial \\mathbf {x} _ {t} ^ {i}} - \\epsilon^ {i}) \\frac {\\delta g (\\theta , c ^ {i})}{\\delta \\theta} ], \\\\ \\end{array}", + "image_path": "ad004f9352cde38e89f1e837b44b1369eedc4093c83108f2a463f98e9def38d6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 459, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 459, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 131, + 459, + 482, + 591 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 459, + 482, + 591 + ], + "type": "inline_equation", + "content": "\\{\\epsilon^i\\}_{i = 1}^V" + }, + { + "bbox": [ + 131, + 459, + 482, + 591 + ], + "type": "text", + "content": " are noises during score matching for different views. The proof can be found in the Appendix. To intuitively compare SDS with JSD, we sample multi-view images from the forward pass of pre-trained diffusion model [38] as shown in Fig. 2(c). For each view " + }, + { + "bbox": [ + 131, + 459, + 482, + 591 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 131, + 459, + 482, + 591 + ], + "type": "text", + "content": ", we randomly select " + }, + { + "bbox": [ + 131, + 459, + 482, + 591 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^\\prime" + }, + { + "bbox": [ + 131, + 459, + 482, + 591 + ], + "type": "text", + "content": " from a different view and adopt the binary classification model presented in Sec. 4.2 as the energy function to measure coherence across the two views. The results illustrate that JSD significantly enhances the correspondence with directional prompts and consistency across different views, particularly for the initially biased views such as side and back views. These theoretical and empirical results prove that the multi-face Janus issue in SDS is rooted in the view-agnostic and view-biased 2D image distribution, a challenge effectively mitigated by JSD." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 609, + 413, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 413, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 413, + 621 + ], + "type": "text", + "content": "4.2 Universal View-Aware Models as Energy Function" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 479, + 665 + ], + "type": "text", + "content": "JSD requires an energy function " + }, + { + "bbox": [ + 131, + 629, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\tilde{\\mathbf{x}},\\tilde{\\mathbf{c}})" + }, + { + "bbox": [ + 131, + 629, + 479, + 665 + ], + "type": "text", + "content": " to measure coherence across denoised images, as presented in Eq. (3). This energy function plays a crucial role in assessing the consistency between different views in image distribution. To demonstrate" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "JointDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 242, + 278 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 242, + 278 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 242, + 278 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 242, + 278 + ], + "type": "image", + "image_path": "e09f1c9ad9fd4190ae30fb8f1027fa4246413df1e8eb1bcb7d934abc5ec3222e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "type": "text", + "content": "Fig. 4: Illustration of the binary classification model and qualitative results with JSD. (a) The classification model " + }, + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{CLS}}" + }, + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "type": "text", + "content": " produces the binary logit to measure the consistency between two input views " + }, + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "type": "inline_equation", + "content": "x^i" + }, + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "type": "inline_equation", + "content": "x^j" + }, + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "type": "text", + "content": ". (b) JSD integrated with the classification model effectively alleviates Janus issues compared to SDS." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 244, + 114, + 478, + 279 + ], + "blocks": [ + { + "bbox": [ + 244, + 114, + 478, + 279 + ], + "lines": [ + { + "bbox": [ + 244, + 114, + 478, + 279 + ], + "spans": [ + { + "bbox": [ + 244, + 114, + 478, + 279 + ], + "type": "image", + "image_path": "8310201a10e7cba4b295123a8f7c531556b987951d0cd38f7015d09ecab44872.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 350, + 480, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 480, + 373 + ], + "type": "text", + "content": "the compatibility of JSD, we employ three different types of models trained for various representative multi-view tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "text", + "content": "Binary classification model " + }, + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{CLS}}" + }, + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "text", + "content": ". The binary classification model " + }, + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{CLS}}" + }, + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "text", + "content": " is designed to classify the content consistency between two input images based on their relative camera pose. To ensure computational efficiency, we introduce a dedicated classification model, as shown in Fig. 4 (a). The training process takes two days with a single A800 GPU on Objaverse dataset [8]. Further details can be found in the appendix. The classification model " + }, + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{CLS}}" + }, + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "text", + "content": " processes pairs of images " + }, + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^i" + }, + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^j" + }, + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "text", + "content": " captured from different viewpoints " + }, + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "inline_equation", + "content": "c^i" + }, + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "inline_equation", + "content": "c^j" + }, + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "text", + "content": ". It extracts image features using the DINO-ViT/s16 backbone [3]. These image features are conditioned on the camera feature obtained through MLP layers from the relative camera transformation matrix " + }, + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\Delta(c^j, c^i)" + }, + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "text", + "content": ". Finally, the classification head produces the binary score. To incorporate with JSD, we only consider neighboring views in " + }, + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "text", + "content": " as image pairs for coherence measurement, which is denoted as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 221, + 526, + 481, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 526, + 481, + 550 + ], + "spans": [ + { + "bbox": [ + 221, + 526, + 481, + 550 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {\\mathrm {C L S}} \\left(\\tilde {\\mathbf {x}}, \\tilde {\\mathbf {c}}\\right) = \\sum_ {i, j \\in 1, \\dots , V; i \\neq j} M _ {\\mathrm {C L S}} \\left(\\mathbf {x} _ {t} ^ {i}, \\mathbf {x} _ {t} ^ {j}, \\Delta \\left(c ^ {j}, c ^ {i}\\right)\\right), \\tag {6}", + "image_path": "4f2b404b3b559e3f27ea621e42745e4b21a8d8e8016746b964ddc94408e1e432.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 428, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 428, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 428, + 569 + ], + "type": "text", + "content": "where the higher logit indicates the stronger geometric consistency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Image-to-image translation model " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{I2I}}" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ". The image-to-image translation model " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{I2I}}" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " is tailored for novel view synthesis [26, 27, 29, 40]. We employ the most recent model, Wonder3D [29] as " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{I2I}}" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ", which is a viewpoint-conditioned image translation model and generates consistent content in the target viewpoint. When integrated with JSD, a random reference view " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{ref}}" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " is selected from the set of 3D rendered images. Then we input the relative camera transformation " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\Delta (c^{i}, c^{\\mathrm{ref}})" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " and rendered images to " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{I2I}}" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ". The measure of consistency is determined by calculating the reconstruction loss between the synthesized new" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "C. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 112, + 482, + 381 + ], + "blocks": [ + { + "bbox": [ + 130, + 112, + 482, + 381 + ], + "lines": [ + { + "bbox": [ + 130, + 112, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 112, + 482, + 381 + ], + "type": "image", + "image_path": "05ee31543744ce6424a8ff61b64d3a716f75d5b6b3cc881732c67aa4b1f6f364.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 384, + 475, + 396 + ], + "lines": [ + { + "bbox": [ + 137, + 384, + 475, + 396 + ], + "spans": [ + { + "bbox": [ + 137, + 384, + 475, + 396 + ], + "type": "text", + "content": "Fig. 5: Comparison of text-to-3D generation. See Appendix for more results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 422, + 330, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 330, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 330, + 434 + ], + "type": "text", + "content": "image and its corresponding rendered image:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 205, + 442, + 481, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 442, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 205, + 442, + 481, + 467 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {\\mathrm {I 2 I}} (\\tilde {\\mathbf {x}}, \\tilde {\\mathbf {c}}) = - \\sum_ {i \\in 1, \\dots , V} | | M _ {\\mathrm {I 2 I}} (\\mathbf {x} _ {t} ^ {\\mathbf {r e f}}, \\Delta (c ^ {i}, c ^ {\\mathbf {r e f}})) - \\mathbf {x} _ {t} ^ {i} | | _ {2} ^ {2}, \\tag {7}", + "image_path": "b37dbcf31ce7738efddb34f1c7f9b9d3c776a6599629798164f3111c3a71bd47.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 477, + 480, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 477, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 477, + 480, + 501 + ], + "type": "text", + "content": "where a smaller reconstruction loss indicates stronger geometric consistency under the estimation of " + }, + { + "bbox": [ + 130, + 477, + 480, + 501 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{I2I}}" + }, + { + "bbox": [ + 130, + 477, + 480, + 501 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 502, + 480, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 480, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 480, + 550 + ], + "type": "text", + "content": "Multi-view synthesis model " + }, + { + "bbox": [ + 130, + 502, + 480, + 550 + ], + "type": "inline_equation", + "content": "M_{\\mathbf{MVS}}" + }, + { + "bbox": [ + 130, + 502, + 480, + 550 + ], + "type": "text", + "content": ". The multi-view synthesis model is designed to generate multiple images conditioned on text prompts and camera poses, wherein we employ very recent work MVDream [41] as " + }, + { + "bbox": [ + 130, + 502, + 480, + 550 + ], + "type": "inline_equation", + "content": "M_{\\mathbf{MVS}}" + }, + { + "bbox": [ + 130, + 502, + 480, + 550 + ], + "type": "text", + "content": ". We compute the reconstruction loss for multiple generative views and rendered views:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 230, + 563, + 481, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 563, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 230, + 563, + 481, + 577 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {\\mathbf {M V S}} (\\tilde {\\mathbf {x}}, \\tilde {\\mathbf {c}}) = - | | M _ {\\mathbf {M V S}} (y, \\tilde {\\mathbf {c}}) - \\tilde {\\mathbf {x}} | | _ {2} ^ {2} \\tag {8}", + "image_path": "413646e589bccec3b830358627fab01169d1f06d51229fba9f066c6340bce1a1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "The smaller reconstruction loss signifies better geometric consistency within " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "M_{\\mathbf{MVS}}" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": ". These view-aware models measure the coherence across views according to their own 3D-aware insights. Incorporated with JSD, they provide distinct constraints, resulting in different 3D generations that nonetheless contribute to enhanced geometric consistency. We believe that JSD can be adapted to more universal view-aware models, which enables us to progressively redefine the benchmark of 3D generation with the advancement of multi-view tasks." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "JointDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 121, + 218, + 167 + ], + "blocks": [ + { + "bbox": [ + 133, + 121, + 218, + 167 + ], + "lines": [ + { + "bbox": [ + 133, + 121, + 218, + 167 + ], + "spans": [ + { + "bbox": [ + 133, + 121, + 218, + 167 + ], + "type": "image", + "image_path": "102b4a6ebb3b696522f4e2d89ea052289fe8e8ecb09e08d9c900169d2752a6fa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 218, + 167, + 430, + 174 + ], + "lines": [ + { + "bbox": [ + 218, + 167, + 430, + 174 + ], + "spans": [ + { + "bbox": [ + 218, + 167, + 430, + 174 + ], + "type": "text", + "content": "A pink Spiderman dancing ballet, Marvel character HD, highly detailed 3D model" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 220, + 121, + 307, + 166 + ], + "blocks": [ + { + "bbox": [ + 238, + 114, + 392, + 121 + ], + "lines": [ + { + "bbox": [ + 238, + 114, + 392, + 121 + ], + "spans": [ + { + "bbox": [ + 238, + 114, + 392, + 121 + ], + "type": "text", + "content": "a blue jay standing on a large basket of rainbow macarons" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 220, + 121, + 307, + 166 + ], + "lines": [ + { + "bbox": [ + 220, + 121, + 307, + 166 + ], + "spans": [ + { + "bbox": [ + 220, + 121, + 307, + 166 + ], + "type": "image", + "image_path": "3af7a027f877022bb7afa73cd7789c9983710496e07cac7a9cd7fb3cb47c2aa0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 120, + 391, + 166 + ], + "blocks": [ + { + "bbox": [ + 310, + 120, + 391, + 166 + ], + "lines": [ + { + "bbox": [ + 310, + 120, + 391, + 166 + ], + "spans": [ + { + "bbox": [ + 310, + 120, + 391, + 166 + ], + "type": "image", + "image_path": "1998689416a7d15587413fd8edaa95a77158f265e79006ff00bce141c91294ce.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 394, + 120, + 477, + 167 + ], + "blocks": [ + { + "bbox": [ + 394, + 120, + 477, + 167 + ], + "lines": [ + { + "bbox": [ + 394, + 120, + 477, + 167 + ], + "spans": [ + { + "bbox": [ + 394, + 120, + 477, + 167 + ], + "type": "image", + "image_path": "3bf4446cea63ec8c1e91bed4a1f1ea8f725aaf73698439af7c00f5c0f6e68594.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 133, + 167, + 218, + 218 + ], + "blocks": [ + { + "bbox": [ + 133, + 167, + 218, + 218 + ], + "lines": [ + { + "bbox": [ + 133, + 167, + 218, + 218 + ], + "spans": [ + { + "bbox": [ + 133, + 167, + 218, + 218 + ], + "type": "image", + "image_path": "01eac03f88f6d4290468fb46d98dfca58dd4a52be5656373f7c754c23eb28048.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 220, + 178, + 225 + ], + "lines": [ + { + "bbox": [ + 165, + 220, + 178, + 225 + ], + "spans": [ + { + "bbox": [ + 165, + 220, + 178, + 225 + ], + "type": "text", + "content": "SDS" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 227, + 173, + 306, + 220 + ], + "blocks": [ + { + "bbox": [ + 227, + 173, + 306, + 220 + ], + "lines": [ + { + "bbox": [ + 227, + 173, + 306, + 220 + ], + "spans": [ + { + "bbox": [ + 227, + 173, + 306, + 220 + ], + "type": "image", + "image_path": "69e36279d7daf173dbe8bab8cefae7d3f48863bbf6f2e02b3906265b591d5f1f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 220, + 280, + 227 + ], + "lines": [ + { + "bbox": [ + 252, + 220, + 280, + 227 + ], + "spans": [ + { + "bbox": [ + 252, + 220, + 280, + 227 + ], + "type": "text", + "content": "JSD w/ " + }, + { + "bbox": [ + 252, + 220, + 280, + 227 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{CLS}}" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 235, + 481, + 269 + ], + "lines": [ + { + "bbox": [ + 131, + 235, + 481, + 269 + ], + "spans": [ + { + "bbox": [ + 131, + 235, + 481, + 269 + ], + "type": "text", + "content": "Fig. 6: Ablations on JSD incorporated with different energy function " + }, + { + "bbox": [ + 131, + 235, + 481, + 269 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 131, + 235, + 481, + 269 + ], + "type": "text", + "content": ". CLS: Binary Classification model; I2I: Image-to-Image Translation model (Wonder3D [29]); MVS: Multi-View Image Synthesis model (MVDream [41])." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 173, + 391, + 220 + ], + "blocks": [ + { + "bbox": [ + 315, + 173, + 391, + 220 + ], + "lines": [ + { + "bbox": [ + 315, + 173, + 391, + 220 + ], + "spans": [ + { + "bbox": [ + 315, + 173, + 391, + 220 + ], + "type": "image", + "image_path": "d72f5478b2d4cf76bd21e429cf0f6c55e92504ce12e7261538d26229a13e2bb3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 220, + 367, + 227 + ], + "lines": [ + { + "bbox": [ + 340, + 220, + 367, + 227 + ], + "spans": [ + { + "bbox": [ + 340, + 220, + 367, + 227 + ], + "type": "text", + "content": "JSD w/C12I" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 399, + 173, + 478, + 220 + ], + "blocks": [ + { + "bbox": [ + 399, + 173, + 478, + 220 + ], + "lines": [ + { + "bbox": [ + 399, + 173, + 478, + 220 + ], + "spans": [ + { + "bbox": [ + 399, + 173, + 478, + 220 + ], + "type": "image", + "image_path": "1fc711d4cbf5a4eef051a0c6dc9d7104b6a863b16842ee8a70fe681190930047.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 420, + 220, + 449, + 227 + ], + "lines": [ + { + "bbox": [ + 420, + 220, + 449, + 227 + ], + "spans": [ + { + "bbox": [ + 420, + 220, + 449, + 227 + ], + "type": "text", + "content": "JSD w/C_MVS" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 292, + 303, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 292, + 303, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 292, + 303, + 304 + ], + "type": "text", + "content": "4.3 Framework of JointDreamer" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 311, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 478 + ], + "type": "text", + "content": "Building upon the JSD optimization, we propose the overall framework JointDreamer. The optimization is based on neural radiance field (NeRF) [31], adopting Instant-NGP [32] with volume renderer. We adopt the multi-view synthesis model " + }, + { + "bbox": [ + 130, + 311, + 482, + 478 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{MVS}}" + }, + { + "bbox": [ + 130, + 311, + 482, + 478 + ], + "type": "text", + "content": " as the energy function to integrate with JSD in JointDreamer. During optimization training, we utilize the common techniques including time-annealing and resolution scaling-up following [41, 42]. Besides, we propose two novel techniques to further enhance the generation quality, including a Geometry Fading scheme and a Classifier-Free Guidance Scale (CFG) Switching strategy. Geometry Fading. We aim to shift attention between geometric structure and texture details during optimization. Specifically, starting from iteration 5K, we reduce the learning rate of the density network of NeRF from " + }, + { + "bbox": [ + 130, + 311, + 482, + 478 + ], + "type": "inline_equation", + "content": "1e - 2" + }, + { + "bbox": [ + 130, + 311, + 482, + 478 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 311, + 482, + 478 + ], + "type": "inline_equation", + "content": "1e - 6" + }, + { + "bbox": [ + 130, + 311, + 482, + 478 + ], + "type": "text", + "content": " and set orientation loss to 0. Consequently, it benefits geometric convergence in the early phase of optimization, while allowing for decreased attention on geometry and increased attention on texture enhancement in the later stages." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 479, + 482, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 599 + ], + "type": "text", + "content": "CFG Switching. CFG scheduling strategies have been employed in 2D domain [15,39] to enhance quality. In this work, we propose to modify the CFG scale " + }, + { + "bbox": [ + 130, + 479, + 482, + 599 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 479, + 482, + 599 + ], + "type": "text", + "content": " during the training for 3D generation. We are motivated by the observation that a large CFG scale can lead to accelerated geometric convergence but may result in under-optimized geometry and distorted texture. Unlike the annealed CFG approach in CLIP-Sculptor [39], our increasing strategy prioritizes texture while maintaining accurate geometry. Specifically, a smaller " + }, + { + "bbox": [ + 130, + 479, + 482, + 599 + ], + "type": "inline_equation", + "content": "s = 30" + }, + { + "bbox": [ + 130, + 479, + 482, + 599 + ], + "type": "text", + "content": " is employed in the early stages to preserve shape integrity, which allows for stronger coherence guidance from JSD. After 5K iterations, we increase the " + }, + { + "bbox": [ + 130, + 479, + 482, + 599 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 479, + 482, + 599 + ], + "type": "text", + "content": " to 50, enhancing texture fidelity and overall quality." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 616, + 224, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 224, + 630 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 224, + 630 + ], + "type": "text", + "content": "5 Experiment" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "In this section, we present the text-to-3D generation results of JointDreamer with qualitative and quantitative evaluations, illustrating state-of-the-art per" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "C. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 249, + 114, + 378, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 114, + 378, + 121 + ], + "spans": [ + { + "bbox": [ + 249, + 114, + 378, + 121 + ], + "type": "text", + "content": "a rabbit cutting grass with a lawnmower" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 123, + 200, + 186 + ], + "blocks": [ + { + "bbox": [ + 133, + 123, + 200, + 186 + ], + "lines": [ + { + "bbox": [ + 133, + 123, + 200, + 186 + ], + "spans": [ + { + "bbox": [ + 133, + 123, + 200, + 186 + ], + "type": "image", + "image_path": "0371e9faf6e2e8e83e1302f14370651347026ce9da59593af65db1532a4204e1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 186, + 172, + 193 + ], + "lines": [ + { + "bbox": [ + 160, + 186, + 172, + 193 + ], + "spans": [ + { + "bbox": [ + 160, + 186, + 172, + 193 + ], + "type": "text", + "content": "SDS" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 198, + 480, + 220 + ], + "lines": [ + { + "bbox": [ + 131, + 198, + 480, + 220 + ], + "spans": [ + { + "bbox": [ + 131, + 198, + 480, + 220 + ], + "type": "text", + "content": "Fig. 7: Ablations on CFG Switching and Geometry Fading techniques in JointDreamer pipeline, demonstrating the effectiveness of quality enhancement." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 202, + 121, + 293, + 186 + ], + "blocks": [ + { + "bbox": [ + 202, + 121, + 293, + 186 + ], + "lines": [ + { + "bbox": [ + 202, + 121, + 293, + 186 + ], + "spans": [ + { + "bbox": [ + 202, + 121, + 293, + 186 + ], + "type": "image", + "image_path": "42c9d79ec65651d1664022af306902beb92bac090fbc22f4db7319ca0e7cbc2d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 186, + 258, + 193 + ], + "lines": [ + { + "bbox": [ + 227, + 186, + 258, + 193 + ], + "spans": [ + { + "bbox": [ + 227, + 186, + 258, + 193 + ], + "type": "text", + "content": "original JSD" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 294, + 121, + 386, + 186 + ], + "blocks": [ + { + "bbox": [ + 294, + 121, + 386, + 186 + ], + "lines": [ + { + "bbox": [ + 294, + 121, + 386, + 186 + ], + "spans": [ + { + "bbox": [ + 294, + 121, + 386, + 186 + ], + "type": "image", + "image_path": "cebe4868e9d9921597ed69400d942ae6a6489b2a36e5573d18a8247e7637d85e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 186, + 360, + 194 + ], + "lines": [ + { + "bbox": [ + 319, + 186, + 360, + 194 + ], + "spans": [ + { + "bbox": [ + 319, + 186, + 360, + 194 + ], + "type": "text", + "content": "+CFG Switching" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 387, + 121, + 480, + 186 + ], + "blocks": [ + { + "bbox": [ + 387, + 121, + 480, + 186 + ], + "lines": [ + { + "bbox": [ + 387, + 121, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 387, + 121, + 480, + 186 + ], + "type": "image", + "image_path": "3fdc32514adaa3be98d6b95b0af59b380c4d4cd51a361f0d083cda149af1bd97.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 186, + 456, + 194 + ], + "lines": [ + { + "bbox": [ + 422, + 186, + 456, + 194 + ], + "spans": [ + { + "bbox": [ + 422, + 186, + 456, + 194 + ], + "type": "text", + "content": "+ Geo Fading" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 244, + 479, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 244, + 479, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 479, + 268 + ], + "type": "text", + "content": "performance. We also make further ablation analysis on the proposed JSD. More training and evaluation details can be found in the Appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 285, + 276, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 285, + 276, + 296 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 276, + 296 + ], + "type": "text", + "content": "5.1 Text-to-3D Generation" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 137, + 338, + 331, + 390 + ], + "blocks": [ + { + "bbox": [ + 130, + 304, + 334, + 338 + ], + "lines": [ + { + "bbox": [ + 130, + 304, + 334, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 334, + 338 + ], + "type": "text", + "content": "Table 1: Quantitative results on textual consistency and user preference, tested on object-centric subset of MS-COCO [25]." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 338, + 331, + 390 + ], + "lines": [ + { + "bbox": [ + 137, + 338, + 331, + 390 + ], + "spans": [ + { + "bbox": [ + 137, + 338, + 331, + 390 + ], + "type": "table", + "html": "
MethodCLIP Score ↑R-Precision ↑User Study↑
DreamFusion [37]20.127.718.2
ProlificDreamer [42]25.018.716.2
MVDream [41]20.833.623.5
JointDreamer27.788.542.1
", + "image_path": "51f9d1580bf1f5e25e8157774556395ba0088d10d44e943c57071816056103c0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 343, + 338, + 473, + 390 + ], + "blocks": [ + { + "bbox": [ + 336, + 304, + 477, + 338 + ], + "lines": [ + { + "bbox": [ + 336, + 304, + 477, + 338 + ], + "spans": [ + { + "bbox": [ + 336, + 304, + 477, + 338 + ], + "type": "text", + "content": "Table 2: Ablation study on CFG Switching (CFGS) and Geometry Fading (GF)." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 343, + 338, + 473, + 390 + ], + "lines": [ + { + "bbox": [ + 343, + 338, + 473, + 390 + ], + "spans": [ + { + "bbox": [ + 343, + 338, + 473, + 390 + ], + "type": "table", + "html": "
SDS JSD CFGS GF|CLIP Score↑ FID↓
✓20.0 429.2
✓27.6 360.7
✓ ✓28.2 357.6
✓ ✓ ✓28.8 353.9
", + "image_path": "088b748911caa73510d1906497d96864c0a9c19f59804325ae7af353a6d86a43.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 391, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 666 + ], + "type": "text", + "content": "Qualitative Comparisons. We compare with several representative baselines on threestudio [11] project, which modularizes the text-to-3D framework allowing for fair comparisons by ablating individual components. The generative samples are shown in Fig. 5. (i) DreamFusion [37]: Compared to DreamFusion which utilizes traditional SDS optimization, JointDreamer significantly improves the Multi-Face Janus issues in the generations of DreamFusion by introducing an energy term to ensure inter-view coherence. (ii) Magic3D [24]: Magic3D introduces a two-stage generation pipeline, transferring NeRF to DMTet in the second stage to enhance generation quality. We also transfer NeRF to DMTet as JointDreamer-DMTet, showcasing consistent superiority in geometry and texture quality by utilizing JSD optimization instead of SDS. (iii) ProlificDreamer [42]: ProlificDreamer presents variational score distillation (VSD) as a variant of score distillation function. While VSD enhances photorealism in 3D renderings by introducing a LoRA model, the ill-posed association of pose and images during LoRA training deepens the geometric inconsistency of 3D representation, resulting in severe multi-view artifacts in generations. In contrast, JointDreamer with JSD achieves geometric consistency while maintaining high-fidelity texture quality. (iv) MVDream [41]: Compared to direct distillation from a finetuned model as MVDream, JointDreamer employs view-aware models as the coherence constraint in JSD, while still inherit the generalization capabilities of original diffusion models [38]. The results indicate that JointDreamer mitigates the overfitting issues of MVDream, which accurately responds to complex input text and enhances the 3D consistency of generations." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "JointDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": "Quantitative Comparisons. Following [18,37], we evaluate CLIP Score [14], CLIP R-Precision [36] and user preference on the object-centric caption subset of MS-COCO [25] with 153 prompts to measure the text congruence and generative quality. For computational efficiency, we generate each 3D asset using 5K iterations of " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": " rendered images and render 20 images per caption for evaluation. CLIP ViT-B/32 is adopted as the feature extractor for Clip Score and Clip R-Precision. As illustrated by the results in Table 1, JointDreamer can outperform all baselines on CLIP Score and CLIP R-Precision by large margins. Specifically, JointDreamer achieves an improvement of the R-Precision by " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "60.8\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "54.9\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": " over DreamFusion and MVDream, demonstrating its superior corresponding to textual description. Notably, the severe Janus artifacts in ProlificDreamer compromise the quality of rendering with noisy background and semantic distortion, resulting in the lowest R-Precision. We also conduct a user study about shape preference on these prompts in Table 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 304, + 253, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 253, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 253, + 316 + ], + "type": "text", + "content": "5.2 Ablation Analyses" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 323, + 481, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 323, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 323, + 481, + 335 + ], + "type": "text", + "content": "Ablations on Energy Functions. Sec. 4.2 discusses the varying interview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 335, + 315, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 315, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 315, + 453 + ], + "type": "text", + "content": "coherency measurements provided by view-aware models trained. When incorporated with JSD, these models have distinct impacts on 3D generations, as shown in Fig. 6. The binary classifier effectively corrects inaccurate geometric structures in SDS. However, it cannot introduce additional imaginative elements as a discriminative model, resulting in oversaturated and monotonous textures. In con" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 326, + 402, + 479, + 445 + ], + "blocks": [ + { + "bbox": [ + 321, + 346, + 482, + 401 + ], + "lines": [ + { + "bbox": [ + 321, + 346, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 321, + 346, + 482, + 401 + ], + "type": "text", + "content": "Table 3: Quantitative comparison of energy functions, showcasing the effectiveness of JSD in mitigating Janus artifacts with comparable computational efficiency to SDS." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 402, + 479, + 445 + ], + "lines": [ + { + "bbox": [ + 326, + 402, + 479, + 445 + ], + "spans": [ + { + "bbox": [ + 326, + 402, + 479, + 445 + ], + "type": "table", + "html": "
MethodsJanus Rate ↓ GPU Memory ↓ Train Time ↓
SDS100%16.1 G50 min.
JSD w/CCLS12.5%22.1 G80 min.
JSD w/C2I31.2%16.0 G119 min.
JSD w/C_MVS6.2%19.4 G54 min.
", + "image_path": "07988e616f80e452f474179d33e7e988608c31405acf13acb23e5d6287550fd9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 453, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 574 + ], + "type": "text", + "content": "trast, as generative models, Wonder3D [29] and MVDream [41] employ reconstruction loss to estimate 3D consistency. Hence, they not only guide geometric structural modifications but also influence texture quality. we further conduct a quantitative comparison using 16 complex multi-Janus prompts, as outlined in Table 3. The experimental setup details can be found in the Appendix. The results indicate that our JSD consistently mitigates Janus artifacts across different view-aware models, with only a slight increase in computational requirements compared to SDS. We find that the energy function " + }, + { + "bbox": [ + 130, + 453, + 482, + 574 + ], + "type": "inline_equation", + "content": "C_{12\\mathrm{I}}" + }, + { + "bbox": [ + 130, + 453, + 482, + 574 + ], + "type": "text", + "content": " derived from the imagedto-image translation model exhibits poor performance, likely attributed to a mismatch in camera range and inaccurate translated results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Analysis on Geometry Fading and CFG Switching Mechanisms. We conduct incremental ablations on our proposed techniques in JointDreamer, including the Geometry Fading scheme and CFG Switching strategy. As shown in Fig. 7, increasing the CFG value enhances texture detail compared to the original JSD, but over-optimizes the geometry, resulting in a bumpier shape. Compared to “+CFG Switching”, the Geometry fading effectively protects shape when larger CFG guidance. We also conduct a quantitative evaluation on a " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "30\\%" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "C. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 480, + 232 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 480, + 232 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 480, + 232 + ], + "type": "image", + "image_path": "e157a862a628d12ee2b43482add96ef10c550bad5e80f4812bd18bd605ff4218.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 237, + 480, + 260 + ], + "lines": [ + { + "bbox": [ + 131, + 237, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 131, + 237, + 480, + 260 + ], + "type": "text", + "content": "Fig. 9: Discuss the impact of prompts and random seeds, demonstrating the robustness of JointDreamer. " + }, + { + "bbox": [ + 131, + 237, + 480, + 260 + ], + "type": "inline_equation", + "content": "/..." + }, + { + "bbox": [ + 131, + 237, + 480, + 260 + ], + "type": "text", + "content": " means the same content as the prompt in Fig. 7." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 281, + 479, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 479, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 479, + 317 + ], + "type": "text", + "content": "MS-COCO subset in Table 2. JSD demonstrates superior texture quality compared to SDS, as evidenced by the Clip Score (CS) and FID metrics. The two proposed mechanisms further enhance the texture quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 322, + 480, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 322, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 131, + 322, + 480, + 334 + ], + "type": "text", + "content": "Discussions on Training Loss. To make further comparisons with JSD" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 335, + 298, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 298, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 298, + 453 + ], + "type": "text", + "content": "and SDS, we aggregate training losses from multiple prompts on two optimization functions and visualize the training loss curve as illustrated in Fig. 8. We observe that SDS experiences significant fluctuations due to the randomness of single-view optimization. In contrast, JSD can converge gradually and smoothly, demonstrating the introduction of multi" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 337, + 481, + 406 + ], + "blocks": [ + { + "bbox": [ + 306, + 337, + 481, + 406 + ], + "lines": [ + { + "bbox": [ + 306, + 337, + 481, + 406 + ], + "spans": [ + { + "bbox": [ + 306, + 337, + 481, + 406 + ], + "type": "image", + "image_path": "bbe2af54d2d91285ab5177d2cb8a553833888e958203ae0ca9e1017c2c00c2c4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 411, + 482, + 445 + ], + "lines": [ + { + "bbox": [ + 304, + 411, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 482, + 445 + ], + "type": "text", + "content": "Fig.8: Training loss comparisons of JSD and SDS. JSD eliminates randomness fluctuations in SDS convergence." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 454, + 479, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 479, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 479, + 478 + ], + "type": "text", + "content": "view optimization with inter-view coherence in JSD can reduce the randomness of optimization and contribute to better convergence for 3D representation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 483, + 479, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 479, + 533 + ], + "type": "text", + "content": "Discussions on Robustness for Prompts and Random Seed. To show the impact of seed and prompt, we modify seeds and key prompt components in Fig. 7, such as subject, object and verb. Results in Fig. 9 demonstrate the robustness of JointDreamer for different seed. Note that the default seed is 0." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 536, + 480, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 480, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 480, + 635 + ], + "type": "text", + "content": "The Effectiveness of Classification Model. The classification model surpasses MVDream in training speed by a factor of 48. Fig.4 (b) provides additional high-quality results obtained using JSD in conjunction with the binary classification model " + }, + { + "bbox": [ + 130, + 536, + 480, + 635 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{CLS}}" + }, + { + "bbox": [ + 130, + 536, + 480, + 635 + ], + "type": "text", + "content": ". Furthermore, we conduct an ablation study by replacing " + }, + { + "bbox": [ + 130, + 536, + 480, + 635 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\mathcal{C}(\\hat{\\mathbf{x}})}{\\partial\\mathbf{x}_i^i}" + }, + { + "bbox": [ + 130, + 536, + 480, + 635 + ], + "type": "text", + "content": " in Eq.5 with a randomly generated value between [0,1]. However, it yields shapeless results due to unrelated disturbances in the optimization process. These findings highlight the significance of the proposed classification model in achieving a balance between computational efficiency and generation quality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "content": "More Comparison with MVDream. View-aware models can be incorporated into 3D generation by combining them with JSD or using them directly with" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "JointDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 476, + 210 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 476, + 210 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 476, + 210 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 476, + 210 + ], + "type": "image", + "image_path": "5449a51b441ed97c9047f6bbe0da35a093cb2df9d2136515ff4b5f87a604bbc0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 213, + 482, + 248 + ], + "lines": [ + { + "bbox": [ + 131, + 213, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 482, + 248 + ], + "type": "text", + "content": "Fig. 10: Comparison of JointDreamer and SDS+MVDream, illustrating the superiority of JointDreamer in mitigating the Janus problem. SDS+MVDream consistently exhibits semantic missing or Janus issues with different weight options." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 266, + 479, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 479, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 479, + 326 + ], + "type": "text", + "content": "SDS. In this discussion, we use MVDream [41] as an example to demonstrate the superiority of JSD. MVDream can generate consistent shapes when calculating the SDS loss directly. However, it may miss components in the complete input text due to its fine-tuning on limited 3D data, as illustrated in Fig. 5. To address this limitation, a straightforward approach is to \"SDS+MVDream\"." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 327, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 460 + ], + "type": "text", + "content": "However, balancing the impact of SDS and MVDream is challenging. As shown in Fig. 10, setting " + }, + { + "bbox": [ + 130, + 327, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{SDS}}" + }, + { + "bbox": [ + 130, + 327, + 482, + 460 + ], + "type": "text", + "content": " to 0 degrades to MVDream alone, while setting " + }, + { + "bbox": [ + 130, + 327, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{MVDream}}" + }, + { + "bbox": [ + 130, + 327, + 482, + 460 + ], + "type": "text", + "content": " to 0 yields a combination similar to DreamFusion [37]. Achieving a balanced impact between SDS and MVDream through a simple combination is difficult. When " + }, + { + "bbox": [ + 130, + 327, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{MVDream}}" + }, + { + "bbox": [ + 130, + 327, + 482, + 460 + ], + "type": "text", + "content": " is large, textual consistency remains constrained, whereas decreasing " + }, + { + "bbox": [ + 130, + 327, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{MVDream}}" + }, + { + "bbox": [ + 130, + 327, + 482, + 460 + ], + "type": "text", + "content": " leads to the Janus problem. This is due to gradient misalignment across multiple views in the \"SDS+MVDream\" combination, as SDS lacks multi-view information and cannot derive our objective outlined in Eq. 5. In contrast, JSD based on the joint image distribution provides supervision for text consistency and high-fidelity texture. JSD also promotes inter-view consistency by introducing an energy term with the view-aware model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 480, + 220, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 480, + 220, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 480, + 220, + 493 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 509, + 491, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 491, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 491, + 581 + ], + "type": "text", + "content": "In this work, we introduce Joint Score Distillation (JSD) as a new paradigm for text-to-3D generation, which conducts multi-view optimization jointly and accounts for inter-view coherence. We demonstrate that JSD can significantly enhance 3D coherence while maintaining generalizability. With other proposed techniques, our overall framework, JointDreamer, is capable of geometric-consistent and high-fidelity 3D generation adhering to complex text input." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Limitations. While the training time of JointDreamer is comparable to existing SDS pipelines, there is room for improvement in terms of acceleration. Future work will explore alternative 3D representations, such as 3D Gaussian [20]. Additionally, JSD utilizes view-aware models to ensure geometry consistency and mitigate the impact of limited data. However, view-aware models still require 3D data for training, and further efficient 3D data collection or reconstruction from multi-view images is also worth investigating." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "C. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 140, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "text", + "content": "1. Armandpour, M., Zheng, H., Sadeghian, A., Sadeghian, A., Zhou, M.: Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. In: ICLR (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "type": "text", + "content": "2. Cao, Z., Hong, F., Wu, T., Pan, L., Liu, Z.: Large-vocabulary 3d diffusion model with transformer. In: ICLR (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 480, + 226 + ], + "type": "text", + "content": "3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV. pp. 9650-9660 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "type": "text", + "content": "4. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: CVPR. pp. 16123-16133 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "type": "text", + "content": "5. Chen, H., Gu, J., Chen, A., Tian, W., Tu, Z., Liu, L., Su, H.: Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction. arXiv preprint arXiv:2304.06714 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "type": "text", + "content": "6. Chen, R., Chen, Y., Jiao, N., Jia, K.: *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation*. In: ICCV. pp. 22246–22256 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "type": "text", + "content": "7. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objverse-xl: A universe of " + }, + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "type": "text", + "content": " objects. In: NeurIPS (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 359, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 359, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 359, + 481, + 392 + ], + "type": "text", + "content": "8. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objverse: A universe of annotated 3d objects. In: CVPR. pp. 13142-13153 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "type": "text", + "content": "9. Deng, Y., Yang, J., Xiang, J., Tong, X.: Gram: Generative radiance manifolds for 3d-aware image generation. In: CVPR. pp. 10673-10683 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 445 + ], + "type": "text", + "content": "10. Gao, J., Shen, T., Wang, Z., Chen, W., Yin, K., Li, D., Litany, O., Gojcic, Z., Fidler, S.: Get3d: A generative model of high quality 3d textured shapes learned from images. NeurIPS 35, 31841-31854 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 447, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 447, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 134, + 447, + 480, + 489 + ], + "type": "text", + "content": "11. Guo, Y.C., Liu, Y.T., Shao, R., Laforte, C., Voleti, V., Luo, G., Chen, C.H., Zou, Z.X., Wang, C., Cao, Y.P., Zhang, S.H.: threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 134, + 491, + 480, + 511 + ], + "type": "text", + "content": "12. Henderson, P., Ferrari, V.: Learning single-image 3d reconstruction by generative modelling of shape, pose and shading. IJCV 128(4), 835-854 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 134, + 512, + 480, + 533 + ], + "type": "text", + "content": "13. Henderson, P., Tsiminaki, V., Lampert, C.H.: Leveraging 2d data to learn textured 3d mesh generation. In: CVPR. pp. 7498-7507 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 134, + 534, + 480, + 555 + ], + "type": "text", + "content": "14. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: CLIPScore: a reference-free evaluation metric for image captioning. In: EMNLP (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 556, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 556, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 134, + 556, + 480, + 588 + ], + "type": "text", + "content": "15. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 589, + 480, + 621 + ], + "type": "text", + "content": "16. Hu, Z., Zhao, M., Zhao, C., Liang, X., Li, L., Zhao, Z., Fan, C., Zhou, X., Yu, X.: Efficientdreamer: High-fidelity and robust 3d creation via orthogonal-view diffusion priors. In: CVPR. pp. 4949-4958 (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 480, + 643 + ], + "type": "text", + "content": "17. Huang, Y., Wang, J., Shi, Y., Tang, B., Qi, X., Zhang, L.: Dreamtime: An improved optimization strategy for diffusion-guided 3d generation. In: ICLR (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "text", + "content": "18. Jain, A., Mildenhall, B., Barron, J.T., Abbeel, P., Poole, B.: Zero-shot text-guided object generation with dream fields. In: CVPR. pp. 867-876 (2022)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "JointDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "19. Jun, H., Nichol, A.: Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "text", + "content": "20. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4), 1-14 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "type": "text", + "content": "21. LeCun, Y., Chopra, S., Hadsell, R., Ranzato, M., Huang, F.: A tutorial on energy-based learning. Predicting structured data 1(0) (2006)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 192, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 480, + 213 + ], + "type": "text", + "content": "22. Li, M., Zhou, P., Liu, J.W., Keppo, J., Lin, M., Yan, S., Xu, X.: Instant3d: Instant text-to-3d generation. arXiv preprint arXiv:2311.08403 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 213, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 213, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 133, + 213, + 480, + 235 + ], + "type": "text", + "content": "23. Li, W., Chen, R., Chen, X., Tan, P.: Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. In: ICLR (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 235, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 235, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 133, + 235, + 480, + 266 + ], + "type": "text", + "content": "24. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: CVPR (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 266, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 266, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 133, + 266, + 480, + 311 + ], + "type": "text", + "content": "25. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 311, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 311, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 133, + 311, + 480, + 332 + ], + "type": "text", + "content": "26. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: ICCV. pp. 9298-9309 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 332, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 332, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 133, + 332, + 480, + 354 + ], + "type": "text", + "content": "27. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. In: ICLR (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 133, + 354, + 480, + 374 + ], + "type": "text", + "content": "28. Liu, Z., Feng, Y., Black, M.J., Nowrouzezahrai, D., Paull, L., Liu, W.: Meshdiffusion: Score-based generative 3d mesh modeling. In: ICLR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 374, + 480, + 407 + ], + "type": "text", + "content": "29. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3d: Single image to 3d using cross-domain diffusion. In: CVPR (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 407, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 407, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 133, + 407, + 480, + 438 + ], + "type": "text", + "content": "30. Luo, W., Hu, T., Zhang, S., Sun, J., Li, Z., Zhang, Z.: Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. NeurIPS 36 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 438, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 438, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 133, + 438, + 480, + 472 + ], + "type": "text", + "content": "31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 472, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 472, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 133, + 472, + 480, + 503 + ], + "type": "text", + "content": "32. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 503, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 503, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 133, + 503, + 480, + 536 + ], + "type": "text", + "content": "33. Nguyen-Phuoc, T., Li, C., Theis, L., Richardt, C., Yang, Y.L.: Hologan: Unsupervised learning of 3d representations from natural images. In: ICCV. pp. 7588-7597 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 133, + 536, + 480, + 568 + ], + "type": "text", + "content": "34. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 568, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 568, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 133, + 568, + 480, + 590 + ], + "type": "text", + "content": "35. Niemeyer, M., Geiger, A.: Giraffe: Representing scenes as compositional generative neural feature fields. In: CVPR. pp. 11453-11464 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 590, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 590, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 133, + 590, + 480, + 622 + ], + "type": "text", + "content": "36. Park, D.H., Azadi, S., Liu, X., Darrell, T., Rohrbach, A.: Benchmark for compositional text-to-image synthesis. In: NeurIPS Datasets and Benchmarks Track (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 622, + 480, + 643 + ], + "type": "text", + "content": "37. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: ICLR (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 133, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 480, + 666 + ], + "type": "text", + "content": "38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "C. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 336 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "39. Sanghi, A., Fu, R., Liu, V., Willis, K.D., Shayani, H., Khasahmadi, A.H., Sridhar, S., Ritchie, D.: Clip-sculptor: Zero-shot generation of high-fidelity and diverse shapes from natural language. In: CVPR. pp. 18339-18348 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "40. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "text", + "content": "41. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. In: ICLR (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "text", + "content": "42. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In: NeurIPS (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "text", + "content": "43. Watson, D., Chan, W., Martin-Brualla, R., Ho, J., Tagliasacchi, A., Norouzi, M.: Novel view synthesis with diffusion models. In: ICLR (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 259, + 482, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 304 + ], + "type": "text", + "content": "44. Weese, J., Kaus, M., Lorenz, C., Lobregt, S., Truyen, R., Pekar, V.: Shape constrained deformable models for 3d medical image segmentation. In: Information Processing in Medical Imaging: 17th International Conference, IPMI 2001 Davis, CA, USA, June 18-22, 2001 Proceedings 17. pp. 380-387. Springer (2001)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 304, + 482, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 336 + ], + "type": "text", + "content": "45. Zhao, W., Yan, L., Zhang, Y.: Geometric-constrained multi-view image matching method based on semi-global optimization. Geo-spatial information science 21(2), 115-126 (2018)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "JointDreamer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Just a Hint_ Point-Supervised Camouflaged Object Detection/6f40032b-910c-409c-99b0-064b71681182_content_list.json b/2024/Just a Hint_ Point-Supervised Camouflaged Object Detection/6f40032b-910c-409c-99b0-064b71681182_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a28e78ae38e96e66b355576fb612a2a61b6e6c66 --- /dev/null +++ b/2024/Just a Hint_ Point-Supervised Camouflaged Object Detection/6f40032b-910c-409c-99b0-064b71681182_content_list.json @@ -0,0 +1,1687 @@ +[ + { + "type": "text", + "text": "Just a Hint: Point-Supervised Camouflaged Object Detection", + "text_level": 1, + "bbox": [ + 254, + 140, + 750, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huafeng Chen $^{1,2}$ , Dian Shao $^{1,2}$ , Guangqian Guo $^{1,2}$ , and Shan Gao $^{1,2}$", + "bbox": [ + 236, + 210, + 764, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Unmanned System Research Institute, Northwestern Polytechnical University", + "bbox": [ + 236, + 239, + 764, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National Key Laboratory of Unmanned Aerial Vehicle Technology {chf,guogq21}@mail.nwpu.edu.cn, {shaodian,gaoshan}@nwpu.edu.cn", + "bbox": [ + 259, + 253, + 740, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Camouflaged Object Detection (COD) demands models to expeditiously and accurately distinguish objects which conceal themselves seamlessly in the environment. Owing to the subtle differences and ambiguous boundaries, COD is not only a remarkably challenging task for models but also for human annotators, requiring huge efforts to provide pixel-wise annotations. To alleviate the heavy annotation burden, we propose to fulfill this task with the help of only one point supervision. Specifically, by swiftly clicking on each object, we first adaptively expand the original point-based annotation to a reasonable hint area. Then, to avoid partial localization around discriminative parts, we propose an attention regulator to scatter model attention to the whole object through partially masking labeled regions. Moreover, to solve the unstable feature representation of camouflaged objects under only point-based annotation, we perform unsupervised contrastive learning based on differently augmented image pairs (e.g. changing color or doing translation). On three mainstream COD benchmarks, experimental results show that our model outperforms several weakly-supervised methods by a large margin across various metrics.", + "bbox": [ + 261, + 316, + 743, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Camouflaged Object Detection $\\cdot$ Point-supervised $\\cdot$ Unsupervised Contrastive Learning", + "bbox": [ + 259, + 580, + 740, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 633, + 375, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Camouflaged Object Detection (COD) aims to distinguish the disguised objects hiding carefully in the environment [8-10, 29]. Recently, it has attracted ever-growing research interest from the computer vision community and facilitates valuable real-life applications, such as search and rescue [8], species discovery [14], and medical image analysis [11, 12]. Specifically, we tentatively figure out two key-points for this task: 1) Quickly probing the existence of camouflaged objects, and 2) Accurately detecting the corresponding locations pixelwise within the given image. Towards this end, accurate pixel-level annotations are required, based on which several fully-supervised approaches achieve good performance [9, 25]. However, since spotting camouflaged objects is also difficult", + "bbox": [ + 212, + 664, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding Authors", + "bbox": [ + 230, + 825, + 410, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/57d61a1a1001791ba97ba3d5b3b87e3d92074ef775c0ab74fc32226a2d1df18b.jpg", + "image_caption": [ + "Fig. 1: Different types of annotation in camouflaged object detection task. The mask annotation takes about 60 minutes for each image. The scribbles take about 10 seconds but have diversity and boundary issues. The point takes just 2 seconds and only needs to point the most discriminative part of the camouflaged object." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 782, + 236 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for humans, such an annotation process is not only laborious but also time-consuming, e.g., 60 minutes for each image [9].", + "bbox": [ + 212, + 335, + 782, + 367 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, weakly-supervised COD has been explored recently. In [19], the authors try to use scribbles to replace pixel-wise annotations and propose a novel CRNet to solve this problem. But such a setting falls into a dilemma: 1) efficiency, the quality of scribble could not be ensured, arbitrary patterns could harm the learning process, 2) effectiveness, scribbles are expected to indicate the outline of boundaries, which also increases the annotation difficulty, as shown in Fig. 1.", + "bbox": [ + 212, + 368, + 784, + 473 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose to fulfill the COD task with only point-based supervision. Such an idea is motivated by the cognitive process of the human vision system: we tend to recognize the most \"discriminative part\" of one camouflaged object [25], and then scatter our attention to figure out the whole object mask. Even sometimes the camouflaged object is not perceived at first glance, we could quickly notice the target with just a little hint when a friend points to the image and says \"carefully scrutinize there!\". Importantly, the whole point-supervised annotation process only takes within 2 seconds.", + "bbox": [ + 212, + 474, + 787, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, supervision provided by a single point is rather limited, making the model easily collapse during training. In a similar but different task, Salient Object Detection (SOD), widely adopted solutions include detecting the edge map [16] first, or generating the prediction map [37] as the proposal region at the first stage. Unfortunately, all the previously mentioned solutions for generating proposal regions in SOD are not applicable in point-supervised COD, owing to the low contrast and ambiguous edges between the camouflaged objects and backgrounds (shown in S.M.). To address it, we explore a new point-to-region supervision generation strategy, hint area generator. Specifically, starting from the single point annotation, we adaptively expand it to a certain region, which obviously enriches the limited supervision, as shown in Fig. 2.", + "bbox": [ + 212, + 597, + 787, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, weakly-supervised COD methods suffer from partial detection, i.e., only discriminative parts of the object could gain adequate attention, while the whole shape is not accurately perceived, especially with the supervision of a single point. To make the model not only focus on local distinct ones, but nondiscriminatory parts, we propose an attention regulator module. By masking", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b396ff78523e13ee16cb6724538428cfff2ecf485919f1cd1f5167a4bce2ea29.jpg", + "image_caption": [ + "Fig. 2: Overview of our method. The first row shows the process of hint area generation from just a point to a hypothetical area. The second row shows the training process of attention regulator and presentation optimizer from the hint area to an accurate mask." + ], + "image_footnote": [], + "bbox": [ + 223, + 146, + 776, + 290 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "labeled region randomly, the model could scatter attention to the whole object, as shown in Fig. 2.", + "bbox": [ + 212, + 372, + 782, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, the learned feature representations with only point-based annotation tend to be unstable. As shown in Fig. 3, although the two images appear to be similar, the prediction accuracy largely varies. This is largely due to that the learned features are not robust enough. To obtain invariant patterns and stable representations during training, we apply Unsupervised Contrastive Learning (UCL) strategy on differently augmented data pairs. Such UCL process facilitates the model to learn more reliable feature representations under point supervision.", + "bbox": [ + 212, + 402, + 784, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, our contributions are as follows:", + "bbox": [ + 238, + 523, + 581, + 539 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel point-based learning paradigm for the challenging COD task, and construct the first weakly-supervised COD dataset with point annotation, P-COD, which contains 3040 images from the training set of COD10K and 1000 images from the training set of CAMO. Annotators are only required to swiftly click a point within the camouflaged object according to their first impression without localizing the whole object.", + "- We develop a point-supervised COD method by imitating the cognitive process of the human vision system, which contains several novel designs, 1) hint area generator, 2) attention regulator, and 3) representation optimizer.", + "- Experimental results show that our method outperforms existing weakly-supervised COD methods largely and even surpasses many fully-supervised methods. Moreover, when migrated to scribble-supervised COD and SOD tasks, our method also achieves competitive results." + ], + "bbox": [ + 225, + 553, + 785, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 775, + 387, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Weakly Supervised Camouflaged Detection. Pixel-level annotation in COD is very time-consuming, in order to reduce the cost of labeling, recent work has", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Just a Hint: Point-Supervised Camouflaged Object Detection", + "bbox": [ + 323, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/711ed9e88dd5c17064465aa157de7e0cb2dfa1181a112b05975902cd827af9e5.jpg", + "image_caption": [ + "Fig. 3: Comparison of prediction accuracy for similar images during training. Although two images $I_{1}$ and $I_{2}$ are very similar, their prediction accuracy trends are opposite. It is largely due to that the learned features are not robust enough in weak supervision." + ], + "image_footnote": [], + "bbox": [ + 223, + 145, + 777, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "begun to explore weakly supervised camouflaged object detection. CRNet [19] firstly tries to use scribble labels to train the model. However, scribble labels have the problems of label diversity and poor control of label quality. We try to use simpler and more natural point labels.", + "bbox": [ + 212, + 340, + 782, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Point Annotation. There have been several works on point annotations in weakly-supervised segmentation [1,4,32,34] and instance segmentation [2,24]. Semantic segmentation focuses on class-specific categories, which is different from COD task. PSOD [16] proposes a novel weakly-supervised Salient Object Detection (SOD) method using point supervision. But the drawbacks of it are also obvious: 1) extra edge detectors and edge information 2) flood fill algorithm introducing additional computation, and not appropriate for COD since camouflaged objects are not salient.", + "bbox": [ + 212, + 402, + 784, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contrastive Learning. The core of contrastive learning is to attract the positive sample pairs and repulse the negative sample pairs. Moco [18] maintained a queue of negative samples and turned one branch of siamese network into a momentum encoder to improve the consistency of the queue. BYOL [17] proposed a self-supervised image representation learning approach, achieving state-of-the-art results without negative samples. The mainstream paradigm of contrastive learning later shifted to only using positive samples. SimSiam [3] built an efficient and simple contrastive learning framework only using positive samples, which is free from the limitations of large batch size, momentum encoder, and negative samples. The above work [3,17,18] has proved that UCL can completely approach or even exceed the effect of supervised manner, and learn representations with strong generalization. Therefore, we try to introduce unsupervised contrastive learning to point-supervised COD to strengthen the representation of the model.", + "bbox": [ + 212, + 523, + 787, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 761, + 380, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall architecture of the proposed framework is shown in Fig. 4, which includes a) a hint area generator to extend supervision from a single point to the small hint region, b) an attention regulator to guide the model to focus", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/910730c2889d5f0baa168e045b9e7a6ce2ad29489ccc0f4c3d6791fe84d07ce4.jpg", + "image_caption": [ + "Fig. 4: Details of our method. The main components consist of three parts: (a) hint area generator, (b) attention regulator, and (c) representation optimizer. Hint area generator extends the point label from a single point to the small hint region. Attention regulator enforce the model to focus on the whole object instead of being stuck in most discriminative part. Representation optimizer uses the unsupervised contrastive learning to learn a stable feature representation for disentangling camouflaged objects and background." + ], + "image_footnote": [], + "bbox": [ + 223, + 145, + 777, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "on the whole object instead of being stuck in local discriminative parts, and c) a representation optimizer based on unsupervised contrastive learning for disentangling camouflaged objects and background.", + "bbox": [ + 212, + 531, + 787, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Hint Area Generator", + "text_level": 1, + "bbox": [ + 214, + 599, + 434, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared with other forms of mask supervision, supervision provided by a point is rather limited, which easily leads to model collapse, as shown in Tab. 5. To explain a little, an annotated point just functions like a \"hint\", only indicating the existence of the object, with no knowledge of the area, shape, or boundary. To solve this, we design a hint area generator to adaptively expand a single point to a relatively larger region. One of the most crucial things is to control the degree of point expansion, avoiding the resultant hint area being too small (limited effect) or too large (out of boundary). Especially avoiding excessive expansion to introduce noise, place quality over quantity, because incorrect supervision is fatal for WSCOD, as shown in Tab. 5.", + "bbox": [ + 212, + 628, + 787, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, as shown in Fig. 4(a), we initialize the desired annotation region with a small square area around the annotated point. It just prevents the model training from collapsing if only using one pixel as a point. The side length of the square area is $d$ . Then, after training the encoder $w$ epochs as a warm-", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Just a Hint: Point-Supervised Camouflaged Object Detection", + "bbox": [ + 323, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "start (Short-term training), we hope this encoder could have a certain extent of semantic sense towards the object. Based on this, we estimate the pseudo area size of the camouflaged object, which is calculated by: $P = E_w(I)$ , where $E_w$ denotes the encoder after training $w$ epochs, $I$ represents the input image, and $P$ is the obtained prediction map used to serve as the rough object area hint.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We choose circular to be the basic area shape, which could be intuitively regarded as \"a large point\". Then the key is to determine the reasonable radius of the circular area. Based on the object area hint $P$ , the initialized radius $r$ is defined as:", + "bbox": [ + 212, + 222, + 787, + 281 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nr = \\sqrt {\\frac {S u m (\\mathbb {I} (P > \\tau))}{N}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 281, + 785, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tau$ is a threshold, $\\mathbb{I}(\\cdot)$ is an indicator function, and $N$ is the number of objects. Note that $N$ could be achieved by counting the annotated point and thus is a deterministic parameter.", + "bbox": [ + 212, + 320, + 787, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, to prevent the calculated circular from being out of the object boundary, we use a hyperparameter, $\\alpha$ , to alter the radius $R$ using: $R = \\frac{r}{\\alpha}$ . Specifically, the annotation for an object $n$ is enlarged from a single point $(x_{n},y_{n})$ , to a small circle region around that point with the estimated radius $R$ , denoted as $C_{(x_n,y_n)}^R$ . It is noteworthy that the background is also annotated by such a small circular region, unfolding around the point $(x_b,y_b)$ , which is randomly picked from the background region.", + "bbox": [ + 212, + 366, + 787, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As a result, with the help of the hint area generator, the original point supervision for the image $I$ is enriched into several small regions, expanding supervision while minimizing the introduction of incorrect supervision as much as possible, represented as:", + "bbox": [ + 212, + 472, + 787, + 532 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS (I) = C _ {\\left(x _ {b}, y _ {b}\\right)} ^ {R} \\cup \\sum_ {n} C _ {\\left(x _ {n}, y _ {n}\\right)} ^ {R}, \\quad n = 1, \\dots , N \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 546, + 785, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $S(I)$ is the obtained supervision, as shown in Fig. 4.", + "bbox": [ + 214, + 588, + 640, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Attention Regulator", + "text_level": 1, + "bbox": [ + 214, + 628, + 431, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Weakly-supervised methods often suffer from partial detection problems, owing to the fact that model tends to focus too much on those discriminative regions. Inspired by the HaS method [21], the random mask is able to inhibit the strong response, we further design an attention regulator module for PCOD. Because in camouflaged scenes, people often annotate point labels on the most visually discriminative regions for the camouflaged objects, we leverage this prior to guide mask generation to scatter the attention from the directly supervised discriminative part to the whole object area, as shown in Fig. 4(b). Firstly, based on the given supervision region, $C^R$ , we could obtain a mask $M$ :", + "bbox": [ + 212, + 655, + 787, + 792 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nM = \\left\\{ \\begin{array}{l l} M _ {i} = h (Z), & i \\in C ^ {R} \\\\ M _ {i} = 1, & i \\notin C ^ {R} \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 804, + 785, + 844 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $Z$ is a logical matrix consisting of zeros and ones with the same shape as $I$ , $h(\\cdot)$ denotes a shuffle operation, $i$ is the pixel index.", + "bbox": [ + 214, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Then, we apply this mask to the original image $I$ : $I_{*} = I * M$ , where * indicates element-wise multiplication. Through such formulation, the annotated discriminative area will be randomly masked. During the training process, the model needs to explore the recognition of the surrounding foreground areas of the mask to restore recognition of the masked foreground area, helping the model scatter attention to other regions within the foreground objects.", + "bbox": [ + 212, + 176, + 787, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Representation Optimizer", + "text_level": 1, + "bbox": [ + 214, + 286, + 475, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In COD, the foreground and the background share high similarity. The subtle difference sometimes comes from the texture abruption or the color inconsistency. Based on such observations, we hope at least from one aspect the discrepancy could be exposed.", + "bbox": [ + 212, + 308, + 782, + 368 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Toward this goal, we propose to use unsupervised contrastive learning strategy to optimize the feature space, aiming to make the foreground object distinguished from the background through the learned invariant patterns and stable representations.", + "bbox": [ + 212, + 369, + 784, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, as shown in Fig. 4(c), for image $I$ , we adopt two different visual transformations $T_{1}$ and $T_{2}$ , selecting from colorjitter, Gaussianblur, flip, etc. These visual transformations are able to change the appearance of images. For example, Gaussian blur will smooth the texture, and changing color is relatively harmless to the texture patterns. This process results in two transformed images $I^{1}$ and $I^{2}$ : $I^{1} = T_{1}(I)$ , and $I^{2} = T_{2}(I)$ .", + "bbox": [ + 214, + 430, + 782, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We encode the transformed images differently into two prediction maps $P^{(1)}$ and $P^{(2)}$ , denoting as $P^1 = g(f(I^1))$ , and $P^2 = f(I^2)$ , where $f$ represents the encoder to be learned, and $g$ is a small network cascaded over $f$ to increase encoding diversity. When a position-related or size-related transformation $T$ (e.g., scale, crop) is applied to the image $I^1$ , this $T$ should be applied to $P^2$ to be aligned with $P^1$ , and vice versa. Our objective is to minimize the distance between these two prediction maps:", + "bbox": [ + 212, + 520, + 787, + 626 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\min \\mathcal {D} (P ^ {1}, P ^ {2}) = \\sum_ {i} \\left| P _ {i} ^ {1} - P _ {i} ^ {2} \\right|, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 633, + 784, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $i$ is the pixel index. Here we follow the design of UCL, i.e., stopping the gradient (stopgrad) update at one end, so the contrastive loss function is defined as:", + "bbox": [ + 214, + 671, + 782, + 714 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\mathcal {D} \\left(P ^ {1}, \\operatorname {s t o p g r a d} \\left(P ^ {2}\\right)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 715, + 784, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The prediction gap using different transformations will be narrowed by minimizing the above loss, making the learned feature representation robust.", + "bbox": [ + 214, + 737, + 782, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Network", + "text_level": 1, + "bbox": [ + 214, + 787, + 333, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Encoder. The structure of our feature encoder is simple compared to previous work. Specifically, in order to capture long-distance feature dependency and", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Just a Hint: Point-Supervised Camouflaged Object Detection", + "bbox": [ + 323, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2316259abe1c3346c6cdfa99bfc7355e4b70e6c090de7cf62ff20082a99672fc.jpg", + "table_caption": [ + "Table 1: Quantitative comparison with state-of-the-arts on four benchmarks. \"F\", \"U\", \"S\", \"P\" denote fully-supervised, unsupervised, scribble-supervised, and point-supervised methods, respectively. The best results are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
MethodsSup.CAMO (250)COD10K (2026)NC4K (4121)
MAE\\( \\downarrow S_m \\)↑\\( E_m \\)↑\\( F_w^\\beta \\)↑MAE\\( \\downarrow S_m \\)↑\\( E_m \\)↑\\( F_w^\\beta \\)↑MAE\\( \\downarrow S_m \\)↑\\( E_m \\)↑\\( F_w^\\beta \\)↑
F3Net [38]F0.1090.7110.7410.5640.0510.7390.7950.5440.0690.7820.8250.706
CSNet [15]F0.0920.7710.7950.6410.0470.7780.8090.5690.0610.8190.8450.748
ITSD [45]F0.1020.7500.7790.6100.0510.7670.8080.5570.0640.8110.8450.729
MINet [30]F0.0900.7480.7910.6370.0420.770.8320.608----
PraNet [11]F0.0940.7690.8250.6630.0450.7890.8610.629----
UCNet [42]F0.0940.7390.7870.6400.0420.7760.8570.6330.0550.8130.8720.777
SINet [9]F0.0920.7450.8040.6440.0430.7760.8640.6310.0580.8080.8710.723
MGL-R [41]F0.0880.7750.8120.6730.0350.8140.8510.6660.0520.8330.8670.740
PFNet [27]F0.0850.7820.8410.6950.0400.8000.8770.6600.0530.8290.8870.745
UJSC [23]F0.0730.8000.8590.7280.0350.8090.8840.6840.0470.8420.8980.771
UGTR [39]F0.0860.7840.8220.6840.0360.8170.8520.6660.0520.8390.8740.747
ZoomNet [29]F0.0660.8200.8920.7520.0290.8380.9110.7290.0430.8530.8960.784
DUSD [44]U0.1660.5510.5940.3080.1070.5800.6460.276----
USPS [28]U0.2070.5680.6410.3990.1960.5190.5360.265----
SAM [20]U0.1320.6840.6870.6060.0500.7830.7980.7010.0780.7670.7760.696
SS [43]S0.1180.6960.7860.5620.0710.6840.7700.461----
SCSOD [40]S0.1020.7130.7950.6180.0550.7100.8050.546----
CRNet [19]S0.0920.7350.8150.6410.0490.7330.8320.5760.0630.7750.8550.688
OursP0.0740.7980.8720.7270.0420.7840.8590.6500.0510.8220.8890.748
", + "bbox": [ + 215, + 208, + 792, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "obtain multi-scale information, we use PVT [36] as the backbone, for an input image $I \\in \\mathbb{R}^{3 \\times H \\times W}$ , we put it into the backbone to get the output features $F_{eat_i}$ for the $i$ -th. Then, we get the multi-scale features $(F_{eat1}, F_{eat2}, F_{eat3}, F_{eat4})$ with $\\left(\\frac{1}{4}, \\frac{1}{8}, \\frac{1}{16}, \\frac{1}{32}\\right)$ resolution of input image $I$ . We downsize the channel dimension of $F_{eat_i}$ into 64 by using $3 \\times 3$ convolutional layers. Next, these feature maps are unified into the same size by an up-sampling operation, and combined through the concatenation. Finally the output map $\\hat{S} \\in \\mathbb{R}^{1 \\times W \\times H}$ is obtained by the $3 \\times 3$ convolution layer, and the detailed network structure is provided in supplementary materials.", + "bbox": [ + 212, + 482, + 787, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Loss. Compared to other weak-supervised methods [19,35,40,43], we have fewer losses, only two. The final loss $\\mathcal{L}$ includes contrastive loss $\\mathcal{L}_c$ and partial cross-entropy loss $\\mathcal{L}_{pce}$ , the former is defined in Eq. 5, and the latter is defined as:", + "bbox": [ + 212, + 619, + 787, + 667 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p c e} = - \\sum_ {i \\in \\tilde {S}} S _ {i} \\log (\\hat {S} _ {i}) + (1 - S _ {i}) \\log (1 - \\hat {S} _ {i}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 683, + 785, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\tilde{S}$ represents the labeled area of $S$ . $S_{i}$ is the true class of pixel $i$ , and $\\hat{S}_i$ is the prediction of pixel $i$ . So the final loss $\\mathcal{L}$ can be defined as:", + "bbox": [ + 212, + 729, + 787, + 762 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c} + \\mathcal {L} _ {p c e}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 447, + 777, + 785, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Even though we only use point label in training, our loss is generalizable to the other weakly-supervised COD models by contrastive learning.", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f976a0ff99afb29e8cf41d88adcb7c1c1111fcf55234bd329eccfdbd98a42d9b.jpg", + "image_caption": [ + "Fig. 5: Visual comparison with representative scribble-supervised and fully-supervised models." + ], + "image_footnote": [], + "bbox": [ + 267, + 143, + 733, + 359 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/0db3fda2cfadadb49837aa994fc8d63c0d9e43c39efd46f61a078c96388f3b86.jpg", + "table_caption": [ + "Table 2: Ablation on point position." + ], + "table_footnote": [], + "table_body": "
TypeMAE↓Sm↑Em↑Fwβ↑
Center0.0790.7890.8580.716
Random0.0760.7910.8690.722
Ours0.0740.7980.8720.727
", + "bbox": [ + 230, + 441, + 488, + 496 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/5ce5f7ae8abe58eda3acca947f8d6194944b68bc3b1b84bb01a5608a52a93632.jpg", + "table_caption": [ + "Table 3: Ablation on point numbers." + ], + "table_footnote": [], + "table_body": "
Number of pointsMAE↓Sm↑Em↑Fwβ↑
One-point0.0760.7910.8690.722
Two-points0.0740.8020.8810.737
Three-points0.0740.8050.8870.741
", + "bbox": [ + 500, + 441, + 787, + 496 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 526, + 374, + 544 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 214, + 558, + 426, + 574 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Point Dataset. Our experiments are conducted on three COD benchmarks, CAMO [22], COD10K [9], and NC4K [25]. In order to evaluate our point-supervised COD method, we relabel 4040 images (3040 from COD10K, 1000 from CAMO) and propose the Point-supervised Dataset (P-COD) for training and the remaining is for testing. In Fig. 1, we show an example of point annotation and compare it with other annotation methods. For each image, we simulate the hunting process to label only one point for each camouflaged object, without needing to worry about ambiguous boundaries, making the process of the annotation easy and natural.", + "bbox": [ + 212, + 583, + 785, + 718 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Metrics. We adopt four evaluation metrics: mean absolute error (MAE), S-measure $(\\mathrm{S}_m)$ [6], E-measure $(\\mathrm{E}_m)$ [7], weighted F-measure $(\\mathrm{F}_{\\beta}^{w})$ [26]. Implementation Details. We implement our method with PyTorch and conduct experiments on one GeForce RTX4090 GPU. We use the stochastic gradient descent optimizer with a momentum of 0.9, a weight decay of $5e - 4$ , and triangle learning rate schedule with maximum learning rate $1e - 3$ . The batch size is 8, and the training epoch is 60. It takes around 7 hours to train. During training and inference, input images are resized to $512\\times 512$ .", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Just a Hint: Point-Supervised Camouflaged Object Detection", + "bbox": [ + 323, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/066f1f6f08500199256c7a4f23a01648e5a25afb09cecd3702bd5f44fa8c09b3.jpg", + "table_caption": [ + "Table 4: Comparison of parameters and MACs. * is reset in the same input size." + ], + "table_footnote": [], + "table_body": "
MethodsSup.Input sizePara(M)MACs(G)MAE↓Sm↑Em↑Fwβ↑
SINetF352248.9519.420.0920.7450.8040.644
CRNet*S352232.6514.270.0950.7300.8040.634
OursP352225.4410.220.0880.7610.8360.677
", + "bbox": [ + 217, + 171, + 785, + 241 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/79ae56ea26309a2111b9e413dd73c1bf3808c10834fbcc6279c17dd2430427aa.jpg", + "table_caption": [ + "Table 5: Effect of hint area generator with different settings. Point, I, P and SAM denote point label, initial square area, prediction map P and P obtained by SAM as proposal area, respectively." + ], + "table_footnote": [], + "table_body": "
SettingsMAE↓Sm↑Em↑Fwβ↑
w/ Pointcollapse
w/ I0.0920.7540.8020.655
w/ P0.1030.7260.8130.627
w/ SAM0.1080.6940.7710.603
w/ Ours0.0850.7590.8420.677
", + "bbox": [ + 228, + 356, + 493, + 431 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/81f8c8e04b8d493154f873e883557b38ace3d4c4638452b6ce96850b3247aa4a.jpg", + "table_caption": [ + "Table 6: Ablation study on contrastive losses. MSE, KL, COS, L1 mean Mean Square Error, Kullback-Leibler divergence, Consine similarity, and L1 loss, respectively." + ], + "table_footnote": [], + "table_body": "
LossMAE↓Sm↑Em↑Fwβ↑
w/o0.0850.7590.8420.677
MSE0.0790.7910.8510.708
KL0.0850.7690.8440.686
COS0.0870.7660.8440.680
L10.0740.7980.8720.727
", + "bbox": [ + 504, + 354, + 769, + 431 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Compare with State-of-the-arts", + "text_level": 1, + "bbox": [ + 215, + 467, + 521, + 483 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Quantitative Comparison. Being the first point supervised COD method, our proposed approach is primarily benchmarked against scribble-supervised and fully-supervised methods. As demonstrated in Tab. 1, our method achieves substantial improvements, with an average enhancement of $17.6\\%$ for MAE, $7.2\\%$ for Sm, $4.7\\%$ for Em, and $11.7\\%$ for $\\mathrm{F}_{\\beta}^{w}$ compared to the state-of-the-art weakly-supervised COD method CRNet [19]. This highlights our capability to achieve better performance with fewer annotations. Our approach even outperforms fully supervised methods in several metrics across multiple datasets. Specifically, it excels against the majority of fully-supervised methods on CAMO and outpaces nearly half of them on other datasets.", + "bbox": [ + 212, + 502, + 787, + 654 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Qualitative Evaluation. Our method produces prediction maps characterized by clearer and more complete object regions, along with sharper contours, significantly outperforming state-of-the-art weakly supervised COD method CRNet [19], as shown in Fig. 5. Our method performs well in various challenging scenarios, including scenarios with tiny objects (row 4), high intrinsic similarities (row 3, 5), indefinable boundaries (row 1), and complex backgrounds (row 2, 6).", + "bbox": [ + 212, + 656, + 787, + 747 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Parameter Complexity. With lower parameter complexity and minimal computational cost, our model (25.44M parameters and 10.22G MACs) consistently outperforms the fully supervised SINet (48.95M parameters and 19.42G MACs) by $4.3\\%$ for MAE, $2.1\\%$ for Sm, and $4.0\\%$ for Em on the challenging CAMO dataset. It also outperforms the scribble supervised CRNet (32.65M parameters and 14.27G MACs) by $7.4\\%$ for MAE, $4.2\\%$ for Sm, $4.0\\%$ for Em, and $6.8\\%$", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/706e18944d507ffe6f22dcf181193e39008fc227bda5db63aa63e3d5679427e0.jpg", + "image_caption": [ + "Fig. 6: Visualization of the component ablation study. I, II, III respectively represent hint area generator, representation optimizer, and attention regulator." + ], + "image_footnote": [], + "bbox": [ + 267, + 143, + 736, + 301 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/9f257ad084ebdf1169a766d674959e90c30b094ba22601f0a215bdd6f9ff3a5c.jpg", + "table_caption": [ + "Table 7: The ablation study for different augmentations in contrastive learning. \"S\", \"C\", \"F\", \"T\", \"R\", \"G\", \"J\" are Scale, Crop, Flip, Translate, Attention Regulator, Guassblur, Colorjitter." + ], + "table_footnote": [], + "table_body": "
SCFTRGJMAE↓Sm↑Em↑Fwβ↑
0.0850.7590.8420.677
✓0.0820.7790.8530.701
✓✓0.0800.7820.8600.713
✓✓✓0.0800.7710.8450.707
✓✓✓✓0.0770.7810.8530.711
✓✓✓✓✓0.0760.7900.8660.724
✓✓✓✓✓✓0.0740.7940.8670.729
✓✓✓✓✓✓✓0.0740.7980.8720.727
", + "bbox": [ + 264, + 409, + 738, + 513 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "for $\\mathrm{F}_{\\beta}^{w}$ , as shown in Tab. 4. These results underscore the superior lightweight performance of our model.", + "bbox": [ + 212, + 555, + 784, + 585 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 606, + 388, + 622 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As CAMO dataset is the most challenging one, indicated by the lowest scores in Tab. 1, all following ablation experiments are performed on it.", + "bbox": [ + 212, + 628, + 784, + 657 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ablation Study on Point Supervision. To reduce annotation costs, we intuitively annotated a point for each camouflaged object by humans. We conduct experiments on point positions, as shown in Tab. 2. For Random, we randomly choose a point from real labels, and repeat three times, taking the average of the results. which shows our method far exceeds the center annotation and random annotation under conditions of low annotation cost. We attribute it to the model's utilization of the priors implicit in point annotations to simulate human cognitive processes. In addition, we conduct experiments for point numbers (Random), as shown in Tab. 3. One point yields satisfactory performance. Increasing the points shows diminishing returns.", + "bbox": [ + 212, + 659, + 785, + 809 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effectiveness of Hint Area Generator. As shown in Tab. 5, our proposed hint area generator obtains a significant improvement over the direct use of point", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Just a Hint: Point-Supervised Camouflaged Object Detection", + "bbox": [ + 323, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/aaafe84bbaa85f4067d8ef9fc788c2a13b3bb890f85cda6112dedcb121cad936.jpg", + "table_caption": [ + "Table 8: Effect of different masking setting in Attention regulator." + ], + "table_footnote": [], + "table_body": "
Basic settingTypeMAE↓Sm↑Em↑FWβ↑
w/ aug (S, C, T, F)HaS [21]0.0780.7850.8640.711
w/ aug (S, C, T, F)Cutout [5]0.0760.7890.8620.721
w/ aug (S, C, T, F)Ours0.0760.7900.8660.724
", + "bbox": [ + 302, + 171, + 702, + 227 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/99006204fde38a373c62e1520f5124a76382f379d02e4b03d1a562c8ef2b2a73.jpg", + "table_caption": [ + "Table 9: Effect of key components in optimizer.", + "Table 10: Transferability studies on scribble datasets." + ], + "table_footnote": [], + "table_body": "
stop-grad predictorMAE↓Sm↑Em↑Fwβ↑MethodsSup.MAE↓Sm↑Em↑Fwβ↑
✓×unstableZoomNet [29]F0.0660.8200.8920.752
×✓0.0810.7680.8520.706CRNet [19]S0.0920.7350.8150.641
✓✓0.0740.7980.8720.727Ours + scribbleS0.0650.8160.8940.761
", + "bbox": [ + 222, + 303, + 785, + 357 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "label, Init square area and Prediction map. Because it has more supervision compared to the Init square area and fewer errors (e.g., background noise) compared to prediction map and prediction map obtained by SAM. Interestingly, the init square area performs much better than prediction map P and P obtained by SAM, and we think that the accuracy of supervision in WSCOD is more crucial than the quantity, since init square area has better accuracy and prediction map has high quantity in area generation but contains a lot of background noise. In addition, compared with prediction map, a relatively complete prediction map can be obtained by using hint area generator, as shown in Fig. 6.", + "bbox": [ + 212, + 395, + 784, + 532 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effectiveness of Attention Regulator. We propose attention regulator and apply it with UCL. As shown in Tab. 7, the results (Bold) are significantly improved after using attention regulator. In addition, to further demonstrate the superiority of regulator in weakly-supervised COD, we compare it with the previous mask approaches HaS [21] and Cutout [5], results are shown in Tab. 8.", + "bbox": [ + 212, + 534, + 785, + 611 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effectiveness of Representation Optimizer. We conduct ablation experiments for representation optimizer. Firstly, we test different types of contrastive loss and find that L1 loss performs best, with a significant improvement over the baseline, as shown in Tab. 6. In addition, we conduct exhaustive experiments for data augmentation, which is important for UCL in representation optimizer. Specifically, the augmentation is divided into three aspects, color-texture related (ColorJitter, GuassBlur), position related (Flip, Translate), and size related (Crop, Scale). It shows that color-texture and size augmentation can improve performance, but position augmentation yields unsatisfactory results, as shown in Tab. 7. We employ UCL exclusively with positive samples and test the key components in ablation study, as shown in Tab. 9. Fig. 6 intuitively illustrates that the representation optimizer enhances the precision of prediction maps. Fig. 7 and 8 show that our model is able to continuously optimize representation through representation optimizer and separate entangled objects from background, making the model eventually learn robust representations.", + "bbox": [ + 212, + 613, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/c40725fee41b6903bf12d2b0463e75f9c75bc6c1712266aefbd1343c40192b0d.jpg", + "table_caption": [ + "Table 11: The impact of $\\tau$ , $\\alpha$ , $w$ , and $d$ in the hint area generator." + ], + "table_footnote": [], + "table_body": "
dMAE↓Sm↑τMAE↓Sm↑αMAE↓Sm↑wMAE↓Sm↑
50.0850.7581000.0860.75730.0910.73950.0890.755
100.0850.7591500.0850.75940.0850.759100.0870.759
150.0860.7602000.0850.75850.0920.736150.0850.759
", + "bbox": [ + 250, + 171, + 753, + 224 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1a23bc3f81a574b984a8002492b824a021489b79aaae3996aa22caade3a4c6a0.jpg", + "table_caption": [ + "Table 12: Quantitative comparison with state-of-the-arts on four popular SOD benchmarks. \"F\", \"S\", \"P\" denote fully-, scribble-, and point-supervised methods, respectively. The best are in bold." + ], + "table_footnote": [], + "table_body": "
MethodsSup.ECSSDDUT-OHKU-ISDUTS-TE
MAE↓Sm↑FβmaxMAE↓Sm↑FβmaxMAE↓Sm↑FβmaxMAE↓Sm↑Fβmax
AFNet [13]F0.0420.9130.9350.0570.8260.7970.0360.9050.9230.0460.8670.863
BASNet [33]F0.0370.9160.9430.0570.8360.8050.0320.9090.9280.0480.8660.859
MFNet [31]S0.0840.8340.8790.0870.7410.7060.0590.8460.8760.7650.7740.770
SCSOD [40]S0.0490.8810.9140.0600.8110.7820.0380.8820.9080.0490.8530.858
PSOD [16]P0.0360.9130.9350.0640.8240.8080.0330.9010.9230.0450.8530.858
OursP0.0450.8950.9150.0590.8260.8240.0310.9060.9260.0420.8640.861
", + "bbox": [ + 215, + 318, + 802, + 414 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Impact of the Hyperparameter $\\tau, \\alpha, w, d$ . Four hyperparameters exist in hint area generator: $\\tau, \\alpha, w, d$ . We test diverse values for these parameters, as shown in Tab. 11. So we set $\\tau = 150$ , $\\alpha = 4$ , $w = 15$ and $d = 10$ . It is worth noting that the values of $\\tau, w$ , and $d$ do not have much impact on the results with different settings. In addition, the $\\tau$ can be replaced with unsupervised k-means clustering method (in S.M.). Our model is robust to parameters. It is because the hint area generator is designed to approximate the size of the camouflaged object through the prediction map. Whether the contour is fine and details are perfect, in the prediction map, has little influence on the estimation of the object size. The accurate estimation and localization of the camouflaged object are explored by the attention regulator and the representation optimizer modules.", + "bbox": [ + 212, + 458, + 787, + 625 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Transferability Studies", + "text_level": 1, + "bbox": [ + 214, + 647, + 450, + 662 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Extension to Scribble. We apply our method to the scribble dataset, and the results are shown in Tab. 10 (More results in supplementary materials) reveal that our model, utilizing scribble labeling, achieves competitive results compared to the state-of-the-art fully supervised method ZoomNet [29]. It showcases the generalization ability of our method.", + "bbox": [ + 212, + 672, + 787, + 748 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Extension to SOD task. Our method excels not only in COD but also demonstrates competitive performance in SOD, as shown in Tab. 12. Considering the difference between these two tasks, we may wonder why our method consistently performs well on such two seemingly different tasks. We attribute this to the generality and rationality of the designed structure. In fact, point-supervised SOD and COD have a clear commonality, i.e., the demands of sufficient and accurate", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Just a Hint: Point-Supervised Camouflaged Object Detection", + "bbox": [ + 323, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d350be7a340903c426077ae45634faaf88c5e47c202d044adc3c808396ac563f.jpg", + "image_caption": [ + "Fig. 7: Visualization of the feature space. Entangled features from foregrounds and backgrounds could be well separated by representation optimizer (visualized by t-SNE)." + ], + "image_footnote": [], + "bbox": [ + 269, + 143, + 733, + 236 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/6f27c4442cf677f4aec15f1aa56003aeb49177f619a735755b40a3ee9999ff2c.jpg", + "image_caption": [ + "Fig. 8: Visualization of results in different iterations. As training proceeds, the entangled features from the foreground and background can be progressively separated by the representation optimizer (visualization by t-SNE)." + ], + "image_footnote": [], + "bbox": [ + 267, + 289, + 736, + 402 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "supervision information. Through an effective point label expansion strategy, our model can provide more supervision information than a single point and more accurate supervision information than pseudo-labels, aiding in the precise segmentation of objects. Additionally, the proposed UCL optimizes the representation of weakly-supervised models, reducing segmentation instability. Despite the saliency of objects in WSSOD, it can also benefit from UCL. All these components are built on the common demands of the two tasks, providing a solid foundation for performance.", + "bbox": [ + 212, + 484, + 787, + 606 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 628, + 359, + 643 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we proposed an effective solution for COD by utilizing point-based supervision. The introduction of the P-COD dataset, based on human cognitive processes, streamlines the annotation process. To overcome the challenge of single-point supervision, we designed a hint area generator bolstering supervision. Representation optimizer optimized model representation by UCL. The attention regulator promoted a holistic object perception, ensuring a comprehensive understanding of camouflaged objects without fixating on distinct parts. Experimental results show that our method outperforms existing weakly-supervised methods largely and even surpasses many fully-supervised ones.", + "bbox": [ + 212, + 657, + 787, + 794 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement. This work is supported by the National Natural Science Foundation of China(NSFC) under Grant 62306239 and the Young Talent Fund of Association for Science and Technology in Shaanxi, China.", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bearman, A., Russakovsky, O., Ferrari, V., Fei-Fei, L.: What's the point: Semantic segmentation with point supervision. In: European conference on computer vision. pp. 549-565. Springer (2016)", + "2. Benenson, R., Popov, S., Ferrari, V.: Large-scale interactive object segmentation with human annotators. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11700-11709 (2019)", + "3. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15750-15758 (2021)", + "4. Cheng, B., Parkhi, O., Kirillov, A.: Pointly-supervised instance segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2617-2626 (2022)", + "5. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017)", + "6. Fan, D.P., Cheng, M.M., Liu, Y., Li, T., Borji, A.: Structure-measure: A new way to evaluate foreground maps. In: Proceedings of the IEEE international conference on computer vision. pp. 4548-4557 (2017)", + "7. Fan, D.P., Gong, C., Cao, Y., Ren, B., Cheng, M.M., Borji, A.: Enhanced-alignment measure for binary foreground map evaluation. arXiv preprint arXiv:1805.10421 (2018)", + "8. Fan, D.P., Ji, G.P., Cheng, M.M., Shao, L.: Concealed object detection. IEEE transactions on pattern analysis and machine intelligence 44(10), 6024-6042 (2021)", + "9. Fan, D.P., Ji, G.P., Sun, G., Cheng, M.M., Shen, J., Shao, L.: Camouflaged object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2777-2787 (2020)", + "10. Fan, D.P., Ji, G.P., Xu, P., Cheng, M.M., Sakaridis, C., Van Gool, L.: Advances in deep concealed scene understanding. Visual Intelligence 1(1), 16 (2023)", + "1. Fan, D.P., Ji, G.P., Zhou, T., Chen, G., Fu, H., Shen, J., Shao, L.: Pranet: Parallel reverse attention network for polyp segmentation. In: International conference on medical image computing and computer-assisted intervention. pp. 263-273. Springer (2020)", + "2. Fan, D.P., Zhou, T., Ji, G.P., Zhou, Y., Chen, G., Fu, H., Shen, J., Shao, L.: Inf-net: Automatic Covid-19 lung infection segmentation from ct images. IEEE transactions on medical imaging 39(8), 2626-2637 (2020)", + "3. Feng, M., Lu, H., Ding, E.: Attentive feedback network for boundary-aware salient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1623-1632 (2019)", + "4. Pérez-de la Fuente, R., Delclos, X., Penalver, E., Speranza, M., Wierzchos, J., Ascaso, C., Engel, M.S.: Early evolution and ecology of camouflage in insects. Proceedings of the National Academy of Sciences 109(52), 21414-21419 (2012)", + "5. Gao, S.H., Tan, Y.Q., Cheng, M.M., Lu, C., Chen, Y., Yan, S.: Highly efficient salient object detection with 100k parameters. In: European Conference on Computer Vision. pp. 702-721. Springer (2020)", + "6. Gao, S., Zhang, W., Wang, Y., Guo, Q., Zhang, C., He, Y., Zhang, W.: Weakly-supervised salient object detection using point supervision. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 670-678 (2022)", + "7. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Just a Hint: Point-Supervised Camouflaged Object Detection", + "bbox": [ + 323, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020)", + "18. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)", + "19. He, R., Dong, Q., Lin, J., Lau, R.W.: Weakly-supervised camouflaged object detection with scribble annotations. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 781-789 (2023)", + "20. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)", + "21. Kumar Singh, K., Jae Lee, Y.: Hide-and-seek: Forcing a network to be meticulous for weakly-supervised object and action localization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3524-3533 (2017)", + "22. Le, T.N., Nguyen, T.V., Nie, Z., Tran, M.T., Sugimoto, A.: Anabranch network for camouflaged object segmentation. Computer vision and image understanding 184, 45-56 (2019)", + "23. Li, A., Zhang, J., Lv, Y., Liu, B., Zhang, T., Dai, Y.: Uncertainty-aware joint salient object and camouflaged object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10071-10081 (2021)", + "24. Liew, J., Wei, Y., Xiong, W., Ong, S.H., Feng, J.: Regional interactive image segmentation networks. In: 2017 IEEE international conference on computer vision (ICCV). pp. 2746-2754. IEEE (2017)", + "25. Lv, Y., Zhang, J., Dai, Y., Li, A., Liu, B., Barnes, N., Fan, D.P.: Simultaneously localize, segment and rank the camouflaged objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11591-11601 (2021)", + "26. Margolin, R., Zelnik-Manor, L., Tal, A.: How to evaluate foreground maps? In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 248-255 (2014)", + "27. Mei, H., Ji, G.P., Wei, Z., Yang, X., Wei, X., Fan, D.P.: Camouflaged object segmentation with distraction mining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8772-8781 (2021)", + "28. Nguyen, T., Dax, M., Mummadi, C.K., Ngo, N., Nguyen, T.H.P., Lou, Z., Brox, T.: Deepusps: Deep robust unsupervised saliency prediction via self-supervision. Advances in Neural Information Processing Systems 32 (2019)", + "29. Pang, Y., Zhao, X., Xiang, T.Z., Zhang, L., Lu, H.: Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 2160-2170 (2022)", + "30. Pang, Y., Zhao, X., Zhang, L., Lu, H.: Multi-scale interactive network for salient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9413-9422 (2020)", + "31. Piao, Y., Wang, J., Zhang, M., Lu, H.: Mfnet: Multi-filter directive network for weakly supervised salient object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4136-4145 (2021)", + "32. Qian, R., Wei, Y., Shi, H., Li, J., Liu, J., Huang, T.: Weakly supervised scene parsing with point-based distance metric learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 8843-8850 (2019)", + "33. Qin, X., Zhang, Z., Huang, C., Gao, C., Dehghan, M., Jagersand, M.: Basnet: Boundary-aware salient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7479-7489 (2019)" + ], + "bbox": [ + 215, + 147, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Shin, G., Xie, W., Albanie, S.: All you need are a few pixels: semantic segmentation with pixelpick. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1687-1697 (2021)", + "35. Wang, B., Zhao, Y., Yang, L., Long, T., Li, X.: Temporal action localization in the deep learning era: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "36. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 568-578 (2021)", + "37. Wei, J., Wang, Q., Li, Z., Wang, S., Zhou, S.K., Cui, S.: Shallow feature matters for weakly supervised object localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5993-6001 (2021)", + "38. Wei, J., Wang, S., Huang, Q.: $\\mathbf{F}^3$ net: fusion, feedback and focus for salient object detection. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 12321-12328 (2020)", + "39. Yang, F., Zhai, Q., Li, X., Huang, R., Luo, A., Cheng, H., Fan, D.P.: Uncertainty-guided transformer reasoning for camouflaged object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4146-4155 (2021)", + "40. Yu, S., Zhang, B., Xiao, J., Lim, E.G.: Structure-consistent weakly supervised salient object detection with local saliency coherence. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 3234-3242 (2021)", + "41. Zhai, Q., Li, X., Yang, F., Chen, C., Cheng, H., Fan, D.P.: Mutual graph learning for camouflaged object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12997-13007 (2021)", + "42. Zhang, J., Fan, D.P., Dai, Y., Anwar, S., Saleh, F.S., Zhang, T., Barnes, N.: Ucnet: Uncertainty inspired rgb-d saliency detection via conditional variational autoencoders. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8582-8591 (2020)", + "43. Zhang, J., Yu, X., Li, A., Song, P., Liu, B., Dai, Y.: Weakly-supervised salient object detection via scribble annotations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12546-12555 (2020)", + "44. Zhang, J., Zhang, T., Dai, Y., Harandi, M., Hartley, R.: Deep unsupervised saliency detection: A multiple noisy labeling perspective. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9029-9038 (2018)", + "45. Zhou, H., Xie, X., Lai, J.H., Chen, Z., Yang, L.: Interactive two-stream decoder for accurate and fast saliency detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9141-9150 (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 688 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Just a Hint: Point-Supervised Camouflaged Object Detection", + "bbox": [ + 323, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Just a Hint_ Point-Supervised Camouflaged Object Detection/6f40032b-910c-409c-99b0-064b71681182_model.json b/2024/Just a Hint_ Point-Supervised Camouflaged Object Detection/6f40032b-910c-409c-99b0-064b71681182_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cc306186eb8f6b3c3b7016e75d418e9cd155a723 --- /dev/null +++ b/2024/Just a Hint_ Point-Supervised Camouflaged Object Detection/6f40032b-910c-409c-99b0-064b71681182_model.json @@ -0,0 +1,2313 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.255, + 0.141, + 0.75, + 0.187 + ], + "angle": 0, + "content": "Just a Hint: Point-Supervised Camouflaged Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.212, + 0.765, + 0.229 + ], + "angle": 0, + "content": "Huafeng Chen\\(^{1,2}\\), Dian Shao\\(^{1,2}\\), Guangqian Guo\\(^{1,2}\\), and Shan Gao\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.24, + 0.765, + 0.254 + ], + "angle": 0, + "content": "1 Unmanned System Research Institute, Northwestern Polytechnical University" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.254, + 0.741, + 0.282 + ], + "angle": 0, + "content": "\\(^{2}\\) National Key Laboratory of Unmanned Aerial Vehicle Technology {chf,guogq21}@mail.nwpu.edu.cn, {shaodian,gaoshan}@nwpu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.318, + 0.744, + 0.568 + ], + "angle": 0, + "content": "Abstract. Camouflaged Object Detection (COD) demands models to expeditiously and accurately distinguish objects which conceal themselves seamlessly in the environment. Owing to the subtle differences and ambiguous boundaries, COD is not only a remarkably challenging task for models but also for human annotators, requiring huge efforts to provide pixel-wise annotations. To alleviate the heavy annotation burden, we propose to fulfill this task with the help of only one point supervision. Specifically, by swiftly clicking on each object, we first adaptively expand the original point-based annotation to a reasonable hint area. Then, to avoid partial localization around discriminative parts, we propose an attention regulator to scatter model attention to the whole object through partially masking labeled regions. Moreover, to solve the unstable feature representation of camouflaged objects under only point-based annotation, we perform unsupervised contrastive learning based on differently augmented image pairs (e.g. changing color or doing translation). On three mainstream COD benchmarks, experimental results show that our model outperforms several weakly-supervised methods by a large margin across various metrics." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.581, + 0.741, + 0.61 + ], + "angle": 0, + "content": "Keywords: Camouflaged Object Detection \\(\\cdot\\) Point-supervised \\(\\cdot\\) Unsupervised Contrastive Learning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.634, + 0.377, + 0.65 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.665, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Camouflaged Object Detection (COD) aims to distinguish the disguised objects hiding carefully in the environment [8-10, 29]. Recently, it has attracted ever-growing research interest from the computer vision community and facilitates valuable real-life applications, such as search and rescue [8], species discovery [14], and medical image analysis [11, 12]. Specifically, we tentatively figure out two key-points for this task: 1) Quickly probing the existence of camouflaged objects, and 2) Accurately detecting the corresponding locations pixelwise within the given image. Towards this end, accurate pixel-level annotations are required, based on which several fully-supervised approaches achieve good performance [9, 25]. However, since spotting camouflaged objects is also difficult" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.411, + 0.84 + ], + "angle": 0, + "content": "Corresponding Authors" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "H. Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.784, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.249, + 0.788, + 0.306 + ], + "angle": 0, + "content": "Fig. 1: Different types of annotation in camouflaged object detection task. The mask annotation takes about 60 minutes for each image. The scribbles take about 10 seconds but have diversity and boundary issues. The point takes just 2 seconds and only needs to point the most discriminative part of the camouflaged object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.784, + 0.368 + ], + "angle": 0, + "content": "for humans, such an annotation process is not only laborious but also time-consuming, e.g., 60 minutes for each image [9]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.785, + 0.474 + ], + "angle": 0, + "content": "To address this issue, weakly-supervised COD has been explored recently. In [19], the authors try to use scribbles to replace pixel-wise annotations and propose a novel CRNet to solve this problem. But such a setting falls into a dilemma: 1) efficiency, the quality of scribble could not be ensured, arbitrary patterns could harm the learning process, 2) effectiveness, scribbles are expected to indicate the outline of boundaries, which also increases the annotation difficulty, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.596 + ], + "angle": 0, + "content": "In this work, we propose to fulfill the COD task with only point-based supervision. Such an idea is motivated by the cognitive process of the human vision system: we tend to recognize the most \"discriminative part\" of one camouflaged object [25], and then scatter our attention to figure out the whole object mask. Even sometimes the camouflaged object is not perceived at first glance, we could quickly notice the target with just a little hint when a friend points to the image and says \"carefully scrutinize there!\". Importantly, the whole point-supervised annotation process only takes within 2 seconds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.764 + ], + "angle": 0, + "content": "However, supervision provided by a single point is rather limited, making the model easily collapse during training. In a similar but different task, Salient Object Detection (SOD), widely adopted solutions include detecting the edge map [16] first, or generating the prediction map [37] as the proposal region at the first stage. Unfortunately, all the previously mentioned solutions for generating proposal regions in SOD are not applicable in point-supervised COD, owing to the low contrast and ambiguous edges between the camouflaged objects and backgrounds (shown in S.M.). To address it, we explore a new point-to-region supervision generation strategy, hint area generator. Specifically, starting from the single point annotation, we adaptively expand it to a certain region, which obviously enriches the limited supervision, as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Moreover, weakly-supervised COD methods suffer from partial detection, i.e., only discriminative parts of the object could gain adequate attention, while the whole shape is not accurately perceived, especially with the supervision of a single point. To make the model not only focus on local distinct ones, but nondiscriminatory parts, we propose an attention regulator module. By masking" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Just a Hint: Point-Supervised Camouflaged Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.147, + 0.777, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.301, + 0.785, + 0.343 + ], + "angle": 0, + "content": "Fig. 2: Overview of our method. The first row shows the process of hint area generation from just a point to a hypothetical area. The second row shows the training process of attention regulator and presentation optimizer from the hint area to an accurate mask." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.784, + 0.403 + ], + "angle": 0, + "content": "labeled region randomly, the model could scatter attention to the whole object, as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.785, + 0.524 + ], + "angle": 0, + "content": "Furthermore, the learned feature representations with only point-based annotation tend to be unstable. As shown in Fig. 3, although the two images appear to be similar, the prediction accuracy largely varies. This is largely due to that the learned features are not robust enough. To obtain invariant patterns and stable representations during training, we apply Unsupervised Contrastive Learning (UCL) strategy on differently augmented data pairs. Such UCL process facilitates the model to learn more reliable feature representations under point supervision." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.525, + 0.583, + 0.54 + ], + "angle": 0, + "content": "To summarize, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.554, + 0.787, + 0.645 + ], + "angle": 0, + "content": "- We propose a novel point-based learning paradigm for the challenging COD task, and construct the first weakly-supervised COD dataset with point annotation, P-COD, which contains 3040 images from the training set of COD10K and 1000 images from the training set of CAMO. Annotators are only required to swiftly click a point within the camouflaged object according to their first impression without localizing the whole object." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.646, + 0.785, + 0.691 + ], + "angle": 0, + "content": "- We develop a point-supervised COD method by imitating the cognitive process of the human vision system, which contains several novel designs, 1) hint area generator, 2) attention regulator, and 3) representation optimizer." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.692, + 0.785, + 0.751 + ], + "angle": 0, + "content": "- Experimental results show that our method outperforms existing weakly-supervised COD methods largely and even surpasses many fully-supervised methods. Moreover, when migrated to scribble-supervised COD and SOD tasks, our method also achieves competitive results." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.554, + 0.787, + 0.751 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.776, + 0.388, + 0.792 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Weakly Supervised Camouflaged Detection. Pixel-level annotation in COD is very time-consuming, in order to reduce the cost of labeling, recent work has" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "H. Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.146, + 0.778, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.27, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Fig. 3: Comparison of prediction accuracy for similar images during training. Although two images \\( I_{1} \\) and \\( I_{2} \\) are very similar, their prediction accuracy trends are opposite. It is largely due to that the learned features are not robust enough in weak supervision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.784, + 0.403 + ], + "angle": 0, + "content": "begun to explore weakly supervised camouflaged object detection. CRNet [19] firstly tries to use scribble labels to train the model. However, scribble labels have the problems of label diversity and poor control of label quality. We try to use simpler and more natural point labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.785, + 0.524 + ], + "angle": 0, + "content": "Point Annotation. There have been several works on point annotations in weakly-supervised segmentation [1,4,32,34] and instance segmentation [2,24]. Semantic segmentation focuses on class-specific categories, which is different from COD task. PSOD [16] proposes a novel weakly-supervised Salient Object Detection (SOD) method using point supervision. But the drawbacks of it are also obvious: 1) extra edge detectors and edge information 2) flood fill algorithm introducing additional computation, and not appropriate for COD since camouflaged objects are not salient." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.788, + 0.737 + ], + "angle": 0, + "content": "Contrastive Learning. The core of contrastive learning is to attract the positive sample pairs and repulse the negative sample pairs. Moco [18] maintained a queue of negative samples and turned one branch of siamese network into a momentum encoder to improve the consistency of the queue. BYOL [17] proposed a self-supervised image representation learning approach, achieving state-of-the-art results without negative samples. The mainstream paradigm of contrastive learning later shifted to only using positive samples. SimSiam [3] built an efficient and simple contrastive learning framework only using positive samples, which is free from the limitations of large batch size, momentum encoder, and negative samples. The above work [3,17,18] has proved that UCL can completely approach or even exceed the effect of supervised manner, and learn representations with strong generalization. Therefore, we try to introduce unsupervised contrastive learning to point-supervised COD to strengthen the representation of the model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.762, + 0.381, + 0.78 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The overall architecture of the proposed framework is shown in Fig. 4, which includes a) a hint area generator to extend supervision from a single point to the small hint region, b) an attention regulator to guide the model to focus" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Just a Hint: Point-Supervised Camouflaged Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.146, + 0.778, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.403, + 0.788, + 0.501 + ], + "angle": 0, + "content": "Fig. 4: Details of our method. The main components consist of three parts: (a) hint area generator, (b) attention regulator, and (c) representation optimizer. Hint area generator extends the point label from a single point to the small hint region. Attention regulator enforce the model to focus on the whole object instead of being stuck in most discriminative part. Representation optimizer uses the unsupervised contrastive learning to learn a stable feature representation for disentangling camouflaged objects and background." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.532, + 0.788, + 0.579 + ], + "angle": 0, + "content": "on the whole object instead of being stuck in local discriminative parts, and c) a representation optimizer based on unsupervised contrastive learning for disentangling camouflaged objects and background." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.601, + 0.436, + 0.616 + ], + "angle": 0, + "content": "3.1 Hint Area Generator" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Compared with other forms of mask supervision, supervision provided by a point is rather limited, which easily leads to model collapse, as shown in Tab. 5. To explain a little, an annotated point just functions like a \"hint\", only indicating the existence of the object, with no knowledge of the area, shape, or boundary. To solve this, we design a hint area generator to adaptively expand a single point to a relatively larger region. One of the most crucial things is to control the degree of point expansion, avoiding the resultant hint area being too small (limited effect) or too large (out of boundary). Especially avoiding excessive expansion to introduce noise, place quality over quantity, because incorrect supervision is fatal for WSCOD, as shown in Tab. 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Specifically, as shown in Fig. 4(a), we initialize the desired annotation region with a small square area around the annotated point. It just prevents the model training from collapsing if only using one pixel as a point. The side length of the square area is \\( d \\). Then, after training the encoder \\( w \\) epochs as a warm-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "H. Chen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "start (Short-term training), we hope this encoder could have a certain extent of semantic sense towards the object. Based on this, we estimate the pseudo area size of the camouflaged object, which is calculated by: \\( P = E_w(I) \\), where \\( E_w \\) denotes the encoder after training \\( w \\) epochs, \\( I \\) represents the input image, and \\( P \\) is the obtained prediction map used to serve as the rough object area hint." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.788, + 0.282 + ], + "angle": 0, + "content": "We choose circular to be the basic area shape, which could be intuitively regarded as \"a large point\". Then the key is to determine the reasonable radius of the circular area. Based on the object area hint \\( P \\), the initialized radius \\( r \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.282, + 0.786, + 0.314 + ], + "angle": 0, + "content": "\\[\nr = \\sqrt {\\frac {S u m (\\mathbb {I} (P > \\tau))}{N}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.321, + 0.789, + 0.367 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a threshold, \\(\\mathbb{I}(\\cdot)\\) is an indicator function, and \\(N\\) is the number of objects. Note that \\(N\\) could be achieved by counting the annotated point and thus is a deterministic parameter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.789, + 0.472 + ], + "angle": 0, + "content": "Then, to prevent the calculated circular from being out of the object boundary, we use a hyperparameter, \\(\\alpha\\), to alter the radius \\(R\\) using: \\(R = \\frac{r}{\\alpha}\\). Specifically, the annotation for an object \\(n\\) is enlarged from a single point \\((x_{n},y_{n})\\), to a small circle region around that point with the estimated radius \\(R\\), denoted as \\(C_{(x_n,y_n)}^R\\). It is noteworthy that the background is also annotated by such a small circular region, unfolding around the point \\((x_b,y_b)\\), which is randomly picked from the background region." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.789, + 0.534 + ], + "angle": 0, + "content": "As a result, with the help of the hint area generator, the original point supervision for the image \\(I\\) is enriched into several small regions, expanding supervision while minimizing the introduction of incorrect supervision as much as possible, represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.547, + 0.786, + 0.577 + ], + "angle": 0, + "content": "\\[\nS (I) = C _ {\\left(x _ {b}, y _ {b}\\right)} ^ {R} \\cup \\sum_ {n} C _ {\\left(x _ {n}, y _ {n}\\right)} ^ {R}, \\quad n = 1, \\dots , N \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.589, + 0.642, + 0.605 + ], + "angle": 0, + "content": "where \\(S(I)\\) is the obtained supervision, as shown in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.629, + 0.432, + 0.644 + ], + "angle": 0, + "content": "3.2 Attention Regulator" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.788, + 0.793 + ], + "angle": 0, + "content": "Weakly-supervised methods often suffer from partial detection problems, owing to the fact that model tends to focus too much on those discriminative regions. Inspired by the HaS method [21], the random mask is able to inhibit the strong response, we further design an attention regulator module for PCOD. Because in camouflaged scenes, people often annotate point labels on the most visually discriminative regions for the camouflaged objects, we leverage this prior to guide mask generation to scatter the attention from the directly supervised discriminative part to the whole object area, as shown in Fig. 4(b). Firstly, based on the given supervision region, \\( C^R \\), we could obtain a mask \\( M \\):" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.805, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\nM = \\left\\{ \\begin{array}{l l} M _ {i} = h (Z), & i \\in C ^ {R} \\\\ M _ {i} = 1, & i \\notin C ^ {R} \\end{array} \\right. \\tag {3}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Just a Hint: Point-Supervised Camouflaged Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "where \\( Z \\) is a logical matrix consisting of zeros and ones with the same shape as \\( I \\), \\( h(\\cdot) \\) denotes a shuffle operation, \\( i \\) is the pixel index." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.269 + ], + "angle": 0, + "content": "Then, we apply this mask to the original image \\( I \\): \\( I_{*} = I * M \\), where * indicates element-wise multiplication. Through such formulation, the annotated discriminative area will be randomly masked. During the training process, the model needs to explore the recognition of the surrounding foreground areas of the mask to restore recognition of the masked foreground area, helping the model scatter attention to other regions within the foreground objects." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.287, + 0.476, + 0.303 + ], + "angle": 0, + "content": "3.3 Representation Optimizer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.309, + 0.784, + 0.369 + ], + "angle": 0, + "content": "In COD, the foreground and the background share high similarity. The subtle difference sometimes comes from the texture abruption or the color inconsistency. Based on such observations, we hope at least from one aspect the discrepancy could be exposed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.785, + 0.429 + ], + "angle": 0, + "content": "Toward this goal, we propose to use unsupervised contrastive learning strategy to optimize the feature space, aiming to make the foreground object distinguished from the background through the learned invariant patterns and stable representations." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.431, + 0.784, + 0.52 + ], + "angle": 0, + "content": "Specifically, as shown in Fig. 4(c), for image \\( I \\), we adopt two different visual transformations \\( T_{1} \\) and \\( T_{2} \\), selecting from colorjitter, Gaussianblur, flip, etc. These visual transformations are able to change the appearance of images. For example, Gaussian blur will smooth the texture, and changing color is relatively harmless to the texture patterns. This process results in two transformed images \\( I^{1} \\) and \\( I^{2} \\): \\( I^{1} = T_{1}(I) \\), and \\( I^{2} = T_{2}(I) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.627 + ], + "angle": 0, + "content": "We encode the transformed images differently into two prediction maps \\( P^{(1)} \\) and \\( P^{(2)} \\), denoting as \\( P^1 = g(f(I^1)) \\), and \\( P^2 = f(I^2) \\), where \\( f \\) represents the encoder to be learned, and \\( g \\) is a small network cascaded over \\( f \\) to increase encoding diversity. When a position-related or size-related transformation \\( T \\) (e.g., scale, crop) is applied to the image \\( I^1 \\), this \\( T \\) should be applied to \\( P^2 \\) to be aligned with \\( P^1 \\), and vice versa. Our objective is to minimize the distance between these two prediction maps:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.635, + 0.785, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\min \\mathcal {D} (P ^ {1}, P ^ {2}) = \\sum_ {i} \\left| P _ {i} ^ {1} - P _ {i} ^ {2} \\right|, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.672, + 0.784, + 0.715 + ], + "angle": 0, + "content": "where \\( i \\) is the pixel index. Here we follow the design of UCL, i.e., stopping the gradient (stopgrad) update at one end, so the contrastive loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.717, + 0.785, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\mathcal {D} \\left(P ^ {1}, \\operatorname {s t o p g r a d} \\left(P ^ {2}\\right)\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.738, + 0.784, + 0.768 + ], + "angle": 0, + "content": "The prediction gap using different transformations will be narrowed by minimizing the above loss, making the learned feature representation robust." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.788, + 0.334, + 0.802 + ], + "angle": 0, + "content": "3.4 Network" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Encoder. The structure of our feature encoder is simple compared to previous work. Specifically, in order to capture long-distance feature dependency and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "H. Chen et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison with state-of-the-arts on four benchmarks. \"F\", \"U\", \"S\", \"P\" denote fully-supervised, unsupervised, scribble-supervised, and point-supervised methods, respectively. The best results are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.209, + 0.794, + 0.436 + ], + "angle": 0, + "content": "
MethodsSup.CAMO (250)COD10K (2026)NC4K (4121)
MAE\\( \\downarrow S_m \\)↑\\( E_m \\)↑\\( F_w^\\beta \\)↑MAE\\( \\downarrow S_m \\)↑\\( E_m \\)↑\\( F_w^\\beta \\)↑MAE\\( \\downarrow S_m \\)↑\\( E_m \\)↑\\( F_w^\\beta \\)↑
F3Net [38]F0.1090.7110.7410.5640.0510.7390.7950.5440.0690.7820.8250.706
CSNet [15]F0.0920.7710.7950.6410.0470.7780.8090.5690.0610.8190.8450.748
ITSD [45]F0.1020.7500.7790.6100.0510.7670.8080.5570.0640.8110.8450.729
MINet [30]F0.0900.7480.7910.6370.0420.770.8320.608----
PraNet [11]F0.0940.7690.8250.6630.0450.7890.8610.629----
UCNet [42]F0.0940.7390.7870.6400.0420.7760.8570.6330.0550.8130.8720.777
SINet [9]F0.0920.7450.8040.6440.0430.7760.8640.6310.0580.8080.8710.723
MGL-R [41]F0.0880.7750.8120.6730.0350.8140.8510.6660.0520.8330.8670.740
PFNet [27]F0.0850.7820.8410.6950.0400.8000.8770.6600.0530.8290.8870.745
UJSC [23]F0.0730.8000.8590.7280.0350.8090.8840.6840.0470.8420.8980.771
UGTR [39]F0.0860.7840.8220.6840.0360.8170.8520.6660.0520.8390.8740.747
ZoomNet [29]F0.0660.8200.8920.7520.0290.8380.9110.7290.0430.8530.8960.784
DUSD [44]U0.1660.5510.5940.3080.1070.5800.6460.276----
USPS [28]U0.2070.5680.6410.3990.1960.5190.5360.265----
SAM [20]U0.1320.6840.6870.6060.0500.7830.7980.7010.0780.7670.7760.696
SS [43]S0.1180.6960.7860.5620.0710.6840.7700.461----
SCSOD [40]S0.1020.7130.7950.6180.0550.7100.8050.546----
CRNet [19]S0.0920.7350.8150.6410.0490.7330.8320.5760.0630.7750.8550.688
OursP0.0740.7980.8720.7270.0420.7840.8590.6500.0510.8220.8890.748
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.789, + 0.619 + ], + "angle": 0, + "content": "obtain multi-scale information, we use PVT [36] as the backbone, for an input image \\(I \\in \\mathbb{R}^{3 \\times H \\times W}\\), we put it into the backbone to get the output features \\(F_{eat_i}\\) for the \\(i\\)-th. Then, we get the multi-scale features \\((F_{eat1}, F_{eat2}, F_{eat3}, F_{eat4})\\) with \\(\\left(\\frac{1}{4}, \\frac{1}{8}, \\frac{1}{16}, \\frac{1}{32}\\right)\\) resolution of input image \\(I\\). We downsize the channel dimension of \\(F_{eat_i}\\) into 64 by using \\(3 \\times 3\\) convolutional layers. Next, these feature maps are unified into the same size by an up-sampling operation, and combined through the concatenation. Finally the output map \\(\\hat{S} \\in \\mathbb{R}^{1 \\times W \\times H}\\) is obtained by the \\(3 \\times 3\\) convolution layer, and the detailed network structure is provided in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.789, + 0.669 + ], + "angle": 0, + "content": "Loss. Compared to other weak-supervised methods [19,35,40,43], we have fewer losses, only two. The final loss \\(\\mathcal{L}\\) includes contrastive loss \\(\\mathcal{L}_c\\) and partial cross-entropy loss \\(\\mathcal{L}_{pce}\\), the former is defined in Eq. 5, and the latter is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.684, + 0.787, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p c e} = - \\sum_ {i \\in \\tilde {S}} S _ {i} \\log (\\hat {S} _ {i}) + (1 - S _ {i}) \\log (1 - \\hat {S} _ {i}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.73, + 0.788, + 0.763 + ], + "angle": 0, + "content": "where \\(\\tilde{S}\\) represents the labeled area of \\(S\\). \\(S_{i}\\) is the true class of pixel \\(i\\), and \\(\\hat{S}_i\\) is the prediction of pixel \\(i\\). So the final loss \\(\\mathcal{L}\\) can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.449, + 0.778, + 0.787, + 0.795 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c} + \\mathcal {L} _ {p c e}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Even though we only use point label in training, our loss is generalizable to the other weakly-supervised COD models by contrastive learning." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Just a Hint: Point-Supervised Camouflaged Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.145, + 0.734, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.373, + 0.785, + 0.401 + ], + "angle": 0, + "content": "Fig. 5: Visual comparison with representative scribble-supervised and fully-supervised models." + }, + { + "type": "table_caption", + "bbox": [ + 0.235, + 0.416, + 0.484, + 0.43 + ], + "angle": 0, + "content": "Table 2: Ablation on point position." + }, + { + "type": "table", + "bbox": [ + 0.231, + 0.442, + 0.49, + 0.497 + ], + "angle": 0, + "content": "
TypeMAE↓Sm↑Em↑Fwβ↑
Center0.0790.7890.8580.716
Random0.0760.7910.8690.722
Ours0.0740.7980.8720.727
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.416, + 0.763, + 0.431 + ], + "angle": 0, + "content": "Table 3: Ablation on point numbers." + }, + { + "type": "table", + "bbox": [ + 0.5, + 0.442, + 0.789, + 0.497 + ], + "angle": 0, + "content": "
Number of pointsMAE↓Sm↑Em↑Fwβ↑
One-point0.0760.7910.8690.722
Two-points0.0740.8020.8810.737
Three-points0.0740.8050.8870.741
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.527, + 0.375, + 0.545 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.559, + 0.427, + 0.575 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.787, + 0.719 + ], + "angle": 0, + "content": "Point Dataset. Our experiments are conducted on three COD benchmarks, CAMO [22], COD10K [9], and NC4K [25]. In order to evaluate our point-supervised COD method, we relabel 4040 images (3040 from COD10K, 1000 from CAMO) and propose the Point-supervised Dataset (P-COD) for training and the remaining is for testing. In Fig. 1, we show an example of point annotation and compare it with other annotation methods. For each image, we simulate the hunting process to label only one point for each camouflaged object, without needing to worry about ambiguous boundaries, making the process of the annotation easy and natural." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Evaluation Metrics. We adopt four evaluation metrics: mean absolute error (MAE), S-measure \\((\\mathrm{S}_m)\\) [6], E-measure \\((\\mathrm{E}_m)\\) [7], weighted F-measure \\((\\mathrm{F}_{\\beta}^{w})\\) [26]. Implementation Details. We implement our method with PyTorch and conduct experiments on one GeForce RTX4090 GPU. We use the stochastic gradient descent optimizer with a momentum of 0.9, a weight decay of \\(5e - 4\\), and triangle learning rate schedule with maximum learning rate \\(1e - 3\\). The batch size is 8, and the training epoch is 60. It takes around 7 hours to train. During training and inference, input images are resized to \\(512\\times 512\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "H. Chen et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.23, + 0.145, + 0.772, + 0.16 + ], + "angle": 0, + "content": "Table 4: Comparison of parameters and MACs. * is reset in the same input size." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.172, + 0.787, + 0.242 + ], + "angle": 0, + "content": "
MethodsSup.Input sizePara(M)MACs(G)MAE↓Sm↑Em↑Fwβ↑
SINetF352248.9519.420.0920.7450.8040.644
CRNet*S352232.6514.270.0950.7300.8040.634
OursP352225.4410.220.0880.7610.8360.677
" + }, + { + "type": "table_caption", + "bbox": [ + 0.222, + 0.274, + 0.497, + 0.345 + ], + "angle": 0, + "content": "Table 5: Effect of hint area generator with different settings. Point, I, P and SAM denote point label, initial square area, prediction map P and P obtained by SAM as proposal area, respectively." + }, + { + "type": "table_caption", + "bbox": [ + 0.504, + 0.274, + 0.771, + 0.344 + ], + "angle": 0, + "content": "Table 6: Ablation study on contrastive losses. MSE, KL, COS, L1 mean Mean Square Error, Kullback-Leibler divergence, Consine similarity, and L1 loss, respectively." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.357, + 0.495, + 0.432 + ], + "angle": 0, + "content": "
SettingsMAE↓Sm↑Em↑Fwβ↑
w/ Pointcollapse
w/ I0.0920.7540.8020.655
w/ P0.1030.7260.8130.627
w/ SAM0.1080.6940.7710.603
w/ Ours0.0850.7590.8420.677
" + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.356, + 0.771, + 0.432 + ], + "angle": 0, + "content": "
LossMAE↓Sm↑Em↑Fwβ↑
w/o0.0850.7590.8420.677
MSE0.0790.7910.8510.708
KL0.0850.7690.8440.686
COS0.0870.7660.8440.680
L10.0740.7980.8720.727
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.468, + 0.522, + 0.484 + ], + "angle": 0, + "content": "4.2 Compare with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.788, + 0.655 + ], + "angle": 0, + "content": "Quantitative Comparison. Being the first point supervised COD method, our proposed approach is primarily benchmarked against scribble-supervised and fully-supervised methods. As demonstrated in Tab. 1, our method achieves substantial improvements, with an average enhancement of \\(17.6\\%\\) for MAE, \\(7.2\\%\\) for Sm, \\(4.7\\%\\) for Em, and \\(11.7\\%\\) for \\(\\mathrm{F}_{\\beta}^{w}\\) compared to the state-of-the-art weakly-supervised COD method CRNet [19]. This highlights our capability to achieve better performance with fewer annotations. Our approach even outperforms fully supervised methods in several metrics across multiple datasets. Specifically, it excels against the majority of fully-supervised methods on CAMO and outpaces nearly half of them on other datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.788, + 0.748 + ], + "angle": 0, + "content": "Qualitative Evaluation. Our method produces prediction maps characterized by clearer and more complete object regions, along with sharper contours, significantly outperforming state-of-the-art weakly supervised COD method CRNet [19], as shown in Fig. 5. Our method performs well in various challenging scenarios, including scenarios with tiny objects (row 4), high intrinsic similarities (row 3, 5), indefinable boundaries (row 1), and complex backgrounds (row 2, 6)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Parameter Complexity. With lower parameter complexity and minimal computational cost, our model (25.44M parameters and 10.22G MACs) consistently outperforms the fully supervised SINet (48.95M parameters and 19.42G MACs) by \\(4.3\\%\\) for MAE, \\(2.1\\%\\) for Sm, and \\(4.0\\%\\) for Em on the challenging CAMO dataset. It also outperforms the scribble supervised CRNet (32.65M parameters and 14.27G MACs) by \\(7.4\\%\\) for MAE, \\(4.2\\%\\) for Sm, \\(4.0\\%\\) for Em, and \\(6.8\\%\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Just a Hint: Point-Supervised Camouflaged Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.145, + 0.737, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.313, + 0.785, + 0.342 + ], + "angle": 0, + "content": "Fig. 6: Visualization of the component ablation study. I, II, III respectively represent hint area generator, representation optimizer, and attention regulator." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.355, + 0.785, + 0.397 + ], + "angle": 0, + "content": "Table 7: The ablation study for different augmentations in contrastive learning. \"S\", \"C\", \"F\", \"T\", \"R\", \"G\", \"J\" are Scale, Crop, Flip, Translate, Attention Regulator, Guassblur, Colorjitter." + }, + { + "type": "table", + "bbox": [ + 0.265, + 0.41, + 0.74, + 0.514 + ], + "angle": 0, + "content": "
SCFTRGJMAE↓Sm↑Em↑Fwβ↑
0.0850.7590.8420.677
✓0.0820.7790.8530.701
✓✓0.0800.7820.8600.713
✓✓✓0.0800.7710.8450.707
✓✓✓✓0.0770.7810.8530.711
✓✓✓✓✓0.0760.7900.8660.724
✓✓✓✓✓✓0.0740.7940.8670.729
✓✓✓✓✓✓✓0.0740.7980.8720.727
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.785, + 0.587 + ], + "angle": 0, + "content": "for \\(\\mathrm{F}_{\\beta}^{w}\\), as shown in Tab. 4. These results underscore the superior lightweight performance of our model." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.607, + 0.389, + 0.623 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.785, + 0.659 + ], + "angle": 0, + "content": "As CAMO dataset is the most challenging one, indicated by the lowest scores in Tab. 1, all following ablation experiments are performed on it." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.787, + 0.81 + ], + "angle": 0, + "content": "Ablation Study on Point Supervision. To reduce annotation costs, we intuitively annotated a point for each camouflaged object by humans. We conduct experiments on point positions, as shown in Tab. 2. For Random, we randomly choose a point from real labels, and repeat three times, taking the average of the results. which shows our method far exceeds the center annotation and random annotation under conditions of low annotation cost. We attribute it to the model's utilization of the priors implicit in point annotations to simulate human cognitive processes. In addition, we conduct experiments for point numbers (Random), as shown in Tab. 3. One point yields satisfactory performance. Increasing the points shows diminishing returns." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Effectiveness of Hint Area Generator. As shown in Tab. 5, our proposed hint area generator obtains a significant improvement over the direct use of point" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "H. Chen et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.277, + 0.145, + 0.725, + 0.16 + ], + "angle": 0, + "content": "Table 8: Effect of different masking setting in Attention regulator." + }, + { + "type": "table", + "bbox": [ + 0.303, + 0.172, + 0.703, + 0.228 + ], + "angle": 0, + "content": "
Basic settingTypeMAE↓Sm↑Em↑FWβ↑
w/ aug (S, C, T, F)HaS [21]0.0780.7850.8640.711
w/ aug (S, C, T, F)Cutout [5]0.0760.7890.8620.721
w/ aug (S, C, T, F)Ours0.0760.7900.8660.724
" + }, + { + "type": "table_caption", + "bbox": [ + 0.223, + 0.262, + 0.496, + 0.289 + ], + "angle": 0, + "content": "Table 9: Effect of key components in optimizer." + }, + { + "type": "table_caption", + "bbox": [ + 0.505, + 0.262, + 0.768, + 0.29 + ], + "angle": 0, + "content": "Table 10: Transferability studies on scribble datasets." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.304, + 0.787, + 0.358 + ], + "angle": 0, + "content": "
stop-grad predictorMAE↓Sm↑Em↑Fwβ↑MethodsSup.MAE↓Sm↑Em↑Fwβ↑
✓×unstableZoomNet [29]F0.0660.8200.8920.752
×✓0.0810.7680.8520.706CRNet [19]S0.0920.7350.8150.641
✓✓0.0740.7980.8720.727Ours + scribbleS0.0650.8160.8940.761
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.785, + 0.533 + ], + "angle": 0, + "content": "label, Init square area and Prediction map. Because it has more supervision compared to the Init square area and fewer errors (e.g., background noise) compared to prediction map and prediction map obtained by SAM. Interestingly, the init square area performs much better than prediction map P and P obtained by SAM, and we think that the accuracy of supervision in WSCOD is more crucial than the quantity, since init square area has better accuracy and prediction map has high quantity in area generation but contains a lot of background noise. In addition, compared with prediction map, a relatively complete prediction map can be obtained by using hint area generator, as shown in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.787, + 0.612 + ], + "angle": 0, + "content": "Effectiveness of Attention Regulator. We propose attention regulator and apply it with UCL. As shown in Tab. 7, the results (Bold) are significantly improved after using attention regulator. In addition, to further demonstrate the superiority of regulator in weakly-supervised COD, we compare it with the previous mask approaches HaS [21] and Cutout [5], results are shown in Tab. 8." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Effectiveness of Representation Optimizer. We conduct ablation experiments for representation optimizer. Firstly, we test different types of contrastive loss and find that L1 loss performs best, with a significant improvement over the baseline, as shown in Tab. 6. In addition, we conduct exhaustive experiments for data augmentation, which is important for UCL in representation optimizer. Specifically, the augmentation is divided into three aspects, color-texture related (ColorJitter, GuassBlur), position related (Flip, Translate), and size related (Crop, Scale). It shows that color-texture and size augmentation can improve performance, but position augmentation yields unsatisfactory results, as shown in Tab. 7. We employ UCL exclusively with positive samples and test the key components in ablation study, as shown in Tab. 9. Fig. 6 intuitively illustrates that the representation optimizer enhances the precision of prediction maps. Fig. 7 and 8 show that our model is able to continuously optimize representation through representation optimizer and separate entangled objects from background, making the model eventually learn robust representations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Just a Hint: Point-Supervised Camouflaged Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.275, + 0.145, + 0.726, + 0.16 + ], + "angle": 0, + "content": "Table 11: The impact of \\(\\tau\\), \\(\\alpha\\), \\(w\\), and \\(d\\) in the hint area generator." + }, + { + "type": "table", + "bbox": [ + 0.25, + 0.172, + 0.754, + 0.226 + ], + "angle": 0, + "content": "
dMAE↓Sm↑τMAE↓Sm↑αMAE↓Sm↑wMAE↓Sm↑
50.0850.7581000.0860.75730.0910.73950.0890.755
100.0850.7591500.0850.75940.0850.759100.0870.759
150.0860.7602000.0850.75850.0920.736150.0850.759
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.254, + 0.785, + 0.296 + ], + "angle": 0, + "content": "Table 12: Quantitative comparison with state-of-the-arts on four popular SOD benchmarks. \"F\", \"S\", \"P\" denote fully-, scribble-, and point-supervised methods, respectively. The best are in bold." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.319, + 0.803, + 0.415 + ], + "angle": 0, + "content": "
MethodsSup.ECSSDDUT-OHKU-ISDUTS-TE
MAE↓Sm↑FβmaxMAE↓Sm↑FβmaxMAE↓Sm↑FβmaxMAE↓Sm↑Fβmax
AFNet [13]F0.0420.9130.9350.0570.8260.7970.0360.9050.9230.0460.8670.863
BASNet [33]F0.0370.9160.9430.0570.8360.8050.0320.9090.9280.0480.8660.859
MFNet [31]S0.0840.8340.8790.0870.7410.7060.0590.8460.8760.7650.7740.770
SCSOD [40]S0.0490.8810.9140.0600.8110.7820.0380.8820.9080.0490.8530.858
PSOD [16]P0.0360.9130.9350.0640.8240.8080.0330.9010.9230.0450.8530.858
OursP0.0450.8950.9150.0590.8260.8240.0310.9060.9260.0420.8640.861
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.626 + ], + "angle": 0, + "content": "Impact of the Hyperparameter \\(\\tau, \\alpha, w, d\\). Four hyperparameters exist in hint area generator: \\(\\tau, \\alpha, w, d\\). We test diverse values for these parameters, as shown in Tab. 11. So we set \\(\\tau = 150\\), \\(\\alpha = 4\\), \\(w = 15\\) and \\(d = 10\\). It is worth noting that the values of \\(\\tau, w\\), and \\(d\\) do not have much impact on the results with different settings. In addition, the \\(\\tau\\) can be replaced with unsupervised k-means clustering method (in S.M.). Our model is robust to parameters. It is because the hint area generator is designed to approximate the size of the camouflaged object through the prediction map. Whether the contour is fine and details are perfect, in the prediction map, has little influence on the estimation of the object size. The accurate estimation and localization of the camouflaged object are explored by the attention regulator and the representation optimizer modules." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.648, + 0.451, + 0.664 + ], + "angle": 0, + "content": "4.4 Transferability Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Extension to Scribble. We apply our method to the scribble dataset, and the results are shown in Tab. 10 (More results in supplementary materials) reveal that our model, utilizing scribble labeling, achieves competitive results compared to the state-of-the-art fully supervised method ZoomNet [29]. It showcases the generalization ability of our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Extension to SOD task. Our method excels not only in COD but also demonstrates competitive performance in SOD, as shown in Tab. 12. Considering the difference between these two tasks, we may wonder why our method consistently performs well on such two seemingly different tasks. We attribute this to the generality and rationality of the designed structure. In fact, point-supervised SOD and COD have a clear commonality, i.e., the demands of sufficient and accurate" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.127 + ], + "angle": 0, + "content": "H. Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.145, + 0.734, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.248, + 0.785, + 0.276 + ], + "angle": 0, + "content": "Fig. 7: Visualization of the feature space. Entangled features from foregrounds and backgrounds could be well separated by representation optimizer (visualized by t-SNE)." + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.29, + 0.737, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.415, + 0.785, + 0.457 + ], + "angle": 0, + "content": "Fig. 8: Visualization of results in different iterations. As training proceeds, the entangled features from the foreground and background can be progressively separated by the representation optimizer (visualization by t-SNE)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.607 + ], + "angle": 0, + "content": "supervision information. Through an effective point label expansion strategy, our model can provide more supervision information than a single point and more accurate supervision information than pseudo-labels, aiding in the precise segmentation of objects. Additionally, the proposed UCL optimizes the representation of weakly-supervised models, reducing segmentation instability. Despite the saliency of objects in WSSOD, it can also benefit from UCL. All these components are built on the common demands of the two tasks, providing a solid foundation for performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.629, + 0.36, + 0.645 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.795 + ], + "angle": 0, + "content": "In this paper, we proposed an effective solution for COD by utilizing point-based supervision. The introduction of the P-COD dataset, based on human cognitive processes, streamlines the annotation process. To overcome the challenge of single-point supervision, we designed a hint area generator bolstering supervision. Representation optimizer optimized model representation by UCL. The attention regulator promoted a holistic object perception, ensuring a comprehensive understanding of camouflaged objects without fixating on distinct parts. Experimental results show that our method outperforms existing weakly-supervised methods largely and even surpasses many fully-supervised ones." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Acknowledgement. This work is supported by the National Natural Science Foundation of China(NSFC) under Grant 62306239 and the Young Talent Fund of Association for Science and Technology in Shaanxi, China." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Just a Hint: Point-Supervised Camouflaged Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Bearman, A., Russakovsky, O., Ferrari, V., Fei-Fei, L.: What's the point: Semantic segmentation with point supervision. In: European conference on computer vision. pp. 549-565. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.26 + ], + "angle": 0, + "content": "2. Benenson, R., Popov, S., Ferrari, V.: Large-scale interactive object segmentation with human annotators. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11700-11709 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.3 + ], + "angle": 0, + "content": "3. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15750-15758 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.301, + 0.785, + 0.342 + ], + "angle": 0, + "content": "4. Cheng, B., Parkhi, O., Kirillov, A.: Pointly-supervised instance segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2617-2626 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "5. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.412 + ], + "angle": 0, + "content": "6. Fan, D.P., Cheng, M.M., Liu, Y., Li, T., Borji, A.: Structure-measure: A new way to evaluate foreground maps. In: Proceedings of the IEEE international conference on computer vision. pp. 4548-4557 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.453 + ], + "angle": 0, + "content": "7. Fan, D.P., Gong, C., Cao, Y., Ren, B., Cheng, M.M., Borji, A.: Enhanced-alignment measure for binary foreground map evaluation. arXiv preprint arXiv:1805.10421 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.453, + 0.785, + 0.481 + ], + "angle": 0, + "content": "8. Fan, D.P., Ji, G.P., Cheng, M.M., Shao, L.: Concealed object detection. IEEE transactions on pattern analysis and machine intelligence 44(10), 6024-6042 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "9. Fan, D.P., Ji, G.P., Sun, G., Cheng, M.M., Shen, J., Shao, L.: Camouflaged object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2777-2787 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.55 + ], + "angle": 0, + "content": "10. Fan, D.P., Ji, G.P., Xu, P., Cheng, M.M., Sakaridis, C., Van Gool, L.: Advances in deep concealed scene understanding. Visual Intelligence 1(1), 16 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.55, + 0.785, + 0.605 + ], + "angle": 0, + "content": "1. Fan, D.P., Ji, G.P., Zhou, T., Chen, G., Fu, H., Shen, J., Shao, L.: Pranet: Parallel reverse attention network for polyp segmentation. In: International conference on medical image computing and computer-assisted intervention. pp. 263-273. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.647 + ], + "angle": 0, + "content": "2. Fan, D.P., Zhou, T., Ji, G.P., Zhou, Y., Chen, G., Fu, H., Shen, J., Shao, L.: Inf-net: Automatic Covid-19 lung infection segmentation from ct images. IEEE transactions on medical imaging 39(8), 2626-2637 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "3. Feng, M., Lu, H., Ding, E.: Attentive feedback network for boundary-aware salient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1623-1632 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.73 + ], + "angle": 0, + "content": "4. Pérez-de la Fuente, R., Delclos, X., Penalver, E., Speranza, M., Wierzchos, J., Ascaso, C., Engel, M.S.: Early evolution and ecology of camouflage in insects. Proceedings of the National Academy of Sciences 109(52), 21414-21419 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.785, + 0.771 + ], + "angle": 0, + "content": "5. Gao, S.H., Tan, Y.Q., Cheng, M.M., Lu, C., Chen, Y., Yan, S.: Highly efficient salient object detection with 100k parameters. In: European Conference on Computer Vision. pp. 702-721. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "6. Gao, S., Zhang, W., Wang, Y., Guo, Q., Zhang, C., He, Y., Zhang, W.: Weakly-supervised salient object detection using point supervision. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 670-678 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.812, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.127 + ], + "angle": 0, + "content": "H. Chen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.786, + 0.216 + ], + "angle": 0, + "content": "18. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.216, + 0.786, + 0.257 + ], + "angle": 0, + "content": "19. He, R., Dong, Q., Lin, J., Lau, R.W.: Weakly-supervised camouflaged object detection with scribble annotations. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 781-789 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.786, + 0.297 + ], + "angle": 0, + "content": "20. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.297, + 0.786, + 0.338 + ], + "angle": 0, + "content": "21. Kumar Singh, K., Jae Lee, Y.: Hide-and-seek: Forcing a network to be meticulous for weakly-supervised object and action localization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3524-3533 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.786, + 0.379 + ], + "angle": 0, + "content": "22. Le, T.N., Nguyen, T.V., Nie, Z., Tran, M.T., Sugimoto, A.: Anabranch network for camouflaged object segmentation. Computer vision and image understanding 184, 45-56 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.786, + 0.42 + ], + "angle": 0, + "content": "23. Li, A., Zhang, J., Lv, Y., Liu, B., Zhang, T., Dai, Y.: Uncertainty-aware joint salient object and camouflaged object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10071-10081 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.786, + 0.461 + ], + "angle": 0, + "content": "24. Liew, J., Wei, Y., Xiong, W., Ong, S.H., Feng, J.: Regional interactive image segmentation networks. In: 2017 IEEE international conference on computer vision (ICCV). pp. 2746-2754. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.786, + 0.514 + ], + "angle": 0, + "content": "25. Lv, Y., Zhang, J., Dai, Y., Li, A., Liu, B., Barnes, N., Fan, D.P.: Simultaneously localize, segment and rank the camouflaged objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11591-11601 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.786, + 0.555 + ], + "angle": 0, + "content": "26. Margolin, R., Zelnik-Manor, L., Tal, A.: How to evaluate foreground maps? In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 248-255 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.786, + 0.597 + ], + "angle": 0, + "content": "27. Mei, H., Ji, G.P., Wei, Z., Yang, X., Wei, X., Fan, D.P.: Camouflaged object segmentation with distraction mining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8772-8781 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.786, + 0.637 + ], + "angle": 0, + "content": "28. Nguyen, T., Dax, M., Mummadi, C.K., Ngo, N., Nguyen, T.H.P., Lou, Z., Brox, T.: Deepusps: Deep robust unsupervised saliency prediction via self-supervision. Advances in Neural Information Processing Systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.786, + 0.678 + ], + "angle": 0, + "content": "29. Pang, Y., Zhao, X., Xiang, T.Z., Zhang, L., Lu, H.: Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 2160-2170 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.786, + 0.718 + ], + "angle": 0, + "content": "30. Pang, Y., Zhao, X., Zhang, L., Lu, H.: Multi-scale interactive network for salient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9413-9422 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.786, + 0.759 + ], + "angle": 0, + "content": "31. Piao, Y., Wang, J., Zhang, M., Lu, H.: Mfnet: Multi-filter directive network for weakly supervised salient object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4136-4145 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.786, + 0.8 + ], + "angle": 0, + "content": "32. Qian, R., Wei, Y., Shi, H., Li, J., Liu, J., Huang, T.: Weakly supervised scene parsing with point-based distance metric learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 8843-8850 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.786, + 0.841 + ], + "angle": 0, + "content": "33. Qin, X., Zhang, Z., Huang, C., Gao, C., Dehghan, M., Jagersand, M.: Basnet: Boundary-aware salient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7479-7489 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.786, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.324, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Just a Hint: Point-Supervised Camouflaged Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "34. Shin, G., Xie, W., Albanie, S.: All you need are a few pixels: semantic segmentation with pixelpick. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1687-1697 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "35. Wang, B., Zhao, Y., Yang, L., Long, T., Li, X.: Temporal action localization in the deep learning era: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.286 + ], + "angle": 0, + "content": "36. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 568-578 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.328 + ], + "angle": 0, + "content": "37. Wei, J., Wang, Q., Li, Z., Wang, S., Zhou, S.K., Cui, S.: Shallow feature matters for weakly supervised object localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5993-6001 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "38. Wei, J., Wang, S., Huang, Q.: \\(\\mathbf{F}^3\\) net: fusion, feedback and focus for salient object detection. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 12321-12328 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.424 + ], + "angle": 0, + "content": "39. Yang, F., Zhai, Q., Li, X., Huang, R., Luo, A., Cheng, H., Fan, D.P.: Uncertainty-guided transformer reasoning for camouflaged object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4146-4155 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.467 + ], + "angle": 0, + "content": "40. Yu, S., Zhang, B., Xiao, J., Lim, E.G.: Structure-consistent weakly supervised salient object detection with local saliency coherence. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 3234-3242 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.508 + ], + "angle": 0, + "content": "41. Zhai, Q., Li, X., Yang, F., Chen, C., Cheng, H., Fan, D.P.: Mutual graph learning for camouflaged object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12997-13007 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.564 + ], + "angle": 0, + "content": "42. Zhang, J., Fan, D.P., Dai, Y., Anwar, S., Saleh, F.S., Zhang, T., Barnes, N.: Ucnet: Uncertainty inspired rgb-d saliency detection via conditional variational autoencoders. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8582-8591 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.605 + ], + "angle": 0, + "content": "43. Zhang, J., Yu, X., Li, A., Song, P., Liu, B., Dai, Y.: Weakly-supervised salient object detection via scribble annotations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12546-12555 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.647 + ], + "angle": 0, + "content": "44. Zhang, J., Zhang, T., Dai, Y., Harandi, M., Hartley, R.: Deep unsupervised saliency detection: A multiple noisy labeling perspective. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9029-9038 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.647, + 0.788, + 0.689 + ], + "angle": 0, + "content": "45. Zhou, H., Xie, X., Lai, J.H., Chen, Z., Yang, L.: Interactive two-stream decoder for accurate and fast saliency detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9141-9150 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.689 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Just a Hint_ Point-Supervised Camouflaged Object Detection/6f40032b-910c-409c-99b0-064b71681182_origin.pdf b/2024/Just a Hint_ Point-Supervised Camouflaged Object Detection/6f40032b-910c-409c-99b0-064b71681182_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..81afe59dd63adde57a60e3d63a7c5e9e85d2cd4e --- /dev/null +++ b/2024/Just a Hint_ Point-Supervised Camouflaged Object Detection/6f40032b-910c-409c-99b0-064b71681182_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b4b7f6bec373941c1a5a6fc7635d70e5b4eced729264b9c7db58771902f36857 +size 7437237 diff --git a/2024/Just a Hint_ Point-Supervised Camouflaged Object Detection/full.md b/2024/Just a Hint_ Point-Supervised Camouflaged Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8a01315345c1f3e9b22a37bc48a1f40ad5c0058e --- /dev/null +++ b/2024/Just a Hint_ Point-Supervised Camouflaged Object Detection/full.md @@ -0,0 +1,299 @@ +# Just a Hint: Point-Supervised Camouflaged Object Detection + +Huafeng Chen $^{1,2}$ , Dian Shao $^{1,2}$ , Guangqian Guo $^{1,2}$ , and Shan Gao $^{1,2}$ + +1 Unmanned System Research Institute, Northwestern Polytechnical University + +$^{2}$ National Key Laboratory of Unmanned Aerial Vehicle Technology {chf,guogq21}@mail.nwpu.edu.cn, {shaodian,gaoshan}@nwpu.edu.cn + +Abstract. Camouflaged Object Detection (COD) demands models to expeditiously and accurately distinguish objects which conceal themselves seamlessly in the environment. Owing to the subtle differences and ambiguous boundaries, COD is not only a remarkably challenging task for models but also for human annotators, requiring huge efforts to provide pixel-wise annotations. To alleviate the heavy annotation burden, we propose to fulfill this task with the help of only one point supervision. Specifically, by swiftly clicking on each object, we first adaptively expand the original point-based annotation to a reasonable hint area. Then, to avoid partial localization around discriminative parts, we propose an attention regulator to scatter model attention to the whole object through partially masking labeled regions. Moreover, to solve the unstable feature representation of camouflaged objects under only point-based annotation, we perform unsupervised contrastive learning based on differently augmented image pairs (e.g. changing color or doing translation). On three mainstream COD benchmarks, experimental results show that our model outperforms several weakly-supervised methods by a large margin across various metrics. + +Keywords: Camouflaged Object Detection $\cdot$ Point-supervised $\cdot$ Unsupervised Contrastive Learning + +# 1 Introduction + +Camouflaged Object Detection (COD) aims to distinguish the disguised objects hiding carefully in the environment [8-10, 29]. Recently, it has attracted ever-growing research interest from the computer vision community and facilitates valuable real-life applications, such as search and rescue [8], species discovery [14], and medical image analysis [11, 12]. Specifically, we tentatively figure out two key-points for this task: 1) Quickly probing the existence of camouflaged objects, and 2) Accurately detecting the corresponding locations pixelwise within the given image. Towards this end, accurate pixel-level annotations are required, based on which several fully-supervised approaches achieve good performance [9, 25]. However, since spotting camouflaged objects is also difficult + +![](images/57d61a1a1001791ba97ba3d5b3b87e3d92074ef775c0ab74fc32226a2d1df18b.jpg) +Fig. 1: Different types of annotation in camouflaged object detection task. The mask annotation takes about 60 minutes for each image. The scribbles take about 10 seconds but have diversity and boundary issues. The point takes just 2 seconds and only needs to point the most discriminative part of the camouflaged object. + +for humans, such an annotation process is not only laborious but also time-consuming, e.g., 60 minutes for each image [9]. + +To address this issue, weakly-supervised COD has been explored recently. In [19], the authors try to use scribbles to replace pixel-wise annotations and propose a novel CRNet to solve this problem. But such a setting falls into a dilemma: 1) efficiency, the quality of scribble could not be ensured, arbitrary patterns could harm the learning process, 2) effectiveness, scribbles are expected to indicate the outline of boundaries, which also increases the annotation difficulty, as shown in Fig. 1. + +In this work, we propose to fulfill the COD task with only point-based supervision. Such an idea is motivated by the cognitive process of the human vision system: we tend to recognize the most "discriminative part" of one camouflaged object [25], and then scatter our attention to figure out the whole object mask. Even sometimes the camouflaged object is not perceived at first glance, we could quickly notice the target with just a little hint when a friend points to the image and says "carefully scrutinize there!". Importantly, the whole point-supervised annotation process only takes within 2 seconds. + +However, supervision provided by a single point is rather limited, making the model easily collapse during training. In a similar but different task, Salient Object Detection (SOD), widely adopted solutions include detecting the edge map [16] first, or generating the prediction map [37] as the proposal region at the first stage. Unfortunately, all the previously mentioned solutions for generating proposal regions in SOD are not applicable in point-supervised COD, owing to the low contrast and ambiguous edges between the camouflaged objects and backgrounds (shown in S.M.). To address it, we explore a new point-to-region supervision generation strategy, hint area generator. Specifically, starting from the single point annotation, we adaptively expand it to a certain region, which obviously enriches the limited supervision, as shown in Fig. 2. + +Moreover, weakly-supervised COD methods suffer from partial detection, i.e., only discriminative parts of the object could gain adequate attention, while the whole shape is not accurately perceived, especially with the supervision of a single point. To make the model not only focus on local distinct ones, but nondiscriminatory parts, we propose an attention regulator module. By masking + +![](images/b396ff78523e13ee16cb6724538428cfff2ecf485919f1cd1f5167a4bce2ea29.jpg) +Fig. 2: Overview of our method. The first row shows the process of hint area generation from just a point to a hypothetical area. The second row shows the training process of attention regulator and presentation optimizer from the hint area to an accurate mask. + +labeled region randomly, the model could scatter attention to the whole object, as shown in Fig. 2. + +Furthermore, the learned feature representations with only point-based annotation tend to be unstable. As shown in Fig. 3, although the two images appear to be similar, the prediction accuracy largely varies. This is largely due to that the learned features are not robust enough. To obtain invariant patterns and stable representations during training, we apply Unsupervised Contrastive Learning (UCL) strategy on differently augmented data pairs. Such UCL process facilitates the model to learn more reliable feature representations under point supervision. + +To summarize, our contributions are as follows: + +- We propose a novel point-based learning paradigm for the challenging COD task, and construct the first weakly-supervised COD dataset with point annotation, P-COD, which contains 3040 images from the training set of COD10K and 1000 images from the training set of CAMO. Annotators are only required to swiftly click a point within the camouflaged object according to their first impression without localizing the whole object. +- We develop a point-supervised COD method by imitating the cognitive process of the human vision system, which contains several novel designs, 1) hint area generator, 2) attention regulator, and 3) representation optimizer. +- Experimental results show that our method outperforms existing weakly-supervised COD methods largely and even surpasses many fully-supervised methods. Moreover, when migrated to scribble-supervised COD and SOD tasks, our method also achieves competitive results. + +# 2 Related Work + +Weakly Supervised Camouflaged Detection. Pixel-level annotation in COD is very time-consuming, in order to reduce the cost of labeling, recent work has + +![](images/711ed9e88dd5c17064465aa157de7e0cb2dfa1181a112b05975902cd827af9e5.jpg) +Fig. 3: Comparison of prediction accuracy for similar images during training. Although two images $I_{1}$ and $I_{2}$ are very similar, their prediction accuracy trends are opposite. It is largely due to that the learned features are not robust enough in weak supervision. + +begun to explore weakly supervised camouflaged object detection. CRNet [19] firstly tries to use scribble labels to train the model. However, scribble labels have the problems of label diversity and poor control of label quality. We try to use simpler and more natural point labels. + +Point Annotation. There have been several works on point annotations in weakly-supervised segmentation [1,4,32,34] and instance segmentation [2,24]. Semantic segmentation focuses on class-specific categories, which is different from COD task. PSOD [16] proposes a novel weakly-supervised Salient Object Detection (SOD) method using point supervision. But the drawbacks of it are also obvious: 1) extra edge detectors and edge information 2) flood fill algorithm introducing additional computation, and not appropriate for COD since camouflaged objects are not salient. + +Contrastive Learning. The core of contrastive learning is to attract the positive sample pairs and repulse the negative sample pairs. Moco [18] maintained a queue of negative samples and turned one branch of siamese network into a momentum encoder to improve the consistency of the queue. BYOL [17] proposed a self-supervised image representation learning approach, achieving state-of-the-art results without negative samples. The mainstream paradigm of contrastive learning later shifted to only using positive samples. SimSiam [3] built an efficient and simple contrastive learning framework only using positive samples, which is free from the limitations of large batch size, momentum encoder, and negative samples. The above work [3,17,18] has proved that UCL can completely approach or even exceed the effect of supervised manner, and learn representations with strong generalization. Therefore, we try to introduce unsupervised contrastive learning to point-supervised COD to strengthen the representation of the model. + +# 3 Methodology + +The overall architecture of the proposed framework is shown in Fig. 4, which includes a) a hint area generator to extend supervision from a single point to the small hint region, b) an attention regulator to guide the model to focus + +![](images/910730c2889d5f0baa168e045b9e7a6ce2ad29489ccc0f4c3d6791fe84d07ce4.jpg) +Fig. 4: Details of our method. The main components consist of three parts: (a) hint area generator, (b) attention regulator, and (c) representation optimizer. Hint area generator extends the point label from a single point to the small hint region. Attention regulator enforce the model to focus on the whole object instead of being stuck in most discriminative part. Representation optimizer uses the unsupervised contrastive learning to learn a stable feature representation for disentangling camouflaged objects and background. + +on the whole object instead of being stuck in local discriminative parts, and c) a representation optimizer based on unsupervised contrastive learning for disentangling camouflaged objects and background. + +# 3.1 Hint Area Generator + +Compared with other forms of mask supervision, supervision provided by a point is rather limited, which easily leads to model collapse, as shown in Tab. 5. To explain a little, an annotated point just functions like a "hint", only indicating the existence of the object, with no knowledge of the area, shape, or boundary. To solve this, we design a hint area generator to adaptively expand a single point to a relatively larger region. One of the most crucial things is to control the degree of point expansion, avoiding the resultant hint area being too small (limited effect) or too large (out of boundary). Especially avoiding excessive expansion to introduce noise, place quality over quantity, because incorrect supervision is fatal for WSCOD, as shown in Tab. 5. + +Specifically, as shown in Fig. 4(a), we initialize the desired annotation region with a small square area around the annotated point. It just prevents the model training from collapsing if only using one pixel as a point. The side length of the square area is $d$ . Then, after training the encoder $w$ epochs as a warm- + +start (Short-term training), we hope this encoder could have a certain extent of semantic sense towards the object. Based on this, we estimate the pseudo area size of the camouflaged object, which is calculated by: $P = E_w(I)$ , where $E_w$ denotes the encoder after training $w$ epochs, $I$ represents the input image, and $P$ is the obtained prediction map used to serve as the rough object area hint. + +We choose circular to be the basic area shape, which could be intuitively regarded as "a large point". Then the key is to determine the reasonable radius of the circular area. Based on the object area hint $P$ , the initialized radius $r$ is defined as: + +$$ +r = \sqrt {\frac {S u m (\mathbb {I} (P > \tau))}{N}}, \tag {1} +$$ + +where $\tau$ is a threshold, $\mathbb{I}(\cdot)$ is an indicator function, and $N$ is the number of objects. Note that $N$ could be achieved by counting the annotated point and thus is a deterministic parameter. + +Then, to prevent the calculated circular from being out of the object boundary, we use a hyperparameter, $\alpha$ , to alter the radius $R$ using: $R = \frac{r}{\alpha}$ . Specifically, the annotation for an object $n$ is enlarged from a single point $(x_{n},y_{n})$ , to a small circle region around that point with the estimated radius $R$ , denoted as $C_{(x_n,y_n)}^R$ . It is noteworthy that the background is also annotated by such a small circular region, unfolding around the point $(x_b,y_b)$ , which is randomly picked from the background region. + +As a result, with the help of the hint area generator, the original point supervision for the image $I$ is enriched into several small regions, expanding supervision while minimizing the introduction of incorrect supervision as much as possible, represented as: + +$$ +S (I) = C _ {\left(x _ {b}, y _ {b}\right)} ^ {R} \cup \sum_ {n} C _ {\left(x _ {n}, y _ {n}\right)} ^ {R}, \quad n = 1, \dots , N \tag {2} +$$ + +where $S(I)$ is the obtained supervision, as shown in Fig. 4. + +# 3.2 Attention Regulator + +Weakly-supervised methods often suffer from partial detection problems, owing to the fact that model tends to focus too much on those discriminative regions. Inspired by the HaS method [21], the random mask is able to inhibit the strong response, we further design an attention regulator module for PCOD. Because in camouflaged scenes, people often annotate point labels on the most visually discriminative regions for the camouflaged objects, we leverage this prior to guide mask generation to scatter the attention from the directly supervised discriminative part to the whole object area, as shown in Fig. 4(b). Firstly, based on the given supervision region, $C^R$ , we could obtain a mask $M$ : + +$$ +M = \left\{ \begin{array}{l l} M _ {i} = h (Z), & i \in C ^ {R} \\ M _ {i} = 1, & i \notin C ^ {R} \end{array} \right. \tag {3} +$$ + +where $Z$ is a logical matrix consisting of zeros and ones with the same shape as $I$ , $h(\cdot)$ denotes a shuffle operation, $i$ is the pixel index. + +Then, we apply this mask to the original image $I$ : $I_{*} = I * M$ , where * indicates element-wise multiplication. Through such formulation, the annotated discriminative area will be randomly masked. During the training process, the model needs to explore the recognition of the surrounding foreground areas of the mask to restore recognition of the masked foreground area, helping the model scatter attention to other regions within the foreground objects. + +# 3.3 Representation Optimizer + +In COD, the foreground and the background share high similarity. The subtle difference sometimes comes from the texture abruption or the color inconsistency. Based on such observations, we hope at least from one aspect the discrepancy could be exposed. + +Toward this goal, we propose to use unsupervised contrastive learning strategy to optimize the feature space, aiming to make the foreground object distinguished from the background through the learned invariant patterns and stable representations. + +Specifically, as shown in Fig. 4(c), for image $I$ , we adopt two different visual transformations $T_{1}$ and $T_{2}$ , selecting from colorjitter, Gaussianblur, flip, etc. These visual transformations are able to change the appearance of images. For example, Gaussian blur will smooth the texture, and changing color is relatively harmless to the texture patterns. This process results in two transformed images $I^{1}$ and $I^{2}$ : $I^{1} = T_{1}(I)$ , and $I^{2} = T_{2}(I)$ . + +We encode the transformed images differently into two prediction maps $P^{(1)}$ and $P^{(2)}$ , denoting as $P^1 = g(f(I^1))$ , and $P^2 = f(I^2)$ , where $f$ represents the encoder to be learned, and $g$ is a small network cascaded over $f$ to increase encoding diversity. When a position-related or size-related transformation $T$ (e.g., scale, crop) is applied to the image $I^1$ , this $T$ should be applied to $P^2$ to be aligned with $P^1$ , and vice versa. Our objective is to minimize the distance between these two prediction maps: + +$$ +\min \mathcal {D} (P ^ {1}, P ^ {2}) = \sum_ {i} \left| P _ {i} ^ {1} - P _ {i} ^ {2} \right|, \tag {4} +$$ + +where $i$ is the pixel index. Here we follow the design of UCL, i.e., stopping the gradient (stopgrad) update at one end, so the contrastive loss function is defined as: + +$$ +\mathcal {L} _ {c} = \mathcal {D} \left(P ^ {1}, \operatorname {s t o p g r a d} \left(P ^ {2}\right)\right). \tag {5} +$$ + +The prediction gap using different transformations will be narrowed by minimizing the above loss, making the learned feature representation robust. + +# 3.4 Network + +Encoder. The structure of our feature encoder is simple compared to previous work. Specifically, in order to capture long-distance feature dependency and + +Table 1: Quantitative comparison with state-of-the-arts on four benchmarks. "F", "U", "S", "P" denote fully-supervised, unsupervised, scribble-supervised, and point-supervised methods, respectively. The best results are highlighted in bold. + +
MethodsSup.CAMO (250)COD10K (2026)NC4K (4121)
MAE\( \downarrow S_m \)↑\( E_m \)↑\( F_w^\beta \)↑MAE\( \downarrow S_m \)↑\( E_m \)↑\( F_w^\beta \)↑MAE\( \downarrow S_m \)↑\( E_m \)↑\( F_w^\beta \)↑
F3Net [38]F0.1090.7110.7410.5640.0510.7390.7950.5440.0690.7820.8250.706
CSNet [15]F0.0920.7710.7950.6410.0470.7780.8090.5690.0610.8190.8450.748
ITSD [45]F0.1020.7500.7790.6100.0510.7670.8080.5570.0640.8110.8450.729
MINet [30]F0.0900.7480.7910.6370.0420.770.8320.608----
PraNet [11]F0.0940.7690.8250.6630.0450.7890.8610.629----
UCNet [42]F0.0940.7390.7870.6400.0420.7760.8570.6330.0550.8130.8720.777
SINet [9]F0.0920.7450.8040.6440.0430.7760.8640.6310.0580.8080.8710.723
MGL-R [41]F0.0880.7750.8120.6730.0350.8140.8510.6660.0520.8330.8670.740
PFNet [27]F0.0850.7820.8410.6950.0400.8000.8770.6600.0530.8290.8870.745
UJSC [23]F0.0730.8000.8590.7280.0350.8090.8840.6840.0470.8420.8980.771
UGTR [39]F0.0860.7840.8220.6840.0360.8170.8520.6660.0520.8390.8740.747
ZoomNet [29]F0.0660.8200.8920.7520.0290.8380.9110.7290.0430.8530.8960.784
DUSD [44]U0.1660.5510.5940.3080.1070.5800.6460.276----
USPS [28]U0.2070.5680.6410.3990.1960.5190.5360.265----
SAM [20]U0.1320.6840.6870.6060.0500.7830.7980.7010.0780.7670.7760.696
SS [43]S0.1180.6960.7860.5620.0710.6840.7700.461----
SCSOD [40]S0.1020.7130.7950.6180.0550.7100.8050.546----
CRNet [19]S0.0920.7350.8150.6410.0490.7330.8320.5760.0630.7750.8550.688
OursP0.0740.7980.8720.7270.0420.7840.8590.6500.0510.8220.8890.748
+ +obtain multi-scale information, we use PVT [36] as the backbone, for an input image $I \in \mathbb{R}^{3 \times H \times W}$ , we put it into the backbone to get the output features $F_{eat_i}$ for the $i$ -th. Then, we get the multi-scale features $(F_{eat1}, F_{eat2}, F_{eat3}, F_{eat4})$ with $\left(\frac{1}{4}, \frac{1}{8}, \frac{1}{16}, \frac{1}{32}\right)$ resolution of input image $I$ . We downsize the channel dimension of $F_{eat_i}$ into 64 by using $3 \times 3$ convolutional layers. Next, these feature maps are unified into the same size by an up-sampling operation, and combined through the concatenation. Finally the output map $\hat{S} \in \mathbb{R}^{1 \times W \times H}$ is obtained by the $3 \times 3$ convolution layer, and the detailed network structure is provided in supplementary materials. + +Loss. Compared to other weak-supervised methods [19,35,40,43], we have fewer losses, only two. The final loss $\mathcal{L}$ includes contrastive loss $\mathcal{L}_c$ and partial cross-entropy loss $\mathcal{L}_{pce}$ , the former is defined in Eq. 5, and the latter is defined as: + +$$ +\mathcal {L} _ {p c e} = - \sum_ {i \in \tilde {S}} S _ {i} \log (\hat {S} _ {i}) + (1 - S _ {i}) \log (1 - \hat {S} _ {i}), \tag {6} +$$ + +where $\tilde{S}$ represents the labeled area of $S$ . $S_{i}$ is the true class of pixel $i$ , and $\hat{S}_i$ is the prediction of pixel $i$ . So the final loss $\mathcal{L}$ can be defined as: + +$$ +\mathcal {L} = \mathcal {L} _ {c} + \mathcal {L} _ {p c e}. \tag {7} +$$ + +Even though we only use point label in training, our loss is generalizable to the other weakly-supervised COD models by contrastive learning. + +![](images/f976a0ff99afb29e8cf41d88adcb7c1c1111fcf55234bd329eccfdbd98a42d9b.jpg) +Fig. 5: Visual comparison with representative scribble-supervised and fully-supervised models. + +Table 2: Ablation on point position. + +
TypeMAE↓Sm↑Em↑Fwβ↑
Center0.0790.7890.8580.716
Random0.0760.7910.8690.722
Ours0.0740.7980.8720.727
+ +Table 3: Ablation on point numbers. + +
Number of pointsMAE↓Sm↑Em↑Fwβ↑
One-point0.0760.7910.8690.722
Two-points0.0740.8020.8810.737
Three-points0.0740.8050.8870.741
+ +# 4 Experiments + +# 4.1 Experimental Setup + +Point Dataset. Our experiments are conducted on three COD benchmarks, CAMO [22], COD10K [9], and NC4K [25]. In order to evaluate our point-supervised COD method, we relabel 4040 images (3040 from COD10K, 1000 from CAMO) and propose the Point-supervised Dataset (P-COD) for training and the remaining is for testing. In Fig. 1, we show an example of point annotation and compare it with other annotation methods. For each image, we simulate the hunting process to label only one point for each camouflaged object, without needing to worry about ambiguous boundaries, making the process of the annotation easy and natural. + +Evaluation Metrics. We adopt four evaluation metrics: mean absolute error (MAE), S-measure $(\mathrm{S}_m)$ [6], E-measure $(\mathrm{E}_m)$ [7], weighted F-measure $(\mathrm{F}_{\beta}^{w})$ [26]. Implementation Details. We implement our method with PyTorch and conduct experiments on one GeForce RTX4090 GPU. We use the stochastic gradient descent optimizer with a momentum of 0.9, a weight decay of $5e - 4$ , and triangle learning rate schedule with maximum learning rate $1e - 3$ . The batch size is 8, and the training epoch is 60. It takes around 7 hours to train. During training and inference, input images are resized to $512\times 512$ . + +Table 4: Comparison of parameters and MACs. * is reset in the same input size. + +
MethodsSup.Input sizePara(M)MACs(G)MAE↓Sm↑Em↑Fwβ↑
SINetF352248.9519.420.0920.7450.8040.644
CRNet*S352232.6514.270.0950.7300.8040.634
OursP352225.4410.220.0880.7610.8360.677
+ +Table 5: Effect of hint area generator with different settings. Point, I, P and SAM denote point label, initial square area, prediction map P and P obtained by SAM as proposal area, respectively. + +
SettingsMAE↓Sm↑Em↑Fwβ↑
w/ Pointcollapse
w/ I0.0920.7540.8020.655
w/ P0.1030.7260.8130.627
w/ SAM0.1080.6940.7710.603
w/ Ours0.0850.7590.8420.677
+ +Table 6: Ablation study on contrastive losses. MSE, KL, COS, L1 mean Mean Square Error, Kullback-Leibler divergence, Consine similarity, and L1 loss, respectively. + +
LossMAE↓Sm↑Em↑Fwβ↑
w/o0.0850.7590.8420.677
MSE0.0790.7910.8510.708
KL0.0850.7690.8440.686
COS0.0870.7660.8440.680
L10.0740.7980.8720.727
+ +# 4.2 Compare with State-of-the-arts + +Quantitative Comparison. Being the first point supervised COD method, our proposed approach is primarily benchmarked against scribble-supervised and fully-supervised methods. As demonstrated in Tab. 1, our method achieves substantial improvements, with an average enhancement of $17.6\%$ for MAE, $7.2\%$ for Sm, $4.7\%$ for Em, and $11.7\%$ for $\mathrm{F}_{\beta}^{w}$ compared to the state-of-the-art weakly-supervised COD method CRNet [19]. This highlights our capability to achieve better performance with fewer annotations. Our approach even outperforms fully supervised methods in several metrics across multiple datasets. Specifically, it excels against the majority of fully-supervised methods on CAMO and outpaces nearly half of them on other datasets. + +Qualitative Evaluation. Our method produces prediction maps characterized by clearer and more complete object regions, along with sharper contours, significantly outperforming state-of-the-art weakly supervised COD method CRNet [19], as shown in Fig. 5. Our method performs well in various challenging scenarios, including scenarios with tiny objects (row 4), high intrinsic similarities (row 3, 5), indefinable boundaries (row 1), and complex backgrounds (row 2, 6). + +Parameter Complexity. With lower parameter complexity and minimal computational cost, our model (25.44M parameters and 10.22G MACs) consistently outperforms the fully supervised SINet (48.95M parameters and 19.42G MACs) by $4.3\%$ for MAE, $2.1\%$ for Sm, and $4.0\%$ for Em on the challenging CAMO dataset. It also outperforms the scribble supervised CRNet (32.65M parameters and 14.27G MACs) by $7.4\%$ for MAE, $4.2\%$ for Sm, $4.0\%$ for Em, and $6.8\%$ + +![](images/706e18944d507ffe6f22dcf181193e39008fc227bda5db63aa63e3d5679427e0.jpg) +Fig. 6: Visualization of the component ablation study. I, II, III respectively represent hint area generator, representation optimizer, and attention regulator. + +Table 7: The ablation study for different augmentations in contrastive learning. "S", "C", "F", "T", "R", "G", "J" are Scale, Crop, Flip, Translate, Attention Regulator, Guassblur, Colorjitter. + +
SCFTRGJMAE↓Sm↑Em↑Fwβ↑
0.0850.7590.8420.677
✓0.0820.7790.8530.701
✓✓0.0800.7820.8600.713
✓✓✓0.0800.7710.8450.707
✓✓✓✓0.0770.7810.8530.711
✓✓✓✓✓0.0760.7900.8660.724
✓✓✓✓✓✓0.0740.7940.8670.729
✓✓✓✓✓✓✓0.0740.7980.8720.727
+ +for $\mathrm{F}_{\beta}^{w}$ , as shown in Tab. 4. These results underscore the superior lightweight performance of our model. + +# 4.3 Ablation Study + +As CAMO dataset is the most challenging one, indicated by the lowest scores in Tab. 1, all following ablation experiments are performed on it. + +Ablation Study on Point Supervision. To reduce annotation costs, we intuitively annotated a point for each camouflaged object by humans. We conduct experiments on point positions, as shown in Tab. 2. For Random, we randomly choose a point from real labels, and repeat three times, taking the average of the results. which shows our method far exceeds the center annotation and random annotation under conditions of low annotation cost. We attribute it to the model's utilization of the priors implicit in point annotations to simulate human cognitive processes. In addition, we conduct experiments for point numbers (Random), as shown in Tab. 3. One point yields satisfactory performance. Increasing the points shows diminishing returns. + +Effectiveness of Hint Area Generator. As shown in Tab. 5, our proposed hint area generator obtains a significant improvement over the direct use of point + +Table 8: Effect of different masking setting in Attention regulator. + +
Basic settingTypeMAE↓Sm↑Em↑FWβ↑
w/ aug (S, C, T, F)HaS [21]0.0780.7850.8640.711
w/ aug (S, C, T, F)Cutout [5]0.0760.7890.8620.721
w/ aug (S, C, T, F)Ours0.0760.7900.8660.724
+ +Table 9: Effect of key components in optimizer. +Table 10: Transferability studies on scribble datasets. + +
stop-grad predictorMAE↓Sm↑Em↑Fwβ↑MethodsSup.MAE↓Sm↑Em↑Fwβ↑
✓×unstableZoomNet [29]F0.0660.8200.8920.752
×✓0.0810.7680.8520.706CRNet [19]S0.0920.7350.8150.641
✓✓0.0740.7980.8720.727Ours + scribbleS0.0650.8160.8940.761
+ +label, Init square area and Prediction map. Because it has more supervision compared to the Init square area and fewer errors (e.g., background noise) compared to prediction map and prediction map obtained by SAM. Interestingly, the init square area performs much better than prediction map P and P obtained by SAM, and we think that the accuracy of supervision in WSCOD is more crucial than the quantity, since init square area has better accuracy and prediction map has high quantity in area generation but contains a lot of background noise. In addition, compared with prediction map, a relatively complete prediction map can be obtained by using hint area generator, as shown in Fig. 6. + +Effectiveness of Attention Regulator. We propose attention regulator and apply it with UCL. As shown in Tab. 7, the results (Bold) are significantly improved after using attention regulator. In addition, to further demonstrate the superiority of regulator in weakly-supervised COD, we compare it with the previous mask approaches HaS [21] and Cutout [5], results are shown in Tab. 8. + +Effectiveness of Representation Optimizer. We conduct ablation experiments for representation optimizer. Firstly, we test different types of contrastive loss and find that L1 loss performs best, with a significant improvement over the baseline, as shown in Tab. 6. In addition, we conduct exhaustive experiments for data augmentation, which is important for UCL in representation optimizer. Specifically, the augmentation is divided into three aspects, color-texture related (ColorJitter, GuassBlur), position related (Flip, Translate), and size related (Crop, Scale). It shows that color-texture and size augmentation can improve performance, but position augmentation yields unsatisfactory results, as shown in Tab. 7. We employ UCL exclusively with positive samples and test the key components in ablation study, as shown in Tab. 9. Fig. 6 intuitively illustrates that the representation optimizer enhances the precision of prediction maps. Fig. 7 and 8 show that our model is able to continuously optimize representation through representation optimizer and separate entangled objects from background, making the model eventually learn robust representations. + +Table 11: The impact of $\tau$ , $\alpha$ , $w$ , and $d$ in the hint area generator. + +
dMAE↓Sm↑τMAE↓Sm↑αMAE↓Sm↑wMAE↓Sm↑
50.0850.7581000.0860.75730.0910.73950.0890.755
100.0850.7591500.0850.75940.0850.759100.0870.759
150.0860.7602000.0850.75850.0920.736150.0850.759
+ +Table 12: Quantitative comparison with state-of-the-arts on four popular SOD benchmarks. "F", "S", "P" denote fully-, scribble-, and point-supervised methods, respectively. The best are in bold. + +
MethodsSup.ECSSDDUT-OHKU-ISDUTS-TE
MAE↓Sm↑FβmaxMAE↓Sm↑FβmaxMAE↓Sm↑FβmaxMAE↓Sm↑Fβmax
AFNet [13]F0.0420.9130.9350.0570.8260.7970.0360.9050.9230.0460.8670.863
BASNet [33]F0.0370.9160.9430.0570.8360.8050.0320.9090.9280.0480.8660.859
MFNet [31]S0.0840.8340.8790.0870.7410.7060.0590.8460.8760.7650.7740.770
SCSOD [40]S0.0490.8810.9140.0600.8110.7820.0380.8820.9080.0490.8530.858
PSOD [16]P0.0360.9130.9350.0640.8240.8080.0330.9010.9230.0450.8530.858
OursP0.0450.8950.9150.0590.8260.8240.0310.9060.9260.0420.8640.861
+ +Impact of the Hyperparameter $\tau, \alpha, w, d$ . Four hyperparameters exist in hint area generator: $\tau, \alpha, w, d$ . We test diverse values for these parameters, as shown in Tab. 11. So we set $\tau = 150$ , $\alpha = 4$ , $w = 15$ and $d = 10$ . It is worth noting that the values of $\tau, w$ , and $d$ do not have much impact on the results with different settings. In addition, the $\tau$ can be replaced with unsupervised k-means clustering method (in S.M.). Our model is robust to parameters. It is because the hint area generator is designed to approximate the size of the camouflaged object through the prediction map. Whether the contour is fine and details are perfect, in the prediction map, has little influence on the estimation of the object size. The accurate estimation and localization of the camouflaged object are explored by the attention regulator and the representation optimizer modules. + +# 4.4 Transferability Studies + +Extension to Scribble. We apply our method to the scribble dataset, and the results are shown in Tab. 10 (More results in supplementary materials) reveal that our model, utilizing scribble labeling, achieves competitive results compared to the state-of-the-art fully supervised method ZoomNet [29]. It showcases the generalization ability of our method. + +Extension to SOD task. Our method excels not only in COD but also demonstrates competitive performance in SOD, as shown in Tab. 12. Considering the difference between these two tasks, we may wonder why our method consistently performs well on such two seemingly different tasks. We attribute this to the generality and rationality of the designed structure. In fact, point-supervised SOD and COD have a clear commonality, i.e., the demands of sufficient and accurate + +![](images/d350be7a340903c426077ae45634faaf88c5e47c202d044adc3c808396ac563f.jpg) +Fig. 7: Visualization of the feature space. Entangled features from foregrounds and backgrounds could be well separated by representation optimizer (visualized by t-SNE). + +![](images/6f27c4442cf677f4aec15f1aa56003aeb49177f619a735755b40a3ee9999ff2c.jpg) +Fig. 8: Visualization of results in different iterations. As training proceeds, the entangled features from the foreground and background can be progressively separated by the representation optimizer (visualization by t-SNE). + +supervision information. Through an effective point label expansion strategy, our model can provide more supervision information than a single point and more accurate supervision information than pseudo-labels, aiding in the precise segmentation of objects. Additionally, the proposed UCL optimizes the representation of weakly-supervised models, reducing segmentation instability. Despite the saliency of objects in WSSOD, it can also benefit from UCL. All these components are built on the common demands of the two tasks, providing a solid foundation for performance. + +# 5 Conclusion + +In this paper, we proposed an effective solution for COD by utilizing point-based supervision. The introduction of the P-COD dataset, based on human cognitive processes, streamlines the annotation process. To overcome the challenge of single-point supervision, we designed a hint area generator bolstering supervision. Representation optimizer optimized model representation by UCL. The attention regulator promoted a holistic object perception, ensuring a comprehensive understanding of camouflaged objects without fixating on distinct parts. Experimental results show that our method outperforms existing weakly-supervised methods largely and even surpasses many fully-supervised ones. + +Acknowledgement. This work is supported by the National Natural Science Foundation of China(NSFC) under Grant 62306239 and the Young Talent Fund of Association for Science and Technology in Shaanxi, China. + +# References + +1. Bearman, A., Russakovsky, O., Ferrari, V., Fei-Fei, L.: What's the point: Semantic segmentation with point supervision. In: European conference on computer vision. pp. 549-565. Springer (2016) +2. Benenson, R., Popov, S., Ferrari, V.: Large-scale interactive object segmentation with human annotators. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11700-11709 (2019) +3. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15750-15758 (2021) +4. Cheng, B., Parkhi, O., Kirillov, A.: Pointly-supervised instance segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2617-2626 (2022) +5. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017) +6. Fan, D.P., Cheng, M.M., Liu, Y., Li, T., Borji, A.: Structure-measure: A new way to evaluate foreground maps. In: Proceedings of the IEEE international conference on computer vision. pp. 4548-4557 (2017) +7. Fan, D.P., Gong, C., Cao, Y., Ren, B., Cheng, M.M., Borji, A.: Enhanced-alignment measure for binary foreground map evaluation. arXiv preprint arXiv:1805.10421 (2018) +8. Fan, D.P., Ji, G.P., Cheng, M.M., Shao, L.: Concealed object detection. IEEE transactions on pattern analysis and machine intelligence 44(10), 6024-6042 (2021) +9. Fan, D.P., Ji, G.P., Sun, G., Cheng, M.M., Shen, J., Shao, L.: Camouflaged object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2777-2787 (2020) +10. Fan, D.P., Ji, G.P., Xu, P., Cheng, M.M., Sakaridis, C., Van Gool, L.: Advances in deep concealed scene understanding. Visual Intelligence 1(1), 16 (2023) +1. Fan, D.P., Ji, G.P., Zhou, T., Chen, G., Fu, H., Shen, J., Shao, L.: Pranet: Parallel reverse attention network for polyp segmentation. In: International conference on medical image computing and computer-assisted intervention. pp. 263-273. Springer (2020) +2. Fan, D.P., Zhou, T., Ji, G.P., Zhou, Y., Chen, G., Fu, H., Shen, J., Shao, L.: Inf-net: Automatic Covid-19 lung infection segmentation from ct images. IEEE transactions on medical imaging 39(8), 2626-2637 (2020) +3. Feng, M., Lu, H., Ding, E.: Attentive feedback network for boundary-aware salient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1623-1632 (2019) +4. Pérez-de la Fuente, R., Delclos, X., Penalver, E., Speranza, M., Wierzchos, J., Ascaso, C., Engel, M.S.: Early evolution and ecology of camouflage in insects. Proceedings of the National Academy of Sciences 109(52), 21414-21419 (2012) +5. Gao, S.H., Tan, Y.Q., Cheng, M.M., Lu, C., Chen, Y., Yan, S.: Highly efficient salient object detection with 100k parameters. In: European Conference on Computer Vision. pp. 702-721. Springer (2020) +6. Gao, S., Zhang, W., Wang, Y., Guo, Q., Zhang, C., He, Y., Zhang, W.: Weakly-supervised salient object detection using point supervision. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 670-678 (2022) +7. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own + +latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020) +18. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) +19. He, R., Dong, Q., Lin, J., Lau, R.W.: Weakly-supervised camouflaged object detection with scribble annotations. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 781-789 (2023) +20. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) +21. Kumar Singh, K., Jae Lee, Y.: Hide-and-seek: Forcing a network to be meticulous for weakly-supervised object and action localization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3524-3533 (2017) +22. Le, T.N., Nguyen, T.V., Nie, Z., Tran, M.T., Sugimoto, A.: Anabranch network for camouflaged object segmentation. Computer vision and image understanding 184, 45-56 (2019) +23. Li, A., Zhang, J., Lv, Y., Liu, B., Zhang, T., Dai, Y.: Uncertainty-aware joint salient object and camouflaged object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10071-10081 (2021) +24. Liew, J., Wei, Y., Xiong, W., Ong, S.H., Feng, J.: Regional interactive image segmentation networks. In: 2017 IEEE international conference on computer vision (ICCV). pp. 2746-2754. IEEE (2017) +25. Lv, Y., Zhang, J., Dai, Y., Li, A., Liu, B., Barnes, N., Fan, D.P.: Simultaneously localize, segment and rank the camouflaged objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11591-11601 (2021) +26. Margolin, R., Zelnik-Manor, L., Tal, A.: How to evaluate foreground maps? In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 248-255 (2014) +27. Mei, H., Ji, G.P., Wei, Z., Yang, X., Wei, X., Fan, D.P.: Camouflaged object segmentation with distraction mining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8772-8781 (2021) +28. Nguyen, T., Dax, M., Mummadi, C.K., Ngo, N., Nguyen, T.H.P., Lou, Z., Brox, T.: Deepusps: Deep robust unsupervised saliency prediction via self-supervision. Advances in Neural Information Processing Systems 32 (2019) +29. Pang, Y., Zhao, X., Xiang, T.Z., Zhang, L., Lu, H.: Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 2160-2170 (2022) +30. Pang, Y., Zhao, X., Zhang, L., Lu, H.: Multi-scale interactive network for salient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9413-9422 (2020) +31. Piao, Y., Wang, J., Zhang, M., Lu, H.: Mfnet: Multi-filter directive network for weakly supervised salient object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4136-4145 (2021) +32. Qian, R., Wei, Y., Shi, H., Li, J., Liu, J., Huang, T.: Weakly supervised scene parsing with point-based distance metric learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 8843-8850 (2019) +33. Qin, X., Zhang, Z., Huang, C., Gao, C., Dehghan, M., Jagersand, M.: Basnet: Boundary-aware salient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7479-7489 (2019) + +34. Shin, G., Xie, W., Albanie, S.: All you need are a few pixels: semantic segmentation with pixelpick. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1687-1697 (2021) +35. Wang, B., Zhao, Y., Yang, L., Long, T., Li, X.: Temporal action localization in the deep learning era: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +36. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 568-578 (2021) +37. Wei, J., Wang, Q., Li, Z., Wang, S., Zhou, S.K., Cui, S.: Shallow feature matters for weakly supervised object localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5993-6001 (2021) +38. Wei, J., Wang, S., Huang, Q.: $\mathbf{F}^3$ net: fusion, feedback and focus for salient object detection. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 12321-12328 (2020) +39. Yang, F., Zhai, Q., Li, X., Huang, R., Luo, A., Cheng, H., Fan, D.P.: Uncertainty-guided transformer reasoning for camouflaged object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4146-4155 (2021) +40. Yu, S., Zhang, B., Xiao, J., Lim, E.G.: Structure-consistent weakly supervised salient object detection with local saliency coherence. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 3234-3242 (2021) +41. Zhai, Q., Li, X., Yang, F., Chen, C., Cheng, H., Fan, D.P.: Mutual graph learning for camouflaged object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12997-13007 (2021) +42. Zhang, J., Fan, D.P., Dai, Y., Anwar, S., Saleh, F.S., Zhang, T., Barnes, N.: Ucnet: Uncertainty inspired rgb-d saliency detection via conditional variational autoencoders. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8582-8591 (2020) +43. Zhang, J., Yu, X., Li, A., Song, P., Liu, B., Dai, Y.: Weakly-supervised salient object detection via scribble annotations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12546-12555 (2020) +44. Zhang, J., Zhang, T., Dai, Y., Harandi, M., Hartley, R.: Deep unsupervised saliency detection: A multiple noisy labeling perspective. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9029-9038 (2018) +45. Zhou, H., Xie, X., Lai, J.H., Chen, Z., Yang, L.: Interactive two-stream decoder for accurate and fast saliency detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9141-9150 (2020) \ No newline at end of file diff --git a/2024/Just a Hint_ Point-Supervised Camouflaged Object Detection/images.zip b/2024/Just a Hint_ Point-Supervised Camouflaged Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a0233b8215ef52ab870ca0f33f92a3abe9de4438 --- /dev/null +++ b/2024/Just a Hint_ Point-Supervised Camouflaged Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:390e96cc62be4ca7c60c7d3c086942a2825fd18f91e0d6640b091cd63b8a3522 +size 754305 diff --git a/2024/Just a Hint_ Point-Supervised Camouflaged Object Detection/layout.json b/2024/Just a Hint_ Point-Supervised Camouflaged Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..80fefce96bb6850f196a671b227afe36b9e86eeb --- /dev/null +++ b/2024/Just a Hint_ Point-Supervised Camouflaged Object Detection/layout.json @@ -0,0 +1,9424 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "spans": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "type": "text", + "content": "Just a Hint: Point-Supervised Camouflaged Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 167, + 468, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 167, + 468, + 181 + ], + "spans": [ + { + "bbox": [ + 145, + 167, + 468, + 181 + ], + "type": "text", + "content": "Huafeng Chen" + }, + { + "bbox": [ + 145, + 167, + 468, + 181 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 145, + 167, + 468, + 181 + ], + "type": "text", + "content": ", Dian Shao" + }, + { + "bbox": [ + 145, + 167, + 468, + 181 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 145, + 167, + 468, + 181 + ], + "type": "text", + "content": ", Guangqian Guo" + }, + { + "bbox": [ + 145, + 167, + 468, + 181 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 145, + 167, + 468, + 181 + ], + "type": "text", + "content": ", and Shan Gao" + }, + { + "bbox": [ + 145, + 167, + 468, + 181 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 190, + 468, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 190, + 468, + 201 + ], + "spans": [ + { + "bbox": [ + 145, + 190, + 468, + 201 + ], + "type": "text", + "content": "1 Unmanned System Research Institute, Northwestern Polytechnical University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 201, + 453, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 201, + 453, + 223 + ], + "spans": [ + { + "bbox": [ + 159, + 201, + 453, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 159, + 201, + 453, + 223 + ], + "type": "text", + "content": " National Key Laboratory of Unmanned Aerial Vehicle Technology {chf,guogq21}@mail.nwpu.edu.cn, {shaodian,gaoshan}@nwpu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 251, + 455, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 251, + 455, + 449 + ], + "spans": [ + { + "bbox": [ + 160, + 251, + 455, + 449 + ], + "type": "text", + "content": "Abstract. Camouflaged Object Detection (COD) demands models to expeditiously and accurately distinguish objects which conceal themselves seamlessly in the environment. Owing to the subtle differences and ambiguous boundaries, COD is not only a remarkably challenging task for models but also for human annotators, requiring huge efforts to provide pixel-wise annotations. To alleviate the heavy annotation burden, we propose to fulfill this task with the help of only one point supervision. Specifically, by swiftly clicking on each object, we first adaptively expand the original point-based annotation to a reasonable hint area. Then, to avoid partial localization around discriminative parts, we propose an attention regulator to scatter model attention to the whole object through partially masking labeled regions. Moreover, to solve the unstable feature representation of camouflaged objects under only point-based annotation, we perform unsupervised contrastive learning based on differently augmented image pairs (e.g. changing color or doing translation). On three mainstream COD benchmarks, experimental results show that our model outperforms several weakly-supervised methods by a large margin across various metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 460, + 453, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 460, + 453, + 483 + ], + "spans": [ + { + "bbox": [ + 159, + 460, + 453, + 483 + ], + "type": "text", + "content": "Keywords: Camouflaged Object Detection " + }, + { + "bbox": [ + 159, + 460, + 453, + 483 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 460, + 453, + 483 + ], + "type": "text", + "content": " Point-supervised " + }, + { + "bbox": [ + 159, + 460, + 453, + 483 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 460, + 453, + 483 + ], + "type": "text", + "content": " Unsupervised Contrastive Learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 502, + 230, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 230, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 230, + 514 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 482, + 647 + ], + "type": "text", + "content": "Camouflaged Object Detection (COD) aims to distinguish the disguised objects hiding carefully in the environment [8-10, 29]. Recently, it has attracted ever-growing research interest from the computer vision community and facilitates valuable real-life applications, such as search and rescue [8], species discovery [14], and medical image analysis [11, 12]. Specifically, we tentatively figure out two key-points for this task: 1) Quickly probing the existence of camouflaged objects, and 2) Accurately detecting the corresponding locations pixelwise within the given image. Towards this end, accurate pixel-level annotations are required, based on which several fully-supervised approaches achieve good performance [9, 25]. However, since spotting camouflaged objects is also difficult" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 251, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 251, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 251, + 665 + ], + "type": "text", + "content": "Corresponding Authors" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 479, + 187 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 479, + 187 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 479, + 187 + ], + "type": "image", + "image_path": "57d61a1a1001791ba97ba3d5b3b87e3d92074ef775c0ab74fc32226a2d1df18b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 197, + 482, + 242 + ], + "lines": [ + { + "bbox": [ + 130, + 197, + 482, + 242 + ], + "spans": [ + { + "bbox": [ + 130, + 197, + 482, + 242 + ], + "type": "text", + "content": "Fig. 1: Different types of annotation in camouflaged object detection task. The mask annotation takes about 60 minutes for each image. The scribbles take about 10 seconds but have diversity and boundary issues. The point takes just 2 seconds and only needs to point the most discriminative part of the camouflaged object." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 266, + 479, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 479, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 479, + 291 + ], + "type": "text", + "content": "for humans, such an annotation process is not only laborious but also time-consuming, e.g., 60 minutes for each image [9]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 292, + 480, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 480, + 375 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 480, + 375 + ], + "type": "text", + "content": "To address this issue, weakly-supervised COD has been explored recently. In [19], the authors try to use scribbles to replace pixel-wise annotations and propose a novel CRNet to solve this problem. But such a setting falls into a dilemma: 1) efficiency, the quality of scribble could not be ensured, arbitrary patterns could harm the learning process, 2) effectiveness, scribbles are expected to indicate the outline of boundaries, which also increases the annotation difficulty, as shown in Fig. 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "content": "In this work, we propose to fulfill the COD task with only point-based supervision. Such an idea is motivated by the cognitive process of the human vision system: we tend to recognize the most \"discriminative part\" of one camouflaged object [25], and then scatter our attention to figure out the whole object mask. Even sometimes the camouflaged object is not perceived at first glance, we could quickly notice the target with just a little hint when a friend points to the image and says \"carefully scrutinize there!\". Importantly, the whole point-supervised annotation process only takes within 2 seconds." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 473, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 605 + ], + "type": "text", + "content": "However, supervision provided by a single point is rather limited, making the model easily collapse during training. In a similar but different task, Salient Object Detection (SOD), widely adopted solutions include detecting the edge map [16] first, or generating the prediction map [37] as the proposal region at the first stage. Unfortunately, all the previously mentioned solutions for generating proposal regions in SOD are not applicable in point-supervised COD, owing to the low contrast and ambiguous edges between the camouflaged objects and backgrounds (shown in S.M.). To address it, we explore a new point-to-region supervision generation strategy, hint area generator. Specifically, starting from the single point annotation, we adaptively expand it to a certain region, which obviously enriches the limited supervision, as shown in Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Moreover, weakly-supervised COD methods suffer from partial detection, i.e., only discriminative parts of the object could gain adequate attention, while the whole shape is not accurately perceived, especially with the supervision of a single point. To make the model not only focus on local distinct ones, but nondiscriminatory parts, we propose an attention regulator module. By masking" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 116, + 475, + 230 + ], + "blocks": [ + { + "bbox": [ + 137, + 116, + 475, + 230 + ], + "lines": [ + { + "bbox": [ + 137, + 116, + 475, + 230 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 475, + 230 + ], + "type": "image", + "image_path": "b396ff78523e13ee16cb6724538428cfff2ecf485919f1cd1f5167a4bce2ea29.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 238, + 480, + 271 + ], + "lines": [ + { + "bbox": [ + 131, + 238, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 131, + 238, + 480, + 271 + ], + "type": "text", + "content": "Fig. 2: Overview of our method. The first row shows the process of hint area generation from just a point to a hypothetical area. The second row shows the training process of attention regulator and presentation optimizer from the hint area to an accurate mask." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 295, + 479, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 479, + 319 + ], + "type": "text", + "content": "labeled region randomly, the model could scatter attention to the whole object, as shown in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 480, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 480, + 415 + ], + "type": "text", + "content": "Furthermore, the learned feature representations with only point-based annotation tend to be unstable. As shown in Fig. 3, although the two images appear to be similar, the prediction accuracy largely varies. This is largely due to that the learned features are not robust enough. To obtain invariant patterns and stable representations during training, we apply Unsupervised Contrastive Learning (UCL) strategy on differently augmented data pairs. Such UCL process facilitates the model to learn more reliable feature representations under point supervision." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 415, + 356, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 415, + 356, + 427 + ], + "spans": [ + { + "bbox": [ + 146, + 415, + 356, + 427 + ], + "type": "text", + "content": "To summarize, our contributions are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 438, + 481, + 594 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 138, + 438, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 438, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 438, + 481, + 510 + ], + "type": "text", + "content": "- We propose a novel point-based learning paradigm for the challenging COD task, and construct the first weakly-supervised COD dataset with point annotation, P-COD, which contains 3040 images from the training set of COD10K and 1000 images from the training set of CAMO. Annotators are only required to swiftly click a point within the camouflaged object according to their first impression without localizing the whole object." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 511, + 480, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 511, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 138, + 511, + 480, + 547 + ], + "type": "text", + "content": "- We develop a point-supervised COD method by imitating the cognitive process of the human vision system, which contains several novel designs, 1) hint area generator, 2) attention regulator, and 3) representation optimizer." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 548, + 480, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 548, + 480, + 594 + ], + "spans": [ + { + "bbox": [ + 138, + 548, + 480, + 594 + ], + "type": "text", + "content": "- Experimental results show that our method outperforms existing weakly-supervised COD methods largely and even surpasses many fully-supervised methods. Moreover, when migrated to scribble-supervised COD and SOD tasks, our method also achieves competitive results." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 614, + 237, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 614, + 237, + 627 + ], + "spans": [ + { + "bbox": [ + 132, + 614, + 237, + 627 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Weakly Supervised Camouflaged Detection. Pixel-level annotation in COD is very time-consuming, in order to reduce the cost of labeling, recent work has" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "text", + "content": "Just a Hint: Point-Supervised Camouflaged Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 115, + 476, + 205 + ], + "blocks": [ + { + "bbox": [ + 137, + 115, + 476, + 205 + ], + "lines": [ + { + "bbox": [ + 137, + 115, + 476, + 205 + ], + "spans": [ + { + "bbox": [ + 137, + 115, + 476, + 205 + ], + "type": "image", + "image_path": "711ed9e88dd5c17064465aa157de7e0cb2dfa1181a112b05975902cd827af9e5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 213, + 482, + 248 + ], + "lines": [ + { + "bbox": [ + 131, + 213, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 482, + 248 + ], + "type": "text", + "content": "Fig. 3: Comparison of prediction accuracy for similar images during training. Although two images " + }, + { + "bbox": [ + 131, + 213, + 482, + 248 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 131, + 213, + 482, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 213, + 482, + 248 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 131, + 213, + 482, + 248 + ], + "type": "text", + "content": " are very similar, their prediction accuracy trends are opposite. It is largely due to that the learned features are not robust enough in weak supervision." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 270, + 479, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 479, + 319 + ], + "type": "text", + "content": "begun to explore weakly supervised camouflaged object detection. CRNet [19] firstly tries to use scribble labels to train the model. However, scribble labels have the problems of label diversity and poor control of label quality. We try to use simpler and more natural point labels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 480, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 480, + 415 + ], + "type": "text", + "content": "Point Annotation. There have been several works on point annotations in weakly-supervised segmentation [1,4,32,34] and instance segmentation [2,24]. Semantic segmentation focuses on class-specific categories, which is different from COD task. PSOD [16] proposes a novel weakly-supervised Salient Object Detection (SOD) method using point supervision. But the drawbacks of it are also obvious: 1) extra edge detectors and edge information 2) flood fill algorithm introducing additional computation, and not appropriate for COD since camouflaged objects are not salient." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 415, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 583 + ], + "type": "text", + "content": "Contrastive Learning. The core of contrastive learning is to attract the positive sample pairs and repulse the negative sample pairs. Moco [18] maintained a queue of negative samples and turned one branch of siamese network into a momentum encoder to improve the consistency of the queue. BYOL [17] proposed a self-supervised image representation learning approach, achieving state-of-the-art results without negative samples. The mainstream paradigm of contrastive learning later shifted to only using positive samples. SimSiam [3] built an efficient and simple contrastive learning framework only using positive samples, which is free from the limitations of large batch size, momentum encoder, and negative samples. The above work [3,17,18] has proved that UCL can completely approach or even exceed the effect of supervised manner, and learn representations with strong generalization. Therefore, we try to introduce unsupervised contrastive learning to point-supervised COD to strengthen the representation of the model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 603, + 233, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 603, + 233, + 617 + ], + "spans": [ + { + "bbox": [ + 132, + 603, + 233, + 617 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "The overall architecture of the proposed framework is shown in Fig. 4, which includes a) a hint area generator to extend supervision from a single point to the small hint region, b) an attention regulator to guide the model to focus" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 115, + 476, + 309 + ], + "blocks": [ + { + "bbox": [ + 137, + 115, + 476, + 309 + ], + "lines": [ + { + "bbox": [ + 137, + 115, + 476, + 309 + ], + "spans": [ + { + "bbox": [ + 137, + 115, + 476, + 309 + ], + "type": "image", + "image_path": "910730c2889d5f0baa168e045b9e7a6ce2ad29489ccc0f4c3d6791fe84d07ce4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 319, + 482, + 396 + ], + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 396 + ], + "type": "text", + "content": "Fig. 4: Details of our method. The main components consist of three parts: (a) hint area generator, (b) attention regulator, and (c) representation optimizer. Hint area generator extends the point label from a single point to the small hint region. Attention regulator enforce the model to focus on the whole object instead of being stuck in most discriminative part. Representation optimizer uses the unsupervised contrastive learning to learn a stable feature representation for disentangling camouflaged objects and background." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 421, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 482, + 458 + ], + "type": "text", + "content": "on the whole object instead of being stuck in local discriminative parts, and c) a representation optimizer based on unsupervised contrastive learning for disentangling camouflaged objects and background." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 475, + 266, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 475, + 266, + 487 + ], + "spans": [ + { + "bbox": [ + 131, + 475, + 266, + 487 + ], + "type": "text", + "content": "3.1 Hint Area Generator" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 616 + ], + "type": "text", + "content": "Compared with other forms of mask supervision, supervision provided by a point is rather limited, which easily leads to model collapse, as shown in Tab. 5. To explain a little, an annotated point just functions like a \"hint\", only indicating the existence of the object, with no knowledge of the area, shape, or boundary. To solve this, we design a hint area generator to adaptively expand a single point to a relatively larger region. One of the most crucial things is to control the degree of point expansion, avoiding the resultant hint area being too small (limited effect) or too large (out of boundary). Especially avoiding excessive expansion to introduce noise, place quality over quantity, because incorrect supervision is fatal for WSCOD, as shown in Tab. 5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Specifically, as shown in Fig. 4(a), we initialize the desired annotation region with a small square area around the annotated point. It just prevents the model training from collapsing if only using one pixel as a point. The side length of the square area is " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ". Then, after training the encoder " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " epochs as a warm-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "text", + "content": "Just a Hint: Point-Supervised Camouflaged Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "start (Short-term training), we hope this encoder could have a certain extent of semantic sense towards the object. Based on this, we estimate the pseudo area size of the camouflaged object, which is calculated by: " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "P = E_w(I)" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "E_w" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": " denotes the encoder after training " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": " epochs, " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": " represents the input image, and " + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": " is the obtained prediction map used to serve as the rough object area hint." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 482, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 482, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 482, + 223 + ], + "type": "text", + "content": "We choose circular to be the basic area shape, which could be intuitively regarded as \"a large point\". Then the key is to determine the reasonable radius of the circular area. Based on the object area hint " + }, + { + "bbox": [ + 130, + 176, + 482, + 223 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 176, + 482, + 223 + ], + "type": "text", + "content": ", the initialized radius " + }, + { + "bbox": [ + 130, + 176, + 482, + 223 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 176, + 482, + 223 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 256, + 223, + 481, + 248 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 223, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 256, + 223, + 481, + 248 + ], + "type": "interline_equation", + "content": "r = \\sqrt {\\frac {S u m (\\mathbb {I} (P > \\tau))}{N}}, \\tag {1}", + "image_path": "2081dec493de88dcef5a6583451eb1000cd90c22b7dcce24005f39f9ce824818.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 254, + 482, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 290 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 290 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 254, + 482, + 290 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 254, + 482, + 290 + ], + "type": "text", + "content": " is a threshold, " + }, + { + "bbox": [ + 130, + 254, + 482, + 290 + ], + "type": "inline_equation", + "content": "\\mathbb{I}(\\cdot)" + }, + { + "bbox": [ + 130, + 254, + 482, + 290 + ], + "type": "text", + "content": " is an indicator function, and " + }, + { + "bbox": [ + 130, + 254, + 482, + 290 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 254, + 482, + 290 + ], + "type": "text", + "content": " is the number of objects. Note that " + }, + { + "bbox": [ + 130, + 254, + 482, + 290 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 254, + 482, + 290 + ], + "type": "text", + "content": " could be achieved by counting the annotated point and thus is a deterministic parameter." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "text", + "content": "Then, to prevent the calculated circular from being out of the object boundary, we use a hyperparameter, " + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "text", + "content": ", to alter the radius " + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "text", + "content": " using: " + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "inline_equation", + "content": "R = \\frac{r}{\\alpha}" + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "text", + "content": ". Specifically, the annotation for an object " + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "text", + "content": " is enlarged from a single point " + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "inline_equation", + "content": "(x_{n},y_{n})" + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "text", + "content": ", to a small circle region around that point with the estimated radius " + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "inline_equation", + "content": "C_{(x_n,y_n)}^R" + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "text", + "content": ". It is noteworthy that the background is also annotated by such a small circular region, unfolding around the point " + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "inline_equation", + "content": "(x_b,y_b)" + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "text", + "content": ", which is randomly picked from the background region." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 374, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 422 + ], + "type": "text", + "content": "As a result, with the help of the hint area generator, the original point supervision for the image " + }, + { + "bbox": [ + 130, + 374, + 482, + 422 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 374, + 482, + 422 + ], + "type": "text", + "content": " is enriched into several small regions, expanding supervision while minimizing the introduction of incorrect supervision as much as possible, represented as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 212, + 433, + 481, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 433, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 212, + 433, + 481, + 456 + ], + "type": "interline_equation", + "content": "S (I) = C _ {\\left(x _ {b}, y _ {b}\\right)} ^ {R} \\cup \\sum_ {n} C _ {\\left(x _ {n}, y _ {n}\\right)} ^ {R}, \\quad n = 1, \\dots , N \\tag {2}", + "image_path": "77cba96d8db28189cc943322528893db1fb7adff289e7e1b8e9b3091c6c79c89.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 466, + 392, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 466, + 392, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 466, + 392, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 466, + 392, + 479 + ], + "type": "inline_equation", + "content": "S(I)" + }, + { + "bbox": [ + 131, + 466, + 392, + 479 + ], + "type": "text", + "content": " is the obtained supervision, as shown in Fig. 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 498, + 264, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 264, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 264, + 510 + ], + "type": "text", + "content": "3.2 Attention Regulator" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 519, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 482, + 628 + ], + "type": "text", + "content": "Weakly-supervised methods often suffer from partial detection problems, owing to the fact that model tends to focus too much on those discriminative regions. Inspired by the HaS method [21], the random mask is able to inhibit the strong response, we further design an attention regulator module for PCOD. Because in camouflaged scenes, people often annotate point labels on the most visually discriminative regions for the camouflaged objects, we leverage this prior to guide mask generation to scatter the attention from the directly supervised discriminative part to the whole object area, as shown in Fig. 4(b). Firstly, based on the given supervision region, " + }, + { + "bbox": [ + 130, + 519, + 482, + 628 + ], + "type": "inline_equation", + "content": "C^R" + }, + { + "bbox": [ + 130, + 519, + 482, + 628 + ], + "type": "text", + "content": ", we could obtain a mask " + }, + { + "bbox": [ + 130, + 519, + 482, + 628 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 519, + 482, + 628 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 246, + 637, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 637, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 246, + 637, + 481, + 669 + ], + "type": "interline_equation", + "content": "M = \\left\\{ \\begin{array}{l l} M _ {i} = h (Z), & i \\in C ^ {R} \\\\ M _ {i} = 1, & i \\notin C ^ {R} \\end{array} \\right. \\tag {3}", + "image_path": "3e3eaaedd0b4aaff978dd08774644c1efa9c77b9f6488ea1dcd764d7e4dedc19.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": " is a logical matrix consisting of zeros and ones with the same shape as " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": " denotes a shuffle operation, " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": " is the pixel index." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "content": "Then, we apply this mask to the original image " + }, + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "inline_equation", + "content": "I_{*} = I * M" + }, + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "content": ", where * indicates element-wise multiplication. Through such formulation, the annotated discriminative area will be randomly masked. During the training process, the model needs to explore the recognition of the surrounding foreground areas of the mask to restore recognition of the masked foreground area, helping the model scatter attention to other regions within the foreground objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 227, + 291, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 227, + 291, + 239 + ], + "spans": [ + { + "bbox": [ + 131, + 227, + 291, + 239 + ], + "type": "text", + "content": "3.3 Representation Optimizer" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 244, + 479, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 244, + 479, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 479, + 292 + ], + "type": "text", + "content": "In COD, the foreground and the background share high similarity. The subtle difference sometimes comes from the texture abruption or the color inconsistency. Based on such observations, we hope at least from one aspect the discrepancy could be exposed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 293, + 480, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 480, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 480, + 339 + ], + "type": "text", + "content": "Toward this goal, we propose to use unsupervised contrastive learning strategy to optimize the feature space, aiming to make the foreground object distinguished from the background through the learned invariant patterns and stable representations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 341, + 479, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 341, + 479, + 411 + ], + "spans": [ + { + "bbox": [ + 131, + 341, + 479, + 411 + ], + "type": "text", + "content": "Specifically, as shown in Fig. 4(c), for image " + }, + { + "bbox": [ + 131, + 341, + 479, + 411 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 131, + 341, + 479, + 411 + ], + "type": "text", + "content": ", we adopt two different visual transformations " + }, + { + "bbox": [ + 131, + 341, + 479, + 411 + ], + "type": "inline_equation", + "content": "T_{1}" + }, + { + "bbox": [ + 131, + 341, + 479, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 341, + 479, + 411 + ], + "type": "inline_equation", + "content": "T_{2}" + }, + { + "bbox": [ + 131, + 341, + 479, + 411 + ], + "type": "text", + "content": ", selecting from colorjitter, Gaussianblur, flip, etc. These visual transformations are able to change the appearance of images. For example, Gaussian blur will smooth the texture, and changing color is relatively harmless to the texture patterns. This process results in two transformed images " + }, + { + "bbox": [ + 131, + 341, + 479, + 411 + ], + "type": "inline_equation", + "content": "I^{1}" + }, + { + "bbox": [ + 131, + 341, + 479, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 341, + 479, + 411 + ], + "type": "inline_equation", + "content": "I^{2}" + }, + { + "bbox": [ + 131, + 341, + 479, + 411 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 131, + 341, + 479, + 411 + ], + "type": "inline_equation", + "content": "I^{1} = T_{1}(I)" + }, + { + "bbox": [ + 131, + 341, + 479, + 411 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 341, + 479, + 411 + ], + "type": "inline_equation", + "content": "I^{2} = T_{2}(I)" + }, + { + "bbox": [ + 131, + 341, + 479, + 411 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "text", + "content": "We encode the transformed images differently into two prediction maps " + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "inline_equation", + "content": "P^{(1)}" + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "inline_equation", + "content": "P^{(2)}" + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "text", + "content": ", denoting as " + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "inline_equation", + "content": "P^1 = g(f(I^1))" + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "inline_equation", + "content": "P^2 = f(I^2)" + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "text", + "content": " represents the encoder to be learned, and " + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "text", + "content": " is a small network cascaded over " + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "text", + "content": " to increase encoding diversity. When a position-related or size-related transformation " + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "text", + "content": " (e.g., scale, crop) is applied to the image " + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "inline_equation", + "content": "I^1" + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "text", + "content": ", this " + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "text", + "content": " should be applied to " + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "inline_equation", + "content": "P^2" + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "text", + "content": " to be aligned with " + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "inline_equation", + "content": "P^1" + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "text", + "content": ", and vice versa. Our objective is to minimize the distance between these two prediction maps:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 238, + 502, + 480, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 502, + 480, + 526 + ], + "spans": [ + { + "bbox": [ + 238, + 502, + 480, + 526 + ], + "type": "interline_equation", + "content": "\\min \\mathcal {D} (P ^ {1}, P ^ {2}) = \\sum_ {i} \\left| P _ {i} ^ {1} - P _ {i} ^ {2} \\right|, \\tag {4}", + "image_path": "f9bd2c4b6074f7751e9d9c0df3c0d0c6ad09b75de832a085267f036d79c8bcfe.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 532, + 479, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 532, + 479, + 566 + ], + "spans": [ + { + "bbox": [ + 131, + 532, + 479, + 566 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 532, + 479, + 566 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 532, + 479, + 566 + ], + "type": "text", + "content": " is the pixel index. Here we follow the design of UCL, i.e., stopping the gradient (stopgrad) update at one end, so the contrastive loss function is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 249, + 567, + 480, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 567, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 249, + 567, + 480, + 581 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\mathcal {D} \\left(P ^ {1}, \\operatorname {s t o p g r a d} \\left(P ^ {2}\\right)\\right). \\tag {5}", + "image_path": "9723a285aec93f67be682434da8b372e9ce6e02b6cb943af92b7904a9ba2d0cc.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 584, + 479, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 584, + 479, + 608 + ], + "spans": [ + { + "bbox": [ + 131, + 584, + 479, + 608 + ], + "type": "text", + "content": "The prediction gap using different transformations will be narrowed by minimizing the above loss, making the learned feature representation robust." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 624, + 204, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 204, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 204, + 635 + ], + "type": "text", + "content": "3.4 Network" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "Encoder. The structure of our feature encoder is simple compared to previous work. Specifically, in order to capture long-distance feature dependency and" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 103 + ], + "type": "text", + "content": "Just a Hint: Point-Supervised Camouflaged Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 165, + 485, + 345 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 1: Quantitative comparison with state-of-the-arts on four benchmarks. \"F\", \"U\", \"S\", \"P\" denote fully-supervised, unsupervised, scribble-supervised, and point-supervised methods, respectively. The best results are highlighted in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 165, + 485, + 345 + ], + "lines": [ + { + "bbox": [ + 132, + 165, + 485, + 345 + ], + "spans": [ + { + "bbox": [ + 132, + 165, + 485, + 345 + ], + "type": "table", + "html": "
MethodsSup.CAMO (250)COD10K (2026)NC4K (4121)
MAE\\( \\downarrow S_m \\)↑\\( E_m \\)↑\\( F_w^\\beta \\)↑MAE\\( \\downarrow S_m \\)↑\\( E_m \\)↑\\( F_w^\\beta \\)↑MAE\\( \\downarrow S_m \\)↑\\( E_m \\)↑\\( F_w^\\beta \\)↑
F3Net [38]F0.1090.7110.7410.5640.0510.7390.7950.5440.0690.7820.8250.706
CSNet [15]F0.0920.7710.7950.6410.0470.7780.8090.5690.0610.8190.8450.748
ITSD [45]F0.1020.7500.7790.6100.0510.7670.8080.5570.0640.8110.8450.729
MINet [30]F0.0900.7480.7910.6370.0420.770.8320.608----
PraNet [11]F0.0940.7690.8250.6630.0450.7890.8610.629----
UCNet [42]F0.0940.7390.7870.6400.0420.7760.8570.6330.0550.8130.8720.777
SINet [9]F0.0920.7450.8040.6440.0430.7760.8640.6310.0580.8080.8710.723
MGL-R [41]F0.0880.7750.8120.6730.0350.8140.8510.6660.0520.8330.8670.740
PFNet [27]F0.0850.7820.8410.6950.0400.8000.8770.6600.0530.8290.8870.745
UJSC [23]F0.0730.8000.8590.7280.0350.8090.8840.6840.0470.8420.8980.771
UGTR [39]F0.0860.7840.8220.6840.0360.8170.8520.6660.0520.8390.8740.747
ZoomNet [29]F0.0660.8200.8920.7520.0290.8380.9110.7290.0430.8530.8960.784
DUSD [44]U0.1660.5510.5940.3080.1070.5800.6460.276----
USPS [28]U0.2070.5680.6410.3990.1960.5190.5360.265----
SAM [20]U0.1320.6840.6870.6060.0500.7830.7980.7010.0780.7670.7760.696
SS [43]S0.1180.6960.7860.5620.0710.6840.7700.461----
SCSOD [40]S0.1020.7130.7950.6180.0550.7100.8050.546----
CRNet [19]S0.0920.7350.8150.6410.0490.7330.8320.5760.0630.7750.8550.688
OursP0.0740.7980.8720.7270.0420.7840.8590.6500.0510.8220.8890.748
", + "image_path": "2316259abe1c3346c6cdfa99bfc7355e4b70e6c090de7cf62ff20082a99672fc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "text", + "content": "obtain multi-scale information, we use PVT [36] as the backbone, for an input image " + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "text", + "content": ", we put it into the backbone to get the output features " + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "inline_equation", + "content": "F_{eat_i}" + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "text", + "content": "-th. Then, we get the multi-scale features " + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "inline_equation", + "content": "(F_{eat1}, F_{eat2}, F_{eat3}, F_{eat4})" + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "inline_equation", + "content": "\\left(\\frac{1}{4}, \\frac{1}{8}, \\frac{1}{16}, \\frac{1}{32}\\right)" + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "text", + "content": " resolution of input image " + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "text", + "content": ". We downsize the channel dimension of " + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "inline_equation", + "content": "F_{eat_i}" + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "text", + "content": " into 64 by using " + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "text", + "content": " convolutional layers. Next, these feature maps are unified into the same size by an up-sampling operation, and combined through the concatenation. Finally the output map " + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "inline_equation", + "content": "\\hat{S} \\in \\mathbb{R}^{1 \\times W \\times H}" + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "text", + "content": " is obtained by the " + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "text", + "content": " convolution layer, and the detailed network structure is provided in supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 491, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 529 + ], + "type": "text", + "content": "Loss. Compared to other weak-supervised methods [19,35,40,43], we have fewer losses, only two. The final loss " + }, + { + "bbox": [ + 130, + 491, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 130, + 491, + 482, + 529 + ], + "type": "text", + "content": " includes contrastive loss " + }, + { + "bbox": [ + 130, + 491, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 130, + 491, + 482, + 529 + ], + "type": "text", + "content": " and partial cross-entropy loss " + }, + { + "bbox": [ + 130, + 491, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pce}" + }, + { + "bbox": [ + 130, + 491, + 482, + 529 + ], + "type": "text", + "content": ", the former is defined in Eq. 5, and the latter is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 209, + 541, + 481, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 541, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 209, + 541, + 481, + 569 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p c e} = - \\sum_ {i \\in \\tilde {S}} S _ {i} \\log (\\hat {S} _ {i}) + (1 - S _ {i}) \\log (1 - \\hat {S} _ {i}), \\tag {6}", + "image_path": "ba7787a3877ff8f114a5d47c4fc4eccfc61c04d6c7a2e25a05c1ddf9503e6dd3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 578, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 482, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 578, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\tilde{S}" + }, + { + "bbox": [ + 130, + 578, + 482, + 604 + ], + "type": "text", + "content": " represents the labeled area of " + }, + { + "bbox": [ + 130, + 578, + 482, + 604 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 578, + 482, + 604 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 578, + 482, + 604 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 130, + 578, + 482, + 604 + ], + "type": "text", + "content": " is the true class of pixel " + }, + { + "bbox": [ + 130, + 578, + 482, + 604 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 578, + 482, + 604 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 578, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\hat{S}_i" + }, + { + "bbox": [ + 130, + 578, + 482, + 604 + ], + "type": "text", + "content": " is the prediction of pixel " + }, + { + "bbox": [ + 130, + 578, + 482, + 604 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 578, + 482, + 604 + ], + "type": "text", + "content": ". So the final loss " + }, + { + "bbox": [ + 130, + 578, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 130, + 578, + 482, + 604 + ], + "type": "text", + "content": " can be defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 274, + 616, + 481, + 629 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 616, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 274, + 616, + 481, + 629 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c} + \\mathcal {L} _ {p c e}. \\tag {7}", + "image_path": "cf4f82561bea702d73ebdab1fa959a30e700d0ae2d82f9386bb6abe4c9d17bc1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Even though we only use point label in training, our loss is generalizable to the other weakly-supervised COD models by contrastive learning." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 164, + 114, + 449, + 285 + ], + "blocks": [ + { + "bbox": [ + 164, + 114, + 449, + 285 + ], + "lines": [ + { + "bbox": [ + 164, + 114, + 449, + 285 + ], + "spans": [ + { + "bbox": [ + 164, + 114, + 449, + 285 + ], + "type": "image", + "image_path": "f976a0ff99afb29e8cf41d88adcb7c1c1111fcf55234bd329eccfdbd98a42d9b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 295, + 480, + 317 + ], + "lines": [ + { + "bbox": [ + 131, + 295, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 131, + 295, + 480, + 317 + ], + "type": "text", + "content": "Fig. 5: Visual comparison with representative scribble-supervised and fully-supervised models." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 141, + 350, + 299, + 393 + ], + "blocks": [ + { + "bbox": [ + 143, + 329, + 296, + 340 + ], + "lines": [ + { + "bbox": [ + 143, + 329, + 296, + 340 + ], + "spans": [ + { + "bbox": [ + 143, + 329, + 296, + 340 + ], + "type": "text", + "content": "Table 2: Ablation on point position." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 350, + 299, + 393 + ], + "lines": [ + { + "bbox": [ + 141, + 350, + 299, + 393 + ], + "spans": [ + { + "bbox": [ + 141, + 350, + 299, + 393 + ], + "type": "table", + "html": "
TypeMAE↓Sm↑Em↑Fwβ↑
Center0.0790.7890.8580.716
Random0.0760.7910.8690.722
Ours0.0740.7980.8720.727
", + "image_path": "0db3fda2cfadadb49837aa994fc8d63c0d9e43c39efd46f61a078c96388f3b86.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 306, + 350, + 482, + 393 + ], + "blocks": [ + { + "bbox": [ + 312, + 329, + 466, + 341 + ], + "lines": [ + { + "bbox": [ + 312, + 329, + 466, + 341 + ], + "spans": [ + { + "bbox": [ + 312, + 329, + 466, + 341 + ], + "type": "text", + "content": "Table 3: Ablation on point numbers." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 350, + 482, + 393 + ], + "lines": [ + { + "bbox": [ + 306, + 350, + 482, + 393 + ], + "spans": [ + { + "bbox": [ + 306, + 350, + 482, + 393 + ], + "type": "table", + "html": "
Number of pointsMAE↓Sm↑Em↑Fwβ↑
One-point0.0760.7910.8690.722
Two-points0.0740.8020.8810.737
Three-points0.0740.8050.8870.741
", + "image_path": "5ce5f7ae8abe58eda3acca947f8d6194944b68bc3b1b84bb01a5608a52a93632.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 417, + 229, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 417, + 229, + 431 + ], + "spans": [ + { + "bbox": [ + 131, + 417, + 229, + 431 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 442, + 261, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 442, + 261, + 455 + ], + "spans": [ + { + "bbox": [ + 131, + 442, + 261, + 455 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 462, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 481, + 569 + ], + "type": "text", + "content": "Point Dataset. Our experiments are conducted on three COD benchmarks, CAMO [22], COD10K [9], and NC4K [25]. In order to evaluate our point-supervised COD method, we relabel 4040 images (3040 from COD10K, 1000 from CAMO) and propose the Point-supervised Dataset (P-COD) for training and the remaining is for testing. In Fig. 1, we show an example of point annotation and compare it with other annotation methods. For each image, we simulate the hunting process to label only one point for each camouflaged object, without needing to worry about ambiguous boundaries, making the process of the annotation easy and natural." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Evaluation Metrics. We adopt four evaluation metrics: mean absolute error (MAE), S-measure " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "(\\mathrm{S}_m)" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " [6], E-measure " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "(\\mathrm{E}_m)" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " [7], weighted F-measure " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "(\\mathrm{F}_{\\beta}^{w})" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " [26]. Implementation Details. We implement our method with PyTorch and conduct experiments on one GeForce RTX4090 GPU. We use the stochastic gradient descent optimizer with a momentum of 0.9, a weight decay of " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "5e - 4" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ", and triangle learning rate schedule with maximum learning rate " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "1e - 3" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ". The batch size is 8, and the training epoch is 60. It takes around 7 hours to train. During training and inference, input images are resized to " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "512\\times 512" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "text", + "content": "Just a Hint: Point-Supervised Camouflaged Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 136, + 481, + 191 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 472, + 126 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 472, + 126 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 472, + 126 + ], + "type": "text", + "content": "Table 4: Comparison of parameters and MACs. * is reset in the same input size." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 136, + 481, + 191 + ], + "lines": [ + { + "bbox": [ + 133, + 136, + 481, + 191 + ], + "spans": [ + { + "bbox": [ + 133, + 136, + 481, + 191 + ], + "type": "table", + "html": "
MethodsSup.Input sizePara(M)MACs(G)MAE↓Sm↑Em↑Fwβ↑
SINetF352248.9519.420.0920.7450.8040.644
CRNet*S352232.6514.270.0950.7300.8040.634
OursP352225.4410.220.0880.7610.8360.677
", + "image_path": "066f1f6f08500199256c7a4f23a01648e5a25afb09cecd3702bd5f44fa8c09b3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 140, + 282, + 302, + 342 + ], + "blocks": [ + { + "bbox": [ + 135, + 217, + 304, + 273 + ], + "lines": [ + { + "bbox": [ + 135, + 217, + 304, + 273 + ], + "spans": [ + { + "bbox": [ + 135, + 217, + 304, + 273 + ], + "type": "text", + "content": "Table 5: Effect of hint area generator with different settings. Point, I, P and SAM denote point label, initial square area, prediction map P and P obtained by SAM as proposal area, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 282, + 302, + 342 + ], + "lines": [ + { + "bbox": [ + 140, + 282, + 302, + 342 + ], + "spans": [ + { + "bbox": [ + 140, + 282, + 302, + 342 + ], + "type": "table", + "html": "
SettingsMAE↓Sm↑Em↑Fwβ↑
w/ Pointcollapse
w/ I0.0920.7540.8020.655
w/ P0.1030.7260.8130.627
w/ SAM0.1080.6940.7710.603
w/ Ours0.0850.7590.8420.677
", + "image_path": "79ae56ea26309a2111b9e413dd73c1bf3808c10834fbcc6279c17dd2430427aa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 281, + 471, + 342 + ], + "blocks": [ + { + "bbox": [ + 308, + 217, + 471, + 272 + ], + "lines": [ + { + "bbox": [ + 308, + 217, + 471, + 272 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 471, + 272 + ], + "type": "text", + "content": "Table 6: Ablation study on contrastive losses. MSE, KL, COS, L1 mean Mean Square Error, Kullback-Leibler divergence, Consine similarity, and L1 loss, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 281, + 471, + 342 + ], + "lines": [ + { + "bbox": [ + 309, + 281, + 471, + 342 + ], + "spans": [ + { + "bbox": [ + 309, + 281, + 471, + 342 + ], + "type": "table", + "html": "
LossMAE↓Sm↑Em↑Fwβ↑
w/o0.0850.7590.8420.677
MSE0.0790.7910.8510.708
KL0.0850.7690.8440.686
COS0.0870.7660.8440.680
L10.0740.7980.8720.727
", + "image_path": "81f8c8e04b8d493154f873e883557b38ace3d4c4638452b6ce96850b3247aa4a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 370, + 319, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 319, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 319, + 383 + ], + "type": "text", + "content": "4.2 Compare with State-of-the-arts" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 398, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 482, + 518 + ], + "type": "text", + "content": "Quantitative Comparison. Being the first point supervised COD method, our proposed approach is primarily benchmarked against scribble-supervised and fully-supervised methods. As demonstrated in Tab. 1, our method achieves substantial improvements, with an average enhancement of " + }, + { + "bbox": [ + 130, + 398, + 482, + 518 + ], + "type": "inline_equation", + "content": "17.6\\%" + }, + { + "bbox": [ + 130, + 398, + 482, + 518 + ], + "type": "text", + "content": " for MAE, " + }, + { + "bbox": [ + 130, + 398, + 482, + 518 + ], + "type": "inline_equation", + "content": "7.2\\%" + }, + { + "bbox": [ + 130, + 398, + 482, + 518 + ], + "type": "text", + "content": " for Sm, " + }, + { + "bbox": [ + 130, + 398, + 482, + 518 + ], + "type": "inline_equation", + "content": "4.7\\%" + }, + { + "bbox": [ + 130, + 398, + 482, + 518 + ], + "type": "text", + "content": " for Em, and " + }, + { + "bbox": [ + 130, + 398, + 482, + 518 + ], + "type": "inline_equation", + "content": "11.7\\%" + }, + { + "bbox": [ + 130, + 398, + 482, + 518 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 398, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_{\\beta}^{w}" + }, + { + "bbox": [ + 130, + 398, + 482, + 518 + ], + "type": "text", + "content": " compared to the state-of-the-art weakly-supervised COD method CRNet [19]. This highlights our capability to achieve better performance with fewer annotations. Our approach even outperforms fully supervised methods in several metrics across multiple datasets. Specifically, it excels against the majority of fully-supervised methods on CAMO and outpaces nearly half of them on other datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 520, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 592 + ], + "type": "text", + "content": "Qualitative Evaluation. Our method produces prediction maps characterized by clearer and more complete object regions, along with sharper contours, significantly outperforming state-of-the-art weakly supervised COD method CRNet [19], as shown in Fig. 5. Our method performs well in various challenging scenarios, including scenarios with tiny objects (row 4), high intrinsic similarities (row 3, 5), indefinable boundaries (row 1), and complex backgrounds (row 2, 6)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Parameter Complexity. With lower parameter complexity and minimal computational cost, our model (25.44M parameters and 10.22G MACs) consistently outperforms the fully supervised SINet (48.95M parameters and 19.42G MACs) by " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " for MAE, " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " for Sm, and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "4.0\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " for Em on the challenging CAMO dataset. It also outperforms the scribble supervised CRNet (32.65M parameters and 14.27G MACs) by " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "7.4\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " for MAE, " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " for Sm, " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "4.0\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " for Em, and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "6.8\\%" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 164, + 114, + 451, + 239 + ], + "blocks": [ + { + "bbox": [ + 164, + 114, + 451, + 239 + ], + "lines": [ + { + "bbox": [ + 164, + 114, + 451, + 239 + ], + "spans": [ + { + "bbox": [ + 164, + 114, + 451, + 239 + ], + "type": "image", + "image_path": "706e18944d507ffe6f22dcf181193e39008fc227bda5db63aa63e3d5679427e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 247, + 480, + 270 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 480, + 270 + ], + "type": "text", + "content": "Fig. 6: Visualization of the component ablation study. I, II, III respectively represent hint area generator, representation optimizer, and attention regulator." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 162, + 324, + 452, + 407 + ], + "blocks": [ + { + "bbox": [ + 130, + 281, + 480, + 314 + ], + "lines": [ + { + "bbox": [ + 130, + 281, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 480, + 314 + ], + "type": "text", + "content": "Table 7: The ablation study for different augmentations in contrastive learning. \"S\", \"C\", \"F\", \"T\", \"R\", \"G\", \"J\" are Scale, Crop, Flip, Translate, Attention Regulator, Guassblur, Colorjitter." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 162, + 324, + 452, + 407 + ], + "lines": [ + { + "bbox": [ + 162, + 324, + 452, + 407 + ], + "spans": [ + { + "bbox": [ + 162, + 324, + 452, + 407 + ], + "type": "table", + "html": "
SCFTRGJMAE↓Sm↑Em↑Fwβ↑
0.0850.7590.8420.677
✓0.0820.7790.8530.701
✓✓0.0800.7820.8600.713
✓✓✓0.0800.7710.8450.707
✓✓✓✓0.0770.7810.8530.711
✓✓✓✓✓0.0760.7900.8660.724
✓✓✓✓✓✓0.0740.7940.8670.729
✓✓✓✓✓✓✓0.0740.7980.8720.727
", + "image_path": "9f257ad084ebdf1169a766d674959e90c30b094ba22601f0a215bdd6f9ff3a5c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_{\\beta}^{w}" + }, + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "type": "text", + "content": ", as shown in Tab. 4. These results underscore the superior lightweight performance of our model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 480, + 238, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 480, + 238, + 493 + ], + "spans": [ + { + "bbox": [ + 131, + 480, + 238, + 493 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 480, + 521 + ], + "type": "text", + "content": "As CAMO dataset is the most challenging one, indicated by the lowest scores in Tab. 1, all following ablation experiments are performed on it." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 522, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 481, + 641 + ], + "type": "text", + "content": "Ablation Study on Point Supervision. To reduce annotation costs, we intuitively annotated a point for each camouflaged object by humans. We conduct experiments on point positions, as shown in Tab. 2. For Random, we randomly choose a point from real labels, and repeat three times, taking the average of the results. which shows our method far exceeds the center annotation and random annotation under conditions of low annotation cost. We attribute it to the model's utilization of the priors implicit in point annotations to simulate human cognitive processes. In addition, we conduct experiments for point numbers (Random), as shown in Tab. 3. One point yields satisfactory performance. Increasing the points shows diminishing returns." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Effectiveness of Hint Area Generator. As shown in Tab. 5, our proposed hint area generator obtains a significant improvement over the direct use of point" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 448, + 102 + ], + "type": "text", + "content": "Just a Hint: Point-Supervised Camouflaged Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 185, + 136, + 430, + 180 + ], + "blocks": [ + { + "bbox": [ + 169, + 114, + 443, + 126 + ], + "lines": [ + { + "bbox": [ + 169, + 114, + 443, + 126 + ], + "spans": [ + { + "bbox": [ + 169, + 114, + 443, + 126 + ], + "type": "text", + "content": "Table 8: Effect of different masking setting in Attention regulator." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 185, + 136, + 430, + 180 + ], + "lines": [ + { + "bbox": [ + 185, + 136, + 430, + 180 + ], + "spans": [ + { + "bbox": [ + 185, + 136, + 430, + 180 + ], + "type": "table", + "html": "
Basic settingTypeMAE↓Sm↑Em↑FWβ↑
w/ aug (S, C, T, F)HaS [21]0.0780.7850.8640.711
w/ aug (S, C, T, F)Cutout [5]0.0760.7890.8620.721
w/ aug (S, C, T, F)Ours0.0760.7900.8660.724
", + "image_path": "aaafe84bbaa85f4067d8ef9fc788c2a13b3bb890f85cda6112dedcb121cad936.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 136, + 240, + 481, + 283 + ], + "blocks": [ + { + "bbox": [ + 136, + 207, + 303, + 228 + ], + "lines": [ + { + "bbox": [ + 136, + 207, + 303, + 228 + ], + "spans": [ + { + "bbox": [ + 136, + 207, + 303, + 228 + ], + "type": "text", + "content": "Table 9: Effect of key components in optimizer." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 207, + 470, + 229 + ], + "lines": [ + { + "bbox": [ + 309, + 207, + 470, + 229 + ], + "spans": [ + { + "bbox": [ + 309, + 207, + 470, + 229 + ], + "type": "text", + "content": "Table 10: Transferability studies on scribble datasets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 240, + 481, + 283 + ], + "lines": [ + { + "bbox": [ + 136, + 240, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 136, + 240, + 481, + 283 + ], + "type": "table", + "html": "
stop-grad predictorMAE↓Sm↑Em↑Fwβ↑MethodsSup.MAE↓Sm↑Em↑Fwβ↑
✓×unstableZoomNet [29]F0.0660.8200.8920.752
×✓0.0810.7680.8520.706CRNet [19]S0.0920.7350.8150.641
✓✓0.0740.7980.8720.727Ours + scribbleS0.0650.8160.8940.761
", + "image_path": "99006204fde38a373c62e1520f5124a76382f379d02e4b03d1a562c8ef2b2a73.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 313, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 480, + 422 + ], + "type": "text", + "content": "label, Init square area and Prediction map. Because it has more supervision compared to the Init square area and fewer errors (e.g., background noise) compared to prediction map and prediction map obtained by SAM. Interestingly, the init square area performs much better than prediction map P and P obtained by SAM, and we think that the accuracy of supervision in WSCOD is more crucial than the quantity, since init square area has better accuracy and prediction map has high quantity in area generation but contains a lot of background noise. In addition, compared with prediction map, a relatively complete prediction map can be obtained by using hint area generator, as shown in Fig. 6." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 423, + 481, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 481, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 481, + 484 + ], + "type": "text", + "content": "Effectiveness of Attention Regulator. We propose attention regulator and apply it with UCL. As shown in Tab. 7, the results (Bold) are significantly improved after using attention regulator. In addition, to further demonstrate the superiority of regulator in weakly-supervised COD, we compare it with the previous mask approaches HaS [21] and Cutout [5], results are shown in Tab. 8." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "content": "Effectiveness of Representation Optimizer. We conduct ablation experiments for representation optimizer. Firstly, we test different types of contrastive loss and find that L1 loss performs best, with a significant improvement over the baseline, as shown in Tab. 6. In addition, we conduct exhaustive experiments for data augmentation, which is important for UCL in representation optimizer. Specifically, the augmentation is divided into three aspects, color-texture related (ColorJitter, GuassBlur), position related (Flip, Translate), and size related (Crop, Scale). It shows that color-texture and size augmentation can improve performance, but position augmentation yields unsatisfactory results, as shown in Tab. 7. We employ UCL exclusively with positive samples and test the key components in ablation study, as shown in Tab. 9. Fig. 6 intuitively illustrates that the representation optimizer enhances the precision of prediction maps. Fig. 7 and 8 show that our model is able to continuously optimize representation through representation optimizer and separate entangled objects from background, making the model eventually learn robust representations." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "H. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 153, + 136, + 461, + 178 + ], + "blocks": [ + { + "bbox": [ + 168, + 114, + 444, + 126 + ], + "lines": [ + { + "bbox": [ + 168, + 114, + 444, + 126 + ], + "spans": [ + { + "bbox": [ + 168, + 114, + 444, + 126 + ], + "type": "text", + "content": "Table 11: The impact of " + }, + { + "bbox": [ + 168, + 114, + 444, + 126 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 168, + 114, + 444, + 126 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 168, + 114, + 444, + 126 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 168, + 114, + 444, + 126 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 168, + 114, + 444, + 126 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 168, + 114, + 444, + 126 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 168, + 114, + 444, + 126 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 168, + 114, + 444, + 126 + ], + "type": "text", + "content": " in the hint area generator." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 136, + 461, + 178 + ], + "lines": [ + { + "bbox": [ + 153, + 136, + 461, + 178 + ], + "spans": [ + { + "bbox": [ + 153, + 136, + 461, + 178 + ], + "type": "table", + "html": "
dMAE↓Sm↑τMAE↓Sm↑αMAE↓Sm↑wMAE↓Sm↑
50.0850.7581000.0860.75730.0910.73950.0890.755
100.0850.7591500.0850.75940.0850.759100.0870.759
150.0860.7602000.0850.75850.0920.736150.0850.759
", + "image_path": "c40725fee41b6903bf12d2b0463e75f9c75bc6c1712266aefbd1343c40192b0d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 132, + 252, + 491, + 328 + ], + "blocks": [ + { + "bbox": [ + 130, + 201, + 480, + 234 + ], + "lines": [ + { + "bbox": [ + 130, + 201, + 480, + 234 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 480, + 234 + ], + "type": "text", + "content": "Table 12: Quantitative comparison with state-of-the-arts on four popular SOD benchmarks. \"F\", \"S\", \"P\" denote fully-, scribble-, and point-supervised methods, respectively. The best are in bold." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 252, + 491, + 328 + ], + "lines": [ + { + "bbox": [ + 132, + 252, + 491, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 491, + 328 + ], + "type": "table", + "html": "
MethodsSup.ECSSDDUT-OHKU-ISDUTS-TE
MAE↓Sm↑FβmaxMAE↓Sm↑FβmaxMAE↓Sm↑FβmaxMAE↓Sm↑Fβmax
AFNet [13]F0.0420.9130.9350.0570.8260.7970.0360.9050.9230.0460.8670.863
BASNet [33]F0.0370.9160.9430.0570.8360.8050.0320.9090.9280.0480.8660.859
MFNet [31]S0.0840.8340.8790.0870.7410.7060.0590.8460.8760.7650.7740.770
SCSOD [40]S0.0490.8810.9140.0600.8110.7820.0380.8820.9080.0490.8530.858
PSOD [16]P0.0360.9130.9350.0640.8240.8080.0330.9010.9230.0450.8530.858
OursP0.0450.8950.9150.0590.8260.8240.0310.9060.9260.0420.8640.861
", + "image_path": "1a23bc3f81a574b984a8002492b824a021489b79aaae3996aa22caade3a4c6a0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 363, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 495 + ], + "type": "text", + "content": "Impact of the Hyperparameter " + }, + { + "bbox": [ + 130, + 363, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\tau, \\alpha, w, d" + }, + { + "bbox": [ + 130, + 363, + 482, + 495 + ], + "type": "text", + "content": ". Four hyperparameters exist in hint area generator: " + }, + { + "bbox": [ + 130, + 363, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\tau, \\alpha, w, d" + }, + { + "bbox": [ + 130, + 363, + 482, + 495 + ], + "type": "text", + "content": ". We test diverse values for these parameters, as shown in Tab. 11. So we set " + }, + { + "bbox": [ + 130, + 363, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\tau = 150" + }, + { + "bbox": [ + 130, + 363, + 482, + 495 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 363, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\alpha = 4" + }, + { + "bbox": [ + 130, + 363, + 482, + 495 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 363, + 482, + 495 + ], + "type": "inline_equation", + "content": "w = 15" + }, + { + "bbox": [ + 130, + 363, + 482, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 363, + 482, + 495 + ], + "type": "inline_equation", + "content": "d = 10" + }, + { + "bbox": [ + 130, + 363, + 482, + 495 + ], + "type": "text", + "content": ". It is worth noting that the values of " + }, + { + "bbox": [ + 130, + 363, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\tau, w" + }, + { + "bbox": [ + 130, + 363, + 482, + 495 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 363, + 482, + 495 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 363, + 482, + 495 + ], + "type": "text", + "content": " do not have much impact on the results with different settings. In addition, the " + }, + { + "bbox": [ + 130, + 363, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 363, + 482, + 495 + ], + "type": "text", + "content": " can be replaced with unsupervised k-means clustering method (in S.M.). Our model is robust to parameters. It is because the hint area generator is designed to approximate the size of the camouflaged object through the prediction map. Whether the contour is fine and details are perfect, in the prediction map, has little influence on the estimation of the object size. The accurate estimation and localization of the camouflaged object are explored by the attention regulator and the representation optimizer modules." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 513, + 276, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 513, + 276, + 525 + ], + "spans": [ + { + "bbox": [ + 131, + 513, + 276, + 525 + ], + "type": "text", + "content": "4.4 Transferability Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": "Extension to Scribble. We apply our method to the scribble dataset, and the results are shown in Tab. 10 (More results in supplementary materials) reveal that our model, utilizing scribble labeling, achieves competitive results compared to the state-of-the-art fully supervised method ZoomNet [29]. It showcases the generalization ability of our method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Extension to SOD task. Our method excels not only in COD but also demonstrates competitive performance in SOD, as shown in Tab. 12. Considering the difference between these two tasks, we may wonder why our method consistently performs well on such two seemingly different tasks. We attribute this to the generality and rationality of the designed structure. In fact, point-supervised SOD and COD have a clear commonality, i.e., the demands of sufficient and accurate" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "text", + "content": "Just a Hint: Point-Supervised Camouflaged Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 165, + 114, + 449, + 187 + ], + "blocks": [ + { + "bbox": [ + 165, + 114, + 449, + 187 + ], + "lines": [ + { + "bbox": [ + 165, + 114, + 449, + 187 + ], + "spans": [ + { + "bbox": [ + 165, + 114, + 449, + 187 + ], + "type": "image", + "image_path": "d350be7a340903c426077ae45634faaf88c5e47c202d044adc3c808396ac563f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 196, + 480, + 218 + ], + "lines": [ + { + "bbox": [ + 131, + 196, + 480, + 218 + ], + "spans": [ + { + "bbox": [ + 131, + 196, + 480, + 218 + ], + "type": "text", + "content": "Fig. 7: Visualization of the feature space. Entangled features from foregrounds and backgrounds could be well separated by representation optimizer (visualized by t-SNE)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 164, + 229, + 451, + 319 + ], + "blocks": [ + { + "bbox": [ + 164, + 229, + 451, + 319 + ], + "lines": [ + { + "bbox": [ + 164, + 229, + 451, + 319 + ], + "spans": [ + { + "bbox": [ + 164, + 229, + 451, + 319 + ], + "type": "image", + "image_path": "6f27c4442cf677f4aec15f1aa56003aeb49177f619a735755b40a3ee9999ff2c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 328, + 480, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 328, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 480, + 361 + ], + "type": "text", + "content": "Fig. 8: Visualization of results in different iterations. As training proceeds, the entangled features from the foreground and background can be progressively separated by the representation optimizer (visualization by t-SNE)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 384, + 482, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 480 + ], + "type": "text", + "content": "supervision information. Through an effective point label expansion strategy, our model can provide more supervision information than a single point and more accurate supervision information than pseudo-labels, aiding in the precise segmentation of objects. Additionally, the proposed UCL optimizes the representation of weakly-supervised models, reducing segmentation instability. Despite the saliency of objects in WSSOD, it can also benefit from UCL. All these components are built on the common demands of the two tasks, providing a solid foundation for performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 498, + 220, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 220, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 220, + 510 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "text", + "content": "In this paper, we proposed an effective solution for COD by utilizing point-based supervision. The introduction of the P-COD dataset, based on human cognitive processes, streamlines the annotation process. To overcome the challenge of single-point supervision, we designed a hint area generator bolstering supervision. Representation optimizer optimized model representation by UCL. The attention regulator promoted a holistic object perception, ensuring a comprehensive understanding of camouflaged objects without fixating on distinct parts. Experimental results show that our method outperforms existing weakly-supervised methods largely and even surpasses many fully-supervised ones." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "Acknowledgement. This work is supported by the National Natural Science Foundation of China(NSFC) under Grant 62306239 and the Young Talent Fund of Association for Science and Technology in Shaanxi, China." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "H. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "text", + "content": "1. Bearman, A., Russakovsky, O., Ferrari, V., Fei-Fei, L.: What's the point: Semantic segmentation with point supervision. In: European conference on computer vision. pp. 549-565. Springer (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "text", + "content": "2. Benenson, R., Popov, S., Ferrari, V.: Large-scale interactive object segmentation with human annotators. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11700-11709 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "type": "text", + "content": "3. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15750-15758 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "type": "text", + "content": "4. Cheng, B., Parkhi, O., Kirillov, A.: Pointly-supervised instance segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2617-2626 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "text", + "content": "5. DeVries, T., Taylor, G.W.: Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 293, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 326 + ], + "type": "text", + "content": "6. Fan, D.P., Cheng, M.M., Liu, Y., Li, T., Borji, A.: Structure-measure: A new way to evaluate foreground maps. In: Proceedings of the IEEE international conference on computer vision. pp. 4548-4557 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "type": "text", + "content": "7. Fan, D.P., Gong, C., Cao, Y., Ren, B., Cheng, M.M., Borji, A.: Enhanced-alignment measure for binary foreground map evaluation. arXiv preprint arXiv:1805.10421 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 358, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 358, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 358, + 480, + 380 + ], + "type": "text", + "content": "8. Fan, D.P., Ji, G.P., Cheng, M.M., Shao, L.: Concealed object detection. IEEE transactions on pattern analysis and machine intelligence 44(10), 6024-6042 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 413 + ], + "type": "text", + "content": "9. Fan, D.P., Ji, G.P., Sun, G., Cheng, M.M., Shen, J., Shao, L.: Camouflaged object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2777-2787 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 414, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 435 + ], + "type": "text", + "content": "10. Fan, D.P., Ji, G.P., Xu, P., Cheng, M.M., Sakaridis, C., Van Gool, L.: Advances in deep concealed scene understanding. Visual Intelligence 1(1), 16 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 435, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 480, + 479 + ], + "type": "text", + "content": "1. Fan, D.P., Ji, G.P., Zhou, T., Chen, G., Fu, H., Shen, J., Shao, L.: Pranet: Parallel reverse attention network for polyp segmentation. In: International conference on medical image computing and computer-assisted intervention. pp. 263-273. Springer (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 479, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 512 + ], + "type": "text", + "content": "2. Fan, D.P., Zhou, T., Ji, G.P., Zhou, Y., Chen, G., Fu, H., Shen, J., Shao, L.: Inf-net: Automatic Covid-19 lung infection segmentation from ct images. IEEE transactions on medical imaging 39(8), 2626-2637 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "text", + "content": "3. Feng, M., Lu, H., Ding, E.: Attentive feedback network for boundary-aware salient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1623-1632 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "type": "text", + "content": "4. Pérez-de la Fuente, R., Delclos, X., Penalver, E., Speranza, M., Wierzchos, J., Ascaso, C., Engel, M.S.: Early evolution and ecology of camouflage in insects. Proceedings of the National Academy of Sciences 109(52), 21414-21419 (2012)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 610 + ], + "type": "text", + "content": "5. Gao, S.H., Tan, Y.Q., Cheng, M.M., Lu, C., Chen, Y., Yan, S.: Highly efficient salient object detection with 100k parameters. In: European Conference on Computer Vision. pp. 702-721. Springer (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "text", + "content": "6. Gao, S., Zhang, W., Wang, Y., Guo, Q., Zhang, C., He, Y., Zhang, W.: Weakly-supervised salient object detection using point supervision. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 670-678 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "7. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "text", + "content": "Just a Hint: Point-Supervised Camouflaged Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 147, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 117, + 480, + 138 + ], + "type": "text", + "content": "latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 481, + 171 + ], + "type": "text", + "content": "18. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 481, + 203 + ], + "type": "text", + "content": "19. He, R., Dong, Q., Lin, J., Lau, R.W.: Weakly-supervised camouflaged object detection with scribble annotations. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 781-789 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 481, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 481, + 235 + ], + "type": "text", + "content": "20. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 235, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 481, + 267 + ], + "type": "text", + "content": "21. Kumar Singh, K., Jae Lee, Y.: Hide-and-seek: Forcing a network to be meticulous for weakly-supervised object and action localization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 3524-3533 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "type": "text", + "content": "22. Le, T.N., Nguyen, T.V., Nie, Z., Tran, M.T., Sugimoto, A.: Anabranch network for camouflaged object segmentation. Computer vision and image understanding 184, 45-56 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "text", + "content": "23. Li, A., Zhang, J., Lv, Y., Liu, B., Zhang, T., Dai, Y.: Uncertainty-aware joint salient object and camouflaged object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10071-10081 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "text", + "content": "24. Liew, J., Wei, Y., Xiong, W., Ong, S.H., Feng, J.: Regional interactive image segmentation networks. In: 2017 IEEE international conference on computer vision (ICCV). pp. 2746-2754. IEEE (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 407 + ], + "type": "text", + "content": "25. Lv, Y., Zhang, J., Dai, Y., Li, A., Liu, B., Barnes, N., Fan, D.P.: Simultaneously localize, segment and rank the camouflaged objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11591-11601 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "text", + "content": "26. Margolin, R., Zelnik-Manor, L., Tal, A.: How to evaluate foreground maps? In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 248-255 (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "text", + "content": "27. Mei, H., Ji, G.P., Wei, Z., Yang, X., Wei, X., Fan, D.P.: Camouflaged object segmentation with distraction mining. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8772-8781 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "text", + "content": "28. Nguyen, T., Dax, M., Mummadi, C.K., Ngo, N., Nguyen, T.H.P., Lou, Z., Brox, T.: Deepusps: Deep robust unsupervised saliency prediction via self-supervision. Advances in Neural Information Processing Systems 32 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "text", + "content": "29. Pang, Y., Zhao, X., Xiang, T.Z., Zhang, L., Lu, H.: Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 2160-2170 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "text", + "content": "30. Pang, Y., Zhao, X., Zhang, L., Lu, H.: Multi-scale interactive network for salient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9413-9422 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "31. Piao, Y., Wang, J., Zhang, M., Lu, H.: Mfnet: Multi-filter directive network for weakly supervised salient object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4136-4145 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "32. Qian, R., Wei, Y., Shi, H., Li, J., Liu, J., Huang, T.: Weakly supervised scene parsing with point-based distance metric learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 8843-8850 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "33. Qin, X., Zhang, Z., Huang, C., Gao, C., Dehghan, M., Jagersand, M.: Basnet: Boundary-aware salient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7479-7489 (2019)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "H. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 545 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "34. Shin, G., Xie, W., Albanie, S.: All you need are a few pixels: semantic segmentation with pixelpick. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1687-1697 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "text", + "content": "35. Wang, B., Zhao, Y., Yang, L., Long, T., Li, X.: Temporal action localization in the deep learning era: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 226 + ], + "type": "text", + "content": "36. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 568-578 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "text", + "content": "37. Wei, J., Wang, Q., Li, Z., Wang, S., Zhou, S.K., Cui, S.: Shallow feature matters for weakly supervised object localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5993-6001 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "text", + "content": "38. Wei, J., Wang, S., Huang, Q.: " + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^3" + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "text", + "content": " net: fusion, feedback and focus for salient object detection. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 12321-12328 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 335 + ], + "type": "text", + "content": "39. Yang, F., Zhai, Q., Li, X., Huang, R., Luo, A., Cheng, H., Fan, D.P.: Uncertainty-guided transformer reasoning for camouflaged object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4146-4155 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "type": "text", + "content": "40. Yu, S., Zhang, B., Xiao, J., Lim, E.G.: Structure-consistent weakly supervised salient object detection with local saliency coherence. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 3234-3242 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "type": "text", + "content": "41. Zhai, Q., Li, X., Yang, F., Chen, C., Cheng, H., Fan, D.P.: Mutual graph learning for camouflaged object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12997-13007 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 402, + 482, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 446 + ], + "type": "text", + "content": "42. Zhang, J., Fan, D.P., Dai, Y., Anwar, S., Saleh, F.S., Zhang, T., Barnes, N.: Ucnet: Uncertainty inspired rgb-d saliency detection via conditional variational autoencoders. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8582-8591 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "text", + "content": "43. Zhang, J., Yu, X., Li, A., Song, P., Liu, B., Dai, Y.: Weakly-supervised salient object detection via scribble annotations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12546-12555 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 479, + 482, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 512 + ], + "type": "text", + "content": "44. Zhang, J., Zhang, T., Dai, Y., Harandi, M., Hartley, R.: Deep unsupervised saliency detection: A multiple noisy labeling perspective. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9029-9038 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 512, + 482, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 482, + 545 + ], + "type": "text", + "content": "45. Zhou, H., Xie, X., Lai, J.H., Chen, Z., Yang, L.: Interactive two-stream decoder for accurate and fast saliency detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9141-9150 (2020)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 198, + 91, + 447, + 102 + ], + "type": "text", + "content": "Just a Hint: Point-Supervised Camouflaged Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/KDProR_ A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval/8778d22d-a6f9-4fed-bc5f-fddacba7cf0b_content_list.json b/2024/KDProR_ A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval/8778d22d-a6f9-4fed-bc5f-fddacba7cf0b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..038a509f191833ea0290e90db54f7063e99959f5 --- /dev/null +++ b/2024/KDProR_ A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval/8778d22d-a6f9-4fed-bc5f-fddacba7cf0b_content_list.json @@ -0,0 +1,1880 @@ +[ + { + "type": "text", + "text": "KDProR: A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval", + "text_level": 1, + "bbox": [ + 222, + 140, + 779, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xianwei Zhuang*, Hongxiang Li*, Xuxin Cheng, Zhihong Zhu, Yuxin Xie, Yuexian Zou†", + "bbox": [ + 326, + 212, + 673, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ADSPLAB, School of ECE, Peking University, China {xwzhuang, lihongxiang}@stu.pku.edu.cn, zouyx@pku.edu.cn", + "bbox": [ + 282, + 255, + 717, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Existing video-text retrieval methods predominantly focus on designing diverse cross-modal interaction mechanisms between captions and videos. However, those approaches diverge from human learning paradigms, where humans possess the capability to seek and associate knowledge from an open set, rather than rote memorizing all text-video instances. Motivated by this, we attempt to decouple knowledge from retrieval models through multi-grained knowledge stores and identify two significant benefits of our knowledge-decoupling strategy: (1) it ensures a harmonious balance between knowledge memorization and retrieval optimization, thereby improving retrieval performance; and (2) it can promote incorporating diverse open-world knowledge to augment video-text retrieval. To efficiently integrate information from knowledge stores, we further introduce a novel retrieval framework termed KDProR, which utilizes our proposed Expectation-Knowledge-Maximization (EKM) algorithm for optimization. Specifically, in E-step, KDProR obtains relevant contextual semantics from knowledge stores and achieves efficient knowledge injection through interpolation and alignment correction. During the K-step, KDProR calculates the knowledge KNN distribution by indexing the top-K acquired knowledge to refine the retrieval distribution, and in M-step, KDProR optimizes the retrieval model by maximizing the likelihood of the objective. Extensive experiments on various benchmarks prove that KDProR significantly outperforms previous state-of-the-art methods across all metrics. Remarkably, KDProR can uniformly and efficiently incorporate diverse open-world knowledge and is compatible with different interaction mechanisms and architectures.", + "bbox": [ + 261, + 321, + 738, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Video Retrieval $\\cdot$ Knowledge Decoupling $\\cdot$ Open World", + "bbox": [ + 261, + 681, + 715, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 722, + 375, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video-text retrieval (VTR) is a significant and challenging task in cross-modal interaction [1,12,31,42], aiming to align video (text) candidates with text (video) queries to identify the most relevant instances accurately [61]. The standard", + "bbox": [ + 212, + 753, + 787, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* These authors contributed equally.", + "bbox": [ + 232, + 810, + 477, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 232, + 824, + 395, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a382c3ae873f7e8bea857310715212eaf1e2152084b31d3febae47a8e86298ca.jpg", + "image_caption": [ + "Fig. 1: The illustration of different VTR paradigms. (a) denotes vanilla VTR methods; (b) introduces additional data only during training. (c) achieves \"open examination\" with knowledge decoupling, thereby enhancing VTR in both closed and open sets." + ], + "image_footnote": [], + "bbox": [ + 240, + 143, + 763, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "paradigm of VTR for addressing cross-modal matching involves aligning videotext features acquired through visual and textual encoders via contrastive learning. With the advent of large-scale image-language pre-trained models [20, 44, 60, 62], recent VTR methods [3, 13, 35] have realized substantial advancements in retrieval performance by leveraging pre-trained image-language models, such as CLIP [44]. Benefiting from powerful CLIP, most of these methods focus on devising various cross-modal interaction strategies to enhance visual-text alignment, including cross-attention [13], multi-grained interaction [36], disentangled representation learning [48], etc, and have achieved considerable success.", + "bbox": [ + 212, + 369, + 787, + 505 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, if we conceptualize the training set as a \"book\" intuitively, this training-test paradigm for VTR can be likened to \"video-text memorization\" and \"closed-book VTR examination\", which diverges from the human learning process [37, 65]. Humans possess the capability to seek and associate knowledge from an open set, rather than relying on rote memorization of all text-video instances [47]. This \"closed-book examination\" approach may be further hindered by dense memorization overhead [37] and challenges in memorizing difficult samples [9, 11], echoing the proverb: \"The palest ink is better than the best memory.\" Furthermore, previous open-set fine-tuning [18, 52] and pre-training [27, 59] methods have shown that incorporating auxiliary data from the open world can improve VTR performance. These insights encourage us to explore a novel VTR paradigm: \"open-set examination\", which involves constructing knowledge stores from the closed training set or open-world sources to decouple knowledge. As shown in Figure 1, this strategy introduces relevant knowledge from stores as reinforcement signals to help the model strike a balance between generalization and memory. However, this also brings two new challenges: (1) How to decouple knowledge from models and unify closed- and open-world knowledge; (2) How to efficiently incorporate decoupled knowledge to improve retrieval performance.", + "bbox": [ + 212, + 506, + 787, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a novel knowledge-decoupling probabilistic framework for VTR (KDProR), tackling the challenges with the following strategies:", + "bbox": [ + 214, + 779, + 785, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(1) KDProR utilizes multi-grained stores to decouple knowledge and standardize closed- and open-world knowledge sources. We con", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "X. Zhuang et al.", + "bbox": [ + 271, + 114, + 382, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "struct multi-scale knowledge stores composed of local and global neural representations to decouple knowledge. This approach allows rare patterns to be memorized explicitly, rather than implicitly in model parameters. In addition, knowledge stores can be updated asynchronously and endowed with knowledge from closed- or open-world sources. In this work, we explore three different sources of additional knowledge, including spatio-temporal knowledge from pre-trained models, entity-relation structured knowledge and outputs of large language models (LLMs, e.g., GPT [40]) prompted by chain-of-thoughts (CoT) [50]. Moreover, we identify two core advantages of the knowledge decoupling strategy: it (a) ensures a harmonious balance between memorization and retrieval optimization; and (b) opens up a unified interface for injecting diverse open-world knowledge.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(2) KDProR utilizes a novel Expectation-Knowledge-Maximization algorithm to achieve efficient knowledge injection and retrieval optimization. We extend the traditional Expectation-Maximization (EM) algorithm [5] to the Expectation-Knowledge-Maximization (EKM) algorithm for optimizing KDProR. Specifically, in the E-step, KDProR obtains relevant contextual semantics and achieves efficient knowledge injection through interpolation and alignment correction. In the K-step, we calculate the k-nearest neighbor (KNN) distribution by indexing the top-K knowledge obtained to calibrate the original retrieval distribution. And in the M-step, we optimize the model by maximizing the likelihood of the target. Additionally, we provide a theoretical analysis of our KDProR framework from a probabilistic perspective in Section 3.7.", + "bbox": [ + 212, + 313, + 787, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extensive experiments on four benchmarks, i.e., MSR-VTT [55], DiDeMo [16], LSMDC [45] and ActivityNet [15], prove that KDProR significantly outperforms previous state-of-the-art methods in both closed and open-world settings. In summary, our KDProR has several appealing facets: (1) Effectiveness: It can significantly improve retrieval performance under both closed-set and open-set settings. (2) Universality: It can be used to introduce various open-world knowledge, while being compatible with various interaction modules and pretrained foundation models (e.g., [27, 59]). (3) Explainability: The effectiveness and convergence of retrieval optimization via our EKM algorithm are theoretically guaranteed. The main contributions of this paper are presented as follows:", + "bbox": [ + 212, + 479, + 787, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(1) We explore a novel knowledge-decoupling paradigm for VTR and construct multi-grained knowledge stores to uniformly introduce knowledge from closed sets and diverse open-world sources. (2) We propose a novel VTR framework termed KDProR, which utilizes a principled EKM algorithm to inject various knowledge into the model to improve retrieval performance. Theoretical analysis provides an underlying insight into the effectiveness and convergence of KDProR. (3) Experiments conducted on four benchmarks show that KDProR achieves new state-of-the-art results under both closed- and open-world settings.", + "bbox": [ + 212, + 630, + 787, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 776, + 382, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature Representation for Video-Text Retrieval. Recent studies [1, 12, 31, 42] on VTR employ a bi-encoder architecture to extract text and video rep", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval", + "bbox": [ + 245, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "resentation. The selection of text and visual encoders changes with the development of research on feature extraction upon NLP and CV, e.g., Word2Vec [38] and BERT-based models [7, 32] for textual representation, and CNN-based [14] and ViT-based models [8, 63, 64] for visual tasks. These researches use independently pre-trained dual encoders for visual-textual semantic extraction [1, 12, 42, 57]. Recently, CLIP [43] pre-trained on a large-scale dataset with 400 million pairs has achieved excellent capability of cross-modal representation. And CLIP-based VTR methods [3, 13, 35, 53] have rapidly developed and achieved great success. Our KDProR also benefits from existing pre-training, while exploring a novel perspective for enhancing representation, i.e., knowledge decoupling.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Interaction Mechanism for Video-Text Retrieval. Many VTR methods explore different cross-modal interaction mechanisms for cosine similarity [35] to improve retrieval performance. Some studies [35] seek to enhance the spatiotemporal understanding ability of image encoders through LSTM [17] or temporal Transformers. MoE-based [3] and Cross-Transformers-based [13,54] methods significantly improve the cross-modal interaction. [25, 26, 29, 36, 49, 56] explore hierarchical and multi-grained interactions, respectively. DRL [48] proposes disentangled representation learning for cross-modal interaction. Our KDProR is compatible with these cross-modal interaction strategies.", + "bbox": [ + 212, + 297, + 787, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Open-World Knowledge for Video-Text Retrieval. The previous methods mainly introduce open-world knowledge in the pre-training and fine-tuning stages to enhance VTR performance. CLIP-ViP [59] conducts secondary pretraining of CLIP on large-scale video-text datasets, and UMT [27] achieves efficient pre-training to improve the spatio-temporal understanding and inject open-set knowledge into VTR tasks. TEFAL [18] utilizes additional audio features to enhance VTR retrieval performance and Cap4Video [52] utilizes auxiliary captions obtained by Web Search or generated by LLMs to enhance cross-modal interaction. KDProR is compatible with these pre-training methods, and we do not specify the type of open-world knowledge to be introduced, but instead achieve universal open-set knowledge injection through knowledge decoupling.", + "bbox": [ + 212, + 433, + 787, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methods", + "text_level": 1, + "bbox": [ + 215, + 619, + 339, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Settings and Feature Extraction.", + "text_level": 1, + "bbox": [ + 215, + 650, + 532, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following previous work [35, 48], we utilize the dual-encoder architecture of CLIP [44] to achieve video and text feature extraction on training set $\\mathcal{D} = \\{x_i\\}_i^N = \\{(T_i,V_i)\\}_i^N$ , where $T_{i}$ and $V_{i}$ represent text and video instances respectively, and $N$ is the size of the training set. Given a text query $T_{i}$ , we initially add identifiers [CLS] and [SEP] to the sentence, and leverage the text encoder of CLIP to encode the text representation: $t_i = \\mathrm{TextEncoder}(T_i)$ . For each video $V_{i}$ , we uniformly select $n_v$ frames $[F_1,\\dots ,F_{n_v}]$ as keyframes, and employ the transformer-based encoder of CLIP to extract sequential features $v_{i} = \\mathrm{VideoEncoder}([F_{1},\\dots ,F_{n_{v}}])$ .", + "bbox": [ + 212, + 672, + 787, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Extensive work explores various cross-modal interaction strategies to improve retrieval performance [30,35,48]. Our work focuses on decoupling knowledge from", + "bbox": [ + 212, + 810, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "X. Zhuang et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/906932314491d8f437564da914c9aa9f23a0d7e03a0b1830842f4c20535591a2.jpg", + "image_caption": [ + "Fig. 2: The illustration of the proposed framework, consisting of (1) constructing a multi-grained knowledge store using open-world knowledge or closed-world knowledge (training dataset); (2) EKM algorithm that sequentially implements knowledge injection, distribution calibration, and target likelihood maximization." + ], + "image_footnote": [], + "bbox": [ + 238, + 143, + 767, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "models and exploring the impact of different knowledge sources on retrieval performance, rather than pre-training or interaction strategies. Thus we adopt the weighted token-wise interaction module of DRL [48] to achieve cross-modal interaction. Note that KDProR is also compatible with other interaction strategies.", + "bbox": [ + 214, + 425, + 782, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Decoupling Knowledge Store", + "text_level": 1, + "bbox": [ + 215, + 503, + 500, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our KDProR involves augmenting CLIP-based retrieval models with semantic information from decoupled knowledge stores. Thus, the first step of our methods is to build a multi-grained knowledge store. As illustrated in Figure 2, our stores contain fine-grained and coarse-grained caches, where each key and value (text-video) are representations of knowledge.", + "bbox": [ + 212, + 526, + 782, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fine-grained Knowledge Store. The fine-grained knowledge store can be represented as $\\mathcal{S}^f = \\{\\mathcal{S}_i^f\\}_{i=1}^{K^f} = \\{(\\mathcal{K}_i^f, \\mathcal{V}_i^f)\\}_{i=1}^{K^f}$ , where $\\mathcal{S}_i^f = (\\mathcal{K}_i^f, \\mathcal{V}_i^f)$ denotes the $i$ -th key-value pair in the fine-grained knowledge store, and $\\mathcal{K}_i^f$ is the size of our fine-grained knowledge store. Under the closed world setting, we utilize text and visual encoders of CLIP to obtain the feature key-value pairs on the training dataset as $\\mathcal{K}_i^f = t_i$ and $\\mathcal{V}_i^f = v_i$ .", + "bbox": [ + 214, + 602, + 784, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Coarse-grained Knowledge Store. The coarse-grained knowledge store is designed to encapsulate global semantics. It equips retrieval models with global, coarse-grained insights from the knowledge store, enhancing the comprehensive knowledge understanding. Similarly, the coarse-grained storage can be represented as: $\\mathcal{S}^c = \\{\\mathcal{S}_i^c\\}_{i=1}^{K^c} = \\{(\\mathcal{K}_i^c, \\mathcal{V}_i^c)\\}_{i=1}^{K^c}$ , where $\\mathcal{S}_i^c = (\\mathcal{K}_i^c, \\mathcal{V}_i^c)$ denotes the $i$ -th key-value pair in the coarse-grained knowledge store, and $\\mathcal{K}_i^c$ is the size of our coarse-grained knowledge store. We adopt the K-mean strategy to extract knowledge from fine-grained storage into more advanced representations:", + "bbox": [ + 214, + 696, + 787, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {K} _ {i} ^ {c} = \\phi \\left(c _ {i} ^ {t}\\right), \\forall c _ {i} ^ {t} \\in \\mathcal {C} ^ {t}, \\quad \\mathcal {V} _ {i} ^ {c} = \\phi \\left(c _ {i} ^ {v}\\right), \\forall c _ {i} ^ {v} \\in \\mathcal {C} ^ {v}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 325, + 823, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval", + "bbox": [ + 245, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where, $\\mathcal{C}^t = \\{c_1^t,\\dots ,c_{K^c}^t\\}$ denotes the cluster set obtained by K-Mean over fine-grained knowledge $\\{\\mathcal{K}_i^f\\}$ , $\\mathcal{C}^v = \\{c_1^v,\\dots ,c_{K^c}^v\\}$ represents the cluster set obtained by K-Mean over $\\{\\mathcal{V}_i^f\\}$ , and $\\phi (\\cdot)$ represents the Max-Pooling aggregate function.", + "bbox": [ + 212, + 145, + 782, + 195 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that our framework is universal, and we will introduce more applications for constructing diverse knowledge stores in Section 3.6.", + "bbox": [ + 212, + 196, + 784, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 E-step for Multi-Grained Knowledge Injection", + "text_level": 1, + "bbox": [ + 214, + 247, + 643, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Knowledge Interpolation. In E-step, KDProR initially attempts to inject fine-grained and coarse-grained knowledge into the model by knowledge interpolation of features. We treat with $t_i$ and $v_i$ as anchors to retrieve the fine-grained knowledge store, and obtain two sets of top-K similar vector pairs $\\mathcal{N}_{t_i}^f$ and $\\mathcal{N}_{v_i}^f$ , respectively. $\\mathcal{N}_{t_i}^f = \\{(\\mathcal{K}_{t_i}^f, \\mathcal{V}_{t_i}^f)\\}$ and $\\mathcal{N}_{v_i}^f = \\{(\\mathcal{K}_{v_i}^f, \\mathcal{V}_{v_i}^f)\\}$ are two sets of key-value knowledge pairs, where $\\mathcal{K}_{t_i}^f$ and $\\mathcal{V}_{v_i}^f$ are the nearest neighbors of $t_i$ and $v_i$ , respectively. After the model retrieves the top-K candidates for each $t_i$ and $v_i$ , their corresponding representation $\\mathcal{K}_{t_i}^f$ and $\\mathcal{V}_{v_i}^f$ obtained from knowledge-store will be incorporated into original features $t_i$ and $v_i$ to act as demonstration learning. Specifically, we adopt the nearest neighbor sets to inject decoupled knowledge into text and video features as:", + "bbox": [ + 212, + 272, + 787, + 443 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nt _ {i} ^ {f} = \\rho t _ {i} + (1 - \\rho) \\frac {\\sum_ {(\\mathcal {K} , \\mathcal {V}) \\in \\mathcal {N} _ {t _ {i}} ^ {f}} \\mathcal {K}}{| \\mathcal {N} _ {t _ {i}} ^ {f} |}, v _ {i} ^ {f} = \\rho v _ {i} + (1 - \\rho) \\frac {\\sum_ {(\\mathcal {K} , \\mathcal {V}) \\in \\mathcal {N} _ {v _ {i}} ^ {f}} \\mathcal {V}}{| \\mathcal {N} _ {v _ {i}} ^ {f} |}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 250, + 452, + 785, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where, $\\rho$ is a hyperparameter used to express the interpolation ratio. Subsequently, we utilize the weighted token-wise interaction module [48] to enhance textual and visual representations injected with additional knowledge.", + "bbox": [ + 212, + 502, + 782, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, the textual and visual knowledge within the top-K sets, obtained by $t_i$ and $v_i$ respectively, may not correspond on a one-to-one basis due to their derivation from two independent top-K operations. To mitigate the issue of mismatched video-text features during knowledge injection, we utilize the optimal transport (OT) algorithm to facilitate alignment correction within a batch.", + "bbox": [ + 212, + 547, + 784, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Alignment Correction via Optimal Transport. We first define $\\langle \\mathbf{Q},\\mathbf{S}^f\\rangle =$ $\\mathrm{tr}(\\mathbf{Q}^\\top \\mathbf{S}^f)\\in \\mathbb{R}^{B\\times B}$ as the similarity matrix calculated from the updated text and video pairs $\\{(t_i^f,v_i^f)\\}_{i = 1}^B$ in a batch. $\\mathbf{Q}\\in \\mathbb{R}^{B\\times B}$ represents the corresponding transport assignment matrix, where $\\mathbf{Q}[i,j]$ denotes the probabilities of aligning $t_i^f$ with $v_{i}^{f}$ . The optimal transport aims to establish flexible alignment between videos and captions by maximizing global text-video similarity $\\mathrm{tr}(\\mathbf{Q}^\\top \\mathbf{S}^f)$ . The objective of optimal transport can be formulated as:", + "bbox": [ + 214, + 623, + 787, + 733 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\mathbf {Q} \\in \\mathcal {Q}} \\langle \\mathbf {Q}, \\mathbf {S} ^ {f} \\rangle + \\varepsilon H (\\mathbf {Q}) \\quad \\text {s . t .} \\mathcal {Q} = \\left\\{\\mathbf {Q} \\in \\mathbb {R} ^ {B \\times B} \\mid \\mathbf {Q} \\mathbf {1} _ {B} = \\boldsymbol {\\mu}, \\mathbf {Q} ^ {\\top} \\mathbf {1} _ {B} = \\boldsymbol {\\nu} \\right\\} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 240, + 742, + 784, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where, $\\pmb{\\mu} \\in \\mathbb{R}^{B}$ and $\\pmb{\\nu} \\in \\mathbb{R}^{B}$ indicate the relative importance of each video and caption, $\\mathbf{1}_{B}$ represents the vector of ones in dimension $B$ , $H(\\mathbf{Q})$ denotes a entropy regular term [4] and $\\varepsilon$ controls its smoothness. Following [46], we initialize $\\pmb{\\mu}$ and $\\pmb{\\nu}$ to a uniform distribution $\\mathbf{1}_{B}$ . The optimal transport problem", + "bbox": [ + 212, + 777, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "X. Zhuang et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in Eq. 3 can be solved using the Sinkhorn-Knopp algorithm [46] to obtain the optimal assignment scheme $\\mathbf{Q}^*$ . Based on this, we can obtain the similarity matrix after adding alignment correction $\\mathbf{Q}^*$ as follows:", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} ^ {f ^ {*}} = \\left(\\left(1 - \\beta\\right) \\mathbf {I} + \\beta \\mathbf {Q} ^ {*}\\right) \\mathbf {S} ^ {f}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 198, + 785, + 215 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where, $\\mathbf{I}$ is the identity matrix and $\\beta$ is a hyperparameter.", + "bbox": [ + 214, + 223, + 635, + 238 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We can implement the coarse-grained knowledge injection and obtain the realigned similarity matrix $\\mathbf{S}^{c*}$ in a similar way.", + "bbox": [ + 212, + 238, + 782, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 K-step for Retrieval Distribution Calibration", + "text_level": 1, + "bbox": [ + 214, + 287, + 633, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Existing studies [23, 41] have shown that non-parametric KNN classification or nearest-neighbor interpolation methods have extraordinary potential in improving robustness and generalization. It is intuitively to leverage the KNN's retrieval results as the prior knowledge to guide the parameters training for video-text pairs that are difficult to establish deep correlations. In practice, we aggregate the prediction of KNN into the original probability distribution of video-text retrieval to achieve the multi-grained calibration of retrieval distribution.", + "bbox": [ + 212, + 309, + 784, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fine-grained Distribution Calibration. We utilize the set of k-nearest neighbors $\\mathcal{N}_{t_i}^f$ and $\\mathcal{N}_{v_i}^f$ retrieved by querying the open-book knowledge store in E-step as fine-grained top-K knowledge. Subsequently, we calculate the KNN distribution according to the softmax of the similarity and aggregation probability mass of each video-text pair in the retrieved target as:", + "bbox": [ + 212, + 415, + 784, + 491 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np _ {k n n} ^ {f} \\left(v _ {i} \\mid t _ {i}\\right) = \\frac {\\sum_ {\\left(\\mathcal {K} , \\mathcal {V}\\right) \\in N _ {t _ {i}} ^ {f}} \\varphi \\left(v _ {i} , \\mathcal {V}\\right) e ^ {d \\left(v _ {i} , \\mathcal {V}\\right)}}{\\sum_ {\\left(\\mathcal {K} , \\mathcal {V}\\right) \\in N _ {t _ {i}} ^ {f}} \\varphi \\left(v _ {i} , \\mathcal {V}\\right)}, \\quad \\varphi \\left(v _ {i}, \\mathcal {V}\\right) = \\left\\{ \\begin{array}{l l} 1, & \\mathcal {V} \\in N _ {v _ {i}} ^ {f} \\\\ 0, & \\mathcal {V} \\notin N _ {v _ {i}} ^ {f} \\end{array} , \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 496, + 784, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where, $d(\\cdot ,\\cdot)\\in [-1,1]$ denotes the cosine similarity function, and $\\varphi (v_{i},\\mathcal{V})$ is an indicator function used to determine whether $\\mathcal{V}$ is within the top-K nearest neighbor set $N_{v_i}^f$ of the $v_{i}$ .", + "bbox": [ + 212, + 547, + 782, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We can further obtain the KNN probability distribution $p_{knn}^{f}(t_{i} \\mid v_{i})$ for video-to-text retrieval in a completely symmetrical manner.", + "bbox": [ + 212, + 594, + 782, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Coarse-grained Distribution Calibration. Consistent with Eq. 5, we employ the KNN algorithm to obtain top-K knowledge, utilizing closed-world or open-world knowledge from the coarse-grained knowledge store to refine the probability distribution function pertinent to the text-to-video retrieval. Nevertheless, since the coarse-grained knowledge store is derived from the distillation of fine-grained knowledge, it is impracticable to apply indicator functions (cf. Eq. 5) to ascertain the presence of an inclusion relation between textual features and their nearest neighbor sets. To tackle this challenge, we introduce a novel indicator function, denoted as $\\phi(\\cdot, \\cdot)$ , parameterized by a threshold $\\alpha$ . $\\phi(\\cdot, \\cdot)$ is designed to evaluate whether $\\mathcal{V}$ resides within the top-K nearest neighbor set $N_{v_i}^c$ of the $v_i$ via the cosine similarity function $d(\\cdot, \\cdot)$ , and is formulated as:", + "bbox": [ + 212, + 626, + 784, + 792 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\phi \\left(v _ {i}, \\mathcal {V}\\right) = \\left\\{ \\begin{array}{l l} 1, & d \\left(\\mathcal {V}, \\mathcal {V} ^ {c}\\right) < \\alpha , \\exists \\left(\\mathcal {K} ^ {c}, \\mathcal {V} ^ {c}\\right) \\in N _ {v _ {i}} ^ {c} \\\\ 0, & o t h e r w i s e s \\end{array} . \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 804, + 784, + 844 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval", + "bbox": [ + 245, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on this, we can obtain a closed-form expression for the KNN distribution used for coarse-grained distribution calibration as follows:", + "bbox": [ + 215, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\np _ {k n n} ^ {c} \\left(v _ {i} \\mid t _ {i}\\right) = \\frac {\\sum_ {\\left(\\mathcal {K} , \\mathcal {V}\\right) \\in N _ {t _ {i}} ^ {c}} \\phi \\left(v _ {i} , \\mathcal {V}\\right) e ^ {d \\left(v _ {i} , \\mathcal {V}\\right)}}{\\sum_ {\\left(\\mathcal {K} , \\mathcal {V}\\right) \\in N _ {t _ {i}} ^ {c}} \\phi \\left(v _ {i} , \\mathcal {V}\\right)}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 185, + 785, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We can further obtain the KNN probability distribution $p_{knn}^{c}(t_{i} \\mid v_{i})$ for coarse-grained video-to-text retrieval in a completely symmetrical manner.", + "bbox": [ + 215, + 236, + 782, + 266 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 M-step for Unified Retrieval Optimization", + "text_level": 1, + "bbox": [ + 215, + 286, + 609, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the M-step, we optimize the whole model by maximizing the retrieval likelihood objective over the realigned similarity matrix. Specifically, we utilize the re-aligned multi-grained similarity matrix in Eq. 4 obtained from E-step to calculate the InfoNCE loss:", + "bbox": [ + 212, + 309, + 782, + 368 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t 2 v} = - \\frac {1}{B} \\sum_ {i} ^ {B} \\left(\\log \\frac {\\exp \\left(\\mathbf {S} ^ {f _ {i i} ^ {*}}\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\mathbf {S} ^ {f _ {i j} ^ {*}}\\right)} + \\log \\frac {\\exp \\left(\\mathbf {S} _ {i i} ^ {c ^ {*}}\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\mathbf {S} _ {i j} ^ {c ^ {*}}\\right)}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 274, + 376, + 784, + 433 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v 2 t} = - \\frac {1}{B} \\sum_ {i} ^ {B} \\left(\\log \\frac {\\exp \\left(\\mathbf {S} _ {i i} ^ {f ^ {*}}\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\mathbf {S} _ {j i} ^ {f ^ {*}}\\right)} + \\log \\frac {\\exp \\left(\\mathbf {S} _ {i i} ^ {c ^ {*}}\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\mathbf {S} _ {j i} ^ {c ^ {*}}\\right)}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 274, + 429, + 728, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $B$ is the batch size. Subsequently, we propose utilizing the KNN distribution (cf. Eq. 5 and 7) to guide the training process. The KNN calibrator reweights the InfoNCE loss by adjusting the relative loss for the correctly-matched or mismatched video-text instances identified by KNN, respectively. We focus on exploiting the results of KNN distribution for calibrating the training of retrieval models. Specifically, we apply the negative log-likelihood of $p_{knn}^{f}$ in Eq 5 and $p_{knn}^{c}$ in Eq. 7 as the calibration factor:", + "bbox": [ + 212, + 483, + 784, + 589 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nF _ {t 2 v} ^ {f} = - \\log p _ {k n n} ^ {f} \\left(v _ {i} \\mid t _ {i}\\right), \\quad F _ {v 2 t} ^ {f} = - \\log p _ {k n n} ^ {f} \\left(t _ {i} \\mid v _ {i}\\right); \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 601, + 784, + 626 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nF _ {t 2 v} ^ {c} = - \\log p _ {k n n} ^ {c} \\left(v _ {i} \\mid t _ {i}\\right), \\quad F _ {v 2 t} ^ {c} = - \\log p _ {k n n} ^ {c} \\left(t _ {i} \\mid v _ {i}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 623, + 694, + 637 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we can optimize the retrieval model by maximizing the following knowledge-guided cross-modal contrastive loss as:", + "bbox": [ + 212, + 642, + 782, + 671 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v 2 t} ^ {*} = \\left(1 + \\lambda_ {1} F _ {v 2 t} ^ {f} + \\lambda_ {2} F _ {v 2 t} ^ {c}\\right) \\mathcal {L} _ {v 2 t}; \\quad \\mathcal {L} _ {t 2 v} ^ {*} = \\left(1 + \\lambda_ {1} F _ {t 2 v} ^ {f} + \\lambda_ {2} F _ {t 2 v} ^ {c}\\right) \\mathcal {L} _ {t 2 v}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 680, + 784, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ are hyper-parameters to determine the proportion of each loss term. Hence, in the M-step, the complete training objective of KDProR can be formulated as $\\mathcal{L} = \\mathcal{L}_{v2t}^{*} + \\mathcal{L}_{t2v}^{*}$ .", + "bbox": [ + 212, + 705, + 782, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.6 Applications of Our KDProR Framework", + "text_level": 1, + "bbox": [ + 215, + 771, + 601, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, as illustrated in Figure 3, we present three distinct applications for constructing open-world knowledge stores, including pre-trained knowledge, entity-relation structured knowledge, and CoT knowledge within LLMs.", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "X. Zhuang et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c4524fa7d5fdadbc17b735a712b83c4ad55e4d985731291e859026e15c0e9f2c.jpg", + "image_caption": [ + "Fig. 3: The illustration of proposed open-world knowledge stores, including (1) video-text pretraining knowledge, (2) entity-relation structure knowledge, and (3) knowledge within LLMs prompted by the CoT technique." + ], + "image_footnote": [], + "bbox": [ + 254, + 143, + 743, + 266 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Video-Text Pretraining. As shown in Figure 3, we utilize CLIP-ViP [59], which is further pre-trained on HD-VILA-100M [58] dataset, to extract pretraining video-text knowledge for constructing our knowledge stores.", + "bbox": [ + 212, + 345, + 782, + 391 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Entity-Relation Structure. We attempt to parse captions to scene graphs $\\mathcal{G}$ through the Scene Graph Parser [51], then employ $\\mathcal{G}$ to obtain subject-relation-object triplets. Subsequently, we use these triplets as substitutes for the original captions and then employ CLIP to extract text-video features to construct stores.", + "bbox": [ + 212, + 392, + 784, + 450 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "CoT Generation of LLMs. In this work, we attempt to use a strategy similar to CoT [50] to generate auxiliary captions for each video frame by prompting LLMs. Subsequently, we utilize CLIP to extract neural representations of auxiliary captions and corresponding video frames to construct knowledge stores.", + "bbox": [ + 212, + 452, + 784, + 512 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.7 Theoretical Analysis", + "text_level": 1, + "bbox": [ + 214, + 547, + 429, + 561 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "From a probabilistic perspective, the optimization goal of the VTR model $\\Theta$ trained on dataset $\\mathcal{D} = \\{x_{i}\\}_{i}^{N} = \\{(T_{i},V_{i})\\}_{i}^{N}$ is to maximize the log-likelihood $\\mathcal{L}(\\Theta)$ for text-video pairs. We consider the knowledge introduced from fine-grained and coarse-grained stores as two related latent variables $z_{i}^{f}$ and $z_{i}^{c}$ . In a closed world, the posterior distribution of $z_{i}^{f}$ and $z_{i}^{c}$ can be represented as:", + "bbox": [ + 214, + 569, + 784, + 648 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nQ \\left(z _ {i} ^ {f}\\right) = p \\left(z _ {i} ^ {f}; x _ {i}, \\Theta\\right), \\quad Q \\left(z _ {i} ^ {c}\\right) = p \\left(z _ {i} ^ {c}; z _ {i} ^ {f}, x _ {i}, \\Theta\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 655, + 784, + 676 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Thus we can rewrite the objective function of the retrieval model $\\Theta$ according to Jensen's inequality [19] as follows:", + "bbox": [ + 214, + 681, + 782, + 712 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta^ {*} = \\arg \\max \\sum_ {i} \\sum_ {z _ {i} ^ {f}, z _ {i} ^ {c}} Q \\left(z _ {i} ^ {f}\\right) Q \\left(z _ {i} ^ {c}\\right) \\log p \\left(x _ {i}, z _ {i} ^ {f}, z _ {i} ^ {c}; \\Theta\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 302, + 719, + 784, + 758 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To sum up, the E-step and K-step aim to estimate the posterior probability $Q(z_{i}^{f})$ and $Q(z_{i}^{c})$ on fine-grained and coarse-grained knowledge stores with given current $\\Theta^{(t)}$ . Then KDProR achieves efficient alignment of videos and texts and obtains $\\Theta^{(t + 1)}$ under knowledge guidance at M-step by maximizing the log-likelihood in Eq. 12 with known $Q(z_{i}^{f})$ and $Q(z_{i}^{c})$ , and satisfies the Theorem 1:", + "bbox": [ + 214, + 763, + 784, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval", + "bbox": [ + 245, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/38bea03cc3f70225f9c77c8359f18cf76a214e99e73cd644e03f386d8a3c35c4.jpg", + "table_caption": [ + "Table 1: Results on MSR-VTT without any post-processing (e.g., [3] and [2]). ${}^{ + }$ and ${}^{ \\top }$ are closed-set settings and denote constructing knowledge stores on the training set using the original CLIP and the fine-tuned CLIP upon the training set, respectively. *, † and ‡ denote pre-trained, structural and CoT knowledge in Section 3.6, respectively." + ], + "table_footnote": [], + "table_body": "
BackboneMethodText-to-Video RetrievalVideo-to-Text Retrieval
R@1↑R@5↑R@10↑MdR↓MnR↓R@1↑R@5↑R@10↑MdR↓MnR↓
Non-CLIPCE [31]20.948.862.46.028.220.650.364.05.325.1
MMT [12]26.657.169.64.024.027.057.569.73.721.3
Support [42]27.456.367.73.0-26.655.167.53.0-
Frozen [1]31.059.670.53.0------
CLIP ViT-B/32CLIP4clip [35]44.571.481.62.015.342.770.980.62.011.6
X-pool [13]46.972.882.22.014.3-----
EMCL [21]46.873.183.12.0-46.573.583.52.0-
TS2-Net [33]47.074.583.82.013.045.374.183.72.09.2
DRL [48]47.474.683.82.0-45.373.983.32.0-
STAN [28]46.972.882.82.0------
UATVR [10]47.573.983.52.012.946.973.883.82.08.6
PromptSwitch [6]47.873.982.2-14.146.074.384.8-8.5
KDProR+48.474.484.22.012.247.174.684.42.08.9
KDProR+48.774.484.12.012.047.274.684.42.08.7
KDProR*49.274.284.42.012.047.174.084.12.08.9
KDProR†49.074.684.32.012.047.374.383.92.08.7
KDProR‡49.675.184.42.011.648.274.284.22.08.1
", + "bbox": [ + 220, + 212, + 787, + 452 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Theorem 1 The retrieval model $\\Theta$ optimized iteratively through the proposed EKM algorithm of step $t$ satisfies the following properties:", + "bbox": [ + 215, + 481, + 789, + 513 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Monotonic Increasing: $\\forall t \\geq 0$ , $\\mathcal{L}(\\Theta^{(t + 1)}) \\geq \\mathcal{L}(\\Theta^{(t)})$ ;", + "- Convergence: $\\forall \\epsilon > 0, \\exists \\delta > 0$ , if $t > \\delta$ , $|\\Theta^{(t + 1)} - \\Theta^{(t)}| \\leq \\epsilon$ ." + ], + "bbox": [ + 225, + 518, + 691, + 550 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Theorem 1 indicates that our EKM algorithm can ensure the model is positively optimized in each iteration, while also maintaining the convergence properties of vanilla EM algorithms.", + "bbox": [ + 212, + 566, + 787, + 612 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 633, + 375, + 652 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experiments Settings", + "text_level": 1, + "bbox": [ + 215, + 664, + 439, + 681 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on four benchmarks for video-text retrieval tasks including: MSR-VTT [55], DiDeMo [16], LSMDC [45] and ActivityNet [15]. MSR-VTT [55] comprises 10,000 videos with 20 captions each. We train on 9,000 videos and their captions, testing on 1k-A with 1,000 videotext pairs. DiDeMo [16] provides 10,000 videos annotated with 40,000 sentences. We concatenate all descriptions of a video into one paragraph for evaluation. LSMDC [45] includes 118,081 movie-extracted videos, with our dataset partitioned into 109,673 for training, 7,408 for validation, and 1,000 for testing. ActivityNet [15] includes 20,000 YouTube videos. Following [35], we concatenate all captions similar to DiDeMo and evaluate on the 'val1' split.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "X. Zhuang et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/56dd927188e6b562656f81513048cb985ada27134f7e12432ea4e9be90880e02.jpg", + "table_caption": [ + "Table 2: Text-to-Video retrieval results on DiDeMo, ActivityNet and LSMDC datasets without any post-processing (e.g., [3] and [2]). $^+$ and $\\top$ are closed-set settings and $*$ , $\\dagger$ and $\\ddagger$ are open-world settings, whose meanings are all consistent with those in Table 1." + ], + "table_footnote": [], + "table_body": "
MethodDiDeMoActivityNetLSMDC
R@1↑R@5↑R@10↑MdR↓R@1↑R@5↑R@10↑MdR↓R@1↑R@5↑R@10↑MdR↓
Text-to-Video Retrieval using CLIP (ViT-B/32)
CLIP4clip [35]43.470.280.62.040.572.483.62.022.641.049.111.0
X-CLIP [36]45.274.0--44.374.1--23.343.0--
TS2-Net [33]41.871.682.02.041.073.684.52.023.442.350.99.0
DRL [48]47.973.882.72.044.274.586.12.024.945.755.37.0
PromptSwitch [10]--------23.141.750.5-
UATVR [10]43.171.882.32.0--------
DiffusionRet [22]46.774.782.72.045.875.686.32.024.443.154.38.0
\\(KDProR^{+}\\)48.876.184.12.045.875.485.02.026.145.957.37.0
\\(KDProR^{\\top}\\)49.275.383.62.046.476.285.72.026.547.857.86.0
\\(KDProR^{*}\\)51.778.384.91.048.376.886.12.026.950.258.35.0
\\(KDProR^{\\dagger}\\)50.576.385.11.047.576.185.72.027.650.658.15.0
\\(KDProR^{\\ddagger}\\)53.279.285.01.049.177.986.82.028.250.659.25.0
", + "bbox": [ + 218, + 198, + 787, + 391 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Metrics. Following [35, 48], we utilize recall at rank K (R@K), median rank (MdR) and mean rank (MnR) as metrics to validate the effectiveness of KDProR. Implementation Details. Following [48], we adopt CLIP's ViT-B/32 as the visual encoder and fine-tune the model with 5 epochs on all training sets. We adopt Adam [24] as the optimizer with a cosine warm-up method [34]. Following [35, 48], the frame and caption length are 12 and 32 for MSR-VTT and LSMDC, 64 and 64 for ActivityNet and DiDeMo. The initial learning rate for vision and text encoder is set to $10^{-7}$ and other modules are $10^{-4}$ . The hyperparameter $\\rho$ in Eq. 2 is set to 0.2 on MSR-VTT and LSMDC, 0.1 on ActivityNet and DiDeMo for optimal performance. $\\beta$ in Eq. 4 are set to 0.2, and $\\lambda_{1}$ and $\\lambda_{2}$ in Eq. 9 are set to 0.2 and 0.3, respectively. We use grid search to obtain the optimal parameters. We uniformly set the top-5 relevant knowledge for KNN retrieval, and the size of the coarse-grained knowledge store $K^{c}$ is set to 256. We adopt the gpt-3.5-turbo-1106 [39] to obtain the CoT knowledge in Section 3.6. All experiments are conducted on 2 Tesla A100 GPUs (80G) and 8 RTX 4090 GPUs.", + "bbox": [ + 212, + 421, + 787, + 662 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 215, + 689, + 372, + 704 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparisons to State-of-the-arts. In this section, we compare KDProR with state-of-the-art methods on four datasets, as results shown in Table 1 and 2. Analysis of Table 1 and 2 yields several insights:", + "bbox": [ + 212, + 718, + 785, + 762 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(1) Our KDProR, employing multi-grained knowledge stores constructed from vanilla and fine-tuned CLIP, can achieve performance gains of $1\\%$ and $1.3\\%$ R@1 on MSR-VTT, $1.6\\%$ and $2.2\\%$ R@1 on ActivityNet compared to DRL [48], respectively. This indicates that KDProR can achieve a harmonious balance between memorization and optimization, thereby improving performance.", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval", + "bbox": [ + 245, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/428b2045b346449166cec7427409141bec3d88379882cb66d7db402f9237ec94.jpg", + "table_caption": [ + "Table 3: Comparisons to Other Baseline Methods in Text-to-Video Retrieval. $^\\ddagger$ denotes the utilization of open-world CoT knowledge in Section 3.6." + ], + "table_footnote": [], + "table_body": "
Baseline MethodsKnowledge SourceMSR-VTTLSMDCDiDeMo
R@1↑R@5↑MnR↓R@1↑R@5↑MnR↓R@1↑R@5↑MnR↓
Data AugmentationCaptions from GPT47.374.712.126.446.943.949.776.314.2
Feature ConcatenatePretrained CLIP-ViP47.874.511.926.146.144.550.376.313.4
Cap4Video [52]Captions from GPT49.374.312.0---52.079.410.5
TEFAL [18]Audio Signal49.475.912.026.846.144.4---
Our KDProR‡Captions from GPT49.675.111.628.250.634.953.279.210.2
", + "bbox": [ + 218, + 185, + 787, + 289 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/8c723305f89e86fd41a0d0b2f93510468dda5b7b8fa438c63d8de38ffb56449e.jpg", + "table_caption": [ + "Table 4: Text-to-Video retrieval results on MSR-VTT using our re-implemented CLIP4clip [35] and EMCL [21] as base models. $^+$ and $\\top$ are closed-set settings, *, $^\\dagger$ and $\\ddagger$ are open-world settings, whose meanings are consistent with those in Table 1." + ], + "table_footnote": [], + "table_body": "
MethodCLIP4clip [35] as Base ModelEMCL [21] as Base Model
R@1↑R@5↑R@10↑MdR↓MnR↓R@1↑R@5↑R@10↑MdR↓MnR↓
Base Model43.670.681.02.016.146.673.582.72.013.6
+ KDProR+45.573.482.72.014.147.873.583.12.013.8
+ KDProRT46.474.082.92.013.848.174.083.52.013.7
+ KDProR*47.074.482.92.013.648.773.683.62.013.7
+ KDProR†46.974.082.82.013.648.473.783.82.013.3
+ KDProR‡47.374.683.32.013.449.074.883.82.012.9
", + "bbox": [ + 251, + 359, + 750, + 474 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(2) The integration of additional open-world knowledge significantly bolsters the predictive capabilities of retrieval models. KDProR of employing three open-world knowledge (i.e., pre-trained, structural, and CoT knowledge) in Section 3.6 achieves superior performance, e.g., $+1.8\\%$ , $+1.6\\%$ and $+2.2\\%$ R@1 on MSR-VTT, $3.8\\%$ , $2.6\\%$ and $5.3\\%$ R@1 on DiDeMo compared to DRL [48], respectively. We attribute the performance gain to incorporating open-set knowledge.", + "(3) KDProR demonstrates superior performance compared to all baselines across all benchmarks. This demonstrates that our KDProR offers a twofold advantage: it can unify and incorporate diverse additional knowledge and balance memorization and retrieval to significantly improve performance." + ], + "bbox": [ + 212, + 503, + 785, + 657 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparisons to Other Baseline Methods. We further compare our method with other baselines in Table 3, including Cap4Video [52] and TEFAL [18]. In addition, we compare it with two baseline methods that introduce open-world knowledge, namely using auxiliary captions generated by LLM as data augmentation and directly concatenating the features generated by CLIP-VIP with the original text-video features. The results indicate that introducing open-world knowledge can enhance performance, and our KDProR achieves better performance gains than Data Augmentation and Feature Concatenate. Impressively, KDProR significantly outperforms state-of-the-art open-set baselines, e.g., $+1.4\\%$ R@1 by Cap4Video on LSMDC and $+1.2\\%$ R@1 by TEFAL on DiDeMo.", + "bbox": [ + 212, + 657, + 787, + 810 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Generalization Analysis. We further equip KDProR with two powerful base models in addition to DRL, i.e., CLIP4clip [35] and EMCL [21], and evaluate", + "bbox": [ + 215, + 810, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "X. Zhuang et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/0e4d32936c80c94036c3574b2e0558dc6d8609d32ced0f86853de7e1c3df9539.jpg", + "table_caption": [ + "Table 5: Ablation experiments on MSR-VTT 9k dataset. ${}^{ \\ddagger }$ denotes the utilization of open-world CoT knowledge in Section 3.6." + ], + "table_footnote": [], + "table_body": "
G. MethodText-to-Video RetrievalVideo-to-Text Retrieval
R@1↑R@5↑R@10↑MnR↓R@1↑R@5↑R@10↑MnR↓
w/o all Knowledge Stores (baseline)46.974.082.412.545.874.084.38.8
w/o Fine-grained Knowledge Store48.174.483.312.146.873.584.08.5
w/o Coarse-grained Knowledge Store48.474.583.711.947.273.884.08.4
w/o the whole E-step47.773.983.412.046.273.584.18.9
w/o Knowledge Interpolation E-step48.273.683.712.047.473.983.88.7
w/o OT Alignment Correction in E-step48.974.684.211.846.673.483.68.9
w/o the whole K-step47.973.883.112.347.073.684.08.7
w/o Fine-grained Calibration in K-step48.774.583.811.847.673.884.08.5
w/o Coarse-grained Calibration in K-step48.574.383.311.847.373.983.68.5
4 Our Full KDProR‡49.675.184.411.648.274.284.28.1
", + "bbox": [ + 220, + 184, + 787, + 345 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/fc88c9de3ce35d1afb69c10a09b10f0d80fbf06b2be8fb6710b16ec85bc6dc2e.jpg", + "image_caption": [ + "Fig. 4: Ablation experiment on MSR-VTT under closed-world settings. Effect of (a) the number K of top-K knowledge; (b) the knowledge interpolation coefficient $\\rho$ ." + ], + "image_footnote": [], + "bbox": [ + 261, + 359, + 500, + 479 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f2fdab8ca87fac4712c877274d2a5506bfb6b4dd17aa65e41ce06ecaab3386df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 361, + 743, + 481 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "on MSR-VTT. Table 4 shows that KDProR can be applied to successfully boost all baselines as a plug-and-play module under both closed-set or open-world settings. The significant improvements prove the generalization ability of KDProR.", + "bbox": [ + 212, + 551, + 787, + 599 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Ablation Study and Analysis", + "text_level": 1, + "bbox": [ + 214, + 626, + 501, + 642 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of the Multi-Grained Knowledge Store. As shown in Groups 1 and 4 in Table 5, removing either fine-grained or coarse-grained knowledge stores will result in a significant decrease in retrieval performance, e.g., $-1.5\\%$ and $-1.2\\%$ R@1 text-to-video retrieval on MSR-VTT, respectively. When all stores are removed, the model will lose the ability to decouple knowledge and introduce additional knowledge, resulting in a significant drop in performance. This verifies the effectiveness of our multi-grained knowledge decoupling strategy.", + "bbox": [ + 212, + 657, + 787, + 763 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of the Knowledge Injection in E-step. As shown in Groups 2 and 4 in Table 5, we evaluate the effect of the knowledge injection in E-step (i.e., Eq. 2) by removing different components. We can observe that when the knowledge interpolation strategy is removed, the performance drops sharply, indicating that our strategy can effectively inject knowledge from stores. We further remove the", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval", + "bbox": [ + 245, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f0cbf02afbdb2e22fd8706c4b5e215e083cd9bb2c6a0b4d4b0f7c91e38be7464.jpg", + "image_caption": [ + "Fig. 5: The text-video results on the MSR-VTT 1K-A test set. $\\top$ and $\\ddagger$ are close- and open-world settings respectively, whose meanings are consistent with those in Table 1." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 774, + 232 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "OT alignment correction strategy (i.e., Eq. 4) and observe a significant decrease in video-to-text retrieval performance, proving our strategy's effectiveness.", + "bbox": [ + 212, + 301, + 782, + 330 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effect of the Distribution Calibration in K-step. We conduct experiments to study the effect of the KNN distribution calibration strategy as illustrated in Table 5 with Groups 3 and 4. It can be seen that the performance of KDProR significantly decreases without KNN distribution to calibrate the retrieval distribution in Eq. 9. Moreover, the lack of distribution calibration (fine-grained or coarse-grained) in the K-step can also affect retrieval performance. This verifies that the multi-grained distribution calibration strategy can calibrate the retrieval distribution to inject additional knowledge sufficiently and effectively.", + "bbox": [ + 212, + 332, + 784, + 450 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effect of the number K of top-K Knowledge. In Figure 4 (left), we show the effect of the number K of top-K knowledge. On the one hand, we find that fewer K ( $K < 5$ ) means less knowledge is selected, which limits our KDProR's generalization ability for different knowledge. On the other hand, a larger K ( $K > 5$ ) requires more retrieval time, which increases the cost of learning.", + "bbox": [ + 212, + 452, + 784, + 527 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effect of the Knowledge Interpolation Coefficient. In Figure 4 (right), we show the effect of the knowledge interpolation coefficient $\\rho$ on KDProR. We can observe that KDProR achieves the optimal balance between knowledge injection and optimization when $\\rho = 0.2$ . Deviations from this optimal $\\rho$ , either towards a smaller or larger number, result in performance degradation, which is attributed to inadequate knowledge acquisition and insufficient optimization, respectively. Cases Study. We provide two cases of videos retrieved by KDProR and DRL [48]. As illustrated in Figure 5, KDProR successfully retrieves ground-truth videos in both closed- and open-set settings, whereas DRL fails to do so.", + "bbox": [ + 212, + 527, + 784, + 662 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 686, + 359, + 704 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we explore a novel knowledge-decoupling VTR paradigm that utilizes multi-grained knowledge stores to unify closed- and open-world knowledge. This strategy allows VTR models to achieve an efficient balance between memorization and retrieval optimization, while also opening up a unified interface for injecting various open-world knowledge. We further propose a principled EKM algorithm to achieve an efficient injection of knowledge. Our method could be applied to other visual-language tasks, e.g., visual question answering, and injecting various open-world knowledge for augmenting cross-modal understanding.", + "bbox": [ + 212, + 718, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "X. Zhuang et al.", + "bbox": [ + 271, + 114, + 382, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We would like to thank all reviewers for their insightful comments. This paper was partially supported by NSFC (No: 62176008).", + "bbox": [ + 215, + 176, + 784, + 205 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 229, + 321, + 244 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1728-1738 (2021)", + "2. Bogolin, S.V., Croitoru, I., Jin, H., Liu, Y., Albanie, S.: Cross modal retrieval with querybank normalisation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5194-5205 (2022)", + "3. Cheng, X., Lin, H., Wu, X., Yang, F., Shen, D.: Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290 (2021)", + "4. Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems 26 (2013)", + "5. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the em-algorithm plus discussions on the paper. In:\"\" (1977)", + "6. Deng, C., Chen, Q., Qin, P., Chen, D., Wu, Q.: Prompt switch: Efficient clip adaptation for text-video retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15648-15658 (2023)", + "7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: North American Chapter of the Association for Computational Linguistics (2019)", + "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv abs/2010.11929 (2020)", + "9. Elangovan, A., He, J., Verspoor, K.M.: Memorization vs. generalization : Quantifying data leakage in nlp performance evaluation. In: Conference of the European Chapter of the Association for Computational Linguistics (2021)", + "10. Fang, B., Liu, C., Zhou, Y., Yang, M., Song, Y., Li, F., Wang, W., Ji, X., Ouyang, W., et al.: Uatvr: Uncertainty-adaptive text-video retrieval. arXiv preprint arXiv:2301.06309 (2023)", + "1. Feldman, V.: Does learning require memorization? a short tale about a long tail. Proceedings of the 52nd Annual ACM SIGACT Symposium on Theory of Computing (2019)", + "2. Gabeur, V., Sun, C., Alahari, K., Schmid, C.: Multi-modal transformer for video retrieval. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 214-229. Springer (2020)", + "3. Gorti, S.K., Vouitsis, N., Ma, J., Golestan, K., Volkovs, M., Garg, A., Yu, G.: X-pool: Cross-modal language-video attention for text-video retrieval. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5006-5015 (2022)", + "4. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 770-778 (2015)" + ], + "bbox": [ + 225, + 258, + 784, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval", + "bbox": [ + 245, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 961-970 (2015)", + "16. Hendricks, L.A., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.C.: Localizing moments in video with natural language. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 5804-5813 (2017)", + "17. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Computation 9, 1735-1780 (1997)", + "18. Ibrahimi, S., Sun, X., Wang, P., Garg, A., Sanan, A., Omar, M.: Audio-enhanced text-to-video retrieval using text-conditioned feature alignment. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 12020-12030 (2023)", + "19. Jensen, J.L.W.V.: Sur les fonctions convexes et les inégalités entre les valeurs moyennes. Acta Mathematica 30, 175-193 (1906)", + "20. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q.V., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. ArXiv abs/2102.05918 (2021)", + "21. Jin, P., Huang, J., Liu, F., Wu, X., Ge, S., Song, G., Clifton, D., Chen, J.: Expectation-maximization contrastive learning for compact video-and-language representations. Advances in Neural Information Processing Systems 35, 30291-30306 (2022)", + "22. Jin, P., Li, H., Cheng, Z.L., Li, K., Ji, X., Liu, C., ming Yuan, L., Chen, J.: Diffusionret: Generative text-video retrieval with diffusion model. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 2470-2481 (2023)", + "23. Khandelwal, U., Levy, O., Jurafsky, D., Zettlemoyer, L., Lewis, M.: Generalization through memorization: Nearest neighbor language models. arXiv preprint arXiv:1911.00172 (2019)", + "24. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. CoRR abs/1412.6980 (2014)", + "25. Li, H., Cao, M., Cheng, X., Li, Y., Zhu, Z., Zou, Y.: G2l: Semantically aligned and uniform video grounding via geodesic and game theory. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 11998-12008 (2023), https://api-semanticscholar.org/CorpusID:260164585", + "26. Li, H., Cao, M., Cheng, X., Zhu, Z., Li, Y., Zou, Y.: Exploiting auxiliary caption for video grounding. In: AAAI Conference on Artificial Intelligence (2023), https://apisemantic scholar.org/CorpusID:257772084", + "27. Li, K., Wang, Y., Li, Y., Wang, Y., He, Y., Wang, L., Qiao, Y.: Unmasked teacher: Towards training-efficient video foundation models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 19891-19903 (2023)", + "28. Liu, R., Huang, J., Li, G., Feng, J., Wu, X., Li, T.H.: Revisiting temporal modeling for clip-based image-to-video knowledge transferring. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6555-6564 (2023)", + "29. Liu, S., Fan, H., Qian, S., Chen, Y., Ding, W., Wang, Z.: Hit: Hierarchical transformer with momentum contrast for video-text retrieval. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 11895-11905 (2021)", + "30. Liu, S., Fan, H., Qian, S., Chen, Y., Ding, W., Wang, Z.: Hit: Hierarchical transformer with momentum contrast for video-text retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11915-11925 (2021)", + "31. Liu, Y., Albanie, S., Nagrani, A., Zisserman, A.: Use what you have: Video retrieval using representations from collaborative experts. arXiv preprint arXiv:1907.13487 (2019)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "X. Zhuang et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. ArXiv abs/1907.11692 (2019)", + "33. Liu, Y., Xiong, P., Xu, L., Cao, S., Jin, Q.: Ts2-net: Token shift and selection transformer for text-video retrieval. In: European Conference on Computer Vision. pp. 319-335. Springer (2022)", + "34. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv: Learning (2016)", + "35. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing 508, 293-304 (2022)", + "36. Ma, Y., Xu, G., Sun, X., Yan, M., Zhang, J., Ji, R.: X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. Proceedings of the 30th ACM International Conference on Multimedia (2022)", + "37. Meng, Y., Zong, S., Li, X., Sun, X., Zhang, T., Wu, F., Li, J.: Gnn-lm: Language modeling based on global contexts via gnn. ArXiv abs/2110.08743 (2021)", + "38. Mikolov, T., Chen, K., Corrado, G.S., Dean, J.: Efficient estimation of word representations in vector space. In: International Conference on Learning Representations (2013)", + "39. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L.E., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.J.: Training language models to follow instructions with human feedback. ArXiv abs/2203.02155 (2022)", + "40. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)", + "41. Papernot, N., McDaniel, P.: Deep k-nearest neighbors: Towards confident, interpretable and robust deep learning. arXiv preprint arXiv:1803.04765 (2018)", + "42. Patrick, M., Huang, P.Y., Asano, Y., Metze, F., Hauptmann, A., Henriques, J., Vedaldi, A.: Support-set bottlenecks for video-text representation learning. arXiv preprint arXiv:2010.02824 (2020)", + "43. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021)", + "44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "45. Rohrbach, A., Rohrbach, M., Tandon, N., Schiele, B.: A dataset for movie description. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3202-3212 (2015)", + "46. Su, B., Hua, G.: Order-preserving Wasserstein distance for sequence matching. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1049-1057 (2017)", + "47. Tänzer, M., Ruder, S., Rei, M.: Memorisation versus generalisation in pre-trained language models. In: Annual Meeting of the Association for Computational Linguistics (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval", + "bbox": [ + 245, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Wang, Q., Zhang, Y., Zheng, Y., Pan, P., Hua, X.: Disentangled representation learning for text-video retrieval. ArXiv abs/2203.07111 (2022)", + "49. Wang, Z., Sung, Y.L., Cheng, F., Bertasius, G., Bansal, M.: Unified coarse-to-fine alignment for video-text retrieval. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 2804-2815 (2023)", + "50. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)", + "51. Wu, H., Mao, J., Zhang, Y., Jiang, Y., Li, L., Sun, W., Ma, W.Y.: Unified visual-semantic embeddings: Bridging vision and language with structured meaning representations. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6602-6611 (2019)", + "52. Wu, W., Luo, H., Fang, B., Wang, J., Ouyang, W.: Cap4video: What can auxiliary captions do for text-video retrieval? 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 10704-10713 (2022)", + "53. Wu, X., Li, H., Luo, Y., Cheng, X., Zhuang, X., Cao, M., Fu, K.: Uncertainty-aware sign language video retrieval with probability distribution modeling. ArXiv abs/2405.19689 (2024), https://api.sementicscholar.org/CorpusID: 270123137", + "54. Xie, Y., Zhu, Z., Zhuang, X., Liang, L., Wang, Z., Zou, Y.: Gpa: Global and prototype alignment for audio-text retrieval. In: Interspeech (2024)", + "55. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5288-5296 (2016)", + "56. Xu, Z., Chen, Z., Zhang, Y., Song, Y., Wan, X., Li, G.: Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17503-17512 (2023)", + "57. Xu, Z., Huang, J., Liu, T., Liu, Y., Han, H., Yuan, K., Li, X.: Enhancing fine-grained multi-modal alignment via adapters: A parameter-efficient training framework for referring image segmentation. In: 2nd Workshop on Advancing Neural Network Training: Computational Efficiency, Scalability, and Resource Optimization (WANT@ ICML 2024) (2024)", + "58. Xue, H., Hang, T., Zeng, Y., Sun, Y., Liu, B., Yang, H., Fu, J., Guo, B.: Advancing high-resolution video-language representation with large-scale video transcriptions. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5026-5035 (2021)", + "59. Xue, H., Sun, Y., Liu, B., Fu, J., Song, R., Li, H., Luo, J.: Clip-vip: Adapting pre-trained image-text model to video-language alignment. In: International Conference on Learning Representations (2023)", + "60. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. Trans. Mach. Learn. Res. 2022 (2022)", + "61. Yu, Y., Kim, J., Kim, G.: A joint sequence fusion model for video question answering and retrieval. In: European Conference on Computer Vision (2018)", + "62. Yuan, L., Chen, D., Chen, Y.L., Codella, N.C.F., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., Liu, C., Liu, M., Liu, Z., Lu, Y., Shi, Y., Wang, L., Wang, J., Xiao, B., Xiao, Z., Yang, J., Zeng, M., Zhou, L., Zhang, P.: Florence: A new foundation model for computer vision. ArXiv abs/2111.11432 (2021)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "X. Zhuang et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "63. Zhao, Y., Li, K., Cheng, Z., Qiao, P., Zheng, X., Ji, R., Liu, C., Yuan, L., Chen, J.: Graco: Granularity-controllable interactive segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3501-3510 (June 2024)", + "64. Zhao, Y., Lv, W., Xu, S., Wei, J., Wang, G., Dang, Q., Liu, Y., Chen, J.: Detrs beat yolos on real-time object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16965-16974 (June 2024)", + "65. Zheng, X., Jiang, J.: An empirical study of memorization in nlp. In: Annual Meeting of the Association for Computational Linguistics (2022)" + ], + "bbox": [ + 215, + 146, + 787, + 287 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval", + "bbox": [ + 245, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/KDProR_ A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval/8778d22d-a6f9-4fed-bc5f-fddacba7cf0b_model.json b/2024/KDProR_ A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval/8778d22d-a6f9-4fed-bc5f-fddacba7cf0b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..52e60664274b3e91e3e7d251d8804421f629e23b --- /dev/null +++ b/2024/KDProR_ A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval/8778d22d-a6f9-4fed-bc5f-fddacba7cf0b_model.json @@ -0,0 +1,2636 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.141, + 0.78, + 0.185 + ], + "angle": 0, + "content": "KDProR: A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.213, + 0.674, + 0.244 + ], + "angle": 0, + "content": "Xianwei Zhuang*, Hongxiang Li*, Xuxin Cheng, Zhihong Zhu, Yuxin Xie, Yuexian Zou†" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.256, + 0.718, + 0.284 + ], + "angle": 0, + "content": "ADSPLAB, School of ECE, Peking University, China {xwzhuang, lihongxiang}@stu.pku.edu.cn, zouyx@pku.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.323, + 0.74, + 0.668 + ], + "angle": 0, + "content": "Abstract. Existing video-text retrieval methods predominantly focus on designing diverse cross-modal interaction mechanisms between captions and videos. However, those approaches diverge from human learning paradigms, where humans possess the capability to seek and associate knowledge from an open set, rather than rote memorizing all text-video instances. Motivated by this, we attempt to decouple knowledge from retrieval models through multi-grained knowledge stores and identify two significant benefits of our knowledge-decoupling strategy: (1) it ensures a harmonious balance between knowledge memorization and retrieval optimization, thereby improving retrieval performance; and (2) it can promote incorporating diverse open-world knowledge to augment video-text retrieval. To efficiently integrate information from knowledge stores, we further introduce a novel retrieval framework termed KDProR, which utilizes our proposed Expectation-Knowledge-Maximization (EKM) algorithm for optimization. Specifically, in E-step, KDProR obtains relevant contextual semantics from knowledge stores and achieves efficient knowledge injection through interpolation and alignment correction. During the K-step, KDProR calculates the knowledge KNN distribution by indexing the top-K acquired knowledge to refine the retrieval distribution, and in M-step, KDProR optimizes the retrieval model by maximizing the likelihood of the objective. Extensive experiments on various benchmarks prove that KDProR significantly outperforms previous state-of-the-art methods across all metrics. Remarkably, KDProR can uniformly and efficiently incorporate diverse open-world knowledge and is compatible with different interaction mechanisms and architectures." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.682, + 0.717, + 0.697 + ], + "angle": 0, + "content": "Keywords: Video Retrieval \\(\\cdot\\) Knowledge Decoupling \\(\\cdot\\) Open World" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.723, + 0.376, + 0.74 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.755, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Video-text retrieval (VTR) is a significant and challenging task in cross-modal interaction [1,12,31,42], aiming to align video (text) candidates with text (video) queries to identify the most relevant instances accurately [61]. The standard" + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.811, + 0.478, + 0.825 + ], + "angle": 0, + "content": "* These authors contributed equally." + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.825, + 0.396, + 0.84 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.233, + 0.811, + 0.478, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.129 + ], + "angle": 0, + "content": "X. Zhuang et al." + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.145, + 0.764, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.298, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Fig. 1: The illustration of different VTR paradigms. (a) denotes vanilla VTR methods; (b) introduces additional data only during training. (c) achieves \"open examination\" with knowledge decoupling, thereby enhancing VTR in both closed and open sets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.506 + ], + "angle": 0, + "content": "paradigm of VTR for addressing cross-modal matching involves aligning videotext features acquired through visual and textual encoders via contrastive learning. With the advent of large-scale image-language pre-trained models [20, 44, 60, 62], recent VTR methods [3, 13, 35] have realized substantial advancements in retrieval performance by leveraging pre-trained image-language models, such as CLIP [44]. Benefiting from powerful CLIP, most of these methods focus on devising various cross-modal interaction strategies to enhance visual-text alignment, including cross-attention [13], multi-grained interaction [36], disentangled representation learning [48], etc, and have achieved considerable success." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.779 + ], + "angle": 0, + "content": "However, if we conceptualize the training set as a \"book\" intuitively, this training-test paradigm for VTR can be likened to \"video-text memorization\" and \"closed-book VTR examination\", which diverges from the human learning process [37, 65]. Humans possess the capability to seek and associate knowledge from an open set, rather than relying on rote memorization of all text-video instances [47]. This \"closed-book examination\" approach may be further hindered by dense memorization overhead [37] and challenges in memorizing difficult samples [9, 11], echoing the proverb: \"The palest ink is better than the best memory.\" Furthermore, previous open-set fine-tuning [18, 52] and pre-training [27, 59] methods have shown that incorporating auxiliary data from the open world can improve VTR performance. These insights encourage us to explore a novel VTR paradigm: \"open-set examination\", which involves constructing knowledge stores from the closed training set or open-world sources to decouple knowledge. As shown in Figure 1, this strategy introduces relevant knowledge from stores as reinforcement signals to help the model strike a balance between generalization and memory. However, this also brings two new challenges: (1) How to decouple knowledge from models and unify closed- and open-world knowledge; (2) How to efficiently incorporate decoupled knowledge to improve retrieval performance." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.787, + 0.81 + ], + "angle": 0, + "content": "In this work, we propose a novel knowledge-decoupling probabilistic framework for VTR (KDProR), tackling the challenges with the following strategies:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "(1) KDProR utilizes multi-grained stores to decouple knowledge and standardize closed- and open-world knowledge sources. We con" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "struct multi-scale knowledge stores composed of local and global neural representations to decouple knowledge. This approach allows rare patterns to be memorized explicitly, rather than implicitly in model parameters. In addition, knowledge stores can be updated asynchronously and endowed with knowledge from closed- or open-world sources. In this work, we explore three different sources of additional knowledge, including spatio-temporal knowledge from pre-trained models, entity-relation structured knowledge and outputs of large language models (LLMs, e.g., GPT [40]) prompted by chain-of-thoughts (CoT) [50]. Moreover, we identify two core advantages of the knowledge decoupling strategy: it (a) ensures a harmonious balance between memorization and retrieval optimization; and (b) opens up a unified interface for injecting diverse open-world knowledge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.479 + ], + "angle": 0, + "content": "(2) KDProR utilizes a novel Expectation-Knowledge-Maximization algorithm to achieve efficient knowledge injection and retrieval optimization. We extend the traditional Expectation-Maximization (EM) algorithm [5] to the Expectation-Knowledge-Maximization (EKM) algorithm for optimizing KDProR. Specifically, in the E-step, KDProR obtains relevant contextual semantics and achieves efficient knowledge injection through interpolation and alignment correction. In the K-step, we calculate the k-nearest neighbor (KNN) distribution by indexing the top-K knowledge obtained to calibrate the original retrieval distribution. And in the M-step, we optimize the model by maximizing the likelihood of the target. Additionally, we provide a theoretical analysis of our KDProR framework from a probabilistic perspective in Section 3.7." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.789, + 0.631 + ], + "angle": 0, + "content": "Extensive experiments on four benchmarks, i.e., MSR-VTT [55], DiDeMo [16], LSMDC [45] and ActivityNet [15], prove that KDProR significantly outperforms previous state-of-the-art methods in both closed and open-world settings. In summary, our KDProR has several appealing facets: (1) Effectiveness: It can significantly improve retrieval performance under both closed-set and open-set settings. (2) Universality: It can be used to introduce various open-world knowledge, while being compatible with various interaction modules and pretrained foundation models (e.g., [27, 59]). (3) Explainability: The effectiveness and convergence of retrieval optimization via our EKM algorithm are theoretically guaranteed. The main contributions of this paper are presented as follows:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.631, + 0.788, + 0.753 + ], + "angle": 0, + "content": "(1) We explore a novel knowledge-decoupling paradigm for VTR and construct multi-grained knowledge stores to uniformly introduce knowledge from closed sets and diverse open-world sources. (2) We propose a novel VTR framework termed KDProR, which utilizes a principled EKM algorithm to inject various knowledge into the model to improve retrieval performance. Theoretical analysis provides an underlying insight into the effectiveness and convergence of KDProR. (3) Experiments conducted on four benchmarks show that KDProR achieves new state-of-the-art results under both closed- and open-world settings." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.777, + 0.383, + 0.792 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Feature Representation for Video-Text Retrieval. Recent studies [1, 12, 31, 42] on VTR employ a bi-encoder architecture to extract text and video rep" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.382, + 0.129 + ], + "angle": 0, + "content": "X. Zhuang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "resentation. The selection of text and visual encoders changes with the development of research on feature extraction upon NLP and CV, e.g., Word2Vec [38] and BERT-based models [7, 32] for textual representation, and CNN-based [14] and ViT-based models [8, 63, 64] for visual tasks. These researches use independently pre-trained dual encoders for visual-textual semantic extraction [1, 12, 42, 57]. Recently, CLIP [43] pre-trained on a large-scale dataset with 400 million pairs has achieved excellent capability of cross-modal representation. And CLIP-based VTR methods [3, 13, 35, 53] have rapidly developed and achieved great success. Our KDProR also benefits from existing pre-training, while exploring a novel perspective for enhancing representation, i.e., knowledge decoupling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.433 + ], + "angle": 0, + "content": "Interaction Mechanism for Video-Text Retrieval. Many VTR methods explore different cross-modal interaction mechanisms for cosine similarity [35] to improve retrieval performance. Some studies [35] seek to enhance the spatiotemporal understanding ability of image encoders through LSTM [17] or temporal Transformers. MoE-based [3] and Cross-Transformers-based [13,54] methods significantly improve the cross-modal interaction. [25, 26, 29, 36, 49, 56] explore hierarchical and multi-grained interactions, respectively. DRL [48] proposes disentangled representation learning for cross-modal interaction. Our KDProR is compatible with these cross-modal interaction strategies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.603 + ], + "angle": 0, + "content": "Open-World Knowledge for Video-Text Retrieval. The previous methods mainly introduce open-world knowledge in the pre-training and fine-tuning stages to enhance VTR performance. CLIP-ViP [59] conducts secondary pretraining of CLIP on large-scale video-text datasets, and UMT [27] achieves efficient pre-training to improve the spatio-temporal understanding and inject open-set knowledge into VTR tasks. TEFAL [18] utilizes additional audio features to enhance VTR retrieval performance and Cap4Video [52] utilizes auxiliary captions obtained by Web Search or generated by LLMs to enhance cross-modal interaction. KDProR is compatible with these pre-training methods, and we do not specify the type of open-world knowledge to be introduced, but instead achieve universal open-set knowledge injection through knowledge decoupling." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.621, + 0.341, + 0.637 + ], + "angle": 0, + "content": "3 Methods" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.651, + 0.533, + 0.667 + ], + "angle": 0, + "content": "3.1 Settings and Feature Extraction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Following previous work [35, 48], we utilize the dual-encoder architecture of CLIP [44] to achieve video and text feature extraction on training set \\(\\mathcal{D} = \\{x_i\\}_i^N = \\{(T_i,V_i)\\}_i^N\\), where \\(T_{i}\\) and \\(V_{i}\\) represent text and video instances respectively, and \\(N\\) is the size of the training set. Given a text query \\(T_{i}\\), we initially add identifiers [CLS] and [SEP] to the sentence, and leverage the text encoder of CLIP to encode the text representation: \\(t_i = \\mathrm{TextEncoder}(T_i)\\). For each video \\(V_{i}\\), we uniformly select \\(n_v\\) frames \\([F_1,\\dots ,F_{n_v}]\\) as keyframes, and employ the transformer-based encoder of CLIP to extract sequential features \\(v_{i} = \\mathrm{VideoEncoder}([F_{1},\\dots ,F_{n_{v}}])\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Extensive work explores various cross-modal interaction strategies to improve retrieval performance [30,35,48]. Our work focuses on decoupling knowledge from" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.145, + 0.768, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.343, + 0.788, + 0.398 + ], + "angle": 0, + "content": "Fig. 2: The illustration of the proposed framework, consisting of (1) constructing a multi-grained knowledge store using open-world knowledge or closed-world knowledge (training dataset); (2) EKM algorithm that sequentially implements knowledge injection, distribution calibration, and target likelihood maximization." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.426, + 0.784, + 0.486 + ], + "angle": 0, + "content": "models and exploring the impact of different knowledge sources on retrieval performance, rather than pre-training or interaction strategies. Thus we adopt the weighted token-wise interaction module of DRL [48] to achieve cross-modal interaction. Note that KDProR is also compatible with other interaction strategies." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.505, + 0.501, + 0.52 + ], + "angle": 0, + "content": "3.2 Decoupling Knowledge Store" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.784, + 0.602 + ], + "angle": 0, + "content": "Our KDProR involves augmenting CLIP-based retrieval models with semantic information from decoupled knowledge stores. Thus, the first step of our methods is to build a multi-grained knowledge store. As illustrated in Figure 2, our stores contain fine-grained and coarse-grained caches, where each key and value (text-video) are representations of knowledge." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.603, + 0.785, + 0.696 + ], + "angle": 0, + "content": "Fine-grained Knowledge Store. The fine-grained knowledge store can be represented as \\(\\mathcal{S}^f = \\{\\mathcal{S}_i^f\\}_{i=1}^{K^f} = \\{(\\mathcal{K}_i^f, \\mathcal{V}_i^f)\\}_{i=1}^{K^f}\\), where \\(\\mathcal{S}_i^f = (\\mathcal{K}_i^f, \\mathcal{V}_i^f)\\) denotes the \\(i\\)-th key-value pair in the fine-grained knowledge store, and \\(\\mathcal{K}_i^f\\) is the size of our fine-grained knowledge store. Under the closed world setting, we utilize text and visual encoders of CLIP to obtain the feature key-value pairs on the training dataset as \\(\\mathcal{K}_i^f = t_i\\) and \\(\\mathcal{V}_i^f = v_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.697, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Coarse-grained Knowledge Store. The coarse-grained knowledge store is designed to encapsulate global semantics. It equips retrieval models with global, coarse-grained insights from the knowledge store, enhancing the comprehensive knowledge understanding. Similarly, the coarse-grained storage can be represented as: \\(\\mathcal{S}^c = \\{\\mathcal{S}_i^c\\}_{i=1}^{K^c} = \\{(\\mathcal{K}_i^c, \\mathcal{V}_i^c)\\}_{i=1}^{K^c}\\), where \\(\\mathcal{S}_i^c = (\\mathcal{K}_i^c, \\mathcal{V}_i^c)\\) denotes the \\(i\\)-th key-value pair in the coarse-grained knowledge store, and \\(\\mathcal{K}_i^c\\) is the size of our coarse-grained knowledge store. We adopt the K-mean strategy to extract knowledge from fine-grained storage into more advanced representations:" + }, + { + "type": "equation", + "bbox": [ + 0.326, + 0.824, + 0.786, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathcal {K} _ {i} ^ {c} = \\phi \\left(c _ {i} ^ {t}\\right), \\forall c _ {i} ^ {t} \\in \\mathcal {C} ^ {t}, \\quad \\mathcal {V} _ {i} ^ {c} = \\phi \\left(c _ {i} ^ {v}\\right), \\forall c _ {i} ^ {v} \\in \\mathcal {C} ^ {v}, \\tag {1}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.382, + 0.129 + ], + "angle": 0, + "content": "X. Zhuang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.196 + ], + "angle": 0, + "content": "where, \\(\\mathcal{C}^t = \\{c_1^t,\\dots ,c_{K^c}^t\\}\\) denotes the cluster set obtained by K-Mean over fine-grained knowledge \\(\\{\\mathcal{K}_i^f\\}\\), \\(\\mathcal{C}^v = \\{c_1^v,\\dots ,c_{K^c}^v\\}\\) represents the cluster set obtained by K-Mean over \\(\\{\\mathcal{V}_i^f\\}\\), and \\(\\phi (\\cdot)\\) represents the Max-Pooling aggregate function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.197, + 0.785, + 0.226 + ], + "angle": 0, + "content": "Note that our framework is universal, and we will introduce more applications for constructing diverse knowledge stores in Section 3.6." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.248, + 0.644, + 0.264 + ], + "angle": 0, + "content": "3.3 E-step for Multi-Grained Knowledge Injection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.444 + ], + "angle": 0, + "content": "Knowledge Interpolation. In E-step, KDProR initially attempts to inject fine-grained and coarse-grained knowledge into the model by knowledge interpolation of features. We treat with \\( t_i \\) and \\( v_i \\) as anchors to retrieve the fine-grained knowledge store, and obtain two sets of top-K similar vector pairs \\( \\mathcal{N}_{t_i}^f \\) and \\( \\mathcal{N}_{v_i}^f \\), respectively. \\( \\mathcal{N}_{t_i}^f = \\{(\\mathcal{K}_{t_i}^f, \\mathcal{V}_{t_i}^f)\\} \\) and \\( \\mathcal{N}_{v_i}^f = \\{(\\mathcal{K}_{v_i}^f, \\mathcal{V}_{v_i}^f)\\} \\) are two sets of key-value knowledge pairs, where \\( \\mathcal{K}_{t_i}^f \\) and \\( \\mathcal{V}_{v_i}^f \\) are the nearest neighbors of \\( t_i \\) and \\( v_i \\), respectively. After the model retrieves the top-K candidates for each \\( t_i \\) and \\( v_i \\), their corresponding representation \\( \\mathcal{K}_{t_i}^f \\) and \\( \\mathcal{V}_{v_i}^f \\) obtained from knowledge-store will be incorporated into original features \\( t_i \\) and \\( v_i \\) to act as demonstration learning. Specifically, we adopt the nearest neighbor sets to inject decoupled knowledge into text and video features as:" + }, + { + "type": "equation", + "bbox": [ + 0.251, + 0.453, + 0.786, + 0.493 + ], + "angle": 0, + "content": "\\[\nt _ {i} ^ {f} = \\rho t _ {i} + (1 - \\rho) \\frac {\\sum_ {(\\mathcal {K} , \\mathcal {V}) \\in \\mathcal {N} _ {t _ {i}} ^ {f}} \\mathcal {K}}{| \\mathcal {N} _ {t _ {i}} ^ {f} |}, v _ {i} ^ {f} = \\rho v _ {i} + (1 - \\rho) \\frac {\\sum_ {(\\mathcal {K} , \\mathcal {V}) \\in \\mathcal {N} _ {v _ {i}} ^ {f}} \\mathcal {V}}{| \\mathcal {N} _ {v _ {i}} ^ {f} |}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.784, + 0.548 + ], + "angle": 0, + "content": "where, \\(\\rho\\) is a hyperparameter used to express the interpolation ratio. Subsequently, we utilize the weighted token-wise interaction module [48] to enhance textual and visual representations injected with additional knowledge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.785, + 0.623 + ], + "angle": 0, + "content": "However, the textual and visual knowledge within the top-K sets, obtained by \\( t_i \\) and \\( v_i \\) respectively, may not correspond on a one-to-one basis due to their derivation from two independent top-K operations. To mitigate the issue of mismatched video-text features during knowledge injection, we utilize the optimal transport (OT) algorithm to facilitate alignment correction within a batch." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.624, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Alignment Correction via Optimal Transport. We first define \\(\\langle \\mathbf{Q},\\mathbf{S}^f\\rangle =\\) \\(\\mathrm{tr}(\\mathbf{Q}^\\top \\mathbf{S}^f)\\in \\mathbb{R}^{B\\times B}\\) as the similarity matrix calculated from the updated text and video pairs \\(\\{(t_i^f,v_i^f)\\}_{i = 1}^B\\) in a batch. \\(\\mathbf{Q}\\in \\mathbb{R}^{B\\times B}\\) represents the corresponding transport assignment matrix, where \\(\\mathbf{Q}[i,j]\\) denotes the probabilities of aligning \\(t_i^f\\) with \\(v_{i}^{f}\\). The optimal transport aims to establish flexible alignment between videos and captions by maximizing global text-video similarity \\(\\mathrm{tr}(\\mathbf{Q}^\\top \\mathbf{S}^f)\\). The objective of optimal transport can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.241, + 0.743, + 0.785, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\mathbf {Q} \\in \\mathcal {Q}} \\langle \\mathbf {Q}, \\mathbf {S} ^ {f} \\rangle + \\varepsilon H (\\mathbf {Q}) \\quad \\text {s . t .} \\mathcal {Q} = \\left\\{\\mathbf {Q} \\in \\mathbb {R} ^ {B \\times B} \\mid \\mathbf {Q} \\mathbf {1} _ {B} = \\boldsymbol {\\mu}, \\mathbf {Q} ^ {\\top} \\mathbf {1} _ {B} = \\boldsymbol {\\nu} \\right\\} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.779, + 0.788, + 0.842 + ], + "angle": 0, + "content": "where, \\(\\pmb{\\mu} \\in \\mathbb{R}^{B}\\) and \\(\\pmb{\\nu} \\in \\mathbb{R}^{B}\\) indicate the relative importance of each video and caption, \\(\\mathbf{1}_{B}\\) represents the vector of ones in dimension \\(B\\), \\(H(\\mathbf{Q})\\) denotes a entropy regular term [4] and \\(\\varepsilon\\) controls its smoothness. Following [46], we initialize \\(\\pmb{\\mu}\\) and \\(\\pmb{\\nu}\\) to a uniform distribution \\(\\mathbf{1}_{B}\\). The optimal transport problem" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "in Eq. 3 can be solved using the Sinkhorn-Knopp algorithm [46] to obtain the optimal assignment scheme \\(\\mathbf{Q}^*\\). Based on this, we can obtain the similarity matrix after adding alignment correction \\(\\mathbf{Q}^*\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.199, + 0.786, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} ^ {f ^ {*}} = \\left(\\left(1 - \\beta\\right) \\mathbf {I} + \\beta \\mathbf {Q} ^ {*}\\right) \\mathbf {S} ^ {f}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.224, + 0.637, + 0.239 + ], + "angle": 0, + "content": "where, \\(\\mathbf{I}\\) is the identity matrix and \\(\\beta\\) is a hyperparameter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.784, + 0.269 + ], + "angle": 0, + "content": "We can implement the coarse-grained knowledge injection and obtain the realigned similarity matrix \\(\\mathbf{S}^{c*}\\) in a similar way." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.288, + 0.634, + 0.303 + ], + "angle": 0, + "content": "3.4 K-step for Retrieval Distribution Calibration" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.785, + 0.414 + ], + "angle": 0, + "content": "Existing studies [23, 41] have shown that non-parametric KNN classification or nearest-neighbor interpolation methods have extraordinary potential in improving robustness and generalization. It is intuitively to leverage the KNN's retrieval results as the prior knowledge to guide the parameters training for video-text pairs that are difficult to establish deep correlations. In practice, we aggregate the prediction of KNN into the original probability distribution of video-text retrieval to achieve the multi-grained calibration of retrieval distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.785, + 0.492 + ], + "angle": 0, + "content": "Fine-grained Distribution Calibration. We utilize the set of k-nearest neighbors \\(\\mathcal{N}_{t_i}^f\\) and \\(\\mathcal{N}_{v_i}^f\\) retrieved by querying the open-book knowledge store in E-step as fine-grained top-K knowledge. Subsequently, we calculate the KNN distribution according to the softmax of the similarity and aggregation probability mass of each video-text pair in the retrieved target as:" + }, + { + "type": "equation", + "bbox": [ + 0.231, + 0.497, + 0.785, + 0.542 + ], + "angle": 0, + "content": "\\[\np _ {k n n} ^ {f} \\left(v _ {i} \\mid t _ {i}\\right) = \\frac {\\sum_ {\\left(\\mathcal {K} , \\mathcal {V}\\right) \\in N _ {t _ {i}} ^ {f}} \\varphi \\left(v _ {i} , \\mathcal {V}\\right) e ^ {d \\left(v _ {i} , \\mathcal {V}\\right)}}{\\sum_ {\\left(\\mathcal {K} , \\mathcal {V}\\right) \\in N _ {t _ {i}} ^ {f}} \\varphi \\left(v _ {i} , \\mathcal {V}\\right)}, \\quad \\varphi \\left(v _ {i}, \\mathcal {V}\\right) = \\left\\{ \\begin{array}{l l} 1, & \\mathcal {V} \\in N _ {v _ {i}} ^ {f} \\\\ 0, & \\mathcal {V} \\notin N _ {v _ {i}} ^ {f} \\end{array} , \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.784, + 0.594 + ], + "angle": 0, + "content": "where, \\( d(\\cdot ,\\cdot)\\in [-1,1] \\) denotes the cosine similarity function, and \\( \\varphi (v_{i},\\mathcal{V}) \\) is an indicator function used to determine whether \\( \\mathcal{V} \\) is within the top-K nearest neighbor set \\( N_{v_i}^f \\) of the \\( v_{i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.784, + 0.626 + ], + "angle": 0, + "content": "We can further obtain the KNN probability distribution \\( p_{knn}^{f}(t_{i} \\mid v_{i}) \\) for video-to-text retrieval in a completely symmetrical manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.785, + 0.794 + ], + "angle": 0, + "content": "Coarse-grained Distribution Calibration. Consistent with Eq. 5, we employ the KNN algorithm to obtain top-K knowledge, utilizing closed-world or open-world knowledge from the coarse-grained knowledge store to refine the probability distribution function pertinent to the text-to-video retrieval. Nevertheless, since the coarse-grained knowledge store is derived from the distillation of fine-grained knowledge, it is impracticable to apply indicator functions (cf. Eq. 5) to ascertain the presence of an inclusion relation between textual features and their nearest neighbor sets. To tackle this challenge, we introduce a novel indicator function, denoted as \\(\\phi(\\cdot, \\cdot)\\), parameterized by a threshold \\(\\alpha\\). \\(\\phi(\\cdot, \\cdot)\\) is designed to evaluate whether \\(\\mathcal{V}\\) resides within the top-K nearest neighbor set \\(N_{v_i}^c\\) of the \\(v_i\\) via the cosine similarity function \\(d(\\cdot, \\cdot)\\), and is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.805, + 0.785, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\phi \\left(v _ {i}, \\mathcal {V}\\right) = \\left\\{ \\begin{array}{l l} 1, & d \\left(\\mathcal {V}, \\mathcal {V} ^ {c}\\right) < \\alpha , \\exists \\left(\\mathcal {K} ^ {c}, \\mathcal {V} ^ {c}\\right) \\in N _ {v _ {i}} ^ {c} \\\\ 0, & o t h e r w i s e s \\end{array} . \\right. \\tag {6}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.382, + 0.129 + ], + "angle": 0, + "content": "X. Zhuang et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "Based on this, we can obtain a closed-form expression for the KNN distribution used for coarse-grained distribution calibration as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.186, + 0.786, + 0.23 + ], + "angle": 0, + "content": "\\[\np _ {k n n} ^ {c} \\left(v _ {i} \\mid t _ {i}\\right) = \\frac {\\sum_ {\\left(\\mathcal {K} , \\mathcal {V}\\right) \\in N _ {t _ {i}} ^ {c}} \\phi \\left(v _ {i} , \\mathcal {V}\\right) e ^ {d \\left(v _ {i} , \\mathcal {V}\\right)}}{\\sum_ {\\left(\\mathcal {K} , \\mathcal {V}\\right) \\in N _ {t _ {i}} ^ {c}} \\phi \\left(v _ {i} , \\mathcal {V}\\right)}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.237, + 0.784, + 0.267 + ], + "angle": 0, + "content": "We can further obtain the KNN probability distribution \\( p_{knn}^{c}(t_{i} \\mid v_{i}) \\) for coarse-grained video-to-text retrieval in a completely symmetrical manner." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.287, + 0.61, + 0.303 + ], + "angle": 0, + "content": "3.5 M-step for Unified Retrieval Optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.784, + 0.369 + ], + "angle": 0, + "content": "In the M-step, we optimize the whole model by maximizing the retrieval likelihood objective over the realigned similarity matrix. Specifically, we utilize the re-aligned multi-grained similarity matrix in Eq. 4 obtained from E-step to calculate the InfoNCE loss:" + }, + { + "type": "equation", + "bbox": [ + 0.276, + 0.377, + 0.785, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t 2 v} = - \\frac {1}{B} \\sum_ {i} ^ {B} \\left(\\log \\frac {\\exp \\left(\\mathbf {S} ^ {f _ {i i} ^ {*}}\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\mathbf {S} ^ {f _ {i j} ^ {*}}\\right)} + \\log \\frac {\\exp \\left(\\mathbf {S} _ {i i} ^ {c ^ {*}}\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\mathbf {S} _ {i j} ^ {c ^ {*}}\\right)}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.276, + 0.43, + 0.729, + 0.478 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v 2 t} = - \\frac {1}{B} \\sum_ {i} ^ {B} \\left(\\log \\frac {\\exp \\left(\\mathbf {S} _ {i i} ^ {f ^ {*}}\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\mathbf {S} _ {j i} ^ {f ^ {*}}\\right)} + \\log \\frac {\\exp \\left(\\mathbf {S} _ {i i} ^ {c ^ {*}}\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\mathbf {S} _ {j i} ^ {c ^ {*}}\\right)}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.484, + 0.785, + 0.59 + ], + "angle": 0, + "content": "where \\( B \\) is the batch size. Subsequently, we propose utilizing the KNN distribution (cf. Eq. 5 and 7) to guide the training process. The KNN calibrator reweights the InfoNCE loss by adjusting the relative loss for the correctly-matched or mismatched video-text instances identified by KNN, respectively. We focus on exploiting the results of KNN distribution for calibrating the training of retrieval models. Specifically, we apply the negative log-likelihood of \\( p_{knn}^{f} \\) in Eq 5 and \\( p_{knn}^{c} \\) in Eq. 7 as the calibration factor:" + }, + { + "type": "equation", + "bbox": [ + 0.31, + 0.602, + 0.785, + 0.627 + ], + "angle": 0, + "content": "\\[\nF _ {t 2 v} ^ {f} = - \\log p _ {k n n} ^ {f} \\left(v _ {i} \\mid t _ {i}\\right), \\quad F _ {v 2 t} ^ {f} = - \\log p _ {k n n} ^ {f} \\left(t _ {i} \\mid v _ {i}\\right); \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.31, + 0.624, + 0.695, + 0.638 + ], + "angle": 0, + "content": "\\[\nF _ {t 2 v} ^ {c} = - \\log p _ {k n n} ^ {c} \\left(v _ {i} \\mid t _ {i}\\right), \\quad F _ {v 2 t} ^ {c} = - \\log p _ {k n n} ^ {c} \\left(t _ {i} \\mid v _ {i}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.784, + 0.672 + ], + "angle": 0, + "content": "Finally, we can optimize the retrieval model by maximizing the following knowledge-guided cross-modal contrastive loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.238, + 0.681, + 0.785, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v 2 t} ^ {*} = \\left(1 + \\lambda_ {1} F _ {v 2 t} ^ {f} + \\lambda_ {2} F _ {v 2 t} ^ {c}\\right) \\mathcal {L} _ {v 2 t}; \\quad \\mathcal {L} _ {t 2 v} ^ {*} = \\left(1 + \\lambda_ {1} F _ {t 2 v} ^ {f} + \\lambda_ {2} F _ {t 2 v} ^ {c}\\right) \\mathcal {L} _ {t 2 v}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.707, + 0.784, + 0.753 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are hyper-parameters to determine the proportion of each loss term. Hence, in the M-step, the complete training objective of KDProR can be formulated as \\(\\mathcal{L} = \\mathcal{L}_{v2t}^{*} + \\mathcal{L}_{t2v}^{*}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.772, + 0.602, + 0.788 + ], + "angle": 0, + "content": "3.6 Applications of Our KDProR Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In this section, as illustrated in Figure 3, we present three distinct applications for constructing open-world knowledge stores, including pre-trained knowledge, entity-relation structured knowledge, and CoT knowledge within LLMs." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.145, + 0.744, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.279, + 0.785, + 0.321 + ], + "angle": 0, + "content": "Fig. 3: The illustration of proposed open-world knowledge stores, including (1) video-text pretraining knowledge, (2) entity-relation structure knowledge, and (3) knowledge within LLMs prompted by the CoT technique." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.347, + 0.784, + 0.392 + ], + "angle": 0, + "content": "Video-Text Pretraining. As shown in Figure 3, we utilize CLIP-ViP [59], which is further pre-trained on HD-VILA-100M [58] dataset, to extract pretraining video-text knowledge for constructing our knowledge stores." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.785, + 0.452 + ], + "angle": 0, + "content": "Entity-Relation Structure. We attempt to parse captions to scene graphs \\(\\mathcal{G}\\) through the Scene Graph Parser [51], then employ \\(\\mathcal{G}\\) to obtain subject-relation-object triplets. Subsequently, we use these triplets as substitutes for the original captions and then employ CLIP to extract text-video features to construct stores." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.785, + 0.513 + ], + "angle": 0, + "content": "CoT Generation of LLMs. In this work, we attempt to use a strategy similar to CoT [50] to generate auxiliary captions for each video frame by prompting LLMs. Subsequently, we utilize CLIP to extract neural representations of auxiliary captions and corresponding video frames to construct knowledge stores." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.548, + 0.431, + 0.563 + ], + "angle": 0, + "content": "3.7 Theoretical Analysis" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.57, + 0.785, + 0.65 + ], + "angle": 0, + "content": "From a probabilistic perspective, the optimization goal of the VTR model \\(\\Theta\\) trained on dataset \\(\\mathcal{D} = \\{x_{i}\\}_{i}^{N} = \\{(T_{i},V_{i})\\}_{i}^{N}\\) is to maximize the log-likelihood \\(\\mathcal{L}(\\Theta)\\) for text-video pairs. We consider the knowledge introduced from fine-grained and coarse-grained stores as two related latent variables \\(z_{i}^{f}\\) and \\(z_{i}^{c}\\). In a closed world, the posterior distribution of \\(z_{i}^{f}\\) and \\(z_{i}^{c}\\) can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.656, + 0.785, + 0.677 + ], + "angle": 0, + "content": "\\[\nQ \\left(z _ {i} ^ {f}\\right) = p \\left(z _ {i} ^ {f}; x _ {i}, \\Theta\\right), \\quad Q \\left(z _ {i} ^ {c}\\right) = p \\left(z _ {i} ^ {c}; z _ {i} ^ {f}, x _ {i}, \\Theta\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.682, + 0.784, + 0.713 + ], + "angle": 0, + "content": "Thus we can rewrite the objective function of the retrieval model \\(\\Theta\\) according to Jensen's inequality [19] as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.303, + 0.72, + 0.785, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\Theta^ {*} = \\arg \\max \\sum_ {i} \\sum_ {z _ {i} ^ {f}, z _ {i} ^ {c}} Q \\left(z _ {i} ^ {f}\\right) Q \\left(z _ {i} ^ {c}\\right) \\log p \\left(x _ {i}, z _ {i} ^ {f}, z _ {i} ^ {c}; \\Theta\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.785, + 0.841 + ], + "angle": 0, + "content": "To sum up, the E-step and K-step aim to estimate the posterior probability \\( Q(z_{i}^{f}) \\) and \\( Q(z_{i}^{c}) \\) on fine-grained and coarse-grained knowledge stores with given current \\( \\Theta^{(t)} \\). Then KDProR achieves efficient alignment of videos and texts and obtains \\( \\Theta^{(t + 1)} \\) under knowledge guidance at M-step by maximizing the log-likelihood in Eq. 12 with known \\( Q(z_{i}^{f}) \\) and \\( Q(z_{i}^{c}) \\), and satisfies the Theorem 1:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.382, + 0.129 + ], + "angle": 0, + "content": "X. Zhuang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.144, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Results on MSR-VTT without any post-processing (e.g., [3] and [2]). \\( {}^{ + } \\) and \\( {}^{ \\top } \\) are closed-set settings and denote constructing knowledge stores on the training set using the original CLIP and the fine-tuned CLIP upon the training set, respectively. *, † and ‡ denote pre-trained, structural and CoT knowledge in Section 3.6, respectively." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.213, + 0.789, + 0.453 + ], + "angle": 0, + "content": "
BackboneMethodText-to-Video RetrievalVideo-to-Text Retrieval
R@1↑R@5↑R@10↑MdR↓MnR↓R@1↑R@5↑R@10↑MdR↓MnR↓
Non-CLIPCE [31]20.948.862.46.028.220.650.364.05.325.1
MMT [12]26.657.169.64.024.027.057.569.73.721.3
Support [42]27.456.367.73.0-26.655.167.53.0-
Frozen [1]31.059.670.53.0------
CLIP ViT-B/32CLIP4clip [35]44.571.481.62.015.342.770.980.62.011.6
X-pool [13]46.972.882.22.014.3-----
EMCL [21]46.873.183.12.0-46.573.583.52.0-
TS2-Net [33]47.074.583.82.013.045.374.183.72.09.2
DRL [48]47.474.683.82.0-45.373.983.32.0-
STAN [28]46.972.882.82.0------
UATVR [10]47.573.983.52.012.946.973.883.82.08.6
PromptSwitch [6]47.873.982.2-14.146.074.384.8-8.5
KDProR+48.474.484.22.012.247.174.684.42.08.9
KDProR+48.774.484.12.012.047.274.684.42.08.7
KDProR*49.274.284.42.012.047.174.084.12.08.9
KDProR†49.074.684.32.012.047.374.383.92.08.7
KDProR‡49.675.184.42.011.648.274.284.22.08.1
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.482, + 0.79, + 0.514 + ], + "angle": 0, + "content": "Theorem 1 The retrieval model \\(\\Theta\\) optimized iteratively through the proposed EKM algorithm of step \\(t\\) satisfies the following properties:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.519, + 0.663, + 0.534 + ], + "angle": 0, + "content": "- Monotonic Increasing: \\(\\forall t \\geq 0\\), \\(\\mathcal{L}(\\Theta^{(t + 1)}) \\geq \\mathcal{L}(\\Theta^{(t)})\\);" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.534, + 0.692, + 0.551 + ], + "angle": 0, + "content": "- Convergence: \\(\\forall \\epsilon > 0, \\exists \\delta > 0\\), if \\(t > \\delta\\), \\(|\\Theta^{(t + 1)} - \\Theta^{(t)}| \\leq \\epsilon\\)." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.519, + 0.692, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Theorem 1 indicates that our EKM algorithm can ensure the model is positively optimized in each iteration, while also maintaining the convergence properties of vanilla EM algorithms." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.635, + 0.377, + 0.653 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.665, + 0.44, + 0.682 + ], + "angle": 0, + "content": "4.1 Experiments Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on four benchmarks for video-text retrieval tasks including: MSR-VTT [55], DiDeMo [16], LSMDC [45] and ActivityNet [15]. MSR-VTT [55] comprises 10,000 videos with 20 captions each. We train on 9,000 videos and their captions, testing on 1k-A with 1,000 videotext pairs. DiDeMo [16] provides 10,000 videos annotated with 40,000 sentences. We concatenate all descriptions of a video into one paragraph for evaluation. LSMDC [45] includes 118,081 movie-extracted videos, with our dataset partitioned into 109,673 for training, 7,408 for validation, and 1,000 for testing. ActivityNet [15] includes 20,000 YouTube videos. Following [35], we concatenate all captions similar to DiDeMo and evaluate on the 'val1' split." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Text-to-Video retrieval results on DiDeMo, ActivityNet and LSMDC datasets without any post-processing (e.g., [3] and [2]). \\(^+\\) and \\(\\top\\) are closed-set settings and \\(*\\), \\(\\dagger\\) and \\(\\ddagger\\) are open-world settings, whose meanings are all consistent with those in Table 1." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.199, + 0.789, + 0.392 + ], + "angle": 0, + "content": "
MethodDiDeMoActivityNetLSMDC
R@1↑R@5↑R@10↑MdR↓R@1↑R@5↑R@10↑MdR↓R@1↑R@5↑R@10↑MdR↓
Text-to-Video Retrieval using CLIP (ViT-B/32)
CLIP4clip [35]43.470.280.62.040.572.483.62.022.641.049.111.0
X-CLIP [36]45.274.0--44.374.1--23.343.0--
TS2-Net [33]41.871.682.02.041.073.684.52.023.442.350.99.0
DRL [48]47.973.882.72.044.274.586.12.024.945.755.37.0
PromptSwitch [10]--------23.141.750.5-
UATVR [10]43.171.882.32.0--------
DiffusionRet [22]46.774.782.72.045.875.686.32.024.443.154.38.0
\\(KDProR^{+}\\)48.876.184.12.045.875.485.02.026.145.957.37.0
\\(KDProR^{\\top}\\)49.275.383.62.046.476.285.72.026.547.857.86.0
\\(KDProR^{*}\\)51.778.384.91.048.376.886.12.026.950.258.35.0
\\(KDProR^{\\dagger}\\)50.576.385.11.047.576.185.72.027.650.658.15.0
\\(KDProR^{\\ddagger}\\)53.279.285.01.049.177.986.82.028.250.659.25.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.789, + 0.664 + ], + "angle": 0, + "content": "Metrics. Following [35, 48], we utilize recall at rank K (R@K), median rank (MdR) and mean rank (MnR) as metrics to validate the effectiveness of KDProR. Implementation Details. Following [48], we adopt CLIP's ViT-B/32 as the visual encoder and fine-tune the model with 5 epochs on all training sets. We adopt Adam [24] as the optimizer with a cosine warm-up method [34]. Following [35, 48], the frame and caption length are 12 and 32 for MSR-VTT and LSMDC, 64 and 64 for ActivityNet and DiDeMo. The initial learning rate for vision and text encoder is set to \\(10^{-7}\\) and other modules are \\(10^{-4}\\). The hyperparameter \\(\\rho\\) in Eq. 2 is set to 0.2 on MSR-VTT and LSMDC, 0.1 on ActivityNet and DiDeMo for optimal performance. \\(\\beta\\) in Eq. 4 are set to 0.2, and \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) in Eq. 9 are set to 0.2 and 0.3, respectively. We use grid search to obtain the optimal parameters. We uniformly set the top-5 relevant knowledge for KNN retrieval, and the size of the coarse-grained knowledge store \\(K^{c}\\) is set to 256. We adopt the gpt-3.5-turbo-1106 [39] to obtain the CoT knowledge in Section 3.6. All experiments are conducted on 2 Tesla A100 GPUs (80G) and 8 RTX 4090 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.69, + 0.373, + 0.705 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.786, + 0.763 + ], + "angle": 0, + "content": "Comparisons to State-of-the-arts. In this section, we compare KDProR with state-of-the-art methods on four datasets, as results shown in Table 1 and 2. Analysis of Table 1 and 2 yields several insights:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "(1) Our KDProR, employing multi-grained knowledge stores constructed from vanilla and fine-tuned CLIP, can achieve performance gains of \\(1\\%\\) and \\(1.3\\%\\) R@1 on MSR-VTT, \\(1.6\\%\\) and \\(2.2\\%\\) R@1 on ActivityNet compared to DRL [48], respectively. This indicates that KDProR can achieve a harmonious balance between memorization and optimization, thereby improving performance." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.382, + 0.129 + ], + "angle": 0, + "content": "X. Zhuang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 3: Comparisons to Other Baseline Methods in Text-to-Video Retrieval. \\(^\\ddagger\\) denotes the utilization of open-world CoT knowledge in Section 3.6." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.186, + 0.789, + 0.29 + ], + "angle": 0, + "content": "
Baseline MethodsKnowledge SourceMSR-VTTLSMDCDiDeMo
R@1↑R@5↑MnR↓R@1↑R@5↑MnR↓R@1↑R@5↑MnR↓
Data AugmentationCaptions from GPT47.374.712.126.446.943.949.776.314.2
Feature ConcatenatePretrained CLIP-ViP47.874.511.926.146.144.550.376.313.4
Cap4Video [52]Captions from GPT49.374.312.0---52.079.410.5
TEFAL [18]Audio Signal49.475.912.026.846.144.4---
Our KDProR‡Captions from GPT49.675.111.628.250.634.953.279.210.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.306, + 0.789, + 0.349 + ], + "angle": 0, + "content": "Table 4: Text-to-Video retrieval results on MSR-VTT using our re-implemented CLIP4clip [35] and EMCL [21] as base models. \\(^+\\) and \\(\\top\\) are closed-set settings, *, \\(^\\dagger\\) and \\(\\ddagger\\) are open-world settings, whose meanings are consistent with those in Table 1." + }, + { + "type": "table", + "bbox": [ + 0.253, + 0.36, + 0.751, + 0.475 + ], + "angle": 0, + "content": "
MethodCLIP4clip [35] as Base ModelEMCL [21] as Base Model
R@1↑R@5↑R@10↑MdR↓MnR↓R@1↑R@5↑R@10↑MdR↓MnR↓
Base Model43.670.681.02.016.146.673.582.72.013.6
+ KDProR+45.573.482.72.014.147.873.583.12.013.8
+ KDProRT46.474.082.92.013.848.174.083.52.013.7
+ KDProR*47.074.482.92.013.648.773.683.62.013.7
+ KDProR†46.974.082.82.013.648.473.783.82.013.3
+ KDProR‡47.374.683.32.013.449.074.883.82.012.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.787, + 0.596 + ], + "angle": 0, + "content": "(2) The integration of additional open-world knowledge significantly bolsters the predictive capabilities of retrieval models. KDProR of employing three open-world knowledge (i.e., pre-trained, structural, and CoT knowledge) in Section 3.6 achieves superior performance, e.g., \\(+1.8\\%\\), \\(+1.6\\%\\) and \\(+2.2\\%\\) R@1 on MSR-VTT, \\(3.8\\%\\), \\(2.6\\%\\) and \\(5.3\\%\\) R@1 on DiDeMo compared to DRL [48], respectively. We attribute the performance gain to incorporating open-set knowledge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.787, + 0.658 + ], + "angle": 0, + "content": "(3) KDProR demonstrates superior performance compared to all baselines across all benchmarks. This demonstrates that our KDProR offers a twofold advantage: it can unify and incorporate diverse additional knowledge and balance memorization and retrieval to significantly improve performance." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.505, + 0.787, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.811 + ], + "angle": 0, + "content": "Comparisons to Other Baseline Methods. We further compare our method with other baselines in Table 3, including Cap4Video [52] and TEFAL [18]. In addition, we compare it with two baseline methods that introduce open-world knowledge, namely using auxiliary captions generated by LLM as data augmentation and directly concatenating the features generated by CLIP-VIP with the original text-video features. The results indicate that introducing open-world knowledge can enhance performance, and our KDProR achieves better performance gains than Data Augmentation and Feature Concatenate. Impressively, KDProR significantly outperforms state-of-the-art open-set baselines, e.g., \\(+1.4\\%\\) R@1 by Cap4Video on LSMDC and \\(+1.2\\%\\) R@1 by TEFAL on DiDeMo." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.811, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Generalization Analysis. We further equip KDProR with two powerful base models in addition to DRL, i.e., CLIP4clip [35] and EMCL [21], and evaluate" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 5: Ablation experiments on MSR-VTT 9k dataset. \\( {}^{ \\ddagger } \\) denotes the utilization of open-world CoT knowledge in Section 3.6." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.185, + 0.789, + 0.347 + ], + "angle": 0, + "content": "
G. MethodText-to-Video RetrievalVideo-to-Text Retrieval
R@1↑R@5↑R@10↑MnR↓R@1↑R@5↑R@10↑MnR↓
w/o all Knowledge Stores (baseline)46.974.082.412.545.874.084.38.8
w/o Fine-grained Knowledge Store48.174.483.312.146.873.584.08.5
w/o Coarse-grained Knowledge Store48.474.583.711.947.273.884.08.4
w/o the whole E-step47.773.983.412.046.273.584.18.9
w/o Knowledge Interpolation E-step48.273.683.712.047.473.983.88.7
w/o OT Alignment Correction in E-step48.974.684.211.846.673.483.68.9
w/o the whole K-step47.973.883.112.347.073.684.08.7
w/o Fine-grained Calibration in K-step48.774.583.811.847.673.884.08.5
w/o Coarse-grained Calibration in K-step48.574.383.311.847.373.983.68.5
4 Our Full KDProR‡49.675.184.411.648.274.284.28.1
" + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.361, + 0.501, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.362, + 0.744, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.492, + 0.785, + 0.521 + ], + "angle": 0, + "content": "Fig. 4: Ablation experiment on MSR-VTT under closed-world settings. Effect of (a) the number K of top-K knowledge; (b) the knowledge interpolation coefficient \\(\\rho\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.6 + ], + "angle": 0, + "content": "on MSR-VTT. Table 4 shows that KDProR can be applied to successfully boost all baselines as a plug-and-play module under both closed-set or open-world settings. The significant improvements prove the generalization ability of KDProR." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.627, + 0.502, + 0.643 + ], + "angle": 0, + "content": "4.3 Ablation Study and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Effect of the Multi-Grained Knowledge Store. As shown in Groups 1 and 4 in Table 5, removing either fine-grained or coarse-grained knowledge stores will result in a significant decrease in retrieval performance, e.g., \\(-1.5\\%\\) and \\(-1.2\\%\\) R@1 text-to-video retrieval on MSR-VTT, respectively. When all stores are removed, the model will lose the ability to decouple knowledge and introduce additional knowledge, resulting in a significant drop in performance. This verifies the effectiveness of our multi-grained knowledge decoupling strategy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Effect of the Knowledge Injection in E-step. As shown in Groups 2 and 4 in Table 5, we evaluate the effect of the knowledge injection in E-step (i.e., Eq. 2) by removing different components. We can observe that when the knowledge interpolation strategy is removed, the performance drops sharply, indicating that our strategy can effectively inject knowledge from stores. We further remove the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.129 + ], + "angle": 0, + "content": "X. Zhuang et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.144, + 0.775, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.243, + 0.785, + 0.273 + ], + "angle": 0, + "content": "Fig. 5: The text-video results on the MSR-VTT 1K-A test set. \\(\\top\\) and \\(\\ddagger\\) are close- and open-world settings respectively, whose meanings are consistent with those in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.302, + 0.784, + 0.331 + ], + "angle": 0, + "content": "OT alignment correction strategy (i.e., Eq. 4) and observe a significant decrease in video-to-text retrieval performance, proving our strategy's effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.785, + 0.452 + ], + "angle": 0, + "content": "Effect of the Distribution Calibration in K-step. We conduct experiments to study the effect of the KNN distribution calibration strategy as illustrated in Table 5 with Groups 3 and 4. It can be seen that the performance of KDProR significantly decreases without KNN distribution to calibrate the retrieval distribution in Eq. 9. Moreover, the lack of distribution calibration (fine-grained or coarse-grained) in the K-step can also affect retrieval performance. This verifies that the multi-grained distribution calibration strategy can calibrate the retrieval distribution to inject additional knowledge sufficiently and effectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.785, + 0.528 + ], + "angle": 0, + "content": "Effect of the number K of top-K Knowledge. In Figure 4 (left), we show the effect of the number K of top-K knowledge. On the one hand, we find that fewer K (\\(K < 5\\)) means less knowledge is selected, which limits our KDProR's generalization ability for different knowledge. On the other hand, a larger K (\\(K > 5\\)) requires more retrieval time, which increases the cost of learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.785, + 0.664 + ], + "angle": 0, + "content": "Effect of the Knowledge Interpolation Coefficient. In Figure 4 (right), we show the effect of the knowledge interpolation coefficient \\(\\rho\\) on KDProR. We can observe that KDProR achieves the optimal balance between knowledge injection and optimization when \\(\\rho = 0.2\\). Deviations from this optimal \\(\\rho\\), either towards a smaller or larger number, result in performance degradation, which is attributed to inadequate knowledge acquisition and insufficient optimization, respectively. Cases Study. We provide two cases of videos retrieved by KDProR and DRL [48]. As illustrated in Figure 5, KDProR successfully retrieves ground-truth videos in both closed- and open-set settings, whereas DRL fails to do so." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.688, + 0.36, + 0.705 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this work, we explore a novel knowledge-decoupling VTR paradigm that utilizes multi-grained knowledge stores to unify closed- and open-world knowledge. This strategy allows VTR models to achieve an efficient balance between memorization and retrieval optimization, while also opening up a unified interface for injecting various open-world knowledge. We further propose a principled EKM algorithm to achieve an efficient injection of knowledge. Our method could be applied to other visual-language tasks, e.g., visual question answering, and injecting various open-world knowledge for augmenting cross-modal understanding." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.207 + ], + "angle": 0, + "content": "We would like to thank all reviewers for their insightful comments. This paper was partially supported by NSFC (No: 62176008)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.23, + 0.323, + 0.246 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.785, + 0.302 + ], + "angle": 0, + "content": "1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1728-1738 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.302, + 0.785, + 0.343 + ], + "angle": 0, + "content": "2. Bogolin, S.V., Croitoru, I., Jin, H., Liu, Y., Albanie, S.: Cross modal retrieval with querybank normalisation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5194-5205 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.344, + 0.785, + 0.385 + ], + "angle": 0, + "content": "3. Cheng, X., Lin, H., Wu, X., Yang, F., Shen, D.: Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.785, + 0.413 + ], + "angle": 0, + "content": "4. Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems 26 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.413, + 0.785, + 0.44 + ], + "angle": 0, + "content": "5. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the em-algorithm plus discussions on the paper. In:\"\" (1977)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.785, + 0.481 + ], + "angle": 0, + "content": "6. Deng, C., Chen, Q., Qin, P., Chen, D., Wu, Q.: Prompt switch: Efficient clip adaptation for text-video retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15648-15658 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.785, + 0.523 + ], + "angle": 0, + "content": "7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: North American Chapter of the Association for Computational Linguistics (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.579 + ], + "angle": 0, + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv abs/2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.579, + 0.785, + 0.62 + ], + "angle": 0, + "content": "9. Elangovan, A., He, J., Verspoor, K.M.: Memorization vs. generalization : Quantifying data leakage in nlp performance evaluation. In: Conference of the European Chapter of the Association for Computational Linguistics (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.661 + ], + "angle": 0, + "content": "10. Fang, B., Liu, C., Zhou, Y., Yang, M., Song, Y., Li, F., Wang, W., Ji, X., Ouyang, W., et al.: Uatvr: Uncertainty-adaptive text-video retrieval. arXiv preprint arXiv:2301.06309 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "1. Feldman, V.: Does learning require memorization? a short tale about a long tail. Proceedings of the 52nd Annual ACM SIGACT Symposium on Theory of Computing (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.744 + ], + "angle": 0, + "content": "2. Gabeur, V., Sun, C., Alahari, K., Schmid, C.: Multi-modal transformer for video retrieval. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 214-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.799 + ], + "angle": 0, + "content": "3. Gorti, S.K., Vouitsis, N., Ma, J., Golestan, K., Volkovs, M., Garg, A., Yu, G.: X-pool: Cross-modal language-video attention for text-video retrieval. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5006-5015 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "4. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 770-778 (2015)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.26, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.382, + 0.129 + ], + "angle": 0, + "content": "X. Zhuang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "15. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 961-970 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "16. Hendricks, L.A., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.C.: Localizing moments in video with natural language. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 5804-5813 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.787, + 0.259 + ], + "angle": 0, + "content": "17. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Computation 9, 1735-1780 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.259, + 0.787, + 0.3 + ], + "angle": 0, + "content": "18. Ibrahimi, S., Sun, X., Wang, P., Garg, A., Sanan, A., Omar, M.: Audio-enhanced text-to-video retrieval using text-conditioned feature alignment. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 12020-12030 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.3, + 0.787, + 0.328 + ], + "angle": 0, + "content": "19. Jensen, J.L.W.V.: Sur les fonctions convexes et les inégalités entre les valeurs moyennes. Acta Mathematica 30, 175-193 (1906)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "20. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q.V., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. ArXiv abs/2102.05918 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.424 + ], + "angle": 0, + "content": "21. Jin, P., Huang, J., Liu, F., Wu, X., Ge, S., Song, G., Clifton, D., Chen, J.: Expectation-maximization contrastive learning for compact video-and-language representations. Advances in Neural Information Processing Systems 35, 30291-30306 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.424, + 0.787, + 0.467 + ], + "angle": 0, + "content": "22. Jin, P., Li, H., Cheng, Z.L., Li, K., Ji, X., Liu, C., ming Yuan, L., Chen, J.: Diffusionret: Generative text-video retrieval with diffusion model. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 2470-2481 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.509 + ], + "angle": 0, + "content": "23. Khandelwal, U., Levy, O., Jurafsky, D., Zettlemoyer, L., Lewis, M.: Generalization through memorization: Nearest neighbor language models. arXiv preprint arXiv:1911.00172 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.536 + ], + "angle": 0, + "content": "24. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. CoRR abs/1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.591 + ], + "angle": 0, + "content": "25. Li, H., Cao, M., Cheng, X., Li, Y., Zhu, Z., Zou, Y.: G2l: Semantically aligned and uniform video grounding via geodesic and game theory. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 11998-12008 (2023), https://api-semanticscholar.org/CorpusID:260164585" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.787, + 0.633 + ], + "angle": 0, + "content": "26. Li, H., Cao, M., Cheng, X., Zhu, Z., Li, Y., Zou, Y.: Exploiting auxiliary caption for video grounding. In: AAAI Conference on Artificial Intelligence (2023), https://apisemantic scholar.org/CorpusID:257772084" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.675 + ], + "angle": 0, + "content": "27. Li, K., Wang, Y., Li, Y., Wang, Y., He, Y., Wang, L., Qiao, Y.: Unmasked teacher: Towards training-efficient video foundation models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 19891-19903 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.716 + ], + "angle": 0, + "content": "28. Liu, R., Huang, J., Li, G., Feng, J., Wu, X., Li, T.H.: Revisiting temporal modeling for clip-based image-to-video knowledge transferring. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6555-6564 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.758 + ], + "angle": 0, + "content": "29. Liu, S., Fan, H., Qian, S., Chen, Y., Ding, W., Wang, Z.: Hit: Hierarchical transformer with momentum contrast for video-text retrieval. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 11895-11905 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "30. Liu, S., Fan, H., Qian, S., Chen, Y., Ding, W., Wang, Z.: Hit: Hierarchical transformer with momentum contrast for video-text retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11915-11925 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "31. Liu, Y., Albanie, S., Nagrani, A., Zisserman, A.: Use what you have: Video retrieval using representations from collaborative experts. arXiv preprint arXiv:1907.13487 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "32. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. ArXiv abs/1907.11692 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "33. Liu, Y., Xiong, P., Xu, L., Cao, S., Jin, Q.: Ts2-net: Token shift and selection transformer for text-video retrieval. In: European Conference on Computer Vision. pp. 319-335. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.786, + 0.259 + ], + "angle": 0, + "content": "34. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv: Learning (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.786, + 0.3 + ], + "angle": 0, + "content": "35. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing 508, 293-304 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.786, + 0.342 + ], + "angle": 0, + "content": "36. Ma, Y., Xu, G., Sun, X., Yan, M., Zhang, J., Ji, R.: X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. Proceedings of the 30th ACM International Conference on Multimedia (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.786, + 0.37 + ], + "angle": 0, + "content": "37. Meng, Y., Zong, S., Li, X., Sun, X., Zhang, T., Wu, F., Li, J.: Gnn-lm: Language modeling based on global contexts via gnn. ArXiv abs/2110.08743 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.786, + 0.411 + ], + "angle": 0, + "content": "38. Mikolov, T., Chen, K., Corrado, G.S., Dean, J.: Efficient estimation of word representations in vector space. In: International Conference on Learning Representations (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.786, + 0.48 + ], + "angle": 0, + "content": "39. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L.E., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.J.: Training language models to follow instructions with human feedback. ArXiv abs/2203.02155 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.786, + 0.535 + ], + "angle": 0, + "content": "40. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.786, + 0.563 + ], + "angle": 0, + "content": "41. Papernot, N., McDaniel, P.: Deep k-nearest neighbors: Towards confident, interpretable and robust deep learning. arXiv preprint arXiv:1803.04765 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.786, + 0.605 + ], + "angle": 0, + "content": "42. Patrick, M., Huang, P.Y., Asano, Y., Metze, F., Hauptmann, A., Henriques, J., Vedaldi, A.: Support-set bottlenecks for video-text representation learning. arXiv preprint arXiv:2010.02824 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.786, + 0.66 + ], + "angle": 0, + "content": "43. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.786, + 0.715 + ], + "angle": 0, + "content": "44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.786, + 0.758 + ], + "angle": 0, + "content": "45. Rohrbach, A., Rohrbach, M., Tandon, N., Schiele, B.: A dataset for movie description. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3202-3212 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.786, + 0.799 + ], + "angle": 0, + "content": "46. Su, B., Hua, G.: Order-preserving Wasserstein distance for sequence matching. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1049-1057 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.786, + 0.841 + ], + "angle": 0, + "content": "47. Tänzer, M., Ruder, S., Rei, M.: Memorisation versus generalisation in pre-trained language models. In: Annual Meeting of the Association for Computational Linguistics (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.382, + 0.129 + ], + "angle": 0, + "content": "X. Zhuang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "48. Wang, Q., Zhang, Y., Zheng, Y., Pan, P., Hua, X.: Disentangled representation learning for text-video retrieval. ArXiv abs/2203.07111 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.178, + 0.785, + 0.218 + ], + "angle": 0, + "content": "49. Wang, Z., Sung, Y.L., Cheng, F., Bertasius, G., Bansal, M.: Unified coarse-to-fine alignment for video-text retrieval. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 2804-2815 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.22, + 0.784, + 0.261 + ], + "angle": 0, + "content": "50. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.263, + 0.785, + 0.317 + ], + "angle": 0, + "content": "51. Wu, H., Mao, J., Zhang, Y., Jiang, Y., Li, L., Sun, W., Ma, W.Y.: Unified visual-semantic embeddings: Bridging vision and language with structured meaning representations. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6602-6611 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.319, + 0.785, + 0.359 + ], + "angle": 0, + "content": "52. Wu, W., Luo, H., Fang, B., Wang, J., Ouyang, W.: Cap4video: What can auxiliary captions do for text-video retrieval? 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 10704-10713 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.361, + 0.785, + 0.413 + ], + "angle": 0, + "content": "53. Wu, X., Li, H., Luo, Y., Cheng, X., Zhuang, X., Cao, M., Fu, K.: Uncertainty-aware sign language video retrieval with probability distribution modeling. ArXiv abs/2405.19689 (2024), https://api.sementicscholar.org/CorpusID: 270123137" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.417, + 0.785, + 0.445 + ], + "angle": 0, + "content": "54. Xie, Y., Zhu, Z., Zhuang, X., Liang, L., Wang, Z., Zou, Y.: Gpa: Global and prototype alignment for audio-text retrieval. In: Interspeech (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.446, + 0.785, + 0.488 + ], + "angle": 0, + "content": "55. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5288-5296 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.489, + 0.785, + 0.543 + ], + "angle": 0, + "content": "56. Xu, Z., Chen, Z., Zhang, Y., Song, Y., Wan, X., Li, G.: Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17503-17512 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.545, + 0.785, + 0.613 + ], + "angle": 0, + "content": "57. Xu, Z., Huang, J., Liu, T., Liu, Y., Han, H., Yuan, K., Li, X.: Enhancing fine-grained multi-modal alignment via adapters: A parameter-efficient training framework for referring image segmentation. In: 2nd Workshop on Advancing Neural Network Training: Computational Efficiency, Scalability, and Resource Optimization (WANT@ ICML 2024) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.615, + 0.785, + 0.671 + ], + "angle": 0, + "content": "58. Xue, H., Hang, T., Zeng, Y., Sun, Y., Liu, B., Yang, H., Fu, J., Guo, B.: Advancing high-resolution video-language representation with large-scale video transcriptions. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5026-5035 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.672, + 0.785, + 0.712 + ], + "angle": 0, + "content": "59. Xue, H., Sun, Y., Liu, B., Fu, J., Song, R., Li, H., Luo, J.: Clip-vip: Adapting pre-trained image-text model to video-language alignment. In: International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.714, + 0.785, + 0.755 + ], + "angle": 0, + "content": "60. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. Trans. Mach. Learn. Res. 2022 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.757, + 0.785, + 0.784 + ], + "angle": 0, + "content": "61. Yu, Y., Kim, J., Kim, G.: A joint sequence fusion model for video question answering and retrieval. In: European Conference on Computer Vision (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "62. Yuan, L., Chen, D., Chen, Y.L., Codella, N.C.F., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., Liu, C., Liu, M., Liu, Z., Lu, Y., Shi, Y., Wang, L., Wang, J., Xiao, B., Xiao, Z., Yang, J., Zeng, M., Zhou, L., Zhang, P.: Florence: A new foundation model for computer vision. ArXiv abs/2111.11432 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.205 + ], + "angle": 0, + "content": "63. Zhao, Y., Li, K., Cheng, Z., Qiao, P., Zheng, X., Ji, R., Liu, C., Yuan, L., Chen, J.: Graco: Granularity-controllable interactive segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3501-3510 (June 2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.788, + 0.259 + ], + "angle": 0, + "content": "64. Zhao, Y., Lv, W., Xu, S., Wei, J., Wang, G., Dang, Q., Liu, Y., Chen, J.: Detrs beat yolos on real-time object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16965-16974 (June 2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.788, + 0.288 + ], + "angle": 0, + "content": "65. Zheng, X., Jiang, J.: An empirical study of memorization in nlp. In: Annual Meeting of the Association for Computational Linguistics (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.288 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/KDProR_ A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval/8778d22d-a6f9-4fed-bc5f-fddacba7cf0b_origin.pdf b/2024/KDProR_ A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval/8778d22d-a6f9-4fed-bc5f-fddacba7cf0b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d880f6cd35c762ba3f7bf15d601874f908f894c1 --- /dev/null +++ b/2024/KDProR_ A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval/8778d22d-a6f9-4fed-bc5f-fddacba7cf0b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae66befd273fb2e17aac3f8a76e72639ee2614733adbdb2627dc893c80e408e7 +size 1095444 diff --git a/2024/KDProR_ A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval/full.md b/2024/KDProR_ A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dbfdd03fd868298c1d66068e4ba5e3633c7fcf99 --- /dev/null +++ b/2024/KDProR_ A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval/full.md @@ -0,0 +1,340 @@ +# KDProR: A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval + +Xianwei Zhuang*, Hongxiang Li*, Xuxin Cheng, Zhihong Zhu, Yuxin Xie, Yuexian Zou† + +ADSPLAB, School of ECE, Peking University, China {xwzhuang, lihongxiang}@stu.pku.edu.cn, zouyx@pku.edu.cn + +Abstract. Existing video-text retrieval methods predominantly focus on designing diverse cross-modal interaction mechanisms between captions and videos. However, those approaches diverge from human learning paradigms, where humans possess the capability to seek and associate knowledge from an open set, rather than rote memorizing all text-video instances. Motivated by this, we attempt to decouple knowledge from retrieval models through multi-grained knowledge stores and identify two significant benefits of our knowledge-decoupling strategy: (1) it ensures a harmonious balance between knowledge memorization and retrieval optimization, thereby improving retrieval performance; and (2) it can promote incorporating diverse open-world knowledge to augment video-text retrieval. To efficiently integrate information from knowledge stores, we further introduce a novel retrieval framework termed KDProR, which utilizes our proposed Expectation-Knowledge-Maximization (EKM) algorithm for optimization. Specifically, in E-step, KDProR obtains relevant contextual semantics from knowledge stores and achieves efficient knowledge injection through interpolation and alignment correction. During the K-step, KDProR calculates the knowledge KNN distribution by indexing the top-K acquired knowledge to refine the retrieval distribution, and in M-step, KDProR optimizes the retrieval model by maximizing the likelihood of the objective. Extensive experiments on various benchmarks prove that KDProR significantly outperforms previous state-of-the-art methods across all metrics. Remarkably, KDProR can uniformly and efficiently incorporate diverse open-world knowledge and is compatible with different interaction mechanisms and architectures. + +Keywords: Video Retrieval $\cdot$ Knowledge Decoupling $\cdot$ Open World + +# 1 Introduction + +Video-text retrieval (VTR) is a significant and challenging task in cross-modal interaction [1,12,31,42], aiming to align video (text) candidates with text (video) queries to identify the most relevant instances accurately [61]. The standard + +![](images/a382c3ae873f7e8bea857310715212eaf1e2152084b31d3febae47a8e86298ca.jpg) +Fig. 1: The illustration of different VTR paradigms. (a) denotes vanilla VTR methods; (b) introduces additional data only during training. (c) achieves "open examination" with knowledge decoupling, thereby enhancing VTR in both closed and open sets. + +paradigm of VTR for addressing cross-modal matching involves aligning videotext features acquired through visual and textual encoders via contrastive learning. With the advent of large-scale image-language pre-trained models [20, 44, 60, 62], recent VTR methods [3, 13, 35] have realized substantial advancements in retrieval performance by leveraging pre-trained image-language models, such as CLIP [44]. Benefiting from powerful CLIP, most of these methods focus on devising various cross-modal interaction strategies to enhance visual-text alignment, including cross-attention [13], multi-grained interaction [36], disentangled representation learning [48], etc, and have achieved considerable success. + +However, if we conceptualize the training set as a "book" intuitively, this training-test paradigm for VTR can be likened to "video-text memorization" and "closed-book VTR examination", which diverges from the human learning process [37, 65]. Humans possess the capability to seek and associate knowledge from an open set, rather than relying on rote memorization of all text-video instances [47]. This "closed-book examination" approach may be further hindered by dense memorization overhead [37] and challenges in memorizing difficult samples [9, 11], echoing the proverb: "The palest ink is better than the best memory." Furthermore, previous open-set fine-tuning [18, 52] and pre-training [27, 59] methods have shown that incorporating auxiliary data from the open world can improve VTR performance. These insights encourage us to explore a novel VTR paradigm: "open-set examination", which involves constructing knowledge stores from the closed training set or open-world sources to decouple knowledge. As shown in Figure 1, this strategy introduces relevant knowledge from stores as reinforcement signals to help the model strike a balance between generalization and memory. However, this also brings two new challenges: (1) How to decouple knowledge from models and unify closed- and open-world knowledge; (2) How to efficiently incorporate decoupled knowledge to improve retrieval performance. + +In this work, we propose a novel knowledge-decoupling probabilistic framework for VTR (KDProR), tackling the challenges with the following strategies: + +(1) KDProR utilizes multi-grained stores to decouple knowledge and standardize closed- and open-world knowledge sources. We con + +struct multi-scale knowledge stores composed of local and global neural representations to decouple knowledge. This approach allows rare patterns to be memorized explicitly, rather than implicitly in model parameters. In addition, knowledge stores can be updated asynchronously and endowed with knowledge from closed- or open-world sources. In this work, we explore three different sources of additional knowledge, including spatio-temporal knowledge from pre-trained models, entity-relation structured knowledge and outputs of large language models (LLMs, e.g., GPT [40]) prompted by chain-of-thoughts (CoT) [50]. Moreover, we identify two core advantages of the knowledge decoupling strategy: it (a) ensures a harmonious balance between memorization and retrieval optimization; and (b) opens up a unified interface for injecting diverse open-world knowledge. + +(2) KDProR utilizes a novel Expectation-Knowledge-Maximization algorithm to achieve efficient knowledge injection and retrieval optimization. We extend the traditional Expectation-Maximization (EM) algorithm [5] to the Expectation-Knowledge-Maximization (EKM) algorithm for optimizing KDProR. Specifically, in the E-step, KDProR obtains relevant contextual semantics and achieves efficient knowledge injection through interpolation and alignment correction. In the K-step, we calculate the k-nearest neighbor (KNN) distribution by indexing the top-K knowledge obtained to calibrate the original retrieval distribution. And in the M-step, we optimize the model by maximizing the likelihood of the target. Additionally, we provide a theoretical analysis of our KDProR framework from a probabilistic perspective in Section 3.7. + +Extensive experiments on four benchmarks, i.e., MSR-VTT [55], DiDeMo [16], LSMDC [45] and ActivityNet [15], prove that KDProR significantly outperforms previous state-of-the-art methods in both closed and open-world settings. In summary, our KDProR has several appealing facets: (1) Effectiveness: It can significantly improve retrieval performance under both closed-set and open-set settings. (2) Universality: It can be used to introduce various open-world knowledge, while being compatible with various interaction modules and pretrained foundation models (e.g., [27, 59]). (3) Explainability: The effectiveness and convergence of retrieval optimization via our EKM algorithm are theoretically guaranteed. The main contributions of this paper are presented as follows: + +(1) We explore a novel knowledge-decoupling paradigm for VTR and construct multi-grained knowledge stores to uniformly introduce knowledge from closed sets and diverse open-world sources. (2) We propose a novel VTR framework termed KDProR, which utilizes a principled EKM algorithm to inject various knowledge into the model to improve retrieval performance. Theoretical analysis provides an underlying insight into the effectiveness and convergence of KDProR. (3) Experiments conducted on four benchmarks show that KDProR achieves new state-of-the-art results under both closed- and open-world settings. + +# 2 Related work + +Feature Representation for Video-Text Retrieval. Recent studies [1, 12, 31, 42] on VTR employ a bi-encoder architecture to extract text and video rep + +resentation. The selection of text and visual encoders changes with the development of research on feature extraction upon NLP and CV, e.g., Word2Vec [38] and BERT-based models [7, 32] for textual representation, and CNN-based [14] and ViT-based models [8, 63, 64] for visual tasks. These researches use independently pre-trained dual encoders for visual-textual semantic extraction [1, 12, 42, 57]. Recently, CLIP [43] pre-trained on a large-scale dataset with 400 million pairs has achieved excellent capability of cross-modal representation. And CLIP-based VTR methods [3, 13, 35, 53] have rapidly developed and achieved great success. Our KDProR also benefits from existing pre-training, while exploring a novel perspective for enhancing representation, i.e., knowledge decoupling. + +Interaction Mechanism for Video-Text Retrieval. Many VTR methods explore different cross-modal interaction mechanisms for cosine similarity [35] to improve retrieval performance. Some studies [35] seek to enhance the spatiotemporal understanding ability of image encoders through LSTM [17] or temporal Transformers. MoE-based [3] and Cross-Transformers-based [13,54] methods significantly improve the cross-modal interaction. [25, 26, 29, 36, 49, 56] explore hierarchical and multi-grained interactions, respectively. DRL [48] proposes disentangled representation learning for cross-modal interaction. Our KDProR is compatible with these cross-modal interaction strategies. + +Open-World Knowledge for Video-Text Retrieval. The previous methods mainly introduce open-world knowledge in the pre-training and fine-tuning stages to enhance VTR performance. CLIP-ViP [59] conducts secondary pretraining of CLIP on large-scale video-text datasets, and UMT [27] achieves efficient pre-training to improve the spatio-temporal understanding and inject open-set knowledge into VTR tasks. TEFAL [18] utilizes additional audio features to enhance VTR retrieval performance and Cap4Video [52] utilizes auxiliary captions obtained by Web Search or generated by LLMs to enhance cross-modal interaction. KDProR is compatible with these pre-training methods, and we do not specify the type of open-world knowledge to be introduced, but instead achieve universal open-set knowledge injection through knowledge decoupling. + +# 3 Methods + +# 3.1 Settings and Feature Extraction. + +Following previous work [35, 48], we utilize the dual-encoder architecture of CLIP [44] to achieve video and text feature extraction on training set $\mathcal{D} = \{x_i\}_i^N = \{(T_i,V_i)\}_i^N$ , where $T_{i}$ and $V_{i}$ represent text and video instances respectively, and $N$ is the size of the training set. Given a text query $T_{i}$ , we initially add identifiers [CLS] and [SEP] to the sentence, and leverage the text encoder of CLIP to encode the text representation: $t_i = \mathrm{TextEncoder}(T_i)$ . For each video $V_{i}$ , we uniformly select $n_v$ frames $[F_1,\dots ,F_{n_v}]$ as keyframes, and employ the transformer-based encoder of CLIP to extract sequential features $v_{i} = \mathrm{VideoEncoder}([F_{1},\dots ,F_{n_{v}}])$ . + +Extensive work explores various cross-modal interaction strategies to improve retrieval performance [30,35,48]. Our work focuses on decoupling knowledge from + +![](images/906932314491d8f437564da914c9aa9f23a0d7e03a0b1830842f4c20535591a2.jpg) +Fig. 2: The illustration of the proposed framework, consisting of (1) constructing a multi-grained knowledge store using open-world knowledge or closed-world knowledge (training dataset); (2) EKM algorithm that sequentially implements knowledge injection, distribution calibration, and target likelihood maximization. + +models and exploring the impact of different knowledge sources on retrieval performance, rather than pre-training or interaction strategies. Thus we adopt the weighted token-wise interaction module of DRL [48] to achieve cross-modal interaction. Note that KDProR is also compatible with other interaction strategies. + +# 3.2 Decoupling Knowledge Store + +Our KDProR involves augmenting CLIP-based retrieval models with semantic information from decoupled knowledge stores. Thus, the first step of our methods is to build a multi-grained knowledge store. As illustrated in Figure 2, our stores contain fine-grained and coarse-grained caches, where each key and value (text-video) are representations of knowledge. + +Fine-grained Knowledge Store. The fine-grained knowledge store can be represented as $\mathcal{S}^f = \{\mathcal{S}_i^f\}_{i=1}^{K^f} = \{(\mathcal{K}_i^f, \mathcal{V}_i^f)\}_{i=1}^{K^f}$ , where $\mathcal{S}_i^f = (\mathcal{K}_i^f, \mathcal{V}_i^f)$ denotes the $i$ -th key-value pair in the fine-grained knowledge store, and $\mathcal{K}_i^f$ is the size of our fine-grained knowledge store. Under the closed world setting, we utilize text and visual encoders of CLIP to obtain the feature key-value pairs on the training dataset as $\mathcal{K}_i^f = t_i$ and $\mathcal{V}_i^f = v_i$ . + +Coarse-grained Knowledge Store. The coarse-grained knowledge store is designed to encapsulate global semantics. It equips retrieval models with global, coarse-grained insights from the knowledge store, enhancing the comprehensive knowledge understanding. Similarly, the coarse-grained storage can be represented as: $\mathcal{S}^c = \{\mathcal{S}_i^c\}_{i=1}^{K^c} = \{(\mathcal{K}_i^c, \mathcal{V}_i^c)\}_{i=1}^{K^c}$ , where $\mathcal{S}_i^c = (\mathcal{K}_i^c, \mathcal{V}_i^c)$ denotes the $i$ -th key-value pair in the coarse-grained knowledge store, and $\mathcal{K}_i^c$ is the size of our coarse-grained knowledge store. We adopt the K-mean strategy to extract knowledge from fine-grained storage into more advanced representations: + +$$ +\mathcal {K} _ {i} ^ {c} = \phi \left(c _ {i} ^ {t}\right), \forall c _ {i} ^ {t} \in \mathcal {C} ^ {t}, \quad \mathcal {V} _ {i} ^ {c} = \phi \left(c _ {i} ^ {v}\right), \forall c _ {i} ^ {v} \in \mathcal {C} ^ {v}, \tag {1} +$$ + +where, $\mathcal{C}^t = \{c_1^t,\dots ,c_{K^c}^t\}$ denotes the cluster set obtained by K-Mean over fine-grained knowledge $\{\mathcal{K}_i^f\}$ , $\mathcal{C}^v = \{c_1^v,\dots ,c_{K^c}^v\}$ represents the cluster set obtained by K-Mean over $\{\mathcal{V}_i^f\}$ , and $\phi (\cdot)$ represents the Max-Pooling aggregate function. + +Note that our framework is universal, and we will introduce more applications for constructing diverse knowledge stores in Section 3.6. + +# 3.3 E-step for Multi-Grained Knowledge Injection + +Knowledge Interpolation. In E-step, KDProR initially attempts to inject fine-grained and coarse-grained knowledge into the model by knowledge interpolation of features. We treat with $t_i$ and $v_i$ as anchors to retrieve the fine-grained knowledge store, and obtain two sets of top-K similar vector pairs $\mathcal{N}_{t_i}^f$ and $\mathcal{N}_{v_i}^f$ , respectively. $\mathcal{N}_{t_i}^f = \{(\mathcal{K}_{t_i}^f, \mathcal{V}_{t_i}^f)\}$ and $\mathcal{N}_{v_i}^f = \{(\mathcal{K}_{v_i}^f, \mathcal{V}_{v_i}^f)\}$ are two sets of key-value knowledge pairs, where $\mathcal{K}_{t_i}^f$ and $\mathcal{V}_{v_i}^f$ are the nearest neighbors of $t_i$ and $v_i$ , respectively. After the model retrieves the top-K candidates for each $t_i$ and $v_i$ , their corresponding representation $\mathcal{K}_{t_i}^f$ and $\mathcal{V}_{v_i}^f$ obtained from knowledge-store will be incorporated into original features $t_i$ and $v_i$ to act as demonstration learning. Specifically, we adopt the nearest neighbor sets to inject decoupled knowledge into text and video features as: + +$$ +t _ {i} ^ {f} = \rho t _ {i} + (1 - \rho) \frac {\sum_ {(\mathcal {K} , \mathcal {V}) \in \mathcal {N} _ {t _ {i}} ^ {f}} \mathcal {K}}{| \mathcal {N} _ {t _ {i}} ^ {f} |}, v _ {i} ^ {f} = \rho v _ {i} + (1 - \rho) \frac {\sum_ {(\mathcal {K} , \mathcal {V}) \in \mathcal {N} _ {v _ {i}} ^ {f}} \mathcal {V}}{| \mathcal {N} _ {v _ {i}} ^ {f} |}, \tag {2} +$$ + +where, $\rho$ is a hyperparameter used to express the interpolation ratio. Subsequently, we utilize the weighted token-wise interaction module [48] to enhance textual and visual representations injected with additional knowledge. + +However, the textual and visual knowledge within the top-K sets, obtained by $t_i$ and $v_i$ respectively, may not correspond on a one-to-one basis due to their derivation from two independent top-K operations. To mitigate the issue of mismatched video-text features during knowledge injection, we utilize the optimal transport (OT) algorithm to facilitate alignment correction within a batch. + +Alignment Correction via Optimal Transport. We first define $\langle \mathbf{Q},\mathbf{S}^f\rangle =$ $\mathrm{tr}(\mathbf{Q}^\top \mathbf{S}^f)\in \mathbb{R}^{B\times B}$ as the similarity matrix calculated from the updated text and video pairs $\{(t_i^f,v_i^f)\}_{i = 1}^B$ in a batch. $\mathbf{Q}\in \mathbb{R}^{B\times B}$ represents the corresponding transport assignment matrix, where $\mathbf{Q}[i,j]$ denotes the probabilities of aligning $t_i^f$ with $v_{i}^{f}$ . The optimal transport aims to establish flexible alignment between videos and captions by maximizing global text-video similarity $\mathrm{tr}(\mathbf{Q}^\top \mathbf{S}^f)$ . The objective of optimal transport can be formulated as: + +$$ +\max _ {\mathbf {Q} \in \mathcal {Q}} \langle \mathbf {Q}, \mathbf {S} ^ {f} \rangle + \varepsilon H (\mathbf {Q}) \quad \text {s . t .} \mathcal {Q} = \left\{\mathbf {Q} \in \mathbb {R} ^ {B \times B} \mid \mathbf {Q} \mathbf {1} _ {B} = \boldsymbol {\mu}, \mathbf {Q} ^ {\top} \mathbf {1} _ {B} = \boldsymbol {\nu} \right\} \tag {3} +$$ + +where, $\pmb{\mu} \in \mathbb{R}^{B}$ and $\pmb{\nu} \in \mathbb{R}^{B}$ indicate the relative importance of each video and caption, $\mathbf{1}_{B}$ represents the vector of ones in dimension $B$ , $H(\mathbf{Q})$ denotes a entropy regular term [4] and $\varepsilon$ controls its smoothness. Following [46], we initialize $\pmb{\mu}$ and $\pmb{\nu}$ to a uniform distribution $\mathbf{1}_{B}$ . The optimal transport problem + +in Eq. 3 can be solved using the Sinkhorn-Knopp algorithm [46] to obtain the optimal assignment scheme $\mathbf{Q}^*$ . Based on this, we can obtain the similarity matrix after adding alignment correction $\mathbf{Q}^*$ as follows: + +$$ +\mathbf {S} ^ {f ^ {*}} = \left(\left(1 - \beta\right) \mathbf {I} + \beta \mathbf {Q} ^ {*}\right) \mathbf {S} ^ {f}, \tag {4} +$$ + +where, $\mathbf{I}$ is the identity matrix and $\beta$ is a hyperparameter. + +We can implement the coarse-grained knowledge injection and obtain the realigned similarity matrix $\mathbf{S}^{c*}$ in a similar way. + +# 3.4 K-step for Retrieval Distribution Calibration + +Existing studies [23, 41] have shown that non-parametric KNN classification or nearest-neighbor interpolation methods have extraordinary potential in improving robustness and generalization. It is intuitively to leverage the KNN's retrieval results as the prior knowledge to guide the parameters training for video-text pairs that are difficult to establish deep correlations. In practice, we aggregate the prediction of KNN into the original probability distribution of video-text retrieval to achieve the multi-grained calibration of retrieval distribution. + +Fine-grained Distribution Calibration. We utilize the set of k-nearest neighbors $\mathcal{N}_{t_i}^f$ and $\mathcal{N}_{v_i}^f$ retrieved by querying the open-book knowledge store in E-step as fine-grained top-K knowledge. Subsequently, we calculate the KNN distribution according to the softmax of the similarity and aggregation probability mass of each video-text pair in the retrieved target as: + +$$ +p _ {k n n} ^ {f} \left(v _ {i} \mid t _ {i}\right) = \frac {\sum_ {\left(\mathcal {K} , \mathcal {V}\right) \in N _ {t _ {i}} ^ {f}} \varphi \left(v _ {i} , \mathcal {V}\right) e ^ {d \left(v _ {i} , \mathcal {V}\right)}}{\sum_ {\left(\mathcal {K} , \mathcal {V}\right) \in N _ {t _ {i}} ^ {f}} \varphi \left(v _ {i} , \mathcal {V}\right)}, \quad \varphi \left(v _ {i}, \mathcal {V}\right) = \left\{ \begin{array}{l l} 1, & \mathcal {V} \in N _ {v _ {i}} ^ {f} \\ 0, & \mathcal {V} \notin N _ {v _ {i}} ^ {f} \end{array} , \right. \tag {5} +$$ + +where, $d(\cdot ,\cdot)\in [-1,1]$ denotes the cosine similarity function, and $\varphi (v_{i},\mathcal{V})$ is an indicator function used to determine whether $\mathcal{V}$ is within the top-K nearest neighbor set $N_{v_i}^f$ of the $v_{i}$ . + +We can further obtain the KNN probability distribution $p_{knn}^{f}(t_{i} \mid v_{i})$ for video-to-text retrieval in a completely symmetrical manner. + +Coarse-grained Distribution Calibration. Consistent with Eq. 5, we employ the KNN algorithm to obtain top-K knowledge, utilizing closed-world or open-world knowledge from the coarse-grained knowledge store to refine the probability distribution function pertinent to the text-to-video retrieval. Nevertheless, since the coarse-grained knowledge store is derived from the distillation of fine-grained knowledge, it is impracticable to apply indicator functions (cf. Eq. 5) to ascertain the presence of an inclusion relation between textual features and their nearest neighbor sets. To tackle this challenge, we introduce a novel indicator function, denoted as $\phi(\cdot, \cdot)$ , parameterized by a threshold $\alpha$ . $\phi(\cdot, \cdot)$ is designed to evaluate whether $\mathcal{V}$ resides within the top-K nearest neighbor set $N_{v_i}^c$ of the $v_i$ via the cosine similarity function $d(\cdot, \cdot)$ , and is formulated as: + +$$ +\phi \left(v _ {i}, \mathcal {V}\right) = \left\{ \begin{array}{l l} 1, & d \left(\mathcal {V}, \mathcal {V} ^ {c}\right) < \alpha , \exists \left(\mathcal {K} ^ {c}, \mathcal {V} ^ {c}\right) \in N _ {v _ {i}} ^ {c} \\ 0, & o t h e r w i s e s \end{array} . \right. \tag {6} +$$ + +Based on this, we can obtain a closed-form expression for the KNN distribution used for coarse-grained distribution calibration as follows: + +$$ +p _ {k n n} ^ {c} \left(v _ {i} \mid t _ {i}\right) = \frac {\sum_ {\left(\mathcal {K} , \mathcal {V}\right) \in N _ {t _ {i}} ^ {c}} \phi \left(v _ {i} , \mathcal {V}\right) e ^ {d \left(v _ {i} , \mathcal {V}\right)}}{\sum_ {\left(\mathcal {K} , \mathcal {V}\right) \in N _ {t _ {i}} ^ {c}} \phi \left(v _ {i} , \mathcal {V}\right)}. \tag {7} +$$ + +We can further obtain the KNN probability distribution $p_{knn}^{c}(t_{i} \mid v_{i})$ for coarse-grained video-to-text retrieval in a completely symmetrical manner. + +# 3.5 M-step for Unified Retrieval Optimization + +In the M-step, we optimize the whole model by maximizing the retrieval likelihood objective over the realigned similarity matrix. Specifically, we utilize the re-aligned multi-grained similarity matrix in Eq. 4 obtained from E-step to calculate the InfoNCE loss: + +$$ +\mathcal {L} _ {t 2 v} = - \frac {1}{B} \sum_ {i} ^ {B} \left(\log \frac {\exp \left(\mathbf {S} ^ {f _ {i i} ^ {*}}\right)}{\sum_ {j = 1} ^ {B} \exp \left(\mathbf {S} ^ {f _ {i j} ^ {*}}\right)} + \log \frac {\exp \left(\mathbf {S} _ {i i} ^ {c ^ {*}}\right)}{\sum_ {j = 1} ^ {B} \exp \left(\mathbf {S} _ {i j} ^ {c ^ {*}}\right)}\right), \tag {8} +$$ + +$$ +\mathcal {L} _ {v 2 t} = - \frac {1}{B} \sum_ {i} ^ {B} \left(\log \frac {\exp \left(\mathbf {S} _ {i i} ^ {f ^ {*}}\right)}{\sum_ {j = 1} ^ {B} \exp \left(\mathbf {S} _ {j i} ^ {f ^ {*}}\right)} + \log \frac {\exp \left(\mathbf {S} _ {i i} ^ {c ^ {*}}\right)}{\sum_ {j = 1} ^ {B} \exp \left(\mathbf {S} _ {j i} ^ {c ^ {*}}\right)}\right), +$$ + +where $B$ is the batch size. Subsequently, we propose utilizing the KNN distribution (cf. Eq. 5 and 7) to guide the training process. The KNN calibrator reweights the InfoNCE loss by adjusting the relative loss for the correctly-matched or mismatched video-text instances identified by KNN, respectively. We focus on exploiting the results of KNN distribution for calibrating the training of retrieval models. Specifically, we apply the negative log-likelihood of $p_{knn}^{f}$ in Eq 5 and $p_{knn}^{c}$ in Eq. 7 as the calibration factor: + +$$ +F _ {t 2 v} ^ {f} = - \log p _ {k n n} ^ {f} \left(v _ {i} \mid t _ {i}\right), \quad F _ {v 2 t} ^ {f} = - \log p _ {k n n} ^ {f} \left(t _ {i} \mid v _ {i}\right); \tag {9} +$$ + +$$ +F _ {t 2 v} ^ {c} = - \log p _ {k n n} ^ {c} \left(v _ {i} \mid t _ {i}\right), \quad F _ {v 2 t} ^ {c} = - \log p _ {k n n} ^ {c} \left(t _ {i} \mid v _ {i}\right). +$$ + +Finally, we can optimize the retrieval model by maximizing the following knowledge-guided cross-modal contrastive loss as: + +$$ +\mathcal {L} _ {v 2 t} ^ {*} = \left(1 + \lambda_ {1} F _ {v 2 t} ^ {f} + \lambda_ {2} F _ {v 2 t} ^ {c}\right) \mathcal {L} _ {v 2 t}; \quad \mathcal {L} _ {t 2 v} ^ {*} = \left(1 + \lambda_ {1} F _ {t 2 v} ^ {f} + \lambda_ {2} F _ {t 2 v} ^ {c}\right) \mathcal {L} _ {t 2 v}, \tag {10} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are hyper-parameters to determine the proportion of each loss term. Hence, in the M-step, the complete training objective of KDProR can be formulated as $\mathcal{L} = \mathcal{L}_{v2t}^{*} + \mathcal{L}_{t2v}^{*}$ . + +# 3.6 Applications of Our KDProR Framework + +In this section, as illustrated in Figure 3, we present three distinct applications for constructing open-world knowledge stores, including pre-trained knowledge, entity-relation structured knowledge, and CoT knowledge within LLMs. + +![](images/c4524fa7d5fdadbc17b735a712b83c4ad55e4d985731291e859026e15c0e9f2c.jpg) +Fig. 3: The illustration of proposed open-world knowledge stores, including (1) video-text pretraining knowledge, (2) entity-relation structure knowledge, and (3) knowledge within LLMs prompted by the CoT technique. + +Video-Text Pretraining. As shown in Figure 3, we utilize CLIP-ViP [59], which is further pre-trained on HD-VILA-100M [58] dataset, to extract pretraining video-text knowledge for constructing our knowledge stores. + +Entity-Relation Structure. We attempt to parse captions to scene graphs $\mathcal{G}$ through the Scene Graph Parser [51], then employ $\mathcal{G}$ to obtain subject-relation-object triplets. Subsequently, we use these triplets as substitutes for the original captions and then employ CLIP to extract text-video features to construct stores. + +CoT Generation of LLMs. In this work, we attempt to use a strategy similar to CoT [50] to generate auxiliary captions for each video frame by prompting LLMs. Subsequently, we utilize CLIP to extract neural representations of auxiliary captions and corresponding video frames to construct knowledge stores. + +# 3.7 Theoretical Analysis + +From a probabilistic perspective, the optimization goal of the VTR model $\Theta$ trained on dataset $\mathcal{D} = \{x_{i}\}_{i}^{N} = \{(T_{i},V_{i})\}_{i}^{N}$ is to maximize the log-likelihood $\mathcal{L}(\Theta)$ for text-video pairs. We consider the knowledge introduced from fine-grained and coarse-grained stores as two related latent variables $z_{i}^{f}$ and $z_{i}^{c}$ . In a closed world, the posterior distribution of $z_{i}^{f}$ and $z_{i}^{c}$ can be represented as: + +$$ +Q \left(z _ {i} ^ {f}\right) = p \left(z _ {i} ^ {f}; x _ {i}, \Theta\right), \quad Q \left(z _ {i} ^ {c}\right) = p \left(z _ {i} ^ {c}; z _ {i} ^ {f}, x _ {i}, \Theta\right). \tag {11} +$$ + +Thus we can rewrite the objective function of the retrieval model $\Theta$ according to Jensen's inequality [19] as follows: + +$$ +\Theta^ {*} = \arg \max \sum_ {i} \sum_ {z _ {i} ^ {f}, z _ {i} ^ {c}} Q \left(z _ {i} ^ {f}\right) Q \left(z _ {i} ^ {c}\right) \log p \left(x _ {i}, z _ {i} ^ {f}, z _ {i} ^ {c}; \Theta\right). \tag {12} +$$ + +To sum up, the E-step and K-step aim to estimate the posterior probability $Q(z_{i}^{f})$ and $Q(z_{i}^{c})$ on fine-grained and coarse-grained knowledge stores with given current $\Theta^{(t)}$ . Then KDProR achieves efficient alignment of videos and texts and obtains $\Theta^{(t + 1)}$ under knowledge guidance at M-step by maximizing the log-likelihood in Eq. 12 with known $Q(z_{i}^{f})$ and $Q(z_{i}^{c})$ , and satisfies the Theorem 1: + +Table 1: Results on MSR-VTT without any post-processing (e.g., [3] and [2]). ${}^{ + }$ and ${}^{ \top }$ are closed-set settings and denote constructing knowledge stores on the training set using the original CLIP and the fine-tuned CLIP upon the training set, respectively. *, † and ‡ denote pre-trained, structural and CoT knowledge in Section 3.6, respectively. + +
BackboneMethodText-to-Video RetrievalVideo-to-Text Retrieval
R@1↑R@5↑R@10↑MdR↓MnR↓R@1↑R@5↑R@10↑MdR↓MnR↓
Non-CLIPCE [31]20.948.862.46.028.220.650.364.05.325.1
MMT [12]26.657.169.64.024.027.057.569.73.721.3
Support [42]27.456.367.73.0-26.655.167.53.0-
Frozen [1]31.059.670.53.0------
CLIP ViT-B/32CLIP4clip [35]44.571.481.62.015.342.770.980.62.011.6
X-pool [13]46.972.882.22.014.3-----
EMCL [21]46.873.183.12.0-46.573.583.52.0-
TS2-Net [33]47.074.583.82.013.045.374.183.72.09.2
DRL [48]47.474.683.82.0-45.373.983.32.0-
STAN [28]46.972.882.82.0------
UATVR [10]47.573.983.52.012.946.973.883.82.08.6
PromptSwitch [6]47.873.982.2-14.146.074.384.8-8.5
KDProR+48.474.484.22.012.247.174.684.42.08.9
KDProR+48.774.484.12.012.047.274.684.42.08.7
KDProR*49.274.284.42.012.047.174.084.12.08.9
KDProR†49.074.684.32.012.047.374.383.92.08.7
KDProR‡49.675.184.42.011.648.274.284.22.08.1
+ +Theorem 1 The retrieval model $\Theta$ optimized iteratively through the proposed EKM algorithm of step $t$ satisfies the following properties: + +- Monotonic Increasing: $\forall t \geq 0$ , $\mathcal{L}(\Theta^{(t + 1)}) \geq \mathcal{L}(\Theta^{(t)})$ ; +- Convergence: $\forall \epsilon > 0, \exists \delta > 0$ , if $t > \delta$ , $|\Theta^{(t + 1)} - \Theta^{(t)}| \leq \epsilon$ . + +Theorem 1 indicates that our EKM algorithm can ensure the model is positively optimized in each iteration, while also maintaining the convergence properties of vanilla EM algorithms. + +# 4 Experiments + +# 4.1 Experiments Settings + +Datasets. We conduct experiments on four benchmarks for video-text retrieval tasks including: MSR-VTT [55], DiDeMo [16], LSMDC [45] and ActivityNet [15]. MSR-VTT [55] comprises 10,000 videos with 20 captions each. We train on 9,000 videos and their captions, testing on 1k-A with 1,000 videotext pairs. DiDeMo [16] provides 10,000 videos annotated with 40,000 sentences. We concatenate all descriptions of a video into one paragraph for evaluation. LSMDC [45] includes 118,081 movie-extracted videos, with our dataset partitioned into 109,673 for training, 7,408 for validation, and 1,000 for testing. ActivityNet [15] includes 20,000 YouTube videos. Following [35], we concatenate all captions similar to DiDeMo and evaluate on the 'val1' split. + +Table 2: Text-to-Video retrieval results on DiDeMo, ActivityNet and LSMDC datasets without any post-processing (e.g., [3] and [2]). $^+$ and $\top$ are closed-set settings and $*$ , $\dagger$ and $\ddagger$ are open-world settings, whose meanings are all consistent with those in Table 1. + +
MethodDiDeMoActivityNetLSMDC
R@1↑R@5↑R@10↑MdR↓R@1↑R@5↑R@10↑MdR↓R@1↑R@5↑R@10↑MdR↓
Text-to-Video Retrieval using CLIP (ViT-B/32)
CLIP4clip [35]43.470.280.62.040.572.483.62.022.641.049.111.0
X-CLIP [36]45.274.0--44.374.1--23.343.0--
TS2-Net [33]41.871.682.02.041.073.684.52.023.442.350.99.0
DRL [48]47.973.882.72.044.274.586.12.024.945.755.37.0
PromptSwitch [10]--------23.141.750.5-
UATVR [10]43.171.882.32.0--------
DiffusionRet [22]46.774.782.72.045.875.686.32.024.443.154.38.0
\(KDProR^{+}\)48.876.184.12.045.875.485.02.026.145.957.37.0
\(KDProR^{\top}\)49.275.383.62.046.476.285.72.026.547.857.86.0
\(KDProR^{*}\)51.778.384.91.048.376.886.12.026.950.258.35.0
\(KDProR^{\dagger}\)50.576.385.11.047.576.185.72.027.650.658.15.0
\(KDProR^{\ddagger}\)53.279.285.01.049.177.986.82.028.250.659.25.0
+ +Metrics. Following [35, 48], we utilize recall at rank K (R@K), median rank (MdR) and mean rank (MnR) as metrics to validate the effectiveness of KDProR. Implementation Details. Following [48], we adopt CLIP's ViT-B/32 as the visual encoder and fine-tune the model with 5 epochs on all training sets. We adopt Adam [24] as the optimizer with a cosine warm-up method [34]. Following [35, 48], the frame and caption length are 12 and 32 for MSR-VTT and LSMDC, 64 and 64 for ActivityNet and DiDeMo. The initial learning rate for vision and text encoder is set to $10^{-7}$ and other modules are $10^{-4}$ . The hyperparameter $\rho$ in Eq. 2 is set to 0.2 on MSR-VTT and LSMDC, 0.1 on ActivityNet and DiDeMo for optimal performance. $\beta$ in Eq. 4 are set to 0.2, and $\lambda_{1}$ and $\lambda_{2}$ in Eq. 9 are set to 0.2 and 0.3, respectively. We use grid search to obtain the optimal parameters. We uniformly set the top-5 relevant knowledge for KNN retrieval, and the size of the coarse-grained knowledge store $K^{c}$ is set to 256. We adopt the gpt-3.5-turbo-1106 [39] to obtain the CoT knowledge in Section 3.6. All experiments are conducted on 2 Tesla A100 GPUs (80G) and 8 RTX 4090 GPUs. + +# 4.2 Main Results + +Comparisons to State-of-the-arts. In this section, we compare KDProR with state-of-the-art methods on four datasets, as results shown in Table 1 and 2. Analysis of Table 1 and 2 yields several insights: + +(1) Our KDProR, employing multi-grained knowledge stores constructed from vanilla and fine-tuned CLIP, can achieve performance gains of $1\%$ and $1.3\%$ R@1 on MSR-VTT, $1.6\%$ and $2.2\%$ R@1 on ActivityNet compared to DRL [48], respectively. This indicates that KDProR can achieve a harmonious balance between memorization and optimization, thereby improving performance. + +Table 3: Comparisons to Other Baseline Methods in Text-to-Video Retrieval. $^\ddagger$ denotes the utilization of open-world CoT knowledge in Section 3.6. + +
Baseline MethodsKnowledge SourceMSR-VTTLSMDCDiDeMo
R@1↑R@5↑MnR↓R@1↑R@5↑MnR↓R@1↑R@5↑MnR↓
Data AugmentationCaptions from GPT47.374.712.126.446.943.949.776.314.2
Feature ConcatenatePretrained CLIP-ViP47.874.511.926.146.144.550.376.313.4
Cap4Video [52]Captions from GPT49.374.312.0---52.079.410.5
TEFAL [18]Audio Signal49.475.912.026.846.144.4---
Our KDProR‡Captions from GPT49.675.111.628.250.634.953.279.210.2
+ +Table 4: Text-to-Video retrieval results on MSR-VTT using our re-implemented CLIP4clip [35] and EMCL [21] as base models. $^+$ and $\top$ are closed-set settings, *, $^\dagger$ and $\ddagger$ are open-world settings, whose meanings are consistent with those in Table 1. + +
MethodCLIP4clip [35] as Base ModelEMCL [21] as Base Model
R@1↑R@5↑R@10↑MdR↓MnR↓R@1↑R@5↑R@10↑MdR↓MnR↓
Base Model43.670.681.02.016.146.673.582.72.013.6
+ KDProR+45.573.482.72.014.147.873.583.12.013.8
+ KDProRT46.474.082.92.013.848.174.083.52.013.7
+ KDProR*47.074.482.92.013.648.773.683.62.013.7
+ KDProR†46.974.082.82.013.648.473.783.82.013.3
+ KDProR‡47.374.683.32.013.449.074.883.82.012.9
+ +(2) The integration of additional open-world knowledge significantly bolsters the predictive capabilities of retrieval models. KDProR of employing three open-world knowledge (i.e., pre-trained, structural, and CoT knowledge) in Section 3.6 achieves superior performance, e.g., $+1.8\%$ , $+1.6\%$ and $+2.2\%$ R@1 on MSR-VTT, $3.8\%$ , $2.6\%$ and $5.3\%$ R@1 on DiDeMo compared to DRL [48], respectively. We attribute the performance gain to incorporating open-set knowledge. +(3) KDProR demonstrates superior performance compared to all baselines across all benchmarks. This demonstrates that our KDProR offers a twofold advantage: it can unify and incorporate diverse additional knowledge and balance memorization and retrieval to significantly improve performance. + +Comparisons to Other Baseline Methods. We further compare our method with other baselines in Table 3, including Cap4Video [52] and TEFAL [18]. In addition, we compare it with two baseline methods that introduce open-world knowledge, namely using auxiliary captions generated by LLM as data augmentation and directly concatenating the features generated by CLIP-VIP with the original text-video features. The results indicate that introducing open-world knowledge can enhance performance, and our KDProR achieves better performance gains than Data Augmentation and Feature Concatenate. Impressively, KDProR significantly outperforms state-of-the-art open-set baselines, e.g., $+1.4\%$ R@1 by Cap4Video on LSMDC and $+1.2\%$ R@1 by TEFAL on DiDeMo. + +Generalization Analysis. We further equip KDProR with two powerful base models in addition to DRL, i.e., CLIP4clip [35] and EMCL [21], and evaluate + +Table 5: Ablation experiments on MSR-VTT 9k dataset. ${}^{ \ddagger }$ denotes the utilization of open-world CoT knowledge in Section 3.6. + +
G. MethodText-to-Video RetrievalVideo-to-Text Retrieval
R@1↑R@5↑R@10↑MnR↓R@1↑R@5↑R@10↑MnR↓
w/o all Knowledge Stores (baseline)46.974.082.412.545.874.084.38.8
w/o Fine-grained Knowledge Store48.174.483.312.146.873.584.08.5
w/o Coarse-grained Knowledge Store48.474.583.711.947.273.884.08.4
w/o the whole E-step47.773.983.412.046.273.584.18.9
w/o Knowledge Interpolation E-step48.273.683.712.047.473.983.88.7
w/o OT Alignment Correction in E-step48.974.684.211.846.673.483.68.9
w/o the whole K-step47.973.883.112.347.073.684.08.7
w/o Fine-grained Calibration in K-step48.774.583.811.847.673.884.08.5
w/o Coarse-grained Calibration in K-step48.574.383.311.847.373.983.68.5
4 Our Full KDProR‡49.675.184.411.648.274.284.28.1
+ +![](images/fc88c9de3ce35d1afb69c10a09b10f0d80fbf06b2be8fb6710b16ec85bc6dc2e.jpg) +Fig. 4: Ablation experiment on MSR-VTT under closed-world settings. Effect of (a) the number K of top-K knowledge; (b) the knowledge interpolation coefficient $\rho$ . + +![](images/f2fdab8ca87fac4712c877274d2a5506bfb6b4dd17aa65e41ce06ecaab3386df.jpg) + +on MSR-VTT. Table 4 shows that KDProR can be applied to successfully boost all baselines as a plug-and-play module under both closed-set or open-world settings. The significant improvements prove the generalization ability of KDProR. + +# 4.3 Ablation Study and Analysis + +Effect of the Multi-Grained Knowledge Store. As shown in Groups 1 and 4 in Table 5, removing either fine-grained or coarse-grained knowledge stores will result in a significant decrease in retrieval performance, e.g., $-1.5\%$ and $-1.2\%$ R@1 text-to-video retrieval on MSR-VTT, respectively. When all stores are removed, the model will lose the ability to decouple knowledge and introduce additional knowledge, resulting in a significant drop in performance. This verifies the effectiveness of our multi-grained knowledge decoupling strategy. + +Effect of the Knowledge Injection in E-step. As shown in Groups 2 and 4 in Table 5, we evaluate the effect of the knowledge injection in E-step (i.e., Eq. 2) by removing different components. We can observe that when the knowledge interpolation strategy is removed, the performance drops sharply, indicating that our strategy can effectively inject knowledge from stores. We further remove the + +![](images/f0cbf02afbdb2e22fd8706c4b5e215e083cd9bb2c6a0b4d4b0f7c91e38be7464.jpg) +Fig. 5: The text-video results on the MSR-VTT 1K-A test set. $\top$ and $\ddagger$ are close- and open-world settings respectively, whose meanings are consistent with those in Table 1. + +OT alignment correction strategy (i.e., Eq. 4) and observe a significant decrease in video-to-text retrieval performance, proving our strategy's effectiveness. + +Effect of the Distribution Calibration in K-step. We conduct experiments to study the effect of the KNN distribution calibration strategy as illustrated in Table 5 with Groups 3 and 4. It can be seen that the performance of KDProR significantly decreases without KNN distribution to calibrate the retrieval distribution in Eq. 9. Moreover, the lack of distribution calibration (fine-grained or coarse-grained) in the K-step can also affect retrieval performance. This verifies that the multi-grained distribution calibration strategy can calibrate the retrieval distribution to inject additional knowledge sufficiently and effectively. + +Effect of the number K of top-K Knowledge. In Figure 4 (left), we show the effect of the number K of top-K knowledge. On the one hand, we find that fewer K ( $K < 5$ ) means less knowledge is selected, which limits our KDProR's generalization ability for different knowledge. On the other hand, a larger K ( $K > 5$ ) requires more retrieval time, which increases the cost of learning. + +Effect of the Knowledge Interpolation Coefficient. In Figure 4 (right), we show the effect of the knowledge interpolation coefficient $\rho$ on KDProR. We can observe that KDProR achieves the optimal balance between knowledge injection and optimization when $\rho = 0.2$ . Deviations from this optimal $\rho$ , either towards a smaller or larger number, result in performance degradation, which is attributed to inadequate knowledge acquisition and insufficient optimization, respectively. Cases Study. We provide two cases of videos retrieved by KDProR and DRL [48]. As illustrated in Figure 5, KDProR successfully retrieves ground-truth videos in both closed- and open-set settings, whereas DRL fails to do so. + +# 5 Conclusion + +In this work, we explore a novel knowledge-decoupling VTR paradigm that utilizes multi-grained knowledge stores to unify closed- and open-world knowledge. This strategy allows VTR models to achieve an efficient balance between memorization and retrieval optimization, while also opening up a unified interface for injecting various open-world knowledge. We further propose a principled EKM algorithm to achieve an efficient injection of knowledge. Our method could be applied to other visual-language tasks, e.g., visual question answering, and injecting various open-world knowledge for augmenting cross-modal understanding. + +# Acknowledgements + +We would like to thank all reviewers for their insightful comments. This paper was partially supported by NSFC (No: 62176008). + +# References + +1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1728-1738 (2021) +2. Bogolin, S.V., Croitoru, I., Jin, H., Liu, Y., Albanie, S.: Cross modal retrieval with querybank normalisation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5194-5205 (2022) +3. Cheng, X., Lin, H., Wu, X., Yang, F., Shen, D.: Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290 (2021) +4. Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems 26 (2013) +5. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the em-algorithm plus discussions on the paper. In:"" (1977) +6. Deng, C., Chen, Q., Qin, P., Chen, D., Wu, Q.: Prompt switch: Efficient clip adaptation for text-video retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15648-15658 (2023) +7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: North American Chapter of the Association for Computational Linguistics (2019) +8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv abs/2010.11929 (2020) +9. Elangovan, A., He, J., Verspoor, K.M.: Memorization vs. generalization : Quantifying data leakage in nlp performance evaluation. In: Conference of the European Chapter of the Association for Computational Linguistics (2021) +10. Fang, B., Liu, C., Zhou, Y., Yang, M., Song, Y., Li, F., Wang, W., Ji, X., Ouyang, W., et al.: Uatvr: Uncertainty-adaptive text-video retrieval. arXiv preprint arXiv:2301.06309 (2023) +1. Feldman, V.: Does learning require memorization? a short tale about a long tail. Proceedings of the 52nd Annual ACM SIGACT Symposium on Theory of Computing (2019) +2. Gabeur, V., Sun, C., Alahari, K., Schmid, C.: Multi-modal transformer for video retrieval. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 214-229. Springer (2020) +3. Gorti, S.K., Vouitsis, N., Ma, J., Golestan, K., Volkovs, M., Garg, A., Yu, G.: X-pool: Cross-modal language-video attention for text-video retrieval. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5006-5015 (2022) +4. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 770-778 (2015) + +15. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 961-970 (2015) +16. Hendricks, L.A., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.C.: Localizing moments in video with natural language. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 5804-5813 (2017) +17. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Computation 9, 1735-1780 (1997) +18. Ibrahimi, S., Sun, X., Wang, P., Garg, A., Sanan, A., Omar, M.: Audio-enhanced text-to-video retrieval using text-conditioned feature alignment. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 12020-12030 (2023) +19. Jensen, J.L.W.V.: Sur les fonctions convexes et les inégalités entre les valeurs moyennes. Acta Mathematica 30, 175-193 (1906) +20. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q.V., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. ArXiv abs/2102.05918 (2021) +21. Jin, P., Huang, J., Liu, F., Wu, X., Ge, S., Song, G., Clifton, D., Chen, J.: Expectation-maximization contrastive learning for compact video-and-language representations. Advances in Neural Information Processing Systems 35, 30291-30306 (2022) +22. Jin, P., Li, H., Cheng, Z.L., Li, K., Ji, X., Liu, C., ming Yuan, L., Chen, J.: Diffusionret: Generative text-video retrieval with diffusion model. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 2470-2481 (2023) +23. Khandelwal, U., Levy, O., Jurafsky, D., Zettlemoyer, L., Lewis, M.: Generalization through memorization: Nearest neighbor language models. arXiv preprint arXiv:1911.00172 (2019) +24. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. CoRR abs/1412.6980 (2014) +25. Li, H., Cao, M., Cheng, X., Li, Y., Zhu, Z., Zou, Y.: G2l: Semantically aligned and uniform video grounding via geodesic and game theory. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 11998-12008 (2023), https://api-semanticscholar.org/CorpusID:260164585 +26. Li, H., Cao, M., Cheng, X., Zhu, Z., Li, Y., Zou, Y.: Exploiting auxiliary caption for video grounding. In: AAAI Conference on Artificial Intelligence (2023), https://apisemantic scholar.org/CorpusID:257772084 +27. Li, K., Wang, Y., Li, Y., Wang, Y., He, Y., Wang, L., Qiao, Y.: Unmasked teacher: Towards training-efficient video foundation models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 19891-19903 (2023) +28. Liu, R., Huang, J., Li, G., Feng, J., Wu, X., Li, T.H.: Revisiting temporal modeling for clip-based image-to-video knowledge transferring. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6555-6564 (2023) +29. Liu, S., Fan, H., Qian, S., Chen, Y., Ding, W., Wang, Z.: Hit: Hierarchical transformer with momentum contrast for video-text retrieval. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 11895-11905 (2021) +30. Liu, S., Fan, H., Qian, S., Chen, Y., Ding, W., Wang, Z.: Hit: Hierarchical transformer with momentum contrast for video-text retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11915-11925 (2021) +31. Liu, Y., Albanie, S., Nagrani, A., Zisserman, A.: Use what you have: Video retrieval using representations from collaborative experts. arXiv preprint arXiv:1907.13487 (2019) + +32. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. ArXiv abs/1907.11692 (2019) +33. Liu, Y., Xiong, P., Xu, L., Cao, S., Jin, Q.: Ts2-net: Token shift and selection transformer for text-video retrieval. In: European Conference on Computer Vision. pp. 319-335. Springer (2022) +34. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv: Learning (2016) +35. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing 508, 293-304 (2022) +36. Ma, Y., Xu, G., Sun, X., Yan, M., Zhang, J., Ji, R.: X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. Proceedings of the 30th ACM International Conference on Multimedia (2022) +37. Meng, Y., Zong, S., Li, X., Sun, X., Zhang, T., Wu, F., Li, J.: Gnn-lm: Language modeling based on global contexts via gnn. ArXiv abs/2110.08743 (2021) +38. Mikolov, T., Chen, K., Corrado, G.S., Dean, J.: Efficient estimation of word representations in vector space. In: International Conference on Learning Representations (2013) +39. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L.E., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.J.: Training language models to follow instructions with human feedback. ArXiv abs/2203.02155 (2022) +40. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022) +41. Papernot, N., McDaniel, P.: Deep k-nearest neighbors: Towards confident, interpretable and robust deep learning. arXiv preprint arXiv:1803.04765 (2018) +42. Patrick, M., Huang, P.Y., Asano, Y., Metze, F., Hauptmann, A., Henriques, J., Vedaldi, A.: Support-set bottlenecks for video-text representation learning. arXiv preprint arXiv:2010.02824 (2020) +43. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021) +44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +45. Rohrbach, A., Rohrbach, M., Tandon, N., Schiele, B.: A dataset for movie description. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3202-3212 (2015) +46. Su, B., Hua, G.: Order-preserving Wasserstein distance for sequence matching. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1049-1057 (2017) +47. Tänzer, M., Ruder, S., Rei, M.: Memorisation versus generalisation in pre-trained language models. In: Annual Meeting of the Association for Computational Linguistics (2021) + +48. Wang, Q., Zhang, Y., Zheng, Y., Pan, P., Hua, X.: Disentangled representation learning for text-video retrieval. ArXiv abs/2203.07111 (2022) +49. Wang, Z., Sung, Y.L., Cheng, F., Bertasius, G., Bansal, M.: Unified coarse-to-fine alignment for video-text retrieval. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 2804-2815 (2023) +50. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022) +51. Wu, H., Mao, J., Zhang, Y., Jiang, Y., Li, L., Sun, W., Ma, W.Y.: Unified visual-semantic embeddings: Bridging vision and language with structured meaning representations. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6602-6611 (2019) +52. Wu, W., Luo, H., Fang, B., Wang, J., Ouyang, W.: Cap4video: What can auxiliary captions do for text-video retrieval? 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 10704-10713 (2022) +53. Wu, X., Li, H., Luo, Y., Cheng, X., Zhuang, X., Cao, M., Fu, K.: Uncertainty-aware sign language video retrieval with probability distribution modeling. ArXiv abs/2405.19689 (2024), https://api.sementicscholar.org/CorpusID: 270123137 +54. Xie, Y., Zhu, Z., Zhuang, X., Liang, L., Wang, Z., Zou, Y.: Gpa: Global and prototype alignment for audio-text retrieval. In: Interspeech (2024) +55. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5288-5296 (2016) +56. Xu, Z., Chen, Z., Zhang, Y., Song, Y., Wan, X., Li, G.: Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17503-17512 (2023) +57. Xu, Z., Huang, J., Liu, T., Liu, Y., Han, H., Yuan, K., Li, X.: Enhancing fine-grained multi-modal alignment via adapters: A parameter-efficient training framework for referring image segmentation. In: 2nd Workshop on Advancing Neural Network Training: Computational Efficiency, Scalability, and Resource Optimization (WANT@ ICML 2024) (2024) +58. Xue, H., Hang, T., Zeng, Y., Sun, Y., Liu, B., Yang, H., Fu, J., Guo, B.: Advancing high-resolution video-language representation with large-scale video transcriptions. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5026-5035 (2021) +59. Xue, H., Sun, Y., Liu, B., Fu, J., Song, R., Li, H., Luo, J.: Clip-vip: Adapting pre-trained image-text model to video-language alignment. In: International Conference on Learning Representations (2023) +60. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. Trans. Mach. Learn. Res. 2022 (2022) +61. Yu, Y., Kim, J., Kim, G.: A joint sequence fusion model for video question answering and retrieval. In: European Conference on Computer Vision (2018) +62. Yuan, L., Chen, D., Chen, Y.L., Codella, N.C.F., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., Liu, C., Liu, M., Liu, Z., Lu, Y., Shi, Y., Wang, L., Wang, J., Xiao, B., Xiao, Z., Yang, J., Zeng, M., Zhou, L., Zhang, P.: Florence: A new foundation model for computer vision. ArXiv abs/2111.11432 (2021) + +63. Zhao, Y., Li, K., Cheng, Z., Qiao, P., Zheng, X., Ji, R., Liu, C., Yuan, L., Chen, J.: Graco: Granularity-controllable interactive segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3501-3510 (June 2024) +64. Zhao, Y., Lv, W., Xu, S., Wei, J., Wang, G., Dang, Q., Liu, Y., Chen, J.: Detrs beat yolos on real-time object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16965-16974 (June 2024) +65. Zheng, X., Jiang, J.: An empirical study of memorization in nlp. In: Annual Meeting of the Association for Computational Linguistics (2022) \ No newline at end of file diff --git a/2024/KDProR_ A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval/images.zip b/2024/KDProR_ A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ed99cd61b177bf07efacd092934238c4037148f0 --- /dev/null +++ b/2024/KDProR_ A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab3850135aa77a7ef229fbdb6aea01a195f00c57f7fb31c0b946eac93136f3be +size 724931 diff --git a/2024/KDProR_ A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval/layout.json b/2024/KDProR_ A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a62202bc95b31f86499147b23ec8a02c607e84a6 --- /dev/null +++ b/2024/KDProR_ A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval/layout.json @@ -0,0 +1,11078 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 111, + 477, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 111, + 477, + 146 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 477, + 146 + ], + "type": "text", + "content": "KDProR: A Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 200, + 168, + 412, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 168, + 412, + 193 + ], + "spans": [ + { + "bbox": [ + 200, + 168, + 412, + 193 + ], + "type": "text", + "content": "Xianwei Zhuang*, Hongxiang Li*, Xuxin Cheng, Zhihong Zhu, Yuxin Xie, Yuexian Zou†" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 173, + 202, + 439, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 202, + 439, + 224 + ], + "spans": [ + { + "bbox": [ + 173, + 202, + 439, + 224 + ], + "type": "text", + "content": "ADSPLAB, School of ECE, Peking University, China {xwzhuang, lihongxiang}@stu.pku.edu.cn, zouyx@pku.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 255, + 452, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 255, + 452, + 529 + ], + "spans": [ + { + "bbox": [ + 160, + 255, + 452, + 529 + ], + "type": "text", + "content": "Abstract. Existing video-text retrieval methods predominantly focus on designing diverse cross-modal interaction mechanisms between captions and videos. However, those approaches diverge from human learning paradigms, where humans possess the capability to seek and associate knowledge from an open set, rather than rote memorizing all text-video instances. Motivated by this, we attempt to decouple knowledge from retrieval models through multi-grained knowledge stores and identify two significant benefits of our knowledge-decoupling strategy: (1) it ensures a harmonious balance between knowledge memorization and retrieval optimization, thereby improving retrieval performance; and (2) it can promote incorporating diverse open-world knowledge to augment video-text retrieval. To efficiently integrate information from knowledge stores, we further introduce a novel retrieval framework termed KDProR, which utilizes our proposed Expectation-Knowledge-Maximization (EKM) algorithm for optimization. Specifically, in E-step, KDProR obtains relevant contextual semantics from knowledge stores and achieves efficient knowledge injection through interpolation and alignment correction. During the K-step, KDProR calculates the knowledge KNN distribution by indexing the top-K acquired knowledge to refine the retrieval distribution, and in M-step, KDProR optimizes the retrieval model by maximizing the likelihood of the objective. Extensive experiments on various benchmarks prove that KDProR significantly outperforms previous state-of-the-art methods across all metrics. Remarkably, KDProR can uniformly and efficiently incorporate diverse open-world knowledge and is compatible with different interaction mechanisms and architectures." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 540, + 438, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 540, + 438, + 552 + ], + "spans": [ + { + "bbox": [ + 160, + 540, + 438, + 552 + ], + "type": "text", + "content": "Keywords: Video Retrieval " + }, + { + "bbox": [ + 160, + 540, + 438, + 552 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 540, + 438, + 552 + ], + "type": "text", + "content": " Knowledge Decoupling " + }, + { + "bbox": [ + 160, + 540, + 438, + 552 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 540, + 438, + 552 + ], + "type": "text", + "content": " Open World" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 572, + 230, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 572, + 230, + 586 + ], + "spans": [ + { + "bbox": [ + 133, + 572, + 230, + 586 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 597, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 597, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 482, + 635 + ], + "type": "text", + "content": "Video-text retrieval (VTR) is a significant and challenging task in cross-modal interaction [1,12,31,42], aiming to align video (text) candidates with text (video) queries to identify the most relevant instances accurately [61]. The standard" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 642, + 292, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 642, + 292, + 653 + ], + "spans": [ + { + "bbox": [ + 142, + 642, + 292, + 653 + ], + "type": "text", + "content": "* These authors contributed equally." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 653, + 242, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 653, + 242, + 665 + ], + "spans": [ + { + "bbox": [ + 142, + 653, + 242, + 665 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 147, + 114, + 467, + 228 + ], + "blocks": [ + { + "bbox": [ + 147, + 114, + 467, + 228 + ], + "lines": [ + { + "bbox": [ + 147, + 114, + 467, + 228 + ], + "spans": [ + { + "bbox": [ + 147, + 114, + 467, + 228 + ], + "type": "image", + "image_path": "a382c3ae873f7e8bea857310715212eaf1e2152084b31d3febae47a8e86298ca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 236, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 132, + 236, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 482, + 269 + ], + "type": "text", + "content": "Fig. 1: The illustration of different VTR paradigms. (a) denotes vanilla VTR methods; (b) introduces additional data only during training. (c) achieves \"open examination\" with knowledge decoupling, thereby enhancing VTR in both closed and open sets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 400 + ], + "type": "text", + "content": "paradigm of VTR for addressing cross-modal matching involves aligning videotext features acquired through visual and textual encoders via contrastive learning. With the advent of large-scale image-language pre-trained models [20, 44, 60, 62], recent VTR methods [3, 13, 35] have realized substantial advancements in retrieval performance by leveraging pre-trained image-language models, such as CLIP [44]. Benefiting from powerful CLIP, most of these methods focus on devising various cross-modal interaction strategies to enhance visual-text alignment, including cross-attention [13], multi-grained interaction [36], disentangled representation learning [48], etc, and have achieved considerable success." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 401, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 616 + ], + "type": "text", + "content": "However, if we conceptualize the training set as a \"book\" intuitively, this training-test paradigm for VTR can be likened to \"video-text memorization\" and \"closed-book VTR examination\", which diverges from the human learning process [37, 65]. Humans possess the capability to seek and associate knowledge from an open set, rather than relying on rote memorization of all text-video instances [47]. This \"closed-book examination\" approach may be further hindered by dense memorization overhead [37] and challenges in memorizing difficult samples [9, 11], echoing the proverb: \"The palest ink is better than the best memory.\" Furthermore, previous open-set fine-tuning [18, 52] and pre-training [27, 59] methods have shown that incorporating auxiliary data from the open world can improve VTR performance. These insights encourage us to explore a novel VTR paradigm: \"open-set examination\", which involves constructing knowledge stores from the closed training set or open-world sources to decouple knowledge. As shown in Figure 1, this strategy introduces relevant knowledge from stores as reinforcement signals to help the model strike a balance between generalization and memory. However, this also brings two new challenges: (1) How to decouple knowledge from models and unify closed- and open-world knowledge; (2) How to efficiently incorporate decoupled knowledge to improve retrieval performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 617, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 481, + 641 + ], + "type": "text", + "content": "In this work, we propose a novel knowledge-decoupling probabilistic framework for VTR (KDProR), tackling the challenges with the following strategies:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "(1) KDProR utilizes multi-grained stores to decouple knowledge and standardize closed- and open-world knowledge sources. We con" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 102 + ], + "type": "text", + "content": "X. Zhuang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "struct multi-scale knowledge stores composed of local and global neural representations to decouple knowledge. This approach allows rare patterns to be memorized explicitly, rather than implicitly in model parameters. In addition, knowledge stores can be updated asynchronously and endowed with knowledge from closed- or open-world sources. In this work, we explore three different sources of additional knowledge, including spatio-temporal knowledge from pre-trained models, entity-relation structured knowledge and outputs of large language models (LLMs, e.g., GPT [40]) prompted by chain-of-thoughts (CoT) [50]. Moreover, we identify two core advantages of the knowledge decoupling strategy: it (a) ensures a harmonious balance between memorization and retrieval optimization; and (b) opens up a unified interface for injecting diverse open-world knowledge." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 379 + ], + "type": "text", + "content": "(2) KDProR utilizes a novel Expectation-Knowledge-Maximization algorithm to achieve efficient knowledge injection and retrieval optimization. We extend the traditional Expectation-Maximization (EM) algorithm [5] to the Expectation-Knowledge-Maximization (EKM) algorithm for optimizing KDProR. Specifically, in the E-step, KDProR obtains relevant contextual semantics and achieves efficient knowledge injection through interpolation and alignment correction. In the K-step, we calculate the k-nearest neighbor (KNN) distribution by indexing the top-K knowledge obtained to calibrate the original retrieval distribution. And in the M-step, we optimize the model by maximizing the likelihood of the target. Additionally, we provide a theoretical analysis of our KDProR framework from a probabilistic perspective in Section 3.7." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 380, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 499 + ], + "type": "text", + "content": "Extensive experiments on four benchmarks, i.e., MSR-VTT [55], DiDeMo [16], LSMDC [45] and ActivityNet [15], prove that KDProR significantly outperforms previous state-of-the-art methods in both closed and open-world settings. In summary, our KDProR has several appealing facets: (1) Effectiveness: It can significantly improve retrieval performance under both closed-set and open-set settings. (2) Universality: It can be used to introduce various open-world knowledge, while being compatible with various interaction modules and pretrained foundation models (e.g., [27, 59]). (3) Explainability: The effectiveness and convergence of retrieval optimization via our EKM algorithm are theoretically guaranteed. The main contributions of this paper are presented as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 499, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 499, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 499, + 482, + 596 + ], + "type": "text", + "content": "(1) We explore a novel knowledge-decoupling paradigm for VTR and construct multi-grained knowledge stores to uniformly introduce knowledge from closed sets and diverse open-world sources. (2) We propose a novel VTR framework termed KDProR, which utilizes a principled EKM algorithm to inject various knowledge into the model to improve retrieval performance. Theoretical analysis provides an underlying insight into the effectiveness and convergence of KDProR. (3) Experiments conducted on four benchmarks show that KDProR achieves new state-of-the-art results under both closed- and open-world settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 615, + 234, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 615, + 234, + 627 + ], + "spans": [ + { + "bbox": [ + 132, + 615, + 234, + 627 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Feature Representation for Video-Text Retrieval. Recent studies [1, 12, 31, 42] on VTR employ a bi-encoder architecture to extract text and video rep" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "text", + "content": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "resentation. The selection of text and visual encoders changes with the development of research on feature extraction upon NLP and CV, e.g., Word2Vec [38] and BERT-based models [7, 32] for textual representation, and CNN-based [14] and ViT-based models [8, 63, 64] for visual tasks. These researches use independently pre-trained dual encoders for visual-textual semantic extraction [1, 12, 42, 57]. Recently, CLIP [43] pre-trained on a large-scale dataset with 400 million pairs has achieved excellent capability of cross-modal representation. And CLIP-based VTR methods [3, 13, 35, 53] have rapidly developed and achieved great success. Our KDProR also benefits from existing pre-training, while exploring a novel perspective for enhancing representation, i.e., knowledge decoupling." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 342 + ], + "type": "text", + "content": "Interaction Mechanism for Video-Text Retrieval. Many VTR methods explore different cross-modal interaction mechanisms for cosine similarity [35] to improve retrieval performance. Some studies [35] seek to enhance the spatiotemporal understanding ability of image encoders through LSTM [17] or temporal Transformers. MoE-based [3] and Cross-Transformers-based [13,54] methods significantly improve the cross-modal interaction. [25, 26, 29, 36, 49, 56] explore hierarchical and multi-grained interactions, respectively. DRL [48] proposes disentangled representation learning for cross-modal interaction. Our KDProR is compatible with these cross-modal interaction strategies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 343, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 477 + ], + "type": "text", + "content": "Open-World Knowledge for Video-Text Retrieval. The previous methods mainly introduce open-world knowledge in the pre-training and fine-tuning stages to enhance VTR performance. CLIP-ViP [59] conducts secondary pretraining of CLIP on large-scale video-text datasets, and UMT [27] achieves efficient pre-training to improve the spatio-temporal understanding and inject open-set knowledge into VTR tasks. TEFAL [18] utilizes additional audio features to enhance VTR retrieval performance and Cap4Video [52] utilizes auxiliary captions obtained by Web Search or generated by LLMs to enhance cross-modal interaction. KDProR is compatible with these pre-training methods, and we do not specify the type of open-world knowledge to be introduced, but instead achieve universal open-set knowledge injection through knowledge decoupling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 491, + 208, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 208, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 208, + 504 + ], + "type": "text", + "content": "3 Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 515, + 326, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 326, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 326, + 528 + ], + "type": "text", + "content": "3.1 Settings and Feature Extraction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": "Following previous work [35, 48], we utilize the dual-encoder architecture of CLIP [44] to achieve video and text feature extraction on training set " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{x_i\\}_i^N = \\{(T_i,V_i)\\}_i^N" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": " represent text and video instances respectively, and " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": " is the size of the training set. Given a text query " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": ", we initially add identifiers [CLS] and [SEP] to the sentence, and leverage the text encoder of CLIP to encode the text representation: " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "t_i = \\mathrm{TextEncoder}(T_i)" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": ". For each video " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": ", we uniformly select " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "n_v" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": " frames " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "[F_1,\\dots ,F_{n_v}]" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": " as keyframes, and employ the transformer-based encoder of CLIP to extract sequential features " + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "inline_equation", + "content": "v_{i} = \\mathrm{VideoEncoder}([F_{1},\\dots ,F_{n_{v}}])" + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 481, + 666 + ], + "type": "text", + "content": "Extensive work explores various cross-modal interaction strategies to improve retrieval performance [30,35,48]. Our work focuses on decoupling knowledge from" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "X. Zhuang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 146, + 114, + 470, + 262 + ], + "blocks": [ + { + "bbox": [ + 146, + 114, + 470, + 262 + ], + "lines": [ + { + "bbox": [ + 146, + 114, + 470, + 262 + ], + "spans": [ + { + "bbox": [ + 146, + 114, + 470, + 262 + ], + "type": "image", + "image_path": "906932314491d8f437564da914c9aa9f23a0d7e03a0b1830842f4c20535591a2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 271, + 482, + 315 + ], + "lines": [ + { + "bbox": [ + 131, + 271, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 131, + 271, + 482, + 315 + ], + "type": "text", + "content": "Fig. 2: The illustration of the proposed framework, consisting of (1) constructing a multi-grained knowledge store using open-world knowledge or closed-world knowledge (training dataset); (2) EKM algorithm that sequentially implements knowledge injection, distribution calibration, and target likelihood maximization." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 337, + 479, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 337, + 479, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 337, + 479, + 384 + ], + "type": "text", + "content": "models and exploring the impact of different knowledge sources on retrieval performance, rather than pre-training or interaction strategies. Thus we adopt the weighted token-wise interaction module of DRL [48] to achieve cross-modal interaction. Note that KDProR is also compatible with other interaction strategies." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 399, + 306, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 399, + 306, + 411 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 306, + 411 + ], + "type": "text", + "content": "3.2 Decoupling Knowledge Store" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 417, + 479, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 479, + 476 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 479, + 476 + ], + "type": "text", + "content": "Our KDProR involves augmenting CLIP-based retrieval models with semantic information from decoupled knowledge stores. Thus, the first step of our methods is to build a multi-grained knowledge store. As illustrated in Figure 2, our stores contain fine-grained and coarse-grained caches, where each key and value (text-video) are representations of knowledge." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 477, + 480, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 477, + 480, + 551 + ], + "spans": [ + { + "bbox": [ + 131, + 477, + 480, + 551 + ], + "type": "text", + "content": "Fine-grained Knowledge Store. The fine-grained knowledge store can be represented as " + }, + { + "bbox": [ + 131, + 477, + 480, + 551 + ], + "type": "inline_equation", + "content": "\\mathcal{S}^f = \\{\\mathcal{S}_i^f\\}_{i=1}^{K^f} = \\{(\\mathcal{K}_i^f, \\mathcal{V}_i^f)\\}_{i=1}^{K^f}" + }, + { + "bbox": [ + 131, + 477, + 480, + 551 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 477, + 480, + 551 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_i^f = (\\mathcal{K}_i^f, \\mathcal{V}_i^f)" + }, + { + "bbox": [ + 131, + 477, + 480, + 551 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 131, + 477, + 480, + 551 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 477, + 480, + 551 + ], + "type": "text", + "content": "-th key-value pair in the fine-grained knowledge store, and " + }, + { + "bbox": [ + 131, + 477, + 480, + 551 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_i^f" + }, + { + "bbox": [ + 131, + 477, + 480, + 551 + ], + "type": "text", + "content": " is the size of our fine-grained knowledge store. Under the closed world setting, we utilize text and visual encoders of CLIP to obtain the feature key-value pairs on the training dataset as " + }, + { + "bbox": [ + 131, + 477, + 480, + 551 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_i^f = t_i" + }, + { + "bbox": [ + 131, + 477, + 480, + 551 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 477, + 480, + 551 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i^f = v_i" + }, + { + "bbox": [ + 131, + 477, + 480, + 551 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 552, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 482, + 647 + ], + "type": "text", + "content": "Coarse-grained Knowledge Store. The coarse-grained knowledge store is designed to encapsulate global semantics. It equips retrieval models with global, coarse-grained insights from the knowledge store, enhancing the comprehensive knowledge understanding. Similarly, the coarse-grained storage can be represented as: " + }, + { + "bbox": [ + 131, + 552, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{S}^c = \\{\\mathcal{S}_i^c\\}_{i=1}^{K^c} = \\{(\\mathcal{K}_i^c, \\mathcal{V}_i^c)\\}_{i=1}^{K^c}" + }, + { + "bbox": [ + 131, + 552, + 482, + 647 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 552, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_i^c = (\\mathcal{K}_i^c, \\mathcal{V}_i^c)" + }, + { + "bbox": [ + 131, + 552, + 482, + 647 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 131, + 552, + 482, + 647 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 552, + 482, + 647 + ], + "type": "text", + "content": "-th key-value pair in the coarse-grained knowledge store, and " + }, + { + "bbox": [ + 131, + 552, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_i^c" + }, + { + "bbox": [ + 131, + 552, + 482, + 647 + ], + "type": "text", + "content": " is the size of our coarse-grained knowledge store. We adopt the K-mean strategy to extract knowledge from fine-grained storage into more advanced representations:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 199, + 652, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 652, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 199, + 652, + 481, + 666 + ], + "type": "interline_equation", + "content": "\\mathcal {K} _ {i} ^ {c} = \\phi \\left(c _ {i} ^ {t}\\right), \\forall c _ {i} ^ {t} \\in \\mathcal {C} ^ {t}, \\quad \\mathcal {V} _ {i} ^ {c} = \\phi \\left(c _ {i} ^ {v}\\right), \\forall c _ {i} ^ {v} \\in \\mathcal {C} ^ {v}, \\tag {1}", + "image_path": "ac41273a47735808d1e96920c518f49c60551adaf9a6f63740223012bfa64091.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "text", + "content": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 155 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 155 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 130, + 115, + 479, + 155 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^t = \\{c_1^t,\\dots ,c_{K^c}^t\\}" + }, + { + "bbox": [ + 130, + 115, + 479, + 155 + ], + "type": "text", + "content": " denotes the cluster set obtained by K-Mean over fine-grained knowledge " + }, + { + "bbox": [ + 130, + 115, + 479, + 155 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{K}_i^f\\}" + }, + { + "bbox": [ + 130, + 115, + 479, + 155 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 115, + 479, + 155 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^v = \\{c_1^v,\\dots ,c_{K^c}^v\\}" + }, + { + "bbox": [ + 130, + 115, + 479, + 155 + ], + "type": "text", + "content": " represents the cluster set obtained by K-Mean over " + }, + { + "bbox": [ + 130, + 115, + 479, + 155 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{V}_i^f\\}" + }, + { + "bbox": [ + 130, + 115, + 479, + 155 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 115, + 479, + 155 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot)" + }, + { + "bbox": [ + 130, + 115, + 479, + 155 + ], + "type": "text", + "content": " represents the Max-Pooling aggregate function." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 156, + 480, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 156, + 480, + 178 + ], + "spans": [ + { + "bbox": [ + 130, + 156, + 480, + 178 + ], + "type": "text", + "content": "Note that our framework is universal, and we will introduce more applications for constructing diverse knowledge stores in Section 3.6." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 196, + 394, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 196, + 394, + 209 + ], + "spans": [ + { + "bbox": [ + 131, + 196, + 394, + 209 + ], + "type": "text", + "content": "3.3 E-step for Multi-Grained Knowledge Injection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "text", + "content": "Knowledge Interpolation. In E-step, KDProR initially attempts to inject fine-grained and coarse-grained knowledge into the model by knowledge interpolation of features. We treat with " + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "text", + "content": " as anchors to retrieve the fine-grained knowledge store, and obtain two sets of top-K similar vector pairs " + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{t_i}^f" + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{v_i}^f" + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{t_i}^f = \\{(\\mathcal{K}_{t_i}^f, \\mathcal{V}_{t_i}^f)\\}" + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{v_i}^f = \\{(\\mathcal{K}_{v_i}^f, \\mathcal{V}_{v_i}^f)\\}" + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "text", + "content": " are two sets of key-value knowledge pairs, where " + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_{t_i}^f" + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{v_i}^f" + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "text", + "content": " are the nearest neighbors of " + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "text", + "content": ", respectively. After the model retrieves the top-K candidates for each " + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "text", + "content": ", their corresponding representation " + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_{t_i}^f" + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{v_i}^f" + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "text", + "content": " obtained from knowledge-store will be incorporated into original features " + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 130, + 216, + 482, + 351 + ], + "type": "text", + "content": " to act as demonstration learning. Specifically, we adopt the nearest neighbor sets to inject decoupled knowledge into text and video features as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 153, + 358, + 481, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 153, + 358, + 481, + 390 + ], + "type": "interline_equation", + "content": "t _ {i} ^ {f} = \\rho t _ {i} + (1 - \\rho) \\frac {\\sum_ {(\\mathcal {K} , \\mathcal {V}) \\in \\mathcal {N} _ {t _ {i}} ^ {f}} \\mathcal {K}}{| \\mathcal {N} _ {t _ {i}} ^ {f} |}, v _ {i} ^ {f} = \\rho v _ {i} + (1 - \\rho) \\frac {\\sum_ {(\\mathcal {K} , \\mathcal {V}) \\in \\mathcal {N} _ {v _ {i}} ^ {f}} \\mathcal {V}}{| \\mathcal {N} _ {v _ {i}} ^ {f} |}, \\tag {2}", + "image_path": "8f0329931f06a8d8f003b9f64e6fa81d33cda333323ba5c8914bd554f3b4c84b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 398, + 479, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 479, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 479, + 434 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 130, + 398, + 479, + 434 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 398, + 479, + 434 + ], + "type": "text", + "content": " is a hyperparameter used to express the interpolation ratio. Subsequently, we utilize the weighted token-wise interaction module [48] to enhance textual and visual representations injected with additional knowledge." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 434, + 480, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 480, + 493 + ], + "type": "text", + "content": "However, the textual and visual knowledge within the top-K sets, obtained by " + }, + { + "bbox": [ + 130, + 434, + 480, + 493 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 434, + 480, + 493 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 434, + 480, + 493 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 130, + 434, + 480, + 493 + ], + "type": "text", + "content": " respectively, may not correspond on a one-to-one basis due to their derivation from two independent top-K operations. To mitigate the issue of mismatched video-text features during knowledge injection, we utilize the optimal transport (OT) algorithm to facilitate alignment correction within a batch." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 494, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 494, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 131, + 494, + 482, + 581 + ], + "type": "text", + "content": "Alignment Correction via Optimal Transport. We first define " + }, + { + "bbox": [ + 131, + 494, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\langle \\mathbf{Q},\\mathbf{S}^f\\rangle =" + }, + { + "bbox": [ + 131, + 494, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{tr}(\\mathbf{Q}^\\top \\mathbf{S}^f)\\in \\mathbb{R}^{B\\times B}" + }, + { + "bbox": [ + 131, + 494, + 482, + 581 + ], + "type": "text", + "content": " as the similarity matrix calculated from the updated text and video pairs " + }, + { + "bbox": [ + 131, + 494, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\{(t_i^f,v_i^f)\\}_{i = 1}^B" + }, + { + "bbox": [ + 131, + 494, + 482, + 581 + ], + "type": "text", + "content": " in a batch. " + }, + { + "bbox": [ + 131, + 494, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}\\in \\mathbb{R}^{B\\times B}" + }, + { + "bbox": [ + 131, + 494, + 482, + 581 + ], + "type": "text", + "content": " represents the corresponding transport assignment matrix, where " + }, + { + "bbox": [ + 131, + 494, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}[i,j]" + }, + { + "bbox": [ + 131, + 494, + 482, + 581 + ], + "type": "text", + "content": " denotes the probabilities of aligning " + }, + { + "bbox": [ + 131, + 494, + 482, + 581 + ], + "type": "inline_equation", + "content": "t_i^f" + }, + { + "bbox": [ + 131, + 494, + 482, + 581 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 131, + 494, + 482, + 581 + ], + "type": "inline_equation", + "content": "v_{i}^{f}" + }, + { + "bbox": [ + 131, + 494, + 482, + 581 + ], + "type": "text", + "content": ". The optimal transport aims to establish flexible alignment between videos and captions by maximizing global text-video similarity " + }, + { + "bbox": [ + 131, + 494, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{tr}(\\mathbf{Q}^\\top \\mathbf{S}^f)" + }, + { + "bbox": [ + 131, + 494, + 482, + 581 + ], + "type": "text", + "content": ". The objective of optimal transport can be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 147, + 588, + 480, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 588, + 480, + 608 + ], + "spans": [ + { + "bbox": [ + 147, + 588, + 480, + 608 + ], + "type": "interline_equation", + "content": "\\max _ {\\mathbf {Q} \\in \\mathcal {Q}} \\langle \\mathbf {Q}, \\mathbf {S} ^ {f} \\rangle + \\varepsilon H (\\mathbf {Q}) \\quad \\text {s . t .} \\mathcal {Q} = \\left\\{\\mathbf {Q} \\in \\mathbb {R} ^ {B \\times B} \\mid \\mathbf {Q} \\mathbf {1} _ {B} = \\boldsymbol {\\mu}, \\mathbf {Q} ^ {\\top} \\mathbf {1} _ {B} = \\boldsymbol {\\nu} \\right\\} \\tag {3}", + "image_path": "be2874892cd0ccf4a563e97a25db424a2d9f36a14540cfdd9fee68438eb8032b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu} \\in \\mathbb{R}^{B}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{\\nu} \\in \\mathbb{R}^{B}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " indicate the relative importance of each video and caption, " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_{B}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " represents the vector of ones in dimension " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "H(\\mathbf{Q})" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " denotes a entropy regular term [4] and " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " controls its smoothness. Following [46], we initialize " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{\\nu}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " to a uniform distribution " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_{B}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": ". The optimal transport problem" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "X. Zhuang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "in Eq. 3 can be solved using the Sinkhorn-Knopp algorithm [46] to obtain the optimal assignment scheme " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^*" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ". Based on this, we can obtain the similarity matrix after adding alignment correction " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^*" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 246, + 157, + 481, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 157, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 246, + 157, + 481, + 171 + ], + "type": "interline_equation", + "content": "\\mathbf {S} ^ {f ^ {*}} = \\left(\\left(1 - \\beta\\right) \\mathbf {I} + \\beta \\mathbf {Q} ^ {*}\\right) \\mathbf {S} ^ {f}, \\tag {4}", + "image_path": "7f954c1159cdc3a754639f1e88dde76aead8706fa89a5cb99487f1f6e823cdeb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 177, + 389, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 177, + 389, + 189 + ], + "spans": [ + { + "bbox": [ + 131, + 177, + 389, + 189 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 131, + 177, + 389, + 189 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 131, + 177, + 389, + 189 + ], + "type": "text", + "content": " is the identity matrix and " + }, + { + "bbox": [ + 131, + 177, + 389, + 189 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 131, + 177, + 389, + 189 + ], + "type": "text", + "content": " is a hyperparameter." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 189, + 479, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 479, + 213 + ], + "type": "text", + "content": "We can implement the coarse-grained knowledge injection and obtain the realigned similarity matrix " + }, + { + "bbox": [ + 130, + 189, + 479, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{c*}" + }, + { + "bbox": [ + 130, + 189, + 479, + 213 + ], + "type": "text", + "content": " in a similar way." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 228, + 388, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 228, + 388, + 239 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 388, + 239 + ], + "type": "text", + "content": "3.4 K-step for Retrieval Distribution Calibration" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 245, + 480, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 480, + 327 + ], + "type": "text", + "content": "Existing studies [23, 41] have shown that non-parametric KNN classification or nearest-neighbor interpolation methods have extraordinary potential in improving robustness and generalization. It is intuitively to leverage the KNN's retrieval results as the prior knowledge to guide the parameters training for video-text pairs that are difficult to establish deep correlations. In practice, we aggregate the prediction of KNN into the original probability distribution of video-text retrieval to achieve the multi-grained calibration of retrieval distribution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 329, + 480, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 480, + 389 + ], + "type": "text", + "content": "Fine-grained Distribution Calibration. We utilize the set of k-nearest neighbors " + }, + { + "bbox": [ + 130, + 329, + 480, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{t_i}^f" + }, + { + "bbox": [ + 130, + 329, + 480, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 329, + 480, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{v_i}^f" + }, + { + "bbox": [ + 130, + 329, + 480, + 389 + ], + "type": "text", + "content": " retrieved by querying the open-book knowledge store in E-step as fine-grained top-K knowledge. Subsequently, we calculate the KNN distribution according to the softmax of the similarity and aggregation probability mass of each video-text pair in the retrieved target as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 393, + 480, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 393, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 141, + 393, + 480, + 429 + ], + "type": "interline_equation", + "content": "p _ {k n n} ^ {f} \\left(v _ {i} \\mid t _ {i}\\right) = \\frac {\\sum_ {\\left(\\mathcal {K} , \\mathcal {V}\\right) \\in N _ {t _ {i}} ^ {f}} \\varphi \\left(v _ {i} , \\mathcal {V}\\right) e ^ {d \\left(v _ {i} , \\mathcal {V}\\right)}}{\\sum_ {\\left(\\mathcal {K} , \\mathcal {V}\\right) \\in N _ {t _ {i}} ^ {f}} \\varphi \\left(v _ {i} , \\mathcal {V}\\right)}, \\quad \\varphi \\left(v _ {i}, \\mathcal {V}\\right) = \\left\\{ \\begin{array}{l l} 1, & \\mathcal {V} \\in N _ {v _ {i}} ^ {f} \\\\ 0, & \\mathcal {V} \\notin N _ {v _ {i}} ^ {f} \\end{array} , \\right. \\tag {5}", + "image_path": "0189bf8876308f33d737ba80c6f0ca790db22fb3511e12aefa09047eb47408ad.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 434, + 479, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 479, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 479, + 470 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 130, + 434, + 479, + 470 + ], + "type": "inline_equation", + "content": "d(\\cdot ,\\cdot)\\in [-1,1]" + }, + { + "bbox": [ + 130, + 434, + 479, + 470 + ], + "type": "text", + "content": " denotes the cosine similarity function, and " + }, + { + "bbox": [ + 130, + 434, + 479, + 470 + ], + "type": "inline_equation", + "content": "\\varphi (v_{i},\\mathcal{V})" + }, + { + "bbox": [ + 130, + 434, + 479, + 470 + ], + "type": "text", + "content": " is an indicator function used to determine whether " + }, + { + "bbox": [ + 130, + 434, + 479, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 434, + 479, + 470 + ], + "type": "text", + "content": " is within the top-K nearest neighbor set " + }, + { + "bbox": [ + 130, + 434, + 479, + 470 + ], + "type": "inline_equation", + "content": "N_{v_i}^f" + }, + { + "bbox": [ + 130, + 434, + 479, + 470 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 130, + 434, + 479, + 470 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 130, + 434, + 479, + 470 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 471, + 479, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 479, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 479, + 495 + ], + "type": "text", + "content": "We can further obtain the KNN probability distribution " + }, + { + "bbox": [ + 130, + 471, + 479, + 495 + ], + "type": "inline_equation", + "content": "p_{knn}^{f}(t_{i} \\mid v_{i})" + }, + { + "bbox": [ + 130, + 471, + 479, + 495 + ], + "type": "text", + "content": " for video-to-text retrieval in a completely symmetrical manner." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 496, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 480, + 628 + ], + "type": "text", + "content": "Coarse-grained Distribution Calibration. Consistent with Eq. 5, we employ the KNN algorithm to obtain top-K knowledge, utilizing closed-world or open-world knowledge from the coarse-grained knowledge store to refine the probability distribution function pertinent to the text-to-video retrieval. Nevertheless, since the coarse-grained knowledge store is derived from the distillation of fine-grained knowledge, it is impracticable to apply indicator functions (cf. Eq. 5) to ascertain the presence of an inclusion relation between textual features and their nearest neighbor sets. To tackle this challenge, we introduce a novel indicator function, denoted as " + }, + { + "bbox": [ + 130, + 496, + 480, + 628 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 496, + 480, + 628 + ], + "type": "text", + "content": ", parameterized by a threshold " + }, + { + "bbox": [ + 130, + 496, + 480, + 628 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 496, + 480, + 628 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 496, + 480, + 628 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 496, + 480, + 628 + ], + "type": "text", + "content": " is designed to evaluate whether " + }, + { + "bbox": [ + 130, + 496, + 480, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 496, + 480, + 628 + ], + "type": "text", + "content": " resides within the top-K nearest neighbor set " + }, + { + "bbox": [ + 130, + 496, + 480, + 628 + ], + "type": "inline_equation", + "content": "N_{v_i}^c" + }, + { + "bbox": [ + 130, + 496, + 480, + 628 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 130, + 496, + 480, + 628 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 130, + 496, + 480, + 628 + ], + "type": "text", + "content": " via the cosine similarity function " + }, + { + "bbox": [ + 130, + 496, + 480, + 628 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 496, + 480, + 628 + ], + "type": "text", + "content": ", and is formulated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 203, + 637, + 480, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 637, + 480, + 669 + ], + "spans": [ + { + "bbox": [ + 203, + 637, + 480, + 669 + ], + "type": "interline_equation", + "content": "\\phi \\left(v _ {i}, \\mathcal {V}\\right) = \\left\\{ \\begin{array}{l l} 1, & d \\left(\\mathcal {V}, \\mathcal {V} ^ {c}\\right) < \\alpha , \\exists \\left(\\mathcal {K} ^ {c}, \\mathcal {V} ^ {c}\\right) \\in N _ {v _ {i}} ^ {c} \\\\ 0, & o t h e r w i s e s \\end{array} . \\right. \\tag {6}", + "image_path": "71525edb29060b4005eeac41c1e2d677052a25fdd9c82546c2f370145d2ba090.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "type": "text", + "content": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": "Based on this, we can obtain a closed-form expression for the KNN distribution used for coarse-grained distribution calibration as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 212, + 147, + 481, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 147, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 212, + 147, + 481, + 182 + ], + "type": "interline_equation", + "content": "p _ {k n n} ^ {c} \\left(v _ {i} \\mid t _ {i}\\right) = \\frac {\\sum_ {\\left(\\mathcal {K} , \\mathcal {V}\\right) \\in N _ {t _ {i}} ^ {c}} \\phi \\left(v _ {i} , \\mathcal {V}\\right) e ^ {d \\left(v _ {i} , \\mathcal {V}\\right)}}{\\sum_ {\\left(\\mathcal {K} , \\mathcal {V}\\right) \\in N _ {t _ {i}} ^ {c}} \\phi \\left(v _ {i} , \\mathcal {V}\\right)}. \\tag {7}", + "image_path": "d882ccf9a6aa1fcb246d1adeac1c00381eacdf5cfd7635fdd04c020138a939c6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 187, + 479, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 187, + 479, + 211 + ], + "spans": [ + { + "bbox": [ + 132, + 187, + 479, + 211 + ], + "type": "text", + "content": "We can further obtain the KNN probability distribution " + }, + { + "bbox": [ + 132, + 187, + 479, + 211 + ], + "type": "inline_equation", + "content": "p_{knn}^{c}(t_{i} \\mid v_{i})" + }, + { + "bbox": [ + 132, + 187, + 479, + 211 + ], + "type": "text", + "content": " for coarse-grained video-to-text retrieval in a completely symmetrical manner." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 373, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 373, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 373, + 239 + ], + "type": "text", + "content": "3.5 M-step for Unified Retrieval Optimization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 245, + 479, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 479, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 479, + 292 + ], + "type": "text", + "content": "In the M-step, we optimize the whole model by maximizing the retrieval likelihood objective over the realigned similarity matrix. Specifically, we utilize the re-aligned multi-grained similarity matrix in Eq. 4 obtained from E-step to calculate the InfoNCE loss:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 168, + 298, + 480, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 298, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 168, + 298, + 480, + 343 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t 2 v} = - \\frac {1}{B} \\sum_ {i} ^ {B} \\left(\\log \\frac {\\exp \\left(\\mathbf {S} ^ {f _ {i i} ^ {*}}\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\mathbf {S} ^ {f _ {i j} ^ {*}}\\right)} + \\log \\frac {\\exp \\left(\\mathbf {S} _ {i i} ^ {c ^ {*}}\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\mathbf {S} _ {i j} ^ {c ^ {*}}\\right)}\\right), \\tag {8}", + "image_path": "2df63b446eb8e9ec96c88ce3faf8189945feae853d319fbcac473782fd94b494.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 168, + 340, + 446, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 340, + 446, + 378 + ], + "spans": [ + { + "bbox": [ + 168, + 340, + 446, + 378 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v 2 t} = - \\frac {1}{B} \\sum_ {i} ^ {B} \\left(\\log \\frac {\\exp \\left(\\mathbf {S} _ {i i} ^ {f ^ {*}}\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\mathbf {S} _ {j i} ^ {f ^ {*}}\\right)} + \\log \\frac {\\exp \\left(\\mathbf {S} _ {i i} ^ {c ^ {*}}\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\mathbf {S} _ {j i} ^ {c ^ {*}}\\right)}\\right),", + "image_path": "4d3f0a97035d945a90b0cf1ad9451d3be592c570c5a046e4875fa469e44db96a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 383, + 480, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 480, + 467 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 383, + 480, + 467 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 383, + 480, + 467 + ], + "type": "text", + "content": " is the batch size. Subsequently, we propose utilizing the KNN distribution (cf. Eq. 5 and 7) to guide the training process. The KNN calibrator reweights the InfoNCE loss by adjusting the relative loss for the correctly-matched or mismatched video-text instances identified by KNN, respectively. We focus on exploiting the results of KNN distribution for calibrating the training of retrieval models. Specifically, we apply the negative log-likelihood of " + }, + { + "bbox": [ + 130, + 383, + 480, + 467 + ], + "type": "inline_equation", + "content": "p_{knn}^{f}" + }, + { + "bbox": [ + 130, + 383, + 480, + 467 + ], + "type": "text", + "content": " in Eq 5 and " + }, + { + "bbox": [ + 130, + 383, + 480, + 467 + ], + "type": "inline_equation", + "content": "p_{knn}^{c}" + }, + { + "bbox": [ + 130, + 383, + 480, + 467 + ], + "type": "text", + "content": " in Eq. 7 as the calibration factor:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 189, + 476, + 480, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 476, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 189, + 476, + 480, + 496 + ], + "type": "interline_equation", + "content": "F _ {t 2 v} ^ {f} = - \\log p _ {k n n} ^ {f} \\left(v _ {i} \\mid t _ {i}\\right), \\quad F _ {v 2 t} ^ {f} = - \\log p _ {k n n} ^ {f} \\left(t _ {i} \\mid v _ {i}\\right); \\tag {9}", + "image_path": "98ff786395de5484315729a6172b0e7cb827291883544096203c269e32281a4d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 189, + 494, + 425, + 505 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 494, + 425, + 505 + ], + "spans": [ + { + "bbox": [ + 189, + 494, + 425, + 505 + ], + "type": "interline_equation", + "content": "F _ {t 2 v} ^ {c} = - \\log p _ {k n n} ^ {c} \\left(v _ {i} \\mid t _ {i}\\right), \\quad F _ {v 2 t} ^ {c} = - \\log p _ {k n n} ^ {c} \\left(t _ {i} \\mid v _ {i}\\right).", + "image_path": "739dc7d969915b0dd998d464cbfe754fbf4b8d2f88dd8a899505eb272bd3c0f0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 509, + 479, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 479, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 479, + 532 + ], + "type": "text", + "content": "Finally, we can optimize the retrieval model by maximizing the following knowledge-guided cross-modal contrastive loss as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 145, + 539, + 480, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 539, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 145, + 539, + 480, + 553 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v 2 t} ^ {*} = \\left(1 + \\lambda_ {1} F _ {v 2 t} ^ {f} + \\lambda_ {2} F _ {v 2 t} ^ {c}\\right) \\mathcal {L} _ {v 2 t}; \\quad \\mathcal {L} _ {t 2 v} ^ {*} = \\left(1 + \\lambda_ {1} F _ {t 2 v} ^ {f} + \\lambda_ {2} F _ {t 2 v} ^ {c}\\right) \\mathcal {L} _ {t 2 v}, \\tag {10}", + "image_path": "13104ceb89a1e7c9c1e573533ecf901062d9f271380c074cf4efe5b738a16d05.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 559, + 479, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 479, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 479, + 596 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 559, + 479, + 596 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 559, + 479, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 559, + 479, + 596 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 559, + 479, + 596 + ], + "type": "text", + "content": " are hyper-parameters to determine the proportion of each loss term. Hence, in the M-step, the complete training objective of KDProR can be formulated as " + }, + { + "bbox": [ + 130, + 559, + 479, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{v2t}^{*} + \\mathcal{L}_{t2v}^{*}" + }, + { + "bbox": [ + 130, + 559, + 479, + 596 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 611, + 368, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 368, + 624 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 368, + 624 + ], + "type": "text", + "content": "3.6 Applications of Our KDProR Framework" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "In this section, as illustrated in Figure 3, we present three distinct applications for constructing open-world knowledge stores, including pre-trained knowledge, entity-relation structured knowledge, and CoT knowledge within LLMs." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "X. Zhuang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 156, + 114, + 455, + 211 + ], + "blocks": [ + { + "bbox": [ + 156, + 114, + 455, + 211 + ], + "lines": [ + { + "bbox": [ + 156, + 114, + 455, + 211 + ], + "spans": [ + { + "bbox": [ + 156, + 114, + 455, + 211 + ], + "type": "image", + "image_path": "c4524fa7d5fdadbc17b735a712b83c4ad55e4d985731291e859026e15c0e9f2c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 220, + 480, + 254 + ], + "lines": [ + { + "bbox": [ + 131, + 220, + 480, + 254 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 480, + 254 + ], + "type": "text", + "content": "Fig. 3: The illustration of proposed open-world knowledge stores, including (1) video-text pretraining knowledge, (2) entity-relation structure knowledge, and (3) knowledge within LLMs prompted by the CoT technique." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 274, + 479, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 479, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 479, + 310 + ], + "type": "text", + "content": "Video-Text Pretraining. As shown in Figure 3, we utilize CLIP-ViP [59], which is further pre-trained on HD-VILA-100M [58] dataset, to extract pretraining video-text knowledge for constructing our knowledge stores." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 311, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 480, + 357 + ], + "type": "text", + "content": "Entity-Relation Structure. We attempt to parse captions to scene graphs " + }, + { + "bbox": [ + 130, + 311, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 311, + 480, + 357 + ], + "type": "text", + "content": " through the Scene Graph Parser [51], then employ " + }, + { + "bbox": [ + 130, + 311, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 311, + 480, + 357 + ], + "type": "text", + "content": " to obtain subject-relation-object triplets. Subsequently, we use these triplets as substitutes for the original captions and then employ CLIP to extract text-video features to construct stores." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 358, + 480, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 480, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 480, + 406 + ], + "type": "text", + "content": "CoT Generation of LLMs. In this work, we attempt to use a strategy similar to CoT [50] to generate auxiliary captions for each video frame by prompting LLMs. Subsequently, we utilize CLIP to extract neural representations of auxiliary captions and corresponding video frames to construct knowledge stores." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 434, + 263, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 434, + 263, + 445 + ], + "spans": [ + { + "bbox": [ + 131, + 434, + 263, + 445 + ], + "type": "text", + "content": "3.7 Theoretical Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 451, + 480, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 451, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 451, + 480, + 514 + ], + "type": "text", + "content": "From a probabilistic perspective, the optimization goal of the VTR model " + }, + { + "bbox": [ + 131, + 451, + 480, + 514 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 131, + 451, + 480, + 514 + ], + "type": "text", + "content": " trained on dataset " + }, + { + "bbox": [ + 131, + 451, + 480, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{x_{i}\\}_{i}^{N} = \\{(T_{i},V_{i})\\}_{i}^{N}" + }, + { + "bbox": [ + 131, + 451, + 480, + 514 + ], + "type": "text", + "content": " is to maximize the log-likelihood " + }, + { + "bbox": [ + 131, + 451, + 480, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\Theta)" + }, + { + "bbox": [ + 131, + 451, + 480, + 514 + ], + "type": "text", + "content": " for text-video pairs. We consider the knowledge introduced from fine-grained and coarse-grained stores as two related latent variables " + }, + { + "bbox": [ + 131, + 451, + 480, + 514 + ], + "type": "inline_equation", + "content": "z_{i}^{f}" + }, + { + "bbox": [ + 131, + 451, + 480, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 451, + 480, + 514 + ], + "type": "inline_equation", + "content": "z_{i}^{c}" + }, + { + "bbox": [ + 131, + 451, + 480, + 514 + ], + "type": "text", + "content": ". In a closed world, the posterior distribution of " + }, + { + "bbox": [ + 131, + 451, + 480, + 514 + ], + "type": "inline_equation", + "content": "z_{i}^{f}" + }, + { + "bbox": [ + 131, + 451, + 480, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 451, + 480, + 514 + ], + "type": "inline_equation", + "content": "z_{i}^{c}" + }, + { + "bbox": [ + 131, + 451, + 480, + 514 + ], + "type": "text", + "content": " can be represented as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 201, + 519, + 480, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 519, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 201, + 519, + 480, + 536 + ], + "type": "interline_equation", + "content": "Q \\left(z _ {i} ^ {f}\\right) = p \\left(z _ {i} ^ {f}; x _ {i}, \\Theta\\right), \\quad Q \\left(z _ {i} ^ {c}\\right) = p \\left(z _ {i} ^ {c}; z _ {i} ^ {f}, x _ {i}, \\Theta\\right). \\tag {11}", + "image_path": "42c76cb23c669b22ceccd78c634371ca86856a5855601431c0a0278cdd92da77.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 540, + 479, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 540, + 479, + 564 + ], + "spans": [ + { + "bbox": [ + 131, + 540, + 479, + 564 + ], + "type": "text", + "content": "Thus we can rewrite the objective function of the retrieval model " + }, + { + "bbox": [ + 131, + 540, + 479, + 564 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 131, + 540, + 479, + 564 + ], + "type": "text", + "content": " according to Jensen's inequality [19] as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 185, + 570, + 480, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 570, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 185, + 570, + 480, + 601 + ], + "type": "interline_equation", + "content": "\\Theta^ {*} = \\arg \\max \\sum_ {i} \\sum_ {z _ {i} ^ {f}, z _ {i} ^ {c}} Q \\left(z _ {i} ^ {f}\\right) Q \\left(z _ {i} ^ {c}\\right) \\log p \\left(x _ {i}, z _ {i} ^ {f}, z _ {i} ^ {c}; \\Theta\\right). \\tag {12}", + "image_path": "7b13f210605026414574bbf3e5130f444973bd319dcbcab0c49f9b697f724dc8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 605, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 480, + 666 + ], + "type": "text", + "content": "To sum up, the E-step and K-step aim to estimate the posterior probability " + }, + { + "bbox": [ + 131, + 605, + 480, + 666 + ], + "type": "inline_equation", + "content": "Q(z_{i}^{f})" + }, + { + "bbox": [ + 131, + 605, + 480, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 605, + 480, + 666 + ], + "type": "inline_equation", + "content": "Q(z_{i}^{c})" + }, + { + "bbox": [ + 131, + 605, + 480, + 666 + ], + "type": "text", + "content": " on fine-grained and coarse-grained knowledge stores with given current " + }, + { + "bbox": [ + 131, + 605, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\Theta^{(t)}" + }, + { + "bbox": [ + 131, + 605, + 480, + 666 + ], + "type": "text", + "content": ". Then KDProR achieves efficient alignment of videos and texts and obtains " + }, + { + "bbox": [ + 131, + 605, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\Theta^{(t + 1)}" + }, + { + "bbox": [ + 131, + 605, + 480, + 666 + ], + "type": "text", + "content": " under knowledge guidance at M-step by maximizing the log-likelihood in Eq. 12 with known " + }, + { + "bbox": [ + 131, + 605, + 480, + 666 + ], + "type": "inline_equation", + "content": "Q(z_{i}^{f})" + }, + { + "bbox": [ + 131, + 605, + 480, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 605, + 480, + 666 + ], + "type": "inline_equation", + "content": "Q(z_{i}^{c})" + }, + { + "bbox": [ + 131, + 605, + 480, + 666 + ], + "type": "text", + "content": ", and satisfies the Theorem 1:" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "text", + "content": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 168, + 482, + 358 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Results on MSR-VTT without any post-processing (e.g., [3] and [2]). " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "{}^{ + }" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "{}^{ \\top }" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": " are closed-set settings and denote constructing knowledge stores on the training set using the original CLIP and the fine-tuned CLIP upon the training set, respectively. *, † and ‡ denote pre-trained, structural and CoT knowledge in Section 3.6, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 168, + 482, + 358 + ], + "lines": [ + { + "bbox": [ + 135, + 168, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 135, + 168, + 482, + 358 + ], + "type": "table", + "html": "
BackboneMethodText-to-Video RetrievalVideo-to-Text Retrieval
R@1↑R@5↑R@10↑MdR↓MnR↓R@1↑R@5↑R@10↑MdR↓MnR↓
Non-CLIPCE [31]20.948.862.46.028.220.650.364.05.325.1
MMT [12]26.657.169.64.024.027.057.569.73.721.3
Support [42]27.456.367.73.0-26.655.167.53.0-
Frozen [1]31.059.670.53.0------
CLIP ViT-B/32CLIP4clip [35]44.571.481.62.015.342.770.980.62.011.6
X-pool [13]46.972.882.22.014.3-----
EMCL [21]46.873.183.12.0-46.573.583.52.0-
TS2-Net [33]47.074.583.82.013.045.374.183.72.09.2
DRL [48]47.474.683.82.0-45.373.983.32.0-
STAN [28]46.972.882.82.0------
UATVR [10]47.573.983.52.012.946.973.883.82.08.6
PromptSwitch [6]47.873.982.2-14.146.074.384.8-8.5
KDProR+48.474.484.22.012.247.174.684.42.08.9
KDProR+48.774.484.12.012.047.274.684.42.08.7
KDProR*49.274.284.42.012.047.174.084.12.08.9
KDProR†49.074.684.32.012.047.374.383.92.08.7
KDProR‡49.675.184.42.011.648.274.284.22.08.1
", + "image_path": "38bea03cc3f70225f9c77c8359f18cf76a214e99e73cd644e03f386d8a3c35c4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 381, + 483, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 381, + 483, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 381, + 483, + 407 + ], + "type": "text", + "content": "Theorem 1 The retrieval model " + }, + { + "bbox": [ + 132, + 381, + 483, + 407 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 132, + 381, + 483, + 407 + ], + "type": "text", + "content": " optimized iteratively through the proposed EKM algorithm of step " + }, + { + "bbox": [ + 132, + 381, + 483, + 407 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 381, + 483, + 407 + ], + "type": "text", + "content": " satisfies the following properties:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 411, + 423, + 436 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 411, + 405, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 405, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 405, + 422 + ], + "type": "text", + "content": "- Monotonic Increasing: " + }, + { + "bbox": [ + 138, + 411, + 405, + 422 + ], + "type": "inline_equation", + "content": "\\forall t \\geq 0" + }, + { + "bbox": [ + 138, + 411, + 405, + 422 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 138, + 411, + 405, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\Theta^{(t + 1)}) \\geq \\mathcal{L}(\\Theta^{(t)})" + }, + { + "bbox": [ + 138, + 411, + 405, + 422 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 422, + 423, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 423, + 436 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 423, + 436 + ], + "type": "text", + "content": "- Convergence: " + }, + { + "bbox": [ + 138, + 422, + 423, + 436 + ], + "type": "inline_equation", + "content": "\\forall \\epsilon > 0, \\exists \\delta > 0" + }, + { + "bbox": [ + 138, + 422, + 423, + 436 + ], + "type": "text", + "content": ", if " + }, + { + "bbox": [ + 138, + 422, + 423, + 436 + ], + "type": "inline_equation", + "content": "t > \\delta" + }, + { + "bbox": [ + 138, + 422, + 423, + 436 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 138, + 422, + 423, + 436 + ], + "type": "inline_equation", + "content": "|\\Theta^{(t + 1)} - \\Theta^{(t)}| \\leq \\epsilon" + }, + { + "bbox": [ + 138, + 422, + 423, + 436 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 449, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 485 + ], + "type": "text", + "content": "Theorem 1 indicates that our EKM algorithm can ensure the model is positively optimized in each iteration, while also maintaining the convergence properties of vanilla EM algorithms." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 502, + 230, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 230, + 517 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 230, + 517 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 526, + 269, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 269, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 269, + 540 + ], + "type": "text", + "content": "4.1 Experiments Settings" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Datasets. We conduct experiments on four benchmarks for video-text retrieval tasks including: MSR-VTT [55], DiDeMo [16], LSMDC [45] and ActivityNet [15]. MSR-VTT [55] comprises 10,000 videos with 20 captions each. We train on 9,000 videos and their captions, testing on 1k-A with 1,000 videotext pairs. DiDeMo [16] provides 10,000 videos annotated with 40,000 sentences. We concatenate all descriptions of a video into one paragraph for evaluation. LSMDC [45] includes 118,081 movie-extracted videos, with our dataset partitioned into 109,673 for training, 7,408 for validation, and 1,000 for testing. ActivityNet [15] includes 20,000 YouTube videos. Following [35], we concatenate all captions similar to DiDeMo and evaluate on the 'val1' split." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "X. Zhuang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 157, + 482, + 310 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Text-to-Video retrieval results on DiDeMo, ActivityNet and LSMDC datasets without any post-processing (e.g., [3] and [2]). " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\top" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " are closed-set settings and " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " are open-world settings, whose meanings are all consistent with those in Table 1." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 157, + 482, + 310 + ], + "lines": [ + { + "bbox": [ + 134, + 157, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 134, + 157, + 482, + 310 + ], + "type": "table", + "html": "
MethodDiDeMoActivityNetLSMDC
R@1↑R@5↑R@10↑MdR↓R@1↑R@5↑R@10↑MdR↓R@1↑R@5↑R@10↑MdR↓
Text-to-Video Retrieval using CLIP (ViT-B/32)
CLIP4clip [35]43.470.280.62.040.572.483.62.022.641.049.111.0
X-CLIP [36]45.274.0--44.374.1--23.343.0--
TS2-Net [33]41.871.682.02.041.073.684.52.023.442.350.99.0
DRL [48]47.973.882.72.044.274.586.12.024.945.755.37.0
PromptSwitch [10]--------23.141.750.5-
UATVR [10]43.171.882.32.0--------
DiffusionRet [22]46.774.782.72.045.875.686.32.024.443.154.38.0
\\(KDProR^{+}\\)48.876.184.12.045.875.485.02.026.145.957.37.0
\\(KDProR^{\\top}\\)49.275.383.62.046.476.285.72.026.547.857.86.0
\\(KDProR^{*}\\)51.778.384.91.048.376.886.12.026.950.258.35.0
\\(KDProR^{\\dagger}\\)50.576.385.11.047.576.185.72.027.650.658.15.0
\\(KDProR^{\\ddagger}\\)53.279.285.01.049.177.986.82.028.250.659.25.0
", + "image_path": "56dd927188e6b562656f81513048cb985ada27134f7e12432ea4e9be90880e02.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 334, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 525 + ], + "type": "text", + "content": "Metrics. Following [35, 48], we utilize recall at rank K (R@K), median rank (MdR) and mean rank (MnR) as metrics to validate the effectiveness of KDProR. Implementation Details. Following [48], we adopt CLIP's ViT-B/32 as the visual encoder and fine-tune the model with 5 epochs on all training sets. We adopt Adam [24] as the optimizer with a cosine warm-up method [34]. Following [35, 48], the frame and caption length are 12 and 32 for MSR-VTT and LSMDC, 64 and 64 for ActivityNet and DiDeMo. The initial learning rate for vision and text encoder is set to " + }, + { + "bbox": [ + 130, + 334, + 482, + 525 + ], + "type": "inline_equation", + "content": "10^{-7}" + }, + { + "bbox": [ + 130, + 334, + 482, + 525 + ], + "type": "text", + "content": " and other modules are " + }, + { + "bbox": [ + 130, + 334, + 482, + 525 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 334, + 482, + 525 + ], + "type": "text", + "content": ". The hyperparameter " + }, + { + "bbox": [ + 130, + 334, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 334, + 482, + 525 + ], + "type": "text", + "content": " in Eq. 2 is set to 0.2 on MSR-VTT and LSMDC, 0.1 on ActivityNet and DiDeMo for optimal performance. " + }, + { + "bbox": [ + 130, + 334, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 334, + 482, + 525 + ], + "type": "text", + "content": " in Eq. 4 are set to 0.2, and " + }, + { + "bbox": [ + 130, + 334, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 334, + 482, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 334, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 334, + 482, + 525 + ], + "type": "text", + "content": " in Eq. 9 are set to 0.2 and 0.3, respectively. We use grid search to obtain the optimal parameters. We uniformly set the top-5 relevant knowledge for KNN retrieval, and the size of the coarse-grained knowledge store " + }, + { + "bbox": [ + 130, + 334, + 482, + 525 + ], + "type": "inline_equation", + "content": "K^{c}" + }, + { + "bbox": [ + 130, + 334, + 482, + 525 + ], + "type": "text", + "content": " is set to 256. We adopt the gpt-3.5-turbo-1106 [39] to obtain the CoT knowledge in Section 3.6. All experiments are conducted on 2 Tesla A100 GPUs (80G) and 8 RTX 4090 GPUs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 546, + 228, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 228, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 228, + 558 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 569, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 481, + 604 + ], + "type": "text", + "content": "Comparisons to State-of-the-arts. In this section, we compare KDProR with state-of-the-art methods on four datasets, as results shown in Table 1 and 2. Analysis of Table 1 and 2 yields several insights:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "(1) Our KDProR, employing multi-grained knowledge stores constructed from vanilla and fine-tuned CLIP, can achieve performance gains of " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " R@1 on MSR-VTT, " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " R@1 on ActivityNet compared to DRL [48], respectively. This indicates that KDProR can achieve a harmonious balance between memorization and optimization, thereby improving performance." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "text", + "content": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 147, + 482, + 229 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 3: Comparisons to Other Baseline Methods in Text-to-Video Retrieval. " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "^\\ddagger" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": " denotes the utilization of open-world CoT knowledge in Section 3.6." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 147, + 482, + 229 + ], + "lines": [ + { + "bbox": [ + 134, + 147, + 482, + 229 + ], + "spans": [ + { + "bbox": [ + 134, + 147, + 482, + 229 + ], + "type": "table", + "html": "
Baseline MethodsKnowledge SourceMSR-VTTLSMDCDiDeMo
R@1↑R@5↑MnR↓R@1↑R@5↑MnR↓R@1↑R@5↑MnR↓
Data AugmentationCaptions from GPT47.374.712.126.446.943.949.776.314.2
Feature ConcatenatePretrained CLIP-ViP47.874.511.926.146.144.550.376.313.4
Cap4Video [52]Captions from GPT49.374.312.0---52.079.410.5
TEFAL [18]Audio Signal49.475.912.026.846.144.4---
Our KDProR‡Captions from GPT49.675.111.628.250.634.953.279.210.2
", + "image_path": "428b2045b346449166cec7427409141bec3d88379882cb66d7db402f9237ec94.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 154, + 285, + 459, + 376 + ], + "blocks": [ + { + "bbox": [ + 132, + 242, + 482, + 276 + ], + "lines": [ + { + "bbox": [ + 132, + 242, + 482, + 276 + ], + "spans": [ + { + "bbox": [ + 132, + 242, + 482, + 276 + ], + "type": "text", + "content": "Table 4: Text-to-Video retrieval results on MSR-VTT using our re-implemented CLIP4clip [35] and EMCL [21] as base models. " + }, + { + "bbox": [ + 132, + 242, + 482, + 276 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 132, + 242, + 482, + 276 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 242, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\top" + }, + { + "bbox": [ + 132, + 242, + 482, + 276 + ], + "type": "text", + "content": " are closed-set settings, *, " + }, + { + "bbox": [ + 132, + 242, + 482, + 276 + ], + "type": "inline_equation", + "content": "^\\dagger" + }, + { + "bbox": [ + 132, + 242, + 482, + 276 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 242, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 132, + 242, + 482, + 276 + ], + "type": "text", + "content": " are open-world settings, whose meanings are consistent with those in Table 1." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 154, + 285, + 459, + 376 + ], + "lines": [ + { + "bbox": [ + 154, + 285, + 459, + 376 + ], + "spans": [ + { + "bbox": [ + 154, + 285, + 459, + 376 + ], + "type": "table", + "html": "
MethodCLIP4clip [35] as Base ModelEMCL [21] as Base Model
R@1↑R@5↑R@10↑MdR↓MnR↓R@1↑R@5↑R@10↑MdR↓MnR↓
Base Model43.670.681.02.016.146.673.582.72.013.6
+ KDProR+45.573.482.72.014.147.873.583.12.013.8
+ KDProRT46.474.082.92.013.848.174.083.52.013.7
+ KDProR*47.074.482.92.013.648.773.683.62.013.7
+ KDProR†46.974.082.82.013.648.473.783.82.013.3
+ KDProR‡47.374.683.32.013.449.074.883.82.012.9
", + "image_path": "8c723305f89e86fd41a0d0b2f93510468dda5b7b8fa438c63d8de38ffb56449e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 399, + 481, + 521 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 130, + 399, + 481, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 481, + 472 + ], + "type": "text", + "content": "(2) The integration of additional open-world knowledge significantly bolsters the predictive capabilities of retrieval models. KDProR of employing three open-world knowledge (i.e., pre-trained, structural, and CoT knowledge) in Section 3.6 achieves superior performance, e.g., " + }, + { + "bbox": [ + 130, + 399, + 481, + 472 + ], + "type": "inline_equation", + "content": "+1.8\\%" + }, + { + "bbox": [ + 130, + 399, + 481, + 472 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 399, + 481, + 472 + ], + "type": "inline_equation", + "content": "+1.6\\%" + }, + { + "bbox": [ + 130, + 399, + 481, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 399, + 481, + 472 + ], + "type": "inline_equation", + "content": "+2.2\\%" + }, + { + "bbox": [ + 130, + 399, + 481, + 472 + ], + "type": "text", + "content": " R@1 on MSR-VTT, " + }, + { + "bbox": [ + 130, + 399, + 481, + 472 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 130, + 399, + 481, + 472 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 399, + 481, + 472 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 130, + 399, + 481, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 399, + 481, + 472 + ], + "type": "inline_equation", + "content": "5.3\\%" + }, + { + "bbox": [ + 130, + 399, + 481, + 472 + ], + "type": "text", + "content": " R@1 on DiDeMo compared to DRL [48], respectively. We attribute the performance gain to incorporating open-set knowledge." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 472, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 521 + ], + "type": "text", + "content": "(3) KDProR demonstrates superior performance compared to all baselines across all benchmarks. This demonstrates that our KDProR offers a twofold advantage: it can unify and incorporate diverse additional knowledge and balance memorization and retrieval to significantly improve performance." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 521, + 482, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 642 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 642 + ], + "type": "text", + "content": "Comparisons to Other Baseline Methods. We further compare our method with other baselines in Table 3, including Cap4Video [52] and TEFAL [18]. In addition, we compare it with two baseline methods that introduce open-world knowledge, namely using auxiliary captions generated by LLM as data augmentation and directly concatenating the features generated by CLIP-VIP with the original text-video features. The results indicate that introducing open-world knowledge can enhance performance, and our KDProR achieves better performance gains than Data Augmentation and Feature Concatenate. Impressively, KDProR significantly outperforms state-of-the-art open-set baselines, e.g., " + }, + { + "bbox": [ + 130, + 521, + 482, + 642 + ], + "type": "inline_equation", + "content": "+1.4\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 642 + ], + "type": "text", + "content": " R@1 by Cap4Video on LSMDC and " + }, + { + "bbox": [ + 130, + 521, + 482, + 642 + ], + "type": "inline_equation", + "content": "+1.2\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 642 + ], + "type": "text", + "content": " R@1 by TEFAL on DiDeMo." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 642, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 642, + 481, + 666 + ], + "type": "text", + "content": "Generalization Analysis. We further equip KDProR with two powerful base models in addition to DRL, i.e., CLIP4clip [35] and EMCL [21], and evaluate" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "X. Zhuang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 146, + 482, + 274 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 5: Ablation experiments on MSR-VTT 9k dataset. " + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "{}^{ \\ddagger }" + }, + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": " denotes the utilization of open-world CoT knowledge in Section 3.6." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 146, + 482, + 274 + ], + "lines": [ + { + "bbox": [ + 135, + 146, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 135, + 146, + 482, + 274 + ], + "type": "table", + "html": "
G. MethodText-to-Video RetrievalVideo-to-Text Retrieval
R@1↑R@5↑R@10↑MnR↓R@1↑R@5↑R@10↑MnR↓
w/o all Knowledge Stores (baseline)46.974.082.412.545.874.084.38.8
w/o Fine-grained Knowledge Store48.174.483.312.146.873.584.08.5
w/o Coarse-grained Knowledge Store48.474.583.711.947.273.884.08.4
w/o the whole E-step47.773.983.412.046.273.584.18.9
w/o Knowledge Interpolation E-step48.273.683.712.047.473.983.88.7
w/o OT Alignment Correction in E-step48.974.684.211.846.673.483.68.9
w/o the whole K-step47.973.883.112.347.073.684.08.7
w/o Fine-grained Calibration in K-step48.774.583.811.847.673.884.08.5
w/o Coarse-grained Calibration in K-step48.574.383.311.847.373.983.68.5
4 Our Full KDProR‡49.675.184.411.648.274.284.28.1
", + "image_path": "0e4d32936c80c94036c3574b2e0558dc6d8609d32ced0f86853de7e1c3df9539.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 160, + 285, + 306, + 380 + ], + "blocks": [ + { + "bbox": [ + 160, + 285, + 306, + 380 + ], + "lines": [ + { + "bbox": [ + 160, + 285, + 306, + 380 + ], + "spans": [ + { + "bbox": [ + 160, + 285, + 306, + 380 + ], + "type": "image", + "image_path": "fc88c9de3ce35d1afb69c10a09b10f0d80fbf06b2be8fb6710b16ec85bc6dc2e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 389, + 480, + 412 + ], + "lines": [ + { + "bbox": [ + 130, + 389, + 480, + 412 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 480, + 412 + ], + "type": "text", + "content": "Fig. 4: Ablation experiment on MSR-VTT under closed-world settings. Effect of (a) the number K of top-K knowledge; (b) the knowledge interpolation coefficient " + }, + { + "bbox": [ + 130, + 389, + 480, + 412 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 389, + 480, + 412 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 286, + 455, + 381 + ], + "blocks": [ + { + "bbox": [ + 307, + 286, + 455, + 381 + ], + "lines": [ + { + "bbox": [ + 307, + 286, + 455, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 455, + 381 + ], + "type": "image", + "image_path": "f2fdab8ca87fac4712c877274d2a5506bfb6b4dd17aa65e41ce06ecaab3386df.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "text", + "content": "on MSR-VTT. Table 4 shows that KDProR can be applied to successfully boost all baselines as a plug-and-play module under both closed-set or open-world settings. The significant improvements prove the generalization ability of KDProR." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 496, + 307, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 496, + 307, + 509 + ], + "spans": [ + { + "bbox": [ + 131, + 496, + 307, + 509 + ], + "type": "text", + "content": "4.3 Ablation Study and Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": "Effect of the Multi-Grained Knowledge Store. As shown in Groups 1 and 4 in Table 5, removing either fine-grained or coarse-grained knowledge stores will result in a significant decrease in retrieval performance, e.g., " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "-1.5\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "-1.2\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": " R@1 text-to-video retrieval on MSR-VTT, respectively. When all stores are removed, the model will lose the ability to decouple knowledge and introduce additional knowledge, resulting in a significant drop in performance. This verifies the effectiveness of our multi-grained knowledge decoupling strategy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Effect of the Knowledge Injection in E-step. As shown in Groups 2 and 4 in Table 5, we evaluate the effect of the knowledge injection in E-step (i.e., Eq. 2) by removing different components. We can observe that when the knowledge interpolation strategy is removed, the performance drops sharply, indicating that our strategy can effectively inject knowledge from stores. We further remove the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "text", + "content": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 474, + 184 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 474, + 184 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 474, + 184 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 474, + 184 + ], + "type": "image", + "image_path": "f0cbf02afbdb2e22fd8706c4b5e215e083cd9bb2c6a0b4d4b0f7c91e38be7464.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 192, + 480, + 216 + ], + "lines": [ + { + "bbox": [ + 130, + 192, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 480, + 216 + ], + "type": "text", + "content": "Fig. 5: The text-video results on the MSR-VTT 1K-A test set. " + }, + { + "bbox": [ + 130, + 192, + 480, + 216 + ], + "type": "inline_equation", + "content": "\\top" + }, + { + "bbox": [ + 130, + 192, + 480, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 192, + 480, + 216 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 130, + 192, + 480, + 216 + ], + "type": "text", + "content": " are close- and open-world settings respectively, whose meanings are consistent with those in Table 1." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 239, + 479, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 479, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 479, + 262 + ], + "type": "text", + "content": "OT alignment correction strategy (i.e., Eq. 4) and observe a significant decrease in video-to-text retrieval performance, proving our strategy's effectiveness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 263, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 480, + 357 + ], + "type": "text", + "content": "Effect of the Distribution Calibration in K-step. We conduct experiments to study the effect of the KNN distribution calibration strategy as illustrated in Table 5 with Groups 3 and 4. It can be seen that the performance of KDProR significantly decreases without KNN distribution to calibrate the retrieval distribution in Eq. 9. Moreover, the lack of distribution calibration (fine-grained or coarse-grained) in the K-step can also affect retrieval performance. This verifies that the multi-grained distribution calibration strategy can calibrate the retrieval distribution to inject additional knowledge sufficiently and effectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 358, + 480, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 480, + 418 + ], + "type": "text", + "content": "Effect of the number K of top-K Knowledge. In Figure 4 (left), we show the effect of the number K of top-K knowledge. On the one hand, we find that fewer K (" + }, + { + "bbox": [ + 130, + 358, + 480, + 418 + ], + "type": "inline_equation", + "content": "K < 5" + }, + { + "bbox": [ + 130, + 358, + 480, + 418 + ], + "type": "text", + "content": ") means less knowledge is selected, which limits our KDProR's generalization ability for different knowledge. On the other hand, a larger K (" + }, + { + "bbox": [ + 130, + 358, + 480, + 418 + ], + "type": "inline_equation", + "content": "K > 5" + }, + { + "bbox": [ + 130, + 358, + 480, + 418 + ], + "type": "text", + "content": ") requires more retrieval time, which increases the cost of learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 418, + 480, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 480, + 525 + ], + "type": "text", + "content": "Effect of the Knowledge Interpolation Coefficient. In Figure 4 (right), we show the effect of the knowledge interpolation coefficient " + }, + { + "bbox": [ + 130, + 418, + 480, + 525 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 418, + 480, + 525 + ], + "type": "text", + "content": " on KDProR. We can observe that KDProR achieves the optimal balance between knowledge injection and optimization when " + }, + { + "bbox": [ + 130, + 418, + 480, + 525 + ], + "type": "inline_equation", + "content": "\\rho = 0.2" + }, + { + "bbox": [ + 130, + 418, + 480, + 525 + ], + "type": "text", + "content": ". Deviations from this optimal " + }, + { + "bbox": [ + 130, + 418, + 480, + 525 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 418, + 480, + 525 + ], + "type": "text", + "content": ", either towards a smaller or larger number, result in performance degradation, which is attributed to inadequate knowledge acquisition and insufficient optimization, respectively. Cases Study. We provide two cases of videos retrieved by KDProR and DRL [48]. As illustrated in Figure 5, KDProR successfully retrieves ground-truth videos in both closed- and open-set settings, whereas DRL fails to do so." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 544, + 220, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 220, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 220, + 558 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 569, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 481, + 665 + ], + "type": "text", + "content": "In this work, we explore a novel knowledge-decoupling VTR paradigm that utilizes multi-grained knowledge stores to unify closed- and open-world knowledge. This strategy allows VTR models to achieve an efficient balance between memorization and retrieval optimization, while also opening up a unified interface for injecting various open-world knowledge. We further propose a principled EKM algorithm to achieve an efficient injection of knowledge. Our method could be applied to other visual-language tasks, e.g., visual question answering, and injecting various open-world knowledge for augmenting cross-modal understanding." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 102 + ], + "type": "text", + "content": "X. Zhuang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 163 + ], + "type": "text", + "content": "We would like to thank all reviewers for their insightful comments. This paper was partially supported by NSFC (No: 62176008)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 205, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 205, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 480, + 239 + ], + "type": "text", + "content": "1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1728-1738 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 239, + 480, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 271 + ], + "type": "text", + "content": "2. Bogolin, S.V., Croitoru, I., Jin, H., Liu, Y., Albanie, S.: Cross modal retrieval with querybank normalisation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5194-5205 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 272, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 272, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 272, + 480, + 304 + ], + "type": "text", + "content": "3. Cheng, X., Lin, H., Wu, X., Yang, F., Shen, D.: Improving video-text retrieval by multi-stream corpus alignment and dual softmax loss. arXiv preprint arXiv:2109.04290 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 304, + 480, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 327 + ], + "type": "text", + "content": "4. Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems 26 (2013)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 327, + 480, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 480, + 348 + ], + "type": "text", + "content": "5. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the em-algorithm plus discussions on the paper. In:\"\" (1977)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "type": "text", + "content": "6. Deng, C., Chen, Q., Qin, P., Chen, D., Wu, Q.: Prompt switch: Efficient clip adaptation for text-video retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15648-15658 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 381, + 480, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 480, + 414 + ], + "type": "text", + "content": "7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: North American Chapter of the Association for Computational Linguistics (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 414, + 480, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 458 + ], + "type": "text", + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv abs/2010.11929 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 458, + 480, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 480, + 491 + ], + "type": "text", + "content": "9. Elangovan, A., He, J., Verspoor, K.M.: Memorization vs. generalization : Quantifying data leakage in nlp performance evaluation. In: Conference of the European Chapter of the Association for Computational Linguistics (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 491, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 523 + ], + "type": "text", + "content": "10. Fang, B., Liu, C., Zhou, Y., Yang, M., Song, Y., Li, F., Wang, W., Ji, X., Ouyang, W., et al.: Uatvr: Uncertainty-adaptive text-video retrieval. arXiv preprint arXiv:2301.06309 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "text", + "content": "1. Feldman, V.: Does learning require memorization? a short tale about a long tail. Proceedings of the 52nd Annual ACM SIGACT Symposium on Theory of Computing (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 555, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 589 + ], + "type": "text", + "content": "2. Gabeur, V., Sun, C., Alahari, K., Schmid, C.: Multi-modal transformer for video retrieval. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. pp. 214-229. Springer (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 589, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 632 + ], + "type": "text", + "content": "3. Gorti, S.K., Vouitsis, N., Ma, J., Golestan, K., Volkovs, M., Garg, A., Yu, G.: X-pool: Cross-modal language-video attention for text-video retrieval. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5006-5015 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 666 + ], + "type": "text", + "content": "4. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 770-778 (2015)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "text", + "content": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "15. Heilbron, F.C., Escorcia, V., Ghanem, B., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 961-970 (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "16. Hendricks, L.A., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.C.: Localizing moments in video with natural language. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 5804-5813 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 481, + 205 + ], + "type": "text", + "content": "17. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Computation 9, 1735-1780 (1997)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 481, + 237 + ], + "type": "text", + "content": "18. Ibrahimi, S., Sun, X., Wang, P., Garg, A., Sanan, A., Omar, M.: Audio-enhanced text-to-video retrieval using text-conditioned feature alignment. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 12020-12030 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 481, + 259 + ], + "type": "text", + "content": "19. Jensen, J.L.W.V.: Sur les fonctions convexes et les inégalités entre les valeurs moyennes. Acta Mathematica 30, 175-193 (1906)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 481, + 293 + ], + "type": "text", + "content": "20. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q.V., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. ArXiv abs/2102.05918 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "text", + "content": "21. Jin, P., Huang, J., Liu, F., Wu, X., Ge, S., Song, G., Clifton, D., Chen, J.: Expectation-maximization contrastive learning for compact video-and-language representations. Advances in Neural Information Processing Systems 35, 30291-30306 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 335, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 481, + 369 + ], + "type": "text", + "content": "22. Jin, P., Li, H., Cheng, Z.L., Li, K., Ji, X., Liu, C., ming Yuan, L., Chen, J.: Diffusionret: Generative text-video retrieval with diffusion model. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 2470-2481 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 403 + ], + "type": "text", + "content": "23. Khandelwal, U., Levy, O., Jurafsky, D., Zettlemoyer, L., Lewis, M.: Generalization through memorization: Nearest neighbor language models. arXiv preprint arXiv:1911.00172 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "text", + "content": "24. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. CoRR abs/1412.6980 (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 468 + ], + "type": "text", + "content": "25. Li, H., Cao, M., Cheng, X., Li, Y., Zhu, Z., Zou, Y.: G2l: Semantically aligned and uniform video grounding via geodesic and game theory. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 11998-12008 (2023), https://api-semanticscholar.org/CorpusID:260164585" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "26. Li, H., Cao, M., Cheng, X., Zhu, Z., Li, Y., Zou, Y.: Exploiting auxiliary caption for video grounding. In: AAAI Conference on Artificial Intelligence (2023), https://apisemantic scholar.org/CorpusID:257772084" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "type": "text", + "content": "27. Li, K., Wang, Y., Li, Y., Wang, Y., He, Y., Wang, L., Qiao, Y.: Unmasked teacher: Towards training-efficient video foundation models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 19891-19903 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "text", + "content": "28. Liu, R., Huang, J., Li, G., Feng, J., Wu, X., Li, T.H.: Revisiting temporal modeling for clip-based image-to-video knowledge transferring. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6555-6564 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "text", + "content": "29. Liu, S., Fan, H., Qian, S., Chen, Y., Ding, W., Wang, Z.: Hit: Hierarchical transformer with momentum contrast for video-text retrieval. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 11895-11905 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "30. Liu, S., Fan, H., Qian, S., Chen, Y., Ding, W., Wang, Z.: Hit: Hierarchical transformer with momentum contrast for video-text retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11915-11925 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "31. Liu, Y., Albanie, S., Nagrani, A., Zisserman, A.: Use what you have: Video retrieval using representations from collaborative experts. arXiv preprint arXiv:1907.13487 (2019)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "X. Zhuang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "text", + "content": "32. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. ArXiv abs/1907.11692 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "33. Liu, Y., Xiong, P., Xu, L., Cao, S., Jin, Q.: Ts2-net: Token shift and selection transformer for text-video retrieval. In: European Conference on Computer Vision. pp. 319-335. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "34. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv: Learning (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "35. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing 508, 293-304 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "36. Ma, Y., Xu, G., Sun, X., Yan, M., Zhang, J., Ji, R.: X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. Proceedings of the 30th ACM International Conference on Multimedia (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "text", + "content": "37. Meng, Y., Zong, S., Li, X., Sun, X., Zhang, T., Wu, F., Li, J.: Gnn-lm: Language modeling based on global contexts via gnn. ArXiv abs/2110.08743 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "text", + "content": "38. Mikolov, T., Chen, K., Corrado, G.S., Dean, J.: Efficient estimation of word representations in vector space. In: International Conference on Learning Representations (2013)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 380 + ], + "type": "text", + "content": "39. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L.E., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.J.: Training language models to follow instructions with human feedback. ArXiv abs/2203.02155 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 423 + ], + "type": "text", + "content": "40. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 445 + ], + "type": "text", + "content": "41. Papernot, N., McDaniel, P.: Deep k-nearest neighbors: Towards confident, interpretable and robust deep learning. arXiv preprint arXiv:1803.04765 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "42. Patrick, M., Huang, P.Y., Asano, Y., Metze, F., Hauptmann, A., Henriques, J., Vedaldi, A.: Support-set bottlenecks for video-text representation learning. arXiv preprint arXiv:2010.02824 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 522 + ], + "type": "text", + "content": "43. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 522, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 566 + ], + "type": "text", + "content": "44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "text", + "content": "45. Rohrbach, A., Rohrbach, M., Tandon, N., Schiele, B.: A dataset for movie description. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3202-3212 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "type": "text", + "content": "46. Su, B., Hua, G.: Order-preserving Wasserstein distance for sequence matching. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1049-1057 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "47. Tänzer, M., Ruder, S., Rei, M.: Memorisation versus generalisation in pre-trained language models. In: Annual Meeting of the Association for Computational Linguistics (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "text", + "content": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "48. Wang, Q., Zhang, Y., Zheng, Y., Pan, P., Hua, X.: Disentangled representation learning for text-video retrieval. ArXiv abs/2203.07111 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "text", + "content": "49. Wang, Z., Sung, Y.L., Cheng, F., Bertasius, G., Bansal, M.: Unified coarse-to-fine alignment for video-text retrieval. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 2804-2815 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 174, + 479, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 174, + 479, + 206 + ], + "spans": [ + { + "bbox": [ + 133, + 174, + 479, + 206 + ], + "type": "text", + "content": "50. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 208, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 208, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 134, + 208, + 480, + 251 + ], + "type": "text", + "content": "51. Wu, H., Mao, J., Zhang, Y., Jiang, Y., Li, L., Sun, W., Ma, W.Y.: Unified visual-semantic embeddings: Bridging vision and language with structured meaning representations. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6602-6611 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 252, + 480, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 252, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 134, + 252, + 480, + 284 + ], + "type": "text", + "content": "52. Wu, W., Luo, H., Fang, B., Wang, J., Ouyang, W.: Cap4video: What can auxiliary captions do for text-video retrieval? 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 10704-10713 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 285, + 480, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 285, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 134, + 285, + 480, + 327 + ], + "type": "text", + "content": "53. Wu, X., Li, H., Luo, Y., Cheng, X., Zhuang, X., Cao, M., Fu, K.: Uncertainty-aware sign language video retrieval with probability distribution modeling. ArXiv abs/2405.19689 (2024), https://api.sementicscholar.org/CorpusID: 270123137" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 330, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 330, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 133, + 330, + 480, + 352 + ], + "type": "text", + "content": "54. Xie, Y., Zhu, Z., Zhuang, X., Liang, L., Wang, Z., Zou, Y.: Gpa: Global and prototype alignment for audio-text retrieval. In: Interspeech (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 353, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 353, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 133, + 353, + 480, + 386 + ], + "type": "text", + "content": "55. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5288-5296 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 387, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 387, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 134, + 387, + 480, + 430 + ], + "type": "text", + "content": "56. Xu, Z., Chen, Z., Zhang, Y., Song, Y., Wan, X., Li, G.: Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17503-17512 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 431, + 480, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 431, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 134, + 431, + 480, + 485 + ], + "type": "text", + "content": "57. Xu, Z., Huang, J., Liu, T., Liu, Y., Han, H., Yuan, K., Li, X.: Enhancing fine-grained multi-modal alignment via adapters: A parameter-efficient training framework for referring image segmentation. In: 2nd Workshop on Advancing Neural Network Training: Computational Efficiency, Scalability, and Resource Optimization (WANT@ ICML 2024) (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 487, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 487, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 134, + 487, + 480, + 531 + ], + "type": "text", + "content": "58. Xue, H., Hang, T., Zeng, Y., Sun, Y., Liu, B., Yang, H., Fu, J., Guo, B.: Advancing high-resolution video-language representation with large-scale video transcriptions. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5026-5035 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 532, + 480, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 532, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 134, + 532, + 480, + 563 + ], + "type": "text", + "content": "59. Xue, H., Sun, Y., Liu, B., Fu, J., Song, R., Li, H., Luo, J.: Clip-vip: Adapting pre-trained image-text model to video-language alignment. In: International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 565, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 565, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 134, + 565, + 480, + 597 + ], + "type": "text", + "content": "60. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. Trans. Mach. Learn. Res. 2022 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 599, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 599, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 133, + 599, + 480, + 620 + ], + "type": "text", + "content": "61. Yu, Y., Kim, J., Kim, G.: A joint sequence fusion model for video question answering and retrieval. In: European Conference on Computer Vision (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 621, + 480, + 665 + ], + "type": "text", + "content": "62. Yuan, L., Chen, D., Chen, Y.L., Codella, N.C.F., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., Liu, C., Liu, M., Liu, Z., Lu, Y., Shi, Y., Wang, L., Wang, J., Xiao, B., Xiao, Z., Yang, J., Zeng, M., Zhou, L., Zhang, P.: Florence: A new foundation model for computer vision. ArXiv abs/2111.11432 (2021)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "X. Zhuang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 228 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 162 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 162 + ], + "type": "text", + "content": "63. Zhao, Y., Li, K., Cheng, Z., Qiao, P., Zheng, X., Ji, R., Liu, C., Yuan, L., Chen, J.: Graco: Granularity-controllable interactive segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3501-3510 (June 2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 482, + 205 + ], + "type": "text", + "content": "64. Zhao, Y., Lv, W., Xu, S., Wei, J., Wang, G., Dang, Q., Liu, Y., Chen, J.: Detrs beat yolos on real-time object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16965-16974 (June 2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 482, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 482, + 228 + ], + "type": "text", + "content": "65. Zheng, X., Jiang, J.: An empirical study of memorization in nlp. In: Annual Meeting of the Association for Computational Linguistics (2022)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 447, + 102 + ], + "type": "text", + "content": "Knowledge-Decoupling Probabilistic Framework for Video-Text Retrieval" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/KFD-NeRF_ Rethinking Dynamic NeRF with Kalman Filter/3bae14dd-a29b-4634-bf7b-d258a05d1f55_content_list.json b/2024/KFD-NeRF_ Rethinking Dynamic NeRF with Kalman Filter/3bae14dd-a29b-4634-bf7b-d258a05d1f55_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c1d475977355d208c67d95fc42b42876fd929ba9 --- /dev/null +++ b/2024/KFD-NeRF_ Rethinking Dynamic NeRF with Kalman Filter/3bae14dd-a29b-4634-bf7b-d258a05d1f55_content_list.json @@ -0,0 +1,1927 @@ +[ + { + "type": "text", + "text": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter", + "text_level": 1, + "bbox": [ + 241, + 140, + 759, + 183 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yifan Zhan $^{1}$ , Zhuoxiao Li $^{1}$ , Muyao Niu $^{1}$ , Zhihang Zhong $^{3}$ , Shohei Nobuhara $^{2}$ , Ko Nishino $^{2}$ , and Yinqiang Zheng $^{\\star 1}$", + "bbox": [ + 220, + 212, + 781, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The University of Tokyo", + "bbox": [ + 411, + 253, + 589, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Kyoto University", + "bbox": [ + 436, + 268, + 565, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Shanghai Artificial Intelligence Laboratory", + "bbox": [ + 352, + 282, + 650, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We introduce KFD-NeRF, a novel dynamic neural radiance field integrated with an efficient and high-quality motion reconstruction framework based on Kalman filtering. Our key idea is to model the dynamic radiance field as a dynamic system whose temporally varying states are estimated based on two sources of knowledge: observations and predictions. We introduce a novel plug-in Kalman filter guided deformation field that enables accurate deformation estimation from scene observations and predictions. We use a shallow Multi-Layer Perceptron (MLP) for observations and model the motion as locally linear to calculate predictions with motion equations. To further enhance the performance of the observation MLP, we introduce regularization in the canonical space to facilitate the network's ability to learn warping for different frames. Additionally, we employ an efficient tri-plane representation for encoding the canonical space, which has been experimentally demonstrated to converge quickly with high quality. This enables us to use a shallower observation MLP, consisting of just two layers in our implementation. We conduct experiments on synthetic and real data and compare with past dynamic NeRF methods. Our KFD-NeRF demonstrates similar or even superior rendering performance within comparable computational time and achieves state-of-the-art view synthesis performance with thorough training. Github page: https://github.com/Yifever20002/KFD-NeRF.", + "bbox": [ + 259, + 339, + 743, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Dynamic NeRF $\\cdot$ Deformable Network $\\cdot$ Kalman Filter", + "bbox": [ + 261, + 643, + 715, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 686, + 375, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRF) [30] have demonstrated outstanding success as a versatile and accurate 3D representation of real-world scenes, which has led to its wide adoption in daily and industrial applications in numerous domains. One of the remaining key desiderata for NeRFs is its extension to dynamic scenes.", + "bbox": [ + 212, + 722, + 787, + 781 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing dynamic NeRF methods can be broadly categorized into two approaches. One is to learn deformation fields for motion warping (e.g., D-NeRF [36]", + "bbox": [ + 212, + 782, + 789, + 814 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* denotes corresponding author.", + "bbox": [ + 217, + 825, + 434, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cd76fcb37d8442f2af9a36fa968ce6d3d58bfc27e340375e385fa84fc65371a8.jpg", + "image_caption": [ + "Fig. 1: In contrast to a vanilla deformation field, our plug-in Kalman filter guided deformation field consists of a prediction branch along with the direct observations from input data. From noise related terms $\\varepsilon_{t_i}$ and $\\varepsilon_{t_{i - 1}}$ we learn Kalman gain $K_{t_i}$ , weighting observations $y_{t_i}$ and predictions $\\hat{d\\pmb{x}_{t_i}}$ for more accurate deformation estimations." + ], + "image_footnote": [], + "bbox": [ + 225, + 148, + 776, + 397 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and TiNeuVox [10]). Another is to disregard motion priors and directly interpolate time in the feature space (e.g., DyNeRF [24] and KPlanes [12]). These approaches, however, often overlook the characteristics of a dynamic radiance field as a time-state sequence, missing the opportunity to fully leverage temporal contextual information.", + "bbox": [ + 212, + 505, + 787, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we draw inspirations from control theory and model the 4D radiance field as a dynamic system with temporally-varying states. Our state estimates of this dynamic system come from two sources of knowledge: observations based on the input data and predictions based on the system physical dynamics. Optimal state estimates cannot be achieved with only one of these two sources of knowledge. On the one hand, observations, as commonly used in previous dynamic NeRF works, are inherently error-prone due to the discrete temporal sampling of the dynamic scenes. On the other hand, predictions are governed by the correctness of the assumed kinematic model and may struggle to maintain accuracy for real dynamic scenes.", + "bbox": [ + 212, + 582, + 787, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To maximize combined potential of both observations and predictions, we introduce an efficient plug-and-play Kalman filter [43] module to optimize state estimations of our dynamic system. In Fig. 1 we illustrate our plug-in Kalman filter guided deformation field. We model the 4D radiance field as a single-state system, with the state denoted as $d\\pmb{x}_{t_i}$ for the current frame deformation. In contrast to a vanilla deformation field that only considers system observations in the current frame, our approach incorporates richer information from previous", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y.Zhan, Z.Li et al.", + "bbox": [ + 271, + 114, + 397, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "frames by introducing a prediction branch based on a motion equation. Given the absence of prior trajectory regarding the scene's motion, we employ local linear motion.", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Both observations and predictions are weighted using a learnable Kalman gain to calculate precise deformation estimations. During the initial stages of training, predictions primarily influence the process, facilitating the convergence of frames with substantial motion. In the later stages of training, observations take precedence, allowing for the recovery of more precise and fine-grained motion details.", + "bbox": [ + 212, + 191, + 785, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "All points in the real space are warped to a time-independent canonical space according to the estimated deformations. To further improve the performance of our two-branch deformable model, we employ an efficient tri-plane spatial representation for encoding our canonical space. Experimental evidence shows that this representation permits a shallower observation MLP with only two layers in our implementation. At the same time, we improve the warping ability of the observation MLP by regularizing the learning of the radiance field in the canonical space.", + "bbox": [ + 212, + 282, + 787, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 238, + 402, + 566, + 417 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) The first method for modeling 4D radiance fields as dynamic systems by integrating a Kalman filter into the deformation field formulation, which results in a plug-in, efficient method for estimating deformations;", + "2) KFD-NeRF, a novel deformable NeRF with the Kalman filter plug-in and a tri-plane spatial representation, trained with a novel strategy of gradually releasing temporal information to facilitate learning dynamic systems:", + "3) and regularization in the canonical space for enhancing the learning capacity of a shallow observation MLP. We achieve state-of-the-art results on both synthetic and real data with all these designs, compared to dynamic NeRFs." + ], + "bbox": [ + 220, + 426, + 782, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 583, + 395, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Radiance Fields Representation. NeRF [30] represents 3D scenes based on an implicit representation encoded by an MLP. Given multi-view images and corresponding camera poses as input, the MLP is trained by \"analysis by synthesis,\" achieving novel view synthesis and coarse geometry reconstruction. Vanilla NeRF leaves room for improvement that has attracted abundant research.", + "bbox": [ + 212, + 613, + 782, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Many methods [13, 20, 26, 39, 48] use sparse voxel grids to represent the 3D scenes. While grid-based modeling achieves fast training and rendering, fine details require high-resolution grids leading to excessive memory consumption. The optimization process for grids is also unstable, failing to leverage the smoothness bias brought by MLP. Instant-NGP [31] proposes an efficient hash encoding for mapping spatial and feature domains. For sparsely-observed regions, however, this one-to-many mapping can easily introduce noise. The tri-plane representation [5, 6, 35] significantly reduces the memory footprint by leveraging its low-rank decomposition. We use this representation to build a fine-detailed canonical space.", + "bbox": [ + 212, + 688, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic Neural Rendering. For dynamic scenarios, modeling the time dimension together with the spatial representation becomes the main challenge. One approach to dynamic NeRF [4, 12, 14, 15, 21, 24, 25, 27, 34, 38, 45] is to add timestamps as an extra dimension to the 3D space and formulate 4D interpolation. Though intuitive, this leads to temporal incoherence due to the lack of supervision and priors between frames. Another approach [7, 9, 10, 17, 18, 25, 32, 33, 36, 40, 49] is to model the motion with deformation fields that warp points in arbitrary frames to the canonical space conditioned on the timestamps. This design can take advantage of the MLP smoothness bias in accordance with the motion smoothness prior. Nevertheless, deformation fields fail to find correspondences in frames with significant motion. We emphasize the performance improvement resulting from the switch in backbone spatial representation from D-NeRF [36] (MLP-based) to TiNeuVox [10] (voxel grids based) and later analyze the impact of different spatial representations on the deformation field. We also compare with latest point-based 4D rendering works [29, 44, 46] inspired by 3D Gaussian Splatting [23].", + "bbox": [ + 212, + 146, + 787, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Deep Kalman Filter. The combination of deep learning and Kalman filtering has been explored to address the challenges of incomplete observations in various scenarios. Some works [28, 52] use Convolutional Neural Network (CNN)-based Kalman filtering for video compression and camera localization. Recurrent Neural Network (RNN)-based Kalman filter has also been used in some studies [8, 37] to improve state estimation and to optimize pose regularization. The Transformer in [50] enables a more comprehensive exploitation of temporal contextual information. All of these works along with [3, 11, 19], however, model dynamics only in the learned latent space without taking into account physical priors, with the exception of [16] which fuses known physical priors and network outputs to construct a Kalman filter for video prediction. In this paper, for the first time, we employ a neural Kalman filter to assist in the dynamic NeRF task. We derive a two-stream method consisting of a shallow MLP and physical priors to achieve efficient motion estimation.", + "bbox": [ + 212, + 402, + 787, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 637, + 382, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 NeRF and Volume Rendering Revisited", + "text_level": 1, + "bbox": [ + 215, + 667, + 589, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NeRF [30] consists of three parts: sampling, volume mapping, and rendering. In the sampling process, points $\\pmb{x} \\in \\mathbb{R}^3$ are sampled along rays calculated from the camera pose. Then, in the volume mapping process, each $\\pmb{x}$ in the 3D world, as well as its viewing direction $\\pmb{v} \\in \\mathbb{R}^3$ , is queried to output the volume density $\\sigma$ and radiance $\\pmb{c} = (r, g, b)$ of $\\pmb{x}$ . Finally, in the rendering process, the color of each ray is computed with volume rendering [22]. The expected color of the ray $r(s) = o + sv$ becomes [30]", + "bbox": [ + 212, + 693, + 787, + 800 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC (\\boldsymbol {r}) = \\int_ {s _ {n}} ^ {s _ {f}} T (s) \\sigma (\\boldsymbol {r} (s)) \\boldsymbol {c} (\\boldsymbol {r} (s), \\boldsymbol {v}) d s, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 810, + 785, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y.Zhan, Z.Li et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s_n$ and $s_f$ are the near and far bounds, respectively, and", + "bbox": [ + 215, + 146, + 671, + 162 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT (s) = \\exp \\left(- \\int_ {s _ {n}} ^ {s} \\sigma (\\boldsymbol {r} (k)) d k\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 174, + 785, + 208 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The only supervision for training NeRF comes from the ground truth color $C_{gt}(\\boldsymbol{r})$ of each ray", + "bbox": [ + 215, + 217, + 782, + 250 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {i m a g e}} = \\sum_ {\\boldsymbol {r} \\in \\mathcal {R}} \\| C (\\boldsymbol {r}) - C _ {g t} (\\boldsymbol {r}) \\| _ {2} ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 260, + 785, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{R}$ is the ray batch.", + "bbox": [ + 215, + 304, + 401, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Kalman Filter", + "text_level": 1, + "bbox": [ + 215, + 342, + 380, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consider a dynamic system with input $u_{t}$ , output $y_{t}$ , process noise $n_t$ and measurement noise $m_t$ . We want to obtain its state $x_{t}$ at each frame $t$ , assuming the state equation follows", + "bbox": [ + 215, + 367, + 785, + 412 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = A x _ {t - 1} + B u _ {t} + n _ {t}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 426, + 785, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the output equation follows", + "bbox": [ + 215, + 453, + 450, + 468 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny _ {t} = C x _ {t} + m _ {t}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 446, + 482, + 785, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $A$ , $B$ and $C$ are the system matrix, control matrix and output matrix, respectively.", + "bbox": [ + 212, + 508, + 782, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "There are two methods for obtaining the system's state $x_{t}$ at any given $t$ : observation and prediction. The observation method tries to rely on $y_{t}$ in Eq. (5) while the prediction method tries to mathematically model the system dynamics to predict the state. Both of these methods, however, come with non-negligible errors due to the existence of $m_{t}$ and $n_{t}$ and for inaccurate system modeling.", + "bbox": [ + 212, + 539, + 784, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Kalman filter algorithm posits that the state $x_{t}$ is obtained by a weighted sum of the prediction based on the state $x_{t-1}$ and the observation $y_{t}$ . For a specific $t$ , estimating the state $x_{t}$ can be divided into two steps: prediction step and update step. Assuming that the process noise and measurement noise follow zero-mean Gaussian distributions with variances $Q$ and $R$ , respectively, at the prediction step, it first calculates a predicted state based on the estimated state at $t-1$", + "bbox": [ + 212, + 614, + 784, + 718 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {t} ^ {-} = A \\hat {x} _ {t - 1} + B u _ {t}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 720, + 784, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and its error covariance", + "bbox": [ + 215, + 744, + 388, + 757 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {t} ^ {-} = A P _ {t - 1} A ^ {T} + Q. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 758, + 784, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the update step, this predicted state is combined with observation $y_{t}$ to obtain updated state estimate", + "bbox": [ + 215, + 781, + 782, + 811 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {t} = \\hat {x} _ {t} ^ {-} + K _ {t} \\left(y _ {t} - C \\hat {x} _ {t} ^ {-}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 825, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter", + "bbox": [ + 326, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e9b5666b0bf697a6a488424ba18708fc1e2de9adfa55e02f7848b4f8a2746b90.jpg", + "image_caption": [ + "Fig. 2: Visualization of feature planes learned by the feature interpolation method ① and the deformation fields method ②. We show a point $P$ in the real world space and its corresponding four points $P_{1}, P_{2}, P_{3}$ and $P_{4}$ in the canonical space at four different timestamps. The feature plane in ② exhibits better smoothness compared to ①, so we use ② to construct our backbone." + ], + "image_footnote": [], + "bbox": [ + 222, + 148, + 460, + 277 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ebcc21e63339fbf3db054c13e464e62b6269b166bd8bb8f2fc46697238dc5bb9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 157, + 774, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and its error covariance", + "bbox": [ + 215, + 393, + 388, + 407 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP _ {t} = \\left(I - K _ {t} C\\right) P _ {t} ^ {-}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 409, + 784, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 215, + 434, + 263, + 448 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nK _ {t} = \\frac {P _ {t} ^ {-} C ^ {T}}{C P _ {t} ^ {-} C ^ {T} + R} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 445, + 784, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is the Kalman gain, reflecting the weights assigned to the observation and prediction components.", + "bbox": [ + 214, + 486, + 784, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 542, + 330, + 558 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 3 illustrates the three stages of the complete pipeline of KFD-NeRF. In this section, we will first analyze the advantages of using deformation fields as the motion representation over feature interpolation. We will then introduce KFD-NeRF based on the Kalman filter to achieve accurate deformation estimations. Finally, we will discuss spatial reconstruction details, the training strategy, and the incorporation of regularization.", + "bbox": [ + 214, + 575, + 787, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Motion Representation: Deformation Fields vs. Feature Interpolation", + "text_level": 1, + "bbox": [ + 215, + 691, + 725, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4D NeRFs generally employ two approaches to temporal modeling: motion deformation fields or time-conditioned feature interpolation. A deformation field $D(\\pmb{x}, t) \\rightarrow \\Delta \\pmb{x}$ calculates the deformation $\\Delta \\pmb{x}$ at each timestamp $t$ , which can be used to warp a 3D point at $t$ to its corresponding position in the canonical space. In contrast, time-conditioned feature interpolation $F(\\pmb{x}, t) \\rightarrow \\pmb{f}$ directly learns a feature vector $\\pmb{f}$ from a given 3D point at $t$ , which is sent to a decoder for RGB and density calculation.", + "bbox": [ + 214, + 734, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y.Zhan, Z.Li et al.", + "bbox": [ + 271, + 114, + 397, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b587c683564864b326b922bb7e11c8da4bd82e7fe1666680f15827eaf33d1462.jpg", + "image_caption": [ + "Fig. 3: The overall pipeline of our KFD-NeRF. Our designed deformation field calculates final deformations based on two sources of knowledge. At the observation stage, system observations are directly output by a shallow observation MLP. At the prediction and fusion stage, we calculate predictions based on system dynamics, and further fuse observations and predictions to obtain final deformation estimations. At the spatial reconstruction stage, tri-plane encoded canonical points are concatenated with raw positions and timestamps, which are further decoded to obtain predicted colors for loss calculation. (*f represents a linear layer, used to obtain the Kalman gain from noise-related terms $\\varepsilon_{t_i}$ and $\\varepsilon_{t_{i - 1}}$ .)" + ], + "image_footnote": [], + "bbox": [ + 223, + 147, + 782, + 392 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 2, we illustrate the difference between these two methods with a motion example from time $t = 0$ to time $t = 1$ . Given a 3D point $P$ in the real world space, we can identify its corresponding four points in the canonical space at four different timestamps, named $P_{1}, P_{2}, P_{3}$ , and $P_{4}$ . ① shows the feature to be learned at $P$ which changes over time. Notice that when the radiance undergoes abrupt changes, the feature space exhibits many high-frequency signals, which are hard to fit. ② shows the backward deformation from the real world space to the canonical space to be learned at $P$ which changes over time. These signals are smoother and easier to fit by leveraging the motion's smoothness prior. MLP representations have inductive smoothness bias and are well-suited for learning such deformations. Hence, we employ MLP-based deformation fields to represent motion in this work.", + "bbox": [ + 212, + 556, + 787, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Kalman Filter Guided Motion Prediction", + "text_level": 1, + "bbox": [ + 215, + 756, + 602, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we derive the system equations for the dynamic radiance field following Eqs. (4) and (5). As we are trying to model non-rigid motions, each 3D point in the dynamic radiance field could be independently seen as a dynamic system, with its velocity $\\pmb{v}_{t_i}$ at each frame $t_i$ as a single input $u_t$ . Also note that", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter", + "bbox": [ + 328, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "this dynamic system is a single-state system since we only focus on estimating deformation at each frame $t_i$ , denoted as $d\\pmb{x}_{t_i}$ .", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the next step, we determine the coefficients matrix $A$ , $B$ , and $C$ in Eqs. (4) and (5), each degenerating to a value under a single-input and single-state system. As we directly observe $d\\pmb{x}_{t_i}$ , the output matrix $C$ becomes 1. Since the motion trajectory is unknown, we use a locally-linear motion model to describe this dynamic system: $d\\pmb{x}_{t_i} = d\\pmb{x}_{t_{i-1}} + \\Delta t_i \\cdot \\pmb{v}_{t_i}$ ( $A = 1$ and $B = \\Delta t_i$ ). Further considering noise, our state and output equations become", + "bbox": [ + 212, + 176, + 784, + 267 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} d \\boldsymbol {x} _ {t _ {i}} = d \\boldsymbol {x} _ {t _ {i - 1}} + \\Delta t _ {i} \\cdot \\boldsymbol {v} _ {t _ {i}} + n _ {t _ {i}}, \\\\ y _ {t _ {i}} = d \\boldsymbol {x} _ {t _ {i}} + m _ {t _ {i}}. \\end{array} \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 279, + 785, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 3, our designed deformation field consists of two parts: an MLP-based observer and a system-dynamics-based predictor. As for the observer, by querying 3D points coordinates $\\mathbf{x}$ and timestamps $t_i$ , a two-layer shallow MLP is used to output the mean observation $y_{t_i}$ (assuming zero-mean Gaussian measurement noise $m_{t_i}$ ) and noise-related terms $\\varepsilon_{t_i}$ . For the predictor, we follow the locally-linear motion model and first calculate the input $\\mathbf{v}_{t_i}$ . Because the deformation of the current frame $t_i$ is being estimated, we only use the information from frame $t_{i-1}$ and frame $t_{i-2}$ to approximate the velocity", + "bbox": [ + 212, + 328, + 784, + 449 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {t _ {i}} = \\left(d \\boldsymbol {x} _ {t _ {i - 1}} - d \\boldsymbol {x} _ {t _ {i - 2}}\\right) / \\Delta t _ {i}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 462, + 784, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Based on the estimated velocity, we compute the predicted deformation state $\\hat{d\\pmb{x}}_{t_i}^{-}$ . The final estimation $\\hat{d\\pmb{x}}_{t_i}$ is the sum of $\\hat{d\\pmb{x}}_{t_i}^{-}$ and $y_{t_i}$ weighted by a Kalman gain $K_{t_i}$ . The Kalman gain in Eq. (10) is calculated by current measurement noise and past process noise. In our implementation, we obtain $K_{t_i}$ utilizing a linear layer that takes noise-related terms $\\varepsilon_{t_i}$ and $\\varepsilon_{t_{i-1}}$ and timestamps $t_i$ and $t_{i-1}$ as inputs, incorporating historical information. Algorithm 1 summarizes this deformation estimation.", + "bbox": [ + 212, + 488, + 784, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Volume is warped to the canonical space based on the estimated deformation, which is further decoded to obtain density $\\sigma$ and color. We compute the loss by comparing the rendered values with the ground truth and use it to update the network, following Eq. (3). In addition, we note that the estimated deformation serves as a good supervision for learning the observation MLP, in analogy with observation updating process in Kalman filter. Therefore, we try to minimize the error between the current observation and the estimated deformation with", + "bbox": [ + 212, + 597, + 784, + 702 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {k f} = \\frac {1}{\\mathcal {N}} \\sum_ {\\boldsymbol {x} \\in \\mathcal {N}} \\left\\| y _ {t _ {i}} - \\hat {d} \\boldsymbol {x} _ {t _ {i}} \\right\\| _ {2} ^ {2}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 714, + 784, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The weighted $y_{t_i}$ and $\\hat{d\\pmb{x}}_{t_i}^{-}$ stand for two sources of knowledge from the whole system. $y_{t_i}$ represents the state directly observed by the observation MLP, which lacks the information from past frames and thus has a low confidence in the early stages of training. In contrast, $\\hat{d\\pmb{x}}_{t_i}^{-}$ represents the current frame information predicted from history, which provides a good prior in the early stages of training.", + "bbox": [ + 212, + 760, + 784, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y.Zhan, Z.Li et al.", + "bbox": [ + 271, + 114, + 397, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/706867a460b0ed58c33827051379fa4c80b4685ef4614f834b0520ffa27bf722.jpg", + "table_caption": [ + "Algorithm 1 Deformation Estimations" + ], + "table_footnote": [], + "table_body": "
Prediction:Update:
dxti-1, εti-1 = observer(x, ti-1)yti, εti = observer(x, ti)
dxti-2, εti-2 = observer(x, ti-2)Kti = f(εti, εti-1, ti, ti-1)
vti = (dxti-1 - dxti-2) / Δtidxti = dxti - Kti(yti - dxti - )
dxti = dxti-1 + Δti · vt = 2dxti-1 - dxti-2
", + "bbox": [ + 215, + 162, + 759, + 241 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the later stages of training, however, the confidence of the predictor drops significantly due to the inherent errors of the modeled motion dynamics, which is gradually compensated by the well learned observation MLP. Our Kalman Filter guided model automatically strikes the balance of both leveraging $\\hat{d\\mathbf{x}}_{t_i}^{-}$ in the early stages to accelerate convergence and avoid local minimum and also using $y_{t_i}$ in the later stages to learn fine details.", + "bbox": [ + 212, + 270, + 787, + 364 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Spatial Representation", + "text_level": 1, + "bbox": [ + 215, + 387, + 450, + 402 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "MLP-based spatial representations suffer from slow convergence and require a deeper deformation network (e.g., eight layers MLP in [36]) to model 4D scenes. Otherwise, the deformation field can get stuck in a local optimum before learning the radiance field in the canonical space. Voxel grids based spatial representations converge very quickly (e.g., TiNeuVox [10]) but demand high spatial resolution to store fine details, which requires significant memory footprint. Under memory constraints, achieving high resolution in the canonical space can be challenging, resulting in losing scene details and deteriorating the warping quality for each frame. We employ a tri-plane representation, which uses three sets of mutually orthogonal 2D planes to represent 3D space. This low-rank model allows for rapid convergence while significantly reducing memory consumption, making it possible to achieve fast and high-quality canonical space reconstruction.", + "bbox": [ + 212, + 412, + 787, + 594 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "What's more, there are some inevitable errors due to coordinate shifts when $\\pmb{x}'$ are encoded by finite resolution tri-plane. To enhance the raw coordinate information, we follow [10] by concatenating encoded tri-plane features with raw coordinate inputs.", + "bbox": [ + 212, + 594, + 787, + 656 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 Training Strategy and Regularization", + "text_level": 1, + "bbox": [ + 215, + 678, + 570, + 694 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our model takes into account temporal information so the learning of the current frame partially relies on the results of previous frames. To ensure that previous frames can offer ample priors, we employ a training strategy of gradually releasing the training images in chronological order.", + "bbox": [ + 212, + 704, + 787, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We also notice that the lack of constraints or priors in the canonical space can affect the performance of the observation branch, which could be regularized by pre-setting the shape in the canonical space. D-NeRF [36] sets frame $t_0$ to be the canonical space and force the deformation output of frame $t_0$ to be masked to 0. Such a hard mask, however, does not allow learning of the deformation", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter", + "bbox": [ + 326, + 113, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "at frame $t_0$ and may cause the canonical space to be too complex for motions warping.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We instead design a soft regularization term to normalize the difference between the canonical space and the radiance field in the real world at frame $t_0$ to improve the observation branch. The canonical observation loss of a points batch $\\mathcal{N}$ is", + "bbox": [ + 212, + 176, + 784, + 234 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c o} = \\frac {1}{\\mathcal {N}} \\sum_ {\\boldsymbol {x} \\in \\mathcal {N}} \\mathbf {1} (t) d \\boldsymbol {x}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 234, + 784, + 268 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\mathbf{1}(t) = 1$ when $t = 0$ and $\\mathbf{1}(t) = 0$ when $t\\neq 0$", + "bbox": [ + 214, + 275, + 593, + 290 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We use a proposal sampling strategy from Mip-NeRF 360 [2] by distilling the density field for occupancy estimation. This online distillation necessitates a loss function $\\mathcal{L}_{prop}$ to ensure consistency between the proposal network and our learned model. Please refer to Section 3 in [2] for detailed definition.", + "bbox": [ + 212, + 291, + 784, + 349 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Total variation loss $\\mathcal{L}_{tv}$ is a common regularizer in inverse problems, which encourages sparse edges in space. We apply this loss to each of our tri-plane to get $\\mathcal{L}_{tv}(\\mathbf{x}^{\\prime})$ , where $\\mathbf{x}^{\\prime}$ are warped 3D points in the canonical space. The total variation loss is", + "bbox": [ + 212, + 352, + 784, + 410 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t v} (\\mathbf {x}) = \\frac {1}{| C |} \\sum_ {c, i, j} \\left(\\left\\| \\mathbf {x} _ {c} ^ {i, j} - \\mathbf {x} _ {c} ^ {i - 1, j} \\right\\| _ {2} ^ {2} + \\left\\| \\mathbf {x} _ {c} ^ {i, j} - \\mathbf {x} _ {c} ^ {i, j - 1} \\right\\| _ {2} ^ {2}\\right), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 422, + 784, + 454 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $c$ is a certain plane from the tri-plane collection $C$ and $i, j$ are indices on the plane's resolution.", + "bbox": [ + 212, + 458, + 782, + 487 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The final loss function becomes", + "bbox": [ + 240, + 489, + 468, + 502 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {i m a g e}} + \\mathcal {L} _ {k f} + \\mathcal {L} _ {c o} + \\mathcal {L} _ {\\text {p r o p}} + \\lambda_ {t v} \\mathcal {L} _ {t v}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 521, + 782, + 536 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "and we experimentally choose $\\lambda_{tv}$ to be $1\\times 10^{-4}$ .", + "bbox": [ + 214, + 542, + 573, + 558 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experimental Results", + "text_level": 1, + "bbox": [ + 214, + 580, + 460, + 598 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Dataset", + "text_level": 1, + "bbox": [ + 214, + 614, + 326, + 628 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For synthetic data, we use the Dynamic NeRF Synthetic Dataset provided by D-NeRF [36], whose training and testing splits have already been well-organized. For each scene in the synthetic dataset, a photo from an arbitrary view with corresponding camera pose is provided at each timestamp. For real data, we use the Nvidia Real Dynamic Scenes Dataset [47] which consists of 8 dynamic scenes recorded by 12 synchronized cameras. We use 11 camera videos for training and the remaining one for testing.", + "bbox": [ + 212, + 640, + 784, + 746 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Baselines and Metrics", + "text_level": 1, + "bbox": [ + 214, + 768, + 442, + 782 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Due to differences in the format of synthetic and real data, we carefully select cutting-edge baselines to thoroughly validate our method based on comparative experiments. For synthetic data, we test deformation based methods", + "bbox": [ + 212, + 794, + 782, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y.Zhan, Z.Li et al.", + "bbox": [ + 271, + 114, + 397, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f0fd7359dd0441ddaecd72c7010e52aabe06fbd6820fa00e1b348d1a2f62f483.jpg", + "image_caption": [ + "Fig. 4: Qualitative Comparison of our KFD-NeRF against other dynamic NeRF methods on synthetic data. Zoom in for better details." + ], + "image_footnote": [], + "bbox": [ + 250, + 157, + 751, + 372 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/58c955a9b809652629e70559cf6ab53dca88c2dd7affd03183a3726483e2fe9d.jpg", + "image_caption": [ + "Fig. 5: Qualitative Comparison of our KFD-NeRF against other dynamic NeRF methods on real data. Zoom in for better details." + ], + "image_footnote": [], + "bbox": [ + 223, + 440, + 777, + 781 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter", + "bbox": [ + 328, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ee34d9b83a0b012c92a451d07b07ef976f16bb813e737c49d5def2c67f874465.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSynthetic Data
PSNR(dB)↑SSIM↑LPIPS↓Average↓
D-NeRF [36]30.250.9540.0760.0288
TiNeuVox-B [10]32.930.9730.0430.0173
KPlanes [12]30.470.9630.0720.0261
NDVG [17]30.510.9640.0560.0223
V4D [14]33.540.9780.0280.0141
4D-GS [44]33.310.9790.0260.0136
KFD-NeRF35.730.9840.0450.0131
", + "bbox": [ + 218, + 145, + 493, + 250 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/bcf3c3b07f6e52e580a429f62650b181ae038c0c826ebb223a119577bcd9827b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelReal Data
PSNR(dB)↑SSIM↑LPIPS↓Average↓
TiNeuVox-B [10]24.390.7340.3030.0869
KPlanes [12]28.250.8810.1370.0433
Mixvoxels-L [41]26.950.8280.2220.0619
V4D [14]28.160.8840.1360.0426
KFD-NeRF28.750.8910.1150.0389
", + "bbox": [ + 506, + 154, + 782, + 239 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/5ece2e7e4d85ede3788a72577e2960f5945bff51f9c8ca430baddb2f84d31831.jpg", + "table_caption": [ + "Table 1: Quantitative comparison of our KFD-NeRF against other dynamic NeRF methods. We show results on both synthetic and real data. See Sec. 5.2 for more details." + ], + "table_footnote": [], + "table_body": "
D-NeRFTiNeuVox-BKPlanesNDVGV4D4D-GSMixvoxels-LKFD-NeRF-SKFD-NeRF-L
Training Time↓48hrs40mins60mins35mins7hrs25mins1.5hrs30mins3hrs
Params(MB)↓1650340150275163140175175
", + "bbox": [ + 220, + 329, + 785, + 367 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2: Training time and parameter size for different methods.", + "bbox": [ + 281, + 376, + 718, + 391 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "D-NeRF [36] (MLP based spatial representation), TiNeuVox-B [10] (voxel grids based spatial representation), NDVG [17] (voxel grids based spatial representation) and 4D-GS [44] (Gaussian points based spatial representation), and feature interpolation based methods KPlanes [12] and V4D [14]. For real data, besides TiNeuVox-B and KPlanes, we further compare multi-view videos reconstruction methods MixVoxels [41]. We train all these methods on a single GeForce RTX3090. See Tab. 2 for detailed training time and parameters consumption of ours and other methods.", + "bbox": [ + 212, + 431, + 787, + 551 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We provide an exhaustive qualitative and quantitative comparison of our KFD-NeRF with these baseline methods. Three main metrics are reported, namely the peak signal-to-noise ratio (PSNR), the structural similarity index measure (SSIM) [42], and the learned perceptual image patch similarity (LPIPS) [51]. To provide more intuitive results, we further calculate metric \"average\" [1], which is the geometric mean of $\\mathrm{MSE} = 10^{-\\mathrm{PSNR} / 10}$ , $\\sqrt{1 - \\mathrm{SSIM}}$ , and LPIPS. Please see Tab. 1 for quantitative comparison and Figs. 4 and 5 for qualitative comparison. Per-scene results can be found in the supplemental material. We strongly recommend readers to watch the supplemental videos for a more intuitive understanding of the results.", + "bbox": [ + 212, + 553, + 787, + 705 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 724, + 401, + 739 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conduct ablation studies on both synthetic and real data to validate our various proposed system components. We compare our full model with variants related to $\\mathcal{L}_{co}$ , $\\mathcal{L}_{kf}$ and prediction branch.", + "bbox": [ + 212, + 750, + 785, + 795 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Canonical observation loss $\\mathcal{L}_{co}$ . This loss is designed to regularize a continuous and smooth volume shape in the canonical space for better observations. Specifically, we guide the shape in the canonical space to be close to the shape", + "bbox": [ + 212, + 795, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y.Zhan, Z.Li et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/cd705bcd6785d78392fc29bb7bd5776ba1be0f1a51535c4cb5b18492165995bc.jpg", + "image_caption": [ + "Fig. 6: Two sources of knowledge at frame $t_i$ from our pre-trained dynamic system with inputs up to and including frame $t_{i-1}$ . Notice that during the early stages of training, our observation MLP has insufficient initialization, resulting in the loss of many details (green boxes on the right). Fortunately, our predictions provide additional priors (green boxes on the left) for the current frame based on information from previous frames, compensating for the loss in observations." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 470, + 215 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/be19e179074b0b8bd52e65f67996a17fabedbb3b99dad15ca89c8053438f4cc7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LkfLcoPSNR(dB)↑SSIM↑LPIPS↓
a)30.150.9140.110
b)✓30.600.9200.106
c)✓31.780.9310.086
d)✓✓32.240.9380.080
", + "bbox": [ + 506, + 143, + 774, + 233 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/717e8e70efb40096f50d068fe062bbf926ab7bb1b6df9ba973c0b079f9106800.jpg", + "table_caption": [ + "Table 3: Ablation Studies on the $\\mathcal{L}_{kf}$ and $\\mathcal{L}_{co}$ . We show all the ablation combinations and report the average results on synthetic and real data." + ], + "table_footnote": [], + "table_body": "
ModelDynamicStatic
PSNR(dB)↑SSIM↑PSNR(dB)↑SSIM↑
P.N., (xi, ti) Input25.610.96329.580.812
P.N., (xi, ti, ti-1, ti-2) Input25.660.96429.550.812
w/ Prediction Branch (full)26.710.96729.570.813
", + "bbox": [ + 506, + 299, + 771, + 348 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 4: The ablation study on our prediction branch for real data in average, where dynamic and static areas are counted separately. P.N. is short for Pure MLP Network.", + "bbox": [ + 501, + 349, + 774, + 416 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "at frame $t_0$ . In Tab. 3 line $b$ , we remove $\\mathcal{L}_{co}$ and observe a decrease in model performance.", + "bbox": [ + 212, + 460, + 784, + 489 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Estimation update loss $\\mathcal{L}_{kf}$ . This loss tries to minimize the difference between estimated deformation and current observation. This updating process ensures that with each iteration, the observation acquire progressively more accurate information to guide the learning process. In Tab. 3 line $c$ ), we remove $\\mathcal{L}_{kf}$ and witness a decrease in model performance.", + "bbox": [ + 212, + 489, + 785, + 566 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Extra Studies on Prediction Branch", + "text_level": 1, + "bbox": [ + 215, + 588, + 558, + 602 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "This study aims to demonstrate the effectiveness of our plug-in Kalman filter. The most crucial step in Kalman filter is to fuse the original network observations with the predictions from system dynamics. Therefore, we directly remove the \"Prediction and Fusion Stage\" in our pipeline and use pure shallow observation MLP to generate deformations. We further conduct an experiment where a pure deformable network takes the current frame and the previous two states as inputs to ablate the effectiveness of the Kalman motion modeling. In Tab. 4, we see a clear drop in performance by pure MLP network-based methods, indicating that simply inputting previous Kalman filter parameters without modeling motion prediction is insufficient.", + "bbox": [ + 212, + 613, + 787, + 763 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The prediction branch we have designed can offer reliable priors in the early stages of training. We visualize these priors through an experiment to gain further insight. Specifically, we train our system with only inputs up to and including frame $t_{i-1}$ and then let our system produce rendering results at frame $t_i$ with two different branches: predictions and observations. This experiment", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter", + "bbox": [ + 328, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d745ebbb4ae374a9c051a1d4d6aba042933f85bde751b1b06031d5e63f3b81d5.jpg", + "image_caption": [ + "Fig. 7: An example in Nvidia Dataset [47] to illustrate scale and topological issues in choosing canonical space. From $①$ to $②$ there is a scale change due to moving closer to the camera. From $③$ to $②$ there is a topological change caused by target disappearance. According to this, choosing canonical space $①$ will suffer from low resolution and choosing canonical space $③$ will suffer from incomplete topology while $②$ is a better canonical space to choose." + ], + "image_footnote": [], + "bbox": [ + 223, + 147, + 779, + 224 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "effectively simulates the amount of knowledge contained in predictions and observations in the early stages of training when the system encounters new input data. In Fig. 6 we show two sources of knowledge from prediction branch and from direct output of observation MLP at frame $t_i$ . We see an obvious deficiency in the observation MLP when initializing new input frames and are pleased to find that the prediction branch compensates for this loss based on information from previous frames.", + "bbox": [ + 212, + 349, + 787, + 455 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 215, + 477, + 509, + 494 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. Our method relies on a well-reconstructed radiance field in the canonical space, which in our pipeline is guided by $\\mathcal{L}_{co}$ . This design, however, will partially fail, if the chosen canonical space exhibits significant scale changes or even topological changes related to other frames. In Fig. 7 we use an example to illustrate the influences caused by choosing canonical space. This phenomenon, however, cannot be mitigated by precisely setting the canonical space as we lack a priori access to the input data. We note that some works [1,33] focus on addressing scale or topological issues in radiance fields reconstruction. Nevertheless, these issues are not the main focus of our paper and will be explored in future works to further refine our model.", + "bbox": [ + 212, + 507, + 787, + 657 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Conclusion. In this work, we present KFD-NeRF, a Kalman filter guided NeRF, for 4D dynamic view synthesis. We model the dynamic radiance field as a dynamic system in control theory and use Kalman filter to estimate the deformation states based on both observations and predictions. We further enhance our observation by encoding canonical space with an efficient tri-plane and by regularizing the shape in the canonical space. Through our temporal training strategy and newly derived pipeline, KFD-NeRF achieves state-of-the-art view synthesis performance among a variety of dynamic NeRF methods. We hope the dynamic system modeling of 4D radiance fields will encourage researchers to explore motion contextual information. KFD-NeRF hopefully inspires the use of existing sequential methods mainly in control theory and visual state estimation to further improve 4D view synthesis and deformation estimations tasks.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y.Zhan, Z.Li et al.", + "bbox": [ + 271, + 114, + 397, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research was supported in part by JSPS KAKENHI Grant Numbers 24K22318, 22H00529, 20H05951, JST-Mirai Program JPMJMI23G1.", + "bbox": [ + 215, + 176, + 784, + 205 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 229, + 321, + 244 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A Multiscale Representation for Anti-aliasing Neural Radiance Fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)", + "2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: MipNeRF 360: Unbounded Anti-aliased Neural Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)", + "3. Becker, P., Pandya, H., Gebhardt, G., Zhao, C., Taylor, C.J., Neumann, G.: Recurrent Kalman Networks: Factorized Inference in High-dimensional Deep Feature Spaces. In: International conference on machine learning. pp. 544-552. PMLR (2019)", + "4. Cao, A., Johnson, J.: Hexplane: A Fast Representation for Dynamic Scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2023)", + "5. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient Geometry-aware 3D Generative Adversarial Networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)", + "6. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial Radiance Fields. In: European Conference on Computer Vision. pp. 333-350. Springer (2022)", + "7. Choe, J., Choy, C., Park, J., Kweon, I.S., Anandkumar, A.: Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17871-17881 (2023)", + "8. Coskun, H., Achilles, F., DiPietro, R., Navab, N., Tombari, F.: Long Short-term Memory Kalman Filters: Recurrent Neural Estimators for Pose Regularization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5524-5532 (2017)", + "9. Du, Y., Zhang, Y., Yu, H.X., Tenenbaum, J.B., Wu, J.: Neural Radiance Flow for 4D View Synthesis and Video Processing. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14304-14314. IEEE Computer Society (2021)", + "10. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Niefsner, M., Tian, Q.: Fast Dynamic Radiance Fields with Time-aware Neural Voxels. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)", + "1. Fraccaro, M., Kamronn, S., Paquet, U., Winther, O.: A Disentangled Recognition and Nonlinear Dynamics Model for Unsupervised Learning. Advances in neural information processing systems 30 (2017)", + "2. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit Radiance Fields in Space, Time, and Appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023)" + ], + "bbox": [ + 225, + 258, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter", + "bbox": [ + 326, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance Fields without Neural Networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)", + "14. Gan, W., Xu, H., Huang, Y., Chen, S., Yokoya, N.: V4D: Voxel for 4D Novel View Synthesis. IEEE Transactions on Visualization and Computer Graphics (2023)", + "15. Gao, C., Saraf, A., Kopf, J., Huang, J.B.: Dynamic View Synthesis from Dynamic Monocular Video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5712-5721 (2021)", + "16. Guen, V.L., Thome, N.: Disentangling Physical Dynamics from Unknown Factors for Unsupervised Video Prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11474-11484 (2020)", + "17. Guo, X., Chen, G., Dai, Y., Ye, X., Sun, J., Tan, X., Ding, E.: Neural Deformable Voxel Grid for Fast Optimization of Dynamic View Synthesis. In: Proceedings of the Asian Conference on Computer Vision. pp. 3757-3775 (2022)", + "18. Guo, X., Sun, J., Dai, Y., Chen, G., Ye, X., Tan, X., Ding, E., Zhang, Y., Wang, J.: Forward Flow for Novel View Synthesis of Dynamic Scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16022-16033 (2023)", + "19. Haarnoja, T., Ajay, A., Levine, S., Abbeel, P.: Backprop Kf: Learning Discriminative Deterministic State Estimators. Advances in neural information processing systems 29 (2016)", + "20. Hu, T., Liu, S., Chen, Y., Shen, T., Jia, J.: EfficientNeRF: Efficient Neural Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12902-12911 (2022)", + "21. Jang, H., Kim, D.: D-tensorf: Tensorial Radiance Fields for Dynamic Scenes. arXiv preprint arXiv:2212.02375 (2022)", + "22. Kajiya, J.T., Von Herzen, B.P.: Ray Tracing Volume Densities. ACM SIGGRAPH computer graphics 18(3), 165-174 (1984)", + "23. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D Gaussian Splatting for Real-time Radiance Field Rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023)", + "24. Li, T., Slavcheva, M., Zollhoefer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., et al.: Neural 3D Video Synthesis from Multi-view Video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5521-5531 (2022)", + "25. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural Scene Flow Fields for Space-time View Synthesis of Dynamic Scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6498-6508 (2021)", + "26. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural Sparse Voxel Fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)", + "27. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust Dynamic Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13-23 (2023)", + "28. Lu, G., Ouyang, W., Xu, D., Zhang, X., Gao, Z., Sun, M.T.: Deep Kalman Filtering Network for Video Compression Artifact Reduction. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 568-584 (2018)", + "29. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d Gaussians: Tracking by Persistent Dynamic View Synthesis. arXiv preprint arXiv:2308.09713 (2023)" + ], + "bbox": [ + 217, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y.Zhan, Z.Li et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "31. Müller, T., Evans, A., Schied, C., Keller, A.: Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)", + "32. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable Neural Radiance Fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021)", + "33. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A Higher-dimensional Representation for Topologically Varying Neural Radiance Fields. arXiv preprint arXiv:2106.13228 (2021)", + "34. Park, S., Son, M., Jang, S., Ahn, Y.C., Kim, J.Y., Kang, N.: Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4212-4221 (2023)", + "35. Peng, S., Niemeyer, M., Mescheder, L., Pollefeys, M., Geiger, A.: Convolutional Occupancy Networks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 523-540. Springer (2020)", + "36. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-neRF: Neural Radiance Fields for Dynamic Scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021)", + "37. Revach, G., Shlezinger, N., Ni, X., Escoriza, A.L., Van Sloun, R.J., Eldar, Y.C.: KalmanNet: Neural Network Aided Kalman Filtering for Partially Known Dynamics. IEEE Transactions on Signal Processing 70, 1532-1547 (2022)", + "38. Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., Liu, Y.: Tensor4d: Efficient Neural 4D Decomposition for High-fidelity Dynamic Reconstruction and Rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16632-16642 (2023)", + "39. Sun, C., Sun, M., Chen, H.T.: Direct Voxel Grid Optimization: Super-fast CO convergence for Radiance Fields Reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5459-5469 (2022)", + "40. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid Neural Radiance Fields: Reconstruction and Novel View Synthesis of a Dynamic Scene from Monocular Video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12959-12970 (2021)", + "41. Wang, F., Tan, S., Li, X., Tian, Z., Song, Y., Liu, H.: Mixed Neural Voxels for Fast Multi-view Video Synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19706-19716 (2023)", + "42. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image Quality Assessment: from Error Visibility to Structural Similarity. IEEE transactions on image processing 13(4), 600-612 (2004)", + "43. Welch, G., Bishop, G., et al.: An Introduction to the Kalman Filter (1995)", + "44. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering. arXiv preprint arXiv:2310.08528 (2023)", + "45. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time Neural Irradiance Fields for Free-viewpoint Video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9421-9431 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter", + "bbox": [ + 326, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "46. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction. arXiv preprint arXiv:2309.13101 (2023)", + "47. Yoon, J.S., Kim, K., Gallo, O., Park, H.S., Kautz, J.: Novel View Synthesis of Dynamic Scenes with Globally Coherent Depths from a Monocular Camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5336-5345 (2020)", + "48. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for Real-time Rendering of Neural Radiance Fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5752-5761 (2021)", + "49. Yuan, W., Lv, Z., Schmidt, T., Lovegrove, S.: Star: Self-supervised Tracking and Reconstruction of Rigid Objects in Motion with Neural Rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13144-13152 (2021)", + "50. Zhang, H., Yang, Z., Xiong, H., Zhu, T., Long, Z., Wu, W.: Transformer Aided Adaptive Extended Kalman Filter for Autonomous Vehicle Mass Estimation. Processes 11(3), 887 (2023)", + "51. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)", + "52. Zhou, L., Luo, Z., Shen, T., Zhang, J., Zhen, M., Yao, Y., Fang, T., Quan, L.: Kfnet: Learning Temporal Camera Relocalization Using Kalman Filtering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4919-4928 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 479 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y.Zhan, Z.Li et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/KFD-NeRF_ Rethinking Dynamic NeRF with Kalman Filter/3bae14dd-a29b-4634-bf7b-d258a05d1f55_model.json b/2024/KFD-NeRF_ Rethinking Dynamic NeRF with Kalman Filter/3bae14dd-a29b-4634-bf7b-d258a05d1f55_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c17093f055f1d69b03f652a2c3f16c3eee8a2540 --- /dev/null +++ b/2024/KFD-NeRF_ Rethinking Dynamic NeRF with Kalman Filter/3bae14dd-a29b-4634-bf7b-d258a05d1f55_model.json @@ -0,0 +1,2524 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.243, + 0.141, + 0.761, + 0.184 + ], + "angle": 0, + "content": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.213, + 0.782, + 0.245 + ], + "angle": 0, + "content": "Yifan Zhan\\(^{1}\\), Zhuoxiao Li\\(^{1}\\), Muyao Niu\\(^{1}\\), Zhihang Zhong\\(^{3}\\), Shohei Nobuhara\\(^{2}\\), Ko Nishino\\(^{2}\\), and Yinqiang Zheng\\(^{\\star 1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.413, + 0.255, + 0.59, + 0.269 + ], + "angle": 0, + "content": "1 The University of Tokyo" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.27, + 0.566, + 0.284 + ], + "angle": 0, + "content": "2 Kyoto University" + }, + { + "type": "text", + "bbox": [ + 0.353, + 0.284, + 0.651, + 0.298 + ], + "angle": 0, + "content": "3 Shanghai Artificial Intelligence Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.34, + 0.744, + 0.631 + ], + "angle": 0, + "content": "Abstract. We introduce KFD-NeRF, a novel dynamic neural radiance field integrated with an efficient and high-quality motion reconstruction framework based on Kalman filtering. Our key idea is to model the dynamic radiance field as a dynamic system whose temporally varying states are estimated based on two sources of knowledge: observations and predictions. We introduce a novel plug-in Kalman filter guided deformation field that enables accurate deformation estimation from scene observations and predictions. We use a shallow Multi-Layer Perceptron (MLP) for observations and model the motion as locally linear to calculate predictions with motion equations. To further enhance the performance of the observation MLP, we introduce regularization in the canonical space to facilitate the network's ability to learn warping for different frames. Additionally, we employ an efficient tri-plane representation for encoding the canonical space, which has been experimentally demonstrated to converge quickly with high quality. This enables us to use a shallower observation MLP, consisting of just two layers in our implementation. We conduct experiments on synthetic and real data and compare with past dynamic NeRF methods. Our KFD-NeRF demonstrates similar or even superior rendering performance within comparable computational time and achieves state-of-the-art view synthesis performance with thorough training. Github page: https://github.com/Yifever20002/KFD-NeRF." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.645, + 0.717, + 0.66 + ], + "angle": 0, + "content": "Keywords: Dynamic NeRF \\(\\cdot\\) Deformable Network \\(\\cdot\\) Kalman Filter" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.688, + 0.377, + 0.705 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.723, + 0.788, + 0.782 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRF) [30] have demonstrated outstanding success as a versatile and accurate 3D representation of real-world scenes, which has led to its wide adoption in daily and industrial applications in numerous domains. One of the remaining key desiderata for NeRFs is its extension to dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.784, + 0.79, + 0.815 + ], + "angle": 0, + "content": "Existing dynamic NeRF methods can be broadly categorized into two approaches. One is to learn deformation fields for motion warping (e.g., D-NeRF [36]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.826, + 0.436, + 0.84 + ], + "angle": 0, + "content": "* denotes corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.398, + 0.128 + ], + "angle": 0, + "content": "Y.Zhan, Z.Li et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.149, + 0.777, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.414, + 0.788, + 0.475 + ], + "angle": 0, + "content": "Fig. 1: In contrast to a vanilla deformation field, our plug-in Kalman filter guided deformation field consists of a prediction branch along with the direct observations from input data. From noise related terms \\(\\varepsilon_{t_i}\\) and \\(\\varepsilon_{t_{i - 1}}\\) we learn Kalman gain \\(K_{t_i}\\), weighting observations \\(y_{t_i}\\) and predictions \\(\\hat{d\\pmb{x}_{t_i}}\\) for more accurate deformation estimations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.581 + ], + "angle": 0, + "content": "and TiNeuVox [10]). Another is to disregard motion priors and directly interpolate time in the feature space (e.g., DyNeRF [24] and KPlanes [12]). These approaches, however, often overlook the characteristics of a dynamic radiance field as a time-state sequence, missing the opportunity to fully leverage temporal contextual information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.734 + ], + "angle": 0, + "content": "In this work, we draw inspirations from control theory and model the 4D radiance field as a dynamic system with temporally-varying states. Our state estimates of this dynamic system come from two sources of knowledge: observations based on the input data and predictions based on the system physical dynamics. Optimal state estimates cannot be achieved with only one of these two sources of knowledge. On the one hand, observations, as commonly used in previous dynamic NeRF works, are inherently error-prone due to the discrete temporal sampling of the dynamic scenes. On the other hand, predictions are governed by the correctness of the assumed kinematic model and may struggle to maintain accuracy for real dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To maximize combined potential of both observations and predictions, we introduce an efficient plug-and-play Kalman filter [43] module to optimize state estimations of our dynamic system. In Fig. 1 we illustrate our plug-in Kalman filter guided deformation field. We model the 4D radiance field as a single-state system, with the state denoted as \\(d\\pmb{x}_{t_i}\\) for the current frame deformation. In contrast to a vanilla deformation field that only considers system observations in the current frame, our approach incorporates richer information from previous" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.329, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "frames by introducing a prediction branch based on a motion equation. Given the absence of prior trajectory regarding the scene's motion, we employ local linear motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.787, + 0.281 + ], + "angle": 0, + "content": "Both observations and predictions are weighted using a learnable Kalman gain to calculate precise deformation estimations. During the initial stages of training, predictions primarily influence the process, facilitating the convergence of frames with substantial motion. In the later stages of training, observations take precedence, allowing for the recovery of more precise and fine-grained motion details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.403 + ], + "angle": 0, + "content": "All points in the real space are warped to a time-independent canonical space according to the estimated deformations. To further improve the performance of our two-branch deformable model, we employ an efficient tri-plane spatial representation for encoding our canonical space. Experimental evidence shows that this representation permits a shallower observation MLP with only two layers in our implementation. At the same time, we improve the warping ability of the observation MLP by regularizing the learning of the radiance field in the canonical space." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.404, + 0.568, + 0.419 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.428, + 0.784, + 0.471 + ], + "angle": 0, + "content": "1) The first method for modeling 4D radiance fields as dynamic systems by integrating a Kalman filter into the deformation field formulation, which results in a plug-in, efficient method for estimating deformations;" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.473, + 0.784, + 0.515 + ], + "angle": 0, + "content": "2) KFD-NeRF, a novel deformable NeRF with the Kalman filter plug-in and a tri-plane spatial representation, trained with a novel strategy of gradually releasing temporal information to facilitate learning dynamic systems:" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.518, + 0.784, + 0.56 + ], + "angle": 0, + "content": "3) and regularization in the canonical space for enhancing the learning capacity of a shallow observation MLP. We achieve state-of-the-art results on both synthetic and real data with all these designs, compared to dynamic NeRFs." + }, + { + "type": "list", + "bbox": [ + 0.221, + 0.428, + 0.784, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.584, + 0.396, + 0.6 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.784, + 0.688 + ], + "angle": 0, + "content": "Neural Radiance Fields Representation. NeRF [30] represents 3D scenes based on an implicit representation encoded by an MLP. Given multi-view images and corresponding camera poses as input, the MLP is trained by \"analysis by synthesis,\" achieving novel view synthesis and coarse geometry reconstruction. Vanilla NeRF leaves room for improvement that has attracted abundant research." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Many methods [13, 20, 26, 39, 48] use sparse voxel grids to represent the 3D scenes. While grid-based modeling achieves fast training and rendering, fine details require high-resolution grids leading to excessive memory consumption. The optimization process for grids is also unstable, failing to leverage the smoothness bias brought by MLP. Instant-NGP [31] proposes an efficient hash encoding for mapping spatial and feature domains. For sparsely-observed regions, however, this one-to-many mapping can easily introduce noise. The tri-plane representation [5, 6, 35] significantly reduces the memory footprint by leveraging its low-rank decomposition. We use this representation to build a fine-detailed canonical space." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.128 + ], + "angle": 0, + "content": "Y.Zhan, Z.Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.39 + ], + "angle": 0, + "content": "Dynamic Neural Rendering. For dynamic scenarios, modeling the time dimension together with the spatial representation becomes the main challenge. One approach to dynamic NeRF [4, 12, 14, 15, 21, 24, 25, 27, 34, 38, 45] is to add timestamps as an extra dimension to the 3D space and formulate 4D interpolation. Though intuitive, this leads to temporal incoherence due to the lack of supervision and priors between frames. Another approach [7, 9, 10, 17, 18, 25, 32, 33, 36, 40, 49] is to model the motion with deformation fields that warp points in arbitrary frames to the canonical space conditioned on the timestamps. This design can take advantage of the MLP smoothness bias in accordance with the motion smoothness prior. Nevertheless, deformation fields fail to find correspondences in frames with significant motion. We emphasize the performance improvement resulting from the switch in backbone spatial representation from D-NeRF [36] (MLP-based) to TiNeuVox [10] (voxel grids based) and later analyze the impact of different spatial representations on the deformation field. We also compare with latest point-based 4D rendering works [29, 44, 46] inspired by 3D Gaussian Splatting [23]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.788, + 0.614 + ], + "angle": 0, + "content": "Deep Kalman Filter. The combination of deep learning and Kalman filtering has been explored to address the challenges of incomplete observations in various scenarios. Some works [28, 52] use Convolutional Neural Network (CNN)-based Kalman filtering for video compression and camera localization. Recurrent Neural Network (RNN)-based Kalman filter has also been used in some studies [8, 37] to improve state estimation and to optimize pose regularization. The Transformer in [50] enables a more comprehensive exploitation of temporal contextual information. All of these works along with [3, 11, 19], however, model dynamics only in the learned latent space without taking into account physical priors, with the exception of [16] which fuses known physical priors and network outputs to construct a Kalman filter for video prediction. In this paper, for the first time, we employ a neural Kalman filter to assist in the dynamic NeRF task. We derive a two-stream method consisting of a shallow MLP and physical priors to achieve efficient motion estimation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.638, + 0.383, + 0.654 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.669, + 0.591, + 0.684 + ], + "angle": 0, + "content": "3.1 NeRF and Volume Rendering Revisited" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.694, + 0.788, + 0.801 + ], + "angle": 0, + "content": "NeRF [30] consists of three parts: sampling, volume mapping, and rendering. In the sampling process, points \\( \\pmb{x} \\in \\mathbb{R}^3 \\) are sampled along rays calculated from the camera pose. Then, in the volume mapping process, each \\( \\pmb{x} \\) in the 3D world, as well as its viewing direction \\( \\pmb{v} \\in \\mathbb{R}^3 \\), is queried to output the volume density \\( \\sigma \\) and radiance \\( \\pmb{c} = (r, g, b) \\) of \\( \\pmb{x} \\). Finally, in the rendering process, the color of each ray is computed with volume rendering [22]. The expected color of the ray \\( r(s) = o + sv \\) becomes [30]" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.811, + 0.786, + 0.844 + ], + "angle": 0, + "content": "\\[\nC (\\boldsymbol {r}) = \\int_ {s _ {n}} ^ {s _ {f}} T (s) \\sigma (\\boldsymbol {r} (s)) \\boldsymbol {c} (\\boldsymbol {r} (s), \\boldsymbol {v}) d s, \\tag {1}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.328, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.673, + 0.163 + ], + "angle": 0, + "content": "where \\( s_n \\) and \\( s_f \\) are the near and far bounds, respectively, and" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.175, + 0.786, + 0.209 + ], + "angle": 0, + "content": "\\[\nT (s) = \\exp \\left(- \\int_ {s _ {n}} ^ {s} \\sigma (\\boldsymbol {r} (k)) d k\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.218, + 0.784, + 0.25 + ], + "angle": 0, + "content": "The only supervision for training NeRF comes from the ground truth color \\( C_{gt}(\\boldsymbol{r}) \\) of each ray" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.261, + 0.786, + 0.293 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {i m a g e}} = \\sum_ {\\boldsymbol {r} \\in \\mathcal {R}} \\| C (\\boldsymbol {r}) - C _ {g t} (\\boldsymbol {r}) \\| _ {2} ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.305, + 0.403, + 0.319 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}\\) is the ray batch." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.343, + 0.381, + 0.357 + ], + "angle": 0, + "content": "3.2 Kalman Filter" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.368, + 0.787, + 0.413 + ], + "angle": 0, + "content": "Consider a dynamic system with input \\( u_{t} \\), output \\( y_{t} \\), process noise \\( n_t \\) and measurement noise \\( m_t \\). We want to obtain its state \\( x_{t} \\) at each frame \\( t \\), assuming the state equation follows" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.428, + 0.786, + 0.442 + ], + "angle": 0, + "content": "\\[\nx _ {t} = A x _ {t - 1} + B u _ {t} + n _ {t}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.454, + 0.452, + 0.469 + ], + "angle": 0, + "content": "and the output equation follows" + }, + { + "type": "equation", + "bbox": [ + 0.447, + 0.483, + 0.786, + 0.497 + ], + "angle": 0, + "content": "\\[\ny _ {t} = C x _ {t} + m _ {t}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.784, + 0.539 + ], + "angle": 0, + "content": "where \\( A \\), \\( B \\) and \\( C \\) are the system matrix, control matrix and output matrix, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.785, + 0.614 + ], + "angle": 0, + "content": "There are two methods for obtaining the system's state \\( x_{t} \\) at any given \\( t \\): observation and prediction. The observation method tries to rely on \\( y_{t} \\) in Eq. (5) while the prediction method tries to mathematically model the system dynamics to predict the state. Both of these methods, however, come with non-negligible errors due to the existence of \\( m_{t} \\) and \\( n_{t} \\) and for inaccurate system modeling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.785, + 0.719 + ], + "angle": 0, + "content": "The Kalman filter algorithm posits that the state \\( x_{t} \\) is obtained by a weighted sum of the prediction based on the state \\( x_{t-1} \\) and the observation \\( y_{t} \\). For a specific \\( t \\), estimating the state \\( x_{t} \\) can be divided into two steps: prediction step and update step. Assuming that the process noise and measurement noise follow zero-mean Gaussian distributions with variances \\( Q \\) and \\( R \\), respectively, at the prediction step, it first calculates a predicted state based on the estimated state at \\( t-1 \\)" + }, + { + "type": "equation", + "bbox": [ + 0.431, + 0.722, + 0.785, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {t} ^ {-} = A \\hat {x} _ {t - 1} + B u _ {t}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.745, + 0.39, + 0.758 + ], + "angle": 0, + "content": "and its error covariance" + }, + { + "type": "equation", + "bbox": [ + 0.426, + 0.759, + 0.785, + 0.775 + ], + "angle": 0, + "content": "\\[\nP _ {t} ^ {-} = A P _ {t - 1} A ^ {T} + Q. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.782, + 0.784, + 0.813 + ], + "angle": 0, + "content": "In the update step, this predicted state is combined with observation \\( y_{t} \\) to obtain updated state estimate" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.826, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {t} = \\hat {x} _ {t} ^ {-} + K _ {t} \\left(y _ {t} - C \\hat {x} _ {t} ^ {-}\\right), \\tag {8}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.398, + 0.128 + ], + "angle": 0, + "content": "Y.Zhan, Z.Li et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.149, + 0.462, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.158, + 0.775, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.293, + 0.789, + 0.364 + ], + "angle": 0, + "content": "Fig. 2: Visualization of feature planes learned by the feature interpolation method ① and the deformation fields method ②. We show a point \\( P \\) in the real world space and its corresponding four points \\( P_{1}, P_{2}, P_{3} \\) and \\( P_{4} \\) in the canonical space at four different timestamps. The feature plane in ② exhibits better smoothness compared to ①, so we use ② to construct our backbone." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.395, + 0.39, + 0.408 + ], + "angle": 0, + "content": "and its error covariance" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.41, + 0.785, + 0.427 + ], + "angle": 0, + "content": "\\[\nP _ {t} = \\left(I - K _ {t} C\\right) P _ {t} ^ {-}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.435, + 0.264, + 0.449 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.431, + 0.446, + 0.785, + 0.481 + ], + "angle": 0, + "content": "\\[\nK _ {t} = \\frac {P _ {t} ^ {-} C ^ {T}}{C P _ {t} ^ {-} C ^ {T} + R} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.487, + 0.785, + 0.518 + ], + "angle": 0, + "content": "is the Kalman gain, reflecting the weights assigned to the observation and prediction components." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.543, + 0.331, + 0.559 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.577, + 0.788, + 0.669 + ], + "angle": 0, + "content": "Figure 3 illustrates the three stages of the complete pipeline of KFD-NeRF. In this section, we will first analyze the advantages of using deformation fields as the motion representation over feature interpolation. We will then introduce KFD-NeRF based on the Kalman filter to achieve accurate deformation estimations. Finally, we will discuss spatial reconstruction details, the training strategy, and the incorporation of regularization." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.692, + 0.727, + 0.723 + ], + "angle": 0, + "content": "4.1 Motion Representation: Deformation Fields vs. Feature Interpolation" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "4D NeRFs generally employ two approaches to temporal modeling: motion deformation fields or time-conditioned feature interpolation. A deformation field \\( D(\\pmb{x}, t) \\rightarrow \\Delta \\pmb{x} \\) calculates the deformation \\( \\Delta \\pmb{x} \\) at each timestamp \\( t \\), which can be used to warp a 3D point at \\( t \\) to its corresponding position in the canonical space. In contrast, time-conditioned feature interpolation \\( F(\\pmb{x}, t) \\rightarrow \\pmb{f} \\) directly learns a feature vector \\( \\pmb{f} \\) from a given 3D point at \\( t \\), which is sent to a decoder for RGB and density calculation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.329, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.148, + 0.784, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.403, + 0.788, + 0.53 + ], + "angle": 0, + "content": "Fig. 3: The overall pipeline of our KFD-NeRF. Our designed deformation field calculates final deformations based on two sources of knowledge. At the observation stage, system observations are directly output by a shallow observation MLP. At the prediction and fusion stage, we calculate predictions based on system dynamics, and further fuse observations and predictions to obtain final deformation estimations. At the spatial reconstruction stage, tri-plane encoded canonical points are concatenated with raw positions and timestamps, which are further decoded to obtain predicted colors for loss calculation. (*f represents a linear layer, used to obtain the Kalman gain from noise-related terms \\(\\varepsilon_{t_i}\\) and \\(\\varepsilon_{t_{i - 1}}\\).)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.788, + 0.737 + ], + "angle": 0, + "content": "In Fig. 2, we illustrate the difference between these two methods with a motion example from time \\( t = 0 \\) to time \\( t = 1 \\). Given a 3D point \\( P \\) in the real world space, we can identify its corresponding four points in the canonical space at four different timestamps, named \\( P_{1}, P_{2}, P_{3} \\), and \\( P_{4} \\). ① shows the feature to be learned at \\( P \\) which changes over time. Notice that when the radiance undergoes abrupt changes, the feature space exhibits many high-frequency signals, which are hard to fit. ② shows the backward deformation from the real world space to the canonical space to be learned at \\( P \\) which changes over time. These signals are smoother and easier to fit by leveraging the motion's smoothness prior. MLP representations have inductive smoothness bias and are well-suited for learning such deformations. Hence, we employ MLP-based deformation fields to represent motion in this work." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.757, + 0.603, + 0.772 + ], + "angle": 0, + "content": "4.2 Kalman Filter Guided Motion Prediction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Next, we derive the system equations for the dynamic radiance field following Eqs. (4) and (5). As we are trying to model non-rigid motions, each 3D point in the dynamic radiance field could be independently seen as a dynamic system, with its velocity \\(\\pmb{v}_{t_i}\\) at each frame \\(t_i\\) as a single input \\(u_t\\). Also note that" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.398, + 0.128 + ], + "angle": 0, + "content": "Y.Zhan, Z.Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "this dynamic system is a single-state system since we only focus on estimating deformation at each frame \\( t_i \\), denoted as \\( d\\pmb{x}_{t_i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.268 + ], + "angle": 0, + "content": "In the next step, we determine the coefficients matrix \\(A\\), \\(B\\), and \\(C\\) in Eqs. (4) and (5), each degenerating to a value under a single-input and single-state system. As we directly observe \\(d\\pmb{x}_{t_i}\\), the output matrix \\(C\\) becomes 1. Since the motion trajectory is unknown, we use a locally-linear motion model to describe this dynamic system: \\(d\\pmb{x}_{t_i} = d\\pmb{x}_{t_{i-1}} + \\Delta t_i \\cdot \\pmb{v}_{t_i}\\) (\\(A = 1\\) and \\(B = \\Delta t_i\\)). Further considering noise, our state and output equations become" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.28, + 0.786, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} d \\boldsymbol {x} _ {t _ {i}} = d \\boldsymbol {x} _ {t _ {i - 1}} + \\Delta t _ {i} \\cdot \\boldsymbol {v} _ {t _ {i}} + n _ {t _ {i}}, \\\\ y _ {t _ {i}} = d \\boldsymbol {x} _ {t _ {i}} + m _ {t _ {i}}. \\end{array} \\right. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.785, + 0.45 + ], + "angle": 0, + "content": "In Fig. 3, our designed deformation field consists of two parts: an MLP-based observer and a system-dynamics-based predictor. As for the observer, by querying 3D points coordinates \\(\\mathbf{x}\\) and timestamps \\(t_i\\), a two-layer shallow MLP is used to output the mean observation \\(y_{t_i}\\) (assuming zero-mean Gaussian measurement noise \\(m_{t_i}\\)) and noise-related terms \\(\\varepsilon_{t_i}\\). For the predictor, we follow the locally-linear motion model and first calculate the input \\(\\mathbf{v}_{t_i}\\). Because the deformation of the current frame \\(t_i\\) is being estimated, we only use the information from frame \\(t_{i-1}\\) and frame \\(t_{i-2}\\) to approximate the velocity" + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.463, + 0.785, + 0.478 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {t _ {i}} = \\left(d \\boldsymbol {x} _ {t _ {i - 1}} - d \\boldsymbol {x} _ {t _ {i - 2}}\\right) / \\Delta t _ {i}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.785, + 0.596 + ], + "angle": 0, + "content": "Based on the estimated velocity, we compute the predicted deformation state \\( \\hat{d\\pmb{x}}_{t_i}^{-} \\). The final estimation \\( \\hat{d\\pmb{x}}_{t_i} \\) is the sum of \\( \\hat{d\\pmb{x}}_{t_i}^{-} \\) and \\( y_{t_i} \\) weighted by a Kalman gain \\( K_{t_i} \\). The Kalman gain in Eq. (10) is calculated by current measurement noise and past process noise. In our implementation, we obtain \\( K_{t_i} \\) utilizing a linear layer that takes noise-related terms \\( \\varepsilon_{t_i} \\) and \\( \\varepsilon_{t_{i-1}} \\) and timestamps \\( t_i \\) and \\( t_{i-1} \\) as inputs, incorporating historical information. Algorithm 1 summarizes this deformation estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.785, + 0.703 + ], + "angle": 0, + "content": "Volume is warped to the canonical space based on the estimated deformation, which is further decoded to obtain density \\(\\sigma\\) and color. We compute the loss by comparing the rendered values with the ground truth and use it to update the network, following Eq. (3). In addition, we note that the estimated deformation serves as a good supervision for learning the observation MLP, in analogy with observation updating process in Kalman filter. Therefore, we try to minimize the error between the current observation and the estimated deformation with" + }, + { + "type": "equation", + "bbox": [ + 0.396, + 0.715, + 0.785, + 0.75 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {k f} = \\frac {1}{\\mathcal {N}} \\sum_ {\\boldsymbol {x} \\in \\mathcal {N}} \\left\\| y _ {t _ {i}} - \\hat {d} \\boldsymbol {x} _ {t _ {i}} \\right\\| _ {2} ^ {2}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.761, + 0.785, + 0.841 + ], + "angle": 0, + "content": "The weighted \\( y_{t_i} \\) and \\( \\hat{d\\pmb{x}}_{t_i}^{-} \\) stand for two sources of knowledge from the whole system. \\( y_{t_i} \\) represents the state directly observed by the observation MLP, which lacks the information from past frames and thus has a low confidence in the early stages of training. In contrast, \\( \\hat{d\\pmb{x}}_{t_i}^{-} \\) represents the current frame information predicted from history, which provides a good prior in the early stages of training." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.328, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.506, + 0.162 + ], + "angle": 0, + "content": "Algorithm 1 Deformation Estimations" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.163, + 0.761, + 0.242 + ], + "angle": 0, + "content": "
Prediction:Update:
dxti-1, εti-1 = observer(x, ti-1)yti, εti = observer(x, ti)
dxti-2, εti-2 = observer(x, ti-2)Kti = f(εti, εti-1, ti, ti-1)
vti = (dxti-1 - dxti-2) / Δtidxti = dxti - Kti(yti - dxti - )
dxti = dxti-1 + Δti · vt = 2dxti-1 - dxti-2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.366 + ], + "angle": 0, + "content": "In the later stages of training, however, the confidence of the predictor drops significantly due to the inherent errors of the modeled motion dynamics, which is gradually compensated by the well learned observation MLP. Our Kalman Filter guided model automatically strikes the balance of both leveraging \\( \\hat{d\\mathbf{x}}_{t_i}^{-} \\) in the early stages to accelerate convergence and avoid local minimum and also using \\( y_{t_i} \\) in the later stages to learn fine details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.388, + 0.452, + 0.404 + ], + "angle": 0, + "content": "4.3 Spatial Representation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.789, + 0.595 + ], + "angle": 0, + "content": "MLP-based spatial representations suffer from slow convergence and require a deeper deformation network (e.g., eight layers MLP in [36]) to model 4D scenes. Otherwise, the deformation field can get stuck in a local optimum before learning the radiance field in the canonical space. Voxel grids based spatial representations converge very quickly (e.g., TiNeuVox [10]) but demand high spatial resolution to store fine details, which requires significant memory footprint. Under memory constraints, achieving high resolution in the canonical space can be challenging, resulting in losing scene details and deteriorating the warping quality for each frame. We employ a tri-plane representation, which uses three sets of mutually orthogonal 2D planes to represent 3D space. This low-rank model allows for rapid convergence while significantly reducing memory consumption, making it possible to achieve fast and high-quality canonical space reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.789, + 0.657 + ], + "angle": 0, + "content": "What's more, there are some inevitable errors due to coordinate shifts when \\( \\pmb{x}' \\) are encoded by finite resolution tri-plane. To enhance the raw coordinate information, we follow [10] by concatenating encoded tri-plane features with raw coordinate inputs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.679, + 0.571, + 0.695 + ], + "angle": 0, + "content": "4.4 Training Strategy and Regularization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.765 + ], + "angle": 0, + "content": "Our model takes into account temporal information so the learning of the current frame partially relies on the results of previous frames. To ensure that previous frames can offer ample priors, we employ a training strategy of gradually releasing the training images in chronological order." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We also notice that the lack of constraints or priors in the canonical space can affect the performance of the observation branch, which could be regularized by pre-setting the shape in the canonical space. D-NeRF [36] sets frame \\( t_0 \\) to be the canonical space and force the deformation output of frame \\( t_0 \\) to be masked to 0. Such a hard mask, however, does not allow learning of the deformation" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.398, + 0.128 + ], + "angle": 0, + "content": "Y.Zhan, Z.Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "at frame \\( t_0 \\) and may cause the canonical space to be too complex for motions warping." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.235 + ], + "angle": 0, + "content": "We instead design a soft regularization term to normalize the difference between the canonical space and the radiance field in the real world at frame \\( t_0 \\) to improve the observation branch. The canonical observation loss of a points batch \\( \\mathcal{N} \\) is" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.235, + 0.785, + 0.269 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c o} = \\frac {1}{\\mathcal {N}} \\sum_ {\\boldsymbol {x} \\in \\mathcal {N}} \\mathbf {1} (t) d \\boldsymbol {x}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.276, + 0.594, + 0.291 + ], + "angle": 0, + "content": "where \\(\\mathbf{1}(t) = 1\\) when \\(t = 0\\) and \\(\\mathbf{1}(t) = 0\\) when \\(t\\neq 0\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.292, + 0.785, + 0.351 + ], + "angle": 0, + "content": "We use a proposal sampling strategy from Mip-NeRF 360 [2] by distilling the density field for occupancy estimation. This online distillation necessitates a loss function \\(\\mathcal{L}_{prop}\\) to ensure consistency between the proposal network and our learned model. Please refer to Section 3 in [2] for detailed definition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.785, + 0.411 + ], + "angle": 0, + "content": "Total variation loss \\(\\mathcal{L}_{tv}\\) is a common regularizer in inverse problems, which encourages sparse edges in space. We apply this loss to each of our tri-plane to get \\(\\mathcal{L}_{tv}(\\mathbf{x}^{\\prime})\\), where \\(\\mathbf{x}^{\\prime}\\) are warped 3D points in the canonical space. The total variation loss is" + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.424, + 0.785, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t v} (\\mathbf {x}) = \\frac {1}{| C |} \\sum_ {c, i, j} \\left(\\left\\| \\mathbf {x} _ {c} ^ {i, j} - \\mathbf {x} _ {c} ^ {i - 1, j} \\right\\| _ {2} ^ {2} + \\left\\| \\mathbf {x} _ {c} ^ {i, j} - \\mathbf {x} _ {c} ^ {i, j - 1} \\right\\| _ {2} ^ {2}\\right), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.784, + 0.488 + ], + "angle": 0, + "content": "where \\(c\\) is a certain plane from the tri-plane collection \\(C\\) and \\(i, j\\) are indices on the plane's resolution." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.49, + 0.47, + 0.503 + ], + "angle": 0, + "content": "The final loss function becomes" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.522, + 0.784, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {i m a g e}} + \\mathcal {L} _ {k f} + \\mathcal {L} _ {c o} + \\mathcal {L} _ {\\text {p r o p}} + \\lambda_ {t v} \\mathcal {L} _ {t v}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.543, + 0.575, + 0.559 + ], + "angle": 0, + "content": "and we experimentally choose \\(\\lambda_{tv}\\) to be \\(1\\times 10^{-4}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.582, + 0.462, + 0.599 + ], + "angle": 0, + "content": "5 Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.615, + 0.327, + 0.629 + ], + "angle": 0, + "content": "5.1 Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.785, + 0.747 + ], + "angle": 0, + "content": "For synthetic data, we use the Dynamic NeRF Synthetic Dataset provided by D-NeRF [36], whose training and testing splits have already been well-organized. For each scene in the synthetic dataset, a photo from an arbitrary view with corresponding camera pose is provided at each timestamp. For real data, we use the Nvidia Real Dynamic Scenes Dataset [47] which consists of 8 dynamic scenes recorded by 12 synchronized cameras. We use 11 camera videos for training and the remaining one for testing." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.769, + 0.444, + 0.784 + ], + "angle": 0, + "content": "5.2 Baselines and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.841 + ], + "angle": 0, + "content": "Due to differences in the format of synthetic and real data, we carefully select cutting-edge baselines to thoroughly validate our method based on comparative experiments. For synthetic data, we test deformation based methods" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.329, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.158, + 0.752, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.386, + 0.785, + 0.413 + ], + "angle": 0, + "content": "Fig. 4: Qualitative Comparison of our KFD-NeRF against other dynamic NeRF methods on synthetic data. Zoom in for better details." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.441, + 0.779, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.801, + 0.785, + 0.827 + ], + "angle": 0, + "content": "Fig. 5: Qualitative Comparison of our KFD-NeRF against other dynamic NeRF methods on real data. Zoom in for better details." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.128 + ], + "angle": 0, + "content": "Y.Zhan, Z.Li et al." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.146, + 0.495, + 0.25 + ], + "angle": 0, + "content": "
ModelSynthetic Data
PSNR(dB)↑SSIM↑LPIPS↓Average↓
D-NeRF [36]30.250.9540.0760.0288
TiNeuVox-B [10]32.930.9730.0430.0173
KPlanes [12]30.470.9630.0720.0261
NDVG [17]30.510.9640.0560.0223
V4D [14]33.540.9780.0280.0141
4D-GS [44]33.310.9790.0260.0136
KFD-NeRF35.730.9840.0450.0131
" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.155, + 0.783, + 0.24 + ], + "angle": 0, + "content": "
ModelReal Data
PSNR(dB)↑SSIM↑LPIPS↓Average↓
TiNeuVox-B [10]24.390.7340.3030.0869
KPlanes [12]28.250.8810.1370.0433
Mixvoxels-L [41]26.950.8280.2220.0619
V4D [14]28.160.8840.1360.0426
KFD-NeRF28.750.8910.1150.0389
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.263, + 0.788, + 0.304 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison of our KFD-NeRF against other dynamic NeRF methods. We show results on both synthetic and real data. See Sec. 5.2 for more details." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.33, + 0.787, + 0.368 + ], + "angle": 0, + "content": "
D-NeRFTiNeuVox-BKPlanesNDVGV4D4D-GSMixvoxels-LKFD-NeRF-SKFD-NeRF-L
Training Time↓48hrs40mins60mins35mins7hrs25mins1.5hrs30mins3hrs
Params(MB)↓1650340150275163140175175
" + }, + { + "type": "table_caption", + "bbox": [ + 0.282, + 0.377, + 0.72, + 0.392 + ], + "angle": 0, + "content": "Table 2: Training time and parameter size for different methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.553 + ], + "angle": 0, + "content": "D-NeRF [36] (MLP based spatial representation), TiNeuVox-B [10] (voxel grids based spatial representation), NDVG [17] (voxel grids based spatial representation) and 4D-GS [44] (Gaussian points based spatial representation), and feature interpolation based methods KPlanes [12] and V4D [14]. For real data, besides TiNeuVox-B and KPlanes, we further compare multi-view videos reconstruction methods MixVoxels [41]. We train all these methods on a single GeForce RTX3090. See Tab. 2 for detailed training time and parameters consumption of ours and other methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.706 + ], + "angle": 0, + "content": "We provide an exhaustive qualitative and quantitative comparison of our KFD-NeRF with these baseline methods. Three main metrics are reported, namely the peak signal-to-noise ratio (PSNR), the structural similarity index measure (SSIM) [42], and the learned perceptual image patch similarity (LPIPS) [51]. To provide more intuitive results, we further calculate metric \"average\" [1], which is the geometric mean of \\(\\mathrm{MSE} = 10^{-\\mathrm{PSNR} / 10}\\), \\(\\sqrt{1 - \\mathrm{SSIM}}\\), and LPIPS. Please see Tab. 1 for quantitative comparison and Figs. 4 and 5 for qualitative comparison. Per-scene results can be found in the supplemental material. We strongly recommend readers to watch the supplemental videos for a more intuitive understanding of the results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.726, + 0.403, + 0.74 + ], + "angle": 0, + "content": "5.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.796 + ], + "angle": 0, + "content": "We conduct ablation studies on both synthetic and real data to validate our various proposed system components. We compare our full model with variants related to \\(\\mathcal{L}_{co}\\), \\(\\mathcal{L}_{kf}\\) and prediction branch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Canonical observation loss \\(\\mathcal{L}_{co}\\). This loss is designed to regularize a continuous and smooth volume shape in the canonical space for better observations. Specifically, we guide the shape in the canonical space to be close to the shape" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.329, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.145, + 0.472, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.219, + 0.477, + 0.398 + ], + "angle": 0, + "content": "Fig. 6: Two sources of knowledge at frame \\( t_i \\) from our pre-trained dynamic system with inputs up to and including frame \\( t_{i-1} \\). Notice that during the early stages of training, our observation MLP has insufficient initialization, resulting in the loss of many details (green boxes on the right). Fortunately, our predictions provide additional priors (green boxes on the left) for the current frame based on information from previous frames, compensating for the loss in observations." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.145, + 0.775, + 0.234 + ], + "angle": 0, + "content": "
LkfLcoPSNR(dB)↑SSIM↑LPIPS↓
a)30.150.9140.110
b)✓30.600.9200.106
c)✓31.780.9310.086
d)✓✓32.240.9380.080
" + }, + { + "type": "table_caption", + "bbox": [ + 0.503, + 0.237, + 0.776, + 0.293 + ], + "angle": 0, + "content": "Table 3: Ablation Studies on the \\(\\mathcal{L}_{kf}\\) and \\(\\mathcal{L}_{co}\\). We show all the ablation combinations and report the average results on synthetic and real data." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.3, + 0.772, + 0.349 + ], + "angle": 0, + "content": "
ModelDynamicStatic
PSNR(dB)↑SSIM↑PSNR(dB)↑SSIM↑
P.N., (xi, ti) Input25.610.96329.580.812
P.N., (xi, ti, ti-1, ti-2) Input25.660.96429.550.812
w/ Prediction Branch (full)26.710.96729.570.813
" + }, + { + "type": "table_caption", + "bbox": [ + 0.502, + 0.35, + 0.776, + 0.417 + ], + "angle": 0, + "content": "Table 4: The ablation study on our prediction branch for real data in average, where dynamic and static areas are counted separately. P.N. is short for Pure MLP Network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.785, + 0.49 + ], + "angle": 0, + "content": "at frame \\( t_0 \\). In Tab. 3 line \\( b \\), we remove \\( \\mathcal{L}_{co} \\) and observe a decrease in model performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.787, + 0.567 + ], + "angle": 0, + "content": "Estimation update loss \\(\\mathcal{L}_{kf}\\). This loss tries to minimize the difference between estimated deformation and current observation. This updating process ensures that with each iteration, the observation acquire progressively more accurate information to guide the learning process. In Tab. 3 line \\(c\\)), we remove \\(\\mathcal{L}_{kf}\\) and witness a decrease in model performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.589, + 0.559, + 0.603 + ], + "angle": 0, + "content": "5.4 Extra Studies on Prediction Branch" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.764 + ], + "angle": 0, + "content": "This study aims to demonstrate the effectiveness of our plug-in Kalman filter. The most crucial step in Kalman filter is to fuse the original network observations with the predictions from system dynamics. Therefore, we directly remove the \"Prediction and Fusion Stage\" in our pipeline and use pure shallow observation MLP to generate deformations. We further conduct an experiment where a pure deformable network takes the current frame and the previous two states as inputs to ablate the effectiveness of the Kalman motion modeling. In Tab. 4, we see a clear drop in performance by pure MLP network-based methods, indicating that simply inputting previous Kalman filter parameters without modeling motion prediction is insufficient." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The prediction branch we have designed can offer reliable priors in the early stages of training. We visualize these priors through an experiment to gain further insight. Specifically, we train our system with only inputs up to and including frame \\( t_{i-1} \\) and then let our system produce rendering results at frame \\( t_i \\) with two different branches: predictions and observations. This experiment" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.398, + 0.128 + ], + "angle": 0, + "content": "Y.Zhan, Z.Li et al." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.148, + 0.781, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.239, + 0.788, + 0.323 + ], + "angle": 0, + "content": "Fig. 7: An example in Nvidia Dataset [47] to illustrate scale and topological issues in choosing canonical space. From \\(①\\) to \\(②\\) there is a scale change due to moving closer to the camera. From \\(③\\) to \\(②\\) there is a topological change caused by target disappearance. According to this, choosing canonical space \\(①\\) will suffer from low resolution and choosing canonical space \\(③\\) will suffer from incomplete topology while \\(②\\) is a better canonical space to choose." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.788, + 0.457 + ], + "angle": 0, + "content": "effectively simulates the amount of knowledge contained in predictions and observations in the early stages of training when the system encounters new input data. In Fig. 6 we show two sources of knowledge from prediction branch and from direct output of observation MLP at frame \\( t_i \\). We see an obvious deficiency in the observation MLP when initializing new input frames and are pleased to find that the prediction branch compensates for this loss based on information from previous frames." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.478, + 0.51, + 0.495 + ], + "angle": 0, + "content": "6 Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.658 + ], + "angle": 0, + "content": "Limitations. Our method relies on a well-reconstructed radiance field in the canonical space, which in our pipeline is guided by \\(\\mathcal{L}_{co}\\). This design, however, will partially fail, if the chosen canonical space exhibits significant scale changes or even topological changes related to other frames. In Fig. 7 we use an example to illustrate the influences caused by choosing canonical space. This phenomenon, however, cannot be mitigated by precisely setting the canonical space as we lack a priori access to the input data. We note that some works [1,33] focus on addressing scale or topological issues in radiance fields reconstruction. Nevertheless, these issues are not the main focus of our paper and will be explored in future works to further refine our model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Conclusion. In this work, we present KFD-NeRF, a Kalman filter guided NeRF, for 4D dynamic view synthesis. We model the dynamic radiance field as a dynamic system in control theory and use Kalman filter to estimate the deformation states based on both observations and predictions. We further enhance our observation by encoding canonical space with an efficient tri-plane and by regularizing the shape in the canonical space. Through our temporal training strategy and newly derived pipeline, KFD-NeRF achieves state-of-the-art view synthesis performance among a variety of dynamic NeRF methods. We hope the dynamic system modeling of 4D radiance fields will encourage researchers to explore motion contextual information. KFD-NeRF hopefully inspires the use of existing sequential methods mainly in control theory and visual state estimation to further improve 4D view synthesis and deformation estimations tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.328, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.207 + ], + "angle": 0, + "content": "This research was supported in part by JSPS KAKENHI Grant Numbers 24K22318, 22H00529, 20H05951, JST-Mirai Program JPMJMI23G1." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.23, + 0.323, + 0.246 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.785, + 0.315 + ], + "angle": 0, + "content": "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A Multiscale Representation for Anti-aliasing Neural Radiance Fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.316, + 0.785, + 0.37 + ], + "angle": 0, + "content": "2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: MipNeRF 360: Unbounded Anti-aliased Neural Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.426 + ], + "angle": 0, + "content": "3. Becker, P., Pandya, H., Gebhardt, G., Zhao, C., Taylor, C.J., Neumann, G.: Recurrent Kalman Networks: Factorized Inference in High-dimensional Deep Feature Spaces. In: International conference on machine learning. pp. 544-552. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.468 + ], + "angle": 0, + "content": "4. Cao, A., Johnson, J.: Hexplane: A Fast Representation for Dynamic Scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.785, + 0.523 + ], + "angle": 0, + "content": "5. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient Geometry-aware 3D Generative Adversarial Networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.551 + ], + "angle": 0, + "content": "6. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial Radiance Fields. In: European Conference on Computer Vision. pp. 333-350. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.785, + 0.592 + ], + "angle": 0, + "content": "7. Choe, J., Choy, C., Park, J., Kweon, I.S., Anandkumar, A.: Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17871-17881 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.647 + ], + "angle": 0, + "content": "8. Coskun, H., Achilles, F., DiPietro, R., Navab, N., Tombari, F.: Long Short-term Memory Kalman Filters: Recurrent Neural Estimators for Pose Regularization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5524-5532 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.702 + ], + "angle": 0, + "content": "9. Du, Y., Zhang, Y., Yu, H.X., Tenenbaum, J.B., Wu, J.: Neural Radiance Flow for 4D View Synthesis and Video Processing. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14304-14314. IEEE Computer Society (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.743 + ], + "angle": 0, + "content": "10. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Niefsner, M., Tian, Q.: Fast Dynamic Radiance Fields with Time-aware Neural Voxels. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.785, + 0.785 + ], + "angle": 0, + "content": "1. Fraccaro, M., Kamronn, S., Paquet, U., Winther, O.: A Disentangled Recognition and Nonlinear Dynamics Model for Unsupervised Learning. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "2. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit Radiance Fields in Space, Time, and Appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.26, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.128 + ], + "angle": 0, + "content": "Y.Zhan, Z.Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "13. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance Fields without Neural Networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "14. Gan, W., Xu, H., Huang, Y., Chen, S., Yokoya, N.: V4D: Voxel for 4D Novel View Synthesis. IEEE Transactions on Visualization and Computer Graphics (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.219, + 0.785, + 0.26 + ], + "angle": 0, + "content": "15. Gao, C., Saraf, A., Kopf, J., Huang, J.B.: Dynamic View Synthesis from Dynamic Monocular Video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5712-5721 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.261, + 0.785, + 0.303 + ], + "angle": 0, + "content": "16. Guen, V.L., Thome, N.: Disentangling Physical Dynamics from Unknown Factors for Unsupervised Video Prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11474-11484 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.304, + 0.785, + 0.346 + ], + "angle": 0, + "content": "17. Guo, X., Chen, G., Dai, Y., Ye, X., Sun, J., Tan, X., Ding, E.: Neural Deformable Voxel Grid for Fast Optimization of Dynamic View Synthesis. In: Proceedings of the Asian Conference on Computer Vision. pp. 3757-3775 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.346, + 0.785, + 0.401 + ], + "angle": 0, + "content": "18. Guo, X., Sun, J., Dai, Y., Chen, G., Ye, X., Tan, X., Ding, E., Zhang, Y., Wang, J.: Forward Flow for Novel View Synthesis of Dynamic Scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16022-16033 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.402, + 0.785, + 0.444 + ], + "angle": 0, + "content": "19. Haarnoja, T., Ajay, A., Levine, S., Abbeel, P.: Backprop Kf: Learning Discriminative Deterministic State Estimators. Advances in neural information processing systems 29 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.445, + 0.785, + 0.487 + ], + "angle": 0, + "content": "20. Hu, T., Liu, S., Chen, Y., Shen, T., Jia, J.: EfficientNeRF: Efficient Neural Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12902-12911 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.488, + 0.785, + 0.515 + ], + "angle": 0, + "content": "21. Jang, H., Kim, D.: D-tensorf: Tensorial Radiance Fields for Dynamic Scenes. arXiv preprint arXiv:2212.02375 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.516, + 0.785, + 0.544 + ], + "angle": 0, + "content": "22. Kajiya, J.T., Von Herzen, B.P.: Ray Tracing Volume Densities. ACM SIGGRAPH computer graphics 18(3), 165-174 (1984)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.545, + 0.785, + 0.585 + ], + "angle": 0, + "content": "23. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D Gaussian Splatting for Real-time Radiance Field Rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.587, + 0.785, + 0.642 + ], + "angle": 0, + "content": "24. Li, T., Slavcheva, M., Zollhoefer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., et al.: Neural 3D Video Synthesis from Multi-view Video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5521-5531 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.643, + 0.785, + 0.685 + ], + "angle": 0, + "content": "25. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural Scene Flow Fields for Space-time View Synthesis of Dynamic Scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6498-6508 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.686, + 0.785, + 0.713 + ], + "angle": 0, + "content": "26. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural Sparse Voxel Fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.714, + 0.785, + 0.769 + ], + "angle": 0, + "content": "27. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust Dynamic Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13-23 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.77, + 0.785, + 0.812 + ], + "angle": 0, + "content": "28. Lu, G., Ouyang, W., Xu, D., Zhang, X., Gao, Z., Sun, M.T.: Deep Kalman Filtering Network for Video Compression Artifact Reduction. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 568-584 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "29. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d Gaussians: Tracking by Persistent Dynamic View Synthesis. arXiv preprint arXiv:2308.09713 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.328, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "30. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.231 + ], + "angle": 0, + "content": "31. Müller, T., Evans, A., Schied, C., Keller, A.: Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.272 + ], + "angle": 0, + "content": "32. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable Neural Radiance Fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.314 + ], + "angle": 0, + "content": "33. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A Higher-dimensional Representation for Topologically Varying Neural Radiance Fields. arXiv preprint arXiv:2106.13228 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.369 + ], + "angle": 0, + "content": "34. Park, S., Son, M., Jang, S., Ahn, Y.C., Kim, J.Y., Kang, N.: Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4212-4221 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.424 + ], + "angle": 0, + "content": "35. Peng, S., Niemeyer, M., Mescheder, L., Pollefeys, M., Geiger, A.: Convolutional Occupancy Networks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 523-540. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.467 + ], + "angle": 0, + "content": "36. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-neRF: Neural Radiance Fields for Dynamic Scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.508 + ], + "angle": 0, + "content": "37. Revach, G., Shlezinger, N., Ni, X., Escoriza, A.L., Van Sloun, R.J., Eldar, Y.C.: KalmanNet: Neural Network Aided Kalman Filtering for Partially Known Dynamics. IEEE Transactions on Signal Processing 70, 1532-1547 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.563 + ], + "angle": 0, + "content": "38. Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., Liu, Y.: Tensor4d: Efficient Neural 4D Decomposition for High-fidelity Dynamic Reconstruction and Rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16632-16642 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.605 + ], + "angle": 0, + "content": "39. Sun, C., Sun, M., Chen, H.T.: Direct Voxel Grid Optimization: Super-fast CO convergence for Radiance Fields Reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5459-5469 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.66 + ], + "angle": 0, + "content": "40. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid Neural Radiance Fields: Reconstruction and Novel View Synthesis of a Dynamic Scene from Monocular Video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12959-12970 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.702 + ], + "angle": 0, + "content": "41. Wang, F., Tan, S., Li, X., Tian, Z., Song, Y., Liu, H.: Mixed Neural Voxels for Fast Multi-view Video Synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19706-19716 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.702, + 0.788, + 0.743 + ], + "angle": 0, + "content": "42. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image Quality Assessment: from Error Visibility to Structural Similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.743, + 0.743, + 0.758 + ], + "angle": 0, + "content": "43. Welch, G., Bishop, G., et al.: An Introduction to the Kalman Filter (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.758, + 0.788, + 0.799 + ], + "angle": 0, + "content": "44. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering. arXiv preprint arXiv:2310.08528 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.841 + ], + "angle": 0, + "content": "45. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time Neural Irradiance Fields for Free-viewpoint Video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9421-9431 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.128 + ], + "angle": 0, + "content": "Y.Zhan, Z.Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "46. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction. arXiv preprint arXiv:2309.13101 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.787, + 0.245 + ], + "angle": 0, + "content": "47. Yoon, J.S., Kim, K., Gallo, O., Park, H.S., Kautz, J.: Novel View Synthesis of Dynamic Scenes with Globally Coherent Depths from a Monocular Camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5336-5345 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "48. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for Real-time Rendering of Neural Radiance Fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5752-5761 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.34 + ], + "angle": 0, + "content": "49. Yuan, W., Lv, Z., Schmidt, T., Lovegrove, S.: Star: Self-supervised Tracking and Reconstruction of Rigid Objects in Motion with Neural Rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13144-13152 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.384 + ], + "angle": 0, + "content": "50. Zhang, H., Yang, Z., Xiong, H., Zhu, T., Long, Z., Wu, W.: Transformer Aided Adaptive Extended Kalman Filter for Autonomous Vehicle Mass Estimation. Processes 11(3), 887 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.425 + ], + "angle": 0, + "content": "51. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.48 + ], + "angle": 0, + "content": "52. Zhou, L., Luo, Z., Shen, T., Zhang, J., Zhen, M., Yao, Y., Fang, T., Quan, L.: Kfnet: Learning Temporal Camera Relocalization Using Kalman Filtering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4919-4928 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.48 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/KFD-NeRF_ Rethinking Dynamic NeRF with Kalman Filter/3bae14dd-a29b-4634-bf7b-d258a05d1f55_origin.pdf b/2024/KFD-NeRF_ Rethinking Dynamic NeRF with Kalman Filter/3bae14dd-a29b-4634-bf7b-d258a05d1f55_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a5510cb34811f4c1b6e6921e419def463c0cd4ce --- /dev/null +++ b/2024/KFD-NeRF_ Rethinking Dynamic NeRF with Kalman Filter/3bae14dd-a29b-4634-bf7b-d258a05d1f55_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a0362d2e51b2251ced079f8545d948cff82710f0bfc0820f476b4b91c003330 +size 22970738 diff --git a/2024/KFD-NeRF_ Rethinking Dynamic NeRF with Kalman Filter/full.md b/2024/KFD-NeRF_ Rethinking Dynamic NeRF with Kalman Filter/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e31a7ea622ff4b7bc96a928eb65d56688fe057ad --- /dev/null +++ b/2024/KFD-NeRF_ Rethinking Dynamic NeRF with Kalman Filter/full.md @@ -0,0 +1,346 @@ +# KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter + +Yifan Zhan $^{1}$ , Zhuoxiao Li $^{1}$ , Muyao Niu $^{1}$ , Zhihang Zhong $^{3}$ , Shohei Nobuhara $^{2}$ , Ko Nishino $^{2}$ , and Yinqiang Zheng $^{\star 1}$ + +1 The University of Tokyo + +2 Kyoto University + +3 Shanghai Artificial Intelligence Laboratory + +Abstract. We introduce KFD-NeRF, a novel dynamic neural radiance field integrated with an efficient and high-quality motion reconstruction framework based on Kalman filtering. Our key idea is to model the dynamic radiance field as a dynamic system whose temporally varying states are estimated based on two sources of knowledge: observations and predictions. We introduce a novel plug-in Kalman filter guided deformation field that enables accurate deformation estimation from scene observations and predictions. We use a shallow Multi-Layer Perceptron (MLP) for observations and model the motion as locally linear to calculate predictions with motion equations. To further enhance the performance of the observation MLP, we introduce regularization in the canonical space to facilitate the network's ability to learn warping for different frames. Additionally, we employ an efficient tri-plane representation for encoding the canonical space, which has been experimentally demonstrated to converge quickly with high quality. This enables us to use a shallower observation MLP, consisting of just two layers in our implementation. We conduct experiments on synthetic and real data and compare with past dynamic NeRF methods. Our KFD-NeRF demonstrates similar or even superior rendering performance within comparable computational time and achieves state-of-the-art view synthesis performance with thorough training. Github page: https://github.com/Yifever20002/KFD-NeRF. + +Keywords: Dynamic NeRF $\cdot$ Deformable Network $\cdot$ Kalman Filter + +# 1 Introduction + +Neural Radiance Fields (NeRF) [30] have demonstrated outstanding success as a versatile and accurate 3D representation of real-world scenes, which has led to its wide adoption in daily and industrial applications in numerous domains. One of the remaining key desiderata for NeRFs is its extension to dynamic scenes. + +Existing dynamic NeRF methods can be broadly categorized into two approaches. One is to learn deformation fields for motion warping (e.g., D-NeRF [36] + +![](images/cd76fcb37d8442f2af9a36fa968ce6d3d58bfc27e340375e385fa84fc65371a8.jpg) +Fig. 1: In contrast to a vanilla deformation field, our plug-in Kalman filter guided deformation field consists of a prediction branch along with the direct observations from input data. From noise related terms $\varepsilon_{t_i}$ and $\varepsilon_{t_{i - 1}}$ we learn Kalman gain $K_{t_i}$ , weighting observations $y_{t_i}$ and predictions $\hat{d\pmb{x}_{t_i}}$ for more accurate deformation estimations. + +and TiNeuVox [10]). Another is to disregard motion priors and directly interpolate time in the feature space (e.g., DyNeRF [24] and KPlanes [12]). These approaches, however, often overlook the characteristics of a dynamic radiance field as a time-state sequence, missing the opportunity to fully leverage temporal contextual information. + +In this work, we draw inspirations from control theory and model the 4D radiance field as a dynamic system with temporally-varying states. Our state estimates of this dynamic system come from two sources of knowledge: observations based on the input data and predictions based on the system physical dynamics. Optimal state estimates cannot be achieved with only one of these two sources of knowledge. On the one hand, observations, as commonly used in previous dynamic NeRF works, are inherently error-prone due to the discrete temporal sampling of the dynamic scenes. On the other hand, predictions are governed by the correctness of the assumed kinematic model and may struggle to maintain accuracy for real dynamic scenes. + +To maximize combined potential of both observations and predictions, we introduce an efficient plug-and-play Kalman filter [43] module to optimize state estimations of our dynamic system. In Fig. 1 we illustrate our plug-in Kalman filter guided deformation field. We model the 4D radiance field as a single-state system, with the state denoted as $d\pmb{x}_{t_i}$ for the current frame deformation. In contrast to a vanilla deformation field that only considers system observations in the current frame, our approach incorporates richer information from previous + +frames by introducing a prediction branch based on a motion equation. Given the absence of prior trajectory regarding the scene's motion, we employ local linear motion. + +Both observations and predictions are weighted using a learnable Kalman gain to calculate precise deformation estimations. During the initial stages of training, predictions primarily influence the process, facilitating the convergence of frames with substantial motion. In the later stages of training, observations take precedence, allowing for the recovery of more precise and fine-grained motion details. + +All points in the real space are warped to a time-independent canonical space according to the estimated deformations. To further improve the performance of our two-branch deformable model, we employ an efficient tri-plane spatial representation for encoding our canonical space. Experimental evidence shows that this representation permits a shallower observation MLP with only two layers in our implementation. At the same time, we improve the warping ability of the observation MLP by regularizing the learning of the radiance field in the canonical space. + +In summary, our contributions are as follows: + +1) The first method for modeling 4D radiance fields as dynamic systems by integrating a Kalman filter into the deformation field formulation, which results in a plug-in, efficient method for estimating deformations; +2) KFD-NeRF, a novel deformable NeRF with the Kalman filter plug-in and a tri-plane spatial representation, trained with a novel strategy of gradually releasing temporal information to facilitate learning dynamic systems: +3) and regularization in the canonical space for enhancing the learning capacity of a shallow observation MLP. We achieve state-of-the-art results on both synthetic and real data with all these designs, compared to dynamic NeRFs. + +# 2 Related Works + +Neural Radiance Fields Representation. NeRF [30] represents 3D scenes based on an implicit representation encoded by an MLP. Given multi-view images and corresponding camera poses as input, the MLP is trained by "analysis by synthesis," achieving novel view synthesis and coarse geometry reconstruction. Vanilla NeRF leaves room for improvement that has attracted abundant research. + +Many methods [13, 20, 26, 39, 48] use sparse voxel grids to represent the 3D scenes. While grid-based modeling achieves fast training and rendering, fine details require high-resolution grids leading to excessive memory consumption. The optimization process for grids is also unstable, failing to leverage the smoothness bias brought by MLP. Instant-NGP [31] proposes an efficient hash encoding for mapping spatial and feature domains. For sparsely-observed regions, however, this one-to-many mapping can easily introduce noise. The tri-plane representation [5, 6, 35] significantly reduces the memory footprint by leveraging its low-rank decomposition. We use this representation to build a fine-detailed canonical space. + +Dynamic Neural Rendering. For dynamic scenarios, modeling the time dimension together with the spatial representation becomes the main challenge. One approach to dynamic NeRF [4, 12, 14, 15, 21, 24, 25, 27, 34, 38, 45] is to add timestamps as an extra dimension to the 3D space and formulate 4D interpolation. Though intuitive, this leads to temporal incoherence due to the lack of supervision and priors between frames. Another approach [7, 9, 10, 17, 18, 25, 32, 33, 36, 40, 49] is to model the motion with deformation fields that warp points in arbitrary frames to the canonical space conditioned on the timestamps. This design can take advantage of the MLP smoothness bias in accordance with the motion smoothness prior. Nevertheless, deformation fields fail to find correspondences in frames with significant motion. We emphasize the performance improvement resulting from the switch in backbone spatial representation from D-NeRF [36] (MLP-based) to TiNeuVox [10] (voxel grids based) and later analyze the impact of different spatial representations on the deformation field. We also compare with latest point-based 4D rendering works [29, 44, 46] inspired by 3D Gaussian Splatting [23]. + +Deep Kalman Filter. The combination of deep learning and Kalman filtering has been explored to address the challenges of incomplete observations in various scenarios. Some works [28, 52] use Convolutional Neural Network (CNN)-based Kalman filtering for video compression and camera localization. Recurrent Neural Network (RNN)-based Kalman filter has also been used in some studies [8, 37] to improve state estimation and to optimize pose regularization. The Transformer in [50] enables a more comprehensive exploitation of temporal contextual information. All of these works along with [3, 11, 19], however, model dynamics only in the learned latent space without taking into account physical priors, with the exception of [16] which fuses known physical priors and network outputs to construct a Kalman filter for video prediction. In this paper, for the first time, we employ a neural Kalman filter to assist in the dynamic NeRF task. We derive a two-stream method consisting of a shallow MLP and physical priors to achieve efficient motion estimation. + +# 3 Preliminaries + +# 3.1 NeRF and Volume Rendering Revisited + +NeRF [30] consists of three parts: sampling, volume mapping, and rendering. In the sampling process, points $\pmb{x} \in \mathbb{R}^3$ are sampled along rays calculated from the camera pose. Then, in the volume mapping process, each $\pmb{x}$ in the 3D world, as well as its viewing direction $\pmb{v} \in \mathbb{R}^3$ , is queried to output the volume density $\sigma$ and radiance $\pmb{c} = (r, g, b)$ of $\pmb{x}$ . Finally, in the rendering process, the color of each ray is computed with volume rendering [22]. The expected color of the ray $r(s) = o + sv$ becomes [30] + +$$ +C (\boldsymbol {r}) = \int_ {s _ {n}} ^ {s _ {f}} T (s) \sigma (\boldsymbol {r} (s)) \boldsymbol {c} (\boldsymbol {r} (s), \boldsymbol {v}) d s, \tag {1} +$$ + +where $s_n$ and $s_f$ are the near and far bounds, respectively, and + +$$ +T (s) = \exp \left(- \int_ {s _ {n}} ^ {s} \sigma (\boldsymbol {r} (k)) d k\right). \tag {2} +$$ + +The only supervision for training NeRF comes from the ground truth color $C_{gt}(\boldsymbol{r})$ of each ray + +$$ +\mathcal {L} _ {\text {i m a g e}} = \sum_ {\boldsymbol {r} \in \mathcal {R}} \| C (\boldsymbol {r}) - C _ {g t} (\boldsymbol {r}) \| _ {2} ^ {2}, \tag {3} +$$ + +where $\mathcal{R}$ is the ray batch. + +# 3.2 Kalman Filter + +Consider a dynamic system with input $u_{t}$ , output $y_{t}$ , process noise $n_t$ and measurement noise $m_t$ . We want to obtain its state $x_{t}$ at each frame $t$ , assuming the state equation follows + +$$ +x _ {t} = A x _ {t - 1} + B u _ {t} + n _ {t}, \tag {4} +$$ + +and the output equation follows + +$$ +y _ {t} = C x _ {t} + m _ {t}, \tag {5} +$$ + +where $A$ , $B$ and $C$ are the system matrix, control matrix and output matrix, respectively. + +There are two methods for obtaining the system's state $x_{t}$ at any given $t$ : observation and prediction. The observation method tries to rely on $y_{t}$ in Eq. (5) while the prediction method tries to mathematically model the system dynamics to predict the state. Both of these methods, however, come with non-negligible errors due to the existence of $m_{t}$ and $n_{t}$ and for inaccurate system modeling. + +The Kalman filter algorithm posits that the state $x_{t}$ is obtained by a weighted sum of the prediction based on the state $x_{t-1}$ and the observation $y_{t}$ . For a specific $t$ , estimating the state $x_{t}$ can be divided into two steps: prediction step and update step. Assuming that the process noise and measurement noise follow zero-mean Gaussian distributions with variances $Q$ and $R$ , respectively, at the prediction step, it first calculates a predicted state based on the estimated state at $t-1$ + +$$ +\hat {x} _ {t} ^ {-} = A \hat {x} _ {t - 1} + B u _ {t}, \tag {6} +$$ + +and its error covariance + +$$ +P _ {t} ^ {-} = A P _ {t - 1} A ^ {T} + Q. \tag {7} +$$ + +In the update step, this predicted state is combined with observation $y_{t}$ to obtain updated state estimate + +$$ +\hat {x} _ {t} = \hat {x} _ {t} ^ {-} + K _ {t} \left(y _ {t} - C \hat {x} _ {t} ^ {-}\right), \tag {8} +$$ + +![](images/e9b5666b0bf697a6a488424ba18708fc1e2de9adfa55e02f7848b4f8a2746b90.jpg) +Fig. 2: Visualization of feature planes learned by the feature interpolation method ① and the deformation fields method ②. We show a point $P$ in the real world space and its corresponding four points $P_{1}, P_{2}, P_{3}$ and $P_{4}$ in the canonical space at four different timestamps. The feature plane in ② exhibits better smoothness compared to ①, so we use ② to construct our backbone. + +![](images/ebcc21e63339fbf3db054c13e464e62b6269b166bd8bb8f2fc46697238dc5bb9.jpg) + +and its error covariance + +$$ +P _ {t} = \left(I - K _ {t} C\right) P _ {t} ^ {-}, \tag {9} +$$ + +where + +$$ +K _ {t} = \frac {P _ {t} ^ {-} C ^ {T}}{C P _ {t} ^ {-} C ^ {T} + R} \tag {10} +$$ + +is the Kalman gain, reflecting the weights assigned to the observation and prediction components. + +# 4 Method + +Figure 3 illustrates the three stages of the complete pipeline of KFD-NeRF. In this section, we will first analyze the advantages of using deformation fields as the motion representation over feature interpolation. We will then introduce KFD-NeRF based on the Kalman filter to achieve accurate deformation estimations. Finally, we will discuss spatial reconstruction details, the training strategy, and the incorporation of regularization. + +# 4.1 Motion Representation: Deformation Fields vs. Feature Interpolation + +4D NeRFs generally employ two approaches to temporal modeling: motion deformation fields or time-conditioned feature interpolation. A deformation field $D(\pmb{x}, t) \rightarrow \Delta \pmb{x}$ calculates the deformation $\Delta \pmb{x}$ at each timestamp $t$ , which can be used to warp a 3D point at $t$ to its corresponding position in the canonical space. In contrast, time-conditioned feature interpolation $F(\pmb{x}, t) \rightarrow \pmb{f}$ directly learns a feature vector $\pmb{f}$ from a given 3D point at $t$ , which is sent to a decoder for RGB and density calculation. + +![](images/b587c683564864b326b922bb7e11c8da4bd82e7fe1666680f15827eaf33d1462.jpg) +Fig. 3: The overall pipeline of our KFD-NeRF. Our designed deformation field calculates final deformations based on two sources of knowledge. At the observation stage, system observations are directly output by a shallow observation MLP. At the prediction and fusion stage, we calculate predictions based on system dynamics, and further fuse observations and predictions to obtain final deformation estimations. At the spatial reconstruction stage, tri-plane encoded canonical points are concatenated with raw positions and timestamps, which are further decoded to obtain predicted colors for loss calculation. (*f represents a linear layer, used to obtain the Kalman gain from noise-related terms $\varepsilon_{t_i}$ and $\varepsilon_{t_{i - 1}}$ .) + +In Fig. 2, we illustrate the difference between these two methods with a motion example from time $t = 0$ to time $t = 1$ . Given a 3D point $P$ in the real world space, we can identify its corresponding four points in the canonical space at four different timestamps, named $P_{1}, P_{2}, P_{3}$ , and $P_{4}$ . ① shows the feature to be learned at $P$ which changes over time. Notice that when the radiance undergoes abrupt changes, the feature space exhibits many high-frequency signals, which are hard to fit. ② shows the backward deformation from the real world space to the canonical space to be learned at $P$ which changes over time. These signals are smoother and easier to fit by leveraging the motion's smoothness prior. MLP representations have inductive smoothness bias and are well-suited for learning such deformations. Hence, we employ MLP-based deformation fields to represent motion in this work. + +# 4.2 Kalman Filter Guided Motion Prediction + +Next, we derive the system equations for the dynamic radiance field following Eqs. (4) and (5). As we are trying to model non-rigid motions, each 3D point in the dynamic radiance field could be independently seen as a dynamic system, with its velocity $\pmb{v}_{t_i}$ at each frame $t_i$ as a single input $u_t$ . Also note that + +this dynamic system is a single-state system since we only focus on estimating deformation at each frame $t_i$ , denoted as $d\pmb{x}_{t_i}$ . + +In the next step, we determine the coefficients matrix $A$ , $B$ , and $C$ in Eqs. (4) and (5), each degenerating to a value under a single-input and single-state system. As we directly observe $d\pmb{x}_{t_i}$ , the output matrix $C$ becomes 1. Since the motion trajectory is unknown, we use a locally-linear motion model to describe this dynamic system: $d\pmb{x}_{t_i} = d\pmb{x}_{t_{i-1}} + \Delta t_i \cdot \pmb{v}_{t_i}$ ( $A = 1$ and $B = \Delta t_i$ ). Further considering noise, our state and output equations become + +$$ +\left\{ \begin{array}{l} d \boldsymbol {x} _ {t _ {i}} = d \boldsymbol {x} _ {t _ {i - 1}} + \Delta t _ {i} \cdot \boldsymbol {v} _ {t _ {i}} + n _ {t _ {i}}, \\ y _ {t _ {i}} = d \boldsymbol {x} _ {t _ {i}} + m _ {t _ {i}}. \end{array} \right. \tag {11} +$$ + +In Fig. 3, our designed deformation field consists of two parts: an MLP-based observer and a system-dynamics-based predictor. As for the observer, by querying 3D points coordinates $\mathbf{x}$ and timestamps $t_i$ , a two-layer shallow MLP is used to output the mean observation $y_{t_i}$ (assuming zero-mean Gaussian measurement noise $m_{t_i}$ ) and noise-related terms $\varepsilon_{t_i}$ . For the predictor, we follow the locally-linear motion model and first calculate the input $\mathbf{v}_{t_i}$ . Because the deformation of the current frame $t_i$ is being estimated, we only use the information from frame $t_{i-1}$ and frame $t_{i-2}$ to approximate the velocity + +$$ +\boldsymbol {v} _ {t _ {i}} = \left(d \boldsymbol {x} _ {t _ {i - 1}} - d \boldsymbol {x} _ {t _ {i - 2}}\right) / \Delta t _ {i}. \tag {12} +$$ + +Based on the estimated velocity, we compute the predicted deformation state $\hat{d\pmb{x}}_{t_i}^{-}$ . The final estimation $\hat{d\pmb{x}}_{t_i}$ is the sum of $\hat{d\pmb{x}}_{t_i}^{-}$ and $y_{t_i}$ weighted by a Kalman gain $K_{t_i}$ . The Kalman gain in Eq. (10) is calculated by current measurement noise and past process noise. In our implementation, we obtain $K_{t_i}$ utilizing a linear layer that takes noise-related terms $\varepsilon_{t_i}$ and $\varepsilon_{t_{i-1}}$ and timestamps $t_i$ and $t_{i-1}$ as inputs, incorporating historical information. Algorithm 1 summarizes this deformation estimation. + +Volume is warped to the canonical space based on the estimated deformation, which is further decoded to obtain density $\sigma$ and color. We compute the loss by comparing the rendered values with the ground truth and use it to update the network, following Eq. (3). In addition, we note that the estimated deformation serves as a good supervision for learning the observation MLP, in analogy with observation updating process in Kalman filter. Therefore, we try to minimize the error between the current observation and the estimated deformation with + +$$ +\mathcal {L} _ {k f} = \frac {1}{\mathcal {N}} \sum_ {\boldsymbol {x} \in \mathcal {N}} \left\| y _ {t _ {i}} - \hat {d} \boldsymbol {x} _ {t _ {i}} \right\| _ {2} ^ {2}. \tag {13} +$$ + +The weighted $y_{t_i}$ and $\hat{d\pmb{x}}_{t_i}^{-}$ stand for two sources of knowledge from the whole system. $y_{t_i}$ represents the state directly observed by the observation MLP, which lacks the information from past frames and thus has a low confidence in the early stages of training. In contrast, $\hat{d\pmb{x}}_{t_i}^{-}$ represents the current frame information predicted from history, which provides a good prior in the early stages of training. + +Algorithm 1 Deformation Estimations + +
Prediction:Update:
dxti-1, εti-1 = observer(x, ti-1)yti, εti = observer(x, ti)
dxti-2, εti-2 = observer(x, ti-2)Kti = f(εti, εti-1, ti, ti-1)
vti = (dxti-1 - dxti-2) / Δtidxti = dxti - Kti(yti - dxti - )
dxti = dxti-1 + Δti · vt = 2dxti-1 - dxti-2
+ +In the later stages of training, however, the confidence of the predictor drops significantly due to the inherent errors of the modeled motion dynamics, which is gradually compensated by the well learned observation MLP. Our Kalman Filter guided model automatically strikes the balance of both leveraging $\hat{d\mathbf{x}}_{t_i}^{-}$ in the early stages to accelerate convergence and avoid local minimum and also using $y_{t_i}$ in the later stages to learn fine details. + +# 4.3 Spatial Representation + +MLP-based spatial representations suffer from slow convergence and require a deeper deformation network (e.g., eight layers MLP in [36]) to model 4D scenes. Otherwise, the deformation field can get stuck in a local optimum before learning the radiance field in the canonical space. Voxel grids based spatial representations converge very quickly (e.g., TiNeuVox [10]) but demand high spatial resolution to store fine details, which requires significant memory footprint. Under memory constraints, achieving high resolution in the canonical space can be challenging, resulting in losing scene details and deteriorating the warping quality for each frame. We employ a tri-plane representation, which uses three sets of mutually orthogonal 2D planes to represent 3D space. This low-rank model allows for rapid convergence while significantly reducing memory consumption, making it possible to achieve fast and high-quality canonical space reconstruction. + +What's more, there are some inevitable errors due to coordinate shifts when $\pmb{x}'$ are encoded by finite resolution tri-plane. To enhance the raw coordinate information, we follow [10] by concatenating encoded tri-plane features with raw coordinate inputs. + +# 4.4 Training Strategy and Regularization + +Our model takes into account temporal information so the learning of the current frame partially relies on the results of previous frames. To ensure that previous frames can offer ample priors, we employ a training strategy of gradually releasing the training images in chronological order. + +We also notice that the lack of constraints or priors in the canonical space can affect the performance of the observation branch, which could be regularized by pre-setting the shape in the canonical space. D-NeRF [36] sets frame $t_0$ to be the canonical space and force the deformation output of frame $t_0$ to be masked to 0. Such a hard mask, however, does not allow learning of the deformation + +at frame $t_0$ and may cause the canonical space to be too complex for motions warping. + +We instead design a soft regularization term to normalize the difference between the canonical space and the radiance field in the real world at frame $t_0$ to improve the observation branch. The canonical observation loss of a points batch $\mathcal{N}$ is + +$$ +\mathcal {L} _ {c o} = \frac {1}{\mathcal {N}} \sum_ {\boldsymbol {x} \in \mathcal {N}} \mathbf {1} (t) d \boldsymbol {x}, \tag {14} +$$ + +where $\mathbf{1}(t) = 1$ when $t = 0$ and $\mathbf{1}(t) = 0$ when $t\neq 0$ + +We use a proposal sampling strategy from Mip-NeRF 360 [2] by distilling the density field for occupancy estimation. This online distillation necessitates a loss function $\mathcal{L}_{prop}$ to ensure consistency between the proposal network and our learned model. Please refer to Section 3 in [2] for detailed definition. + +Total variation loss $\mathcal{L}_{tv}$ is a common regularizer in inverse problems, which encourages sparse edges in space. We apply this loss to each of our tri-plane to get $\mathcal{L}_{tv}(\mathbf{x}^{\prime})$ , where $\mathbf{x}^{\prime}$ are warped 3D points in the canonical space. The total variation loss is + +$$ +\mathcal {L} _ {t v} (\mathbf {x}) = \frac {1}{| C |} \sum_ {c, i, j} \left(\left\| \mathbf {x} _ {c} ^ {i, j} - \mathbf {x} _ {c} ^ {i - 1, j} \right\| _ {2} ^ {2} + \left\| \mathbf {x} _ {c} ^ {i, j} - \mathbf {x} _ {c} ^ {i, j - 1} \right\| _ {2} ^ {2}\right), \tag {15} +$$ + +where $c$ is a certain plane from the tri-plane collection $C$ and $i, j$ are indices on the plane's resolution. + +The final loss function becomes + +$$ +\mathcal {L} = \mathcal {L} _ {\text {i m a g e}} + \mathcal {L} _ {k f} + \mathcal {L} _ {c o} + \mathcal {L} _ {\text {p r o p}} + \lambda_ {t v} \mathcal {L} _ {t v}, \tag {16} +$$ + +and we experimentally choose $\lambda_{tv}$ to be $1\times 10^{-4}$ . + +# 5 Experimental Results + +# 5.1 Dataset + +For synthetic data, we use the Dynamic NeRF Synthetic Dataset provided by D-NeRF [36], whose training and testing splits have already been well-organized. For each scene in the synthetic dataset, a photo from an arbitrary view with corresponding camera pose is provided at each timestamp. For real data, we use the Nvidia Real Dynamic Scenes Dataset [47] which consists of 8 dynamic scenes recorded by 12 synchronized cameras. We use 11 camera videos for training and the remaining one for testing. + +# 5.2 Baselines and Metrics + +Due to differences in the format of synthetic and real data, we carefully select cutting-edge baselines to thoroughly validate our method based on comparative experiments. For synthetic data, we test deformation based methods + +![](images/f0fd7359dd0441ddaecd72c7010e52aabe06fbd6820fa00e1b348d1a2f62f483.jpg) +Fig. 4: Qualitative Comparison of our KFD-NeRF against other dynamic NeRF methods on synthetic data. Zoom in for better details. + +![](images/58c955a9b809652629e70559cf6ab53dca88c2dd7affd03183a3726483e2fe9d.jpg) +Fig. 5: Qualitative Comparison of our KFD-NeRF against other dynamic NeRF methods on real data. Zoom in for better details. + +
ModelSynthetic Data
PSNR(dB)↑SSIM↑LPIPS↓Average↓
D-NeRF [36]30.250.9540.0760.0288
TiNeuVox-B [10]32.930.9730.0430.0173
KPlanes [12]30.470.9630.0720.0261
NDVG [17]30.510.9640.0560.0223
V4D [14]33.540.9780.0280.0141
4D-GS [44]33.310.9790.0260.0136
KFD-NeRF35.730.9840.0450.0131
+ +
ModelReal Data
PSNR(dB)↑SSIM↑LPIPS↓Average↓
TiNeuVox-B [10]24.390.7340.3030.0869
KPlanes [12]28.250.8810.1370.0433
Mixvoxels-L [41]26.950.8280.2220.0619
V4D [14]28.160.8840.1360.0426
KFD-NeRF28.750.8910.1150.0389
+ +Table 1: Quantitative comparison of our KFD-NeRF against other dynamic NeRF methods. We show results on both synthetic and real data. See Sec. 5.2 for more details. + +
D-NeRFTiNeuVox-BKPlanesNDVGV4D4D-GSMixvoxels-LKFD-NeRF-SKFD-NeRF-L
Training Time↓48hrs40mins60mins35mins7hrs25mins1.5hrs30mins3hrs
Params(MB)↓1650340150275163140175175
+ +Table 2: Training time and parameter size for different methods. + +D-NeRF [36] (MLP based spatial representation), TiNeuVox-B [10] (voxel grids based spatial representation), NDVG [17] (voxel grids based spatial representation) and 4D-GS [44] (Gaussian points based spatial representation), and feature interpolation based methods KPlanes [12] and V4D [14]. For real data, besides TiNeuVox-B and KPlanes, we further compare multi-view videos reconstruction methods MixVoxels [41]. We train all these methods on a single GeForce RTX3090. See Tab. 2 for detailed training time and parameters consumption of ours and other methods. + +We provide an exhaustive qualitative and quantitative comparison of our KFD-NeRF with these baseline methods. Three main metrics are reported, namely the peak signal-to-noise ratio (PSNR), the structural similarity index measure (SSIM) [42], and the learned perceptual image patch similarity (LPIPS) [51]. To provide more intuitive results, we further calculate metric "average" [1], which is the geometric mean of $\mathrm{MSE} = 10^{-\mathrm{PSNR} / 10}$ , $\sqrt{1 - \mathrm{SSIM}}$ , and LPIPS. Please see Tab. 1 for quantitative comparison and Figs. 4 and 5 for qualitative comparison. Per-scene results can be found in the supplemental material. We strongly recommend readers to watch the supplemental videos for a more intuitive understanding of the results. + +# 5.3 Ablation Studies + +We conduct ablation studies on both synthetic and real data to validate our various proposed system components. We compare our full model with variants related to $\mathcal{L}_{co}$ , $\mathcal{L}_{kf}$ and prediction branch. + +Canonical observation loss $\mathcal{L}_{co}$ . This loss is designed to regularize a continuous and smooth volume shape in the canonical space for better observations. Specifically, we guide the shape in the canonical space to be close to the shape + +![](images/cd705bcd6785d78392fc29bb7bd5776ba1be0f1a51535c4cb5b18492165995bc.jpg) +Fig. 6: Two sources of knowledge at frame $t_i$ from our pre-trained dynamic system with inputs up to and including frame $t_{i-1}$ . Notice that during the early stages of training, our observation MLP has insufficient initialization, resulting in the loss of many details (green boxes on the right). Fortunately, our predictions provide additional priors (green boxes on the left) for the current frame based on information from previous frames, compensating for the loss in observations. + +
LkfLcoPSNR(dB)↑SSIM↑LPIPS↓
a)30.150.9140.110
b)✓30.600.9200.106
c)✓31.780.9310.086
d)✓✓32.240.9380.080
+ +Table 3: Ablation Studies on the $\mathcal{L}_{kf}$ and $\mathcal{L}_{co}$ . We show all the ablation combinations and report the average results on synthetic and real data. + +
ModelDynamicStatic
PSNR(dB)↑SSIM↑PSNR(dB)↑SSIM↑
P.N., (xi, ti) Input25.610.96329.580.812
P.N., (xi, ti, ti-1, ti-2) Input25.660.96429.550.812
w/ Prediction Branch (full)26.710.96729.570.813
+ +Table 4: The ablation study on our prediction branch for real data in average, where dynamic and static areas are counted separately. P.N. is short for Pure MLP Network. + +at frame $t_0$ . In Tab. 3 line $b$ , we remove $\mathcal{L}_{co}$ and observe a decrease in model performance. + +Estimation update loss $\mathcal{L}_{kf}$ . This loss tries to minimize the difference between estimated deformation and current observation. This updating process ensures that with each iteration, the observation acquire progressively more accurate information to guide the learning process. In Tab. 3 line $c$ ), we remove $\mathcal{L}_{kf}$ and witness a decrease in model performance. + +# 5.4 Extra Studies on Prediction Branch + +This study aims to demonstrate the effectiveness of our plug-in Kalman filter. The most crucial step in Kalman filter is to fuse the original network observations with the predictions from system dynamics. Therefore, we directly remove the "Prediction and Fusion Stage" in our pipeline and use pure shallow observation MLP to generate deformations. We further conduct an experiment where a pure deformable network takes the current frame and the previous two states as inputs to ablate the effectiveness of the Kalman motion modeling. In Tab. 4, we see a clear drop in performance by pure MLP network-based methods, indicating that simply inputting previous Kalman filter parameters without modeling motion prediction is insufficient. + +The prediction branch we have designed can offer reliable priors in the early stages of training. We visualize these priors through an experiment to gain further insight. Specifically, we train our system with only inputs up to and including frame $t_{i-1}$ and then let our system produce rendering results at frame $t_i$ with two different branches: predictions and observations. This experiment + +![](images/d745ebbb4ae374a9c051a1d4d6aba042933f85bde751b1b06031d5e63f3b81d5.jpg) +Fig. 7: An example in Nvidia Dataset [47] to illustrate scale and topological issues in choosing canonical space. From $①$ to $②$ there is a scale change due to moving closer to the camera. From $③$ to $②$ there is a topological change caused by target disappearance. According to this, choosing canonical space $①$ will suffer from low resolution and choosing canonical space $③$ will suffer from incomplete topology while $②$ is a better canonical space to choose. + +effectively simulates the amount of knowledge contained in predictions and observations in the early stages of training when the system encounters new input data. In Fig. 6 we show two sources of knowledge from prediction branch and from direct output of observation MLP at frame $t_i$ . We see an obvious deficiency in the observation MLP when initializing new input frames and are pleased to find that the prediction branch compensates for this loss based on information from previous frames. + +# 6 Discussion and Conclusion + +Limitations. Our method relies on a well-reconstructed radiance field in the canonical space, which in our pipeline is guided by $\mathcal{L}_{co}$ . This design, however, will partially fail, if the chosen canonical space exhibits significant scale changes or even topological changes related to other frames. In Fig. 7 we use an example to illustrate the influences caused by choosing canonical space. This phenomenon, however, cannot be mitigated by precisely setting the canonical space as we lack a priori access to the input data. We note that some works [1,33] focus on addressing scale or topological issues in radiance fields reconstruction. Nevertheless, these issues are not the main focus of our paper and will be explored in future works to further refine our model. + +Conclusion. In this work, we present KFD-NeRF, a Kalman filter guided NeRF, for 4D dynamic view synthesis. We model the dynamic radiance field as a dynamic system in control theory and use Kalman filter to estimate the deformation states based on both observations and predictions. We further enhance our observation by encoding canonical space with an efficient tri-plane and by regularizing the shape in the canonical space. Through our temporal training strategy and newly derived pipeline, KFD-NeRF achieves state-of-the-art view synthesis performance among a variety of dynamic NeRF methods. We hope the dynamic system modeling of 4D radiance fields will encourage researchers to explore motion contextual information. KFD-NeRF hopefully inspires the use of existing sequential methods mainly in control theory and visual state estimation to further improve 4D view synthesis and deformation estimations tasks. + +# Acknowledgements + +This research was supported in part by JSPS KAKENHI Grant Numbers 24K22318, 22H00529, 20H05951, JST-Mirai Program JPMJMI23G1. + +# References + +1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A Multiscale Representation for Anti-aliasing Neural Radiance Fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021) +2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: MipNeRF 360: Unbounded Anti-aliased Neural Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022) +3. Becker, P., Pandya, H., Gebhardt, G., Zhao, C., Taylor, C.J., Neumann, G.: Recurrent Kalman Networks: Factorized Inference in High-dimensional Deep Feature Spaces. In: International conference on machine learning. pp. 544-552. PMLR (2019) +4. Cao, A., Johnson, J.: Hexplane: A Fast Representation for Dynamic Scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2023) +5. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient Geometry-aware 3D Generative Adversarial Networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022) +6. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial Radiance Fields. In: European Conference on Computer Vision. pp. 333-350. Springer (2022) +7. Choe, J., Choy, C., Park, J., Kweon, I.S., Anandkumar, A.: Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17871-17881 (2023) +8. Coskun, H., Achilles, F., DiPietro, R., Navab, N., Tombari, F.: Long Short-term Memory Kalman Filters: Recurrent Neural Estimators for Pose Regularization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5524-5532 (2017) +9. Du, Y., Zhang, Y., Yu, H.X., Tenenbaum, J.B., Wu, J.: Neural Radiance Flow for 4D View Synthesis and Video Processing. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14304-14314. IEEE Computer Society (2021) +10. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Niefsner, M., Tian, Q.: Fast Dynamic Radiance Fields with Time-aware Neural Voxels. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022) +1. Fraccaro, M., Kamronn, S., Paquet, U., Winther, O.: A Disentangled Recognition and Nonlinear Dynamics Model for Unsupervised Learning. Advances in neural information processing systems 30 (2017) +2. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit Radiance Fields in Space, Time, and Appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023) + +13. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance Fields without Neural Networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022) +14. Gan, W., Xu, H., Huang, Y., Chen, S., Yokoya, N.: V4D: Voxel for 4D Novel View Synthesis. IEEE Transactions on Visualization and Computer Graphics (2023) +15. Gao, C., Saraf, A., Kopf, J., Huang, J.B.: Dynamic View Synthesis from Dynamic Monocular Video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5712-5721 (2021) +16. Guen, V.L., Thome, N.: Disentangling Physical Dynamics from Unknown Factors for Unsupervised Video Prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11474-11484 (2020) +17. Guo, X., Chen, G., Dai, Y., Ye, X., Sun, J., Tan, X., Ding, E.: Neural Deformable Voxel Grid for Fast Optimization of Dynamic View Synthesis. In: Proceedings of the Asian Conference on Computer Vision. pp. 3757-3775 (2022) +18. Guo, X., Sun, J., Dai, Y., Chen, G., Ye, X., Tan, X., Ding, E., Zhang, Y., Wang, J.: Forward Flow for Novel View Synthesis of Dynamic Scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16022-16033 (2023) +19. Haarnoja, T., Ajay, A., Levine, S., Abbeel, P.: Backprop Kf: Learning Discriminative Deterministic State Estimators. Advances in neural information processing systems 29 (2016) +20. Hu, T., Liu, S., Chen, Y., Shen, T., Jia, J.: EfficientNeRF: Efficient Neural Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12902-12911 (2022) +21. Jang, H., Kim, D.: D-tensorf: Tensorial Radiance Fields for Dynamic Scenes. arXiv preprint arXiv:2212.02375 (2022) +22. Kajiya, J.T., Von Herzen, B.P.: Ray Tracing Volume Densities. ACM SIGGRAPH computer graphics 18(3), 165-174 (1984) +23. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D Gaussian Splatting for Real-time Radiance Field Rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023) +24. Li, T., Slavcheva, M., Zollhoefer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., et al.: Neural 3D Video Synthesis from Multi-view Video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5521-5531 (2022) +25. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural Scene Flow Fields for Space-time View Synthesis of Dynamic Scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6498-6508 (2021) +26. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural Sparse Voxel Fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020) +27. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust Dynamic Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13-23 (2023) +28. Lu, G., Ouyang, W., Xu, D., Zhang, X., Gao, Z., Sun, M.T.: Deep Kalman Filtering Network for Video Compression Artifact Reduction. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 568-584 (2018) +29. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d Gaussians: Tracking by Persistent Dynamic View Synthesis. arXiv preprint arXiv:2308.09713 (2023) + +30. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. Communications of the ACM 65(1), 99-106 (2021) +31. Müller, T., Evans, A., Schied, C., Keller, A.: Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022) +32. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable Neural Radiance Fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021) +33. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A Higher-dimensional Representation for Topologically Varying Neural Radiance Fields. arXiv preprint arXiv:2106.13228 (2021) +34. Park, S., Son, M., Jang, S., Ahn, Y.C., Kim, J.Y., Kang, N.: Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4212-4221 (2023) +35. Peng, S., Niemeyer, M., Mescheder, L., Pollefeys, M., Geiger, A.: Convolutional Occupancy Networks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 523-540. Springer (2020) +36. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-neRF: Neural Radiance Fields for Dynamic Scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021) +37. Revach, G., Shlezinger, N., Ni, X., Escoriza, A.L., Van Sloun, R.J., Eldar, Y.C.: KalmanNet: Neural Network Aided Kalman Filtering for Partially Known Dynamics. IEEE Transactions on Signal Processing 70, 1532-1547 (2022) +38. Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., Liu, Y.: Tensor4d: Efficient Neural 4D Decomposition for High-fidelity Dynamic Reconstruction and Rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16632-16642 (2023) +39. Sun, C., Sun, M., Chen, H.T.: Direct Voxel Grid Optimization: Super-fast CO convergence for Radiance Fields Reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5459-5469 (2022) +40. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid Neural Radiance Fields: Reconstruction and Novel View Synthesis of a Dynamic Scene from Monocular Video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12959-12970 (2021) +41. Wang, F., Tan, S., Li, X., Tian, Z., Song, Y., Liu, H.: Mixed Neural Voxels for Fast Multi-view Video Synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19706-19716 (2023) +42. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image Quality Assessment: from Error Visibility to Structural Similarity. IEEE transactions on image processing 13(4), 600-612 (2004) +43. Welch, G., Bishop, G., et al.: An Introduction to the Kalman Filter (1995) +44. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering. arXiv preprint arXiv:2310.08528 (2023) +45. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time Neural Irradiance Fields for Free-viewpoint Video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9421-9431 (2021) + +46. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction. arXiv preprint arXiv:2309.13101 (2023) +47. Yoon, J.S., Kim, K., Gallo, O., Park, H.S., Kautz, J.: Novel View Synthesis of Dynamic Scenes with Globally Coherent Depths from a Monocular Camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5336-5345 (2020) +48. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for Real-time Rendering of Neural Radiance Fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5752-5761 (2021) +49. Yuan, W., Lv, Z., Schmidt, T., Lovegrove, S.: Star: Self-supervised Tracking and Reconstruction of Rigid Objects in Motion with Neural Rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13144-13152 (2021) +50. Zhang, H., Yang, Z., Xiong, H., Zhu, T., Long, Z., Wu, W.: Transformer Aided Adaptive Extended Kalman Filter for Autonomous Vehicle Mass Estimation. Processes 11(3), 887 (2023) +51. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018) +52. Zhou, L., Luo, Z., Shen, T., Zhang, J., Zhen, M., Yao, Y., Fang, T., Quan, L.: Kfnet: Learning Temporal Camera Relocalization Using Kalman Filtering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4919-4928 (2020) \ No newline at end of file diff --git a/2024/KFD-NeRF_ Rethinking Dynamic NeRF with Kalman Filter/images.zip b/2024/KFD-NeRF_ Rethinking Dynamic NeRF with Kalman Filter/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..06046b92dc86abaf6bbb3026ceba1e6271d46074 --- /dev/null +++ b/2024/KFD-NeRF_ Rethinking Dynamic NeRF with Kalman Filter/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87e2855fa160fc9fd0100329d8dae83c6ee0b6c4519afa03bfc5628b621a863d +size 605735 diff --git a/2024/KFD-NeRF_ Rethinking Dynamic NeRF with Kalman Filter/layout.json b/2024/KFD-NeRF_ Rethinking Dynamic NeRF with Kalman Filter/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..abfb2486ad23485183b1f49a98576b5c6d61c232 --- /dev/null +++ b/2024/KFD-NeRF_ Rethinking Dynamic NeRF with Kalman Filter/layout.json @@ -0,0 +1,10767 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 111, + 465, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 111, + 465, + 145 + ], + "spans": [ + { + "bbox": [ + 148, + 111, + 465, + 145 + ], + "type": "text", + "content": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 135, + 168, + 478, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 168, + 478, + 194 + ], + "spans": [ + { + "bbox": [ + 135, + 168, + 478, + 194 + ], + "type": "text", + "content": "Yifan Zhan" + }, + { + "bbox": [ + 135, + 168, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 168, + 478, + 194 + ], + "type": "text", + "content": ", Zhuoxiao Li" + }, + { + "bbox": [ + 135, + 168, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 168, + 478, + 194 + ], + "type": "text", + "content": ", Muyao Niu" + }, + { + "bbox": [ + 135, + 168, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 168, + 478, + 194 + ], + "type": "text", + "content": ", Zhihang Zhong" + }, + { + "bbox": [ + 135, + 168, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 135, + 168, + 478, + 194 + ], + "type": "text", + "content": ", Shohei Nobuhara" + }, + { + "bbox": [ + 135, + 168, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 168, + 478, + 194 + ], + "type": "text", + "content": ", Ko Nishino" + }, + { + "bbox": [ + 135, + 168, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 168, + 478, + 194 + ], + "type": "text", + "content": ", and Yinqiang Zheng" + }, + { + "bbox": [ + 135, + 168, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{\\star 1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 252, + 201, + 361, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 201, + 361, + 213 + ], + "spans": [ + { + "bbox": [ + 252, + 201, + 361, + 213 + ], + "type": "text", + "content": "1 The University of Tokyo" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 267, + 213, + 346, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 213, + 346, + 224 + ], + "spans": [ + { + "bbox": [ + 267, + 213, + 346, + 224 + ], + "type": "text", + "content": "2 Kyoto University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 216, + 224, + 398, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 224, + 398, + 236 + ], + "spans": [ + { + "bbox": [ + 216, + 224, + 398, + 236 + ], + "type": "text", + "content": "3 Shanghai Artificial Intelligence Laboratory" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 269, + 455, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 269, + 455, + 499 + ], + "spans": [ + { + "bbox": [ + 159, + 269, + 455, + 499 + ], + "type": "text", + "content": "Abstract. We introduce KFD-NeRF, a novel dynamic neural radiance field integrated with an efficient and high-quality motion reconstruction framework based on Kalman filtering. Our key idea is to model the dynamic radiance field as a dynamic system whose temporally varying states are estimated based on two sources of knowledge: observations and predictions. We introduce a novel plug-in Kalman filter guided deformation field that enables accurate deformation estimation from scene observations and predictions. We use a shallow Multi-Layer Perceptron (MLP) for observations and model the motion as locally linear to calculate predictions with motion equations. To further enhance the performance of the observation MLP, we introduce regularization in the canonical space to facilitate the network's ability to learn warping for different frames. Additionally, we employ an efficient tri-plane representation for encoding the canonical space, which has been experimentally demonstrated to converge quickly with high quality. This enables us to use a shallower observation MLP, consisting of just two layers in our implementation. We conduct experiments on synthetic and real data and compare with past dynamic NeRF methods. Our KFD-NeRF demonstrates similar or even superior rendering performance within comparable computational time and achieves state-of-the-art view synthesis performance with thorough training. Github page: https://github.com/Yifever20002/KFD-NeRF." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 510, + 438, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 510, + 438, + 522 + ], + "spans": [ + { + "bbox": [ + 160, + 510, + 438, + 522 + ], + "type": "text", + "content": "Keywords: Dynamic NeRF " + }, + { + "bbox": [ + 160, + 510, + 438, + 522 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 510, + 438, + 522 + ], + "type": "text", + "content": " Deformable Network " + }, + { + "bbox": [ + 160, + 510, + 438, + 522 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 510, + 438, + 522 + ], + "type": "text", + "content": " Kalman Filter" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 544, + 230, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 230, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 230, + 558 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 572, + 482, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 572, + 482, + 619 + ], + "spans": [ + { + "bbox": [ + 130, + 572, + 482, + 619 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRF) [30] have demonstrated outstanding success as a versatile and accurate 3D representation of real-world scenes, which has led to its wide adoption in daily and industrial applications in numerous domains. One of the remaining key desiderata for NeRFs is its extension to dynamic scenes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 620, + 483, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 620, + 483, + 645 + ], + "spans": [ + { + "bbox": [ + 130, + 620, + 483, + 645 + ], + "type": "text", + "content": "Existing dynamic NeRF methods can be broadly categorized into two approaches. One is to learn deformation fields for motion warping (e.g., D-NeRF [36]" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 654, + 266, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 654, + 266, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 654, + 266, + 665 + ], + "type": "text", + "content": "* denotes corresponding author." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 118, + 475, + 315 + ], + "blocks": [ + { + "bbox": [ + 138, + 118, + 475, + 315 + ], + "lines": [ + { + "bbox": [ + 138, + 118, + 475, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 118, + 475, + 315 + ], + "type": "image", + "image_path": "cd76fcb37d8442f2af9a36fa968ce6d3d58bfc27e340375e385fa84fc65371a8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "text", + "content": "Fig. 1: In contrast to a vanilla deformation field, our plug-in Kalman filter guided deformation field consists of a prediction branch along with the direct observations from input data. From noise related terms " + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\varepsilon_{t_i}" + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\varepsilon_{t_{i - 1}}" + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "text", + "content": " we learn Kalman gain " + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "inline_equation", + "content": "K_{t_i}" + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "text", + "content": ", weighting observations " + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "inline_equation", + "content": "y_{t_i}" + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "text", + "content": " and predictions " + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\hat{d\\pmb{x}_{t_i}}" + }, + { + "bbox": [ + 130, + 327, + 482, + 376 + ], + "type": "text", + "content": " for more accurate deformation estimations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 400, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 460 + ], + "type": "text", + "content": "and TiNeuVox [10]). Another is to disregard motion priors and directly interpolate time in the feature space (e.g., DyNeRF [24] and KPlanes [12]). These approaches, however, often overlook the characteristics of a dynamic radiance field as a time-state sequence, missing the opportunity to fully leverage temporal contextual information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 461, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 581 + ], + "type": "text", + "content": "In this work, we draw inspirations from control theory and model the 4D radiance field as a dynamic system with temporally-varying states. Our state estimates of this dynamic system come from two sources of knowledge: observations based on the input data and predictions based on the system physical dynamics. Optimal state estimates cannot be achieved with only one of these two sources of knowledge. On the one hand, observations, as commonly used in previous dynamic NeRF works, are inherently error-prone due to the discrete temporal sampling of the dynamic scenes. On the other hand, predictions are governed by the correctness of the assumed kinematic model and may struggle to maintain accuracy for real dynamic scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "To maximize combined potential of both observations and predictions, we introduce an efficient plug-and-play Kalman filter [43] module to optimize state estimations of our dynamic system. In Fig. 1 we illustrate our plug-in Kalman filter guided deformation field. We model the 4D radiance field as a single-state system, with the state denoted as " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "d\\pmb{x}_{t_i}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " for the current frame deformation. In contrast to a vanilla deformation field that only considers system observations in the current frame, our approach incorporates richer information from previous" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "type": "text", + "content": "Y.Zhan, Z.Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "frames by introducing a prediction branch based on a motion equation. Given the absence of prior trajectory regarding the scene's motion, we employ local linear motion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 481, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 481, + 222 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 481, + 222 + ], + "type": "text", + "content": "Both observations and predictions are weighted using a learnable Kalman gain to calculate precise deformation estimations. During the initial stages of training, predictions primarily influence the process, facilitating the convergence of frames with substantial motion. In the later stages of training, observations take precedence, allowing for the recovery of more precise and fine-grained motion details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 319 + ], + "type": "text", + "content": "All points in the real space are warped to a time-independent canonical space according to the estimated deformations. To further improve the performance of our two-branch deformable model, we employ an efficient tri-plane spatial representation for encoding our canonical space. Experimental evidence shows that this representation permits a shallower observation MLP with only two layers in our implementation. At the same time, we improve the warping ability of the observation MLP by regularizing the learning of the radiance field in the canonical space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 319, + 347, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 319, + 347, + 331 + ], + "spans": [ + { + "bbox": [ + 146, + 319, + 347, + 331 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 338, + 479, + 443 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 135, + 338, + 479, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 338, + 479, + 373 + ], + "spans": [ + { + "bbox": [ + 135, + 338, + 479, + 373 + ], + "type": "text", + "content": "1) The first method for modeling 4D radiance fields as dynamic systems by integrating a Kalman filter into the deformation field formulation, which results in a plug-in, efficient method for estimating deformations;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 374, + 479, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 374, + 479, + 407 + ], + "spans": [ + { + "bbox": [ + 135, + 374, + 479, + 407 + ], + "type": "text", + "content": "2) KFD-NeRF, a novel deformable NeRF with the Kalman filter plug-in and a tri-plane spatial representation, trained with a novel strategy of gradually releasing temporal information to facilitate learning dynamic systems:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 135, + 410, + 479, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 410, + 479, + 443 + ], + "spans": [ + { + "bbox": [ + 135, + 410, + 479, + 443 + ], + "type": "text", + "content": "3) and regularization in the canonical space for enhancing the learning capacity of a shallow observation MLP. We achieve state-of-the-art results on both synthetic and real data with all these designs, compared to dynamic NeRFs." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 462, + 242, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 462, + 242, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 462, + 242, + 475 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 486, + 479, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 479, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 479, + 544 + ], + "type": "text", + "content": "Neural Radiance Fields Representation. NeRF [30] represents 3D scenes based on an implicit representation encoded by an MLP. Given multi-view images and corresponding camera poses as input, the MLP is trained by \"analysis by synthesis,\" achieving novel view synthesis and coarse geometry reconstruction. Vanilla NeRF leaves room for improvement that has attracted abundant research." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "content": "Many methods [13, 20, 26, 39, 48] use sparse voxel grids to represent the 3D scenes. While grid-based modeling achieves fast training and rendering, fine details require high-resolution grids leading to excessive memory consumption. The optimization process for grids is also unstable, failing to leverage the smoothness bias brought by MLP. Instant-NGP [31] proposes an efficient hash encoding for mapping spatial and feature domains. For sparsely-observed regions, however, this one-to-many mapping can easily introduce noise. The tri-plane representation [5, 6, 35] significantly reduces the memory footprint by leveraging its low-rank decomposition. We use this representation to build a fine-detailed canonical space." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": "Dynamic Neural Rendering. For dynamic scenarios, modeling the time dimension together with the spatial representation becomes the main challenge. One approach to dynamic NeRF [4, 12, 14, 15, 21, 24, 25, 27, 34, 38, 45] is to add timestamps as an extra dimension to the 3D space and formulate 4D interpolation. Though intuitive, this leads to temporal incoherence due to the lack of supervision and priors between frames. Another approach [7, 9, 10, 17, 18, 25, 32, 33, 36, 40, 49] is to model the motion with deformation fields that warp points in arbitrary frames to the canonical space conditioned on the timestamps. This design can take advantage of the MLP smoothness bias in accordance with the motion smoothness prior. Nevertheless, deformation fields fail to find correspondences in frames with significant motion. We emphasize the performance improvement resulting from the switch in backbone spatial representation from D-NeRF [36] (MLP-based) to TiNeuVox [10] (voxel grids based) and later analyze the impact of different spatial representations on the deformation field. We also compare with latest point-based 4D rendering works [29, 44, 46] inspired by 3D Gaussian Splatting [23]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 319, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 486 + ], + "type": "text", + "content": "Deep Kalman Filter. The combination of deep learning and Kalman filtering has been explored to address the challenges of incomplete observations in various scenarios. Some works [28, 52] use Convolutional Neural Network (CNN)-based Kalman filtering for video compression and camera localization. Recurrent Neural Network (RNN)-based Kalman filter has also been used in some studies [8, 37] to improve state estimation and to optimize pose regularization. The Transformer in [50] enables a more comprehensive exploitation of temporal contextual information. All of these works along with [3, 11, 19], however, model dynamics only in the learned latent space without taking into account physical priors, with the exception of [16] which fuses known physical priors and network outputs to construct a Kalman filter for video prediction. In this paper, for the first time, we employ a neural Kalman filter to assist in the dynamic NeRF task. We derive a two-stream method consisting of a shallow MLP and physical priors to achieve efficient motion estimation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 505, + 234, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 505, + 234, + 517 + ], + "spans": [ + { + "bbox": [ + 132, + 505, + 234, + 517 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 529, + 361, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 529, + 361, + 541 + ], + "spans": [ + { + "bbox": [ + 132, + 529, + 361, + 541 + ], + "type": "text", + "content": "3.1 NeRF and Volume Rendering Revisited" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 549, + 482, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 549, + 482, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 549, + 482, + 634 + ], + "type": "text", + "content": "NeRF [30] consists of three parts: sampling, volume mapping, and rendering. In the sampling process, points " + }, + { + "bbox": [ + 130, + 549, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 549, + 482, + 634 + ], + "type": "text", + "content": " are sampled along rays calculated from the camera pose. Then, in the volume mapping process, each " + }, + { + "bbox": [ + 130, + 549, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 549, + 482, + 634 + ], + "type": "text", + "content": " in the 3D world, as well as its viewing direction " + }, + { + "bbox": [ + 130, + 549, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\pmb{v} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 549, + 482, + 634 + ], + "type": "text", + "content": ", is queried to output the volume density " + }, + { + "bbox": [ + 130, + 549, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 549, + 482, + 634 + ], + "type": "text", + "content": " and radiance " + }, + { + "bbox": [ + 130, + 549, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\pmb{c} = (r, g, b)" + }, + { + "bbox": [ + 130, + 549, + 482, + 634 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 549, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 549, + 482, + 634 + ], + "type": "text", + "content": ". Finally, in the rendering process, the color of each ray is computed with volume rendering [22]. The expected color of the ray " + }, + { + "bbox": [ + 130, + 549, + 482, + 634 + ], + "type": "inline_equation", + "content": "r(s) = o + sv" + }, + { + "bbox": [ + 130, + 549, + 482, + 634 + ], + "type": "text", + "content": " becomes [30]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 224, + 642, + 481, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 642, + 481, + 668 + ], + "spans": [ + { + "bbox": [ + 224, + 642, + 481, + 668 + ], + "type": "interline_equation", + "content": "C (\\boldsymbol {r}) = \\int_ {s _ {n}} ^ {s _ {f}} T (s) \\sigma (\\boldsymbol {r} (s)) \\boldsymbol {c} (\\boldsymbol {r} (s), \\boldsymbol {v}) d s, \\tag {1}", + "image_path": "0983550cc9ef05de537ff7bf3a759e459808865afade7ff8042bf02f2dcd3ca5.jpg" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "Y.Zhan, Z.Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 411, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 411, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 411, + 129 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 116, + 411, + 129 + ], + "type": "inline_equation", + "content": "s_n" + }, + { + "bbox": [ + 132, + 116, + 411, + 129 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 116, + 411, + 129 + ], + "type": "inline_equation", + "content": "s_f" + }, + { + "bbox": [ + 132, + 116, + 411, + 129 + ], + "type": "text", + "content": " are the near and far bounds, respectively, and" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 235, + 138, + 481, + 165 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 138, + 481, + 165 + ], + "spans": [ + { + "bbox": [ + 235, + 138, + 481, + 165 + ], + "type": "interline_equation", + "content": "T (s) = \\exp \\left(- \\int_ {s _ {n}} ^ {s} \\sigma (\\boldsymbol {r} (k)) d k\\right). \\tag {2}", + "image_path": "94ad04f5e81f66e8a9f6b45181209ff5af7261dda4dcbafc762b165111f43fcb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 479, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 479, + 198 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 479, + 198 + ], + "type": "text", + "content": "The only supervision for training NeRF comes from the ground truth color " + }, + { + "bbox": [ + 132, + 172, + 479, + 198 + ], + "type": "inline_equation", + "content": "C_{gt}(\\boldsymbol{r})" + }, + { + "bbox": [ + 132, + 172, + 479, + 198 + ], + "type": "text", + "content": " of each ray" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 233, + 206, + 481, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 206, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 233, + 206, + 481, + 232 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {i m a g e}} = \\sum_ {\\boldsymbol {r} \\in \\mathcal {R}} \\| C (\\boldsymbol {r}) - C _ {g t} (\\boldsymbol {r}) \\| _ {2} ^ {2}, \\tag {3}", + "image_path": "13ed0e5d7a3d69c39cbe65320094d1c6cc53b3e4a13310f10b8f58c9f51750cc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 241, + 246, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 241, + 246, + 252 + ], + "spans": [ + { + "bbox": [ + 132, + 241, + 246, + 252 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 241, + 246, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 132, + 241, + 246, + 252 + ], + "type": "text", + "content": " is the ray batch." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 233, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 233, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 233, + 282 + ], + "type": "text", + "content": "3.2 Kalman Filter" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 291, + 481, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 291, + 481, + 327 + ], + "spans": [ + { + "bbox": [ + 132, + 291, + 481, + 327 + ], + "type": "text", + "content": "Consider a dynamic system with input " + }, + { + "bbox": [ + 132, + 291, + 481, + 327 + ], + "type": "inline_equation", + "content": "u_{t}" + }, + { + "bbox": [ + 132, + 291, + 481, + 327 + ], + "type": "text", + "content": ", output " + }, + { + "bbox": [ + 132, + 291, + 481, + 327 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 132, + 291, + 481, + 327 + ], + "type": "text", + "content": ", process noise " + }, + { + "bbox": [ + 132, + 291, + 481, + 327 + ], + "type": "inline_equation", + "content": "n_t" + }, + { + "bbox": [ + 132, + 291, + 481, + 327 + ], + "type": "text", + "content": " and measurement noise " + }, + { + "bbox": [ + 132, + 291, + 481, + 327 + ], + "type": "inline_equation", + "content": "m_t" + }, + { + "bbox": [ + 132, + 291, + 481, + 327 + ], + "type": "text", + "content": ". We want to obtain its state " + }, + { + "bbox": [ + 132, + 291, + 481, + 327 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 132, + 291, + 481, + 327 + ], + "type": "text", + "content": " at each frame " + }, + { + "bbox": [ + 132, + 291, + 481, + 327 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 291, + 481, + 327 + ], + "type": "text", + "content": ", assuming the state equation follows" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 254, + 338, + 481, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 338, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 254, + 338, + 481, + 350 + ], + "type": "interline_equation", + "content": "x _ {t} = A x _ {t - 1} + B u _ {t} + n _ {t}, \\tag {4}", + "image_path": "4baa02acdb33aaa96f9bf0073be4ca3b24874f7748ce69f70bace026043ff55d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 359, + 276, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 359, + 276, + 371 + ], + "spans": [ + { + "bbox": [ + 132, + 359, + 276, + 371 + ], + "type": "text", + "content": "and the output equation follows" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 273, + 382, + 481, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 382, + 481, + 393 + ], + "spans": [ + { + "bbox": [ + 273, + 382, + 481, + 393 + ], + "type": "interline_equation", + "content": "y _ {t} = C x _ {t} + m _ {t}, \\tag {5}", + "image_path": "beb61de502b533620422c4124aee82f92b63660c0fd11af772d6404041414adf.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 403, + 479, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 479, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 479, + 426 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 403, + 479, + 426 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 403, + 479, + 426 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 403, + 479, + 426 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 403, + 479, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 403, + 479, + 426 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 403, + 479, + 426 + ], + "type": "text", + "content": " are the system matrix, control matrix and output matrix, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 427, + 480, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 480, + 486 + ], + "type": "text", + "content": "There are two methods for obtaining the system's state " + }, + { + "bbox": [ + 130, + 427, + 480, + 486 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 427, + 480, + 486 + ], + "type": "text", + "content": " at any given " + }, + { + "bbox": [ + 130, + 427, + 480, + 486 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 427, + 480, + 486 + ], + "type": "text", + "content": ": observation and prediction. The observation method tries to rely on " + }, + { + "bbox": [ + 130, + 427, + 480, + 486 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 130, + 427, + 480, + 486 + ], + "type": "text", + "content": " in Eq. (5) while the prediction method tries to mathematically model the system dynamics to predict the state. Both of these methods, however, come with non-negligible errors due to the existence of " + }, + { + "bbox": [ + 130, + 427, + 480, + 486 + ], + "type": "inline_equation", + "content": "m_{t}" + }, + { + "bbox": [ + 130, + 427, + 480, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 427, + 480, + 486 + ], + "type": "inline_equation", + "content": "n_{t}" + }, + { + "bbox": [ + 130, + 427, + 480, + 486 + ], + "type": "text", + "content": " and for inaccurate system modeling." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 487, + 480, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 480, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 480, + 569 + ], + "type": "text", + "content": "The Kalman filter algorithm posits that the state " + }, + { + "bbox": [ + 130, + 487, + 480, + 569 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 487, + 480, + 569 + ], + "type": "text", + "content": " is obtained by a weighted sum of the prediction based on the state " + }, + { + "bbox": [ + 130, + 487, + 480, + 569 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 130, + 487, + 480, + 569 + ], + "type": "text", + "content": " and the observation " + }, + { + "bbox": [ + 130, + 487, + 480, + 569 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 130, + 487, + 480, + 569 + ], + "type": "text", + "content": ". For a specific " + }, + { + "bbox": [ + 130, + 487, + 480, + 569 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 487, + 480, + 569 + ], + "type": "text", + "content": ", estimating the state " + }, + { + "bbox": [ + 130, + 487, + 480, + 569 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 487, + 480, + 569 + ], + "type": "text", + "content": " can be divided into two steps: prediction step and update step. Assuming that the process noise and measurement noise follow zero-mean Gaussian distributions with variances " + }, + { + "bbox": [ + 130, + 487, + 480, + 569 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 487, + 480, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 487, + 480, + 569 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 487, + 480, + 569 + ], + "type": "text", + "content": ", respectively, at the prediction step, it first calculates a predicted state based on the estimated state at " + }, + { + "bbox": [ + 130, + 487, + 480, + 569 + ], + "type": "inline_equation", + "content": "t-1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 263, + 571, + 480, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 571, + 480, + 583 + ], + "spans": [ + { + "bbox": [ + 263, + 571, + 480, + 583 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {t} ^ {-} = A \\hat {x} _ {t - 1} + B u _ {t}, \\tag {6}", + "image_path": "d95d0eaf2534afc48ff44a8db354ab8dcaffa498c82cd20e04e7dd26e7c8c598.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 590, + 238, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 238, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 238, + 600 + ], + "type": "text", + "content": "and its error covariance" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 260, + 601, + 480, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 601, + 480, + 613 + ], + "spans": [ + { + "bbox": [ + 260, + 601, + 480, + 613 + ], + "type": "interline_equation", + "content": "P _ {t} ^ {-} = A P _ {t - 1} A ^ {T} + Q. \\tag {7}", + "image_path": "15f988d6f1d9430df5fc5ced50ae9534ad1c9f5992b543c522763efa3401b486.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 619, + 479, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 619, + 479, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 619, + 479, + 643 + ], + "type": "text", + "content": "In the update step, this predicted state is combined with observation " + }, + { + "bbox": [ + 132, + 619, + 479, + 643 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 132, + 619, + 479, + 643 + ], + "type": "text", + "content": " to obtain updated state estimate" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 251, + 654, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 654, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 251, + 654, + 480, + 666 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {t} = \\hat {x} _ {t} ^ {-} + K _ {t} \\left(y _ {t} - C \\hat {x} _ {t} ^ {-}\\right), \\tag {8}", + "image_path": "563f8719402bcc255b5ff3af7a05f65e7a888566b34dad74c4668a6f469dbc54.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "type": "text", + "content": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 118, + 282, + 220 + ], + "blocks": [ + { + "bbox": [ + 136, + 118, + 282, + 220 + ], + "lines": [ + { + "bbox": [ + 136, + 118, + 282, + 220 + ], + "spans": [ + { + "bbox": [ + 136, + 118, + 282, + 220 + ], + "type": "image", + "image_path": "e9b5666b0bf697a6a488424ba18708fc1e2de9adfa55e02f7848b4f8a2746b90.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 232, + 482, + 288 + ], + "lines": [ + { + "bbox": [ + 131, + 232, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 131, + 232, + 482, + 288 + ], + "type": "text", + "content": "Fig. 2: Visualization of feature planes learned by the feature interpolation method ① and the deformation fields method ②. We show a point " + }, + { + "bbox": [ + 131, + 232, + 482, + 288 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 131, + 232, + 482, + 288 + ], + "type": "text", + "content": " in the real world space and its corresponding four points " + }, + { + "bbox": [ + 131, + 232, + 482, + 288 + ], + "type": "inline_equation", + "content": "P_{1}, P_{2}, P_{3}" + }, + { + "bbox": [ + 131, + 232, + 482, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 232, + 482, + 288 + ], + "type": "inline_equation", + "content": "P_{4}" + }, + { + "bbox": [ + 131, + 232, + 482, + 288 + ], + "type": "text", + "content": " in the canonical space at four different timestamps. The feature plane in ② exhibits better smoothness compared to ①, so we use ② to construct our backbone." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 125, + 474, + 219 + ], + "blocks": [ + { + "bbox": [ + 306, + 125, + 474, + 219 + ], + "lines": [ + { + "bbox": [ + 306, + 125, + 474, + 219 + ], + "spans": [ + { + "bbox": [ + 306, + 125, + 474, + 219 + ], + "type": "image", + "image_path": "ebcc21e63339fbf3db054c13e464e62b6269b166bd8bb8f2fc46697238dc5bb9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 312, + 238, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 238, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 238, + 323 + ], + "type": "text", + "content": "and its error covariance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 261, + 324, + 480, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 324, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 261, + 324, + 480, + 338 + ], + "type": "interline_equation", + "content": "P _ {t} = \\left(I - K _ {t} C\\right) P _ {t} ^ {-}, \\tag {9}", + "image_path": "35c8d1e05baffaabd0f4cb14dafd3ae97919fcc21a370c0264ab70adfe18a2fa.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 344, + 161, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 344, + 161, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 344, + 161, + 355 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 263, + 353, + 480, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 353, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 263, + 353, + 480, + 380 + ], + "type": "interline_equation", + "content": "K _ {t} = \\frac {P _ {t} ^ {-} C ^ {T}}{C P _ {t} ^ {-} C ^ {T} + R} \\tag {10}", + "image_path": "cdbb1aaa1c16fd31acc5a028b1627a4398e8b918f927e24ca18f7ad73444acdb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 385, + 480, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 385, + 480, + 410 + ], + "spans": [ + { + "bbox": [ + 131, + 385, + 480, + 410 + ], + "type": "text", + "content": "is the Kalman gain, reflecting the weights assigned to the observation and prediction components." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 430, + 202, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 202, + 442 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 202, + 442 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 456, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 456, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 131, + 456, + 482, + 529 + ], + "type": "text", + "content": "Figure 3 illustrates the three stages of the complete pipeline of KFD-NeRF. In this section, we will first analyze the advantages of using deformation fields as the motion representation over feature interpolation. We will then introduce KFD-NeRF based on the Kalman filter to achieve accurate deformation estimations. Finally, we will discuss spatial reconstruction details, the training strategy, and the incorporation of regularization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 548, + 444, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 444, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 444, + 572 + ], + "type": "text", + "content": "4.1 Motion Representation: Deformation Fields vs. Feature Interpolation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 582, + 482, + 666 + ], + "type": "text", + "content": "4D NeRFs generally employ two approaches to temporal modeling: motion deformation fields or time-conditioned feature interpolation. A deformation field " + }, + { + "bbox": [ + 131, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "D(\\pmb{x}, t) \\rightarrow \\Delta \\pmb{x}" + }, + { + "bbox": [ + 131, + 582, + 482, + 666 + ], + "type": "text", + "content": " calculates the deformation " + }, + { + "bbox": [ + 131, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta \\pmb{x}" + }, + { + "bbox": [ + 131, + 582, + 482, + 666 + ], + "type": "text", + "content": " at each timestamp " + }, + { + "bbox": [ + 131, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 582, + 482, + 666 + ], + "type": "text", + "content": ", which can be used to warp a 3D point at " + }, + { + "bbox": [ + 131, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 582, + 482, + 666 + ], + "type": "text", + "content": " to its corresponding position in the canonical space. In contrast, time-conditioned feature interpolation " + }, + { + "bbox": [ + 131, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "F(\\pmb{x}, t) \\rightarrow \\pmb{f}" + }, + { + "bbox": [ + 131, + 582, + 482, + 666 + ], + "type": "text", + "content": " directly learns a feature vector " + }, + { + "bbox": [ + 131, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + }, + { + "bbox": [ + 131, + 582, + 482, + 666 + ], + "type": "text", + "content": " from a given 3D point at " + }, + { + "bbox": [ + 131, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 582, + 482, + 666 + ], + "type": "text", + "content": ", which is sent to a decoder for RGB and density calculation." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "type": "text", + "content": "Y.Zhan, Z.Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 117, + 479, + 311 + ], + "blocks": [ + { + "bbox": [ + 137, + 117, + 479, + 311 + ], + "lines": [ + { + "bbox": [ + 137, + 117, + 479, + 311 + ], + "spans": [ + { + "bbox": [ + 137, + 117, + 479, + 311 + ], + "type": "image", + "image_path": "b587c683564864b326b922bb7e11c8da4bd82e7fe1666680f15827eaf33d1462.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 319, + 482, + 419 + ], + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 419 + ], + "type": "text", + "content": "Fig. 3: The overall pipeline of our KFD-NeRF. Our designed deformation field calculates final deformations based on two sources of knowledge. At the observation stage, system observations are directly output by a shallow observation MLP. At the prediction and fusion stage, we calculate predictions based on system dynamics, and further fuse observations and predictions to obtain final deformation estimations. At the spatial reconstruction stage, tri-plane encoded canonical points are concatenated with raw positions and timestamps, which are further decoded to obtain predicted colors for loss calculation. (*f represents a linear layer, used to obtain the Kalman gain from noise-related terms " + }, + { + "bbox": [ + 130, + 319, + 482, + 419 + ], + "type": "inline_equation", + "content": "\\varepsilon_{t_i}" + }, + { + "bbox": [ + 130, + 319, + 482, + 419 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 319, + 482, + 419 + ], + "type": "inline_equation", + "content": "\\varepsilon_{t_{i - 1}}" + }, + { + "bbox": [ + 130, + 319, + 482, + 419 + ], + "type": "text", + "content": ".)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 441, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 583 + ], + "type": "text", + "content": "In Fig. 2, we illustrate the difference between these two methods with a motion example from time " + }, + { + "bbox": [ + 130, + 441, + 482, + 583 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 130, + 441, + 482, + 583 + ], + "type": "text", + "content": " to time " + }, + { + "bbox": [ + 130, + 441, + 482, + 583 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 130, + 441, + 482, + 583 + ], + "type": "text", + "content": ". Given a 3D point " + }, + { + "bbox": [ + 130, + 441, + 482, + 583 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 441, + 482, + 583 + ], + "type": "text", + "content": " in the real world space, we can identify its corresponding four points in the canonical space at four different timestamps, named " + }, + { + "bbox": [ + 130, + 441, + 482, + 583 + ], + "type": "inline_equation", + "content": "P_{1}, P_{2}, P_{3}" + }, + { + "bbox": [ + 130, + 441, + 482, + 583 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 441, + 482, + 583 + ], + "type": "inline_equation", + "content": "P_{4}" + }, + { + "bbox": [ + 130, + 441, + 482, + 583 + ], + "type": "text", + "content": ". ① shows the feature to be learned at " + }, + { + "bbox": [ + 130, + 441, + 482, + 583 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 441, + 482, + 583 + ], + "type": "text", + "content": " which changes over time. Notice that when the radiance undergoes abrupt changes, the feature space exhibits many high-frequency signals, which are hard to fit. ② shows the backward deformation from the real world space to the canonical space to be learned at " + }, + { + "bbox": [ + 130, + 441, + 482, + 583 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 441, + 482, + 583 + ], + "type": "text", + "content": " which changes over time. These signals are smoother and easier to fit by leveraging the motion's smoothness prior. MLP representations have inductive smoothness bias and are well-suited for learning such deformations. Hence, we employ MLP-based deformation fields to represent motion in this work." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 599, + 369, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 369, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 369, + 611 + ], + "type": "text", + "content": "4.2 Kalman Filter Guided Motion Prediction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Next, we derive the system equations for the dynamic radiance field following Eqs. (4) and (5). As we are trying to model non-rigid motions, each 3D point in the dynamic radiance field could be independently seen as a dynamic system, with its velocity " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{t_i}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " at each frame " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " as a single input " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "u_t" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ". Also note that" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "text", + "content": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "this dynamic system is a single-state system since we only focus on estimating deformation at each frame " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "d\\pmb{x}_{t_i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": "In the next step, we determine the coefficients matrix " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " in Eqs. (4) and (5), each degenerating to a value under a single-input and single-state system. As we directly observe " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "d\\pmb{x}_{t_i}" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": ", the output matrix " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " becomes 1. Since the motion trajectory is unknown, we use a locally-linear motion model to describe this dynamic system: " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "d\\pmb{x}_{t_i} = d\\pmb{x}_{t_{i-1}} + \\Delta t_i \\cdot \\pmb{v}_{t_i}" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "A = 1" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "B = \\Delta t_i" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": "). Further considering noise, our state and output equations become" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 236, + 221, + 481, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 221, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 236, + 221, + 481, + 251 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} d \\boldsymbol {x} _ {t _ {i}} = d \\boldsymbol {x} _ {t _ {i - 1}} + \\Delta t _ {i} \\cdot \\boldsymbol {v} _ {t _ {i}} + n _ {t _ {i}}, \\\\ y _ {t _ {i}} = d \\boldsymbol {x} _ {t _ {i}} + m _ {t _ {i}}. \\end{array} \\right. \\tag {11}", + "image_path": "5411bafdfc85b4be8cbd9426f7c201978c6c11b312b855b443b96678e1bfe266.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 260, + 480, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 480, + 356 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 480, + 356 + ], + "type": "text", + "content": "In Fig. 3, our designed deformation field consists of two parts: an MLP-based observer and a system-dynamics-based predictor. As for the observer, by querying 3D points coordinates " + }, + { + "bbox": [ + 130, + 260, + 480, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 260, + 480, + 356 + ], + "type": "text", + "content": " and timestamps " + }, + { + "bbox": [ + 130, + 260, + 480, + 356 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 260, + 480, + 356 + ], + "type": "text", + "content": ", a two-layer shallow MLP is used to output the mean observation " + }, + { + "bbox": [ + 130, + 260, + 480, + 356 + ], + "type": "inline_equation", + "content": "y_{t_i}" + }, + { + "bbox": [ + 130, + 260, + 480, + 356 + ], + "type": "text", + "content": " (assuming zero-mean Gaussian measurement noise " + }, + { + "bbox": [ + 130, + 260, + 480, + 356 + ], + "type": "inline_equation", + "content": "m_{t_i}" + }, + { + "bbox": [ + 130, + 260, + 480, + 356 + ], + "type": "text", + "content": ") and noise-related terms " + }, + { + "bbox": [ + 130, + 260, + 480, + 356 + ], + "type": "inline_equation", + "content": "\\varepsilon_{t_i}" + }, + { + "bbox": [ + 130, + 260, + 480, + 356 + ], + "type": "text", + "content": ". For the predictor, we follow the locally-linear motion model and first calculate the input " + }, + { + "bbox": [ + 130, + 260, + 480, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{t_i}" + }, + { + "bbox": [ + 130, + 260, + 480, + 356 + ], + "type": "text", + "content": ". Because the deformation of the current frame " + }, + { + "bbox": [ + 130, + 260, + 480, + 356 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 260, + 480, + 356 + ], + "type": "text", + "content": " is being estimated, we only use the information from frame " + }, + { + "bbox": [ + 130, + 260, + 480, + 356 + ], + "type": "inline_equation", + "content": "t_{i-1}" + }, + { + "bbox": [ + 130, + 260, + 480, + 356 + ], + "type": "text", + "content": " and frame " + }, + { + "bbox": [ + 130, + 260, + 480, + 356 + ], + "type": "inline_equation", + "content": "t_{i-2}" + }, + { + "bbox": [ + 130, + 260, + 480, + 356 + ], + "type": "text", + "content": " to approximate the velocity" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 245, + 366, + 480, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 366, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 245, + 366, + 480, + 378 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {t _ {i}} = \\left(d \\boldsymbol {x} _ {t _ {i - 1}} - d \\boldsymbol {x} _ {t _ {i - 2}}\\right) / \\Delta t _ {i}. \\tag {12}", + "image_path": "a6d984ae74259530595de52b8caea240dbbc1eb67def7f7655656363e6972a26.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "text", + "content": "Based on the estimated velocity, we compute the predicted deformation state " + }, + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "inline_equation", + "content": "\\hat{d\\pmb{x}}_{t_i}^{-}" + }, + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "text", + "content": ". The final estimation " + }, + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "inline_equation", + "content": "\\hat{d\\pmb{x}}_{t_i}" + }, + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "text", + "content": " is the sum of " + }, + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "inline_equation", + "content": "\\hat{d\\pmb{x}}_{t_i}^{-}" + }, + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "inline_equation", + "content": "y_{t_i}" + }, + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "text", + "content": " weighted by a Kalman gain " + }, + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "inline_equation", + "content": "K_{t_i}" + }, + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "text", + "content": ". The Kalman gain in Eq. (10) is calculated by current measurement noise and past process noise. In our implementation, we obtain " + }, + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "inline_equation", + "content": "K_{t_i}" + }, + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "text", + "content": " utilizing a linear layer that takes noise-related terms " + }, + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "inline_equation", + "content": "\\varepsilon_{t_i}" + }, + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "inline_equation", + "content": "\\varepsilon_{t_{i-1}}" + }, + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "text", + "content": " and timestamps " + }, + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "inline_equation", + "content": "t_{i-1}" + }, + { + "bbox": [ + 130, + 387, + 480, + 472 + ], + "type": "text", + "content": " as inputs, incorporating historical information. Algorithm 1 summarizes this deformation estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 473, + 480, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 480, + 556 + ], + "type": "text", + "content": "Volume is warped to the canonical space based on the estimated deformation, which is further decoded to obtain density " + }, + { + "bbox": [ + 130, + 473, + 480, + 556 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 473, + 480, + 556 + ], + "type": "text", + "content": " and color. We compute the loss by comparing the rendered values with the ground truth and use it to update the network, following Eq. (3). In addition, we note that the estimated deformation serves as a good supervision for learning the observation MLP, in analogy with observation updating process in Kalman filter. Therefore, we try to minimize the error between the current observation and the estimated deformation with" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 242, + 566, + 480, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 566, + 480, + 594 + ], + "spans": [ + { + "bbox": [ + 242, + 566, + 480, + 594 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {k f} = \\frac {1}{\\mathcal {N}} \\sum_ {\\boldsymbol {x} \\in \\mathcal {N}} \\left\\| y _ {t _ {i}} - \\hat {d} \\boldsymbol {x} _ {t _ {i}} \\right\\| _ {2} ^ {2}. \\tag {13}", + "image_path": "d6ca4bd9dbd047d62cf5cd7a7c53a1d55596c5aa8d856b7ce48da511f6f46dc2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 602, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 602, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 602, + 480, + 666 + ], + "type": "text", + "content": "The weighted " + }, + { + "bbox": [ + 130, + 602, + 480, + 666 + ], + "type": "inline_equation", + "content": "y_{t_i}" + }, + { + "bbox": [ + 130, + 602, + 480, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 602, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\hat{d\\pmb{x}}_{t_i}^{-}" + }, + { + "bbox": [ + 130, + 602, + 480, + 666 + ], + "type": "text", + "content": " stand for two sources of knowledge from the whole system. " + }, + { + "bbox": [ + 130, + 602, + 480, + 666 + ], + "type": "inline_equation", + "content": "y_{t_i}" + }, + { + "bbox": [ + 130, + 602, + 480, + 666 + ], + "type": "text", + "content": " represents the state directly observed by the observation MLP, which lacks the information from past frames and thus has a low confidence in the early stages of training. In contrast, " + }, + { + "bbox": [ + 130, + 602, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\hat{d\\pmb{x}}_{t_i}^{-}" + }, + { + "bbox": [ + 130, + 602, + 480, + 666 + ], + "type": "text", + "content": " represents the current frame information predicted from history, which provides a good prior in the early stages of training." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "type": "text", + "content": "Y.Zhan, Z.Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 129, + 465, + 191 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 309, + 128 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 309, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 309, + 128 + ], + "type": "text", + "content": "Algorithm 1 Deformation Estimations" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 129, + 465, + 191 + ], + "lines": [ + { + "bbox": [ + 132, + 129, + 465, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 129, + 465, + 191 + ], + "type": "table", + "html": "
Prediction:Update:
dxti-1, εti-1 = observer(x, ti-1)yti, εti = observer(x, ti)
dxti-2, εti-2 = observer(x, ti-2)Kti = f(εti, εti-1, ti, ti-1)
vti = (dxti-1 - dxti-2) / Δtidxti = dxti - Kti(yti - dxti - )
dxti = dxti-1 + Δti · vt = 2dxti-1 - dxti-2
", + "image_path": "706867a460b0ed58c33827051379fa4c80b4685ef4614f834b0520ffa27bf722.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 214, + 482, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 289 + ], + "type": "text", + "content": "In the later stages of training, however, the confidence of the predictor drops significantly due to the inherent errors of the modeled motion dynamics, which is gradually compensated by the well learned observation MLP. Our Kalman Filter guided model automatically strikes the balance of both leveraging " + }, + { + "bbox": [ + 130, + 214, + 482, + 289 + ], + "type": "inline_equation", + "content": "\\hat{d\\mathbf{x}}_{t_i}^{-}" + }, + { + "bbox": [ + 130, + 214, + 482, + 289 + ], + "type": "text", + "content": " in the early stages to accelerate convergence and avoid local minimum and also using " + }, + { + "bbox": [ + 130, + 214, + 482, + 289 + ], + "type": "inline_equation", + "content": "y_{t_i}" + }, + { + "bbox": [ + 130, + 214, + 482, + 289 + ], + "type": "text", + "content": " in the later stages to learn fine details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 307, + 276, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 276, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 276, + 319 + ], + "type": "text", + "content": "4.3 Spatial Representation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 327, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 471 + ], + "type": "text", + "content": "MLP-based spatial representations suffer from slow convergence and require a deeper deformation network (e.g., eight layers MLP in [36]) to model 4D scenes. Otherwise, the deformation field can get stuck in a local optimum before learning the radiance field in the canonical space. Voxel grids based spatial representations converge very quickly (e.g., TiNeuVox [10]) but demand high spatial resolution to store fine details, which requires significant memory footprint. Under memory constraints, achieving high resolution in the canonical space can be challenging, resulting in losing scene details and deteriorating the warping quality for each frame. We employ a tri-plane representation, which uses three sets of mutually orthogonal 2D planes to represent 3D space. This low-rank model allows for rapid convergence while significantly reducing memory consumption, making it possible to achieve fast and high-quality canonical space reconstruction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 471, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 520 + ], + "type": "text", + "content": "What's more, there are some inevitable errors due to coordinate shifts when " + }, + { + "bbox": [ + 130, + 471, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\pmb{x}'" + }, + { + "bbox": [ + 130, + 471, + 482, + 520 + ], + "type": "text", + "content": " are encoded by finite resolution tri-plane. To enhance the raw coordinate information, we follow [10] by concatenating encoded tri-plane features with raw coordinate inputs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 537, + 349, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 349, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 349, + 550 + ], + "type": "text", + "content": "4.4 Training Strategy and Regularization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 605 + ], + "type": "text", + "content": "Our model takes into account temporal information so the learning of the current frame partially relies on the results of previous frames. To ensure that previous frames can offer ample priors, we employ a training strategy of gradually releasing the training images in chronological order." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "We also notice that the lack of constraints or priors in the canonical space can affect the performance of the observation branch, which could be regularized by pre-setting the shape in the canonical space. D-NeRF [36] sets frame " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " to be the canonical space and force the deformation output of frame " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " to be masked to 0. Such a hard mask, however, does not allow learning of the deformation" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 90, + 448, + 102 + ], + "type": "text", + "content": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "at frame " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " and may cause the canonical space to be too complex for motions warping." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 186 + ], + "type": "text", + "content": "We instead design a soft regularization term to normalize the difference between the canonical space and the radiance field in the real world at frame " + }, + { + "bbox": [ + 130, + 140, + 480, + 186 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 130, + 140, + 480, + 186 + ], + "type": "text", + "content": " to improve the observation branch. The canonical observation loss of a points batch " + }, + { + "bbox": [ + 130, + 140, + 480, + 186 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 130, + 140, + 480, + 186 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 260, + 186, + 480, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 186, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 260, + 186, + 480, + 213 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c o} = \\frac {1}{\\mathcal {N}} \\sum_ {\\boldsymbol {x} \\in \\mathcal {N}} \\mathbf {1} (t) d \\boldsymbol {x}, \\tag {14}", + "image_path": "104044ddb5c140a8d5fe832417abf690191b1b83fa5df4519ed3a8e7d9376f9c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 218, + 363, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 218, + 363, + 230 + ], + "spans": [ + { + "bbox": [ + 131, + 218, + 363, + 230 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 218, + 363, + 230 + ], + "type": "inline_equation", + "content": "\\mathbf{1}(t) = 1" + }, + { + "bbox": [ + 131, + 218, + 363, + 230 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 131, + 218, + 363, + 230 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 131, + 218, + 363, + 230 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 218, + 363, + 230 + ], + "type": "inline_equation", + "content": "\\mathbf{1}(t) = 0" + }, + { + "bbox": [ + 131, + 218, + 363, + 230 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 131, + 218, + 363, + 230 + ], + "type": "inline_equation", + "content": "t\\neq 0" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 231, + 480, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 231, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 480, + 277 + ], + "type": "text", + "content": "We use a proposal sampling strategy from Mip-NeRF 360 [2] by distilling the density field for occupancy estimation. This online distillation necessitates a loss function " + }, + { + "bbox": [ + 130, + 231, + 480, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{prop}" + }, + { + "bbox": [ + 130, + 231, + 480, + 277 + ], + "type": "text", + "content": " to ensure consistency between the proposal network and our learned model. Please refer to Section 3 in [2] for detailed definition." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 279, + 480, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 480, + 325 + ], + "type": "text", + "content": "Total variation loss " + }, + { + "bbox": [ + 130, + 279, + 480, + 325 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{tv}" + }, + { + "bbox": [ + 130, + 279, + 480, + 325 + ], + "type": "text", + "content": " is a common regularizer in inverse problems, which encourages sparse edges in space. We apply this loss to each of our tri-plane to get " + }, + { + "bbox": [ + 130, + 279, + 480, + 325 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{tv}(\\mathbf{x}^{\\prime})" + }, + { + "bbox": [ + 130, + 279, + 480, + 325 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 279, + 480, + 325 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\prime}" + }, + { + "bbox": [ + 130, + 279, + 480, + 325 + ], + "type": "text", + "content": " are warped 3D points in the canonical space. The total variation loss is" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 197, + 335, + 480, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 335, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 197, + 335, + 480, + 360 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t v} (\\mathbf {x}) = \\frac {1}{| C |} \\sum_ {c, i, j} \\left(\\left\\| \\mathbf {x} _ {c} ^ {i, j} - \\mathbf {x} _ {c} ^ {i - 1, j} \\right\\| _ {2} ^ {2} + \\left\\| \\mathbf {x} _ {c} ^ {i, j} - \\mathbf {x} _ {c} ^ {i, j - 1} \\right\\| _ {2} ^ {2}\\right), \\tag {15}", + "image_path": "f03b8dcc286ec9980c648d099f23bfaa034f883a9457595acaac69516281dc03.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 363, + 479, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 479, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 479, + 386 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 363, + 479, + 386 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 363, + 479, + 386 + ], + "type": "text", + "content": " is a certain plane from the tri-plane collection " + }, + { + "bbox": [ + 130, + 363, + 479, + 386 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 363, + 479, + 386 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 363, + 479, + 386 + ], + "type": "inline_equation", + "content": "i, j" + }, + { + "bbox": [ + 130, + 363, + 479, + 386 + ], + "type": "text", + "content": " are indices on the plane's resolution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 147, + 388, + 287, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 388, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 147, + 388, + 287, + 398 + ], + "type": "text", + "content": "The final loss function becomes" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 214, + 413, + 479, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 413, + 479, + 425 + ], + "spans": [ + { + "bbox": [ + 214, + 413, + 479, + 425 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {i m a g e}} + \\mathcal {L} _ {k f} + \\mathcal {L} _ {c o} + \\mathcal {L} _ {\\text {p r o p}} + \\lambda_ {t v} \\mathcal {L} _ {t v}, \\tag {16}", + "image_path": "d5c47bdd79f86bb30d57739274b860ceb24f4a4e8ac81d7522b75651240ab311.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 430, + 351, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 430, + 351, + 442 + ], + "spans": [ + { + "bbox": [ + 131, + 430, + 351, + 442 + ], + "type": "text", + "content": "and we experimentally choose " + }, + { + "bbox": [ + 131, + 430, + 351, + 442 + ], + "type": "inline_equation", + "content": "\\lambda_{tv}" + }, + { + "bbox": [ + 131, + 430, + 351, + 442 + ], + "type": "text", + "content": " to be " + }, + { + "bbox": [ + 131, + 430, + 351, + 442 + ], + "type": "inline_equation", + "content": "1\\times 10^{-4}" + }, + { + "bbox": [ + 131, + 430, + 351, + 442 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 460, + 282, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 460, + 282, + 474 + ], + "spans": [ + { + "bbox": [ + 131, + 460, + 282, + 474 + ], + "type": "text", + "content": "5 Experimental Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 487, + 200, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 487, + 200, + 498 + ], + "spans": [ + { + "bbox": [ + 131, + 487, + 200, + 498 + ], + "type": "text", + "content": "5.1 Dataset" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 507, + 480, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 480, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 480, + 591 + ], + "type": "text", + "content": "For synthetic data, we use the Dynamic NeRF Synthetic Dataset provided by D-NeRF [36], whose training and testing splits have already been well-organized. For each scene in the synthetic dataset, a photo from an arbitrary view with corresponding camera pose is provided at each timestamp. For real data, we use the Nvidia Real Dynamic Scenes Dataset [47] which consists of 8 dynamic scenes recorded by 12 synchronized cameras. We use 11 camera videos for training and the remaining one for testing." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 609, + 271, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 271, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 271, + 620 + ], + "type": "text", + "content": "5.2 Baselines and Metrics" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": "Due to differences in the format of synthetic and real data, we carefully select cutting-edge baselines to thoroughly validate our method based on comparative experiments. For synthetic data, we test deformation based methods" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "type": "text", + "content": "Y.Zhan, Z.Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 125, + 460, + 295 + ], + "blocks": [ + { + "bbox": [ + 153, + 125, + 460, + 295 + ], + "lines": [ + { + "bbox": [ + 153, + 125, + 460, + 295 + ], + "spans": [ + { + "bbox": [ + 153, + 125, + 460, + 295 + ], + "type": "image", + "image_path": "f0fd7359dd0441ddaecd72c7010e52aabe06fbd6820fa00e1b348d1a2f62f483.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 305, + 480, + 327 + ], + "lines": [ + { + "bbox": [ + 131, + 305, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 131, + 305, + 480, + 327 + ], + "type": "text", + "content": "Fig. 4: Qualitative Comparison of our KFD-NeRF against other dynamic NeRF methods on synthetic data. Zoom in for better details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 137, + 349, + 476, + 619 + ], + "blocks": [ + { + "bbox": [ + 137, + 349, + 476, + 619 + ], + "lines": [ + { + "bbox": [ + 137, + 349, + 476, + 619 + ], + "spans": [ + { + "bbox": [ + 137, + 349, + 476, + 619 + ], + "type": "image", + "image_path": "58c955a9b809652629e70559cf6ab53dca88c2dd7affd03183a3726483e2fe9d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 634, + 480, + 654 + ], + "lines": [ + { + "bbox": [ + 130, + 634, + 480, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 634, + 480, + 654 + ], + "type": "text", + "content": "Fig. 5: Qualitative Comparison of our KFD-NeRF against other dynamic NeRF methods on real data. Zoom in for better details." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "text", + "content": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 115, + 302, + 198 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 302, + 198 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 302, + 198 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 302, + 198 + ], + "type": "table", + "html": "
ModelSynthetic Data
PSNR(dB)↑SSIM↑LPIPS↓Average↓
D-NeRF [36]30.250.9540.0760.0288
TiNeuVox-B [10]32.930.9730.0430.0173
KPlanes [12]30.470.9630.0720.0261
NDVG [17]30.510.9640.0560.0223
V4D [14]33.540.9780.0280.0141
4D-GS [44]33.310.9790.0260.0136
KFD-NeRF35.730.9840.0450.0131
", + "image_path": "ee34d9b83a0b012c92a451d07b07ef976f16bb813e737c49d5def2c67f874465.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 310, + 122, + 479, + 190 + ], + "blocks": [ + { + "bbox": [ + 310, + 122, + 479, + 190 + ], + "lines": [ + { + "bbox": [ + 310, + 122, + 479, + 190 + ], + "spans": [ + { + "bbox": [ + 310, + 122, + 479, + 190 + ], + "type": "table", + "html": "
ModelReal Data
PSNR(dB)↑SSIM↑LPIPS↓Average↓
TiNeuVox-B [10]24.390.7340.3030.0869
KPlanes [12]28.250.8810.1370.0433
Mixvoxels-L [41]26.950.8280.2220.0619
V4D [14]28.160.8840.1360.0426
KFD-NeRF28.750.8910.1150.0389
", + "image_path": "bcf3c3b07f6e52e580a429f62650b181ae038c0c826ebb223a119577bcd9827b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 261, + 481, + 291 + ], + "blocks": [ + { + "bbox": [ + 132, + 208, + 482, + 240 + ], + "lines": [ + { + "bbox": [ + 132, + 208, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 132, + 208, + 482, + 240 + ], + "type": "text", + "content": "Table 1: Quantitative comparison of our KFD-NeRF against other dynamic NeRF methods. We show results on both synthetic and real data. See Sec. 5.2 for more details." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 261, + 481, + 291 + ], + "lines": [ + { + "bbox": [ + 135, + 261, + 481, + 291 + ], + "spans": [ + { + "bbox": [ + 135, + 261, + 481, + 291 + ], + "type": "table", + "html": "
D-NeRFTiNeuVox-BKPlanesNDVGV4D4D-GSMixvoxels-LKFD-NeRF-SKFD-NeRF-L
Training Time↓48hrs40mins60mins35mins7hrs25mins1.5hrs30mins3hrs
Params(MB)↓1650340150275163140175175
", + "image_path": "5ece2e7e4d85ede3788a72577e2960f5945bff51f9c8ca430baddb2f84d31831.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 172, + 298, + 440, + 310 + ], + "lines": [ + { + "bbox": [ + 172, + 298, + 440, + 310 + ], + "spans": [ + { + "bbox": [ + 172, + 298, + 440, + 310 + ], + "type": "text", + "content": "Table 2: Training time and parameter size for different methods." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 342, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 437 + ], + "type": "text", + "content": "D-NeRF [36] (MLP based spatial representation), TiNeuVox-B [10] (voxel grids based spatial representation), NDVG [17] (voxel grids based spatial representation) and 4D-GS [44] (Gaussian points based spatial representation), and feature interpolation based methods KPlanes [12] and V4D [14]. For real data, besides TiNeuVox-B and KPlanes, we further compare multi-view videos reconstruction methods MixVoxels [41]. We train all these methods on a single GeForce RTX3090. See Tab. 2 for detailed training time and parameters consumption of ours and other methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 438, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 559 + ], + "type": "text", + "content": "We provide an exhaustive qualitative and quantitative comparison of our KFD-NeRF with these baseline methods. Three main metrics are reported, namely the peak signal-to-noise ratio (PSNR), the structural similarity index measure (SSIM) [42], and the learned perceptual image patch similarity (LPIPS) [51]. To provide more intuitive results, we further calculate metric \"average\" [1], which is the geometric mean of " + }, + { + "bbox": [ + 130, + 438, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{MSE} = 10^{-\\mathrm{PSNR} / 10}" + }, + { + "bbox": [ + 130, + 438, + 482, + 559 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 438, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\sqrt{1 - \\mathrm{SSIM}}" + }, + { + "bbox": [ + 130, + 438, + 482, + 559 + ], + "type": "text", + "content": ", and LPIPS. Please see Tab. 1 for quantitative comparison and Figs. 4 and 5 for qualitative comparison. Per-scene results can be found in the supplemental material. We strongly recommend readers to watch the supplemental videos for a more intuitive understanding of the results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 574, + 246, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 246, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 246, + 586 + ], + "type": "text", + "content": "5.3 Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 481, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 630 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 630 + ], + "type": "text", + "content": "We conduct ablation studies on both synthetic and real data to validate our various proposed system components. We compare our full model with variants related to " + }, + { + "bbox": [ + 130, + 594, + 481, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{co}" + }, + { + "bbox": [ + 130, + 594, + 481, + 630 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 594, + 481, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{kf}" + }, + { + "bbox": [ + 130, + 594, + 481, + 630 + ], + "type": "text", + "content": " and prediction branch." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "content": "Canonical observation loss " + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{co}" + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "content": ". This loss is designed to regularize a continuous and smooth volume shape in the canonical space for better observations. Specifically, we guide the shape in the canonical space to be close to the shape" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "Y.Zhan, Z.Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 288, + 171 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 288, + 171 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 288, + 171 + ], + "type": "image", + "image_path": "cd705bcd6785d78392fc29bb7bd5776ba1be0f1a51535c4cb5b18492165995bc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 173, + 291, + 315 + ], + "lines": [ + { + "bbox": [ + 132, + 173, + 291, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 291, + 315 + ], + "type": "text", + "content": "Fig. 6: Two sources of knowledge at frame " + }, + { + "bbox": [ + 132, + 173, + 291, + 315 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 132, + 173, + 291, + 315 + ], + "type": "text", + "content": " from our pre-trained dynamic system with inputs up to and including frame " + }, + { + "bbox": [ + 132, + 173, + 291, + 315 + ], + "type": "inline_equation", + "content": "t_{i-1}" + }, + { + "bbox": [ + 132, + 173, + 291, + 315 + ], + "type": "text", + "content": ". Notice that during the early stages of training, our observation MLP has insufficient initialization, resulting in the loss of many details (green boxes on the right). Fortunately, our predictions provide additional priors (green boxes on the left) for the current frame based on information from previous frames, compensating for the loss in observations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 310, + 114, + 474, + 185 + ], + "blocks": [ + { + "bbox": [ + 310, + 114, + 474, + 185 + ], + "lines": [ + { + "bbox": [ + 310, + 114, + 474, + 185 + ], + "spans": [ + { + "bbox": [ + 310, + 114, + 474, + 185 + ], + "type": "table", + "html": "
LkfLcoPSNR(dB)↑SSIM↑LPIPS↓
a)30.150.9140.110
b)✓30.600.9200.106
c)✓31.780.9310.086
d)✓✓32.240.9380.080
", + "image_path": "be19e179074b0b8bd52e65f67996a17fabedbb3b99dad15ca89c8053438f4cc7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 310, + 237, + 472, + 276 + ], + "blocks": [ + { + "bbox": [ + 307, + 187, + 474, + 232 + ], + "lines": [ + { + "bbox": [ + 307, + 187, + 474, + 232 + ], + "spans": [ + { + "bbox": [ + 307, + 187, + 474, + 232 + ], + "type": "text", + "content": "Table 3: Ablation Studies on the " + }, + { + "bbox": [ + 307, + 187, + 474, + 232 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{kf}" + }, + { + "bbox": [ + 307, + 187, + 474, + 232 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 307, + 187, + 474, + 232 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{co}" + }, + { + "bbox": [ + 307, + 187, + 474, + 232 + ], + "type": "text", + "content": ". We show all the ablation combinations and report the average results on synthetic and real data." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 237, + 472, + 276 + ], + "lines": [ + { + "bbox": [ + 310, + 237, + 472, + 276 + ], + "spans": [ + { + "bbox": [ + 310, + 237, + 472, + 276 + ], + "type": "table", + "html": "
ModelDynamicStatic
PSNR(dB)↑SSIM↑PSNR(dB)↑SSIM↑
P.N., (xi, ti) Input25.610.96329.580.812
P.N., (xi, ti, ti-1, ti-2) Input25.660.96429.550.812
w/ Prediction Branch (full)26.710.96729.570.813
", + "image_path": "717e8e70efb40096f50d068fe062bbf926ab7bb1b6df9ba973c0b079f9106800.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 277, + 474, + 330 + ], + "lines": [ + { + "bbox": [ + 307, + 277, + 474, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 474, + 330 + ], + "type": "text", + "content": "Table 4: The ablation study on our prediction branch for real data in average, where dynamic and static areas are counted separately. P.N. is short for Pure MLP Network." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 365, + 480, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 480, + 388 + ], + "type": "text", + "content": "at frame " + }, + { + "bbox": [ + 130, + 365, + 480, + 388 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 130, + 365, + 480, + 388 + ], + "type": "text", + "content": ". In Tab. 3 line " + }, + { + "bbox": [ + 130, + 365, + 480, + 388 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 365, + 480, + 388 + ], + "type": "text", + "content": ", we remove " + }, + { + "bbox": [ + 130, + 365, + 480, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{co}" + }, + { + "bbox": [ + 130, + 365, + 480, + 388 + ], + "type": "text", + "content": " and observe a decrease in model performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 388, + 481, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 481, + 449 + ], + "type": "text", + "content": "Estimation update loss " + }, + { + "bbox": [ + 130, + 388, + 481, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{kf}" + }, + { + "bbox": [ + 130, + 388, + 481, + 449 + ], + "type": "text", + "content": ". This loss tries to minimize the difference between estimated deformation and current observation. This updating process ensures that with each iteration, the observation acquire progressively more accurate information to guide the learning process. In Tab. 3 line " + }, + { + "bbox": [ + 130, + 388, + 481, + 449 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 388, + 481, + 449 + ], + "type": "text", + "content": "), we remove " + }, + { + "bbox": [ + 130, + 388, + 481, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{kf}" + }, + { + "bbox": [ + 130, + 388, + 481, + 449 + ], + "type": "text", + "content": " and witness a decrease in model performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 466, + 342, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 466, + 342, + 477 + ], + "spans": [ + { + "bbox": [ + 132, + 466, + 342, + 477 + ], + "type": "text", + "content": "5.4 Extra Studies on Prediction Branch" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 486, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 605 + ], + "type": "text", + "content": "This study aims to demonstrate the effectiveness of our plug-in Kalman filter. The most crucial step in Kalman filter is to fuse the original network observations with the predictions from system dynamics. Therefore, we directly remove the \"Prediction and Fusion Stage\" in our pipeline and use pure shallow observation MLP to generate deformations. We further conduct an experiment where a pure deformable network takes the current frame and the previous two states as inputs to ablate the effectiveness of the Kalman motion modeling. In Tab. 4, we see a clear drop in performance by pure MLP network-based methods, indicating that simply inputting previous Kalman filter parameters without modeling motion prediction is insufficient." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "The prediction branch we have designed can offer reliable priors in the early stages of training. We visualize these priors through an experiment to gain further insight. Specifically, we train our system with only inputs up to and including frame " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "t_{i-1}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " and then let our system produce rendering results at frame " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " with two different branches: predictions and observations. This experiment" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "text", + "content": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 117, + 477, + 178 + ], + "blocks": [ + { + "bbox": [ + 137, + 117, + 477, + 178 + ], + "lines": [ + { + "bbox": [ + 137, + 117, + 477, + 178 + ], + "spans": [ + { + "bbox": [ + 137, + 117, + 477, + 178 + ], + "type": "image", + "image_path": "d745ebbb4ae374a9c051a1d4d6aba042933f85bde751b1b06031d5e63f3b81d5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 189, + 482, + 255 + ], + "lines": [ + { + "bbox": [ + 130, + 189, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 482, + 255 + ], + "type": "text", + "content": "Fig. 7: An example in Nvidia Dataset [47] to illustrate scale and topological issues in choosing canonical space. From " + }, + { + "bbox": [ + 130, + 189, + 482, + 255 + ], + "type": "inline_equation", + "content": "①" + }, + { + "bbox": [ + 130, + 189, + 482, + 255 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 189, + 482, + 255 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 130, + 189, + 482, + 255 + ], + "type": "text", + "content": " there is a scale change due to moving closer to the camera. From " + }, + { + "bbox": [ + 130, + 189, + 482, + 255 + ], + "type": "inline_equation", + "content": "③" + }, + { + "bbox": [ + 130, + 189, + 482, + 255 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 189, + 482, + 255 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 130, + 189, + 482, + 255 + ], + "type": "text", + "content": " there is a topological change caused by target disappearance. According to this, choosing canonical space " + }, + { + "bbox": [ + 130, + 189, + 482, + 255 + ], + "type": "inline_equation", + "content": "①" + }, + { + "bbox": [ + 130, + 189, + 482, + 255 + ], + "type": "text", + "content": " will suffer from low resolution and choosing canonical space " + }, + { + "bbox": [ + 130, + 189, + 482, + 255 + ], + "type": "inline_equation", + "content": "③" + }, + { + "bbox": [ + 130, + 189, + 482, + 255 + ], + "type": "text", + "content": " will suffer from incomplete topology while " + }, + { + "bbox": [ + 130, + 189, + 482, + 255 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 130, + 189, + 482, + 255 + ], + "type": "text", + "content": " is a better canonical space to choose." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "text", + "content": "effectively simulates the amount of knowledge contained in predictions and observations in the early stages of training when the system encounters new input data. In Fig. 6 we show two sources of knowledge from prediction branch and from direct output of observation MLP at frame " + }, + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 277, + 482, + 361 + ], + "type": "text", + "content": ". We see an obvious deficiency in the observation MLP when initializing new input frames and are pleased to find that the prediction branch compensates for this loss based on information from previous frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 378, + 312, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 378, + 312, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 378, + 312, + 392 + ], + "type": "text", + "content": "6 Discussion and Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "content": "Limitations. Our method relies on a well-reconstructed radiance field in the canonical space, which in our pipeline is guided by " + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{co}" + }, + { + "bbox": [ + 130, + 402, + 482, + 521 + ], + "type": "text", + "content": ". This design, however, will partially fail, if the chosen canonical space exhibits significant scale changes or even topological changes related to other frames. In Fig. 7 we use an example to illustrate the influences caused by choosing canonical space. This phenomenon, however, cannot be mitigated by precisely setting the canonical space as we lack a priori access to the input data. We note that some works [1,33] focus on addressing scale or topological issues in radiance fields reconstruction. Nevertheless, these issues are not the main focus of our paper and will be explored in future works to further refine our model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Conclusion. In this work, we present KFD-NeRF, a Kalman filter guided NeRF, for 4D dynamic view synthesis. We model the dynamic radiance field as a dynamic system in control theory and use Kalman filter to estimate the deformation states based on both observations and predictions. We further enhance our observation by encoding canonical space with an efficient tri-plane and by regularizing the shape in the canonical space. Through our temporal training strategy and newly derived pipeline, KFD-NeRF achieves state-of-the-art view synthesis performance among a variety of dynamic NeRF methods. We hope the dynamic system modeling of 4D radiance fields will encourage researchers to explore motion contextual information. KFD-NeRF hopefully inspires the use of existing sequential methods mainly in control theory and visual state estimation to further improve 4D view synthesis and deformation estimations tasks." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 243, + 101 + ], + "type": "text", + "content": "Y.Zhan, Z.Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 163 + ], + "type": "text", + "content": "This research was supported in part by JSPS KAKENHI Grant Numbers 24K22318, 22H00529, 20H05951, JST-Mirai Program JPMJMI23G1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 205, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 205, + 480, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 480, + 249 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 480, + 249 + ], + "type": "text", + "content": "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A Multiscale Representation for Anti-aliasing Neural Radiance Fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 250, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 250, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 250, + 480, + 293 + ], + "type": "text", + "content": "2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: MipNeRF 360: Unbounded Anti-aliased Neural Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 293, + 480, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 337 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 337 + ], + "type": "text", + "content": "3. Becker, P., Pandya, H., Gebhardt, G., Zhao, C., Taylor, C.J., Neumann, G.: Recurrent Kalman Networks: Factorized Inference in High-dimensional Deep Feature Spaces. In: International conference on machine learning. pp. 544-552. PMLR (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 337, + 480, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 480, + 370 + ], + "type": "text", + "content": "4. Cao, A., Johnson, J.: Hexplane: A Fast Representation for Dynamic Scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 130-141 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 370, + 480, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 480, + 414 + ], + "type": "text", + "content": "5. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient Geometry-aware 3D Generative Adversarial Networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 414, + 480, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 436 + ], + "type": "text", + "content": "6. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial Radiance Fields. In: European Conference on Computer Vision. pp. 333-350. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 436, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 480, + 468 + ], + "type": "text", + "content": "7. Choe, J., Choy, C., Park, J., Kweon, I.S., Anandkumar, A.: Spacetime Surface Regularization for Neural Dynamic Scene Reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17871-17881 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 468, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 512 + ], + "type": "text", + "content": "8. Coskun, H., Achilles, F., DiPietro, R., Navab, N., Tombari, F.: Long Short-term Memory Kalman Filters: Recurrent Neural Estimators for Pose Regularization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5524-5532 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 512, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 555 + ], + "type": "text", + "content": "9. Du, Y., Zhang, Y., Yu, H.X., Tenenbaum, J.B., Wu, J.: Neural Radiance Flow for 4D View Synthesis and Video Processing. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14304-14314. IEEE Computer Society (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 555, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 588 + ], + "type": "text", + "content": "10. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Niefsner, M., Tian, Q.: Fast Dynamic Radiance Fields with Time-aware Neural Voxels. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 588, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 480, + 621 + ], + "type": "text", + "content": "1. Fraccaro, M., Kamronn, S., Paquet, U., Winther, O.: A Disentangled Recognition and Nonlinear Dynamics Model for Unsupervised Learning. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "text", + "content": "2. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit Radiance Fields in Space, Time, and Appearance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12479-12488 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "text", + "content": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "13. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance Fields without Neural Networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "text", + "content": "14. Gan, W., Xu, H., Huang, Y., Chen, S., Yokoya, N.: V4D: Voxel for 4D Novel View Synthesis. IEEE Transactions on Visualization and Computer Graphics (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 173, + 480, + 205 + ], + "type": "text", + "content": "15. Gao, C., Saraf, A., Kopf, J., Huang, J.B.: Dynamic View Synthesis from Dynamic Monocular Video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5712-5721 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 206, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 206, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 134, + 206, + 480, + 239 + ], + "type": "text", + "content": "16. Guen, V.L., Thome, N.: Disentangling Physical Dynamics from Unknown Factors for Unsupervised Video Prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11474-11484 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 240, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 240, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 133, + 240, + 480, + 274 + ], + "type": "text", + "content": "17. Guo, X., Chen, G., Dai, Y., Ye, X., Sun, J., Tan, X., Ding, E.: Neural Deformable Voxel Grid for Fast Optimization of Dynamic View Synthesis. In: Proceedings of the Asian Conference on Computer Vision. pp. 3757-3775 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 274, + 480, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 274, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 134, + 274, + 480, + 317 + ], + "type": "text", + "content": "18. Guo, X., Sun, J., Dai, Y., Chen, G., Ye, X., Tan, X., Ding, E., Zhang, Y., Wang, J.: Forward Flow for Novel View Synthesis of Dynamic Scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16022-16033 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 318, + 480, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 318, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 134, + 318, + 480, + 351 + ], + "type": "text", + "content": "19. Haarnoja, T., Ajay, A., Levine, S., Abbeel, P.: Backprop Kf: Learning Discriminative Deterministic State Estimators. Advances in neural information processing systems 29 (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 352, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 352, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 133, + 352, + 480, + 385 + ], + "type": "text", + "content": "20. Hu, T., Liu, S., Chen, Y., Shen, T., Jia, J.: EfficientNeRF: Efficient Neural Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12902-12911 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 386, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 386, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 386, + 480, + 407 + ], + "type": "text", + "content": "21. Jang, H., Kim, D.: D-tensorf: Tensorial Radiance Fields for Dynamic Scenes. arXiv preprint arXiv:2212.02375 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 408, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 408, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 133, + 408, + 480, + 430 + ], + "type": "text", + "content": "22. Kajiya, J.T., Von Herzen, B.P.: Ray Tracing Volume Densities. ACM SIGGRAPH computer graphics 18(3), 165-174 (1984)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 431, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 431, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 133, + 431, + 480, + 463 + ], + "type": "text", + "content": "23. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D Gaussian Splatting for Real-time Radiance Field Rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 464, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 464, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 133, + 464, + 480, + 508 + ], + "type": "text", + "content": "24. Li, T., Slavcheva, M., Zollhoefer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., et al.: Neural 3D Video Synthesis from Multi-view Video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5521-5531 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 509, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 509, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 133, + 509, + 480, + 542 + ], + "type": "text", + "content": "25. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural Scene Flow Fields for Space-time View Synthesis of Dynamic Scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6498-6508 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 543, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 543, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 133, + 543, + 480, + 564 + ], + "type": "text", + "content": "26. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural Sparse Voxel Fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 565, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 565, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 565, + 480, + 609 + ], + "type": "text", + "content": "27. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust Dynamic Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13-23 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 609, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 609, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 609, + 480, + 643 + ], + "type": "text", + "content": "28. Lu, G., Ouyang, W., Xu, D., Zhang, X., Gao, Z., Sun, M.T.: Deep Kalman Filtering Network for Video Compression Artifact Reduction. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 568-584 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "text", + "content": "29. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d Gaussians: Tracking by Persistent Dynamic View Synthesis. arXiv preprint arXiv:2308.09713 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "Y.Zhan, Z.Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "30. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "31. Müller, T., Evans, A., Schied, C., Keller, A.: Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Transactions on Graphics (ToG) 41(4), 1-15 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "text", + "content": "32. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable Neural Radiance Fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "type": "text", + "content": "33. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A Higher-dimensional Representation for Topologically Varying Neural Radiance Fields. arXiv preprint arXiv:2106.13228 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 248, + 482, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 292 + ], + "type": "text", + "content": "34. Park, S., Son, M., Jang, S., Ahn, Y.C., Kim, J.Y., Kang, N.: Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4212-4221 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 335 + ], + "type": "text", + "content": "35. Peng, S., Niemeyer, M., Mescheder, L., Pollefeys, M., Geiger, A.: Convolutional Occupancy Networks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16. pp. 523-540. Springer (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "type": "text", + "content": "36. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-neRF: Neural Radiance Fields for Dynamic Scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 402 + ], + "type": "text", + "content": "37. Revach, G., Shlezinger, N., Ni, X., Escoriza, A.L., Van Sloun, R.J., Eldar, Y.C.: KalmanNet: Neural Network Aided Kalman Filtering for Partially Known Dynamics. IEEE Transactions on Signal Processing 70, 1532-1547 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 402, + 482, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 445 + ], + "type": "text", + "content": "38. Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., Liu, Y.: Tensor4d: Efficient Neural 4D Decomposition for High-fidelity Dynamic Reconstruction and Rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16632-16642 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "text", + "content": "39. Sun, C., Sun, M., Chen, H.T.: Direct Voxel Grid Optimization: Super-fast CO convergence for Radiance Fields Reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5459-5469 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 479, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 522 + ], + "type": "text", + "content": "40. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid Neural Radiance Fields: Reconstruction and Novel View Synthesis of a Dynamic Scene from Monocular Video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12959-12970 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "text", + "content": "41. Wang, F., Tan, S., Li, X., Tian, Z., Song, Y., Liu, H.: Mixed Neural Voxels for Fast Multi-view Video Synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19706-19716 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "type": "text", + "content": "42. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image Quality Assessment: from Error Visibility to Structural Similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 588, + 454, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 588, + 454, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 454, + 600 + ], + "type": "text", + "content": "43. Welch, G., Bishop, G., et al.: An Introduction to the Kalman Filter (1995)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 600, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 482, + 632 + ], + "type": "text", + "content": "44. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering. arXiv preprint arXiv:2310.08528 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "45. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time Neural Irradiance Fields for Free-viewpoint Video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9421-9431 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "text", + "content": "KFD-NeRF: Rethinking Dynamic NeRF with Kalman Filter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 380 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 148 + ], + "type": "text", + "content": "46. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction. arXiv preprint arXiv:2309.13101 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 481, + 194 + ], + "type": "text", + "content": "47. Yoon, J.S., Kim, K., Gallo, O., Park, H.S., Kautz, J.: Novel View Synthesis of Dynamic Scenes with Globally Coherent Depths from a Monocular Camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5336-5345 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "48. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for Real-time Rendering of Neural Radiance Fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5752-5761 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 269 + ], + "type": "text", + "content": "49. Yuan, W., Lv, Z., Schmidt, T., Lovegrove, S.: Star: Self-supervised Tracking and Reconstruction of Rigid Objects in Motion with Neural Rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13144-13152 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "type": "text", + "content": "50. Zhang, H., Yang, Z., Xiong, H., Zhu, T., Long, Z., Wu, W.: Transformer Aided Adaptive Extended Kalman Filter for Autonomous Vehicle Mass Estimation. Processes 11(3), 887 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "type": "text", + "content": "51. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 380 + ], + "type": "text", + "content": "52. Zhou, L., Luo, Z., Shen, T., Zhang, J., Zhen, M., Yao, Y., Fang, T., Quan, L.: Kfnet: Learning Temporal Camera Relocalization Using Kalman Filtering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4919-4928 (2020)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "Y.Zhan, Z.Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/KMTalk_ Speech-Driven 3D Facial Animation with Key Motion Embedding/8f222025-36f0-4f3b-ab78-77e262291c3e_content_list.json b/2024/KMTalk_ Speech-Driven 3D Facial Animation with Key Motion Embedding/8f222025-36f0-4f3b-ab78-77e262291c3e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..948960dbe942c34fd90dccb92570081bc0666a5f --- /dev/null +++ b/2024/KMTalk_ Speech-Driven 3D Facial Animation with Key Motion Embedding/8f222025-36f0-4f3b-ab78-77e262291c3e_content_list.json @@ -0,0 +1,1540 @@ +[ + { + "type": "text", + "text": "KMTalk: Speech-Driven 3D Facial Animation with Key Motion Embedding", + "text_level": 1, + "bbox": [ + 241, + 140, + 759, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhihao $\\mathrm{Xu}^{1^*}$ , Shengjie Gong $^{1^*}$ , Jiapeng Tang $^{2}$ , Lingyu Liang $^{1}$ , Yining Huang $^{1}$ , Haojie Li $^{1}$ , and Shuangping Huang $^{1,3\\dagger}$", + "bbox": [ + 215, + 218, + 782, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ South China University of Technology", + "bbox": [ + 367, + 270, + 635, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Technical University of Munich", + "bbox": [ + 390, + 285, + 611, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Pazhou Laboratory", + "bbox": [ + 429, + 299, + 573, + 313 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{eezhihaoxu,eeshengjiegong}@mail.scut.edu.cn, eehsp@scut.edu.cn", + "bbox": [ + 261, + 314, + 741, + 327 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We present a novel approach for synthesizing 3D facial motions from audio sequences using key motion embeddings. Despite recent advancements in data-driven techniques, accurately mapping between audio signals and 3D facial meshes remains challenging. Direct regression of the entire sequence often leads to over-smoothed results due to the ill-posed nature of the problem. To this end, we propose a progressive learning mechanism that generates 3D facial animations by introducing key motion capture to decrease cross-modal mapping uncertainty and learning complexity. Concretely, our method integrates linguistic and data-driven priors through two modules: the linguistic-based key motion acquisition and the cross-modal motion completion. The former identifies key motions and learns the associated 3D facial expressions, ensuring accurate lip-speech synchronization. The latter extends key motions into a full sequence of 3D talking faces guided by audio features, improving temporal coherence and audio-visual consistency. Extensive experimental comparisons against existing state-of-the-art methods demonstrate the superiority of our approach in generating more vivid and consistent talking face animations. Consistent enhancements in results through the integration of our proposed learning scheme with existing methods underscore the efficacy of our approach.", + "bbox": [ + 259, + 364, + 738, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Speech-driven $\\cdot$ 3D Facial Animation $\\cdot$ Key Motion", + "bbox": [ + 261, + 654, + 689, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 695, + 374, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Speech-driven 3D facial animation aims to create realistic talking heads that synchronize with input speech. It plays a significant role in many applications of virtual reality, like film production, computer gaming, and education [24, 43].", + "bbox": [ + 212, + 726, + 782, + 771 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The main challenge of speech-driven 3D talking faces lies in the ill-posed problem caused by the cross-modal mapping uncertainty from the speech domain to", + "bbox": [ + 212, + 772, + 784, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Authors contributed equally.", + "bbox": [ + 236, + 810, + 437, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author.", + "bbox": [ + 240, + 824, + 397, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/79288203aa1c38b920ac9f2d7ce605aaac642a3c4776e940c5f5fb8a7be84a81.jpg", + "image_caption": [ + "Fig. 1: Compared to the state-of-the-art method Selftalk, our approach can produce more vivid lip motions from speeches, since we introduce linguistic priors to characterize key motions and utilize data-driven priors to interpolate non-key motions." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 787, + 338 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the 3D motion domain. Since there may be multiple plausible outputs for input audio, effective regularizations and constraints should be integrated into the system, to generate vivid facial motions. The related methods can be roughly divided into linguistic-based methods [3, 11, 29, 46, 54] and data-driven methods [8, 12, 16, 26, 34, 36, 41, 49, 52, 53]. For linguistic-based methods [3, 11, 29, 46, 54], a set of intricate phoneme-to-viseme mapping rules is manually designed to generate the talking mouth based on priors from visemes or linguistic knowledge. While these methods explicitly control the animation of articulation processes, such as procedural lip sync with animation curves, their focus is mainly on localized facial movements, like those of the mouth area, lacking a systematic approach for modeling comprehensive facial motion. Thanks to the established audio-to-face datasets, learning-based methods [12, 16, 34, 36, 41, 49, 52, 53] choose to map audio signals into 3D facial meshes in a data-driven manner. Most of these works [8, 12, 16, 26, 34, 36, 41, 48, 49, 52, 53] typically formulated the cross-modal mapping of 3D talking face generation as a regression task, such as MeshTalk [41], FaceFormer [12], and SelfTalk [36]. While achieving impressive performance, they exhibit common limitations in their learning schemes. Firstly, they directly learn the ambiguous cross-modal mapping between audio and facial expression sequences, always leading to sub-optimal results in terms of temporal coherence and audio-visual consistency. Secondly, these methods typically regress the entire sequence without considering key motion cues, hindering the capture of detailed facial dynamics and accurate lip movements, particularly in complex facial expressions such as puckering or opening the mouth (as depicted in Fig. 1). Lastly, they overlook linguistic priors essential for simulating the articulation process, thereby limiting their ability to achieve precise lip-speech synchronization.", + "bbox": [ + 212, + 429, + 787, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, inspired by the keyframe-based video generation techniques observed in recent studies [27, 33, 51, 56], which prioritize the generation of", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Xu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "keyframes before adding detailed elements, we introduce a progressive learning mechanism that generates realistic 3D facial animations from audio inputs by incorporating key motion embeddings. The key idea is to initially generate key facial expressions, and then interpolate the intermediate motions to obtain the entire motion sequence, which significantly reduces the uncertainty of cross-modal mapping and eases the learning difficulty. Concretely, our method integrates linguistic and data-driven priors through two modules: the linguistic-based key motion acquisition and the cross-modal motion completion. The linguistic-based key motion acquisition module utilizes phoneme-based localization methods to identify temporal indices of key motion, which correspond to significant motion snapshots aligned with phoneme changes in the audio. Once the key motion indexes are determined, a key motion decoder interprets associated 3D facial meshes from corresponding audio features. This highlights those distinct facial expressions and facilitates lip-speech synchronization. The cross-modal motion completion module expands non-continuous key motions into a full sequence of continuous face motions using audio features as guidance. This process enhances audio-mesh alignment and improves the temporal smoothness of output facial mesh sequences. The contributions of our work are summarized as follows:", + "bbox": [ + 212, + 146, + 787, + 419 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a progressive learning mechanism to generate speech-driven 3D talking faces. It uses linguistic priors to initially generate key motions, and then interpolate key motions into complete motions via data-driven priors.", + "- We propose the use of phoneme-based localization methods to capture key facial motions. It effectively captures significant expression transitions aligned with phoneme changes in audio, improving lip-speech synchronization.", + "- We design a cross-modal facial motion completion module to produce full sequences of 3D talking faces using synthesized key motions and audio features. It further enhances lip-speech synchronization accuracy while facilitating temporal coherence in facial motions." + ], + "bbox": [ + 225, + 422, + 782, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extensive experimental comparisons on the BIWI [13] and VOCASET [8] datasets demonstrate that our method outperforms existing state-of-the-art approaches in more accurate and realistic talking face generation. Detailed ablation studies confirm the effectiveness of our proposed key motion capture technique. Additionally, consistent improvements in results by combining our proposed scheme with existing methods validate the efficacy of our design. Our code and weights will be at the project website: https://github.com/ffxzh/KMTalk.", + "bbox": [ + 212, + 575, + 789, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 699, + 387, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While existing research [1,5,6,10,15,18,19,23,25,35,38,42,47,50,55,59,61] focuses on 2D talking heads, we focus on audio-driven 3D facial animations in this work, which can be roughly categorized into linguistics-based and data-driven methods.", + "bbox": [ + 212, + 726, + 782, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Linguistic-based Methods", + "text_level": 1, + "bbox": [ + 215, + 789, + 472, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Linguistic-based methods [3, 7, 11, 14, 22, 29, 30, 46, 54] establish a set of intricate phoneme-to-viseme mapping rules for animating the mouth. For example, the", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "KMTalk", + "bbox": [ + 673, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dynamic viseme model proposed by Taylor et al. [46] exploits the one-to-many mapping of phonemes to lip motions. JALI [11] considers the many-to-many mapping between phonemes and visemes. More recently, Bao et al. [3] introduced a novel parameterized viseme fitting algorithm that extracts viseme parameters from speech videos using phonemic priors. Leveraging linguistic priors, these methods indicate the articulation process by providing animators with explicit control over animation, thus boosting their performance in lip-speech synchronization. However, these approaches primarily focus on animating the lip region, lacking a comprehensive strategy for animating the entire face. In our work, we leverage linguistic priors to detect key frames with significant expression changes from audio in an analytical manner, without the need for supervised training.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Data-driven Methods", + "text_level": 1, + "bbox": [ + 214, + 330, + 439, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the development of deep learning technology and the availability of high-quality datasets, data-driven methods [4, 8, 12, 16, 20, 26, 34, 36, 37, 41, 44, 45, 49, 52, 53, 60] were proposed to synthesize entire 3D facial animation. Some methods [8, 12, 20, 40, 52] attempt to establish a direct audio-to-visual mapping through regression. Person-specific approaches [20, 40] can usually obtain plausible facial motions because of the relatively consistent talking style. VOCA [8] incorporates a robust audio feature extraction model capable of capturing various speaking styles, which can generate realistic speaker-independent animation and shows its wide applicability. MeshTalk [41] constructed a categorical latent space to adaptively generate motions based on the separated audio-correlated and audiouncorrelated facial information. FaceFormer [12] introduced two biased attention mechanisms and integrated the self-supervised pre-trained speech representations for the ill-posed and data scarcity issues. CodeTalker [53] proposed the discrete motion prior which regards the cross-modal mapping as a code query task in a finite proxy space of the learned codebook. SelfTalk [36] proposed a self-supervised approach to construct a lip-reading interpreter and speech recognizer to enhance the comprehensibility of generated lip movements. While data-driven approaches have shown impressive performance, accurately learning cross-modal audio-visual mappings remains challenging due to inherent uncertainties. These methods often regress the entire sequence, leading to over-smoothing and a lack of detailed facial dynamics. In contrast, our approach employs a coarse-to-fine learning mechanism that separates the problem into key motion capture and motion completion stages. This approach effectively mitigates cross-modal mapping uncertainties and reduces learning complexity, resulting in more precise and dynamic facial animation synthesis.", + "bbox": [ + 212, + 349, + 787, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 746, + 331, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 214, + 773, + 341, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $\\mathbf{x}$ represents the raw audio input and $\\hat{\\mathbf{Y}} = (\\hat{\\mathbf{y}}_1,\\dots,\\hat{\\mathbf{y}}_N)\\in \\mathbb{R}^{N\\times V\\times 3}$ denotes the corresponding ground-truth sequence of facial movement over a neutral template, where $N$ indicates the number of visual frames and $V$ denotes the number", + "bbox": [ + 212, + 792, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Xu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/732b3f9865d772653c6d7af7d745fbc23d87894c3accf04f47592faefe020760.jpg", + "image_caption": [ + "(a) The overview pipeline of our proposed KMTalk." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 781, + 268 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/575248935e24c38bbf8194ce56065892ff51a23bdb35467e4b9c07603f61ebce.jpg", + "image_caption": [ + "(b) The details of two key modules in KMTalk.", + "Fig. 2: Fig. 2a illustrates the overview pipeline of our proposed KMTalk. Initially, the Audio Encoder takes the input raw audio $\\mathbf{x}$ and encodes it into audio features $\\mathbf{A}$ . Subsequently, in the LKMA module, key motions $\\mathbf{K}$ are generated from the audio $\\mathbf{x}$ and $\\mathbf{A}$ . Finally, the CMC module reintroduces audio features $\\mathbf{A}$ to extend these key motions $\\mathbf{K}$ into a full sequence $\\mathbf{Y}$ . Fig. 2b presents the details of two key modules in KMTalk. In the Linguistic-based Key Motion Acquisition, a Phoneme-based Localization Method is used to identify key motion indices $\\mathcal{I}$ from raw audio $\\mathbf{x}$ . Based on audio features $\\mathbf{A}$ and $\\mathcal{I}$ , the Key Motion-focused Decoder generates key motions $\\mathbf{K}$ . In the Cross-modal Motion Completion, the Motion Flow Encoder processes $\\mathbf{K}$ and $\\mathcal{I}$ , producing motion flow features $\\Phi$ . Then, with the dynamic fusion weight $\\mathbf{G}$ , the Multimodal-Guided Decoder combines $\\Phi$ and $\\mathbf{A}$ to decode the final motion sequence $\\mathbf{Y}$ ." + ], + "image_footnote": [], + "bbox": [ + 217, + 282, + 782, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of vertices in the facial mesh. The objective is to synthesize $\\mathbf{Y} = (\\mathbf{y}_1, \\dots, \\mathbf{y}_N)$ that is similar to $\\hat{\\mathbf{Y}}$ , driven by the raw audio $\\mathbf{x}$ . The generated sequence should ensure lip synchronization with the audio while exhibiting natural facial movements.", + "bbox": [ + 212, + 585, + 785, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the domain gap between modalities and the ill-posed nature of directly translating audio to facial movement sequences, it is a challenging task that often results in over-smooth or poorly synchronized lip movements. To address these issues, this paper introduces a coarse-to-fine approach with key motion embedding, integrating both linguistic and data-driven priors. The overview pipeline is presented in Fig. 2a. In the LKMA module, linguistic priors are introduced to locate and generate higher-quality key motions (see Sec. 3.2), followed by the CMC module, where these key motions are fleshed out into a complete sequence of facial motions (see Sec. 3.3).", + "bbox": [ + 212, + 631, + 787, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Linguistic-based Key Motion Acquisition", + "text_level": 1, + "bbox": [ + 214, + 787, + 599, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the realm of audio-driven 3D facial animation, it presents significant challenges to precisely define which frames constitute key motions. An alternative", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "KMTalk", + "bbox": [ + 673, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and simple solution is to use uniform or random sampling to determine the positions of these key motions. Although these approaches can boost performance to a certain degree due to the reduced learning complexity (refer to Sec. 4.2 for ablation studies), they fail to utilize the correlation between audio content and facial movements. However, we can leverage linguistic priors to capture pronounced articulatory actions, which are identifiable at phoneme boundaries. This approach circumvents the issue of over-smooth in the output sequence, thereby enhancing the overall quality of the results.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in the left of Fig. 2b, the Linguistic-based Key Motion Acquisition (LKMA) module receives as inputs the raw audio $\\mathbf{x}$ and the audio features $\\mathbf{A} = (\\mathbf{a}_1,\\dots,\\mathbf{a}_N)\\in \\mathbb{R}^{N\\times d}$ , where the audio features are derived from the Audio Encoder that utilizes the wav2vec 2.0 pre-trained model [2], processing the raw audio $\\mathbf{x}$ as its input. Then it takes raw audio $\\mathbf{x}$ as input to produce the key motion indices $\\mathcal{I} = \\{i_1,\\dots,i_m\\}$ , where $i_j\\in \\{1,\\dots,N\\},\\forall i_j\\in \\mathcal{I}$ , through the proposed Phoneme-based Localization method. Subsequently, the Key Motion-focused Decoder utilizes the audio features $\\mathbf{A}$ and the key motion indices $\\mathcal{I}$ to generate key motions $\\mathbf{K} = (\\mathbf{k}_{i_1},\\dots,\\mathbf{k}_{i_m})\\in \\mathbb{R}^{m\\times V\\times 3}$ consisting of $m$ frames of facial movement which are located on the key motion positions, where $\\mathbf{k}_{i_j}\\simeq \\hat{\\mathbf{y}}_{i_j},\\forall i_j\\in \\mathcal{I}$ . The process of the LKMA module is expressed as:", + "bbox": [ + 212, + 268, + 789, + 435 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I}, \\mathbf {K} = \\operatorname {L K M A} (\\mathbf {x}, \\mathbf {A}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 448, + 784, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Phoneme-based Localization. At phoneme boundaries, a noticeable offset is observed in the articulator movement, with visualization results available in supplementary materials. Furthermore, the phoneme boundary effects underscore the ease with which the boundaries of phonemes can be perceived [17]. Both experimental and theoretical analyses have demonstrated a distinct position-mapping relationship between the phoneme boundaries in the audio and the significant elements in the motion sequence, specifically the key motions. The mapping relationship can be harnessed to facilitate the initial alignment between the audio and visual modalities.", + "bbox": [ + 212, + 476, + 787, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, an Automatic Speech Recognition model [28] is first utilized to obtain the text content from the raw audio $\\mathbf{x}$ . Then, a Montreal Forced Aligner [31] is adopted to align the audio and the text, producing the start and the end timestamps for each phoneme. Finally, the indices of these motion frames corresponding to the timestamps are regarded as key motion indices $\\mathcal{I}$ , and the corresponding motions compose the key motions $\\mathbf{K}$ .", + "bbox": [ + 212, + 612, + 787, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Key Motion-focused Decoder. It is utilized to synthesize key motions $\\mathbf{K}$ of superior quality. Initially, employing $\\mathcal{I}$ as indices, we extract the corresponding aligned audio features $\\mathbf{A_k} = (\\mathbf{a}_{i_1},\\dots,\\mathbf{a}_{i_m})\\in \\mathbb{R}^{m\\times d}$ from the comprehensive audio features $\\mathbf{A}$ . Subsequently, it adopts a modified transformer-based architecture, processes $\\mathbf{A_k}$ to generate the key motions $\\mathbf{K}$ .", + "bbox": [ + 212, + 703, + 787, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Loss Function. Intuitively, a straightforward approach to optimize the key motions $\\mathbf{K}$ involves utilizing $\\mathcal{I}$ to index $\\hat{\\mathbf{Y}}$ , resulting in $\\hat{\\mathbf{Y}}_{\\mathbf{k}} = (\\hat{\\mathbf{y}}_{i_1},\\dots ,\\hat{\\mathbf{y}}_{i_m})$ , which serves as the supervision for the training process. However, key motions typically occupy non-adjacent positions within the entire sequence. Hence, given the lack", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Xu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of inter-frame contextual information, attempting direct frame-by-frame regression of $\\mathbf{K}$ towards $\\hat{\\mathbf{Y}}_{\\mathbf{k}}$ may fall short in achieving accurate facial expressions, as well as producing smooth and realistic animation.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To address this limitation, we adopt a pseudo-complete sequence training method that utilizes the ground-truth frame labels $\\hat{\\mathbf{Y}}$ at the non-key indices $\\mathcal{I}' = \\{1,\\dots,N\\} \\setminus \\mathcal{I}$ and the generated key motions $\\mathbf{K}$ at the key indices $\\mathcal{I}$ to form a predicted pseudo-complete sequence $\\mathbf{Y}_{\\mathbf{p}} \\in \\mathbb{R}^{N \\times V \\times 3}$ . Then, the model is trained by minimizing the loss between the pseudo-complete sequence $\\mathbf{Y}_{\\mathbf{p}}$ and the ground-truth sequence $\\hat{\\mathbf{Y}}$ . This enables the model to capture subtle changes between key motions and adjacent ground-truth frames, thereby mitigating inter-frame jitter and achieving more accurate regression of facial expressions. Following the SelfTalk [36], the loss function is formulated as:", + "bbox": [ + 212, + 191, + 784, + 330 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {L K M A} = \\lambda_ {1} \\mathcal {L} _ {r e c} + \\lambda_ {2} \\mathcal {L} _ {v e l} + \\lambda_ {3} \\mathcal {L} _ {l a t} + \\lambda_ {4} \\mathcal {L} _ {c t c}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 343, + 784, + 358 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lambda_{1} = 1000.0, \\lambda_{2} = 1000.0, \\lambda_{3} = 0.001$ , and $\\lambda_{4} = 0.0001$ in all of our experiments. $\\mathcal{L}_{rec}$ , $\\mathcal{L}_{vel}$ , and $\\mathcal{L}_{lat}$ are measured by mean square error, while $\\mathcal{L}_{ctc}$ is quantified by CTC Loss. The reconstruction loss $\\mathcal{L}_{rec}$ quantifies the discrepancy between the predicted and the ground-truth facial movements. The velocity loss $\\mathcal{L}_{vel}$ reduces frame jitter, ensuring smooth and natural lip movements. The latent consistency loss $\\mathcal{L}_{lat}$ assesses the variance between latent features extracted from both the audio and lip shape encoders, aiming to align the learned audio and lip features. Lastly, the text consistency loss $\\mathcal{L}_{ctc}$ evaluates the difference between the lip-reading decoder's output and the original text, ensuring the intelligibility of the lip-reading results.", + "bbox": [ + 212, + 369, + 784, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Cross-modal Motion Completion", + "text_level": 1, + "bbox": [ + 214, + 545, + 532, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A straightforward method to obtain a complete talking face sequence is to directly integrate key motions into the entire face movements. However, it is important to note that while key motions capture essential facial dynamics, they may not encompass all details of non-key motions. This direct integration could result in mismatches between augmented non-key motions and their corresponding audio segments (see Sec. 4.2 for ablation studies). To mitigate this issue, we introduce a Cross-modal Motion Completion (CMC) module that jointly combines the audio features $\\mathbf{A}$ , key motions $\\mathbf{K}$ , and key motion indices $\\mathcal{I}$ to generate a complete sequence of 3D facial meshes $\\mathbf{Y}$ . The process can be formulated as:", + "bbox": [ + 212, + 571, + 782, + 708 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} = \\operatorname {C M C} (\\mathbf {A}, \\mathbf {K}, \\mathcal {I}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 428, + 720, + 784, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The details of the CMC module are illustrated on the right of Fig. 2b.", + "bbox": [ + 212, + 750, + 720, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motion Flow Encoder. Key motions serve as a kinematic prior for the remaining frames, offering valuable insights into facial dynamics. To effectively capture the motion flow information provided by key motions, we draw inspiration from some manifold methods [32] to acquire motion flow features $\\Phi \\in \\mathbb{R}^{N\\times d}$ from the key motions $\\mathbf{K}$ . Specifically, we first encode the key motions $\\mathbf{K}$ into key motion", + "bbox": [ + 212, + 763, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "KMTalk", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "context tokens $\\Phi_{\\mathbf{k}} \\in \\mathbb{R}^{m \\times d}$ by multiple transformer-encoder layers. At the same time, the indices of non-key motions $\\mathcal{T}'$ are encoded into positional encodings by a sinusoidal position embedding layer, representing the non-key frame positions. Then, we adopt the cross-attention layers to extract the intermediate tokens $\\Phi_{\\mathrm{non - key}} \\in \\mathbb{R}^{(N - m) \\times d}$ , with key and value from linear transformations of $\\Phi_{\\mathbf{k}}$ and the query is the positional encodings of non-key frames indices. Above all, the implicit motion manifold proposed in CITL [32] is utilized to arrange $\\Phi_{\\mathbf{k}}$ and $\\Phi_{\\mathrm{non - key}}$ based on their indices, followed by a 1D convolution for fusion, ultimately obtaining the motion flow features $\\Phi$ .", + "bbox": [ + 212, + 146, + 787, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multimodal-Guided Decoder. The non-key motions' features in the complete motion sequence feature estimation are derived from the global context interpolation of key motions, which has a certain degree of information loss due to the audio feature selection process in the Key Motion-focused Decoder (Sec. 3.2). Hence, we have devised a multimodal-guided decoding approach that incorporates audio modalities to furnish comprehensive information across the entire temporal scale, alongside motion flow to offer facial motion priors. These elements serve to guide and constrain the decoding process, thereby facilitating the precise generation of the motion sequence. Technically, we simply employ the gated mechanism [9,57,58] for the modality fusion, which can be formulated as:", + "bbox": [ + 212, + 282, + 787, + 434 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} = \\sigma ([ \\mathbf {A}, \\boldsymbol {\\Phi} ] \\mathbf {W}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 449, + 785, + 465 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} = \\mathbf {G} \\odot \\mathbf {A} + (1 - \\mathbf {G}) \\odot \\boldsymbol {\\Phi}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 487, + 784, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\odot$ represents the element-wise multiplication operation, $[\\cdot, \\cdot]$ denotes the concatenation operation, $\\mathbf{W} \\in \\mathbb{R}^{2d \\times d}$ is a parameter and $\\sigma$ is a sigmoid function, while $\\mathbf{G} \\in \\mathbb{R}^{N \\times d}$ dynamically selects features from the audio features $\\mathbf{A}$ and the motion flow features $\\Phi$ . Then, the Motion Decoder, a transformer-based structural model, is employed to transform the fusion features $\\mathbf{Z}$ into the complete 3D facial movement sequence $\\hat{\\mathbf{Y}}$ .", + "bbox": [ + 212, + 508, + 787, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Loss Function. We train the CMC module utilizing the reconstruction loss and velocity loss. The total loss function is defined as:", + "bbox": [ + 212, + 599, + 787, + 631 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C M C} = \\mathcal {L} _ {r e c} + \\mathcal {L} _ {v e l}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 646, + 784, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 214, + 686, + 366, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset. BIWI [13] consists of 40 paired audio-visual sentences from 14 subjects. The 3D facial geometries, consisting of 23370 vertices, were captured at a frame rate of 25fps, and the average duration of each sequence was 4.67 seconds. We adopt the same evaluation protocol as FaceFormer [12] on the BIWI dataset. Specifically, the training set (BIWI-Train) comprises 190 sentences, while the validation set (BIWI-Val) encompasses 24 sentences. The dataset is divided into two testing sets: BIWI-Test-A, which comprises 24 sentences articulated by six subjects observed during training, and BIWI-Test-B, which consists of 32 sentences", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Xu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e9a7c9d2061499b48cb00f06f7e6658c1dad926205a959da47760c1440ed76f8.jpg", + "table_caption": [ + "Table 1: Quantitative comparisons on the BIWI-Test-A and VOCA-Test datasets. The results of Lip-Vertex Error (LVE) and the upper-Face Dynamics Deviation (FDD) are reported. For both metrics, the lower the better." + ], + "table_footnote": [], + "table_body": "
MethodsBIWI-Test-AVOCA-Test
LVE↓×10-4mmFDD↓×10-5mmLVE↓×10-5mmFDD↓×10-7mm
VOCA [8]6.55638.18164.92454.8447
MeshTalk [41]5.91815.10254.54415.2062
FaceFormer [12]5.30774.64084.10904.6675
CodeTalker [53]4.79144.11703.94454.5422
SelfTalk [36]4.24853.57613.22384.0912
KMTalk (Ours)3.96542.54462.26394.0594
", + "bbox": [ + 220, + 196, + 785, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "uttered by eight unseen subjects. VOCASET [8] consists of 480 paired audiovisual sequences from 12 subjects. Each sequence is recorded at a frame rate of 60fps and ranges in duration from 3 to 4 seconds. The 3D face mesh for each sequence consists of 5023 vertices. To ensure a fair comparison, we used identical training (VOCA-Train), validation (VOCA-Val), and testing (VOCA-Test) partitions as methods [12,36,53].", + "bbox": [ + 212, + 340, + 784, + 430 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines. We compare against current state-of-the-arts method, including VOCA [8], MeshTalk [41]. FaceFormer [12], CodeTalker [53], and SelfTalk [36]. Faceformer [12] employs a transformer-based model to incorporate long-term audio context and synthesizes sequential motions in an autoregressive manner. CodeTalker [53] introduces discrete motion priors to enable self-reconstruction of real facial movements, mitigating the issue of excessive smoothing in facial motion. SelfTalk [36] designs a learning-based recognizer to minimize the domain gap between diverse modalities.", + "bbox": [ + 212, + 431, + 785, + 551 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following CodeTalker [53] and SelfTalk [36], we adopt two metrics for the quantitative evaluation of speech-driven facial animation: lip vertex error (LVE) to measure lip synchronization and upper-face dynamics deviation (FDD) to assess the overall facial dynamics. The LVE for each frame is defined as the maximal L2 error among all lip vertices for each frame and takes the average over all frames. This L2 error is computed by comparing the predictions with the processed 3D face geometry data. FDD is introduced to quantify the variation in facial dynamics between a synthetic motion sequence and the reference sequence. The implementation of FDD is to calculate the difference between the variances of vertex offsets in the upper-face region and the variances of ground truth vertex offsets. In addition, we visualize the prediction results for qualitative evaluation.", + "bbox": [ + 212, + 551, + 785, + 733 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. For a fair comparison, KMTalk operates at a frame rate of 30 fps on VOCASET and 25 fps on BIWI, following the setting of previous methods [12, 36, 53]. Also, it can naturally adapt to a higher frame rate, as shown in the supplementary materials. In the LKMA module, we first employ the Phoneme-based Localization method to process the raw audio and obtain key motion indices for data preprocessing, which costs less than 10 minutes on two datasets [8, 13]. Secondly, we train the Key Motion Decoder on a single NVIDIA", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "KMTalk", + "bbox": [ + 673, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/626f767d17967288bffb105075a792a0cb286ae1b11d81e744077f7135ada470.jpg", + "image_caption": [ + "Fig. 3: Qualitative comparisons on VOCA-Test (left) and BIWI-Test-B (right). We provide visual comparisons of facial animations synchronized with six syllables extracted from the test speech sequences. The 1st, 3rd, and 5th rows display synthesized meshes and their corresponding ground-truths, while the 2nd, 4th, and 6th rows visualize the L2 loss for individual frames. Our method demonstrates more precise mouth movement on syllables like /æ/ that require a wide-open mouth. For syllables that start with a closed mouth and then slightly open, such as /bɪ/, our KMTalk generates more synchronized motion sequences visually. The last row visualizes the mean square errors of different methods across all sentences in the test set for a specific subject." + ], + "image_footnote": [], + "bbox": [ + 225, + 141, + 774, + 452 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "RTX 3090 for 200 epochs (about 2 hours) using the Adam optimizer [21]. The learning rate is initialized as $10^{-4}$ , and the mini-batch size is set to 1. In the CMC module, we train for 200 epochs (approximately 2 hours) with the same training settings as the Key Motion Decoder. It is noteworthy that, since the training of the two modules is independent of each other, we can train both modules concurrently to enhance training efficiency.", + "bbox": [ + 212, + 619, + 787, + 710 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Comparisons against State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 214, + 736, + 645, + 752 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Quantitative Comparisons. We computed the lip vertex error (LVE) and facial dynamics deviation (FDD) for all sequences within the BIWI-Test-A and VOCA-Test datasets. According to Table 1, our proposed KMTalk demonstrated lower errors compared to the alternative methods examined. Notably, the lip vertex error for our method on the VOCA-Test dataset is $30\\%$ lower than the", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Xu et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "recently introduced SelfTalk [36], and the FDD is $27\\%$ lower than SelfTalk on the BIWI-Test-A dataset, providing compelling evidence for the advantages of our proposed KMTalk. This indicates that our approach is more effective in achieving audio-visual alignment, thereby leading to improved lip synchronization.", + "bbox": [ + 212, + 145, + 782, + 205 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative Comparisons. In Fig. 3, we visualize the output facial meshes from different methods as well as ground truths for reference. Additionally, we display error maps calculated from the vertex L2 loss between the generated and ground truth meshes. It is evident that our method consistently yields lower errors across different speech sequences, demonstrating its ability to generate more accurate facial animation sequences. Notably, for representative syllables (e.g. $/a\\mathrm{e}/$ ), KMTalk closely approximates ground truth, excelling in synthesizing accurate lip movements for syllables requiring significant mouth opening. Additionally, for syllables starting with mouth closure followed by a slight opening (e.g., $/b\\mathrm{i}/$ ), KMTalk produces more natural and synchronized motions. We recommend that readers watch the supplementary video for more detailed comparisons. It showcases KMTalk's capability to generate coherent, realistic animations with precise lip synchronization.", + "bbox": [ + 212, + 207, + 784, + 402 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "User Studies. A user study stands as a dependable evaluation method in the context of 3D talking faces. Following the strategy of Faceformer [12], we conduct pairwise comparisons between our method and baselines [12,36,53], as well as ground truths. This study encompassed the assessment of two key metrics: perceptual lip synchronization and facial realism. Participants were presented with side-by-side comparisons and were tasked with selecting the better facial animation based on their personal preferences. We computed the ratio of user preferences as a measurement of satisfaction evaluation on BIWI-Test-B and VOCA-Test. We randomly sampled 30 examples from each test set and compared the performance of KMTalk with four aforementioned settings on each sample. Therefore, we constructed a total of 240 different video pairs and randomly selected 24 video pairs for the two metrics assessments for each participant. Our user study involved 30 participants with a strong capability for audio-visual perception, resulting in 720 effective evaluation entries. As demonstrated in Table 2, our approach indicates superior perceptual lip synchronization and facial realism. For instance, a noteworthy $60.0\\%$ of users favored our lip synchronization method on BIWI-Test-B in comparison to SelfTalk [36]. Overall, it shows that KMTalk can generate more favorable facial animations from speeches.", + "bbox": [ + 212, + 402, + 785, + 675 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 712, + 401, + 726 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we perform ablation studies on to evaluate the influence of different components within our proposed KMTalk framework on the quality of the generated 3D talking faces. The quantitative results on BIWI are in Table 3, and the qualitative results are in Fig. 4. In addition, the results of the ablation study on VOCA-Test can be found in the supplementary material. In Table 4, we further investigate the robustness of our approach to different Phoneme-based localization methods and the possible errors during phoneme extraction.", + "bbox": [ + 212, + 734, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "KMTalk", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/24f418aeb43f306d67f11794a19ceeaaee8fd76ace2ac2de7933fe4320aed7c3.jpg", + "table_caption": [ + "Table 2: User study results on BIWI-Test-B and VOCA-Test." + ], + "table_footnote": [], + "table_body": "
MethodMetricBIWI-Test-BVOCA-Test
competitorourscompetitorours
Ours vs. FaceFormerLip Sync24.4%75.6%26.7%73.3%
Realism25.6%74.4%28.9%71.1%
Ours vs. CodeTalkerLip Sync31.1%68.9%37.8%62.2%
Realism27.8%72.2%35.6%64.4%
Ours vs. SelfTalkLip Sync40.0%60.0%43.3%56.7%
Realism38.9%61.1%41.1%58.9%
Ours vs. GTLip Sync54.4%45.6%56.7%43.3%
Realism52.2%47.8%56.7%43.3%
", + "bbox": [ + 261, + 170, + 753, + 339 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/cee9155627e2d006166bb9239cd0ccc495a5bc84db907c904b689b40424259ab.jpg", + "table_caption": [ + "Table 3: Ablation study for our components on BIWI-Test-A." + ], + "table_footnote": [], + "table_body": "
Phoneme-based Localization MethodKey Motion-focused Audio in CMCLVE↓FDD↓
Decoder✓4.24853.5761
---4.16482.8713
-✓✓4.13812.9546
✓-✓4.88593.2780
✓✓-3.96542.5446
", + "bbox": [ + 272, + 380, + 730, + 474 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "What's the effect of the Phoneme-based localization method for key motion capture? The Phoneme-based Localization Method enables us to identify key frames of speech with notable facial expression transitions. We can replace it with uniform sampling, neglecting the crucial content information of the audio. Specifically, we experimented using uniform sampling at a rate of $33\\%$ to closely align with the number of key motions. For further comparisons under different numbers of sampled elements, please refer to the supplementary material. In Table 3, we observe degradation in all metrics with uniform sampling. This underscores the importance of accurate key motion capture in speech-driven talking face generation. Additionally, it showcases that the linguistic-based key motion capture is better equipped to mitigate audio-visual uncertainty and recover more precise facial motions.", + "bbox": [ + 212, + 506, + 785, + 686 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "What's the effect of the Key Motion Decoder? Our method employs a specialized key motion decoder to generate facial meshes based on keyframe indices obtained from phoneme-based localization methods. An alternative approach to generating key motions is to select corresponding facial meshes from a complete motion sequence produced by an existing method such as SelfTalk [36]. Table 3 demonstrates that both metrics deteriorated but still outperformed the state-of-the-art method SelfTalk. This indicates that the Key Motion Decoder can enhance the quality of key motion generation, resulting in more plausible facial animations. It also suggests that even if the captured key motions are not accurate enough, the CMC module can further refine output full motions.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Xu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/47975d6ec9b6d62e30e566b978fbc70e53c4b4ea2aa0ba8b68ef5a5206e109cf.jpg", + "image_caption": [ + "Fig. 4: Qualitative ablation studies on the input speech \"specifically\". For each method variant, we removed one of three modules: PLM (Phoneme-based Localization Method), KMD (Key Motion-focused Decoder), and AG (Audio Guidance in CMC). Error maps between generated and the ground-truth mesh sequence were visualized. Our final model yielded the best results, showcasing the effectiveness of each module." + ], + "image_footnote": [], + "bbox": [ + 218, + 148, + 782, + 393 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "What's the effect of audio guidance in the Cross-modal Motion Completion? The Cross-modal Motion Completion leverages audio features to guide the completion of the full motion sequence. To assess its usefulness, we implement a method variant that removes the audio feature guidance. Table 3 demonstrates a notable degradation in both metrics, particularly with a $24\\%$ increase in Lip Vertex Error (LVE) and a $27\\%$ increase in upper-face dynamics deviation (FDD). This suggests that audio information plays a crucial role in refining fine-grained lip movements and enhancing audio-visual consistency and temporal smoothness. Despite the degradation, the FDD metric can still outperform state-of-the-art methods, underscoring the significance of key motion capture for achieving temporally coherent full motion synthesis.", + "bbox": [ + 212, + 516, + 787, + 683 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Is our method sensitive to different Phomene-based Localization methods? We experimented with different Automatic Speech Recognition (ASR) models, such as Auto-avs r [28] and Whisper [39]. The results in the first three rows of Table 4 show that our KMTalk method consistently maintained high performance, achieving at least a $21\\%$ improvement in FDD regardless of the ASR model used [28, 39]. This highlights the robustness of our approach across various ASR models. Besides, to simulate phoneme localization deviations, we shifted all key motion indices extracted by Auto-avs r [28] one frame to the right. The results in the last row of Table 4 indicate negligible variations. This demonstrates the robustness of our method to inaccurate key frame localization.", + "bbox": [ + 212, + 689, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "KMTalk", + "bbox": [ + 673, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b3d462ccbb68e84ad47bac09de429e964dd9afc31f8de4bded6d589b85e333dd.jpg", + "table_caption": [ + "Table 4: Robust analysis of Phoneme-based Localization on BIWI-Test-A." + ], + "table_footnote": [], + "table_body": "
MethodsLVE↓ ×10-4mmFDD↓ ×10-5mm
Auto-avsr [28]3.96542.5446
Whisper-large [39]4.07182.8141
Whisper-tiny [39]4.06432.8083
Auto-avsr+offset3.99912.6420
", + "bbox": [ + 232, + 210, + 485, + 340 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/bd1dd33089855e014118ae2055aedf4d8ef828d9fb5e6f531c8800e7058149b1.jpg", + "table_caption": [ + "Table 5: The results of integrating our proposed KMTalk with existing methods on BIWI-Test-A." + ], + "table_footnote": [], + "table_body": "
MethodsLVE↓ ×10-4mmFDD↓ ×10-5mm
FaceFormer [12]Original5.30774.6408
After5.27934.2654
CodeTalker [53]Original4.79144.1170
After4.50963.9043
SelfTalk [36]Original4.24853.5761
After4.11222.8668
", + "bbox": [ + 498, + 210, + 763, + 340 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.3 Integration with Existing Methods", + "text_level": 1, + "bbox": [ + 215, + 359, + 549, + 376 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Existing approaches focus on enhancing prediction outcomes by designing elaborate priors or the learning-based recognizer, which may be highly coupled with the proposed architecture of these methods. Our KMTalk introduces a new learning strategy of speech-driven talking face generation, which is orthogonal to these approaches. Therefore, we can explore whether performance can be enhanced by applying our progressive learning scheme without the need for additional fine-tuning of their models. Detailed implementation is in the supplementary material. As shown in Table 5, the results of existing methods are improved after integration with our proposed progressive learning mechanism utilizing key motion embeddings. This further emphasizes the efficacy of our design.", + "bbox": [ + 212, + 382, + 787, + 532 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 554, + 359, + 569 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we introduce KMTalk, a novel method for progressively learning 3D facial animation from speeches using key motion embeddings. It incorporates linguistic priors for key motion generation and extends them to a full motion sequence via data-driven priors. We propose phoneme-based localization methods to determine the temporal position of key facial motions, improving lip-speech synchronization by aligning motion transitions with phoneme changes. Additionally, we design a cross-modal facial motion completion module that synthesizes the entire motion sequence from key motions and audio features, enhancing lip-speech synchronization and motion coherence. Extensive evaluations of the datasets demonstrate KMTalk's superiority over existing methods, producing more accurate and realistic animations. Moreover, coupling our idea with existing methods consistently improves performance, further verifying the efficacy of our proposed progressive learning mechanism based on key motion acquisition. Although the proposed method has demonstrated its robustness to inaccurate keyframe localization, it may encounter errors in dialect variations. Integrating advanced ASR(Automatic Speech Recognition) technology in the future could enhance its adaptability to various speech patterns.", + "bbox": [ + 212, + 582, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Xu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement The research is partially supported by National Key Research and Development Program of China (2023YFC3502900), National Natural Science Foundation of China (No.62176093, 61673182), Key Realm Research and Development Program of Guangzhou (No.202206030001), Guangdong-Hong Kong-Macao Joint Innovation Project (No.2023A0505030016), Guangdong Natural Science Foundation (No. 2024A1515012217).", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 258, + 321, + 275 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alghamdi, M.M., Wang, H., Bulpitt, A.J., Hogg, D.C.: Talking head from speech audio using a pre-trained image generator. In: ACM MM. pp. 5228-5236 (2022)", + "2. Baevski, A., Zhou, Y., Mohamed, A., Auli, M.: wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems 33, 12449-12460 (2020)", + "3. Bao, L., Zhang, H., Qian, Y., Xue, T., Chen, C., Zhe, X., Kang, D.: Learning audio-driven viseme dynamics for 3d face animation. arXiv preprint arXiv:2301.06059 (2023)", + "4. Cao, Y., Tien, W.C., Faloutsos, P., Pighin, F.: Expressive speech-driven facial animation. ACM Trans. Graph 24(4), 1283-1302 (2005)", + "5. Chen, L., Cui, G., Liu, C., Li, Z., Kou, Z., Xu, Y., Xu, C.: Talking-head generation with rhythmic head motion. In: ECCV. pp. 35-51. Springer (2020)", + "6. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: ACCV. pp. 251-263. Springer (2017)", + "7. Cohen, M.M., Clark, R., Massaro, D.W.: Animated speech: Research progress and applications. In: AVSP (2001)", + "8. Cudeiro, D., Bolkart, T., Laidlaw, C., Ranjan, A., Black, M.J.: Capture, learning, and synthesis of 3d speaking styles. In: CVPR. pp. 10101-10111 (2019)", + "9. Dai, G., Zhang, Y., Wang, Q., Du, Q., Yu, Z., Liu, Z., Huang, S.: Disentangling writer and character styles for handwriting generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5977-5986 (2023)", + "10. Das, D., Biswas, S., Sinha, S., Bhowmick, B.: Speech-driven facial animation using cascaded gans for learning of motion and texture. In: ECCV. pp. 408-424. Springer (2020)", + "11. Edwards, P., Landreth, C., Fiume, E., Singh, K.: Jali: an animator-centric viseme model for expressive lip synchronization. ACM Trans. Graph 35(4), 1-11 (2016)", + "12. Fan, Y., Lin, Z., Saito, J., Wang, W., Komura, T.: Faceformer: Speech-driven 3d facial animation with transformers. In: CVPR. pp. 18770-18780 (2022)", + "13. Fanelli, G., Gall, J., Romsdorfer, H., Weise, T., Van Gool, L.: A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia 12(6), 591-598 (2010)", + "14. Fisher, C.G.: Confusions among visually perceived consonants. Journal of speech and hearing research 11(4), 796-804 (1968)", + "15. Guo, Y., Chen, K., Liang, S., Liu, Y.J., Bao, H., Zhang, J.: Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In: ICCV. pp. 5784-5794 (2021)", + "16. Habibie, I., Xu, W., Mehta, D., Liu, L., Seidel, H.P., Pons-Moll, G., Elgharib, M., Theobalt, C.: Learning speech-driven 3d conversational gestures from video. In: Proceedings of the 21st ACM International Conference on Intelligent Virtual Agents. pp. 101-108 (2021)" + ], + "bbox": [ + 218, + 287, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "KMTalk", + "bbox": [ + 673, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Iverson, P., Kuhl, P.K.: Perceptual magnet and phoneme boundary effects in speech perception: Do they arise from a common mechanism? Perception & psychophysics 62, 874-886 (2000)", + "18. Ji, X., Zhou, H., Wang, K., Wu, Q., Wu, W., Xu, F., Cao, X.: Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In: ACM SIGGRAPH. pp. 1-10 (2022)", + "19. Ji, X., Zhou, H., Wang, K., Wu, W., Loy, C.C., Cao, X., Xu, F.: Audio-driven emotional video portraits. In: CVPR. pp. 14080-14089 (2021)", + "20. Karras, T., Aila, T., Laine, S., Herva, A., Lehtinen, J.: Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Trans. Graph 36(4), 1-12 (2017)", + "21. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "22. Lewis, J.: Automated lip-sync: Background and techniques. IEEE Trans Vis Comput Graph 2(4), 118-122 (1991)", + "23. Liang, B., Pan, Y., Guo, Z., Zhou, H., Hong, Z., Han, X., Han, J., Liu, J., Ding, E., Wang, J.: Expressive talking head generation with granular audio-visual control. In: CVPR. pp. 3387-3396 (2022)", + "24. Liu, C.: An analysis of the current and future state of 3d facial animation techniques and systems (2009)", + "25. Liu, X., Xu, Y., Wu, Q., Zhou, H., Wu, W., Zhou, B.: Semantic-aware implicit neural audio-driven video portrait generation. In: ECCV. pp. 106-125. Springer (2022)", + "26. Liu, Y., Xu, F., Chai, J., Tong, X., Wang, L., Huo, Q.: Video-audio driven real-time facial animation. ACM Trans. Graph 34(6), 1-10 (2015)", + "27. Lu, L., Wu, R., Lin, H., Lu, J., Jia, J.: Video frame interpolation with transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3532-3542 (2022)", + "28. Ma, P., Haliassos, A., Fernandez-Lopez, A., Chen, H., Petridis, S., Pantic, M.: Auto-avsr: Audio-visual speech recognition with automatic labels. arxiv 2023. arXiv preprint arXiv:2303.14307", + "29. Massaro, D., Cohen, M., Tabain, M., Beskow, J., Clark, R.: 12 animated speech: research progress and applications (2012)", + "30. Mattheyses, W., Verhelst, W.: Audiovisual speech synthesis: An overview of the state-of-the-art. Speech Communication 66, 182-217 (2015)", + "31. McAuliffe, M., Sonderegger, M.: English mfa acoustic model v2.2.1. Tech. rep., https://mfa-models.readthedocs.io/acoustic/English/EnglishMFAacousticmodelv2_2_1.html (May 2023)", + "32. Mo, C.A., Hu, K., Long, C., Wang, Z.: Continuous intermediate token learning with implicit motion manifold for keyframe based motion interpolation. In: CVPR. pp. 13894-13903 (2023)", + "33. Niklaus, S., Mai, L., Liu, F.: Video frame interpolation via adaptive convolution. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 670-679 (2017)", + "34. Nocentini, F., Ferrari, C., Berretti, S.: Learning landmarks motion from speech for speaker-agnostic 3d talking heads generation. arXiv preprint arXiv:2306.01415 (2023)", + "35. Pang, Y., Zhang, Y., Quan, W., Fan, Y., Cun, X., Shan, Y., Yan, D.m.: Dpe: Disentanglement of pose and expression for general video portrait editing. In: CVPR. pp. 427-436 (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Xu et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "36. Peng, Z., Luo, Y., Shi, Y., Xu, H., Zhu, X., Liu, H., He, J., Fan, Z.: Selftalk: A self-supervised commutative training diagram to comprehend 3d talking faces. arXiv preprint arXiv:2306.10799 (2023)", + "37. Pham, H.X., Wang, Y., Pavlovic, V.: End-to-end learning for 3d facial animation from speech. In: ICMI. pp. 361-365 (2018)", + "38. Prajwal, K., Mukhopadhyay, R., Namboodiri, V.P., Jawahar, C.: A lip sync expert is all you need for speech to lip generation in the wild. In: ACM MM. pp. 484-492 (2020)", + "39. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision. In: International Conference on Machine Learning. pp. 28492-28518. PMLR (2023)", + "40. Richard, A., Lea, C., Ma, S., Gall, J., De la Torre, F., Sheikh, Y.: Audio-and gaze-driven facial animation of codec avatars. In: WACV. pp. 41-50 (2021)", + "41. Richard, A., Zollhöfer, M., Wen, Y., De la Torre, F., Sheikh, Y.: Meshtalk: 3d face animation from speech using cross-modality disentanglement. In: ICCV. pp. 1173-1182 (2021)", + "42. Shen, S., Li, W., Zhu, Z., Duan, Y., Zhou, J., Lu, J.: Learning dynamic facial radiance fields for few-shot talking head synthesis. In: ECCV. pp. 666-682. Springer (2022)", + "43. Tanaka, H., Nakamura, S., et al.: The acceptability of virtual characters as social skills trainers: usability study. JMIR human factors 9(1), e35358 (2022)", + "44. Tang, J., Dai, A., Nie, Y., Markhasin, L., Thies, J., Nießner, M.: Dphms: Diffusion parametric head models for depth-based tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1111-1122 (2024)", + "45. Taylor, S., Kim, T., Yue, Y., Mahler, M., Krahe, J., Rodriguez, A.G., Hodgins, J., Matthews, I.: A deep learning approach for generalized speech animation. ACM Trans. Graph 36(4), 1-11 (2017)", + "46. Taylor, S.L., Mahler, M., Theobald, B.J., Matthews, I.: Dynamic units of visual speech. In: ACM SIGGRAPH. pp. 275-284 (2012)", + "47. Vougioukas, K., Petridis, S., Pantic, M.: Realistic speech-driven facial animation with gans. International Journal of Computer Vision 128, 1398-1413 (2020)", + "48. Wang, J., Qian, X., Zhang, M., Tan, R.T., Li, H.: Seeing what you said: Talking face generation guided by a lip reading expert. In: CVPR. pp. 14653-14662 (2023)", + "49. Wang, Q., Fan, Z., Xia, S.: 3d-talkemo: Learning to synthesize 3d emotional talking head. arXiv preprint arXiv:2104.12051 (2021)", + "50. Wang, S., Li, L., Ding, Y., Yu, X.: One-shot talking face generation from single-speaker audio-visual correlation learning. In: AAAI. vol. 36, pp. 2531-2539 (2022)", + "51. Wen, S., Liu, W., Yang, Y., Huang, T., Zeng, Z.: Generating realistic videos from keyframes with concatenated gans. IEEE Transactions on Circuits and Systems for Video Technology 29(8), 2337-2348 (2018)", + "52. Wu, H., Zhou, S., Jia, J., Xing, J., Wen, Q., Wen, X.: Speech-driven 3d face animation with composite and regional facial movements. arXiv preprint arXiv:2308.05428 (2023)", + "53. Xing, J., Xia, M., Zhang, Y., Cun, X., Wang, J., Wong, T.T.: Codetalker: Speechdriven 3d facial animation with discrete motion prior. In: CVPR. pp. 12780-12790 (2023)", + "54. Xu, Y., Feng, A.W., Marsella, S., Shapiro, A.: A practical and configurable lip sync method for games. In: Proceedings of Motion on Games, pp. 131-140 (2013)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "KMTalk", + "bbox": [ + 673, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "55. Yi, R., Ye, Z., Zhang, J., Bao, H., Liu, Y.J.: Audio-driven talking face video generation with learning-based personalized head pose. arXiv preprint arXiv:2002.10137 (2020)", + "56. Yin, S., Wu, C., Yang, H., Wang, J., Wang, X., Ni, M., Yang, Z., Li, L., Liu, S., Yang, F., et al.: Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv preprint arXiv:2303.12346 (2023)", + "57. Yu, D., Li, X., Zhang, C., Liu, T., Han, J., Liu, J., Ding, E.: Towards accurate scene text recognition with semantic reasoning networks. In: CVPR. pp. 12113-12122 (2020)", + "58. Yue, X., Kuang, Z., Lin, C., Sun, H., Zhang, W.: Robustscanner: Dynamically enhancing positional clues for robust text recognition. In: ECCV. pp. 135-151. Springer (2020)", + "59. Zhang, B., Qi, C., Zhang, P., Zhang, B., Wu, H., Chen, D., Chen, Q., Wang, Y., Wen, F.: Metaportrait: Identity-preserving talking head generation with fast personalized adaptation. In: CVPR. pp. 22096-22105 (2023)", + "60. Zhang, C., Ni, S., Fan, Z., Li, H., Zeng, M., Budagavi, M., Guo, X.: 3d talking face with personalized pose dynamics. IEEE Trans Vis Comput Graph. (2021)", + "61. Zhou, H., Sun, Y., Wu, W., Loy, C.C., Wang, X., Liu, Z.: Pose-controllable talking face generation by implicitly modularized audio-visual representation. In: CVPR. pp. 4176-4186 (2021)" + ], + "bbox": [ + 215, + 146, + 787, + 424 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Z. Xu et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/KMTalk_ Speech-Driven 3D Facial Animation with Key Motion Embedding/8f222025-36f0-4f3b-ab78-77e262291c3e_model.json b/2024/KMTalk_ Speech-Driven 3D Facial Animation with Key Motion Embedding/8f222025-36f0-4f3b-ab78-77e262291c3e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e663578f3fab18b65e26a3d9ce1fac2d0a2f7464 --- /dev/null +++ b/2024/KMTalk_ Speech-Driven 3D Facial Animation with Key Motion Embedding/8f222025-36f0-4f3b-ab78-77e262291c3e_model.json @@ -0,0 +1,2260 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.243, + 0.141, + 0.761, + 0.186 + ], + "angle": 0, + "content": "KMTalk: Speech-Driven 3D Facial Animation with Key Motion Embedding" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.219, + 0.784, + 0.26 + ], + "angle": 0, + "content": "Zhihao \\(\\mathrm{Xu}^{1^*}\\), Shengjie Gong\\(^{1^*}\\), Jiapeng Tang\\(^{2}\\), Lingyu Liang\\(^{1}\\), Yining Huang\\(^{1}\\), Haojie Li\\(^{1}\\), and Shuangping Huang\\(^{1,3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.271, + 0.636, + 0.287 + ], + "angle": 0, + "content": "\\(^{1}\\) South China University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.391, + 0.286, + 0.612, + 0.3 + ], + "angle": 0, + "content": "2 Technical University of Munich" + }, + { + "type": "text", + "bbox": [ + 0.43, + 0.3, + 0.574, + 0.314 + ], + "angle": 0, + "content": "3 Pazhou Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.315, + 0.742, + 0.328 + ], + "angle": 0, + "content": "{eezhihaoxu,eeshengjiegong}@mail.scut.edu.cn, eehsp@scut.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.365, + 0.74, + 0.642 + ], + "angle": 0, + "content": "Abstract. We present a novel approach for synthesizing 3D facial motions from audio sequences using key motion embeddings. Despite recent advancements in data-driven techniques, accurately mapping between audio signals and 3D facial meshes remains challenging. Direct regression of the entire sequence often leads to over-smoothed results due to the ill-posed nature of the problem. To this end, we propose a progressive learning mechanism that generates 3D facial animations by introducing key motion capture to decrease cross-modal mapping uncertainty and learning complexity. Concretely, our method integrates linguistic and data-driven priors through two modules: the linguistic-based key motion acquisition and the cross-modal motion completion. The former identifies key motions and learns the associated 3D facial expressions, ensuring accurate lip-speech synchronization. The latter extends key motions into a full sequence of 3D talking faces guided by audio features, improving temporal coherence and audio-visual consistency. Extensive experimental comparisons against existing state-of-the-art methods demonstrate the superiority of our approach in generating more vivid and consistent talking face animations. Consistent enhancements in results through the integration of our proposed learning scheme with existing methods underscore the efficacy of our approach." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.655, + 0.69, + 0.67 + ], + "angle": 0, + "content": "Keywords: Speech-driven \\(\\cdot\\) 3D Facial Animation \\(\\cdot\\) Key Motion" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.696, + 0.375, + 0.71 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.727, + 0.784, + 0.772 + ], + "angle": 0, + "content": "Speech-driven 3D facial animation aims to create realistic talking heads that synchronize with input speech. It plays a significant role in many applications of virtual reality, like film production, computer gaming, and education [24, 43]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.773, + 0.785, + 0.803 + ], + "angle": 0, + "content": "The main challenge of speech-driven 3D talking faces lies in the ill-posed problem caused by the cross-modal mapping uncertainty from the speech domain to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.238, + 0.811, + 0.438, + 0.825 + ], + "angle": 0, + "content": "*Authors contributed equally." + }, + { + "type": "page_footnote", + "bbox": [ + 0.241, + 0.825, + 0.398, + 0.84 + ], + "angle": 0, + "content": "†Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "Z. Xu et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.788, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.351, + 0.788, + 0.395 + ], + "angle": 0, + "content": "Fig. 1: Compared to the state-of-the-art method Selftalk, our approach can produce more vivid lip motions from speeches, since we introduce linguistic priors to characterize key motions and utilize data-driven priors to interpolate non-key motions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.789, + 0.808 + ], + "angle": 0, + "content": "the 3D motion domain. Since there may be multiple plausible outputs for input audio, effective regularizations and constraints should be integrated into the system, to generate vivid facial motions. The related methods can be roughly divided into linguistic-based methods [3, 11, 29, 46, 54] and data-driven methods [8, 12, 16, 26, 34, 36, 41, 49, 52, 53]. For linguistic-based methods [3, 11, 29, 46, 54], a set of intricate phoneme-to-viseme mapping rules is manually designed to generate the talking mouth based on priors from visemes or linguistic knowledge. While these methods explicitly control the animation of articulation processes, such as procedural lip sync with animation curves, their focus is mainly on localized facial movements, like those of the mouth area, lacking a systematic approach for modeling comprehensive facial motion. Thanks to the established audio-to-face datasets, learning-based methods [12, 16, 34, 36, 41, 49, 52, 53] choose to map audio signals into 3D facial meshes in a data-driven manner. Most of these works [8, 12, 16, 26, 34, 36, 41, 48, 49, 52, 53] typically formulated the cross-modal mapping of 3D talking face generation as a regression task, such as MeshTalk [41], FaceFormer [12], and SelfTalk [36]. While achieving impressive performance, they exhibit common limitations in their learning schemes. Firstly, they directly learn the ambiguous cross-modal mapping between audio and facial expression sequences, always leading to sub-optimal results in terms of temporal coherence and audio-visual consistency. Secondly, these methods typically regress the entire sequence without considering key motion cues, hindering the capture of detailed facial dynamics and accurate lip movements, particularly in complex facial expressions such as puckering or opening the mouth (as depicted in Fig. 1). Lastly, they overlook linguistic priors essential for simulating the articulation process, thereby limiting their ability to achieve precise lip-speech synchronization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "To this end, inspired by the keyframe-based video generation techniques observed in recent studies [27, 33, 51, 56], which prioritize the generation of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.674, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KMTalk" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.42 + ], + "angle": 0, + "content": "keyframes before adding detailed elements, we introduce a progressive learning mechanism that generates realistic 3D facial animations from audio inputs by incorporating key motion embeddings. The key idea is to initially generate key facial expressions, and then interpolate the intermediate motions to obtain the entire motion sequence, which significantly reduces the uncertainty of cross-modal mapping and eases the learning difficulty. Concretely, our method integrates linguistic and data-driven priors through two modules: the linguistic-based key motion acquisition and the cross-modal motion completion. The linguistic-based key motion acquisition module utilizes phoneme-based localization methods to identify temporal indices of key motion, which correspond to significant motion snapshots aligned with phoneme changes in the audio. Once the key motion indexes are determined, a key motion decoder interprets associated 3D facial meshes from corresponding audio features. This highlights those distinct facial expressions and facilitates lip-speech synchronization. The cross-modal motion completion module expands non-continuous key motions into a full sequence of continuous face motions using audio features as guidance. This process enhances audio-mesh alignment and improves the temporal smoothness of output facial mesh sequences. The contributions of our work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.423, + 0.784, + 0.467 + ], + "angle": 0, + "content": "- We propose a progressive learning mechanism to generate speech-driven 3D talking faces. It uses linguistic priors to initially generate key motions, and then interpolate key motions into complete motions via data-driven priors." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.468, + 0.784, + 0.51 + ], + "angle": 0, + "content": "- We propose the use of phoneme-based localization methods to capture key facial motions. It effectively captures significant expression transitions aligned with phoneme changes in audio, improving lip-speech synchronization." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.511, + 0.784, + 0.57 + ], + "angle": 0, + "content": "- We design a cross-modal facial motion completion module to produce full sequences of 3D talking faces using synthesized key motions and audio features. It further enhances lip-speech synchronization accuracy while facilitating temporal coherence in facial motions." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.423, + 0.784, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.576, + 0.79, + 0.683 + ], + "angle": 0, + "content": "Extensive experimental comparisons on the BIWI [13] and VOCASET [8] datasets demonstrate that our method outperforms existing state-of-the-art approaches in more accurate and realistic talking face generation. Detailed ablation studies confirm the effectiveness of our proposed key motion capture technique. Additionally, consistent improvements in results by combining our proposed scheme with existing methods validate the efficacy of our design. Our code and weights will be at the project website: https://github.com/ffxzh/KMTalk." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.7, + 0.388, + 0.716 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.727, + 0.784, + 0.773 + ], + "angle": 0, + "content": "While existing research [1,5,6,10,15,18,19,23,25,35,38,42,47,50,55,59,61] focuses on 2D talking heads, we focus on audio-driven 3D facial animations in this work, which can be roughly categorized into linguistics-based and data-driven methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.79, + 0.473, + 0.805 + ], + "angle": 0, + "content": "2.1 Linguistic-based Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Linguistic-based methods [3, 7, 11, 14, 22, 29, 30, 46, 54] establish a set of intricate phoneme-to-viseme mapping rules for animating the mouth. For example, the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Z. Xu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "dynamic viseme model proposed by Taylor et al. [46] exploits the one-to-many mapping of phonemes to lip motions. JALI [11] considers the many-to-many mapping between phonemes and visemes. More recently, Bao et al. [3] introduced a novel parameterized viseme fitting algorithm that extracts viseme parameters from speech videos using phonemic priors. Leveraging linguistic priors, these methods indicate the articulation process by providing animators with explicit control over animation, thus boosting their performance in lip-speech synchronization. However, these approaches primarily focus on animating the lip region, lacking a comprehensive strategy for animating the entire face. In our work, we leverage linguistic priors to detect key frames with significant expression changes from audio in an analytical manner, without the need for supervised training." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.331, + 0.441, + 0.345 + ], + "angle": 0, + "content": "2.2 Data-driven Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.789, + 0.732 + ], + "angle": 0, + "content": "With the development of deep learning technology and the availability of high-quality datasets, data-driven methods [4, 8, 12, 16, 20, 26, 34, 36, 37, 41, 44, 45, 49, 52, 53, 60] were proposed to synthesize entire 3D facial animation. Some methods [8, 12, 20, 40, 52] attempt to establish a direct audio-to-visual mapping through regression. Person-specific approaches [20, 40] can usually obtain plausible facial motions because of the relatively consistent talking style. VOCA [8] incorporates a robust audio feature extraction model capable of capturing various speaking styles, which can generate realistic speaker-independent animation and shows its wide applicability. MeshTalk [41] constructed a categorical latent space to adaptively generate motions based on the separated audio-correlated and audiouncorrelated facial information. FaceFormer [12] introduced two biased attention mechanisms and integrated the self-supervised pre-trained speech representations for the ill-posed and data scarcity issues. CodeTalker [53] proposed the discrete motion prior which regards the cross-modal mapping as a code query task in a finite proxy space of the learned codebook. SelfTalk [36] proposed a self-supervised approach to construct a lip-reading interpreter and speech recognizer to enhance the comprehensibility of generated lip movements. While data-driven approaches have shown impressive performance, accurately learning cross-modal audio-visual mappings remains challenging due to inherent uncertainties. These methods often regress the entire sequence, leading to over-smoothing and a lack of detailed facial dynamics. In contrast, our approach employs a coarse-to-fine learning mechanism that separates the problem into key motion capture and motion completion stages. This approach effectively mitigates cross-modal mapping uncertainties and reduces learning complexity, resulting in more precise and dynamic facial animation synthesis." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.747, + 0.332, + 0.763 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.775, + 0.342, + 0.789 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.794, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Let \\(\\mathbf{x}\\) represents the raw audio input and \\(\\hat{\\mathbf{Y}} = (\\hat{\\mathbf{y}}_1,\\dots,\\hat{\\mathbf{y}}_N)\\in \\mathbb{R}^{N\\times V\\times 3}\\) denotes the corresponding ground-truth sequence of facial movement over a neutral template, where \\(N\\) indicates the number of visual frames and \\(V\\) denotes the number" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.674, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KMTalk" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.145, + 0.782, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.271, + 0.646, + 0.282 + ], + "angle": 0, + "content": "(a) The overview pipeline of our proposed KMTalk." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.283, + 0.784, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.384, + 0.638, + 0.395 + ], + "angle": 0, + "content": "(b) The details of two key modules in KMTalk." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.407, + 0.788, + 0.56 + ], + "angle": 0, + "content": "Fig. 2: Fig. 2a illustrates the overview pipeline of our proposed KMTalk. Initially, the Audio Encoder takes the input raw audio \\(\\mathbf{x}\\) and encodes it into audio features \\(\\mathbf{A}\\). Subsequently, in the LKMA module, key motions \\(\\mathbf{K}\\) are generated from the audio \\(\\mathbf{x}\\) and \\(\\mathbf{A}\\). Finally, the CMC module reintroduces audio features \\(\\mathbf{A}\\) to extend these key motions \\(\\mathbf{K}\\) into a full sequence \\(\\mathbf{Y}\\). Fig. 2b presents the details of two key modules in KMTalk. In the Linguistic-based Key Motion Acquisition, a Phoneme-based Localization Method is used to identify key motion indices \\(\\mathcal{I}\\) from raw audio \\(\\mathbf{x}\\). Based on audio features \\(\\mathbf{A}\\) and \\(\\mathcal{I}\\), the Key Motion-focused Decoder generates key motions \\(\\mathbf{K}\\). In the Cross-modal Motion Completion, the Motion Flow Encoder processes \\(\\mathbf{K}\\) and \\(\\mathcal{I}\\), producing motion flow features \\(\\Phi\\). Then, with the dynamic fusion weight \\(\\mathbf{G}\\), the Multimodal-Guided Decoder combines \\(\\Phi\\) and \\(\\mathbf{A}\\) to decode the final motion sequence \\(\\mathbf{Y}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.787, + 0.632 + ], + "angle": 0, + "content": "of vertices in the facial mesh. The objective is to synthesize \\(\\mathbf{Y} = (\\mathbf{y}_1, \\dots, \\mathbf{y}_N)\\) that is similar to \\(\\hat{\\mathbf{Y}}\\), driven by the raw audio \\(\\mathbf{x}\\). The generated sequence should ensure lip synchronization with the audio while exhibiting natural facial movements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.632, + 0.788, + 0.768 + ], + "angle": 0, + "content": "Due to the domain gap between modalities and the ill-posed nature of directly translating audio to facial movement sequences, it is a challenging task that often results in over-smooth or poorly synchronized lip movements. To address these issues, this paper introduces a coarse-to-fine approach with key motion embedding, integrating both linguistic and data-driven priors. The overview pipeline is presented in Fig. 2a. In the LKMA module, linguistic priors are introduced to locate and generate higher-quality key motions (see Sec. 3.2), followed by the CMC module, where these key motions are fleshed out into a complete sequence of facial motions (see Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.788, + 0.6, + 0.803 + ], + "angle": 0, + "content": "3.2 Linguistic-based Key Motion Acquisition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In the realm of audio-driven 3D facial animation, it presents significant challenges to precisely define which frames constitute key motions. An alternative" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Z. Xu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "and simple solution is to use uniform or random sampling to determine the positions of these key motions. Although these approaches can boost performance to a certain degree due to the reduced learning complexity (refer to Sec. 4.2 for ablation studies), they fail to utilize the correlation between audio content and facial movements. However, we can leverage linguistic priors to capture pronounced articulatory actions, which are identifiable at phoneme boundaries. This approach circumvents the issue of over-smooth in the output sequence, thereby enhancing the overall quality of the results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.79, + 0.436 + ], + "angle": 0, + "content": "As shown in the left of Fig. 2b, the Linguistic-based Key Motion Acquisition (LKMA) module receives as inputs the raw audio \\(\\mathbf{x}\\) and the audio features \\(\\mathbf{A} = (\\mathbf{a}_1,\\dots,\\mathbf{a}_N)\\in \\mathbb{R}^{N\\times d}\\), where the audio features are derived from the Audio Encoder that utilizes the wav2vec 2.0 pre-trained model [2], processing the raw audio \\(\\mathbf{x}\\) as its input. Then it takes raw audio \\(\\mathbf{x}\\) as input to produce the key motion indices \\(\\mathcal{I} = \\{i_1,\\dots,i_m\\}\\), where \\(i_j\\in \\{1,\\dots,N\\},\\forall i_j\\in \\mathcal{I}\\), through the proposed Phoneme-based Localization method. Subsequently, the Key Motion-focused Decoder utilizes the audio features \\(\\mathbf{A}\\) and the key motion indices \\(\\mathcal{I}\\) to generate key motions \\(\\mathbf{K} = (\\mathbf{k}_{i_1},\\dots,\\mathbf{k}_{i_m})\\in \\mathbb{R}^{m\\times V\\times 3}\\) consisting of \\(m\\) frames of facial movement which are located on the key motion positions, where \\(\\mathbf{k}_{i_j}\\simeq \\hat{\\mathbf{y}}_{i_j},\\forall i_j\\in \\mathcal{I}\\). The process of the LKMA module is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.449, + 0.785, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I}, \\mathbf {K} = \\operatorname {L K M A} (\\mathbf {x}, \\mathbf {A}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.788, + 0.611 + ], + "angle": 0, + "content": "Phoneme-based Localization. At phoneme boundaries, a noticeable offset is observed in the articulator movement, with visualization results available in supplementary materials. Furthermore, the phoneme boundary effects underscore the ease with which the boundaries of phonemes can be perceived [17]. Both experimental and theoretical analyses have demonstrated a distinct position-mapping relationship between the phoneme boundaries in the audio and the significant elements in the motion sequence, specifically the key motions. The mapping relationship can be harnessed to facilitate the initial alignment between the audio and visual modalities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Specifically, an Automatic Speech Recognition model [28] is first utilized to obtain the text content from the raw audio \\(\\mathbf{x}\\). Then, a Montreal Forced Aligner [31] is adopted to align the audio and the text, producing the start and the end timestamps for each phoneme. Finally, the indices of these motion frames corresponding to the timestamps are regarded as key motion indices \\(\\mathcal{I}\\), and the corresponding motions compose the key motions \\(\\mathbf{K}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Key Motion-focused Decoder. It is utilized to synthesize key motions \\(\\mathbf{K}\\) of superior quality. Initially, employing \\(\\mathcal{I}\\) as indices, we extract the corresponding aligned audio features \\(\\mathbf{A_k} = (\\mathbf{a}_{i_1},\\dots,\\mathbf{a}_{i_m})\\in \\mathbb{R}^{m\\times d}\\) from the comprehensive audio features \\(\\mathbf{A}\\). Subsequently, it adopts a modified transformer-based architecture, processes \\(\\mathbf{A_k}\\) to generate the key motions \\(\\mathbf{K}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Loss Function. Intuitively, a straightforward approach to optimize the key motions \\(\\mathbf{K}\\) involves utilizing \\(\\mathcal{I}\\) to index \\(\\hat{\\mathbf{Y}}\\), resulting in \\(\\hat{\\mathbf{Y}}_{\\mathbf{k}} = (\\hat{\\mathbf{y}}_{i_1},\\dots ,\\hat{\\mathbf{y}}_{i_m})\\), which serves as the supervision for the training process. However, key motions typically occupy non-adjacent positions within the entire sequence. Hence, given the lack" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KMTalk" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "of inter-frame contextual information, attempting direct frame-by-frame regression of \\(\\mathbf{K}\\) towards \\(\\hat{\\mathbf{Y}}_{\\mathbf{k}}\\) may fall short in achieving accurate facial expressions, as well as producing smooth and realistic animation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.331 + ], + "angle": 0, + "content": "To address this limitation, we adopt a pseudo-complete sequence training method that utilizes the ground-truth frame labels \\(\\hat{\\mathbf{Y}}\\) at the non-key indices \\(\\mathcal{I}' = \\{1,\\dots,N\\} \\setminus \\mathcal{I}\\) and the generated key motions \\(\\mathbf{K}\\) at the key indices \\(\\mathcal{I}\\) to form a predicted pseudo-complete sequence \\(\\mathbf{Y}_{\\mathbf{p}} \\in \\mathbb{R}^{N \\times V \\times 3}\\). Then, the model is trained by minimizing the loss between the pseudo-complete sequence \\(\\mathbf{Y}_{\\mathbf{p}}\\) and the ground-truth sequence \\(\\hat{\\mathbf{Y}}\\). This enables the model to capture subtle changes between key motions and adjacent ground-truth frames, thereby mitigating inter-frame jitter and achieving more accurate regression of facial expressions. Following the SelfTalk [36], the loss function is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.344, + 0.785, + 0.359 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {L K M A} = \\lambda_ {1} \\mathcal {L} _ {r e c} + \\lambda_ {2} \\mathcal {L} _ {v e l} + \\lambda_ {3} \\mathcal {L} _ {l a t} + \\lambda_ {4} \\mathcal {L} _ {c t c}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.785, + 0.522 + ], + "angle": 0, + "content": "where \\(\\lambda_{1} = 1000.0, \\lambda_{2} = 1000.0, \\lambda_{3} = 0.001\\), and \\(\\lambda_{4} = 0.0001\\) in all of our experiments. \\(\\mathcal{L}_{rec}\\), \\(\\mathcal{L}_{vel}\\), and \\(\\mathcal{L}_{lat}\\) are measured by mean square error, while \\(\\mathcal{L}_{ctc}\\) is quantified by CTC Loss. The reconstruction loss \\(\\mathcal{L}_{rec}\\) quantifies the discrepancy between the predicted and the ground-truth facial movements. The velocity loss \\(\\mathcal{L}_{vel}\\) reduces frame jitter, ensuring smooth and natural lip movements. The latent consistency loss \\(\\mathcal{L}_{lat}\\) assesses the variance between latent features extracted from both the audio and lip shape encoders, aiming to align the learned audio and lip features. Lastly, the text consistency loss \\(\\mathcal{L}_{ctc}\\) evaluates the difference between the lip-reading decoder's output and the original text, ensuring the intelligibility of the lip-reading results." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.546, + 0.533, + 0.561 + ], + "angle": 0, + "content": "3.3 Cross-modal Motion Completion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.784, + 0.709 + ], + "angle": 0, + "content": "A straightforward method to obtain a complete talking face sequence is to directly integrate key motions into the entire face movements. However, it is important to note that while key motions capture essential facial dynamics, they may not encompass all details of non-key motions. This direct integration could result in mismatches between augmented non-key motions and their corresponding audio segments (see Sec. 4.2 for ablation studies). To mitigate this issue, we introduce a Cross-modal Motion Completion (CMC) module that jointly combines the audio features \\(\\mathbf{A}\\), key motions \\(\\mathbf{K}\\), and key motion indices \\(\\mathcal{I}\\) to generate a complete sequence of 3D facial meshes \\(\\mathbf{Y}\\). The process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.429, + 0.722, + 0.785, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} = \\operatorname {C M C} (\\mathbf {A}, \\mathbf {K}, \\mathcal {I}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.721, + 0.764 + ], + "angle": 0, + "content": "The details of the CMC module are illustrated on the right of Fig. 2b." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Motion Flow Encoder. Key motions serve as a kinematic prior for the remaining frames, offering valuable insights into facial dynamics. To effectively capture the motion flow information provided by key motions, we draw inspiration from some manifold methods [32] to acquire motion flow features \\(\\Phi \\in \\mathbb{R}^{N\\times d}\\) from the key motions \\(\\mathbf{K}\\). Specifically, we first encode the key motions \\(\\mathbf{K}\\) into key motion" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Z. Xu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.282 + ], + "angle": 0, + "content": "context tokens \\(\\Phi_{\\mathbf{k}} \\in \\mathbb{R}^{m \\times d}\\) by multiple transformer-encoder layers. At the same time, the indices of non-key motions \\(\\mathcal{T}'\\) are encoded into positional encodings by a sinusoidal position embedding layer, representing the non-key frame positions. Then, we adopt the cross-attention layers to extract the intermediate tokens \\(\\Phi_{\\mathrm{non - key}} \\in \\mathbb{R}^{(N - m) \\times d}\\), with key and value from linear transformations of \\(\\Phi_{\\mathbf{k}}\\) and the query is the positional encodings of non-key frames indices. Above all, the implicit motion manifold proposed in CITL [32] is utilized to arrange \\(\\Phi_{\\mathbf{k}}\\) and \\(\\Phi_{\\mathrm{non - key}}\\) based on their indices, followed by a 1D convolution for fusion, ultimately obtaining the motion flow features \\(\\Phi\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Multimodal-Guided Decoder. The non-key motions' features in the complete motion sequence feature estimation are derived from the global context interpolation of key motions, which has a certain degree of information loss due to the audio feature selection process in the Key Motion-focused Decoder (Sec. 3.2). Hence, we have devised a multimodal-guided decoding approach that incorporates audio modalities to furnish comprehensive information across the entire temporal scale, alongside motion flow to offer facial motion priors. These elements serve to guide and constrain the decoding process, thereby facilitating the precise generation of the motion sequence. Technically, we simply employ the gated mechanism [9,57,58] for the modality fusion, which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.438, + 0.45, + 0.786, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} = \\sigma ([ \\mathbf {A}, \\boldsymbol {\\Phi} ] \\mathbf {W}), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.488, + 0.785, + 0.504 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} = \\mathbf {G} \\odot \\mathbf {A} + (1 - \\mathbf {G}) \\odot \\boldsymbol {\\Phi}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.601 + ], + "angle": 0, + "content": "where \\(\\odot\\) represents the element-wise multiplication operation, \\([\\cdot, \\cdot]\\) denotes the concatenation operation, \\(\\mathbf{W} \\in \\mathbb{R}^{2d \\times d}\\) is a parameter and \\(\\sigma\\) is a sigmoid function, while \\(\\mathbf{G} \\in \\mathbb{R}^{N \\times d}\\) dynamically selects features from the audio features \\(\\mathbf{A}\\) and the motion flow features \\(\\Phi\\). Then, the Motion Decoder, a transformer-based structural model, is employed to transform the fusion features \\(\\mathbf{Z}\\) into the complete 3D facial movement sequence \\(\\hat{\\mathbf{Y}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.788, + 0.632 + ], + "angle": 0, + "content": "Loss Function. We train the CMC module utilizing the reconstruction loss and velocity loss. The total loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.426, + 0.647, + 0.785, + 0.662 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C M C} = \\mathcal {L} _ {r e c} + \\mathcal {L} _ {v e l}. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.687, + 0.367, + 0.705 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Dataset. BIWI [13] consists of 40 paired audio-visual sentences from 14 subjects. The 3D facial geometries, consisting of 23370 vertices, were captured at a frame rate of 25fps, and the average duration of each sequence was 4.67 seconds. We adopt the same evaluation protocol as FaceFormer [12] on the BIWI dataset. Specifically, the training set (BIWI-Train) comprises 190 sentences, while the validation set (BIWI-Val) encompasses 24 sentences. The dataset is divided into two testing sets: BIWI-Test-A, which comprises 24 sentences articulated by six subjects observed during training, and BIWI-Test-B, which consists of 32 sentences" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.674, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KMTalk" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.186 + ], + "angle": 0, + "content": "Table 1: Quantitative comparisons on the BIWI-Test-A and VOCA-Test datasets. The results of Lip-Vertex Error (LVE) and the upper-Face Dynamics Deviation (FDD) are reported. For both metrics, the lower the better." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.198, + 0.787, + 0.312 + ], + "angle": 0, + "content": "
MethodsBIWI-Test-AVOCA-Test
LVE↓×10-4mmFDD↓×10-5mmLVE↓×10-5mmFDD↓×10-7mm
VOCA [8]6.55638.18164.92454.8447
MeshTalk [41]5.91815.10254.54415.2062
FaceFormer [12]5.30774.64084.10904.6675
CodeTalker [53]4.79144.11703.94454.5422
SelfTalk [36]4.24853.57613.22384.0912
KMTalk (Ours)3.96542.54462.26394.0594
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.785, + 0.431 + ], + "angle": 0, + "content": "uttered by eight unseen subjects. VOCASET [8] consists of 480 paired audiovisual sequences from 12 subjects. Each sequence is recorded at a frame rate of 60fps and ranges in duration from 3 to 4 seconds. The 3D face mesh for each sequence consists of 5023 vertices. To ensure a fair comparison, we used identical training (VOCA-Train), validation (VOCA-Val), and testing (VOCA-Test) partitions as methods [12,36,53]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.787, + 0.552 + ], + "angle": 0, + "content": "Baselines. We compare against current state-of-the-arts method, including VOCA [8], MeshTalk [41]. FaceFormer [12], CodeTalker [53], and SelfTalk [36]. Faceformer [12] employs a transformer-based model to incorporate long-term audio context and synthesizes sequential motions in an autoregressive manner. CodeTalker [53] introduces discrete motion priors to enable self-reconstruction of real facial movements, mitigating the issue of excessive smoothing in facial motion. SelfTalk [36] designs a learning-based recognizer to minimize the domain gap between diverse modalities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.787, + 0.734 + ], + "angle": 0, + "content": "Evaluation Metrics. Following CodeTalker [53] and SelfTalk [36], we adopt two metrics for the quantitative evaluation of speech-driven facial animation: lip vertex error (LVE) to measure lip synchronization and upper-face dynamics deviation (FDD) to assess the overall facial dynamics. The LVE for each frame is defined as the maximal L2 error among all lip vertices for each frame and takes the average over all frames. This L2 error is computed by comparing the predictions with the processed 3D face geometry data. FDD is introduced to quantify the variation in facial dynamics between a synthetic motion sequence and the reference sequence. The implementation of FDD is to calculate the difference between the variances of vertex offsets in the upper-face region and the variances of ground truth vertex offsets. In addition, we visualize the prediction results for qualitative evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Implementation Details. For a fair comparison, KMTalk operates at a frame rate of 30 fps on VOCASET and 25 fps on BIWI, following the setting of previous methods [12, 36, 53]. Also, it can naturally adapt to a higher frame rate, as shown in the supplementary materials. In the LKMA module, we first employ the Phoneme-based Localization method to process the raw audio and obtain key motion indices for data preprocessing, which costs less than 10 minutes on two datasets [8, 13]. Secondly, we train the Key Motion Decoder on a single NVIDIA" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "Z. Xu et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.142, + 0.775, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.464, + 0.788, + 0.59 + ], + "angle": 0, + "content": "Fig. 3: Qualitative comparisons on VOCA-Test (left) and BIWI-Test-B (right). We provide visual comparisons of facial animations synchronized with six syllables extracted from the test speech sequences. The 1st, 3rd, and 5th rows display synthesized meshes and their corresponding ground-truths, while the 2nd, 4th, and 6th rows visualize the L2 loss for individual frames. Our method demonstrates more precise mouth movement on syllables like /æ/ that require a wide-open mouth. For syllables that start with a closed mouth and then slightly open, such as /bɪ/, our KMTalk generates more synchronized motion sequences visually. The last row visualizes the mean square errors of different methods across all sentences in the test set for a specific subject." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.788, + 0.712 + ], + "angle": 0, + "content": "RTX 3090 for 200 epochs (about 2 hours) using the Adam optimizer [21]. The learning rate is initialized as \\(10^{-4}\\), and the mini-batch size is set to 1. In the CMC module, we train for 200 epochs (approximately 2 hours) with the same training settings as the Key Motion Decoder. It is noteworthy that, since the training of the two modules is independent of each other, we can train both modules concurrently to enhance training efficiency." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.737, + 0.646, + 0.753 + ], + "angle": 0, + "content": "4.1 Comparisons against State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Quantitative Comparisons. We computed the lip vertex error (LVE) and facial dynamics deviation (FDD) for all sequences within the BIWI-Test-A and VOCA-Test datasets. According to Table 1, our proposed KMTalk demonstrated lower errors compared to the alternative methods examined. Notably, the lip vertex error for our method on the VOCA-Test dataset is \\(30\\%\\) lower than the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KMTalk" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.207 + ], + "angle": 0, + "content": "recently introduced SelfTalk [36], and the FDD is \\(27\\%\\) lower than SelfTalk on the BIWI-Test-A dataset, providing compelling evidence for the advantages of our proposed KMTalk. This indicates that our approach is more effective in achieving audio-visual alignment, thereby leading to improved lip synchronization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.403 + ], + "angle": 0, + "content": "Qualitative Comparisons. In Fig. 3, we visualize the output facial meshes from different methods as well as ground truths for reference. Additionally, we display error maps calculated from the vertex L2 loss between the generated and ground truth meshes. It is evident that our method consistently yields lower errors across different speech sequences, demonstrating its ability to generate more accurate facial animation sequences. Notably, for representative syllables (e.g. \\(/a\\mathrm{e}/\\)), KMTalk closely approximates ground truth, excelling in synthesizing accurate lip movements for syllables requiring significant mouth opening. Additionally, for syllables starting with mouth closure followed by a slight opening (e.g., \\(/b\\mathrm{i}/\\)), KMTalk produces more natural and synchronized motions. We recommend that readers watch the supplementary video for more detailed comparisons. It showcases KMTalk's capability to generate coherent, realistic animations with precise lip synchronization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.787, + 0.676 + ], + "angle": 0, + "content": "User Studies. A user study stands as a dependable evaluation method in the context of 3D talking faces. Following the strategy of Faceformer [12], we conduct pairwise comparisons between our method and baselines [12,36,53], as well as ground truths. This study encompassed the assessment of two key metrics: perceptual lip synchronization and facial realism. Participants were presented with side-by-side comparisons and were tasked with selecting the better facial animation based on their personal preferences. We computed the ratio of user preferences as a measurement of satisfaction evaluation on BIWI-Test-B and VOCA-Test. We randomly sampled 30 examples from each test set and compared the performance of KMTalk with four aforementioned settings on each sample. Therefore, we constructed a total of 240 different video pairs and randomly selected 24 video pairs for the two metrics assessments for each participant. Our user study involved 30 participants with a strong capability for audio-visual perception, resulting in 720 effective evaluation entries. As demonstrated in Table 2, our approach indicates superior perceptual lip synchronization and facial realism. For instance, a noteworthy \\(60.0\\%\\) of users favored our lip synchronization method on BIWI-Test-B in comparison to SelfTalk [36]. Overall, it shows that KMTalk can generate more favorable facial animations from speeches." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.713, + 0.403, + 0.727 + ], + "angle": 0, + "content": "4.2 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.785, + 0.84 + ], + "angle": 0, + "content": "In this section, we perform ablation studies on to evaluate the influence of different components within our proposed KMTalk framework on the quality of the generated 3D talking faces. The quantitative results on BIWI are in Table 3, and the qualitative results are in Fig. 4. In addition, the results of the ablation study on VOCA-Test can be found in the supplementary material. In Table 4, we further investigate the robustness of our approach to different Phoneme-based localization methods and the possible errors during phoneme extraction." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "Z. Xu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.292, + 0.145, + 0.71, + 0.159 + ], + "angle": 0, + "content": "Table 2: User study results on BIWI-Test-B and VOCA-Test." + }, + { + "type": "table", + "bbox": [ + 0.262, + 0.171, + 0.754, + 0.34 + ], + "angle": 0, + "content": "
MethodMetricBIWI-Test-BVOCA-Test
competitorourscompetitorours
Ours vs. FaceFormerLip Sync24.4%75.6%26.7%73.3%
Realism25.6%74.4%28.9%71.1%
Ours vs. CodeTalkerLip Sync31.1%68.9%37.8%62.2%
Realism27.8%72.2%35.6%64.4%
Ours vs. SelfTalkLip Sync40.0%60.0%43.3%56.7%
Realism38.9%61.1%41.1%58.9%
Ours vs. GTLip Sync54.4%45.6%56.7%43.3%
Realism52.2%47.8%56.7%43.3%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.292, + 0.355, + 0.71, + 0.37 + ], + "angle": 0, + "content": "Table 3: Ablation study for our components on BIWI-Test-A." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.381, + 0.731, + 0.476 + ], + "angle": 0, + "content": "
Phoneme-based Localization MethodKey Motion-focused Audio in CMCLVE↓FDD↓
Decoder✓4.24853.5761
---4.16482.8713
-✓✓4.13812.9546
✓-✓4.88593.2780
✓✓-3.96542.5446
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.787, + 0.687 + ], + "angle": 0, + "content": "What's the effect of the Phoneme-based localization method for key motion capture? The Phoneme-based Localization Method enables us to identify key frames of speech with notable facial expression transitions. We can replace it with uniform sampling, neglecting the crucial content information of the audio. Specifically, we experimented using uniform sampling at a rate of \\(33\\%\\) to closely align with the number of key motions. For further comparisons under different numbers of sampled elements, please refer to the supplementary material. In Table 3, we observe degradation in all metrics with uniform sampling. This underscores the importance of accurate key motion capture in speech-driven talking face generation. Additionally, it showcases that the linguistic-based key motion capture is better equipped to mitigate audio-visual uncertainty and recover more precise facial motions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "What's the effect of the Key Motion Decoder? Our method employs a specialized key motion decoder to generate facial meshes based on keyframe indices obtained from phoneme-based localization methods. An alternative approach to generating key motions is to select corresponding facial meshes from a complete motion sequence produced by an existing method such as SelfTalk [36]. Table 3 demonstrates that both metrics deteriorated but still outperformed the state-of-the-art method SelfTalk. This indicates that the Key Motion Decoder can enhance the quality of key motion generation, resulting in more plausible facial animations. It also suggests that even if the captured key motions are not accurate enough, the CMC module can further refine output full motions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.674, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KMTalk" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.149, + 0.784, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.476 + ], + "angle": 0, + "content": "Fig. 4: Qualitative ablation studies on the input speech \"specifically\". For each method variant, we removed one of three modules: PLM (Phoneme-based Localization Method), KMD (Key Motion-focused Decoder), and AG (Audio Guidance in CMC). Error maps between generated and the ground-truth mesh sequence were visualized. Our final model yielded the best results, showcasing the effectiveness of each module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.788, + 0.684 + ], + "angle": 0, + "content": "What's the effect of audio guidance in the Cross-modal Motion Completion? The Cross-modal Motion Completion leverages audio features to guide the completion of the full motion sequence. To assess its usefulness, we implement a method variant that removes the audio feature guidance. Table 3 demonstrates a notable degradation in both metrics, particularly with a \\(24\\%\\) increase in Lip Vertex Error (LVE) and a \\(27\\%\\) increase in upper-face dynamics deviation (FDD). This suggests that audio information plays a crucial role in refining fine-grained lip movements and enhancing audio-visual consistency and temporal smoothness. Despite the degradation, the FDD metric can still outperform state-of-the-art methods, underscoring the significance of key motion capture for achieving temporally coherent full motion synthesis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Is our method sensitive to different Phomene-based Localization methods? We experimented with different Automatic Speech Recognition (ASR) models, such as Auto-avs r [28] and Whisper [39]. The results in the first three rows of Table 4 show that our KMTalk method consistently maintained high performance, achieving at least a \\(21\\%\\) improvement in FDD regardless of the ASR model used [28, 39]. This highlights the robustness of our approach across various ASR models. Besides, to simulate phoneme localization deviations, we shifted all key motion indices extracted by Auto-avs r [28] one frame to the right. The results in the last row of Table 4 indicate negligible variations. This demonstrates the robustness of our method to inaccurate key frame localization." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Z. Xu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.227, + 0.158, + 0.461, + 0.199 + ], + "angle": 0, + "content": "Table 4: Robust analysis of Phoneme-based Localization on BIWI-Test-A." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.211, + 0.486, + 0.342 + ], + "angle": 0, + "content": "
MethodsLVE↓ ×10-4mmFDD↓ ×10-5mm
Auto-avsr [28]3.96542.5446
Whisper-large [39]4.07182.8141
Whisper-tiny [39]4.06432.8083
Auto-avsr+offset3.99912.6420
" + }, + { + "type": "table_caption", + "bbox": [ + 0.487, + 0.158, + 0.776, + 0.199 + ], + "angle": 0, + "content": "Table 5: The results of integrating our proposed KMTalk with existing methods on BIWI-Test-A." + }, + { + "type": "table", + "bbox": [ + 0.499, + 0.211, + 0.764, + 0.342 + ], + "angle": 0, + "content": "
MethodsLVE↓ ×10-4mmFDD↓ ×10-5mm
FaceFormer [12]Original5.30774.6408
After5.27934.2654
CodeTalker [53]Original4.79144.1170
After4.50963.9043
SelfTalk [36]Original4.24853.5761
After4.11222.8668
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.36, + 0.55, + 0.377 + ], + "angle": 0, + "content": "4.3 Integration with Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.534 + ], + "angle": 0, + "content": "Existing approaches focus on enhancing prediction outcomes by designing elaborate priors or the learning-based recognizer, which may be highly coupled with the proposed architecture of these methods. Our KMTalk introduces a new learning strategy of speech-driven talking face generation, which is orthogonal to these approaches. Therefore, we can explore whether performance can be enhanced by applying our progressive learning scheme without the need for additional fine-tuning of their models. Detailed implementation is in the supplementary material. As shown in Table 5, the results of existing methods are improved after integration with our proposed progressive learning mechanism utilizing key motion embeddings. This further emphasizes the efficacy of our design." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.555, + 0.36, + 0.57 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In this work, we introduce KMTalk, a novel method for progressively learning 3D facial animation from speeches using key motion embeddings. It incorporates linguistic priors for key motion generation and extends them to a full motion sequence via data-driven priors. We propose phoneme-based localization methods to determine the temporal position of key facial motions, improving lip-speech synchronization by aligning motion transitions with phoneme changes. Additionally, we design a cross-modal facial motion completion module that synthesizes the entire motion sequence from key motions and audio features, enhancing lip-speech synchronization and motion coherence. Extensive evaluations of the datasets demonstrate KMTalk's superiority over existing methods, producing more accurate and realistic animations. Moreover, coupling our idea with existing methods consistently improves performance, further verifying the efficacy of our proposed progressive learning mechanism based on key motion acquisition. Although the proposed method has demonstrated its robustness to inaccurate keyframe localization, it may encounter errors in dialect variations. Integrating advanced ASR(Automatic Speech Recognition) technology in the future could enhance its adaptability to various speech patterns." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.674, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KMTalk" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Acknowledgement The research is partially supported by National Key Research and Development Program of China (2023YFC3502900), National Natural Science Foundation of China (No.62176093, 61673182), Key Realm Research and Development Program of Guangzhou (No.202206030001), Guangdong-Hong Kong-Macao Joint Innovation Project (No.2023A0505030016), Guangdong Natural Science Foundation (No. 2024A1515012217)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.26, + 0.323, + 0.276 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.318 + ], + "angle": 0, + "content": "1. Alghamdi, M.M., Wang, H., Bulpitt, A.J., Hogg, D.C.: Talking head from speech audio using a pre-trained image generator. In: ACM MM. pp. 5228-5236 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.319, + 0.785, + 0.359 + ], + "angle": 0, + "content": "2. Baevski, A., Zhou, Y., Mohamed, A., Auli, M.: wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems 33, 12449-12460 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.36, + 0.785, + 0.4 + ], + "angle": 0, + "content": "3. Bao, L., Zhang, H., Qian, Y., Xue, T., Chen, C., Zhe, X., Kang, D.: Learning audio-driven viseme dynamics for 3d face animation. arXiv preprint arXiv:2301.06059 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.401, + 0.785, + 0.428 + ], + "angle": 0, + "content": "4. Cao, Y., Tien, W.C., Faloutsos, P., Pighin, F.: Expressive speech-driven facial animation. ACM Trans. Graph 24(4), 1283-1302 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.429, + 0.785, + 0.455 + ], + "angle": 0, + "content": "5. Chen, L., Cui, G., Liu, C., Li, Z., Kou, Z., Xu, Y., Xu, C.: Talking-head generation with rhythmic head motion. In: ECCV. pp. 35-51. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.457, + 0.785, + 0.483 + ], + "angle": 0, + "content": "6. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: ACCV. pp. 251-263. Springer (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.484, + 0.785, + 0.51 + ], + "angle": 0, + "content": "7. Cohen, M.M., Clark, R., Massaro, D.W.: Animated speech: Research progress and applications. In: AVSP (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.785, + 0.538 + ], + "angle": 0, + "content": "8. Cudeiro, D., Bolkart, T., Laidlaw, C., Ranjan, A., Black, M.J.: Capture, learning, and synthesis of 3d speaking styles. In: CVPR. pp. 10101-10111 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.539, + 0.785, + 0.592 + ], + "angle": 0, + "content": "9. Dai, G., Zhang, Y., Wang, Q., Du, Q., Yu, Z., Liu, Z., Huang, S.: Disentangling writer and character styles for handwriting generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5977-5986 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.594, + 0.785, + 0.634 + ], + "angle": 0, + "content": "10. Das, D., Biswas, S., Sinha, S., Bhowmick, B.: Speech-driven facial animation using cascaded gans for learning of motion and texture. In: ECCV. pp. 408-424. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.635, + 0.785, + 0.662 + ], + "angle": 0, + "content": "11. Edwards, P., Landreth, C., Fiume, E., Singh, K.: Jali: an animator-centric viseme model for expressive lip synchronization. ACM Trans. Graph 35(4), 1-11 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.663, + 0.785, + 0.689 + ], + "angle": 0, + "content": "12. Fan, Y., Lin, Z., Saito, J., Wang, W., Komura, T.: Faceformer: Speech-driven 3d facial animation with transformers. In: CVPR. pp. 18770-18780 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.689, + 0.785, + 0.731 + ], + "angle": 0, + "content": "13. Fanelli, G., Gall, J., Romsdorfer, H., Weise, T., Van Gool, L.: A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia 12(6), 591-598 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.731, + 0.785, + 0.758 + ], + "angle": 0, + "content": "14. Fisher, C.G.: Confusions among visually perceived consonants. Journal of speech and hearing research 11(4), 796-804 (1968)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.759, + 0.785, + 0.785 + ], + "angle": 0, + "content": "15. Guo, Y., Chen, K., Liang, S., Liu, Y.J., Bao, H., Zhang, J.: Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In: ICCV. pp. 5784-5794 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "16. Habibie, I., Xu, W., Mehta, D., Liu, L., Seidel, H.P., Pons-Moll, G., Elgharib, M., Theobalt, C.: Learning speech-driven 3d conversational gestures from video. In: Proceedings of the 21st ACM International Conference on Intelligent Virtual Agents. pp. 101-108 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.219, + 0.289, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "Z. Xu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "17. Iverson, P., Kuhl, P.K.: Perceptual magnet and phoneme boundary effects in speech perception: Do they arise from a common mechanism? Perception & psychophysics 62, 874-886 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "18. Ji, X., Zhou, H., Wang, K., Wu, Q., Wu, W., Xu, F., Cao, X.: Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In: ACM SIGGRAPH. pp. 1-10 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "19. Ji, X., Zhou, H., Wang, K., Wu, W., Loy, C.C., Cao, X., Xu, F.: Audio-driven emotional video portraits. In: CVPR. pp. 14080-14089 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "20. Karras, T., Aila, T., Laine, S., Herva, A., Lehtinen, J.: Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Trans. Graph 36(4), 1-12 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.328 + ], + "angle": 0, + "content": "21. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.785, + 0.356 + ], + "angle": 0, + "content": "22. Lewis, J.: Automated lip-sync: Background and techniques. IEEE Trans Vis Comput Graph 2(4), 118-122 (1991)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.398 + ], + "angle": 0, + "content": "23. Liang, B., Pan, Y., Guo, Z., Zhou, H., Hong, Z., Han, X., Han, J., Liu, J., Ding, E., Wang, J.: Expressive talking head generation with granular audio-visual control. In: CVPR. pp. 3387-3396 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.399, + 0.785, + 0.424 + ], + "angle": 0, + "content": "24. Liu, C.: An analysis of the current and future state of 3d facial animation techniques and systems (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.467 + ], + "angle": 0, + "content": "25. Liu, X., Xu, Y., Wu, Q., Zhou, H., Wu, W., Zhou, B.: Semantic-aware implicit neural audio-driven video portrait generation. In: ECCV. pp. 106-125. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.785, + 0.494 + ], + "angle": 0, + "content": "26. Liu, Y., Xu, F., Chai, J., Tong, X., Wang, L., Huo, Q.: Video-audio driven real-time facial animation. ACM Trans. Graph 34(6), 1-10 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.496, + 0.785, + 0.536 + ], + "angle": 0, + "content": "27. Lu, L., Wu, R., Lin, H., Lu, J., Jia, J.: Video frame interpolation with transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3532-3542 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.577 + ], + "angle": 0, + "content": "28. Ma, P., Haliassos, A., Fernandez-Lopez, A., Chen, H., Petridis, S., Pantic, M.: Auto-avsr: Audio-visual speech recognition with automatic labels. arxiv 2023. arXiv preprint arXiv:2303.14307" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "29. Massaro, D., Cohen, M., Tabain, M., Beskow, J., Clark, R.: 12 animated speech: research progress and applications (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "30. Mattheyses, W., Verhelst, W.: Audiovisual speech synthesis: An overview of the state-of-the-art. Speech Communication 66, 182-217 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.785, + 0.674 + ], + "angle": 0, + "content": "31. McAuliffe, M., Sonderegger, M.: English mfa acoustic model v2.2.1. Tech. rep., https://mfa-models.readthedocs.io/acoustic/English/EnglishMFAacousticmodelv2_2_1.html (May 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "32. Mo, C.A., Hu, K., Long, C., Wang, Z.: Continuous intermediate token learning with implicit motion manifold for keyframe based motion interpolation. In: CVPR. pp. 13894-13903 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.758 + ], + "angle": 0, + "content": "33. Niklaus, S., Mai, L., Liu, F.: Video frame interpolation via adaptive convolution. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 670-679 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "34. Nocentini, F., Ferrari, C., Berretti, S.: Learning landmarks motion from speech for speaker-agnostic 3d talking heads generation. arXiv preprint arXiv:2306.01415 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "35. Pang, Y., Zhang, Y., Quan, W., Fan, Y., Cun, X., Shan, Y., Yan, D.m.: Dpe: Disentanglement of pose and expression for general video portrait editing. In: CVPR. pp. 427-436 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.674, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KMTalk" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "36. Peng, Z., Luo, Y., Shi, Y., Xu, H., Zhu, X., Liu, H., He, J., Fan, Z.: Selftalk: A self-supervised commutative training diagram to comprehend 3d talking faces. arXiv preprint arXiv:2306.10799 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "37. Pham, H.X., Wang, Y., Pavlovic, V.: End-to-end learning for 3d facial animation from speech. In: ICMI. pp. 361-365 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.219, + 0.786, + 0.26 + ], + "angle": 0, + "content": "38. Prajwal, K., Mukhopadhyay, R., Namboodiri, V.P., Jawahar, C.: A lip sync expert is all you need for speech to lip generation in the wild. In: ACM MM. pp. 484-492 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.787, + 0.302 + ], + "angle": 0, + "content": "39. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision. In: International Conference on Machine Learning. pp. 28492-28518. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.303, + 0.787, + 0.332 + ], + "angle": 0, + "content": "40. Richard, A., Lea, C., Ma, S., Gall, J., De la Torre, F., Sheikh, Y.: Audio-and gaze-driven facial animation of codec avatars. In: WACV. pp. 41-50 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.333, + 0.786, + 0.374 + ], + "angle": 0, + "content": "41. Richard, A., Zollhöfer, M., Wen, Y., De la Torre, F., Sheikh, Y.: Meshtalk: 3d face animation from speech using cross-modality disentanglement. In: ICCV. pp. 1173-1182 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.787, + 0.415 + ], + "angle": 0, + "content": "42. Shen, S., Li, W., Zhu, Z., Duan, Y., Zhou, J., Lu, J.: Learning dynamic facial radiance fields for few-shot talking head synthesis. In: ECCV. pp. 666-682. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.416, + 0.787, + 0.444 + ], + "angle": 0, + "content": "43. Tanaka, H., Nakamura, S., et al.: The acceptability of virtual characters as social skills trainers: usability study. JMIR human factors 9(1), e35358 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.445, + 0.787, + 0.5 + ], + "angle": 0, + "content": "44. Tang, J., Dai, A., Nie, Y., Markhasin, L., Thies, J., Nießner, M.: Dphms: Diffusion parametric head models for depth-based tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1111-1122 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.501, + 0.787, + 0.543 + ], + "angle": 0, + "content": "45. Taylor, S., Kim, T., Yue, Y., Mahler, M., Krahe, J., Rodriguez, A.G., Hodgins, J., Matthews, I.: A deep learning approach for generalized speech animation. ACM Trans. Graph 36(4), 1-11 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.544, + 0.787, + 0.571 + ], + "angle": 0, + "content": "46. Taylor, S.L., Mahler, M., Theobald, B.J., Matthews, I.: Dynamic units of visual speech. In: ACM SIGGRAPH. pp. 275-284 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.572, + 0.787, + 0.599 + ], + "angle": 0, + "content": "47. Vougioukas, K., Petridis, S., Pantic, M.: Realistic speech-driven facial animation with gans. International Journal of Computer Vision 128, 1398-1413 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.6, + 0.787, + 0.628 + ], + "angle": 0, + "content": "48. Wang, J., Qian, X., Zhang, M., Tan, R.T., Li, H.: Seeing what you said: Talking face generation guided by a lip reading expert. In: CVPR. pp. 14653-14662 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.629, + 0.787, + 0.656 + ], + "angle": 0, + "content": "49. Wang, Q., Fan, Z., Xia, S.: 3d-talkemo: Learning to synthesize 3d emotional talking head. arXiv preprint arXiv:2104.12051 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.657, + 0.787, + 0.685 + ], + "angle": 0, + "content": "50. Wang, S., Li, L., Ding, Y., Yu, X.: One-shot talking face generation from single-speaker audio-visual correlation learning. In: AAAI. vol. 36, pp. 2531-2539 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.686, + 0.787, + 0.727 + ], + "angle": 0, + "content": "51. Wen, S., Liu, W., Yang, Y., Huang, T., Zeng, Z.: Generating realistic videos from keyframes with concatenated gans. IEEE Transactions on Circuits and Systems for Video Technology 29(8), 2337-2348 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.728, + 0.787, + 0.769 + ], + "angle": 0, + "content": "52. Wu, H., Zhou, S., Jia, J., Xing, J., Wen, Q., Wen, X.: Speech-driven 3d face animation with composite and regional facial movements. arXiv preprint arXiv:2308.05428 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.77, + 0.787, + 0.812 + ], + "angle": 0, + "content": "53. Xing, J., Xia, M., Zhang, Y., Cun, X., Wang, J., Wong, T.T.: Codetalker: Speechdriven 3d facial animation with discrete motion prior. In: CVPR. pp. 12780-12790 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "54. Xu, Y., Feng, A.W., Marsella, S., Shapiro, A.: A practical and configurable lip sync method for games. In: Proceedings of Motion on Games, pp. 131-140 (2013)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "Z. Xu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "55. Yi, R., Ye, Z., Zhang, J., Bao, H., Liu, Y.J.: Audio-driven talking face video generation with learning-based personalized head pose. arXiv preprint arXiv:2002.10137 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "56. Yin, S., Wu, C., Yang, H., Wang, J., Wang, X., Ni, M., Yang, Z., Li, L., Liu, S., Yang, F., et al.: Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv preprint arXiv:2303.12346 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "57. Yu, D., Li, X., Zhang, C., Liu, T., Han, J., Liu, J., Ding, E.: Towards accurate scene text recognition with semantic reasoning networks. In: CVPR. pp. 12113-12122 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "58. Yue, X., Kuang, Z., Lin, C., Sun, H., Zhang, W.: Robustscanner: Dynamically enhancing positional clues for robust text recognition. In: ECCV. pp. 135-151. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.357 + ], + "angle": 0, + "content": "59. Zhang, B., Qi, C., Zhang, P., Zhang, B., Wu, H., Chen, D., Chen, Q., Wang, Y., Wen, F.: Metaportrait: Identity-preserving talking head generation with fast personalized adaptation. In: CVPR. pp. 22096-22105 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.384 + ], + "angle": 0, + "content": "60. Zhang, C., Ni, S., Fan, Z., Li, H., Zeng, M., Budagavi, M., Guo, X.: 3d talking face with personalized pose dynamics. IEEE Trans Vis Comput Graph. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.425 + ], + "angle": 0, + "content": "61. Zhou, H., Sun, Y., Wu, W., Loy, C.C., Wang, X., Liu, Z.: Pose-controllable talking face generation by implicitly modularized audio-visual representation. In: CVPR. pp. 4176-4186 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.425 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/KMTalk_ Speech-Driven 3D Facial Animation with Key Motion Embedding/8f222025-36f0-4f3b-ab78-77e262291c3e_origin.pdf b/2024/KMTalk_ Speech-Driven 3D Facial Animation with Key Motion Embedding/8f222025-36f0-4f3b-ab78-77e262291c3e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c28248b6b0e9e273fc98720036a78e3adef1a089 --- /dev/null +++ b/2024/KMTalk_ Speech-Driven 3D Facial Animation with Key Motion Embedding/8f222025-36f0-4f3b-ab78-77e262291c3e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c960052a38d214a3e6a1478b221fb5a61e44d04bf9c989c27fb5abd6efba4146 +size 4251984 diff --git a/2024/KMTalk_ Speech-Driven 3D Facial Animation with Key Motion Embedding/full.md b/2024/KMTalk_ Speech-Driven 3D Facial Animation with Key Motion Embedding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..58f5b04a7e22eba0ad09ce99c6ed61d2319bbc3d --- /dev/null +++ b/2024/KMTalk_ Speech-Driven 3D Facial Animation with Key Motion Embedding/full.md @@ -0,0 +1,268 @@ +# KMTalk: Speech-Driven 3D Facial Animation with Key Motion Embedding + +Zhihao $\mathrm{Xu}^{1^*}$ , Shengjie Gong $^{1^*}$ , Jiapeng Tang $^{2}$ , Lingyu Liang $^{1}$ , Yining Huang $^{1}$ , Haojie Li $^{1}$ , and Shuangping Huang $^{1,3\dagger}$ + +$^{1}$ South China University of Technology + +2 Technical University of Munich + +3 Pazhou Laboratory + +{eezhihaoxu,eeshengjiegong}@mail.scut.edu.cn, eehsp@scut.edu.cn + +Abstract. We present a novel approach for synthesizing 3D facial motions from audio sequences using key motion embeddings. Despite recent advancements in data-driven techniques, accurately mapping between audio signals and 3D facial meshes remains challenging. Direct regression of the entire sequence often leads to over-smoothed results due to the ill-posed nature of the problem. To this end, we propose a progressive learning mechanism that generates 3D facial animations by introducing key motion capture to decrease cross-modal mapping uncertainty and learning complexity. Concretely, our method integrates linguistic and data-driven priors through two modules: the linguistic-based key motion acquisition and the cross-modal motion completion. The former identifies key motions and learns the associated 3D facial expressions, ensuring accurate lip-speech synchronization. The latter extends key motions into a full sequence of 3D talking faces guided by audio features, improving temporal coherence and audio-visual consistency. Extensive experimental comparisons against existing state-of-the-art methods demonstrate the superiority of our approach in generating more vivid and consistent talking face animations. Consistent enhancements in results through the integration of our proposed learning scheme with existing methods underscore the efficacy of our approach. + +Keywords: Speech-driven $\cdot$ 3D Facial Animation $\cdot$ Key Motion + +# 1 Introduction + +Speech-driven 3D facial animation aims to create realistic talking heads that synchronize with input speech. It plays a significant role in many applications of virtual reality, like film production, computer gaming, and education [24, 43]. + +The main challenge of speech-driven 3D talking faces lies in the ill-posed problem caused by the cross-modal mapping uncertainty from the speech domain to + +![](images/79288203aa1c38b920ac9f2d7ce605aaac642a3c4776e940c5f5fb8a7be84a81.jpg) +Fig. 1: Compared to the state-of-the-art method Selftalk, our approach can produce more vivid lip motions from speeches, since we introduce linguistic priors to characterize key motions and utilize data-driven priors to interpolate non-key motions. + +the 3D motion domain. Since there may be multiple plausible outputs for input audio, effective regularizations and constraints should be integrated into the system, to generate vivid facial motions. The related methods can be roughly divided into linguistic-based methods [3, 11, 29, 46, 54] and data-driven methods [8, 12, 16, 26, 34, 36, 41, 49, 52, 53]. For linguistic-based methods [3, 11, 29, 46, 54], a set of intricate phoneme-to-viseme mapping rules is manually designed to generate the talking mouth based on priors from visemes or linguistic knowledge. While these methods explicitly control the animation of articulation processes, such as procedural lip sync with animation curves, their focus is mainly on localized facial movements, like those of the mouth area, lacking a systematic approach for modeling comprehensive facial motion. Thanks to the established audio-to-face datasets, learning-based methods [12, 16, 34, 36, 41, 49, 52, 53] choose to map audio signals into 3D facial meshes in a data-driven manner. Most of these works [8, 12, 16, 26, 34, 36, 41, 48, 49, 52, 53] typically formulated the cross-modal mapping of 3D talking face generation as a regression task, such as MeshTalk [41], FaceFormer [12], and SelfTalk [36]. While achieving impressive performance, they exhibit common limitations in their learning schemes. Firstly, they directly learn the ambiguous cross-modal mapping between audio and facial expression sequences, always leading to sub-optimal results in terms of temporal coherence and audio-visual consistency. Secondly, these methods typically regress the entire sequence without considering key motion cues, hindering the capture of detailed facial dynamics and accurate lip movements, particularly in complex facial expressions such as puckering or opening the mouth (as depicted in Fig. 1). Lastly, they overlook linguistic priors essential for simulating the articulation process, thereby limiting their ability to achieve precise lip-speech synchronization. + +To this end, inspired by the keyframe-based video generation techniques observed in recent studies [27, 33, 51, 56], which prioritize the generation of + +keyframes before adding detailed elements, we introduce a progressive learning mechanism that generates realistic 3D facial animations from audio inputs by incorporating key motion embeddings. The key idea is to initially generate key facial expressions, and then interpolate the intermediate motions to obtain the entire motion sequence, which significantly reduces the uncertainty of cross-modal mapping and eases the learning difficulty. Concretely, our method integrates linguistic and data-driven priors through two modules: the linguistic-based key motion acquisition and the cross-modal motion completion. The linguistic-based key motion acquisition module utilizes phoneme-based localization methods to identify temporal indices of key motion, which correspond to significant motion snapshots aligned with phoneme changes in the audio. Once the key motion indexes are determined, a key motion decoder interprets associated 3D facial meshes from corresponding audio features. This highlights those distinct facial expressions and facilitates lip-speech synchronization. The cross-modal motion completion module expands non-continuous key motions into a full sequence of continuous face motions using audio features as guidance. This process enhances audio-mesh alignment and improves the temporal smoothness of output facial mesh sequences. The contributions of our work are summarized as follows: + +- We propose a progressive learning mechanism to generate speech-driven 3D talking faces. It uses linguistic priors to initially generate key motions, and then interpolate key motions into complete motions via data-driven priors. +- We propose the use of phoneme-based localization methods to capture key facial motions. It effectively captures significant expression transitions aligned with phoneme changes in audio, improving lip-speech synchronization. +- We design a cross-modal facial motion completion module to produce full sequences of 3D talking faces using synthesized key motions and audio features. It further enhances lip-speech synchronization accuracy while facilitating temporal coherence in facial motions. + +Extensive experimental comparisons on the BIWI [13] and VOCASET [8] datasets demonstrate that our method outperforms existing state-of-the-art approaches in more accurate and realistic talking face generation. Detailed ablation studies confirm the effectiveness of our proposed key motion capture technique. Additionally, consistent improvements in results by combining our proposed scheme with existing methods validate the efficacy of our design. Our code and weights will be at the project website: https://github.com/ffxzh/KMTalk. + +# 2 Related Work + +While existing research [1,5,6,10,15,18,19,23,25,35,38,42,47,50,55,59,61] focuses on 2D talking heads, we focus on audio-driven 3D facial animations in this work, which can be roughly categorized into linguistics-based and data-driven methods. + +# 2.1 Linguistic-based Methods + +Linguistic-based methods [3, 7, 11, 14, 22, 29, 30, 46, 54] establish a set of intricate phoneme-to-viseme mapping rules for animating the mouth. For example, the + +dynamic viseme model proposed by Taylor et al. [46] exploits the one-to-many mapping of phonemes to lip motions. JALI [11] considers the many-to-many mapping between phonemes and visemes. More recently, Bao et al. [3] introduced a novel parameterized viseme fitting algorithm that extracts viseme parameters from speech videos using phonemic priors. Leveraging linguistic priors, these methods indicate the articulation process by providing animators with explicit control over animation, thus boosting their performance in lip-speech synchronization. However, these approaches primarily focus on animating the lip region, lacking a comprehensive strategy for animating the entire face. In our work, we leverage linguistic priors to detect key frames with significant expression changes from audio in an analytical manner, without the need for supervised training. + +# 2.2 Data-driven Methods + +With the development of deep learning technology and the availability of high-quality datasets, data-driven methods [4, 8, 12, 16, 20, 26, 34, 36, 37, 41, 44, 45, 49, 52, 53, 60] were proposed to synthesize entire 3D facial animation. Some methods [8, 12, 20, 40, 52] attempt to establish a direct audio-to-visual mapping through regression. Person-specific approaches [20, 40] can usually obtain plausible facial motions because of the relatively consistent talking style. VOCA [8] incorporates a robust audio feature extraction model capable of capturing various speaking styles, which can generate realistic speaker-independent animation and shows its wide applicability. MeshTalk [41] constructed a categorical latent space to adaptively generate motions based on the separated audio-correlated and audiouncorrelated facial information. FaceFormer [12] introduced two biased attention mechanisms and integrated the self-supervised pre-trained speech representations for the ill-posed and data scarcity issues. CodeTalker [53] proposed the discrete motion prior which regards the cross-modal mapping as a code query task in a finite proxy space of the learned codebook. SelfTalk [36] proposed a self-supervised approach to construct a lip-reading interpreter and speech recognizer to enhance the comprehensibility of generated lip movements. While data-driven approaches have shown impressive performance, accurately learning cross-modal audio-visual mappings remains challenging due to inherent uncertainties. These methods often regress the entire sequence, leading to over-smoothing and a lack of detailed facial dynamics. In contrast, our approach employs a coarse-to-fine learning mechanism that separates the problem into key motion capture and motion completion stages. This approach effectively mitigates cross-modal mapping uncertainties and reduces learning complexity, resulting in more precise and dynamic facial animation synthesis. + +# 3 Method + +# 3.1 Overview + +Let $\mathbf{x}$ represents the raw audio input and $\hat{\mathbf{Y}} = (\hat{\mathbf{y}}_1,\dots,\hat{\mathbf{y}}_N)\in \mathbb{R}^{N\times V\times 3}$ denotes the corresponding ground-truth sequence of facial movement over a neutral template, where $N$ indicates the number of visual frames and $V$ denotes the number + +![](images/732b3f9865d772653c6d7af7d745fbc23d87894c3accf04f47592faefe020760.jpg) +(a) The overview pipeline of our proposed KMTalk. + +![](images/575248935e24c38bbf8194ce56065892ff51a23bdb35467e4b9c07603f61ebce.jpg) +(b) The details of two key modules in KMTalk. +Fig. 2: Fig. 2a illustrates the overview pipeline of our proposed KMTalk. Initially, the Audio Encoder takes the input raw audio $\mathbf{x}$ and encodes it into audio features $\mathbf{A}$ . Subsequently, in the LKMA module, key motions $\mathbf{K}$ are generated from the audio $\mathbf{x}$ and $\mathbf{A}$ . Finally, the CMC module reintroduces audio features $\mathbf{A}$ to extend these key motions $\mathbf{K}$ into a full sequence $\mathbf{Y}$ . Fig. 2b presents the details of two key modules in KMTalk. In the Linguistic-based Key Motion Acquisition, a Phoneme-based Localization Method is used to identify key motion indices $\mathcal{I}$ from raw audio $\mathbf{x}$ . Based on audio features $\mathbf{A}$ and $\mathcal{I}$ , the Key Motion-focused Decoder generates key motions $\mathbf{K}$ . In the Cross-modal Motion Completion, the Motion Flow Encoder processes $\mathbf{K}$ and $\mathcal{I}$ , producing motion flow features $\Phi$ . Then, with the dynamic fusion weight $\mathbf{G}$ , the Multimodal-Guided Decoder combines $\Phi$ and $\mathbf{A}$ to decode the final motion sequence $\mathbf{Y}$ . + +of vertices in the facial mesh. The objective is to synthesize $\mathbf{Y} = (\mathbf{y}_1, \dots, \mathbf{y}_N)$ that is similar to $\hat{\mathbf{Y}}$ , driven by the raw audio $\mathbf{x}$ . The generated sequence should ensure lip synchronization with the audio while exhibiting natural facial movements. + +Due to the domain gap between modalities and the ill-posed nature of directly translating audio to facial movement sequences, it is a challenging task that often results in over-smooth or poorly synchronized lip movements. To address these issues, this paper introduces a coarse-to-fine approach with key motion embedding, integrating both linguistic and data-driven priors. The overview pipeline is presented in Fig. 2a. In the LKMA module, linguistic priors are introduced to locate and generate higher-quality key motions (see Sec. 3.2), followed by the CMC module, where these key motions are fleshed out into a complete sequence of facial motions (see Sec. 3.3). + +# 3.2 Linguistic-based Key Motion Acquisition + +In the realm of audio-driven 3D facial animation, it presents significant challenges to precisely define which frames constitute key motions. An alternative + +and simple solution is to use uniform or random sampling to determine the positions of these key motions. Although these approaches can boost performance to a certain degree due to the reduced learning complexity (refer to Sec. 4.2 for ablation studies), they fail to utilize the correlation between audio content and facial movements. However, we can leverage linguistic priors to capture pronounced articulatory actions, which are identifiable at phoneme boundaries. This approach circumvents the issue of over-smooth in the output sequence, thereby enhancing the overall quality of the results. + +As shown in the left of Fig. 2b, the Linguistic-based Key Motion Acquisition (LKMA) module receives as inputs the raw audio $\mathbf{x}$ and the audio features $\mathbf{A} = (\mathbf{a}_1,\dots,\mathbf{a}_N)\in \mathbb{R}^{N\times d}$ , where the audio features are derived from the Audio Encoder that utilizes the wav2vec 2.0 pre-trained model [2], processing the raw audio $\mathbf{x}$ as its input. Then it takes raw audio $\mathbf{x}$ as input to produce the key motion indices $\mathcal{I} = \{i_1,\dots,i_m\}$ , where $i_j\in \{1,\dots,N\},\forall i_j\in \mathcal{I}$ , through the proposed Phoneme-based Localization method. Subsequently, the Key Motion-focused Decoder utilizes the audio features $\mathbf{A}$ and the key motion indices $\mathcal{I}$ to generate key motions $\mathbf{K} = (\mathbf{k}_{i_1},\dots,\mathbf{k}_{i_m})\in \mathbb{R}^{m\times V\times 3}$ consisting of $m$ frames of facial movement which are located on the key motion positions, where $\mathbf{k}_{i_j}\simeq \hat{\mathbf{y}}_{i_j},\forall i_j\in \mathcal{I}$ . The process of the LKMA module is expressed as: + +$$ +\mathcal {I}, \mathbf {K} = \operatorname {L K M A} (\mathbf {x}, \mathbf {A}). \tag {1} +$$ + +Phoneme-based Localization. At phoneme boundaries, a noticeable offset is observed in the articulator movement, with visualization results available in supplementary materials. Furthermore, the phoneme boundary effects underscore the ease with which the boundaries of phonemes can be perceived [17]. Both experimental and theoretical analyses have demonstrated a distinct position-mapping relationship between the phoneme boundaries in the audio and the significant elements in the motion sequence, specifically the key motions. The mapping relationship can be harnessed to facilitate the initial alignment between the audio and visual modalities. + +Specifically, an Automatic Speech Recognition model [28] is first utilized to obtain the text content from the raw audio $\mathbf{x}$ . Then, a Montreal Forced Aligner [31] is adopted to align the audio and the text, producing the start and the end timestamps for each phoneme. Finally, the indices of these motion frames corresponding to the timestamps are regarded as key motion indices $\mathcal{I}$ , and the corresponding motions compose the key motions $\mathbf{K}$ . + +Key Motion-focused Decoder. It is utilized to synthesize key motions $\mathbf{K}$ of superior quality. Initially, employing $\mathcal{I}$ as indices, we extract the corresponding aligned audio features $\mathbf{A_k} = (\mathbf{a}_{i_1},\dots,\mathbf{a}_{i_m})\in \mathbb{R}^{m\times d}$ from the comprehensive audio features $\mathbf{A}$ . Subsequently, it adopts a modified transformer-based architecture, processes $\mathbf{A_k}$ to generate the key motions $\mathbf{K}$ . + +Loss Function. Intuitively, a straightforward approach to optimize the key motions $\mathbf{K}$ involves utilizing $\mathcal{I}$ to index $\hat{\mathbf{Y}}$ , resulting in $\hat{\mathbf{Y}}_{\mathbf{k}} = (\hat{\mathbf{y}}_{i_1},\dots ,\hat{\mathbf{y}}_{i_m})$ , which serves as the supervision for the training process. However, key motions typically occupy non-adjacent positions within the entire sequence. Hence, given the lack + +of inter-frame contextual information, attempting direct frame-by-frame regression of $\mathbf{K}$ towards $\hat{\mathbf{Y}}_{\mathbf{k}}$ may fall short in achieving accurate facial expressions, as well as producing smooth and realistic animation. + +To address this limitation, we adopt a pseudo-complete sequence training method that utilizes the ground-truth frame labels $\hat{\mathbf{Y}}$ at the non-key indices $\mathcal{I}' = \{1,\dots,N\} \setminus \mathcal{I}$ and the generated key motions $\mathbf{K}$ at the key indices $\mathcal{I}$ to form a predicted pseudo-complete sequence $\mathbf{Y}_{\mathbf{p}} \in \mathbb{R}^{N \times V \times 3}$ . Then, the model is trained by minimizing the loss between the pseudo-complete sequence $\mathbf{Y}_{\mathbf{p}}$ and the ground-truth sequence $\hat{\mathbf{Y}}$ . This enables the model to capture subtle changes between key motions and adjacent ground-truth frames, thereby mitigating inter-frame jitter and achieving more accurate regression of facial expressions. Following the SelfTalk [36], the loss function is formulated as: + +$$ +\mathcal {L} _ {L K M A} = \lambda_ {1} \mathcal {L} _ {r e c} + \lambda_ {2} \mathcal {L} _ {v e l} + \lambda_ {3} \mathcal {L} _ {l a t} + \lambda_ {4} \mathcal {L} _ {c t c}, \tag {2} +$$ + +where $\lambda_{1} = 1000.0, \lambda_{2} = 1000.0, \lambda_{3} = 0.001$ , and $\lambda_{4} = 0.0001$ in all of our experiments. $\mathcal{L}_{rec}$ , $\mathcal{L}_{vel}$ , and $\mathcal{L}_{lat}$ are measured by mean square error, while $\mathcal{L}_{ctc}$ is quantified by CTC Loss. The reconstruction loss $\mathcal{L}_{rec}$ quantifies the discrepancy between the predicted and the ground-truth facial movements. The velocity loss $\mathcal{L}_{vel}$ reduces frame jitter, ensuring smooth and natural lip movements. The latent consistency loss $\mathcal{L}_{lat}$ assesses the variance between latent features extracted from both the audio and lip shape encoders, aiming to align the learned audio and lip features. Lastly, the text consistency loss $\mathcal{L}_{ctc}$ evaluates the difference between the lip-reading decoder's output and the original text, ensuring the intelligibility of the lip-reading results. + +# 3.3 Cross-modal Motion Completion + +A straightforward method to obtain a complete talking face sequence is to directly integrate key motions into the entire face movements. However, it is important to note that while key motions capture essential facial dynamics, they may not encompass all details of non-key motions. This direct integration could result in mismatches between augmented non-key motions and their corresponding audio segments (see Sec. 4.2 for ablation studies). To mitigate this issue, we introduce a Cross-modal Motion Completion (CMC) module that jointly combines the audio features $\mathbf{A}$ , key motions $\mathbf{K}$ , and key motion indices $\mathcal{I}$ to generate a complete sequence of 3D facial meshes $\mathbf{Y}$ . The process can be formulated as: + +$$ +\mathbf {Y} = \operatorname {C M C} (\mathbf {A}, \mathbf {K}, \mathcal {I}). \tag {3} +$$ + +The details of the CMC module are illustrated on the right of Fig. 2b. + +Motion Flow Encoder. Key motions serve as a kinematic prior for the remaining frames, offering valuable insights into facial dynamics. To effectively capture the motion flow information provided by key motions, we draw inspiration from some manifold methods [32] to acquire motion flow features $\Phi \in \mathbb{R}^{N\times d}$ from the key motions $\mathbf{K}$ . Specifically, we first encode the key motions $\mathbf{K}$ into key motion + +context tokens $\Phi_{\mathbf{k}} \in \mathbb{R}^{m \times d}$ by multiple transformer-encoder layers. At the same time, the indices of non-key motions $\mathcal{T}'$ are encoded into positional encodings by a sinusoidal position embedding layer, representing the non-key frame positions. Then, we adopt the cross-attention layers to extract the intermediate tokens $\Phi_{\mathrm{non - key}} \in \mathbb{R}^{(N - m) \times d}$ , with key and value from linear transformations of $\Phi_{\mathbf{k}}$ and the query is the positional encodings of non-key frames indices. Above all, the implicit motion manifold proposed in CITL [32] is utilized to arrange $\Phi_{\mathbf{k}}$ and $\Phi_{\mathrm{non - key}}$ based on their indices, followed by a 1D convolution for fusion, ultimately obtaining the motion flow features $\Phi$ . + +Multimodal-Guided Decoder. The non-key motions' features in the complete motion sequence feature estimation are derived from the global context interpolation of key motions, which has a certain degree of information loss due to the audio feature selection process in the Key Motion-focused Decoder (Sec. 3.2). Hence, we have devised a multimodal-guided decoding approach that incorporates audio modalities to furnish comprehensive information across the entire temporal scale, alongside motion flow to offer facial motion priors. These elements serve to guide and constrain the decoding process, thereby facilitating the precise generation of the motion sequence. Technically, we simply employ the gated mechanism [9,57,58] for the modality fusion, which can be formulated as: + +$$ +\mathbf {G} = \sigma ([ \mathbf {A}, \boldsymbol {\Phi} ] \mathbf {W}), \tag {4} +$$ + +$$ +\mathbf {Z} = \mathbf {G} \odot \mathbf {A} + (1 - \mathbf {G}) \odot \boldsymbol {\Phi}, \tag {5} +$$ + +where $\odot$ represents the element-wise multiplication operation, $[\cdot, \cdot]$ denotes the concatenation operation, $\mathbf{W} \in \mathbb{R}^{2d \times d}$ is a parameter and $\sigma$ is a sigmoid function, while $\mathbf{G} \in \mathbb{R}^{N \times d}$ dynamically selects features from the audio features $\mathbf{A}$ and the motion flow features $\Phi$ . Then, the Motion Decoder, a transformer-based structural model, is employed to transform the fusion features $\mathbf{Z}$ into the complete 3D facial movement sequence $\hat{\mathbf{Y}}$ . + +Loss Function. We train the CMC module utilizing the reconstruction loss and velocity loss. The total loss function is defined as: + +$$ +\mathcal {L} _ {C M C} = \mathcal {L} _ {r e c} + \mathcal {L} _ {v e l}. \tag {6} +$$ + +# 4 Experiment + +Dataset. BIWI [13] consists of 40 paired audio-visual sentences from 14 subjects. The 3D facial geometries, consisting of 23370 vertices, were captured at a frame rate of 25fps, and the average duration of each sequence was 4.67 seconds. We adopt the same evaluation protocol as FaceFormer [12] on the BIWI dataset. Specifically, the training set (BIWI-Train) comprises 190 sentences, while the validation set (BIWI-Val) encompasses 24 sentences. The dataset is divided into two testing sets: BIWI-Test-A, which comprises 24 sentences articulated by six subjects observed during training, and BIWI-Test-B, which consists of 32 sentences + +Table 1: Quantitative comparisons on the BIWI-Test-A and VOCA-Test datasets. The results of Lip-Vertex Error (LVE) and the upper-Face Dynamics Deviation (FDD) are reported. For both metrics, the lower the better. + +
MethodsBIWI-Test-AVOCA-Test
LVE↓×10-4mmFDD↓×10-5mmLVE↓×10-5mmFDD↓×10-7mm
VOCA [8]6.55638.18164.92454.8447
MeshTalk [41]5.91815.10254.54415.2062
FaceFormer [12]5.30774.64084.10904.6675
CodeTalker [53]4.79144.11703.94454.5422
SelfTalk [36]4.24853.57613.22384.0912
KMTalk (Ours)3.96542.54462.26394.0594
+ +uttered by eight unseen subjects. VOCASET [8] consists of 480 paired audiovisual sequences from 12 subjects. Each sequence is recorded at a frame rate of 60fps and ranges in duration from 3 to 4 seconds. The 3D face mesh for each sequence consists of 5023 vertices. To ensure a fair comparison, we used identical training (VOCA-Train), validation (VOCA-Val), and testing (VOCA-Test) partitions as methods [12,36,53]. + +Baselines. We compare against current state-of-the-arts method, including VOCA [8], MeshTalk [41]. FaceFormer [12], CodeTalker [53], and SelfTalk [36]. Faceformer [12] employs a transformer-based model to incorporate long-term audio context and synthesizes sequential motions in an autoregressive manner. CodeTalker [53] introduces discrete motion priors to enable self-reconstruction of real facial movements, mitigating the issue of excessive smoothing in facial motion. SelfTalk [36] designs a learning-based recognizer to minimize the domain gap between diverse modalities. + +Evaluation Metrics. Following CodeTalker [53] and SelfTalk [36], we adopt two metrics for the quantitative evaluation of speech-driven facial animation: lip vertex error (LVE) to measure lip synchronization and upper-face dynamics deviation (FDD) to assess the overall facial dynamics. The LVE for each frame is defined as the maximal L2 error among all lip vertices for each frame and takes the average over all frames. This L2 error is computed by comparing the predictions with the processed 3D face geometry data. FDD is introduced to quantify the variation in facial dynamics between a synthetic motion sequence and the reference sequence. The implementation of FDD is to calculate the difference between the variances of vertex offsets in the upper-face region and the variances of ground truth vertex offsets. In addition, we visualize the prediction results for qualitative evaluation. + +Implementation Details. For a fair comparison, KMTalk operates at a frame rate of 30 fps on VOCASET and 25 fps on BIWI, following the setting of previous methods [12, 36, 53]. Also, it can naturally adapt to a higher frame rate, as shown in the supplementary materials. In the LKMA module, we first employ the Phoneme-based Localization method to process the raw audio and obtain key motion indices for data preprocessing, which costs less than 10 minutes on two datasets [8, 13]. Secondly, we train the Key Motion Decoder on a single NVIDIA + +![](images/626f767d17967288bffb105075a792a0cb286ae1b11d81e744077f7135ada470.jpg) +Fig. 3: Qualitative comparisons on VOCA-Test (left) and BIWI-Test-B (right). We provide visual comparisons of facial animations synchronized with six syllables extracted from the test speech sequences. The 1st, 3rd, and 5th rows display synthesized meshes and their corresponding ground-truths, while the 2nd, 4th, and 6th rows visualize the L2 loss for individual frames. Our method demonstrates more precise mouth movement on syllables like /æ/ that require a wide-open mouth. For syllables that start with a closed mouth and then slightly open, such as /bɪ/, our KMTalk generates more synchronized motion sequences visually. The last row visualizes the mean square errors of different methods across all sentences in the test set for a specific subject. + +RTX 3090 for 200 epochs (about 2 hours) using the Adam optimizer [21]. The learning rate is initialized as $10^{-4}$ , and the mini-batch size is set to 1. In the CMC module, we train for 200 epochs (approximately 2 hours) with the same training settings as the Key Motion Decoder. It is noteworthy that, since the training of the two modules is independent of each other, we can train both modules concurrently to enhance training efficiency. + +# 4.1 Comparisons against State-of-the-art Methods + +Quantitative Comparisons. We computed the lip vertex error (LVE) and facial dynamics deviation (FDD) for all sequences within the BIWI-Test-A and VOCA-Test datasets. According to Table 1, our proposed KMTalk demonstrated lower errors compared to the alternative methods examined. Notably, the lip vertex error for our method on the VOCA-Test dataset is $30\%$ lower than the + +recently introduced SelfTalk [36], and the FDD is $27\%$ lower than SelfTalk on the BIWI-Test-A dataset, providing compelling evidence for the advantages of our proposed KMTalk. This indicates that our approach is more effective in achieving audio-visual alignment, thereby leading to improved lip synchronization. + +Qualitative Comparisons. In Fig. 3, we visualize the output facial meshes from different methods as well as ground truths for reference. Additionally, we display error maps calculated from the vertex L2 loss between the generated and ground truth meshes. It is evident that our method consistently yields lower errors across different speech sequences, demonstrating its ability to generate more accurate facial animation sequences. Notably, for representative syllables (e.g. $/a\mathrm{e}/$ ), KMTalk closely approximates ground truth, excelling in synthesizing accurate lip movements for syllables requiring significant mouth opening. Additionally, for syllables starting with mouth closure followed by a slight opening (e.g., $/b\mathrm{i}/$ ), KMTalk produces more natural and synchronized motions. We recommend that readers watch the supplementary video for more detailed comparisons. It showcases KMTalk's capability to generate coherent, realistic animations with precise lip synchronization. + +User Studies. A user study stands as a dependable evaluation method in the context of 3D talking faces. Following the strategy of Faceformer [12], we conduct pairwise comparisons between our method and baselines [12,36,53], as well as ground truths. This study encompassed the assessment of two key metrics: perceptual lip synchronization and facial realism. Participants were presented with side-by-side comparisons and were tasked with selecting the better facial animation based on their personal preferences. We computed the ratio of user preferences as a measurement of satisfaction evaluation on BIWI-Test-B and VOCA-Test. We randomly sampled 30 examples from each test set and compared the performance of KMTalk with four aforementioned settings on each sample. Therefore, we constructed a total of 240 different video pairs and randomly selected 24 video pairs for the two metrics assessments for each participant. Our user study involved 30 participants with a strong capability for audio-visual perception, resulting in 720 effective evaluation entries. As demonstrated in Table 2, our approach indicates superior perceptual lip synchronization and facial realism. For instance, a noteworthy $60.0\%$ of users favored our lip synchronization method on BIWI-Test-B in comparison to SelfTalk [36]. Overall, it shows that KMTalk can generate more favorable facial animations from speeches. + +# 4.2 Ablation Studies + +In this section, we perform ablation studies on to evaluate the influence of different components within our proposed KMTalk framework on the quality of the generated 3D talking faces. The quantitative results on BIWI are in Table 3, and the qualitative results are in Fig. 4. In addition, the results of the ablation study on VOCA-Test can be found in the supplementary material. In Table 4, we further investigate the robustness of our approach to different Phoneme-based localization methods and the possible errors during phoneme extraction. + +Table 2: User study results on BIWI-Test-B and VOCA-Test. + +
MethodMetricBIWI-Test-BVOCA-Test
competitorourscompetitorours
Ours vs. FaceFormerLip Sync24.4%75.6%26.7%73.3%
Realism25.6%74.4%28.9%71.1%
Ours vs. CodeTalkerLip Sync31.1%68.9%37.8%62.2%
Realism27.8%72.2%35.6%64.4%
Ours vs. SelfTalkLip Sync40.0%60.0%43.3%56.7%
Realism38.9%61.1%41.1%58.9%
Ours vs. GTLip Sync54.4%45.6%56.7%43.3%
Realism52.2%47.8%56.7%43.3%
+ +Table 3: Ablation study for our components on BIWI-Test-A. + +
Phoneme-based Localization MethodKey Motion-focused Audio in CMCLVE↓FDD↓
Decoder✓4.24853.5761
---4.16482.8713
-✓✓4.13812.9546
✓-✓4.88593.2780
✓✓-3.96542.5446
+ +What's the effect of the Phoneme-based localization method for key motion capture? The Phoneme-based Localization Method enables us to identify key frames of speech with notable facial expression transitions. We can replace it with uniform sampling, neglecting the crucial content information of the audio. Specifically, we experimented using uniform sampling at a rate of $33\%$ to closely align with the number of key motions. For further comparisons under different numbers of sampled elements, please refer to the supplementary material. In Table 3, we observe degradation in all metrics with uniform sampling. This underscores the importance of accurate key motion capture in speech-driven talking face generation. Additionally, it showcases that the linguistic-based key motion capture is better equipped to mitigate audio-visual uncertainty and recover more precise facial motions. + +What's the effect of the Key Motion Decoder? Our method employs a specialized key motion decoder to generate facial meshes based on keyframe indices obtained from phoneme-based localization methods. An alternative approach to generating key motions is to select corresponding facial meshes from a complete motion sequence produced by an existing method such as SelfTalk [36]. Table 3 demonstrates that both metrics deteriorated but still outperformed the state-of-the-art method SelfTalk. This indicates that the Key Motion Decoder can enhance the quality of key motion generation, resulting in more plausible facial animations. It also suggests that even if the captured key motions are not accurate enough, the CMC module can further refine output full motions. + +![](images/47975d6ec9b6d62e30e566b978fbc70e53c4b4ea2aa0ba8b68ef5a5206e109cf.jpg) +Fig. 4: Qualitative ablation studies on the input speech "specifically". For each method variant, we removed one of three modules: PLM (Phoneme-based Localization Method), KMD (Key Motion-focused Decoder), and AG (Audio Guidance in CMC). Error maps between generated and the ground-truth mesh sequence were visualized. Our final model yielded the best results, showcasing the effectiveness of each module. + +What's the effect of audio guidance in the Cross-modal Motion Completion? The Cross-modal Motion Completion leverages audio features to guide the completion of the full motion sequence. To assess its usefulness, we implement a method variant that removes the audio feature guidance. Table 3 demonstrates a notable degradation in both metrics, particularly with a $24\%$ increase in Lip Vertex Error (LVE) and a $27\%$ increase in upper-face dynamics deviation (FDD). This suggests that audio information plays a crucial role in refining fine-grained lip movements and enhancing audio-visual consistency and temporal smoothness. Despite the degradation, the FDD metric can still outperform state-of-the-art methods, underscoring the significance of key motion capture for achieving temporally coherent full motion synthesis. + +Is our method sensitive to different Phomene-based Localization methods? We experimented with different Automatic Speech Recognition (ASR) models, such as Auto-avs r [28] and Whisper [39]. The results in the first three rows of Table 4 show that our KMTalk method consistently maintained high performance, achieving at least a $21\%$ improvement in FDD regardless of the ASR model used [28, 39]. This highlights the robustness of our approach across various ASR models. Besides, to simulate phoneme localization deviations, we shifted all key motion indices extracted by Auto-avs r [28] one frame to the right. The results in the last row of Table 4 indicate negligible variations. This demonstrates the robustness of our method to inaccurate key frame localization. + +Table 4: Robust analysis of Phoneme-based Localization on BIWI-Test-A. + +
MethodsLVE↓ ×10-4mmFDD↓ ×10-5mm
Auto-avsr [28]3.96542.5446
Whisper-large [39]4.07182.8141
Whisper-tiny [39]4.06432.8083
Auto-avsr+offset3.99912.6420
+ +Table 5: The results of integrating our proposed KMTalk with existing methods on BIWI-Test-A. + +
MethodsLVE↓ ×10-4mmFDD↓ ×10-5mm
FaceFormer [12]Original5.30774.6408
After5.27934.2654
CodeTalker [53]Original4.79144.1170
After4.50963.9043
SelfTalk [36]Original4.24853.5761
After4.11222.8668
+ +# 4.3 Integration with Existing Methods + +Existing approaches focus on enhancing prediction outcomes by designing elaborate priors or the learning-based recognizer, which may be highly coupled with the proposed architecture of these methods. Our KMTalk introduces a new learning strategy of speech-driven talking face generation, which is orthogonal to these approaches. Therefore, we can explore whether performance can be enhanced by applying our progressive learning scheme without the need for additional fine-tuning of their models. Detailed implementation is in the supplementary material. As shown in Table 5, the results of existing methods are improved after integration with our proposed progressive learning mechanism utilizing key motion embeddings. This further emphasizes the efficacy of our design. + +# 5 Conclusion + +In this work, we introduce KMTalk, a novel method for progressively learning 3D facial animation from speeches using key motion embeddings. It incorporates linguistic priors for key motion generation and extends them to a full motion sequence via data-driven priors. We propose phoneme-based localization methods to determine the temporal position of key facial motions, improving lip-speech synchronization by aligning motion transitions with phoneme changes. Additionally, we design a cross-modal facial motion completion module that synthesizes the entire motion sequence from key motions and audio features, enhancing lip-speech synchronization and motion coherence. Extensive evaluations of the datasets demonstrate KMTalk's superiority over existing methods, producing more accurate and realistic animations. Moreover, coupling our idea with existing methods consistently improves performance, further verifying the efficacy of our proposed progressive learning mechanism based on key motion acquisition. Although the proposed method has demonstrated its robustness to inaccurate keyframe localization, it may encounter errors in dialect variations. Integrating advanced ASR(Automatic Speech Recognition) technology in the future could enhance its adaptability to various speech patterns. + +Acknowledgement The research is partially supported by National Key Research and Development Program of China (2023YFC3502900), National Natural Science Foundation of China (No.62176093, 61673182), Key Realm Research and Development Program of Guangzhou (No.202206030001), Guangdong-Hong Kong-Macao Joint Innovation Project (No.2023A0505030016), Guangdong Natural Science Foundation (No. 2024A1515012217). + +# References + +1. Alghamdi, M.M., Wang, H., Bulpitt, A.J., Hogg, D.C.: Talking head from speech audio using a pre-trained image generator. In: ACM MM. pp. 5228-5236 (2022) +2. Baevski, A., Zhou, Y., Mohamed, A., Auli, M.: wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems 33, 12449-12460 (2020) +3. Bao, L., Zhang, H., Qian, Y., Xue, T., Chen, C., Zhe, X., Kang, D.: Learning audio-driven viseme dynamics for 3d face animation. arXiv preprint arXiv:2301.06059 (2023) +4. Cao, Y., Tien, W.C., Faloutsos, P., Pighin, F.: Expressive speech-driven facial animation. ACM Trans. Graph 24(4), 1283-1302 (2005) +5. Chen, L., Cui, G., Liu, C., Li, Z., Kou, Z., Xu, Y., Xu, C.: Talking-head generation with rhythmic head motion. In: ECCV. pp. 35-51. Springer (2020) +6. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: ACCV. pp. 251-263. Springer (2017) +7. Cohen, M.M., Clark, R., Massaro, D.W.: Animated speech: Research progress and applications. In: AVSP (2001) +8. Cudeiro, D., Bolkart, T., Laidlaw, C., Ranjan, A., Black, M.J.: Capture, learning, and synthesis of 3d speaking styles. In: CVPR. pp. 10101-10111 (2019) +9. Dai, G., Zhang, Y., Wang, Q., Du, Q., Yu, Z., Liu, Z., Huang, S.: Disentangling writer and character styles for handwriting generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5977-5986 (2023) +10. Das, D., Biswas, S., Sinha, S., Bhowmick, B.: Speech-driven facial animation using cascaded gans for learning of motion and texture. In: ECCV. pp. 408-424. Springer (2020) +11. Edwards, P., Landreth, C., Fiume, E., Singh, K.: Jali: an animator-centric viseme model for expressive lip synchronization. ACM Trans. Graph 35(4), 1-11 (2016) +12. Fan, Y., Lin, Z., Saito, J., Wang, W., Komura, T.: Faceformer: Speech-driven 3d facial animation with transformers. In: CVPR. pp. 18770-18780 (2022) +13. Fanelli, G., Gall, J., Romsdorfer, H., Weise, T., Van Gool, L.: A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia 12(6), 591-598 (2010) +14. Fisher, C.G.: Confusions among visually perceived consonants. Journal of speech and hearing research 11(4), 796-804 (1968) +15. Guo, Y., Chen, K., Liang, S., Liu, Y.J., Bao, H., Zhang, J.: Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In: ICCV. pp. 5784-5794 (2021) +16. Habibie, I., Xu, W., Mehta, D., Liu, L., Seidel, H.P., Pons-Moll, G., Elgharib, M., Theobalt, C.: Learning speech-driven 3d conversational gestures from video. In: Proceedings of the 21st ACM International Conference on Intelligent Virtual Agents. pp. 101-108 (2021) + +17. Iverson, P., Kuhl, P.K.: Perceptual magnet and phoneme boundary effects in speech perception: Do they arise from a common mechanism? Perception & psychophysics 62, 874-886 (2000) +18. Ji, X., Zhou, H., Wang, K., Wu, Q., Wu, W., Xu, F., Cao, X.: Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In: ACM SIGGRAPH. pp. 1-10 (2022) +19. Ji, X., Zhou, H., Wang, K., Wu, W., Loy, C.C., Cao, X., Xu, F.: Audio-driven emotional video portraits. In: CVPR. pp. 14080-14089 (2021) +20. Karras, T., Aila, T., Laine, S., Herva, A., Lehtinen, J.: Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Trans. Graph 36(4), 1-12 (2017) +21. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +22. Lewis, J.: Automated lip-sync: Background and techniques. IEEE Trans Vis Comput Graph 2(4), 118-122 (1991) +23. Liang, B., Pan, Y., Guo, Z., Zhou, H., Hong, Z., Han, X., Han, J., Liu, J., Ding, E., Wang, J.: Expressive talking head generation with granular audio-visual control. In: CVPR. pp. 3387-3396 (2022) +24. Liu, C.: An analysis of the current and future state of 3d facial animation techniques and systems (2009) +25. Liu, X., Xu, Y., Wu, Q., Zhou, H., Wu, W., Zhou, B.: Semantic-aware implicit neural audio-driven video portrait generation. In: ECCV. pp. 106-125. Springer (2022) +26. Liu, Y., Xu, F., Chai, J., Tong, X., Wang, L., Huo, Q.: Video-audio driven real-time facial animation. ACM Trans. Graph 34(6), 1-10 (2015) +27. Lu, L., Wu, R., Lin, H., Lu, J., Jia, J.: Video frame interpolation with transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3532-3542 (2022) +28. Ma, P., Haliassos, A., Fernandez-Lopez, A., Chen, H., Petridis, S., Pantic, M.: Auto-avsr: Audio-visual speech recognition with automatic labels. arxiv 2023. arXiv preprint arXiv:2303.14307 +29. Massaro, D., Cohen, M., Tabain, M., Beskow, J., Clark, R.: 12 animated speech: research progress and applications (2012) +30. Mattheyses, W., Verhelst, W.: Audiovisual speech synthesis: An overview of the state-of-the-art. Speech Communication 66, 182-217 (2015) +31. McAuliffe, M., Sonderegger, M.: English mfa acoustic model v2.2.1. Tech. rep., https://mfa-models.readthedocs.io/acoustic/English/EnglishMFAacousticmodelv2_2_1.html (May 2023) +32. Mo, C.A., Hu, K., Long, C., Wang, Z.: Continuous intermediate token learning with implicit motion manifold for keyframe based motion interpolation. In: CVPR. pp. 13894-13903 (2023) +33. Niklaus, S., Mai, L., Liu, F.: Video frame interpolation via adaptive convolution. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 670-679 (2017) +34. Nocentini, F., Ferrari, C., Berretti, S.: Learning landmarks motion from speech for speaker-agnostic 3d talking heads generation. arXiv preprint arXiv:2306.01415 (2023) +35. Pang, Y., Zhang, Y., Quan, W., Fan, Y., Cun, X., Shan, Y., Yan, D.m.: Dpe: Disentanglement of pose and expression for general video portrait editing. In: CVPR. pp. 427-436 (2023) + +36. Peng, Z., Luo, Y., Shi, Y., Xu, H., Zhu, X., Liu, H., He, J., Fan, Z.: Selftalk: A self-supervised commutative training diagram to comprehend 3d talking faces. arXiv preprint arXiv:2306.10799 (2023) +37. Pham, H.X., Wang, Y., Pavlovic, V.: End-to-end learning for 3d facial animation from speech. In: ICMI. pp. 361-365 (2018) +38. Prajwal, K., Mukhopadhyay, R., Namboodiri, V.P., Jawahar, C.: A lip sync expert is all you need for speech to lip generation in the wild. In: ACM MM. pp. 484-492 (2020) +39. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision. In: International Conference on Machine Learning. pp. 28492-28518. PMLR (2023) +40. Richard, A., Lea, C., Ma, S., Gall, J., De la Torre, F., Sheikh, Y.: Audio-and gaze-driven facial animation of codec avatars. In: WACV. pp. 41-50 (2021) +41. Richard, A., Zollhöfer, M., Wen, Y., De la Torre, F., Sheikh, Y.: Meshtalk: 3d face animation from speech using cross-modality disentanglement. In: ICCV. pp. 1173-1182 (2021) +42. Shen, S., Li, W., Zhu, Z., Duan, Y., Zhou, J., Lu, J.: Learning dynamic facial radiance fields for few-shot talking head synthesis. In: ECCV. pp. 666-682. Springer (2022) +43. Tanaka, H., Nakamura, S., et al.: The acceptability of virtual characters as social skills trainers: usability study. JMIR human factors 9(1), e35358 (2022) +44. Tang, J., Dai, A., Nie, Y., Markhasin, L., Thies, J., Nießner, M.: Dphms: Diffusion parametric head models for depth-based tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1111-1122 (2024) +45. Taylor, S., Kim, T., Yue, Y., Mahler, M., Krahe, J., Rodriguez, A.G., Hodgins, J., Matthews, I.: A deep learning approach for generalized speech animation. ACM Trans. Graph 36(4), 1-11 (2017) +46. Taylor, S.L., Mahler, M., Theobald, B.J., Matthews, I.: Dynamic units of visual speech. In: ACM SIGGRAPH. pp. 275-284 (2012) +47. Vougioukas, K., Petridis, S., Pantic, M.: Realistic speech-driven facial animation with gans. International Journal of Computer Vision 128, 1398-1413 (2020) +48. Wang, J., Qian, X., Zhang, M., Tan, R.T., Li, H.: Seeing what you said: Talking face generation guided by a lip reading expert. In: CVPR. pp. 14653-14662 (2023) +49. Wang, Q., Fan, Z., Xia, S.: 3d-talkemo: Learning to synthesize 3d emotional talking head. arXiv preprint arXiv:2104.12051 (2021) +50. Wang, S., Li, L., Ding, Y., Yu, X.: One-shot talking face generation from single-speaker audio-visual correlation learning. In: AAAI. vol. 36, pp. 2531-2539 (2022) +51. Wen, S., Liu, W., Yang, Y., Huang, T., Zeng, Z.: Generating realistic videos from keyframes with concatenated gans. IEEE Transactions on Circuits and Systems for Video Technology 29(8), 2337-2348 (2018) +52. Wu, H., Zhou, S., Jia, J., Xing, J., Wen, Q., Wen, X.: Speech-driven 3d face animation with composite and regional facial movements. arXiv preprint arXiv:2308.05428 (2023) +53. Xing, J., Xia, M., Zhang, Y., Cun, X., Wang, J., Wong, T.T.: Codetalker: Speechdriven 3d facial animation with discrete motion prior. In: CVPR. pp. 12780-12790 (2023) +54. Xu, Y., Feng, A.W., Marsella, S., Shapiro, A.: A practical and configurable lip sync method for games. In: Proceedings of Motion on Games, pp. 131-140 (2013) + +55. Yi, R., Ye, Z., Zhang, J., Bao, H., Liu, Y.J.: Audio-driven talking face video generation with learning-based personalized head pose. arXiv preprint arXiv:2002.10137 (2020) +56. Yin, S., Wu, C., Yang, H., Wang, J., Wang, X., Ni, M., Yang, Z., Li, L., Liu, S., Yang, F., et al.: Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv preprint arXiv:2303.12346 (2023) +57. Yu, D., Li, X., Zhang, C., Liu, T., Han, J., Liu, J., Ding, E.: Towards accurate scene text recognition with semantic reasoning networks. In: CVPR. pp. 12113-12122 (2020) +58. Yue, X., Kuang, Z., Lin, C., Sun, H., Zhang, W.: Robustscanner: Dynamically enhancing positional clues for robust text recognition. In: ECCV. pp. 135-151. Springer (2020) +59. Zhang, B., Qi, C., Zhang, P., Zhang, B., Wu, H., Chen, D., Chen, Q., Wang, Y., Wen, F.: Metaportrait: Identity-preserving talking head generation with fast personalized adaptation. In: CVPR. pp. 22096-22105 (2023) +60. Zhang, C., Ni, S., Fan, Z., Li, H., Zeng, M., Budagavi, M., Guo, X.: 3d talking face with personalized pose dynamics. IEEE Trans Vis Comput Graph. (2021) +61. Zhou, H., Sun, Y., Wu, W., Loy, C.C., Wang, X., Liu, Z.: Pose-controllable talking face generation by implicitly modularized audio-visual representation. In: CVPR. pp. 4176-4186 (2021) \ No newline at end of file diff --git a/2024/KMTalk_ Speech-Driven 3D Facial Animation with Key Motion Embedding/images.zip b/2024/KMTalk_ Speech-Driven 3D Facial Animation with Key Motion Embedding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f321d02cf30552b3f272e8bd01cd407496140b71 --- /dev/null +++ b/2024/KMTalk_ Speech-Driven 3D Facial Animation with Key Motion Embedding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48f9735720746509e34fad3023144c0e817769de55ed21e1b59cadd004a506ec +size 531046 diff --git a/2024/KMTalk_ Speech-Driven 3D Facial Animation with Key Motion Embedding/layout.json b/2024/KMTalk_ Speech-Driven 3D Facial Animation with Key Motion Embedding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..850b82f5ec64f97a785cae8819aa245e954a9fe5 --- /dev/null +++ b/2024/KMTalk_ Speech-Driven 3D Facial Animation with Key Motion Embedding/layout.json @@ -0,0 +1,9146 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 111, + 465, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 111, + 465, + 147 + ], + "spans": [ + { + "bbox": [ + 148, + 111, + 465, + 147 + ], + "type": "text", + "content": "KMTalk: Speech-Driven 3D Facial Animation with Key Motion Embedding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 173, + 479, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 479, + 205 + ], + "type": "text", + "content": "Zhihao " + }, + { + "bbox": [ + 132, + 173, + 479, + 205 + ], + "type": "inline_equation", + "content": "\\mathrm{Xu}^{1^*}" + }, + { + "bbox": [ + 132, + 173, + 479, + 205 + ], + "type": "text", + "content": ", Shengjie Gong" + }, + { + "bbox": [ + 132, + 173, + 479, + 205 + ], + "type": "inline_equation", + "content": "^{1^*}" + }, + { + "bbox": [ + 132, + 173, + 479, + 205 + ], + "type": "text", + "content": ", Jiapeng Tang" + }, + { + "bbox": [ + 132, + 173, + 479, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 173, + 479, + 205 + ], + "type": "text", + "content": ", Lingyu Liang" + }, + { + "bbox": [ + 132, + 173, + 479, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 173, + 479, + 205 + ], + "type": "text", + "content": ", Yining Huang" + }, + { + "bbox": [ + 132, + 173, + 479, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 173, + 479, + 205 + ], + "type": "text", + "content": ", Haojie Li" + }, + { + "bbox": [ + 132, + 173, + 479, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 173, + 479, + 205 + ], + "type": "text", + "content": ", and Shuangping Huang" + }, + { + "bbox": [ + 132, + 173, + 479, + 205 + ], + "type": "inline_equation", + "content": "^{1,3\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 225, + 214, + 389, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 214, + 389, + 227 + ], + "spans": [ + { + "bbox": [ + 225, + 214, + 389, + 227 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 225, + 214, + 389, + 227 + ], + "type": "text", + "content": " South China University of Technology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 239, + 226, + 374, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 226, + 374, + 237 + ], + "spans": [ + { + "bbox": [ + 239, + 226, + 374, + 237 + ], + "type": "text", + "content": "2 Technical University of Munich" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 263, + 237, + 351, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 237, + 351, + 248 + ], + "spans": [ + { + "bbox": [ + 263, + 237, + 351, + 248 + ], + "type": "text", + "content": "3 Pazhou Laboratory" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 249, + 454, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 249, + 454, + 259 + ], + "spans": [ + { + "bbox": [ + 160, + 249, + 454, + 259 + ], + "type": "text", + "content": "{eezhihaoxu,eeshengjiegong}@mail.scut.edu.cn, eehsp@scut.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 289, + 452, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 289, + 452, + 508 + ], + "spans": [ + { + "bbox": [ + 159, + 289, + 452, + 508 + ], + "type": "text", + "content": "Abstract. We present a novel approach for synthesizing 3D facial motions from audio sequences using key motion embeddings. Despite recent advancements in data-driven techniques, accurately mapping between audio signals and 3D facial meshes remains challenging. Direct regression of the entire sequence often leads to over-smoothed results due to the ill-posed nature of the problem. To this end, we propose a progressive learning mechanism that generates 3D facial animations by introducing key motion capture to decrease cross-modal mapping uncertainty and learning complexity. Concretely, our method integrates linguistic and data-driven priors through two modules: the linguistic-based key motion acquisition and the cross-modal motion completion. The former identifies key motions and learns the associated 3D facial expressions, ensuring accurate lip-speech synchronization. The latter extends key motions into a full sequence of 3D talking faces guided by audio features, improving temporal coherence and audio-visual consistency. Extensive experimental comparisons against existing state-of-the-art methods demonstrate the superiority of our approach in generating more vivid and consistent talking face animations. Consistent enhancements in results through the integration of our proposed learning scheme with existing methods underscore the efficacy of our approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 518, + 422, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 518, + 422, + 530 + ], + "spans": [ + { + "bbox": [ + 160, + 518, + 422, + 530 + ], + "type": "text", + "content": "Keywords: Speech-driven " + }, + { + "bbox": [ + 160, + 518, + 422, + 530 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 518, + 422, + 530 + ], + "type": "text", + "content": " 3D Facial Animation " + }, + { + "bbox": [ + 160, + 518, + 422, + 530 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 518, + 422, + 530 + ], + "type": "text", + "content": " Key Motion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 551, + 229, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 551, + 229, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 551, + 229, + 562 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 575, + 479, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 575, + 479, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 575, + 479, + 611 + ], + "type": "text", + "content": "Speech-driven 3D facial animation aims to create realistic talking heads that synchronize with input speech. It plays a significant role in many applications of virtual reality, like film production, computer gaming, and education [24, 43]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 612, + 480, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 612, + 480, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 612, + 480, + 635 + ], + "type": "text", + "content": "The main challenge of speech-driven 3D talking faces lies in the ill-posed problem caused by the cross-modal mapping uncertainty from the speech domain to" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 642, + 268, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 642, + 268, + 653 + ], + "spans": [ + { + "bbox": [ + 145, + 642, + 268, + 653 + ], + "type": "text", + "content": "*Authors contributed equally." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 147, + 653, + 243, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 653, + 243, + 665 + ], + "spans": [ + { + "bbox": [ + 147, + 653, + 243, + 665 + ], + "type": "text", + "content": "†Corresponding author." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 482, + 268 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 482, + 268 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 482, + 268 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 482, + 268 + ], + "type": "image", + "image_path": "79288203aa1c38b920ac9f2d7ce605aaac642a3c4776e940c5f5fb8a7be84a81.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 277, + 482, + 312 + ], + "lines": [ + { + "bbox": [ + 131, + 277, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 131, + 277, + 482, + 312 + ], + "type": "text", + "content": "Fig. 1: Compared to the state-of-the-art method Selftalk, our approach can produce more vivid lip motions from speeches, since we introduce linguistic priors to characterize key motions and utilize data-driven priors to interpolate non-key motions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 340, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 639 + ], + "type": "text", + "content": "the 3D motion domain. Since there may be multiple plausible outputs for input audio, effective regularizations and constraints should be integrated into the system, to generate vivid facial motions. The related methods can be roughly divided into linguistic-based methods [3, 11, 29, 46, 54] and data-driven methods [8, 12, 16, 26, 34, 36, 41, 49, 52, 53]. For linguistic-based methods [3, 11, 29, 46, 54], a set of intricate phoneme-to-viseme mapping rules is manually designed to generate the talking mouth based on priors from visemes or linguistic knowledge. While these methods explicitly control the animation of articulation processes, such as procedural lip sync with animation curves, their focus is mainly on localized facial movements, like those of the mouth area, lacking a systematic approach for modeling comprehensive facial motion. Thanks to the established audio-to-face datasets, learning-based methods [12, 16, 34, 36, 41, 49, 52, 53] choose to map audio signals into 3D facial meshes in a data-driven manner. Most of these works [8, 12, 16, 26, 34, 36, 41, 48, 49, 52, 53] typically formulated the cross-modal mapping of 3D talking face generation as a regression task, such as MeshTalk [41], FaceFormer [12], and SelfTalk [36]. While achieving impressive performance, they exhibit common limitations in their learning schemes. Firstly, they directly learn the ambiguous cross-modal mapping between audio and facial expression sequences, always leading to sub-optimal results in terms of temporal coherence and audio-visual consistency. Secondly, these methods typically regress the entire sequence without considering key motion cues, hindering the capture of detailed facial dynamics and accurate lip movements, particularly in complex facial expressions such as puckering or opening the mouth (as depicted in Fig. 1). Lastly, they overlook linguistic priors essential for simulating the articulation process, thereby limiting their ability to achieve precise lip-speech synchronization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "To this end, inspired by the keyframe-based video generation techniques observed in recent studies [27, 33, 51, 56], which prioritize the generation of" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "Z. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": "keyframes before adding detailed elements, we introduce a progressive learning mechanism that generates realistic 3D facial animations from audio inputs by incorporating key motion embeddings. The key idea is to initially generate key facial expressions, and then interpolate the intermediate motions to obtain the entire motion sequence, which significantly reduces the uncertainty of cross-modal mapping and eases the learning difficulty. Concretely, our method integrates linguistic and data-driven priors through two modules: the linguistic-based key motion acquisition and the cross-modal motion completion. The linguistic-based key motion acquisition module utilizes phoneme-based localization methods to identify temporal indices of key motion, which correspond to significant motion snapshots aligned with phoneme changes in the audio. Once the key motion indexes are determined, a key motion decoder interprets associated 3D facial meshes from corresponding audio features. This highlights those distinct facial expressions and facilitates lip-speech synchronization. The cross-modal motion completion module expands non-continuous key motions into a full sequence of continuous face motions using audio features as guidance. This process enhances audio-mesh alignment and improves the temporal smoothness of output facial mesh sequences. The contributions of our work are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 335, + 479, + 451 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 335, + 479, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 479, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 479, + 369 + ], + "type": "text", + "content": "- We propose a progressive learning mechanism to generate speech-driven 3D talking faces. It uses linguistic priors to initially generate key motions, and then interpolate key motions into complete motions via data-driven priors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 370, + 479, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 479, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 479, + 403 + ], + "type": "text", + "content": "- We propose the use of phoneme-based localization methods to capture key facial motions. It effectively captures significant expression transitions aligned with phoneme changes in audio, improving lip-speech synchronization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 404, + 479, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 479, + 451 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 479, + 451 + ], + "type": "text", + "content": "- We design a cross-modal facial motion completion module to produce full sequences of 3D talking faces using synthesized key motions and audio features. It further enhances lip-speech synchronization accuracy while facilitating temporal coherence in facial motions." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 456, + 483, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 483, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 483, + 540 + ], + "type": "text", + "content": "Extensive experimental comparisons on the BIWI [13] and VOCASET [8] datasets demonstrate that our method outperforms existing state-of-the-art approaches in more accurate and realistic talking face generation. Detailed ablation studies confirm the effectiveness of our proposed key motion capture technique. Additionally, consistent improvements in results by combining our proposed scheme with existing methods validate the efficacy of our design. Our code and weights will be at the project website: https://github.com/ffxzh/KMTalk." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 554, + 237, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 237, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 237, + 567 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 575, + 479, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 575, + 479, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 575, + 479, + 612 + ], + "type": "text", + "content": "While existing research [1,5,6,10,15,18,19,23,25,35,38,42,47,50,55,59,61] focuses on 2D talking heads, we focus on audio-driven 3D facial animations in this work, which can be roughly categorized into linguistics-based and data-driven methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 625, + 289, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 625, + 289, + 637 + ], + "spans": [ + { + "bbox": [ + 132, + 625, + 289, + 637 + ], + "type": "text", + "content": "2.1 Linguistic-based Methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Linguistic-based methods [3, 7, 11, 14, 22, 29, 30, 46, 54] establish a set of intricate phoneme-to-viseme mapping rules for animating the mouth. For example, the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "text", + "content": "KMTalk" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "dynamic viseme model proposed by Taylor et al. [46] exploits the one-to-many mapping of phonemes to lip motions. JALI [11] considers the many-to-many mapping between phonemes and visemes. More recently, Bao et al. [3] introduced a novel parameterized viseme fitting algorithm that extracts viseme parameters from speech videos using phonemic priors. Leveraging linguistic priors, these methods indicate the articulation process by providing animators with explicit control over animation, thus boosting their performance in lip-speech synchronization. However, these approaches primarily focus on animating the lip region, lacking a comprehensive strategy for animating the entire face. In our work, we leverage linguistic priors to detect key frames with significant expression changes from audio in an analytical manner, without the need for supervised training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 262, + 269, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 262, + 269, + 273 + ], + "spans": [ + { + "bbox": [ + 131, + 262, + 269, + 273 + ], + "type": "text", + "content": "2.2 Data-driven Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 277, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 579 + ], + "type": "text", + "content": "With the development of deep learning technology and the availability of high-quality datasets, data-driven methods [4, 8, 12, 16, 20, 26, 34, 36, 37, 41, 44, 45, 49, 52, 53, 60] were proposed to synthesize entire 3D facial animation. Some methods [8, 12, 20, 40, 52] attempt to establish a direct audio-to-visual mapping through regression. Person-specific approaches [20, 40] can usually obtain plausible facial motions because of the relatively consistent talking style. VOCA [8] incorporates a robust audio feature extraction model capable of capturing various speaking styles, which can generate realistic speaker-independent animation and shows its wide applicability. MeshTalk [41] constructed a categorical latent space to adaptively generate motions based on the separated audio-correlated and audiouncorrelated facial information. FaceFormer [12] introduced two biased attention mechanisms and integrated the self-supervised pre-trained speech representations for the ill-posed and data scarcity issues. CodeTalker [53] proposed the discrete motion prior which regards the cross-modal mapping as a code query task in a finite proxy space of the learned codebook. SelfTalk [36] proposed a self-supervised approach to construct a lip-reading interpreter and speech recognizer to enhance the comprehensibility of generated lip movements. While data-driven approaches have shown impressive performance, accurately learning cross-modal audio-visual mappings remains challenging due to inherent uncertainties. These methods often regress the entire sequence, leading to over-smoothing and a lack of detailed facial dynamics. In contrast, our approach employs a coarse-to-fine learning mechanism that separates the problem into key motion capture and motion completion stages. This approach effectively mitigates cross-modal mapping uncertainties and reduces learning complexity, resulting in more precise and dynamic facial animation synthesis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 591, + 203, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 591, + 203, + 604 + ], + "spans": [ + { + "bbox": [ + 131, + 591, + 203, + 604 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 613, + 209, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 613, + 209, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 613, + 209, + 624 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 628, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 628, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 481, + 666 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 628, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 628, + 481, + 666 + ], + "type": "text", + "content": " represents the raw audio input and " + }, + { + "bbox": [ + 130, + 628, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}} = (\\hat{\\mathbf{y}}_1,\\dots,\\hat{\\mathbf{y}}_N)\\in \\mathbb{R}^{N\\times V\\times 3}" + }, + { + "bbox": [ + 130, + 628, + 481, + 666 + ], + "type": "text", + "content": " denotes the corresponding ground-truth sequence of facial movement over a neutral template, where " + }, + { + "bbox": [ + 130, + 628, + 481, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 628, + 481, + 666 + ], + "type": "text", + "content": " indicates the number of visual frames and " + }, + { + "bbox": [ + 130, + 628, + 481, + 666 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 628, + 481, + 666 + ], + "type": "text", + "content": " denotes the number" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Z. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 478, + 213 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 478, + 213 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 478, + 213 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 478, + 213 + ], + "type": "image", + "image_path": "732b3f9865d772653c6d7af7d745fbc23d87894c3accf04f47592faefe020760.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 214, + 395, + 223 + ], + "lines": [ + { + "bbox": [ + 217, + 214, + 395, + 223 + ], + "spans": [ + { + "bbox": [ + 217, + 214, + 395, + 223 + ], + "type": "text", + "content": "(a) The overview pipeline of our proposed KMTalk." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 224, + 479, + 302 + ], + "blocks": [ + { + "bbox": [ + 133, + 224, + 479, + 302 + ], + "lines": [ + { + "bbox": [ + 133, + 224, + 479, + 302 + ], + "spans": [ + { + "bbox": [ + 133, + 224, + 479, + 302 + ], + "type": "image", + "image_path": "575248935e24c38bbf8194ce56065892ff51a23bdb35467e4b9c07603f61ebce.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 304, + 390, + 312 + ], + "lines": [ + { + "bbox": [ + 223, + 304, + 390, + 312 + ], + "spans": [ + { + "bbox": [ + 223, + 304, + 390, + 312 + ], + "type": "text", + "content": "(b) The details of two key modules in KMTalk." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "text", + "content": "Fig. 2: Fig. 2a illustrates the overview pipeline of our proposed KMTalk. Initially, the Audio Encoder takes the input raw audio " + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "text", + "content": " and encodes it into audio features " + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "text", + "content": ". Subsequently, in the LKMA module, key motions " + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "text", + "content": " are generated from the audio " + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "text", + "content": ". Finally, the CMC module reintroduces audio features " + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "text", + "content": " to extend these key motions " + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "text", + "content": " into a full sequence " + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "text", + "content": ". Fig. 2b presents the details of two key modules in KMTalk. In the Linguistic-based Key Motion Acquisition, a Phoneme-based Localization Method is used to identify key motion indices " + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "text", + "content": " from raw audio " + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "text", + "content": ". Based on audio features " + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "text", + "content": ", the Key Motion-focused Decoder generates key motions " + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "text", + "content": ". In the Cross-modal Motion Completion, the Motion Flow Encoder processes " + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "text", + "content": ", producing motion flow features " + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "text", + "content": ". Then, with the dynamic fusion weight " + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "text", + "content": ", the Multimodal-Guided Decoder combines " + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "text", + "content": " to decode the final motion sequence " + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 130, + 322, + 482, + 443 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 464, + 481, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 481, + 500 + ], + "type": "text", + "content": "of vertices in the facial mesh. The objective is to synthesize " + }, + { + "bbox": [ + 130, + 464, + 481, + 500 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} = (\\mathbf{y}_1, \\dots, \\mathbf{y}_N)" + }, + { + "bbox": [ + 130, + 464, + 481, + 500 + ], + "type": "text", + "content": " that is similar to " + }, + { + "bbox": [ + 130, + 464, + 481, + 500 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}" + }, + { + "bbox": [ + 130, + 464, + 481, + 500 + ], + "type": "text", + "content": ", driven by the raw audio " + }, + { + "bbox": [ + 130, + 464, + 481, + 500 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 464, + 481, + 500 + ], + "type": "text", + "content": ". The generated sequence should ensure lip synchronization with the audio while exhibiting natural facial movements." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 500, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 482, + 608 + ], + "type": "text", + "content": "Due to the domain gap between modalities and the ill-posed nature of directly translating audio to facial movement sequences, it is a challenging task that often results in over-smooth or poorly synchronized lip movements. To address these issues, this paper introduces a coarse-to-fine approach with key motion embedding, integrating both linguistic and data-driven priors. The overview pipeline is presented in Fig. 2a. In the LKMA module, linguistic priors are introduced to locate and generate higher-quality key motions (see Sec. 3.2), followed by the CMC module, where these key motions are fleshed out into a complete sequence of facial motions (see Sec. 3.3)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 624, + 367, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 367, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 367, + 635 + ], + "type": "text", + "content": "3.2 Linguistic-based Key Motion Acquisition" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "In the realm of audio-driven 3D facial animation, it presents significant challenges to precisely define which frames constitute key motions. An alternative" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "text", + "content": "KMTalk" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "and simple solution is to use uniform or random sampling to determine the positions of these key motions. Although these approaches can boost performance to a certain degree due to the reduced learning complexity (refer to Sec. 4.2 for ablation studies), they fail to utilize the correlation between audio content and facial movements. However, we can leverage linguistic priors to capture pronounced articulatory actions, which are identifiable at phoneme boundaries. This approach circumvents the issue of over-smooth in the output sequence, thereby enhancing the overall quality of the results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "text", + "content": "As shown in the left of Fig. 2b, the Linguistic-based Key Motion Acquisition (LKMA) module receives as inputs the raw audio " + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "text", + "content": " and the audio features " + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{A} = (\\mathbf{a}_1,\\dots,\\mathbf{a}_N)\\in \\mathbb{R}^{N\\times d}" + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "text", + "content": ", where the audio features are derived from the Audio Encoder that utilizes the wav2vec 2.0 pre-trained model [2], processing the raw audio " + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "text", + "content": " as its input. Then it takes raw audio " + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "text", + "content": " as input to produce the key motion indices " + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{i_1,\\dots,i_m\\}" + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "inline_equation", + "content": "i_j\\in \\{1,\\dots,N\\},\\forall i_j\\in \\mathcal{I}" + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "text", + "content": ", through the proposed Phoneme-based Localization method. Subsequently, the Key Motion-focused Decoder utilizes the audio features " + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "text", + "content": " and the key motion indices " + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "text", + "content": " to generate key motions " + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{K} = (\\mathbf{k}_{i_1},\\dots,\\mathbf{k}_{i_m})\\in \\mathbb{R}^{m\\times V\\times 3}" + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "text", + "content": " frames of facial movement which are located on the key motion positions, where " + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_{i_j}\\simeq \\hat{\\mathbf{y}}_{i_j},\\forall i_j\\in \\mathcal{I}" + }, + { + "bbox": [ + 130, + 213, + 483, + 345 + ], + "type": "text", + "content": ". The process of the LKMA module is expressed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 260, + 355, + 480, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 355, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 260, + 355, + 480, + 366 + ], + "type": "interline_equation", + "content": "\\mathcal {I}, \\mathbf {K} = \\operatorname {L K M A} (\\mathbf {x}, \\mathbf {A}). \\tag {1}", + "image_path": "533afb50772b2d8e5af2a8ccfd09c8f601abd2e0ba72d8dcefbbb5142998778b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 377, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 483 + ], + "type": "text", + "content": "Phoneme-based Localization. At phoneme boundaries, a noticeable offset is observed in the articulator movement, with visualization results available in supplementary materials. Furthermore, the phoneme boundary effects underscore the ease with which the boundaries of phonemes can be perceived [17]. Both experimental and theoretical analyses have demonstrated a distinct position-mapping relationship between the phoneme boundaries in the audio and the significant elements in the motion sequence, specifically the key motions. The mapping relationship can be harnessed to facilitate the initial alignment between the audio and visual modalities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 485, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 557 + ], + "type": "text", + "content": "Specifically, an Automatic Speech Recognition model [28] is first utilized to obtain the text content from the raw audio " + }, + { + "bbox": [ + 130, + 485, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 485, + 482, + 557 + ], + "type": "text", + "content": ". Then, a Montreal Forced Aligner [31] is adopted to align the audio and the text, producing the start and the end timestamps for each phoneme. Finally, the indices of these motion frames corresponding to the timestamps are regarded as key motion indices " + }, + { + "bbox": [ + 130, + 485, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 485, + 482, + 557 + ], + "type": "text", + "content": ", and the corresponding motions compose the key motions " + }, + { + "bbox": [ + 130, + 485, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 485, + 482, + 557 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 617 + ], + "type": "text", + "content": "Key Motion-focused Decoder. It is utilized to synthesize key motions " + }, + { + "bbox": [ + 130, + 557, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 557, + 482, + 617 + ], + "type": "text", + "content": " of superior quality. Initially, employing " + }, + { + "bbox": [ + 130, + 557, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 557, + 482, + 617 + ], + "type": "text", + "content": " as indices, we extract the corresponding aligned audio features " + }, + { + "bbox": [ + 130, + 557, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{A_k} = (\\mathbf{a}_{i_1},\\dots,\\mathbf{a}_{i_m})\\in \\mathbb{R}^{m\\times d}" + }, + { + "bbox": [ + 130, + 557, + 482, + 617 + ], + "type": "text", + "content": " from the comprehensive audio features " + }, + { + "bbox": [ + 130, + 557, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 557, + 482, + 617 + ], + "type": "text", + "content": ". Subsequently, it adopts a modified transformer-based architecture, processes " + }, + { + "bbox": [ + 130, + 557, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{A_k}" + }, + { + "bbox": [ + 130, + 557, + 482, + 617 + ], + "type": "text", + "content": " to generate the key motions " + }, + { + "bbox": [ + 130, + 557, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 557, + 482, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Loss Function. Intuitively, a straightforward approach to optimize the key motions " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " involves utilizing " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " to index " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}_{\\mathbf{k}} = (\\hat{\\mathbf{y}}_{i_1},\\dots ,\\hat{\\mathbf{y}}_{i_m})" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", which serves as the supervision for the training process. However, key motions typically occupy non-adjacent positions within the entire sequence. Hence, given the lack" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Z. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "of inter-frame contextual information, attempting direct frame-by-frame regression of " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " towards " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}_{\\mathbf{k}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " may fall short in achieving accurate facial expressions, as well as producing smooth and realistic animation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 262 + ], + "type": "text", + "content": "To address this limitation, we adopt a pseudo-complete sequence training method that utilizes the ground-truth frame labels " + }, + { + "bbox": [ + 130, + 152, + 480, + 262 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}" + }, + { + "bbox": [ + 130, + 152, + 480, + 262 + ], + "type": "text", + "content": " at the non-key indices " + }, + { + "bbox": [ + 130, + 152, + 480, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{I}' = \\{1,\\dots,N\\} \\setminus \\mathcal{I}" + }, + { + "bbox": [ + 130, + 152, + 480, + 262 + ], + "type": "text", + "content": " and the generated key motions " + }, + { + "bbox": [ + 130, + 152, + 480, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 152, + 480, + 262 + ], + "type": "text", + "content": " at the key indices " + }, + { + "bbox": [ + 130, + 152, + 480, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 152, + 480, + 262 + ], + "type": "text", + "content": " to form a predicted pseudo-complete sequence " + }, + { + "bbox": [ + 130, + 152, + 480, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{\\mathbf{p}} \\in \\mathbb{R}^{N \\times V \\times 3}" + }, + { + "bbox": [ + 130, + 152, + 480, + 262 + ], + "type": "text", + "content": ". Then, the model is trained by minimizing the loss between the pseudo-complete sequence " + }, + { + "bbox": [ + 130, + 152, + 480, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{\\mathbf{p}}" + }, + { + "bbox": [ + 130, + 152, + 480, + 262 + ], + "type": "text", + "content": " and the ground-truth sequence " + }, + { + "bbox": [ + 130, + 152, + 480, + 262 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}" + }, + { + "bbox": [ + 130, + 152, + 480, + 262 + ], + "type": "text", + "content": ". This enables the model to capture subtle changes between key motions and adjacent ground-truth frames, thereby mitigating inter-frame jitter and achieving more accurate regression of facial expressions. Following the SelfTalk [36], the loss function is formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 207, + 272, + 480, + 284 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 272, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 207, + 272, + 480, + 284 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {L K M A} = \\lambda_ {1} \\mathcal {L} _ {r e c} + \\lambda_ {2} \\mathcal {L} _ {v e l} + \\lambda_ {3} \\mathcal {L} _ {l a t} + \\lambda_ {4} \\mathcal {L} _ {c t c}, \\tag {2}", + "image_path": "9621d49d647fbd4d73e2e8184679f88c1796cd1f2358ebf7b5992dc34ce48d06.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 1000.0, \\lambda_{2} = 1000.0, \\lambda_{3} = 0.001" + }, + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\lambda_{4} = 0.0001" + }, + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "text", + "content": " in all of our experiments. " + }, + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{vel}" + }, + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{lat}" + }, + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "text", + "content": " are measured by mean square error, while " + }, + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ctc}" + }, + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "text", + "content": " is quantified by CTC Loss. The reconstruction loss " + }, + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "text", + "content": " quantifies the discrepancy between the predicted and the ground-truth facial movements. The velocity loss " + }, + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{vel}" + }, + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "text", + "content": " reduces frame jitter, ensuring smooth and natural lip movements. The latent consistency loss " + }, + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{lat}" + }, + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "text", + "content": " assesses the variance between latent features extracted from both the audio and lip shape encoders, aiming to align the learned audio and lip features. Lastly, the text consistency loss " + }, + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ctc}" + }, + { + "bbox": [ + 130, + 293, + 480, + 413 + ], + "type": "text", + "content": " evaluates the difference between the lip-reading decoder's output and the original text, ensuring the intelligibility of the lip-reading results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 432, + 326, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 432, + 326, + 444 + ], + "spans": [ + { + "bbox": [ + 131, + 432, + 326, + 444 + ], + "type": "text", + "content": "3.3 Cross-modal Motion Completion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 453, + 479, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 479, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 479, + 561 + ], + "type": "text", + "content": "A straightforward method to obtain a complete talking face sequence is to directly integrate key motions into the entire face movements. However, it is important to note that while key motions capture essential facial dynamics, they may not encompass all details of non-key motions. This direct integration could result in mismatches between augmented non-key motions and their corresponding audio segments (see Sec. 4.2 for ablation studies). To mitigate this issue, we introduce a Cross-modal Motion Completion (CMC) module that jointly combines the audio features " + }, + { + "bbox": [ + 130, + 453, + 479, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 453, + 479, + 561 + ], + "type": "text", + "content": ", key motions " + }, + { + "bbox": [ + 130, + 453, + 479, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 453, + 479, + 561 + ], + "type": "text", + "content": ", and key motion indices " + }, + { + "bbox": [ + 130, + 453, + 479, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 453, + 479, + 561 + ], + "type": "text", + "content": " to generate a complete sequence of 3D facial meshes " + }, + { + "bbox": [ + 130, + 453, + 479, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 130, + 453, + 479, + 561 + ], + "type": "text", + "content": ". The process can be formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 262, + 571, + 480, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 571, + 480, + 583 + ], + "spans": [ + { + "bbox": [ + 262, + 571, + 480, + 583 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} = \\operatorname {C M C} (\\mathbf {A}, \\mathbf {K}, \\mathcal {I}). \\tag {3}", + "image_path": "25805a4a2abf6079ef588a9dd1130ea91499d3527cc5d3074a18583a8e85163b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 441, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 441, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 441, + 605 + ], + "type": "text", + "content": "The details of the CMC module are illustrated on the right of Fig. 2b." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": "Motion Flow Encoder. Key motions serve as a kinematic prior for the remaining frames, offering valuable insights into facial dynamics. To effectively capture the motion flow information provided by key motions, we draw inspiration from some manifold methods [32] to acquire motion flow features " + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\Phi \\in \\mathbb{R}^{N\\times d}" + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": " from the key motions " + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": ". Specifically, we first encode the key motions " + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": " into key motion" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "KMTalk" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "content": "context tokens " + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathbf{k}} \\in \\mathbb{R}^{m \\times d}" + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "content": " by multiple transformer-encoder layers. At the same time, the indices of non-key motions " + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{T}'" + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "content": " are encoded into positional encodings by a sinusoidal position embedding layer, representing the non-key frame positions. Then, we adopt the cross-attention layers to extract the intermediate tokens " + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{non - key}} \\in \\mathbb{R}^{(N - m) \\times d}" + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "content": ", with key and value from linear transformations of " + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathbf{k}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "content": " and the query is the positional encodings of non-key frames indices. Above all, the implicit motion manifold proposed in CITL [32] is utilized to arrange " + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathbf{k}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{non - key}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "content": " based on their indices, followed by a 1D convolution for fusion, ultimately obtaining the motion flow features " + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 344 + ], + "type": "text", + "content": "Multimodal-Guided Decoder. The non-key motions' features in the complete motion sequence feature estimation are derived from the global context interpolation of key motions, which has a certain degree of information loss due to the audio feature selection process in the Key Motion-focused Decoder (Sec. 3.2). Hence, we have devised a multimodal-guided decoding approach that incorporates audio modalities to furnish comprehensive information across the entire temporal scale, alongside motion flow to offer facial motion priors. These elements serve to guide and constrain the decoding process, thereby facilitating the precise generation of the motion sequence. Technically, we simply employ the gated mechanism [9,57,58] for the modality fusion, which can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 268, + 356, + 481, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 356, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 268, + 356, + 481, + 369 + ], + "type": "interline_equation", + "content": "\\mathbf {G} = \\sigma ([ \\mathbf {A}, \\boldsymbol {\\Phi} ] \\mathbf {W}), \\tag {4}", + "image_path": "e36c15d6501df165696044e19b4603b61d715ab30ebe278c5cd86c78d61c7fe7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 246, + 386, + 480, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 386, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 246, + 386, + 480, + 399 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} = \\mathbf {G} \\odot \\mathbf {A} + (1 - \\mathbf {G}) \\odot \\boldsymbol {\\Phi}, \\tag {5}", + "image_path": "fb3597d1c2a10f7cf669bffb2515aa4997a04b7361e74a18f856b5b0d56146d2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "content": " represents the element-wise multiplication operation, " + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "inline_equation", + "content": "[\\cdot, \\cdot]" + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "content": " denotes the concatenation operation, " + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{2d \\times d}" + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "content": " is a parameter and " + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "content": " is a sigmoid function, while " + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{G} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "content": " dynamically selects features from the audio features " + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "content": " and the motion flow features " + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "content": ". Then, the Motion Decoder, a transformer-based structural model, is employed to transform the fusion features " + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "content": " into the complete 3D facial movement sequence " + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}" + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 475, + 482, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 500 + ], + "type": "text", + "content": "Loss Function. We train the CMC module utilizing the reconstruction loss and velocity loss. The total loss function is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 260, + 512, + 480, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 512, + 480, + 524 + ], + "spans": [ + { + "bbox": [ + 260, + 512, + 480, + 524 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C M C} = \\mathcal {L} _ {r e c} + \\mathcal {L} _ {v e l}. \\tag {6}", + "image_path": "7e01659847c9e51cd8298aa663185fd22f8dd3030056a5a234fe642f3d0c6d02.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 544, + 224, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 544, + 224, + 558 + ], + "spans": [ + { + "bbox": [ + 131, + 544, + 224, + 558 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Dataset. BIWI [13] consists of 40 paired audio-visual sentences from 14 subjects. The 3D facial geometries, consisting of 23370 vertices, were captured at a frame rate of 25fps, and the average duration of each sequence was 4.67 seconds. We adopt the same evaluation protocol as FaceFormer [12] on the BIWI dataset. Specifically, the training set (BIWI-Train) comprises 190 sentences, while the validation set (BIWI-Val) encompasses 24 sentences. The dataset is divided into two testing sets: BIWI-Test-A, which comprises 24 sentences articulated by six subjects observed during training, and BIWI-Test-B, which consists of 32 sentences" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Z. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 156, + 481, + 247 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 147 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 147 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 147 + ], + "type": "text", + "content": "Table 1: Quantitative comparisons on the BIWI-Test-A and VOCA-Test datasets. The results of Lip-Vertex Error (LVE) and the upper-Face Dynamics Deviation (FDD) are reported. For both metrics, the lower the better." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 156, + 481, + 247 + ], + "lines": [ + { + "bbox": [ + 135, + 156, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 135, + 156, + 481, + 247 + ], + "type": "table", + "html": "
MethodsBIWI-Test-AVOCA-Test
LVE↓×10-4mmFDD↓×10-5mmLVE↓×10-5mmFDD↓×10-7mm
VOCA [8]6.55638.18164.92454.8447
MeshTalk [41]5.91815.10254.54415.2062
FaceFormer [12]5.30774.64084.10904.6675
CodeTalker [53]4.79144.11703.94454.5422
SelfTalk [36]4.24853.57613.22384.0912
KMTalk (Ours)3.96542.54462.26394.0594
", + "image_path": "e9a7c9d2061499b48cb00f06f7e6658c1dad926205a959da47760c1440ed76f8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 270, + 480, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 480, + 341 + ], + "type": "text", + "content": "uttered by eight unseen subjects. VOCASET [8] consists of 480 paired audiovisual sequences from 12 subjects. Each sequence is recorded at a frame rate of 60fps and ranges in duration from 3 to 4 seconds. The 3D face mesh for each sequence consists of 5023 vertices. To ensure a fair comparison, we used identical training (VOCA-Train), validation (VOCA-Val), and testing (VOCA-Test) partitions as methods [12,36,53]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 481, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 481, + 437 + ], + "type": "text", + "content": "Baselines. We compare against current state-of-the-arts method, including VOCA [8], MeshTalk [41]. FaceFormer [12], CodeTalker [53], and SelfTalk [36]. Faceformer [12] employs a transformer-based model to incorporate long-term audio context and synthesizes sequential motions in an autoregressive manner. CodeTalker [53] introduces discrete motion priors to enable self-reconstruction of real facial movements, mitigating the issue of excessive smoothing in facial motion. SelfTalk [36] designs a learning-based recognizer to minimize the domain gap between diverse modalities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 481, + 581 + ], + "type": "text", + "content": "Evaluation Metrics. Following CodeTalker [53] and SelfTalk [36], we adopt two metrics for the quantitative evaluation of speech-driven facial animation: lip vertex error (LVE) to measure lip synchronization and upper-face dynamics deviation (FDD) to assess the overall facial dynamics. The LVE for each frame is defined as the maximal L2 error among all lip vertices for each frame and takes the average over all frames. This L2 error is computed by comparing the predictions with the processed 3D face geometry data. FDD is introduced to quantify the variation in facial dynamics between a synthetic motion sequence and the reference sequence. The implementation of FDD is to calculate the difference between the variances of vertex offsets in the upper-face region and the variances of ground truth vertex offsets. In addition, we visualize the prediction results for qualitative evaluation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "Implementation Details. For a fair comparison, KMTalk operates at a frame rate of 30 fps on VOCASET and 25 fps on BIWI, following the setting of previous methods [12, 36, 53]. Also, it can naturally adapt to a higher frame rate, as shown in the supplementary materials. In the LKMA module, we first employ the Phoneme-based Localization method to process the raw audio and obtain key motion indices for data preprocessing, which costs less than 10 minutes on two datasets [8, 13]. Secondly, we train the Key Motion Decoder on a single NVIDIA" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "text", + "content": "KMTalk" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 112, + 474, + 358 + ], + "blocks": [ + { + "bbox": [ + 138, + 112, + 474, + 358 + ], + "lines": [ + { + "bbox": [ + 138, + 112, + 474, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 112, + 474, + 358 + ], + "type": "image", + "image_path": "626f767d17967288bffb105075a792a0cb286ae1b11d81e744077f7135ada470.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 367, + 482, + 467 + ], + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 467 + ], + "type": "text", + "content": "Fig. 3: Qualitative comparisons on VOCA-Test (left) and BIWI-Test-B (right). We provide visual comparisons of facial animations synchronized with six syllables extracted from the test speech sequences. The 1st, 3rd, and 5th rows display synthesized meshes and their corresponding ground-truths, while the 2nd, 4th, and 6th rows visualize the L2 loss for individual frames. Our method demonstrates more precise mouth movement on syllables like /æ/ that require a wide-open mouth. For syllables that start with a closed mouth and then slightly open, such as /bɪ/, our KMTalk generates more synchronized motion sequences visually. The last row visualizes the mean square errors of different methods across all sentences in the test set for a specific subject." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 491, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 563 + ], + "type": "text", + "content": "RTX 3090 for 200 epochs (about 2 hours) using the Adam optimizer [21]. The learning rate is initialized as " + }, + { + "bbox": [ + 130, + 491, + 482, + 563 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 491, + 482, + 563 + ], + "type": "text", + "content": ", and the mini-batch size is set to 1. In the CMC module, we train for 200 epochs (approximately 2 hours) with the same training settings as the Key Motion Decoder. It is noteworthy that, since the training of the two modules is independent of each other, we can train both modules concurrently to enhance training efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 583, + 395, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 583, + 395, + 596 + ], + "spans": [ + { + "bbox": [ + 131, + 583, + 395, + 596 + ], + "type": "text", + "content": "4.1 Comparisons against State-of-the-art Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Quantitative Comparisons. We computed the lip vertex error (LVE) and facial dynamics deviation (FDD) for all sequences within the BIWI-Test-A and VOCA-Test datasets. According to Table 1, our proposed KMTalk demonstrated lower errors compared to the alternative methods examined. Notably, the lip vertex error for our method on the VOCA-Test dataset is " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " lower than the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "Z. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": "recently introduced SelfTalk [36], and the FDD is " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": " lower than SelfTalk on the BIWI-Test-A dataset, providing compelling evidence for the advantages of our proposed KMTalk. This indicates that our approach is more effective in achieving audio-visual alignment, thereby leading to improved lip synchronization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 319 + ], + "type": "text", + "content": "Qualitative Comparisons. In Fig. 3, we visualize the output facial meshes from different methods as well as ground truths for reference. Additionally, we display error maps calculated from the vertex L2 loss between the generated and ground truth meshes. It is evident that our method consistently yields lower errors across different speech sequences, demonstrating its ability to generate more accurate facial animation sequences. Notably, for representative syllables (e.g. " + }, + { + "bbox": [ + 130, + 164, + 480, + 319 + ], + "type": "inline_equation", + "content": "/a\\mathrm{e}/" + }, + { + "bbox": [ + 130, + 164, + 480, + 319 + ], + "type": "text", + "content": "), KMTalk closely approximates ground truth, excelling in synthesizing accurate lip movements for syllables requiring significant mouth opening. Additionally, for syllables starting with mouth closure followed by a slight opening (e.g., " + }, + { + "bbox": [ + 130, + 164, + 480, + 319 + ], + "type": "inline_equation", + "content": "/b\\mathrm{i}/" + }, + { + "bbox": [ + 130, + 164, + 480, + 319 + ], + "type": "text", + "content": "), KMTalk produces more natural and synchronized motions. We recommend that readers watch the supplementary video for more detailed comparisons. It showcases KMTalk's capability to generate coherent, realistic animations with precise lip synchronization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 319, + 481, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 535 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 535 + ], + "type": "text", + "content": "User Studies. A user study stands as a dependable evaluation method in the context of 3D talking faces. Following the strategy of Faceformer [12], we conduct pairwise comparisons between our method and baselines [12,36,53], as well as ground truths. This study encompassed the assessment of two key metrics: perceptual lip synchronization and facial realism. Participants were presented with side-by-side comparisons and were tasked with selecting the better facial animation based on their personal preferences. We computed the ratio of user preferences as a measurement of satisfaction evaluation on BIWI-Test-B and VOCA-Test. We randomly sampled 30 examples from each test set and compared the performance of KMTalk with four aforementioned settings on each sample. Therefore, we constructed a total of 240 different video pairs and randomly selected 24 video pairs for the two metrics assessments for each participant. Our user study involved 30 participants with a strong capability for audio-visual perception, resulting in 720 effective evaluation entries. As demonstrated in Table 2, our approach indicates superior perceptual lip synchronization and facial realism. For instance, a noteworthy " + }, + { + "bbox": [ + 130, + 319, + 481, + 535 + ], + "type": "inline_equation", + "content": "60.0\\%" + }, + { + "bbox": [ + 130, + 319, + 481, + 535 + ], + "type": "text", + "content": " of users favored our lip synchronization method on BIWI-Test-B in comparison to SelfTalk [36]. Overall, it shows that KMTalk can generate more favorable facial animations from speeches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 564, + 246, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 564, + 246, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 564, + 246, + 575 + ], + "type": "text", + "content": "4.2 Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "type": "text", + "content": "In this section, we perform ablation studies on to evaluate the influence of different components within our proposed KMTalk framework on the quality of the generated 3D talking faces. The quantitative results on BIWI are in Table 3, and the qualitative results are in Fig. 4. In addition, the results of the ablation study on VOCA-Test can be found in the supplementary material. In Table 4, we further investigate the robustness of our approach to different Phoneme-based localization methods and the possible errors during phoneme extraction." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "KMTalk" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 160, + 135, + 461, + 269 + ], + "blocks": [ + { + "bbox": [ + 178, + 114, + 434, + 125 + ], + "lines": [ + { + "bbox": [ + 178, + 114, + 434, + 125 + ], + "spans": [ + { + "bbox": [ + 178, + 114, + 434, + 125 + ], + "type": "text", + "content": "Table 2: User study results on BIWI-Test-B and VOCA-Test." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 160, + 135, + 461, + 269 + ], + "lines": [ + { + "bbox": [ + 160, + 135, + 461, + 269 + ], + "spans": [ + { + "bbox": [ + 160, + 135, + 461, + 269 + ], + "type": "table", + "html": "
MethodMetricBIWI-Test-BVOCA-Test
competitorourscompetitorours
Ours vs. FaceFormerLip Sync24.4%75.6%26.7%73.3%
Realism25.6%74.4%28.9%71.1%
Ours vs. CodeTalkerLip Sync31.1%68.9%37.8%62.2%
Realism27.8%72.2%35.6%64.4%
Ours vs. SelfTalkLip Sync40.0%60.0%43.3%56.7%
Realism38.9%61.1%41.1%58.9%
Ours vs. GTLip Sync54.4%45.6%56.7%43.3%
Realism52.2%47.8%56.7%43.3%
", + "image_path": "24f418aeb43f306d67f11794a19ceeaaee8fd76ace2ac2de7933fe4320aed7c3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 167, + 301, + 447, + 376 + ], + "blocks": [ + { + "bbox": [ + 178, + 281, + 434, + 293 + ], + "lines": [ + { + "bbox": [ + 178, + 281, + 434, + 293 + ], + "spans": [ + { + "bbox": [ + 178, + 281, + 434, + 293 + ], + "type": "text", + "content": "Table 3: Ablation study for our components on BIWI-Test-A." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 301, + 447, + 376 + ], + "lines": [ + { + "bbox": [ + 167, + 301, + 447, + 376 + ], + "spans": [ + { + "bbox": [ + 167, + 301, + 447, + 376 + ], + "type": "table", + "html": "
Phoneme-based Localization MethodKey Motion-focused Audio in CMCLVE↓FDD↓
Decoder✓4.24853.5761
---4.16482.8713
-✓✓4.13812.9546
✓-✓4.88593.2780
✓✓-3.96542.5446
", + "image_path": "cee9155627e2d006166bb9239cd0ccc495a5bc84db907c904b689b40424259ab.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 401, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 481, + 544 + ], + "type": "text", + "content": "What's the effect of the Phoneme-based localization method for key motion capture? The Phoneme-based Localization Method enables us to identify key frames of speech with notable facial expression transitions. We can replace it with uniform sampling, neglecting the crucial content information of the audio. Specifically, we experimented using uniform sampling at a rate of " + }, + { + "bbox": [ + 130, + 401, + 481, + 544 + ], + "type": "inline_equation", + "content": "33\\%" + }, + { + "bbox": [ + 130, + 401, + 481, + 544 + ], + "type": "text", + "content": " to closely align with the number of key motions. For further comparisons under different numbers of sampled elements, please refer to the supplementary material. In Table 3, we observe degradation in all metrics with uniform sampling. This underscores the importance of accurate key motion capture in speech-driven talking face generation. Additionally, it showcases that the linguistic-based key motion capture is better equipped to mitigate audio-visual uncertainty and recover more precise facial motions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "What's the effect of the Key Motion Decoder? Our method employs a specialized key motion decoder to generate facial meshes based on keyframe indices obtained from phoneme-based localization methods. An alternative approach to generating key motions is to select corresponding facial meshes from a complete motion sequence produced by an existing method such as SelfTalk [36]. Table 3 demonstrates that both metrics deteriorated but still outperformed the state-of-the-art method SelfTalk. This indicates that the Key Motion Decoder can enhance the quality of key motion generation, resulting in more plausible facial animations. It also suggests that even if the captured key motions are not accurate enough, the CMC module can further refine output full motions." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "Z. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 118, + 479, + 312 + ], + "blocks": [ + { + "bbox": [ + 134, + 118, + 479, + 312 + ], + "lines": [ + { + "bbox": [ + 134, + 118, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 134, + 118, + 479, + 312 + ], + "type": "image", + "image_path": "47975d6ec9b6d62e30e566b978fbc70e53c4b4ea2aa0ba8b68ef5a5206e109cf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 320, + 482, + 376 + ], + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 376 + ], + "type": "text", + "content": "Fig. 4: Qualitative ablation studies on the input speech \"specifically\". For each method variant, we removed one of three modules: PLM (Phoneme-based Localization Method), KMD (Key Motion-focused Decoder), and AG (Audio Guidance in CMC). Error maps between generated and the ground-truth mesh sequence were visualized. Our final model yielded the best results, showcasing the effectiveness of each module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 409, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 541 + ], + "type": "text", + "content": "What's the effect of audio guidance in the Cross-modal Motion Completion? The Cross-modal Motion Completion leverages audio features to guide the completion of the full motion sequence. To assess its usefulness, we implement a method variant that removes the audio feature guidance. Table 3 demonstrates a notable degradation in both metrics, particularly with a " + }, + { + "bbox": [ + 130, + 409, + 482, + 541 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 130, + 409, + 482, + 541 + ], + "type": "text", + "content": " increase in Lip Vertex Error (LVE) and a " + }, + { + "bbox": [ + 130, + 409, + 482, + 541 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 130, + 409, + 482, + 541 + ], + "type": "text", + "content": " increase in upper-face dynamics deviation (FDD). This suggests that audio information plays a crucial role in refining fine-grained lip movements and enhancing audio-visual consistency and temporal smoothness. Despite the degradation, the FDD metric can still outperform state-of-the-art methods, underscoring the significance of key motion capture for achieving temporally coherent full motion synthesis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "text", + "content": "Is our method sensitive to different Phomene-based Localization methods? We experimented with different Automatic Speech Recognition (ASR) models, such as Auto-avs r [28] and Whisper [39]. The results in the first three rows of Table 4 show that our KMTalk method consistently maintained high performance, achieving at least a " + }, + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "inline_equation", + "content": "21\\%" + }, + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "text", + "content": " improvement in FDD regardless of the ASR model used [28, 39]. This highlights the robustness of our approach across various ASR models. Besides, to simulate phoneme localization deviations, we shifted all key motion indices extracted by Auto-avs r [28] one frame to the right. The results in the last row of Table 4 indicate negligible variations. This demonstrates the robustness of our method to inaccurate key frame localization." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "text", + "content": "KMTalk" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 142, + 167, + 297, + 270 + ], + "blocks": [ + { + "bbox": [ + 138, + 125, + 282, + 157 + ], + "lines": [ + { + "bbox": [ + 138, + 125, + 282, + 157 + ], + "spans": [ + { + "bbox": [ + 138, + 125, + 282, + 157 + ], + "type": "text", + "content": "Table 4: Robust analysis of Phoneme-based Localization on BIWI-Test-A." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 167, + 297, + 270 + ], + "lines": [ + { + "bbox": [ + 142, + 167, + 297, + 270 + ], + "spans": [ + { + "bbox": [ + 142, + 167, + 297, + 270 + ], + "type": "table", + "html": "
MethodsLVE↓ ×10-4mmFDD↓ ×10-5mm
Auto-avsr [28]3.96542.5446
Whisper-large [39]4.07182.8141
Whisper-tiny [39]4.06432.8083
Auto-avsr+offset3.99912.6420
", + "image_path": "b3d462ccbb68e84ad47bac09de429e964dd9afc31f8de4bded6d589b85e333dd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 305, + 167, + 467, + 270 + ], + "blocks": [ + { + "bbox": [ + 298, + 125, + 474, + 157 + ], + "lines": [ + { + "bbox": [ + 298, + 125, + 474, + 157 + ], + "spans": [ + { + "bbox": [ + 298, + 125, + 474, + 157 + ], + "type": "text", + "content": "Table 5: The results of integrating our proposed KMTalk with existing methods on BIWI-Test-A." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 305, + 167, + 467, + 270 + ], + "lines": [ + { + "bbox": [ + 305, + 167, + 467, + 270 + ], + "spans": [ + { + "bbox": [ + 305, + 167, + 467, + 270 + ], + "type": "table", + "html": "
MethodsLVE↓ ×10-4mmFDD↓ ×10-5mm
FaceFormer [12]Original5.30774.6408
After5.27934.2654
CodeTalker [53]Original4.79144.1170
After4.50963.9043
SelfTalk [36]Original4.24853.5761
After4.11222.8668
", + "image_path": "bd1dd33089855e014118ae2055aedf4d8ef828d9fb5e6f531c8800e7058149b1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 285, + 336, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 336, + 298 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 336, + 298 + ], + "type": "text", + "content": "4.3 Integration with Existing Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 422 + ], + "type": "text", + "content": "Existing approaches focus on enhancing prediction outcomes by designing elaborate priors or the learning-based recognizer, which may be highly coupled with the proposed architecture of these methods. Our KMTalk introduces a new learning strategy of speech-driven talking face generation, which is orthogonal to these approaches. Therefore, we can explore whether performance can be enhanced by applying our progressive learning scheme without the need for additional fine-tuning of their models. Detailed implementation is in the supplementary material. As shown in Table 5, the results of existing methods are improved after integration with our proposed progressive learning mechanism utilizing key motion embeddings. This further emphasizes the efficacy of our design." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 439, + 220, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 220, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 220, + 451 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 666 + ], + "type": "text", + "content": "In this work, we introduce KMTalk, a novel method for progressively learning 3D facial animation from speeches using key motion embeddings. It incorporates linguistic priors for key motion generation and extends them to a full motion sequence via data-driven priors. We propose phoneme-based localization methods to determine the temporal position of key facial motions, improving lip-speech synchronization by aligning motion transitions with phoneme changes. Additionally, we design a cross-modal facial motion completion module that synthesizes the entire motion sequence from key motions and audio features, enhancing lip-speech synchronization and motion coherence. Extensive evaluations of the datasets demonstrate KMTalk's superiority over existing methods, producing more accurate and realistic animations. Moreover, coupling our idea with existing methods consistently improves performance, further verifying the efficacy of our proposed progressive learning mechanism based on key motion acquisition. Although the proposed method has demonstrated its robustness to inaccurate keyframe localization, it may encounter errors in dialect variations. Integrating advanced ASR(Automatic Speech Recognition) technology in the future could enhance its adaptability to various speech patterns." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Z. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Acknowledgement The research is partially supported by National Key Research and Development Program of China (2023YFC3502900), National Natural Science Foundation of China (No.62176093, 61673182), Key Realm Research and Development Program of Guangzhou (No.202206030001), Guangdong-Hong Kong-Macao Joint Innovation Project (No.2023A0505030016), Guangdong Natural Science Foundation (No. 2024A1515012217)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 205, + 197, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 197, + 218 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 197, + 218 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 228, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "type": "text", + "content": "1. Alghamdi, M.M., Wang, H., Bulpitt, A.J., Hogg, D.C.: Talking head from speech audio using a pre-trained image generator. In: ACM MM. pp. 5228-5236 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 252, + 480, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 252, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 138, + 252, + 480, + 284 + ], + "type": "text", + "content": "2. Baevski, A., Zhou, Y., Mohamed, A., Auli, M.: wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems 33, 12449-12460 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 285, + 480, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 480, + 316 + ], + "type": "text", + "content": "3. Bao, L., Zhang, H., Qian, Y., Xue, T., Chen, C., Zhe, X., Kang, D.: Learning audio-driven viseme dynamics for 3d face animation. arXiv preprint arXiv:2301.06059 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 317, + 480, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 317, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 138, + 317, + 480, + 338 + ], + "type": "text", + "content": "4. Cao, Y., Tien, W.C., Faloutsos, P., Pighin, F.: Expressive speech-driven facial animation. ACM Trans. Graph 24(4), 1283-1302 (2005)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 339, + 480, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 339, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 339, + 480, + 360 + ], + "type": "text", + "content": "5. Chen, L., Cui, G., Liu, C., Li, Z., Kou, Z., Xu, Y., Xu, C.: Talking-head generation with rhythmic head motion. In: ECCV. pp. 35-51. Springer (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 361, + 480, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 480, + 382 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 480, + 382 + ], + "type": "text", + "content": "6. Chung, J.S., Zisserman, A.: Out of time: automated lip sync in the wild. In: ACCV. pp. 251-263. Springer (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 383, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 383, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 383, + 480, + 403 + ], + "type": "text", + "content": "7. Cohen, M.M., Clark, R., Massaro, D.W.: Animated speech: Research progress and applications. In: AVSP (2001)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 404, + 480, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 480, + 426 + ], + "type": "text", + "content": "8. Cudeiro, D., Bolkart, T., Laidlaw, C., Ranjan, A., Black, M.J.: Capture, learning, and synthesis of 3d speaking styles. In: CVPR. pp. 10101-10111 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 426, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 480, + 468 + ], + "type": "text", + "content": "9. Dai, G., Zhang, Y., Wang, Q., Du, Q., Yu, Z., Liu, Z., Huang, S.: Disentangling writer and character styles for handwriting generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5977-5986 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 470, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 470, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 134, + 470, + 480, + 502 + ], + "type": "text", + "content": "10. Das, D., Biswas, S., Sinha, S., Bhowmick, B.: Speech-driven facial animation using cascaded gans for learning of motion and texture. In: ECCV. pp. 408-424. Springer (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 502, + 480, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 502, + 480, + 524 + ], + "spans": [ + { + "bbox": [ + 134, + 502, + 480, + 524 + ], + "type": "text", + "content": "11. Edwards, P., Landreth, C., Fiume, E., Singh, K.: Jali: an animator-centric viseme model for expressive lip synchronization. ACM Trans. Graph 35(4), 1-11 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 525, + 480, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 525, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 134, + 525, + 480, + 545 + ], + "type": "text", + "content": "12. Fan, Y., Lin, Z., Saito, J., Wang, W., Komura, T.: Faceformer: Speech-driven 3d facial animation with transformers. In: CVPR. pp. 18770-18780 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 545, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 545, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 134, + 545, + 480, + 578 + ], + "type": "text", + "content": "13. Fanelli, G., Gall, J., Romsdorfer, H., Weise, T., Van Gool, L.: A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia 12(6), 591-598 (2010)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 578, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 578, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 134, + 578, + 480, + 600 + ], + "type": "text", + "content": "14. Fisher, C.G.: Confusions among visually perceived consonants. Journal of speech and hearing research 11(4), 796-804 (1968)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 601, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 480, + 621 + ], + "type": "text", + "content": "15. Guo, Y., Chen, K., Liang, S., Liu, Y.J., Bao, H., Zhang, J.: Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In: ICCV. pp. 5784-5794 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "type": "text", + "content": "16. Habibie, I., Xu, W., Mehta, D., Liu, L., Seidel, H.P., Pons-Moll, G., Elgharib, M., Theobalt, C.: Learning speech-driven 3d conversational gestures from video. In: Proceedings of the 21st ACM International Conference on Intelligent Virtual Agents. pp. 101-108 (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "text", + "content": "KMTalk" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "17. Iverson, P., Kuhl, P.K.: Perceptual magnet and phoneme boundary effects in speech perception: Do they arise from a common mechanism? Perception & psychophysics 62, 874-886 (2000)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "text", + "content": "18. Ji, X., Zhou, H., Wang, K., Wu, Q., Wu, W., Xu, F., Cao, X.: Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In: ACM SIGGRAPH. pp. 1-10 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "19. Ji, X., Zhou, H., Wang, K., Wu, W., Loy, C.C., Cao, X., Xu, F.: Audio-driven emotional video portraits. In: CVPR. pp. 14080-14089 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "20. Karras, T., Aila, T., Laine, S., Herva, A., Lehtinen, J.: Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Trans. Graph 36(4), 1-12 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "type": "text", + "content": "21. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "type": "text", + "content": "22. Lewis, J.: Automated lip-sync: Background and techniques. IEEE Trans Vis Comput Graph 2(4), 118-122 (1991)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 315 + ], + "type": "text", + "content": "23. Liang, B., Pan, Y., Guo, Z., Zhou, H., Hong, Z., Han, X., Han, J., Liu, J., Ding, E., Wang, J.: Expressive talking head generation with granular audio-visual control. In: CVPR. pp. 3387-3396 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 316, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 480, + 335 + ], + "type": "text", + "content": "24. Liu, C.: An analysis of the current and future state of 3d facial animation techniques and systems (2009)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "25. Liu, X., Xu, Y., Wu, Q., Zhou, H., Wu, W., Zhou, B.: Semantic-aware implicit neural audio-driven video portrait generation. In: ECCV. pp. 106-125. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 370, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 480, + 391 + ], + "type": "text", + "content": "26. Liu, Y., Xu, F., Chai, J., Tong, X., Wang, L., Huo, Q.: Video-audio driven real-time facial animation. ACM Trans. Graph 34(6), 1-10 (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "text", + "content": "27. Lu, L., Wu, R., Lin, H., Lu, J., Jia, J.: Video frame interpolation with transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3532-3542 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "text", + "content": "28. Ma, P., Haliassos, A., Fernandez-Lopez, A., Chen, H., Petridis, S., Pantic, M.: Auto-avsr: Audio-visual speech recognition with automatic labels. arxiv 2023. arXiv preprint arXiv:2303.14307" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "text", + "content": "29. Massaro, D., Cohen, M., Tabain, M., Beskow, J., Clark, R.: 12 animated speech: research progress and applications (2012)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "30. Mattheyses, W., Verhelst, W.: Audiovisual speech synthesis: An overview of the state-of-the-art. Speech Communication 66, 182-217 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "text", + "content": "31. McAuliffe, M., Sonderegger, M.: English mfa acoustic model v2.2.1. Tech. rep., https://mfa-models.readthedocs.io/acoustic/English/EnglishMFAacousticmodelv2_2_1.html (May 2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "32. Mo, C.A., Hu, K., Long, C., Wang, Z.: Continuous intermediate token learning with implicit motion manifold for keyframe based motion interpolation. In: CVPR. pp. 13894-13903 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "text", + "content": "33. Niklaus, S., Mai, L., Liu, F.: Video frame interpolation via adaptive convolution. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 670-679 (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "text", + "content": "34. Nocentini, F., Ferrari, C., Berretti, S.: Learning landmarks motion from speech for speaker-agnostic 3d talking heads generation. arXiv preprint arXiv:2306.01415 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "35. Pang, Y., Zhang, Y., Quan, W., Fan, Y., Cun, X., Shan, Y., Yan, D.m.: Dpe: Disentanglement of pose and expression for general video portrait editing. In: CVPR. pp. 427-436 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "Z. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "36. Peng, Z., Luo, Y., Shi, Y., Xu, H., Zhu, X., Liu, H., He, J., Fan, Z.: Selftalk: A self-supervised commutative training diagram to comprehend 3d talking faces. arXiv preprint arXiv:2306.10799 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "text", + "content": "37. Pham, H.X., Wang, Y., Pavlovic, V.: End-to-end learning for 3d facial animation from speech. In: ICMI. pp. 361-365 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 173, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 481, + 205 + ], + "type": "text", + "content": "38. Prajwal, K., Mukhopadhyay, R., Namboodiri, V.P., Jawahar, C.: A lip sync expert is all you need for speech to lip generation in the wild. In: ACM MM. pp. 484-492 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 206, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 481, + 239 + ], + "type": "text", + "content": "39. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision. In: International Conference on Machine Learning. pp. 28492-28518. PMLR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 239, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 481, + 262 + ], + "type": "text", + "content": "40. Richard, A., Lea, C., Ma, S., Gall, J., De la Torre, F., Sheikh, Y.: Audio-and gaze-driven facial animation of codec avatars. In: WACV. pp. 41-50 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 263, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 481, + 296 + ], + "type": "text", + "content": "41. Richard, A., Zollhöfer, M., Wen, Y., De la Torre, F., Sheikh, Y.: Meshtalk: 3d face animation from speech using cross-modality disentanglement. In: ICCV. pp. 1173-1182 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 297, + 481, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 328 + ], + "type": "text", + "content": "42. Shen, S., Li, W., Zhu, Z., Duan, Y., Zhou, J., Lu, J.: Learning dynamic facial radiance fields for few-shot talking head synthesis. In: ECCV. pp. 666-682. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 329, + 481, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 329, + 481, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 329, + 481, + 351 + ], + "type": "text", + "content": "43. Tanaka, H., Nakamura, S., et al.: The acceptability of virtual characters as social skills trainers: usability study. JMIR human factors 9(1), e35358 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 352, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 352, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 133, + 352, + 481, + 396 + ], + "type": "text", + "content": "44. Tang, J., Dai, A., Nie, Y., Markhasin, L., Thies, J., Nießner, M.: Dphms: Diffusion parametric head models for depth-based tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1111-1122 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 396, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 396, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 133, + 396, + 481, + 430 + ], + "type": "text", + "content": "45. Taylor, S., Kim, T., Yue, Y., Mahler, M., Krahe, J., Rodriguez, A.G., Hodgins, J., Matthews, I.: A deep learning approach for generalized speech animation. ACM Trans. Graph 36(4), 1-11 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 430, + 481, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 430, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 133, + 430, + 481, + 452 + ], + "type": "text", + "content": "46. Taylor, S.L., Mahler, M., Theobald, B.J., Matthews, I.: Dynamic units of visual speech. In: ACM SIGGRAPH. pp. 275-284 (2012)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 453, + 481, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 453, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 133, + 453, + 481, + 474 + ], + "type": "text", + "content": "47. Vougioukas, K., Petridis, S., Pantic, M.: Realistic speech-driven facial animation with gans. International Journal of Computer Vision 128, 1398-1413 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 475, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 475, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 133, + 475, + 481, + 497 + ], + "type": "text", + "content": "48. Wang, J., Qian, X., Zhang, M., Tan, R.T., Li, H.: Seeing what you said: Talking face generation guided by a lip reading expert. In: CVPR. pp. 14653-14662 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 498, + 481, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 498, + 481, + 519 + ], + "spans": [ + { + "bbox": [ + 133, + 498, + 481, + 519 + ], + "type": "text", + "content": "49. Wang, Q., Fan, Z., Xia, S.: 3d-talkemo: Learning to synthesize 3d emotional talking head. arXiv preprint arXiv:2104.12051 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 520, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 520, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 133, + 520, + 481, + 542 + ], + "type": "text", + "content": "50. Wang, S., Li, L., Ding, Y., Yu, X.: One-shot talking face generation from single-speaker audio-visual correlation learning. In: AAAI. vol. 36, pp. 2531-2539 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 543, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 543, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 133, + 543, + 481, + 575 + ], + "type": "text", + "content": "51. Wen, S., Liu, W., Yang, Y., Huang, T., Zeng, Z.: Generating realistic videos from keyframes with concatenated gans. IEEE Transactions on Circuits and Systems for Video Technology 29(8), 2337-2348 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 576, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 576, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 576, + 481, + 609 + ], + "type": "text", + "content": "52. Wu, H., Zhou, S., Jia, J., Xing, J., Wen, Q., Wen, X.: Speech-driven 3d face animation with composite and regional facial movements. arXiv preprint arXiv:2308.05428 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 609, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 609, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 609, + 481, + 643 + ], + "type": "text", + "content": "53. Xing, J., Xia, M., Zhang, Y., Cun, X., Wang, J., Wong, T.T.: Codetalker: Speechdriven 3d facial animation with discrete motion prior. In: CVPR. pp. 12780-12790 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 481, + 665 + ], + "type": "text", + "content": "54. Xu, Y., Feng, A.W., Marsella, S., Shapiro, A.: A practical and configurable lip sync method for games. In: Proceedings of Motion on Games, pp. 131-140 (2013)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "text", + "content": "KMTalk" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 336 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "55. Yi, R., Ye, Z., Zhang, J., Bao, H., Liu, Y.J.: Audio-driven talking face video generation with learning-based personalized head pose. arXiv preprint arXiv:2002.10137 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "56. Yin, S., Wu, C., Yang, H., Wang, J., Wang, X., Ni, M., Yang, Z., Li, L., Liu, S., Yang, F., et al.: Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv preprint arXiv:2303.12346 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "text", + "content": "57. Yu, D., Li, X., Zhang, C., Liu, T., Han, J., Liu, J., Ding, E.: Towards accurate scene text recognition with semantic reasoning networks. In: CVPR. pp. 12113-12122 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "58. Yue, X., Kuang, Z., Lin, C., Sun, H., Zhang, W.: Robustscanner: Dynamically enhancing positional clues for robust text recognition. In: ECCV. pp. 135-151. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "text", + "content": "59. Zhang, B., Qi, C., Zhang, P., Zhang, B., Wu, H., Chen, D., Chen, Q., Wang, Y., Wen, F.: Metaportrait: Identity-preserving talking head generation with fast personalized adaptation. In: CVPR. pp. 22096-22105 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "text", + "content": "60. Zhang, C., Ni, S., Fan, Z., Li, H., Zeng, M., Budagavi, M., Guo, X.: 3d talking face with personalized pose dynamics. IEEE Trans Vis Comput Graph. (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "text", + "content": "61. Zhou, H., Sun, Y., Wu, W., Loy, C.C., Wang, X., Liu, Z.: Pose-controllable talking face generation by implicitly modularized audio-visual representation. In: CVPR. pp. 4176-4186 (2021)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "Z. Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Kalman-Inspired Feature Propagation for Video Face Super-Resolution/86d2672e-c2f0-471a-a7a6-101117941b3f_content_list.json b/2024/Kalman-Inspired Feature Propagation for Video Face Super-Resolution/86d2672e-c2f0-471a-a7a6-101117941b3f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6b365f53a4e11bcc44731ff7fec1d9624a520b0e --- /dev/null +++ b/2024/Kalman-Inspired Feature Propagation for Video Face Super-Resolution/86d2672e-c2f0-471a-a7a6-101117941b3f_content_list.json @@ -0,0 +1,1811 @@ +[ + { + "type": "text", + "text": "Kalman-Inspired Feature Propagation for Video Face Super-Resolution", + "text_level": 1, + "bbox": [ + 264, + 140, + 736, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruicheng Feng, Chongyi Li, and Chen Change Loy", + "bbox": [ + 303, + 212, + 700, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "S-Lab, Nanyang Technological University, Singapore {ruicheng002, ccloy}@ntu.edu.sg lichongyi25@gmail.com", + "bbox": [ + 328, + 239, + 674, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Despite the promising progress of face image super-resolution, video face super-resolution remains relatively under-explored. Existing approaches either adapt general video super-resolution networks to face datasets or apply established face image super-resolution models independently on individual video frames. These paradigms encounter challenges either in reconstructing facial details or maintaining temporal consistency. To address these issues, we introduce a novel framework called Kalman-inspired Feature Propagation (KEEP), designed to maintain a stable face prior over time. The Kalman filtering principles offer our method a recurrent ability to use the information from previously restored frames to guide and regulate the restoration process of the current frame. Extensive experiments demonstrate the effectiveness of our method in capturing facial details consistently across video frames. Code and video demo are available at https://jnjaby.github.io/projects/KEEP/.", + "bbox": [ + 264, + 316, + 740, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 218, + 551, + 372, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The field of Face Super-Resolution (FSR), which focuses on reconstructing high-resolution (HR) face images from highly degraded versions, has witnessed remarkable progress. In particular, numerous studies have successfully leveraged various types of prior information, such as geometric facial priors [7,8,53], reference priors [26-28], generative priors [2,4,45,52], and codebook priors [13,46,56]. These approaches have significantly advanced the realism and quality of generated face images. However, the majority of these studies are confined to still images, with the extension to Video Face Super-Resolution (VFSR) remaining relatively under-explored. Despite the substantial potential benefits of video face restoration in various practical domains, such as the restoration of old films, VFSR has yet to receive the same level of attention and development as its image-based counterpart.", + "bbox": [ + 218, + 582, + 785, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Two main strategies emerge for implementing VFSR. The first approach involves adapting general Video Super-Resolution (VSR) networks, such as EDVR [44], BasicVSR [3], BasicVSR++ [5], and RVRT [30], to large-scale face video datasets [34, 48]. These methods exploit temporal information and propagate features across video frames. However, they are not specifically tailored for", + "bbox": [ + 218, + 765, + 784, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d0990cb8d6bd312c9850737ee484d46a046f3d917329cb81e114178078088ed4.jpg", + "image_caption": [ + "Fig. 1: Comparing main VFSR strategies. We show seven frames with an interval of 6. Generic VSR model BasicVSR [3] fails to reconstruct facial components faithfully. Single-image FSR model CodeFormer [56] hallucinates unnatural and inconsistent face details. Our method, in contrast, enables consistent restoration of low-quality face video while preserving temporal coherence across frames." + ], + "image_footnote": [], + "bbox": [ + 228, + 143, + 781, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "face restoration and often fall short in reconstructing detailed facial features, particularly in severely degraded scenarios, as depicted in Fig. 1(second row). The second method involves applying existing face image SR models to process each video frame independently. This frame-by-frame approach, while straightforward, introduces temporal inconsistencies in the video, as demonstrated in Fig. 1(third row). This problem arises because FSR is inherently ill-posed and the existing priors may not suffice to maintain appearance consistency throughout the video sequence. Specifically, a single degraded image could correspond to multiple high-resolution interpretations, leading to discrepancies and inconsistent structures across independently processed video frames.", + "bbox": [ + 212, + 503, + 787, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we wish to devise an effective framework for maintaining a stable face prior over time for VFSR. We use CodeFormer [56], a representative model that exploits codebook priors for FSR, to demonstrate how face priors can be consistently preserved across different time frames. A pivotal aspect of our approach is the idea that frames previously restored can act as references, guiding and regulating the restoration process of the current frame. This strategy helps minimize the divergence between consecutive frames. Moreover, this reliance on previously restored frames naturally suggests a recurrent framework, enabling the effective use of information from past restorations. This intuition aligns closely with Kalman filtering principles, or linear quadratic estimation, which involves using a sequence of time-based measurements, despite their statistical noise and inaccuracies, to produce more accurate estimates of unknown", + "bbox": [ + 212, + 657, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "R. Feng et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "variables than would be possible with a single measurement. Similarly, in VFSR, faces observed over time are often noisy and inaccurate, making them suitable for refinement using Kalman filtering techniques.", + "bbox": [ + 212, + 146, + 784, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Driven by these insights, we formulate a novel method, Kalman-inspired fEaturE Propagation (KEEP), which recurrently updates the current latent state in CodeFormer by incorporating information from preceding frames. This method of temporal propagation within the latent space ensures the stability of the face prior over time, thereby capturing facial details that consistently match in appearance. The effectiveness of KEEP is shown in Fig. 1, where it is evident that our method delivers high-quality restoration with superior consistency, compared to both generic video restoration methods fine-tuned for faces and approaches that restore frames independently. Please refer to the supplementary video to appreciate the superiority of our approach in terms of temporal consistency. A key advantage of KEEP is its robustness in handling severe video-based degradation, outperforming single-image models. In addition, our model demonstrates enhanced performance on non-frontal faces by providing more stable estimations of face priors.", + "bbox": [ + 212, + 191, + 787, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, the main contribution of this work is a novel framework for maintaining a stable and meaningful face prior for VFSR. While we demonstrate its application using the CodeFormer method as a case study, the underlying principles of our framework, inspired by the Kalman filtering approach, are applicable to other approaches. Extensive experimental results on the VFHQ dataset [48] and real-world data demonstrate the effectiveness of our approach in improving both the fidelity and coherence of VFSR outputs. Compared to other state-of-the-art methods, our model achieves superior performance with a large margin of 0.8 dB in PSNR, while also significantly maintaining temporal coherence.", + "bbox": [ + 212, + 404, + 787, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 564, + 387, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Blind Face Restoration. Blind face restoration aims at recovering severely degraded face images in the wild. Unlike natural images, faces are highly structured. This property allows researchers to incorporate prior information into the restoration models, which have demonstrated remarkable progress in capabilities to restore high-quality faces. Most existing FSR methods can be categorized into four classes: geometric priors, reference priors, generative priors, and codebook priors. Geometric priors usually include facial elements, such as face landmarks [8], parsing maps [7], and facial component heatmaps [53]. Another major line is reference-based methods that require high-quality exemplar images. GFRNet [28] and ASFFNet [27] leverage a warped high-quality image to extract rich details to improve facial detail restoration. DFDNet [26] constructs deep dictionaries with facial components from large-scale images to recover fine details. Generative priors from pre-trained GAN, e.g., StyleGAN [20] and StyleGAN2 [21], are employed through iterative latent optimization of GAN inversion [12, 33, 37]. Still, they produce face images with low fidelity and are computationally expensive. To address this, GLEAN [2], GPEN [52], and GFP-", + "bbox": [ + 212, + 598, + 787, + 843 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "KEEP", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "GAN [45] integrate generative priors into encoder-decoder architectures, which estimate latent priors in one-forward pass. These methods achieve great tradeoff between quality and fidelity but usually fail when the corruption is severe. Codebook priors [13,46,56] can be regarded as a special case of generative priors. In contrast to continuous generative priors, they squeeze the latent space into a small finite codebook space and improve the robustness to severe degradation. However, most existing FSR methods are image-based and thus they cannot guarantee temporally consistent details for VFSR.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video Super-Resolution. Most existing video restoration techniques can be categorized into two paradigms based on their parallelizability: parallel and recurrent methods. Parallel models estimate all frames simultaneously, and the restoration of each frame does not rely on the update of other frames. These methods typically involve feature extraction, feature alignment, feature fusion, and reconstruction. Representative works, FSTRN [25] and VESCPN [1], introduce fast spatio-temporal networks using 3D convolutions to enhance alignment, combining motion compensation and super-resolution. EDVR [44] and TDAN [42] leverage deformable convolutions for aligning adjacent frames. Besides, Transformer-based methods [29,32] are proposed to reconstruct all frames simultaneously by jointly extracting, aligning, and fusing features. In addition, RVRT [30] and TTVSR [31] integrate optical flow into Transformer and enable long-range models in videos. Empowered by the great expressive capability of Transformer, this line of work exhibits remarkable performance improvements over previous methods. However, they suffer from large model sizes and high memory consumption. Recurrent methods [3,16,18] do not aggregate information solely from adjacent frames. Instead, they maintain hidden states to convey relevant information from previous frames and propagate latent features sequentially, accumulating information for later restoration. For example, RBPN [14] treated each frame as a separate source, combined iteratively in a refinement framework. RSDN [17] divided the input into structure and detail components, proposing a two-stream structure-detail block to learn textures. BasicVSR [3] and BasicVSR++ [5] fused bidirectional hidden states from both past and future frames, bring significant improvements. They aim to fully utilize information from the entire sequence, synchronously updating the hidden state through the weights of the reconstruction network. Due to the recurrent nature of feature propagation, recurrent methods experience information loss.", + "bbox": [ + 212, + 273, + 787, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video Consistency. In addition to VSR, previous works also attempt to inflate image models into video models [43, 47, 50, 51] or improve temporal consistency with the implicit representation of the given videos [23, 24, 35]. For example, Stitch-it-in-Time [43] discovers locally consistent pivots in the latent space to provide spatially consistent transitions. Tune-A-Video [47] adopts cross-frame attention and fine-tune it on a single video. All-in-One deflicker [23] learns a neural atlas for each video to solve long-term inconsistency. These works either require post-processing or optimization on a video basis. In contrast, we aim at repurposing image FSR models and enforcing temporal consistency in the restored face videos.", + "bbox": [ + 212, + 688, + 789, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "R. Feng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 143, + 380, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In a corrupted face video, local textures and facial details are irrevocably lost. Therefore, the latent codes, usually estimated by an encoder, are inaccurate to match the real underlying priors of ground-truth video. Different from image-based autoencoder, models in the video domain enable us to exploit evidence accumulated from preceding frames for better restoration. In this paper, we repurpose the image-based CodeFormer for video FSR and propose KEEP to estimate stable face priors in latent space over time given the noisy and inaccurate estimations, which consequently enable temporally coherent restoration. An appealing idea to realize KEEP is through reformulating the method in a Kalman Filter framework.", + "bbox": [ + 212, + 174, + 787, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Formulation", + "text_level": 1, + "bbox": [ + 215, + 345, + 362, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "State Space Model. We consider observations of low-quality (LQ) video sequence $X = \\{\\pmb{x}_{t}\\}_{t=1}^{T}$ with length of $T$ , where $\\pmb{x}_t \\in \\mathbb{R}^{H \\times W \\times 3}$ , and underlying high-quality (HQ) sequences $Y = \\{\\pmb{y}_t\\}_{t=1}^T$ . Kalman filter [19] assumes linear dynamic systems that are characterized by a state space model driven by Gaussian noise", + "bbox": [ + 212, + 366, + 787, + 441 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {t} = \\boldsymbol {F} _ {t} \\boldsymbol {y} _ {t - 1} + \\boldsymbol {q} _ {t}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 436, + 444, + 785, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{F}_t$ is the transition matrix and $\\pmb{q}_t$ denotes process noise drawn from Gaussian noise. The observation $\\pmb{x}_t$ is measured by", + "bbox": [ + 212, + 463, + 785, + 494 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t} = \\boldsymbol {H} \\boldsymbol {y} _ {t} + \\boldsymbol {r} _ {t}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 446, + 503, + 784, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{H}$ is the measurement matrix and $\\pmb{r}_t$ represents measurement noise. However, the linear assumption does not hold in some complex real-world scenarios. Hence, the non-linear Kalman filter can be reformulated as", + "bbox": [ + 212, + 527, + 785, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {t} = d \\left(\\boldsymbol {y} _ {t - 1}, \\boldsymbol {q} _ {t}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 442, + 582, + 784, + 597 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t} = h \\left(\\boldsymbol {y} _ {t}\\right) + \\boldsymbol {r} _ {t}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 442, + 602, + 784, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d(\\cdot)$ and $h(\\cdot)$ are non-linear transition and measurement models. Specifically, $d(\\cdot)$ can be represented by any explicit motion estimation (e.g., optical flow), which defines how the current frame transits to the next one. $h(\\cdot)$ commonly models the degradation in video restoration problems. As opposed to the classical assumptions in Kalman filter, the measurement function $h(\\cdot)$ is unknown in a blind setting. This is regarded as partially known dynamic models [39].", + "bbox": [ + 212, + 625, + 787, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by VQGAN [10] and Stable Diffusion [40], we estimate underlying latent representations $Z = \\{z_{t}\\}_{t=1}^{T}$ such that $z_{t}$ can correspond to $y_{t}$ by a generative model $g_{\\theta}$ , given by", + "bbox": [ + 212, + 715, + 787, + 762 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {t} = g _ {\\theta} \\left(\\boldsymbol {z} _ {t}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 457, + 771, + 784, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instead of directly estimating individual pixels, modeling the low-dimensional latent code is computationally more efficient and focuses on more perceptually significant variations of the data. The graphical model is depicted in Fig. 2 (a).", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "KEEP", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b24db573ef9e18763e701385155b15c194b9262dbc57157da9c67a056394191a.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 444, + 263 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3243931ce86ba68f7e0872371ad01dfa3bde02ef51ec1347563adb3fd76e6268.jpg", + "image_caption": [ + "(b)", + "Fig.2: (a) Graphical model of state space. It defines the underlying dynamic system model, where $f$ describes how the latent states $z_{t}$ transit over time, $g$ is a generative model, and $h$ models the degradation from clean frame $y_{t}$ to degraded frame $x_{t}$ . (b) Block diagram of Kalman filter model. In each time step, a predictive state from previous frame $\\hat{z}_{t-1}^{+}$ (Blue dash box) and new observed state of current frame $x_{t}$ (Red dash box) are fused by Kalman gain $\\mathcal{K}_t$ from Kalman Gain Network (KGN) to produce more accurate estimates. The combined state $\\hat{z}_t^+$ is then used to generate the estimated clean frame $\\hat{y}_t$ by $g_{\\theta}$ . Note that $\\tilde{z}_1$ goes along with $\\tilde{z}_{t-1}$ as an anchor and it is omitted in the diagram for simplicity." + ], + "image_footnote": [], + "bbox": [ + 452, + 146, + 776, + 265 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Kalman Filter Model. The principles of Kalman filter can be formulated by a two-step procedure, i.e., state prediction and state update. The overview diagram is illustrated in Fig. 2 (b). In this problem, the observation is a face image $\\mathbf{x}_t$ and the state is $\\mathbf{z}_t$ .", + "bbox": [ + 214, + 446, + 787, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the state prediction step, the model predicts the prior estimation $\\hat{z}_t^-$ of the current state $z_{t}$ based on the posterior estimation $\\hat{z}_{t - 1}^{+}$ of the previous state and the dynamic model. Specifically, the prior estimation of latent state $z_{t}$ and estimation of observation $\\pmb{x}_t$ are computed as", + "bbox": [ + 214, + 508, + 789, + 568 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {t} ^ {-} = f \\left(\\hat {z} _ {t - 1} ^ {+}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 578, + 784, + 595 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {x}} _ {t} ^ {-} = h \\left(g _ {\\theta} \\left(\\hat {\\boldsymbol {z}} _ {t} ^ {-}\\right)\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 599, + 784, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The system dynamics $f$ define how the latent state $Z$ evolves over time, and it incorporates any control inputs that might affect the current state.", + "bbox": [ + 214, + 627, + 785, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the state update step, a posterior state estimation $\\hat{z}_t^+$ is computed based on the prior estimation and new observations $\\pmb{x}_t$ as", + "bbox": [ + 214, + 657, + 785, + 688 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {t} ^ {+} = \\hat {z} _ {t} ^ {-} + \\mathcal {K} _ {t} \\Delta z _ {t}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 696, + 784, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{K}_t$ is conceptually referred to as the Kalman gain and $\\Delta z_{t}$ as the innovation, i.e., the residual between the prior estimation $\\hat{\\pmb{z}}_t^-$ and approximation of current state from $\\pmb{x}_{t}$ , given by", + "bbox": [ + 214, + 726, + 787, + 772 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\boldsymbol {z} _ {t} = \\hat {\\boldsymbol {z}} _ {t} ^ {-} - \\tilde {\\boldsymbol {z}} _ {t}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 442, + 782, + 784, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tilde{\\pmb{z}}_t = e(\\pmb{x}_t)$ . Note that this formula differs from the original Kalman filter which minimizes the innovation of $\\Delta \\pmb{x}_t$ (i.e., residual between $\\pmb{x}_t$ and $\\hat{\\pmb{x}}_t^{-}$ ), since", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "R. Feng et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f066a8fe1c1daebbee5240bd34542656ce6bb37eef9f54e7f19a22480a7627ed.jpg", + "image_caption": [ + "Fig. 3: Overview of the proposed KEEP. It consists of four modules: encoder $\\mathcal{E}_L$ , decoder $\\mathcal{D}_Q$ , Kalman filter network, and CFA. We illustrate the information flow in one timestep." + ], + "image_footnote": [], + "bbox": [ + 241, + 146, + 754, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the assumption of available measurement function $h(\\cdot)$ is no longer valid in our setting. An original measurement system models how observation $\\boldsymbol{x}_t$ is derived from the latent state $\\boldsymbol{z}_t$ , formally $\\boldsymbol{x}_t = h(\\boldsymbol{z}_t)$ . Inspired by KFNet [55], we directly estimate the state $\\boldsymbol{z}_t$ given new observation $\\boldsymbol{x}_t$ , i.e., mapping $\\boldsymbol{x}_t$ to $\\tilde{\\boldsymbol{z}}_t$ with an estimator $e$ .", + "bbox": [ + 212, + 387, + 782, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The remaining problem is to compute the Kalman gain $\\mathcal{K}_t$ . As discussed in KalmanNet [39], covariance estimation is intractable when dealing with high-dimensional signals. Additionally, the second-order statistical moments are only used for calculating Kalman gain. Inspired by this, we directly learn the gains from the data distribution and do not explicitly maintain an estimation of covariances. Additionally, we follow [51] to include $\\tilde{\\pmb{z}}_1$ of the first frame as anchor into KGN for Kalman gain estimation. Then, the final predicted $\\hat{\\pmb{y}}_t$ can be derived by", + "bbox": [ + 212, + 463, + 784, + 584 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {y}} _ {t} = g _ {\\theta} \\left(\\hat {\\boldsymbol {z}} _ {t} ^ {+}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 452, + 585, + 784, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Parameterized Models", + "text_level": 1, + "bbox": [ + 215, + 627, + 446, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Fig. 2 (b), we will parameterize or define the system dynamics $f$ , observation estimator $e$ , Kalman Gain Nets (KGN), and generative model $g_{\\theta}$ in this section. The overall framework is shown in Fig. 3.", + "bbox": [ + 212, + 656, + 782, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generative Model. The generative model $g_{\\theta}$ is generally parameterized by a backbone of CodeFormer [56], which consists of a LQ encoder $\\mathcal{E}_L$ , a HQ encoder $\\mathcal{E}_H$ , a codebook lookup Transformer and quantization layer $T_{Q}$ , and a decoder $\\mathcal{D}$ . For simplicity, $\\mathcal{D}_Q$ denotes the decoder with the codebook lookup Transformer and quantization layer absorbed. Basically, the predicted $\\hat{\\pmb{y}}_t$ is given by $\\mathcal{D}_Q(\\hat{\\pmb{z}}_t^+)$ , and the observed state $\\tilde{\\pmb{z}}_t$ is approximated by", + "bbox": [ + 214, + 719, + 782, + 810 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {z} _ {t} = e \\left(\\boldsymbol {x} _ {t}\\right) = \\mathcal {E} _ {L} \\left(\\boldsymbol {x} _ {t}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 423, + 824, + 784, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "KEEP", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "State Dynamic System. The system dynamics define how the system evolves over time, and it incorporates any control inputs that might affect the current state. The prediction for the state $\\boldsymbol{z}_t$ at the current timestep is achieved via state extrapolation. In particular, given the posterior estimation of previous state $\\hat{\\boldsymbol{z}}_{t-1}^+$ , we define the dynamic model by", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {z}} _ {t} ^ {-} = f \\left(\\hat {\\boldsymbol {z}} _ {t - 1} ^ {+}\\right) = \\mathcal {E} _ {H} \\left(\\omega \\left(\\mathcal {D} _ {Q} \\left(\\hat {\\boldsymbol {z}} _ {t - 1} ^ {+}\\right), \\Phi_ {t - 1 \\rightarrow t}\\right)\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 234, + 785, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\varPhi_{t-1\\rightarrow t}$ denotes the flow estimated from LQ frames $\\pmb{x}_{t-1}$ to $\\pmb{x}_t$ and $\\omega$ the spatial warping modules. Specifically, we first decode the estimated code $\\hat{\\pmb{z}}_{t-1}^{+}$ of the previous frame to get the estimation of $\\hat{\\pmb{y}}_{t-1} = \\mathcal{D}_Q(\\hat{\\pmb{z}}_{t-1}^+)$ and warp it to the current frame. Then, it was encoded back in the latent space to obtain the prediction of the current state $\\hat{\\pmb{z}}_t^{-}$ .", + "bbox": [ + 212, + 263, + 787, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Kalman Filter System. Given the approximated observed state $\\tilde{z}_t$ and prior estimation $\\hat{z}_t^-$ from system dynamics, the filter system aims to promote temporal information propagation and maintain stable latent code priors. In particular, the filter recursively fuses both the estimation to form a more accurate posterior estimate of the current state $\\hat{z}_t^+$ , which is also known as state update.", + "bbox": [ + 212, + 354, + 787, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "According to Eqn. 8, a more intuitive way to express the updated state is a linear interpolation $\\hat{z}_t^+ = \\mathcal{K}_t\\hat{z}_t^- + (1 - \\mathcal{K}_t)\\tilde{z}_t$ for $\\mathcal{K}_t$ normalized in the range of $[0,1]$ . The Kalman Gain $\\mathcal{K}_t$ measures the estimated accuracy of the predicted states compared to the approximated observed states, to update the state and reduce the uncertainty. As illustrated in Fig. 3, the Kalman gain is approximated via Kalman Gain Network (KGN), consisting of two distinct parameterized modules, i.e., uncertainty network and gain network. The uncertainty associated with the current prediction is implicitly estimated by an uncertainty network constructed by spatial-temporal attention [47] and temporal attention layers (or any other recurrent networks). Then, a gain network calculates Kalman gain $\\mathcal{K}_t$ for each code token. Please refer to the supplementary files for the detailed architectures.", + "bbox": [ + 212, + 431, + 789, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Local Temporal Consistency", + "text_level": 1, + "bbox": [ + 215, + 619, + 496, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inspired by [51], we adopt cross-frame attention (CFA) layers in the decoder to further promote local temporal consistency to regularize the information propagation. Specifically, given the latent features from the previous frame $v_{t-1}$ and current frame $v_t$ . They are projected onto the embedding space and output the features $v_i'$ by $v_i' = \\mathrm{Attn}(Q, K, V) = \\mathrm{softmax}\\left(\\frac{QK^T}{\\sqrt{d}}\\right) \\cdot V$ , where", + "bbox": [ + 212, + 643, + 787, + 727 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nQ = W _ {Q} \\cdot v _ {t}, K = W _ {K} \\cdot v _ {t - 1}, V = W _ {V} \\cdot v _ {t - 1}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 738, + 785, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Intuitively, cross-frame attention modules can be regarded as searching and matching similar patches from the previous frame and fusing them correspondingly. This module facilitates temporal information propagation in the decoder. We adopt CFA modules on features of small scale 16 and 32 to avoid introducing blur to the decoded results.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "R. Feng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 145, + 375, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset. VFHQ [48] contains over 15,000 high-quality video clips of diverse interviews and talk shows, where 15,381 clips are used for training and 50 clips are reserved for testing. Each sequence consists of 100 to 900 frames of resolution $512 \\times 512$ . Following common practice [6,36,45,48], we adopt blind settings in all experiments. Specifically, we apply random blur, resize, and noise as image-based degradations. Moreover, video compression is adopted to control the video quality by changing streaming bitrate. For a comprehensive evaluation, we synthesize three splits of the VFHQ-Test dataset containing different levels of degradation, denoted as VFHQ-mild, VFHQ-medium, and VFHQ-heavy. They follow the same degradation model but differ in the degree of noise, blur, and compression. In addition to synthetic data, we also collect real corrupted video for testing. Please refer to the supplementary files for more details.", + "bbox": [ + 212, + 194, + 787, + 376 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Alignment. Pre-trained image models for face restoration are trained on the FFHQ dataset [20], where each image is automatically cropped and aligned. Hence, employing pre-trained models requires a similar alignment phase on VFHQ dataset [48]. However, the discrete step (i.e., cropping) is sensitive to the detected locations of landmarks, which can consequently result in unintentional temporal inconsistencies. Inspired by the work of Fox et al. [11] and Tzaban et al. [43], we employ a Gaussian lowpass filter over the landmarks. We find that this smoothing can significantly attenuate the inconsistencies induced by the alignment step. See supplementary files for more details.", + "bbox": [ + 212, + 378, + 787, + 516 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementations. For all stages of training, we initialize all networks with Kaiming Normal [15] and train them using Adam optimizer [22], and a batch size of 4 for all the experiments. The learning rate is set to $2 \\times 10^{-4}$ for stages I and II, and $1 \\times 10^{-4}$ for stage III. The models are trained with $800k$ , $400k$ , and $50k$ iterations for three stages, respectively. We implement our models with PyTorch [38] and train them using NVIDIA Tesla V100 GPUs. Hyper-parameters $\\lambda_{1}$ , $\\lambda_{VGG}$ , and $\\lambda_{GAN}$ are set to $10^{-2}$ , 1, and $1 \\times 10^{-2}$ . We use GMFlow [49] for optical flow estimation.", + "bbox": [ + 212, + 518, + 787, + 640 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Metrics. For quantitative evaluation, we evaluate the fidelity of restoration using PSNR, SSIM, and LPIPS [54]. In addition, we evaluate the identity preservation scores, termed as $IDS$ , by cosine similarity of the off-the-shelf identity detection network ArcFace [9]. Besides, we follow Tzaban et al. [43] to measure the pose consistency using Average Keypoint Distance (AKD), which is quantified by the average distance of detected landmarks between the generated and ground-truth video frames. In addition to the above single-frame quality evaluation, we also measure the fluctuation of identity/landmarks across frames. Hence, $\\sigma_{IDS}$ measures the standard deviation of identity similarity over the entire video. We expect a considerable identity drift in the generated videos without local identity jitter, where $\\sigma_{IDS}$ is supposed to be low. Similarly, we use $\\sigma_{AKD}$ to measure the standard deviation of keypoint distances over the video, which quantifies the temporal consistency of the pose.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "KEEP", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/5e7eaf8c4efaa15eff7bbd6e7ebd03af808f10c4cc8d596e912eb93e224f4ce1.jpg", + "table_caption": [ + "Table 1: Quantitative comparison on the VFHQ-mild. Red and Blue indicate the best and the second best results. Full results on other test partitions (medium and heavy) are presented in the supplementary material." + ], + "table_footnote": [], + "table_body": "
Method|PSNR↑ SSIM↑ LPIPS↓| IDS↑ AKD↓ |σIDS(×10-2) ↓ σAKD ↓
GPEN [52]25.51930.75170.29880.714211.46914.7416
GFPGAN [45]26.29330.77950.24820.743710.54674.5700
RestoreFormer [46]25.57200.73440.31950.753010.53544.7159
CodeFormer [56]24.65970.74540.27420.627211.49836.3726
EDVR [44]26.60510.78580.24840.719511.62204.8048
BasicVSR [3]26.04580.77650.24960.697311.36795.0343
BasicVSR++ [5]27.19960.80570.19580.764111.31365.2543
KEEP (Ours)27.99940.82670.16190.79608.81823.6866
", + "bbox": [ + 248, + 198, + 754, + 334 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 215, + 375, + 620, + 390 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines. We compare our method to two categories of approaches. i) Image-based Face SR models (CodeFormer [56], GPEN [52], GFPGAN [45], Restore-Former [46]) are used to generate face videos frame-by-frame. ii) We retrain the general VSR models (EDVR [44], BasicVSR [3], BasicVSR++ [5]) on VFHQ dataset [48]. The degradation settings remain unchanged as our experiments while other training settings follow their original papers.", + "bbox": [ + 212, + 433, + 787, + 525 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Quantitative Evaluation. The quantitative results are listed in Table 1. We observe that our method achieves better results than existing methods across all the metrics. The results indicate that KEEP can faithfully recover facial details while preserving the identity. Our method also maintains temporal coherence across frames, as quantified by $\\sigma_{IDS}$ and $\\sigma_{AKD}$ , which represent the fluctuation of restored face identities and facial shapes. Though exhibiting structural distortions and artifacts (See Fig. 4), general VSR models (EDVR, BasicVSR, BasicVSR++) typically achieve higher performance on fidelity metrics (PNSR, SSIM, LPIPS) than single-image FSR models. This suggests that image-based models produce high-quality but relatively low-fidelity results. Inconsistency could be introduced when latent code estimation is noisy and inaccurate.", + "bbox": [ + 212, + 531, + 787, + 698 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Qualitative Evaluation. In Fig. 4, we can observe that the compared methods fail to reconstruct consistent appearances with perceptually pleasant details. For example, GFPGAN tends to hallucinate facial details (e.g., ears in the second frame and incomplete glass in the last frame of the left example). CodeFormer produces unnatural facial shapes (e.g., eyes), and BasicVSR leaves severe artifacts on the face images. In the last frame of the right example, both GFPGAN and CodeFormer generate unpleasant eyes (see yellow arrows). In contrast, our KEEP exploits temporal information and restores finer and coherent facial details. We refer readers to supplementary files for more video results.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "R. Feng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/fc23af2204512c17dda0e651fb9a307526d9781ac2362fa99183d8339e29b254.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparison on the VFHQ. Our KEEP produces high-fidelity face videos with faithful and consistent details. See arrows for details." + ], + "image_footnote": [], + "bbox": [ + 250, + 143, + 759, + 435 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b646ac0f5ad2c369438e50314ba317c27cbe4f15bb101450d802806be25830a2.jpg", + "table_caption": [ + "Table 2: Ablation study of variant networks." + ], + "table_footnote": [], + "table_body": "
ModelsLPIPS↓IDS↑AKD↓
w/o CFA0.16210.79708.9029
w/o KFN0.17210.77739.1952
Full model0.16190.79608.8182
", + "bbox": [ + 248, + 532, + 460, + 601 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/1fa49b1218c0cb7e75d0430f94eae01454010028f424c6443c0f64a7560e3f37.jpg", + "table_caption": [ + "Table 3: Ablation study on optical flow estimator." + ], + "table_footnote": [], + "table_body": "
EstimatorLPIPS↓IDS↑AKD↓
PWC-Net [41]0.16230.79578.7839
GMFlow [49] (Ours)0.16190.79608.8182
", + "bbox": [ + 527, + 532, + 767, + 580 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 More Analysis", + "text_level": 1, + "bbox": [ + 215, + 633, + 382, + 648 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effectiveness of KFN. We first investigate the effectiveness of the Kalman Filter Network (KFN). As shown in Table 2, removing KFN results in worse performance on LPIPS, IDS, and AKD. The results suggest the design of KFN is the key to promoting temporal consistency and identity preservation.", + "bbox": [ + 212, + 671, + 784, + 733 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effectiveness of CFA. Table 2 also shows that Cross-Frame Attention (CFA) can further improve the performance. Though not clearly reflected in the number, we further qualitatively show the effectiveness of KFN and CFA in the supplementary video. From the demo video, we can observe that 1) adopting KFN can ensure consistency in global style and maintain the global appearance of the recovered faces. 2) adding CFA can further coherently render local texture details (e.g., hair), and suppress flicker.", + "bbox": [ + 212, + 734, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "KEEP", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3c1afdfa52b4f2ab68793f92afec9aeba7e4a89e495e11a6bba058a2d9678cac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 145, + 344, + 224 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d9f67801607784ec21ce04e0148a2799d7ff888633b7870cfe3fed703681eaaa.jpg", + "image_caption": [ + "Fig. 5: Comparison of temporal flicker. We select each frame's column (red lines) and show the changes across time. Image-based models (GFPGAN, CodeFormer, and RestoreFormer) have obvious discontinuity around the eyes and wrinkles, and general VSR methods leave artifacts behind. In contrast, by maintaining stable facial priors and aggregating temporal information, our method remarkably suppresses temporal jitters and promotes coherent local details." + ], + "image_footnote": [], + "bbox": [ + 246, + 224, + 344, + 306 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a769f221c8655c2c4d31b4aba16fa4d25a3273f51948f8d8b6c594a201093b86.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 145, + 449, + 224 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/705b2c2e9d5f7f413a65527283e01a4d374813157d95352c859ae2a2cc1e0458.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 224, + 449, + 306 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/370b3538f8035b080f502b022a025b798d7b5e40657439ba67484d75a7c5ca30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 145, + 552, + 224 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5ec4c63cd7e3b9ffa055bdb6d17abef62435df9641965d9b4b6046552a732d3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 224, + 550, + 305 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ff53ce63c05153754f218a588abed31c9199f8170fbaac93d3f96ad3138505e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 552, + 145, + 653, + 224 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/919718bfb026b6fc12898b988e9ae1a595e4cce48ca701e7cfca532704aca6a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 552, + 224, + 653, + 305 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8d8c7915cbbe0a92ac8880633aa446f8d5d2ea164b2915b50770e2b3b8a6dd16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 145, + 756, + 224 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/707ba80d225e97027b97952a3a63fe31a413399557a6bdd53917aeed221c6a5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 224, + 756, + 305 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/636d1b31318759a86d6ff643d192b591408efdfbfe8b843d27b958397d23d519.jpg", + "image_caption": [ + "(a) GT", + "Fig.6: Illustration of predicted state and observed state through Decoder $\\mathcal{D}_Q$ . The teeth in red boxes illustrate a slight difference between $\\tilde{\\boldsymbol{z}}_t$ and $\\hat{\\boldsymbol{z}}_t^-$ when decoded in pixel space, suggesting the potential to supplement and fuse information to obtain a stable latent code." + ], + "image_footnote": [], + "bbox": [ + 245, + 417, + 349, + 500 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c611cdf9682c7466939eb0b3fe82e0b3a5c5446344724cdfe586f9fa853bf429.jpg", + "image_caption": [ + "(b) $\\mathcal{D}_Q(\\tilde{z}_t)$" + ], + "image_footnote": [], + "bbox": [ + 349, + 417, + 450, + 500 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/fffc8b6a527dde660c741a3f521cbdaf92ab36a1c4b342089bf0ee8f9711d4fd.jpg", + "image_caption": [ + "(c) Error map of (b)" + ], + "image_footnote": [], + "bbox": [ + 450, + 417, + 552, + 500 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1a710af11c863fa37ef2a4ae3fc4a48c781cfadeb9563ce21e8a519af6050322.jpg", + "image_caption": [ + "(d) $\\mathcal{D}_Q(\\hat{z}_t^-)$" + ], + "image_footnote": [], + "bbox": [ + 552, + 417, + 653, + 500 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d1906e3de3369fdd83d7d162f71d3c89bf3c44a5bc2ebc81fcc493bb2472804e.jpg", + "image_caption": [ + "(e) Error map of (d)" + ], + "image_footnote": [], + "bbox": [ + 653, + 417, + 756, + 500 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effectiveness of Various Flow Estimator. We compare models with different flow estimators in Table 3. The results suggest that the accuracy of estimated flows does not significantly affect the final performance. We conjecture this can be attributed to two factors: 1) Minor misalignment in pixel space can be reasonably diminished as the latent code is highly downsampled by a factor of $32 \\times$ , at which level the latent representations are less sensitive to small spatial discrepancies present in the pixel space. 2) Other modules can compensate for the inaccuracy caused by flow estimators in a joint training fashion.", + "bbox": [ + 212, + 612, + 787, + 733 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Analysis of Flickering. We extract a short vertical segment of pixels from each frame and stack them horizontally to visualize the jittering issues within the video. In particular, existing methods demonstrate clear jitter and flicker across time, while our method shows better temporal consistency. Fig. 5 demonstrates that other image-based models bring obvious jitters around eyes, and general VSR methods leave behind artifacts, while our method could remarkably suppress temporal jitters and promote coherent local details.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "R. Feng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/17aa6da14616d029dc97340e685d3e6374af2657b6192604c5a6ed073be8682c.jpg", + "image_caption": [ + "Fig. 7: Identity similarity across frames. Our method preserves the identity of input images and exhibits less fluctuation over time. CodeFormer results of two frames (green and red dashed line) exemplify an abrupt change of identity in the right figure, while our method maintains a stable identity both quantitatively and qualitatively." + ], + "image_footnote": [], + "bbox": [ + 258, + 143, + 509, + 262 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9d2cef2b1af797ad5c641485037945a0673da9178bfb937c7557ee2f0d40b110.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 143, + 746, + 263 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/898a5a68770ff2ec7a5586367486dd2c2f5175f4c56ae7106b96fb7b1071d2c9.jpg", + "image_caption": [ + "Fig. 8: Qualitative comparison on different levels of degradation. Our KEEP maintains high-fidelity in various degradations." + ], + "image_footnote": [], + "bbox": [ + 254, + 352, + 751, + 513 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Analysis of Latent Space. Since the true state $\\boldsymbol{z}_t$ is unavailable, we indirectly analyze the predicted state $\\tilde{\\boldsymbol{z}}_t$ and observed state $\\hat{\\boldsymbol{z}}_t^-$ by decoding them to the pixel space through Decoder $\\mathcal{D}_Q$ . As shown in Fig. 6, the areas around teeth exhibit a slight difference, while most remaining parts in error maps show similar decoded results. This indicates that the predicted and observed states could supplement each other to obtain a more accurate estimation of $\\boldsymbol{z}_t$ , which is where the power of Kalman filter lies.", + "bbox": [ + 212, + 583, + 787, + 686 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Identity Preservation. We show the identity similarity across frames of one representative video clip in Fig. 7 As illustrated in the left figure, our method achieves better identity preservation and less identity jitter within the video, compared to the single-image methods. We also exemplify that the identity of CodeFormer results can change abruptly within several frames. The identity score of CodeFormer increases from 0.4552 to 0.6392, and turns down in later frames (See red curve), demonstrating great fluctuation across time. In contrast, our method maintains a stable identity both quantitatively and qualitatively.", + "bbox": [ + 212, + 688, + 787, + 809 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Various Degradations. Fig. 9a shows that KEEP is consistently better than the compared methods across different difficulty levels. Fig. 8 demonstrates re", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "KEEP", + "bbox": [ + 686, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/17c6a6f32090f5ba3e07d3bc6bde30371f59407bebd810a12c91229a3f93bfd0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 150, + 480, + 268 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/458dc2f108eacd6d5c58937491bf6fb7d9cd762d2c2e8eb54214ca17d6797b14.jpg", + "image_caption": [ + "(a) PNSR and IDS on different partitions." + ], + "image_footnote": [], + "bbox": [ + 232, + 277, + 480, + 395 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1f233c901cc9f5fd087b4ef429260c4c91ab67f0c8b6417530003c1809221778.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 157, + 782, + 234 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ab73be52ebc623cbb65365755e714283da6f4571c9967723c2590042cbcbb88a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 234, + 782, + 309 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1ebcfac9ccb6174376d8145b59177eff45dafab47885fee29070080894b41b2c.jpg", + "image_caption": [ + "(b) Comparisons on non-frontal faces.", + "Fig. 9: (a) Our methods achieve consistently better performance on various levels of degradation. (b) While CodeFormer fails to restore eyes to these challenging cases, our method can still produce plausible facial elements." + ], + "image_footnote": [], + "bbox": [ + 496, + 310, + 782, + 395 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "sults on different levels of degradation. We draw the following observations. 1) Even GFPGAN and CodeFormer can restore plausible results on frames with mild degradation, the performance significantly deteriorates (e.g., eyes) upon heavier degradation. 2) Our method achieves better results by considering complementary information between adjacent frames and maintaining stable face priors. In addition, our method is appealing in handling heavy degradation.", + "bbox": [ + 212, + 500, + 787, + 590 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Non-Frontal-View Faces. In Fig. 9b, our model shows enhanced performance on non-frontal faces by providing more stable face priors estimations. While the single-image model CodeFormer cannot recover the eyes, our KEEP is still able to show robustness to these challenging cases.", + "bbox": [ + 212, + 592, + 787, + 654 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 681, + 359, + 698 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We present a novel framework, KEEP, aiming at resolving the challenges associated with facial detail and temporal consistency in video face restoration. The proposed method demonstrates a unique capability to maintain a stable face prior over time, which is achieved by Kalman filtering principles, where our approach recurrently incorporates information from previously restored frames to guide and regulate the restoration process of the current frame. Extensive experiments demonstrate the efficacy of KEEP in consistently capturing facial details across video frames and keeping the temporal stability of face videos.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "R. Feng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 217, + 143, + 382, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).", + "bbox": [ + 212, + 176, + 787, + 222 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 243, + 321, + 258 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Caballero, J., Ledig, C., Aitken, A., Acosta, A., Totz, J., Wang, Z., Shi, W.: Real-time video super-resolution with spatio-temporal networks and motion compensation. In: CVPR (2017)", + "2. Chan, K.C., Wang, X., Xu, X., Gu, J., Loy, C.C.: GLEAN: Generative latent bank for large-factor image super-resolution. In: CVPR (2021)", + "3. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: BasicVSR: The search for essential components in video super-resolution and beyond. In: CVPR (2021)", + "4. Chan, K.C., Xu, X., Wang, X., Gu, J., Loy, C.C.: GLEAN: Generative latent bank for large-factor image super-resolution and beyond. TPAMI (2022)", + "5. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: CVPR (2022)", + "6. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Investigating tradeoffs in real-world video super-resolution. In: CVPR (2022)", + "7. Chen, C., Li, X., Yang, L., Lin, X., Zhang, L., Wong, K.Y.K.: Progressive semantic-aware style transformation for blind face restoration. In: CVPR (2021)", + "8. Chen, Y., Tai, Y., Liu, X., Shen, C., Yang, J.: FSRNet: End-to-end learning face super-resolution with facial priors. In: CVPR (2018)", + "9. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: ArcFace: Additive angular margin loss for deep face recognition. In: CVPR (2019)", + "10. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: CVPR (2021)", + "1. Fox, G., Tewari, A., Elgharib, M., Theobalt, C.: StyleVideoGAN: A temporal generative model using a pretrained stylegan. In: BMVC (2021)", + "2. Gu, J., Shen, Y., Zhou, B.: Image processing using multi-code gan prior. In: CVPR (2020)", + "3. Gu, Y., Wang, X., Xie, L., Dong, C., Li, G., Shan, Y., Cheng, M.M.: VQFR: Blind face restoration with vector-quantized dictionary and parallel decoder. In: ECCV (2022)", + "4. Haris, M., Shakhnarovich, G., Ukita, N.: Recurrent back-projection network for video super-resolution. In: CVPR (2019)", + "5. He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In: ICCV (2015)", + "6. Huang, Y., Wang, W., Wang, L.: Bidirectional recurrent convolutional networks for multi-frame super-resolution. In: NeurIPS (2015)", + "7. Isobe, T., Jia, X., Gu, S., Li, S., Wang, S., Tian, Q.: Video super-resolution with recurrent structure-detail network. In: ECCV (2020)", + "8. Isobe, T., Zhu, F., Wang, S.: Revisiting temporal modeling for video superresolution. In: BMVC (2020)", + "9. Kalman, R.E.: A new approach to linear filtering and prediction problems (1960)", + "20. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019)" + ], + "bbox": [ + 225, + 273, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "KEEP", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "21. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: CVPR (2020)", + "22. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015)", + "23. Lei, C., Ren, X., Zhang, Z., Chen, Q.: Blind video deflickering by neural filtering with a flawed atlas. In: CVPR (2023)", + "24. Lei, C., Xing, Y., Chen, Q.: Blind video temporal consistency via deep video prior. In: NeurIPS (2020)", + "25. Li, S., He, F., Du, B., Zhang, L., Xu, Y., Tao, D.: Fast spatio-temporal residual network for video super-resolution. In: CVPR (2019)", + "26. Li, X., Chen, C., Zhou, S., Lin, X., Zuo, W., Zhang, L.: Blind face restoration via deep multi-scale component dictionaries. In: ECCV (2020)", + "27. Li, X., Li, W., Ren, D., Zhang, H., Wang, M., Zuo, W.: Enhanced blind face restoration with multi-exemplar images and adaptive spatial feature fusion. In: CVPR (2020)", + "28. Li, X., Liu, M., Ye, Y., Zuo, W., Lin, L., Yang, R.: Learning warped guidance for blind face restoration. In: ECCV (2018)", + "29. Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: Vrt: A video restoration transformer. IEEE Transactions on Image Processing (2024)", + "30. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided deformable attention. NeurIPS (2022)", + "31. Liu, C., Yang, H., Fu, J., Qian, X.: Learning trajectory-aware transformer for video super-resolution. In: CVPR (2022)", + "32. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: CVPR (2022)", + "33. Menon, S., Damian, A., Hu, S., Ravi, N., Rudin, C.: Pulse: Self-supervised photoupsampling via latent space exploration of generative models. In: CVPR (2020)", + "34. Nagrani, A., Chung, J.S., Zisserman, A.: Voxceleb: A large-scale speaker identification dataset. In: Interspeech (2017)", + "35. Ouyang, H., Wang, Q., Xiao, Y., Bai, Q., Zhang, J., Zheng, K., Zhou, X., Chen, Q., Shen, Y.: Codef: Content deformation fields for temporally consistent video processing. In: CVPR (2024)", + "36. Pan, J., Bai, H., Dong, J., Zhang, J., Tang, J.: Deep blind video super-resolution. In: ICCV (2021)", + "37. Pan, X., Zhan, X., Dai, B., Lin, D., Loy, C.C., Luo, P.: Exploiting deep generative prior for versatile image restoration and manipulation. TPAMI 44(11), 7474-7489 (2021)", + "38. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017)", + "39. Revach, G., Shlezinger, N., Ni, X., Escoriza, A.L., Van Sloun, R.J., Eldar, Y.C.: Kalmannet: Neural network aided kalman filtering for partially known dynamics. IEEE Transactions on Signal Processing (2022)", + "40. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)", + "41. Sun, D., Yang, X., Liu, M.Y., Kautz, J.: Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In: CVPR (2018)", + "42. Tian, Y., Zhang, Y., Fu, Y., Xu, C.: TDAN: Temporally-deformable alignment network for video super-resolution. In: CVPR (2020)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "R. Feng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Tzaban, R., Mokady, R., Gal, R., Bermano, A., Cohen-Or, D.: Stitch it in time: Gan-based facial editing of real videos. In: SIGGRAPH Asia (2022)", + "44. Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: EDVR: Video restoration with enhanced deformable convolutional networks. In: CVPRW (2019)", + "45. Wang, X., Li, Y., Zhang, H., Shan, Y.: Towards real-world blind face restoration with generative facial prior. In: CVPR (2021)", + "46. Wang, Z., Zhang, J., Chen, R., Wang, W., Luo, P.: Restoreformer: High-quality blind face restoration from undegraded key-value pairs. In: CVPR (2022)", + "47. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: ICCV (2023)", + "48. Xie, L., Wang, X., Zhang, H., Dong, C., Shan, Y.: VFHQ: A high-quality dataset and benchmark for video face super-resolution. In: CVPR (2022)", + "49. Xu, H., Zhang, J., Cai, J., Rezatofighi, H., Tao, D.: Gmflow: Learning optical flow via global matching. In: CVPR (2022)", + "50. Xu, Y., AlBahar, B., Huang, J.B.: Temporally consistent semantic video editing. In: ECCV (2022)", + "51. Yang, S., Zhou, Y., Liu, Z.,, Loy, C.C.: Rerender a video: Zero-shot text-guided video-to-video translation. In: SIGGRAPH Asia (2023)", + "52. Yang, T., Ren, P., Xie, X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: CVPR (2021)", + "53. Yu, X., Fernando, B., Ghanem, B., Porikli, F., Hartley, R.: Face super-resolution guided by facial component heatmaps. In: ECCV (2018)", + "54. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)", + "55. Zhou, L., Luo, Z., Shen, T., Zhang, J., Zhen, M., Yao, Y., Fang, T., Quan, L.: KFNet: Learning temporal camera relocalization using Kalman filtering. In: CVPR (2020)", + "56. Zhou, S., Chan, K.C., Li, C., Loy, C.C.: Towards robust blind face restoration with codebook lookup transformer. In: NeurIPS (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 561 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "KEEP", + "bbox": [ + 684, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Kalman-Inspired Feature Propagation for Video Face Super-Resolution/86d2672e-c2f0-471a-a7a6-101117941b3f_model.json b/2024/Kalman-Inspired Feature Propagation for Video Face Super-Resolution/86d2672e-c2f0-471a-a7a6-101117941b3f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2f0d5530c9cab209f8bc68a6a48efc0794f28cd9 --- /dev/null +++ b/2024/Kalman-Inspired Feature Propagation for Video Face Super-Resolution/86d2672e-c2f0-471a-a7a6-101117941b3f_model.json @@ -0,0 +1,2500 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.266, + 0.141, + 0.737, + 0.186 + ], + "angle": 0, + "content": "Kalman-Inspired Feature Propagation for Video Face Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.304, + 0.213, + 0.701, + 0.228 + ], + "angle": 0, + "content": "Ruicheng Feng, Chongyi Li, and Chen Change Loy" + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.24, + 0.675, + 0.282 + ], + "angle": 0, + "content": "S-Lab, Nanyang Technological University, Singapore {ruicheng002, ccloy}@ntu.edu.sg lichongyi25@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.318, + 0.741, + 0.525 + ], + "angle": 0, + "content": "Abstract. Despite the promising progress of face image super-resolution, video face super-resolution remains relatively under-explored. Existing approaches either adapt general video super-resolution networks to face datasets or apply established face image super-resolution models independently on individual video frames. These paradigms encounter challenges either in reconstructing facial details or maintaining temporal consistency. To address these issues, we introduce a novel framework called Kalman-inspired Feature Propagation (KEEP), designed to maintain a stable face prior over time. The Kalman filtering principles offer our method a recurrent ability to use the information from previously restored frames to guide and regulate the restoration process of the current frame. Extensive experiments demonstrate the effectiveness of our method in capturing facial details consistently across video frames. Code and video demo are available at https://jnjaby.github.io/projects/KEEP/." + }, + { + "type": "title", + "bbox": [ + 0.22, + 0.553, + 0.373, + 0.568 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.583, + 0.786, + 0.764 + ], + "angle": 0, + "content": "The field of Face Super-Resolution (FSR), which focuses on reconstructing high-resolution (HR) face images from highly degraded versions, has witnessed remarkable progress. In particular, numerous studies have successfully leveraged various types of prior information, such as geometric facial priors [7,8,53], reference priors [26-28], generative priors [2,4,45,52], and codebook priors [13,46,56]. These approaches have significantly advanced the realism and quality of generated face images. However, the majority of these studies are confined to still images, with the extension to Video Face Super-Resolution (VFSR) remaining relatively under-explored. Despite the substantial potential benefits of video face restoration in various practical domains, such as the restoration of old films, VFSR has yet to receive the same level of attention and development as its image-based counterpart." + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.766, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Two main strategies emerge for implementing VFSR. The first approach involves adapting general Video Super-Resolution (VSR) networks, such as EDVR [44], BasicVSR [3], BasicVSR++ [5], and RVRT [30], to large-scale face video datasets [34, 48]. These methods exploit temporal information and propagate features across video frames. However, they are not specifically tailored for" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "R. Feng et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.144, + 0.782, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.395, + 0.788, + 0.466 + ], + "angle": 0, + "content": "Fig. 1: Comparing main VFSR strategies. We show seven frames with an interval of 6. Generic VSR model BasicVSR [3] fails to reconstruct facial components faithfully. Single-image FSR model CodeFormer [56] hallucinates unnatural and inconsistent face details. Our method, in contrast, enables consistent restoration of low-quality face video while preserving temporal coherence across frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.788, + 0.655 + ], + "angle": 0, + "content": "face restoration and often fall short in reconstructing detailed facial features, particularly in severely degraded scenarios, as depicted in Fig. 1(second row). The second method involves applying existing face image SR models to process each video frame independently. This frame-by-frame approach, while straightforward, introduces temporal inconsistencies in the video, as demonstrated in Fig. 1(third row). This problem arises because FSR is inherently ill-posed and the existing priors may not suffice to maintain appearance consistency throughout the video sequence. Specifically, a single degraded image could correspond to multiple high-resolution interpretations, leading to discrepancies and inconsistent structures across independently processed video frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this paper, we wish to devise an effective framework for maintaining a stable face prior over time for VFSR. We use CodeFormer [56], a representative model that exploits codebook priors for FSR, to demonstrate how face priors can be consistently preserved across different time frames. A pivotal aspect of our approach is the idea that frames previously restored can act as references, guiding and regulating the restoration process of the current frame. This strategy helps minimize the divergence between consecutive frames. Moreover, this reliance on previously restored frames naturally suggests a recurrent framework, enabling the effective use of information from past restorations. This intuition aligns closely with Kalman filtering principles, or linear quadratic estimation, which involves using a sequence of time-based measurements, despite their statistical noise and inaccuracies, to produce more accurate estimates of unknown" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KEEP" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.193 + ], + "angle": 0, + "content": "variables than would be possible with a single measurement. Similarly, in VFSR, faces observed over time are often noisy and inaccurate, making them suitable for refinement using Kalman filtering techniques." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.403 + ], + "angle": 0, + "content": "Driven by these insights, we formulate a novel method, Kalman-inspired fEaturE Propagation (KEEP), which recurrently updates the current latent state in CodeFormer by incorporating information from preceding frames. This method of temporal propagation within the latent space ensures the stability of the face prior over time, thereby capturing facial details that consistently match in appearance. The effectiveness of KEEP is shown in Fig. 1, where it is evident that our method delivers high-quality restoration with superior consistency, compared to both generic video restoration methods fine-tuned for faces and approaches that restore frames independently. Please refer to the supplementary video to appreciate the superiority of our approach in terms of temporal consistency. A key advantage of KEEP is its robustness in handling severe video-based degradation, outperforming single-image models. In addition, our model demonstrates enhanced performance on non-frontal faces by providing more stable estimations of face priors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.541 + ], + "angle": 0, + "content": "In summary, the main contribution of this work is a novel framework for maintaining a stable and meaningful face prior for VFSR. While we demonstrate its application using the CodeFormer method as a case study, the underlying principles of our framework, inspired by the Kalman filtering approach, are applicable to other approaches. Extensive experimental results on the VFHQ dataset [48] and real-world data demonstrate the effectiveness of our approach in improving both the fidelity and coherence of VFSR outputs. Compared to other state-of-the-art methods, our model achieves superior performance with a large margin of 0.8 dB in PSNR, while also significantly maintaining temporal coherence." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.565, + 0.388, + 0.58 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Blind Face Restoration. Blind face restoration aims at recovering severely degraded face images in the wild. Unlike natural images, faces are highly structured. This property allows researchers to incorporate prior information into the restoration models, which have demonstrated remarkable progress in capabilities to restore high-quality faces. Most existing FSR methods can be categorized into four classes: geometric priors, reference priors, generative priors, and codebook priors. Geometric priors usually include facial elements, such as face landmarks [8], parsing maps [7], and facial component heatmaps [53]. Another major line is reference-based methods that require high-quality exemplar images. GFRNet [28] and ASFFNet [27] leverage a warped high-quality image to extract rich details to improve facial detail restoration. DFDNet [26] constructs deep dictionaries with facial components from large-scale images to recover fine details. Generative priors from pre-trained GAN, e.g., StyleGAN [20] and StyleGAN2 [21], are employed through iterative latent optimization of GAN inversion [12, 33, 37]. Still, they produce face images with low fidelity and are computationally expensive. To address this, GLEAN [2], GPEN [52], and GFP-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "R. Feng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "GAN [45] integrate generative priors into encoder-decoder architectures, which estimate latent priors in one-forward pass. These methods achieve great tradeoff between quality and fidelity but usually fail when the corruption is severe. Codebook priors [13,46,56] can be regarded as a special case of generative priors. In contrast to continuous generative priors, they squeeze the latent space into a small finite codebook space and improve the robustness to severe degradation. However, most existing FSR methods are image-based and thus they cannot guarantee temporally consistent details for VFSR." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.275, + 0.789, + 0.683 + ], + "angle": 0, + "content": "Video Super-Resolution. Most existing video restoration techniques can be categorized into two paradigms based on their parallelizability: parallel and recurrent methods. Parallel models estimate all frames simultaneously, and the restoration of each frame does not rely on the update of other frames. These methods typically involve feature extraction, feature alignment, feature fusion, and reconstruction. Representative works, FSTRN [25] and VESCPN [1], introduce fast spatio-temporal networks using 3D convolutions to enhance alignment, combining motion compensation and super-resolution. EDVR [44] and TDAN [42] leverage deformable convolutions for aligning adjacent frames. Besides, Transformer-based methods [29,32] are proposed to reconstruct all frames simultaneously by jointly extracting, aligning, and fusing features. In addition, RVRT [30] and TTVSR [31] integrate optical flow into Transformer and enable long-range models in videos. Empowered by the great expressive capability of Transformer, this line of work exhibits remarkable performance improvements over previous methods. However, they suffer from large model sizes and high memory consumption. Recurrent methods [3,16,18] do not aggregate information solely from adjacent frames. Instead, they maintain hidden states to convey relevant information from previous frames and propagate latent features sequentially, accumulating information for later restoration. For example, RBPN [14] treated each frame as a separate source, combined iteratively in a refinement framework. RSDN [17] divided the input into structure and detail components, proposing a two-stream structure-detail block to learn textures. BasicVSR [3] and BasicVSR++ [5] fused bidirectional hidden states from both past and future frames, bring significant improvements. They aim to fully utilize information from the entire sequence, synchronously updating the hidden state through the weights of the reconstruction network. Due to the recurrent nature of feature propagation, recurrent methods experience information loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.79, + 0.84 + ], + "angle": 0, + "content": "Video Consistency. In addition to VSR, previous works also attempt to inflate image models into video models [43, 47, 50, 51] or improve temporal consistency with the implicit representation of the given videos [23, 24, 35]. For example, Stitch-it-in-Time [43] discovers locally consistent pivots in the latent space to provide spatially consistent transitions. Tune-A-Video [47] adopts cross-frame attention and fine-tune it on a single video. All-in-One deflicker [23] learns a neural atlas for each video to solve long-term inconsistency. These works either require post-processing or optimization on a video basis. In contrast, we aim at repurposing image FSR models and enforcing temporal consistency in the restored face videos." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KEEP" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.381, + 0.163 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.326 + ], + "angle": 0, + "content": "In a corrupted face video, local textures and facial details are irrevocably lost. Therefore, the latent codes, usually estimated by an encoder, are inaccurate to match the real underlying priors of ground-truth video. Different from image-based autoencoder, models in the video domain enable us to exploit evidence accumulated from preceding frames for better restoration. In this paper, we repurpose the image-based CodeFormer for video FSR and propose KEEP to estimate stable face priors in latent space over time given the noisy and inaccurate estimations, which consequently enable temporally coherent restoration. An appealing idea to realize KEEP is through reformulating the method in a Kalman Filter framework." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.346, + 0.364, + 0.359 + ], + "angle": 0, + "content": "3.1 Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.442 + ], + "angle": 0, + "content": "State Space Model. We consider observations of low-quality (LQ) video sequence \\( X = \\{\\pmb{x}_{t}\\}_{t=1}^{T} \\) with length of \\( T \\), where \\( \\pmb{x}_t \\in \\mathbb{R}^{H \\times W \\times 3} \\), and underlying high-quality (HQ) sequences \\( Y = \\{\\pmb{y}_t\\}_{t=1}^T \\). Kalman filter [19] assumes linear dynamic systems that are characterized by a state space model driven by Gaussian noise" + }, + { + "type": "equation", + "bbox": [ + 0.437, + 0.445, + 0.786, + 0.46 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {t} = \\boldsymbol {F} _ {t} \\boldsymbol {y} _ {t - 1} + \\boldsymbol {q} _ {t}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.786, + 0.495 + ], + "angle": 0, + "content": "where \\(\\pmb{F}_t\\) is the transition matrix and \\(\\pmb{q}_t\\) denotes process noise drawn from Gaussian noise. The observation \\(\\pmb{x}_t\\) is measured by" + }, + { + "type": "equation", + "bbox": [ + 0.447, + 0.505, + 0.785, + 0.52 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t} = \\boldsymbol {H} \\boldsymbol {y} _ {t} + \\boldsymbol {r} _ {t}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.786, + 0.573 + ], + "angle": 0, + "content": "where \\(\\pmb{H}\\) is the measurement matrix and \\(\\pmb{r}_t\\) represents measurement noise. However, the linear assumption does not hold in some complex real-world scenarios. Hence, the non-linear Kalman filter can be reformulated as" + }, + { + "type": "equation", + "bbox": [ + 0.444, + 0.583, + 0.785, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {t} = d \\left(\\boldsymbol {y} _ {t - 1}, \\boldsymbol {q} _ {t}\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.444, + 0.603, + 0.785, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t} = h \\left(\\boldsymbol {y} _ {t}\\right) + \\boldsymbol {r} _ {t}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.788, + 0.716 + ], + "angle": 0, + "content": "where \\( d(\\cdot) \\) and \\( h(\\cdot) \\) are non-linear transition and measurement models. Specifically, \\( d(\\cdot) \\) can be represented by any explicit motion estimation (e.g., optical flow), which defines how the current frame transits to the next one. \\( h(\\cdot) \\) commonly models the degradation in video restoration problems. As opposed to the classical assumptions in Kalman filter, the measurement function \\( h(\\cdot) \\) is unknown in a blind setting. This is regarded as partially known dynamic models [39]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.717, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Inspired by VQGAN [10] and Stable Diffusion [40], we estimate underlying latent representations \\( Z = \\{z_{t}\\}_{t=1}^{T} \\) such that \\( z_{t} \\) can correspond to \\( y_{t} \\) by a generative model \\( g_{\\theta} \\), given by" + }, + { + "type": "equation", + "bbox": [ + 0.459, + 0.772, + 0.785, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {t} = g _ {\\theta} \\left(\\boldsymbol {z} _ {t}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Instead of directly estimating individual pixels, modeling the low-dimensional latent code is computationally more efficient and focuses on more perceptually significant variations of the data. The graphical model is depicted in Fig. 2 (a)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "R. Feng et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.145, + 0.445, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.323, + 0.266, + 0.339, + 0.277 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.147, + 0.777, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.267, + 0.612, + 0.277 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.293, + 0.789, + 0.419 + ], + "angle": 0, + "content": "Fig.2: (a) Graphical model of state space. It defines the underlying dynamic system model, where \\( f \\) describes how the latent states \\( z_{t} \\) transit over time, \\( g \\) is a generative model, and \\( h \\) models the degradation from clean frame \\( y_{t} \\) to degraded frame \\( x_{t} \\). (b) Block diagram of Kalman filter model. In each time step, a predictive state from previous frame \\( \\hat{z}_{t-1}^{+} \\) (Blue dash box) and new observed state of current frame \\( x_{t} \\) (Red dash box) are fused by Kalman gain \\( \\mathcal{K}_t \\) from Kalman Gain Network (KGN) to produce more accurate estimates. The combined state \\( \\hat{z}_t^+ \\) is then used to generate the estimated clean frame \\( \\hat{y}_t \\) by \\( g_{\\theta} \\). Note that \\( \\tilde{z}_1 \\) goes along with \\( \\tilde{z}_{t-1} \\) as an anchor and it is omitted in the diagram for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.448, + 0.789, + 0.508 + ], + "angle": 0, + "content": "Kalman Filter Model. The principles of Kalman filter can be formulated by a two-step procedure, i.e., state prediction and state update. The overview diagram is illustrated in Fig. 2 (b). In this problem, the observation is a face image \\( \\mathbf{x}_t \\) and the state is \\( \\mathbf{z}_t \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.509, + 0.79, + 0.569 + ], + "angle": 0, + "content": "In the state prediction step, the model predicts the prior estimation \\(\\hat{z}_t^-\\) of the current state \\(z_{t}\\) based on the posterior estimation \\(\\hat{z}_{t - 1}^{+}\\) of the previous state and the dynamic model. Specifically, the prior estimation of latent state \\(z_{t}\\) and estimation of observation \\(\\pmb{x}_t\\) are computed as" + }, + { + "type": "equation", + "bbox": [ + 0.439, + 0.579, + 0.785, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {t} ^ {-} = f \\left(\\hat {z} _ {t - 1} ^ {+}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.44, + 0.601, + 0.785, + 0.618 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {x}} _ {t} ^ {-} = h \\left(g _ {\\theta} \\left(\\hat {\\boldsymbol {z}} _ {t} ^ {-}\\right)\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.628, + 0.787, + 0.658 + ], + "angle": 0, + "content": "The system dynamics \\( f \\) define how the latent state \\( Z \\) evolves over time, and it incorporates any control inputs that might affect the current state." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.658, + 0.787, + 0.689 + ], + "angle": 0, + "content": "In the state update step, a posterior state estimation \\(\\hat{z}_t^+\\) is computed based on the prior estimation and new observations \\(\\pmb{x}_t\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.431, + 0.698, + 0.785, + 0.716 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {t} ^ {+} = \\hat {z} _ {t} ^ {-} + \\mathcal {K} _ {t} \\Delta z _ {t}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.727, + 0.788, + 0.773 + ], + "angle": 0, + "content": "where \\(\\mathcal{K}_t\\) is conceptually referred to as the Kalman gain and \\(\\Delta z_{t}\\) as the innovation, i.e., the residual between the prior estimation \\(\\hat{\\pmb{z}}_t^-\\) and approximation of current state from \\(\\pmb{x}_{t}\\), given by" + }, + { + "type": "equation", + "bbox": [ + 0.444, + 0.784, + 0.785, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\Delta \\boldsymbol {z} _ {t} = \\hat {\\boldsymbol {z}} _ {t} ^ {-} - \\tilde {\\boldsymbol {z}} _ {t}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "where \\(\\tilde{\\pmb{z}}_t = e(\\pmb{x}_t)\\). Note that this formula differs from the original Kalman filter which minimizes the innovation of \\(\\Delta \\pmb{x}_t\\) (i.e., residual between \\(\\pmb{x}_t\\) and \\(\\hat{\\pmb{x}}_t^{-}\\)), since" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KEEP" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.243, + 0.147, + 0.756, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.314, + 0.785, + 0.357 + ], + "angle": 0, + "content": "Fig. 3: Overview of the proposed KEEP. It consists of four modules: encoder \\(\\mathcal{E}_L\\), decoder \\(\\mathcal{D}_Q\\), Kalman filter network, and CFA. We illustrate the information flow in one timestep." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.784, + 0.461 + ], + "angle": 0, + "content": "the assumption of available measurement function \\( h(\\cdot) \\) is no longer valid in our setting. An original measurement system models how observation \\( \\boldsymbol{x}_t \\) is derived from the latent state \\( \\boldsymbol{z}_t \\), formally \\( \\boldsymbol{x}_t = h(\\boldsymbol{z}_t) \\). Inspired by KFNet [55], we directly estimate the state \\( \\boldsymbol{z}_t \\) given new observation \\( \\boldsymbol{x}_t \\), i.e., mapping \\( \\boldsymbol{x}_t \\) to \\( \\tilde{\\boldsymbol{z}}_t \\) with an estimator \\( e \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.785, + 0.585 + ], + "angle": 0, + "content": "The remaining problem is to compute the Kalman gain \\(\\mathcal{K}_t\\). As discussed in KalmanNet [39], covariance estimation is intractable when dealing with high-dimensional signals. Additionally, the second-order statistical moments are only used for calculating Kalman gain. Inspired by this, we directly learn the gains from the data distribution and do not explicitly maintain an estimation of covariances. Additionally, we follow [51] to include \\(\\tilde{\\pmb{z}}_1\\) of the first frame as anchor into KGN for Kalman gain estimation. Then, the final predicted \\(\\hat{\\pmb{y}}_t\\) can be derived by" + }, + { + "type": "equation", + "bbox": [ + 0.453, + 0.586, + 0.785, + 0.603 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {y}} _ {t} = g _ {\\theta} \\left(\\hat {\\boldsymbol {z}} _ {t} ^ {+}\\right). \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.628, + 0.447, + 0.642 + ], + "angle": 0, + "content": "3.2 Parameterized Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.784, + 0.703 + ], + "angle": 0, + "content": "As shown in Fig. 2 (b), we will parameterize or define the system dynamics \\( f \\), observation estimator \\( e \\), Kalman Gain Nets (KGN), and generative model \\( g_{\\theta} \\) in this section. The overall framework is shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.72, + 0.784, + 0.811 + ], + "angle": 0, + "content": "Generative Model. The generative model \\( g_{\\theta} \\) is generally parameterized by a backbone of CodeFormer [56], which consists of a LQ encoder \\( \\mathcal{E}_L \\), a HQ encoder \\( \\mathcal{E}_H \\), a codebook lookup Transformer and quantization layer \\( T_{Q} \\), and a decoder \\( \\mathcal{D} \\). For simplicity, \\( \\mathcal{D}_Q \\) denotes the decoder with the codebook lookup Transformer and quantization layer absorbed. Basically, the predicted \\( \\hat{\\pmb{y}}_t \\) is given by \\( \\mathcal{D}_Q(\\hat{\\pmb{z}}_t^+) \\), and the observed state \\( \\tilde{\\pmb{z}}_t \\) is approximated by" + }, + { + "type": "equation", + "bbox": [ + 0.424, + 0.825, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\tilde {z} _ {t} = e \\left(\\boldsymbol {x} _ {t}\\right) = \\mathcal {E} _ {L} \\left(\\boldsymbol {x} _ {t}\\right). \\tag {11}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "R. Feng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "State Dynamic System. The system dynamics define how the system evolves over time, and it incorporates any control inputs that might affect the current state. The prediction for the state \\( \\boldsymbol{z}_t \\) at the current timestep is achieved via state extrapolation. In particular, given the posterior estimation of previous state \\( \\hat{\\boldsymbol{z}}_{t-1}^+ \\), we define the dynamic model by" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.236, + 0.786, + 0.253 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {z}} _ {t} ^ {-} = f \\left(\\hat {\\boldsymbol {z}} _ {t - 1} ^ {+}\\right) = \\mathcal {E} _ {H} \\left(\\omega \\left(\\mathcal {D} _ {Q} \\left(\\hat {\\boldsymbol {z}} _ {t - 1} ^ {+}\\right), \\Phi_ {t - 1 \\rightarrow t}\\right)\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.264, + 0.789, + 0.342 + ], + "angle": 0, + "content": "where \\(\\varPhi_{t-1\\rightarrow t}\\) denotes the flow estimated from LQ frames \\(\\pmb{x}_{t-1}\\) to \\(\\pmb{x}_t\\) and \\(\\omega\\) the spatial warping modules. Specifically, we first decode the estimated code \\(\\hat{\\pmb{z}}_{t-1}^{+}\\) of the previous frame to get the estimation of \\(\\hat{\\pmb{y}}_{t-1} = \\mathcal{D}_Q(\\hat{\\pmb{z}}_{t-1}^+)\\) and warp it to the current frame. Then, it was encoded back in the latent space to obtain the prediction of the current state \\(\\hat{\\pmb{z}}_t^{-}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.789, + 0.431 + ], + "angle": 0, + "content": "Kalman Filter System. Given the approximated observed state \\(\\tilde{z}_t\\) and prior estimation \\(\\hat{z}_t^-\\) from system dynamics, the filter system aims to promote temporal information propagation and maintain stable latent code priors. In particular, the filter recursively fuses both the estimation to form a more accurate posterior estimate of the current state \\(\\hat{z}_t^+\\), which is also known as state update." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.79, + 0.598 + ], + "angle": 0, + "content": "According to Eqn. 8, a more intuitive way to express the updated state is a linear interpolation \\(\\hat{z}_t^+ = \\mathcal{K}_t\\hat{z}_t^- + (1 - \\mathcal{K}_t)\\tilde{z}_t\\) for \\(\\mathcal{K}_t\\) normalized in the range of \\([0,1]\\). The Kalman Gain \\(\\mathcal{K}_t\\) measures the estimated accuracy of the predicted states compared to the approximated observed states, to update the state and reduce the uncertainty. As illustrated in Fig. 3, the Kalman gain is approximated via Kalman Gain Network (KGN), consisting of two distinct parameterized modules, i.e., uncertainty network and gain network. The uncertainty associated with the current prediction is implicitly estimated by an uncertainty network constructed by spatial-temporal attention [47] and temporal attention layers (or any other recurrent networks). Then, a gain network calculates Kalman gain \\(\\mathcal{K}_t\\) for each code token. Please refer to the supplementary files for the detailed architectures." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.62, + 0.497, + 0.636 + ], + "angle": 0, + "content": "3.3 Local Temporal Consistency" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.728 + ], + "angle": 0, + "content": "Inspired by [51], we adopt cross-frame attention (CFA) layers in the decoder to further promote local temporal consistency to regularize the information propagation. Specifically, given the latent features from the previous frame \\( v_{t-1} \\) and current frame \\( v_t \\). They are projected onto the embedding space and output the features \\( v_i' \\) by \\( v_i' = \\mathrm{Attn}(Q, K, V) = \\mathrm{softmax}\\left(\\frac{QK^T}{\\sqrt{d}}\\right) \\cdot V \\), where" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.739, + 0.786, + 0.754 + ], + "angle": 0, + "content": "\\[\nQ = W _ {Q} \\cdot v _ {t}, K = W _ {K} \\cdot v _ {t - 1}, V = W _ {V} \\cdot v _ {t - 1}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Intuitively, cross-frame attention modules can be regarded as searching and matching similar patches from the previous frame and fusing them correspondingly. This module facilitates temporal information propagation in the decoder. We adopt CFA modules on features of small scale 16 and 32 to avoid introducing blur to the decoded results." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KEEP" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.146, + 0.377, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.195, + 0.788, + 0.377 + ], + "angle": 0, + "content": "Dataset. VFHQ [48] contains over 15,000 high-quality video clips of diverse interviews and talk shows, where 15,381 clips are used for training and 50 clips are reserved for testing. Each sequence consists of 100 to 900 frames of resolution \\(512 \\times 512\\). Following common practice [6,36,45,48], we adopt blind settings in all experiments. Specifically, we apply random blur, resize, and noise as image-based degradations. Moreover, video compression is adopted to control the video quality by changing streaming bitrate. For a comprehensive evaluation, we synthesize three splits of the VFHQ-Test dataset containing different levels of degradation, denoted as VFHQ-mild, VFHQ-medium, and VFHQ-heavy. They follow the same degradation model but differ in the degree of noise, blur, and compression. In addition to synthetic data, we also collect real corrupted video for testing. Please refer to the supplementary files for more details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.788, + 0.517 + ], + "angle": 0, + "content": "Alignment. Pre-trained image models for face restoration are trained on the FFHQ dataset [20], where each image is automatically cropped and aligned. Hence, employing pre-trained models requires a similar alignment phase on VFHQ dataset [48]. However, the discrete step (i.e., cropping) is sensitive to the detected locations of landmarks, which can consequently result in unintentional temporal inconsistencies. Inspired by the work of Fox et al. [11] and Tzaban et al. [43], we employ a Gaussian lowpass filter over the landmarks. We find that this smoothing can significantly attenuate the inconsistencies induced by the alignment step. See supplementary files for more details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.788, + 0.641 + ], + "angle": 0, + "content": "Implementations. For all stages of training, we initialize all networks with Kaiming Normal [15] and train them using Adam optimizer [22], and a batch size of 4 for all the experiments. The learning rate is set to \\(2 \\times 10^{-4}\\) for stages I and II, and \\(1 \\times 10^{-4}\\) for stage III. The models are trained with \\(800k\\), \\(400k\\), and \\(50k\\) iterations for three stages, respectively. We implement our models with PyTorch [38] and train them using NVIDIA Tesla V100 GPUs. Hyper-parameters \\(\\lambda_{1}\\), \\(\\lambda_{VGG}\\), and \\(\\lambda_{GAN}\\) are set to \\(10^{-2}\\), 1, and \\(1 \\times 10^{-2}\\). We use GMFlow [49] for optical flow estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Metrics. For quantitative evaluation, we evaluate the fidelity of restoration using PSNR, SSIM, and LPIPS [54]. In addition, we evaluate the identity preservation scores, termed as \\( IDS \\), by cosine similarity of the off-the-shelf identity detection network ArcFace [9]. Besides, we follow Tzaban et al. [43] to measure the pose consistency using Average Keypoint Distance (AKD), which is quantified by the average distance of detected landmarks between the generated and ground-truth video frames. In addition to the above single-frame quality evaluation, we also measure the fluctuation of identity/landmarks across frames. Hence, \\( \\sigma_{IDS} \\) measures the standard deviation of identity similarity over the entire video. We expect a considerable identity drift in the generated videos without local identity jitter, where \\( \\sigma_{IDS} \\) is supposed to be low. Similarly, we use \\( \\sigma_{AKD} \\) to measure the standard deviation of keypoint distances over the video, which quantifies the temporal consistency of the pose." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "R. Feng et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison on the VFHQ-mild. Red and Blue indicate the best and the second best results. Full results on other test partitions (medium and heavy) are presented in the supplementary material." + }, + { + "type": "table", + "bbox": [ + 0.249, + 0.199, + 0.756, + 0.335 + ], + "angle": 0, + "content": "
Method|PSNR↑ SSIM↑ LPIPS↓| IDS↑ AKD↓ |σIDS(×10-2) ↓ σAKD ↓
GPEN [52]25.51930.75170.29880.714211.46914.7416
GFPGAN [45]26.29330.77950.24820.743710.54674.5700
RestoreFormer [46]25.57200.73440.31950.753010.53544.7159
CodeFormer [56]24.65970.74540.27420.627211.49836.3726
EDVR [44]26.60510.78580.24840.719511.62204.8048
BasicVSR [3]26.04580.77650.24960.697311.36795.0343
BasicVSR++ [5]27.19960.80570.19580.764111.31365.2543
KEEP (Ours)27.99940.82670.16190.79608.81823.6866
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.376, + 0.621, + 0.391 + ], + "angle": 0, + "content": "4.1 Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.526 + ], + "angle": 0, + "content": "Baselines. We compare our method to two categories of approaches. i) Image-based Face SR models (CodeFormer [56], GPEN [52], GFPGAN [45], Restore-Former [46]) are used to generate face videos frame-by-frame. ii) We retrain the general VSR models (EDVR [44], BasicVSR [3], BasicVSR++ [5]) on VFHQ dataset [48]. The degradation settings remain unchanged as our experiments while other training settings follow their original papers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.532, + 0.788, + 0.699 + ], + "angle": 0, + "content": "Quantitative Evaluation. The quantitative results are listed in Table 1. We observe that our method achieves better results than existing methods across all the metrics. The results indicate that KEEP can faithfully recover facial details while preserving the identity. Our method also maintains temporal coherence across frames, as quantified by \\(\\sigma_{IDS}\\) and \\(\\sigma_{AKD}\\), which represent the fluctuation of restored face identities and facial shapes. Though exhibiting structural distortions and artifacts (See Fig. 4), general VSR models (EDVR, BasicVSR, BasicVSR++) typically achieve higher performance on fidelity metrics (PNSR, SSIM, LPIPS) than single-image FSR models. This suggests that image-based models produce high-quality but relatively low-fidelity results. Inconsistency could be introduced when latent code estimation is noisy and inaccurate." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Qualitative Evaluation. In Fig. 4, we can observe that the compared methods fail to reconstruct consistent appearances with perceptually pleasant details. For example, GFPGAN tends to hallucinate facial details (e.g., ears in the second frame and incomplete glass in the last frame of the left example). CodeFormer produces unnatural facial shapes (e.g., eyes), and BasicVSR leaves severe artifacts on the face images. In the last frame of the right example, both GFPGAN and CodeFormer generate unpleasant eyes (see yellow arrows). In contrast, our KEEP exploits temporal information and restores finer and coherent facial details. We refer readers to supplementary files for more video results." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KEEP" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.144, + 0.761, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.445, + 0.785, + 0.473 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparison on the VFHQ. Our KEEP produces high-fidelity face videos with faithful and consistent details. See arrows for details." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.493, + 0.492, + 0.52 + ], + "angle": 0, + "content": "Table 2: Ablation study of variant networks." + }, + { + "type": "table", + "bbox": [ + 0.249, + 0.534, + 0.461, + 0.602 + ], + "angle": 0, + "content": "
ModelsLPIPS↓IDS↑AKD↓
w/o CFA0.16210.79708.9029
w/o KFN0.17210.77739.1952
Full model0.16190.79608.8182
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.493, + 0.785, + 0.52 + ], + "angle": 0, + "content": "Table 3: Ablation study on optical flow estimator." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.534, + 0.769, + 0.582 + ], + "angle": 0, + "content": "
EstimatorLPIPS↓IDS↑AKD↓
PWC-Net [41]0.16230.79578.7839
GMFlow [49] (Ours)0.16190.79608.8182
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.635, + 0.383, + 0.65 + ], + "angle": 0, + "content": "4.2 More Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.672, + 0.785, + 0.734 + ], + "angle": 0, + "content": "Effectiveness of KFN. We first investigate the effectiveness of the Kalman Filter Network (KFN). As shown in Table 2, removing KFN results in worse performance on LPIPS, IDS, and AKD. The results suggest the design of KFN is the key to promoting temporal consistency and identity preservation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Effectiveness of CFA. Table 2 also shows that Cross-Frame Attention (CFA) can further improve the performance. Though not clearly reflected in the number, we further qualitatively show the effectiveness of KFN and CFA in the supplementary video. From the demo video, we can observe that 1) adopting KFN can ensure consistency in global style and maintain the global appearance of the recovered faces. 2) adding CFA can further coherently render local texture details (e.g., hair), and suppress flicker." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "R. Feng et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.146, + 0.346, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.226, + 0.346, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.348, + 0.146, + 0.45, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.348, + 0.226, + 0.45, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.146, + 0.553, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.226, + 0.552, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.146, + 0.654, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.226, + 0.654, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.146, + 0.757, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.226, + 0.757, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.32, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Fig. 5: Comparison of temporal flicker. We select each frame's column (red lines) and show the changes across time. Image-based models (GFPGAN, CodeFormer, and RestoreFormer) have obvious discontinuity around the eyes and wrinkles, and general VSR methods leave artifacts behind. In contrast, by maintaining stable facial priors and aggregating temporal information, our method remarkably suppresses temporal jitters and promotes coherent local details." + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.419, + 0.35, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.501, + 0.313, + 0.51 + ], + "angle": 0, + "content": "(a) GT" + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.419, + 0.451, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.378, + 0.501, + 0.422, + 0.512 + ], + "angle": 0, + "content": "(b) \\(\\mathcal{D}_Q(\\tilde{z}_t)\\)" + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.419, + 0.553, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.461, + 0.501, + 0.54, + 0.512 + ], + "angle": 0, + "content": "(c) Error map of (b)" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.419, + 0.655, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.501, + 0.627, + 0.512 + ], + "angle": 0, + "content": "(d) \\(\\mathcal{D}_Q(\\hat{z}_t^-)\\)" + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.419, + 0.757, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.501, + 0.743, + 0.511 + ], + "angle": 0, + "content": "(e) Error map of (d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.526, + 0.788, + 0.581 + ], + "angle": 0, + "content": "Fig.6: Illustration of predicted state and observed state through Decoder \\(\\mathcal{D}_Q\\). The teeth in red boxes illustrate a slight difference between \\(\\tilde{\\boldsymbol{z}}_t\\) and \\(\\hat{\\boldsymbol{z}}_t^-\\) when decoded in pixel space, suggesting the potential to supplement and fuse information to obtain a stable latent code." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Effectiveness of Various Flow Estimator. We compare models with different flow estimators in Table 3. The results suggest that the accuracy of estimated flows does not significantly affect the final performance. We conjecture this can be attributed to two factors: 1) Minor misalignment in pixel space can be reasonably diminished as the latent code is highly downsampled by a factor of \\(32 \\times\\), at which level the latent representations are less sensitive to small spatial discrepancies present in the pixel space. 2) Other modules can compensate for the inaccuracy caused by flow estimators in a joint training fashion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Analysis of Flickering. We extract a short vertical segment of pixels from each frame and stack them horizontally to visualize the jittering issues within the video. In particular, existing methods demonstrate clear jitter and flicker across time, while our method shows better temporal consistency. Fig. 5 demonstrates that other image-based models bring obvious jitters around eyes, and general VSR methods leave behind artifacts, while our method could remarkably suppress temporal jitters and promote coherent local details." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.687, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KEEP" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.144, + 0.51, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.144, + 0.747, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.279, + 0.788, + 0.336 + ], + "angle": 0, + "content": "Fig. 7: Identity similarity across frames. Our method preserves the identity of input images and exhibits less fluctuation over time. CodeFormer results of two frames (green and red dashed line) exemplify an abrupt change of identity in the right figure, while our method maintains a stable identity both quantitatively and qualitatively." + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.353, + 0.753, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.525, + 0.785, + 0.553 + ], + "angle": 0, + "content": "Fig. 8: Qualitative comparison on different levels of degradation. Our KEEP maintains high-fidelity in various degradations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.688 + ], + "angle": 0, + "content": "Analysis of Latent Space. Since the true state \\( \\boldsymbol{z}_t \\) is unavailable, we indirectly analyze the predicted state \\( \\tilde{\\boldsymbol{z}}_t \\) and observed state \\( \\hat{\\boldsymbol{z}}_t^- \\) by decoding them to the pixel space through Decoder \\( \\mathcal{D}_Q \\). As shown in Fig. 6, the areas around teeth exhibit a slight difference, while most remaining parts in error maps show similar decoded results. This indicates that the predicted and observed states could supplement each other to obtain a more accurate estimation of \\( \\boldsymbol{z}_t \\), which is where the power of Kalman filter lies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Identity Preservation. We show the identity similarity across frames of one representative video clip in Fig. 7 As illustrated in the left figure, our method achieves better identity preservation and less identity jitter within the video, compared to the single-image methods. We also exemplify that the identity of CodeFormer results can change abruptly within several frames. The identity score of CodeFormer increases from 0.4552 to 0.6392, and turns down in later frames (See red curve), demonstrating great fluctuation across time. In contrast, our method maintains a stable identity both quantitatively and qualitatively." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Various Degradations. Fig. 9a shows that KEEP is consistently better than the compared methods across different difficulty levels. Fig. 8 demonstrates re" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "R. Feng et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.151, + 0.481, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.279, + 0.482, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.403, + 0.478, + 0.415 + ], + "angle": 0, + "content": "(a) PNSR and IDS on different partitions." + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.158, + 0.783, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.235, + 0.783, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.311, + 0.783, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.403, + 0.751, + 0.415 + ], + "angle": 0, + "content": "(b) Comparisons on non-frontal faces." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.427, + 0.788, + 0.469 + ], + "angle": 0, + "content": "Fig. 9: (a) Our methods achieve consistently better performance on various levels of degradation. (b) While CodeFormer fails to restore eyes to these challenging cases, our method can still produce plausible facial elements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.788, + 0.592 + ], + "angle": 0, + "content": "sults on different levels of degradation. We draw the following observations. 1) Even GFPGAN and CodeFormer can restore plausible results on frames with mild degradation, the performance significantly deteriorates (e.g., eyes) upon heavier degradation. 2) Our method achieves better results by considering complementary information between adjacent frames and maintaining stable face priors. In addition, our method is appealing in handling heavy degradation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.593, + 0.788, + 0.655 + ], + "angle": 0, + "content": "Non-Frontal-View Faces. In Fig. 9b, our model shows enhanced performance on non-frontal faces by providing more stable face priors estimations. While the single-image model CodeFormer cannot recover the eyes, our KEEP is still able to show robustness to these challenging cases." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.682, + 0.36, + 0.699 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We present a novel framework, KEEP, aiming at resolving the challenges associated with facial detail and temporal consistency in video face restoration. The proposed method demonstrates a unique capability to maintain a stable face prior over time, which is achieved by Kalman filtering principles, where our approach recurrently incorporates information from previously restored frames to guide and regulate the restoration process of the current frame. Extensive experiments demonstrate the efficacy of KEEP in consistently capturing facial details across video frames and keeping the temporal stability of face videos." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KEEP" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.383, + 0.163 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.223 + ], + "angle": 0, + "content": "This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.244, + 0.323, + 0.26 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.787, + 0.317 + ], + "angle": 0, + "content": "1. Caballero, J., Ledig, C., Aitken, A., Acosta, A., Totz, J., Wang, Z., Shi, W.: Real-time video super-resolution with spatio-temporal networks and motion compensation. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.787, + 0.344 + ], + "angle": 0, + "content": "2. Chan, K.C., Wang, X., Xu, X., Gu, J., Loy, C.C.: GLEAN: Generative latent bank for large-factor image super-resolution. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.346, + 0.787, + 0.372 + ], + "angle": 0, + "content": "3. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: BasicVSR: The search for essential components in video super-resolution and beyond. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.373, + 0.787, + 0.4 + ], + "angle": 0, + "content": "4. Chan, K.C., Xu, X., Wang, X., Gu, J., Loy, C.C.: GLEAN: Generative latent bank for large-factor image super-resolution and beyond. TPAMI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.401, + 0.787, + 0.427 + ], + "angle": 0, + "content": "5. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.428, + 0.787, + 0.454 + ], + "angle": 0, + "content": "6. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Investigating tradeoffs in real-world video super-resolution. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.456, + 0.787, + 0.482 + ], + "angle": 0, + "content": "7. Chen, C., Li, X., Yang, L., Lin, X., Zhang, L., Wong, K.Y.K.: Progressive semantic-aware style transformation for blind face restoration. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.483, + 0.787, + 0.51 + ], + "angle": 0, + "content": "8. Chen, Y., Tai, Y., Liu, X., Shen, C., Yang, J.: FSRNet: End-to-end learning face super-resolution with facial priors. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.787, + 0.537 + ], + "angle": 0, + "content": "9. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: ArcFace: Additive angular margin loss for deep face recognition. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.538, + 0.787, + 0.565 + ], + "angle": 0, + "content": "10. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.566, + 0.787, + 0.592 + ], + "angle": 0, + "content": "1. Fox, G., Tewari, A., Elgharib, M., Theobalt, C.: StyleVideoGAN: A temporal generative model using a pretrained stylegan. In: BMVC (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.594, + 0.787, + 0.62 + ], + "angle": 0, + "content": "2. Gu, J., Shen, Y., Zhou, B.: Image processing using multi-code gan prior. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.621, + 0.787, + 0.661 + ], + "angle": 0, + "content": "3. Gu, Y., Wang, X., Xie, L., Dong, C., Li, G., Shan, Y., Cheng, M.M.: VQFR: Blind face restoration with vector-quantized dictionary and parallel decoder. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.662, + 0.787, + 0.689 + ], + "angle": 0, + "content": "4. Haris, M., Shakhnarovich, G., Ukita, N.: Recurrent back-projection network for video super-resolution. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.69, + 0.787, + 0.716 + ], + "angle": 0, + "content": "5. He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In: ICCV (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.787, + 0.744 + ], + "angle": 0, + "content": "6. Huang, Y., Wang, W., Wang, L.: Bidirectional recurrent convolutional networks for multi-frame super-resolution. In: NeurIPS (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.787, + 0.771 + ], + "angle": 0, + "content": "7. Isobe, T., Jia, X., Gu, S., Li, S., Wang, S., Tian, Q.: Video super-resolution with recurrent structure-detail network. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "8. Isobe, T., Zhu, F., Wang, S.: Revisiting temporal modeling for video superresolution. In: BMVC (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.787, + 0.813 + ], + "angle": 0, + "content": "9. Kalman, R.E.: A new approach to linear filtering and prediction problems (1960)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.787, + 0.84 + ], + "angle": 0, + "content": "20. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.274, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "R. Feng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "21. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "22. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "23. Lei, C., Ren, X., Zhang, Z., Chen, Q.: Blind video deflickering by neural filtering with a flawed atlas. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "24. Lei, C., Xing, Y., Chen, Q.: Blind video temporal consistency via deep video prior. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "25. Li, S., He, F., Du, B., Zhang, L., Xu, Y., Tao, D.: Fast spatio-temporal residual network for video super-resolution. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "26. Li, X., Chen, C., Zhou, S., Lin, X., Zuo, W., Zhang, L.: Blind face restoration via deep multi-scale component dictionaries. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "27. Li, X., Li, W., Ren, D., Zhang, H., Wang, M., Zuo, W.: Enhanced blind face restoration with multi-exemplar images and adaptive spatial feature fusion. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.383 + ], + "angle": 0, + "content": "28. Li, X., Liu, M., Ye, Y., Zuo, W., Lin, L., Yang, R.: Learning warped guidance for blind face restoration. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "29. Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: Vrt: A video restoration transformer. IEEE Transactions on Image Processing (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "30. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided deformable attention. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "31. Liu, C., Yang, H., Fu, J., Qian, X.: Learning trajectory-aware transformer for video super-resolution. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "32. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.55 + ], + "angle": 0, + "content": "33. Menon, S., Damian, A., Hu, S., Ravi, N., Rudin, C.: Pulse: Self-supervised photoupsampling via latent space exploration of generative models. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.551, + 0.785, + 0.577 + ], + "angle": 0, + "content": "34. Nagrani, A., Chung, J.S., Zisserman, A.: Voxceleb: A large-scale speaker identification dataset. In: Interspeech (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.619 + ], + "angle": 0, + "content": "35. Ouyang, H., Wang, Q., Xiao, Y., Bai, Q., Zhang, J., Zheng, K., Zhou, X., Chen, Q., Shen, Y.: Codef: Content deformation fields for temporally consistent video processing. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "36. Pan, J., Bai, H., Dong, J., Zhang, J., Tang, J.: Deep blind video super-resolution. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "37. Pan, X., Zhan, X., Dai, B., Lin, D., Loy, C.C., Luo, P.: Exploiting deep generative prior for versatile image restoration and manipulation. TPAMI 44(11), 7474-7489 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "38. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.785, + 0.757 + ], + "angle": 0, + "content": "39. Revach, G., Shlezinger, N., Ni, X., Escoriza, A.L., Van Sloun, R.J., Eldar, Y.C.: Kalmannet: Neural network aided kalman filtering for partially known dynamics. IEEE Transactions on Signal Processing (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "40. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "41. Sun, D., Yang, X., Liu, M.Y., Kautz, J.: Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "42. Tian, Y., Zhang, Y., Fu, Y., Xu, C.: TDAN: Temporally-deformable alignment network for video super-resolution. In: CVPR (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.686, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KEEP" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "43. Tzaban, R., Mokady, R., Gal, R., Bermano, A., Cohen-Or, D.: Stitch it in time: Gan-based facial editing of real videos. In: SIGGRAPH Asia (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "44. Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: EDVR: Video restoration with enhanced deformable convolutional networks. In: CVPRW (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "45. Wang, X., Li, Y., Zhang, H., Shan, Y.: Towards real-world blind face restoration with generative facial prior. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "46. Wang, Z., Zhang, J., Chen, R., Wang, W., Luo, P.: Restoreformer: High-quality blind face restoration from undegraded key-value pairs. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "47. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.328 + ], + "angle": 0, + "content": "48. Xie, L., Wang, X., Zhang, H., Dong, C., Shan, Y.: VFHQ: A high-quality dataset and benchmark for video face super-resolution. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.785, + 0.356 + ], + "angle": 0, + "content": "49. Xu, H., Zhang, J., Cai, J., Rezatofighi, H., Tao, D.: Gmflow: Learning optical flow via global matching. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.384 + ], + "angle": 0, + "content": "50. Xu, Y., AlBahar, B., Huang, J.B.: Temporally consistent semantic video editing. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.385, + 0.785, + 0.411 + ], + "angle": 0, + "content": "51. Yang, S., Zhou, Y., Liu, Z.,, Loy, C.C.: Rerender a video: Zero-shot text-guided video-to-video translation. In: SIGGRAPH Asia (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.438 + ], + "angle": 0, + "content": "52. Yang, T., Ren, P., Xie, X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.785, + 0.466 + ], + "angle": 0, + "content": "53. Yu, X., Fernando, B., Ghanem, B., Porikli, F., Hartley, R.: Face super-resolution guided by facial component heatmaps. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.785, + 0.494 + ], + "angle": 0, + "content": "54. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.535 + ], + "angle": 0, + "content": "55. Zhou, L., Luo, Z., Shen, T., Zhang, J., Zhen, M., Yao, Y., Fang, T., Quan, L.: KFNet: Learning temporal camera relocalization using Kalman filtering. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.563 + ], + "angle": 0, + "content": "56. Zhou, S., Chan, K.C., Li, C., Loy, C.C.: Towards robust blind face restoration with codebook lookup transformer. In: NeurIPS (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.563 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Kalman-Inspired Feature Propagation for Video Face Super-Resolution/86d2672e-c2f0-471a-a7a6-101117941b3f_origin.pdf b/2024/Kalman-Inspired Feature Propagation for Video Face Super-Resolution/86d2672e-c2f0-471a-a7a6-101117941b3f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4b27c5e39da5e5515b79edfbb7422a6e2056847e --- /dev/null +++ b/2024/Kalman-Inspired Feature Propagation for Video Face Super-Resolution/86d2672e-c2f0-471a-a7a6-101117941b3f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4fb082265c8015a89a221047cd853d3fdb52306f66d0216efe86570584473be6 +size 8233597 diff --git a/2024/Kalman-Inspired Feature Propagation for Video Face Super-Resolution/full.md b/2024/Kalman-Inspired Feature Propagation for Video Face Super-Resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..308271804cd06dfd63322aa3840435198bf38127 --- /dev/null +++ b/2024/Kalman-Inspired Feature Propagation for Video Face Super-Resolution/full.md @@ -0,0 +1,332 @@ +# Kalman-Inspired Feature Propagation for Video Face Super-Resolution + +Ruicheng Feng, Chongyi Li, and Chen Change Loy + +S-Lab, Nanyang Technological University, Singapore {ruicheng002, ccloy}@ntu.edu.sg lichongyi25@gmail.com + +Abstract. Despite the promising progress of face image super-resolution, video face super-resolution remains relatively under-explored. Existing approaches either adapt general video super-resolution networks to face datasets or apply established face image super-resolution models independently on individual video frames. These paradigms encounter challenges either in reconstructing facial details or maintaining temporal consistency. To address these issues, we introduce a novel framework called Kalman-inspired Feature Propagation (KEEP), designed to maintain a stable face prior over time. The Kalman filtering principles offer our method a recurrent ability to use the information from previously restored frames to guide and regulate the restoration process of the current frame. Extensive experiments demonstrate the effectiveness of our method in capturing facial details consistently across video frames. Code and video demo are available at https://jnjaby.github.io/projects/KEEP/. + +# 1 Introduction + +The field of Face Super-Resolution (FSR), which focuses on reconstructing high-resolution (HR) face images from highly degraded versions, has witnessed remarkable progress. In particular, numerous studies have successfully leveraged various types of prior information, such as geometric facial priors [7,8,53], reference priors [26-28], generative priors [2,4,45,52], and codebook priors [13,46,56]. These approaches have significantly advanced the realism and quality of generated face images. However, the majority of these studies are confined to still images, with the extension to Video Face Super-Resolution (VFSR) remaining relatively under-explored. Despite the substantial potential benefits of video face restoration in various practical domains, such as the restoration of old films, VFSR has yet to receive the same level of attention and development as its image-based counterpart. + +Two main strategies emerge for implementing VFSR. The first approach involves adapting general Video Super-Resolution (VSR) networks, such as EDVR [44], BasicVSR [3], BasicVSR++ [5], and RVRT [30], to large-scale face video datasets [34, 48]. These methods exploit temporal information and propagate features across video frames. However, they are not specifically tailored for + +![](images/d0990cb8d6bd312c9850737ee484d46a046f3d917329cb81e114178078088ed4.jpg) +Fig. 1: Comparing main VFSR strategies. We show seven frames with an interval of 6. Generic VSR model BasicVSR [3] fails to reconstruct facial components faithfully. Single-image FSR model CodeFormer [56] hallucinates unnatural and inconsistent face details. Our method, in contrast, enables consistent restoration of low-quality face video while preserving temporal coherence across frames. + +face restoration and often fall short in reconstructing detailed facial features, particularly in severely degraded scenarios, as depicted in Fig. 1(second row). The second method involves applying existing face image SR models to process each video frame independently. This frame-by-frame approach, while straightforward, introduces temporal inconsistencies in the video, as demonstrated in Fig. 1(third row). This problem arises because FSR is inherently ill-posed and the existing priors may not suffice to maintain appearance consistency throughout the video sequence. Specifically, a single degraded image could correspond to multiple high-resolution interpretations, leading to discrepancies and inconsistent structures across independently processed video frames. + +In this paper, we wish to devise an effective framework for maintaining a stable face prior over time for VFSR. We use CodeFormer [56], a representative model that exploits codebook priors for FSR, to demonstrate how face priors can be consistently preserved across different time frames. A pivotal aspect of our approach is the idea that frames previously restored can act as references, guiding and regulating the restoration process of the current frame. This strategy helps minimize the divergence between consecutive frames. Moreover, this reliance on previously restored frames naturally suggests a recurrent framework, enabling the effective use of information from past restorations. This intuition aligns closely with Kalman filtering principles, or linear quadratic estimation, which involves using a sequence of time-based measurements, despite their statistical noise and inaccuracies, to produce more accurate estimates of unknown + +variables than would be possible with a single measurement. Similarly, in VFSR, faces observed over time are often noisy and inaccurate, making them suitable for refinement using Kalman filtering techniques. + +Driven by these insights, we formulate a novel method, Kalman-inspired fEaturE Propagation (KEEP), which recurrently updates the current latent state in CodeFormer by incorporating information from preceding frames. This method of temporal propagation within the latent space ensures the stability of the face prior over time, thereby capturing facial details that consistently match in appearance. The effectiveness of KEEP is shown in Fig. 1, where it is evident that our method delivers high-quality restoration with superior consistency, compared to both generic video restoration methods fine-tuned for faces and approaches that restore frames independently. Please refer to the supplementary video to appreciate the superiority of our approach in terms of temporal consistency. A key advantage of KEEP is its robustness in handling severe video-based degradation, outperforming single-image models. In addition, our model demonstrates enhanced performance on non-frontal faces by providing more stable estimations of face priors. + +In summary, the main contribution of this work is a novel framework for maintaining a stable and meaningful face prior for VFSR. While we demonstrate its application using the CodeFormer method as a case study, the underlying principles of our framework, inspired by the Kalman filtering approach, are applicable to other approaches. Extensive experimental results on the VFHQ dataset [48] and real-world data demonstrate the effectiveness of our approach in improving both the fidelity and coherence of VFSR outputs. Compared to other state-of-the-art methods, our model achieves superior performance with a large margin of 0.8 dB in PSNR, while also significantly maintaining temporal coherence. + +# 2 Related Work + +Blind Face Restoration. Blind face restoration aims at recovering severely degraded face images in the wild. Unlike natural images, faces are highly structured. This property allows researchers to incorporate prior information into the restoration models, which have demonstrated remarkable progress in capabilities to restore high-quality faces. Most existing FSR methods can be categorized into four classes: geometric priors, reference priors, generative priors, and codebook priors. Geometric priors usually include facial elements, such as face landmarks [8], parsing maps [7], and facial component heatmaps [53]. Another major line is reference-based methods that require high-quality exemplar images. GFRNet [28] and ASFFNet [27] leverage a warped high-quality image to extract rich details to improve facial detail restoration. DFDNet [26] constructs deep dictionaries with facial components from large-scale images to recover fine details. Generative priors from pre-trained GAN, e.g., StyleGAN [20] and StyleGAN2 [21], are employed through iterative latent optimization of GAN inversion [12, 33, 37]. Still, they produce face images with low fidelity and are computationally expensive. To address this, GLEAN [2], GPEN [52], and GFP- + +GAN [45] integrate generative priors into encoder-decoder architectures, which estimate latent priors in one-forward pass. These methods achieve great tradeoff between quality and fidelity but usually fail when the corruption is severe. Codebook priors [13,46,56] can be regarded as a special case of generative priors. In contrast to continuous generative priors, they squeeze the latent space into a small finite codebook space and improve the robustness to severe degradation. However, most existing FSR methods are image-based and thus they cannot guarantee temporally consistent details for VFSR. + +Video Super-Resolution. Most existing video restoration techniques can be categorized into two paradigms based on their parallelizability: parallel and recurrent methods. Parallel models estimate all frames simultaneously, and the restoration of each frame does not rely on the update of other frames. These methods typically involve feature extraction, feature alignment, feature fusion, and reconstruction. Representative works, FSTRN [25] and VESCPN [1], introduce fast spatio-temporal networks using 3D convolutions to enhance alignment, combining motion compensation and super-resolution. EDVR [44] and TDAN [42] leverage deformable convolutions for aligning adjacent frames. Besides, Transformer-based methods [29,32] are proposed to reconstruct all frames simultaneously by jointly extracting, aligning, and fusing features. In addition, RVRT [30] and TTVSR [31] integrate optical flow into Transformer and enable long-range models in videos. Empowered by the great expressive capability of Transformer, this line of work exhibits remarkable performance improvements over previous methods. However, they suffer from large model sizes and high memory consumption. Recurrent methods [3,16,18] do not aggregate information solely from adjacent frames. Instead, they maintain hidden states to convey relevant information from previous frames and propagate latent features sequentially, accumulating information for later restoration. For example, RBPN [14] treated each frame as a separate source, combined iteratively in a refinement framework. RSDN [17] divided the input into structure and detail components, proposing a two-stream structure-detail block to learn textures. BasicVSR [3] and BasicVSR++ [5] fused bidirectional hidden states from both past and future frames, bring significant improvements. They aim to fully utilize information from the entire sequence, synchronously updating the hidden state through the weights of the reconstruction network. Due to the recurrent nature of feature propagation, recurrent methods experience information loss. + +Video Consistency. In addition to VSR, previous works also attempt to inflate image models into video models [43, 47, 50, 51] or improve temporal consistency with the implicit representation of the given videos [23, 24, 35]. For example, Stitch-it-in-Time [43] discovers locally consistent pivots in the latent space to provide spatially consistent transitions. Tune-A-Video [47] adopts cross-frame attention and fine-tune it on a single video. All-in-One deflicker [23] learns a neural atlas for each video to solve long-term inconsistency. These works either require post-processing or optimization on a video basis. In contrast, we aim at repurposing image FSR models and enforcing temporal consistency in the restored face videos. + +# 3 Methodology + +In a corrupted face video, local textures and facial details are irrevocably lost. Therefore, the latent codes, usually estimated by an encoder, are inaccurate to match the real underlying priors of ground-truth video. Different from image-based autoencoder, models in the video domain enable us to exploit evidence accumulated from preceding frames for better restoration. In this paper, we repurpose the image-based CodeFormer for video FSR and propose KEEP to estimate stable face priors in latent space over time given the noisy and inaccurate estimations, which consequently enable temporally coherent restoration. An appealing idea to realize KEEP is through reformulating the method in a Kalman Filter framework. + +# 3.1 Formulation + +State Space Model. We consider observations of low-quality (LQ) video sequence $X = \{\pmb{x}_{t}\}_{t=1}^{T}$ with length of $T$ , where $\pmb{x}_t \in \mathbb{R}^{H \times W \times 3}$ , and underlying high-quality (HQ) sequences $Y = \{\pmb{y}_t\}_{t=1}^T$ . Kalman filter [19] assumes linear dynamic systems that are characterized by a state space model driven by Gaussian noise + +$$ +\boldsymbol {y} _ {t} = \boldsymbol {F} _ {t} \boldsymbol {y} _ {t - 1} + \boldsymbol {q} _ {t}, \tag {1} +$$ + +where $\pmb{F}_t$ is the transition matrix and $\pmb{q}_t$ denotes process noise drawn from Gaussian noise. The observation $\pmb{x}_t$ is measured by + +$$ +\boldsymbol {x} _ {t} = \boldsymbol {H} \boldsymbol {y} _ {t} + \boldsymbol {r} _ {t}, \tag {2} +$$ + +where $\pmb{H}$ is the measurement matrix and $\pmb{r}_t$ represents measurement noise. However, the linear assumption does not hold in some complex real-world scenarios. Hence, the non-linear Kalman filter can be reformulated as + +$$ +\boldsymbol {y} _ {t} = d \left(\boldsymbol {y} _ {t - 1}, \boldsymbol {q} _ {t}\right) \tag {3} +$$ + +$$ +\boldsymbol {x} _ {t} = h \left(\boldsymbol {y} _ {t}\right) + \boldsymbol {r} _ {t}, \tag {4} +$$ + +where $d(\cdot)$ and $h(\cdot)$ are non-linear transition and measurement models. Specifically, $d(\cdot)$ can be represented by any explicit motion estimation (e.g., optical flow), which defines how the current frame transits to the next one. $h(\cdot)$ commonly models the degradation in video restoration problems. As opposed to the classical assumptions in Kalman filter, the measurement function $h(\cdot)$ is unknown in a blind setting. This is regarded as partially known dynamic models [39]. + +Inspired by VQGAN [10] and Stable Diffusion [40], we estimate underlying latent representations $Z = \{z_{t}\}_{t=1}^{T}$ such that $z_{t}$ can correspond to $y_{t}$ by a generative model $g_{\theta}$ , given by + +$$ +\boldsymbol {y} _ {t} = g _ {\theta} \left(\boldsymbol {z} _ {t}\right). \tag {5} +$$ + +Instead of directly estimating individual pixels, modeling the low-dimensional latent code is computationally more efficient and focuses on more perceptually significant variations of the data. The graphical model is depicted in Fig. 2 (a). + +![](images/b24db573ef9e18763e701385155b15c194b9262dbc57157da9c67a056394191a.jpg) +(a) + +![](images/3243931ce86ba68f7e0872371ad01dfa3bde02ef51ec1347563adb3fd76e6268.jpg) +(b) +Fig.2: (a) Graphical model of state space. It defines the underlying dynamic system model, where $f$ describes how the latent states $z_{t}$ transit over time, $g$ is a generative model, and $h$ models the degradation from clean frame $y_{t}$ to degraded frame $x_{t}$ . (b) Block diagram of Kalman filter model. In each time step, a predictive state from previous frame $\hat{z}_{t-1}^{+}$ (Blue dash box) and new observed state of current frame $x_{t}$ (Red dash box) are fused by Kalman gain $\mathcal{K}_t$ from Kalman Gain Network (KGN) to produce more accurate estimates. The combined state $\hat{z}_t^+$ is then used to generate the estimated clean frame $\hat{y}_t$ by $g_{\theta}$ . Note that $\tilde{z}_1$ goes along with $\tilde{z}_{t-1}$ as an anchor and it is omitted in the diagram for simplicity. + +Kalman Filter Model. The principles of Kalman filter can be formulated by a two-step procedure, i.e., state prediction and state update. The overview diagram is illustrated in Fig. 2 (b). In this problem, the observation is a face image $\mathbf{x}_t$ and the state is $\mathbf{z}_t$ . + +In the state prediction step, the model predicts the prior estimation $\hat{z}_t^-$ of the current state $z_{t}$ based on the posterior estimation $\hat{z}_{t - 1}^{+}$ of the previous state and the dynamic model. Specifically, the prior estimation of latent state $z_{t}$ and estimation of observation $\pmb{x}_t$ are computed as + +$$ +\hat {z} _ {t} ^ {-} = f \left(\hat {z} _ {t - 1} ^ {+}\right), \tag {6} +$$ + +$$ +\hat {\boldsymbol {x}} _ {t} ^ {-} = h \left(g _ {\theta} \left(\hat {\boldsymbol {z}} _ {t} ^ {-}\right)\right). \tag {7} +$$ + +The system dynamics $f$ define how the latent state $Z$ evolves over time, and it incorporates any control inputs that might affect the current state. + +In the state update step, a posterior state estimation $\hat{z}_t^+$ is computed based on the prior estimation and new observations $\pmb{x}_t$ as + +$$ +\hat {z} _ {t} ^ {+} = \hat {z} _ {t} ^ {-} + \mathcal {K} _ {t} \Delta z _ {t}, \tag {8} +$$ + +where $\mathcal{K}_t$ is conceptually referred to as the Kalman gain and $\Delta z_{t}$ as the innovation, i.e., the residual between the prior estimation $\hat{\pmb{z}}_t^-$ and approximation of current state from $\pmb{x}_{t}$ , given by + +$$ +\Delta \boldsymbol {z} _ {t} = \hat {\boldsymbol {z}} _ {t} ^ {-} - \tilde {\boldsymbol {z}} _ {t}, \tag {9} +$$ + +where $\tilde{\pmb{z}}_t = e(\pmb{x}_t)$ . Note that this formula differs from the original Kalman filter which minimizes the innovation of $\Delta \pmb{x}_t$ (i.e., residual between $\pmb{x}_t$ and $\hat{\pmb{x}}_t^{-}$ ), since + +![](images/f066a8fe1c1daebbee5240bd34542656ce6bb37eef9f54e7f19a22480a7627ed.jpg) +Fig. 3: Overview of the proposed KEEP. It consists of four modules: encoder $\mathcal{E}_L$ , decoder $\mathcal{D}_Q$ , Kalman filter network, and CFA. We illustrate the information flow in one timestep. + +the assumption of available measurement function $h(\cdot)$ is no longer valid in our setting. An original measurement system models how observation $\boldsymbol{x}_t$ is derived from the latent state $\boldsymbol{z}_t$ , formally $\boldsymbol{x}_t = h(\boldsymbol{z}_t)$ . Inspired by KFNet [55], we directly estimate the state $\boldsymbol{z}_t$ given new observation $\boldsymbol{x}_t$ , i.e., mapping $\boldsymbol{x}_t$ to $\tilde{\boldsymbol{z}}_t$ with an estimator $e$ . + +The remaining problem is to compute the Kalman gain $\mathcal{K}_t$ . As discussed in KalmanNet [39], covariance estimation is intractable when dealing with high-dimensional signals. Additionally, the second-order statistical moments are only used for calculating Kalman gain. Inspired by this, we directly learn the gains from the data distribution and do not explicitly maintain an estimation of covariances. Additionally, we follow [51] to include $\tilde{\pmb{z}}_1$ of the first frame as anchor into KGN for Kalman gain estimation. Then, the final predicted $\hat{\pmb{y}}_t$ can be derived by + +$$ +\hat {\boldsymbol {y}} _ {t} = g _ {\theta} \left(\hat {\boldsymbol {z}} _ {t} ^ {+}\right). \tag {10} +$$ + +# 3.2 Parameterized Models + +As shown in Fig. 2 (b), we will parameterize or define the system dynamics $f$ , observation estimator $e$ , Kalman Gain Nets (KGN), and generative model $g_{\theta}$ in this section. The overall framework is shown in Fig. 3. + +Generative Model. The generative model $g_{\theta}$ is generally parameterized by a backbone of CodeFormer [56], which consists of a LQ encoder $\mathcal{E}_L$ , a HQ encoder $\mathcal{E}_H$ , a codebook lookup Transformer and quantization layer $T_{Q}$ , and a decoder $\mathcal{D}$ . For simplicity, $\mathcal{D}_Q$ denotes the decoder with the codebook lookup Transformer and quantization layer absorbed. Basically, the predicted $\hat{\pmb{y}}_t$ is given by $\mathcal{D}_Q(\hat{\pmb{z}}_t^+)$ , and the observed state $\tilde{\pmb{z}}_t$ is approximated by + +$$ +\tilde {z} _ {t} = e \left(\boldsymbol {x} _ {t}\right) = \mathcal {E} _ {L} \left(\boldsymbol {x} _ {t}\right). \tag {11} +$$ + +State Dynamic System. The system dynamics define how the system evolves over time, and it incorporates any control inputs that might affect the current state. The prediction for the state $\boldsymbol{z}_t$ at the current timestep is achieved via state extrapolation. In particular, given the posterior estimation of previous state $\hat{\boldsymbol{z}}_{t-1}^+$ , we define the dynamic model by + +$$ +\hat {\boldsymbol {z}} _ {t} ^ {-} = f \left(\hat {\boldsymbol {z}} _ {t - 1} ^ {+}\right) = \mathcal {E} _ {H} \left(\omega \left(\mathcal {D} _ {Q} \left(\hat {\boldsymbol {z}} _ {t - 1} ^ {+}\right), \Phi_ {t - 1 \rightarrow t}\right)\right). \tag {12} +$$ + +where $\varPhi_{t-1\rightarrow t}$ denotes the flow estimated from LQ frames $\pmb{x}_{t-1}$ to $\pmb{x}_t$ and $\omega$ the spatial warping modules. Specifically, we first decode the estimated code $\hat{\pmb{z}}_{t-1}^{+}$ of the previous frame to get the estimation of $\hat{\pmb{y}}_{t-1} = \mathcal{D}_Q(\hat{\pmb{z}}_{t-1}^+)$ and warp it to the current frame. Then, it was encoded back in the latent space to obtain the prediction of the current state $\hat{\pmb{z}}_t^{-}$ . + +Kalman Filter System. Given the approximated observed state $\tilde{z}_t$ and prior estimation $\hat{z}_t^-$ from system dynamics, the filter system aims to promote temporal information propagation and maintain stable latent code priors. In particular, the filter recursively fuses both the estimation to form a more accurate posterior estimate of the current state $\hat{z}_t^+$ , which is also known as state update. + +According to Eqn. 8, a more intuitive way to express the updated state is a linear interpolation $\hat{z}_t^+ = \mathcal{K}_t\hat{z}_t^- + (1 - \mathcal{K}_t)\tilde{z}_t$ for $\mathcal{K}_t$ normalized in the range of $[0,1]$ . The Kalman Gain $\mathcal{K}_t$ measures the estimated accuracy of the predicted states compared to the approximated observed states, to update the state and reduce the uncertainty. As illustrated in Fig. 3, the Kalman gain is approximated via Kalman Gain Network (KGN), consisting of two distinct parameterized modules, i.e., uncertainty network and gain network. The uncertainty associated with the current prediction is implicitly estimated by an uncertainty network constructed by spatial-temporal attention [47] and temporal attention layers (or any other recurrent networks). Then, a gain network calculates Kalman gain $\mathcal{K}_t$ for each code token. Please refer to the supplementary files for the detailed architectures. + +# 3.3 Local Temporal Consistency + +Inspired by [51], we adopt cross-frame attention (CFA) layers in the decoder to further promote local temporal consistency to regularize the information propagation. Specifically, given the latent features from the previous frame $v_{t-1}$ and current frame $v_t$ . They are projected onto the embedding space and output the features $v_i'$ by $v_i' = \mathrm{Attn}(Q, K, V) = \mathrm{softmax}\left(\frac{QK^T}{\sqrt{d}}\right) \cdot V$ , where + +$$ +Q = W _ {Q} \cdot v _ {t}, K = W _ {K} \cdot v _ {t - 1}, V = W _ {V} \cdot v _ {t - 1}. \tag {13} +$$ + +Intuitively, cross-frame attention modules can be regarded as searching and matching similar patches from the previous frame and fusing them correspondingly. This module facilitates temporal information propagation in the decoder. We adopt CFA modules on features of small scale 16 and 32 to avoid introducing blur to the decoded results. + +# 4 Experiments + +Dataset. VFHQ [48] contains over 15,000 high-quality video clips of diverse interviews and talk shows, where 15,381 clips are used for training and 50 clips are reserved for testing. Each sequence consists of 100 to 900 frames of resolution $512 \times 512$ . Following common practice [6,36,45,48], we adopt blind settings in all experiments. Specifically, we apply random blur, resize, and noise as image-based degradations. Moreover, video compression is adopted to control the video quality by changing streaming bitrate. For a comprehensive evaluation, we synthesize three splits of the VFHQ-Test dataset containing different levels of degradation, denoted as VFHQ-mild, VFHQ-medium, and VFHQ-heavy. They follow the same degradation model but differ in the degree of noise, blur, and compression. In addition to synthetic data, we also collect real corrupted video for testing. Please refer to the supplementary files for more details. + +Alignment. Pre-trained image models for face restoration are trained on the FFHQ dataset [20], where each image is automatically cropped and aligned. Hence, employing pre-trained models requires a similar alignment phase on VFHQ dataset [48]. However, the discrete step (i.e., cropping) is sensitive to the detected locations of landmarks, which can consequently result in unintentional temporal inconsistencies. Inspired by the work of Fox et al. [11] and Tzaban et al. [43], we employ a Gaussian lowpass filter over the landmarks. We find that this smoothing can significantly attenuate the inconsistencies induced by the alignment step. See supplementary files for more details. + +Implementations. For all stages of training, we initialize all networks with Kaiming Normal [15] and train them using Adam optimizer [22], and a batch size of 4 for all the experiments. The learning rate is set to $2 \times 10^{-4}$ for stages I and II, and $1 \times 10^{-4}$ for stage III. The models are trained with $800k$ , $400k$ , and $50k$ iterations for three stages, respectively. We implement our models with PyTorch [38] and train them using NVIDIA Tesla V100 GPUs. Hyper-parameters $\lambda_{1}$ , $\lambda_{VGG}$ , and $\lambda_{GAN}$ are set to $10^{-2}$ , 1, and $1 \times 10^{-2}$ . We use GMFlow [49] for optical flow estimation. + +Metrics. For quantitative evaluation, we evaluate the fidelity of restoration using PSNR, SSIM, and LPIPS [54]. In addition, we evaluate the identity preservation scores, termed as $IDS$ , by cosine similarity of the off-the-shelf identity detection network ArcFace [9]. Besides, we follow Tzaban et al. [43] to measure the pose consistency using Average Keypoint Distance (AKD), which is quantified by the average distance of detected landmarks between the generated and ground-truth video frames. In addition to the above single-frame quality evaluation, we also measure the fluctuation of identity/landmarks across frames. Hence, $\sigma_{IDS}$ measures the standard deviation of identity similarity over the entire video. We expect a considerable identity drift in the generated videos without local identity jitter, where $\sigma_{IDS}$ is supposed to be low. Similarly, we use $\sigma_{AKD}$ to measure the standard deviation of keypoint distances over the video, which quantifies the temporal consistency of the pose. + +Table 1: Quantitative comparison on the VFHQ-mild. Red and Blue indicate the best and the second best results. Full results on other test partitions (medium and heavy) are presented in the supplementary material. + +
Method|PSNR↑ SSIM↑ LPIPS↓| IDS↑ AKD↓ |σIDS(×10-2) ↓ σAKD ↓
GPEN [52]25.51930.75170.29880.714211.46914.7416
GFPGAN [45]26.29330.77950.24820.743710.54674.5700
RestoreFormer [46]25.57200.73440.31950.753010.53544.7159
CodeFormer [56]24.65970.74540.27420.627211.49836.3726
EDVR [44]26.60510.78580.24840.719511.62204.8048
BasicVSR [3]26.04580.77650.24960.697311.36795.0343
BasicVSR++ [5]27.19960.80570.19580.764111.31365.2543
KEEP (Ours)27.99940.82670.16190.79608.81823.6866
+ +# 4.1 Comparison with State-of-the-Art Methods + +Baselines. We compare our method to two categories of approaches. i) Image-based Face SR models (CodeFormer [56], GPEN [52], GFPGAN [45], Restore-Former [46]) are used to generate face videos frame-by-frame. ii) We retrain the general VSR models (EDVR [44], BasicVSR [3], BasicVSR++ [5]) on VFHQ dataset [48]. The degradation settings remain unchanged as our experiments while other training settings follow their original papers. + +Quantitative Evaluation. The quantitative results are listed in Table 1. We observe that our method achieves better results than existing methods across all the metrics. The results indicate that KEEP can faithfully recover facial details while preserving the identity. Our method also maintains temporal coherence across frames, as quantified by $\sigma_{IDS}$ and $\sigma_{AKD}$ , which represent the fluctuation of restored face identities and facial shapes. Though exhibiting structural distortions and artifacts (See Fig. 4), general VSR models (EDVR, BasicVSR, BasicVSR++) typically achieve higher performance on fidelity metrics (PNSR, SSIM, LPIPS) than single-image FSR models. This suggests that image-based models produce high-quality but relatively low-fidelity results. Inconsistency could be introduced when latent code estimation is noisy and inaccurate. + +Qualitative Evaluation. In Fig. 4, we can observe that the compared methods fail to reconstruct consistent appearances with perceptually pleasant details. For example, GFPGAN tends to hallucinate facial details (e.g., ears in the second frame and incomplete glass in the last frame of the left example). CodeFormer produces unnatural facial shapes (e.g., eyes), and BasicVSR leaves severe artifacts on the face images. In the last frame of the right example, both GFPGAN and CodeFormer generate unpleasant eyes (see yellow arrows). In contrast, our KEEP exploits temporal information and restores finer and coherent facial details. We refer readers to supplementary files for more video results. + +![](images/fc23af2204512c17dda0e651fb9a307526d9781ac2362fa99183d8339e29b254.jpg) +Fig. 4: Qualitative comparison on the VFHQ. Our KEEP produces high-fidelity face videos with faithful and consistent details. See arrows for details. + +Table 2: Ablation study of variant networks. + +
ModelsLPIPS↓IDS↑AKD↓
w/o CFA0.16210.79708.9029
w/o KFN0.17210.77739.1952
Full model0.16190.79608.8182
+ +Table 3: Ablation study on optical flow estimator. + +
EstimatorLPIPS↓IDS↑AKD↓
PWC-Net [41]0.16230.79578.7839
GMFlow [49] (Ours)0.16190.79608.8182
+ +# 4.2 More Analysis + +Effectiveness of KFN. We first investigate the effectiveness of the Kalman Filter Network (KFN). As shown in Table 2, removing KFN results in worse performance on LPIPS, IDS, and AKD. The results suggest the design of KFN is the key to promoting temporal consistency and identity preservation. + +Effectiveness of CFA. Table 2 also shows that Cross-Frame Attention (CFA) can further improve the performance. Though not clearly reflected in the number, we further qualitatively show the effectiveness of KFN and CFA in the supplementary video. From the demo video, we can observe that 1) adopting KFN can ensure consistency in global style and maintain the global appearance of the recovered faces. 2) adding CFA can further coherently render local texture details (e.g., hair), and suppress flicker. + +![](images/3c1afdfa52b4f2ab68793f92afec9aeba7e4a89e495e11a6bba058a2d9678cac.jpg) + +![](images/d9f67801607784ec21ce04e0148a2799d7ff888633b7870cfe3fed703681eaaa.jpg) +Fig. 5: Comparison of temporal flicker. We select each frame's column (red lines) and show the changes across time. Image-based models (GFPGAN, CodeFormer, and RestoreFormer) have obvious discontinuity around the eyes and wrinkles, and general VSR methods leave artifacts behind. In contrast, by maintaining stable facial priors and aggregating temporal information, our method remarkably suppresses temporal jitters and promotes coherent local details. + +![](images/a769f221c8655c2c4d31b4aba16fa4d25a3273f51948f8d8b6c594a201093b86.jpg) + +![](images/705b2c2e9d5f7f413a65527283e01a4d374813157d95352c859ae2a2cc1e0458.jpg) + +![](images/370b3538f8035b080f502b022a025b798d7b5e40657439ba67484d75a7c5ca30.jpg) + +![](images/5ec4c63cd7e3b9ffa055bdb6d17abef62435df9641965d9b4b6046552a732d3f.jpg) + +![](images/ff53ce63c05153754f218a588abed31c9199f8170fbaac93d3f96ad3138505e2.jpg) + +![](images/919718bfb026b6fc12898b988e9ae1a595e4cce48ca701e7cfca532704aca6a8.jpg) + +![](images/8d8c7915cbbe0a92ac8880633aa446f8d5d2ea164b2915b50770e2b3b8a6dd16.jpg) + +![](images/707ba80d225e97027b97952a3a63fe31a413399557a6bdd53917aeed221c6a5d.jpg) + +![](images/636d1b31318759a86d6ff643d192b591408efdfbfe8b843d27b958397d23d519.jpg) +(a) GT +Fig.6: Illustration of predicted state and observed state through Decoder $\mathcal{D}_Q$ . The teeth in red boxes illustrate a slight difference between $\tilde{\boldsymbol{z}}_t$ and $\hat{\boldsymbol{z}}_t^-$ when decoded in pixel space, suggesting the potential to supplement and fuse information to obtain a stable latent code. + +![](images/c611cdf9682c7466939eb0b3fe82e0b3a5c5446344724cdfe586f9fa853bf429.jpg) +(b) $\mathcal{D}_Q(\tilde{z}_t)$ + +![](images/fffc8b6a527dde660c741a3f521cbdaf92ab36a1c4b342089bf0ee8f9711d4fd.jpg) +(c) Error map of (b) + +![](images/1a710af11c863fa37ef2a4ae3fc4a48c781cfadeb9563ce21e8a519af6050322.jpg) +(d) $\mathcal{D}_Q(\hat{z}_t^-)$ + +![](images/d1906e3de3369fdd83d7d162f71d3c89bf3c44a5bc2ebc81fcc493bb2472804e.jpg) +(e) Error map of (d) + +Effectiveness of Various Flow Estimator. We compare models with different flow estimators in Table 3. The results suggest that the accuracy of estimated flows does not significantly affect the final performance. We conjecture this can be attributed to two factors: 1) Minor misalignment in pixel space can be reasonably diminished as the latent code is highly downsampled by a factor of $32 \times$ , at which level the latent representations are less sensitive to small spatial discrepancies present in the pixel space. 2) Other modules can compensate for the inaccuracy caused by flow estimators in a joint training fashion. + +Analysis of Flickering. We extract a short vertical segment of pixels from each frame and stack them horizontally to visualize the jittering issues within the video. In particular, existing methods demonstrate clear jitter and flicker across time, while our method shows better temporal consistency. Fig. 5 demonstrates that other image-based models bring obvious jitters around eyes, and general VSR methods leave behind artifacts, while our method could remarkably suppress temporal jitters and promote coherent local details. + +![](images/17aa6da14616d029dc97340e685d3e6374af2657b6192604c5a6ed073be8682c.jpg) +Fig. 7: Identity similarity across frames. Our method preserves the identity of input images and exhibits less fluctuation over time. CodeFormer results of two frames (green and red dashed line) exemplify an abrupt change of identity in the right figure, while our method maintains a stable identity both quantitatively and qualitatively. + +![](images/9d2cef2b1af797ad5c641485037945a0673da9178bfb937c7557ee2f0d40b110.jpg) + +![](images/898a5a68770ff2ec7a5586367486dd2c2f5175f4c56ae7106b96fb7b1071d2c9.jpg) +Fig. 8: Qualitative comparison on different levels of degradation. Our KEEP maintains high-fidelity in various degradations. + +Analysis of Latent Space. Since the true state $\boldsymbol{z}_t$ is unavailable, we indirectly analyze the predicted state $\tilde{\boldsymbol{z}}_t$ and observed state $\hat{\boldsymbol{z}}_t^-$ by decoding them to the pixel space through Decoder $\mathcal{D}_Q$ . As shown in Fig. 6, the areas around teeth exhibit a slight difference, while most remaining parts in error maps show similar decoded results. This indicates that the predicted and observed states could supplement each other to obtain a more accurate estimation of $\boldsymbol{z}_t$ , which is where the power of Kalman filter lies. + +Identity Preservation. We show the identity similarity across frames of one representative video clip in Fig. 7 As illustrated in the left figure, our method achieves better identity preservation and less identity jitter within the video, compared to the single-image methods. We also exemplify that the identity of CodeFormer results can change abruptly within several frames. The identity score of CodeFormer increases from 0.4552 to 0.6392, and turns down in later frames (See red curve), demonstrating great fluctuation across time. In contrast, our method maintains a stable identity both quantitatively and qualitatively. + +Various Degradations. Fig. 9a shows that KEEP is consistently better than the compared methods across different difficulty levels. Fig. 8 demonstrates re + +![](images/17c6a6f32090f5ba3e07d3bc6bde30371f59407bebd810a12c91229a3f93bfd0.jpg) + +![](images/458dc2f108eacd6d5c58937491bf6fb7d9cd762d2c2e8eb54214ca17d6797b14.jpg) +(a) PNSR and IDS on different partitions. + +![](images/1f233c901cc9f5fd087b4ef429260c4c91ab67f0c8b6417530003c1809221778.jpg) + +![](images/ab73be52ebc623cbb65365755e714283da6f4571c9967723c2590042cbcbb88a.jpg) + +![](images/1ebcfac9ccb6174376d8145b59177eff45dafab47885fee29070080894b41b2c.jpg) +(b) Comparisons on non-frontal faces. +Fig. 9: (a) Our methods achieve consistently better performance on various levels of degradation. (b) While CodeFormer fails to restore eyes to these challenging cases, our method can still produce plausible facial elements. + +sults on different levels of degradation. We draw the following observations. 1) Even GFPGAN and CodeFormer can restore plausible results on frames with mild degradation, the performance significantly deteriorates (e.g., eyes) upon heavier degradation. 2) Our method achieves better results by considering complementary information between adjacent frames and maintaining stable face priors. In addition, our method is appealing in handling heavy degradation. + +Non-Frontal-View Faces. In Fig. 9b, our model shows enhanced performance on non-frontal faces by providing more stable face priors estimations. While the single-image model CodeFormer cannot recover the eyes, our KEEP is still able to show robustness to these challenging cases. + +# 5 Conclusion + +We present a novel framework, KEEP, aiming at resolving the challenges associated with facial detail and temporal consistency in video face restoration. The proposed method demonstrates a unique capability to maintain a stable face prior over time, which is achieved by Kalman filtering principles, where our approach recurrently incorporates information from previously restored frames to guide and regulate the restoration process of the current frame. Extensive experiments demonstrate the efficacy of KEEP in consistently capturing facial details across video frames and keeping the temporal stability of face videos. + +# Acknowledgment + +This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). + +# References + +1. Caballero, J., Ledig, C., Aitken, A., Acosta, A., Totz, J., Wang, Z., Shi, W.: Real-time video super-resolution with spatio-temporal networks and motion compensation. In: CVPR (2017) +2. Chan, K.C., Wang, X., Xu, X., Gu, J., Loy, C.C.: GLEAN: Generative latent bank for large-factor image super-resolution. In: CVPR (2021) +3. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: BasicVSR: The search for essential components in video super-resolution and beyond. In: CVPR (2021) +4. Chan, K.C., Xu, X., Wang, X., Gu, J., Loy, C.C.: GLEAN: Generative latent bank for large-factor image super-resolution and beyond. TPAMI (2022) +5. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: CVPR (2022) +6. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Investigating tradeoffs in real-world video super-resolution. In: CVPR (2022) +7. Chen, C., Li, X., Yang, L., Lin, X., Zhang, L., Wong, K.Y.K.: Progressive semantic-aware style transformation for blind face restoration. In: CVPR (2021) +8. Chen, Y., Tai, Y., Liu, X., Shen, C., Yang, J.: FSRNet: End-to-end learning face super-resolution with facial priors. In: CVPR (2018) +9. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: ArcFace: Additive angular margin loss for deep face recognition. In: CVPR (2019) +10. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: CVPR (2021) +1. Fox, G., Tewari, A., Elgharib, M., Theobalt, C.: StyleVideoGAN: A temporal generative model using a pretrained stylegan. In: BMVC (2021) +2. Gu, J., Shen, Y., Zhou, B.: Image processing using multi-code gan prior. In: CVPR (2020) +3. Gu, Y., Wang, X., Xie, L., Dong, C., Li, G., Shan, Y., Cheng, M.M.: VQFR: Blind face restoration with vector-quantized dictionary and parallel decoder. In: ECCV (2022) +4. Haris, M., Shakhnarovich, G., Ukita, N.: Recurrent back-projection network for video super-resolution. In: CVPR (2019) +5. He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In: ICCV (2015) +6. Huang, Y., Wang, W., Wang, L.: Bidirectional recurrent convolutional networks for multi-frame super-resolution. In: NeurIPS (2015) +7. Isobe, T., Jia, X., Gu, S., Li, S., Wang, S., Tian, Q.: Video super-resolution with recurrent structure-detail network. In: ECCV (2020) +8. Isobe, T., Zhu, F., Wang, S.: Revisiting temporal modeling for video superresolution. In: BMVC (2020) +9. Kalman, R.E.: A new approach to linear filtering and prediction problems (1960) +20. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019) + +21. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: CVPR (2020) +22. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015) +23. Lei, C., Ren, X., Zhang, Z., Chen, Q.: Blind video deflickering by neural filtering with a flawed atlas. In: CVPR (2023) +24. Lei, C., Xing, Y., Chen, Q.: Blind video temporal consistency via deep video prior. In: NeurIPS (2020) +25. Li, S., He, F., Du, B., Zhang, L., Xu, Y., Tao, D.: Fast spatio-temporal residual network for video super-resolution. In: CVPR (2019) +26. Li, X., Chen, C., Zhou, S., Lin, X., Zuo, W., Zhang, L.: Blind face restoration via deep multi-scale component dictionaries. In: ECCV (2020) +27. Li, X., Li, W., Ren, D., Zhang, H., Wang, M., Zuo, W.: Enhanced blind face restoration with multi-exemplar images and adaptive spatial feature fusion. In: CVPR (2020) +28. Li, X., Liu, M., Ye, Y., Zuo, W., Lin, L., Yang, R.: Learning warped guidance for blind face restoration. In: ECCV (2018) +29. Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: Vrt: A video restoration transformer. IEEE Transactions on Image Processing (2024) +30. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided deformable attention. NeurIPS (2022) +31. Liu, C., Yang, H., Fu, J., Qian, X.: Learning trajectory-aware transformer for video super-resolution. In: CVPR (2022) +32. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: CVPR (2022) +33. Menon, S., Damian, A., Hu, S., Ravi, N., Rudin, C.: Pulse: Self-supervised photoupsampling via latent space exploration of generative models. In: CVPR (2020) +34. Nagrani, A., Chung, J.S., Zisserman, A.: Voxceleb: A large-scale speaker identification dataset. In: Interspeech (2017) +35. Ouyang, H., Wang, Q., Xiao, Y., Bai, Q., Zhang, J., Zheng, K., Zhou, X., Chen, Q., Shen, Y.: Codef: Content deformation fields for temporally consistent video processing. In: CVPR (2024) +36. Pan, J., Bai, H., Dong, J., Zhang, J., Tang, J.: Deep blind video super-resolution. In: ICCV (2021) +37. Pan, X., Zhan, X., Dai, B., Lin, D., Loy, C.C., Luo, P.: Exploiting deep generative prior for versatile image restoration and manipulation. TPAMI 44(11), 7474-7489 (2021) +38. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017) +39. Revach, G., Shlezinger, N., Ni, X., Escoriza, A.L., Van Sloun, R.J., Eldar, Y.C.: Kalmannet: Neural network aided kalman filtering for partially known dynamics. IEEE Transactions on Signal Processing (2022) +40. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) +41. Sun, D., Yang, X., Liu, M.Y., Kautz, J.: Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In: CVPR (2018) +42. Tian, Y., Zhang, Y., Fu, Y., Xu, C.: TDAN: Temporally-deformable alignment network for video super-resolution. In: CVPR (2020) + +43. Tzaban, R., Mokady, R., Gal, R., Bermano, A., Cohen-Or, D.: Stitch it in time: Gan-based facial editing of real videos. In: SIGGRAPH Asia (2022) +44. Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: EDVR: Video restoration with enhanced deformable convolutional networks. In: CVPRW (2019) +45. Wang, X., Li, Y., Zhang, H., Shan, Y.: Towards real-world blind face restoration with generative facial prior. In: CVPR (2021) +46. Wang, Z., Zhang, J., Chen, R., Wang, W., Luo, P.: Restoreformer: High-quality blind face restoration from undegraded key-value pairs. In: CVPR (2022) +47. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: ICCV (2023) +48. Xie, L., Wang, X., Zhang, H., Dong, C., Shan, Y.: VFHQ: A high-quality dataset and benchmark for video face super-resolution. In: CVPR (2022) +49. Xu, H., Zhang, J., Cai, J., Rezatofighi, H., Tao, D.: Gmflow: Learning optical flow via global matching. In: CVPR (2022) +50. Xu, Y., AlBahar, B., Huang, J.B.: Temporally consistent semantic video editing. In: ECCV (2022) +51. Yang, S., Zhou, Y., Liu, Z.,, Loy, C.C.: Rerender a video: Zero-shot text-guided video-to-video translation. In: SIGGRAPH Asia (2023) +52. Yang, T., Ren, P., Xie, X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: CVPR (2021) +53. Yu, X., Fernando, B., Ghanem, B., Porikli, F., Hartley, R.: Face super-resolution guided by facial component heatmaps. In: ECCV (2018) +54. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) +55. Zhou, L., Luo, Z., Shen, T., Zhang, J., Zhen, M., Yao, Y., Fang, T., Quan, L.: KFNet: Learning temporal camera relocalization using Kalman filtering. In: CVPR (2020) +56. Zhou, S., Chan, K.C., Li, C., Loy, C.C.: Towards robust blind face restoration with codebook lookup transformer. In: NeurIPS (2022) \ No newline at end of file diff --git a/2024/Kalman-Inspired Feature Propagation for Video Face Super-Resolution/images.zip b/2024/Kalman-Inspired Feature Propagation for Video Face Super-Resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..586222112f60309748e66b5bd8de4baba30afd3a --- /dev/null +++ b/2024/Kalman-Inspired Feature Propagation for Video Face Super-Resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:145363a52f7b764873a83fe0c6481fb7a87aadd47855933bd1e797afe9168e80 +size 600130 diff --git a/2024/Kalman-Inspired Feature Propagation for Video Face Super-Resolution/layout.json b/2024/Kalman-Inspired Feature Propagation for Video Face Super-Resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4415190f5db8b6b6ea6531dea6e67f30614dfa0e --- /dev/null +++ b/2024/Kalman-Inspired Feature Propagation for Video Face Super-Resolution/layout.json @@ -0,0 +1,10415 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 162, + 111, + 451, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 111, + 451, + 147 + ], + "spans": [ + { + "bbox": [ + 162, + 111, + 451, + 147 + ], + "type": "text", + "content": "Kalman-Inspired Feature Propagation for Video Face Super-Resolution" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 186, + 168, + 429, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 168, + 429, + 180 + ], + "spans": [ + { + "bbox": [ + 186, + 168, + 429, + 180 + ], + "type": "text", + "content": "Ruicheng Feng, Chongyi Li, and Chen Change Loy" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 201, + 190, + 413, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 190, + 413, + 223 + ], + "spans": [ + { + "bbox": [ + 201, + 190, + 413, + 223 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University, Singapore {ruicheng002, ccloy}@ntu.edu.sg lichongyi25@gmail.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 251, + 453, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 251, + 453, + 415 + ], + "spans": [ + { + "bbox": [ + 162, + 251, + 453, + 415 + ], + "type": "text", + "content": "Abstract. Despite the promising progress of face image super-resolution, video face super-resolution remains relatively under-explored. Existing approaches either adapt general video super-resolution networks to face datasets or apply established face image super-resolution models independently on individual video frames. These paradigms encounter challenges either in reconstructing facial details or maintaining temporal consistency. To address these issues, we introduce a novel framework called Kalman-inspired Feature Propagation (KEEP), designed to maintain a stable face prior over time. The Kalman filtering principles offer our method a recurrent ability to use the information from previously restored frames to guide and regulate the restoration process of the current frame. Extensive experiments demonstrate the effectiveness of our method in capturing facial details consistently across video frames. Code and video demo are available at https://jnjaby.github.io/projects/KEEP/." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 437, + 228, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 437, + 228, + 449 + ], + "spans": [ + { + "bbox": [ + 134, + 437, + 228, + 449 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 461, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 461, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 134, + 461, + 481, + 605 + ], + "type": "text", + "content": "The field of Face Super-Resolution (FSR), which focuses on reconstructing high-resolution (HR) face images from highly degraded versions, has witnessed remarkable progress. In particular, numerous studies have successfully leveraged various types of prior information, such as geometric facial priors [7,8,53], reference priors [26-28], generative priors [2,4,45,52], and codebook priors [13,46,56]. These approaches have significantly advanced the realism and quality of generated face images. However, the majority of these studies are confined to still images, with the extension to Video Face Super-Resolution (VFSR) remaining relatively under-explored. Despite the substantial potential benefits of video face restoration in various practical domains, such as the restoration of old films, VFSR has yet to receive the same level of attention and development as its image-based counterpart." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 606, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 606, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 606, + 480, + 665 + ], + "type": "text", + "content": "Two main strategies emerge for implementing VFSR. The first approach involves adapting general Video Super-Resolution (VSR) networks, such as EDVR [44], BasicVSR [3], BasicVSR++ [5], and RVRT [30], to large-scale face video datasets [34, 48]. These methods exploit temporal information and propagate features across video frames. However, they are not specifically tailored for" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 478, + 304 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 478, + 304 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 478, + 304 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 478, + 304 + ], + "type": "image", + "image_path": "d0990cb8d6bd312c9850737ee484d46a046f3d917329cb81e114178078088ed4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 312, + 482, + 369 + ], + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 369 + ], + "type": "text", + "content": "Fig. 1: Comparing main VFSR strategies. We show seven frames with an interval of 6. Generic VSR model BasicVSR [3] fails to reconstruct facial components faithfully. Single-image FSR model CodeFormer [56] hallucinates unnatural and inconsistent face details. Our method, in contrast, enables consistent restoration of low-quality face video while preserving temporal coherence across frames." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 399, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 518 + ], + "type": "text", + "content": "face restoration and often fall short in reconstructing detailed facial features, particularly in severely degraded scenarios, as depicted in Fig. 1(second row). The second method involves applying existing face image SR models to process each video frame independently. This frame-by-frame approach, while straightforward, introduces temporal inconsistencies in the video, as demonstrated in Fig. 1(third row). This problem arises because FSR is inherently ill-posed and the existing priors may not suffice to maintain appearance consistency throughout the video sequence. Specifically, a single degraded image could correspond to multiple high-resolution interpretations, leading to discrepancies and inconsistent structures across independently processed video frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": "In this paper, we wish to devise an effective framework for maintaining a stable face prior over time for VFSR. We use CodeFormer [56], a representative model that exploits codebook priors for FSR, to demonstrate how face priors can be consistently preserved across different time frames. A pivotal aspect of our approach is the idea that frames previously restored can act as references, guiding and regulating the restoration process of the current frame. This strategy helps minimize the divergence between consecutive frames. Moreover, this reliance on previously restored frames naturally suggests a recurrent framework, enabling the effective use of information from past restorations. This intuition aligns closely with Kalman filtering principles, or linear quadratic estimation, which involves using a sequence of time-based measurements, despite their statistical noise and inaccuracies, to produce more accurate estimates of unknown" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "R. Feng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": "variables than would be possible with a single measurement. Similarly, in VFSR, faces observed over time are often noisy and inaccurate, making them suitable for refinement using Kalman filtering techniques." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "type": "text", + "content": "Driven by these insights, we formulate a novel method, Kalman-inspired fEaturE Propagation (KEEP), which recurrently updates the current latent state in CodeFormer by incorporating information from preceding frames. This method of temporal propagation within the latent space ensures the stability of the face prior over time, thereby capturing facial details that consistently match in appearance. The effectiveness of KEEP is shown in Fig. 1, where it is evident that our method delivers high-quality restoration with superior consistency, compared to both generic video restoration methods fine-tuned for faces and approaches that restore frames independently. Please refer to the supplementary video to appreciate the superiority of our approach in terms of temporal consistency. A key advantage of KEEP is its robustness in handling severe video-based degradation, outperforming single-image models. In addition, our model demonstrates enhanced performance on non-frontal faces by providing more stable estimations of face priors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 320, + 482, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 428 + ], + "type": "text", + "content": "In summary, the main contribution of this work is a novel framework for maintaining a stable and meaningful face prior for VFSR. While we demonstrate its application using the CodeFormer method as a case study, the underlying principles of our framework, inspired by the Kalman filtering approach, are applicable to other approaches. Extensive experimental results on the VFHQ dataset [48] and real-world data demonstrate the effectiveness of our approach in improving both the fidelity and coherence of VFSR outputs. Compared to other state-of-the-art methods, our model achieves superior performance with a large margin of 0.8 dB in PSNR, while also significantly maintaining temporal coherence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 447, + 237, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 237, + 459 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 237, + 459 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": "Blind Face Restoration. Blind face restoration aims at recovering severely degraded face images in the wild. Unlike natural images, faces are highly structured. This property allows researchers to incorporate prior information into the restoration models, which have demonstrated remarkable progress in capabilities to restore high-quality faces. Most existing FSR methods can be categorized into four classes: geometric priors, reference priors, generative priors, and codebook priors. Geometric priors usually include facial elements, such as face landmarks [8], parsing maps [7], and facial component heatmaps [53]. Another major line is reference-based methods that require high-quality exemplar images. GFRNet [28] and ASFFNet [27] leverage a warped high-quality image to extract rich details to improve facial detail restoration. DFDNet [26] constructs deep dictionaries with facial components from large-scale images to recover fine details. Generative priors from pre-trained GAN, e.g., StyleGAN [20] and StyleGAN2 [21], are employed through iterative latent optimization of GAN inversion [12, 33, 37]. Still, they produce face images with low fidelity and are computationally expensive. To address this, GLEAN [2], GPEN [52], and GFP-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "KEEP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "GAN [45] integrate generative priors into encoder-decoder architectures, which estimate latent priors in one-forward pass. These methods achieve great tradeoff between quality and fidelity but usually fail when the corruption is severe. Codebook priors [13,46,56] can be regarded as a special case of generative priors. In contrast to continuous generative priors, they squeeze the latent space into a small finite codebook space and improve the robustness to severe degradation. However, most existing FSR methods are image-based and thus they cannot guarantee temporally consistent details for VFSR." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 217, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 540 + ], + "type": "text", + "content": "Video Super-Resolution. Most existing video restoration techniques can be categorized into two paradigms based on their parallelizability: parallel and recurrent methods. Parallel models estimate all frames simultaneously, and the restoration of each frame does not rely on the update of other frames. These methods typically involve feature extraction, feature alignment, feature fusion, and reconstruction. Representative works, FSTRN [25] and VESCPN [1], introduce fast spatio-temporal networks using 3D convolutions to enhance alignment, combining motion compensation and super-resolution. EDVR [44] and TDAN [42] leverage deformable convolutions for aligning adjacent frames. Besides, Transformer-based methods [29,32] are proposed to reconstruct all frames simultaneously by jointly extracting, aligning, and fusing features. In addition, RVRT [30] and TTVSR [31] integrate optical flow into Transformer and enable long-range models in videos. Empowered by the great expressive capability of Transformer, this line of work exhibits remarkable performance improvements over previous methods. However, they suffer from large model sizes and high memory consumption. Recurrent methods [3,16,18] do not aggregate information solely from adjacent frames. Instead, they maintain hidden states to convey relevant information from previous frames and propagate latent features sequentially, accumulating information for later restoration. For example, RBPN [14] treated each frame as a separate source, combined iteratively in a refinement framework. RSDN [17] divided the input into structure and detail components, proposing a two-stream structure-detail block to learn textures. BasicVSR [3] and BasicVSR++ [5] fused bidirectional hidden states from both past and future frames, bring significant improvements. They aim to fully utilize information from the entire sequence, synchronously updating the hidden state through the weights of the reconstruction network. Due to the recurrent nature of feature propagation, recurrent methods experience information loss." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 545, + 483, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 483, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 483, + 665 + ], + "type": "text", + "content": "Video Consistency. In addition to VSR, previous works also attempt to inflate image models into video models [43, 47, 50, 51] or improve temporal consistency with the implicit representation of the given videos [23, 24, 35]. For example, Stitch-it-in-Time [43] discovers locally consistent pivots in the latent space to provide spatially consistent transitions. Tune-A-Video [47] adopts cross-frame attention and fine-tune it on a single video. All-in-One deflicker [23] learns a neural atlas for each video to solve long-term inconsistency. These works either require post-processing or optimization on a video basis. In contrast, we aim at repurposing image FSR models and enforcing temporal consistency in the restored face videos." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "R. Feng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 233, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 233, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 233, + 129 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 258 + ], + "type": "text", + "content": "In a corrupted face video, local textures and facial details are irrevocably lost. Therefore, the latent codes, usually estimated by an encoder, are inaccurate to match the real underlying priors of ground-truth video. Different from image-based autoencoder, models in the video domain enable us to exploit evidence accumulated from preceding frames for better restoration. In this paper, we repurpose the image-based CodeFormer for video FSR and propose KEEP to estimate stable face priors in latent space over time given the noisy and inaccurate estimations, which consequently enable temporally coherent restoration. An appealing idea to realize KEEP is through reformulating the method in a Kalman Filter framework." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 274, + 222, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 222, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 222, + 284 + ], + "type": "text", + "content": "3.1 Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 290, + 482, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 350 + ], + "type": "text", + "content": "State Space Model. We consider observations of low-quality (LQ) video sequence " + }, + { + "bbox": [ + 130, + 290, + 482, + 350 + ], + "type": "inline_equation", + "content": "X = \\{\\pmb{x}_{t}\\}_{t=1}^{T}" + }, + { + "bbox": [ + 130, + 290, + 482, + 350 + ], + "type": "text", + "content": " with length of " + }, + { + "bbox": [ + 130, + 290, + 482, + 350 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 290, + 482, + 350 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 290, + 482, + 350 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 290, + 482, + 350 + ], + "type": "text", + "content": ", and underlying high-quality (HQ) sequences " + }, + { + "bbox": [ + 130, + 290, + 482, + 350 + ], + "type": "inline_equation", + "content": "Y = \\{\\pmb{y}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 290, + 482, + 350 + ], + "type": "text", + "content": ". Kalman filter [19] assumes linear dynamic systems that are characterized by a state space model driven by Gaussian noise" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 267, + 352, + 481, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 352, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 267, + 352, + 481, + 364 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {t} = \\boldsymbol {F} _ {t} \\boldsymbol {y} _ {t - 1} + \\boldsymbol {q} _ {t}, \\tag {1}", + "image_path": "c68f9e7999ee18965f4c5965f595654ff5056b13accadf56664e398b7f97fa73.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 367, + 481, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 481, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 367, + 481, + 392 + ], + "type": "inline_equation", + "content": "\\pmb{F}_t" + }, + { + "bbox": [ + 130, + 367, + 481, + 392 + ], + "type": "text", + "content": " is the transition matrix and " + }, + { + "bbox": [ + 130, + 367, + 481, + 392 + ], + "type": "inline_equation", + "content": "\\pmb{q}_t" + }, + { + "bbox": [ + 130, + 367, + 481, + 392 + ], + "type": "text", + "content": " denotes process noise drawn from Gaussian noise. The observation " + }, + { + "bbox": [ + 130, + 367, + 481, + 392 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 130, + 367, + 481, + 392 + ], + "type": "text", + "content": " is measured by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 273, + 399, + 480, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 399, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 273, + 399, + 480, + 411 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t} = \\boldsymbol {H} \\boldsymbol {y} _ {t} + \\boldsymbol {r} _ {t}, \\tag {2}", + "image_path": "a95607e538058d34ca830c20bd35379e92353bd722eb0ddc745d79252e92e294.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 418, + 481, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 481, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 418, + 481, + 453 + ], + "type": "inline_equation", + "content": "\\pmb{H}" + }, + { + "bbox": [ + 130, + 418, + 481, + 453 + ], + "type": "text", + "content": " is the measurement matrix and " + }, + { + "bbox": [ + 130, + 418, + 481, + 453 + ], + "type": "inline_equation", + "content": "\\pmb{r}_t" + }, + { + "bbox": [ + 130, + 418, + 481, + 453 + ], + "type": "text", + "content": " represents measurement noise. However, the linear assumption does not hold in some complex real-world scenarios. Hence, the non-linear Kalman filter can be reformulated as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 271, + 461, + 480, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 461, + 480, + 473 + ], + "spans": [ + { + "bbox": [ + 271, + 461, + 480, + 473 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {t} = d \\left(\\boldsymbol {y} _ {t - 1}, \\boldsymbol {q} _ {t}\\right) \\tag {3}", + "image_path": "40fd7e0146e83c61d9a4aa0f09a1c1747e93a1c3b2b66d1c85d48e990cb61c83.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 271, + 477, + 480, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 477, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 271, + 477, + 480, + 488 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t} = h \\left(\\boldsymbol {y} _ {t}\\right) + \\boldsymbol {r} _ {t}, \\tag {4}", + "image_path": "dc98e7c9769155ef454a5239013f6314088d2a53fb7d98d4b1e6abd093057257.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 495, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 495, + 482, + 567 + ], + "type": "inline_equation", + "content": "d(\\cdot)" + }, + { + "bbox": [ + 130, + 495, + 482, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 495, + 482, + 567 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 130, + 495, + 482, + 567 + ], + "type": "text", + "content": " are non-linear transition and measurement models. Specifically, " + }, + { + "bbox": [ + 130, + 495, + 482, + 567 + ], + "type": "inline_equation", + "content": "d(\\cdot)" + }, + { + "bbox": [ + 130, + 495, + 482, + 567 + ], + "type": "text", + "content": " can be represented by any explicit motion estimation (e.g., optical flow), which defines how the current frame transits to the next one. " + }, + { + "bbox": [ + 130, + 495, + 482, + 567 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 130, + 495, + 482, + 567 + ], + "type": "text", + "content": " commonly models the degradation in video restoration problems. As opposed to the classical assumptions in Kalman filter, the measurement function " + }, + { + "bbox": [ + 130, + 495, + 482, + 567 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 130, + 495, + 482, + 567 + ], + "type": "text", + "content": " is unknown in a blind setting. This is regarded as partially known dynamic models [39]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 567, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 604 + ], + "type": "text", + "content": "Inspired by VQGAN [10] and Stable Diffusion [40], we estimate underlying latent representations " + }, + { + "bbox": [ + 130, + 567, + 482, + 604 + ], + "type": "inline_equation", + "content": "Z = \\{z_{t}\\}_{t=1}^{T}" + }, + { + "bbox": [ + 130, + 567, + 482, + 604 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 130, + 567, + 482, + 604 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 567, + 482, + 604 + ], + "type": "text", + "content": " can correspond to " + }, + { + "bbox": [ + 130, + 567, + 482, + 604 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 130, + 567, + 482, + 604 + ], + "type": "text", + "content": " by a generative model " + }, + { + "bbox": [ + 130, + 567, + 482, + 604 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 130, + 567, + 482, + 604 + ], + "type": "text", + "content": ", given by" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 280, + 611, + 480, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 280, + 611, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 280, + 611, + 480, + 622 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {t} = g _ {\\theta} \\left(\\boldsymbol {z} _ {t}\\right). \\tag {5}", + "image_path": "ce4cdf2e13ae1b47c455e4cfc1ae7add37d2125188167230a627b1646c57354b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Instead of directly estimating individual pixels, modeling the low-dimensional latent code is computationally more efficient and focuses on more perceptually significant variations of the data. The graphical model is depicted in Fig. 2 (a)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "KEEP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 272, + 209 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 272, + 209 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 272, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 272, + 209 + ], + "type": "image", + "image_path": "b24db573ef9e18763e701385155b15c194b9262dbc57157da9c67a056394191a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 197, + 210, + 207, + 219 + ], + "lines": [ + { + "bbox": [ + 197, + 210, + 207, + 219 + ], + "spans": [ + { + "bbox": [ + 197, + 210, + 207, + 219 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 277, + 116, + 475, + 210 + ], + "blocks": [ + { + "bbox": [ + 277, + 116, + 475, + 210 + ], + "lines": [ + { + "bbox": [ + 277, + 116, + 475, + 210 + ], + "spans": [ + { + "bbox": [ + 277, + 116, + 475, + 210 + ], + "type": "image", + "image_path": "3243931ce86ba68f7e0872371ad01dfa3bde02ef51ec1347563adb3fd76e6268.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 211, + 374, + 219 + ], + "lines": [ + { + "bbox": [ + 364, + 211, + 374, + 219 + ], + "spans": [ + { + "bbox": [ + 364, + 211, + 374, + 219 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "text", + "content": "Fig.2: (a) Graphical model of state space. It defines the underlying dynamic system model, where " + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "text", + "content": " describes how the latent states " + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "text", + "content": " transit over time, " + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "text", + "content": " is a generative model, and " + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "text", + "content": " models the degradation from clean frame " + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "text", + "content": " to degraded frame " + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "text", + "content": ". (b) Block diagram of Kalman filter model. In each time step, a predictive state from previous frame " + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\hat{z}_{t-1}^{+}" + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "text", + "content": " (Blue dash box) and new observed state of current frame " + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "text", + "content": " (Red dash box) are fused by Kalman gain " + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_t" + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "text", + "content": " from Kalman Gain Network (KGN) to produce more accurate estimates. The combined state " + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\hat{z}_t^+" + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "text", + "content": " is then used to generate the estimated clean frame " + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\hat{y}_t" + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\tilde{z}_1" + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "text", + "content": " goes along with " + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{t-1}" + }, + { + "bbox": [ + 131, + 232, + 482, + 331 + ], + "type": "text", + "content": " as an anchor and it is omitted in the diagram for simplicity." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 354, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 354, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 131, + 354, + 482, + 402 + ], + "type": "text", + "content": "Kalman Filter Model. The principles of Kalman filter can be formulated by a two-step procedure, i.e., state prediction and state update. The overview diagram is illustrated in Fig. 2 (b). In this problem, the observation is a face image " + }, + { + "bbox": [ + 131, + 354, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 131, + 354, + 482, + 402 + ], + "type": "text", + "content": " and the state is " + }, + { + "bbox": [ + 131, + 354, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 131, + 354, + 482, + 402 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 403, + 483, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 403, + 483, + 450 + ], + "spans": [ + { + "bbox": [ + 131, + 403, + 483, + 450 + ], + "type": "text", + "content": "In the state prediction step, the model predicts the prior estimation " + }, + { + "bbox": [ + 131, + 403, + 483, + 450 + ], + "type": "inline_equation", + "content": "\\hat{z}_t^-" + }, + { + "bbox": [ + 131, + 403, + 483, + 450 + ], + "type": "text", + "content": " of the current state " + }, + { + "bbox": [ + 131, + 403, + 483, + 450 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 131, + 403, + 483, + 450 + ], + "type": "text", + "content": " based on the posterior estimation " + }, + { + "bbox": [ + 131, + 403, + 483, + 450 + ], + "type": "inline_equation", + "content": "\\hat{z}_{t - 1}^{+}" + }, + { + "bbox": [ + 131, + 403, + 483, + 450 + ], + "type": "text", + "content": " of the previous state and the dynamic model. Specifically, the prior estimation of latent state " + }, + { + "bbox": [ + 131, + 403, + 483, + 450 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 131, + 403, + 483, + 450 + ], + "type": "text", + "content": " and estimation of observation " + }, + { + "bbox": [ + 131, + 403, + 483, + 450 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 131, + 403, + 483, + 450 + ], + "type": "text", + "content": " are computed as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 268, + 458, + 480, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 458, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 268, + 458, + 480, + 472 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {t} ^ {-} = f \\left(\\hat {z} _ {t - 1} ^ {+}\\right), \\tag {6}", + "image_path": "2472841f7c8bd2ef04063a177d9a03cb1756716b60e832d7544a7331cb713055.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 269, + 475, + 480, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 475, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 269, + 475, + 480, + 489 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {x}} _ {t} ^ {-} = h \\left(g _ {\\theta} \\left(\\hat {\\boldsymbol {z}} _ {t} ^ {-}\\right)\\right). \\tag {7}", + "image_path": "166188b494dda65616c978faed1acc4939a5c0868ed8beb0295dbbdfc1ba1c7f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 497, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 497, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 131, + 497, + 481, + 521 + ], + "type": "text", + "content": "The system dynamics " + }, + { + "bbox": [ + 131, + 497, + 481, + 521 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 131, + 497, + 481, + 521 + ], + "type": "text", + "content": " define how the latent state " + }, + { + "bbox": [ + 131, + 497, + 481, + 521 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 131, + 497, + 481, + 521 + ], + "type": "text", + "content": " evolves over time, and it incorporates any control inputs that might affect the current state." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 521, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 481, + 545 + ], + "type": "text", + "content": "In the state update step, a posterior state estimation " + }, + { + "bbox": [ + 131, + 521, + 481, + 545 + ], + "type": "inline_equation", + "content": "\\hat{z}_t^+" + }, + { + "bbox": [ + 131, + 521, + 481, + 545 + ], + "type": "text", + "content": " is computed based on the prior estimation and new observations " + }, + { + "bbox": [ + 131, + 521, + 481, + 545 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 131, + 521, + 481, + 545 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 263, + 552, + 480, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 552, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 263, + 552, + 480, + 567 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {t} ^ {+} = \\hat {z} _ {t} ^ {-} + \\mathcal {K} _ {t} \\Delta z _ {t}, \\tag {8}", + "image_path": "19aeb85ce1841db2da3946fa2d28caf1890923a63045694b228f4b45a638acb1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 575, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 575, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 131, + 575, + 482, + 612 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 575, + 482, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_t" + }, + { + "bbox": [ + 131, + 575, + 482, + 612 + ], + "type": "text", + "content": " is conceptually referred to as the Kalman gain and " + }, + { + "bbox": [ + 131, + 575, + 482, + 612 + ], + "type": "inline_equation", + "content": "\\Delta z_{t}" + }, + { + "bbox": [ + 131, + 575, + 482, + 612 + ], + "type": "text", + "content": " as the innovation, i.e., the residual between the prior estimation " + }, + { + "bbox": [ + 131, + 575, + 482, + 612 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{z}}_t^-" + }, + { + "bbox": [ + 131, + 575, + 482, + 612 + ], + "type": "text", + "content": " and approximation of current state from " + }, + { + "bbox": [ + 131, + 575, + 482, + 612 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t}" + }, + { + "bbox": [ + 131, + 575, + 482, + 612 + ], + "type": "text", + "content": ", given by" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 271, + 620, + 480, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 620, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 271, + 620, + 480, + 633 + ], + "type": "interline_equation", + "content": "\\Delta \\boldsymbol {z} _ {t} = \\hat {\\boldsymbol {z}} _ {t} ^ {-} - \\tilde {\\boldsymbol {z}} _ {t}, \\tag {9}", + "image_path": "e3f457bf0448da81f7847d0ca089c9dca2454e879b9be5ecb83a547996725af4.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{z}}_t = e(\\pmb{x}_t)" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": ". Note that this formula differs from the original Kalman filter which minimizes the innovation of " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\Delta \\pmb{x}_t" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " (i.e., residual between " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_t^{-}" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "), since" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "R. Feng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 148, + 116, + 462, + 240 + ], + "blocks": [ + { + "bbox": [ + 148, + 116, + 462, + 240 + ], + "lines": [ + { + "bbox": [ + 148, + 116, + 462, + 240 + ], + "spans": [ + { + "bbox": [ + 148, + 116, + 462, + 240 + ], + "type": "image", + "image_path": "f066a8fe1c1daebbee5240bd34542656ce6bb37eef9f54e7f19a22480a7627ed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 248, + 480, + 282 + ], + "lines": [ + { + "bbox": [ + 131, + 248, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 131, + 248, + 480, + 282 + ], + "type": "text", + "content": "Fig. 3: Overview of the proposed KEEP. It consists of four modules: encoder " + }, + { + "bbox": [ + 131, + 248, + 480, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_L" + }, + { + "bbox": [ + 131, + 248, + 480, + 282 + ], + "type": "text", + "content": ", decoder " + }, + { + "bbox": [ + 131, + 248, + 480, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_Q" + }, + { + "bbox": [ + 131, + 248, + 480, + 282 + ], + "type": "text", + "content": ", Kalman filter network, and CFA. We illustrate the information flow in one timestep." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 307, + 479, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 479, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 479, + 365 + ], + "type": "text", + "content": "the assumption of available measurement function " + }, + { + "bbox": [ + 130, + 307, + 479, + 365 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 130, + 307, + 479, + 365 + ], + "type": "text", + "content": " is no longer valid in our setting. An original measurement system models how observation " + }, + { + "bbox": [ + 130, + 307, + 479, + 365 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t" + }, + { + "bbox": [ + 130, + 307, + 479, + 365 + ], + "type": "text", + "content": " is derived from the latent state " + }, + { + "bbox": [ + 130, + 307, + 479, + 365 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_t" + }, + { + "bbox": [ + 130, + 307, + 479, + 365 + ], + "type": "text", + "content": ", formally " + }, + { + "bbox": [ + 130, + 307, + 479, + 365 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t = h(\\boldsymbol{z}_t)" + }, + { + "bbox": [ + 130, + 307, + 479, + 365 + ], + "type": "text", + "content": ". Inspired by KFNet [55], we directly estimate the state " + }, + { + "bbox": [ + 130, + 307, + 479, + 365 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_t" + }, + { + "bbox": [ + 130, + 307, + 479, + 365 + ], + "type": "text", + "content": " given new observation " + }, + { + "bbox": [ + 130, + 307, + 479, + 365 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t" + }, + { + "bbox": [ + 130, + 307, + 479, + 365 + ], + "type": "text", + "content": ", i.e., mapping " + }, + { + "bbox": [ + 130, + 307, + 479, + 365 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t" + }, + { + "bbox": [ + 130, + 307, + 479, + 365 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 307, + 479, + 365 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{z}}_t" + }, + { + "bbox": [ + 130, + 307, + 479, + 365 + ], + "type": "text", + "content": " with an estimator " + }, + { + "bbox": [ + 130, + 307, + 479, + 365 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 130, + 307, + 479, + 365 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 367, + 480, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 480, + 463 + ], + "type": "text", + "content": "The remaining problem is to compute the Kalman gain " + }, + { + "bbox": [ + 130, + 367, + 480, + 463 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_t" + }, + { + "bbox": [ + 130, + 367, + 480, + 463 + ], + "type": "text", + "content": ". As discussed in KalmanNet [39], covariance estimation is intractable when dealing with high-dimensional signals. Additionally, the second-order statistical moments are only used for calculating Kalman gain. Inspired by this, we directly learn the gains from the data distribution and do not explicitly maintain an estimation of covariances. Additionally, we follow [51] to include " + }, + { + "bbox": [ + 130, + 367, + 480, + 463 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{z}}_1" + }, + { + "bbox": [ + 130, + 367, + 480, + 463 + ], + "type": "text", + "content": " of the first frame as anchor into KGN for Kalman gain estimation. Then, the final predicted " + }, + { + "bbox": [ + 130, + 367, + 480, + 463 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_t" + }, + { + "bbox": [ + 130, + 367, + 480, + 463 + ], + "type": "text", + "content": " can be derived by" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 277, + 464, + 480, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 464, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 277, + 464, + 480, + 477 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {y}} _ {t} = g _ {\\theta} \\left(\\hat {\\boldsymbol {z}} _ {t} ^ {+}\\right). \\tag {10}", + "image_path": "526437d7f4a846292ad71f02de247df62ed6a391a4db27259b487e93719c66d1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 497, + 273, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 273, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 273, + 508 + ], + "type": "text", + "content": "3.2 Parameterized Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 520, + 479, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 479, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 479, + 556 + ], + "type": "text", + "content": "As shown in Fig. 2 (b), we will parameterize or define the system dynamics " + }, + { + "bbox": [ + 130, + 520, + 479, + 556 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 520, + 479, + 556 + ], + "type": "text", + "content": ", observation estimator " + }, + { + "bbox": [ + 130, + 520, + 479, + 556 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 130, + 520, + 479, + 556 + ], + "type": "text", + "content": ", Kalman Gain Nets (KGN), and generative model " + }, + { + "bbox": [ + 130, + 520, + 479, + 556 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 130, + 520, + 479, + 556 + ], + "type": "text", + "content": " in this section. The overall framework is shown in Fig. 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 570, + 479, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 570, + 479, + 642 + ], + "spans": [ + { + "bbox": [ + 131, + 570, + 479, + 642 + ], + "type": "text", + "content": "Generative Model. The generative model " + }, + { + "bbox": [ + 131, + 570, + 479, + 642 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 131, + 570, + 479, + 642 + ], + "type": "text", + "content": " is generally parameterized by a backbone of CodeFormer [56], which consists of a LQ encoder " + }, + { + "bbox": [ + 131, + 570, + 479, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_L" + }, + { + "bbox": [ + 131, + 570, + 479, + 642 + ], + "type": "text", + "content": ", a HQ encoder " + }, + { + "bbox": [ + 131, + 570, + 479, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_H" + }, + { + "bbox": [ + 131, + 570, + 479, + 642 + ], + "type": "text", + "content": ", a codebook lookup Transformer and quantization layer " + }, + { + "bbox": [ + 131, + 570, + 479, + 642 + ], + "type": "inline_equation", + "content": "T_{Q}" + }, + { + "bbox": [ + 131, + 570, + 479, + 642 + ], + "type": "text", + "content": ", and a decoder " + }, + { + "bbox": [ + 131, + 570, + 479, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 131, + 570, + 479, + 642 + ], + "type": "text", + "content": ". For simplicity, " + }, + { + "bbox": [ + 131, + 570, + 479, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_Q" + }, + { + "bbox": [ + 131, + 570, + 479, + 642 + ], + "type": "text", + "content": " denotes the decoder with the codebook lookup Transformer and quantization layer absorbed. Basically, the predicted " + }, + { + "bbox": [ + 131, + 570, + 479, + 642 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_t" + }, + { + "bbox": [ + 131, + 570, + 479, + 642 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 131, + 570, + 479, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_Q(\\hat{\\pmb{z}}_t^+)" + }, + { + "bbox": [ + 131, + 570, + 479, + 642 + ], + "type": "text", + "content": ", and the observed state " + }, + { + "bbox": [ + 131, + 570, + 479, + 642 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{z}}_t" + }, + { + "bbox": [ + 131, + 570, + 479, + 642 + ], + "type": "text", + "content": " is approximated by" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 259, + 653, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 653, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 259, + 653, + 480, + 666 + ], + "type": "interline_equation", + "content": "\\tilde {z} _ {t} = e \\left(\\boldsymbol {x} _ {t}\\right) = \\mathcal {E} _ {L} \\left(\\boldsymbol {x} _ {t}\\right). \\tag {11}", + "image_path": "35654f7af8d27ee92e67fd703658e8fb1a5724e845791f55e5f5d7ed33116981.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "KEEP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "State Dynamic System. The system dynamics define how the system evolves over time, and it incorporates any control inputs that might affect the current state. The prediction for the state " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_t" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " at the current timestep is achieved via state extrapolation. In particular, given the posterior estimation of previous state " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{z}}_{t-1}^+" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ", we define the dynamic model by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 213, + 186, + 481, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 186, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 213, + 186, + 481, + 200 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {z}} _ {t} ^ {-} = f \\left(\\hat {\\boldsymbol {z}} _ {t - 1} ^ {+}\\right) = \\mathcal {E} _ {H} \\left(\\omega \\left(\\mathcal {D} _ {Q} \\left(\\hat {\\boldsymbol {z}} _ {t - 1} ^ {+}\\right), \\Phi_ {t - 1 \\rightarrow t}\\right)\\right). \\tag {12}", + "image_path": "5cb474c4e8d1460b3d2d965425dee16413c17898ed5742421b11708588514297.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 209, + 482, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 482, + 270 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 209, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\varPhi_{t-1\\rightarrow t}" + }, + { + "bbox": [ + 130, + 209, + 482, + 270 + ], + "type": "text", + "content": " denotes the flow estimated from LQ frames " + }, + { + "bbox": [ + 130, + 209, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t-1}" + }, + { + "bbox": [ + 130, + 209, + 482, + 270 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 209, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 130, + 209, + 482, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 209, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 209, + 482, + 270 + ], + "type": "text", + "content": " the spatial warping modules. Specifically, we first decode the estimated code " + }, + { + "bbox": [ + 130, + 209, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{z}}_{t-1}^{+}" + }, + { + "bbox": [ + 130, + 209, + 482, + 270 + ], + "type": "text", + "content": " of the previous frame to get the estimation of " + }, + { + "bbox": [ + 130, + 209, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_{t-1} = \\mathcal{D}_Q(\\hat{\\pmb{z}}_{t-1}^+)" + }, + { + "bbox": [ + 130, + 209, + 482, + 270 + ], + "type": "text", + "content": " and warp it to the current frame. Then, it was encoded back in the latent space to obtain the prediction of the current state " + }, + { + "bbox": [ + 130, + 209, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{z}}_t^{-}" + }, + { + "bbox": [ + 130, + 209, + 482, + 270 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 281, + 482, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 341 + ], + "type": "text", + "content": "Kalman Filter System. Given the approximated observed state " + }, + { + "bbox": [ + 130, + 281, + 482, + 341 + ], + "type": "inline_equation", + "content": "\\tilde{z}_t" + }, + { + "bbox": [ + 130, + 281, + 482, + 341 + ], + "type": "text", + "content": " and prior estimation " + }, + { + "bbox": [ + 130, + 281, + 482, + 341 + ], + "type": "inline_equation", + "content": "\\hat{z}_t^-" + }, + { + "bbox": [ + 130, + 281, + 482, + 341 + ], + "type": "text", + "content": " from system dynamics, the filter system aims to promote temporal information propagation and maintain stable latent code priors. In particular, the filter recursively fuses both the estimation to form a more accurate posterior estimate of the current state " + }, + { + "bbox": [ + 130, + 281, + 482, + 341 + ], + "type": "inline_equation", + "content": "\\hat{z}_t^+" + }, + { + "bbox": [ + 130, + 281, + 482, + 341 + ], + "type": "text", + "content": ", which is also known as state update." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 342, + 483, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 483, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 483, + 473 + ], + "type": "text", + "content": "According to Eqn. 8, a more intuitive way to express the updated state is a linear interpolation " + }, + { + "bbox": [ + 130, + 342, + 483, + 473 + ], + "type": "inline_equation", + "content": "\\hat{z}_t^+ = \\mathcal{K}_t\\hat{z}_t^- + (1 - \\mathcal{K}_t)\\tilde{z}_t" + }, + { + "bbox": [ + 130, + 342, + 483, + 473 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 342, + 483, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_t" + }, + { + "bbox": [ + 130, + 342, + 483, + 473 + ], + "type": "text", + "content": " normalized in the range of " + }, + { + "bbox": [ + 130, + 342, + 483, + 473 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 130, + 342, + 483, + 473 + ], + "type": "text", + "content": ". The Kalman Gain " + }, + { + "bbox": [ + 130, + 342, + 483, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_t" + }, + { + "bbox": [ + 130, + 342, + 483, + 473 + ], + "type": "text", + "content": " measures the estimated accuracy of the predicted states compared to the approximated observed states, to update the state and reduce the uncertainty. As illustrated in Fig. 3, the Kalman gain is approximated via Kalman Gain Network (KGN), consisting of two distinct parameterized modules, i.e., uncertainty network and gain network. The uncertainty associated with the current prediction is implicitly estimated by an uncertainty network constructed by spatial-temporal attention [47] and temporal attention layers (or any other recurrent networks). Then, a gain network calculates Kalman gain " + }, + { + "bbox": [ + 130, + 342, + 483, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_t" + }, + { + "bbox": [ + 130, + 342, + 483, + 473 + ], + "type": "text", + "content": " for each code token. Please refer to the supplementary files for the detailed architectures." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 491, + 304, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 304, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 304, + 503 + ], + "type": "text", + "content": "3.3 Local Temporal Consistency" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 576 + ], + "type": "text", + "content": "Inspired by [51], we adopt cross-frame attention (CFA) layers in the decoder to further promote local temporal consistency to regularize the information propagation. Specifically, given the latent features from the previous frame " + }, + { + "bbox": [ + 130, + 510, + 482, + 576 + ], + "type": "inline_equation", + "content": "v_{t-1}" + }, + { + "bbox": [ + 130, + 510, + 482, + 576 + ], + "type": "text", + "content": " and current frame " + }, + { + "bbox": [ + 130, + 510, + 482, + 576 + ], + "type": "inline_equation", + "content": "v_t" + }, + { + "bbox": [ + 130, + 510, + 482, + 576 + ], + "type": "text", + "content": ". They are projected onto the embedding space and output the features " + }, + { + "bbox": [ + 130, + 510, + 482, + 576 + ], + "type": "inline_equation", + "content": "v_i'" + }, + { + "bbox": [ + 130, + 510, + 482, + 576 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 510, + 482, + 576 + ], + "type": "inline_equation", + "content": "v_i' = \\mathrm{Attn}(Q, K, V) = \\mathrm{softmax}\\left(\\frac{QK^T}{\\sqrt{d}}\\right) \\cdot V" + }, + { + "bbox": [ + 130, + 510, + 482, + 576 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 211, + 585, + 481, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 585, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 211, + 585, + 481, + 597 + ], + "type": "interline_equation", + "content": "Q = W _ {Q} \\cdot v _ {t}, K = W _ {K} \\cdot v _ {t - 1}, V = W _ {V} \\cdot v _ {t - 1}. \\tag {13}", + "image_path": "3e2b06f5d47057b74ea603cefe15dc9e2843c3a0f10c385ebd37f5465c1d9fa4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Intuitively, cross-frame attention modules can be regarded as searching and matching similar patches from the previous frame and fusing them correspondingly. This module facilitates temporal information propagation in the decoder. We adopt CFA modules on features of small scale 16 and 32 to avoid introducing blur to the decoded results." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "R. Feng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 230, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 154, + 482, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 154, + 482, + 298 + ], + "spans": [ + { + "bbox": [ + 130, + 154, + 482, + 298 + ], + "type": "text", + "content": "Dataset. VFHQ [48] contains over 15,000 high-quality video clips of diverse interviews and talk shows, where 15,381 clips are used for training and 50 clips are reserved for testing. Each sequence consists of 100 to 900 frames of resolution " + }, + { + "bbox": [ + 130, + 154, + 482, + 298 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 154, + 482, + 298 + ], + "type": "text", + "content": ". Following common practice [6,36,45,48], we adopt blind settings in all experiments. Specifically, we apply random blur, resize, and noise as image-based degradations. Moreover, video compression is adopted to control the video quality by changing streaming bitrate. For a comprehensive evaluation, we synthesize three splits of the VFHQ-Test dataset containing different levels of degradation, denoted as VFHQ-mild, VFHQ-medium, and VFHQ-heavy. They follow the same degradation model but differ in the degree of noise, blur, and compression. In addition to synthetic data, we also collect real corrupted video for testing. Please refer to the supplementary files for more details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 300, + 482, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 409 + ], + "type": "text", + "content": "Alignment. Pre-trained image models for face restoration are trained on the FFHQ dataset [20], where each image is automatically cropped and aligned. Hence, employing pre-trained models requires a similar alignment phase on VFHQ dataset [48]. However, the discrete step (i.e., cropping) is sensitive to the detected locations of landmarks, which can consequently result in unintentional temporal inconsistencies. Inspired by the work of Fox et al. [11] and Tzaban et al. [43], we employ a Gaussian lowpass filter over the landmarks. We find that this smoothing can significantly attenuate the inconsistencies induced by the alignment step. See supplementary files for more details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "text", + "content": "Implementations. For all stages of training, we initialize all networks with Kaiming Normal [15] and train them using Adam optimizer [22], and a batch size of 4 for all the experiments. The learning rate is set to " + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "text", + "content": " for stages I and II, and " + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "text", + "content": " for stage III. The models are trained with " + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "inline_equation", + "content": "800k" + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "inline_equation", + "content": "400k" + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "inline_equation", + "content": "50k" + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "text", + "content": " iterations for three stages, respectively. We implement our models with PyTorch [38] and train them using NVIDIA Tesla V100 GPUs. Hyper-parameters " + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "inline_equation", + "content": "\\lambda_{VGG}" + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "inline_equation", + "content": "\\lambda_{GAN}" + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "text", + "content": " are set to " + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "text", + "content": ", 1, and " + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-2}" + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "text", + "content": ". We use GMFlow [49] for optical flow estimation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Metrics. For quantitative evaluation, we evaluate the fidelity of restoration using PSNR, SSIM, and LPIPS [54]. In addition, we evaluate the identity preservation scores, termed as " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "IDS" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ", by cosine similarity of the off-the-shelf identity detection network ArcFace [9]. Besides, we follow Tzaban et al. [43] to measure the pose consistency using Average Keypoint Distance (AKD), which is quantified by the average distance of detected landmarks between the generated and ground-truth video frames. In addition to the above single-frame quality evaluation, we also measure the fluctuation of identity/landmarks across frames. Hence, " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma_{IDS}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " measures the standard deviation of identity similarity over the entire video. We expect a considerable identity drift in the generated videos without local identity jitter, where " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma_{IDS}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " is supposed to be low. Similarly, we use " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma_{AKD}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " to measure the standard deviation of keypoint distances over the video, which quantifies the temporal consistency of the pose." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "KEEP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 152, + 157, + 462, + 265 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Quantitative comparison on the VFHQ-mild. Red and Blue indicate the best and the second best results. Full results on other test partitions (medium and heavy) are presented in the supplementary material." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 152, + 157, + 462, + 265 + ], + "lines": [ + { + "bbox": [ + 152, + 157, + 462, + 265 + ], + "spans": [ + { + "bbox": [ + 152, + 157, + 462, + 265 + ], + "type": "table", + "html": "
Method|PSNR↑ SSIM↑ LPIPS↓| IDS↑ AKD↓ |σIDS(×10-2) ↓ σAKD ↓
GPEN [52]25.51930.75170.29880.714211.46914.7416
GFPGAN [45]26.29330.77950.24820.743710.54674.5700
RestoreFormer [46]25.57200.73440.31950.753010.53544.7159
CodeFormer [56]24.65970.74540.27420.627211.49836.3726
EDVR [44]26.60510.78580.24840.719511.62204.8048
BasicVSR [3]26.04580.77650.24960.697311.36795.0343
BasicVSR++ [5]27.19960.80570.19580.764111.31365.2543
KEEP (Ours)27.99940.82670.16190.79608.81823.6866
", + "image_path": "5e7eaf8c4efaa15eff7bbd6e7ebd03af808f10c4cc8d596e912eb93e224f4ce1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 297, + 380, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 380, + 309 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 380, + 309 + ], + "type": "text", + "content": "4.1 Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 343, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 416 + ], + "type": "text", + "content": "Baselines. We compare our method to two categories of approaches. i) Image-based Face SR models (CodeFormer [56], GPEN [52], GFPGAN [45], Restore-Former [46]) are used to generate face videos frame-by-frame. ii) We retrain the general VSR models (EDVR [44], BasicVSR [3], BasicVSR++ [5]) on VFHQ dataset [48]. The degradation settings remain unchanged as our experiments while other training settings follow their original papers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 421, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 482, + 553 + ], + "type": "text", + "content": "Quantitative Evaluation. The quantitative results are listed in Table 1. We observe that our method achieves better results than existing methods across all the metrics. The results indicate that KEEP can faithfully recover facial details while preserving the identity. Our method also maintains temporal coherence across frames, as quantified by " + }, + { + "bbox": [ + 130, + 421, + 482, + 553 + ], + "type": "inline_equation", + "content": "\\sigma_{IDS}" + }, + { + "bbox": [ + 130, + 421, + 482, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 421, + 482, + 553 + ], + "type": "inline_equation", + "content": "\\sigma_{AKD}" + }, + { + "bbox": [ + 130, + 421, + 482, + 553 + ], + "type": "text", + "content": ", which represent the fluctuation of restored face identities and facial shapes. Though exhibiting structural distortions and artifacts (See Fig. 4), general VSR models (EDVR, BasicVSR, BasicVSR++) typically achieve higher performance on fidelity metrics (PNSR, SSIM, LPIPS) than single-image FSR models. This suggests that image-based models produce high-quality but relatively low-fidelity results. Inconsistency could be introduced when latent code estimation is noisy and inaccurate." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Qualitative Evaluation. In Fig. 4, we can observe that the compared methods fail to reconstruct consistent appearances with perceptually pleasant details. For example, GFPGAN tends to hallucinate facial details (e.g., ears in the second frame and incomplete glass in the last frame of the left example). CodeFormer produces unnatural facial shapes (e.g., eyes), and BasicVSR leaves severe artifacts on the face images. In the last frame of the right example, both GFPGAN and CodeFormer generate unpleasant eyes (see yellow arrows). In contrast, our KEEP exploits temporal information and restores finer and coherent facial details. We refer readers to supplementary files for more video results." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "R. Feng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 114, + 465, + 345 + ], + "blocks": [ + { + "bbox": [ + 153, + 114, + 465, + 345 + ], + "lines": [ + { + "bbox": [ + 153, + 114, + 465, + 345 + ], + "spans": [ + { + "bbox": [ + 153, + 114, + 465, + 345 + ], + "type": "image", + "image_path": "fc23af2204512c17dda0e651fb9a307526d9781ac2362fa99183d8339e29b254.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 352, + 480, + 374 + ], + "lines": [ + { + "bbox": [ + 131, + 352, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 131, + 352, + 480, + 374 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparison on the VFHQ. Our KEEP produces high-fidelity face videos with faithful and consistent details. See arrows for details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 152, + 422, + 282, + 476 + ], + "blocks": [ + { + "bbox": [ + 132, + 390, + 301, + 411 + ], + "lines": [ + { + "bbox": [ + 132, + 390, + 301, + 411 + ], + "spans": [ + { + "bbox": [ + 132, + 390, + 301, + 411 + ], + "type": "text", + "content": "Table 2: Ablation study of variant networks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 152, + 422, + 282, + 476 + ], + "lines": [ + { + "bbox": [ + 152, + 422, + 282, + 476 + ], + "spans": [ + { + "bbox": [ + 152, + 422, + 282, + 476 + ], + "type": "table", + "html": "
ModelsLPIPS↓IDS↑AKD↓
w/o CFA0.16210.79708.9029
w/o KFN0.17210.77739.1952
Full model0.16190.79608.8182
", + "image_path": "b646ac0f5ad2c369438e50314ba317c27cbe4f15bb101450d802806be25830a2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 323, + 422, + 470, + 460 + ], + "blocks": [ + { + "bbox": [ + 312, + 390, + 480, + 411 + ], + "lines": [ + { + "bbox": [ + 312, + 390, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 312, + 390, + 480, + 411 + ], + "type": "text", + "content": "Table 3: Ablation study on optical flow estimator." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 422, + 470, + 460 + ], + "lines": [ + { + "bbox": [ + 323, + 422, + 470, + 460 + ], + "spans": [ + { + "bbox": [ + 323, + 422, + 470, + 460 + ], + "type": "table", + "html": "
EstimatorLPIPS↓IDS↑AKD↓
PWC-Net [41]0.16230.79578.7839
GMFlow [49] (Ours)0.16190.79608.8182
", + "image_path": "1fa49b1218c0cb7e75d0430f94eae01454010028f424c6443c0f64a7560e3f37.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 502, + 234, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 234, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 234, + 514 + ], + "type": "text", + "content": "4.2 More Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 532, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 480, + 581 + ], + "type": "text", + "content": "Effectiveness of KFN. We first investigate the effectiveness of the Kalman Filter Network (KFN). As shown in Table 2, removing KFN results in worse performance on LPIPS, IDS, and AKD. The results suggest the design of KFN is the key to promoting temporal consistency and identity preservation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "type": "text", + "content": "Effectiveness of CFA. Table 2 also shows that Cross-Frame Attention (CFA) can further improve the performance. Though not clearly reflected in the number, we further qualitatively show the effectiveness of KFN and CFA in the supplementary video. From the demo video, we can observe that 1) adopting KFN can ensure consistency in global style and maintain the global appearance of the recovered faces. 2) adding CFA can further coherently render local texture details (e.g., hair), and suppress flicker." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "KEEP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 115, + 211, + 178 + ], + "blocks": [ + { + "bbox": [ + 151, + 115, + 211, + 178 + ], + "lines": [ + { + "bbox": [ + 151, + 115, + 211, + 178 + ], + "spans": [ + { + "bbox": [ + 151, + 115, + 211, + 178 + ], + "type": "image", + "image_path": "3c1afdfa52b4f2ab68793f92afec9aeba7e4a89e495e11a6bba058a2d9678cac.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 151, + 178, + 211, + 243 + ], + "blocks": [ + { + "bbox": [ + 151, + 178, + 211, + 243 + ], + "lines": [ + { + "bbox": [ + 151, + 178, + 211, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 178, + 211, + 243 + ], + "type": "image", + "image_path": "d9f67801607784ec21ce04e0148a2799d7ff888633b7870cfe3fed703681eaaa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 253, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 253, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 482, + 319 + ], + "type": "text", + "content": "Fig. 5: Comparison of temporal flicker. We select each frame's column (red lines) and show the changes across time. Image-based models (GFPGAN, CodeFormer, and RestoreFormer) have obvious discontinuity around the eyes and wrinkles, and general VSR methods leave artifacts behind. In contrast, by maintaining stable facial priors and aggregating temporal information, our method remarkably suppresses temporal jitters and promotes coherent local details." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 212, + 115, + 275, + 178 + ], + "blocks": [ + { + "bbox": [ + 212, + 115, + 275, + 178 + ], + "lines": [ + { + "bbox": [ + 212, + 115, + 275, + 178 + ], + "spans": [ + { + "bbox": [ + 212, + 115, + 275, + 178 + ], + "type": "image", + "image_path": "a769f221c8655c2c4d31b4aba16fa4d25a3273f51948f8d8b6c594a201093b86.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 212, + 178, + 275, + 243 + ], + "blocks": [ + { + "bbox": [ + 212, + 178, + 275, + 243 + ], + "lines": [ + { + "bbox": [ + 212, + 178, + 275, + 243 + ], + "spans": [ + { + "bbox": [ + 212, + 178, + 275, + 243 + ], + "type": "image", + "image_path": "705b2c2e9d5f7f413a65527283e01a4d374813157d95352c859ae2a2cc1e0458.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 276, + 115, + 338, + 178 + ], + "blocks": [ + { + "bbox": [ + 276, + 115, + 338, + 178 + ], + "lines": [ + { + "bbox": [ + 276, + 115, + 338, + 178 + ], + "spans": [ + { + "bbox": [ + 276, + 115, + 338, + 178 + ], + "type": "image", + "image_path": "370b3538f8035b080f502b022a025b798d7b5e40657439ba67484d75a7c5ca30.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 276, + 178, + 337, + 242 + ], + "blocks": [ + { + "bbox": [ + 276, + 178, + 337, + 242 + ], + "lines": [ + { + "bbox": [ + 276, + 178, + 337, + 242 + ], + "spans": [ + { + "bbox": [ + 276, + 178, + 337, + 242 + ], + "type": "image", + "image_path": "5ec4c63cd7e3b9ffa055bdb6d17abef62435df9641965d9b4b6046552a732d3f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 338, + 115, + 400, + 178 + ], + "blocks": [ + { + "bbox": [ + 338, + 115, + 400, + 178 + ], + "lines": [ + { + "bbox": [ + 338, + 115, + 400, + 178 + ], + "spans": [ + { + "bbox": [ + 338, + 115, + 400, + 178 + ], + "type": "image", + "image_path": "ff53ce63c05153754f218a588abed31c9199f8170fbaac93d3f96ad3138505e2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 338, + 178, + 400, + 242 + ], + "blocks": [ + { + "bbox": [ + 338, + 178, + 400, + 242 + ], + "lines": [ + { + "bbox": [ + 338, + 178, + 400, + 242 + ], + "spans": [ + { + "bbox": [ + 338, + 178, + 400, + 242 + ], + "type": "image", + "image_path": "919718bfb026b6fc12898b988e9ae1a595e4cce48ca701e7cfca532704aca6a8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 400, + 115, + 463, + 178 + ], + "blocks": [ + { + "bbox": [ + 400, + 115, + 463, + 178 + ], + "lines": [ + { + "bbox": [ + 400, + 115, + 463, + 178 + ], + "spans": [ + { + "bbox": [ + 400, + 115, + 463, + 178 + ], + "type": "image", + "image_path": "8d8c7915cbbe0a92ac8880633aa446f8d5d2ea164b2915b50770e2b3b8a6dd16.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 400, + 178, + 463, + 242 + ], + "blocks": [ + { + "bbox": [ + 400, + 178, + 463, + 242 + ], + "lines": [ + { + "bbox": [ + 400, + 178, + 463, + 242 + ], + "spans": [ + { + "bbox": [ + 400, + 178, + 463, + 242 + ], + "type": "image", + "image_path": "707ba80d225e97027b97952a3a63fe31a413399557a6bdd53917aeed221c6a5d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 150, + 331, + 214, + 396 + ], + "blocks": [ + { + "bbox": [ + 150, + 331, + 214, + 396 + ], + "lines": [ + { + "bbox": [ + 150, + 331, + 214, + 396 + ], + "spans": [ + { + "bbox": [ + 150, + 331, + 214, + 396 + ], + "type": "image", + "image_path": "636d1b31318759a86d6ff643d192b591408efdfbfe8b843d27b958397d23d519.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 396, + 191, + 403 + ], + "lines": [ + { + "bbox": [ + 174, + 396, + 191, + 403 + ], + "spans": [ + { + "bbox": [ + 174, + 396, + 191, + 403 + ], + "type": "text", + "content": "(a) GT" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 416, + 482, + 460 + ], + "lines": [ + { + "bbox": [ + 130, + 416, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 482, + 460 + ], + "type": "text", + "content": "Fig.6: Illustration of predicted state and observed state through Decoder " + }, + { + "bbox": [ + 130, + 416, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_Q" + }, + { + "bbox": [ + 130, + 416, + 482, + 460 + ], + "type": "text", + "content": ". The teeth in red boxes illustrate a slight difference between " + }, + { + "bbox": [ + 130, + 416, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{z}}_t" + }, + { + "bbox": [ + 130, + 416, + 482, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 416, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{z}}_t^-" + }, + { + "bbox": [ + 130, + 416, + 482, + 460 + ], + "type": "text", + "content": " when decoded in pixel space, suggesting the potential to supplement and fuse information to obtain a stable latent code." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 214, + 331, + 276, + 396 + ], + "blocks": [ + { + "bbox": [ + 214, + 331, + 276, + 396 + ], + "lines": [ + { + "bbox": [ + 214, + 331, + 276, + 396 + ], + "spans": [ + { + "bbox": [ + 214, + 331, + 276, + 396 + ], + "type": "image", + "image_path": "c611cdf9682c7466939eb0b3fe82e0b3a5c5446344724cdfe586f9fa853bf429.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 396, + 258, + 405 + ], + "lines": [ + { + "bbox": [ + 231, + 396, + 258, + 405 + ], + "spans": [ + { + "bbox": [ + 231, + 396, + 258, + 405 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 231, + 396, + 258, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_Q(\\tilde{z}_t)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 276, + 331, + 338, + 396 + ], + "blocks": [ + { + "bbox": [ + 276, + 331, + 338, + 396 + ], + "lines": [ + { + "bbox": [ + 276, + 331, + 338, + 396 + ], + "spans": [ + { + "bbox": [ + 276, + 331, + 338, + 396 + ], + "type": "image", + "image_path": "fffc8b6a527dde660c741a3f521cbdaf92ab36a1c4b342089bf0ee8f9711d4fd.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 396, + 330, + 405 + ], + "lines": [ + { + "bbox": [ + 282, + 396, + 330, + 405 + ], + "spans": [ + { + "bbox": [ + 282, + 396, + 330, + 405 + ], + "type": "text", + "content": "(c) Error map of (b)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 338, + 331, + 400, + 396 + ], + "blocks": [ + { + "bbox": [ + 338, + 331, + 400, + 396 + ], + "lines": [ + { + "bbox": [ + 338, + 331, + 400, + 396 + ], + "spans": [ + { + "bbox": [ + 338, + 331, + 400, + 396 + ], + "type": "image", + "image_path": "1a710af11c863fa37ef2a4ae3fc4a48c781cfadeb9563ce21e8a519af6050322.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 396, + 383, + 405 + ], + "lines": [ + { + "bbox": [ + 355, + 396, + 383, + 405 + ], + "spans": [ + { + "bbox": [ + 355, + 396, + 383, + 405 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 355, + 396, + 383, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_Q(\\hat{z}_t^-)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 400, + 331, + 463, + 396 + ], + "blocks": [ + { + "bbox": [ + 400, + 331, + 463, + 396 + ], + "lines": [ + { + "bbox": [ + 400, + 331, + 463, + 396 + ], + "spans": [ + { + "bbox": [ + 400, + 331, + 463, + 396 + ], + "type": "image", + "image_path": "d1906e3de3369fdd83d7d162f71d3c89bf3c44a5bc2ebc81fcc493bb2472804e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 396, + 454, + 404 + ], + "lines": [ + { + "bbox": [ + 406, + 396, + 454, + 404 + ], + "spans": [ + { + "bbox": [ + 406, + 396, + 454, + 404 + ], + "type": "text", + "content": "(e) Error map of (d)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "type": "text", + "content": "Effectiveness of Various Flow Estimator. We compare models with different flow estimators in Table 3. The results suggest that the accuracy of estimated flows does not significantly affect the final performance. We conjecture this can be attributed to two factors: 1) Minor misalignment in pixel space can be reasonably diminished as the latent code is highly downsampled by a factor of " + }, + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "type": "inline_equation", + "content": "32 \\times" + }, + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "type": "text", + "content": ", at which level the latent representations are less sensitive to small spatial discrepancies present in the pixel space. 2) Other modules can compensate for the inaccuracy caused by flow estimators in a joint training fashion." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Analysis of Flickering. We extract a short vertical segment of pixels from each frame and stack them horizontally to visualize the jittering issues within the video. In particular, existing methods demonstrate clear jitter and flicker across time, while our method shows better temporal consistency. Fig. 5 demonstrates that other image-based models bring obvious jitters around eyes, and general VSR methods leave behind artifacts, while our method could remarkably suppress temporal jitters and promote coherent local details." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "R. Feng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 158, + 114, + 312, + 208 + ], + "blocks": [ + { + "bbox": [ + 158, + 114, + 312, + 208 + ], + "lines": [ + { + "bbox": [ + 158, + 114, + 312, + 208 + ], + "spans": [ + { + "bbox": [ + 158, + 114, + 312, + 208 + ], + "type": "image", + "image_path": "17aa6da14616d029dc97340e685d3e6374af2657b6192604c5a6ed073be8682c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 220, + 482, + 266 + ], + "lines": [ + { + "bbox": [ + 131, + 220, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 482, + 266 + ], + "type": "text", + "content": "Fig. 7: Identity similarity across frames. Our method preserves the identity of input images and exhibits less fluctuation over time. CodeFormer results of two frames (green and red dashed line) exemplify an abrupt change of identity in the right figure, while our method maintains a stable identity both quantitatively and qualitatively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 114, + 457, + 209 + ], + "blocks": [ + { + "bbox": [ + 313, + 114, + 457, + 209 + ], + "lines": [ + { + "bbox": [ + 313, + 114, + 457, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 457, + 209 + ], + "type": "image", + "image_path": "9d2cef2b1af797ad5c641485037945a0673da9178bfb937c7557ee2f0d40b110.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 156, + 279, + 460, + 407 + ], + "blocks": [ + { + "bbox": [ + 156, + 279, + 460, + 407 + ], + "lines": [ + { + "bbox": [ + 156, + 279, + 460, + 407 + ], + "spans": [ + { + "bbox": [ + 156, + 279, + 460, + 407 + ], + "type": "image", + "image_path": "898a5a68770ff2ec7a5586367486dd2c2f5175f4c56ae7106b96fb7b1071d2c9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 415, + 480, + 437 + ], + "lines": [ + { + "bbox": [ + 131, + 415, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 480, + 437 + ], + "type": "text", + "content": "Fig. 8: Qualitative comparison on different levels of degradation. Our KEEP maintains high-fidelity in various degradations." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "type": "text", + "content": "Analysis of Latent Space. Since the true state " + }, + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_t" + }, + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "type": "text", + "content": " is unavailable, we indirectly analyze the predicted state " + }, + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{z}}_t" + }, + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "type": "text", + "content": " and observed state " + }, + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{z}}_t^-" + }, + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "type": "text", + "content": " by decoding them to the pixel space through Decoder " + }, + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_Q" + }, + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "type": "text", + "content": ". As shown in Fig. 6, the areas around teeth exhibit a slight difference, while most remaining parts in error maps show similar decoded results. This indicates that the predicted and observed states could supplement each other to obtain a more accurate estimation of " + }, + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_t" + }, + { + "bbox": [ + 130, + 462, + 482, + 544 + ], + "type": "text", + "content": ", which is where the power of Kalman filter lies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "type": "text", + "content": "Identity Preservation. We show the identity similarity across frames of one representative video clip in Fig. 7 As illustrated in the left figure, our method achieves better identity preservation and less identity jitter within the video, compared to the single-image methods. We also exemplify that the identity of CodeFormer results can change abruptly within several frames. The identity score of CodeFormer increases from 0.4552 to 0.6392, and turns down in later frames (See red curve), demonstrating great fluctuation across time. In contrast, our method maintains a stable identity both quantitatively and qualitatively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Various Degradations. Fig. 9a shows that KEEP is consistently better than the compared methods across different difficulty levels. Fig. 8 demonstrates re" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 420, + 91, + 447, + 100 + ], + "type": "text", + "content": "KEEP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 119, + 294, + 213 + ], + "blocks": [ + { + "bbox": [ + 140, + 119, + 294, + 213 + ], + "lines": [ + { + "bbox": [ + 140, + 119, + 294, + 213 + ], + "spans": [ + { + "bbox": [ + 140, + 119, + 294, + 213 + ], + "type": "image", + "image_path": "17c6a6f32090f5ba3e07d3bc6bde30371f59407bebd810a12c91229a3f93bfd0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 142, + 220, + 294, + 313 + ], + "blocks": [ + { + "bbox": [ + 142, + 220, + 294, + 313 + ], + "lines": [ + { + "bbox": [ + 142, + 220, + 294, + 313 + ], + "spans": [ + { + "bbox": [ + 142, + 220, + 294, + 313 + ], + "type": "image", + "image_path": "458dc2f108eacd6d5c58937491bf6fb7d9cd762d2c2e8eb54214ca17d6797b14.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 319, + 292, + 328 + ], + "lines": [ + { + "bbox": [ + 141, + 319, + 292, + 328 + ], + "spans": [ + { + "bbox": [ + 141, + 319, + 292, + 328 + ], + "type": "text", + "content": "(a) PNSR and IDS on different partitions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 304, + 125, + 479, + 186 + ], + "blocks": [ + { + "bbox": [ + 304, + 125, + 479, + 186 + ], + "lines": [ + { + "bbox": [ + 304, + 125, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 479, + 186 + ], + "type": "image", + "image_path": "1f233c901cc9f5fd087b4ef429260c4c91ab67f0c8b6417530003c1809221778.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 305, + 186, + 479, + 245 + ], + "blocks": [ + { + "bbox": [ + 305, + 186, + 479, + 245 + ], + "lines": [ + { + "bbox": [ + 305, + 186, + 479, + 245 + ], + "spans": [ + { + "bbox": [ + 305, + 186, + 479, + 245 + ], + "type": "image", + "image_path": "ab73be52ebc623cbb65365755e714283da6f4571c9967723c2590042cbcbb88a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 304, + 246, + 479, + 313 + ], + "blocks": [ + { + "bbox": [ + 304, + 246, + 479, + 313 + ], + "lines": [ + { + "bbox": [ + 304, + 246, + 479, + 313 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 479, + 313 + ], + "type": "image", + "image_path": "1ebcfac9ccb6174376d8145b59177eff45dafab47885fee29070080894b41b2c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 319, + 459, + 328 + ], + "lines": [ + { + "bbox": [ + 324, + 319, + 459, + 328 + ], + "spans": [ + { + "bbox": [ + 324, + 319, + 459, + 328 + ], + "type": "text", + "content": "(b) Comparisons on non-frontal faces." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 338, + 482, + 371 + ], + "lines": [ + { + "bbox": [ + 131, + 338, + 482, + 371 + ], + "spans": [ + { + "bbox": [ + 131, + 338, + 482, + 371 + ], + "type": "text", + "content": "Fig. 9: (a) Our methods achieve consistently better performance on various levels of degradation. (b) While CodeFormer fails to restore eyes to these challenging cases, our method can still produce plausible facial elements." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "content": "sults on different levels of degradation. We draw the following observations. 1) Even GFPGAN and CodeFormer can restore plausible results on frames with mild degradation, the performance significantly deteriorates (e.g., eyes) upon heavier degradation. 2) Our method achieves better results by considering complementary information between adjacent frames and maintaining stable face priors. In addition, our method is appealing in handling heavy degradation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 469, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 469, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 469, + 482, + 518 + ], + "type": "text", + "content": "Non-Frontal-View Faces. In Fig. 9b, our model shows enhanced performance on non-frontal faces by providing more stable face priors estimations. While the single-image model CodeFormer cannot recover the eyes, our KEEP is still able to show robustness to these challenging cases." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 540, + 220, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 540, + 220, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 540, + 220, + 553 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "We present a novel framework, KEEP, aiming at resolving the challenges associated with facial detail and temporal consistency in video face restoration. The proposed method demonstrates a unique capability to maintain a stable face prior over time, which is achieved by Kalman filtering principles, where our approach recurrently incorporates information from previously restored frames to guide and regulate the restoration process of the current frame. Extensive experiments demonstrate the efficacy of KEEP in consistently capturing facial details across video frames and keeping the temporal stability of face videos." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "R. Feng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 176 + ], + "type": "text", + "content": "This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 217, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 138, + 217, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 481, + 251 + ], + "type": "text", + "content": "1. Caballero, J., Ledig, C., Aitken, A., Acosta, A., Totz, J., Wang, Z., Shi, W.: Real-time video super-resolution with spatio-temporal networks and motion compensation. In: CVPR (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 481, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 481, + 272 + ], + "type": "text", + "content": "2. Chan, K.C., Wang, X., Xu, X., Gu, J., Loy, C.C.: GLEAN: Generative latent bank for large-factor image super-resolution. In: CVPR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 274, + 481, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 481, + 294 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 481, + 294 + ], + "type": "text", + "content": "3. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: BasicVSR: The search for essential components in video super-resolution and beyond. In: CVPR (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 295, + 481, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 295, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 295, + 481, + 316 + ], + "type": "text", + "content": "4. Chan, K.C., Xu, X., Wang, X., Gu, J., Loy, C.C.: GLEAN: Generative latent bank for large-factor image super-resolution and beyond. TPAMI (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 317, + 481, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 317, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 138, + 317, + 481, + 338 + ], + "type": "text", + "content": "5. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: CVPR (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 338, + 481, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 338, + 481, + 359 + ], + "spans": [ + { + "bbox": [ + 138, + 338, + 481, + 359 + ], + "type": "text", + "content": "6. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Investigating tradeoffs in real-world video super-resolution. In: CVPR (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 361, + 481, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 481, + 381 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 481, + 381 + ], + "type": "text", + "content": "7. Chen, C., Li, X., Yang, L., Lin, X., Zhang, L., Wong, K.Y.K.: Progressive semantic-aware style transformation for blind face restoration. In: CVPR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 382, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 382, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 382, + 481, + 403 + ], + "type": "text", + "content": "8. Chen, Y., Tai, Y., Liu, X., Shen, C., Yang, J.: FSRNet: End-to-end learning face super-resolution with facial priors. In: CVPR (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 404, + 481, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 481, + 425 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 481, + 425 + ], + "type": "text", + "content": "9. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: ArcFace: Additive angular margin loss for deep face recognition. In: CVPR (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 426, + 481, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 481, + 447 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 481, + 447 + ], + "type": "text", + "content": "10. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: CVPR (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 448, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 448, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 448, + 481, + 468 + ], + "type": "text", + "content": "1. Fox, G., Tewari, A., Elgharib, M., Theobalt, C.: StyleVideoGAN: A temporal generative model using a pretrained stylegan. In: BMVC (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 470, + 481, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 470, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 470, + 481, + 491 + ], + "type": "text", + "content": "2. Gu, J., Shen, Y., Zhou, B.: Image processing using multi-code gan prior. In: CVPR (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 491, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 481, + 523 + ], + "type": "text", + "content": "3. Gu, Y., Wang, X., Xie, L., Dong, C., Li, G., Shan, Y., Cheng, M.M.: VQFR: Blind face restoration with vector-quantized dictionary and parallel decoder. In: ECCV (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 524, + 481, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 524, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 524, + 481, + 545 + ], + "type": "text", + "content": "4. Haris, M., Shakhnarovich, G., Ukita, N.: Recurrent back-projection network for video super-resolution. In: CVPR (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 546, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 481, + 567 + ], + "type": "text", + "content": "5. He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In: ICCV (2015)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 567, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 589 + ], + "type": "text", + "content": "6. Huang, Y., Wang, W., Wang, L.: Bidirectional recurrent convolutional networks for multi-frame super-resolution. In: NeurIPS (2015)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 590, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 481, + 610 + ], + "type": "text", + "content": "7. Isobe, T., Jia, X., Gu, S., Li, S., Wang, S., Tian, Q.: Video super-resolution with recurrent structure-detail network. In: ECCV (2020)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "text", + "content": "8. Isobe, T., Zhu, F., Wang, S.: Revisiting temporal modeling for video superresolution. In: BMVC (2020)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 633, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 643 + ], + "type": "text", + "content": "9. Kalman, R.E.: A new approach to linear filtering and prediction problems (1960)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 138, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 481, + 665 + ], + "type": "text", + "content": "20. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "KEEP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "21. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: CVPR (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "22. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "23. Lei, C., Ren, X., Zhang, Z., Chen, Q.: Blind video deflickering by neural filtering with a flawed atlas. In: CVPR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "24. Lei, C., Xing, Y., Chen, Q.: Blind video temporal consistency via deep video prior. In: NeurIPS (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "25. Li, S., He, F., Du, B., Zhang, L., Xu, Y., Tao, D.: Fast spatio-temporal residual network for video super-resolution. In: CVPR (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "26. Li, X., Chen, C., Zhou, S., Lin, X., Zuo, W., Zhang, L.: Blind face restoration via deep multi-scale component dictionaries. In: ECCV (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "27. Li, X., Li, W., Ren, D., Zhang, H., Wang, M., Zuo, W.: Enhanced blind face restoration with multi-exemplar images and adaptive spatial feature fusion. In: CVPR (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "text", + "content": "28. Li, X., Liu, M., Ye, Y., Zuo, W., Lin, L., Yang, R.: Learning warped guidance for blind face restoration. In: ECCV (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "29. Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: Vrt: A video restoration transformer. IEEE Transactions on Image Processing (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "30. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided deformable attention. NeurIPS (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "31. Liu, C., Yang, H., Fu, J., Qian, X.: Learning trajectory-aware transformer for video super-resolution. In: CVPR (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "32. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: CVPR (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 414, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 435 + ], + "type": "text", + "content": "33. Menon, S., Damian, A., Hu, S., Ravi, N., Rudin, C.: Pulse: Self-supervised photoupsampling via latent space exploration of generative models. In: CVPR (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 436, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 480, + 456 + ], + "type": "text", + "content": "34. Nagrani, A., Chung, J.S., Zisserman, A.: Voxceleb: A large-scale speaker identification dataset. In: Interspeech (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 457, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 490 + ], + "type": "text", + "content": "35. Ouyang, H., Wang, Q., Xiao, Y., Bai, Q., Zhang, J., Zheng, K., Zhou, X., Chen, Q., Shen, Y.: Codef: Content deformation fields for temporally consistent video processing. In: CVPR (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "text", + "content": "36. Pan, J., Bai, H., Dong, J., Zhang, J., Tang, J.: Deep blind video super-resolution. In: ICCV (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "37. Pan, X., Zhan, X., Dai, B., Lin, D., Loy, C.C., Luo, P.: Exploiting deep generative prior for versatile image restoration and manipulation. TPAMI 44(11), 7474-7489 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "text", + "content": "38. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "39. Revach, G., Shlezinger, N., Ni, X., Escoriza, A.L., Van Sloun, R.J., Eldar, Y.C.: Kalmannet: Neural network aided kalman filtering for partially known dynamics. IEEE Transactions on Signal Processing (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "text", + "content": "40. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "41. Sun, D., Yang, X., Liu, M.Y., Kautz, J.: Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In: CVPR (2018)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "42. Tian, Y., Zhang, Y., Fu, Y., Xu, C.: TDAN: Temporally-deformable alignment network for video super-resolution. In: CVPR (2020)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "R. Feng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 445 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "43. Tzaban, R., Mokady, R., Gal, R., Bermano, A., Cohen-Or, D.: Stitch it in time: Gan-based facial editing of real videos. In: SIGGRAPH Asia (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "text", + "content": "44. Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: EDVR: Video restoration with enhanced deformable convolutional networks. In: CVPRW (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "text", + "content": "45. Wang, X., Li, Y., Zhang, H., Shan, Y.: Towards real-world blind face restoration with generative facial prior. In: CVPR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "46. Wang, Z., Zhang, J., Chen, R., Wang, W., Luo, P.: Restoreformer: High-quality blind face restoration from undegraded key-value pairs. In: CVPR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "47. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: ICCV (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "type": "text", + "content": "48. Xie, L., Wang, X., Zhang, H., Dong, C., Shan, Y.: VFHQ: A high-quality dataset and benchmark for video face super-resolution. In: CVPR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "type": "text", + "content": "49. Xu, H., Zhang, J., Cai, J., Rezatofighi, H., Tao, D.: Gmflow: Learning optical flow via global matching. In: CVPR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "type": "text", + "content": "50. Xu, Y., AlBahar, B., Huang, J.B.: Temporally consistent semantic video editing. In: ECCV (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "51. Yang, S., Zhou, Y., Liu, Z.,, Loy, C.C.: Rerender a video: Zero-shot text-guided video-to-video translation. In: SIGGRAPH Asia (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "text", + "content": "52. Yang, T., Ren, P., Xie, X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: CVPR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "type": "text", + "content": "53. Yu, X., Fernando, B., Ghanem, B., Porikli, F., Hartley, R.: Face super-resolution guided by facial component heatmaps. In: ECCV (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 370, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 480, + 391 + ], + "type": "text", + "content": "54. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "text", + "content": "55. Zhou, L., Luo, Z., Shen, T., Zhang, J., Zhen, M., Yao, Y., Fang, T., Quan, L.: KFNet: Learning temporal camera relocalization using Kalman filtering. In: CVPR (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "text", + "content": "56. Zhou, S., Chan, K.C., Li, C., Loy, C.C.: Towards robust blind face restoration with codebook lookup transformer. In: NeurIPS (2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 419, + 91, + 447, + 100 + ], + "type": "text", + "content": "KEEP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Kernel Diffusion_ An Alternate Approach to Blind Deconvolution/0f568881-c0ff-47c0-b18b-592b7cbf5a93_content_list.json b/2024/Kernel Diffusion_ An Alternate Approach to Blind Deconvolution/0f568881-c0ff-47c0-b18b-592b7cbf5a93_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e19a9bc0076970879d33497ed2fbb158fdbd8ba0 --- /dev/null +++ b/2024/Kernel Diffusion_ An Alternate Approach to Blind Deconvolution/0f568881-c0ff-47c0-b18b-592b7cbf5a93_content_list.json @@ -0,0 +1,2071 @@ +[ + { + "type": "text", + "text": "Kernel Diffusion: An Alternate Approach to Blind Deconvolution", + "text_level": 1, + "bbox": [ + 250, + 140, + 754, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yash Sanghvi1, Yiheng Chi1, and Stanley H. Chan1", + "bbox": [ + 295, + 210, + 705, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Elmore Family School of Electrical and Computer Engineering, Purdue University, West Lafayette, Indiana, USA {ysanghvi,chi14,stanchan}@purdue.edu", + "bbox": [ + 225, + 239, + 774, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Blind deconvolution problems are severely ill-posed because neither the underlying signal nor the forward operator are not known exactly. Conventionally, these problems are solved by alternating between estimation of the image and kernel while keeping the other fixed. In this paper, we show that this framework is flawed because of its tendency to get trapped in local minima and, instead, suggest the use of a kernel estimation strategy with a non-blind solver. This framework is employed by a diffusion method which is trained to sample the blur kernel from the conditional distribution with guidance from a pre-trained non-blind solver. The proposed diffusion method leads to state-of-the-art results on both synthetic and real blur datasets.", + "bbox": [ + 261, + 319, + 743, + 472 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Blind Deconvolution $\\cdot$ Diffusion Models $\\cdot$ Deblurring", + "bbox": [ + 261, + 484, + 700, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 526, + 375, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Images taken by a hand-held camera under long exposure suffer from motion blur. Assuming this blur is spatially invariant, the captured image $\\mathbf{y}$ and the underlying sharp image $\\mathbf{x}$ are related as follows:", + "bbox": [ + 212, + 556, + 787, + 603 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} = \\mathbf {k} * \\mathbf {x} + \\mathbf {n}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 446, + 616, + 785, + 631 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $\\mathbf{k}$ , $\\mathbf{n}$ refer to the blur kernel corresponding to the camera motion and noise, respectively, and the operator $\\ast$ is convolution. The inverse problem of estimating $\\mathbf{x}$ given $\\mathbf{y}$ is called deconvolution, which can be classified into non-blind and blind.", + "bbox": [ + 212, + 643, + 787, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the non-blind case, the blur kernel $\\mathbf{k}$ is known. There are plenty of iterative methods which have been successfully applied to various image restoration tasks. Usually, these approaches, such as Plug-and-Play [1,8,34,57,70,73] and Regularization by Denoising (RED) [20,53,56], use a pre-trained image denoiser as a substitute for a complex image prior. Recent deep learning counterparts to these approaches involve unrolling the iterative methods followed by end-to-end training [45,48,82,86].", + "bbox": [ + 212, + 704, + 787, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In blind deconvolution, the blur kernel $\\mathbf{k}$ is not known. A natural strategy for the blind inverse problem is to jointly estimate the blur kernel $\\mathbf{k}$ , and the", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "image $\\mathbf{x}$ by maximizing the posterior probability:", + "bbox": [ + 215, + 146, + 573, + 162 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\mathbf {x}}, \\hat {\\mathbf {k}} = \\underset {\\mathbf {x}, \\mathbf {k}} {\\arg \\max} \\left[ p (\\mathbf {x}, \\mathbf {k} | \\mathbf {y}) \\right] \\\\ = \\underset {\\mathbf {x}, \\mathbf {k}} {\\arg \\max } \\left[ \\log p (\\mathbf {y} | \\mathbf {k}, \\mathbf {x}) + \\log p (\\mathbf {x}) + \\log p (\\mathbf {k}) \\right]. \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 172, + 785, + 232 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Maximum-A-Posteriori (MAP) for the pair $(\\mathbf{x},\\mathbf{k})$ , i.e. $\\mathrm{MAP}_{\\mathbf{x},\\mathbf{k}}$ , can be obtained by an iterative scheme which alternates between the following minimization steps:", + "bbox": [ + 214, + 244, + 785, + 289 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} ^ {i + 1} = \\arg \\min _ {\\mathbf {x}} \\left[ \\| \\mathbf {y} - \\mathbf {x} * \\mathbf {k} ^ {i} \\| ^ {2} + \\phi (\\mathbf {x}) \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 300, + 785, + 325 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {k} ^ {i + 1} = \\arg \\operatorname * {m i n} _ {\\mathbf {k}} \\left[ \\| \\mathbf {y} - \\mathbf {x} ^ {i + 1} * \\mathbf {\\Phi} \\mathbf {k} \\| ^ {2} + \\psi (\\mathbf {k}) \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 328, + 785, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\phi(\\cdot), \\psi(\\cdot)$ represent the image and kernel priors. Many classical deconvolution techniques can be framed as variations of this framework [2,9,63,80]. Even neural network based iterative methods such as Self-Deblur [54] and diffusion models [15,49] have achieved state-of-the-art performance in blind deconvolution using the stated framework.", + "bbox": [ + 214, + 364, + 782, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite recent success of diffusion methods, they are based on $\\mathrm{MAP}_{\\mathbf{x},\\mathbf{k}}$ approach, which was first brought into question by Levin et al. [40] They argued that the number of variables to be estimated, i.e., $\\{\\mathbf{x},\\mathbf{k}\\}$ , are much larger than the number of known variables, i.e., $\\mathbf{y}$ , and as a result, (2) favors the no-blur solution $\\hat{\\mathbf{x}},\\hat{\\mathbf{k}} = (\\mathbf{y},\\mathbf{I})$ . [40] proposed an alternate strategy of solving the kernel $\\mathbf{k}$ first to exploit the inherent symmetry in the problem. This is achieved by marginalizing out the image space as follows:", + "bbox": [ + 214, + 440, + 787, + 547 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {k}} = \\underset {\\mathbf {k}} {\\arg \\max } \\left[ p (\\mathbf {k} | \\mathbf {y}) \\right] = \\underset {\\mathbf {k}} {\\arg \\max } \\left[ \\int p (\\mathbf {k}, \\mathbf {x} | \\mathbf {y}) d \\mathbf {x} \\right]. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 558, + 785, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although $\\mathrm{MAP_k}$ optimization can lead to significantly improved reconstruction (as shown in Figure 1(c)) it has not been adopted in blind deconvolution solutions, which we surmise because of the intrinsic computational difficulty of the integration in (5).", + "bbox": [ + 214, + 597, + 787, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.1 Contributions", + "text_level": 1, + "bbox": [ + 215, + 679, + 377, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we address the problem with a kernel estimation approach, and we demonstrate that this strategy can be implemented in practice. Our contribution in this paper can be summarized as follows: We propose a novel diffusion method, called Kernel-Diffusion (Kernel-Diff) which solves the $\\mathrm{MAP_k}$ optimization problem by sampling from the conditional distribution $p(\\mathbf{k}|\\mathbf{y})$ . This diffusion method coupled with a differentiable non-blind solver achieves state-of-the-art deconvolution performance. We show with numerical experiments on real and synthetic data that the proposed approach creates significant performance improvement. We will soon release the source code and model weights for Kernel-Diff.", + "bbox": [ + 214, + 704, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Sanghvi et al.", + "bbox": [ + 271, + 114, + 385, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6bd8424048f4acdff77721ad1edb5dd90169c5ff1dce62f4fed01bdee083e1a4.jpg", + "image_caption": [ + "(a) Blurred image Inset: true kernel" + ], + "image_footnote": [], + "bbox": [ + 218, + 142, + 357, + 251 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/59cecb71a103eb3e0bc94f99fd9b97c891c0087b1445c9b8fd681fdb65aa1f3a.jpg", + "image_caption": [ + "(b) Blind-DPS [15] based on alternating minimization" + ], + "image_footnote": [], + "bbox": [ + 359, + 142, + 500, + 250 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dd7f9059013ace52f84651fad535dd383568eb78988b192323343882a47889d9.jpg", + "image_caption": [ + "(c) Kernel-Diff (ours) based on kernel estimation" + ], + "image_footnote": [], + "bbox": [ + 504, + 143, + 642, + 250 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1adf6c343a729100a94d8573902c4351359370b2ac3cb917221e3d5fff3c3671.jpg", + "image_caption": [ + "(d) Ground Truth", + "Fig. 1: In this paper, we present a novel diffusion-based approach to the blind deconvolution problem. The conventional approaches, such as Blind-DPS, iterate between between kernel and image space minimization. Our approach uses a kernel-first estimation strategy in the context of diffusion models to sample kernels from the conditional distribution $p(\\mathbf{k}|\\mathbf{y})$ ." + ], + "image_footnote": [], + "bbox": [ + 643, + 143, + 782, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Background", + "text_level": 1, + "bbox": [ + 215, + 410, + 367, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Diffusion Methods", + "text_level": 1, + "bbox": [ + 215, + 445, + 413, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models [29, 64, 67, 74] are a popular generative method which can be used to sample any target diffusion $p_{data}(\\mathbf{x})$ . The diffusion models begin with a forward process, which starts from $\\mathbf{x}_0 \\sim p_{data}(\\mathbf{x})$ and gradually adds noise according to variance schedule $\\{\\beta_1, \\beta_2, \\ldots, \\beta_T\\}$ , formulated as", + "bbox": [ + 214, + 472, + 784, + 532 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t} \\sim \\mathcal {N} \\left(\\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 546, + 784, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "at time steps $t \\in \\{1, 2, \\dots, T\\}$ . This forward process converges to $\\mathbf{x}_T \\sim \\mathcal{N}(0, \\mathbf{I})$ . Next, the reverse diffusion process is used to sample from the distribution of $p_{data}$ by starting with $\\mathbf{x}_T \\sim \\mathcal{N}(0, \\mathbf{I})$ and performing the iterative updates", + "bbox": [ + 214, + 577, + 784, + 623 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {x} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {x} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta^ {*}} \\left(\\mathbf {x} _ {t}, t\\right)\\right) + \\sigma_ {t} \\mathbf {z} _ {t}, \\tag {7} \\\\ \\mathbf {z} _ {t} \\sim \\mathcal {N} (0, \\mathbf {I}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 633, + 784, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{t} \\stackrel{\\mathrm{def}}{=} 1 - \\beta_{t}$ , $\\bar{\\alpha}_{t} \\stackrel{\\mathrm{def}}{=} \\prod_{i=1}^{t} \\alpha_{i}$ , and $\\epsilon_{\\theta^{*}}(\\mathbf{x}_{t}, t)$ is the trained diffusion network used to learn to denoise $\\mathbf{x}_{t}$ at any given time step, trained by minimizing the following loss function.", + "bbox": [ + 214, + 702, + 782, + 760 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\theta) \\stackrel {\\text {d e f}} {=} \\mathbb {E} _ {t, \\mathbf {x} _ {0}, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t\\right) \\| ^ {2} \\right]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 772, + 784, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inverse Problems and Diffusion Models. Score based diffusion methods have been extensively used in solving inverse problems [17, 21, 33, 35-37, 89] by", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Kernel Diffusion", + "bbox": [ + 620, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sampling from the posterior $p(\\mathbf{x}|\\mathbf{y})$ and the corresponding score is calculated using the Bayes' formula", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t} | \\mathbf {y}) = \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t}) + \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {y} | \\mathbf {x} _ {t}). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 186, + 785, + 202 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For example, [31] uses an approximation to the posterior score in (9) and used the Langevin dynamics to sample from the distribution $p(\\mathbf{x}|\\mathbf{y})$ for the MRI compressed sensing problem; [18,19] solve the inpainting and super-resolution inverse problems with a modified reverse diffusion scheme which includes a projection step to reduce the consistency metric $\\| \\mathbf{y} - \\mathbf{A}\\mathbf{x}\\|$ in each iteration; diffusion Posterior Sampling (DPS) [16] solves the noisy inverse problem with alternating diffusion and gradient descent steps.", + "bbox": [ + 212, + 209, + 782, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image Restoration and Diffusion Models Diffusion models sample from a target distribution while MMSE estimators average over them [37]. As a result, the former has been widely used in image restoration tasks such as deblurring [78], super-resolution [41, 58], and low-light enhancement [32, 76]. For an exhaustive review, we refer the reader to [42].", + "bbox": [ + 212, + 315, + 782, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Guidance in Diffusion Models. Pre-trained diffusion models are often guided at each iteration to produce desired results. This is first proposed in [22] where a classifier gradient is added to each iteration for the final output to belong to the given class. ILVR [14] guides the image diffusion process by constraining the low-frequency information of an image. In the same spirit, [75] guides the diffusion process by refining only the null-space of the corresponding inverse problem. [78] guides a deblurring diffusion model using a pretrained structure module.", + "bbox": [ + 212, + 390, + 784, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Blind Deconvolution", + "text_level": 1, + "bbox": [ + 215, + 530, + 429, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Traditional blind deconvolution methods can be categorized under the $\\mathrm{MAP}_{\\mathbf{x},\\mathbf{k}}$ framework [2,9,10,44,52,63,69,80,81]. For example, [13] uses salient edges for the kernel estimation and then performs image estimation using a Total Variation (TV) prior. [79] refines the kernel estimation process by iterative support detection. [28] performs alternating image and kernel minimization but uses a motion density representation for the blur kernel instead of for pixels. Self-Deblur [54] uses an alternating minimization scheme where a latent deep image prior [72] is used to represent the kernel and image space.", + "bbox": [ + 212, + 551, + 782, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Deep learning methods have replaced the optimization-based approaches and usually involve end-to-end training of image restoration networks [38, 62, 68, 71, 77, 83-85] using large datasets of pairs of sharp and blurry images, e.g., GoPro [50] and RealBlur [55] datasets. For a more exhaustive review of deep learning based approaches, we refer the readers to [87].", + "bbox": [ + 212, + 674, + 782, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A few methods have utilized the $\\mathrm{MAP_k}$ approach, even though not always explicitly stated by the authors. [7] uses a fully convolutional network to predict the Fourier coefficients of the underlying kernel which are used to deconvolve the blurry image in Fourier space. [5,6] train a Kernel-Prediction Network for all pixels in the image to deal with spatially varying blur. [60] leverages the differentiability of a non-blind solver [59] to frame the blind deconvolution problem", + "bbox": [ + 212, + 750, + 782, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Sanghvi et al.", + "bbox": [ + 271, + 114, + 385, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "entirely in terms of kernel estimation problems. This strategy is improved by [61] using a key-points based representation to further reduce the dimensionality of the problem.", + "bbox": [ + 214, + 146, + 787, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 213, + 330, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Blind-DPS [15] proposed a diffusion model based scheme for solving the blind inverse problem. Using separate diffusion models to sample from the image and kernel prior, a new scheme was proposed to sample solutions from the joint conditional distribution", + "bbox": [ + 214, + 244, + 782, + 304 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}}, \\hat {\\mathbf {k}} \\sim p (\\mathbf {x}, \\mathbf {k} | \\mathbf {y}) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 446, + 316, + 785, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This method is similar in principle to the conventional blind deconvolution techniques where the goal is to estimate the maximum of the conditional distribution i.e.", + "bbox": [ + 214, + 342, + 782, + 387 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\mathbf {x}}, \\hat {\\mathbf {k}} = \\underset {\\mathbf {x}, \\mathbf {k}} {\\arg \\max } \\left[ p (\\mathbf {x}, \\mathbf {k} | \\mathbf {y}) \\right] (11) \\\\ = \\underset {\\mathbf {x}, \\mathbf {k}} {\\arg \\min } \\left[ \\left\\| \\mathbf {y} - \\mathbf {k} * \\mathbf {x} \\right\\| ^ {2} + R _ {\\mathbf {x}} (\\mathbf {x}) + R _ {\\mathbf {k}} (\\mathbf {k}) \\right] (12) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 395, + 785, + 457 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $R_{\\mathbf{x}}(.)$ and $R_{\\mathbf{k}}(.)$ represent the image and kernel priors respectively. This minimization is achieved using alternating minimization techniques where the kernel is estimated while keeping the image estimate fixed and vice versa.", + "bbox": [ + 214, + 467, + 782, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recently, Blind-DPS proposed a diffusion method using a similar framework. In this approach, after each parallel reverse diffusion step from the image and kernel models, the estimates are refined using single gradient descent on $||\\mathbf{y} - \\mathbf{x} * \\mathbf{k}||^2$ . Specifically, after each time step $t$ of parallel reverse diffusion on kernel and image, i.e. $\\mathbf{k}_t$ and $\\mathbf{x}_t$ , temporary estimates of the kernel and image are obtained as follows:", + "bbox": [ + 214, + 513, + 782, + 602 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {0} \\leftarrow \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left[ \\mathbf {x} _ {t} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\hat {s} _ {\\mathbf {x}} (\\mathbf {x} _ {t}, t) \\right] \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 611, + 785, + 642 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {k}} _ {0} \\leftarrow \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left[ \\mathbf {k} _ {t} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\hat {s} _ {\\mathbf {k}} (\\mathbf {k} _ {t}, t) \\right] \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 645, + 785, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{s}_{\\mathbf{x}}, \\hat{s}_{\\mathbf{k}}$ represent the corresponding score functions. This is followed by a gradient descent step on the reblurring loss for each update step", + "bbox": [ + 214, + 684, + 782, + 715 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {t} ^ {\\prime} \\leftarrow \\mathbf {x} _ {t} - \\delta \\nabla_ {\\mathbf {x} _ {t}} \\| \\mathbf {y} - \\hat {\\mathbf {k}} _ {0} * \\hat {\\mathbf {x}} _ {0} \\| ^ {2} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 724, + 785, + 742 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {k}} _ {t} ^ {\\prime} \\leftarrow \\mathbf {k} _ {t} - \\delta \\nabla_ {\\mathbf {k} _ {t}} \\left[ \\| \\mathbf {y} - \\hat {\\mathbf {k}} _ {0} * \\hat {\\mathbf {x}} _ {0} \\| ^ {2} + \\lambda R _ {\\mathbf {k}} (\\hat {\\mathbf {k}} _ {0}) \\right] \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 744, + 785, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$R_{\\mathbf{k}}(.)$ represents the $\\ell_0 / \\ell_1$ regularization on the kernel to impose sparsity and $\\delta$ represents the step size change.", + "bbox": [ + 214, + 779, + 782, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While this alternating minimization approach to blind deconvolution is prevalent, this framework was first put into question by Levin et. al [40]. They showed", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Kernel Diffusion", + "bbox": [ + 620, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "that the conventional blind deconvolution techniques in this framework tended to converge to a no-blur degenerate solution, because the number of variables to be estimated i.e. $\\{\\mathbf{x},\\mathbf{k}\\}$ exceeded the number of knowns i.e. $\\mathbf{y}$ . Instead they recommended marginalization of the image space and first estimating the kernel using the conditional dsiitribution", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {k}} = \\underset {\\mathbf {k}} {\\arg \\max } \\left[ p (\\mathbf {k} | \\mathbf {y}) \\right] = \\underset {\\mathbf {k}} {\\arg \\max } \\left[ \\int p (\\mathbf {k}, \\mathbf {x} | \\mathbf {y}) d \\mathbf {x} \\right] \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 233, + 785, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "followed by estimation of image as follows:", + "bbox": [ + 215, + 272, + 524, + 287 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} = \\underset {\\mathbf {x}} {\\arg \\max } \\left[ p (\\mathbf {x} | \\mathbf {y}, \\hat {\\mathbf {k}}) \\right] \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 297, + 785, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since the number of unknowns is fewer than the number of knowns in the first sub-problem in (17), this leads to a better kernel estimate and hence, a better image estimate. Applying this principle to deconvolution with diffusion models, we aim to propose a method which samples from the conditional distribution $p(\\mathbf{k}|\\mathbf{y})$ instead of $p(\\mathbf{k},\\mathbf{x}|\\mathbf{y})$ since the former provides more stable solutions.", + "bbox": [ + 212, + 334, + 784, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Marginalization of Image Space", + "text_level": 1, + "bbox": [ + 215, + 431, + 522, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In practice, the estimation of $p(\\mathbf{k}|\\mathbf{y})$ can be difficult because the marginalization in (5) requires a weighted average over all images. We surmise that due to lack of a feasible method to evaluate the likelihood, the approach from [40] was not explored by the imaging community for the blind deconvolution. To mitigate this issue, we first discuss estimation of the conditional distribution $p(\\mathbf{k}|\\mathbf{y})$ using a non-blind solver.", + "bbox": [ + 212, + 455, + 787, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, we explicitly define a non-blind solver. Given a blurred image $\\mathbf{y}$ and the corresponding blur kernel $\\mathbf{k}$ , we can easily estimate the latent image using a non-blind solver which performs the following image estimation:", + "bbox": [ + 212, + 547, + 787, + 592 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} (\\mathbf {k}) = \\arg \\max _ {\\mathbf {x}} \\left[ p (\\mathbf {x} | \\mathbf {k}, \\mathbf {y}) \\right] \\stackrel {\\mathrm {d e f}} {=} F (\\mathbf {y}, \\mathbf {k}). \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 602, + 785, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Traditionally, these non-blind solvers are implemented using an iterative method where the corresponding cost function for MAP estimation in (19) is minimized using gradient descent or proximal updates. Convolutional neural networks with Wiener-deconvolution processing have been used with great success also [12, 23, 24, 51, 59].", + "bbox": [ + 212, + 636, + 787, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Remark 1: Given a non-blind solver $F_{\\phi}(.)$ , we can approximate the conditional probability $p(\\mathbf{k}|\\mathbf{y})$ as follows:", + "bbox": [ + 212, + 712, + 787, + 743 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np (\\mathbf {k} | \\mathbf {y}) = \\int p (\\mathbf {x}, \\mathbf {k} | \\mathbf {y}) d \\mathbf {x} \\approx \\frac {1}{Z (\\mathbf {y})} p (\\mathbf {y} \\mid F _ {\\phi} (\\mathbf {y}, \\mathbf {k}), \\mathbf {k}) \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 753, + 785, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $Z(\\mathbf{y})$ is independent of the kernel $\\mathbf{k}$ . Through the approximation, we can avoid the integration over the image space and express the likelihood $p(\\mathbf{k}|\\mathbf{y})$ in a tractable manner, up to the unknown the normalization constant $Z(\\mathbf{y})$ .", + "bbox": [ + 215, + 794, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Sanghvi et al.", + "bbox": [ + 271, + 114, + 385, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ac695fc85bf3d2f5a1e5fb782905aae5781bcf81bf8f0b0046504fa8a32e6de2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 142, + 767, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/58303bbcaccb426ca74ae88de8380d40f78f2535622e4a36af6b5dbf48020e20.jpg", + "image_caption": [ + "Fig. 2: Proposed Kernel-Diff. Each iteration of the proposed kernel-based diffusion contains two sub-steps. First, we perform a reverse diffusion step using the trained diffusion model $\\epsilon_{\\theta}(\\mathbf{k}_t,\\mathbf{y},t)$ . Next we plug the temporary kernel estimate $\\hat{\\mathbf{k}}_0$ in the differentiable non-blind solver $F(\\cdot)$ and perform a gradient descent update to guide the diffusion process to minimizing the reburring loss $\\| \\mathbf{y} - \\hat{\\mathbf{k}}_0\\circledast \\hat{\\mathbf{x}}_0\\|^2$ ." + ], + "image_footnote": [], + "bbox": [ + 228, + 213, + 774, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Kernel-Based Diffusion", + "text_level": 1, + "bbox": [ + 215, + 489, + 452, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To sample from $p(\\mathbf{k}|\\mathbf{y})$ , we propose to use a kernel-based diffusion network. This is different from existing methods using diffusion models, e.g., Blind-DPS [15], that resemble the alternating minimization approach, aiming to sample $\\{\\mathbf{x},\\mathbf{k}\\}$ from the conditional distribution $p(\\mathbf{x},\\mathbf{k}|\\mathbf{y})$ using implicit score models for image and kernel spaces.", + "bbox": [ + 212, + 513, + 787, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Diffusion Network. We use a modified version of the U-Net architecture as opposed to the one described in [22]. Since we are sampling from a conditional distribution, the diffusion model $\\epsilon_{\\theta}(\\mathbf{k},\\mathbf{y},t)$ takes as input both the current kernel estimate $\\mathbf{k}_t$ and the blurred image $\\mathbf{y}$ . This is achieved in a similar fashion as [58]. Specifically, our U-Net architecture has two encoders, one for the image $\\mathbf{y}$ and the other for the kernel $\\mathbf{k}$ . In our implementation, we use images of size $256\\times 256$ and kernels of size $64\\times 64$ . At the end of the encoders, the encoded features of the kernel branch are upscaled by a factor of 4 and concatenated to the encoded features of the image branch. The concatenated features are then passed through a common U-Net decoder stage to predict the kernel at the next time step.", + "bbox": [ + 212, + 589, + 787, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The diffusion model $\\epsilon_{\\theta}(\\cdot)$ is trained by minimizing the following loss function:", + "bbox": [ + 238, + 739, + 785, + 755 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {k e r n e l}} (\\theta) \\stackrel {\\mathrm {d e f}} {=} \\mathbb {E} _ {t, \\mathbf {k}, \\mathbf {y}, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {k} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\mathbf {y}, t\\right) \\| _ {2} ^ {2} \\right] \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 302, + 775, + 785, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the expectation is taken over all timesteps $t$ , blurred images $\\mathbf{y}$ and corresponding blur kernels $\\mathbf{k}$ . With trained parameters $\\theta^{*}$ , we sample $p(\\mathbf{k}|\\mathbf{y})$ using", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Kernel Diffusion", + "bbox": [ + 620, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "a series of reverse diffusion steps starting from $\\mathbf{k}_T\\sim \\mathcal{N}(0,\\mathbf{I})$ and the following reverse-diffusion iteration:", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {k} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {k} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta^ {*}} \\left(\\mathbf {k} _ {t}, \\mathbf {y}, t\\right)\\right) + \\bar {\\sigma} _ {t} \\mathbf {z} _ {t} \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 185, + 785, + 218 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Guidance using Non-Blind Solver. Once we have sampled the kernel $\\mathbf{k}_0$ , we estimate the latent image as $\\mathbf{x}_0 = F_{\\phi}(\\mathbf{y},\\mathbf{k}_0)$ , where $F_{\\phi}(\\cdot)$ is a neural network non-blind solver. There is a vast selection of candidate non-blind solver, including many network-based methods proposed recently [24-27, 59, 88]. These networks usually convert the blurred images to features space, apply Wiener deconvolution using the kernel followed by converting the deconvolved features into a single image using a decoder.", + "bbox": [ + 212, + 226, + 782, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While diffusion processes are effective in generating samples from the conditional distribution $p(\\mathbf{k}|\\mathbf{y})$ , they do not guarantee minimization of the reblurring loss, i.e., $\\| \\mathbf{y} - \\mathbf{k} * \\mathbf{x} \\|^2$ . To minimize reblurring loss while also generating a plausible sample $p(\\mathbf{k}|\\mathbf{y})$ , we guide the reverse diffusion process using the pre-trained non-blind solver as follows.", + "bbox": [ + 212, + 333, + 784, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "At step $t$ of the reverse diffusion process, the kernel estimate is given as $\\mathbf{k}_t$ . Using the trained diffusion model $\\epsilon_{\\theta^*}(\\mathbf{k}_t,\\mathbf{y},t)$ and $\\mathbf{k}_t$ , we can compute a temporary estimate of the sampled kernel $\\mathbf{k}_0$ as follows:", + "bbox": [ + 212, + 409, + 784, + 454 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {k}} _ {0} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\mathbf {k} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta^ {*}} \\left(\\mathbf {k} _ {t}, \\mathbf {y}, t\\right)\\right). \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 464, + 784, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We then estimate the image using this kernel estimate $\\hat{\\mathbf{k}}_0$ by $\\hat{\\mathbf{x}}_0 \\stackrel{\\mathrm{def}}{=} F(\\mathbf{y}, \\hat{\\mathbf{k}}_0)$ , plug into the reburring loss, and perform gradient descent by", + "bbox": [ + 212, + 507, + 782, + 541 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {k} _ {t} \\leftarrow \\mathbf {k} _ {t} - \\delta \\nabla_ {\\mathbf {k} _ {t}} \\| \\mathbf {y} - \\hat {\\mathbf {k}} _ {0} * F (\\mathbf {y}, \\hat {\\mathbf {k}} _ {0}) \\| ^ {2}. \\tag {24}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 565, + 784, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This step is equivalent to maximizing log likelihood of the approximation provided in (20). Taking the log of approximation, we get,", + "bbox": [ + 212, + 594, + 782, + 625 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\mathbf {k}} \\log p (\\mathbf {k} | \\mathbf {y}) \\approx \\nabla_ {\\mathbf {k}} \\left[ \\| \\mathbf {y} - F (\\mathbf {y}, \\mathbf {k}) * * \\mathbf {k} \\| _ {2} ^ {2} \\right] \\tag {25}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 635, + 784, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall procedure. We summarize the overall procedure in Algorithm 1 and pictorially represent it in Figure 2. At any given iteration $t$ , we first perform reverse diffusion process using the denoising network $\\epsilon_{\\theta}(\\cdot)$ to output $\\mathbf{k}_{t-1/2}$ (Step 5). Next, we perform a gradient descent update on $\\mathbf{k}_{t-1/2}$ the reburrowing loss $\\| \\mathbf{y} - \\hat{\\mathbf{k}}_0 * \\hat{\\mathbf{x}}_0 \\|^2$ (Step 7-9). Similar to Blind-DPS [15], we set the step size $\\delta$ according to following rule", + "bbox": [ + 212, + 681, + 784, + 775 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\delta = \\frac {\\rho}{\\left\\| \\mathbf {y} - \\hat {\\mathbf {x}} _ {0} * \\hat {\\mathbf {k}} _ {0} \\right\\|} \\tag {26}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 785, + 784, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\rho = 0.05$ using trial and error.", + "bbox": [ + 212, + 824, + 480, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Sanghvi et al.", + "bbox": [ + 271, + 114, + 385, + 128 + ], + "page_idx": 7 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Kernel-Diff" + ], + "code_body": "1: Input: Blurred Image y, Diffusion UNet $\\epsilon_{\\theta}(.)$ , Non-Blind Solver $F_{\\phi}(.)$ \n2: // Stage I: Reverse Diffusion \n3: $\\mathbf{k}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ \n4: for $t = T,\\dots ,1$ do \n5: $\\mathbf{z}_t\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ \n6: $\\mathbf{k}_{t - 1 / 2} = \\frac{1}{\\sqrt{\\alpha_t}}\\Bigg(\\mathbf{k}_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\bar{\\alpha}_t}}\\epsilon_\\theta (\\mathbf{k}_t,\\mathbf{y},t)\\Bigg) + \\sigma_t\\mathbf{z}_t$ \n7: $\\hat{\\mathbf{k}}_0 = \\frac{1}{\\sqrt{\\bar{\\alpha}_t}}\\big(\\mathbf{k}_t - \\sqrt{1 - \\bar{\\alpha}_t}\\epsilon_\\theta (\\mathbf{k}_t,\\mathbf{y},t)\\big)$ \n8: $\\hat{\\mathbf{x}}_0 = F_{\\phi}(\\mathbf{y},\\hat{\\mathbf{k}}_0)$ \n9: $\\mathbf{k}_{t - 1} = \\mathbf{k}_{t - 1 / 2} - \\delta \\nabla_{\\mathbf{k}_t}(\\| \\mathbf{y} - \\hat{\\mathbf{k}}_0*\\hat{\\mathbf{x}}_0\\| _2^2)$ \n10: end for \n11: // Stage II: Refinement \n12: for $j = 1,\\dots ,J$ do \n13: $\\mathbf{k}_0\\gets \\mathbf{k}_0 - \\delta \\nabla_{\\mathbf{k}_0}(\\| \\mathbf{y} - \\mathbf{k}_0*\\boldsymbol {F}_\\phi (\\mathbf{y},\\mathbf{k}_0)\\| _2^2 +\\gamma \\| \\mathbf{k}\\| _1)$ \n14: end for \n15: return $\\mathbf{k}_0$ and ${\\bf x}_0 = F_{\\phi}({\\bf y},{\\bf k}_0)$", + "bbox": [ + 218, + 162, + 781, + 417 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For further performance boost, once the reverse diffusion process is over, we repeat the gradient descent steps (Step 13) to refine the kernel, similar to Stage II in [61]. Here the overall reburring loss is minimized along with $\\ell_1$ penalty on the kernel with sparisty weight $\\gamma = 10^{-4}$ . Since this stage is meant to only fine-tune the kernel, we keep the number of total iterations in second stage, i.e. $J$ , to only 20.", + "bbox": [ + 212, + 450, + 787, + 541 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 566, + 375, + 584 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experiment Setup", + "text_level": 1, + "bbox": [ + 215, + 599, + 413, + 616 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation of Kernel-Diff. The diffusion network we use is a modified U-Net [22]. To train the diffusion model $\\epsilon_{\\theta}(\\mathbf{k}_t,\\mathbf{y},t)$ , we generate blurred data using synthetic motion blur kernels, with details of data synthesis to be elaborated in the next section. We assume symmetric boundary conditions during the convolution process. We train the diffusion U-Net architecture for 500,000 iterations with a batch size of 16. The non-blind solver in our scheme, i.e. $F_{\\phi}(.)$ , can be any differentiable neural network based method. In this paper, we plug Deep Wiener Deconvolution (DWDN) [24] to guide the reverse diffusion process. Training Dataset. The diffusion U-Net in Kernel-Diff as well as all methods we compare in our experiments are trained using the same set of blurred and clean image data. We synthesize the blurred images by applying random blur kernels from a kernel database on clean images, similar to [15]. The kernel database consists of $60,000\\times 64$ blur kernels generated using an open-sourced motion blur model [4] and blur kernel generation method provided by [3]. We note that", + "bbox": [ + 212, + 627, + 787, + 843 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Kernel Diffusion", + "bbox": [ + 620, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "the former method from [4] produces larger blur kernels than usually observed in practice. However, we include them in our analysis for sake of fair comparison to Blind-DPS which uses method to train diffusion model for blur kernel generation. For clean images, we use the full-resolution images from Flickr2K [46] and LSIDR [43] since diffusion models can be require large amounts of training data. To train the diffusion U-Net for Kernel-Diff, we additionally randomly crop the images to $256 \\times 256$ patches and augment the data by random horizontal flipping.", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/c9b9f947d61e70ab7cf3be10d1e0a91ead7565446fd4fde350f0d7d33b7ea0d3.jpg", + "table_caption": [ + "Table 1: Performance on BSD100 dataset with synthetic blur. Bold and underline refer to overall and second best blind deconvolution method respectively. Maximum normalized correlation (MNC) is provided to evaluate estimated kernels, when explicitly provided in case of Self-Deblur, Blind-DPS etc. DWDN is a non-blind method which knows the ground-truth kernel (i.e. $\\mathrm{MNC} = 1.00$ ) and is provided as an empirical upper bound only." + ], + "table_footnote": [], + "table_body": "
Method → Metric ↓DeblurGAN [39]DMPHN [85]MPR-Net [84]Self-Deblur [54]Fast-Deblur [13]Blind-DPS [15]Kernel-Diff (Ours)DWDN [24]
PSNR ↑15.4418.7318.4913.5314.0217.5619.0722.95
SSIM ↑0.3960.4270.4350.2270.2960.3870.5000.810
LPIPS ↓0.5630.6170.6320.6650.5830.6700.3550.169
FID ↓287.44265.05265.60259.40304.16280.53172.3379.11
MNC ↑---0.2060.5200.2830.6631.0
", + "bbox": [ + 217, + 376, + 815, + 468 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/7b13bc40de187e83352233b113079be04eee3a36fd10ca4a0be4e6c6dfac754c.jpg", + "table_caption": [ + "Table 2: Performance on RealBlur-50. Bold and underline refer to overall and second best blind deconvolution method respectively. The corresponding ground truth kernel is not provided for the RealBlur [55] dataset since the blur kernel may not be spatially invariant. As a result the metric MNC along with the performance on DWDN is not provided." + ], + "table_footnote": [], + "table_body": "
Method → Metric ↓DeblurGAN [39]Fast-Deblur [13]Self-Deblur [54]Blind-DPS [85]LMG [11]DMPHN [85]MPR-Net [84]Kernel-Diff (Ours)
PSNR ↑14.9818.5416.1519.8224.2024.3723.6325.57
SSIM ↑0.5830.4980.4830.6070.7890.7190.7150.798
LPIPS ↓0.2890.3920.4590.3200.1830.2010.2130.171
FID ↓129.06236.21172.32175.95106.35135.33113.1198.89
", + "bbox": [ + 217, + 608, + 805, + 683 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Quantitative Comparison", + "text_level": 1, + "bbox": [ + 215, + 737, + 470, + 752 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we evaluate the proposed method Kernel-Diff on synthetic and real blurred images and compare the performance with other state-of-the-art deconvolution methods. We evaluate our method on the BSD100 and RealBlur datasets as described below along with the following deconvolution/deblurring methods - Deep Hierarchical Multi-Patch Network (DHMPN) [85], Self-Deblur", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Sanghvi et al.", + "bbox": [ + 271, + 114, + 383, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c36b0ffc10d152c4048ee83f800e638cac306bb2e2d672b3b0ea71a8151b4754.jpg", + "image_caption": [ + "Fig. 3: Qualitative Results: Reconstruction results on synthetically blurred images from the BSD100 dataset. For Self-Deblur, Blind-DPS and Kernel-Diff, we show the estimated kernels in inset. $\\dagger$ DWDN is a non-blind deconvolution method assuming knowledge of the ground-truth kernel; its results are provided for reference only." + ], + "image_footnote": [], + "bbox": [ + 218, + 176, + 805, + 734 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Kernel Diffusion", + "bbox": [ + 620, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[54], Blind-DPS [15], MPR-Net [84], and DeblurGANv2 [39]. We use these methods as a representative set of contemporary deconvolution methods which include both iterative and end-to-end trained solutions. For the sake of a fair comparison, we retrain the end-to-end methods using synthetically blurred datasets with details provided in Section B of the supplementary document.", + "bbox": [ + 212, + 145, + 782, + 219 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "BSD100 Dataset. We first evaluate our method on 100 synthetically blurred images from the BSD dataset [47]. Specifically, we crop out random patches of size $256 \\times 256$ and then blur them with synthetic motion blur kernels using symmetric boundary conditions. The performance of Kernel-Diff and other deblurring methods is shown in Table 1 along with some qualitative examples in Figure 3. We also compare estimated kernels from Self-Deblur, Blind-DPS and Kernel-Diff in Table 1 using maximum of normalized convolution (MNC) [30].", + "bbox": [ + 212, + 220, + 782, + 325 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "RealBlur-50. To evaluate our blind deconvolution method on real blurred images, we use the RealBlur-J dataset [55]. However, since the focus of this paper is only spatially invariant blur and not other forms of blur, we crop random patches of size $256 \\times 256$ from 50 blurred images in the dataset to reduce the possibility of violating the spatially invarying blur assumption. The results of this comparison are shown in Table 2. We note that we for improved performance on the RealBlur dataset, we use a diffusion model trained only using the realistic blur kernels generated from [3] and increase the sparsity penalty $\\gamma$ to $5 \\times 10^{-4}$", + "bbox": [ + 212, + 328, + 784, + 460 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 483, + 388, + 498 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effect of Non-Blind Solver Guidance Each iteration of the diffusion consists of two sub-steps: (1) reverse diffusion using the trained denoiser $\\epsilon_{\\theta}(\\cdot)$ and (2) gradient descent on $\\| \\mathbf{y} - \\hat{\\mathbf{k}}_0 * \\hat{\\mathbf{x}}_0 \\|^2$ using the non-blind solver. In Table 3, we show the effect on performance of the second sub-step. Through the giant gap in performance with and without the guidance, we find that reverse diffusion is insufficient to converge to a good kernel estimate and hence, the experiment justifies the extra complexity required to guide the diffusion process to minimize the reblurring loss. In the same ablation study, we show how using the correct step size is important for the performance and how the second stage of finetuning impacts the overall scheme's performance.", + "bbox": [ + 212, + 506, + 782, + 654 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Computational Comparison The method proposed in this paper, Kernel-Diff is a kernel estimation analogue of the Blind-DPS method as both of them aim to solve the blind deconvolution problem using diffusion models. Therefore, we compare the computational costs of both the techniques in Table 4. Both techniques use a diffusion model trained for $T = 1000$ steps and our method performs much better than Blind-DPS which uses the alternating minimization approach using fewer parameters and similar time. Future work may involve reducing the inference time using DDIM [65] or consistency models [66].", + "bbox": [ + 214, + 657, + 782, + 777 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Reblurring Loss during Diffusion Levin et. al [40] show that alternate minimization approach is more likely to end up in a false local minima. We empirically demonstrate this by comparing the residual across the iterations, i.e., $\\| \\mathbf{y} - \\hat{\\mathbf{k}}_0\\otimes \\hat{\\mathbf{x}}_0\\|$ with Blind-DPS [15] in Figure 5. From the figure we observe that", + "bbox": [ + 214, + 779, + 782, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Sanghvi et al.", + "bbox": [ + 271, + 114, + 383, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "the residual in our proposed kernel estimation scheme converges to a solution with a much lower residual value. Assuming the kernel and image likelihoods to be approximately equal for both solutions, we conclude that our method leads to solutions with a higher log-likelihood, and hence, better quality results.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9e0e5e2c59910da4a00c34a75d59c57f57f538231c6444d91e86ef919665036c.jpg", + "image_caption": [ + "w/o non-blind solver guidance" + ], + "image_footnote": [], + "bbox": [ + 241, + 232, + 411, + 366 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/64e324936f037901ee3719ea1fd566b2fdd0973739bbae463f6936071e10f7e0.jpg", + "image_caption": [ + "with non-blind solver guidance" + ], + "image_footnote": [], + "bbox": [ + 415, + 232, + 586, + 366 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2ae7b294a0177dd6ff332cc044075edc2597a665db99e24846b3a50e0fc23f93.jpg", + "image_caption": [ + "DWDN + True kernel", + "Fig. 4: Effect of Non-Blind Solver Guidance. Through this ablation study, we examine the effect of the second sub-step in the diffusion process which guides the diffusion process to minimize the reburring loss using the non-blind solver DWDN [24]." + ], + "image_footnote": [], + "bbox": [ + 589, + 233, + 761, + 366 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/203c76c10d2b5bde55a1c838deb99db41b1d21244eefa2d399f29cf44ff7723d.jpg", + "table_caption": [ + "Table 3: Ablation study comparing effect of different hyperparameters of the scheme on the performance. We observe that removing the guidance of the non-blind solver reduces the performance across all metrics. While using gradient descent using non-blind solver in (24), for optimal performance, the step size shouldn't be set too large. Lastly, the effect of finetuning kernel in Stage-II can also provide a significant performance boost for PSNR, however, not so much for other metrics." + ], + "table_footnote": [], + "table_body": "
Ablation Study ↓ Metric →PSNR ↑SSIM ↑LPIPS ↓FID ↓MNC ↑
ρ = 0, No DWDN21.900.6500.251152.510.714
ρ = 0.0522.640.6800.234140.280.724
ρ = 0.121.970.6510.256156.040.567
ρ = 0.520.870.6090.300173.360.270
J = 0, No Stage II21.760.6710.241143.580.675
DWDN, known kernel30.540.9060.08749.291.00
", + "bbox": [ + 225, + 607, + 772, + 724 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 777, + 359, + 792 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this paper, we proposed Kernel-Diff, a blind deconvolution solution in the tradition of Levin et. al [40]. We show how the kernel minimization approach can be", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Kernel Diffusion", + "bbox": [ + 620, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8b081afa8d5e605ab5d803911d7f0664d0f6b9ef1b3c528ab73332ef5f677531.jpg", + "image_caption": [ + "Fig. 5: Residual $\\| \\mathbf{y} - \\hat{\\mathbf{k}}_0\\otimes \\hat{\\mathbf{x}}_0\\|^2$ for Blind-DPS and Kernel-Diff. This experiments confirms our hypothesis that the proposed kernel estimation scheme converges to a better local minimum. The shaded region shows the range of the residual over 50 trials of each scheme." + ], + "image_footnote": [], + "bbox": [ + 331, + 133, + 669, + 330 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/1b8f239832f27305c2dc7e0f3bf99ea415329a497291f5facdb1dc27406ecee8.jpg", + "table_caption": [ + "Table 4: Comparison of computational cost with Blind-DPS in terms of total number of parameters and time taken to deblur a $256 \\times 256$ image respectively. Both Blind-DPS and Kernel-Diff use a diffusion model approach to the blind deconvolution problem, our kernel estimation approach uses a lightweight diffusion model and non-blind solver for the reverse diffusion process. Using significantly fewer parameters and smaller computational time, we achieve significant performance boost over the alternating minimization approach." + ], + "table_footnote": [], + "table_body": "
Metric → Method ↓Total Parameter CountTotal Time
Blind-DPS110M337s
Kernel-Diff + DWDN43M305s
", + "bbox": [ + 310, + 526, + 691, + 599 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "implemented in practice using a pre-trained non-blind solver in diffusion models setting and propose a guided diffusion method that samples from $p(\\mathbf{k}|\\mathbf{y})$ to solve the blind deconvolution problem and show extensive numerical experiments on blurred images.", + "bbox": [ + 212, + 627, + 785, + 686 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "For future work, the method can be refined to deblur images with spatially varying and motion blur i.e. degradation which do not fit the single-blur convolutional model. Alternately, the non-blind solver can be modified to be robust to inaccuracies in kernel estimation.", + "bbox": [ + 212, + 688, + 785, + 748 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 771, + 403, + 789 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work is supported, in part, by the DARPA / SRC CogniSense JUMP 2.0 Center, NSF IIS-2133032, and NSF ECCS-2030570.", + "bbox": [ + 212, + 803, + 785, + 833 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Sanghvi et al.", + "bbox": [ + 271, + 114, + 383, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahmad, R., Bouman, C.A., Buzzard, G.T., Chan, S., Liu, S., Reehorst, E.T., Schniter, P.: Plug-and-play methods for magnetic resonance imaging: Using de-noisers for image recovery. IEEE signal processing magazine 37(1), 105-116 (2020)", + "2. Anger, J., Facciolo, G., Delbracio, M.: Blind image deblurring using the 10 gradient prior. Image processing on line 9, 124-142 (2019)", + "3. Boracchi, G., Foi, A.: Modeling the performance of image restoration from motion blur. Image Processing, IEEE Transactions on 21(8), 3502-3517 (aug 2012). https://doi.org/10.1109/TIP.2012.2192126", + "4. Borodenko, L.: Motion blur (May 2020), https://github.com/LeviBorodenko/motionblur", + "5. Carbajal, G., Vitoria, P., Delbracio, M., Muse, P., Lezama, J.: Non-uniform blur kernel estimation via adaptive basis decomposition. arXiv preprint arXiv:2102.01026 (2021)", + "6. Carbajal, G., Vitoria, P., Lezama, J., Muse, P.: Blind motion deblurring with pixel-wise kernel estimation via kernel prediction networks. IEEE Transactions on Computational Imaging (2023)", + "7. Chakrabarti, A.: A neural approach to blind motion deblurring. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14. pp. 221-235. Springer (2016)", + "8. Chan, S.H., Wang, X., Elgendy, O.A.: Plug-and-play admm for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging 3(1), 84-98 (2016)", + "9. Chan, T.F., Wong, C.K.: Total variation blind deconvolution. IEEE transactions on Image Processing 7(3), 370-375 (1998)", + "0. Chen, L., Fang, F., Wang, T., Zhang, G.: Blind image deblurring with local maximum gradient prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1742-1750 (2019)", + "1. Chen, L., Fang, F., Wang, T., Zhang, G.: Blind image deblurring with local maximum gradient prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1742-1750 (2019)", + "2. Chen, L., Zhang, J., Li, Z., Wei, Y., Fang, F., Ren, J., Pan, J.: Deep richardson-lucy deconvolution for low-light image deblurring. International Journal of Computer Vision pp. 1-18 (2023)", + "3. Cho, S., Lee, S.: Fast motion deblurring. ACM Transactions on Graphics pp. 1-8 (2009)", + "4. Choi, J., Kim, S., Jeong, Y., Gwon, Y., Yoon, S.: Ilvr: Conditioning method for denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14367-14376 (October 2021)", + "5. Chung, H., Kim, J., Kim, S., Ye, J.C.: Parallel diffusion models of operator and image for blind inverse problems. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6059-6069 (2023)", + "6. Chung, H., Kim, J., Mccann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=OnD9zGAGT0k", + "7. Chung, H., Ryu, D., McCann, M.T., Klasky, M.L., Ye, J.C.: Solving 3d inverse problems using pre-trained 2d diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22542-22551 (2023)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Kernel Diffusion", + "bbox": [ + 620, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Chung, H., Sim, B., Ryu, D., Ye, J.C.: Improving diffusion models for inverse problems using manifold constraints. Advances in Neural Information Processing Systems 35, 25683-25696 (2022)", + "19. Chung, H., Sim, B., Ye, J.C.: Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12413-12422 (2022)", + "20. Cohen, R., Elad, M., Milanfar, P.: Regularization by denoising via fixed-point projection (red-pro). SIAM Journal on Imaging Sciences 14(3), 1374-1406 (2021)", + "21. Daras, G., Delbracio, M., Talebi, H., Dimakis, A., Milanfar, P.: Soft diffusion: Score matching with general corruptions. Transactions on Machine Learning Research (2023)", + "22. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)", + "23. Dong, J., Pan, J., Sun, D., Su, Z., Yang, M.H.: Learning data terms for non-blind deblurring. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 748-763 (2018)", + "24. Dong, J., Roth, S., Schiele, B.: Deep wiener deconvolution: Wiener meets deep learning for image deblurring. Advances in Neural Information Processing Systems 33, 1048-1059 (2020)", + "25. Dong, J., Roth, S., Schiele, B.: Learning spatially-variant map models for non-blind image deblurring. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4886-4895 (2021)", + "26. Eboli, T., Sun, J., Ponce, J.: End-to-end interpretable learning of non-blind image deblurring. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVII 16. pp. 314-331. Springer (2020)", + "27. Gong, D., Zhang, Z., Shi, Q., van den Hengel, A., Shen, C., Zhang, Y.: Learning deep gradient descent optimization for image deconvolution. IEEE transactions on neural networks and learning systems 31(12), 5468-5482 (2020)", + "28. Gupta, A., Joshi, N., Lawrence Zitnick, C., Cohen, M., Curless, B.: Single image deblurring using motion density functions. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part I 11. pp. 171-184. Springer (2010)", + "29. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)", + "30. Hu, Z., Yang, M.H.: Good regions to deblur. In: Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12. pp. 59-72. Springer (2012)", + "31. Jalal, A., Arvinte, M., Daras, G., Price, E., Dimakis, A.G., Tamir, J.: Robust compressed sensing mri with deep generative priors. Advances in Neural Information Processing Systems 34, 14938-14954 (2021)", + "32. Jiang, H., Luo, A., Han, S., Fan, H., Liu, S.: Low-light image enhancement with wavelet-based diffusion models. arXiv preprint arXiv:2306.00306 (2023)", + "33. Kadkhodaie, Z., Simoncelli, E.: Stochastic solutions for linear inverse problems using the prior implicit in a denoiser. Advances in Neural Information Processing Systems 34, 13242-13254 (2021)", + "34. Kamilov, U.S., Mansour, H., Wohlberg, B.: A plug-and-play priors approach for solving nonlinear imaging inverse problems. IEEE Signal Processing Letters 24(12), 1872-1876 (2017)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Sanghvi et al.", + "bbox": [ + 271, + 114, + 383, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. Advances in Neural Information Processing Systems 35, 23593-23606 (2022)", + "36. Kawar, B., Vaksman, G., Elad, M.: Snips: Solving noisy inverse problems stochastically. Advances in Neural Information Processing Systems 34, 21757-21769 (2021)", + "37. Kawar, B., Vaksman, G., Elad, M.: Stochastic image denoising by sampling from the posterior distribution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1866-1875 (2021)", + "38. Kupyn, O., Budzan, V., Mykhailych, M., Mishkin, D., Matas, J.: Deblurgan: Blind motion deblurring using conditional adversarial networks. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8183-8192 (2018). https://doi.org/10.1109/CVPR.2018.00854", + "39. Kupyn, O., Martyniuk, T., Wu, J., Wang, Z.: Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In: The IEEE International Conference on Computer Vision (ICCV) (Oct 2019)", + "40. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding blind deconvolution algorithms. IEEE Transactions on Pattern Analysis and Machine Intelligence 33(12), 2354-2367 (2011)", + "41. Li, H., Yang, Y., Chang, M., Chen, S., Feng, H., Xu, Z., Li, Q., Chen, Y.: Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing 479, 47-59 (2022)", + "42. Li, X., Ren, Y., Jin, X., Lan, C., Wang, X., Zeng, W., Wang, X., Chen, Z.: Diffusion models for image restoration and enhancement-a comprehensive survey. arXiv preprint arXiv:2308.09388 (2023)", + "43. Li, Y., Zhang, K., Liang, J., Cao, J., Liu, C., Gong, R., Zhang, Y., Tang, H., Liu, Y., Demandolx, D., et al.: LSDIR: A large scale dataset for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1775-1787 (2023)", + "44. Li, Y., Tofighi, M., Geng, J., Monga, V., Eldar, Y.C.: Efficient and interpretable deep blind image deblurring via algorithm unrolling. IEEE Transactions on Computational Imaging 6, 666-681 (2020)", + "45. Li, Y., Tofighi, M., Monga, V., Eldar, Y.C.: An algorithm unrolling approach to deep image deblurring. In: ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 7675-7679. IEEE (2019)", + "46. Lim, B., Son, S., Kim, H., Nah, S., Lee, K.M.: Enhanced deep residual networks for single image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (July 2017)", + "47. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: Proc. 8th Int'l Conf. Computer Vision. vol. 2, pp. 416-423 (July 2001)", + "48. Monga, V., Li, Y., Eldar, Y.C.: Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine 38(2), 18-44 (2021)", + "49. Murata, N., Saito, K., Lai, C.H., Takida, Y., Uesaka, T., Mitsufuji, Y., Ermon, S.: Gibbsddrm: A partially collapsed gibbs sampler for solving blind inverse problems with denoising diffusion restoration. arXiv preprint arXiv:2301.12686 (2023)", + "50. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017)" + ], + "bbox": [ + 212, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Kernel Diffusion", + "bbox": [ + 620, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "51. Nan, Y., Ji, H.: Deep learning for handling kernel/model uncertainty in image deconvolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2388-2397 (2020)", + "52. Pan, J., Sun, D., Pfister, H., Yang, M.H.: Blind image deblurring using dark channel prior. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1628-1636 (2016)", + "53. Reehorst, E.T., Schniter, P.: Regularization by denoising: Clarifications and new interpretations. IEEE transactions on computational imaging 5(1), 52-67 (2018)", + "54. Ren, D., Zhang, K., Wang, Q., Hu, Q., Zuo, W.: Neural blind deconvolution using deep priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3341-3350 (2020)", + "55. Rim, J., Lee, H., Won, J., Cho, S.: Real-world blur dataset for learning and benchmarking deblurring algorithms. In: Proceedings of the European Conference on Computer Vision (ECCV) (2020)", + "56. Romano, Y., Elad, M., Milanfar, P.: The little engine that could: Regularization by denoising (red). SIAM Journal on Imaging Sciences 10(4), 1804-1844 (2017)", + "57. Rond, A., Giryes, R., Elad, M.: Poisson inverse problems by the plug-and-play scheme. Journal of Visual Communication and Image Representation 41, 96-108 (2016)", + "58. Saharia, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. arXiv:2104.07636 (2021)", + "59. Sanghvi, Y., Gnanasambandam, A., Chan, S.H.: Photon limited non-blind deblurring using algorithm unrolling. IEEE Transactions on Computational Imaging 8, 851-864 (2022)", + "60. Sanghvi, Y., Gnanasambandam, A., Mao, Z., Chan, S.H.: Photon-limited blind deconvolution using unsupervised iterative kernel estimation. IEEE Transactions on Computational Imaging 8, 1051-1062 (2022)", + "61. Sanghvi, Y., Mao, Z., Chan, S.H.: Structured kernel estimation for photon-limited deconvolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9863-9872 (2023)", + "62. Schuler, C.J., Hirsch, M., Harmeling, S., Scholkopf, B.: Learning to deblur. IEEE transactions on pattern analysis and machine intelligence 38(7), 1439-1451 (2015)", + "63. Shan, Q., Jia, J., Agarwala, A.: High-quality motion deblurring from a single image. ACM Transactions on Graphics 27(3), 1-10 (2008)", + "64. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)", + "65. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv:2010.02502 (October 2020), https://arxiv.org/abs/2010.02502", + "66. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models (2023)", + "67. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=PxTIG12RRHS", + "68. Sun, J., Cao, W., Xu, Z., Ponce, J.: Learning a convolutional neural network for non-uniform motion blur removal. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 769-777 (2015)", + "69. Sun, L., Cho, S., Wang, J., Hays, J.: Edge-based blur kernel estimation using patch priors. In: IEEE international conference on computational photography (ICCP). pp. 1-8. IEEE (2013)" + ], + "bbox": [ + 215, + 147, + 784, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Sanghvi et al.", + "bbox": [ + 271, + 114, + 383, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "70. Sun, Y., Wohlberg, B., Kamilov, U.S.: An online plug-and-play algorithm for regularized image reconstruction. IEEE Transactions on Computational Imaging 5(3), 395-408 (2019)", + "71. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-recurrent network for deep image deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8174-8182 (2018)", + "72. Ulyanov, D., Vedaldi, A., Lempitsky, V.: Deep image prior. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9446-9454 (2018)", + "73. Venkatakrishnan, S.V., Bouman, C.A., Wohlberg, B.: Plug-and-play priors for model based reconstruction. In: 2013 IEEE global conference on signal and information processing. pp. 945-948. IEEE (2013)", + "74. Vincent, P.: A connection between score matching and denoising autoencoders. Neural computation 23(7), 1661-1674 (2011)", + "75. Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. The Eleventh International Conference on Learning Representations (2023)", + "76. Wang, Y., Yu, Y., Yang, W., Guo, L., Chau, L.P., Kot, A.C., Wen, B.: Exposurediffusion: Learning to expose for low-light image enhancement. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12438-12448 (2023)", + "77. Wang, Z., Cun, X., Bao, J., Liu, J.: Uformer: A general u-shaped transformer for image restoration. in 2022 IEEE. In: CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17662-17672 (2021)", + "78. Whang, J., Delbracio, M., Talebi, H., Sahara, C., Dimakis, A.G., Milanfar, P.: Deblurring via stochastic refinement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16293-16303 (2022)", + "79. Xu, L., Jia, J.: Two-phase kernel estimation for robust motion deblurring. In: European Conference on Computer Vision. pp. 157-170. Springer (2010)", + "80. Xu, L., Zheng, S., Jia, J.: Unnatural l0 sparse representation for natural image deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1107-1114 (2013)", + "81. Yan, Y., Ren, W., Guo, Y., Wang, R., Cao, X.: Image deblurring via extreme channels prior. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4003-4011 (2017)", + "82. yang, y., Sun, J., Li, H., Xu, Z.: Deep admm-net for compressive sensing mri. In: Lee, D., Sugiyama, M., Luxburg, U., Guyon, I., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 29. Curran Associates, Inc. (2016), https://proceedings.neurips.cc/paper_files/paper/2016/file/1679091c5a880faf6fb5e6087eb1b2dc-Paper.pdf", + "83. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022)", + "84. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14821-14831 (2021)", + "85. Zhang, H., Dai, Y., Li, H., Koniusz, P.: Deep stacked hierarchical multi-patch network for image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5978-5986 (2019)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Kernel Diffusion", + "bbox": [ + 620, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "86. Zhang, K., Zuo, W., Gu, S., Zhang, L.: Learning deep cnn denoiser prior for image restoration. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3929-3938 (2017)", + "87. Zhang, K., Ren, W., Luo, W., Lai, W.S., Stenger, B., Yang, M.H., Li, H.: Deep image deblurring: A survey. International Journal of Computer Vision 130(9), 2103-2130 (2022)", + "88. Zhang, Z., Cheng, Y., Suo, J., Bian, L., Dai, Q.: Infwide: Image and feature space wiener deconvolution network for non-blind image deblurring in low-light conditions. IEEE Transactions on Image Processing 32, 1390-1402 (2023)", + "89. Zhu, Y., Zhang, K., Liang, J., Cao, J., Wen, B., Timofte, R., Van Gool, L.: Denoising diffusion models for plug-and-play image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1219-1229 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 327 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "Y. Sanghvi et al.", + "bbox": [ + 271, + 114, + 383, + 128 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/Kernel Diffusion_ An Alternate Approach to Blind Deconvolution/0f568881-c0ff-47c0-b18b-592b7cbf5a93_model.json b/2024/Kernel Diffusion_ An Alternate Approach to Blind Deconvolution/0f568881-c0ff-47c0-b18b-592b7cbf5a93_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d70e081a97b436c456c3d518b393d34e85daf7de --- /dev/null +++ b/2024/Kernel Diffusion_ An Alternate Approach to Blind Deconvolution/0f568881-c0ff-47c0-b18b-592b7cbf5a93_model.json @@ -0,0 +1,3067 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.25, + 0.141, + 0.756, + 0.187 + ], + "angle": 0, + "content": "Kernel Diffusion: An Alternate Approach to Blind Deconvolution" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.212, + 0.707, + 0.229 + ], + "angle": 0, + "content": "Yash Sanghvi1, Yiheng Chi1, and Stanley H. Chan1" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.24, + 0.776, + 0.283 + ], + "angle": 0, + "content": "Elmore Family School of Electrical and Computer Engineering, Purdue University, West Lafayette, Indiana, USA {ysanghvi,chi14,stanchan}@purdue.edu" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.32, + 0.744, + 0.473 + ], + "angle": 0, + "content": "Abstract. Blind deconvolution problems are severely ill-posed because neither the underlying signal nor the forward operator are not known exactly. Conventionally, these problems are solved by alternating between estimation of the image and kernel while keeping the other fixed. In this paper, we show that this framework is flawed because of its tendency to get trapped in local minima and, instead, suggest the use of a kernel estimation strategy with a non-blind solver. This framework is employed by a diffusion method which is trained to sample the blur kernel from the conditional distribution with guidance from a pre-trained non-blind solver. The proposed diffusion method leads to state-of-the-art results on both synthetic and real blur datasets." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.486, + 0.702, + 0.501 + ], + "angle": 0, + "content": "Keywords: Blind Deconvolution \\(\\cdot\\) Diffusion Models \\(\\cdot\\) Deblurring" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.527, + 0.377, + 0.543 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.558, + 0.788, + 0.604 + ], + "angle": 0, + "content": "Images taken by a hand-held camera under long exposure suffer from motion blur. Assuming this blur is spatially invariant, the captured image \\(\\mathbf{y}\\) and the underlying sharp image \\(\\mathbf{x}\\) are related as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.447, + 0.617, + 0.786, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} = \\mathbf {k} * \\mathbf {x} + \\mathbf {n}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.703 + ], + "angle": 0, + "content": "where \\(\\mathbf{k}\\), \\(\\mathbf{n}\\) refer to the blur kernel corresponding to the camera motion and noise, respectively, and the operator \\(\\ast\\) is convolution. The inverse problem of estimating \\(\\mathbf{x}\\) given \\(\\mathbf{y}\\) is called deconvolution, which can be classified into non-blind and blind." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.81 + ], + "angle": 0, + "content": "In the non-blind case, the blur kernel \\(\\mathbf{k}\\) is known. There are plenty of iterative methods which have been successfully applied to various image restoration tasks. Usually, these approaches, such as Plug-and-Play [1,8,34,57,70,73] and Regularization by Denoising (RED) [20,53,56], use a pre-trained image denoiser as a substitute for a complex image prior. Recent deep learning counterparts to these approaches involve unrolling the iterative methods followed by end-to-end training [45,48,82,86]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In blind deconvolution, the blur kernel \\(\\mathbf{k}\\) is not known. A natural strategy for the blind inverse problem is to jointly estimate the blur kernel \\(\\mathbf{k}\\), and the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.386, + 0.129 + ], + "angle": 0, + "content": "Y. Sanghvi et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.574, + 0.163 + ], + "angle": 0, + "content": "image \\(\\mathbf{x}\\) by maximizing the posterior probability:" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.173, + 0.786, + 0.233 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\mathbf {x}}, \\hat {\\mathbf {k}} = \\underset {\\mathbf {x}, \\mathbf {k}} {\\arg \\max} \\left[ p (\\mathbf {x}, \\mathbf {k} | \\mathbf {y}) \\right] \\\\ = \\underset {\\mathbf {x}, \\mathbf {k}} {\\arg \\max } \\left[ \\log p (\\mathbf {y} | \\mathbf {k}, \\mathbf {x}) + \\log p (\\mathbf {x}) + \\log p (\\mathbf {k}) \\right]. \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.245, + 0.786, + 0.29 + ], + "angle": 0, + "content": "The Maximum-A-Posteriori (MAP) for the pair \\((\\mathbf{x},\\mathbf{k})\\), i.e. \\(\\mathrm{MAP}_{\\mathbf{x},\\mathbf{k}}\\), can be obtained by an iterative scheme which alternates between the following minimization steps:" + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.301, + 0.786, + 0.327 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} ^ {i + 1} = \\arg \\min _ {\\mathbf {x}} \\left[ \\| \\mathbf {y} - \\mathbf {x} * \\mathbf {k} ^ {i} \\| ^ {2} + \\phi (\\mathbf {x}) \\right], \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.329, + 0.786, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\mathbf {k} ^ {i + 1} = \\arg \\operatorname * {m i n} _ {\\mathbf {k}} \\left[ \\| \\mathbf {y} - \\mathbf {x} ^ {i + 1} * \\mathbf {\\Phi} \\mathbf {k} \\| ^ {2} + \\psi (\\mathbf {k}) \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.365, + 0.784, + 0.439 + ], + "angle": 0, + "content": "where \\(\\phi(\\cdot), \\psi(\\cdot)\\) represent the image and kernel priors. Many classical deconvolution techniques can be framed as variations of this framework [2,9,63,80]. Even neural network based iterative methods such as Self-Deblur [54] and diffusion models [15,49] have achieved state-of-the-art performance in blind deconvolution using the stated framework." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.441, + 0.788, + 0.548 + ], + "angle": 0, + "content": "Despite recent success of diffusion methods, they are based on \\(\\mathrm{MAP}_{\\mathbf{x},\\mathbf{k}}\\) approach, which was first brought into question by Levin et al. [40] They argued that the number of variables to be estimated, i.e., \\(\\{\\mathbf{x},\\mathbf{k}\\}\\), are much larger than the number of known variables, i.e., \\(\\mathbf{y}\\), and as a result, (2) favors the no-blur solution \\(\\hat{\\mathbf{x}},\\hat{\\mathbf{k}} = (\\mathbf{y},\\mathbf{I})\\). [40] proposed an alternate strategy of solving the kernel \\(\\mathbf{k}\\) first to exploit the inherent symmetry in the problem. This is achieved by marginalizing out the image space as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.559, + 0.786, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {k}} = \\underset {\\mathbf {k}} {\\arg \\max } \\left[ p (\\mathbf {k} | \\mathbf {y}) \\right] = \\underset {\\mathbf {k}} {\\arg \\max } \\left[ \\int p (\\mathbf {k}, \\mathbf {x} | \\mathbf {y}) d \\mathbf {x} \\right]. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.598, + 0.788, + 0.659 + ], + "angle": 0, + "content": "Although \\(\\mathrm{MAP_k}\\) optimization can lead to significantly improved reconstruction (as shown in Figure 1(c)) it has not been adopted in blind deconvolution solutions, which we surmise because of the intrinsic computational difficulty of the integration in (5)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.68, + 0.378, + 0.695 + ], + "angle": 0, + "content": "1.1 Contributions" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we address the problem with a kernel estimation approach, and we demonstrate that this strategy can be implemented in practice. Our contribution in this paper can be summarized as follows: We propose a novel diffusion method, called Kernel-Diffusion (Kernel-Diff) which solves the \\(\\mathrm{MAP_k}\\) optimization problem by sampling from the conditional distribution \\(p(\\mathbf{k}|\\mathbf{y})\\). This diffusion method coupled with a differentiable non-blind solver achieves state-of-the-art deconvolution performance. We show with numerical experiments on real and synthetic data that the proposed approach creates significant performance improvement. We will soon release the source code and model weights for Kernel-Diff." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.621, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Kernel Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.143, + 0.359, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.255, + 0.348, + 0.283 + ], + "angle": 0, + "content": "(a) Blurred image Inset: true kernel" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.143, + 0.501, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.255, + 0.499, + 0.298 + ], + "angle": 0, + "content": "(b) Blind-DPS [15] based on alternating minimization" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.144, + 0.643, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.255, + 0.634, + 0.298 + ], + "angle": 0, + "content": "(c) Kernel-Diff (ours) based on kernel estimation" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.144, + 0.784, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.255, + 0.753, + 0.283 + ], + "angle": 0, + "content": "(d) Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.312, + 0.785, + 0.381 + ], + "angle": 0, + "content": "Fig. 1: In this paper, we present a novel diffusion-based approach to the blind deconvolution problem. The conventional approaches, such as Blind-DPS, iterate between between kernel and image space minimization. Our approach uses a kernel-first estimation strategy in the context of diffusion models to sample kernels from the conditional distribution \\( p(\\mathbf{k}|\\mathbf{y}) \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.411, + 0.368, + 0.429 + ], + "angle": 0, + "content": "2 Background" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.446, + 0.415, + 0.46 + ], + "angle": 0, + "content": "2.1 Diffusion Methods" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.473, + 0.785, + 0.534 + ], + "angle": 0, + "content": "Diffusion models [29, 64, 67, 74] are a popular generative method which can be used to sample any target diffusion \\( p_{data}(\\mathbf{x}) \\). The diffusion models begin with a forward process, which starts from \\( \\mathbf{x}_0 \\sim p_{data}(\\mathbf{x}) \\) and gradually adds noise according to variance schedule \\( \\{\\beta_1, \\beta_2, \\ldots, \\beta_T\\} \\), formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.404, + 0.547, + 0.785, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t} \\sim \\mathcal {N} \\left(\\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.578, + 0.785, + 0.624 + ], + "angle": 0, + "content": "at time steps \\( t \\in \\{1, 2, \\dots, T\\} \\). This forward process converges to \\( \\mathbf{x}_T \\sim \\mathcal{N}(0, \\mathbf{I}) \\). Next, the reverse diffusion process is used to sample from the distribution of \\( p_{data} \\) by starting with \\( \\mathbf{x}_T \\sim \\mathcal{N}(0, \\mathbf{I}) \\) and performing the iterative updates" + }, + { + "type": "equation", + "bbox": [ + 0.336, + 0.635, + 0.785, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {x} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {x} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta^ {*}} \\left(\\mathbf {x} _ {t}, t\\right)\\right) + \\sigma_ {t} \\mathbf {z} _ {t}, \\tag {7} \\\\ \\mathbf {z} _ {t} \\sim \\mathcal {N} (0, \\mathbf {I}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.703, + 0.784, + 0.761 + ], + "angle": 0, + "content": "where \\(\\alpha_{t} \\stackrel{\\mathrm{def}}{=} 1 - \\beta_{t}\\), \\(\\bar{\\alpha}_{t} \\stackrel{\\mathrm{def}}{=} \\prod_{i=1}^{t} \\alpha_{i}\\), and \\(\\epsilon_{\\theta^{*}}(\\mathbf{x}_{t}, t)\\) is the trained diffusion network used to learn to denoise \\(\\mathbf{x}_{t}\\) at any given time step, trained by minimizing the following loss function." + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.773, + 0.785, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\theta) \\stackrel {\\text {d e f}} {=} \\mathbb {E} _ {t, \\mathbf {x} _ {0}, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t\\right) \\| ^ {2} \\right]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Inverse Problems and Diffusion Models. Score based diffusion methods have been extensively used in solving inverse problems [17, 21, 33, 35-37, 89] by" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.386, + 0.129 + ], + "angle": 0, + "content": "Y. Sanghvi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "sampling from the posterior \\( p(\\mathbf{x}|\\mathbf{y}) \\) and the corresponding score is calculated using the Bayes' formula" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.187, + 0.786, + 0.203 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t} | \\mathbf {y}) = \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t}) + \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {y} | \\mathbf {x} _ {t}). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.21, + 0.784, + 0.315 + ], + "angle": 0, + "content": "For example, [31] uses an approximation to the posterior score in (9) and used the Langevin dynamics to sample from the distribution \\( p(\\mathbf{x}|\\mathbf{y}) \\) for the MRI compressed sensing problem; [18,19] solve the inpainting and super-resolution inverse problems with a modified reverse diffusion scheme which includes a projection step to reduce the consistency metric \\( \\| \\mathbf{y} - \\mathbf{A}\\mathbf{x}\\| \\) in each iteration; diffusion Posterior Sampling (DPS) [16] solves the noisy inverse problem with alternating diffusion and gradient descent steps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.316, + 0.784, + 0.39 + ], + "angle": 0, + "content": "Image Restoration and Diffusion Models Diffusion models sample from a target distribution while MMSE estimators average over them [37]. As a result, the former has been widely used in image restoration tasks such as deblurring [78], super-resolution [41, 58], and low-light enhancement [32, 76]. For an exhaustive review, we refer the reader to [42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.785, + 0.511 + ], + "angle": 0, + "content": "Guidance in Diffusion Models. Pre-trained diffusion models are often guided at each iteration to produce desired results. This is first proposed in [22] where a classifier gradient is added to each iteration for the final output to belong to the given class. ILVR [14] guides the image diffusion process by constraining the low-frequency information of an image. In the same spirit, [75] guides the diffusion process by refining only the null-space of the corresponding inverse problem. [78] guides a deblurring diffusion model using a pretrained structure module." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.531, + 0.431, + 0.545 + ], + "angle": 0, + "content": "2.2 Blind Deconvolution" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.784, + 0.674 + ], + "angle": 0, + "content": "Traditional blind deconvolution methods can be categorized under the \\(\\mathrm{MAP}_{\\mathbf{x},\\mathbf{k}}\\) framework [2,9,10,44,52,63,69,80,81]. For example, [13] uses salient edges for the kernel estimation and then performs image estimation using a Total Variation (TV) prior. [79] refines the kernel estimation process by iterative support detection. [28] performs alternating image and kernel minimization but uses a motion density representation for the blur kernel instead of for pixels. Self-Deblur [54] uses an alternating minimization scheme where a latent deep image prior [72] is used to represent the kernel and image space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.784, + 0.749 + ], + "angle": 0, + "content": "Deep learning methods have replaced the optimization-based approaches and usually involve end-to-end training of image restoration networks [38, 62, 68, 71, 77, 83-85] using large datasets of pairs of sharp and blurry images, e.g., GoPro [50] and RealBlur [55] datasets. For a more exhaustive review of deep learning based approaches, we refer the readers to [87]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.784, + 0.84 + ], + "angle": 0, + "content": "A few methods have utilized the \\(\\mathrm{MAP_k}\\) approach, even though not always explicitly stated by the authors. [7] uses a fully convolutional network to predict the Fourier coefficients of the underlying kernel which are used to deconvolve the blurry image in Fourier space. [5,6] train a Kernel-Prediction Network for all pixels in the image to deal with spatially varying blur. [60] leverages the differentiability of a non-blind solver [59] to frame the blind deconvolution problem" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.621, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Kernel Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "entirely in terms of kernel estimation problems. This strategy is improved by [61] using a key-points based representation to further reduce the dimensionality of the problem." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.214, + 0.331, + 0.23 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.245, + 0.784, + 0.305 + ], + "angle": 0, + "content": "Blind-DPS [15] proposed a diffusion model based scheme for solving the blind inverse problem. Using separate diffusion models to sample from the image and kernel prior, a new scheme was proposed to sample solutions from the joint conditional distribution" + }, + { + "type": "equation", + "bbox": [ + 0.447, + 0.317, + 0.786, + 0.334 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}}, \\hat {\\mathbf {k}} \\sim p (\\mathbf {x}, \\mathbf {k} | \\mathbf {y}) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.343, + 0.784, + 0.388 + ], + "angle": 0, + "content": "This method is similar in principle to the conventional blind deconvolution techniques where the goal is to estimate the maximum of the conditional distribution i.e." + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.396, + 0.786, + 0.458 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\mathbf {x}}, \\hat {\\mathbf {k}} = \\underset {\\mathbf {x}, \\mathbf {k}} {\\arg \\max } \\left[ p (\\mathbf {x}, \\mathbf {k} | \\mathbf {y}) \\right] (11) \\\\ = \\underset {\\mathbf {x}, \\mathbf {k}} {\\arg \\min } \\left[ \\left\\| \\mathbf {y} - \\mathbf {k} * \\mathbf {x} \\right\\| ^ {2} + R _ {\\mathbf {x}} (\\mathbf {x}) + R _ {\\mathbf {k}} (\\mathbf {k}) \\right] (12) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.468, + 0.784, + 0.514 + ], + "angle": 0, + "content": "where \\( R_{\\mathbf{x}}(.) \\) and \\( R_{\\mathbf{k}}(.) \\) represent the image and kernel priors respectively. This minimization is achieved using alternating minimization techniques where the kernel is estimated while keeping the image estimate fixed and vice versa." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.514, + 0.784, + 0.603 + ], + "angle": 0, + "content": "Recently, Blind-DPS proposed a diffusion method using a similar framework. In this approach, after each parallel reverse diffusion step from the image and kernel models, the estimates are refined using single gradient descent on \\(||\\mathbf{y} - \\mathbf{x} * \\mathbf{k}||^2\\). Specifically, after each time step \\(t\\) of parallel reverse diffusion on kernel and image, i.e. \\(\\mathbf{k}_t\\) and \\(\\mathbf{x}_t\\), temporary estimates of the kernel and image are obtained as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.612, + 0.786, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {0} \\leftarrow \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left[ \\mathbf {x} _ {t} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\hat {s} _ {\\mathbf {x}} (\\mathbf {x} _ {t}, t) \\right] \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.646, + 0.786, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {k}} _ {0} \\leftarrow \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left[ \\mathbf {k} _ {t} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\hat {s} _ {\\mathbf {k}} (\\mathbf {k} _ {t}, t) \\right] \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.685, + 0.784, + 0.716 + ], + "angle": 0, + "content": "where \\(\\hat{s}_{\\mathbf{x}}, \\hat{s}_{\\mathbf{k}}\\) represent the corresponding score functions. This is followed by a gradient descent step on the reblurring loss for each update step" + }, + { + "type": "equation", + "bbox": [ + 0.441, + 0.725, + 0.786, + 0.743 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {t} ^ {\\prime} \\leftarrow \\mathbf {x} _ {t} - \\delta \\nabla_ {\\mathbf {x} _ {t}} \\| \\mathbf {y} - \\hat {\\mathbf {k}} _ {0} * \\hat {\\mathbf {x}} _ {0} \\| ^ {2} \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.746, + 0.786, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {k}} _ {t} ^ {\\prime} \\leftarrow \\mathbf {k} _ {t} - \\delta \\nabla_ {\\mathbf {k} _ {t}} \\left[ \\| \\mathbf {y} - \\hat {\\mathbf {k}} _ {0} * \\hat {\\mathbf {x}} _ {0} \\| ^ {2} + \\lambda R _ {\\mathbf {k}} (\\hat {\\mathbf {k}} _ {0}) \\right] \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.784, + 0.81 + ], + "angle": 0, + "content": "\\(R_{\\mathbf{k}}(.)\\) represents the \\(\\ell_0 / \\ell_1\\) regularization on the kernel to impose sparsity and \\(\\delta\\) represents the step size change." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "While this alternating minimization approach to blind deconvolution is prevalent, this framework was first put into question by Levin et. al [40]. They showed" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.386, + 0.129 + ], + "angle": 0, + "content": "Y. Sanghvi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "that the conventional blind deconvolution techniques in this framework tended to converge to a no-blur degenerate solution, because the number of variables to be estimated i.e. \\(\\{\\mathbf{x},\\mathbf{k}\\}\\) exceeded the number of knowns i.e. \\(\\mathbf{y}\\). Instead they recommended marginalization of the image space and first estimating the kernel using the conditional dsiitribution" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.234, + 0.786, + 0.264 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {k}} = \\underset {\\mathbf {k}} {\\arg \\max } \\left[ p (\\mathbf {k} | \\mathbf {y}) \\right] = \\underset {\\mathbf {k}} {\\arg \\max } \\left[ \\int p (\\mathbf {k}, \\mathbf {x} | \\mathbf {y}) d \\mathbf {x} \\right] \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.273, + 0.525, + 0.289 + ], + "angle": 0, + "content": "followed by estimation of image as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.298, + 0.786, + 0.326 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} = \\underset {\\mathbf {x}} {\\arg \\max } \\left[ p (\\mathbf {x} | \\mathbf {y}, \\hat {\\mathbf {k}}) \\right] \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.785, + 0.412 + ], + "angle": 0, + "content": "Since the number of unknowns is fewer than the number of knowns in the first sub-problem in (17), this leads to a better kernel estimate and hence, a better image estimate. Applying this principle to deconvolution with diffusion models, we aim to propose a method which samples from the conditional distribution \\( p(\\mathbf{k}|\\mathbf{y}) \\) instead of \\( p(\\mathbf{k},\\mathbf{x}|\\mathbf{y}) \\) since the former provides more stable solutions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.433, + 0.523, + 0.449 + ], + "angle": 0, + "content": "3.1 Marginalization of Image Space" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.546 + ], + "angle": 0, + "content": "In practice, the estimation of \\( p(\\mathbf{k}|\\mathbf{y}) \\) can be difficult because the marginalization in (5) requires a weighted average over all images. We surmise that due to lack of a feasible method to evaluate the likelihood, the approach from [40] was not explored by the imaging community for the blind deconvolution. To mitigate this issue, we first discuss estimation of the conditional distribution \\( p(\\mathbf{k}|\\mathbf{y}) \\) using a non-blind solver." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.593 + ], + "angle": 0, + "content": "First, we explicitly define a non-blind solver. Given a blurred image \\(\\mathbf{y}\\) and the corresponding blur kernel \\(\\mathbf{k}\\), we can easily estimate the latent image using a non-blind solver which performs the following image estimation:" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.603, + 0.786, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} (\\mathbf {k}) = \\arg \\max _ {\\mathbf {x}} \\left[ p (\\mathbf {x} | \\mathbf {k}, \\mathbf {y}) \\right] \\stackrel {\\mathrm {d e f}} {=} F (\\mathbf {y}, \\mathbf {k}). \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.712 + ], + "angle": 0, + "content": "Traditionally, these non-blind solvers are implemented using an iterative method where the corresponding cost function for MAP estimation in (19) is minimized using gradient descent or proximal updates. Convolutional neural networks with Wiener-deconvolution processing have been used with great success also [12, 23, 24, 51, 59]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.713, + 0.788, + 0.744 + ], + "angle": 0, + "content": "Remark 1: Given a non-blind solver \\( F_{\\phi}(.) \\), we can approximate the conditional probability \\( p(\\mathbf{k}|\\mathbf{y}) \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.332, + 0.755, + 0.786, + 0.786 + ], + "angle": 0, + "content": "\\[\np (\\mathbf {k} | \\mathbf {y}) = \\int p (\\mathbf {x}, \\mathbf {k} | \\mathbf {y}) d \\mathbf {x} \\approx \\frac {1}{Z (\\mathbf {y})} p (\\mathbf {y} \\mid F _ {\\phi} (\\mathbf {y}, \\mathbf {k}), \\mathbf {k}) \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\( Z(\\mathbf{y}) \\) is independent of the kernel \\( \\mathbf{k} \\). Through the approximation, we can avoid the integration over the image space and express the likelihood \\( p(\\mathbf{k}|\\mathbf{y}) \\) in a tractable manner, up to the unknown the normalization constant \\( Z(\\mathbf{y}) \\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.621, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Kernel Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.143, + 0.768, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.214, + 0.776, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.392, + 0.788, + 0.463 + ], + "angle": 0, + "content": "Fig. 2: Proposed Kernel-Diff. Each iteration of the proposed kernel-based diffusion contains two sub-steps. First, we perform a reverse diffusion step using the trained diffusion model \\(\\epsilon_{\\theta}(\\mathbf{k}_t,\\mathbf{y},t)\\). Next we plug the temporary kernel estimate \\(\\hat{\\mathbf{k}}_0\\) in the differentiable non-blind solver \\(F(\\cdot)\\) and perform a gradient descent update to guide the diffusion process to minimizing the reburring loss \\(\\| \\mathbf{y} - \\hat{\\mathbf{k}}_0\\circledast \\hat{\\mathbf{x}}_0\\|^2\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.491, + 0.454, + 0.505 + ], + "angle": 0, + "content": "3.2 Kernel-Based Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.589 + ], + "angle": 0, + "content": "To sample from \\( p(\\mathbf{k}|\\mathbf{y}) \\), we propose to use a kernel-based diffusion network. This is different from existing methods using diffusion models, e.g., Blind-DPS [15], that resemble the alternating minimization approach, aiming to sample \\( \\{\\mathbf{x},\\mathbf{k}\\} \\) from the conditional distribution \\( p(\\mathbf{x},\\mathbf{k}|\\mathbf{y}) \\) using implicit score models for image and kernel spaces." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.788, + 0.74 + ], + "angle": 0, + "content": "Diffusion Network. We use a modified version of the U-Net architecture as opposed to the one described in [22]. Since we are sampling from a conditional distribution, the diffusion model \\(\\epsilon_{\\theta}(\\mathbf{k},\\mathbf{y},t)\\) takes as input both the current kernel estimate \\(\\mathbf{k}_t\\) and the blurred image \\(\\mathbf{y}\\). This is achieved in a similar fashion as [58]. Specifically, our U-Net architecture has two encoders, one for the image \\(\\mathbf{y}\\) and the other for the kernel \\(\\mathbf{k}\\). In our implementation, we use images of size \\(256\\times 256\\) and kernels of size \\(64\\times 64\\). At the end of the encoders, the encoded features of the kernel branch are upscaled by a factor of 4 and concatenated to the encoded features of the image branch. The concatenated features are then passed through a common U-Net decoder stage to predict the kernel at the next time step." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.74, + 0.786, + 0.756 + ], + "angle": 0, + "content": "The diffusion model \\(\\epsilon_{\\theta}(\\cdot)\\) is trained by minimizing the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.303, + 0.776, + 0.787, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {k e r n e l}} (\\theta) \\stackrel {\\mathrm {d e f}} {=} \\mathbb {E} _ {t, \\mathbf {k}, \\mathbf {y}, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {k} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\mathbf {y}, t\\right) \\| _ {2} ^ {2} \\right] \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "where the expectation is taken over all timesteps \\(t\\), blurred images \\(\\mathbf{y}\\) and corresponding blur kernels \\(\\mathbf{k}\\). With trained parameters \\(\\theta^{*}\\), we sample \\(p(\\mathbf{k}|\\mathbf{y})\\) using" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.386, + 0.129 + ], + "angle": 0, + "content": "Y. Sanghvi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "a series of reverse diffusion steps starting from \\(\\mathbf{k}_T\\sim \\mathcal{N}(0,\\mathbf{I})\\) and the following reverse-diffusion iteration:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.186, + 0.786, + 0.219 + ], + "angle": 0, + "content": "\\[\n\\mathbf {k} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {k} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta^ {*}} \\left(\\mathbf {k} _ {t}, \\mathbf {y}, t\\right)\\right) + \\bar {\\sigma} _ {t} \\mathbf {z} _ {t} \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.227, + 0.784, + 0.333 + ], + "angle": 0, + "content": "Guidance using Non-Blind Solver. Once we have sampled the kernel \\(\\mathbf{k}_0\\), we estimate the latent image as \\(\\mathbf{x}_0 = F_{\\phi}(\\mathbf{y},\\mathbf{k}_0)\\), where \\(F_{\\phi}(\\cdot)\\) is a neural network non-blind solver. There is a vast selection of candidate non-blind solver, including many network-based methods proposed recently [24-27, 59, 88]. These networks usually convert the blurred images to features space, apply Wiener deconvolution using the kernel followed by converting the deconvolved features into a single image using a decoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.334, + 0.785, + 0.408 + ], + "angle": 0, + "content": "While diffusion processes are effective in generating samples from the conditional distribution \\( p(\\mathbf{k}|\\mathbf{y}) \\), they do not guarantee minimization of the reblurring loss, i.e., \\( \\| \\mathbf{y} - \\mathbf{k} * \\mathbf{x} \\|^2 \\). To minimize reblurring loss while also generating a plausible sample \\( p(\\mathbf{k}|\\mathbf{y}) \\), we guide the reverse diffusion process using the pre-trained non-blind solver as follows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.785, + 0.455 + ], + "angle": 0, + "content": "At step \\(t\\) of the reverse diffusion process, the kernel estimate is given as \\(\\mathbf{k}_t\\). Using the trained diffusion model \\(\\epsilon_{\\theta^*}(\\mathbf{k}_t,\\mathbf{y},t)\\) and \\(\\mathbf{k}_t\\), we can compute a temporary estimate of the sampled kernel \\(\\mathbf{k}_0\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.369, + 0.465, + 0.785, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {k}} _ {0} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\mathbf {k} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta^ {*}} \\left(\\mathbf {k} _ {t}, \\mathbf {y}, t\\right)\\right). \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.784, + 0.542 + ], + "angle": 0, + "content": "We then estimate the image using this kernel estimate \\(\\hat{\\mathbf{k}}_0\\) by \\(\\hat{\\mathbf{x}}_0 \\stackrel{\\mathrm{def}}{=} F(\\mathbf{y}, \\hat{\\mathbf{k}}_0)\\), plug into the reburring loss, and perform gradient descent by" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.566, + 0.785, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\mathbf {k} _ {t} \\leftarrow \\mathbf {k} _ {t} - \\delta \\nabla_ {\\mathbf {k} _ {t}} \\| \\mathbf {y} - \\hat {\\mathbf {k}} _ {0} * F (\\mathbf {y}, \\hat {\\mathbf {k}} _ {0}) \\| ^ {2}. \\tag {24}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.784, + 0.626 + ], + "angle": 0, + "content": "This step is equivalent to maximizing log likelihood of the approximation provided in (20). Taking the log of approximation, we get," + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.636, + 0.785, + 0.662 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\mathbf {k}} \\log p (\\mathbf {k} | \\mathbf {y}) \\approx \\nabla_ {\\mathbf {k}} \\left[ \\| \\mathbf {y} - F (\\mathbf {y}, \\mathbf {k}) * * \\mathbf {k} \\| _ {2} ^ {2} \\right] \\tag {25}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.682, + 0.785, + 0.776 + ], + "angle": 0, + "content": "Overall procedure. We summarize the overall procedure in Algorithm 1 and pictorially represent it in Figure 2. At any given iteration \\( t \\), we first perform reverse diffusion process using the denoising network \\( \\epsilon_{\\theta}(\\cdot) \\) to output \\( \\mathbf{k}_{t-1/2} \\) (Step 5). Next, we perform a gradient descent update on \\( \\mathbf{k}_{t-1/2} \\) the reburrowing loss \\( \\| \\mathbf{y} - \\hat{\\mathbf{k}}_0 * \\hat{\\mathbf{x}}_0 \\|^2 \\) (Step 7-9). Similar to Blind-DPS [15], we set the step size \\( \\delta \\) according to following rule" + }, + { + "type": "equation", + "bbox": [ + 0.435, + 0.786, + 0.785, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\delta = \\frac {\\rho}{\\left\\| \\mathbf {y} - \\hat {\\mathbf {x}} _ {0} * \\hat {\\mathbf {k}} _ {0} \\right\\|} \\tag {26}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.825, + 0.481, + 0.841 + ], + "angle": 0, + "content": "where \\(\\rho = 0.05\\) using trial and error." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.621, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Kernel Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.146, + 0.408, + 0.162 + ], + "angle": 0, + "content": "Algorithm 1 Kernel-Diff" + }, + { + "type": "algorithm", + "bbox": [ + 0.219, + 0.163, + 0.782, + 0.419 + ], + "angle": 0, + "content": "1: Input: Blurred Image y, Diffusion UNet \\(\\epsilon_{\\theta}(.)\\) , Non-Blind Solver \\(F_{\\phi}(.)\\) \n2: // Stage I: Reverse Diffusion \n3: \\(\\mathbf{k}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) \n4: for \\(t = T,\\dots ,1\\) do \n5: \\(\\mathbf{z}_t\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) \n6: \\(\\mathbf{k}_{t - 1 / 2} = \\frac{1}{\\sqrt{\\alpha_t}}\\Bigg(\\mathbf{k}_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\bar{\\alpha}_t}}\\epsilon_\\theta (\\mathbf{k}_t,\\mathbf{y},t)\\Bigg) + \\sigma_t\\mathbf{z}_t\\) \n7: \\(\\hat{\\mathbf{k}}_0 = \\frac{1}{\\sqrt{\\bar{\\alpha}_t}}\\big(\\mathbf{k}_t - \\sqrt{1 - \\bar{\\alpha}_t}\\epsilon_\\theta (\\mathbf{k}_t,\\mathbf{y},t)\\big)\\) \n8: \\(\\hat{\\mathbf{x}}_0 = F_{\\phi}(\\mathbf{y},\\hat{\\mathbf{k}}_0)\\) \n9: \\(\\mathbf{k}_{t - 1} = \\mathbf{k}_{t - 1 / 2} - \\delta \\nabla_{\\mathbf{k}_t}(\\| \\mathbf{y} - \\hat{\\mathbf{k}}_0*\\hat{\\mathbf{x}}_0\\| _2^2)\\) \n10: end for \n11: // Stage II: Refinement \n12: for \\(j = 1,\\dots ,J\\) do \n13: \\(\\mathbf{k}_0\\gets \\mathbf{k}_0 - \\delta \\nabla_{\\mathbf{k}_0}(\\| \\mathbf{y} - \\mathbf{k}_0*\\boldsymbol {F}_\\phi (\\mathbf{y},\\mathbf{k}_0)\\| _2^2 +\\gamma \\| \\mathbf{k}\\| _1)\\) \n14: end for \n15: return \\(\\mathbf{k}_0\\) and \\({\\bf x}_0 = F_{\\phi}({\\bf y},{\\bf k}_0)\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.451, + 0.788, + 0.542 + ], + "angle": 0, + "content": "For further performance boost, once the reverse diffusion process is over, we repeat the gradient descent steps (Step 13) to refine the kernel, similar to Stage II in [61]. Here the overall reburring loss is minimized along with \\(\\ell_1\\) penalty on the kernel with sparisty weight \\(\\gamma = 10^{-4}\\). Since this stage is meant to only fine-tune the kernel, we keep the number of total iterations in second stage, i.e. \\(J\\), to only 20." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.568, + 0.376, + 0.585 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.601, + 0.414, + 0.617 + ], + "angle": 0, + "content": "4.1 Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Implementation of Kernel-Diff. The diffusion network we use is a modified U-Net [22]. To train the diffusion model \\(\\epsilon_{\\theta}(\\mathbf{k}_t,\\mathbf{y},t)\\), we generate blurred data using synthetic motion blur kernels, with details of data synthesis to be elaborated in the next section. We assume symmetric boundary conditions during the convolution process. We train the diffusion U-Net architecture for 500,000 iterations with a batch size of 16. The non-blind solver in our scheme, i.e. \\(F_{\\phi}(.)\\), can be any differentiable neural network based method. In this paper, we plug Deep Wiener Deconvolution (DWDN) [24] to guide the reverse diffusion process. Training Dataset. The diffusion U-Net in Kernel-Diff as well as all methods we compare in our experiments are trained using the same set of blurred and clean image data. We synthesize the blurred images by applying random blur kernels from a kernel database on clean images, similar to [15]. The kernel database consists of \\(60,000\\times 64\\) blur kernels generated using an open-sourced motion blur model [4] and blur kernel generation method provided by [3]. We note that" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.129 + ], + "angle": 0, + "content": "Y. Sanghvi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "the former method from [4] produces larger blur kernels than usually observed in practice. However, we include them in our analysis for sake of fair comparison to Blind-DPS which uses method to train diffusion model for blur kernel generation. For clean images, we use the full-resolution images from Flickr2K [46] and LSIDR [43] since diffusion models can be require large amounts of training data. To train the diffusion U-Net for Kernel-Diff, we additionally randomly crop the images to \\(256 \\times 256\\) patches and augment the data by random horizontal flipping." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.28, + 0.788, + 0.364 + ], + "angle": 0, + "content": "Table 1: Performance on BSD100 dataset with synthetic blur. Bold and underline refer to overall and second best blind deconvolution method respectively. Maximum normalized correlation (MNC) is provided to evaluate estimated kernels, when explicitly provided in case of Self-Deblur, Blind-DPS etc. DWDN is a non-blind method which knows the ground-truth kernel (i.e. \\(\\mathrm{MNC} = 1.00\\)) and is provided as an empirical upper bound only." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.377, + 0.816, + 0.469 + ], + "angle": 0, + "content": "
Method → Metric ↓DeblurGAN [39]DMPHN [85]MPR-Net [84]Self-Deblur [54]Fast-Deblur [13]Blind-DPS [15]Kernel-Diff (Ours)DWDN [24]
PSNR ↑15.4418.7318.4913.5314.0217.5619.0722.95
SSIM ↑0.3960.4270.4350.2270.2960.3870.5000.810
LPIPS ↓0.5630.6170.6320.6650.5830.6700.3550.169
FID ↓287.44265.05265.60259.40304.16280.53172.3379.11
MNC ↑---0.2060.5200.2830.6631.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.526, + 0.788, + 0.596 + ], + "angle": 0, + "content": "Table 2: Performance on RealBlur-50. Bold and underline refer to overall and second best blind deconvolution method respectively. The corresponding ground truth kernel is not provided for the RealBlur [55] dataset since the blur kernel may not be spatially invariant. As a result the metric MNC along with the performance on DWDN is not provided." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.609, + 0.806, + 0.684 + ], + "angle": 0, + "content": "
Method → Metric ↓DeblurGAN [39]Fast-Deblur [13]Self-Deblur [54]Blind-DPS [85]LMG [11]DMPHN [85]MPR-Net [84]Kernel-Diff (Ours)
PSNR ↑14.9818.5416.1519.8224.2024.3723.6325.57
SSIM ↑0.5830.4980.4830.6070.7890.7190.7150.798
LPIPS ↓0.2890.3920.4590.3200.1830.2010.2130.171
FID ↓129.06236.21172.32175.95106.35135.33113.1198.89
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.738, + 0.471, + 0.753 + ], + "angle": 0, + "content": "4.2 Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this section, we evaluate the proposed method Kernel-Diff on synthetic and real blurred images and compare the performance with other state-of-the-art deconvolution methods. We evaluate our method on the BSD100 and RealBlur datasets as described below along with the following deconvolution/deblurring methods - Deep Hierarchical Multi-Patch Network (DHMPN) [85], Self-Deblur" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.622, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Kernel Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.178, + 0.806, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.747, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Fig. 3: Qualitative Results: Reconstruction results on synthetically blurred images from the BSD100 dataset. For Self-Deblur, Blind-DPS and Kernel-Diff, we show the estimated kernels in inset. \\(\\dagger\\)DWDN is a non-blind deconvolution method assuming knowledge of the ground-truth kernel; its results are provided for reference only." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.129 + ], + "angle": 0, + "content": "Y. Sanghvi et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.783, + 0.22 + ], + "angle": 0, + "content": "[54], Blind-DPS [15], MPR-Net [84], and DeblurGANv2 [39]. We use these methods as a representative set of contemporary deconvolution methods which include both iterative and end-to-end trained solutions. For the sake of a fair comparison, we retrain the end-to-end methods using synthetically blurred datasets with details provided in Section B of the supplementary document." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.784, + 0.327 + ], + "angle": 0, + "content": "BSD100 Dataset. We first evaluate our method on 100 synthetically blurred images from the BSD dataset [47]. Specifically, we crop out random patches of size \\(256 \\times 256\\) and then blur them with synthetic motion blur kernels using symmetric boundary conditions. The performance of Kernel-Diff and other deblurring methods is shown in Table 1 along with some qualitative examples in Figure 3. We also compare estimated kernels from Self-Deblur, Blind-DPS and Kernel-Diff in Table 1 using maximum of normalized convolution (MNC) [30]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.785, + 0.461 + ], + "angle": 0, + "content": "RealBlur-50. To evaluate our blind deconvolution method on real blurred images, we use the RealBlur-J dataset [55]. However, since the focus of this paper is only spatially invariant blur and not other forms of blur, we crop random patches of size \\(256 \\times 256\\) from 50 blurred images in the dataset to reduce the possibility of violating the spatially invarying blur assumption. The results of this comparison are shown in Table 2. We note that we for improved performance on the RealBlur dataset, we use a diffusion model trained only using the realistic blur kernels generated from [3] and increase the sparsity penalty \\(\\gamma\\) to \\(5 \\times 10^{-4}\\)" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.484, + 0.389, + 0.499 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.784, + 0.655 + ], + "angle": 0, + "content": "Effect of Non-Blind Solver Guidance Each iteration of the diffusion consists of two sub-steps: (1) reverse diffusion using the trained denoiser \\(\\epsilon_{\\theta}(\\cdot)\\) and (2) gradient descent on \\(\\| \\mathbf{y} - \\hat{\\mathbf{k}}_0 * \\hat{\\mathbf{x}}_0 \\|^2\\) using the non-blind solver. In Table 3, we show the effect on performance of the second sub-step. Through the giant gap in performance with and without the guidance, we find that reverse diffusion is insufficient to converge to a good kernel estimate and hence, the experiment justifies the extra complexity required to guide the diffusion process to minimize the reblurring loss. In the same ablation study, we show how using the correct step size is important for the performance and how the second stage of finetuning impacts the overall scheme's performance." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.659, + 0.784, + 0.778 + ], + "angle": 0, + "content": "Computational Comparison The method proposed in this paper, Kernel-Diff is a kernel estimation analogue of the Blind-DPS method as both of them aim to solve the blind deconvolution problem using diffusion models. Therefore, we compare the computational costs of both the techniques in Table 4. Both techniques use a diffusion model trained for \\( T = 1000 \\) steps and our method performs much better than Blind-DPS which uses the alternating minimization approach using fewer parameters and similar time. Future work may involve reducing the inference time using DDIM [65] or consistency models [66]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Reblurring Loss during Diffusion Levin et. al [40] show that alternate minimization approach is more likely to end up in a false local minima. We empirically demonstrate this by comparing the residual across the iterations, i.e., \\(\\| \\mathbf{y} - \\hat{\\mathbf{k}}_0\\otimes \\hat{\\mathbf{x}}_0\\|\\) with Blind-DPS [15] in Figure 5. From the figure we observe that" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.621, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Kernel Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "the residual in our proposed kernel estimation scheme converges to a solution with a much lower residual value. Assuming the kernel and image likelihoods to be approximately equal for both solutions, we conclude that our method leads to solutions with a higher log-likelihood, and hence, better quality results." + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.233, + 0.413, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.371, + 0.38, + 0.4 + ], + "angle": 0, + "content": "w/o non-blind solver guidance" + }, + { + "type": "image", + "bbox": [ + 0.416, + 0.233, + 0.588, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.449, + 0.371, + 0.554, + 0.4 + ], + "angle": 0, + "content": "with non-blind solver guidance" + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.234, + 0.763, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.371, + 0.724, + 0.398 + ], + "angle": 0, + "content": "DWDN + True kernel" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.412, + 0.789, + 0.456 + ], + "angle": 0, + "content": "Fig. 4: Effect of Non-Blind Solver Guidance. Through this ablation study, we examine the effect of the second sub-step in the diffusion process which guides the diffusion process to minimize the reburring loss using the non-blind solver DWDN [24]." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.511, + 0.788, + 0.596 + ], + "angle": 0, + "content": "Table 3: Ablation study comparing effect of different hyperparameters of the scheme on the performance. We observe that removing the guidance of the non-blind solver reduces the performance across all metrics. While using gradient descent using non-blind solver in (24), for optimal performance, the step size shouldn't be set too large. Lastly, the effect of finetuning kernel in Stage-II can also provide a significant performance boost for PSNR, however, not so much for other metrics." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.608, + 0.773, + 0.726 + ], + "angle": 0, + "content": "
Ablation Study ↓ Metric →PSNR ↑SSIM ↑LPIPS ↓FID ↓MNC ↑
ρ = 0, No DWDN21.900.6500.251152.510.714
ρ = 0.0522.640.6800.234140.280.724
ρ = 0.121.970.6510.256156.040.567
ρ = 0.520.870.6090.300173.360.270
J = 0, No Stage II21.760.6710.241143.580.675
DWDN, known kernel30.540.9060.08749.291.00
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.778, + 0.36, + 0.794 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In this paper, we proposed Kernel-Diff, a blind deconvolution solution in the tradition of Levin et. al [40]. We show how the kernel minimization approach can be" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.129 + ], + "angle": 0, + "content": "Y. Sanghvi et al." + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.135, + 0.67, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.403 + ], + "angle": 0, + "content": "Fig. 5: Residual \\(\\| \\mathbf{y} - \\hat{\\mathbf{k}}_0\\otimes \\hat{\\mathbf{x}}_0\\|^2\\) for Blind-DPS and Kernel-Diff. This experiments confirms our hypothesis that the proposed kernel estimation scheme converges to a better local minimum. The shaded region shows the range of the residual over 50 trials of each scheme." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.515 + ], + "angle": 0, + "content": "Table 4: Comparison of computational cost with Blind-DPS in terms of total number of parameters and time taken to deblur a \\(256 \\times 256\\) image respectively. Both Blind-DPS and Kernel-Diff use a diffusion model approach to the blind deconvolution problem, our kernel estimation approach uses a lightweight diffusion model and non-blind solver for the reverse diffusion process. Using significantly fewer parameters and smaller computational time, we achieve significant performance boost over the alternating minimization approach." + }, + { + "type": "table", + "bbox": [ + 0.312, + 0.527, + 0.692, + 0.6 + ], + "angle": 0, + "content": "
Metric → Method ↓Total Parameter CountTotal Time
Blind-DPS110M337s
Kernel-Diff + DWDN43M305s
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.787, + 0.688 + ], + "angle": 0, + "content": "implemented in practice using a pre-trained non-blind solver in diffusion models setting and propose a guided diffusion method that samples from \\( p(\\mathbf{k}|\\mathbf{y}) \\) to solve the blind deconvolution problem and show extensive numerical experiments on blurred images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.749 + ], + "angle": 0, + "content": "For future work, the method can be refined to deblur images with spatially varying and motion blur i.e. degradation which do not fit the single-blur convolutional model. Alternately, the non-blind solver can be modified to be robust to inaccuracies in kernel estimation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.772, + 0.405, + 0.79 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.804, + 0.787, + 0.834 + ], + "angle": 0, + "content": "This work is supported, in part, by the DARPA / SRC CogniSense JUMP 2.0 Center, NSF IIS-2133032, and NSF ECCS-2030570." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.622, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Kernel Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Ahmad, R., Bouman, C.A., Buzzard, G.T., Chan, S., Liu, S., Reehorst, E.T., Schniter, P.: Plug-and-play methods for magnetic resonance imaging: Using de-noisers for image recovery. IEEE signal processing magazine 37(1), 105-116 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.247 + ], + "angle": 0, + "content": "2. Anger, J., Facciolo, G., Delbracio, M.: Blind image deblurring using the 10 gradient prior. Image processing on line 9, 124-142 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.785, + 0.287 + ], + "angle": 0, + "content": "3. Boracchi, G., Foi, A.: Modeling the performance of image restoration from motion blur. Image Processing, IEEE Transactions on 21(8), 3502-3517 (aug 2012). https://doi.org/10.1109/TIP.2012.2192126" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.314 + ], + "angle": 0, + "content": "4. Borodenko, L.: Motion blur (May 2020), https://github.com/LeviBorodenko/motionblur" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.315, + 0.785, + 0.357 + ], + "angle": 0, + "content": "5. Carbajal, G., Vitoria, P., Delbracio, M., Muse, P., Lezama, J.: Non-uniform blur kernel estimation via adaptive basis decomposition. arXiv preprint arXiv:2102.01026 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.357, + 0.785, + 0.398 + ], + "angle": 0, + "content": "6. Carbajal, G., Vitoria, P., Lezama, J., Muse, P.: Blind motion deblurring with pixel-wise kernel estimation via kernel prediction networks. IEEE Transactions on Computational Imaging (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.785, + 0.44 + ], + "angle": 0, + "content": "7. Chakrabarti, A.: A neural approach to blind motion deblurring. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14. pp. 221-235. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.785, + 0.482 + ], + "angle": 0, + "content": "8. Chan, S.H., Wang, X., Elgendy, O.A.: Plug-and-play admm for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging 3(1), 84-98 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.785, + 0.509 + ], + "angle": 0, + "content": "9. Chan, T.F., Wong, C.K.: Total variation blind deconvolution. IEEE transactions on Image Processing 7(3), 370-375 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "0. Chen, L., Fang, F., Wang, T., Zhang, G.: Blind image deblurring with local maximum gradient prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1742-1750 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.55, + 0.785, + 0.592 + ], + "angle": 0, + "content": "1. Chen, L., Fang, F., Wang, T., Zhang, G.: Blind image deblurring with local maximum gradient prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1742-1750 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.633 + ], + "angle": 0, + "content": "2. Chen, L., Zhang, J., Li, Z., Wei, Y., Fang, F., Ren, J., Pan, J.: Deep richardson-lucy deconvolution for low-light image deblurring. International Journal of Computer Vision pp. 1-18 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.661 + ], + "angle": 0, + "content": "3. Cho, S., Lee, S.: Fast motion deblurring. ACM Transactions on Graphics pp. 1-8 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "4. Choi, J., Kim, S., Jeong, Y., Gwon, Y., Yoon, S.: Ilvr: Conditioning method for denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14367-14376 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.744 + ], + "angle": 0, + "content": "5. Chung, H., Kim, J., Kim, S., Ye, J.C.: Parallel diffusion models of operator and image for blind inverse problems. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6059-6069 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.797 + ], + "angle": 0, + "content": "6. Chung, H., Kim, J., Mccann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=OnD9zGAGT0k" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.798, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Chung, H., Ryu, D., McCann, M.T., Klasky, M.L., Ye, J.C.: Solving 3d inverse problems using pre-trained 2d diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22542-22551 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.129 + ], + "angle": 0, + "content": "Y. Sanghvi et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "18. Chung, H., Sim, B., Ryu, D., Ye, J.C.: Improving diffusion models for inverse problems using manifold constraints. Advances in Neural Information Processing Systems 35, 25683-25696 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "19. Chung, H., Sim, B., Ye, J.C.: Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12413-12422 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.273 + ], + "angle": 0, + "content": "20. Cohen, R., Elad, M., Milanfar, P.: Regularization by denoising via fixed-point projection (red-pro). SIAM Journal on Imaging Sciences 14(3), 1374-1406 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "21. Daras, G., Delbracio, M., Talebi, H., Dimakis, A., Milanfar, P.: Soft diffusion: Score matching with general corruptions. Transactions on Machine Learning Research (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.342 + ], + "angle": 0, + "content": "22. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.383 + ], + "angle": 0, + "content": "23. Dong, J., Pan, J., Sun, D., Su, Z., Yang, M.H.: Learning data terms for non-blind deblurring. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 748-763 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.424 + ], + "angle": 0, + "content": "24. Dong, J., Roth, S., Schiele, B.: Deep wiener deconvolution: Wiener meets deep learning for image deblurring. Advances in Neural Information Processing Systems 33, 1048-1059 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.467 + ], + "angle": 0, + "content": "25. Dong, J., Roth, S., Schiele, B.: Learning spatially-variant map models for non-blind image deblurring. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4886-4895 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.787, + 0.522 + ], + "angle": 0, + "content": "26. Eboli, T., Sun, J., Ponce, J.: End-to-end interpretable learning of non-blind image deblurring. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVII 16. pp. 314-331. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.787, + 0.563 + ], + "angle": 0, + "content": "27. Gong, D., Zhang, Z., Shi, Q., van den Hengel, A., Shen, C., Zhang, Y.: Learning deep gradient descent optimization for image deconvolution. IEEE transactions on neural networks and learning systems 31(12), 5468-5482 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.787, + 0.619 + ], + "angle": 0, + "content": "28. Gupta, A., Joshi, N., Lawrence Zitnick, C., Cohen, M., Curless, B.: Single image deblurring using motion density functions. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part I 11. pp. 171-184. Springer (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.787, + 0.646 + ], + "angle": 0, + "content": "29. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "30. Hu, Z., Yang, M.H.: Good regions to deblur. In: Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12. pp. 59-72. Springer (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.73 + ], + "angle": 0, + "content": "31. Jalal, A., Arvinte, M., Daras, G., Price, E., Dimakis, A.G., Tamir, J.: Robust compressed sensing mri with deep generative priors. Advances in Neural Information Processing Systems 34, 14938-14954 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.787, + 0.757 + ], + "angle": 0, + "content": "32. Jiang, H., Luo, A., Han, S., Fan, H., Liu, S.: Low-light image enhancement with wavelet-based diffusion models. arXiv preprint arXiv:2306.00306 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "33. Kadkhodaie, Z., Simoncelli, E.: Stochastic solutions for linear inverse problems using the prior implicit in a denoiser. Advances in Neural Information Processing Systems 34, 13242-13254 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "34. Kamilov, U.S., Mansour, H., Wohlberg, B.: A plug-and-play priors approach for solving nonlinear imaging inverse problems. IEEE Signal Processing Letters 24(12), 1872-1876 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.622, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Kernel Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "35. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. Advances in Neural Information Processing Systems 35, 23593-23606 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.177, + 0.787, + 0.204 + ], + "angle": 0, + "content": "36. Kawar, B., Vaksman, G., Elad, M.: Snips: Solving noisy inverse problems stochastically. Advances in Neural Information Processing Systems 34, 21757-21769 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.206, + 0.785, + 0.247 + ], + "angle": 0, + "content": "37. Kawar, B., Vaksman, G., Elad, M.: Stochastic image denoising by sampling from the posterior distribution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1866-1875 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.249, + 0.785, + 0.303 + ], + "angle": 0, + "content": "38. Kupyn, O., Budzan, V., Mykhailych, M., Mishkin, D., Matas, J.: Deblurgan: Blind motion deblurring using conditional adversarial networks. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8183-8192 (2018). https://doi.org/10.1109/CVPR.2018.00854" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.305, + 0.785, + 0.346 + ], + "angle": 0, + "content": "39. Kupyn, O., Martyniuk, T., Wu, J., Wang, Z.: Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In: The IEEE International Conference on Computer Vision (ICCV) (Oct 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.348, + 0.785, + 0.388 + ], + "angle": 0, + "content": "40. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding blind deconvolution algorithms. IEEE Transactions on Pattern Analysis and Machine Intelligence 33(12), 2354-2367 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.39, + 0.785, + 0.431 + ], + "angle": 0, + "content": "41. Li, H., Yang, Y., Chang, M., Chen, S., Feng, H., Xu, Z., Li, Q., Chen, Y.: Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing 479, 47-59 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.433, + 0.785, + 0.473 + ], + "angle": 0, + "content": "42. Li, X., Ren, Y., Jin, X., Lan, C., Wang, X., Zeng, W., Wang, X., Chen, Z.: Diffusion models for image restoration and enhancement-a comprehensive survey. arXiv preprint arXiv:2308.09388 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.475, + 0.785, + 0.529 + ], + "angle": 0, + "content": "43. Li, Y., Zhang, K., Liang, J., Cao, J., Liu, C., Gong, R., Zhang, Y., Tang, H., Liu, Y., Demandolx, D., et al.: LSDIR: A large scale dataset for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1775-1787 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.531, + 0.785, + 0.572 + ], + "angle": 0, + "content": "44. Li, Y., Tofighi, M., Geng, J., Monga, V., Eldar, Y.C.: Efficient and interpretable deep blind image deblurring via algorithm unrolling. IEEE Transactions on Computational Imaging 6, 666-681 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.574, + 0.785, + 0.614 + ], + "angle": 0, + "content": "45. Li, Y., Tofighi, M., Monga, V., Eldar, Y.C.: An algorithm unrolling approach to deep image deblurring. In: ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 7675-7679. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.616, + 0.785, + 0.657 + ], + "angle": 0, + "content": "46. Lim, B., Son, S., Kim, H., Nah, S., Lee, K.M.: Enhanced deep residual networks for single image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (July 2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.659, + 0.785, + 0.712 + ], + "angle": 0, + "content": "47. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: Proc. 8th Int'l Conf. Computer Vision. vol. 2, pp. 416-423 (July 2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.715, + 0.785, + 0.755 + ], + "angle": 0, + "content": "48. Monga, V., Li, Y., Eldar, Y.C.: Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine 38(2), 18-44 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "49. Murata, N., Saito, K., Lai, C.H., Takida, Y., Uesaka, T., Mitsufuji, Y., Ermon, S.: Gibbsddrm: A partially collapsed gibbs sampler for solving blind inverse problems with denoising diffusion restoration. arXiv preprint arXiv:2301.12686 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "50. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.129 + ], + "angle": 0, + "content": "Y. Sanghvi et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "51. Nan, Y., Ji, H.: Deep learning for handling kernel/model uncertainty in image deconvolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2388-2397 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "52. Pan, J., Sun, D., Pfister, H., Yang, M.H.: Blind image deblurring using dark channel prior. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1628-1636 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "53. Reehorst, E.T., Schniter, P.: Regularization by denoising: Clarifications and new interpretations. IEEE transactions on computational imaging 5(1), 52-67 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "54. Ren, D., Zhang, K., Wang, Q., Hu, Q., Zuo, W.: Neural blind deconvolution using deep priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3341-3350 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "55. Rim, J., Lee, H., Won, J., Cho, S.: Real-world blur dataset for learning and benchmarking deblurring algorithms. In: Proceedings of the European Conference on Computer Vision (ECCV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.37 + ], + "angle": 0, + "content": "56. Romano, Y., Elad, M., Milanfar, P.: The little engine that could: Regularization by denoising (red). SIAM Journal on Imaging Sciences 10(4), 1804-1844 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "57. Rond, A., Giryes, R., Elad, M.: Poisson inverse problems by the plug-and-play scheme. Journal of Visual Communication and Image Representation 41, 96-108 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.438 + ], + "angle": 0, + "content": "58. Saharia, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. arXiv:2104.07636 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.48 + ], + "angle": 0, + "content": "59. Sanghvi, Y., Gnanasambandam, A., Chan, S.H.: Photon limited non-blind deblurring using algorithm unrolling. IEEE Transactions on Computational Imaging 8, 851-864 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "60. Sanghvi, Y., Gnanasambandam, A., Mao, Z., Chan, S.H.: Photon-limited blind deconvolution using unsupervised iterative kernel estimation. IEEE Transactions on Computational Imaging 8, 1051-1062 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "61. Sanghvi, Y., Mao, Z., Chan, S.H.: Structured kernel estimation for photon-limited deconvolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9863-9872 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.592 + ], + "angle": 0, + "content": "62. Schuler, C.J., Hirsch, M., Harmeling, S., Scholkopf, B.: Learning to deblur. IEEE transactions on pattern analysis and machine intelligence 38(7), 1439-1451 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "63. Shan, Q., Jia, J., Agarwala, A.: High-quality motion deblurring from a single image. ACM Transactions on Graphics 27(3), 1-10 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "64. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.688 + ], + "angle": 0, + "content": "65. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv:2010.02502 (October 2020), https://arxiv.org/abs/2010.02502" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.742, + 0.702 + ], + "angle": 0, + "content": "66. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.757 + ], + "angle": 0, + "content": "67. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=PxTIG12RRHS" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "68. Sun, J., Cao, W., Xu, Z., Ponce, J.: Learning a convolutional neural network for non-uniform motion blur removal. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 769-777 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "69. Sun, L., Cho, S., Wang, J., Hays, J.: Edge-based blur kernel estimation using patch priors. In: IEEE international conference on computational photography (ICCP). pp. 1-8. IEEE (2013)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.622, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Kernel Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "70. Sun, Y., Wohlberg, B., Kamilov, U.S.: An online plug-and-play algorithm for regularized image reconstruction. IEEE Transactions on Computational Imaging 5(3), 395-408 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.191, + 0.787, + 0.233 + ], + "angle": 0, + "content": "71. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-recurrent network for deep image deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8174-8182 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.787, + 0.261 + ], + "angle": 0, + "content": "72. Ulyanov, D., Vedaldi, A., Lempitsky, V.: Deep image prior. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9446-9454 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.262, + 0.787, + 0.303 + ], + "angle": 0, + "content": "73. Venkatakrishnan, S.V., Bouman, C.A., Wohlberg, B.: Plug-and-play priors for model based reconstruction. In: 2013 IEEE global conference on signal and information processing. pp. 945-948. IEEE (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.305, + 0.787, + 0.332 + ], + "angle": 0, + "content": "74. Vincent, P.: A connection between score matching and denoising autoencoders. Neural computation 23(7), 1661-1674 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.333, + 0.787, + 0.374 + ], + "angle": 0, + "content": "75. Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. The Eleventh International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.375, + 0.787, + 0.431 + ], + "angle": 0, + "content": "76. Wang, Y., Yu, Y., Yang, W., Guo, L., Chau, L.P., Kot, A.C., Wen, B.: Exposurediffusion: Learning to expose for low-light image enhancement. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12438-12448 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.432, + 0.787, + 0.473 + ], + "angle": 0, + "content": "77. Wang, Z., Cun, X., Bao, J., Liu, J.: Uformer: A general u-shaped transformer for image restoration. in 2022 IEEE. In: CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17662-17672 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.475, + 0.787, + 0.515 + ], + "angle": 0, + "content": "78. Whang, J., Delbracio, M., Talebi, H., Sahara, C., Dimakis, A.G., Milanfar, P.: Deblurring via stochastic refinement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16293-16303 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.517, + 0.787, + 0.544 + ], + "angle": 0, + "content": "79. Xu, L., Jia, J.: Two-phase kernel estimation for robust motion deblurring. In: European Conference on Computer Vision. pp. 157-170. Springer (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.546, + 0.787, + 0.586 + ], + "angle": 0, + "content": "80. Xu, L., Zheng, S., Jia, J.: Unnatural l0 sparse representation for natural image deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1107-1114 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.588, + 0.787, + 0.629 + ], + "angle": 0, + "content": "81. Yan, Y., Ren, W., Guo, Y., Wang, R., Cao, X.: Image deblurring via extreme channels prior. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4003-4011 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.63, + 0.787, + 0.699 + ], + "angle": 0, + "content": "82. yang, y., Sun, J., Li, H., Xu, Z.: Deep admm-net for compressive sensing mri. In: Lee, D., Sugiyama, M., Luxburg, U., Guyon, I., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 29. Curran Associates, Inc. (2016), https://proceedings.neurips.cc/paper_files/paper/2016/file/1679091c5a880faf6fb5e6087eb1b2dc-Paper.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.7, + 0.787, + 0.755 + ], + "angle": 0, + "content": "83. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "84. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14821-14831 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "85. Zhang, H., Dai, Y., Li, H., Koniusz, P.: Deep stacked hierarchical multi-patch network for image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5978-5986 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.385, + 0.129 + ], + "angle": 0, + "content": "Y. Sanghvi et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "86. Zhang, K., Zuo, W., Gu, S., Zhang, L.: Learning deep cnn denoiser prior for image restoration. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3929-3938 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "87. Zhang, K., Ren, W., Luo, W., Lai, W.S., Stenger, B., Yang, M.H., Li, H.: Deep image deblurring: A survey. International Journal of Computer Vision 130(9), 2103-2130 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "88. Zhang, Z., Cheng, Y., Suo, J., Bian, L., Dai, Q.: Infwide: Image and feature space wiener deconvolution network for non-blind image deblurring in low-light conditions. IEEE Transactions on Image Processing 32, 1390-1402 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.328 + ], + "angle": 0, + "content": "89. Zhu, Y., Zhang, K., Liang, J., Cao, J., Wen, B., Timofte, R., Van Gool, L.: Denoising diffusion models for plug-and-play image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1219-1229 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.328 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Kernel Diffusion_ An Alternate Approach to Blind Deconvolution/0f568881-c0ff-47c0-b18b-592b7cbf5a93_origin.pdf b/2024/Kernel Diffusion_ An Alternate Approach to Blind Deconvolution/0f568881-c0ff-47c0-b18b-592b7cbf5a93_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6db503cb34722e7a89e7270fbe187cb8e925350e --- /dev/null +++ b/2024/Kernel Diffusion_ An Alternate Approach to Blind Deconvolution/0f568881-c0ff-47c0-b18b-592b7cbf5a93_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dfa131043af2aa318bde6768441ddbd924813516a9979103ee8e4e7d0f2c62f3 +size 11184382 diff --git a/2024/Kernel Diffusion_ An Alternate Approach to Blind Deconvolution/full.md b/2024/Kernel Diffusion_ An Alternate Approach to Blind Deconvolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..27be707d4d27a1cb50e283cd0d2036c8a36e4051 --- /dev/null +++ b/2024/Kernel Diffusion_ An Alternate Approach to Blind Deconvolution/full.md @@ -0,0 +1,429 @@ +# Kernel Diffusion: An Alternate Approach to Blind Deconvolution + +Yash Sanghvi1, Yiheng Chi1, and Stanley H. Chan1 + +Elmore Family School of Electrical and Computer Engineering, Purdue University, West Lafayette, Indiana, USA {ysanghvi,chi14,stanchan}@purdue.edu + +Abstract. Blind deconvolution problems are severely ill-posed because neither the underlying signal nor the forward operator are not known exactly. Conventionally, these problems are solved by alternating between estimation of the image and kernel while keeping the other fixed. In this paper, we show that this framework is flawed because of its tendency to get trapped in local minima and, instead, suggest the use of a kernel estimation strategy with a non-blind solver. This framework is employed by a diffusion method which is trained to sample the blur kernel from the conditional distribution with guidance from a pre-trained non-blind solver. The proposed diffusion method leads to state-of-the-art results on both synthetic and real blur datasets. + +Keywords: Blind Deconvolution $\cdot$ Diffusion Models $\cdot$ Deblurring + +# 1 Introduction + +Images taken by a hand-held camera under long exposure suffer from motion blur. Assuming this blur is spatially invariant, the captured image $\mathbf{y}$ and the underlying sharp image $\mathbf{x}$ are related as follows: + +$$ +\mathbf {y} = \mathbf {k} * \mathbf {x} + \mathbf {n}, \tag {1} +$$ + +where $\mathbf{k}$ , $\mathbf{n}$ refer to the blur kernel corresponding to the camera motion and noise, respectively, and the operator $\ast$ is convolution. The inverse problem of estimating $\mathbf{x}$ given $\mathbf{y}$ is called deconvolution, which can be classified into non-blind and blind. + +In the non-blind case, the blur kernel $\mathbf{k}$ is known. There are plenty of iterative methods which have been successfully applied to various image restoration tasks. Usually, these approaches, such as Plug-and-Play [1,8,34,57,70,73] and Regularization by Denoising (RED) [20,53,56], use a pre-trained image denoiser as a substitute for a complex image prior. Recent deep learning counterparts to these approaches involve unrolling the iterative methods followed by end-to-end training [45,48,82,86]. + +In blind deconvolution, the blur kernel $\mathbf{k}$ is not known. A natural strategy for the blind inverse problem is to jointly estimate the blur kernel $\mathbf{k}$ , and the + +image $\mathbf{x}$ by maximizing the posterior probability: + +$$ +\begin{array}{l} \hat {\mathbf {x}}, \hat {\mathbf {k}} = \underset {\mathbf {x}, \mathbf {k}} {\arg \max} \left[ p (\mathbf {x}, \mathbf {k} | \mathbf {y}) \right] \\ = \underset {\mathbf {x}, \mathbf {k}} {\arg \max } \left[ \log p (\mathbf {y} | \mathbf {k}, \mathbf {x}) + \log p (\mathbf {x}) + \log p (\mathbf {k}) \right]. \tag {2} \\ \end{array} +$$ + +The Maximum-A-Posteriori (MAP) for the pair $(\mathbf{x},\mathbf{k})$ , i.e. $\mathrm{MAP}_{\mathbf{x},\mathbf{k}}$ , can be obtained by an iterative scheme which alternates between the following minimization steps: + +$$ +\mathbf {x} ^ {i + 1} = \arg \min _ {\mathbf {x}} \left[ \| \mathbf {y} - \mathbf {x} * \mathbf {k} ^ {i} \| ^ {2} + \phi (\mathbf {x}) \right], \tag {3} +$$ + +$$ +\mathbf {k} ^ {i + 1} = \arg \operatorname * {m i n} _ {\mathbf {k}} \left[ \| \mathbf {y} - \mathbf {x} ^ {i + 1} * \mathbf {\Phi} \mathbf {k} \| ^ {2} + \psi (\mathbf {k}) \right], \tag {4} +$$ + +where $\phi(\cdot), \psi(\cdot)$ represent the image and kernel priors. Many classical deconvolution techniques can be framed as variations of this framework [2,9,63,80]. Even neural network based iterative methods such as Self-Deblur [54] and diffusion models [15,49] have achieved state-of-the-art performance in blind deconvolution using the stated framework. + +Despite recent success of diffusion methods, they are based on $\mathrm{MAP}_{\mathbf{x},\mathbf{k}}$ approach, which was first brought into question by Levin et al. [40] They argued that the number of variables to be estimated, i.e., $\{\mathbf{x},\mathbf{k}\}$ , are much larger than the number of known variables, i.e., $\mathbf{y}$ , and as a result, (2) favors the no-blur solution $\hat{\mathbf{x}},\hat{\mathbf{k}} = (\mathbf{y},\mathbf{I})$ . [40] proposed an alternate strategy of solving the kernel $\mathbf{k}$ first to exploit the inherent symmetry in the problem. This is achieved by marginalizing out the image space as follows: + +$$ +\hat {\mathbf {k}} = \underset {\mathbf {k}} {\arg \max } \left[ p (\mathbf {k} | \mathbf {y}) \right] = \underset {\mathbf {k}} {\arg \max } \left[ \int p (\mathbf {k}, \mathbf {x} | \mathbf {y}) d \mathbf {x} \right]. \tag {5} +$$ + +Although $\mathrm{MAP_k}$ optimization can lead to significantly improved reconstruction (as shown in Figure 1(c)) it has not been adopted in blind deconvolution solutions, which we surmise because of the intrinsic computational difficulty of the integration in (5). + +# 1.1 Contributions + +In this paper, we address the problem with a kernel estimation approach, and we demonstrate that this strategy can be implemented in practice. Our contribution in this paper can be summarized as follows: We propose a novel diffusion method, called Kernel-Diffusion (Kernel-Diff) which solves the $\mathrm{MAP_k}$ optimization problem by sampling from the conditional distribution $p(\mathbf{k}|\mathbf{y})$ . This diffusion method coupled with a differentiable non-blind solver achieves state-of-the-art deconvolution performance. We show with numerical experiments on real and synthetic data that the proposed approach creates significant performance improvement. We will soon release the source code and model weights for Kernel-Diff. + +![](images/6bd8424048f4acdff77721ad1edb5dd90169c5ff1dce62f4fed01bdee083e1a4.jpg) +(a) Blurred image Inset: true kernel + +![](images/59cecb71a103eb3e0bc94f99fd9b97c891c0087b1445c9b8fd681fdb65aa1f3a.jpg) +(b) Blind-DPS [15] based on alternating minimization + +![](images/dd7f9059013ace52f84651fad535dd383568eb78988b192323343882a47889d9.jpg) +(c) Kernel-Diff (ours) based on kernel estimation + +![](images/1adf6c343a729100a94d8573902c4351359370b2ac3cb917221e3d5fff3c3671.jpg) +(d) Ground Truth +Fig. 1: In this paper, we present a novel diffusion-based approach to the blind deconvolution problem. The conventional approaches, such as Blind-DPS, iterate between between kernel and image space minimization. Our approach uses a kernel-first estimation strategy in the context of diffusion models to sample kernels from the conditional distribution $p(\mathbf{k}|\mathbf{y})$ . + +# 2 Background + +# 2.1 Diffusion Methods + +Diffusion models [29, 64, 67, 74] are a popular generative method which can be used to sample any target diffusion $p_{data}(\mathbf{x})$ . The diffusion models begin with a forward process, which starts from $\mathbf{x}_0 \sim p_{data}(\mathbf{x})$ and gradually adds noise according to variance schedule $\{\beta_1, \beta_2, \ldots, \beta_T\}$ , formulated as + +$$ +\mathbf {x} _ {t} \sim \mathcal {N} \left(\sqrt {1 - \beta_ {t}} \mathbf {x} _ {t - 1}, \beta_ {t} \mathbf {I}\right) \tag {6} +$$ + +at time steps $t \in \{1, 2, \dots, T\}$ . This forward process converges to $\mathbf{x}_T \sim \mathcal{N}(0, \mathbf{I})$ . Next, the reverse diffusion process is used to sample from the distribution of $p_{data}$ by starting with $\mathbf{x}_T \sim \mathcal{N}(0, \mathbf{I})$ and performing the iterative updates + +$$ +\begin{array}{l} \mathbf {x} _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(\mathbf {x} _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\theta^ {*}} \left(\mathbf {x} _ {t}, t\right)\right) + \sigma_ {t} \mathbf {z} _ {t}, \tag {7} \\ \mathbf {z} _ {t} \sim \mathcal {N} (0, \mathbf {I}), \\ \end{array} +$$ + +where $\alpha_{t} \stackrel{\mathrm{def}}{=} 1 - \beta_{t}$ , $\bar{\alpha}_{t} \stackrel{\mathrm{def}}{=} \prod_{i=1}^{t} \alpha_{i}$ , and $\epsilon_{\theta^{*}}(\mathbf{x}_{t}, t)$ is the trained diffusion network used to learn to denoise $\mathbf{x}_{t}$ at any given time step, trained by minimizing the following loss function. + +$$ +\mathcal {L} (\theta) \stackrel {\text {d e f}} {=} \mathbb {E} _ {t, \mathbf {x} _ {0}, \epsilon} \left[ \| \epsilon - \epsilon_ {\theta} \left(\sqrt {\bar {\alpha} _ {t}} \mathbf {x} _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , t\right) \| ^ {2} \right]. \tag {8} +$$ + +Inverse Problems and Diffusion Models. Score based diffusion methods have been extensively used in solving inverse problems [17, 21, 33, 35-37, 89] by + +sampling from the posterior $p(\mathbf{x}|\mathbf{y})$ and the corresponding score is calculated using the Bayes' formula + +$$ +\nabla_ {\mathbf {x} _ {t}} \log p (\mathbf {x} _ {t} | \mathbf {y}) = \nabla_ {\mathbf {x} _ {t}} \log p (\mathbf {x} _ {t}) + \nabla_ {\mathbf {x} _ {t}} \log p (\mathbf {y} | \mathbf {x} _ {t}). \tag {9} +$$ + +For example, [31] uses an approximation to the posterior score in (9) and used the Langevin dynamics to sample from the distribution $p(\mathbf{x}|\mathbf{y})$ for the MRI compressed sensing problem; [18,19] solve the inpainting and super-resolution inverse problems with a modified reverse diffusion scheme which includes a projection step to reduce the consistency metric $\| \mathbf{y} - \mathbf{A}\mathbf{x}\|$ in each iteration; diffusion Posterior Sampling (DPS) [16] solves the noisy inverse problem with alternating diffusion and gradient descent steps. + +Image Restoration and Diffusion Models Diffusion models sample from a target distribution while MMSE estimators average over them [37]. As a result, the former has been widely used in image restoration tasks such as deblurring [78], super-resolution [41, 58], and low-light enhancement [32, 76]. For an exhaustive review, we refer the reader to [42]. + +Guidance in Diffusion Models. Pre-trained diffusion models are often guided at each iteration to produce desired results. This is first proposed in [22] where a classifier gradient is added to each iteration for the final output to belong to the given class. ILVR [14] guides the image diffusion process by constraining the low-frequency information of an image. In the same spirit, [75] guides the diffusion process by refining only the null-space of the corresponding inverse problem. [78] guides a deblurring diffusion model using a pretrained structure module. + +# 2.2 Blind Deconvolution + +Traditional blind deconvolution methods can be categorized under the $\mathrm{MAP}_{\mathbf{x},\mathbf{k}}$ framework [2,9,10,44,52,63,69,80,81]. For example, [13] uses salient edges for the kernel estimation and then performs image estimation using a Total Variation (TV) prior. [79] refines the kernel estimation process by iterative support detection. [28] performs alternating image and kernel minimization but uses a motion density representation for the blur kernel instead of for pixels. Self-Deblur [54] uses an alternating minimization scheme where a latent deep image prior [72] is used to represent the kernel and image space. + +Deep learning methods have replaced the optimization-based approaches and usually involve end-to-end training of image restoration networks [38, 62, 68, 71, 77, 83-85] using large datasets of pairs of sharp and blurry images, e.g., GoPro [50] and RealBlur [55] datasets. For a more exhaustive review of deep learning based approaches, we refer the readers to [87]. + +A few methods have utilized the $\mathrm{MAP_k}$ approach, even though not always explicitly stated by the authors. [7] uses a fully convolutional network to predict the Fourier coefficients of the underlying kernel which are used to deconvolve the blurry image in Fourier space. [5,6] train a Kernel-Prediction Network for all pixels in the image to deal with spatially varying blur. [60] leverages the differentiability of a non-blind solver [59] to frame the blind deconvolution problem + +entirely in terms of kernel estimation problems. This strategy is improved by [61] using a key-points based representation to further reduce the dimensionality of the problem. + +# 3 Method + +Blind-DPS [15] proposed a diffusion model based scheme for solving the blind inverse problem. Using separate diffusion models to sample from the image and kernel prior, a new scheme was proposed to sample solutions from the joint conditional distribution + +$$ +\hat {\mathbf {x}}, \hat {\mathbf {k}} \sim p (\mathbf {x}, \mathbf {k} | \mathbf {y}) \tag {10} +$$ + +This method is similar in principle to the conventional blind deconvolution techniques where the goal is to estimate the maximum of the conditional distribution i.e. + +$$ +\begin{array}{l} \hat {\mathbf {x}}, \hat {\mathbf {k}} = \underset {\mathbf {x}, \mathbf {k}} {\arg \max } \left[ p (\mathbf {x}, \mathbf {k} | \mathbf {y}) \right] (11) \\ = \underset {\mathbf {x}, \mathbf {k}} {\arg \min } \left[ \left\| \mathbf {y} - \mathbf {k} * \mathbf {x} \right\| ^ {2} + R _ {\mathbf {x}} (\mathbf {x}) + R _ {\mathbf {k}} (\mathbf {k}) \right] (12) \\ \end{array} +$$ + +where $R_{\mathbf{x}}(.)$ and $R_{\mathbf{k}}(.)$ represent the image and kernel priors respectively. This minimization is achieved using alternating minimization techniques where the kernel is estimated while keeping the image estimate fixed and vice versa. + +Recently, Blind-DPS proposed a diffusion method using a similar framework. In this approach, after each parallel reverse diffusion step from the image and kernel models, the estimates are refined using single gradient descent on $||\mathbf{y} - \mathbf{x} * \mathbf{k}||^2$ . Specifically, after each time step $t$ of parallel reverse diffusion on kernel and image, i.e. $\mathbf{k}_t$ and $\mathbf{x}_t$ , temporary estimates of the kernel and image are obtained as follows: + +$$ +\hat {\mathbf {x}} _ {0} \leftarrow \frac {1}{\sqrt {\bar {\alpha} _ {t}}} \left[ \mathbf {x} _ {t} + \sqrt {1 - \bar {\alpha} _ {t}} \hat {s} _ {\mathbf {x}} (\mathbf {x} _ {t}, t) \right] \tag {13} +$$ + +$$ +\hat {\mathbf {k}} _ {0} \leftarrow \frac {1}{\sqrt {\bar {\alpha} _ {t}}} \left[ \mathbf {k} _ {t} + \sqrt {1 - \bar {\alpha} _ {t}} \hat {s} _ {\mathbf {k}} (\mathbf {k} _ {t}, t) \right] \tag {14} +$$ + +where $\hat{s}_{\mathbf{x}}, \hat{s}_{\mathbf{k}}$ represent the corresponding score functions. This is followed by a gradient descent step on the reblurring loss for each update step + +$$ +\hat {\mathbf {x}} _ {t} ^ {\prime} \leftarrow \mathbf {x} _ {t} - \delta \nabla_ {\mathbf {x} _ {t}} \| \mathbf {y} - \hat {\mathbf {k}} _ {0} * \hat {\mathbf {x}} _ {0} \| ^ {2} \tag {15} +$$ + +$$ +\hat {\mathbf {k}} _ {t} ^ {\prime} \leftarrow \mathbf {k} _ {t} - \delta \nabla_ {\mathbf {k} _ {t}} \left[ \| \mathbf {y} - \hat {\mathbf {k}} _ {0} * \hat {\mathbf {x}} _ {0} \| ^ {2} + \lambda R _ {\mathbf {k}} (\hat {\mathbf {k}} _ {0}) \right] \tag {16} +$$ + +$R_{\mathbf{k}}(.)$ represents the $\ell_0 / \ell_1$ regularization on the kernel to impose sparsity and $\delta$ represents the step size change. + +While this alternating minimization approach to blind deconvolution is prevalent, this framework was first put into question by Levin et. al [40]. They showed + +that the conventional blind deconvolution techniques in this framework tended to converge to a no-blur degenerate solution, because the number of variables to be estimated i.e. $\{\mathbf{x},\mathbf{k}\}$ exceeded the number of knowns i.e. $\mathbf{y}$ . Instead they recommended marginalization of the image space and first estimating the kernel using the conditional dsiitribution + +$$ +\hat {\mathbf {k}} = \underset {\mathbf {k}} {\arg \max } \left[ p (\mathbf {k} | \mathbf {y}) \right] = \underset {\mathbf {k}} {\arg \max } \left[ \int p (\mathbf {k}, \mathbf {x} | \mathbf {y}) d \mathbf {x} \right] \tag {17} +$$ + +followed by estimation of image as follows: + +$$ +\hat {\mathbf {x}} = \underset {\mathbf {x}} {\arg \max } \left[ p (\mathbf {x} | \mathbf {y}, \hat {\mathbf {k}}) \right] \tag {18} +$$ + +Since the number of unknowns is fewer than the number of knowns in the first sub-problem in (17), this leads to a better kernel estimate and hence, a better image estimate. Applying this principle to deconvolution with diffusion models, we aim to propose a method which samples from the conditional distribution $p(\mathbf{k}|\mathbf{y})$ instead of $p(\mathbf{k},\mathbf{x}|\mathbf{y})$ since the former provides more stable solutions. + +# 3.1 Marginalization of Image Space + +In practice, the estimation of $p(\mathbf{k}|\mathbf{y})$ can be difficult because the marginalization in (5) requires a weighted average over all images. We surmise that due to lack of a feasible method to evaluate the likelihood, the approach from [40] was not explored by the imaging community for the blind deconvolution. To mitigate this issue, we first discuss estimation of the conditional distribution $p(\mathbf{k}|\mathbf{y})$ using a non-blind solver. + +First, we explicitly define a non-blind solver. Given a blurred image $\mathbf{y}$ and the corresponding blur kernel $\mathbf{k}$ , we can easily estimate the latent image using a non-blind solver which performs the following image estimation: + +$$ +\hat {\mathbf {x}} (\mathbf {k}) = \arg \max _ {\mathbf {x}} \left[ p (\mathbf {x} | \mathbf {k}, \mathbf {y}) \right] \stackrel {\mathrm {d e f}} {=} F (\mathbf {y}, \mathbf {k}). \tag {19} +$$ + +Traditionally, these non-blind solvers are implemented using an iterative method where the corresponding cost function for MAP estimation in (19) is minimized using gradient descent or proximal updates. Convolutional neural networks with Wiener-deconvolution processing have been used with great success also [12, 23, 24, 51, 59]. + +Remark 1: Given a non-blind solver $F_{\phi}(.)$ , we can approximate the conditional probability $p(\mathbf{k}|\mathbf{y})$ as follows: + +$$ +p (\mathbf {k} | \mathbf {y}) = \int p (\mathbf {x}, \mathbf {k} | \mathbf {y}) d \mathbf {x} \approx \frac {1}{Z (\mathbf {y})} p (\mathbf {y} \mid F _ {\phi} (\mathbf {y}, \mathbf {k}), \mathbf {k}) \tag {20} +$$ + +where $Z(\mathbf{y})$ is independent of the kernel $\mathbf{k}$ . Through the approximation, we can avoid the integration over the image space and express the likelihood $p(\mathbf{k}|\mathbf{y})$ in a tractable manner, up to the unknown the normalization constant $Z(\mathbf{y})$ . + +![](images/ac695fc85bf3d2f5a1e5fb782905aae5781bcf81bf8f0b0046504fa8a32e6de2.jpg) + +![](images/58303bbcaccb426ca74ae88de8380d40f78f2535622e4a36af6b5dbf48020e20.jpg) +Fig. 2: Proposed Kernel-Diff. Each iteration of the proposed kernel-based diffusion contains two sub-steps. First, we perform a reverse diffusion step using the trained diffusion model $\epsilon_{\theta}(\mathbf{k}_t,\mathbf{y},t)$ . Next we plug the temporary kernel estimate $\hat{\mathbf{k}}_0$ in the differentiable non-blind solver $F(\cdot)$ and perform a gradient descent update to guide the diffusion process to minimizing the reburring loss $\| \mathbf{y} - \hat{\mathbf{k}}_0\circledast \hat{\mathbf{x}}_0\|^2$ . + +# 3.2 Kernel-Based Diffusion + +To sample from $p(\mathbf{k}|\mathbf{y})$ , we propose to use a kernel-based diffusion network. This is different from existing methods using diffusion models, e.g., Blind-DPS [15], that resemble the alternating minimization approach, aiming to sample $\{\mathbf{x},\mathbf{k}\}$ from the conditional distribution $p(\mathbf{x},\mathbf{k}|\mathbf{y})$ using implicit score models for image and kernel spaces. + +Diffusion Network. We use a modified version of the U-Net architecture as opposed to the one described in [22]. Since we are sampling from a conditional distribution, the diffusion model $\epsilon_{\theta}(\mathbf{k},\mathbf{y},t)$ takes as input both the current kernel estimate $\mathbf{k}_t$ and the blurred image $\mathbf{y}$ . This is achieved in a similar fashion as [58]. Specifically, our U-Net architecture has two encoders, one for the image $\mathbf{y}$ and the other for the kernel $\mathbf{k}$ . In our implementation, we use images of size $256\times 256$ and kernels of size $64\times 64$ . At the end of the encoders, the encoded features of the kernel branch are upscaled by a factor of 4 and concatenated to the encoded features of the image branch. The concatenated features are then passed through a common U-Net decoder stage to predict the kernel at the next time step. + +The diffusion model $\epsilon_{\theta}(\cdot)$ is trained by minimizing the following loss function: + +$$ +\mathcal {L} _ {\mathrm {k e r n e l}} (\theta) \stackrel {\mathrm {d e f}} {=} \mathbb {E} _ {t, \mathbf {k}, \mathbf {y}, \epsilon} \left[ \| \epsilon - \epsilon_ {\theta} \left(\sqrt {\bar {\alpha} _ {t}} \mathbf {k} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , \mathbf {y}, t\right) \| _ {2} ^ {2} \right] \tag {21} +$$ + +where the expectation is taken over all timesteps $t$ , blurred images $\mathbf{y}$ and corresponding blur kernels $\mathbf{k}$ . With trained parameters $\theta^{*}$ , we sample $p(\mathbf{k}|\mathbf{y})$ using + +a series of reverse diffusion steps starting from $\mathbf{k}_T\sim \mathcal{N}(0,\mathbf{I})$ and the following reverse-diffusion iteration: + +$$ +\mathbf {k} _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(\mathbf {k} _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\theta^ {*}} \left(\mathbf {k} _ {t}, \mathbf {y}, t\right)\right) + \bar {\sigma} _ {t} \mathbf {z} _ {t} \tag {22} +$$ + +Guidance using Non-Blind Solver. Once we have sampled the kernel $\mathbf{k}_0$ , we estimate the latent image as $\mathbf{x}_0 = F_{\phi}(\mathbf{y},\mathbf{k}_0)$ , where $F_{\phi}(\cdot)$ is a neural network non-blind solver. There is a vast selection of candidate non-blind solver, including many network-based methods proposed recently [24-27, 59, 88]. These networks usually convert the blurred images to features space, apply Wiener deconvolution using the kernel followed by converting the deconvolved features into a single image using a decoder. + +While diffusion processes are effective in generating samples from the conditional distribution $p(\mathbf{k}|\mathbf{y})$ , they do not guarantee minimization of the reblurring loss, i.e., $\| \mathbf{y} - \mathbf{k} * \mathbf{x} \|^2$ . To minimize reblurring loss while also generating a plausible sample $p(\mathbf{k}|\mathbf{y})$ , we guide the reverse diffusion process using the pre-trained non-blind solver as follows. + +At step $t$ of the reverse diffusion process, the kernel estimate is given as $\mathbf{k}_t$ . Using the trained diffusion model $\epsilon_{\theta^*}(\mathbf{k}_t,\mathbf{y},t)$ and $\mathbf{k}_t$ , we can compute a temporary estimate of the sampled kernel $\mathbf{k}_0$ as follows: + +$$ +\hat {\mathbf {k}} _ {0} = \frac {1}{\sqrt {\bar {\alpha} _ {t}}} \left(\mathbf {k} _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {\theta^ {*}} \left(\mathbf {k} _ {t}, \mathbf {y}, t\right)\right). \tag {23} +$$ + +We then estimate the image using this kernel estimate $\hat{\mathbf{k}}_0$ by $\hat{\mathbf{x}}_0 \stackrel{\mathrm{def}}{=} F(\mathbf{y}, \hat{\mathbf{k}}_0)$ , plug into the reburring loss, and perform gradient descent by + +$$ +\mathbf {k} _ {t} \leftarrow \mathbf {k} _ {t} - \delta \nabla_ {\mathbf {k} _ {t}} \| \mathbf {y} - \hat {\mathbf {k}} _ {0} * F (\mathbf {y}, \hat {\mathbf {k}} _ {0}) \| ^ {2}. \tag {24} +$$ + +This step is equivalent to maximizing log likelihood of the approximation provided in (20). Taking the log of approximation, we get, + +$$ +\nabla_ {\mathbf {k}} \log p (\mathbf {k} | \mathbf {y}) \approx \nabla_ {\mathbf {k}} \left[ \| \mathbf {y} - F (\mathbf {y}, \mathbf {k}) * * \mathbf {k} \| _ {2} ^ {2} \right] \tag {25} +$$ + +Overall procedure. We summarize the overall procedure in Algorithm 1 and pictorially represent it in Figure 2. At any given iteration $t$ , we first perform reverse diffusion process using the denoising network $\epsilon_{\theta}(\cdot)$ to output $\mathbf{k}_{t-1/2}$ (Step 5). Next, we perform a gradient descent update on $\mathbf{k}_{t-1/2}$ the reburrowing loss $\| \mathbf{y} - \hat{\mathbf{k}}_0 * \hat{\mathbf{x}}_0 \|^2$ (Step 7-9). Similar to Blind-DPS [15], we set the step size $\delta$ according to following rule + +$$ +\delta = \frac {\rho}{\left\| \mathbf {y} - \hat {\mathbf {x}} _ {0} * \hat {\mathbf {k}} _ {0} \right\|} \tag {26} +$$ + +where $\rho = 0.05$ using trial and error. + +Algorithm 1 Kernel-Diff +1: Input: Blurred Image y, Diffusion UNet $\epsilon_{\theta}(.)$ , Non-Blind Solver $F_{\phi}(.)$ +2: // Stage I: Reverse Diffusion +3: $\mathbf{k}_T\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ +4: for $t = T,\dots ,1$ do +5: $\mathbf{z}_t\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ +6: $\mathbf{k}_{t - 1 / 2} = \frac{1}{\sqrt{\alpha_t}}\Bigg(\mathbf{k}_t - \frac{1 - \alpha_t}{\sqrt{1 - \bar{\alpha}_t}}\epsilon_\theta (\mathbf{k}_t,\mathbf{y},t)\Bigg) + \sigma_t\mathbf{z}_t$ +7: $\hat{\mathbf{k}}_0 = \frac{1}{\sqrt{\bar{\alpha}_t}}\big(\mathbf{k}_t - \sqrt{1 - \bar{\alpha}_t}\epsilon_\theta (\mathbf{k}_t,\mathbf{y},t)\big)$ +8: $\hat{\mathbf{x}}_0 = F_{\phi}(\mathbf{y},\hat{\mathbf{k}}_0)$ +9: $\mathbf{k}_{t - 1} = \mathbf{k}_{t - 1 / 2} - \delta \nabla_{\mathbf{k}_t}(\| \mathbf{y} - \hat{\mathbf{k}}_0*\hat{\mathbf{x}}_0\| _2^2)$ +10: end for +11: // Stage II: Refinement +12: for $j = 1,\dots ,J$ do +13: $\mathbf{k}_0\gets \mathbf{k}_0 - \delta \nabla_{\mathbf{k}_0}(\| \mathbf{y} - \mathbf{k}_0*\boldsymbol {F}_\phi (\mathbf{y},\mathbf{k}_0)\| _2^2 +\gamma \| \mathbf{k}\| _1)$ +14: end for +15: return $\mathbf{k}_0$ and ${\bf x}_0 = F_{\phi}({\bf y},{\bf k}_0)$ + +For further performance boost, once the reverse diffusion process is over, we repeat the gradient descent steps (Step 13) to refine the kernel, similar to Stage II in [61]. Here the overall reburring loss is minimized along with $\ell_1$ penalty on the kernel with sparisty weight $\gamma = 10^{-4}$ . Since this stage is meant to only fine-tune the kernel, we keep the number of total iterations in second stage, i.e. $J$ , to only 20. + +# 4 Experiments + +# 4.1 Experiment Setup + +Implementation of Kernel-Diff. The diffusion network we use is a modified U-Net [22]. To train the diffusion model $\epsilon_{\theta}(\mathbf{k}_t,\mathbf{y},t)$ , we generate blurred data using synthetic motion blur kernels, with details of data synthesis to be elaborated in the next section. We assume symmetric boundary conditions during the convolution process. We train the diffusion U-Net architecture for 500,000 iterations with a batch size of 16. The non-blind solver in our scheme, i.e. $F_{\phi}(.)$ , can be any differentiable neural network based method. In this paper, we plug Deep Wiener Deconvolution (DWDN) [24] to guide the reverse diffusion process. Training Dataset. The diffusion U-Net in Kernel-Diff as well as all methods we compare in our experiments are trained using the same set of blurred and clean image data. We synthesize the blurred images by applying random blur kernels from a kernel database on clean images, similar to [15]. The kernel database consists of $60,000\times 64$ blur kernels generated using an open-sourced motion blur model [4] and blur kernel generation method provided by [3]. We note that + +the former method from [4] produces larger blur kernels than usually observed in practice. However, we include them in our analysis for sake of fair comparison to Blind-DPS which uses method to train diffusion model for blur kernel generation. For clean images, we use the full-resolution images from Flickr2K [46] and LSIDR [43] since diffusion models can be require large amounts of training data. To train the diffusion U-Net for Kernel-Diff, we additionally randomly crop the images to $256 \times 256$ patches and augment the data by random horizontal flipping. + +Table 1: Performance on BSD100 dataset with synthetic blur. Bold and underline refer to overall and second best blind deconvolution method respectively. Maximum normalized correlation (MNC) is provided to evaluate estimated kernels, when explicitly provided in case of Self-Deblur, Blind-DPS etc. DWDN is a non-blind method which knows the ground-truth kernel (i.e. $\mathrm{MNC} = 1.00$ ) and is provided as an empirical upper bound only. + +
Method → Metric ↓DeblurGAN [39]DMPHN [85]MPR-Net [84]Self-Deblur [54]Fast-Deblur [13]Blind-DPS [15]Kernel-Diff (Ours)DWDN [24]
PSNR ↑15.4418.7318.4913.5314.0217.5619.0722.95
SSIM ↑0.3960.4270.4350.2270.2960.3870.5000.810
LPIPS ↓0.5630.6170.6320.6650.5830.6700.3550.169
FID ↓287.44265.05265.60259.40304.16280.53172.3379.11
MNC ↑---0.2060.5200.2830.6631.0
+ +Table 2: Performance on RealBlur-50. Bold and underline refer to overall and second best blind deconvolution method respectively. The corresponding ground truth kernel is not provided for the RealBlur [55] dataset since the blur kernel may not be spatially invariant. As a result the metric MNC along with the performance on DWDN is not provided. + +
Method → Metric ↓DeblurGAN [39]Fast-Deblur [13]Self-Deblur [54]Blind-DPS [85]LMG [11]DMPHN [85]MPR-Net [84]Kernel-Diff (Ours)
PSNR ↑14.9818.5416.1519.8224.2024.3723.6325.57
SSIM ↑0.5830.4980.4830.6070.7890.7190.7150.798
LPIPS ↓0.2890.3920.4590.3200.1830.2010.2130.171
FID ↓129.06236.21172.32175.95106.35135.33113.1198.89
+ +# 4.2 Quantitative Comparison + +In this section, we evaluate the proposed method Kernel-Diff on synthetic and real blurred images and compare the performance with other state-of-the-art deconvolution methods. We evaluate our method on the BSD100 and RealBlur datasets as described below along with the following deconvolution/deblurring methods - Deep Hierarchical Multi-Patch Network (DHMPN) [85], Self-Deblur + +![](images/c36b0ffc10d152c4048ee83f800e638cac306bb2e2d672b3b0ea71a8151b4754.jpg) +Fig. 3: Qualitative Results: Reconstruction results on synthetically blurred images from the BSD100 dataset. For Self-Deblur, Blind-DPS and Kernel-Diff, we show the estimated kernels in inset. $\dagger$ DWDN is a non-blind deconvolution method assuming knowledge of the ground-truth kernel; its results are provided for reference only. + +[54], Blind-DPS [15], MPR-Net [84], and DeblurGANv2 [39]. We use these methods as a representative set of contemporary deconvolution methods which include both iterative and end-to-end trained solutions. For the sake of a fair comparison, we retrain the end-to-end methods using synthetically blurred datasets with details provided in Section B of the supplementary document. + +BSD100 Dataset. We first evaluate our method on 100 synthetically blurred images from the BSD dataset [47]. Specifically, we crop out random patches of size $256 \times 256$ and then blur them with synthetic motion blur kernels using symmetric boundary conditions. The performance of Kernel-Diff and other deblurring methods is shown in Table 1 along with some qualitative examples in Figure 3. We also compare estimated kernels from Self-Deblur, Blind-DPS and Kernel-Diff in Table 1 using maximum of normalized convolution (MNC) [30]. + +RealBlur-50. To evaluate our blind deconvolution method on real blurred images, we use the RealBlur-J dataset [55]. However, since the focus of this paper is only spatially invariant blur and not other forms of blur, we crop random patches of size $256 \times 256$ from 50 blurred images in the dataset to reduce the possibility of violating the spatially invarying blur assumption. The results of this comparison are shown in Table 2. We note that we for improved performance on the RealBlur dataset, we use a diffusion model trained only using the realistic blur kernels generated from [3] and increase the sparsity penalty $\gamma$ to $5 \times 10^{-4}$ + +# 4.3 Ablation Study + +Effect of Non-Blind Solver Guidance Each iteration of the diffusion consists of two sub-steps: (1) reverse diffusion using the trained denoiser $\epsilon_{\theta}(\cdot)$ and (2) gradient descent on $\| \mathbf{y} - \hat{\mathbf{k}}_0 * \hat{\mathbf{x}}_0 \|^2$ using the non-blind solver. In Table 3, we show the effect on performance of the second sub-step. Through the giant gap in performance with and without the guidance, we find that reverse diffusion is insufficient to converge to a good kernel estimate and hence, the experiment justifies the extra complexity required to guide the diffusion process to minimize the reblurring loss. In the same ablation study, we show how using the correct step size is important for the performance and how the second stage of finetuning impacts the overall scheme's performance. + +Computational Comparison The method proposed in this paper, Kernel-Diff is a kernel estimation analogue of the Blind-DPS method as both of them aim to solve the blind deconvolution problem using diffusion models. Therefore, we compare the computational costs of both the techniques in Table 4. Both techniques use a diffusion model trained for $T = 1000$ steps and our method performs much better than Blind-DPS which uses the alternating minimization approach using fewer parameters and similar time. Future work may involve reducing the inference time using DDIM [65] or consistency models [66]. + +Reblurring Loss during Diffusion Levin et. al [40] show that alternate minimization approach is more likely to end up in a false local minima. We empirically demonstrate this by comparing the residual across the iterations, i.e., $\| \mathbf{y} - \hat{\mathbf{k}}_0\otimes \hat{\mathbf{x}}_0\|$ with Blind-DPS [15] in Figure 5. From the figure we observe that + +the residual in our proposed kernel estimation scheme converges to a solution with a much lower residual value. Assuming the kernel and image likelihoods to be approximately equal for both solutions, we conclude that our method leads to solutions with a higher log-likelihood, and hence, better quality results. + +![](images/9e0e5e2c59910da4a00c34a75d59c57f57f538231c6444d91e86ef919665036c.jpg) +w/o non-blind solver guidance + +![](images/64e324936f037901ee3719ea1fd566b2fdd0973739bbae463f6936071e10f7e0.jpg) +with non-blind solver guidance + +![](images/2ae7b294a0177dd6ff332cc044075edc2597a665db99e24846b3a50e0fc23f93.jpg) +DWDN + True kernel +Fig. 4: Effect of Non-Blind Solver Guidance. Through this ablation study, we examine the effect of the second sub-step in the diffusion process which guides the diffusion process to minimize the reburring loss using the non-blind solver DWDN [24]. + +Table 3: Ablation study comparing effect of different hyperparameters of the scheme on the performance. We observe that removing the guidance of the non-blind solver reduces the performance across all metrics. While using gradient descent using non-blind solver in (24), for optimal performance, the step size shouldn't be set too large. Lastly, the effect of finetuning kernel in Stage-II can also provide a significant performance boost for PSNR, however, not so much for other metrics. + +
Ablation Study ↓ Metric →PSNR ↑SSIM ↑LPIPS ↓FID ↓MNC ↑
ρ = 0, No DWDN21.900.6500.251152.510.714
ρ = 0.0522.640.6800.234140.280.724
ρ = 0.121.970.6510.256156.040.567
ρ = 0.520.870.6090.300173.360.270
J = 0, No Stage II21.760.6710.241143.580.675
DWDN, known kernel30.540.9060.08749.291.00
+ +# 5 Conclusion + +In this paper, we proposed Kernel-Diff, a blind deconvolution solution in the tradition of Levin et. al [40]. We show how the kernel minimization approach can be + +![](images/8b081afa8d5e605ab5d803911d7f0664d0f6b9ef1b3c528ab73332ef5f677531.jpg) +Fig. 5: Residual $\| \mathbf{y} - \hat{\mathbf{k}}_0\otimes \hat{\mathbf{x}}_0\|^2$ for Blind-DPS and Kernel-Diff. This experiments confirms our hypothesis that the proposed kernel estimation scheme converges to a better local minimum. The shaded region shows the range of the residual over 50 trials of each scheme. + +Table 4: Comparison of computational cost with Blind-DPS in terms of total number of parameters and time taken to deblur a $256 \times 256$ image respectively. Both Blind-DPS and Kernel-Diff use a diffusion model approach to the blind deconvolution problem, our kernel estimation approach uses a lightweight diffusion model and non-blind solver for the reverse diffusion process. Using significantly fewer parameters and smaller computational time, we achieve significant performance boost over the alternating minimization approach. + +
Metric → Method ↓Total Parameter CountTotal Time
Blind-DPS110M337s
Kernel-Diff + DWDN43M305s
+ +implemented in practice using a pre-trained non-blind solver in diffusion models setting and propose a guided diffusion method that samples from $p(\mathbf{k}|\mathbf{y})$ to solve the blind deconvolution problem and show extensive numerical experiments on blurred images. + +For future work, the method can be refined to deblur images with spatially varying and motion blur i.e. degradation which do not fit the single-blur convolutional model. Alternately, the non-blind solver can be modified to be robust to inaccuracies in kernel estimation. + +# Acknowledgements + +This work is supported, in part, by the DARPA / SRC CogniSense JUMP 2.0 Center, NSF IIS-2133032, and NSF ECCS-2030570. + +# References + +1. Ahmad, R., Bouman, C.A., Buzzard, G.T., Chan, S., Liu, S., Reehorst, E.T., Schniter, P.: Plug-and-play methods for magnetic resonance imaging: Using de-noisers for image recovery. IEEE signal processing magazine 37(1), 105-116 (2020) +2. Anger, J., Facciolo, G., Delbracio, M.: Blind image deblurring using the 10 gradient prior. Image processing on line 9, 124-142 (2019) +3. Boracchi, G., Foi, A.: Modeling the performance of image restoration from motion blur. Image Processing, IEEE Transactions on 21(8), 3502-3517 (aug 2012). https://doi.org/10.1109/TIP.2012.2192126 +4. Borodenko, L.: Motion blur (May 2020), https://github.com/LeviBorodenko/motionblur +5. Carbajal, G., Vitoria, P., Delbracio, M., Muse, P., Lezama, J.: Non-uniform blur kernel estimation via adaptive basis decomposition. arXiv preprint arXiv:2102.01026 (2021) +6. Carbajal, G., Vitoria, P., Lezama, J., Muse, P.: Blind motion deblurring with pixel-wise kernel estimation via kernel prediction networks. IEEE Transactions on Computational Imaging (2023) +7. Chakrabarti, A.: A neural approach to blind motion deblurring. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14. pp. 221-235. Springer (2016) +8. Chan, S.H., Wang, X., Elgendy, O.A.: Plug-and-play admm for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging 3(1), 84-98 (2016) +9. Chan, T.F., Wong, C.K.: Total variation blind deconvolution. IEEE transactions on Image Processing 7(3), 370-375 (1998) +0. Chen, L., Fang, F., Wang, T., Zhang, G.: Blind image deblurring with local maximum gradient prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1742-1750 (2019) +1. Chen, L., Fang, F., Wang, T., Zhang, G.: Blind image deblurring with local maximum gradient prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1742-1750 (2019) +2. Chen, L., Zhang, J., Li, Z., Wei, Y., Fang, F., Ren, J., Pan, J.: Deep richardson-lucy deconvolution for low-light image deblurring. International Journal of Computer Vision pp. 1-18 (2023) +3. Cho, S., Lee, S.: Fast motion deblurring. ACM Transactions on Graphics pp. 1-8 (2009) +4. Choi, J., Kim, S., Jeong, Y., Gwon, Y., Yoon, S.: Ilvr: Conditioning method for denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14367-14376 (October 2021) +5. Chung, H., Kim, J., Kim, S., Ye, J.C.: Parallel diffusion models of operator and image for blind inverse problems. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6059-6069 (2023) +6. Chung, H., Kim, J., Mccann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=OnD9zGAGT0k +7. Chung, H., Ryu, D., McCann, M.T., Klasky, M.L., Ye, J.C.: Solving 3d inverse problems using pre-trained 2d diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22542-22551 (2023) + +18. Chung, H., Sim, B., Ryu, D., Ye, J.C.: Improving diffusion models for inverse problems using manifold constraints. Advances in Neural Information Processing Systems 35, 25683-25696 (2022) +19. Chung, H., Sim, B., Ye, J.C.: Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12413-12422 (2022) +20. Cohen, R., Elad, M., Milanfar, P.: Regularization by denoising via fixed-point projection (red-pro). SIAM Journal on Imaging Sciences 14(3), 1374-1406 (2021) +21. Daras, G., Delbracio, M., Talebi, H., Dimakis, A., Milanfar, P.: Soft diffusion: Score matching with general corruptions. Transactions on Machine Learning Research (2023) +22. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021) +23. Dong, J., Pan, J., Sun, D., Su, Z., Yang, M.H.: Learning data terms for non-blind deblurring. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 748-763 (2018) +24. Dong, J., Roth, S., Schiele, B.: Deep wiener deconvolution: Wiener meets deep learning for image deblurring. Advances in Neural Information Processing Systems 33, 1048-1059 (2020) +25. Dong, J., Roth, S., Schiele, B.: Learning spatially-variant map models for non-blind image deblurring. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4886-4895 (2021) +26. Eboli, T., Sun, J., Ponce, J.: End-to-end interpretable learning of non-blind image deblurring. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVII 16. pp. 314-331. Springer (2020) +27. Gong, D., Zhang, Z., Shi, Q., van den Hengel, A., Shen, C., Zhang, Y.: Learning deep gradient descent optimization for image deconvolution. IEEE transactions on neural networks and learning systems 31(12), 5468-5482 (2020) +28. Gupta, A., Joshi, N., Lawrence Zitnick, C., Cohen, M., Curless, B.: Single image deblurring using motion density functions. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part I 11. pp. 171-184. Springer (2010) +29. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) +30. Hu, Z., Yang, M.H.: Good regions to deblur. In: Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12. pp. 59-72. Springer (2012) +31. Jalal, A., Arvinte, M., Daras, G., Price, E., Dimakis, A.G., Tamir, J.: Robust compressed sensing mri with deep generative priors. Advances in Neural Information Processing Systems 34, 14938-14954 (2021) +32. Jiang, H., Luo, A., Han, S., Fan, H., Liu, S.: Low-light image enhancement with wavelet-based diffusion models. arXiv preprint arXiv:2306.00306 (2023) +33. Kadkhodaie, Z., Simoncelli, E.: Stochastic solutions for linear inverse problems using the prior implicit in a denoiser. Advances in Neural Information Processing Systems 34, 13242-13254 (2021) +34. Kamilov, U.S., Mansour, H., Wohlberg, B.: A plug-and-play priors approach for solving nonlinear imaging inverse problems. IEEE Signal Processing Letters 24(12), 1872-1876 (2017) + +35. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. Advances in Neural Information Processing Systems 35, 23593-23606 (2022) +36. Kawar, B., Vaksman, G., Elad, M.: Snips: Solving noisy inverse problems stochastically. Advances in Neural Information Processing Systems 34, 21757-21769 (2021) +37. Kawar, B., Vaksman, G., Elad, M.: Stochastic image denoising by sampling from the posterior distribution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1866-1875 (2021) +38. Kupyn, O., Budzan, V., Mykhailych, M., Mishkin, D., Matas, J.: Deblurgan: Blind motion deblurring using conditional adversarial networks. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8183-8192 (2018). https://doi.org/10.1109/CVPR.2018.00854 +39. Kupyn, O., Martyniuk, T., Wu, J., Wang, Z.: Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In: The IEEE International Conference on Computer Vision (ICCV) (Oct 2019) +40. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding blind deconvolution algorithms. IEEE Transactions on Pattern Analysis and Machine Intelligence 33(12), 2354-2367 (2011) +41. Li, H., Yang, Y., Chang, M., Chen, S., Feng, H., Xu, Z., Li, Q., Chen, Y.: Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing 479, 47-59 (2022) +42. Li, X., Ren, Y., Jin, X., Lan, C., Wang, X., Zeng, W., Wang, X., Chen, Z.: Diffusion models for image restoration and enhancement-a comprehensive survey. arXiv preprint arXiv:2308.09388 (2023) +43. Li, Y., Zhang, K., Liang, J., Cao, J., Liu, C., Gong, R., Zhang, Y., Tang, H., Liu, Y., Demandolx, D., et al.: LSDIR: A large scale dataset for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1775-1787 (2023) +44. Li, Y., Tofighi, M., Geng, J., Monga, V., Eldar, Y.C.: Efficient and interpretable deep blind image deblurring via algorithm unrolling. IEEE Transactions on Computational Imaging 6, 666-681 (2020) +45. Li, Y., Tofighi, M., Monga, V., Eldar, Y.C.: An algorithm unrolling approach to deep image deblurring. In: ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 7675-7679. IEEE (2019) +46. Lim, B., Son, S., Kim, H., Nah, S., Lee, K.M.: Enhanced deep residual networks for single image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (July 2017) +47. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: Proc. 8th Int'l Conf. Computer Vision. vol. 2, pp. 416-423 (July 2001) +48. Monga, V., Li, Y., Eldar, Y.C.: Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine 38(2), 18-44 (2021) +49. Murata, N., Saito, K., Lai, C.H., Takida, Y., Uesaka, T., Mitsufuji, Y., Ermon, S.: Gibbsddrm: A partially collapsed gibbs sampler for solving blind inverse problems with denoising diffusion restoration. arXiv preprint arXiv:2301.12686 (2023) +50. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017) + +51. Nan, Y., Ji, H.: Deep learning for handling kernel/model uncertainty in image deconvolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2388-2397 (2020) +52. Pan, J., Sun, D., Pfister, H., Yang, M.H.: Blind image deblurring using dark channel prior. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1628-1636 (2016) +53. Reehorst, E.T., Schniter, P.: Regularization by denoising: Clarifications and new interpretations. IEEE transactions on computational imaging 5(1), 52-67 (2018) +54. Ren, D., Zhang, K., Wang, Q., Hu, Q., Zuo, W.: Neural blind deconvolution using deep priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3341-3350 (2020) +55. Rim, J., Lee, H., Won, J., Cho, S.: Real-world blur dataset for learning and benchmarking deblurring algorithms. In: Proceedings of the European Conference on Computer Vision (ECCV) (2020) +56. Romano, Y., Elad, M., Milanfar, P.: The little engine that could: Regularization by denoising (red). SIAM Journal on Imaging Sciences 10(4), 1804-1844 (2017) +57. Rond, A., Giryes, R., Elad, M.: Poisson inverse problems by the plug-and-play scheme. Journal of Visual Communication and Image Representation 41, 96-108 (2016) +58. Saharia, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. arXiv:2104.07636 (2021) +59. Sanghvi, Y., Gnanasambandam, A., Chan, S.H.: Photon limited non-blind deblurring using algorithm unrolling. IEEE Transactions on Computational Imaging 8, 851-864 (2022) +60. Sanghvi, Y., Gnanasambandam, A., Mao, Z., Chan, S.H.: Photon-limited blind deconvolution using unsupervised iterative kernel estimation. IEEE Transactions on Computational Imaging 8, 1051-1062 (2022) +61. Sanghvi, Y., Mao, Z., Chan, S.H.: Structured kernel estimation for photon-limited deconvolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9863-9872 (2023) +62. Schuler, C.J., Hirsch, M., Harmeling, S., Scholkopf, B.: Learning to deblur. IEEE transactions on pattern analysis and machine intelligence 38(7), 1439-1451 (2015) +63. Shan, Q., Jia, J., Agarwala, A.: High-quality motion deblurring from a single image. ACM Transactions on Graphics 27(3), 1-10 (2008) +64. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015) +65. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv:2010.02502 (October 2020), https://arxiv.org/abs/2010.02502 +66. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models (2023) +67. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=PxTIG12RRHS +68. Sun, J., Cao, W., Xu, Z., Ponce, J.: Learning a convolutional neural network for non-uniform motion blur removal. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 769-777 (2015) +69. Sun, L., Cho, S., Wang, J., Hays, J.: Edge-based blur kernel estimation using patch priors. In: IEEE international conference on computational photography (ICCP). pp. 1-8. IEEE (2013) + +70. Sun, Y., Wohlberg, B., Kamilov, U.S.: An online plug-and-play algorithm for regularized image reconstruction. IEEE Transactions on Computational Imaging 5(3), 395-408 (2019) +71. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-recurrent network for deep image deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8174-8182 (2018) +72. Ulyanov, D., Vedaldi, A., Lempitsky, V.: Deep image prior. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9446-9454 (2018) +73. Venkatakrishnan, S.V., Bouman, C.A., Wohlberg, B.: Plug-and-play priors for model based reconstruction. In: 2013 IEEE global conference on signal and information processing. pp. 945-948. IEEE (2013) +74. Vincent, P.: A connection between score matching and denoising autoencoders. Neural computation 23(7), 1661-1674 (2011) +75. Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. The Eleventh International Conference on Learning Representations (2023) +76. Wang, Y., Yu, Y., Yang, W., Guo, L., Chau, L.P., Kot, A.C., Wen, B.: Exposurediffusion: Learning to expose for low-light image enhancement. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12438-12448 (2023) +77. Wang, Z., Cun, X., Bao, J., Liu, J.: Uformer: A general u-shaped transformer for image restoration. in 2022 IEEE. In: CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17662-17672 (2021) +78. Whang, J., Delbracio, M., Talebi, H., Sahara, C., Dimakis, A.G., Milanfar, P.: Deblurring via stochastic refinement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16293-16303 (2022) +79. Xu, L., Jia, J.: Two-phase kernel estimation for robust motion deblurring. In: European Conference on Computer Vision. pp. 157-170. Springer (2010) +80. Xu, L., Zheng, S., Jia, J.: Unnatural l0 sparse representation for natural image deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1107-1114 (2013) +81. Yan, Y., Ren, W., Guo, Y., Wang, R., Cao, X.: Image deblurring via extreme channels prior. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4003-4011 (2017) +82. yang, y., Sun, J., Li, H., Xu, Z.: Deep admm-net for compressive sensing mri. In: Lee, D., Sugiyama, M., Luxburg, U., Guyon, I., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 29. Curran Associates, Inc. (2016), https://proceedings.neurips.cc/paper_files/paper/2016/file/1679091c5a880faf6fb5e6087eb1b2dc-Paper.pdf +83. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022) +84. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14821-14831 (2021) +85. Zhang, H., Dai, Y., Li, H., Koniusz, P.: Deep stacked hierarchical multi-patch network for image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5978-5986 (2019) + +86. Zhang, K., Zuo, W., Gu, S., Zhang, L.: Learning deep cnn denoiser prior for image restoration. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3929-3938 (2017) +87. Zhang, K., Ren, W., Luo, W., Lai, W.S., Stenger, B., Yang, M.H., Li, H.: Deep image deblurring: A survey. International Journal of Computer Vision 130(9), 2103-2130 (2022) +88. Zhang, Z., Cheng, Y., Suo, J., Bian, L., Dai, Q.: Infwide: Image and feature space wiener deconvolution network for non-blind image deblurring in low-light conditions. IEEE Transactions on Image Processing 32, 1390-1402 (2023) +89. Zhu, Y., Zhang, K., Liang, J., Cao, J., Wen, B., Timofte, R., Van Gool, L.: Denoising diffusion models for plug-and-play image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1219-1229 (2023) \ No newline at end of file diff --git a/2024/Kernel Diffusion_ An Alternate Approach to Blind Deconvolution/images.zip b/2024/Kernel Diffusion_ An Alternate Approach to Blind Deconvolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..da24d76859fa13fa2a3d70e60cf2b8beb204b91e --- /dev/null +++ b/2024/Kernel Diffusion_ An Alternate Approach to Blind Deconvolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df18c04d00cbde808f1f6cafc38ceb3801696bab4645d91577ffab5d44a36a32 +size 702196 diff --git a/2024/Kernel Diffusion_ An Alternate Approach to Blind Deconvolution/layout.json b/2024/Kernel Diffusion_ An Alternate Approach to Blind Deconvolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8e13500decbe50abb81e52792f0c0e53d8cb2483 --- /dev/null +++ b/2024/Kernel Diffusion_ An Alternate Approach to Blind Deconvolution/layout.json @@ -0,0 +1,12158 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 111, + 462, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 111, + 462, + 148 + ], + "spans": [ + { + "bbox": [ + 153, + 111, + 462, + 148 + ], + "type": "text", + "content": "Kernel Diffusion: An Alternate Approach to Blind Deconvolution" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 181, + 167, + 432, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 167, + 432, + 181 + ], + "spans": [ + { + "bbox": [ + 181, + 167, + 432, + 181 + ], + "type": "text", + "content": "Yash Sanghvi1, Yiheng Chi1, and Stanley H. Chan1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 190, + 474, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 190, + 474, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 190, + 474, + 224 + ], + "type": "text", + "content": "Elmore Family School of Electrical and Computer Engineering, Purdue University, West Lafayette, Indiana, USA {ysanghvi,chi14,stanchan}@purdue.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 253, + 455, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 253, + 455, + 374 + ], + "spans": [ + { + "bbox": [ + 160, + 253, + 455, + 374 + ], + "type": "text", + "content": "Abstract. Blind deconvolution problems are severely ill-posed because neither the underlying signal nor the forward operator are not known exactly. Conventionally, these problems are solved by alternating between estimation of the image and kernel while keeping the other fixed. In this paper, we show that this framework is flawed because of its tendency to get trapped in local minima and, instead, suggest the use of a kernel estimation strategy with a non-blind solver. This framework is employed by a diffusion method which is trained to sample the blur kernel from the conditional distribution with guidance from a pre-trained non-blind solver. The proposed diffusion method leads to state-of-the-art results on both synthetic and real blur datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 384, + 429, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 384, + 429, + 396 + ], + "spans": [ + { + "bbox": [ + 160, + 384, + 429, + 396 + ], + "type": "text", + "content": "Keywords: Blind Deconvolution " + }, + { + "bbox": [ + 160, + 384, + 429, + 396 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 384, + 429, + 396 + ], + "type": "text", + "content": " Diffusion Models " + }, + { + "bbox": [ + 160, + 384, + 429, + 396 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 384, + 429, + 396 + ], + "type": "text", + "content": " Deblurring" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 417, + 230, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 417, + 230, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 417, + 230, + 430 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 441, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 478 + ], + "type": "text", + "content": "Images taken by a hand-held camera under long exposure suffer from motion blur. Assuming this blur is spatially invariant, the captured image " + }, + { + "bbox": [ + 130, + 441, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 441, + 482, + 478 + ], + "type": "text", + "content": " and the underlying sharp image " + }, + { + "bbox": [ + 130, + 441, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 441, + 482, + 478 + ], + "type": "text", + "content": " are related as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 273, + 488, + 481, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 488, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 273, + 488, + 481, + 500 + ], + "type": "interline_equation", + "content": "\\mathbf {y} = \\mathbf {k} * \\mathbf {x} + \\mathbf {n}, \\tag {1}", + "image_path": "3f9eb9e65d27a3d9db899354a04b4d3464b61dad60d229243b0c4919964e3239.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 556 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 510, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{k}" + }, + { + "bbox": [ + 130, + 510, + 482, + 556 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 510, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 510, + 482, + 556 + ], + "type": "text", + "content": " refer to the blur kernel corresponding to the camera motion and noise, respectively, and the operator " + }, + { + "bbox": [ + 130, + 510, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\ast" + }, + { + "bbox": [ + 130, + 510, + 482, + 556 + ], + "type": "text", + "content": " is convolution. The inverse problem of estimating " + }, + { + "bbox": [ + 130, + 510, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 510, + 482, + 556 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 130, + 510, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 510, + 482, + 556 + ], + "type": "text", + "content": " is called deconvolution, which can be classified into non-blind and blind." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "text", + "content": "In the non-blind case, the blur kernel " + }, + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{k}" + }, + { + "bbox": [ + 130, + 558, + 482, + 641 + ], + "type": "text", + "content": " is known. There are plenty of iterative methods which have been successfully applied to various image restoration tasks. Usually, these approaches, such as Plug-and-Play [1,8,34,57,70,73] and Regularization by Denoising (RED) [20,53,56], use a pre-trained image denoiser as a substitute for a complex image prior. Recent deep learning counterparts to these approaches involve unrolling the iterative methods followed by end-to-end training [45,48,82,86]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "In blind deconvolution, the blur kernel " + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{k}" + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": " is not known. A natural strategy for the blind inverse problem is to jointly estimate the blur kernel " + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{k}" + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": ", and the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 351, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 351, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 351, + 129 + ], + "type": "text", + "content": "image " + }, + { + "bbox": [ + 132, + 116, + 351, + 129 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 132, + 116, + 351, + 129 + ], + "type": "text", + "content": " by maximizing the posterior probability:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 195, + 137, + 481, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 137, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 195, + 137, + 481, + 184 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\mathbf {x}}, \\hat {\\mathbf {k}} = \\underset {\\mathbf {x}, \\mathbf {k}} {\\arg \\max} \\left[ p (\\mathbf {x}, \\mathbf {k} | \\mathbf {y}) \\right] \\\\ = \\underset {\\mathbf {x}, \\mathbf {k}} {\\arg \\max } \\left[ \\log p (\\mathbf {y} | \\mathbf {k}, \\mathbf {x}) + \\log p (\\mathbf {x}) + \\log p (\\mathbf {k}) \\right]. \\tag {2} \\\\ \\end{array}", + "image_path": "1d1c3b81e45e8723a856db48f917e579c77b5cc14c841b203b82c53fd2a85e8f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 194, + 481, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 194, + 481, + 229 + ], + "spans": [ + { + "bbox": [ + 131, + 194, + 481, + 229 + ], + "type": "text", + "content": "The Maximum-A-Posteriori (MAP) for the pair " + }, + { + "bbox": [ + 131, + 194, + 481, + 229 + ], + "type": "inline_equation", + "content": "(\\mathbf{x},\\mathbf{k})" + }, + { + "bbox": [ + 131, + 194, + 481, + 229 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 131, + 194, + 481, + 229 + ], + "type": "inline_equation", + "content": "\\mathrm{MAP}_{\\mathbf{x},\\mathbf{k}}" + }, + { + "bbox": [ + 131, + 194, + 481, + 229 + ], + "type": "text", + "content": ", can be obtained by an iterative scheme which alternates between the following minimization steps:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 215, + 238, + 481, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 238, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 215, + 238, + 481, + 258 + ], + "type": "interline_equation", + "content": "\\mathbf {x} ^ {i + 1} = \\arg \\min _ {\\mathbf {x}} \\left[ \\| \\mathbf {y} - \\mathbf {x} * \\mathbf {k} ^ {i} \\| ^ {2} + \\phi (\\mathbf {x}) \\right], \\tag {3}", + "image_path": "11b82e0b2e16d7752fca5d7429b4dec7fa51a4ba508ff2831da028bef3bf29c7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 215, + 260, + 481, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 260, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 215, + 260, + 481, + 281 + ], + "type": "interline_equation", + "content": "\\mathbf {k} ^ {i + 1} = \\arg \\operatorname * {m i n} _ {\\mathbf {k}} \\left[ \\| \\mathbf {y} - \\mathbf {x} ^ {i + 1} * \\mathbf {\\Phi} \\mathbf {k} \\| ^ {2} + \\psi (\\mathbf {k}) \\right], \\tag {4}", + "image_path": "c5c27d3b145ed28b941a6634a804ab26f9fa8fb3148e4f597abd5c232b8ddcde.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 289, + 479, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 479, + 347 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 479, + 347 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 289, + 479, + 347 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot), \\psi(\\cdot)" + }, + { + "bbox": [ + 131, + 289, + 479, + 347 + ], + "type": "text", + "content": " represent the image and kernel priors. Many classical deconvolution techniques can be framed as variations of this framework [2,9,63,80]. Even neural network based iterative methods such as Self-Deblur [54] and diffusion models [15,49] have achieved state-of-the-art performance in blind deconvolution using the stated framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 349, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 349, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 131, + 349, + 482, + 434 + ], + "type": "text", + "content": "Despite recent success of diffusion methods, they are based on " + }, + { + "bbox": [ + 131, + 349, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathrm{MAP}_{\\mathbf{x},\\mathbf{k}}" + }, + { + "bbox": [ + 131, + 349, + 482, + 434 + ], + "type": "text", + "content": " approach, which was first brought into question by Levin et al. [40] They argued that the number of variables to be estimated, i.e., " + }, + { + "bbox": [ + 131, + 349, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x},\\mathbf{k}\\}" + }, + { + "bbox": [ + 131, + 349, + 482, + 434 + ], + "type": "text", + "content": ", are much larger than the number of known variables, i.e., " + }, + { + "bbox": [ + 131, + 349, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 131, + 349, + 482, + 434 + ], + "type": "text", + "content": ", and as a result, (2) favors the no-blur solution " + }, + { + "bbox": [ + 131, + 349, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}},\\hat{\\mathbf{k}} = (\\mathbf{y},\\mathbf{I})" + }, + { + "bbox": [ + 131, + 349, + 482, + 434 + ], + "type": "text", + "content": ". [40] proposed an alternate strategy of solving the kernel " + }, + { + "bbox": [ + 131, + 349, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{k}" + }, + { + "bbox": [ + 131, + 349, + 482, + 434 + ], + "type": "text", + "content": " first to exploit the inherent symmetry in the problem. This is achieved by marginalizing out the image space as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 200, + 442, + 481, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 442, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 200, + 442, + 481, + 466 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {k}} = \\underset {\\mathbf {k}} {\\arg \\max } \\left[ p (\\mathbf {k} | \\mathbf {y}) \\right] = \\underset {\\mathbf {k}} {\\arg \\max } \\left[ \\int p (\\mathbf {k}, \\mathbf {x} | \\mathbf {y}) d \\mathbf {x} \\right]. \\tag {5}", + "image_path": "c014d4df2ddbf1aab6767bbf394975d68c8546e31d8b9dd5319f867fb0ed65b0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 473, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 473, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 131, + 473, + 482, + 521 + ], + "type": "text", + "content": "Although " + }, + { + "bbox": [ + 131, + 473, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathrm{MAP_k}" + }, + { + "bbox": [ + 131, + 473, + 482, + 521 + ], + "type": "text", + "content": " optimization can lead to significantly improved reconstruction (as shown in Figure 1(c)) it has not been adopted in blind deconvolution solutions, which we surmise because of the intrinsic computational difficulty of the integration in (5)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 538, + 231, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 538, + 231, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 538, + 231, + 550 + ], + "type": "text", + "content": "1.1 Contributions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 558, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we address the problem with a kernel estimation approach, and we demonstrate that this strategy can be implemented in practice. Our contribution in this paper can be summarized as follows: We propose a novel diffusion method, called Kernel-Diffusion (Kernel-Diff) which solves the " + }, + { + "bbox": [ + 131, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{MAP_k}" + }, + { + "bbox": [ + 131, + 558, + 482, + 666 + ], + "type": "text", + "content": " optimization problem by sampling from the conditional distribution " + }, + { + "bbox": [ + 131, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "p(\\mathbf{k}|\\mathbf{y})" + }, + { + "bbox": [ + 131, + 558, + 482, + 666 + ], + "type": "text", + "content": ". This diffusion method coupled with a differentiable non-blind solver achieves state-of-the-art deconvolution performance. We show with numerical experiments on real and synthetic data that the proposed approach creates significant performance improvement. We will soon release the source code and model weights for Kernel-Diff." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "type": "text", + "content": "Y. Sanghvi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 113, + 219, + 199 + ], + "blocks": [ + { + "bbox": [ + 134, + 113, + 219, + 199 + ], + "lines": [ + { + "bbox": [ + 134, + 113, + 219, + 199 + ], + "spans": [ + { + "bbox": [ + 134, + 113, + 219, + 199 + ], + "type": "image", + "image_path": "6bd8424048f4acdff77721ad1edb5dd90169c5ff1dce62f4fed01bdee083e1a4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 139, + 201, + 212, + 224 + ], + "lines": [ + { + "bbox": [ + 139, + 201, + 212, + 224 + ], + "spans": [ + { + "bbox": [ + 139, + 201, + 212, + 224 + ], + "type": "text", + "content": "(a) Blurred image Inset: true kernel" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 220, + 113, + 306, + 198 + ], + "blocks": [ + { + "bbox": [ + 220, + 113, + 306, + 198 + ], + "lines": [ + { + "bbox": [ + 220, + 113, + 306, + 198 + ], + "spans": [ + { + "bbox": [ + 220, + 113, + 306, + 198 + ], + "type": "image", + "image_path": "59cecb71a103eb3e0bc94f99fd9b97c891c0087b1445c9b8fd681fdb65aa1f3a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 201, + 305, + 236 + ], + "lines": [ + { + "bbox": [ + 220, + 201, + 305, + 236 + ], + "spans": [ + { + "bbox": [ + 220, + 201, + 305, + 236 + ], + "type": "text", + "content": "(b) Blind-DPS [15] based on alternating minimization" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 393, + 198 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 393, + 198 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 393, + 198 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 393, + 198 + ], + "type": "image", + "image_path": "dd7f9059013ace52f84651fad535dd383568eb78988b192323343882a47889d9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 201, + 388, + 236 + ], + "lines": [ + { + "bbox": [ + 316, + 201, + 388, + 236 + ], + "spans": [ + { + "bbox": [ + 316, + 201, + 388, + 236 + ], + "type": "text", + "content": "(c) Kernel-Diff (ours) based on kernel estimation" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 394, + 114, + 479, + 199 + ], + "blocks": [ + { + "bbox": [ + 394, + 114, + 479, + 199 + ], + "lines": [ + { + "bbox": [ + 394, + 114, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 394, + 114, + 479, + 199 + ], + "type": "image", + "image_path": "1adf6c343a729100a94d8573902c4351359370b2ac3cb917221e3d5fff3c3671.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 201, + 460, + 224 + ], + "lines": [ + { + "bbox": [ + 413, + 201, + 460, + 224 + ], + "spans": [ + { + "bbox": [ + 413, + 201, + 460, + 224 + ], + "type": "text", + "content": "(d) Ground Truth" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 247, + 480, + 301 + ], + "lines": [ + { + "bbox": [ + 131, + 247, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 131, + 247, + 480, + 301 + ], + "type": "text", + "content": "Fig. 1: In this paper, we present a novel diffusion-based approach to the blind deconvolution problem. The conventional approaches, such as Blind-DPS, iterate between between kernel and image space minimization. Our approach uses a kernel-first estimation strategy in the context of diffusion models to sample kernels from the conditional distribution " + }, + { + "bbox": [ + 131, + 247, + 480, + 301 + ], + "type": "inline_equation", + "content": "p(\\mathbf{k}|\\mathbf{y})" + }, + { + "bbox": [ + 131, + 247, + 480, + 301 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 225, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 225, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 225, + 339 + ], + "type": "text", + "content": "2 Background" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 353, + 253, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 253, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 253, + 364 + ], + "type": "text", + "content": "2.1 Diffusion Methods" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 374, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 374, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 131, + 374, + 480, + 422 + ], + "type": "text", + "content": "Diffusion models [29, 64, 67, 74] are a popular generative method which can be used to sample any target diffusion " + }, + { + "bbox": [ + 131, + 374, + 480, + 422 + ], + "type": "inline_equation", + "content": "p_{data}(\\mathbf{x})" + }, + { + "bbox": [ + 131, + 374, + 480, + 422 + ], + "type": "text", + "content": ". The diffusion models begin with a forward process, which starts from " + }, + { + "bbox": [ + 131, + 374, + 480, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 \\sim p_{data}(\\mathbf{x})" + }, + { + "bbox": [ + 131, + 374, + 480, + 422 + ], + "type": "text", + "content": " and gradually adds noise according to variance schedule " + }, + { + "bbox": [ + 131, + 374, + 480, + 422 + ], + "type": "inline_equation", + "content": "\\{\\beta_1, \\beta_2, \\ldots, \\beta_T\\}" + }, + { + "bbox": [ + 131, + 374, + 480, + 422 + ], + "type": "text", + "content": ", formulated as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 247, + 433, + 480, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 433, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 247, + 433, + 480, + 447 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t} \\sim \\mathcal {N} \\left(\\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right) \\tag {6}", + "image_path": "b15b816a0251a4c98638b93d16a19668460be9491e8daaeaf48b98ea4b7a838c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 457, + 480, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 457, + 480, + 494 + ], + "spans": [ + { + "bbox": [ + 131, + 457, + 480, + 494 + ], + "type": "text", + "content": "at time steps " + }, + { + "bbox": [ + 131, + 457, + 480, + 494 + ], + "type": "inline_equation", + "content": "t \\in \\{1, 2, \\dots, T\\}" + }, + { + "bbox": [ + 131, + 457, + 480, + 494 + ], + "type": "text", + "content": ". This forward process converges to " + }, + { + "bbox": [ + 131, + 457, + 480, + 494 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 131, + 457, + 480, + 494 + ], + "type": "text", + "content": ". Next, the reverse diffusion process is used to sample from the distribution of " + }, + { + "bbox": [ + 131, + 457, + 480, + 494 + ], + "type": "inline_equation", + "content": "p_{data}" + }, + { + "bbox": [ + 131, + 457, + 480, + 494 + ], + "type": "text", + "content": " by starting with " + }, + { + "bbox": [ + 131, + 457, + 480, + 494 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 131, + 457, + 480, + 494 + ], + "type": "text", + "content": " and performing the iterative updates" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 205, + 502, + 480, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 502, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 205, + 502, + 480, + 544 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {x} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {x} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta^ {*}} \\left(\\mathbf {x} _ {t}, t\\right)\\right) + \\sigma_ {t} \\mathbf {z} _ {t}, \\tag {7} \\\\ \\mathbf {z} _ {t} \\sim \\mathcal {N} (0, \\mathbf {I}), \\\\ \\end{array}", + "image_path": "9ef80254912999ce2903622b1e0bb18de4731ef033284e7126908e23d6d6a768.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 556, + 479, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 556, + 479, + 602 + ], + "spans": [ + { + "bbox": [ + 131, + 556, + 479, + 602 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 556, + 479, + 602 + ], + "type": "inline_equation", + "content": "\\alpha_{t} \\stackrel{\\mathrm{def}}{=} 1 - \\beta_{t}" + }, + { + "bbox": [ + 131, + 556, + 479, + 602 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 556, + 479, + 602 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t} \\stackrel{\\mathrm{def}}{=} \\prod_{i=1}^{t} \\alpha_{i}" + }, + { + "bbox": [ + 131, + 556, + 479, + 602 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 556, + 479, + 602 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta^{*}}(\\mathbf{x}_{t}, t)" + }, + { + "bbox": [ + 131, + 556, + 479, + 602 + ], + "type": "text", + "content": " is the trained diffusion network used to learn to denoise " + }, + { + "bbox": [ + 131, + 556, + 479, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t}" + }, + { + "bbox": [ + 131, + 556, + 479, + 602 + ], + "type": "text", + "content": " at any given time step, trained by minimizing the following loss function." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 200, + 612, + 480, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 612, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 200, + 612, + 480, + 632 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\theta) \\stackrel {\\text {d e f}} {=} \\mathbb {E} _ {t, \\mathbf {x} _ {0}, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t\\right) \\| ^ {2} \\right]. \\tag {8}", + "image_path": "e6bff523caa1a7360dffe51e66227778372915e00e1f3350f3d5b40d202384a1.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "Inverse Problems and Diffusion Models. Score based diffusion methods have been extensively used in solving inverse problems [17, 21, 33, 35-37, 89] by" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "text", + "content": "Kernel Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "sampling from the posterior " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}|\\mathbf{y})" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " and the corresponding score is calculated using the Bayes' formula" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 200, + 148, + 481, + 160 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 148, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 200, + 148, + 481, + 160 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t} | \\mathbf {y}) = \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t}) + \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {y} | \\mathbf {x} _ {t}). \\tag {9}", + "image_path": "18388bc2dc900b19275bfb453e170dd0ca1330c1a5f0badcd5101558bc6c6ca2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 166, + 479, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 166, + 479, + 249 + ], + "spans": [ + { + "bbox": [ + 130, + 166, + 479, + 249 + ], + "type": "text", + "content": "For example, [31] uses an approximation to the posterior score in (9) and used the Langevin dynamics to sample from the distribution " + }, + { + "bbox": [ + 130, + 166, + 479, + 249 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}|\\mathbf{y})" + }, + { + "bbox": [ + 130, + 166, + 479, + 249 + ], + "type": "text", + "content": " for the MRI compressed sensing problem; [18,19] solve the inpainting and super-resolution inverse problems with a modified reverse diffusion scheme which includes a projection step to reduce the consistency metric " + }, + { + "bbox": [ + 130, + 166, + 479, + 249 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{y} - \\mathbf{A}\\mathbf{x}\\|" + }, + { + "bbox": [ + 130, + 166, + 479, + 249 + ], + "type": "text", + "content": " in each iteration; diffusion Posterior Sampling (DPS) [16] solves the noisy inverse problem with alternating diffusion and gradient descent steps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 250, + 479, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 250, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 479, + 308 + ], + "type": "text", + "content": "Image Restoration and Diffusion Models Diffusion models sample from a target distribution while MMSE estimators average over them [37]. As a result, the former has been widely used in image restoration tasks such as deblurring [78], super-resolution [41, 58], and low-light enhancement [32, 76]. For an exhaustive review, we refer the reader to [42]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 309, + 480, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 480, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 480, + 404 + ], + "type": "text", + "content": "Guidance in Diffusion Models. Pre-trained diffusion models are often guided at each iteration to produce desired results. This is first proposed in [22] where a classifier gradient is added to each iteration for the final output to belong to the given class. ILVR [14] guides the image diffusion process by constraining the low-frequency information of an image. In the same spirit, [75] guides the diffusion process by refining only the null-space of the corresponding inverse problem. [78] guides a deblurring diffusion model using a pretrained structure module." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 420, + 263, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 263, + 431 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 263, + 431 + ], + "type": "text", + "content": "2.2 Blind Deconvolution" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 437, + 479, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 479, + 533 + ], + "type": "text", + "content": "Traditional blind deconvolution methods can be categorized under the " + }, + { + "bbox": [ + 130, + 437, + 479, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{MAP}_{\\mathbf{x},\\mathbf{k}}" + }, + { + "bbox": [ + 130, + 437, + 479, + 533 + ], + "type": "text", + "content": " framework [2,9,10,44,52,63,69,80,81]. For example, [13] uses salient edges for the kernel estimation and then performs image estimation using a Total Variation (TV) prior. [79] refines the kernel estimation process by iterative support detection. [28] performs alternating image and kernel minimization but uses a motion density representation for the blur kernel instead of for pixels. Self-Deblur [54] uses an alternating minimization scheme where a latent deep image prior [72] is used to represent the kernel and image space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 534, + 479, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 479, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 479, + 593 + ], + "type": "text", + "content": "Deep learning methods have replaced the optimization-based approaches and usually involve end-to-end training of image restoration networks [38, 62, 68, 71, 77, 83-85] using large datasets of pairs of sharp and blurry images, e.g., GoPro [50] and RealBlur [55] datasets. For a more exhaustive review of deep learning based approaches, we refer the readers to [87]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "content": "A few methods have utilized the " + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{MAP_k}" + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "content": " approach, even though not always explicitly stated by the authors. [7] uses a fully convolutional network to predict the Fourier coefficients of the underlying kernel which are used to deconvolve the blurry image in Fourier space. [5,6] train a Kernel-Prediction Network for all pixels in the image to deal with spatially varying blur. [60] leverages the differentiability of a non-blind solver [59] to frame the blind deconvolution problem" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "type": "text", + "content": "Y. Sanghvi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 152 + ], + "type": "text", + "content": "entirely in terms of kernel estimation problems. This strategy is improved by [61] using a key-points based representation to further reduce the dimensionality of the problem." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 169, + 202, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 169, + 202, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 169, + 202, + 182 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 194, + 479, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 194, + 479, + 241 + ], + "spans": [ + { + "bbox": [ + 131, + 194, + 479, + 241 + ], + "type": "text", + "content": "Blind-DPS [15] proposed a diffusion model based scheme for solving the blind inverse problem. Using separate diffusion models to sample from the image and kernel prior, a new scheme was proposed to sample solutions from the joint conditional distribution" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 273, + 251, + 481, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 251, + 481, + 264 + ], + "spans": [ + { + "bbox": [ + 273, + 251, + 481, + 264 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}}, \\hat {\\mathbf {k}} \\sim p (\\mathbf {x}, \\mathbf {k} | \\mathbf {y}) \\tag {10}", + "image_path": "6c17e9c284ec7c9c381c4caaa615c661aaa10aea00e93914d81365940a615ba0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 271, + 479, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 271, + 479, + 307 + ], + "spans": [ + { + "bbox": [ + 131, + 271, + 479, + 307 + ], + "type": "text", + "content": "This method is similar in principle to the conventional blind deconvolution techniques where the goal is to estimate the maximum of the conditional distribution i.e." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 203, + 313, + 481, + 362 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 313, + 481, + 362 + ], + "spans": [ + { + "bbox": [ + 203, + 313, + 481, + 362 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\mathbf {x}}, \\hat {\\mathbf {k}} = \\underset {\\mathbf {x}, \\mathbf {k}} {\\arg \\max } \\left[ p (\\mathbf {x}, \\mathbf {k} | \\mathbf {y}) \\right] (11) \\\\ = \\underset {\\mathbf {x}, \\mathbf {k}} {\\arg \\min } \\left[ \\left\\| \\mathbf {y} - \\mathbf {k} * \\mathbf {x} \\right\\| ^ {2} + R _ {\\mathbf {x}} (\\mathbf {x}) + R _ {\\mathbf {k}} (\\mathbf {k}) \\right] (12) \\\\ \\end{array}", + "image_path": "68cfd5dc82ec2e5a87ab51416a5811fe8936e025564844e35dfe845435cd9a8b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 370, + 479, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 370, + 479, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 370, + 479, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 370, + 479, + 407 + ], + "type": "inline_equation", + "content": "R_{\\mathbf{x}}(.)" + }, + { + "bbox": [ + 131, + 370, + 479, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 370, + 479, + 407 + ], + "type": "inline_equation", + "content": "R_{\\mathbf{k}}(.)" + }, + { + "bbox": [ + 131, + 370, + 479, + 407 + ], + "type": "text", + "content": " represent the image and kernel priors respectively. This minimization is achieved using alternating minimization techniques where the kernel is estimated while keeping the image estimate fixed and vice versa." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 407, + 479, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 407, + 479, + 477 + ], + "spans": [ + { + "bbox": [ + 131, + 407, + 479, + 477 + ], + "type": "text", + "content": "Recently, Blind-DPS proposed a diffusion method using a similar framework. In this approach, after each parallel reverse diffusion step from the image and kernel models, the estimates are refined using single gradient descent on " + }, + { + "bbox": [ + 131, + 407, + 479, + 477 + ], + "type": "inline_equation", + "content": "||\\mathbf{y} - \\mathbf{x} * \\mathbf{k}||^2" + }, + { + "bbox": [ + 131, + 407, + 479, + 477 + ], + "type": "text", + "content": ". Specifically, after each time step " + }, + { + "bbox": [ + 131, + 407, + 479, + 477 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 407, + 479, + 477 + ], + "type": "text", + "content": " of parallel reverse diffusion on kernel and image, i.e. " + }, + { + "bbox": [ + 131, + 407, + 479, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_t" + }, + { + "bbox": [ + 131, + 407, + 479, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 407, + 479, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 131, + 407, + 479, + 477 + ], + "type": "text", + "content": ", temporary estimates of the kernel and image are obtained as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 233, + 484, + 481, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 484, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 233, + 484, + 481, + 509 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {0} \\leftarrow \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left[ \\mathbf {x} _ {t} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\hat {s} _ {\\mathbf {x}} (\\mathbf {x} _ {t}, t) \\right] \\tag {13}", + "image_path": "d50e745139d0dcd067fda6304fc1bdf422f00868b3ef40da86473d4429e66425.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 233, + 511, + 481, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 511, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 233, + 511, + 481, + 536 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {k}} _ {0} \\leftarrow \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left[ \\mathbf {k} _ {t} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\hat {s} _ {\\mathbf {k}} (\\mathbf {k} _ {t}, t) \\right] \\tag {14}", + "image_path": "c24d97b86e36972b1506f026a0223e7c617451bcb3a8569bc2af6dec449b5318.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 542, + 479, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 542, + 479, + 567 + ], + "spans": [ + { + "bbox": [ + 131, + 542, + 479, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 542, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\hat{s}_{\\mathbf{x}}, \\hat{s}_{\\mathbf{k}}" + }, + { + "bbox": [ + 131, + 542, + 479, + 567 + ], + "type": "text", + "content": " represent the corresponding score functions. This is followed by a gradient descent step on the reblurring loss for each update step" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 269, + 574, + 481, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 574, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 269, + 574, + 481, + 588 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {t} ^ {\\prime} \\leftarrow \\mathbf {x} _ {t} - \\delta \\nabla_ {\\mathbf {x} _ {t}} \\| \\mathbf {y} - \\hat {\\mathbf {k}} _ {0} * \\hat {\\mathbf {x}} _ {0} \\| ^ {2} \\tag {15}", + "image_path": "bad3de2bf348b2866df1e328408a5d7d910c9744a84b65ff089d5221ac6475e2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 211, + 590, + 481, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 590, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 211, + 590, + 481, + 611 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {k}} _ {t} ^ {\\prime} \\leftarrow \\mathbf {k} _ {t} - \\delta \\nabla_ {\\mathbf {k} _ {t}} \\left[ \\| \\mathbf {y} - \\hat {\\mathbf {k}} _ {0} * \\hat {\\mathbf {x}} _ {0} \\| ^ {2} + \\lambda R _ {\\mathbf {k}} (\\hat {\\mathbf {k}} _ {0}) \\right] \\tag {16}", + "image_path": "1fa085ccb15fcbbbdecd0231f5e035e0bf9d9757cbf1e538fcc4122aefaf71cf.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 617, + 479, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 479, + 641 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 479, + 641 + ], + "type": "inline_equation", + "content": "R_{\\mathbf{k}}(.)" + }, + { + "bbox": [ + 131, + 617, + 479, + 641 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 131, + 617, + 479, + 641 + ], + "type": "inline_equation", + "content": "\\ell_0 / \\ell_1" + }, + { + "bbox": [ + 131, + 617, + 479, + 641 + ], + "type": "text", + "content": " regularization on the kernel to impose sparsity and " + }, + { + "bbox": [ + 131, + 617, + 479, + 641 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 131, + 617, + 479, + 641 + ], + "type": "text", + "content": " represents the step size change." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "While this alternating minimization approach to blind deconvolution is prevalent, this framework was first put into question by Levin et. al [40]. They showed" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 101 + ], + "type": "text", + "content": "Kernel Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "that the conventional blind deconvolution techniques in this framework tended to converge to a no-blur degenerate solution, because the number of variables to be estimated i.e. " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x},\\mathbf{k}\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " exceeded the number of knowns i.e. " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": ". Instead they recommended marginalization of the image space and first estimating the kernel using the conditional dsiitribution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 201, + 185, + 481, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 185, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 201, + 185, + 481, + 209 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {k}} = \\underset {\\mathbf {k}} {\\arg \\max } \\left[ p (\\mathbf {k} | \\mathbf {y}) \\right] = \\underset {\\mathbf {k}} {\\arg \\max } \\left[ \\int p (\\mathbf {k}, \\mathbf {x} | \\mathbf {y}) d \\mathbf {x} \\right] \\tag {17}", + "image_path": "b0b72c13fd8d5da93cd81d38b0822cf3fccbe691a5379289177bc4e5d2125246.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 216, + 321, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 321, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 321, + 228 + ], + "type": "text", + "content": "followed by estimation of image as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 256, + 236, + 481, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 236, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 256, + 236, + 481, + 258 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} = \\underset {\\mathbf {x}} {\\arg \\max } \\left[ p (\\mathbf {x} | \\mathbf {y}, \\hat {\\mathbf {k}}) \\right] \\tag {18}", + "image_path": "cbc2979fa3eac2e5b050d21fb95abd4de8638e88a70f57f01e178092cf1db27c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 265, + 480, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 480, + 326 + ], + "type": "text", + "content": "Since the number of unknowns is fewer than the number of knowns in the first sub-problem in (17), this leads to a better kernel estimate and hence, a better image estimate. Applying this principle to deconvolution with diffusion models, we aim to propose a method which samples from the conditional distribution " + }, + { + "bbox": [ + 130, + 265, + 480, + 326 + ], + "type": "inline_equation", + "content": "p(\\mathbf{k}|\\mathbf{y})" + }, + { + "bbox": [ + 130, + 265, + 480, + 326 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 130, + 265, + 480, + 326 + ], + "type": "inline_equation", + "content": "p(\\mathbf{k},\\mathbf{x}|\\mathbf{y})" + }, + { + "bbox": [ + 130, + 265, + 480, + 326 + ], + "type": "text", + "content": " since the former provides more stable solutions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 342, + 320, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 320, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 320, + 355 + ], + "type": "text", + "content": "3.1 Marginalization of Image Space" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 361, + 482, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 432 + ], + "type": "text", + "content": "In practice, the estimation of " + }, + { + "bbox": [ + 130, + 361, + 482, + 432 + ], + "type": "inline_equation", + "content": "p(\\mathbf{k}|\\mathbf{y})" + }, + { + "bbox": [ + 130, + 361, + 482, + 432 + ], + "type": "text", + "content": " can be difficult because the marginalization in (5) requires a weighted average over all images. We surmise that due to lack of a feasible method to evaluate the likelihood, the approach from [40] was not explored by the imaging community for the blind deconvolution. To mitigate this issue, we first discuss estimation of the conditional distribution " + }, + { + "bbox": [ + 130, + 361, + 482, + 432 + ], + "type": "inline_equation", + "content": "p(\\mathbf{k}|\\mathbf{y})" + }, + { + "bbox": [ + 130, + 361, + 482, + 432 + ], + "type": "text", + "content": " using a non-blind solver." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 434, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 469 + ], + "type": "text", + "content": "First, we explicitly define a non-blind solver. Given a blurred image " + }, + { + "bbox": [ + 130, + 434, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 434, + 482, + 469 + ], + "type": "text", + "content": " and the corresponding blur kernel " + }, + { + "bbox": [ + 130, + 434, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{k}" + }, + { + "bbox": [ + 130, + 434, + 482, + 469 + ], + "type": "text", + "content": ", we can easily estimate the latent image using a non-blind solver which performs the following image estimation:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 223, + 477, + 481, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 477, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 223, + 477, + 481, + 498 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} (\\mathbf {k}) = \\arg \\max _ {\\mathbf {x}} \\left[ p (\\mathbf {x} | \\mathbf {k}, \\mathbf {y}) \\right] \\stackrel {\\mathrm {d e f}} {=} F (\\mathbf {y}, \\mathbf {k}). \\tag {19}", + "image_path": "b3e54ebb8aad4367555da7b8f7989b70f0493f1534e20fc6b4b58ca62ebeddf0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 504, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 563 + ], + "type": "text", + "content": "Traditionally, these non-blind solvers are implemented using an iterative method where the corresponding cost function for MAP estimation in (19) is minimized using gradient descent or proximal updates. Convolutional neural networks with Wiener-deconvolution processing have been used with great success also [12, 23, 24, 51, 59]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 564, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 482, + 589 + ], + "type": "text", + "content": "Remark 1: Given a non-blind solver " + }, + { + "bbox": [ + 130, + 564, + 482, + 589 + ], + "type": "inline_equation", + "content": "F_{\\phi}(.)" + }, + { + "bbox": [ + 130, + 564, + 482, + 589 + ], + "type": "text", + "content": ", we can approximate the conditional probability " + }, + { + "bbox": [ + 130, + 564, + 482, + 589 + ], + "type": "inline_equation", + "content": "p(\\mathbf{k}|\\mathbf{y})" + }, + { + "bbox": [ + 130, + 564, + 482, + 589 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 203, + 597, + 481, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 597, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 203, + 597, + 481, + 622 + ], + "type": "interline_equation", + "content": "p (\\mathbf {k} | \\mathbf {y}) = \\int p (\\mathbf {x}, \\mathbf {k} | \\mathbf {y}) d \\mathbf {x} \\approx \\frac {1}{Z (\\mathbf {y})} p (\\mathbf {y} \\mid F _ {\\phi} (\\mathbf {y}, \\mathbf {k}), \\mathbf {k}) \\tag {20}", + "image_path": "ce54bc617c29a49fc493f812b552a1661cd20b9c292c7a40f75d2babab0f27b4.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "Z(\\mathbf{y})" + }, + { + "bbox": [ + 132, + 629, + 482, + 667 + ], + "type": "text", + "content": " is independent of the kernel " + }, + { + "bbox": [ + 132, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{k}" + }, + { + "bbox": [ + 132, + 629, + 482, + 667 + ], + "type": "text", + "content": ". Through the approximation, we can avoid the integration over the image space and express the likelihood " + }, + { + "bbox": [ + 132, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "p(\\mathbf{k}|\\mathbf{y})" + }, + { + "bbox": [ + 132, + 629, + 482, + 667 + ], + "type": "text", + "content": " in a tractable manner, up to the unknown the normalization constant " + }, + { + "bbox": [ + 132, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "Z(\\mathbf{y})" + }, + { + "bbox": [ + 132, + 629, + 482, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "type": "text", + "content": "Y. Sanghvi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 113, + 470, + 167 + ], + "blocks": [ + { + "bbox": [ + 143, + 113, + 470, + 167 + ], + "lines": [ + { + "bbox": [ + 143, + 113, + 470, + 167 + ], + "spans": [ + { + "bbox": [ + 143, + 113, + 470, + 167 + ], + "type": "image", + "image_path": "ac695fc85bf3d2f5a1e5fb782905aae5781bcf81bf8f0b0046504fa8a32e6de2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 140, + 169, + 474, + 301 + ], + "blocks": [ + { + "bbox": [ + 140, + 169, + 474, + 301 + ], + "lines": [ + { + "bbox": [ + 140, + 169, + 474, + 301 + ], + "spans": [ + { + "bbox": [ + 140, + 169, + 474, + 301 + ], + "type": "image", + "image_path": "58303bbcaccb426ca74ae88de8380d40f78f2535622e4a36af6b5dbf48020e20.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 310, + 482, + 366 + ], + "lines": [ + { + "bbox": [ + 130, + 310, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 482, + 366 + ], + "type": "text", + "content": "Fig. 2: Proposed Kernel-Diff. Each iteration of the proposed kernel-based diffusion contains two sub-steps. First, we perform a reverse diffusion step using the trained diffusion model " + }, + { + "bbox": [ + 130, + 310, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{k}_t,\\mathbf{y},t)" + }, + { + "bbox": [ + 130, + 310, + 482, + 366 + ], + "type": "text", + "content": ". Next we plug the temporary kernel estimate " + }, + { + "bbox": [ + 130, + 310, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{k}}_0" + }, + { + "bbox": [ + 130, + 310, + 482, + 366 + ], + "type": "text", + "content": " in the differentiable non-blind solver " + }, + { + "bbox": [ + 130, + 310, + 482, + 366 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 130, + 310, + 482, + 366 + ], + "type": "text", + "content": " and perform a gradient descent update to guide the diffusion process to minimizing the reburring loss " + }, + { + "bbox": [ + 130, + 310, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{y} - \\hat{\\mathbf{k}}_0\\circledast \\hat{\\mathbf{x}}_0\\|^2" + }, + { + "bbox": [ + 130, + 310, + 482, + 366 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 388, + 277, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 388, + 277, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 388, + 277, + 399 + ], + "type": "text", + "content": "3.2 Kernel-Based Diffusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 407, + 482, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 466 + ], + "type": "text", + "content": "To sample from " + }, + { + "bbox": [ + 130, + 407, + 482, + 466 + ], + "type": "inline_equation", + "content": "p(\\mathbf{k}|\\mathbf{y})" + }, + { + "bbox": [ + 130, + 407, + 482, + 466 + ], + "type": "text", + "content": ", we propose to use a kernel-based diffusion network. This is different from existing methods using diffusion models, e.g., Blind-DPS [15], that resemble the alternating minimization approach, aiming to sample " + }, + { + "bbox": [ + 130, + 407, + 482, + 466 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x},\\mathbf{k}\\}" + }, + { + "bbox": [ + 130, + 407, + 482, + 466 + ], + "type": "text", + "content": " from the conditional distribution " + }, + { + "bbox": [ + 130, + 407, + 482, + 466 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x},\\mathbf{k}|\\mathbf{y})" + }, + { + "bbox": [ + 130, + 407, + 482, + 466 + ], + "type": "text", + "content": " using implicit score models for image and kernel spaces." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 467, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 482, + 586 + ], + "type": "text", + "content": "Diffusion Network. We use a modified version of the U-Net architecture as opposed to the one described in [22]. Since we are sampling from a conditional distribution, the diffusion model " + }, + { + "bbox": [ + 130, + 467, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{k},\\mathbf{y},t)" + }, + { + "bbox": [ + 130, + 467, + 482, + 586 + ], + "type": "text", + "content": " takes as input both the current kernel estimate " + }, + { + "bbox": [ + 130, + 467, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_t" + }, + { + "bbox": [ + 130, + 467, + 482, + 586 + ], + "type": "text", + "content": " and the blurred image " + }, + { + "bbox": [ + 130, + 467, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 467, + 482, + 586 + ], + "type": "text", + "content": ". This is achieved in a similar fashion as [58]. Specifically, our U-Net architecture has two encoders, one for the image " + }, + { + "bbox": [ + 130, + 467, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 467, + 482, + 586 + ], + "type": "text", + "content": " and the other for the kernel " + }, + { + "bbox": [ + 130, + 467, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{k}" + }, + { + "bbox": [ + 130, + 467, + 482, + 586 + ], + "type": "text", + "content": ". In our implementation, we use images of size " + }, + { + "bbox": [ + 130, + 467, + 482, + 586 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 130, + 467, + 482, + 586 + ], + "type": "text", + "content": " and kernels of size " + }, + { + "bbox": [ + 130, + 467, + 482, + 586 + ], + "type": "inline_equation", + "content": "64\\times 64" + }, + { + "bbox": [ + 130, + 467, + 482, + 586 + ], + "type": "text", + "content": ". At the end of the encoders, the encoded features of the kernel branch are upscaled by a factor of 4 and concatenated to the encoded features of the image branch. The concatenated features are then passed through a common U-Net decoder stage to predict the kernel at the next time step." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 146, + 586, + 481, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 586, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 146, + 586, + 481, + 598 + ], + "type": "text", + "content": "The diffusion model " + }, + { + "bbox": [ + 146, + 586, + 481, + 598 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 146, + 586, + 481, + 598 + ], + "type": "text", + "content": " is trained by minimizing the following loss function:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 185, + 614, + 481, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 614, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 185, + 614, + 481, + 635 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {k e r n e l}} (\\theta) \\stackrel {\\mathrm {d e f}} {=} \\mathbb {E} _ {t, \\mathbf {k}, \\mathbf {y}, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {k} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\mathbf {y}, t\\right) \\| _ {2} ^ {2} \\right] \\tag {21}", + "image_path": "f5e2c3695af0f183182c05d75ae6303f9f5dda8c1559c5f1cff1a2b52c50e095.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "where the expectation is taken over all timesteps " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": ", blurred images " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " and corresponding blur kernels " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{k}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": ". With trained parameters " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": ", we sample " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "p(\\mathbf{k}|\\mathbf{y})" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " using" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "text", + "content": "Kernel Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "a series of reverse diffusion steps starting from " + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_T\\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": " and the following reverse-diffusion iteration:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 203, + 147, + 481, + 173 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 147, + 481, + 173 + ], + "spans": [ + { + "bbox": [ + 203, + 147, + 481, + 173 + ], + "type": "interline_equation", + "content": "\\mathbf {k} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {k} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta^ {*}} \\left(\\mathbf {k} _ {t}, \\mathbf {y}, t\\right)\\right) + \\bar {\\sigma} _ {t} \\mathbf {z} _ {t} \\tag {22}", + "image_path": "36ccd20b6bf7770ae5a7dbe40203ad263da695023c06dd32b73686829d63137e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 179, + 479, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 179, + 479, + 263 + ], + "spans": [ + { + "bbox": [ + 130, + 179, + 479, + 263 + ], + "type": "text", + "content": "Guidance using Non-Blind Solver. Once we have sampled the kernel " + }, + { + "bbox": [ + 130, + 179, + 479, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_0" + }, + { + "bbox": [ + 130, + 179, + 479, + 263 + ], + "type": "text", + "content": ", we estimate the latent image as " + }, + { + "bbox": [ + 130, + 179, + 479, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 = F_{\\phi}(\\mathbf{y},\\mathbf{k}_0)" + }, + { + "bbox": [ + 130, + 179, + 479, + 263 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 179, + 479, + 263 + ], + "type": "inline_equation", + "content": "F_{\\phi}(\\cdot)" + }, + { + "bbox": [ + 130, + 179, + 479, + 263 + ], + "type": "text", + "content": " is a neural network non-blind solver. There is a vast selection of candidate non-blind solver, including many network-based methods proposed recently [24-27, 59, 88]. These networks usually convert the blurred images to features space, apply Wiener deconvolution using the kernel followed by converting the deconvolved features into a single image using a decoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 264, + 480, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 264, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 480, + 323 + ], + "type": "text", + "content": "While diffusion processes are effective in generating samples from the conditional distribution " + }, + { + "bbox": [ + 130, + 264, + 480, + 323 + ], + "type": "inline_equation", + "content": "p(\\mathbf{k}|\\mathbf{y})" + }, + { + "bbox": [ + 130, + 264, + 480, + 323 + ], + "type": "text", + "content": ", they do not guarantee minimization of the reblurring loss, i.e., " + }, + { + "bbox": [ + 130, + 264, + 480, + 323 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{y} - \\mathbf{k} * \\mathbf{x} \\|^2" + }, + { + "bbox": [ + 130, + 264, + 480, + 323 + ], + "type": "text", + "content": ". To minimize reblurring loss while also generating a plausible sample " + }, + { + "bbox": [ + 130, + 264, + 480, + 323 + ], + "type": "inline_equation", + "content": "p(\\mathbf{k}|\\mathbf{y})" + }, + { + "bbox": [ + 130, + 264, + 480, + 323 + ], + "type": "text", + "content": ", we guide the reverse diffusion process using the pre-trained non-blind solver as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 324, + 480, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 480, + 360 + ], + "type": "text", + "content": "At step " + }, + { + "bbox": [ + 130, + 324, + 480, + 360 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 324, + 480, + 360 + ], + "type": "text", + "content": " of the reverse diffusion process, the kernel estimate is given as " + }, + { + "bbox": [ + 130, + 324, + 480, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_t" + }, + { + "bbox": [ + 130, + 324, + 480, + 360 + ], + "type": "text", + "content": ". Using the trained diffusion model " + }, + { + "bbox": [ + 130, + 324, + 480, + 360 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta^*}(\\mathbf{k}_t,\\mathbf{y},t)" + }, + { + "bbox": [ + 130, + 324, + 480, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 324, + 480, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_t" + }, + { + "bbox": [ + 130, + 324, + 480, + 360 + ], + "type": "text", + "content": ", we can compute a temporary estimate of the sampled kernel " + }, + { + "bbox": [ + 130, + 324, + 480, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_0" + }, + { + "bbox": [ + 130, + 324, + 480, + 360 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 225, + 368, + 480, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 368, + 480, + 393 + ], + "spans": [ + { + "bbox": [ + 225, + 368, + 480, + 393 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {k}} _ {0} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\mathbf {k} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta^ {*}} \\left(\\mathbf {k} _ {t}, \\mathbf {y}, t\\right)\\right). \\tag {23}", + "image_path": "d4162497663c4cffe35a78b7a7bb0b1ff9c7a16b0cc21bade157baecf12b4cdd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 402, + 479, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 479, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 479, + 429 + ], + "type": "text", + "content": "We then estimate the image using this kernel estimate " + }, + { + "bbox": [ + 130, + 402, + 479, + 429 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{k}}_0" + }, + { + "bbox": [ + 130, + 402, + 479, + 429 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 402, + 479, + 429 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0 \\stackrel{\\mathrm{def}}{=} F(\\mathbf{y}, \\hat{\\mathbf{k}}_0)" + }, + { + "bbox": [ + 130, + 402, + 479, + 429 + ], + "type": "text", + "content": ", plug into the reburring loss, and perform gradient descent by" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 225, + 448, + 480, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 448, + 480, + 462 + ], + "spans": [ + { + "bbox": [ + 225, + 448, + 480, + 462 + ], + "type": "interline_equation", + "content": "\\mathbf {k} _ {t} \\leftarrow \\mathbf {k} _ {t} - \\delta \\nabla_ {\\mathbf {k} _ {t}} \\| \\mathbf {y} - \\hat {\\mathbf {k}} _ {0} * F (\\mathbf {y}, \\hat {\\mathbf {k}} _ {0}) \\| ^ {2}. \\tag {24}", + "image_path": "f226d68f98458069dbf7c003c2e6e01f1948065626bdaf84f70f4094fff57a0b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 471, + 479, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 479, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 479, + 495 + ], + "type": "text", + "content": "This step is equivalent to maximizing log likelihood of the approximation provided in (20). Taking the log of approximation, we get," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 219, + 503, + 480, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 503, + 480, + 524 + ], + "spans": [ + { + "bbox": [ + 219, + 503, + 480, + 524 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\mathbf {k}} \\log p (\\mathbf {k} | \\mathbf {y}) \\approx \\nabla_ {\\mathbf {k}} \\left[ \\| \\mathbf {y} - F (\\mathbf {y}, \\mathbf {k}) * * \\mathbf {k} \\| _ {2} ^ {2} \\right] \\tag {25}", + "image_path": "9d4c2d3a9e9d04ebffcc33dfb937a8501c41af6b7afff87f517aa7f6d69ce0a8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 540, + 480, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 480, + 614 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 480, + 614 + ], + "type": "text", + "content": "Overall procedure. We summarize the overall procedure in Algorithm 1 and pictorially represent it in Figure 2. At any given iteration " + }, + { + "bbox": [ + 130, + 540, + 480, + 614 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 540, + 480, + 614 + ], + "type": "text", + "content": ", we first perform reverse diffusion process using the denoising network " + }, + { + "bbox": [ + 130, + 540, + 480, + 614 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 130, + 540, + 480, + 614 + ], + "type": "text", + "content": " to output " + }, + { + "bbox": [ + 130, + 540, + 480, + 614 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_{t-1/2}" + }, + { + "bbox": [ + 130, + 540, + 480, + 614 + ], + "type": "text", + "content": " (Step 5). Next, we perform a gradient descent update on " + }, + { + "bbox": [ + 130, + 540, + 480, + 614 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_{t-1/2}" + }, + { + "bbox": [ + 130, + 540, + 480, + 614 + ], + "type": "text", + "content": " the reburrowing loss " + }, + { + "bbox": [ + 130, + 540, + 480, + 614 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{y} - \\hat{\\mathbf{k}}_0 * \\hat{\\mathbf{x}}_0 \\|^2" + }, + { + "bbox": [ + 130, + 540, + 480, + 614 + ], + "type": "text", + "content": " (Step 7-9). Similar to Blind-DPS [15], we set the step size " + }, + { + "bbox": [ + 130, + 540, + 480, + 614 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 540, + 480, + 614 + ], + "type": "text", + "content": " according to following rule" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 266, + 622, + 480, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 622, + 480, + 648 + ], + "spans": [ + { + "bbox": [ + 266, + 622, + 480, + 648 + ], + "type": "interline_equation", + "content": "\\delta = \\frac {\\rho}{\\left\\| \\mathbf {y} - \\hat {\\mathbf {x}} _ {0} * \\hat {\\mathbf {k}} _ {0} \\right\\|} \\tag {26}", + "image_path": "b5eb8ba294b8c4b465e8ac21c56adbac0dee6404fc1ec671da7dfbcd54f024b9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 653, + 294, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 653, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 653, + 294, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 653, + 294, + 666 + ], + "type": "inline_equation", + "content": "\\rho = 0.05" + }, + { + "bbox": [ + 130, + 653, + 294, + 666 + ], + "type": "text", + "content": " using trial and error." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 236, + 102 + ], + "type": "text", + "content": "Y. Sanghvi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 134, + 129, + 478, + 331 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 249, + 128 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 249, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 249, + 128 + ], + "type": "text", + "content": "Algorithm 1 Kernel-Diff" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "lines": [ + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "spans": [ + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "text", + "content": "1: Input: Blurred Image y, Diffusion UNet " + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(.)" + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "text", + "content": " , Non-Blind Solver " + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "inline_equation", + "content": "F_{\\phi}(.)" + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "text", + "content": " \n2: // Stage I: Reverse Diffusion \n3: " + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "text", + "content": " \n4: for " + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "inline_equation", + "content": "t = T,\\dots ,1" + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "text", + "content": " do \n5: " + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_{t - 1 / 2} = \\frac{1}{\\sqrt{\\alpha_t}}\\Bigg(\\mathbf{k}_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\bar{\\alpha}_t}}\\epsilon_\\theta (\\mathbf{k}_t,\\mathbf{y},t)\\Bigg) + \\sigma_t\\mathbf{z}_t" + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{k}}_0 = \\frac{1}{\\sqrt{\\bar{\\alpha}_t}}\\big(\\mathbf{k}_t - \\sqrt{1 - \\bar{\\alpha}_t}\\epsilon_\\theta (\\mathbf{k}_t,\\mathbf{y},t)\\big)" + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0 = F_{\\phi}(\\mathbf{y},\\hat{\\mathbf{k}}_0)" + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "text", + "content": " \n9: " + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_{t - 1} = \\mathbf{k}_{t - 1 / 2} - \\delta \\nabla_{\\mathbf{k}_t}(\\| \\mathbf{y} - \\hat{\\mathbf{k}}_0*\\hat{\\mathbf{x}}_0\\| _2^2)" + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "text", + "content": " \n10: end for \n11: // Stage II: Refinement \n12: for " + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "inline_equation", + "content": "j = 1,\\dots ,J" + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "text", + "content": " do \n13: " + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_0\\gets \\mathbf{k}_0 - \\delta \\nabla_{\\mathbf{k}_0}(\\| \\mathbf{y} - \\mathbf{k}_0*\\boldsymbol {F}_\\phi (\\mathbf{y},\\mathbf{k}_0)\\| _2^2 +\\gamma \\| \\mathbf{k}\\| _1)" + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "text", + "content": " \n14: end for \n15: return " + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_0" + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 134, + 129, + 478, + 331 + ], + "type": "inline_equation", + "content": "{\\bf x}_0 = F_{\\phi}({\\bf y},{\\bf k}_0)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 357, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 429 + ], + "type": "text", + "content": "For further performance boost, once the reverse diffusion process is over, we repeat the gradient descent steps (Step 13) to refine the kernel, similar to Stage II in [61]. Here the overall reburring loss is minimized along with " + }, + { + "bbox": [ + 130, + 357, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 130, + 357, + 482, + 429 + ], + "type": "text", + "content": " penalty on the kernel with sparisty weight " + }, + { + "bbox": [ + 130, + 357, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\gamma = 10^{-4}" + }, + { + "bbox": [ + 130, + 357, + 482, + 429 + ], + "type": "text", + "content": ". Since this stage is meant to only fine-tune the kernel, we keep the number of total iterations in second stage, i.e. " + }, + { + "bbox": [ + 130, + 357, + 482, + 429 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 357, + 482, + 429 + ], + "type": "text", + "content": ", to only 20." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 449, + 230, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 230, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 230, + 463 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 475, + 253, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 253, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 253, + 488 + ], + "type": "text", + "content": "4.1 Experiment Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 497, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 668 + ], + "type": "text", + "content": "Implementation of Kernel-Diff. The diffusion network we use is a modified U-Net [22]. To train the diffusion model " + }, + { + "bbox": [ + 130, + 497, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{k}_t,\\mathbf{y},t)" + }, + { + "bbox": [ + 130, + 497, + 482, + 668 + ], + "type": "text", + "content": ", we generate blurred data using synthetic motion blur kernels, with details of data synthesis to be elaborated in the next section. We assume symmetric boundary conditions during the convolution process. We train the diffusion U-Net architecture for 500,000 iterations with a batch size of 16. The non-blind solver in our scheme, i.e. " + }, + { + "bbox": [ + 130, + 497, + 482, + 668 + ], + "type": "inline_equation", + "content": "F_{\\phi}(.)" + }, + { + "bbox": [ + 130, + 497, + 482, + 668 + ], + "type": "text", + "content": ", can be any differentiable neural network based method. In this paper, we plug Deep Wiener Deconvolution (DWDN) [24] to guide the reverse diffusion process. Training Dataset. The diffusion U-Net in Kernel-Diff as well as all methods we compare in our experiments are trained using the same set of blurred and clean image data. We synthesize the blurred images by applying random blur kernels from a kernel database on clean images, similar to [15]. The kernel database consists of " + }, + { + "bbox": [ + 130, + 497, + 482, + 668 + ], + "type": "inline_equation", + "content": "60,000\\times 64" + }, + { + "bbox": [ + 130, + 497, + 482, + 668 + ], + "type": "text", + "content": " blur kernels generated using an open-sourced motion blur model [4] and blur kernel generation method provided by [3]. We note that" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 101 + ], + "type": "text", + "content": "Kernel Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "the former method from [4] produces larger blur kernels than usually observed in practice. However, we include them in our analysis for sake of fair comparison to Blind-DPS which uses method to train diffusion model for blur kernel generation. For clean images, we use the full-resolution images from Flickr2K [46] and LSIDR [43] since diffusion models can be require large amounts of training data. To train the diffusion U-Net for Kernel-Diff, we additionally randomly crop the images to " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " patches and augment the data by random horizontal flipping." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 133, + 298, + 499, + 371 + ], + "blocks": [ + { + "bbox": [ + 130, + 221, + 482, + 288 + ], + "lines": [ + { + "bbox": [ + 130, + 221, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 482, + 288 + ], + "type": "text", + "content": "Table 1: Performance on BSD100 dataset with synthetic blur. Bold and underline refer to overall and second best blind deconvolution method respectively. Maximum normalized correlation (MNC) is provided to evaluate estimated kernels, when explicitly provided in case of Self-Deblur, Blind-DPS etc. DWDN is a non-blind method which knows the ground-truth kernel (i.e. " + }, + { + "bbox": [ + 130, + 221, + 482, + 288 + ], + "type": "inline_equation", + "content": "\\mathrm{MNC} = 1.00" + }, + { + "bbox": [ + 130, + 221, + 482, + 288 + ], + "type": "text", + "content": ") and is provided as an empirical upper bound only." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 298, + 499, + 371 + ], + "lines": [ + { + "bbox": [ + 133, + 298, + 499, + 371 + ], + "spans": [ + { + "bbox": [ + 133, + 298, + 499, + 371 + ], + "type": "table", + "html": "
Method → Metric ↓DeblurGAN [39]DMPHN [85]MPR-Net [84]Self-Deblur [54]Fast-Deblur [13]Blind-DPS [15]Kernel-Diff (Ours)DWDN [24]
PSNR ↑15.4418.7318.4913.5314.0217.5619.0722.95
SSIM ↑0.3960.4270.4350.2270.2960.3870.5000.810
LPIPS ↓0.5630.6170.6320.6650.5830.6700.3550.169
FID ↓287.44265.05265.60259.40304.16280.53172.3379.11
MNC ↑---0.2060.5200.2830.6631.0
", + "image_path": "c9b9f947d61e70ab7cf3be10d1e0a91ead7565446fd4fde350f0d7d33b7ea0d3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 133, + 482, + 493, + 541 + ], + "blocks": [ + { + "bbox": [ + 130, + 416, + 482, + 472 + ], + "lines": [ + { + "bbox": [ + 130, + 416, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 482, + 472 + ], + "type": "text", + "content": "Table 2: Performance on RealBlur-50. Bold and underline refer to overall and second best blind deconvolution method respectively. The corresponding ground truth kernel is not provided for the RealBlur [55] dataset since the blur kernel may not be spatially invariant. As a result the metric MNC along with the performance on DWDN is not provided." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 482, + 493, + 541 + ], + "lines": [ + { + "bbox": [ + 133, + 482, + 493, + 541 + ], + "spans": [ + { + "bbox": [ + 133, + 482, + 493, + 541 + ], + "type": "table", + "html": "
Method → Metric ↓DeblurGAN [39]Fast-Deblur [13]Self-Deblur [54]Blind-DPS [85]LMG [11]DMPHN [85]MPR-Net [84]Kernel-Diff (Ours)
PSNR ↑14.9818.5416.1519.8224.2024.3723.6325.57
SSIM ↑0.5830.4980.4830.6070.7890.7190.7150.798
LPIPS ↓0.2890.3920.4590.3200.1830.2010.2130.171
FID ↓129.06236.21172.32175.95106.35135.33113.1198.89
", + "image_path": "7b13bc40de187e83352233b113079be04eee3a36fd10ca4a0be4e6c6dfac754c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 584, + 288, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 584, + 288, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 584, + 288, + 596 + ], + "type": "text", + "content": "4.2 Quantitative Comparison" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "In this section, we evaluate the proposed method Kernel-Diff on synthetic and real blurred images and compare the performance with other state-of-the-art deconvolution methods. We evaluate our method on the BSD100 and RealBlur datasets as described below along with the following deconvolution/deblurring methods - Deep Hierarchical Multi-Patch Network (DHMPN) [85], Self-Deblur" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "text", + "content": "Y. Sanghvi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 140, + 493, + 582 + ], + "blocks": [ + { + "bbox": [ + 134, + 140, + 493, + 582 + ], + "lines": [ + { + "bbox": [ + 134, + 140, + 493, + 582 + ], + "spans": [ + { + "bbox": [ + 134, + 140, + 493, + 582 + ], + "type": "image", + "image_path": "c36b0ffc10d152c4048ee83f800e638cac306bb2e2d672b3b0ea71a8151b4754.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 591, + 482, + 635 + ], + "lines": [ + { + "bbox": [ + 132, + 591, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 591, + 482, + 635 + ], + "type": "text", + "content": "Fig. 3: Qualitative Results: Reconstruction results on synthetically blurred images from the BSD100 dataset. For Self-Deblur, Blind-DPS and Kernel-Diff, we show the estimated kernels in inset. " + }, + { + "bbox": [ + 132, + 591, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 132, + 591, + 482, + 635 + ], + "type": "text", + "content": "DWDN is a non-blind deconvolution method assuming knowledge of the ground-truth kernel; its results are provided for reference only." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "text", + "content": "Kernel Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 174 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 174 + ], + "type": "text", + "content": "[54], Blind-DPS [15], MPR-Net [84], and DeblurGANv2 [39]. We use these methods as a representative set of contemporary deconvolution methods which include both iterative and end-to-end trained solutions. For the sake of a fair comparison, we retrain the end-to-end methods using synthetically blurred datasets with details provided in Section B of the supplementary document." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 175, + 479, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 479, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 479, + 258 + ], + "type": "text", + "content": "BSD100 Dataset. We first evaluate our method on 100 synthetically blurred images from the BSD dataset [47]. Specifically, we crop out random patches of size " + }, + { + "bbox": [ + 130, + 175, + 479, + 258 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 175, + 479, + 258 + ], + "type": "text", + "content": " and then blur them with synthetic motion blur kernels using symmetric boundary conditions. The performance of Kernel-Diff and other deblurring methods is shown in Table 1 along with some qualitative examples in Figure 3. We also compare estimated kernels from Self-Deblur, Blind-DPS and Kernel-Diff in Table 1 using maximum of normalized convolution (MNC) [30]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 260, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 480, + 365 + ], + "type": "text", + "content": "RealBlur-50. To evaluate our blind deconvolution method on real blurred images, we use the RealBlur-J dataset [55]. However, since the focus of this paper is only spatially invariant blur and not other forms of blur, we crop random patches of size " + }, + { + "bbox": [ + 130, + 260, + 480, + 365 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 260, + 480, + 365 + ], + "type": "text", + "content": " from 50 blurred images in the dataset to reduce the possibility of violating the spatially invarying blur assumption. The results of this comparison are shown in Table 2. We note that we for improved performance on the RealBlur dataset, we use a diffusion model trained only using the realistic blur kernels generated from [3] and increase the sparsity penalty " + }, + { + "bbox": [ + 130, + 260, + 480, + 365 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 260, + 480, + 365 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 260, + 480, + 365 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 383, + 238, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 383, + 238, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 383, + 238, + 395 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 401, + 479, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 479, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 479, + 518 + ], + "type": "text", + "content": "Effect of Non-Blind Solver Guidance Each iteration of the diffusion consists of two sub-steps: (1) reverse diffusion using the trained denoiser " + }, + { + "bbox": [ + 130, + 401, + 479, + 518 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 130, + 401, + 479, + 518 + ], + "type": "text", + "content": " and (2) gradient descent on " + }, + { + "bbox": [ + 130, + 401, + 479, + 518 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{y} - \\hat{\\mathbf{k}}_0 * \\hat{\\mathbf{x}}_0 \\|^2" + }, + { + "bbox": [ + 130, + 401, + 479, + 518 + ], + "type": "text", + "content": " using the non-blind solver. In Table 3, we show the effect on performance of the second sub-step. Through the giant gap in performance with and without the guidance, we find that reverse diffusion is insufficient to converge to a good kernel estimate and hence, the experiment justifies the extra complexity required to guide the diffusion process to minimize the reblurring loss. In the same ablation study, we show how using the correct step size is important for the performance and how the second stage of finetuning impacts the overall scheme's performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 521, + 479, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 479, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 479, + 616 + ], + "type": "text", + "content": "Computational Comparison The method proposed in this paper, Kernel-Diff is a kernel estimation analogue of the Blind-DPS method as both of them aim to solve the blind deconvolution problem using diffusion models. Therefore, we compare the computational costs of both the techniques in Table 4. Both techniques use a diffusion model trained for " + }, + { + "bbox": [ + 131, + 521, + 479, + 616 + ], + "type": "inline_equation", + "content": "T = 1000" + }, + { + "bbox": [ + 131, + 521, + 479, + 616 + ], + "type": "text", + "content": " steps and our method performs much better than Blind-DPS which uses the alternating minimization approach using fewer parameters and similar time. Future work may involve reducing the inference time using DDIM [65] or consistency models [66]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 617, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 479, + 665 + ], + "type": "text", + "content": "Reblurring Loss during Diffusion Levin et. al [40] show that alternate minimization approach is more likely to end up in a false local minima. We empirically demonstrate this by comparing the residual across the iterations, i.e., " + }, + { + "bbox": [ + 131, + 617, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{y} - \\hat{\\mathbf{k}}_0\\otimes \\hat{\\mathbf{x}}_0\\|" + }, + { + "bbox": [ + 131, + 617, + 479, + 665 + ], + "type": "text", + "content": " with Blind-DPS [15] in Figure 5. From the figure we observe that" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "text", + "content": "Y. Sanghvi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "the residual in our proposed kernel estimation scheme converges to a solution with a much lower residual value. Assuming the kernel and image likelihoods to be approximately equal for both solutions, we conclude that our method leads to solutions with a higher log-likelihood, and hence, better quality results." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 148, + 184, + 252, + 290 + ], + "blocks": [ + { + "bbox": [ + 148, + 184, + 252, + 290 + ], + "lines": [ + { + "bbox": [ + 148, + 184, + 252, + 290 + ], + "spans": [ + { + "bbox": [ + 148, + 184, + 252, + 290 + ], + "type": "image", + "image_path": "9e0e5e2c59910da4a00c34a75d59c57f57f538231c6444d91e86ef919665036c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 293, + 232, + 316 + ], + "lines": [ + { + "bbox": [ + 168, + 293, + 232, + 316 + ], + "spans": [ + { + "bbox": [ + 168, + 293, + 232, + 316 + ], + "type": "text", + "content": "w/o non-blind solver guidance" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 254, + 184, + 359, + 290 + ], + "blocks": [ + { + "bbox": [ + 254, + 184, + 359, + 290 + ], + "lines": [ + { + "bbox": [ + 254, + 184, + 359, + 290 + ], + "spans": [ + { + "bbox": [ + 254, + 184, + 359, + 290 + ], + "type": "image", + "image_path": "64e324936f037901ee3719ea1fd566b2fdd0973739bbae463f6936071e10f7e0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 293, + 339, + 316 + ], + "lines": [ + { + "bbox": [ + 274, + 293, + 339, + 316 + ], + "spans": [ + { + "bbox": [ + 274, + 293, + 339, + 316 + ], + "type": "text", + "content": "with non-blind solver guidance" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 361, + 185, + 466, + 290 + ], + "blocks": [ + { + "bbox": [ + 361, + 185, + 466, + 290 + ], + "lines": [ + { + "bbox": [ + 361, + 185, + 466, + 290 + ], + "spans": [ + { + "bbox": [ + 361, + 185, + 466, + 290 + ], + "type": "image", + "image_path": "2ae7b294a0177dd6ff332cc044075edc2597a665db99e24846b3a50e0fc23f93.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 293, + 443, + 315 + ], + "lines": [ + { + "bbox": [ + 384, + 293, + 443, + 315 + ], + "spans": [ + { + "bbox": [ + 384, + 293, + 443, + 315 + ], + "type": "text", + "content": "DWDN + True kernel" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 326, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 361 + ], + "type": "text", + "content": "Fig. 4: Effect of Non-Blind Solver Guidance. Through this ablation study, we examine the effect of the second sub-step in the diffusion process which guides the diffusion process to minimize the reburring loss using the non-blind solver DWDN [24]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 138, + 481, + 473, + 574 + ], + "blocks": [ + { + "bbox": [ + 130, + 404, + 482, + 472 + ], + "lines": [ + { + "bbox": [ + 130, + 404, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 482, + 472 + ], + "type": "text", + "content": "Table 3: Ablation study comparing effect of different hyperparameters of the scheme on the performance. We observe that removing the guidance of the non-blind solver reduces the performance across all metrics. While using gradient descent using non-blind solver in (24), for optimal performance, the step size shouldn't be set too large. Lastly, the effect of finetuning kernel in Stage-II can also provide a significant performance boost for PSNR, however, not so much for other metrics." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 481, + 473, + 574 + ], + "lines": [ + { + "bbox": [ + 138, + 481, + 473, + 574 + ], + "spans": [ + { + "bbox": [ + 138, + 481, + 473, + 574 + ], + "type": "table", + "html": "
Ablation Study ↓ Metric →PSNR ↑SSIM ↑LPIPS ↓FID ↓MNC ↑
ρ = 0, No DWDN21.900.6500.251152.510.714
ρ = 0.0522.640.6800.234140.280.724
ρ = 0.121.970.6510.256156.040.567
ρ = 0.520.870.6090.300173.360.270
J = 0, No Stage II21.760.6710.241143.580.675
DWDN, known kernel30.540.9060.08749.291.00
", + "image_path": "203c76c10d2b5bde55a1c838deb99db41b1d21244eefa2d399f29cf44ff7723d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 616, + 220, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 220, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 220, + 628 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "In this paper, we proposed Kernel-Diff, a blind deconvolution solution in the tradition of Levin et. al [40]. We show how the kernel minimization approach can be" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 101 + ], + "type": "text", + "content": "Kernel Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 203, + 106, + 410, + 262 + ], + "blocks": [ + { + "bbox": [ + 203, + 106, + 410, + 262 + ], + "lines": [ + { + "bbox": [ + 203, + 106, + 410, + 262 + ], + "spans": [ + { + "bbox": [ + 203, + 106, + 410, + 262 + ], + "type": "image", + "image_path": "8b081afa8d5e605ab5d803911d7f0664d0f6b9ef1b3c528ab73332ef5f677531.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 274, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 319 + ], + "type": "text", + "content": "Fig. 5: Residual " + }, + { + "bbox": [ + 130, + 274, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{y} - \\hat{\\mathbf{k}}_0\\otimes \\hat{\\mathbf{x}}_0\\|^2" + }, + { + "bbox": [ + 130, + 274, + 482, + 319 + ], + "type": "text", + "content": " for Blind-DPS and Kernel-Diff. This experiments confirms our hypothesis that the proposed kernel estimation scheme converges to a better local minimum. The shaded region shows the range of the residual over 50 trials of each scheme." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 190, + 417, + 423, + 475 + ], + "blocks": [ + { + "bbox": [ + 130, + 330, + 482, + 407 + ], + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 407 + ], + "type": "text", + "content": "Table 4: Comparison of computational cost with Blind-DPS in terms of total number of parameters and time taken to deblur a " + }, + { + "bbox": [ + 130, + 330, + 482, + 407 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 330, + 482, + 407 + ], + "type": "text", + "content": " image respectively. Both Blind-DPS and Kernel-Diff use a diffusion model approach to the blind deconvolution problem, our kernel estimation approach uses a lightweight diffusion model and non-blind solver for the reverse diffusion process. Using significantly fewer parameters and smaller computational time, we achieve significant performance boost over the alternating minimization approach." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 190, + 417, + 423, + 475 + ], + "lines": [ + { + "bbox": [ + 190, + 417, + 423, + 475 + ], + "spans": [ + { + "bbox": [ + 190, + 417, + 423, + 475 + ], + "type": "table", + "html": "
Metric → Method ↓Total Parameter CountTotal Time
Blind-DPS110M337s
Kernel-Diff + DWDN43M305s
", + "image_path": "1b8f239832f27305c2dc7e0f3bf99ea415329a497291f5facdb1dc27406ecee8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 497, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 481, + 544 + ], + "type": "text", + "content": "implemented in practice using a pre-trained non-blind solver in diffusion models setting and propose a guided diffusion method that samples from " + }, + { + "bbox": [ + 130, + 497, + 481, + 544 + ], + "type": "inline_equation", + "content": "p(\\mathbf{k}|\\mathbf{y})" + }, + { + "bbox": [ + 130, + 497, + 481, + 544 + ], + "type": "text", + "content": " to solve the blind deconvolution problem and show extensive numerical experiments on blurred images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 593 + ], + "type": "text", + "content": "For future work, the method can be refined to deblur images with spatially varying and motion blur i.e. degradation which do not fit the single-blur convolutional model. Alternately, the non-blind solver can be modified to be robust to inaccuracies in kernel estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 611, + 247, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 247, + 625 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 247, + 625 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 636, + 481, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 636, + 481, + 660 + ], + "spans": [ + { + "bbox": [ + 130, + 636, + 481, + 660 + ], + "type": "text", + "content": "This work is supported, in part, by the DARPA / SRC CogniSense JUMP 2.0 Center, NSF IIS-2133032, and NSF ECCS-2030570." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "text", + "content": "Y. Sanghvi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "text", + "content": "1. Ahmad, R., Bouman, C.A., Buzzard, G.T., Chan, S., Liu, S., Reehorst, E.T., Schniter, P.: Plug-and-play methods for magnetic resonance imaging: Using de-noisers for image recovery. IEEE signal processing magazine 37(1), 105-116 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 195 + ], + "type": "text", + "content": "2. Anger, J., Facciolo, G., Delbracio, M.: Blind image deblurring using the 10 gradient prior. Image processing on line 9, 124-142 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 195, + 480, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 480, + 227 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 480, + 227 + ], + "type": "text", + "content": "3. Boracchi, G., Foi, A.: Modeling the performance of image restoration from motion blur. Image Processing, IEEE Transactions on 21(8), 3502-3517 (aug 2012). https://doi.org/10.1109/TIP.2012.2192126" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 248 + ], + "type": "text", + "content": "4. Borodenko, L.: Motion blur (May 2020), https://github.com/LeviBorodenko/motionblur" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 249, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 249, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 480, + 282 + ], + "type": "text", + "content": "5. Carbajal, G., Vitoria, P., Delbracio, M., Muse, P., Lezama, J.: Non-uniform blur kernel estimation via adaptive basis decomposition. arXiv preprint arXiv:2102.01026 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 282, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 282, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 282, + 480, + 315 + ], + "type": "text", + "content": "6. Carbajal, G., Vitoria, P., Lezama, J., Muse, P.: Blind motion deblurring with pixel-wise kernel estimation via kernel prediction networks. IEEE Transactions on Computational Imaging (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 316, + 480, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 480, + 348 + ], + "type": "text", + "content": "7. Chakrabarti, A.: A neural approach to blind motion deblurring. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14. pp. 221-235. Springer (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 348, + 480, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 480, + 381 + ], + "type": "text", + "content": "8. Chan, S.H., Wang, X., Elgendy, O.A.: Plug-and-play admm for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging 3(1), 84-98 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 381, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 480, + 403 + ], + "type": "text", + "content": "9. Chan, T.F., Wong, C.K.: Total variation blind deconvolution. IEEE transactions on Image Processing 7(3), 370-375 (1998)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "text", + "content": "0. Chen, L., Fang, F., Wang, T., Zhang, G.: Blind image deblurring with local maximum gradient prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1742-1750 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "type": "text", + "content": "1. Chen, L., Fang, F., Wang, T., Zhang, G.: Blind image deblurring with local maximum gradient prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1742-1750 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "type": "text", + "content": "2. Chen, L., Zhang, J., Li, Z., Wei, Y., Fang, F., Ren, J., Pan, J.: Deep richardson-lucy deconvolution for low-light image deblurring. International Journal of Computer Vision pp. 1-18 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 501, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 523 + ], + "type": "text", + "content": "3. Cho, S., Lee, S.: Fast motion deblurring. ACM Transactions on Graphics pp. 1-8 (2009)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "text", + "content": "4. Choi, J., Kim, S., Jeong, Y., Gwon, Y., Yoon, S.: Ilvr: Conditioning method for denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14367-14376 (October 2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 555, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 589 + ], + "type": "text", + "content": "5. Chung, H., Kim, J., Kim, S., Ye, J.C.: Parallel diffusion models of operator and image for blind inverse problems. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6059-6069 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 589, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 631 + ], + "type": "text", + "content": "6. Chung, H., Kim, J., Mccann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=OnD9zGAGT0k" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "7. Chung, H., Ryu, D., McCann, M.T., Klasky, M.L., Ye, J.C.: Solving 3d inverse problems using pre-trained 2d diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22542-22551 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "text", + "content": "Kernel Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "18. Chung, H., Sim, B., Ryu, D., Ye, J.C.: Improving diffusion models for inverse problems using manifold constraints. Advances in Neural Information Processing Systems 35, 25683-25696 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "19. Chung, H., Sim, B., Ye, J.C.: Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12413-12422 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "20. Cohen, R., Elad, M., Milanfar, P.: Regularization by denoising via fixed-point projection (red-pro). SIAM Journal on Imaging Sciences 14(3), 1374-1406 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "21. Daras, G., Delbracio, M., Talebi, H., Dimakis, A., Milanfar, P.: Soft diffusion: Score matching with general corruptions. Transactions on Machine Learning Research (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "text", + "content": "22. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "23. Dong, J., Pan, J., Sun, D., Su, Z., Yang, M.H.: Learning data terms for non-blind deblurring. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 748-763 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "24. Dong, J., Roth, S., Schiele, B.: Deep wiener deconvolution: Wiener meets deep learning for image deblurring. Advances in Neural Information Processing Systems 33, 1048-1059 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "25. Dong, J., Roth, S., Schiele, B.: Learning spatially-variant map models for non-blind image deblurring. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4886-4895 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 370, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 413 + ], + "type": "text", + "content": "26. Eboli, T., Sun, J., Ponce, J.: End-to-end interpretable learning of non-blind image deblurring. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVII 16. pp. 314-331. Springer (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "text", + "content": "27. Gong, D., Zhang, Z., Shi, Q., van den Hengel, A., Shen, C., Zhang, Y.: Learning deep gradient descent optimization for image deconvolution. IEEE transactions on neural networks and learning systems 31(12), 5468-5482 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 447, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 490 + ], + "type": "text", + "content": "28. Gupta, A., Joshi, N., Lawrence Zitnick, C., Cohen, M., Curless, B.: Single image deblurring using motion density functions. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part I 11. pp. 171-184. Springer (2010)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 491, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 511 + ], + "type": "text", + "content": "29. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "30. Hu, Z., Yang, M.H.: Good regions to deblur. In: Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12. pp. 59-72. Springer (2012)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "text", + "content": "31. Jalal, A., Arvinte, M., Daras, G., Price, E., Dimakis, A.G., Tamir, J.: Robust compressed sensing mri with deep generative priors. Advances in Neural Information Processing Systems 34, 14938-14954 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 599 + ], + "type": "text", + "content": "32. Jiang, H., Luo, A., Han, S., Fan, H., Liu, S.: Low-light image enhancement with wavelet-based diffusion models. arXiv preprint arXiv:2306.00306 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "33. Kadkhodaie, Z., Simoncelli, E.: Stochastic solutions for linear inverse problems using the prior implicit in a denoiser. Advances in Neural Information Processing Systems 34, 13242-13254 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "34. Kamilov, U.S., Mansour, H., Wohlberg, B.: A plug-and-play priors approach for solving nonlinear imaging inverse problems. IEEE Signal Processing Letters 24(12), 1872-1876 (2017)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "text", + "content": "Y. Sanghvi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "35. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. Advances in Neural Information Processing Systems 35, 23593-23606 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 161 + ], + "type": "text", + "content": "36. Kawar, B., Vaksman, G., Elad, M.: Snips: Solving noisy inverse problems stochastically. Advances in Neural Information Processing Systems 34, 21757-21769 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 163, + 480, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 163, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 163, + 480, + 195 + ], + "type": "text", + "content": "37. Kawar, B., Vaksman, G., Elad, M.: Stochastic image denoising by sampling from the posterior distribution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1866-1875 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 197, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 197, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 197, + 480, + 239 + ], + "type": "text", + "content": "38. Kupyn, O., Budzan, V., Mykhailych, M., Mishkin, D., Matas, J.: Deblurgan: Blind motion deblurring using conditional adversarial networks. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8183-8192 (2018). https://doi.org/10.1109/CVPR.2018.00854" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 241, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 241, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 241, + 480, + 274 + ], + "type": "text", + "content": "39. Kupyn, O., Martyniuk, T., Wu, J., Wang, Z.: Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In: The IEEE International Conference on Computer Vision (ICCV) (Oct 2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 275, + 480, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 275, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 133, + 275, + 480, + 307 + ], + "type": "text", + "content": "40. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding blind deconvolution algorithms. IEEE Transactions on Pattern Analysis and Machine Intelligence 33(12), 2354-2367 (2011)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 308, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 308, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 133, + 308, + 480, + 341 + ], + "type": "text", + "content": "41. Li, H., Yang, Y., Chang, M., Chen, S., Feng, H., Xu, Z., Li, Q., Chen, Y.: Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing 479, 47-59 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 342, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 342, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 133, + 342, + 480, + 374 + ], + "type": "text", + "content": "42. Li, X., Ren, Y., Jin, X., Lan, C., Wang, X., Zeng, W., Wang, X., Chen, Z.: Diffusion models for image restoration and enhancement-a comprehensive survey. arXiv preprint arXiv:2308.09388 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 376, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 376, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 133, + 376, + 480, + 418 + ], + "type": "text", + "content": "43. Li, Y., Zhang, K., Liang, J., Cao, J., Liu, C., Gong, R., Zhang, Y., Tang, H., Liu, Y., Demandolx, D., et al.: LSDIR: A large scale dataset for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1775-1787 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 420, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 420, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 133, + 420, + 480, + 453 + ], + "type": "text", + "content": "44. Li, Y., Tofighi, M., Geng, J., Monga, V., Eldar, Y.C.: Efficient and interpretable deep blind image deblurring via algorithm unrolling. IEEE Transactions on Computational Imaging 6, 666-681 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 454, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 454, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 133, + 454, + 480, + 486 + ], + "type": "text", + "content": "45. Li, Y., Tofighi, M., Monga, V., Eldar, Y.C.: An algorithm unrolling approach to deep image deblurring. In: ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 7675-7679. IEEE (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 487, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 487, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 133, + 487, + 480, + 520 + ], + "type": "text", + "content": "46. Lim, B., Son, S., Kim, H., Nah, S., Lee, K.M.: Enhanced deep residual networks for single image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (July 2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 521, + 480, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 521, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 133, + 521, + 480, + 563 + ], + "type": "text", + "content": "47. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: Proc. 8th Int'l Conf. Computer Vision. vol. 2, pp. 416-423 (July 2001)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 566, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 566, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 133, + 566, + 480, + 597 + ], + "type": "text", + "content": "48. Monga, V., Li, Y., Eldar, Y.C.: Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine 38(2), 18-44 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 599, + 480, + 632 + ], + "type": "text", + "content": "49. Murata, N., Saito, K., Lai, C.H., Takida, Y., Uesaka, T., Mitsufuji, Y., Ermon, S.: Gibbsddrm: A partially collapsed gibbs sampler for solving blind inverse problems with denoising diffusion restoration. arXiv preprint arXiv:2301.12686 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "text", + "content": "50. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "text", + "content": "Kernel Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 149 + ], + "type": "text", + "content": "51. Nan, Y., Ji, H.: Deep learning for handling kernel/model uncertainty in image deconvolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2388-2397 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "52. Pan, J., Sun, D., Pfister, H., Yang, M.H.: Blind image deblurring using dark channel prior. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1628-1636 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "53. Reehorst, E.T., Schniter, P.: Regularization by denoising: Clarifications and new interpretations. IEEE transactions on computational imaging 5(1), 52-67 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "54. Ren, D., Zhang, K., Wang, Q., Hu, Q., Zuo, W.: Neural blind deconvolution using deep priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3341-3350 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "55. Rim, J., Lee, H., Won, J., Cho, S.: Real-world blur dataset for learning and benchmarking deblurring algorithms. In: Proceedings of the European Conference on Computer Vision (ECCV) (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "text", + "content": "56. Romano, Y., Elad, M., Milanfar, P.: The little engine that could: Regularization by denoising (red). SIAM Journal on Imaging Sciences 10(4), 1804-1844 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "57. Rond, A., Giryes, R., Elad, M.: Poisson inverse problems by the plug-and-play scheme. Journal of Visual Communication and Image Representation 41, 96-108 (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "type": "text", + "content": "58. Saharia, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. arXiv:2104.07636 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "59. Sanghvi, Y., Gnanasambandam, A., Chan, S.H.: Photon limited non-blind deblurring using algorithm unrolling. IEEE Transactions on Computational Imaging 8, 851-864 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "60. Sanghvi, Y., Gnanasambandam, A., Mao, Z., Chan, S.H.: Photon-limited blind deconvolution using unsupervised iterative kernel estimation. IEEE Transactions on Computational Imaging 8, 1051-1062 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "61. Sanghvi, Y., Mao, Z., Chan, S.H.: Structured kernel estimation for photon-limited deconvolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9863-9872 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "text", + "content": "62. Schuler, C.J., Hirsch, M., Harmeling, S., Scholkopf, B.: Learning to deblur. IEEE transactions on pattern analysis and machine intelligence 38(7), 1439-1451 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "63. Shan, Q., Jia, J., Agarwala, A.: High-quality motion deblurring from a single image. ACM Transactions on Graphics 27(3), 1-10 (2008)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "text", + "content": "64. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "text", + "content": "65. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv:2010.02502 (October 2020), https://arxiv.org/abs/2010.02502" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 545, + 454, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 454, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 454, + 555 + ], + "type": "text", + "content": "66. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "type": "text", + "content": "67. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=PxTIG12RRHS" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "68. Sun, J., Cao, W., Xu, Z., Ponce, J.: Learning a convolutional neural network for non-uniform motion blur removal. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 769-777 (2015)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "69. Sun, L., Cho, S., Wang, J., Hays, J.: Edge-based blur kernel estimation using patch priors. In: IEEE international conference on computational photography (ICCP). pp. 1-8. IEEE (2013)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "text", + "content": "Y. Sanghvi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "70. Sun, Y., Wohlberg, B., Kamilov, U.S.: An online plug-and-play algorithm for regularized image reconstruction. IEEE Transactions on Computational Imaging 5(3), 395-408 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "type": "text", + "content": "71. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-recurrent network for deep image deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8174-8182 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "type": "text", + "content": "72. Ulyanov, D., Vedaldi, A., Lempitsky, V.: Deep image prior. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9446-9454 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 207, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 207, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 133, + 207, + 481, + 239 + ], + "type": "text", + "content": "73. Venkatakrishnan, S.V., Bouman, C.A., Wohlberg, B.: Plug-and-play priors for model based reconstruction. In: 2013 IEEE global conference on signal and information processing. pp. 945-948. IEEE (2013)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 241, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 241, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 133, + 241, + 481, + 262 + ], + "type": "text", + "content": "74. Vincent, P.: A connection between score matching and denoising autoencoders. Neural computation 23(7), 1661-1674 (2011)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 263, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 263, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 133, + 263, + 481, + 296 + ], + "type": "text", + "content": "75. Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. The Eleventh International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 297, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 297, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 133, + 297, + 481, + 341 + ], + "type": "text", + "content": "76. Wang, Y., Yu, Y., Yang, W., Guo, L., Chau, L.P., Kot, A.C., Wen, B.: Exposurediffusion: Learning to expose for low-light image enhancement. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12438-12448 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 342, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 342, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 133, + 342, + 481, + 374 + ], + "type": "text", + "content": "77. Wang, Z., Cun, X., Bao, J., Liu, J.: Uformer: A general u-shaped transformer for image restoration. in 2022 IEEE. In: CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17662-17672 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 376, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 376, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 376, + 481, + 407 + ], + "type": "text", + "content": "78. Whang, J., Delbracio, M., Talebi, H., Sahara, C., Dimakis, A.G., Milanfar, P.: Deblurring via stochastic refinement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16293-16303 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 409, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 409, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 133, + 409, + 481, + 430 + ], + "type": "text", + "content": "79. Xu, L., Jia, J.: Two-phase kernel estimation for robust motion deblurring. In: European Conference on Computer Vision. pp. 157-170. Springer (2010)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 432, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 432, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 133, + 432, + 481, + 464 + ], + "type": "text", + "content": "80. Xu, L., Zheng, S., Jia, J.: Unnatural l0 sparse representation for natural image deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1107-1114 (2013)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 465, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 465, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 133, + 465, + 481, + 498 + ], + "type": "text", + "content": "81. Yan, Y., Ren, W., Guo, Y., Wang, R., Cao, X.: Image deblurring via extreme channels prior. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4003-4011 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 498, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 498, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 133, + 498, + 481, + 553 + ], + "type": "text", + "content": "82. yang, y., Sun, J., Li, H., Xu, Z.: Deep admm-net for compressive sensing mri. In: Lee, D., Sugiyama, M., Luxburg, U., Guyon, I., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 29. Curran Associates, Inc. (2016), https://proceedings.neurips.cc/paper_files/paper/2016/file/1679091c5a880faf6fb5e6087eb1b2dc-Paper.pdf" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 554, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 554, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 133, + 554, + 481, + 597 + ], + "type": "text", + "content": "83. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 599, + 481, + 632 + ], + "type": "text", + "content": "84. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14821-14831 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "type": "text", + "content": "85. Zhang, H., Dai, Y., Li, H., Koniusz, P.: Deep stacked hierarchical multi-patch network for image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5978-5986 (2019)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "text", + "content": "Kernel Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 259 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "86. Zhang, K., Zuo, W., Gu, S., Zhang, L.: Learning deep cnn denoiser prior for image restoration. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3929-3938 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "text", + "content": "87. Zhang, K., Ren, W., Luo, W., Lai, W.S., Stenger, B., Yang, M.H., Li, H.: Deep image deblurring: A survey. International Journal of Computer Vision 130(9), 2103-2130 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "88. Zhang, Z., Cheng, Y., Suo, J., Bian, L., Dai, Q.: Infwide: Image and feature space wiener deconvolution network for non-blind image deblurring in low-light conditions. IEEE Transactions on Image Processing 32, 1390-1402 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 259 + ], + "type": "text", + "content": "89. Zhu, Y., Zhang, K., Liang, J., Cao, J., Wen, B., Timofte, R., Van Gool, L.: Denoising diffusion models for plug-and-play image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1219-1229 (2023)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 102 + ], + "type": "text", + "content": "Y. Sanghvi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Keypoint Promptable Re-Identification/21228ec1-ee24-4ba9-acb3-211d122a99e2_content_list.json b/2024/Keypoint Promptable Re-Identification/21228ec1-ee24-4ba9-acb3-211d122a99e2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9cae5a37bc95f92e796147f4c857127fbda07389 --- /dev/null +++ b/2024/Keypoint Promptable Re-Identification/21228ec1-ee24-4ba9-acb3-211d122a99e2_content_list.json @@ -0,0 +1,1751 @@ +[ + { + "type": "text", + "text": "Keypoint Promptable Re-Identification", + "text_level": 1, + "bbox": [ + 277, + 141, + 725, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vladimir Somers1,2,3, Alexandre Alahi1, and Christophe De Vleeschouwer2", + "bbox": [ + 215, + 188, + 787, + 207 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Ecole Polytechnique Fédérale de Lausanne (EPFL), Switzerland", + "bbox": [ + 279, + 215, + 723, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Université Catholique de Louvain (UCLouvain), Belgium", + "bbox": [ + 305, + 229, + 697, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Sportradar, Switzerland", + "bbox": [ + 413, + 244, + 589, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Occluded Person Re-Identification (ReID) is a metric learning task that involves matching occluded individuals based on their appearance. While many studies have tackled occlusions caused by objects, multi-person occlusions remain less explored. In this work, we identify and address a critical challenge overlooked by previous occluded ReID methods: the Multi-Person Ambiguity (MPA) arising when multiple individuals are visible in the same bounding box, making it impossible to determine the intended ReID target among the candidates. Inspired by recent work on prompting in vision, we introduce Keypoint Promptable ReID (KPR), a novel formulation of the ReID problem that explicitly complements the input bounding box with a set of semantic keypoints indicating the intended target. Since promptable re-identification is an unexplored paradigm, existing ReID datasets lack the pixel-level annotations necessary for prompting. To bridge this gap and foster further research on this topic, we introduce Occluded PoseTrack-ReID, a novel ReID dataset with keypoints labels, that features strong inter-person occlusions. Furthermore, we release custom keypoints labels for four popular ReID benchmarks. Experiments on person retrieval, but also on pose tracking, demonstrate that our method systematically surpasses previous state-of-the-art approaches on various occluded scenarios. Our code, dataset and annotations are available at https://github.com/V1Somers/Keypoint_prompts_reidentification", + "bbox": [ + 259, + 297, + 743, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Person Re-Identification $\\cdot$ Vision Prompting $\\cdot$ Pose Tracking", + "bbox": [ + 261, + 614, + 740, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 656, + 375, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Person re-identification (ReID) 47 is a challenging retrieval task that involves matching a query image of a person of interest with other person images. ReID finds wide-ranging applications in multi-object tracking 32, pedestrian flow analysis 2, sport understanding 11,14,28,33,38 and video-surveillance 53. However, ReID is a difficult task due to various factors including inaccurate bounding boxes, luminosity changes, poor image quality, and occlusions.", + "bbox": [ + 212, + 688, + 787, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, there has been a growing interest in addressing the occluded re-identification task [30], where the ReID target might be occluded by various objects or other people. In this work, we identify and explicitly address a critical challenge that has been overlooked by previous occluded ReID methods [34, 41]:", + "bbox": [ + 212, + 779, + 787, + 842 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1ad9df224b74aacc1b686ebbf4388b878fee5e4767ba9672bc84d64a38a9575f.jpg", + "image_caption": [ + "(a) Keypoint Promptable ReID (KPR)" + ], + "image_footnote": [], + "bbox": [ + 348, + 143, + 647, + 236 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4b0aa7420085aaac44a1af8e4a9c3f46db364f2299ebfc01a9aa11829c19271f.jpg", + "image_caption": [ + "Fig. 1: Overview of our proposed Keypoint Promptable ReID (KPR) method. KPR takes an image with keypoints prompts as input and produces part-based features of the prompted target. The prompt instructs the model to focus on a specific individual, i.e. the back blue jacket man (b) or the front black t-shirt man (c) in this example. Colored dots illustrate the positive keypoints prompt, with one color per body part." + ], + "image_footnote": [], + "bbox": [ + 243, + 251, + 759, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the presence of Multi-Person Ambiguity (MPA) inherent to bounding box annotations. The multi-person ambiguity arises when a bounding box image contains multiple individuals, making it challenging even for humans to accurately identify the intended ReID target among all the candidates. Figure 2a shows how MPA can negatively impact the re-identification model, causing feature mixing among different individuals or even a focus on an unintended person.", + "bbox": [ + 212, + 463, + 787, + 555 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle this multi-person ambiguity, it is necessary to incorporate additional pixel-level information, such as keypoints or segmentation masks, provided by an upstream human operator or vision model. This additional data plays a vital role in assisting the downstream ReID method to distinguish the intended target from other candidates.", + "bbox": [ + 212, + 558, + 789, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by recent advances in promptable $^{4}$ methods for vision transformers (e.g., SAM [18]), we propose to explicitly use additional semantic $^{5}$ keypoints as inputs to disambiguate images with multiple individuals. Semantic keypoints are suitable for two crucial use cases: 1) manual prompting by a human operator with a few clicks on an image, and 2) automated prompting by an upstream pose estimation model. Nevertheless, our architecture could be easily extended to support segmentation masks as prompts.", + "bbox": [ + 212, + 638, + 787, + 744 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "V. Somers et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "4 A \"prompt\" typically refers to an input instruction that guides the model's response. In vision tasks, e.g. segmentation, prompts can be any kind of data that specifies what the model should focus on [18].", + "bbox": [ + 217, + 768, + 785, + 811 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "5 Semantic keypoints are distinct points within a person image that carry semantic information about specific body parts, typically obtained with pose estimation [22].", + "bbox": [ + 218, + 811, + 785, + 842 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f6b8068b9db777551cc624f85484631d293c415cc09021a88e77ae37c3cf6be2.jpg", + "image_caption": [ + "(a) Existing ReID models (e.g. BPBreID 34) are confused by MPA" + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 491, + 226 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c40029641172a0923771296c580d38c1a0bfb111c02ae15721403ee6da5485e9.jpg", + "image_caption": [ + "(b) KPR explicitly addresses MPA with positive (pastel) and negative (red) keypoints prompt", + "Fig. 2: Person retrieval with Multi-Person Ambiguity (MPA). Green/red borders are correct/incorrect matches. Red/pastel dots indicate negative/positive prompts." + ], + "image_footnote": [], + "bbox": [ + 504, + 143, + 805, + 228 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prior research [8,39,40] on pose-guided methods has explored the use of additional information from external pose estimation models. However, none of these works explicitly tackled the multi-pedestrian ambiguity issue.", + "bbox": [ + 212, + 327, + 782, + 373 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following above insights, we propose KPR, a novel Keypoint-Promptable transformer for part-based person Re-identification, illustrated in Fig. 1a. KPR is a transformer-based model that takes an image along with semantic keypoints as input. It then produces part-based features, each representing a distinct body part of the ReID target, along with their respective visibility scores. Part-based methods have demonstrated superior performance [34,57] in re-identifying occluded individuals because they utilize only the visible parts for comparison. Our method can process both positive and negative keypoints, which respectively represent the target and non-target pedestrians. As illustrated in Fig. 1b, prompts can be used to re-identify the desired (front or back) person in case of occlusion. Furthermore, KPR is designed to be prompt optional to offer more practical flexibility. This means the same model can be used without prompt on non-ambiguous images, or with prompt when dealing with occlusions, while consistently achieving state-of-the-art performance in both cases. To demonstrate the advantages of keypoint prompts for re-identification, we evaluate KPR on two tasks: person retrieval and multi-person pose tracking.", + "bbox": [ + 212, + 376, + 787, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, since no current ReID dataset is provided with explicit target identification in case of multi-person ambiguity, we introduce a new dataset called Occluded PoseTrack-ReID, nicknamed \"Occ-PTrack\", and derived from the PoseTrack21 dataset [6]. This dataset offers high-quality keypoints annotations and features a substantial number of images with multi-person occlusions.", + "bbox": [ + 212, + 621, + 787, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall, we summarize the main contributions of our work as follows:", + "bbox": [ + 238, + 700, + 740, + 715 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose $KPR$ , a novel prompt-optional model for part-based ReID. To our knowledge, we are the first to introduce a visual prompting mechanism for person re-identification, and the first to explicitly address the ambiguity caused by multi-person occlusions.", + "2. We introduce Occ-PTrack: the first multi-person occluded ReID dataset with explicit target identification through manual keypoints annotations. Furthermore, we propose new keypoint annotations for four popular re-identification datasets." + ], + "bbox": [ + 220, + 719, + 785, + 838 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Keypoint Promptable Re-Identification", + "bbox": [ + 465, + 114, + 727, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our method outperforms all previous state-of-the-art methods on the occluded ReID task, demonstrating the effectiveness of the keypoint promptable approach.", + "bbox": [ + 220, + 146, + 785, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our ambition is to advocate for a shift from the ambiguous bounding-box approach to ReID and promote a keypoint-centric paradigm. To this end, we publicly release our codebase, proposed dataset, and annotations, to provide a common setup for evaluating promptable ReID methods.", + "bbox": [ + 212, + 191, + 787, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 277, + 388, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-Person Ambiguity: A previous study [41] addressed a related issue called \"Non-Target Pedestrians (NTP),\" similar to our concept of \"Multi-Person Ambiguity\" (MPA). However, their focus was on mitigating the negative impact of NTP during training. In contrast, we provide a method to explicitly disambiguate the intended ReID target during both training and testing phases.", + "bbox": [ + 212, + 310, + 785, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pose-guided ReID: We detail here the fundamental difference between our novel keypoint promptable approach and the traditional pose-guided methods. Although none of the previous works explicitly address the multi-person ambiguity (MPA) issue, we identify two categories within these pose-guided approaches: 1) those inherently unable to address MPA [48,15,27,29,34,44,45], and 2) those with the potential to overcome MPA [39,40]. In the first category, methods like BPBreID [34] or Pirt [27] utilize pose labels only during training, and can not leverage additional information at test time to disambiguate multiple persons. Moreover, methods relying on horizontal stripes (VGTri [45], PGFA [29]) or the affinity fields of a pose model (PVPM [8]), are meant to tackle object-occlusions, but struggle to disambiguate multiple persons. In the second category, methods like HOReID [39] and PFD [40] directly employ keypoints from a single person, enabling focus on a specified target. However, HOReID [39] is restricted to utilizing spatial features at the exact keypoints locations, thus overlooking critical appearance cues on entire body regions. Finally, PFD [40], and all previous pose-guided methods including HOReID, incorporate pose information after appearance encoding has occurred, e.g., for local pooling of a spatial feature map, but do not leverage pose data to guide the appearance encoding process as our method do. Indeed, our approach implements a prompt-aware appearance encoding, since it conditions feature extraction on the input keypoints. This enables better feature disentanglement between the target and the occluders from the early feature extraction stages, resulting in more accurate representations. Furthermore, our method is prompt optional, setting it apart for its flexibility from these previous pose-guided methods. Finally, our method can ingest additional negative keypoints, to further disambiguate a multi-person occlusion scenario.", + "bbox": [ + 212, + 387, + 787, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompting in Vision: Recent works have introduced promptable transformers to tackle object segmentation. For instance, SAM [18] employs two networks: a transformer encoder to generate generic spatial features and a lightweight promptable decoder to integrate user inputs for interactive segmentation. In contrast, our approach directly prompts a feature encoder, so that feature extrac", + "bbox": [ + 212, + 765, + 790, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "V. Somers et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ccd792ff4b3f7f393d4b555253e4b3874bb29def2fee108a0847d6dce535b7a7.jpg", + "image_caption": [ + "Fig. 3: Architecture overview of our proposed Keypoint Promptable ReID (KPR) model. The Batch-wise Inter-Person Occlusion (BIPO) data augmentation is first applied to generate artificial inter-person occlusions (Sec. 3.5). The image and the optional positive/negative $(\\oplus/\\ominus)$ prompts are then tokenized and summed (Sec. 3.1). Tokens are then fed to our proposed Multi-Stage feature Fusion (MSF) Swin transformer to generate high-resolution feature maps (3.2). The feature map is then fed to a Part-based Head (PBH), which assigns a part (or the background) to each token with a part classifier, and then averages all tokens of the same part to produce the final $K$ part-based embeddings $\\{f_1, \\dots, f_K\\}$ and their binary visibility scores $\\{v_1, \\dots, v_K\\}$ , with $v_i \\in \\{0,1\\}$ , and visually denoted by $\\{\\mathbf{x}, \\check{\\mathbf{v}}\\}$ (Sec. 3.3). A part $i$ with no token assigned is considered invisible (i.e., $v_i = 0$ ), and ignored when computing two samples' similarity. KPR is illustrated here for $K = 8$ , with a unique color for each body part: $\\{\\text{head}, \\text{torso}, \\text{right/left arm}, \\text{right/left leg}, \\text{right/left feet}\\}$ . Finally, the entire pipeline is trained with two losses: a Part-Prediction (PP) Loss and a ReID Loss (Sec. 3.4)." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 782, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tion is explicitly driven by the prompts. Unlike SimpleClick [23], which prompts a transformer with a single 2D point and a previous mask for iterative segmentation, our model ingests multiple keypoints concurrently, enhancing body part localization and representation by leveraging their semantic information.", + "bbox": [ + 212, + 482, + 787, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 565, + 380, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall architecture of our model KPR is illustrated in Fig. 3 and is based on a Swin transformer backbone [24]. The next subsections first describe the various components of the architecture. We then present the training process, including our proposed data augmentation. Finally, we describe the inference procedure and conclude with an intuitive discussion of our design choices.", + "bbox": [ + 212, + 595, + 787, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Tokenization", + "text_level": 1, + "bbox": [ + 215, + 694, + 369, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The tokenization procedure coverts raw image and keypoints information into $C_i$ -dimensional tokens, to be subsequently processed by the Swin encoder.", + "bbox": [ + 212, + 719, + 785, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image tokenization: The input image, with dimensions $H \\times W$ , undergoes tokenization using the conventional Swin patch embed module. This process generates $H_{t} \\times W_{t} = \\frac{H}{4} \\times \\frac{W}{4}$ spatial tokens of dimensions $C_{i}$ by embedding each $4 \\times 4$ image patch through a linear layer.", + "bbox": [ + 212, + 750, + 785, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Keypoints tokenization: As previously mentioned, our model uses two types of semantic keypoints as prompts: positive keypoints indicating the ReID", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Keypoint Promptable Re-Identification", + "bbox": [ + 465, + 114, + 727, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "target and negative keypoints indicating other individuals. We organize these prompt keypoints into $\\mathrm{K} + 1$ groups: the first group contains all negative keypoints, and the remaining $\\mathrm{K}$ groups contain subsets of the positive keypoints. Each subset of positive keypoints actually corresponds to a semantic body region. For instance, with $K = 8$ , we define groups like: $\\{\\text{head, torso, right/left arm, right/left leg, and right/left feet}\\}$ . For each group $i \\in \\{1, \\dots, K + 1\\}$ , a unique heatmap $M_i$ of size $H \\times W$ is created by drawing a Gaussian 2D kernel at each of its keypoints's center location $(x, y)$ . The kernel's standard deviation is set to a fraction $\\alpha_G$ of the image width. Invisible keypoints are ignored. These $K + 1$ heatmaps are then concatenated into a global spatial tensor $M$ of size $H \\times W \\times (K + 1)$ . Finally, each $4 \\times 4$ patch in $M$ is linearly projected to produce prompt tokens of dimension $C_i$ . Like the image tokenizer, the keypoints tokenizer outputs therefore $H_t \\times W_t$ spatial tokens of dimension $C_i$ .", + "bbox": [ + 212, + 146, + 787, + 342 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Token fusion: Images and keypoints tokens are fused by summing tokens at the same spatial location, producing $H_{t} \\times W_{t}$ spatial tokens with dimension $C_{i}$ . These tokens are then forwarded to the Swin-based encoder.", + "bbox": [ + 212, + 342, + 787, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Design motivations: Prompted keypoints are organized into $K + 1$ channels to preserve semantic meaning before tokenization. It allows the model to differentiate positive and negative prompts, and to distinguish prompt information of different body parts, for better part localization. Prompt-aware appearance encoding: Moreover, feeding the network with keypoint information before the encoding stage is critical in enabling the network to disentangle the ReID target features from other persons' features. Finally, this design offers the flexibility to support segmentation mask prompts by replacing $M$ with a double-channel map of positive and negative segmentation masks. Regarding token fusion, the sum operation maintains the input dimension for the Swin backbone, permitting the use of pre-trained weights from the image-only model. Prompt optionality: This design also offers the flexibility of running the model without prompts, by processing the image tokens alone, with no architectural adjustments required.", + "bbox": [ + 212, + 388, + 787, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 MSF Swin Encoder", + "text_level": 1, + "bbox": [ + 215, + 604, + 423, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A Swin-transformer [24] is employed as the backbone feature extractor for its great performance in vision tasks and its ability to capture high-frequency details, which is beneficial to ReID [49]. Moreover, we propose to enhance Swin with a Multi-Stage feature Fusion strategy, denoted as \"MSF\", to output feature maps with a high spatial resolution. Indeed, similar to CNN, each Swin stage reduces the number of tokens by a factor of 4 while doubling the channel dimension. To obtain a high-resolution token map, the outputs of all stages are bilinearly upsampled to the same resolution $H_{t} \\times W_{t}$ as the output of the tokenization module, then concatenated into a single feature map along the channel dimension, and finally passed through a linear layer to obtain the final spatial token map of size $H_{t} \\times W_{t} \\times C_{o}$ . Different from the standard Swin, our MSF Swin therefore preserves the input tokens resolution. As pointed out in prior works [26,57], a high-resolution feature map enriches the granularity of the features and brings significant performance improvements (see Tab. 3, Exp. 3).", + "bbox": [ + 212, + 628, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "V. Somers et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Part-based Head", + "text_level": 1, + "bbox": [ + 215, + 146, + 401, + 159 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The encoder yields $H_{t} \\times W_{t}$ tokens, which are sent to the Part-based Head (PBH) module, as illustrated in Fig. 3. Each token is first labeled either as one of the $K$ body parts or as background, by a token-wise part classifier. The part classifier is implemented as a linear layer with $K + 1$ output logits followed by the softmax operation, and is applied individually on each token. The assignment of each token to one of the $K$ body parts (or the background) is therefore formulated as a classification task with $K + 1$ classes. We term the probability map of size $H_{t} \\times W_{t} \\times (K + 1)$ produced by this part classifier over all tokens as \"part-attention maps\". These maps are showcased as colored heatmaps in Fig. 1a and Fig. 6b with one color per channel (the background channel is ignored). Finally, tokens having the same assigned body part are averaged together, resulting in $K$ part-based embeddings $\\{f_{1}, \\dots, f_{K}\\}$ of size $C_{o}$ . Background tokens are ignored. Parts with no tokens assigned are considered invisible and given a visibility score $v_{i} = 0$ . The corresponding invisible embedding $f_{i}$ , with $v_{i} = 0$ , is therefore ignored for downstream computation. The PBH module thus produces $K$ part-based embeddings $\\{f_{1}, \\dots, f_{K}\\}$ and their binary visibility scores $\\{v_{1}, \\dots, v_{K}\\}$ .", + "bbox": [ + 212, + 169, + 787, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human parsing labels: To train the token-wise part classifier, each spatial token is assigned a ground-truth class, i.e., part label, derived from the coarse human parsing labels provided by [34]. These human parsing labels are illustrated on the right of Fig. 3 and in Fig. 4a.", + "bbox": [ + 212, + 411, + 785, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Training Process", + "text_level": 1, + "bbox": [ + 214, + 492, + 401, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The overall objective function used to train the model is:", + "bbox": [ + 214, + 513, + 627, + 529 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {\\text {R e I D}} + \\lambda_ {p p} L _ {p p}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 537, + 785, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $L_{ReID}$ is the ReID loss supervised with identity labels, $L_{pp}$ is the token-wise part-prediction loss, and $\\lambda_{pp}$ is a weight parameter empirically set to 0.3.", + "bbox": [ + 214, + 561, + 785, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Part Prediction Loss: The token-wise part classifier introduced in Sec. 3.3 is supervised with a Part Prediction (PP) Loss that is actually a cross-entropy loss applied on each spatial token as in 34.", + "bbox": [ + 212, + 590, + 787, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Re-identification loss: As a ReID objective, we employ the GiLt Loss [34], that is specifically designed to train part-based ReID models. It has the benefit of being robust to occlusions and non-discriminative body parts. GiLt combines an identity loss [26] and a batch-hard triplet loss [13]. The distance function employed for the triplet loss is the average of all local part-based distances.", + "bbox": [ + 212, + 636, + 787, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5 BIPO Data Augmentation", + "text_level": 1, + "bbox": [ + 214, + 732, + 480, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since most ReID datasets feature limited occlusion in their training set, a model often struggles to handle the occlusions it encounters during testing. To address this, we propose a novel Batch-wise Inter-Person Occlusion data augmentation (BIPO) to generate artificial inter-person occlusions on training images,", + "bbox": [ + 212, + 756, + 785, + 820 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Keypoint Promptable Re-Identification", + "bbox": [ + 464, + 114, + 727, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "Inspired by the copy-paste technique for segmentation [10].", + "bbox": [ + 217, + 823, + 624, + 842 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f920778012ac8ec4370cc21f7bcba16027afeb32e5903055918ca03bf4ed620a.jpg", + "image_caption": [ + "Fig. 4: (a) Our proposed Batch-wise Inter-Person Occlusion (BIPO) data augmentation creates artificial person occlusions that are consistent across image, prompt, and human parsing labels. BIPO is crucial to enforce the model to rely on the input prompts. (b) Four identities with their keypoints labels from our Occ-PTrack dataset." + ], + "image_footnote": [], + "bbox": [ + 0, + 143, + 163, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a2200e0b78efa351fdf741dffc48eda9af5d1d4c8a77abb1c1ffaeeed093d1c0.jpg", + "image_caption": [ + "(a) BIPO" + ], + "image_footnote": [], + "bbox": [ + 251, + 143, + 411, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ff517f87e4513a9e9f3ada78b005867e1cffe85282c53d2ca0a87b31fee112e9.jpg", + "image_caption": [ + "(b) Occluded PoseTrack-ReID" + ], + "image_footnote": [], + "bbox": [ + 485, + 143, + 736, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/abf05153db381133534dd36eaef671a57cf8d8e7572d428bd7274a85f83ab5b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 826, + 145, + 985, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "prompts, and human parsing labels. When provided with a training sample, BIPO randomly selects a different person's image (the occluder) from the same training batch, contours it with a segmentation mask derived from the human parsing labels, and overlays it on the main image. The human parsing label and keypoints prompt of the training image are then updated accordingly. Finally, all positive keypoints from the occluder are added to the training prompt to serve as additional negative points.", + "bbox": [ + 212, + 383, + 787, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Design motivations: Given that a triplet loss with batch-hard mining [13] is employed during training, using images from the same training batch to generate occluders yields hard positive and negative samples for mining triplets, thereby creating challenging scenarios for contrastive learning. BIPO is illustrated in Fig. 4a and its significance for KPR is further discussed in Sec. 3.7.", + "bbox": [ + 212, + 489, + 787, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.6 Inference", + "text_level": 1, + "bbox": [ + 214, + 585, + 339, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "At inference, KPR processes each input image together with their respective keypoints prompts, and produces $K$ part-based representations $\\{f_1,\\dots,f_K\\}$ of the ReID target and their binary visibility scores $\\{v_{1},\\ldots ,v_{K}\\}$ . Similar to previous part-based works [45, 57], the global distance $dist_{\\mathrm{global}}^{gg}$ between a query image $q$ and a gallery image $g$ is computed as the average cosine distance $dist_{\\mathrm{cos}}$ of the part-based embeddings $f_{i}$ that are visible in both samples:", + "bbox": [ + 212, + 608, + 787, + 700 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nd i s t _ {\\text {g l o b a l}} ^ {q g} = \\frac {\\sum_ {i \\in \\{1 , \\dots , K \\}} \\left(v _ {i} ^ {q} \\cdot v _ {i} ^ {g} \\cdot d i s t _ {\\cos} \\left(f _ {i} ^ {q} , f _ {i} ^ {g}\\right)\\right)}{\\sum_ {i \\in \\{1 , \\dots , K \\}} \\left(v _ {i} ^ {q} \\cdot v _ {i} ^ {g}\\right)}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 708, + 785, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.7 KPR Pipeline Motivation and Intuition", + "text_level": 1, + "bbox": [ + 214, + 786, + 589, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our pipeline comprises two closely related architectural components. First, the prompt tokenizer, which encodes semantic keypoints information to disambiguate", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "V. Somers et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the ReID target, thereby providing prior information about body part location to drive feature extraction. Second, the Part-based Head (PBH), responsible for localizing the body parts of the target and constructing its part-based features.", + "bbox": [ + 212, + 146, + 784, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While the PBH is explicitly supervised with human parsing labels to learn human topology, there is no dedicated loss incentivizing the model to leverage the input prompt. Additionally, multi-person occlusions are rare in most ReID training sets, giving the model little incentive to rely on these prompts to localize the target person. For this reason, just adding a prompt tokenizer and training the pipeline as is produces a KPR model with weak reliance on the input prompts, and therefore low performance in multi-person occlusion scenarios.", + "bbox": [ + 212, + 191, + 785, + 296 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To address these issues, our proposed BIPO data augmentation plays a critical role. By generating artificial multi-person occlusions during training, the model is compelled to rely more on the prompts, as they represent the sole reliable source of information to distinguish the true ReID target from the occluder.", + "bbox": [ + 212, + 297, + 785, + 357 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Consequently, incorporating BIPO during training is crucial to produce a model that is proficient at leveraging prompt information and effectively handling multi-person occlusions. Finally, designing the Part-based Head without explicit reliance on the prompt offers an additional advantage: it renders our model prompt-optional, enhancing robustness against noisy or missing prompts while still enabling effective localization of the target's body parts.", + "bbox": [ + 212, + 358, + 785, + 448 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 470, + 375, + 488 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Evaluation Benchmarks and Metrics", + "text_level": 1, + "bbox": [ + 214, + 503, + 560, + 517 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We assess our model's performance on three types of person retrieval setups: non-occluded (i.e., holistic) on Market-1501 [53], object occlusions on Occluded-ReID [58] and Partial-ReID [54], and finally multi-person occlusions on our below introduced Occluded PoseTrack-ReID dataset. Occluded-Duke [29] was taken down and removed from our study; check out our GitHub for more information. Two standard person retrieval metrics are reported: the CMC at Rank-1 and the $mAP$ . Performances are evaluated in a single query setting and without re-ranking [55]. Furthermore, we evaluate our model on multi-person pose tracking with the PoseTrack21 [6] dataset. We report the HOTA [25] performance metric, and relevant sub-metrics such as association accuracy ( $AssA$ ), detection accuracy ( $DetA$ ), and the number of identity switches ( $IDs$ ) [20].", + "bbox": [ + 212, + 527, + 787, + 696 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Proposed Occluded PoseTrack-ReID Dataset and Annotations", + "text_level": 1, + "bbox": [ + 214, + 715, + 774, + 732 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We introduce Occluded PoseTrack-ReID (or simply Occ-PTrack), a new ReID dataset we built out of the annotation available with PoseTrack21 [6], a popular video benchmark for multi-person pose tracking, that features keypoints and cross-video identity annotations. Unlike previous ReID datasets focused on street surveillance, Occ-PTrack consists of images from everyday life videos, primarily", + "bbox": [ + 212, + 739, + 785, + 816 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Keypoint Promptable Re-Identification", + "bbox": [ + 464, + 114, + 727, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "7 We use the occluded (non-cropped) images as queries.", + "bbox": [ + 217, + 824, + 594, + 840 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/76c589f3888b3de884e1225ecbcfa7cac48d525cc493864cb0c3476f0f1707e9.jpg", + "table_caption": [ + "Table 1: Comparison of KPR with SOTA methods. Results in Italic are not provided in the original paper but reproduced by ourselves. The $1^{st}/2^{nd}/3^{rd}$ best scores are indicated with $1/2/3$ ." + ], + "table_footnote": [], + "table_body": "
DatasetsMarket-1501Occluded-reIDPartial-reIDOccluded-PoseTrack
TypeHolisticOccluded
Object Occlusions✓✓
Person Occlusions✓
MethodsR-1mAPR-1mAPR-1R-3R-1mAP
BoT94.585.958.452.3--78.869.7
PCB93.881.6----81.771.2
VGTri--81.071.085.793.73--
PVPM--66.859.578.3---
HOReID94.284.980.370.285.391.0--
ISP95.388.6------
PAT95.488.081.672.188.0392.3--
TRANS95.288.9----83.573.4
SOLIDER96.9193.91----84.476.13
SSGR96.189.378.572.9----
FED95.086.386.3179.3384.6---
BPBreed95.789.482.975.2--84.975.5
PFD95.589.783.081.52----
KPRIN w/o prompt95.688.783.378.281.786.085.375.4
KPRIN95.989.685.4279.186.090.092.3182.31
KPRSOL w/o prompt96.6393.0380.078.590.3293.7286.1375.8
KPRSOL96.6293.2284.8382.6190.7194.0190.6281.22
", + "bbox": [ + 277, + 199, + 725, + 498 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "from sports activities, as illustrated in Fig. 4b. Occ-PTrack is divided into a train/test that includes 1000/1411 identities with 17.898/13.412 images from 474/170 videos, which is roughly equivalent in terms of scale to other popular ReID datasets [29,53] (Tab. 5 of the supp. materials). To assess the ReID model's performance in multi-person occlusion scenarios, we select the most cluttered images of each identity in the test set as query samples, and the remaining test images as gallery samples. Cluttered images correspond to multi-person occlusions scenarios where either the front (occluding) or back (occluded) person is the ReID target, to evaluate the model ability to re-identify individuals in both scenarios. We provide further details about our proposed dataset in the supp. materials. Occ-PTrack is challenging as persons within the same video exhibit a high degree of visual resemblance since they often wear similar sports kits.", + "bbox": [ + 212, + 532, + 787, + 715 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Keypoint Annotations for Standard ReID Datasets: Since conventional ReID datasets [29,53,58] do not contain keypoint annotations, we generate pseudo-labels for them with the PifPaf [19] pose estimator. These pseudo-labels include keypoint annotations in COCO [22] format for each image, as illustrated in Fig. 4a. When multiple skeletons are detected in an image, we make the assumption the one with its head closer to the top center part of the image is the intended target, and mark it with a \"is_target\" attribute. Other skeletons are used as negatives. More details are provided in the supplementary materials.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "V. Somers et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 146, + 455, + 161 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We evaluate our model with both the ImageNet [31] (KPRIN) and the SOLIDER [5] (KPRSOL) pre-trained Swin backbones. SOLIDER is a human-centric foundation model trained on the large-scale LUPerson [7] dataset. Following [34], the number of body parts $K$ is set to 8 for Occ-PTrack, and 5 otherwise. The training procedure is mainly adopted from TransReID [12]. We apply our BIPO data augmentation (Sec. 3.5) with a 0.3 probability. The model is trained for 120 epochs with a batch size of 64 and a cosine annealing lr scheduler. For KPRIN/KPRSOL, images are resized to a width of 128 and a height of 256/384. The Gaussian kernel standard deviation $\\alpha_{G}$ (Sec. 3.1) is set to 0.1.", + "bbox": [ + 212, + 169, + 787, + 308 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 215, + 325, + 620, + 340 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Table $\\boxed{1}$ , we present a comparative analysis of KPR, when using prompts (KPR) and not using prompts $^{8}$ (KPR w/o prompt), against all leading occluded ReID methods. Our method emerges as the top performer overall.", + "bbox": [ + 212, + 349, + 784, + 395 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Occluded PoseTrack-ReID: Unlike the other three datasets, Occ-PTrack contain a substantial number of images with multi-person occlusions, making them suitable environments to assess the effectiveness of our promptable method. On this challenging occluded dataset, our proposed model KPR outperforms all previous methods. Our keypoint promptable approach offers several key advantages over the previous pose-guided methods, including: 1) the ability to handle multi-person occlusions; 2) the capability to process both positive and negative keypoints; 3) a prompt-aware appearance encoding, leading to better feature disentanglement between multiple persons; and 4) the flexibility of prompts being optional. We refer readers to Sec. 2 for a detailed discussion about these key differences and a comparison to previous methods. On our proposed Occ-PTrack, we can see that the best-performing method is the foundation model SOLIDER, but it still lies about $6\\%$ behind our KPR solution. Even BPBreID [34], which is specialized for occluded scenarios, demonstrates far lower performance, since it is not designed to address the Multi-Person Ambiguity. Our experiments show that integrating keypoint prompts boosts performance by at least $7.0\\%$ Rank-1 and $6.2\\%$ mAP, since it helps in mitigating the negative impact of MPA.", + "bbox": [ + 212, + 396, + 787, + 652 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Market-1501, Occluded-ReID and Partial-reID: These three datasets primarily feature single individuals in their images, which means multi-person ambiguities are rare, and therefore prompts are not exploited to their full potential. On Market-1501, our method performs on par with the state-of-the-art. However, performance levels have plateaued in recent years on this dataset. In contrast, on Occluded-ReID and Partial-reID, using a prompt noticeably improves performance by helping distinguish the individual features from occluding objects. Additionally, using prompt is beneficial when facing this cross-domain setting. These findings are consistent with earlier studies [8, 39-41] that also leveraged pose estimation to attain strong performance on this dataset.", + "bbox": [ + 212, + 652, + 787, + 803 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Keypoint Promptable Re-Identification", + "bbox": [ + 464, + 114, + 727, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_footnote", + "text": "8 The prompt tokenizer is completely removed at both training and inference.", + "bbox": [ + 217, + 810, + 735, + 824 + ], + "page_idx": 10 + }, + { + "type": "page_footnote", + "text": "9 Occ-ReID and Partial-ReID have no train set, so Market-1501 is used for training.", + "bbox": [ + 218, + 824, + 779, + 839 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/c8eadf0f669b692a07308d8c80fcd1f7d29d7887adc9a31370e8aeeeb93c46c8.jpg", + "table_caption": [ + "Table 2: Multi-person pose tracking performance in videos on PoseTrack21 [6]." + ], + "table_footnote": [], + "table_body": "
MethodHOTA↑DetA↑AssA↑IDs↓
TRMOT 4246.840.955.0-
FairMOT 5153.547.461.4-
Tracktor++ 358.352v.765.4-
CorrTrack + ReID 656.951.364.2-
KPRTrack w/o prompt61.558.465.83830
KPRTrack63.0 (+1.5)59.0 (+0.6)68.3 (+2.5)3344 (-486)
", + "bbox": [ + 217, + 190, + 503, + 282 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4ad084a3f993e83a4eb377bd7d17afb441dacc4a0d8472fe64f3fd5bdab9578e.jpg", + "image_caption": [ + "Fig. 5: Prompt optionality study" + ], + "image_footnote": [], + "bbox": [ + 526, + 143, + 785, + 260 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Multi-Person Pose Tracking in Videos", + "text_level": 1, + "bbox": [ + 215, + 352, + 576, + 368 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We introduce a simple yet robust pose tracker, referred to as KPRTrack, built upon our promptable ReID method. KPRTrack combines the YOLOX object detector with the HRNet pose estimator. For each detected person, it extracts part-based ReID features utilizing their keypoints as positive prompts and keypoints from other individuals within the detected person bbox as negative prompts. Detections are matched in an online fashion with the Hungarian algorithm based on the distance function introduced in Eq. (2). Only tracklet-detection pairs with a distance below a threshold of 0.2 are considered for matching. Therefore, our proposed tracker relies solely on appearance information, meaning spatio-temporal cues are omitted. Pose tracking results are reported in Tab. Despite relying on ReID only, KPRTrack achieves state-of-the-art performance, demonstrating that an occlusion-robust ReID model has strong tracking capabilities. Finally, our two last experiments in Tab. reveal that employing keypoint prompts significantly enhances tracking performance. This improvement can be attributed to the enhanced communication between the detector and the ReID via the keypoint prompts. This leads to improved association accuracy and reduced identity switches, especially when multiple targets intersect paths (i.e., multi-person occlusion scenarios). Further information regarding our baseline tracker is provided in the supplementary materials.", + "bbox": [ + 212, + 377, + 787, + 664 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 686, + 401, + 700 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we analyze the performance impact of our architectural choices.", + "bbox": [ + 215, + 710, + 785, + 727 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Components of KPR As demonstrated in Table 3, all our proposed modules introduced in Sec. 3 play a crucial role in improving the overall ReID performance. Exp. 2 and Exp. 8 demonstrate the importance of the Part-based Head to compare only mutually visible body parts when facing occluded persons, compared to using a single global feature (as in Exp. 1 and 7). PBH degrades performance when applied alone (Exp. 2 vs Exp. 1), but is beneficial when jointly", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "V. Somers et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e436790e18885671bd9fb677bbfb4809b0dd55a13260be73cc20d314f56232e1.jpg", + "table_caption": [ + "Table 3: Ablation study of our proposed KPR architecture. PBH stands for \"Part-based Head\", meaning we use multiple part-based embeddings with visibility scores instead of a single global embedding (Sec. 3.3). MSF stands for the \"Multi-Scale Features\" module added to Swin (Sec. 3.2). Prompt $\\oplus/\\ominus$ refers to positive/negative prompts (Sec. 3.1). BIPO stands for the \"Batch-wise Inter-Person Occlusion\" data augmentation (Sec. 3.5). SOL stands for the pre-trained weights from SOLIDER [5] (Sec. 4.3)." + ], + "table_footnote": [], + "table_body": "
IdMain components of KPRDataset
PBHMSFPromptBIPOSOLOccluded-PoseTrack
⊕∅R-1mAP
184.874.7
2✓83.3 (-1.5)72.2 (-2.5)
3✓✓84.8 (+0.0)76.1 (+1.4)
4✓✓✓88.8 (+4.0)80.4 (+5.7)
5✓✓✓✓90.0 (+5.2)80.7 (+6.0)
6✓✓✓86.1 (+1.3)76.7 (+2.0)
7✓✓✓✓80.4 (+5.1)89.9 (+5.7)
8✓✓✓✓✓92.3 (+6.5)82.3 (+7.6)
9✓✓✓✓✓✓90.6 (+5.8)81.2 (+6.5)
", + "bbox": [ + 320, + 239, + 683, + 395 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "enabled with other modules (Exp. 8 vs Exp 7). Exp. 3 is consistent with claims from previous works [26,57] that increasing the feature map resolution is crucial to achieve fine-grained re-identification and better overall performance. Exp. 4 and 5 illustrate the benefits of positive and negative prompts to disambiguate the intended ReID target from other persons. Exp. 6 demonstrates the importance of our BIPO augmentation to generate artificial occlusions at training, since most occluded ReID benchmarks feature mainly occluded samples in their test set and not in their training set. Exp. 8, in comparison with Exp. 5 and 6, demonstrates how generating these artificial multi-person occlusions at training is crucial to teach the model to rely on the input prompt to identify the intended ReID target among all candidates. The usage of the SOLIDER foundation model as pre-trained weights is illustrated in Exp. 9. Unfortunately, SOLIDER does not enhance performance on the Occ-PTrack dataset, likely because of a domain gap: SOLIDER focuses on street surveillance, while Occ-PTrack predominantly features sports images from handheld cameras. However, as demonstrated in Tab. [1], SOLIDER does improve performance for all other datasets.", + "bbox": [ + 212, + 422, + 787, + 667 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A Prompt-Optional Method In Figure 5 we analyze the test-time robustness of KPR against noisy or missing prompts on Occ-PTrack, by increasingly removing a higher percentage of keypoints from the input prompt. The red curve represents the performance of our proposed promptable KPR model, while the blue constant line illustrates the performance of the non-promptable version, which is unaffected by the proportion of keypoints removed from the input prompt. As illustrated, KPR significantly outperforms the non-promptable baseline with a full prompt and remains competitive with few or no keypoints, highlighting its prompt optional capability. We provide an additional study of the performance w.r.t. the amount of multi-person occlusion in the supplementary materials.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Keypoint Promptable Re-Identification", + "bbox": [ + 464, + 114, + 727, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/df82a5220764d57d243a469ecf4ca035a7c064c43370a24b3e1e50a6af70e90f.jpg", + "image_caption": [ + "(b) Part-attention maps (c) Pose tracking results: all track IDs, depicted by unique skeleton for various input prompts. colors, are maintained after a long occlusion by the front cyclist.", + "Fig.6: Qualitative results on multi-person pose tracking and person retrieval. Red/pastel dots indicate negative/positive prompts. Colored heatmaps illustrates the part-attention maps of KPR described in Sec. 3.3 with one color per body part." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 844, + 368 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Qualitative Assessment Figure demonstrates the impact of keypoint prompting on the part-attention maps, on person retrieval, and on pose tracking. As illustrated, KPR effectively leverages the input prompts to localize the intended target, and is therefore robust to occlusions and multi-person ambiguities.", + "bbox": [ + 212, + 468, + 792, + 530 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 551, + 359, + 566 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work we propose KPR, the first promptable ReID model, designed to address occluded scenarios and multi-person ambiguities arising from bounding-box-based inputs. Our model offers practical flexibility by being prompt-optional, achieving state-of-the-art performance without prompts, and outperforming this baseline with prompts. Furthermore, our model outperforms the state-of-the-art on three popular ReID benchmarks and on our novel Occ-PTrack dataset. We finally demonstrate KPR's potential for pose tracking in videos. Our codebase, keypoint labels and proposed dataset will be released to encourage further research on this new promptable ReID paradigm.", + "bbox": [ + 212, + 580, + 787, + 715 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment: This work has been funded by Sportradar, by the Walloon region project ReconNAssance, and by the FNRS.", + "bbox": [ + 212, + 717, + 785, + 748 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 768, + 323, + 784 + ], + "page_idx": 13 + }, + { + "type": "ref_text", + "text": "1. Adaimi, G., Kreiss, S., Alahi, A.: Deep visual re-identification with confidence. Transportation Research Part C: Emerging Technologies 126, 103067 (May 2021). https://doi.org/10.1016/j.trc.2021.103067", + "bbox": [ + 222, + 797, + 785, + 842 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "V. Somers et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2. Alahi, A., Ramanathan, V., Fei-Fei, L.: Tracking millions of humans in crowded spaces. Group and Crowd Behavior for Computer Vision pp. 115-135 (2017). https://doi.org/10.1016/b978-0-12-809276-7.00007-2", + "3. Bergmann, P., Meinhardt, T., Leal-Taixe, L.: Tracking without bells and whistles. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 941-951. Inst. Electr. Electron. Eng. (IEEE) (Oct 2019). https://doi.org/10.1109/iccv.2019.00103", + "4. Chen, P., Liu, W., Dai, P., Liu, J., Ye, Q., Xu, M., Chen, Q., Ji, R.: Occlude them all: Occlusion-aware attention network for occluded person re-ID. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 11813-11822. Inst. Electr. Electron. Eng. (IEEE), Montreal, QC, Canada (Oct 2021). https://doi.org/10.1109/icc48922.2021.01162", + "5. Chen, W., Xu, X., Jia, J., Luo, H., Wang, Y., Wang, F., Jin, R., Sun, X.: Beyond appearance: A semantic controllable self-supervised learning framework for human-centric visual tasks. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 15050-15061. Inst. Electr. Electron. Eng. (IEEE), Vancouver, Can. (Jun 2023). https://doi.org/10.1109/cvpr52729.2023.01445", + "6. Doering, A., Chen, D., Zhang, S., Schiele, B., Gall, J.: PoseTrack21: A dataset for person search, multi-object tracking and multi-person pose tracking. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 20931-20940. Inst. Electr. Electron. Eng. (IEEE), New Orleans, LA, USA (Jun 2022). https://doi.org/10.1109/cvpr52688.2022.02029", + "7. Fu, D., Chen, D., Bao, J., Yang, H., Yuan, L., Zhang, L., Li, H., Chen, D.: Unsupervised pre-training for person re-identification. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 14745-14754. Inst. Electr. Electron. Eng. (IEEE), Nashville, TN, USA (Jun 2021). https://doi.org/10.1109/cvpr46437.2021.01451", + "8. Gao, S., Wang, J., Lu, H., Liu, Z.: Pose-guided visible part matching for occluded person ReID. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 11741-11749. Inst. Electr. Electron. Eng. (IEEE), Seattle, WA, USA (Jun 2020). https://doi.org/10.1109/cvpr42600.2020.01176", + "9. Ge, Z., Liu, S., Wang, F., Li, Z., Sun, J.: YOLOX: Exceeding YOLO series in 2021. arXiv abs/2107.08430 (2021). https://doi.org/10.48550/arXiv.2107.08430", + "10. Ghiasi, G., Cui, Y., Srinivas, A., Qian, R., Lin, T.Y., Cubuk, E.D., Le, Q.V., Zoph, B.: Simple copy-paste is a strong data augmentation method for instance segmentation. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 2917-2927. Inst. Electr. Electron. Eng. (IEEE), Nashville, TN, USA (Jun 2021). https://doi.org/10.1109/cvpr46437.2021.00294", + "11. Giancola, S., Cioppa, A., Deliège, A., Magera, F., Somers, V., Kang, L., Zhou, X., Barnich, O., De Vleeschouwer, C., Alahi, A., Ghanem, B., Van Droogenbroeck, M., Darwish, A., Maglo, A., Clapés, A., Luyts, A., Boiarov, A., Xarles, A., Orcesi, A., Shah, A., Fan, B., Comandur, B., Chen, C., Zhang, C., Zhao, C., Lin, C., Chan, C.Y., Hui, C.C., Li, D., Yang, F., Liang, F., Da, F., Yan, F., Yu, F., Wang, G., Chan, H.A., Zhu, H., Kan, H., Chu, J., Hu, J., Gu, J., Chen, J., Soares, J.V.B., Theiner, J., De Corte, J., Brito, J.H., Zhang, J., Li, J., Liang, J., Shen, L., Ma, L., Chen, L., Santos Marques, M., Azatov, M., Kasatkin, N., Wang, N., Jia, Q., Pham, Q.C., Ewerth, R., Song, R., Li, R., Gade, R., Debien, R., Zhang, R., Lee, S., Escalera, S., Jiang, S., Odashima, S., Chen, S., Masui, S., Ding, S., Chan, S.w., Chen, S., El-Shabrawy, T., He, T., Moeslund, T.B., Siu, W.C., Zhang, W., Li, W., Wang, X., Tan, X., Li, X., Wei, X., Ye, X., Liu, X., Wang, X., Guo, Y., Zhao, Y., Yu, Y., Li, Y., He, Y., Zhong, Y., Guo, Z., Li, Z.: SoccerNet 2022" + ], + "bbox": [ + 217, + 146, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Keypoint Promptable Re-Identification", + "bbox": [ + 464, + 114, + 727, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "challenges results. In: Proceedings of the 5th International ACM Workshop on Multimedia Content Analysis in Sports. ACM (Oct 2022). https://doi.org/10.1145/3552437.3558545", + "12. He, S., Luo, H., Wang, P., Wang, F., Li, H., Jiang, W.: TransReID: Transformer-based object re-identification. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 14993–15002. Inst. Electr. Electron. Eng. (IEEE), Montreal, QC, Canada (Oct 2021). https://doi.org/10.1109/iccv48922.2021.01474", + "13. Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re-identification. arXiv abs/1703.07737 (2017). https://doi.org/10.48550/arXiv.1703.07737", + "14. Istasse, M., Somers, V., Elancheliyan, P., De, J., Zambrano, D.: DeepSportradar-v2: A multi-sport computer vision dataset for sport understandings. In: Proceedings of the 6th International Workshop on Multimedia Content Analysis in Sports. ACM (Oct 2023). https://doi.org/10.1145/3606038.3616160", + "15. Jia, M., Cheng, X., Zhai, Y., Lu, S., Ma, S., Tian, Y., Zhang, J.: Matching on sets: Conquer occluded person re-identification without alignment. In: AAAI Conf. Artif. Intell. vol. 35, pp. 1673-1681. Association for the Advancement of Artificial Intelligence (AAAI) (May 2021). https://doi.org/10.1609/aaai.v35i2.16260", + "16. Joos, V., Somers, V., Standaert, B.: TrackLab. https://github.com/TrackingLaboratory/tracklab (2024)", + "17. Kiran, M., Praveen, R.G., Nguyen-Meidine, L.T., Belharbi, S., Blais-Morin, L.A., Granger, E.: Holistic guidance for occluded person re-identification. arXiv abs/2104.06524 (2021). https://doi.org/10.48550/arXiv.2104.06524", + "18. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv abs/2304.02643 (2023). https://doi.org/10.48550/arXiv.2304.02643", + "19. Kreiss, S., Bertoni, L., Alahi, A.: PifPaf: Composite fields for human pose estimation. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 11969-11978. Inst. Electr. Electron. Eng. (IEEE), Long Beach, CA, USA (Jun 2019). https://doi.org/10.1109/cvpr.2019.01225", + "20. Leal-Taixe, L., Milan, A., Reid, I., Roth, S., Schindler, K.: MOTChallenge 2015: Towards a benchmark for multi-target tracking. arXiv abs/1504.01942 (2015). https://doi.org/10.48550/arXiv.1504.01942", + "21. Li, Y., He, J., Zhang, T., Liu, X., Zhang, Y., Wu, F.: Diverse part discovery: Occluded person re-identification with part-aware transformer. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 2897-2906. Inst. Electr. Electron. Eng. (IEEE), Nashville, TN, USA (Jun 2021). https://doi.org/10.1109/cvpr46437.2021.00292", + "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: Eur. Conf. Comput. Vis. (ECCV). Lect. Notes Comput. Sci., vol. 8693, pp. 740-755. Springer Int. Publ. (2014). https://doi.org/10.1007/978-3-319-10602-1_48", + "23. Liu, Q., Xu, Z., Bertasius, G., Niethammer, M.: SimpleClick: Interactive image segmentation with simple vision transformers. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 22233-22243. Inst. Electr. Electron. Eng. (IEEE), Paris, Fr. (Oct 2023). https://doi.org/10.1109/iccv51070.2023.02037", + "24. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 9992-10002. Inst. Electr. Electron. Eng. (IEEE), Montreal, QC, Canada (Oct 2021). https://doi.org/10.1109/iccv48922.2021.00986" + ], + "bbox": [ + 217, + 146, + 785, + 838 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "V. Somers et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "25. Luiten, J., Osuep, A., Dendorfer, P., Torr, P., Geiger, A., Leal-Taixe, L., Leibe, B.: HOTA: A higher order metric for evaluating multi-object tracking. Int. J. Comput. Vis. 129(2), 548-578 (Oct 2020). https://doi.org/10.1007/s11263-020-01375-2", + "26. Luo, H., Gu, Y., Liao, X., Lai, S., Jiang, W.: Bag of tricks and a strong baseline for deep person re-identification. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Work. (CVPRW). pp. 1487-1495. Inst. Electr. Electron. Eng. (IEEE), Long Beach, CA, USA (Jun 2019). https://doi.org/10.1109/cvprw.2019.00190", + "27. Ma, Z., Zhao, Y., Li, J.: Pose-guided inter- and intra-part relational transformer for occluded person re-identification. In: Proceedings of the 29th ACM International Conference on Multimedia. ACM (Oct 2021). https://doi.org/10.1145/3474085.3475283", + "28. Mansourian, A.M., Somers, V., De Vleeschouwer, C., Kasaei, S.: Multi-task learning for joint re-identification, team affiliation, and role classification for sports visual tracking. In: Proceedings of the 6th International Workshop on Multimedia Content Analysis in Sports. ACM (Oct 2023). https://doi.org/10.1145/3606038.3616172", + "29. Miao, J., Wu, Y., Liu, P., Ding, Y., Yang, Y.: Pose-guided feature alignment for occluded person re-identification. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 542-551. Inst. Electr. Electron. Eng. (IEEE) (Oct 2019). https://doi.org/10.1109/iccv.2019.00063", + "30. Peng, Y., Hou, S., Cao, C., Liu, X., Huang, Y., He, Z.: Deep learning-based occluded person re-identification: A survey. arXiv abs/2207.14452 (2022). https://doi.org/10.48550/arXiv.2207.14452", + "31. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet large scale visual recognition challenge. Int. J. Comput. Vis. 115(3), 211-252 (Apr 2015). https://doi.org/10.1007/s11263-015-0816-y", + "32. Seidenschwarz, J., Brasó, G., Serrano, V.C., Elezi, I., Leal-Taixe, L.: Simple cues lead to a strong multi-object tracker. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 13813-13823. Inst. Electr. Electron. Eng. (IEEE), Vancouver, Can. (Jun 2023). https://doi.org/10.1109/cvpr52729.2023.01327", + "33. Somers, V., Joos, V., Giancola, S., Cioppa, A., Ghasemzadeh, S.A., Magera, F., Standaert, B., Mansourian, A.M., Zhou, X., Kasaei, S., Ghanem, B., Alahi, A., Van Droogenbroeck, M., De Vleeschouwer, C.: SoccerNet game state reconstruction: End-to-end athlete tracking and identification on a minimap. In: IEEE Int. Conf. Comput. Vis. Pattern Recognit. Work. (CVPRW), CVsports. Seattle, WA, USA (Jun 2024). https://doi.org/10.48550/ARXIV.2404.11335", + "34. Somers, V., Vleeschouwer, C.D., Alahi, A.: Body part-based representation learning for occluded person re-identification. In: IEEE/CVF Winter Conf. Appl. Comput. Vis. (WACV). pp. 1613–1623. Inst. Electr. Electron. Eng. (IEEE), Waikoloa, HI, USA (Jan 2023). https://doi.org/10.1109/wacv56688.2023.00166", + "35. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 5686-5696. Inst. Electr. Electron. Eng. (IEEE), Long Beach, CA, USA (Jun 2019). https://doi.org/10.1109/cvpr.2019.00584", + "36. Sun, Y., Zheng, L., Yang, Y., Tian, Q., Wang, S.: Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline). In: Eur. Conf. Comput. Vis. (ECCV). Lect. Notes Comput. Sci., vol. 11208, pp. 501-518. Springer Int. Publ. (2018). https://doi.org/10.1007/978-3-030-01225-0_30" + ], + "bbox": [ + 215, + 146, + 787, + 842 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Keypoint Promptable Re-Identification", + "bbox": [ + 464, + 114, + 727, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Tan, H., Liu, X., Yin, B., Li, X.: MHSA-net: Multihead self-attention network for occluded person re-identification. IEEE Trans. Neural Networks Learn. Syst. 34(11), 8210-8224 (Nov 2023). https://doi.org/10.1109/tnnls.2022.3144163", + "38. Van Zandycke, G., Somers, V., Istasse, M., Don, C.D., Zambrano, D.: DeepSportradar-v1: Computer vision dataset for sports understanding with high quality annotations. In: Proceedings of the 5th International ACM Workshop on Multimedia Content Analysis in Sports. ACM (Oct 2022). https://doi.org/10.1145/3552437.3555699", + "39. Wang, G., Yang, S., Liu, H., Wang, Z., Yang, Y., Wang, S., Yu, G., Zhou, E., Sun, J.: High-order information matters: Learning relation and topology for occluded person re-identification. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 6448-6457. Inst. Electr. Electron. Eng. (IEEE), Seattle, WA, USA (Jun 2020). https://doi.org/10.1109/cvpr42600.2020.00648", + "40. Wang, T., Liu, H., Song, P., Guo, T., Shi, W.: Pose-guided feature disentangling for occluded person re-identification based on transformer. In: AAAI Conf. Artif. Intell. vol. 36, pp. 2540-2549. Association for the Advancement of Artificial Intelligence (AAAI) (Jun 2022). https://doi.org/10.1609/aaai.v36i3.20155", + "41. Wang, Z., Zhu, F., Tang, S., Zhao, R., He, L., Song, J.: Feature erasing and diffusion network for occluded person re-identification. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 4744-4753. Inst. Electr. Electron. Eng. (IEEE), New Orleans, LA, USA (Jun 2022). https://doi.org/10.1109/cvpr52688.2022.00471", + "42. Wang, Z., Zheng, L., Liu, Y., Li, Y., Wang, S.: Towards real-time multi-object tracking. In: Eur. Conf. Comput. Vis. (ECCV). Lect. Notes Comput. Sci., vol. 12356, pp. 107-122. Springer Int. Publ. (2020). https://doi.org/10.1007/978-3-030-58621-8_7", + "43. Wei, L., Zhang, S., Gao, W., Tian, Q.: Person transfer GAN to bridge domain gap for person re-identification. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 79–88. Inst. Electr. Electron. Eng. (IEEE), Salt Lake City, UT, USA (Jun 2018). https://doi.org/10.1109/cvpr.2018.00016", + "44. Yan, C., Pang, G., Jiao, J., Bai, X., Feng, X., Shen, C.: Occluded person re-identification with single-scale global representations. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 11855-11864. Inst. Electr. Electron. Eng. (IEEE), Montreal, QC, Canada (Oct 2021). https://doi.org/10.1109/iccv48922.2021.01166", + "45. Yang, J., Zhang, J., Yu, F., Jiang, X., Zhang, M., Sun, X., Chen, Y., Zheng, W.S.: Learning to know where to see: A visibility-aware approach for occluded person re-identification. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 11865-11874. Inst. Electr. Electron. Eng. (IEEE), Montreal, QC, Canada (Oct 2021). https://doi.org/10.1109/iccv48922.2021.01167", + "46. Yang, Q., Wang, P., Fang, Z., Lu, Q.: Focus on the visible regions: Semantic-guided alignment model for occluded person re-identification. Sensors 20(16), 1-15 (Aug 2020). https://doi.org/10.3390/s20164431", + "47. Ye, M., Shen, J., Lin, G., Xiang, T., Shao, L., Hoi, S.C.H.: Deep learning for person re-identification: A survey and outlook. IEEE Trans. Pattern Anal. Mach. Intell. 44(6), 2872–2893 (Jun 2022). https://doi.org/10.1109/tpami.2021.3054775", + "48. Zang, X., Li, G., Gao, W., Shu, X.: Learning to disentangle scenes for person re-identification. Image Vis. Comput. 116, 104330 (Dec 2021). https://doi.org/10.1016/j.imavis.2021.104330", + "49. Zhang, G., Zhang, Y., Zhang, T., Li, B., Pu, S.: PHA: Patch-wise high-frequency augmentation for transformer-based person re-identification. In: IEEE/CVF Conf." + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "V. Somers et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Comput. Vis. Pattern Recognit. (CVPR). pp. 14133-14142. Inst. Electr. Electron. Eng. (IEEE), Vancouver, Can. (Jun 2023). https://doi.org/10.1109/cvpr52729.2023.01358", + "50. Zhang, Y., Sun, P., Jiang, Y., Yu, D., Weng, F., Yuan, Z., Luo, P., Liu, W., Wang, X.: ByteTrack: Multi-object tracking by associating every detection box. In: Eur. Conf. Comput. Vis. (ECCV). Lect. Notes Comput. Sci., vol. 13682, pp. 1-21. Springer Nat. Switz. (2022). https://doi.org/10.1007/978-3-031-20047-2_1, https://doi.org/10.1007/978-3-031-20047-2_1", + "51. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: FairMOT: On the fairness of detection and re-identification in multiple object tracking. Int. J. Comput. Vis. 129(11), 3069–3087 (Sept 2021). https://doi.org/10.1007/s11263-021-01513-4", + "52. Zheng, K., Lan, C., Zeng, W., Liu, J., Zhang, Z., Zha, Z.J.: Pose-guided feature learning with knowledge distillation for occluded person re-identification. arXiv abs/2108.00139 (2021). https://doi.org/10.48550/arXiv.2108.00139", + "53. Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., Tian, Q.: Scalable person re-identification: A benchmark. In: IEEE Int. Conf. Comput. Vis. (ICCV). pp. 1116-1124. Inst. Electr. Electron. Eng. (IEEE), Santiago, Chile (Dec 2015). https://doi.org/10.1109/iccv.2015.133", + "54. Zheng, W.S., Li, X., Xiang, T., Liao, S., Lai, J., Gong, S.: Partial person re-identification. In: IEEE Int. Conf. Comput. Vis. (ICCV). pp. 4678-4686. Inst. Electr. Electron. Eng. (IEEE), Santiago, Chile (Dec 2015). https://doi.org/10.1109/iccv.2015.531", + "55. Zhong, Z., Zheng, L., Cao, D., Li, S.: Re-ranking person re-identification with k-reciprocal encoding. In: IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 3652-3661. Inst. Electr. Electron. Eng. (IEEE), Honolulu, HI, USA (Jul 2017). https://doi.org/10.1109/cvpr.2017.389", + "56. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. In: AAAI Conf. Artif. Intell. vol. 34, pp. 13001-13008. Association for the Advancement of Artificial Intelligence (AAAI) (Apr 2020). https://doi.org/10.1609/aaai.v34i07.7000", + "57. Zhu, K., Guo, H., Liu, Z., Tang, M., Wang, J.: Identity-guided human semantic parsing for person re-identification. In: Eur. Conf. Comput. Vis. (ECCV). Lect. Notes Comput. Sci., vol. 12348, pp. 346-363. Springer Int. Publ. (2020). https://doi.org/10.1007/978-3-030-58580-8_21", + "58. Zhuo, J., Chen, Z., Lai, J., Wang, G.: Occluded person re-identification. arXiv abs/1804.02792 (2018). https://doi.org/10.48550/arXiv.1804.02792" + ], + "bbox": [ + 215, + 146, + 787, + 660 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Keypoint Promptable Re-Identification", + "bbox": [ + 465, + 114, + 727, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Keypoint Promptable Re-Identification/21228ec1-ee24-4ba9-acb3-211d122a99e2_model.json b/2024/Keypoint Promptable Re-Identification/21228ec1-ee24-4ba9-acb3-211d122a99e2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c8676e6c862f1f35c8c8d0803681cc8dfc0a8c0c --- /dev/null +++ b/2024/Keypoint Promptable Re-Identification/21228ec1-ee24-4ba9-acb3-211d122a99e2_model.json @@ -0,0 +1,2493 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.278, + 0.142, + 0.726, + 0.163 + ], + "angle": 0, + "content": "Keypoint Promptable Re-Identification" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.189, + 0.788, + 0.208 + ], + "angle": 0, + "content": "Vladimir Somers1,2,3, Alexandre Alahi1, and Christophe De Vleeschouwer2" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.217, + 0.725, + 0.231 + ], + "angle": 0, + "content": "1 Ecole Polytechnique Fédérale de Lausanne (EPFL), Switzerland" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.231, + 0.699, + 0.246 + ], + "angle": 0, + "content": "2 Université Catholique de Louvain (UCLouvain), Belgium" + }, + { + "type": "text", + "bbox": [ + 0.415, + 0.246, + 0.59, + 0.259 + ], + "angle": 0, + "content": "3 Sportradar, Switzerland" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.298, + 0.744, + 0.604 + ], + "angle": 0, + "content": "Abstract. Occluded Person Re-Identification (ReID) is a metric learning task that involves matching occluded individuals based on their appearance. While many studies have tackled occlusions caused by objects, multi-person occlusions remain less explored. In this work, we identify and address a critical challenge overlooked by previous occluded ReID methods: the Multi-Person Ambiguity (MPA) arising when multiple individuals are visible in the same bounding box, making it impossible to determine the intended ReID target among the candidates. Inspired by recent work on prompting in vision, we introduce Keypoint Promptable ReID (KPR), a novel formulation of the ReID problem that explicitly complements the input bounding box with a set of semantic keypoints indicating the intended target. Since promptable re-identification is an unexplored paradigm, existing ReID datasets lack the pixel-level annotations necessary for prompting. To bridge this gap and foster further research on this topic, we introduce Occluded PoseTrack-ReID, a novel ReID dataset with keypoints labels, that features strong inter-person occlusions. Furthermore, we release custom keypoints labels for four popular ReID benchmarks. Experiments on person retrieval, but also on pose tracking, demonstrate that our method systematically surpasses previous state-of-the-art approaches on various occluded scenarios. Our code, dataset and annotations are available at https://github.com/V1Somers/Keypoint_prompts_reidentification" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.616, + 0.741, + 0.631 + ], + "angle": 0, + "content": "Keywords: Person Re-Identification \\(\\cdot\\) Vision Prompting \\(\\cdot\\) Pose Tracking" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.657, + 0.377, + 0.673 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Person re-identification (ReID) 47 is a challenging retrieval task that involves matching a query image of a person of interest with other person images. ReID finds wide-ranging applications in multi-object tracking 32, pedestrian flow analysis 2, sport understanding 11,14,28,33,38 and video-surveillance 53. However, ReID is a difficult task due to various factors including inaccurate bounding boxes, luminosity changes, poor image quality, and occlusions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.843 + ], + "angle": 0, + "content": "In recent years, there has been a growing interest in addressing the occluded re-identification task [30], where the ReID target might be occluded by various objects or other people. In this work, we identify and explicitly address a critical challenge that has been overlooked by previous occluded ReID methods [34, 41]:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "V. Somers et al." + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.145, + 0.648, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.239, + 0.615, + 0.251 + ], + "angle": 0, + "content": "(a) Keypoint Promptable ReID (KPR)" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.252, + 0.761, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.356, + 0.789, + 0.427 + ], + "angle": 0, + "content": "Fig. 1: Overview of our proposed Keypoint Promptable ReID (KPR) method. KPR takes an image with keypoints prompts as input and produces part-based features of the prompted target. The prompt instructs the model to focus on a specific individual, i.e. the back blue jacket man (b) or the front black t-shirt man (c) in this example. Colored dots illustrate the positive keypoints prompt, with one color per body part." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.789, + 0.556 + ], + "angle": 0, + "content": "the presence of Multi-Person Ambiguity (MPA) inherent to bounding box annotations. The multi-person ambiguity arises when a bounding box image contains multiple individuals, making it challenging even for humans to accurately identify the intended ReID target among all the candidates. Figure 2a shows how MPA can negatively impact the re-identification model, causing feature mixing among different individuals or even a focus on an unintended person." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.559, + 0.79, + 0.634 + ], + "angle": 0, + "content": "To tackle this multi-person ambiguity, it is necessary to incorporate additional pixel-level information, such as keypoints or segmentation masks, provided by an upstream human operator or vision model. This additional data plays a vital role in assisting the downstream ReID method to distinguish the intended target from other candidates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.639, + 0.789, + 0.746 + ], + "angle": 0, + "content": "Inspired by recent advances in promptable\\(^{4}\\) methods for vision transformers (e.g., SAM [18]), we propose to explicitly use additional semantic\\(^{5}\\) keypoints as inputs to disambiguate images with multiple individuals. Semantic keypoints are suitable for two crucial use cases: 1) manual prompting by a human operator with a few clicks on an image, and 2) automated prompting by an upstream pose estimation model. Nevertheless, our architecture could be easily extended to support segmentation masks as prompts." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.77, + 0.787, + 0.812 + ], + "angle": 0, + "content": "4 A \"prompt\" typically refers to an input instruction that guides the model's response. In vision tasks, e.g. segmentation, prompts can be any kind of data that specifies what the model should focus on [18]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.812, + 0.787, + 0.843 + ], + "angle": 0, + "content": "5 Semantic keypoints are distinct points within a person image that carry semantic information about specific body parts, typically obtained with pose estimation [22]." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.77, + 0.787, + 0.843 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.466, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Keypoint Promptable Re-Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.493, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.23, + 0.49, + 0.252 + ], + "angle": 0, + "content": "(a) Existing ReID models (e.g. BPBreID 34) are confused by MPA" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.144, + 0.807, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.23, + 0.786, + 0.253 + ], + "angle": 0, + "content": "(b) KPR explicitly addresses MPA with positive (pastel) and negative (red) keypoints prompt" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.263, + 0.788, + 0.293 + ], + "angle": 0, + "content": "Fig. 2: Person retrieval with Multi-Person Ambiguity (MPA). Green/red borders are correct/incorrect matches. Red/pastel dots indicate negative/positive prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.784, + 0.374 + ], + "angle": 0, + "content": "Prior research [8,39,40] on pose-guided methods has explored the use of additional information from external pose estimation models. However, none of these works explicitly tackled the multi-pedestrian ambiguity issue." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.789, + 0.62 + ], + "angle": 0, + "content": "Following above insights, we propose KPR, a novel Keypoint-Promptable transformer for part-based person Re-identification, illustrated in Fig. 1a. KPR is a transformer-based model that takes an image along with semantic keypoints as input. It then produces part-based features, each representing a distinct body part of the ReID target, along with their respective visibility scores. Part-based methods have demonstrated superior performance [34,57] in re-identifying occluded individuals because they utilize only the visible parts for comparison. Our method can process both positive and negative keypoints, which respectively represent the target and non-target pedestrians. As illustrated in Fig. 1b, prompts can be used to re-identify the desired (front or back) person in case of occlusion. Furthermore, KPR is designed to be prompt optional to offer more practical flexibility. This means the same model can be used without prompt on non-ambiguous images, or with prompt when dealing with occlusions, while consistently achieving state-of-the-art performance in both cases. To demonstrate the advantages of keypoint prompts for re-identification, we evaluate KPR on two tasks: person retrieval and multi-person pose tracking." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.788, + 0.698 + ], + "angle": 0, + "content": "Moreover, since no current ReID dataset is provided with explicit target identification in case of multi-person ambiguity, we introduce a new dataset called Occluded PoseTrack-ReID, nicknamed \"Occ-PTrack\", and derived from the PoseTrack21 dataset [6]. This dataset offers high-quality keypoints annotations and features a substantial number of images with multi-person occlusions." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.701, + 0.741, + 0.716 + ], + "angle": 0, + "content": "Overall, we summarize the main contributions of our work as follows:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.72, + 0.787, + 0.779 + ], + "angle": 0, + "content": "1. We propose \\( KPR \\), a novel prompt-optional model for part-based ReID. To our knowledge, we are the first to introduce a visual prompting mechanism for person re-identification, and the first to explicitly address the ambiguity caused by multi-person occlusions." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.78, + 0.787, + 0.839 + ], + "angle": 0, + "content": "2. We introduce Occ-PTrack: the first multi-person occluded ReID dataset with explicit target identification through manual keypoints annotations. Furthermore, we propose new keypoint annotations for four popular re-identification datasets." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.72, + 0.787, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "V. Somers et al." + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.147, + 0.786, + 0.192 + ], + "angle": 0, + "content": "3. Our method outperforms all previous state-of-the-art methods on the occluded ReID task, demonstrating the effectiveness of the keypoint promptable approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.254 + ], + "angle": 0, + "content": "Our ambition is to advocate for a shift from the ambiguous bounding-box approach to ReID and promote a keypoint-centric paradigm. To this end, we publicly release our codebase, proposed dataset, and annotations, to provide a common setup for evaluating promptable ReID methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.278, + 0.389, + 0.294 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.787, + 0.386 + ], + "angle": 0, + "content": "Multi-Person Ambiguity: A previous study [41] addressed a related issue called \"Non-Target Pedestrians (NTP),\" similar to our concept of \"Multi-Person Ambiguity\" (MPA). However, their focus was on mitigating the negative impact of NTP during training. In contrast, we provide a method to explicitly disambiguate the intended ReID target during both training and testing phases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.789, + 0.765 + ], + "angle": 0, + "content": "Pose-guided ReID: We detail here the fundamental difference between our novel keypoint promptable approach and the traditional pose-guided methods. Although none of the previous works explicitly address the multi-person ambiguity (MPA) issue, we identify two categories within these pose-guided approaches: 1) those inherently unable to address MPA [48,15,27,29,34,44,45], and 2) those with the potential to overcome MPA [39,40]. In the first category, methods like BPBreID [34] or Pirt [27] utilize pose labels only during training, and can not leverage additional information at test time to disambiguate multiple persons. Moreover, methods relying on horizontal stripes (VGTri [45], PGFA [29]) or the affinity fields of a pose model (PVPM [8]), are meant to tackle object-occlusions, but struggle to disambiguate multiple persons. In the second category, methods like HOReID [39] and PFD [40] directly employ keypoints from a single person, enabling focus on a specified target. However, HOReID [39] is restricted to utilizing spatial features at the exact keypoints locations, thus overlooking critical appearance cues on entire body regions. Finally, PFD [40], and all previous pose-guided methods including HOReID, incorporate pose information after appearance encoding has occurred, e.g., for local pooling of a spatial feature map, but do not leverage pose data to guide the appearance encoding process as our method do. Indeed, our approach implements a prompt-aware appearance encoding, since it conditions feature extraction on the input keypoints. This enables better feature disentanglement between the target and the occluders from the early feature extraction stages, resulting in more accurate representations. Furthermore, our method is prompt optional, setting it apart for its flexibility from these previous pose-guided methods. Finally, our method can ingest additional negative keypoints, to further disambiguate a multi-person occlusion scenario." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.792, + 0.842 + ], + "angle": 0, + "content": "Prompting in Vision: Recent works have introduced promptable transformers to tackle object segmentation. For instance, SAM [18] employs two networks: a transformer encoder to generate generic spatial features and a lightweight promptable decoder to integrate user inputs for interactive segmentation. In contrast, our approach directly prompts a feature encoder, so that feature extrac" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.466, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Keypoint Promptable Re-Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.145, + 0.784, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.457 + ], + "angle": 0, + "content": "Fig. 3: Architecture overview of our proposed Keypoint Promptable ReID (KPR) model. The Batch-wise Inter-Person Occlusion (BIPO) data augmentation is first applied to generate artificial inter-person occlusions (Sec. 3.5). The image and the optional positive/negative \\((\\oplus/\\ominus)\\) prompts are then tokenized and summed (Sec. 3.1). Tokens are then fed to our proposed Multi-Stage feature Fusion (MSF) Swin transformer to generate high-resolution feature maps (3.2). The feature map is then fed to a Part-based Head (PBH), which assigns a part (or the background) to each token with a part classifier, and then averages all tokens of the same part to produce the final \\(K\\) part-based embeddings \\(\\{f_1, \\dots, f_K\\}\\) and their binary visibility scores \\(\\{v_1, \\dots, v_K\\}\\), with \\(v_i \\in \\{0,1\\}\\), and visually denoted by \\(\\{\\mathbf{x}, \\check{\\mathbf{v}}\\}\\) (Sec. 3.3). A part \\(i\\) with no token assigned is considered invisible (i.e., \\(v_i = 0\\)), and ignored when computing two samples' similarity. KPR is illustrated here for \\(K = 8\\), with a unique color for each body part: \\(\\{\\text{head}, \\text{torso}, \\text{right/left arm}, \\text{right/left leg}, \\text{right/left feet}\\}\\). Finally, the entire pipeline is trained with two losses: a Part-Prediction (PP) Loss and a ReID Loss (Sec. 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.788, + 0.545 + ], + "angle": 0, + "content": "tion is explicitly driven by the prompts. Unlike SimpleClick [23], which prompts a transformer with a single 2D point and a previous mask for iterative segmentation, our model ingests multiple keypoints concurrently, enhancing body part localization and representation by leveraging their semantic information." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.566, + 0.381, + 0.584 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.675 + ], + "angle": 0, + "content": "The overall architecture of our model KPR is illustrated in Fig. 3 and is based on a Swin transformer backbone [24]. The next subsections first describe the various components of the architecture. We then present the training process, including our proposed data augmentation. Finally, we describe the inference procedure and conclude with an intuitive discussion of our design choices." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.695, + 0.37, + 0.709 + ], + "angle": 0, + "content": "3.1 Tokenization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.749 + ], + "angle": 0, + "content": "The tokenization procedure coverts raw image and keypoints information into \\( C_i \\)-dimensional tokens, to be subsequently processed by the Swin encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.81 + ], + "angle": 0, + "content": "Image tokenization: The input image, with dimensions \\( H \\times W \\), undergoes tokenization using the conventional Swin patch embed module. This process generates \\( H_{t} \\times W_{t} = \\frac{H}{4} \\times \\frac{W}{4} \\) spatial tokens of dimensions \\( C_{i} \\) by embedding each \\( 4 \\times 4 \\) image patch through a linear layer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Keypoints tokenization: As previously mentioned, our model uses two types of semantic keypoints as prompts: positive keypoints indicating the ReID" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "V. Somers et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.343 + ], + "angle": 0, + "content": "target and negative keypoints indicating other individuals. We organize these prompt keypoints into \\(\\mathrm{K} + 1\\) groups: the first group contains all negative keypoints, and the remaining \\(\\mathrm{K}\\) groups contain subsets of the positive keypoints. Each subset of positive keypoints actually corresponds to a semantic body region. For instance, with \\(K = 8\\), we define groups like: \\(\\{\\text{head, torso, right/left arm, right/left leg, and right/left feet}\\}\\). For each group \\(i \\in \\{1, \\dots, K + 1\\}\\), a unique heatmap \\(M_i\\) of size \\(H \\times W\\) is created by drawing a Gaussian 2D kernel at each of its keypoints's center location \\((x, y)\\). The kernel's standard deviation is set to a fraction \\(\\alpha_G\\) of the image width. Invisible keypoints are ignored. These \\(K + 1\\) heatmaps are then concatenated into a global spatial tensor \\(M\\) of size \\(H \\times W \\times (K + 1)\\). Finally, each \\(4 \\times 4\\) patch in \\(M\\) is linearly projected to produce prompt tokens of dimension \\(C_i\\). Like the image tokenizer, the keypoints tokenizer outputs therefore \\(H_t \\times W_t\\) spatial tokens of dimension \\(C_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Token fusion: Images and keypoints tokens are fused by summing tokens at the same spatial location, producing \\( H_{t} \\times W_{t} \\) spatial tokens with dimension \\( C_{i} \\). These tokens are then forwarded to the Swin-based encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.585 + ], + "angle": 0, + "content": "Design motivations: Prompted keypoints are organized into \\( K + 1 \\) channels to preserve semantic meaning before tokenization. It allows the model to differentiate positive and negative prompts, and to distinguish prompt information of different body parts, for better part localization. Prompt-aware appearance encoding: Moreover, feeding the network with keypoint information before the encoding stage is critical in enabling the network to disentangle the ReID target features from other persons' features. Finally, this design offers the flexibility to support segmentation mask prompts by replacing \\( M \\) with a double-channel map of positive and negative segmentation masks. Regarding token fusion, the sum operation maintains the input dimension for the Swin backbone, permitting the use of pre-trained weights from the image-only model. Prompt optionality: This design also offers the flexibility of running the model without prompts, by processing the image tokens alone, with no architectural adjustments required." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.605, + 0.424, + 0.62 + ], + "angle": 0, + "content": "3.2 MSF Swin Encoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.843 + ], + "angle": 0, + "content": "A Swin-transformer [24] is employed as the backbone feature extractor for its great performance in vision tasks and its ability to capture high-frequency details, which is beneficial to ReID [49]. Moreover, we propose to enhance Swin with a Multi-Stage feature Fusion strategy, denoted as \"MSF\", to output feature maps with a high spatial resolution. Indeed, similar to CNN, each Swin stage reduces the number of tokens by a factor of 4 while doubling the channel dimension. To obtain a high-resolution token map, the outputs of all stages are bilinearly upsampled to the same resolution \\( H_{t} \\times W_{t} \\) as the output of the tokenization module, then concatenated into a single feature map along the channel dimension, and finally passed through a linear layer to obtain the final spatial token map of size \\( H_{t} \\times W_{t} \\times C_{o} \\). Different from the standard Swin, our MSF Swin therefore preserves the input tokens resolution. As pointed out in prior works [26,57], a high-resolution feature map enriches the granularity of the features and brings significant performance improvements (see Tab. 3, Exp. 3)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.465, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Keypoint Promptable Re-Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.402, + 0.16 + ], + "angle": 0, + "content": "3.3 Part-based Head" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.412 + ], + "angle": 0, + "content": "The encoder yields \\( H_{t} \\times W_{t} \\) tokens, which are sent to the Part-based Head (PBH) module, as illustrated in Fig. 3. Each token is first labeled either as one of the \\( K \\) body parts or as background, by a token-wise part classifier. The part classifier is implemented as a linear layer with \\( K + 1 \\) output logits followed by the softmax operation, and is applied individually on each token. The assignment of each token to one of the \\( K \\) body parts (or the background) is therefore formulated as a classification task with \\( K + 1 \\) classes. We term the probability map of size \\( H_{t} \\times W_{t} \\times (K + 1) \\) produced by this part classifier over all tokens as \"part-attention maps\". These maps are showcased as colored heatmaps in Fig. 1a and Fig. 6b with one color per channel (the background channel is ignored). Finally, tokens having the same assigned body part are averaged together, resulting in \\( K \\) part-based embeddings \\( \\{f_{1}, \\dots, f_{K}\\} \\) of size \\( C_{o} \\). Background tokens are ignored. Parts with no tokens assigned are considered invisible and given a visibility score \\( v_{i} = 0 \\). The corresponding invisible embedding \\( f_{i} \\), with \\( v_{i} = 0 \\), is therefore ignored for downstream computation. The PBH module thus produces \\( K \\) part-based embeddings \\( \\{f_{1}, \\dots, f_{K}\\} \\) and their binary visibility scores \\( \\{v_{1}, \\dots, v_{K}\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.787, + 0.475 + ], + "angle": 0, + "content": "Human parsing labels: To train the token-wise part classifier, each spatial token is assigned a ground-truth class, i.e., part label, derived from the coarse human parsing labels provided by [34]. These human parsing labels are illustrated on the right of Fig. 3 and in Fig. 4a." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.493, + 0.403, + 0.509 + ], + "angle": 0, + "content": "3.4 Training Process" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.515, + 0.629, + 0.53 + ], + "angle": 0, + "content": "The overall objective function used to train the model is:" + }, + { + "type": "equation", + "bbox": [ + 0.427, + 0.539, + 0.786, + 0.555 + ], + "angle": 0, + "content": "\\[\nL = L _ {\\text {R e I D}} + \\lambda_ {p p} L _ {p p}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.562, + 0.787, + 0.592 + ], + "angle": 0, + "content": "where \\( L_{ReID} \\) is the ReID loss supervised with identity labels, \\( L_{pp} \\) is the token-wise part-prediction loss, and \\( \\lambda_{pp} \\) is a weight parameter empirically set to 0.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.789, + 0.637 + ], + "angle": 0, + "content": "Part Prediction Loss: The token-wise part classifier introduced in Sec. 3.3 is supervised with a Part Prediction (PP) Loss that is actually a cross-entropy loss applied on each spatial token as in 34." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.713 + ], + "angle": 0, + "content": "Re-identification loss: As a ReID objective, we employ the GiLt Loss [34], that is specifically designed to train part-based ReID models. It has the benefit of being robust to occlusions and non-discriminative body parts. GiLt combines an identity loss [26] and a batch-hard triplet loss [13]. The distance function employed for the triplet loss is the average of all local part-based distances." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.733, + 0.481, + 0.749 + ], + "angle": 0, + "content": "3.5 BIPO Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.757, + 0.787, + 0.821 + ], + "angle": 0, + "content": "Since most ReID datasets feature limited occlusion in their training set, a model often struggles to handle the occlusions it encounters during testing. To address this, we propose a novel Batch-wise Inter-Person Occlusion data augmentation (BIPO) to generate artificial inter-person occlusions on training images," + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.625, + 0.843 + ], + "angle": 0, + "content": "Inspired by the copy-paste technique for segmentation [10]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "V. Somers et al." + }, + { + "type": "image", + "bbox": [ + 0.0, + 0.145, + 0.165, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.145, + 0.413, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.277, + 0.362, + 0.289 + ], + "angle": 0, + "content": "(a) BIPO" + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.145, + 0.738, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.277, + 0.7, + 0.289 + ], + "angle": 0, + "content": "(b) Occluded PoseTrack-ReID" + }, + { + "type": "image", + "bbox": [ + 0.827, + 0.146, + 0.986, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.357 + ], + "angle": 0, + "content": "Fig. 4: (a) Our proposed Batch-wise Inter-Person Occlusion (BIPO) data augmentation creates artificial person occlusions that are consistent across image, prompt, and human parsing labels. BIPO is crucial to enforce the model to rely on the input prompts. (b) Four identities with their keypoints labels from our Occ-PTrack dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.489 + ], + "angle": 0, + "content": "prompts, and human parsing labels. When provided with a training sample, BIPO randomly selects a different person's image (the occluder) from the same training batch, contours it with a segmentation mask derived from the human parsing labels, and overlays it on the main image. The human parsing label and keypoints prompt of the training image are then updated accordingly. Finally, all positive keypoints from the occluder are added to the training prompt to serve as additional negative points." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.788, + 0.567 + ], + "angle": 0, + "content": "Design motivations: Given that a triplet loss with batch-hard mining [13] is employed during training, using images from the same training batch to generate occluders yields hard positive and negative samples for mining triplets, thereby creating challenging scenarios for contrastive learning. BIPO is illustrated in Fig. 4a and its significance for KPR is further discussed in Sec. 3.7." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.586, + 0.341, + 0.6 + ], + "angle": 0, + "content": "3.6 Inference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.788, + 0.701 + ], + "angle": 0, + "content": "At inference, KPR processes each input image together with their respective keypoints prompts, and produces \\(K\\) part-based representations \\(\\{f_1,\\dots,f_K\\}\\) of the ReID target and their binary visibility scores \\(\\{v_{1},\\ldots ,v_{K}\\}\\). Similar to previous part-based works [45, 57], the global distance \\(dist_{\\mathrm{global}}^{gg}\\) between a query image \\(q\\) and a gallery image \\(g\\) is computed as the average cosine distance \\(dist_{\\mathrm{cos}}\\) of the part-based embeddings \\(f_{i}\\) that are visible in both samples:" + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.709, + 0.786, + 0.765 + ], + "angle": 0, + "content": "\\[\nd i s t _ {\\text {g l o b a l}} ^ {q g} = \\frac {\\sum_ {i \\in \\{1 , \\dots , K \\}} \\left(v _ {i} ^ {q} \\cdot v _ {i} ^ {g} \\cdot d i s t _ {\\cos} \\left(f _ {i} ^ {q} , f _ {i} ^ {g}\\right)\\right)}{\\sum_ {i \\in \\{1 , \\dots , K \\}} \\left(v _ {i} ^ {q} \\cdot v _ {i} ^ {g}\\right)}. \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.787, + 0.591, + 0.803 + ], + "angle": 0, + "content": "3.7 KPR Pipeline Motivation and Intuition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our pipeline comprises two closely related architectural components. First, the prompt tokenizer, which encodes semantic keypoints information to disambiguate" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.465, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Keypoint Promptable Re-Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.192 + ], + "angle": 0, + "content": "the ReID target, thereby providing prior information about body part location to drive feature extraction. Second, the Part-based Head (PBH), responsible for localizing the body parts of the target and constructing its part-based features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.787, + 0.297 + ], + "angle": 0, + "content": "While the PBH is explicitly supervised with human parsing labels to learn human topology, there is no dedicated loss incentivizing the model to leverage the input prompt. Additionally, multi-person occlusions are rare in most ReID training sets, giving the model little incentive to rely on these prompts to localize the target person. For this reason, just adding a prompt tokenizer and training the pipeline as is produces a KPR model with weak reliance on the input prompts, and therefore low performance in multi-person occlusion scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.787, + 0.358 + ], + "angle": 0, + "content": "To address these issues, our proposed BIPO data augmentation plays a critical role. By generating artificial multi-person occlusions during training, the model is compelled to rely more on the prompts, as they represent the sole reliable source of information to distinguish the true ReID target from the occluder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.787, + 0.449 + ], + "angle": 0, + "content": "Consequently, incorporating BIPO during training is crucial to produce a model that is proficient at leveraging prompt information and effectively handling multi-person occlusions. Finally, designing the Part-based Head without explicit reliance on the prompt offers an additional advantage: it renders our model prompt-optional, enhancing robustness against noisy or missing prompts while still enabling effective localization of the target's body parts." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.472, + 0.376, + 0.489 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.504, + 0.562, + 0.518 + ], + "angle": 0, + "content": "4.1 Evaluation Benchmarks and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.788, + 0.697 + ], + "angle": 0, + "content": "We assess our model's performance on three types of person retrieval setups: non-occluded (i.e., holistic) on Market-1501 [53], object occlusions on Occluded-ReID [58] and Partial-ReID [54], and finally multi-person occlusions on our below introduced Occluded PoseTrack-ReID dataset. Occluded-Duke [29] was taken down and removed from our study; check out our GitHub for more information. Two standard person retrieval metrics are reported: the CMC at Rank-1 and the \\(mAP\\). Performances are evaluated in a single query setting and without re-ranking [55]. Furthermore, we evaluate our model on multi-person pose tracking with the PoseTrack21 [6] dataset. We report the HOTA [25] performance metric, and relevant sub-metrics such as association accuracy (\\(AssA\\)), detection accuracy (\\(DetA\\)), and the number of identity switches (\\(IDs\\)) [20]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.717, + 0.776, + 0.733 + ], + "angle": 0, + "content": "4.2 Proposed Occluded PoseTrack-ReID Dataset and Annotations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.741, + 0.787, + 0.817 + ], + "angle": 0, + "content": "We introduce Occluded PoseTrack-ReID (or simply Occ-PTrack), a new ReID dataset we built out of the annotation available with PoseTrack21 [6], a popular video benchmark for multi-person pose tracking, that features keypoints and cross-video identity annotations. Unlike previous ReID datasets focused on street surveillance, Occ-PTrack consists of images from everyday life videos, primarily" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.595, + 0.841 + ], + "angle": 0, + "content": "7 We use the occluded (non-cropped) images as queries." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "V. Somers et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Comparison of KPR with SOTA methods. Results in Italic are not provided in the original paper but reproduced by ourselves. The \\(1^{st}/2^{nd}/3^{rd}\\) best scores are indicated with \\(1/2/3\\)." + }, + { + "type": "table", + "bbox": [ + 0.279, + 0.2, + 0.727, + 0.499 + ], + "angle": 0, + "content": "
DatasetsMarket-1501Occluded-reIDPartial-reIDOccluded-PoseTrack
TypeHolisticOccluded
Object Occlusions✓✓
Person Occlusions✓
MethodsR-1mAPR-1mAPR-1R-3R-1mAP
BoT94.585.958.452.3--78.869.7
PCB93.881.6----81.771.2
VGTri--81.071.085.793.73--
PVPM--66.859.578.3---
HOReID94.284.980.370.285.391.0--
ISP95.388.6------
PAT95.488.081.672.188.0392.3--
TRANS95.288.9----83.573.4
SOLIDER96.9193.91----84.476.13
SSGR96.189.378.572.9----
FED95.086.386.3179.3384.6---
BPBreed95.789.482.975.2--84.975.5
PFD95.589.783.081.52----
KPRIN w/o prompt95.688.783.378.281.786.085.375.4
KPRIN95.989.685.4279.186.090.092.3182.31
KPRSOL w/o prompt96.6393.0380.078.590.3293.7286.1375.8
KPRSOL96.6293.2284.8382.6190.7194.0190.6281.22
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.788, + 0.716 + ], + "angle": 0, + "content": "from sports activities, as illustrated in Fig. 4b. Occ-PTrack is divided into a train/test that includes 1000/1411 identities with 17.898/13.412 images from 474/170 videos, which is roughly equivalent in terms of scale to other popular ReID datasets [29,53] (Tab. 5 of the supp. materials). To assess the ReID model's performance in multi-person occlusion scenarios, we select the most cluttered images of each identity in the test set as query samples, and the remaining test images as gallery samples. Cluttered images correspond to multi-person occlusions scenarios where either the front (occluding) or back (occluded) person is the ReID target, to evaluate the model ability to re-identify individuals in both scenarios. We provide further details about our proposed dataset in the supp. materials. Occ-PTrack is challenging as persons within the same video exhibit a high degree of visual resemblance since they often wear similar sports kits." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Keypoint Annotations for Standard ReID Datasets: Since conventional ReID datasets [29,53,58] do not contain keypoint annotations, we generate pseudo-labels for them with the PifPaf [19] pose estimator. These pseudo-labels include keypoint annotations in COCO [22] format for each image, as illustrated in Fig. 4a. When multiple skeletons are detected in an image, we make the assumption the one with its head closer to the top center part of the image is the intended target, and mark it with a \"is_target\" attribute. Other skeletons are used as negatives. More details are provided in the supplementary materials." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.465, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Keypoint Promptable Re-Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.457, + 0.162 + ], + "angle": 0, + "content": "4.3 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.309 + ], + "angle": 0, + "content": "We evaluate our model with both the ImageNet [31] (KPRIN) and the SOLIDER [5] (KPRSOL) pre-trained Swin backbones. SOLIDER is a human-centric foundation model trained on the large-scale LUPerson [7] dataset. Following [34], the number of body parts \\( K \\) is set to 8 for Occ-PTrack, and 5 otherwise. The training procedure is mainly adopted from TransReID [12]. We apply our BIPO data augmentation (Sec. 3.5) with a 0.3 probability. The model is trained for 120 epochs with a batch size of 64 and a cosine annealing lr scheduler. For KPRIN/KPRSOL, images are resized to a width of 128 and a height of 256/384. The Gaussian kernel standard deviation \\( \\alpha_{G} \\) (Sec. 3.1) is set to 0.1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.327, + 0.621, + 0.342 + ], + "angle": 0, + "content": "4.4 Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.785, + 0.396 + ], + "angle": 0, + "content": "In Table \\(\\boxed{1}\\), we present a comparative analysis of KPR, when using prompts (KPR) and not using prompts\\(^{8}\\) (KPR w/o prompt), against all leading occluded ReID methods. Our method emerges as the top performer overall." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.789, + 0.653 + ], + "angle": 0, + "content": "Occluded PoseTrack-ReID: Unlike the other three datasets, Occ-PTrack contain a substantial number of images with multi-person occlusions, making them suitable environments to assess the effectiveness of our promptable method. On this challenging occluded dataset, our proposed model KPR outperforms all previous methods. Our keypoint promptable approach offers several key advantages over the previous pose-guided methods, including: 1) the ability to handle multi-person occlusions; 2) the capability to process both positive and negative keypoints; 3) a prompt-aware appearance encoding, leading to better feature disentanglement between multiple persons; and 4) the flexibility of prompts being optional. We refer readers to Sec. 2 for a detailed discussion about these key differences and a comparison to previous methods. On our proposed Occ-PTrack, we can see that the best-performing method is the foundation model SOLIDER, but it still lies about \\(6\\%\\) behind our KPR solution. Even BPBreID [34], which is specialized for occluded scenarios, demonstrates far lower performance, since it is not designed to address the Multi-Person Ambiguity. Our experiments show that integrating keypoint prompts boosts performance by at least \\(7.0\\%\\) Rank-1 and \\(6.2\\%\\) mAP, since it helps in mitigating the negative impact of MPA." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.653, + 0.789, + 0.804 + ], + "angle": 0, + "content": "Market-1501, Occluded-ReID and Partial-reID: These three datasets primarily feature single individuals in their images, which means multi-person ambiguities are rare, and therefore prompts are not exploited to their full potential. On Market-1501, our method performs on par with the state-of-the-art. However, performance levels have plateaued in recent years on this dataset. In contrast, on Occluded-ReID and Partial-reID, using a prompt noticeably improves performance by helping distinguish the individual features from occluding objects. Additionally, using prompt is beneficial when facing this cross-domain setting. These findings are consistent with earlier studies [8, 39-41] that also leveraged pose estimation to attain strong performance on this dataset." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.736, + 0.825 + ], + "angle": 0, + "content": "8 The prompt tokenizer is completely removed at both training and inference." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.825, + 0.78, + 0.84 + ], + "angle": 0, + "content": "9 Occ-ReID and Partial-ReID have no train set, so Market-1501 is used for training." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.811, + 0.78, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "V. Somers et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.162, + 0.504, + 0.191 + ], + "angle": 0, + "content": "Table 2: Multi-person pose tracking performance in videos on PoseTrack21 [6]." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.191, + 0.504, + 0.283 + ], + "angle": 0, + "content": "
MethodHOTA↑DetA↑AssA↑IDs↓
TRMOT 4246.840.955.0-
FairMOT 5153.547.461.4-
Tracktor++ 358.352v.765.4-
CorrTrack + ReID 656.951.364.2-
KPRTrack w/o prompt61.558.465.83830
KPRTrack63.0 (+1.5)59.0 (+0.6)68.3 (+2.5)3344 (-486)
" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.145, + 0.787, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.272, + 0.771, + 0.287 + ], + "angle": 0, + "content": "Fig. 5: Prompt optionality study" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.353, + 0.577, + 0.369 + ], + "angle": 0, + "content": "4.5 Multi-Person Pose Tracking in Videos" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.789, + 0.665 + ], + "angle": 0, + "content": "We introduce a simple yet robust pose tracker, referred to as KPRTrack, built upon our promptable ReID method. KPRTrack combines the YOLOX object detector with the HRNet pose estimator. For each detected person, it extracts part-based ReID features utilizing their keypoints as positive prompts and keypoints from other individuals within the detected person bbox as negative prompts. Detections are matched in an online fashion with the Hungarian algorithm based on the distance function introduced in Eq. (2). Only tracklet-detection pairs with a distance below a threshold of 0.2 are considered for matching. Therefore, our proposed tracker relies solely on appearance information, meaning spatio-temporal cues are omitted. Pose tracking results are reported in Tab. Despite relying on ReID only, KPRTrack achieves state-of-the-art performance, demonstrating that an occlusion-robust ReID model has strong tracking capabilities. Finally, our two last experiments in Tab. reveal that employing keypoint prompts significantly enhances tracking performance. This improvement can be attributed to the enhanced communication between the detector and the ReID via the keypoint prompts. This leads to improved association accuracy and reduced identity switches, especially when multiple targets intersect paths (i.e., multi-person occlusion scenarios). Further information regarding our baseline tracker is provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.687, + 0.403, + 0.701 + ], + "angle": 0, + "content": "4.6 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.712, + 0.787, + 0.728 + ], + "angle": 0, + "content": "In this section, we analyze the performance impact of our architectural choices." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Components of KPR As demonstrated in Table 3, all our proposed modules introduced in Sec. 3 play a crucial role in improving the overall ReID performance. Exp. 2 and Exp. 8 demonstrate the importance of the Part-based Head to compare only mutually visible body parts when facing occluded persons, compared to using a single global feature (as in Exp. 1 and 7). PBH degrades performance when applied alone (Exp. 2 vs Exp. 1), but is beneficial when jointly" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.465, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Keypoint Promptable Re-Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.231 + ], + "angle": 0, + "content": "Table 3: Ablation study of our proposed KPR architecture. PBH stands for \"Part-based Head\", meaning we use multiple part-based embeddings with visibility scores instead of a single global embedding (Sec. 3.3). MSF stands for the \"Multi-Scale Features\" module added to Swin (Sec. 3.2). Prompt \\(\\oplus/\\ominus\\) refers to positive/negative prompts (Sec. 3.1). BIPO stands for the \"Batch-wise Inter-Person Occlusion\" data augmentation (Sec. 3.5). SOL stands for the pre-trained weights from SOLIDER [5] (Sec. 4.3)." + }, + { + "type": "table", + "bbox": [ + 0.321, + 0.241, + 0.684, + 0.396 + ], + "angle": 0, + "content": "
IdMain components of KPRDataset
PBHMSFPromptBIPOSOLOccluded-PoseTrack
⊕∅R-1mAP
184.874.7
2✓83.3 (-1.5)72.2 (-2.5)
3✓✓84.8 (+0.0)76.1 (+1.4)
4✓✓✓88.8 (+4.0)80.4 (+5.7)
5✓✓✓✓90.0 (+5.2)80.7 (+6.0)
6✓✓✓86.1 (+1.3)76.7 (+2.0)
7✓✓✓✓80.4 (+5.1)89.9 (+5.7)
8✓✓✓✓✓92.3 (+6.5)82.3 (+7.6)
9✓✓✓✓✓✓90.6 (+5.8)81.2 (+6.5)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.789, + 0.668 + ], + "angle": 0, + "content": "enabled with other modules (Exp. 8 vs Exp 7). Exp. 3 is consistent with claims from previous works [26,57] that increasing the feature map resolution is crucial to achieve fine-grained re-identification and better overall performance. Exp. 4 and 5 illustrate the benefits of positive and negative prompts to disambiguate the intended ReID target from other persons. Exp. 6 demonstrates the importance of our BIPO augmentation to generate artificial occlusions at training, since most occluded ReID benchmarks feature mainly occluded samples in their test set and not in their training set. Exp. 8, in comparison with Exp. 5 and 6, demonstrates how generating these artificial multi-person occlusions at training is crucial to teach the model to rely on the input prompt to identify the intended ReID target among all candidates. The usage of the SOLIDER foundation model as pre-trained weights is illustrated in Exp. 9. Unfortunately, SOLIDER does not enhance performance on the Occ-PTrack dataset, likely because of a domain gap: SOLIDER focuses on street surveillance, while Occ-PTrack predominantly features sports images from handheld cameras. However, as demonstrated in Tab. [1], SOLIDER does improve performance for all other datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "A Prompt-Optional Method In Figure 5 we analyze the test-time robustness of KPR against noisy or missing prompts on Occ-PTrack, by increasingly removing a higher percentage of keypoints from the input prompt. The red curve represents the performance of our proposed promptable KPR model, while the blue constant line illustrates the performance of the non-promptable version, which is unaffected by the proportion of keypoints removed from the input prompt. As illustrated, KPR significantly outperforms the non-promptable baseline with a full prompt and remains competitive with few or no keypoints, highlighting its prompt optional capability. We provide an additional study of the performance w.r.t. the amount of multi-person occlusion in the supplementary materials." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "V. Somers et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.143, + 0.846, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.369, + 0.786, + 0.388 + ], + "angle": 0, + "content": "(b) Part-attention maps (c) Pose tracking results: all track IDs, depicted by unique skeleton for various input prompts. colors, are maintained after a long occlusion by the front cyclist." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.4, + 0.788, + 0.445 + ], + "angle": 0, + "content": "Fig.6: Qualitative results on multi-person pose tracking and person retrieval. Red/pastel dots indicate negative/positive prompts. Colored heatmaps illustrates the part-attention maps of KPR described in Sec. 3.3 with one color per body part." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.793, + 0.531 + ], + "angle": 0, + "content": "Qualitative Assessment Figure demonstrates the impact of keypoint prompting on the part-attention maps, on person retrieval, and on pose tracking. As illustrated, KPR effectively leverages the input prompts to localize the intended target, and is therefore robust to occlusions and multi-person ambiguities." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.552, + 0.36, + 0.568 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.788, + 0.717 + ], + "angle": 0, + "content": "In this work we propose KPR, the first promptable ReID model, designed to address occluded scenarios and multi-person ambiguities arising from bounding-box-based inputs. Our model offers practical flexibility by being prompt-optional, achieving state-of-the-art performance without prompts, and outperforming this baseline with prompts. Furthermore, our model outperforms the state-of-the-art on three popular ReID benchmarks and on our novel Occ-PTrack dataset. We finally demonstrate KPR's potential for pose tracking in videos. Our codebase, keypoint labels and proposed dataset will be released to encourage further research on this new promptable ReID paradigm." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.787, + 0.749 + ], + "angle": 0, + "content": "Acknowledgment: This work has been funded by Sportradar, by the Walloon region project ReconNAssance, and by the FNRS." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.769, + 0.325, + 0.785 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.798, + 0.787, + 0.843 + ], + "angle": 0, + "content": "1. Adaimi, G., Kreiss, S., Alahi, A.: Deep visual re-identification with confidence. Transportation Research Part C: Emerging Technologies 126, 103067 (May 2021). https://doi.org/10.1016/j.trc.2021.103067" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.465, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Keypoint Promptable Re-Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "2. Alahi, A., Ramanathan, V., Fei-Fei, L.: Tracking millions of humans in crowded spaces. Group and Crowd Behavior for Computer Vision pp. 115-135 (2017). https://doi.org/10.1016/b978-0-12-809276-7.00007-2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "3. Bergmann, P., Meinhardt, T., Leal-Taixe, L.: Tracking without bells and whistles. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 941-951. Inst. Electr. Electron. Eng. (IEEE) (Oct 2019). https://doi.org/10.1109/iccv.2019.00103" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.232, + 0.785, + 0.299 + ], + "angle": 0, + "content": "4. Chen, P., Liu, W., Dai, P., Liu, J., Ye, Q., Xu, M., Chen, Q., Ji, R.: Occlude them all: Occlusion-aware attention network for occluded person re-ID. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 11813-11822. Inst. Electr. Electron. Eng. (IEEE), Montreal, QC, Canada (Oct 2021). https://doi.org/10.1109/icc48922.2021.01162" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.3, + 0.785, + 0.37 + ], + "angle": 0, + "content": "5. Chen, W., Xu, X., Jia, J., Luo, H., Wang, Y., Wang, F., Jin, R., Sun, X.: Beyond appearance: A semantic controllable self-supervised learning framework for human-centric visual tasks. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 15050-15061. Inst. Electr. Electron. Eng. (IEEE), Vancouver, Can. (Jun 2023). https://doi.org/10.1109/cvpr52729.2023.01445" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.438 + ], + "angle": 0, + "content": "6. Doering, A., Chen, D., Zhang, S., Schiele, B., Gall, J.: PoseTrack21: A dataset for person search, multi-object tracking and multi-person pose tracking. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 20931-20940. Inst. Electr. Electron. Eng. (IEEE), New Orleans, LA, USA (Jun 2022). https://doi.org/10.1109/cvpr52688.2022.02029" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.439, + 0.785, + 0.507 + ], + "angle": 0, + "content": "7. Fu, D., Chen, D., Bao, J., Yang, H., Yuan, L., Zhang, L., Li, H., Chen, D.: Unsupervised pre-training for person re-identification. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 14745-14754. Inst. Electr. Electron. Eng. (IEEE), Nashville, TN, USA (Jun 2021). https://doi.org/10.1109/cvpr46437.2021.01451" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.508, + 0.785, + 0.563 + ], + "angle": 0, + "content": "8. Gao, S., Wang, J., Lu, H., Liu, Z.: Pose-guided visible part matching for occluded person ReID. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 11741-11749. Inst. Electr. Electron. Eng. (IEEE), Seattle, WA, USA (Jun 2020). https://doi.org/10.1109/cvpr42600.2020.01176" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.564, + 0.785, + 0.592 + ], + "angle": 0, + "content": "9. Ge, Z., Liu, S., Wang, F., Li, Z., Sun, J.: YOLOX: Exceeding YOLO series in 2021. arXiv abs/2107.08430 (2021). https://doi.org/10.48550/arXiv.2107.08430" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.592, + 0.785, + 0.66 + ], + "angle": 0, + "content": "10. Ghiasi, G., Cui, Y., Srinivas, A., Qian, R., Lin, T.Y., Cubuk, E.D., Le, Q.V., Zoph, B.: Simple copy-paste is a strong data augmentation method for instance segmentation. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 2917-2927. Inst. Electr. Electron. Eng. (IEEE), Nashville, TN, USA (Jun 2021). https://doi.org/10.1109/cvpr46437.2021.00294" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.66, + 0.785, + 0.84 + ], + "angle": 0, + "content": "11. Giancola, S., Cioppa, A., Deliège, A., Magera, F., Somers, V., Kang, L., Zhou, X., Barnich, O., De Vleeschouwer, C., Alahi, A., Ghanem, B., Van Droogenbroeck, M., Darwish, A., Maglo, A., Clapés, A., Luyts, A., Boiarov, A., Xarles, A., Orcesi, A., Shah, A., Fan, B., Comandur, B., Chen, C., Zhang, C., Zhao, C., Lin, C., Chan, C.Y., Hui, C.C., Li, D., Yang, F., Liang, F., Da, F., Yan, F., Yu, F., Wang, G., Chan, H.A., Zhu, H., Kan, H., Chu, J., Hu, J., Gu, J., Chen, J., Soares, J.V.B., Theiner, J., De Corte, J., Brito, J.H., Zhang, J., Li, J., Liang, J., Shen, L., Ma, L., Chen, L., Santos Marques, M., Azatov, M., Kasatkin, N., Wang, N., Jia, Q., Pham, Q.C., Ewerth, R., Song, R., Li, R., Gade, R., Debien, R., Zhang, R., Lee, S., Escalera, S., Jiang, S., Odashima, S., Chen, S., Masui, S., Ding, S., Chan, S.w., Chen, S., El-Shabrawy, T., He, T., Moeslund, T.B., Siu, W.C., Zhang, W., Li, W., Wang, X., Tan, X., Li, X., Wei, X., Ye, X., Liu, X., Wang, X., Guo, Y., Zhao, Y., Yu, Y., Li, Y., He, Y., Zhong, Y., Guo, Z., Li, Z.: SoccerNet 2022" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "V. Somers et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.787, + 0.188 + ], + "angle": 0, + "content": "challenges results. In: Proceedings of the 5th International ACM Workshop on Multimedia Content Analysis in Sports. ACM (Oct 2022). https://doi.org/10.1145/3552437.3558545" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.243 + ], + "angle": 0, + "content": "12. He, S., Luo, H., Wang, P., Wang, F., Li, H., Jiang, W.: TransReID: Transformer-based object re-identification. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 14993–15002. Inst. Electr. Electron. Eng. (IEEE), Montreal, QC, Canada (Oct 2021). https://doi.org/10.1109/iccv48922.2021.01474" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.244, + 0.787, + 0.283 + ], + "angle": 0, + "content": "13. Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re-identification. arXiv abs/1703.07737 (2017). https://doi.org/10.48550/arXiv.1703.07737" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.284, + 0.787, + 0.338 + ], + "angle": 0, + "content": "14. Istasse, M., Somers, V., Elancheliyan, P., De, J., Zambrano, D.: DeepSportradar-v2: A multi-sport computer vision dataset for sport understandings. In: Proceedings of the 6th International Workshop on Multimedia Content Analysis in Sports. ACM (Oct 2023). https://doi.org/10.1145/3606038.3616160" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.338, + 0.787, + 0.393 + ], + "angle": 0, + "content": "15. Jia, M., Cheng, X., Zhai, Y., Lu, S., Ma, S., Tian, Y., Zhang, J.: Matching on sets: Conquer occluded person re-identification without alignment. In: AAAI Conf. Artif. Intell. vol. 35, pp. 1673-1681. Association for the Advancement of Artificial Intelligence (AAAI) (May 2021). https://doi.org/10.1609/aaai.v35i2.16260" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.393, + 0.787, + 0.419 + ], + "angle": 0, + "content": "16. Joos, V., Somers, V., Standaert, B.: TrackLab. https://github.com/TrackingLaboratory/tracklab (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.42, + 0.787, + 0.46 + ], + "angle": 0, + "content": "17. Kiran, M., Praveen, R.G., Nguyen-Meidine, L.T., Belharbi, S., Blais-Morin, L.A., Granger, E.: Holistic guidance for occluded person re-identification. arXiv abs/2104.06524 (2021). https://doi.org/10.48550/arXiv.2104.06524" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.46, + 0.787, + 0.5 + ], + "angle": 0, + "content": "18. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv abs/2304.02643 (2023). https://doi.org/10.48550/arXiv.2304.02643" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.501, + 0.787, + 0.554 + ], + "angle": 0, + "content": "19. Kreiss, S., Bertoni, L., Alahi, A.: PifPaf: Composite fields for human pose estimation. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 11969-11978. Inst. Electr. Electron. Eng. (IEEE), Long Beach, CA, USA (Jun 2019). https://doi.org/10.1109/cvpr.2019.01225" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.555, + 0.787, + 0.595 + ], + "angle": 0, + "content": "20. Leal-Taixe, L., Milan, A., Reid, I., Roth, S., Schindler, K.: MOTChallenge 2015: Towards a benchmark for multi-target tracking. arXiv abs/1504.01942 (2015). https://doi.org/10.48550/arXiv.1504.01942" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.595, + 0.787, + 0.663 + ], + "angle": 0, + "content": "21. Li, Y., He, J., Zhang, T., Liu, X., Zhang, Y., Wu, F.: Diverse part discovery: Occluded person re-identification with part-aware transformer. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 2897-2906. Inst. Electr. Electron. Eng. (IEEE), Nashville, TN, USA (Jun 2021). https://doi.org/10.1109/cvpr46437.2021.00292" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.663, + 0.787, + 0.718 + ], + "angle": 0, + "content": "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: Eur. Conf. Comput. Vis. (ECCV). Lect. Notes Comput. Sci., vol. 8693, pp. 740-755. Springer Int. Publ. (2014). https://doi.org/10.1007/978-3-319-10602-1_48" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.718, + 0.787, + 0.771 + ], + "angle": 0, + "content": "23. Liu, Q., Xu, Z., Bertasius, G., Niethammer, M.: SimpleClick: Interactive image segmentation with simple vision transformers. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 22233-22243. Inst. Electr. Electron. Eng. (IEEE), Paris, Fr. (Oct 2023). https://doi.org/10.1109/iccv51070.2023.02037" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.787, + 0.839 + ], + "angle": 0, + "content": "24. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 9992-10002. Inst. Electr. Electron. Eng. (IEEE), Montreal, QC, Canada (Oct 2021). https://doi.org/10.1109/iccv48922.2021.00986" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.465, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Keypoint Promptable Re-Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.202 + ], + "angle": 0, + "content": "25. Luiten, J., Osuep, A., Dendorfer, P., Torr, P., Geiger, A., Leal-Taixe, L., Leibe, B.: HOTA: A higher order metric for evaluating multi-object tracking. Int. J. Comput. Vis. 129(2), 548-578 (Oct 2020). https://doi.org/10.1007/s11263-020-01375-2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.787, + 0.259 + ], + "angle": 0, + "content": "26. Luo, H., Gu, Y., Liao, X., Lai, S., Jiang, W.: Bag of tricks and a strong baseline for deep person re-identification. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Work. (CVPRW). pp. 1487-1495. Inst. Electr. Electron. Eng. (IEEE), Long Beach, CA, USA (Jun 2019). https://doi.org/10.1109/cvprw.2019.00190" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.313 + ], + "angle": 0, + "content": "27. Ma, Z., Zhao, Y., Li, J.: Pose-guided inter- and intra-part relational transformer for occluded person re-identification. In: Proceedings of the 29th ACM International Conference on Multimedia. ACM (Oct 2021). https://doi.org/10.1145/3474085.3475283" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.383 + ], + "angle": 0, + "content": "28. Mansourian, A.M., Somers, V., De Vleeschouwer, C., Kasaei, S.: Multi-task learning for joint re-identification, team affiliation, and role classification for sports visual tracking. In: Proceedings of the 6th International Workshop on Multimedia Content Analysis in Sports. ACM (Oct 2023). https://doi.org/10.1145/3606038.3616172" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.438 + ], + "angle": 0, + "content": "29. Miao, J., Wu, Y., Liu, P., Ding, Y., Yang, Y.: Pose-guided feature alignment for occluded person re-identification. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 542-551. Inst. Electr. Electron. Eng. (IEEE) (Oct 2019). https://doi.org/10.1109/iccv.2019.00063" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "30. Peng, Y., Hou, S., Cao, C., Liu, X., Huang, Y., He, Z.: Deep learning-based occluded person re-identification: A survey. arXiv abs/2207.14452 (2022). https://doi.org/10.48550/arXiv.2207.14452" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.536 + ], + "angle": 0, + "content": "31. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet large scale visual recognition challenge. Int. J. Comput. Vis. 115(3), 211-252 (Apr 2015). https://doi.org/10.1007/s11263-015-0816-y" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.592 + ], + "angle": 0, + "content": "32. Seidenschwarz, J., Brasó, G., Serrano, V.C., Elezi, I., Leal-Taixe, L.: Simple cues lead to a strong multi-object tracker. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 13813-13823. Inst. Electr. Electron. Eng. (IEEE), Vancouver, Can. (Jun 2023). https://doi.org/10.1109/cvpr52729.2023.01327" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.675 + ], + "angle": 0, + "content": "33. Somers, V., Joos, V., Giancola, S., Cioppa, A., Ghasemzadeh, S.A., Magera, F., Standaert, B., Mansourian, A.M., Zhou, X., Kasaei, S., Ghanem, B., Alahi, A., Van Droogenbroeck, M., De Vleeschouwer, C.: SoccerNet game state reconstruction: End-to-end athlete tracking and identification on a minimap. In: IEEE Int. Conf. Comput. Vis. Pattern Recognit. Work. (CVPRW), CVsports. Seattle, WA, USA (Jun 2024). https://doi.org/10.48550/ARXIV.2404.11335" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.731 + ], + "angle": 0, + "content": "34. Somers, V., Vleeschouwer, C.D., Alahi, A.: Body part-based representation learning for occluded person re-identification. In: IEEE/CVF Winter Conf. Appl. Comput. Vis. (WACV). pp. 1613–1623. Inst. Electr. Electron. Eng. (IEEE), Waikoloa, HI, USA (Jan 2023). https://doi.org/10.1109/wacv56688.2023.00166" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.787, + 0.786 + ], + "angle": 0, + "content": "35. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 5686-5696. Inst. Electr. Electron. Eng. (IEEE), Long Beach, CA, USA (Jun 2019). https://doi.org/10.1109/cvpr.2019.00584" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.843 + ], + "angle": 0, + "content": "36. Sun, Y., Zheng, L., Yang, Y., Tian, Q., Wang, S.: Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline). In: Eur. Conf. Comput. Vis. (ECCV). Lect. Notes Comput. Sci., vol. 11208, pp. 501-518. Springer Int. Publ. (2018). https://doi.org/10.1007/978-3-030-01225-0_30" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.843 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "V. Somers et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "37. Tan, H., Liu, X., Yin, B., Li, X.: MHSA-net: Multihead self-attention network for occluded person re-identification. IEEE Trans. Neural Networks Learn. Syst. 34(11), 8210-8224 (Nov 2023). https://doi.org/10.1109/tnnls.2022.3144163" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.258 + ], + "angle": 0, + "content": "38. Van Zandycke, G., Somers, V., Istasse, M., Don, C.D., Zambrano, D.: DeepSportradar-v1: Computer vision dataset for sports understanding with high quality annotations. In: Proceedings of the 5th International ACM Workshop on Multimedia Content Analysis in Sports. ACM (Oct 2022). https://doi.org/10.1145/3552437.3555699" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.786, + 0.328 + ], + "angle": 0, + "content": "39. Wang, G., Yang, S., Liu, H., Wang, Z., Yang, Y., Wang, S., Yu, G., Zhou, E., Sun, J.: High-order information matters: Learning relation and topology for occluded person re-identification. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 6448-6457. Inst. Electr. Electron. Eng. (IEEE), Seattle, WA, USA (Jun 2020). https://doi.org/10.1109/cvpr42600.2020.00648" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.384 + ], + "angle": 0, + "content": "40. Wang, T., Liu, H., Song, P., Guo, T., Shi, W.: Pose-guided feature disentangling for occluded person re-identification based on transformer. In: AAAI Conf. Artif. Intell. vol. 36, pp. 2540-2549. Association for the Advancement of Artificial Intelligence (AAAI) (Jun 2022). https://doi.org/10.1609/aaai.v36i3.20155" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.786, + 0.452 + ], + "angle": 0, + "content": "41. Wang, Z., Zhu, F., Tang, S., Zhao, R., He, L., Song, J.: Feature erasing and diffusion network for occluded person re-identification. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 4744-4753. Inst. Electr. Electron. Eng. (IEEE), New Orleans, LA, USA (Jun 2022). https://doi.org/10.1109/cvpr52688.2022.00471" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.507 + ], + "angle": 0, + "content": "42. Wang, Z., Zheng, L., Liu, Y., Li, Y., Wang, S.: Towards real-time multi-object tracking. In: Eur. Conf. Comput. Vis. (ECCV). Lect. Notes Comput. Sci., vol. 12356, pp. 107-122. Springer Int. Publ. (2020). https://doi.org/10.1007/978-3-030-58621-8_7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.786, + 0.564 + ], + "angle": 0, + "content": "43. Wei, L., Zhang, S., Gao, W., Tian, Q.: Person transfer GAN to bridge domain gap for person re-identification. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 79–88. Inst. Electr. Electron. Eng. (IEEE), Salt Lake City, UT, USA (Jun 2018). https://doi.org/10.1109/cvpr.2018.00016" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.786, + 0.619 + ], + "angle": 0, + "content": "44. Yan, C., Pang, G., Jiao, J., Bai, X., Feng, X., Shen, C.: Occluded person re-identification with single-scale global representations. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 11855-11864. Inst. Electr. Electron. Eng. (IEEE), Montreal, QC, Canada (Oct 2021). https://doi.org/10.1109/iccv48922.2021.01166" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.786, + 0.687 + ], + "angle": 0, + "content": "45. Yang, J., Zhang, J., Yu, F., Jiang, X., Zhang, M., Sun, X., Chen, Y., Zheng, W.S.: Learning to know where to see: A visibility-aware approach for occluded person re-identification. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 11865-11874. Inst. Electr. Electron. Eng. (IEEE), Montreal, QC, Canada (Oct 2021). https://doi.org/10.1109/iccv48922.2021.01167" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.786, + 0.73 + ], + "angle": 0, + "content": "46. Yang, Q., Wang, P., Fang, Z., Lu, Q.: Focus on the visible regions: Semantic-guided alignment model for occluded person re-identification. Sensors 20(16), 1-15 (Aug 2020). https://doi.org/10.3390/s20164431" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.786, + 0.772 + ], + "angle": 0, + "content": "47. Ye, M., Shen, J., Lin, G., Xiang, T., Shao, L., Hoi, S.C.H.: Deep learning for person re-identification: A survey and outlook. IEEE Trans. Pattern Anal. Mach. Intell. 44(6), 2872–2893 (Jun 2022). https://doi.org/10.1109/tpami.2021.3054775" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.786, + 0.813 + ], + "angle": 0, + "content": "48. Zang, X., Li, G., Gao, W., Shu, X.: Learning to disentangle scenes for person re-identification. Image Vis. Comput. 116, 104330 (Dec 2021). https://doi.org/10.1016/j.imavis.2021.104330" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "49. Zhang, G., Zhang, Y., Zhang, T., Li, B., Pu, S.: PHA: Patch-wise high-frequency augmentation for transformer-based person re-identification. In: IEEE/CVF Conf." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.466, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Keypoint Promptable Re-Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Comput. Vis. Pattern Recognit. (CVPR). pp. 14133-14142. Inst. Electr. Electron. Eng. (IEEE), Vancouver, Can. (Jun 2023). https://doi.org/10.1109/cvpr52729.2023.01358" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.259 + ], + "angle": 0, + "content": "50. Zhang, Y., Sun, P., Jiang, Y., Yu, D., Weng, F., Yuan, Z., Luo, P., Liu, W., Wang, X.: ByteTrack: Multi-object tracking by associating every detection box. In: Eur. Conf. Comput. Vis. (ECCV). Lect. Notes Comput. Sci., vol. 13682, pp. 1-21. Springer Nat. Switz. (2022). https://doi.org/10.1007/978-3-031-20047-2_1, https://doi.org/10.1007/978-3-031-20047-2_1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.313 + ], + "angle": 0, + "content": "51. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: FairMOT: On the fairness of detection and re-identification in multiple object tracking. Int. J. Comput. Vis. 129(11), 3069–3087 (Sept 2021). https://doi.org/10.1007/s11263-021-01513-4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.357 + ], + "angle": 0, + "content": "52. Zheng, K., Lan, C., Zeng, W., Liu, J., Zhang, Z., Zha, Z.J.: Pose-guided feature learning with knowledge distillation for occluded person re-identification. arXiv abs/2108.00139 (2021). https://doi.org/10.48550/arXiv.2108.00139" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.412 + ], + "angle": 0, + "content": "53. Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., Tian, Q.: Scalable person re-identification: A benchmark. In: IEEE Int. Conf. Comput. Vis. (ICCV). pp. 1116-1124. Inst. Electr. Electron. Eng. (IEEE), Santiago, Chile (Dec 2015). https://doi.org/10.1109/iccv.2015.133" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.466 + ], + "angle": 0, + "content": "54. Zheng, W.S., Li, X., Xiang, T., Liao, S., Lai, J., Gong, S.: Partial person re-identification. In: IEEE Int. Conf. Comput. Vis. (ICCV). pp. 4678-4686. Inst. Electr. Electron. Eng. (IEEE), Santiago, Chile (Dec 2015). https://doi.org/10.1109/iccv.2015.531" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.467, + 0.787, + 0.522 + ], + "angle": 0, + "content": "55. Zhong, Z., Zheng, L., Cao, D., Li, S.: Re-ranking person re-identification with k-reciprocal encoding. In: IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 3652-3661. Inst. Electr. Electron. Eng. (IEEE), Honolulu, HI, USA (Jul 2017). https://doi.org/10.1109/cvpr.2017.389" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.522, + 0.787, + 0.577 + ], + "angle": 0, + "content": "56. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. In: AAAI Conf. Artif. Intell. vol. 34, pp. 13001-13008. Association for the Advancement of Artificial Intelligence (AAAI) (Apr 2020). https://doi.org/10.1609/aaai.v34i07.7000" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.578, + 0.787, + 0.633 + ], + "angle": 0, + "content": "57. Zhu, K., Guo, H., Liu, Z., Tang, M., Wang, J.: Identity-guided human semantic parsing for person re-identification. In: Eur. Conf. Comput. Vis. (ECCV). Lect. Notes Comput. Sci., vol. 12348, pp. 346-363. Springer Int. Publ. (2020). https://doi.org/10.1007/978-3-030-58580-8_21" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.661 + ], + "angle": 0, + "content": "58. Zhuo, J., Chen, Z., Lai, J., Wang, G.: Occluded person re-identification. arXiv abs/1804.02792 (2018). https://doi.org/10.48550/arXiv.1804.02792" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.661 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Keypoint Promptable Re-Identification/21228ec1-ee24-4ba9-acb3-211d122a99e2_origin.pdf b/2024/Keypoint Promptable Re-Identification/21228ec1-ee24-4ba9-acb3-211d122a99e2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dbea05988a5cca05354eaa763ddb417f4e4f97e7 --- /dev/null +++ b/2024/Keypoint Promptable Re-Identification/21228ec1-ee24-4ba9-acb3-211d122a99e2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c546901f331a47de76eac94da786e1a4abc724a33da37fbea2b5c750a43862de +size 2676216 diff --git a/2024/Keypoint Promptable Re-Identification/full.md b/2024/Keypoint Promptable Re-Identification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e97d9cd79900f5bbcea1c4e5f3eecb0bb3bdda5f --- /dev/null +++ b/2024/Keypoint Promptable Re-Identification/full.md @@ -0,0 +1,282 @@ +# Keypoint Promptable Re-Identification + +Vladimir Somers1,2,3, Alexandre Alahi1, and Christophe De Vleeschouwer2 + +1 Ecole Polytechnique Fédérale de Lausanne (EPFL), Switzerland + +2 Université Catholique de Louvain (UCLouvain), Belgium + +3 Sportradar, Switzerland + +Abstract. Occluded Person Re-Identification (ReID) is a metric learning task that involves matching occluded individuals based on their appearance. While many studies have tackled occlusions caused by objects, multi-person occlusions remain less explored. In this work, we identify and address a critical challenge overlooked by previous occluded ReID methods: the Multi-Person Ambiguity (MPA) arising when multiple individuals are visible in the same bounding box, making it impossible to determine the intended ReID target among the candidates. Inspired by recent work on prompting in vision, we introduce Keypoint Promptable ReID (KPR), a novel formulation of the ReID problem that explicitly complements the input bounding box with a set of semantic keypoints indicating the intended target. Since promptable re-identification is an unexplored paradigm, existing ReID datasets lack the pixel-level annotations necessary for prompting. To bridge this gap and foster further research on this topic, we introduce Occluded PoseTrack-ReID, a novel ReID dataset with keypoints labels, that features strong inter-person occlusions. Furthermore, we release custom keypoints labels for four popular ReID benchmarks. Experiments on person retrieval, but also on pose tracking, demonstrate that our method systematically surpasses previous state-of-the-art approaches on various occluded scenarios. Our code, dataset and annotations are available at https://github.com/V1Somers/Keypoint_prompts_reidentification + +Keywords: Person Re-Identification $\cdot$ Vision Prompting $\cdot$ Pose Tracking + +# 1 Introduction + +Person re-identification (ReID) 47 is a challenging retrieval task that involves matching a query image of a person of interest with other person images. ReID finds wide-ranging applications in multi-object tracking 32, pedestrian flow analysis 2, sport understanding 11,14,28,33,38 and video-surveillance 53. However, ReID is a difficult task due to various factors including inaccurate bounding boxes, luminosity changes, poor image quality, and occlusions. + +In recent years, there has been a growing interest in addressing the occluded re-identification task [30], where the ReID target might be occluded by various objects or other people. In this work, we identify and explicitly address a critical challenge that has been overlooked by previous occluded ReID methods [34, 41]: + +![](images/1ad9df224b74aacc1b686ebbf4388b878fee5e4767ba9672bc84d64a38a9575f.jpg) +(a) Keypoint Promptable ReID (KPR) + +![](images/4b0aa7420085aaac44a1af8e4a9c3f46db364f2299ebfc01a9aa11829c19271f.jpg) +Fig. 1: Overview of our proposed Keypoint Promptable ReID (KPR) method. KPR takes an image with keypoints prompts as input and produces part-based features of the prompted target. The prompt instructs the model to focus on a specific individual, i.e. the back blue jacket man (b) or the front black t-shirt man (c) in this example. Colored dots illustrate the positive keypoints prompt, with one color per body part. + +the presence of Multi-Person Ambiguity (MPA) inherent to bounding box annotations. The multi-person ambiguity arises when a bounding box image contains multiple individuals, making it challenging even for humans to accurately identify the intended ReID target among all the candidates. Figure 2a shows how MPA can negatively impact the re-identification model, causing feature mixing among different individuals or even a focus on an unintended person. + +To tackle this multi-person ambiguity, it is necessary to incorporate additional pixel-level information, such as keypoints or segmentation masks, provided by an upstream human operator or vision model. This additional data plays a vital role in assisting the downstream ReID method to distinguish the intended target from other candidates. + +Inspired by recent advances in promptable $^{4}$ methods for vision transformers (e.g., SAM [18]), we propose to explicitly use additional semantic $^{5}$ keypoints as inputs to disambiguate images with multiple individuals. Semantic keypoints are suitable for two crucial use cases: 1) manual prompting by a human operator with a few clicks on an image, and 2) automated prompting by an upstream pose estimation model. Nevertheless, our architecture could be easily extended to support segmentation masks as prompts. + +![](images/f6b8068b9db777551cc624f85484631d293c415cc09021a88e77ae37c3cf6be2.jpg) +(a) Existing ReID models (e.g. BPBreID 34) are confused by MPA + +![](images/c40029641172a0923771296c580d38c1a0bfb111c02ae15721403ee6da5485e9.jpg) +(b) KPR explicitly addresses MPA with positive (pastel) and negative (red) keypoints prompt +Fig. 2: Person retrieval with Multi-Person Ambiguity (MPA). Green/red borders are correct/incorrect matches. Red/pastel dots indicate negative/positive prompts. + +Prior research [8,39,40] on pose-guided methods has explored the use of additional information from external pose estimation models. However, none of these works explicitly tackled the multi-pedestrian ambiguity issue. + +Following above insights, we propose KPR, a novel Keypoint-Promptable transformer for part-based person Re-identification, illustrated in Fig. 1a. KPR is a transformer-based model that takes an image along with semantic keypoints as input. It then produces part-based features, each representing a distinct body part of the ReID target, along with their respective visibility scores. Part-based methods have demonstrated superior performance [34,57] in re-identifying occluded individuals because they utilize only the visible parts for comparison. Our method can process both positive and negative keypoints, which respectively represent the target and non-target pedestrians. As illustrated in Fig. 1b, prompts can be used to re-identify the desired (front or back) person in case of occlusion. Furthermore, KPR is designed to be prompt optional to offer more practical flexibility. This means the same model can be used without prompt on non-ambiguous images, or with prompt when dealing with occlusions, while consistently achieving state-of-the-art performance in both cases. To demonstrate the advantages of keypoint prompts for re-identification, we evaluate KPR on two tasks: person retrieval and multi-person pose tracking. + +Moreover, since no current ReID dataset is provided with explicit target identification in case of multi-person ambiguity, we introduce a new dataset called Occluded PoseTrack-ReID, nicknamed "Occ-PTrack", and derived from the PoseTrack21 dataset [6]. This dataset offers high-quality keypoints annotations and features a substantial number of images with multi-person occlusions. + +Overall, we summarize the main contributions of our work as follows: + +1. We propose $KPR$ , a novel prompt-optional model for part-based ReID. To our knowledge, we are the first to introduce a visual prompting mechanism for person re-identification, and the first to explicitly address the ambiguity caused by multi-person occlusions. +2. We introduce Occ-PTrack: the first multi-person occluded ReID dataset with explicit target identification through manual keypoints annotations. Furthermore, we propose new keypoint annotations for four popular re-identification datasets. + +3. Our method outperforms all previous state-of-the-art methods on the occluded ReID task, demonstrating the effectiveness of the keypoint promptable approach. + +Our ambition is to advocate for a shift from the ambiguous bounding-box approach to ReID and promote a keypoint-centric paradigm. To this end, we publicly release our codebase, proposed dataset, and annotations, to provide a common setup for evaluating promptable ReID methods. + +# 2 Related Work + +Multi-Person Ambiguity: A previous study [41] addressed a related issue called "Non-Target Pedestrians (NTP)," similar to our concept of "Multi-Person Ambiguity" (MPA). However, their focus was on mitigating the negative impact of NTP during training. In contrast, we provide a method to explicitly disambiguate the intended ReID target during both training and testing phases. + +Pose-guided ReID: We detail here the fundamental difference between our novel keypoint promptable approach and the traditional pose-guided methods. Although none of the previous works explicitly address the multi-person ambiguity (MPA) issue, we identify two categories within these pose-guided approaches: 1) those inherently unable to address MPA [48,15,27,29,34,44,45], and 2) those with the potential to overcome MPA [39,40]. In the first category, methods like BPBreID [34] or Pirt [27] utilize pose labels only during training, and can not leverage additional information at test time to disambiguate multiple persons. Moreover, methods relying on horizontal stripes (VGTri [45], PGFA [29]) or the affinity fields of a pose model (PVPM [8]), are meant to tackle object-occlusions, but struggle to disambiguate multiple persons. In the second category, methods like HOReID [39] and PFD [40] directly employ keypoints from a single person, enabling focus on a specified target. However, HOReID [39] is restricted to utilizing spatial features at the exact keypoints locations, thus overlooking critical appearance cues on entire body regions. Finally, PFD [40], and all previous pose-guided methods including HOReID, incorporate pose information after appearance encoding has occurred, e.g., for local pooling of a spatial feature map, but do not leverage pose data to guide the appearance encoding process as our method do. Indeed, our approach implements a prompt-aware appearance encoding, since it conditions feature extraction on the input keypoints. This enables better feature disentanglement between the target and the occluders from the early feature extraction stages, resulting in more accurate representations. Furthermore, our method is prompt optional, setting it apart for its flexibility from these previous pose-guided methods. Finally, our method can ingest additional negative keypoints, to further disambiguate a multi-person occlusion scenario. + +Prompting in Vision: Recent works have introduced promptable transformers to tackle object segmentation. For instance, SAM [18] employs two networks: a transformer encoder to generate generic spatial features and a lightweight promptable decoder to integrate user inputs for interactive segmentation. In contrast, our approach directly prompts a feature encoder, so that feature extrac + +![](images/ccd792ff4b3f7f393d4b555253e4b3874bb29def2fee108a0847d6dce535b7a7.jpg) +Fig. 3: Architecture overview of our proposed Keypoint Promptable ReID (KPR) model. The Batch-wise Inter-Person Occlusion (BIPO) data augmentation is first applied to generate artificial inter-person occlusions (Sec. 3.5). The image and the optional positive/negative $(\oplus/\ominus)$ prompts are then tokenized and summed (Sec. 3.1). Tokens are then fed to our proposed Multi-Stage feature Fusion (MSF) Swin transformer to generate high-resolution feature maps (3.2). The feature map is then fed to a Part-based Head (PBH), which assigns a part (or the background) to each token with a part classifier, and then averages all tokens of the same part to produce the final $K$ part-based embeddings $\{f_1, \dots, f_K\}$ and their binary visibility scores $\{v_1, \dots, v_K\}$ , with $v_i \in \{0,1\}$ , and visually denoted by $\{\mathbf{x}, \check{\mathbf{v}}\}$ (Sec. 3.3). A part $i$ with no token assigned is considered invisible (i.e., $v_i = 0$ ), and ignored when computing two samples' similarity. KPR is illustrated here for $K = 8$ , with a unique color for each body part: $\{\text{head}, \text{torso}, \text{right/left arm}, \text{right/left leg}, \text{right/left feet}\}$ . Finally, the entire pipeline is trained with two losses: a Part-Prediction (PP) Loss and a ReID Loss (Sec. 3.4). + +tion is explicitly driven by the prompts. Unlike SimpleClick [23], which prompts a transformer with a single 2D point and a previous mask for iterative segmentation, our model ingests multiple keypoints concurrently, enhancing body part localization and representation by leveraging their semantic information. + +# 3 Methodology + +The overall architecture of our model KPR is illustrated in Fig. 3 and is based on a Swin transformer backbone [24]. The next subsections first describe the various components of the architecture. We then present the training process, including our proposed data augmentation. Finally, we describe the inference procedure and conclude with an intuitive discussion of our design choices. + +# 3.1 Tokenization + +The tokenization procedure coverts raw image and keypoints information into $C_i$ -dimensional tokens, to be subsequently processed by the Swin encoder. + +Image tokenization: The input image, with dimensions $H \times W$ , undergoes tokenization using the conventional Swin patch embed module. This process generates $H_{t} \times W_{t} = \frac{H}{4} \times \frac{W}{4}$ spatial tokens of dimensions $C_{i}$ by embedding each $4 \times 4$ image patch through a linear layer. + +Keypoints tokenization: As previously mentioned, our model uses two types of semantic keypoints as prompts: positive keypoints indicating the ReID + +target and negative keypoints indicating other individuals. We organize these prompt keypoints into $\mathrm{K} + 1$ groups: the first group contains all negative keypoints, and the remaining $\mathrm{K}$ groups contain subsets of the positive keypoints. Each subset of positive keypoints actually corresponds to a semantic body region. For instance, with $K = 8$ , we define groups like: $\{\text{head, torso, right/left arm, right/left leg, and right/left feet}\}$ . For each group $i \in \{1, \dots, K + 1\}$ , a unique heatmap $M_i$ of size $H \times W$ is created by drawing a Gaussian 2D kernel at each of its keypoints's center location $(x, y)$ . The kernel's standard deviation is set to a fraction $\alpha_G$ of the image width. Invisible keypoints are ignored. These $K + 1$ heatmaps are then concatenated into a global spatial tensor $M$ of size $H \times W \times (K + 1)$ . Finally, each $4 \times 4$ patch in $M$ is linearly projected to produce prompt tokens of dimension $C_i$ . Like the image tokenizer, the keypoints tokenizer outputs therefore $H_t \times W_t$ spatial tokens of dimension $C_i$ . + +Token fusion: Images and keypoints tokens are fused by summing tokens at the same spatial location, producing $H_{t} \times W_{t}$ spatial tokens with dimension $C_{i}$ . These tokens are then forwarded to the Swin-based encoder. + +Design motivations: Prompted keypoints are organized into $K + 1$ channels to preserve semantic meaning before tokenization. It allows the model to differentiate positive and negative prompts, and to distinguish prompt information of different body parts, for better part localization. Prompt-aware appearance encoding: Moreover, feeding the network with keypoint information before the encoding stage is critical in enabling the network to disentangle the ReID target features from other persons' features. Finally, this design offers the flexibility to support segmentation mask prompts by replacing $M$ with a double-channel map of positive and negative segmentation masks. Regarding token fusion, the sum operation maintains the input dimension for the Swin backbone, permitting the use of pre-trained weights from the image-only model. Prompt optionality: This design also offers the flexibility of running the model without prompts, by processing the image tokens alone, with no architectural adjustments required. + +# 3.2 MSF Swin Encoder + +A Swin-transformer [24] is employed as the backbone feature extractor for its great performance in vision tasks and its ability to capture high-frequency details, which is beneficial to ReID [49]. Moreover, we propose to enhance Swin with a Multi-Stage feature Fusion strategy, denoted as "MSF", to output feature maps with a high spatial resolution. Indeed, similar to CNN, each Swin stage reduces the number of tokens by a factor of 4 while doubling the channel dimension. To obtain a high-resolution token map, the outputs of all stages are bilinearly upsampled to the same resolution $H_{t} \times W_{t}$ as the output of the tokenization module, then concatenated into a single feature map along the channel dimension, and finally passed through a linear layer to obtain the final spatial token map of size $H_{t} \times W_{t} \times C_{o}$ . Different from the standard Swin, our MSF Swin therefore preserves the input tokens resolution. As pointed out in prior works [26,57], a high-resolution feature map enriches the granularity of the features and brings significant performance improvements (see Tab. 3, Exp. 3). + +# 3.3 Part-based Head + +The encoder yields $H_{t} \times W_{t}$ tokens, which are sent to the Part-based Head (PBH) module, as illustrated in Fig. 3. Each token is first labeled either as one of the $K$ body parts or as background, by a token-wise part classifier. The part classifier is implemented as a linear layer with $K + 1$ output logits followed by the softmax operation, and is applied individually on each token. The assignment of each token to one of the $K$ body parts (or the background) is therefore formulated as a classification task with $K + 1$ classes. We term the probability map of size $H_{t} \times W_{t} \times (K + 1)$ produced by this part classifier over all tokens as "part-attention maps". These maps are showcased as colored heatmaps in Fig. 1a and Fig. 6b with one color per channel (the background channel is ignored). Finally, tokens having the same assigned body part are averaged together, resulting in $K$ part-based embeddings $\{f_{1}, \dots, f_{K}\}$ of size $C_{o}$ . Background tokens are ignored. Parts with no tokens assigned are considered invisible and given a visibility score $v_{i} = 0$ . The corresponding invisible embedding $f_{i}$ , with $v_{i} = 0$ , is therefore ignored for downstream computation. The PBH module thus produces $K$ part-based embeddings $\{f_{1}, \dots, f_{K}\}$ and their binary visibility scores $\{v_{1}, \dots, v_{K}\}$ . + +Human parsing labels: To train the token-wise part classifier, each spatial token is assigned a ground-truth class, i.e., part label, derived from the coarse human parsing labels provided by [34]. These human parsing labels are illustrated on the right of Fig. 3 and in Fig. 4a. + +# 3.4 Training Process + +The overall objective function used to train the model is: + +$$ +L = L _ {\text {R e I D}} + \lambda_ {p p} L _ {p p}, \tag {1} +$$ + +where $L_{ReID}$ is the ReID loss supervised with identity labels, $L_{pp}$ is the token-wise part-prediction loss, and $\lambda_{pp}$ is a weight parameter empirically set to 0.3. + +Part Prediction Loss: The token-wise part classifier introduced in Sec. 3.3 is supervised with a Part Prediction (PP) Loss that is actually a cross-entropy loss applied on each spatial token as in 34. + +Re-identification loss: As a ReID objective, we employ the GiLt Loss [34], that is specifically designed to train part-based ReID models. It has the benefit of being robust to occlusions and non-discriminative body parts. GiLt combines an identity loss [26] and a batch-hard triplet loss [13]. The distance function employed for the triplet loss is the average of all local part-based distances. + +# 3.5 BIPO Data Augmentation + +Since most ReID datasets feature limited occlusion in their training set, a model often struggles to handle the occlusions it encounters during testing. To address this, we propose a novel Batch-wise Inter-Person Occlusion data augmentation (BIPO) to generate artificial inter-person occlusions on training images, + +![](images/f920778012ac8ec4370cc21f7bcba16027afeb32e5903055918ca03bf4ed620a.jpg) +Fig. 4: (a) Our proposed Batch-wise Inter-Person Occlusion (BIPO) data augmentation creates artificial person occlusions that are consistent across image, prompt, and human parsing labels. BIPO is crucial to enforce the model to rely on the input prompts. (b) Four identities with their keypoints labels from our Occ-PTrack dataset. + +![](images/a2200e0b78efa351fdf741dffc48eda9af5d1d4c8a77abb1c1ffaeeed093d1c0.jpg) +(a) BIPO + +![](images/ff517f87e4513a9e9f3ada78b005867e1cffe85282c53d2ca0a87b31fee112e9.jpg) +(b) Occluded PoseTrack-ReID + +![](images/abf05153db381133534dd36eaef671a57cf8d8e7572d428bd7274a85f83ab5b7.jpg) + +prompts, and human parsing labels. When provided with a training sample, BIPO randomly selects a different person's image (the occluder) from the same training batch, contours it with a segmentation mask derived from the human parsing labels, and overlays it on the main image. The human parsing label and keypoints prompt of the training image are then updated accordingly. Finally, all positive keypoints from the occluder are added to the training prompt to serve as additional negative points. + +Design motivations: Given that a triplet loss with batch-hard mining [13] is employed during training, using images from the same training batch to generate occluders yields hard positive and negative samples for mining triplets, thereby creating challenging scenarios for contrastive learning. BIPO is illustrated in Fig. 4a and its significance for KPR is further discussed in Sec. 3.7. + +# 3.6 Inference + +At inference, KPR processes each input image together with their respective keypoints prompts, and produces $K$ part-based representations $\{f_1,\dots,f_K\}$ of the ReID target and their binary visibility scores $\{v_{1},\ldots ,v_{K}\}$ . Similar to previous part-based works [45, 57], the global distance $dist_{\mathrm{global}}^{gg}$ between a query image $q$ and a gallery image $g$ is computed as the average cosine distance $dist_{\mathrm{cos}}$ of the part-based embeddings $f_{i}$ that are visible in both samples: + +$$ +d i s t _ {\text {g l o b a l}} ^ {q g} = \frac {\sum_ {i \in \{1 , \dots , K \}} \left(v _ {i} ^ {q} \cdot v _ {i} ^ {g} \cdot d i s t _ {\cos} \left(f _ {i} ^ {q} , f _ {i} ^ {g}\right)\right)}{\sum_ {i \in \{1 , \dots , K \}} \left(v _ {i} ^ {q} \cdot v _ {i} ^ {g}\right)}. \tag {2} +$$ + +# 3.7 KPR Pipeline Motivation and Intuition + +Our pipeline comprises two closely related architectural components. First, the prompt tokenizer, which encodes semantic keypoints information to disambiguate + +the ReID target, thereby providing prior information about body part location to drive feature extraction. Second, the Part-based Head (PBH), responsible for localizing the body parts of the target and constructing its part-based features. + +While the PBH is explicitly supervised with human parsing labels to learn human topology, there is no dedicated loss incentivizing the model to leverage the input prompt. Additionally, multi-person occlusions are rare in most ReID training sets, giving the model little incentive to rely on these prompts to localize the target person. For this reason, just adding a prompt tokenizer and training the pipeline as is produces a KPR model with weak reliance on the input prompts, and therefore low performance in multi-person occlusion scenarios. + +To address these issues, our proposed BIPO data augmentation plays a critical role. By generating artificial multi-person occlusions during training, the model is compelled to rely more on the prompts, as they represent the sole reliable source of information to distinguish the true ReID target from the occluder. + +Consequently, incorporating BIPO during training is crucial to produce a model that is proficient at leveraging prompt information and effectively handling multi-person occlusions. Finally, designing the Part-based Head without explicit reliance on the prompt offers an additional advantage: it renders our model prompt-optional, enhancing robustness against noisy or missing prompts while still enabling effective localization of the target's body parts. + +# 4 Experiments + +# 4.1 Evaluation Benchmarks and Metrics + +We assess our model's performance on three types of person retrieval setups: non-occluded (i.e., holistic) on Market-1501 [53], object occlusions on Occluded-ReID [58] and Partial-ReID [54], and finally multi-person occlusions on our below introduced Occluded PoseTrack-ReID dataset. Occluded-Duke [29] was taken down and removed from our study; check out our GitHub for more information. Two standard person retrieval metrics are reported: the CMC at Rank-1 and the $mAP$ . Performances are evaluated in a single query setting and without re-ranking [55]. Furthermore, we evaluate our model on multi-person pose tracking with the PoseTrack21 [6] dataset. We report the HOTA [25] performance metric, and relevant sub-metrics such as association accuracy ( $AssA$ ), detection accuracy ( $DetA$ ), and the number of identity switches ( $IDs$ ) [20]. + +# 4.2 Proposed Occluded PoseTrack-ReID Dataset and Annotations + +We introduce Occluded PoseTrack-ReID (or simply Occ-PTrack), a new ReID dataset we built out of the annotation available with PoseTrack21 [6], a popular video benchmark for multi-person pose tracking, that features keypoints and cross-video identity annotations. Unlike previous ReID datasets focused on street surveillance, Occ-PTrack consists of images from everyday life videos, primarily + +Table 1: Comparison of KPR with SOTA methods. Results in Italic are not provided in the original paper but reproduced by ourselves. The $1^{st}/2^{nd}/3^{rd}$ best scores are indicated with $1/2/3$ . + +
DatasetsMarket-1501Occluded-reIDPartial-reIDOccluded-PoseTrack
TypeHolisticOccluded
Object Occlusions✓✓
Person Occlusions✓
MethodsR-1mAPR-1mAPR-1R-3R-1mAP
BoT94.585.958.452.3--78.869.7
PCB93.881.6----81.771.2
VGTri--81.071.085.793.73--
PVPM--66.859.578.3---
HOReID94.284.980.370.285.391.0--
ISP95.388.6------
PAT95.488.081.672.188.0392.3--
TRANS95.288.9----83.573.4
SOLIDER96.9193.91----84.476.13
SSGR96.189.378.572.9----
FED95.086.386.3179.3384.6---
BPBreed95.789.482.975.2--84.975.5
PFD95.589.783.081.52----
KPRIN w/o prompt95.688.783.378.281.786.085.375.4
KPRIN95.989.685.4279.186.090.092.3182.31
KPRSOL w/o prompt96.6393.0380.078.590.3293.7286.1375.8
KPRSOL96.6293.2284.8382.6190.7194.0190.6281.22
+ +from sports activities, as illustrated in Fig. 4b. Occ-PTrack is divided into a train/test that includes 1000/1411 identities with 17.898/13.412 images from 474/170 videos, which is roughly equivalent in terms of scale to other popular ReID datasets [29,53] (Tab. 5 of the supp. materials). To assess the ReID model's performance in multi-person occlusion scenarios, we select the most cluttered images of each identity in the test set as query samples, and the remaining test images as gallery samples. Cluttered images correspond to multi-person occlusions scenarios where either the front (occluding) or back (occluded) person is the ReID target, to evaluate the model ability to re-identify individuals in both scenarios. We provide further details about our proposed dataset in the supp. materials. Occ-PTrack is challenging as persons within the same video exhibit a high degree of visual resemblance since they often wear similar sports kits. + +Keypoint Annotations for Standard ReID Datasets: Since conventional ReID datasets [29,53,58] do not contain keypoint annotations, we generate pseudo-labels for them with the PifPaf [19] pose estimator. These pseudo-labels include keypoint annotations in COCO [22] format for each image, as illustrated in Fig. 4a. When multiple skeletons are detected in an image, we make the assumption the one with its head closer to the top center part of the image is the intended target, and mark it with a "is_target" attribute. Other skeletons are used as negatives. More details are provided in the supplementary materials. + +# 4.3 Implementation Details + +We evaluate our model with both the ImageNet [31] (KPRIN) and the SOLIDER [5] (KPRSOL) pre-trained Swin backbones. SOLIDER is a human-centric foundation model trained on the large-scale LUPerson [7] dataset. Following [34], the number of body parts $K$ is set to 8 for Occ-PTrack, and 5 otherwise. The training procedure is mainly adopted from TransReID [12]. We apply our BIPO data augmentation (Sec. 3.5) with a 0.3 probability. The model is trained for 120 epochs with a batch size of 64 and a cosine annealing lr scheduler. For KPRIN/KPRSOL, images are resized to a width of 128 and a height of 256/384. The Gaussian kernel standard deviation $\alpha_{G}$ (Sec. 3.1) is set to 0.1. + +# 4.4 Comparison with State-of-the-Art Methods + +In Table $\boxed{1}$ , we present a comparative analysis of KPR, when using prompts (KPR) and not using prompts $^{8}$ (KPR w/o prompt), against all leading occluded ReID methods. Our method emerges as the top performer overall. + +Occluded PoseTrack-ReID: Unlike the other three datasets, Occ-PTrack contain a substantial number of images with multi-person occlusions, making them suitable environments to assess the effectiveness of our promptable method. On this challenging occluded dataset, our proposed model KPR outperforms all previous methods. Our keypoint promptable approach offers several key advantages over the previous pose-guided methods, including: 1) the ability to handle multi-person occlusions; 2) the capability to process both positive and negative keypoints; 3) a prompt-aware appearance encoding, leading to better feature disentanglement between multiple persons; and 4) the flexibility of prompts being optional. We refer readers to Sec. 2 for a detailed discussion about these key differences and a comparison to previous methods. On our proposed Occ-PTrack, we can see that the best-performing method is the foundation model SOLIDER, but it still lies about $6\%$ behind our KPR solution. Even BPBreID [34], which is specialized for occluded scenarios, demonstrates far lower performance, since it is not designed to address the Multi-Person Ambiguity. Our experiments show that integrating keypoint prompts boosts performance by at least $7.0\%$ Rank-1 and $6.2\%$ mAP, since it helps in mitigating the negative impact of MPA. + +Market-1501, Occluded-ReID and Partial-reID: These three datasets primarily feature single individuals in their images, which means multi-person ambiguities are rare, and therefore prompts are not exploited to their full potential. On Market-1501, our method performs on par with the state-of-the-art. However, performance levels have plateaued in recent years on this dataset. In contrast, on Occluded-ReID and Partial-reID, using a prompt noticeably improves performance by helping distinguish the individual features from occluding objects. Additionally, using prompt is beneficial when facing this cross-domain setting. These findings are consistent with earlier studies [8, 39-41] that also leveraged pose estimation to attain strong performance on this dataset. + +Table 2: Multi-person pose tracking performance in videos on PoseTrack21 [6]. + +
MethodHOTA↑DetA↑AssA↑IDs↓
TRMOT 4246.840.955.0-
FairMOT 5153.547.461.4-
Tracktor++ 358.352v.765.4-
CorrTrack + ReID 656.951.364.2-
KPRTrack w/o prompt61.558.465.83830
KPRTrack63.0 (+1.5)59.0 (+0.6)68.3 (+2.5)3344 (-486)
+ +![](images/4ad084a3f993e83a4eb377bd7d17afb441dacc4a0d8472fe64f3fd5bdab9578e.jpg) +Fig. 5: Prompt optionality study + +# 4.5 Multi-Person Pose Tracking in Videos + +We introduce a simple yet robust pose tracker, referred to as KPRTrack, built upon our promptable ReID method. KPRTrack combines the YOLOX object detector with the HRNet pose estimator. For each detected person, it extracts part-based ReID features utilizing their keypoints as positive prompts and keypoints from other individuals within the detected person bbox as negative prompts. Detections are matched in an online fashion with the Hungarian algorithm based on the distance function introduced in Eq. (2). Only tracklet-detection pairs with a distance below a threshold of 0.2 are considered for matching. Therefore, our proposed tracker relies solely on appearance information, meaning spatio-temporal cues are omitted. Pose tracking results are reported in Tab. Despite relying on ReID only, KPRTrack achieves state-of-the-art performance, demonstrating that an occlusion-robust ReID model has strong tracking capabilities. Finally, our two last experiments in Tab. reveal that employing keypoint prompts significantly enhances tracking performance. This improvement can be attributed to the enhanced communication between the detector and the ReID via the keypoint prompts. This leads to improved association accuracy and reduced identity switches, especially when multiple targets intersect paths (i.e., multi-person occlusion scenarios). Further information regarding our baseline tracker is provided in the supplementary materials. + +# 4.6 Ablation Studies + +In this section, we analyze the performance impact of our architectural choices. + +Components of KPR As demonstrated in Table 3, all our proposed modules introduced in Sec. 3 play a crucial role in improving the overall ReID performance. Exp. 2 and Exp. 8 demonstrate the importance of the Part-based Head to compare only mutually visible body parts when facing occluded persons, compared to using a single global feature (as in Exp. 1 and 7). PBH degrades performance when applied alone (Exp. 2 vs Exp. 1), but is beneficial when jointly + +Table 3: Ablation study of our proposed KPR architecture. PBH stands for "Part-based Head", meaning we use multiple part-based embeddings with visibility scores instead of a single global embedding (Sec. 3.3). MSF stands for the "Multi-Scale Features" module added to Swin (Sec. 3.2). Prompt $\oplus/\ominus$ refers to positive/negative prompts (Sec. 3.1). BIPO stands for the "Batch-wise Inter-Person Occlusion" data augmentation (Sec. 3.5). SOL stands for the pre-trained weights from SOLIDER [5] (Sec. 4.3). + +
IdMain components of KPRDataset
PBHMSFPromptBIPOSOLOccluded-PoseTrack
⊕∅R-1mAP
184.874.7
2✓83.3 (-1.5)72.2 (-2.5)
3✓✓84.8 (+0.0)76.1 (+1.4)
4✓✓✓88.8 (+4.0)80.4 (+5.7)
5✓✓✓✓90.0 (+5.2)80.7 (+6.0)
6✓✓✓86.1 (+1.3)76.7 (+2.0)
7✓✓✓✓80.4 (+5.1)89.9 (+5.7)
8✓✓✓✓✓92.3 (+6.5)82.3 (+7.6)
9✓✓✓✓✓✓90.6 (+5.8)81.2 (+6.5)
+ +enabled with other modules (Exp. 8 vs Exp 7). Exp. 3 is consistent with claims from previous works [26,57] that increasing the feature map resolution is crucial to achieve fine-grained re-identification and better overall performance. Exp. 4 and 5 illustrate the benefits of positive and negative prompts to disambiguate the intended ReID target from other persons. Exp. 6 demonstrates the importance of our BIPO augmentation to generate artificial occlusions at training, since most occluded ReID benchmarks feature mainly occluded samples in their test set and not in their training set. Exp. 8, in comparison with Exp. 5 and 6, demonstrates how generating these artificial multi-person occlusions at training is crucial to teach the model to rely on the input prompt to identify the intended ReID target among all candidates. The usage of the SOLIDER foundation model as pre-trained weights is illustrated in Exp. 9. Unfortunately, SOLIDER does not enhance performance on the Occ-PTrack dataset, likely because of a domain gap: SOLIDER focuses on street surveillance, while Occ-PTrack predominantly features sports images from handheld cameras. However, as demonstrated in Tab. [1], SOLIDER does improve performance for all other datasets. + +A Prompt-Optional Method In Figure 5 we analyze the test-time robustness of KPR against noisy or missing prompts on Occ-PTrack, by increasingly removing a higher percentage of keypoints from the input prompt. The red curve represents the performance of our proposed promptable KPR model, while the blue constant line illustrates the performance of the non-promptable version, which is unaffected by the proportion of keypoints removed from the input prompt. As illustrated, KPR significantly outperforms the non-promptable baseline with a full prompt and remains competitive with few or no keypoints, highlighting its prompt optional capability. We provide an additional study of the performance w.r.t. the amount of multi-person occlusion in the supplementary materials. + +![](images/df82a5220764d57d243a469ecf4ca035a7c064c43370a24b3e1e50a6af70e90f.jpg) +(b) Part-attention maps (c) Pose tracking results: all track IDs, depicted by unique skeleton for various input prompts. colors, are maintained after a long occlusion by the front cyclist. +Fig.6: Qualitative results on multi-person pose tracking and person retrieval. Red/pastel dots indicate negative/positive prompts. Colored heatmaps illustrates the part-attention maps of KPR described in Sec. 3.3 with one color per body part. + +Qualitative Assessment Figure demonstrates the impact of keypoint prompting on the part-attention maps, on person retrieval, and on pose tracking. As illustrated, KPR effectively leverages the input prompts to localize the intended target, and is therefore robust to occlusions and multi-person ambiguities. + +# 5 Conclusion + +In this work we propose KPR, the first promptable ReID model, designed to address occluded scenarios and multi-person ambiguities arising from bounding-box-based inputs. Our model offers practical flexibility by being prompt-optional, achieving state-of-the-art performance without prompts, and outperforming this baseline with prompts. Furthermore, our model outperforms the state-of-the-art on three popular ReID benchmarks and on our novel Occ-PTrack dataset. We finally demonstrate KPR's potential for pose tracking in videos. Our codebase, keypoint labels and proposed dataset will be released to encourage further research on this new promptable ReID paradigm. + +Acknowledgment: This work has been funded by Sportradar, by the Walloon region project ReconNAssance, and by the FNRS. + +# References + +1. Adaimi, G., Kreiss, S., Alahi, A.: Deep visual re-identification with confidence. Transportation Research Part C: Emerging Technologies 126, 103067 (May 2021). https://doi.org/10.1016/j.trc.2021.103067 + +2. Alahi, A., Ramanathan, V., Fei-Fei, L.: Tracking millions of humans in crowded spaces. Group and Crowd Behavior for Computer Vision pp. 115-135 (2017). https://doi.org/10.1016/b978-0-12-809276-7.00007-2 +3. Bergmann, P., Meinhardt, T., Leal-Taixe, L.: Tracking without bells and whistles. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 941-951. Inst. Electr. Electron. Eng. (IEEE) (Oct 2019). https://doi.org/10.1109/iccv.2019.00103 +4. Chen, P., Liu, W., Dai, P., Liu, J., Ye, Q., Xu, M., Chen, Q., Ji, R.: Occlude them all: Occlusion-aware attention network for occluded person re-ID. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 11813-11822. Inst. Electr. Electron. Eng. (IEEE), Montreal, QC, Canada (Oct 2021). https://doi.org/10.1109/icc48922.2021.01162 +5. Chen, W., Xu, X., Jia, J., Luo, H., Wang, Y., Wang, F., Jin, R., Sun, X.: Beyond appearance: A semantic controllable self-supervised learning framework for human-centric visual tasks. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 15050-15061. Inst. Electr. Electron. Eng. (IEEE), Vancouver, Can. (Jun 2023). https://doi.org/10.1109/cvpr52729.2023.01445 +6. Doering, A., Chen, D., Zhang, S., Schiele, B., Gall, J.: PoseTrack21: A dataset for person search, multi-object tracking and multi-person pose tracking. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 20931-20940. Inst. Electr. Electron. Eng. (IEEE), New Orleans, LA, USA (Jun 2022). https://doi.org/10.1109/cvpr52688.2022.02029 +7. Fu, D., Chen, D., Bao, J., Yang, H., Yuan, L., Zhang, L., Li, H., Chen, D.: Unsupervised pre-training for person re-identification. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 14745-14754. Inst. Electr. Electron. Eng. (IEEE), Nashville, TN, USA (Jun 2021). https://doi.org/10.1109/cvpr46437.2021.01451 +8. Gao, S., Wang, J., Lu, H., Liu, Z.: Pose-guided visible part matching for occluded person ReID. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 11741-11749. Inst. Electr. Electron. Eng. (IEEE), Seattle, WA, USA (Jun 2020). https://doi.org/10.1109/cvpr42600.2020.01176 +9. Ge, Z., Liu, S., Wang, F., Li, Z., Sun, J.: YOLOX: Exceeding YOLO series in 2021. arXiv abs/2107.08430 (2021). https://doi.org/10.48550/arXiv.2107.08430 +10. Ghiasi, G., Cui, Y., Srinivas, A., Qian, R., Lin, T.Y., Cubuk, E.D., Le, Q.V., Zoph, B.: Simple copy-paste is a strong data augmentation method for instance segmentation. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 2917-2927. Inst. Electr. Electron. Eng. (IEEE), Nashville, TN, USA (Jun 2021). https://doi.org/10.1109/cvpr46437.2021.00294 +11. Giancola, S., Cioppa, A., Deliège, A., Magera, F., Somers, V., Kang, L., Zhou, X., Barnich, O., De Vleeschouwer, C., Alahi, A., Ghanem, B., Van Droogenbroeck, M., Darwish, A., Maglo, A., Clapés, A., Luyts, A., Boiarov, A., Xarles, A., Orcesi, A., Shah, A., Fan, B., Comandur, B., Chen, C., Zhang, C., Zhao, C., Lin, C., Chan, C.Y., Hui, C.C., Li, D., Yang, F., Liang, F., Da, F., Yan, F., Yu, F., Wang, G., Chan, H.A., Zhu, H., Kan, H., Chu, J., Hu, J., Gu, J., Chen, J., Soares, J.V.B., Theiner, J., De Corte, J., Brito, J.H., Zhang, J., Li, J., Liang, J., Shen, L., Ma, L., Chen, L., Santos Marques, M., Azatov, M., Kasatkin, N., Wang, N., Jia, Q., Pham, Q.C., Ewerth, R., Song, R., Li, R., Gade, R., Debien, R., Zhang, R., Lee, S., Escalera, S., Jiang, S., Odashima, S., Chen, S., Masui, S., Ding, S., Chan, S.w., Chen, S., El-Shabrawy, T., He, T., Moeslund, T.B., Siu, W.C., Zhang, W., Li, W., Wang, X., Tan, X., Li, X., Wei, X., Ye, X., Liu, X., Wang, X., Guo, Y., Zhao, Y., Yu, Y., Li, Y., He, Y., Zhong, Y., Guo, Z., Li, Z.: SoccerNet 2022 + +challenges results. In: Proceedings of the 5th International ACM Workshop on Multimedia Content Analysis in Sports. ACM (Oct 2022). https://doi.org/10.1145/3552437.3558545 +12. He, S., Luo, H., Wang, P., Wang, F., Li, H., Jiang, W.: TransReID: Transformer-based object re-identification. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 14993–15002. Inst. Electr. Electron. Eng. (IEEE), Montreal, QC, Canada (Oct 2021). https://doi.org/10.1109/iccv48922.2021.01474 +13. Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re-identification. arXiv abs/1703.07737 (2017). https://doi.org/10.48550/arXiv.1703.07737 +14. Istasse, M., Somers, V., Elancheliyan, P., De, J., Zambrano, D.: DeepSportradar-v2: A multi-sport computer vision dataset for sport understandings. In: Proceedings of the 6th International Workshop on Multimedia Content Analysis in Sports. ACM (Oct 2023). https://doi.org/10.1145/3606038.3616160 +15. Jia, M., Cheng, X., Zhai, Y., Lu, S., Ma, S., Tian, Y., Zhang, J.: Matching on sets: Conquer occluded person re-identification without alignment. In: AAAI Conf. Artif. Intell. vol. 35, pp. 1673-1681. Association for the Advancement of Artificial Intelligence (AAAI) (May 2021). https://doi.org/10.1609/aaai.v35i2.16260 +16. Joos, V., Somers, V., Standaert, B.: TrackLab. https://github.com/TrackingLaboratory/tracklab (2024) +17. Kiran, M., Praveen, R.G., Nguyen-Meidine, L.T., Belharbi, S., Blais-Morin, L.A., Granger, E.: Holistic guidance for occluded person re-identification. arXiv abs/2104.06524 (2021). https://doi.org/10.48550/arXiv.2104.06524 +18. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv abs/2304.02643 (2023). https://doi.org/10.48550/arXiv.2304.02643 +19. Kreiss, S., Bertoni, L., Alahi, A.: PifPaf: Composite fields for human pose estimation. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 11969-11978. Inst. Electr. Electron. Eng. (IEEE), Long Beach, CA, USA (Jun 2019). https://doi.org/10.1109/cvpr.2019.01225 +20. Leal-Taixe, L., Milan, A., Reid, I., Roth, S., Schindler, K.: MOTChallenge 2015: Towards a benchmark for multi-target tracking. arXiv abs/1504.01942 (2015). https://doi.org/10.48550/arXiv.1504.01942 +21. Li, Y., He, J., Zhang, T., Liu, X., Zhang, Y., Wu, F.: Diverse part discovery: Occluded person re-identification with part-aware transformer. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 2897-2906. Inst. Electr. Electron. Eng. (IEEE), Nashville, TN, USA (Jun 2021). https://doi.org/10.1109/cvpr46437.2021.00292 +22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: Eur. Conf. Comput. Vis. (ECCV). Lect. Notes Comput. Sci., vol. 8693, pp. 740-755. Springer Int. Publ. (2014). https://doi.org/10.1007/978-3-319-10602-1_48 +23. Liu, Q., Xu, Z., Bertasius, G., Niethammer, M.: SimpleClick: Interactive image segmentation with simple vision transformers. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 22233-22243. Inst. Electr. Electron. Eng. (IEEE), Paris, Fr. (Oct 2023). https://doi.org/10.1109/iccv51070.2023.02037 +24. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 9992-10002. Inst. Electr. Electron. Eng. (IEEE), Montreal, QC, Canada (Oct 2021). https://doi.org/10.1109/iccv48922.2021.00986 + +25. Luiten, J., Osuep, A., Dendorfer, P., Torr, P., Geiger, A., Leal-Taixe, L., Leibe, B.: HOTA: A higher order metric for evaluating multi-object tracking. Int. J. Comput. Vis. 129(2), 548-578 (Oct 2020). https://doi.org/10.1007/s11263-020-01375-2 +26. Luo, H., Gu, Y., Liao, X., Lai, S., Jiang, W.: Bag of tricks and a strong baseline for deep person re-identification. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Work. (CVPRW). pp. 1487-1495. Inst. Electr. Electron. Eng. (IEEE), Long Beach, CA, USA (Jun 2019). https://doi.org/10.1109/cvprw.2019.00190 +27. Ma, Z., Zhao, Y., Li, J.: Pose-guided inter- and intra-part relational transformer for occluded person re-identification. In: Proceedings of the 29th ACM International Conference on Multimedia. ACM (Oct 2021). https://doi.org/10.1145/3474085.3475283 +28. Mansourian, A.M., Somers, V., De Vleeschouwer, C., Kasaei, S.: Multi-task learning for joint re-identification, team affiliation, and role classification for sports visual tracking. In: Proceedings of the 6th International Workshop on Multimedia Content Analysis in Sports. ACM (Oct 2023). https://doi.org/10.1145/3606038.3616172 +29. Miao, J., Wu, Y., Liu, P., Ding, Y., Yang, Y.: Pose-guided feature alignment for occluded person re-identification. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 542-551. Inst. Electr. Electron. Eng. (IEEE) (Oct 2019). https://doi.org/10.1109/iccv.2019.00063 +30. Peng, Y., Hou, S., Cao, C., Liu, X., Huang, Y., He, Z.: Deep learning-based occluded person re-identification: A survey. arXiv abs/2207.14452 (2022). https://doi.org/10.48550/arXiv.2207.14452 +31. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet large scale visual recognition challenge. Int. J. Comput. Vis. 115(3), 211-252 (Apr 2015). https://doi.org/10.1007/s11263-015-0816-y +32. Seidenschwarz, J., Brasó, G., Serrano, V.C., Elezi, I., Leal-Taixe, L.: Simple cues lead to a strong multi-object tracker. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 13813-13823. Inst. Electr. Electron. Eng. (IEEE), Vancouver, Can. (Jun 2023). https://doi.org/10.1109/cvpr52729.2023.01327 +33. Somers, V., Joos, V., Giancola, S., Cioppa, A., Ghasemzadeh, S.A., Magera, F., Standaert, B., Mansourian, A.M., Zhou, X., Kasaei, S., Ghanem, B., Alahi, A., Van Droogenbroeck, M., De Vleeschouwer, C.: SoccerNet game state reconstruction: End-to-end athlete tracking and identification on a minimap. In: IEEE Int. Conf. Comput. Vis. Pattern Recognit. Work. (CVPRW), CVsports. Seattle, WA, USA (Jun 2024). https://doi.org/10.48550/ARXIV.2404.11335 +34. Somers, V., Vleeschouwer, C.D., Alahi, A.: Body part-based representation learning for occluded person re-identification. In: IEEE/CVF Winter Conf. Appl. Comput. Vis. (WACV). pp. 1613–1623. Inst. Electr. Electron. Eng. (IEEE), Waikoloa, HI, USA (Jan 2023). https://doi.org/10.1109/wacv56688.2023.00166 +35. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 5686-5696. Inst. Electr. Electron. Eng. (IEEE), Long Beach, CA, USA (Jun 2019). https://doi.org/10.1109/cvpr.2019.00584 +36. Sun, Y., Zheng, L., Yang, Y., Tian, Q., Wang, S.: Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline). In: Eur. Conf. Comput. Vis. (ECCV). Lect. Notes Comput. Sci., vol. 11208, pp. 501-518. Springer Int. Publ. (2018). https://doi.org/10.1007/978-3-030-01225-0_30 + +37. Tan, H., Liu, X., Yin, B., Li, X.: MHSA-net: Multihead self-attention network for occluded person re-identification. IEEE Trans. Neural Networks Learn. Syst. 34(11), 8210-8224 (Nov 2023). https://doi.org/10.1109/tnnls.2022.3144163 +38. Van Zandycke, G., Somers, V., Istasse, M., Don, C.D., Zambrano, D.: DeepSportradar-v1: Computer vision dataset for sports understanding with high quality annotations. In: Proceedings of the 5th International ACM Workshop on Multimedia Content Analysis in Sports. ACM (Oct 2022). https://doi.org/10.1145/3552437.3555699 +39. Wang, G., Yang, S., Liu, H., Wang, Z., Yang, Y., Wang, S., Yu, G., Zhou, E., Sun, J.: High-order information matters: Learning relation and topology for occluded person re-identification. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 6448-6457. Inst. Electr. Electron. Eng. (IEEE), Seattle, WA, USA (Jun 2020). https://doi.org/10.1109/cvpr42600.2020.00648 +40. Wang, T., Liu, H., Song, P., Guo, T., Shi, W.: Pose-guided feature disentangling for occluded person re-identification based on transformer. In: AAAI Conf. Artif. Intell. vol. 36, pp. 2540-2549. Association for the Advancement of Artificial Intelligence (AAAI) (Jun 2022). https://doi.org/10.1609/aaai.v36i3.20155 +41. Wang, Z., Zhu, F., Tang, S., Zhao, R., He, L., Song, J.: Feature erasing and diffusion network for occluded person re-identification. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 4744-4753. Inst. Electr. Electron. Eng. (IEEE), New Orleans, LA, USA (Jun 2022). https://doi.org/10.1109/cvpr52688.2022.00471 +42. Wang, Z., Zheng, L., Liu, Y., Li, Y., Wang, S.: Towards real-time multi-object tracking. In: Eur. Conf. Comput. Vis. (ECCV). Lect. Notes Comput. Sci., vol. 12356, pp. 107-122. Springer Int. Publ. (2020). https://doi.org/10.1007/978-3-030-58621-8_7 +43. Wei, L., Zhang, S., Gao, W., Tian, Q.: Person transfer GAN to bridge domain gap for person re-identification. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 79–88. Inst. Electr. Electron. Eng. (IEEE), Salt Lake City, UT, USA (Jun 2018). https://doi.org/10.1109/cvpr.2018.00016 +44. Yan, C., Pang, G., Jiao, J., Bai, X., Feng, X., Shen, C.: Occluded person re-identification with single-scale global representations. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 11855-11864. Inst. Electr. Electron. Eng. (IEEE), Montreal, QC, Canada (Oct 2021). https://doi.org/10.1109/iccv48922.2021.01166 +45. Yang, J., Zhang, J., Yu, F., Jiang, X., Zhang, M., Sun, X., Chen, Y., Zheng, W.S.: Learning to know where to see: A visibility-aware approach for occluded person re-identification. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 11865-11874. Inst. Electr. Electron. Eng. (IEEE), Montreal, QC, Canada (Oct 2021). https://doi.org/10.1109/iccv48922.2021.01167 +46. Yang, Q., Wang, P., Fang, Z., Lu, Q.: Focus on the visible regions: Semantic-guided alignment model for occluded person re-identification. Sensors 20(16), 1-15 (Aug 2020). https://doi.org/10.3390/s20164431 +47. Ye, M., Shen, J., Lin, G., Xiang, T., Shao, L., Hoi, S.C.H.: Deep learning for person re-identification: A survey and outlook. IEEE Trans. Pattern Anal. Mach. Intell. 44(6), 2872–2893 (Jun 2022). https://doi.org/10.1109/tpami.2021.3054775 +48. Zang, X., Li, G., Gao, W., Shu, X.: Learning to disentangle scenes for person re-identification. Image Vis. Comput. 116, 104330 (Dec 2021). https://doi.org/10.1016/j.imavis.2021.104330 +49. Zhang, G., Zhang, Y., Zhang, T., Li, B., Pu, S.: PHA: Patch-wise high-frequency augmentation for transformer-based person re-identification. In: IEEE/CVF Conf. + +Comput. Vis. Pattern Recognit. (CVPR). pp. 14133-14142. Inst. Electr. Electron. Eng. (IEEE), Vancouver, Can. (Jun 2023). https://doi.org/10.1109/cvpr52729.2023.01358 +50. Zhang, Y., Sun, P., Jiang, Y., Yu, D., Weng, F., Yuan, Z., Luo, P., Liu, W., Wang, X.: ByteTrack: Multi-object tracking by associating every detection box. In: Eur. Conf. Comput. Vis. (ECCV). Lect. Notes Comput. Sci., vol. 13682, pp. 1-21. Springer Nat. Switz. (2022). https://doi.org/10.1007/978-3-031-20047-2_1, https://doi.org/10.1007/978-3-031-20047-2_1 +51. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: FairMOT: On the fairness of detection and re-identification in multiple object tracking. Int. J. Comput. Vis. 129(11), 3069–3087 (Sept 2021). https://doi.org/10.1007/s11263-021-01513-4 +52. Zheng, K., Lan, C., Zeng, W., Liu, J., Zhang, Z., Zha, Z.J.: Pose-guided feature learning with knowledge distillation for occluded person re-identification. arXiv abs/2108.00139 (2021). https://doi.org/10.48550/arXiv.2108.00139 +53. Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., Tian, Q.: Scalable person re-identification: A benchmark. In: IEEE Int. Conf. Comput. Vis. (ICCV). pp. 1116-1124. Inst. Electr. Electron. Eng. (IEEE), Santiago, Chile (Dec 2015). https://doi.org/10.1109/iccv.2015.133 +54. Zheng, W.S., Li, X., Xiang, T., Liao, S., Lai, J., Gong, S.: Partial person re-identification. In: IEEE Int. Conf. Comput. Vis. (ICCV). pp. 4678-4686. Inst. Electr. Electron. Eng. (IEEE), Santiago, Chile (Dec 2015). https://doi.org/10.1109/iccv.2015.531 +55. Zhong, Z., Zheng, L., Cao, D., Li, S.: Re-ranking person re-identification with k-reciprocal encoding. In: IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 3652-3661. Inst. Electr. Electron. Eng. (IEEE), Honolulu, HI, USA (Jul 2017). https://doi.org/10.1109/cvpr.2017.389 +56. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. In: AAAI Conf. Artif. Intell. vol. 34, pp. 13001-13008. Association for the Advancement of Artificial Intelligence (AAAI) (Apr 2020). https://doi.org/10.1609/aaai.v34i07.7000 +57. Zhu, K., Guo, H., Liu, Z., Tang, M., Wang, J.: Identity-guided human semantic parsing for person re-identification. In: Eur. Conf. Comput. Vis. (ECCV). Lect. Notes Comput. Sci., vol. 12348, pp. 346-363. Springer Int. Publ. (2020). https://doi.org/10.1007/978-3-030-58580-8_21 +58. Zhuo, J., Chen, Z., Lai, J., Wang, G.: Occluded person re-identification. arXiv abs/1804.02792 (2018). https://doi.org/10.48550/arXiv.1804.02792 \ No newline at end of file diff --git a/2024/Keypoint Promptable Re-Identification/images.zip b/2024/Keypoint Promptable Re-Identification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..51314f0e8eee5fa2aeab597b08af71f6c157bd3b --- /dev/null +++ b/2024/Keypoint Promptable Re-Identification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3504af4d11841048a8608ea9a302bcb02797f27a1d6286aeafed86c9cc3d1ca1 +size 528821 diff --git a/2024/Keypoint Promptable Re-Identification/layout.json b/2024/Keypoint Promptable Re-Identification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e6e101410cd6dc4496420e48712c3081d5c1eb6c --- /dev/null +++ b/2024/Keypoint Promptable Re-Identification/layout.json @@ -0,0 +1,9226 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 170, + 112, + 444, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 112, + 444, + 129 + ], + "spans": [ + { + "bbox": [ + 170, + 112, + 444, + 129 + ], + "type": "text", + "content": "Keypoint Promptable Re-Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 149, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 482, + 164 + ], + "type": "text", + "content": "Vladimir Somers1,2,3, Alexandre Alahi1, and Christophe De Vleeschouwer2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 171, + 171, + 443, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 171, + 443, + 182 + ], + "spans": [ + { + "bbox": [ + 171, + 171, + 443, + 182 + ], + "type": "text", + "content": "1 Ecole Polytechnique Fédérale de Lausanne (EPFL), Switzerland" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 187, + 182, + 427, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 182, + 427, + 194 + ], + "spans": [ + { + "bbox": [ + 187, + 182, + 427, + 194 + ], + "type": "text", + "content": "2 Université Catholique de Louvain (UCLouvain), Belgium" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 253, + 194, + 361, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 194, + 361, + 205 + ], + "spans": [ + { + "bbox": [ + 253, + 194, + 361, + 205 + ], + "type": "text", + "content": "3 Sportradar, Switzerland" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 236, + 455, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 236, + 455, + 478 + ], + "spans": [ + { + "bbox": [ + 159, + 236, + 455, + 478 + ], + "type": "text", + "content": "Abstract. Occluded Person Re-Identification (ReID) is a metric learning task that involves matching occluded individuals based on their appearance. While many studies have tackled occlusions caused by objects, multi-person occlusions remain less explored. In this work, we identify and address a critical challenge overlooked by previous occluded ReID methods: the Multi-Person Ambiguity (MPA) arising when multiple individuals are visible in the same bounding box, making it impossible to determine the intended ReID target among the candidates. Inspired by recent work on prompting in vision, we introduce Keypoint Promptable ReID (KPR), a novel formulation of the ReID problem that explicitly complements the input bounding box with a set of semantic keypoints indicating the intended target. Since promptable re-identification is an unexplored paradigm, existing ReID datasets lack the pixel-level annotations necessary for prompting. To bridge this gap and foster further research on this topic, we introduce Occluded PoseTrack-ReID, a novel ReID dataset with keypoints labels, that features strong inter-person occlusions. Furthermore, we release custom keypoints labels for four popular ReID benchmarks. Experiments on person retrieval, but also on pose tracking, demonstrate that our method systematically surpasses previous state-of-the-art approaches on various occluded scenarios. Our code, dataset and annotations are available at https://github.com/V1Somers/Keypoint_prompts_reidentification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 487, + 453, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 487, + 453, + 499 + ], + "spans": [ + { + "bbox": [ + 160, + 487, + 453, + 499 + ], + "type": "text", + "content": "Keywords: Person Re-Identification " + }, + { + "bbox": [ + 160, + 487, + 453, + 499 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 487, + 453, + 499 + ], + "type": "text", + "content": " Vision Prompting " + }, + { + "bbox": [ + 160, + 487, + 453, + 499 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 487, + 453, + 499 + ], + "type": "text", + "content": " Pose Tracking" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 520, + 230, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 230, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 230, + 533 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 616 + ], + "type": "text", + "content": "Person re-identification (ReID) 47 is a challenging retrieval task that involves matching a query image of a person of interest with other person images. ReID finds wide-ranging applications in multi-object tracking 32, pedestrian flow analysis 2, sport understanding 11,14,28,33,38 and video-surveillance 53. However, ReID is a difficult task due to various factors including inaccurate bounding boxes, luminosity changes, poor image quality, and occlusions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": "In recent years, there has been a growing interest in addressing the occluded re-identification task [30], where the ReID target might be occluded by various objects or other people. In this work, we identify and explicitly address a critical challenge that has been overlooked by previous occluded ReID methods [34, 41]:" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 213, + 114, + 396, + 187 + ], + "blocks": [ + { + "bbox": [ + 213, + 114, + 396, + 187 + ], + "lines": [ + { + "bbox": [ + 213, + 114, + 396, + 187 + ], + "spans": [ + { + "bbox": [ + 213, + 114, + 396, + 187 + ], + "type": "image", + "image_path": "1ad9df224b74aacc1b686ebbf4388b878fee5e4767ba9672bc84d64a38a9575f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 189, + 376, + 198 + ], + "lines": [ + { + "bbox": [ + 239, + 189, + 376, + 198 + ], + "spans": [ + { + "bbox": [ + 239, + 189, + 376, + 198 + ], + "type": "text", + "content": "(a) Keypoint Promptable ReID (KPR)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 149, + 199, + 465, + 274 + ], + "blocks": [ + { + "bbox": [ + 149, + 199, + 465, + 274 + ], + "lines": [ + { + "bbox": [ + 149, + 199, + 465, + 274 + ], + "spans": [ + { + "bbox": [ + 149, + 199, + 465, + 274 + ], + "type": "image", + "image_path": "4b0aa7420085aaac44a1af8e4a9c3f46db364f2299ebfc01a9aa11829c19271f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 281, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 131, + 281, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 131, + 281, + 482, + 338 + ], + "type": "text", + "content": "Fig. 1: Overview of our proposed Keypoint Promptable ReID (KPR) method. KPR takes an image with keypoints prompts as input and produces part-based features of the prompted target. The prompt instructs the model to focus on a specific individual, i.e. the back blue jacket man (b) or the front black t-shirt man (c) in this example. Colored dots illustrate the positive keypoints prompt, with one color per body part." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 367, + 482, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 440 + ], + "type": "text", + "content": "the presence of Multi-Person Ambiguity (MPA) inherent to bounding box annotations. The multi-person ambiguity arises when a bounding box image contains multiple individuals, making it challenging even for humans to accurately identify the intended ReID target among all the candidates. Figure 2a shows how MPA can negatively impact the re-identification model, causing feature mixing among different individuals or even a focus on an unintended person." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 442, + 483, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 442, + 483, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 442, + 483, + 502 + ], + "type": "text", + "content": "To tackle this multi-person ambiguity, it is necessary to incorporate additional pixel-level information, such as keypoints or segmentation masks, provided by an upstream human operator or vision model. This additional data plays a vital role in assisting the downstream ReID method to distinguish the intended target from other candidates." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 506, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 590 + ], + "type": "text", + "content": "Inspired by recent advances in promptable" + }, + { + "bbox": [ + 130, + 506, + 482, + 590 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 130, + 506, + 482, + 590 + ], + "type": "text", + "content": " methods for vision transformers (e.g., SAM [18]), we propose to explicitly use additional semantic" + }, + { + "bbox": [ + 130, + 506, + 482, + 590 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 130, + 506, + 482, + 590 + ], + "type": "text", + "content": " keypoints as inputs to disambiguate images with multiple individuals. Semantic keypoints are suitable for two crucial use cases: 1) manual prompting by a human operator with a few clicks on an image, and 2) automated prompting by an upstream pose estimation model. Nevertheless, our architecture could be easily extended to support segmentation masks as prompts." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "V. Somers et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 609, + 481, + 643 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 609, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 609, + 481, + 643 + ], + "type": "text", + "content": "4 A \"prompt\" typically refers to an input instruction that guides the model's response. In vision tasks, e.g. segmentation, prompts can be any kind of data that specifies what the model should focus on [18]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 643, + 481, + 667 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 481, + 667 + ], + "type": "text", + "content": "5 Semantic keypoints are distinct points within a person image that carry semantic information about specific body parts, typically obtained with pose estimation [22]." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 301, + 179 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 301, + 179 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 301, + 179 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 301, + 179 + ], + "type": "image", + "image_path": "f6b8068b9db777551cc624f85484631d293c415cc09021a88e77ae37c3cf6be2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 182, + 299, + 199 + ], + "lines": [ + { + "bbox": [ + 137, + 182, + 299, + 199 + ], + "spans": [ + { + "bbox": [ + 137, + 182, + 299, + 199 + ], + "type": "text", + "content": "(a) Existing ReID models (e.g. BPBreID 34) are confused by MPA" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 493, + 181 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 493, + 181 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 493, + 181 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 493, + 181 + ], + "type": "image", + "image_path": "c40029641172a0923771296c580d38c1a0bfb111c02ae15721403ee6da5485e9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 182, + 481, + 200 + ], + "lines": [ + { + "bbox": [ + 310, + 182, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 310, + 182, + 481, + 200 + ], + "type": "text", + "content": "(b) KPR explicitly addresses MPA with positive (pastel) and negative (red) keypoints prompt" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 208, + 482, + 232 + ], + "lines": [ + { + "bbox": [ + 130, + 208, + 482, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 208, + 482, + 232 + ], + "type": "text", + "content": "Fig. 2: Person retrieval with Multi-Person Ambiguity (MPA). Green/red borders are correct/incorrect matches. Red/pastel dots indicate negative/positive prompts." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 259, + 479, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 479, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 479, + 296 + ], + "type": "text", + "content": "Prior research [8,39,40] on pose-guided methods has explored the use of additional information from external pose estimation models. However, none of these works explicitly tackled the multi-pedestrian ambiguity issue." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 298, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 491 + ], + "type": "text", + "content": "Following above insights, we propose KPR, a novel Keypoint-Promptable transformer for part-based person Re-identification, illustrated in Fig. 1a. KPR is a transformer-based model that takes an image along with semantic keypoints as input. It then produces part-based features, each representing a distinct body part of the ReID target, along with their respective visibility scores. Part-based methods have demonstrated superior performance [34,57] in re-identifying occluded individuals because they utilize only the visible parts for comparison. Our method can process both positive and negative keypoints, which respectively represent the target and non-target pedestrians. As illustrated in Fig. 1b, prompts can be used to re-identify the desired (front or back) person in case of occlusion. Furthermore, KPR is designed to be prompt optional to offer more practical flexibility. This means the same model can be used without prompt on non-ambiguous images, or with prompt when dealing with occlusions, while consistently achieving state-of-the-art performance in both cases. To demonstrate the advantages of keypoint prompts for re-identification, we evaluate KPR on two tasks: person retrieval and multi-person pose tracking." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 492, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 482, + 552 + ], + "type": "text", + "content": "Moreover, since no current ReID dataset is provided with explicit target identification in case of multi-person ambiguity, we introduce a new dataset called Occluded PoseTrack-ReID, nicknamed \"Occ-PTrack\", and derived from the PoseTrack21 dataset [6]. This dataset offers high-quality keypoints annotations and features a substantial number of images with multi-person occlusions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 146, + 555, + 453, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 555, + 453, + 567 + ], + "spans": [ + { + "bbox": [ + 146, + 555, + 453, + 567 + ], + "type": "text", + "content": "Overall, we summarize the main contributions of our work as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 135, + 570, + 481, + 664 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 136, + 570, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 570, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 136, + 570, + 481, + 616 + ], + "type": "text", + "content": "1. We propose " + }, + { + "bbox": [ + 136, + 570, + 481, + 616 + ], + "type": "inline_equation", + "content": "KPR" + }, + { + "bbox": [ + 136, + 570, + 481, + 616 + ], + "type": "text", + "content": ", a novel prompt-optional model for part-based ReID. To our knowledge, we are the first to introduce a visual prompting mechanism for person re-identification, and the first to explicitly address the ambiguity caused by multi-person occlusions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 135, + 617, + 481, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 617, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 135, + 617, + 481, + 664 + ], + "type": "text", + "content": "2. We introduce Occ-PTrack: the first multi-person occluded ReID dataset with explicit target identification through manual keypoints annotations. Furthermore, we propose new keypoint annotations for four popular re-identification datasets." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 285, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 285, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 285, + 91, + 445, + 102 + ], + "type": "text", + "content": "Keypoint Promptable Re-Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 135, + 116, + 481, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 116, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 481, + 152 + ], + "type": "text", + "content": "3. Our method outperforms all previous state-of-the-art methods on the occluded ReID task, demonstrating the effectiveness of the keypoint promptable approach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 201 + ], + "type": "text", + "content": "Our ambition is to advocate for a shift from the ambiguous bounding-box approach to ReID and promote a keypoint-centric paradigm. To this end, we publicly release our codebase, proposed dataset, and annotations, to provide a common setup for evaluating promptable ReID methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 220, + 238, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 220, + 238, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 238, + 232 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 246, + 481, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 481, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 481, + 305 + ], + "type": "text", + "content": "Multi-Person Ambiguity: A previous study [41] addressed a related issue called \"Non-Target Pedestrians (NTP),\" similar to our concept of \"Multi-Person Ambiguity\" (MPA). However, their focus was on mitigating the negative impact of NTP during training. In contrast, we provide a method to explicitly disambiguate the intended ReID target during both training and testing phases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 307, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 482, + 605 + ], + "type": "text", + "content": "Pose-guided ReID: We detail here the fundamental difference between our novel keypoint promptable approach and the traditional pose-guided methods. Although none of the previous works explicitly address the multi-person ambiguity (MPA) issue, we identify two categories within these pose-guided approaches: 1) those inherently unable to address MPA [48,15,27,29,34,44,45], and 2) those with the potential to overcome MPA [39,40]. In the first category, methods like BPBreID [34] or Pirt [27] utilize pose labels only during training, and can not leverage additional information at test time to disambiguate multiple persons. Moreover, methods relying on horizontal stripes (VGTri [45], PGFA [29]) or the affinity fields of a pose model (PVPM [8]), are meant to tackle object-occlusions, but struggle to disambiguate multiple persons. In the second category, methods like HOReID [39] and PFD [40] directly employ keypoints from a single person, enabling focus on a specified target. However, HOReID [39] is restricted to utilizing spatial features at the exact keypoints locations, thus overlooking critical appearance cues on entire body regions. Finally, PFD [40], and all previous pose-guided methods including HOReID, incorporate pose information after appearance encoding has occurred, e.g., for local pooling of a spatial feature map, but do not leverage pose data to guide the appearance encoding process as our method do. Indeed, our approach implements a prompt-aware appearance encoding, since it conditions feature extraction on the input keypoints. This enables better feature disentanglement between the target and the occluders from the early feature extraction stages, resulting in more accurate representations. Furthermore, our method is prompt optional, setting it apart for its flexibility from these previous pose-guided methods. Finally, our method can ingest additional negative keypoints, to further disambiguate a multi-person occlusion scenario." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 606, + 484, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 484, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 484, + 666 + ], + "type": "text", + "content": "Prompting in Vision: Recent works have introduced promptable transformers to tackle object segmentation. For instance, SAM [18] employs two networks: a transformer encoder to generate generic spatial features and a lightweight promptable decoder to integrate user inputs for interactive segmentation. In contrast, our approach directly prompts a feature encoder, so that feature extrac" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "V. Somers et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 479, + 196 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 479, + 196 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 479, + 196 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 479, + 196 + ], + "type": "image", + "image_path": "ccd792ff4b3f7f393d4b555253e4b3874bb29def2fee108a0847d6dce535b7a7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "type": "text", + "content": "Fig. 3: Architecture overview of our proposed Keypoint Promptable ReID (KPR) model. The Batch-wise Inter-Person Occlusion (BIPO) data augmentation is first applied to generate artificial inter-person occlusions (Sec. 3.5). The image and the optional positive/negative " + }, + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "type": "inline_equation", + "content": "(\\oplus/\\ominus)" + }, + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "type": "text", + "content": " prompts are then tokenized and summed (Sec. 3.1). Tokens are then fed to our proposed Multi-Stage feature Fusion (MSF) Swin transformer to generate high-resolution feature maps (3.2). The feature map is then fed to a Part-based Head (PBH), which assigns a part (or the background) to each token with a part classifier, and then averages all tokens of the same part to produce the final " + }, + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "type": "text", + "content": " part-based embeddings " + }, + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\{f_1, \\dots, f_K\\}" + }, + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "type": "text", + "content": " and their binary visibility scores " + }, + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\{v_1, \\dots, v_K\\}" + }, + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "type": "inline_equation", + "content": "v_i \\in \\{0,1\\}" + }, + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "type": "text", + "content": ", and visually denoted by " + }, + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}, \\check{\\mathbf{v}}\\}" + }, + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "type": "text", + "content": " (Sec. 3.3). A part " + }, + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "type": "text", + "content": " with no token assigned is considered invisible (i.e., " + }, + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "type": "inline_equation", + "content": "v_i = 0" + }, + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "type": "text", + "content": "), and ignored when computing two samples' similarity. KPR is illustrated here for " + }, + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "type": "text", + "content": ", with a unique color for each body part: " + }, + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\{\\text{head}, \\text{torso}, \\text{right/left arm}, \\text{right/left leg}, \\text{right/left feet}\\}" + }, + { + "bbox": [ + 130, + 205, + 482, + 361 + ], + "type": "text", + "content": ". Finally, the entire pipeline is trained with two losses: a Part-Prediction (PP) Loss and a ReID Loss (Sec. 3.4)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 382, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 482, + 431 + ], + "type": "text", + "content": "tion is explicitly driven by the prompts. Unlike SimpleClick [23], which prompts a transformer with a single 2D point and a previous mask for iterative segmentation, our model ingests multiple keypoints concurrently, enhancing body part localization and representation by leveraging their semantic information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 448, + 233, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 448, + 233, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 448, + 233, + 462 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 472, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 534 + ], + "type": "text", + "content": "The overall architecture of our model KPR is illustrated in Fig. 3 and is based on a Swin transformer backbone [24]. The next subsections first describe the various components of the architecture. We then present the training process, including our proposed data augmentation. Finally, we describe the inference procedure and conclude with an intuitive discussion of our design choices." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 550, + 226, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 550, + 226, + 561 + ], + "spans": [ + { + "bbox": [ + 132, + 550, + 226, + 561 + ], + "type": "text", + "content": "3.1 Tokenization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 593 + ], + "type": "text", + "content": "The tokenization procedure coverts raw image and keypoints information into " + }, + { + "bbox": [ + 130, + 570, + 481, + 593 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 130, + 570, + 481, + 593 + ], + "type": "text", + "content": "-dimensional tokens, to be subsequently processed by the Swin encoder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 641 + ], + "type": "text", + "content": "Image tokenization: The input image, with dimensions " + }, + { + "bbox": [ + 130, + 594, + 481, + 641 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 130, + 594, + 481, + 641 + ], + "type": "text", + "content": ", undergoes tokenization using the conventional Swin patch embed module. This process generates " + }, + { + "bbox": [ + 130, + 594, + 481, + 641 + ], + "type": "inline_equation", + "content": "H_{t} \\times W_{t} = \\frac{H}{4} \\times \\frac{W}{4}" + }, + { + "bbox": [ + 130, + 594, + 481, + 641 + ], + "type": "text", + "content": " spatial tokens of dimensions " + }, + { + "bbox": [ + 130, + 594, + 481, + 641 + ], + "type": "inline_equation", + "content": "C_{i}" + }, + { + "bbox": [ + 130, + 594, + 481, + 641 + ], + "type": "text", + "content": " by embedding each " + }, + { + "bbox": [ + 130, + 594, + 481, + 641 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 130, + 594, + 481, + 641 + ], + "type": "text", + "content": " image patch through a linear layer." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Keypoints tokenization: As previously mentioned, our model uses two types of semantic keypoints as prompts: positive keypoints indicating the ReID" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 285, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 285, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 285, + 91, + 445, + 102 + ], + "type": "text", + "content": "Keypoint Promptable Re-Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": "target and negative keypoints indicating other individuals. We organize these prompt keypoints into " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\mathrm{K} + 1" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " groups: the first group contains all negative keypoints, and the remaining " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\mathrm{K}" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " groups contain subsets of the positive keypoints. Each subset of positive keypoints actually corresponds to a semantic body region. For instance, with " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": ", we define groups like: " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\{\\text{head, torso, right/left arm, right/left leg, and right/left feet}\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": ". For each group " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "i \\in \\{1, \\dots, K + 1\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": ", a unique heatmap " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "M_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " is created by drawing a Gaussian 2D kernel at each of its keypoints's center location " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": ". The kernel's standard deviation is set to a fraction " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\alpha_G" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " of the image width. Invisible keypoints are ignored. These " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "K + 1" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " heatmaps are then concatenated into a global spatial tensor " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "H \\times W \\times (K + 1)" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": ". Finally, each " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " patch in " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " is linearly projected to produce prompt tokens of dimension " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": ". Like the image tokenizer, the keypoints tokenizer outputs therefore " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "H_t \\times W_t" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " spatial tokens of dimension " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 271, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 307 + ], + "type": "text", + "content": "Token fusion: Images and keypoints tokens are fused by summing tokens at the same spatial location, producing " + }, + { + "bbox": [ + 130, + 271, + 482, + 307 + ], + "type": "inline_equation", + "content": "H_{t} \\times W_{t}" + }, + { + "bbox": [ + 130, + 271, + 482, + 307 + ], + "type": "text", + "content": " spatial tokens with dimension " + }, + { + "bbox": [ + 130, + 271, + 482, + 307 + ], + "type": "inline_equation", + "content": "C_{i}" + }, + { + "bbox": [ + 130, + 271, + 482, + 307 + ], + "type": "text", + "content": ". These tokens are then forwarded to the Swin-based encoder." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 482, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 463 + ], + "type": "text", + "content": "Design motivations: Prompted keypoints are organized into " + }, + { + "bbox": [ + 130, + 308, + 482, + 463 + ], + "type": "inline_equation", + "content": "K + 1" + }, + { + "bbox": [ + 130, + 308, + 482, + 463 + ], + "type": "text", + "content": " channels to preserve semantic meaning before tokenization. It allows the model to differentiate positive and negative prompts, and to distinguish prompt information of different body parts, for better part localization. Prompt-aware appearance encoding: Moreover, feeding the network with keypoint information before the encoding stage is critical in enabling the network to disentangle the ReID target features from other persons' features. Finally, this design offers the flexibility to support segmentation mask prompts by replacing " + }, + { + "bbox": [ + 130, + 308, + 482, + 463 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 308, + 482, + 463 + ], + "type": "text", + "content": " with a double-channel map of positive and negative segmentation masks. Regarding token fusion, the sum operation maintains the input dimension for the Swin backbone, permitting the use of pre-trained weights from the image-only model. Prompt optionality: This design also offers the flexibility of running the model without prompts, by processing the image tokens alone, with no architectural adjustments required." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 479, + 259, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 259, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 259, + 491 + ], + "type": "text", + "content": "3.2 MSF Swin Encoder" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": "A Swin-transformer [24] is employed as the backbone feature extractor for its great performance in vision tasks and its ability to capture high-frequency details, which is beneficial to ReID [49]. Moreover, we propose to enhance Swin with a Multi-Stage feature Fusion strategy, denoted as \"MSF\", to output feature maps with a high spatial resolution. Indeed, similar to CNN, each Swin stage reduces the number of tokens by a factor of 4 while doubling the channel dimension. To obtain a high-resolution token map, the outputs of all stages are bilinearly upsampled to the same resolution " + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "inline_equation", + "content": "H_{t} \\times W_{t}" + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": " as the output of the tokenization module, then concatenated into a single feature map along the channel dimension, and finally passed through a linear layer to obtain the final spatial token map of size " + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "inline_equation", + "content": "H_{t} \\times W_{t} \\times C_{o}" + }, + { + "bbox": [ + 130, + 498, + 482, + 667 + ], + "type": "text", + "content": ". Different from the standard Swin, our MSF Swin therefore preserves the input tokens resolution. As pointed out in prior works [26,57], a high-resolution feature map enriches the granularity of the features and brings significant performance improvements (see Tab. 3, Exp. 3)." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "V. Somers et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 246, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 246, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 246, + 126 + ], + "type": "text", + "content": "3.3 Part-based Head" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": "The encoder yields " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "H_{t} \\times W_{t}" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": " tokens, which are sent to the Part-based Head (PBH) module, as illustrated in Fig. 3. Each token is first labeled either as one of the " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": " body parts or as background, by a token-wise part classifier. The part classifier is implemented as a linear layer with " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "K + 1" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": " output logits followed by the softmax operation, and is applied individually on each token. The assignment of each token to one of the " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": " body parts (or the background) is therefore formulated as a classification task with " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "K + 1" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": " classes. We term the probability map of size " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "H_{t} \\times W_{t} \\times (K + 1)" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": " produced by this part classifier over all tokens as \"part-attention maps\". These maps are showcased as colored heatmaps in Fig. 1a and Fig. 6b with one color per channel (the background channel is ignored). Finally, tokens having the same assigned body part are averaged together, resulting in " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": " part-based embeddings " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "\\{f_{1}, \\dots, f_{K}\\}" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "C_{o}" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": ". Background tokens are ignored. Parts with no tokens assigned are considered invisible and given a visibility score " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "v_{i} = 0" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": ". The corresponding invisible embedding " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "v_{i} = 0" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": ", is therefore ignored for downstream computation. The PBH module thus produces " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": " part-based embeddings " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "\\{f_{1}, \\dots, f_{K}\\}" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": " and their binary visibility scores " + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "inline_equation", + "content": "\\{v_{1}, \\dots, v_{K}\\}" + }, + { + "bbox": [ + 130, + 134, + 482, + 326 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 326, + 481, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 481, + 376 + ], + "type": "text", + "content": "Human parsing labels: To train the token-wise part classifier, each spatial token is assigned a ground-truth class, i.e., part label, derived from the coarse human parsing labels provided by [34]. These human parsing labels are illustrated on the right of Fig. 3 and in Fig. 4a." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 390, + 246, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 390, + 246, + 403 + ], + "spans": [ + { + "bbox": [ + 131, + 390, + 246, + 403 + ], + "type": "text", + "content": "3.4 Training Process" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 407, + 384, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 407, + 384, + 419 + ], + "spans": [ + { + "bbox": [ + 131, + 407, + 384, + 419 + ], + "type": "text", + "content": "The overall objective function used to train the model is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 261, + 426, + 481, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 426, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 261, + 426, + 481, + 439 + ], + "type": "interline_equation", + "content": "L = L _ {\\text {R e I D}} + \\lambda_ {p p} L _ {p p}, \\tag {1}", + "image_path": "4a8e98b7be26e2ccffceca503864f7d40f16bccc76474f89a0dc3a8588ef08cc.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 445, + 481, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 445, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 131, + 445, + 481, + 468 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 445, + 481, + 468 + ], + "type": "inline_equation", + "content": "L_{ReID}" + }, + { + "bbox": [ + 131, + 445, + 481, + 468 + ], + "type": "text", + "content": " is the ReID loss supervised with identity labels, " + }, + { + "bbox": [ + 131, + 445, + 481, + 468 + ], + "type": "inline_equation", + "content": "L_{pp}" + }, + { + "bbox": [ + 131, + 445, + 481, + 468 + ], + "type": "text", + "content": " is the token-wise part-prediction loss, and " + }, + { + "bbox": [ + 131, + 445, + 481, + 468 + ], + "type": "inline_equation", + "content": "\\lambda_{pp}" + }, + { + "bbox": [ + 131, + 445, + 481, + 468 + ], + "type": "text", + "content": " is a weight parameter empirically set to 0.3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 468, + 482, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 504 + ], + "type": "text", + "content": "Part Prediction Loss: The token-wise part classifier introduced in Sec. 3.3 is supervised with a Part Prediction (PP) Loss that is actually a cross-entropy loss applied on each spatial token as in 34." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 504, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 564 + ], + "type": "text", + "content": "Re-identification loss: As a ReID objective, we employ the GiLt Loss [34], that is specifically designed to train part-based ReID models. It has the benefit of being robust to occlusions and non-discriminative body parts. GiLt combines an identity loss [26] and a batch-hard triplet loss [13]. The distance function employed for the triplet loss is the average of all local part-based distances." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 580, + 294, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 580, + 294, + 593 + ], + "spans": [ + { + "bbox": [ + 131, + 580, + 294, + 593 + ], + "type": "text", + "content": "3.5 BIPO Data Augmentation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 599, + 481, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 481, + 650 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 481, + 650 + ], + "type": "text", + "content": "Since most ReID datasets feature limited occlusion in their training set, a model often struggles to handle the occlusions it encounters during testing. To address this, we propose a novel Batch-wise Inter-Person Occlusion data augmentation (BIPO) to generate artificial inter-person occlusions on training images," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 284, + 91, + 445, + 102 + ], + "type": "text", + "content": "Keypoint Promptable Re-Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 652, + 382, + 667 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 382, + 667 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 382, + 667 + ], + "type": "text", + "content": "Inspired by the copy-paste technique for segmentation [10]." + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 0, + 114, + 100, + 217 + ], + "blocks": [ + { + "bbox": [ + 0, + 114, + 100, + 217 + ], + "lines": [ + { + "bbox": [ + 0, + 114, + 100, + 217 + ], + "spans": [ + { + "bbox": [ + 0, + 114, + 100, + 217 + ], + "type": "image", + "image_path": "f920778012ac8ec4370cc21f7bcba16027afeb32e5903055918ca03bf4ed620a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 237, + 482, + 282 + ], + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 282 + ], + "type": "text", + "content": "Fig. 4: (a) Our proposed Batch-wise Inter-Person Occlusion (BIPO) data augmentation creates artificial person occlusions that are consistent across image, prompt, and human parsing labels. BIPO is crucial to enforce the model to rely on the input prompts. (b) Four identities with their keypoints labels from our Occ-PTrack dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 154, + 114, + 252, + 217 + ], + "blocks": [ + { + "bbox": [ + 154, + 114, + 252, + 217 + ], + "lines": [ + { + "bbox": [ + 154, + 114, + 252, + 217 + ], + "spans": [ + { + "bbox": [ + 154, + 114, + 252, + 217 + ], + "type": "image", + "image_path": "a2200e0b78efa351fdf741dffc48eda9af5d1d4c8a77abb1c1ffaeeed093d1c0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 219, + 221, + 228 + ], + "lines": [ + { + "bbox": [ + 186, + 219, + 221, + 228 + ], + "spans": [ + { + "bbox": [ + 186, + 219, + 221, + 228 + ], + "type": "text", + "content": "(a) BIPO" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 297, + 114, + 451, + 217 + ], + "blocks": [ + { + "bbox": [ + 297, + 114, + 451, + 217 + ], + "lines": [ + { + "bbox": [ + 297, + 114, + 451, + 217 + ], + "spans": [ + { + "bbox": [ + 297, + 114, + 451, + 217 + ], + "type": "image", + "image_path": "ff517f87e4513a9e9f3ada78b005867e1cffe85282c53d2ca0a87b31fee112e9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 219, + 428, + 228 + ], + "lines": [ + { + "bbox": [ + 321, + 219, + 428, + 228 + ], + "spans": [ + { + "bbox": [ + 321, + 219, + 428, + 228 + ], + "type": "text", + "content": "(b) Occluded PoseTrack-ReID" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 506, + 115, + 603, + 217 + ], + "blocks": [ + { + "bbox": [ + 506, + 115, + 603, + 217 + ], + "lines": [ + { + "bbox": [ + 506, + 115, + 603, + 217 + ], + "spans": [ + { + "bbox": [ + 506, + 115, + 603, + 217 + ], + "type": "image", + "image_path": "abf05153db381133534dd36eaef671a57cf8d8e7572d428bd7274a85f83ab5b7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 304, + 482, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 387 + ], + "type": "text", + "content": "prompts, and human parsing labels. When provided with a training sample, BIPO randomly selects a different person's image (the occluder) from the same training batch, contours it with a segmentation mask derived from the human parsing labels, and overlays it on the main image. The human parsing label and keypoints prompt of the training image are then updated accordingly. Finally, all positive keypoints from the occluder are added to the training prompt to serve as additional negative points." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "type": "text", + "content": "Design motivations: Given that a triplet loss with batch-hard mining [13] is employed during training, using images from the same training batch to generate occluders yields hard positive and negative samples for mining triplets, thereby creating challenging scenarios for contrastive learning. BIPO is illustrated in Fig. 4a and its significance for KPR is further discussed in Sec. 3.7." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 464, + 208, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 464, + 208, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 464, + 208, + 475 + ], + "type": "text", + "content": "3.6 Inference" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "type": "text", + "content": "At inference, KPR processes each input image together with their respective keypoints prompts, and produces " + }, + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "type": "text", + "content": " part-based representations " + }, + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\{f_1,\\dots,f_K\\}" + }, + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "type": "text", + "content": " of the ReID target and their binary visibility scores " + }, + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\{v_{1},\\ldots ,v_{K}\\}" + }, + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "type": "text", + "content": ". Similar to previous part-based works [45, 57], the global distance " + }, + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "type": "inline_equation", + "content": "dist_{\\mathrm{global}}^{gg}" + }, + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "type": "text", + "content": " between a query image " + }, + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "type": "text", + "content": " and a gallery image " + }, + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "type": "text", + "content": " is computed as the average cosine distance " + }, + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "type": "inline_equation", + "content": "dist_{\\mathrm{cos}}" + }, + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "type": "text", + "content": " of the part-based embeddings " + }, + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 130, + 482, + 482, + 555 + ], + "type": "text", + "content": " that are visible in both samples:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 208, + 561, + 481, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 561, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 208, + 561, + 481, + 605 + ], + "type": "interline_equation", + "content": "d i s t _ {\\text {g l o b a l}} ^ {q g} = \\frac {\\sum_ {i \\in \\{1 , \\dots , K \\}} \\left(v _ {i} ^ {q} \\cdot v _ {i} ^ {g} \\cdot d i s t _ {\\cos} \\left(f _ {i} ^ {q} , f _ {i} ^ {g}\\right)\\right)}{\\sum_ {i \\in \\{1 , \\dots , K \\}} \\left(v _ {i} ^ {q} \\cdot v _ {i} ^ {g}\\right)}. \\tag {2}", + "image_path": "5c82b2dd818c39a329ceaab404fb6533ad9e3fe17fbe6cc1739e5f9e226d7d0f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 623, + 361, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 361, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 361, + 635 + ], + "type": "text", + "content": "3.7 KPR Pipeline Motivation and Intuition" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Our pipeline comprises two closely related architectural components. First, the prompt tokenizer, which encodes semantic keypoints information to disambiguate" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "V. Somers et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": "the ReID target, thereby providing prior information about body part location to drive feature extraction. Second, the Part-based Head (PBH), responsible for localizing the body parts of the target and constructing its part-based features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 481, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 481, + 235 + ], + "type": "text", + "content": "While the PBH is explicitly supervised with human parsing labels to learn human topology, there is no dedicated loss incentivizing the model to leverage the input prompt. Additionally, multi-person occlusions are rare in most ReID training sets, giving the model little incentive to rely on these prompts to localize the target person. For this reason, just adding a prompt tokenizer and training the pipeline as is produces a KPR model with weak reliance on the input prompts, and therefore low performance in multi-person occlusion scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 481, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 481, + 283 + ], + "type": "text", + "content": "To address these issues, our proposed BIPO data augmentation plays a critical role. By generating artificial multi-person occlusions during training, the model is compelled to rely more on the prompts, as they represent the sole reliable source of information to distinguish the true ReID target from the occluder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 284, + 481, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 481, + 355 + ], + "type": "text", + "content": "Consequently, incorporating BIPO during training is crucial to produce a model that is proficient at leveraging prompt information and effectively handling multi-person occlusions. Finally, designing the Part-based Head without explicit reliance on the prompt offers an additional advantage: it renders our model prompt-optional, enhancing robustness against noisy or missing prompts while still enabling effective localization of the target's body parts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 373, + 230, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 373, + 230, + 387 + ], + "spans": [ + { + "bbox": [ + 131, + 373, + 230, + 387 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 399, + 343, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 399, + 343, + 410 + ], + "spans": [ + { + "bbox": [ + 131, + 399, + 343, + 410 + ], + "type": "text", + "content": "4.1 Evaluation Benchmarks and Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": "We assess our model's performance on three types of person retrieval setups: non-occluded (i.e., holistic) on Market-1501 [53], object occlusions on Occluded-ReID [58] and Partial-ReID [54], and finally multi-person occlusions on our below introduced Occluded PoseTrack-ReID dataset. Occluded-Duke [29] was taken down and removed from our study; check out our GitHub for more information. Two standard person retrieval metrics are reported: the CMC at Rank-1 and the " + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "inline_equation", + "content": "mAP" + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": ". Performances are evaluated in a single query setting and without re-ranking [55]. Furthermore, we evaluate our model on multi-person pose tracking with the PoseTrack21 [6] dataset. We report the HOTA [25] performance metric, and relevant sub-metrics such as association accuracy (" + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "inline_equation", + "content": "AssA" + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": "), detection accuracy (" + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "inline_equation", + "content": "DetA" + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": "), and the number of identity switches (" + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "inline_equation", + "content": "IDs" + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": ") [20]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 567, + 474, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 567, + 474, + 580 + ], + "spans": [ + { + "bbox": [ + 131, + 567, + 474, + 580 + ], + "type": "text", + "content": "4.2 Proposed Occluded PoseTrack-ReID Dataset and Annotations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 586, + 481, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 481, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 481, + 647 + ], + "type": "text", + "content": "We introduce Occluded PoseTrack-ReID (or simply Occ-PTrack), a new ReID dataset we built out of the annotation available with PoseTrack21 [6], a popular video benchmark for multi-person pose tracking, that features keypoints and cross-video identity annotations. Unlike previous ReID datasets focused on street surveillance, Occ-PTrack consists of images from everyday life videos, primarily" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 284, + 91, + 445, + 102 + ], + "type": "text", + "content": "Keypoint Promptable Re-Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 364, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 364, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 364, + 666 + ], + "type": "text", + "content": "7 We use the occluded (non-cropped) images as queries." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 170, + 158, + 444, + 395 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Comparison of KPR with SOTA methods. Results in Italic are not provided in the original paper but reproduced by ourselves. The " + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "1^{st}/2^{nd}/3^{rd}" + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": " best scores are indicated with " + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "1/2/3" + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 170, + 158, + 444, + 395 + ], + "lines": [ + { + "bbox": [ + 170, + 158, + 444, + 395 + ], + "spans": [ + { + "bbox": [ + 170, + 158, + 444, + 395 + ], + "type": "table", + "html": "
DatasetsMarket-1501Occluded-reIDPartial-reIDOccluded-PoseTrack
TypeHolisticOccluded
Object Occlusions✓✓
Person Occlusions✓
MethodsR-1mAPR-1mAPR-1R-3R-1mAP
BoT94.585.958.452.3--78.869.7
PCB93.881.6----81.771.2
VGTri--81.071.085.793.73--
PVPM--66.859.578.3---
HOReID94.284.980.370.285.391.0--
ISP95.388.6------
PAT95.488.081.672.188.0392.3--
TRANS95.288.9----83.573.4
SOLIDER96.9193.91----84.476.13
SSGR96.189.378.572.9----
FED95.086.386.3179.3384.6---
BPBreed95.789.482.975.2--84.975.5
PFD95.589.783.081.52----
KPRIN w/o prompt95.688.783.378.281.786.085.375.4
KPRIN95.989.685.4279.186.090.092.3182.31
KPRSOL w/o prompt96.6393.0380.078.590.3293.7286.1375.8
KPRSOL96.6293.2284.8382.6190.7194.0190.6281.22
", + "image_path": "76c589f3888b3de884e1225ecbcfa7cac48d525cc493864cb0c3476f0f1707e9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 422, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 567 + ], + "type": "text", + "content": "from sports activities, as illustrated in Fig. 4b. Occ-PTrack is divided into a train/test that includes 1000/1411 identities with 17.898/13.412 images from 474/170 videos, which is roughly equivalent in terms of scale to other popular ReID datasets [29,53] (Tab. 5 of the supp. materials). To assess the ReID model's performance in multi-person occlusion scenarios, we select the most cluttered images of each identity in the test set as query samples, and the remaining test images as gallery samples. Cluttered images correspond to multi-person occlusions scenarios where either the front (occluding) or back (occluded) person is the ReID target, to evaluate the model ability to re-identify individuals in both scenarios. We provide further details about our proposed dataset in the supp. materials. Occ-PTrack is challenging as persons within the same video exhibit a high degree of visual resemblance since they often wear similar sports kits." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Keypoint Annotations for Standard ReID Datasets: Since conventional ReID datasets [29,53,58] do not contain keypoint annotations, we generate pseudo-labels for them with the PifPaf [19] pose estimator. These pseudo-labels include keypoint annotations in COCO [22] format for each image, as illustrated in Fig. 4a. When multiple skeletons are detected in an image, we make the assumption the one with its head closer to the top center part of the image is the intended target, and mark it with a \"is_target\" attribute. Other skeletons are used as negatives. More details are provided in the supplementary materials." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "V. Somers et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 279, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 279, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 279, + 128 + ], + "type": "text", + "content": "4.3 Implementation Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 244 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 244 + ], + "type": "text", + "content": "We evaluate our model with both the ImageNet [31] (KPRIN) and the SOLIDER [5] (KPRSOL) pre-trained Swin backbones. SOLIDER is a human-centric foundation model trained on the large-scale LUPerson [7] dataset. Following [34], the number of body parts " + }, + { + "bbox": [ + 130, + 134, + 482, + 244 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 134, + 482, + 244 + ], + "type": "text", + "content": " is set to 8 for Occ-PTrack, and 5 otherwise. The training procedure is mainly adopted from TransReID [12]. We apply our BIPO data augmentation (Sec. 3.5) with a 0.3 probability. The model is trained for 120 epochs with a batch size of 64 and a cosine annealing lr scheduler. For KPRIN/KPRSOL, images are resized to a width of 128 and a height of 256/384. The Gaussian kernel standard deviation " + }, + { + "bbox": [ + 130, + 134, + 482, + 244 + ], + "type": "inline_equation", + "content": "\\alpha_{G}" + }, + { + "bbox": [ + 130, + 134, + 482, + 244 + ], + "type": "text", + "content": " (Sec. 3.1) is set to 0.1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 258, + 380, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 258, + 380, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 380, + 270 + ], + "type": "text", + "content": "4.4 Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 277, + 480, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 480, + 313 + ], + "type": "text", + "content": "In Table " + }, + { + "bbox": [ + 130, + 277, + 480, + 313 + ], + "type": "inline_equation", + "content": "\\boxed{1}" + }, + { + "bbox": [ + 130, + 277, + 480, + 313 + ], + "type": "text", + "content": ", we present a comparative analysis of KPR, when using prompts (KPR) and not using prompts" + }, + { + "bbox": [ + 130, + 277, + 480, + 313 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 130, + 277, + 480, + 313 + ], + "type": "text", + "content": " (KPR w/o prompt), against all leading occluded ReID methods. Our method emerges as the top performer overall." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 314, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 517 + ], + "type": "text", + "content": "Occluded PoseTrack-ReID: Unlike the other three datasets, Occ-PTrack contain a substantial number of images with multi-person occlusions, making them suitable environments to assess the effectiveness of our promptable method. On this challenging occluded dataset, our proposed model KPR outperforms all previous methods. Our keypoint promptable approach offers several key advantages over the previous pose-guided methods, including: 1) the ability to handle multi-person occlusions; 2) the capability to process both positive and negative keypoints; 3) a prompt-aware appearance encoding, leading to better feature disentanglement between multiple persons; and 4) the flexibility of prompts being optional. We refer readers to Sec. 2 for a detailed discussion about these key differences and a comparison to previous methods. On our proposed Occ-PTrack, we can see that the best-performing method is the foundation model SOLIDER, but it still lies about " + }, + { + "bbox": [ + 130, + 314, + 482, + 517 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 130, + 314, + 482, + 517 + ], + "type": "text", + "content": " behind our KPR solution. Even BPBreID [34], which is specialized for occluded scenarios, demonstrates far lower performance, since it is not designed to address the Multi-Person Ambiguity. Our experiments show that integrating keypoint prompts boosts performance by at least " + }, + { + "bbox": [ + 130, + 314, + 482, + 517 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 130, + 314, + 482, + 517 + ], + "type": "text", + "content": " Rank-1 and " + }, + { + "bbox": [ + 130, + 314, + 482, + 517 + ], + "type": "inline_equation", + "content": "6.2\\%" + }, + { + "bbox": [ + 130, + 314, + 482, + 517 + ], + "type": "text", + "content": " mAP, since it helps in mitigating the negative impact of MPA." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 517, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 636 + ], + "type": "text", + "content": "Market-1501, Occluded-ReID and Partial-reID: These three datasets primarily feature single individuals in their images, which means multi-person ambiguities are rare, and therefore prompts are not exploited to their full potential. On Market-1501, our method performs on par with the state-of-the-art. However, performance levels have plateaued in recent years on this dataset. In contrast, on Occluded-ReID and Partial-reID, using a prompt noticeably improves performance by helping distinguish the individual features from occluding objects. Additionally, using prompt is beneficial when facing this cross-domain setting. These findings are consistent with earlier studies [8, 39-41] that also leveraged pose estimation to attain strong performance on this dataset." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 284, + 91, + 445, + 102 + ], + "type": "text", + "content": "Keypoint Promptable Re-Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 450, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 450, + 653 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 450, + 653 + ], + "type": "text", + "content": "8 The prompt tokenizer is completely removed at both training and inference." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 653, + 477, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 653, + 477, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 653, + 477, + 665 + ], + "type": "text", + "content": "9 Occ-ReID and Partial-ReID have no train set, so Market-1501 is used for training." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 151, + 308, + 224 + ], + "blocks": [ + { + "bbox": [ + 132, + 128, + 308, + 151 + ], + "lines": [ + { + "bbox": [ + 132, + 128, + 308, + 151 + ], + "spans": [ + { + "bbox": [ + 132, + 128, + 308, + 151 + ], + "type": "text", + "content": "Table 2: Multi-person pose tracking performance in videos on PoseTrack21 [6]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 151, + 308, + 224 + ], + "lines": [ + { + "bbox": [ + 133, + 151, + 308, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 151, + 308, + 224 + ], + "type": "table", + "html": "
MethodHOTA↑DetA↑AssA↑IDs↓
TRMOT 4246.840.955.0-
FairMOT 5153.547.461.4-
Tracktor++ 358.352v.765.4-
CorrTrack + ReID 656.951.364.2-
KPRTrack w/o prompt61.558.465.83830
KPRTrack63.0 (+1.5)59.0 (+0.6)68.3 (+2.5)3344 (-486)
", + "image_path": "c8eadf0f669b692a07308d8c80fcd1f7d29d7887adc9a31370e8aeeeb93c46c8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 322, + 114, + 481, + 206 + ], + "blocks": [ + { + "bbox": [ + 322, + 114, + 481, + 206 + ], + "lines": [ + { + "bbox": [ + 322, + 114, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 322, + 114, + 481, + 206 + ], + "type": "image", + "image_path": "4ad084a3f993e83a4eb377bd7d17afb441dacc4a0d8472fe64f3fd5bdab9578e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 215, + 471, + 227 + ], + "lines": [ + { + "bbox": [ + 333, + 215, + 471, + 227 + ], + "spans": [ + { + "bbox": [ + 333, + 215, + 471, + 227 + ], + "type": "text", + "content": "Fig. 5: Prompt optionality study" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 279, + 353, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 353, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 353, + 292 + ], + "type": "text", + "content": "4.5 Multi-Person Pose Tracking in Videos" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 299, + 482, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 482, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 482, + 526 + ], + "type": "text", + "content": "We introduce a simple yet robust pose tracker, referred to as KPRTrack, built upon our promptable ReID method. KPRTrack combines the YOLOX object detector with the HRNet pose estimator. For each detected person, it extracts part-based ReID features utilizing their keypoints as positive prompts and keypoints from other individuals within the detected person bbox as negative prompts. Detections are matched in an online fashion with the Hungarian algorithm based on the distance function introduced in Eq. (2). Only tracklet-detection pairs with a distance below a threshold of 0.2 are considered for matching. Therefore, our proposed tracker relies solely on appearance information, meaning spatio-temporal cues are omitted. Pose tracking results are reported in Tab. Despite relying on ReID only, KPRTrack achieves state-of-the-art performance, demonstrating that an occlusion-robust ReID model has strong tracking capabilities. Finally, our two last experiments in Tab. reveal that employing keypoint prompts significantly enhances tracking performance. This improvement can be attributed to the enhanced communication between the detector and the ReID via the keypoint prompts. This leads to improved association accuracy and reduced identity switches, especially when multiple targets intersect paths (i.e., multi-person occlusion scenarios). Further information regarding our baseline tracker is provided in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 544, + 246, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 246, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 246, + 555 + ], + "type": "text", + "content": "4.6 Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 563, + 481, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 481, + 576 + ], + "type": "text", + "content": "In this section, we analyze the performance impact of our architectural choices." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "Components of KPR As demonstrated in Table 3, all our proposed modules introduced in Sec. 3 play a crucial role in improving the overall ReID performance. Exp. 2 and Exp. 8 demonstrate the importance of the Part-based Head to compare only mutually visible body parts when facing occluded persons, compared to using a single global feature (as in Exp. 1 and 7). PBH degrades performance when applied alone (Exp. 2 vs Exp. 1), but is beneficial when jointly" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "V. Somers et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 196, + 190, + 418, + 313 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 182 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 182 + ], + "type": "text", + "content": "Table 3: Ablation study of our proposed KPR architecture. PBH stands for \"Part-based Head\", meaning we use multiple part-based embeddings with visibility scores instead of a single global embedding (Sec. 3.3). MSF stands for the \"Multi-Scale Features\" module added to Swin (Sec. 3.2). Prompt " + }, + { + "bbox": [ + 130, + 114, + 482, + 182 + ], + "type": "inline_equation", + "content": "\\oplus/\\ominus" + }, + { + "bbox": [ + 130, + 114, + 482, + 182 + ], + "type": "text", + "content": " refers to positive/negative prompts (Sec. 3.1). BIPO stands for the \"Batch-wise Inter-Person Occlusion\" data augmentation (Sec. 3.5). SOL stands for the pre-trained weights from SOLIDER [5] (Sec. 4.3)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 196, + 190, + 418, + 313 + ], + "lines": [ + { + "bbox": [ + 196, + 190, + 418, + 313 + ], + "spans": [ + { + "bbox": [ + 196, + 190, + 418, + 313 + ], + "type": "table", + "html": "
IdMain components of KPRDataset
PBHMSFPromptBIPOSOLOccluded-PoseTrack
⊕∅R-1mAP
184.874.7
2✓83.3 (-1.5)72.2 (-2.5)
3✓✓84.8 (+0.0)76.1 (+1.4)
4✓✓✓88.8 (+4.0)80.4 (+5.7)
5✓✓✓✓90.0 (+5.2)80.7 (+6.0)
6✓✓✓86.1 (+1.3)76.7 (+2.0)
7✓✓✓✓80.4 (+5.1)89.9 (+5.7)
8✓✓✓✓✓92.3 (+6.5)82.3 (+7.6)
9✓✓✓✓✓✓90.6 (+5.8)81.2 (+6.5)
", + "image_path": "e436790e18885671bd9fb677bbfb4809b0dd55a13260be73cc20d314f56232e1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 335, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 529 + ], + "type": "text", + "content": "enabled with other modules (Exp. 8 vs Exp 7). Exp. 3 is consistent with claims from previous works [26,57] that increasing the feature map resolution is crucial to achieve fine-grained re-identification and better overall performance. Exp. 4 and 5 illustrate the benefits of positive and negative prompts to disambiguate the intended ReID target from other persons. Exp. 6 demonstrates the importance of our BIPO augmentation to generate artificial occlusions at training, since most occluded ReID benchmarks feature mainly occluded samples in their test set and not in their training set. Exp. 8, in comparison with Exp. 5 and 6, demonstrates how generating these artificial multi-person occlusions at training is crucial to teach the model to rely on the input prompt to identify the intended ReID target among all candidates. The usage of the SOLIDER foundation model as pre-trained weights is illustrated in Exp. 9. Unfortunately, SOLIDER does not enhance performance on the Occ-PTrack dataset, likely because of a domain gap: SOLIDER focuses on street surveillance, while Occ-PTrack predominantly features sports images from handheld cameras. However, as demonstrated in Tab. [1], SOLIDER does improve performance for all other datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "A Prompt-Optional Method In Figure 5 we analyze the test-time robustness of KPR against noisy or missing prompts on Occ-PTrack, by increasingly removing a higher percentage of keypoints from the input prompt. The red curve represents the performance of our proposed promptable KPR model, while the blue constant line illustrates the performance of the non-promptable version, which is unaffected by the proportion of keypoints removed from the input prompt. As illustrated, KPR significantly outperforms the non-promptable baseline with a full prompt and remains competitive with few or no keypoints, highlighting its prompt optional capability. We provide an additional study of the performance w.r.t. the amount of multi-person occlusion in the supplementary materials." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 284, + 91, + 445, + 102 + ], + "type": "text", + "content": "Keypoint Promptable Re-Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 517, + 292 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 517, + 292 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 517, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 517, + 292 + ], + "type": "image", + "image_path": "df82a5220764d57d243a469ecf4ca035a7c064c43370a24b3e1e50a6af70e90f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 292, + 481, + 307 + ], + "lines": [ + { + "bbox": [ + 132, + 292, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 292, + 481, + 307 + ], + "type": "text", + "content": "(b) Part-attention maps (c) Pose tracking results: all track IDs, depicted by unique skeleton for various input prompts. colors, are maintained after a long occlusion by the front cyclist." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 316, + 482, + 352 + ], + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 352 + ], + "type": "text", + "content": "Fig.6: Qualitative results on multi-person pose tracking and person retrieval. Red/pastel dots indicate negative/positive prompts. Colored heatmaps illustrates the part-attention maps of KPR described in Sec. 3.3 with one color per body part." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 371, + 485, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 485, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 485, + 420 + ], + "type": "text", + "content": "Qualitative Assessment Figure demonstrates the impact of keypoint prompting on the part-attention maps, on person retrieval, and on pose tracking. As illustrated, KPR effectively leverages the input prompts to localize the intended target, and is therefore robust to occlusions and multi-person ambiguities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 437, + 220, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 220, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 220, + 449 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 460, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 567 + ], + "type": "text", + "content": "In this work we propose KPR, the first promptable ReID model, designed to address occluded scenarios and multi-person ambiguities arising from bounding-box-based inputs. Our model offers practical flexibility by being prompt-optional, achieving state-of-the-art performance without prompts, and outperforming this baseline with prompts. Furthermore, our model outperforms the state-of-the-art on three popular ReID benchmarks and on our novel Occ-PTrack dataset. We finally demonstrate KPR's potential for pose tracking in videos. Our codebase, keypoint labels and proposed dataset will be released to encourage further research on this new promptable ReID paradigm." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 568, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 481, + 593 + ], + "type": "text", + "content": "Acknowledgment: This work has been funded by Sportradar, by the Walloon region project ReconNAssance, and by the FNRS." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 609, + 198, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 198, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 198, + 621 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 136, + 632, + 481, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 632, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 136, + 632, + 481, + 667 + ], + "type": "text", + "content": "1. Adaimi, G., Kreiss, S., Alahi, A.: Deep visual re-identification with confidence. Transportation Research Part C: Emerging Technologies 126, 103067 (May 2021). https://doi.org/10.1016/j.trc.2021.103067" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "V. Somers et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "text", + "content": "2. Alahi, A., Ramanathan, V., Fei-Fei, L.: Tracking millions of humans in crowded spaces. Group and Crowd Behavior for Computer Vision pp. 115-135 (2017). https://doi.org/10.1016/b978-0-12-809276-7.00007-2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 150, + 480, + 183 + ], + "type": "text", + "content": "3. Bergmann, P., Meinhardt, T., Leal-Taixe, L.: Tracking without bells and whistles. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 941-951. Inst. Electr. Electron. Eng. (IEEE) (Oct 2019). https://doi.org/10.1109/iccv.2019.00103" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 183, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 183, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 138, + 183, + 480, + 236 + ], + "type": "text", + "content": "4. Chen, P., Liu, W., Dai, P., Liu, J., Ye, Q., Xu, M., Chen, Q., Ji, R.: Occlude them all: Occlusion-aware attention network for occluded person re-ID. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 11813-11822. Inst. Electr. Electron. Eng. (IEEE), Montreal, QC, Canada (Oct 2021). https://doi.org/10.1109/icc48922.2021.01162" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 237, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 237, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 237, + 480, + 293 + ], + "type": "text", + "content": "5. Chen, W., Xu, X., Jia, J., Luo, H., Wang, Y., Wang, F., Jin, R., Sun, X.: Beyond appearance: A semantic controllable self-supervised learning framework for human-centric visual tasks. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 15050-15061. Inst. Electr. Electron. Eng. (IEEE), Vancouver, Can. (Jun 2023). https://doi.org/10.1109/cvpr52729.2023.01445" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 293, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 346 + ], + "type": "text", + "content": "6. Doering, A., Chen, D., Zhang, S., Schiele, B., Gall, J.: PoseTrack21: A dataset for person search, multi-object tracking and multi-person pose tracking. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 20931-20940. Inst. Electr. Electron. Eng. (IEEE), New Orleans, LA, USA (Jun 2022). https://doi.org/10.1109/cvpr52688.2022.02029" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 347, + 480, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 347, + 480, + 401 + ], + "spans": [ + { + "bbox": [ + 138, + 347, + 480, + 401 + ], + "type": "text", + "content": "7. Fu, D., Chen, D., Bao, J., Yang, H., Yuan, L., Zhang, L., Li, H., Chen, D.: Unsupervised pre-training for person re-identification. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 14745-14754. Inst. Electr. Electron. Eng. (IEEE), Nashville, TN, USA (Jun 2021). https://doi.org/10.1109/cvpr46437.2021.01451" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 402, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 402, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 402, + 480, + 445 + ], + "type": "text", + "content": "8. Gao, S., Wang, J., Lu, H., Liu, Z.: Pose-guided visible part matching for occluded person ReID. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 11741-11749. Inst. Electr. Electron. Eng. (IEEE), Seattle, WA, USA (Jun 2020). https://doi.org/10.1109/cvpr42600.2020.01176" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 446, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 446, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 446, + 480, + 468 + ], + "type": "text", + "content": "9. Ge, Z., Liu, S., Wang, F., Li, Z., Sun, J.: YOLOX: Exceeding YOLO series in 2021. arXiv abs/2107.08430 (2021). https://doi.org/10.48550/arXiv.2107.08430" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 468, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 468, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 134, + 468, + 480, + 522 + ], + "type": "text", + "content": "10. Ghiasi, G., Cui, Y., Srinivas, A., Qian, R., Lin, T.Y., Cubuk, E.D., Le, Q.V., Zoph, B.: Simple copy-paste is a strong data augmentation method for instance segmentation. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 2917-2927. Inst. Electr. Electron. Eng. (IEEE), Nashville, TN, USA (Jun 2021). https://doi.org/10.1109/cvpr46437.2021.00294" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 522, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 522, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 522, + 480, + 665 + ], + "type": "text", + "content": "11. Giancola, S., Cioppa, A., Deliège, A., Magera, F., Somers, V., Kang, L., Zhou, X., Barnich, O., De Vleeschouwer, C., Alahi, A., Ghanem, B., Van Droogenbroeck, M., Darwish, A., Maglo, A., Clapés, A., Luyts, A., Boiarov, A., Xarles, A., Orcesi, A., Shah, A., Fan, B., Comandur, B., Chen, C., Zhang, C., Zhao, C., Lin, C., Chan, C.Y., Hui, C.C., Li, D., Yang, F., Liang, F., Da, F., Yan, F., Yu, F., Wang, G., Chan, H.A., Zhu, H., Kan, H., Chu, J., Hu, J., Gu, J., Chen, J., Soares, J.V.B., Theiner, J., De Corte, J., Brito, J.H., Zhang, J., Li, J., Liang, J., Shen, L., Ma, L., Chen, L., Santos Marques, M., Azatov, M., Kasatkin, N., Wang, N., Jia, Q., Pham, Q.C., Ewerth, R., Song, R., Li, R., Gade, R., Debien, R., Zhang, R., Lee, S., Escalera, S., Jiang, S., Odashima, S., Chen, S., Masui, S., Ding, S., Chan, S.w., Chen, S., El-Shabrawy, T., He, T., Moeslund, T.B., Siu, W.C., Zhang, W., Li, W., Wang, X., Tan, X., Li, X., Wei, X., Ye, X., Liu, X., Wang, X., Guo, Y., Zhao, Y., Yu, Y., Li, Y., He, Y., Zhong, Y., Guo, Z., Li, Z.: SoccerNet 2022" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 284, + 91, + 445, + 102 + ], + "type": "text", + "content": "Keypoint Promptable Re-Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 664 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 149, + 116, + 481, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 481, + 148 + ], + "type": "text", + "content": "challenges results. In: Proceedings of the 5th International ACM Workshop on Multimedia Content Analysis in Sports. ACM (Oct 2022). https://doi.org/10.1145/3552437.3558545" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 192 + ], + "type": "text", + "content": "12. He, S., Luo, H., Wang, P., Wang, F., Li, H., Jiang, W.: TransReID: Transformer-based object re-identification. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 14993–15002. Inst. Electr. Electron. Eng. (IEEE), Montreal, QC, Canada (Oct 2021). https://doi.org/10.1109/iccv48922.2021.01474" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 481, + 224 + ], + "type": "text", + "content": "13. Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re-identification. arXiv abs/1703.07737 (2017). https://doi.org/10.48550/arXiv.1703.07737" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 224, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 224, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 224, + 481, + 267 + ], + "type": "text", + "content": "14. Istasse, M., Somers, V., Elancheliyan, P., De, J., Zambrano, D.: DeepSportradar-v2: A multi-sport computer vision dataset for sport understandings. In: Proceedings of the 6th International Workshop on Multimedia Content Analysis in Sports. ACM (Oct 2023). https://doi.org/10.1145/3606038.3616160" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 267, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 267, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 133, + 267, + 481, + 311 + ], + "type": "text", + "content": "15. Jia, M., Cheng, X., Zhai, Y., Lu, S., Ma, S., Tian, Y., Zhang, J.: Matching on sets: Conquer occluded person re-identification without alignment. In: AAAI Conf. Artif. Intell. vol. 35, pp. 1673-1681. Association for the Advancement of Artificial Intelligence (AAAI) (May 2021). https://doi.org/10.1609/aaai.v35i2.16260" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 311, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 311, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 133, + 311, + 481, + 331 + ], + "type": "text", + "content": "16. Joos, V., Somers, V., Standaert, B.: TrackLab. https://github.com/TrackingLaboratory/tracklab (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 332, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 332, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 133, + 332, + 481, + 364 + ], + "type": "text", + "content": "17. Kiran, M., Praveen, R.G., Nguyen-Meidine, L.T., Belharbi, S., Blais-Morin, L.A., Granger, E.: Holistic guidance for occluded person re-identification. arXiv abs/2104.06524 (2021). https://doi.org/10.48550/arXiv.2104.06524" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 364, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 364, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 133, + 364, + 481, + 396 + ], + "type": "text", + "content": "18. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv abs/2304.02643 (2023). https://doi.org/10.48550/arXiv.2304.02643" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 396, + 481, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 396, + 481, + 438 + ], + "spans": [ + { + "bbox": [ + 133, + 396, + 481, + 438 + ], + "type": "text", + "content": "19. Kreiss, S., Bertoni, L., Alahi, A.: PifPaf: Composite fields for human pose estimation. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 11969-11978. Inst. Electr. Electron. Eng. (IEEE), Long Beach, CA, USA (Jun 2019). https://doi.org/10.1109/cvpr.2019.01225" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 439, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 439, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 133, + 439, + 481, + 471 + ], + "type": "text", + "content": "20. Leal-Taixe, L., Milan, A., Reid, I., Roth, S., Schindler, K.: MOTChallenge 2015: Towards a benchmark for multi-target tracking. arXiv abs/1504.01942 (2015). https://doi.org/10.48550/arXiv.1504.01942" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 471, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 471, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 133, + 471, + 481, + 525 + ], + "type": "text", + "content": "21. Li, Y., He, J., Zhang, T., Liu, X., Zhang, Y., Wu, F.: Diverse part discovery: Occluded person re-identification with part-aware transformer. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 2897-2906. Inst. Electr. Electron. Eng. (IEEE), Nashville, TN, USA (Jun 2021). https://doi.org/10.1109/cvpr46437.2021.00292" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 525, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 525, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 133, + 525, + 481, + 568 + ], + "type": "text", + "content": "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: Eur. Conf. Comput. Vis. (ECCV). Lect. Notes Comput. Sci., vol. 8693, pp. 740-755. Springer Int. Publ. (2014). https://doi.org/10.1007/978-3-319-10602-1_48" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 568, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 568, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 133, + 568, + 481, + 610 + ], + "type": "text", + "content": "23. Liu, Q., Xu, Z., Bertasius, G., Niethammer, M.: SimpleClick: Interactive image segmentation with simple vision transformers. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 22233-22243. Inst. Electr. Electron. Eng. (IEEE), Paris, Fr. (Oct 2023). https://doi.org/10.1109/iccv51070.2023.02037" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 611, + 481, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 481, + 664 + ], + "type": "text", + "content": "24. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 9992-10002. Inst. Electr. Electron. Eng. (IEEE), Montreal, QC, Canada (Oct 2021). https://doi.org/10.1109/iccv48922.2021.00986" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "V. Somers et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 667 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 159 + ], + "type": "text", + "content": "25. Luiten, J., Osuep, A., Dendorfer, P., Torr, P., Geiger, A., Leal-Taixe, L., Leibe, B.: HOTA: A higher order metric for evaluating multi-object tracking. Int. J. Comput. Vis. 129(2), 548-578 (Oct 2020). https://doi.org/10.1007/s11263-020-01375-2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 481, + 205 + ], + "type": "text", + "content": "26. Luo, H., Gu, Y., Liao, X., Lai, S., Jiang, W.: Bag of tricks and a strong baseline for deep person re-identification. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Work. (CVPRW). pp. 1487-1495. Inst. Electr. Electron. Eng. (IEEE), Long Beach, CA, USA (Jun 2019). https://doi.org/10.1109/cvprw.2019.00190" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "type": "text", + "content": "27. Ma, Z., Zhao, Y., Li, J.: Pose-guided inter- and intra-part relational transformer for occluded person re-identification. In: Proceedings of the 29th ACM International Conference on Multimedia. ACM (Oct 2021). https://doi.org/10.1145/3474085.3475283" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 249, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 303 + ], + "type": "text", + "content": "28. Mansourian, A.M., Somers, V., De Vleeschouwer, C., Kasaei, S.: Multi-task learning for joint re-identification, team affiliation, and role classification for sports visual tracking. In: Proceedings of the 6th International Workshop on Multimedia Content Analysis in Sports. ACM (Oct 2023). https://doi.org/10.1145/3606038.3616172" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "type": "text", + "content": "29. Miao, J., Wu, Y., Liu, P., Ding, Y., Yang, Y.: Pose-guided feature alignment for occluded person re-identification. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 542-551. Inst. Electr. Electron. Eng. (IEEE) (Oct 2019). https://doi.org/10.1109/iccv.2019.00063" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "30. Peng, Y., Hou, S., Cao, C., Liu, X., Huang, Y., He, Z.: Deep learning-based occluded person re-identification: A survey. arXiv abs/2207.14452 (2022). https://doi.org/10.48550/arXiv.2207.14452" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 380, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 424 + ], + "type": "text", + "content": "31. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet large scale visual recognition challenge. Int. J. Comput. Vis. 115(3), 211-252 (Apr 2015). https://doi.org/10.1007/s11263-015-0816-y" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 424, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 468 + ], + "type": "text", + "content": "32. Seidenschwarz, J., Brasó, G., Serrano, V.C., Elezi, I., Leal-Taixe, L.: Simple cues lead to a strong multi-object tracker. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 13813-13823. Inst. Electr. Electron. Eng. (IEEE), Vancouver, Can. (Jun 2023). https://doi.org/10.1109/cvpr52729.2023.01327" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 468, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 534 + ], + "type": "text", + "content": "33. Somers, V., Joos, V., Giancola, S., Cioppa, A., Ghasemzadeh, S.A., Magera, F., Standaert, B., Mansourian, A.M., Zhou, X., Kasaei, S., Ghanem, B., Alahi, A., Van Droogenbroeck, M., De Vleeschouwer, C.: SoccerNet game state reconstruction: End-to-end athlete tracking and identification on a minimap. In: IEEE Int. Conf. Comput. Vis. Pattern Recognit. Work. (CVPRW), CVsports. Seattle, WA, USA (Jun 2024). https://doi.org/10.48550/ARXIV.2404.11335" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 534, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 578 + ], + "type": "text", + "content": "34. Somers, V., Vleeschouwer, C.D., Alahi, A.: Body part-based representation learning for occluded person re-identification. In: IEEE/CVF Winter Conf. Appl. Comput. Vis. (WACV). pp. 1613–1623. Inst. Electr. Electron. Eng. (IEEE), Waikoloa, HI, USA (Jan 2023). https://doi.org/10.1109/wacv56688.2023.00166" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 578, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 622 + ], + "type": "text", + "content": "35. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 5686-5696. Inst. Electr. Electron. Eng. (IEEE), Long Beach, CA, USA (Jun 2019). https://doi.org/10.1109/cvpr.2019.00584" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 622, + 481, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 667 + ], + "type": "text", + "content": "36. Sun, Y., Zheng, L., Yang, Y., Tian, Q., Wang, S.: Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline). In: Eur. Conf. Comput. Vis. (ECCV). Lect. Notes Comput. Sci., vol. 11208, pp. 501-518. Springer Int. Publ. (2018). https://doi.org/10.1007/978-3-030-01225-0_30" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 284, + 91, + 445, + 102 + ], + "type": "text", + "content": "Keypoint Promptable Re-Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "text", + "content": "37. Tan, H., Liu, X., Yin, B., Li, X.: MHSA-net: Multihead self-attention network for occluded person re-identification. IEEE Trans. Neural Networks Learn. Syst. 34(11), 8210-8224 (Nov 2023). https://doi.org/10.1109/tnnls.2022.3144163" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 204 + ], + "type": "text", + "content": "38. Van Zandycke, G., Somers, V., Istasse, M., Don, C.D., Zambrano, D.: DeepSportradar-v1: Computer vision dataset for sports understanding with high quality annotations. In: Proceedings of the 5th International ACM Workshop on Multimedia Content Analysis in Sports. ACM (Oct 2022). https://doi.org/10.1145/3552437.3555699" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 259 + ], + "type": "text", + "content": "39. Wang, G., Yang, S., Liu, H., Wang, Z., Yang, Y., Wang, S., Yu, G., Zhou, E., Sun, J.: High-order information matters: Learning relation and topology for occluded person re-identification. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 6448-6457. Inst. Electr. Electron. Eng. (IEEE), Seattle, WA, USA (Jun 2020). https://doi.org/10.1109/cvpr42600.2020.00648" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 259, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 304 + ], + "type": "text", + "content": "40. Wang, T., Liu, H., Song, P., Guo, T., Shi, W.: Pose-guided feature disentangling for occluded person re-identification based on transformer. In: AAAI Conf. Artif. Intell. vol. 36, pp. 2540-2549. Association for the Advancement of Artificial Intelligence (AAAI) (Jun 2022). https://doi.org/10.1609/aaai.v36i3.20155" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 304, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 357 + ], + "type": "text", + "content": "41. Wang, Z., Zhu, F., Tang, S., Zhao, R., He, L., Song, J.: Feature erasing and diffusion network for occluded person re-identification. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 4744-4753. Inst. Electr. Electron. Eng. (IEEE), New Orleans, LA, USA (Jun 2022). https://doi.org/10.1109/cvpr52688.2022.00471" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 358, + 481, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 401 + ], + "type": "text", + "content": "42. Wang, Z., Zheng, L., Liu, Y., Li, Y., Wang, S.: Towards real-time multi-object tracking. In: Eur. Conf. Comput. Vis. (ECCV). Lect. Notes Comput. Sci., vol. 12356, pp. 107-122. Springer Int. Publ. (2020). https://doi.org/10.1007/978-3-030-58621-8_7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 402, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 446 + ], + "type": "text", + "content": "43. Wei, L., Zhang, S., Gao, W., Tian, Q.: Person transfer GAN to bridge domain gap for person re-identification. In: IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 79–88. Inst. Electr. Electron. Eng. (IEEE), Salt Lake City, UT, USA (Jun 2018). https://doi.org/10.1109/cvpr.2018.00016" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 446, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 490 + ], + "type": "text", + "content": "44. Yan, C., Pang, G., Jiao, J., Bai, X., Feng, X., Shen, C.: Occluded person re-identification with single-scale global representations. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 11855-11864. Inst. Electr. Electron. Eng. (IEEE), Montreal, QC, Canada (Oct 2021). https://doi.org/10.1109/iccv48922.2021.01166" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 491, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 544 + ], + "type": "text", + "content": "45. Yang, J., Zhang, J., Yu, F., Jiang, X., Zhang, M., Sun, X., Chen, Y., Zheng, W.S.: Learning to know where to see: A visibility-aware approach for occluded person re-identification. In: IEEE/CVF Int. Conf. Comput. Vis. (ICCV). pp. 11865-11874. Inst. Electr. Electron. Eng. (IEEE), Montreal, QC, Canada (Oct 2021). https://doi.org/10.1109/iccv48922.2021.01167" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 544, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 578 + ], + "type": "text", + "content": "46. Yang, Q., Wang, P., Fang, Z., Lu, Q.: Focus on the visible regions: Semantic-guided alignment model for occluded person re-identification. Sensors 20(16), 1-15 (Aug 2020). https://doi.org/10.3390/s20164431" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "text", + "content": "47. Ye, M., Shen, J., Lin, G., Xiang, T., Shao, L., Hoi, S.C.H.: Deep learning for person re-identification: A survey and outlook. IEEE Trans. Pattern Anal. Mach. Intell. 44(6), 2872–2893 (Jun 2022). https://doi.org/10.1109/tpami.2021.3054775" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "48. Zang, X., Li, G., Gao, W., Shu, X.: Learning to disentangle scenes for person re-identification. Image Vis. Comput. 116, 104330 (Dec 2021). https://doi.org/10.1016/j.imavis.2021.104330" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "49. Zhang, G., Zhang, Y., Zhang, T., Li, B., Pu, S.: PHA: Patch-wise high-frequency augmentation for transformer-based person re-identification. In: IEEE/CVF Conf." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "V. Somers et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 523 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 147, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 482, + 149 + ], + "type": "text", + "content": "Comput. Vis. Pattern Recognit. (CVPR). pp. 14133-14142. Inst. Electr. Electron. Eng. (IEEE), Vancouver, Can. (Jun 2023). https://doi.org/10.1109/cvpr52729.2023.01358" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 205 + ], + "type": "text", + "content": "50. Zhang, Y., Sun, P., Jiang, Y., Yu, D., Weng, F., Yuan, Z., Luo, P., Liu, W., Wang, X.: ByteTrack: Multi-object tracking by associating every detection box. In: Eur. Conf. Comput. Vis. (ECCV). Lect. Notes Comput. Sci., vol. 13682, pp. 1-21. Springer Nat. Switz. (2022). https://doi.org/10.1007/978-3-031-20047-2_1, https://doi.org/10.1007/978-3-031-20047-2_1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "type": "text", + "content": "51. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: FairMOT: On the fairness of detection and re-identification in multiple object tracking. Int. J. Comput. Vis. 129(11), 3069–3087 (Sept 2021). https://doi.org/10.1007/s11263-021-01513-4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "text", + "content": "52. Zheng, K., Lan, C., Zeng, W., Liu, J., Zhang, Z., Zha, Z.J.: Pose-guided feature learning with knowledge distillation for occluded person re-identification. arXiv abs/2108.00139 (2021). https://doi.org/10.48550/arXiv.2108.00139" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 282, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 326 + ], + "type": "text", + "content": "53. Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., Tian, Q.: Scalable person re-identification: A benchmark. In: IEEE Int. Conf. Comput. Vis. (ICCV). pp. 1116-1124. Inst. Electr. Electron. Eng. (IEEE), Santiago, Chile (Dec 2015). https://doi.org/10.1109/iccv.2015.133" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "text", + "content": "54. Zheng, W.S., Li, X., Xiang, T., Liao, S., Lai, J., Gong, S.: Partial person re-identification. In: IEEE Int. Conf. Comput. Vis. (ICCV). pp. 4678-4686. Inst. Electr. Electron. Eng. (IEEE), Santiago, Chile (Dec 2015). https://doi.org/10.1109/iccv.2015.531" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 369, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 369, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 133, + 369, + 481, + 413 + ], + "type": "text", + "content": "55. Zhong, Z., Zheng, L., Cao, D., Li, S.: Re-ranking person re-identification with k-reciprocal encoding. In: IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR). pp. 3652-3661. Inst. Electr. Electron. Eng. (IEEE), Honolulu, HI, USA (Jul 2017). https://doi.org/10.1109/cvpr.2017.389" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 413, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 413, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 133, + 413, + 481, + 456 + ], + "type": "text", + "content": "56. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. In: AAAI Conf. Artif. Intell. vol. 34, pp. 13001-13008. Association for the Advancement of Artificial Intelligence (AAAI) (Apr 2020). https://doi.org/10.1609/aaai.v34i07.7000" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 457, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 457, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 133, + 457, + 481, + 501 + ], + "type": "text", + "content": "57. Zhu, K., Guo, H., Liu, Z., Tang, M., Wang, J.: Identity-guided human semantic parsing for person re-identification. In: Eur. Conf. Comput. Vis. (ECCV). Lect. Notes Comput. Sci., vol. 12348, pp. 346-363. Springer Int. Publ. (2020). https://doi.org/10.1007/978-3-030-58580-8_21" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 501, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 523 + ], + "type": "text", + "content": "58. Zhuo, J., Chen, Z., Lai, J., Wang, G.: Occluded person re-identification. arXiv abs/1804.02792 (2018). https://doi.org/10.48550/arXiv.1804.02792" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 285, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 285, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 285, + 91, + 445, + 102 + ], + "type": "text", + "content": "Keypoint Promptable Re-Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/KeypointDETR_ An End-to-End 3D Keypoint Detector/d05c910b-e5c7-4df0-b517-a9b20e7a9b64_content_list.json b/2024/KeypointDETR_ An End-to-End 3D Keypoint Detector/d05c910b-e5c7-4df0-b517-a9b20e7a9b64_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9397450fb3574c18bdf6323fb1cd1b5b170b6ab8 --- /dev/null +++ b/2024/KeypointDETR_ An End-to-End 3D Keypoint Detector/d05c910b-e5c7-4df0-b517-a9b20e7a9b64_content_list.json @@ -0,0 +1,1556 @@ +[ + { + "type": "text", + "text": "KeypointDETR: An End-to-End 3D Keypoint Detector", + "text_level": 1, + "bbox": [ + 238, + 140, + 763, + 183 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hairong Jin $^{1,2\\star}$ , Yuefan Shen $^{1}$ , Jianwen Lou $^{3}$ , Kun Zhou $^{1}$ , and Youyi Zheng $^{1(\\boxtimes)}$", + "bbox": [ + 250, + 210, + 754, + 244 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ State Key Lab of CAD&CG, Zhejiang University, Hangzhou, China hairong_jin, jhonve, kunzhou, youyizheng@zju.edu.cn", + "2 Chohotech Co.,Ltd., Hangzhou, China", + "3 School of Software Technology, Zhejiang University, Hangzhou, China jianwen.lou@zju.edu.cn" + ], + "bbox": [ + 261, + 253, + 738, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. 3D keypoint detection plays a pivotal role in 3D shape analysis. The majority of prevalent methods depend on producing a shared heatmap. This approach necessitates subsequent post-processing techniques such as clustering or non-maximum suppression (NMS) to pinpoint keypoints within high-confidence regions, resulting in performance inefficiencies. To address this issue, we introduce KeypointDETR, an end-to-end 3D keypoint detection framework. KeypointDETR is predominantly trained with a bipartite matching loss, which compels the network to forecast sets of heatmaps and probabilities for potential keypoints. Each heatmap highlights one keypoint's location, and the associated probability indicates not only the presence of that specific keypoint but also its semantic consistency. Together with the bipartite matching loss, we utilize a transformer-based network architecture, which incorporates both point-wise and query-wise self-attention within the encoder and decoder, respectively. The point-wise encoder leverages the self-attention mechanism on a dynamic graph derived from the local feature space of each point, resulting in the generation of heatmap features. As a key part of our framework, the query-wise decoder facilitates inter-query information exchange. It captures the underlying connections among keypoints' heatmaps, positions, and semantic attributes via the cross-attention mechanism, enabling the prediction of heatmaps and probabilities. Extensive experiments conducted on the KeypointNet dataset reveal that KeypointDETR outperforms competitive baselines, demonstrating superior performance in keypoint saliency and correspondence estimation tasks. (The code will be released at github.com/bibi547/KeypointDETR)", + "bbox": [ + 261, + 349, + 738, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D keypoint $\\cdot$ Point cloud $\\cdot$ Heatmap", + "bbox": [ + 261, + 709, + 599, + 723 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 744, + 375, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The automatic extraction of 3D keypoints presents a significant challenge in the realms of computer vision and computer graphics. These keypoints are instrumental in various geometric processing tasks, such as object detection [20], shape", + "bbox": [ + 215, + 772, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* This work was done during the author's internship at Chohotech Co.,Ltd..", + "bbox": [ + 218, + 825, + 725, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/628c584447c38ac4ab09e2ddbe83460aaffe7bdc560e6faa094a6cf7cef82eb8.jpg", + "image_caption": [ + "Fig. 1: Illustration of the differences between the existing methods and KeypointDETR for keypoint saliency and correspondence estimation. Given a 3D model, Keypoint-DETR can predict its multi-heatmaps and probabilities. Heatmaps provide accurate inference of keypoints' positions. Probabilities indicate whether the specific keypoints exist in the 3D model." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 769, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "segmentation [13, 21], shape matching [39, 44], point registration [1], and point cloud completion [36], etc.", + "bbox": [ + 212, + 429, + 782, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent breakthroughs in deep learning, especially in 3D shape analysis and geometric feature extraction [5,11,27,28,40,43], have catalyzed the emergence of a multitude of learning-based approaches for 3D keypoint detection. A straightforward technique [45,47] is point classification on the surface of a 3D shape, but this method grapples with significant sample imbalances. To mitigate this, many existing keypoint detection strategies [32,41,42] utilize Gaussian kernel functions to create a ground truth heatmap. This shared heatmap is generated based on the distances between points and keypoints, guiding the extraction of keypoints from areas identified as salient. However, these techniques typically require post-processing steps such as clustering [29] or non-maximum suppression [22] to sift out high-confidence points and finalize the keypoint extraction. These processes often involve computing distance matrices, which can considerably slow down inference speeds. Additionally, the selection of parameters can profoundly affect the accuracy of keypoint extraction. Many landmark detection methods for 2D or 3D images [4] prefer using multi-heatmaps for keypoints, as opposed to relying on a single shared heatmap. This approach assigns each keypoint its unique heatmap, facilitating the extraction of the point with the highest value from each map. Although their strategy eliminates the need for post-processing, it can limit the algorithm's generalizability. One significant challenge is ensuring a consistent number of keypoints across all 3D models within a category.", + "bbox": [ + 212, + 460, + 787, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle these challenges, we introduce an innovative end-to-end 3D keypoint detection model named KeypointDETR. This approach not only eliminates the need for any post-processing steps but also exhibits remarkable generalization capabilities. As illustrated in Fig. 1, our framework takes point coordinates as input and predicts multi-heatmaps along with their associated probabilities,", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H.Jin et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "providing a consistent paradigm for keypoint saliency and correspondence estimation. For keypoint saliency estimation, we can derive binary $yes / no$ labels from the probabilities, while for keypoint correspondence estimation, semantic labels can be inferred. Our pipeline utilizes a loss function based on the Hungarian algorithm [15], which facilitates the optimal bipartite matching between the predictions and the ground truth heatmaps. Consistent with this strategy, we develop a transformer-based encoder-decoder network architecture. The encoder employs dynamic graph transformer modules to extract geometric features from 3D models, specifically focusing on the generation of heatmap features. We regard the learnable heatmap features as intrinsic geometric features, eliminating the need for direct supervision. The dynamic graph transformer module aggregates the features from neighboring points within the feature space and assigns self-attention weights, effectively extending the range of perception by stacking these modules. In the decoder, we introduce a transformer refinement strategy that utilizes the query-wise self-attention mechanism to facilitate information exchange among potential keypoints. Additionally, it incorporates the cross-attention mechanism to capture the latent connections among keypoints' heatmaps, positions, and semantics. The query-wise decoder optimizes the heatmap and query features through the analysis of inter-keypoint and inter-attribute relationships, thereby enabling the parallel prediction of multi-heatmaps and probabilities.", + "bbox": [ + 212, + 146, + 787, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We validate our method on the KeypointNet dataset [47], achieving outstanding results in two essential tasks: keypoint saliency and correspondence estimation. Furthermore, our approach of assigning a specific heatmap to each keypoint leads to significantly higher accuracy compared to methods that use a shared heatmap.", + "bbox": [ + 212, + 464, + 787, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as follows:", + "bbox": [ + 238, + 540, + 632, + 554 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce KeypointDETR, an end-to-end 3D keypoint detection framework. It is trained with a loss function based on the optimal bipartite matching derived from the cost matrix of multi-heatmaps and probabilities.", + "- We construct a novel transformer-based network architecture. The point-wise encoder harnesses self-attention on the dynamic graph of points' local feature space to extract geometric features. The query-wise decoder incorporates global self-attention to enhance the interaction between queries.", + "- We conduct a comprehensive evaluation that validates our method and provides insights into the impact of principal components, such as the bipartite matching loss and query-wise self-attention." + ], + "bbox": [ + 225, + 569, + 784, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 747, + 387, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 3D Keypoint Detection", + "text_level": 1, + "bbox": [ + 214, + 782, + 454, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In previous studies on 3D keypoint detection [23,33,37], algorithms heavily rely on a variety of hand-crafted geometric features, such as ISS [30], HKS [34],", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "KeypointDETR", + "bbox": [ + 622, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CGF [14], Salient Points [3], and Mesh Saliency [16]. These traditional approaches primarily focus on local geometric features of 3D shapes, often overlooking the integration of global and semantic information.", + "bbox": [ + 212, + 145, + 782, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With advancements in deep learning, there has been a notable surge in the development of data-driven algorithms for keypoint detection [10, 35, 54]. Numerous unsupervised 3D keypoint detection algorithms [1, 7, 12, 18, 31, 50, 53] are proposed for specific tasks such as point registration and reconstruction. For example, UKPGAN [46] utilizes a GAN-based module to manage keypoint sparsity and reconstructs 3D models by distilling saliency information. Unsupervised methods tend to focus on identifying keypoints in areas with distinct geometric features. While some of these methods are capable of attaining semantic consistency, they typically require the predetermination of the number of keypoints. Furthermore, in many scenarios, it is essential to identify keypoints that may not have prominent geometric features but possess solid semantic importance. Methods like SyncSpecCNN [45] and PRA-Net [5] address keypoint detection by classifying points on 3D shapes. These methods often result in an imbalance between positive and negative samples, presenting a significant challenge. Consequently, current popular methods employ a heatmap-based approach [32, 41], where the value assigned to each point indicates its proximity to the nearest keypoint. Wei et al. [42] proposes a multi-task learning framework that combines point-to-keypoint offsets with a confidence map, facilitating effective 3D keypoint saliency and correspondence estimation. Heatmap-based methods for 3D shapes predominantly depend on a shared heatmap for all keypoints and require post-processing. These methods employ post-processing techniques like non-maximum suppression [22] or clustering [29] to isolate keypoints from regions with high scores during the inference process.", + "bbox": [ + 212, + 193, + 787, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Transformer in Vision", + "text_level": 1, + "bbox": [ + 215, + 566, + 442, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformer [38] was initially introduced in the realm of natural language processing. Recent advancements in visual research [6, 19] have explored various models centered around self-attention [38], achieving unprecedented performance. DETR [2] pioneers an innovative end-to-end object detection approach. It formulates a loss matching rule employing the Hungarian algorithm [15], facilitating bipartite matching between predictions and ground truth. This method serves as the cornerstone for a series of object detection methods [4,17,55] based on set prediction. Li et al. [17] develop a facial landmark detection method using cascaded transformer decoders. They devise parallel decoders utilizing deformable self-attention [55] to optimize facial landmark coordinates through offset predictions. Chen et al. [4] leverage self-attention to create a structure-aware LSTM framework tailored for predicting 3D heatmaps in CBCT. Keypoint transformer [9] efficiently extracts all potential keypoints within the image from a shared heatmap, subsequently employing the DETR decoder to predict 3D hand poses. Given the transformer's inherent capacity to handle input element permutations in sequences, it is exceptionally well-suited for point cloud processing [8, 24, 26].", + "bbox": [ + 212, + 598, + 787, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H.Jin et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d3926970f31692af47b56440d3e54f96bb96946fb089ed5910f49310668f2a29.jpg", + "image_caption": [ + "Fig. 2: Pipeline of KeypointDETR for keypoint saliency and correspondence estimation. Our framework takes point coordinates as input. The point-wise encoder first extracts geometric features from the point cloud by stacking dynamic graph transformer modules (DGT). Then, the query-wise decoder accepts heatmap features, which are transformed into heatmap and query features to generate multi-heatmaps and probabilities. These predictions enable the direct inference of both keypoints positions and semantic labels." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 787, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Nonetheless, applying global self-attention to 3D point cloud models incurs substantial computational costs. Consequently, existing research [48, 52] frequently adopts local self-attention mechanisms, expanding receptive fields through sampling and grouping strategies. In contrast, our point-wise encoder integrates local self-attention within the dynamic graph of the point feature space, thereby eliminating the necessity for sampling and grouping operations. Distinct from most existing DETR style works, our query-wise decoder is meticulously engineered to predict multi-heatmaps and probabilities, enabling efficient end-to-end 3D keypoint detection.", + "bbox": [ + 212, + 503, + 787, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 659, + 330, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we delve into the specifics of our end-to-end 3D keypoint detector, KeypointDETR, which consists of two essential components: a network architecture based on the transformer and a loss function formulated using bipartite matching. As illustrated in Fig. 2, our approach is applied to tasks involving keypoint saliency and correspondence estimation. Upon receiving a 3D point cloud as input, our point-wise encoder utilizes the dynamic graph transformer modules to extract heatmap features, which can be considered intrinsic geometric features. We then deploy a query-wise transformer decoder designed to enhance information exchange between queries and capture the latent relationships between attributes through cross-attention, thereby optimizing heatmap", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "KeypointDETR", + "bbox": [ + 622, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a172b7cf989f94f61f4dfb7be805e5e4f0d0929863e53e084955cb4fe6015a29.jpg", + "image_caption": [ + "(a) The structure of the dynamic graph transformer module." + ], + "image_footnote": [], + "bbox": [ + 215, + 146, + 537, + 250 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/64d5d4d97f9c41e14488bef42082634eff88c72bb8921cbd0873f3538faa0e5c.jpg", + "image_caption": [ + "(b) The structure of the transformer refinement module.", + "Fig. 3: The structure of the dynamic graph transformer module (DGT) and transformer refinement module." + ], + "image_footnote": [], + "bbox": [ + 544, + 143, + 759, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and query features. This network yields heatmaps and probabilities as its outputs. The prediction sets are aligned with ground truth heatmaps through a cost matrix, leading to the computation of the final loss based on optimal bipartite matching. During the inference stage, we select heatmaps corresponding to keypoints, as indicated by the predicted probabilities, and extract keypoints exhibiting the highest heat values from these maps.", + "bbox": [ + 212, + 343, + 787, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Data Preparation", + "text_level": 1, + "bbox": [ + 215, + 455, + 408, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To tackle the issue of imbalanced samples concerning keypoints on the surface of 3D models, heatmap-based 3D keypoint detection methods establish the ground truth heat value for each point based on its distances to the keypoints. Specifically, the heatmap is generated by applying a Gaussian kernel in conjunction with the geodesic distance matrix:", + "bbox": [ + 212, + 482, + 787, + 556 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i, j} = \\exp^ {- \\frac {1}{2 \\sigma^ {2}} \\mathrm {d} \\left(p _ {i}, k _ {j}\\right)}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 566, + 785, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given a 3D point cloud with $N_{p}$ points $P = \\{p_{i}\\}_{i=1}^{N_{p}}$ and $N_{k}$ keypoints $K = \\{k_{j}\\}_{j=1}^{N_{k}}$ , our process begins with the computation of the geodesic distance matrix from each point to the keypoints. Then, we derive the heatmap matrix $\\mathcal{H} = \\{h_{i,j}\\} \\in \\mathbb{R}^{N_{p} \\times N_{k}}$ using the Gaussian kernel function. Contrasting with methods that use a shared heatmap, where the maximum value for each point is considered, our approach distinctively employs multi-heatmaps, each corresponding to specific keypoints, as the ground truth.", + "bbox": [ + 214, + 599, + 787, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Network Architecture", + "text_level": 1, + "bbox": [ + 215, + 731, + 444, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 2, our transformer-based encoder-decoder network architecture primarily comprises a point-wise encoder and a query-wise decoder.", + "bbox": [ + 214, + 756, + 785, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Point-wise Encoder The main objective of the encoder is to extract geometric features from the 3D point cloud. Drawing inspiration from DGCNN [40] and", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H.Jin et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the Point Transformer [52], we devise a point-wise encoder constructed using dynamic graph transformer modules (DGT).", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As depicted in Fig. 3 (a), we employ local self-attention within the dynamic graph transformer module to operate on the feature space. To be specific, for the feature map $\\mathbf{F}$ at each layer, we extract a $k$ -nearest neighbors feature set $\\mathbf{F}_i = \\{\\mathbf{f}_j\\}_{j=1}^k$ for $\\mathbf{f}_i$ . Thereafter, we assign self-attention weights to these $k$ -nn features and aggregate them using the following equation:", + "bbox": [ + 212, + 176, + 784, + 253 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {i} ^ {\\prime} = \\sum_ {\\mathbf {f} _ {j} \\in \\mathbf {F} _ {i}} \\rho (\\theta (\\phi (\\mathbf {f} _ {i}) - \\varphi (\\mathbf {f} _ {j}) + \\delta)) \\odot (\\psi (\\mathbf {f} _ {j}) + \\delta). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 263, + 785, + 296 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, $\\theta$ is implemented as an MLP, while $\\phi$ , $\\varphi$ , $\\psi$ are realized using Conv1d. $\\rho$ represents the softmax operation applied to self-attention weights, and $\\odot$ denotes the element-wise product operation. The term $\\delta$ represents the position encoding, which is generated from the point coordinates $P_{i} = \\{p_{j}\\}_{j=1}^{k}$ corresponding to the neighborhood features, as expressed in the following:", + "bbox": [ + 214, + 308, + 782, + 383 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\delta = \\operatorname {M L P} \\left(p _ {i} - p _ {j}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 397, + 784, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we concatenate the features extracted from each dynamic graph transformer module to obtain point-wise features, which are decoded into heatmap features through MLPs.", + "bbox": [ + 214, + 424, + 782, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Query-wise Decoder The decoder's output comprises a predefined number of predictions, referred to as keypoint queries, with the query count $M$ set significantly higher than the number of keypoints within a 3D model. Each prediction query consists of both a heatmap and a probability. Heatmaps serve to indicate the position of keypoints within the 3D model. Probabilities reflect the semantic consistency of potential keypoints. In keypoint saliency estimation, they determine the presence or absence of specific keypoints. In keypoint correspondence estimation, such probabilities are correlated with their respective semantic labels.", + "bbox": [ + 212, + 492, + 784, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Guided by these objectives, we develop a query-wise decoder that adheres to the proven DETR-style architecture [2], as illustrated in Fig. 3 (b). Initially, we take the heatmap features $\\mathbf{F}_{heat} \\in \\mathbb{R}^{N \\times M}$ from the encoder. Considering the inherent relationship between the heatmaps and semantics of keypoints, the heatmap features are directly transposed and then decoded into query features $\\mathbf{F}_{prob} \\in \\mathbb{R}^{M \\times 128}$ through an MLP. Following this, both the heatmap $\\mathbf{F}_{heat}$ and query $\\mathbf{F}_{prob}$ features are fed into the transformer decoder. To facilitate the exchange of information among queries, a multi-head self-attention module is applied to $\\mathbf{F}_{prob}$ . For the position encoding, we extract the coordinates of the reference points corresponding to the peaks of heat values in $\\mathbf{F}_{heat}$ , which are subsequently processed through an MLP to produce the position encoding. The transposed $\\mathbf{F}_{heat}$ and $\\mathbf{F}_{prob}$ serve as inputs to the cross-attention module, enhancing the interplay of information between probability and heatmap features. Ultimately, through the application of Feed-Forward Network modules (FFN),", + "bbox": [ + 212, + 628, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "KeypointDETR", + "bbox": [ + 622, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the refined $\\mathbf{F}_{heat}$ and $\\mathbf{F}_{prob}$ are decoded into multi-heatmaps $\\mathbf{M}_h$ and probabilities $\\mathbf{M}_p$ , respectively.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The decoder utilizes a query-wise self-attention mechanism across the queries, leading to $M$ distinct predictions. This mechanism operates as a global self-attention centered around potential keypoints. Such a globally parallel decoding approach efficiently captures the interactions among the predicted queries, thereby optimizing heat values and reducing redundancy in the identification of keypoints. Additionally, considering the potential relationship between keypoint heatmaps and semantics, cross-attention is employed to facilitate the interaction between heatmap and query features.", + "bbox": [ + 212, + 176, + 787, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Loss Function", + "text_level": 1, + "bbox": [ + 215, + 318, + 377, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inspired by DETR [2], we integrate the bipartite matching loss into 3D keypoint detection to enable end-to-end inference, eliminating the need for postprocessing. During the bipartite matching phase, we calculate the pairwise matching costs by assessing the discrepancies between the predicted queries and ground truth. We utilize this cost matrix to consolidate the prediction sets, distinguishing between \"background\" queries and keypoint queries, in order to compute the final loss.", + "bbox": [ + 212, + 342, + 787, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given a 3D model with $N_{k}$ ground truth keypoints, KeypointDETR generates a set of $M$ predictions, where $M$ is preconfigured to be significantly larger than $N_{k}$ . The ground truth can be represented as $\\{y_{i}\\}_{i = 1}^{N_{k}} = \\{h_{i},p_{i}\\}_{i = 1}^{N_{k}}$ , where $h_i$ signifies the ground truth heatmap and $p_i$ denotes the ground truth probability. The predicted queries are denoted as $\\{\\hat{y}_j\\}_{j = 1}^M = \\{\\hat{h_j},\\hat{p_j}\\}_{j = 1}^M$ , with $\\hat{h_j}$ representing the predicted heatmap, and $\\hat{p_j}$ corresponding to the predicted probabilities associated with keypoints categories $\\{c_i\\}_{i = 1}^{N_k}$ . In the task of keypoint saliency estimation, $p_i$ remains fixed at 1, while $\\hat{p_j}(c_i)$ signifies the probability of being predicted as a keypoint. However, in the task of keypoint correspondence estimation, $\\hat{p_j}(c_i)$ is defined as the probability of $c_i$ , where $c_i$ denotes the semantic label corresponding to the ground truth keypoint $k_i$ . Expanding upon these notations, in order to achieve the optimal bipartite matching between ground truth and predictions, it is imperative to establish a matching cost matrix $\\mathcal{M}_c \\in \\mathbb{R}^{N_k \\times M}$ . This cost matrix takes into account both the heatmaps and probabilities. Precisely, each element $\\mathcal{M}_c(i,j)$ of the cost matrix can be calculated as follows:", + "bbox": [ + 212, + 449, + 787, + 678 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {c} (i, j) = \\mathbf {W} _ {\\mathbf {h}} \\cdot \\sum | h _ {i} - \\hat {h} _ {j} | + \\mathbf {W} _ {\\mathbf {p}} \\cdot \\left(1 - \\log p _ {\\hat {j}} \\left(c _ {i}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 688, + 785, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{W_h}$ represents the weight assigned to heatmaps, and $\\mathbf{W_p}$ is the weight attributed to probabilities. $\\sum |h_i - \\hat{h_j}|$ signifies the dissimilarity between the ground truth $h_i$ and predicted $\\hat{h_j}$ heatmaps.", + "bbox": [ + 212, + 715, + 784, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Following the application of the Hungarian algorithm [15], we obtain the optimal matching and categorize the predicted queries into two groups: positive queries and negative queries. Positive queries are those that align with the ground truth. For a positive query $\\hat{y}_i$ , we determine the matched ground truth category as $g_i$ and the corresponding heatmap as $h_i$ . Then, we compute the probability", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H.Jin et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "loss using the Focal Loss (FL) and the heatmap loss employing the Mean Squared Error Loss (MSE). For a negative query, only the probability loss is considered. The associated ground truth $g_{i}$ is designated as 0, signifying nonexistence. Thus, the overall loss can be calculated as follows:", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {i = 1} ^ {M} \\left(\\lambda_ {p} \\operatorname {F L} \\left(g _ {i}, \\hat {p _ {i}}\\right) + \\lambda_ {h} \\mathbb {1} _ {\\{g _ {i} \\neq 0 \\}} \\operatorname {M S E} \\left(h _ {i}, \\hat {h _ {i}}\\right)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 214, + 785, + 252 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 272, + 375, + 289 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we conduct a series of comprehensive experiments to validate the superior performance of our approach in both keypoint saliency and correspondence estimation tasks. We provide comparisons with other state-of-the-art methods to demonstrate our method's efficacy.", + "bbox": [ + 212, + 301, + 782, + 361 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 382, + 455, + 397 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "KeypointDETR takes 3D point coordinates as input. In the point-wise encoder, we utilize 5 dynamic graph transformer modules, each with point features of 64 dimensions, the $k$ value of $k$ -nearest neighbors set to 20, and a global embedding dimension of 1024. In the query-wise decoder, we set the number of queries $M$ to 50. The decoder's output comprises two branches: the heatmap head generates multi-heatmaps $\\mathbf{M}_h \\in \\mathbb{R}^{M \\times N_p}$ , and the classification head provides probabilities $\\mathbf{M}_p \\in \\mathbb{R}^{M \\times C}$ . For keypoint saliency estimation, $C$ is held constant at 2. However, for keypoint correspondence estimation, $C$ is configured as $(N_c + 1)$ , where $N_c$ signifies the number of keypoint categories. More details can be found in the supplementary material.", + "bbox": [ + 212, + 405, + 784, + 555 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Dataset", + "text_level": 1, + "bbox": [ + 215, + 575, + 326, + 589 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our experiments are conducted on the KeypointNet dataset [47], which is designed for both keypoint saliency and correspondence estimation tasks. We focus on four classical categories: airplane, chair, guitar, and table. These models are divided into training, validation, and test sets, with a distribution ratio of 7:1:2 for each category. We use 2,048 sampled points from each 3D model as input data. The results of other categories and applications of KeypointDETR can be found in the supplementary material.", + "bbox": [ + 212, + 599, + 782, + 704 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Metrics", + "text_level": 1, + "bbox": [ + 215, + 724, + 326, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate our method alongside comparison methods on the keypoint saliency estimation using mIoU and Chamfer Distance (CD) while utilizing mIoU and Dual Alignment Score (DAS) [31] for keypoint correspondence estimation. The IoU is calculated as follows:", + "bbox": [ + 212, + 748, + 782, + 806 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nI o U = \\frac {T P}{T P + F P + F N}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 813, + 784, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "KeypointDETR", + "bbox": [ + 622, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $TP$ stands for true positives, $FP$ represents false positives, and $FN$ corresponds to false negatives. Given the ground truth and predicted keypoints sets $G$ and $K$ , CD can be computed as:", + "bbox": [ + 212, + 146, + 784, + 191 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nC D = \\frac {1}{| G |} \\sum_ {p \\in G} \\min _ {q \\in K} \\| p - q \\| _ {2} ^ {2} + \\frac {1}{| K |} \\sum_ {p \\in K} \\min _ {q \\in G} \\| p - q \\| _ {2} ^ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 204, + 785, + 241 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The settings for clustering and NMS parameters profoundly affect the repeatability and omission rates of predicted keypoints. However, the mIoU metric used in prior methods fails to effectively capture these erroneous cases. To ensure a comprehensive comparison, we design a mIoU metric based on the Hungarian algorithm [15]. For the keypoint saliency estimation, we begin by calculating the Euclidean distance matrix between the ground truth and the predicted keypoints. Subsequently, the Hungarian algorithm is employed to determine their optimal bipartite matching. Each matched distance falling below the predefined threshold $\\epsilon$ is considered a true positive. For keypoint correspondence estimation, a ground truth keypoint is considered a true positive if there is a predicted keypoint with the same semantic label whose distance to it is less than the threshold $\\epsilon$ . Additionally, we further measure whether the closest keypoints between predictions and ground truth have the same semantic labels using DAS [31].", + "bbox": [ + 212, + 253, + 787, + 450 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 Comparison on Keypoint Saliency Estimation", + "text_level": 1, + "bbox": [ + 214, + 477, + 637, + 493 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We conduct a comparative analysis of our method against six advanced algorithms, including two unsupervised keypoint detection methods (Skeleton Merger [31] and UKPGAN [46]) and three point cloud geometric feature extraction algorithms (PointNet++ [28], PRA-Net [5], and DGCNN [40]), and a fully-supervised keypoint detection method (Wei et al. [42]). For the three point cloud geometric feature extraction algorithms on saliency analysis, the network's output is a shared heatmap, wherein each point is assigned the maximum value of the multi-heatmaps.", + "bbox": [ + 212, + 506, + 812, + 627 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The quantitative results for mIoU and Chamfer Distance (CD) are presented in Tab. 1, where we compare our KeypointDETR with other algorithms using a threshold $\\epsilon$ of 0.1. Fig. 6 (a) depicts the mIoU curves for various methods as $\\epsilon$ varies from 0 to 0.1. It is evident that KeypointDETR consistently outperforms other methods in the Hungarian mIoU evaluation across all categories. This significant advantage stems from our method's capacity to bypass post-processing techniques like NMS or clustering algorithms, leading to direct 'one-to-one' predictions that effectively tackle redundancy and omissions in 3D keypoint detection. In contrast, methods reliant on a shared heatmap face challenges with parameter selection of post-processing, which predominantly depends on prior knowledge factors such as the expected number of keypoints and the 3D model's scale. Unfortunately, even optimal parameter settings for NMS (radius=0.1) in these methods do not entirely resolve issues of omissions and redundant predictions, as further illustrated in Fig. 4.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H.Jin et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e9d4af2281c9871703b8ec87417582c1043b1bfc8d8ed5e64fc4de01820bfea3.jpg", + "table_caption": [ + "Table 1: Quantitative results of keypoint saliency estimation using Hungarian mIoU (\\%) and CD." + ], + "table_footnote": [], + "table_body": "
AirplaneChairGuitarTableAverage
mIoUCDmIoUCDmIoUCDmIoUCDmIoUCD
SkeletonMerger [31]60.730.10742.740.18947.710.13227.840.30244.750.182
UKPGAN [46]62.630.11825.160.16854.250.13117.780.24039.950.164
PRA-Net [5]82.880.04970.440.06073.710.05369.690.05374.180.053
PointNet++ [28]82.310.05057.500.07772.340.06060.010.06268.040.062
DGCNN [40]78.490.05859.490.07663.880.07259.920.06665.440.068
Wei et al. [42]79.860.04886.280.05578.080.06686.610.03982.700.052
Ours91.370.04591.890.05593.960.04598.050.02793.810.043
", + "bbox": [ + 215, + 184, + 795, + 339 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/da1962fca4c8e3590bda244efc3a23c4c0223b2aec5bd070bbb4e446bf948610.jpg", + "image_caption": [ + "Fig. 4: Visualization examples of keypoint saliency estimation." + ], + "image_footnote": [], + "bbox": [ + 215, + 348, + 787, + 551 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.5 Comparison on Keypoint Correspondence Estimation", + "text_level": 1, + "bbox": [ + 215, + 604, + 700, + 619 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Keypoint correspondence estimation represents a more significant challenge compared to keypoint saliency estimation, as it not only involves locating keypoints but also predicting their semantic labels. Additionally, 3D models within the same category may not have identical semantic keypoints. Prior approaches [35, 45,47] typically involve direct classification of points within a point cloud, which does not effectively address the issue of imbalanced positive and negative samples in keypoint detection. Some methods [42, 54] attempt to add a semantic branch to keypoint saliency estimation, aiming to assign labels to keypoints in a two-stage process.", + "bbox": [ + 212, + 628, + 785, + 763 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As depicted in Tab. 2, we conduct a comparison of our method with five methods using a threshold $\\epsilon$ of 0.1: PointConv [43], PointNet++ [28], DGCNN [40], PRA-Net [5], and Wei et al. [42]. Additionally, we present the mIoU curves for various methods, with the $\\epsilon$ ranging from 0 to 0.1, in Fig. 6 (b). Fig. 5 showcases visualization results of different methods alongside the ground truth. Notably,", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "KeypointDETR", + "bbox": [ + 622, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/9d66a3c820742232df2df361abe05de6d2549eccc35075af2254cfcb99ad8eaa.jpg", + "table_caption": [ + "Table 2: Quantitative results of keypoint correspondence estimation using Hungarian mIoU (%) and DAS [31] (%)." + ], + "table_footnote": [], + "table_body": "
AirplaneChairGuitarTableAverage
mIoUDASmIoUDASmIoUDASmIoUDASmIoUDAS
DGCNN [40]77.7782.6963.1679.5661.1378.5957.9978.8365.0179.91
PointNet++ [28]79.1084.7458.2478.1967.6981.9262.9781.5967.0081.61
PointConv [43]81.5485.7464.9282.1576.3988.9166.2083.7372.2685.13
PRA-Net [5]81.1085.8068.3583.6272.0785.1570.1684.8972.9284.86
Wei et al. [42]78.5488.3479.7790.4475.6493.6986.5697.3180.1292.44
Ours85.7191.7479.3389.5390.9393.8992.4097.7987.0993.23
", + "bbox": [ + 220, + 184, + 782, + 323 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0b90e31dcf5ceaea378fae405d2c9b7f7502fbbac83dac648efd28d13eddb5bc.jpg", + "image_caption": [ + "Fig. 5: Visualization examples of keypoint correspondence estimation." + ], + "image_footnote": [], + "bbox": [ + 217, + 333, + 787, + 508 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "our method generates no redundant points with the same label in close proximity to predicted keypoints. This demonstrates that our method adeptly accomplishes semantic correspondence for keypoints, with KeypointDETR showcasing superior performance in both the heatmap and probability branch.", + "bbox": [ + 212, + 561, + 784, + 622 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 643, + 390, + 657 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conduct a series of ablation experiments to validate the performance of KeypointDETR and deeply explore the impact of each component. Throughout these experiments, we consistently use four representative categories from the KeypointNet dataset for keypoint saliency estimation. The evaluation adopts mIoU with a threshold $\\epsilon$ of 0.1.", + "bbox": [ + 212, + 666, + 782, + 742 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effects of point-wise encoder. Our point-wise encoder primarily consists of dynamic graph transformer modules that extract geometric features from the point cloud. To validate the effectiveness of our point-wise encoder, we compare KeypointDETR with its variants, which utilize different backbones for geometric feature extraction, namely DGCNN [40], PointNet++ [28], and Point Transformer [52]. The DGCNN [40] backbone (DG.Enc.) is implemented as a five-layer", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H.Jin et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5272ec86e824043d21ad877d71e1dc6aa7d8a4ce5f80e266cd8b661fb74c95e7.jpg", + "image_caption": [ + "(a) Hungarian mIoU curves for keypoint saliency estimation." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 455, + 284 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6638a544fd8bea7b654de57d2c5544416436791a87ad1688af1823ae8fa958c7.jpg", + "image_caption": [ + "(b) Hungarian mIoU curves for keypoint correspondence estimation.", + "Fig. 6: Hungarian mIoU curves for keypoint saliency and correspondence estimation under various distance thresholds (0-0.1)." + ], + "image_footnote": [], + "bbox": [ + 526, + 143, + 767, + 284 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/797800f87d2e26fb50c5c2c58126a5c65c29821691182de8cdcccc4c0c69eaab.jpg", + "table_caption": [ + "Table 3: Quantitative results of various variations in the ablation study using Hungarian mIoU (%) and CD." + ], + "table_footnote": [], + "table_body": "
AirplaneChairGuitarTableAverage
mIoUCDmIoUCDmIoUCDmIoUCDmIoUCD
DG.Enc. [40]86.750.04882.120.06292.760.05297.620.03789.810.049
PN2.Enc. [28]76.300.07168.310.10172.090.09976.420.09973.280.092
PT.Enc. [52]82.550.05973.040.08568.920.09197.320.03280.450.066
Point-BERT Enc. [49]91.110.03493.570.03493.510.03293.710.03392.880.033
I2P-MAE Enc. [51]79.060.05080.400.06788.790.04886.050.04883.570.053
Point-MAE Enc. [25]89.820.03993.980.05583.790.04395.950.04390.880.045
w/o Trans.Dec.87.960.04488.230.05892.740.04597.770.02991.670.044
w/o PE.88.050.04291.430.05492.440.04797.970.02892.470.042
Coord-based66.720.10844.550.19037.990.16450.080.18749.830.162
Coord+offsets68.400.09463.620.13642.450.15160.770.15958.810.135
Ours91.370.04591.890.05593.960.04598.050.02793.810.043
", + "bbox": [ + 243, + 402, + 759, + 577 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "stack of EdgeConv. PointNet++ [28] (PN2.Enc.) and Point Transformer [52] (PT.Enc.) employ eight layers due to their U-net structures. Furthermore, we explore replacing our encoder with pre-trained models, including Point-BERT, Point-MAE, and I2P-MAE, and achieve favorable results. The overall quantitative comparison results are presented in the first sub-table of Tab. 3.", + "bbox": [ + 212, + 604, + 784, + 680 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effects of query-wise decoder. In the decoder phase, we develop a query-wise decoder utilizing self-attention to facilitate enhanced information exchange among queries. This decoder is tasked with decoding the heatmap and probability feature maps, resulting in the generation of multi-heatmaps and probabilities. A straightforward approach involves employing MLPs to decode these feature maps separately. However, to demonstrate the efficacy and reasoning behind our query-wise decoder design, we conduct an experiment where we replace our self-attention decoder (Trans.Dec.) with MLPs. The comparison results, as shown in the second sub-table of Tab. 3, indicate that the comprehensive performance of our KeypointDETR significantly surpasses that of its MLPs decoder variant.", + "bbox": [ + 212, + 689, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "KeypointDETR", + "bbox": [ + 622, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effects of position encoding in decoder. Position encoding holds a pivotal role in the self-attention mechanism and is a key factor that makes the transformer architecture suitable for 3D point cloud models, as previously validated in the Point Transformer [52]. In our investigation, we further explore the role of position encoding in our query-wise decoder, which is generated from heatmap features and point coordinates. We compare the results with and without position encoding (PE.) in the third sub-table of Tab. 3.", + "bbox": [ + 215, + 145, + 785, + 252 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Coordinate-based or Heatmap-based. We further explore a coordinate-based architecture within our KeypointDETR approach. In this variant, KeypointDETR predicts the coordinates and probabilities of keypoints, and we subsequently select the closest matching points from the point cloud as the final keypoints. To enhance the accuracy of these coordinates, we incorporate a cascade module designed to predict offsets. Comparative analysis of the heatmap-based KeypointDETR and its coordinate-based variant, as shown in the fourth sub-table of Tab. 3, reveals that while the addition of offsets improves the results of the coordinate-based model, there remains a notable performance gap when compared to our primary heatmap-based KeypointDETR. This discrepancy arises because our geometric feature extraction backbone is better suited for processing 3D shapes rather than 3D spatial coordinates. Furthermore, it validates that the heatmap-based approach can achieve better performance on keypoints detection.", + "bbox": [ + 215, + 263, + 785, + 474 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 508, + 357, + 523 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce KeypointDETR, an end-to-end 3D keypoint detector that obviates the need for post-processing steps, such as clustering and NMS, commonly required in heatmap-based keypoint detection methods. The training strategy of KeypointDETR involves computing a cost matrix between the predictions and ground truth, deriving an optimal bipartite matching loss. This innovative approach enables KeypointDETR to accurately predict multi-heatmaps and probabilities, which can provide the positions and semantic consistency of potential keypoints, respectively. Complementing this framework is a transformer-based network architecture. Extensive quantitative experiments and analyses substantiate the effectiveness and rationality of KeypointDETR, underlining its potential as a state-of-the-art solution in keypoint detection.", + "bbox": [ + 215, + 551, + 785, + 718 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 751, + 403, + 768 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work is partially supported by the following grants: National Natural Science Foundation of China (No. 62172363), Management Center of the School of Software (Ningbo) at Zhejiang University under Grant (No. Z24001).", + "bbox": [ + 215, + 794, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H.Jin et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bai, X., Luo, Z., Zhou, L., Fu, H., Quan, L., Tai, C.L.: D3feat: Joint learning of dense detection and description of 3d local features. In: CVPR. pp. 6359-6367 (2020)", + "2. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV. pp. 213-229. Springer (2020)", + "3. Castellani, U., Cristani, M., Fantoni, S., Murino, V.: Sparse points matching by combining 3d mesh saliency with statistical descriptors. In: Comput. Graph. Forum. vol. 27, pp. 643-652. Wiley Online Library (2008)", + "4. Chen, R., Ma, Y., Chen, N., Liu, L., Cui, Z., Lin, Y., Wang, W.: Structure-aware long short-term memory network for 3d cephalometric landmark detection. IEEE Transactions on Medical Imaging 41(7), 1791-1801 (2022)", + "5. Cheng, S., Chen, X., He, X., Liu, Z., Bai, X.: Pra-net: Point relation-aware network for 3d point cloud analysis. IEEE TIP 30, 4436-4448 (2021)", + "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "7. Fernandez-Labrador, C., Chhatkuli, A., Paudel, D.P., Guerrero, J.J., Demonceaux, C., Gool, L.V.: Unsupervised learning of category-specific symmetric 3d keypoints from point sets. In: ECCV. pp. 546-563. Springer (2020)", + "8. Guo, M.H., Cai, J.X., Liu, Z.N., Mu, T.J., Martin, R.R., Hu, S.M.: Pct: Point cloud transformer. Computational Visual Media 7, 187-199 (2021)", + "9. Hampali, S., Sarkar, S.D., Rad, M., Lepetit, V.: Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In: CVPR. pp. 11090-11100 (2022)", + "0. He, Y., Sun, W., Huang, H., Liu, J., Fan, H., Sun, J.: Pvn3d: A deep point-wise 3d keypoints voting network for 6 dof pose estimation. In: CVPR. pp. 11632-11641 (2020)", + "1. Huang, Q., Wang, W., Neumann, U.: Recurrent slice networks for 3d segmentation of point clouds. In: CVPR. pp. 2626-2635 (2018)", + "2. Jakab, T., Tucker, R., Makadia, A., Wu, J., Snavely, N., Kanazawa, A.: Keypoint-deformer: Unsupervised 3d keypoint discovery for shape control. In: CVPR. pp. 12783-12792 (2021)", + "3. Katz, S., Leifman, G., Tal, A.: Mesh segmentation using feature point and core extraction. The Vis. Comput. 21, 649-658 (2005)", + "4. Khoury, M., Zhou, Q.Y., Koltun, V.: Learning compact geometric features. In: ICCV. pp. 153-161 (2017)", + "5. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955)", + "6. Lee, C.H., Varshney, A., Jacobs, D.W.: Mesh saliency. In: ACM SIGGRAPH 2005 Papers, pp. 659-666 (2005)", + "7. Li, H., Guo, Z., Rhee, S.M., Han, S., Han, J.J.: Towards accurate facial landmark detection via cascaded transformers. In: CVPR. pp. 4176-4185 (2022)", + "8. Li, J., Lee, G.H.: Usip: Unsupervised stable interest point detection from 3d point clouds. In: ICCV. pp. 361-370 (2019)", + "9. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV. pp. 10012-10022 (2021)" + ], + "bbox": [ + 225, + 176, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "KeypointDETR", + "bbox": [ + 622, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "20. Liu, Z., Wu, Z., Toth, R.: Smoke: Single-stage monocular 3d object detection via keypoint estimation. In: CVPR. pp. 996-997 (2020)", + "21. Mian, A.S., Bennamoun, M., Owens, R.: Three-dimensional model-based object recognition and segmentation in cluttered scenes. IEEE TPAMI 28(10), 1584-1601 (2006)", + "22. Neubeck, A., Van Gool, L.: Efficient non-maximum suppression. In: ICPR. vol. 3, pp. 850-855. IEEE (2006)", + "23. Novatnack, J., Nishino, K.: Scale-dependent 3d geometric features. In: ICCV. pp. 1-8. IEEE (2007)", + "24. Pan, X., Xia, Z., Song, S., Li, L.E., Huang, G.: 3d object detection with pointformer. In: CVPR. pp. 7463-7472 (2021)", + "25. Pang, Y., Wang, W., Tay, F.E., Liu, W., Tian, Y., Yuan, L.: Masked autoencoders for point cloud self-supervised learning. In: ECCV. pp. 604-621. Springer (2022)", + "26. Park, C., Jeong, Y., Cho, M., Park, J.: Fast point transformer. In: CVPR. pp. 16949-16958 (2022)", + "27. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR. pp. 652-660 (2017)", + "28. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS 30 (2017)", + "29. Rodriguez, A., Laio, A.: Clustering by fast search and find of density peaks. science 344(6191), 1492-1496 (2014)", + "30. Salti, S., Tombari, F., Spezialietti, R., Di Stefano, L.: Learning a descriptor-specific 3d keypoint detector. In: ICCV. pp. 2318-2326 (2015)", + "31. Shi, R., Xue, Z., You, Y., Lu, C.: Skeleton merger: an unsupervised aligned keypoint detector. In: CVPR. pp. 43-52 (2021)", + "32. Shu, Z., Yu, J., Chao, K., Xin, S., Liu, L.: A multi-modal attention-based approach for points of interest detection on 3d shapes. IEEE TVCG (2024)", + "33. Sipiran, I., Bustos, B.: Harris 3d: a robust extension of the harris operator for interest point detection on 3d meshes. The Vis. Comput. 27, 963-976 (2011)", + "34. Sun, J., Ovsjanikov, M., Guibas, L.: A concise and provably informative multiscale signature based on heat diffusion. In: Comput. Graph. Forum. vol. 28, pp. 1383-1392. Wiley Online Library (2009)", + "35. Sung, M., Su, H., Yu, R., Guibas, L.J.: Deep functional dictionaries: Learning consistent semantic structures on 3d models from functions. NeurIPS 31 (2018)", + "36. Tang, J., Gong, Z., Yi, R., Xie, Y., Ma, L.: Lake-net: Topology-aware point cloud completion by localizing aligned keypoints. In: CVPR. pp. 1726-1735 (2022)", + "37. Tombari, F., Salti, S., Di Stefano, L.: Unique signatures of histograms for local surface description. In: ECCV. pp. 356-369. Springer (2010)", + "38. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)", + "39. Wang, H., Guo, J., Yan, D.M., Quan, W., Zhang, X.: Learning 3d keypoint descriptors for non-rigid shape matching. In: ECCV. pp. 3-19 (2018)", + "40. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. ACM TOG 38(5), 1-12 (2019)", + "41. Wei, G., Cui, Z., Zhu, J., Yang, L., Zhou, Y., Singh, P., Gu, M., Wang, W.: Dense representative tooth landmark/axis detection network on 3d model. Computer Aided Geometric Design 94, 102077 (2022)", + "42. Wei, G., Ma, L., Wang, C., Desrosiers, C., Zhou, Y.: Multi-task joint learning of 3d keypoint saliency and correspondence estimation. Computer-Aided Design 141, 103105 (2021)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H.Jin et al.", + "bbox": [ + 271, + 114, + 348, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Wu, W., Qi, Z., Fuxin, L.: Pointconv: Deep convolutional networks on 3d point clouds. In: CVPR. pp. 9621-9630 (2019)", + "44. Yew, Z.J., Lee, G.H.: 3dfeat-net: Weakly supervised local 3d features for point cloud registration. In: ECCV. pp. 607-623 (2018)", + "45. Yi, L., Su, H., Guo, X., Guibas, L.J.: Syncspeccnn: Synchronized spectral cnn for 3d shape segmentation. In: CVPR. pp. 2282-2290 (2017)", + "46. You, Y., Liu, W., Ze, Y., Li, Y.L., Wang, W., Lu, C.: Ukpgan: A general self-supervised keypoint detector. In: CVPR. pp. 17042-17051 (2022)", + "47. You, Y., Lou, Y., Li, C., Cheng, Z., Li, L., Ma, L., Lu, C., Wang, W.: Keypointnet: A large-scale 3d keypoint dataset aggregated from numerous human annotations. In: CVPR. pp. 13647-13656 (2020)", + "48. Yu, X., Rao, Y., Wang, Z., Liu, Z., Lu, J., Zhou, J.: Pointr: Diverse point cloud completion with geometry-aware transformers. In: ICCV. pp. 12498-12507 (2021)", + "49. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In: CVPR. pp. 19313-19322 (2022)", + "50. Yuan, H., Zhao, C., Fan, S., Jiang, J., Yang, J.: Unsupervised learning of 3d semantic keypoints with mutual reconstruction. In: ECCV. pp. 534-549. Springer (2022)", + "51. Zhang, R., Wang, L., Qiao, Y., Gao, P., Li, H.: Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In: CVPR. pp. 21769-21780 (2023)", + "52. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV. pp. 16259-16268 (2021)", + "53. Zhong, C., You, P., Chen, X., Zhao, H., Sun, F., Zhou, G., Mu, X., Gan, C., Huang, W.: Snake: Shape-aware neural 3d keypoint field. NeurIPS 35, 7052-7064 (2022)", + "54. Zhu, X., Du, D., Huang, H., Ma, C., Han, X.: 3d keypoint estimation using implicit representation learning. arXiv preprint arXiv:2306.11529 (2023)", + "55. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 575 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "KeypointDETR", + "bbox": [ + 622, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/KeypointDETR_ An End-to-End 3D Keypoint Detector/d05c910b-e5c7-4df0-b517-a9b20e7a9b64_model.json b/2024/KeypointDETR_ An End-to-End 3D Keypoint Detector/d05c910b-e5c7-4df0-b517-a9b20e7a9b64_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0346a3ab3b4ead3ada0372c737f2c60dea199cac --- /dev/null +++ b/2024/KeypointDETR_ An End-to-End 3D Keypoint Detector/d05c910b-e5c7-4df0-b517-a9b20e7a9b64_model.json @@ -0,0 +1,2258 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.24, + 0.141, + 0.764, + 0.184 + ], + "angle": 0, + "content": "KeypointDETR: An End-to-End 3D Keypoint Detector" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.212, + 0.756, + 0.245 + ], + "angle": 0, + "content": "Hairong Jin\\(^{1,2\\star}\\), Yuefan Shen\\(^{1}\\), Jianwen Lou\\(^{3}\\), Kun Zhou\\(^{1}\\), and Youyi Zheng\\(^{1(\\boxtimes)}\\)" + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.255, + 0.733, + 0.283 + ], + "angle": 0, + "content": "\\(^{1}\\) State Key Lab of CAD&CG, Zhejiang University, Hangzhou, China hairong_jin, jhonve, kunzhou, youyizheng@zju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.367, + 0.283, + 0.638, + 0.297 + ], + "angle": 0, + "content": "2 Chohotech Co.,Ltd., Hangzhou, China" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.297, + 0.739, + 0.325 + ], + "angle": 0, + "content": "3 School of Software Technology, Zhejiang University, Hangzhou, China jianwen.lou@zju.edu.cn" + }, + { + "type": "list", + "bbox": [ + 0.263, + 0.255, + 0.739, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.351, + 0.74, + 0.698 + ], + "angle": 0, + "content": "Abstract. 3D keypoint detection plays a pivotal role in 3D shape analysis. The majority of prevalent methods depend on producing a shared heatmap. This approach necessitates subsequent post-processing techniques such as clustering or non-maximum suppression (NMS) to pinpoint keypoints within high-confidence regions, resulting in performance inefficiencies. To address this issue, we introduce KeypointDETR, an end-to-end 3D keypoint detection framework. KeypointDETR is predominantly trained with a bipartite matching loss, which compels the network to forecast sets of heatmaps and probabilities for potential keypoints. Each heatmap highlights one keypoint's location, and the associated probability indicates not only the presence of that specific keypoint but also its semantic consistency. Together with the bipartite matching loss, we utilize a transformer-based network architecture, which incorporates both point-wise and query-wise self-attention within the encoder and decoder, respectively. The point-wise encoder leverages the self-attention mechanism on a dynamic graph derived from the local feature space of each point, resulting in the generation of heatmap features. As a key part of our framework, the query-wise decoder facilitates inter-query information exchange. It captures the underlying connections among keypoints' heatmaps, positions, and semantic attributes via the cross-attention mechanism, enabling the prediction of heatmaps and probabilities. Extensive experiments conducted on the KeypointNet dataset reveal that KeypointDETR outperforms competitive baselines, demonstrating superior performance in keypoint saliency and correspondence estimation tasks. (The code will be released at github.com/bibi547/KeypointDETR)" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.71, + 0.6, + 0.724 + ], + "angle": 0, + "content": "Keywords: 3D keypoint \\(\\cdot\\) Point cloud \\(\\cdot\\) Heatmap" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.746, + 0.377, + 0.761 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.773, + 0.788, + 0.819 + ], + "angle": 0, + "content": "The automatic extraction of 3D keypoints presents a significant challenge in the realms of computer vision and computer graphics. These keypoints are instrumental in various geometric processing tasks, such as object detection [20], shape" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.727, + 0.84 + ], + "angle": 0, + "content": "* This work was done during the author's internship at Chohotech Co.,Ltd.." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "H.Jin et al." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.144, + 0.77, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.398 + ], + "angle": 0, + "content": "Fig. 1: Illustration of the differences between the existing methods and KeypointDETR for keypoint saliency and correspondence estimation. Given a 3D model, Keypoint-DETR can predict its multi-heatmaps and probabilities. Heatmaps provide accurate inference of keypoints' positions. Probabilities indicate whether the specific keypoints exist in the 3D model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.784, + 0.461 + ], + "angle": 0, + "content": "segmentation [13, 21], shape matching [39, 44], point registration [1], and point cloud completion [36], etc." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Recent breakthroughs in deep learning, especially in 3D shape analysis and geometric feature extraction [5,11,27,28,40,43], have catalyzed the emergence of a multitude of learning-based approaches for 3D keypoint detection. A straightforward technique [45,47] is point classification on the surface of a 3D shape, but this method grapples with significant sample imbalances. To mitigate this, many existing keypoint detection strategies [32,41,42] utilize Gaussian kernel functions to create a ground truth heatmap. This shared heatmap is generated based on the distances between points and keypoints, guiding the extraction of keypoints from areas identified as salient. However, these techniques typically require post-processing steps such as clustering [29] or non-maximum suppression [22] to sift out high-confidence points and finalize the keypoint extraction. These processes often involve computing distance matrices, which can considerably slow down inference speeds. Additionally, the selection of parameters can profoundly affect the accuracy of keypoint extraction. Many landmark detection methods for 2D or 3D images [4] prefer using multi-heatmaps for keypoints, as opposed to relying on a single shared heatmap. This approach assigns each keypoint its unique heatmap, facilitating the extraction of the point with the highest value from each map. Although their strategy eliminates the need for post-processing, it can limit the algorithm's generalizability. One significant challenge is ensuring a consistent number of keypoints across all 3D models within a category." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To tackle these challenges, we introduce an innovative end-to-end 3D keypoint detection model named KeypointDETR. This approach not only eliminates the need for any post-processing steps but also exhibits remarkable generalization capabilities. As illustrated in Fig. 1, our framework takes point coordinates as input and predicts multi-heatmaps along with their associated probabilities," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.624, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "KeypointDETR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.463 + ], + "angle": 0, + "content": "providing a consistent paradigm for keypoint saliency and correspondence estimation. For keypoint saliency estimation, we can derive binary \\( yes / no \\) labels from the probabilities, while for keypoint correspondence estimation, semantic labels can be inferred. Our pipeline utilizes a loss function based on the Hungarian algorithm [15], which facilitates the optimal bipartite matching between the predictions and the ground truth heatmaps. Consistent with this strategy, we develop a transformer-based encoder-decoder network architecture. The encoder employs dynamic graph transformer modules to extract geometric features from 3D models, specifically focusing on the generation of heatmap features. We regard the learnable heatmap features as intrinsic geometric features, eliminating the need for direct supervision. The dynamic graph transformer module aggregates the features from neighboring points within the feature space and assigns self-attention weights, effectively extending the range of perception by stacking these modules. In the decoder, we introduce a transformer refinement strategy that utilizes the query-wise self-attention mechanism to facilitate information exchange among potential keypoints. Additionally, it incorporates the cross-attention mechanism to capture the latent connections among keypoints' heatmaps, positions, and semantics. The query-wise decoder optimizes the heatmap and query features through the analysis of inter-keypoint and inter-attribute relationships, thereby enabling the parallel prediction of multi-heatmaps and probabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.54 + ], + "angle": 0, + "content": "We validate our method on the KeypointNet dataset [47], achieving outstanding results in two essential tasks: keypoint saliency and correspondence estimation. Furthermore, our approach of assigning a specific heatmap to each keypoint leads to significantly higher accuracy compared to methods that use a shared heatmap." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.541, + 0.633, + 0.555 + ], + "angle": 0, + "content": "Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.57, + 0.785, + 0.616 + ], + "angle": 0, + "content": "- We introduce KeypointDETR, an end-to-end 3D keypoint detection framework. It is trained with a loss function based on the optimal bipartite matching derived from the cost matrix of multi-heatmaps and probabilities." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.617, + 0.785, + 0.675 + ], + "angle": 0, + "content": "- We construct a novel transformer-based network architecture. The point-wise encoder harnesses self-attention on the dynamic graph of points' local feature space to extract geometric features. The query-wise decoder incorporates global self-attention to enhance the interaction between queries." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.678, + 0.785, + 0.722 + ], + "angle": 0, + "content": "- We conduct a comprehensive evaluation that validates our method and provides insights into the impact of principal components, such as the bipartite matching loss and query-wise self-attention." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.57, + 0.785, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.748, + 0.388, + 0.765 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.783, + 0.455, + 0.798 + ], + "angle": 0, + "content": "2.1 3D Keypoint Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In previous studies on 3D keypoint detection [23,33,37], algorithms heavily rely on a variety of hand-crafted geometric features, such as ISS [30], HKS [34]," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "H.Jin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.192 + ], + "angle": 0, + "content": "CGF [14], Salient Points [3], and Mesh Saliency [16]. These traditional approaches primarily focus on local geometric features of 3D shapes, often overlooking the integration of global and semantic information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.194, + 0.788, + 0.542 + ], + "angle": 0, + "content": "With advancements in deep learning, there has been a notable surge in the development of data-driven algorithms for keypoint detection [10, 35, 54]. Numerous unsupervised 3D keypoint detection algorithms [1, 7, 12, 18, 31, 50, 53] are proposed for specific tasks such as point registration and reconstruction. For example, UKPGAN [46] utilizes a GAN-based module to manage keypoint sparsity and reconstructs 3D models by distilling saliency information. Unsupervised methods tend to focus on identifying keypoints in areas with distinct geometric features. While some of these methods are capable of attaining semantic consistency, they typically require the predetermination of the number of keypoints. Furthermore, in many scenarios, it is essential to identify keypoints that may not have prominent geometric features but possess solid semantic importance. Methods like SyncSpecCNN [45] and PRA-Net [5] address keypoint detection by classifying points on 3D shapes. These methods often result in an imbalance between positive and negative samples, presenting a significant challenge. Consequently, current popular methods employ a heatmap-based approach [32, 41], where the value assigned to each point indicates its proximity to the nearest keypoint. Wei et al. [42] proposes a multi-task learning framework that combines point-to-keypoint offsets with a confidence map, facilitating effective 3D keypoint saliency and correspondence estimation. Heatmap-based methods for 3D shapes predominantly depend on a shared heatmap for all keypoints and require post-processing. These methods employ post-processing techniques like non-maximum suppression [22] or clustering [29] to isolate keypoints from regions with high scores during the inference process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.568, + 0.444, + 0.582 + ], + "angle": 0, + "content": "2.2 Transformer in Vision" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Transformer [38] was initially introduced in the realm of natural language processing. Recent advancements in visual research [6, 19] have explored various models centered around self-attention [38], achieving unprecedented performance. DETR [2] pioneers an innovative end-to-end object detection approach. It formulates a loss matching rule employing the Hungarian algorithm [15], facilitating bipartite matching between predictions and ground truth. This method serves as the cornerstone for a series of object detection methods [4,17,55] based on set prediction. Li et al. [17] develop a facial landmark detection method using cascaded transformer decoders. They devise parallel decoders utilizing deformable self-attention [55] to optimize facial landmark coordinates through offset predictions. Chen et al. [4] leverage self-attention to create a structure-aware LSTM framework tailored for predicting 3D heatmaps in CBCT. Keypoint transformer [9] efficiently extracts all potential keypoints within the image from a shared heatmap, subsequently employing the DETR decoder to predict 3D hand poses. Given the transformer's inherent capacity to handle input element permutations in sequences, it is exceptionally well-suited for point cloud processing [8, 24, 26]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.624, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "KeypointDETR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.788, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.476 + ], + "angle": 0, + "content": "Fig. 2: Pipeline of KeypointDETR for keypoint saliency and correspondence estimation. Our framework takes point coordinates as input. The point-wise encoder first extracts geometric features from the point cloud by stacking dynamic graph transformer modules (DGT). Then, the query-wise decoder accepts heatmap features, which are transformed into heatmap and query features to generate multi-heatmaps and probabilities. These predictions enable the direct inference of both keypoints positions and semantic labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.788, + 0.64 + ], + "angle": 0, + "content": "Nonetheless, applying global self-attention to 3D point cloud models incurs substantial computational costs. Consequently, existing research [48, 52] frequently adopts local self-attention mechanisms, expanding receptive fields through sampling and grouping strategies. In contrast, our point-wise encoder integrates local self-attention within the dynamic graph of the point feature space, thereby eliminating the necessity for sampling and grouping operations. Distinct from most existing DETR style works, our query-wise decoder is meticulously engineered to predict multi-heatmaps and probabilities, enabling efficient end-to-end 3D keypoint detection." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.66, + 0.331, + 0.675 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this section, we delve into the specifics of our end-to-end 3D keypoint detector, KeypointDETR, which consists of two essential components: a network architecture based on the transformer and a loss function formulated using bipartite matching. As illustrated in Fig. 2, our approach is applied to tasks involving keypoint saliency and correspondence estimation. Upon receiving a 3D point cloud as input, our point-wise encoder utilizes the dynamic graph transformer modules to extract heatmap features, which can be considered intrinsic geometric features. We then deploy a query-wise transformer decoder designed to enhance information exchange between queries and capture the latent relationships between attributes through cross-attention, thereby optimizing heatmap" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "H.Jin et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.147, + 0.538, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.252, + 0.541, + 0.274 + ], + "angle": 0, + "content": "(a) The structure of the dynamic graph transformer module." + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.144, + 0.761, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.253, + 0.786, + 0.274 + ], + "angle": 0, + "content": "(b) The structure of the transformer refinement module." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.286, + 0.784, + 0.314 + ], + "angle": 0, + "content": "Fig. 3: The structure of the dynamic graph transformer module (DGT) and transformer refinement module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.435 + ], + "angle": 0, + "content": "and query features. This network yields heatmaps and probabilities as its outputs. The prediction sets are aligned with ground truth heatmaps through a cost matrix, leading to the computation of the final loss based on optimal bipartite matching. During the inference stage, we select heatmaps corresponding to keypoints, as indicated by the predicted probabilities, and extract keypoints exhibiting the highest heat values from these maps." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.457, + 0.409, + 0.473 + ], + "angle": 0, + "content": "3.1 Data Preparation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.788, + 0.558 + ], + "angle": 0, + "content": "To tackle the issue of imbalanced samples concerning keypoints on the surface of 3D models, heatmap-based 3D keypoint detection methods establish the ground truth heat value for each point based on its distances to the keypoints. Specifically, the heatmap is generated by applying a Gaussian kernel in conjunction with the geodesic distance matrix:" + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.568, + 0.786, + 0.589 + ], + "angle": 0, + "content": "\\[\nh _ {i, j} = \\exp^ {- \\frac {1}{2 \\sigma^ {2}} \\mathrm {d} \\left(p _ {i}, k _ {j}\\right)}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.601, + 0.788, + 0.71 + ], + "angle": 0, + "content": "Given a 3D point cloud with \\( N_{p} \\) points \\( P = \\{p_{i}\\}_{i=1}^{N_{p}} \\) and \\( N_{k} \\) keypoints \\( K = \\{k_{j}\\}_{j=1}^{N_{k}} \\), our process begins with the computation of the geodesic distance matrix from each point to the keypoints. Then, we derive the heatmap matrix \\( \\mathcal{H} = \\{h_{i,j}\\} \\in \\mathbb{R}^{N_{p} \\times N_{k}} \\) using the Gaussian kernel function. Contrasting with methods that use a shared heatmap, where the maximum value for each point is considered, our approach distinctively employs multi-heatmaps, each corresponding to specific keypoints, as the ground truth." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.732, + 0.445, + 0.746 + ], + "angle": 0, + "content": "3.2 Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.757, + 0.787, + 0.789 + ], + "angle": 0, + "content": "As shown in Fig. 2, our transformer-based encoder-decoder network architecture primarily comprises a point-wise encoder and a query-wise decoder." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Point-wise Encoder The main objective of the encoder is to extract geometric features from the 3D point cloud. Drawing inspiration from DGCNN [40] and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.624, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "KeypointDETR" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "the Point Transformer [52], we devise a point-wise encoder constructed using dynamic graph transformer modules (DGT)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.254 + ], + "angle": 0, + "content": "As depicted in Fig. 3 (a), we employ local self-attention within the dynamic graph transformer module to operate on the feature space. To be specific, for the feature map \\(\\mathbf{F}\\) at each layer, we extract a \\(k\\)-nearest neighbors feature set \\(\\mathbf{F}_i = \\{\\mathbf{f}_j\\}_{j=1}^k\\) for \\(\\mathbf{f}_i\\). Thereafter, we assign self-attention weights to these \\(k\\)-nn features and aggregate them using the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.265, + 0.786, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {i} ^ {\\prime} = \\sum_ {\\mathbf {f} _ {j} \\in \\mathbf {F} _ {i}} \\rho (\\theta (\\phi (\\mathbf {f} _ {i}) - \\varphi (\\mathbf {f} _ {j}) + \\delta)) \\odot (\\psi (\\mathbf {f} _ {j}) + \\delta). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.309, + 0.784, + 0.384 + ], + "angle": 0, + "content": "Here, \\(\\theta\\) is implemented as an MLP, while \\(\\phi\\), \\(\\varphi\\), \\(\\psi\\) are realized using Conv1d. \\(\\rho\\) represents the softmax operation applied to self-attention weights, and \\(\\odot\\) denotes the element-wise product operation. The term \\(\\delta\\) represents the position encoding, which is generated from the point coordinates \\(P_{i} = \\{p_{j}\\}_{j=1}^{k}\\) corresponding to the neighborhood features, as expressed in the following:" + }, + { + "type": "equation", + "bbox": [ + 0.436, + 0.398, + 0.785, + 0.414 + ], + "angle": 0, + "content": "\\[\n\\delta = \\operatorname {M L P} \\left(p _ {i} - p _ {j}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.425, + 0.784, + 0.47 + ], + "angle": 0, + "content": "Finally, we concatenate the features extracted from each dynamic graph transformer module to obtain point-wise features, which are decoded into heatmap features through MLPs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.785, + 0.627 + ], + "angle": 0, + "content": "Query-wise Decoder The decoder's output comprises a predefined number of predictions, referred to as keypoint queries, with the query count \\( M \\) set significantly higher than the number of keypoints within a 3D model. Each prediction query consists of both a heatmap and a probability. Heatmaps serve to indicate the position of keypoints within the 3D model. Probabilities reflect the semantic consistency of potential keypoints. In keypoint saliency estimation, they determine the presence or absence of specific keypoints. In keypoint correspondence estimation, such probabilities are correlated with their respective semantic labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.786, + 0.841 + ], + "angle": 0, + "content": "Guided by these objectives, we develop a query-wise decoder that adheres to the proven DETR-style architecture [2], as illustrated in Fig. 3 (b). Initially, we take the heatmap features \\(\\mathbf{F}_{heat} \\in \\mathbb{R}^{N \\times M}\\) from the encoder. Considering the inherent relationship between the heatmaps and semantics of keypoints, the heatmap features are directly transposed and then decoded into query features \\(\\mathbf{F}_{prob} \\in \\mathbb{R}^{M \\times 128}\\) through an MLP. Following this, both the heatmap \\(\\mathbf{F}_{heat}\\) and query \\(\\mathbf{F}_{prob}\\) features are fed into the transformer decoder. To facilitate the exchange of information among queries, a multi-head self-attention module is applied to \\(\\mathbf{F}_{prob}\\). For the position encoding, we extract the coordinates of the reference points corresponding to the peaks of heat values in \\(\\mathbf{F}_{heat}\\), which are subsequently processed through an MLP to produce the position encoding. The transposed \\(\\mathbf{F}_{heat}\\) and \\(\\mathbf{F}_{prob}\\) serve as inputs to the cross-attention module, enhancing the interplay of information between probability and heatmap features. Ultimately, through the application of Feed-Forward Network modules (FFN)," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "H.Jin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "the refined \\(\\mathbf{F}_{heat}\\) and \\(\\mathbf{F}_{prob}\\) are decoded into multi-heatmaps \\(\\mathbf{M}_h\\) and probabilities \\(\\mathbf{M}_p\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.299 + ], + "angle": 0, + "content": "The decoder utilizes a query-wise self-attention mechanism across the queries, leading to \\( M \\) distinct predictions. This mechanism operates as a global self-attention centered around potential keypoints. Such a globally parallel decoding approach efficiently captures the interactions among the predicted queries, thereby optimizing heat values and reducing redundancy in the identification of keypoints. Additionally, considering the potential relationship between keypoint heatmaps and semantics, cross-attention is employed to facilitate the interaction between heatmap and query features." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.319, + 0.379, + 0.333 + ], + "angle": 0, + "content": "3.3 Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.447 + ], + "angle": 0, + "content": "Inspired by DETR [2], we integrate the bipartite matching loss into 3D keypoint detection to enable end-to-end inference, eliminating the need for postprocessing. During the bipartite matching phase, we calculate the pairwise matching costs by assessing the discrepancies between the predicted queries and ground truth. We utilize this cost matrix to consolidate the prediction sets, distinguishing between \"background\" queries and keypoint queries, in order to compute the final loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.789, + 0.679 + ], + "angle": 0, + "content": "Given a 3D model with \\( N_{k} \\) ground truth keypoints, KeypointDETR generates a set of \\( M \\) predictions, where \\( M \\) is preconfigured to be significantly larger than \\( N_{k} \\). The ground truth can be represented as \\( \\{y_{i}\\}_{i = 1}^{N_{k}} = \\{h_{i},p_{i}\\}_{i = 1}^{N_{k}} \\), where \\( h_i \\) signifies the ground truth heatmap and \\( p_i \\) denotes the ground truth probability. The predicted queries are denoted as \\( \\{\\hat{y}_j\\}_{j = 1}^M = \\{\\hat{h_j},\\hat{p_j}\\}_{j = 1}^M \\), with \\( \\hat{h_j} \\) representing the predicted heatmap, and \\( \\hat{p_j} \\) corresponding to the predicted probabilities associated with keypoints categories \\( \\{c_i\\}_{i = 1}^{N_k} \\). In the task of keypoint saliency estimation, \\( p_i \\) remains fixed at 1, while \\( \\hat{p_j}(c_i) \\) signifies the probability of being predicted as a keypoint. However, in the task of keypoint correspondence estimation, \\( \\hat{p_j}(c_i) \\) is defined as the probability of \\( c_i \\), where \\( c_i \\) denotes the semantic label corresponding to the ground truth keypoint \\( k_i \\). Expanding upon these notations, in order to achieve the optimal bipartite matching between ground truth and predictions, it is imperative to establish a matching cost matrix \\( \\mathcal{M}_c \\in \\mathbb{R}^{N_k \\times M} \\). This cost matrix takes into account both the heatmaps and probabilities. Precisely, each element \\( \\mathcal{M}_c(i,j) \\) of the cost matrix can be calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.689, + 0.787, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {c} (i, j) = \\mathbf {W} _ {\\mathbf {h}} \\cdot \\sum | h _ {i} - \\hat {h} _ {j} | + \\mathbf {W} _ {\\mathbf {p}} \\cdot \\left(1 - \\log p _ {\\hat {j}} \\left(c _ {i}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.716, + 0.785, + 0.765 + ], + "angle": 0, + "content": "where \\(\\mathbf{W_h}\\) represents the weight assigned to heatmaps, and \\(\\mathbf{W_p}\\) is the weight attributed to probabilities. \\(\\sum |h_i - \\hat{h_j}|\\) signifies the dissimilarity between the ground truth \\(h_i\\) and predicted \\(\\hat{h_j}\\) heatmaps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Following the application of the Hungarian algorithm [15], we obtain the optimal matching and categorize the predicted queries into two groups: positive queries and negative queries. Positive queries are those that align with the ground truth. For a positive query \\(\\hat{y}_i\\), we determine the matched ground truth category as \\(g_i\\) and the corresponding heatmap as \\(h_i\\). Then, we compute the probability" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.623, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "KeypointDETR" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "loss using the Focal Loss (FL) and the heatmap loss employing the Mean Squared Error Loss (MSE). For a negative query, only the probability loss is considered. The associated ground truth \\( g_{i} \\) is designated as 0, signifying nonexistence. Thus, the overall loss can be calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.215, + 0.786, + 0.253 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {i = 1} ^ {M} \\left(\\lambda_ {p} \\operatorname {F L} \\left(g _ {i}, \\hat {p _ {i}}\\right) + \\lambda_ {h} \\mathbb {1} _ {\\{g _ {i} \\neq 0 \\}} \\operatorname {M S E} \\left(h _ {i}, \\hat {h _ {i}}\\right)\\right). \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.273, + 0.376, + 0.29 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.303, + 0.784, + 0.362 + ], + "angle": 0, + "content": "In this section, we conduct a series of comprehensive experiments to validate the superior performance of our approach in both keypoint saliency and correspondence estimation tasks. We provide comparisons with other state-of-the-art methods to demonstrate our method's efficacy." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.383, + 0.457, + 0.398 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.785, + 0.556 + ], + "angle": 0, + "content": "KeypointDETR takes 3D point coordinates as input. In the point-wise encoder, we utilize 5 dynamic graph transformer modules, each with point features of 64 dimensions, the \\(k\\) value of \\(k\\)-nearest neighbors set to 20, and a global embedding dimension of 1024. In the query-wise decoder, we set the number of queries \\(M\\) to 50. The decoder's output comprises two branches: the heatmap head generates multi-heatmaps \\(\\mathbf{M}_h \\in \\mathbb{R}^{M \\times N_p}\\), and the classification head provides probabilities \\(\\mathbf{M}_p \\in \\mathbb{R}^{M \\times C}\\). For keypoint saliency estimation, \\(C\\) is held constant at 2. However, for keypoint correspondence estimation, \\(C\\) is configured as \\((N_c + 1)\\), where \\(N_c\\) signifies the number of keypoint categories. More details can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.577, + 0.327, + 0.59 + ], + "angle": 0, + "content": "4.2 Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.784, + 0.705 + ], + "angle": 0, + "content": "Our experiments are conducted on the KeypointNet dataset [47], which is designed for both keypoint saliency and correspondence estimation tasks. We focus on four classical categories: airplane, chair, guitar, and table. These models are divided into training, validation, and test sets, with a distribution ratio of 7:1:2 for each category. We use 2,048 sampled points from each 3D model as input data. The results of other categories and applications of KeypointDETR can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.726, + 0.327, + 0.74 + ], + "angle": 0, + "content": "4.3 Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.784, + 0.807 + ], + "angle": 0, + "content": "We evaluate our method alongside comparison methods on the keypoint saliency estimation using mIoU and Chamfer Distance (CD) while utilizing mIoU and Dual Alignment Score (DAS) [31] for keypoint correspondence estimation. The IoU is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.814, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\nI o U = \\frac {T P}{T P + F P + F N}, \\tag {6}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "H.Jin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.193 + ], + "angle": 0, + "content": "where \\( TP \\) stands for true positives, \\( FP \\) represents false positives, and \\( FN \\) corresponds to false negatives. Given the ground truth and predicted keypoints sets \\( G \\) and \\( K \\), CD can be computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.306, + 0.205, + 0.786, + 0.242 + ], + "angle": 0, + "content": "\\[\nC D = \\frac {1}{| G |} \\sum_ {p \\in G} \\min _ {q \\in K} \\| p - q \\| _ {2} ^ {2} + \\frac {1}{| K |} \\sum_ {p \\in K} \\min _ {q \\in G} \\| p - q \\| _ {2} ^ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.788, + 0.452 + ], + "angle": 0, + "content": "The settings for clustering and NMS parameters profoundly affect the repeatability and omission rates of predicted keypoints. However, the mIoU metric used in prior methods fails to effectively capture these erroneous cases. To ensure a comprehensive comparison, we design a mIoU metric based on the Hungarian algorithm [15]. For the keypoint saliency estimation, we begin by calculating the Euclidean distance matrix between the ground truth and the predicted keypoints. Subsequently, the Hungarian algorithm is employed to determine their optimal bipartite matching. Each matched distance falling below the predefined threshold \\(\\epsilon\\) is considered a true positive. For keypoint correspondence estimation, a ground truth keypoint is considered a true positive if there is a predicted keypoint with the same semantic label whose distance to it is less than the threshold \\(\\epsilon\\). Additionally, we further measure whether the closest keypoints between predictions and ground truth have the same semantic labels using DAS [31]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.478, + 0.638, + 0.494 + ], + "angle": 0, + "content": "4.4 Comparison on Keypoint Saliency Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.813, + 0.628 + ], + "angle": 0, + "content": "We conduct a comparative analysis of our method against six advanced algorithms, including two unsupervised keypoint detection methods (Skeleton Merger [31] and UKPGAN [46]) and three point cloud geometric feature extraction algorithms (PointNet++ [28], PRA-Net [5], and DGCNN [40]), and a fully-supervised keypoint detection method (Wei et al. [42]). For the three point cloud geometric feature extraction algorithms on saliency analysis, the network's output is a shared heatmap, wherein each point is assigned the maximum value of the multi-heatmaps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The quantitative results for mIoU and Chamfer Distance (CD) are presented in Tab. 1, where we compare our KeypointDETR with other algorithms using a threshold \\(\\epsilon\\) of 0.1. Fig. 6 (a) depicts the mIoU curves for various methods as \\(\\epsilon\\) varies from 0 to 0.1. It is evident that KeypointDETR consistently outperforms other methods in the Hungarian mIoU evaluation across all categories. This significant advantage stems from our method's capacity to bypass post-processing techniques like NMS or clustering algorithms, leading to direct 'one-to-one' predictions that effectively tackle redundancy and omissions in 3D keypoint detection. In contrast, methods reliant on a shared heatmap face challenges with parameter selection of post-processing, which predominantly depends on prior knowledge factors such as the expected number of keypoints and the 3D model's scale. Unfortunately, even optimal parameter settings for NMS (radius=0.1) in these methods do not entirely resolve issues of omissions and redundant predictions, as further illustrated in Fig. 4." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.624, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "KeypointDETR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 1: Quantitative results of keypoint saliency estimation using Hungarian mIoU (\\%) and CD." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.185, + 0.797, + 0.34 + ], + "angle": 0, + "content": "
AirplaneChairGuitarTableAverage
mIoUCDmIoUCDmIoUCDmIoUCDmIoUCD
SkeletonMerger [31]60.730.10742.740.18947.710.13227.840.30244.750.182
UKPGAN [46]62.630.11825.160.16854.250.13117.780.24039.950.164
PRA-Net [5]82.880.04970.440.06073.710.05369.690.05374.180.053
PointNet++ [28]82.310.05057.500.07772.340.06060.010.06268.040.062
DGCNN [40]78.490.05859.490.07663.880.07259.920.06665.440.068
Wei et al. [42]79.860.04886.280.05578.080.06686.610.03982.700.052
Ours91.370.04591.890.05593.960.04598.050.02793.810.043
" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.349, + 0.788, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.562, + 0.71, + 0.576 + ], + "angle": 0, + "content": "Fig. 4: Visualization examples of keypoint saliency estimation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.605, + 0.702, + 0.62 + ], + "angle": 0, + "content": "4.5 Comparison on Keypoint Correspondence Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.786, + 0.764 + ], + "angle": 0, + "content": "Keypoint correspondence estimation represents a more significant challenge compared to keypoint saliency estimation, as it not only involves locating keypoints but also predicting their semantic labels. Additionally, 3D models within the same category may not have identical semantic keypoints. Prior approaches [35, 45,47] typically involve direct classification of points within a point cloud, which does not effectively address the issue of imbalanced positive and negative samples in keypoint detection. Some methods [42, 54] attempt to add a semantic branch to keypoint saliency estimation, aiming to assign labels to keypoints in a two-stage process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "As depicted in Tab. 2, we conduct a comparison of our method with five methods using a threshold \\(\\epsilon\\) of 0.1: PointConv [43], PointNet++ [28], DGCNN [40], PRA-Net [5], and Wei et al. [42]. Additionally, we present the mIoU curves for various methods, with the \\(\\epsilon\\) ranging from 0 to 0.1, in Fig. 6 (b). Fig. 5 showcases visualization results of different methods alongside the ground truth. Notably," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "H.Jin et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 2: Quantitative results of keypoint correspondence estimation using Hungarian mIoU (%) and DAS [31] (%)." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.185, + 0.784, + 0.324 + ], + "angle": 0, + "content": "
AirplaneChairGuitarTableAverage
mIoUDASmIoUDASmIoUDASmIoUDASmIoUDAS
DGCNN [40]77.7782.6963.1679.5661.1378.5957.9978.8365.0179.91
PointNet++ [28]79.1084.7458.2478.1967.6981.9262.9781.5967.0081.61
PointConv [43]81.5485.7464.9282.1576.3988.9166.2083.7372.2685.13
PRA-Net [5]81.1085.8068.3583.6272.0785.1570.1684.8972.9284.86
Wei et al. [42]78.5488.3479.7790.4475.6493.6986.5697.3180.1292.44
Ours85.7191.7479.3389.5390.9393.8992.4097.7987.0993.23
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.334, + 0.788, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.52, + 0.735, + 0.534 + ], + "angle": 0, + "content": "Fig. 5: Visualization examples of keypoint correspondence estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.785, + 0.623 + ], + "angle": 0, + "content": "our method generates no redundant points with the same label in close proximity to predicted keypoints. This demonstrates that our method adeptly accomplishes semantic correspondence for keypoints, with KeypointDETR showcasing superior performance in both the heatmap and probability branch." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.644, + 0.391, + 0.659 + ], + "angle": 0, + "content": "4.6 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.784, + 0.743 + ], + "angle": 0, + "content": "We conduct a series of ablation experiments to validate the performance of KeypointDETR and deeply explore the impact of each component. Throughout these experiments, we consistently use four representative categories from the KeypointNet dataset for keypoint saliency estimation. The evaluation adopts mIoU with a threshold \\(\\epsilon\\) of 0.1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Effects of point-wise encoder. Our point-wise encoder primarily consists of dynamic graph transformer modules that extract geometric features from the point cloud. To validate the effectiveness of our point-wise encoder, we compare KeypointDETR with its variants, which utilize different backbones for geometric feature extraction, namely DGCNN [40], PointNet++ [28], and Point Transformer [52]. The DGCNN [40] backbone (DG.Enc.) is implemented as a five-layer" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.624, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "KeypointDETR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.457, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.287, + 0.475, + 0.309 + ], + "angle": 0, + "content": "(a) Hungarian mIoU curves for keypoint saliency estimation." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.144, + 0.768, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.287, + 0.785, + 0.309 + ], + "angle": 0, + "content": "(b) Hungarian mIoU curves for keypoint correspondence estimation." + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.32, + 0.785, + 0.349 + ], + "angle": 0, + "content": "Fig. 6: Hungarian mIoU curves for keypoint saliency and correspondence estimation under various distance thresholds (0-0.1)." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.363, + 0.785, + 0.391 + ], + "angle": 0, + "content": "Table 3: Quantitative results of various variations in the ablation study using Hungarian mIoU (%) and CD." + }, + { + "type": "table", + "bbox": [ + 0.244, + 0.404, + 0.761, + 0.578 + ], + "angle": 0, + "content": "
AirplaneChairGuitarTableAverage
mIoUCDmIoUCDmIoUCDmIoUCDmIoUCD
DG.Enc. [40]86.750.04882.120.06292.760.05297.620.03789.810.049
PN2.Enc. [28]76.300.07168.310.10172.090.09976.420.09973.280.092
PT.Enc. [52]82.550.05973.040.08568.920.09197.320.03280.450.066
Point-BERT Enc. [49]91.110.03493.570.03493.510.03293.710.03392.880.033
I2P-MAE Enc. [51]79.060.05080.400.06788.790.04886.050.04883.570.053
Point-MAE Enc. [25]89.820.03993.980.05583.790.04395.950.04390.880.045
w/o Trans.Dec.87.960.04488.230.05892.740.04597.770.02991.670.044
w/o PE.88.050.04291.430.05492.440.04797.970.02892.470.042
Coord-based66.720.10844.550.19037.990.16450.080.18749.830.162
Coord+offsets68.400.09463.620.13642.450.15160.770.15958.810.135
Ours91.370.04591.890.05593.960.04598.050.02793.810.043
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.606, + 0.785, + 0.681 + ], + "angle": 0, + "content": "stack of EdgeConv. PointNet++ [28] (PN2.Enc.) and Point Transformer [52] (PT.Enc.) employ eight layers due to their U-net structures. Furthermore, we explore replacing our encoder with pre-trained models, including Point-BERT, Point-MAE, and I2P-MAE, and achieve favorable results. The overall quantitative comparison results are presented in the first sub-table of Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Effects of query-wise decoder. In the decoder phase, we develop a query-wise decoder utilizing self-attention to facilitate enhanced information exchange among queries. This decoder is tasked with decoding the heatmap and probability feature maps, resulting in the generation of multi-heatmaps and probabilities. A straightforward approach involves employing MLPs to decode these feature maps separately. However, to demonstrate the efficacy and reasoning behind our query-wise decoder design, we conduct an experiment where we replace our self-attention decoder (Trans.Dec.) with MLPs. The comparison results, as shown in the second sub-table of Tab. 3, indicate that the comprehensive performance of our KeypointDETR significantly surpasses that of its MLPs decoder variant." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "H.Jin et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.146, + 0.787, + 0.253 + ], + "angle": 0, + "content": "Effects of position encoding in decoder. Position encoding holds a pivotal role in the self-attention mechanism and is a key factor that makes the transformer architecture suitable for 3D point cloud models, as previously validated in the Point Transformer [52]. In our investigation, we further explore the role of position encoding in our query-wise decoder, which is generated from heatmap features and point coordinates. We compare the results with and without position encoding (PE.) in the third sub-table of Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.264, + 0.787, + 0.476 + ], + "angle": 0, + "content": "Coordinate-based or Heatmap-based. We further explore a coordinate-based architecture within our KeypointDETR approach. In this variant, KeypointDETR predicts the coordinates and probabilities of keypoints, and we subsequently select the closest matching points from the point cloud as the final keypoints. To enhance the accuracy of these coordinates, we incorporate a cascade module designed to predict offsets. Comparative analysis of the heatmap-based KeypointDETR and its coordinate-based variant, as shown in the fourth sub-table of Tab. 3, reveals that while the addition of offsets improves the results of the coordinate-based model, there remains a notable performance gap when compared to our primary heatmap-based KeypointDETR. This discrepancy arises because our geometric feature extraction backbone is better suited for processing 3D shapes rather than 3D spatial coordinates. Furthermore, it validates that the heatmap-based approach can achieve better performance on keypoints detection." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.509, + 0.359, + 0.525 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.552, + 0.787, + 0.719 + ], + "angle": 0, + "content": "In this paper, we introduce KeypointDETR, an end-to-end 3D keypoint detector that obviates the need for post-processing steps, such as clustering and NMS, commonly required in heatmap-based keypoint detection methods. The training strategy of KeypointDETR involves computing a cost matrix between the predictions and ground truth, deriving an optimal bipartite matching loss. This innovative approach enables KeypointDETR to accurately predict multi-heatmaps and probabilities, which can provide the positions and semantic consistency of potential keypoints, respectively. Complementing this framework is a transformer-based network architecture. Extensive quantitative experiments and analyses substantiate the effectiveness and rationality of KeypointDETR, underlining its potential as a state-of-the-art solution in keypoint detection." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.752, + 0.405, + 0.77 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.795, + 0.787, + 0.841 + ], + "angle": 0, + "content": "This work is partially supported by the following grants: National Natural Science Foundation of China (No. 62172363), Management Center of the School of Software (Ningbo) at Zhejiang University under Grant (No. Z24001)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.624, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "KeypointDETR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.786, + 0.218 + ], + "angle": 0, + "content": "1. Bai, X., Luo, Z., Zhou, L., Fu, H., Quan, L., Tai, C.L.: D3feat: Joint learning of dense detection and description of 3d local features. In: CVPR. pp. 6359-6367 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.786, + 0.247 + ], + "angle": 0, + "content": "2. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV. pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.786, + 0.287 + ], + "angle": 0, + "content": "3. Castellani, U., Cristani, M., Fantoni, S., Murino, V.: Sparse points matching by combining 3d mesh saliency with statistical descriptors. In: Comput. Graph. Forum. vol. 27, pp. 643-652. Wiley Online Library (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.786, + 0.328 + ], + "angle": 0, + "content": "4. Chen, R., Ma, Y., Chen, N., Liu, L., Cui, Z., Lin, Y., Wang, W.: Structure-aware long short-term memory network for 3d cephalometric landmark detection. IEEE Transactions on Medical Imaging 41(7), 1791-1801 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.786, + 0.356 + ], + "angle": 0, + "content": "5. Cheng, S., Chen, X., He, X., Liu, Z., Bai, X.: Pra-net: Point relation-aware network for 3d point cloud analysis. IEEE TIP 30, 4436-4448 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.357, + 0.786, + 0.412 + ], + "angle": 0, + "content": "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.786, + 0.453 + ], + "angle": 0, + "content": "7. Fernandez-Labrador, C., Chhatkuli, A., Paudel, D.P., Guerrero, J.J., Demonceaux, C., Gool, L.V.: Unsupervised learning of category-specific symmetric 3d keypoints from point sets. In: ECCV. pp. 546-563. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.453, + 0.786, + 0.481 + ], + "angle": 0, + "content": "8. Guo, M.H., Cai, J.X., Liu, Z.N., Mu, T.J., Martin, R.R., Hu, S.M.: Pct: Point cloud transformer. Computational Visual Media 7, 187-199 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.786, + 0.522 + ], + "angle": 0, + "content": "9. Hampali, S., Sarkar, S.D., Rad, M., Lepetit, V.: Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In: CVPR. pp. 11090-11100 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.786, + 0.564 + ], + "angle": 0, + "content": "0. He, Y., Sun, W., Huang, H., Liu, J., Fan, H., Sun, J.: Pvn3d: A deep point-wise 3d keypoints voting network for 6 dof pose estimation. In: CVPR. pp. 11632-11641 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.786, + 0.592 + ], + "angle": 0, + "content": "1. Huang, Q., Wang, W., Neumann, U.: Recurrent slice networks for 3d segmentation of point clouds. In: CVPR. pp. 2626-2635 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.786, + 0.633 + ], + "angle": 0, + "content": "2. Jakab, T., Tucker, R., Makadia, A., Wu, J., Snavely, N., Kanazawa, A.: Keypoint-deformer: Unsupervised 3d keypoint discovery for shape control. In: CVPR. pp. 12783-12792 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.786, + 0.661 + ], + "angle": 0, + "content": "3. Katz, S., Leifman, G., Tal, A.: Mesh segmentation using feature point and core extraction. The Vis. Comput. 21, 649-658 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.786, + 0.688 + ], + "angle": 0, + "content": "4. Khoury, M., Zhou, Q.Y., Koltun, V.: Learning compact geometric features. In: ICCV. pp. 153-161 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.786, + 0.716 + ], + "angle": 0, + "content": "5. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.786, + 0.744 + ], + "angle": 0, + "content": "6. Lee, C.H., Varshney, A., Jacobs, D.W.: Mesh saliency. In: ACM SIGGRAPH 2005 Papers, pp. 659-666 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.786, + 0.771 + ], + "angle": 0, + "content": "7. Li, H., Guo, Z., Rhee, S.M., Han, S., Han, J.J.: Towards accurate facial landmark detection via cascaded transformers. In: CVPR. pp. 4176-4185 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.786, + 0.799 + ], + "angle": 0, + "content": "8. Li, J., Lee, G.H.: Usip: Unsupervised stable interest point detection from 3d point clouds. In: ICCV. pp. 361-370 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "9. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV. pp. 10012-10022 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.349, + 0.127 + ], + "angle": 0, + "content": "H.Jin et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "20. Liu, Z., Wu, Z., Toth, R.: Smoke: Single-stage monocular 3d object detection via keypoint estimation. In: CVPR. pp. 996-997 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "21. Mian, A.S., Bennamoun, M., Owens, R.: Three-dimensional model-based object recognition and segmentation in cluttered scenes. IEEE TPAMI 28(10), 1584-1601 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.245 + ], + "angle": 0, + "content": "22. Neubeck, A., Van Gool, L.: Efficient non-maximum suppression. In: ICPR. vol. 3, pp. 850-855. IEEE (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.273 + ], + "angle": 0, + "content": "23. Novatnack, J., Nishino, K.: Scale-dependent 3d geometric features. In: ICCV. pp. 1-8. IEEE (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "24. Pan, X., Xia, Z., Song, S., Li, L.E., Huang, G.: 3d object detection with pointformer. In: CVPR. pp. 7463-7472 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.3, + 0.785, + 0.328 + ], + "angle": 0, + "content": "25. Pang, Y., Wang, W., Tay, F.E., Liu, W., Tian, Y., Yuan, L.: Masked autoencoders for point cloud self-supervised learning. In: ECCV. pp. 604-621. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "26. Park, C., Jeong, Y., Cho, M., Park, J.: Fast point transformer. In: CVPR. pp. 16949-16958 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.356, + 0.785, + 0.384 + ], + "angle": 0, + "content": "27. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR. pp. 652-660 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "28. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.412, + 0.785, + 0.438 + ], + "angle": 0, + "content": "29. Rodriguez, A., Laio, A.: Clustering by fast search and find of density peaks. science 344(6191), 1492-1496 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "30. Salti, S., Tombari, F., Spezialietti, R., Di Stefano, L.: Learning a descriptor-specific 3d keypoint detector. In: ICCV. pp. 2318-2326 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "31. Shi, R., Xue, Z., You, Y., Lu, C.: Skeleton merger: an unsupervised aligned keypoint detector. In: CVPR. pp. 43-52 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "32. Shu, Z., Yu, J., Chao, K., Xin, S., Liu, L.: A multi-modal attention-based approach for points of interest detection on 3d shapes. IEEE TVCG (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.523, + 0.785, + 0.55 + ], + "angle": 0, + "content": "33. Sipiran, I., Bustos, B.: Harris 3d: a robust extension of the harris operator for interest point detection on 3d meshes. The Vis. Comput. 27, 963-976 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "34. Sun, J., Ovsjanikov, M., Guibas, L.: A concise and provably informative multiscale signature based on heat diffusion. In: Comput. Graph. Forum. vol. 28, pp. 1383-1392. Wiley Online Library (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "35. Sung, M., Su, H., Yu, R., Guibas, L.J.: Deep functional dictionaries: Learning consistent semantic structures on 3d models from functions. NeurIPS 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "36. Tang, J., Gong, Z., Yi, R., Xie, Y., Ma, L.: Lake-net: Topology-aware point cloud completion by localizing aligned keypoints. In: CVPR. pp. 1726-1735 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "37. Tombari, F., Salti, S., Di Stefano, L.: Unique signatures of histograms for local surface description. In: ECCV. pp. 356-369. Springer (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "38. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.702, + 0.785, + 0.73 + ], + "angle": 0, + "content": "39. Wang, H., Guo, J., Yan, D.M., Quan, W., Zhang, X.: Learning 3d keypoint descriptors for non-rigid shape matching. In: ECCV. pp. 3-19 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.73, + 0.785, + 0.757 + ], + "angle": 0, + "content": "40. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. ACM TOG 38(5), 1-12 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "41. Wei, G., Cui, Z., Zhu, J., Yang, L., Zhou, Y., Singh, P., Gu, M., Wang, W.: Dense representative tooth landmark/axis detection network on 3d model. Computer Aided Geometric Design 94, 102077 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "42. Wei, G., Ma, L., Wang, C., Desrosiers, C., Zhou, Y.: Multi-task joint learning of 3d keypoint saliency and correspondence estimation. Computer-Aided Design 141, 103105 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.624, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "KeypointDETR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "43. Wu, W., Qi, Z., Fuxin, L.: Pointconv: Deep convolutional networks on 3d point clouds. In: CVPR. pp. 9621-9630 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.203 + ], + "angle": 0, + "content": "44. Yew, Z.J., Lee, G.H.: 3dfeat-net: Weakly supervised local 3d features for point cloud registration. In: ECCV. pp. 607-623 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.788, + 0.231 + ], + "angle": 0, + "content": "45. Yi, L., Su, H., Guo, X., Guibas, L.J.: Syncspeccnn: Synchronized spectral cnn for 3d shape segmentation. In: CVPR. pp. 2282-2290 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.259 + ], + "angle": 0, + "content": "46. You, Y., Liu, W., Ze, Y., Li, Y.L., Wang, W., Lu, C.: Ukpgan: A general self-supervised keypoint detector. In: CVPR. pp. 17042-17051 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.3 + ], + "angle": 0, + "content": "47. You, Y., Lou, Y., Li, C., Cheng, Z., Li, L., Ma, L., Lu, C., Wang, W.: Keypointnet: A large-scale 3d keypoint dataset aggregated from numerous human annotations. In: CVPR. pp. 13647-13656 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.328 + ], + "angle": 0, + "content": "48. Yu, X., Rao, Y., Wang, Z., Liu, Z., Lu, J., Zhou, J.: Pointr: Diverse point cloud completion with geometry-aware transformers. In: ICCV. pp. 12498-12507 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.369 + ], + "angle": 0, + "content": "49. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In: CVPR. pp. 19313-19322 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.41 + ], + "angle": 0, + "content": "50. Yuan, H., Zhao, C., Fan, S., Jiang, J., Yang, J.: Unsupervised learning of 3d semantic keypoints with mutual reconstruction. In: ECCV. pp. 534-549. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.411, + 0.788, + 0.452 + ], + "angle": 0, + "content": "51. Zhang, R., Wang, L., Qiao, Y., Gao, P., Li, H.: Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In: CVPR. pp. 21769-21780 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.479 + ], + "angle": 0, + "content": "52. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV. pp. 16259-16268 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.508 + ], + "angle": 0, + "content": "53. Zhong, C., You, P., Chen, X., Zhao, H., Sun, F., Zhou, G., Mu, X., Gan, C., Huang, W.: Snake: Shape-aware neural 3d keypoint field. NeurIPS 35, 7052-7064 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.536 + ], + "angle": 0, + "content": "54. Zhu, X., Du, D., Huang, H., Ma, C., Han, X.: 3d keypoint estimation using implicit representation learning. arXiv preprint arXiv:2306.11529 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.577 + ], + "angle": 0, + "content": "55. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.577 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/KeypointDETR_ An End-to-End 3D Keypoint Detector/d05c910b-e5c7-4df0-b517-a9b20e7a9b64_origin.pdf b/2024/KeypointDETR_ An End-to-End 3D Keypoint Detector/d05c910b-e5c7-4df0-b517-a9b20e7a9b64_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..875d288f050e82ab28bdbf7d668cf6dd13124dd1 --- /dev/null +++ b/2024/KeypointDETR_ An End-to-End 3D Keypoint Detector/d05c910b-e5c7-4df0-b517-a9b20e7a9b64_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b250b86f232542379c622e3f0c4d59f33be0ad6184ca2e88b85a87c298673e19 +size 3038831 diff --git a/2024/KeypointDETR_ An End-to-End 3D Keypoint Detector/full.md b/2024/KeypointDETR_ An End-to-End 3D Keypoint Detector/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0911a03afc1490cf1c1f3dc80341e8cc4843529c --- /dev/null +++ b/2024/KeypointDETR_ An End-to-End 3D Keypoint Detector/full.md @@ -0,0 +1,275 @@ +# KeypointDETR: An End-to-End 3D Keypoint Detector + +Hairong Jin $^{1,2\star}$ , Yuefan Shen $^{1}$ , Jianwen Lou $^{3}$ , Kun Zhou $^{1}$ , and Youyi Zheng $^{1(\boxtimes)}$ + +$^{1}$ State Key Lab of CAD&CG, Zhejiang University, Hangzhou, China hairong_jin, jhonve, kunzhou, youyizheng@zju.edu.cn +2 Chohotech Co.,Ltd., Hangzhou, China +3 School of Software Technology, Zhejiang University, Hangzhou, China jianwen.lou@zju.edu.cn + +Abstract. 3D keypoint detection plays a pivotal role in 3D shape analysis. The majority of prevalent methods depend on producing a shared heatmap. This approach necessitates subsequent post-processing techniques such as clustering or non-maximum suppression (NMS) to pinpoint keypoints within high-confidence regions, resulting in performance inefficiencies. To address this issue, we introduce KeypointDETR, an end-to-end 3D keypoint detection framework. KeypointDETR is predominantly trained with a bipartite matching loss, which compels the network to forecast sets of heatmaps and probabilities for potential keypoints. Each heatmap highlights one keypoint's location, and the associated probability indicates not only the presence of that specific keypoint but also its semantic consistency. Together with the bipartite matching loss, we utilize a transformer-based network architecture, which incorporates both point-wise and query-wise self-attention within the encoder and decoder, respectively. The point-wise encoder leverages the self-attention mechanism on a dynamic graph derived from the local feature space of each point, resulting in the generation of heatmap features. As a key part of our framework, the query-wise decoder facilitates inter-query information exchange. It captures the underlying connections among keypoints' heatmaps, positions, and semantic attributes via the cross-attention mechanism, enabling the prediction of heatmaps and probabilities. Extensive experiments conducted on the KeypointNet dataset reveal that KeypointDETR outperforms competitive baselines, demonstrating superior performance in keypoint saliency and correspondence estimation tasks. (The code will be released at github.com/bibi547/KeypointDETR) + +Keywords: 3D keypoint $\cdot$ Point cloud $\cdot$ Heatmap + +# 1 Introduction + +The automatic extraction of 3D keypoints presents a significant challenge in the realms of computer vision and computer graphics. These keypoints are instrumental in various geometric processing tasks, such as object detection [20], shape + +![](images/628c584447c38ac4ab09e2ddbe83460aaffe7bdc560e6faa094a6cf7cef82eb8.jpg) +Fig. 1: Illustration of the differences between the existing methods and KeypointDETR for keypoint saliency and correspondence estimation. Given a 3D model, Keypoint-DETR can predict its multi-heatmaps and probabilities. Heatmaps provide accurate inference of keypoints' positions. Probabilities indicate whether the specific keypoints exist in the 3D model. + +segmentation [13, 21], shape matching [39, 44], point registration [1], and point cloud completion [36], etc. + +Recent breakthroughs in deep learning, especially in 3D shape analysis and geometric feature extraction [5,11,27,28,40,43], have catalyzed the emergence of a multitude of learning-based approaches for 3D keypoint detection. A straightforward technique [45,47] is point classification on the surface of a 3D shape, but this method grapples with significant sample imbalances. To mitigate this, many existing keypoint detection strategies [32,41,42] utilize Gaussian kernel functions to create a ground truth heatmap. This shared heatmap is generated based on the distances between points and keypoints, guiding the extraction of keypoints from areas identified as salient. However, these techniques typically require post-processing steps such as clustering [29] or non-maximum suppression [22] to sift out high-confidence points and finalize the keypoint extraction. These processes often involve computing distance matrices, which can considerably slow down inference speeds. Additionally, the selection of parameters can profoundly affect the accuracy of keypoint extraction. Many landmark detection methods for 2D or 3D images [4] prefer using multi-heatmaps for keypoints, as opposed to relying on a single shared heatmap. This approach assigns each keypoint its unique heatmap, facilitating the extraction of the point with the highest value from each map. Although their strategy eliminates the need for post-processing, it can limit the algorithm's generalizability. One significant challenge is ensuring a consistent number of keypoints across all 3D models within a category. + +To tackle these challenges, we introduce an innovative end-to-end 3D keypoint detection model named KeypointDETR. This approach not only eliminates the need for any post-processing steps but also exhibits remarkable generalization capabilities. As illustrated in Fig. 1, our framework takes point coordinates as input and predicts multi-heatmaps along with their associated probabilities, + +providing a consistent paradigm for keypoint saliency and correspondence estimation. For keypoint saliency estimation, we can derive binary $yes / no$ labels from the probabilities, while for keypoint correspondence estimation, semantic labels can be inferred. Our pipeline utilizes a loss function based on the Hungarian algorithm [15], which facilitates the optimal bipartite matching between the predictions and the ground truth heatmaps. Consistent with this strategy, we develop a transformer-based encoder-decoder network architecture. The encoder employs dynamic graph transformer modules to extract geometric features from 3D models, specifically focusing on the generation of heatmap features. We regard the learnable heatmap features as intrinsic geometric features, eliminating the need for direct supervision. The dynamic graph transformer module aggregates the features from neighboring points within the feature space and assigns self-attention weights, effectively extending the range of perception by stacking these modules. In the decoder, we introduce a transformer refinement strategy that utilizes the query-wise self-attention mechanism to facilitate information exchange among potential keypoints. Additionally, it incorporates the cross-attention mechanism to capture the latent connections among keypoints' heatmaps, positions, and semantics. The query-wise decoder optimizes the heatmap and query features through the analysis of inter-keypoint and inter-attribute relationships, thereby enabling the parallel prediction of multi-heatmaps and probabilities. + +We validate our method on the KeypointNet dataset [47], achieving outstanding results in two essential tasks: keypoint saliency and correspondence estimation. Furthermore, our approach of assigning a specific heatmap to each keypoint leads to significantly higher accuracy compared to methods that use a shared heatmap. + +Our main contributions can be summarized as follows: + +- We introduce KeypointDETR, an end-to-end 3D keypoint detection framework. It is trained with a loss function based on the optimal bipartite matching derived from the cost matrix of multi-heatmaps and probabilities. +- We construct a novel transformer-based network architecture. The point-wise encoder harnesses self-attention on the dynamic graph of points' local feature space to extract geometric features. The query-wise decoder incorporates global self-attention to enhance the interaction between queries. +- We conduct a comprehensive evaluation that validates our method and provides insights into the impact of principal components, such as the bipartite matching loss and query-wise self-attention. + +# 2 Related Work + +# 2.1 3D Keypoint Detection + +In previous studies on 3D keypoint detection [23,33,37], algorithms heavily rely on a variety of hand-crafted geometric features, such as ISS [30], HKS [34], + +CGF [14], Salient Points [3], and Mesh Saliency [16]. These traditional approaches primarily focus on local geometric features of 3D shapes, often overlooking the integration of global and semantic information. + +With advancements in deep learning, there has been a notable surge in the development of data-driven algorithms for keypoint detection [10, 35, 54]. Numerous unsupervised 3D keypoint detection algorithms [1, 7, 12, 18, 31, 50, 53] are proposed for specific tasks such as point registration and reconstruction. For example, UKPGAN [46] utilizes a GAN-based module to manage keypoint sparsity and reconstructs 3D models by distilling saliency information. Unsupervised methods tend to focus on identifying keypoints in areas with distinct geometric features. While some of these methods are capable of attaining semantic consistency, they typically require the predetermination of the number of keypoints. Furthermore, in many scenarios, it is essential to identify keypoints that may not have prominent geometric features but possess solid semantic importance. Methods like SyncSpecCNN [45] and PRA-Net [5] address keypoint detection by classifying points on 3D shapes. These methods often result in an imbalance between positive and negative samples, presenting a significant challenge. Consequently, current popular methods employ a heatmap-based approach [32, 41], where the value assigned to each point indicates its proximity to the nearest keypoint. Wei et al. [42] proposes a multi-task learning framework that combines point-to-keypoint offsets with a confidence map, facilitating effective 3D keypoint saliency and correspondence estimation. Heatmap-based methods for 3D shapes predominantly depend on a shared heatmap for all keypoints and require post-processing. These methods employ post-processing techniques like non-maximum suppression [22] or clustering [29] to isolate keypoints from regions with high scores during the inference process. + +# 2.2 Transformer in Vision + +Transformer [38] was initially introduced in the realm of natural language processing. Recent advancements in visual research [6, 19] have explored various models centered around self-attention [38], achieving unprecedented performance. DETR [2] pioneers an innovative end-to-end object detection approach. It formulates a loss matching rule employing the Hungarian algorithm [15], facilitating bipartite matching between predictions and ground truth. This method serves as the cornerstone for a series of object detection methods [4,17,55] based on set prediction. Li et al. [17] develop a facial landmark detection method using cascaded transformer decoders. They devise parallel decoders utilizing deformable self-attention [55] to optimize facial landmark coordinates through offset predictions. Chen et al. [4] leverage self-attention to create a structure-aware LSTM framework tailored for predicting 3D heatmaps in CBCT. Keypoint transformer [9] efficiently extracts all potential keypoints within the image from a shared heatmap, subsequently employing the DETR decoder to predict 3D hand poses. Given the transformer's inherent capacity to handle input element permutations in sequences, it is exceptionally well-suited for point cloud processing [8, 24, 26]. + +![](images/d3926970f31692af47b56440d3e54f96bb96946fb089ed5910f49310668f2a29.jpg) +Fig. 2: Pipeline of KeypointDETR for keypoint saliency and correspondence estimation. Our framework takes point coordinates as input. The point-wise encoder first extracts geometric features from the point cloud by stacking dynamic graph transformer modules (DGT). Then, the query-wise decoder accepts heatmap features, which are transformed into heatmap and query features to generate multi-heatmaps and probabilities. These predictions enable the direct inference of both keypoints positions and semantic labels. + +Nonetheless, applying global self-attention to 3D point cloud models incurs substantial computational costs. Consequently, existing research [48, 52] frequently adopts local self-attention mechanisms, expanding receptive fields through sampling and grouping strategies. In contrast, our point-wise encoder integrates local self-attention within the dynamic graph of the point feature space, thereby eliminating the necessity for sampling and grouping operations. Distinct from most existing DETR style works, our query-wise decoder is meticulously engineered to predict multi-heatmaps and probabilities, enabling efficient end-to-end 3D keypoint detection. + +# 3 Method + +In this section, we delve into the specifics of our end-to-end 3D keypoint detector, KeypointDETR, which consists of two essential components: a network architecture based on the transformer and a loss function formulated using bipartite matching. As illustrated in Fig. 2, our approach is applied to tasks involving keypoint saliency and correspondence estimation. Upon receiving a 3D point cloud as input, our point-wise encoder utilizes the dynamic graph transformer modules to extract heatmap features, which can be considered intrinsic geometric features. We then deploy a query-wise transformer decoder designed to enhance information exchange between queries and capture the latent relationships between attributes through cross-attention, thereby optimizing heatmap + +![](images/a172b7cf989f94f61f4dfb7be805e5e4f0d0929863e53e084955cb4fe6015a29.jpg) +(a) The structure of the dynamic graph transformer module. + +![](images/64d5d4d97f9c41e14488bef42082634eff88c72bb8921cbd0873f3538faa0e5c.jpg) +(b) The structure of the transformer refinement module. +Fig. 3: The structure of the dynamic graph transformer module (DGT) and transformer refinement module. + +and query features. This network yields heatmaps and probabilities as its outputs. The prediction sets are aligned with ground truth heatmaps through a cost matrix, leading to the computation of the final loss based on optimal bipartite matching. During the inference stage, we select heatmaps corresponding to keypoints, as indicated by the predicted probabilities, and extract keypoints exhibiting the highest heat values from these maps. + +# 3.1 Data Preparation + +To tackle the issue of imbalanced samples concerning keypoints on the surface of 3D models, heatmap-based 3D keypoint detection methods establish the ground truth heat value for each point based on its distances to the keypoints. Specifically, the heatmap is generated by applying a Gaussian kernel in conjunction with the geodesic distance matrix: + +$$ +h _ {i, j} = \exp^ {- \frac {1}{2 \sigma^ {2}} \mathrm {d} \left(p _ {i}, k _ {j}\right)}. \tag {1} +$$ + +Given a 3D point cloud with $N_{p}$ points $P = \{p_{i}\}_{i=1}^{N_{p}}$ and $N_{k}$ keypoints $K = \{k_{j}\}_{j=1}^{N_{k}}$ , our process begins with the computation of the geodesic distance matrix from each point to the keypoints. Then, we derive the heatmap matrix $\mathcal{H} = \{h_{i,j}\} \in \mathbb{R}^{N_{p} \times N_{k}}$ using the Gaussian kernel function. Contrasting with methods that use a shared heatmap, where the maximum value for each point is considered, our approach distinctively employs multi-heatmaps, each corresponding to specific keypoints, as the ground truth. + +# 3.2 Network Architecture + +As shown in Fig. 2, our transformer-based encoder-decoder network architecture primarily comprises a point-wise encoder and a query-wise decoder. + +Point-wise Encoder The main objective of the encoder is to extract geometric features from the 3D point cloud. Drawing inspiration from DGCNN [40] and + +the Point Transformer [52], we devise a point-wise encoder constructed using dynamic graph transformer modules (DGT). + +As depicted in Fig. 3 (a), we employ local self-attention within the dynamic graph transformer module to operate on the feature space. To be specific, for the feature map $\mathbf{F}$ at each layer, we extract a $k$ -nearest neighbors feature set $\mathbf{F}_i = \{\mathbf{f}_j\}_{j=1}^k$ for $\mathbf{f}_i$ . Thereafter, we assign self-attention weights to these $k$ -nn features and aggregate them using the following equation: + +$$ +\mathbf {f} _ {i} ^ {\prime} = \sum_ {\mathbf {f} _ {j} \in \mathbf {F} _ {i}} \rho (\theta (\phi (\mathbf {f} _ {i}) - \varphi (\mathbf {f} _ {j}) + \delta)) \odot (\psi (\mathbf {f} _ {j}) + \delta). \tag {2} +$$ + +Here, $\theta$ is implemented as an MLP, while $\phi$ , $\varphi$ , $\psi$ are realized using Conv1d. $\rho$ represents the softmax operation applied to self-attention weights, and $\odot$ denotes the element-wise product operation. The term $\delta$ represents the position encoding, which is generated from the point coordinates $P_{i} = \{p_{j}\}_{j=1}^{k}$ corresponding to the neighborhood features, as expressed in the following: + +$$ +\delta = \operatorname {M L P} \left(p _ {i} - p _ {j}\right). \tag {3} +$$ + +Finally, we concatenate the features extracted from each dynamic graph transformer module to obtain point-wise features, which are decoded into heatmap features through MLPs. + +Query-wise Decoder The decoder's output comprises a predefined number of predictions, referred to as keypoint queries, with the query count $M$ set significantly higher than the number of keypoints within a 3D model. Each prediction query consists of both a heatmap and a probability. Heatmaps serve to indicate the position of keypoints within the 3D model. Probabilities reflect the semantic consistency of potential keypoints. In keypoint saliency estimation, they determine the presence or absence of specific keypoints. In keypoint correspondence estimation, such probabilities are correlated with their respective semantic labels. + +Guided by these objectives, we develop a query-wise decoder that adheres to the proven DETR-style architecture [2], as illustrated in Fig. 3 (b). Initially, we take the heatmap features $\mathbf{F}_{heat} \in \mathbb{R}^{N \times M}$ from the encoder. Considering the inherent relationship between the heatmaps and semantics of keypoints, the heatmap features are directly transposed and then decoded into query features $\mathbf{F}_{prob} \in \mathbb{R}^{M \times 128}$ through an MLP. Following this, both the heatmap $\mathbf{F}_{heat}$ and query $\mathbf{F}_{prob}$ features are fed into the transformer decoder. To facilitate the exchange of information among queries, a multi-head self-attention module is applied to $\mathbf{F}_{prob}$ . For the position encoding, we extract the coordinates of the reference points corresponding to the peaks of heat values in $\mathbf{F}_{heat}$ , which are subsequently processed through an MLP to produce the position encoding. The transposed $\mathbf{F}_{heat}$ and $\mathbf{F}_{prob}$ serve as inputs to the cross-attention module, enhancing the interplay of information between probability and heatmap features. Ultimately, through the application of Feed-Forward Network modules (FFN), + +the refined $\mathbf{F}_{heat}$ and $\mathbf{F}_{prob}$ are decoded into multi-heatmaps $\mathbf{M}_h$ and probabilities $\mathbf{M}_p$ , respectively. + +The decoder utilizes a query-wise self-attention mechanism across the queries, leading to $M$ distinct predictions. This mechanism operates as a global self-attention centered around potential keypoints. Such a globally parallel decoding approach efficiently captures the interactions among the predicted queries, thereby optimizing heat values and reducing redundancy in the identification of keypoints. Additionally, considering the potential relationship between keypoint heatmaps and semantics, cross-attention is employed to facilitate the interaction between heatmap and query features. + +# 3.3 Loss Function + +Inspired by DETR [2], we integrate the bipartite matching loss into 3D keypoint detection to enable end-to-end inference, eliminating the need for postprocessing. During the bipartite matching phase, we calculate the pairwise matching costs by assessing the discrepancies between the predicted queries and ground truth. We utilize this cost matrix to consolidate the prediction sets, distinguishing between "background" queries and keypoint queries, in order to compute the final loss. + +Given a 3D model with $N_{k}$ ground truth keypoints, KeypointDETR generates a set of $M$ predictions, where $M$ is preconfigured to be significantly larger than $N_{k}$ . The ground truth can be represented as $\{y_{i}\}_{i = 1}^{N_{k}} = \{h_{i},p_{i}\}_{i = 1}^{N_{k}}$ , where $h_i$ signifies the ground truth heatmap and $p_i$ denotes the ground truth probability. The predicted queries are denoted as $\{\hat{y}_j\}_{j = 1}^M = \{\hat{h_j},\hat{p_j}\}_{j = 1}^M$ , with $\hat{h_j}$ representing the predicted heatmap, and $\hat{p_j}$ corresponding to the predicted probabilities associated with keypoints categories $\{c_i\}_{i = 1}^{N_k}$ . In the task of keypoint saliency estimation, $p_i$ remains fixed at 1, while $\hat{p_j}(c_i)$ signifies the probability of being predicted as a keypoint. However, in the task of keypoint correspondence estimation, $\hat{p_j}(c_i)$ is defined as the probability of $c_i$ , where $c_i$ denotes the semantic label corresponding to the ground truth keypoint $k_i$ . Expanding upon these notations, in order to achieve the optimal bipartite matching between ground truth and predictions, it is imperative to establish a matching cost matrix $\mathcal{M}_c \in \mathbb{R}^{N_k \times M}$ . This cost matrix takes into account both the heatmaps and probabilities. Precisely, each element $\mathcal{M}_c(i,j)$ of the cost matrix can be calculated as follows: + +$$ +\mathcal {M} _ {c} (i, j) = \mathbf {W} _ {\mathbf {h}} \cdot \sum | h _ {i} - \hat {h} _ {j} | + \mathbf {W} _ {\mathbf {p}} \cdot \left(1 - \log p _ {\hat {j}} \left(c _ {i}\right)\right), \tag {4} +$$ + +where $\mathbf{W_h}$ represents the weight assigned to heatmaps, and $\mathbf{W_p}$ is the weight attributed to probabilities. $\sum |h_i - \hat{h_j}|$ signifies the dissimilarity between the ground truth $h_i$ and predicted $\hat{h_j}$ heatmaps. + +Following the application of the Hungarian algorithm [15], we obtain the optimal matching and categorize the predicted queries into two groups: positive queries and negative queries. Positive queries are those that align with the ground truth. For a positive query $\hat{y}_i$ , we determine the matched ground truth category as $g_i$ and the corresponding heatmap as $h_i$ . Then, we compute the probability + +loss using the Focal Loss (FL) and the heatmap loss employing the Mean Squared Error Loss (MSE). For a negative query, only the probability loss is considered. The associated ground truth $g_{i}$ is designated as 0, signifying nonexistence. Thus, the overall loss can be calculated as follows: + +$$ +\mathcal {L} = \sum_ {i = 1} ^ {M} \left(\lambda_ {p} \operatorname {F L} \left(g _ {i}, \hat {p _ {i}}\right) + \lambda_ {h} \mathbb {1} _ {\{g _ {i} \neq 0 \}} \operatorname {M S E} \left(h _ {i}, \hat {h _ {i}}\right)\right). \tag {5} +$$ + +# 4 Experiments + +In this section, we conduct a series of comprehensive experiments to validate the superior performance of our approach in both keypoint saliency and correspondence estimation tasks. We provide comparisons with other state-of-the-art methods to demonstrate our method's efficacy. + +# 4.1 Implementation Details + +KeypointDETR takes 3D point coordinates as input. In the point-wise encoder, we utilize 5 dynamic graph transformer modules, each with point features of 64 dimensions, the $k$ value of $k$ -nearest neighbors set to 20, and a global embedding dimension of 1024. In the query-wise decoder, we set the number of queries $M$ to 50. The decoder's output comprises two branches: the heatmap head generates multi-heatmaps $\mathbf{M}_h \in \mathbb{R}^{M \times N_p}$ , and the classification head provides probabilities $\mathbf{M}_p \in \mathbb{R}^{M \times C}$ . For keypoint saliency estimation, $C$ is held constant at 2. However, for keypoint correspondence estimation, $C$ is configured as $(N_c + 1)$ , where $N_c$ signifies the number of keypoint categories. More details can be found in the supplementary material. + +# 4.2 Dataset + +Our experiments are conducted on the KeypointNet dataset [47], which is designed for both keypoint saliency and correspondence estimation tasks. We focus on four classical categories: airplane, chair, guitar, and table. These models are divided into training, validation, and test sets, with a distribution ratio of 7:1:2 for each category. We use 2,048 sampled points from each 3D model as input data. The results of other categories and applications of KeypointDETR can be found in the supplementary material. + +# 4.3 Metrics + +We evaluate our method alongside comparison methods on the keypoint saliency estimation using mIoU and Chamfer Distance (CD) while utilizing mIoU and Dual Alignment Score (DAS) [31] for keypoint correspondence estimation. The IoU is calculated as follows: + +$$ +I o U = \frac {T P}{T P + F P + F N}, \tag {6} +$$ + +where $TP$ stands for true positives, $FP$ represents false positives, and $FN$ corresponds to false negatives. Given the ground truth and predicted keypoints sets $G$ and $K$ , CD can be computed as: + +$$ +C D = \frac {1}{| G |} \sum_ {p \in G} \min _ {q \in K} \| p - q \| _ {2} ^ {2} + \frac {1}{| K |} \sum_ {p \in K} \min _ {q \in G} \| p - q \| _ {2} ^ {2}. \tag {7} +$$ + +The settings for clustering and NMS parameters profoundly affect the repeatability and omission rates of predicted keypoints. However, the mIoU metric used in prior methods fails to effectively capture these erroneous cases. To ensure a comprehensive comparison, we design a mIoU metric based on the Hungarian algorithm [15]. For the keypoint saliency estimation, we begin by calculating the Euclidean distance matrix between the ground truth and the predicted keypoints. Subsequently, the Hungarian algorithm is employed to determine their optimal bipartite matching. Each matched distance falling below the predefined threshold $\epsilon$ is considered a true positive. For keypoint correspondence estimation, a ground truth keypoint is considered a true positive if there is a predicted keypoint with the same semantic label whose distance to it is less than the threshold $\epsilon$ . Additionally, we further measure whether the closest keypoints between predictions and ground truth have the same semantic labels using DAS [31]. + +# 4.4 Comparison on Keypoint Saliency Estimation + +We conduct a comparative analysis of our method against six advanced algorithms, including two unsupervised keypoint detection methods (Skeleton Merger [31] and UKPGAN [46]) and three point cloud geometric feature extraction algorithms (PointNet++ [28], PRA-Net [5], and DGCNN [40]), and a fully-supervised keypoint detection method (Wei et al. [42]). For the three point cloud geometric feature extraction algorithms on saliency analysis, the network's output is a shared heatmap, wherein each point is assigned the maximum value of the multi-heatmaps. + +The quantitative results for mIoU and Chamfer Distance (CD) are presented in Tab. 1, where we compare our KeypointDETR with other algorithms using a threshold $\epsilon$ of 0.1. Fig. 6 (a) depicts the mIoU curves for various methods as $\epsilon$ varies from 0 to 0.1. It is evident that KeypointDETR consistently outperforms other methods in the Hungarian mIoU evaluation across all categories. This significant advantage stems from our method's capacity to bypass post-processing techniques like NMS or clustering algorithms, leading to direct 'one-to-one' predictions that effectively tackle redundancy and omissions in 3D keypoint detection. In contrast, methods reliant on a shared heatmap face challenges with parameter selection of post-processing, which predominantly depends on prior knowledge factors such as the expected number of keypoints and the 3D model's scale. Unfortunately, even optimal parameter settings for NMS (radius=0.1) in these methods do not entirely resolve issues of omissions and redundant predictions, as further illustrated in Fig. 4. + +Table 1: Quantitative results of keypoint saliency estimation using Hungarian mIoU (\%) and CD. + +
AirplaneChairGuitarTableAverage
mIoUCDmIoUCDmIoUCDmIoUCDmIoUCD
SkeletonMerger [31]60.730.10742.740.18947.710.13227.840.30244.750.182
UKPGAN [46]62.630.11825.160.16854.250.13117.780.24039.950.164
PRA-Net [5]82.880.04970.440.06073.710.05369.690.05374.180.053
PointNet++ [28]82.310.05057.500.07772.340.06060.010.06268.040.062
DGCNN [40]78.490.05859.490.07663.880.07259.920.06665.440.068
Wei et al. [42]79.860.04886.280.05578.080.06686.610.03982.700.052
Ours91.370.04591.890.05593.960.04598.050.02793.810.043
+ +![](images/da1962fca4c8e3590bda244efc3a23c4c0223b2aec5bd070bbb4e446bf948610.jpg) +Fig. 4: Visualization examples of keypoint saliency estimation. + +# 4.5 Comparison on Keypoint Correspondence Estimation + +Keypoint correspondence estimation represents a more significant challenge compared to keypoint saliency estimation, as it not only involves locating keypoints but also predicting their semantic labels. Additionally, 3D models within the same category may not have identical semantic keypoints. Prior approaches [35, 45,47] typically involve direct classification of points within a point cloud, which does not effectively address the issue of imbalanced positive and negative samples in keypoint detection. Some methods [42, 54] attempt to add a semantic branch to keypoint saliency estimation, aiming to assign labels to keypoints in a two-stage process. + +As depicted in Tab. 2, we conduct a comparison of our method with five methods using a threshold $\epsilon$ of 0.1: PointConv [43], PointNet++ [28], DGCNN [40], PRA-Net [5], and Wei et al. [42]. Additionally, we present the mIoU curves for various methods, with the $\epsilon$ ranging from 0 to 0.1, in Fig. 6 (b). Fig. 5 showcases visualization results of different methods alongside the ground truth. Notably, + +Table 2: Quantitative results of keypoint correspondence estimation using Hungarian mIoU (%) and DAS [31] (%). + +
AirplaneChairGuitarTableAverage
mIoUDASmIoUDASmIoUDASmIoUDASmIoUDAS
DGCNN [40]77.7782.6963.1679.5661.1378.5957.9978.8365.0179.91
PointNet++ [28]79.1084.7458.2478.1967.6981.9262.9781.5967.0081.61
PointConv [43]81.5485.7464.9282.1576.3988.9166.2083.7372.2685.13
PRA-Net [5]81.1085.8068.3583.6272.0785.1570.1684.8972.9284.86
Wei et al. [42]78.5488.3479.7790.4475.6493.6986.5697.3180.1292.44
Ours85.7191.7479.3389.5390.9393.8992.4097.7987.0993.23
+ +![](images/0b90e31dcf5ceaea378fae405d2c9b7f7502fbbac83dac648efd28d13eddb5bc.jpg) +Fig. 5: Visualization examples of keypoint correspondence estimation. + +our method generates no redundant points with the same label in close proximity to predicted keypoints. This demonstrates that our method adeptly accomplishes semantic correspondence for keypoints, with KeypointDETR showcasing superior performance in both the heatmap and probability branch. + +# 4.6 Ablation Study + +We conduct a series of ablation experiments to validate the performance of KeypointDETR and deeply explore the impact of each component. Throughout these experiments, we consistently use four representative categories from the KeypointNet dataset for keypoint saliency estimation. The evaluation adopts mIoU with a threshold $\epsilon$ of 0.1. + +Effects of point-wise encoder. Our point-wise encoder primarily consists of dynamic graph transformer modules that extract geometric features from the point cloud. To validate the effectiveness of our point-wise encoder, we compare KeypointDETR with its variants, which utilize different backbones for geometric feature extraction, namely DGCNN [40], PointNet++ [28], and Point Transformer [52]. The DGCNN [40] backbone (DG.Enc.) is implemented as a five-layer + +![](images/5272ec86e824043d21ad877d71e1dc6aa7d8a4ce5f80e266cd8b661fb74c95e7.jpg) +(a) Hungarian mIoU curves for keypoint saliency estimation. + +![](images/6638a544fd8bea7b654de57d2c5544416436791a87ad1688af1823ae8fa958c7.jpg) +(b) Hungarian mIoU curves for keypoint correspondence estimation. +Fig. 6: Hungarian mIoU curves for keypoint saliency and correspondence estimation under various distance thresholds (0-0.1). + +Table 3: Quantitative results of various variations in the ablation study using Hungarian mIoU (%) and CD. + +
AirplaneChairGuitarTableAverage
mIoUCDmIoUCDmIoUCDmIoUCDmIoUCD
DG.Enc. [40]86.750.04882.120.06292.760.05297.620.03789.810.049
PN2.Enc. [28]76.300.07168.310.10172.090.09976.420.09973.280.092
PT.Enc. [52]82.550.05973.040.08568.920.09197.320.03280.450.066
Point-BERT Enc. [49]91.110.03493.570.03493.510.03293.710.03392.880.033
I2P-MAE Enc. [51]79.060.05080.400.06788.790.04886.050.04883.570.053
Point-MAE Enc. [25]89.820.03993.980.05583.790.04395.950.04390.880.045
w/o Trans.Dec.87.960.04488.230.05892.740.04597.770.02991.670.044
w/o PE.88.050.04291.430.05492.440.04797.970.02892.470.042
Coord-based66.720.10844.550.19037.990.16450.080.18749.830.162
Coord+offsets68.400.09463.620.13642.450.15160.770.15958.810.135
Ours91.370.04591.890.05593.960.04598.050.02793.810.043
+ +stack of EdgeConv. PointNet++ [28] (PN2.Enc.) and Point Transformer [52] (PT.Enc.) employ eight layers due to their U-net structures. Furthermore, we explore replacing our encoder with pre-trained models, including Point-BERT, Point-MAE, and I2P-MAE, and achieve favorable results. The overall quantitative comparison results are presented in the first sub-table of Tab. 3. + +Effects of query-wise decoder. In the decoder phase, we develop a query-wise decoder utilizing self-attention to facilitate enhanced information exchange among queries. This decoder is tasked with decoding the heatmap and probability feature maps, resulting in the generation of multi-heatmaps and probabilities. A straightforward approach involves employing MLPs to decode these feature maps separately. However, to demonstrate the efficacy and reasoning behind our query-wise decoder design, we conduct an experiment where we replace our self-attention decoder (Trans.Dec.) with MLPs. The comparison results, as shown in the second sub-table of Tab. 3, indicate that the comprehensive performance of our KeypointDETR significantly surpasses that of its MLPs decoder variant. + +Effects of position encoding in decoder. Position encoding holds a pivotal role in the self-attention mechanism and is a key factor that makes the transformer architecture suitable for 3D point cloud models, as previously validated in the Point Transformer [52]. In our investigation, we further explore the role of position encoding in our query-wise decoder, which is generated from heatmap features and point coordinates. We compare the results with and without position encoding (PE.) in the third sub-table of Tab. 3. + +Coordinate-based or Heatmap-based. We further explore a coordinate-based architecture within our KeypointDETR approach. In this variant, KeypointDETR predicts the coordinates and probabilities of keypoints, and we subsequently select the closest matching points from the point cloud as the final keypoints. To enhance the accuracy of these coordinates, we incorporate a cascade module designed to predict offsets. Comparative analysis of the heatmap-based KeypointDETR and its coordinate-based variant, as shown in the fourth sub-table of Tab. 3, reveals that while the addition of offsets improves the results of the coordinate-based model, there remains a notable performance gap when compared to our primary heatmap-based KeypointDETR. This discrepancy arises because our geometric feature extraction backbone is better suited for processing 3D shapes rather than 3D spatial coordinates. Furthermore, it validates that the heatmap-based approach can achieve better performance on keypoints detection. + +# 5 Conclusion + +In this paper, we introduce KeypointDETR, an end-to-end 3D keypoint detector that obviates the need for post-processing steps, such as clustering and NMS, commonly required in heatmap-based keypoint detection methods. The training strategy of KeypointDETR involves computing a cost matrix between the predictions and ground truth, deriving an optimal bipartite matching loss. This innovative approach enables KeypointDETR to accurately predict multi-heatmaps and probabilities, which can provide the positions and semantic consistency of potential keypoints, respectively. Complementing this framework is a transformer-based network architecture. Extensive quantitative experiments and analyses substantiate the effectiveness and rationality of KeypointDETR, underlining its potential as a state-of-the-art solution in keypoint detection. + +# Acknowledgements + +This work is partially supported by the following grants: National Natural Science Foundation of China (No. 62172363), Management Center of the School of Software (Ningbo) at Zhejiang University under Grant (No. Z24001). + +# References + +1. Bai, X., Luo, Z., Zhou, L., Fu, H., Quan, L., Tai, C.L.: D3feat: Joint learning of dense detection and description of 3d local features. In: CVPR. pp. 6359-6367 (2020) +2. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV. pp. 213-229. Springer (2020) +3. Castellani, U., Cristani, M., Fantoni, S., Murino, V.: Sparse points matching by combining 3d mesh saliency with statistical descriptors. In: Comput. Graph. Forum. vol. 27, pp. 643-652. Wiley Online Library (2008) +4. Chen, R., Ma, Y., Chen, N., Liu, L., Cui, Z., Lin, Y., Wang, W.: Structure-aware long short-term memory network for 3d cephalometric landmark detection. IEEE Transactions on Medical Imaging 41(7), 1791-1801 (2022) +5. Cheng, S., Chen, X., He, X., Liu, Z., Bai, X.: Pra-net: Point relation-aware network for 3d point cloud analysis. IEEE TIP 30, 4436-4448 (2021) +6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +7. Fernandez-Labrador, C., Chhatkuli, A., Paudel, D.P., Guerrero, J.J., Demonceaux, C., Gool, L.V.: Unsupervised learning of category-specific symmetric 3d keypoints from point sets. In: ECCV. pp. 546-563. Springer (2020) +8. Guo, M.H., Cai, J.X., Liu, Z.N., Mu, T.J., Martin, R.R., Hu, S.M.: Pct: Point cloud transformer. Computational Visual Media 7, 187-199 (2021) +9. Hampali, S., Sarkar, S.D., Rad, M., Lepetit, V.: Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In: CVPR. pp. 11090-11100 (2022) +0. He, Y., Sun, W., Huang, H., Liu, J., Fan, H., Sun, J.: Pvn3d: A deep point-wise 3d keypoints voting network for 6 dof pose estimation. In: CVPR. pp. 11632-11641 (2020) +1. Huang, Q., Wang, W., Neumann, U.: Recurrent slice networks for 3d segmentation of point clouds. In: CVPR. pp. 2626-2635 (2018) +2. Jakab, T., Tucker, R., Makadia, A., Wu, J., Snavely, N., Kanazawa, A.: Keypoint-deformer: Unsupervised 3d keypoint discovery for shape control. In: CVPR. pp. 12783-12792 (2021) +3. Katz, S., Leifman, G., Tal, A.: Mesh segmentation using feature point and core extraction. The Vis. Comput. 21, 649-658 (2005) +4. Khoury, M., Zhou, Q.Y., Koltun, V.: Learning compact geometric features. In: ICCV. pp. 153-161 (2017) +5. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955) +6. Lee, C.H., Varshney, A., Jacobs, D.W.: Mesh saliency. In: ACM SIGGRAPH 2005 Papers, pp. 659-666 (2005) +7. Li, H., Guo, Z., Rhee, S.M., Han, S., Han, J.J.: Towards accurate facial landmark detection via cascaded transformers. In: CVPR. pp. 4176-4185 (2022) +8. Li, J., Lee, G.H.: Usip: Unsupervised stable interest point detection from 3d point clouds. In: ICCV. pp. 361-370 (2019) +9. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV. pp. 10012-10022 (2021) + +20. Liu, Z., Wu, Z., Toth, R.: Smoke: Single-stage monocular 3d object detection via keypoint estimation. In: CVPR. pp. 996-997 (2020) +21. Mian, A.S., Bennamoun, M., Owens, R.: Three-dimensional model-based object recognition and segmentation in cluttered scenes. IEEE TPAMI 28(10), 1584-1601 (2006) +22. Neubeck, A., Van Gool, L.: Efficient non-maximum suppression. In: ICPR. vol. 3, pp. 850-855. IEEE (2006) +23. Novatnack, J., Nishino, K.: Scale-dependent 3d geometric features. In: ICCV. pp. 1-8. IEEE (2007) +24. Pan, X., Xia, Z., Song, S., Li, L.E., Huang, G.: 3d object detection with pointformer. In: CVPR. pp. 7463-7472 (2021) +25. Pang, Y., Wang, W., Tay, F.E., Liu, W., Tian, Y., Yuan, L.: Masked autoencoders for point cloud self-supervised learning. In: ECCV. pp. 604-621. Springer (2022) +26. Park, C., Jeong, Y., Cho, M., Park, J.: Fast point transformer. In: CVPR. pp. 16949-16958 (2022) +27. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR. pp. 652-660 (2017) +28. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS 30 (2017) +29. Rodriguez, A., Laio, A.: Clustering by fast search and find of density peaks. science 344(6191), 1492-1496 (2014) +30. Salti, S., Tombari, F., Spezialietti, R., Di Stefano, L.: Learning a descriptor-specific 3d keypoint detector. In: ICCV. pp. 2318-2326 (2015) +31. Shi, R., Xue, Z., You, Y., Lu, C.: Skeleton merger: an unsupervised aligned keypoint detector. In: CVPR. pp. 43-52 (2021) +32. Shu, Z., Yu, J., Chao, K., Xin, S., Liu, L.: A multi-modal attention-based approach for points of interest detection on 3d shapes. IEEE TVCG (2024) +33. Sipiran, I., Bustos, B.: Harris 3d: a robust extension of the harris operator for interest point detection on 3d meshes. The Vis. Comput. 27, 963-976 (2011) +34. Sun, J., Ovsjanikov, M., Guibas, L.: A concise and provably informative multiscale signature based on heat diffusion. In: Comput. Graph. Forum. vol. 28, pp. 1383-1392. Wiley Online Library (2009) +35. Sung, M., Su, H., Yu, R., Guibas, L.J.: Deep functional dictionaries: Learning consistent semantic structures on 3d models from functions. NeurIPS 31 (2018) +36. Tang, J., Gong, Z., Yi, R., Xie, Y., Ma, L.: Lake-net: Topology-aware point cloud completion by localizing aligned keypoints. In: CVPR. pp. 1726-1735 (2022) +37. Tombari, F., Salti, S., Di Stefano, L.: Unique signatures of histograms for local surface description. In: ECCV. pp. 356-369. Springer (2010) +38. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017) +39. Wang, H., Guo, J., Yan, D.M., Quan, W., Zhang, X.: Learning 3d keypoint descriptors for non-rigid shape matching. In: ECCV. pp. 3-19 (2018) +40. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. ACM TOG 38(5), 1-12 (2019) +41. Wei, G., Cui, Z., Zhu, J., Yang, L., Zhou, Y., Singh, P., Gu, M., Wang, W.: Dense representative tooth landmark/axis detection network on 3d model. Computer Aided Geometric Design 94, 102077 (2022) +42. Wei, G., Ma, L., Wang, C., Desrosiers, C., Zhou, Y.: Multi-task joint learning of 3d keypoint saliency and correspondence estimation. Computer-Aided Design 141, 103105 (2021) + +43. Wu, W., Qi, Z., Fuxin, L.: Pointconv: Deep convolutional networks on 3d point clouds. In: CVPR. pp. 9621-9630 (2019) +44. Yew, Z.J., Lee, G.H.: 3dfeat-net: Weakly supervised local 3d features for point cloud registration. In: ECCV. pp. 607-623 (2018) +45. Yi, L., Su, H., Guo, X., Guibas, L.J.: Syncspeccnn: Synchronized spectral cnn for 3d shape segmentation. In: CVPR. pp. 2282-2290 (2017) +46. You, Y., Liu, W., Ze, Y., Li, Y.L., Wang, W., Lu, C.: Ukpgan: A general self-supervised keypoint detector. In: CVPR. pp. 17042-17051 (2022) +47. You, Y., Lou, Y., Li, C., Cheng, Z., Li, L., Ma, L., Lu, C., Wang, W.: Keypointnet: A large-scale 3d keypoint dataset aggregated from numerous human annotations. In: CVPR. pp. 13647-13656 (2020) +48. Yu, X., Rao, Y., Wang, Z., Liu, Z., Lu, J., Zhou, J.: Pointr: Diverse point cloud completion with geometry-aware transformers. In: ICCV. pp. 12498-12507 (2021) +49. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In: CVPR. pp. 19313-19322 (2022) +50. Yuan, H., Zhao, C., Fan, S., Jiang, J., Yang, J.: Unsupervised learning of 3d semantic keypoints with mutual reconstruction. In: ECCV. pp. 534-549. Springer (2022) +51. Zhang, R., Wang, L., Qiao, Y., Gao, P., Li, H.: Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In: CVPR. pp. 21769-21780 (2023) +52. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV. pp. 16259-16268 (2021) +53. Zhong, C., You, P., Chen, X., Zhao, H., Sun, F., Zhou, G., Mu, X., Gan, C., Huang, W.: Snake: Shape-aware neural 3d keypoint field. NeurIPS 35, 7052-7064 (2022) +54. Zhu, X., Du, D., Huang, H., Ma, C., Han, X.: 3d keypoint estimation using implicit representation learning. arXiv preprint arXiv:2306.11529 (2023) +55. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020) \ No newline at end of file diff --git a/2024/KeypointDETR_ An End-to-End 3D Keypoint Detector/images.zip b/2024/KeypointDETR_ An End-to-End 3D Keypoint Detector/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3ee91cd361bc90ab58565f37e17cb63f44b6bd28 --- /dev/null +++ b/2024/KeypointDETR_ An End-to-End 3D Keypoint Detector/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b91b8275b7eff2ecb40e2cb2c2b88040772f63b577c7ee27f773a2cfc87066b0 +size 525469 diff --git a/2024/KeypointDETR_ An End-to-End 3D Keypoint Detector/layout.json b/2024/KeypointDETR_ An End-to-End 3D Keypoint Detector/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d33139565079741cbd06faee9df3288edb844089 --- /dev/null +++ b/2024/KeypointDETR_ An End-to-End 3D Keypoint Detector/layout.json @@ -0,0 +1,8686 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 111, + 467, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 111, + 467, + 145 + ], + "spans": [ + { + "bbox": [ + 146, + 111, + 467, + 145 + ], + "type": "text", + "content": "KeypointDETR: An End-to-End 3D Keypoint Detector" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "spans": [ + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "text", + "content": "Hairong Jin" + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "inline_equation", + "content": "^{1,2\\star}" + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "text", + "content": ", Yuefan Shen" + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "text", + "content": ", Jianwen Lou" + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "text", + "content": ", Kun Zhou" + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "text", + "content": ", and Youyi Zheng" + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "inline_equation", + "content": "^{1(\\boxtimes)}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 160, + 201, + 452, + 257 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 164, + 201, + 448, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 201, + 448, + 224 + ], + "spans": [ + { + "bbox": [ + 164, + 201, + 448, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 164, + 201, + 448, + 224 + ], + "type": "text", + "content": " State Key Lab of CAD&CG, Zhejiang University, Hangzhou, China hairong_jin, jhonve, kunzhou, youyizheng@zju.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 224, + 224, + 390, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 224, + 390, + 235 + ], + "spans": [ + { + "bbox": [ + 224, + 224, + 390, + 235 + ], + "type": "text", + "content": "2 Chohotech Co.,Ltd., Hangzhou, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 235, + 452, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 235, + 452, + 257 + ], + "spans": [ + { + "bbox": [ + 160, + 235, + 452, + 257 + ], + "type": "text", + "content": "3 School of Software Technology, Zhejiang University, Hangzhou, China jianwen.lou@zju.edu.cn" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 160, + 277, + 452, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 277, + 452, + 552 + ], + "spans": [ + { + "bbox": [ + 160, + 277, + 452, + 552 + ], + "type": "text", + "content": "Abstract. 3D keypoint detection plays a pivotal role in 3D shape analysis. The majority of prevalent methods depend on producing a shared heatmap. This approach necessitates subsequent post-processing techniques such as clustering or non-maximum suppression (NMS) to pinpoint keypoints within high-confidence regions, resulting in performance inefficiencies. To address this issue, we introduce KeypointDETR, an end-to-end 3D keypoint detection framework. KeypointDETR is predominantly trained with a bipartite matching loss, which compels the network to forecast sets of heatmaps and probabilities for potential keypoints. Each heatmap highlights one keypoint's location, and the associated probability indicates not only the presence of that specific keypoint but also its semantic consistency. Together with the bipartite matching loss, we utilize a transformer-based network architecture, which incorporates both point-wise and query-wise self-attention within the encoder and decoder, respectively. The point-wise encoder leverages the self-attention mechanism on a dynamic graph derived from the local feature space of each point, resulting in the generation of heatmap features. As a key part of our framework, the query-wise decoder facilitates inter-query information exchange. It captures the underlying connections among keypoints' heatmaps, positions, and semantic attributes via the cross-attention mechanism, enabling the prediction of heatmaps and probabilities. Extensive experiments conducted on the KeypointNet dataset reveal that KeypointDETR outperforms competitive baselines, demonstrating superior performance in keypoint saliency and correspondence estimation tasks. (The code will be released at github.com/bibi547/KeypointDETR)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 562, + 367, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 562, + 367, + 573 + ], + "spans": [ + { + "bbox": [ + 160, + 562, + 367, + 573 + ], + "type": "text", + "content": "Keywords: 3D keypoint " + }, + { + "bbox": [ + 160, + 562, + 367, + 573 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 562, + 367, + 573 + ], + "type": "text", + "content": " Point cloud " + }, + { + "bbox": [ + 160, + 562, + 367, + 573 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 562, + 367, + 573 + ], + "type": "text", + "content": " Heatmap" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 590, + 230, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 590, + 230, + 602 + ], + "spans": [ + { + "bbox": [ + 133, + 590, + 230, + 602 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 612, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 482, + 648 + ], + "type": "text", + "content": "The automatic extraction of 3D keypoints presents a significant challenge in the realms of computer vision and computer graphics. These keypoints are instrumental in various geometric processing tasks, such as object detection [20], shape" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 654, + 444, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 444, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 444, + 665 + ], + "type": "text", + "content": "* This work was done during the author's internship at Chohotech Co.,Ltd.." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 471, + 251 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 471, + 251 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 471, + 251 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 471, + 251 + ], + "type": "image", + "image_path": "628c584447c38ac4ab09e2ddbe83460aaffe7bdc560e6faa094a6cf7cef82eb8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 259, + 482, + 315 + ], + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 315 + ], + "type": "text", + "content": "Fig. 1: Illustration of the differences between the existing methods and KeypointDETR for keypoint saliency and correspondence estimation. Given a 3D model, Keypoint-DETR can predict its multi-heatmaps and probabilities. Heatmaps provide accurate inference of keypoints' positions. Probabilities indicate whether the specific keypoints exist in the 3D model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 340, + 479, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 479, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 479, + 365 + ], + "type": "text", + "content": "segmentation [13, 21], shape matching [39, 44], point registration [1], and point cloud completion [36], etc." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 605 + ], + "type": "text", + "content": "Recent breakthroughs in deep learning, especially in 3D shape analysis and geometric feature extraction [5,11,27,28,40,43], have catalyzed the emergence of a multitude of learning-based approaches for 3D keypoint detection. A straightforward technique [45,47] is point classification on the surface of a 3D shape, but this method grapples with significant sample imbalances. To mitigate this, many existing keypoint detection strategies [32,41,42] utilize Gaussian kernel functions to create a ground truth heatmap. This shared heatmap is generated based on the distances between points and keypoints, guiding the extraction of keypoints from areas identified as salient. However, these techniques typically require post-processing steps such as clustering [29] or non-maximum suppression [22] to sift out high-confidence points and finalize the keypoint extraction. These processes often involve computing distance matrices, which can considerably slow down inference speeds. Additionally, the selection of parameters can profoundly affect the accuracy of keypoint extraction. Many landmark detection methods for 2D or 3D images [4] prefer using multi-heatmaps for keypoints, as opposed to relying on a single shared heatmap. This approach assigns each keypoint its unique heatmap, facilitating the extraction of the point with the highest value from each map. Although their strategy eliminates the need for post-processing, it can limit the algorithm's generalizability. One significant challenge is ensuring a consistent number of keypoints across all 3D models within a category." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "To tackle these challenges, we introduce an innovative end-to-end 3D keypoint detection model named KeypointDETR. This approach not only eliminates the need for any post-processing steps but also exhibits remarkable generalization capabilities. As illustrated in Fig. 1, our framework takes point coordinates as input and predicts multi-heatmaps along with their associated probabilities," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "H.Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 366 + ], + "type": "text", + "content": "providing a consistent paradigm for keypoint saliency and correspondence estimation. For keypoint saliency estimation, we can derive binary " + }, + { + "bbox": [ + 130, + 116, + 482, + 366 + ], + "type": "inline_equation", + "content": "yes / no" + }, + { + "bbox": [ + 130, + 116, + 482, + 366 + ], + "type": "text", + "content": " labels from the probabilities, while for keypoint correspondence estimation, semantic labels can be inferred. Our pipeline utilizes a loss function based on the Hungarian algorithm [15], which facilitates the optimal bipartite matching between the predictions and the ground truth heatmaps. Consistent with this strategy, we develop a transformer-based encoder-decoder network architecture. The encoder employs dynamic graph transformer modules to extract geometric features from 3D models, specifically focusing on the generation of heatmap features. We regard the learnable heatmap features as intrinsic geometric features, eliminating the need for direct supervision. The dynamic graph transformer module aggregates the features from neighboring points within the feature space and assigns self-attention weights, effectively extending the range of perception by stacking these modules. In the decoder, we introduce a transformer refinement strategy that utilizes the query-wise self-attention mechanism to facilitate information exchange among potential keypoints. Additionally, it incorporates the cross-attention mechanism to capture the latent connections among keypoints' heatmaps, positions, and semantics. The query-wise decoder optimizes the heatmap and query features through the analysis of inter-keypoint and inter-attribute relationships, thereby enabling the parallel prediction of multi-heatmaps and probabilities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 368, + 482, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 427 + ], + "type": "text", + "content": "We validate our method on the KeypointNet dataset [47], achieving outstanding results in two essential tasks: keypoint saliency and correspondence estimation. Furthermore, our approach of assigning a specific heatmap to each keypoint leads to significantly higher accuracy compared to methods that use a shared heatmap." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 428, + 387, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 428, + 387, + 439 + ], + "spans": [ + { + "bbox": [ + 146, + 428, + 387, + 439 + ], + "type": "text", + "content": "Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 451, + 480, + 571 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 451, + 480, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 451, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 451, + 480, + 487 + ], + "type": "text", + "content": "- We introduce KeypointDETR, an end-to-end 3D keypoint detection framework. It is trained with a loss function based on the optimal bipartite matching derived from the cost matrix of multi-heatmaps and probabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 488, + 480, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 480, + 534 + ], + "type": "text", + "content": "- We construct a novel transformer-based network architecture. The point-wise encoder harnesses self-attention on the dynamic graph of points' local feature space to extract geometric features. The query-wise decoder incorporates global self-attention to enhance the interaction between queries." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 536, + 480, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 571 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 571 + ], + "type": "text", + "content": "- We conduct a comprehensive evaluation that validates our method and provides insights into the impact of principal components, such as the bipartite matching loss and query-wise self-attention." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 592, + 237, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 592, + 237, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 592, + 237, + 605 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 620, + 278, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 278, + 632 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 278, + 632 + ], + "type": "text", + "content": "2.1 3D Keypoint Detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "In previous studies on 3D keypoint detection [23,33,37], algorithms heavily rely on a variety of hand-crafted geometric features, such as ISS [30], HKS [34]," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "type": "text", + "content": "KeypointDETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": "CGF [14], Salient Points [3], and Mesh Saliency [16]. These traditional approaches primarily focus on local geometric features of 3D shapes, often overlooking the integration of global and semantic information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 153, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 153, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 153, + 482, + 429 + ], + "type": "text", + "content": "With advancements in deep learning, there has been a notable surge in the development of data-driven algorithms for keypoint detection [10, 35, 54]. Numerous unsupervised 3D keypoint detection algorithms [1, 7, 12, 18, 31, 50, 53] are proposed for specific tasks such as point registration and reconstruction. For example, UKPGAN [46] utilizes a GAN-based module to manage keypoint sparsity and reconstructs 3D models by distilling saliency information. Unsupervised methods tend to focus on identifying keypoints in areas with distinct geometric features. While some of these methods are capable of attaining semantic consistency, they typically require the predetermination of the number of keypoints. Furthermore, in many scenarios, it is essential to identify keypoints that may not have prominent geometric features but possess solid semantic importance. Methods like SyncSpecCNN [45] and PRA-Net [5] address keypoint detection by classifying points on 3D shapes. These methods often result in an imbalance between positive and negative samples, presenting a significant challenge. Consequently, current popular methods employ a heatmap-based approach [32, 41], where the value assigned to each point indicates its proximity to the nearest keypoint. Wei et al. [42] proposes a multi-task learning framework that combines point-to-keypoint offsets with a confidence map, facilitating effective 3D keypoint saliency and correspondence estimation. Heatmap-based methods for 3D shapes predominantly depend on a shared heatmap for all keypoints and require post-processing. These methods employ post-processing techniques like non-maximum suppression [22] or clustering [29] to isolate keypoints from regions with high scores during the inference process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 449, + 271, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 271, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 271, + 460 + ], + "type": "text", + "content": "2.2 Transformer in Vision" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": "Transformer [38] was initially introduced in the realm of natural language processing. Recent advancements in visual research [6, 19] have explored various models centered around self-attention [38], achieving unprecedented performance. DETR [2] pioneers an innovative end-to-end object detection approach. It formulates a loss matching rule employing the Hungarian algorithm [15], facilitating bipartite matching between predictions and ground truth. This method serves as the cornerstone for a series of object detection methods [4,17,55] based on set prediction. Li et al. [17] develop a facial landmark detection method using cascaded transformer decoders. They devise parallel decoders utilizing deformable self-attention [55] to optimize facial landmark coordinates through offset predictions. Chen et al. [4] leverage self-attention to create a structure-aware LSTM framework tailored for predicting 3D heatmaps in CBCT. Keypoint transformer [9] efficiently extracts all potential keypoints within the image from a shared heatmap, subsequently employing the DETR decoder to predict 3D hand poses. Given the transformer's inherent capacity to handle input element permutations in sequences, it is exceptionally well-suited for point cloud processing [8, 24, 26]." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "H.Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 482, + 292 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 292 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 292 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 292 + ], + "type": "image", + "image_path": "d3926970f31692af47b56440d3e54f96bb96946fb089ed5910f49310668f2a29.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 300, + 482, + 376 + ], + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 376 + ], + "type": "text", + "content": "Fig. 2: Pipeline of KeypointDETR for keypoint saliency and correspondence estimation. Our framework takes point coordinates as input. The point-wise encoder first extracts geometric features from the point cloud by stacking dynamic graph transformer modules (DGT). Then, the query-wise decoder accepts heatmap features, which are transformed into heatmap and query features to generate multi-heatmaps and probabilities. These predictions enable the direct inference of both keypoints positions and semantic labels." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 399, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 506 + ], + "type": "text", + "content": "Nonetheless, applying global self-attention to 3D point cloud models incurs substantial computational costs. Consequently, existing research [48, 52] frequently adopts local self-attention mechanisms, expanding receptive fields through sampling and grouping strategies. In contrast, our point-wise encoder integrates local self-attention within the dynamic graph of the point feature space, thereby eliminating the necessity for sampling and grouping operations. Distinct from most existing DETR style works, our query-wise decoder is meticulously engineered to predict multi-heatmaps and probabilities, enabling efficient end-to-end 3D keypoint detection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 522, + 202, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 202, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 202, + 534 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this section, we delve into the specifics of our end-to-end 3D keypoint detector, KeypointDETR, which consists of two essential components: a network architecture based on the transformer and a loss function formulated using bipartite matching. As illustrated in Fig. 2, our approach is applied to tasks involving keypoint saliency and correspondence estimation. Upon receiving a 3D point cloud as input, our point-wise encoder utilizes the dynamic graph transformer modules to extract heatmap features, which can be considered intrinsic geometric features. We then deploy a query-wise transformer decoder designed to enhance information exchange between queries and capture the latent relationships between attributes through cross-attention, thereby optimizing heatmap" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "type": "text", + "content": "KeypointDETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 116, + 329, + 198 + ], + "blocks": [ + { + "bbox": [ + 132, + 116, + 329, + 198 + ], + "lines": [ + { + "bbox": [ + 132, + 116, + 329, + 198 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 329, + 198 + ], + "type": "image", + "image_path": "a172b7cf989f94f61f4dfb7be805e5e4f0d0929863e53e084955cb4fe6015a29.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 199, + 331, + 217 + ], + "lines": [ + { + "bbox": [ + 132, + 199, + 331, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 199, + 331, + 217 + ], + "type": "text", + "content": "(a) The structure of the dynamic graph transformer module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 333, + 114, + 465, + 198 + ], + "blocks": [ + { + "bbox": [ + 333, + 114, + 465, + 198 + ], + "lines": [ + { + "bbox": [ + 333, + 114, + 465, + 198 + ], + "spans": [ + { + "bbox": [ + 333, + 114, + 465, + 198 + ], + "type": "image", + "image_path": "64d5d4d97f9c41e14488bef42082634eff88c72bb8921cbd0873f3538faa0e5c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 200, + 481, + 217 + ], + "lines": [ + { + "bbox": [ + 334, + 200, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 334, + 200, + 481, + 217 + ], + "type": "text", + "content": "(b) The structure of the transformer refinement module." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 226, + 479, + 248 + ], + "lines": [ + { + "bbox": [ + 131, + 226, + 479, + 248 + ], + "spans": [ + { + "bbox": [ + 131, + 226, + 479, + 248 + ], + "type": "text", + "content": "Fig. 3: The structure of the dynamic graph transformer module (DGT) and transformer refinement module." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 344 + ], + "type": "text", + "content": "and query features. This network yields heatmaps and probabilities as its outputs. The prediction sets are aligned with ground truth heatmaps through a cost matrix, leading to the computation of the final loss based on optimal bipartite matching. During the inference stage, we select heatmaps corresponding to keypoints, as indicated by the predicted probabilities, and extract keypoints exhibiting the highest heat values from these maps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 361, + 250, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 250, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 250, + 374 + ], + "type": "text", + "content": "3.1 Data Preparation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 382, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 482, + 441 + ], + "type": "text", + "content": "To tackle the issue of imbalanced samples concerning keypoints on the surface of 3D models, heatmap-based 3D keypoint detection methods establish the ground truth heat value for each point based on its distances to the keypoints. Specifically, the heatmap is generated by applying a Gaussian kernel in conjunction with the geodesic distance matrix:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 255, + 449, + 481, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 449, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 255, + 449, + 481, + 466 + ], + "type": "interline_equation", + "content": "h _ {i, j} = \\exp^ {- \\frac {1}{2 \\sigma^ {2}} \\mathrm {d} \\left(p _ {i}, k _ {j}\\right)}. \\tag {1}", + "image_path": "118368bbc826b23019478c484c3800669cc9595b8838892fcf34c9cc20ef4423.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 475, + 482, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 475, + 482, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 475, + 482, + 562 + ], + "type": "text", + "content": "Given a 3D point cloud with " + }, + { + "bbox": [ + 131, + 475, + 482, + 562 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 131, + 475, + 482, + 562 + ], + "type": "text", + "content": " points " + }, + { + "bbox": [ + 131, + 475, + 482, + 562 + ], + "type": "inline_equation", + "content": "P = \\{p_{i}\\}_{i=1}^{N_{p}}" + }, + { + "bbox": [ + 131, + 475, + 482, + 562 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 475, + 482, + 562 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 131, + 475, + 482, + 562 + ], + "type": "text", + "content": " keypoints " + }, + { + "bbox": [ + 131, + 475, + 482, + 562 + ], + "type": "inline_equation", + "content": "K = \\{k_{j}\\}_{j=1}^{N_{k}}" + }, + { + "bbox": [ + 131, + 475, + 482, + 562 + ], + "type": "text", + "content": ", our process begins with the computation of the geodesic distance matrix from each point to the keypoints. Then, we derive the heatmap matrix " + }, + { + "bbox": [ + 131, + 475, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{H} = \\{h_{i,j}\\} \\in \\mathbb{R}^{N_{p} \\times N_{k}}" + }, + { + "bbox": [ + 131, + 475, + 482, + 562 + ], + "type": "text", + "content": " using the Gaussian kernel function. Contrasting with methods that use a shared heatmap, where the maximum value for each point is considered, our approach distinctively employs multi-heatmaps, each corresponding to specific keypoints, as the ground truth." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 579, + 272, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 272, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 272, + 590 + ], + "type": "text", + "content": "3.2 Network Architecture" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 599, + 481, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 599, + 481, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 599, + 481, + 624 + ], + "type": "text", + "content": "As shown in Fig. 2, our transformer-based encoder-decoder network architecture primarily comprises a point-wise encoder and a query-wise decoder." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Point-wise Encoder The main objective of the encoder is to extract geometric features from the 3D point cloud. Drawing inspiration from DGCNN [40] and" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "H.Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "the Point Transformer [52], we devise a point-wise encoder constructed using dynamic graph transformer modules (DGT)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "content": "As depicted in Fig. 3 (a), we employ local self-attention within the dynamic graph transformer module to operate on the feature space. To be specific, for the feature map " + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "content": " at each layer, we extract a " + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "content": "-nearest neighbors feature set " + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i = \\{\\mathbf{f}_j\\}_{j=1}^k" + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "content": ". Thereafter, we assign self-attention weights to these " + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 140, + 480, + 201 + ], + "type": "text", + "content": "-nn features and aggregate them using the following equation:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 205, + 209, + 481, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 209, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 205, + 209, + 481, + 235 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {i} ^ {\\prime} = \\sum_ {\\mathbf {f} _ {j} \\in \\mathbf {F} _ {i}} \\rho (\\theta (\\phi (\\mathbf {f} _ {i}) - \\varphi (\\mathbf {f} _ {j}) + \\delta)) \\odot (\\psi (\\mathbf {f} _ {j}) + \\delta). \\tag {2}", + "image_path": "f588490a37713f01b0f47ba8c9260e3a0a34c74c0e04b533c84bb72b8e19b007.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "text", + "content": " is implemented as an MLP, while " + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "text", + "content": " are realized using Conv1d. " + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "text", + "content": " represents the softmax operation applied to self-attention weights, and " + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "text", + "content": " denotes the element-wise product operation. The term " + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "text", + "content": " represents the position encoding, which is generated from the point coordinates " + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "inline_equation", + "content": "P_{i} = \\{p_{j}\\}_{j=1}^{k}" + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "text", + "content": " corresponding to the neighborhood features, as expressed in the following:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 266, + 315, + 480, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 315, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 266, + 315, + 480, + 327 + ], + "type": "interline_equation", + "content": "\\delta = \\operatorname {M L P} \\left(p _ {i} - p _ {j}\\right). \\tag {3}", + "image_path": "5a20d08d9466db49096965282c419aea27221b6f8357d44f4adf9e9da464038c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 336, + 479, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 336, + 479, + 372 + ], + "spans": [ + { + "bbox": [ + 131, + 336, + 479, + 372 + ], + "type": "text", + "content": "Finally, we concatenate the features extracted from each dynamic graph transformer module to obtain point-wise features, which are decoded into heatmap features through MLPs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 390, + 480, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 480, + 496 + ], + "type": "text", + "content": "Query-wise Decoder The decoder's output comprises a predefined number of predictions, referred to as keypoint queries, with the query count " + }, + { + "bbox": [ + 130, + 390, + 480, + 496 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 390, + 480, + 496 + ], + "type": "text", + "content": " set significantly higher than the number of keypoints within a 3D model. Each prediction query consists of both a heatmap and a probability. Heatmaps serve to indicate the position of keypoints within the 3D model. Probabilities reflect the semantic consistency of potential keypoints. In keypoint saliency estimation, they determine the presence or absence of specific keypoints. In keypoint correspondence estimation, such probabilities are correlated with their respective semantic labels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "text", + "content": "Guided by these objectives, we develop a query-wise decoder that adheres to the proven DETR-style architecture [2], as illustrated in Fig. 3 (b). Initially, we take the heatmap features " + }, + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{heat} \\in \\mathbb{R}^{N \\times M}" + }, + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "text", + "content": " from the encoder. Considering the inherent relationship between the heatmaps and semantics of keypoints, the heatmap features are directly transposed and then decoded into query features " + }, + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{prob} \\in \\mathbb{R}^{M \\times 128}" + }, + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "text", + "content": " through an MLP. Following this, both the heatmap " + }, + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{heat}" + }, + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "text", + "content": " and query " + }, + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{prob}" + }, + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "text", + "content": " features are fed into the transformer decoder. To facilitate the exchange of information among queries, a multi-head self-attention module is applied to " + }, + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{prob}" + }, + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "text", + "content": ". For the position encoding, we extract the coordinates of the reference points corresponding to the peaks of heat values in " + }, + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{heat}" + }, + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "text", + "content": ", which are subsequently processed through an MLP to produce the position encoding. The transposed " + }, + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{heat}" + }, + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{prob}" + }, + { + "bbox": [ + 130, + 498, + 481, + 666 + ], + "type": "text", + "content": " serve as inputs to the cross-attention module, enhancing the interplay of information between probability and heatmap features. Ultimately, through the application of Feed-Forward Network modules (FFN)," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "type": "text", + "content": "KeypointDETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "the refined " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{heat}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{prob}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " are decoded into multi-heatmaps " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_h" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " and probabilities " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_p" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "content": "The decoder utilizes a query-wise self-attention mechanism across the queries, leading to " + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "content": " distinct predictions. This mechanism operates as a global self-attention centered around potential keypoints. Such a globally parallel decoding approach efficiently captures the interactions among the predicted queries, thereby optimizing heat values and reducing redundancy in the identification of keypoints. Additionally, considering the potential relationship between keypoint heatmaps and semantics, cross-attention is employed to facilitate the interaction between heatmap and query features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 252, + 231, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 252, + 231, + 263 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 231, + 263 + ], + "type": "text", + "content": "3.3 Loss Function" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 271, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 354 + ], + "type": "text", + "content": "Inspired by DETR [2], we integrate the bipartite matching loss into 3D keypoint detection to enable end-to-end inference, eliminating the need for postprocessing. During the bipartite matching phase, we calculate the pairwise matching costs by assessing the discrepancies between the predicted queries and ground truth. We utilize this cost matrix to consolidate the prediction sets, distinguishing between \"background\" queries and keypoint queries, in order to compute the final loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "text", + "content": "Given a 3D model with " + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "text", + "content": " ground truth keypoints, KeypointDETR generates a set of " + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "text", + "content": " predictions, where " + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "text", + "content": " is preconfigured to be significantly larger than " + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "text", + "content": ". The ground truth can be represented as " + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\{y_{i}\\}_{i = 1}^{N_{k}} = \\{h_{i},p_{i}\\}_{i = 1}^{N_{k}}" + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "text", + "content": " signifies the ground truth heatmap and " + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "text", + "content": " denotes the ground truth probability. The predicted queries are denoted as " + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\{\\hat{y}_j\\}_{j = 1}^M = \\{\\hat{h_j},\\hat{p_j}\\}_{j = 1}^M" + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\hat{h_j}" + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "text", + "content": " representing the predicted heatmap, and " + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\hat{p_j}" + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "text", + "content": " corresponding to the predicted probabilities associated with keypoints categories " + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\{c_i\\}_{i = 1}^{N_k}" + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "text", + "content": ". In the task of keypoint saliency estimation, " + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "text", + "content": " remains fixed at 1, while " + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\hat{p_j}(c_i)" + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "text", + "content": " signifies the probability of being predicted as a keypoint. However, in the task of keypoint correspondence estimation, " + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\hat{p_j}(c_i)" + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "text", + "content": " is defined as the probability of " + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "text", + "content": " denotes the semantic label corresponding to the ground truth keypoint " + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "inline_equation", + "content": "k_i" + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "text", + "content": ". Expanding upon these notations, in order to achieve the optimal bipartite matching between ground truth and predictions, it is imperative to establish a matching cost matrix " + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c \\in \\mathbb{R}^{N_k \\times M}" + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "text", + "content": ". This cost matrix takes into account both the heatmaps and probabilities. Precisely, each element " + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c(i,j)" + }, + { + "bbox": [ + 130, + 356, + 482, + 537 + ], + "type": "text", + "content": " of the cost matrix can be calculated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 188, + 545, + 481, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 545, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 188, + 545, + 481, + 563 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {c} (i, j) = \\mathbf {W} _ {\\mathbf {h}} \\cdot \\sum | h _ {i} - \\hat {h} _ {j} | + \\mathbf {W} _ {\\mathbf {p}} \\cdot \\left(1 - \\log p _ {\\hat {j}} \\left(c _ {i}\\right)\\right), \\tag {4}", + "image_path": "5ca52304197918503ec2a7af6587c0cf8db277b8176876cd5c47b88639db2320.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 567, + 480, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 480, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 567, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{W_h}" + }, + { + "bbox": [ + 130, + 567, + 480, + 605 + ], + "type": "text", + "content": " represents the weight assigned to heatmaps, and " + }, + { + "bbox": [ + 130, + 567, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{W_p}" + }, + { + "bbox": [ + 130, + 567, + 480, + 605 + ], + "type": "text", + "content": " is the weight attributed to probabilities. " + }, + { + "bbox": [ + 130, + 567, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\sum |h_i - \\hat{h_j}|" + }, + { + "bbox": [ + 130, + 567, + 480, + 605 + ], + "type": "text", + "content": " signifies the dissimilarity between the ground truth " + }, + { + "bbox": [ + 130, + 567, + 480, + 605 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 130, + 567, + 480, + 605 + ], + "type": "text", + "content": " and predicted " + }, + { + "bbox": [ + 130, + 567, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\hat{h_j}" + }, + { + "bbox": [ + 130, + 567, + 480, + 605 + ], + "type": "text", + "content": " heatmaps." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "Following the application of the Hungarian algorithm [15], we obtain the optimal matching and categorize the predicted queries into two groups: positive queries and negative queries. Positive queries are those that align with the ground truth. For a positive query " + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": ", we determine the matched ground truth category as " + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "inline_equation", + "content": "g_i" + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": " and the corresponding heatmap as " + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": ". Then, we compute the probability" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "H.Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "loss using the Focal Loss (FL) and the heatmap loss employing the Mean Squared Error Loss (MSE). For a negative query, only the probability loss is considered. The associated ground truth " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " is designated as 0, signifying nonexistence. Thus, the overall loss can be calculated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 205, + 170, + 481, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 170, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 205, + 170, + 481, + 200 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {i = 1} ^ {M} \\left(\\lambda_ {p} \\operatorname {F L} \\left(g _ {i}, \\hat {p _ {i}}\\right) + \\lambda_ {h} \\mathbb {1} _ {\\{g _ {i} \\neq 0 \\}} \\operatorname {M S E} \\left(h _ {i}, \\hat {h _ {i}}\\right)\\right). \\tag {5}", + "image_path": "21969b0ebe596049bf1a1ddeac9e1934d3a1709d55277381691e9da5949926ab.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 216, + 230, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 230, + 229 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 230, + 229 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 239, + 479, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 479, + 286 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 479, + 286 + ], + "type": "text", + "content": "In this section, we conduct a series of comprehensive experiments to validate the superior performance of our approach in both keypoint saliency and correspondence estimation tasks. We provide comparisons with other state-of-the-art methods to demonstrate our method's efficacy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 303, + 279, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 279, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 279, + 315 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 321, + 480, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 480, + 440 + ], + "type": "text", + "content": "KeypointDETR takes 3D point coordinates as input. In the point-wise encoder, we utilize 5 dynamic graph transformer modules, each with point features of 64 dimensions, the " + }, + { + "bbox": [ + 130, + 321, + 480, + 440 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 321, + 480, + 440 + ], + "type": "text", + "content": " value of " + }, + { + "bbox": [ + 130, + 321, + 480, + 440 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 321, + 480, + 440 + ], + "type": "text", + "content": "-nearest neighbors set to 20, and a global embedding dimension of 1024. In the query-wise decoder, we set the number of queries " + }, + { + "bbox": [ + 130, + 321, + 480, + 440 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 321, + 480, + 440 + ], + "type": "text", + "content": " to 50. The decoder's output comprises two branches: the heatmap head generates multi-heatmaps " + }, + { + "bbox": [ + 130, + 321, + 480, + 440 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_h \\in \\mathbb{R}^{M \\times N_p}" + }, + { + "bbox": [ + 130, + 321, + 480, + 440 + ], + "type": "text", + "content": ", and the classification head provides probabilities " + }, + { + "bbox": [ + 130, + 321, + 480, + 440 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_p \\in \\mathbb{R}^{M \\times C}" + }, + { + "bbox": [ + 130, + 321, + 480, + 440 + ], + "type": "text", + "content": ". For keypoint saliency estimation, " + }, + { + "bbox": [ + 130, + 321, + 480, + 440 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 321, + 480, + 440 + ], + "type": "text", + "content": " is held constant at 2. However, for keypoint correspondence estimation, " + }, + { + "bbox": [ + 130, + 321, + 480, + 440 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 321, + 480, + 440 + ], + "type": "text", + "content": " is configured as " + }, + { + "bbox": [ + 130, + 321, + 480, + 440 + ], + "type": "inline_equation", + "content": "(N_c + 1)" + }, + { + "bbox": [ + 130, + 321, + 480, + 440 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 321, + 480, + 440 + ], + "type": "inline_equation", + "content": "N_c" + }, + { + "bbox": [ + 130, + 321, + 480, + 440 + ], + "type": "text", + "content": " signifies the number of keypoint categories. More details can be found in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 456, + 200, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 200, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 200, + 467 + ], + "type": "text", + "content": "4.2 Dataset" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 475, + 479, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 479, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 479, + 558 + ], + "type": "text", + "content": "Our experiments are conducted on the KeypointNet dataset [47], which is designed for both keypoint saliency and correspondence estimation tasks. We focus on four classical categories: airplane, chair, guitar, and table. These models are divided into training, validation, and test sets, with a distribution ratio of 7:1:2 for each category. We use 2,048 sampled points from each 3D model as input data. The results of other categories and applications of KeypointDETR can be found in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 574, + 200, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 200, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 200, + 586 + ], + "type": "text", + "content": "4.3 Metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 593, + 479, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 479, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 479, + 639 + ], + "type": "text", + "content": "We evaluate our method alongside comparison methods on the keypoint saliency estimation using mIoU and Chamfer Distance (CD) while utilizing mIoU and Dual Alignment Score (DAS) [31] for keypoint correspondence estimation. The IoU is calculated as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 253, + 644, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 644, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 253, + 644, + 480, + 667 + ], + "type": "interline_equation", + "content": "I o U = \\frac {T P}{T P + F P + F N}, \\tag {6}", + "image_path": "99ce16f87c6149ba973a43bbcfb5859dd7476763649813caef5662397c0e27d5.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "type": "text", + "content": "KeypointDETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "TP" + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": " stands for true positives, " + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "FP" + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": " represents false positives, and " + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "FN" + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": " corresponds to false negatives. Given the ground truth and predicted keypoints sets " + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": ", CD can be computed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 187, + 162, + 481, + 191 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 162, + 481, + 191 + ], + "spans": [ + { + "bbox": [ + 187, + 162, + 481, + 191 + ], + "type": "interline_equation", + "content": "C D = \\frac {1}{| G |} \\sum_ {p \\in G} \\min _ {q \\in K} \\| p - q \\| _ {2} ^ {2} + \\frac {1}{| K |} \\sum_ {p \\in K} \\min _ {q \\in G} \\| p - q \\| _ {2} ^ {2}. \\tag {7}", + "image_path": "ced7d336cc94c019295fb801cd461671adfc31c3bca99667a47f8f88575280a8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 201, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 357 + ], + "type": "text", + "content": "The settings for clustering and NMS parameters profoundly affect the repeatability and omission rates of predicted keypoints. However, the mIoU metric used in prior methods fails to effectively capture these erroneous cases. To ensure a comprehensive comparison, we design a mIoU metric based on the Hungarian algorithm [15]. For the keypoint saliency estimation, we begin by calculating the Euclidean distance matrix between the ground truth and the predicted keypoints. Subsequently, the Hungarian algorithm is employed to determine their optimal bipartite matching. Each matched distance falling below the predefined threshold " + }, + { + "bbox": [ + 130, + 201, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 201, + 482, + 357 + ], + "type": "text", + "content": " is considered a true positive. For keypoint correspondence estimation, a ground truth keypoint is considered a true positive if there is a predicted keypoint with the same semantic label whose distance to it is less than the threshold " + }, + { + "bbox": [ + 130, + 201, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 201, + 482, + 357 + ], + "type": "text", + "content": ". Additionally, we further measure whether the closest keypoints between predictions and ground truth have the same semantic labels using DAS [31]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 378, + 390, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 378, + 390, + 391 + ], + "spans": [ + { + "bbox": [ + 131, + 378, + 390, + 391 + ], + "type": "text", + "content": "4.4 Comparison on Keypoint Saliency Estimation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 401, + 497, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 497, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 497, + 497 + ], + "type": "text", + "content": "We conduct a comparative analysis of our method against six advanced algorithms, including two unsupervised keypoint detection methods (Skeleton Merger [31] and UKPGAN [46]) and three point cloud geometric feature extraction algorithms (PointNet++ [28], PRA-Net [5], and DGCNN [40]), and a fully-supervised keypoint detection method (Wei et al. [42]). For the three point cloud geometric feature extraction algorithms on saliency analysis, the network's output is a shared heatmap, wherein each point is assigned the maximum value of the multi-heatmaps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "The quantitative results for mIoU and Chamfer Distance (CD) are presented in Tab. 1, where we compare our KeypointDETR with other algorithms using a threshold " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " of 0.1. Fig. 6 (a) depicts the mIoU curves for various methods as " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " varies from 0 to 0.1. It is evident that KeypointDETR consistently outperforms other methods in the Hungarian mIoU evaluation across all categories. This significant advantage stems from our method's capacity to bypass post-processing techniques like NMS or clustering algorithms, leading to direct 'one-to-one' predictions that effectively tackle redundancy and omissions in 3D keypoint detection. In contrast, methods reliant on a shared heatmap face challenges with parameter selection of post-processing, which predominantly depends on prior knowledge factors such as the expected number of keypoints and the 3D model's scale. Unfortunately, even optimal parameter settings for NMS (radius=0.1) in these methods do not entirely resolve issues of omissions and redundant predictions, as further illustrated in Fig. 4." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "H.Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 146, + 487, + 269 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 1: Quantitative results of keypoint saliency estimation using Hungarian mIoU (\\%) and CD." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 146, + 487, + 269 + ], + "lines": [ + { + "bbox": [ + 132, + 146, + 487, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 487, + 269 + ], + "type": "table", + "html": "
AirplaneChairGuitarTableAverage
mIoUCDmIoUCDmIoUCDmIoUCDmIoUCD
SkeletonMerger [31]60.730.10742.740.18947.710.13227.840.30244.750.182
UKPGAN [46]62.630.11825.160.16854.250.13117.780.24039.950.164
PRA-Net [5]82.880.04970.440.06073.710.05369.690.05374.180.053
PointNet++ [28]82.310.05057.500.07772.340.06060.010.06268.040.062
DGCNN [40]78.490.05859.490.07663.880.07259.920.06665.440.068
Wei et al. [42]79.860.04886.280.05578.080.06686.610.03982.700.052
Ours91.370.04591.890.05593.960.04598.050.02793.810.043
", + "image_path": "e9d4af2281c9871703b8ec87417582c1043b1bfc8d8ed5e64fc4de01820bfea3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 132, + 276, + 482, + 437 + ], + "blocks": [ + { + "bbox": [ + 132, + 276, + 482, + 437 + ], + "lines": [ + { + "bbox": [ + 132, + 276, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 132, + 276, + 482, + 437 + ], + "type": "image", + "image_path": "da1962fca4c8e3590bda244efc3a23c4c0223b2aec5bd070bbb4e446bf948610.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 177, + 445, + 434, + 456 + ], + "lines": [ + { + "bbox": [ + 177, + 445, + 434, + 456 + ], + "spans": [ + { + "bbox": [ + 177, + 445, + 434, + 456 + ], + "type": "text", + "content": "Fig. 4: Visualization examples of keypoint saliency estimation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 479, + 429, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 429, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 429, + 491 + ], + "type": "text", + "content": "4.5 Comparison on Keypoint Correspondence Estimation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 605 + ], + "type": "text", + "content": "Keypoint correspondence estimation represents a more significant challenge compared to keypoint saliency estimation, as it not only involves locating keypoints but also predicting their semantic labels. Additionally, 3D models within the same category may not have identical semantic keypoints. Prior approaches [35, 45,47] typically involve direct classification of points within a point cloud, which does not effectively address the issue of imbalanced positive and negative samples in keypoint detection. Some methods [42, 54] attempt to add a semantic branch to keypoint saliency estimation, aiming to assign labels to keypoints in a two-stage process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "As depicted in Tab. 2, we conduct a comparison of our method with five methods using a threshold " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " of 0.1: PointConv [43], PointNet++ [28], DGCNN [40], PRA-Net [5], and Wei et al. [42]. Additionally, we present the mIoU curves for various methods, with the " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " ranging from 0 to 0.1, in Fig. 6 (b). Fig. 5 showcases visualization results of different methods alongside the ground truth. Notably," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "type": "text", + "content": "KeypointDETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 146, + 479, + 256 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 2: Quantitative results of keypoint correspondence estimation using Hungarian mIoU (%) and DAS [31] (%)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 146, + 479, + 256 + ], + "lines": [ + { + "bbox": [ + 135, + 146, + 479, + 256 + ], + "spans": [ + { + "bbox": [ + 135, + 146, + 479, + 256 + ], + "type": "table", + "html": "
AirplaneChairGuitarTableAverage
mIoUDASmIoUDASmIoUDASmIoUDASmIoUDAS
DGCNN [40]77.7782.6963.1679.5661.1378.5957.9978.8365.0179.91
PointNet++ [28]79.1084.7458.2478.1967.6981.9262.9781.5967.0081.61
PointConv [43]81.5485.7464.9282.1576.3988.9166.2083.7372.2685.13
PRA-Net [5]81.1085.8068.3583.6272.0785.1570.1684.8972.9284.86
Wei et al. [42]78.5488.3479.7790.4475.6493.6986.5697.3180.1292.44
Ours85.7191.7479.3389.5390.9393.8992.4097.7987.0993.23
", + "image_path": "9d66a3c820742232df2df361abe05de6d2549eccc35075af2254cfcb99ad8eaa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 264, + 482, + 403 + ], + "blocks": [ + { + "bbox": [ + 133, + 264, + 482, + 403 + ], + "lines": [ + { + "bbox": [ + 133, + 264, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 133, + 264, + 482, + 403 + ], + "type": "image", + "image_path": "0b90e31dcf5ceaea378fae405d2c9b7f7502fbbac83dac648efd28d13eddb5bc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 411, + 449, + 422 + ], + "lines": [ + { + "bbox": [ + 162, + 411, + 449, + 422 + ], + "spans": [ + { + "bbox": [ + 162, + 411, + 449, + 422 + ], + "type": "text", + "content": "Fig. 5: Visualization examples of keypoint correspondence estimation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 445, + 480, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 480, + 493 + ], + "type": "text", + "content": "our method generates no redundant points with the same label in close proximity to predicted keypoints. This demonstrates that our method adeptly accomplishes semantic correspondence for keypoints, with KeypointDETR showcasing superior performance in both the heatmap and probability branch." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 510, + 239, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 239, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 239, + 521 + ], + "type": "text", + "content": "4.6 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 528, + 479, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 479, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 479, + 588 + ], + "type": "text", + "content": "We conduct a series of ablation experiments to validate the performance of KeypointDETR and deeply explore the impact of each component. Throughout these experiments, we consistently use four representative categories from the KeypointNet dataset for keypoint saliency estimation. The evaluation adopts mIoU with a threshold " + }, + { + "bbox": [ + 130, + 528, + 479, + 588 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 528, + 479, + 588 + ], + "type": "text", + "content": " of 0.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Effects of point-wise encoder. Our point-wise encoder primarily consists of dynamic graph transformer modules that extract geometric features from the point cloud. To validate the effectiveness of our point-wise encoder, we compare KeypointDETR with its variants, which utilize different backbones for geometric feature extraction, namely DGCNN [40], PointNet++ [28], and Point Transformer [52]. The DGCNN [40] backbone (DG.Enc.) is implemented as a five-layer" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "H.Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 279, + 225 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 279, + 225 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 279, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 279, + 225 + ], + "type": "image", + "image_path": "5272ec86e824043d21ad877d71e1dc6aa7d8a4ce5f80e266cd8b661fb74c95e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 227, + 290, + 244 + ], + "lines": [ + { + "bbox": [ + 133, + 227, + 290, + 244 + ], + "spans": [ + { + "bbox": [ + 133, + 227, + 290, + 244 + ], + "type": "text", + "content": "(a) Hungarian mIoU curves for keypoint saliency estimation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 322, + 114, + 470, + 225 + ], + "blocks": [ + { + "bbox": [ + 322, + 114, + 470, + 225 + ], + "lines": [ + { + "bbox": [ + 322, + 114, + 470, + 225 + ], + "spans": [ + { + "bbox": [ + 322, + 114, + 470, + 225 + ], + "type": "image", + "image_path": "6638a544fd8bea7b654de57d2c5544416436791a87ad1688af1823ae8fa958c7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 227, + 480, + 244 + ], + "lines": [ + { + "bbox": [ + 323, + 227, + 480, + 244 + ], + "spans": [ + { + "bbox": [ + 323, + 227, + 480, + 244 + ], + "type": "text", + "content": "(b) Hungarian mIoU curves for keypoint correspondence estimation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 253, + 480, + 276 + ], + "lines": [ + { + "bbox": [ + 132, + 253, + 480, + 276 + ], + "spans": [ + { + "bbox": [ + 132, + 253, + 480, + 276 + ], + "type": "text", + "content": "Fig. 6: Hungarian mIoU curves for keypoint saliency and correspondence estimation under various distance thresholds (0-0.1)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 149, + 319, + 465, + 457 + ], + "blocks": [ + { + "bbox": [ + 132, + 287, + 480, + 309 + ], + "lines": [ + { + "bbox": [ + 132, + 287, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 132, + 287, + 480, + 309 + ], + "type": "text", + "content": "Table 3: Quantitative results of various variations in the ablation study using Hungarian mIoU (%) and CD." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 319, + 465, + 457 + ], + "lines": [ + { + "bbox": [ + 149, + 319, + 465, + 457 + ], + "spans": [ + { + "bbox": [ + 149, + 319, + 465, + 457 + ], + "type": "table", + "html": "
AirplaneChairGuitarTableAverage
mIoUCDmIoUCDmIoUCDmIoUCDmIoUCD
DG.Enc. [40]86.750.04882.120.06292.760.05297.620.03789.810.049
PN2.Enc. [28]76.300.07168.310.10172.090.09976.420.09973.280.092
PT.Enc. [52]82.550.05973.040.08568.920.09197.320.03280.450.066
Point-BERT Enc. [49]91.110.03493.570.03493.510.03293.710.03392.880.033
I2P-MAE Enc. [51]79.060.05080.400.06788.790.04886.050.04883.570.053
Point-MAE Enc. [25]89.820.03993.980.05583.790.04395.950.04390.880.045
w/o Trans.Dec.87.960.04488.230.05892.740.04597.770.02991.670.044
w/o PE.88.050.04291.430.05492.440.04797.970.02892.470.042
Coord-based66.720.10844.550.19037.990.16450.080.18749.830.162
Coord+offsets68.400.09463.620.13642.450.15160.770.15958.810.135
Ours91.370.04591.890.05593.960.04598.050.02793.810.043
", + "image_path": "797800f87d2e26fb50c5c2c58126a5c65c29821691182de8cdcccc4c0c69eaab.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 479, + 480, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 480, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 480, + 539 + ], + "type": "text", + "content": "stack of EdgeConv. PointNet++ [28] (PN2.Enc.) and Point Transformer [52] (PT.Enc.) employ eight layers due to their U-net structures. Furthermore, we explore replacing our encoder with pre-trained models, including Point-BERT, Point-MAE, and I2P-MAE, and achieve favorable results. The overall quantitative comparison results are presented in the first sub-table of Tab. 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "text", + "content": "Effects of query-wise decoder. In the decoder phase, we develop a query-wise decoder utilizing self-attention to facilitate enhanced information exchange among queries. This decoder is tasked with decoding the heatmap and probability feature maps, resulting in the generation of multi-heatmaps and probabilities. A straightforward approach involves employing MLPs to decode these feature maps separately. However, to demonstrate the efficacy and reasoning behind our query-wise decoder design, we conduct an experiment where we replace our self-attention decoder (Trans.Dec.) with MLPs. The comparison results, as shown in the second sub-table of Tab. 3, indicate that the comprehensive performance of our KeypointDETR significantly surpasses that of its MLPs decoder variant." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "type": "text", + "content": "KeypointDETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 481, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 481, + 200 + ], + "type": "text", + "content": "Effects of position encoding in decoder. Position encoding holds a pivotal role in the self-attention mechanism and is a key factor that makes the transformer architecture suitable for 3D point cloud models, as previously validated in the Point Transformer [52]. In our investigation, we further explore the role of position encoding in our query-wise decoder, which is generated from heatmap features and point coordinates. We compare the results with and without position encoding (PE.) in the third sub-table of Tab. 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 209, + 481, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 209, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 481, + 376 + ], + "type": "text", + "content": "Coordinate-based or Heatmap-based. We further explore a coordinate-based architecture within our KeypointDETR approach. In this variant, KeypointDETR predicts the coordinates and probabilities of keypoints, and we subsequently select the closest matching points from the point cloud as the final keypoints. To enhance the accuracy of these coordinates, we incorporate a cascade module designed to predict offsets. Comparative analysis of the heatmap-based KeypointDETR and its coordinate-based variant, as shown in the fourth sub-table of Tab. 3, reveals that while the addition of offsets improves the results of the coordinate-based model, there remains a notable performance gap when compared to our primary heatmap-based KeypointDETR. This discrepancy arises because our geometric feature extraction backbone is better suited for processing 3D shapes rather than 3D spatial coordinates. Furthermore, it validates that the heatmap-based approach can achieve better performance on keypoints detection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 403, + 219, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 219, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 219, + 415 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 437, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 481, + 569 + ], + "type": "text", + "content": "In this paper, we introduce KeypointDETR, an end-to-end 3D keypoint detector that obviates the need for post-processing steps, such as clustering and NMS, commonly required in heatmap-based keypoint detection methods. The training strategy of KeypointDETR involves computing a cost matrix between the predictions and ground truth, deriving an optimal bipartite matching loss. This innovative approach enables KeypointDETR to accurately predict multi-heatmaps and probabilities, which can provide the positions and semantic consistency of potential keypoints, respectively. Complementing this framework is a transformer-based network architecture. Extensive quantitative experiments and analyses substantiate the effectiveness and rationality of KeypointDETR, underlining its potential as a state-of-the-art solution in keypoint detection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 595, + 247, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 595, + 247, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 595, + 247, + 609 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 481, + 666 + ], + "type": "text", + "content": "This work is partially supported by the following grants: National Natural Science Foundation of China (No. 62172363), Management Center of the School of Software (Ningbo) at Zhejiang University under Grant (No. Z24001)." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "H.Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 140, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 481, + 172 + ], + "type": "text", + "content": "1. Bai, X., Luo, Z., Zhou, L., Fu, H., Quan, L., Tai, C.L.: D3feat: Joint learning of dense detection and description of 3d local features. In: CVPR. pp. 6359-6367 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 481, + 195 + ], + "type": "text", + "content": "2. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV. pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 195, + 481, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 481, + 227 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 481, + 227 + ], + "type": "text", + "content": "3. Castellani, U., Cristani, M., Fantoni, S., Murino, V.: Sparse points matching by combining 3d mesh saliency with statistical descriptors. In: Comput. Graph. Forum. vol. 27, pp. 643-652. Wiley Online Library (2008)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 481, + 259 + ], + "type": "text", + "content": "4. Chen, R., Ma, Y., Chen, N., Liu, L., Cui, Z., Lin, Y., Wang, W.: Structure-aware long short-term memory network for 3d cephalometric landmark detection. IEEE Transactions on Medical Imaging 41(7), 1791-1801 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 261, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 481, + 281 + ], + "type": "text", + "content": "5. Cheng, S., Chen, X., He, X., Liu, Z., Bai, X.: Pra-net: Point relation-aware network for 3d point cloud analysis. IEEE TIP 30, 4436-4448 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 282, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 282, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 282, + 481, + 326 + ], + "type": "text", + "content": "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 326, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 481, + 358 + ], + "type": "text", + "content": "7. Fernandez-Labrador, C., Chhatkuli, A., Paudel, D.P., Guerrero, J.J., Demonceaux, C., Gool, L.V.: Unsupervised learning of category-specific symmetric 3d keypoints from point sets. In: ECCV. pp. 546-563. Springer (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 358, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 358, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 358, + 481, + 380 + ], + "type": "text", + "content": "8. Guo, M.H., Cai, J.X., Liu, Z.N., Mu, T.J., Martin, R.R., Hu, S.M.: Pct: Point cloud transformer. Computational Visual Media 7, 187-199 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 381, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 481, + 413 + ], + "type": "text", + "content": "9. Hampali, S., Sarkar, S.D., Rad, M., Lepetit, V.: Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In: CVPR. pp. 11090-11100 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 414, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 481, + 446 + ], + "type": "text", + "content": "0. He, Y., Sun, W., Huang, H., Liu, J., Fan, H., Sun, J.: Pvn3d: A deep point-wise 3d keypoints voting network for 6 dof pose estimation. In: CVPR. pp. 11632-11641 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 447, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 481, + 468 + ], + "type": "text", + "content": "1. Huang, Q., Wang, W., Neumann, U.: Recurrent slice networks for 3d segmentation of point clouds. In: CVPR. pp. 2626-2635 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 481, + 501 + ], + "type": "text", + "content": "2. Jakab, T., Tucker, R., Makadia, A., Wu, J., Snavely, N., Kanazawa, A.: Keypoint-deformer: Unsupervised 3d keypoint discovery for shape control. In: CVPR. pp. 12783-12792 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 501, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 481, + 523 + ], + "type": "text", + "content": "3. Katz, S., Leifman, G., Tal, A.: Mesh segmentation using feature point and core extraction. The Vis. Comput. 21, 649-658 (2005)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 523, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 481, + 544 + ], + "type": "text", + "content": "4. Khoury, M., Zhou, Q.Y., Koltun, V.: Learning compact geometric features. In: ICCV. pp. 153-161 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 545, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 481, + 567 + ], + "type": "text", + "content": "5. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 567, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 589 + ], + "type": "text", + "content": "6. Lee, C.H., Varshney, A., Jacobs, D.W.: Mesh saliency. In: ACM SIGGRAPH 2005 Papers, pp. 659-666 (2005)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 589, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 481, + 610 + ], + "type": "text", + "content": "7. Li, H., Guo, Z., Rhee, S.M., Han, S., Han, J.J.: Towards accurate facial landmark detection via cascaded transformers. In: CVPR. pp. 4176-4185 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "text", + "content": "8. Li, J., Lee, G.H.: Usip: Unsupervised stable interest point detection from 3d point clouds. In: ICCV. pp. 361-370 (2019)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "9. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "type": "text", + "content": "KeypointDETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "20. Liu, Z., Wu, Z., Toth, R.: Smoke: Single-stage monocular 3d object detection via keypoint estimation. In: CVPR. pp. 996-997 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "text", + "content": "21. Mian, A.S., Bennamoun, M., Owens, R.: Three-dimensional model-based object recognition and segmentation in cluttered scenes. IEEE TPAMI 28(10), 1584-1601 (2006)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 194 + ], + "type": "text", + "content": "22. Neubeck, A., Van Gool, L.: Efficient non-maximum suppression. In: ICPR. vol. 3, pp. 850-855. IEEE (2006)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 216 + ], + "type": "text", + "content": "23. Novatnack, J., Nishino, K.: Scale-dependent 3d geometric features. In: ICCV. pp. 1-8. IEEE (2007)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "type": "text", + "content": "24. Pan, X., Xia, Z., Song, S., Li, L.E., Huang, G.: 3d object detection with pointformer. In: CVPR. pp. 7463-7472 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 237, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 237, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 480, + 259 + ], + "type": "text", + "content": "25. Pang, Y., Wang, W., Tay, F.E., Liu, W., Tian, Y., Yuan, L.: Masked autoencoders for point cloud self-supervised learning. In: ECCV. pp. 604-621. Springer (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 480, + 281 + ], + "type": "text", + "content": "26. Park, C., Jeong, Y., Cho, M., Park, J.: Fast point transformer. In: CVPR. pp. 16949-16958 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 281, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 281, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 133, + 281, + 480, + 304 + ], + "type": "text", + "content": "27. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR. pp. 652-660 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 133, + 304, + 480, + 325 + ], + "type": "text", + "content": "28. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS 30 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 326, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 326, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 133, + 326, + 480, + 346 + ], + "type": "text", + "content": "29. Rodriguez, A., Laio, A.: Clustering by fast search and find of density peaks. science 344(6191), 1492-1496 (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 133, + 347, + 480, + 369 + ], + "type": "text", + "content": "30. Salti, S., Tombari, F., Spezialietti, R., Di Stefano, L.: Learning a descriptor-specific 3d keypoint detector. In: ICCV. pp. 2318-2326 (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 133, + 369, + 480, + 391 + ], + "type": "text", + "content": "31. Shi, R., Xue, Z., You, Y., Lu, C.: Skeleton merger: an unsupervised aligned keypoint detector. In: CVPR. pp. 43-52 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 133, + 392, + 480, + 413 + ], + "type": "text", + "content": "32. Shu, Z., Yu, J., Chao, K., Xin, S., Liu, L.: A multi-modal attention-based approach for points of interest detection on 3d shapes. IEEE TVCG (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 414, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 414, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 133, + 414, + 480, + 435 + ], + "type": "text", + "content": "33. Sipiran, I., Bustos, B.: Harris 3d: a robust extension of the harris operator for interest point detection on 3d meshes. The Vis. Comput. 27, 963-976 (2011)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 133, + 435, + 480, + 468 + ], + "type": "text", + "content": "34. Sun, J., Ovsjanikov, M., Guibas, L.: A concise and provably informative multiscale signature based on heat diffusion. In: Comput. Graph. Forum. vol. 28, pp. 1383-1392. Wiley Online Library (2009)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 133, + 468, + 480, + 490 + ], + "type": "text", + "content": "35. Sung, M., Su, H., Yu, R., Guibas, L.J.: Deep functional dictionaries: Learning consistent semantic structures on 3d models from functions. NeurIPS 31 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 133, + 491, + 480, + 511 + ], + "type": "text", + "content": "36. Tang, J., Gong, Z., Yi, R., Xie, Y., Ma, L.: Lake-net: Topology-aware point cloud completion by localizing aligned keypoints. In: CVPR. pp. 1726-1735 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 133, + 512, + 480, + 533 + ], + "type": "text", + "content": "37. Tombari, F., Salti, S., Di Stefano, L.: Unique signatures of histograms for local surface description. In: ECCV. pp. 356-369. Springer (2010)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 133, + 534, + 480, + 555 + ], + "type": "text", + "content": "38. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 133, + 555, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 555, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 133, + 555, + 480, + 578 + ], + "type": "text", + "content": "39. Wang, H., Guo, J., Yan, D.M., Quan, W., Zhang, X.: Learning 3d keypoint descriptors for non-rigid shape matching. In: ECCV. pp. 3-19 (2018)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 133, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 133, + 578, + 480, + 599 + ], + "type": "text", + "content": "40. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. ACM TOG 38(5), 1-12 (2019)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 133, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 600, + 480, + 632 + ], + "type": "text", + "content": "41. Wei, G., Cui, Z., Zhu, J., Yang, L., Zhou, Y., Singh, P., Gu, M., Wang, W.: Dense representative tooth landmark/axis detection network on 3d model. Computer Aided Geometric Design 94, 102077 (2022)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "text", + "content": "42. Wei, G., Ma, L., Wang, C., Desrosiers, C., Zhou, Y.: Multi-task joint learning of 3d keypoint saliency and correspondence estimation. Computer-Aided Design 141, 103105 (2021)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 213, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 213, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 213, + 100 + ], + "type": "text", + "content": "H.Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 456 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "43. Wu, W., Qi, Z., Fuxin, L.: Pointconv: Deep convolutional networks on 3d point clouds. In: CVPR. pp. 9621-9630 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "text", + "content": "44. Yew, Z.J., Lee, G.H.: 3dfeat-net: Weakly supervised local 3d features for point cloud registration. In: ECCV. pp. 607-623 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 182 + ], + "type": "text", + "content": "45. Yi, L., Su, H., Guo, X., Guibas, L.J.: Syncspeccnn: Synchronized spectral cnn for 3d shape segmentation. In: CVPR. pp. 2282-2290 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "type": "text", + "content": "46. You, Y., Liu, W., Ze, Y., Li, Y.L., Wang, W., Lu, C.: Ukpgan: A general self-supervised keypoint detector. In: CVPR. pp. 17042-17051 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "text", + "content": "47. You, Y., Lou, Y., Li, C., Cheng, Z., Li, L., Ma, L., Lu, C., Wang, W.: Keypointnet: A large-scale 3d keypoint dataset aggregated from numerous human annotations. In: CVPR. pp. 13647-13656 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "text", + "content": "48. Yu, X., Rao, Y., Wang, Z., Liu, Z., Lu, J., Zhou, J.: Pointr: Diverse point cloud completion with geometry-aware transformers. In: ICCV. pp. 12498-12507 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 259, + 482, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 292 + ], + "type": "text", + "content": "49. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In: CVPR. pp. 19313-19322 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 293, + 482, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 324 + ], + "type": "text", + "content": "50. Yuan, H., Zhao, C., Fan, S., Jiang, J., Yang, J.: Unsupervised learning of 3d semantic keypoints with mutual reconstruction. In: ECCV. pp. 534-549. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 325, + 482, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 357 + ], + "type": "text", + "content": "51. Zhang, R., Wang, L., Qiao, Y., Gao, P., Li, H.: Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In: CVPR. pp. 21769-21780 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 358, + 482, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 379 + ], + "type": "text", + "content": "52. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV. pp. 16259-16268 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 380, + 482, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 402 + ], + "type": "text", + "content": "53. Zhong, C., You, P., Chen, X., Zhao, H., Sun, F., Zhou, G., Mu, X., Gan, C., Huang, W.: Snake: Shape-aware neural 3d keypoint field. NeurIPS 35, 7052-7064 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 402, + 482, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 424 + ], + "type": "text", + "content": "54. Zhu, X., Du, D., Huang, H., Ma, C., Han, X.: 3d keypoint estimation using implicit representation learning. arXiv preprint arXiv:2306.11529 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 424, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 456 + ], + "type": "text", + "content": "55. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 381, + 91, + 447, + 102 + ], + "type": "text", + "content": "KeypointDETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Kinetic Typography Diffusion Model/d7a51e44-9061-4ac8-9c76-37fd7685fe8a_content_list.json b/2024/Kinetic Typography Diffusion Model/d7a51e44-9061-4ac8-9c76-37fd7685fe8a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..63af23438a81199184dfc209532d12fc9cee8486 --- /dev/null +++ b/2024/Kinetic Typography Diffusion Model/d7a51e44-9061-4ac8-9c76-37fd7685fe8a_content_list.json @@ -0,0 +1,2490 @@ +[ + { + "type": "text", + "text": "Kinetic Typography Diffusion Model", + "text_level": 1, + "bbox": [ + 290, + 141, + 710, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seonmi Park, Inhwan Bae, Seunghyun Shin, and Hae-Gon Jeon", + "bbox": [ + 243, + 189, + 754, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "AI Graduate School, GIST, South Korea \n{bluesky1000, inhwanbae, seunghyuns98}@gm.gist.ac.kr, haegonj@gist.ac.kr \nhttps://seonmip.github.io/kinety", + "bbox": [ + 220, + 217, + 777, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. This paper introduces a method for realistic kinetic typography that generates user-preferred animatable \"text content\". We draw on recent advances in guided video diffusion models to achieve visually-pleasing text appearances. To do this, we first construct a kinetic typography dataset, comprising about 600K videos. Our dataset is made from a variety of combinations in 584 templates designed by professional motion graphics designers and involves changing each letter's position, glyph, and size (i.e., flying, glitches, chromatic aberration, reflecting effects, etc.). Next, we propose a video diffusion model for kinetic typography. For this, there are three requirements: aesthetic appearances, motion effects, and readable letters. This paper identifies the requirements. For this, we present static and dynamic captions used as spatial and temporal guidance of a video diffusion model, respectively. The static caption describes the overall appearance of the video, such as colors, texture and glyph which represent a shape of each letter. The dynamic caption accounts for the movements of letters and backgrounds. We add one more guidance with zero convolution to determine which text content should be visible in the video. We apply the zero convolution to the text content, and impose it on the diffusion model. Lastly, our glyph loss, only minimizing a difference between the predicted word and its ground-truth, is proposed to make the prediction letters readable. Experiments show that our model generates kinetic typography videos with legible and artistic letter motions based on text prompts.", + "bbox": [ + 261, + 291, + 738, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Kinetic Typography $\\cdot$ Video Diffusion $\\cdot$ Visual Text Generation", + "bbox": [ + 261, + 623, + 738, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 681, + 374, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kinetic typography is an artistic motion graphics design combining text and animations [13]. Based on the word's meaning, motions are generated to convey information in videos. The main goal is eye-catching and to improve message retention. With the rise of video media, it has become an essential element in TV programs, commercials, music videos and film leaders.", + "bbox": [ + 212, + 710, + 785, + 786 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kinetic typography controls letters' shape (glyph), color and texture over time, and transforms their positions. Professional motion graphic designers", + "bbox": [ + 215, + 787, + 785, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 218, + 824, + 380, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1b7e4304a1c291bb2fa7195c891d6b6d3b7e5cbc489435d4b6e83f0f094e8022.jpg", + "image_caption": [ + "(a) Designer's workflow" + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 557, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8e0ab68d1e3e2732eb91503e1f6aaa3c140574b36324ae03bd1c67ae76ac8705.jpg", + "image_caption": [ + "(b) Our Kinety model", + "Fig. 1: An overview of our KineTy pipeline, which is motivated by the designer's workflow. Our key idea is to generate eye-catching and aesthetic animatable words based on user instructions." + ], + "image_footnote": [], + "bbox": [ + 575, + 143, + 787, + 218 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "use commercial software like 'Adobe After Effects' to make and render kinetic typography videos. The conventional pipeline is as follows [14, 50]; (1) Define a text box in a workspace called composition and enter a text (i.e. word or message). (2) Edit the static text by tuning its font, color and texture (3) Build the background if needed (4) Apply motion effects for all text and background (5) Repeat, optimize, and fine-tune the whole process until it satisfies user intention. This process is time-consuming and labor-intensive. It takes three hours for simple motions to several days for sophisticated effects per single kinetic typography video [14, 26]. To efficiently do this, designers may need to consider either various design references or pre-defined design options called templates.", + "bbox": [ + 212, + 310, + 784, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the advancement of generative models, there have been attempts to produce dynamic typography. The literature on typography generation has mostly focused on creating static single-letter images. Starting from transforming the shape of the object into a single letter [23, 39, 52, 56], there have been works to create multiple letters and multiple words by introducing a concept of layout [10, 11, 74]. Furthermore, DynTypo [37] and Shape-Matching GAN++ [72] adopt style-transfer techniques to produce animatable effects on a single letter. They focus on only animatable effects without any actual motions of letters.", + "bbox": [ + 212, + 462, + 785, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, the recent advancement of the video generation model has been actively studied with the success of text-guided video diffusion models, especially user's description conditions video frames [7,16,18,20,49]. Although these models have opened up the possibility of creating kinetic typography, there is a critical issue that the video generations show a weak understanding of the letters' shapes and motions.", + "bbox": [ + 212, + 583, + 785, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose the Kinetic Typography diffusion model (called KineTy model) that generates kinetic typography from user-provided text prompts. Inspired by the video diffusion model, we allow users to input comprehensive descriptions for color, font, size, position and motion effects of letters. To better represent text motions, we first introduce our kinetic typography video dataset. 600K videos are rendered by combining randomly generated text contents with 584 templates made by professional motion graphic designers. These videos are labeled with static and dynamic captions that describe the video with respect to its appearance and motion characteristics, respectively.", + "bbox": [ + 212, + 672, + 787, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Next, we present KineTy model that effectively synthesizes videos from the text prompt. Here, we reevaluate how to effectively condition the caption guidance", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Park et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "into the video diffusion model at a fundamental level. We enforce that each effect is incorporated by separately inserting static and dynamic captions into the spatial attention and temporal attention, respectively. To strengthen these attentions, we apply a zero convolution [76] into the word caption, and add it to them. We use a mask loss term only for the generated video by masking out the background contents, which makes the text contents more readable.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Experimental results show that our Kinecty robustly generates kinetic typography video with multi-letter legibility, while accurately representing captions. Furthermore, extensive and meticulous user studies support our claim that Kinecty produces more aesthetic outcomes than general-purpose video generations.", + "bbox": [ + 212, + 237, + 789, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 321, + 387, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Typography Generation", + "text_level": 1, + "bbox": [ + 215, + 354, + 460, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The key of typography is to make readable and visuallyappealing text contents to readers. Most research in this field has focused on transferring the style of a single-letter design, including elements like color, glyph, font and effects, to other letters [1,2,5,8,9,15,30-34,40,42,45,53,57,60,62,64,70,75]. Additionally, there have been works on transferring an image style to the text contents [73]. Similarly, research on scene text editing, which changes text content while maintaining its own style in a scene, is also underway [27,43,54,69], whose extended version to video is available in [51]. In addition to changing text contents, color editing [47] and text segmentation [59,68] are proposed.", + "bbox": [ + 212, + 381, + 787, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the recent success of the high-fidelity text-to-image diffusion model [46], typography generation has gained interest. Works in [23,52,56] apply image styles into the target letter in an unsupervised manner.", + "bbox": [ + 212, + 517, + 787, + 563 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are concurrent works that display multi-letters, beyond the single-letter generation. They follow the multi-step approach that firstly generates layouts at specific positions, arrange them with multi-letters or multi-words with the same fonts and colors, and contextually infers the background [10, 11, 24, 25, 48, 61, 67, 74, 77]. However, these models are not specialized for typography. Since all letters are generated together, it is not editable for the letters, which makes it difficult to animate and add movement to each character.", + "bbox": [ + 212, + 563, + 787, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Typography Video Generation", + "text_level": 1, + "bbox": [ + 215, + 691, + 513, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As we mentioned above, this work is the first attempt to generate kinetic typography. Although no existing works directly align with this, there are some works related to typography videos [26, 28, 29, 38, 63, 65]. Here, we would like to introduce two methods related to ours.", + "bbox": [ + 212, + 719, + 787, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DynTypo [37] proposes a dynamic typography model that transfers the dynamic effect with realistic movements like fire and water on a specific uppercase English letter to others. Shape-Matching GAN++ [72] transfers an image style into a target letter by matching these shapes with structures of the target letter.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Kinetic Typography Diffusion Model", + "bbox": [ + 483, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/d1b4c9ba3cf95e64c2d1f8c450c7d5a54b64609e7f273c61475e237e46011116.jpg", + "table_caption": [ + "Table 1: A summary of the kinetic typography datasets." + ], + "table_footnote": [], + "table_body": "
DatasetDomain#SamplesVideoStatic CaptionDynamic CaptionMulti-letterText AppearanceBackground Text Movements
Multi-Content [2]Typography10K----✓-
TextLogo3K [61]Typography3K----✓-
TE141K [71]Typography141K---✓✓✓
TextSeg [68]Text Segmentation4K---✓✓✓
LAION-Glyph-10M [74]Text-to-Image10M-✓-✓✓✓
MARIO-10M [11]Text-to-Image10M-✓-✓✓✓
AnyWord-3M [54]Text-to-Image3M-✓-✓✓✓
CATER-GEN [21]Image-to-Video35K✓-✓N/AN/A△
WebVid-10M [3]Text-to-Video10M✓✓△N/AN/A✓
Vimeo25M [58]Text-to-Video25M✓✓△N/AN/A✓
OursKinetic Typography600K✓✓✓✓✓✓
", + "bbox": [ + 217, + 160, + 782, + 291 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the position and glyph of the letter are still fixed, with no actual movement between frames. Even under the condition of a single uppercase letter, representing various static and dynamic effects through text prompts remains a challenging issue.", + "bbox": [ + 212, + 315, + 785, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By substituting the motion description with external user input, the task becomes more tractable. A work, Wakey-Wakey [66], transfers the source GIF (Graphics Interchange Format) motion to the target text. Here, graphic designers manually assign additional corresponding key points to achieve motion transfer. However, this process still requires direct human intervention.", + "bbox": [ + 212, + 376, + 787, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Text-to-Video Diffusion Models", + "text_level": 1, + "bbox": [ + 215, + 473, + 522, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The video diffusion model produces a visually-plausible and photo-realistic video based on text conditioning [7, 12, 16, 18, 20, 36, 49, 58]. The main challenge in the video diffusion model is to maintain the temporal coherency between frames. The pioneer works for video diffusion is to add temporal blocks to the text-to-image model and to learn temporal coherency between video frames [16, 20, 49]. Another type of relevant studies use pretrained models to leverage the temporal block [6, 7, 18]. In particular, Guo et al. [18] propose a motion module as the temporal block. It is designed as a plug-and-play module. After joint training with the pre-trained weight of stable diffusion [46], the module enables any text-to-image model to generate the video.", + "bbox": [ + 212, + 498, + 787, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing typography has been mainly studied using static single letters. The similar work to kinetic typography is also single letters, with only variation of motion effect and almost no letter's movement exists. To initiate kinetic typography generation, we present our KineTy dataset and model that allows dynamic motion, effects and glyph deformation of multi-letters, which will be explained at the next section.", + "bbox": [ + 212, + 648, + 787, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Kinetic Typography Dataset", + "text_level": 1, + "bbox": [ + 215, + 762, + 529, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We describe how to build the KineTy dataset. Unlike previous datasets that only cover single-lettered images [71], our KineTy dataset has not only visual effects on multiple letters, but also their animations. We first provide detailed process", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Park et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/17bfe6dbb3cc3a454ad1d956c7f83b1ee090a34b8f06abfd6aa7426fe10f2c0d.jpg", + "image_caption": [], + "image_footnote": [ + "pink to white background." + ], + "bbox": [ + 217, + 143, + 500, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a8c6f10e5518deddc1b8517ab215d2f867a21cbedfb2b03e9924e7a26ca338df.jpg", + "image_caption": [], + "image_footnote": [ + "giving it a vibrant and dimensional appearance on a white background." + ], + "bbox": [ + 501, + 143, + 785, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/299721f2b7e5c6502d2dd9d0195d331e114e3c8dc30a871d08d8a31d9f93ba22.jpg", + "image_caption": [], + "image_footnote": [ + "The word \"LgLed\" forms with a glossy, liquid effect, where each letter appears in sequence from left to right to gradually complete the word.", + "The word \"ADI\" prominently displayed in bold, 3D-styled letters set against an orange textured background." + ], + "bbox": [ + 217, + 220, + 500, + 263 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2f0ec059ee10fafcea3c2de8e886848c1f78f46e4b2b3da8d162dcc82e135faf.jpg", + "image_caption": [], + "image_footnote": [ + "The video illustrates the word \"HHtR\" materializing in layers, with each letter representing a layer and enhancing the previous ones to gradually construct the complete word.", + "The word \"FLRNQVOIZ\" with a stylized pink colored glitch effect on a dark background." + ], + "bbox": [ + 501, + 220, + 785, + 263 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/20d6eadc6c3330e13c5e80df381d0202d712502f31ff79bb93691c2a5cccbe69.jpg", + "image_caption": [ + "Fig. 2: Examples of our Kinecty dataset. Our dataset provides high-quality kinetic typography video created by professional motion graphic designers, along with captions that describe the visual appearance and motion effects. To aid visualization, we provide three frames from each video clip." + ], + "image_footnote": [ + "The word \"ADI\" gets bigger.", + "The word \"INSZG\" in bold letters casting a long shadow on a grey background.", + "The word \"INSZG\" is revealed in sequential order with each character appearing in order until the full word is displayed." + ], + "bbox": [ + 217, + 295, + 500, + 339 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/252a0fd273ddb66c019680d90258fa14e633cb8bbecb2d066b18acb077fc91bf.jpg", + "image_caption": [], + "image_footnote": [ + "The word \"FRLNQVOZ\" starts from a scattered formation and converging to form the word clearly.", + "A colorful, distorted graphic with the word \"CmfyTSr\" in stylized lettering on a dark background.", + "The video depicts a fluid-like stream that gradually forms the word \"CmfyTSR.\" The flow is slow and smooth in a smooth flowing sequence giving the impression of a liquid in motion." + ], + "bbox": [ + 501, + 295, + 785, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to render kinetic typography videos using templates made by professional editors in Sec. 3.1. Next, we introduce an way to describe their appearance and motions through text prompts in Sec. 3.2. Lastly, we explain how to make ground-truth kinetic typography video for strictly fair comparisons in Sec. 3.3. We summarize the difference between ours and existing datasets in Tab. 1.", + "bbox": [ + 214, + 463, + 785, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Video Rendering", + "text_level": 1, + "bbox": [ + 215, + 566, + 401, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Employing templates for video rendering. The template is a pre-designed project file which contains a visual effect on letters. Many editors prefer to use the templates because it saves their time and labor cost in practice. Following the best practice, we utilize 584 kinetic typography templates from professional graphic designers for our dataset construction, as visualized in Fig. 2.", + "bbox": [ + 212, + 595, + 785, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Set of multiple letters. Next, we utilize the kinetic typography templates and randomly replace the text contents for augmentation. We employ multiple letters, in contrast to existing typography datasets with static single-letters [2,31,70]. Text contents are randomly generated by arranging up to 12 letters sampled from a set of 52 letters, including both uppercase and lowercase alphabets. Through this, we can expect rich letter-by-letter effects with various arrangements, while also keeping consistent styles across the multiple letters.", + "bbox": [ + 212, + 672, + 787, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Speculation. We render videos with $1,920 \\times 1,080$ resolution for 3 sec., and 1,000 random words from 584 templates. Subsequently, all videos are downsampled to $512 \\times 288$ resolution with 8fps for training. It takes a month to render the whole dataset with four i9 13900KF CPUs and an NVIDIA 2080ti GPU.", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Kinetic Typography Diffusion Model", + "bbox": [ + 485, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/36479d975f307eba64d998dee15e397e39683cd323c01c92b4271c547caa9ef8.jpg", + "image_caption": [ + "Fig. 3: Statistics of our proposed dataset." + ], + "image_footnote": [], + "bbox": [ + 230, + 143, + 787, + 265 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Video Captioning.", + "text_level": 1, + "bbox": [ + 215, + 303, + 411, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Static and dynamic effect separation. In the next step, we caption the rendered kinetic typography for text-to-video generation. The professional editors usually design the static appearances at first, and then apply dynamic effects. Following this practice, we use two type of captions: static and dynamic captions, and labeled them accordingly.", + "bbox": [ + 212, + 325, + 787, + 400 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Static effect captioning. To label static captions, we focus on letter's appearance which contains spatial information of typography. As shown in Fig. 2, the static captions $\\mathcal{C}_{\\text {static }}$ describe the color of the letters and the background, the glyph of the letters (e.g., outlined with yellow color or bold font), the characteristics of the background (e.g., textured and shiny background) and the arrangement of the letters (e.g., in a diagonal way) based on the last frame of the video where all the letters are displayed.", + "bbox": [ + 212, + 402, + 787, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dynamic effect captioning. Similarly, we concentrate on the temporal change of motions for each frame to write dynamic captions. Dynamic captions $\\mathcal{C}_{dynamic}$ describe the motion part of the video, such as whether each letter appears in sequential or random order, can be rotated, or has a fade-in effect. For better systematic process, we initially label the videos with the GPT-4Vision model [41], and then manually verify and refine them to fix missing and wrongly labeled components.", + "bbox": [ + 212, + 508, + 787, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Statistics. The statistics of our dataset is summarized in Fig. 3. There are three main categories for the static appearance in Fig. 3 (a): Text type, text effect and background. Here, the text fill effects are dominant in the text type because it is the most obvious way to clearly show the text content. Especially, professional designers tend to prefer using solid colors, rather than text outlines. The reason why $57.7\\%$ of our dataset has no text effect is the readability of text contents. Since kinetic typography makes dynamic motions of the contents, users do not need for fancy visual effects on it.", + "bbox": [ + 212, + 613, + 787, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, we categorize dynamic motions into two-fold in Fig. 3 (b): motion type, including entrance, emphasis and movement, and sequence type: Our dataset supports various effect of entrance and emphasis to deliver striking messages. In contrast, the line has the majority in the movement because users' intention is usually conveyed after all letters in a scene are arranged in a line. In the same vein, when words come inside all at one and appear one-by-one in sequence, it is readable. They thus have the high portion in the sequence category.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Park et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/332008dee99c2e7634aad4e753ccc52e7fb96b6e40cada6f1eb788ca418087de.jpg", + "image_caption": [ + "Fig. 4: An architecture of our KineTy model." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 787, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Ground-truth Video Generation", + "text_level": 1, + "bbox": [ + 214, + 276, + 529, + 291 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unlike generating random synthetic words in a training phase, we use real-world letters for evaluation by leveraging the templates. We can render ground-truth videos corresponding to user-input captions.", + "bbox": [ + 212, + 299, + 782, + 344 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For evaluation, individual letters are selected for each alphabet letter from A to Z, and the first letters are capitalized for checking case sensitivity (e.g., Apple, Ball, ..., Zebra). These 26 words are used for rendering 584 templates, so that a total of 15,184 videos are used for the evaluation.", + "bbox": [ + 212, + 345, + 787, + 404 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Kinetic Typography Diffusion Model", + "text_level": 1, + "bbox": [ + 214, + 425, + 609, + 443 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we present our model for creating kinetic typography videos based on textual prompts. The key concept is to show animate visual effects on texts, and our challenges in this work are summarized as: (1) the effect should be visually-pleasing and eye-catching to viewers; (2) the transition between frames should be smooth and align with the captions; (3) the text must be readable.", + "bbox": [ + 212, + 454, + 784, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We begin by defining a kinetic typography generation problem in Sec. 4.1. We then describe how to model static and dynamic captions that efficiently guide appearance and motions in Sec. 4.2. We lastly discuss how to improve the glyph legibility with respect to the model design and its learning strategy in Sec. 4.3. Implementation details are provided in Sec. 4.4.", + "bbox": [ + 212, + 529, + 787, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Preliminary", + "text_level": 1, + "bbox": [ + 214, + 625, + 362, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Conditional Latent Diffusion Models. Diffusion models train the data distribution by progressively refining a noisy initial state $z_{M} \\sim \\mathcal{N}(0,1)$ into the target data representation $z_{0}$ for $M$ diffusion steps. Recent advancements, particularly in Latent Diffusion Models (LDMs) [46], enhance the efficiency by encoding an image $x$ into a compact latent representation $z_{0} = \\mathcal{E}(x)$ using an encoder $\\mathcal{E}$ , and is transformed back to the image $\\tilde{x} = \\mathcal{D}(z_{0})$ using a decoder $\\mathcal{D}$ . These adding noise and its subsequent removal are done with U-Net-style denoising network $\\epsilon_{\\pi}$ . In addition, a condition $y$ is mapped to the hidden state $h$ of $\\epsilon_{\\pi}$ via attention:", + "bbox": [ + 212, + 646, + 787, + 782 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t n} _ {\\theta} (h, y) = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {\\theta} (h) K _ {\\theta} (y) ^ {\\top}}{\\sqrt {d}}\\right) \\cdot V _ {\\theta} (y) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 787, + 787, + 824 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {s . t .} Q _ {\\theta} (h) = W _ {\\theta , Q} \\cdot h, K _ {\\theta} (y) = W _ {\\theta , K} \\cdot y, V _ {\\theta} (y) = W _ {\\theta , V} \\cdot y,\n$$\n", + "text_format": "latex", + "bbox": [ + 285, + 825, + 715, + 843 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Kinetic Typography Diffusion Model", + "bbox": [ + 483, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $W_{\\theta, Q}$ , $W_{\\theta, K}$ and $W_{\\theta, V}$ denote the learnable parameters for queries, keys and values. $d$ is the number of dimensions of the keys. The objective function for the denoising network $\\epsilon_{\\pi}$ is formulated as:", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {l d m} = \\mathbb {E} _ {\\mathcal {E} (x), y, \\epsilon \\sim \\mathcal {N} (0, 1), m} \\left[ \\left\\| \\epsilon - \\epsilon_ {\\pi} \\left(z _ {m}, m, y\\right) \\right\\| _ {2} ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 204, + 787, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\|_2$ is the $L_2$ distance and diffusion step $m \\in \\{1, \\dots, M\\}$ is uniformly sampled during training.", + "bbox": [ + 212, + 241, + 785, + 271 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Extending image diffusion models to video. Extending the capability of text-to-image diffusion models to video generation has been more feasible. By incorporating a temporal self-attention called motion module [18], it learns a temporal coherency between $T$ frames in a latent sequence $\\mathbf{z}_0^{0:T}$ , which corresponds to the video sequence $\\tilde{\\mathbf{x}}^{0:T} = \\mathcal{D}(\\mathbf{z}_0^{0:T})$ . This allows diffusion models to generate smoothly changed sequential images over time, whose temporal self-attention can be defined as follows:", + "bbox": [ + 212, + 273, + 787, + 378 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t n} _ {\\phi} \\left(\\mathbf {h} ^ {0: T}, \\mathbf {h} ^ {0: T}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {\\phi} \\left(\\mathbf {h} ^ {0 : T}\\right) K _ {\\phi} \\left(\\mathbf {h} ^ {0 : T}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\cdot V _ {\\phi} \\left(\\mathbf {h} ^ {0: T}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 276, + 388, + 787, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Spatial and Temporal Guidance", + "text_level": 1, + "bbox": [ + 215, + 448, + 524, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Static caption incorporation. Motivated by the best practice of professional designers who handle appearance and motion effects separately, we divide the caption into static and dynamic elements.", + "bbox": [ + 212, + 477, + 785, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When existing text-to-image models learn the distribution of image data, they are conditioned on texts related to the image's appearance. In the same manner, our model is also guided by captions describing the appearance of the text contents and background in each frame of video. Using Eq. (1), we define a spatial attention block with a self-attention followed by a cross-attention with a static caption $\\mathcal{C}_{\\text{static}}$ as:", + "bbox": [ + 212, + 523, + 787, + 613 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} ^ {\\prime 0: T} = \\left\\{\\operatorname {A t t n} _ {\\vartheta} \\left(\\operatorname {A t t n} _ {\\theta} \\left(\\mathbf {h} ^ {t}, \\mathbf {h} ^ {t}\\right), \\tau \\left(\\mathcal {C} _ {\\text {s t a t i c}}\\right)\\right) \\right\\} _ {t = 0} ^ {T - 1}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 626, + 787, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\tau (\\cdot)$ is the CLIP text encoder [44].", + "bbox": [ + 212, + 659, + 513, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dynamic caption incorporation. A previous work [18] uses a motion module that only utilizes a self-attention between frames to learn the temporal consistency. On the other hand, in kinetic typography, it is essential to accurately display the dynamic motion effects of each letter in a video, following user's textual description. To do this, we extend Eq. (3) by adding the cross-attention with a dynamic caption as follows:", + "bbox": [ + 212, + 675, + 787, + 766 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} ^ {\\prime 0: T} = \\operatorname {A t t n} _ {\\psi} \\big (\\operatorname {A t t n} _ {\\phi} (\\mathbf {h} ^ {0: T}, \\mathbf {h} ^ {0: T}), \\tau \\big (\\mathcal {C} _ {\\text {d y n a m i c}} \\big) \\big), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 777, + 787, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Through this process, the model becomes capable of maintaining temporal consistency as well as direct control over dynamic movements.", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Park et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Enhancing Glyph Legibility.", + "text_level": 1, + "bbox": [ + 215, + 146, + 493, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Readability improvement. As a next step, we aim to make the legibility of our model better. To do this, we first classify descriptions for text contents and prompt because diffusion models often have a difficulty in distinguishing them. In this work, we put down a delimiter between each character for text contents.", + "bbox": [ + 212, + 171, + 787, + 231 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Let a text content $\\mathcal{L} = \\{l_1, l_2, \\dots, l_L\\}$ contain $L$ letters. To condition each letter separately, we join each letter $l_i$ in the text content with a delimiter symbol $|\\cdot\\rangle$ . Further, we add a symbol '\\~' behind upper case letters because the clip encoder does not care of upper cases. To the end, when the word $\\mathcal{L}$ denotes \"Apple\", $\\mathcal{L}'$ is represented as \"A^{|p|p|l|e}\". We denote the whole $\\mathcal{L}$ as $\\mathcal{L}'$ for better representation.", + "bbox": [ + 212, + 232, + 787, + 306 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Attention on text contents. It can be challenging for the attention module to guide both text contents and effects at once, while disentangling the text content information in CLIP feature space. What is worse, the text content can be overwhelmed by relatively long appearance and motion effects in CLIP text encoder. To tackle this problem, we introduce an additional cross-attention branch to the text content. Inspired by ControlNet [76], we regard the text content as conditions through the zero convolution operation. Starting with the definition of word caption $\\mathcal{C}_{word}$ via a prompt template like \"The word $\\{\\mathcal{L}'\\}$ \", we extend Eq. (4) by adding the cross-attention module between the word caption and hidden feature, weighted by the zero-initialized convolutions $\\rho$ as follows:", + "bbox": [ + 212, + 306, + 787, + 459 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {h} _ {\\text {s t a t i c}} ^ {\\prime \\prime 0: T} = \\left\\{\\operatorname {A t t n} _ {\\vartheta} \\left(\\mathbf {h} ^ {\\prime}, \\tau \\left(\\mathcal {C} _ {\\text {s t a t i c}}\\right)\\right) + \\rho_ {v} \\left(\\operatorname {A t t n} _ {v} \\left(\\mathbf {h} ^ {\\prime}, \\tau \\left(\\mathcal {C} _ {\\text {w o r d}}\\right)\\right)\\right) \\right\\} _ {t = 0} ^ {T - 1}, \\tag {6} \\\\ w h e r e \\quad \\mathbf {h} ^ {\\prime} = \\operatorname {A t t n} _ {\\theta} \\left(\\mathbf {h} ^ {t}, \\mathbf {h} ^ {t}\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 274, + 469, + 787, + 517 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This allows the network to gradually evaluate the usefulness of this additional caption based on the condition. In the same way, we incorporate the word caption into the temporal cross-attention in Eq. (5) as:", + "bbox": [ + 212, + 521, + 787, + 568 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {h} _ {\\text {d y n a m i c}} ^ {\\prime \\prime 0: T} = \\operatorname {A t t n} _ {\\psi} \\left(\\mathbf {h} ^ {\\prime}, \\tau \\left(\\mathcal {C} _ {\\text {d y n a m i c}}\\right)\\right) + \\rho_ {\\varphi} \\left(\\operatorname {A t t n} _ {\\varphi} \\left(\\mathbf {h} ^ {\\prime}, \\tau \\left(\\mathcal {C} _ {\\text {w o r d}}\\right)\\right)\\right), \\tag {7} \\\\ w h e r e \\quad \\mathbf {h} ^ {\\prime} = \\operatorname {A t t n} _ {\\phi} \\left(\\mathbf {h} ^ {0: T}, \\mathbf {h} ^ {0: T}\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 281, + 579, + 787, + 623 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Glyph loss. To train our model, we use a common loss function $L_{ldm}$ . Additionally, we impose an extra penalty on the letter regions to enforce a sharp and correct glyph of text contents. We first use a binary mask $B$ for the letters in the last frame $V^T$ from a text segmentation model [68]. The mask is then blurred to cover its surrounding effects. We then define a glyph loss based on $L_{LDM}$ with the additional pixel-wise weighting strategy using the blurred mask as follows:", + "bbox": [ + 212, + 627, + 787, + 719 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {g l y p h} = \\mathbb {E} _ {\\mathcal {E} (x), y, \\epsilon \\sim \\mathcal {N} (0, 1), m} \\left[ \\left\\| B \\odot \\left(\\epsilon - \\epsilon_ {\\pi} \\left(z _ {m} ^ {T}, m, y\\right)\\right) \\right\\| _ {2} ^ {2} \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 729, + 787, + 755 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\odot$ is the element-wise multiplication.", + "bbox": [ + 212, + 763, + 529, + 777 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "With the glyph loss, we enhance the legibility of text contents, enabling the precise creation of multi-letter formations. Through the linear combination of both loss terms, we can formulate the final loss function as $L = L_{ldm} + \\alpha L_{glyph}$ . Here, we empirically set $\\alpha$ to 0.01.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Kinetic Typography Diffusion Model", + "bbox": [ + 483, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 146, + 455, + 161 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To train our model, we use a two-step training strategy similar to that of human trainees [14]. First, we pre-train the network to generate spatial appearances using only static caption and last video frame pairs. Here, we detach the temporal attention modules to make them work as text-to-image diffusion models during 30 epochs with a batch size of 200. Next, we train the full model while freezing the spatial attention using whole video frames and captions for an epoch with a batch size of 8. For training, we resize the height $H$ and width $W$ of the video into $256 \\times 256$ with $T = 24$ frames. The training is performed with AdamW optimizer [35], with a diffusion step $M$ of 1000 and a learning rate of 0.0001, which usually takes about 20 hours for training on 8 NVIDIA A100 GPUs. The inference time is about 20 seconds when the number of sampling steps is 25.", + "bbox": [ + 212, + 170, + 787, + 335 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 358, + 375, + 375 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we conduct comprehensive experiments to verify the effectiveness of our model for kinetic typography. We first describe the experimental setup in Sec. 5.1. We then provide comparison results with relevant typography and video generation models, and report user-study to highlight the practical applicability of our model in Sec. 5.2. We lastly carry out an extensive ablation study to validate the effect of each component in our model in Sec. 5.3", + "bbox": [ + 212, + 387, + 787, + 479 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 500, + 426, + 513 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparison methods. We compare ours with state-of-the-art generative models, which consist of 3 two-stage methods and 2 one-stage methods. The two-stage methods are based on combinations of text-to-image models, including DS-Fusion [52], GlyphControl [74] and Text-Diffuser [11], and an image-to-video model, SparseCtrl [17]. DS-Fusion outputs a stylized letter based on the given style phrase and a letter. Since our caption is a sentence, we extract a style keyword from it and consider it as a style phrase. Note that DS-Fusion is able to generate one letter at a time, so we concatenate each letter to make a full-text content. GlyphControl takes an instruction to generate a glyph image for text contents, then uses the image as a condition for the final output image. Text-Diffuser first finds a proper layout for words and produce the output based on a caption and the layout. Since the comparison methods yield images, we combine them with an image-to-video model, SparseCtrl, to compare with ours. SparseCtrl uses a caption and a guidance such as sketches, depth maps and images to generate a stylized video.", + "bbox": [ + 212, + 522, + 787, + 748 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "On the other hand, one-stage-methods,AnimateDiff [18] and Lavie [58], are based on text-to-video models.AnimateDiff extends pre-trained text-to-image diffusion models with a motion module, and trains only the motion module to make fully use of the massive information of video dataset. Lavie leverages a temporal self-attention block to enhance a temporal consistency between frames while keeping the generation performance.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Park et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/c63be7d77e6c5de12bd0ab9380bf79fff5fdf6025100b463f4534e33a86ade0a.jpg", + "table_caption": [ + "Table 2: Quantitative evaluation of two-stage generation methods, and one-stage text-to-video generation methods. Bold: Best, Underline: Second-best." + ], + "table_footnote": [], + "table_body": "
ModelEntranceEmphasisMotionAverage
FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑
DS-F[52]+Sparse[17]1636.35.710.690.583184.95.970.730.581236.55.420.760.561843.105.660.700.58
Glyph[74]+Sparse[17]2415.64.300.760.682667.15.060.760.751468.64.480.840.802361.984.460.760.69
T-Diff[11]+Sparse[17]2937.25.610.780.903107.55.160.790.891208.24.060.820.952815.65.350.780.90
AnimateDiff[18]1613.84.710.590.012106.04.560.660.032160.64.510.670.071727.554.680.610.01
Lavie [58]825.12.250.820.541835.43.830.830.41505.443.330.850.35956.972.680.820.49
Ours147.41.790.870.71177.521.750.880.5436.161.620.940.98125.901.770.880.76
", + "bbox": [ + 218, + 174, + 782, + 268 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/71782326842b53f99dee087045070277f47a92419be858169d0b1c441cd3bdb2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 282, + 331, + 334 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6641917464a0de8178c6ea8044efebe2e111f39e81a4eaef998a2f5518fb1b54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 282, + 444, + 334 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/05fce8642548fec43bc3020bec1932a97a35155db0895a0f08126a8813a9375e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 282, + 558, + 334 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/430afc07bae35518711bf141055c3fcf1dab309f032232f8fe2a76cfeeb55957.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 282, + 671, + 334 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/52669d860eb45a89f7fb1ca1a130deaf098cd2d654a4e6e09de00d4184b55b1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 282, + 784, + 334 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Static caption: The word 'Dog' in bold, green-stripped letters against a textured coral background. Dynamic caption: The sequential addition of letters one by one to spell out the word 'Dog'.", + "bbox": [ + 217, + 334, + 575, + 345 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/725d9f0177562488b62051c1829de0480da90cf8738b5a6eab434fba78ed85d1.jpg", + "image_caption": [ + "Fig. 5: Qualitative results from the comparison models and ours. Obviously, the results from ours reflect the contents of captions better than the others." + ], + "image_footnote": [], + "bbox": [ + 217, + 347, + 331, + 395 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/18c86c52a02390aa315b9aeff28d706f1e9c14eee3ad5986533c4856cabeda9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 347, + 444, + 395 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f89da7733b5becbc819a3b1246035ed2db8de3a689ef745483c5cc71bd0ee9cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 347, + 558, + 395 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0feb4f8301ead937bbda8bae32405b40814e4149d36d3d653dff3f1825692a69.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 347, + 671, + 395 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c71bf862a28e003657c5b309f564ac140b75dc6dc4534a6c9d97ab038373c49e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 347, + 784, + 395 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Static caption: The word Mermaid in a blurred, tunnel-like visual effect, creating a sense of depth and distortion on a dark background. Dynamic caption: The word Mermaid gradually comes into focus, with each letter appearing in a random order, intensifying the sense of depth and creating a captivating rev", + "bbox": [ + 217, + 396, + 774, + 409 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/db57bd5ce63cf6aac5ac85486c3296071025225634e4224bb07fd10af688e203.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 411, + 331, + 460 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/dc63aa8892f1dc22fe0f3adc810c83059f39d91ff9b920b0ca92cb85dd29d5e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 411, + 444, + 460 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5b8f5443d35da0d2affedee1bfb4b78f430e4ea1e40c6979a63b11ba075fe648.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 411, + 558, + 460 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2f89c99544f55493f992466f0075e9762afa9f6ff646f601aa6c689eb0f3e9b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 411, + 671, + 460 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8b83ac5a8b4f61f82533292a21623fca2fc3fd4adddc2dd716ba3dd2718460b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 411, + 784, + 460 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Static caption: The word 'Watermelon' in white letters against a black background, accompanied by two smiling emojis faces with heart eyes. Dynamic caption: Each character of the word 'Watermelon' appears sequentially, with various smiling emojis popping up randomly around the text.", + "bbox": [ + 217, + 460, + 750, + 470 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(a) DS-Fusion", + "bbox": [ + 243, + 474, + 303, + 484 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "+ SparseCtrl", + "bbox": [ + 246, + 484, + 302, + 491 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(b) GlyphControl", + "bbox": [ + 349, + 474, + 423, + 484 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "+SparseCtrl", + "bbox": [ + 359, + 484, + 413, + 492 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(c) Text-Diffuser", + "bbox": [ + 462, + 474, + 535, + 484 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "+ SparseCtrl", + "bbox": [ + 473, + 484, + 527, + 492 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(d) Lavie", + "bbox": [ + 594, + 474, + 635, + 483 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(e) Ours", + "bbox": [ + 710, + 474, + 746, + 483 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For a fair comparison, we utilize the official source codes and pre-trained weights provided by the authors. For more analysis, we categorize the motion effects into three groups, 'Entrance', 'Emphasis' and 'Motion', and conduct experiments to check the quality of each result.", + "bbox": [ + 212, + 545, + 784, + 604 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation metrics. Since the final products of kinetic typography depend on the designer's preferences, we adopt four feature-based metrics rather than photoconsistency measures like PSNR. (1) Fréchet Video Distance (FVD) qualifies a similarity between two videos by comparing their feature representations extracted from a pre-trained deep neural network [55]. (2) Inception Score (IS) evaluates the quality and diversity of images generated by a model [4]. (3) CLIPScore measures the semantic alignment between generated images and their textual descriptions, using the CLIP model [19, 22, 44]. (4) Optical Character Recognition (OCR) checks the clarity of the output by comparing recognized text to a reference text using the F1-score.", + "bbox": [ + 212, + 606, + 785, + 757 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Evaluation Results", + "text_level": 1, + "bbox": [ + 214, + 782, + 416, + 795 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative results. As shown in Tab. 2, our Kinety model achieves the better performance than the comparison methods in the most metrics on all the", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Kinetic Typography Diffusion Model", + "bbox": [ + 485, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ab40cc94a030dfde61128a6056884fbfeda639dd999542f473ce3bba82c57d86.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 545, + 289 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6f66d48c83b9388c2cda852f8b1135f976d91323de4cf8d1d1c6b4e07734c7dd.jpg", + "image_caption": [ + "Fig. 6: Visualization of attention maps between letters from a caption and feature map of a hidden layer." + ], + "image_footnote": [], + "bbox": [ + 218, + 289, + 267, + 299 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/263d4876eca264f3cbaea87f31cfa9e74ae2c6377af224a23698643ecb7adefd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 289, + 310, + 299 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/497865e5ac049905fc5ac22ec87f42c9aaaebb5bf1a18f75d6d3c96755934e68.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 289, + 361, + 299 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8f775c883874303c4be34e3807bd5ea41820861fd1e14a10226ea1db42c59ec8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 289, + 393, + 299 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b2880aa9c7ff5380ededae83965f46f4a08ce62526fda4eaf87269b6f408007d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 289, + 433, + 299 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8f2abb81af309ba4df511cd93a97736f0009516dd2c0e7f3c7e300de2133def8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 454, + 289, + 480, + 299 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b8e0bb6acd9f9357508616b74839a754397ff3de6dd32700c6574a55fea150b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 289, + 521, + 299 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/057b2b05a8f49d404d77940ebbb804c48ec1a767d0b87df4f88dd7728e67512d.jpg", + "image_caption": [ + "The word" + ], + "image_footnote": [], + "bbox": [ + 547, + 143, + 785, + 289 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/326f57c1f7443e98bae158cb2562753b2a1fc71e21c364f1ede53cad23911da9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 289, + 630, + 299 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6758910690e47a18537e1f9c04724d3fa496f988a1624c0651a1b40518a5119d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 289, + 660, + 299 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f23cfb53c5f2760e5045299d91958b0037d79dcf314a890586d68f66c0e17808.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 289, + 689, + 299 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ad572e26d114e4a06a400706820b16b29fa619ccfa96641c7d224ed341c92722.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 289, + 718, + 299 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/46d6dbec202424b4b31cdbefa4e66af86378bb54c115cae1c572ac71fc87112a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 289, + 748, + 299 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/bb8d3ea1b9ffe45b07f06f3e881003ba83a74be905dec44a8d3bd77027f7dc04.jpg", + "image_caption": [ + "(a) Initial kinetic typography video", + "(b) Changing text color" + ], + "image_footnote": [], + "bbox": [ + 217, + 345, + 785, + 383 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d317fddc2596f93d21cdf4f1de59c186de4ac3c8f697a5eaae18e2797c135a45.jpg", + "image_caption": [ + "(c) Changing text font", + "(d) Changing background" + ], + "image_footnote": [], + "bbox": [ + 217, + 398, + 785, + 435 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/342fba2ed5aa57c5035b6b2205b0eb8f0ef20438f84fe7c3a4580c3a4eeae615.jpg", + "image_caption": [ + "(e) Changing motion effect", + "(f) Final product", + "Fig. 7: The demonstration of the editable capability of our Kinecty model after generating initial kinetic typography." + ], + "image_footnote": [], + "bbox": [ + 217, + 450, + 785, + 487 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "categories. The comparison methods, despite their impressive performance in general image and video generation, show unsatisfactory performance in creating kinetic typography. The reasons are mainly two-folds: (1) Text-to-video models are not specialized for kinetic typography and often fail to exactly distinguish the text content in the long captions. Different from them, we additionally feed $\\mathcal{C}_{word}$ which imposes a high attention to the text content in the video. (2) They have challenges in creating letter-level motions. Since they are unaware of each letter in the text content, all the letters move as a whole or inconsistently. We handle this issue by providing a delimiter between each letter and using zero-convolution in Fig. 6.", + "bbox": [ + 212, + 551, + 787, + 700 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Our model shows promising results in generating legible text content, but it sometimes falls short in OCR accuracy. This is because the two-stage methods algorithmically generate a glyph image of a text content, and then deform glyphs conditioned on the glyph image. As a result, they have the better performance in OCR metrics than ours and one-stage methods. Nevertheless, our method outperforms in the most cases, demonstrating the effectiveness of our method in generating the clear text content and motion effects.", + "bbox": [ + 212, + 703, + 787, + 808 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative results. In Fig. 5, we display three examples of the comparison methods. Compared to ours, the two-stage methods do not understand the user's", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Park et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/83d5fafd730e51519212aba942becefc4c64c5b8f48fc510576e35d818066b56.jpg", + "table_caption": [ + "Table 3: The result of user-study. The score ranges from $1 \\sim 6$ for Study 1 & 2, and 0 $\\sim 1$ for Study 3. Bold: Best, Underline: Second-best." + ], + "table_footnote": [], + "table_body": "
User StudyDynamic EffectGeneralExpert
DS-F +SparseGlyph +SparseT-Diff +SparseAnimate DiffLavieOursDS-F +SparseGlyph +SparseT-Diff +SparseAnimate DiffLavieOurs
Study 1Entrance3.3843.6873.8912.4063.1534.4783.4814.0254.1311.0193.0815.263
Emphasis3.5623.7134.1192.5632.7254.3193.6254.1754.2251.0502.6255.300
Motion2.3634.1384.1133.3382.3754.6751.7003.5504.0004.0002.0005.750
Average3.3183.7133.9232.5183.0364.4943.3353.9954.0901.3252.9055.350
Study 2Entrance3.5143.7753.9412.4133.2704.0883.6504.0814.2441.0193.1064.900
Emphasis3.5313.8814.0752.2693.0504.1943.7254.1504.0501.0252.6755.375
Motion2.8383.9503.8753.2502.8504.2382.1504.0004.4503.4501.9505.000
Average3.4393.7993.9482.4603.2204.1353.4854.0604.2551.2652.9704.995
Study 3Entrance0.6280.6200.6520.0630.1690.7250.8440.8560.8870.0190.1380.969
Emphasis0.5810.6060.6380.0880.1060.7810.7500.7000.9000.0000.0001.000
Motion0.0750.7130.7500.3620.2880.7810.0000.9000.9000.5000.2501.000
Average0.5530.6260.6580.0960.1740.7440.7300.8350.8800.0650.1350.975
", + "bbox": [ + 220, + 174, + 782, + 337 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "instruction with respect to either the whole caption or the letter's movement. Here, we need to know how to perform our Kinecty model. Figure 6 visualizes the attention map between the caption and features from the hidden state on the corresponding images. Thanks to the cross-attention between the word caption and the noisy latent, and the delimiter to separate the words, our model successfully embeds features for each letter in the noise.", + "bbox": [ + 212, + 371, + 784, + 460 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Lastly, to show the generality of our KineTy model, we conduct an additional experiment. After generating an initial kinetic typography, our KineTy model allows users to modify its style and motions such as the content color, background, font and motion, and even to combine them. As demonstrated in Fig. 7, ours can render the modified outcomes corresponding to the additional captions.", + "bbox": [ + 212, + 465, + 785, + 542 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "User Study. To assess the practical utility of our results in the field of kinetic typography, we conduct a user study using Amazon MTurk. We ask 20 questionnaires to 50 participants. Since their familiarity of typography can vary significantly, we divide them into two groups: experts and non-experts, consisting of 10 and 40 individuals, respectively. The study involves three sections: (1) caption alignment; (2) kinetic typography suitability; (3) Word readability. We make two video clips by randomly choosing two words for each effect. In total, we make 20 clips for this user study.", + "bbox": [ + 212, + 547, + 787, + 667 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In the first study, we measure how well ours and the comparison methods generate videos that align with provided captions. The participants see 6 videos from different models with the same corresponding caption in a random order. They are asked to rank the videos from 1 to 6, and the rankings are subsequently converted into scores, ranging from 6 (Best alignment) to 1 (Worst alignment). The following study evaluates how proper the generated videos are for motion graphic applications. Since some participants might be unfamiliar with kinetic typography, all participants watch 4 example videos from online in advance. After that, they check six videos in a random order and rank them. Scores were assigned based on these rankings, from 6 (the most proper) to 1 (the least proper). The last study assesses how readable the outcomes are. After watching 6 videos, the", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Kinetic Typography Diffusion Model", + "bbox": [ + 485, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/71eef853e8a15d196a592c730a7bf4333456d22cbea3dc8cee47c983ae0cd90e.jpg", + "table_caption": [ + "Table 4: The result of ablation study. Bold: Best, Underline: Second-best." + ], + "table_footnote": [], + "table_body": "
ModelEntranceEmphasisMotionAverage
FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑
Ours147.41.790.870.71177.521.750.880.5436.161.620.940.98125.901.770.880.76
-Effect separation789.82.750.740.49880.12.880.750.49607.12.540.820.23729.82.730.750.48
-Cword239.11.770.860.70247.11.730.870.5462.71.600.950.95201.81.740.870.74
-Lglph188.91.770.860.70226.81.750.880.5550.11.590.930.98160.81.750.880.75
", + "bbox": [ + 220, + 160, + 782, + 234 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "participants are needed to vote for videos with readable text contents. Of course, multiple voting is available.", + "bbox": [ + 212, + 258, + 785, + 287 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The performance of each method is measured based on scores for the first two studies, and on the selection ratio for third one. As demonstrated in Tab. 3, our approach shows the promising performance across all studies. The results highlight our model's practical usability in creating kinetic typography, especially for the domain experts who give our model the highest scores.", + "bbox": [ + 212, + 289, + 785, + 364 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 388, + 388, + 405 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Without caption separation. First of all, we train our network without dividing static and dynamic caption. The performances drop significantly for every metrics in Tab. 4. Since we train the text-to-image backbone using only static captions, dynamic captions are considered as noisy input. Similarly, the motion module, used to train dynamic motions, is effective when the simple motion guidance is given, compared to using static and dynamic caption together.", + "bbox": [ + 212, + 417, + 785, + 507 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Text contents incorporation. We observe that text-to-image diffusion models often fail to catch up on text contents in the prompt when a long description is given. We thus use $\\mathcal{C}_{word}$ as an additional input to give a strong attention to each letter. As shown in Tab. 4, without $\\mathcal{C}_{word}$ , the performance degradation is observed as expected.", + "bbox": [ + 212, + 508, + 785, + 582 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Without glyph loss. Lastly, we evaluate the performance without $L_{glyph}$ . The inferior FVD scores come from the blur effect at the edge of the word along with color bleeding effects.", + "bbox": [ + 212, + 584, + 785, + 628 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 655, + 359, + 671 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose a generative kinetic typography model, named KineTy. To achieve this, we first build a large-scale kinetic typography dataset by collecting 584 templates from professional designers, and use them to train our diffusion-based network. To effectively handle the text prompt, we split it into static and dynamic captions. They are used to directly guide the spatial and temporal cross-attention for better appearance and motion effects, respectively. Our glyph loss also strengthens the legible and artistic letter generation. We lastly demonstrate that our KineTy produces the visually compelling and semantically coherent kinetic typography videos. The extensive user studies further validate the effectiveness of our model with respect to its practical utility.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Park et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements This research was supported by 'Project for Science and Technology Opens the Future of the Region' program through the INNOPOLIS FOUNDATION funded by Ministry of Science and ICT (Project Number: 2022-DD-UP-0312), the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(RS-2024-00338439), GIST-MIT Research Collaboration grant funded by the GIST in 2024, and the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-01842, Artificial Intelligence Graduate School Program (GIST)).", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 304, + 321, + 319 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Anderson, D., Shamir, A., Fried, O.: Neural font rendering. arXiv preprint arXiv:2211.14802 (2022)", + "2. Azadi, S., Fisher, M., Kim, V.G., Wang, Z., Shechtman, E., Darrell, T.: Multicontent gan for few-shot font style transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018)", + "3. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)", + "4. Barratt, S., Sharma, R.: A note on the inception score. arXiv preprint arXiv:1801.01973 (2018)", + "5. Berio, D., Leymarie, F.F., Asente, P., Echevarria, J.: Strokestyles: Stroke-based segmentation and stylization of fonts. ACM Transactions on Graphics (TOG) (2022)", + "6. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)", + "7. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "8. Campbell, N.D., Kautz, J.: Learning a manifold of fonts. ACM Transactions on Graphics (TOG) (2014)", + "9. Chen, C.H., Liu, Y.T., Zhang, Z., Guo, Y.C., Zhang, S.H.: Joint implicit neural representation for high-fidelity and compact vector fonts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)", + "10. Chen, J., Huang, Y., Lv, T., Cui, L., Chen, Q., Wei, F.: Textdiffuser-2: Unleashing the power of language models for text rendering. arXiv preprint arXiv:2311.16465 (2023)", + "11. Chen, J., Huang, Y., Lv, T., Cui, L., Chen, Q., Wei, F.: Textdiffuser: Diffusion models as text painters. Proceedings of the Neural Information Processing Systems (NeurIPS) (2023)", + "12. Esser, P., Chiu, J., Atighehchian, P., Granskog, J., Germanidis, A.: Structure and content-guided video synthesis with diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)", + "13. Ford, S., Forlizzi, J., Ishizaki, S.: Kinetic typography: issues in time-based presentation of text. In: CHI'97 extended abstracts on Human factors in computing systems, pp. 269-270. ACM Digital Library (1997)" + ], + "bbox": [ + 218, + 332, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Kinetic Typography Diffusion Model", + "bbox": [ + 485, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Fridsma, L., Gyncild, B.: Adobe After Effects CC Classroom in a Book. Adobe Press (2019)", + "15. Fu, B., He, J., Wang, J., Qiao, Y.: Neural transformation fields for arbitrary-styled font generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "16. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)", + "17. Guo, Y., Yang, C., Rao, A., Agrawala, M., Lin, D., Dai, B.: Sparsectrl: Adding sparse controls to text-to-video diffusion models. arXiv preprint arXiv:2311.16933 (2023)", + "18. Guo, Y., Yang, C., Rao, A., Wang, Y., Qiao, Y., Lin, D., Dai, B.: AnimateDiff: Animate your personalized text-to-image diffusion models without specific tuning arXiv preprint arXiv:2307.04725 (2023)", + "19. Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (2021)", + "20. Ho, J., Chan, W., Saharia, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)", + "21. Hu, Y., Luo, C., Chen, Z.: Make it move: controllable image-to-video generation with text descriptions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "22. Huang, Y., Xue, H., Liu, B., Lu, Y.: Unifying multimodal transformer for bidirectional image and text generation. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 1138-1147 (2021)", + "23. Iluz, S., Vinker, Y., Hertz, A., Berio, D., Cohen-Or, D., Shamir, A.: Word-as-image for semantic typography. ACM Transactions on Graphics (TOG) (2023)", + "24. Jahanian, A., Liu, J., Lin, Q., Tretter, D., O'Brien-Strain, E., Lee, S.C., Lyons, N., Allebach, J.: Recommendation system for automatic design of magazine covers In: Proceedings of the 2013 International Conference on Intelligent User Interfaces (IUI) (2013)", + "25. Jia, P., Li, C., Liu, Z., Shen, Y., Chen, X., Yuan, Y., Zheng, Y., Chen, D., Li, J., Xie, X., et al.: Cole: A hierarchical generation framework for graphic design. arXiv preprint arXiv:2311.16974 (2023)", + "26. Kato, J., Nakano, T., Goto, M.: Textalive: Integrated design environment for kinetic typography. In: Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems (CHI) (2015)", + "27. Krishnan, P., Kovvuri, R., Pang, G., Vassilev, B., Hassner, T.: Textstylebrush: Transfer of text aesthetics from a single example. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2023)", + "28. Lee, J.C., Forlizzi, J., Hudson, S.E.: The kinetic typography engine: an extensible system for animating expressive text. In: Proceedings of the 15th annual ACM symposium on User Interface Software and Technology (UIST) (2002)", + "29. Lee, J., Jun, S., Forlizzi, J., Hudson, S.E.: Using kinetic typography to convey emotion in text-based interpersonal communication. In: Proceedings of the 6th conference on Designing Interactive systems (DIS) (2006)", + "30. Li, C., Taniguchi, Y., Lu, M., Konomi, S.: Few-shot font style transfer between different languages. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Park et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Li, X., Wu, L., Wang, C., Meng, L., Meng, X.: Compositional zero-shot artistic font synthesis. In: Proceedings of the International Joint Conferences on Artificial Intelligence (IJCAI) (2023)", + "32. Liu, X., Meng, G., Chang, J., Hu, R., Xiang, S., Pan, C.: Decoupled representation learning for character glyph synthesis. IEEE Transactions on Multimedia (TMM) (2021)", + "33. Liu, Y.T., Guo, Y.C., Li, Y.X., Wang, C., Zhang, S.H.: Learning implicit glyph shape representation. IEEE Transactions on Visualization and Computer Graphics (TVCG) (2022)", + "34. Liu, Y.T., Zhang, Z., Guo, Y.C., Fisher, M., Wang, Z., Zhang, S.H.: Dualvector: Unsupervised vector font synthesis with dual-part representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "35. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: Proceedings of the International Conference on Learning Representations (ICLR) (2018)", + "36. Luo, Z., Chen, D., Zhang, Y., Huang, Y., Wang, L., Shen, Y., Zhao, D., Zhou, J., Tan, T.: Videofusion: Decomposed diffusion models for high-quality video generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "37. Men, Y., Lian, Z., Tang, Y., Xiao, J.: Dyntypo: Example-based dynamic text effects transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)", + "38. Minakuchi, M., Tanaka, K.: Automatic kinetic typography composer. In: Proceedings of the ACM SIGCHI International Conference on Advances in Computer Entertainment Technology (ACE) (2005)", + "39. Mu, X., Chen, L., Chen, B., Gu, S., Bao, J., Chen, D., Li, J., Yuan, Y.: Fontstudio: Shape-adaptive diffusion model for coherent and consistent font effect generation. arXiv preprint arXiv:2406.08392 (2024)", + "40. Nagata, Y., Iwana, B.K., Uchida, S.: Contour completion by transformers and its application to vector font data. arXiv preprint arXiv:2304.13988 (2023)", + "41. OpenAI: GPT-4V(ision) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf (2023)", + "42. Pan, W., Zhu, A., Zhou, X., Iwana, B.K., Li, S.: Few shot font generation via transferring similarity guided global style and quantization local style. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)", + "43. Qu, Y., Tan, Q., Xie, H., Xu, J., Wang, Y., Zhang, Y.: Exploring stroke-level modifications for scene text editing. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) (2023)", + "44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proceedings of the International Conference on Machine Learning (ICML) (2021)", + "45. Reddy, P., Zhang, Z., Wang, Z., Fisher, M., Jin, H., Mitra, N.: A multi-implicit neural representation for fonts. Proceedings of the Neural Information Processing Systems (NeurIPS) (2021)", + "46. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "47. Shimoda, W., Haraguchi, D., Uchida, S., Yamaguchi, K.: De-rendering stylized texts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Kinetic Typography Diffusion Model", + "bbox": [ + 485, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Shimoda, W., Haraguchi, D., Uchida, S., Yamaguchi, K.: Towards diverse and consistent typography generation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2024)", + "49. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. In: Proceedings of the International Conference on Learning Representations (ICLR) (2023)", + "50. Smith, J., Team, A.C.: Adobe After Effects CS6 Digital Classroom. John Wiley & Sons (2012)", + "51. Subramanian, J., Chordia, V., Bart, E., Fang, S., Guan, K., Bala, R., et al.: Strive: Scene text replacement in videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)", + "52. Tanveer, M., Wang, Y., Mahdavi-Amiri, A., Zhang, H.: Ds-fusion: Artistic typography via discriminated and stylized diffusion. arXiv preprint arXiv:2303.09604 (2023)", + "53. Thamizharasan, V., Liu, D., Agarwal, S., Fisher, M., Gharbi, M., Wang, O., Jacobson, A., Kalogerakis, E.: Vecfusion: Vector font generation with diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)", + "54. Tuo, Y., Xiang, W., He, J.Y., Geng, Y., Xie, X.: Anytext: Multilingual visual text generation and editing. In: Proceedings of the International Conference on Learning Representations (ICLR) (2024)", + "55. Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717 (2018)", + "56. Wang, C., Wu, L., Liu, X., Li, X., Meng, L., Meng, X.: Anything to glyph: Artistic font synthesis via text-to-image diffusion model. In: SIGGRAPH Asia 2023 Conference Papers (2023)", + "57. Wang, C., Zhou, M., Ge, T., Jiang, Y., Bao, H., Xu, W.: Cf-font: Content fusion for few-shot font generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "58. Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., Yang, P., et al.: Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103 (2023)", + "59. Wang, Y., Ye, Y., Mao, Y., Yu, Y., Song, Y.: Self-supervised scene text segmentation with object-centric layered representations augmented by text regions. In: Proceedings of the 30th ACM International Conference on Multimedia (2022)", + "60. Wang, Y., Lian, Z.: Deepvecfont: Synthesizing high-quality vector fonts via dual-modality learning. ACM Transactions on Graphics (TOG) (2021)", + "61. Wang, Y., Pu, G., Luo, W., Wang, Y., Xiong, P., Kang, H., Lian, Z.: Aesthetic text logo synthesis via content-aware layout inferring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "62. Wang, Y., Wang, Y., Yu, L., Zhu, Y., Lian, Z.: Deepvecfont-v2: Exploiting transformers to synthesize vector fonts with higher quality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "63. Wong, Y.Y.: Temporal typography: a proposal to enrich written expression. In: Proceedings of the Conference Companion on Human Factors in Computing Systems (CHI) (1996)", + "64. Xia, Z., Xiong, B., Lian, Z.: Vecfontsf: Learning to reconstruct and synthesize high-quality vector fonts via signed distance functions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "S. Park et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "65. Xie, L., Shu, X., Su, J.C., Wang, Y., Chen, S., Qu, H.: Creating emordle: Animating word cloud for emotion expression. IEEE Transactions on Visualization and Computer Graphics (TVCG) (2023)", + "66. Xie, L., Zhou, Z., Yu, K., Wang, Y., Qu, H., Chen, S.: Wakey-wakey: animate text by mimicking characters in a gif. In: Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (2023)", + "67. Xu, C., Zhou, M., Ge, T., Jiang, Y., Xu, W.: Unsupervised domain adaption with pixel-level discriminator for image-aware layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "68. Xu, X., Zhang, Z., Wang, Z., Price, B., Wang, Z., Shi, H.: Rethinking text segmentation: A novel dataset and a text-specific refinement approach. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "69. Yang, Q., Huang, J., Lin, W.: Swaptext: Image based texts transfer in scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "70. Yang, S., Liu, J., Lian, Z., Guo, Z.: Awesome typography: Statistics-based text effects transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2017)", + "71. Yang, S., Wang, W., Liu, J.: Te141k: artistic text benchmark for text effect transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2020)", + "72. Yang, S., Wang, Z., Liu, J.: Shape-matching $\\mathrm{gan} + +$ : Scale controllable dynamic artistic text style transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2021)", + "73. Yang, S., Wang, Z., Wang, Z., Xu, N., Liu, J., Guo, Z.: Controllable artistic text style transfer via shape-matching gan. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2019)", + "74. Yang, Y., Gui, D., Yuan, Y., Liang, W., Ding, H., Hu, H., Chen, K.: Glyphcontrol: Glyph conditional control for visual text generation. In: Proceedings of the Neural Information Processing Systems (NeurIPS) (2023)", + "75. Yang, Z., Peng, D., Kong, Y., Zhang, Y., Yao, C., Jin, L.: Fontdiffuser: One-shot font generation via denoising diffusion with multi-scale content aggregation and style contrastive learning. arXiv preprint arXiv:2312.12142 (2023)", + "76. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)", + "77. Zhang, S., Ma, J., Wu, J., Ritchie, D., Agrawala, M.: Editing motion graphics video via motion vectorization and transformation. ACM Transactions on Graphics (TOG) (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 700 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Kinetic Typography Diffusion Model", + "bbox": [ + 485, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Kinetic Typography Diffusion Model/d7a51e44-9061-4ac8-9c76-37fd7685fe8a_model.json b/2024/Kinetic Typography Diffusion Model/d7a51e44-9061-4ac8-9c76-37fd7685fe8a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f4bcbde40423566aae5e460e63eaf157631fc35e --- /dev/null +++ b/2024/Kinetic Typography Diffusion Model/d7a51e44-9061-4ac8-9c76-37fd7685fe8a_model.json @@ -0,0 +1,3527 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.292, + 0.142, + 0.711, + 0.163 + ], + "angle": 0, + "content": "Kinetic Typography Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.19, + 0.756, + 0.206 + ], + "angle": 0, + "content": "Seonmi Park, Inhwan Bae, Seunghyun Shin, and Hae-Gon Jeon" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.218, + 0.778, + 0.26 + ], + "angle": 0, + "content": "AI Graduate School, GIST, South Korea \n{bluesky1000, inhwanbae, seunghyuns98}@gm.gist.ac.kr, haegonj@gist.ac.kr \nhttps://seonmip.github.io/kinety" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.292, + 0.74, + 0.61 + ], + "angle": 0, + "content": "Abstract. This paper introduces a method for realistic kinetic typography that generates user-preferred animatable \"text content\". We draw on recent advances in guided video diffusion models to achieve visually-pleasing text appearances. To do this, we first construct a kinetic typography dataset, comprising about 600K videos. Our dataset is made from a variety of combinations in 584 templates designed by professional motion graphics designers and involves changing each letter's position, glyph, and size (i.e., flying, glitches, chromatic aberration, reflecting effects, etc.). Next, we propose a video diffusion model for kinetic typography. For this, there are three requirements: aesthetic appearances, motion effects, and readable letters. This paper identifies the requirements. For this, we present static and dynamic captions used as spatial and temporal guidance of a video diffusion model, respectively. The static caption describes the overall appearance of the video, such as colors, texture and glyph which represent a shape of each letter. The dynamic caption accounts for the movements of letters and backgrounds. We add one more guidance with zero convolution to determine which text content should be visible in the video. We apply the zero convolution to the text content, and impose it on the diffusion model. Lastly, our glyph loss, only minimizing a difference between the predicted word and its ground-truth, is proposed to make the prediction letters readable. Experiments show that our model generates kinetic typography videos with legible and artistic letter motions based on text prompts." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.624, + 0.739, + 0.637 + ], + "angle": 0, + "content": "Keywords: Kinetic Typography \\(\\cdot\\) Video Diffusion \\(\\cdot\\) Visual Text Generation" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.682, + 0.375, + 0.697 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.786, + 0.787 + ], + "angle": 0, + "content": "Kinetic typography is an artistic motion graphics design combining text and animations [13]. Based on the word's meaning, motions are generated to convey information in videos. The main goal is eye-catching and to improve message retention. With the rise of video media, it has become an essential element in TV programs, commercials, music videos and film leaders." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.788, + 0.787, + 0.818 + ], + "angle": 0, + "content": "Kinetic typography controls letters' shape (glyph), color and texture over time, and transforms their positions. Professional motion graphic designers" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.825, + 0.382, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "S. Park et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.145, + 0.558, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.219, + 0.446, + 0.231 + ], + "angle": 0, + "content": "(a) Designer's workflow" + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.144, + 0.788, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.22, + 0.736, + 0.232 + ], + "angle": 0, + "content": "(b) Our Kinety model" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.238, + 0.785, + 0.279 + ], + "angle": 0, + "content": "Fig. 1: An overview of our KineTy pipeline, which is motivated by the designer's workflow. Our key idea is to generate eye-catching and aesthetic animatable words based on user instructions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.785, + 0.462 + ], + "angle": 0, + "content": "use commercial software like 'Adobe After Effects' to make and render kinetic typography videos. The conventional pipeline is as follows [14, 50]; (1) Define a text box in a workspace called composition and enter a text (i.e. word or message). (2) Edit the static text by tuning its font, color and texture (3) Build the background if needed (4) Apply motion effects for all text and background (5) Repeat, optimize, and fine-tune the whole process until it satisfies user intention. This process is time-consuming and labor-intensive. It takes three hours for simple motions to several days for sophisticated effects per single kinetic typography video [14, 26]. To efficiently do this, designers may need to consider either various design references or pre-defined design options called templates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.787, + 0.582 + ], + "angle": 0, + "content": "With the advancement of generative models, there have been attempts to produce dynamic typography. The literature on typography generation has mostly focused on creating static single-letter images. Starting from transforming the shape of the object into a single letter [23, 39, 52, 56], there have been works to create multiple letters and multiple words by introducing a concept of layout [10, 11, 74]. Furthermore, DynTypo [37] and Shape-Matching GAN++ [72] adopt style-transfer techniques to produce animatable effects on a single letter. They focus on only animatable effects without any actual motions of letters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.787, + 0.672 + ], + "angle": 0, + "content": "Meanwhile, the recent advancement of the video generation model has been actively studied with the success of text-guided video diffusion models, especially user's description conditions video frames [7,16,18,20,49]. Although these models have opened up the possibility of creating kinetic typography, there is a critical issue that the video generations show a weak understanding of the letters' shapes and motions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.809 + ], + "angle": 0, + "content": "In this paper, we propose the Kinetic Typography diffusion model (called KineTy model) that generates kinetic typography from user-provided text prompts. Inspired by the video diffusion model, we allow users to input comprehensive descriptions for color, font, size, position and motion effects of letters. To better represent text motions, we first introduce our kinetic typography video dataset. 600K videos are rendered by combining randomly generated text contents with 584 templates made by professional motion graphic designers. These videos are labeled with static and dynamic captions that describe the video with respect to its appearance and motion characteristics, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Next, we present KineTy model that effectively synthesizes videos from the text prompt. Here, we reevaluate how to effectively condition the caption guidance" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.485, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Kinetic Typography Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "into the video diffusion model at a fundamental level. We enforce that each effect is incorporated by separately inserting static and dynamic captions into the spatial attention and temporal attention, respectively. To strengthen these attentions, we apply a zero convolution [76] into the word caption, and add it to them. We use a mask loss term only for the generated video by masking out the background contents, which makes the text contents more readable." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.79, + 0.3 + ], + "angle": 0, + "content": "Experimental results show that our Kinecty robustly generates kinetic typography video with multi-letter legibility, while accurately representing captions. Furthermore, extensive and meticulous user studies support our claim that Kinecty produces more aesthetic outcomes than general-purpose video generations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.323, + 0.388, + 0.339 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.356, + 0.461, + 0.372 + ], + "angle": 0, + "content": "2.1 Typography Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.789, + 0.518 + ], + "angle": 0, + "content": "The key of typography is to make readable and visuallyappealing text contents to readers. Most research in this field has focused on transferring the style of a single-letter design, including elements like color, glyph, font and effects, to other letters [1,2,5,8,9,15,30-34,40,42,45,53,57,60,62,64,70,75]. Additionally, there have been works on transferring an image style to the text contents [73]. Similarly, research on scene text editing, which changes text content while maintaining its own style in a scene, is also underway [27,43,54,69], whose extended version to video is available in [51]. In addition to changing text contents, color editing [47] and text segmentation [59,68] are proposed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.789, + 0.564 + ], + "angle": 0, + "content": "With the recent success of the high-fidelity text-to-image diffusion model [46], typography generation has gained interest. Works in [23,52,56] apply image styles into the target letter in an unsupervised manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.789, + 0.669 + ], + "angle": 0, + "content": "There are concurrent works that display multi-letters, beyond the single-letter generation. They follow the multi-step approach that firstly generates layouts at specific positions, arrange them with multi-letters or multi-words with the same fonts and colors, and contextually infers the background [10, 11, 24, 25, 48, 61, 67, 74, 77]. However, these models are not specialized for typography. Since all letters are generated together, it is not editable for the letters, which makes it difficult to animate and add movement to each character." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.693, + 0.514, + 0.709 + ], + "angle": 0, + "content": "2.2 Typography Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.779 + ], + "angle": 0, + "content": "As we mentioned above, this work is the first attempt to generate kinetic typography. Although no existing works directly align with this, there are some works related to typography videos [26, 28, 29, 38, 63, 65]. Here, we would like to introduce two methods related to ours." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.841 + ], + "angle": 0, + "content": "DynTypo [37] proposes a dynamic typography model that transfers the dynamic effect with realistic movements like fire and water on a specific uppercase English letter to others. Shape-Matching GAN++ [72] transfers an image style into a target letter by matching these shapes with structures of the target letter." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "S. Park et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.31, + 0.145, + 0.692, + 0.159 + ], + "angle": 0, + "content": "Table 1: A summary of the kinetic typography datasets." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.161, + 0.784, + 0.292 + ], + "angle": 0, + "content": "
DatasetDomain#SamplesVideoStatic CaptionDynamic CaptionMulti-letterText AppearanceBackground Text Movements
Multi-Content [2]Typography10K----✓-
TextLogo3K [61]Typography3K----✓-
TE141K [71]Typography141K---✓✓✓
TextSeg [68]Text Segmentation4K---✓✓✓
LAION-Glyph-10M [74]Text-to-Image10M-✓-✓✓✓
MARIO-10M [11]Text-to-Image10M-✓-✓✓✓
AnyWord-3M [54]Text-to-Image3M-✓-✓✓✓
CATER-GEN [21]Image-to-Video35K✓-✓N/AN/A△
WebVid-10M [3]Text-to-Video10M✓✓△N/AN/A✓
Vimeo25M [58]Text-to-Video25M✓✓△N/AN/A✓
OursKinetic Typography600K✓✓✓✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.316, + 0.787, + 0.376 + ], + "angle": 0, + "content": "However, the position and glyph of the letter are still fixed, with no actual movement between frames. Even under the condition of a single uppercase letter, representing various static and dynamic effects through text prompts remains a challenging issue." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.788, + 0.453 + ], + "angle": 0, + "content": "By substituting the motion description with external user input, the task becomes more tractable. A work, Wakey-Wakey [66], transfers the source GIF (Graphics Interchange Format) motion to the target text. Here, graphic designers manually assign additional corresponding key points to achieve motion transfer. However, this process still requires direct human intervention." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.474, + 0.523, + 0.489 + ], + "angle": 0, + "content": "2.3 Text-to-Video Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.499, + 0.788, + 0.65 + ], + "angle": 0, + "content": "The video diffusion model produces a visually-plausible and photo-realistic video based on text conditioning [7, 12, 16, 18, 20, 36, 49, 58]. The main challenge in the video diffusion model is to maintain the temporal coherency between frames. The pioneer works for video diffusion is to add temporal blocks to the text-to-image model and to learn temporal coherency between video frames [16, 20, 49]. Another type of relevant studies use pretrained models to leverage the temporal block [6, 7, 18]. In particular, Guo et al. [18] propose a motion module as the temporal block. It is designed as a plug-and-play module. After joint training with the pre-trained weight of stable diffusion [46], the module enables any text-to-image model to generate the video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.741 + ], + "angle": 0, + "content": "Existing typography has been mainly studied using static single letters. The similar work to kinetic typography is also single letters, with only variation of motion effect and almost no letter's movement exists. To initiate kinetic typography generation, we present our KineTy dataset and model that allows dynamic motion, effects and glyph deformation of multi-letters, which will be explained at the next section." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.763, + 0.53, + 0.781 + ], + "angle": 0, + "content": "3 Kinetic Typography Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We describe how to build the KineTy dataset. Unlike previous datasets that only cover single-lettered images [71], our KineTy dataset has not only visual effects on multiple letters, but also their animations. We first provide detailed process" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.486, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Kinetic Typography Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.501, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.218, + 0.188, + 0.501, + 0.204 + ], + "angle": 0, + "content": "pink to white background." + }, + { + "type": "image_footnote", + "bbox": [ + 0.218, + 0.204, + 0.476, + 0.219 + ], + "angle": 0, + "content": "The word \"LgLed\" forms with a glossy, liquid effect, where each letter appears in sequence from left to right to gradually complete the word." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.188, + 0.501, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.787, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.503, + 0.188, + 0.787, + 0.203 + ], + "angle": 0, + "content": "giving it a vibrant and dimensional appearance on a white background." + }, + { + "type": "image_footnote", + "bbox": [ + 0.503, + 0.203, + 0.787, + 0.219 + ], + "angle": 0, + "content": "The video illustrates the word \"HHtR\" materializing in layers, with each letter representing a layer and enhancing the previous ones to gradually construct the complete word." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.188, + 0.787, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.221, + 0.501, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.218, + 0.265, + 0.499, + 0.28 + ], + "angle": 0, + "content": "The word \"ADI\" prominently displayed in bold, 3D-styled letters set against an orange textured background." + }, + { + "type": "image_footnote", + "bbox": [ + 0.218, + 0.28, + 0.328, + 0.293 + ], + "angle": 0, + "content": "The word \"ADI\" gets bigger." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.265, + 0.499, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.221, + 0.787, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.503, + 0.265, + 0.771, + 0.279 + ], + "angle": 0, + "content": "The word \"FLRNQVOIZ\" with a stylized pink colored glitch effect on a dark background." + }, + { + "type": "image_footnote", + "bbox": [ + 0.503, + 0.279, + 0.779, + 0.293 + ], + "angle": 0, + "content": "The word \"FRLNQVOZ\" starts from a scattered formation and converging to form the word clearly." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.265, + 0.779, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.296, + 0.501, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.218, + 0.34, + 0.464, + 0.354 + ], + "angle": 0, + "content": "The word \"INSZG\" in bold letters casting a long shadow on a grey background." + }, + { + "type": "image_footnote", + "bbox": [ + 0.218, + 0.354, + 0.478, + 0.368 + ], + "angle": 0, + "content": "The word \"INSZG\" is revealed in sequential order with each character appearing in order until the full word is displayed." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.34, + 0.478, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.296, + 0.787, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.503, + 0.34, + 0.786, + 0.354 + ], + "angle": 0, + "content": "A colorful, distorted graphic with the word \"CmfyTSr\" in stylized lettering on a dark background." + }, + { + "type": "image_footnote", + "bbox": [ + 0.503, + 0.354, + 0.786, + 0.368 + ], + "angle": 0, + "content": "The video depicts a fluid-like stream that gradually forms the word \"CmfyTSR.\" The flow is slow and smooth in a smooth flowing sequence giving the impression of a liquid in motion." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.34, + 0.786, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.377, + 0.787, + 0.432 + ], + "angle": 0, + "content": "Fig. 2: Examples of our Kinecty dataset. Our dataset provides high-quality kinetic typography video created by professional motion graphic designers, along with captions that describe the visual appearance and motion effects. To aid visualization, we provide three frames from each video clip." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.464, + 0.787, + 0.541 + ], + "angle": 0, + "content": "to render kinetic typography videos using templates made by professional editors in Sec. 3.1. Next, we introduce an way to describe their appearance and motions through text prompts in Sec. 3.2. Lastly, we explain how to make ground-truth kinetic typography video for strictly fair comparisons in Sec. 3.3. We summarize the difference between ours and existing datasets in Tab. 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.567, + 0.403, + 0.582 + ], + "angle": 0, + "content": "3.1 Video Rendering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.787, + 0.673 + ], + "angle": 0, + "content": "Employing templates for video rendering. The template is a pre-designed project file which contains a visual effect on letters. Many editors prefer to use the templates because it saves their time and labor cost in practice. Following the best practice, we utilize 584 kinetic typography templates from professional graphic designers for our dataset construction, as visualized in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Set of multiple letters. Next, we utilize the kinetic typography templates and randomly replace the text contents for augmentation. We employ multiple letters, in contrast to existing typography datasets with static single-letters [2,31,70]. Text contents are randomly generated by arranging up to 12 letters sampled from a set of 52 letters, including both uppercase and lowercase alphabets. Through this, we can expect rich letter-by-letter effects with various arrangements, while also keeping consistent styles across the multiple letters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Speculation. We render videos with \\(1,920 \\times 1,080\\) resolution for 3 sec., and 1,000 random words from 584 templates. Subsequently, all videos are downsampled to \\(512 \\times 288\\) resolution with 8fps for training. It takes a month to render the whole dataset with four i9 13900KF CPUs and an NVIDIA 2080ti GPU." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "S. Park et al." + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.144, + 0.788, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.269, + 0.642, + 0.283 + ], + "angle": 0, + "content": "Fig. 3: Statistics of our proposed dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.304, + 0.413, + 0.32 + ], + "angle": 0, + "content": "3.2 Video Captioning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.788, + 0.401 + ], + "angle": 0, + "content": "Static and dynamic effect separation. In the next step, we caption the rendered kinetic typography for text-to-video generation. The professional editors usually design the static appearances at first, and then apply dynamic effects. Following this practice, we use two type of captions: static and dynamic captions, and labeled them accordingly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.788, + 0.508 + ], + "angle": 0, + "content": "Static effect captioning. To label static captions, we focus on letter's appearance which contains spatial information of typography. As shown in Fig. 2, the static captions \\(\\mathcal{C}_{\\text {static }}\\) describe the color of the letters and the background, the glyph of the letters (e.g., outlined with yellow color or bold font), the characteristics of the background (e.g., textured and shiny background) and the arrangement of the letters (e.g., in a diagonal way) based on the last frame of the video where all the letters are displayed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Dynamic effect captioning. Similarly, we concentrate on the temporal change of motions for each frame to write dynamic captions. Dynamic captions \\(\\mathcal{C}_{dynamic}\\) describe the motion part of the video, such as whether each letter appears in sequential or random order, can be rotated, or has a fade-in effect. For better systematic process, we initially label the videos with the GPT-4Vision model [41], and then manually verify and refine them to fix missing and wrongly labeled components." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Statistics. The statistics of our dataset is summarized in Fig. 3. There are three main categories for the static appearance in Fig. 3 (a): Text type, text effect and background. Here, the text fill effects are dominant in the text type because it is the most obvious way to clearly show the text content. Especially, professional designers tend to prefer using solid colors, rather than text outlines. The reason why \\(57.7\\%\\) of our dataset has no text effect is the readability of text contents. Since kinetic typography makes dynamic motions of the contents, users do not need for fancy visual effects on it." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In addition, we categorize dynamic motions into two-fold in Fig. 3 (b): motion type, including entrance, emphasis and movement, and sequence type: Our dataset supports various effect of entrance and emphasis to deliver striking messages. In contrast, the line has the majority in the movement because users' intention is usually conveyed after all letters in a scene are arranged in a line. In the same vein, when words come inside all at one and appear one-by-one in sequence, it is readable. They thus have the high portion in the sequence category." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.485, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Kinetic Typography Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.788, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.235, + 0.655, + 0.25 + ], + "angle": 0, + "content": "Fig. 4: An architecture of our KineTy model." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.277, + 0.53, + 0.292 + ], + "angle": 0, + "content": "3.3 Ground-truth Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.3, + 0.784, + 0.345 + ], + "angle": 0, + "content": "Unlike generating random synthetic words in a training phase, we use real-world letters for evaluation by leveraging the templates. We can render ground-truth videos corresponding to user-input captions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.405 + ], + "angle": 0, + "content": "For evaluation, individual letters are selected for each alphabet letter from A to Z, and the first letters are capitalized for checking case sensitivity (e.g., Apple, Ball, ..., Zebra). These 26 words are used for rendering 584 templates, so that a total of 15,184 videos are used for the evaluation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.426, + 0.611, + 0.444 + ], + "angle": 0, + "content": "4 Kinetic Typography Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.785, + 0.53 + ], + "angle": 0, + "content": "In this section, we present our model for creating kinetic typography videos based on textual prompts. The key concept is to show animate visual effects on texts, and our challenges in this work are summarized as: (1) the effect should be visually-pleasing and eye-catching to viewers; (2) the transition between frames should be smooth and align with the captions; (3) the text must be readable." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.789, + 0.605 + ], + "angle": 0, + "content": "We begin by defining a kinetic typography generation problem in Sec. 4.1. We then describe how to model static and dynamic captions that efficiently guide appearance and motions in Sec. 4.2. We lastly discuss how to improve the glyph legibility with respect to the model design and its learning strategy in Sec. 4.3. Implementation details are provided in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.625, + 0.363, + 0.642 + ], + "angle": 0, + "content": "4.1 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.647, + 0.789, + 0.784 + ], + "angle": 0, + "content": "Conditional Latent Diffusion Models. Diffusion models train the data distribution by progressively refining a noisy initial state \\( z_{M} \\sim \\mathcal{N}(0,1) \\) into the target data representation \\( z_{0} \\) for \\( M \\) diffusion steps. Recent advancements, particularly in Latent Diffusion Models (LDMs) [46], enhance the efficiency by encoding an image \\( x \\) into a compact latent representation \\( z_{0} = \\mathcal{E}(x) \\) using an encoder \\( \\mathcal{E} \\), and is transformed back to the image \\( \\tilde{x} = \\mathcal{D}(z_{0}) \\) using a decoder \\( \\mathcal{D} \\). These adding noise and its subsequent removal are done with U-Net-style denoising network \\( \\epsilon_{\\pi} \\). In addition, a condition \\( y \\) is mapped to the hidden state \\( h \\) of \\( \\epsilon_{\\pi} \\) via attention:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.789, + 0.788, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t n} _ {\\theta} (h, y) = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {\\theta} (h) K _ {\\theta} (y) ^ {\\top}}{\\sqrt {d}}\\right) \\cdot V _ {\\theta} (y) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.287, + 0.827, + 0.717, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\mathrm {s . t .} Q _ {\\theta} (h) = W _ {\\theta , Q} \\cdot h, K _ {\\theta} (y) = W _ {\\theta , K} \\cdot y, V _ {\\theta} (y) = W _ {\\theta , V} \\cdot y,\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.363, + 0.128 + ], + "angle": 0, + "content": "S. Park et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "where \\( W_{\\theta, Q} \\), \\( W_{\\theta, K} \\) and \\( W_{\\theta, V} \\) denote the learnable parameters for queries, keys and values. \\( d \\) is the number of dimensions of the keys. The objective function for the denoising network \\( \\epsilon_{\\pi} \\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.205, + 0.788, + 0.23 + ], + "angle": 0, + "content": "\\[\nL _ {l d m} = \\mathbb {E} _ {\\mathcal {E} (x), y, \\epsilon \\sim \\mathcal {N} (0, 1), m} \\left[ \\left\\| \\epsilon - \\epsilon_ {\\pi} \\left(z _ {m}, m, y\\right) \\right\\| _ {2} ^ {2} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.242, + 0.786, + 0.272 + ], + "angle": 0, + "content": "where \\( \\| \\cdot \\|_2 \\) is the \\( L_2 \\) distance and diffusion step \\( m \\in \\{1, \\dots, M\\} \\) is uniformly sampled during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.274, + 0.789, + 0.379 + ], + "angle": 0, + "content": "Extending image diffusion models to video. Extending the capability of text-to-image diffusion models to video generation has been more feasible. By incorporating a temporal self-attention called motion module [18], it learns a temporal coherency between \\( T \\) frames in a latent sequence \\( \\mathbf{z}_0^{0:T} \\), which corresponds to the video sequence \\( \\tilde{\\mathbf{x}}^{0:T} = \\mathcal{D}(\\mathbf{z}_0^{0:T}) \\). This allows diffusion models to generate smoothly changed sequential images over time, whose temporal self-attention can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.277, + 0.39, + 0.788, + 0.426 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t n} _ {\\phi} \\left(\\mathbf {h} ^ {0: T}, \\mathbf {h} ^ {0: T}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {\\phi} \\left(\\mathbf {h} ^ {0 : T}\\right) K _ {\\phi} \\left(\\mathbf {h} ^ {0 : T}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\cdot V _ {\\phi} \\left(\\mathbf {h} ^ {0: T}\\right). \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.449, + 0.525, + 0.465 + ], + "angle": 0, + "content": "4.2 Spatial and Temporal Guidance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.786, + 0.522 + ], + "angle": 0, + "content": "Static caption incorporation. Motivated by the best practice of professional designers who handle appearance and motion effects separately, we divide the caption into static and dynamic elements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.614 + ], + "angle": 0, + "content": "When existing text-to-image models learn the distribution of image data, they are conditioned on texts related to the image's appearance. In the same manner, our model is also guided by captions describing the appearance of the text contents and background in each frame of video. Using Eq. (1), we define a spatial attention block with a self-attention followed by a cross-attention with a static caption \\(\\mathcal{C}_{\\text{static}}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.627, + 0.788, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} ^ {\\prime 0: T} = \\left\\{\\operatorname {A t t n} _ {\\vartheta} \\left(\\operatorname {A t t n} _ {\\theta} \\left(\\mathbf {h} ^ {t}, \\mathbf {h} ^ {t}\\right), \\tau \\left(\\mathcal {C} _ {\\text {s t a t i c}}\\right)\\right) \\right\\} _ {t = 0} ^ {T - 1}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.514, + 0.675 + ], + "angle": 0, + "content": "where \\(\\tau (\\cdot)\\) is the CLIP text encoder [44]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.676, + 0.789, + 0.767 + ], + "angle": 0, + "content": "Dynamic caption incorporation. A previous work [18] uses a motion module that only utilizes a self-attention between frames to learn the temporal consistency. On the other hand, in kinetic typography, it is essential to accurately display the dynamic motion effects of each letter in a video, following user's textual description. To do this, we extend Eq. (3) by adding the cross-attention with a dynamic caption as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.336, + 0.779, + 0.788, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} ^ {\\prime 0: T} = \\operatorname {A t t n} _ {\\psi} \\big (\\operatorname {A t t n} _ {\\phi} (\\mathbf {h} ^ {0: T}, \\mathbf {h} ^ {0: T}), \\tau \\big (\\mathcal {C} _ {\\text {d y n a m i c}} \\big) \\big), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Through this process, the model becomes capable of maintaining temporal consistency as well as direct control over dynamic movements." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.485, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Kinetic Typography Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.494, + 0.163 + ], + "angle": 0, + "content": "4.3 Enhancing Glyph Legibility." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.232 + ], + "angle": 0, + "content": "Readability improvement. As a next step, we aim to make the legibility of our model better. To do this, we first classify descriptions for text contents and prompt because diffusion models often have a difficulty in distinguishing them. In this work, we put down a delimiter between each character for text contents." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.233, + 0.789, + 0.308 + ], + "angle": 0, + "content": "Let a text content \\(\\mathcal{L} = \\{l_1, l_2, \\dots, l_L\\}\\) contain \\(L\\) letters. To condition each letter separately, we join each letter \\(l_i\\) in the text content with a delimiter symbol \\(|\\cdot\\rangle\\). Further, we add a symbol '\\~' behind upper case letters because the clip encoder does not care of upper cases. To the end, when the word \\(\\mathcal{L}\\) denotes \"Apple\", \\(\\mathcal{L}'\\) is represented as \"A^{|p|p|l|e}\". We denote the whole \\(\\mathcal{L}\\) as \\(\\mathcal{L}'\\) for better representation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.789, + 0.46 + ], + "angle": 0, + "content": "Attention on text contents. It can be challenging for the attention module to guide both text contents and effects at once, while disentangling the text content information in CLIP feature space. What is worse, the text content can be overwhelmed by relatively long appearance and motion effects in CLIP text encoder. To tackle this problem, we introduce an additional cross-attention branch to the text content. Inspired by ControlNet [76], we regard the text content as conditions through the zero convolution operation. Starting with the definition of word caption \\(\\mathcal{C}_{word}\\) via a prompt template like \"The word \\(\\{\\mathcal{L}'\\}\\)\", we extend Eq. (4) by adding the cross-attention module between the word caption and hidden feature, weighted by the zero-initialized convolutions \\(\\rho\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.276, + 0.47, + 0.788, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {h} _ {\\text {s t a t i c}} ^ {\\prime \\prime 0: T} = \\left\\{\\operatorname {A t t n} _ {\\vartheta} \\left(\\mathbf {h} ^ {\\prime}, \\tau \\left(\\mathcal {C} _ {\\text {s t a t i c}}\\right)\\right) + \\rho_ {v} \\left(\\operatorname {A t t n} _ {v} \\left(\\mathbf {h} ^ {\\prime}, \\tau \\left(\\mathcal {C} _ {\\text {w o r d}}\\right)\\right)\\right) \\right\\} _ {t = 0} ^ {T - 1}, \\tag {6} \\\\ w h e r e \\quad \\mathbf {h} ^ {\\prime} = \\operatorname {A t t n} _ {\\theta} \\left(\\mathbf {h} ^ {t}, \\mathbf {h} ^ {t}\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.569 + ], + "angle": 0, + "content": "This allows the network to gradually evaluate the usefulness of this additional caption based on the condition. In the same way, we incorporate the word caption into the temporal cross-attention in Eq. (5) as:" + }, + { + "type": "equation", + "bbox": [ + 0.282, + 0.58, + 0.788, + 0.624 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {h} _ {\\text {d y n a m i c}} ^ {\\prime \\prime 0: T} = \\operatorname {A t t n} _ {\\psi} \\left(\\mathbf {h} ^ {\\prime}, \\tau \\left(\\mathcal {C} _ {\\text {d y n a m i c}}\\right)\\right) + \\rho_ {\\varphi} \\left(\\operatorname {A t t n} _ {\\varphi} \\left(\\mathbf {h} ^ {\\prime}, \\tau \\left(\\mathcal {C} _ {\\text {w o r d}}\\right)\\right)\\right), \\tag {7} \\\\ w h e r e \\quad \\mathbf {h} ^ {\\prime} = \\operatorname {A t t n} _ {\\phi} \\left(\\mathbf {h} ^ {0: T}, \\mathbf {h} ^ {0: T}\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.72 + ], + "angle": 0, + "content": "Glyph loss. To train our model, we use a common loss function \\( L_{ldm} \\). Additionally, we impose an extra penalty on the letter regions to enforce a sharp and correct glyph of text contents. We first use a binary mask \\( B \\) for the letters in the last frame \\( V^T \\) from a text segmentation model [68]. The mask is then blurred to cover its surrounding effects. We then define a glyph loss based on \\( L_{LDM} \\) with the additional pixel-wise weighting strategy using the blurred mask as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.301, + 0.73, + 0.788, + 0.756 + ], + "angle": 0, + "content": "\\[\nL _ {g l y p h} = \\mathbb {E} _ {\\mathcal {E} (x), y, \\epsilon \\sim \\mathcal {N} (0, 1), m} \\left[ \\left\\| B \\odot \\left(\\epsilon - \\epsilon_ {\\pi} \\left(z _ {m} ^ {T}, m, y\\right)\\right) \\right\\| _ {2} ^ {2} \\right], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.531, + 0.779 + ], + "angle": 0, + "content": "where \\(\\odot\\) is the element-wise multiplication." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.841 + ], + "angle": 0, + "content": "With the glyph loss, we enhance the legibility of text contents, enabling the precise creation of multi-letter formations. Through the linear combination of both loss terms, we can formulate the final loss function as \\( L = L_{ldm} + \\alpha L_{glyph} \\). Here, we empirically set \\( \\alpha \\) to 0.01." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "S. Park et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.456, + 0.162 + ], + "angle": 0, + "content": "4.4 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.337 + ], + "angle": 0, + "content": "To train our model, we use a two-step training strategy similar to that of human trainees [14]. First, we pre-train the network to generate spatial appearances using only static caption and last video frame pairs. Here, we detach the temporal attention modules to make them work as text-to-image diffusion models during 30 epochs with a batch size of 200. Next, we train the full model while freezing the spatial attention using whole video frames and captions for an epoch with a batch size of 8. For training, we resize the height \\( H \\) and width \\( W \\) of the video into \\( 256 \\times 256 \\) with \\( T = 24 \\) frames. The training is performed with AdamW optimizer [35], with a diffusion step \\( M \\) of 1000 and a learning rate of 0.0001, which usually takes about 20 hours for training on 8 NVIDIA A100 GPUs. The inference time is about 20 seconds when the number of sampling steps is 25." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.359, + 0.376, + 0.375 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.788, + 0.48 + ], + "angle": 0, + "content": "In this section, we conduct comprehensive experiments to verify the effectiveness of our model for kinetic typography. We first describe the experimental setup in Sec. 5.1. We then provide comparison results with relevant typography and video generation models, and report user-study to highlight the practical applicability of our model in Sec. 5.2. We lastly carry out an extensive ablation study to validate the effect of each component in our model in Sec. 5.3" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.5, + 0.427, + 0.515 + ], + "angle": 0, + "content": "5.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Comparison methods. We compare ours with state-of-the-art generative models, which consist of 3 two-stage methods and 2 one-stage methods. The two-stage methods are based on combinations of text-to-image models, including DS-Fusion [52], GlyphControl [74] and Text-Diffuser [11], and an image-to-video model, SparseCtrl [17]. DS-Fusion outputs a stylized letter based on the given style phrase and a letter. Since our caption is a sentence, we extract a style keyword from it and consider it as a style phrase. Note that DS-Fusion is able to generate one letter at a time, so we concatenate each letter to make a full-text content. GlyphControl takes an instruction to generate a glyph image for text contents, then uses the image as a condition for the final output image. Text-Diffuser first finds a proper layout for words and produce the output based on a caption and the layout. Since the comparison methods yield images, we combine them with an image-to-video model, SparseCtrl, to compare with ours. SparseCtrl uses a caption and a guidance such as sketches, depth maps and images to generate a stylized video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "On the other hand, one-stage-methods,AnimateDiff [18] and Lavie [58], are based on text-to-video models.AnimateDiff extends pre-trained text-to-image diffusion models with a motion module, and trains only the motion module to make fully use of the massive information of video dataset. Lavie leverages a temporal self-attention block to enhance a temporal consistency between frames while keeping the generation performance." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.486, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Kinetic Typography Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 2: Quantitative evaluation of two-stage generation methods, and one-stage text-to-video generation methods. Bold: Best, Underline: Second-best." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.175, + 0.784, + 0.269 + ], + "angle": 0, + "content": "
ModelEntranceEmphasisMotionAverage
FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑
DS-F[52]+Sparse[17]1636.35.710.690.583184.95.970.730.581236.55.420.760.561843.105.660.700.58
Glyph[74]+Sparse[17]2415.64.300.760.682667.15.060.760.751468.64.480.840.802361.984.460.760.69
T-Diff[11]+Sparse[17]2937.25.610.780.903107.55.160.790.891208.24.060.820.952815.65.350.780.90
AnimateDiff[18]1613.84.710.590.012106.04.560.660.032160.64.510.670.071727.554.680.610.01
Lavie [58]825.12.250.820.541835.43.830.830.41505.443.330.850.35956.972.680.820.49
Ours147.41.790.870.71177.521.750.880.5436.161.620.940.98125.901.770.880.76
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.283, + 0.332, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.283, + 0.445, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.283, + 0.559, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.283, + 0.672, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.283, + 0.785, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.335, + 0.576, + 0.346 + ], + "angle": 0, + "content": "Static caption: The word 'Dog' in bold, green-stripped letters against a textured coral background. Dynamic caption: The sequential addition of letters one by one to spell out the word 'Dog'." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.348, + 0.332, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.348, + 0.445, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.348, + 0.559, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.348, + 0.672, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.348, + 0.785, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.397, + 0.776, + 0.41 + ], + "angle": 0, + "content": "Static caption: The word Mermaid in a blurred, tunnel-like visual effect, creating a sense of depth and distortion on a dark background. Dynamic caption: The word Mermaid gradually comes into focus, with each letter appearing in a random order, intensifying the sense of depth and creating a captivating rev" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.412, + 0.332, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.412, + 0.445, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.412, + 0.559, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.412, + 0.672, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.412, + 0.785, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.462, + 0.751, + 0.472 + ], + "angle": 0, + "content": "Static caption: The word 'Watermelon' in white letters against a black background, accompanied by two smiling emojis faces with heart eyes. Dynamic caption: Each character of the word 'Watermelon' appears sequentially, with various smiling emojis popping up randomly around the text." + }, + { + "type": "text", + "bbox": [ + 0.244, + 0.475, + 0.304, + 0.485 + ], + "angle": 0, + "content": "(a) DS-Fusion" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.485, + 0.303, + 0.492 + ], + "angle": 0, + "content": "+ SparseCtrl" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.475, + 0.424, + 0.485 + ], + "angle": 0, + "content": "(b) GlyphControl" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.485, + 0.415, + 0.493 + ], + "angle": 0, + "content": "+SparseCtrl" + }, + { + "type": "text", + "bbox": [ + 0.464, + 0.475, + 0.537, + 0.485 + ], + "angle": 0, + "content": "(c) Text-Diffuser" + }, + { + "type": "text", + "bbox": [ + 0.475, + 0.485, + 0.529, + 0.493 + ], + "angle": 0, + "content": "+ SparseCtrl" + }, + { + "type": "text", + "bbox": [ + 0.595, + 0.475, + 0.636, + 0.484 + ], + "angle": 0, + "content": "(d) Lavie" + }, + { + "type": "text", + "bbox": [ + 0.711, + 0.475, + 0.747, + 0.484 + ], + "angle": 0, + "content": "(e) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.498, + 0.785, + 0.525 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results from the comparison models and ours. Obviously, the results from ours reflect the contents of captions better than the others." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.785, + 0.606 + ], + "angle": 0, + "content": "For a fair comparison, we utilize the official source codes and pre-trained weights provided by the authors. For more analysis, we categorize the motion effects into three groups, 'Entrance', 'Emphasis' and 'Motion', and conduct experiments to check the quality of each result." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.787, + 0.758 + ], + "angle": 0, + "content": "Evaluation metrics. Since the final products of kinetic typography depend on the designer's preferences, we adopt four feature-based metrics rather than photoconsistency measures like PSNR. (1) Fréchet Video Distance (FVD) qualifies a similarity between two videos by comparing their feature representations extracted from a pre-trained deep neural network [55]. (2) Inception Score (IS) evaluates the quality and diversity of images generated by a model [4]. (3) CLIPScore measures the semantic alignment between generated images and their textual descriptions, using the CLIP model [19, 22, 44]. (4) Optical Character Recognition (OCR) checks the clarity of the output by comparing recognized text to a reference text using the F1-score." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.783, + 0.417, + 0.796 + ], + "angle": 0, + "content": "5.2 Evaluation Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Quantitative results. As shown in Tab. 2, our Kinety model achieves the better performance than the comparison methods in the most metrics on all the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "S. Park et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.547, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.29, + 0.268, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.29, + 0.311, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.29, + 0.362, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.29, + 0.394, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.407, + 0.29, + 0.434, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.29, + 0.482, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.29, + 0.522, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.144, + 0.787, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.29, + 0.6, + 0.3 + ], + "angle": 0, + "content": "The word" + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.29, + 0.631, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.29, + 0.661, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.29, + 0.69, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.29, + 0.72, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.29, + 0.749, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.304, + 0.784, + 0.332 + ], + "angle": 0, + "content": "Fig. 6: Visualization of attention maps between letters from a caption and feature map of a hidden layer." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.347, + 0.787, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.385, + 0.446, + 0.396 + ], + "angle": 0, + "content": "(a) Initial kinetic typography video" + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.385, + 0.701, + 0.395 + ], + "angle": 0, + "content": "(b) Changing text color" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.399, + 0.787, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.437, + 0.415, + 0.448 + ], + "angle": 0, + "content": "(c) Changing text font" + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.437, + 0.708, + 0.448 + ], + "angle": 0, + "content": "(d) Changing background" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.452, + 0.787, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.489, + 0.428, + 0.5 + ], + "angle": 0, + "content": "(e) Changing motion effect" + }, + { + "type": "image_caption", + "bbox": [ + 0.603, + 0.489, + 0.686, + 0.5 + ], + "angle": 0, + "content": "(f) Final product" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.504, + 0.787, + 0.532 + ], + "angle": 0, + "content": "Fig. 7: The demonstration of the editable capability of our Kinecty model after generating initial kinetic typography." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.702 + ], + "angle": 0, + "content": "categories. The comparison methods, despite their impressive performance in general image and video generation, show unsatisfactory performance in creating kinetic typography. The reasons are mainly two-folds: (1) Text-to-video models are not specialized for kinetic typography and often fail to exactly distinguish the text content in the long captions. Different from them, we additionally feed \\( \\mathcal{C}_{word} \\) which imposes a high attention to the text content in the video. (2) They have challenges in creating letter-level motions. Since they are unaware of each letter in the text content, all the letters move as a whole or inconsistently. We handle this issue by providing a delimiter between each letter and using zero-convolution in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Our model shows promising results in generating legible text content, but it sometimes falls short in OCR accuracy. This is because the two-stage methods algorithmically generate a glyph image of a text content, and then deform glyphs conditioned on the glyph image. As a result, they have the better performance in OCR metrics than ours and one-stage methods. Nevertheless, our method outperforms in the most cases, demonstrating the effectiveness of our method in generating the clear text content and motion effects." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Qualitative results. In Fig. 5, we display three examples of the comparison methods. Compared to ours, the two-stage methods do not understand the user's" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.486, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Kinetic Typography Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.787, + 0.173 + ], + "angle": 0, + "content": "Table 3: The result of user-study. The score ranges from \\(1 \\sim 6\\) for Study 1 & 2, and 0 \\(\\sim 1\\) for Study 3. Bold: Best, Underline: Second-best." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.175, + 0.784, + 0.338 + ], + "angle": 0, + "content": "
User StudyDynamic EffectGeneralExpert
DS-F +SparseGlyph +SparseT-Diff +SparseAnimate DiffLavieOursDS-F +SparseGlyph +SparseT-Diff +SparseAnimate DiffLavieOurs
Study 1Entrance3.3843.6873.8912.4063.1534.4783.4814.0254.1311.0193.0815.263
Emphasis3.5623.7134.1192.5632.7254.3193.6254.1754.2251.0502.6255.300
Motion2.3634.1384.1133.3382.3754.6751.7003.5504.0004.0002.0005.750
Average3.3183.7133.9232.5183.0364.4943.3353.9954.0901.3252.9055.350
Study 2Entrance3.5143.7753.9412.4133.2704.0883.6504.0814.2441.0193.1064.900
Emphasis3.5313.8814.0752.2693.0504.1943.7254.1504.0501.0252.6755.375
Motion2.8383.9503.8753.2502.8504.2382.1504.0004.4503.4501.9505.000
Average3.4393.7993.9482.4603.2204.1353.4854.0604.2551.2652.9704.995
Study 3Entrance0.6280.6200.6520.0630.1690.7250.8440.8560.8870.0190.1380.969
Emphasis0.5810.6060.6380.0880.1060.7810.7500.7000.9000.0000.0001.000
Motion0.0750.7130.7500.3620.2880.7810.0000.9000.9000.5000.2501.000
Average0.5530.6260.6580.0960.1740.7440.7300.8350.8800.0650.1350.975
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.785, + 0.462 + ], + "angle": 0, + "content": "instruction with respect to either the whole caption or the letter's movement. Here, we need to know how to perform our Kinecty model. Figure 6 visualizes the attention map between the caption and features from the hidden state on the corresponding images. Thanks to the cross-attention between the word caption and the noisy latent, and the delimiter to separate the words, our model successfully embeds features for each letter in the noise." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.787, + 0.543 + ], + "angle": 0, + "content": "Lastly, to show the generality of our KineTy model, we conduct an additional experiment. After generating an initial kinetic typography, our KineTy model allows users to modify its style and motions such as the content color, background, font and motion, and even to combine them. As demonstrated in Fig. 7, ours can render the modified outcomes corresponding to the additional captions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.669 + ], + "angle": 0, + "content": "User Study. To assess the practical utility of our results in the field of kinetic typography, we conduct a user study using Amazon MTurk. We ask 20 questionnaires to 50 participants. Since their familiarity of typography can vary significantly, we divide them into two groups: experts and non-experts, consisting of 10 and 40 individuals, respectively. The study involves three sections: (1) caption alignment; (2) kinetic typography suitability; (3) Word readability. We make two video clips by randomly choosing two words for each effect. In total, we make 20 clips for this user study." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In the first study, we measure how well ours and the comparison methods generate videos that align with provided captions. The participants see 6 videos from different models with the same corresponding caption in a random order. They are asked to rank the videos from 1 to 6, and the rankings are subsequently converted into scores, ranging from 6 (Best alignment) to 1 (Worst alignment). The following study evaluates how proper the generated videos are for motion graphic applications. Since some participants might be unfamiliar with kinetic typography, all participants watch 4 example videos from online in advance. After that, they check six videos in a random order and rank them. Scores were assigned based on these rankings, from 6 (the most proper) to 1 (the least proper). The last study assesses how readable the outcomes are. After watching 6 videos, the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "S. Park et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.251, + 0.145, + 0.75, + 0.159 + ], + "angle": 0, + "content": "Table 4: The result of ablation study. Bold: Best, Underline: Second-best." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.161, + 0.784, + 0.236 + ], + "angle": 0, + "content": "
ModelEntranceEmphasisMotionAverage
FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑
Ours147.41.790.870.71177.521.750.880.5436.161.620.940.98125.901.770.880.76
-Effect separation789.82.750.740.49880.12.880.750.49607.12.540.820.23729.82.730.750.48
-Cword239.11.770.860.70247.11.730.870.5462.71.600.950.95201.81.740.870.74
-Lglph188.91.770.860.70226.81.750.880.5550.11.590.930.98160.81.750.880.75
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.26, + 0.786, + 0.289 + ], + "angle": 0, + "content": "participants are needed to vote for videos with readable text contents. Of course, multiple voting is available." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.29, + 0.787, + 0.366 + ], + "angle": 0, + "content": "The performance of each method is measured based on scores for the first two studies, and on the selection ratio for third one. As demonstrated in Tab. 3, our approach shows the promising performance across all studies. The results highlight our model's practical usability in creating kinetic typography, especially for the domain experts who give our model the highest scores." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.39, + 0.389, + 0.406 + ], + "angle": 0, + "content": "5.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.787, + 0.508 + ], + "angle": 0, + "content": "Without caption separation. First of all, we train our network without dividing static and dynamic caption. The performances drop significantly for every metrics in Tab. 4. Since we train the text-to-image backbone using only static captions, dynamic captions are considered as noisy input. Similarly, the motion module, used to train dynamic motions, is effective when the simple motion guidance is given, compared to using static and dynamic caption together." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.787, + 0.583 + ], + "angle": 0, + "content": "Text contents incorporation. We observe that text-to-image diffusion models often fail to catch up on text contents in the prompt when a long description is given. We thus use \\( \\mathcal{C}_{word} \\) as an additional input to give a strong attention to each letter. As shown in Tab. 4, without \\( \\mathcal{C}_{word} \\), the performance degradation is observed as expected." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.787, + 0.63 + ], + "angle": 0, + "content": "Without glyph loss. Lastly, we evaluate the performance without \\( L_{glyph} \\). The inferior FVD scores come from the blur effect at the edge of the word along with color bleeding effects." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.656, + 0.36, + 0.672 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we propose a generative kinetic typography model, named KineTy. To achieve this, we first build a large-scale kinetic typography dataset by collecting 584 templates from professional designers, and use them to train our diffusion-based network. To effectively handle the text prompt, we split it into static and dynamic captions. They are used to directly guide the spatial and temporal cross-attention for better appearance and motion effects, respectively. Our glyph loss also strengthens the legible and artistic letter generation. We lastly demonstrate that our KineTy produces the visually compelling and semantically coherent kinetic typography videos. The extensive user studies further validate the effectiveness of our model with respect to its practical utility." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.486, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Kinetic Typography Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "Acknowledgements This research was supported by 'Project for Science and Technology Opens the Future of the Region' program through the INNOPOLIS FOUNDATION funded by Ministry of Science and ICT (Project Number: 2022-DD-UP-0312), the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(RS-2024-00338439), GIST-MIT Research Collaboration grant funded by the GIST in 2024, and the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-01842, Artificial Intelligence Graduate School Program (GIST))." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.305, + 0.323, + 0.32 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.333, + 0.785, + 0.36 + ], + "angle": 0, + "content": "1. Anderson, D., Shamir, A., Fried, O.: Neural font rendering. arXiv preprint arXiv:2211.14802 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.361, + 0.787, + 0.403 + ], + "angle": 0, + "content": "2. Azadi, S., Fisher, M., Kim, V.G., Wang, Z., Shechtman, E., Darrell, T.: Multicontent gan for few-shot font style transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.403, + 0.787, + 0.443 + ], + "angle": 0, + "content": "3. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.444, + 0.785, + 0.47 + ], + "angle": 0, + "content": "4. Barratt, S., Sharma, R.: A note on the inception score. arXiv preprint arXiv:1801.01973 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.471, + 0.785, + 0.498 + ], + "angle": 0, + "content": "5. Berio, D., Leymarie, F.F., Asente, P., Echevarria, J.: Strokestyles: Stroke-based segmentation and stylization of fonts. ACM Transactions on Graphics (TOG) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.499, + 0.787, + 0.552 + ], + "angle": 0, + "content": "6. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.553, + 0.787, + 0.607 + ], + "angle": 0, + "content": "7. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.608, + 0.785, + 0.636 + ], + "angle": 0, + "content": "8. Campbell, N.D., Kautz, J.: Learning a manifold of fonts. ACM Transactions on Graphics (TOG) (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.636, + 0.785, + 0.676 + ], + "angle": 0, + "content": "9. Chen, C.H., Liu, Y.T., Zhang, Z., Guo, Y.C., Zhang, S.H.: Joint implicit neural representation for high-fidelity and compact vector fonts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.677, + 0.785, + 0.717 + ], + "angle": 0, + "content": "10. Chen, J., Huang, Y., Lv, T., Cui, L., Chen, Q., Wei, F.: Textdiffuser-2: Unleashing the power of language models for text rendering. arXiv preprint arXiv:2311.16465 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.718, + 0.785, + 0.758 + ], + "angle": 0, + "content": "11. Chen, J., Huang, Y., Lv, T., Cui, L., Chen, Q., Wei, F.: Textdiffuser: Diffusion models as text painters. Proceedings of the Neural Information Processing Systems (NeurIPS) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "12. Esser, P., Chiu, J., Atighehchian, P., Granskog, J., Germanidis, A.: Structure and content-guided video synthesis with diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "13. Ford, S., Forlizzi, J., Ishizaki, S.: Kinetic typography: issues in time-based presentation of text. In: CHI'97 extended abstracts on Human factors in computing systems, pp. 269-270. ACM Digital Library (1997)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.333, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "S. Park et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "14. Fridsma, L., Gyncild, B.: Adobe After Effects CC Classroom in a Book. Adobe Press (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.785, + 0.216 + ], + "angle": 0, + "content": "15. Fu, B., He, J., Wang, J., Qiao, Y.: Neural transformation fields for arbitrary-styled font generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.216, + 0.785, + 0.271 + ], + "angle": 0, + "content": "16. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.785, + 0.311 + ], + "angle": 0, + "content": "17. Guo, Y., Yang, C., Rao, A., Agrawala, M., Lin, D., Dai, B.: Sparsectrl: Adding sparse controls to text-to-video diffusion models. arXiv preprint arXiv:2311.16933 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.311, + 0.785, + 0.352 + ], + "angle": 0, + "content": "18. Guo, Y., Yang, C., Rao, A., Wang, Y., Qiao, Y., Lin, D., Dai, B.: AnimateDiff: Animate your personalized text-to-image diffusion models without specific tuning arXiv preprint arXiv:2307.04725 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.352, + 0.785, + 0.393 + ], + "angle": 0, + "content": "19. Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.434 + ], + "angle": 0, + "content": "20. Ho, J., Chan, W., Saharia, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.473 + ], + "angle": 0, + "content": "21. Hu, Y., Luo, C., Chen, Z.: Make it move: controllable image-to-video generation with text descriptions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "22. Huang, Y., Xue, H., Liu, B., Lu, Y.: Unifying multimodal transformer for bidirectional image and text generation. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 1138-1147 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.542 + ], + "angle": 0, + "content": "23. Iluz, S., Vinker, Y., Hertz, A., Berio, D., Cohen-Or, D., Shamir, A.: Word-as-image for semantic typography. ACM Transactions on Graphics (TOG) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.595 + ], + "angle": 0, + "content": "24. Jahanian, A., Liu, J., Lin, Q., Tretter, D., O'Brien-Strain, E., Lee, S.C., Lyons, N., Allebach, J.: Recommendation system for automatic design of magazine covers In: Proceedings of the 2013 International Conference on Intelligent User Interfaces (IUI) (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.785, + 0.637 + ], + "angle": 0, + "content": "25. Jia, P., Li, C., Liu, Z., Shen, Y., Chen, X., Yuan, Y., Zheng, Y., Chen, D., Li, J., Xie, X., et al.: Cole: A hierarchical generation framework for graphic design. arXiv preprint arXiv:2311.16974 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.678 + ], + "angle": 0, + "content": "26. Kato, J., Nakano, T., Goto, M.: Textalive: Integrated design environment for kinetic typography. In: Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems (CHI) (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "27. Krishnan, P., Kovvuri, R., Pang, G., Vassilev, B., Hassner, T.: Textstylebrush: Transfer of text aesthetics from a single example. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "28. Lee, J.C., Forlizzi, J., Hudson, S.E.: The kinetic typography engine: an extensible system for animating expressive text. In: Proceedings of the 15th annual ACM symposium on User Interface Software and Technology (UIST) (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "29. Lee, J., Jun, S., Forlizzi, J., Hudson, S.E.: Using kinetic typography to convey emotion in text-based interpersonal communication. In: Proceedings of the 6th conference on Designing Interactive systems (DIS) (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "30. Li, C., Taniguchi, Y., Lu, M., Konomi, S.: Few-shot font style transfer between different languages. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.486, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Kinetic Typography Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "31. Li, X., Wu, L., Wang, C., Meng, L., Meng, X.: Compositional zero-shot artistic font synthesis. In: Proceedings of the International Joint Conferences on Artificial Intelligence (IJCAI) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.229 + ], + "angle": 0, + "content": "32. Liu, X., Meng, G., Chang, J., Hu, R., Xiang, S., Pan, C.: Decoupled representation learning for character glyph synthesis. IEEE Transactions on Multimedia (TMM) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.229, + 0.788, + 0.271 + ], + "angle": 0, + "content": "33. Liu, Y.T., Guo, Y.C., Li, Y.X., Wang, C., Zhang, S.H.: Learning implicit glyph shape representation. IEEE Transactions on Visualization and Computer Graphics (TVCG) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.325 + ], + "angle": 0, + "content": "34. Liu, Y.T., Zhang, Z., Guo, Y.C., Fisher, M., Wang, Z., Zhang, S.H.: Dualvector: Unsupervised vector font synthesis with dual-part representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.352 + ], + "angle": 0, + "content": "35. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: Proceedings of the International Conference on Learning Representations (ICLR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.407 + ], + "angle": 0, + "content": "36. Luo, Z., Chen, D., Zhang, Y., Huang, Y., Wang, L., Shen, Y., Zhao, D., Zhou, J., Tan, T.: Videofusion: Decomposed diffusion models for high-quality video generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.407, + 0.788, + 0.447 + ], + "angle": 0, + "content": "37. Men, Y., Lian, Z., Tang, Y., Xiao, J.: Dyntypo: Example-based dynamic text effects transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.488 + ], + "angle": 0, + "content": "38. Minakuchi, M., Tanaka, K.: Automatic kinetic typography composer. In: Proceedings of the ACM SIGCHI International Conference on Advances in Computer Entertainment Technology (ACE) (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.529 + ], + "angle": 0, + "content": "39. Mu, X., Chen, L., Chen, B., Gu, S., Bao, J., Chen, D., Li, J., Yuan, Y.: Fontstudio: Shape-adaptive diffusion model for coherent and consistent font effect generation. arXiv preprint arXiv:2406.08392 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.529, + 0.788, + 0.555 + ], + "angle": 0, + "content": "40. Nagata, Y., Iwana, B.K., Uchida, S.: Contour completion by transformers and its application to vector font data. arXiv preprint arXiv:2304.13988 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.555, + 0.788, + 0.582 + ], + "angle": 0, + "content": "41. OpenAI: GPT-4V(ision) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.582, + 0.788, + 0.623 + ], + "angle": 0, + "content": "42. Pan, W., Zhu, A., Zhou, X., Iwana, B.K., Li, S.: Few shot font generation via transferring similarity guided global style and quantization local style. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.623, + 0.788, + 0.663 + ], + "angle": 0, + "content": "43. Qu, Y., Tan, Q., Xie, H., Xu, J., Wang, Y., Zhang, Y.: Exploring stroke-level modifications for scene text editing. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.663, + 0.788, + 0.718 + ], + "angle": 0, + "content": "44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proceedings of the International Conference on Machine Learning (ICML) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.759 + ], + "angle": 0, + "content": "45. Reddy, P., Zhang, Z., Wang, Z., Fisher, M., Jin, H., Mitra, N.: A multi-implicit neural representation for fonts. Proceedings of the Neural Information Processing Systems (NeurIPS) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.8 + ], + "angle": 0, + "content": "46. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.8, + 0.788, + 0.84 + ], + "angle": 0, + "content": "47. Shimoda, W., Haraguchi, D., Uchida, S., Yamaguchi, K.: De-rendering stylized texts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "S. Park et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "48. Shimoda, W., Haraguchi, D., Uchida, S., Yamaguchi, K.: Towards diverse and consistent typography generation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.243 + ], + "angle": 0, + "content": "49. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. In: Proceedings of the International Conference on Learning Representations (ICLR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.786, + 0.271 + ], + "angle": 0, + "content": "50. Smith, J., Team, A.C.: Adobe After Effects CS6 Digital Classroom. John Wiley & Sons (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.787, + 0.311 + ], + "angle": 0, + "content": "51. Subramanian, J., Chordia, V., Bart, E., Fang, S., Guan, K., Bala, R., et al.: Strive: Scene text replacement in videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.787, + 0.352 + ], + "angle": 0, + "content": "52. Tanveer, M., Wang, Y., Mahdavi-Amiri, A., Zhang, H.: Ds-fusion: Artistic typography via discriminated and stylized diffusion. arXiv preprint arXiv:2303.09604 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.787, + 0.407 + ], + "angle": 0, + "content": "53. Thamizharasan, V., Liu, D., Agarwal, S., Fisher, M., Gharbi, M., Wang, O., Jacobson, A., Kalogerakis, E.: Vecfusion: Vector font generation with diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.787, + 0.447 + ], + "angle": 0, + "content": "54. Tuo, Y., Xiang, W., He, J.Y., Geng, Y., Xie, X.: Anytext: Multilingual visual text generation and editing. In: Proceedings of the International Conference on Learning Representations (ICLR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.787, + 0.488 + ], + "angle": 0, + "content": "55. Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.528 + ], + "angle": 0, + "content": "56. Wang, C., Wu, L., Liu, X., Li, X., Meng, L., Meng, X.: Anything to glyph: Artistic font synthesis via text-to-image diffusion model. In: SIGGRAPH Asia 2023 Conference Papers (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.787, + 0.57 + ], + "angle": 0, + "content": "57. Wang, C., Zhou, M., Ge, T., Jiang, Y., Bao, H., Xu, W.: Cf-font: Content fusion for few-shot font generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.787, + 0.609 + ], + "angle": 0, + "content": "58. Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., Yang, P., et al.: Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.787, + 0.651 + ], + "angle": 0, + "content": "59. Wang, Y., Ye, Y., Mao, Y., Yu, Y., Song, Y.: Self-supervised scene text segmentation with object-centric layered representations augmented by text regions. In: Proceedings of the 30th ACM International Conference on Multimedia (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.787, + 0.677 + ], + "angle": 0, + "content": "60. Wang, Y., Lian, Z.: Deepvecfont: Synthesizing high-quality vector fonts via dual-modality learning. ACM Transactions on Graphics (TOG) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.787, + 0.718 + ], + "angle": 0, + "content": "61. Wang, Y., Pu, G., Luo, W., Wang, Y., Xiong, P., Kang, H., Lian, Z.: Aesthetic text logo synthesis via content-aware layout inferring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "62. Wang, Y., Wang, Y., Yu, L., Zhu, Y., Lian, Z.: Deepvecfont-v2: Exploiting transformers to synthesize vector fonts with higher quality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "63. Wong, Y.Y.: Temporal typography: a proposal to enrich written expression. In: Proceedings of the Conference Companion on Human Factors in Computing Systems (CHI) (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "64. Xia, Z., Xiong, B., Lian, Z.: Vecfontsf: Learning to reconstruct and synthesize high-quality vector fonts via signed distance functions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.486, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Kinetic Typography Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "65. Xie, L., Shu, X., Su, J.C., Wang, Y., Chen, S., Qu, H.: Creating emordle: Animating word cloud for emotion expression. IEEE Transactions on Visualization and Computer Graphics (TVCG) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "66. Xie, L., Zhou, Z., Yu, K., Wang, Y., Qu, H., Chen, S.: Wakey-wakey: animate text by mimicking characters in a gif. In: Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.286 + ], + "angle": 0, + "content": "67. Xu, C., Zhou, M., Ge, T., Jiang, Y., Xu, W.: Unsupervised domain adaption with pixel-level discriminator for image-aware layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.342 + ], + "angle": 0, + "content": "68. Xu, X., Zhang, Z., Wang, Z., Price, B., Wang, Z., Shi, H.: Rethinking text segmentation: A novel dataset and a text-specific refinement approach. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.383 + ], + "angle": 0, + "content": "69. Yang, Q., Huang, J., Lin, W.: Swaptext: Image based texts transfer in scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.425 + ], + "angle": 0, + "content": "70. Yang, S., Liu, J., Lian, Z., Guo, Z.: Awesome typography: Statistics-based text effects transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.452 + ], + "angle": 0, + "content": "71. Yang, S., Wang, W., Liu, J.: Te141k: artistic text benchmark for text effect transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.452, + 0.788, + 0.493 + ], + "angle": 0, + "content": "72. Yang, S., Wang, Z., Liu, J.: Shape-matching \\(\\mathrm{gan} + +\\): Scale controllable dynamic artistic text style transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.535 + ], + "angle": 0, + "content": "73. Yang, S., Wang, Z., Wang, Z., Xu, N., Liu, J., Guo, Z.: Controllable artistic text style transfer via shape-matching gan. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.577 + ], + "angle": 0, + "content": "74. Yang, Y., Gui, D., Yuan, Y., Liang, W., Ding, H., Hu, H., Chen, K.: Glyphcontrol: Glyph conditional control for visual text generation. In: Proceedings of the Neural Information Processing Systems (NeurIPS) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.619 + ], + "angle": 0, + "content": "75. Yang, Z., Peng, D., Kong, Y., Zhang, Y., Yao, C., Jin, L.: Fontdiffuser: One-shot font generation via denoising diffusion with multi-scale content aggregation and style contrastive learning. arXiv preprint arXiv:2312.12142 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.66 + ], + "angle": 0, + "content": "76. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.702 + ], + "angle": 0, + "content": "77. Zhang, S., Ma, J., Wu, J., Ritchie, D., Agrawala, M.: Editing motion graphics video via motion vectorization and transformation. ACM Transactions on Graphics (TOG) (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.702 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Kinetic Typography Diffusion Model/d7a51e44-9061-4ac8-9c76-37fd7685fe8a_origin.pdf b/2024/Kinetic Typography Diffusion Model/d7a51e44-9061-4ac8-9c76-37fd7685fe8a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dc8fa2d2e84dd2be8240d2f010901a925d0b940a --- /dev/null +++ b/2024/Kinetic Typography Diffusion Model/d7a51e44-9061-4ac8-9c76-37fd7685fe8a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de5d554dabbdd6dfec75b8d743c101bb31f6f923dea40e778239d8cfd5f7d547 +size 9223589 diff --git a/2024/Kinetic Typography Diffusion Model/full.md b/2024/Kinetic Typography Diffusion Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f603e8d923989a191a05a83521a7a2b07fdddad4 --- /dev/null +++ b/2024/Kinetic Typography Diffusion Model/full.md @@ -0,0 +1,457 @@ +# Kinetic Typography Diffusion Model + +Seonmi Park, Inhwan Bae, Seunghyun Shin, and Hae-Gon Jeon + +AI Graduate School, GIST, South Korea +{bluesky1000, inhwanbae, seunghyuns98}@gm.gist.ac.kr, haegonj@gist.ac.kr +https://seonmip.github.io/kinety + +Abstract. This paper introduces a method for realistic kinetic typography that generates user-preferred animatable "text content". We draw on recent advances in guided video diffusion models to achieve visually-pleasing text appearances. To do this, we first construct a kinetic typography dataset, comprising about 600K videos. Our dataset is made from a variety of combinations in 584 templates designed by professional motion graphics designers and involves changing each letter's position, glyph, and size (i.e., flying, glitches, chromatic aberration, reflecting effects, etc.). Next, we propose a video diffusion model for kinetic typography. For this, there are three requirements: aesthetic appearances, motion effects, and readable letters. This paper identifies the requirements. For this, we present static and dynamic captions used as spatial and temporal guidance of a video diffusion model, respectively. The static caption describes the overall appearance of the video, such as colors, texture and glyph which represent a shape of each letter. The dynamic caption accounts for the movements of letters and backgrounds. We add one more guidance with zero convolution to determine which text content should be visible in the video. We apply the zero convolution to the text content, and impose it on the diffusion model. Lastly, our glyph loss, only minimizing a difference between the predicted word and its ground-truth, is proposed to make the prediction letters readable. Experiments show that our model generates kinetic typography videos with legible and artistic letter motions based on text prompts. + +Keywords: Kinetic Typography $\cdot$ Video Diffusion $\cdot$ Visual Text Generation + +# 1 Introduction + +Kinetic typography is an artistic motion graphics design combining text and animations [13]. Based on the word's meaning, motions are generated to convey information in videos. The main goal is eye-catching and to improve message retention. With the rise of video media, it has become an essential element in TV programs, commercials, music videos and film leaders. + +Kinetic typography controls letters' shape (glyph), color and texture over time, and transforms their positions. Professional motion graphic designers + +![](images/1b7e4304a1c291bb2fa7195c891d6b6d3b7e5cbc489435d4b6e83f0f094e8022.jpg) +(a) Designer's workflow + +![](images/8e0ab68d1e3e2732eb91503e1f6aaa3c140574b36324ae03bd1c67ae76ac8705.jpg) +(b) Our Kinety model +Fig. 1: An overview of our KineTy pipeline, which is motivated by the designer's workflow. Our key idea is to generate eye-catching and aesthetic animatable words based on user instructions. + +use commercial software like 'Adobe After Effects' to make and render kinetic typography videos. The conventional pipeline is as follows [14, 50]; (1) Define a text box in a workspace called composition and enter a text (i.e. word or message). (2) Edit the static text by tuning its font, color and texture (3) Build the background if needed (4) Apply motion effects for all text and background (5) Repeat, optimize, and fine-tune the whole process until it satisfies user intention. This process is time-consuming and labor-intensive. It takes three hours for simple motions to several days for sophisticated effects per single kinetic typography video [14, 26]. To efficiently do this, designers may need to consider either various design references or pre-defined design options called templates. + +With the advancement of generative models, there have been attempts to produce dynamic typography. The literature on typography generation has mostly focused on creating static single-letter images. Starting from transforming the shape of the object into a single letter [23, 39, 52, 56], there have been works to create multiple letters and multiple words by introducing a concept of layout [10, 11, 74]. Furthermore, DynTypo [37] and Shape-Matching GAN++ [72] adopt style-transfer techniques to produce animatable effects on a single letter. They focus on only animatable effects without any actual motions of letters. + +Meanwhile, the recent advancement of the video generation model has been actively studied with the success of text-guided video diffusion models, especially user's description conditions video frames [7,16,18,20,49]. Although these models have opened up the possibility of creating kinetic typography, there is a critical issue that the video generations show a weak understanding of the letters' shapes and motions. + +In this paper, we propose the Kinetic Typography diffusion model (called KineTy model) that generates kinetic typography from user-provided text prompts. Inspired by the video diffusion model, we allow users to input comprehensive descriptions for color, font, size, position and motion effects of letters. To better represent text motions, we first introduce our kinetic typography video dataset. 600K videos are rendered by combining randomly generated text contents with 584 templates made by professional motion graphic designers. These videos are labeled with static and dynamic captions that describe the video with respect to its appearance and motion characteristics, respectively. + +Next, we present KineTy model that effectively synthesizes videos from the text prompt. Here, we reevaluate how to effectively condition the caption guidance + +into the video diffusion model at a fundamental level. We enforce that each effect is incorporated by separately inserting static and dynamic captions into the spatial attention and temporal attention, respectively. To strengthen these attentions, we apply a zero convolution [76] into the word caption, and add it to them. We use a mask loss term only for the generated video by masking out the background contents, which makes the text contents more readable. + +Experimental results show that our Kinecty robustly generates kinetic typography video with multi-letter legibility, while accurately representing captions. Furthermore, extensive and meticulous user studies support our claim that Kinecty produces more aesthetic outcomes than general-purpose video generations. + +# 2 Related Work + +# 2.1 Typography Generation + +The key of typography is to make readable and visuallyappealing text contents to readers. Most research in this field has focused on transferring the style of a single-letter design, including elements like color, glyph, font and effects, to other letters [1,2,5,8,9,15,30-34,40,42,45,53,57,60,62,64,70,75]. Additionally, there have been works on transferring an image style to the text contents [73]. Similarly, research on scene text editing, which changes text content while maintaining its own style in a scene, is also underway [27,43,54,69], whose extended version to video is available in [51]. In addition to changing text contents, color editing [47] and text segmentation [59,68] are proposed. + +With the recent success of the high-fidelity text-to-image diffusion model [46], typography generation has gained interest. Works in [23,52,56] apply image styles into the target letter in an unsupervised manner. + +There are concurrent works that display multi-letters, beyond the single-letter generation. They follow the multi-step approach that firstly generates layouts at specific positions, arrange them with multi-letters or multi-words with the same fonts and colors, and contextually infers the background [10, 11, 24, 25, 48, 61, 67, 74, 77]. However, these models are not specialized for typography. Since all letters are generated together, it is not editable for the letters, which makes it difficult to animate and add movement to each character. + +# 2.2 Typography Video Generation + +As we mentioned above, this work is the first attempt to generate kinetic typography. Although no existing works directly align with this, there are some works related to typography videos [26, 28, 29, 38, 63, 65]. Here, we would like to introduce two methods related to ours. + +DynTypo [37] proposes a dynamic typography model that transfers the dynamic effect with realistic movements like fire and water on a specific uppercase English letter to others. Shape-Matching GAN++ [72] transfers an image style into a target letter by matching these shapes with structures of the target letter. + +Table 1: A summary of the kinetic typography datasets. + +
DatasetDomain#SamplesVideoStatic CaptionDynamic CaptionMulti-letterText AppearanceBackground Text Movements
Multi-Content [2]Typography10K----✓-
TextLogo3K [61]Typography3K----✓-
TE141K [71]Typography141K---✓✓✓
TextSeg [68]Text Segmentation4K---✓✓✓
LAION-Glyph-10M [74]Text-to-Image10M-✓-✓✓✓
MARIO-10M [11]Text-to-Image10M-✓-✓✓✓
AnyWord-3M [54]Text-to-Image3M-✓-✓✓✓
CATER-GEN [21]Image-to-Video35K✓-✓N/AN/A△
WebVid-10M [3]Text-to-Video10M✓✓△N/AN/A✓
Vimeo25M [58]Text-to-Video25M✓✓△N/AN/A✓
OursKinetic Typography600K✓✓✓✓✓✓
+ +However, the position and glyph of the letter are still fixed, with no actual movement between frames. Even under the condition of a single uppercase letter, representing various static and dynamic effects through text prompts remains a challenging issue. + +By substituting the motion description with external user input, the task becomes more tractable. A work, Wakey-Wakey [66], transfers the source GIF (Graphics Interchange Format) motion to the target text. Here, graphic designers manually assign additional corresponding key points to achieve motion transfer. However, this process still requires direct human intervention. + +# 2.3 Text-to-Video Diffusion Models + +The video diffusion model produces a visually-plausible and photo-realistic video based on text conditioning [7, 12, 16, 18, 20, 36, 49, 58]. The main challenge in the video diffusion model is to maintain the temporal coherency between frames. The pioneer works for video diffusion is to add temporal blocks to the text-to-image model and to learn temporal coherency between video frames [16, 20, 49]. Another type of relevant studies use pretrained models to leverage the temporal block [6, 7, 18]. In particular, Guo et al. [18] propose a motion module as the temporal block. It is designed as a plug-and-play module. After joint training with the pre-trained weight of stable diffusion [46], the module enables any text-to-image model to generate the video. + +Existing typography has been mainly studied using static single letters. The similar work to kinetic typography is also single letters, with only variation of motion effect and almost no letter's movement exists. To initiate kinetic typography generation, we present our KineTy dataset and model that allows dynamic motion, effects and glyph deformation of multi-letters, which will be explained at the next section. + +# 3 Kinetic Typography Dataset + +We describe how to build the KineTy dataset. Unlike previous datasets that only cover single-lettered images [71], our KineTy dataset has not only visual effects on multiple letters, but also their animations. We first provide detailed process + +![](images/17bfe6dbb3cc3a454ad1d956c7f83b1ee090a34b8f06abfd6aa7426fe10f2c0d.jpg) +pink to white background. + +![](images/a8c6f10e5518deddc1b8517ab215d2f867a21cbedfb2b03e9924e7a26ca338df.jpg) +giving it a vibrant and dimensional appearance on a white background. + +![](images/299721f2b7e5c6502d2dd9d0195d331e114e3c8dc30a871d08d8a31d9f93ba22.jpg) +The word "LgLed" forms with a glossy, liquid effect, where each letter appears in sequence from left to right to gradually complete the word. +The word "ADI" prominently displayed in bold, 3D-styled letters set against an orange textured background. + +![](images/2f0ec059ee10fafcea3c2de8e886848c1f78f46e4b2b3da8d162dcc82e135faf.jpg) +The video illustrates the word "HHtR" materializing in layers, with each letter representing a layer and enhancing the previous ones to gradually construct the complete word. +The word "FLRNQVOIZ" with a stylized pink colored glitch effect on a dark background. + +Fig. 2: Examples of our Kinecty dataset. Our dataset provides high-quality kinetic typography video created by professional motion graphic designers, along with captions that describe the visual appearance and motion effects. To aid visualization, we provide three frames from each video clip. +![](images/20d6eadc6c3330e13c5e80df381d0202d712502f31ff79bb93691c2a5cccbe69.jpg) +The word "ADI" gets bigger. +The word "INSZG" in bold letters casting a long shadow on a grey background. +The word "INSZG" is revealed in sequential order with each character appearing in order until the full word is displayed. + +![](images/252a0fd273ddb66c019680d90258fa14e633cb8bbecb2d066b18acb077fc91bf.jpg) +The word "FRLNQVOZ" starts from a scattered formation and converging to form the word clearly. +A colorful, distorted graphic with the word "CmfyTSr" in stylized lettering on a dark background. +The video depicts a fluid-like stream that gradually forms the word "CmfyTSR." The flow is slow and smooth in a smooth flowing sequence giving the impression of a liquid in motion. + +to render kinetic typography videos using templates made by professional editors in Sec. 3.1. Next, we introduce an way to describe their appearance and motions through text prompts in Sec. 3.2. Lastly, we explain how to make ground-truth kinetic typography video for strictly fair comparisons in Sec. 3.3. We summarize the difference between ours and existing datasets in Tab. 1. + +# 3.1 Video Rendering + +Employing templates for video rendering. The template is a pre-designed project file which contains a visual effect on letters. Many editors prefer to use the templates because it saves their time and labor cost in practice. Following the best practice, we utilize 584 kinetic typography templates from professional graphic designers for our dataset construction, as visualized in Fig. 2. + +Set of multiple letters. Next, we utilize the kinetic typography templates and randomly replace the text contents for augmentation. We employ multiple letters, in contrast to existing typography datasets with static single-letters [2,31,70]. Text contents are randomly generated by arranging up to 12 letters sampled from a set of 52 letters, including both uppercase and lowercase alphabets. Through this, we can expect rich letter-by-letter effects with various arrangements, while also keeping consistent styles across the multiple letters. + +Speculation. We render videos with $1,920 \times 1,080$ resolution for 3 sec., and 1,000 random words from 584 templates. Subsequently, all videos are downsampled to $512 \times 288$ resolution with 8fps for training. It takes a month to render the whole dataset with four i9 13900KF CPUs and an NVIDIA 2080ti GPU. + +![](images/36479d975f307eba64d998dee15e397e39683cd323c01c92b4271c547caa9ef8.jpg) +Fig. 3: Statistics of our proposed dataset. + +# 3.2 Video Captioning. + +Static and dynamic effect separation. In the next step, we caption the rendered kinetic typography for text-to-video generation. The professional editors usually design the static appearances at first, and then apply dynamic effects. Following this practice, we use two type of captions: static and dynamic captions, and labeled them accordingly. + +Static effect captioning. To label static captions, we focus on letter's appearance which contains spatial information of typography. As shown in Fig. 2, the static captions $\mathcal{C}_{\text {static }}$ describe the color of the letters and the background, the glyph of the letters (e.g., outlined with yellow color or bold font), the characteristics of the background (e.g., textured and shiny background) and the arrangement of the letters (e.g., in a diagonal way) based on the last frame of the video where all the letters are displayed. + +Dynamic effect captioning. Similarly, we concentrate on the temporal change of motions for each frame to write dynamic captions. Dynamic captions $\mathcal{C}_{dynamic}$ describe the motion part of the video, such as whether each letter appears in sequential or random order, can be rotated, or has a fade-in effect. For better systematic process, we initially label the videos with the GPT-4Vision model [41], and then manually verify and refine them to fix missing and wrongly labeled components. + +Statistics. The statistics of our dataset is summarized in Fig. 3. There are three main categories for the static appearance in Fig. 3 (a): Text type, text effect and background. Here, the text fill effects are dominant in the text type because it is the most obvious way to clearly show the text content. Especially, professional designers tend to prefer using solid colors, rather than text outlines. The reason why $57.7\%$ of our dataset has no text effect is the readability of text contents. Since kinetic typography makes dynamic motions of the contents, users do not need for fancy visual effects on it. + +In addition, we categorize dynamic motions into two-fold in Fig. 3 (b): motion type, including entrance, emphasis and movement, and sequence type: Our dataset supports various effect of entrance and emphasis to deliver striking messages. In contrast, the line has the majority in the movement because users' intention is usually conveyed after all letters in a scene are arranged in a line. In the same vein, when words come inside all at one and appear one-by-one in sequence, it is readable. They thus have the high portion in the sequence category. + +![](images/332008dee99c2e7634aad4e753ccc52e7fb96b6e40cada6f1eb788ca418087de.jpg) +Fig. 4: An architecture of our KineTy model. + +# 3.3 Ground-truth Video Generation + +Unlike generating random synthetic words in a training phase, we use real-world letters for evaluation by leveraging the templates. We can render ground-truth videos corresponding to user-input captions. + +For evaluation, individual letters are selected for each alphabet letter from A to Z, and the first letters are capitalized for checking case sensitivity (e.g., Apple, Ball, ..., Zebra). These 26 words are used for rendering 584 templates, so that a total of 15,184 videos are used for the evaluation. + +# 4 Kinetic Typography Diffusion Model + +In this section, we present our model for creating kinetic typography videos based on textual prompts. The key concept is to show animate visual effects on texts, and our challenges in this work are summarized as: (1) the effect should be visually-pleasing and eye-catching to viewers; (2) the transition between frames should be smooth and align with the captions; (3) the text must be readable. + +We begin by defining a kinetic typography generation problem in Sec. 4.1. We then describe how to model static and dynamic captions that efficiently guide appearance and motions in Sec. 4.2. We lastly discuss how to improve the glyph legibility with respect to the model design and its learning strategy in Sec. 4.3. Implementation details are provided in Sec. 4.4. + +# 4.1 Preliminary + +Conditional Latent Diffusion Models. Diffusion models train the data distribution by progressively refining a noisy initial state $z_{M} \sim \mathcal{N}(0,1)$ into the target data representation $z_{0}$ for $M$ diffusion steps. Recent advancements, particularly in Latent Diffusion Models (LDMs) [46], enhance the efficiency by encoding an image $x$ into a compact latent representation $z_{0} = \mathcal{E}(x)$ using an encoder $\mathcal{E}$ , and is transformed back to the image $\tilde{x} = \mathcal{D}(z_{0})$ using a decoder $\mathcal{D}$ . These adding noise and its subsequent removal are done with U-Net-style denoising network $\epsilon_{\pi}$ . In addition, a condition $y$ is mapped to the hidden state $h$ of $\epsilon_{\pi}$ via attention: + +$$ +\operatorname {A t t n} _ {\theta} (h, y) = \operatorname {S o f t m a x} \left(\frac {Q _ {\theta} (h) K _ {\theta} (y) ^ {\top}}{\sqrt {d}}\right) \cdot V _ {\theta} (y) \tag {1} +$$ + +$$ +\mathrm {s . t .} Q _ {\theta} (h) = W _ {\theta , Q} \cdot h, K _ {\theta} (y) = W _ {\theta , K} \cdot y, V _ {\theta} (y) = W _ {\theta , V} \cdot y, +$$ + +where $W_{\theta, Q}$ , $W_{\theta, K}$ and $W_{\theta, V}$ denote the learnable parameters for queries, keys and values. $d$ is the number of dimensions of the keys. The objective function for the denoising network $\epsilon_{\pi}$ is formulated as: + +$$ +L _ {l d m} = \mathbb {E} _ {\mathcal {E} (x), y, \epsilon \sim \mathcal {N} (0, 1), m} \left[ \left\| \epsilon - \epsilon_ {\pi} \left(z _ {m}, m, y\right) \right\| _ {2} ^ {2} \right], \tag {2} +$$ + +where $\| \cdot \|_2$ is the $L_2$ distance and diffusion step $m \in \{1, \dots, M\}$ is uniformly sampled during training. + +Extending image diffusion models to video. Extending the capability of text-to-image diffusion models to video generation has been more feasible. By incorporating a temporal self-attention called motion module [18], it learns a temporal coherency between $T$ frames in a latent sequence $\mathbf{z}_0^{0:T}$ , which corresponds to the video sequence $\tilde{\mathbf{x}}^{0:T} = \mathcal{D}(\mathbf{z}_0^{0:T})$ . This allows diffusion models to generate smoothly changed sequential images over time, whose temporal self-attention can be defined as follows: + +$$ +\operatorname {A t t n} _ {\phi} \left(\mathbf {h} ^ {0: T}, \mathbf {h} ^ {0: T}\right) = \operatorname {S o f t m a x} \left(\frac {Q _ {\phi} \left(\mathbf {h} ^ {0 : T}\right) K _ {\phi} \left(\mathbf {h} ^ {0 : T}\right) ^ {\top}}{\sqrt {d}}\right) \cdot V _ {\phi} \left(\mathbf {h} ^ {0: T}\right). \tag {3} +$$ + +# 4.2 Spatial and Temporal Guidance + +Static caption incorporation. Motivated by the best practice of professional designers who handle appearance and motion effects separately, we divide the caption into static and dynamic elements. + +When existing text-to-image models learn the distribution of image data, they are conditioned on texts related to the image's appearance. In the same manner, our model is also guided by captions describing the appearance of the text contents and background in each frame of video. Using Eq. (1), we define a spatial attention block with a self-attention followed by a cross-attention with a static caption $\mathcal{C}_{\text{static}}$ as: + +$$ +\mathbf {h} ^ {\prime 0: T} = \left\{\operatorname {A t t n} _ {\vartheta} \left(\operatorname {A t t n} _ {\theta} \left(\mathbf {h} ^ {t}, \mathbf {h} ^ {t}\right), \tau \left(\mathcal {C} _ {\text {s t a t i c}}\right)\right) \right\} _ {t = 0} ^ {T - 1}, \tag {4} +$$ + +where $\tau (\cdot)$ is the CLIP text encoder [44]. + +Dynamic caption incorporation. A previous work [18] uses a motion module that only utilizes a self-attention between frames to learn the temporal consistency. On the other hand, in kinetic typography, it is essential to accurately display the dynamic motion effects of each letter in a video, following user's textual description. To do this, we extend Eq. (3) by adding the cross-attention with a dynamic caption as follows: + +$$ +\mathbf {h} ^ {\prime 0: T} = \operatorname {A t t n} _ {\psi} \big (\operatorname {A t t n} _ {\phi} (\mathbf {h} ^ {0: T}, \mathbf {h} ^ {0: T}), \tau \big (\mathcal {C} _ {\text {d y n a m i c}} \big) \big), \tag {5} +$$ + +Through this process, the model becomes capable of maintaining temporal consistency as well as direct control over dynamic movements. + +# 4.3 Enhancing Glyph Legibility. + +Readability improvement. As a next step, we aim to make the legibility of our model better. To do this, we first classify descriptions for text contents and prompt because diffusion models often have a difficulty in distinguishing them. In this work, we put down a delimiter between each character for text contents. + +Let a text content $\mathcal{L} = \{l_1, l_2, \dots, l_L\}$ contain $L$ letters. To condition each letter separately, we join each letter $l_i$ in the text content with a delimiter symbol $|\cdot\rangle$ . Further, we add a symbol '\~' behind upper case letters because the clip encoder does not care of upper cases. To the end, when the word $\mathcal{L}$ denotes "Apple", $\mathcal{L}'$ is represented as "A^{|p|p|l|e}". We denote the whole $\mathcal{L}$ as $\mathcal{L}'$ for better representation. + +Attention on text contents. It can be challenging for the attention module to guide both text contents and effects at once, while disentangling the text content information in CLIP feature space. What is worse, the text content can be overwhelmed by relatively long appearance and motion effects in CLIP text encoder. To tackle this problem, we introduce an additional cross-attention branch to the text content. Inspired by ControlNet [76], we regard the text content as conditions through the zero convolution operation. Starting with the definition of word caption $\mathcal{C}_{word}$ via a prompt template like "The word $\{\mathcal{L}'\}$ ", we extend Eq. (4) by adding the cross-attention module between the word caption and hidden feature, weighted by the zero-initialized convolutions $\rho$ as follows: + +$$ +\begin{array}{l} \mathbf {h} _ {\text {s t a t i c}} ^ {\prime \prime 0: T} = \left\{\operatorname {A t t n} _ {\vartheta} \left(\mathbf {h} ^ {\prime}, \tau \left(\mathcal {C} _ {\text {s t a t i c}}\right)\right) + \rho_ {v} \left(\operatorname {A t t n} _ {v} \left(\mathbf {h} ^ {\prime}, \tau \left(\mathcal {C} _ {\text {w o r d}}\right)\right)\right) \right\} _ {t = 0} ^ {T - 1}, \tag {6} \\ w h e r e \quad \mathbf {h} ^ {\prime} = \operatorname {A t t n} _ {\theta} \left(\mathbf {h} ^ {t}, \mathbf {h} ^ {t}\right). \\ \end{array} +$$ + +This allows the network to gradually evaluate the usefulness of this additional caption based on the condition. In the same way, we incorporate the word caption into the temporal cross-attention in Eq. (5) as: + +$$ +\begin{array}{l} \mathbf {h} _ {\text {d y n a m i c}} ^ {\prime \prime 0: T} = \operatorname {A t t n} _ {\psi} \left(\mathbf {h} ^ {\prime}, \tau \left(\mathcal {C} _ {\text {d y n a m i c}}\right)\right) + \rho_ {\varphi} \left(\operatorname {A t t n} _ {\varphi} \left(\mathbf {h} ^ {\prime}, \tau \left(\mathcal {C} _ {\text {w o r d}}\right)\right)\right), \tag {7} \\ w h e r e \quad \mathbf {h} ^ {\prime} = \operatorname {A t t n} _ {\phi} \left(\mathbf {h} ^ {0: T}, \mathbf {h} ^ {0: T}\right). \\ \end{array} +$$ + +Glyph loss. To train our model, we use a common loss function $L_{ldm}$ . Additionally, we impose an extra penalty on the letter regions to enforce a sharp and correct glyph of text contents. We first use a binary mask $B$ for the letters in the last frame $V^T$ from a text segmentation model [68]. The mask is then blurred to cover its surrounding effects. We then define a glyph loss based on $L_{LDM}$ with the additional pixel-wise weighting strategy using the blurred mask as follows: + +$$ +L _ {g l y p h} = \mathbb {E} _ {\mathcal {E} (x), y, \epsilon \sim \mathcal {N} (0, 1), m} \left[ \left\| B \odot \left(\epsilon - \epsilon_ {\pi} \left(z _ {m} ^ {T}, m, y\right)\right) \right\| _ {2} ^ {2} \right], \tag {8} +$$ + +where $\odot$ is the element-wise multiplication. + +With the glyph loss, we enhance the legibility of text contents, enabling the precise creation of multi-letter formations. Through the linear combination of both loss terms, we can formulate the final loss function as $L = L_{ldm} + \alpha L_{glyph}$ . Here, we empirically set $\alpha$ to 0.01. + +# 4.4 Implementation Details + +To train our model, we use a two-step training strategy similar to that of human trainees [14]. First, we pre-train the network to generate spatial appearances using only static caption and last video frame pairs. Here, we detach the temporal attention modules to make them work as text-to-image diffusion models during 30 epochs with a batch size of 200. Next, we train the full model while freezing the spatial attention using whole video frames and captions for an epoch with a batch size of 8. For training, we resize the height $H$ and width $W$ of the video into $256 \times 256$ with $T = 24$ frames. The training is performed with AdamW optimizer [35], with a diffusion step $M$ of 1000 and a learning rate of 0.0001, which usually takes about 20 hours for training on 8 NVIDIA A100 GPUs. The inference time is about 20 seconds when the number of sampling steps is 25. + +# 5 Experiments + +In this section, we conduct comprehensive experiments to verify the effectiveness of our model for kinetic typography. We first describe the experimental setup in Sec. 5.1. We then provide comparison results with relevant typography and video generation models, and report user-study to highlight the practical applicability of our model in Sec. 5.2. We lastly carry out an extensive ablation study to validate the effect of each component in our model in Sec. 5.3 + +# 5.1 Experimental Setup + +Comparison methods. We compare ours with state-of-the-art generative models, which consist of 3 two-stage methods and 2 one-stage methods. The two-stage methods are based on combinations of text-to-image models, including DS-Fusion [52], GlyphControl [74] and Text-Diffuser [11], and an image-to-video model, SparseCtrl [17]. DS-Fusion outputs a stylized letter based on the given style phrase and a letter. Since our caption is a sentence, we extract a style keyword from it and consider it as a style phrase. Note that DS-Fusion is able to generate one letter at a time, so we concatenate each letter to make a full-text content. GlyphControl takes an instruction to generate a glyph image for text contents, then uses the image as a condition for the final output image. Text-Diffuser first finds a proper layout for words and produce the output based on a caption and the layout. Since the comparison methods yield images, we combine them with an image-to-video model, SparseCtrl, to compare with ours. SparseCtrl uses a caption and a guidance such as sketches, depth maps and images to generate a stylized video. + +On the other hand, one-stage-methods,AnimateDiff [18] and Lavie [58], are based on text-to-video models.AnimateDiff extends pre-trained text-to-image diffusion models with a motion module, and trains only the motion module to make fully use of the massive information of video dataset. Lavie leverages a temporal self-attention block to enhance a temporal consistency between frames while keeping the generation performance. + +Table 2: Quantitative evaluation of two-stage generation methods, and one-stage text-to-video generation methods. Bold: Best, Underline: Second-best. + +
ModelEntranceEmphasisMotionAverage
FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑
DS-F[52]+Sparse[17]1636.35.710.690.583184.95.970.730.581236.55.420.760.561843.105.660.700.58
Glyph[74]+Sparse[17]2415.64.300.760.682667.15.060.760.751468.64.480.840.802361.984.460.760.69
T-Diff[11]+Sparse[17]2937.25.610.780.903107.55.160.790.891208.24.060.820.952815.65.350.780.90
AnimateDiff[18]1613.84.710.590.012106.04.560.660.032160.64.510.670.071727.554.680.610.01
Lavie [58]825.12.250.820.541835.43.830.830.41505.443.330.850.35956.972.680.820.49
Ours147.41.790.870.71177.521.750.880.5436.161.620.940.98125.901.770.880.76
+ +![](images/71782326842b53f99dee087045070277f47a92419be858169d0b1c441cd3bdb2.jpg) + +![](images/6641917464a0de8178c6ea8044efebe2e111f39e81a4eaef998a2f5518fb1b54.jpg) + +![](images/05fce8642548fec43bc3020bec1932a97a35155db0895a0f08126a8813a9375e.jpg) + +![](images/430afc07bae35518711bf141055c3fcf1dab309f032232f8fe2a76cfeeb55957.jpg) + +![](images/52669d860eb45a89f7fb1ca1a130deaf098cd2d654a4e6e09de00d4184b55b1a.jpg) + +Static caption: The word 'Dog' in bold, green-stripped letters against a textured coral background. Dynamic caption: The sequential addition of letters one by one to spell out the word 'Dog'. + +![](images/725d9f0177562488b62051c1829de0480da90cf8738b5a6eab434fba78ed85d1.jpg) +Fig. 5: Qualitative results from the comparison models and ours. Obviously, the results from ours reflect the contents of captions better than the others. + +![](images/18c86c52a02390aa315b9aeff28d706f1e9c14eee3ad5986533c4856cabeda9f.jpg) + +![](images/f89da7733b5becbc819a3b1246035ed2db8de3a689ef745483c5cc71bd0ee9cd.jpg) + +![](images/0feb4f8301ead937bbda8bae32405b40814e4149d36d3d653dff3f1825692a69.jpg) + +![](images/c71bf862a28e003657c5b309f564ac140b75dc6dc4534a6c9d97ab038373c49e.jpg) + +Static caption: The word Mermaid in a blurred, tunnel-like visual effect, creating a sense of depth and distortion on a dark background. Dynamic caption: The word Mermaid gradually comes into focus, with each letter appearing in a random order, intensifying the sense of depth and creating a captivating rev + +![](images/db57bd5ce63cf6aac5ac85486c3296071025225634e4224bb07fd10af688e203.jpg) + +![](images/dc63aa8892f1dc22fe0f3adc810c83059f39d91ff9b920b0ca92cb85dd29d5e5.jpg) + +![](images/5b8f5443d35da0d2affedee1bfb4b78f430e4ea1e40c6979a63b11ba075fe648.jpg) + +![](images/2f89c99544f55493f992466f0075e9762afa9f6ff646f601aa6c689eb0f3e9b6.jpg) + +![](images/8b83ac5a8b4f61f82533292a21623fca2fc3fd4adddc2dd716ba3dd2718460b9.jpg) + +Static caption: The word 'Watermelon' in white letters against a black background, accompanied by two smiling emojis faces with heart eyes. Dynamic caption: Each character of the word 'Watermelon' appears sequentially, with various smiling emojis popping up randomly around the text. + +(a) DS-Fusion + ++ SparseCtrl + +(b) GlyphControl + ++SparseCtrl + +(c) Text-Diffuser + ++ SparseCtrl + +(d) Lavie + +(e) Ours + +For a fair comparison, we utilize the official source codes and pre-trained weights provided by the authors. For more analysis, we categorize the motion effects into three groups, 'Entrance', 'Emphasis' and 'Motion', and conduct experiments to check the quality of each result. + +Evaluation metrics. Since the final products of kinetic typography depend on the designer's preferences, we adopt four feature-based metrics rather than photoconsistency measures like PSNR. (1) Fréchet Video Distance (FVD) qualifies a similarity between two videos by comparing their feature representations extracted from a pre-trained deep neural network [55]. (2) Inception Score (IS) evaluates the quality and diversity of images generated by a model [4]. (3) CLIPScore measures the semantic alignment between generated images and their textual descriptions, using the CLIP model [19, 22, 44]. (4) Optical Character Recognition (OCR) checks the clarity of the output by comparing recognized text to a reference text using the F1-score. + +# 5.2 Evaluation Results + +Quantitative results. As shown in Tab. 2, our Kinety model achieves the better performance than the comparison methods in the most metrics on all the + +![](images/ab40cc94a030dfde61128a6056884fbfeda639dd999542f473ce3bba82c57d86.jpg) + +![](images/6f66d48c83b9388c2cda852f8b1135f976d91323de4cf8d1d1c6b4e07734c7dd.jpg) +Fig. 6: Visualization of attention maps between letters from a caption and feature map of a hidden layer. + +![](images/263d4876eca264f3cbaea87f31cfa9e74ae2c6377af224a23698643ecb7adefd.jpg) + +![](images/497865e5ac049905fc5ac22ec87f42c9aaaebb5bf1a18f75d6d3c96755934e68.jpg) + +![](images/8f775c883874303c4be34e3807bd5ea41820861fd1e14a10226ea1db42c59ec8.jpg) + +![](images/b2880aa9c7ff5380ededae83965f46f4a08ce62526fda4eaf87269b6f408007d.jpg) + +![](images/8f2abb81af309ba4df511cd93a97736f0009516dd2c0e7f3c7e300de2133def8.jpg) + +![](images/b8e0bb6acd9f9357508616b74839a754397ff3de6dd32700c6574a55fea150b1.jpg) + +![](images/057b2b05a8f49d404d77940ebbb804c48ec1a767d0b87df4f88dd7728e67512d.jpg) +The word + +![](images/326f57c1f7443e98bae158cb2562753b2a1fc71e21c364f1ede53cad23911da9.jpg) + +![](images/6758910690e47a18537e1f9c04724d3fa496f988a1624c0651a1b40518a5119d.jpg) + +![](images/f23cfb53c5f2760e5045299d91958b0037d79dcf314a890586d68f66c0e17808.jpg) + +![](images/ad572e26d114e4a06a400706820b16b29fa619ccfa96641c7d224ed341c92722.jpg) + +![](images/46d6dbec202424b4b31cdbefa4e66af86378bb54c115cae1c572ac71fc87112a.jpg) + +![](images/bb8d3ea1b9ffe45b07f06f3e881003ba83a74be905dec44a8d3bd77027f7dc04.jpg) +(a) Initial kinetic typography video +(b) Changing text color + +![](images/d317fddc2596f93d21cdf4f1de59c186de4ac3c8f697a5eaae18e2797c135a45.jpg) +(c) Changing text font +(d) Changing background + +![](images/342fba2ed5aa57c5035b6b2205b0eb8f0ef20438f84fe7c3a4580c3a4eeae615.jpg) +(e) Changing motion effect +(f) Final product +Fig. 7: The demonstration of the editable capability of our Kinecty model after generating initial kinetic typography. + +categories. The comparison methods, despite their impressive performance in general image and video generation, show unsatisfactory performance in creating kinetic typography. The reasons are mainly two-folds: (1) Text-to-video models are not specialized for kinetic typography and often fail to exactly distinguish the text content in the long captions. Different from them, we additionally feed $\mathcal{C}_{word}$ which imposes a high attention to the text content in the video. (2) They have challenges in creating letter-level motions. Since they are unaware of each letter in the text content, all the letters move as a whole or inconsistently. We handle this issue by providing a delimiter between each letter and using zero-convolution in Fig. 6. + +Our model shows promising results in generating legible text content, but it sometimes falls short in OCR accuracy. This is because the two-stage methods algorithmically generate a glyph image of a text content, and then deform glyphs conditioned on the glyph image. As a result, they have the better performance in OCR metrics than ours and one-stage methods. Nevertheless, our method outperforms in the most cases, demonstrating the effectiveness of our method in generating the clear text content and motion effects. + +Qualitative results. In Fig. 5, we display three examples of the comparison methods. Compared to ours, the two-stage methods do not understand the user's + +Table 3: The result of user-study. The score ranges from $1 \sim 6$ for Study 1 & 2, and 0 $\sim 1$ for Study 3. Bold: Best, Underline: Second-best. + +
User StudyDynamic EffectGeneralExpert
DS-F +SparseGlyph +SparseT-Diff +SparseAnimate DiffLavieOursDS-F +SparseGlyph +SparseT-Diff +SparseAnimate DiffLavieOurs
Study 1Entrance3.3843.6873.8912.4063.1534.4783.4814.0254.1311.0193.0815.263
Emphasis3.5623.7134.1192.5632.7254.3193.6254.1754.2251.0502.6255.300
Motion2.3634.1384.1133.3382.3754.6751.7003.5504.0004.0002.0005.750
Average3.3183.7133.9232.5183.0364.4943.3353.9954.0901.3252.9055.350
Study 2Entrance3.5143.7753.9412.4133.2704.0883.6504.0814.2441.0193.1064.900
Emphasis3.5313.8814.0752.2693.0504.1943.7254.1504.0501.0252.6755.375
Motion2.8383.9503.8753.2502.8504.2382.1504.0004.4503.4501.9505.000
Average3.4393.7993.9482.4603.2204.1353.4854.0604.2551.2652.9704.995
Study 3Entrance0.6280.6200.6520.0630.1690.7250.8440.8560.8870.0190.1380.969
Emphasis0.5810.6060.6380.0880.1060.7810.7500.7000.9000.0000.0001.000
Motion0.0750.7130.7500.3620.2880.7810.0000.9000.9000.5000.2501.000
Average0.5530.6260.6580.0960.1740.7440.7300.8350.8800.0650.1350.975
+ +instruction with respect to either the whole caption or the letter's movement. Here, we need to know how to perform our Kinecty model. Figure 6 visualizes the attention map between the caption and features from the hidden state on the corresponding images. Thanks to the cross-attention between the word caption and the noisy latent, and the delimiter to separate the words, our model successfully embeds features for each letter in the noise. + +Lastly, to show the generality of our KineTy model, we conduct an additional experiment. After generating an initial kinetic typography, our KineTy model allows users to modify its style and motions such as the content color, background, font and motion, and even to combine them. As demonstrated in Fig. 7, ours can render the modified outcomes corresponding to the additional captions. + +User Study. To assess the practical utility of our results in the field of kinetic typography, we conduct a user study using Amazon MTurk. We ask 20 questionnaires to 50 participants. Since their familiarity of typography can vary significantly, we divide them into two groups: experts and non-experts, consisting of 10 and 40 individuals, respectively. The study involves three sections: (1) caption alignment; (2) kinetic typography suitability; (3) Word readability. We make two video clips by randomly choosing two words for each effect. In total, we make 20 clips for this user study. + +In the first study, we measure how well ours and the comparison methods generate videos that align with provided captions. The participants see 6 videos from different models with the same corresponding caption in a random order. They are asked to rank the videos from 1 to 6, and the rankings are subsequently converted into scores, ranging from 6 (Best alignment) to 1 (Worst alignment). The following study evaluates how proper the generated videos are for motion graphic applications. Since some participants might be unfamiliar with kinetic typography, all participants watch 4 example videos from online in advance. After that, they check six videos in a random order and rank them. Scores were assigned based on these rankings, from 6 (the most proper) to 1 (the least proper). The last study assesses how readable the outcomes are. After watching 6 videos, the + +Table 4: The result of ablation study. Bold: Best, Underline: Second-best. + +
ModelEntranceEmphasisMotionAverage
FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑
Ours147.41.790.870.71177.521.750.880.5436.161.620.940.98125.901.770.880.76
-Effect separation789.82.750.740.49880.12.880.750.49607.12.540.820.23729.82.730.750.48
-Cword239.11.770.860.70247.11.730.870.5462.71.600.950.95201.81.740.870.74
-Lglph188.91.770.860.70226.81.750.880.5550.11.590.930.98160.81.750.880.75
+ +participants are needed to vote for videos with readable text contents. Of course, multiple voting is available. + +The performance of each method is measured based on scores for the first two studies, and on the selection ratio for third one. As demonstrated in Tab. 3, our approach shows the promising performance across all studies. The results highlight our model's practical usability in creating kinetic typography, especially for the domain experts who give our model the highest scores. + +# 5.3 Ablation Study + +Without caption separation. First of all, we train our network without dividing static and dynamic caption. The performances drop significantly for every metrics in Tab. 4. Since we train the text-to-image backbone using only static captions, dynamic captions are considered as noisy input. Similarly, the motion module, used to train dynamic motions, is effective when the simple motion guidance is given, compared to using static and dynamic caption together. + +Text contents incorporation. We observe that text-to-image diffusion models often fail to catch up on text contents in the prompt when a long description is given. We thus use $\mathcal{C}_{word}$ as an additional input to give a strong attention to each letter. As shown in Tab. 4, without $\mathcal{C}_{word}$ , the performance degradation is observed as expected. + +Without glyph loss. Lastly, we evaluate the performance without $L_{glyph}$ . The inferior FVD scores come from the blur effect at the edge of the word along with color bleeding effects. + +# 6 Conclusion + +In this paper, we propose a generative kinetic typography model, named KineTy. To achieve this, we first build a large-scale kinetic typography dataset by collecting 584 templates from professional designers, and use them to train our diffusion-based network. To effectively handle the text prompt, we split it into static and dynamic captions. They are used to directly guide the spatial and temporal cross-attention for better appearance and motion effects, respectively. Our glyph loss also strengthens the legible and artistic letter generation. We lastly demonstrate that our KineTy produces the visually compelling and semantically coherent kinetic typography videos. The extensive user studies further validate the effectiveness of our model with respect to its practical utility. + +Acknowledgements This research was supported by 'Project for Science and Technology Opens the Future of the Region' program through the INNOPOLIS FOUNDATION funded by Ministry of Science and ICT (Project Number: 2022-DD-UP-0312), the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(RS-2024-00338439), GIST-MIT Research Collaboration grant funded by the GIST in 2024, and the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-01842, Artificial Intelligence Graduate School Program (GIST)). + +# References + +1. Anderson, D., Shamir, A., Fried, O.: Neural font rendering. arXiv preprint arXiv:2211.14802 (2022) +2. Azadi, S., Fisher, M., Kim, V.G., Wang, Z., Shechtman, E., Darrell, T.: Multicontent gan for few-shot font style transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018) +3. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021) +4. Barratt, S., Sharma, R.: A note on the inception score. arXiv preprint arXiv:1801.01973 (2018) +5. Berio, D., Leymarie, F.F., Asente, P., Echevarria, J.: Strokestyles: Stroke-based segmentation and stylization of fonts. ACM Transactions on Graphics (TOG) (2022) +6. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023) +7. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +8. Campbell, N.D., Kautz, J.: Learning a manifold of fonts. ACM Transactions on Graphics (TOG) (2014) +9. Chen, C.H., Liu, Y.T., Zhang, Z., Guo, Y.C., Zhang, S.H.: Joint implicit neural representation for high-fidelity and compact vector fonts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) +10. Chen, J., Huang, Y., Lv, T., Cui, L., Chen, Q., Wei, F.: Textdiffuser-2: Unleashing the power of language models for text rendering. arXiv preprint arXiv:2311.16465 (2023) +11. Chen, J., Huang, Y., Lv, T., Cui, L., Chen, Q., Wei, F.: Textdiffuser: Diffusion models as text painters. Proceedings of the Neural Information Processing Systems (NeurIPS) (2023) +12. Esser, P., Chiu, J., Atighehchian, P., Granskog, J., Germanidis, A.: Structure and content-guided video synthesis with diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) +13. Ford, S., Forlizzi, J., Ishizaki, S.: Kinetic typography: issues in time-based presentation of text. In: CHI'97 extended abstracts on Human factors in computing systems, pp. 269-270. ACM Digital Library (1997) + +14. Fridsma, L., Gyncild, B.: Adobe After Effects CC Classroom in a Book. Adobe Press (2019) +15. Fu, B., He, J., Wang, J., Qiao, Y.: Neural transformation fields for arbitrary-styled font generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +16. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) +17. Guo, Y., Yang, C., Rao, A., Agrawala, M., Lin, D., Dai, B.: Sparsectrl: Adding sparse controls to text-to-video diffusion models. arXiv preprint arXiv:2311.16933 (2023) +18. Guo, Y., Yang, C., Rao, A., Wang, Y., Qiao, Y., Lin, D., Dai, B.: AnimateDiff: Animate your personalized text-to-image diffusion models without specific tuning arXiv preprint arXiv:2307.04725 (2023) +19. Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (2021) +20. Ho, J., Chan, W., Saharia, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022) +21. Hu, Y., Luo, C., Chen, Z.: Make it move: controllable image-to-video generation with text descriptions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +22. Huang, Y., Xue, H., Liu, B., Lu, Y.: Unifying multimodal transformer for bidirectional image and text generation. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 1138-1147 (2021) +23. Iluz, S., Vinker, Y., Hertz, A., Berio, D., Cohen-Or, D., Shamir, A.: Word-as-image for semantic typography. ACM Transactions on Graphics (TOG) (2023) +24. Jahanian, A., Liu, J., Lin, Q., Tretter, D., O'Brien-Strain, E., Lee, S.C., Lyons, N., Allebach, J.: Recommendation system for automatic design of magazine covers In: Proceedings of the 2013 International Conference on Intelligent User Interfaces (IUI) (2013) +25. Jia, P., Li, C., Liu, Z., Shen, Y., Chen, X., Yuan, Y., Zheng, Y., Chen, D., Li, J., Xie, X., et al.: Cole: A hierarchical generation framework for graphic design. arXiv preprint arXiv:2311.16974 (2023) +26. Kato, J., Nakano, T., Goto, M.: Textalive: Integrated design environment for kinetic typography. In: Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems (CHI) (2015) +27. Krishnan, P., Kovvuri, R., Pang, G., Vassilev, B., Hassner, T.: Textstylebrush: Transfer of text aesthetics from a single example. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2023) +28. Lee, J.C., Forlizzi, J., Hudson, S.E.: The kinetic typography engine: an extensible system for animating expressive text. In: Proceedings of the 15th annual ACM symposium on User Interface Software and Technology (UIST) (2002) +29. Lee, J., Jun, S., Forlizzi, J., Hudson, S.E.: Using kinetic typography to convey emotion in text-based interpersonal communication. In: Proceedings of the 6th conference on Designing Interactive systems (DIS) (2006) +30. Li, C., Taniguchi, Y., Lu, M., Konomi, S.: Few-shot font style transfer between different languages. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2021) + +31. Li, X., Wu, L., Wang, C., Meng, L., Meng, X.: Compositional zero-shot artistic font synthesis. In: Proceedings of the International Joint Conferences on Artificial Intelligence (IJCAI) (2023) +32. Liu, X., Meng, G., Chang, J., Hu, R., Xiang, S., Pan, C.: Decoupled representation learning for character glyph synthesis. IEEE Transactions on Multimedia (TMM) (2021) +33. Liu, Y.T., Guo, Y.C., Li, Y.X., Wang, C., Zhang, S.H.: Learning implicit glyph shape representation. IEEE Transactions on Visualization and Computer Graphics (TVCG) (2022) +34. Liu, Y.T., Zhang, Z., Guo, Y.C., Fisher, M., Wang, Z., Zhang, S.H.: Dualvector: Unsupervised vector font synthesis with dual-part representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +35. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: Proceedings of the International Conference on Learning Representations (ICLR) (2018) +36. Luo, Z., Chen, D., Zhang, Y., Huang, Y., Wang, L., Shen, Y., Zhao, D., Zhou, J., Tan, T.: Videofusion: Decomposed diffusion models for high-quality video generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +37. Men, Y., Lian, Z., Tang, Y., Xiao, J.: Dyntypo: Example-based dynamic text effects transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019) +38. Minakuchi, M., Tanaka, K.: Automatic kinetic typography composer. In: Proceedings of the ACM SIGCHI International Conference on Advances in Computer Entertainment Technology (ACE) (2005) +39. Mu, X., Chen, L., Chen, B., Gu, S., Bao, J., Chen, D., Li, J., Yuan, Y.: Fontstudio: Shape-adaptive diffusion model for coherent and consistent font effect generation. arXiv preprint arXiv:2406.08392 (2024) +40. Nagata, Y., Iwana, B.K., Uchida, S.: Contour completion by transformers and its application to vector font data. arXiv preprint arXiv:2304.13988 (2023) +41. OpenAI: GPT-4V(ision) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf (2023) +42. Pan, W., Zhu, A., Zhou, X., Iwana, B.K., Li, S.: Few shot font generation via transferring similarity guided global style and quantization local style. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) +43. Qu, Y., Tan, Q., Xie, H., Xu, J., Wang, Y., Zhang, Y.: Exploring stroke-level modifications for scene text editing. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) (2023) +44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proceedings of the International Conference on Machine Learning (ICML) (2021) +45. Reddy, P., Zhang, Z., Wang, Z., Fisher, M., Jin, H., Mitra, N.: A multi-implicit neural representation for fonts. Proceedings of the Neural Information Processing Systems (NeurIPS) (2021) +46. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +47. Shimoda, W., Haraguchi, D., Uchida, S., Yamaguchi, K.: De-rendering stylized texts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021) + +48. Shimoda, W., Haraguchi, D., Uchida, S., Yamaguchi, K.: Towards diverse and consistent typography generation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2024) +49. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. In: Proceedings of the International Conference on Learning Representations (ICLR) (2023) +50. Smith, J., Team, A.C.: Adobe After Effects CS6 Digital Classroom. John Wiley & Sons (2012) +51. Subramanian, J., Chordia, V., Bart, E., Fang, S., Guan, K., Bala, R., et al.: Strive: Scene text replacement in videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021) +52. Tanveer, M., Wang, Y., Mahdavi-Amiri, A., Zhang, H.: Ds-fusion: Artistic typography via discriminated and stylized diffusion. arXiv preprint arXiv:2303.09604 (2023) +53. Thamizharasan, V., Liu, D., Agarwal, S., Fisher, M., Gharbi, M., Wang, O., Jacobson, A., Kalogerakis, E.: Vecfusion: Vector font generation with diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) +54. Tuo, Y., Xiang, W., He, J.Y., Geng, Y., Xie, X.: Anytext: Multilingual visual text generation and editing. In: Proceedings of the International Conference on Learning Representations (ICLR) (2024) +55. Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717 (2018) +56. Wang, C., Wu, L., Liu, X., Li, X., Meng, L., Meng, X.: Anything to glyph: Artistic font synthesis via text-to-image diffusion model. In: SIGGRAPH Asia 2023 Conference Papers (2023) +57. Wang, C., Zhou, M., Ge, T., Jiang, Y., Bao, H., Xu, W.: Cf-font: Content fusion for few-shot font generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +58. Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., Yang, P., et al.: Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103 (2023) +59. Wang, Y., Ye, Y., Mao, Y., Yu, Y., Song, Y.: Self-supervised scene text segmentation with object-centric layered representations augmented by text regions. In: Proceedings of the 30th ACM International Conference on Multimedia (2022) +60. Wang, Y., Lian, Z.: Deepvecfont: Synthesizing high-quality vector fonts via dual-modality learning. ACM Transactions on Graphics (TOG) (2021) +61. Wang, Y., Pu, G., Luo, W., Wang, Y., Xiong, P., Kang, H., Lian, Z.: Aesthetic text logo synthesis via content-aware layout inferring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +62. Wang, Y., Wang, Y., Yu, L., Zhu, Y., Lian, Z.: Deepvecfont-v2: Exploiting transformers to synthesize vector fonts with higher quality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +63. Wong, Y.Y.: Temporal typography: a proposal to enrich written expression. In: Proceedings of the Conference Companion on Human Factors in Computing Systems (CHI) (1996) +64. Xia, Z., Xiong, B., Lian, Z.: Vecfontsf: Learning to reconstruct and synthesize high-quality vector fonts via signed distance functions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) + +65. Xie, L., Shu, X., Su, J.C., Wang, Y., Chen, S., Qu, H.: Creating emordle: Animating word cloud for emotion expression. IEEE Transactions on Visualization and Computer Graphics (TVCG) (2023) +66. Xie, L., Zhou, Z., Yu, K., Wang, Y., Qu, H., Chen, S.: Wakey-wakey: animate text by mimicking characters in a gif. In: Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (2023) +67. Xu, C., Zhou, M., Ge, T., Jiang, Y., Xu, W.: Unsupervised domain adaption with pixel-level discriminator for image-aware layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +68. Xu, X., Zhang, Z., Wang, Z., Price, B., Wang, Z., Shi, H.: Rethinking text segmentation: A novel dataset and a text-specific refinement approach. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +69. Yang, Q., Huang, J., Lin, W.: Swaptext: Image based texts transfer in scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +70. Yang, S., Liu, J., Lian, Z., Guo, Z.: Awesome typography: Statistics-based text effects transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2017) +71. Yang, S., Wang, W., Liu, J.: Te141k: artistic text benchmark for text effect transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2020) +72. Yang, S., Wang, Z., Liu, J.: Shape-matching $\mathrm{gan} + +$ : Scale controllable dynamic artistic text style transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2021) +73. Yang, S., Wang, Z., Wang, Z., Xu, N., Liu, J., Guo, Z.: Controllable artistic text style transfer via shape-matching gan. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2019) +74. Yang, Y., Gui, D., Yuan, Y., Liang, W., Ding, H., Hu, H., Chen, K.: Glyphcontrol: Glyph conditional control for visual text generation. In: Proceedings of the Neural Information Processing Systems (NeurIPS) (2023) +75. Yang, Z., Peng, D., Kong, Y., Zhang, Y., Yao, C., Jin, L.: Fontdiffuser: One-shot font generation via denoising diffusion with multi-scale content aggregation and style contrastive learning. arXiv preprint arXiv:2312.12142 (2023) +76. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) +77. Zhang, S., Ma, J., Wu, J., Ritchie, D., Agrawala, M.: Editing motion graphics video via motion vectorization and transformation. ACM Transactions on Graphics (TOG) (2023) \ No newline at end of file diff --git a/2024/Kinetic Typography Diffusion Model/images.zip b/2024/Kinetic Typography Diffusion Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8780698843adf103cca02a3b7eeac0805aff66e2 --- /dev/null +++ b/2024/Kinetic Typography Diffusion Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:735ae1ac884531cead0bbcff6bb946066e9ccb2cbae8c4a5f862c86a081cf8a0 +size 608168 diff --git a/2024/Kinetic Typography Diffusion Model/layout.json b/2024/Kinetic Typography Diffusion Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8481161a1c8d02105f3857ea51a9fc6f9d991ad6 --- /dev/null +++ b/2024/Kinetic Typography Diffusion Model/layout.json @@ -0,0 +1,12188 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 178, + 112, + 435, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 112, + 435, + 129 + ], + "spans": [ + { + "bbox": [ + 178, + 112, + 435, + 129 + ], + "type": "text", + "content": "Kinetic Typography Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 149, + 150, + 462, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 150, + 462, + 163 + ], + "spans": [ + { + "bbox": [ + 149, + 150, + 462, + 163 + ], + "type": "text", + "content": "Seonmi Park, Inhwan Bae, Seunghyun Shin, and Hae-Gon Jeon" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 135, + 172, + 476, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 172, + 476, + 205 + ], + "spans": [ + { + "bbox": [ + 135, + 172, + 476, + 205 + ], + "type": "text", + "content": "AI Graduate School, GIST, South Korea \n{bluesky1000, inhwanbae, seunghyuns98}@gm.gist.ac.kr, haegonj@gist.ac.kr \nhttps://seonmip.github.io/kinety" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 231, + 452, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 231, + 452, + 483 + ], + "spans": [ + { + "bbox": [ + 160, + 231, + 452, + 483 + ], + "type": "text", + "content": "Abstract. This paper introduces a method for realistic kinetic typography that generates user-preferred animatable \"text content\". We draw on recent advances in guided video diffusion models to achieve visually-pleasing text appearances. To do this, we first construct a kinetic typography dataset, comprising about 600K videos. Our dataset is made from a variety of combinations in 584 templates designed by professional motion graphics designers and involves changing each letter's position, glyph, and size (i.e., flying, glitches, chromatic aberration, reflecting effects, etc.). Next, we propose a video diffusion model for kinetic typography. For this, there are three requirements: aesthetic appearances, motion effects, and readable letters. This paper identifies the requirements. For this, we present static and dynamic captions used as spatial and temporal guidance of a video diffusion model, respectively. The static caption describes the overall appearance of the video, such as colors, texture and glyph which represent a shape of each letter. The dynamic caption accounts for the movements of letters and backgrounds. We add one more guidance with zero convolution to determine which text content should be visible in the video. We apply the zero convolution to the text content, and impose it on the diffusion model. Lastly, our glyph loss, only minimizing a difference between the predicted word and its ground-truth, is proposed to make the prediction letters readable. Experiments show that our model generates kinetic typography videos with legible and artistic letter motions based on text prompts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 494, + 452, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 494, + 452, + 504 + ], + "spans": [ + { + "bbox": [ + 160, + 494, + 452, + 504 + ], + "type": "text", + "content": "Keywords: Kinetic Typography " + }, + { + "bbox": [ + 160, + 494, + 452, + 504 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 494, + 452, + 504 + ], + "type": "text", + "content": " Video Diffusion " + }, + { + "bbox": [ + 160, + 494, + 452, + 504 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 494, + 452, + 504 + ], + "type": "text", + "content": " Visual Text Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 540, + 229, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 540, + 229, + 552 + ], + "spans": [ + { + "bbox": [ + 133, + 540, + 229, + 552 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 563, + 481, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 481, + 623 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 481, + 623 + ], + "type": "text", + "content": "Kinetic typography is an artistic motion graphics design combining text and animations [13]. Based on the word's meaning, motions are generated to convey information in videos. The main goal is eye-catching and to improve message retention. With the rise of video media, it has become an essential element in TV programs, commercials, music videos and film leaders." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 624, + 481, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 481, + 647 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 481, + 647 + ], + "type": "text", + "content": "Kinetic typography controls letters' shape (glyph), color and texture over time, and transforms their positions. Professional motion graphic designers" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 653, + 233, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 653, + 233, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 653, + 233, + 665 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 341, + 171 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 341, + 171 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 341, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 341, + 171 + ], + "type": "image", + "image_path": "1b7e4304a1c291bb2fa7195c891d6b6d3b7e5cbc489435d4b6e83f0f094e8022.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 173, + 272, + 182 + ], + "lines": [ + { + "bbox": [ + 198, + 173, + 272, + 182 + ], + "spans": [ + { + "bbox": [ + 198, + 173, + 272, + 182 + ], + "type": "text", + "content": "(a) Designer's workflow" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 352, + 114, + 482, + 173 + ], + "blocks": [ + { + "bbox": [ + 352, + 114, + 482, + 173 + ], + "lines": [ + { + "bbox": [ + 352, + 114, + 482, + 173 + ], + "spans": [ + { + "bbox": [ + 352, + 114, + 482, + 173 + ], + "type": "image", + "image_path": "8e0ab68d1e3e2732eb91503e1f6aaa3c140574b36324ae03bd1c67ae76ac8705.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 174, + 450, + 183 + ], + "lines": [ + { + "bbox": [ + 381, + 174, + 450, + 183 + ], + "spans": [ + { + "bbox": [ + 381, + 174, + 450, + 183 + ], + "type": "text", + "content": "(b) Our Kinety model" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 188, + 480, + 220 + ], + "lines": [ + { + "bbox": [ + 130, + 188, + 480, + 220 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 480, + 220 + ], + "type": "text", + "content": "Fig. 1: An overview of our KineTy pipeline, which is motivated by the designer's workflow. Our key idea is to generate eye-catching and aesthetic animatable words based on user instructions." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 246, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 480, + 365 + ], + "type": "text", + "content": "use commercial software like 'Adobe After Effects' to make and render kinetic typography videos. The conventional pipeline is as follows [14, 50]; (1) Define a text box in a workspace called composition and enter a text (i.e. word or message). (2) Edit the static text by tuning its font, color and texture (3) Build the background if needed (4) Apply motion effects for all text and background (5) Repeat, optimize, and fine-tune the whole process until it satisfies user intention. This process is time-consuming and labor-intensive. It takes three hours for simple motions to several days for sophisticated effects per single kinetic typography video [14, 26]. To efficiently do this, designers may need to consider either various design references or pre-defined design options called templates." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 366, + 481, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 481, + 460 + ], + "type": "text", + "content": "With the advancement of generative models, there have been attempts to produce dynamic typography. The literature on typography generation has mostly focused on creating static single-letter images. Starting from transforming the shape of the object into a single letter [23, 39, 52, 56], there have been works to create multiple letters and multiple words by introducing a concept of layout [10, 11, 74]. Furthermore, DynTypo [37] and Shape-Matching GAN++ [72] adopt style-transfer techniques to produce animatable effects on a single letter. They focus on only animatable effects without any actual motions of letters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 462, + 481, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 481, + 532 + ], + "type": "text", + "content": "Meanwhile, the recent advancement of the video generation model has been actively studied with the success of text-guided video diffusion models, especially user's description conditions video frames [7,16,18,20,49]. Although these models have opened up the possibility of creating kinetic typography, there is a critical issue that the video generations show a weak understanding of the letters' shapes and motions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 640 + ], + "type": "text", + "content": "In this paper, we propose the Kinetic Typography diffusion model (called KineTy model) that generates kinetic typography from user-provided text prompts. Inspired by the video diffusion model, we allow users to input comprehensive descriptions for color, font, size, position and motion effects of letters. To better represent text motions, we first introduce our kinetic typography video dataset. 600K videos are rendered by combining randomly generated text contents with 584 templates made by professional motion graphic designers. These videos are labeled with static and dynamic captions that describe the video with respect to its appearance and motion characteristics, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Next, we present KineTy model that effectively synthesizes videos from the text prompt. Here, we reevaluate how to effectively condition the caption guidance" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "S. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "into the video diffusion model at a fundamental level. We enforce that each effect is incorporated by separately inserting static and dynamic captions into the spatial attention and temporal attention, respectively. To strengthen these attentions, we apply a zero convolution [76] into the word caption, and add it to them. We use a mask loss term only for the generated video by masking out the background contents, which makes the text contents more readable." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 483, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 483, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 483, + 237 + ], + "type": "text", + "content": "Experimental results show that our Kinecty robustly generates kinetic typography video with multi-letter legibility, while accurately representing captions. Furthermore, extensive and meticulous user studies support our claim that Kinecty produces more aesthetic outcomes than general-purpose video generations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 255, + 237, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 237, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 237, + 268 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 281, + 282, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 282, + 294 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 282, + 294 + ], + "type": "text", + "content": "2.1 Typography Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 302, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 482, + 410 + ], + "type": "text", + "content": "The key of typography is to make readable and visuallyappealing text contents to readers. Most research in this field has focused on transferring the style of a single-letter design, including elements like color, glyph, font and effects, to other letters [1,2,5,8,9,15,30-34,40,42,45,53,57,60,62,64,70,75]. Additionally, there have been works on transferring an image style to the text contents [73]. Similarly, research on scene text editing, which changes text content while maintaining its own style in a scene, is also underway [27,43,54,69], whose extended version to video is available in [51]. In addition to changing text contents, color editing [47] and text segmentation [59,68] are proposed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 410, + 482, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 446 + ], + "type": "text", + "content": "With the recent success of the high-fidelity text-to-image diffusion model [46], typography generation has gained interest. Works in [23,52,56] apply image styles into the target letter in an unsupervised manner." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 446, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 529 + ], + "type": "text", + "content": "There are concurrent works that display multi-letters, beyond the single-letter generation. They follow the multi-step approach that firstly generates layouts at specific positions, arrange them with multi-letters or multi-words with the same fonts and colors, and contextually infers the background [10, 11, 24, 25, 48, 61, 67, 74, 77]. However, these models are not specialized for typography. Since all letters are generated together, it is not editable for the letters, which makes it difficult to animate and add movement to each character." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 548, + 314, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 314, + 561 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 314, + 561 + ], + "type": "text", + "content": "2.2 Typography Video Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 616 + ], + "type": "text", + "content": "As we mentioned above, this work is the first attempt to generate kinetic typography. Although no existing works directly align with this, there are some works related to typography videos [26, 28, 29, 38, 63, 65]. Here, we would like to introduce two methods related to ours." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "DynTypo [37] proposes a dynamic typography model that transfers the dynamic effect with realistic movements like fire and water on a specific uppercase English letter to others. Shape-Matching GAN++ [72] transfers an image style into a target letter by matching these shapes with structures of the target letter." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 296, + 91, + 447, + 102 + ], + "type": "text", + "content": "Kinetic Typography Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 127, + 479, + 231 + ], + "blocks": [ + { + "bbox": [ + 189, + 114, + 423, + 125 + ], + "lines": [ + { + "bbox": [ + 189, + 114, + 423, + 125 + ], + "spans": [ + { + "bbox": [ + 189, + 114, + 423, + 125 + ], + "type": "text", + "content": "Table 1: A summary of the kinetic typography datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 127, + 479, + 231 + ], + "lines": [ + { + "bbox": [ + 133, + 127, + 479, + 231 + ], + "spans": [ + { + "bbox": [ + 133, + 127, + 479, + 231 + ], + "type": "table", + "html": "
DatasetDomain#SamplesVideoStatic CaptionDynamic CaptionMulti-letterText AppearanceBackground Text Movements
Multi-Content [2]Typography10K----✓-
TextLogo3K [61]Typography3K----✓-
TE141K [71]Typography141K---✓✓✓
TextSeg [68]Text Segmentation4K---✓✓✓
LAION-Glyph-10M [74]Text-to-Image10M-✓-✓✓✓
MARIO-10M [11]Text-to-Image10M-✓-✓✓✓
AnyWord-3M [54]Text-to-Image3M-✓-✓✓✓
CATER-GEN [21]Image-to-Video35K✓-✓N/AN/A△
WebVid-10M [3]Text-to-Video10M✓✓△N/AN/A✓
Vimeo25M [58]Text-to-Video25M✓✓△N/AN/A✓
OursKinetic Typography600K✓✓✓✓✓✓
", + "image_path": "d1b4c9ba3cf95e64c2d1f8c450c7d5a54b64609e7f273c61475e237e46011116.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 250, + 481, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 250, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 481, + 297 + ], + "type": "text", + "content": "However, the position and glyph of the letter are still fixed, with no actual movement between frames. Even under the condition of a single uppercase letter, representing various static and dynamic effects through text prompts remains a challenging issue." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 298, + 482, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 358 + ], + "type": "text", + "content": "By substituting the motion description with external user input, the task becomes more tractable. A work, Wakey-Wakey [66], transfers the source GIF (Graphics Interchange Format) motion to the target text. Here, graphic designers manually assign additional corresponding key points to achieve motion transfer. However, this process still requires direct human intervention." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 375, + 320, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 320, + 387 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 320, + 387 + ], + "type": "text", + "content": "2.3 Text-to-Video Diffusion Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 395, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 395, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 395, + 482, + 514 + ], + "type": "text", + "content": "The video diffusion model produces a visually-plausible and photo-realistic video based on text conditioning [7, 12, 16, 18, 20, 36, 49, 58]. The main challenge in the video diffusion model is to maintain the temporal coherency between frames. The pioneer works for video diffusion is to add temporal blocks to the text-to-image model and to learn temporal coherency between video frames [16, 20, 49]. Another type of relevant studies use pretrained models to leverage the temporal block [6, 7, 18]. In particular, Guo et al. [18] propose a motion module as the temporal block. It is designed as a plug-and-play module. After joint training with the pre-trained weight of stable diffusion [46], the module enables any text-to-image model to generate the video." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 514, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 586 + ], + "type": "text", + "content": "Existing typography has been mainly studied using static single letters. The similar work to kinetic typography is also single letters, with only variation of motion effect and almost no letter's movement exists. To initiate kinetic typography generation, we present our KineTy dataset and model that allows dynamic motion, effects and glyph deformation of multi-letters, which will be explained at the next section." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 604, + 324, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 604, + 324, + 618 + ], + "spans": [ + { + "bbox": [ + 132, + 604, + 324, + 618 + ], + "type": "text", + "content": "3 Kinetic Typography Dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "We describe how to build the KineTy dataset. Unlike previous datasets that only cover single-lettered images [71], our KineTy dataset has not only visual effects on multiple letters, but also their animations. We first provide detailed process" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "S. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 306, + 148 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 306, + 148 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 306, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 306, + 148 + ], + "type": "image", + "image_path": "17bfe6dbb3cc3a454ad1d956c7f83b1ee090a34b8f06abfd6aa7426fe10f2c0d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 148, + 306, + 161 + ], + "lines": [ + { + "bbox": [ + 133, + 148, + 306, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 148, + 306, + 161 + ], + "type": "text", + "content": "pink to white background." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 481, + 148 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 481, + 148 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 481, + 148 + ], + "type": "image", + "image_path": "a8c6f10e5518deddc1b8517ab215d2f867a21cbedfb2b03e9924e7a26ca338df.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 148, + 481, + 160 + ], + "lines": [ + { + "bbox": [ + 307, + 148, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 148, + 481, + 160 + ], + "type": "text", + "content": "giving it a vibrant and dimensional appearance on a white background." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 133, + 175, + 306, + 209 + ], + "blocks": [ + { + "bbox": [ + 133, + 161, + 291, + 173 + ], + "lines": [ + { + "bbox": [ + 133, + 161, + 291, + 173 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 291, + 173 + ], + "type": "text", + "content": "The word \"LgLed\" forms with a glossy, liquid effect, where each letter appears in sequence from left to right to gradually complete the word." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 133, + 175, + 306, + 209 + ], + "lines": [ + { + "bbox": [ + 133, + 175, + 306, + 209 + ], + "spans": [ + { + "bbox": [ + 133, + 175, + 306, + 209 + ], + "type": "image", + "image_path": "299721f2b7e5c6502d2dd9d0195d331e114e3c8dc30a871d08d8a31d9f93ba22.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 209, + 305, + 221 + ], + "lines": [ + { + "bbox": [ + 133, + 209, + 305, + 221 + ], + "spans": [ + { + "bbox": [ + 133, + 209, + 305, + 221 + ], + "type": "text", + "content": "The word \"ADI\" prominently displayed in bold, 3D-styled letters set against an orange textured background." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 175, + 481, + 209 + ], + "blocks": [ + { + "bbox": [ + 307, + 160, + 481, + 173 + ], + "lines": [ + { + "bbox": [ + 307, + 160, + 481, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 160, + 481, + 173 + ], + "type": "text", + "content": "The video illustrates the word \"HHtR\" materializing in layers, with each letter representing a layer and enhancing the previous ones to gradually construct the complete word." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 307, + 175, + 481, + 209 + ], + "lines": [ + { + "bbox": [ + 307, + 175, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 481, + 209 + ], + "type": "image", + "image_path": "2f0ec059ee10fafcea3c2de8e886848c1f78f46e4b2b3da8d162dcc82e135faf.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 209, + 471, + 220 + ], + "lines": [ + { + "bbox": [ + 307, + 209, + 471, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 471, + 220 + ], + "type": "text", + "content": "The word \"FLRNQVOIZ\" with a stylized pink colored glitch effect on a dark background." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 133, + 234, + 306, + 269 + ], + "blocks": [ + { + "bbox": [ + 133, + 221, + 200, + 232 + ], + "lines": [ + { + "bbox": [ + 133, + 221, + 200, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 221, + 200, + 232 + ], + "type": "text", + "content": "The word \"ADI\" gets bigger." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 133, + 234, + 306, + 269 + ], + "lines": [ + { + "bbox": [ + 133, + 234, + 306, + 269 + ], + "spans": [ + { + "bbox": [ + 133, + 234, + 306, + 269 + ], + "type": "image", + "image_path": "20d6eadc6c3330e13c5e80df381d0202d712502f31ff79bb93691c2a5cccbe69.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 269, + 283, + 280 + ], + "lines": [ + { + "bbox": [ + 133, + 269, + 283, + 280 + ], + "spans": [ + { + "bbox": [ + 133, + 269, + 283, + 280 + ], + "type": "text", + "content": "The word \"INSZG\" in bold letters casting a long shadow on a grey background." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 133, + 280, + 292, + 291 + ], + "lines": [ + { + "bbox": [ + 133, + 280, + 292, + 291 + ], + "spans": [ + { + "bbox": [ + 133, + 280, + 292, + 291 + ], + "type": "text", + "content": "The word \"INSZG\" is revealed in sequential order with each character appearing in order until the full word is displayed." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 132, + 298, + 481, + 342 + ], + "lines": [ + { + "bbox": [ + 132, + 298, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 298, + 481, + 342 + ], + "type": "text", + "content": "Fig. 2: Examples of our Kinecty dataset. Our dataset provides high-quality kinetic typography video created by professional motion graphic designers, along with captions that describe the visual appearance and motion effects. To aid visualization, we provide three frames from each video clip." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 307, + 234, + 481, + 269 + ], + "blocks": [ + { + "bbox": [ + 307, + 220, + 476, + 232 + ], + "lines": [ + { + "bbox": [ + 307, + 220, + 476, + 232 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 476, + 232 + ], + "type": "text", + "content": "The word \"FRLNQVOZ\" starts from a scattered formation and converging to form the word clearly." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 307, + 234, + 481, + 269 + ], + "lines": [ + { + "bbox": [ + 307, + 234, + 481, + 269 + ], + "spans": [ + { + "bbox": [ + 307, + 234, + 481, + 269 + ], + "type": "image", + "image_path": "252a0fd273ddb66c019680d90258fa14e633cb8bbecb2d066b18acb077fc91bf.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 269, + 481, + 280 + ], + "lines": [ + { + "bbox": [ + 307, + 269, + 481, + 280 + ], + "spans": [ + { + "bbox": [ + 307, + 269, + 481, + 280 + ], + "type": "text", + "content": "A colorful, distorted graphic with the word \"CmfyTSr\" in stylized lettering on a dark background." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 307, + 280, + 481, + 291 + ], + "lines": [ + { + "bbox": [ + 307, + 280, + 481, + 291 + ], + "spans": [ + { + "bbox": [ + 307, + 280, + 481, + 291 + ], + "type": "text", + "content": "The video depicts a fluid-like stream that gradually forms the word \"CmfyTSR.\" The flow is slow and smooth in a smooth flowing sequence giving the impression of a liquid in motion." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 22 + }, + { + "bbox": [ + 131, + 367, + 481, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 367, + 481, + 428 + ], + "spans": [ + { + "bbox": [ + 131, + 367, + 481, + 428 + ], + "type": "text", + "content": "to render kinetic typography videos using templates made by professional editors in Sec. 3.1. Next, we introduce an way to describe their appearance and motions through text prompts in Sec. 3.2. Lastly, we explain how to make ground-truth kinetic typography video for strictly fair comparisons in Sec. 3.3. We summarize the difference between ours and existing datasets in Tab. 1." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 132, + 449, + 246, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 246, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 246, + 460 + ], + "type": "text", + "content": "3.1 Video Rendering" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 130, + 472, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 533 + ], + "type": "text", + "content": "Employing templates for video rendering. The template is a pre-designed project file which contains a visual effect on letters. Many editors prefer to use the templates because it saves their time and labor cost in practice. Following the best practice, we utilize 584 kinetic typography templates from professional graphic designers for our dataset construction, as visualized in Fig. 2." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 130, + 533, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 617 + ], + "type": "text", + "content": "Set of multiple letters. Next, we utilize the kinetic typography templates and randomly replace the text contents for augmentation. We employ multiple letters, in contrast to existing typography datasets with static single-letters [2,31,70]. Text contents are randomly generated by arranging up to 12 letters sampled from a set of 52 letters, including both uppercase and lowercase alphabets. Through this, we can expect rich letter-by-letter effects with various arrangements, while also keeping consistent styles across the multiple letters." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "Speculation. We render videos with " + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "inline_equation", + "content": "1,920 \\times 1,080" + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": " resolution for 3 sec., and 1,000 random words from 584 templates. Subsequently, all videos are downsampled to " + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "inline_equation", + "content": "512 \\times 288" + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": " resolution with 8fps for training. It takes a month to render the whole dataset with four i9 13900KF CPUs and an NVIDIA 2080ti GPU." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 297, + 91, + 447, + 102 + ], + "type": "text", + "content": "Kinetic Typography Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 114, + 482, + 210 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 482, + 210 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 482, + 210 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 482, + 210 + ], + "type": "image", + "image_path": "36479d975f307eba64d998dee15e397e39683cd323c01c92b4271c547caa9ef8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 213, + 392, + 224 + ], + "lines": [ + { + "bbox": [ + 220, + 213, + 392, + 224 + ], + "spans": [ + { + "bbox": [ + 220, + 213, + 392, + 224 + ], + "type": "text", + "content": "Fig. 3: Statistics of our proposed dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 240, + 252, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 252, + 253 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 252, + 253 + ], + "type": "text", + "content": "3.2 Video Captioning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 258, + 482, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 317 + ], + "type": "text", + "content": "Static and dynamic effect separation. In the next step, we caption the rendered kinetic typography for text-to-video generation. The professional editors usually design the static appearances at first, and then apply dynamic effects. Following this practice, we use two type of captions: static and dynamic captions, and labeled them accordingly." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 319, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 402 + ], + "type": "text", + "content": "Static effect captioning. To label static captions, we focus on letter's appearance which contains spatial information of typography. As shown in Fig. 2, the static captions " + }, + { + "bbox": [ + 130, + 319, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\text {static }}" + }, + { + "bbox": [ + 130, + 319, + 482, + 402 + ], + "type": "text", + "content": " describe the color of the letters and the background, the glyph of the letters (e.g., outlined with yellow color or bold font), the characteristics of the background (e.g., textured and shiny background) and the arrangement of the letters (e.g., in a diagonal way) based on the last frame of the video where all the letters are displayed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 403, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 485 + ], + "type": "text", + "content": "Dynamic effect captioning. Similarly, we concentrate on the temporal change of motions for each frame to write dynamic captions. Dynamic captions " + }, + { + "bbox": [ + 130, + 403, + 482, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{dynamic}" + }, + { + "bbox": [ + 130, + 403, + 482, + 485 + ], + "type": "text", + "content": " describe the motion part of the video, such as whether each letter appears in sequential or random order, can be rotated, or has a fade-in effect. For better systematic process, we initially label the videos with the GPT-4Vision model [41], and then manually verify and refine them to fix missing and wrongly labeled components." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "text", + "content": "Statistics. The statistics of our dataset is summarized in Fig. 3. There are three main categories for the static appearance in Fig. 3 (a): Text type, text effect and background. Here, the text fill effects are dominant in the text type because it is the most obvious way to clearly show the text content. Especially, professional designers tend to prefer using solid colors, rather than text outlines. The reason why " + }, + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "inline_equation", + "content": "57.7\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 581 + ], + "type": "text", + "content": " of our dataset has no text effect is the readability of text contents. Since kinetic typography makes dynamic motions of the contents, users do not need for fancy visual effects on it." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "In addition, we categorize dynamic motions into two-fold in Fig. 3 (b): motion type, including entrance, emphasis and movement, and sequence type: Our dataset supports various effect of entrance and emphasis to deliver striking messages. In contrast, the line has the majority in the movement because users' intention is usually conveyed after all letters in a scene are arranged in a line. In the same vein, when words come inside all at one and appear one-by-one in sequence, it is readable. They thus have the high portion in the sequence category." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "S. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 482, + 182 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 182 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 182 + ], + "type": "image", + "image_path": "332008dee99c2e7634aad4e753ccc52e7fb96b6e40cada6f1eb788ca418087de.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 186, + 400, + 198 + ], + "lines": [ + { + "bbox": [ + 212, + 186, + 400, + 198 + ], + "spans": [ + { + "bbox": [ + 212, + 186, + 400, + 198 + ], + "type": "text", + "content": "Fig. 4: An architecture of our KineTy model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 219, + 324, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 219, + 324, + 231 + ], + "spans": [ + { + "bbox": [ + 131, + 219, + 324, + 231 + ], + "type": "text", + "content": "3.3 Ground-truth Video Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 237, + 479, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 479, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 479, + 273 + ], + "type": "text", + "content": "Unlike generating random synthetic words in a training phase, we use real-world letters for evaluation by leveraging the templates. We can render ground-truth videos corresponding to user-input captions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 274, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 320 + ], + "type": "text", + "content": "For evaluation, individual letters are selected for each alphabet letter from A to Z, and the first letters are capitalized for checking case sensitivity (e.g., Apple, Ball, ..., Zebra). These 26 words are used for rendering 584 templates, so that a total of 15,184 videos are used for the evaluation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 337, + 373, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 337, + 373, + 351 + ], + "spans": [ + { + "bbox": [ + 131, + 337, + 373, + 351 + ], + "type": "text", + "content": "4 Kinetic Typography Diffusion Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 360, + 480, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 480, + 419 + ], + "type": "text", + "content": "In this section, we present our model for creating kinetic typography videos based on textual prompts. The key concept is to show animate visual effects on texts, and our challenges in this work are summarized as: (1) the effect should be visually-pleasing and eye-catching to viewers; (2) the transition between frames should be smooth and align with the captions; (3) the text must be readable." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 482, + 479 + ], + "type": "text", + "content": "We begin by defining a kinetic typography generation problem in Sec. 4.1. We then describe how to model static and dynamic captions that efficiently guide appearance and motions in Sec. 4.2. We lastly discuss how to improve the glyph legibility with respect to the model design and its learning strategy in Sec. 4.3. Implementation details are provided in Sec. 4.4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 495, + 222, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 495, + 222, + 508 + ], + "spans": [ + { + "bbox": [ + 131, + 495, + 222, + 508 + ], + "type": "text", + "content": "4.1 Preliminary" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "text", + "content": "Conditional Latent Diffusion Models. Diffusion models train the data distribution by progressively refining a noisy initial state " + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "inline_equation", + "content": "z_{M} \\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "text", + "content": " into the target data representation " + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "text", + "content": " diffusion steps. Recent advancements, particularly in Latent Diffusion Models (LDMs) [46], enhance the efficiency by encoding an image " + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "text", + "content": " into a compact latent representation " + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "inline_equation", + "content": "z_{0} = \\mathcal{E}(x)" + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "text", + "content": " using an encoder " + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "text", + "content": ", and is transformed back to the image " + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "inline_equation", + "content": "\\tilde{x} = \\mathcal{D}(z_{0})" + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "text", + "content": " using a decoder " + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "text", + "content": ". These adding noise and its subsequent removal are done with U-Net-style denoising network " + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\pi}" + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "text", + "content": ". In addition, a condition " + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "text", + "content": " is mapped to the hidden state " + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\pi}" + }, + { + "bbox": [ + 130, + 512, + 482, + 620 + ], + "type": "text", + "content": " via attention:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 203, + 624, + 482, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 624, + 482, + 653 + ], + "spans": [ + { + "bbox": [ + 203, + 624, + 482, + 653 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t n} _ {\\theta} (h, y) = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {\\theta} (h) K _ {\\theta} (y) ^ {\\top}}{\\sqrt {d}}\\right) \\cdot V _ {\\theta} (y) \\tag {1}", + "image_path": "072137385e69c583cb6d9ac12998cd8908ffef4ac3057645317c5adf28241283.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 175, + 654, + 438, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 654, + 438, + 668 + ], + "spans": [ + { + "bbox": [ + 175, + 654, + 438, + 668 + ], + "type": "interline_equation", + "content": "\\mathrm {s . t .} Q _ {\\theta} (h) = W _ {\\theta , Q} \\cdot h, K _ {\\theta} (y) = W _ {\\theta , K} \\cdot y, V _ {\\theta} (y) = W _ {\\theta , V} \\cdot y,", + "image_path": "2228921ca5ecc792b34acab76dd95322adb490f1dca3ffa31cfd2f9bac586575.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 296, + 91, + 448, + 102 + ], + "type": "text", + "content": "Kinetic Typography Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "W_{\\theta, Q}" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "W_{\\theta, K}" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "W_{\\theta, V}" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " denote the learnable parameters for queries, keys and values. " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " is the number of dimensions of the keys. The objective function for the denoising network " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\pi}" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 201, + 162, + 482, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 162, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 201, + 162, + 482, + 182 + ], + "type": "interline_equation", + "content": "L _ {l d m} = \\mathbb {E} _ {\\mathcal {E} (x), y, \\epsilon \\sim \\mathcal {N} (0, 1), m} \\left[ \\left\\| \\epsilon - \\epsilon_ {\\pi} \\left(z _ {m}, m, y\\right) \\right\\| _ {2} ^ {2} \\right], \\tag {2}", + "image_path": "84e633f94f70a776ed18cddf3ddf21f08815adcf76cb18a384a97278dd1bec13.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 191, + 481, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 191, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 481, + 215 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 191, + 481, + 215 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_2" + }, + { + "bbox": [ + 130, + 191, + 481, + 215 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 191, + 481, + 215 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 130, + 191, + 481, + 215 + ], + "type": "text", + "content": " distance and diffusion step " + }, + { + "bbox": [ + 130, + 191, + 481, + 215 + ], + "type": "inline_equation", + "content": "m \\in \\{1, \\dots, M\\}" + }, + { + "bbox": [ + 130, + 191, + 481, + 215 + ], + "type": "text", + "content": " is uniformly sampled during training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 217, + 482, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 300 + ], + "type": "text", + "content": "Extending image diffusion models to video. Extending the capability of text-to-image diffusion models to video generation has been more feasible. By incorporating a temporal self-attention called motion module [18], it learns a temporal coherency between " + }, + { + "bbox": [ + 130, + 217, + 482, + 300 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 217, + 482, + 300 + ], + "type": "text", + "content": " frames in a latent sequence " + }, + { + "bbox": [ + 130, + 217, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0^{0:T}" + }, + { + "bbox": [ + 130, + 217, + 482, + 300 + ], + "type": "text", + "content": ", which corresponds to the video sequence " + }, + { + "bbox": [ + 130, + 217, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}^{0:T} = \\mathcal{D}(\\mathbf{z}_0^{0:T})" + }, + { + "bbox": [ + 130, + 217, + 482, + 300 + ], + "type": "text", + "content": ". This allows diffusion models to generate smoothly changed sequential images over time, whose temporal self-attention can be defined as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 169, + 308, + 482, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 308, + 482, + 337 + ], + "spans": [ + { + "bbox": [ + 169, + 308, + 482, + 337 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t n} _ {\\phi} \\left(\\mathbf {h} ^ {0: T}, \\mathbf {h} ^ {0: T}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {\\phi} \\left(\\mathbf {h} ^ {0 : T}\\right) K _ {\\phi} \\left(\\mathbf {h} ^ {0 : T}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\cdot V _ {\\phi} \\left(\\mathbf {h} ^ {0: T}\\right). \\tag {3}", + "image_path": "e90ce47e9235da125cb7ae1a93aa5422f3aac7771ae817b12f5e36aabc621627.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 355, + 321, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 355, + 321, + 368 + ], + "spans": [ + { + "bbox": [ + 132, + 355, + 321, + 368 + ], + "type": "text", + "content": "4.2 Spatial and Temporal Guidance" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 378, + 481, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 481, + 413 + ], + "type": "text", + "content": "Static caption incorporation. Motivated by the best practice of professional designers who handle appearance and motion effects separately, we divide the caption into static and dynamic elements." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 415, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 486 + ], + "type": "text", + "content": "When existing text-to-image models learn the distribution of image data, they are conditioned on texts related to the image's appearance. In the same manner, our model is also guided by captions describing the appearance of the text contents and background in each frame of video. Using Eq. (1), we define a spatial attention block with a self-attention followed by a cross-attention with a static caption " + }, + { + "bbox": [ + 130, + 415, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\text{static}}" + }, + { + "bbox": [ + 130, + 415, + 482, + 486 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 207, + 496, + 482, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 496, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 207, + 496, + 482, + 514 + ], + "type": "interline_equation", + "content": "\\mathbf {h} ^ {\\prime 0: T} = \\left\\{\\operatorname {A t t n} _ {\\vartheta} \\left(\\operatorname {A t t n} _ {\\theta} \\left(\\mathbf {h} ^ {t}, \\mathbf {h} ^ {t}\\right), \\tau \\left(\\mathcal {C} _ {\\text {s t a t i c}}\\right)\\right) \\right\\} _ {t = 0} ^ {T - 1}, \\tag {4}", + "image_path": "ae37afd363c465920600cefde0834af513d959b41f20b26c9bb3aacad5f90f9b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 522, + 314, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 314, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 314, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 522, + 314, + 534 + ], + "type": "inline_equation", + "content": "\\tau (\\cdot)" + }, + { + "bbox": [ + 130, + 522, + 314, + 534 + ], + "type": "text", + "content": " is the CLIP text encoder [44]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 535, + 482, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 535, + 482, + 607 + ], + "spans": [ + { + "bbox": [ + 130, + 535, + 482, + 607 + ], + "type": "text", + "content": "Dynamic caption incorporation. A previous work [18] uses a motion module that only utilizes a self-attention between frames to learn the temporal consistency. On the other hand, in kinetic typography, it is essential to accurately display the dynamic motion effects of each letter in a video, following user's textual description. To do this, we extend Eq. (3) by adding the cross-attention with a dynamic caption as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 205, + 616, + 482, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 616, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 205, + 616, + 482, + 632 + ], + "type": "interline_equation", + "content": "\\mathbf {h} ^ {\\prime 0: T} = \\operatorname {A t t n} _ {\\psi} \\big (\\operatorname {A t t n} _ {\\phi} (\\mathbf {h} ^ {0: T}, \\mathbf {h} ^ {0: T}), \\tau \\big (\\mathcal {C} _ {\\text {d y n a m i c}} \\big) \\big), \\tag {5}", + "image_path": "87b0cfc70eb398b29d5cf64dee45adaaaf4cfffb6677c2bf78524fda9c545cac.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Through this process, the model becomes capable of maintaining temporal consistency as well as direct control over dynamic movements." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "S. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 302, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 302, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 302, + 129 + ], + "type": "text", + "content": "4.3 Enhancing Glyph Legibility." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 183 + ], + "type": "text", + "content": "Readability improvement. As a next step, we aim to make the legibility of our model better. To do this, we first classify descriptions for text contents and prompt because diffusion models often have a difficulty in distinguishing them. In this work, we put down a delimiter between each character for text contents." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 184, + 482, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 184, + 482, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 184, + 482, + 243 + ], + "type": "text", + "content": "Let a text content " + }, + { + "bbox": [ + 130, + 184, + 482, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\{l_1, l_2, \\dots, l_L\\}" + }, + { + "bbox": [ + 130, + 184, + 482, + 243 + ], + "type": "text", + "content": " contain " + }, + { + "bbox": [ + 130, + 184, + 482, + 243 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 184, + 482, + 243 + ], + "type": "text", + "content": " letters. To condition each letter separately, we join each letter " + }, + { + "bbox": [ + 130, + 184, + 482, + 243 + ], + "type": "inline_equation", + "content": "l_i" + }, + { + "bbox": [ + 130, + 184, + 482, + 243 + ], + "type": "text", + "content": " in the text content with a delimiter symbol " + }, + { + "bbox": [ + 130, + 184, + 482, + 243 + ], + "type": "inline_equation", + "content": "|\\cdot\\rangle" + }, + { + "bbox": [ + 130, + 184, + 482, + 243 + ], + "type": "text", + "content": ". Further, we add a symbol '\\~' behind upper case letters because the clip encoder does not care of upper cases. To the end, when the word " + }, + { + "bbox": [ + 130, + 184, + 482, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 130, + 184, + 482, + 243 + ], + "type": "text", + "content": " denotes \"Apple\", " + }, + { + "bbox": [ + 130, + 184, + 482, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{L}'" + }, + { + "bbox": [ + 130, + 184, + 482, + 243 + ], + "type": "text", + "content": " is represented as \"A^{|p|p|l|e}\". We denote the whole " + }, + { + "bbox": [ + 130, + 184, + 482, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 130, + 184, + 482, + 243 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 184, + 482, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{L}'" + }, + { + "bbox": [ + 130, + 184, + 482, + 243 + ], + "type": "text", + "content": " for better representation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 243, + 482, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 364 + ], + "type": "text", + "content": "Attention on text contents. It can be challenging for the attention module to guide both text contents and effects at once, while disentangling the text content information in CLIP feature space. What is worse, the text content can be overwhelmed by relatively long appearance and motion effects in CLIP text encoder. To tackle this problem, we introduce an additional cross-attention branch to the text content. Inspired by ControlNet [76], we regard the text content as conditions through the zero convolution operation. Starting with the definition of word caption " + }, + { + "bbox": [ + 130, + 243, + 482, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{word}" + }, + { + "bbox": [ + 130, + 243, + 482, + 364 + ], + "type": "text", + "content": " via a prompt template like \"The word " + }, + { + "bbox": [ + 130, + 243, + 482, + 364 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{L}'\\}" + }, + { + "bbox": [ + 130, + 243, + 482, + 364 + ], + "type": "text", + "content": "\", we extend Eq. (4) by adding the cross-attention module between the word caption and hidden feature, weighted by the zero-initialized convolutions " + }, + { + "bbox": [ + 130, + 243, + 482, + 364 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 243, + 482, + 364 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 168, + 372, + 482, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 372, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 168, + 372, + 482, + 410 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {h} _ {\\text {s t a t i c}} ^ {\\prime \\prime 0: T} = \\left\\{\\operatorname {A t t n} _ {\\vartheta} \\left(\\mathbf {h} ^ {\\prime}, \\tau \\left(\\mathcal {C} _ {\\text {s t a t i c}}\\right)\\right) + \\rho_ {v} \\left(\\operatorname {A t t n} _ {v} \\left(\\mathbf {h} ^ {\\prime}, \\tau \\left(\\mathcal {C} _ {\\text {w o r d}}\\right)\\right)\\right) \\right\\} _ {t = 0} ^ {T - 1}, \\tag {6} \\\\ w h e r e \\quad \\mathbf {h} ^ {\\prime} = \\operatorname {A t t n} _ {\\theta} \\left(\\mathbf {h} ^ {t}, \\mathbf {h} ^ {t}\\right). \\\\ \\end{array}", + "image_path": "c0e59a2b8514cf0c02a626d4a03ec67f424ed2b5adccdab3970cb17f4f4aa1d7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 413, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 450 + ], + "type": "text", + "content": "This allows the network to gradually evaluate the usefulness of this additional caption based on the condition. In the same way, we incorporate the word caption into the temporal cross-attention in Eq. (5) as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 172, + 459, + 482, + 494 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 459, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 172, + 459, + 482, + 494 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {h} _ {\\text {d y n a m i c}} ^ {\\prime \\prime 0: T} = \\operatorname {A t t n} _ {\\psi} \\left(\\mathbf {h} ^ {\\prime}, \\tau \\left(\\mathcal {C} _ {\\text {d y n a m i c}}\\right)\\right) + \\rho_ {\\varphi} \\left(\\operatorname {A t t n} _ {\\varphi} \\left(\\mathbf {h} ^ {\\prime}, \\tau \\left(\\mathcal {C} _ {\\text {w o r d}}\\right)\\right)\\right), \\tag {7} \\\\ w h e r e \\quad \\mathbf {h} ^ {\\prime} = \\operatorname {A t t n} _ {\\phi} \\left(\\mathbf {h} ^ {0: T}, \\mathbf {h} ^ {0: T}\\right). \\\\ \\end{array}", + "image_path": "4c3fc02a2b8d74987d259c7f3c3023dfbc7b6fdffda0e1f01eee391882bb0f19.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 497, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 570 + ], + "type": "text", + "content": "Glyph loss. To train our model, we use a common loss function " + }, + { + "bbox": [ + 130, + 497, + 482, + 570 + ], + "type": "inline_equation", + "content": "L_{ldm}" + }, + { + "bbox": [ + 130, + 497, + 482, + 570 + ], + "type": "text", + "content": ". Additionally, we impose an extra penalty on the letter regions to enforce a sharp and correct glyph of text contents. We first use a binary mask " + }, + { + "bbox": [ + 130, + 497, + 482, + 570 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 497, + 482, + 570 + ], + "type": "text", + "content": " for the letters in the last frame " + }, + { + "bbox": [ + 130, + 497, + 482, + 570 + ], + "type": "inline_equation", + "content": "V^T" + }, + { + "bbox": [ + 130, + 497, + 482, + 570 + ], + "type": "text", + "content": " from a text segmentation model [68]. The mask is then blurred to cover its surrounding effects. We then define a glyph loss based on " + }, + { + "bbox": [ + 130, + 497, + 482, + 570 + ], + "type": "inline_equation", + "content": "L_{LDM}" + }, + { + "bbox": [ + 130, + 497, + 482, + 570 + ], + "type": "text", + "content": " with the additional pixel-wise weighting strategy using the blurred mask as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 184, + 578, + 482, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 578, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 184, + 578, + 482, + 598 + ], + "type": "interline_equation", + "content": "L _ {g l y p h} = \\mathbb {E} _ {\\mathcal {E} (x), y, \\epsilon \\sim \\mathcal {N} (0, 1), m} \\left[ \\left\\| B \\odot \\left(\\epsilon - \\epsilon_ {\\pi} \\left(z _ {m} ^ {T}, m, y\\right)\\right) \\right\\| _ {2} ^ {2} \\right], \\tag {8}", + "image_path": "740848666a5b8e222ba8a44e064b589f0c5e6f2186376a6bc79f101f01699f6f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 324, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 324, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 324, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 605, + 324, + 616 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 605, + 324, + 616 + ], + "type": "text", + "content": " is the element-wise multiplication." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "With the glyph loss, we enhance the legibility of text contents, enabling the precise creation of multi-letter formations. Through the linear combination of both loss terms, we can formulate the final loss function as " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "L = L_{ldm} + \\alpha L_{glyph}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ". Here, we empirically set " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " to 0.01." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 296, + 91, + 447, + 102 + ], + "type": "text", + "content": "Kinetic Typography Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 279, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 279, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 279, + 128 + ], + "type": "text", + "content": "4.4 Implementation Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 266 + ], + "type": "text", + "content": "To train our model, we use a two-step training strategy similar to that of human trainees [14]. First, we pre-train the network to generate spatial appearances using only static caption and last video frame pairs. Here, we detach the temporal attention modules to make them work as text-to-image diffusion models during 30 epochs with a batch size of 200. Next, we train the full model while freezing the spatial attention using whole video frames and captions for an epoch with a batch size of 8. For training, we resize the height " + }, + { + "bbox": [ + 130, + 135, + 482, + 266 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 135, + 482, + 266 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 130, + 135, + 482, + 266 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 135, + 482, + 266 + ], + "type": "text", + "content": " of the video into " + }, + { + "bbox": [ + 130, + 135, + 482, + 266 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 135, + 482, + 266 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 135, + 482, + 266 + ], + "type": "inline_equation", + "content": "T = 24" + }, + { + "bbox": [ + 130, + 135, + 482, + 266 + ], + "type": "text", + "content": " frames. The training is performed with AdamW optimizer [35], with a diffusion step " + }, + { + "bbox": [ + 130, + 135, + 482, + 266 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 135, + 482, + 266 + ], + "type": "text", + "content": " of 1000 and a learning rate of 0.0001, which usually takes about 20 hours for training on 8 NVIDIA A100 GPUs. The inference time is about 20 seconds when the number of sampling steps is 25." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 284, + 230, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 284, + 230, + 297 + ], + "spans": [ + { + "bbox": [ + 132, + 284, + 230, + 297 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 307, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 482, + 380 + ], + "type": "text", + "content": "In this section, we conduct comprehensive experiments to verify the effectiveness of our model for kinetic typography. We first describe the experimental setup in Sec. 5.1. We then provide comparison results with relevant typography and video generation models, and report user-study to highlight the practical applicability of our model in Sec. 5.2. We lastly carry out an extensive ablation study to validate the effect of each component in our model in Sec. 5.3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 396, + 261, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 261, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 261, + 407 + ], + "type": "text", + "content": "5.1 Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 414, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 593 + ], + "type": "text", + "content": "Comparison methods. We compare ours with state-of-the-art generative models, which consist of 3 two-stage methods and 2 one-stage methods. The two-stage methods are based on combinations of text-to-image models, including DS-Fusion [52], GlyphControl [74] and Text-Diffuser [11], and an image-to-video model, SparseCtrl [17]. DS-Fusion outputs a stylized letter based on the given style phrase and a letter. Since our caption is a sentence, we extract a style keyword from it and consider it as a style phrase. Note that DS-Fusion is able to generate one letter at a time, so we concatenate each letter to make a full-text content. GlyphControl takes an instruction to generate a glyph image for text contents, then uses the image as a condition for the final output image. Text-Diffuser first finds a proper layout for words and produce the output based on a caption and the layout. Since the comparison methods yield images, we combine them with an image-to-video model, SparseCtrl, to compare with ours. SparseCtrl uses a caption and a guidance such as sketches, depth maps and images to generate a stylized video." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "On the other hand, one-stage-methods,AnimateDiff [18] and Lavie [58], are based on text-to-video models.AnimateDiff extends pre-trained text-to-image diffusion models with a motion module, and trains only the motion module to make fully use of the massive information of video dataset. Lavie leverages a temporal self-attention block to enhance a temporal consistency between frames while keeping the generation performance." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "S. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 138, + 479, + 213 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 2: Quantitative evaluation of two-stage generation methods, and one-stage text-to-video generation methods. Bold: Best, Underline: Second-best." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 138, + 479, + 213 + ], + "lines": [ + { + "bbox": [ + 134, + 138, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 134, + 138, + 479, + 213 + ], + "type": "table", + "html": "
ModelEntranceEmphasisMotionAverage
FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑
DS-F[52]+Sparse[17]1636.35.710.690.583184.95.970.730.581236.55.420.760.561843.105.660.700.58
Glyph[74]+Sparse[17]2415.64.300.760.682667.15.060.760.751468.64.480.840.802361.984.460.760.69
T-Diff[11]+Sparse[17]2937.25.610.780.903107.55.160.790.891208.24.060.820.952815.65.350.780.90
AnimateDiff[18]1613.84.710.590.012106.04.560.660.032160.64.510.670.071727.554.680.610.01
Lavie [58]825.12.250.820.541835.43.830.830.41505.443.330.850.35956.972.680.820.49
Ours147.41.790.870.71177.521.750.880.5436.161.620.940.98125.901.770.880.76
", + "image_path": "c63be7d77e6c5de12bd0ab9380bf79fff5fdf6025100b463f4534e33a86ade0a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 224, + 203, + 265 + ], + "blocks": [ + { + "bbox": [ + 133, + 224, + 203, + 265 + ], + "lines": [ + { + "bbox": [ + 133, + 224, + 203, + 265 + ], + "spans": [ + { + "bbox": [ + 133, + 224, + 203, + 265 + ], + "type": "image", + "image_path": "71782326842b53f99dee087045070277f47a92419be858169d0b1c441cd3bdb2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 203, + 224, + 272, + 265 + ], + "blocks": [ + { + "bbox": [ + 203, + 224, + 272, + 265 + ], + "lines": [ + { + "bbox": [ + 203, + 224, + 272, + 265 + ], + "spans": [ + { + "bbox": [ + 203, + 224, + 272, + 265 + ], + "type": "image", + "image_path": "6641917464a0de8178c6ea8044efebe2e111f39e81a4eaef998a2f5518fb1b54.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 272, + 224, + 342, + 265 + ], + "blocks": [ + { + "bbox": [ + 272, + 224, + 342, + 265 + ], + "lines": [ + { + "bbox": [ + 272, + 224, + 342, + 265 + ], + "spans": [ + { + "bbox": [ + 272, + 224, + 342, + 265 + ], + "type": "image", + "image_path": "05fce8642548fec43bc3020bec1932a97a35155db0895a0f08126a8813a9375e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 342, + 224, + 411, + 265 + ], + "blocks": [ + { + "bbox": [ + 342, + 224, + 411, + 265 + ], + "lines": [ + { + "bbox": [ + 342, + 224, + 411, + 265 + ], + "spans": [ + { + "bbox": [ + 342, + 224, + 411, + 265 + ], + "type": "image", + "image_path": "430afc07bae35518711bf141055c3fcf1dab309f032232f8fe2a76cfeeb55957.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 411, + 224, + 480, + 265 + ], + "blocks": [ + { + "bbox": [ + 411, + 224, + 480, + 265 + ], + "lines": [ + { + "bbox": [ + 411, + 224, + 480, + 265 + ], + "spans": [ + { + "bbox": [ + 411, + 224, + 480, + 265 + ], + "type": "image", + "image_path": "52669d860eb45a89f7fb1ca1a130deaf098cd2d654a4e6e09de00d4184b55b1a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 265, + 352, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 265, + 352, + 274 + ], + "spans": [ + { + "bbox": [ + 133, + 265, + 352, + 274 + ], + "type": "text", + "content": "Static caption: The word 'Dog' in bold, green-stripped letters against a textured coral background. Dynamic caption: The sequential addition of letters one by one to spell out the word 'Dog'." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 133, + 275, + 203, + 313 + ], + "blocks": [ + { + "bbox": [ + 133, + 275, + 203, + 313 + ], + "lines": [ + { + "bbox": [ + 133, + 275, + 203, + 313 + ], + "spans": [ + { + "bbox": [ + 133, + 275, + 203, + 313 + ], + "type": "image", + "image_path": "725d9f0177562488b62051c1829de0480da90cf8738b5a6eab434fba78ed85d1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 394, + 480, + 415 + ], + "lines": [ + { + "bbox": [ + 131, + 394, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 131, + 394, + 480, + 415 + ], + "type": "text", + "content": "Fig. 5: Qualitative results from the comparison models and ours. Obviously, the results from ours reflect the contents of captions better than the others." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 203, + 275, + 272, + 313 + ], + "blocks": [ + { + "bbox": [ + 203, + 275, + 272, + 313 + ], + "lines": [ + { + "bbox": [ + 203, + 275, + 272, + 313 + ], + "spans": [ + { + "bbox": [ + 203, + 275, + 272, + 313 + ], + "type": "image", + "image_path": "18c86c52a02390aa315b9aeff28d706f1e9c14eee3ad5986533c4856cabeda9f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 272, + 275, + 342, + 313 + ], + "blocks": [ + { + "bbox": [ + 272, + 275, + 342, + 313 + ], + "lines": [ + { + "bbox": [ + 272, + 275, + 342, + 313 + ], + "spans": [ + { + "bbox": [ + 272, + 275, + 342, + 313 + ], + "type": "image", + "image_path": "f89da7733b5becbc819a3b1246035ed2db8de3a689ef745483c5cc71bd0ee9cd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 342, + 275, + 411, + 313 + ], + "blocks": [ + { + "bbox": [ + 342, + 275, + 411, + 313 + ], + "lines": [ + { + "bbox": [ + 342, + 275, + 411, + 313 + ], + "spans": [ + { + "bbox": [ + 342, + 275, + 411, + 313 + ], + "type": "image", + "image_path": "0feb4f8301ead937bbda8bae32405b40814e4149d36d3d653dff3f1825692a69.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 411, + 275, + 480, + 313 + ], + "blocks": [ + { + "bbox": [ + 411, + 275, + 480, + 313 + ], + "lines": [ + { + "bbox": [ + 411, + 275, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 411, + 275, + 480, + 313 + ], + "type": "image", + "image_path": "c71bf862a28e003657c5b309f564ac140b75dc6dc4534a6c9d97ab038373c49e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 314, + 474, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 314, + 474, + 324 + ], + "spans": [ + { + "bbox": [ + 133, + 314, + 474, + 324 + ], + "type": "text", + "content": "Static caption: The word Mermaid in a blurred, tunnel-like visual effect, creating a sense of depth and distortion on a dark background. Dynamic caption: The word Mermaid gradually comes into focus, with each letter appearing in a random order, intensifying the sense of depth and creating a captivating rev" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 133, + 326, + 203, + 365 + ], + "blocks": [ + { + "bbox": [ + 133, + 326, + 203, + 365 + ], + "lines": [ + { + "bbox": [ + 133, + 326, + 203, + 365 + ], + "spans": [ + { + "bbox": [ + 133, + 326, + 203, + 365 + ], + "type": "image", + "image_path": "db57bd5ce63cf6aac5ac85486c3296071025225634e4224bb07fd10af688e203.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 203, + 326, + 272, + 365 + ], + "blocks": [ + { + "bbox": [ + 203, + 326, + 272, + 365 + ], + "lines": [ + { + "bbox": [ + 203, + 326, + 272, + 365 + ], + "spans": [ + { + "bbox": [ + 203, + 326, + 272, + 365 + ], + "type": "image", + "image_path": "dc63aa8892f1dc22fe0f3adc810c83059f39d91ff9b920b0ca92cb85dd29d5e5.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 272, + 326, + 342, + 365 + ], + "blocks": [ + { + "bbox": [ + 272, + 326, + 342, + 365 + ], + "lines": [ + { + "bbox": [ + 272, + 326, + 342, + 365 + ], + "spans": [ + { + "bbox": [ + 272, + 326, + 342, + 365 + ], + "type": "image", + "image_path": "5b8f5443d35da0d2affedee1bfb4b78f430e4ea1e40c6979a63b11ba075fe648.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 342, + 326, + 411, + 365 + ], + "blocks": [ + { + "bbox": [ + 342, + 326, + 411, + 365 + ], + "lines": [ + { + "bbox": [ + 342, + 326, + 411, + 365 + ], + "spans": [ + { + "bbox": [ + 342, + 326, + 411, + 365 + ], + "type": "image", + "image_path": "2f89c99544f55493f992466f0075e9762afa9f6ff646f601aa6c689eb0f3e9b6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 411, + 326, + 480, + 365 + ], + "blocks": [ + { + "bbox": [ + 411, + 326, + 480, + 365 + ], + "lines": [ + { + "bbox": [ + 411, + 326, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 411, + 326, + 480, + 365 + ], + "type": "image", + "image_path": "8b83ac5a8b4f61f82533292a21623fca2fc3fd4adddc2dd716ba3dd2718460b9.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 133, + 365, + 459, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 365, + 459, + 373 + ], + "spans": [ + { + "bbox": [ + 133, + 365, + 459, + 373 + ], + "type": "text", + "content": "Static caption: The word 'Watermelon' in white letters against a black background, accompanied by two smiling emojis faces with heart eyes. Dynamic caption: Each character of the word 'Watermelon' appears sequentially, with various smiling emojis popping up randomly around the text." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 149, + 376, + 186, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 376, + 186, + 384 + ], + "spans": [ + { + "bbox": [ + 149, + 376, + 186, + 384 + ], + "type": "text", + "content": "(a) DS-Fusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 151, + 384, + 185, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 384, + 185, + 389 + ], + "spans": [ + { + "bbox": [ + 151, + 384, + 185, + 389 + ], + "type": "text", + "content": "+ SparseCtrl" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 214, + 376, + 259, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 376, + 259, + 384 + ], + "spans": [ + { + "bbox": [ + 214, + 376, + 259, + 384 + ], + "type": "text", + "content": "(b) GlyphControl" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 220, + 384, + 253, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 384, + 253, + 390 + ], + "spans": [ + { + "bbox": [ + 220, + 384, + 253, + 390 + ], + "type": "text", + "content": "+SparseCtrl" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 283, + 376, + 328, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 283, + 376, + 328, + 384 + ], + "spans": [ + { + "bbox": [ + 283, + 376, + 328, + 384 + ], + "type": "text", + "content": "(c) Text-Diffuser" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 290, + 384, + 323, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 384, + 323, + 390 + ], + "spans": [ + { + "bbox": [ + 290, + 384, + 323, + 390 + ], + "type": "text", + "content": "+ SparseCtrl" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 364, + 376, + 389, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 376, + 389, + 383 + ], + "spans": [ + { + "bbox": [ + 364, + 376, + 389, + 383 + ], + "type": "text", + "content": "(d) Lavie" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 435, + 376, + 457, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 435, + 376, + 457, + 383 + ], + "spans": [ + { + "bbox": [ + 435, + 376, + 457, + 383 + ], + "type": "text", + "content": "(e) Ours" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 130, + 432, + 480, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 480, + 479 + ], + "type": "text", + "content": "For a fair comparison, we utilize the official source codes and pre-trained weights provided by the authors. For more analysis, we categorize the motion effects into three groups, 'Entrance', 'Emphasis' and 'Motion', and conduct experiments to check the quality of each result." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 130, + 480, + 481, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 481, + 600 + ], + "type": "text", + "content": "Evaluation metrics. Since the final products of kinetic typography depend on the designer's preferences, we adopt four feature-based metrics rather than photoconsistency measures like PSNR. (1) Fréchet Video Distance (FVD) qualifies a similarity between two videos by comparing their feature representations extracted from a pre-trained deep neural network [55]. (2) Inception Score (IS) evaluates the quality and diversity of images generated by a model [4]. (3) CLIPScore measures the semantic alignment between generated images and their textual descriptions, using the CLIP model [19, 22, 44]. (4) Optical Character Recognition (OCR) checks the clarity of the output by comparing recognized text to a reference text using the F1-score." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 131, + 620, + 255, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 255, + 630 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 255, + 630 + ], + "type": "text", + "content": "5.2 Evaluation Results" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "Quantitative results. As shown in Tab. 2, our Kinety model achieves the better performance than the comparison methods in the most metrics on all the" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 297, + 91, + 447, + 102 + ], + "type": "text", + "content": "Kinetic Typography Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 334, + 229 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 334, + 229 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 334, + 229 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 334, + 229 + ], + "type": "image", + "image_path": "ab40cc94a030dfde61128a6056884fbfeda639dd999542f473ce3bba82c57d86.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 229, + 164, + 237 + ], + "blocks": [ + { + "bbox": [ + 134, + 229, + 164, + 237 + ], + "lines": [ + { + "bbox": [ + 134, + 229, + 164, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 229, + 164, + 237 + ], + "type": "image", + "image_path": "6f66d48c83b9388c2cda852f8b1135f976d91323de4cf8d1d1c6b4e07734c7dd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 240, + 479, + 262 + ], + "lines": [ + { + "bbox": [ + 132, + 240, + 479, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 479, + 262 + ], + "type": "text", + "content": "Fig. 6: Visualization of attention maps between letters from a caption and feature map of a hidden layer." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 173, + 229, + 190, + 237 + ], + "blocks": [ + { + "bbox": [ + 173, + 229, + 190, + 237 + ], + "lines": [ + { + "bbox": [ + 173, + 229, + 190, + 237 + ], + "spans": [ + { + "bbox": [ + 173, + 229, + 190, + 237 + ], + "type": "image", + "image_path": "263d4876eca264f3cbaea87f31cfa9e74ae2c6377af224a23698643ecb7adefd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 198, + 229, + 221, + 237 + ], + "blocks": [ + { + "bbox": [ + 198, + 229, + 221, + 237 + ], + "lines": [ + { + "bbox": [ + 198, + 229, + 221, + 237 + ], + "spans": [ + { + "bbox": [ + 198, + 229, + 221, + 237 + ], + "type": "image", + "image_path": "497865e5ac049905fc5ac22ec87f42c9aaaebb5bf1a18f75d6d3c96755934e68.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 230, + 229, + 241, + 237 + ], + "blocks": [ + { + "bbox": [ + 230, + 229, + 241, + 237 + ], + "lines": [ + { + "bbox": [ + 230, + 229, + 241, + 237 + ], + "spans": [ + { + "bbox": [ + 230, + 229, + 241, + 237 + ], + "type": "image", + "image_path": "8f775c883874303c4be34e3807bd5ea41820861fd1e14a10226ea1db42c59ec8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 249, + 229, + 265, + 237 + ], + "blocks": [ + { + "bbox": [ + 249, + 229, + 265, + 237 + ], + "lines": [ + { + "bbox": [ + 249, + 229, + 265, + 237 + ], + "spans": [ + { + "bbox": [ + 249, + 229, + 265, + 237 + ], + "type": "image", + "image_path": "b2880aa9c7ff5380ededae83965f46f4a08ce62526fda4eaf87269b6f408007d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 278, + 229, + 294, + 237 + ], + "blocks": [ + { + "bbox": [ + 278, + 229, + 294, + 237 + ], + "lines": [ + { + "bbox": [ + 278, + 229, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 278, + 229, + 294, + 237 + ], + "type": "image", + "image_path": "8f2abb81af309ba4df511cd93a97736f0009516dd2c0e7f3c7e300de2133def8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 302, + 229, + 319, + 237 + ], + "blocks": [ + { + "bbox": [ + 302, + 229, + 319, + 237 + ], + "lines": [ + { + "bbox": [ + 302, + 229, + 319, + 237 + ], + "spans": [ + { + "bbox": [ + 302, + 229, + 319, + 237 + ], + "type": "image", + "image_path": "b8e0bb6acd9f9357508616b74839a754397ff3de6dd32700c6574a55fea150b1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 335, + 114, + 481, + 229 + ], + "blocks": [ + { + "bbox": [ + 335, + 114, + 481, + 229 + ], + "lines": [ + { + "bbox": [ + 335, + 114, + 481, + 229 + ], + "spans": [ + { + "bbox": [ + 335, + 114, + 481, + 229 + ], + "type": "image", + "image_path": "057b2b05a8f49d404d77940ebbb804c48ec1a767d0b87df4f88dd7728e67512d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 229, + 367, + 237 + ], + "lines": [ + { + "bbox": [ + 336, + 229, + 367, + 237 + ], + "spans": [ + { + "bbox": [ + 336, + 229, + 367, + 237 + ], + "type": "text", + "content": "The word" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 375, + 229, + 386, + 237 + ], + "blocks": [ + { + "bbox": [ + 375, + 229, + 386, + 237 + ], + "lines": [ + { + "bbox": [ + 375, + 229, + 386, + 237 + ], + "spans": [ + { + "bbox": [ + 375, + 229, + 386, + 237 + ], + "type": "image", + "image_path": "326f57c1f7443e98bae158cb2562753b2a1fc71e21c364f1ede53cad23911da9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 394, + 229, + 404, + 237 + ], + "blocks": [ + { + "bbox": [ + 394, + 229, + 404, + 237 + ], + "lines": [ + { + "bbox": [ + 394, + 229, + 404, + 237 + ], + "spans": [ + { + "bbox": [ + 394, + 229, + 404, + 237 + ], + "type": "image", + "image_path": "6758910690e47a18537e1f9c04724d3fa496f988a1624c0651a1b40518a5119d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 412, + 229, + 422, + 237 + ], + "blocks": [ + { + "bbox": [ + 412, + 229, + 422, + 237 + ], + "lines": [ + { + "bbox": [ + 412, + 229, + 422, + 237 + ], + "spans": [ + { + "bbox": [ + 412, + 229, + 422, + 237 + ], + "type": "image", + "image_path": "f23cfb53c5f2760e5045299d91958b0037d79dcf314a890586d68f66c0e17808.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 430, + 229, + 440, + 237 + ], + "blocks": [ + { + "bbox": [ + 430, + 229, + 440, + 237 + ], + "lines": [ + { + "bbox": [ + 430, + 229, + 440, + 237 + ], + "spans": [ + { + "bbox": [ + 430, + 229, + 440, + 237 + ], + "type": "image", + "image_path": "ad572e26d114e4a06a400706820b16b29fa619ccfa96641c7d224ed341c92722.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 449, + 229, + 458, + 237 + ], + "blocks": [ + { + "bbox": [ + 449, + 229, + 458, + 237 + ], + "lines": [ + { + "bbox": [ + 449, + 229, + 458, + 237 + ], + "spans": [ + { + "bbox": [ + 449, + 229, + 458, + 237 + ], + "type": "image", + "image_path": "46d6dbec202424b4b31cdbefa4e66af86378bb54c115cae1c572ac71fc87112a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 133, + 274, + 481, + 304 + ], + "blocks": [ + { + "bbox": [ + 133, + 274, + 481, + 304 + ], + "lines": [ + { + "bbox": [ + 133, + 274, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 133, + 274, + 481, + 304 + ], + "type": "image", + "image_path": "bb8d3ea1b9ffe45b07f06f3e881003ba83a74be905dec44a8d3bd77027f7dc04.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 304, + 272, + 313 + ], + "lines": [ + { + "bbox": [ + 168, + 304, + 272, + 313 + ], + "spans": [ + { + "bbox": [ + 168, + 304, + 272, + 313 + ], + "type": "text", + "content": "(a) Initial kinetic typography video" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 359, + 304, + 429, + 312 + ], + "lines": [ + { + "bbox": [ + 359, + 304, + 429, + 312 + ], + "spans": [ + { + "bbox": [ + 359, + 304, + 429, + 312 + ], + "type": "text", + "content": "(b) Changing text color" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 133, + 316, + 481, + 345 + ], + "blocks": [ + { + "bbox": [ + 133, + 316, + 481, + 345 + ], + "lines": [ + { + "bbox": [ + 133, + 316, + 481, + 345 + ], + "spans": [ + { + "bbox": [ + 133, + 316, + 481, + 345 + ], + "type": "image", + "image_path": "d317fddc2596f93d21cdf4f1de59c186de4ac3c8f697a5eaae18e2797c135a45.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 346, + 253, + 354 + ], + "lines": [ + { + "bbox": [ + 187, + 346, + 253, + 354 + ], + "spans": [ + { + "bbox": [ + 187, + 346, + 253, + 354 + ], + "type": "text", + "content": "(c) Changing text font" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 356, + 346, + 433, + 354 + ], + "lines": [ + { + "bbox": [ + 356, + 346, + 433, + 354 + ], + "spans": [ + { + "bbox": [ + 356, + 346, + 433, + 354 + ], + "type": "text", + "content": "(d) Changing background" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 133, + 357, + 481, + 386 + ], + "blocks": [ + { + "bbox": [ + 133, + 357, + 481, + 386 + ], + "lines": [ + { + "bbox": [ + 133, + 357, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 133, + 357, + 481, + 386 + ], + "type": "image", + "image_path": "342fba2ed5aa57c5035b6b2205b0eb8f0ef20438f84fe7c3a4580c3a4eeae615.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 387, + 261, + 396 + ], + "lines": [ + { + "bbox": [ + 179, + 387, + 261, + 396 + ], + "spans": [ + { + "bbox": [ + 179, + 387, + 261, + 396 + ], + "type": "text", + "content": "(e) Changing motion effect" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 369, + 387, + 419, + 396 + ], + "lines": [ + { + "bbox": [ + 369, + 387, + 419, + 396 + ], + "spans": [ + { + "bbox": [ + 369, + 387, + 419, + 396 + ], + "type": "text", + "content": "(f) Final product" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 399, + 481, + 421 + ], + "lines": [ + { + "bbox": [ + 132, + 399, + 481, + 421 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 481, + 421 + ], + "type": "text", + "content": "Fig. 7: The demonstration of the editable capability of our Kinecty model after generating initial kinetic typography." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 437, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 555 + ], + "type": "text", + "content": "categories. The comparison methods, despite their impressive performance in general image and video generation, show unsatisfactory performance in creating kinetic typography. The reasons are mainly two-folds: (1) Text-to-video models are not specialized for kinetic typography and often fail to exactly distinguish the text content in the long captions. Different from them, we additionally feed " + }, + { + "bbox": [ + 130, + 437, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{word}" + }, + { + "bbox": [ + 130, + 437, + 482, + 555 + ], + "type": "text", + "content": " which imposes a high attention to the text content in the video. (2) They have challenges in creating letter-level motions. Since they are unaware of each letter in the text content, all the letters move as a whole or inconsistently. We handle this issue by providing a delimiter between each letter and using zero-convolution in Fig. 6." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 640 + ], + "type": "text", + "content": "Our model shows promising results in generating legible text content, but it sometimes falls short in OCR accuracy. This is because the two-stage methods algorithmically generate a glyph image of a text content, and then deform glyphs conditioned on the glyph image. As a result, they have the better performance in OCR metrics than ours and one-stage methods. Nevertheless, our method outperforms in the most cases, demonstrating the effectiveness of our method in generating the clear text content and motion effects." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Qualitative results. In Fig. 5, we display three examples of the comparison methods. Compared to ours, the two-stage methods do not understand the user's" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "S. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 138, + 479, + 267 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 3: The result of user-study. The score ranges from " + }, + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "inline_equation", + "content": "1 \\sim 6" + }, + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": " for Study 1 & 2, and 0 " + }, + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "inline_equation", + "content": "\\sim 1" + }, + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": " for Study 3. Bold: Best, Underline: Second-best." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 138, + 479, + 267 + ], + "lines": [ + { + "bbox": [ + 135, + 138, + 479, + 267 + ], + "spans": [ + { + "bbox": [ + 135, + 138, + 479, + 267 + ], + "type": "table", + "html": "
User StudyDynamic EffectGeneralExpert
DS-F +SparseGlyph +SparseT-Diff +SparseAnimate DiffLavieOursDS-F +SparseGlyph +SparseT-Diff +SparseAnimate DiffLavieOurs
Study 1Entrance3.3843.6873.8912.4063.1534.4783.4814.0254.1311.0193.0815.263
Emphasis3.5623.7134.1192.5632.7254.3193.6254.1754.2251.0502.6255.300
Motion2.3634.1384.1133.3382.3754.6751.7003.5504.0004.0002.0005.750
Average3.3183.7133.9232.5183.0364.4943.3353.9954.0901.3252.9055.350
Study 2Entrance3.5143.7753.9412.4133.2704.0883.6504.0814.2441.0193.1064.900
Emphasis3.5313.8814.0752.2693.0504.1943.7254.1504.0501.0252.6755.375
Motion2.8383.9503.8753.2502.8504.2382.1504.0004.4503.4501.9505.000
Average3.4393.7993.9482.4603.2204.1353.4854.0604.2551.2652.9704.995
Study 3Entrance0.6280.6200.6520.0630.1690.7250.8440.8560.8870.0190.1380.969
Emphasis0.5810.6060.6380.0880.1060.7810.7500.7000.9000.0000.0001.000
Motion0.0750.7130.7500.3620.2880.7810.0000.9000.9000.5000.2501.000
Average0.5530.6260.6580.0960.1740.7440.7300.8350.8800.0650.1350.975
", + "image_path": "83d5fafd730e51519212aba942becefc4c64c5b8f48fc510576e35d818066b56.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 294, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 480, + 365 + ], + "type": "text", + "content": "instruction with respect to either the whole caption or the letter's movement. Here, we need to know how to perform our Kinecty model. Figure 6 visualizes the attention map between the caption and features from the hidden state on the corresponding images. Thanks to the cross-attention between the word caption and the noisy latent, and the delimiter to separate the words, our model successfully embeds features for each letter in the noise." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 369, + 481, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 481, + 430 + ], + "type": "text", + "content": "Lastly, to show the generality of our KineTy model, we conduct an additional experiment. After generating an initial kinetic typography, our KineTy model allows users to modify its style and motions such as the content color, background, font and motion, and even to combine them. As demonstrated in Fig. 7, ours can render the modified outcomes corresponding to the additional captions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 434, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 529 + ], + "type": "text", + "content": "User Study. To assess the practical utility of our results in the field of kinetic typography, we conduct a user study using Amazon MTurk. We ask 20 questionnaires to 50 participants. Since their familiarity of typography can vary significantly, we divide them into two groups: experts and non-experts, consisting of 10 and 40 individuals, respectively. The study involves three sections: (1) caption alignment; (2) kinetic typography suitability; (3) Word readability. We make two video clips by randomly choosing two words for each effect. In total, we make 20 clips for this user study." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "In the first study, we measure how well ours and the comparison methods generate videos that align with provided captions. The participants see 6 videos from different models with the same corresponding caption in a random order. They are asked to rank the videos from 1 to 6, and the rankings are subsequently converted into scores, ranging from 6 (Best alignment) to 1 (Worst alignment). The following study evaluates how proper the generated videos are for motion graphic applications. Since some participants might be unfamiliar with kinetic typography, all participants watch 4 example videos from online in advance. After that, they check six videos in a random order and rank them. Scores were assigned based on these rankings, from 6 (the most proper) to 1 (the least proper). The last study assesses how readable the outcomes are. After watching 6 videos, the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 297, + 91, + 447, + 102 + ], + "type": "text", + "content": "Kinetic Typography Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 127, + 479, + 186 + ], + "blocks": [ + { + "bbox": [ + 153, + 114, + 459, + 125 + ], + "lines": [ + { + "bbox": [ + 153, + 114, + 459, + 125 + ], + "spans": [ + { + "bbox": [ + 153, + 114, + 459, + 125 + ], + "type": "text", + "content": "Table 4: The result of ablation study. Bold: Best, Underline: Second-best." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 127, + 479, + 186 + ], + "lines": [ + { + "bbox": [ + 135, + 127, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 135, + 127, + 479, + 186 + ], + "type": "table", + "html": "
ModelEntranceEmphasisMotionAverage
FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑FVD↓IS↓CLIP↑OCR↑
Ours147.41.790.870.71177.521.750.880.5436.161.620.940.98125.901.770.880.76
-Effect separation789.82.750.740.49880.12.880.750.49607.12.540.820.23729.82.730.750.48
-Cword239.11.770.860.70247.11.730.870.5462.71.600.950.95201.81.740.870.74
-Lglph188.91.770.860.70226.81.750.880.5550.11.590.930.98160.81.750.880.75
", + "image_path": "71eef853e8a15d196a592c730a7bf4333456d22cbea3dc8cee47c983ae0cd90e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 481, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 481, + 228 + ], + "type": "text", + "content": "participants are needed to vote for videos with readable text contents. Of course, multiple voting is available." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 229, + 481, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 229, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 229, + 481, + 289 + ], + "type": "text", + "content": "The performance of each method is measured based on scores for the first two studies, and on the selection ratio for third one. As demonstrated in Tab. 3, our approach shows the promising performance across all studies. The results highlight our model's practical usability in creating kinetic typography, especially for the domain experts who give our model the highest scores." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 308, + 238, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 238, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 238, + 321 + ], + "type": "text", + "content": "5.3 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 331, + 481, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 481, + 402 + ], + "type": "text", + "content": "Without caption separation. First of all, we train our network without dividing static and dynamic caption. The performances drop significantly for every metrics in Tab. 4. Since we train the text-to-image backbone using only static captions, dynamic captions are considered as noisy input. Similarly, the motion module, used to train dynamic motions, is effective when the simple motion guidance is given, compared to using static and dynamic caption together." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 403, + 481, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 481, + 461 + ], + "type": "text", + "content": "Text contents incorporation. We observe that text-to-image diffusion models often fail to catch up on text contents in the prompt when a long description is given. We thus use " + }, + { + "bbox": [ + 130, + 403, + 481, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{word}" + }, + { + "bbox": [ + 130, + 403, + 481, + 461 + ], + "type": "text", + "content": " as an additional input to give a strong attention to each letter. As shown in Tab. 4, without " + }, + { + "bbox": [ + 130, + 403, + 481, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{word}" + }, + { + "bbox": [ + 130, + 403, + 481, + 461 + ], + "type": "text", + "content": ", the performance degradation is observed as expected." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 463, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 481, + 498 + ], + "type": "text", + "content": "Without glyph loss. Lastly, we evaluate the performance without " + }, + { + "bbox": [ + 130, + 463, + 481, + 498 + ], + "type": "inline_equation", + "content": "L_{glyph}" + }, + { + "bbox": [ + 130, + 463, + 481, + 498 + ], + "type": "text", + "content": ". The inferior FVD scores come from the blur effect at the edge of the word along with color bleeding effects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 519, + 220, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 519, + 220, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 220, + 532 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we propose a generative kinetic typography model, named KineTy. To achieve this, we first build a large-scale kinetic typography dataset by collecting 584 templates from professional designers, and use them to train our diffusion-based network. To effectively handle the text prompt, we split it into static and dynamic captions. They are used to directly guide the spatial and temporal cross-attention for better appearance and motion effects, respectively. Our glyph loss also strengthens the legible and artistic letter generation. We lastly demonstrate that our KineTy produces the visually compelling and semantically coherent kinetic typography videos. The extensive user studies further validate the effectiveness of our model with respect to its practical utility." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "S. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "Acknowledgements This research was supported by 'Project for Science and Technology Opens the Future of the Region' program through the INNOPOLIS FOUNDATION funded by Ministry of Science and ICT (Project Number: 2022-DD-UP-0312), the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(RS-2024-00338439), GIST-MIT Research Collaboration grant funded by the GIST in 2024, and the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-01842, Artificial Intelligence Graduate School Program (GIST))." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 241, + 197, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 241, + 197, + 253 + ], + "spans": [ + { + "bbox": [ + 133, + 241, + 197, + 253 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 263, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 263, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 263, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 263, + 480, + 285 + ], + "type": "text", + "content": "1. Anderson, D., Shamir, A., Fried, O.: Neural font rendering. arXiv preprint arXiv:2211.14802 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 285, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 481, + 319 + ], + "type": "text", + "content": "2. Azadi, S., Fisher, M., Kim, V.G., Wang, Z., Shechtman, E., Darrell, T.: Multicontent gan for few-shot font style transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 319, + 481, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 319, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 138, + 319, + 481, + 350 + ], + "type": "text", + "content": "3. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 351, + 480, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 351, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 138, + 351, + 480, + 372 + ], + "type": "text", + "content": "4. Barratt, S., Sharma, R.: A note on the inception score. arXiv preprint arXiv:1801.01973 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 373, + 480, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 373, + 480, + 394 + ], + "spans": [ + { + "bbox": [ + 138, + 373, + 480, + 394 + ], + "type": "text", + "content": "5. Berio, D., Leymarie, F.F., Asente, P., Echevarria, J.: Strokestyles: Stroke-based segmentation and stylization of fonts. ACM Transactions on Graphics (TOG) (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 395, + 481, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 395, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 138, + 395, + 481, + 437 + ], + "type": "text", + "content": "6. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 437, + 481, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 481, + 480 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 481, + 480 + ], + "type": "text", + "content": "7. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 481, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 481, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 481, + 480, + 503 + ], + "type": "text", + "content": "8. Campbell, N.D., Kautz, J.: Learning a manifold of fonts. ACM Transactions on Graphics (TOG) (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 503, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 480, + 535 + ], + "type": "text", + "content": "9. Chen, C.H., Liu, Y.T., Zhang, Z., Guo, Y.C., Zhang, S.H.: Joint implicit neural representation for high-fidelity and compact vector fonts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 134, + 536, + 480, + 567 + ], + "type": "text", + "content": "10. Chen, J., Huang, Y., Lv, T., Cui, L., Chen, Q., Wei, F.: Textdiffuser-2: Unleashing the power of language models for text rendering. arXiv preprint arXiv:2311.16465 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 568, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 568, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 134, + 568, + 480, + 600 + ], + "type": "text", + "content": "11. Chen, J., Huang, Y., Lv, T., Cui, L., Chen, Q., Wei, F.: Textdiffuser: Diffusion models as text painters. Proceedings of the Neural Information Processing Systems (NeurIPS) (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "text", + "content": "12. Esser, P., Chiu, J., Atighehchian, P., Granskog, J., Germanidis, A.: Structure and content-guided video synthesis with diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "text", + "content": "13. Ford, S., Forlizzi, J., Ishizaki, S.: Kinetic typography: issues in time-based presentation of text. In: CHI'97 extended abstracts on Human factors in computing systems, pp. 269-270. ACM Digital Library (1997)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 297, + 91, + 447, + 102 + ], + "type": "text", + "content": "Kinetic Typography Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "14. Fridsma, L., Gyncild, B.: Adobe After Effects CC Classroom in a Book. Adobe Press (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 480, + 171 + ], + "type": "text", + "content": "15. Fu, B., He, J., Wang, J., Qiao, Y.: Neural transformation fields for arbitrary-styled font generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 214 + ], + "type": "text", + "content": "16. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 480, + 246 + ], + "type": "text", + "content": "17. Guo, Y., Yang, C., Rao, A., Agrawala, M., Lin, D., Dai, B.: Sparsectrl: Adding sparse controls to text-to-video diffusion models. arXiv preprint arXiv:2311.16933 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 246, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 246, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 133, + 246, + 480, + 278 + ], + "type": "text", + "content": "18. Guo, Y., Yang, C., Rao, A., Wang, Y., Qiao, Y., Lin, D., Dai, B.: AnimateDiff: Animate your personalized text-to-image diffusion models without specific tuning arXiv preprint arXiv:2307.04725 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 278, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 278, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 133, + 278, + 480, + 311 + ], + "type": "text", + "content": "19. Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 311, + 480, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 343 + ], + "type": "text", + "content": "20. Ho, J., Chan, W., Saharia, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 343, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 374 + ], + "type": "text", + "content": "21. Hu, Y., Luo, C., Chen, Z.: Make it move: controllable image-to-video generation with text descriptions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "text", + "content": "22. Huang, Y., Xue, H., Liu, B., Lu, Y.: Unifying multimodal transformer for bidirectional image and text generation. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 1138-1147 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "type": "text", + "content": "23. Iluz, S., Vinker, Y., Hertz, A., Berio, D., Cohen-Or, D., Shamir, A.: Word-as-image for semantic typography. ACM Transactions on Graphics (TOG) (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 429, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 471 + ], + "type": "text", + "content": "24. Jahanian, A., Liu, J., Lin, Q., Tretter, D., O'Brien-Strain, E., Lee, S.C., Lyons, N., Allebach, J.: Recommendation system for automatic design of magazine covers In: Proceedings of the 2013 International Conference on Intelligent User Interfaces (IUI) (2013)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 471, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 480, + 504 + ], + "type": "text", + "content": "25. Jia, P., Li, C., Liu, Z., Shen, Y., Chen, X., Yuan, Y., Zheng, Y., Chen, D., Li, J., Xie, X., et al.: Cole: A hierarchical generation framework for graphic design. arXiv preprint arXiv:2311.16974 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "26. Kato, J., Nakano, T., Goto, M.: Textalive: Integrated design environment for kinetic typography. In: Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems (CHI) (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "text", + "content": "27. Krishnan, P., Kovvuri, R., Pang, G., Vassilev, B., Hassner, T.: Textstylebrush: Transfer of text aesthetics from a single example. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "text", + "content": "28. Lee, J.C., Forlizzi, J., Hudson, S.E.: The kinetic typography engine: an extensible system for animating expressive text. In: Proceedings of the 15th annual ACM symposium on User Interface Software and Technology (UIST) (2002)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "text", + "content": "29. Lee, J., Jun, S., Forlizzi, J., Hudson, S.E.: Using kinetic typography to convey emotion in text-based interpersonal communication. In: Proceedings of the 6th conference on Designing Interactive systems (DIS) (2006)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "30. Li, C., Taniguchi, Y., Lu, M., Konomi, S.: Few-shot font style transfer between different languages. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "S. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "31. Li, X., Wu, L., Wang, C., Meng, L., Meng, X.: Compositional zero-shot artistic font synthesis. In: Proceedings of the International Joint Conferences on Artificial Intelligence (IJCAI) (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "type": "text", + "content": "32. Liu, X., Meng, G., Chang, J., Hu, R., Xiang, S., Pan, C.: Decoupled representation learning for character glyph synthesis. IEEE Transactions on Multimedia (TMM) (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 181, + 482, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 181, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 181, + 482, + 214 + ], + "type": "text", + "content": "33. Liu, Y.T., Guo, Y.C., Li, Y.X., Wang, C., Zhang, S.H.: Learning implicit glyph shape representation. IEEE Transactions on Visualization and Computer Graphics (TVCG) (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 214, + 482, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 257 + ], + "type": "text", + "content": "34. Liu, Y.T., Zhang, Z., Guo, Y.C., Fisher, M., Wang, Z., Zhang, S.H.: Dualvector: Unsupervised vector font synthesis with dual-part representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 257, + 482, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 278 + ], + "type": "text", + "content": "35. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: Proceedings of the International Conference on Learning Representations (ICLR) (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 278, + 482, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 322 + ], + "type": "text", + "content": "36. Luo, Z., Chen, D., Zhang, Y., Huang, Y., Wang, L., Shen, Y., Zhao, D., Zhou, J., Tan, T.: Videofusion: Decomposed diffusion models for high-quality video generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 322, + 482, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 354 + ], + "type": "text", + "content": "37. Men, Y., Lian, Z., Tang, Y., Xiao, J.: Dyntypo: Example-based dynamic text effects transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 354, + 482, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 386 + ], + "type": "text", + "content": "38. Minakuchi, M., Tanaka, K.: Automatic kinetic typography composer. In: Proceedings of the ACM SIGCHI International Conference on Advances in Computer Entertainment Technology (ACE) (2005)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 386, + 482, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 418 + ], + "type": "text", + "content": "39. Mu, X., Chen, L., Chen, B., Gu, S., Bao, J., Chen, D., Li, J., Yuan, Y.: Fontstudio: Shape-adaptive diffusion model for coherent and consistent font effect generation. arXiv preprint arXiv:2406.08392 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 418, + 482, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 439 + ], + "type": "text", + "content": "40. Nagata, Y., Iwana, B.K., Uchida, S.: Contour completion by transformers and its application to vector font data. arXiv preprint arXiv:2304.13988 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 439, + 482, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 460 + ], + "type": "text", + "content": "41. OpenAI: GPT-4V(ision) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 460, + 482, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 493 + ], + "type": "text", + "content": "42. Pan, W., Zhu, A., Zhou, X., Iwana, B.K., Li, S.: Few shot font generation via transferring similarity guided global style and quantization local style. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 493, + 482, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 482, + 525 + ], + "type": "text", + "content": "43. Qu, Y., Tan, Q., Xie, H., Xu, J., Wang, Y., Zhang, Y.: Exploring stroke-level modifications for scene text editing. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 525, + 482, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 568 + ], + "type": "text", + "content": "44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proceedings of the International Conference on Machine Learning (ICML) (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "type": "text", + "content": "45. Reddy, P., Zhang, Z., Wang, Z., Fisher, M., Jin, H., Mitra, N.: A multi-implicit neural representation for fonts. Proceedings of the Neural Information Processing Systems (NeurIPS) (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "text", + "content": "46. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 633, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 633, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 633, + 482, + 665 + ], + "type": "text", + "content": "47. Shimoda, W., Haraguchi, D., Uchida, S., Yamaguchi, K.: De-rendering stylized texts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 297, + 91, + 447, + 102 + ], + "type": "text", + "content": "Kinetic Typography Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "48. Shimoda, W., Haraguchi, D., Uchida, S., Yamaguchi, K.: Towards diverse and consistent typography generation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 192 + ], + "type": "text", + "content": "49. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. In: Proceedings of the International Conference on Learning Representations (ICLR) (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 192, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 481, + 214 + ], + "type": "text", + "content": "50. Smith, J., Team, A.C.: Adobe After Effects CS6 Digital Classroom. John Wiley & Sons (2012)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 481, + 246 + ], + "type": "text", + "content": "51. Subramanian, J., Chordia, V., Bart, E., Fang, S., Guan, K., Bala, R., et al.: Strive: Scene text replacement in videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "text", + "content": "52. Tanveer, M., Wang, Y., Mahdavi-Amiri, A., Zhang, H.: Ds-fusion: Artistic typography via discriminated and stylized diffusion. arXiv preprint arXiv:2303.09604 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 278, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 322 + ], + "type": "text", + "content": "53. Thamizharasan, V., Liu, D., Agarwal, S., Fisher, M., Gharbi, M., Wang, O., Jacobson, A., Kalogerakis, E.: Vecfusion: Vector font generation with diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 322, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 481, + 354 + ], + "type": "text", + "content": "54. Tuo, Y., Xiang, W., He, J.Y., Geng, Y., Xie, X.: Anytext: Multilingual visual text generation and editing. In: Proceedings of the International Conference on Learning Representations (ICLR) (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "text", + "content": "55. Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "56. Wang, C., Wu, L., Liu, X., Li, X., Meng, L., Meng, X.: Anything to glyph: Artistic font synthesis via text-to-image diffusion model. In: SIGGRAPH Asia 2023 Conference Papers (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 418, + 481, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 451 + ], + "type": "text", + "content": "57. Wang, C., Zhou, M., Ge, T., Jiang, Y., Bao, H., Xu, W.: Cf-font: Content fusion for few-shot font generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "type": "text", + "content": "58. Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., Yang, P., et al.: Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 483, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 481, + 515 + ], + "type": "text", + "content": "59. Wang, Y., Ye, Y., Mao, Y., Yu, Y., Song, Y.: Self-supervised scene text segmentation with object-centric layered representations augmented by text regions. In: Proceedings of the 30th ACM International Conference on Multimedia (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 515, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 481, + 536 + ], + "type": "text", + "content": "60. Wang, Y., Lian, Z.: Deepvecfont: Synthesizing high-quality vector fonts via dual-modality learning. ACM Transactions on Graphics (TOG) (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "text", + "content": "61. Wang, Y., Pu, G., Luo, W., Wang, Y., Xiong, P., Kang, H., Lian, Z.: Aesthetic text logo synthesis via content-aware layout inferring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "62. Wang, Y., Wang, Y., Yu, L., Zhu, Y., Lian, Z.: Deepvecfont-v2: Exploiting transformers to synthesize vector fonts with higher quality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "63. Wong, Y.Y.: Temporal typography: a proposal to enrich written expression. In: Proceedings of the Conference Companion on Human Factors in Computing Systems (CHI) (1996)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "64. Xia, Z., Xiong, B., Lian, Z.: Vecfontsf: Learning to reconstruct and synthesize high-quality vector fonts via signed distance functions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "S. Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 555 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "65. Xie, L., Shu, X., Su, J.C., Wang, Y., Chen, S., Qu, H.: Creating emordle: Animating word cloud for emotion expression. IEEE Transactions on Visualization and Computer Graphics (TVCG) (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "66. Xie, L., Zhou, Z., Yu, K., Wang, Y., Qu, H., Chen, S.: Wakey-wakey: animate text by mimicking characters in a gif. In: Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 226 + ], + "type": "text", + "content": "67. Xu, C., Zhou, M., Ge, T., Jiang, Y., Xu, W.: Unsupervised domain adaption with pixel-level discriminator for image-aware layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "type": "text", + "content": "68. Xu, X., Zhang, Z., Wang, Z., Price, B., Wang, Z., Shi, H.: Rethinking text segmentation: A novel dataset and a text-specific refinement approach. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 270, + 482, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 303 + ], + "type": "text", + "content": "69. Yang, Q., Huang, J., Lin, W.: Swaptext: Image based texts transfer in scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 303, + 482, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 336 + ], + "type": "text", + "content": "70. Yang, S., Liu, J., Lian, Z., Guo, Z.: Awesome typography: Statistics-based text effects transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 336, + 482, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 357 + ], + "type": "text", + "content": "71. Yang, S., Wang, W., Liu, J.: Te141k: artistic text benchmark for text effect transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "type": "text", + "content": "72. Yang, S., Wang, Z., Liu, J.: Shape-matching " + }, + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\mathrm{gan} + +" + }, + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "type": "text", + "content": ": Scale controllable dynamic artistic text style transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 390, + 482, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 423 + ], + "type": "text", + "content": "73. Yang, S., Wang, Z., Wang, Z., Xu, N., Liu, J., Guo, Z.: Controllable artistic text style transfer via shape-matching gan. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 423, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 456 + ], + "type": "text", + "content": "74. Yang, Y., Gui, D., Yuan, Y., Liang, W., Ding, H., Hu, H., Chen, K.: Glyphcontrol: Glyph conditional control for visual text generation. In: Proceedings of the Neural Information Processing Systems (NeurIPS) (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 456, + 482, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 490 + ], + "type": "text", + "content": "75. Yang, Z., Peng, D., Kong, Y., Zhang, Y., Yao, C., Jin, L.: Fontdiffuser: One-shot font generation via denoising diffusion with multi-scale content aggregation and style contrastive learning. arXiv preprint arXiv:2312.12142 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 490, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 522 + ], + "type": "text", + "content": "76. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "text", + "content": "77. Zhang, S., Ma, J., Wu, J., Ritchie, D., Agrawala, M.: Editing motion graphics video via motion vectorization and transformation. ACM Transactions on Graphics (TOG) (2023)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 297, + 91, + 447, + 102 + ], + "type": "text", + "content": "Kinetic Typography Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation/41d0edfc-3cb8-4f88-8e86-7449093dd76b_content_list.json b/2024/Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation/41d0edfc-3cb8-4f88-8e86-7449093dd76b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5d59b69f867abdd8bc69f5213ce4a2aa8bcb5163 --- /dev/null +++ b/2024/Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation/41d0edfc-3cb8-4f88-8e86-7449093dd76b_content_list.json @@ -0,0 +1,1616 @@ +[ + { + "type": "text", + "text": "Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation", + "text_level": 1, + "bbox": [ + 227, + 140, + 779, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tao Chen $^{1}$ , Xiruo Jiang $^{1}$ , Gensheng Pei $^{1}$ , Zeren Sun $^{1}$ , Yucheng Wang $^{2}$ , and Yazhou Yao $^{1(\\text{忍})}$", + "bbox": [ + 302, + 234, + 700, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Nanjing University of Science and Technology, Nanjing 210094, China {taochen,jiangxiruo622,peigsh,zerens,yazhou.yao}@njust.edu.cn $^{2}$ Horizon Robotics, Beijing 100089, China yucheng.wang@horizon.cc", + "bbox": [ + 261, + 277, + 743, + 335 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Though adversarial erasing has prevailed in weakly supervised semantic segmentation to help activate integral object regions, existing approaches still suffer from the dilemma of under-activation and over-expansion due to the difficulty in determining when to stop erasing. In this paper, we propose a Knowledge Transfer with Simulated InterImage Erasing (KTSE) approach for weakly supervised semantic segmentation to alleviate the above problem. In contrast to existing erasing-based methods that remove the discriminative part for more object discovery, we propose a simulated inter-image erasing scenario to weaken the original activation by introducing extra object information. Then, object knowledge is transferred from the anchor image to the consequent less activated localization map to strengthen network localization ability. Considering the adopted bidirectional alignment will also weaken the anchor image activation if appropriate constraints are missing, we propose a self-supervised regularization module to maintain the reliable activation in discriminative regions and improve the inter-class object boundary recognition for complex images with multiple categories of objects. In addition, we resort to intra-image erasing and propose a multi-granularity alignment module to gently enlarge the object activation to boost the object knowledge transfer. Extensive experiments and ablation studies on PAS-CAL VOC 2012 and COCO datasets demonstrate the superiority of our proposed approach. Codes and models are available at https://nust-machine-intelligence-laboratory.github.io/project-KTSE.", + "bbox": [ + 261, + 373, + 743, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Weakly Supervised Learning $\\cdot$ Semantic Segmentation $\\cdot$ InterImage Erasing", + "bbox": [ + 261, + 705, + 750, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 761, + 375, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With huge progress in the era of deep learning, semantic segmentation has been widely applied in fields like autonomous driving and image editing [5,41]. Since deep learning thus far is data-driven, the model training typically involves a large", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/58f2061d0629499ba4986e9c776a3ac1bfa5f966e43e207d7b25862355899a78.jpg", + "image_caption": [ + "(a) Previous Adversarial Erasing" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 496, + 186 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9de728cea173eb22073b8f9b7700208595d731074360f0bde3e7937f5d92c335.jpg", + "image_caption": [ + "(c) Result CAM vs Ground Truth", + "Fig. 1: (a) Previous adversarial erasing-based approaches typically suffer from the over-expansion problem, which is hard to constrain. (b) Different from their information removal strategy, we propose to add extra object knowledge from a paired image to weaken the current object activation. The localization ability of the network is then enhanced by improving the consequent less activated attention map through learning from the object knowledge of the anchor branch. (C) Result comparison." + ], + "image_footnote": [], + "bbox": [ + 217, + 199, + 496, + 241 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/577d128e86287d349d5eb7649899060911529a0aafb216ca283234d19d0a2355.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 143, + 785, + 191 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4bea42c8c3cc3ec67b863124745c05367d582458fc549d5fa261ebe703ba6e73.jpg", + "image_caption": [ + "(b) Our Simulated Inter-Image Erasing" + ], + "image_footnote": [], + "bbox": [ + 500, + 193, + 785, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "abundance of labeled images. However, collecting accurate pixel-level annotation for segmentation tasks is highly labor-intensive and time-consuming [3, 9, 16]. Therefore, as a promising direction to alleviate such annotation burden, weakly supervised learning has attracted the attention of many researchers. In this paper, we focus on the weakly supervised semantic segmentation (WSSS) under the supervision of image-level labels.", + "bbox": [ + 212, + 381, + 787, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The recent advance of WSSS typically follows the three-step pipeline of: 1) transforming image tags into pixel-level coarse labels, 2) refining the pseudo labels, and 3) training the final segmentation model with the refined labels. Speaking of segmentation label generation, the technique of class activation map (CAM) [71] has been the de facto paradigm for object localization with image-level labels. Unfortunately, the naive CAM can only highlight the most discriminative area of objects and thus its small and sparse activation will lead to incomplete object mining. Since the first step of seed generation is the foundation of later processes, numerous works have been developed to expand CAM activation for high-quality pseudo labels. Among them, adversarial erasing is one of the mainstream approaches, aiming to discover new and complementary object regions by masking the currently detected area in an adversarial manner [60,67,69]. For example, the pioneering work of AE-PSL [60] performs the erasing multiple times to progressively discover increasingly activated areas. The recent work of AEFT [67] proposes to learn the concept of erasing with the triplet loss between the input image, erased image, and negatively sampled image. Though improvement is obtained, existing adversarial erasing-based methods still suffer from the difficulty in determining when to stop erasing: excessive removal will lead to over-expansion and insufficient erasure will result in under-activation.", + "bbox": [ + 212, + 474, + 787, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a Knowledge Transfer with Simulated Inter-Image Erasing (KTSE) approach for WSSS to alleviate the problem mentioned above. Existing erasing-based methods typically mask the most discriminative object part to force the classifier into seeking other regions. Based on the observation that information removal leads to more activation (which might also incur the", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "T.Chen et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "over-expansion illustrated in Fig. 1 (a)), we act in a diametrically opposite way. Specifically, as shown in Fig. 1 (b), we propose to simulate an inter-image erasing (SIE) scenario, introducing extra object information by concatenating a paired image and rendering the anchor image as the erased one. We then strengthen the object localization ability of the network by transferring the object knowledge from the anchor CAM to the consequent less activated localization map. Intuitively, our knowledge transfer with simulated inter-image erasing will not cause the over-expansion problem if the object in the anchor branch is well-discovered.", + "bbox": [ + 212, + 146, + 782, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, since the above knowledge transfer is bidirectional, it will also weaken object mining in anchor branch when learning from the sparse activation of simulated branch. Therefore, as shown in the right of Fig. 2, we propose a self-supervised regularization module for the anchor CAM feature to maintain its reliable activation in discriminative regions. Specifically, we first generate pseudo labels from the CAM feature by locating confident foreground and background area, which are used to directly supervise the learning of CAM features. However, extracted labels of the foreground are quite noisy in complex images with multiple categories of objects due to the inter-class confusion. Therefore, we also design an inter-class loss to implicitly encourage activation consistency in complex images to improve recognition of the inter-class object boundary.", + "bbox": [ + 212, + 268, + 784, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Though our proposed self-supervised regularization can effectively constrain and facilitate the knowledge transfer in our simulated inter-image erasing module, the improvement of the network's object localization ability will be severely limited by the quality of anchor CAM feature which is usually under-activated. Therefore, we resort to the traditional intra-image erasing and propose a multi-granularity alignment (MGA) module to gently expand the object activation without introducing much background noise. Specifically, as shown in the bottom of Fig. 2, we first leverage the image-level global alignment to distill the soft object confidence from the anchor branch to the masked one for object activation enlargement. Then, we leverage a pixel-level local alignment to transfer the newly discovered object information back to the anchor branch. With the gentle and constrained activation enlargement, the proposed multi-granularity alignment module further boosts the performance of our simulated inter-image erasing. As can be seen in Fig. 2, our SIE and MGA are designed in a symmetric way that play the adversarial role of expanding and constraining the anchor image activation respectively for mining complete and compact object regions.", + "bbox": [ + 212, + 438, + 782, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 238, + 681, + 591, + 695 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) We propose a knowledge transfer with simulated inter-image erasing approach for weakly supervised semantic segmentation to alleviate the over-expansion problem of existing adversarial erasing-based methods.", + "(2) A self-supervised regularization module is proposed to constrain the knowledge transfer for maintaining reliable activation and improving recognition of the inter-class object boundary.", + "(3) We propose a multi-granularity alignment module to gently enlarge the object activation via image-level global alignment and pixel-level local alignment, boosting the knowledge transfer with simulated inter-image erasing." + ], + "bbox": [ + 214, + 699, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS", + "bbox": [ + 285, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Weakly Supervised Semantic Segmentation", + "text_level": 1, + "bbox": [ + 215, + 176, + 617, + 193 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Weakly supervised semantic segmentation (WSSS) is the task that aims to obtain a high-performance segmentation model by learning only from weak labels. Compared with bounding boxes [16, 25, 27, 53], scribbles [37, 56] and points [3], the image-level label [1, 4, 10, 12, 24, 34, 42, 50, 66, 73] is the cheapest weak supervision to collect, making it the most popular annotation format for WSSS. The recent advance of WSSS [1, 24, 34, 66, 73] typically relies on the class activation technique of CAM [71] to locate the target object, which helps transform the image-level label to pixel-level dense annotation for segmentation network training. Since the object area highlighted in CAM is usually small and sparse, enlarging the CAM activation becomes the focus of the WSSS task. For example, RDC [61] and DRS [28] propagate the object information to less discriminative regions via dilated convolution with varied rates and attention suppression of most discriminative areas, respectively. Popular contrastive learning [18, 48, 73] and self-supervised approaches [8, 51, 59] are also applied to facilitate object mining by improving representation learning. Moreover, cross-image information [21, 57, 72] is exploited to help the classifier discover more object patterns to obtain consistent and integral target regions. Besides attention expansion, Ahn and Kwak propose AffinityNet [2] for CAM refinement, which propagates semantic affinity via a random walk to recover delicate shapes of objects. IRN [1] is then proposed to exploit class boundary detection for discovering the entire instance areas with more accurate boundaries.", + "bbox": [ + 212, + 202, + 787, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Erasing-based Approaches", + "text_level": 1, + "bbox": [ + 215, + 542, + 478, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering the original CAM technique usually leads to suboptimal performance due to failure to localize all object parts, erasing-based approaches are commonly employed to expand the highlighted region [30, 52, 60, 67, 69, 70]. They mask areas in a training image to force the network to seek other relevant parts. While direct erasing methods like Hide-and-Seek [52] hide image patches randomly, adversarial erasing-based approaches mask the most discriminative regions with the attention information, demonstrating more promising activation expansion potential. The first adversarial erasing-based work for WSSS is proposed by Wei et al. [60], which progressively discovers new object discriminative regions by erasing the currently activated area. Such a heuristic erasing strategy then attracts the attention of many researchers. For example, ACoL [69] proposes adversarial complementary learning with two adversary classifiers. Recently, Kweon et al. propose to apply the adversarial erasing framework to exploit the potential of the pre-trained classifier [30] and learn the concept of erasing with the triplet loss [67]. While adversarial erasing approaches achieve great success in enlarging the CAM activation, they tend to suffer from the over-expansion problem due to difficulty in determining when to stop erasing. Unlike the information removal strategy adopted in these approaches, we propose a simulated", + "bbox": [ + 212, + 566, + 787, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "T.Chen et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/de5f5feb92fccd6629646b1a111135f69f07405507455e38d47c9b38f7d5dac3.jpg", + "image_caption": [ + "Fig. 2: The architecture of our proposed approach. We propose a simulated inter-image erasing (SIE) scenario where extra object information is introduced from another paired image. We then strengthen the object localization ability of the network by improving the consequent less activated localization map through learning object knowledge from the anchor image. A self-supervised regularization (SSR) module is also proposed to avoid weakening the anchor activation due to bidirectional alignment and improve the inter-class object boundary recognition for complex images. In addition, we propose a multi-granularity alignment (MGA) module to gently enlarge the object activation to further boost the object knowledge transfer." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "inter-image erasing scenario to weaken the activation by adding extra discriminative object information. The localization ability of network is then enhanced by improving the consequent less activated CAM through knowledge transfer.", + "bbox": [ + 212, + 482, + 787, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 550, + 330, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we propose a knowledge transfer with simulated inter-image erasing approach for weakly supervised semantic segmentation to alleviate the over-expansion and under-activation problems of existing adversarial erasing-based methods. Our framework is illustrated in Fig. 2. As demonstrated in the anchor branch, we train a classification network with the given image-level weak labels, which consists of a backbone feature extractor and a pooling classification head. In contrast to existing adversarial erasing-based approaches that expand the activation through masking the most discriminative object part, as shown in the top-left of Fig. 2, we propose to simulate an inter-image erasing scenario where extra object information is introduced through concatenating a paired image. Object knowledge is then transferred from the anchor image to the consequent less activated localization map for strengthening the object localization ability of network. We also propose a self-supervised regularization module to avoid weakening the anchor activation due to bidirectional alignment and improve the inter-class object boundary recognition for complex multi-category images. In addition, we propose a multi-granularity alignment module to gently enlarge the object activation for boosting the object knowledge transfer.", + "bbox": [ + 212, + 583, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS", + "bbox": [ + 287, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 CAM Generation", + "text_level": 1, + "bbox": [ + 215, + 146, + 405, + 159 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the architecture of classification network, we follow the previous erasing-based work ACoL [69] and remove the final fully-connected layer. We set the output channel of the backbone to $C + 1$ , where $C$ is the number of foreground categories and 1 is added for the background. We can thus directly generate object localization maps from the class-aware CAM features $F$ in the forward pass to facilitate the object knowledge transfer. To obtain CAM for each foreground class $c$ , we feed the attention map $F^c$ into a ReLU layer and then normalize it to the range from 0 to 1:", + "bbox": [ + 212, + 176, + 787, + 297 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nA ^ {c} = \\frac {\\operatorname {R e L U} \\left(F ^ {c}\\right)}{\\operatorname* {m a x} \\left(F ^ {c}\\right)}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 297, + 785, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To improve the quality of the CAM by capturing the global context and local details of the image, we adopt the Gated Pyramid Pooling (GPP) layer [67] as the final pooling head. With the classification logits $q^{c}$ generated from the pooling head, we train the classification network with the multi-label soft margin loss as follows:", + "bbox": [ + 214, + 338, + 787, + 412 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c l s} = - \\frac {1}{C} \\sum_ {c = 1} ^ {C} y ^ {c} \\log \\sigma \\left(q ^ {c}\\right) + \\left(1 - y ^ {c}\\right) \\log \\left[ 1 - \\sigma \\left(q ^ {c}\\right) \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 426, + 785, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\sigma (\\cdot)$ is the sigmoid function. $y^{c}$ is the image-level label for the $c$ -th class.", + "bbox": [ + 215, + 479, + 782, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Simulated Inter-Image Erasing", + "text_level": 1, + "bbox": [ + 215, + 521, + 516, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Due to the sparsity of the CAM activation, previous adversarial erasing-based methods typically enlarge the object activation by maintaining the classification confidence on masked images or features, where the most discriminative regions are erased. However, due to the absence of guidelines on when to stop erasing, these approaches easily incur over-expansion with excessive removal and might also still suffer from under-activation with insufficient erasure. Therefore, we propose a knowledge transfer with simulated inter-image erasing approach for weakly supervised semantic segmentation to alleviate the above problem.", + "bbox": [ + 212, + 551, + 787, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In contrast to existing erasing-based approaches that remove the discriminative part for more object mining, we propose a simulated inter-image erasing (SIE) scenario to enable the network to benefit from introducing extra discriminative object information. Specifically, as shown in the top-left of Fig. 2, a larger synthetic image, created by concatenating the anchor image and another paired one, plays the role of the original image needing masking. Correspondingly, the anchor image can be treated as the erased one where the paired image is masked. Due to the introduction of extra discriminative object information, fewer object regions will be highlighted in the anchor part of the synthetic image. Therefore, we then strengthen the object localization ability of the network by transferring the more integral object knowledge from the original anchor CAM to the", + "bbox": [ + 214, + 674, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "T.Chen et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "consequently less activated anchor part of the synthetic image. The knowledge transfer loss for the simulated inter-image erasing can be formulated as follows:", + "bbox": [ + 214, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {k t} = \\operatorname {R e L U} \\left(\\hat {F} _ {a} - \\hat {F} _ {s}\\right), \\quad \\text {w h e r e} \\hat {F} _ {a} = C F E \\left(F _ {a}, y\\right), \\hat {F} _ {s} = C F E \\left(F _ {s}, y\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 184, + 784, + 208 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, $F_{a}$ and $F_{s}$ denote the CAM features of the anchor and simulated branches, respectively. CFE represents class feature extraction, which extracts the feature channel related to the classes that exist in the image. The ReLU operation means that our alignment focuses on the more highlighted object region rather than the background. Note that no rotation or resize is performed for the concatenated image. We only introduce extra object information from the paired image and control other variables to guarantee the weakened activation for the anchor part. Intuitively, our knowledge transfer with simulated inter-image erasing will not cause the over-expansion problem when the object in the anchor branch is well-discovered. Our experiments demonstrate that such knowledge transfer can effectively strengthen the localization ability of the network to alleviate the under-activation, especially when multiple object instances co-exist in the image. Besides, benefiting from the bidirectional alignment, the anchor branch will also learn from the weakened activation of the simulated one and produce compact attention maps for small objects to mitigate the over-expansion problem.", + "bbox": [ + 214, + 213, + 787, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Self-Supervised Regularization", + "text_level": 1, + "bbox": [ + 215, + 459, + 511, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our simulated inter-image erasing aims to improve the object localization ability of the network by maintaining the high activation given extra discriminative object information. However, the adopted knowledge transfer is bidirectional, which will also weaken the anchor branch object mining when learning from the sparse activation of the simulated branch. Since simply cutting the gradient propagation of the anchor branch makes the training unstable, we propose a self-supervised regularization module for the anchor CAM feature to maintain its reliable activation in discriminative regions. Specifically, with the generated CAM feature $F_{a}$ and the corresponding CAM $A_{a}$ , we first leverage two thresholds $\\beta_{h} = 0.3$ and $\\beta_{l} = 0.15$ to locate the confident foreground and background as follows (the subscript $a$ is omitted for simplicity):", + "bbox": [ + 214, + 481, + 787, + 648 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {Y} _ {i, j} = \\left\\{ \\begin{array}{l l} \\operatorname {a r g m a x} \\left(F _ {i, j,:}\\right), & \\text {i f} \\max \\left(A _ {i, j,:}\\right) \\geq \\beta_ {h}, \\\\ 0, & \\text {i f} \\max \\left(A _ {i, j,:}\\right) \\leq \\beta_ {l}, \\\\ 2 5 5, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 656, + 784, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where 255 denote the ignored labels for uncertain regions and $\\arg\\max(\\cdot)$ extracts the semantic class with the maximum activation value. After refining the pseudo label $\\hat{Y}$ with the pixel-adaptive refinement module [47], we leverage it to directly supervise the learning of CAM features with the cross-entropy loss (coordinates $i, j$ are omitted):", + "bbox": [ + 214, + 718, + 787, + 794 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c e} = - \\sum_ {c = 0} ^ {C} \\hat {Y} _ {b g} ^ {c} \\log \\Gamma (F) ^ {c} + \\omega_ {f g} \\hat {Y} _ {f g} ^ {c} \\log \\Gamma (F) ^ {c}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 803, + 784, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\Gamma(\\cdot)$ denotes the softmax function. We also leverage the pseudo label $\\hat{Y}$ generated from the anchor branch to guide the learning of the simulated CAM feature with the same cross-entropy loss. However, we notice that the extracted labels of the foreground are quite noisy, especially for complex images with multiple categories of objects where the inter-class boundary is blurred. Here, we define images with objects of only one category as simple images and others containing two or more categories as complex ones. Therefore, we discount the influence of foreground label $\\hat{Y}_{fg}$ with the weight of $\\omega_{fg} = 0.0125$ . Besides, as shown in the right of Fig. 2, we design an inter-class loss to implicitly encourage the activation consistency for the multiple foreground classes in complex images as follows:", + "bbox": [ + 212, + 145, + 787, + 297 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {i n t e r}} = \\frac {\\sum \\left(V ^ {2 n d} - V ^ {\\max }\\right) \\cdot M ^ {f}}{\\left| M ^ {f} \\right|}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 309, + 785, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, $M^f$ denotes the foreground mask obtained with a threshold $\\beta_o = 0.2$ and $|\\cdot|$ return the value of $l_1$ -norm. $V^{max}$ and $V^{2nd}$ denotes the largest and second largest activation value along the channel dimension of the CAM feature $F_a$ , which can be obtained as follows:", + "bbox": [ + 214, + 354, + 787, + 415 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nV _ {i, j} ^ {m a x} = \\max (F _ {i, j,:}), \\quad F ^ {l e f t} = \\mathsf {r e m o v e} (F, V _ {m a x}), \\quad V _ {i, j} ^ {2 n d} = \\max (F _ {i, j,:} ^ {l e f t}). \\qquad (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 245, + 428, + 785, + 449 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Eq. (6) improves the exclusivity of the class activation for each pixel, leading to more accurate boundaries between foreground objects. Our proposed self-supervised regularization module treats the simple and complex images differently, intending to maintain the reliable activation of the main branch without deteriorating the inter-class boundary caused by inaccurate pseudo labels.", + "bbox": [ + 212, + 459, + 787, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Multi-Granularity Alignment", + "text_level": 1, + "bbox": [ + 214, + 556, + 503, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our proposed self-supervised regularization module can effectively constrain and boost the knowledge transfer in our simulated inter-image erasing module. However, the improvement of the network's object localization ability will be severely limited by the quality of the anchor CAM, which is usually under-activated. Therefore, it is natural to resort to the traditional intra-image erasing, which aims to expand the object area. Following AEFT [67], we leverage a threshold of 0.6 to mask the most discriminative region. To avoid introducing too much background noise, we propose a multi-granularity alignment module to gently expand the object activation, which encourages the CAM expansion of the erased image and then transfers the learned object knowledge back to the anchor branch. As shown in the bottom of Fig. 2, we first input the anchor feature $F_{a}$ and masked feature $F_{m}$ into a class feature extraction module to obtain the corresponding features of the existing target categories $(\\hat{F}_{a}$ and $\\hat{F}_{m}$ ). Following the experimental finding of AEFT [67] that rigid classification supervision [69] for the masked branch tends to trigger the over-expansion problem in adversarial erasing, we resort to the soft class confidence guidance from the anchor branch to the masked one. Specifically, we directly adopt a global average pooling (GAP) operation", + "bbox": [ + 212, + 583, + 787, + 843 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "T.Chen et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to obtain the final class confidence for each branch and an image-level global alignment loss can be formulated as follows:", + "bbox": [ + 214, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g l o b a l} = G A P \\left(R e L U \\left(\\hat {F} _ {a}\\right)\\right) - G A P \\left(R e L U \\left(\\hat {F} _ {m}\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 188, + 696, + 213 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} \\quad \\hat {F _ {m}} = C F E \\left(F _ {m}, y\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 217, + 782, + 233 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "$\\hat{F}_a$ and CFE have been defined in Eq. (3). Another difference with rigid classification supervision is that benefiting from the class feature extraction, our global alignment loss only focuses on the logit confidence for classes that exist in the image, enabling more effective and efficient gradient propagation. Unlike AEFT [67] that utilizes the feature space of the GPP layer as embedding space for loss construction, our experiments reveal that direct alignment with the CAM feature leads to more promising performance. With class information learned from the anchor branch, we further leverage a pixel-level local activation alignment to transfer the newly discovered object information from the erased image to guide the anchor branch learning.", + "bbox": [ + 214, + 246, + 784, + 398 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {l o c a l}} = \\operatorname {R e L U} \\left(\\hat {F _ {m}} - \\hat {F _ {a}}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 411, + 782, + 434 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Such pixel-level local alignment can also hinder the erased branch from activating the unwanted background area by learning from the lowly-activated regions of the anchor branch. With the gentle and constrained activation enlargement, the multi-granularity alignment module further boosts the performance of our proposed simulated inter-image erasing approach. As can be seen in Fig. 2, our proposed SIE and MGA branches are designed in a symmetric way that play the adversarial role of expanding and constraining the anchor image activation, respectively. Their synergically cooperation facilitates the convergence for mining complete and compact object regions.", + "bbox": [ + 214, + 445, + 784, + 580 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Training Objective", + "text_level": 1, + "bbox": [ + 214, + 603, + 416, + 618 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The overall training loss is as follows:", + "bbox": [ + 215, + 628, + 486, + 643 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c l s} + \\mathcal {L} _ {k t} + \\mathcal {L} _ {g l o b a l} + \\mathcal {L} _ {l o c a l} + \\mathcal {L} _ {c e} + \\lambda_ {i n t e r} \\mathcal {L} _ {i n t e r}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 657, + 782, + 672 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We empirically set $\\lambda_{inter} = 0.005$ as the hyperparameter that controls the weight of inter-class loss.", + "bbox": [ + 214, + 684, + 782, + 712 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 737, + 366, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets and Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 768, + 568, + 784 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate our approach on the PASCAL VOC 2012 [19] and COCO [38] datasets. The PASCAL VOC 2012 dataset contains 21 classes (20 object categories and the background) for semantic segmentation. It has 10,582 images for", + "bbox": [ + 214, + 794, + 782, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/4f066a10f473d2c107fc5a93cbbba97bb23699205c43b79edafe50cfda5a8279.jpg", + "table_caption": [ + "Table 1: Accuracy of pseudo-masks evaluated on PASCAL VOC 2012 training set." + ], + "table_footnote": [], + "table_body": "
MethodsSeedw/IRN [1]
IRN [1]CVPR1948.366.3
MBMNet [40]MM2050.266.8
CONTA [68]NIPS2048.867.9
AdvCAM [32]CVPR2155.669.9
RIB [31]NIPS2156.670.6
ReCAM [14]CVPR2256.670.5
ESOL [35]NIPS2253.668.7
CLIMS [63]CVPR2256.670.5
AEFT [67]ECCV2256.071.0
ACR [15]CVPR2360.372.3
FPR [6]ICCV2363.8-
KTSE (Ours)67.073.8
", + "bbox": [ + 217, + 184, + 488, + 369 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/9934c9b4287ab9ed4405b6f657289198a03988abb114ccb3a14e2dc18437b3db.jpg", + "table_caption": [ + "Table 2: Quantitative comparisons on PASCAL VOC 2012 val and test sets with VGG backbone. Sup: Supervision, I: Image-level label, S: Saliency maps." + ], + "table_footnote": [], + "table_body": "
MethodsSupValTest
DRS [28]AAAI21I+S63.664.4
GSM [36]AAAI21I+S63.363.6
NSROM [66]CVPR21I+S65.565.3
EPS [34]CVPR21I+S67.067.3
L2G [24]CVPR22I+S68.568.9
AffinityNet [2]CVPR18I58.460.5
ICD [20]CVPR20I61.260.9
BES [7]ECCV20I60.161.1
ECS [55]ICCV21I62.163.4
KTSE (Ours)I67.367.0
", + "bbox": [ + 506, + 210, + 784, + 369 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "training (expanded with SBD [22]), 1,449 for validation and 1,456 for testing. The COCO dataset is a more challenging benchmark with 80 semantic classes and the background. Following previous works [36, 58, 68], we use the default train/val splits (80k images for training and 40k for validation) in the experiment. Mean intersection over union (mIoU) is adopted as the metric to evaluate the quality of our pseudo labels and segmentation results.", + "bbox": [ + 212, + 398, + 782, + 488 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For the classification network, we follow the work of AEFT [67] and employ ResNet38 [62] as the backbone. We adopt a poly learning rate with an initial value of $10^{-2}$ and a power of 0.9. For the second stage training of WSSS, following the recent work of BECO [45], we adopt DeeplabV2 [5] as the segmentation network, which uses ResNet101 [23] as the backbone. The momentum and weight decay of the SGD optimizer are 0.9 and $10^{-4}$ . The initial learning rate is set to $10^{-2}$ and is decreased using polynomial decay. The segmentation model is trained for 80 epochs and 40 epochs on VOC and COCO datasets, respectively, with a common batch size of 16. We also follow the default setting of DeeplabV2 [5] and conduct experiments with VGG16 backbone for a more comprehensive comparison with previous approaches [7, 20, 24, 34, 55]. All our backbones are pre-trained on ImageNet [17].", + "bbox": [ + 212, + 489, + 787, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Comparisons to the State-of-the-arts", + "text_level": 1, + "bbox": [ + 215, + 693, + 565, + 709 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Accuracy of Pseudo-Masks. We first report the quality of the segmentation seeds and the generated pseudo-masks derived from our approach. The comparison with other state-of-the-arts is presented in Tab. 1. As can be seen, our segmentation seed can reach the mIoU of $67.0\\%$ , bringing a gain of $18.7\\%$ compared to the baseline reported by IRN [1]. Our method can outperform the state-of-the-art method FPR [6] by $3.2\\%$ . With the further refinement of IRN [1], the mIoU of our generated pseudo-masks can arrive at $73.8\\%$ , surpassing the previous SOTA of AEFT [67] and ACR [15] by more than $1.5\\%$ .", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "T.Chen et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/98125d948255bc9d8f381042a1a524784c8d54153782d9ffeb04088a6bb0a58d.jpg", + "table_caption": [ + "Table 3: Quantitative comparisons on PASCAL VOC 2012 val and test sets with ResNet backbone. Sup: Supervision, I: Image-level label, S: Saliency maps, O: Out-of-distribution data, L: Language." + ], + "table_footnote": [], + "table_body": "
MethodsSupValTest
DRS [28]AAAI21I+S71.271.4
NSROM [66]CVPR21I+S70.470.2
EPS [34]CVPR21I+S71.071.8
AuxSeg [64]ICCV21I+S69.068.6
PPC [18]CVPR22I+S72.673.6
RCA [73]CVPR22I+S72.272.8
L2G [24]CVPR22I+S72.171.7
W-OoD [33]CVPR22I+O69.869.9
CLIP-ES [39]CVPR23I+L71.171.4
LPCAM [13]CVPR23I+S71.872.1
AdvCAM [32]CVPR21I68.168.0
CDA [54]ICCV21I66.166.8
CSE [30]ICCV21I68.468.2
RIB [31]NIPS21I68.368.6
AMR [44]AAAI22I68.869.1
MCT [65]CVPR22I71.971.6
AFA [47]CVPR22I66.066.3
SIPE [8]CVPR22I68.869.7
ReCAM [14]CVPR22I68.568.4
PPC [18]CVPR22I67.767.4
ViT-PCM [46]ECCV22I70.370.9
S-BCE [46]ECCV22I70.071.3
AEFT [67]ECCV22I70.971.7
TOCO [48]CVPR23I69.870.5
OCR [15]CVPR23I72.772.0
ACR [15]CVPR23I71.971.9
BECO [45]CVPR23I72.171.8
FPR [6]ICCV23I70.370.1
MARS [26]ICCV23I70.371.2
KTSE (Ours)I73.072.9
", + "bbox": [ + 215, + 226, + 500, + 699 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/92bbdc6bb1847b295da10172ad4ea6e8f1e802cc4231750db5f8b66424fc98a0.jpg", + "table_caption": [ + "Table 4: Quantitative comparisons on COCO val set with VGG backbone. Sup: Supervision, I: Image-level label, S: Saliency maps." + ], + "table_footnote": [], + "table_body": "
MethodsSupVal
IAL [58]IJCV20I+S27.7
GWSM [36]AAAI21I+S28.4
EPS [34]CVPR21I+S35.7
I2CRC [73]TMM22I+S31.2
RCA [73]CVPR22I+S36.8
MDBA [11]TIP23I+S37.8
BFBP [49]ECCV16I20.4
SEC [29]ECCV16I22.4
CONTA [68]NIPS20I23.7
KTSE (Ours)I37.2
", + "bbox": [ + 522, + 214, + 784, + 386 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/4defe851b87d29314c297784531fc0dff7248bc0cb8ef85e24f954b92300346a.jpg", + "table_caption": [ + "Table 5: Quantitative comparisons on COCO val set with ResNet backbone. Sup: Supervision, I: Image-level label, S: Saliency maps, L: Language." + ], + "table_footnote": [], + "table_body": "
MethodsSupVal
L2G [24]CVPR22I+S44.2
CLIP-ES [39]CVPR23I+L45.4
LPCAM [13]CVPR23I+S42.1
OC-CSE [30]ICCV21I36.4
CDA [54]ICCV21I33.2
RIB [31]NIPS21I43.8
MCT [65]CVPR22I42.0
SIPE [8]CVPR22I40.6
TOCO [48]CVPR23I41.3
OCR [15]CVPR23I42.5
ACR [15]CVPR23I45.3
BECO [45]CVPR23I45.1
FPR [6]ICCV23I43.9
USAGE [43]ICCV23I44.3
KTSE (Ours)I45.9
", + "bbox": [ + 524, + 452, + 782, + 698 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Accuracy of Segmentation Maps on PASCAL VOC 2012. Our segmentation results on the PASCAL VOC 2012 dataset with the backbone of VGG and ResNet are demonstrated in Tab. 2 and Tab. 3, respectively. As can be seen, with the VGG backbone, our approach achieves the performance of $67.3\\%$ on the validation set and $67.0\\%$ on the test set, better than other state-of-the-art methods with only image-level labels. Our segmentation results are also competitive to many approaches that rely on saliency maps, e.g., NSROM [66] and EPS [34]. With the ResNet backbone, we can improve the results to $73.0\\%$ and $72.9\\%$ on", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS", + "bbox": [ + 285, + 113, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/3ca663f4f7fa382033020949a3f466a14de2a05267f4fc2c2424d5a5a89dec93.jpg", + "table_caption": [ + "Table 6: Element-wise component analysis. SIE: Simulated Inter-Image Erasing, SSR: Self-Supervised Regularization, MGA: Multi-Granularity Alignment." + ], + "table_footnote": [], + "table_body": "
BaseSIESSRMGAmIoU
✓54.2
✓✓57.1
✓✓57.4
✓✓60.2
✓✓✓63.8
✓✓✓✓67.0
", + "bbox": [ + 222, + 212, + 498, + 314 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/60f1110ce0c27a7df56ba24a5fbe0f998b92cf599ee240aa3fb863bab451633b.jpg", + "table_caption": [ + "Table 7: Comparison of Our Multi-Granularity Alignment with Previous Erasing-based Approaches. GA: Global Alignment; LA: Local Alignment." + ], + "table_footnote": [], + "table_body": "
MethodmIoU
Base54.2
Rigid Classification [69]55.6
Soft GPP Feature [67]54.9
Our GA58.7
Our GA & LA60.2
", + "bbox": [ + 529, + 212, + 782, + 313 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "the validation and test sets, respectively. As can be seen in Tab. 3, our proposed approach can achieve better performance compared with recent SOTA methods. For example, our approach outperforms OCR [15] and ACR [15] by about $1\\%$ on the test set, demonstrating the superiority of our knowledge transfer with simulated inter-image erasing.", + "bbox": [ + 212, + 340, + 782, + 416 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Accuracy of Segmentation Maps on COCO. For the more challenging COCO dataset, we provide performance comparisons with state-of-the-art WSSS methods using the backbone of VGG and ResNet in Tab. 4 and Tab. 5, respectively. As shown in Tab. 4, our proposed KTSE with VGG backbone can achieve the performance of $37.2\\%$ mIoU, much better than previous methods supervised with only image-level labels, e.g., $13.5\\%$ mIoU higher than CONTA [68]. Besides, our results are also competitive compared to previous SOTA methods with additional saliency guidance like RCA [73] and MDBA [11]. With the ResNet backbone, our proposed KTSE reaches the best result of $45.9\\%$ mIoU compared to previous SOTA WSSS methods. Specifically, our approach outperforms ACR [15] and BECO [45] by $0.6\\%$ and $0.8\\%$ mIoU, respectively.", + "bbox": [ + 212, + 417, + 787, + 583 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 604, + 401, + 618 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Element-Wise Component Analysis. In this part, we demonstrate the contribution of each component proposed in our approach to improving the quality of pseudo-masks. As mentioned in Sec. 3.1, we incorporate the GPP [67] into the classification network as our strong baseline. As shown in Tab. 6, with our simulated inter-image erasing (SIE) module, we can improve the accuracy of segmentation seeds from the baseline of $54.2\\%$ to $57.1\\%$ , which demonstrates the benefit of strengthening the localization ability of the network through weakening the activation with extra object information and then learning from the anchor branch. By maintaining reliable activation in discriminative regions, our self-supervised regularization (SSR) module can achieve the mIoU of $57.4\\%$ . Moreover, the combination of our proposed SIE and SSR can significantly boost the performance to $63.8\\%$ mIoU. With our proposed multi-granularity alignment module to gently expand the object activation, the mIoU of pseudo masks finally arrives at $67.0\\%$ , highlighting the importance of improving the quality of", + "bbox": [ + 212, + 628, + 787, + 843 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "T.Chen et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c35b984ca6892c1d4ed49a980056c572e34d7615035221993edc7a9d95ead474.jpg", + "image_caption": [ + "Fig. 3: Example localization maps on the PASCAL VOC 2012 training set. For each (a) image, we show (b) ground truth, localization maps produced by (c) previous work of AEFT [67], (d) our baseline, (e) baseline + SIE, (f) baseline + SIE + SSR, and (g) baseline + SIE + SSR + MGA. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 787, + 305 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "anchor CAM to boost the performance of our knowledge transfer with simulated inter-image erasing.", + "bbox": [ + 212, + 409, + 782, + 439 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Some example localization maps on the PASCAL VOC 2012 training set can be viewed in Fig. 3. As seen from the first two rows, after strengthening the localization ability of the network with our simulated inter-image erasing (SIE) module, we can successfully expand the object activation and discover the less discriminative object in the image (e.g., the smaller animals on the right). We can notice that our SIE can also help alleviate the over-expansion problem and generate more compact activation, e.g., the cow in the first row. However, as demonstrated in the last two rows, the bidirectional alignment in SIE will also lead to more severe under-activation when the anchor branch learns from the sparse activation of the simulated one, especially for larger objects. Fortunately, our proposed self-supervised regularization (SSR) module can help maintain reliable activation in discriminative regions to guarantee effective knowledge transfer. Moreover, as seen in the last row, our multi-granularity alignment (MGA) module can further help discover the integral target region by gently expanding the object activation. Comparing the results of ours with that of AEFT [67], Fig. 3 (g) vs Fig. 3 (c), we can see that our proposed approach can significantly alleviate the under-activation and over-expansion of previous approaches due to insufficient or excessive erasure.", + "bbox": [ + 212, + 443, + 787, + 715 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "MGA vs. Previous Erasing-based Approaches. For the multi-granularity alignment, we resort to the soft class confidence knowledge from the anchor branch to guide the gentle activation expansion of the masked image. In Tab. 7, we compare performance improvement from different guidance settings. As can be seen, our global alignment of CAM feature can improve the baseline from $54.2\\%$ to $58.7\\%$ , outperforming the rigid classification guidance in ACoL [69] and GPP feature alignment adopted in AEFT [67] by $3.1\\%$ and $3.8\\%$ , respectively. It highlights the importance of direct global alignment in CAM features", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS", + "bbox": [ + 287, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "rather than embedding space with extra projection layers. With our pixel-level local alignment, our multi-granularity alignment module can finally improve the performance to $60.2\\%$ , demonstrating the benefit of our gentle alignment strategy compared with previous erasing-based approaches.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Phenomenon of SIE. Based on the observation of classic adversarial erasing methods that removing the most discriminative region will lead to the extra activation of other object parts, the reversal thinking comes naturally—what if we introduce extra discriminative object information? The answer is the original highly activated area will become less discriminative and be lowly activated. We then enhance this weakened activation by learning from the original CAM. When the network learns to increase the attention for the less-discriminative region in the concatenated image, it will also learn to activate less-discriminative object region in the original image to locate more objects.", + "bbox": [ + 212, + 207, + 785, + 343 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "SIE vs. Data Augmentation. Though data augmentations like CutMix will also change the anchor image and lead to activation perturbation, they do not guarantee the weakened activation (which might cause over-expansion) for the anchor part of the concatenated image like ours. In contrast, we only introduce extra object information from the paired image and control other variables. Most importantly, our novelty lies in constructing the simulated inter-image erasing scenario (reversal thinking on existing masking methods) rather than simply augmenting the data input. It enhances the network's localization ability by improving the consequent less activated attention map through learning from the object knowledge of the anchor branch. Accompanied by our proposed MGA, they are designed in a symmetric way that play the adversarial role of expanding and constraining the anchor image activation respectively for mining complete and compact object regions. Note that our approach significantly outperforms the CutMix-based augmentation method of CDA [54] with 73.0 vs. 66.1 on VOC (Tab. 3) and 45.9 vs. 33.2 on COCO (Tab. 5).", + "bbox": [ + 212, + 343, + 787, + 570 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 594, + 359, + 611 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper proposed a Knowledge Transfer with Simulated Inter-Image Erasing (KTSE) approach for weakly supervised semantic segmentation. Specifically, in contrast to existing adversarial erasing-based methods that remove the discriminative part for mining less-discriminative areas, we proposed to simulate an inter-image erasing scenario where extra object information was added through concatenating a paired image. We then strengthened the object localization ability of the network by enhancing the consequent less activated localization map. In addition, we proposed a self-supervised regularization module to maintain reliable activation in the discriminative regions and improve the inter-class object boundary recognition for complex images. Moreover, we proposed a multi-granularity alignment module to gently enlarge the object activation via intra-image erasing for boosting the object knowledge transfer. Extensive experiments and ablation studies were conducted on PASCAL VOC 2012 and COCO datasets to demonstrate the superiority of our proposed approach.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "T.Chen et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (No. 62202227 and 62102182), Natural Science Foundation of Jiangsu Province (No. BK20220938 and BK20220934), China Postdoctoral Science Foundation (No. 2022M711635), Jiangsu Funding Program for Excellent Postdoctoral Talent (No. 2022ZB267), Fundamental Research Funds for the Central Universities (No. 30923010303).", + "bbox": [ + 212, + 176, + 787, + 268 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 292, + 321, + 306 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahn, J., Cho, S., Kwak, S.: Weakly supervised learning of instance segmentation with inter-pixel relations. In: CVPR. pp. 2209-2218 (2019)", + "2. Ahn, J., Kwak, S.: Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In: CVPR. pp. 4981-4990 (2018)", + "3. Bearman, A., Russakovsky, O., Ferrari, V., Fei-Fei, L.: What's the point: Semantic segmentation with point supervision. In: ECCV. pp. 549-565 (2016)", + "4. Cai, X., Lai, Q., Wang, Y., Wang, W., Sun, Z., Yao, Y.: Poly kernel inception network for remote sensing detection. In: CVPR. pp. 27706-27716 (2024)", + "5. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. In: IEEE TPAMI. vol. 40, pp. 834-848 (2017)", + "6. Chen, L., Lei, C., Li, R., Li, S., Zhang, Z., Zhang, L.: Fpr: False positive rectification for weakly supervised semantic segmentation. In: ICCV. pp. 1108-1118 (2023)", + "7. Chen, L., Wu, W., Fu, C., Han, X., Zhang, Y.T.: Weakly supervised semantic segmentation with boundary exploration. In: ECCV (2020)", + "8. Chen, Q., Yang, L., Lai, J.H., Xie, X.: Self-supervised image-specific prototype exploration for weakly supervised semantic segmentation. In: CVPR. pp. 4288-4298 (2022)", + "9. Chen, T., Xie, G.S., Yao, Y., Wang, Q., Shen, F., Tang, Z., Zhang, J.: Semantically meaningful class prototype learning for one-shot image segmentation. IEEE TMM 24, 968-980 (2021)", + "0. Chen, T., Yao, Y., Huang, X., Li, Z., Nie, L., Tang, J.: Spatial structure constraints for weakly supervised semantic segmentation. IEEE TIP 33, 1136-1148 (2024)", + "1. Chen, T., Yao, Y., Tang, J.: Multi-granularity denoising and bidirectional alignment for weakly supervised semantic segmentation. IEEE TIP 32, 2960-2971 (2023)", + "2. Chen, T., Yao, Y., Zhang, L., Wang, Q., Xie, G.S., Shen, F.: Saliency guided inter- and intra-class relation constraints for weakly supervised semantic segmentation. IEEE TMM 25, 1727–1737 (2022)", + "3. Chen, Z., Sun, Q.: Extracting class activation maps from non-discriminative features as well. In: CVPR. pp. 3135-3144 (2023)", + "4. Chen, Z., Wang, T., Wu, X., Hua, X.S., Zhang, H., Sun, Q.: Class re-activation maps for weakly-supervised semantic segmentation. In: CVPR. pp. 969–978 (2022)", + "5. Cheng, Z., Qiao, P., Li, K., Li, S., Wei, P., Ji, X., Yuan, L., Liu, C., Chen, J.: Out-of-candidate rectification for weakly supervised semantic segmentation. In: CVPR. pp. 23673-23684 (2023)" + ], + "bbox": [ + 225, + 324, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS", + "bbox": [ + 285, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Dai, J., He, K., Sun, J.: Boxsup: Exploiting bounding boxes to supervise convolutional networks for semantic segmentation. In: ICCV. pp. 1635-1643 (2015)", + "17. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255 (2009)", + "18. Du, Y., Fu, Z., Liu, Q., Wang, Y.: Weakly supervised semantic segmentation by pixel-to-prototype contrast. In: CVPR. pp. 4320-4329 (2022)", + "19. Everingham, M., Gool, L.V., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. In: IJCV. vol. 88, pp. 303-338 (2010)", + "20. Fan, J., Zhang, Z., Song, C., Tan, T.: Learning integral objects with intra-class discriminator for weakly-supervised semantic segmentation. In: CVPR. pp. 4283-4292 (2020)", + "21. Fan, J., Zhang, Z., Tan, T., Song, C., Xiao, J.: Cian: Cross-image affinity net for weakly supervised semantic segmentation. In: AAAI. vol. 34, pp. 10762-10769 (2020)", + "22. Hariharan, B., Arbeláez, P., Bourdev, L.D., Maji, S., Malik, J.: Semantic contours from inverse detectors. In: ICCV. pp. 991-998 (2011)", + "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)", + "24. Jiang, P.T., Yang, Y., Hou, Q., Wei, Y.: L2g: A simple local-to-global knowledge transfer framework for weakly supervised semantic segmentation. In: CVPR. pp. 16886-16896 (2022)", + "25. Jing, L., Chen, Y., Tian, Y.: Coarse-to-fine semantic segmentation from image-level labels. IEEE TIP 29, 225-236 (2019)", + "26. Jo, S., Yu, I.J., Kim, K.: Mars: Model-agnostic biased object removal without additional supervision for weakly-supervised semantic segmentation. arXiv preprint arXiv:2304.09913 (2023)", + "27. Khoreva, A., Benenson, R., Hosang, J., Hein, M., Schiele, B.: Simple does it: Weakly supervised instance and semantic segmentation. In: CVPR. pp. 876-885 (2017)", + "28. Kim, B., Han, S., Kim, J.: Discriminative region suppression for weakly-supervised semantic segmentation. In: AAAI. vol. 35, pp. 1754-1761 (2021)", + "29. Kolesnikov, A., Lampert, C.: Seed, expand and constrain: Three principles for weakly-supervised image segmentation. In: ECCV. pp. 695-711 (2016)", + "30. Kweon, H., Yoon, S.H., Kim, H., Park, D., Yoon, K.J.: Unlocking the potential of ordinary classifier: Class-specific adversarial erasing framework for weakly supervised semantic segmentation. In: ICCV. pp. 6994-7003 (2021)", + "31. Lee, J., Choi, J., Mok, J., Yoon, S.: Reducing information bottleneck for weakly supervised semantic segmentation. In: NeurIPS. vol. 34, pp. 27408-27421 (2021)", + "32. Lee, J., Kim, E., Yoon, S.: Anti-adversarily manipulated attributions for weakly and semi-supervised semantic segmentation. In: CVPR. pp. 4071-4080 (2021)", + "33. Lee, J., Oh, S.J., Yun, S., Choe, J., Kim, E., Yoon, S.: Weakly supervised semantic segmentation using out-of-distribution data. In: CVPR. pp. 16897-16906 (2022)", + "34. Lee, S., Lee, M., Lee, J., Shim, H.: Railroad is not a train: Saliency as pseudo-pixel supervision for weakly supervised semantic segmentation. In: CVPR. pp. 5495-5505 (2021)", + "35. Li, J., Jie, Z., Wang, X., Wei, X., Ma, L.: Expansion and shrinkage of localization for weakly-supervised semantic segmentation. vol. 35, pp. 16037-16051 (2022)", + "36. Li, X., Zhou, T., Li, J., Zhou, Y., Zhang, Z.: Group-wise semantic mining for weakly supervised semantic segmentation. In: AAAI. pp. 1984-1992 (2021)", + "37. Lin, D., Dai, J., Jia, J., He, K., Sun, J.: Scribblesup: Scribble-supervised convolutional networks for semantic segmentation. In: CVPR. pp. 3159-3167 (2016)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "T.Chen et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, L.: Microsoft coco: Common objects in context. In: ECCV. pp. 740-755 (2014)", + "39. Lin, Y., Chen, M., Wang, W., Wu, B., Li, K., Lin, B., Liu, H., He, X.: Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In: CVPR. pp. 15305-15314 (2023)", + "40. Liu, W., Zhang, C., Lin, G., Hung, T.Y., Miao, C.: Weakly supervised segmentation with maximum bipartite graph matching. In: ACM MM. pp. 2085-2094 (2020)", + "41. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR. pp. 3431-3440 (2015)", + "42. Pei, G., Chen, T., Jiang, X., Liu, H., Sun, Z., Yao, Y.: Videomac: Video masked autoencoders meet convnets. In: CVPR. pp. 22733-22743 (2024)", + "43. Peng, Z., Wang, G., Xie, L., Jiang, D., Shen, W., Tian, Q.: Usage: A unified seed area generation paradigm for weakly supervised semantic segmentation. arXiv preprint arXiv:2303.07806 (2023)", + "44. Qin, J., Wu, J., Xiao, X., Li, L., Wang, X.: Activation modulation and recalibration scheme for weakly supervised semantic segmentation. In: AAAI. vol. 36, pp. 2117-2125 (2022)", + "45. Rong, S., Tu, B., Wang, Z., Li, J.: Boundary-enhanced co-training for weakly supervised semantic segmentation. In: CVPR. pp. 19574-19584 (2023)", + "46. Rossetti, S., Zappia, D., Sanzari, M., Schaerf, M., Pirri, F.: Max pooling with vision transformers reconciles class and shape in weakly supervised semantic segmentation. In: ECCV. pp. 446-463. Springer (2022)", + "47. Ru, L., Zhan, Y., Yu, B., Du, B.: Learning affinity from attention: End-to-end weakly-supervised semantic segmentation with transformers. In: CVPR. pp. 16846-16855 (2022)", + "48. Ru, L., Zheng, H., Zhan, Y., Du, B.: Token contrast for weakly-supervised semantic segmentation. In: CVPR. pp. 3093-3102 (2023)", + "49. Saleh, F., Aliakbarian, M.S., Salzmann, M., Petersson, L., Gould, S., Alvarez, J.M.: Built-in foreground/background prior for weakly-supervised semantic segmentation. In: ECCV. pp. 413-432 (2016)", + "50. Sheng, M., Sun, Z., Cai, Z., Chen, T., Zhou, Y., Yao, Y.: Adaptive integration of partial label learning and negative learning for enhanced noisy label learning. In: AAAI. vol. 38, pp. 4820-4828 (2024)", + "51. Shimoda, W., Yanai, K.: Self-supervised difference detection for weakly-supervised semantic segmentation. In: ICCV. pp. 5208-5217 (2019)", + "52. Singh, K.K., Yu, H., Sarmasi, A., Pradeep, G., Lee, Y.J.: Hide-and-seek: A data augmentation technique for weakly-supervised localization and beyond. arXiv preprint arXiv:1811.02545 (2018)", + "53. Song, C., Huang, Y., Ouyang, W., Wang, L.: Box-driven class-wise region masking and filling rate guided loss for weakly supervised semantic segmentation. In: CVPR. pp. 3136-3145 (2019)", + "54. Su, Y., Sun, R., Lin, G., Wu, Q.: Context decoupling augmentation for weakly supervised semantic segmentation. In: ICCV (2021)", + "55. Sun, K., Shi, H., Zhang, Z., Huang, Y.: ECS-net: Improving weakly supervised semantic segmentation by using connections between class activation maps. In: ICCV. pp. 7283-7292 (2021)", + "56. Vernaza, P., Chandraker, M.: Learning random-walk label propagation for weakly-supervised semantic segmentation. In: CVPR. pp. 7158-7166 (2017)", + "57. Wang, W., Sun, G., Van Gool, L.: Looking beyond single images for weakly supervised semantic segmentation learning. IEEE TPAMI (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS", + "bbox": [ + 287, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "58. Wang, X., Liu, S., Ma, H., Yang, M.H.: Weakly-supervised semantic segmentation by iterative affinity learning. vol. 128, pp. 1736-1749. Springer (2020)", + "59. Wang, Y., Zhang, J., Kan, M., Shan, S., Chen, X.: Self-supervised equivariant attention mechanism for weakly supervised semantic segmentation. In: CVPR. pp. 12275-12284 (2020)", + "60. Wei, Y., Feng, J., Liang, X., Cheng, M.M., Zhao, Y., Yan, S.: Object region mining with adversarial erasing: A simple classification to semantic segmentation approach. In: CVPR. pp. 1568-1576 (2017)", + "61. Wei, Y., Xiao, H., Shi, H., Jie, Z., Feng, J., Huang, T.: Revisiting dilated convolution: A simple approach for weakly-and semi-supervised semantic segmentation. In: CVPR. pp. 7268-7277 (2018)", + "62. Wu, Z., Shen, C., Van Den Hengel, A.: Wider or deeper: Revisiting the resnet model for visual recognition. PR 90, 119-133 (2019)", + "63. Xie, J., Hou, X., Ye, K., Shen, L.: Clims: Cross language image matching for weakly supervised semantic segmentation. In: CVPR. pp. 4483-4492 (2022)", + "64. Xu, L., Ouyang, W., Bennamoun, M., Boussaid, F., Sohel, F., Xu, D.: Leveraging auxiliary tasks with affinity learning for weakly supervised semantic segmentation. In: ICCV. pp. 6984-6993 (2021)", + "65. Xu, L., Ouyang, W., Bennamoun, M., Boussaid, F., Xu, D.: Multi-class token transformer for weakly supervised semantic segmentation. In: CVPR. pp. 4310-4319 (2022)", + "66. Yao, Y., Chen, T., Xie, G.S., Zhang, C., Shen, F., Wu, Q., Tang, Z., Zhang, J.: Non-salient region object mining for weakly supervised semantic segmentation. In: CVPR. pp. 2623-2632 (2021)", + "67. Yoon, S.H., Kweon, H., Cho, J., Kim, S., Yoon, K.J.: Adversarial erasing framework via triplet with gated pyramid pooling layer for weakly supervised semantic segmentation. In: ECCV. pp. 326-344. Springer (2022)", + "68. Zhang, D., Zhang, H., Tang, J., Hua, X.S., Sun, Q.: Causal intervention for weakly-supervised semantic segmentation. In: NeurIPS. vol. 33 (2020)", + "69. Zhang, X., Wei, Y., Feng, J., Yang, Y., Huang, T.S.: Adversarial complementary learning for weakly supervised object localization. In: CVPR. pp. 1325-1334 (2018)", + "70. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. In: AAAI. vol. 34, pp. 13001-13008 (2020)", + "71. Zhou, B., Khosla, A., Lapedriza, Å., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: CVPR. pp. 2921-2929 (2016)", + "72. Zhou, T., Li, L., Li, X., Feng, C.M., Li, J., Shao, L.: Group-wise learning for weakly supervised semantic segmentation. IEEE TIP 31, 799-811 (2021)", + "73. Zhou, T., Zhang, M., Zhao, F., Li, J.: Regional semantic contrast and aggregation for weakly supervised semantic segmentation. In: CVPR. pp. 4299-4309 (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 686 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "T.Chen et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation/41d0edfc-3cb8-4f88-8e86-7449093dd76b_model.json b/2024/Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation/41d0edfc-3cb8-4f88-8e86-7449093dd76b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a15c267d0da370e2c5a2731af1a0102eef071f8a --- /dev/null +++ b/2024/Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation/41d0edfc-3cb8-4f88-8e86-7449093dd76b_model.json @@ -0,0 +1,2458 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.228, + 0.141, + 0.78, + 0.21 + ], + "angle": 0, + "content": "Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.303, + 0.235, + 0.702, + 0.267 + ], + "angle": 0, + "content": "Tao Chen\\(^{1}\\), Xiruo Jiang\\(^{1}\\), Gensheng Pei\\(^{1}\\), Zeren Sun\\(^{1}\\), Yucheng Wang\\(^{2}\\), and Yazhou Yao\\(^{1(\\text{忍})}\\)" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.278, + 0.744, + 0.337 + ], + "angle": 0, + "content": "\\(^{1}\\) Nanjing University of Science and Technology, Nanjing 210094, China {taochen,jiangxiruo622,peigsh,zerens,yazhou.yao}@njust.edu.cn \\(^{2}\\) Horizon Robotics, Beijing 100089, China yucheng.wang@horizon.cc" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.374, + 0.744, + 0.694 + ], + "angle": 0, + "content": "Abstract. Though adversarial erasing has prevailed in weakly supervised semantic segmentation to help activate integral object regions, existing approaches still suffer from the dilemma of under-activation and over-expansion due to the difficulty in determining when to stop erasing. In this paper, we propose a Knowledge Transfer with Simulated InterImage Erasing (KTSE) approach for weakly supervised semantic segmentation to alleviate the above problem. In contrast to existing erasing-based methods that remove the discriminative part for more object discovery, we propose a simulated inter-image erasing scenario to weaken the original activation by introducing extra object information. Then, object knowledge is transferred from the anchor image to the consequent less activated localization map to strengthen network localization ability. Considering the adopted bidirectional alignment will also weaken the anchor image activation if appropriate constraints are missing, we propose a self-supervised regularization module to maintain the reliable activation in discriminative regions and improve the inter-class object boundary recognition for complex images with multiple categories of objects. In addition, we resort to intra-image erasing and propose a multi-granularity alignment module to gently enlarge the object activation to boost the object knowledge transfer. Extensive experiments and ablation studies on PAS-CAL VOC 2012 and COCO datasets demonstrate the superiority of our proposed approach. Codes and models are available at https://nust-machine-intelligence-laboratory.github.io/project-KTSE." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.706, + 0.751, + 0.737 + ], + "angle": 0, + "content": "Keywords: Weakly Supervised Learning \\(\\cdot\\) Semantic Segmentation \\(\\cdot\\) InterImage Erasing" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.762, + 0.377, + 0.779 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.842 + ], + "angle": 0, + "content": "With huge progress in the era of deep learning, semantic segmentation has been widely applied in fields like autonomous driving and image editing [5,41]. Since deep learning thus far is data-driven, the model training typically involves a large" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "T.Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.498, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.276, + 0.189, + 0.427, + 0.199 + ], + "angle": 0, + "content": "(a) Previous Adversarial Erasing" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.2, + 0.498, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.242, + 0.429, + 0.253 + ], + "angle": 0, + "content": "(c) Result CAM vs Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.145, + 0.787, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.194, + 0.787, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.241, + 0.733, + 0.253 + ], + "angle": 0, + "content": "(b) Our Simulated Inter-Image Erasing" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.263, + 0.788, + 0.348 + ], + "angle": 0, + "content": "Fig. 1: (a) Previous adversarial erasing-based approaches typically suffer from the over-expansion problem, which is hard to constrain. (b) Different from their information removal strategy, we propose to add extra object knowledge from a paired image to weaken the current object activation. The localization ability of the network is then enhanced by improving the consequent less activated attention map through learning from the object knowledge of the anchor branch. (C) Result comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.788, + 0.473 + ], + "angle": 0, + "content": "abundance of labeled images. However, collecting accurate pixel-level annotation for segmentation tasks is highly labor-intensive and time-consuming [3, 9, 16]. Therefore, as a promising direction to alleviate such annotation burden, weakly supervised learning has attracted the attention of many researchers. In this paper, we focus on the weakly supervised semantic segmentation (WSSS) under the supervision of image-level labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.789, + 0.763 + ], + "angle": 0, + "content": "The recent advance of WSSS typically follows the three-step pipeline of: 1) transforming image tags into pixel-level coarse labels, 2) refining the pseudo labels, and 3) training the final segmentation model with the refined labels. Speaking of segmentation label generation, the technique of class activation map (CAM) [71] has been the de facto paradigm for object localization with image-level labels. Unfortunately, the naive CAM can only highlight the most discriminative area of objects and thus its small and sparse activation will lead to incomplete object mining. Since the first step of seed generation is the foundation of later processes, numerous works have been developed to expand CAM activation for high-quality pseudo labels. Among them, adversarial erasing is one of the mainstream approaches, aiming to discover new and complementary object regions by masking the currently detected area in an adversarial manner [60,67,69]. For example, the pioneering work of AE-PSL [60] performs the erasing multiple times to progressively discover increasingly activated areas. The recent work of AEFT [67] proposes to learn the concept of erasing with the triplet loss between the input image, erased image, and negatively sampled image. Though improvement is obtained, existing adversarial erasing-based methods still suffer from the difficulty in determining when to stop erasing: excessive removal will lead to over-expansion and insufficient erasure will result in under-activation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this paper, we propose a Knowledge Transfer with Simulated Inter-Image Erasing (KTSE) approach for WSSS to alleviate the problem mentioned above. Existing erasing-based methods typically mask the most discriminative object part to force the classifier into seeking other regions. Based on the observation that information removal leads to more activation (which might also incur the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.287, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.265 + ], + "angle": 0, + "content": "over-expansion illustrated in Fig. 1 (a)), we act in a diametrically opposite way. Specifically, as shown in Fig. 1 (b), we propose to simulate an inter-image erasing (SIE) scenario, introducing extra object information by concatenating a paired image and rendering the anchor image as the erased one. We then strengthen the object localization ability of the network by transferring the object knowledge from the anchor CAM to the consequent less activated localization map. Intuitively, our knowledge transfer with simulated inter-image erasing will not cause the over-expansion problem if the object in the anchor branch is well-discovered." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.27, + 0.785, + 0.435 + ], + "angle": 0, + "content": "However, since the above knowledge transfer is bidirectional, it will also weaken object mining in anchor branch when learning from the sparse activation of simulated branch. Therefore, as shown in the right of Fig. 2, we propose a self-supervised regularization module for the anchor CAM feature to maintain its reliable activation in discriminative regions. Specifically, we first generate pseudo labels from the CAM feature by locating confident foreground and background area, which are used to directly supervise the learning of CAM features. However, extracted labels of the foreground are quite noisy in complex images with multiple categories of objects due to the inter-class confusion. Therefore, we also design an inter-class loss to implicitly encourage activation consistency in complex images to improve recognition of the inter-class object boundary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.784, + 0.68 + ], + "angle": 0, + "content": "Though our proposed self-supervised regularization can effectively constrain and facilitate the knowledge transfer in our simulated inter-image erasing module, the improvement of the network's object localization ability will be severely limited by the quality of anchor CAM feature which is usually under-activated. Therefore, we resort to the traditional intra-image erasing and propose a multi-granularity alignment (MGA) module to gently expand the object activation without introducing much background noise. Specifically, as shown in the bottom of Fig. 2, we first leverage the image-level global alignment to distill the soft object confidence from the anchor branch to the masked one for object activation enlargement. Then, we leverage a pixel-level local alignment to transfer the newly discovered object information back to the anchor branch. With the gentle and constrained activation enlargement, the proposed multi-granularity alignment module further boosts the performance of our simulated inter-image erasing. As can be seen in Fig. 2, our SIE and MGA are designed in a symmetric way that play the adversarial role of expanding and constraining the anchor image activation respectively for mining complete and compact object regions." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.683, + 0.593, + 0.696 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.7, + 0.784, + 0.744 + ], + "angle": 0, + "content": "(1) We propose a knowledge transfer with simulated inter-image erasing approach for weakly supervised semantic segmentation to alleviate the over-expansion problem of existing adversarial erasing-based methods." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.747, + 0.784, + 0.792 + ], + "angle": 0, + "content": "(2) A self-supervised regularization module is proposed to constrain the knowledge transfer for maintaining reliable activation and improving recognition of the inter-class object boundary." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "(3) We propose a multi-granularity alignment module to gently enlarge the object activation via image-level global alignment and pixel-level local alignment, boosting the knowledge transfer with simulated inter-image erasing." + }, + { + "type": "list", + "bbox": [ + 0.215, + 0.7, + 0.784, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.363, + 0.127 + ], + "angle": 0, + "content": "T.Chen et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.177, + 0.618, + 0.194 + ], + "angle": 0, + "content": "2.1 Weakly Supervised Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.203, + 0.789, + 0.522 + ], + "angle": 0, + "content": "Weakly supervised semantic segmentation (WSSS) is the task that aims to obtain a high-performance segmentation model by learning only from weak labels. Compared with bounding boxes [16, 25, 27, 53], scribbles [37, 56] and points [3], the image-level label [1, 4, 10, 12, 24, 34, 42, 50, 66, 73] is the cheapest weak supervision to collect, making it the most popular annotation format for WSSS. The recent advance of WSSS [1, 24, 34, 66, 73] typically relies on the class activation technique of CAM [71] to locate the target object, which helps transform the image-level label to pixel-level dense annotation for segmentation network training. Since the object area highlighted in CAM is usually small and sparse, enlarging the CAM activation becomes the focus of the WSSS task. For example, RDC [61] and DRS [28] propagate the object information to less discriminative regions via dilated convolution with varied rates and attention suppression of most discriminative areas, respectively. Popular contrastive learning [18, 48, 73] and self-supervised approaches [8, 51, 59] are also applied to facilitate object mining by improving representation learning. Moreover, cross-image information [21, 57, 72] is exploited to help the classifier discover more object patterns to obtain consistent and integral target regions. Besides attention expansion, Ahn and Kwak propose AffinityNet [2] for CAM refinement, which propagates semantic affinity via a random walk to recover delicate shapes of objects. IRN [1] is then proposed to exploit class boundary detection for discovering the entire instance areas with more accurate boundaries." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.543, + 0.48, + 0.559 + ], + "angle": 0, + "content": "2.2 Erasing-based Approaches" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Considering the original CAM technique usually leads to suboptimal performance due to failure to localize all object parts, erasing-based approaches are commonly employed to expand the highlighted region [30, 52, 60, 67, 69, 70]. They mask areas in a training image to force the network to seek other relevant parts. While direct erasing methods like Hide-and-Seek [52] hide image patches randomly, adversarial erasing-based approaches mask the most discriminative regions with the attention information, demonstrating more promising activation expansion potential. The first adversarial erasing-based work for WSSS is proposed by Wei et al. [60], which progressively discovers new object discriminative regions by erasing the currently activated area. Such a heuristic erasing strategy then attracts the attention of many researchers. For example, ACoL [69] proposes adversarial complementary learning with two adversary classifiers. Recently, Kweon et al. propose to apply the adversarial erasing framework to exploit the potential of the pre-trained classifier [30] and learn the concept of erasing with the triplet loss [67]. While adversarial erasing approaches achieve great success in enlarging the CAM activation, they tend to suffer from the over-expansion problem due to difficulty in determining when to stop erasing. Unlike the information removal strategy adopted in these approaches, we propose a simulated" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.288, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.455 + ], + "angle": 0, + "content": "Fig. 2: The architecture of our proposed approach. We propose a simulated inter-image erasing (SIE) scenario where extra object information is introduced from another paired image. We then strengthen the object localization ability of the network by improving the consequent less activated localization map through learning object knowledge from the anchor image. A self-supervised regularization (SSR) module is also proposed to avoid weakening the anchor activation due to bidirectional alignment and improve the inter-class object boundary recognition for complex images. In addition, we propose a multi-granularity alignment (MGA) module to gently enlarge the object activation to further boost the object knowledge transfer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.788, + 0.53 + ], + "angle": 0, + "content": "inter-image erasing scenario to weaken the activation by adding extra discriminative object information. The localization ability of network is then enhanced by improving the consequent less activated CAM through knowledge transfer." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.551, + 0.331, + 0.567 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In this paper, we propose a knowledge transfer with simulated inter-image erasing approach for weakly supervised semantic segmentation to alleviate the over-expansion and under-activation problems of existing adversarial erasing-based methods. Our framework is illustrated in Fig. 2. As demonstrated in the anchor branch, we train a classification network with the given image-level weak labels, which consists of a backbone feature extractor and a pooling classification head. In contrast to existing adversarial erasing-based approaches that expand the activation through masking the most discriminative object part, as shown in the top-left of Fig. 2, we propose to simulate an inter-image erasing scenario where extra object information is introduced through concatenating a paired image. Object knowledge is then transferred from the anchor image to the consequent less activated localization map for strengthening the object localization ability of network. We also propose a self-supervised regularization module to avoid weakening the anchor activation due to bidirectional alignment and improve the inter-class object boundary recognition for complex multi-category images. In addition, we propose a multi-granularity alignment module to gently enlarge the object activation for boosting the object knowledge transfer." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.363, + 0.128 + ], + "angle": 0, + "content": "T.Chen et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.406, + 0.16 + ], + "angle": 0, + "content": "3.1 CAM Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.298 + ], + "angle": 0, + "content": "For the architecture of classification network, we follow the previous erasing-based work ACoL [69] and remove the final fully-connected layer. We set the output channel of the backbone to \\( C + 1 \\), where \\( C \\) is the number of foreground categories and 1 is added for the background. We can thus directly generate object localization maps from the class-aware CAM features \\( F \\) in the forward pass to facilitate the object knowledge transfer. To obtain CAM for each foreground class \\( c \\), we feed the attention map \\( F^c \\) into a ReLU layer and then normalize it to the range from 0 to 1:" + }, + { + "type": "equation", + "bbox": [ + 0.439, + 0.299, + 0.786, + 0.332 + ], + "angle": 0, + "content": "\\[\nA ^ {c} = \\frac {\\operatorname {R e L U} \\left(F ^ {c}\\right)}{\\operatorname* {m a x} \\left(F ^ {c}\\right)}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.339, + 0.788, + 0.414 + ], + "angle": 0, + "content": "To improve the quality of the CAM by capturing the global context and local details of the image, we adopt the Gated Pyramid Pooling (GPP) layer [67] as the final pooling head. With the classification logits \\( q^{c} \\) generated from the pooling head, we train the classification network with the multi-label soft margin loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.314, + 0.427, + 0.786, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c l s} = - \\frac {1}{C} \\sum_ {c = 1} ^ {C} y ^ {c} \\log \\sigma \\left(q ^ {c}\\right) + \\left(1 - y ^ {c}\\right) \\log \\left[ 1 - \\sigma \\left(q ^ {c}\\right) \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.48, + 0.784, + 0.496 + ], + "angle": 0, + "content": "Here, \\(\\sigma (\\cdot)\\) is the sigmoid function. \\(y^{c}\\) is the image-level label for the \\(c\\)-th class." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.522, + 0.517, + 0.538 + ], + "angle": 0, + "content": "3.2 Simulated Inter-Image Erasing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Due to the sparsity of the CAM activation, previous adversarial erasing-based methods typically enlarge the object activation by maintaining the classification confidence on masked images or features, where the most discriminative regions are erased. However, due to the absence of guidelines on when to stop erasing, these approaches easily incur over-expansion with excessive removal and might also still suffer from under-activation with insufficient erasure. Therefore, we propose a knowledge transfer with simulated inter-image erasing approach for weakly supervised semantic segmentation to alleviate the above problem." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.675, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In contrast to existing erasing-based approaches that remove the discriminative part for more object mining, we propose a simulated inter-image erasing (SIE) scenario to enable the network to benefit from introducing extra discriminative object information. Specifically, as shown in the top-left of Fig. 2, a larger synthetic image, created by concatenating the anchor image and another paired one, plays the role of the original image needing masking. Correspondingly, the anchor image can be treated as the erased one where the paired image is masked. Due to the introduction of extra discriminative object information, fewer object regions will be highlighted in the anchor part of the synthetic image. Therefore, we then strengthen the object localization ability of the network by transferring the more integral object knowledge from the original anchor CAM to the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.287, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "consequently less activated anchor part of the synthetic image. The knowledge transfer loss for the simulated inter-image erasing can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.185, + 0.785, + 0.209 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {k t} = \\operatorname {R e L U} \\left(\\hat {F} _ {a} - \\hat {F} _ {s}\\right), \\quad \\text {w h e r e} \\hat {F} _ {a} = C F E \\left(F _ {a}, y\\right), \\hat {F} _ {s} = C F E \\left(F _ {s}, y\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.214, + 0.788, + 0.442 + ], + "angle": 0, + "content": "Here, \\( F_{a} \\) and \\( F_{s} \\) denote the CAM features of the anchor and simulated branches, respectively. CFE represents class feature extraction, which extracts the feature channel related to the classes that exist in the image. The ReLU operation means that our alignment focuses on the more highlighted object region rather than the background. Note that no rotation or resize is performed for the concatenated image. We only introduce extra object information from the paired image and control other variables to guarantee the weakened activation for the anchor part. Intuitively, our knowledge transfer with simulated inter-image erasing will not cause the over-expansion problem when the object in the anchor branch is well-discovered. Our experiments demonstrate that such knowledge transfer can effectively strengthen the localization ability of the network to alleviate the under-activation, especially when multiple object instances co-exist in the image. Besides, benefiting from the bidirectional alignment, the anchor branch will also learn from the weakened activation of the simulated one and produce compact attention maps for small objects to mitigate the over-expansion problem." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.46, + 0.513, + 0.475 + ], + "angle": 0, + "content": "3.3 Self-Supervised Regularization" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.482, + 0.789, + 0.65 + ], + "angle": 0, + "content": "Our simulated inter-image erasing aims to improve the object localization ability of the network by maintaining the high activation given extra discriminative object information. However, the adopted knowledge transfer is bidirectional, which will also weaken the anchor branch object mining when learning from the sparse activation of the simulated branch. Since simply cutting the gradient propagation of the anchor branch makes the training unstable, we propose a self-supervised regularization module for the anchor CAM feature to maintain its reliable activation in discriminative regions. Specifically, with the generated CAM feature \\( F_{a} \\) and the corresponding CAM \\( A_{a} \\), we first leverage two thresholds \\( \\beta_{h} = 0.3 \\) and \\( \\beta_{l} = 0.15 \\) to locate the confident foreground and background as follows (the subscript \\( a \\) is omitted for simplicity):" + }, + { + "type": "equation", + "bbox": [ + 0.346, + 0.657, + 0.785, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\hat {Y} _ {i, j} = \\left\\{ \\begin{array}{l l} \\operatorname {a r g m a x} \\left(F _ {i, j,:}\\right), & \\text {i f} \\max \\left(A _ {i, j,:}\\right) \\geq \\beta_ {h}, \\\\ 0, & \\text {i f} \\max \\left(A _ {i, j,:}\\right) \\leq \\beta_ {l}, \\\\ 2 5 5, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.719, + 0.788, + 0.795 + ], + "angle": 0, + "content": "where 255 denote the ignored labels for uncertain regions and \\(\\arg\\max(\\cdot)\\) extracts the semantic class with the maximum activation value. After refining the pseudo label \\(\\hat{Y}\\) with the pixel-adaptive refinement module [47], we leverage it to directly supervise the learning of CAM features with the cross-entropy loss (coordinates \\(i, j\\) are omitted):" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.804, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c e} = - \\sum_ {c = 0} ^ {C} \\hat {Y} _ {b g} ^ {c} \\log \\Gamma (F) ^ {c} + \\omega_ {f g} \\hat {Y} _ {f g} ^ {c} \\log \\Gamma (F) ^ {c}. \\tag {5}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.363, + 0.128 + ], + "angle": 0, + "content": "T.Chen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.299 + ], + "angle": 0, + "content": "\\(\\Gamma(\\cdot)\\) denotes the softmax function. We also leverage the pseudo label \\(\\hat{Y}\\) generated from the anchor branch to guide the learning of the simulated CAM feature with the same cross-entropy loss. However, we notice that the extracted labels of the foreground are quite noisy, especially for complex images with multiple categories of objects where the inter-class boundary is blurred. Here, we define images with objects of only one category as simple images and others containing two or more categories as complex ones. Therefore, we discount the influence of foreground label \\(\\hat{Y}_{fg}\\) with the weight of \\(\\omega_{fg} = 0.0125\\). Besides, as shown in the right of Fig. 2, we design an inter-class loss to implicitly encourage the activation consistency for the multiple foreground classes in complex images as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.31, + 0.786, + 0.343 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {i n t e r}} = \\frac {\\sum \\left(V ^ {2 n d} - V ^ {\\max }\\right) \\cdot M ^ {f}}{\\left| M ^ {f} \\right|}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.355, + 0.788, + 0.416 + ], + "angle": 0, + "content": "Here, \\( M^f \\) denotes the foreground mask obtained with a threshold \\( \\beta_o = 0.2 \\) and \\( |\\cdot| \\) return the value of \\( l_1 \\)-norm. \\( V^{max} \\) and \\( V^{2nd} \\) denotes the largest and second largest activation value along the channel dimension of the CAM feature \\( F_a \\), which can be obtained as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.246, + 0.429, + 0.786, + 0.45 + ], + "angle": 0, + "content": "\\[\nV _ {i, j} ^ {m a x} = \\max (F _ {i, j,:}), \\quad F ^ {l e f t} = \\mathsf {r e m o v e} (F, V _ {m a x}), \\quad V _ {i, j} ^ {2 n d} = \\max (F _ {i, j,:} ^ {l e f t}). \\qquad (7)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.788, + 0.537 + ], + "angle": 0, + "content": "Eq. (6) improves the exclusivity of the class activation for each pixel, leading to more accurate boundaries between foreground objects. Our proposed self-supervised regularization module treats the simple and complex images differently, intending to maintain the reliable activation of the main branch without deteriorating the inter-class boundary caused by inaccurate pseudo labels." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.558, + 0.504, + 0.574 + ], + "angle": 0, + "content": "3.4 Multi-Granularity Alignment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Our proposed self-supervised regularization module can effectively constrain and boost the knowledge transfer in our simulated inter-image erasing module. However, the improvement of the network's object localization ability will be severely limited by the quality of the anchor CAM, which is usually under-activated. Therefore, it is natural to resort to the traditional intra-image erasing, which aims to expand the object area. Following AEFT [67], we leverage a threshold of 0.6 to mask the most discriminative region. To avoid introducing too much background noise, we propose a multi-granularity alignment module to gently expand the object activation, which encourages the CAM expansion of the erased image and then transfers the learned object knowledge back to the anchor branch. As shown in the bottom of Fig. 2, we first input the anchor feature \\( F_{a} \\) and masked feature \\( F_{m} \\) into a class feature extraction module to obtain the corresponding features of the existing target categories \\( (\\hat{F}_{a} \\) and \\( \\hat{F}_{m} \\)). Following the experimental finding of AEFT [67] that rigid classification supervision [69] for the masked branch tends to trigger the over-expansion problem in adversarial erasing, we resort to the soft class confidence guidance from the anchor branch to the masked one. Specifically, we directly adopt a global average pooling (GAP) operation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.287, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "to obtain the final class confidence for each branch and an image-level global alignment loss can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.307, + 0.189, + 0.697, + 0.214 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g l o b a l} = G A P \\left(R e L U \\left(\\hat {F} _ {a}\\right)\\right) - G A P \\left(R e L U \\left(\\hat {F} _ {m}\\right)\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.218, + 0.784, + 0.234 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} \\quad \\hat {F _ {m}} = C F E \\left(F _ {m}, y\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.247, + 0.785, + 0.399 + ], + "angle": 0, + "content": "\\(\\hat{F}_a\\) and CFE have been defined in Eq. (3). Another difference with rigid classification supervision is that benefiting from the class feature extraction, our global alignment loss only focuses on the logit confidence for classes that exist in the image, enabling more effective and efficient gradient propagation. Unlike AEFT [67] that utilizes the feature space of the GPP layer as embedding space for loss construction, our experiments reveal that direct alignment with the CAM feature leads to more promising performance. With class information learned from the anchor branch, we further leverage a pixel-level local activation alignment to transfer the newly discovered object information from the erased image to guide the anchor branch learning." + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.412, + 0.784, + 0.435 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {l o c a l}} = \\operatorname {R e L U} \\left(\\hat {F _ {m}} - \\hat {F _ {a}}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.446, + 0.785, + 0.581 + ], + "angle": 0, + "content": "Such pixel-level local alignment can also hinder the erased branch from activating the unwanted background area by learning from the lowly-activated regions of the anchor branch. With the gentle and constrained activation enlargement, the multi-granularity alignment module further boosts the performance of our proposed simulated inter-image erasing approach. As can be seen in Fig. 2, our proposed SIE and MGA branches are designed in a symmetric way that play the adversarial role of expanding and constraining the anchor image activation, respectively. Their synergically cooperation facilitates the convergence for mining complete and compact object regions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.604, + 0.417, + 0.619 + ], + "angle": 0, + "content": "3.5 Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.629, + 0.487, + 0.644 + ], + "angle": 0, + "content": "The overall training loss is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.311, + 0.658, + 0.784, + 0.673 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c l s} + \\mathcal {L} _ {k t} + \\mathcal {L} _ {g l o b a l} + \\mathcal {L} _ {l o c a l} + \\mathcal {L} _ {c e} + \\lambda_ {i n t e r} \\mathcal {L} _ {i n t e r}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.685, + 0.784, + 0.713 + ], + "angle": 0, + "content": "We empirically set \\(\\lambda_{inter} = 0.005\\) as the hyperparameter that controls the weight of inter-class loss." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.738, + 0.367, + 0.755 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.77, + 0.57, + 0.785 + ], + "angle": 0, + "content": "4.1 Datasets and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.784, + 0.841 + ], + "angle": 0, + "content": "We evaluate our approach on the PASCAL VOC 2012 [19] and COCO [38] datasets. The PASCAL VOC 2012 dataset contains 21 classes (20 object categories and the background) for semantic segmentation. It has 10,582 images for" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "T.Chen et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.492, + 0.173 + ], + "angle": 0, + "content": "Table 1: Accuracy of pseudo-masks evaluated on PASCAL VOC 2012 training set." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.185, + 0.49, + 0.371 + ], + "angle": 0, + "content": "
MethodsSeedw/IRN [1]
IRN [1]CVPR1948.366.3
MBMNet [40]MM2050.266.8
CONTA [68]NIPS2048.867.9
AdvCAM [32]CVPR2155.669.9
RIB [31]NIPS2156.670.6
ReCAM [14]CVPR2256.670.5
ESOL [35]NIPS2253.668.7
CLIMS [63]CVPR2256.670.5
AEFT [67]ECCV2256.071.0
ACR [15]CVPR2360.372.3
FPR [6]ICCV2363.8-
KTSE (Ours)67.073.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.503, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 2: Quantitative comparisons on PASCAL VOC 2012 val and test sets with VGG backbone. Sup: Supervision, I: Image-level label, S: Saliency maps." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.212, + 0.785, + 0.371 + ], + "angle": 0, + "content": "
MethodsSupValTest
DRS [28]AAAI21I+S63.664.4
GSM [36]AAAI21I+S63.363.6
NSROM [66]CVPR21I+S65.565.3
EPS [34]CVPR21I+S67.067.3
L2G [24]CVPR22I+S68.568.9
AffinityNet [2]CVPR18I58.460.5
ICD [20]CVPR20I61.260.9
BES [7]ECCV20I60.161.1
ECS [55]ICCV21I62.163.4
KTSE (Ours)I67.367.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.784, + 0.489 + ], + "angle": 0, + "content": "training (expanded with SBD [22]), 1,449 for validation and 1,456 for testing. The COCO dataset is a more challenging benchmark with 80 semantic classes and the background. Following previous works [36, 58, 68], we use the default train/val splits (80k images for training and 40k for validation) in the experiment. Mean intersection over union (mIoU) is adopted as the metric to evaluate the quality of our pseudo labels and segmentation results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.788, + 0.673 + ], + "angle": 0, + "content": "For the classification network, we follow the work of AEFT [67] and employ ResNet38 [62] as the backbone. We adopt a poly learning rate with an initial value of \\(10^{-2}\\) and a power of 0.9. For the second stage training of WSSS, following the recent work of BECO [45], we adopt DeeplabV2 [5] as the segmentation network, which uses ResNet101 [23] as the backbone. The momentum and weight decay of the SGD optimizer are 0.9 and \\(10^{-4}\\). The initial learning rate is set to \\(10^{-2}\\) and is decreased using polynomial decay. The segmentation model is trained for 80 epochs and 40 epochs on VOC and COCO datasets, respectively, with a common batch size of 16. We also follow the default setting of DeeplabV2 [5] and conduct experiments with VGG16 backbone for a more comprehensive comparison with previous approaches [7, 20, 24, 34, 55]. All our backbones are pre-trained on ImageNet [17]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.694, + 0.566, + 0.71 + ], + "angle": 0, + "content": "4.2 Comparisons to the State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Accuracy of Pseudo-Masks. We first report the quality of the segmentation seeds and the generated pseudo-masks derived from our approach. The comparison with other state-of-the-arts is presented in Tab. 1. As can be seen, our segmentation seed can reach the mIoU of \\(67.0\\%\\), bringing a gain of \\(18.7\\%\\) compared to the baseline reported by IRN [1]. Our method can outperform the state-of-the-art method FPR [6] by \\(3.2\\%\\). With the further refinement of IRN [1], the mIoU of our generated pseudo-masks can arrive at \\(73.8\\%\\), surpassing the previous SOTA of AEFT [67] and ACR [15] by more than \\(1.5\\%\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.287, + 0.114, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.158, + 0.499, + 0.226 + ], + "angle": 0, + "content": "Table 3: Quantitative comparisons on PASCAL VOC 2012 val and test sets with ResNet backbone. Sup: Supervision, I: Image-level label, S: Saliency maps, O: Out-of-distribution data, L: Language." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.227, + 0.501, + 0.7 + ], + "angle": 0, + "content": "
MethodsSupValTest
DRS [28]AAAI21I+S71.271.4
NSROM [66]CVPR21I+S70.470.2
EPS [34]CVPR21I+S71.071.8
AuxSeg [64]ICCV21I+S69.068.6
PPC [18]CVPR22I+S72.673.6
RCA [73]CVPR22I+S72.272.8
L2G [24]CVPR22I+S72.171.7
W-OoD [33]CVPR22I+O69.869.9
CLIP-ES [39]CVPR23I+L71.171.4
LPCAM [13]CVPR23I+S71.872.1
AdvCAM [32]CVPR21I68.168.0
CDA [54]ICCV21I66.166.8
CSE [30]ICCV21I68.468.2
RIB [31]NIPS21I68.368.6
AMR [44]AAAI22I68.869.1
MCT [65]CVPR22I71.971.6
AFA [47]CVPR22I66.066.3
SIPE [8]CVPR22I68.869.7
ReCAM [14]CVPR22I68.568.4
PPC [18]CVPR22I67.767.4
ViT-PCM [46]ECCV22I70.370.9
S-BCE [46]ECCV22I70.071.3
AEFT [67]ECCV22I70.971.7
TOCO [48]CVPR23I69.870.5
OCR [15]CVPR23I72.772.0
ACR [15]CVPR23I71.971.9
BECO [45]CVPR23I72.171.8
FPR [6]ICCV23I70.370.1
MARS [26]ICCV23I70.371.2
KTSE (Ours)I73.072.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.159, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 4: Quantitative comparisons on COCO val set with VGG backbone. Sup: Supervision, I: Image-level label, S: Saliency maps." + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.215, + 0.785, + 0.387 + ], + "angle": 0, + "content": "
MethodsSupVal
IAL [58]IJCV20I+S27.7
GWSM [36]AAAI21I+S28.4
EPS [34]CVPR21I+S35.7
I2CRC [73]TMM22I+S31.2
RCA [73]CVPR22I+S36.8
MDBA [11]TIP23I+S37.8
BFBP [49]ECCV16I20.4
SEC [29]ECCV16I22.4
CONTA [68]NIPS20I23.7
KTSE (Ours)I37.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.397, + 0.788, + 0.452 + ], + "angle": 0, + "content": "Table 5: Quantitative comparisons on COCO val set with ResNet backbone. Sup: Supervision, I: Image-level label, S: Saliency maps, L: Language." + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.453, + 0.784, + 0.699 + ], + "angle": 0, + "content": "
MethodsSupVal
L2G [24]CVPR22I+S44.2
CLIP-ES [39]CVPR23I+L45.4
LPCAM [13]CVPR23I+S42.1
OC-CSE [30]ICCV21I36.4
CDA [54]ICCV21I33.2
RIB [31]NIPS21I43.8
MCT [65]CVPR22I42.0
SIPE [8]CVPR22I40.6
TOCO [48]CVPR23I41.3
OCR [15]CVPR23I42.5
ACR [15]CVPR23I45.3
BECO [45]CVPR23I45.1
FPR [6]ICCV23I43.9
USAGE [43]ICCV23I44.3
KTSE (Ours)I45.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Accuracy of Segmentation Maps on PASCAL VOC 2012. Our segmentation results on the PASCAL VOC 2012 dataset with the backbone of VGG and ResNet are demonstrated in Tab. 2 and Tab. 3, respectively. As can be seen, with the VGG backbone, our approach achieves the performance of \\(67.3\\%\\) on the validation set and \\(67.0\\%\\) on the test set, better than other state-of-the-art methods with only image-level labels. Our segmentation results are also competitive to many approaches that rely on saliency maps, e.g., NSROM [66] and EPS [34]. With the ResNet backbone, we can improve the results to \\(73.0\\%\\) and \\(72.9\\%\\) on" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "T.Chen et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.505, + 0.201 + ], + "angle": 0, + "content": "Table 6: Element-wise component analysis. SIE: Simulated Inter-Image Erasing, SSR: Self-Supervised Regularization, MGA: Multi-Granularity Alignment." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.213, + 0.499, + 0.315 + ], + "angle": 0, + "content": "
BaseSIESSRMGAmIoU
✓54.2
✓✓57.1
✓✓57.4
✓✓60.2
✓✓✓63.8
✓✓✓✓67.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.527, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 7: Comparison of Our Multi-Granularity Alignment with Previous Erasing-based Approaches. GA: Global Alignment; LA: Local Alignment." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.213, + 0.784, + 0.314 + ], + "angle": 0, + "content": "
MethodmIoU
Base54.2
Rigid Classification [69]55.6
Soft GPP Feature [67]54.9
Our GA58.7
Our GA & LA60.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.784, + 0.417 + ], + "angle": 0, + "content": "the validation and test sets, respectively. As can be seen in Tab. 3, our proposed approach can achieve better performance compared with recent SOTA methods. For example, our approach outperforms OCR [15] and ACR [15] by about \\(1\\%\\) on the test set, demonstrating the superiority of our knowledge transfer with simulated inter-image erasing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.788, + 0.584 + ], + "angle": 0, + "content": "Accuracy of Segmentation Maps on COCO. For the more challenging COCO dataset, we provide performance comparisons with state-of-the-art WSSS methods using the backbone of VGG and ResNet in Tab. 4 and Tab. 5, respectively. As shown in Tab. 4, our proposed KTSE with VGG backbone can achieve the performance of \\(37.2\\%\\) mIoU, much better than previous methods supervised with only image-level labels, e.g., \\(13.5\\%\\) mIoU higher than CONTA [68]. Besides, our results are also competitive compared to previous SOTA methods with additional saliency guidance like RCA [73] and MDBA [11]. With the ResNet backbone, our proposed KTSE reaches the best result of \\(45.9\\%\\) mIoU compared to previous SOTA WSSS methods. Specifically, our approach outperforms ACR [15] and BECO [45] by \\(0.6\\%\\) and \\(0.8\\%\\) mIoU, respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.605, + 0.403, + 0.619 + ], + "angle": 0, + "content": "4.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Element-Wise Component Analysis. In this part, we demonstrate the contribution of each component proposed in our approach to improving the quality of pseudo-masks. As mentioned in Sec. 3.1, we incorporate the GPP [67] into the classification network as our strong baseline. As shown in Tab. 6, with our simulated inter-image erasing (SIE) module, we can improve the accuracy of segmentation seeds from the baseline of \\(54.2\\%\\) to \\(57.1\\%\\), which demonstrates the benefit of strengthening the localization ability of the network through weakening the activation with extra object information and then learning from the anchor branch. By maintaining reliable activation in discriminative regions, our self-supervised regularization (SSR) module can achieve the mIoU of \\(57.4\\%\\). Moreover, the combination of our proposed SIE and SSR can significantly boost the performance to \\(63.8\\%\\) mIoU. With our proposed multi-granularity alignment module to gently expand the object activation, the mIoU of pseudo masks finally arrives at \\(67.0\\%\\), highlighting the importance of improving the quality of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.288, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.788, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.317, + 0.785, + 0.371 + ], + "angle": 0, + "content": "Fig. 3: Example localization maps on the PASCAL VOC 2012 training set. For each (a) image, we show (b) ground truth, localization maps produced by (c) previous work of AEFT [67], (d) our baseline, (e) baseline + SIE, (f) baseline + SIE + SSR, and (g) baseline + SIE + SSR + MGA. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.784, + 0.44 + ], + "angle": 0, + "content": "anchor CAM to boost the performance of our knowledge transfer with simulated inter-image erasing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.788, + 0.716 + ], + "angle": 0, + "content": "Some example localization maps on the PASCAL VOC 2012 training set can be viewed in Fig. 3. As seen from the first two rows, after strengthening the localization ability of the network with our simulated inter-image erasing (SIE) module, we can successfully expand the object activation and discover the less discriminative object in the image (e.g., the smaller animals on the right). We can notice that our SIE can also help alleviate the over-expansion problem and generate more compact activation, e.g., the cow in the first row. However, as demonstrated in the last two rows, the bidirectional alignment in SIE will also lead to more severe under-activation when the anchor branch learns from the sparse activation of the simulated one, especially for larger objects. Fortunately, our proposed self-supervised regularization (SSR) module can help maintain reliable activation in discriminative regions to guarantee effective knowledge transfer. Moreover, as seen in the last row, our multi-granularity alignment (MGA) module can further help discover the integral target region by gently expanding the object activation. Comparing the results of ours with that of AEFT [67], Fig. 3 (g) vs Fig. 3 (c), we can see that our proposed approach can significantly alleviate the under-activation and over-expansion of previous approaches due to insufficient or excessive erasure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "MGA vs. Previous Erasing-based Approaches. For the multi-granularity alignment, we resort to the soft class confidence knowledge from the anchor branch to guide the gentle activation expansion of the masked image. In Tab. 7, we compare performance improvement from different guidance settings. As can be seen, our global alignment of CAM feature can improve the baseline from \\(54.2\\%\\) to \\(58.7\\%\\), outperforming the rigid classification guidance in ACoL [69] and GPP feature alignment adopted in AEFT [67] by \\(3.1\\%\\) and \\(3.8\\%\\), respectively. It highlights the importance of direct global alignment in CAM features" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "T.Chen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.207 + ], + "angle": 0, + "content": "rather than embedding space with extra projection layers. With our pixel-level local alignment, our multi-granularity alignment module can finally improve the performance to \\(60.2\\%\\), demonstrating the benefit of our gentle alignment strategy compared with previous erasing-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.787, + 0.344 + ], + "angle": 0, + "content": "Phenomenon of SIE. Based on the observation of classic adversarial erasing methods that removing the most discriminative region will lead to the extra activation of other object parts, the reversal thinking comes naturally—what if we introduce extra discriminative object information? The answer is the original highly activated area will become less discriminative and be lowly activated. We then enhance this weakened activation by learning from the original CAM. When the network learns to increase the attention for the less-discriminative region in the concatenated image, it will also learn to activate less-discriminative object region in the original image to locate more objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.571 + ], + "angle": 0, + "content": "SIE vs. Data Augmentation. Though data augmentations like CutMix will also change the anchor image and lead to activation perturbation, they do not guarantee the weakened activation (which might cause over-expansion) for the anchor part of the concatenated image like ours. In contrast, we only introduce extra object information from the paired image and control other variables. Most importantly, our novelty lies in constructing the simulated inter-image erasing scenario (reversal thinking on existing masking methods) rather than simply augmenting the data input. It enhances the network's localization ability by improving the consequent less activated attention map through learning from the object knowledge of the anchor branch. Accompanied by our proposed MGA, they are designed in a symmetric way that play the adversarial role of expanding and constraining the anchor image activation respectively for mining complete and compact object regions. Note that our approach significantly outperforms the CutMix-based augmentation method of CDA [54] with 73.0 vs. 66.1 on VOC (Tab. 3) and 45.9 vs. 33.2 on COCO (Tab. 5)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.595, + 0.36, + 0.612 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This paper proposed a Knowledge Transfer with Simulated Inter-Image Erasing (KTSE) approach for weakly supervised semantic segmentation. Specifically, in contrast to existing adversarial erasing-based methods that remove the discriminative part for mining less-discriminative areas, we proposed to simulate an inter-image erasing scenario where extra object information was added through concatenating a paired image. We then strengthened the object localization ability of the network by enhancing the consequent less activated localization map. In addition, we proposed a self-supervised regularization module to maintain reliable activation in the discriminative regions and improve the inter-class object boundary recognition for complex images. Moreover, we proposed a multi-granularity alignment module to gently enlarge the object activation via intra-image erasing for boosting the object knowledge transfer. Extensive experiments and ablation studies were conducted on PASCAL VOC 2012 and COCO datasets to demonstrate the superiority of our proposed approach." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.287, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.269 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (No. 62202227 and 62102182), Natural Science Foundation of Jiangsu Province (No. BK20220938 and BK20220934), China Postdoctoral Science Foundation (No. 2022M711635), Jiangsu Funding Program for Excellent Postdoctoral Talent (No. 2022ZB267), Fundamental Research Funds for the Central Universities (No. 30923010303)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.293, + 0.323, + 0.308 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.325, + 0.785, + 0.353 + ], + "angle": 0, + "content": "1. Ahn, J., Cho, S., Kwak, S.: Weakly supervised learning of instance segmentation with inter-pixel relations. In: CVPR. pp. 2209-2218 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.354, + 0.785, + 0.393 + ], + "angle": 0, + "content": "2. Ahn, J., Kwak, S.: Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In: CVPR. pp. 4981-4990 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.395, + 0.785, + 0.422 + ], + "angle": 0, + "content": "3. Bearman, A., Russakovsky, O., Ferrari, V., Fei-Fei, L.: What's the point: Semantic segmentation with point supervision. In: ECCV. pp. 549-565 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.424, + 0.785, + 0.45 + ], + "angle": 0, + "content": "4. Cai, X., Lai, Q., Wang, Y., Wang, W., Sun, Z., Yao, Y.: Poly kernel inception network for remote sensing detection. In: CVPR. pp. 27706-27716 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.451, + 0.785, + 0.491 + ], + "angle": 0, + "content": "5. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. In: IEEE TPAMI. vol. 40, pp. 834-848 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.493, + 0.785, + 0.52 + ], + "angle": 0, + "content": "6. Chen, L., Lei, C., Li, R., Li, S., Zhang, Z., Zhang, L.: Fpr: False positive rectification for weakly supervised semantic segmentation. In: ICCV. pp. 1108-1118 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.521, + 0.785, + 0.548 + ], + "angle": 0, + "content": "7. Chen, L., Wu, W., Fu, C., Han, X., Zhang, Y.T.: Weakly supervised semantic segmentation with boundary exploration. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.549, + 0.785, + 0.589 + ], + "angle": 0, + "content": "8. Chen, Q., Yang, L., Lai, J.H., Xie, X.: Self-supervised image-specific prototype exploration for weakly supervised semantic segmentation. In: CVPR. pp. 4288-4298 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.59, + 0.785, + 0.631 + ], + "angle": 0, + "content": "9. Chen, T., Xie, G.S., Yao, Y., Wang, Q., Shen, F., Tang, Z., Zhang, J.: Semantically meaningful class prototype learning for one-shot image segmentation. IEEE TMM 24, 968-980 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.632, + 0.785, + 0.659 + ], + "angle": 0, + "content": "0. Chen, T., Yao, Y., Huang, X., Li, Z., Nie, L., Tang, J.: Spatial structure constraints for weakly supervised semantic segmentation. IEEE TIP 33, 1136-1148 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.785, + 0.7 + ], + "angle": 0, + "content": "1. Chen, T., Yao, Y., Tang, J.: Multi-granularity denoising and bidirectional alignment for weakly supervised semantic segmentation. IEEE TIP 32, 2960-2971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.701, + 0.785, + 0.742 + ], + "angle": 0, + "content": "2. Chen, T., Yao, Y., Zhang, L., Wang, Q., Xie, G.S., Shen, F.: Saliency guided inter- and intra-class relation constraints for weakly supervised semantic segmentation. IEEE TMM 25, 1727–1737 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.785, + 0.77 + ], + "angle": 0, + "content": "3. Chen, Z., Sun, Q.: Extracting class activation maps from non-discriminative features as well. In: CVPR. pp. 3135-3144 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.798 + ], + "angle": 0, + "content": "4. Chen, Z., Wang, T., Wu, X., Hua, X.S., Zhang, H., Sun, Q.: Class re-activation maps for weakly-supervised semantic segmentation. In: CVPR. pp. 969–978 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Cheng, Z., Qiao, P., Li, K., Li, S., Wei, P., Ji, X., Yuan, L., Liu, C., Chen, J.: Out-of-candidate rectification for weakly supervised semantic segmentation. In: CVPR. pp. 23673-23684 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.325, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "T.Chen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Dai, J., He, K., Sun, J.: Boxsup: Exploiting bounding boxes to supervise convolutional networks for semantic segmentation. In: ICCV. pp. 1635-1643 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "17. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "18. Du, Y., Fu, Z., Liu, Q., Wang, Y.: Weakly supervised semantic segmentation by pixel-to-prototype contrast. In: CVPR. pp. 4320-4329 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "19. Everingham, M., Gool, L.V., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. In: IJCV. vol. 88, pp. 303-338 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.299 + ], + "angle": 0, + "content": "20. Fan, J., Zhang, Z., Song, C., Tan, T.: Learning integral objects with intra-class discriminator for weakly-supervised semantic segmentation. In: CVPR. pp. 4283-4292 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "21. Fan, J., Zhang, Z., Tan, T., Song, C., Xiao, J.: Cian: Cross-image affinity net for weakly supervised semantic segmentation. In: AAAI. vol. 34, pp. 10762-10769 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "22. Hariharan, B., Arbeláez, P., Bourdev, L.D., Maji, S., Malik, J.: Semantic contours from inverse detectors. In: ICCV. pp. 991-998 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "24. Jiang, P.T., Yang, Y., Hou, Q., Wei, Y.: L2g: A simple local-to-global knowledge transfer framework for weakly supervised semantic segmentation. In: CVPR. pp. 16886-16896 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "25. Jing, L., Chen, Y., Tian, Y.: Coarse-to-fine semantic segmentation from image-level labels. IEEE TIP 29, 225-236 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "26. Jo, S., Yu, I.J., Kim, K.: Mars: Model-agnostic biased object removal without additional supervision for weakly-supervised semantic segmentation. arXiv preprint arXiv:2304.09913 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "27. Khoreva, A., Benenson, R., Hosang, J., Hein, M., Schiele, B.: Simple does it: Weakly supervised instance and semantic segmentation. In: CVPR. pp. 876-885 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "28. Kim, B., Han, S., Kim, J.: Discriminative region suppression for weakly-supervised semantic segmentation. In: AAAI. vol. 35, pp. 1754-1761 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "29. Kolesnikov, A., Lampert, C.: Seed, expand and constrain: Three principles for weakly-supervised image segmentation. In: ECCV. pp. 695-711 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.632 + ], + "angle": 0, + "content": "30. Kweon, H., Yoon, S.H., Kim, H., Park, D., Yoon, K.J.: Unlocking the potential of ordinary classifier: Class-specific adversarial erasing framework for weakly supervised semantic segmentation. In: ICCV. pp. 6994-7003 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.66 + ], + "angle": 0, + "content": "31. Lee, J., Choi, J., Mok, J., Yoon, S.: Reducing information bottleneck for weakly supervised semantic segmentation. In: NeurIPS. vol. 34, pp. 27408-27421 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "32. Lee, J., Kim, E., Yoon, S.: Anti-adversarily manipulated attributions for weakly and semi-supervised semantic segmentation. In: CVPR. pp. 4071-4080 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.715 + ], + "angle": 0, + "content": "33. Lee, J., Oh, S.J., Yun, S., Choe, J., Kim, E., Yoon, S.: Weakly supervised semantic segmentation using out-of-distribution data. In: CVPR. pp. 16897-16906 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "34. Lee, S., Lee, M., Lee, J., Shim, H.: Railroad is not a train: Saliency as pseudo-pixel supervision for weakly supervised semantic segmentation. In: CVPR. pp. 5495-5505 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.784 + ], + "angle": 0, + "content": "35. Li, J., Jie, Z., Wang, X., Wei, X., Ma, L.: Expansion and shrinkage of localization for weakly-supervised semantic segmentation. vol. 35, pp. 16037-16051 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "36. Li, X., Zhou, T., Li, J., Zhou, Y., Zhang, Z.: Group-wise semantic mining for weakly supervised semantic segmentation. In: AAAI. pp. 1984-1992 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "37. Lin, D., Dai, J., Jia, J., He, K., Sun, J.: Scribblesup: Scribble-supervised convolutional networks for semantic segmentation. In: CVPR. pp. 3159-3167 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.288, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "38. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, L.: Microsoft coco: Common objects in context. In: ECCV. pp. 740-755 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "39. Lin, Y., Chen, M., Wang, W., Wu, B., Li, K., Lin, B., Liu, H., He, X.: Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In: CVPR. pp. 15305-15314 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.257 + ], + "angle": 0, + "content": "40. Liu, W., Zhang, C., Lin, G., Hung, T.Y., Miao, C.: Weakly supervised segmentation with maximum bipartite graph matching. In: ACM MM. pp. 2085-2094 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.284 + ], + "angle": 0, + "content": "41. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR. pp. 3431-3440 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.311 + ], + "angle": 0, + "content": "42. Pei, G., Chen, T., Jiang, X., Liu, H., Sun, Z., Yao, Y.: Videomac: Video masked autoencoders meet convnets. In: CVPR. pp. 22733-22743 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.352 + ], + "angle": 0, + "content": "43. Peng, Z., Wang, G., Xie, L., Jiang, D., Shen, W., Tian, Q.: Usage: A unified seed area generation paradigm for weakly supervised semantic segmentation. arXiv preprint arXiv:2303.07806 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.392 + ], + "angle": 0, + "content": "44. Qin, J., Wu, J., Xiao, X., Li, L., Wang, X.: Activation modulation and recalibration scheme for weakly supervised semantic segmentation. In: AAAI. vol. 36, pp. 2117-2125 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.785, + 0.42 + ], + "angle": 0, + "content": "45. Rong, S., Tu, B., Wang, Z., Li, J.: Boundary-enhanced co-training for weakly supervised semantic segmentation. In: CVPR. pp. 19574-19584 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.46 + ], + "angle": 0, + "content": "46. Rossetti, S., Zappia, D., Sanzari, M., Schaerf, M., Pirri, F.: Max pooling with vision transformers reconciles class and shape in weakly supervised semantic segmentation. In: ECCV. pp. 446-463. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.501 + ], + "angle": 0, + "content": "47. Ru, L., Zhan, Y., Yu, B., Du, B.: Learning affinity from attention: End-to-end weakly-supervised semantic segmentation with transformers. In: CVPR. pp. 16846-16855 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.528 + ], + "angle": 0, + "content": "48. Ru, L., Zheng, H., Zhan, Y., Du, B.: Token contrast for weakly-supervised semantic segmentation. In: CVPR. pp. 3093-3102 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.569 + ], + "angle": 0, + "content": "49. Saleh, F., Aliakbarian, M.S., Salzmann, M., Petersson, L., Gould, S., Alvarez, J.M.: Built-in foreground/background prior for weakly-supervised semantic segmentation. In: ECCV. pp. 413-432 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.61 + ], + "angle": 0, + "content": "50. Sheng, M., Sun, Z., Cai, Z., Chen, T., Zhou, Y., Yao, Y.: Adaptive integration of partial label learning and negative learning for enhanced noisy label learning. In: AAAI. vol. 38, pp. 4820-4828 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.637 + ], + "angle": 0, + "content": "51. Shimoda, W., Yanai, K.: Self-supervised difference detection for weakly-supervised semantic segmentation. In: ICCV. pp. 5208-5217 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.678 + ], + "angle": 0, + "content": "52. Singh, K.K., Yu, H., Sarmasi, A., Pradeep, G., Lee, Y.J.: Hide-and-seek: A data augmentation technique for weakly-supervised localization and beyond. arXiv preprint arXiv:1811.02545 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "53. Song, C., Huang, Y., Ouyang, W., Wang, L.: Box-driven class-wise region masking and filling rate guided loss for weakly supervised semantic segmentation. In: CVPR. pp. 3136-3145 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.746 + ], + "angle": 0, + "content": "54. Su, Y., Sun, R., Lin, G., Wu, Q.: Context decoupling augmentation for weakly supervised semantic segmentation. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.786 + ], + "angle": 0, + "content": "55. Sun, K., Shi, H., Zhang, Z., Huang, Y.: ECS-net: Improving weakly supervised semantic segmentation by using connections between class activation maps. In: ICCV. pp. 7283-7292 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "56. Vernaza, P., Chandraker, M.: Learning random-walk label propagation for weakly-supervised semantic segmentation. In: CVPR. pp. 7158-7166 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "57. Wang, W., Sun, G., Van Gool, L.: Looking beyond single images for weakly supervised semantic segmentation learning. IEEE TPAMI (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "T.Chen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "58. Wang, X., Liu, S., Ma, H., Yang, M.H.: Weakly-supervised semantic segmentation by iterative affinity learning. vol. 128, pp. 1736-1749. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "59. Wang, Y., Zhang, J., Kan, M., Shan, S., Chen, X.: Self-supervised equivariant attention mechanism for weakly supervised semantic segmentation. In: CVPR. pp. 12275-12284 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "60. Wei, Y., Feng, J., Liang, X., Cheng, M.M., Zhao, Y., Yan, S.: Object region mining with adversarial erasing: A simple classification to semantic segmentation approach. In: CVPR. pp. 1568-1576 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "61. Wei, Y., Xiao, H., Shi, H., Jie, Z., Feng, J., Huang, T.: Revisiting dilated convolution: A simple approach for weakly-and semi-supervised semantic segmentation. In: CVPR. pp. 7268-7277 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.327 + ], + "angle": 0, + "content": "62. Wu, Z., Shen, C., Van Den Hengel, A.: Wider or deeper: Revisiting the resnet model for visual recognition. PR 90, 119-133 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.356 + ], + "angle": 0, + "content": "63. Xie, J., Hou, X., Ye, K., Shen, L.: Clims: Cross language image matching for weakly supervised semantic segmentation. In: CVPR. pp. 4483-4492 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.397 + ], + "angle": 0, + "content": "64. Xu, L., Ouyang, W., Bennamoun, M., Boussaid, F., Sohel, F., Xu, D.: Leveraging auxiliary tasks with affinity learning for weakly supervised semantic segmentation. In: ICCV. pp. 6984-6993 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "65. Xu, L., Ouyang, W., Bennamoun, M., Boussaid, F., Xu, D.: Multi-class token transformer for weakly supervised semantic segmentation. In: CVPR. pp. 4310-4319 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "66. Yao, Y., Chen, T., Xie, G.S., Zhang, C., Shen, F., Wu, Q., Tang, Z., Zhang, J.: Non-salient region object mining for weakly supervised semantic segmentation. In: CVPR. pp. 2623-2632 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "67. Yoon, S.H., Kweon, H., Cho, J., Kim, S., Yoon, K.J.: Adversarial erasing framework via triplet with gated pyramid pooling layer for weakly supervised semantic segmentation. In: ECCV. pp. 326-344. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.55 + ], + "angle": 0, + "content": "68. Zhang, D., Zhang, H., Tang, J., Hua, X.S., Sun, Q.: Causal intervention for weakly-supervised semantic segmentation. In: NeurIPS. vol. 33 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.578 + ], + "angle": 0, + "content": "69. Zhang, X., Wei, Y., Feng, J., Yang, Y., Huang, T.S.: Adversarial complementary learning for weakly supervised object localization. In: CVPR. pp. 1325-1334 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "70. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. In: AAAI. vol. 34, pp. 13001-13008 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "71. Zhou, B., Khosla, A., Lapedriza, Å., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: CVPR. pp. 2921-2929 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.66 + ], + "angle": 0, + "content": "72. Zhou, T., Li, L., Li, X., Feng, C.M., Li, J., Shao, L.: Group-wise learning for weakly supervised semantic segmentation. IEEE TIP 31, 799-811 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.688 + ], + "angle": 0, + "content": "73. Zhou, T., Zhang, M., Zhao, F., Li, J.: Regional semantic contrast and aggregation for weakly supervised semantic segmentation. In: CVPR. pp. 4299-4309 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.688 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation/41d0edfc-3cb8-4f88-8e86-7449093dd76b_origin.pdf b/2024/Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation/41d0edfc-3cb8-4f88-8e86-7449093dd76b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1bbc14d040405298c4c45894a57e84a0e7c6c521 --- /dev/null +++ b/2024/Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation/41d0edfc-3cb8-4f88-8e86-7449093dd76b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:711b51c2e92357ed4330759dae9d77b45698e2cc40e9df9675367ac1fa1d76c2 +size 1796658 diff --git a/2024/Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation/full.md b/2024/Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eafe44cd6ba9db659d94dfaff0bebb7f03283e11 --- /dev/null +++ b/2024/Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation/full.md @@ -0,0 +1,306 @@ +# Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation + +Tao Chen $^{1}$ , Xiruo Jiang $^{1}$ , Gensheng Pei $^{1}$ , Zeren Sun $^{1}$ , Yucheng Wang $^{2}$ , and Yazhou Yao $^{1(\text{忍})}$ + +$^{1}$ Nanjing University of Science and Technology, Nanjing 210094, China {taochen,jiangxiruo622,peigsh,zerens,yazhou.yao}@njust.edu.cn $^{2}$ Horizon Robotics, Beijing 100089, China yucheng.wang@horizon.cc + +Abstract. Though adversarial erasing has prevailed in weakly supervised semantic segmentation to help activate integral object regions, existing approaches still suffer from the dilemma of under-activation and over-expansion due to the difficulty in determining when to stop erasing. In this paper, we propose a Knowledge Transfer with Simulated InterImage Erasing (KTSE) approach for weakly supervised semantic segmentation to alleviate the above problem. In contrast to existing erasing-based methods that remove the discriminative part for more object discovery, we propose a simulated inter-image erasing scenario to weaken the original activation by introducing extra object information. Then, object knowledge is transferred from the anchor image to the consequent less activated localization map to strengthen network localization ability. Considering the adopted bidirectional alignment will also weaken the anchor image activation if appropriate constraints are missing, we propose a self-supervised regularization module to maintain the reliable activation in discriminative regions and improve the inter-class object boundary recognition for complex images with multiple categories of objects. In addition, we resort to intra-image erasing and propose a multi-granularity alignment module to gently enlarge the object activation to boost the object knowledge transfer. Extensive experiments and ablation studies on PAS-CAL VOC 2012 and COCO datasets demonstrate the superiority of our proposed approach. Codes and models are available at https://nust-machine-intelligence-laboratory.github.io/project-KTSE. + +Keywords: Weakly Supervised Learning $\cdot$ Semantic Segmentation $\cdot$ InterImage Erasing + +# 1 Introduction + +With huge progress in the era of deep learning, semantic segmentation has been widely applied in fields like autonomous driving and image editing [5,41]. Since deep learning thus far is data-driven, the model training typically involves a large + +![](images/58f2061d0629499ba4986e9c776a3ac1bfa5f966e43e207d7b25862355899a78.jpg) +(a) Previous Adversarial Erasing + +![](images/9de728cea173eb22073b8f9b7700208595d731074360f0bde3e7937f5d92c335.jpg) +(c) Result CAM vs Ground Truth +Fig. 1: (a) Previous adversarial erasing-based approaches typically suffer from the over-expansion problem, which is hard to constrain. (b) Different from their information removal strategy, we propose to add extra object knowledge from a paired image to weaken the current object activation. The localization ability of the network is then enhanced by improving the consequent less activated attention map through learning from the object knowledge of the anchor branch. (C) Result comparison. + +![](images/577d128e86287d349d5eb7649899060911529a0aafb216ca283234d19d0a2355.jpg) + +![](images/4bea42c8c3cc3ec67b863124745c05367d582458fc549d5fa261ebe703ba6e73.jpg) +(b) Our Simulated Inter-Image Erasing + +abundance of labeled images. However, collecting accurate pixel-level annotation for segmentation tasks is highly labor-intensive and time-consuming [3, 9, 16]. Therefore, as a promising direction to alleviate such annotation burden, weakly supervised learning has attracted the attention of many researchers. In this paper, we focus on the weakly supervised semantic segmentation (WSSS) under the supervision of image-level labels. + +The recent advance of WSSS typically follows the three-step pipeline of: 1) transforming image tags into pixel-level coarse labels, 2) refining the pseudo labels, and 3) training the final segmentation model with the refined labels. Speaking of segmentation label generation, the technique of class activation map (CAM) [71] has been the de facto paradigm for object localization with image-level labels. Unfortunately, the naive CAM can only highlight the most discriminative area of objects and thus its small and sparse activation will lead to incomplete object mining. Since the first step of seed generation is the foundation of later processes, numerous works have been developed to expand CAM activation for high-quality pseudo labels. Among them, adversarial erasing is one of the mainstream approaches, aiming to discover new and complementary object regions by masking the currently detected area in an adversarial manner [60,67,69]. For example, the pioneering work of AE-PSL [60] performs the erasing multiple times to progressively discover increasingly activated areas. The recent work of AEFT [67] proposes to learn the concept of erasing with the triplet loss between the input image, erased image, and negatively sampled image. Though improvement is obtained, existing adversarial erasing-based methods still suffer from the difficulty in determining when to stop erasing: excessive removal will lead to over-expansion and insufficient erasure will result in under-activation. + +In this paper, we propose a Knowledge Transfer with Simulated Inter-Image Erasing (KTSE) approach for WSSS to alleviate the problem mentioned above. Existing erasing-based methods typically mask the most discriminative object part to force the classifier into seeking other regions. Based on the observation that information removal leads to more activation (which might also incur the + +over-expansion illustrated in Fig. 1 (a)), we act in a diametrically opposite way. Specifically, as shown in Fig. 1 (b), we propose to simulate an inter-image erasing (SIE) scenario, introducing extra object information by concatenating a paired image and rendering the anchor image as the erased one. We then strengthen the object localization ability of the network by transferring the object knowledge from the anchor CAM to the consequent less activated localization map. Intuitively, our knowledge transfer with simulated inter-image erasing will not cause the over-expansion problem if the object in the anchor branch is well-discovered. + +However, since the above knowledge transfer is bidirectional, it will also weaken object mining in anchor branch when learning from the sparse activation of simulated branch. Therefore, as shown in the right of Fig. 2, we propose a self-supervised regularization module for the anchor CAM feature to maintain its reliable activation in discriminative regions. Specifically, we first generate pseudo labels from the CAM feature by locating confident foreground and background area, which are used to directly supervise the learning of CAM features. However, extracted labels of the foreground are quite noisy in complex images with multiple categories of objects due to the inter-class confusion. Therefore, we also design an inter-class loss to implicitly encourage activation consistency in complex images to improve recognition of the inter-class object boundary. + +Though our proposed self-supervised regularization can effectively constrain and facilitate the knowledge transfer in our simulated inter-image erasing module, the improvement of the network's object localization ability will be severely limited by the quality of anchor CAM feature which is usually under-activated. Therefore, we resort to the traditional intra-image erasing and propose a multi-granularity alignment (MGA) module to gently expand the object activation without introducing much background noise. Specifically, as shown in the bottom of Fig. 2, we first leverage the image-level global alignment to distill the soft object confidence from the anchor branch to the masked one for object activation enlargement. Then, we leverage a pixel-level local alignment to transfer the newly discovered object information back to the anchor branch. With the gentle and constrained activation enlargement, the proposed multi-granularity alignment module further boosts the performance of our simulated inter-image erasing. As can be seen in Fig. 2, our SIE and MGA are designed in a symmetric way that play the adversarial role of expanding and constraining the anchor image activation respectively for mining complete and compact object regions. + +Our contributions can be summarized as follows: + +(1) We propose a knowledge transfer with simulated inter-image erasing approach for weakly supervised semantic segmentation to alleviate the over-expansion problem of existing adversarial erasing-based methods. +(2) A self-supervised regularization module is proposed to constrain the knowledge transfer for maintaining reliable activation and improving recognition of the inter-class object boundary. +(3) We propose a multi-granularity alignment module to gently enlarge the object activation via image-level global alignment and pixel-level local alignment, boosting the knowledge transfer with simulated inter-image erasing. + +# 2 Related Work + +# 2.1 Weakly Supervised Semantic Segmentation + +Weakly supervised semantic segmentation (WSSS) is the task that aims to obtain a high-performance segmentation model by learning only from weak labels. Compared with bounding boxes [16, 25, 27, 53], scribbles [37, 56] and points [3], the image-level label [1, 4, 10, 12, 24, 34, 42, 50, 66, 73] is the cheapest weak supervision to collect, making it the most popular annotation format for WSSS. The recent advance of WSSS [1, 24, 34, 66, 73] typically relies on the class activation technique of CAM [71] to locate the target object, which helps transform the image-level label to pixel-level dense annotation for segmentation network training. Since the object area highlighted in CAM is usually small and sparse, enlarging the CAM activation becomes the focus of the WSSS task. For example, RDC [61] and DRS [28] propagate the object information to less discriminative regions via dilated convolution with varied rates and attention suppression of most discriminative areas, respectively. Popular contrastive learning [18, 48, 73] and self-supervised approaches [8, 51, 59] are also applied to facilitate object mining by improving representation learning. Moreover, cross-image information [21, 57, 72] is exploited to help the classifier discover more object patterns to obtain consistent and integral target regions. Besides attention expansion, Ahn and Kwak propose AffinityNet [2] for CAM refinement, which propagates semantic affinity via a random walk to recover delicate shapes of objects. IRN [1] is then proposed to exploit class boundary detection for discovering the entire instance areas with more accurate boundaries. + +# 2.2 Erasing-based Approaches + +Considering the original CAM technique usually leads to suboptimal performance due to failure to localize all object parts, erasing-based approaches are commonly employed to expand the highlighted region [30, 52, 60, 67, 69, 70]. They mask areas in a training image to force the network to seek other relevant parts. While direct erasing methods like Hide-and-Seek [52] hide image patches randomly, adversarial erasing-based approaches mask the most discriminative regions with the attention information, demonstrating more promising activation expansion potential. The first adversarial erasing-based work for WSSS is proposed by Wei et al. [60], which progressively discovers new object discriminative regions by erasing the currently activated area. Such a heuristic erasing strategy then attracts the attention of many researchers. For example, ACoL [69] proposes adversarial complementary learning with two adversary classifiers. Recently, Kweon et al. propose to apply the adversarial erasing framework to exploit the potential of the pre-trained classifier [30] and learn the concept of erasing with the triplet loss [67]. While adversarial erasing approaches achieve great success in enlarging the CAM activation, they tend to suffer from the over-expansion problem due to difficulty in determining when to stop erasing. Unlike the information removal strategy adopted in these approaches, we propose a simulated + +![](images/de5f5feb92fccd6629646b1a111135f69f07405507455e38d47c9b38f7d5dac3.jpg) +Fig. 2: The architecture of our proposed approach. We propose a simulated inter-image erasing (SIE) scenario where extra object information is introduced from another paired image. We then strengthen the object localization ability of the network by improving the consequent less activated localization map through learning object knowledge from the anchor image. A self-supervised regularization (SSR) module is also proposed to avoid weakening the anchor activation due to bidirectional alignment and improve the inter-class object boundary recognition for complex images. In addition, we propose a multi-granularity alignment (MGA) module to gently enlarge the object activation to further boost the object knowledge transfer. + +inter-image erasing scenario to weaken the activation by adding extra discriminative object information. The localization ability of network is then enhanced by improving the consequent less activated CAM through knowledge transfer. + +# 3 Method + +In this paper, we propose a knowledge transfer with simulated inter-image erasing approach for weakly supervised semantic segmentation to alleviate the over-expansion and under-activation problems of existing adversarial erasing-based methods. Our framework is illustrated in Fig. 2. As demonstrated in the anchor branch, we train a classification network with the given image-level weak labels, which consists of a backbone feature extractor and a pooling classification head. In contrast to existing adversarial erasing-based approaches that expand the activation through masking the most discriminative object part, as shown in the top-left of Fig. 2, we propose to simulate an inter-image erasing scenario where extra object information is introduced through concatenating a paired image. Object knowledge is then transferred from the anchor image to the consequent less activated localization map for strengthening the object localization ability of network. We also propose a self-supervised regularization module to avoid weakening the anchor activation due to bidirectional alignment and improve the inter-class object boundary recognition for complex multi-category images. In addition, we propose a multi-granularity alignment module to gently enlarge the object activation for boosting the object knowledge transfer. + +# 3.1 CAM Generation + +For the architecture of classification network, we follow the previous erasing-based work ACoL [69] and remove the final fully-connected layer. We set the output channel of the backbone to $C + 1$ , where $C$ is the number of foreground categories and 1 is added for the background. We can thus directly generate object localization maps from the class-aware CAM features $F$ in the forward pass to facilitate the object knowledge transfer. To obtain CAM for each foreground class $c$ , we feed the attention map $F^c$ into a ReLU layer and then normalize it to the range from 0 to 1: + +$$ +A ^ {c} = \frac {\operatorname {R e L U} \left(F ^ {c}\right)}{\operatorname* {m a x} \left(F ^ {c}\right)}. \tag {1} +$$ + +To improve the quality of the CAM by capturing the global context and local details of the image, we adopt the Gated Pyramid Pooling (GPP) layer [67] as the final pooling head. With the classification logits $q^{c}$ generated from the pooling head, we train the classification network with the multi-label soft margin loss as follows: + +$$ +\mathcal {L} _ {c l s} = - \frac {1}{C} \sum_ {c = 1} ^ {C} y ^ {c} \log \sigma \left(q ^ {c}\right) + \left(1 - y ^ {c}\right) \log \left[ 1 - \sigma \left(q ^ {c}\right) \right]. \tag {2} +$$ + +Here, $\sigma (\cdot)$ is the sigmoid function. $y^{c}$ is the image-level label for the $c$ -th class. + +# 3.2 Simulated Inter-Image Erasing + +Due to the sparsity of the CAM activation, previous adversarial erasing-based methods typically enlarge the object activation by maintaining the classification confidence on masked images or features, where the most discriminative regions are erased. However, due to the absence of guidelines on when to stop erasing, these approaches easily incur over-expansion with excessive removal and might also still suffer from under-activation with insufficient erasure. Therefore, we propose a knowledge transfer with simulated inter-image erasing approach for weakly supervised semantic segmentation to alleviate the above problem. + +In contrast to existing erasing-based approaches that remove the discriminative part for more object mining, we propose a simulated inter-image erasing (SIE) scenario to enable the network to benefit from introducing extra discriminative object information. Specifically, as shown in the top-left of Fig. 2, a larger synthetic image, created by concatenating the anchor image and another paired one, plays the role of the original image needing masking. Correspondingly, the anchor image can be treated as the erased one where the paired image is masked. Due to the introduction of extra discriminative object information, fewer object regions will be highlighted in the anchor part of the synthetic image. Therefore, we then strengthen the object localization ability of the network by transferring the more integral object knowledge from the original anchor CAM to the + +consequently less activated anchor part of the synthetic image. The knowledge transfer loss for the simulated inter-image erasing can be formulated as follows: + +$$ +\mathcal {L} _ {k t} = \operatorname {R e L U} \left(\hat {F} _ {a} - \hat {F} _ {s}\right), \quad \text {w h e r e} \hat {F} _ {a} = C F E \left(F _ {a}, y\right), \hat {F} _ {s} = C F E \left(F _ {s}, y\right). \tag {3} +$$ + +Here, $F_{a}$ and $F_{s}$ denote the CAM features of the anchor and simulated branches, respectively. CFE represents class feature extraction, which extracts the feature channel related to the classes that exist in the image. The ReLU operation means that our alignment focuses on the more highlighted object region rather than the background. Note that no rotation or resize is performed for the concatenated image. We only introduce extra object information from the paired image and control other variables to guarantee the weakened activation for the anchor part. Intuitively, our knowledge transfer with simulated inter-image erasing will not cause the over-expansion problem when the object in the anchor branch is well-discovered. Our experiments demonstrate that such knowledge transfer can effectively strengthen the localization ability of the network to alleviate the under-activation, especially when multiple object instances co-exist in the image. Besides, benefiting from the bidirectional alignment, the anchor branch will also learn from the weakened activation of the simulated one and produce compact attention maps for small objects to mitigate the over-expansion problem. + +# 3.3 Self-Supervised Regularization + +Our simulated inter-image erasing aims to improve the object localization ability of the network by maintaining the high activation given extra discriminative object information. However, the adopted knowledge transfer is bidirectional, which will also weaken the anchor branch object mining when learning from the sparse activation of the simulated branch. Since simply cutting the gradient propagation of the anchor branch makes the training unstable, we propose a self-supervised regularization module for the anchor CAM feature to maintain its reliable activation in discriminative regions. Specifically, with the generated CAM feature $F_{a}$ and the corresponding CAM $A_{a}$ , we first leverage two thresholds $\beta_{h} = 0.3$ and $\beta_{l} = 0.15$ to locate the confident foreground and background as follows (the subscript $a$ is omitted for simplicity): + +$$ +\hat {Y} _ {i, j} = \left\{ \begin{array}{l l} \operatorname {a r g m a x} \left(F _ {i, j,:}\right), & \text {i f} \max \left(A _ {i, j,:}\right) \geq \beta_ {h}, \\ 0, & \text {i f} \max \left(A _ {i, j,:}\right) \leq \beta_ {l}, \\ 2 5 5, & \text {o t h e r w i s e}, \end{array} \right. \tag {4} +$$ + +where 255 denote the ignored labels for uncertain regions and $\arg\max(\cdot)$ extracts the semantic class with the maximum activation value. After refining the pseudo label $\hat{Y}$ with the pixel-adaptive refinement module [47], we leverage it to directly supervise the learning of CAM features with the cross-entropy loss (coordinates $i, j$ are omitted): + +$$ +\mathcal {L} _ {c e} = - \sum_ {c = 0} ^ {C} \hat {Y} _ {b g} ^ {c} \log \Gamma (F) ^ {c} + \omega_ {f g} \hat {Y} _ {f g} ^ {c} \log \Gamma (F) ^ {c}. \tag {5} +$$ + +$\Gamma(\cdot)$ denotes the softmax function. We also leverage the pseudo label $\hat{Y}$ generated from the anchor branch to guide the learning of the simulated CAM feature with the same cross-entropy loss. However, we notice that the extracted labels of the foreground are quite noisy, especially for complex images with multiple categories of objects where the inter-class boundary is blurred. Here, we define images with objects of only one category as simple images and others containing two or more categories as complex ones. Therefore, we discount the influence of foreground label $\hat{Y}_{fg}$ with the weight of $\omega_{fg} = 0.0125$ . Besides, as shown in the right of Fig. 2, we design an inter-class loss to implicitly encourage the activation consistency for the multiple foreground classes in complex images as follows: + +$$ +\mathcal {L} _ {\text {i n t e r}} = \frac {\sum \left(V ^ {2 n d} - V ^ {\max }\right) \cdot M ^ {f}}{\left| M ^ {f} \right|}, \tag {6} +$$ + +Here, $M^f$ denotes the foreground mask obtained with a threshold $\beta_o = 0.2$ and $|\cdot|$ return the value of $l_1$ -norm. $V^{max}$ and $V^{2nd}$ denotes the largest and second largest activation value along the channel dimension of the CAM feature $F_a$ , which can be obtained as follows: + +$$ +V _ {i, j} ^ {m a x} = \max (F _ {i, j,:}), \quad F ^ {l e f t} = \mathsf {r e m o v e} (F, V _ {m a x}), \quad V _ {i, j} ^ {2 n d} = \max (F _ {i, j,:} ^ {l e f t}). \qquad (7) +$$ + +Eq. (6) improves the exclusivity of the class activation for each pixel, leading to more accurate boundaries between foreground objects. Our proposed self-supervised regularization module treats the simple and complex images differently, intending to maintain the reliable activation of the main branch without deteriorating the inter-class boundary caused by inaccurate pseudo labels. + +# 3.4 Multi-Granularity Alignment + +Our proposed self-supervised regularization module can effectively constrain and boost the knowledge transfer in our simulated inter-image erasing module. However, the improvement of the network's object localization ability will be severely limited by the quality of the anchor CAM, which is usually under-activated. Therefore, it is natural to resort to the traditional intra-image erasing, which aims to expand the object area. Following AEFT [67], we leverage a threshold of 0.6 to mask the most discriminative region. To avoid introducing too much background noise, we propose a multi-granularity alignment module to gently expand the object activation, which encourages the CAM expansion of the erased image and then transfers the learned object knowledge back to the anchor branch. As shown in the bottom of Fig. 2, we first input the anchor feature $F_{a}$ and masked feature $F_{m}$ into a class feature extraction module to obtain the corresponding features of the existing target categories $(\hat{F}_{a}$ and $\hat{F}_{m}$ ). Following the experimental finding of AEFT [67] that rigid classification supervision [69] for the masked branch tends to trigger the over-expansion problem in adversarial erasing, we resort to the soft class confidence guidance from the anchor branch to the masked one. Specifically, we directly adopt a global average pooling (GAP) operation + +to obtain the final class confidence for each branch and an image-level global alignment loss can be formulated as follows: + +$$ +\mathcal {L} _ {g l o b a l} = G A P \left(R e L U \left(\hat {F} _ {a}\right)\right) - G A P \left(R e L U \left(\hat {F} _ {m}\right)\right), +$$ + +$$ +\text {w h e r e} \quad \hat {F _ {m}} = C F E \left(F _ {m}, y\right). \tag {8} +$$ + +$\hat{F}_a$ and CFE have been defined in Eq. (3). Another difference with rigid classification supervision is that benefiting from the class feature extraction, our global alignment loss only focuses on the logit confidence for classes that exist in the image, enabling more effective and efficient gradient propagation. Unlike AEFT [67] that utilizes the feature space of the GPP layer as embedding space for loss construction, our experiments reveal that direct alignment with the CAM feature leads to more promising performance. With class information learned from the anchor branch, we further leverage a pixel-level local activation alignment to transfer the newly discovered object information from the erased image to guide the anchor branch learning. + +$$ +\mathcal {L} _ {\text {l o c a l}} = \operatorname {R e L U} \left(\hat {F _ {m}} - \hat {F _ {a}}\right). \tag {9} +$$ + +Such pixel-level local alignment can also hinder the erased branch from activating the unwanted background area by learning from the lowly-activated regions of the anchor branch. With the gentle and constrained activation enlargement, the multi-granularity alignment module further boosts the performance of our proposed simulated inter-image erasing approach. As can be seen in Fig. 2, our proposed SIE and MGA branches are designed in a symmetric way that play the adversarial role of expanding and constraining the anchor image activation, respectively. Their synergically cooperation facilitates the convergence for mining complete and compact object regions. + +# 3.5 Training Objective + +The overall training loss is as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {c l s} + \mathcal {L} _ {k t} + \mathcal {L} _ {g l o b a l} + \mathcal {L} _ {l o c a l} + \mathcal {L} _ {c e} + \lambda_ {i n t e r} \mathcal {L} _ {i n t e r}. \tag {10} +$$ + +We empirically set $\lambda_{inter} = 0.005$ as the hyperparameter that controls the weight of inter-class loss. + +# 4 Experiment + +# 4.1 Datasets and Implementation Details + +We evaluate our approach on the PASCAL VOC 2012 [19] and COCO [38] datasets. The PASCAL VOC 2012 dataset contains 21 classes (20 object categories and the background) for semantic segmentation. It has 10,582 images for + +Table 1: Accuracy of pseudo-masks evaluated on PASCAL VOC 2012 training set. + +
MethodsSeedw/IRN [1]
IRN [1]CVPR1948.366.3
MBMNet [40]MM2050.266.8
CONTA [68]NIPS2048.867.9
AdvCAM [32]CVPR2155.669.9
RIB [31]NIPS2156.670.6
ReCAM [14]CVPR2256.670.5
ESOL [35]NIPS2253.668.7
CLIMS [63]CVPR2256.670.5
AEFT [67]ECCV2256.071.0
ACR [15]CVPR2360.372.3
FPR [6]ICCV2363.8-
KTSE (Ours)67.073.8
+ +Table 2: Quantitative comparisons on PASCAL VOC 2012 val and test sets with VGG backbone. Sup: Supervision, I: Image-level label, S: Saliency maps. + +
MethodsSupValTest
DRS [28]AAAI21I+S63.664.4
GSM [36]AAAI21I+S63.363.6
NSROM [66]CVPR21I+S65.565.3
EPS [34]CVPR21I+S67.067.3
L2G [24]CVPR22I+S68.568.9
AffinityNet [2]CVPR18I58.460.5
ICD [20]CVPR20I61.260.9
BES [7]ECCV20I60.161.1
ECS [55]ICCV21I62.163.4
KTSE (Ours)I67.367.0
+ +training (expanded with SBD [22]), 1,449 for validation and 1,456 for testing. The COCO dataset is a more challenging benchmark with 80 semantic classes and the background. Following previous works [36, 58, 68], we use the default train/val splits (80k images for training and 40k for validation) in the experiment. Mean intersection over union (mIoU) is adopted as the metric to evaluate the quality of our pseudo labels and segmentation results. + +For the classification network, we follow the work of AEFT [67] and employ ResNet38 [62] as the backbone. We adopt a poly learning rate with an initial value of $10^{-2}$ and a power of 0.9. For the second stage training of WSSS, following the recent work of BECO [45], we adopt DeeplabV2 [5] as the segmentation network, which uses ResNet101 [23] as the backbone. The momentum and weight decay of the SGD optimizer are 0.9 and $10^{-4}$ . The initial learning rate is set to $10^{-2}$ and is decreased using polynomial decay. The segmentation model is trained for 80 epochs and 40 epochs on VOC and COCO datasets, respectively, with a common batch size of 16. We also follow the default setting of DeeplabV2 [5] and conduct experiments with VGG16 backbone for a more comprehensive comparison with previous approaches [7, 20, 24, 34, 55]. All our backbones are pre-trained on ImageNet [17]. + +# 4.2 Comparisons to the State-of-the-arts + +Accuracy of Pseudo-Masks. We first report the quality of the segmentation seeds and the generated pseudo-masks derived from our approach. The comparison with other state-of-the-arts is presented in Tab. 1. As can be seen, our segmentation seed can reach the mIoU of $67.0\%$ , bringing a gain of $18.7\%$ compared to the baseline reported by IRN [1]. Our method can outperform the state-of-the-art method FPR [6] by $3.2\%$ . With the further refinement of IRN [1], the mIoU of our generated pseudo-masks can arrive at $73.8\%$ , surpassing the previous SOTA of AEFT [67] and ACR [15] by more than $1.5\%$ . + +Table 3: Quantitative comparisons on PASCAL VOC 2012 val and test sets with ResNet backbone. Sup: Supervision, I: Image-level label, S: Saliency maps, O: Out-of-distribution data, L: Language. + +
MethodsSupValTest
DRS [28]AAAI21I+S71.271.4
NSROM [66]CVPR21I+S70.470.2
EPS [34]CVPR21I+S71.071.8
AuxSeg [64]ICCV21I+S69.068.6
PPC [18]CVPR22I+S72.673.6
RCA [73]CVPR22I+S72.272.8
L2G [24]CVPR22I+S72.171.7
W-OoD [33]CVPR22I+O69.869.9
CLIP-ES [39]CVPR23I+L71.171.4
LPCAM [13]CVPR23I+S71.872.1
AdvCAM [32]CVPR21I68.168.0
CDA [54]ICCV21I66.166.8
CSE [30]ICCV21I68.468.2
RIB [31]NIPS21I68.368.6
AMR [44]AAAI22I68.869.1
MCT [65]CVPR22I71.971.6
AFA [47]CVPR22I66.066.3
SIPE [8]CVPR22I68.869.7
ReCAM [14]CVPR22I68.568.4
PPC [18]CVPR22I67.767.4
ViT-PCM [46]ECCV22I70.370.9
S-BCE [46]ECCV22I70.071.3
AEFT [67]ECCV22I70.971.7
TOCO [48]CVPR23I69.870.5
OCR [15]CVPR23I72.772.0
ACR [15]CVPR23I71.971.9
BECO [45]CVPR23I72.171.8
FPR [6]ICCV23I70.370.1
MARS [26]ICCV23I70.371.2
KTSE (Ours)I73.072.9
+ +Table 4: Quantitative comparisons on COCO val set with VGG backbone. Sup: Supervision, I: Image-level label, S: Saliency maps. + +
MethodsSupVal
IAL [58]IJCV20I+S27.7
GWSM [36]AAAI21I+S28.4
EPS [34]CVPR21I+S35.7
I2CRC [73]TMM22I+S31.2
RCA [73]CVPR22I+S36.8
MDBA [11]TIP23I+S37.8
BFBP [49]ECCV16I20.4
SEC [29]ECCV16I22.4
CONTA [68]NIPS20I23.7
KTSE (Ours)I37.2
+ +Table 5: Quantitative comparisons on COCO val set with ResNet backbone. Sup: Supervision, I: Image-level label, S: Saliency maps, L: Language. + +
MethodsSupVal
L2G [24]CVPR22I+S44.2
CLIP-ES [39]CVPR23I+L45.4
LPCAM [13]CVPR23I+S42.1
OC-CSE [30]ICCV21I36.4
CDA [54]ICCV21I33.2
RIB [31]NIPS21I43.8
MCT [65]CVPR22I42.0
SIPE [8]CVPR22I40.6
TOCO [48]CVPR23I41.3
OCR [15]CVPR23I42.5
ACR [15]CVPR23I45.3
BECO [45]CVPR23I45.1
FPR [6]ICCV23I43.9
USAGE [43]ICCV23I44.3
KTSE (Ours)I45.9
+ +Accuracy of Segmentation Maps on PASCAL VOC 2012. Our segmentation results on the PASCAL VOC 2012 dataset with the backbone of VGG and ResNet are demonstrated in Tab. 2 and Tab. 3, respectively. As can be seen, with the VGG backbone, our approach achieves the performance of $67.3\%$ on the validation set and $67.0\%$ on the test set, better than other state-of-the-art methods with only image-level labels. Our segmentation results are also competitive to many approaches that rely on saliency maps, e.g., NSROM [66] and EPS [34]. With the ResNet backbone, we can improve the results to $73.0\%$ and $72.9\%$ on + +Table 6: Element-wise component analysis. SIE: Simulated Inter-Image Erasing, SSR: Self-Supervised Regularization, MGA: Multi-Granularity Alignment. + +
BaseSIESSRMGAmIoU
✓54.2
✓✓57.1
✓✓57.4
✓✓60.2
✓✓✓63.8
✓✓✓✓67.0
+ +Table 7: Comparison of Our Multi-Granularity Alignment with Previous Erasing-based Approaches. GA: Global Alignment; LA: Local Alignment. + +
MethodmIoU
Base54.2
Rigid Classification [69]55.6
Soft GPP Feature [67]54.9
Our GA58.7
Our GA & LA60.2
+ +the validation and test sets, respectively. As can be seen in Tab. 3, our proposed approach can achieve better performance compared with recent SOTA methods. For example, our approach outperforms OCR [15] and ACR [15] by about $1\%$ on the test set, demonstrating the superiority of our knowledge transfer with simulated inter-image erasing. + +Accuracy of Segmentation Maps on COCO. For the more challenging COCO dataset, we provide performance comparisons with state-of-the-art WSSS methods using the backbone of VGG and ResNet in Tab. 4 and Tab. 5, respectively. As shown in Tab. 4, our proposed KTSE with VGG backbone can achieve the performance of $37.2\%$ mIoU, much better than previous methods supervised with only image-level labels, e.g., $13.5\%$ mIoU higher than CONTA [68]. Besides, our results are also competitive compared to previous SOTA methods with additional saliency guidance like RCA [73] and MDBA [11]. With the ResNet backbone, our proposed KTSE reaches the best result of $45.9\%$ mIoU compared to previous SOTA WSSS methods. Specifically, our approach outperforms ACR [15] and BECO [45] by $0.6\%$ and $0.8\%$ mIoU, respectively. + +# 4.3 Ablation Studies + +Element-Wise Component Analysis. In this part, we demonstrate the contribution of each component proposed in our approach to improving the quality of pseudo-masks. As mentioned in Sec. 3.1, we incorporate the GPP [67] into the classification network as our strong baseline. As shown in Tab. 6, with our simulated inter-image erasing (SIE) module, we can improve the accuracy of segmentation seeds from the baseline of $54.2\%$ to $57.1\%$ , which demonstrates the benefit of strengthening the localization ability of the network through weakening the activation with extra object information and then learning from the anchor branch. By maintaining reliable activation in discriminative regions, our self-supervised regularization (SSR) module can achieve the mIoU of $57.4\%$ . Moreover, the combination of our proposed SIE and SSR can significantly boost the performance to $63.8\%$ mIoU. With our proposed multi-granularity alignment module to gently expand the object activation, the mIoU of pseudo masks finally arrives at $67.0\%$ , highlighting the importance of improving the quality of + +![](images/c35b984ca6892c1d4ed49a980056c572e34d7615035221993edc7a9d95ead474.jpg) +Fig. 3: Example localization maps on the PASCAL VOC 2012 training set. For each (a) image, we show (b) ground truth, localization maps produced by (c) previous work of AEFT [67], (d) our baseline, (e) baseline + SIE, (f) baseline + SIE + SSR, and (g) baseline + SIE + SSR + MGA. Best viewed in color. + +anchor CAM to boost the performance of our knowledge transfer with simulated inter-image erasing. + +Some example localization maps on the PASCAL VOC 2012 training set can be viewed in Fig. 3. As seen from the first two rows, after strengthening the localization ability of the network with our simulated inter-image erasing (SIE) module, we can successfully expand the object activation and discover the less discriminative object in the image (e.g., the smaller animals on the right). We can notice that our SIE can also help alleviate the over-expansion problem and generate more compact activation, e.g., the cow in the first row. However, as demonstrated in the last two rows, the bidirectional alignment in SIE will also lead to more severe under-activation when the anchor branch learns from the sparse activation of the simulated one, especially for larger objects. Fortunately, our proposed self-supervised regularization (SSR) module can help maintain reliable activation in discriminative regions to guarantee effective knowledge transfer. Moreover, as seen in the last row, our multi-granularity alignment (MGA) module can further help discover the integral target region by gently expanding the object activation. Comparing the results of ours with that of AEFT [67], Fig. 3 (g) vs Fig. 3 (c), we can see that our proposed approach can significantly alleviate the under-activation and over-expansion of previous approaches due to insufficient or excessive erasure. + +MGA vs. Previous Erasing-based Approaches. For the multi-granularity alignment, we resort to the soft class confidence knowledge from the anchor branch to guide the gentle activation expansion of the masked image. In Tab. 7, we compare performance improvement from different guidance settings. As can be seen, our global alignment of CAM feature can improve the baseline from $54.2\%$ to $58.7\%$ , outperforming the rigid classification guidance in ACoL [69] and GPP feature alignment adopted in AEFT [67] by $3.1\%$ and $3.8\%$ , respectively. It highlights the importance of direct global alignment in CAM features + +rather than embedding space with extra projection layers. With our pixel-level local alignment, our multi-granularity alignment module can finally improve the performance to $60.2\%$ , demonstrating the benefit of our gentle alignment strategy compared with previous erasing-based approaches. + +Phenomenon of SIE. Based on the observation of classic adversarial erasing methods that removing the most discriminative region will lead to the extra activation of other object parts, the reversal thinking comes naturally—what if we introduce extra discriminative object information? The answer is the original highly activated area will become less discriminative and be lowly activated. We then enhance this weakened activation by learning from the original CAM. When the network learns to increase the attention for the less-discriminative region in the concatenated image, it will also learn to activate less-discriminative object region in the original image to locate more objects. + +SIE vs. Data Augmentation. Though data augmentations like CutMix will also change the anchor image and lead to activation perturbation, they do not guarantee the weakened activation (which might cause over-expansion) for the anchor part of the concatenated image like ours. In contrast, we only introduce extra object information from the paired image and control other variables. Most importantly, our novelty lies in constructing the simulated inter-image erasing scenario (reversal thinking on existing masking methods) rather than simply augmenting the data input. It enhances the network's localization ability by improving the consequent less activated attention map through learning from the object knowledge of the anchor branch. Accompanied by our proposed MGA, they are designed in a symmetric way that play the adversarial role of expanding and constraining the anchor image activation respectively for mining complete and compact object regions. Note that our approach significantly outperforms the CutMix-based augmentation method of CDA [54] with 73.0 vs. 66.1 on VOC (Tab. 3) and 45.9 vs. 33.2 on COCO (Tab. 5). + +# 5 Conclusion + +This paper proposed a Knowledge Transfer with Simulated Inter-Image Erasing (KTSE) approach for weakly supervised semantic segmentation. Specifically, in contrast to existing adversarial erasing-based methods that remove the discriminative part for mining less-discriminative areas, we proposed to simulate an inter-image erasing scenario where extra object information was added through concatenating a paired image. We then strengthened the object localization ability of the network by enhancing the consequent less activated localization map. In addition, we proposed a self-supervised regularization module to maintain reliable activation in the discriminative regions and improve the inter-class object boundary recognition for complex images. Moreover, we proposed a multi-granularity alignment module to gently enlarge the object activation via intra-image erasing for boosting the object knowledge transfer. Extensive experiments and ablation studies were conducted on PASCAL VOC 2012 and COCO datasets to demonstrate the superiority of our proposed approach. + +# Acknowledgements + +This work was supported by the National Natural Science Foundation of China (No. 62202227 and 62102182), Natural Science Foundation of Jiangsu Province (No. BK20220938 and BK20220934), China Postdoctoral Science Foundation (No. 2022M711635), Jiangsu Funding Program for Excellent Postdoctoral Talent (No. 2022ZB267), Fundamental Research Funds for the Central Universities (No. 30923010303). + +# References + +1. Ahn, J., Cho, S., Kwak, S.: Weakly supervised learning of instance segmentation with inter-pixel relations. In: CVPR. pp. 2209-2218 (2019) +2. Ahn, J., Kwak, S.: Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In: CVPR. pp. 4981-4990 (2018) +3. Bearman, A., Russakovsky, O., Ferrari, V., Fei-Fei, L.: What's the point: Semantic segmentation with point supervision. In: ECCV. pp. 549-565 (2016) +4. Cai, X., Lai, Q., Wang, Y., Wang, W., Sun, Z., Yao, Y.: Poly kernel inception network for remote sensing detection. In: CVPR. pp. 27706-27716 (2024) +5. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. In: IEEE TPAMI. vol. 40, pp. 834-848 (2017) +6. Chen, L., Lei, C., Li, R., Li, S., Zhang, Z., Zhang, L.: Fpr: False positive rectification for weakly supervised semantic segmentation. In: ICCV. pp. 1108-1118 (2023) +7. Chen, L., Wu, W., Fu, C., Han, X., Zhang, Y.T.: Weakly supervised semantic segmentation with boundary exploration. In: ECCV (2020) +8. Chen, Q., Yang, L., Lai, J.H., Xie, X.: Self-supervised image-specific prototype exploration for weakly supervised semantic segmentation. In: CVPR. pp. 4288-4298 (2022) +9. Chen, T., Xie, G.S., Yao, Y., Wang, Q., Shen, F., Tang, Z., Zhang, J.: Semantically meaningful class prototype learning for one-shot image segmentation. IEEE TMM 24, 968-980 (2021) +0. Chen, T., Yao, Y., Huang, X., Li, Z., Nie, L., Tang, J.: Spatial structure constraints for weakly supervised semantic segmentation. IEEE TIP 33, 1136-1148 (2024) +1. Chen, T., Yao, Y., Tang, J.: Multi-granularity denoising and bidirectional alignment for weakly supervised semantic segmentation. IEEE TIP 32, 2960-2971 (2023) +2. Chen, T., Yao, Y., Zhang, L., Wang, Q., Xie, G.S., Shen, F.: Saliency guided inter- and intra-class relation constraints for weakly supervised semantic segmentation. IEEE TMM 25, 1727–1737 (2022) +3. Chen, Z., Sun, Q.: Extracting class activation maps from non-discriminative features as well. In: CVPR. pp. 3135-3144 (2023) +4. Chen, Z., Wang, T., Wu, X., Hua, X.S., Zhang, H., Sun, Q.: Class re-activation maps for weakly-supervised semantic segmentation. In: CVPR. pp. 969–978 (2022) +5. Cheng, Z., Qiao, P., Li, K., Li, S., Wei, P., Ji, X., Yuan, L., Liu, C., Chen, J.: Out-of-candidate rectification for weakly supervised semantic segmentation. In: CVPR. pp. 23673-23684 (2023) + +16. Dai, J., He, K., Sun, J.: Boxsup: Exploiting bounding boxes to supervise convolutional networks for semantic segmentation. In: ICCV. pp. 1635-1643 (2015) +17. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255 (2009) +18. Du, Y., Fu, Z., Liu, Q., Wang, Y.: Weakly supervised semantic segmentation by pixel-to-prototype contrast. In: CVPR. pp. 4320-4329 (2022) +19. Everingham, M., Gool, L.V., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. In: IJCV. vol. 88, pp. 303-338 (2010) +20. Fan, J., Zhang, Z., Song, C., Tan, T.: Learning integral objects with intra-class discriminator for weakly-supervised semantic segmentation. In: CVPR. pp. 4283-4292 (2020) +21. Fan, J., Zhang, Z., Tan, T., Song, C., Xiao, J.: Cian: Cross-image affinity net for weakly supervised semantic segmentation. In: AAAI. vol. 34, pp. 10762-10769 (2020) +22. Hariharan, B., Arbeláez, P., Bourdev, L.D., Maji, S., Malik, J.: Semantic contours from inverse detectors. In: ICCV. pp. 991-998 (2011) +23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016) +24. Jiang, P.T., Yang, Y., Hou, Q., Wei, Y.: L2g: A simple local-to-global knowledge transfer framework for weakly supervised semantic segmentation. In: CVPR. pp. 16886-16896 (2022) +25. Jing, L., Chen, Y., Tian, Y.: Coarse-to-fine semantic segmentation from image-level labels. IEEE TIP 29, 225-236 (2019) +26. Jo, S., Yu, I.J., Kim, K.: Mars: Model-agnostic biased object removal without additional supervision for weakly-supervised semantic segmentation. arXiv preprint arXiv:2304.09913 (2023) +27. Khoreva, A., Benenson, R., Hosang, J., Hein, M., Schiele, B.: Simple does it: Weakly supervised instance and semantic segmentation. In: CVPR. pp. 876-885 (2017) +28. Kim, B., Han, S., Kim, J.: Discriminative region suppression for weakly-supervised semantic segmentation. In: AAAI. vol. 35, pp. 1754-1761 (2021) +29. Kolesnikov, A., Lampert, C.: Seed, expand and constrain: Three principles for weakly-supervised image segmentation. In: ECCV. pp. 695-711 (2016) +30. Kweon, H., Yoon, S.H., Kim, H., Park, D., Yoon, K.J.: Unlocking the potential of ordinary classifier: Class-specific adversarial erasing framework for weakly supervised semantic segmentation. In: ICCV. pp. 6994-7003 (2021) +31. Lee, J., Choi, J., Mok, J., Yoon, S.: Reducing information bottleneck for weakly supervised semantic segmentation. In: NeurIPS. vol. 34, pp. 27408-27421 (2021) +32. Lee, J., Kim, E., Yoon, S.: Anti-adversarily manipulated attributions for weakly and semi-supervised semantic segmentation. In: CVPR. pp. 4071-4080 (2021) +33. Lee, J., Oh, S.J., Yun, S., Choe, J., Kim, E., Yoon, S.: Weakly supervised semantic segmentation using out-of-distribution data. In: CVPR. pp. 16897-16906 (2022) +34. Lee, S., Lee, M., Lee, J., Shim, H.: Railroad is not a train: Saliency as pseudo-pixel supervision for weakly supervised semantic segmentation. In: CVPR. pp. 5495-5505 (2021) +35. Li, J., Jie, Z., Wang, X., Wei, X., Ma, L.: Expansion and shrinkage of localization for weakly-supervised semantic segmentation. vol. 35, pp. 16037-16051 (2022) +36. Li, X., Zhou, T., Li, J., Zhou, Y., Zhang, Z.: Group-wise semantic mining for weakly supervised semantic segmentation. In: AAAI. pp. 1984-1992 (2021) +37. Lin, D., Dai, J., Jia, J., He, K., Sun, J.: Scribblesup: Scribble-supervised convolutional networks for semantic segmentation. In: CVPR. pp. 3159-3167 (2016) + +38. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, L.: Microsoft coco: Common objects in context. In: ECCV. pp. 740-755 (2014) +39. Lin, Y., Chen, M., Wang, W., Wu, B., Li, K., Lin, B., Liu, H., He, X.: Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In: CVPR. pp. 15305-15314 (2023) +40. Liu, W., Zhang, C., Lin, G., Hung, T.Y., Miao, C.: Weakly supervised segmentation with maximum bipartite graph matching. In: ACM MM. pp. 2085-2094 (2020) +41. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR. pp. 3431-3440 (2015) +42. Pei, G., Chen, T., Jiang, X., Liu, H., Sun, Z., Yao, Y.: Videomac: Video masked autoencoders meet convnets. In: CVPR. pp. 22733-22743 (2024) +43. Peng, Z., Wang, G., Xie, L., Jiang, D., Shen, W., Tian, Q.: Usage: A unified seed area generation paradigm for weakly supervised semantic segmentation. arXiv preprint arXiv:2303.07806 (2023) +44. Qin, J., Wu, J., Xiao, X., Li, L., Wang, X.: Activation modulation and recalibration scheme for weakly supervised semantic segmentation. In: AAAI. vol. 36, pp. 2117-2125 (2022) +45. Rong, S., Tu, B., Wang, Z., Li, J.: Boundary-enhanced co-training for weakly supervised semantic segmentation. In: CVPR. pp. 19574-19584 (2023) +46. Rossetti, S., Zappia, D., Sanzari, M., Schaerf, M., Pirri, F.: Max pooling with vision transformers reconciles class and shape in weakly supervised semantic segmentation. In: ECCV. pp. 446-463. Springer (2022) +47. Ru, L., Zhan, Y., Yu, B., Du, B.: Learning affinity from attention: End-to-end weakly-supervised semantic segmentation with transformers. In: CVPR. pp. 16846-16855 (2022) +48. Ru, L., Zheng, H., Zhan, Y., Du, B.: Token contrast for weakly-supervised semantic segmentation. In: CVPR. pp. 3093-3102 (2023) +49. Saleh, F., Aliakbarian, M.S., Salzmann, M., Petersson, L., Gould, S., Alvarez, J.M.: Built-in foreground/background prior for weakly-supervised semantic segmentation. In: ECCV. pp. 413-432 (2016) +50. Sheng, M., Sun, Z., Cai, Z., Chen, T., Zhou, Y., Yao, Y.: Adaptive integration of partial label learning and negative learning for enhanced noisy label learning. In: AAAI. vol. 38, pp. 4820-4828 (2024) +51. Shimoda, W., Yanai, K.: Self-supervised difference detection for weakly-supervised semantic segmentation. In: ICCV. pp. 5208-5217 (2019) +52. Singh, K.K., Yu, H., Sarmasi, A., Pradeep, G., Lee, Y.J.: Hide-and-seek: A data augmentation technique for weakly-supervised localization and beyond. arXiv preprint arXiv:1811.02545 (2018) +53. Song, C., Huang, Y., Ouyang, W., Wang, L.: Box-driven class-wise region masking and filling rate guided loss for weakly supervised semantic segmentation. In: CVPR. pp. 3136-3145 (2019) +54. Su, Y., Sun, R., Lin, G., Wu, Q.: Context decoupling augmentation for weakly supervised semantic segmentation. In: ICCV (2021) +55. Sun, K., Shi, H., Zhang, Z., Huang, Y.: ECS-net: Improving weakly supervised semantic segmentation by using connections between class activation maps. In: ICCV. pp. 7283-7292 (2021) +56. Vernaza, P., Chandraker, M.: Learning random-walk label propagation for weakly-supervised semantic segmentation. In: CVPR. pp. 7158-7166 (2017) +57. Wang, W., Sun, G., Van Gool, L.: Looking beyond single images for weakly supervised semantic segmentation learning. IEEE TPAMI (2022) + +58. Wang, X., Liu, S., Ma, H., Yang, M.H.: Weakly-supervised semantic segmentation by iterative affinity learning. vol. 128, pp. 1736-1749. Springer (2020) +59. Wang, Y., Zhang, J., Kan, M., Shan, S., Chen, X.: Self-supervised equivariant attention mechanism for weakly supervised semantic segmentation. In: CVPR. pp. 12275-12284 (2020) +60. Wei, Y., Feng, J., Liang, X., Cheng, M.M., Zhao, Y., Yan, S.: Object region mining with adversarial erasing: A simple classification to semantic segmentation approach. In: CVPR. pp. 1568-1576 (2017) +61. Wei, Y., Xiao, H., Shi, H., Jie, Z., Feng, J., Huang, T.: Revisiting dilated convolution: A simple approach for weakly-and semi-supervised semantic segmentation. In: CVPR. pp. 7268-7277 (2018) +62. Wu, Z., Shen, C., Van Den Hengel, A.: Wider or deeper: Revisiting the resnet model for visual recognition. PR 90, 119-133 (2019) +63. Xie, J., Hou, X., Ye, K., Shen, L.: Clims: Cross language image matching for weakly supervised semantic segmentation. In: CVPR. pp. 4483-4492 (2022) +64. Xu, L., Ouyang, W., Bennamoun, M., Boussaid, F., Sohel, F., Xu, D.: Leveraging auxiliary tasks with affinity learning for weakly supervised semantic segmentation. In: ICCV. pp. 6984-6993 (2021) +65. Xu, L., Ouyang, W., Bennamoun, M., Boussaid, F., Xu, D.: Multi-class token transformer for weakly supervised semantic segmentation. In: CVPR. pp. 4310-4319 (2022) +66. Yao, Y., Chen, T., Xie, G.S., Zhang, C., Shen, F., Wu, Q., Tang, Z., Zhang, J.: Non-salient region object mining for weakly supervised semantic segmentation. In: CVPR. pp. 2623-2632 (2021) +67. Yoon, S.H., Kweon, H., Cho, J., Kim, S., Yoon, K.J.: Adversarial erasing framework via triplet with gated pyramid pooling layer for weakly supervised semantic segmentation. In: ECCV. pp. 326-344. Springer (2022) +68. Zhang, D., Zhang, H., Tang, J., Hua, X.S., Sun, Q.: Causal intervention for weakly-supervised semantic segmentation. In: NeurIPS. vol. 33 (2020) +69. Zhang, X., Wei, Y., Feng, J., Yang, Y., Huang, T.S.: Adversarial complementary learning for weakly supervised object localization. In: CVPR. pp. 1325-1334 (2018) +70. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. In: AAAI. vol. 34, pp. 13001-13008 (2020) +71. Zhou, B., Khosla, A., Lapedriza, Å., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: CVPR. pp. 2921-2929 (2016) +72. Zhou, T., Li, L., Li, X., Feng, C.M., Li, J., Shao, L.: Group-wise learning for weakly supervised semantic segmentation. IEEE TIP 31, 799-811 (2021) +73. Zhou, T., Zhang, M., Zhao, F., Li, J.: Regional semantic contrast and aggregation for weakly supervised semantic segmentation. In: CVPR. pp. 4299-4309 (2022) \ No newline at end of file diff --git a/2024/Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation/images.zip b/2024/Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0749f4ecf79f1b0f2a828201dde63de1f72212f4 --- /dev/null +++ b/2024/Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9803bbdeefccbbfef108a380de71cf4378f0660e61ead1a1911726c76fc3028 +size 542125 diff --git a/2024/Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation/layout.json b/2024/Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b59b7a0812414711ce425548e512f6b159630487 --- /dev/null +++ b/2024/Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation/layout.json @@ -0,0 +1,8982 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 139, + 111, + 477, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 111, + 477, + 166 + ], + "spans": [ + { + "bbox": [ + 139, + 111, + 477, + 166 + ], + "type": "text", + "content": "Knowledge Transfer with Simulated Inter-Image Erasing for Weakly Supervised Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 185, + 186, + 429, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 186, + 429, + 211 + ], + "spans": [ + { + "bbox": [ + 185, + 186, + 429, + 211 + ], + "type": "text", + "content": "Tao Chen" + }, + { + "bbox": [ + 185, + 186, + 429, + 211 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 185, + 186, + 429, + 211 + ], + "type": "text", + "content": ", Xiruo Jiang" + }, + { + "bbox": [ + 185, + 186, + 429, + 211 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 185, + 186, + 429, + 211 + ], + "type": "text", + "content": ", Gensheng Pei" + }, + { + "bbox": [ + 185, + 186, + 429, + 211 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 185, + 186, + 429, + 211 + ], + "type": "text", + "content": ", Zeren Sun" + }, + { + "bbox": [ + 185, + 186, + 429, + 211 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 185, + 186, + 429, + 211 + ], + "type": "text", + "content": ", Yucheng Wang" + }, + { + "bbox": [ + 185, + 186, + 429, + 211 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 185, + 186, + 429, + 211 + ], + "type": "text", + "content": ", and Yazhou Yao" + }, + { + "bbox": [ + 185, + 186, + 429, + 211 + ], + "type": "inline_equation", + "content": "^{1(\\text{忍})}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 160, + 220, + 455, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 220, + 455, + 266 + ], + "spans": [ + { + "bbox": [ + 160, + 220, + 455, + 266 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 160, + 220, + 455, + 266 + ], + "type": "text", + "content": " Nanjing University of Science and Technology, Nanjing 210094, China {taochen,jiangxiruo622,peigsh,zerens,yazhou.yao}@njust.edu.cn " + }, + { + "bbox": [ + 160, + 220, + 455, + 266 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 160, + 220, + 455, + 266 + ], + "type": "text", + "content": " Horizon Robotics, Beijing 100089, China yucheng.wang@horizon.cc" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 296, + 455, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 296, + 455, + 549 + ], + "spans": [ + { + "bbox": [ + 160, + 296, + 455, + 549 + ], + "type": "text", + "content": "Abstract. Though adversarial erasing has prevailed in weakly supervised semantic segmentation to help activate integral object regions, existing approaches still suffer from the dilemma of under-activation and over-expansion due to the difficulty in determining when to stop erasing. In this paper, we propose a Knowledge Transfer with Simulated InterImage Erasing (KTSE) approach for weakly supervised semantic segmentation to alleviate the above problem. In contrast to existing erasing-based methods that remove the discriminative part for more object discovery, we propose a simulated inter-image erasing scenario to weaken the original activation by introducing extra object information. Then, object knowledge is transferred from the anchor image to the consequent less activated localization map to strengthen network localization ability. Considering the adopted bidirectional alignment will also weaken the anchor image activation if appropriate constraints are missing, we propose a self-supervised regularization module to maintain the reliable activation in discriminative regions and improve the inter-class object boundary recognition for complex images with multiple categories of objects. In addition, we resort to intra-image erasing and propose a multi-granularity alignment module to gently enlarge the object activation to boost the object knowledge transfer. Extensive experiments and ablation studies on PAS-CAL VOC 2012 and COCO datasets demonstrate the superiority of our proposed approach. Codes and models are available at https://nust-machine-intelligence-laboratory.github.io/project-KTSE." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 559, + 459, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 559, + 459, + 583 + ], + "spans": [ + { + "bbox": [ + 160, + 559, + 459, + 583 + ], + "type": "text", + "content": "Keywords: Weakly Supervised Learning " + }, + { + "bbox": [ + 160, + 559, + 459, + 583 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 559, + 459, + 583 + ], + "type": "text", + "content": " Semantic Segmentation " + }, + { + "bbox": [ + 160, + 559, + 459, + 583 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 559, + 459, + 583 + ], + "type": "text", + "content": " InterImage Erasing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 603, + 230, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 603, + 230, + 616 + ], + "spans": [ + { + "bbox": [ + 133, + 603, + 230, + 616 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "With huge progress in the era of deep learning, semantic segmentation has been widely applied in fields like autonomous driving and image editing [5,41]. Since deep learning thus far is data-driven, the model training typically involves a large" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 304, + 148 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 304, + 148 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 304, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 304, + 148 + ], + "type": "image", + "image_path": "58f2061d0629499ba4986e9c776a3ac1bfa5f966e43e207d7b25862355899a78.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 149, + 261, + 157 + ], + "lines": [ + { + "bbox": [ + 168, + 149, + 261, + 157 + ], + "spans": [ + { + "bbox": [ + 168, + 149, + 261, + 157 + ], + "type": "text", + "content": "(a) Previous Adversarial Erasing" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 158, + 304, + 191 + ], + "blocks": [ + { + "bbox": [ + 133, + 158, + 304, + 191 + ], + "lines": [ + { + "bbox": [ + 133, + 158, + 304, + 191 + ], + "spans": [ + { + "bbox": [ + 133, + 158, + 304, + 191 + ], + "type": "image", + "image_path": "9de728cea173eb22073b8f9b7700208595d731074360f0bde3e7937f5d92c335.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 191, + 262, + 200 + ], + "lines": [ + { + "bbox": [ + 168, + 191, + 262, + 200 + ], + "spans": [ + { + "bbox": [ + 168, + 191, + 262, + 200 + ], + "type": "text", + "content": "(c) Result CAM vs Ground Truth" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 208, + 482, + 275 + ], + "lines": [ + { + "bbox": [ + 130, + 208, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 208, + 482, + 275 + ], + "type": "text", + "content": "Fig. 1: (a) Previous adversarial erasing-based approaches typically suffer from the over-expansion problem, which is hard to constrain. (b) Different from their information removal strategy, we propose to add extra object knowledge from a paired image to weaken the current object activation. The localization ability of the network is then enhanced by improving the consequent less activated attention map through learning from the object knowledge of the anchor branch. (C) Result comparison." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 114, + 481, + 152 + ], + "blocks": [ + { + "bbox": [ + 306, + 114, + 481, + 152 + ], + "lines": [ + { + "bbox": [ + 306, + 114, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 481, + 152 + ], + "type": "image", + "image_path": "577d128e86287d349d5eb7649899060911529a0aafb216ca283234d19d0a2355.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 153, + 481, + 190 + ], + "blocks": [ + { + "bbox": [ + 306, + 153, + 481, + 190 + ], + "lines": [ + { + "bbox": [ + 306, + 153, + 481, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 153, + 481, + 190 + ], + "type": "image", + "image_path": "4bea42c8c3cc3ec67b863124745c05367d582458fc549d5fa261ebe703ba6e73.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 190, + 448, + 200 + ], + "lines": [ + { + "bbox": [ + 339, + 190, + 448, + 200 + ], + "spans": [ + { + "bbox": [ + 339, + 190, + 448, + 200 + ], + "type": "text", + "content": "(b) Our Simulated Inter-Image Erasing" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 302, + 482, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 482, + 374 + ], + "type": "text", + "content": "abundance of labeled images. However, collecting accurate pixel-level annotation for segmentation tasks is highly labor-intensive and time-consuming [3, 9, 16]. Therefore, as a promising direction to alleviate such annotation burden, weakly supervised learning has attracted the attention of many researchers. In this paper, we focus on the weakly supervised semantic segmentation (WSSS) under the supervision of image-level labels." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 376, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 604 + ], + "type": "text", + "content": "The recent advance of WSSS typically follows the three-step pipeline of: 1) transforming image tags into pixel-level coarse labels, 2) refining the pseudo labels, and 3) training the final segmentation model with the refined labels. Speaking of segmentation label generation, the technique of class activation map (CAM) [71] has been the de facto paradigm for object localization with image-level labels. Unfortunately, the naive CAM can only highlight the most discriminative area of objects and thus its small and sparse activation will lead to incomplete object mining. Since the first step of seed generation is the foundation of later processes, numerous works have been developed to expand CAM activation for high-quality pseudo labels. Among them, adversarial erasing is one of the mainstream approaches, aiming to discover new and complementary object regions by masking the currently detected area in an adversarial manner [60,67,69]. For example, the pioneering work of AE-PSL [60] performs the erasing multiple times to progressively discover increasingly activated areas. The recent work of AEFT [67] proposes to learn the concept of erasing with the triplet loss between the input image, erased image, and negatively sampled image. Though improvement is obtained, existing adversarial erasing-based methods still suffer from the difficulty in determining when to stop erasing: excessive removal will lead to over-expansion and insufficient erasure will result in under-activation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we propose a Knowledge Transfer with Simulated Inter-Image Erasing (KTSE) approach for WSSS to alleviate the problem mentioned above. Existing erasing-based methods typically mask the most discriminative object part to force the classifier into seeking other regions. Based on the observation that information removal leads to more activation (which might also incur the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "T.Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": "over-expansion illustrated in Fig. 1 (a)), we act in a diametrically opposite way. Specifically, as shown in Fig. 1 (b), we propose to simulate an inter-image erasing (SIE) scenario, introducing extra object information by concatenating a paired image and rendering the anchor image as the erased one. We then strengthen the object localization ability of the network by transferring the object knowledge from the anchor CAM to the consequent less activated localization map. Intuitively, our knowledge transfer with simulated inter-image erasing will not cause the over-expansion problem if the object in the anchor branch is well-discovered." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 480, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 480, + 344 + ], + "type": "text", + "content": "However, since the above knowledge transfer is bidirectional, it will also weaken object mining in anchor branch when learning from the sparse activation of simulated branch. Therefore, as shown in the right of Fig. 2, we propose a self-supervised regularization module for the anchor CAM feature to maintain its reliable activation in discriminative regions. Specifically, we first generate pseudo labels from the CAM feature by locating confident foreground and background area, which are used to directly supervise the learning of CAM features. However, extracted labels of the foreground are quite noisy in complex images with multiple categories of objects due to the inter-class confusion. Therefore, we also design an inter-class loss to implicitly encourage activation consistency in complex images to improve recognition of the inter-class object boundary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 347, + 479, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 479, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 479, + 538 + ], + "type": "text", + "content": "Though our proposed self-supervised regularization can effectively constrain and facilitate the knowledge transfer in our simulated inter-image erasing module, the improvement of the network's object localization ability will be severely limited by the quality of anchor CAM feature which is usually under-activated. Therefore, we resort to the traditional intra-image erasing and propose a multi-granularity alignment (MGA) module to gently expand the object activation without introducing much background noise. Specifically, as shown in the bottom of Fig. 2, we first leverage the image-level global alignment to distill the soft object confidence from the anchor branch to the masked one for object activation enlargement. Then, we leverage a pixel-level local alignment to transfer the newly discovered object information back to the anchor branch. With the gentle and constrained activation enlargement, the proposed multi-granularity alignment module further boosts the performance of our simulated inter-image erasing. As can be seen in Fig. 2, our SIE and MGA are designed in a symmetric way that play the adversarial role of expanding and constraining the anchor image activation respectively for mining complete and compact object regions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 540, + 362, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 540, + 362, + 551 + ], + "spans": [ + { + "bbox": [ + 146, + 540, + 362, + 551 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 554, + 479, + 665 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 131, + 554, + 479, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 554, + 479, + 589 + ], + "spans": [ + { + "bbox": [ + 131, + 554, + 479, + 589 + ], + "type": "text", + "content": "(1) We propose a knowledge transfer with simulated inter-image erasing approach for weakly supervised semantic segmentation to alleviate the over-expansion problem of existing adversarial erasing-based methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 591, + 479, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 591, + 479, + 627 + ], + "spans": [ + { + "bbox": [ + 131, + 591, + 479, + 627 + ], + "type": "text", + "content": "(2) A self-supervised regularization module is proposed to constrain the knowledge transfer for maintaining reliable activation and improving recognition of the inter-class object boundary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 479, + 665 + ], + "type": "text", + "content": "(3) We propose a multi-granularity alignment module to gently enlarge the object activation via image-level global alignment and pixel-level local alignment, boosting the knowledge transfer with simulated inter-image erasing." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 448, + 102 + ], + "type": "text", + "content": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 378, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 378, + 153 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 378, + 153 + ], + "type": "text", + "content": "2.1 Weakly Supervised Semantic Segmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 413 + ], + "type": "text", + "content": "Weakly supervised semantic segmentation (WSSS) is the task that aims to obtain a high-performance segmentation model by learning only from weak labels. Compared with bounding boxes [16, 25, 27, 53], scribbles [37, 56] and points [3], the image-level label [1, 4, 10, 12, 24, 34, 42, 50, 66, 73] is the cheapest weak supervision to collect, making it the most popular annotation format for WSSS. The recent advance of WSSS [1, 24, 34, 66, 73] typically relies on the class activation technique of CAM [71] to locate the target object, which helps transform the image-level label to pixel-level dense annotation for segmentation network training. Since the object area highlighted in CAM is usually small and sparse, enlarging the CAM activation becomes the focus of the WSSS task. For example, RDC [61] and DRS [28] propagate the object information to less discriminative regions via dilated convolution with varied rates and attention suppression of most discriminative areas, respectively. Popular contrastive learning [18, 48, 73] and self-supervised approaches [8, 51, 59] are also applied to facilitate object mining by improving representation learning. Moreover, cross-image information [21, 57, 72] is exploited to help the classifier discover more object patterns to obtain consistent and integral target regions. Besides attention expansion, Ahn and Kwak propose AffinityNet [2] for CAM refinement, which propagates semantic affinity via a random walk to recover delicate shapes of objects. IRN [1] is then proposed to exploit class boundary detection for discovering the entire instance areas with more accurate boundaries." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 430, + 293, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 293, + 442 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 293, + 442 + ], + "type": "text", + "content": "2.2 Erasing-based Approaches" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 668 + ], + "type": "text", + "content": "Considering the original CAM technique usually leads to suboptimal performance due to failure to localize all object parts, erasing-based approaches are commonly employed to expand the highlighted region [30, 52, 60, 67, 69, 70]. They mask areas in a training image to force the network to seek other relevant parts. While direct erasing methods like Hide-and-Seek [52] hide image patches randomly, adversarial erasing-based approaches mask the most discriminative regions with the attention information, demonstrating more promising activation expansion potential. The first adversarial erasing-based work for WSSS is proposed by Wei et al. [60], which progressively discovers new object discriminative regions by erasing the currently activated area. Such a heuristic erasing strategy then attracts the attention of many researchers. For example, ACoL [69] proposes adversarial complementary learning with two adversary classifiers. Recently, Kweon et al. propose to apply the adversarial erasing framework to exploit the potential of the pre-trained classifier [30] and learn the concept of erasing with the triplet loss [67]. While adversarial erasing approaches achieve great success in enlarging the CAM activation, they tend to suffer from the over-expansion problem due to difficulty in determining when to stop erasing. Unlike the information removal strategy adopted in these approaches, we propose a simulated" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "T.Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 251 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 251 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 251 + ], + "type": "image", + "image_path": "de5f5feb92fccd6629646b1a111135f69f07405507455e38d47c9b38f7d5dac3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 259, + 482, + 360 + ], + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 360 + ], + "type": "text", + "content": "Fig. 2: The architecture of our proposed approach. We propose a simulated inter-image erasing (SIE) scenario where extra object information is introduced from another paired image. We then strengthen the object localization ability of the network by improving the consequent less activated localization map through learning object knowledge from the anchor image. A self-supervised regularization (SSR) module is also proposed to avoid weakening the anchor activation due to bidirectional alignment and improve the inter-class object boundary recognition for complex images. In addition, we propose a multi-granularity alignment (MGA) module to gently enlarge the object activation to further boost the object knowledge transfer." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 382, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 482, + 419 + ], + "type": "text", + "content": "inter-image erasing scenario to weaken the activation by adding extra discriminative object information. The localization ability of network is then enhanced by improving the consequent less activated CAM through knowledge transfer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 436, + 202, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 202, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 202, + 449 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we propose a knowledge transfer with simulated inter-image erasing approach for weakly supervised semantic segmentation to alleviate the over-expansion and under-activation problems of existing adversarial erasing-based methods. Our framework is illustrated in Fig. 2. As demonstrated in the anchor branch, we train a classification network with the given image-level weak labels, which consists of a backbone feature extractor and a pooling classification head. In contrast to existing adversarial erasing-based approaches that expand the activation through masking the most discriminative object part, as shown in the top-left of Fig. 2, we propose to simulate an inter-image erasing scenario where extra object information is introduced through concatenating a paired image. Object knowledge is then transferred from the anchor image to the consequent less activated localization map for strengthening the object localization ability of network. We also propose a self-supervised regularization module to avoid weakening the anchor activation due to bidirectional alignment and improve the inter-class object boundary recognition for complex multi-category images. In addition, we propose a multi-granularity alignment module to gently enlarge the object activation for boosting the object knowledge transfer." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "text", + "content": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 248, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 248, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 248, + 126 + ], + "type": "text", + "content": "3.1 CAM Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "content": "For the architecture of classification network, we follow the previous erasing-based work ACoL [69] and remove the final fully-connected layer. We set the output channel of the backbone to " + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "inline_equation", + "content": "C + 1" + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "content": " is the number of foreground categories and 1 is added for the background. We can thus directly generate object localization maps from the class-aware CAM features " + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "content": " in the forward pass to facilitate the object knowledge transfer. To obtain CAM for each foreground class " + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "content": ", we feed the attention map " + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "inline_equation", + "content": "F^c" + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "content": " into a ReLU layer and then normalize it to the range from 0 to 1:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 268, + 236, + 481, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 236, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 268, + 236, + 481, + 262 + ], + "type": "interline_equation", + "content": "A ^ {c} = \\frac {\\operatorname {R e L U} \\left(F ^ {c}\\right)}{\\operatorname* {m a x} \\left(F ^ {c}\\right)}. \\tag {1}", + "image_path": "18a83fce9f6b2858de19fe361e4410f5d0247adeec3653d5d094911634529adb.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 268, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 268, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 131, + 268, + 482, + 327 + ], + "type": "text", + "content": "To improve the quality of the CAM by capturing the global context and local details of the image, we adopt the Gated Pyramid Pooling (GPP) layer [67] as the final pooling head. With the classification logits " + }, + { + "bbox": [ + 131, + 268, + 482, + 327 + ], + "type": "inline_equation", + "content": "q^{c}" + }, + { + "bbox": [ + 131, + 268, + 482, + 327 + ], + "type": "text", + "content": " generated from the pooling head, we train the classification network with the multi-label soft margin loss as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 192, + 338, + 481, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 338, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 192, + 338, + 481, + 369 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c l s} = - \\frac {1}{C} \\sum_ {c = 1} ^ {C} y ^ {c} \\log \\sigma \\left(q ^ {c}\\right) + \\left(1 - y ^ {c}\\right) \\log \\left[ 1 - \\sigma \\left(q ^ {c}\\right) \\right]. \\tag {2}", + "image_path": "e5a35c57d2fde0be49830e828df487dd11fa74ed7a73e17f8b9beb9e46b99f8e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 380, + 479, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 479, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 479, + 392 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 132, + 380, + 479, + 392 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 132, + 380, + 479, + 392 + ], + "type": "text", + "content": " is the sigmoid function. " + }, + { + "bbox": [ + 132, + 380, + 479, + 392 + ], + "type": "inline_equation", + "content": "y^{c}" + }, + { + "bbox": [ + 132, + 380, + 479, + 392 + ], + "type": "text", + "content": " is the image-level label for the " + }, + { + "bbox": [ + 132, + 380, + 479, + 392 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 132, + 380, + 479, + 392 + ], + "type": "text", + "content": "-th class." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 413, + 316, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 316, + 426 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 316, + 426 + ], + "type": "text", + "content": "3.2 Simulated Inter-Image Erasing" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 533 + ], + "type": "text", + "content": "Due to the sparsity of the CAM activation, previous adversarial erasing-based methods typically enlarge the object activation by maintaining the classification confidence on masked images or features, where the most discriminative regions are erased. However, due to the absence of guidelines on when to stop erasing, these approaches easily incur over-expansion with excessive removal and might also still suffer from under-activation with insufficient erasure. Therefore, we propose a knowledge transfer with simulated inter-image erasing approach for weakly supervised semantic segmentation to alleviate the above problem." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 534, + 482, + 666 + ], + "type": "text", + "content": "In contrast to existing erasing-based approaches that remove the discriminative part for more object mining, we propose a simulated inter-image erasing (SIE) scenario to enable the network to benefit from introducing extra discriminative object information. Specifically, as shown in the top-left of Fig. 2, a larger synthetic image, created by concatenating the anchor image and another paired one, plays the role of the original image needing masking. Correspondingly, the anchor image can be treated as the erased one where the paired image is masked. Due to the introduction of extra discriminative object information, fewer object regions will be highlighted in the anchor part of the synthetic image. Therefore, we then strengthen the object localization ability of the network by transferring the more integral object knowledge from the original anchor CAM to the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "T.Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": "consequently less activated anchor part of the synthetic image. The knowledge transfer loss for the simulated inter-image erasing can be formulated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 146, + 480, + 165 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 146, + 480, + 165 + ], + "spans": [ + { + "bbox": [ + 143, + 146, + 480, + 165 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {k t} = \\operatorname {R e L U} \\left(\\hat {F} _ {a} - \\hat {F} _ {s}\\right), \\quad \\text {w h e r e} \\hat {F} _ {a} = C F E \\left(F _ {a}, y\\right), \\hat {F} _ {s} = C F E \\left(F _ {s}, y\\right). \\tag {3}", + "image_path": "613ed4c1f18a941075e5a11db354eaaebbe012d48fc56e1cf3abefd54cc6af7b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 169, + 482, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 169, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 131, + 169, + 482, + 350 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 169, + 482, + 350 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 131, + 169, + 482, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 169, + 482, + 350 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 131, + 169, + 482, + 350 + ], + "type": "text", + "content": " denote the CAM features of the anchor and simulated branches, respectively. CFE represents class feature extraction, which extracts the feature channel related to the classes that exist in the image. The ReLU operation means that our alignment focuses on the more highlighted object region rather than the background. Note that no rotation or resize is performed for the concatenated image. We only introduce extra object information from the paired image and control other variables to guarantee the weakened activation for the anchor part. Intuitively, our knowledge transfer with simulated inter-image erasing will not cause the over-expansion problem when the object in the anchor branch is well-discovered. Our experiments demonstrate that such knowledge transfer can effectively strengthen the localization ability of the network to alleviate the under-activation, especially when multiple object instances co-exist in the image. Besides, benefiting from the bidirectional alignment, the anchor branch will also learn from the weakened activation of the simulated one and produce compact attention maps for small objects to mitigate the over-expansion problem." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 364, + 313, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 313, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 313, + 376 + ], + "type": "text", + "content": "3.3 Self-Supervised Regularization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 381, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 381, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 381, + 482, + 514 + ], + "type": "text", + "content": "Our simulated inter-image erasing aims to improve the object localization ability of the network by maintaining the high activation given extra discriminative object information. However, the adopted knowledge transfer is bidirectional, which will also weaken the anchor branch object mining when learning from the sparse activation of the simulated branch. Since simply cutting the gradient propagation of the anchor branch makes the training unstable, we propose a self-supervised regularization module for the anchor CAM feature to maintain its reliable activation in discriminative regions. Specifically, with the generated CAM feature " + }, + { + "bbox": [ + 131, + 381, + 482, + 514 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 131, + 381, + 482, + 514 + ], + "type": "text", + "content": " and the corresponding CAM " + }, + { + "bbox": [ + 131, + 381, + 482, + 514 + ], + "type": "inline_equation", + "content": "A_{a}" + }, + { + "bbox": [ + 131, + 381, + 482, + 514 + ], + "type": "text", + "content": ", we first leverage two thresholds " + }, + { + "bbox": [ + 131, + 381, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\beta_{h} = 0.3" + }, + { + "bbox": [ + 131, + 381, + 482, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 381, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\beta_{l} = 0.15" + }, + { + "bbox": [ + 131, + 381, + 482, + 514 + ], + "type": "text", + "content": " to locate the confident foreground and background as follows (the subscript " + }, + { + "bbox": [ + 131, + 381, + 482, + 514 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 131, + 381, + 482, + 514 + ], + "type": "text", + "content": " is omitted for simplicity):" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 211, + 520, + 480, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 520, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 211, + 520, + 480, + 564 + ], + "type": "interline_equation", + "content": "\\hat {Y} _ {i, j} = \\left\\{ \\begin{array}{l l} \\operatorname {a r g m a x} \\left(F _ {i, j,:}\\right), & \\text {i f} \\max \\left(A _ {i, j,:}\\right) \\geq \\beta_ {h}, \\\\ 0, & \\text {i f} \\max \\left(A _ {i, j,:}\\right) \\leq \\beta_ {l}, \\\\ 2 5 5, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {4}", + "image_path": "84728775f27a38f20db67adbe9990d3bb022d5ff3371db802464c5a19d221770.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 569, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 569, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 131, + 569, + 482, + 629 + ], + "type": "text", + "content": "where 255 denote the ignored labels for uncertain regions and " + }, + { + "bbox": [ + 131, + 569, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\arg\\max(\\cdot)" + }, + { + "bbox": [ + 131, + 569, + 482, + 629 + ], + "type": "text", + "content": " extracts the semantic class with the maximum activation value. After refining the pseudo label " + }, + { + "bbox": [ + 131, + 569, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 131, + 569, + 482, + 629 + ], + "type": "text", + "content": " with the pixel-adaptive refinement module [47], we leverage it to directly supervise the learning of CAM features with the cross-entropy loss (coordinates " + }, + { + "bbox": [ + 131, + 569, + 482, + 629 + ], + "type": "inline_equation", + "content": "i, j" + }, + { + "bbox": [ + 131, + 569, + 482, + 629 + ], + "type": "text", + "content": " are omitted):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 205, + 636, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 636, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 205, + 636, + 480, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c e} = - \\sum_ {c = 0} ^ {C} \\hat {Y} _ {b g} ^ {c} \\log \\Gamma (F) ^ {c} + \\omega_ {f g} \\hat {Y} _ {f g} ^ {c} \\log \\Gamma (F) ^ {c}. \\tag {5}", + "image_path": "db77bef9af4b49713deab585807996d064ed8d24b37482479a0dbb337e9c5f51.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\Gamma(\\cdot)" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": " denotes the softmax function. We also leverage the pseudo label " + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": " generated from the anchor branch to guide the learning of the simulated CAM feature with the same cross-entropy loss. However, we notice that the extracted labels of the foreground are quite noisy, especially for complex images with multiple categories of objects where the inter-class boundary is blurred. Here, we define images with objects of only one category as simple images and others containing two or more categories as complex ones. Therefore, we discount the influence of foreground label " + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\hat{Y}_{fg}" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": " with the weight of " + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\omega_{fg} = 0.0125" + }, + { + "bbox": [ + 130, + 115, + 482, + 236 + ], + "type": "text", + "content": ". Besides, as shown in the right of Fig. 2, we design an inter-class loss to implicitly encourage the activation consistency for the multiple foreground classes in complex images as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 233, + 245, + 481, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 245, + 481, + 271 + ], + "spans": [ + { + "bbox": [ + 233, + 245, + 481, + 271 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {i n t e r}} = \\frac {\\sum \\left(V ^ {2 n d} - V ^ {\\max }\\right) \\cdot M ^ {f}}{\\left| M ^ {f} \\right|}, \\tag {6}", + "image_path": "a004763f6cec3a0e26de804ac33292c31675770984edfd5ff90ce29773e5a18b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 281, + 482, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 281, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 131, + 281, + 482, + 329 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 281, + 482, + 329 + ], + "type": "inline_equation", + "content": "M^f" + }, + { + "bbox": [ + 131, + 281, + 482, + 329 + ], + "type": "text", + "content": " denotes the foreground mask obtained with a threshold " + }, + { + "bbox": [ + 131, + 281, + 482, + 329 + ], + "type": "inline_equation", + "content": "\\beta_o = 0.2" + }, + { + "bbox": [ + 131, + 281, + 482, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 281, + 482, + 329 + ], + "type": "inline_equation", + "content": "|\\cdot|" + }, + { + "bbox": [ + 131, + 281, + 482, + 329 + ], + "type": "text", + "content": " return the value of " + }, + { + "bbox": [ + 131, + 281, + 482, + 329 + ], + "type": "inline_equation", + "content": "l_1" + }, + { + "bbox": [ + 131, + 281, + 482, + 329 + ], + "type": "text", + "content": "-norm. " + }, + { + "bbox": [ + 131, + 281, + 482, + 329 + ], + "type": "inline_equation", + "content": "V^{max}" + }, + { + "bbox": [ + 131, + 281, + 482, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 281, + 482, + 329 + ], + "type": "inline_equation", + "content": "V^{2nd}" + }, + { + "bbox": [ + 131, + 281, + 482, + 329 + ], + "type": "text", + "content": " denotes the largest and second largest activation value along the channel dimension of the CAM feature " + }, + { + "bbox": [ + 131, + 281, + 482, + 329 + ], + "type": "inline_equation", + "content": "F_a" + }, + { + "bbox": [ + 131, + 281, + 482, + 329 + ], + "type": "text", + "content": ", which can be obtained as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 150, + 339, + 481, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 339, + 481, + 356 + ], + "spans": [ + { + "bbox": [ + 150, + 339, + 481, + 356 + ], + "type": "interline_equation", + "content": "V _ {i, j} ^ {m a x} = \\max (F _ {i, j,:}), \\quad F ^ {l e f t} = \\mathsf {r e m o v e} (F, V _ {m a x}), \\quad V _ {i, j} ^ {2 n d} = \\max (F _ {i, j,:} ^ {l e f t}). \\qquad (7)", + "image_path": "6d02094ce70b6a702d18f6a66d2d1fad15357b23252f35a162568b6bcdb60e93.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 364, + 482, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 425 + ], + "type": "text", + "content": "Eq. (6) improves the exclusivity of the class activation for each pixel, leading to more accurate boundaries between foreground objects. Our proposed self-supervised regularization module treats the simple and complex images differently, intending to maintain the reliable activation of the main branch without deteriorating the inter-class boundary caused by inaccurate pseudo labels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 441, + 308, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 441, + 308, + 454 + ], + "spans": [ + { + "bbox": [ + 131, + 441, + 308, + 454 + ], + "type": "text", + "content": "3.4 Multi-Granularity Alignment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": "Our proposed self-supervised regularization module can effectively constrain and boost the knowledge transfer in our simulated inter-image erasing module. However, the improvement of the network's object localization ability will be severely limited by the quality of the anchor CAM, which is usually under-activated. Therefore, it is natural to resort to the traditional intra-image erasing, which aims to expand the object area. Following AEFT [67], we leverage a threshold of 0.6 to mask the most discriminative region. To avoid introducing too much background noise, we propose a multi-granularity alignment module to gently expand the object activation, which encourages the CAM expansion of the erased image and then transfers the learned object knowledge back to the anchor branch. As shown in the bottom of Fig. 2, we first input the anchor feature " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " and masked feature " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "F_{m}" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " into a class feature extraction module to obtain the corresponding features of the existing target categories " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "(\\hat{F}_{a}" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\hat{F}_{m}" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": "). Following the experimental finding of AEFT [67] that rigid classification supervision [69] for the masked branch tends to trigger the over-expansion problem in adversarial erasing, we resort to the soft class confidence guidance from the anchor branch to the masked one. Specifically, we directly adopt a global average pooling (GAP) operation" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "T.Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": "to obtain the final class confidence for each branch and an image-level global alignment loss can be formulated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 187, + 149, + 426, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 149, + 426, + 169 + ], + "spans": [ + { + "bbox": [ + 187, + 149, + 426, + 169 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g l o b a l} = G A P \\left(R e L U \\left(\\hat {F} _ {a}\\right)\\right) - G A P \\left(R e L U \\left(\\hat {F} _ {m}\\right)\\right),", + "image_path": "889b8082fa6cfbdabf7f8bb74e1b994a2b99c4985f20cdf10c137b35e69108be.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 202, + 172, + 479, + 185 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 172, + 479, + 185 + ], + "spans": [ + { + "bbox": [ + 202, + 172, + 479, + 185 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} \\quad \\hat {F _ {m}} = C F E \\left(F _ {m}, y\\right). \\tag {8}", + "image_path": "a9c521cc2586d4e0e62ec9fcc6ef752c3f46d00564b566e7763d13ff40487a3b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 195, + 480, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 195, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 131, + 195, + 480, + 316 + ], + "type": "inline_equation", + "content": "\\hat{F}_a" + }, + { + "bbox": [ + 131, + 195, + 480, + 316 + ], + "type": "text", + "content": " and CFE have been defined in Eq. (3). Another difference with rigid classification supervision is that benefiting from the class feature extraction, our global alignment loss only focuses on the logit confidence for classes that exist in the image, enabling more effective and efficient gradient propagation. Unlike AEFT [67] that utilizes the feature space of the GPP layer as embedding space for loss construction, our experiments reveal that direct alignment with the CAM feature leads to more promising performance. With class information learned from the anchor branch, we further leverage a pixel-level local activation alignment to transfer the newly discovered object information from the erased image to guide the anchor branch learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 247, + 326, + 479, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 326, + 479, + 344 + ], + "spans": [ + { + "bbox": [ + 247, + 326, + 479, + 344 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {l o c a l}} = \\operatorname {R e L U} \\left(\\hat {F _ {m}} - \\hat {F _ {a}}\\right). \\tag {9}", + "image_path": "1d4e7c62207f601d26bee4190442866e7b8a33d127351f5c793be7afe719d84d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 353, + 480, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 353, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 131, + 353, + 480, + 460 + ], + "type": "text", + "content": "Such pixel-level local alignment can also hinder the erased branch from activating the unwanted background area by learning from the lowly-activated regions of the anchor branch. With the gentle and constrained activation enlargement, the multi-granularity alignment module further boosts the performance of our proposed simulated inter-image erasing approach. As can be seen in Fig. 2, our proposed SIE and MGA branches are designed in a symmetric way that play the adversarial role of expanding and constraining the anchor image activation, respectively. Their synergically cooperation facilitates the convergence for mining complete and compact object regions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 478, + 255, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 478, + 255, + 490 + ], + "spans": [ + { + "bbox": [ + 131, + 478, + 255, + 490 + ], + "type": "text", + "content": "3.5 Training Objective" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 498, + 298, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 298, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 298, + 510 + ], + "type": "text", + "content": "The overall training loss is as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 190, + 521, + 479, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 521, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 190, + 521, + 479, + 533 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c l s} + \\mathcal {L} _ {k t} + \\mathcal {L} _ {g l o b a l} + \\mathcal {L} _ {l o c a l} + \\mathcal {L} _ {c e} + \\lambda_ {i n t e r} \\mathcal {L} _ {i n t e r}. \\tag {10}", + "image_path": "1a58728aa0dab9e1347c75034b41f2380edb7ee0be0df06859a356a57dfe9df8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 542, + 479, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 542, + 479, + 564 + ], + "spans": [ + { + "bbox": [ + 131, + 542, + 479, + 564 + ], + "type": "text", + "content": "We empirically set " + }, + { + "bbox": [ + 131, + 542, + 479, + 564 + ], + "type": "inline_equation", + "content": "\\lambda_{inter} = 0.005" + }, + { + "bbox": [ + 131, + 542, + 479, + 564 + ], + "type": "text", + "content": " as the hyperparameter that controls the weight of inter-class loss." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 584, + 224, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 584, + 224, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 584, + 224, + 597 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 609, + 348, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 348, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 348, + 621 + ], + "type": "text", + "content": "4.1 Datasets and Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 629, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 479, + 666 + ], + "type": "text", + "content": "We evaluate our approach on the PASCAL VOC 2012 [19] and COCO [38] datasets. The PASCAL VOC 2012 dataset contains 21 classes (20 object categories and the background) for semantic segmentation. It has 10,582 images for" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 146, + 299, + 293 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 301, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 301, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 301, + 137 + ], + "type": "text", + "content": "Table 1: Accuracy of pseudo-masks evaluated on PASCAL VOC 2012 training set." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 146, + 299, + 293 + ], + "lines": [ + { + "bbox": [ + 133, + 146, + 299, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 146, + 299, + 293 + ], + "type": "table", + "html": "
MethodsSeedw/IRN [1]
IRN [1]CVPR1948.366.3
MBMNet [40]MM2050.266.8
CONTA [68]NIPS2048.867.9
AdvCAM [32]CVPR2155.669.9
RIB [31]NIPS2156.670.6
ReCAM [14]CVPR2256.670.5
ESOL [35]NIPS2253.668.7
CLIMS [63]CVPR2256.670.5
AEFT [67]ECCV2256.071.0
ACR [15]CVPR2360.372.3
FPR [6]ICCV2363.8-
KTSE (Ours)67.073.8
", + "image_path": "4f066a10f473d2c107fc5a93cbbba97bb23699205c43b79edafe50cfda5a8279.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 310, + 167, + 480, + 293 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 2: Quantitative comparisons on PASCAL VOC 2012 val and test sets with VGG backbone. Sup: Supervision, I: Image-level label, S: Saliency maps." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 167, + 480, + 293 + ], + "lines": [ + { + "bbox": [ + 310, + 167, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 310, + 167, + 480, + 293 + ], + "type": "table", + "html": "
MethodsSupValTest
DRS [28]AAAI21I+S63.664.4
GSM [36]AAAI21I+S63.363.6
NSROM [66]CVPR21I+S65.565.3
EPS [34]CVPR21I+S67.067.3
L2G [24]CVPR22I+S68.568.9
AffinityNet [2]CVPR18I58.460.5
ICD [20]CVPR20I61.260.9
BES [7]ECCV20I60.161.1
ECS [55]ICCV21I62.163.4
KTSE (Ours)I67.367.0
", + "image_path": "9934c9b4287ab9ed4405b6f657289198a03988abb114ccb3a14e2dc18437b3db.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 316, + 479, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 479, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 479, + 387 + ], + "type": "text", + "content": "training (expanded with SBD [22]), 1,449 for validation and 1,456 for testing. The COCO dataset is a more challenging benchmark with 80 semantic classes and the background. Following previous works [36, 58, 68], we use the default train/val splits (80k images for training and 40k for validation) in the experiment. Mean intersection over union (mIoU) is adopted as the metric to evaluate the quality of our pseudo labels and segmentation results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "text", + "content": "For the classification network, we follow the work of AEFT [67] and employ ResNet38 [62] as the backbone. We adopt a poly learning rate with an initial value of " + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "text", + "content": " and a power of 0.9. For the second stage training of WSSS, following the recent work of BECO [45], we adopt DeeplabV2 [5] as the segmentation network, which uses ResNet101 [23] as the backbone. The momentum and weight decay of the SGD optimizer are 0.9 and " + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "text", + "content": ". The initial learning rate is set to " + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "text", + "content": " and is decreased using polynomial decay. The segmentation model is trained for 80 epochs and 40 epochs on VOC and COCO datasets, respectively, with a common batch size of 16. We also follow the default setting of DeeplabV2 [5] and conduct experiments with VGG16 backbone for a more comprehensive comparison with previous approaches [7, 20, 24, 34, 55]. All our backbones are pre-trained on ImageNet [17]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 549, + 346, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 549, + 346, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 346, + 562 + ], + "type": "text", + "content": "4.2 Comparisons to the State-of-the-arts" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Accuracy of Pseudo-Masks. We first report the quality of the segmentation seeds and the generated pseudo-masks derived from our approach. The comparison with other state-of-the-arts is presented in Tab. 1. As can be seen, our segmentation seed can reach the mIoU of " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "67.0\\%" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": ", bringing a gain of " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "18.7\\%" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " compared to the baseline reported by IRN [1]. Our method can outperform the state-of-the-art method FPR [6] by " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": ". With the further refinement of IRN [1], the mIoU of our generated pseudo-masks can arrive at " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "73.8\\%" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": ", surpassing the previous SOTA of AEFT [67] and ACR [15] by more than " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "T.Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 179, + 306, + 554 + ], + "blocks": [ + { + "bbox": [ + 132, + 125, + 305, + 178 + ], + "lines": [ + { + "bbox": [ + 132, + 125, + 305, + 178 + ], + "spans": [ + { + "bbox": [ + 132, + 125, + 305, + 178 + ], + "type": "text", + "content": "Table 3: Quantitative comparisons on PASCAL VOC 2012 val and test sets with ResNet backbone. Sup: Supervision, I: Image-level label, S: Saliency maps, O: Out-of-distribution data, L: Language." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 179, + 306, + 554 + ], + "lines": [ + { + "bbox": [ + 132, + 179, + 306, + 554 + ], + "spans": [ + { + "bbox": [ + 132, + 179, + 306, + 554 + ], + "type": "table", + "html": "
MethodsSupValTest
DRS [28]AAAI21I+S71.271.4
NSROM [66]CVPR21I+S70.470.2
EPS [34]CVPR21I+S71.071.8
AuxSeg [64]ICCV21I+S69.068.6
PPC [18]CVPR22I+S72.673.6
RCA [73]CVPR22I+S72.272.8
L2G [24]CVPR22I+S72.171.7
W-OoD [33]CVPR22I+O69.869.9
CLIP-ES [39]CVPR23I+L71.171.4
LPCAM [13]CVPR23I+S71.872.1
AdvCAM [32]CVPR21I68.168.0
CDA [54]ICCV21I66.166.8
CSE [30]ICCV21I68.468.2
RIB [31]NIPS21I68.368.6
AMR [44]AAAI22I68.869.1
MCT [65]CVPR22I71.971.6
AFA [47]CVPR22I66.066.3
SIPE [8]CVPR22I68.869.7
ReCAM [14]CVPR22I68.568.4
PPC [18]CVPR22I67.767.4
ViT-PCM [46]ECCV22I70.370.9
S-BCE [46]ECCV22I70.071.3
AEFT [67]ECCV22I70.971.7
TOCO [48]CVPR23I69.870.5
OCR [15]CVPR23I72.772.0
ACR [15]CVPR23I71.971.9
BECO [45]CVPR23I72.171.8
FPR [6]ICCV23I70.370.1
MARS [26]ICCV23I70.371.2
KTSE (Ours)I73.072.9
", + "image_path": "98125d948255bc9d8f381042a1a524784c8d54153782d9ffeb04088a6bb0a58d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 320, + 170, + 480, + 306 + ], + "blocks": [ + { + "bbox": [ + 318, + 125, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 318, + 125, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 318, + 125, + 482, + 169 + ], + "type": "text", + "content": "Table 4: Quantitative comparisons on COCO val set with VGG backbone. Sup: Supervision, I: Image-level label, S: Saliency maps." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 170, + 480, + 306 + ], + "lines": [ + { + "bbox": [ + 320, + 170, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 320, + 170, + 480, + 306 + ], + "type": "table", + "html": "
MethodsSupVal
IAL [58]IJCV20I+S27.7
GWSM [36]AAAI21I+S28.4
EPS [34]CVPR21I+S35.7
I2CRC [73]TMM22I+S31.2
RCA [73]CVPR22I+S36.8
MDBA [11]TIP23I+S37.8
BFBP [49]ECCV16I20.4
SEC [29]ECCV16I22.4
CONTA [68]NIPS20I23.7
KTSE (Ours)I37.2
", + "image_path": "92bbdc6bb1847b295da10172ad4ea6e8f1e802cc4231750db5f8b66424fc98a0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 321, + 358, + 479, + 553 + ], + "blocks": [ + { + "bbox": [ + 318, + 314, + 482, + 357 + ], + "lines": [ + { + "bbox": [ + 318, + 314, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 318, + 314, + 482, + 357 + ], + "type": "text", + "content": "Table 5: Quantitative comparisons on COCO val set with ResNet backbone. Sup: Supervision, I: Image-level label, S: Saliency maps, L: Language." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 358, + 479, + 553 + ], + "lines": [ + { + "bbox": [ + 321, + 358, + 479, + 553 + ], + "spans": [ + { + "bbox": [ + 321, + 358, + 479, + 553 + ], + "type": "table", + "html": "
MethodsSupVal
L2G [24]CVPR22I+S44.2
CLIP-ES [39]CVPR23I+L45.4
LPCAM [13]CVPR23I+S42.1
OC-CSE [30]ICCV21I36.4
CDA [54]ICCV21I33.2
RIB [31]NIPS21I43.8
MCT [65]CVPR22I42.0
SIPE [8]CVPR22I40.6
TOCO [48]CVPR23I41.3
OCR [15]CVPR23I42.5
ACR [15]CVPR23I45.3
BECO [45]CVPR23I45.1
FPR [6]ICCV23I43.9
USAGE [43]ICCV23I44.3
KTSE (Ours)I45.9
", + "image_path": "4defe851b87d29314c297784531fc0dff7248bc0cb8ef85e24f954b92300346a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Accuracy of Segmentation Maps on PASCAL VOC 2012. Our segmentation results on the PASCAL VOC 2012 dataset with the backbone of VGG and ResNet are demonstrated in Tab. 2 and Tab. 3, respectively. As can be seen, with the VGG backbone, our approach achieves the performance of " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "67.3\\%" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " on the validation set and " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "67.0\\%" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " on the test set, better than other state-of-the-art methods with only image-level labels. Our segmentation results are also competitive to many approaches that rely on saliency maps, e.g., NSROM [66] and EPS [34]. With the ResNet backbone, we can improve the results to " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "73.0\\%" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "72.9\\%" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " on" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 90, + 448, + 102 + ], + "type": "text", + "content": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 168, + 305, + 249 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 309, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 309, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 309, + 159 + ], + "type": "text", + "content": "Table 6: Element-wise component analysis. SIE: Simulated Inter-Image Erasing, SSR: Self-Supervised Regularization, MGA: Multi-Granularity Alignment." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 168, + 305, + 249 + ], + "lines": [ + { + "bbox": [ + 136, + 168, + 305, + 249 + ], + "spans": [ + { + "bbox": [ + 136, + 168, + 305, + 249 + ], + "type": "table", + "html": "
BaseSIESSRMGAmIoU
✓54.2
✓✓57.1
✓✓57.4
✓✓60.2
✓✓✓63.8
✓✓✓✓67.0
", + "image_path": "3ca663f4f7fa382033020949a3f466a14de2a05267f4fc2c2424d5a5a89dec93.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 324, + 168, + 479, + 248 + ], + "blocks": [ + { + "bbox": [ + 322, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 322, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 322, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 7: Comparison of Our Multi-Granularity Alignment with Previous Erasing-based Approaches. GA: Global Alignment; LA: Local Alignment." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 168, + 479, + 248 + ], + "lines": [ + { + "bbox": [ + 324, + 168, + 479, + 248 + ], + "spans": [ + { + "bbox": [ + 324, + 168, + 479, + 248 + ], + "type": "table", + "html": "
MethodmIoU
Base54.2
Rigid Classification [69]55.6
Soft GPP Feature [67]54.9
Our GA58.7
Our GA & LA60.2
", + "image_path": "60f1110ce0c27a7df56ba24a5fbe0f998b92cf599ee240aa3fb863bab451633b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 270, + 479, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 479, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 479, + 330 + ], + "type": "text", + "content": "the validation and test sets, respectively. As can be seen in Tab. 3, our proposed approach can achieve better performance compared with recent SOTA methods. For example, our approach outperforms OCR [15] and ACR [15] by about " + }, + { + "bbox": [ + 130, + 270, + 479, + 330 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 270, + 479, + 330 + ], + "type": "text", + "content": " on the test set, demonstrating the superiority of our knowledge transfer with simulated inter-image erasing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "text", + "content": "Accuracy of Segmentation Maps on COCO. For the more challenging COCO dataset, we provide performance comparisons with state-of-the-art WSSS methods using the backbone of VGG and ResNet in Tab. 4 and Tab. 5, respectively. As shown in Tab. 4, our proposed KTSE with VGG backbone can achieve the performance of " + }, + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "inline_equation", + "content": "37.2\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "text", + "content": " mIoU, much better than previous methods supervised with only image-level labels, e.g., " + }, + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "inline_equation", + "content": "13.5\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "text", + "content": " mIoU higher than CONTA [68]. Besides, our results are also competitive compared to previous SOTA methods with additional saliency guidance like RCA [73] and MDBA [11]. With the ResNet backbone, our proposed KTSE reaches the best result of " + }, + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "inline_equation", + "content": "45.9\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "text", + "content": " mIoU compared to previous SOTA WSSS methods. Specifically, our approach outperforms ACR [15] and BECO [45] by " + }, + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "text", + "content": " mIoU, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 479, + 246, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 246, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 246, + 490 + ], + "type": "text", + "content": "4.3 Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": "Element-Wise Component Analysis. In this part, we demonstrate the contribution of each component proposed in our approach to improving the quality of pseudo-masks. As mentioned in Sec. 3.1, we incorporate the GPP [67] into the classification network as our strong baseline. As shown in Tab. 6, with our simulated inter-image erasing (SIE) module, we can improve the accuracy of segmentation seeds from the baseline of " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "54.2\\%" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "57.1\\%" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": ", which demonstrates the benefit of strengthening the localization ability of the network through weakening the activation with extra object information and then learning from the anchor branch. By maintaining reliable activation in discriminative regions, our self-supervised regularization (SSR) module can achieve the mIoU of " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "57.4\\%" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": ". Moreover, the combination of our proposed SIE and SSR can significantly boost the performance to " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "63.8\\%" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": " mIoU. With our proposed multi-granularity alignment module to gently expand the object activation, the mIoU of pseudo masks finally arrives at " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "67.0\\%" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": ", highlighting the importance of improving the quality of" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "T.Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 482, + 242 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 242 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 242 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 242 + ], + "type": "image", + "image_path": "c35b984ca6892c1d4ed49a980056c572e34d7615035221993edc7a9d95ead474.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 251, + 480, + 293 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 480, + 293 + ], + "type": "text", + "content": "Fig. 3: Example localization maps on the PASCAL VOC 2012 training set. For each (a) image, we show (b) ground truth, localization maps produced by (c) previous work of AEFT [67], (d) our baseline, (e) baseline + SIE, (f) baseline + SIE + SSR, and (g) baseline + SIE + SSR + MGA. Best viewed in color." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 324, + 479, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 479, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 479, + 348 + ], + "type": "text", + "content": "anchor CAM to boost the performance of our knowledge transfer with simulated inter-image erasing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 351, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 567 + ], + "type": "text", + "content": "Some example localization maps on the PASCAL VOC 2012 training set can be viewed in Fig. 3. As seen from the first two rows, after strengthening the localization ability of the network with our simulated inter-image erasing (SIE) module, we can successfully expand the object activation and discover the less discriminative object in the image (e.g., the smaller animals on the right). We can notice that our SIE can also help alleviate the over-expansion problem and generate more compact activation, e.g., the cow in the first row. However, as demonstrated in the last two rows, the bidirectional alignment in SIE will also lead to more severe under-activation when the anchor branch learns from the sparse activation of the simulated one, especially for larger objects. Fortunately, our proposed self-supervised regularization (SSR) module can help maintain reliable activation in discriminative regions to guarantee effective knowledge transfer. Moreover, as seen in the last row, our multi-granularity alignment (MGA) module can further help discover the integral target region by gently expanding the object activation. Comparing the results of ours with that of AEFT [67], Fig. 3 (g) vs Fig. 3 (c), we can see that our proposed approach can significantly alleviate the under-activation and over-expansion of previous approaches due to insufficient or excessive erasure." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "MGA vs. Previous Erasing-based Approaches. For the multi-granularity alignment, we resort to the soft class confidence knowledge from the anchor branch to guide the gentle activation expansion of the masked image. In Tab. 7, we compare performance improvement from different guidance settings. As can be seen, our global alignment of CAM feature can improve the baseline from " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "54.2\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "58.7\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ", outperforming the rigid classification guidance in ACoL [69] and GPP feature alignment adopted in AEFT [67] by " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ", respectively. It highlights the importance of direct global alignment in CAM features" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "text", + "content": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "rather than embedding space with extra projection layers. With our pixel-level local alignment, our multi-granularity alignment module can finally improve the performance to " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "60.2\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": ", demonstrating the benefit of our gentle alignment strategy compared with previous erasing-based approaches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 481, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 481, + 272 + ], + "type": "text", + "content": "Phenomenon of SIE. Based on the observation of classic adversarial erasing methods that removing the most discriminative region will lead to the extra activation of other object parts, the reversal thinking comes naturally—what if we introduce extra discriminative object information? The answer is the original highly activated area will become less discriminative and be lowly activated. We then enhance this weakened activation by learning from the original CAM. When the network learns to increase the attention for the less-discriminative region in the concatenated image, it will also learn to activate less-discriminative object region in the original image to locate more objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 272, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 452 + ], + "type": "text", + "content": "SIE vs. Data Augmentation. Though data augmentations like CutMix will also change the anchor image and lead to activation perturbation, they do not guarantee the weakened activation (which might cause over-expansion) for the anchor part of the concatenated image like ours. In contrast, we only introduce extra object information from the paired image and control other variables. Most importantly, our novelty lies in constructing the simulated inter-image erasing scenario (reversal thinking on existing masking methods) rather than simply augmenting the data input. It enhances the network's localization ability by improving the consequent less activated attention map through learning from the object knowledge of the anchor branch. Accompanied by our proposed MGA, they are designed in a symmetric way that play the adversarial role of expanding and constraining the anchor image activation respectively for mining complete and compact object regions. Note that our approach significantly outperforms the CutMix-based augmentation method of CDA [54] with 73.0 vs. 66.1 on VOC (Tab. 3) and 45.9 vs. 33.2 on COCO (Tab. 5)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 471, + 220, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 220, + 484 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 220, + 484 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "This paper proposed a Knowledge Transfer with Simulated Inter-Image Erasing (KTSE) approach for weakly supervised semantic segmentation. Specifically, in contrast to existing adversarial erasing-based methods that remove the discriminative part for mining less-discriminative areas, we proposed to simulate an inter-image erasing scenario where extra object information was added through concatenating a paired image. We then strengthened the object localization ability of the network by enhancing the consequent less activated localization map. In addition, we proposed a self-supervised regularization module to maintain reliable activation in the discriminative regions and improve the inter-class object boundary recognition for complex images. Moreover, we proposed a multi-granularity alignment module to gently enlarge the object activation via intra-image erasing for boosting the object knowledge transfer. Extensive experiments and ablation studies were conducted on PASCAL VOC 2012 and COCO datasets to demonstrate the superiority of our proposed approach." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "T.Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (No. 62202227 and 62102182), Natural Science Foundation of Jiangsu Province (No. BK20220938 and BK20220934), China Postdoctoral Science Foundation (No. 2022M711635), Jiangsu Funding Program for Excellent Postdoctoral Talent (No. 2022ZB267), Fundamental Research Funds for the Central Universities (No. 30923010303)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 232, + 197, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 197, + 243 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 197, + 243 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 257, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 257, + 480, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 257, + 480, + 279 + ], + "spans": [ + { + "bbox": [ + 138, + 257, + 480, + 279 + ], + "type": "text", + "content": "1. Ahn, J., Cho, S., Kwak, S.: Weakly supervised learning of instance segmentation with inter-pixel relations. In: CVPR. pp. 2209-2218 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 280, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 280, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 280, + 480, + 311 + ], + "type": "text", + "content": "2. Ahn, J., Kwak, S.: Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In: CVPR. pp. 4981-4990 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 312, + 480, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 480, + 334 + ], + "type": "text", + "content": "3. Bearman, A., Russakovsky, O., Ferrari, V., Fei-Fei, L.: What's the point: Semantic segmentation with point supervision. In: ECCV. pp. 549-565 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 335, + 480, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 480, + 356 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 480, + 356 + ], + "type": "text", + "content": "4. Cai, X., Lai, Q., Wang, Y., Wang, W., Sun, Z., Yao, Y.: Poly kernel inception network for remote sensing detection. In: CVPR. pp. 27706-27716 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 357, + 480, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 480, + 388 + ], + "type": "text", + "content": "5. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. In: IEEE TPAMI. vol. 40, pp. 834-848 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 390, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 480, + 411 + ], + "type": "text", + "content": "6. Chen, L., Lei, C., Li, R., Li, S., Zhang, Z., Zhang, L.: Fpr: False positive rectification for weakly supervised semantic segmentation. In: ICCV. pp. 1108-1118 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 412, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 480, + 434 + ], + "type": "text", + "content": "7. Chen, L., Wu, W., Fu, C., Han, X., Zhang, Y.T.: Weakly supervised semantic segmentation with boundary exploration. In: ECCV (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "type": "text", + "content": "8. Chen, Q., Yang, L., Lai, J.H., Xie, X.: Self-supervised image-specific prototype exploration for weakly supervised semantic segmentation. In: CVPR. pp. 4288-4298 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 467, + 480, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 480, + 499 + ], + "type": "text", + "content": "9. Chen, T., Xie, G.S., Yao, Y., Wang, Q., Shen, F., Tang, Z., Zhang, J.: Semantically meaningful class prototype learning for one-shot image segmentation. IEEE TMM 24, 968-980 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 500, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 500, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 500, + 480, + 521 + ], + "type": "text", + "content": "0. Chen, T., Yao, Y., Huang, X., Li, Z., Nie, L., Tang, J.: Spatial structure constraints for weakly supervised semantic segmentation. IEEE TIP 33, 1136-1148 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 522, + 480, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 480, + 554 + ], + "type": "text", + "content": "1. Chen, T., Yao, Y., Tang, J.: Multi-granularity denoising and bidirectional alignment for weakly supervised semantic segmentation. IEEE TIP 32, 2960-2971 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "type": "text", + "content": "2. Chen, T., Yao, Y., Zhang, L., Wang, Q., Xie, G.S., Shen, F.: Saliency guided inter- and intra-class relation constraints for weakly supervised semantic segmentation. IEEE TMM 25, 1727–1737 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 588, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 480, + 609 + ], + "type": "text", + "content": "3. Chen, Z., Sun, Q.: Extracting class activation maps from non-discriminative features as well. In: CVPR. pp. 3135-3144 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "type": "text", + "content": "4. Chen, Z., Wang, T., Wu, X., Hua, X.S., Zhang, H., Sun, Q.: Class re-activation maps for weakly-supervised semantic segmentation. In: CVPR. pp. 969–978 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "5. Cheng, Z., Qiao, P., Li, K., Li, S., Wei, P., Ji, X., Yuan, L., Liu, C., Chen, J.: Out-of-candidate rectification for weakly supervised semantic segmentation. In: CVPR. pp. 23673-23684 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 175, + 91, + 447, + 102 + ], + "type": "text", + "content": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "16. Dai, J., He, K., Sun, J.: Boxsup: Exploiting bounding boxes to supervise convolutional networks for semantic segmentation. In: ICCV. pp. 1635-1643 (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "17. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255 (2009)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "18. Du, Y., Fu, Z., Liu, Q., Wang, Y.: Weakly supervised semantic segmentation by pixel-to-prototype contrast. In: CVPR. pp. 4320-4329 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "19. Everingham, M., Gool, L.V., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. In: IJCV. vol. 88, pp. 303-338 (2010)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 236 + ], + "type": "text", + "content": "20. Fan, J., Zhang, Z., Song, C., Tan, T.: Learning integral objects with intra-class discriminator for weakly-supervised semantic segmentation. In: CVPR. pp. 4283-4292 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "21. Fan, J., Zhang, Z., Tan, T., Song, C., Xiao, J.: Cian: Cross-image affinity net for weakly supervised semantic segmentation. In: AAAI. vol. 34, pp. 10762-10769 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "text", + "content": "22. Hariharan, B., Arbeláez, P., Bourdev, L.D., Maji, S., Malik, J.: Semantic contours from inverse detectors. In: ICCV. pp. 991-998 (2011)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "24. Jiang, P.T., Yang, Y., Hou, Q., Wei, Y.: L2g: A simple local-to-global knowledge transfer framework for weakly supervised semantic segmentation. In: CVPR. pp. 16886-16896 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "25. Jing, L., Chen, Y., Tian, Y.: Coarse-to-fine semantic segmentation from image-level labels. IEEE TIP 29, 225-236 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "26. Jo, S., Yu, I.J., Kim, K.: Mars: Model-agnostic biased object removal without additional supervision for weakly-supervised semantic segmentation. arXiv preprint arXiv:2304.09913 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "27. Khoreva, A., Benenson, R., Hosang, J., Hein, M., Schiele, B.: Simple does it: Weakly supervised instance and semantic segmentation. In: CVPR. pp. 876-885 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "text", + "content": "28. Kim, B., Han, S., Kim, J.: Discriminative region suppression for weakly-supervised semantic segmentation. In: AAAI. vol. 35, pp. 1754-1761 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "text", + "content": "29. Kolesnikov, A., Lampert, C.: Seed, expand and constrain: Three principles for weakly-supervised image segmentation. In: ECCV. pp. 695-711 (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "30. Kweon, H., Yoon, S.H., Kim, H., Park, D., Yoon, K.J.: Unlocking the potential of ordinary classifier: Class-specific adversarial erasing framework for weakly supervised semantic segmentation. In: ICCV. pp. 6994-7003 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "text", + "content": "31. Lee, J., Choi, J., Mok, J., Yoon, S.: Reducing information bottleneck for weakly supervised semantic segmentation. In: NeurIPS. vol. 34, pp. 27408-27421 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "text", + "content": "32. Lee, J., Kim, E., Yoon, S.: Anti-adversarily manipulated attributions for weakly and semi-supervised semantic segmentation. In: CVPR. pp. 4071-4080 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 545, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 566 + ], + "type": "text", + "content": "33. Lee, J., Oh, S.J., Yun, S., Choe, J., Kim, E., Yoon, S.: Weakly supervised semantic segmentation using out-of-distribution data. In: CVPR. pp. 16897-16906 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "34. Lee, S., Lee, M., Lee, J., Shim, H.: Railroad is not a train: Saliency as pseudo-pixel supervision for weakly supervised semantic segmentation. In: CVPR. pp. 5495-5505 (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 600, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 620 + ], + "type": "text", + "content": "35. Li, J., Jie, Z., Wang, X., Wei, X., Ma, L.: Expansion and shrinkage of localization for weakly-supervised semantic segmentation. vol. 35, pp. 16037-16051 (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "text", + "content": "36. Li, X., Zhou, T., Li, J., Zhou, Y., Zhang, Z.: Group-wise semantic mining for weakly supervised semantic segmentation. In: AAAI. pp. 1984-1992 (2021)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "37. Lin, D., Dai, J., Jia, J., He, K., Sun, J.: Scribblesup: Scribble-supervised convolutional networks for semantic segmentation. In: CVPR. pp. 3159-3167 (2016)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "T.Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "38. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, L.: Microsoft coco: Common objects in context. In: ECCV. pp. 740-755 (2014)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "text", + "content": "39. Lin, Y., Chen, M., Wang, W., Wu, B., Li, K., Lin, B., Liu, H., He, X.: Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In: CVPR. pp. 15305-15314 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 203 + ], + "type": "text", + "content": "40. Liu, W., Zhang, C., Lin, G., Hung, T.Y., Miao, C.: Weakly supervised segmentation with maximum bipartite graph matching. In: ACM MM. pp. 2085-2094 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "41. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR. pp. 3431-3440 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 246 + ], + "type": "text", + "content": "42. Pei, G., Chen, T., Jiang, X., Liu, H., Sun, Z., Yao, Y.: Videomac: Video masked autoencoders meet convnets. In: CVPR. pp. 22733-22743 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 246, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 278 + ], + "type": "text", + "content": "43. Peng, Z., Wang, G., Xie, L., Jiang, D., Shen, W., Tian, Q.: Usage: A unified seed area generation paradigm for weakly supervised semantic segmentation. arXiv preprint arXiv:2303.07806 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 310 + ], + "type": "text", + "content": "44. Qin, J., Wu, J., Xiao, X., Li, L., Wang, X.: Activation modulation and recalibration scheme for weakly supervised semantic segmentation. In: AAAI. vol. 36, pp. 2117-2125 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 310, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 480, + 332 + ], + "type": "text", + "content": "45. Rong, S., Tu, B., Wang, Z., Li, J.: Boundary-enhanced co-training for weakly supervised semantic segmentation. In: CVPR. pp. 19574-19584 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "type": "text", + "content": "46. Rossetti, S., Zappia, D., Sanzari, M., Schaerf, M., Pirri, F.: Max pooling with vision transformers reconciles class and shape in weakly supervised semantic segmentation. In: ECCV. pp. 446-463. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "text", + "content": "47. Ru, L., Zhan, Y., Yu, B., Du, B.: Learning affinity from attention: End-to-end weakly-supervised semantic segmentation with transformers. In: CVPR. pp. 16846-16855 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "text", + "content": "48. Ru, L., Zheng, H., Zhan, Y., Du, B.: Token contrast for weakly-supervised semantic segmentation. In: CVPR. pp. 3093-3102 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "type": "text", + "content": "49. Saleh, F., Aliakbarian, M.S., Salzmann, M., Petersson, L., Gould, S., Alvarez, J.M.: Built-in foreground/background prior for weakly-supervised semantic segmentation. In: ECCV. pp. 413-432 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "type": "text", + "content": "50. Sheng, M., Sun, Z., Cai, Z., Chen, T., Zhou, Y., Yao, Y.: Adaptive integration of partial label learning and negative learning for enhanced noisy label learning. In: AAAI. vol. 38, pp. 4820-4828 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "text", + "content": "51. Shimoda, W., Yanai, K.: Self-supervised difference detection for weakly-supervised semantic segmentation. In: ICCV. pp. 5208-5217 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "52. Singh, K.K., Yu, H., Sarmasi, A., Pradeep, G., Lee, Y.J.: Hide-and-seek: A data augmentation technique for weakly-supervised localization and beyond. arXiv preprint arXiv:1811.02545 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "text", + "content": "53. Song, C., Huang, Y., Ouyang, W., Wang, L.: Box-driven class-wise region masking and filling rate guided loss for weakly supervised semantic segmentation. In: CVPR. pp. 3136-3145 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 568, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 590 + ], + "type": "text", + "content": "54. Su, Y., Sun, R., Lin, G., Wu, Q.: Context decoupling augmentation for weakly supervised semantic segmentation. In: ICCV (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "type": "text", + "content": "55. Sun, K., Shi, H., Zhang, Z., Huang, Y.: ECS-net: Improving weakly supervised semantic segmentation by using connections between class activation maps. In: ICCV. pp. 7283-7292 (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "56. Vernaza, P., Chandraker, M.: Learning random-walk label propagation for weakly-supervised semantic segmentation. In: CVPR. pp. 7158-7166 (2017)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "57. Wang, W., Sun, G., Van Gool, L.: Looking beyond single images for weakly supervised semantic segmentation learning. IEEE TPAMI (2022)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 176, + 91, + 447, + 102 + ], + "type": "text", + "content": "Knowledge Transfer with Simulated Inter-Image Erasing for WSSS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 544 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "58. Wang, X., Liu, S., Ma, H., Yang, M.H.: Weakly-supervised semantic segmentation by iterative affinity learning. vol. 128, pp. 1736-1749. Springer (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "text", + "content": "59. Wang, Y., Zhang, J., Kan, M., Shan, S., Chen, X.: Self-supervised equivariant attention mechanism for weakly supervised semantic segmentation. In: CVPR. pp. 12275-12284 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "text", + "content": "60. Wei, Y., Feng, J., Liang, X., Cheng, M.M., Zhao, Y., Yan, S.: Object region mining with adversarial erasing: A simple classification to semantic segmentation approach. In: CVPR. pp. 1568-1576 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "61. Wei, Y., Xiao, H., Shi, H., Jie, Z., Feng, J., Huang, T.: Revisiting dilated convolution: A simple approach for weakly-and semi-supervised semantic segmentation. In: CVPR. pp. 7268-7277 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 258 + ], + "type": "text", + "content": "62. Wu, Z., Shen, C., Van Den Hengel, A.: Wider or deeper: Revisiting the resnet model for visual recognition. PR 90, 119-133 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "63. Xie, J., Hou, X., Ye, K., Shen, L.: Clims: Cross language image matching for weakly supervised semantic segmentation. In: CVPR. pp. 4483-4492 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "type": "text", + "content": "64. Xu, L., Ouyang, W., Bennamoun, M., Boussaid, F., Sohel, F., Xu, D.: Leveraging auxiliary tasks with affinity learning for weakly supervised semantic segmentation. In: ICCV. pp. 6984-6993 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "65. Xu, L., Ouyang, W., Bennamoun, M., Boussaid, F., Xu, D.: Multi-class token transformer for weakly supervised semantic segmentation. In: CVPR. pp. 4310-4319 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "66. Yao, Y., Chen, T., Xie, G.S., Zhang, C., Shen, F., Wu, Q., Tang, Z., Zhang, J.: Non-salient region object mining for weakly supervised semantic segmentation. In: CVPR. pp. 2623-2632 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "67. Yoon, S.H., Kweon, H., Cho, J., Kim, S., Yoon, K.J.: Adversarial erasing framework via triplet with gated pyramid pooling layer for weakly supervised semantic segmentation. In: ECCV. pp. 326-344. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "text", + "content": "68. Zhang, D., Zhang, H., Tang, J., Hua, X.S., Sun, Q.: Causal intervention for weakly-supervised semantic segmentation. In: NeurIPS. vol. 33 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 480, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 457 + ], + "type": "text", + "content": "69. Zhang, X., Wei, Y., Feng, J., Yang, Y., Huang, T.S.: Adversarial complementary learning for weakly supervised object localization. In: CVPR. pp. 1325-1334 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "text", + "content": "70. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. In: AAAI. vol. 34, pp. 13001-13008 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "71. Zhou, B., Khosla, A., Lapedriza, Å., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: CVPR. pp. 2921-2929 (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "text", + "content": "72. Zhou, T., Li, L., Li, X., Feng, C.M., Li, J., Shao, L.: Group-wise learning for weakly supervised semantic segmentation. IEEE TIP 31, 799-811 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "text", + "content": "73. Zhou, T., Zhang, M., Zhao, F., Li, J.: Regional semantic contrast and aggregation for weakly supervised semantic segmentation. In: CVPR. pp. 4299-4309 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "T.Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Knowledge-enhanced Visual-Language Pretraining for Computational Pathology/a04fca37-34eb-47aa-8e9d-2ee68fcc1f47_content_list.json b/2024/Knowledge-enhanced Visual-Language Pretraining for Computational Pathology/a04fca37-34eb-47aa-8e9d-2ee68fcc1f47_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8c1f0c3623bac0d80e951ba5de09141350dbf99d --- /dev/null +++ b/2024/Knowledge-enhanced Visual-Language Pretraining for Computational Pathology/a04fca37-34eb-47aa-8e9d-2ee68fcc1f47_content_list.json @@ -0,0 +1,1775 @@ +[ + { + "type": "text", + "text": "Knowledge-enhanced Visual-Language Pretraining for Computational Pathology", + "text_level": 1, + "bbox": [ + 264, + 140, + 736, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiao Zhou1, Xiaoman Zhang1,2, Chaoyi Wu1,2, Ya Zhang1,2, Weidi Xie1,2, Yanfeng Wang1,2", + "bbox": [ + 321, + 210, + 676, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Artificial Intelligence Laboratory", + "bbox": [ + 349, + 256, + 651, + 270 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Shanghai Jiao Tong University", + "bbox": [ + 390, + 271, + 611, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/MAGIC-AI4Med/KEP", + "bbox": [ + 364, + 286, + 638, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhouxiao@pjlab.org.cn, {xm99sjtu, wtzxxxxwy02, ya_zhang, weidi,", + "bbox": [ + 256, + 300, + 741, + 313 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wangyanfeng} @sjtu.edu.cn", + "bbox": [ + 403, + 314, + 598, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this paper, we consider the problem of visual representation learning for computational pathology, by exploiting large-scale image-text pairs gathered from public resources, along with the domain-specific knowledge in pathology. Specifically, we make the following contributions: (i) We curate a pathology knowledge tree that consists of 50,470 informative attributes for 4,718 diseases requiring pathology diagnosis from 32 human tissues. To our knowledge, this is the first comprehensive structured pathology knowledge base; (ii) We develop a knowledge-enhanced visual-language pretraining approach, where we first project pathology-specific knowledge into latent embedding space via a language model, and use it to guide the visual representation learning; (iii) We conduct thorough experiments to validate the effectiveness of our proposed components, demonstrating significant performance improvement on various downstream tasks, including cross-modal retrieval, zero-shot classification on pathology patches, and zero-shot tumor subtyping on whole slide images (WSIs).", + "bbox": [ + 259, + 364, + 746, + 587 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Pathology knowledge $\\cdot$ Visual-language pretraining", + "bbox": [ + 261, + 599, + 691, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 641, + 375, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pathology diagnosis is currently the golden standard for examining various diseases in clinical applications, especially in the diagnosis of neoplasm [43]. In the last decade, the prosperity of deep learning on computer vision has led to the rapid development of computational pathology, for example, approaches with supervised learning [6,22,34,45,46], and weakly supervised learning [5,11,32,38,59]. Despite being promising, these approaches have been fundamentally limited by the scale of costly labels. Alternatively, self-supervised pretraining on numerous unlabeled pathological images [8,9,26,48] has attracted unprecedented attention, yet it still requires supervised fine-tuning for downstream deployments.", + "bbox": [ + 212, + 672, + 787, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the recent literature [25,41], studies on the multimodal foundation model have demonstrated conspicuous improvement in downstream zero-shot tasks, by", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a239cf8b1e899a359e14ef525856e627b734f565f07d6dc364c42af627c3eec6.jpg", + "image_caption": [ + "Fig. 1: Knowledge-enhanced pathology image-text alignment. The short caption of a pathology image crawled from public websites is typically unstructured and with varying granularities, which introduces noticeable ambiguities for image-text alignment. While the implicit structures and correlations between different image-caption pairs could be constructed by explicit disease attributes (dashed boxes), which can be well-aligned by a pathology knowledge tree. LUSC and LUAD suggest lung squamous cell carcinoma, and lung adenocarcinoma, respectively." + ], + "image_footnote": [], + "bbox": [ + 320, + 143, + 679, + 306 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "simply training to align visual and language embedding space on massive image-text pairs crawled from public websites. In contrast to computer vision [19,25,41], representation learning in pathology requires tremendous expertise and domain knowledge. An ideal training corpus would thus consist of well-structured medical reports from the hospital, however, it is often extremely difficult to acquire, due to privacy concerns. Alternatively, recent works [23, 24, 36, 37] propose to gather large-scale image-caption pairs from public resources (PubMed [42] papers, Twitter, Youtube videos). Compared to discrete image labels, image captions from medical reports and academic articles can potentially provide more valuable medical information without manual annotation.", + "bbox": [ + 212, + 460, + 787, + 612 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For existing works that adopt simple contrastive learning on pathology image and caption pairs, they suffer from the following challenges: First, the short captions from these web-crawled image-text pairs are typically noisy, unstructured, and lack domain knowledge (Fig. 1), which can be sub-optimal for constructing high-quality pathology-specific visual representations; Second, the varying granularities of the free text from captions will inevitably introduce ambiguities when aligning with images, thus causing the model to be highly sensitive to the used text prompts during the inference stage [36]. Thus, effectively harnessing the potential of such web-crawled data remains to be a challenge.", + "bbox": [ + 212, + 619, + 787, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, to tackle the above challenges, we anticipate that introducing pathology knowledge is of great significance to make up for the deficiency of short image captions. To this end, we make the following contributions: (i) We curate a pathology knowledge tree, PathKT, by collecting 50,470 informative pathological attributes of 4,718 diseases in 32 tissues from publicly available educational", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "X. Zhou et al.", + "bbox": [ + 271, + 114, + 366, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "resources and OncoTree $^{3}$ [31]. (ii) We propose a novel knowledge encoder pretraining approach, that projects the attributes of each disease from PathKT into latent embedding space, where the attributes of the same disease, including disease synonyms, definitions, histology, and cytology features, share similar representations. (iii) We develop a knowledge-enhanced pretraining (KEP) approach to align pathology visual-language representations, which freezes the knowledge encoder and continuously injects domain-specific knowledge into the image-text embedding space. To demonstrate the effectiveness of our proposed approach, we conduct thorough experiments on three downstream tasks, including retrieval on three pathological image-caption datasets, zero-shot patch classification on eight patch-level pathology image datasets, and zero-shot WSI tumor subtyping on three datasets from The Cancer Genome Atlas (TCGA) $^{4}$ . Quantitative experiments suggest that knowledge guidance can significantly enhance the performance across different tasks.", + "bbox": [ + 212, + 145, + 787, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 382, + 382, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-Language Pretraining (VLP). Current vision-language pretraining approaches are typically categorized into two groups. The first group, referred to as two-stream approaches [41, 51], involves using two separate encoders to extract features for visual and textual data, respectively. The second group, known as single-stream methods [10, 29], utilizes a cross-modal fusion encoder to enhance interactions between vision and text features. In the field of medical VLP, existing methods mainly adopt the two-stream approach [7, 12, 20, 35, 39, 56, 58]. Despite the valuable contributions of these methods, the reliance on data-driven representation learning restricts the utilization of systematic and structured medical knowledge, leading to less than professional diagnosis.", + "bbox": [ + 212, + 414, + 787, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VLP in Computational Pathology. In the recent literature, Huang et al. proposed a pathology VLP model named PLIP [23] and released OpenPath that contains 200K image-caption pairs from the social media of Twitter and other public resources. To extend the pathology data scale, Ikezogwo et al. curate Quilt1m [24] with over one million histopathology image-text pairs by capturing keyframes and speech from YouTube videos. In addition to evaluating the zero-shot classification on patch-level pathology images, Lu et al. propose MI-Zero [37] and CONCH [36] to extend the transfer ability of pathology VLP models on gigapixel whole slide images (WSIs). In contrast to these existing work, that finetunes a CLIP [41] or CoCa [55] on pathology image-caption pairs, we propose to first train a pathology-specific knowledge encoder, and use it to guide visual-language representation learning.", + "bbox": [ + 212, + 570, + 787, + 751 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Medical Knowledge-enhanced Learning. In the medical community, leveraging external medical knowledge to enhance deep learning models has become an increasingly important topic [49, 50, 57]. Generally, existing approaches can", + "bbox": [ + 212, + 755, + 787, + 801 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "KEP", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3 https://oncotree.info/", + "bbox": [ + 217, + 809, + 406, + 825 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "4 https://portal.gdc.cancer.gov/", + "bbox": [ + 222, + 825, + 468, + 839 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0b346c9114d5f910f02e917024335522a54969e1e35a14d0aabf055f0a15b2e9.jpg", + "image_caption": [ + "Fig. 2: The construction of pathology knowledge tree. OncoTree is adopted as the base architecture to construct the PathKT. The tissue types, disease entities, and attributes are first extracted from web-crawled pathological descriptions, where cancers are then matched to OncoTree based on their tissue types and tumor types/subtypes using UMLS CUIs. Moreover, non-tumor diseases are added to the knowledge tree according to their tissue types. Finally, the pathology knowledge tree integrates 4718 diseases from 32 tissues, with each disease containing various synonyms, definitions, and histological and cytological features. CSCLC in this figure suggests combined small cell lung cancer." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 784, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "be categorized based on the ways of using medical knowledge: model-based approaches [14, 21] adopts the prior knowledge of radiology or diagnosis summarized by doctors to design algorithms; and input-based methods [13, 33, 53, 54] directly exploit knowledge as the external input to guide representation training. However, most of these works are focused on the analysis of chest X-rays.", + "bbox": [ + 212, + 435, + 787, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methods", + "text_level": 1, + "bbox": [ + 215, + 532, + 339, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our primary goal is to leverage structured pathology knowledge to enhance visual-language representation learning. To start with, we construct a **Pathology Knowledge Tree**, termed as **PathKT**, that consists of 50,470 informative attributes of 4718 diseases from 32 human tissues (Sec. 3.1). We then train a knowledge encoder that projects the structured pathology knowledge into an embedding space (Sec. 3.2). We further employ the knowledge encoder to guide visual-language pretraining for computational pathology, termed as KEP (Knowledge-Enhanced Pre-training, Sec. 3.3).", + "bbox": [ + 212, + 564, + 787, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 PathKT Construction", + "text_level": 1, + "bbox": [ + 215, + 705, + 442, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, we detail the procedure for building up a pathology knowledge tree with various online sources, which will be further used for knowledge encoding.", + "bbox": [ + 212, + 729, + 785, + 761 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Knowledge Source. We collect pathology-specific knowledge from publicly available educational resources, such as text books, professional websites, and structured databases (OncoTree [31]). Specifically, we extract pathological descriptions of 884 tumor subtypes from OncoTree and 4360 diseases from text books and professional websites, including all domain knowledge required for", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "X. Zhou et al.", + "bbox": [ + 271, + 114, + 366, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "diagnosis in clinical scenarios, i.e., disease name/ synonyms, definitions, histological features, and cytological features.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Knowledge Tree Construction. We structure these knowledge sources into a knowledge tree by expanding the OncoTree, as shown in Fig. 2. OncoTree is a tree-structure cancer classification system [31], which consists of 884 tumor subtypes from 32 tissues, with each type of tumor linked to a Concept Unique Identifier (CUI) from Unified Medical Language System (UMLS) [2]. Specifically, we first extract tissue types, disease entities, and disease attributes from pathological descriptions of 4360 diseases, in which 168 cancers are found overlapped with OncoTree by using SciSpacy package [40] to link the UMLS CUIs. The rest 4192 diseases are then added to the knowledge tree according to their tissue types. After deduplication and noise reduction (358 diseases without any informative descriptions are deleted), all diseases are organized into the corresponding tissues. The final histopathology knowledge tree contains 4718 diseases from 32 tissues. The attributes of each disease node are constructed by a varying number of synonyms, definitions, and histological and cytological descriptions. The statistics of PathKT is shown in Table S1 and Fig. S1 in Supplementary Materials. In the final PathKT, for instance, the tissue node of the lung connects all lung diseases, including combined small cell lung cancer (CSCLC), lung squamous cell carcinoma (LUSC), and lung adenocarcinoma (LUAD), etc. The LUSC node connects four kinds of attributes, shown in Fig. 2.", + "bbox": [ + 212, + 180, + 787, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Pathology Knowledge Encoding", + "text_level": 1, + "bbox": [ + 215, + 493, + 524, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we describe details for projecting tree-structure pathological knowledge into a latent embedding space, by training a knowledge encoder. Specifically, we align different disease entities with their corresponding free-text attributes via metric learning, such that, the synonyms, definitions, and corresponding histological/cytological features are close in the embedding space. As a consequence, the model can link the pathology images to their implicit disease labels during the visual-language pertaining, since the free-form text descriptions in image-text pairs may contain disease attributes, such as pathological features and disease definitions, which have already been aligned with disease names/synonyms. Therefore, the alignment of pathology knowledge can help the model link to the disease entities for better performance during diagnosis.", + "bbox": [ + 212, + 521, + 787, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem Setting. Given a set of disease entities with their corresponding attributes, $\\mathcal{D} = \\{(d_1,\\mathbf{a}_1),\\ldots ,(d_n,\\mathbf{a}_n)\\}$ , where $d_{i}$ denotes the $i$ -th disease entity, and $\\mathbf{a}_i = \\{a_i^1,\\dots ,a_i^k\\}$ refer to the associated $k$ attributes, both disease and attributes are represented in the format of natural language. Note that, for different disease entities, $k$ also varies, our goal here is to train a model that satisfies:", + "bbox": [ + 212, + 691, + 787, + 768 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s i m} \\left(\\Phi_ {\\mathrm {k}} \\left(a _ {i} ^ {p}\\right), \\Phi_ {\\mathrm {k}} \\left(a _ {i} ^ {q}\\right)\\right) \\gg \\operatorname {s i m} \\left(\\Phi_ {\\mathrm {k}} \\left(a _ {i} ^ {p}\\right), \\Phi_ {\\mathrm {k}} \\left(a _ {j} ^ {t}\\right)\\right), i \\neq j, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 780, + 785, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varPhi_{\\mathbf{k}}(\\cdot)$ denotes the knowledge encoder, $\\mathrm{sim}(\\cdot ,\\cdot)$ refers to the similarity, $a_i^p,a_i^q$ and $a_j^t$ refer to the randomly sampled attributes from the $i,j$ -th disease entity.", + "bbox": [ + 212, + 808, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "KEP", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dd3628686907ec3b9bdb5e0fb419731044620b613a54f42695713b29f2476bf3.jpg", + "image_caption": [ + "Fig. 3: Knowledge encoder pretraining based on metric learning. $n$ disease entities and each with $k$ attributes, including disease synonyms, definitions, cytology and pathology features, construct a mini-batch (left part of the figure), which are fed to a knowledge encoder for pretraining. In the embedding space (right part of the figure), the markers in different shapes represent the embeddings of attributes of different diseases. $S_{i}^{+}$ suggests the max-min positive attribute similarity within the $i$ -th disease, while $S_{i}^{-}$ denotes the maximal attribute similarity between the $i$ -th disease and other diseases. The goal of metric learning is to increase $S_{i}^{+}$ and meanwhile decreasing $S_{i}^{-}$ . The purple-dashed arrow and circle denote the minimal positive attribute similarity in the second class and the hypersphere it spans." + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 777, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Intuitively, the knowledge encoder enables the attributes of the same disease to be pulled together, while attributes from different diseases are pushed apart.", + "bbox": [ + 212, + 465, + 784, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training. In order to achieve the objective defined in Eq. 1, metric learning is exploited to construct an embedding space where the representations of intra/inter-class instances are clustered/separated. In specific, given a mini-batch that contains $n$ random diseases and each with $k$ attributes, we denote the normalized embedding for the $p$ -th attribute of $i$ -th disease as: $\\mathbf{z}_p^i = \\varPhi_{\\mathrm{k}}(a_i^p) / \\| \\varPhi_{\\mathrm{k}}(a_i^p)\\|$ . We adopt the recently proposed AdaSP loss [60], which finds out a max-min positive similarity (Fig. 3) and then shapes a loss with the maximal negative similarity:", + "bbox": [ + 212, + 500, + 787, + 606 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m e t i c}} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\log \\left(1 + e ^ {\\left(S _ {i} ^ {-} - S _ {i} ^ {+}\\right) / \\tau}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 616, + 785, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tau$ is a temperature parameter. $S_{i}^{+}$ and $S_{i}^{-}$ denote the max-min positive and the maximal negative similarity, which can be computed by the soft version:", + "bbox": [ + 212, + 662, + 787, + 694 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS _ {i} ^ {+} = \\max _ {p} \\min _ {q} \\left\\langle \\mathbf {z} _ {p} ^ {i}, \\mathbf {z} _ {q} ^ {i} \\right\\rangle \\approx \\tau \\log \\left(\\sum_ {p = 1} ^ {k} \\frac {1}{\\sum_ {q = 1} ^ {k} e ^ {- \\frac {\\left\\langle \\mathbf {z} _ {p} ^ {i} , \\mathbf {z} _ {q} ^ {i} \\right\\rangle}{\\tau}}}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 704, + 785, + 757 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS _ {i} ^ {-} = \\max _ {j, p, q} \\left\\langle \\mathbf {z} _ {p} ^ {i}, \\mathbf {z} _ {q} ^ {j} \\right\\rangle \\approx \\tau \\log \\left(\\sum_ {p = 1} ^ {k} \\sum_ {j = 1, j \\neq i} ^ {n} \\sum_ {q = 1} ^ {k} e ^ {\\frac {\\left\\langle \\mathbf {z} _ {p} ^ {i} , \\mathbf {z} _ {q} ^ {j} \\right\\rangle}{\\tau}}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 761, + 785, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\langle \\cdot \\rangle$ represents the cosine similarity, the details about the soft version can be found in Supplementary Materials.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "X. Zhou et al.", + "bbox": [ + 271, + 114, + 366, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Discussion. The conventional triplet loss [18] with batch hard mining strategy is widely adopted in metric learning, while it is not suitable in our case due to the following two reasons: First, in the pathology knowledge, each disease entity is associated with at most four types of attributes, the text descriptions of different attribute types might reveal significant divergence, which causes a large intra-class variation, marked by the purple dashed circle in Fig. 3, during training. Second, the fine-grained diseases might share high similarities, such as tumor subtypes: breast invasive ductal carcinoma and breast invasive lobular carcinoma, which cause low inter-class variations during training. When a mini-batch meets these two conditions, traditional triplet loss that enforces the instance similarities could typically produce the bad local minima of optimization [52], which, therefore, undermines the metric learning.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Pathology Knowledge Enhanced Pretraining", + "text_level": 1, + "bbox": [ + 214, + 347, + 627, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we present a simple yet effective pretraining approach, termed KEP, that leverages the established knowledge encoder to guide visual-language pretraining for computational pathology.", + "bbox": [ + 212, + 369, + 785, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual-Language Pretraining. Given paired image and captions, denoted as $\\mathcal{F} = \\{(x_1, c_1), \\ldots, (x_n, c_n)\\}$ , our goal is to construct a visual-language embedding space from paired image-text data, that satisfies:", + "bbox": [ + 214, + 422, + 785, + 468 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\operatorname {s i m} \\left(\\Phi_ {\\mathrm {v}} \\left(x _ {i}\\right), \\Phi_ {\\mathrm {t}} \\left(c _ {i}\\right)\\right) \\gg \\operatorname {s i m} \\left(\\Phi_ {\\mathrm {v}} \\left(x _ {i}\\right), \\Phi_ {\\mathrm {t}} \\left(c _ {j}\\right)\\right), i \\neq j, \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 477, + 784, + 493 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\varPhi_{\\mathrm{v}}(\\cdot)$ and $\\varPhi_{\\mathrm{t}}(\\cdot)$ denote the visual and the textual encoder, respectively. In this work, ViT-B-32/16 is adopted as the backbone for the visual encoder and initialized from the visual weights of CLIP [41] / BiomedCLIP [56]. A projection head is added on top of the visual encoder to bridge the gap caused by the non-pathology initialization, shown in Fig. 4. The text encoder is initialized by the pretrained knowledge encoder, as described in Sec. 3.2. In order to learn effective visual-language representation, we optimise an infoNCE contrastive objective:", + "bbox": [ + 212, + 501, + 787, + 609 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {v t}} = - \\left(\\log \\frac {e ^ {\\mathbf {v} _ {i} ^ {T} \\mathbf {t} _ {i} / \\tau}}{\\sum_ {j = 1} ^ {n} e ^ {\\mathbf {v} _ {i} ^ {T} \\mathbf {t} _ {j} / \\tau}} + \\log \\frac {e ^ {\\mathbf {t} _ {i} ^ {T} \\mathbf {v} _ {i} / \\tau}}{\\sum_ {j = 1} ^ {n} e ^ {\\mathbf {t} _ {i} ^ {T} \\mathbf {v} _ {j} / \\tau}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 614, + 784, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{v}_i = \\varPhi_{\\mathrm{v}}(x_i)$ , and $\\mathbf{t}_i = \\varPhi_{\\mathrm{t}}(c_i)$ , refer to the normalized embedding vectors from visual and text encoders, respectively. $\\tau$ denotes a temperature parameter.", + "bbox": [ + 214, + 657, + 787, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Knowledge Distillation. To keep the alignment between images and free-form captions inside the knowledge space and thus the images can be linked to their implicit disease entities (right part in Fig. 4), we adopt an additional frozen branch to continuously distill pathology knowledge to the text encoder. Specifically, we use the text-knowledge embedding pairs to construct a contrastive loss item $\\mathcal{L}_{\\mathrm{tk}}$ , which can be computed by Eq. 6 with the normalized visual embedding vectors replaced by the knowledge embedding vectors:", + "bbox": [ + 212, + 694, + 787, + 800 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {t k}} = - \\left(\\log \\frac {e ^ {\\mathbf {k} _ {i} ^ {T} \\mathbf {t} _ {i} / \\tau}}{\\sum_ {j = 1} ^ {n} e ^ {\\mathbf {k} _ {i} ^ {T} \\mathbf {t} _ {j} / \\tau}} + \\log \\frac {e ^ {\\mathbf {t} _ {i} ^ {T} \\mathbf {k} _ {i} / \\tau}}{\\sum_ {j = 1} ^ {n} e ^ {\\mathbf {t} _ {i} ^ {T} \\mathbf {k} _ {j} / \\tau}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 806, + 785, + 844 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "KEP", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6614bfe1a57a81c0f11e2fffb9c0ba55b16a9d4704a935fb83700232a46369b4.jpg", + "image_caption": [ + "Fig. 4: Model architecture (left graph). A projection head is added on the top of the visual encoder to bridge the gap between the image and the text encoder. The knowledge encoder is frozen across the whole training stage to distill pathology knowledge to the learnable text encoder. As a result, the pathology images can be aligned with their implicit disease labels (marked by dashed boxes in the right graph) during visual-language pretraining, since the captions contain disease attributes that have been already aligned with disease names/synonyms in the knowledge embedding space." + ], + "image_footnote": [], + "bbox": [ + 230, + 145, + 776, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{k}_i = \\varPhi_{\\mathrm{k}}(c_i)$ , refer to the normalized embedding vectors from the knowledge encoder. The overall training loss can thus be computed as:", + "bbox": [ + 212, + 383, + 787, + 415 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {v t}} + \\alpha \\mathcal {L} _ {\\mathrm {t k}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 442, + 424, + 784, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\alpha$ denotes a weight parameter. It is worth emphasizing that the key contribution in KEP is to initialize the text encoder with the pre-obtained pathology knowledge encoder and adopt it to continuously distill pathology knowledge to the text encoder, while for the other parts, we keep the same as PLIP [23].", + "bbox": [ + 212, + 448, + 787, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 531, + 375, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we first introduce the datasets used for training and evaluation in this paper, followed by the evaluation metrics and implementation details.", + "bbox": [ + 212, + 563, + 787, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Training Datasets", + "text_level": 1, + "bbox": [ + 215, + 614, + 410, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset for Knowledge Encoding. We carry out pathology knowledge encoding with PathKT, where 4,718 disease nodes with a total number of 50,470 attributes are extracted, shown in Table S1 in Supplementary Materials.", + "bbox": [ + 212, + 640, + 787, + 685 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets for KEP Pretraining. Our dataset is composed of two parts. First, we collect the OpenPath data provided by PLIP [23] and obtain 138,874 image-text pairs after denoising and pre-processing. Second, we collect the Quilt1M [24] dataset, which gathers pathology image-text pairs from four public sources: PubMed articles, LAION [44], OpenPath [23], and Youtube videos. Since one of the downstream evaluation datasets Arch-PubMed [15] contains pathology images from PubMed articles, we remove this data source from Quilt1M to avoid data leaking. Considering that many images in Quilt1M are matched with multiple captions, we concatenate the captions related to the same image and finally obtain 576,608 image-text pairs for the Quilt1M dataset.", + "bbox": [ + 212, + 689, + 789, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "X. Zhou et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Downstream Tasks", + "text_level": 1, + "bbox": [ + 215, + 146, + 416, + 159 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate the pretrained models on three tasks, namely, retrieval, zero-shot patch classification, and zero-shot WSI tumor subtyping. The details for all evaluation datasets are exhibited in Table S2 in Supplementary Materials.", + "bbox": [ + 212, + 176, + 782, + 222 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Retrieval. This task involves cross-modal retrieval and disease retrieval. Cross-modal retrieval aims to retrieve the correct caption for a given image and vice versa. Disease retrieval, on the other hand, utilizes the disease names to retrieve captions or images with the same disease label, which is proposed to demonstrate the effectiveness of the knowledge encoder. For cross-modal retrieval, we follow PLIP [23] to split the ARCH [15] dataset into Arch-PubMed and Arch-book. In addition, we also gather image-text pairs from publicly available educational resources and curate a retrieval dataset, termed as PathPair, consisting of 9,358 pathology image-caption pairs with known 1676 disease labels. For disease retrieval, we utilize the captions, images, and their disease labels in PathPair.", + "bbox": [ + 212, + 226, + 784, + 377 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Zero-shot Patch Classification. This task involves zero-shot classification on patch-level pathology images. Specifically, at inference time, we randomly select one template from the 21 templates in CONCH [36] and one type synonym from the corresponding name list (exhibited in Supplementary Materials) to yield a text prompt for each type, e.g. a histopathological image of CLASSNAME. (template) + beast invasive carcinoma (synonym) $\\rightarrow$ a histopathological image of beast invasive carcinoma. This process is repeated 100 times in every experiment. Following PLIP [23] and Quilt1M [24], we adopt these patch-level pathology image datasets: BACH [1], NCT-CRC-HE-100K [27], KatherColon [28], LC25000 [3], RenalCell [4], SICAP [47], SkinCancer [30], WSSS4LUAD [17]. Each dataset includes multiple types of H&E stained cell micrographs.", + "bbox": [ + 212, + 382, + 785, + 547 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Zero-shot WSI Tumor Subtyping. This task involves zero-shot tumor subtyping on pathology whole slide images of common and rare cancers. We follow MI-Zero [37] and CONCH [36] for evaluation. Specifically, we first divide WSI into $256 \\times 256$ patches and then predict the class label of each image patch in a zero-shot manner. The tumor type of the whole slide is then provided by integrating Top-K predictions on patches. For this task, we also employ the templates and tumor synonyms from CONCH [36] to randomly generate 100 text prompts (listed in Supplementary materials) for each tumor subtype. (i) For common cancers, we follow existing research [36,37] and collect 525 WSIs with three tumor types, including 150 breast carcinoma (BRCA), 150 non-small cell lung cancer (NSCLC) histopathology slides, and 225 renal cell carcinoma (RCC) slides, from TCGA. Specifically, BRCA consists of two subtypes: invasive ductal carcinoma (IDC) and invasive lobular carcinoma (IDC). NSCLC contains lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC). RCC is divided into chromophobe renal cell carcinoma (CHRCC), clear-cell renal cell carcinoma (CCRCC), and papillary renal cell carcinoma (PRCC). Each tumor subtype has 75 WSIs. (ii) For rare cancers, we collected 41 WSIs of rare subtypes for breast cancer (6 for Intraductal papillary adenocarcinoma with invasion, 6 for Medullary carcinoma, NOS, 13 for Metaplastic carcinoma, NOS and", + "bbox": [ + 212, + 551, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "KEP", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "16 for Mucinous adenocarcinoma), from TCGA and merge them with common breast cancers (75 for IDC and 75 for ILC).", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Evaluation Metrics.", + "text_level": 1, + "bbox": [ + 215, + 198, + 421, + 213 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Retrieval. To evaluate the retrieval performance, we adopt the Recall@K metric, suggesting the ratio of correctly retrieved queries in Top-K retrieved samples.", + "bbox": [ + 215, + 222, + 784, + 253 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Zero-shot Patch Classification. For zero-shot classification tasks, we adopt the same metric as PLIP [23], namely, weighted F1 (wF1). We report the median, the first, and the third quartile (Q1, Q3) of wF1 across all text 100 prompts.", + "bbox": [ + 215, + 256, + 784, + 301 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Zero-shot WSI Tumor Subtyping. We adopt the same metric as MI-Zero [37], namely, balanced accuracy at Top-K pooling to measure the zero-shot tumor subtyping performance on WSIs. We also report the median, the first, and the third quartile (Q1, Q3) of the balanced accuracy across all 100 text prompts.", + "bbox": [ + 215, + 305, + 785, + 367 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 387, + 455, + 402 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Knowledge Encoder Pretraining. We adopt the architecture of PubMedBERT [16] to encode knowledge. The embedding dimension is set to 512. The temperature parameter $\\tau$ is set to 0.04 in Eq. 2. The batch size is set to 256, including 32 disease entities with 8 instances per entity.", + "bbox": [ + 215, + 415, + 784, + 476 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Pathology Image-text Pretraining. To achieve a fair comparison with PLIP and Quilt1M, we utilize the same visual encoder (ViT-B-32) and initialization of the visual encoder (CLIP [41]), termed by KEP-32, and set the input image size to $224 \\times 224$ . Additionally, we also develop a KEP variant named KEP-16 that is initialized by the visual weights of BiomedCLIP [56] (ViT-B-16). The batch size and learning rate are set to 256 and 1e-5, respectively. The temperature in Eq. 6 and Eq. 7 is set to 0.04 across all experiments. For OpenPath, we conduct the pathology VLP for 30 epochs, While for Quilt1M, 15 epochs are adopted.", + "bbox": [ + 215, + 479, + 785, + 601 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Results", + "text_level": 1, + "bbox": [ + 215, + 623, + 323, + 638 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we show the results on the downstream tasks, to evaluate the effectiveness of proposed knowledge-enhanced representation learning. Note that for a fair comparison with the publicly available model released by the original authors, in all experiments, we separately train KEP on the OpenPath [23] dataset, and Quilt1M [24] dataset, and then report results on downstream tasks.", + "bbox": [ + 215, + 654, + 784, + 729 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Retrieval", + "text_level": 1, + "bbox": [ + 215, + 750, + 336, + 765 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we evaluate the performance of retrieval, including cross-modal retrieval on three datasets and disease retrieval on PathPair.", + "bbox": [ + 215, + 775, + 784, + 805 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Cross-modal Retrieval. In Table 1, we demonstrate the results for different models pretrained on OpenPath and Quilt1M. It can be seen that KEP-32", + "bbox": [ + 215, + 809, + 784, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "X. Zhou et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/55d03c691d24283179ddffbab037a0ebbbe7d48f5ff7cabb9bc65aac3327c8b5.jpg", + "table_caption": [ + "Table 1: Performance comparison with PLIP and QuiltNet on three retrieval datasets. All models are pre-trained on the OpenPath and Quilt1M datasets, respectively. i2t and t2i denote image-to-text and text-to-image retrieval, respectively. Bold fonts suggest the best performance." + ], + "table_footnote": [], + "table_body": "
Training datasetTaskModelKnowledge EnhancementArch-PubMedArch-bookPathPair
Recall@10Recall@50Recall@10Recall@50Recall@10Recall@50
OpenPathi2tPLIPX0.0670.1850.1520.3930.0380.119
KEP-32✓0.0980.2830.1640.4040.0710.205
KEP-16✓0.1630.3980.2440.5370.1080.275
t2iPLIPX0.0670.1810.1650.4190.0470.133
KEP-32✓0.0850.2260.1480.3650.0610.171
KEP-16✓0.1380.3390.2380.5330.0930.247
Quilt1Mi2tQuiltNetX0.1390.3260.1880.4070.0650.166
KEP-32✓0.1400.3270.2400.5210.0840.221
KEP-16✓0.1960.4210.2820.5640.1080.254
t2iQuiltNetX0.1220.2930.2040.4290.0710.195
KEP-32✓0.1350.3260.2750.5680.1060.276
KEP-16✓0.1760.4040.3400.6210.1360.326
", + "bbox": [ + 220, + 212, + 777, + 376 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/1ab6a274e0b5b84d29e7b4869098efba207b4071c5b8d92e1834531599673a87.jpg", + "table_caption": [ + "Table 2: Performance comparison of disease retrieval. 12t and i2l denote label-to-text and image-to-label, respectively. Bold fonts suggest the best performance." + ], + "table_footnote": [], + "table_body": "
TaskMetricsOpenPathQuiltNetQuilt1M
PLIPKEP-32KEP-16KEP-32KEP-16
l2tRecall@100.4080.6930.6990.2110.6480.643
Recall@500.5360.8320.8350.3250.8120.808
i2lRecall@100.1140.1620.2230.1350.2260.259
Recall@500.2920.3810.4780.3220.4730.519
", + "bbox": [ + 235, + 439, + 763, + 522 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "pretrained on OpenPath outperforms PLIP on all datasets with respect to the image-to-text retrieval task, especially on Arch-PubMed and PathPair with more than $3\\%$ boost on R10. As for the text-to-image task, KEP-32 also achieves better performance on the dataset of Arch-PubMed and PathPair. Furthermore, our KEP-16 model improves the retrieval performance by a large margin on all datasets across both retrieval tasks. Similar results can be concluded that KEP-32 pretrained on Quilt1M improves the performance on all datasets for both retrieval tasks, which demonstrates the effectiveness of knowledge guidance for visual-language pretraining on cross-modal retrieval.", + "bbox": [ + 215, + 559, + 787, + 695 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Disease Retrieval. Table 2 shows the performance of disease retrieval on the PathPair dataset for different models. It can be seen that although the scale of Quilt1M is 5 times OpenPath, models pretrained on OpenPath often outperforms that on Quilt1M for the label-to-text task. We conjecture this may be caused by the quality of captions in Quilt1M. Our approach KEP-32 outperforms PLIP and Quilt1M by a large margin on both label-to-text and image-to-label tasks, suggesting that the knowledge encoder contributes to paying attention to the key disease information when encoding captions, and thus improves the alignment between images and their disease labels.", + "bbox": [ + 215, + 703, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "KEP", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/106ec365bde8846e5a5a324aad1d54d9c95098cd352118798569d40f671aee2b.jpg", + "image_caption": [ + "Fig. 5: The comparison of zero-shot patch classification between different models. The left and the right subfigures suggest pretraining on OpenPath and Quilt1M, respectively. The visual encoders of KEP-32 and KEP-16 are initialized by CLIP (ViT-B-32) and BiomedCLIP (ViT-B-16), respectively. The number of points for every box is 100, with each representing the performance of one text prompt. The upper, center, and lower line of each box denote the first, median, and third quartile of the distribution." + ], + "image_footnote": [], + "bbox": [ + 220, + 142, + 504, + 277 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c58d535787e4600fd839579c0cb7a46ccc4dc6c0cf507fd8a17ef3ee144ef61f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 142, + 781, + 277 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Zero-shot Patch Classification", + "text_level": 1, + "bbox": [ + 215, + 398, + 509, + 414 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we evaluate the performance of zero-shot classification on patch-level pathology images from 8 datasets. We report the performance distribution (Fig. 5), where each point in the figure denotes the performance of one prompt.", + "bbox": [ + 212, + 424, + 782, + 469 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison to PLIP. In Fig. 5 (left), we demonstrate the performance comparison with PLIP [23], by only pretraining on OpenPath image-text pairs. Compared with PLIP, KEP-32 achieves better zero-shot classification performance on all datasets, with a maximum enhancement of 0.137 on the BACH dataset. Moreover, with the medical-specific initialization of the visual encoder (BiomedCLIP), KEP-16 can further improve the performance significantly in most datasets. Note that, the boxes of KEP variants are generally shorter than that of PLIP, suggesting that our approach is less sensitive to the varying text prompts than PLIP, which can be demonstrated by the visualization of prompt embeddings, shown in Supplementary Materials.", + "bbox": [ + 212, + 472, + 787, + 625 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison to QuiltNet. As shown in Fig. 5 (right), we exhibit the performance comparison between KEP and QuiltNet [24] pretrained on Quilt1M dataset. Similar to results on OpenPath, KEP-32 outperforms Quilt1M in seven out of eight datasets, especially for the BACH, KatherColon, and SkinCancer datasets. KEP-16 can further improve the weighted F1 score. Most boxes of KEP variants are also shorter than that of Quilt1M. The visualization of prompt embeddings, shown in Supplementary Materials, maintains that pathology knowledge can reduce the ambiguities of image-text alignment and thus improve the zero-shot classification performance.", + "bbox": [ + 212, + 628, + 787, + 763 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Zero-shot WSI Tumor Subtyping", + "text_level": 1, + "bbox": [ + 215, + 785, + 535, + 801 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we evaluate the transfer ability of different models for tumor subtyping on common and rare cancers.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "X. Zhou et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c6f606d380643a708d710e074aa0046d5dd33b2b5b04e325468b78fe21baf948.jpg", + "image_caption": [ + "Fig. 6: The performance comparison of tumor subtyping on TCGA-BRCA (common), TCGA-NSCLC, and TCGA-RCC WSIs. The upper, center and lower line of each box denote the first, median, and third quartile of the performance distribution, respectively. The scattered points represent outliers. KEP-16 and KEP-CTP are trained on OpenPath with the visual encoder initialized by BiomedCLIP [56] and CTransPath [48], respectively. MI-Zero-Bio and MI-Zero-Pub are two variants of MI-Zero with different initialization of text encoder. Their visual encoders are initialized by CTransPath. In addition, MI-Zero adopts in-house pathology reports to pretrain their text encoders." + ], + "image_footnote": [], + "bbox": [ + 225, + 142, + 776, + 265 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/bce2f7d5067914490c5216bca130318b0dfa9f9787cfe2fbcb697769f8990fdb.jpg", + "table_caption": [ + "Table 3: Performance comparison of tumor subtyping on six BRCA subtypes, including 2 common and 4 rare cancers. Bold fonts suggest the best performance." + ], + "table_footnote": [], + "table_body": "
ModelTraining datasetTop-1Top-10
CommonRareCommonRare
PLIPOpenPath0.1130.1150.1190.119
QuiltNetQuilt1M0.2780.0940.2790.102
KEP-16OpenPath0.3430.2170.3680.245
MI-Zero-BioIn-house & Web data0.3900.2140.4030.237
MI-Zero-PubIn-house & Web data0.3310.2820.3470.304
KEP-CTPOpenPath0.4430.2820.4320.301
", + "bbox": [ + 269, + 446, + 733, + 545 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Common cancers. Fig. 6 shows the performance comparison of tumor subtyping on common cancers, including TCGA-BRCA (common), TCGA-NSCLC and TCGA-RCC WSIs. MI-Zero-Bio and MI-Zero-Pub are two variants of MI-Zero with different text encoders. The visual encoders of the two MI-Zero variants are initialized by CTransPath [48]. To achieve a fair comparison, we also train a KEP variant named KEP-CTP with the visual encoder initialized by CTransPath. Both KEP-16 and KEP-CTP are trained on the OpenPath dataset. It can be seen that our approach KEP-16 and KEP-CTP outperform PLIP and QuiltNet on all datasets. Note that, the comparisons between variants of MI-Zero and KEP-CTP are actually not fair for our method, as the MI-Zero variants have also been pretrained on massive in-house data – over 550k pathology reports from hospitals [37]. Yet, KEP-CTP still outperforms MI-Zero-Bio across all datasets and achieves comparable performance with MI-Zero-Pub on NSCLC and RCC.", + "bbox": [ + 212, + 577, + 787, + 773 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Rare cancers. Table 3 exhibits the median balanced accuracy of tumor subtyping on common and rare breast cancers. It can be seen that our method KEP-16 pretrained on OpenPath outperforms PLIP and QuiltNet by a large margin on both common and rare breast cancers. Moreover, KEP-CTP significantly outper", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "KEP", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e374f3ff63dcdd0f9422a7e82edc6d0bf7f5ea0bafa1f3b196337c84425f53d8.jpg", + "table_caption": [ + "Table 4: Experimental results of ablation study on model architecture and initialization. Arch-PubMed and PathPair suggest the text-to-image retrieval and the disease retrieval task, respectively. PMB and KB denote PubMedBERT and our pretrained pathology knowledge encoder, respectively. BCLIP denotes BiomedCLIP. Bold fonts and underline suggest the best and the second-best performance, respectively." + ], + "table_footnote": [], + "table_body": "
Visual Init.Text Init.Projection headKnowledge distillMetric lossKatherColon Median (Q1, Q3)Arch-PubMedPathPair
R10R50R1R5
ScratchPMB0.322 (0.272, 0.392)0.0370.1210.0560.113
CLIPPMB0.407 (0.358, 0.452)0.0810.2140.1610.316
CLIPPMB✓0.486 (0.448, 0.517)0.0730.2100.1610.305
BCLIPPMB✓0.553 (0.503, 0.591)0.1270.3470.1870.350
CLIPPMB✓✓(PMB)0.484 (0.435, 0.520)0.0630.1820.0340.089
CLIPKB✓Adasp0.530 (0.477, 0.600)0.0860.2320.2700.482
CLIPKB✓✓(KB)Adasp0.563 (0.505, 0.610)0.0850.2260.4090.618
CLIPKB✓✓(KB)Triplet0.531 (0.482, 0.580)0.0870.2270.2650.493
BCLIPKB✓✓(KB)Adasp0.580 (0.517, 0.631)0.1380.3390.4240.618
", + "bbox": [ + 225, + 226, + 777, + 363 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "forms MI-Zero variants on common breast cancers while achieving comparable performance on the rare ones.", + "bbox": [ + 212, + 387, + 785, + 417 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.4 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 436, + 388, + 452 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this section, we explore the impact of model architecture, initialization, and other hyperparameters, as shown in Table 4 and Table S3 in Supplementary Materials. Comparing the performance between different models, we can draw the following observations: (i) visual projection head can facilitate the zero-shot classification performance; (ii) our pretrained knowledge encoder can enhance the performance across all tasks; (iii) the frozen knowledge branch can further improve the performance of zero-shot classification and disease retrieval tasks while the frozen PubMedBERT can not; (iv) a better initialization of visual encoder contributes to significant performance improvement on all tasks. (v) the Adasp loss is better than the Triplet loss for the knowledge encoding.", + "bbox": [ + 212, + 458, + 787, + 609 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 630, + 359, + 646 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we address the problem of knowledge-enhanced visual-language pretraining on computational pathology. We first curate a pathology knowledge tree that integrates the informative attributes of diseases requiring pathological diagnosis. We then propose a novel knowledge encoding approach based on metric learning to model structured pathological knowledge. With the guidance of the pretrained knowledge encoder, we conduct extensive visual-language pretraining on pathology image-caption pairs. To demonstrate the effectiveness of our approach, we evaluate pretrained VLP models on three downstream tasks, including retrieval, zero-shot classification on patch-level pathology images, and zero-shot tumor subtyping on pathology WSIs. Quantitative experimental results demonstrate that pathology knowledge can significantly improve the performance across different tasks.", + "bbox": [ + 212, + 657, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "X. Zhou et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported by the National Key R&D Program of China (No. 2022ZD0160702) and China Postdoctoral Science Foundation (Certificate Number: 2023M741850).", + "bbox": [ + 212, + 172, + 785, + 219 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 239, + 321, + 255 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aresta, G., Araújo, T., Kwok, S., Chennamsetty, S.S., Safwan, M., Alex, V., Marami, B., Prastawa, M., Chan, M., Donovan, M., et al.: Bach: Grand challenge on breast cancer histology images. Medical Image Analysis 56, 122-139 (2019)", + "2. Bodenreider, O.: The unified medical language system (uml's): integrating biomedical terminology. Nucleic Acids Research 32(suppl_1), D267-D270 (2004)", + "3. Borkowski, A.A., Bui, M.M., Thomas, L.B., Wilson, C.P., DeLand, L.A., Mastorides, S.M.: Lung and colon cancer histopathological image dataset (lc25000). arXiv preprint arXiv:1912.12142 (2019)", + "4. Brummer, O., Pölönen, P., Mustjoki, S., Brück, O.: Integrative analysis of histological textures and lymphocyte infiltration in renal cell carcinoma using deep learning. bioRxiv pp. 2022-08 (2022)", + "5. Campanella, G., Hanna, M.G., Geneslaw, L., Miraflor, A., Werneck Krauss Silva, V., Busam, K.J., Brogi, E., Reuter, V.E., Klimstra, D.S., Fuchs, T.J.: Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature Medicine 25(8), 1301-1309 (2019)", + "6. Chan, T.H., Cendra, F.J., Ma, L., Yin, G., Yu, L.: Histopathology whole slide image analysis with heterogeneous graph representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15661-15670 (2023)", + "7. Chauhan, G., Liao, R., Wells, W., Andreas, J., Wang, X., Berkowitz, S., Horng, S., Szolovits, P., Golland, P.: Joint modeling of chest radiographs and radiology reports for pulmonary edema assessment. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 529-539. Springer (2020)", + "8. Chen, C., Lu, M.Y., Williamson, D.F., Chen, T.Y., Schaumberg, A.J., Mahmood, F.: Fast and scalable search of whole-slide images via self-supervised deep learning. Nature Biomedical Engineering 6(12), 1420-1434 (2022)", + "9. Chen, R.J., Chen, C., Li, Y., Chen, T.Y., Trister, A.D., Krishnan, R.G., Mahmood, F.: Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16144-16155 (2022)", + "0. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: Uniter: Learning universal image-text representations (2019)", + "1. Chen, Y.C., Lu, C.S.: Rankmix: Data augmentation for weakly supervised learning of classifying whole slide images with diverse sizes and imbalanced categories. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23936-23945 (2023)", + "2. Chen, Z., Li, G., Wan, X.: Align, reason and learn: Enhancing medical vision-and-language pre-training with knowledge. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 5152-5161 (2022)", + "3. Chen, Z., Song, Y., Chang, T.H., Wan, X.: Generating radiology reports via memory-driven transformer. In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). pp. 1439-1449 (2020)" + ], + "bbox": [ + 225, + 266, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "KEP", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Cui, H., Xu, Y., Li, W., Wang, L., Duh, H.: Collaborative learning of cross-channel clinical attention for radiotherapy-related esophageal fistula prediction from ct. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 212-220. Springer (2020)", + "15. Gamper, J., Rajpoot, N.: Multiple instance captioning: Learning representations from histopathology textbooks and articles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16549-16559 (2021)", + "16. Gu, Y., Tinn, R., Cheng, H., Lucas, M., Usuyama, N., Liu, X., Naumann, T., Gao, J., Poon, H.: Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Computing for Healthcare (HEALTH) 3(1), 1-23 (2021)", + "17. Han, C., Pan, X., Yan, L., Lin, H., Li, B., Yao, S., Lv, S., Shi, Z., Mai, J., Lin, J., et al.: Wsss4luad: Grand challenge on weakly-supervised tissue semantic segmentation for lung adenocarcinoma. arXiv preprint arXiv:2204.06455 (2022)", + "18. Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737 (2017)", + "19. Hu, S., Chen, L., Wu, P., Li, H., Yan, J., Tao, D.: St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In: European Conference on Computer Vision. pp. 533-549. Springer (2022)", + "20. Huang, S.C., Shen, L., Lungren, M.P., Yeung, S.: Gloria: A multimodal global-local representation learning framework for label-efficient medical image recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3942-3951 (2021), official Implementation: https://github.com/marshuang80/gloria", + "21. Huang, X., Fang, Y., Lu, M., Yan, F., Yang, J., Xu, Y.: Dual-ray net: automatic diagnosis of thoracic diseases using frontal and lateral chest x-rays. Journal of Medical Imaging and Health Informatics 10(2), 348-355 (2020)", + "22. Huang, Y., Zhao, W., Wang, S., Fu, Y., Jiang, Y., Yu, L.: Conslide: Asynchronous hierarchical interaction transformer with breakup-reorganize rehearsal for continual whole slide image analysis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 21349-21360 (2023)", + "23. Huang, Z., Bianchi, F., Yuksekgonul, M., Montine, T.J., Zou, J.: A visual-language foundation model for pathology image analysis using medical twitter. Nature medicine 29(9), 2307-2316 (2023)", + "24. Ikezogwo, W., Seyfioglu, S., Ghezloo, F., Geva, D., Sheikh Mohammed, F., Anand, P.K., Krishna, R., Shapiro, L.: Quilt-1m: One million image-text pairs for histopathology. Advances in Neural Information Processing Systems 36 (2024)", + "25. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning. pp. 4904-4916. PMLR (2021)", + "26. Kang, M., Song, H., Park, S., Yoo, D., Pereira, S.: Benchmarking self-supervised learning on diverse pathology datasets. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3344-3354 (2023)", + "27. Kather, J.N., Halama, N., Marx, A.: 100,000 histological images of human colorectal cancer and healthy tissue. Zenodo10 5281 (2018)", + "28. Kather, J.N., Krisam, J., Charoentong, P., Luedde, T., Herpel, E., Weis, C.A., Gaiser, T., Marx, A., Valous, N.A., Ferber, D., et al.: Predicting survival from colorectal cancer histology slides using deep learning: A retrospective multicenter study. PLoS Medicine 16(1), e1002730 (2019)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "X. Zhou et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Kim, W., Son, B., Kim, I.: Vilt: Vision-and-language transformer without convolution or region supervision. In: International Conference on Machine Learning. pp. 5583-5594. PMLR (2021)", + "30. Kriegsmann, K., Lobers, F., Zgorzelski, C., Kriegsmann, J., Janssen, C., Meliss, R.R., Muley, T., Sack, U., Steinbuss, G., Kriegsmann, M.: Deep learning for the detection of anatomical tissue structures and neoplasms of the skin on scanned histopathological tissue sections. Frontiers in Oncology 12, 1022967 (2022)", + "31. Kundra, R., Zhang, H., Sheridan, R., Sirlintrapun, S.J., Wang, A., Ochoa, A., Wilson, M., Gross, B., Sun, Y., Madupuri, R., et al.: Oncotree: a cancer classification system for precision oncology. JCO clinical cancer informatics 5, 221-230 (2021)", + "32. Li, H., Zhu, C., Zhang, Y., Sun, Y., Shui, Z., Kuang, W., Zheng, S., Yang, L.: Task-specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7454-7463 (2023)", + "33. Li, M., Cai, W., Verspoor, K., Pan, S., Liang, X., Chang, X.: Cross-modal clinical graph transformer for ophthalmic report generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20656-20665 (2022)", + "34. Lin, T., Yu, Z., Hu, H., Xu, Y., Chen, C.W.: Interventional bag multi-instance learning on whole-slide pathological images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19830-19839 (2023)", + "35. Lin, W., Zhao, Z., Zhang, X., Wu, C., Zhang, Y., Wang, Y., Xie, W.: Pmc-clip: Contrastive language-image pre-training using biomedical documents. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) (2023)", + "36. Lu, M.Y., Chen, B., Williamson, D.F., Chen, R.J., Liang, I., Ding, T., Jaume, G., Odintsov, I., Zhang, A., Le, L.P., et al.: Towards a visual-language foundation model for computational pathology. arXiv preprint arXiv:2307.12914 (2023)", + "37. Lu, M.Y., Chen, B., Zhang, A., Williamson, D.F., Chen, R.J., Ding, T., Le, L.P., Chuang, Y.S., Mahmood, F.: Visual language pretrained multiple instance zero-shot transfer for histopathology images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19764-19775 (2023)", + "38. Lu, M.Y., Williamson, D.F., Chen, T.Y., Chen, R.J., Barbieri, M., Mahmood, F.: Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering 5(6), 555-570 (2021)", + "39. Müller, P., Kaissis, G., Zou, C., Rueckert, D.: Joint learning of localized representations from medical images and reports. In: European Conference on Computer Vision. pp. 685-701. Springer (2022)", + "40. Neumann, M., King, D., Beltagy, I., Ammar, W.: ScispaCy: Fast and robust models for biomedical natural language processing. In: Demner-Fushman, D., Cohen, K.B., Ananiadou, S., Tsujii, J. (eds.) Proceedings of the 18th BioNLP Workshop and Shared Task. pp. 319-327. Association for Computational Linguistics, Florence, Italy (Aug 2019). https://doi.org/10.18653/v1/W19-5034, https://aclanthology.org/W19-5034", + "41. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)", + "42. Roberts, R.J.: Pubmed central: The genbank of the published literature (2001)", + "43. Rorke, L.B.: Pathologic diagnosis as the gold standard (1997)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "KEP", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "44. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278–25294 (2022)", + "45. Shaban, M., Awan, R., Fraz, M.M., Azam, A., Tsang, Y.W., Snead, D., Rajpoot, N.M.: Context-aware convolutional neural network for grading of colorectal cancer histology images. IEEE Transactions on Medical Imaging 39(7), 2395-2405 (2020)", + "46. Shao, Z., Bian, H., Chen, Y., Wang, Y., Zhang, J., Ji, X., et al.: Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in Neural Information Processing Systems 34, 2136-2147 (2021)", + "47. Silva-Rodriguez, J., Colomer, A., Sales, M.A., Molina, R., Naranjo, V.: Going deeper through the gleason scoring scale: An automatic end-to-end system for histology prostate grading and cribriform pattern detection. Computer Methods and Programs in Biomedicine 195, 105637 (2020)", + "48. Wang, X., Yang, S., Zhang, J., Wang, M., Zhang, J., Yang, W., Huang, J., Han, X.: Transformer-based unsupervised contrastive learning for histopathological image classification. Medical Image Analysis 81, 102559 (2022)", + "49. Wu, C., Zhang, X., Zhang, Y., Wang, Y., Xie, W.: Medklip: Medical knowledge enhanced language-image pre-training. medRxiv pp. 2023-01 (2023)", + "50. Xie, X., Niu, J., Liu, X., Chen, Z., Tang, S., Yu, S.: A survey on incorporating domain knowledge into deep learning for medical image analysis. Medical Image Analysis 69, 101985 (2021)", + "51. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettle-moyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 6787-6800 (2021)", + "52. Xuan, H., Stylianou, A., Liu, X., Pless, R.: Hard negative examples are hard, but useful. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 126-142. Springer (2020)", + "53. Yang, S., Wu, X., Ge, S., Zhou, S.K., Xiao, L.: Knowledge matters: Chest radiology report generation with general and specific knowledge. Medical Image Analysis 80, 102510 (2022)", + "54. Yu, F., Tang, J., Yin, W., Sun, Y., Tian, H., Wu, H., Wang, H.: Ernie-vil: Knowledge enhanced vision-language representations through scene graphs. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 3208-3216 (2021)", + "55. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 (2022)", + "56. Zhang, S., Xu, Y., Usuyama, N., Bagga, J., Tinn, R., Preston, S., Rao, R., Wei, M., Valluri, N., Wong, C., et al.: Large-scale domain-specific pretraining for biomedical vision-language processing. arXiv preprint arXiv:2303.00915 (2023)", + "57. Zhang, X., Wu, C., Zhang, Y., Xie, W., Wang, Y.: Knowledge-enhanced visual-language pre-training on chest radiology images. Nature Communications 14(1), 4542 (2023)", + "58. Zhang, Y., Jiang, H., Miura, Y., Manning, C.D., Langlotz, C.P.: Contrastive learning of medical visual representations from paired images and text. In: Machine Learning for Healthcare (2022), highest Starred Implementation: https://github.com/edreisMD/ConVIRT-pytorch", + "59. Zhou, X., Cheng, Z., Gu, M., Chang, F.: Lirnet: Local integral regression network for both strongly and weakly supervised nuclei detection. In: 2020 IEEE Interna" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "X. Zhou et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tional Conference on Bioinformatics and Biomedicine (BIBM). pp. 945-951. IEEE (2020)", + "60. Zhou, X., Zhong, Y., Cheng, Z., Liang, F., Ma, L.: Adaptive sparse pairwise loss for object re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19691-19701 (2023)" + ], + "bbox": [ + 217, + 146, + 785, + 217 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "KEP", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Knowledge-enhanced Visual-Language Pretraining for Computational Pathology/a04fca37-34eb-47aa-8e9d-2ee68fcc1f47_model.json b/2024/Knowledge-enhanced Visual-Language Pretraining for Computational Pathology/a04fca37-34eb-47aa-8e9d-2ee68fcc1f47_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e220b03c1955562d95f7353108ce70a31928e634 --- /dev/null +++ b/2024/Knowledge-enhanced Visual-Language Pretraining for Computational Pathology/a04fca37-34eb-47aa-8e9d-2ee68fcc1f47_model.json @@ -0,0 +1,2460 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.266, + 0.141, + 0.737, + 0.187 + ], + "angle": 0, + "content": "Knowledge-enhanced Visual-Language Pretraining for Computational Pathology" + }, + { + "type": "text", + "bbox": [ + 0.322, + 0.212, + 0.678, + 0.247 + ], + "angle": 0, + "content": "Xiao Zhou1, Xiaoman Zhang1,2, Chaoyi Wu1,2, Ya Zhang1,2, Weidi Xie1,2, Yanfeng Wang1,2" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.257, + 0.652, + 0.271 + ], + "angle": 0, + "content": "\\(^{1}\\) Shanghai Artificial Intelligence Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.392, + 0.272, + 0.612, + 0.286 + ], + "angle": 0, + "content": "2 Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.287, + 0.639, + 0.299 + ], + "angle": 0, + "content": "https://github.com/MAGIC-AI4Med/KEP" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.301, + 0.743, + 0.314 + ], + "angle": 0, + "content": "zhouxiao@pjlab.org.cn, {xm99sjtu, wtzxxxxwy02, ya_zhang, weidi," + }, + { + "type": "text", + "bbox": [ + 0.405, + 0.315, + 0.599, + 0.327 + ], + "angle": 0, + "content": "wangyanfeng} @sjtu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.366, + 0.747, + 0.588 + ], + "angle": 0, + "content": "Abstract. In this paper, we consider the problem of visual representation learning for computational pathology, by exploiting large-scale image-text pairs gathered from public resources, along with the domain-specific knowledge in pathology. Specifically, we make the following contributions: (i) We curate a pathology knowledge tree that consists of 50,470 informative attributes for 4,718 diseases requiring pathology diagnosis from 32 human tissues. To our knowledge, this is the first comprehensive structured pathology knowledge base; (ii) We develop a knowledge-enhanced visual-language pretraining approach, where we first project pathology-specific knowledge into latent embedding space via a language model, and use it to guide the visual representation learning; (iii) We conduct thorough experiments to validate the effectiveness of our proposed components, demonstrating significant performance improvement on various downstream tasks, including cross-modal retrieval, zero-shot classification on pathology patches, and zero-shot tumor subtyping on whole slide images (WSIs)." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.601, + 0.692, + 0.616 + ], + "angle": 0, + "content": "Keywords: Pathology knowledge \\(\\cdot\\) Visual-language pretraining" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.642, + 0.377, + 0.658 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Pathology diagnosis is currently the golden standard for examining various diseases in clinical applications, especially in the diagnosis of neoplasm [43]. In the last decade, the prosperity of deep learning on computer vision has led to the rapid development of computational pathology, for example, approaches with supervised learning [6,22,34,45,46], and weakly supervised learning [5,11,32,38,59]. Despite being promising, these approaches have been fundamentally limited by the scale of costly labels. Alternatively, self-supervised pretraining on numerous unlabeled pathological images [8,9,26,48] has attracted unprecedented attention, yet it still requires supervised fine-tuning for downstream deployments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In the recent literature [25,41], studies on the multimodal foundation model have demonstrated conspicuous improvement in downstream zero-shot tasks, by" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.367, + 0.127 + ], + "angle": 0, + "content": "X. Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.145, + 0.68, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.318, + 0.788, + 0.416 + ], + "angle": 0, + "content": "Fig. 1: Knowledge-enhanced pathology image-text alignment. The short caption of a pathology image crawled from public websites is typically unstructured and with varying granularities, which introduces noticeable ambiguities for image-text alignment. While the implicit structures and correlations between different image-caption pairs could be constructed by explicit disease attributes (dashed boxes), which can be well-aligned by a pathology knowledge tree. LUSC and LUAD suggest lung squamous cell carcinoma, and lung adenocarcinoma, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.613 + ], + "angle": 0, + "content": "simply training to align visual and language embedding space on massive image-text pairs crawled from public websites. In contrast to computer vision [19,25,41], representation learning in pathology requires tremendous expertise and domain knowledge. An ideal training corpus would thus consist of well-structured medical reports from the hospital, however, it is often extremely difficult to acquire, due to privacy concerns. Alternatively, recent works [23, 24, 36, 37] propose to gather large-scale image-caption pairs from public resources (PubMed [42] papers, Twitter, Youtube videos). Compared to discrete image labels, image captions from medical reports and academic articles can potentially provide more valuable medical information without manual annotation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.788, + 0.758 + ], + "angle": 0, + "content": "For existing works that adopt simple contrastive learning on pathology image and caption pairs, they suffer from the following challenges: First, the short captions from these web-crawled image-text pairs are typically noisy, unstructured, and lack domain knowledge (Fig. 1), which can be sub-optimal for constructing high-quality pathology-specific visual representations; Second, the varying granularities of the free text from captions will inevitably introduce ambiguities when aligning with images, thus causing the model to be highly sensitive to the used text prompts during the inference stage [36]. Thus, effectively harnessing the potential of such web-crawled data remains to be a challenge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, to tackle the above challenges, we anticipate that introducing pathology knowledge is of great significance to make up for the deficiency of short image captions. To this end, we make the following contributions: (i) We curate a pathology knowledge tree, PathKT, by collecting 50,470 informative pathological attributes of 4,718 diseases in 32 tissues from publicly available educational" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.696, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KEP" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.358 + ], + "angle": 0, + "content": "resources and OncoTree\\(^{3}\\) [31]. (ii) We propose a novel knowledge encoder pretraining approach, that projects the attributes of each disease from PathKT into latent embedding space, where the attributes of the same disease, including disease synonyms, definitions, histology, and cytology features, share similar representations. (iii) We develop a knowledge-enhanced pretraining (KEP) approach to align pathology visual-language representations, which freezes the knowledge encoder and continuously injects domain-specific knowledge into the image-text embedding space. To demonstrate the effectiveness of our proposed approach, we conduct thorough experiments on three downstream tasks, including retrieval on three pathological image-caption datasets, zero-shot patch classification on eight patch-level pathology image datasets, and zero-shot WSI tumor subtyping on three datasets from The Cancer Genome Atlas (TCGA)\\(^{4}\\). Quantitative experiments suggest that knowledge guidance can significantly enhance the performance across different tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.383, + 0.383, + 0.399 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.788, + 0.568 + ], + "angle": 0, + "content": "Vision-Language Pretraining (VLP). Current vision-language pretraining approaches are typically categorized into two groups. The first group, referred to as two-stream approaches [41, 51], involves using two separate encoders to extract features for visual and textual data, respectively. The second group, known as single-stream methods [10, 29], utilizes a cross-modal fusion encoder to enhance interactions between vision and text features. In the field of medical VLP, existing methods mainly adopt the two-stream approach [7, 12, 20, 35, 39, 56, 58]. Despite the valuable contributions of these methods, the reliance on data-driven representation learning restricts the utilization of systematic and structured medical knowledge, leading to less than professional diagnosis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.571, + 0.788, + 0.752 + ], + "angle": 0, + "content": "VLP in Computational Pathology. In the recent literature, Huang et al. proposed a pathology VLP model named PLIP [23] and released OpenPath that contains 200K image-caption pairs from the social media of Twitter and other public resources. To extend the pathology data scale, Ikezogwo et al. curate Quilt1m [24] with over one million histopathology image-text pairs by capturing keyframes and speech from YouTube videos. In addition to evaluating the zero-shot classification on patch-level pathology images, Lu et al. propose MI-Zero [37] and CONCH [36] to extend the transfer ability of pathology VLP models on gigapixel whole slide images (WSIs). In contrast to these existing work, that finetunes a CLIP [41] or CoCa [55] on pathology image-caption pairs, we propose to first train a pathology-specific knowledge encoder, and use it to guide visual-language representation learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.756, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Medical Knowledge-enhanced Learning. In the medical community, leveraging external medical knowledge to enhance deep learning models has become an increasingly important topic [49, 50, 57]. Generally, existing approaches can" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.81, + 0.408, + 0.826 + ], + "angle": 0, + "content": "3 https://oncotree.info/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.223, + 0.826, + 0.469, + 0.84 + ], + "angle": 0, + "content": "4 https://portal.gdc.cancer.gov/" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.81, + 0.469, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.367, + 0.127 + ], + "angle": 0, + "content": "X. Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.145, + 0.785, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.296, + 0.788, + 0.409 + ], + "angle": 0, + "content": "Fig. 2: The construction of pathology knowledge tree. OncoTree is adopted as the base architecture to construct the PathKT. The tissue types, disease entities, and attributes are first extracted from web-crawled pathological descriptions, where cancers are then matched to OncoTree based on their tissue types and tumor types/subtypes using UMLS CUIs. Moreover, non-tumor diseases are added to the knowledge tree according to their tissue types. Finally, the pathology knowledge tree integrates 4718 diseases from 32 tissues, with each disease containing various synonyms, definitions, and histological and cytological features. CSCLC in this figure suggests combined small cell lung cancer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.788, + 0.514 + ], + "angle": 0, + "content": "be categorized based on the ways of using medical knowledge: model-based approaches [14, 21] adopts the prior knowledge of radiology or diagnosis summarized by doctors to design algorithms; and input-based methods [13, 33, 53, 54] directly exploit knowledge as the external input to guide representation training. However, most of these works are focused on the analysis of chest X-rays." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.534, + 0.341, + 0.55 + ], + "angle": 0, + "content": "3 Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.687 + ], + "angle": 0, + "content": "Our primary goal is to leverage structured pathology knowledge to enhance visual-language representation learning. To start with, we construct a **Pathology Knowledge Tree**, termed as **PathKT**, that consists of 50,470 informative attributes of 4718 diseases from 32 human tissues (Sec. 3.1). We then train a knowledge encoder that projects the structured pathology knowledge into an embedding space (Sec. 3.2). We further employ the knowledge encoder to guide visual-language pretraining for computational pathology, termed as KEP (Knowledge-Enhanced Pre-training, Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.707, + 0.444, + 0.722 + ], + "angle": 0, + "content": "3.1 PathKT Construction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.731, + 0.787, + 0.762 + ], + "angle": 0, + "content": "Here, we detail the procedure for building up a pathology knowledge tree with various online sources, which will be further used for knowledge encoding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Knowledge Source. We collect pathology-specific knowledge from publicly available educational resources, such as text books, professional websites, and structured databases (OncoTree [31]). Specifically, we extract pathological descriptions of 884 tumor subtypes from OncoTree and 4360 diseases from text books and professional websites, including all domain knowledge required for" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.696, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KEP" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "diagnosis in clinical scenarios, i.e., disease name/ synonyms, definitions, histological features, and cytological features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.469 + ], + "angle": 0, + "content": "Knowledge Tree Construction. We structure these knowledge sources into a knowledge tree by expanding the OncoTree, as shown in Fig. 2. OncoTree is a tree-structure cancer classification system [31], which consists of 884 tumor subtypes from 32 tissues, with each type of tumor linked to a Concept Unique Identifier (CUI) from Unified Medical Language System (UMLS) [2]. Specifically, we first extract tissue types, disease entities, and disease attributes from pathological descriptions of 4360 diseases, in which 168 cancers are found overlapped with OncoTree by using SciSpacy package [40] to link the UMLS CUIs. The rest 4192 diseases are then added to the knowledge tree according to their tissue types. After deduplication and noise reduction (358 diseases without any informative descriptions are deleted), all diseases are organized into the corresponding tissues. The final histopathology knowledge tree contains 4718 diseases from 32 tissues. The attributes of each disease node are constructed by a varying number of synonyms, definitions, and histological and cytological descriptions. The statistics of PathKT is shown in Table S1 and Fig. S1 in Supplementary Materials. In the final PathKT, for instance, the tissue node of the lung connects all lung diseases, including combined small cell lung cancer (CSCLC), lung squamous cell carcinoma (LUSC), and lung adenocarcinoma (LUAD), etc. The LUSC node connects four kinds of attributes, shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.494, + 0.525, + 0.51 + ], + "angle": 0, + "content": "3.2 Pathology Knowledge Encoding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.689 + ], + "angle": 0, + "content": "In this section, we describe details for projecting tree-structure pathological knowledge into a latent embedding space, by training a knowledge encoder. Specifically, we align different disease entities with their corresponding free-text attributes via metric learning, such that, the synonyms, definitions, and corresponding histological/cytological features are close in the embedding space. As a consequence, the model can link the pathology images to their implicit disease labels during the visual-language pertaining, since the free-form text descriptions in image-text pairs may contain disease attributes, such as pathological features and disease definitions, which have already been aligned with disease names/synonyms. Therefore, the alignment of pathology knowledge can help the model link to the disease entities for better performance during diagnosis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.692, + 0.788, + 0.769 + ], + "angle": 0, + "content": "Problem Setting. Given a set of disease entities with their corresponding attributes, \\(\\mathcal{D} = \\{(d_1,\\mathbf{a}_1),\\ldots ,(d_n,\\mathbf{a}_n)\\}\\), where \\(d_{i}\\) denotes the \\(i\\)-th disease entity, and \\(\\mathbf{a}_i = \\{a_i^1,\\dots ,a_i^k\\}\\) refer to the associated \\(k\\) attributes, both disease and attributes are represented in the format of natural language. Note that, for different disease entities, \\(k\\) also varies, our goal here is to train a model that satisfies:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.781, + 0.786, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s i m} \\left(\\Phi_ {\\mathrm {k}} \\left(a _ {i} ^ {p}\\right), \\Phi_ {\\mathrm {k}} \\left(a _ {i} ^ {q}\\right)\\right) \\gg \\operatorname {s i m} \\left(\\Phi_ {\\mathrm {k}} \\left(a _ {i} ^ {p}\\right), \\Phi_ {\\mathrm {k}} \\left(a _ {j} ^ {t}\\right)\\right), i \\neq j, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.809, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\(\\varPhi_{\\mathbf{k}}(\\cdot)\\) denotes the knowledge encoder, \\(\\mathrm{sim}(\\cdot ,\\cdot)\\) refers to the similarity, \\(a_i^p,a_i^q\\) and \\(a_j^t\\) refer to the randomly sampled attributes from the \\(i,j\\)-th disease entity." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.367, + 0.127 + ], + "angle": 0, + "content": "X. Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.145, + 0.778, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.296, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Fig. 3: Knowledge encoder pretraining based on metric learning. \\( n \\) disease entities and each with \\( k \\) attributes, including disease synonyms, definitions, cytology and pathology features, construct a mini-batch (left part of the figure), which are fed to a knowledge encoder for pretraining. In the embedding space (right part of the figure), the markers in different shapes represent the embeddings of attributes of different diseases. \\( S_{i}^{+} \\)suggests the max-min positive attribute similarity within the \\( i \\)-th disease, while \\( S_{i}^{-} \\)denotes the maximal attribute similarity between the \\( i \\)-th disease and other diseases. The goal of metric learning is to increase \\( S_{i}^{+} \\)and meanwhile decreasing \\( S_{i}^{-} \\). The purple-dashed arrow and circle denote the minimal positive attribute similarity in the second class and the hypersphere it spans." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.785, + 0.497 + ], + "angle": 0, + "content": "Intuitively, the knowledge encoder enables the attributes of the same disease to be pulled together, while attributes from different diseases are pushed apart." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.788, + 0.607 + ], + "angle": 0, + "content": "Training. In order to achieve the objective defined in Eq. 1, metric learning is exploited to construct an embedding space where the representations of intra/inter-class instances are clustered/separated. In specific, given a mini-batch that contains \\( n \\) random diseases and each with \\( k \\) attributes, we denote the normalized embedding for the \\( p \\)-th attribute of \\( i \\)-th disease as: \\( \\mathbf{z}_p^i = \\varPhi_{\\mathrm{k}}(a_i^p) / \\| \\varPhi_{\\mathrm{k}}(a_i^p)\\| \\). We adopt the recently proposed AdaSP loss [60], which finds out a max-min positive similarity (Fig. 3) and then shapes a loss with the maximal negative similarity:" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.617, + 0.787, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m e t i c}} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\log \\left(1 + e ^ {\\left(S _ {i} ^ {-} - S _ {i} ^ {+}\\right) / \\tau}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.663, + 0.788, + 0.695 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a temperature parameter. \\(S_{i}^{+}\\) and \\(S_{i}^{-}\\) denote the max-min positive and the maximal negative similarity, which can be computed by the soft version:" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.705, + 0.787, + 0.758 + ], + "angle": 0, + "content": "\\[\nS _ {i} ^ {+} = \\max _ {p} \\min _ {q} \\left\\langle \\mathbf {z} _ {p} ^ {i}, \\mathbf {z} _ {q} ^ {i} \\right\\rangle \\approx \\tau \\log \\left(\\sum_ {p = 1} ^ {k} \\frac {1}{\\sum_ {q = 1} ^ {k} e ^ {- \\frac {\\left\\langle \\mathbf {z} _ {p} ^ {i} , \\mathbf {z} _ {q} ^ {i} \\right\\rangle}{\\tau}}}\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.762, + 0.787, + 0.807 + ], + "angle": 0, + "content": "\\[\nS _ {i} ^ {-} = \\max _ {j, p, q} \\left\\langle \\mathbf {z} _ {p} ^ {i}, \\mathbf {z} _ {q} ^ {j} \\right\\rangle \\approx \\tau \\log \\left(\\sum_ {p = 1} ^ {k} \\sum_ {j = 1, j \\neq i} ^ {n} \\sum_ {q = 1} ^ {k} e ^ {\\frac {\\left\\langle \\mathbf {z} _ {p} ^ {i} , \\mathbf {z} _ {q} ^ {j} \\right\\rangle}{\\tau}}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "where \\(\\langle \\cdot \\rangle\\) represents the cosine similarity, the details about the soft version can be found in Supplementary Materials." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.696, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KEP" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "Discussion. The conventional triplet loss [18] with batch hard mining strategy is widely adopted in metric learning, while it is not suitable in our case due to the following two reasons: First, in the pathology knowledge, each disease entity is associated with at most four types of attributes, the text descriptions of different attribute types might reveal significant divergence, which causes a large intra-class variation, marked by the purple dashed circle in Fig. 3, during training. Second, the fine-grained diseases might share high similarities, such as tumor subtypes: breast invasive ductal carcinoma and breast invasive lobular carcinoma, which cause low inter-class variations during training. When a mini-batch meets these two conditions, traditional triplet loss that enforces the instance similarities could typically produce the bad local minima of optimization [52], which, therefore, undermines the metric learning." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.348, + 0.629, + 0.365 + ], + "angle": 0, + "content": "3.3 Pathology Knowledge Enhanced Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.786, + 0.418 + ], + "angle": 0, + "content": "In this section, we present a simple yet effective pretraining approach, termed KEP, that leverages the established knowledge encoder to guide visual-language pretraining for computational pathology." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.423, + 0.787, + 0.469 + ], + "angle": 0, + "content": "Visual-Language Pretraining. Given paired image and captions, denoted as \\(\\mathcal{F} = \\{(x_1, c_1), \\ldots, (x_n, c_n)\\}\\), our goal is to construct a visual-language embedding space from paired image-text data, that satisfies:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.478, + 0.785, + 0.494 + ], + "angle": 0, + "content": "\\[\n\\left. \\operatorname {s i m} \\left(\\Phi_ {\\mathrm {v}} \\left(x _ {i}\\right), \\Phi_ {\\mathrm {t}} \\left(c _ {i}\\right)\\right) \\gg \\operatorname {s i m} \\left(\\Phi_ {\\mathrm {v}} \\left(x _ {i}\\right), \\Phi_ {\\mathrm {t}} \\left(c _ {j}\\right)\\right), i \\neq j, \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.788, + 0.61 + ], + "angle": 0, + "content": "where \\(\\varPhi_{\\mathrm{v}}(\\cdot)\\) and \\(\\varPhi_{\\mathrm{t}}(\\cdot)\\) denote the visual and the textual encoder, respectively. In this work, ViT-B-32/16 is adopted as the backbone for the visual encoder and initialized from the visual weights of CLIP [41] / BiomedCLIP [56]. A projection head is added on top of the visual encoder to bridge the gap caused by the non-pathology initialization, shown in Fig. 4. The text encoder is initialized by the pretrained knowledge encoder, as described in Sec. 3.2. In order to learn effective visual-language representation, we optimise an infoNCE contrastive objective:" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.615, + 0.785, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {v t}} = - \\left(\\log \\frac {e ^ {\\mathbf {v} _ {i} ^ {T} \\mathbf {t} _ {i} / \\tau}}{\\sum_ {j = 1} ^ {n} e ^ {\\mathbf {v} _ {i} ^ {T} \\mathbf {t} _ {j} / \\tau}} + \\log \\frac {e ^ {\\mathbf {t} _ {i} ^ {T} \\mathbf {v} _ {i} / \\tau}}{\\sum_ {j = 1} ^ {n} e ^ {\\mathbf {t} _ {i} ^ {T} \\mathbf {v} _ {j} / \\tau}}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.658, + 0.788, + 0.691 + ], + "angle": 0, + "content": "where \\(\\mathbf{v}_i = \\varPhi_{\\mathrm{v}}(x_i)\\), and \\(\\mathbf{t}_i = \\varPhi_{\\mathrm{t}}(c_i)\\), refer to the normalized embedding vectors from visual and text encoders, respectively. \\(\\tau\\) denotes a temperature parameter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.788, + 0.801 + ], + "angle": 0, + "content": "Knowledge Distillation. To keep the alignment between images and free-form captions inside the knowledge space and thus the images can be linked to their implicit disease entities (right part in Fig. 4), we adopt an additional frozen branch to continuously distill pathology knowledge to the text encoder. Specifically, we use the text-knowledge embedding pairs to construct a contrastive loss item \\(\\mathcal{L}_{\\mathrm{tk}}\\), which can be computed by Eq. 6 with the normalized visual embedding vectors replaced by the knowledge embedding vectors:" + }, + { + "type": "equation", + "bbox": [ + 0.343, + 0.808, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {t k}} = - \\left(\\log \\frac {e ^ {\\mathbf {k} _ {i} ^ {T} \\mathbf {t} _ {i} / \\tau}}{\\sum_ {j = 1} ^ {n} e ^ {\\mathbf {k} _ {i} ^ {T} \\mathbf {t} _ {j} / \\tau}} + \\log \\frac {e ^ {\\mathbf {t} _ {i} ^ {T} \\mathbf {k} _ {i} / \\tau}}{\\sum_ {j = 1} ^ {n} e ^ {\\mathbf {t} _ {i} ^ {T} \\mathbf {k} _ {j} / \\tau}}\\right), \\tag {7}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "X. Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.146, + 0.777, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.258, + 0.789, + 0.357 + ], + "angle": 0, + "content": "Fig. 4: Model architecture (left graph). A projection head is added on the top of the visual encoder to bridge the gap between the image and the text encoder. The knowledge encoder is frozen across the whole training stage to distill pathology knowledge to the learnable text encoder. As a result, the pathology images can be aligned with their implicit disease labels (marked by dashed boxes in the right graph) during visual-language pretraining, since the captions contain disease attributes that have been already aligned with disease names/synonyms in the knowledge embedding space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.789, + 0.416 + ], + "angle": 0, + "content": "where \\(\\mathbf{k}_i = \\varPhi_{\\mathrm{k}}(c_i)\\), refer to the normalized embedding vectors from the knowledge encoder. The overall training loss can thus be computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.444, + 0.425, + 0.785, + 0.44 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {v t}} + \\alpha \\mathcal {L} _ {\\mathrm {t k}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.449, + 0.789, + 0.511 + ], + "angle": 0, + "content": "where \\(\\alpha\\) denotes a weight parameter. It is worth emphasizing that the key contribution in KEP is to initialize the text encoder with the pre-obtained pathology knowledge encoder and adopt it to continuously distill pathology knowledge to the text encoder, while for the other parts, we keep the same as PLIP [23]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.532, + 0.377, + 0.55 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.789, + 0.595 + ], + "angle": 0, + "content": "In this section, we first introduce the datasets used for training and evaluation in this paper, followed by the evaluation metrics and implementation details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.616, + 0.411, + 0.631 + ], + "angle": 0, + "content": "4.1 Training Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.789, + 0.686 + ], + "angle": 0, + "content": "Dataset for Knowledge Encoding. We carry out pathology knowledge encoding with PathKT, where 4,718 disease nodes with a total number of 50,470 attributes are extracted, shown in Table S1 in Supplementary Materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.79, + 0.841 + ], + "angle": 0, + "content": "Datasets for KEP Pretraining. Our dataset is composed of two parts. First, we collect the OpenPath data provided by PLIP [23] and obtain 138,874 image-text pairs after denoising and pre-processing. Second, we collect the Quilt1M [24] dataset, which gathers pathology image-text pairs from four public sources: PubMed articles, LAION [44], OpenPath [23], and Youtube videos. Since one of the downstream evaluation datasets Arch-PubMed [15] contains pathology images from PubMed articles, we remove this data source from Quilt1M to avoid data leaking. Considering that many images in Quilt1M are matched with multiple captions, we concatenate the captions related to the same image and finally obtain 576,608 image-text pairs for the Quilt1M dataset." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.696, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KEP" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.418, + 0.16 + ], + "angle": 0, + "content": "4.2 Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.784, + 0.223 + ], + "angle": 0, + "content": "We evaluate the pretrained models on three tasks, namely, retrieval, zero-shot patch classification, and zero-shot WSI tumor subtyping. The details for all evaluation datasets are exhibited in Table S2 in Supplementary Materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.227, + 0.785, + 0.378 + ], + "angle": 0, + "content": "Retrieval. This task involves cross-modal retrieval and disease retrieval. Cross-modal retrieval aims to retrieve the correct caption for a given image and vice versa. Disease retrieval, on the other hand, utilizes the disease names to retrieve captions or images with the same disease label, which is proposed to demonstrate the effectiveness of the knowledge encoder. For cross-modal retrieval, we follow PLIP [23] to split the ARCH [15] dataset into Arch-PubMed and Arch-book. In addition, we also gather image-text pairs from publicly available educational resources and curate a retrieval dataset, termed as PathPair, consisting of 9,358 pathology image-caption pairs with known 1676 disease labels. For disease retrieval, we utilize the captions, images, and their disease labels in PathPair." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.787, + 0.549 + ], + "angle": 0, + "content": "Zero-shot Patch Classification. This task involves zero-shot classification on patch-level pathology images. Specifically, at inference time, we randomly select one template from the 21 templates in CONCH [36] and one type synonym from the corresponding name list (exhibited in Supplementary Materials) to yield a text prompt for each type, e.g. a histopathological image of CLASSNAME. (template) + beast invasive carcinoma (synonym) \\(\\rightarrow\\) a histopathological image of beast invasive carcinoma. This process is repeated 100 times in every experiment. Following PLIP [23] and Quilt1M [24], we adopt these patch-level pathology image datasets: BACH [1], NCT-CRC-HE-100K [27], KatherColon [28], LC25000 [3], RenalCell [4], SICAP [47], SkinCancer [30], WSSS4LUAD [17]. Each dataset includes multiple types of H&E stained cell micrographs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Zero-shot WSI Tumor Subtyping. This task involves zero-shot tumor subtyping on pathology whole slide images of common and rare cancers. We follow MI-Zero [37] and CONCH [36] for evaluation. Specifically, we first divide WSI into \\(256 \\times 256\\) patches and then predict the class label of each image patch in a zero-shot manner. The tumor type of the whole slide is then provided by integrating Top-K predictions on patches. For this task, we also employ the templates and tumor synonyms from CONCH [36] to randomly generate 100 text prompts (listed in Supplementary materials) for each tumor subtype. (i) For common cancers, we follow existing research [36,37] and collect 525 WSIs with three tumor types, including 150 breast carcinoma (BRCA), 150 non-small cell lung cancer (NSCLC) histopathology slides, and 225 renal cell carcinoma (RCC) slides, from TCGA. Specifically, BRCA consists of two subtypes: invasive ductal carcinoma (IDC) and invasive lobular carcinoma (IDC). NSCLC contains lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC). RCC is divided into chromophobe renal cell carcinoma (CHRCC), clear-cell renal cell carcinoma (CCRCC), and papillary renal cell carcinoma (PRCC). Each tumor subtype has 75 WSIs. (ii) For rare cancers, we collected 41 WSIs of rare subtypes for breast cancer (6 for Intraductal papillary adenocarcinoma with invasion, 6 for Medullary carcinoma, NOS, 13 for Metaplastic carcinoma, NOS and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "X. Zhou et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "16 for Mucinous adenocarcinoma), from TCGA and merge them with common breast cancers (75 for IDC and 75 for ILC)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.199, + 0.423, + 0.214 + ], + "angle": 0, + "content": "4.3 Evaluation Metrics." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.223, + 0.785, + 0.254 + ], + "angle": 0, + "content": "Retrieval. To evaluate the retrieval performance, we adopt the Recall@K metric, suggesting the ratio of correctly retrieved queries in Top-K retrieved samples." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.303 + ], + "angle": 0, + "content": "Zero-shot Patch Classification. For zero-shot classification tasks, we adopt the same metric as PLIP [23], namely, weighted F1 (wF1). We report the median, the first, and the third quartile (Q1, Q3) of wF1 across all text 100 prompts." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.306, + 0.787, + 0.368 + ], + "angle": 0, + "content": "Zero-shot WSI Tumor Subtyping. We adopt the same metric as MI-Zero [37], namely, balanced accuracy at Top-K pooling to measure the zero-shot tumor subtyping performance on WSIs. We also report the median, the first, and the third quartile (Q1, Q3) of the balanced accuracy across all 100 text prompts." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.388, + 0.456, + 0.403 + ], + "angle": 0, + "content": "4.4 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.416, + 0.785, + 0.477 + ], + "angle": 0, + "content": "Knowledge Encoder Pretraining. We adopt the architecture of PubMedBERT [16] to encode knowledge. The embedding dimension is set to 512. The temperature parameter \\(\\tau\\) is set to 0.04 in Eq. 2. The batch size is set to 256, including 32 disease entities with 8 instances per entity." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.48, + 0.787, + 0.602 + ], + "angle": 0, + "content": "Pathology Image-text Pretraining. To achieve a fair comparison with PLIP and Quilt1M, we utilize the same visual encoder (ViT-B-32) and initialization of the visual encoder (CLIP [41]), termed by KEP-32, and set the input image size to \\(224 \\times 224\\). Additionally, we also develop a KEP variant named KEP-16 that is initialized by the visual weights of BiomedCLIP [56] (ViT-B-16). The batch size and learning rate are set to 256 and 1e-5, respectively. The temperature in Eq. 6 and Eq. 7 is set to 0.04 across all experiments. For OpenPath, we conduct the pathology VLP for 30 epochs, While for Quilt1M, 15 epochs are adopted." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.624, + 0.324, + 0.64 + ], + "angle": 0, + "content": "5 Results" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.655, + 0.785, + 0.73 + ], + "angle": 0, + "content": "In this section, we show the results on the downstream tasks, to evaluate the effectiveness of proposed knowledge-enhanced representation learning. Note that for a fair comparison with the publicly available model released by the original authors, in all experiments, we separately train KEP on the OpenPath [23] dataset, and Quilt1M [24] dataset, and then report results on downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.751, + 0.338, + 0.766 + ], + "angle": 0, + "content": "5.1 Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.776, + 0.785, + 0.806 + ], + "angle": 0, + "content": "In this section, we evaluate the performance of retrieval, including cross-modal retrieval on three datasets and disease retrieval on PathPair." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Cross-modal Retrieval. In Table 1, we demonstrate the results for different models pretrained on OpenPath and Quilt1M. It can be seen that KEP-32" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.696, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KEP" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Performance comparison with PLIP and QuiltNet on three retrieval datasets. All models are pre-trained on the OpenPath and Quilt1M datasets, respectively. i2t and t2i denote image-to-text and text-to-image retrieval, respectively. Bold fonts suggest the best performance." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.213, + 0.779, + 0.377 + ], + "angle": 0, + "content": "
Training datasetTaskModelKnowledge EnhancementArch-PubMedArch-bookPathPair
Recall@10Recall@50Recall@10Recall@50Recall@10Recall@50
OpenPathi2tPLIPX0.0670.1850.1520.3930.0380.119
KEP-32✓0.0980.2830.1640.4040.0710.205
KEP-16✓0.1630.3980.2440.5370.1080.275
t2iPLIPX0.0670.1810.1650.4190.0470.133
KEP-32✓0.0850.2260.1480.3650.0610.171
KEP-16✓0.1380.3390.2380.5330.0930.247
Quilt1Mi2tQuiltNetX0.1390.3260.1880.4070.0650.166
KEP-32✓0.1400.3270.2400.5210.0840.221
KEP-16✓0.1960.4210.2820.5640.1080.254
t2iQuiltNetX0.1220.2930.2040.4290.0710.195
KEP-32✓0.1350.3260.2750.5680.1060.276
KEP-16✓0.1760.4040.3400.6210.1360.326
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.399, + 0.785, + 0.428 + ], + "angle": 0, + "content": "Table 2: Performance comparison of disease retrieval. 12t and i2l denote label-to-text and image-to-label, respectively. Bold fonts suggest the best performance." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.44, + 0.764, + 0.523 + ], + "angle": 0, + "content": "
TaskMetricsOpenPathQuiltNetQuilt1M
PLIPKEP-32KEP-16KEP-32KEP-16
l2tRecall@100.4080.6930.6990.2110.6480.643
Recall@500.5360.8320.8350.3250.8120.808
i2lRecall@100.1140.1620.2230.1350.2260.259
Recall@500.2920.3810.4780.3220.4730.519
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.56, + 0.788, + 0.696 + ], + "angle": 0, + "content": "pretrained on OpenPath outperforms PLIP on all datasets with respect to the image-to-text retrieval task, especially on Arch-PubMed and PathPair with more than \\(3\\%\\) boost on R10. As for the text-to-image task, KEP-32 also achieves better performance on the dataset of Arch-PubMed and PathPair. Furthermore, our KEP-16 model improves the retrieval performance by a large margin on all datasets across both retrieval tasks. Similar results can be concluded that KEP-32 pretrained on Quilt1M improves the performance on all datasets for both retrieval tasks, which demonstrates the effectiveness of knowledge guidance for visual-language pretraining on cross-modal retrieval." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.704, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Disease Retrieval. Table 2 shows the performance of disease retrieval on the PathPair dataset for different models. It can be seen that although the scale of Quilt1M is 5 times OpenPath, models pretrained on OpenPath often outperforms that on Quilt1M for the label-to-text task. We conjecture this may be caused by the quality of captions in Quilt1M. Our approach KEP-32 outperforms PLIP and Quilt1M by a large margin on both label-to-text and image-to-label tasks, suggesting that the knowledge encoder contributes to paying attention to the key disease information when encoding captions, and thus improves the alignment between images and their disease labels." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.127 + ], + "angle": 0, + "content": "X. Zhou et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.143, + 0.506, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.143, + 0.782, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.288, + 0.788, + 0.373 + ], + "angle": 0, + "content": "Fig. 5: The comparison of zero-shot patch classification between different models. The left and the right subfigures suggest pretraining on OpenPath and Quilt1M, respectively. The visual encoders of KEP-32 and KEP-16 are initialized by CLIP (ViT-B-32) and BiomedCLIP (ViT-B-16), respectively. The number of points for every box is 100, with each representing the performance of one text prompt. The upper, center, and lower line of each box denote the first, median, and third quartile of the distribution." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.4, + 0.51, + 0.415 + ], + "angle": 0, + "content": "5.2 Zero-shot Patch Classification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.784, + 0.47 + ], + "angle": 0, + "content": "In this section, we evaluate the performance of zero-shot classification on patch-level pathology images from 8 datasets. We report the performance distribution (Fig. 5), where each point in the figure denotes the performance of one prompt." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.625 + ], + "angle": 0, + "content": "Comparison to PLIP. In Fig. 5 (left), we demonstrate the performance comparison with PLIP [23], by only pretraining on OpenPath image-text pairs. Compared with PLIP, KEP-32 achieves better zero-shot classification performance on all datasets, with a maximum enhancement of 0.137 on the BACH dataset. Moreover, with the medical-specific initialization of the visual encoder (BiomedCLIP), KEP-16 can further improve the performance significantly in most datasets. Note that, the boxes of KEP variants are generally shorter than that of PLIP, suggesting that our approach is less sensitive to the varying text prompts than PLIP, which can be demonstrated by the visualization of prompt embeddings, shown in Supplementary Materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Comparison to QuiltNet. As shown in Fig. 5 (right), we exhibit the performance comparison between KEP and QuiltNet [24] pretrained on Quilt1M dataset. Similar to results on OpenPath, KEP-32 outperforms Quilt1M in seven out of eight datasets, especially for the BACH, KatherColon, and SkinCancer datasets. KEP-16 can further improve the weighted F1 score. Most boxes of KEP variants are also shorter than that of Quilt1M. The visualization of prompt embeddings, shown in Supplementary Materials, maintains that pathology knowledge can reduce the ambiguities of image-text alignment and thus improve the zero-shot classification performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.786, + 0.536, + 0.802 + ], + "angle": 0, + "content": "5.3 Zero-shot WSI Tumor Subtyping" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this section, we evaluate the transfer ability of different models for tumor subtyping on common and rare cancers." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.696, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KEP" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.143, + 0.777, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.276, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Fig. 6: The performance comparison of tumor subtyping on TCGA-BRCA (common), TCGA-NSCLC, and TCGA-RCC WSIs. The upper, center and lower line of each box denote the first, median, and third quartile of the performance distribution, respectively. The scattered points represent outliers. KEP-16 and KEP-CTP are trained on OpenPath with the visual encoder initialized by BiomedCLIP [56] and CTransPath [48], respectively. MI-Zero-Bio and MI-Zero-Pub are two variants of MI-Zero with different initialization of text encoder. Their visual encoders are initialized by CTransPath. In addition, MI-Zero adopts in-house pathology reports to pretrain their text encoders." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.406, + 0.785, + 0.434 + ], + "angle": 0, + "content": "Table 3: Performance comparison of tumor subtyping on six BRCA subtypes, including 2 common and 4 rare cancers. Bold fonts suggest the best performance." + }, + { + "type": "table", + "bbox": [ + 0.271, + 0.447, + 0.734, + 0.546 + ], + "angle": 0, + "content": "
ModelTraining datasetTop-1Top-10
CommonRareCommonRare
PLIPOpenPath0.1130.1150.1190.119
QuiltNetQuilt1M0.2780.0940.2790.102
KEP-16OpenPath0.3430.2170.3680.245
MI-Zero-BioIn-house & Web data0.3900.2140.4030.237
MI-Zero-PubIn-house & Web data0.3310.2820.3470.304
KEP-CTPOpenPath0.4430.2820.4320.301
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.788, + 0.775 + ], + "angle": 0, + "content": "Common cancers. Fig. 6 shows the performance comparison of tumor subtyping on common cancers, including TCGA-BRCA (common), TCGA-NSCLC and TCGA-RCC WSIs. MI-Zero-Bio and MI-Zero-Pub are two variants of MI-Zero with different text encoders. The visual encoders of the two MI-Zero variants are initialized by CTransPath [48]. To achieve a fair comparison, we also train a KEP variant named KEP-CTP with the visual encoder initialized by CTransPath. Both KEP-16 and KEP-CTP are trained on the OpenPath dataset. It can be seen that our approach KEP-16 and KEP-CTP outperform PLIP and QuiltNet on all datasets. Note that, the comparisons between variants of MI-Zero and KEP-CTP are actually not fair for our method, as the MI-Zero variants have also been pretrained on massive in-house data – over 550k pathology reports from hospitals [37]. Yet, KEP-CTP still outperforms MI-Zero-Bio across all datasets and achieves comparable performance with MI-Zero-Pub on NSCLC and RCC." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Rare cancers. Table 3 exhibits the median balanced accuracy of tumor subtyping on common and rare breast cancers. It can be seen that our method KEP-16 pretrained on OpenPath outperforms PLIP and QuiltNet by a large margin on both common and rare breast cancers. Moreover, KEP-CTP significantly outper" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "X. Zhou et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 4: Experimental results of ablation study on model architecture and initialization. Arch-PubMed and PathPair suggest the text-to-image retrieval and the disease retrieval task, respectively. PMB and KB denote PubMedBERT and our pretrained pathology knowledge encoder, respectively. BCLIP denotes BiomedCLIP. Bold fonts and underline suggest the best and the second-best performance, respectively." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.227, + 0.779, + 0.364 + ], + "angle": 0, + "content": "
Visual Init.Text Init.Projection headKnowledge distillMetric lossKatherColon Median (Q1, Q3)Arch-PubMedPathPair
R10R50R1R5
ScratchPMB0.322 (0.272, 0.392)0.0370.1210.0560.113
CLIPPMB0.407 (0.358, 0.452)0.0810.2140.1610.316
CLIPPMB✓0.486 (0.448, 0.517)0.0730.2100.1610.305
BCLIPPMB✓0.553 (0.503, 0.591)0.1270.3470.1870.350
CLIPPMB✓✓(PMB)0.484 (0.435, 0.520)0.0630.1820.0340.089
CLIPKB✓Adasp0.530 (0.477, 0.600)0.0860.2320.2700.482
CLIPKB✓✓(KB)Adasp0.563 (0.505, 0.610)0.0850.2260.4090.618
CLIPKB✓✓(KB)Triplet0.531 (0.482, 0.580)0.0870.2270.2650.493
BCLIPKB✓✓(KB)Adasp0.580 (0.517, 0.631)0.1380.3390.4240.618
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.786, + 0.418 + ], + "angle": 0, + "content": "forms MI-Zero variants on common breast cancers while achieving comparable performance on the rare ones." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.438, + 0.39, + 0.453 + ], + "angle": 0, + "content": "5.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.611 + ], + "angle": 0, + "content": "In this section, we explore the impact of model architecture, initialization, and other hyperparameters, as shown in Table 4 and Table S3 in Supplementary Materials. Comparing the performance between different models, we can draw the following observations: (i) visual projection head can facilitate the zero-shot classification performance; (ii) our pretrained knowledge encoder can enhance the performance across all tasks; (iii) the frozen knowledge branch can further improve the performance of zero-shot classification and disease retrieval tasks while the frozen PubMedBERT can not; (iv) a better initialization of visual encoder contributes to significant performance improvement on all tasks. (v) the Adasp loss is better than the Triplet loss for the knowledge encoding." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.631, + 0.36, + 0.647 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this paper, we address the problem of knowledge-enhanced visual-language pretraining on computational pathology. We first curate a pathology knowledge tree that integrates the informative attributes of diseases requiring pathological diagnosis. We then propose a novel knowledge encoding approach based on metric learning to model structured pathological knowledge. With the guidance of the pretrained knowledge encoder, we conduct extensive visual-language pretraining on pathology image-caption pairs. To demonstrate the effectiveness of our approach, we evaluate pretrained VLP models on three downstream tasks, including retrieval, zero-shot classification on patch-level pathology images, and zero-shot tumor subtyping on pathology WSIs. Quantitative experimental results demonstrate that pathology knowledge can significantly improve the performance across different tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.696, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KEP" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.786, + 0.22 + ], + "angle": 0, + "content": "This work is supported by the National Key R&D Program of China (No. 2022ZD0160702) and China Postdoctoral Science Foundation (Certificate Number: 2023M741850)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.24, + 0.323, + 0.256 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.786, + 0.31 + ], + "angle": 0, + "content": "1. Aresta, G., Araújo, T., Kwok, S., Chennamsetty, S.S., Safwan, M., Alex, V., Marami, B., Prastawa, M., Chan, M., Donovan, M., et al.: Bach: Grand challenge on breast cancer histology images. Medical Image Analysis 56, 122-139 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.786, + 0.336 + ], + "angle": 0, + "content": "2. Bodenreider, O.: The unified medical language system (uml's): integrating biomedical terminology. Nucleic Acids Research 32(suppl_1), D267-D270 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.336, + 0.786, + 0.377 + ], + "angle": 0, + "content": "3. Borkowski, A.A., Bui, M.M., Thomas, L.B., Wilson, C.P., DeLand, L.A., Mastorides, S.M.: Lung and colon cancer histopathological image dataset (lc25000). arXiv preprint arXiv:1912.12142 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.377, + 0.786, + 0.418 + ], + "angle": 0, + "content": "4. Brummer, O., Pölönen, P., Mustjoki, S., Brück, O.: Integrative analysis of histological textures and lymphocyte infiltration in renal cell carcinoma using deep learning. bioRxiv pp. 2022-08 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.418, + 0.786, + 0.472 + ], + "angle": 0, + "content": "5. Campanella, G., Hanna, M.G., Geneslaw, L., Miraflor, A., Werneck Krauss Silva, V., Busam, K.J., Brogi, E., Reuter, V.E., Klimstra, D.S., Fuchs, T.J.: Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature Medicine 25(8), 1301-1309 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.472, + 0.786, + 0.527 + ], + "angle": 0, + "content": "6. Chan, T.H., Cendra, F.J., Ma, L., Yin, G., Yu, L.: Histopathology whole slide image analysis with heterogeneous graph representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15661-15670 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.527, + 0.786, + 0.582 + ], + "angle": 0, + "content": "7. Chauhan, G., Liao, R., Wells, W., Andreas, J., Wang, X., Berkowitz, S., Horng, S., Szolovits, P., Golland, P.: Joint modeling of chest radiographs and radiology reports for pulmonary edema assessment. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 529-539. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.786, + 0.623 + ], + "angle": 0, + "content": "8. Chen, C., Lu, M.Y., Williamson, D.F., Chen, T.Y., Schaumberg, A.J., Mahmood, F.: Fast and scalable search of whole-slide images via self-supervised deep learning. Nature Biomedical Engineering 6(12), 1420-1434 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.786, + 0.678 + ], + "angle": 0, + "content": "9. Chen, R.J., Chen, C., Li, Y., Chen, T.Y., Trister, A.D., Krishnan, R.G., Mahmood, F.: Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16144-16155 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.678, + 0.786, + 0.704 + ], + "angle": 0, + "content": "0. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: Uniter: Learning universal image-text representations (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.704, + 0.786, + 0.759 + ], + "angle": 0, + "content": "1. Chen, Y.C., Lu, C.S.: Rankmix: Data augmentation for weakly supervised learning of classifying whole slide images with diverse sizes and imbalanced categories. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23936-23945 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.786, + 0.8 + ], + "angle": 0, + "content": "2. Chen, Z., Li, G., Wan, X.: Align, reason and learn: Enhancing medical vision-and-language pre-training with knowledge. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 5152-5161 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.786, + 0.841 + ], + "angle": 0, + "content": "3. Chen, Z., Song, Y., Chang, T.H., Wan, X.: Generating radiology reports via memory-driven transformer. In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). pp. 1439-1449 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.267, + 0.786, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "X. Zhou et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "14. Cui, H., Xu, Y., Li, W., Wang, L., Duh, H.: Collaborative learning of cross-channel clinical attention for radiotherapy-related esophageal fistula prediction from ct. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 212-220. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.787, + 0.245 + ], + "angle": 0, + "content": "15. Gamper, J., Rajpoot, N.: Multiple instance captioning: Learning representations from histopathology textbooks and articles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16549-16559 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.787, + 0.3 + ], + "angle": 0, + "content": "16. Gu, Y., Tinn, R., Cheng, H., Lucas, M., Usuyama, N., Liu, X., Naumann, T., Gao, J., Poon, H.: Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Computing for Healthcare (HEALTH) 3(1), 1-23 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "17. Han, C., Pan, X., Yan, L., Lin, H., Li, B., Yao, S., Lv, S., Shi, Z., Mai, J., Lin, J., et al.: Wsss4luad: Grand challenge on weakly-supervised tissue semantic segmentation for lung adenocarcinoma. arXiv preprint arXiv:2204.06455 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.343, + 0.787, + 0.37 + ], + "angle": 0, + "content": "18. Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.371, + 0.787, + 0.412 + ], + "angle": 0, + "content": "19. Hu, S., Chen, L., Wu, P., Li, H., Yan, J., Tao, D.: St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In: European Conference on Computer Vision. pp. 533-549. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.412, + 0.787, + 0.48 + ], + "angle": 0, + "content": "20. Huang, S.C., Shen, L., Lungren, M.P., Yeung, S.: Gloria: A multimodal global-local representation learning framework for label-efficient medical image recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3942-3951 (2021), official Implementation: https://github.com/marshuang80/gloria" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "21. Huang, X., Fang, Y., Lu, M., Yan, F., Yang, J., Xu, Y.: Dual-ray net: automatic diagnosis of thoracic diseases using frontal and lateral chest x-rays. Journal of Medical Imaging and Health Informatics 10(2), 348-355 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.523, + 0.787, + 0.578 + ], + "angle": 0, + "content": "22. Huang, Y., Zhao, W., Wang, S., Fu, Y., Jiang, Y., Yu, L.: Conslide: Asynchronous hierarchical interaction transformer with breakup-reorganize rehearsal for continual whole slide image analysis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 21349-21360 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.578, + 0.787, + 0.619 + ], + "angle": 0, + "content": "23. Huang, Z., Bianchi, F., Yuksekgonul, M., Montine, T.J., Zou, J.: A visual-language foundation model for pathology image analysis using medical twitter. Nature medicine 29(9), 2307-2316 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.62, + 0.787, + 0.66 + ], + "angle": 0, + "content": "24. Ikezogwo, W., Seyfioglu, S., Ghezloo, F., Geva, D., Sheikh Mohammed, F., Anand, P.K., Krishna, R., Shapiro, L.: Quilt-1m: One million image-text pairs for histopathology. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.661, + 0.787, + 0.715 + ], + "angle": 0, + "content": "25. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning. pp. 4904-4916. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.716, + 0.787, + 0.758 + ], + "angle": 0, + "content": "26. Kang, M., Song, H., Park, S., Yoo, D., Pereira, S.: Benchmarking self-supervised learning on diverse pathology datasets. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3344-3354 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.759, + 0.787, + 0.785 + ], + "angle": 0, + "content": "27. Kather, J.N., Halama, N., Marx, A.: 100,000 histological images of human colorectal cancer and healthy tissue. Zenodo10 5281 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "28. Kather, J.N., Krisam, J., Charoentong, P., Luedde, T., Herpel, E., Weis, C.A., Gaiser, T., Marx, A., Valous, N.A., Ferber, D., et al.: Predicting survival from colorectal cancer histology slides using deep learning: A retrospective multicenter study. PLoS Medicine 16(1), e1002730 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.696, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KEP" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "29. Kim, W., Son, B., Kim, I.: Vilt: Vision-and-language transformer without convolution or region supervision. In: International Conference on Machine Learning. pp. 5583-5594. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.245 + ], + "angle": 0, + "content": "30. Kriegsmann, K., Lobers, F., Zgorzelski, C., Kriegsmann, J., Janssen, C., Meliss, R.R., Muley, T., Sack, U., Steinbuss, G., Kriegsmann, M.: Deep learning for the detection of anatomical tissue structures and neoplasms of the skin on scanned histopathological tissue sections. Frontiers in Oncology 12, 1022967 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.788, + 0.286 + ], + "angle": 0, + "content": "31. Kundra, R., Zhang, H., Sheridan, R., Sirlintrapun, S.J., Wang, A., Ochoa, A., Wilson, M., Gross, B., Sun, Y., Madupuri, R., et al.: Oncotree: a cancer classification system for precision oncology. JCO clinical cancer informatics 5, 221-230 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.342 + ], + "angle": 0, + "content": "32. Li, H., Zhu, C., Zhang, Y., Sun, Y., Shui, Z., Kuang, W., Zheng, S., Yang, L.: Task-specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7454-7463 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.397 + ], + "angle": 0, + "content": "33. Li, M., Cai, W., Verspoor, K., Pan, S., Liang, X., Chang, X.: Cross-modal clinical graph transformer for ophthalmic report generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20656-20665 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.439 + ], + "angle": 0, + "content": "34. Lin, T., Yu, Z., Hu, H., Xu, Y., Chen, C.W.: Interventional bag multi-instance learning on whole-slide pathological images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19830-19839 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.495 + ], + "angle": 0, + "content": "35. Lin, W., Zhao, Z., Zhang, X., Wu, C., Zhang, Y., Wang, Y., Xie, W.: Pmc-clip: Contrastive language-image pre-training using biomedical documents. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.536 + ], + "angle": 0, + "content": "36. Lu, M.Y., Chen, B., Williamson, D.F., Chen, R.J., Liang, I., Ding, T., Jaume, G., Odintsov, I., Zhang, A., Le, L.P., et al.: Towards a visual-language foundation model for computational pathology. arXiv preprint arXiv:2307.12914 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.592 + ], + "angle": 0, + "content": "37. Lu, M.Y., Chen, B., Zhang, A., Williamson, D.F., Chen, R.J., Ding, T., Le, L.P., Chuang, Y.S., Mahmood, F.: Visual language pretrained multiple instance zero-shot transfer for histopathology images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19764-19775 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.633 + ], + "angle": 0, + "content": "38. Lu, M.Y., Williamson, D.F., Chen, T.Y., Chen, R.J., Barbieri, M., Mahmood, F.: Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering 5(6), 555-570 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.675 + ], + "angle": 0, + "content": "39. Müller, P., Kaissis, G., Zou, C., Rueckert, D.: Joint learning of localized representations from medical images and reports. In: European Conference on Computer Vision. pp. 685-701. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.758 + ], + "angle": 0, + "content": "40. Neumann, M., King, D., Beltagy, I., Ammar, W.: ScispaCy: Fast and robust models for biomedical natural language processing. In: Demner-Fushman, D., Cohen, K.B., Ananiadou, S., Tsujii, J. (eds.) Proceedings of the 18th BioNLP Workshop and Shared Task. pp. 319-327. Association for Computational Linguistics, Florence, Italy (Aug 2019). https://doi.org/10.18653/v1/W19-5034, https://aclanthology.org/W19-5034" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.758, + 0.788, + 0.813 + ], + "angle": 0, + "content": "41. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.813, + 0.788, + 0.827 + ], + "angle": 0, + "content": "42. Roberts, R.J.: Pubmed central: The genbank of the published literature (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.827, + 0.654, + 0.841 + ], + "angle": 0, + "content": "43. Rorke, L.B.: Pathologic diagnosis as the gold standard (1997)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "X. Zhou et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.203 + ], + "angle": 0, + "content": "44. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278–25294 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.245 + ], + "angle": 0, + "content": "45. Shaban, M., Awan, R., Fraz, M.M., Azam, A., Tsang, Y.W., Snead, D., Rajpoot, N.M.: Context-aware convolutional neural network for grading of colorectal cancer histology images. IEEE Transactions on Medical Imaging 39(7), 2395-2405 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "46. Shao, Z., Bian, H., Chen, Y., Wang, Y., Zhang, J., Ji, X., et al.: Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in Neural Information Processing Systems 34, 2136-2147 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.342 + ], + "angle": 0, + "content": "47. Silva-Rodriguez, J., Colomer, A., Sales, M.A., Molina, R., Naranjo, V.: Going deeper through the gleason scoring scale: An automatic end-to-end system for histology prostate grading and cribriform pattern detection. Computer Methods and Programs in Biomedicine 195, 105637 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "48. Wang, X., Yang, S., Zhang, J., Wang, M., Zhang, J., Yang, W., Huang, J., Han, X.: Transformer-based unsupervised contrastive learning for histopathological image classification. Medical Image Analysis 81, 102559 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.412 + ], + "angle": 0, + "content": "49. Wu, C., Zhang, X., Zhang, Y., Wang, Y., Xie, W.: Medklip: Medical knowledge enhanced language-image pre-training. medRxiv pp. 2023-01 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "50. Xie, X., Niu, J., Liu, X., Chen, Z., Tang, S., Yu, S.: A survey on incorporating domain knowledge into deep learning for medical image analysis. Medical Image Analysis 69, 101985 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.508 + ], + "angle": 0, + "content": "51. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettle-moyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 6787-6800 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "52. Xuan, H., Stylianou, A., Liu, X., Pless, R.: Hard negative examples are hard, but useful. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 126-142. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "53. Yang, S., Wu, X., Ge, S., Zhou, S.K., Xiao, L.: Knowledge matters: Chest radiology report generation with general and specific knowledge. Medical Image Analysis 80, 102510 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.633 + ], + "angle": 0, + "content": "54. Yu, F., Tang, J., Yin, W., Sun, Y., Tian, H., Wu, H., Wang, H.: Ernie-vil: Knowledge enhanced vision-language representations through scene graphs. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 3208-3216 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "55. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "56. Zhang, S., Xu, Y., Usuyama, N., Bagga, J., Tinn, R., Preston, S., Rao, R., Wei, M., Valluri, N., Wong, C., et al.: Large-scale domain-specific pretraining for biomedical vision-language processing. arXiv preprint arXiv:2303.00915 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "57. Zhang, X., Wu, C., Zhang, Y., Xie, W., Wang, Y.: Knowledge-enhanced visual-language pre-training on chest radiology images. Nature Communications 14(1), 4542 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.813 + ], + "angle": 0, + "content": "58. Zhang, Y., Jiang, H., Miura, Y., Manning, C.D., Langlotz, C.P.: Contrastive learning of medical visual representations from paired images and text. In: Machine Learning for Healthcare (2022), highest Starred Implementation: https://github.com/edreisMD/ConVIRT-pytorch" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "59. Zhou, X., Cheng, Z., Gu, M., Chang, F.: Lirnet: Local integral regression network for both strongly and weakly supervised nuclei detection. In: 2020 IEEE Interna" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.696, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "KEP" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "tional Conference on Bioinformatics and Biomedicine (BIBM). pp. 945-951. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "60. Zhou, X., Zhong, Y., Cheng, Z., Liang, F., Ma, L.: Adaptive sparse pairwise loss for object re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19691-19701 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.786, + 0.218 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Knowledge-enhanced Visual-Language Pretraining for Computational Pathology/a04fca37-34eb-47aa-8e9d-2ee68fcc1f47_origin.pdf b/2024/Knowledge-enhanced Visual-Language Pretraining for Computational Pathology/a04fca37-34eb-47aa-8e9d-2ee68fcc1f47_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..12e2973f2e3aee5b003349bbe01d1756de3342dc --- /dev/null +++ b/2024/Knowledge-enhanced Visual-Language Pretraining for Computational Pathology/a04fca37-34eb-47aa-8e9d-2ee68fcc1f47_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e5d3e24a67529c4017c8fe26a58f60ef53ec8c9b0f1072ae897cdf5530d15c2 +size 1280868 diff --git a/2024/Knowledge-enhanced Visual-Language Pretraining for Computational Pathology/full.md b/2024/Knowledge-enhanced Visual-Language Pretraining for Computational Pathology/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6727243f734b9db303033f6264441667683f1144 --- /dev/null +++ b/2024/Knowledge-enhanced Visual-Language Pretraining for Computational Pathology/full.md @@ -0,0 +1,305 @@ +# Knowledge-enhanced Visual-Language Pretraining for Computational Pathology + +Xiao Zhou1, Xiaoman Zhang1,2, Chaoyi Wu1,2, Ya Zhang1,2, Weidi Xie1,2, Yanfeng Wang1,2 + +$^{1}$ Shanghai Artificial Intelligence Laboratory + +2 Shanghai Jiao Tong University + +https://github.com/MAGIC-AI4Med/KEP + +zhouxiao@pjlab.org.cn, {xm99sjtu, wtzxxxxwy02, ya_zhang, weidi, + +wangyanfeng} @sjtu.edu.cn + +Abstract. In this paper, we consider the problem of visual representation learning for computational pathology, by exploiting large-scale image-text pairs gathered from public resources, along with the domain-specific knowledge in pathology. Specifically, we make the following contributions: (i) We curate a pathology knowledge tree that consists of 50,470 informative attributes for 4,718 diseases requiring pathology diagnosis from 32 human tissues. To our knowledge, this is the first comprehensive structured pathology knowledge base; (ii) We develop a knowledge-enhanced visual-language pretraining approach, where we first project pathology-specific knowledge into latent embedding space via a language model, and use it to guide the visual representation learning; (iii) We conduct thorough experiments to validate the effectiveness of our proposed components, demonstrating significant performance improvement on various downstream tasks, including cross-modal retrieval, zero-shot classification on pathology patches, and zero-shot tumor subtyping on whole slide images (WSIs). + +Keywords: Pathology knowledge $\cdot$ Visual-language pretraining + +# 1 Introduction + +Pathology diagnosis is currently the golden standard for examining various diseases in clinical applications, especially in the diagnosis of neoplasm [43]. In the last decade, the prosperity of deep learning on computer vision has led to the rapid development of computational pathology, for example, approaches with supervised learning [6,22,34,45,46], and weakly supervised learning [5,11,32,38,59]. Despite being promising, these approaches have been fundamentally limited by the scale of costly labels. Alternatively, self-supervised pretraining on numerous unlabeled pathological images [8,9,26,48] has attracted unprecedented attention, yet it still requires supervised fine-tuning for downstream deployments. + +In the recent literature [25,41], studies on the multimodal foundation model have demonstrated conspicuous improvement in downstream zero-shot tasks, by + +![](images/a239cf8b1e899a359e14ef525856e627b734f565f07d6dc364c42af627c3eec6.jpg) +Fig. 1: Knowledge-enhanced pathology image-text alignment. The short caption of a pathology image crawled from public websites is typically unstructured and with varying granularities, which introduces noticeable ambiguities for image-text alignment. While the implicit structures and correlations between different image-caption pairs could be constructed by explicit disease attributes (dashed boxes), which can be well-aligned by a pathology knowledge tree. LUSC and LUAD suggest lung squamous cell carcinoma, and lung adenocarcinoma, respectively. + +simply training to align visual and language embedding space on massive image-text pairs crawled from public websites. In contrast to computer vision [19,25,41], representation learning in pathology requires tremendous expertise and domain knowledge. An ideal training corpus would thus consist of well-structured medical reports from the hospital, however, it is often extremely difficult to acquire, due to privacy concerns. Alternatively, recent works [23, 24, 36, 37] propose to gather large-scale image-caption pairs from public resources (PubMed [42] papers, Twitter, Youtube videos). Compared to discrete image labels, image captions from medical reports and academic articles can potentially provide more valuable medical information without manual annotation. + +For existing works that adopt simple contrastive learning on pathology image and caption pairs, they suffer from the following challenges: First, the short captions from these web-crawled image-text pairs are typically noisy, unstructured, and lack domain knowledge (Fig. 1), which can be sub-optimal for constructing high-quality pathology-specific visual representations; Second, the varying granularities of the free text from captions will inevitably introduce ambiguities when aligning with images, thus causing the model to be highly sensitive to the used text prompts during the inference stage [36]. Thus, effectively harnessing the potential of such web-crawled data remains to be a challenge. + +In this paper, to tackle the above challenges, we anticipate that introducing pathology knowledge is of great significance to make up for the deficiency of short image captions. To this end, we make the following contributions: (i) We curate a pathology knowledge tree, PathKT, by collecting 50,470 informative pathological attributes of 4,718 diseases in 32 tissues from publicly available educational + +resources and OncoTree $^{3}$ [31]. (ii) We propose a novel knowledge encoder pretraining approach, that projects the attributes of each disease from PathKT into latent embedding space, where the attributes of the same disease, including disease synonyms, definitions, histology, and cytology features, share similar representations. (iii) We develop a knowledge-enhanced pretraining (KEP) approach to align pathology visual-language representations, which freezes the knowledge encoder and continuously injects domain-specific knowledge into the image-text embedding space. To demonstrate the effectiveness of our proposed approach, we conduct thorough experiments on three downstream tasks, including retrieval on three pathological image-caption datasets, zero-shot patch classification on eight patch-level pathology image datasets, and zero-shot WSI tumor subtyping on three datasets from The Cancer Genome Atlas (TCGA) $^{4}$ . Quantitative experiments suggest that knowledge guidance can significantly enhance the performance across different tasks. + +# 2 Related work + +Vision-Language Pretraining (VLP). Current vision-language pretraining approaches are typically categorized into two groups. The first group, referred to as two-stream approaches [41, 51], involves using two separate encoders to extract features for visual and textual data, respectively. The second group, known as single-stream methods [10, 29], utilizes a cross-modal fusion encoder to enhance interactions between vision and text features. In the field of medical VLP, existing methods mainly adopt the two-stream approach [7, 12, 20, 35, 39, 56, 58]. Despite the valuable contributions of these methods, the reliance on data-driven representation learning restricts the utilization of systematic and structured medical knowledge, leading to less than professional diagnosis. + +VLP in Computational Pathology. In the recent literature, Huang et al. proposed a pathology VLP model named PLIP [23] and released OpenPath that contains 200K image-caption pairs from the social media of Twitter and other public resources. To extend the pathology data scale, Ikezogwo et al. curate Quilt1m [24] with over one million histopathology image-text pairs by capturing keyframes and speech from YouTube videos. In addition to evaluating the zero-shot classification on patch-level pathology images, Lu et al. propose MI-Zero [37] and CONCH [36] to extend the transfer ability of pathology VLP models on gigapixel whole slide images (WSIs). In contrast to these existing work, that finetunes a CLIP [41] or CoCa [55] on pathology image-caption pairs, we propose to first train a pathology-specific knowledge encoder, and use it to guide visual-language representation learning. + +Medical Knowledge-enhanced Learning. In the medical community, leveraging external medical knowledge to enhance deep learning models has become an increasingly important topic [49, 50, 57]. Generally, existing approaches can + +![](images/0b346c9114d5f910f02e917024335522a54969e1e35a14d0aabf055f0a15b2e9.jpg) +Fig. 2: The construction of pathology knowledge tree. OncoTree is adopted as the base architecture to construct the PathKT. The tissue types, disease entities, and attributes are first extracted from web-crawled pathological descriptions, where cancers are then matched to OncoTree based on their tissue types and tumor types/subtypes using UMLS CUIs. Moreover, non-tumor diseases are added to the knowledge tree according to their tissue types. Finally, the pathology knowledge tree integrates 4718 diseases from 32 tissues, with each disease containing various synonyms, definitions, and histological and cytological features. CSCLC in this figure suggests combined small cell lung cancer. + +be categorized based on the ways of using medical knowledge: model-based approaches [14, 21] adopts the prior knowledge of radiology or diagnosis summarized by doctors to design algorithms; and input-based methods [13, 33, 53, 54] directly exploit knowledge as the external input to guide representation training. However, most of these works are focused on the analysis of chest X-rays. + +# 3 Methods + +Our primary goal is to leverage structured pathology knowledge to enhance visual-language representation learning. To start with, we construct a **Pathology Knowledge Tree**, termed as **PathKT**, that consists of 50,470 informative attributes of 4718 diseases from 32 human tissues (Sec. 3.1). We then train a knowledge encoder that projects the structured pathology knowledge into an embedding space (Sec. 3.2). We further employ the knowledge encoder to guide visual-language pretraining for computational pathology, termed as KEP (Knowledge-Enhanced Pre-training, Sec. 3.3). + +# 3.1 PathKT Construction + +Here, we detail the procedure for building up a pathology knowledge tree with various online sources, which will be further used for knowledge encoding. + +Knowledge Source. We collect pathology-specific knowledge from publicly available educational resources, such as text books, professional websites, and structured databases (OncoTree [31]). Specifically, we extract pathological descriptions of 884 tumor subtypes from OncoTree and 4360 diseases from text books and professional websites, including all domain knowledge required for + +diagnosis in clinical scenarios, i.e., disease name/ synonyms, definitions, histological features, and cytological features. + +Knowledge Tree Construction. We structure these knowledge sources into a knowledge tree by expanding the OncoTree, as shown in Fig. 2. OncoTree is a tree-structure cancer classification system [31], which consists of 884 tumor subtypes from 32 tissues, with each type of tumor linked to a Concept Unique Identifier (CUI) from Unified Medical Language System (UMLS) [2]. Specifically, we first extract tissue types, disease entities, and disease attributes from pathological descriptions of 4360 diseases, in which 168 cancers are found overlapped with OncoTree by using SciSpacy package [40] to link the UMLS CUIs. The rest 4192 diseases are then added to the knowledge tree according to their tissue types. After deduplication and noise reduction (358 diseases without any informative descriptions are deleted), all diseases are organized into the corresponding tissues. The final histopathology knowledge tree contains 4718 diseases from 32 tissues. The attributes of each disease node are constructed by a varying number of synonyms, definitions, and histological and cytological descriptions. The statistics of PathKT is shown in Table S1 and Fig. S1 in Supplementary Materials. In the final PathKT, for instance, the tissue node of the lung connects all lung diseases, including combined small cell lung cancer (CSCLC), lung squamous cell carcinoma (LUSC), and lung adenocarcinoma (LUAD), etc. The LUSC node connects four kinds of attributes, shown in Fig. 2. + +# 3.2 Pathology Knowledge Encoding + +In this section, we describe details for projecting tree-structure pathological knowledge into a latent embedding space, by training a knowledge encoder. Specifically, we align different disease entities with their corresponding free-text attributes via metric learning, such that, the synonyms, definitions, and corresponding histological/cytological features are close in the embedding space. As a consequence, the model can link the pathology images to their implicit disease labels during the visual-language pertaining, since the free-form text descriptions in image-text pairs may contain disease attributes, such as pathological features and disease definitions, which have already been aligned with disease names/synonyms. Therefore, the alignment of pathology knowledge can help the model link to the disease entities for better performance during diagnosis. + +Problem Setting. Given a set of disease entities with their corresponding attributes, $\mathcal{D} = \{(d_1,\mathbf{a}_1),\ldots ,(d_n,\mathbf{a}_n)\}$ , where $d_{i}$ denotes the $i$ -th disease entity, and $\mathbf{a}_i = \{a_i^1,\dots ,a_i^k\}$ refer to the associated $k$ attributes, both disease and attributes are represented in the format of natural language. Note that, for different disease entities, $k$ also varies, our goal here is to train a model that satisfies: + +$$ +\operatorname {s i m} \left(\Phi_ {\mathrm {k}} \left(a _ {i} ^ {p}\right), \Phi_ {\mathrm {k}} \left(a _ {i} ^ {q}\right)\right) \gg \operatorname {s i m} \left(\Phi_ {\mathrm {k}} \left(a _ {i} ^ {p}\right), \Phi_ {\mathrm {k}} \left(a _ {j} ^ {t}\right)\right), i \neq j, \tag {1} +$$ + +where $\varPhi_{\mathbf{k}}(\cdot)$ denotes the knowledge encoder, $\mathrm{sim}(\cdot ,\cdot)$ refers to the similarity, $a_i^p,a_i^q$ and $a_j^t$ refer to the randomly sampled attributes from the $i,j$ -th disease entity. + +![](images/dd3628686907ec3b9bdb5e0fb419731044620b613a54f42695713b29f2476bf3.jpg) +Fig. 3: Knowledge encoder pretraining based on metric learning. $n$ disease entities and each with $k$ attributes, including disease synonyms, definitions, cytology and pathology features, construct a mini-batch (left part of the figure), which are fed to a knowledge encoder for pretraining. In the embedding space (right part of the figure), the markers in different shapes represent the embeddings of attributes of different diseases. $S_{i}^{+}$ suggests the max-min positive attribute similarity within the $i$ -th disease, while $S_{i}^{-}$ denotes the maximal attribute similarity between the $i$ -th disease and other diseases. The goal of metric learning is to increase $S_{i}^{+}$ and meanwhile decreasing $S_{i}^{-}$ . The purple-dashed arrow and circle denote the minimal positive attribute similarity in the second class and the hypersphere it spans. + +Intuitively, the knowledge encoder enables the attributes of the same disease to be pulled together, while attributes from different diseases are pushed apart. + +Training. In order to achieve the objective defined in Eq. 1, metric learning is exploited to construct an embedding space where the representations of intra/inter-class instances are clustered/separated. In specific, given a mini-batch that contains $n$ random diseases and each with $k$ attributes, we denote the normalized embedding for the $p$ -th attribute of $i$ -th disease as: $\mathbf{z}_p^i = \varPhi_{\mathrm{k}}(a_i^p) / \| \varPhi_{\mathrm{k}}(a_i^p)\|$ . We adopt the recently proposed AdaSP loss [60], which finds out a max-min positive similarity (Fig. 3) and then shapes a loss with the maximal negative similarity: + +$$ +\mathcal {L} _ {\text {m e t i c}} = \frac {1}{n} \sum_ {i = 1} ^ {n} \log \left(1 + e ^ {\left(S _ {i} ^ {-} - S _ {i} ^ {+}\right) / \tau}\right), \tag {2} +$$ + +where $\tau$ is a temperature parameter. $S_{i}^{+}$ and $S_{i}^{-}$ denote the max-min positive and the maximal negative similarity, which can be computed by the soft version: + +$$ +S _ {i} ^ {+} = \max _ {p} \min _ {q} \left\langle \mathbf {z} _ {p} ^ {i}, \mathbf {z} _ {q} ^ {i} \right\rangle \approx \tau \log \left(\sum_ {p = 1} ^ {k} \frac {1}{\sum_ {q = 1} ^ {k} e ^ {- \frac {\left\langle \mathbf {z} _ {p} ^ {i} , \mathbf {z} _ {q} ^ {i} \right\rangle}{\tau}}}\right) \tag {3} +$$ + +$$ +S _ {i} ^ {-} = \max _ {j, p, q} \left\langle \mathbf {z} _ {p} ^ {i}, \mathbf {z} _ {q} ^ {j} \right\rangle \approx \tau \log \left(\sum_ {p = 1} ^ {k} \sum_ {j = 1, j \neq i} ^ {n} \sum_ {q = 1} ^ {k} e ^ {\frac {\left\langle \mathbf {z} _ {p} ^ {i} , \mathbf {z} _ {q} ^ {j} \right\rangle}{\tau}}\right) \tag {4} +$$ + +where $\langle \cdot \rangle$ represents the cosine similarity, the details about the soft version can be found in Supplementary Materials. + +Discussion. The conventional triplet loss [18] with batch hard mining strategy is widely adopted in metric learning, while it is not suitable in our case due to the following two reasons: First, in the pathology knowledge, each disease entity is associated with at most four types of attributes, the text descriptions of different attribute types might reveal significant divergence, which causes a large intra-class variation, marked by the purple dashed circle in Fig. 3, during training. Second, the fine-grained diseases might share high similarities, such as tumor subtypes: breast invasive ductal carcinoma and breast invasive lobular carcinoma, which cause low inter-class variations during training. When a mini-batch meets these two conditions, traditional triplet loss that enforces the instance similarities could typically produce the bad local minima of optimization [52], which, therefore, undermines the metric learning. + +# 3.3 Pathology Knowledge Enhanced Pretraining + +In this section, we present a simple yet effective pretraining approach, termed KEP, that leverages the established knowledge encoder to guide visual-language pretraining for computational pathology. + +Visual-Language Pretraining. Given paired image and captions, denoted as $\mathcal{F} = \{(x_1, c_1), \ldots, (x_n, c_n)\}$ , our goal is to construct a visual-language embedding space from paired image-text data, that satisfies: + +$$ +\left. \operatorname {s i m} \left(\Phi_ {\mathrm {v}} \left(x _ {i}\right), \Phi_ {\mathrm {t}} \left(c _ {i}\right)\right) \gg \operatorname {s i m} \left(\Phi_ {\mathrm {v}} \left(x _ {i}\right), \Phi_ {\mathrm {t}} \left(c _ {j}\right)\right), i \neq j, \right. \tag {5} +$$ + +where $\varPhi_{\mathrm{v}}(\cdot)$ and $\varPhi_{\mathrm{t}}(\cdot)$ denote the visual and the textual encoder, respectively. In this work, ViT-B-32/16 is adopted as the backbone for the visual encoder and initialized from the visual weights of CLIP [41] / BiomedCLIP [56]. A projection head is added on top of the visual encoder to bridge the gap caused by the non-pathology initialization, shown in Fig. 4. The text encoder is initialized by the pretrained knowledge encoder, as described in Sec. 3.2. In order to learn effective visual-language representation, we optimise an infoNCE contrastive objective: + +$$ +\mathcal {L} _ {\mathrm {v t}} = - \left(\log \frac {e ^ {\mathbf {v} _ {i} ^ {T} \mathbf {t} _ {i} / \tau}}{\sum_ {j = 1} ^ {n} e ^ {\mathbf {v} _ {i} ^ {T} \mathbf {t} _ {j} / \tau}} + \log \frac {e ^ {\mathbf {t} _ {i} ^ {T} \mathbf {v} _ {i} / \tau}}{\sum_ {j = 1} ^ {n} e ^ {\mathbf {t} _ {i} ^ {T} \mathbf {v} _ {j} / \tau}}\right), \tag {6} +$$ + +where $\mathbf{v}_i = \varPhi_{\mathrm{v}}(x_i)$ , and $\mathbf{t}_i = \varPhi_{\mathrm{t}}(c_i)$ , refer to the normalized embedding vectors from visual and text encoders, respectively. $\tau$ denotes a temperature parameter. + +Knowledge Distillation. To keep the alignment between images and free-form captions inside the knowledge space and thus the images can be linked to their implicit disease entities (right part in Fig. 4), we adopt an additional frozen branch to continuously distill pathology knowledge to the text encoder. Specifically, we use the text-knowledge embedding pairs to construct a contrastive loss item $\mathcal{L}_{\mathrm{tk}}$ , which can be computed by Eq. 6 with the normalized visual embedding vectors replaced by the knowledge embedding vectors: + +$$ +\mathcal {L} _ {\mathrm {t k}} = - \left(\log \frac {e ^ {\mathbf {k} _ {i} ^ {T} \mathbf {t} _ {i} / \tau}}{\sum_ {j = 1} ^ {n} e ^ {\mathbf {k} _ {i} ^ {T} \mathbf {t} _ {j} / \tau}} + \log \frac {e ^ {\mathbf {t} _ {i} ^ {T} \mathbf {k} _ {i} / \tau}}{\sum_ {j = 1} ^ {n} e ^ {\mathbf {t} _ {i} ^ {T} \mathbf {k} _ {j} / \tau}}\right), \tag {7} +$$ + +![](images/6614bfe1a57a81c0f11e2fffb9c0ba55b16a9d4704a935fb83700232a46369b4.jpg) +Fig. 4: Model architecture (left graph). A projection head is added on the top of the visual encoder to bridge the gap between the image and the text encoder. The knowledge encoder is frozen across the whole training stage to distill pathology knowledge to the learnable text encoder. As a result, the pathology images can be aligned with their implicit disease labels (marked by dashed boxes in the right graph) during visual-language pretraining, since the captions contain disease attributes that have been already aligned with disease names/synonyms in the knowledge embedding space. + +where $\mathbf{k}_i = \varPhi_{\mathrm{k}}(c_i)$ , refer to the normalized embedding vectors from the knowledge encoder. The overall training loss can thus be computed as: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {v t}} + \alpha \mathcal {L} _ {\mathrm {t k}}, \tag {8} +$$ + +where $\alpha$ denotes a weight parameter. It is worth emphasizing that the key contribution in KEP is to initialize the text encoder with the pre-obtained pathology knowledge encoder and adopt it to continuously distill pathology knowledge to the text encoder, while for the other parts, we keep the same as PLIP [23]. + +# 4 Experiments + +In this section, we first introduce the datasets used for training and evaluation in this paper, followed by the evaluation metrics and implementation details. + +# 4.1 Training Datasets + +Dataset for Knowledge Encoding. We carry out pathology knowledge encoding with PathKT, where 4,718 disease nodes with a total number of 50,470 attributes are extracted, shown in Table S1 in Supplementary Materials. + +Datasets for KEP Pretraining. Our dataset is composed of two parts. First, we collect the OpenPath data provided by PLIP [23] and obtain 138,874 image-text pairs after denoising and pre-processing. Second, we collect the Quilt1M [24] dataset, which gathers pathology image-text pairs from four public sources: PubMed articles, LAION [44], OpenPath [23], and Youtube videos. Since one of the downstream evaluation datasets Arch-PubMed [15] contains pathology images from PubMed articles, we remove this data source from Quilt1M to avoid data leaking. Considering that many images in Quilt1M are matched with multiple captions, we concatenate the captions related to the same image and finally obtain 576,608 image-text pairs for the Quilt1M dataset. + +# 4.2 Downstream Tasks + +We evaluate the pretrained models on three tasks, namely, retrieval, zero-shot patch classification, and zero-shot WSI tumor subtyping. The details for all evaluation datasets are exhibited in Table S2 in Supplementary Materials. + +Retrieval. This task involves cross-modal retrieval and disease retrieval. Cross-modal retrieval aims to retrieve the correct caption for a given image and vice versa. Disease retrieval, on the other hand, utilizes the disease names to retrieve captions or images with the same disease label, which is proposed to demonstrate the effectiveness of the knowledge encoder. For cross-modal retrieval, we follow PLIP [23] to split the ARCH [15] dataset into Arch-PubMed and Arch-book. In addition, we also gather image-text pairs from publicly available educational resources and curate a retrieval dataset, termed as PathPair, consisting of 9,358 pathology image-caption pairs with known 1676 disease labels. For disease retrieval, we utilize the captions, images, and their disease labels in PathPair. + +Zero-shot Patch Classification. This task involves zero-shot classification on patch-level pathology images. Specifically, at inference time, we randomly select one template from the 21 templates in CONCH [36] and one type synonym from the corresponding name list (exhibited in Supplementary Materials) to yield a text prompt for each type, e.g. a histopathological image of CLASSNAME. (template) + beast invasive carcinoma (synonym) $\rightarrow$ a histopathological image of beast invasive carcinoma. This process is repeated 100 times in every experiment. Following PLIP [23] and Quilt1M [24], we adopt these patch-level pathology image datasets: BACH [1], NCT-CRC-HE-100K [27], KatherColon [28], LC25000 [3], RenalCell [4], SICAP [47], SkinCancer [30], WSSS4LUAD [17]. Each dataset includes multiple types of H&E stained cell micrographs. + +Zero-shot WSI Tumor Subtyping. This task involves zero-shot tumor subtyping on pathology whole slide images of common and rare cancers. We follow MI-Zero [37] and CONCH [36] for evaluation. Specifically, we first divide WSI into $256 \times 256$ patches and then predict the class label of each image patch in a zero-shot manner. The tumor type of the whole slide is then provided by integrating Top-K predictions on patches. For this task, we also employ the templates and tumor synonyms from CONCH [36] to randomly generate 100 text prompts (listed in Supplementary materials) for each tumor subtype. (i) For common cancers, we follow existing research [36,37] and collect 525 WSIs with three tumor types, including 150 breast carcinoma (BRCA), 150 non-small cell lung cancer (NSCLC) histopathology slides, and 225 renal cell carcinoma (RCC) slides, from TCGA. Specifically, BRCA consists of two subtypes: invasive ductal carcinoma (IDC) and invasive lobular carcinoma (IDC). NSCLC contains lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC). RCC is divided into chromophobe renal cell carcinoma (CHRCC), clear-cell renal cell carcinoma (CCRCC), and papillary renal cell carcinoma (PRCC). Each tumor subtype has 75 WSIs. (ii) For rare cancers, we collected 41 WSIs of rare subtypes for breast cancer (6 for Intraductal papillary adenocarcinoma with invasion, 6 for Medullary carcinoma, NOS, 13 for Metaplastic carcinoma, NOS and + +16 for Mucinous adenocarcinoma), from TCGA and merge them with common breast cancers (75 for IDC and 75 for ILC). + +# 4.3 Evaluation Metrics. + +Retrieval. To evaluate the retrieval performance, we adopt the Recall@K metric, suggesting the ratio of correctly retrieved queries in Top-K retrieved samples. + +Zero-shot Patch Classification. For zero-shot classification tasks, we adopt the same metric as PLIP [23], namely, weighted F1 (wF1). We report the median, the first, and the third quartile (Q1, Q3) of wF1 across all text 100 prompts. + +Zero-shot WSI Tumor Subtyping. We adopt the same metric as MI-Zero [37], namely, balanced accuracy at Top-K pooling to measure the zero-shot tumor subtyping performance on WSIs. We also report the median, the first, and the third quartile (Q1, Q3) of the balanced accuracy across all 100 text prompts. + +# 4.4 Implementation Details + +Knowledge Encoder Pretraining. We adopt the architecture of PubMedBERT [16] to encode knowledge. The embedding dimension is set to 512. The temperature parameter $\tau$ is set to 0.04 in Eq. 2. The batch size is set to 256, including 32 disease entities with 8 instances per entity. + +Pathology Image-text Pretraining. To achieve a fair comparison with PLIP and Quilt1M, we utilize the same visual encoder (ViT-B-32) and initialization of the visual encoder (CLIP [41]), termed by KEP-32, and set the input image size to $224 \times 224$ . Additionally, we also develop a KEP variant named KEP-16 that is initialized by the visual weights of BiomedCLIP [56] (ViT-B-16). The batch size and learning rate are set to 256 and 1e-5, respectively. The temperature in Eq. 6 and Eq. 7 is set to 0.04 across all experiments. For OpenPath, we conduct the pathology VLP for 30 epochs, While for Quilt1M, 15 epochs are adopted. + +# 5 Results + +In this section, we show the results on the downstream tasks, to evaluate the effectiveness of proposed knowledge-enhanced representation learning. Note that for a fair comparison with the publicly available model released by the original authors, in all experiments, we separately train KEP on the OpenPath [23] dataset, and Quilt1M [24] dataset, and then report results on downstream tasks. + +# 5.1 Retrieval + +In this section, we evaluate the performance of retrieval, including cross-modal retrieval on three datasets and disease retrieval on PathPair. + +Cross-modal Retrieval. In Table 1, we demonstrate the results for different models pretrained on OpenPath and Quilt1M. It can be seen that KEP-32 + +Table 1: Performance comparison with PLIP and QuiltNet on three retrieval datasets. All models are pre-trained on the OpenPath and Quilt1M datasets, respectively. i2t and t2i denote image-to-text and text-to-image retrieval, respectively. Bold fonts suggest the best performance. + +
Training datasetTaskModelKnowledge EnhancementArch-PubMedArch-bookPathPair
Recall@10Recall@50Recall@10Recall@50Recall@10Recall@50
OpenPathi2tPLIPX0.0670.1850.1520.3930.0380.119
KEP-32✓0.0980.2830.1640.4040.0710.205
KEP-16✓0.1630.3980.2440.5370.1080.275
t2iPLIPX0.0670.1810.1650.4190.0470.133
KEP-32✓0.0850.2260.1480.3650.0610.171
KEP-16✓0.1380.3390.2380.5330.0930.247
Quilt1Mi2tQuiltNetX0.1390.3260.1880.4070.0650.166
KEP-32✓0.1400.3270.2400.5210.0840.221
KEP-16✓0.1960.4210.2820.5640.1080.254
t2iQuiltNetX0.1220.2930.2040.4290.0710.195
KEP-32✓0.1350.3260.2750.5680.1060.276
KEP-16✓0.1760.4040.3400.6210.1360.326
+ +Table 2: Performance comparison of disease retrieval. 12t and i2l denote label-to-text and image-to-label, respectively. Bold fonts suggest the best performance. + +
TaskMetricsOpenPathQuiltNetQuilt1M
PLIPKEP-32KEP-16KEP-32KEP-16
l2tRecall@100.4080.6930.6990.2110.6480.643
Recall@500.5360.8320.8350.3250.8120.808
i2lRecall@100.1140.1620.2230.1350.2260.259
Recall@500.2920.3810.4780.3220.4730.519
+ +pretrained on OpenPath outperforms PLIP on all datasets with respect to the image-to-text retrieval task, especially on Arch-PubMed and PathPair with more than $3\%$ boost on R10. As for the text-to-image task, KEP-32 also achieves better performance on the dataset of Arch-PubMed and PathPair. Furthermore, our KEP-16 model improves the retrieval performance by a large margin on all datasets across both retrieval tasks. Similar results can be concluded that KEP-32 pretrained on Quilt1M improves the performance on all datasets for both retrieval tasks, which demonstrates the effectiveness of knowledge guidance for visual-language pretraining on cross-modal retrieval. + +Disease Retrieval. Table 2 shows the performance of disease retrieval on the PathPair dataset for different models. It can be seen that although the scale of Quilt1M is 5 times OpenPath, models pretrained on OpenPath often outperforms that on Quilt1M for the label-to-text task. We conjecture this may be caused by the quality of captions in Quilt1M. Our approach KEP-32 outperforms PLIP and Quilt1M by a large margin on both label-to-text and image-to-label tasks, suggesting that the knowledge encoder contributes to paying attention to the key disease information when encoding captions, and thus improves the alignment between images and their disease labels. + +![](images/106ec365bde8846e5a5a324aad1d54d9c95098cd352118798569d40f671aee2b.jpg) +Fig. 5: The comparison of zero-shot patch classification between different models. The left and the right subfigures suggest pretraining on OpenPath and Quilt1M, respectively. The visual encoders of KEP-32 and KEP-16 are initialized by CLIP (ViT-B-32) and BiomedCLIP (ViT-B-16), respectively. The number of points for every box is 100, with each representing the performance of one text prompt. The upper, center, and lower line of each box denote the first, median, and third quartile of the distribution. + +![](images/c58d535787e4600fd839579c0cb7a46ccc4dc6c0cf507fd8a17ef3ee144ef61f.jpg) + +# 5.2 Zero-shot Patch Classification + +In this section, we evaluate the performance of zero-shot classification on patch-level pathology images from 8 datasets. We report the performance distribution (Fig. 5), where each point in the figure denotes the performance of one prompt. + +Comparison to PLIP. In Fig. 5 (left), we demonstrate the performance comparison with PLIP [23], by only pretraining on OpenPath image-text pairs. Compared with PLIP, KEP-32 achieves better zero-shot classification performance on all datasets, with a maximum enhancement of 0.137 on the BACH dataset. Moreover, with the medical-specific initialization of the visual encoder (BiomedCLIP), KEP-16 can further improve the performance significantly in most datasets. Note that, the boxes of KEP variants are generally shorter than that of PLIP, suggesting that our approach is less sensitive to the varying text prompts than PLIP, which can be demonstrated by the visualization of prompt embeddings, shown in Supplementary Materials. + +Comparison to QuiltNet. As shown in Fig. 5 (right), we exhibit the performance comparison between KEP and QuiltNet [24] pretrained on Quilt1M dataset. Similar to results on OpenPath, KEP-32 outperforms Quilt1M in seven out of eight datasets, especially for the BACH, KatherColon, and SkinCancer datasets. KEP-16 can further improve the weighted F1 score. Most boxes of KEP variants are also shorter than that of Quilt1M. The visualization of prompt embeddings, shown in Supplementary Materials, maintains that pathology knowledge can reduce the ambiguities of image-text alignment and thus improve the zero-shot classification performance. + +# 5.3 Zero-shot WSI Tumor Subtyping + +In this section, we evaluate the transfer ability of different models for tumor subtyping on common and rare cancers. + +![](images/c6f606d380643a708d710e074aa0046d5dd33b2b5b04e325468b78fe21baf948.jpg) +Fig. 6: The performance comparison of tumor subtyping on TCGA-BRCA (common), TCGA-NSCLC, and TCGA-RCC WSIs. The upper, center and lower line of each box denote the first, median, and third quartile of the performance distribution, respectively. The scattered points represent outliers. KEP-16 and KEP-CTP are trained on OpenPath with the visual encoder initialized by BiomedCLIP [56] and CTransPath [48], respectively. MI-Zero-Bio and MI-Zero-Pub are two variants of MI-Zero with different initialization of text encoder. Their visual encoders are initialized by CTransPath. In addition, MI-Zero adopts in-house pathology reports to pretrain their text encoders. + +Table 3: Performance comparison of tumor subtyping on six BRCA subtypes, including 2 common and 4 rare cancers. Bold fonts suggest the best performance. + +
ModelTraining datasetTop-1Top-10
CommonRareCommonRare
PLIPOpenPath0.1130.1150.1190.119
QuiltNetQuilt1M0.2780.0940.2790.102
KEP-16OpenPath0.3430.2170.3680.245
MI-Zero-BioIn-house & Web data0.3900.2140.4030.237
MI-Zero-PubIn-house & Web data0.3310.2820.3470.304
KEP-CTPOpenPath0.4430.2820.4320.301
+ +Common cancers. Fig. 6 shows the performance comparison of tumor subtyping on common cancers, including TCGA-BRCA (common), TCGA-NSCLC and TCGA-RCC WSIs. MI-Zero-Bio and MI-Zero-Pub are two variants of MI-Zero with different text encoders. The visual encoders of the two MI-Zero variants are initialized by CTransPath [48]. To achieve a fair comparison, we also train a KEP variant named KEP-CTP with the visual encoder initialized by CTransPath. Both KEP-16 and KEP-CTP are trained on the OpenPath dataset. It can be seen that our approach KEP-16 and KEP-CTP outperform PLIP and QuiltNet on all datasets. Note that, the comparisons between variants of MI-Zero and KEP-CTP are actually not fair for our method, as the MI-Zero variants have also been pretrained on massive in-house data – over 550k pathology reports from hospitals [37]. Yet, KEP-CTP still outperforms MI-Zero-Bio across all datasets and achieves comparable performance with MI-Zero-Pub on NSCLC and RCC. + +Rare cancers. Table 3 exhibits the median balanced accuracy of tumor subtyping on common and rare breast cancers. It can be seen that our method KEP-16 pretrained on OpenPath outperforms PLIP and QuiltNet by a large margin on both common and rare breast cancers. Moreover, KEP-CTP significantly outper + +Table 4: Experimental results of ablation study on model architecture and initialization. Arch-PubMed and PathPair suggest the text-to-image retrieval and the disease retrieval task, respectively. PMB and KB denote PubMedBERT and our pretrained pathology knowledge encoder, respectively. BCLIP denotes BiomedCLIP. Bold fonts and underline suggest the best and the second-best performance, respectively. + +
Visual Init.Text Init.Projection headKnowledge distillMetric lossKatherColon Median (Q1, Q3)Arch-PubMedPathPair
R10R50R1R5
ScratchPMB0.322 (0.272, 0.392)0.0370.1210.0560.113
CLIPPMB0.407 (0.358, 0.452)0.0810.2140.1610.316
CLIPPMB✓0.486 (0.448, 0.517)0.0730.2100.1610.305
BCLIPPMB✓0.553 (0.503, 0.591)0.1270.3470.1870.350
CLIPPMB✓✓(PMB)0.484 (0.435, 0.520)0.0630.1820.0340.089
CLIPKB✓Adasp0.530 (0.477, 0.600)0.0860.2320.2700.482
CLIPKB✓✓(KB)Adasp0.563 (0.505, 0.610)0.0850.2260.4090.618
CLIPKB✓✓(KB)Triplet0.531 (0.482, 0.580)0.0870.2270.2650.493
BCLIPKB✓✓(KB)Adasp0.580 (0.517, 0.631)0.1380.3390.4240.618
+ +forms MI-Zero variants on common breast cancers while achieving comparable performance on the rare ones. + +# 5.4 Ablation Study + +In this section, we explore the impact of model architecture, initialization, and other hyperparameters, as shown in Table 4 and Table S3 in Supplementary Materials. Comparing the performance between different models, we can draw the following observations: (i) visual projection head can facilitate the zero-shot classification performance; (ii) our pretrained knowledge encoder can enhance the performance across all tasks; (iii) the frozen knowledge branch can further improve the performance of zero-shot classification and disease retrieval tasks while the frozen PubMedBERT can not; (iv) a better initialization of visual encoder contributes to significant performance improvement on all tasks. (v) the Adasp loss is better than the Triplet loss for the knowledge encoding. + +# 6 Conclusion + +In this paper, we address the problem of knowledge-enhanced visual-language pretraining on computational pathology. We first curate a pathology knowledge tree that integrates the informative attributes of diseases requiring pathological diagnosis. We then propose a novel knowledge encoding approach based on metric learning to model structured pathological knowledge. With the guidance of the pretrained knowledge encoder, we conduct extensive visual-language pretraining on pathology image-caption pairs. To demonstrate the effectiveness of our approach, we evaluate pretrained VLP models on three downstream tasks, including retrieval, zero-shot classification on patch-level pathology images, and zero-shot tumor subtyping on pathology WSIs. Quantitative experimental results demonstrate that pathology knowledge can significantly improve the performance across different tasks. + +# Acknowledgements + +This work is supported by the National Key R&D Program of China (No. 2022ZD0160702) and China Postdoctoral Science Foundation (Certificate Number: 2023M741850). + +# References + +1. Aresta, G., Araújo, T., Kwok, S., Chennamsetty, S.S., Safwan, M., Alex, V., Marami, B., Prastawa, M., Chan, M., Donovan, M., et al.: Bach: Grand challenge on breast cancer histology images. Medical Image Analysis 56, 122-139 (2019) +2. Bodenreider, O.: The unified medical language system (uml's): integrating biomedical terminology. Nucleic Acids Research 32(suppl_1), D267-D270 (2004) +3. Borkowski, A.A., Bui, M.M., Thomas, L.B., Wilson, C.P., DeLand, L.A., Mastorides, S.M.: Lung and colon cancer histopathological image dataset (lc25000). arXiv preprint arXiv:1912.12142 (2019) +4. Brummer, O., Pölönen, P., Mustjoki, S., Brück, O.: Integrative analysis of histological textures and lymphocyte infiltration in renal cell carcinoma using deep learning. bioRxiv pp. 2022-08 (2022) +5. Campanella, G., Hanna, M.G., Geneslaw, L., Miraflor, A., Werneck Krauss Silva, V., Busam, K.J., Brogi, E., Reuter, V.E., Klimstra, D.S., Fuchs, T.J.: Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature Medicine 25(8), 1301-1309 (2019) +6. Chan, T.H., Cendra, F.J., Ma, L., Yin, G., Yu, L.: Histopathology whole slide image analysis with heterogeneous graph representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15661-15670 (2023) +7. Chauhan, G., Liao, R., Wells, W., Andreas, J., Wang, X., Berkowitz, S., Horng, S., Szolovits, P., Golland, P.: Joint modeling of chest radiographs and radiology reports for pulmonary edema assessment. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 529-539. Springer (2020) +8. Chen, C., Lu, M.Y., Williamson, D.F., Chen, T.Y., Schaumberg, A.J., Mahmood, F.: Fast and scalable search of whole-slide images via self-supervised deep learning. Nature Biomedical Engineering 6(12), 1420-1434 (2022) +9. Chen, R.J., Chen, C., Li, Y., Chen, T.Y., Trister, A.D., Krishnan, R.G., Mahmood, F.: Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16144-16155 (2022) +0. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: Uniter: Learning universal image-text representations (2019) +1. Chen, Y.C., Lu, C.S.: Rankmix: Data augmentation for weakly supervised learning of classifying whole slide images with diverse sizes and imbalanced categories. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23936-23945 (2023) +2. Chen, Z., Li, G., Wan, X.: Align, reason and learn: Enhancing medical vision-and-language pre-training with knowledge. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 5152-5161 (2022) +3. Chen, Z., Song, Y., Chang, T.H., Wan, X.: Generating radiology reports via memory-driven transformer. In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). pp. 1439-1449 (2020) + +14. Cui, H., Xu, Y., Li, W., Wang, L., Duh, H.: Collaborative learning of cross-channel clinical attention for radiotherapy-related esophageal fistula prediction from ct. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 212-220. Springer (2020) +15. Gamper, J., Rajpoot, N.: Multiple instance captioning: Learning representations from histopathology textbooks and articles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16549-16559 (2021) +16. Gu, Y., Tinn, R., Cheng, H., Lucas, M., Usuyama, N., Liu, X., Naumann, T., Gao, J., Poon, H.: Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Computing for Healthcare (HEALTH) 3(1), 1-23 (2021) +17. Han, C., Pan, X., Yan, L., Lin, H., Li, B., Yao, S., Lv, S., Shi, Z., Mai, J., Lin, J., et al.: Wsss4luad: Grand challenge on weakly-supervised tissue semantic segmentation for lung adenocarcinoma. arXiv preprint arXiv:2204.06455 (2022) +18. Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737 (2017) +19. Hu, S., Chen, L., Wu, P., Li, H., Yan, J., Tao, D.: St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In: European Conference on Computer Vision. pp. 533-549. Springer (2022) +20. Huang, S.C., Shen, L., Lungren, M.P., Yeung, S.: Gloria: A multimodal global-local representation learning framework for label-efficient medical image recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3942-3951 (2021), official Implementation: https://github.com/marshuang80/gloria +21. Huang, X., Fang, Y., Lu, M., Yan, F., Yang, J., Xu, Y.: Dual-ray net: automatic diagnosis of thoracic diseases using frontal and lateral chest x-rays. Journal of Medical Imaging and Health Informatics 10(2), 348-355 (2020) +22. Huang, Y., Zhao, W., Wang, S., Fu, Y., Jiang, Y., Yu, L.: Conslide: Asynchronous hierarchical interaction transformer with breakup-reorganize rehearsal for continual whole slide image analysis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 21349-21360 (2023) +23. Huang, Z., Bianchi, F., Yuksekgonul, M., Montine, T.J., Zou, J.: A visual-language foundation model for pathology image analysis using medical twitter. Nature medicine 29(9), 2307-2316 (2023) +24. Ikezogwo, W., Seyfioglu, S., Ghezloo, F., Geva, D., Sheikh Mohammed, F., Anand, P.K., Krishna, R., Shapiro, L.: Quilt-1m: One million image-text pairs for histopathology. Advances in Neural Information Processing Systems 36 (2024) +25. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning. pp. 4904-4916. PMLR (2021) +26. Kang, M., Song, H., Park, S., Yoo, D., Pereira, S.: Benchmarking self-supervised learning on diverse pathology datasets. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3344-3354 (2023) +27. Kather, J.N., Halama, N., Marx, A.: 100,000 histological images of human colorectal cancer and healthy tissue. Zenodo10 5281 (2018) +28. Kather, J.N., Krisam, J., Charoentong, P., Luedde, T., Herpel, E., Weis, C.A., Gaiser, T., Marx, A., Valous, N.A., Ferber, D., et al.: Predicting survival from colorectal cancer histology slides using deep learning: A retrospective multicenter study. PLoS Medicine 16(1), e1002730 (2019) + +29. Kim, W., Son, B., Kim, I.: Vilt: Vision-and-language transformer without convolution or region supervision. In: International Conference on Machine Learning. pp. 5583-5594. PMLR (2021) +30. Kriegsmann, K., Lobers, F., Zgorzelski, C., Kriegsmann, J., Janssen, C., Meliss, R.R., Muley, T., Sack, U., Steinbuss, G., Kriegsmann, M.: Deep learning for the detection of anatomical tissue structures and neoplasms of the skin on scanned histopathological tissue sections. Frontiers in Oncology 12, 1022967 (2022) +31. Kundra, R., Zhang, H., Sheridan, R., Sirlintrapun, S.J., Wang, A., Ochoa, A., Wilson, M., Gross, B., Sun, Y., Madupuri, R., et al.: Oncotree: a cancer classification system for precision oncology. JCO clinical cancer informatics 5, 221-230 (2021) +32. Li, H., Zhu, C., Zhang, Y., Sun, Y., Shui, Z., Kuang, W., Zheng, S., Yang, L.: Task-specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7454-7463 (2023) +33. Li, M., Cai, W., Verspoor, K., Pan, S., Liang, X., Chang, X.: Cross-modal clinical graph transformer for ophthalmic report generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20656-20665 (2022) +34. Lin, T., Yu, Z., Hu, H., Xu, Y., Chen, C.W.: Interventional bag multi-instance learning on whole-slide pathological images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19830-19839 (2023) +35. Lin, W., Zhao, Z., Zhang, X., Wu, C., Zhang, Y., Wang, Y., Xie, W.: Pmc-clip: Contrastive language-image pre-training using biomedical documents. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) (2023) +36. Lu, M.Y., Chen, B., Williamson, D.F., Chen, R.J., Liang, I., Ding, T., Jaume, G., Odintsov, I., Zhang, A., Le, L.P., et al.: Towards a visual-language foundation model for computational pathology. arXiv preprint arXiv:2307.12914 (2023) +37. Lu, M.Y., Chen, B., Zhang, A., Williamson, D.F., Chen, R.J., Ding, T., Le, L.P., Chuang, Y.S., Mahmood, F.: Visual language pretrained multiple instance zero-shot transfer for histopathology images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19764-19775 (2023) +38. Lu, M.Y., Williamson, D.F., Chen, T.Y., Chen, R.J., Barbieri, M., Mahmood, F.: Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering 5(6), 555-570 (2021) +39. Müller, P., Kaissis, G., Zou, C., Rueckert, D.: Joint learning of localized representations from medical images and reports. In: European Conference on Computer Vision. pp. 685-701. Springer (2022) +40. Neumann, M., King, D., Beltagy, I., Ammar, W.: ScispaCy: Fast and robust models for biomedical natural language processing. In: Demner-Fushman, D., Cohen, K.B., Ananiadou, S., Tsujii, J. (eds.) Proceedings of the 18th BioNLP Workshop and Shared Task. pp. 319-327. Association for Computational Linguistics, Florence, Italy (Aug 2019). https://doi.org/10.18653/v1/W19-5034, https://aclanthology.org/W19-5034 +41. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021) +42. Roberts, R.J.: Pubmed central: The genbank of the published literature (2001) +43. Rorke, L.B.: Pathologic diagnosis as the gold standard (1997) + +44. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278–25294 (2022) +45. Shaban, M., Awan, R., Fraz, M.M., Azam, A., Tsang, Y.W., Snead, D., Rajpoot, N.M.: Context-aware convolutional neural network for grading of colorectal cancer histology images. IEEE Transactions on Medical Imaging 39(7), 2395-2405 (2020) +46. Shao, Z., Bian, H., Chen, Y., Wang, Y., Zhang, J., Ji, X., et al.: Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in Neural Information Processing Systems 34, 2136-2147 (2021) +47. Silva-Rodriguez, J., Colomer, A., Sales, M.A., Molina, R., Naranjo, V.: Going deeper through the gleason scoring scale: An automatic end-to-end system for histology prostate grading and cribriform pattern detection. Computer Methods and Programs in Biomedicine 195, 105637 (2020) +48. Wang, X., Yang, S., Zhang, J., Wang, M., Zhang, J., Yang, W., Huang, J., Han, X.: Transformer-based unsupervised contrastive learning for histopathological image classification. Medical Image Analysis 81, 102559 (2022) +49. Wu, C., Zhang, X., Zhang, Y., Wang, Y., Xie, W.: Medklip: Medical knowledge enhanced language-image pre-training. medRxiv pp. 2023-01 (2023) +50. Xie, X., Niu, J., Liu, X., Chen, Z., Tang, S., Yu, S.: A survey on incorporating domain knowledge into deep learning for medical image analysis. Medical Image Analysis 69, 101985 (2021) +51. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettle-moyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 6787-6800 (2021) +52. Xuan, H., Stylianou, A., Liu, X., Pless, R.: Hard negative examples are hard, but useful. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 126-142. Springer (2020) +53. Yang, S., Wu, X., Ge, S., Zhou, S.K., Xiao, L.: Knowledge matters: Chest radiology report generation with general and specific knowledge. Medical Image Analysis 80, 102510 (2022) +54. Yu, F., Tang, J., Yin, W., Sun, Y., Tian, H., Wu, H., Wang, H.: Ernie-vil: Knowledge enhanced vision-language representations through scene graphs. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 3208-3216 (2021) +55. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 (2022) +56. Zhang, S., Xu, Y., Usuyama, N., Bagga, J., Tinn, R., Preston, S., Rao, R., Wei, M., Valluri, N., Wong, C., et al.: Large-scale domain-specific pretraining for biomedical vision-language processing. arXiv preprint arXiv:2303.00915 (2023) +57. Zhang, X., Wu, C., Zhang, Y., Xie, W., Wang, Y.: Knowledge-enhanced visual-language pre-training on chest radiology images. Nature Communications 14(1), 4542 (2023) +58. Zhang, Y., Jiang, H., Miura, Y., Manning, C.D., Langlotz, C.P.: Contrastive learning of medical visual representations from paired images and text. In: Machine Learning for Healthcare (2022), highest Starred Implementation: https://github.com/edreisMD/ConVIRT-pytorch +59. Zhou, X., Cheng, Z., Gu, M., Chang, F.: Lirnet: Local integral regression network for both strongly and weakly supervised nuclei detection. In: 2020 IEEE Interna + +tional Conference on Bioinformatics and Biomedicine (BIBM). pp. 945-951. IEEE (2020) +60. Zhou, X., Zhong, Y., Cheng, Z., Liang, F., Ma, L.: Adaptive sparse pairwise loss for object re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19691-19701 (2023) \ No newline at end of file diff --git a/2024/Knowledge-enhanced Visual-Language Pretraining for Computational Pathology/images.zip b/2024/Knowledge-enhanced Visual-Language Pretraining for Computational Pathology/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d157db5afe1268af2fb09b69709d1de21ac03361 --- /dev/null +++ b/2024/Knowledge-enhanced Visual-Language Pretraining for Computational Pathology/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f9dacf0defd5806d20b43902f910d1ed8cf0264d15f0f5511cd4252e56d34a6 +size 476550 diff --git a/2024/Knowledge-enhanced Visual-Language Pretraining for Computational Pathology/layout.json b/2024/Knowledge-enhanced Visual-Language Pretraining for Computational Pathology/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..28fa6adc911122ce256279e68bd380c5724d4ed7 --- /dev/null +++ b/2024/Knowledge-enhanced Visual-Language Pretraining for Computational Pathology/layout.json @@ -0,0 +1,8400 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 162, + 111, + 451, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 111, + 451, + 148 + ], + "spans": [ + { + "bbox": [ + 162, + 111, + 451, + 148 + ], + "type": "text", + "content": "Knowledge-enhanced Visual-Language Pretraining for Computational Pathology" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 197, + 167, + 414, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 167, + 414, + 195 + ], + "spans": [ + { + "bbox": [ + 197, + 167, + 414, + 195 + ], + "type": "text", + "content": "Xiao Zhou1, Xiaoman Zhang1,2, Chaoyi Wu1,2, Ya Zhang1,2, Weidi Xie1,2, Yanfeng Wang1,2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 214, + 203, + 399, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 203, + 399, + 214 + ], + "spans": [ + { + "bbox": [ + 214, + 203, + 399, + 214 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 214, + 203, + 399, + 214 + ], + "type": "text", + "content": " Shanghai Artificial Intelligence Laboratory" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 239, + 215, + 374, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 215, + 374, + 226 + ], + "spans": [ + { + "bbox": [ + 239, + 215, + 374, + 226 + ], + "type": "text", + "content": "2 Shanghai Jiao Tong University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 223, + 227, + 391, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 227, + 391, + 236 + ], + "spans": [ + { + "bbox": [ + 223, + 227, + 391, + 236 + ], + "type": "text", + "content": "https://github.com/MAGIC-AI4Med/KEP" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 157, + 238, + 454, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 238, + 454, + 248 + ], + "spans": [ + { + "bbox": [ + 157, + 238, + 454, + 248 + ], + "type": "text", + "content": "zhouxiao@pjlab.org.cn, {xm99sjtu, wtzxxxxwy02, ya_zhang, weidi," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 247, + 249, + 366, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 249, + 366, + 258 + ], + "spans": [ + { + "bbox": [ + 247, + 249, + 366, + 258 + ], + "type": "text", + "content": "wangyanfeng} @sjtu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 289, + 457, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 289, + 457, + 465 + ], + "spans": [ + { + "bbox": [ + 159, + 289, + 457, + 465 + ], + "type": "text", + "content": "Abstract. In this paper, we consider the problem of visual representation learning for computational pathology, by exploiting large-scale image-text pairs gathered from public resources, along with the domain-specific knowledge in pathology. Specifically, we make the following contributions: (i) We curate a pathology knowledge tree that consists of 50,470 informative attributes for 4,718 diseases requiring pathology diagnosis from 32 human tissues. To our knowledge, this is the first comprehensive structured pathology knowledge base; (ii) We develop a knowledge-enhanced visual-language pretraining approach, where we first project pathology-specific knowledge into latent embedding space via a language model, and use it to guide the visual representation learning; (iii) We conduct thorough experiments to validate the effectiveness of our proposed components, demonstrating significant performance improvement on various downstream tasks, including cross-modal retrieval, zero-shot classification on pathology patches, and zero-shot tumor subtyping on whole slide images (WSIs)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 160, + 475, + 423, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 475, + 423, + 487 + ], + "spans": [ + { + "bbox": [ + 160, + 475, + 423, + 487 + ], + "type": "text", + "content": "Keywords: Pathology knowledge " + }, + { + "bbox": [ + 160, + 475, + 423, + 487 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 475, + 423, + 487 + ], + "type": "text", + "content": " Visual-language pretraining" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 508, + 230, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 508, + 230, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 230, + 521 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 641 + ], + "type": "text", + "content": "Pathology diagnosis is currently the golden standard for examining various diseases in clinical applications, especially in the diagnosis of neoplasm [43]. In the last decade, the prosperity of deep learning on computer vision has led to the rapid development of computational pathology, for example, approaches with supervised learning [6,22,34,45,46], and weakly supervised learning [5,11,32,38,59]. Despite being promising, these approaches have been fundamentally limited by the scale of costly labels. Alternatively, self-supervised pretraining on numerous unlabeled pathological images [8,9,26,48] has attracted unprecedented attention, yet it still requires supervised fine-tuning for downstream deployments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "In the recent literature [25,41], studies on the multimodal foundation model have demonstrated conspicuous improvement in downstream zero-shot tasks, by" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 196, + 114, + 416, + 243 + ], + "blocks": [ + { + "bbox": [ + 196, + 114, + 416, + 243 + ], + "lines": [ + { + "bbox": [ + 196, + 114, + 416, + 243 + ], + "spans": [ + { + "bbox": [ + 196, + 114, + 416, + 243 + ], + "type": "image", + "image_path": "a239cf8b1e899a359e14ef525856e627b734f565f07d6dc364c42af627c3eec6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 251, + 482, + 329 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 329 + ], + "type": "text", + "content": "Fig. 1: Knowledge-enhanced pathology image-text alignment. The short caption of a pathology image crawled from public websites is typically unstructured and with varying granularities, which introduces noticeable ambiguities for image-text alignment. While the implicit structures and correlations between different image-caption pairs could be constructed by explicit disease attributes (dashed boxes), which can be well-aligned by a pathology knowledge tree. LUSC and LUAD suggest lung squamous cell carcinoma, and lung adenocarcinoma, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 365, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 485 + ], + "type": "text", + "content": "simply training to align visual and language embedding space on massive image-text pairs crawled from public websites. In contrast to computer vision [19,25,41], representation learning in pathology requires tremendous expertise and domain knowledge. An ideal training corpus would thus consist of well-structured medical reports from the hospital, however, it is often extremely difficult to acquire, due to privacy concerns. Alternatively, recent works [23, 24, 36, 37] propose to gather large-scale image-caption pairs from public resources (PubMed [42] papers, Twitter, Youtube videos). Compared to discrete image labels, image captions from medical reports and academic articles can potentially provide more valuable medical information without manual annotation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 491, + 482, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 600 + ], + "type": "text", + "content": "For existing works that adopt simple contrastive learning on pathology image and caption pairs, they suffer from the following challenges: First, the short captions from these web-crawled image-text pairs are typically noisy, unstructured, and lack domain knowledge (Fig. 1), which can be sub-optimal for constructing high-quality pathology-specific visual representations; Second, the varying granularities of the free text from captions will inevitably introduce ambiguities when aligning with images, thus causing the model to be highly sensitive to the used text prompts during the inference stage [36]. Thus, effectively harnessing the potential of such web-crawled data remains to be a challenge." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "In this paper, to tackle the above challenges, we anticipate that introducing pathology knowledge is of great significance to make up for the deficiency of short image captions. To this end, we make the following contributions: (i) We curate a pathology knowledge tree, PathKT, by collecting 50,470 informative pathological attributes of 4,718 diseases in 32 tissues from publicly available educational" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 224, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 224, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 224, + 100 + ], + "type": "text", + "content": "X. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 283 + ], + "type": "text", + "content": "resources and OncoTree" + }, + { + "bbox": [ + 130, + 115, + 482, + 283 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 130, + 115, + 482, + 283 + ], + "type": "text", + "content": " [31]. (ii) We propose a novel knowledge encoder pretraining approach, that projects the attributes of each disease from PathKT into latent embedding space, where the attributes of the same disease, including disease synonyms, definitions, histology, and cytology features, share similar representations. (iii) We develop a knowledge-enhanced pretraining (KEP) approach to align pathology visual-language representations, which freezes the knowledge encoder and continuously injects domain-specific knowledge into the image-text embedding space. To demonstrate the effectiveness of our proposed approach, we conduct thorough experiments on three downstream tasks, including retrieval on three pathological image-caption datasets, zero-shot patch classification on eight patch-level pathology image datasets, and zero-shot WSI tumor subtyping on three datasets from The Cancer Genome Atlas (TCGA)" + }, + { + "bbox": [ + 130, + 115, + 482, + 283 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 130, + 115, + 482, + 283 + ], + "type": "text", + "content": ". Quantitative experiments suggest that knowledge guidance can significantly enhance the performance across different tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 303, + 234, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 234, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 234, + 316 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 328, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 482, + 449 + ], + "type": "text", + "content": "Vision-Language Pretraining (VLP). Current vision-language pretraining approaches are typically categorized into two groups. The first group, referred to as two-stream approaches [41, 51], involves using two separate encoders to extract features for visual and textual data, respectively. The second group, known as single-stream methods [10, 29], utilizes a cross-modal fusion encoder to enhance interactions between vision and text features. In the field of medical VLP, existing methods mainly adopt the two-stream approach [7, 12, 20, 35, 39, 56, 58]. Despite the valuable contributions of these methods, the reliance on data-driven representation learning restricts the utilization of systematic and structured medical knowledge, leading to less than professional diagnosis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 452, + 482, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 452, + 482, + 595 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 482, + 595 + ], + "type": "text", + "content": "VLP in Computational Pathology. In the recent literature, Huang et al. proposed a pathology VLP model named PLIP [23] and released OpenPath that contains 200K image-caption pairs from the social media of Twitter and other public resources. To extend the pathology data scale, Ikezogwo et al. curate Quilt1m [24] with over one million histopathology image-text pairs by capturing keyframes and speech from YouTube videos. In addition to evaluating the zero-shot classification on patch-level pathology images, Lu et al. propose MI-Zero [37] and CONCH [36] to extend the transfer ability of pathology VLP models on gigapixel whole slide images (WSIs). In contrast to these existing work, that finetunes a CLIP [41] or CoCa [55] on pathology image-caption pairs, we propose to first train a pathology-specific knowledge encoder, and use it to guide visual-language representation learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 598, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 598, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 598, + 482, + 635 + ], + "type": "text", + "content": "Medical Knowledge-enhanced Learning. In the medical community, leveraging external medical knowledge to enhance deep learning models has become an increasingly important topic [49, 50, 57]. Generally, existing approaches can" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "KEP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 641, + 249, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 641, + 249, + 654 + ], + "spans": [ + { + "bbox": [ + 133, + 641, + 249, + 654 + ], + "type": "text", + "content": "3 https://oncotree.info/" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 654, + 287, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 654, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 654, + 287, + 665 + ], + "type": "text", + "content": "4 https://portal.gdc.cancer.gov/" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 480, + 225 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 480, + 225 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 480, + 225 + ], + "type": "image", + "image_path": "0b346c9114d5f910f02e917024335522a54969e1e35a14d0aabf055f0a15b2e9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 234, + 482, + 323 + ], + "lines": [ + { + "bbox": [ + 130, + 234, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 482, + 323 + ], + "type": "text", + "content": "Fig. 2: The construction of pathology knowledge tree. OncoTree is adopted as the base architecture to construct the PathKT. The tissue types, disease entities, and attributes are first extracted from web-crawled pathological descriptions, where cancers are then matched to OncoTree based on their tissue types and tumor types/subtypes using UMLS CUIs. Moreover, non-tumor diseases are added to the knowledge tree according to their tissue types. Finally, the pathology knowledge tree integrates 4718 diseases from 32 tissues, with each disease containing various synonyms, definitions, and histological and cytological features. CSCLC in this figure suggests combined small cell lung cancer." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 345, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 482, + 407 + ], + "type": "text", + "content": "be categorized based on the ways of using medical knowledge: model-based approaches [14, 21] adopts the prior knowledge of radiology or diagnosis summarized by doctors to design algorithms; and input-based methods [13, 33, 53, 54] directly exploit knowledge as the external input to guide representation training. However, most of these works are focused on the analysis of chest X-rays." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 422, + 208, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 422, + 208, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 422, + 208, + 435 + ], + "type": "text", + "content": "3 Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 447, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 544 + ], + "type": "text", + "content": "Our primary goal is to leverage structured pathology knowledge to enhance visual-language representation learning. To start with, we construct a **Pathology Knowledge Tree**, termed as **PathKT**, that consists of 50,470 informative attributes of 4718 diseases from 32 human tissues (Sec. 3.1). We then train a knowledge encoder that projects the structured pathology knowledge into an embedding space (Sec. 3.2). We further employ the knowledge encoder to guide visual-language pretraining for computational pathology, termed as KEP (Knowledge-Enhanced Pre-training, Sec. 3.3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 559, + 271, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 271, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 271, + 571 + ], + "type": "text", + "content": "3.1 PathKT Construction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 578, + 481, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 481, + 603 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 481, + 603 + ], + "type": "text", + "content": "Here, we detail the procedure for building up a pathology knowledge tree with various online sources, which will be further used for knowledge encoding." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Knowledge Source. We collect pathology-specific knowledge from publicly available educational resources, such as text books, professional websites, and structured databases (OncoTree [31]). Specifically, we extract pathological descriptions of 884 tumor subtypes from OncoTree and 4360 diseases from text books and professional websites, including all domain knowledge required for" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 224, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 224, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 224, + 100 + ], + "type": "text", + "content": "X. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "diagnosis in clinical scenarios, i.e., disease name/ synonyms, definitions, histological features, and cytological features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 371 + ], + "type": "text", + "content": "Knowledge Tree Construction. We structure these knowledge sources into a knowledge tree by expanding the OncoTree, as shown in Fig. 2. OncoTree is a tree-structure cancer classification system [31], which consists of 884 tumor subtypes from 32 tissues, with each type of tumor linked to a Concept Unique Identifier (CUI) from Unified Medical Language System (UMLS) [2]. Specifically, we first extract tissue types, disease entities, and disease attributes from pathological descriptions of 4360 diseases, in which 168 cancers are found overlapped with OncoTree by using SciSpacy package [40] to link the UMLS CUIs. The rest 4192 diseases are then added to the knowledge tree according to their tissue types. After deduplication and noise reduction (358 diseases without any informative descriptions are deleted), all diseases are organized into the corresponding tissues. The final histopathology knowledge tree contains 4718 diseases from 32 tissues. The attributes of each disease node are constructed by a varying number of synonyms, definitions, and histological and cytological descriptions. The statistics of PathKT is shown in Table S1 and Fig. S1 in Supplementary Materials. In the final PathKT, for instance, the tissue node of the lung connects all lung diseases, including combined small cell lung cancer (CSCLC), lung squamous cell carcinoma (LUSC), and lung adenocarcinoma (LUAD), etc. The LUSC node connects four kinds of attributes, shown in Fig. 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 391, + 321, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 321, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 321, + 403 + ], + "type": "text", + "content": "3.2 Pathology Knowledge Encoding" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 413, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 545 + ], + "type": "text", + "content": "In this section, we describe details for projecting tree-structure pathological knowledge into a latent embedding space, by training a knowledge encoder. Specifically, we align different disease entities with their corresponding free-text attributes via metric learning, such that, the synonyms, definitions, and corresponding histological/cytological features are close in the embedding space. As a consequence, the model can link the pathology images to their implicit disease labels during the visual-language pertaining, since the free-form text descriptions in image-text pairs may contain disease attributes, such as pathological features and disease definitions, which have already been aligned with disease names/synonyms. Therefore, the alignment of pathology knowledge can help the model link to the disease entities for better performance during diagnosis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "text", + "content": "Problem Setting. Given a set of disease entities with their corresponding attributes, " + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(d_1,\\mathbf{a}_1),\\ldots ,(d_n,\\mathbf{a}_n)\\}" + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "text", + "content": "-th disease entity, and " + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i = \\{a_i^1,\\dots ,a_i^k\\}" + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "text", + "content": " refer to the associated " + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "text", + "content": " attributes, both disease and attributes are represented in the format of natural language. Note that, for different disease entities, " + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "text", + "content": " also varies, our goal here is to train a model that satisfies:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 200, + 618, + 481, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 618, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 200, + 618, + 481, + 633 + ], + "type": "interline_equation", + "content": "\\operatorname {s i m} \\left(\\Phi_ {\\mathrm {k}} \\left(a _ {i} ^ {p}\\right), \\Phi_ {\\mathrm {k}} \\left(a _ {i} ^ {q}\\right)\\right) \\gg \\operatorname {s i m} \\left(\\Phi_ {\\mathrm {k}} \\left(a _ {i} ^ {p}\\right), \\Phi_ {\\mathrm {k}} \\left(a _ {j} ^ {t}\\right)\\right), i \\neq j, \\tag {1}", + "image_path": "9658b7ff5713b8b0557ca158ca3b01e8e2feff4079c68bf1e3dbc038889a0019.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 640, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 640, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 640, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\mathbf{k}}(\\cdot)" + }, + { + "bbox": [ + 130, + 640, + 482, + 667 + ], + "type": "text", + "content": " denotes the knowledge encoder, " + }, + { + "bbox": [ + 130, + 640, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 130, + 640, + 482, + 667 + ], + "type": "text", + "content": " refers to the similarity, " + }, + { + "bbox": [ + 130, + 640, + 482, + 667 + ], + "type": "inline_equation", + "content": "a_i^p,a_i^q" + }, + { + "bbox": [ + 130, + 640, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 640, + 482, + 667 + ], + "type": "inline_equation", + "content": "a_j^t" + }, + { + "bbox": [ + 130, + 640, + 482, + 667 + ], + "type": "text", + "content": " refer to the randomly sampled attributes from the " + }, + { + "bbox": [ + 130, + 640, + 482, + 667 + ], + "type": "inline_equation", + "content": "i,j" + }, + { + "bbox": [ + 130, + 640, + 482, + 667 + ], + "type": "text", + "content": "-th disease entity." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "KEP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 476, + 227 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 476, + 227 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 476, + 227 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 476, + 227 + ], + "type": "image", + "image_path": "dd3628686907ec3b9bdb5e0fb419731044620b613a54f42695713b29f2476bf3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 234, + 482, + 344 + ], + "lines": [ + { + "bbox": [ + 130, + 234, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 482, + 344 + ], + "type": "text", + "content": "Fig. 3: Knowledge encoder pretraining based on metric learning. " + }, + { + "bbox": [ + 130, + 234, + 482, + 344 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 234, + 482, + 344 + ], + "type": "text", + "content": " disease entities and each with " + }, + { + "bbox": [ + 130, + 234, + 482, + 344 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 234, + 482, + 344 + ], + "type": "text", + "content": " attributes, including disease synonyms, definitions, cytology and pathology features, construct a mini-batch (left part of the figure), which are fed to a knowledge encoder for pretraining. In the embedding space (right part of the figure), the markers in different shapes represent the embeddings of attributes of different diseases. " + }, + { + "bbox": [ + 130, + 234, + 482, + 344 + ], + "type": "inline_equation", + "content": "S_{i}^{+}" + }, + { + "bbox": [ + 130, + 234, + 482, + 344 + ], + "type": "text", + "content": "suggests the max-min positive attribute similarity within the " + }, + { + "bbox": [ + 130, + 234, + 482, + 344 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 234, + 482, + 344 + ], + "type": "text", + "content": "-th disease, while " + }, + { + "bbox": [ + 130, + 234, + 482, + 344 + ], + "type": "inline_equation", + "content": "S_{i}^{-}" + }, + { + "bbox": [ + 130, + 234, + 482, + 344 + ], + "type": "text", + "content": "denotes the maximal attribute similarity between the " + }, + { + "bbox": [ + 130, + 234, + 482, + 344 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 234, + 482, + 344 + ], + "type": "text", + "content": "-th disease and other diseases. The goal of metric learning is to increase " + }, + { + "bbox": [ + 130, + 234, + 482, + 344 + ], + "type": "inline_equation", + "content": "S_{i}^{+}" + }, + { + "bbox": [ + 130, + 234, + 482, + 344 + ], + "type": "text", + "content": "and meanwhile decreasing " + }, + { + "bbox": [ + 130, + 234, + 482, + 344 + ], + "type": "inline_equation", + "content": "S_{i}^{-}" + }, + { + "bbox": [ + 130, + 234, + 482, + 344 + ], + "type": "text", + "content": ". The purple-dashed arrow and circle denote the minimal positive attribute similarity in the second class and the hypersphere it spans." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 369, + 480, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 480, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 480, + 393 + ], + "type": "text", + "content": "Intuitively, the knowledge encoder enables the attributes of the same disease to be pulled together, while attributes from different diseases are pushed apart." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 396, + 482, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 480 + ], + "type": "text", + "content": "Training. In order to achieve the objective defined in Eq. 1, metric learning is exploited to construct an embedding space where the representations of intra/inter-class instances are clustered/separated. In specific, given a mini-batch that contains " + }, + { + "bbox": [ + 130, + 396, + 482, + 480 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 396, + 482, + 480 + ], + "type": "text", + "content": " random diseases and each with " + }, + { + "bbox": [ + 130, + 396, + 482, + 480 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 396, + 482, + 480 + ], + "type": "text", + "content": " attributes, we denote the normalized embedding for the " + }, + { + "bbox": [ + 130, + 396, + 482, + 480 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 396, + 482, + 480 + ], + "type": "text", + "content": "-th attribute of " + }, + { + "bbox": [ + 130, + 396, + 482, + 480 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 396, + 482, + 480 + ], + "type": "text", + "content": "-th disease as: " + }, + { + "bbox": [ + 130, + 396, + 482, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_p^i = \\varPhi_{\\mathrm{k}}(a_i^p) / \\| \\varPhi_{\\mathrm{k}}(a_i^p)\\|" + }, + { + "bbox": [ + 130, + 396, + 482, + 480 + ], + "type": "text", + "content": ". We adopt the recently proposed AdaSP loss [60], which finds out a max-min positive similarity (Fig. 3) and then shapes a loss with the maximal negative similarity:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 228, + 488, + 481, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 488, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 228, + 488, + 481, + 515 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m e t i c}} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\log \\left(1 + e ^ {\\left(S _ {i} ^ {-} - S _ {i} ^ {+}\\right) / \\tau}\\right), \\tag {2}", + "image_path": "c6c7a6b40a4bd9e720e4115cd94bb17e7846ea81e3230d8b288e7ab9abe84bd3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 525, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 550 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 525, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 525, + 482, + 550 + ], + "type": "text", + "content": " is a temperature parameter. " + }, + { + "bbox": [ + 130, + 525, + 482, + 550 + ], + "type": "inline_equation", + "content": "S_{i}^{+}" + }, + { + "bbox": [ + 130, + 525, + 482, + 550 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 525, + 482, + 550 + ], + "type": "inline_equation", + "content": "S_{i}^{-}" + }, + { + "bbox": [ + 130, + 525, + 482, + 550 + ], + "type": "text", + "content": " denote the max-min positive and the maximal negative similarity, which can be computed by the soft version:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 194, + 558, + 481, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 558, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 194, + 558, + 481, + 600 + ], + "type": "interline_equation", + "content": "S _ {i} ^ {+} = \\max _ {p} \\min _ {q} \\left\\langle \\mathbf {z} _ {p} ^ {i}, \\mathbf {z} _ {q} ^ {i} \\right\\rangle \\approx \\tau \\log \\left(\\sum_ {p = 1} ^ {k} \\frac {1}{\\sum_ {q = 1} ^ {k} e ^ {- \\frac {\\left\\langle \\mathbf {z} _ {p} ^ {i} , \\mathbf {z} _ {q} ^ {i} \\right\\rangle}{\\tau}}}\\right) \\tag {3}", + "image_path": "70f6b788c57154783f6121dade2912a4fb64b7142b75f814289127d34c6e130f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 197, + 603, + 481, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 603, + 481, + 639 + ], + "spans": [ + { + "bbox": [ + 197, + 603, + 481, + 639 + ], + "type": "interline_equation", + "content": "S _ {i} ^ {-} = \\max _ {j, p, q} \\left\\langle \\mathbf {z} _ {p} ^ {i}, \\mathbf {z} _ {q} ^ {j} \\right\\rangle \\approx \\tau \\log \\left(\\sum_ {p = 1} ^ {k} \\sum_ {j = 1, j \\neq i} ^ {n} \\sum_ {q = 1} ^ {k} e ^ {\\frac {\\left\\langle \\mathbf {z} _ {p} ^ {i} , \\mathbf {z} _ {q} ^ {j} \\right\\rangle}{\\tau}}\\right) \\tag {4}", + "image_path": "e5c1933f23bc431df18cbf7181ecd9f41733391a7dd10968cd60f1472ceb7157.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\langle \\cdot \\rangle" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " represents the cosine similarity, the details about the soft version can be found in Supplementary Materials." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 224, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 224, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 224, + 100 + ], + "type": "text", + "content": "X. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "Discussion. The conventional triplet loss [18] with batch hard mining strategy is widely adopted in metric learning, while it is not suitable in our case due to the following two reasons: First, in the pathology knowledge, each disease entity is associated with at most four types of attributes, the text descriptions of different attribute types might reveal significant divergence, which causes a large intra-class variation, marked by the purple dashed circle in Fig. 3, during training. Second, the fine-grained diseases might share high similarities, such as tumor subtypes: breast invasive ductal carcinoma and breast invasive lobular carcinoma, which cause low inter-class variations during training. When a mini-batch meets these two conditions, traditional triplet loss that enforces the instance similarities could typically produce the bad local minima of optimization [52], which, therefore, undermines the metric learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 275, + 384, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 275, + 384, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 275, + 384, + 289 + ], + "type": "text", + "content": "3.3 Pathology Knowledge Enhanced Pretraining" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 293, + 481, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 481, + 331 + ], + "type": "text", + "content": "In this section, we present a simple yet effective pretraining approach, termed KEP, that leverages the established knowledge encoder to guide visual-language pretraining for computational pathology." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 335, + 481, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 335, + 481, + 371 + ], + "spans": [ + { + "bbox": [ + 131, + 335, + 481, + 371 + ], + "type": "text", + "content": "Visual-Language Pretraining. Given paired image and captions, denoted as " + }, + { + "bbox": [ + 131, + 335, + 481, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = \\{(x_1, c_1), \\ldots, (x_n, c_n)\\}" + }, + { + "bbox": [ + 131, + 335, + 481, + 371 + ], + "type": "text", + "content": ", our goal is to construct a visual-language embedding space from paired image-text data, that satisfies:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 203, + 378, + 480, + 391 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 378, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 203, + 378, + 480, + 391 + ], + "type": "interline_equation", + "content": "\\left. \\operatorname {s i m} \\left(\\Phi_ {\\mathrm {v}} \\left(x _ {i}\\right), \\Phi_ {\\mathrm {t}} \\left(c _ {i}\\right)\\right) \\gg \\operatorname {s i m} \\left(\\Phi_ {\\mathrm {v}} \\left(x _ {i}\\right), \\Phi_ {\\mathrm {t}} \\left(c _ {j}\\right)\\right), i \\neq j, \\right. \\tag {5}", + "image_path": "5018db8379402adc3e5a7e5bc46faa7946520697fc60cbe666a7c69230c9b203.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 397, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 397, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\mathrm{v}}(\\cdot)" + }, + { + "bbox": [ + 130, + 397, + 482, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 397, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\mathrm{t}}(\\cdot)" + }, + { + "bbox": [ + 130, + 397, + 482, + 483 + ], + "type": "text", + "content": " denote the visual and the textual encoder, respectively. In this work, ViT-B-32/16 is adopted as the backbone for the visual encoder and initialized from the visual weights of CLIP [41] / BiomedCLIP [56]. A projection head is added on top of the visual encoder to bridge the gap caused by the non-pathology initialization, shown in Fig. 4. The text encoder is initialized by the pretrained knowledge encoder, as described in Sec. 3.2. In order to learn effective visual-language representation, we optimise an infoNCE contrastive objective:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 210, + 487, + 480, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 487, + 480, + 516 + ], + "spans": [ + { + "bbox": [ + 210, + 487, + 480, + 516 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {v t}} = - \\left(\\log \\frac {e ^ {\\mathbf {v} _ {i} ^ {T} \\mathbf {t} _ {i} / \\tau}}{\\sum_ {j = 1} ^ {n} e ^ {\\mathbf {v} _ {i} ^ {T} \\mathbf {t} _ {j} / \\tau}} + \\log \\frac {e ^ {\\mathbf {t} _ {i} ^ {T} \\mathbf {v} _ {i} / \\tau}}{\\sum_ {j = 1} ^ {n} e ^ {\\mathbf {t} _ {i} ^ {T} \\mathbf {v} _ {j} / \\tau}}\\right), \\tag {6}", + "image_path": "94c10403d63858993e5bde1abe501a457feedaa74c76d0459661c8690eca8a95.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 521, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 482, + 547 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 521, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i = \\varPhi_{\\mathrm{v}}(x_i)" + }, + { + "bbox": [ + 131, + 521, + 482, + 547 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 521, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i = \\varPhi_{\\mathrm{t}}(c_i)" + }, + { + "bbox": [ + 131, + 521, + 482, + 547 + ], + "type": "text", + "content": ", refer to the normalized embedding vectors from visual and text encoders, respectively. " + }, + { + "bbox": [ + 131, + 521, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 521, + 482, + 547 + ], + "type": "text", + "content": " denotes a temperature parameter." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 550, + 482, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 482, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 482, + 634 + ], + "type": "text", + "content": "Knowledge Distillation. To keep the alignment between images and free-form captions inside the knowledge space and thus the images can be linked to their implicit disease entities (right part in Fig. 4), we adopt an additional frozen branch to continuously distill pathology knowledge to the text encoder. Specifically, we use the text-knowledge embedding pairs to construct a contrastive loss item " + }, + { + "bbox": [ + 130, + 550, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tk}}" + }, + { + "bbox": [ + 130, + 550, + 482, + 634 + ], + "type": "text", + "content": ", which can be computed by Eq. 6 with the normalized visual embedding vectors replaced by the knowledge embedding vectors:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 209, + 639, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 639, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 209, + 639, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {t k}} = - \\left(\\log \\frac {e ^ {\\mathbf {k} _ {i} ^ {T} \\mathbf {t} _ {i} / \\tau}}{\\sum_ {j = 1} ^ {n} e ^ {\\mathbf {k} _ {i} ^ {T} \\mathbf {t} _ {j} / \\tau}} + \\log \\frac {e ^ {\\mathbf {t} _ {i} ^ {T} \\mathbf {k} _ {i} / \\tau}}{\\sum_ {j = 1} ^ {n} e ^ {\\mathbf {t} _ {i} ^ {T} \\mathbf {k} _ {j} / \\tau}}\\right), \\tag {7}", + "image_path": "22551dd2a922aaa68fb0738d19346e80e6ce8d9bbf94cd489b6c02b8648ffe87.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "KEP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 115, + 475, + 196 + ], + "blocks": [ + { + "bbox": [ + 141, + 115, + 475, + 196 + ], + "lines": [ + { + "bbox": [ + 141, + 115, + 475, + 196 + ], + "spans": [ + { + "bbox": [ + 141, + 115, + 475, + 196 + ], + "type": "image", + "image_path": "6614bfe1a57a81c0f11e2fffb9c0ba55b16a9d4704a935fb83700232a46369b4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 204, + 482, + 282 + ], + "lines": [ + { + "bbox": [ + 130, + 204, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 482, + 282 + ], + "type": "text", + "content": "Fig. 4: Model architecture (left graph). A projection head is added on the top of the visual encoder to bridge the gap between the image and the text encoder. The knowledge encoder is frozen across the whole training stage to distill pathology knowledge to the learnable text encoder. As a result, the pathology images can be aligned with their implicit disease labels (marked by dashed boxes in the right graph) during visual-language pretraining, since the captions contain disease attributes that have been already aligned with disease names/synonyms in the knowledge embedding space." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 304, + 482, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 329 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 304, + 482, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_i = \\varPhi_{\\mathrm{k}}(c_i)" + }, + { + "bbox": [ + 130, + 304, + 482, + 329 + ], + "type": "text", + "content": ", refer to the normalized embedding vectors from the knowledge encoder. The overall training loss can thus be computed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 271, + 336, + 480, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 336, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 271, + 336, + 480, + 348 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {v t}} + \\alpha \\mathcal {L} _ {\\mathrm {t k}}, \\tag {8}", + "image_path": "74070d458ffe2200070b1d1d217f2fbdda4ac8761d8385d018bcaca402b6c25d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 355, + 482, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 355, + 482, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 482, + 404 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 355, + 482, + 404 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 355, + 482, + 404 + ], + "type": "text", + "content": " denotes a weight parameter. It is worth emphasizing that the key contribution in KEP is to initialize the text encoder with the pre-obtained pathology knowledge encoder and adopt it to continuously distill pathology knowledge to the text encoder, while for the other parts, we keep the same as PLIP [23]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 421, + 230, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 421, + 230, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 421, + 230, + 435 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 446, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 471 + ], + "type": "text", + "content": "In this section, we first introduce the datasets used for training and evaluation in this paper, followed by the evaluation metrics and implementation details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 487, + 251, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 251, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 251, + 499 + ], + "type": "text", + "content": "4.1 Training Datasets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 507, + 482, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 543 + ], + "type": "text", + "content": "Dataset for Knowledge Encoding. We carry out pathology knowledge encoding with PathKT, where 4,718 disease nodes with a total number of 50,470 attributes are extracted, shown in Table S1 in Supplementary Materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 546, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 483, + 666 + ], + "type": "text", + "content": "Datasets for KEP Pretraining. Our dataset is composed of two parts. First, we collect the OpenPath data provided by PLIP [23] and obtain 138,874 image-text pairs after denoising and pre-processing. Second, we collect the Quilt1M [24] dataset, which gathers pathology image-text pairs from four public sources: PubMed articles, LAION [44], OpenPath [23], and Youtube videos. Since one of the downstream evaluation datasets Arch-PubMed [15] contains pathology images from PubMed articles, we remove this data source from Quilt1M to avoid data leaking. Considering that many images in Quilt1M are matched with multiple captions, we concatenate the captions related to the same image and finally obtain 576,608 image-text pairs for the Quilt1M dataset." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "X. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 255, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 255, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 255, + 126 + ], + "type": "text", + "content": "4.2 Downstream Tasks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 479, + 176 + ], + "type": "text", + "content": "We evaluate the pretrained models on three tasks, namely, retrieval, zero-shot patch classification, and zero-shot WSI tumor subtyping. The details for all evaluation datasets are exhibited in Table S2 in Supplementary Materials." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 179, + 480, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 179, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 179, + 480, + 299 + ], + "type": "text", + "content": "Retrieval. This task involves cross-modal retrieval and disease retrieval. Cross-modal retrieval aims to retrieve the correct caption for a given image and vice versa. Disease retrieval, on the other hand, utilizes the disease names to retrieve captions or images with the same disease label, which is proposed to demonstrate the effectiveness of the knowledge encoder. For cross-modal retrieval, we follow PLIP [23] to split the ARCH [15] dataset into Arch-PubMed and Arch-book. In addition, we also gather image-text pairs from publicly available educational resources and curate a retrieval dataset, termed as PathPair, consisting of 9,358 pathology image-caption pairs with known 1676 disease labels. For disease retrieval, we utilize the captions, images, and their disease labels in PathPair." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 303, + 481, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 481, + 434 + ], + "type": "text", + "content": "Zero-shot Patch Classification. This task involves zero-shot classification on patch-level pathology images. Specifically, at inference time, we randomly select one template from the 21 templates in CONCH [36] and one type synonym from the corresponding name list (exhibited in Supplementary Materials) to yield a text prompt for each type, e.g. a histopathological image of CLASSNAME. (template) + beast invasive carcinoma (synonym) " + }, + { + "bbox": [ + 130, + 303, + 481, + 434 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 303, + 481, + 434 + ], + "type": "text", + "content": " a histopathological image of beast invasive carcinoma. This process is repeated 100 times in every experiment. Following PLIP [23] and Quilt1M [24], we adopt these patch-level pathology image datasets: BACH [1], NCT-CRC-HE-100K [27], KatherColon [28], LC25000 [3], RenalCell [4], SICAP [47], SkinCancer [30], WSSS4LUAD [17]. Each dataset includes multiple types of H&E stained cell micrographs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 481, + 666 + ], + "type": "text", + "content": "Zero-shot WSI Tumor Subtyping. This task involves zero-shot tumor subtyping on pathology whole slide images of common and rare cancers. We follow MI-Zero [37] and CONCH [36] for evaluation. Specifically, we first divide WSI into " + }, + { + "bbox": [ + 130, + 437, + 481, + 666 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 437, + 481, + 666 + ], + "type": "text", + "content": " patches and then predict the class label of each image patch in a zero-shot manner. The tumor type of the whole slide is then provided by integrating Top-K predictions on patches. For this task, we also employ the templates and tumor synonyms from CONCH [36] to randomly generate 100 text prompts (listed in Supplementary materials) for each tumor subtype. (i) For common cancers, we follow existing research [36,37] and collect 525 WSIs with three tumor types, including 150 breast carcinoma (BRCA), 150 non-small cell lung cancer (NSCLC) histopathology slides, and 225 renal cell carcinoma (RCC) slides, from TCGA. Specifically, BRCA consists of two subtypes: invasive ductal carcinoma (IDC) and invasive lobular carcinoma (IDC). NSCLC contains lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC). RCC is divided into chromophobe renal cell carcinoma (CHRCC), clear-cell renal cell carcinoma (CCRCC), and papillary renal cell carcinoma (PRCC). Each tumor subtype has 75 WSIs. (ii) For rare cancers, we collected 41 WSIs of rare subtypes for breast cancer (6 for Intraductal papillary adenocarcinoma with invasion, 6 for Medullary carcinoma, NOS, 13 for Metaplastic carcinoma, NOS and" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "KEP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "16 for Mucinous adenocarcinoma), from TCGA and merge them with common breast cancers (75 for IDC and 75 for ILC)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 157, + 258, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 157, + 258, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 258, + 169 + ], + "type": "text", + "content": "4.3 Evaluation Metrics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 176, + 480, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 176, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 132, + 176, + 480, + 201 + ], + "type": "text", + "content": "Retrieval. To evaluate the retrieval performance, we adopt the Recall@K metric, suggesting the ratio of correctly retrieved queries in Top-K retrieved samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 480, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 239 + ], + "type": "text", + "content": "Zero-shot Patch Classification. For zero-shot classification tasks, we adopt the same metric as PLIP [23], namely, weighted F1 (wF1). We report the median, the first, and the third quartile (Q1, Q3) of wF1 across all text 100 prompts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 242, + 481, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 242, + 481, + 291 + ], + "spans": [ + { + "bbox": [ + 132, + 242, + 481, + 291 + ], + "type": "text", + "content": "Zero-shot WSI Tumor Subtyping. We adopt the same metric as MI-Zero [37], namely, balanced accuracy at Top-K pooling to measure the zero-shot tumor subtyping performance on WSIs. We also report the median, the first, and the third quartile (Q1, Q3) of the balanced accuracy across all 100 text prompts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 307, + 279, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 279, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 279, + 319 + ], + "type": "text", + "content": "4.4 Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 329, + 480, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 329, + 480, + 377 + ], + "spans": [ + { + "bbox": [ + 132, + 329, + 480, + 377 + ], + "type": "text", + "content": "Knowledge Encoder Pretraining. We adopt the architecture of PubMedBERT [16] to encode knowledge. The embedding dimension is set to 512. The temperature parameter " + }, + { + "bbox": [ + 132, + 329, + 480, + 377 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 132, + 329, + 480, + 377 + ], + "type": "text", + "content": " is set to 0.04 in Eq. 2. The batch size is set to 256, including 32 disease entities with 8 instances per entity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 481, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 476 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 476 + ], + "type": "text", + "content": "Pathology Image-text Pretraining. To achieve a fair comparison with PLIP and Quilt1M, we utilize the same visual encoder (ViT-B-32) and initialization of the visual encoder (CLIP [41]), termed by KEP-32, and set the input image size to " + }, + { + "bbox": [ + 132, + 380, + 481, + 476 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 132, + 380, + 481, + 476 + ], + "type": "text", + "content": ". Additionally, we also develop a KEP variant named KEP-16 that is initialized by the visual weights of BiomedCLIP [56] (ViT-B-16). The batch size and learning rate are set to 256 and 1e-5, respectively. The temperature in Eq. 6 and Eq. 7 is set to 0.04 across all experiments. For OpenPath, we conduct the pathology VLP for 30 epochs, While for Quilt1M, 15 epochs are adopted." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 494, + 198, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 198, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 198, + 506 + ], + "type": "text", + "content": "5 Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 518, + 480, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 518, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 518, + 480, + 578 + ], + "type": "text", + "content": "In this section, we show the results on the downstream tasks, to evaluate the effectiveness of proposed knowledge-enhanced representation learning. Note that for a fair comparison with the publicly available model released by the original authors, in all experiments, we separately train KEP on the OpenPath [23] dataset, and Quilt1M [24] dataset, and then report results on downstream tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 594, + 206, + 606 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 206, + 606 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 206, + 606 + ], + "type": "text", + "content": "5.1 Retrieval" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 614, + 480, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 614, + 480, + 638 + ], + "spans": [ + { + "bbox": [ + 132, + 614, + 480, + 638 + ], + "type": "text", + "content": "In this section, we evaluate the performance of retrieval, including cross-modal retrieval on three datasets and disease retrieval on PathPair." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "type": "text", + "content": "Cross-modal Retrieval. In Table 1, we demonstrate the results for different models pretrained on OpenPath and Quilt1M. It can be seen that KEP-32" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "X. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 168, + 476, + 298 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Performance comparison with PLIP and QuiltNet on three retrieval datasets. All models are pre-trained on the OpenPath and Quilt1M datasets, respectively. i2t and t2i denote image-to-text and text-to-image retrieval, respectively. Bold fonts suggest the best performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 168, + 476, + 298 + ], + "lines": [ + { + "bbox": [ + 135, + 168, + 476, + 298 + ], + "spans": [ + { + "bbox": [ + 135, + 168, + 476, + 298 + ], + "type": "table", + "html": "
Training datasetTaskModelKnowledge EnhancementArch-PubMedArch-bookPathPair
Recall@10Recall@50Recall@10Recall@50Recall@10Recall@50
OpenPathi2tPLIPX0.0670.1850.1520.3930.0380.119
KEP-32✓0.0980.2830.1640.4040.0710.205
KEP-16✓0.1630.3980.2440.5370.1080.275
t2iPLIPX0.0670.1810.1650.4190.0470.133
KEP-32✓0.0850.2260.1480.3650.0610.171
KEP-16✓0.1380.3390.2380.5330.0930.247
Quilt1Mi2tQuiltNetX0.1390.3260.1880.4070.0650.166
KEP-32✓0.1400.3270.2400.5210.0840.221
KEP-16✓0.1960.4210.2820.5640.1080.254
t2iQuiltNetX0.1220.2930.2040.4290.0710.195
KEP-32✓0.1350.3260.2750.5680.1060.276
KEP-16✓0.1760.4040.3400.6210.1360.326
", + "image_path": "55d03c691d24283179ddffbab037a0ebbbe7d48f5ff7cabb9bc65aac3327c8b5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 144, + 348, + 467, + 414 + ], + "blocks": [ + { + "bbox": [ + 132, + 316, + 480, + 338 + ], + "lines": [ + { + "bbox": [ + 132, + 316, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 480, + 338 + ], + "type": "text", + "content": "Table 2: Performance comparison of disease retrieval. 12t and i2l denote label-to-text and image-to-label, respectively. Bold fonts suggest the best performance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 348, + 467, + 414 + ], + "lines": [ + { + "bbox": [ + 144, + 348, + 467, + 414 + ], + "spans": [ + { + "bbox": [ + 144, + 348, + 467, + 414 + ], + "type": "table", + "html": "
TaskMetricsOpenPathQuiltNetQuilt1M
PLIPKEP-32KEP-16KEP-32KEP-16
l2tRecall@100.4080.6930.6990.2110.6480.643
Recall@500.5360.8320.8350.3250.8120.808
i2lRecall@100.1140.1620.2230.1350.2260.259
Recall@500.2920.3810.4780.3220.4730.519
", + "image_path": "1ab6a274e0b5b84d29e7b4869098efba207b4071c5b8d92e1834531599673a87.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 443, + 482, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 443, + 482, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 443, + 482, + 551 + ], + "type": "text", + "content": "pretrained on OpenPath outperforms PLIP on all datasets with respect to the image-to-text retrieval task, especially on Arch-PubMed and PathPair with more than " + }, + { + "bbox": [ + 132, + 443, + 482, + 551 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 132, + 443, + 482, + 551 + ], + "type": "text", + "content": " boost on R10. As for the text-to-image task, KEP-32 also achieves better performance on the dataset of Arch-PubMed and PathPair. Furthermore, our KEP-16 model improves the retrieval performance by a large margin on all datasets across both retrieval tasks. Similar results can be concluded that KEP-32 pretrained on Quilt1M improves the performance on all datasets for both retrieval tasks, which demonstrates the effectiveness of knowledge guidance for visual-language pretraining on cross-modal retrieval." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 557, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 482, + 665 + ], + "type": "text", + "content": "Disease Retrieval. Table 2 shows the performance of disease retrieval on the PathPair dataset for different models. It can be seen that although the scale of Quilt1M is 5 times OpenPath, models pretrained on OpenPath often outperforms that on Quilt1M for the label-to-text task. We conjecture this may be caused by the quality of captions in Quilt1M. Our approach KEP-32 outperforms PLIP and Quilt1M by a large margin on both label-to-text and image-to-label tasks, suggesting that the knowledge encoder contributes to paying attention to the key disease information when encoding captions, and thus improves the alignment between images and their disease labels." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "KEP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 113, + 309, + 220 + ], + "blocks": [ + { + "bbox": [ + 135, + 113, + 309, + 220 + ], + "lines": [ + { + "bbox": [ + 135, + 113, + 309, + 220 + ], + "spans": [ + { + "bbox": [ + 135, + 113, + 309, + 220 + ], + "type": "image", + "image_path": "106ec365bde8846e5a5a324aad1d54d9c95098cd352118798569d40f671aee2b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 228, + 482, + 295 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 295 + ], + "type": "text", + "content": "Fig. 5: The comparison of zero-shot patch classification between different models. The left and the right subfigures suggest pretraining on OpenPath and Quilt1M, respectively. The visual encoders of KEP-32 and KEP-16 are initialized by CLIP (ViT-B-32) and BiomedCLIP (ViT-B-16), respectively. The number of points for every box is 100, with each representing the performance of one text prompt. The upper, center, and lower line of each box denote the first, median, and third quartile of the distribution." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 113, + 478, + 220 + ], + "blocks": [ + { + "bbox": [ + 309, + 113, + 478, + 220 + ], + "lines": [ + { + "bbox": [ + 309, + 113, + 478, + 220 + ], + "spans": [ + { + "bbox": [ + 309, + 113, + 478, + 220 + ], + "type": "image", + "image_path": "c58d535787e4600fd839579c0cb7a46ccc4dc6c0cf507fd8a17ef3ee144ef61f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 316, + 312, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 312, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 312, + 328 + ], + "type": "text", + "content": "5.2 Zero-shot Patch Classification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 336, + 479, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 479, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 479, + 372 + ], + "type": "text", + "content": "In this section, we evaluate the performance of zero-shot classification on patch-level pathology images from 8 datasets. We report the performance distribution (Fig. 5), where each point in the figure denotes the performance of one prompt." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 374, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 495 + ], + "type": "text", + "content": "Comparison to PLIP. In Fig. 5 (left), we demonstrate the performance comparison with PLIP [23], by only pretraining on OpenPath image-text pairs. Compared with PLIP, KEP-32 achieves better zero-shot classification performance on all datasets, with a maximum enhancement of 0.137 on the BACH dataset. Moreover, with the medical-specific initialization of the visual encoder (BiomedCLIP), KEP-16 can further improve the performance significantly in most datasets. Note that, the boxes of KEP variants are generally shorter than that of PLIP, suggesting that our approach is less sensitive to the varying text prompts than PLIP, which can be demonstrated by the visualization of prompt embeddings, shown in Supplementary Materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": "Comparison to QuiltNet. As shown in Fig. 5 (right), we exhibit the performance comparison between KEP and QuiltNet [24] pretrained on Quilt1M dataset. Similar to results on OpenPath, KEP-32 outperforms Quilt1M in seven out of eight datasets, especially for the BACH, KatherColon, and SkinCancer datasets. KEP-16 can further improve the weighted F1 score. Most boxes of KEP variants are also shorter than that of Quilt1M. The visualization of prompt embeddings, shown in Supplementary Materials, maintains that pathology knowledge can reduce the ambiguities of image-text alignment and thus improve the zero-shot classification performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 622, + 328, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 328, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 328, + 635 + ], + "type": "text", + "content": "5.3 Zero-shot WSI Tumor Subtyping" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "In this section, we evaluate the transfer ability of different models for tumor subtyping on common and rare cancers." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "X. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 113, + 475, + 210 + ], + "blocks": [ + { + "bbox": [ + 138, + 113, + 475, + 210 + ], + "lines": [ + { + "bbox": [ + 138, + 113, + 475, + 210 + ], + "spans": [ + { + "bbox": [ + 138, + 113, + 475, + 210 + ], + "type": "image", + "image_path": "c6f606d380643a708d710e074aa0046d5dd33b2b5b04e325468b78fe21baf948.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 218, + 482, + 307 + ], + "lines": [ + { + "bbox": [ + 130, + 218, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 482, + 307 + ], + "type": "text", + "content": "Fig. 6: The performance comparison of tumor subtyping on TCGA-BRCA (common), TCGA-NSCLC, and TCGA-RCC WSIs. The upper, center and lower line of each box denote the first, median, and third quartile of the performance distribution, respectively. The scattered points represent outliers. KEP-16 and KEP-CTP are trained on OpenPath with the visual encoder initialized by BiomedCLIP [56] and CTransPath [48], respectively. MI-Zero-Bio and MI-Zero-Pub are two variants of MI-Zero with different initialization of text encoder. Their visual encoders are initialized by CTransPath. In addition, MI-Zero adopts in-house pathology reports to pretrain their text encoders." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 165, + 354, + 449, + 432 + ], + "blocks": [ + { + "bbox": [ + 132, + 321, + 480, + 343 + ], + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 343 + ], + "type": "text", + "content": "Table 3: Performance comparison of tumor subtyping on six BRCA subtypes, including 2 common and 4 rare cancers. Bold fonts suggest the best performance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 165, + 354, + 449, + 432 + ], + "lines": [ + { + "bbox": [ + 165, + 354, + 449, + 432 + ], + "spans": [ + { + "bbox": [ + 165, + 354, + 449, + 432 + ], + "type": "table", + "html": "
ModelTraining datasetTop-1Top-10
CommonRareCommonRare
PLIPOpenPath0.1130.1150.1190.119
QuiltNetQuilt1M0.2780.0940.2790.102
KEP-16OpenPath0.3430.2170.3680.245
MI-Zero-BioIn-house & Web data0.3900.2140.4030.237
MI-Zero-PubIn-house & Web data0.3310.2820.3470.304
KEP-CTPOpenPath0.4430.2820.4320.301
", + "image_path": "bce2f7d5067914490c5216bca130318b0dfa9f9787cfe2fbcb697769f8990fdb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 457, + 482, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 482, + 613 + ], + "type": "text", + "content": "Common cancers. Fig. 6 shows the performance comparison of tumor subtyping on common cancers, including TCGA-BRCA (common), TCGA-NSCLC and TCGA-RCC WSIs. MI-Zero-Bio and MI-Zero-Pub are two variants of MI-Zero with different text encoders. The visual encoders of the two MI-Zero variants are initialized by CTransPath [48]. To achieve a fair comparison, we also train a KEP variant named KEP-CTP with the visual encoder initialized by CTransPath. Both KEP-16 and KEP-CTP are trained on the OpenPath dataset. It can be seen that our approach KEP-16 and KEP-CTP outperform PLIP and QuiltNet on all datasets. Note that, the comparisons between variants of MI-Zero and KEP-CTP are actually not fair for our method, as the MI-Zero variants have also been pretrained on massive in-house data – over 550k pathology reports from hospitals [37]. Yet, KEP-CTP still outperforms MI-Zero-Bio across all datasets and achieves comparable performance with MI-Zero-Pub on NSCLC and RCC." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Rare cancers. Table 3 exhibits the median balanced accuracy of tumor subtyping on common and rare breast cancers. It can be seen that our method KEP-16 pretrained on OpenPath outperforms PLIP and QuiltNet by a large margin on both common and rare breast cancers. Moreover, KEP-CTP significantly outper" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "KEP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 179, + 476, + 288 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 4: Experimental results of ablation study on model architecture and initialization. Arch-PubMed and PathPair suggest the text-to-image retrieval and the disease retrieval task, respectively. PMB and KB denote PubMedBERT and our pretrained pathology knowledge encoder, respectively. BCLIP denotes BiomedCLIP. Bold fonts and underline suggest the best and the second-best performance, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 179, + 476, + 288 + ], + "lines": [ + { + "bbox": [ + 138, + 179, + 476, + 288 + ], + "spans": [ + { + "bbox": [ + 138, + 179, + 476, + 288 + ], + "type": "table", + "html": "
Visual Init.Text Init.Projection headKnowledge distillMetric lossKatherColon Median (Q1, Q3)Arch-PubMedPathPair
R10R50R1R5
ScratchPMB0.322 (0.272, 0.392)0.0370.1210.0560.113
CLIPPMB0.407 (0.358, 0.452)0.0810.2140.1610.316
CLIPPMB✓0.486 (0.448, 0.517)0.0730.2100.1610.305
BCLIPPMB✓0.553 (0.503, 0.591)0.1270.3470.1870.350
CLIPPMB✓✓(PMB)0.484 (0.435, 0.520)0.0630.1820.0340.089
CLIPKB✓Adasp0.530 (0.477, 0.600)0.0860.2320.2700.482
CLIPKB✓✓(KB)Adasp0.563 (0.505, 0.610)0.0850.2260.4090.618
CLIPKB✓✓(KB)Triplet0.531 (0.482, 0.580)0.0870.2270.2650.493
BCLIPKB✓✓(KB)Adasp0.580 (0.517, 0.631)0.1380.3390.4240.618
", + "image_path": "e374f3ff63dcdd0f9422a7e82edc6d0bf7f5ea0bafa1f3b196337c84425f53d8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 307, + 481, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 481, + 331 + ], + "type": "text", + "content": "forms MI-Zero variants on common breast cancers while achieving comparable performance on the rare ones." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 346, + 238, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 346, + 238, + 358 + ], + "spans": [ + { + "bbox": [ + 131, + 346, + 238, + 358 + ], + "type": "text", + "content": "5.4 Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 363, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 483 + ], + "type": "text", + "content": "In this section, we explore the impact of model architecture, initialization, and other hyperparameters, as shown in Table 4 and Table S3 in Supplementary Materials. Comparing the performance between different models, we can draw the following observations: (i) visual projection head can facilitate the zero-shot classification performance; (ii) our pretrained knowledge encoder can enhance the performance across all tasks; (iii) the frozen knowledge branch can further improve the performance of zero-shot classification and disease retrieval tasks while the frozen PubMedBERT can not; (iv) a better initialization of visual encoder contributes to significant performance improvement on all tasks. (v) the Adasp loss is better than the Triplet loss for the knowledge encoding." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 499, + 220, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 499, + 220, + 512 + ], + "spans": [ + { + "bbox": [ + 131, + 499, + 220, + 512 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": "In this paper, we address the problem of knowledge-enhanced visual-language pretraining on computational pathology. We first curate a pathology knowledge tree that integrates the informative attributes of diseases requiring pathological diagnosis. We then propose a novel knowledge encoding approach based on metric learning to model structured pathological knowledge. With the guidance of the pretrained knowledge encoder, we conduct extensive visual-language pretraining on pathology image-caption pairs. To demonstrate the effectiveness of our approach, we evaluate pretrained VLP models on three downstream tasks, including retrieval, zero-shot classification on patch-level pathology images, and zero-shot tumor subtyping on pathology WSIs. Quantitative experimental results demonstrate that pathology knowledge can significantly improve the performance across different tasks." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "X. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 481, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 481, + 174 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 481, + 174 + ], + "type": "text", + "content": "This work is supported by the National Key R&D Program of China (No. 2022ZD0160702) and China Postdoctoral Science Foundation (Certificate Number: 2023M741850)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 211, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 211, + 481, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 481, + 245 + ], + "type": "text", + "content": "1. Aresta, G., Araújo, T., Kwok, S., Chennamsetty, S.S., Safwan, M., Alex, V., Marami, B., Prastawa, M., Chan, M., Donovan, M., et al.: Bach: Grand challenge on breast cancer histology images. Medical Image Analysis 56, 122-139 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 245, + 481, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 481, + 266 + ], + "type": "text", + "content": "2. Bodenreider, O.: The unified medical language system (uml's): integrating biomedical terminology. Nucleic Acids Research 32(suppl_1), D267-D270 (2004)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 481, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 481, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 481, + 298 + ], + "type": "text", + "content": "3. Borkowski, A.A., Bui, M.M., Thomas, L.B., Wilson, C.P., DeLand, L.A., Mastorides, S.M.: Lung and colon cancer histopathological image dataset (lc25000). arXiv preprint arXiv:1912.12142 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 298, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 298, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 298, + 481, + 331 + ], + "type": "text", + "content": "4. Brummer, O., Pölönen, P., Mustjoki, S., Brück, O.: Integrative analysis of histological textures and lymphocyte infiltration in renal cell carcinoma using deep learning. bioRxiv pp. 2022-08 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 331, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 481, + 373 + ], + "type": "text", + "content": "5. Campanella, G., Hanna, M.G., Geneslaw, L., Miraflor, A., Werneck Krauss Silva, V., Busam, K.J., Brogi, E., Reuter, V.E., Klimstra, D.S., Fuchs, T.J.: Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature Medicine 25(8), 1301-1309 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 373, + 481, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 373, + 481, + 417 + ], + "spans": [ + { + "bbox": [ + 138, + 373, + 481, + 417 + ], + "type": "text", + "content": "6. Chan, T.H., Cendra, F.J., Ma, L., Yin, G., Yu, L.: Histopathology whole slide image analysis with heterogeneous graph representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15661-15670 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 417, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 417, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 417, + 481, + 460 + ], + "type": "text", + "content": "7. Chauhan, G., Liao, R., Wells, W., Andreas, J., Wang, X., Berkowitz, S., Horng, S., Szolovits, P., Golland, P.: Joint modeling of chest radiographs and radiology reports for pulmonary edema assessment. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 529-539. Springer (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 460, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 481, + 493 + ], + "type": "text", + "content": "8. Chen, C., Lu, M.Y., Williamson, D.F., Chen, T.Y., Schaumberg, A.J., Mahmood, F.: Fast and scalable search of whole-slide images via self-supervised deep learning. Nature Biomedical Engineering 6(12), 1420-1434 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 493, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 481, + 536 + ], + "type": "text", + "content": "9. Chen, R.J., Chen, C., Li, Y., Chen, T.Y., Trister, A.D., Krishnan, R.G., Mahmood, F.: Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16144-16155 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 536, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 481, + 557 + ], + "type": "text", + "content": "0. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: Uniter: Learning universal image-text representations (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 557, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 481, + 601 + ], + "type": "text", + "content": "1. Chen, Y.C., Lu, C.S.: Rankmix: Data augmentation for weakly supervised learning of classifying whole slide images with diverse sizes and imbalanced categories. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23936-23945 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "text", + "content": "2. Chen, Z., Li, G., Wan, X.: Align, reason and learn: Enhancing medical vision-and-language pre-training with knowledge. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 5152-5161 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 666 + ], + "type": "text", + "content": "3. Chen, Z., Song, Y., Chang, T.H., Wan, X.: Generating radiology reports via memory-driven transformer. In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). pp. 1439-1449 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "KEP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "14. Cui, H., Xu, Y., Li, W., Wang, L., Duh, H.: Collaborative learning of cross-channel clinical attention for radiotherapy-related esophageal fistula prediction from ct. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 212-220. Springer (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "type": "text", + "content": "15. Gamper, J., Rajpoot, N.: Multiple instance captioning: Learning representations from histopathology textbooks and articles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16549-16559 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "type": "text", + "content": "16. Gu, Y., Tinn, R., Cheng, H., Lucas, M., Usuyama, N., Liu, X., Naumann, T., Gao, J., Poon, H.: Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Computing for Healthcare (HEALTH) 3(1), 1-23 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "type": "text", + "content": "17. Han, C., Pan, X., Yan, L., Lin, H., Li, B., Yao, S., Lv, S., Shi, Z., Mai, J., Lin, J., et al.: Wsss4luad: Grand challenge on weakly-supervised tissue semantic segmentation for lung adenocarcinoma. arXiv preprint arXiv:2204.06455 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 271, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 271, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 271, + 481, + 293 + ], + "type": "text", + "content": "18. Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 133, + 293, + 481, + 326 + ], + "type": "text", + "content": "19. Hu, S., Chen, L., Wu, P., Li, H., Yan, J., Tao, D.: St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In: European Conference on Computer Vision. pp. 533-549. Springer (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 326, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 326, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 133, + 326, + 481, + 380 + ], + "type": "text", + "content": "20. Huang, S.C., Shen, L., Lungren, M.P., Yeung, S.: Gloria: A multimodal global-local representation learning framework for label-efficient medical image recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3942-3951 (2021), official Implementation: https://github.com/marshuang80/gloria" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "21. Huang, X., Fang, Y., Lu, M., Yan, F., Yang, J., Xu, Y.: Dual-ray net: automatic diagnosis of thoracic diseases using frontal and lateral chest x-rays. Journal of Medical Imaging and Health Informatics 10(2), 348-355 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 414, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 414, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 133, + 414, + 481, + 457 + ], + "type": "text", + "content": "22. Huang, Y., Zhao, W., Wang, S., Fu, Y., Jiang, Y., Yu, L.: Conslide: Asynchronous hierarchical interaction transformer with breakup-reorganize rehearsal for continual whole slide image analysis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 21349-21360 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 133, + 457, + 481, + 490 + ], + "type": "text", + "content": "23. Huang, Z., Bianchi, F., Yuksekgonul, M., Montine, T.J., Zou, J.: A visual-language foundation model for pathology image analysis using medical twitter. Nature medicine 29(9), 2307-2316 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 133, + 491, + 481, + 522 + ], + "type": "text", + "content": "24. Ikezogwo, W., Seyfioglu, S., Ghezloo, F., Geva, D., Sheikh Mohammed, F., Anand, P.K., Krishna, R., Shapiro, L.: Quilt-1m: One million image-text pairs for histopathology. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 523, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 523, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 133, + 523, + 481, + 566 + ], + "type": "text", + "content": "25. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning. pp. 4904-4916. PMLR (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 133, + 567, + 481, + 600 + ], + "type": "text", + "content": "26. Kang, M., Song, H., Park, S., Yoo, D., Pereira, S.: Benchmarking self-supervised learning on diverse pathology datasets. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3344-3354 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 601, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 601, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 133, + 601, + 481, + 621 + ], + "type": "text", + "content": "27. Kather, J.N., Halama, N., Marx, A.: 100,000 histological images of human colorectal cancer and healthy tissue. Zenodo10 5281 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 622, + 481, + 665 + ], + "type": "text", + "content": "28. Kather, J.N., Krisam, J., Charoentong, P., Luedde, T., Herpel, E., Weis, C.A., Gaiser, T., Marx, A., Valous, N.A., Ferber, D., et al.: Predicting survival from colorectal cancer histology slides using deep learning: A retrospective multicenter study. PLoS Medicine 16(1), e1002730 (2019)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "X. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "29. Kim, W., Son, B., Kim, I.: Vilt: Vision-and-language transformer without convolution or region supervision. In: International Conference on Machine Learning. pp. 5583-5594. PMLR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 194 + ], + "type": "text", + "content": "30. Kriegsmann, K., Lobers, F., Zgorzelski, C., Kriegsmann, J., Janssen, C., Meliss, R.R., Muley, T., Sack, U., Steinbuss, G., Kriegsmann, M.: Deep learning for the detection of anatomical tissue structures and neoplasms of the skin on scanned histopathological tissue sections. Frontiers in Oncology 12, 1022967 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "text", + "content": "31. Kundra, R., Zhang, H., Sheridan, R., Sirlintrapun, S.J., Wang, A., Ochoa, A., Wilson, M., Gross, B., Sun, Y., Madupuri, R., et al.: Oncotree: a cancer classification system for precision oncology. JCO clinical cancer informatics 5, 221-230 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "type": "text", + "content": "32. Li, H., Zhu, C., Zhang, Y., Sun, Y., Shui, Z., Kuang, W., Zheng, S., Yang, L.: Task-specific fine-tuning via variational information bottleneck for weakly-supervised pathology whole slide image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7454-7463 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 270, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 314 + ], + "type": "text", + "content": "33. Li, M., Cai, W., Verspoor, K., Pan, S., Liang, X., Chang, X.: Cross-modal clinical graph transformer for ophthalmic report generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20656-20665 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 314, + 482, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 347 + ], + "type": "text", + "content": "34. Lin, T., Yu, Z., Hu, H., Xu, Y., Chen, C.W.: Interventional bag multi-instance learning on whole-slide pathological images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19830-19839 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 347, + 482, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 392 + ], + "type": "text", + "content": "35. Lin, W., Zhao, Z., Zhang, X., Wu, C., Zhang, Y., Wang, Y., Xie, W.: Pmc-clip: Contrastive language-image pre-training using biomedical documents. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 392, + 482, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 424 + ], + "type": "text", + "content": "36. Lu, M.Y., Chen, B., Williamson, D.F., Chen, R.J., Liang, I., Ding, T., Jaume, G., Odintsov, I., Zhang, A., Le, L.P., et al.: Towards a visual-language foundation model for computational pathology. arXiv preprint arXiv:2307.12914 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 424, + 482, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 468 + ], + "type": "text", + "content": "37. Lu, M.Y., Chen, B., Zhang, A., Williamson, D.F., Chen, R.J., Ding, T., Le, L.P., Chuang, Y.S., Mahmood, F.: Visual language pretrained multiple instance zero-shot transfer for histopathology images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19764-19775 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "type": "text", + "content": "38. Lu, M.Y., Williamson, D.F., Chen, T.Y., Chen, R.J., Barbieri, M., Mahmood, F.: Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering 5(6), 555-570 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 501, + 482, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 534 + ], + "type": "text", + "content": "39. Müller, P., Kaissis, G., Zou, C., Rueckert, D.: Joint learning of localized representations from medical images and reports. In: European Conference on Computer Vision. pp. 685-701. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 534, + 482, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 600 + ], + "type": "text", + "content": "40. Neumann, M., King, D., Beltagy, I., Ammar, W.: ScispaCy: Fast and robust models for biomedical natural language processing. In: Demner-Fushman, D., Cohen, K.B., Ananiadou, S., Tsujii, J. (eds.) Proceedings of the 18th BioNLP Workshop and Shared Task. pp. 319-327. Association for Computational Linguistics, Florence, Italy (Aug 2019). https://doi.org/10.18653/v1/W19-5034, https://aclanthology.org/W19-5034" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 600, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 482, + 643 + ], + "type": "text", + "content": "41. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 643, + 482, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 482, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 482, + 654 + ], + "type": "text", + "content": "42. Roberts, R.J.: Pubmed central: The genbank of the published literature (2001)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 654, + 400, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 654, + 400, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 654, + 400, + 666 + ], + "type": "text", + "content": "43. Rorke, L.B.: Pathologic diagnosis as the gold standard (1997)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "KEP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 160 + ], + "type": "text", + "content": "44. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278–25294 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "type": "text", + "content": "45. Shaban, M., Awan, R., Fraz, M.M., Azam, A., Tsang, Y.W., Snead, D., Rajpoot, N.M.: Context-aware convolutional neural network for grading of colorectal cancer histology images. IEEE Transactions on Medical Imaging 39(7), 2395-2405 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "text", + "content": "46. Shao, Z., Bian, H., Chen, Y., Wang, Y., Zhang, J., Ji, X., et al.: Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Advances in Neural Information Processing Systems 34, 2136-2147 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "text", + "content": "47. Silva-Rodriguez, J., Colomer, A., Sales, M.A., Molina, R., Naranjo, V.: Going deeper through the gleason scoring scale: An automatic end-to-end system for histology prostate grading and cribriform pattern detection. Computer Methods and Programs in Biomedicine 195, 105637 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "text", + "content": "48. Wang, X., Yang, S., Zhang, J., Wang, M., Zhang, J., Yang, W., Huang, J., Han, X.: Transformer-based unsupervised contrastive learning for histopathological image classification. Medical Image Analysis 81, 102559 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 326 + ], + "type": "text", + "content": "49. Wu, C., Zhang, X., Zhang, Y., Wang, Y., Xie, W.: Medklip: Medical knowledge enhanced language-image pre-training. medRxiv pp. 2023-01 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "50. Xie, X., Niu, J., Liu, X., Chen, Z., Tang, S., Yu, S.: A survey on incorporating domain knowledge into deep learning for medical image analysis. Medical Image Analysis 69, 101985 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "type": "text", + "content": "51. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettle-moyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 6787-6800 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "52. Xuan, H., Stylianou, A., Liu, X., Pless, R.: Hard negative examples are hard, but useful. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 126-142. Springer (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "53. Yang, S., Wu, X., Ge, S., Zhou, S.K., Xiao, L.: Knowledge matters: Chest radiology report generation with general and specific knowledge. Medical Image Analysis 80, 102510 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "text", + "content": "54. Yu, F., Tang, J., Yin, W., Sun, Y., Tian, H., Wu, H., Wang, H.: Ernie-vil: Knowledge enhanced vision-language representations through scene graphs. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 3208-3216 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "55. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "56. Zhang, S., Xu, Y., Usuyama, N., Bagga, J., Tinn, R., Preston, S., Rao, R., Wei, M., Valluri, N., Wong, C., et al.: Large-scale domain-specific pretraining for biomedical vision-language processing. arXiv preprint arXiv:2303.00915 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "57. Zhang, X., Wu, C., Zhang, Y., Xie, W., Wang, Y.: Knowledge-enhanced visual-language pre-training on chest radiology images. Nature Communications 14(1), 4542 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 643 + ], + "type": "text", + "content": "58. Zhang, Y., Jiang, H., Miura, Y., Manning, C.D., Langlotz, C.P.: Contrastive learning of medical visual representations from paired images and text. In: Machine Learning for Healthcare (2022), highest Starred Implementation: https://github.com/edreisMD/ConVIRT-pytorch" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "59. Zhou, X., Cheng, Z., Gu, M., Chang, F.: Lirnet: Local integral regression network for both strongly and weakly supervised nuclei detection. In: 2020 IEEE Interna" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "X. Zhou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 172 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "text", + "content": "tional Conference on Bioinformatics and Biomedicine (BIBM). pp. 945-951. IEEE (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "type": "text", + "content": "60. Zhou, X., Zhong, Y., Cheng, Z., Liang, F., Ma, L.: Adaptive sparse pairwise loss for object re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19691-19701 (2023)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "KEP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/L-DiffER_ Single Image Reflection Removal with Language-based Diffusion Model/84990c91-4cd8-4a1b-83d4-8e97d608b20a_content_list.json b/2024/L-DiffER_ Single Image Reflection Removal with Language-based Diffusion Model/84990c91-4cd8-4a1b-83d4-8e97d608b20a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a2846dcdb08285907571c367d423c5fb8c5ce6ae --- /dev/null +++ b/2024/L-DiffER_ Single Image Reflection Removal with Language-based Diffusion Model/84990c91-4cd8-4a1b-83d4-8e97d608b20a_content_list.json @@ -0,0 +1,1857 @@ +[ + { + "type": "text", + "text": "L-DiffER: Single Image Reflection Removal with Language-based Diffusion Model", + "text_level": 1, + "bbox": [ + 254, + 140, + 750, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuchen Hong $^{1,2\\#}$ Haofeng Zhong $^{1,2,3\\#}$ Shuchen Weng $^{1,2}$ Jinxiu Liang $^{1,2}$ Boxin Shi $^{1,2,3*}$", + "bbox": [ + 284, + 210, + 715, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University", + "bbox": [ + 287, + 253, + 712, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University", + "bbox": [ + 292, + 282, + 705, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ AI Innovation Center, School of Computer Science, Peking University {hfzhong, shuchenweng, cssherryliang, shiboxin}@pku.edu.cn, yuchenhong.cn@gmail.com", + "bbox": [ + 263, + 310, + 738, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this paper, we introduce L-DiffER, a language-based diffusion model designed for the ill-posed single image reflection removal task. Although having shown impressive performance for image generation, existing language-based diffusion models struggle with precise control and faithfulness in image restoration. To overcome these limitations, we propose an iterative condition refinement strategy to resolve the problem of inaccurate control conditions. A multi-condition constraint mechanism is employed to ensure the recovery faithfulness of image color and structure while retaining the generation capability to handle low-transmitted reflections. We demonstrate the superiority of the proposed method through extensive experiments, showcasing both quantitative and qualitative improvements over existing methods.", + "bbox": [ + 261, + 388, + 743, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Reflection removal $\\cdot$ Language-based diffusion model", + "bbox": [ + 261, + 569, + 700, + 584 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 632, + 375, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When photographs are taken through transparent mediums like glass windows, the contaminated mixture image (denoted as $\\mathbf{M}$ ) can be considered as the combination of a transmission layer (denoted as $\\mathbf{T}$ ) and a reflection layer (denoted as $\\mathbf{R}$ ) [61, 62], which impairs the performance of downstream computer vision tasks [1, 47, 60]. Consequently, reflection removal, which aims at removing undesired reflections and recovering clear transmission layers from contaminated mixture images, has become an attractive topic in the field of computational photography [8, 9, 16, 31, 59, 65, 77]. State-of-the-art single-image reflection removal methods [8, 22, 31, 61] predominantly learn deep priors from a mixed dataset of synthetic and real data to mitigate reflections. However, due to insufficient", + "bbox": [ + 212, + 664, + 787, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "# Equal contributions. * Corresponding author.", + "bbox": [ + 215, + 824, + 535, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e7737be1bfb2e3a48f3a32aa083679e9febad2f0975eead01b83fcc556c4e085.jpg", + "image_caption": [ + "Fig. 1: For mixture images with strong reflections, (a-b) single-image reflection removal methods [22, 31] fail in distinguishing reflection and transmission layers. (c) Using language descriptions as guidance and mixture images as conditions, ControlNet [74] generates results with color shifts and structure distortions, while (d) the proposed L-DiffER successfully achieves high-fidelity transmission recovery in low-transmitted reflection regions." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 406, + 318 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3d3eaafcade95108a6cb7e3dfe936b46f442b6e1d2492b7827f3923fbcdef73c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 143, + 787, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "knowledge about transmission and reflection scenes for addressing such a highly ill-posed problem, they often encounter limitations in scenes with complex or low-transmitted reflections, as shown in yellow boxes of Fig. 1(a-b). By leveraging auxiliary scene information from additional inputs obtained with specialized devices [27, 28, 37] or capture settings [34, 35, 43], multi-image reflection removal methods achieve more robust reflection removal than single-image methods. However, the specialized requirements for data acquisition, in contrast, limit their application scopes, especially for mobile devices and images from the Internet. Hence, it is imperative to find a user-friendly auxiliary input that helps to relieve the ill-posedness of reflection removal while maintaining the applicability and accessibility of single-image methods.", + "bbox": [ + 212, + 445, + 787, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fortunately, natural language can exactly serve as the aforementioned auxiliary input, since it provides instructive semantic information about images [36, 63, 72], which bridges the gap between human understanding and machine perception. Thanks to the development of vision-language models [48], various vision tasks (e.g., image editing [55, 56, 67, 68] and image colorization [3-5, 69]) benefit from the semantic guidance of language descriptions. Zhong et al. [81] make the first attempt to introduce language descriptions for reflection removal, but due to the lack of generation capability and generalizability, it encounters difficulties in addressing scenarios containing low-transmitted reflections.", + "bbox": [ + 212, + 612, + 787, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, by combining semantic information from natural language descriptions and generative priors from pre-trained diffusion models [14, 49], language-based diffusion models (e.g., ControlNet [74]) generate visually pleasant results for image restoration tasks like image colorization [4] and super resolution [54]. For reflection removal, it also shows the potential to provide semantic priors to relieve the ill-posedness, where the generative priors in pre-trained diffusion", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Hong et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "models can adeptly manage complex scenarios and make it possible to recover transmission layers under low-transmitted reflections. However, as shown in the red boxes of Fig. 1(c), directly employing the existing language-based diffusion model like ControlNet [74] for reflection removal will encounter two primary challenges: (1) Inaccurate control conditions. When using language-based diffusion models, other tasks such as image colorization [4] often employ a clear (not superimposed) image to provide accurate structure control. However, in the reflection removal task, the image content of the transmission layer in a mixture image is contaminated by the reflection layer, and thus it is unable to provide an initially accurate condition for both structure and color control. Consequently, achieving transmission recovery under such inaccurate conditions presents a significant challenge. (2) Insufficient recovery faithfulness. Since diffusion models are principally designed for image generation, simply applying them for reflection removal may obtain results that are less faithful to conditions (i.e., results with unwanted image content containing color shifts and structural distortions compared with mixture images). Therefore, it is essential to develop a mechanism for constraining the generation capability and improving the recovery faithfulness of diffusion models to recover transmission layers while preventing unwanted alterations in image color and structure.", + "bbox": [ + 212, + 146, + 787, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we propose L-DiffER, the first Language-based Diffusion model for single image rEflection Removal, to achieve high-fidelity transmission recovery with language guidance under the interference of low-transmitted reflections, as the results shown in Fig. 1(d). To leverage the language descriptions of two layers, descriptions of the transmission layer serve as positive prompts, guiding the model to preserve essential image content, whereas descriptions of the reflection layer act as negative prompts, aiding in the suppression of unwanted reflections. To address the aforementioned challenges, we propose an iterative condition refinement strategy within the reverse diffusion process to ensure increasingly accurate representations of color and structure conditions as the process evolves. Furthermore, we introduce a multi-condition constraint mechanism designed to safeguard the faithfulness to the specified conditions, which effectively addresses potential color shifts and structural distortions by applying constraints that align the recovery output with the desired color and structure conditions. The proposed method achieves state-of-the-art performance by making the following contributions:", + "bbox": [ + 212, + 438, + 787, + 681 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present the first language-based diffusion model for reflection removal, leveraging the descriptive power of language to distinguish between transmission and reflection layers.", + "- We propose an iterative condition refinement strategy to provide more accurate control conditions as color and structure guidance for reflection removal.", + "- We design a multi-condition constraint mechanism to ensure the recovery faithfulness of image color and structure, which effectively handles low-transmitted reflections." + ], + "bbox": [ + 225, + 709, + 784, + 838 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "L-DiffER", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 388, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reflection removal. Single image reflection removal aims at suppressing reflections and recovering transmission layers in a single mixture image. Using the presumption that reflections tend to blur and exhibit low intensities, nonlearning approaches employ handcrafted priors into their optimization frameworks [57, 71], such as the gradient sparsity [30], relative smoothness [33], and ghosting cues [51]. With the development of deep learning, researchers attempt to address the task by exploring different network architectures [9, 21, 59, 61, 65], loss functions [76], or the iterative recovery strategy [8, 31, 70, 77]. Concurrently, efforts in data synthesis to meet the data-driven demands is also an attractive topic in this area [9, 22, 24, 39, 66, 79, 80]. Besides, the exploration of special image forms like panoramic images [10, 19, 20, 45] helps to diminish content ambiguity in mixture images. To relieve the ill-posedness of reflection removal, multi-image methods utilize additional input images to introduce auxiliary clues such as polarization information [7, 26, 28, 37, 38, 44, 50], transmission scene information from active illuminations [2, 17, 18, 27, 29], and motion discrepancy between the two layers [32, 34, 35, 52], while special data capture requirements limit their practical applications. Inspired by Zhong et al. [81], we introduce language-based diffusion models to input auxiliary language descriptions for relieving the ill-posedness of reflection removal while maintaining the applicability of the method, and to leverage generative priors for handling low-transmitted reflection regions.", + "bbox": [ + 212, + 172, + 787, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Language-based diffusion models. The advent of diffusion models [14] (DMs) has marked a significant leap forward, particularly in the realms of image generation [49]. The emergence of pre-trained vision-language models, notably CLIP [48], significantly enhances the flexibility and intuitiveness of utilizing language descriptions for vision tasks. Specifically, these pre-trained models have been instrumental in enabling multi-condition image editing [42,74,78], image colorization [4], and devising sophisticated sampling strategies for tailored generation outcomes [12]. Despite the strides in image editing and conditional generation, the specific challenge of single image reflection removal, i.e., recovering clear transmission layers from the guidance of a superimposed mixture images, still remains unexplored for language-based diffusion models [49, 74]. In our work, we seek to address the reflection removal problem by harnessing the generative prowess and flexibility of language-guided diffusion models, which combines the semantic information offered by language descriptions with the sophisticated generative capabilities of DMs while ensuring the recovery fidelity and faithfulness with condition refinement and constraints.", + "bbox": [ + 212, + 474, + 790, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 736, + 380, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we address the problem of language-based reflection removal using diffusion models. We first provide a brief overview of the framework of our language-based reflection removal diffusion model L-DiffER in Sec. 3.1. Sec. 3.2 proposes the iterative condition refinement strategy designed for addressing inaccurate control conditions, and Sec. 3.3 proposes the multi-condition constraint", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Hong et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d6231c5081ffb347fde3bb3531b922f76ac69d15cc683cf5f5b51c2928ac4fb2.jpg", + "image_caption": [ + "Fig. 2: Framework of the proposed language-based reflection removal diffusion model, which leverages language descriptions with the iterative condition refinement strategy and the multi-condition constraint mechanism to achieve faithful transmission recovery." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 583, + 357 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f1a9c04a4c585ba8296882b45ef9f4b3b4354b31dd6fb4ffd7489a0ddcac0830.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 146, + 779, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mechanism for improving the recovery faithfulness. Finally, loss functions and training details are presented in Sec. 3.4 and Sec. 3.5, respectively.", + "bbox": [ + 214, + 441, + 782, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 496, + 339, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem formulation. Given the mixture image $\\mathbf{M} \\in \\mathbb{R}^{H \\times W \\times 3}$ and the language description condition $c$ , we optimize our model to remove reflection contamination and recover transmission layers $\\widetilde{\\mathbf{T}} \\in \\mathbb{R}^{H \\times W \\times 3}$ under the supervision of the ground truth transmission layers $\\mathbf{T} \\in \\mathbb{R}^{H \\times W \\times 3}$ . We apply the commonly-used assumption [21,31,61] as: $\\mathbf{M} = \\mathbf{T} + \\mathbf{R}$ , where the mixture image $\\mathbf{M}$ can be considered as the combination of a transmission layer $\\mathbf{T}$ and a reflection layer $\\mathbf{R}$ . We use the intensity and gradient of the mixture image $\\mathbf{M}$ to initialize the color condition $\\mathbf{I}^{\\mathrm{s}} = \\mathbf{M}$ and the structure condition $\\mathbf{I}^{\\mathrm{s}} = \\nabla \\mathbf{M}$ . Then the color latent $y^{\\mathrm{c}} = \\mathcal{E}^{\\mathrm{c}}(\\mathbf{I}^{\\mathrm{c}})$ and structure latent $y^{\\mathrm{s}} = \\mathcal{E}^{\\mathrm{s}}(\\mathbf{I}^{\\mathrm{s}})$ are extracted by a color and a structure encoder $\\mathcal{E}^{\\mathrm{c}}$ and $\\mathcal{E}^{\\mathrm{s}}$ (denoted by the blue blocks in the middle part of Fig. 2), respectively. Given language descriptions about the content of transmission layers and reflection layers as positive prompt and negative prompt, respectively, the positive and negative latents $c^{\\mathrm{p}}$ and $c^{\\mathrm{n}}$ are extracted from the CLIP text encoder [48]. In the proposed method, the transmission layer $\\widetilde{\\mathbf{T}}$ is recovered from all the spatial and language conditioning information encoded in latents $y = [y^{\\mathrm{s}}, y^{\\mathrm{c}}]$ and $c = [c^{\\mathrm{p}}, c^{\\mathrm{n}}]$ .", + "bbox": [ + 212, + 521, + 787, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffusion models. Diffusion models [14, 53] are probabilistic generative models that have been shown to have powerful data modeling ability for images, which can provide an effective image prior for the estimation of transmission layers. To address the challenges of resource consumption in pixel space, Stable Diffusion (SD) [49] introduces latent diffusion models to improve both the", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "L-DiffER", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "training and sampling efficiency of denoising diffusion models without degrading their quality. Following SD [49], we employ a compression encoder $\\mathcal{E}$ to encode a transmission layer into the latent space as $z = \\mathcal{E}(\\mathbf{T})$ , and a compression decoder $\\mathcal{D}$ to reconstruct the transmission layer from the latent code $z$ as $\\widetilde{\\mathbf{T}} = \\mathcal{D}(z)$ . The diffusion process progressively destructs data by injecting noise, then learns to reverse this process for image generation. During training, the forward diffusion process starts from the target image $z_0 = \\mathcal{E}(\\mathbf{T})$ and generates samples $z_t = \\sqrt{\\overline{\\alpha}_t} z_0 + \\sqrt{1 - \\overline{\\alpha}_t} \\epsilon_t$ , where $\\epsilon_t \\sim \\mathcal{N}(0,1)$ is the Gaussian noise at timestep $t$ , and $\\overline{\\alpha}_t$ represents the noise scheduler introduced in [49]. During inference, the reverse diffusion process starts from a random noise sample $z_T \\sim \\mathcal{N}(0,1)$ , where a noise prediction network denoted as $\\epsilon_\\theta$ is optimized to predict the noise $\\hat{\\epsilon}_t$ at each timestep $t$ given the language condition $c$ and the spatial conditions $y_t^1$ as follows:", + "bbox": [ + 212, + 146, + 787, + 342 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {l d m}} = \\mathbb {E} _ {\\mathcal {E} (x), t, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon_ {t} - \\epsilon_ {\\theta} \\left(z _ {t}, c, y _ {t}, t\\right) \\| _ {2} ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 344, + 785, + 362 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "until it converges to the desired photorealistic transmission layer $\\widetilde{\\mathbf{T}}$ . The overall framework is shown in Fig. 2.", + "bbox": [ + 212, + 369, + 785, + 400 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Condition extraction. During inference, the two conditioning latents start from the initial color and structure latents, i.e., $y_{T}^{c} = y^{c}$ and $y_{T}^{\\mathrm{s}} = y^{\\mathrm{s}}$ , respectively, and they are refined by employing the iterative condition refinement strategy (Sec. 3.2) to provide more accurate conditions. For condition injection, the proposed framework draws inspiration from ControlNet [74], which learns a parameter-efficient, parallel branch on top of SD. Specifically, two trainable copied modules (denoted by the pink blocks in Fig. 2) are utilized to take the color latent $y_{t}^{c}$ and the gradient latent $y_{t}^{\\mathrm{s}}$ with the latent presentation $z_{t}$ as input to extract color and structure features, respectively. These features are then added directly to the corresponding scales of the locked module (denoted by the purple block in Fig. 2), which guides $\\epsilon_{\\theta}$ in learning the noise distribution.", + "bbox": [ + 212, + 402, + 787, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Language guidance. To explicitly instruct the model to suppress reflections from the language guidance, classifier-free guidance [15] is used to employ negative prompts to specify undesired reflection contamination. Specifically, two noise predictions with positive prompts $c^{\\mathrm{p}}$ and negative prompts $c^{\\mathrm{n}}$ are made at each timestep $t$ . The fusion of these two predictions yields the output $\\hat{e}_{t-1}$ :", + "bbox": [ + 212, + 568, + 787, + 643 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} _ {t - 1} = \\epsilon_ {\\theta} \\left(z _ {t}, c ^ {\\mathrm {p}}, y _ {t}, t\\right) + \\lambda_ {\\mathrm {c f g}} \\times \\left(\\epsilon_ {\\theta} \\left(z _ {t}, c ^ {\\mathrm {p}}, y _ {t}, t\\right) - \\epsilon_ {\\theta} \\left(z _ {t}, c ^ {\\mathrm {n}}, y _ {t}, t\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 279, + 657, + 784, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{cfg}}$ is a hyperparameter. In our task, as reflections are not always recognizable [81], we set negative prompts as the empty string $c^{\\mathrm{n}} = \\varnothing$ in the case. For simplicity, we use $\\epsilon_{\\theta}(z_t,c,y_t,t)$ to represent the whole noise prediction process.", + "bbox": [ + 212, + 688, + 787, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Iterative condition refinement strategy", + "text_level": 1, + "bbox": [ + 214, + 758, + 581, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Although with effective data modeling ability, diffusion models probably generate results unfaithful to the input image conditions [40,74]. This issue becomes even", + "bbox": [ + 212, + 784, + 787, + 815 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Hong et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1 Note that the spatial conditions are varied with timesteps in the proposed method.", + "bbox": [ + 217, + 824, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/db5626b361337e2b81256db283df02a4906ce5444fd8c7d716afc2c28c57e045.jpg", + "image_caption": [ + "Fig. 3: Visualization of pseudo transmission layers by using the original mixture image with its gradient as conditions." + ], + "image_footnote": [], + "bbox": [ + 220, + 146, + 782, + 223 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "harder for reflection removal, since the desired transmission layers should have a structure without gradients of reflection contaminations, which is different from the input mixture images. That is, the mixture images itself cannot provide an accurate image condition for diffusion-based reflection removal. Therefore, we turn to the recovered transmission layers in the inference process, which have the potential to be combined with the mixture images and help obtain a refined condition. During inference, given a latent noisy sample $z_{t}$ , we can use the noise prediction network $\\epsilon_{\\theta}$ to predict noise as $\\hat{\\epsilon}_t = \\epsilon_\\theta (z_t,t,c,y_t)$ . Then we can calculate the estimated latent representation of the transmission layer $z_{0|t}$ at the current timestep $t$ as follows:", + "bbox": [ + 212, + 295, + 787, + 446 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nz _ {0 \\mid t} = \\left(z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} * \\hat {\\epsilon} _ {t}\\right) / \\sqrt {\\bar {\\alpha} _ {t}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 459, + 785, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After that, we can obtain the estimated pseudo transmission layer $\\mathbf{T}_{0|t}$ by using the compression decoder $\\mathcal{D}$ to decode $z_{0|t}$ : $\\mathbf{T}_{0|t} = \\mathcal{D}(z_{0|t})$ . Then $\\mathbf{T}_{0|t}$ and its gradient $\\nabla \\mathbf{T}_{0|t}$ can be used as the pseudo-accurate color and structure condition to obtain conditions at the next timestep.", + "bbox": [ + 212, + 489, + 787, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As pseudo transmission layers shown in Fig. 3, we observe two key phenomena: Firstly, the estimated pseudo transmission layers in the first few sampling steps exhibit inaccuracies, with neither structure nor color serving effectively as guiding conditions. However, as the sampling steps progress, the fidelity of the reconstructed images notably improves. Secondly, although employing the mixture image and the gradient of the mixture image as conditions persistently yield inaccurate control, these conditions provide a richer set of usable information than the recovered image at the beginning of sampling. Leveraging the advantages of both observations, we introduce an iterative condition refinement strategy to address the challenge of inaccurate control conditions. Our control conditions transitioned from the mixture image and the gradient of the mixture image to the recovered image and its gradient as the sampling step increases.", + "bbox": [ + 212, + 551, + 787, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, we introduce predefined time-variant coefficients $\\beta_{t}$ and $\\gamma_{t}$ ( $t \\in \\{1,..,T\\}$ ) to combine pseudo-accurate conditions $\\mathbf{T}_{0|t}$ with the one initiated from $\\mathbf{M}$ , which decrease as timestep $t$ increases² as follows:", + "bbox": [ + 215, + 734, + 787, + 779 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {t} ^ {\\mathrm {c}} = \\gamma_ {t} * \\mathbf {T} _ {0 | t} + (1 - \\gamma_ {t}) * \\mathbf {M}, \\quad \\mathbf {I} _ {t} ^ {\\mathrm {s}} = \\beta_ {t} * \\nabla \\mathbf {T} _ {0 | t} + (1 - \\beta_ {t}) * \\nabla \\mathbf {M}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 267, + 794, + 785, + 811 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "L-DiffER", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2 Details of $\\beta_{t}$ and $\\gamma_{t}$ will be explained in the supplementary material.", + "bbox": [ + 217, + 824, + 689, + 839 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f44879e18feae278d2a21e482f7d1266cac3258b01cbc0116e5b4627b9053417.jpg", + "image_caption": [ + "Fig. 4: Visualization of the color condition, structure condition, and the pseudo transmission layer by applying the proposed iterative condition refinement strategy." + ], + "image_footnote": [], + "bbox": [ + 220, + 147, + 784, + 344 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 4 shows intermediate conditions and the pseudo transmission layers during sampling, indicating the effect of the iterative condition refinement strategy.", + "bbox": [ + 212, + 412, + 782, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Meanwhile, we also use the iterative condition refinement strategy with a few modifications during training. We replace the pseudo-accurate transmission layers $\\mathbf{T}_{0|t}$ estimated from sampling with ground truth $\\mathbf{T}$ to refine conditions before estimating noises, which is different from the sampling process. Specifically, given the initial mixture image $\\mathbf{M}$ and its gradient $\\nabla \\mathbf{M}$ , we refine the input conditions at timestep $t$ as follows:", + "bbox": [ + 212, + 441, + 784, + 532 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathbf {I}} _ {t} ^ {\\mathrm {c}} = \\gamma_ {t} * \\mathbf {T} + (1 - \\gamma_ {t}) * \\mathbf {M}, \\quad \\widehat {\\mathbf {I}} _ {t} ^ {\\mathrm {s}} = \\beta_ {t} * \\nabla \\mathbf {T} + (1 - \\beta_ {t}) * \\nabla \\mathbf {M}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 281, + 542, + 784, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\beta_{t}$ and $\\gamma_{t}$ ( $t \\in \\{1, \\dots, T\\}$ ) are the same predefined time-variant coefficients used in the sampling process.", + "bbox": [ + 212, + 570, + 782, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Multi-condition constraint mechanism", + "text_level": 1, + "bbox": [ + 214, + 619, + 575, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In order to mitigate the insufficient recovery faithfulness of diffusion models, we employ a multi-condition mechanism that separately utilizes mixture images and their gradients to obtain initial latent conditions for controlling the color and structure of generated results. As the iterative condition refinement strategy obtains new refined conditions during each timestep, no constraints applying on conditions may cause the sampling process out of control, which also impacts the recovery faithfulness. Hence, it is imperative to conduct constraints on refined conditions. Though inaccurate as color conditions, mixture images provide coarse color information about transmission layers, thus we propose a maximum value constraint to prevent color shifts. Similarly, the gradient of the mixture image provides features showing structure information such as edges and textures. Therefore, we propose a saturation-aware structure constraint to prevent structure distortions when recovering transmission layers. Details are as follows.", + "bbox": [ + 212, + 643, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Hong et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Maximum value constraint (MVC). According to the assumption defined in Sec. 3.1, the pixel values of the mixture image should be greater than or equal to the transmission layer. Therefore, when updating the conditions in the training and sampling processes, we employ maximum value constraint as a function to ensure that any given color condition $\\mathbf{I}_t^{\\mathrm{c}}$ defined in Eq. (4) does not exceed the mixture image $\\mathbf{M}$ in terms of pixel values. This constraint should be adhered to the following:", + "bbox": [ + 214, + 146, + 782, + 252 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M V C} \\left(\\mathbf {I} _ {t} ^ {\\mathrm {c}}\\right) = \\min \\left(\\mathbf {M}, \\mathbf {I} _ {t} ^ {\\mathrm {c}}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 262, + 784, + 277 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Saturation-aware structure constraint (SSC). Due to regional overexposure during photographing, mixture images sometimes contain low-transmitted reflection regions, which lack enough visible content of transmission layers, causing the corresponding gradients to have little usable structure information. Therefore, simply applying the maximum value constraint to edges will force the refined structure condition to abandon generated content in saturated regions, which may hinder optimization. Therefore, we introduce a saturation-aware mask to preserve the generative capability of the structural aspects in regions challenging for recovery. Specifically, we first utilize the mixture image to identify low-transmitted regions, generating a mask $\\mathbf{M}^{\\mathrm{s}}\\in \\mathbb{R}^{H\\times W}$ as:", + "bbox": [ + 214, + 287, + 784, + 436 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Omega} ^ {\\mathrm {s}} = \\left\\{ \\begin{array}{l l} 0, & \\text {w h e r e} \\mathbf {M} > \\tau \\\\ 1, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 446, + 784, + 484 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{M}$ is the mixture image, $\\tau$ is the saturation threshold we empirically set to 0.95. Besides, to ensure that the structure condition does not contain nonexistent edges in non-saturated regions of the original mixture image, we define a mask to indicate the valid edge in the gradient of the mixture image as:", + "bbox": [ + 214, + 491, + 782, + 551 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Omega} ^ {\\mathrm {v}} = \\left\\{ \\begin{array}{l l} 0, & \\text {w h e r e} \\nabla \\mathbf {M} < \\eta \\\\ 1, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 558, + 784, + 598 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\eta$ is the valid edge threshold we empirically set to 0.05. After that, for a given gradient condition $\\mathbf{I}_t^s$ defined in Eq. (4), we define the saturation-aware structure constraint as follows:", + "bbox": [ + 214, + 604, + 782, + 647 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S S C} \\left(\\mathbf {I} _ {t} ^ {\\mathrm {s}}\\right) = \\boldsymbol {\\Omega} ^ {\\mathrm {s}} \\odot \\mathbf {I} _ {t} ^ {\\mathrm {s}} + \\left(1 - \\boldsymbol {\\Omega} ^ {\\mathrm {s}}\\right) \\odot \\boldsymbol {\\Omega} ^ {\\mathrm {v}} \\odot \\mathbf {I} _ {t} ^ {\\mathrm {s}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 659, + 784, + 674 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To extract the color latent $y_{t}^{\\mathrm{c}}$ and the structure latent $y_{t}^{\\mathrm{s}}$ at each timestep $t$ , we apply the maximum value constraint (MVC) and saturation-aware structure constraint (SSC) on the color condition $\\mathbf{I}_t^{\\mathrm{c}}$ and structure condition $\\mathbf{I}_t^{\\mathrm{s}}$ as:", + "bbox": [ + 214, + 684, + 782, + 729 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\ny _ {t - 1} ^ {\\mathrm {c}} = \\mathcal {E} ^ {\\mathrm {c}} \\left(\\operatorname {M V C} \\left(\\mathbf {I} _ {t} ^ {\\mathrm {c}}\\right)\\right), \\quad y _ {t - 1} ^ {\\mathrm {s}} = \\mathcal {E} ^ {\\mathrm {s}} \\left(\\operatorname {S S C} \\left(\\mathbf {I} _ {t} ^ {\\mathrm {s}}\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 739, + 784, + 755 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathcal{E}^{\\mathrm{s}}$ is the structure condition encoder, $\\mathcal{E}^{\\mathrm{c}}$ is the color condition encoder, and MVC and SSC are applied to the color and structure conditions for ensuring recovery faithfulness. Finally, we apply the denoising process with the DDIM [53] sampling scheme to obtain the next noisy sample $z_{t-1}$ . Algorithm 1 outlines the complete sampling process.", + "bbox": [ + 214, + 763, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "L-DiffER", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Iterative condition refinement strategy in the sampling process" + ], + "code_body": "Require: Noise prediction network $\\epsilon_{\\theta}$ Compression decoder $\\mathcal{D}$ Color encoder $\\mathcal{E}^{\\mathrm{c}}$ Structure encoder $\\mathcal{E}^{\\mathrm{s}}$ \nInput: Mixture image M, Language condition $c = [c^{\\mathrm{p}},c^{\\mathrm{n}}]$ , Initial spatial conditions $[y^{\\mathrm{s}} = \\mathcal{E}^{\\mathrm{s}}(\\nabla \\mathbf{M}),y^{\\mathrm{c}} = \\mathcal{E}^{\\mathrm{c}}(\\mathbf{M})]$ , Coefficients $\\{\\overline{\\alpha}_t\\}_{t = 1}^T$ $\\{\\beta_t\\}_{t = 1}^T$ $\\{\\gamma_t\\}_{t = 1}^T$ \nOutput: Recovered latent representation $z_{0}$ \n1: $z_{T}\\sim \\mathcal{N}(0,1)$ \n2: $y_{T} = [y^{s},y^{c}]$ \n3: for each $t\\in [T,1]$ do \n4: $\\hat{\\epsilon}_t = \\epsilon_\\theta (z_t,c,y_t,t)$ \n5: if $t > 1$ then \n6: $z_{0|t} = (z_t - \\sqrt{1 - \\overline{\\alpha}_t} *\\hat{\\epsilon}_t) / \\sqrt{\\overline{\\alpha}_t}$ D Compute $z_{0|t}$ by Eq. (3) \n7: $\\mathbf{T}_{0|t} = \\mathcal{D}(z_{0|t})$ \n8: $y_{t - 1}^{\\mathrm{c}} = \\mathcal{E}^{\\mathrm{c}}(\\mathrm{MVC}(\\gamma_t*\\mathbf{T}_{0|t} + (1 - \\gamma_t)*\\mathbf{M}))$ DRefine color condition \n9: $y_{t - 1}^{\\mathrm{s}} = \\mathcal{E}^{\\mathrm{s}}(\\mathrm{SSC}(\\beta_t*\\nabla \\mathbf{T}_{0|t} + (1 - \\beta_t)*\\nabla \\mathbf{M}))$ DRefine structure condition \n10: $y_{t - 1} = [y_{t - 1}^{\\mathrm{s}},y_{t - 1}^{\\mathrm{c}}]$ DUpdate spatial conditions \n11: end if \n12: $z_{t - 1} = \\sqrt{\\overline{\\alpha}_{t - 1}} z_{0|t} + \\sqrt{1 - \\overline{\\alpha}_{t - 1}}\\hat{\\epsilon}_t$ DDIM \n13: end for", + "bbox": [ + 215, + 164, + 785, + 415 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3.4 Loss functions", + "text_level": 1, + "bbox": [ + 215, + 445, + 379, + 458 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we introduce the latent diffusion model loss to minimize the discrepancy between the learned latent space distribution and the target distribution, as well as the pixel losses to prevent color shifts and structure distortions.", + "bbox": [ + 212, + 469, + 784, + 513 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We follow the training objective of ControlNet [74] but replace task-specific conditions with the mixture image and the gradient of the mixture. Thus our network $\\epsilon_{\\theta}$ learns to predict the noise added to the noisy latent $z_{t}$ with $\\mathcal{L}_{\\mathrm{ldm}}$ defined in Eq. (1). According to Eq. (3), we can obtain $\\mathbf{T}_{0|t}$ in pixel space by decoding $z_{0|t}$ :", + "bbox": [ + 212, + 513, + 784, + 590 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {0 | t} = \\mathcal {D} \\left(z _ {0 | t}\\right) = \\mathcal {D} \\left(\\left(z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} * \\hat {\\epsilon} _ {t}\\right) / \\sqrt {\\bar {\\alpha} _ {t}}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 602, + 784, + 619 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Then RGB loss is introduced to constrain the consistency of the reconstructed transmission $\\mathbf{T}_{0|t}$ and ground truth $\\mathbf{T}$ , minimizing their errors as follows:", + "bbox": [ + 215, + 631, + 782, + 662 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {R G B}} = \\left\\| \\mathbf {T} - \\mathbf {T} _ {0 | t} \\right\\| _ {2} ^ {2}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 672, + 784, + 690 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For better structure consistency, we use gradient loss to constrain the prediction's gradient and its ground truth as below:", + "bbox": [ + 215, + 700, + 782, + 729 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {g r a d}} = \\left\\| \\nabla (\\mathbf {T}) - \\nabla \\left(\\mathbf {T} _ {0 | t}\\right) \\right\\| _ {2} ^ {2}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 739, + 784, + 758 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\nabla (\\cdot)$ denotes the computation of image gradients. Furthermore, we also compute the standard deviation and the mean of the predicted transmission and its ground truth, and use a $\\ell_2$ loss to minimize the color distortion:", + "bbox": [ + 212, + 768, + 784, + 813 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {n u m}} = \\left\\| \\mu (\\mathbf {T}) - \\mu \\left(\\mathbf {T} _ {0 | t}\\right) \\right\\| _ {2} ^ {2} + \\left\\| \\sigma (\\mathbf {T}) - \\sigma \\left(\\mathbf {T} _ {0 | t}\\right) \\right\\| _ {2} ^ {2}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 325, + 823, + 784, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Hong et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\mu (\\cdot)$ and $\\sigma (\\cdot)$ denote the computation of mean and standard deviation, respectively, and $i = 1,\\dots,N$ denotes the $i$ -th sample in the dataset. Overall, the pixel loss is formulated as follows:", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {p i x}} = \\mathcal {L} _ {\\mathrm {R G B}} + \\mathcal {L} _ {\\mathrm {g r a d}} + \\mathcal {L} _ {\\mathrm {n u m}}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 205, + 784, + 219 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Gathering the latent diffusion model loss and pixel losses yields the final objective as:", + "bbox": [ + 212, + 231, + 782, + 260 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l l}} := \\mathcal {L} _ {\\mathrm {l d m}} + \\lambda \\mathcal {L} _ {\\text {p i x}}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 262, + 784, + 277 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $\\lambda$ is set to 1.0 empirically.", + "bbox": [ + 214, + 284, + 450, + 299 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "3.5 Implementation details", + "text_level": 1, + "bbox": [ + 214, + 321, + 452, + 335 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We implement the proposed method using PyTorch [46]. The model is trained for 20 epochs with a batch size of 16 on two NVIDIA GeForce RTX 3090 GPUs. Weights are initialized as in [11] and updated using the Adam optimizer [25]. The learning rate is set to $1 \\times 10^{-5}$ . For sampling, we use the DDIM sampling method [53] with 50 sampling steps.", + "bbox": [ + 212, + 345, + 784, + 421 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Dataset. Our training dataset contains both synthetic and real data. For the synthetic data, following Zhong et al. [81], we generate 50000 triplets of data using COCO Captions [6] and Flickr30k [73] for network training. For real data, we use 200 image pairs from Nature dataset [31] and 90 pairs from Zhang et al. [76] for training and the rest of the real data are used for evaluation. We further collect 50 mixture images from the Internet for qualitative evaluation.", + "bbox": [ + 212, + 422, + 785, + 512 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 535, + 375, + 553 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Comparison with state-of-the-arts", + "text_level": 1, + "bbox": [ + 214, + 566, + 540, + 583 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For evaluating model performances, experiments are conducted on existing real reflection datasets [31, 58, 76] and our collected mixture images from the Internet (with manually annotated language descriptions). We compare the proposed L-DiffER with state-of-the-art single-image reflection removal methods, including Zhang et al. [76], CoRRN [61], ERRNet [65], IBCLN [31], Dong et al. [8], YTMT [21], and DSRNet [22]. To compare with a representative language-based diffusion model ControlNet $[74]^3$ , we make it fit our setting by using mixture images as conditions, and transmission and reflection descriptions as positive and negative prompts, respectively. For fair comparisons, we finetune the above methods on our training data if their training codes are provided.", + "bbox": [ + 212, + 592, + 784, + 742 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative comparison. Quantitative experiments are conducted on three real datasets for reflection removal, i.e., Nature [31], Real20 [76], and SIR² [58] dataset. To evaluate the recovery of transmission layers⁴, we utilize PSNR [23]", + "bbox": [ + 212, + 743, + 784, + 789 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "L-DiffER", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_footnote", + "text": "3 Since ControlNet [74] destructs image color and structure in a generative manner as shown in Fig. 1(c), we only run it for qualitative comparisons.", + "bbox": [ + 217, + 796, + 782, + 825 + ], + "page_idx": 10 + }, + { + "type": "page_footnote", + "text": "4 Evaluations on reflection layers are provided in the supplementary material.", + "bbox": [ + 218, + 825, + 736, + 839 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/81a80b2c51905625dccd26efa85d07feb5843446f91bae90a1401a29c96053e9.jpg", + "table_caption": [ + "Table 1: Comparison of quantitative results on real datasets for evaluating the recovery of transmission layers, compared with several state-of-the-art single-image reflection removal methods [8,21,22,31,61,76]. $\\uparrow (\\downarrow)$ indicates larger (smaller) values are better. Bold numbers indicate the best-performing results." + ], + "table_footnote": [], + "table_body": "
Dataset (size)MetricMethod
Zhang et al.CoRRNERRNetIBCLNDong et al.YTMTDSRNetOurs
SIR2(500)PSNR↑22.4522.9623.1323.3623.0923.0524.9725.18
SSIM↑0.8720.8790.8780.8810.8930.8860.9070.911
LPIPS↓0.1720.1610.1640.1530.1310.1490.1240.121
NIQE↓4.9114.8264.7934.5014.7294.7384.6244.352
FID↓78.6169.5873.3360.9249.3958.6946.6044.66
Real20(20)PSNR↑22.5121.1722.0621.5921.7322.3123.4623.77
SSIM↑0.8060.7860.8030.7710.8110.8050.8060.821
LPIPS↓0.2040.2020.1850.2130.1700.1780.1650.153
NIQE↓3.8193.8913.9893.8953.9593.9774.0653.665
FID↓113.25111.2585.07120.9586.2094.6777.9272.64
Nature(20)PSNR↑20.3720.5421.1123.6923.6121.0321.7023.95
SSIM↑0.7720.7780.8060.8280.8250.8020.8200.831
LPIPS↓0.2170.2050.1830.1700.1570.1860.1680.145
NIQE↓4.4134.4334.4424.3454.5814.3874.3734.223
FID↓109.7789.1279.6875.1771.5581.7774.8468.57
Average(540)PSNR↑22.3822.8023.0223.3123.0622.9524.7925.08
SSIM↑0.8660.8720.8730.8750.8870.8800.9000.905
LPIPS↓0.1750.1640.1650.1560.1330.1510.1270.123
NIQE↓4.8524.7774.7504.4734.6954.6974.5944.322
FID↓81.0571.8574.0063.6751.5760.8848.8146.58
", + "bbox": [ + 215, + 212, + 785, + 513 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "and SSIM [64] as error metrics following reflection removal methods [8, 37]. We further adopts LPIPS [75], NIQE [41], FID [13] to measure the perceptual quality of the recovered results. As quantitative results shown in Table 1, the proposed method achieves the best performance among all competing methods, especially in metrics for perceptual quality, validating its effectiveness.", + "bbox": [ + 215, + 559, + 784, + 635 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative comparison. Qualitative comparisons with the aforementioned single-image reflection removal methods [8,21,22,31,61,76] and a language-based diffusion model ControlNet [74] on recovering transmission layers are shown in Fig. 5. As can be observed, single-image reflection removal methods encounter obstacles in dealing with complex reflections due to the lack of auxiliary semantic information, especially in low-transmitted reflection regions. ControlNet [74] performs significant modifications on the image color and structure in a generative manner, which deviates significantly from the input mixture image, emphasizing the pivotal role of accurate conditions in the sampling process. The proposed method generates visually pleasant reflection removal results with faithful recovery, demonstrating that it capitalizes on the synergy of the auxiliary information from language descriptions, the iterative condition refinement strategy, and the multi-condition constraint mechanism.", + "bbox": [ + 215, + 643, + 785, + 838 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Hong et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/73184fab4595eccaa6f995c8d382bb86c54b531f24f3d23e992c763458005f97.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 145, + 500, + 262 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2e9ec1b2303725df80c73491b414317a0cb5fb966076efc690d9d5141a97cefc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 145, + 784, + 262 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2f5b523a057505ec3b3e0e96ea137bafe12276199accfb90eb3f961c3238ddfa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 263, + 500, + 441 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0b29f5ba22d4dd74e195cec4ab76e7ecbe7210d936be5db5c47da2da13a2cd90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 263, + 784, + 441 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/63c56fa02617f6d90b54a50fd5b7bd660d4cb724c5855eaf87b8f3dd03a53098.jpg", + "image_caption": [ + "Fig. 5: Qualitative comparison of estimated transmission layers on real mixture images collected from the Internet, compared with several single-image methods [8,21,22,31, 61,76] and a diffusion-based method ControlNet [74]. Please zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 218, + 441, + 500, + 513 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e4f183a36e8483299ec0751ecc35af2c6ddfc95f1e0b14567a8477f2b1a3809d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 441, + 784, + 513 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.2 Ablation study", + "text_level": 1, + "bbox": [ + 215, + 601, + 385, + 617 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation on control conditions5. We conduct ablation studies to investigate the effectiveness of the iterative condition refinement strategy and the multi-condition constraint mechanism. As shown in Table 2 and Fig. 6, most reflection contaminations are retained when missing iterative condition refinement (W/o rfn.), and disabling the refinement on color (W/o clr. rfn.) or structure (W/o str. rfn.) will cause inaccurate recovery, which exists color shift or retains the reflection content. In addition, as shown in Table 2 and Fig. 7, due to the generative prior of the diffusion model, the absence of the max value constraint (W/o MVC) or the saturation-aware structure constraint (W/o SSC), or both (W/o cnst.) leads to obvious color shifts, structure distortion, and uncontrolled recovery. Therefore, the degradation observed in the above ablations indicates the necessity of the refinement and constraints on conditions.", + "bbox": [ + 212, + 631, + 787, + 811 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "L-DiffER", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_footnote", + "text": "5 More ablation studies are provided in the supplementary material.", + "bbox": [ + 217, + 824, + 676, + 840 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0a31a070e5b7c8d0797db74623108abe61c38538be89b324e6cae0bf3a8969cb.jpg", + "table_caption": [ + "Table 2: Ablation studies on the iterative condition refinement strategy and the multi-condition constraint mechanism. $\\uparrow (\\downarrow)$ indicates larger (smaller) values are better. Bold numbers indicate the best-performing results." + ], + "table_footnote": [], + "table_body": "
MetricAbl. on iterative condition refinementAbl. on multi-condition constraint
W/o clr. rfn.W/o str. rfn.W/o rfn.W/o MVCW/o SSCW/o cnst.Ours
PSNR↑24.4224.6824.3924.7324.1623.3525.08
SSIM↑0.8980.8960.8910.8970.8870.8790.905
LPIPS↓0.1320.1360.1390.1290.1430.1540.123
NIQE↓4.5014.6454.8684.5534.7094.5614.322
FID↓49.1353.7160.1251.0656.4663.6846.58
", + "bbox": [ + 215, + 198, + 785, + 301 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9f86f49c4ca7b66023cde984905d025df4a3b69d9fefa01730740d5ec21415ff.jpg", + "image_caption": [ + "Fig. 6: The effect of the iterative condition refinement strategy. For each result, we place the corresponding final color and structure condition at its lower left and lower right. Please zoom in for details, respectively." + ], + "image_footnote": [], + "bbox": [ + 218, + 316, + 784, + 393 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0e2c9733bd9676f7a4b1d69337c22d0ab1bbc1100f7b21840669a062f2828af0.jpg", + "image_caption": [ + "Fig. 7: The effect of the color and structure constraints in the multi-condition constraint mechanism. For each result, we place the corresponding final color and structure condition at its lower left and lower right, respectively. Please zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 220, + 463, + 784, + 539 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 621, + 359, + 636 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper introduces the first language-based diffusion model for single image reflection removal, which exploits auxiliary semantic information from language descriptions and generative priors from diffusion models to recover transmission layers from mixture images with low-transmitted reflection regions. An iterative condition refinement strategy is proposed to address the problem of inaccurate control conditions, and a multi-condition constraint mechanism is introduced to ensure recovery faithfulness. Experiments on real data demonstrate the effectiveness of the proposed method.", + "bbox": [ + 212, + 657, + 787, + 777 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. The proposed method may incorrectly identify transmission and reflection layers if they exhibit similar image content. For such cases, more accurate spatial guidance from users such as annotated edges or region masks could help to relieve the ambiguity, which is left as our future work.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Hong et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported by National Natural Science Foundation of China under Grant No. 62136001, 62088102.", + "bbox": [ + 215, + 176, + 782, + 205 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 229, + 321, + 244 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Chang, Y., Jung, C., Sun, J.: Joint reflection removal and depth estimation from a single image. IEEE Transactions on Cybernetics (2020)", + "2. Chang, Y., Jung, C., Sun, J., Wang, F.: Siamese dense network for reflection removal with flash and no-flash image pairs. International Journal of Computer Vision (2020)", + "3. Chang, Z., Weng, S., Li, Y., Li, S., Shi, B.: L-CoDer: Language-based colorization with color-object decoupling transformer. In: Proc. of European Conference on Computer Vision (2022)", + "4. Chang, Z., Weng, S., Zhang, P., Li, Y., Li, S., Shi, B.: L-CAD: Language-based colorization with any-level descriptions using diffusion priors. In: Proc. of Advances in Neural Information Processing Systems (2023)", + "5. Chang, Z., Weng, S., Zhang, P., Li, Y., Li, S., Shi, B.: L-CoIns: Language-based colorization with instance awareness. In: Proc. of Computer Vision and Pattern Recognition (2023)", + "6. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dolkar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015)", + "7. Diamant, Y., Schechner, Y.Y.: Overcoming visual reverberations. In: Proc. of Computer Vision and Pattern Recognition (2008)", + "8. Dong, Z., Xu, K., Yang, Y., Bao, H., Xu, W., Lau, R.W.: Location-aware single image reflection removal. In: Proc. of International Conference on Computer Vision (2021)", + "9. Fan, Q., Yang, J., Hua, G., Chen, B., Wipf, D.: A generic deep architecture for single image reflection removal and image smoothing. In: Proc. of International Conference on Computer Vision (2017)", + "10. Han, B.J., Sim, J.Y.: Zero-shot learning for reflection removal of single 360-degree image. In: Proc. of European Conference on Computer Vision (2022)", + "11. He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In: Proc. of International Conference on Computer Vision (2015)", + "12. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022)", + "13. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Proc. of Advances in Neural Information Processing Systems (2017)", + "14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Proc. of Advances in Neural Information Processing Systems (2020)", + "15. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)", + "16. Hong, Y., Chang, Y., Liang, J., Ma, L., Huang, T., Shi, B.: Light flickering guided reflection removal. International Journal of Computer Vision (2024)" + ], + "bbox": [ + 217, + 260, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "L-DiffER", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Hong, Y., Lyu, Y., Li, S., Cao, G., Shi, B.: Reflection removal with NIR and RGB image feature fusion. IEEE Transactions on Multimedia (2022)", + "18. Hong, Y., Lyu, Y., Li, S., Shi, B.: Near-infrared image guided reflection removal. In: Proc. of International Conference on Multimedia and Expo (2020)", + "19. Hong, Y., Zheng, Q., Zhao, L., Jiang, X., Kot, A.C., Shi, B.: Panoramic image reflection removal. In: Proc. of Computer Vision and Pattern Recognition (2021)", + "20. Hong, Y., Zheng, Q., Zhao, L., Jiang, X., Kot, A.C., Shi, B.: PAR $^2$ Net: End-to-end panoramic image reflection removal. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "21. Hu, Q., Guo, X.:trash or treasure? an interactive dual-stream strategy for single image reflection separation. Proc. of Advances in Neural Information Processing Systems (2021)", + "22. Hu, Q., Guo, X.: Single image reflection separation via component synergy. In: Proc. of International Conference on Computer Vision (2023)", + "23. Huynh-Thu, Q., Ghanbari, M.: Scope of validity of psnr in image/video quality assessment. Electronics letters (2008)", + "24. Kim, S., Huo, Y., Yoon, S.E.: Single image reflection removal with physically-based training images. In: Proc. of Computer Vision and Pattern Recognition (2020)", + "25. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "26. Kong, N., Tai, Y.W., Shin, S.Y.: A physically-based approach to reflection separation. In: Proc. of Computer Vision and Pattern Recognition (2012)", + "27. Lei, C., Chen, Q.: Robust reflection removal with reflection-free flash-only cues. In: Proc. of Computer Vision and Pattern Recognition (2021)", + "28. Lei, C., Huang, X., Zhang, M., Yan, Q., Sun, W., Chen, Q.: Polarized reflection removal with perfect alignment in the wild. In: Proc. of Computer Vision and Pattern Recognition (2020)", + "29. Lei, C., Jiang, X., Chen, Q.: Robust reflection removal with flash-only cues in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "30. Levin, A., Weiss, Y.: User assisted separation of reflections from a single image using a sparsity prior. IEEE Transactions on Pattern Analysis and Machine Intelligence 29(9), 1647-1654 (2007)", + "31. Li, C., Yang, Y., He, K., Lin, S., Hopcroft, J.E.: Single image reflection removal through cascaded refinement. In: Proc. of Computer Vision and Pattern Recognition (2020)", + "32. Li, Y., Brown, M.S.: Exploiting reflection change for automatic reflection removal. In: Proc. of International Conference on Computer Vision (2013)", + "33. Li, Y., Brown, M.S.: Single image layer separation using relative smoothness. In: Proc. of Computer Vision and Pattern Recognition (2014)", + "34. Liu, Y.L., Lai, W.S., Yang, M.H., Chuang, Y.Y., Huang, J.B.: Learning to see through obstructions. In: Proc. of Computer Vision and Pattern Recognition (2020)", + "35. Liu, Y.L., Lai, W.S., Yang, M.H., Chuang, Y.Y., Huang, J.B.: Learning to see through obstructions with layered decomposition. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)", + "36. Luo, J., Fu, J., Kong, X., Gao, C., Ren, H., Shen, H., Xia, H., Liu, S.: 3D-SPS: Single-stage 3D visual grounding via referred point progressive selection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16454-16463 (June 2022)", + "37. Lyu, Y., Cui, Z., Li, S., Pollefeys, M., Shi, B.: Reflection separation using a pair of unpolarized and polarized images. In: Proc. of Advances in Neural Information Processing Systems (2019)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Hong et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. Lyu, Y., Cui, Z., Li, S., Pollefeys, M., Shi, B.: Physics-guided reflection separation from a pair of unpolarized and polarized images. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)", + "39. Ma, D., Wan, R., Shi, B., Kot, A.C., Duan, L.Y.: Learning to jointly generate and separate reflections. In: Proc. of International Conference on Computer Vision (2019)", + "40. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: Proc. of International Conference on Learning Representations (2021)", + "41. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters 20(3), 209-212 (2013)", + "42. Mou, C., Wang, X., Xie, L., Wu, Y., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)", + "43. Nam, S., Brubaker, M.A., Brown, M.S.: Neural image representations for multi-image fusion and layer separation. In: Proc. of European Conference on Computer Vision (2022)", + "44. Nayar, S.K., Fang, X.S., Boult, T.: Separation of reflection components using color and polarization. International Journal of Computer Vision (1997)", + "45. Park, J., Kim, H., Park, E., Sim, J.Y.: Fully-automatic reflection removal for 360-degree images. In: Proc. of Winter Conference on Applications of Computer Vision (2024)", + "46. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: Proc. of Advances in Neural Information Processing Systems (2019)", + "47. Qiu, J., Jiang, P.T., Zhu, Y., Yin, Z.X., Cheng, M.M., Ren, B.: Looking through the glass: Neural surface reconstruction against high specular reflections. In: Proc. of Computer Vision and Pattern Recognition (2023)", + "48. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proc. of International Conference on Machine Learning. PMLR (2021)", + "49. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proc. of Computer Vision and Pattern Recognition (2022)", + "50. Schechner, Y.Y., Kiryati, N., Basri, R.: Separation of transparent layers using focus. International Journal of Computer Vision (2000)", + "51. Shih, Y., Krishnan, D., Durand, F., Freeman, W.T.: Reflection removal using ghosting cues. In: Proc. of Computer Vision and Pattern Recognition (2015)", + "52. Simon, C., Kyu Park, I.: Reflection removal for in-vehicle black box videos. In: Proc. of Computer Vision and Pattern Recognition (2015)", + "53. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)", + "54. Sun, H., Li, W., Liu, J., Chen, H., Pei, R., Zou, X., Yan, Y., Yang, Y.: CoSeR: Bridging image and language for cognitive super-resolution. arXiv preprint arXiv:2311.16512 (2023)", + "55. Sun, J., Weng, S., Chang, Z., Li, S., Shi, B.: UniCoRN: A unified conditional image repainting network. In: Proc. of Computer Vision and Pattern Recognition (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "L-DiffER", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "56. Tang, J., Zhong, H., Weng, S., Shi, B.: LuminAIRe: Illumination-aware conditional image painting for lighting-realistic generation. In: Proc. of Advances in Neural Information Processing Systems (2023)", + "57. Wan, R., Shi, B., Duan, L.Y., Tan, A.H., Gao, W., Kot, A.C.: Region-aware reflection removal with unified content and gradient priors. IEEE Transactions on Image Processing (2018)", + "58. Wan, R., Shi, B., Duan, L.Y., Tan, A.H., Kot, A.C.: Benchmarking single-image reflection removal algorithms. In: Proc. of International Conference on Computer Vision (2017)", + "59. Wan, R., Shi, B., Duan, L.Y., Tan, A.H., Kot, A.C.: CRRN: Multi-scale guided concurrent reflection removal network. In: Proc. of Computer Vision and Pattern Recognition (2018)", + "60. Wan, R., Shi, B., Li, H., Duan, L.Y., Kot, A.C.: Face image reflection removal. International Journal of Computer Vision (2021)", + "61. Wan, R., Shi, B., Li, H., Duan, L.Y., Tan, A.H., Kot, A.C.: CoRRN: Cooperative reflection removal network. IEEE Transactions on Pattern Analysis and Machine Intelligence (2019)", + "62. Wan, R., Shi, B., Li, H., Hong, Y., Duan, L.Y., Kot, A.C.: Benchmarking single-image reflection removal algorithms. In: IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)", + "63. Wang, Z., Lu, Y., Li, Q., Tao, X., Guo, Y., Gong, M., Liu, T.: CRIS: CLIP-driven referring image segmentation. In: Proc. of Computer Vision and Pattern Recognition (2022)", + "64. Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers (2003)", + "65. Wei, K., Yang, J., Fu, Y., Wipf, D., Huang, H.: Single image reflection removal exploiting misaligned training data and network enhancements. In: Proc. of Computer Vision and Pattern Recognition (2019)", + "66. Wen, Q., Tan, Y., Qin, J., Liu, W., Han, G., He, S.: Single image reflection removal beyond linearity. In: Proc. of Computer Vision and Pattern Recognition (2019)", + "67. Weng, S., Li, W., Li, D., Jin, H., Shi, B.: MISC: Multi-condition injection and spatially-adaptive compositing for conditional person image synthesis. In: Proc. of Computer Vision and Pattern Recognition (2020)", + "68. Weng, S., Shi, B.: Conditional image inpainting. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "69. Weng, S., Wu, H., Chang, Z., Tang, J., Li, S., Shi, B.: L-CoDe: Language-based colorization using color-object decoupled conditions. In: Proc. of the AAAI Conference on Artificial Intelligence (2022)", + "70. Yang, J., Gong, D., Liu, L., Shi, Q.: Seeing deeply and bidirectionally: A deep learning approach for single image reflection removal. In: Proc. of European Conference on Computer Vision (2018)", + "71. Yang, Y., Ma, W., Zheng, Y., Cai, J.F., Xu, W.: Fast single image reflection suppression via convex optimization. In: Proc. of Computer Vision and Pattern Recognition (2019)", + "72. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: LAVT: Language-aware vision transformer for referring image segmentation. In: Proc. of Computer Vision and Pattern Recognition (2022)", + "73. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics (2014)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Hong et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "74. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proc. of International Conference on Computer Vision (2023)", + "75. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proc. of Computer Vision and Pattern Recognition (2018)", + "76. Zhang, X., Ng, R., Chen, Q.: Single image reflection separation with perceptual losses. In: Proc. of Computer Vision and Pattern Recognition (2018)", + "77. Zhang, Y.N., Shen, L., Li, Q.: Content and gradient model-driven deep network for single image reflection removal. In: Proc. of ACM International Conference on Multimedia (2022)", + "78. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Unictrlnet: All-in-one control to text-to-image diffusion models. Proc. of Advances in Neural Information Processing Systems (2024)", + "79. Zheng, Q., Chen, J., Lu, Z., Shi, B., Jiang, X., Yap, K.H., Duan, L.Y., Kot, A.C.: What does plate glass reveal about camera calibration? In: Proc. of Computer Vision and Pattern Recognition (2020)", + "80. Zheng, Q., Shi, B., Chen, J., Jiang, X., Duan, L.Y., Kot, A.C.: Single image reflection removal with absorption effect. In: Proc. of Computer Vision and Pattern Recognition (2021)", + "81. Zhong, H., Hong, Y., Weng, S., Liang, J., Shi, B.: Language-guided image reflection separation. In: Proc. of Computer Vision and Pattern Recognition (2024)" + ], + "bbox": [ + 212, + 146, + 787, + 439 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "L-DiffER", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/L-DiffER_ Single Image Reflection Removal with Language-based Diffusion Model/84990c91-4cd8-4a1b-83d4-8e97d608b20a_model.json b/2024/L-DiffER_ Single Image Reflection Removal with Language-based Diffusion Model/84990c91-4cd8-4a1b-83d4-8e97d608b20a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..61073cb42d736a3325d9da4d32ad5e97603449b7 --- /dev/null +++ b/2024/L-DiffER_ Single Image Reflection Removal with Language-based Diffusion Model/84990c91-4cd8-4a1b-83d4-8e97d608b20a_model.json @@ -0,0 +1,2757 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.255, + 0.141, + 0.75, + 0.187 + ], + "angle": 0, + "content": "L-DiffER: Single Image Reflection Removal with Language-based Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.211, + 0.717, + 0.245 + ], + "angle": 0, + "content": "Yuchen Hong\\(^{1,2\\#}\\) Haofeng Zhong\\(^{1,2,3\\#}\\) Shuchen Weng\\(^{1,2}\\) Jinxiu Liang\\(^{1,2}\\) Boxin Shi\\(^{1,2,3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.288, + 0.255, + 0.714, + 0.284 + ], + "angle": 0, + "content": "\\(^{1}\\) State Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.283, + 0.707, + 0.311 + ], + "angle": 0, + "content": "\\(^{2}\\) National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.311, + 0.739, + 0.354 + ], + "angle": 0, + "content": "\\(^{3}\\) AI Innovation Center, School of Computer Science, Peking University {hfzhong, shuchenweng, cssherryliang, shiboxin}@pku.edu.cn, yuchenhong.cn@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.39, + 0.744, + 0.558 + ], + "angle": 0, + "content": "Abstract. In this paper, we introduce L-DiffER, a language-based diffusion model designed for the ill-posed single image reflection removal task. Although having shown impressive performance for image generation, existing language-based diffusion models struggle with precise control and faithfulness in image restoration. To overcome these limitations, we propose an iterative condition refinement strategy to resolve the problem of inaccurate control conditions. A multi-condition constraint mechanism is employed to ensure the recovery faithfulness of image color and structure while retaining the generation capability to handle low-transmitted reflections. We demonstrate the superiority of the proposed method through extensive experiments, showcasing both quantitative and qualitative improvements over existing methods." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.57, + 0.702, + 0.585 + ], + "angle": 0, + "content": "Keywords: Reflection removal \\(\\cdot\\) Language-based diffusion model" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.633, + 0.376, + 0.649 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.665, + 0.788, + 0.816 + ], + "angle": 0, + "content": "When photographs are taken through transparent mediums like glass windows, the contaminated mixture image (denoted as \\(\\mathbf{M}\\)) can be considered as the combination of a transmission layer (denoted as \\(\\mathbf{T}\\)) and a reflection layer (denoted as \\(\\mathbf{R}\\)) [61, 62], which impairs the performance of downstream computer vision tasks [1, 47, 60]. Consequently, reflection removal, which aims at removing undesired reflections and recovering clear transmission layers from contaminated mixture images, has become an attractive topic in the field of computational photography [8, 9, 16, 31, 59, 65, 77]. State-of-the-art single-image reflection removal methods [8, 22, 31, 61] predominantly learn deep priors from a mixed dataset of synthetic and real data to mitigate reflections. However, due to insufficient" + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.825, + 0.537, + 0.84 + ], + "angle": 0, + "content": "# Equal contributions. * Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Y. Hong et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.407, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.145, + 0.788, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.414 + ], + "angle": 0, + "content": "Fig. 1: For mixture images with strong reflections, (a-b) single-image reflection removal methods [22, 31] fail in distinguishing reflection and transmission layers. (c) Using language descriptions as guidance and mixture images as conditions, ControlNet [74] generates results with color shifts and structure distortions, while (d) the proposed L-DiffER successfully achieves high-fidelity transmission recovery in low-transmitted reflection regions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.788, + 0.612 + ], + "angle": 0, + "content": "knowledge about transmission and reflection scenes for addressing such a highly ill-posed problem, they often encounter limitations in scenes with complex or low-transmitted reflections, as shown in yellow boxes of Fig. 1(a-b). By leveraging auxiliary scene information from additional inputs obtained with specialized devices [27, 28, 37] or capture settings [34, 35, 43], multi-image reflection removal methods achieve more robust reflection removal than single-image methods. However, the specialized requirements for data acquisition, in contrast, limit their application scopes, especially for mobile devices and images from the Internet. Hence, it is imperative to find a user-friendly auxiliary input that helps to relieve the ill-posedness of reflection removal while maintaining the applicability and accessibility of single-image methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Fortunately, natural language can exactly serve as the aforementioned auxiliary input, since it provides instructive semantic information about images [36, 63, 72], which bridges the gap between human understanding and machine perception. Thanks to the development of vision-language models [48], various vision tasks (e.g., image editing [55, 56, 67, 68] and image colorization [3-5, 69]) benefit from the semantic guidance of language descriptions. Zhong et al. [81] make the first attempt to introduce language descriptions for reflection removal, but due to the lack of generation capability and generalizability, it encounters difficulties in addressing scenarios containing low-transmitted reflections." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Recently, by combining semantic information from natural language descriptions and generative priors from pre-trained diffusion models [14, 49], language-based diffusion models (e.g., ControlNet [74]) generate visually pleasant results for image restoration tasks like image colorization [4] and super resolution [54]. For reflection removal, it also shows the potential to provide semantic priors to relieve the ill-posedness, where the generative priors in pre-trained diffusion" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "L-DiffER" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.435 + ], + "angle": 0, + "content": "models can adeptly manage complex scenarios and make it possible to recover transmission layers under low-transmitted reflections. However, as shown in the red boxes of Fig. 1(c), directly employing the existing language-based diffusion model like ControlNet [74] for reflection removal will encounter two primary challenges: (1) Inaccurate control conditions. When using language-based diffusion models, other tasks such as image colorization [4] often employ a clear (not superimposed) image to provide accurate structure control. However, in the reflection removal task, the image content of the transmission layer in a mixture image is contaminated by the reflection layer, and thus it is unable to provide an initially accurate condition for both structure and color control. Consequently, achieving transmission recovery under such inaccurate conditions presents a significant challenge. (2) Insufficient recovery faithfulness. Since diffusion models are principally designed for image generation, simply applying them for reflection removal may obtain results that are less faithful to conditions (i.e., results with unwanted image content containing color shifts and structural distortions compared with mixture images). Therefore, it is essential to develop a mechanism for constraining the generation capability and improving the recovery faithfulness of diffusion models to recover transmission layers while preventing unwanted alterations in image color and structure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.683 + ], + "angle": 0, + "content": "In this paper, we propose L-DiffER, the first Language-based Diffusion model for single image rEflection Removal, to achieve high-fidelity transmission recovery with language guidance under the interference of low-transmitted reflections, as the results shown in Fig. 1(d). To leverage the language descriptions of two layers, descriptions of the transmission layer serve as positive prompts, guiding the model to preserve essential image content, whereas descriptions of the reflection layer act as negative prompts, aiding in the suppression of unwanted reflections. To address the aforementioned challenges, we propose an iterative condition refinement strategy within the reverse diffusion process to ensure increasingly accurate representations of color and structure conditions as the process evolves. Furthermore, we introduce a multi-condition constraint mechanism designed to safeguard the faithfulness to the specified conditions, which effectively addresses potential color shifts and structural distortions by applying constraints that align the recovery output with the desired color and structure conditions. The proposed method achieves state-of-the-art performance by making the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.71, + 0.784, + 0.754 + ], + "angle": 0, + "content": "- We present the first language-based diffusion model for reflection removal, leveraging the descriptive power of language to distinguish between transmission and reflection layers." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.76, + 0.785, + 0.789 + ], + "angle": 0, + "content": "- We propose an iterative condition refinement strategy to provide more accurate control conditions as color and structure guidance for reflection removal." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.795, + 0.785, + 0.839 + ], + "angle": 0, + "content": "- We design a multi-condition constraint mechanism to ensure the recovery faithfulness of image color and structure, which effectively handles low-transmitted reflections." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.71, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Y. Hong et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.389, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.475 + ], + "angle": 0, + "content": "Reflection removal. Single image reflection removal aims at suppressing reflections and recovering transmission layers in a single mixture image. Using the presumption that reflections tend to blur and exhibit low intensities, nonlearning approaches employ handcrafted priors into their optimization frameworks [57, 71], such as the gradient sparsity [30], relative smoothness [33], and ghosting cues [51]. With the development of deep learning, researchers attempt to address the task by exploring different network architectures [9, 21, 59, 61, 65], loss functions [76], or the iterative recovery strategy [8, 31, 70, 77]. Concurrently, efforts in data synthesis to meet the data-driven demands is also an attractive topic in this area [9, 22, 24, 39, 66, 79, 80]. Besides, the exploration of special image forms like panoramic images [10, 19, 20, 45] helps to diminish content ambiguity in mixture images. To relieve the ill-posedness of reflection removal, multi-image methods utilize additional input images to introduce auxiliary clues such as polarization information [7, 26, 28, 37, 38, 44, 50], transmission scene information from active illuminations [2, 17, 18, 27, 29], and motion discrepancy between the two layers [32, 34, 35, 52], while special data capture requirements limit their practical applications. Inspired by Zhong et al. [81], we introduce language-based diffusion models to input auxiliary language descriptions for relieving the ill-posedness of reflection removal while maintaining the applicability of the method, and to leverage generative priors for handling low-transmitted reflection regions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.791, + 0.717 + ], + "angle": 0, + "content": "Language-based diffusion models. The advent of diffusion models [14] (DMs) has marked a significant leap forward, particularly in the realms of image generation [49]. The emergence of pre-trained vision-language models, notably CLIP [48], significantly enhances the flexibility and intuitiveness of utilizing language descriptions for vision tasks. Specifically, these pre-trained models have been instrumental in enabling multi-condition image editing [42,74,78], image colorization [4], and devising sophisticated sampling strategies for tailored generation outcomes [12]. Despite the strides in image editing and conditional generation, the specific challenge of single image reflection removal, i.e., recovering clear transmission layers from the guidance of a superimposed mixture images, still remains unexplored for language-based diffusion models [49, 74]. In our work, we seek to address the reflection removal problem by harnessing the generative prowess and flexibility of language-guided diffusion models, which combines the semantic information offered by language descriptions with the sophisticated generative capabilities of DMs while ensuring the recovery fidelity and faithfulness with condition refinement and constraints." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.737, + 0.381, + 0.755 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this paper, we address the problem of language-based reflection removal using diffusion models. We first provide a brief overview of the framework of our language-based reflection removal diffusion model L-DiffER in Sec. 3.1. Sec. 3.2 proposes the iterative condition refinement strategy designed for addressing inaccurate control conditions, and Sec. 3.3 proposes the multi-condition constraint" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "L-DiffER" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.584, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.147, + 0.781, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.37, + 0.785, + 0.412 + ], + "angle": 0, + "content": "Fig. 2: Framework of the proposed language-based reflection removal diffusion model, which leverages language descriptions with the iterative condition refinement strategy and the multi-condition constraint mechanism to achieve faithful transmission recovery." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.443, + 0.784, + 0.473 + ], + "angle": 0, + "content": "mechanism for improving the recovery faithfulness. Finally, loss functions and training details are presented in Sec. 3.4 and Sec. 3.5, respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.497, + 0.341, + 0.51 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Problem formulation. Given the mixture image \\(\\mathbf{M} \\in \\mathbb{R}^{H \\times W \\times 3}\\) and the language description condition \\(c\\), we optimize our model to remove reflection contamination and recover transmission layers \\(\\widetilde{\\mathbf{T}} \\in \\mathbb{R}^{H \\times W \\times 3}\\) under the supervision of the ground truth transmission layers \\(\\mathbf{T} \\in \\mathbb{R}^{H \\times W \\times 3}\\). We apply the commonly-used assumption [21,31,61] as: \\(\\mathbf{M} = \\mathbf{T} + \\mathbf{R}\\), where the mixture image \\(\\mathbf{M}\\) can be considered as the combination of a transmission layer \\(\\mathbf{T}\\) and a reflection layer \\(\\mathbf{R}\\). We use the intensity and gradient of the mixture image \\(\\mathbf{M}\\) to initialize the color condition \\(\\mathbf{I}^{\\mathrm{s}} = \\mathbf{M}\\) and the structure condition \\(\\mathbf{I}^{\\mathrm{s}} = \\nabla \\mathbf{M}\\). Then the color latent \\(y^{\\mathrm{c}} = \\mathcal{E}^{\\mathrm{c}}(\\mathbf{I}^{\\mathrm{c}})\\) and structure latent \\(y^{\\mathrm{s}} = \\mathcal{E}^{\\mathrm{s}}(\\mathbf{I}^{\\mathrm{s}})\\) are extracted by a color and a structure encoder \\(\\mathcal{E}^{\\mathrm{c}}\\) and \\(\\mathcal{E}^{\\mathrm{s}}\\) (denoted by the blue blocks in the middle part of Fig. 2), respectively. Given language descriptions about the content of transmission layers and reflection layers as positive prompt and negative prompt, respectively, the positive and negative latents \\(c^{\\mathrm{p}}\\) and \\(c^{\\mathrm{n}}\\) are extracted from the CLIP text encoder [48]. In the proposed method, the transmission layer \\(\\widetilde{\\mathbf{T}}\\) is recovered from all the spatial and language conditioning information encoded in latents \\(y = [y^{\\mathrm{s}}, y^{\\mathrm{c}}]\\) and \\(c = [c^{\\mathrm{p}}, c^{\\mathrm{n}}]\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Diffusion models. Diffusion models [14, 53] are probabilistic generative models that have been shown to have powerful data modeling ability for images, which can provide an effective image prior for the estimation of transmission layers. To address the challenges of resource consumption in pixel space, Stable Diffusion (SD) [49] introduces latent diffusion models to improve both the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Y. Hong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.343 + ], + "angle": 0, + "content": "training and sampling efficiency of denoising diffusion models without degrading their quality. Following SD [49], we employ a compression encoder \\(\\mathcal{E}\\) to encode a transmission layer into the latent space as \\(z = \\mathcal{E}(\\mathbf{T})\\), and a compression decoder \\(\\mathcal{D}\\) to reconstruct the transmission layer from the latent code \\(z\\) as \\(\\widetilde{\\mathbf{T}} = \\mathcal{D}(z)\\). The diffusion process progressively destructs data by injecting noise, then learns to reverse this process for image generation. During training, the forward diffusion process starts from the target image \\(z_0 = \\mathcal{E}(\\mathbf{T})\\) and generates samples \\(z_t = \\sqrt{\\overline{\\alpha}_t} z_0 + \\sqrt{1 - \\overline{\\alpha}_t} \\epsilon_t\\), where \\(\\epsilon_t \\sim \\mathcal{N}(0,1)\\) is the Gaussian noise at timestep \\(t\\), and \\(\\overline{\\alpha}_t\\) represents the noise scheduler introduced in [49]. During inference, the reverse diffusion process starts from a random noise sample \\(z_T \\sim \\mathcal{N}(0,1)\\), where a noise prediction network denoted as \\(\\epsilon_\\theta\\) is optimized to predict the noise \\(\\hat{\\epsilon}_t\\) at each timestep \\(t\\) given the language condition \\(c\\) and the spatial conditions \\(y_t^1\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.345, + 0.786, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {l d m}} = \\mathbb {E} _ {\\mathcal {E} (x), t, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon_ {t} - \\epsilon_ {\\theta} \\left(z _ {t}, c, y _ {t}, t\\right) \\| _ {2} ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.786, + 0.401 + ], + "angle": 0, + "content": "until it converges to the desired photorealistic transmission layer \\(\\widetilde{\\mathbf{T}}\\). The overall framework is shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.788, + 0.569 + ], + "angle": 0, + "content": "Condition extraction. During inference, the two conditioning latents start from the initial color and structure latents, i.e., \\( y_{T}^{c} = y^{c} \\) and \\( y_{T}^{\\mathrm{s}} = y^{\\mathrm{s}} \\), respectively, and they are refined by employing the iterative condition refinement strategy (Sec. 3.2) to provide more accurate conditions. For condition injection, the proposed framework draws inspiration from ControlNet [74], which learns a parameter-efficient, parallel branch on top of SD. Specifically, two trainable copied modules (denoted by the pink blocks in Fig. 2) are utilized to take the color latent \\( y_{t}^{c} \\) and the gradient latent \\( y_{t}^{\\mathrm{s}} \\) with the latent presentation \\( z_{t} \\) as input to extract color and structure features, respectively. These features are then added directly to the corresponding scales of the locked module (denoted by the purple block in Fig. 2), which guides \\( \\epsilon_{\\theta} \\) in learning the noise distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.645 + ], + "angle": 0, + "content": "Language guidance. To explicitly instruct the model to suppress reflections from the language guidance, classifier-free guidance [15] is used to employ negative prompts to specify undesired reflection contamination. Specifically, two noise predictions with positive prompts \\( c^{\\mathrm{p}} \\) and negative prompts \\( c^{\\mathrm{n}} \\) are made at each timestep \\( t \\). The fusion of these two predictions yields the output \\( \\hat{e}_{t-1} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.28, + 0.659, + 0.785, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} _ {t - 1} = \\epsilon_ {\\theta} \\left(z _ {t}, c ^ {\\mathrm {p}}, y _ {t}, t\\right) + \\lambda_ {\\mathrm {c f g}} \\times \\left(\\epsilon_ {\\theta} \\left(z _ {t}, c ^ {\\mathrm {p}}, y _ {t}, t\\right) - \\epsilon_ {\\theta} \\left(z _ {t}, c ^ {\\mathrm {n}}, y _ {t}, t\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.737 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{cfg}}\\) is a hyperparameter. In our task, as reflections are not always recognizable [81], we set negative prompts as the empty string \\(c^{\\mathrm{n}} = \\varnothing\\) in the case. For simplicity, we use \\(\\epsilon_{\\theta}(z_t,c,y_t,t)\\) to represent the whole noise prediction process." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.759, + 0.583, + 0.775 + ], + "angle": 0, + "content": "3.2 Iterative condition refinement strategy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.785, + 0.788, + 0.816 + ], + "angle": 0, + "content": "Although with effective data modeling ability, diffusion models probably generate results unfaithful to the input image conditions [40,74]. This issue becomes even" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.787, + 0.841 + ], + "angle": 0, + "content": "1 Note that the spatial conditions are varied with timesteps in the proposed method." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "L-DiffER" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.147, + 0.784, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.235, + 0.785, + 0.263 + ], + "angle": 0, + "content": "Fig. 3: Visualization of pseudo transmission layers by using the original mixture image with its gradient as conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.296, + 0.788, + 0.448 + ], + "angle": 0, + "content": "harder for reflection removal, since the desired transmission layers should have a structure without gradients of reflection contaminations, which is different from the input mixture images. That is, the mixture images itself cannot provide an accurate image condition for diffusion-based reflection removal. Therefore, we turn to the recovered transmission layers in the inference process, which have the potential to be combined with the mixture images and help obtain a refined condition. During inference, given a latent noisy sample \\( z_{t} \\), we can use the noise prediction network \\( \\epsilon_{\\theta} \\) to predict noise as \\( \\hat{\\epsilon}_t = \\epsilon_\\theta (z_t,t,c,y_t) \\). Then we can calculate the estimated latent representation of the transmission layer \\( z_{0|t} \\) at the current timestep \\( t \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.46, + 0.786, + 0.478 + ], + "angle": 0, + "content": "\\[\nz _ {0 \\mid t} = \\left(z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} * \\hat {\\epsilon} _ {t}\\right) / \\sqrt {\\bar {\\alpha} _ {t}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.788, + 0.551 + ], + "angle": 0, + "content": "After that, we can obtain the estimated pseudo transmission layer \\(\\mathbf{T}_{0|t}\\) by using the compression decoder \\(\\mathcal{D}\\) to decode \\(z_{0|t}\\): \\(\\mathbf{T}_{0|t} = \\mathcal{D}(z_{0|t})\\). Then \\(\\mathbf{T}_{0|t}\\) and its gradient \\(\\nabla \\mathbf{T}_{0|t}\\) can be used as the pseudo-accurate color and structure condition to obtain conditions at the next timestep." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.734 + ], + "angle": 0, + "content": "As pseudo transmission layers shown in Fig. 3, we observe two key phenomena: Firstly, the estimated pseudo transmission layers in the first few sampling steps exhibit inaccuracies, with neither structure nor color serving effectively as guiding conditions. However, as the sampling steps progress, the fidelity of the reconstructed images notably improves. Secondly, although employing the mixture image and the gradient of the mixture image as conditions persistently yield inaccurate control, these conditions provide a richer set of usable information than the recovered image at the beginning of sampling. Leveraging the advantages of both observations, we introduce an iterative condition refinement strategy to address the challenge of inaccurate control conditions. Our control conditions transitioned from the mixture image and the gradient of the mixture image to the recovered image and its gradient as the sampling step increases." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.735, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Specifically, we introduce predefined time-variant coefficients \\(\\beta_{t}\\) and \\(\\gamma_{t}\\) (\\(t \\in \\{1,..,T\\}\\)) to combine pseudo-accurate conditions \\(\\mathbf{T}_{0|t}\\) with the one initiated from \\(\\mathbf{M}\\), which decrease as timestep \\(t\\) increases² as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.268, + 0.795, + 0.787, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {t} ^ {\\mathrm {c}} = \\gamma_ {t} * \\mathbf {T} _ {0 | t} + (1 - \\gamma_ {t}) * \\mathbf {M}, \\quad \\mathbf {I} _ {t} ^ {\\mathrm {s}} = \\beta_ {t} * \\nabla \\mathbf {T} _ {0 | t} + (1 - \\beta_ {t}) * \\nabla \\mathbf {M}. \\tag {4}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.691, + 0.84 + ], + "angle": 0, + "content": "2 Details of \\(\\beta_{t}\\) and \\(\\gamma_{t}\\) will be explained in the supplementary material." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Y. Hong et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.148, + 0.785, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.357, + 0.785, + 0.385 + ], + "angle": 0, + "content": "Fig. 4: Visualization of the color condition, structure condition, and the pseudo transmission layer by applying the proposed iterative condition refinement strategy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.784, + 0.442 + ], + "angle": 0, + "content": "Fig. 4 shows intermediate conditions and the pseudo transmission layers during sampling, indicating the effect of the iterative condition refinement strategy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.785, + 0.533 + ], + "angle": 0, + "content": "Meanwhile, we also use the iterative condition refinement strategy with a few modifications during training. We replace the pseudo-accurate transmission layers \\(\\mathbf{T}_{0|t}\\) estimated from sampling with ground truth \\(\\mathbf{T}\\) to refine conditions before estimating noises, which is different from the sampling process. Specifically, given the initial mixture image \\(\\mathbf{M}\\) and its gradient \\(\\nabla \\mathbf{M}\\), we refine the input conditions at timestep \\(t\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.282, + 0.543, + 0.785, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathbf {I}} _ {t} ^ {\\mathrm {c}} = \\gamma_ {t} * \\mathbf {T} + (1 - \\gamma_ {t}) * \\mathbf {M}, \\quad \\widehat {\\mathbf {I}} _ {t} ^ {\\mathrm {s}} = \\beta_ {t} * \\nabla \\mathbf {T} + (1 - \\beta_ {t}) * \\nabla \\mathbf {M}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.571, + 0.784, + 0.601 + ], + "angle": 0, + "content": "where \\(\\beta_{t}\\) and \\(\\gamma_{t}\\) (\\(t \\in \\{1, \\dots, T\\}\\)) are the same predefined time-variant coefficients used in the sampling process." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.621, + 0.576, + 0.635 + ], + "angle": 0, + "content": "3.3 Multi-condition constraint mechanism" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In order to mitigate the insufficient recovery faithfulness of diffusion models, we employ a multi-condition mechanism that separately utilizes mixture images and their gradients to obtain initial latent conditions for controlling the color and structure of generated results. As the iterative condition refinement strategy obtains new refined conditions during each timestep, no constraints applying on conditions may cause the sampling process out of control, which also impacts the recovery faithfulness. Hence, it is imperative to conduct constraints on refined conditions. Though inaccurate as color conditions, mixture images provide coarse color information about transmission layers, thus we propose a maximum value constraint to prevent color shifts. Similarly, the gradient of the mixture image provides features showing structure information such as edges and textures. Therefore, we propose a saturation-aware structure constraint to prevent structure distortions when recovering transmission layers. Details are as follows." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "L-DiffER" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "Maximum value constraint (MVC). According to the assumption defined in Sec. 3.1, the pixel values of the mixture image should be greater than or equal to the transmission layer. Therefore, when updating the conditions in the training and sampling processes, we employ maximum value constraint as a function to ensure that any given color condition \\(\\mathbf{I}_t^{\\mathrm{c}}\\) defined in Eq. (4) does not exceed the mixture image \\(\\mathbf{M}\\) in terms of pixel values. This constraint should be adhered to the following:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.263, + 0.785, + 0.279 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M V C} \\left(\\mathbf {I} _ {t} ^ {\\mathrm {c}}\\right) = \\min \\left(\\mathbf {M}, \\mathbf {I} _ {t} ^ {\\mathrm {c}}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.288, + 0.785, + 0.438 + ], + "angle": 0, + "content": "Saturation-aware structure constraint (SSC). Due to regional overexposure during photographing, mixture images sometimes contain low-transmitted reflection regions, which lack enough visible content of transmission layers, causing the corresponding gradients to have little usable structure information. Therefore, simply applying the maximum value constraint to edges will force the refined structure condition to abandon generated content in saturated regions, which may hinder optimization. Therefore, we introduce a saturation-aware mask to preserve the generative capability of the structural aspects in regions challenging for recovery. Specifically, we first utilize the mixture image to identify low-transmitted regions, generating a mask \\(\\mathbf{M}^{\\mathrm{s}}\\in \\mathbb{R}^{H\\times W}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.448, + 0.785, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Omega} ^ {\\mathrm {s}} = \\left\\{ \\begin{array}{l l} 0, & \\text {w h e r e} \\mathbf {M} > \\tau \\\\ 1, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.492, + 0.784, + 0.552 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}\\) is the mixture image, \\(\\tau\\) is the saturation threshold we empirically set to 0.95. Besides, to ensure that the structure condition does not contain nonexistent edges in non-saturated regions of the original mixture image, we define a mask to indicate the valid edge in the gradient of the mixture image as:" + }, + { + "type": "equation", + "bbox": [ + 0.404, + 0.559, + 0.785, + 0.599 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Omega} ^ {\\mathrm {v}} = \\left\\{ \\begin{array}{l l} 0, & \\text {w h e r e} \\nabla \\mathbf {M} < \\eta \\\\ 1, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.605, + 0.784, + 0.648 + ], + "angle": 0, + "content": "where \\(\\eta\\) is the valid edge threshold we empirically set to 0.05. After that, for a given gradient condition \\(\\mathbf{I}_t^s\\) defined in Eq. (4), we define the saturation-aware structure constraint as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.66, + 0.785, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S S C} \\left(\\mathbf {I} _ {t} ^ {\\mathrm {s}}\\right) = \\boldsymbol {\\Omega} ^ {\\mathrm {s}} \\odot \\mathbf {I} _ {t} ^ {\\mathrm {s}} + \\left(1 - \\boldsymbol {\\Omega} ^ {\\mathrm {s}}\\right) \\odot \\boldsymbol {\\Omega} ^ {\\mathrm {v}} \\odot \\mathbf {I} _ {t} ^ {\\mathrm {s}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.685, + 0.784, + 0.731 + ], + "angle": 0, + "content": "To extract the color latent \\( y_{t}^{\\mathrm{c}} \\) and the structure latent \\( y_{t}^{\\mathrm{s}} \\) at each timestep \\( t \\), we apply the maximum value constraint (MVC) and saturation-aware structure constraint (SSC) on the color condition \\( \\mathbf{I}_t^{\\mathrm{c}} \\) and structure condition \\( \\mathbf{I}_t^{\\mathrm{s}} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.74, + 0.785, + 0.756 + ], + "angle": 0, + "content": "\\[\ny _ {t - 1} ^ {\\mathrm {c}} = \\mathcal {E} ^ {\\mathrm {c}} \\left(\\operatorname {M V C} \\left(\\mathbf {I} _ {t} ^ {\\mathrm {c}}\\right)\\right), \\quad y _ {t - 1} ^ {\\mathrm {s}} = \\mathcal {E} ^ {\\mathrm {s}} \\left(\\operatorname {S S C} \\left(\\mathbf {I} _ {t} ^ {\\mathrm {s}}\\right)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.784, + 0.84 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}^{\\mathrm{s}}\\) is the structure condition encoder, \\(\\mathcal{E}^{\\mathrm{c}}\\) is the color condition encoder, and MVC and SSC are applied to the color and structure conditions for ensuring recovery faithfulness. Finally, we apply the denoising process with the DDIM [53] sampling scheme to obtain the next noisy sample \\(z_{t-1}\\). Algorithm 1 outlines the complete sampling process." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "Y. Hong et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.147, + 0.774, + 0.163 + ], + "angle": 0, + "content": "Algorithm 1 Iterative condition refinement strategy in the sampling process" + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.165, + 0.787, + 0.416 + ], + "angle": 0, + "content": "Require: Noise prediction network \\(\\epsilon_{\\theta}\\) Compression decoder \\(\\mathcal{D}\\) Color encoder \\(\\mathcal{E}^{\\mathrm{c}}\\) Structure encoder \\(\\mathcal{E}^{\\mathrm{s}}\\) \nInput: Mixture image M, Language condition \\(c = [c^{\\mathrm{p}},c^{\\mathrm{n}}]\\) , Initial spatial conditions \\([y^{\\mathrm{s}} = \\mathcal{E}^{\\mathrm{s}}(\\nabla \\mathbf{M}),y^{\\mathrm{c}} = \\mathcal{E}^{\\mathrm{c}}(\\mathbf{M})]\\) , Coefficients \\(\\{\\overline{\\alpha}_t\\}_{t = 1}^T\\) \\(\\{\\beta_t\\}_{t = 1}^T\\) \\(\\{\\gamma_t\\}_{t = 1}^T\\) \nOutput: Recovered latent representation \\(z_{0}\\) \n1: \\(z_{T}\\sim \\mathcal{N}(0,1)\\) \n2: \\(y_{T} = [y^{s},y^{c}]\\) \n3: for each \\(t\\in [T,1]\\) do \n4: \\(\\hat{\\epsilon}_t = \\epsilon_\\theta (z_t,c,y_t,t)\\) \n5: if \\(t > 1\\) then \n6: \\(z_{0|t} = (z_t - \\sqrt{1 - \\overline{\\alpha}_t} *\\hat{\\epsilon}_t) / \\sqrt{\\overline{\\alpha}_t}\\) D Compute \\(z_{0|t}\\) by Eq. (3) \n7: \\(\\mathbf{T}_{0|t} = \\mathcal{D}(z_{0|t})\\) \n8: \\(y_{t - 1}^{\\mathrm{c}} = \\mathcal{E}^{\\mathrm{c}}(\\mathrm{MVC}(\\gamma_t*\\mathbf{T}_{0|t} + (1 - \\gamma_t)*\\mathbf{M}))\\) DRefine color condition \n9: \\(y_{t - 1}^{\\mathrm{s}} = \\mathcal{E}^{\\mathrm{s}}(\\mathrm{SSC}(\\beta_t*\\nabla \\mathbf{T}_{0|t} + (1 - \\beta_t)*\\nabla \\mathbf{M}))\\) DRefine structure condition \n10: \\(y_{t - 1} = [y_{t - 1}^{\\mathrm{s}},y_{t - 1}^{\\mathrm{c}}]\\) DUpdate spatial conditions \n11: end if \n12: \\(z_{t - 1} = \\sqrt{\\overline{\\alpha}_{t - 1}} z_{0|t} + \\sqrt{1 - \\overline{\\alpha}_{t - 1}}\\hat{\\epsilon}_t\\) DDIM \n13: end for" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.446, + 0.38, + 0.459 + ], + "angle": 0, + "content": "3.4 Loss functions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.785, + 0.514 + ], + "angle": 0, + "content": "In this section, we introduce the latent diffusion model loss to minimize the discrepancy between the learned latent space distribution and the target distribution, as well as the pixel losses to prevent color shifts and structure distortions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.785, + 0.591 + ], + "angle": 0, + "content": "We follow the training objective of ControlNet [74] but replace task-specific conditions with the mixture image and the gradient of the mixture. Thus our network \\(\\epsilon_{\\theta}\\) learns to predict the noise added to the noisy latent \\(z_{t}\\) with \\(\\mathcal{L}_{\\mathrm{ldm}}\\) defined in Eq. (1). According to Eq. (3), we can obtain \\(\\mathbf{T}_{0|t}\\) in pixel space by decoding \\(z_{0|t}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.337, + 0.603, + 0.785, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {0 | t} = \\mathcal {D} \\left(z _ {0 | t}\\right) = \\mathcal {D} \\left(\\left(z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} * \\hat {\\epsilon} _ {t}\\right) / \\sqrt {\\bar {\\alpha} _ {t}}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.632, + 0.784, + 0.663 + ], + "angle": 0, + "content": "Then RGB loss is introduced to constrain the consistency of the reconstructed transmission \\(\\mathbf{T}_{0|t}\\) and ground truth \\(\\mathbf{T}\\), minimizing their errors as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.422, + 0.673, + 0.785, + 0.691 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {R G B}} = \\left\\| \\mathbf {T} - \\mathbf {T} _ {0 | t} \\right\\| _ {2} ^ {2}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.701, + 0.784, + 0.731 + ], + "angle": 0, + "content": "For better structure consistency, we use gradient loss to constrain the prediction's gradient and its ground truth as below:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.741, + 0.785, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {g r a d}} = \\left\\| \\nabla (\\mathbf {T}) - \\nabla \\left(\\mathbf {T} _ {0 | t}\\right) \\right\\| _ {2} ^ {2}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.769, + 0.785, + 0.814 + ], + "angle": 0, + "content": "where \\(\\nabla (\\cdot)\\) denotes the computation of image gradients. Furthermore, we also compute the standard deviation and the mean of the predicted transmission and its ground truth, and use a \\(\\ell_2\\) loss to minimize the color distortion:" + }, + { + "type": "equation", + "bbox": [ + 0.326, + 0.824, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {n u m}} = \\left\\| \\mu (\\mathbf {T}) - \\mu \\left(\\mathbf {T} _ {0 | t}\\right) \\right\\| _ {2} ^ {2} + \\left\\| \\sigma (\\mathbf {T}) - \\sigma \\left(\\mathbf {T} _ {0 | t}\\right) \\right\\| _ {2} ^ {2}, \\tag {14}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "L-DiffER" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "where \\(\\mu (\\cdot)\\) and \\(\\sigma (\\cdot)\\) denote the computation of mean and standard deviation, respectively, and \\(i = 1,\\dots,N\\) denotes the \\(i\\)-th sample in the dataset. Overall, the pixel loss is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.206, + 0.785, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {p i x}} = \\mathcal {L} _ {\\mathrm {R G B}} + \\mathcal {L} _ {\\mathrm {g r a d}} + \\mathcal {L} _ {\\mathrm {n u m}}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.232, + 0.784, + 0.261 + ], + "angle": 0, + "content": "Gathering the latent diffusion model loss and pixel losses yields the final objective as:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.263, + 0.785, + 0.278 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l l}} := \\mathcal {L} _ {\\mathrm {l d m}} + \\lambda \\mathcal {L} _ {\\text {p i x}}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.285, + 0.452, + 0.3 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is set to 1.0 empirically." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.322, + 0.453, + 0.337 + ], + "angle": 0, + "content": "3.5 Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.347, + 0.785, + 0.422 + ], + "angle": 0, + "content": "We implement the proposed method using PyTorch [46]. The model is trained for 20 epochs with a batch size of 16 on two NVIDIA GeForce RTX 3090 GPUs. Weights are initialized as in [11] and updated using the Adam optimizer [25]. The learning rate is set to \\(1 \\times 10^{-5}\\). For sampling, we use the DDIM sampling method [53] with 50 sampling steps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.423, + 0.787, + 0.513 + ], + "angle": 0, + "content": "Dataset. Our training dataset contains both synthetic and real data. For the synthetic data, following Zhong et al. [81], we generate 50000 triplets of data using COCO Captions [6] and Flickr30k [73] for network training. For real data, we use 200 image pairs from Nature dataset [31] and 90 pairs from Zhang et al. [76] for training and the rest of the real data are used for evaluation. We further collect 50 mixture images from the Internet for qualitative evaluation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.536, + 0.376, + 0.554 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.568, + 0.542, + 0.584 + ], + "angle": 0, + "content": "4.1 Comparison with state-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.593, + 0.785, + 0.743 + ], + "angle": 0, + "content": "For evaluating model performances, experiments are conducted on existing real reflection datasets [31, 58, 76] and our collected mixture images from the Internet (with manually annotated language descriptions). We compare the proposed L-DiffER with state-of-the-art single-image reflection removal methods, including Zhang et al. [76], CoRRN [61], ERRNet [65], IBCLN [31], Dong et al. [8], YTMT [21], and DSRNet [22]. To compare with a representative language-based diffusion model ControlNet \\([74]^3\\), we make it fit our setting by using mixture images as conditions, and transmission and reflection descriptions as positive and negative prompts, respectively. For fair comparisons, we finetune the above methods on our training data if their training codes are provided." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.744, + 0.785, + 0.79 + ], + "angle": 0, + "content": "Quantitative comparison. Quantitative experiments are conducted on three real datasets for reflection removal, i.e., Nature [31], Real20 [76], and SIR² [58] dataset. To evaluate the recovery of transmission layers⁴, we utilize PSNR [23]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.797, + 0.784, + 0.826 + ], + "angle": 0, + "content": "3 Since ControlNet [74] destructs image color and structure in a generative manner as shown in Fig. 1(c), we only run it for qualitative comparisons." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.737, + 0.84 + ], + "angle": 0, + "content": "4 Evaluations on reflection layers are provided in the supplementary material." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.797, + 0.784, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "Y. Hong et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.2 + ], + "angle": 0, + "content": "Table 1: Comparison of quantitative results on real datasets for evaluating the recovery of transmission layers, compared with several state-of-the-art single-image reflection removal methods [8,21,22,31,61,76]. \\(\\uparrow (\\downarrow)\\) indicates larger (smaller) values are better. Bold numbers indicate the best-performing results." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.213, + 0.787, + 0.515 + ], + "angle": 0, + "content": "
Dataset (size)MetricMethod
Zhang et al.CoRRNERRNetIBCLNDong et al.YTMTDSRNetOurs
SIR2(500)PSNR↑22.4522.9623.1323.3623.0923.0524.9725.18
SSIM↑0.8720.8790.8780.8810.8930.8860.9070.911
LPIPS↓0.1720.1610.1640.1530.1310.1490.1240.121
NIQE↓4.9114.8264.7934.5014.7294.7384.6244.352
FID↓78.6169.5873.3360.9249.3958.6946.6044.66
Real20(20)PSNR↑22.5121.1722.0621.5921.7322.3123.4623.77
SSIM↑0.8060.7860.8030.7710.8110.8050.8060.821
LPIPS↓0.2040.2020.1850.2130.1700.1780.1650.153
NIQE↓3.8193.8913.9893.8953.9593.9774.0653.665
FID↓113.25111.2585.07120.9586.2094.6777.9272.64
Nature(20)PSNR↑20.3720.5421.1123.6923.6121.0321.7023.95
SSIM↑0.7720.7780.8060.8280.8250.8020.8200.831
LPIPS↓0.2170.2050.1830.1700.1570.1860.1680.145
NIQE↓4.4134.4334.4424.3454.5814.3874.3734.223
FID↓109.7789.1279.6875.1771.5581.7774.8468.57
Average(540)PSNR↑22.3822.8023.0223.3123.0622.9524.7925.08
SSIM↑0.8660.8720.8730.8750.8870.8800.9000.905
LPIPS↓0.1750.1640.1650.1560.1330.1510.1270.123
NIQE↓4.8524.7774.7504.4734.6954.6974.5944.322
FID↓81.0571.8574.0063.6751.5760.8848.8146.58
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.56, + 0.785, + 0.636 + ], + "angle": 0, + "content": "and SSIM [64] as error metrics following reflection removal methods [8, 37]. We further adopts LPIPS [75], NIQE [41], FID [13] to measure the perceptual quality of the recovered results. As quantitative results shown in Table 1, the proposed method achieves the best performance among all competing methods, especially in metrics for perceptual quality, validating its effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.644, + 0.787, + 0.839 + ], + "angle": 0, + "content": "Qualitative comparison. Qualitative comparisons with the aforementioned single-image reflection removal methods [8,21,22,31,61,76] and a language-based diffusion model ControlNet [74] on recovering transmission layers are shown in Fig. 5. As can be observed, single-image reflection removal methods encounter obstacles in dealing with complex reflections due to the lack of auxiliary semantic information, especially in low-transmitted reflection regions. ControlNet [74] performs significant modifications on the image color and structure in a generative manner, which deviates significantly from the input mixture image, emphasizing the pivotal role of accurate conditions in the sampling process. The proposed method generates visually pleasant reflection removal results with faithful recovery, demonstrating that it capitalizes on the synergy of the auxiliary information from language descriptions, the iterative condition refinement strategy, and the multi-condition constraint mechanism." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "L-DiffER" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.146, + 0.501, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.146, + 0.785, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.264, + 0.501, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.264, + 0.785, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.442, + 0.501, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.442, + 0.785, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.571 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparison of estimated transmission layers on real mixture images collected from the Internet, compared with several single-image methods [8,21,22,31, 61,76] and a diffusion-based method ControlNet [74]. Please zoom in for details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.602, + 0.386, + 0.618 + ], + "angle": 0, + "content": "4.2 Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.632, + 0.788, + 0.813 + ], + "angle": 0, + "content": "Ablation on control conditions5. We conduct ablation studies to investigate the effectiveness of the iterative condition refinement strategy and the multi-condition constraint mechanism. As shown in Table 2 and Fig. 6, most reflection contaminations are retained when missing iterative condition refinement (W/o rfn.), and disabling the refinement on color (W/o clr. rfn.) or structure (W/o str. rfn.) will cause inaccurate recovery, which exists color shift or retains the reflection content. In addition, as shown in Table 2 and Fig. 7, due to the generative prior of the diffusion model, the absence of the max value constraint (W/o MVC) or the saturation-aware structure constraint (W/o SSC), or both (W/o cnst.) leads to obvious color shifts, structure distortion, and uncontrolled recovery. Therefore, the degradation observed in the above ablations indicates the necessity of the refinement and constraints on conditions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.677, + 0.841 + ], + "angle": 0, + "content": "5 More ablation studies are provided in the supplementary material." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Y. Hong et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Ablation studies on the iterative condition refinement strategy and the multi-condition constraint mechanism. \\(\\uparrow (\\downarrow)\\) indicates larger (smaller) values are better. Bold numbers indicate the best-performing results." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.199, + 0.787, + 0.303 + ], + "angle": 0, + "content": "
MetricAbl. on iterative condition refinementAbl. on multi-condition constraint
W/o clr. rfn.W/o str. rfn.W/o rfn.W/o MVCW/o SSCW/o cnst.Ours
PSNR↑24.4224.6824.3924.7324.1623.3525.08
SSIM↑0.8980.8960.8910.8970.8870.8790.905
LPIPS↓0.1320.1360.1390.1290.1430.1540.123
NIQE↓4.5014.6454.8684.5534.7094.5614.322
FID↓49.1353.7160.1251.0656.4663.6846.58
" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.318, + 0.785, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.405, + 0.788, + 0.448 + ], + "angle": 0, + "content": "Fig. 6: The effect of the iterative condition refinement strategy. For each result, we place the corresponding final color and structure condition at its lower left and lower right. Please zoom in for details, respectively." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.464, + 0.785, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.55, + 0.788, + 0.592 + ], + "angle": 0, + "content": "Fig. 7: The effect of the color and structure constraints in the multi-condition constraint mechanism. For each result, we place the corresponding final color and structure condition at its lower left and lower right, respectively. Please zoom in for details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.622, + 0.36, + 0.637 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.779 + ], + "angle": 0, + "content": "This paper introduces the first language-based diffusion model for single image reflection removal, which exploits auxiliary semantic information from language descriptions and generative priors from diffusion models to recover transmission layers from mixture images with low-transmitted reflection regions. An iterative condition refinement strategy is proposed to address the problem of inaccurate control conditions, and a multi-condition constraint mechanism is introduced to ensure recovery faithfulness. Experiments on real data demonstrate the effectiveness of the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Limitations. The proposed method may incorrectly identify transmission and reflection layers if they exhibit similar image content. For such cases, more accurate spatial guidance from users such as annotated edges or region masks could help to relieve the ambiguity, which is left as our future work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "L-DiffER" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.177, + 0.784, + 0.207 + ], + "angle": 0, + "content": "This work is supported by National Natural Science Foundation of China under Grant No. 62136001, 62088102." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.23, + 0.323, + 0.246 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.288 + ], + "angle": 0, + "content": "1. Chang, Y., Jung, C., Sun, J.: Joint reflection removal and depth estimation from a single image. IEEE Transactions on Cybernetics (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.328 + ], + "angle": 0, + "content": "2. Chang, Y., Jung, C., Sun, J., Wang, F.: Siamese dense network for reflection removal with flash and no-flash image pairs. International Journal of Computer Vision (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.785, + 0.371 + ], + "angle": 0, + "content": "3. Chang, Z., Weng, S., Li, Y., Li, S., Shi, B.: L-CoDer: Language-based colorization with color-object decoupling transformer. In: Proc. of European Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.372, + 0.785, + 0.412 + ], + "angle": 0, + "content": "4. Chang, Z., Weng, S., Zhang, P., Li, Y., Li, S., Shi, B.: L-CAD: Language-based colorization with any-level descriptions using diffusion priors. In: Proc. of Advances in Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.413, + 0.785, + 0.453 + ], + "angle": 0, + "content": "5. Chang, Z., Weng, S., Zhang, P., Li, Y., Li, S., Shi, B.: L-CoIns: Language-based colorization with instance awareness. In: Proc. of Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.455, + 0.785, + 0.495 + ], + "angle": 0, + "content": "6. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dolkar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.785, + 0.523 + ], + "angle": 0, + "content": "7. Diamant, Y., Schechner, Y.Y.: Overcoming visual reverberations. In: Proc. of Computer Vision and Pattern Recognition (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.524, + 0.785, + 0.564 + ], + "angle": 0, + "content": "8. Dong, Z., Xu, K., Yang, Y., Bao, H., Xu, W., Lau, R.W.: Location-aware single image reflection removal. In: Proc. of International Conference on Computer Vision (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "9. Fan, Q., Yang, J., Hua, G., Chen, B., Wipf, D.: A generic deep architecture for single image reflection removal and image smoothing. In: Proc. of International Conference on Computer Vision (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.607, + 0.785, + 0.634 + ], + "angle": 0, + "content": "10. Han, B.J., Sim, J.Y.: Zero-shot learning for reflection removal of single 360-degree image. In: Proc. of European Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.635, + 0.785, + 0.674 + ], + "angle": 0, + "content": "11. He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In: Proc. of International Conference on Computer Vision (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.676, + 0.785, + 0.716 + ], + "angle": 0, + "content": "12. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.717, + 0.785, + 0.757 + ], + "angle": 0, + "content": "13. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Proc. of Advances in Neural Information Processing Systems (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Proc. of Advances in Neural Information Processing Systems (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.786, + 0.785, + 0.812 + ], + "angle": 0, + "content": "15. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "16. Hong, Y., Chang, Y., Liang, J., Ma, L., Huang, T., Shi, B.: Light flickering guided reflection removal. International Journal of Computer Vision (2024)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.261, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "Y. Hong et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "17. Hong, Y., Lyu, Y., Li, S., Cao, G., Shi, B.: Reflection removal with NIR and RGB image feature fusion. IEEE Transactions on Multimedia (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.785, + 0.203 + ], + "angle": 0, + "content": "18. Hong, Y., Lyu, Y., Li, S., Shi, B.: Near-infrared image guided reflection removal. In: Proc. of International Conference on Multimedia and Expo (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.785, + 0.23 + ], + "angle": 0, + "content": "19. Hong, Y., Zheng, Q., Zhao, L., Jiang, X., Kot, A.C., Shi, B.: Panoramic image reflection removal. In: Proc. of Computer Vision and Pattern Recognition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.27 + ], + "angle": 0, + "content": "20. Hong, Y., Zheng, Q., Zhao, L., Jiang, X., Kot, A.C., Shi, B.: PAR\\(^2\\)Net: End-to-end panoramic image reflection removal. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.27, + 0.785, + 0.311 + ], + "angle": 0, + "content": "21. Hu, Q., Guo, X.:trash or treasure? an interactive dual-stream strategy for single image reflection separation. Proc. of Advances in Neural Information Processing Systems (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.338 + ], + "angle": 0, + "content": "22. Hu, Q., Guo, X.: Single image reflection separation via component synergy. In: Proc. of International Conference on Computer Vision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.366 + ], + "angle": 0, + "content": "23. Huynh-Thu, Q., Ghanbari, M.: Scope of validity of psnr in image/video quality assessment. Electronics letters (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.785, + 0.393 + ], + "angle": 0, + "content": "24. Kim, S., Huo, Y., Yoon, S.E.: Single image reflection removal with physically-based training images. In: Proc. of Computer Vision and Pattern Recognition (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.42 + ], + "angle": 0, + "content": "25. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.446 + ], + "angle": 0, + "content": "26. Kong, N., Tai, Y.W., Shin, S.Y.: A physically-based approach to reflection separation. In: Proc. of Computer Vision and Pattern Recognition (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.785, + 0.473 + ], + "angle": 0, + "content": "27. Lei, C., Chen, Q.: Robust reflection removal with reflection-free flash-only cues. In: Proc. of Computer Vision and Pattern Recognition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "28. Lei, C., Huang, X., Zhang, M., Yan, Q., Sun, W., Chen, Q.: Polarized reflection removal with perfect alignment in the wild. In: Proc. of Computer Vision and Pattern Recognition (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.542 + ], + "angle": 0, + "content": "29. Lei, C., Jiang, X., Chen, Q.: Robust reflection removal with flash-only cues in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.582 + ], + "angle": 0, + "content": "30. Levin, A., Weiss, Y.: User assisted separation of reflections from a single image using a sparsity prior. IEEE Transactions on Pattern Analysis and Machine Intelligence 29(9), 1647-1654 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.785, + 0.622 + ], + "angle": 0, + "content": "31. Li, C., Yang, Y., He, K., Lin, S., Hopcroft, J.E.: Single image reflection removal through cascaded refinement. In: Proc. of Computer Vision and Pattern Recognition (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.622, + 0.785, + 0.65 + ], + "angle": 0, + "content": "32. Li, Y., Brown, M.S.: Exploiting reflection change for automatic reflection removal. In: Proc. of International Conference on Computer Vision (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.676 + ], + "angle": 0, + "content": "33. Li, Y., Brown, M.S.: Single image layer separation using relative smoothness. In: Proc. of Computer Vision and Pattern Recognition (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.676, + 0.785, + 0.703 + ], + "angle": 0, + "content": "34. Liu, Y.L., Lai, W.S., Yang, M.H., Chuang, Y.Y., Huang, J.B.: Learning to see through obstructions. In: Proc. of Computer Vision and Pattern Recognition (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.785, + 0.744 + ], + "angle": 0, + "content": "35. Liu, Y.L., Lai, W.S., Yang, M.H., Chuang, Y.Y., Huang, J.B.: Learning to see through obstructions with layered decomposition. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.8 + ], + "angle": 0, + "content": "36. Luo, J., Fu, J., Kong, X., Gao, C., Ren, H., Shen, H., Xia, H., Liu, S.: 3D-SPS: Single-stage 3D visual grounding via referred point progressive selection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16454-16463 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "37. Lyu, Y., Cui, Z., Li, S., Pollefeys, M., Shi, B.: Reflection separation using a pair of unpolarized and polarized images. In: Proc. of Advances in Neural Information Processing Systems (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "L-DiffER" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.19 + ], + "angle": 0, + "content": "38. Lyu, Y., Cui, Z., Li, S., Pollefeys, M., Shi, B.: Physics-guided reflection separation from a pair of unpolarized and polarized images. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.233 + ], + "angle": 0, + "content": "39. Ma, D., Wan, R., Shi, B., Kot, A.C., Duan, L.Y.: Learning to jointly generate and separate reflections. In: Proc. of International Conference on Computer Vision (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.787, + 0.275 + ], + "angle": 0, + "content": "40. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: Proc. of International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.787, + 0.303 + ], + "angle": 0, + "content": "41. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters 20(3), 209-212 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.787, + 0.346 + ], + "angle": 0, + "content": "42. Mou, C., Wang, X., Xie, L., Wu, Y., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.787, + 0.388 + ], + "angle": 0, + "content": "43. Nam, S., Brubaker, M.A., Brown, M.S.: Neural image representations for multi-image fusion and layer separation. In: Proc. of European Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.787, + 0.416 + ], + "angle": 0, + "content": "44. Nayar, S.K., Fang, X.S., Boult, T.: Separation of reflection components using color and polarization. International Journal of Computer Vision (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.418, + 0.787, + 0.459 + ], + "angle": 0, + "content": "45. Park, J., Kim, H., Park, E., Sim, J.Y.: Fully-automatic reflection removal for 360-degree images. In: Proc. of Winter Conference on Applications of Computer Vision (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.787, + 0.514 + ], + "angle": 0, + "content": "46. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: Proc. of Advances in Neural Information Processing Systems (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.787, + 0.557 + ], + "angle": 0, + "content": "47. Qiu, J., Jiang, P.T., Zhu, Y., Yin, Z.X., Cheng, M.M., Ren, B.: Looking through the glass: Neural surface reconstruction against high specular reflections. In: Proc. of Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.558, + 0.787, + 0.613 + ], + "angle": 0, + "content": "48. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proc. of International Conference on Machine Learning. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.614, + 0.787, + 0.655 + ], + "angle": 0, + "content": "49. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proc. of Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.656, + 0.787, + 0.684 + ], + "angle": 0, + "content": "50. Schechner, Y.Y., Kiryati, N., Basri, R.: Separation of transparent layers using focus. International Journal of Computer Vision (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.685, + 0.787, + 0.712 + ], + "angle": 0, + "content": "51. Shih, Y., Krishnan, D., Durand, F., Freeman, W.T.: Reflection removal using ghosting cues. In: Proc. of Computer Vision and Pattern Recognition (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.713, + 0.787, + 0.741 + ], + "angle": 0, + "content": "52. Simon, C., Kyu Park, I.: Reflection removal for in-vehicle black box videos. In: Proc. of Computer Vision and Pattern Recognition (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.787, + 0.769 + ], + "angle": 0, + "content": "53. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.787, + 0.811 + ], + "angle": 0, + "content": "54. Sun, H., Li, W., Liu, J., Chen, H., Pei, R., Zou, X., Yan, Y., Yang, Y.: CoSeR: Bridging image and language for cognitive super-resolution. arXiv preprint arXiv:2311.16512 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.787, + 0.84 + ], + "angle": 0, + "content": "55. Sun, J., Weng, S., Chang, Z., Li, S., Shi, B.: UniCoRN: A unified conditional image repainting network. In: Proc. of Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "Y. Hong et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "56. Tang, J., Zhong, H., Weng, S., Shi, B.: LuminAIRe: Illumination-aware conditional image painting for lighting-realistic generation. In: Proc. of Advances in Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "57. Wan, R., Shi, B., Duan, L.Y., Tan, A.H., Gao, W., Kot, A.C.: Region-aware reflection removal with unified content and gradient priors. IEEE Transactions on Image Processing (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.271 + ], + "angle": 0, + "content": "58. Wan, R., Shi, B., Duan, L.Y., Tan, A.H., Kot, A.C.: Benchmarking single-image reflection removal algorithms. In: Proc. of International Conference on Computer Vision (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.312 + ], + "angle": 0, + "content": "59. Wan, R., Shi, B., Duan, L.Y., Tan, A.H., Kot, A.C.: CRRN: Multi-scale guided concurrent reflection removal network. In: Proc. of Computer Vision and Pattern Recognition (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.312, + 0.785, + 0.338 + ], + "angle": 0, + "content": "60. Wan, R., Shi, B., Li, H., Duan, L.Y., Kot, A.C.: Face image reflection removal. International Journal of Computer Vision (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.379 + ], + "angle": 0, + "content": "61. Wan, R., Shi, B., Li, H., Duan, L.Y., Tan, A.H., Kot, A.C.: CoRRN: Cooperative reflection removal network. IEEE Transactions on Pattern Analysis and Machine Intelligence (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.42 + ], + "angle": 0, + "content": "62. Wan, R., Shi, B., Li, H., Hong, Y., Duan, L.Y., Kot, A.C.: Benchmarking single-image reflection removal algorithms. In: IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.461 + ], + "angle": 0, + "content": "63. Wang, Z., Lu, Y., Li, Q., Tao, X., Guo, Y., Gong, M., Liu, T.: CRIS: CLIP-driven referring image segmentation. In: Proc. of Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.502 + ], + "angle": 0, + "content": "64. Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.542 + ], + "angle": 0, + "content": "65. Wei, K., Yang, J., Fu, Y., Wipf, D., Huang, H.: Single image reflection removal exploiting misaligned training data and network enhancements. In: Proc. of Computer Vision and Pattern Recognition (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.569 + ], + "angle": 0, + "content": "66. Wen, Q., Tan, Y., Qin, J., Liu, W., Han, G., He, S.: Single image reflection removal beyond linearity. In: Proc. of Computer Vision and Pattern Recognition (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.61 + ], + "angle": 0, + "content": "67. Weng, S., Li, W., Li, D., Jin, H., Shi, B.: MISC: Multi-condition injection and spatially-adaptive compositing for conditional person image synthesis. In: Proc. of Computer Vision and Pattern Recognition (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.637 + ], + "angle": 0, + "content": "68. Weng, S., Shi, B.: Conditional image inpainting. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.678 + ], + "angle": 0, + "content": "69. Weng, S., Wu, H., Chang, Z., Tang, J., Li, S., Shi, B.: L-CoDe: Language-based colorization using color-object decoupled conditions. In: Proc. of the AAAI Conference on Artificial Intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "70. Yang, J., Gong, D., Liu, L., Shi, Q.: Seeing deeply and bidirectionally: A deep learning approach for single image reflection removal. In: Proc. of European Conference on Computer Vision (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "71. Yang, Y., Ma, W., Zheng, Y., Cai, J.F., Xu, W.: Fast single image reflection suppression via convex optimization. In: Proc. of Computer Vision and Pattern Recognition (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "72. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: LAVT: Language-aware vision transformer for referring image segmentation. In: Proc. of Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.841 + ], + "angle": 0, + "content": "73. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics (2014)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "L-DiffER" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "74. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proc. of International Conference on Computer Vision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.217 + ], + "angle": 0, + "content": "75. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proc. of Computer Vision and Pattern Recognition (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.217, + 0.788, + 0.245 + ], + "angle": 0, + "content": "76. Zhang, X., Ng, R., Chen, Q.: Single image reflection separation with perceptual losses. In: Proc. of Computer Vision and Pattern Recognition (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.788, + 0.286 + ], + "angle": 0, + "content": "77. Zhang, Y.N., Shen, L., Li, Q.: Content and gradient model-driven deep network for single image reflection removal. In: Proc. of ACM International Conference on Multimedia (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.328 + ], + "angle": 0, + "content": "78. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Unictrlnet: All-in-one control to text-to-image diffusion models. Proc. of Advances in Neural Information Processing Systems (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "79. Zheng, Q., Chen, J., Lu, Z., Shi, B., Jiang, X., Yap, K.H., Duan, L.Y., Kot, A.C.: What does plate glass reveal about camera calibration? In: Proc. of Computer Vision and Pattern Recognition (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.411 + ], + "angle": 0, + "content": "80. Zheng, Q., Shi, B., Chen, J., Jiang, X., Duan, L.Y., Kot, A.C.: Single image reflection removal with absorption effect. In: Proc. of Computer Vision and Pattern Recognition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.411, + 0.788, + 0.44 + ], + "angle": 0, + "content": "81. Zhong, H., Hong, Y., Weng, S., Liang, J., Shi, B.: Language-guided image reflection separation. In: Proc. of Computer Vision and Pattern Recognition (2024)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.44 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/L-DiffER_ Single Image Reflection Removal with Language-based Diffusion Model/84990c91-4cd8-4a1b-83d4-8e97d608b20a_origin.pdf b/2024/L-DiffER_ Single Image Reflection Removal with Language-based Diffusion Model/84990c91-4cd8-4a1b-83d4-8e97d608b20a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4781e8aa4158700bcb43e4d353491493146a228b --- /dev/null +++ b/2024/L-DiffER_ Single Image Reflection Removal with Language-based Diffusion Model/84990c91-4cd8-4a1b-83d4-8e97d608b20a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:92177428f232f330874aaa909c7e83a717aed6f83ff42a4936c468598a0579b3 +size 3207834 diff --git a/2024/L-DiffER_ Single Image Reflection Removal with Language-based Diffusion Model/full.md b/2024/L-DiffER_ Single Image Reflection Removal with Language-based Diffusion Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f32661bd2bf77da214d0bb14365de02210698416 --- /dev/null +++ b/2024/L-DiffER_ Single Image Reflection Removal with Language-based Diffusion Model/full.md @@ -0,0 +1,356 @@ +# L-DiffER: Single Image Reflection Removal with Language-based Diffusion Model + +Yuchen Hong $^{1,2\#}$ Haofeng Zhong $^{1,2,3\#}$ Shuchen Weng $^{1,2}$ Jinxiu Liang $^{1,2}$ Boxin Shi $^{1,2,3*}$ + +$^{1}$ State Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University + +$^{2}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University + +$^{3}$ AI Innovation Center, School of Computer Science, Peking University {hfzhong, shuchenweng, cssherryliang, shiboxin}@pku.edu.cn, yuchenhong.cn@gmail.com + +Abstract. In this paper, we introduce L-DiffER, a language-based diffusion model designed for the ill-posed single image reflection removal task. Although having shown impressive performance for image generation, existing language-based diffusion models struggle with precise control and faithfulness in image restoration. To overcome these limitations, we propose an iterative condition refinement strategy to resolve the problem of inaccurate control conditions. A multi-condition constraint mechanism is employed to ensure the recovery faithfulness of image color and structure while retaining the generation capability to handle low-transmitted reflections. We demonstrate the superiority of the proposed method through extensive experiments, showcasing both quantitative and qualitative improvements over existing methods. + +Keywords: Reflection removal $\cdot$ Language-based diffusion model + +# 1 Introduction + +When photographs are taken through transparent mediums like glass windows, the contaminated mixture image (denoted as $\mathbf{M}$ ) can be considered as the combination of a transmission layer (denoted as $\mathbf{T}$ ) and a reflection layer (denoted as $\mathbf{R}$ ) [61, 62], which impairs the performance of downstream computer vision tasks [1, 47, 60]. Consequently, reflection removal, which aims at removing undesired reflections and recovering clear transmission layers from contaminated mixture images, has become an attractive topic in the field of computational photography [8, 9, 16, 31, 59, 65, 77]. State-of-the-art single-image reflection removal methods [8, 22, 31, 61] predominantly learn deep priors from a mixed dataset of synthetic and real data to mitigate reflections. However, due to insufficient + +![](images/e7737be1bfb2e3a48f3a32aa083679e9febad2f0975eead01b83fcc556c4e085.jpg) +Fig. 1: For mixture images with strong reflections, (a-b) single-image reflection removal methods [22, 31] fail in distinguishing reflection and transmission layers. (c) Using language descriptions as guidance and mixture images as conditions, ControlNet [74] generates results with color shifts and structure distortions, while (d) the proposed L-DiffER successfully achieves high-fidelity transmission recovery in low-transmitted reflection regions. + +![](images/3d3eaafcade95108a6cb7e3dfe936b46f442b6e1d2492b7827f3923fbcdef73c.jpg) + +knowledge about transmission and reflection scenes for addressing such a highly ill-posed problem, they often encounter limitations in scenes with complex or low-transmitted reflections, as shown in yellow boxes of Fig. 1(a-b). By leveraging auxiliary scene information from additional inputs obtained with specialized devices [27, 28, 37] or capture settings [34, 35, 43], multi-image reflection removal methods achieve more robust reflection removal than single-image methods. However, the specialized requirements for data acquisition, in contrast, limit their application scopes, especially for mobile devices and images from the Internet. Hence, it is imperative to find a user-friendly auxiliary input that helps to relieve the ill-posedness of reflection removal while maintaining the applicability and accessibility of single-image methods. + +Fortunately, natural language can exactly serve as the aforementioned auxiliary input, since it provides instructive semantic information about images [36, 63, 72], which bridges the gap between human understanding and machine perception. Thanks to the development of vision-language models [48], various vision tasks (e.g., image editing [55, 56, 67, 68] and image colorization [3-5, 69]) benefit from the semantic guidance of language descriptions. Zhong et al. [81] make the first attempt to introduce language descriptions for reflection removal, but due to the lack of generation capability and generalizability, it encounters difficulties in addressing scenarios containing low-transmitted reflections. + +Recently, by combining semantic information from natural language descriptions and generative priors from pre-trained diffusion models [14, 49], language-based diffusion models (e.g., ControlNet [74]) generate visually pleasant results for image restoration tasks like image colorization [4] and super resolution [54]. For reflection removal, it also shows the potential to provide semantic priors to relieve the ill-posedness, where the generative priors in pre-trained diffusion + +models can adeptly manage complex scenarios and make it possible to recover transmission layers under low-transmitted reflections. However, as shown in the red boxes of Fig. 1(c), directly employing the existing language-based diffusion model like ControlNet [74] for reflection removal will encounter two primary challenges: (1) Inaccurate control conditions. When using language-based diffusion models, other tasks such as image colorization [4] often employ a clear (not superimposed) image to provide accurate structure control. However, in the reflection removal task, the image content of the transmission layer in a mixture image is contaminated by the reflection layer, and thus it is unable to provide an initially accurate condition for both structure and color control. Consequently, achieving transmission recovery under such inaccurate conditions presents a significant challenge. (2) Insufficient recovery faithfulness. Since diffusion models are principally designed for image generation, simply applying them for reflection removal may obtain results that are less faithful to conditions (i.e., results with unwanted image content containing color shifts and structural distortions compared with mixture images). Therefore, it is essential to develop a mechanism for constraining the generation capability and improving the recovery faithfulness of diffusion models to recover transmission layers while preventing unwanted alterations in image color and structure. + +In this paper, we propose L-DiffER, the first Language-based Diffusion model for single image rEflection Removal, to achieve high-fidelity transmission recovery with language guidance under the interference of low-transmitted reflections, as the results shown in Fig. 1(d). To leverage the language descriptions of two layers, descriptions of the transmission layer serve as positive prompts, guiding the model to preserve essential image content, whereas descriptions of the reflection layer act as negative prompts, aiding in the suppression of unwanted reflections. To address the aforementioned challenges, we propose an iterative condition refinement strategy within the reverse diffusion process to ensure increasingly accurate representations of color and structure conditions as the process evolves. Furthermore, we introduce a multi-condition constraint mechanism designed to safeguard the faithfulness to the specified conditions, which effectively addresses potential color shifts and structural distortions by applying constraints that align the recovery output with the desired color and structure conditions. The proposed method achieves state-of-the-art performance by making the following contributions: + +- We present the first language-based diffusion model for reflection removal, leveraging the descriptive power of language to distinguish between transmission and reflection layers. +- We propose an iterative condition refinement strategy to provide more accurate control conditions as color and structure guidance for reflection removal. +- We design a multi-condition constraint mechanism to ensure the recovery faithfulness of image color and structure, which effectively handles low-transmitted reflections. + +# 2 Related Work + +Reflection removal. Single image reflection removal aims at suppressing reflections and recovering transmission layers in a single mixture image. Using the presumption that reflections tend to blur and exhibit low intensities, nonlearning approaches employ handcrafted priors into their optimization frameworks [57, 71], such as the gradient sparsity [30], relative smoothness [33], and ghosting cues [51]. With the development of deep learning, researchers attempt to address the task by exploring different network architectures [9, 21, 59, 61, 65], loss functions [76], or the iterative recovery strategy [8, 31, 70, 77]. Concurrently, efforts in data synthesis to meet the data-driven demands is also an attractive topic in this area [9, 22, 24, 39, 66, 79, 80]. Besides, the exploration of special image forms like panoramic images [10, 19, 20, 45] helps to diminish content ambiguity in mixture images. To relieve the ill-posedness of reflection removal, multi-image methods utilize additional input images to introduce auxiliary clues such as polarization information [7, 26, 28, 37, 38, 44, 50], transmission scene information from active illuminations [2, 17, 18, 27, 29], and motion discrepancy between the two layers [32, 34, 35, 52], while special data capture requirements limit their practical applications. Inspired by Zhong et al. [81], we introduce language-based diffusion models to input auxiliary language descriptions for relieving the ill-posedness of reflection removal while maintaining the applicability of the method, and to leverage generative priors for handling low-transmitted reflection regions. + +Language-based diffusion models. The advent of diffusion models [14] (DMs) has marked a significant leap forward, particularly in the realms of image generation [49]. The emergence of pre-trained vision-language models, notably CLIP [48], significantly enhances the flexibility and intuitiveness of utilizing language descriptions for vision tasks. Specifically, these pre-trained models have been instrumental in enabling multi-condition image editing [42,74,78], image colorization [4], and devising sophisticated sampling strategies for tailored generation outcomes [12]. Despite the strides in image editing and conditional generation, the specific challenge of single image reflection removal, i.e., recovering clear transmission layers from the guidance of a superimposed mixture images, still remains unexplored for language-based diffusion models [49, 74]. In our work, we seek to address the reflection removal problem by harnessing the generative prowess and flexibility of language-guided diffusion models, which combines the semantic information offered by language descriptions with the sophisticated generative capabilities of DMs while ensuring the recovery fidelity and faithfulness with condition refinement and constraints. + +# 3 Methodology + +In this paper, we address the problem of language-based reflection removal using diffusion models. We first provide a brief overview of the framework of our language-based reflection removal diffusion model L-DiffER in Sec. 3.1. Sec. 3.2 proposes the iterative condition refinement strategy designed for addressing inaccurate control conditions, and Sec. 3.3 proposes the multi-condition constraint + +![](images/d6231c5081ffb347fde3bb3531b922f76ac69d15cc683cf5f5b51c2928ac4fb2.jpg) +Fig. 2: Framework of the proposed language-based reflection removal diffusion model, which leverages language descriptions with the iterative condition refinement strategy and the multi-condition constraint mechanism to achieve faithful transmission recovery. + +![](images/f1a9c04a4c585ba8296882b45ef9f4b3b4354b31dd6fb4ffd7489a0ddcac0830.jpg) + +mechanism for improving the recovery faithfulness. Finally, loss functions and training details are presented in Sec. 3.4 and Sec. 3.5, respectively. + +# 3.1 Overview + +Problem formulation. Given the mixture image $\mathbf{M} \in \mathbb{R}^{H \times W \times 3}$ and the language description condition $c$ , we optimize our model to remove reflection contamination and recover transmission layers $\widetilde{\mathbf{T}} \in \mathbb{R}^{H \times W \times 3}$ under the supervision of the ground truth transmission layers $\mathbf{T} \in \mathbb{R}^{H \times W \times 3}$ . We apply the commonly-used assumption [21,31,61] as: $\mathbf{M} = \mathbf{T} + \mathbf{R}$ , where the mixture image $\mathbf{M}$ can be considered as the combination of a transmission layer $\mathbf{T}$ and a reflection layer $\mathbf{R}$ . We use the intensity and gradient of the mixture image $\mathbf{M}$ to initialize the color condition $\mathbf{I}^{\mathrm{s}} = \mathbf{M}$ and the structure condition $\mathbf{I}^{\mathrm{s}} = \nabla \mathbf{M}$ . Then the color latent $y^{\mathrm{c}} = \mathcal{E}^{\mathrm{c}}(\mathbf{I}^{\mathrm{c}})$ and structure latent $y^{\mathrm{s}} = \mathcal{E}^{\mathrm{s}}(\mathbf{I}^{\mathrm{s}})$ are extracted by a color and a structure encoder $\mathcal{E}^{\mathrm{c}}$ and $\mathcal{E}^{\mathrm{s}}$ (denoted by the blue blocks in the middle part of Fig. 2), respectively. Given language descriptions about the content of transmission layers and reflection layers as positive prompt and negative prompt, respectively, the positive and negative latents $c^{\mathrm{p}}$ and $c^{\mathrm{n}}$ are extracted from the CLIP text encoder [48]. In the proposed method, the transmission layer $\widetilde{\mathbf{T}}$ is recovered from all the spatial and language conditioning information encoded in latents $y = [y^{\mathrm{s}}, y^{\mathrm{c}}]$ and $c = [c^{\mathrm{p}}, c^{\mathrm{n}}]$ . + +Diffusion models. Diffusion models [14, 53] are probabilistic generative models that have been shown to have powerful data modeling ability for images, which can provide an effective image prior for the estimation of transmission layers. To address the challenges of resource consumption in pixel space, Stable Diffusion (SD) [49] introduces latent diffusion models to improve both the + +training and sampling efficiency of denoising diffusion models without degrading their quality. Following SD [49], we employ a compression encoder $\mathcal{E}$ to encode a transmission layer into the latent space as $z = \mathcal{E}(\mathbf{T})$ , and a compression decoder $\mathcal{D}$ to reconstruct the transmission layer from the latent code $z$ as $\widetilde{\mathbf{T}} = \mathcal{D}(z)$ . The diffusion process progressively destructs data by injecting noise, then learns to reverse this process for image generation. During training, the forward diffusion process starts from the target image $z_0 = \mathcal{E}(\mathbf{T})$ and generates samples $z_t = \sqrt{\overline{\alpha}_t} z_0 + \sqrt{1 - \overline{\alpha}_t} \epsilon_t$ , where $\epsilon_t \sim \mathcal{N}(0,1)$ is the Gaussian noise at timestep $t$ , and $\overline{\alpha}_t$ represents the noise scheduler introduced in [49]. During inference, the reverse diffusion process starts from a random noise sample $z_T \sim \mathcal{N}(0,1)$ , where a noise prediction network denoted as $\epsilon_\theta$ is optimized to predict the noise $\hat{\epsilon}_t$ at each timestep $t$ given the language condition $c$ and the spatial conditions $y_t^1$ as follows: + +$$ +\mathcal {L} _ {\mathrm {l d m}} = \mathbb {E} _ {\mathcal {E} (x), t, \epsilon \sim \mathcal {N} (0, 1)} \left[ \| \epsilon_ {t} - \epsilon_ {\theta} \left(z _ {t}, c, y _ {t}, t\right) \| _ {2} ^ {2} \right], \tag {1} +$$ + +until it converges to the desired photorealistic transmission layer $\widetilde{\mathbf{T}}$ . The overall framework is shown in Fig. 2. + +Condition extraction. During inference, the two conditioning latents start from the initial color and structure latents, i.e., $y_{T}^{c} = y^{c}$ and $y_{T}^{\mathrm{s}} = y^{\mathrm{s}}$ , respectively, and they are refined by employing the iterative condition refinement strategy (Sec. 3.2) to provide more accurate conditions. For condition injection, the proposed framework draws inspiration from ControlNet [74], which learns a parameter-efficient, parallel branch on top of SD. Specifically, two trainable copied modules (denoted by the pink blocks in Fig. 2) are utilized to take the color latent $y_{t}^{c}$ and the gradient latent $y_{t}^{\mathrm{s}}$ with the latent presentation $z_{t}$ as input to extract color and structure features, respectively. These features are then added directly to the corresponding scales of the locked module (denoted by the purple block in Fig. 2), which guides $\epsilon_{\theta}$ in learning the noise distribution. + +Language guidance. To explicitly instruct the model to suppress reflections from the language guidance, classifier-free guidance [15] is used to employ negative prompts to specify undesired reflection contamination. Specifically, two noise predictions with positive prompts $c^{\mathrm{p}}$ and negative prompts $c^{\mathrm{n}}$ are made at each timestep $t$ . The fusion of these two predictions yields the output $\hat{e}_{t-1}$ : + +$$ +\hat {\epsilon} _ {t - 1} = \epsilon_ {\theta} \left(z _ {t}, c ^ {\mathrm {p}}, y _ {t}, t\right) + \lambda_ {\mathrm {c f g}} \times \left(\epsilon_ {\theta} \left(z _ {t}, c ^ {\mathrm {p}}, y _ {t}, t\right) - \epsilon_ {\theta} \left(z _ {t}, c ^ {\mathrm {n}}, y _ {t}, t\right)\right), \tag {2} +$$ + +where $\lambda_{\mathrm{cfg}}$ is a hyperparameter. In our task, as reflections are not always recognizable [81], we set negative prompts as the empty string $c^{\mathrm{n}} = \varnothing$ in the case. For simplicity, we use $\epsilon_{\theta}(z_t,c,y_t,t)$ to represent the whole noise prediction process. + +# 3.2 Iterative condition refinement strategy + +Although with effective data modeling ability, diffusion models probably generate results unfaithful to the input image conditions [40,74]. This issue becomes even + +![](images/db5626b361337e2b81256db283df02a4906ce5444fd8c7d716afc2c28c57e045.jpg) +Fig. 3: Visualization of pseudo transmission layers by using the original mixture image with its gradient as conditions. + +harder for reflection removal, since the desired transmission layers should have a structure without gradients of reflection contaminations, which is different from the input mixture images. That is, the mixture images itself cannot provide an accurate image condition for diffusion-based reflection removal. Therefore, we turn to the recovered transmission layers in the inference process, which have the potential to be combined with the mixture images and help obtain a refined condition. During inference, given a latent noisy sample $z_{t}$ , we can use the noise prediction network $\epsilon_{\theta}$ to predict noise as $\hat{\epsilon}_t = \epsilon_\theta (z_t,t,c,y_t)$ . Then we can calculate the estimated latent representation of the transmission layer $z_{0|t}$ at the current timestep $t$ as follows: + +$$ +z _ {0 \mid t} = \left(z _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} * \hat {\epsilon} _ {t}\right) / \sqrt {\bar {\alpha} _ {t}}. \tag {3} +$$ + +After that, we can obtain the estimated pseudo transmission layer $\mathbf{T}_{0|t}$ by using the compression decoder $\mathcal{D}$ to decode $z_{0|t}$ : $\mathbf{T}_{0|t} = \mathcal{D}(z_{0|t})$ . Then $\mathbf{T}_{0|t}$ and its gradient $\nabla \mathbf{T}_{0|t}$ can be used as the pseudo-accurate color and structure condition to obtain conditions at the next timestep. + +As pseudo transmission layers shown in Fig. 3, we observe two key phenomena: Firstly, the estimated pseudo transmission layers in the first few sampling steps exhibit inaccuracies, with neither structure nor color serving effectively as guiding conditions. However, as the sampling steps progress, the fidelity of the reconstructed images notably improves. Secondly, although employing the mixture image and the gradient of the mixture image as conditions persistently yield inaccurate control, these conditions provide a richer set of usable information than the recovered image at the beginning of sampling. Leveraging the advantages of both observations, we introduce an iterative condition refinement strategy to address the challenge of inaccurate control conditions. Our control conditions transitioned from the mixture image and the gradient of the mixture image to the recovered image and its gradient as the sampling step increases. + +Specifically, we introduce predefined time-variant coefficients $\beta_{t}$ and $\gamma_{t}$ ( $t \in \{1,..,T\}$ ) to combine pseudo-accurate conditions $\mathbf{T}_{0|t}$ with the one initiated from $\mathbf{M}$ , which decrease as timestep $t$ increases² as follows: + +$$ +\mathbf {I} _ {t} ^ {\mathrm {c}} = \gamma_ {t} * \mathbf {T} _ {0 | t} + (1 - \gamma_ {t}) * \mathbf {M}, \quad \mathbf {I} _ {t} ^ {\mathrm {s}} = \beta_ {t} * \nabla \mathbf {T} _ {0 | t} + (1 - \beta_ {t}) * \nabla \mathbf {M}. \tag {4} +$$ + +![](images/f44879e18feae278d2a21e482f7d1266cac3258b01cbc0116e5b4627b9053417.jpg) +Fig. 4: Visualization of the color condition, structure condition, and the pseudo transmission layer by applying the proposed iterative condition refinement strategy. + +Fig. 4 shows intermediate conditions and the pseudo transmission layers during sampling, indicating the effect of the iterative condition refinement strategy. + +Meanwhile, we also use the iterative condition refinement strategy with a few modifications during training. We replace the pseudo-accurate transmission layers $\mathbf{T}_{0|t}$ estimated from sampling with ground truth $\mathbf{T}$ to refine conditions before estimating noises, which is different from the sampling process. Specifically, given the initial mixture image $\mathbf{M}$ and its gradient $\nabla \mathbf{M}$ , we refine the input conditions at timestep $t$ as follows: + +$$ +\widehat {\mathbf {I}} _ {t} ^ {\mathrm {c}} = \gamma_ {t} * \mathbf {T} + (1 - \gamma_ {t}) * \mathbf {M}, \quad \widehat {\mathbf {I}} _ {t} ^ {\mathrm {s}} = \beta_ {t} * \nabla \mathbf {T} + (1 - \beta_ {t}) * \nabla \mathbf {M}, \tag {5} +$$ + +where $\beta_{t}$ and $\gamma_{t}$ ( $t \in \{1, \dots, T\}$ ) are the same predefined time-variant coefficients used in the sampling process. + +# 3.3 Multi-condition constraint mechanism + +In order to mitigate the insufficient recovery faithfulness of diffusion models, we employ a multi-condition mechanism that separately utilizes mixture images and their gradients to obtain initial latent conditions for controlling the color and structure of generated results. As the iterative condition refinement strategy obtains new refined conditions during each timestep, no constraints applying on conditions may cause the sampling process out of control, which also impacts the recovery faithfulness. Hence, it is imperative to conduct constraints on refined conditions. Though inaccurate as color conditions, mixture images provide coarse color information about transmission layers, thus we propose a maximum value constraint to prevent color shifts. Similarly, the gradient of the mixture image provides features showing structure information such as edges and textures. Therefore, we propose a saturation-aware structure constraint to prevent structure distortions when recovering transmission layers. Details are as follows. + +Maximum value constraint (MVC). According to the assumption defined in Sec. 3.1, the pixel values of the mixture image should be greater than or equal to the transmission layer. Therefore, when updating the conditions in the training and sampling processes, we employ maximum value constraint as a function to ensure that any given color condition $\mathbf{I}_t^{\mathrm{c}}$ defined in Eq. (4) does not exceed the mixture image $\mathbf{M}$ in terms of pixel values. This constraint should be adhered to the following: + +$$ +\operatorname {M V C} \left(\mathbf {I} _ {t} ^ {\mathrm {c}}\right) = \min \left(\mathbf {M}, \mathbf {I} _ {t} ^ {\mathrm {c}}\right). \tag {6} +$$ + +Saturation-aware structure constraint (SSC). Due to regional overexposure during photographing, mixture images sometimes contain low-transmitted reflection regions, which lack enough visible content of transmission layers, causing the corresponding gradients to have little usable structure information. Therefore, simply applying the maximum value constraint to edges will force the refined structure condition to abandon generated content in saturated regions, which may hinder optimization. Therefore, we introduce a saturation-aware mask to preserve the generative capability of the structural aspects in regions challenging for recovery. Specifically, we first utilize the mixture image to identify low-transmitted regions, generating a mask $\mathbf{M}^{\mathrm{s}}\in \mathbb{R}^{H\times W}$ as: + +$$ +\boldsymbol {\Omega} ^ {\mathrm {s}} = \left\{ \begin{array}{l l} 0, & \text {w h e r e} \mathbf {M} > \tau \\ 1, & \text {o t h e r w i s e ,} \end{array} \right. \tag {7} +$$ + +where $\mathbf{M}$ is the mixture image, $\tau$ is the saturation threshold we empirically set to 0.95. Besides, to ensure that the structure condition does not contain nonexistent edges in non-saturated regions of the original mixture image, we define a mask to indicate the valid edge in the gradient of the mixture image as: + +$$ +\boldsymbol {\Omega} ^ {\mathrm {v}} = \left\{ \begin{array}{l l} 0, & \text {w h e r e} \nabla \mathbf {M} < \eta \\ 1, & \text {o t h e r w i s e ,} \end{array} \right. \tag {8} +$$ + +where $\eta$ is the valid edge threshold we empirically set to 0.05. After that, for a given gradient condition $\mathbf{I}_t^s$ defined in Eq. (4), we define the saturation-aware structure constraint as follows: + +$$ +\operatorname {S S C} \left(\mathbf {I} _ {t} ^ {\mathrm {s}}\right) = \boldsymbol {\Omega} ^ {\mathrm {s}} \odot \mathbf {I} _ {t} ^ {\mathrm {s}} + \left(1 - \boldsymbol {\Omega} ^ {\mathrm {s}}\right) \odot \boldsymbol {\Omega} ^ {\mathrm {v}} \odot \mathbf {I} _ {t} ^ {\mathrm {s}}. \tag {9} +$$ + +To extract the color latent $y_{t}^{\mathrm{c}}$ and the structure latent $y_{t}^{\mathrm{s}}$ at each timestep $t$ , we apply the maximum value constraint (MVC) and saturation-aware structure constraint (SSC) on the color condition $\mathbf{I}_t^{\mathrm{c}}$ and structure condition $\mathbf{I}_t^{\mathrm{s}}$ as: + +$$ +y _ {t - 1} ^ {\mathrm {c}} = \mathcal {E} ^ {\mathrm {c}} \left(\operatorname {M V C} \left(\mathbf {I} _ {t} ^ {\mathrm {c}}\right)\right), \quad y _ {t - 1} ^ {\mathrm {s}} = \mathcal {E} ^ {\mathrm {s}} \left(\operatorname {S S C} \left(\mathbf {I} _ {t} ^ {\mathrm {s}}\right)\right), \tag {10} +$$ + +where $\mathcal{E}^{\mathrm{s}}$ is the structure condition encoder, $\mathcal{E}^{\mathrm{c}}$ is the color condition encoder, and MVC and SSC are applied to the color and structure conditions for ensuring recovery faithfulness. Finally, we apply the denoising process with the DDIM [53] sampling scheme to obtain the next noisy sample $z_{t-1}$ . Algorithm 1 outlines the complete sampling process. + +Algorithm 1 Iterative condition refinement strategy in the sampling process +Require: Noise prediction network $\epsilon_{\theta}$ Compression decoder $\mathcal{D}$ Color encoder $\mathcal{E}^{\mathrm{c}}$ Structure encoder $\mathcal{E}^{\mathrm{s}}$ +Input: Mixture image M, Language condition $c = [c^{\mathrm{p}},c^{\mathrm{n}}]$ , Initial spatial conditions $[y^{\mathrm{s}} = \mathcal{E}^{\mathrm{s}}(\nabla \mathbf{M}),y^{\mathrm{c}} = \mathcal{E}^{\mathrm{c}}(\mathbf{M})]$ , Coefficients $\{\overline{\alpha}_t\}_{t = 1}^T$ $\{\beta_t\}_{t = 1}^T$ $\{\gamma_t\}_{t = 1}^T$ +Output: Recovered latent representation $z_{0}$ +1: $z_{T}\sim \mathcal{N}(0,1)$ +2: $y_{T} = [y^{s},y^{c}]$ +3: for each $t\in [T,1]$ do +4: $\hat{\epsilon}_t = \epsilon_\theta (z_t,c,y_t,t)$ +5: if $t > 1$ then +6: $z_{0|t} = (z_t - \sqrt{1 - \overline{\alpha}_t} *\hat{\epsilon}_t) / \sqrt{\overline{\alpha}_t}$ D Compute $z_{0|t}$ by Eq. (3) +7: $\mathbf{T}_{0|t} = \mathcal{D}(z_{0|t})$ +8: $y_{t - 1}^{\mathrm{c}} = \mathcal{E}^{\mathrm{c}}(\mathrm{MVC}(\gamma_t*\mathbf{T}_{0|t} + (1 - \gamma_t)*\mathbf{M}))$ DRefine color condition +9: $y_{t - 1}^{\mathrm{s}} = \mathcal{E}^{\mathrm{s}}(\mathrm{SSC}(\beta_t*\nabla \mathbf{T}_{0|t} + (1 - \beta_t)*\nabla \mathbf{M}))$ DRefine structure condition +10: $y_{t - 1} = [y_{t - 1}^{\mathrm{s}},y_{t - 1}^{\mathrm{c}}]$ DUpdate spatial conditions +11: end if +12: $z_{t - 1} = \sqrt{\overline{\alpha}_{t - 1}} z_{0|t} + \sqrt{1 - \overline{\alpha}_{t - 1}}\hat{\epsilon}_t$ DDIM +13: end for + +# 3.4 Loss functions + +In this section, we introduce the latent diffusion model loss to minimize the discrepancy between the learned latent space distribution and the target distribution, as well as the pixel losses to prevent color shifts and structure distortions. + +We follow the training objective of ControlNet [74] but replace task-specific conditions with the mixture image and the gradient of the mixture. Thus our network $\epsilon_{\theta}$ learns to predict the noise added to the noisy latent $z_{t}$ with $\mathcal{L}_{\mathrm{ldm}}$ defined in Eq. (1). According to Eq. (3), we can obtain $\mathbf{T}_{0|t}$ in pixel space by decoding $z_{0|t}$ : + +$$ +\mathbf {T} _ {0 | t} = \mathcal {D} \left(z _ {0 | t}\right) = \mathcal {D} \left(\left(z _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} * \hat {\epsilon} _ {t}\right) / \sqrt {\bar {\alpha} _ {t}}\right). \tag {11} +$$ + +Then RGB loss is introduced to constrain the consistency of the reconstructed transmission $\mathbf{T}_{0|t}$ and ground truth $\mathbf{T}$ , minimizing their errors as follows: + +$$ +\mathcal {L} _ {\mathrm {R G B}} = \left\| \mathbf {T} - \mathbf {T} _ {0 | t} \right\| _ {2} ^ {2}. \tag {12} +$$ + +For better structure consistency, we use gradient loss to constrain the prediction's gradient and its ground truth as below: + +$$ +\mathcal {L} _ {\mathrm {g r a d}} = \left\| \nabla (\mathbf {T}) - \nabla \left(\mathbf {T} _ {0 | t}\right) \right\| _ {2} ^ {2}, \tag {13} +$$ + +where $\nabla (\cdot)$ denotes the computation of image gradients. Furthermore, we also compute the standard deviation and the mean of the predicted transmission and its ground truth, and use a $\ell_2$ loss to minimize the color distortion: + +$$ +\mathcal {L} _ {\mathrm {n u m}} = \left\| \mu (\mathbf {T}) - \mu \left(\mathbf {T} _ {0 | t}\right) \right\| _ {2} ^ {2} + \left\| \sigma (\mathbf {T}) - \sigma \left(\mathbf {T} _ {0 | t}\right) \right\| _ {2} ^ {2}, \tag {14} +$$ + +where $\mu (\cdot)$ and $\sigma (\cdot)$ denote the computation of mean and standard deviation, respectively, and $i = 1,\dots,N$ denotes the $i$ -th sample in the dataset. Overall, the pixel loss is formulated as follows: + +$$ +\mathcal {L} _ {\mathrm {p i x}} = \mathcal {L} _ {\mathrm {R G B}} + \mathcal {L} _ {\mathrm {g r a d}} + \mathcal {L} _ {\mathrm {n u m}}. \tag {15} +$$ + +Gathering the latent diffusion model loss and pixel losses yields the final objective as: + +$$ +\mathcal {L} _ {\text {a l l}} := \mathcal {L} _ {\mathrm {l d m}} + \lambda \mathcal {L} _ {\text {p i x}}, \tag {16} +$$ + +where $\lambda$ is set to 1.0 empirically. + +# 3.5 Implementation details + +We implement the proposed method using PyTorch [46]. The model is trained for 20 epochs with a batch size of 16 on two NVIDIA GeForce RTX 3090 GPUs. Weights are initialized as in [11] and updated using the Adam optimizer [25]. The learning rate is set to $1 \times 10^{-5}$ . For sampling, we use the DDIM sampling method [53] with 50 sampling steps. + +Dataset. Our training dataset contains both synthetic and real data. For the synthetic data, following Zhong et al. [81], we generate 50000 triplets of data using COCO Captions [6] and Flickr30k [73] for network training. For real data, we use 200 image pairs from Nature dataset [31] and 90 pairs from Zhang et al. [76] for training and the rest of the real data are used for evaluation. We further collect 50 mixture images from the Internet for qualitative evaluation. + +# 4 Experiments + +# 4.1 Comparison with state-of-the-arts + +For evaluating model performances, experiments are conducted on existing real reflection datasets [31, 58, 76] and our collected mixture images from the Internet (with manually annotated language descriptions). We compare the proposed L-DiffER with state-of-the-art single-image reflection removal methods, including Zhang et al. [76], CoRRN [61], ERRNet [65], IBCLN [31], Dong et al. [8], YTMT [21], and DSRNet [22]. To compare with a representative language-based diffusion model ControlNet $[74]^3$ , we make it fit our setting by using mixture images as conditions, and transmission and reflection descriptions as positive and negative prompts, respectively. For fair comparisons, we finetune the above methods on our training data if their training codes are provided. + +Quantitative comparison. Quantitative experiments are conducted on three real datasets for reflection removal, i.e., Nature [31], Real20 [76], and SIR² [58] dataset. To evaluate the recovery of transmission layers⁴, we utilize PSNR [23] + +Table 1: Comparison of quantitative results on real datasets for evaluating the recovery of transmission layers, compared with several state-of-the-art single-image reflection removal methods [8,21,22,31,61,76]. $\uparrow (\downarrow)$ indicates larger (smaller) values are better. Bold numbers indicate the best-performing results. + +
Dataset (size)MetricMethod
Zhang et al.CoRRNERRNetIBCLNDong et al.YTMTDSRNetOurs
SIR2(500)PSNR↑22.4522.9623.1323.3623.0923.0524.9725.18
SSIM↑0.8720.8790.8780.8810.8930.8860.9070.911
LPIPS↓0.1720.1610.1640.1530.1310.1490.1240.121
NIQE↓4.9114.8264.7934.5014.7294.7384.6244.352
FID↓78.6169.5873.3360.9249.3958.6946.6044.66
Real20(20)PSNR↑22.5121.1722.0621.5921.7322.3123.4623.77
SSIM↑0.8060.7860.8030.7710.8110.8050.8060.821
LPIPS↓0.2040.2020.1850.2130.1700.1780.1650.153
NIQE↓3.8193.8913.9893.8953.9593.9774.0653.665
FID↓113.25111.2585.07120.9586.2094.6777.9272.64
Nature(20)PSNR↑20.3720.5421.1123.6923.6121.0321.7023.95
SSIM↑0.7720.7780.8060.8280.8250.8020.8200.831
LPIPS↓0.2170.2050.1830.1700.1570.1860.1680.145
NIQE↓4.4134.4334.4424.3454.5814.3874.3734.223
FID↓109.7789.1279.6875.1771.5581.7774.8468.57
Average(540)PSNR↑22.3822.8023.0223.3123.0622.9524.7925.08
SSIM↑0.8660.8720.8730.8750.8870.8800.9000.905
LPIPS↓0.1750.1640.1650.1560.1330.1510.1270.123
NIQE↓4.8524.7774.7504.4734.6954.6974.5944.322
FID↓81.0571.8574.0063.6751.5760.8848.8146.58
+ +and SSIM [64] as error metrics following reflection removal methods [8, 37]. We further adopts LPIPS [75], NIQE [41], FID [13] to measure the perceptual quality of the recovered results. As quantitative results shown in Table 1, the proposed method achieves the best performance among all competing methods, especially in metrics for perceptual quality, validating its effectiveness. + +Qualitative comparison. Qualitative comparisons with the aforementioned single-image reflection removal methods [8,21,22,31,61,76] and a language-based diffusion model ControlNet [74] on recovering transmission layers are shown in Fig. 5. As can be observed, single-image reflection removal methods encounter obstacles in dealing with complex reflections due to the lack of auxiliary semantic information, especially in low-transmitted reflection regions. ControlNet [74] performs significant modifications on the image color and structure in a generative manner, which deviates significantly from the input mixture image, emphasizing the pivotal role of accurate conditions in the sampling process. The proposed method generates visually pleasant reflection removal results with faithful recovery, demonstrating that it capitalizes on the synergy of the auxiliary information from language descriptions, the iterative condition refinement strategy, and the multi-condition constraint mechanism. + +![](images/73184fab4595eccaa6f995c8d382bb86c54b531f24f3d23e992c763458005f97.jpg) + +![](images/2e9ec1b2303725df80c73491b414317a0cb5fb966076efc690d9d5141a97cefc.jpg) + +![](images/2f5b523a057505ec3b3e0e96ea137bafe12276199accfb90eb3f961c3238ddfa.jpg) + +![](images/0b29f5ba22d4dd74e195cec4ab76e7ecbe7210d936be5db5c47da2da13a2cd90.jpg) + +![](images/63c56fa02617f6d90b54a50fd5b7bd660d4cb724c5855eaf87b8f3dd03a53098.jpg) +Fig. 5: Qualitative comparison of estimated transmission layers on real mixture images collected from the Internet, compared with several single-image methods [8,21,22,31, 61,76] and a diffusion-based method ControlNet [74]. Please zoom in for details. + +![](images/e4f183a36e8483299ec0751ecc35af2c6ddfc95f1e0b14567a8477f2b1a3809d.jpg) + +# 4.2 Ablation study + +Ablation on control conditions5. We conduct ablation studies to investigate the effectiveness of the iterative condition refinement strategy and the multi-condition constraint mechanism. As shown in Table 2 and Fig. 6, most reflection contaminations are retained when missing iterative condition refinement (W/o rfn.), and disabling the refinement on color (W/o clr. rfn.) or structure (W/o str. rfn.) will cause inaccurate recovery, which exists color shift or retains the reflection content. In addition, as shown in Table 2 and Fig. 7, due to the generative prior of the diffusion model, the absence of the max value constraint (W/o MVC) or the saturation-aware structure constraint (W/o SSC), or both (W/o cnst.) leads to obvious color shifts, structure distortion, and uncontrolled recovery. Therefore, the degradation observed in the above ablations indicates the necessity of the refinement and constraints on conditions. + +Table 2: Ablation studies on the iterative condition refinement strategy and the multi-condition constraint mechanism. $\uparrow (\downarrow)$ indicates larger (smaller) values are better. Bold numbers indicate the best-performing results. + +
MetricAbl. on iterative condition refinementAbl. on multi-condition constraint
W/o clr. rfn.W/o str. rfn.W/o rfn.W/o MVCW/o SSCW/o cnst.Ours
PSNR↑24.4224.6824.3924.7324.1623.3525.08
SSIM↑0.8980.8960.8910.8970.8870.8790.905
LPIPS↓0.1320.1360.1390.1290.1430.1540.123
NIQE↓4.5014.6454.8684.5534.7094.5614.322
FID↓49.1353.7160.1251.0656.4663.6846.58
+ +![](images/9f86f49c4ca7b66023cde984905d025df4a3b69d9fefa01730740d5ec21415ff.jpg) +Fig. 6: The effect of the iterative condition refinement strategy. For each result, we place the corresponding final color and structure condition at its lower left and lower right. Please zoom in for details, respectively. + +![](images/0e2c9733bd9676f7a4b1d69337c22d0ab1bbc1100f7b21840669a062f2828af0.jpg) +Fig. 7: The effect of the color and structure constraints in the multi-condition constraint mechanism. For each result, we place the corresponding final color and structure condition at its lower left and lower right, respectively. Please zoom in for details. + +# 5 Conclusion + +This paper introduces the first language-based diffusion model for single image reflection removal, which exploits auxiliary semantic information from language descriptions and generative priors from diffusion models to recover transmission layers from mixture images with low-transmitted reflection regions. An iterative condition refinement strategy is proposed to address the problem of inaccurate control conditions, and a multi-condition constraint mechanism is introduced to ensure recovery faithfulness. Experiments on real data demonstrate the effectiveness of the proposed method. + +Limitations. The proposed method may incorrectly identify transmission and reflection layers if they exhibit similar image content. For such cases, more accurate spatial guidance from users such as annotated edges or region masks could help to relieve the ambiguity, which is left as our future work. + +# Acknowledgement + +This work is supported by National Natural Science Foundation of China under Grant No. 62136001, 62088102. + +# References + +1. Chang, Y., Jung, C., Sun, J.: Joint reflection removal and depth estimation from a single image. IEEE Transactions on Cybernetics (2020) +2. Chang, Y., Jung, C., Sun, J., Wang, F.: Siamese dense network for reflection removal with flash and no-flash image pairs. International Journal of Computer Vision (2020) +3. Chang, Z., Weng, S., Li, Y., Li, S., Shi, B.: L-CoDer: Language-based colorization with color-object decoupling transformer. In: Proc. of European Conference on Computer Vision (2022) +4. Chang, Z., Weng, S., Zhang, P., Li, Y., Li, S., Shi, B.: L-CAD: Language-based colorization with any-level descriptions using diffusion priors. In: Proc. of Advances in Neural Information Processing Systems (2023) +5. Chang, Z., Weng, S., Zhang, P., Li, Y., Li, S., Shi, B.: L-CoIns: Language-based colorization with instance awareness. In: Proc. of Computer Vision and Pattern Recognition (2023) +6. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dolkar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015) +7. Diamant, Y., Schechner, Y.Y.: Overcoming visual reverberations. In: Proc. of Computer Vision and Pattern Recognition (2008) +8. Dong, Z., Xu, K., Yang, Y., Bao, H., Xu, W., Lau, R.W.: Location-aware single image reflection removal. In: Proc. of International Conference on Computer Vision (2021) +9. Fan, Q., Yang, J., Hua, G., Chen, B., Wipf, D.: A generic deep architecture for single image reflection removal and image smoothing. In: Proc. of International Conference on Computer Vision (2017) +10. Han, B.J., Sim, J.Y.: Zero-shot learning for reflection removal of single 360-degree image. In: Proc. of European Conference on Computer Vision (2022) +11. He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In: Proc. of International Conference on Computer Vision (2015) +12. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022) +13. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Proc. of Advances in Neural Information Processing Systems (2017) +14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Proc. of Advances in Neural Information Processing Systems (2020) +15. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) +16. Hong, Y., Chang, Y., Liang, J., Ma, L., Huang, T., Shi, B.: Light flickering guided reflection removal. International Journal of Computer Vision (2024) + +17. Hong, Y., Lyu, Y., Li, S., Cao, G., Shi, B.: Reflection removal with NIR and RGB image feature fusion. IEEE Transactions on Multimedia (2022) +18. Hong, Y., Lyu, Y., Li, S., Shi, B.: Near-infrared image guided reflection removal. In: Proc. of International Conference on Multimedia and Expo (2020) +19. Hong, Y., Zheng, Q., Zhao, L., Jiang, X., Kot, A.C., Shi, B.: Panoramic image reflection removal. In: Proc. of Computer Vision and Pattern Recognition (2021) +20. Hong, Y., Zheng, Q., Zhao, L., Jiang, X., Kot, A.C., Shi, B.: PAR $^2$ Net: End-to-end panoramic image reflection removal. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +21. Hu, Q., Guo, X.:trash or treasure? an interactive dual-stream strategy for single image reflection separation. Proc. of Advances in Neural Information Processing Systems (2021) +22. Hu, Q., Guo, X.: Single image reflection separation via component synergy. In: Proc. of International Conference on Computer Vision (2023) +23. Huynh-Thu, Q., Ghanbari, M.: Scope of validity of psnr in image/video quality assessment. Electronics letters (2008) +24. Kim, S., Huo, Y., Yoon, S.E.: Single image reflection removal with physically-based training images. In: Proc. of Computer Vision and Pattern Recognition (2020) +25. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +26. Kong, N., Tai, Y.W., Shin, S.Y.: A physically-based approach to reflection separation. In: Proc. of Computer Vision and Pattern Recognition (2012) +27. Lei, C., Chen, Q.: Robust reflection removal with reflection-free flash-only cues. In: Proc. of Computer Vision and Pattern Recognition (2021) +28. Lei, C., Huang, X., Zhang, M., Yan, Q., Sun, W., Chen, Q.: Polarized reflection removal with perfect alignment in the wild. In: Proc. of Computer Vision and Pattern Recognition (2020) +29. Lei, C., Jiang, X., Chen, Q.: Robust reflection removal with flash-only cues in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +30. Levin, A., Weiss, Y.: User assisted separation of reflections from a single image using a sparsity prior. IEEE Transactions on Pattern Analysis and Machine Intelligence 29(9), 1647-1654 (2007) +31. Li, C., Yang, Y., He, K., Lin, S., Hopcroft, J.E.: Single image reflection removal through cascaded refinement. In: Proc. of Computer Vision and Pattern Recognition (2020) +32. Li, Y., Brown, M.S.: Exploiting reflection change for automatic reflection removal. In: Proc. of International Conference on Computer Vision (2013) +33. Li, Y., Brown, M.S.: Single image layer separation using relative smoothness. In: Proc. of Computer Vision and Pattern Recognition (2014) +34. Liu, Y.L., Lai, W.S., Yang, M.H., Chuang, Y.Y., Huang, J.B.: Learning to see through obstructions. In: Proc. of Computer Vision and Pattern Recognition (2020) +35. Liu, Y.L., Lai, W.S., Yang, M.H., Chuang, Y.Y., Huang, J.B.: Learning to see through obstructions with layered decomposition. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021) +36. Luo, J., Fu, J., Kong, X., Gao, C., Ren, H., Shen, H., Xia, H., Liu, S.: 3D-SPS: Single-stage 3D visual grounding via referred point progressive selection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16454-16463 (June 2022) +37. Lyu, Y., Cui, Z., Li, S., Pollefeys, M., Shi, B.: Reflection separation using a pair of unpolarized and polarized images. In: Proc. of Advances in Neural Information Processing Systems (2019) + +38. Lyu, Y., Cui, Z., Li, S., Pollefeys, M., Shi, B.: Physics-guided reflection separation from a pair of unpolarized and polarized images. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022) +39. Ma, D., Wan, R., Shi, B., Kot, A.C., Duan, L.Y.: Learning to jointly generate and separate reflections. In: Proc. of International Conference on Computer Vision (2019) +40. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: Proc. of International Conference on Learning Representations (2021) +41. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a "completely blind" image quality analyzer. IEEE Signal Processing Letters 20(3), 209-212 (2013) +42. Mou, C., Wang, X., Xie, L., Wu, Y., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023) +43. Nam, S., Brubaker, M.A., Brown, M.S.: Neural image representations for multi-image fusion and layer separation. In: Proc. of European Conference on Computer Vision (2022) +44. Nayar, S.K., Fang, X.S., Boult, T.: Separation of reflection components using color and polarization. International Journal of Computer Vision (1997) +45. Park, J., Kim, H., Park, E., Sim, J.Y.: Fully-automatic reflection removal for 360-degree images. In: Proc. of Winter Conference on Applications of Computer Vision (2024) +46. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: Proc. of Advances in Neural Information Processing Systems (2019) +47. Qiu, J., Jiang, P.T., Zhu, Y., Yin, Z.X., Cheng, M.M., Ren, B.: Looking through the glass: Neural surface reconstruction against high specular reflections. In: Proc. of Computer Vision and Pattern Recognition (2023) +48. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proc. of International Conference on Machine Learning. PMLR (2021) +49. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proc. of Computer Vision and Pattern Recognition (2022) +50. Schechner, Y.Y., Kiryati, N., Basri, R.: Separation of transparent layers using focus. International Journal of Computer Vision (2000) +51. Shih, Y., Krishnan, D., Durand, F., Freeman, W.T.: Reflection removal using ghosting cues. In: Proc. of Computer Vision and Pattern Recognition (2015) +52. Simon, C., Kyu Park, I.: Reflection removal for in-vehicle black box videos. In: Proc. of Computer Vision and Pattern Recognition (2015) +53. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) +54. Sun, H., Li, W., Liu, J., Chen, H., Pei, R., Zou, X., Yan, Y., Yang, Y.: CoSeR: Bridging image and language for cognitive super-resolution. arXiv preprint arXiv:2311.16512 (2023) +55. Sun, J., Weng, S., Chang, Z., Li, S., Shi, B.: UniCoRN: A unified conditional image repainting network. In: Proc. of Computer Vision and Pattern Recognition (2022) + +56. Tang, J., Zhong, H., Weng, S., Shi, B.: LuminAIRe: Illumination-aware conditional image painting for lighting-realistic generation. In: Proc. of Advances in Neural Information Processing Systems (2023) +57. Wan, R., Shi, B., Duan, L.Y., Tan, A.H., Gao, W., Kot, A.C.: Region-aware reflection removal with unified content and gradient priors. IEEE Transactions on Image Processing (2018) +58. Wan, R., Shi, B., Duan, L.Y., Tan, A.H., Kot, A.C.: Benchmarking single-image reflection removal algorithms. In: Proc. of International Conference on Computer Vision (2017) +59. Wan, R., Shi, B., Duan, L.Y., Tan, A.H., Kot, A.C.: CRRN: Multi-scale guided concurrent reflection removal network. In: Proc. of Computer Vision and Pattern Recognition (2018) +60. Wan, R., Shi, B., Li, H., Duan, L.Y., Kot, A.C.: Face image reflection removal. International Journal of Computer Vision (2021) +61. Wan, R., Shi, B., Li, H., Duan, L.Y., Tan, A.H., Kot, A.C.: CoRRN: Cooperative reflection removal network. IEEE Transactions on Pattern Analysis and Machine Intelligence (2019) +62. Wan, R., Shi, B., Li, H., Hong, Y., Duan, L.Y., Kot, A.C.: Benchmarking single-image reflection removal algorithms. In: IEEE Transactions on Pattern Analysis and Machine Intelligence (2022) +63. Wang, Z., Lu, Y., Li, Q., Tao, X., Guo, Y., Gong, M., Liu, T.: CRIS: CLIP-driven referring image segmentation. In: Proc. of Computer Vision and Pattern Recognition (2022) +64. Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers (2003) +65. Wei, K., Yang, J., Fu, Y., Wipf, D., Huang, H.: Single image reflection removal exploiting misaligned training data and network enhancements. In: Proc. of Computer Vision and Pattern Recognition (2019) +66. Wen, Q., Tan, Y., Qin, J., Liu, W., Han, G., He, S.: Single image reflection removal beyond linearity. In: Proc. of Computer Vision and Pattern Recognition (2019) +67. Weng, S., Li, W., Li, D., Jin, H., Shi, B.: MISC: Multi-condition injection and spatially-adaptive compositing for conditional person image synthesis. In: Proc. of Computer Vision and Pattern Recognition (2020) +68. Weng, S., Shi, B.: Conditional image inpainting. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +69. Weng, S., Wu, H., Chang, Z., Tang, J., Li, S., Shi, B.: L-CoDe: Language-based colorization using color-object decoupled conditions. In: Proc. of the AAAI Conference on Artificial Intelligence (2022) +70. Yang, J., Gong, D., Liu, L., Shi, Q.: Seeing deeply and bidirectionally: A deep learning approach for single image reflection removal. In: Proc. of European Conference on Computer Vision (2018) +71. Yang, Y., Ma, W., Zheng, Y., Cai, J.F., Xu, W.: Fast single image reflection suppression via convex optimization. In: Proc. of Computer Vision and Pattern Recognition (2019) +72. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: LAVT: Language-aware vision transformer for referring image segmentation. In: Proc. of Computer Vision and Pattern Recognition (2022) +73. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics (2014) + +74. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proc. of International Conference on Computer Vision (2023) +75. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proc. of Computer Vision and Pattern Recognition (2018) +76. Zhang, X., Ng, R., Chen, Q.: Single image reflection separation with perceptual losses. In: Proc. of Computer Vision and Pattern Recognition (2018) +77. Zhang, Y.N., Shen, L., Li, Q.: Content and gradient model-driven deep network for single image reflection removal. In: Proc. of ACM International Conference on Multimedia (2022) +78. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Unictrlnet: All-in-one control to text-to-image diffusion models. Proc. of Advances in Neural Information Processing Systems (2024) +79. Zheng, Q., Chen, J., Lu, Z., Shi, B., Jiang, X., Yap, K.H., Duan, L.Y., Kot, A.C.: What does plate glass reveal about camera calibration? In: Proc. of Computer Vision and Pattern Recognition (2020) +80. Zheng, Q., Shi, B., Chen, J., Jiang, X., Duan, L.Y., Kot, A.C.: Single image reflection removal with absorption effect. In: Proc. of Computer Vision and Pattern Recognition (2021) +81. Zhong, H., Hong, Y., Weng, S., Liang, J., Shi, B.: Language-guided image reflection separation. In: Proc. of Computer Vision and Pattern Recognition (2024) \ No newline at end of file diff --git a/2024/L-DiffER_ Single Image Reflection Removal with Language-based Diffusion Model/images.zip b/2024/L-DiffER_ Single Image Reflection Removal with Language-based Diffusion Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..968a9379055c53fe3341df9796db9c2e7fa4d94d --- /dev/null +++ b/2024/L-DiffER_ Single Image Reflection Removal with Language-based Diffusion Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ac8a5b64bcf9a6b6f185e660c1be7ed0beb086a7eb87bb9cb41ec6beda80a86 +size 719016 diff --git a/2024/L-DiffER_ Single Image Reflection Removal with Language-based Diffusion Model/layout.json b/2024/L-DiffER_ Single Image Reflection Removal with Language-based Diffusion Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..39356430f2115ce23d0fc32d8dc2fdff8a883a96 --- /dev/null +++ b/2024/L-DiffER_ Single Image Reflection Removal with Language-based Diffusion Model/layout.json @@ -0,0 +1,11282 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "spans": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "type": "text", + "content": "L-DiffER: Single Image Reflection Removal with Language-based Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 174, + 167, + 438, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 167, + 438, + 194 + ], + "spans": [ + { + "bbox": [ + 174, + 167, + 438, + 194 + ], + "type": "text", + "content": "Yuchen Hong" + }, + { + "bbox": [ + 174, + 167, + 438, + 194 + ], + "type": "inline_equation", + "content": "^{1,2\\#}" + }, + { + "bbox": [ + 174, + 167, + 438, + 194 + ], + "type": "text", + "content": " Haofeng Zhong" + }, + { + "bbox": [ + 174, + 167, + 438, + 194 + ], + "type": "inline_equation", + "content": "^{1,2,3\\#}" + }, + { + "bbox": [ + 174, + 167, + 438, + 194 + ], + "type": "text", + "content": " Shuchen Weng" + }, + { + "bbox": [ + 174, + 167, + 438, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 174, + 167, + 438, + 194 + ], + "type": "text", + "content": " Jinxiu Liang" + }, + { + "bbox": [ + 174, + 167, + 438, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 174, + 167, + 438, + 194 + ], + "type": "text", + "content": " Boxin Shi" + }, + { + "bbox": [ + 174, + 167, + 438, + 194 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 176, + 201, + 436, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 201, + 436, + 224 + ], + "spans": [ + { + "bbox": [ + 176, + 201, + 436, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 176, + 201, + 436, + 224 + ], + "type": "text", + "content": " State Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 224, + 432, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 224, + 432, + 246 + ], + "spans": [ + { + "bbox": [ + 179, + 224, + 432, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 179, + 224, + 432, + 246 + ], + "type": "text", + "content": " National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 161, + 246, + 452, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 246, + 452, + 280 + ], + "spans": [ + { + "bbox": [ + 161, + 246, + 452, + 280 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 161, + 246, + 452, + 280 + ], + "type": "text", + "content": " AI Innovation Center, School of Computer Science, Peking University {hfzhong, shuchenweng, cssherryliang, shiboxin}@pku.edu.cn, yuchenhong.cn@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 308, + 455, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 308, + 455, + 441 + ], + "spans": [ + { + "bbox": [ + 160, + 308, + 455, + 441 + ], + "type": "text", + "content": "Abstract. In this paper, we introduce L-DiffER, a language-based diffusion model designed for the ill-posed single image reflection removal task. Although having shown impressive performance for image generation, existing language-based diffusion models struggle with precise control and faithfulness in image restoration. To overcome these limitations, we propose an iterative condition refinement strategy to resolve the problem of inaccurate control conditions. A multi-condition constraint mechanism is employed to ensure the recovery faithfulness of image color and structure while retaining the generation capability to handle low-transmitted reflections. We demonstrate the superiority of the proposed method through extensive experiments, showcasing both quantitative and qualitative improvements over existing methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 451, + 429, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 451, + 429, + 463 + ], + "spans": [ + { + "bbox": [ + 160, + 451, + 429, + 463 + ], + "type": "text", + "content": "Keywords: Reflection removal " + }, + { + "bbox": [ + 160, + 451, + 429, + 463 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 451, + 429, + 463 + ], + "type": "text", + "content": " Language-based diffusion model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 501, + 230, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 230, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 230, + 514 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 526, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 482, + 646 + ], + "type": "text", + "content": "When photographs are taken through transparent mediums like glass windows, the contaminated mixture image (denoted as " + }, + { + "bbox": [ + 130, + 526, + 482, + 646 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 526, + 482, + 646 + ], + "type": "text", + "content": ") can be considered as the combination of a transmission layer (denoted as " + }, + { + "bbox": [ + 130, + 526, + 482, + 646 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 526, + 482, + 646 + ], + "type": "text", + "content": ") and a reflection layer (denoted as " + }, + { + "bbox": [ + 130, + 526, + 482, + 646 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 130, + 526, + 482, + 646 + ], + "type": "text", + "content": ") [61, 62], which impairs the performance of downstream computer vision tasks [1, 47, 60]. Consequently, reflection removal, which aims at removing undesired reflections and recovering clear transmission layers from contaminated mixture images, has become an attractive topic in the field of computational photography [8, 9, 16, 31, 59, 65, 77]. State-of-the-art single-image reflection removal methods [8, 22, 31, 61] predominantly learn deep priors from a mixed dataset of synthetic and real data to mitigate reflections. However, due to insufficient" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 653, + 328, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 653, + 328, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 653, + 328, + 665 + ], + "type": "text", + "content": "# Equal contributions. * Corresponding author." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 249, + 252 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 249, + 252 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 249, + 252 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 249, + 252 + ], + "type": "image", + "image_path": "e7737be1bfb2e3a48f3a32aa083679e9febad2f0975eead01b83fcc556c4e085.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 261, + 482, + 327 + ], + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 327 + ], + "type": "text", + "content": "Fig. 1: For mixture images with strong reflections, (a-b) single-image reflection removal methods [22, 31] fail in distinguishing reflection and transmission layers. (c) Using language descriptions as guidance and mixture images as conditions, ControlNet [74] generates results with color shifts and structure distortions, while (d) the proposed L-DiffER successfully achieves high-fidelity transmission recovery in low-transmitted reflection regions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 249, + 114, + 482, + 252 + ], + "blocks": [ + { + "bbox": [ + 249, + 114, + 482, + 252 + ], + "lines": [ + { + "bbox": [ + 249, + 114, + 482, + 252 + ], + "spans": [ + { + "bbox": [ + 249, + 114, + 482, + 252 + ], + "type": "image", + "image_path": "3d3eaafcade95108a6cb7e3dfe936b46f442b6e1d2492b7827f3923fbcdef73c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 353, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 484 + ], + "type": "text", + "content": "knowledge about transmission and reflection scenes for addressing such a highly ill-posed problem, they often encounter limitations in scenes with complex or low-transmitted reflections, as shown in yellow boxes of Fig. 1(a-b). By leveraging auxiliary scene information from additional inputs obtained with specialized devices [27, 28, 37] or capture settings [34, 35, 43], multi-image reflection removal methods achieve more robust reflection removal than single-image methods. However, the specialized requirements for data acquisition, in contrast, limit their application scopes, especially for mobile devices and images from the Internet. Hence, it is imperative to find a user-friendly auxiliary input that helps to relieve the ill-posedness of reflection removal while maintaining the applicability and accessibility of single-image methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 485, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 593 + ], + "type": "text", + "content": "Fortunately, natural language can exactly serve as the aforementioned auxiliary input, since it provides instructive semantic information about images [36, 63, 72], which bridges the gap between human understanding and machine perception. Thanks to the development of vision-language models [48], various vision tasks (e.g., image editing [55, 56, 67, 68] and image colorization [3-5, 69]) benefit from the semantic guidance of language descriptions. Zhong et al. [81] make the first attempt to introduce language descriptions for reflection removal, but due to the lack of generation capability and generalizability, it encounters difficulties in addressing scenarios containing low-transmitted reflections." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Recently, by combining semantic information from natural language descriptions and generative priors from pre-trained diffusion models [14, 49], language-based diffusion models (e.g., ControlNet [74]) generate visually pleasant results for image restoration tasks like image colorization [4] and super resolution [54]. For reflection removal, it also shows the potential to provide semantic priors to relieve the ill-posedness, where the generative priors in pre-trained diffusion" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Y. Hong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": "models can adeptly manage complex scenarios and make it possible to recover transmission layers under low-transmitted reflections. However, as shown in the red boxes of Fig. 1(c), directly employing the existing language-based diffusion model like ControlNet [74] for reflection removal will encounter two primary challenges: (1) Inaccurate control conditions. When using language-based diffusion models, other tasks such as image colorization [4] often employ a clear (not superimposed) image to provide accurate structure control. However, in the reflection removal task, the image content of the transmission layer in a mixture image is contaminated by the reflection layer, and thus it is unable to provide an initially accurate condition for both structure and color control. Consequently, achieving transmission recovery under such inaccurate conditions presents a significant challenge. (2) Insufficient recovery faithfulness. Since diffusion models are principally designed for image generation, simply applying them for reflection removal may obtain results that are less faithful to conditions (i.e., results with unwanted image content containing color shifts and structural distortions compared with mixture images). Therefore, it is essential to develop a mechanism for constraining the generation capability and improving the recovery faithfulness of diffusion models to recover transmission layers while preventing unwanted alterations in image color and structure." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 347, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 540 + ], + "type": "text", + "content": "In this paper, we propose L-DiffER, the first Language-based Diffusion model for single image rEflection Removal, to achieve high-fidelity transmission recovery with language guidance under the interference of low-transmitted reflections, as the results shown in Fig. 1(d). To leverage the language descriptions of two layers, descriptions of the transmission layer serve as positive prompts, guiding the model to preserve essential image content, whereas descriptions of the reflection layer act as negative prompts, aiding in the suppression of unwanted reflections. To address the aforementioned challenges, we propose an iterative condition refinement strategy within the reverse diffusion process to ensure increasingly accurate representations of color and structure conditions as the process evolves. Furthermore, we introduce a multi-condition constraint mechanism designed to safeguard the faithfulness to the specified conditions, which effectively addresses potential color shifts and structural distortions by applying constraints that align the recovery output with the desired color and structure conditions. The proposed method achieves state-of-the-art performance by making the following contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 562, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 562, + 479, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 562, + 479, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 562, + 479, + 597 + ], + "type": "text", + "content": "- We present the first language-based diffusion model for reflection removal, leveraging the descriptive power of language to distinguish between transmission and reflection layers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 601, + 480, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 624 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 624 + ], + "type": "text", + "content": "- We propose an iterative condition refinement strategy to provide more accurate control conditions as color and structure guidance for reflection removal." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 629, + 480, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 629, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 138, + 629, + 480, + 664 + ], + "type": "text", + "content": "- We design a multi-condition constraint mechanism to ensure the recovery faithfulness of image color and structure, which effectively handles low-transmitted reflections." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "L-DiffER" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 376 + ], + "type": "text", + "content": "Reflection removal. Single image reflection removal aims at suppressing reflections and recovering transmission layers in a single mixture image. Using the presumption that reflections tend to blur and exhibit low intensities, nonlearning approaches employ handcrafted priors into their optimization frameworks [57, 71], such as the gradient sparsity [30], relative smoothness [33], and ghosting cues [51]. With the development of deep learning, researchers attempt to address the task by exploring different network architectures [9, 21, 59, 61, 65], loss functions [76], or the iterative recovery strategy [8, 31, 70, 77]. Concurrently, efforts in data synthesis to meet the data-driven demands is also an attractive topic in this area [9, 22, 24, 39, 66, 79, 80]. Besides, the exploration of special image forms like panoramic images [10, 19, 20, 45] helps to diminish content ambiguity in mixture images. To relieve the ill-posedness of reflection removal, multi-image methods utilize additional input images to introduce auxiliary clues such as polarization information [7, 26, 28, 37, 38, 44, 50], transmission scene information from active illuminations [2, 17, 18, 27, 29], and motion discrepancy between the two layers [32, 34, 35, 52], while special data capture requirements limit their practical applications. Inspired by Zhong et al. [81], we introduce language-based diffusion models to input auxiliary language descriptions for relieving the ill-posedness of reflection removal while maintaining the applicability of the method, and to leverage generative priors for handling low-transmitted reflection regions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 376, + 484, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 484, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 484, + 567 + ], + "type": "text", + "content": "Language-based diffusion models. The advent of diffusion models [14] (DMs) has marked a significant leap forward, particularly in the realms of image generation [49]. The emergence of pre-trained vision-language models, notably CLIP [48], significantly enhances the flexibility and intuitiveness of utilizing language descriptions for vision tasks. Specifically, these pre-trained models have been instrumental in enabling multi-condition image editing [42,74,78], image colorization [4], and devising sophisticated sampling strategies for tailored generation outcomes [12]. Despite the strides in image editing and conditional generation, the specific challenge of single image reflection removal, i.e., recovering clear transmission layers from the guidance of a superimposed mixture images, still remains unexplored for language-based diffusion models [49, 74]. In our work, we seek to address the reflection removal problem by harnessing the generative prowess and flexibility of language-guided diffusion models, which combines the semantic information offered by language descriptions with the sophisticated generative capabilities of DMs while ensuring the recovery fidelity and faithfulness with condition refinement and constraints." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 583, + 233, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 583, + 233, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 583, + 233, + 597 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "In this paper, we address the problem of language-based reflection removal using diffusion models. We first provide a brief overview of the framework of our language-based reflection removal diffusion model L-DiffER in Sec. 3.1. Sec. 3.2 proposes the iterative condition refinement strategy designed for addressing inaccurate control conditions, and Sec. 3.3 proposes the multi-condition constraint" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Y. Hong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 357, + 283 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 357, + 283 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 357, + 283 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 357, + 283 + ], + "type": "image", + "image_path": "d6231c5081ffb347fde3bb3531b922f76ac69d15cc683cf5f5b51c2928ac4fb2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 293, + 480, + 326 + ], + "lines": [ + { + "bbox": [ + 131, + 293, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 131, + 293, + 480, + 326 + ], + "type": "text", + "content": "Fig. 2: Framework of the proposed language-based reflection removal diffusion model, which leverages language descriptions with the iterative condition refinement strategy and the multi-condition constraint mechanism to achieve faithful transmission recovery." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 359, + 116, + 477, + 283 + ], + "blocks": [ + { + "bbox": [ + 359, + 116, + 477, + 283 + ], + "lines": [ + { + "bbox": [ + 359, + 116, + 477, + 283 + ], + "spans": [ + { + "bbox": [ + 359, + 116, + 477, + 283 + ], + "type": "image", + "image_path": "f1a9c04a4c585ba8296882b45ef9f4b3b4354b31dd6fb4ffd7489a0ddcac0830.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 350, + 479, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 350, + 479, + 374 + ], + "spans": [ + { + "bbox": [ + 131, + 350, + 479, + 374 + ], + "type": "text", + "content": "mechanism for improving the recovery faithfulness. Finally, loss functions and training details are presented in Sec. 3.4 and Sec. 3.5, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 393, + 208, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 393, + 208, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 393, + 208, + 403 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "content": "Problem formulation. Given the mixture image " + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "content": " and the language description condition " + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "content": ", we optimize our model to remove reflection contamination and recover transmission layers " + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{T}} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "content": " under the supervision of the ground truth transmission layers " + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{T} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "content": ". We apply the commonly-used assumption [21,31,61] as: " + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = \\mathbf{T} + \\mathbf{R}" + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "content": ", where the mixture image " + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "content": " can be considered as the combination of a transmission layer " + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "content": " and a reflection layer " + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "content": ". We use the intensity and gradient of the mixture image " + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "content": " to initialize the color condition " + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{s}} = \\mathbf{M}" + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "content": " and the structure condition " + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{s}} = \\nabla \\mathbf{M}" + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "content": ". Then the color latent " + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "inline_equation", + "content": "y^{\\mathrm{c}} = \\mathcal{E}^{\\mathrm{c}}(\\mathbf{I}^{\\mathrm{c}})" + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "content": " and structure latent " + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "inline_equation", + "content": "y^{\\mathrm{s}} = \\mathcal{E}^{\\mathrm{s}}(\\mathbf{I}^{\\mathrm{s}})" + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "content": " are extracted by a color and a structure encoder " + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{\\mathrm{c}}" + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{\\mathrm{s}}" + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "content": " (denoted by the blue blocks in the middle part of Fig. 2), respectively. Given language descriptions about the content of transmission layers and reflection layers as positive prompt and negative prompt, respectively, the positive and negative latents " + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "inline_equation", + "content": "c^{\\mathrm{p}}" + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "inline_equation", + "content": "c^{\\mathrm{n}}" + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "content": " are extracted from the CLIP text encoder [48]. In the proposed method, the transmission layer " + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{T}}" + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "content": " is recovered from all the spatial and language conditioning information encoded in latents " + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "inline_equation", + "content": "y = [y^{\\mathrm{s}}, y^{\\mathrm{c}}]" + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "inline_equation", + "content": "c = [c^{\\mathrm{p}}, c^{\\mathrm{n}}]" + }, + { + "bbox": [ + 130, + 413, + 482, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Diffusion models. Diffusion models [14, 53] are probabilistic generative models that have been shown to have powerful data modeling ability for images, which can provide an effective image prior for the estimation of transmission layers. To address the challenges of resource consumption in pixel space, Stable Diffusion (SD) [49] introduces latent diffusion models to improve both the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "L-DiffER" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": "training and sampling efficiency of denoising diffusion models without degrading their quality. Following SD [49], we employ a compression encoder " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " to encode a transmission layer into the latent space as " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "z = \\mathcal{E}(\\mathbf{T})" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": ", and a compression decoder " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " to reconstruct the transmission layer from the latent code " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{T}} = \\mathcal{D}(z)" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": ". The diffusion process progressively destructs data by injecting noise, then learns to reverse this process for image generation. During training, the forward diffusion process starts from the target image " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "z_0 = \\mathcal{E}(\\mathbf{T})" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " and generates samples " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "z_t = \\sqrt{\\overline{\\alpha}_t} z_0 + \\sqrt{1 - \\overline{\\alpha}_t} \\epsilon_t" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\epsilon_t \\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " is the Gaussian noise at timestep " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\overline{\\alpha}_t" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " represents the noise scheduler introduced in [49]. During inference, the reverse diffusion process starts from a random noise sample " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "z_T \\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": ", where a noise prediction network denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\epsilon_\\theta" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " is optimized to predict the noise " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_t" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " at each timestep " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " given the language condition " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " and the spatial conditions " + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "inline_equation", + "content": "y_t^1" + }, + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 206, + 273, + 481, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 273, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 206, + 273, + 481, + 287 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {l d m}} = \\mathbb {E} _ {\\mathcal {E} (x), t, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon_ {t} - \\epsilon_ {\\theta} \\left(z _ {t}, c, y _ {t}, t\\right) \\| _ {2} ^ {2} \\right], \\tag {1}", + "image_path": "25f73576db5d70345d8cef6c5e18455cad0e918e74c1bdf5acb68a4a0276b2b2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 293, + 481, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 481, + 317 + ], + "type": "text", + "content": "until it converges to the desired photorealistic transmission layer " + }, + { + "bbox": [ + 130, + 293, + 481, + 317 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{T}}" + }, + { + "bbox": [ + 130, + 293, + 481, + 317 + ], + "type": "text", + "content": ". The overall framework is shown in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "text", + "content": "Condition extraction. During inference, the two conditioning latents start from the initial color and structure latents, i.e., " + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "inline_equation", + "content": "y_{T}^{c} = y^{c}" + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "inline_equation", + "content": "y_{T}^{\\mathrm{s}} = y^{\\mathrm{s}}" + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "text", + "content": ", respectively, and they are refined by employing the iterative condition refinement strategy (Sec. 3.2) to provide more accurate conditions. For condition injection, the proposed framework draws inspiration from ControlNet [74], which learns a parameter-efficient, parallel branch on top of SD. Specifically, two trainable copied modules (denoted by the pink blocks in Fig. 2) are utilized to take the color latent " + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "inline_equation", + "content": "y_{t}^{c}" + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "text", + "content": " and the gradient latent " + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "inline_equation", + "content": "y_{t}^{\\mathrm{s}}" + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "text", + "content": " with the latent presentation " + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "text", + "content": " as input to extract color and structure features, respectively. These features are then added directly to the corresponding scales of the locked module (denoted by the purple block in Fig. 2), which guides " + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "text", + "content": " in learning the noise distribution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 450, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 510 + ], + "type": "text", + "content": "Language guidance. To explicitly instruct the model to suppress reflections from the language guidance, classifier-free guidance [15] is used to employ negative prompts to specify undesired reflection contamination. Specifically, two noise predictions with positive prompts " + }, + { + "bbox": [ + 130, + 450, + 482, + 510 + ], + "type": "inline_equation", + "content": "c^{\\mathrm{p}}" + }, + { + "bbox": [ + 130, + 450, + 482, + 510 + ], + "type": "text", + "content": " and negative prompts " + }, + { + "bbox": [ + 130, + 450, + 482, + 510 + ], + "type": "inline_equation", + "content": "c^{\\mathrm{n}}" + }, + { + "bbox": [ + 130, + 450, + 482, + 510 + ], + "type": "text", + "content": " are made at each timestep " + }, + { + "bbox": [ + 130, + 450, + 482, + 510 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 450, + 482, + 510 + ], + "type": "text", + "content": ". The fusion of these two predictions yields the output " + }, + { + "bbox": [ + 130, + 450, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\hat{e}_{t-1}" + }, + { + "bbox": [ + 130, + 450, + 482, + 510 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 171, + 521, + 480, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 521, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 171, + 521, + 480, + 536 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} _ {t - 1} = \\epsilon_ {\\theta} \\left(z _ {t}, c ^ {\\mathrm {p}}, y _ {t}, t\\right) + \\lambda_ {\\mathrm {c f g}} \\times \\left(\\epsilon_ {\\theta} \\left(z _ {t}, c ^ {\\mathrm {p}}, y _ {t}, t\\right) - \\epsilon_ {\\theta} \\left(z _ {t}, c ^ {\\mathrm {n}}, y _ {t}, t\\right)\\right), \\tag {2}", + "image_path": "123aba71e43b5f679eaef9833bf6579c313c39b78a004654cd9813a479f5a318.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 583 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 545, + 482, + 583 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{cfg}}" + }, + { + "bbox": [ + 130, + 545, + 482, + 583 + ], + "type": "text", + "content": " is a hyperparameter. In our task, as reflections are not always recognizable [81], we set negative prompts as the empty string " + }, + { + "bbox": [ + 130, + 545, + 482, + 583 + ], + "type": "inline_equation", + "content": "c^{\\mathrm{n}} = \\varnothing" + }, + { + "bbox": [ + 130, + 545, + 482, + 583 + ], + "type": "text", + "content": " in the case. For simplicity, we use " + }, + { + "bbox": [ + 130, + 545, + 482, + 583 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,c,y_t,t)" + }, + { + "bbox": [ + 130, + 545, + 482, + 583 + ], + "type": "text", + "content": " to represent the whole noise prediction process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 601, + 356, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 601, + 356, + 613 + ], + "spans": [ + { + "bbox": [ + 131, + 601, + 356, + 613 + ], + "type": "text", + "content": "3.2 Iterative condition refinement strategy" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 621, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 621, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 621, + 482, + 646 + ], + "type": "text", + "content": "Although with effective data modeling ability, diffusion models probably generate results unfaithful to the input image conditions [40,74]. This issue becomes even" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Y. Hong et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 481, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 481, + 666 + ], + "type": "text", + "content": "1 Note that the spatial conditions are varied with timesteps in the proposed method." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 116, + 479, + 177 + ], + "blocks": [ + { + "bbox": [ + 135, + 116, + 479, + 177 + ], + "lines": [ + { + "bbox": [ + 135, + 116, + 479, + 177 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 479, + 177 + ], + "type": "image", + "image_path": "db5626b361337e2b81256db283df02a4906ce5444fd8c7d716afc2c28c57e045.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 186, + 480, + 208 + ], + "lines": [ + { + "bbox": [ + 132, + 186, + 480, + 208 + ], + "spans": [ + { + "bbox": [ + 132, + 186, + 480, + 208 + ], + "type": "text", + "content": "Fig. 3: Visualization of pseudo transmission layers by using the original mixture image with its gradient as conditions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 234, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 234, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 482, + 354 + ], + "type": "text", + "content": "harder for reflection removal, since the desired transmission layers should have a structure without gradients of reflection contaminations, which is different from the input mixture images. That is, the mixture images itself cannot provide an accurate image condition for diffusion-based reflection removal. Therefore, we turn to the recovered transmission layers in the inference process, which have the potential to be combined with the mixture images and help obtain a refined condition. During inference, given a latent noisy sample " + }, + { + "bbox": [ + 130, + 234, + 482, + 354 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 234, + 482, + 354 + ], + "type": "text", + "content": ", we can use the noise prediction network " + }, + { + "bbox": [ + 130, + 234, + 482, + 354 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 234, + 482, + 354 + ], + "type": "text", + "content": " to predict noise as " + }, + { + "bbox": [ + 130, + 234, + 482, + 354 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_t = \\epsilon_\\theta (z_t,t,c,y_t)" + }, + { + "bbox": [ + 130, + 234, + 482, + 354 + ], + "type": "text", + "content": ". Then we can calculate the estimated latent representation of the transmission layer " + }, + { + "bbox": [ + 130, + 234, + 482, + 354 + ], + "type": "inline_equation", + "content": "z_{0|t}" + }, + { + "bbox": [ + 130, + 234, + 482, + 354 + ], + "type": "text", + "content": " at the current timestep " + }, + { + "bbox": [ + 130, + 234, + 482, + 354 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 234, + 482, + 354 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 238, + 364, + 481, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 364, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 238, + 364, + 481, + 378 + ], + "type": "interline_equation", + "content": "z _ {0 \\mid t} = \\left(z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} * \\hat {\\epsilon} _ {t}\\right) / \\sqrt {\\bar {\\alpha} _ {t}}. \\tag {3}", + "image_path": "11a1e39e4b746ad594959dc592d7b6d8abb07d5bc31ce4294c059e2e536ab870.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 388, + 482, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 436 + ], + "type": "text", + "content": "After that, we can obtain the estimated pseudo transmission layer " + }, + { + "bbox": [ + 130, + 388, + 482, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{0|t}" + }, + { + "bbox": [ + 130, + 388, + 482, + 436 + ], + "type": "text", + "content": " by using the compression decoder " + }, + { + "bbox": [ + 130, + 388, + 482, + 436 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 388, + 482, + 436 + ], + "type": "text", + "content": " to decode " + }, + { + "bbox": [ + 130, + 388, + 482, + 436 + ], + "type": "inline_equation", + "content": "z_{0|t}" + }, + { + "bbox": [ + 130, + 388, + 482, + 436 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 130, + 388, + 482, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{0|t} = \\mathcal{D}(z_{0|t})" + }, + { + "bbox": [ + 130, + 388, + 482, + 436 + ], + "type": "text", + "content": ". Then " + }, + { + "bbox": [ + 130, + 388, + 482, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{0|t}" + }, + { + "bbox": [ + 130, + 388, + 482, + 436 + ], + "type": "text", + "content": " and its gradient " + }, + { + "bbox": [ + 130, + 388, + 482, + 436 + ], + "type": "inline_equation", + "content": "\\nabla \\mathbf{T}_{0|t}" + }, + { + "bbox": [ + 130, + 388, + 482, + 436 + ], + "type": "text", + "content": " can be used as the pseudo-accurate color and structure condition to obtain conditions at the next timestep." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 581 + ], + "type": "text", + "content": "As pseudo transmission layers shown in Fig. 3, we observe two key phenomena: Firstly, the estimated pseudo transmission layers in the first few sampling steps exhibit inaccuracies, with neither structure nor color serving effectively as guiding conditions. However, as the sampling steps progress, the fidelity of the reconstructed images notably improves. Secondly, although employing the mixture image and the gradient of the mixture image as conditions persistently yield inaccurate control, these conditions provide a richer set of usable information than the recovered image at the beginning of sampling. Leveraging the advantages of both observations, we introduce an iterative condition refinement strategy to address the challenge of inaccurate control conditions. Our control conditions transitioned from the mixture image and the gradient of the mixture image to the recovered image and its gradient as the sampling step increases." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 582, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 482, + 617 + ], + "type": "text", + "content": "Specifically, we introduce predefined time-variant coefficients " + }, + { + "bbox": [ + 132, + 582, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 132, + 582, + 482, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 582, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\gamma_{t}" + }, + { + "bbox": [ + 132, + 582, + 482, + 617 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 132, + 582, + 482, + 617 + ], + "type": "inline_equation", + "content": "t \\in \\{1,..,T\\}" + }, + { + "bbox": [ + 132, + 582, + 482, + 617 + ], + "type": "text", + "content": ") to combine pseudo-accurate conditions " + }, + { + "bbox": [ + 132, + 582, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{0|t}" + }, + { + "bbox": [ + 132, + 582, + 482, + 617 + ], + "type": "text", + "content": " with the one initiated from " + }, + { + "bbox": [ + 132, + 582, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 132, + 582, + 482, + 617 + ], + "type": "text", + "content": ", which decrease as timestep " + }, + { + "bbox": [ + 132, + 582, + 482, + 617 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 582, + 482, + 617 + ], + "type": "text", + "content": " increases² as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 164, + 629, + 481, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 629, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 164, + 629, + 481, + 643 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {t} ^ {\\mathrm {c}} = \\gamma_ {t} * \\mathbf {T} _ {0 | t} + (1 - \\gamma_ {t}) * \\mathbf {M}, \\quad \\mathbf {I} _ {t} ^ {\\mathrm {s}} = \\beta_ {t} * \\nabla \\mathbf {T} _ {0 | t} + (1 - \\beta_ {t}) * \\nabla \\mathbf {M}. \\tag {4}", + "image_path": "f4985375b10a6a4f30094bbad372d889cd882d27d3ce68cf7e449a35838cceae.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "L-DiffER" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 422, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 422, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 422, + 665 + ], + "type": "text", + "content": "2 Details of " + }, + { + "bbox": [ + 133, + 653, + 422, + 665 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 133, + 653, + 422, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 133, + 653, + 422, + 665 + ], + "type": "inline_equation", + "content": "\\gamma_{t}" + }, + { + "bbox": [ + 133, + 653, + 422, + 665 + ], + "type": "text", + "content": " will be explained in the supplementary material." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 117, + 480, + 273 + ], + "blocks": [ + { + "bbox": [ + 135, + 117, + 480, + 273 + ], + "lines": [ + { + "bbox": [ + 135, + 117, + 480, + 273 + ], + "spans": [ + { + "bbox": [ + 135, + 117, + 480, + 273 + ], + "type": "image", + "image_path": "f44879e18feae278d2a21e482f7d1266cac3258b01cbc0116e5b4627b9053417.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 282, + 480, + 304 + ], + "lines": [ + { + "bbox": [ + 131, + 282, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 131, + 282, + 480, + 304 + ], + "type": "text", + "content": "Fig. 4: Visualization of the color condition, structure condition, and the pseudo transmission layer by applying the proposed iterative condition refinement strategy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 327, + 479, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 479, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 479, + 350 + ], + "type": "text", + "content": "Fig. 4 shows intermediate conditions and the pseudo transmission layers during sampling, indicating the effect of the iterative condition refinement strategy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 350, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 480, + 422 + ], + "type": "text", + "content": "Meanwhile, we also use the iterative condition refinement strategy with a few modifications during training. We replace the pseudo-accurate transmission layers " + }, + { + "bbox": [ + 130, + 350, + 480, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{0|t}" + }, + { + "bbox": [ + 130, + 350, + 480, + 422 + ], + "type": "text", + "content": " estimated from sampling with ground truth " + }, + { + "bbox": [ + 130, + 350, + 480, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 350, + 480, + 422 + ], + "type": "text", + "content": " to refine conditions before estimating noises, which is different from the sampling process. Specifically, given the initial mixture image " + }, + { + "bbox": [ + 130, + 350, + 480, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 350, + 480, + 422 + ], + "type": "text", + "content": " and its gradient " + }, + { + "bbox": [ + 130, + 350, + 480, + 422 + ], + "type": "inline_equation", + "content": "\\nabla \\mathbf{M}" + }, + { + "bbox": [ + 130, + 350, + 480, + 422 + ], + "type": "text", + "content": ", we refine the input conditions at timestep " + }, + { + "bbox": [ + 130, + 350, + 480, + 422 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 350, + 480, + 422 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 172, + 430, + 480, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 430, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 172, + 430, + 480, + 444 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathbf {I}} _ {t} ^ {\\mathrm {c}} = \\gamma_ {t} * \\mathbf {T} + (1 - \\gamma_ {t}) * \\mathbf {M}, \\quad \\widehat {\\mathbf {I}} _ {t} ^ {\\mathrm {s}} = \\beta_ {t} * \\nabla \\mathbf {T} + (1 - \\beta_ {t}) * \\nabla \\mathbf {M}, \\tag {5}", + "image_path": "eb77ed19f0c0d5abf67b62dea699e5b95863d218ce3f3cdaed332ddc5688d7e7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "inline_equation", + "content": "\\gamma_{t}" + }, + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "inline_equation", + "content": "t \\in \\{1, \\dots, T\\}" + }, + { + "bbox": [ + 130, + 452, + 479, + 475 + ], + "type": "text", + "content": ") are the same predefined time-variant coefficients used in the sampling process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 491, + 352, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 352, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 352, + 502 + ], + "type": "text", + "content": "3.3 Multi-condition constraint mechanism" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": "In order to mitigate the insufficient recovery faithfulness of diffusion models, we employ a multi-condition mechanism that separately utilizes mixture images and their gradients to obtain initial latent conditions for controlling the color and structure of generated results. As the iterative condition refinement strategy obtains new refined conditions during each timestep, no constraints applying on conditions may cause the sampling process out of control, which also impacts the recovery faithfulness. Hence, it is imperative to conduct constraints on refined conditions. Though inaccurate as color conditions, mixture images provide coarse color information about transmission layers, thus we propose a maximum value constraint to prevent color shifts. Similarly, the gradient of the mixture image provides features showing structure information such as edges and textures. Therefore, we propose a saturation-aware structure constraint to prevent structure distortions when recovering transmission layers. Details are as follows." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Y. Hong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 200 + ], + "type": "text", + "content": "Maximum value constraint (MVC). According to the assumption defined in Sec. 3.1, the pixel values of the mixture image should be greater than or equal to the transmission layer. Therefore, when updating the conditions in the training and sampling processes, we employ maximum value constraint as a function to ensure that any given color condition " + }, + { + "bbox": [ + 131, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_t^{\\mathrm{c}}" + }, + { + "bbox": [ + 131, + 116, + 479, + 200 + ], + "type": "text", + "content": " defined in Eq. (4) does not exceed the mixture image " + }, + { + "bbox": [ + 131, + 116, + 479, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 131, + 116, + 479, + 200 + ], + "type": "text", + "content": " in terms of pixel values. This constraint should be adhered to the following:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 255, + 208, + 480, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 208, + 480, + 220 + ], + "spans": [ + { + "bbox": [ + 255, + 208, + 480, + 220 + ], + "type": "interline_equation", + "content": "\\operatorname {M V C} \\left(\\mathbf {I} _ {t} ^ {\\mathrm {c}}\\right) = \\min \\left(\\mathbf {M}, \\mathbf {I} _ {t} ^ {\\mathrm {c}}\\right). \\tag {6}", + "image_path": "df197b5509cd3f0c44842244b35063c834f577a26713e76157e59386b0ceaf01.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 228, + 480, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 228, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 480, + 346 + ], + "type": "text", + "content": "Saturation-aware structure constraint (SSC). Due to regional overexposure during photographing, mixture images sometimes contain low-transmitted reflection regions, which lack enough visible content of transmission layers, causing the corresponding gradients to have little usable structure information. Therefore, simply applying the maximum value constraint to edges will force the refined structure condition to abandon generated content in saturated regions, which may hinder optimization. Therefore, we introduce a saturation-aware mask to preserve the generative capability of the structural aspects in regions challenging for recovery. Specifically, we first utilize the mixture image to identify low-transmitted regions, generating a mask " + }, + { + "bbox": [ + 131, + 228, + 480, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^{\\mathrm{s}}\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 131, + 228, + 480, + 346 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 251, + 354, + 480, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 354, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 251, + 354, + 480, + 384 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Omega} ^ {\\mathrm {s}} = \\left\\{ \\begin{array}{l l} 0, & \\text {w h e r e} \\mathbf {M} > \\tau \\\\ 1, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {7}", + "image_path": "290163b6b20df8a03c117026c72b3a9b3ffe09995696167ae8443a2e6af41a38.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 389, + 479, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 389, + 479, + 437 + ], + "spans": [ + { + "bbox": [ + 131, + 389, + 479, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 389, + 479, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 131, + 389, + 479, + 437 + ], + "type": "text", + "content": " is the mixture image, " + }, + { + "bbox": [ + 131, + 389, + 479, + 437 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 389, + 479, + 437 + ], + "type": "text", + "content": " is the saturation threshold we empirically set to 0.95. Besides, to ensure that the structure condition does not contain nonexistent edges in non-saturated regions of the original mixture image, we define a mask to indicate the valid edge in the gradient of the mixture image as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 247, + 442, + 480, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 442, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 247, + 442, + 480, + 474 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Omega} ^ {\\mathrm {v}} = \\left\\{ \\begin{array}{l l} 0, & \\text {w h e r e} \\nabla \\mathbf {M} < \\eta \\\\ 1, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {8}", + "image_path": "5ffbc4bef08a04fe09d573fc235c7c747b99d2908b665a863b2338076dd1ac58.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 479, + 479, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 479, + 513 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 479, + 513 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 479, + 479, + 513 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 131, + 479, + 479, + 513 + ], + "type": "text", + "content": " is the valid edge threshold we empirically set to 0.05. After that, for a given gradient condition " + }, + { + "bbox": [ + 131, + 479, + 479, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_t^s" + }, + { + "bbox": [ + 131, + 479, + 479, + 513 + ], + "type": "text", + "content": " defined in Eq. (4), we define the saturation-aware structure constraint as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 219, + 522, + 480, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 522, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 219, + 522, + 480, + 534 + ], + "type": "interline_equation", + "content": "\\operatorname {S S C} \\left(\\mathbf {I} _ {t} ^ {\\mathrm {s}}\\right) = \\boldsymbol {\\Omega} ^ {\\mathrm {s}} \\odot \\mathbf {I} _ {t} ^ {\\mathrm {s}} + \\left(1 - \\boldsymbol {\\Omega} ^ {\\mathrm {s}}\\right) \\odot \\boldsymbol {\\Omega} ^ {\\mathrm {v}} \\odot \\mathbf {I} _ {t} ^ {\\mathrm {s}}. \\tag {9}", + "image_path": "80644536b6fb4ee056d0fe274f86bcadffe9d7a0fada11d7f5aa720c4b91c674.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 542, + 479, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 542, + 479, + 578 + ], + "spans": [ + { + "bbox": [ + 131, + 542, + 479, + 578 + ], + "type": "text", + "content": "To extract the color latent " + }, + { + "bbox": [ + 131, + 542, + 479, + 578 + ], + "type": "inline_equation", + "content": "y_{t}^{\\mathrm{c}}" + }, + { + "bbox": [ + 131, + 542, + 479, + 578 + ], + "type": "text", + "content": " and the structure latent " + }, + { + "bbox": [ + 131, + 542, + 479, + 578 + ], + "type": "inline_equation", + "content": "y_{t}^{\\mathrm{s}}" + }, + { + "bbox": [ + 131, + 542, + 479, + 578 + ], + "type": "text", + "content": " at each timestep " + }, + { + "bbox": [ + 131, + 542, + 479, + 578 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 542, + 479, + 578 + ], + "type": "text", + "content": ", we apply the maximum value constraint (MVC) and saturation-aware structure constraint (SSC) on the color condition " + }, + { + "bbox": [ + 131, + 542, + 479, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_t^{\\mathrm{c}}" + }, + { + "bbox": [ + 131, + 542, + 479, + 578 + ], + "type": "text", + "content": " and structure condition " + }, + { + "bbox": [ + 131, + 542, + 479, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_t^{\\mathrm{s}}" + }, + { + "bbox": [ + 131, + 542, + 479, + 578 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 212, + 586, + 480, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 586, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 212, + 586, + 480, + 598 + ], + "type": "interline_equation", + "content": "y _ {t - 1} ^ {\\mathrm {c}} = \\mathcal {E} ^ {\\mathrm {c}} \\left(\\operatorname {M V C} \\left(\\mathbf {I} _ {t} ^ {\\mathrm {c}}\\right)\\right), \\quad y _ {t - 1} ^ {\\mathrm {s}} = \\mathcal {E} ^ {\\mathrm {s}} \\left(\\operatorname {S S C} \\left(\\mathbf {I} _ {t} ^ {\\mathrm {s}}\\right)\\right), \\tag {10}", + "image_path": "b72ad40c7a43b5ec3ceef10204f1faec00cf91b2ea3ac9a1ec7a9800025642a5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{\\mathrm{s}}" + }, + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "text", + "content": " is the structure condition encoder, " + }, + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{\\mathrm{c}}" + }, + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "text", + "content": " is the color condition encoder, and MVC and SSC are applied to the color and structure conditions for ensuring recovery faithfulness. Finally, we apply the denoising process with the DDIM [53] sampling scheme to obtain the next noisy sample " + }, + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "text", + "content": ". Algorithm 1 outlines the complete sampling process." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "L-DiffER" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 130, + 481, + 329 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 473, + 129 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 473, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 473, + 129 + ], + "type": "text", + "content": "Algorithm 1 Iterative condition refinement strategy in the sampling process" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "lines": [ + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "text", + "content": "Require: Noise prediction network " + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "text", + "content": " Compression decoder " + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "text", + "content": " Color encoder " + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{\\mathrm{c}}" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "text", + "content": " Structure encoder " + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{\\mathrm{s}}" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "text", + "content": " \nInput: Mixture image M, Language condition " + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "c = [c^{\\mathrm{p}},c^{\\mathrm{n}}]" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "text", + "content": " , Initial spatial conditions " + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "[y^{\\mathrm{s}} = \\mathcal{E}^{\\mathrm{s}}(\\nabla \\mathbf{M}),y^{\\mathrm{c}} = \\mathcal{E}^{\\mathrm{c}}(\\mathbf{M})]" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "text", + "content": " , Coefficients " + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "\\{\\overline{\\alpha}_t\\}_{t = 1}^T" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "\\{\\beta_t\\}_{t = 1}^T" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "\\{\\gamma_t\\}_{t = 1}^T" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "text", + "content": " \nOutput: Recovered latent representation " + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "text", + "content": " \n1: " + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "z_{T}\\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "text", + "content": " \n2: " + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "y_{T} = [y^{s},y^{c}]" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "text", + "content": " \n3: for each " + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "t\\in [T,1]" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "text", + "content": " do \n4: " + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_t = \\epsilon_\\theta (z_t,c,y_t,t)" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "text", + "content": " \n5: if " + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "t > 1" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "text", + "content": " then \n6: " + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "z_{0|t} = (z_t - \\sqrt{1 - \\overline{\\alpha}_t} *\\hat{\\epsilon}_t) / \\sqrt{\\overline{\\alpha}_t}" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "text", + "content": " D Compute " + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "z_{0|t}" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "text", + "content": " by Eq. (3) \n7: " + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{0|t} = \\mathcal{D}(z_{0|t})" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "y_{t - 1}^{\\mathrm{c}} = \\mathcal{E}^{\\mathrm{c}}(\\mathrm{MVC}(\\gamma_t*\\mathbf{T}_{0|t} + (1 - \\gamma_t)*\\mathbf{M}))" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "text", + "content": " DRefine color condition \n9: " + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "y_{t - 1}^{\\mathrm{s}} = \\mathcal{E}^{\\mathrm{s}}(\\mathrm{SSC}(\\beta_t*\\nabla \\mathbf{T}_{0|t} + (1 - \\beta_t)*\\nabla \\mathbf{M}))" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "text", + "content": " DRefine structure condition \n10: " + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "y_{t - 1} = [y_{t - 1}^{\\mathrm{s}},y_{t - 1}^{\\mathrm{c}}]" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "text", + "content": " DUpdate spatial conditions \n11: end if \n12: " + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "inline_equation", + "content": "z_{t - 1} = \\sqrt{\\overline{\\alpha}_{t - 1}} z_{0|t} + \\sqrt{1 - \\overline{\\alpha}_{t - 1}}\\hat{\\epsilon}_t" + }, + { + "bbox": [ + 132, + 130, + 481, + 329 + ], + "type": "text", + "content": " DDIM \n13: end for" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 132, + 353, + 232, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 232, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 232, + 363 + ], + "type": "text", + "content": "3.4 Loss functions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 372, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 480, + 407 + ], + "type": "text", + "content": "In this section, we introduce the latent diffusion model loss to minimize the discrepancy between the learned latent space distribution and the target distribution, as well as the pixel losses to prevent color shifts and structure distortions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 407, + 480, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 480, + 468 + ], + "type": "text", + "content": "We follow the training objective of ControlNet [74] but replace task-specific conditions with the mixture image and the gradient of the mixture. Thus our network " + }, + { + "bbox": [ + 130, + 407, + 480, + 468 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 407, + 480, + 468 + ], + "type": "text", + "content": " learns to predict the noise added to the noisy latent " + }, + { + "bbox": [ + 130, + 407, + 480, + 468 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 407, + 480, + 468 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 407, + 480, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ldm}}" + }, + { + "bbox": [ + 130, + 407, + 480, + 468 + ], + "type": "text", + "content": " defined in Eq. (1). According to Eq. (3), we can obtain " + }, + { + "bbox": [ + 130, + 407, + 480, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{0|t}" + }, + { + "bbox": [ + 130, + 407, + 480, + 468 + ], + "type": "text", + "content": " in pixel space by decoding " + }, + { + "bbox": [ + 130, + 407, + 480, + 468 + ], + "type": "inline_equation", + "content": "z_{0|t}" + }, + { + "bbox": [ + 130, + 407, + 480, + 468 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 206, + 477, + 480, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 477, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 206, + 477, + 480, + 491 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {0 | t} = \\mathcal {D} \\left(z _ {0 | t}\\right) = \\mathcal {D} \\left(\\left(z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} * \\hat {\\epsilon} _ {t}\\right) / \\sqrt {\\bar {\\alpha} _ {t}}\\right). \\tag {11}", + "image_path": "16d313f4b69292dc177b663a4fc5421f4b13476fb1e22fbe61362a4a8af23c5a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 500, + 479, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 479, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 479, + 525 + ], + "type": "text", + "content": "Then RGB loss is introduced to constrain the consistency of the reconstructed transmission " + }, + { + "bbox": [ + 132, + 500, + 479, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{0|t}" + }, + { + "bbox": [ + 132, + 500, + 479, + 525 + ], + "type": "text", + "content": " and ground truth " + }, + { + "bbox": [ + 132, + 500, + 479, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 132, + 500, + 479, + 525 + ], + "type": "text", + "content": ", minimizing their errors as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 258, + 533, + 480, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 533, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 258, + 533, + 480, + 547 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {R G B}} = \\left\\| \\mathbf {T} - \\mathbf {T} _ {0 | t} \\right\\| _ {2} ^ {2}. \\tag {12}", + "image_path": "7ff05521e38eb2930fae8bf62508a87fb6e7d6f8fc9c5ab6056fce4b07d899db.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 555, + 479, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 479, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 479, + 578 + ], + "type": "text", + "content": "For better structure consistency, we use gradient loss to constrain the prediction's gradient and its ground truth as below:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 242, + 586, + 480, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 586, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 242, + 586, + 480, + 601 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {g r a d}} = \\left\\| \\nabla (\\mathbf {T}) - \\nabla \\left(\\mathbf {T} _ {0 | t}\\right) \\right\\| _ {2} ^ {2}, \\tag {13}", + "image_path": "e59f9a86fc9879c75bdbe614ccb941390d5073d94d8330256f8aa7bb4d9731c7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 609, + 480, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 480, + 644 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 480, + 644 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 609, + 480, + 644 + ], + "type": "inline_equation", + "content": "\\nabla (\\cdot)" + }, + { + "bbox": [ + 130, + 609, + 480, + 644 + ], + "type": "text", + "content": " denotes the computation of image gradients. Furthermore, we also compute the standard deviation and the mean of the predicted transmission and its ground truth, and use a " + }, + { + "bbox": [ + 130, + 609, + 480, + 644 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 609, + 480, + 644 + ], + "type": "text", + "content": " loss to minimize the color distortion:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 199, + 652, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 652, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 199, + 652, + 480, + 666 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {n u m}} = \\left\\| \\mu (\\mathbf {T}) - \\mu \\left(\\mathbf {T} _ {0 | t}\\right) \\right\\| _ {2} ^ {2} + \\left\\| \\sigma (\\mathbf {T}) - \\sigma \\left(\\mathbf {T} _ {0 | t}\\right) \\right\\| _ {2} ^ {2}, \\tag {14}", + "image_path": "916c147f2d504bba65fa5b8556310f2810a357a3f5cbb44baf8523eead3b8f7f.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Y. Hong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mu (\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " denote the computation of mean and standard deviation, respectively, and " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "i = 1,\\dots,N" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "-th sample in the dataset. Overall, the pixel loss is formulated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 242, + 163, + 480, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 163, + 480, + 174 + ], + "spans": [ + { + "bbox": [ + 242, + 163, + 480, + 174 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {p i x}} = \\mathcal {L} _ {\\mathrm {R G B}} + \\mathcal {L} _ {\\mathrm {g r a d}} + \\mathcal {L} _ {\\mathrm {n u m}}. \\tag {15}", + "image_path": "341ea9d56733a26c1d58f76da88681b744ae52be3a96e72e5cab3c7dd9cd6085.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 479, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 479, + 206 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 479, + 206 + ], + "type": "text", + "content": "Gathering the latent diffusion model loss and pixel losses yields the final objective as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 261, + 208, + 480, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 208, + 480, + 220 + ], + "spans": [ + { + "bbox": [ + 261, + 208, + 480, + 220 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l l}} := \\mathcal {L} _ {\\mathrm {l d m}} + \\lambda \\mathcal {L} _ {\\text {p i x}}, \\tag {16}", + "image_path": "cb1d3c86efec5463a7e3b12108679eb2360bc2c1aa90f97ac41c6496dc4d838e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 225, + 276, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 225, + 276, + 237 + ], + "spans": [ + { + "bbox": [ + 131, + 225, + 276, + 237 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 225, + 276, + 237 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 131, + 225, + 276, + 237 + ], + "type": "text", + "content": " is set to 1.0 empirically." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 255, + 277, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 255, + 277, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 255, + 277, + 266 + ], + "type": "text", + "content": "3.5 Implementation details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 274, + 480, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 480, + 334 + ], + "type": "text", + "content": "We implement the proposed method using PyTorch [46]. The model is trained for 20 epochs with a batch size of 16 on two NVIDIA GeForce RTX 3090 GPUs. Weights are initialized as in [11] and updated using the Adam optimizer [25]. The learning rate is set to " + }, + { + "bbox": [ + 130, + 274, + 480, + 334 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 130, + 274, + 480, + 334 + ], + "type": "text", + "content": ". For sampling, we use the DDIM sampling method [53] with 50 sampling steps." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 335, + 481, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 481, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 481, + 406 + ], + "type": "text", + "content": "Dataset. Our training dataset contains both synthetic and real data. For the synthetic data, following Zhong et al. [81], we generate 50000 triplets of data using COCO Captions [6] and Flickr30k [73] for network training. For real data, we use 200 image pairs from Nature dataset [31] and 90 pairs from Zhang et al. [76] for training and the rest of the real data are used for evaluation. We further collect 50 mixture images from the Internet for qualitative evaluation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 424, + 230, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 424, + 230, + 438 + ], + "spans": [ + { + "bbox": [ + 131, + 424, + 230, + 438 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 449, + 331, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 449, + 331, + 462 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 331, + 462 + ], + "type": "text", + "content": "4.1 Comparison with state-of-the-arts" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 469, + 480, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 469, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 469, + 480, + 588 + ], + "type": "text", + "content": "For evaluating model performances, experiments are conducted on existing real reflection datasets [31, 58, 76] and our collected mixture images from the Internet (with manually annotated language descriptions). We compare the proposed L-DiffER with state-of-the-art single-image reflection removal methods, including Zhang et al. [76], CoRRN [61], ERRNet [65], IBCLN [31], Dong et al. [8], YTMT [21], and DSRNet [22]. To compare with a representative language-based diffusion model ControlNet " + }, + { + "bbox": [ + 130, + 469, + 480, + 588 + ], + "type": "inline_equation", + "content": "[74]^3" + }, + { + "bbox": [ + 130, + 469, + 480, + 588 + ], + "type": "text", + "content": ", we make it fit our setting by using mixture images as conditions, and transmission and reflection descriptions as positive and negative prompts, respectively. For fair comparisons, we finetune the above methods on our training data if their training codes are provided." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 589, + 480, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 589, + 480, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 589, + 480, + 625 + ], + "type": "text", + "content": "Quantitative comparison. Quantitative experiments are conducted on three real datasets for reflection removal, i.e., Nature [31], Real20 [76], and SIR² [58] dataset. To evaluate the recovery of transmission layers⁴, we utilize PSNR [23]" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "L-DiffER" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 631, + 479, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 631, + 479, + 654 + ], + "spans": [ + { + "bbox": [ + 133, + 631, + 479, + 654 + ], + "type": "text", + "content": "3 Since ControlNet [74] destructs image color and structure in a generative manner as shown in Fig. 1(c), we only run it for qualitative comparisons." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 654, + 451, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 451, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 451, + 665 + ], + "type": "text", + "content": "4 Evaluations on reflection layers are provided in the supplementary material." + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 168, + 481, + 407 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 158 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 158 + ], + "type": "text", + "content": "Table 1: Comparison of quantitative results on real datasets for evaluating the recovery of transmission layers, compared with several state-of-the-art single-image reflection removal methods [8,21,22,31,61,76]. " + }, + { + "bbox": [ + 132, + 114, + 480, + 158 + ], + "type": "inline_equation", + "content": "\\uparrow (\\downarrow)" + }, + { + "bbox": [ + 132, + 114, + 480, + 158 + ], + "type": "text", + "content": " indicates larger (smaller) values are better. Bold numbers indicate the best-performing results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 168, + 481, + 407 + ], + "lines": [ + { + "bbox": [ + 132, + 168, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 481, + 407 + ], + "type": "table", + "html": "
Dataset (size)MetricMethod
Zhang et al.CoRRNERRNetIBCLNDong et al.YTMTDSRNetOurs
SIR2(500)PSNR↑22.4522.9623.1323.3623.0923.0524.9725.18
SSIM↑0.8720.8790.8780.8810.8930.8860.9070.911
LPIPS↓0.1720.1610.1640.1530.1310.1490.1240.121
NIQE↓4.9114.8264.7934.5014.7294.7384.6244.352
FID↓78.6169.5873.3360.9249.3958.6946.6044.66
Real20(20)PSNR↑22.5121.1722.0621.5921.7322.3123.4623.77
SSIM↑0.8060.7860.8030.7710.8110.8050.8060.821
LPIPS↓0.2040.2020.1850.2130.1700.1780.1650.153
NIQE↓3.8193.8913.9893.8953.9593.9774.0653.665
FID↓113.25111.2585.07120.9586.2094.6777.9272.64
Nature(20)PSNR↑20.3720.5421.1123.6923.6121.0321.7023.95
SSIM↑0.7720.7780.8060.8280.8250.8020.8200.831
LPIPS↓0.2170.2050.1830.1700.1570.1860.1680.145
NIQE↓4.4134.4334.4424.3454.5814.3874.3734.223
FID↓109.7789.1279.6875.1771.5581.7774.8468.57
Average(540)PSNR↑22.3822.8023.0223.3123.0622.9524.7925.08
SSIM↑0.8660.8720.8730.8750.8870.8800.9000.905
LPIPS↓0.1750.1640.1650.1560.1330.1510.1270.123
NIQE↓4.8524.7774.7504.4734.6954.6974.5944.322
FID↓81.0571.8574.0063.6751.5760.8848.8146.58
", + "image_path": "81a80b2c51905625dccd26efa85d07feb5843446f91bae90a1401a29c96053e9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 443, + 480, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 443, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 443, + 480, + 503 + ], + "type": "text", + "content": "and SSIM [64] as error metrics following reflection removal methods [8, 37]. We further adopts LPIPS [75], NIQE [41], FID [13] to measure the perceptual quality of the recovered results. As quantitative results shown in Table 1, the proposed method achieves the best performance among all competing methods, especially in metrics for perceptual quality, validating its effectiveness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 510, + 481, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 481, + 664 + ], + "type": "text", + "content": "Qualitative comparison. Qualitative comparisons with the aforementioned single-image reflection removal methods [8,21,22,31,61,76] and a language-based diffusion model ControlNet [74] on recovering transmission layers are shown in Fig. 5. As can be observed, single-image reflection removal methods encounter obstacles in dealing with complex reflections due to the lack of auxiliary semantic information, especially in low-transmitted reflection regions. ControlNet [74] performs significant modifications on the image color and structure in a generative manner, which deviates significantly from the input mixture image, emphasizing the pivotal role of accurate conditions in the sampling process. The proposed method generates visually pleasant reflection removal results with faithful recovery, demonstrating that it capitalizes on the synergy of the auxiliary information from language descriptions, the iterative condition refinement strategy, and the multi-condition constraint mechanism." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Y. Hong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 306, + 208 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 306, + 208 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 306, + 208 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 306, + 208 + ], + "type": "image", + "image_path": "73184fab4595eccaa6f995c8d382bb86c54b531f24f3d23e992c763458005f97.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 115, + 480, + 208 + ], + "blocks": [ + { + "bbox": [ + 307, + 115, + 480, + 208 + ], + "lines": [ + { + "bbox": [ + 307, + 115, + 480, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 115, + 480, + 208 + ], + "type": "image", + "image_path": "2e9ec1b2303725df80c73491b414317a0cb5fb966076efc690d9d5141a97cefc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 209, + 306, + 350 + ], + "blocks": [ + { + "bbox": [ + 134, + 209, + 306, + 350 + ], + "lines": [ + { + "bbox": [ + 134, + 209, + 306, + 350 + ], + "spans": [ + { + "bbox": [ + 134, + 209, + 306, + 350 + ], + "type": "image", + "image_path": "2f5b523a057505ec3b3e0e96ea137bafe12276199accfb90eb3f961c3238ddfa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 209, + 480, + 350 + ], + "blocks": [ + { + "bbox": [ + 307, + 209, + 480, + 350 + ], + "lines": [ + { + "bbox": [ + 307, + 209, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 480, + 350 + ], + "type": "image", + "image_path": "0b29f5ba22d4dd74e195cec4ab76e7ecbe7210d936be5db5c47da2da13a2cd90.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 134, + 350, + 306, + 407 + ], + "blocks": [ + { + "bbox": [ + 134, + 350, + 306, + 407 + ], + "lines": [ + { + "bbox": [ + 134, + 350, + 306, + 407 + ], + "spans": [ + { + "bbox": [ + 134, + 350, + 306, + 407 + ], + "type": "image", + "image_path": "63c56fa02617f6d90b54a50fd5b7bd660d4cb724c5855eaf87b8f3dd03a53098.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 418, + 482, + 452 + ], + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 452 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparison of estimated transmission layers on real mixture images collected from the Internet, compared with several single-image methods [8,21,22,31, 61,76] and a diffusion-based method ControlNet [74]. Please zoom in for details." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 350, + 480, + 407 + ], + "blocks": [ + { + "bbox": [ + 307, + 350, + 480, + 407 + ], + "lines": [ + { + "bbox": [ + 307, + 350, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 480, + 407 + ], + "type": "image", + "image_path": "e4f183a36e8483299ec0751ecc35af2c6ddfc95f1e0b14567a8477f2b1a3809d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 476, + 236, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 236, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 236, + 489 + ], + "type": "text", + "content": "4.2 Ablation study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 500, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 482, + 643 + ], + "type": "text", + "content": "Ablation on control conditions5. We conduct ablation studies to investigate the effectiveness of the iterative condition refinement strategy and the multi-condition constraint mechanism. As shown in Table 2 and Fig. 6, most reflection contaminations are retained when missing iterative condition refinement (W/o rfn.), and disabling the refinement on color (W/o clr. rfn.) or structure (W/o str. rfn.) will cause inaccurate recovery, which exists color shift or retains the reflection content. In addition, as shown in Table 2 and Fig. 7, due to the generative prior of the diffusion model, the absence of the max value constraint (W/o MVC) or the saturation-aware structure constraint (W/o SSC), or both (W/o cnst.) leads to obvious color shifts, structure distortion, and uncontrolled recovery. Therefore, the degradation observed in the above ablations indicates the necessity of the refinement and constraints on conditions." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "L-DiffER" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 414, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 414, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 414, + 666 + ], + "type": "text", + "content": "5 More ablation studies are provided in the supplementary material." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 157, + 481, + 239 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Ablation studies on the iterative condition refinement strategy and the multi-condition constraint mechanism. " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\uparrow (\\downarrow)" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " indicates larger (smaller) values are better. Bold numbers indicate the best-performing results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 157, + 481, + 239 + ], + "lines": [ + { + "bbox": [ + 132, + 157, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 481, + 239 + ], + "type": "table", + "html": "
MetricAbl. on iterative condition refinementAbl. on multi-condition constraint
W/o clr. rfn.W/o str. rfn.W/o rfn.W/o MVCW/o SSCW/o cnst.Ours
PSNR↑24.4224.6824.3924.7324.1623.3525.08
SSIM↑0.8980.8960.8910.8970.8870.8790.905
LPIPS↓0.1320.1360.1390.1290.1430.1540.123
NIQE↓4.5014.6454.8684.5534.7094.5614.322
FID↓49.1353.7160.1251.0656.4663.6846.58
", + "image_path": "0a31a070e5b7c8d0797db74623108abe61c38538be89b324e6cae0bf3a8969cb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 251, + 480, + 312 + ], + "blocks": [ + { + "bbox": [ + 134, + 251, + 480, + 312 + ], + "lines": [ + { + "bbox": [ + 134, + 251, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 134, + 251, + 480, + 312 + ], + "type": "image", + "image_path": "9f86f49c4ca7b66023cde984905d025df4a3b69d9fefa01730740d5ec21415ff.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 320, + 482, + 354 + ], + "lines": [ + { + "bbox": [ + 132, + 320, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 482, + 354 + ], + "type": "text", + "content": "Fig. 6: The effect of the iterative condition refinement strategy. For each result, we place the corresponding final color and structure condition at its lower left and lower right. Please zoom in for details, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 135, + 367, + 480, + 427 + ], + "blocks": [ + { + "bbox": [ + 135, + 367, + 480, + 427 + ], + "lines": [ + { + "bbox": [ + 135, + 367, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 135, + 367, + 480, + 427 + ], + "type": "image", + "image_path": "0e2c9733bd9676f7a4b1d69337c22d0ab1bbc1100f7b21840669a062f2828af0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 435, + 482, + 468 + ], + "lines": [ + { + "bbox": [ + 132, + 435, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 482, + 468 + ], + "type": "text", + "content": "Fig. 7: The effect of the color and structure constraints in the multi-condition constraint mechanism. For each result, we place the corresponding final color and structure condition at its lower left and lower right, respectively. Please zoom in for details." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 492, + 220, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 220, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 220, + 504 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "text", + "content": "This paper introduces the first language-based diffusion model for single image reflection removal, which exploits auxiliary semantic information from language descriptions and generative priors from diffusion models to recover transmission layers from mixture images with low-transmitted reflection regions. An iterative condition refinement strategy is proposed to address the problem of inaccurate control conditions, and a multi-condition constraint mechanism is introduced to ensure recovery faithfulness. Experiments on real data demonstrate the effectiveness of the proposed method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Limitations. The proposed method may incorrectly identify transmission and reflection layers if they exhibit similar image content. For such cases, more accurate spatial guidance from users such as annotated edges or region masks could help to relieve the ambiguity, which is left as our future work." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Y. Hong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 479, + 163 + ], + "type": "text", + "content": "This work is supported by National Natural Science Foundation of China under Grant No. 62136001, 62088102." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 206, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 206, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 228 + ], + "type": "text", + "content": "1. Chang, Y., Jung, C., Sun, J.: Joint reflection removal and depth estimation from a single image. IEEE Transactions on Cybernetics (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "type": "text", + "content": "2. Chang, Y., Jung, C., Sun, J., Wang, F.: Siamese dense network for reflection removal with flash and no-flash image pairs. International Journal of Computer Vision (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 480, + 293 + ], + "type": "text", + "content": "3. Chang, Z., Weng, S., Li, Y., Li, S., Shi, B.: L-CoDer: Language-based colorization with color-object decoupling transformer. In: Proc. of European Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 294, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 294, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 294, + 480, + 326 + ], + "type": "text", + "content": "4. Chang, Z., Weng, S., Zhang, P., Li, Y., Li, S., Shi, B.: L-CAD: Language-based colorization with any-level descriptions using diffusion priors. In: Proc. of Advances in Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 327, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 480, + 358 + ], + "type": "text", + "content": "5. Chang, Z., Weng, S., Zhang, P., Li, Y., Li, S., Shi, B.: L-CoIns: Language-based colorization with instance awareness. In: Proc. of Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 360, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 360, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 360, + 480, + 392 + ], + "type": "text", + "content": "6. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dolkar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 392, + 480, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 414 + ], + "type": "text", + "content": "7. Diamant, Y., Schechner, Y.Y.: Overcoming visual reverberations. In: Proc. of Computer Vision and Pattern Recognition (2008)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "type": "text", + "content": "8. Dong, Z., Xu, K., Yang, Y., Bao, H., Xu, W., Lau, R.W.: Location-aware single image reflection removal. In: Proc. of International Conference on Computer Vision (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "text", + "content": "9. Fan, Q., Yang, J., Hua, G., Chen, B., Wipf, D.: A generic deep architecture for single image reflection removal and image smoothing. In: Proc. of International Conference on Computer Vision (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 480, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 480, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 133, + 480, + 480, + 502 + ], + "type": "text", + "content": "10. Han, B.J., Sim, J.Y.: Zero-shot learning for reflection removal of single 360-degree image. In: Proc. of European Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 133, + 502, + 480, + 533 + ], + "type": "text", + "content": "11. He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In: Proc. of International Conference on Computer Vision (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 535, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 535, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 133, + 535, + 480, + 567 + ], + "type": "text", + "content": "12. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 133, + 567, + 480, + 599 + ], + "type": "text", + "content": "13. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Proc. of Advances in Neural Information Processing Systems (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 133, + 600, + 480, + 621 + ], + "type": "text", + "content": "14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Proc. of Advances in Neural Information Processing Systems (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 622, + 480, + 643 + ], + "type": "text", + "content": "15. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "text", + "content": "16. Hong, Y., Chang, Y., Liang, J., Ma, L., Huang, T., Shi, B.: Light flickering guided reflection removal. International Journal of Computer Vision (2024)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "L-DiffER" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "17. Hong, Y., Lyu, Y., Li, S., Cao, G., Shi, B.: Reflection removal with NIR and RGB image feature fusion. IEEE Transactions on Multimedia (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 480, + 160 + ], + "type": "text", + "content": "18. Hong, Y., Lyu, Y., Li, S., Shi, B.: Near-infrared image guided reflection removal. In: Proc. of International Conference on Multimedia and Expo (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 160, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 480, + 182 + ], + "type": "text", + "content": "19. Hong, Y., Zheng, Q., Zhao, L., Jiang, X., Kot, A.C., Shi, B.: Panoramic image reflection removal. In: Proc. of Computer Vision and Pattern Recognition (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 213 + ], + "type": "text", + "content": "20. Hong, Y., Zheng, Q., Zhao, L., Jiang, X., Kot, A.C., Shi, B.: PAR" + }, + { + "bbox": [ + 132, + 182, + 480, + 213 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 132, + 182, + 480, + 213 + ], + "type": "text", + "content": "Net: End-to-end panoramic image reflection removal. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 213, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 480, + 246 + ], + "type": "text", + "content": "21. Hu, Q., Guo, X.:trash or treasure? an interactive dual-stream strategy for single image reflection separation. Proc. of Advances in Neural Information Processing Systems (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "text", + "content": "22. Hu, Q., Guo, X.: Single image reflection separation via component synergy. In: Proc. of International Conference on Computer Vision (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "type": "text", + "content": "23. Huynh-Thu, Q., Ghanbari, M.: Scope of validity of psnr in image/video quality assessment. Electronics letters (2008)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "type": "text", + "content": "24. Kim, S., Huo, Y., Yoon, S.E.: Single image reflection removal with physically-based training images. In: Proc. of Computer Vision and Pattern Recognition (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 311, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 332 + ], + "type": "text", + "content": "25. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 332, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 353 + ], + "type": "text", + "content": "26. Kong, N., Tai, Y.W., Shin, S.Y.: A physically-based approach to reflection separation. In: Proc. of Computer Vision and Pattern Recognition (2012)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 353, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 480, + 374 + ], + "type": "text", + "content": "27. Lei, C., Chen, Q.: Robust reflection removal with reflection-free flash-only cues. In: Proc. of Computer Vision and Pattern Recognition (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "text", + "content": "28. Lei, C., Huang, X., Zhang, M., Yan, Q., Sun, W., Chen, Q.: Polarized reflection removal with perfect alignment in the wild. In: Proc. of Computer Vision and Pattern Recognition (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "type": "text", + "content": "29. Lei, C., Jiang, X., Chen, Q.: Robust reflection removal with flash-only cues in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "type": "text", + "content": "30. Levin, A., Weiss, Y.: User assisted separation of reflections from a single image using a sparsity prior. IEEE Transactions on Pattern Analysis and Machine Intelligence 29(9), 1647-1654 (2007)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 460, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 480, + 492 + ], + "type": "text", + "content": "31. Li, C., Yang, Y., He, K., Lin, S., Hopcroft, J.E.: Single image reflection removal through cascaded refinement. In: Proc. of Computer Vision and Pattern Recognition (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 492, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 480, + 514 + ], + "type": "text", + "content": "32. Li, Y., Brown, M.S.: Exploiting reflection change for automatic reflection removal. In: Proc. of International Conference on Computer Vision (2013)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 514, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 535 + ], + "type": "text", + "content": "33. Li, Y., Brown, M.S.: Single image layer separation using relative smoothness. In: Proc. of Computer Vision and Pattern Recognition (2014)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 535, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 535, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 535, + 480, + 556 + ], + "type": "text", + "content": "34. Liu, Y.L., Lai, W.S., Yang, M.H., Chuang, Y.Y., Huang, J.B.: Learning to see through obstructions. In: Proc. of Computer Vision and Pattern Recognition (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 556, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 480, + 589 + ], + "type": "text", + "content": "35. Liu, Y.L., Lai, W.S., Yang, M.H., Chuang, Y.Y., Huang, J.B.: Learning to see through obstructions with layered decomposition. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 589, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 633 + ], + "type": "text", + "content": "36. Luo, J., Fu, J., Kong, X., Gao, C., Ren, H., Shen, H., Xia, H., Liu, S.: 3D-SPS: Single-stage 3D visual grounding via referred point progressive selection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16454-16463 (June 2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "37. Lyu, Y., Cui, Z., Li, S., Pollefeys, M., Shi, B.: Reflection separation using a pair of unpolarized and polarized images. In: Proc. of Advances in Neural Information Processing Systems (2019)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Y. Hong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "text", + "content": "38. Lyu, Y., Cui, Z., Li, S., Pollefeys, M., Shi, B.: Physics-guided reflection separation from a pair of unpolarized and polarized images. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "text", + "content": "39. Ma, D., Wan, R., Shi, B., Kot, A.C., Duan, L.Y.: Learning to jointly generate and separate reflections. In: Proc. of International Conference on Computer Vision (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "text", + "content": "40. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: Proc. of International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "text", + "content": "41. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters 20(3), 209-212 (2013)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "text", + "content": "42. Mou, C., Wang, X., Xie, L., Wu, Y., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 481, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 307 + ], + "type": "text", + "content": "43. Nam, S., Brubaker, M.A., Brown, M.S.: Neural image representations for multi-image fusion and layer separation. In: Proc. of European Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 329 + ], + "type": "text", + "content": "44. Nayar, S.K., Fang, X.S., Boult, T.: Separation of reflection components using color and polarization. International Journal of Computer Vision (1997)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 331, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 481, + 363 + ], + "type": "text", + "content": "45. Park, J., Kim, H., Park, E., Sim, J.Y.: Fully-automatic reflection removal for 360-degree images. In: Proc. of Winter Conference on Applications of Computer Vision (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 363, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 481, + 407 + ], + "type": "text", + "content": "46. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: Proc. of Advances in Neural Information Processing Systems (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 441 + ], + "type": "text", + "content": "47. Qiu, J., Jiang, P.T., Zhu, Y., Yin, Z.X., Cheng, M.M., Ren, B.: Looking through the glass: Neural surface reconstruction against high specular reflections. In: Proc. of Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 441, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 481, + 485 + ], + "type": "text", + "content": "48. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proc. of International Conference on Machine Learning. PMLR (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 486, + 481, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 481, + 518 + ], + "type": "text", + "content": "49. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proc. of Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 519, + 481, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 519, + 481, + 541 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 481, + 541 + ], + "type": "text", + "content": "50. Schechner, Y.Y., Kiryati, N., Basri, R.: Separation of transparent layers using focus. International Journal of Computer Vision (2000)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 542, + 481, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 542, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 542, + 481, + 563 + ], + "type": "text", + "content": "51. Shih, Y., Krishnan, D., Durand, F., Freeman, W.T.: Reflection removal using ghosting cues. In: Proc. of Computer Vision and Pattern Recognition (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 564, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 564, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 564, + 481, + 586 + ], + "type": "text", + "content": "52. Simon, C., Kyu Park, I.: Reflection removal for in-vehicle black box videos. In: Proc. of Computer Vision and Pattern Recognition (2015)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "type": "text", + "content": "53. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 609, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 481, + 642 + ], + "type": "text", + "content": "54. Sun, H., Li, W., Liu, J., Chen, H., Pei, R., Zou, X., Yan, Y., Yang, Y.: CoSeR: Bridging image and language for cognitive super-resolution. arXiv preprint arXiv:2311.16512 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "55. Sun, J., Weng, S., Chang, Z., Li, S., Shi, B.: UniCoRN: A unified conditional image repainting network. In: Proc. of Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "L-DiffER" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "56. Tang, J., Zhong, H., Weng, S., Shi, B.: LuminAIRe: Illumination-aware conditional image painting for lighting-realistic generation. In: Proc. of Advances in Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "text", + "content": "57. Wan, R., Shi, B., Duan, L.Y., Tan, A.H., Gao, W., Kot, A.C.: Region-aware reflection removal with unified content and gradient priors. IEEE Transactions on Image Processing (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "type": "text", + "content": "58. Wan, R., Shi, B., Duan, L.Y., Tan, A.H., Kot, A.C.: Benchmarking single-image reflection removal algorithms. In: Proc. of International Conference on Computer Vision (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 247 + ], + "type": "text", + "content": "59. Wan, R., Shi, B., Duan, L.Y., Tan, A.H., Kot, A.C.: CRRN: Multi-scale guided concurrent reflection removal network. In: Proc. of Computer Vision and Pattern Recognition (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 247, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 480, + 267 + ], + "type": "text", + "content": "60. Wan, R., Shi, B., Li, H., Duan, L.Y., Kot, A.C.: Face image reflection removal. International Journal of Computer Vision (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "text", + "content": "61. Wan, R., Shi, B., Li, H., Duan, L.Y., Tan, A.H., Kot, A.C.: CoRRN: Cooperative reflection removal network. IEEE Transactions on Pattern Analysis and Machine Intelligence (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "text", + "content": "62. Wan, R., Shi, B., Li, H., Hong, Y., Duan, L.Y., Kot, A.C.: Benchmarking single-image reflection removal algorithms. In: IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "type": "text", + "content": "63. Wang, Z., Lu, Y., Li, Q., Tao, X., Guo, Y., Gong, M., Liu, T.: CRIS: CLIP-driven referring image segmentation. In: Proc. of Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 397 + ], + "type": "text", + "content": "64. Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers (2003)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 397, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 429 + ], + "type": "text", + "content": "65. Wei, K., Yang, J., Fu, Y., Wipf, D., Huang, H.: Single image reflection removal exploiting misaligned training data and network enhancements. In: Proc. of Computer Vision and Pattern Recognition (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 429, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 450 + ], + "type": "text", + "content": "66. Wen, Q., Tan, Y., Qin, J., Liu, W., Han, G., He, S.: Single image reflection removal beyond linearity. In: Proc. of Computer Vision and Pattern Recognition (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "type": "text", + "content": "67. Weng, S., Li, W., Li, D., Jin, H., Shi, B.: MISC: Multi-condition injection and spatially-adaptive compositing for conditional person image synthesis. In: Proc. of Computer Vision and Pattern Recognition (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "text", + "content": "68. Weng, S., Shi, B.: Conditional image inpainting. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "69. Weng, S., Wu, H., Chang, Z., Tang, J., Li, S., Shi, B.: L-CoDe: Language-based colorization using color-object decoupled conditions. In: Proc. of the AAAI Conference on Artificial Intelligence (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "text", + "content": "70. Yang, J., Gong, D., Liu, L., Shi, Q.: Seeing deeply and bidirectionally: A deep learning approach for single image reflection removal. In: Proc. of European Conference on Computer Vision (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "text", + "content": "71. Yang, Y., Ma, W., Zheng, Y., Cai, J.F., Xu, W.: Fast single image reflection suppression via convex optimization. In: Proc. of Computer Vision and Pattern Recognition (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "text", + "content": "72. Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: LAVT: Language-aware vision transformer for referring image segmentation. In: Proc. of Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 633, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 666 + ], + "type": "text", + "content": "73. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics (2014)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Y. Hong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 348 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "74. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proc. of International Conference on Computer Vision (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "text", + "content": "75. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proc. of Computer Vision and Pattern Recognition (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 171, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 171, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 482, + 194 + ], + "type": "text", + "content": "76. Zhang, X., Ng, R., Chen, Q.: Single image reflection separation with perceptual losses. In: Proc. of Computer Vision and Pattern Recognition (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "text", + "content": "77. Zhang, Y.N., Shen, L., Li, Q.: Content and gradient model-driven deep network for single image reflection removal. In: Proc. of ACM International Conference on Multimedia (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "text", + "content": "78. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Unictrlnet: All-in-one control to text-to-image diffusion models. Proc. of Advances in Neural Information Processing Systems (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "text", + "content": "79. Zheng, Q., Chen, J., Lu, Z., Shi, B., Jiang, X., Yap, K.H., Duan, L.Y., Kot, A.C.: What does plate glass reveal about camera calibration? In: Proc. of Computer Vision and Pattern Recognition (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "type": "text", + "content": "80. Zheng, Q., Shi, B., Chen, J., Jiang, X., Duan, L.Y., Kot, A.C.: Single image reflection removal with absorption effect. In: Proc. of Computer Vision and Pattern Recognition (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 325, + 482, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 348 + ], + "type": "text", + "content": "81. Zhong, H., Hong, Y., Weng, S., Liang, J., Shi, B.: Language-guided image reflection separation. In: Proc. of Computer Vision and Pattern Recognition (2024)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "L-DiffER" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LAPT_ Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models/64854d06-c2d0-416b-8dda-0ced42b5873e_content_list.json b/2024/LAPT_ Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models/64854d06-c2d0-416b-8dda-0ced42b5873e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b0f9cdf12f6099b3190ec264e042db8b4e2220ff --- /dev/null +++ b/2024/LAPT_ Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models/64854d06-c2d0-416b-8dda-0ced42b5873e_content_list.json @@ -0,0 +1,1891 @@ +[ + { + "type": "text", + "text": "LAPT: Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yabin Zhang $^{1,2}$ , Wenjie Zhu $^{1}$ , Chenhang He $^{1}$ , and Lei Zhang $^{1,2\\star}$", + "bbox": [ + 267, + 210, + 732, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The Hong Kong Polytechnic University", + "bbox": [ + 361, + 238, + 640, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 OPPO Research Institute", + "bbox": [ + 406, + 253, + 594, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{csybzhang,cslzhang}@comp.polyu.edu.hk", + "bbox": [ + 361, + 267, + 640, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Out-of-distribution (OOD) detection is crucial for model reliability, as it identifies samples from unknown classes and reduces errors due to unexpected inputs. Vision-Language Models (VLMs) such as CLIP are emerging as powerful tools for OOD detection by integrating multi-modal information. However, the practical application of such systems is challenged by manual prompt engineering, which demands domain expertise and is sensitive to linguistic nuances. In this paper, we introduce Label-driven Automated Prompt Tuning (LAPT), a novel approach to OOD detection that reduces the need for manual prompt engineering. We develop distribution-aware prompts with in-distribution (ID) class names and negative labels mined automatically. Training samples linked to these class labels are collected autonomously via image synthesis and retrieval methods, allowing for prompt learning without manual effort. We utilize a simple cross-entropy loss for prompt optimization, with cross-modal and cross-distribution mixing strategies to reduce image noise and explore the intermediate space between distributions, respectively. The LAPT framework operates autonomously, requiring only ID class names as input and eliminating the need for manual intervention. With extensive experiments, LAPT consistently outperforms manually crafted prompts, setting a new standard for OOD detection. Moreover, LAPT not only enhances the distinction between ID and OOD samples, but also improves the ID classification accuracy and strengthens the generalization robustness to covariate shifts, resulting in outstanding performance in challenging full-spectrum OOD detection tasks. Codes are available at https://github.com/YBZh/LAPT.", + "bbox": [ + 259, + 321, + 743, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Out-of-distribution detection $\\cdot$ Vision-language models $\\cdot$ Automated prompt tuning $\\cdot$ Label-driven learning", + "bbox": [ + 259, + 681, + 743, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 737, + 375, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In real-world applications, artificial intelligence (AI) systems often encounter data of unknown classes, known as out-of-distribution (OOD) data. When tackling such OOD data, AI models may erroneously remain overconfident in their", + "bbox": [ + 212, + 768, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 217, + 824, + 385, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0efbef56b24ae67870d86badb2c358da0c20e7eb962992ce519e2907ec4db2df.jpg", + "image_caption": [ + "Fig. 1: VLMs-based OOD detection systems are sensitive to linguistic nuances of text prompts, where results are reported on the OpenOOD dataset with a ViTB/16 image encoder. A lower FPR95 metric denotes better OOD detection performance." + ], + "image_footnote": [], + "bbox": [ + 282, + 147, + 720, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "predictions [33, 41], resulting in critical errors and posing security risks. Consequently, it is crucial to recognize OOD data for maintaining the reliability and safety of AI systems in open-world scenarios.", + "bbox": [ + 212, + 401, + 782, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional OOD detection methods are primarily focused on image information [15,24,25], neglecting the rich textual knowledge carried by class names. With the rise of vision-language models (VLMs) [20,36], there has been a growing interest in leveraging textual information to facilitate visual tasks [67,70,72], including OOD detection [21,29,30]. For example, Ming et al. [29] compared test images with linguistic descriptions of in-distribution (ID) classes in feature space, leveraging the maximum probability of a scaled softmax function to identify OOD images. While this method achieves notable generalization across various datasets, it engages only a restricted portion of the textual space, thereby not fully exploiting the textual interpretative capabilities of VLMs. Jiang et al. [21] extended it by exploiting larger textual spaces, extracting negative labels from vast text corpora to enhance OOD detection performance significantly. However, the deployment of this system is hindered by the demanding task of prompt engineering, which requires domain-specific knowledge and is sensitive to fine-grained linguistic variations, as illustrated in Fig. 1.", + "bbox": [ + 212, + 446, + 787, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose Label-driven Automated Prompt Tuning (LAPT), a novel approach to OOD detection that substantially reduces the need of manual prompt engineering. Specifically, given ID class names and automatically extracted negative labels as in [21], we append distribution-aware prompt tokens to these class labels. These learnable tokens are optimized with minimal human intervention by aggregating training images associated with class labels using pre-trained text-to-image generation models [35, 39] or retrieving real images from web-scale dataset [42, 43]. We employ a simple yet effective cross-entropy loss for prompt optimization, enhanced by cross-modal and cross-distribution data mixing techniques. Cross-modal mixing merges image and text features from the same class, mitigating image noise through multi-modal integration.", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Cross-distribution mixing stochastically blends ID and negative features along with their respective labels, exploring the intermediate space between ID and negative domains. During testing, we adopt the OOD score introduced in [21] to utilize both ID and negative labels.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extensive experiments demonstrate that our method outperforms the manually engineered text prompts by a large margin and achieves new state-of-the-art on various OOD detection benchmarks. Especially, on the challenging near-OD detection task using the ImageNet-1k as the ID dataset, our approach exceeds the best hand-crafted prompts by $10.76\\%$ in FPR95 and $5.71\\%$ in AUROC, without any manual label annotations or prompt crafting. Moreover, LAPT not only enhances the distinction between ID and OOD samples, but also improves the ID classification accuracy and strengthens the generalization robustness to covariate shifts. These improvements collectively contribute to the superior performance of LAPT in challenging full-spectrum OOD detection tasks [59, 60]. We summarize our contributions as follows:", + "bbox": [ + 212, + 207, + 784, + 371 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To circumvent the labor-intensive task of manual prompt engineering for OOD detection with VLMs, we introduce the LAPT methodology, which autonomously learns effective prompts using only ID class names.", + "- Utilizing the ID class names and automatically mined negative labels, we devise distribution-aware prompts and autonomously gather the corresponding training images for each class via image synthesis and retrieval.", + "- With the collected training data, we employ a simple yet effective cross-entropy objective for prompt optimization, enhanced by cross-modal data mixing to reduce image noise and cross-distribution data mixing to explore the intermediate space between ID and negative regions.", + "- Extensive experiments demonstrate that our method significantly outperforms the hand-crafted text prompts and achieves new state-of-the-art results across multiple benchmarks. Moreover, LAPT enhances not only the distinction between ID and OOD samples, but also the ID classification accuracy and the generalization robustness to covariate shifts, achieving outstanding results in full-spectrum OOD detection tasks." + ], + "bbox": [ + 225, + 381, + 784, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 643, + 387, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traditional OOD detection aims to address the severe issue of deep learning models in producing overconfident predictions on OOD data [33, 59], which poses significant challenges to the deployment of deep models in real-world applications. To tackle this problem, researchers have developed various OOD detection methods, including score-based [15, 19, 24-26, 45, 52, 54, 55], distance-based [6, 31, 44, 46-48] and generative-based [22, 40] approaches. Score-based methods, which arguably garner the most attentions, distinguish ID from OOD samples using carefully designed scoring functions, such as confidence-based [15, 25, 45, 52, 55], discriminator-based [22], energy-based [26, 54], and gradient-based [18] scores. In contrast, distance-based methods detect OOD samples by measuring the distance in feature space between the test data and their nearest", + "bbox": [ + 212, + 672, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "LAPT for OOD Detection with Vision-Language Models", + "bbox": [ + 352, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ID samples [47] or ID prototypes [48] using metrics like KNN [6,31,46] or Mahalanobis distance [24,44]. While these methods have achieved certain success, they overlook textual information and require training or pre-training on manually labeled ID training data. Moreover, some approaches require manually collecting OOD data for additional training [10,16,63], hindering their real-world applicability. Our method differs from these approaches by effectively leveraging textual information. Utilizing only the ID class names, we mine negative labels, gather training data, and learn effective prompts in an automated fashion, achieving state-of-the-art OOD detection results with minimal human effort.", + "bbox": [ + 212, + 145, + 787, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "OOD detection with VLMs is gaining growing attention as the importance of textual information for OOD image detection becomes increasingly recognized [9, 21, 29, 30, 32, 34, 53]. MCM [29] leverages ID class names for effective zero-shot OOD detection, which is further enhanced by NegLabel [21] through the additional use of mined OOD class names. However, as shown in Fig. 1, NegLabel's performance is sensitive to text prompts, limiting its practical application. ZOC [9] employs VLMs to detect OOD samples by learning a captioner that generates candidate OOD labels. Yet, this captioner struggles to generate effective OOD labels for large-scale ID datasets with numerous ID classes. LoCoOp [32] learns ID prompts using few-shot ID samples, regularizing these ID prompts by mined OOD features from image backgrounds. CLIPN [53] and LSN [34] learn text prompts for OOD classes. Our approach differs significantly from them in two aspects. First, we initialize negative prompts with negative labels instead of initializing negative prompts with 'no' and ID labels. Considering CLIP's difficulty in understanding the meaning of 'no' [34], using OOD labels to initialize negative prompts provides an effective starting point to accelerate convergence. Second, CLIPN and LSN respectively require large-scale multi-modal data and manually collected ID samples, leading to either high training costs or the need of manual annotation. In contrast, our method automatically gathers a small amount of positive/negative training data with ID labels, requiring minimal manual effort yet achieving improved performance.", + "bbox": [ + 212, + 282, + 787, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methods", + "text_level": 1, + "bbox": [ + 215, + 619, + 339, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Problem Setup", + "text_level": 1, + "bbox": [ + 215, + 650, + 388, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $\\mathcal{X}$ and $\\mathcal{Y} = \\{y_1, \\ldots, y_C\\}$ be the image space and ID label space, where $\\mathcal{Y}$ is a set of class names, e.g., $\\mathcal{Y} = \\{cat, dog, \\ldots, bird\\}$ and $C$ is the number of ID classes. Given the ID random variable $\\pmb{x}^{in} \\in \\mathcal{X}$ and the OOD random variable $\\pmb{x}^{odd} \\in \\mathcal{X}$ , we use $\\mathcal{P}_{\\pmb{x}^{in}}$ and $\\mathcal{P}_{\\pmb{x}^{odd}}$ to denote the ID marginal distribution and OOD marginal distribution, respectively.", + "bbox": [ + 212, + 672, + 787, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "OOD detection. In traditional classification tasks, we assume that the test image $\\pmb{x}$ is sampled from the ID distribution and belongs to a specific ID class, i.e., $\\pmb{x} \\in \\mathcal{P}_{\\pmb{x}^{in}}$ and $y \\in \\mathcal{V}$ , where $y$ is the label of $\\pmb{x}$ . However, in real-world applications, AI systems often encounter samples from unknown classes, i.e., $\\pmb{x} \\in \\mathcal{P}_{\\pmb{x}^{ood}}$ and $y \\notin \\mathcal{V}$ . AI models may misclassify these data into known ID classes with high confidence [33,41], posing critical errors and security risks. To tackle", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "this problem, OOD detection is introduced to classify ID samples into correct ID classes and reject OOD samples simultaneously. The ID class classification is achieved with a $C$ -way classifier following traditional classification [12, 23], while the OOD detection is typically achieved with a score function $S$ [24-26]:", + "bbox": [ + 214, + 146, + 782, + 205 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG _ {\\gamma} (\\boldsymbol {x}) = \\operatorname {I D}, \\text {i f} S (\\boldsymbol {x}) \\geq \\gamma ; \\quad \\text {o t h e r w i s e}, G _ {\\gamma} (\\boldsymbol {x}) = \\mathrm {O O D}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 217, + 784, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $G_{\\gamma}$ is the OOD detector with the threshold $\\gamma$ . In other words, the test data $\\mathbf{x}$ is detected as ID samples if and only if $S(\\mathbf{x}) \\geq \\gamma$ .", + "bbox": [ + 214, + 241, + 782, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Reviews on MCM and NegLabel", + "text_level": 1, + "bbox": [ + 214, + 292, + 532, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CLIP-like models. Given the ID test image $\\pmb{x}$ within the label space $\\mathcal{V}$ , we extract the image representation $\\pmb{v} = f_{img}(\\pmb{x}) \\in \\mathcal{R}^D$ and the textual representation $C = f_{txt}(\\rho(\\mathcal{Y})) \\in \\mathcal{R}^{C \\times D}$ with the pre-trained CLIP-like encoders, where $D$ denotes the feature dimension. $f_{img}(\\cdot)$ and $f_{txt}(\\cdot)$ denote the image and text encoders, respectively. $\\rho(\\cdot)$ denotes the text prompt function, which is typically manually defined as 'a photo of a $<\\text{label}>$ .', where the 'label' token is replaced by the specific class name, e.g., 'cat' or 'dog'. Both $\\pmb{v}$ and $C$ are $L_2$ normalized along the $D$ dimension. Then the zero-shot classification probability can be achieved using $C$ as the classifier:", + "bbox": [ + 214, + 314, + 784, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} = \\operatorname {S o f t m a x} \\left(\\boldsymbol {v} \\boldsymbol {C} ^ {T}\\right) \\in \\mathcal {R} ^ {C}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 459, + 784, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the fixed scaling parameter is omitted for simplicity.", + "bbox": [ + 214, + 484, + 640, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MCM. Besides the remarkable classification capabilities, MCM [29] reveals that pre-trained CLIP models also exhibit strong zero-shot OOD detection abilities. Specifically, MCM treats textual representations $C$ as ID prototypes and assesses OOD uncertainty based on the scaled cosine distance between the visual features and the nearest ID prototype. Furthermore, it has been validated that applying a softmax function as a post hoc mechanism enlarges the distinction between ID and OOD samples, leading to the following score function:", + "bbox": [ + 214, + 500, + 782, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {m c m} (\\boldsymbol {v}) = \\max _ {i} \\frac {e ^ {\\cos (\\boldsymbol {v} , \\boldsymbol {c} _ {i}) / \\tau}}{\\sum_ {i = 1} ^ {C} e ^ {\\cos (\\boldsymbol {v} , \\boldsymbol {c} _ {i}) / \\tau}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 614, + 784, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\cos (\\cdot ,\\cdot)$ measures the cosine similarity, $c_{i}$ is the $i$ -th entry of $C$ , and $\\tau >0$ is the scaling temperature.", + "bbox": [ + 214, + 659, + 782, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NegLabel. While it is validated in [29] that leveraging the textual information of ID classes enhances the OOD detection capability, the potential contribution of negative labels is overlooked. To fill this gap, NegLabel [21] explores negative class names $\\mathcal{Y}^{-} = \\{y_{C + 1},\\dots ,y_{C + M}\\}$ from extensive text corpora, where $\\mathcal{Y}^{-}\\cap \\mathcal{Y} = \\emptyset$ . Under the assumption that ID samples are more similar to ID labels and less similar to negative labels than OOD samples, NegLabel employs a scoring function to improve OOD detection:", + "bbox": [ + 214, + 689, + 784, + 794 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {N e g L a b e l} (\\boldsymbol {v}) = \\frac {\\sum_ {i = 1} ^ {C} e ^ {\\cos (\\boldsymbol {v} , \\boldsymbol {c} _ {i}) / \\tau}}{\\sum_ {i = 1} ^ {C} e ^ {\\cos (\\boldsymbol {v} , \\boldsymbol {c} _ {i}) / \\tau} + \\sum_ {j = C + 1} ^ {C + M} e ^ {\\cos (\\boldsymbol {v} , \\boldsymbol {c} _ {j}) / \\tau}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 804, + 784, + 844 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "LAPT for OOD Detection with Vision-Language Models", + "bbox": [ + 352, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/91b5bd354cf89231b19e97f1b85c84e46e1ac22abe9fe8ddbfc2bb010328bfef.jpg", + "image_caption": [ + "Fig. 2: The overall framework of our LAPT method, where $\\mathbf{v}_{dog} / \\mathbf{v}_{boat}$ , $\\mathbf{c}_{dog} / \\mathbf{c}_{boat}$ , and $l_{dog} / l_{boat}$ are image features, textual features, and soft labels of dog/boat samples." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 589, + 334 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bc3f554681f97bdc63a23f7034badf751fbeb1c41f1a288e910402e571b4f914.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 142, + 787, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "NegLabel performs excellently on various OOD detection benchmarks by utilizing both ID and negative text knowledge.", + "bbox": [ + 212, + 407, + 784, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 LAPT: Label-driven Automated Prompt Tuning", + "text_level": 1, + "bbox": [ + 212, + 465, + 661, + 482 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Though NegLabel has achieved a notable success, its performance is highly sensitive to the configuration of text prompts. As illustrated in Fig. 1, given an ID dataset like ImageNet, varying text prompts leads to considerable differences in OOD detection outcomes. For instance, prompts such as 'The nice $<$ label> $^\\prime$ generally improve performance on far-OOD datasets, while 'a photo of a $<$ label> $^\\prime$ shows enhanced effectiveness for near-OOD scenarios. The choice of prompt can lead to over $10\\%$ fluctuation in the FPR95 metric. Hence, manually setting text prompts for optimal results is a labor-intensive and time-consuming process.", + "bbox": [ + 212, + 494, + 787, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We aim to streamline OOD detection by automating prompt generation with minimal manual intervention. Our method learns continuous prompts using autocollected samples, needing only ID class names. The automated process is very useful in practical applications. Details of LAPT are provided as follows.", + "bbox": [ + 212, + 617, + 787, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Distribution-aware prompts. Given the ID class names, we first follow NegLabel [21] to mine the negative labels $\\mathcal{V}^{-}$ from extensive corpus databases. Then, instead of hand-crafting the optimal text prompts, we design a learnable text prompt function in the following form:", + "bbox": [ + 212, + 704, + 807, + 766 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {l} (\\mathrm {L A B E L}) = [ \\boldsymbol {V} ] _ {1} [ \\boldsymbol {V} ] _ {2} \\dots [ \\boldsymbol {V} ] _ {N} [ \\mathrm {L A B E L} ], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 779, + 785, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $[\\mathbf{V}]_n\\in \\mathcal{R}^D (n\\in \\{1,2\\dots ,N\\})$ represents the learnable context tokens, and $N$ is the total number of context tokens.", + "bbox": [ + 212, + 808, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2fbc81e46f72d0789c3b2c69b77dcf1abda30276c86d7efaf9bdf791056a7959.jpg", + "image_caption": [ + "Fig. 3: Visualization of retrieved images (left), generated images (right), and the statistic of their cosine similarity to the class labels (middle). A larger mean indicates greater image-label consistency, and a larger standard deviation implies more diversity." + ], + "image_footnote": [], + "bbox": [ + 264, + 146, + 421, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/73ee1aca0decd463cbe771ec4897987c38cf8d11bf3a77536958dd3397b08ae7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 423, + 146, + 591, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/563bec17de7b0a7b879641ef7d362a874bdce252380b25f09cb43591ae9b4823.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 147, + 738, + 279 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We assess three prompt construction variations for OOD detection: unified, class-specific, and distribution-aware prompts. The unified prompt employs identical context tokens across all classes, and the class-specific prompt utilizes unique tokens for each class, which are explored in the ID classification [72]. Here, we propose a distribution-aware prompt strategy, which differentiates tokens for ID and OOD classes, as shown in Fig. 2. Our empirical results suggest that the distribution-aware prompt notably enhances the distinction between ID and OOD distributions, as analyzed in Sec. 4.3.", + "bbox": [ + 212, + 362, + 784, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Automated sample collection with class labels. Traditional prompt learning methods [32, 72] often optimize the prompts with manually collected image samples, which are typically selected from well-constructed benchmark datasets. Such a manual sampling approach is inefficient for negative label-based OOD detection, primarily because mined negative classes are often not included in existing datasets. Moreover, manually collecting and annotating these data is time-consuming and labor-intensive. To mitigate the manual effort and facilitate practical application, we introduce automated sample collection strategies that require only class names, as detailed subsequently.", + "bbox": [ + 212, + 503, + 784, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We introduce two automated sample collection methods for enhancing OOD detection: text-to-image generation and text-based image retrieval. In the generation approach, class names are fed into pre-trained text-to-image models [35,39] to generate synthetic images with varied appearances using different seeds:", + "bbox": [ + 212, + 638, + 784, + 700 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {X} _ {c o l} ^ {s} = \\mathrm {T 2 I} (\\mathcal {Y} \\cup \\mathcal {Y} ^ {-}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 423, + 709, + 784, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{X}_{col}^{s}$ is the collected set of synthetic images and $\\mathrm{T2I}(\\cdot)$ is the pre-trained generation model. The performance of OOD detection is closely tied to the fidelity of these generated images, depending on the generative power of the underlying model. By harnessing advanced generative models [35, 39], we are able to produce images that are not only visually convincing but also exhibit strong consistency with the input text. Our empirical findings indicate that these synthetic images can significantly enhance the OOD detection capabilities.", + "bbox": [ + 212, + 733, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "LAPT for OOD Detection with Vision-Language Models", + "bbox": [ + 352, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to generating synthetic images with generation models, we also automate the collection of real images through text-based image retrieval. Specifically, we utilize a large-scale real image dataset (e.g., Laion [42, 43]) in conjunction with pre-trained VLMs, such as CLIP, which align images and text within the same feature space. By conducting a search within this space, we retrieve images that are closest to the class names:", + "bbox": [ + 212, + 146, + 787, + 237 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {X} _ {\\text {c o l}} ^ {r} = \\operatorname {C l i p - r e t r i e v a l} (\\mathcal {Y} \\cup \\mathcal {Y} ^ {-}, \\text {W e b D a t a}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 250, + 785, + 266 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{X}_{col}^{r}$ is the retrieved set of real images, Clip-retrieval is the clip-retrieval tool [1], which enables efficient image retrieval from millions of samples, and Data is the web-scale real image set.", + "bbox": [ + 212, + 275, + 785, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We capitalize on the strengths of both synthetic image generation and real image retrieval. Synthetic images ensure high textual fidelity and control but may lack variety, while real images offer diversity but can vary in text consistency, as shown in Fig. 3. Our strategy prioritizes real images with high text-image alignment, quantified by the cosine similarity between image representation and corresponding text label. If real images fall short in text correlation, indicated by cosine similarity below a threshold, synthetic images are used. This hybrid approach is formalized as follows:", + "bbox": [ + 212, + 321, + 787, + 441 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {X} _ {c o l} = \\left\\{\\boldsymbol {x}: \\boldsymbol {x} \\in \\mathcal {X} _ {c o l} ^ {r} \\text {i f} \\cos \\left(f _ {i m g} (\\boldsymbol {x}), f _ {t x t} (y)\\right) > \\kappa \\text {e l s e} \\boldsymbol {x} \\in \\mathcal {X} _ {c o l} ^ {s} \\right\\}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 274, + 454, + 785, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{X}_{col}$ represents the final collection of images. The decision to use a real image or a synthetic one is based on whether the cosine similarity exceeds the threshold $\\kappa$ , which is set as 0.3 by default following [42]. This hybrid approach optimizes the balance between image diversity and text-image consistency, enhancing the robustness of our method.", + "bbox": [ + 212, + 479, + 787, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt tuning with cross-modal and cross-distribution mixing. Given the training images collected on a per-class basis, the most straightforward learning approach is to use a classification objective, which is commonly achieved with the following cross-entropy loss function:", + "bbox": [ + 214, + 578, + 787, + 640 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = - \\frac {1}{| \\mathcal {X} _ {c o l} |} \\sum_ {\\boldsymbol {x} \\in \\mathcal {X} _ {c o l}} \\sum_ {i = 1} ^ {C + M} \\boldsymbol {l} _ {i} \\log \\frac {e ^ {\\cos (\\boldsymbol {v} , (f _ {t x t} (\\rho_ {l} (y))) / \\tau}}{\\sum_ {i = 1} ^ {C + M} e ^ {\\cos (\\boldsymbol {v} , \\cos (f _ {t x t} (\\rho_ {l} (y)))) / \\tau}}, \\boldsymbol {v} = f _ {i m g} (\\boldsymbol {x}) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 650, + 785, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $y$ is the class name of $\\pmb{x}$ . $\\pmb{l}_i$ is the $i$ -th entry of $\\pmb{l} \\in \\mathcal{R}^{C + M}$ , the softlabel of sample $\\pmb{x}$ . In Sec. 4.3, we will demonstrate that this training objective aligns effectively with our proposed distribution-aware prompt construction and automated sample collection strategies.", + "bbox": [ + 212, + 703, + 785, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cross-modal mixing. While the simple classification loss has achieved commendable results, we could further enhance it by addressing potential sample noise and exploring broader data space through carefully designed data mixing strategies. Despite the efforts to collect clean images that match their respective class names, noise is inevitable in the collected images. To mitigate the impact", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of image noise, we introduce a cross-modal mixing strategy to neutralize potential noise in the images. Specifically, for a given training image $\\pmb{x}$ of class $y$ , we introduce the refined training feature $\\pmb{v}_{cm}$ as follows:", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {c m} = \\mathrm {L} _ {2} \\left(\\lambda_ {\\alpha} f _ {i m g} (\\boldsymbol {x}) + (1 - \\lambda_ {\\alpha}) f _ {t x t} (\\rho (y))\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 200, + 784, + 215 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {l} _ {c m} = \\boldsymbol {l}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 220, + 784, + 234 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda_{\\alpha} \\sim \\mathrm{Beta}(\\alpha, \\alpha) \\in [0,1]$ with $\\alpha \\in (0,\\infty)$ . $\\mathrm{L}_2(\\cdot)$ represents the $L_2$ normalization along feature dimension and $\\rho(\\cdot)$ is the hand-crafted text prompts [36]. The softlabel is unchanged after mixing since the image $\\mathbf{x}$ and its label $y$ belong to the same class. When $\\lambda_{\\alpha} = 1$ , we adopt the image features as the training examples, and we use the text features as the training examples when $\\lambda_{\\alpha} = 0$ . We denote the cross-entropy loss of Eq. (9) with $\\pmb{v}_{cm}$ and $\\pmb{l}_{cm}$ as $\\mathcal{L}_{cm}$ . By mixing textual and visual representations, we aim to create a more robust feature set that can help the model learn to focus on relevant features while disregarding noisy information. This strategy not only reinforces the model's resistance to image noise but also encourages it to leverage complementary information from both modalities.", + "bbox": [ + 212, + 242, + 784, + 407 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Cross-distribution mixing. In addition to using cross-modal data mixing to mitigate image noise, we employ the data mixing strategy to explore a broader feature space. In the vanilla NegLabel approach [21], the selected negative classes are typically far from the ID distribution, leaving a substantial intermediate area between ID and negative regions underutilized. To enhance the utilization of these spaces, we introduce a cross-distribution mixing strategy, which combines features and corresponding labels of ID and negative samples to create new training samples. This mixing strategy effectively bridges the distribution gap, generating a continuous spectrum of features that extend from ID to negative areas. The mixing process can be formalized as follows:", + "bbox": [ + 212, + 409, + 784, + 559 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {c d} = \\operatorname {L 2} \\left(\\lambda_ {\\beta} f _ {i m g} (\\boldsymbol {x} _ {i d}) + (1 - \\lambda_ {\\beta}) f _ {i m g} (\\boldsymbol {x} _ {o o d})\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 569, + 784, + 585 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {l} _ {c d} = \\lambda_ {\\beta} \\boldsymbol {l} _ {i d} + (1 - \\lambda_ {\\beta}) \\boldsymbol {l} _ {o o d}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 588, + 784, + 603 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\pmb{x}_{id}$ and $\\pmb{x}_{ood}$ are collected positive and negative images, and $l_{id}$ and $l_{ood}$ are soft labels of $\\pmb{x}_{id}$ and $\\pmb{x}_{ood}$ , respectively. $\\lambda_{\\beta} \\sim \\mathrm{Beta}(\\beta, \\beta) \\in [0,1]$ with $\\beta \\in (0,\\infty)$ . This method not only allows the model to become more aware of the intermediate space between ID and negative regions, but also encourages it to learn more discriminative features that can better generalize to new and unseen OOD samples. We denote the cross-entropy loss of Eq. (9) with $\\pmb{v}_{cd}$ and $l_{cd}$ as $\\mathcal{L}_{cd}$ . The final objective is constructed as:", + "bbox": [ + 212, + 611, + 784, + 715 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a l l} = \\mathcal {L} + \\mathcal {L} _ {c m} + \\mathcal {L} _ {c d}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 726, + 784, + 741 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Remarks. Our proposed data mixing strategies are distinct from existing ones [51,61,62,64] in both implementation and objectives. Traditional data mixing strategies typically aim to enhance ID classification performance by mixing randomly selected ID images [51,62]. In contrast, our approach is carefully designed to include multi-modal and multi-distribution data in the mixing process to enhance the OOD detection capability. While there are some methods", + "bbox": [ + 212, + 750, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "LAPT for OOD Detection with Vision-Language Models", + "bbox": [ + 352, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Algorithm 1 Label-driven Automated Prompt Tuning (LAPT)", + "text_level": 1, + "bbox": [ + 217, + 146, + 679, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Require: ID label space $\\mathcal{V}$", + "bbox": [ + 217, + 166, + 403, + 180 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: $\\mathcal{V}^{-}\\gets \\mathrm{NegMine}(\\mathcal{V})$", + "2: Constructing distribution-aware prompts $\\rho_{l}$ with labels $\\mathcal{V} \\cup \\mathcal{V}^{-}$ using Eq. (5)", + "3: $\\mathcal{X}_{col} \\gets$ Automated sample collection with labels $\\mathcal{Y} \\cup \\mathcal{Y}^{-}$ using Eq. (8)", + "4: Prompt tuning with cross-modal and cross-distribution data mixing using Eq. (14)" + ], + "bbox": [ + 222, + 181, + 782, + 236 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5: return Learned prompts $\\rho_l$", + "bbox": [ + 220, + 250, + 433, + 263 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "that utilize data mixing for OOD detection, they usually mix ID data to simulate OOD representations [61] or combining manually collected ID and OOD data [64]. These methods, however, do not exploit the textual knowledge and require manually labeled samples. Our cross-modal mixing strategy lowers image noise by blending samples from the same class but across different modalities, creating more robust features. Meanwhile, our cross-distribution mixing scheme goes a step further by mixing automatically collected positive and negative data, allowing us to explore the space that lies between ID and negative regions. This exploration enables our model to better understand and characterize the boundary between ID and OOD, leading to more effective detection of anomalies. Our methods are not just mixing data; they are deliberately designed to align with our label-driven automated prompt learning objective. By addressing noise and exploring intermediate feature space through mixing, we develop a tailored solution that boosts the automated prompt tuning for OOD detection.", + "bbox": [ + 212, + 295, + 787, + 507 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In the test stage, we adopt the OOD score of Eq. (4) following [21]. The overall pipeline of our automated prompt learning is summarized in Algorithm 1.", + "bbox": [ + 212, + 507, + 785, + 537 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 559, + 375, + 575 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Setup", + "text_level": 1, + "bbox": [ + 215, + 590, + 310, + 604 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets and benchmarks. Our validation primarily uses the ImageNet-1k dataset [4] as ID data, and more evaluations on smaller-scaled ID datasets are provided in the Supplementary Materials. Aligning with standards from prior research [19, 21, 29], we use four diverse datasets of iNaturalist [49], SUN [56], Places [71], and Textures [3] as OOD test sets. Additionally, we assess our method on the OpenOOD benchmark [58, 65], which differentiates OOD datasets into near-OOD (e.g., SSB-hard [50], NINCO [2]) and far-OOD (e.g., iNaturalist [49], Textures [3], OpenImage-O [52]) based on semantic similarity or difficulty. This grouping allows for a comprehensive evaluation of OOD detectors against various OOD sample types.", + "bbox": [ + 212, + 612, + 787, + 763 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In addition to conventional OOD detection that solely considers semantic shifts, we consider a full-spectrum OOD detection problem [60], which additionally accounts for non-semantic covariate shift generalization. An optimal system should not only be capable of identifying semantically shifted OOD samples but should also exhibit robustness to non-semantic covariate shifted OOD samples.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/24bc101485a7f7339b86a906df9be9c43db6298d3cc9a648e4be6ff4477988a0.jpg", + "table_caption": [ + "Table 1: OOD detection results with ID data of ImageNet-1k and four OOD datasets by using the VITB/16 CLIP encoder." + ], + "table_footnote": [], + "table_body": "
MethodsOOD datasets
INaturalistSunPlacesTexturesAverage
AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓
Requires Manually Labeled Training Samples
MSP [15]87.4458.3679.7373.7279.6774.4179.6971.9381.6369.61
ODIN [25]94.6530.2287.1754.0485.5455.0687.8551.6788.8047.75
Energy [26]95.3326.1292.6635.9791.4139.8786.7657.6191.5439.89
GradNorm [18]72.5681.5072.8682.0073.7080.4170.2679.3672.3580.82
ViM [52]93.1632.1987.1954.0183.7560.6787.1853.9487.8250.20
KNN [46]94.5229.1792.6735.6291.0239.6185.6764.3590.9742.19
VOS [7]94.6228.9992.5736.8891.2338.3986.3361.0291.1941.32
NPOS [48]96.1916.5890.4443.7789.4445.2788.8046.1291.2237.93
LSN [34]95.8321.5694.3526.3291.2534.4890.4238.5492.9630.22
LoCoOp [32]96.8616.0595.0723.4491.9832.8790.1942.2893.5228.66
Does Not Require Manually Labeled Training Samples
Mahalanobis [24]55.8999.3359.9499.4165.9698.5464.2398.4661.5098.94
Energy [26]85.0981.0884.2479.0283.3875.0865.5693.6579.5782.21
ZOC [9]86.0987.3081.2081.5183.3973.0676.4698.9081.7985.19
MCM [29]94.5932.2092.2538.8090.3146.2086.1258.5090.8243.93
CLIPN [53]95.2723.9493.9326.1792.2833.4590.9340.8393.1031.10
NegLabel [21]99.491.9195.4920.5391.6435.5990.2243.5694.2125.40
LAPT (Ours)99.631.1696.0119.1292.0133.0191.0640.3294.6823.40
", + "bbox": [ + 215, + 184, + 792, + 429 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Following the guidelines in [65], we adopt ImageNet-C [14], ImageNet-R [13], and ImageNet-V2 [37] as the covariate-shifted ID data in the full-spectrum OOD detection evaluation.", + "bbox": [ + 215, + 458, + 784, + 502 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Our evaluation metrics include (1) FPR95, the false positive rate of OOD samples with ID detection rate at $95\\%$ ; (2) AUROC, the area under the receiver operating characteristic curve; and (3) ID classification accuracy (ID ACC).", + "bbox": [ + 215, + 503, + 784, + 549 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation details. We adopt the visual encoder of VITB/16 pretrained by CLIP [36]. For automated sample collection, we instantiate the text-to-image generation model with SDXL [35] and employ a subset of the Laion400M dataset [42], approximately 60 million samples, as our WebData. Our model training is conducted using the SGD optimizer with a learning rate schedule that follows cosine annealing, starting from an initial learning rate of 1e-2. We set the batch size to 32 and train the model for 10 epochs in all experiments.", + "bbox": [ + 215, + 550, + 785, + 654 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 217, + 678, + 370, + 691 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results on four OOD datasets. As illustrated in Tab. 1, our approach consistently surpasses existing methods. Specifically, we re-implemented traditional methods [7, 15, 18, 25, 26, 46, 48, 52] by fine-tuning CLIP-encoders with labeled training samples from ImageNet in accordance with [48]. Without utilizing any manually annotated data, our method significantly outperforms these methods, corroborating the effectiveness of our proposed VLMs adaptation technique. Notably, our approach exceeds the performance of NegLabel [21], confirming the superiority of learnable prompts over hand-crafted prompts. The OOD samples in these datasets have a relatively large semantic difference from the ID data", + "bbox": [ + 215, + 704, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "LAPT for OOD Detection with Vision-Language Models", + "bbox": [ + 352, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/02b2a132d0ae1cc7caa6abf8f849d869699e94778488593d033f5b66082dc293.jpg", + "table_caption": [ + "Table 2: OOD detection results on the OpenOOD benchmark, where ImageNet-1k is adopted as ID dataset. Full results are available in the Supplementary Materials." + ], + "table_footnote": [], + "table_body": "
MethodsFPR95 ↓AUROC ↑ACC ↑ ID
Near-OOD Far-OODNear-OOD Far-OOD
Requires Manually Labeled Training Samples
GEN [27]--78.9790.9881.59
AugMix [17] + ReAct [45]--79.9493.7077.63
RMDS [38]--80.0992.6081.14
AugMix [17] + ASH [5]--82.1696.0577.63
Does Not Require Manually Labeled Training Samples
MCM [29]79.0268.5460.1184.7766.28
NegLabel [21]68.1827.3476.9293.3066.82
LAPT (Ours)58.9424.8682.6394.2667.86
", + "bbox": [ + 241, + 184, + 761, + 361 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(e.g., textures vs. animals), making the distinction relatively easy. On more challenging datasets, where the OOD data are semantically closer to the ID data, our advantages are even more pronounced, as shown in Tabs. 2 and 3.", + "bbox": [ + 212, + 387, + 784, + 431 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results on OpenOOD benchmark. A more comprehensive evaluation of our method is provided in Tab. 2 on the OpenOOD benchmark [58,65]. Compared to its closest competitor NegLabel, our method not only enhances the OOD detection performance but also improves the ID classification accuracy. In terms of OOD detection, our approach achieves a slight improvement in the relatively easy far-OOD scenarios (e.g., $0.96\\%$ AUROC) and a substantial improvement in the more challenging near-OOD settings (e.g., $5.71\\%$ AUROC), highlighting the outstanding effectiveness of our method in challenging contexts. The competitors with manually labeled training samples are quoted from [58,65], where models are trained with full ImageNet training data.", + "bbox": [ + 212, + 431, + 785, + 583 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Full-spectrum results on OpenOOD benchmark. In addition to the capabilities of semantic OOD detection and ID sample recognition, we also analyze the generalization ability of our method under covariate shift [68,69,73,74] on the full-spectrum OOD detection task [60] using the OpenOOD benchmark [58,65]. As illustrated in Tab. 3, our method enhances all the three capacities, achieving improved results across the board. Similar to the findings in Tab. 2, the improvement is more pronounced in the challenging near-OOD scenarios.", + "bbox": [ + 212, + 584, + 785, + 689 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Analyses and Discussions", + "text_level": 1, + "bbox": [ + 215, + 710, + 472, + 724 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The analyses are conducted on the OpenOOD dataset with a VITB/16 encoder.", + "bbox": [ + 215, + 733, + 782, + 748 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Prompt construction. We conduct comprehensive analyses on prompt construction as demonstrated in Fig. 4. As shown in Fig. 4a, our proposed distribution-aware prompts are well-matched for OOD detection, outperforming both class-specific and unified prompts. Excessively long prompts slightly impair experimental outcomes, as evidenced in Fig. 4b, possibly due to the increased capacity overfitting noise presented in the training images. Consequently,", + "bbox": [ + 212, + 750, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/f1f53d126a5c627c116cc3826dd4ee419ae329d1fbcd1e9526598ff181881bed.jpg", + "table_caption": [ + "Table 3: Full-spectrum OOD detection results on the OpenOOD benchmark, where ImageNet-1k, ImageNet-C, ImageNet-R, ImageNet-V2 are adopted as ID datasets." + ], + "table_footnote": [], + "table_body": "
MethodsFPR95 ↓AUROC ↑ACC ↑ ID
Near-OOD Far-OODNear-OOD Far-OOD
Requires Manually Labeled Training Samples
DeepAug [13] + SHE [66]83.2668.7068.2778.8557.82
StyAug [11] + GradNorm [18]87.1458.8265.2781.6255.44
AugMix [17] + SHE [66]84.4560.2669.6683.0657.46
ASH [5]93.2759.5660.5286.7554.35
LSA [28]70.5648.0678.2286.8561.74
ISH + SCALE [57]--68.0489.4654.88
Does Not Require Manually Labeled Training Samples
MCM [29]85.3769.8758.9777.1157.69
NegLabel [21]76.2533.3072.7792.0260.11
LAPT (Ours)71.1833.0774.7792.1461.17
", + "bbox": [ + 228, + 184, + 774, + 390 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/eb3482017246b13ee232cc6c9e99ff8e7c953a5f82a7052bca5c7049fac7d883.jpg", + "image_caption": [ + "Fig. 4: Analyses on prompt (a) types, (b) length, (b) initializations, and (d) positions." + ], + "image_footnote": [], + "bbox": [ + 217, + 404, + 356, + 525 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/301e8a53e6506831f7ac50114ec0c9471aea431c0b8ecfc08c2d6d26b2462caf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 402, + 500, + 526 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ba6e84b44e23c9a9cb63d4d65158fa9ab1a8a971bc689e70da6690f52c55b920.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 402, + 640, + 525 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1cc558c13428cc1d4a50181e84c4114f990fcf9ad0a234d4033c8e8beffbce19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 402, + 784, + 525 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "we set the prompt length as $N = 2$ by default. As indicated in Fig. 4c, initializing prompts with manually crafted text introduces textual bias. For instance, 'The nice' yields better results for far-OOD but proves suboptimal for near-OOD, a bias that is perpetuated in our method when learnable prompts are initialized with 'The nice'. We discover that a random initialization, which does not introduce any prior, performs consistently well across different OOD tasks and is thus adopted as the default practice. Finally, in Fig. 4d, we observe that the placement of label tokens has a minimal impact on the results; hence, we by default place the class label at the end of the prompt, as shown in Eq. (5)", + "bbox": [ + 212, + 582, + 787, + 717 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Automated sample collection. As shown in Fig. 5a, the synthetic image generation strategy outperforms realistic image retrieval, possibly because the generator has been exposed to a vast amount of training data (e.g., 5 billion images), which greatly surpasses the scale of our utilized Data (e.g., 60 million images). The best results are obtained by synergistically leveraging both synthetic and real data, confirming their complementary nature. In terms of image generation, employing a stronger image generator can result in better performance, as shown in Fig. 5b. Regarding image retrieval, utilizing a larger-", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "LAPT for OOD Detection with Vision-Language Models", + "bbox": [ + 352, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c8f2e4f3702b8e0796779848690d227a0d8b9a5702fe75d7251f738fed5ab07d.jpg", + "image_caption": [ + "(a) Collection Methods" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 356, + 250 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a29cfcf9dea77bc15da4d5e9ce80b54fb424eeddbe71bf1de87e259aaeb76186.jpg", + "image_caption": [ + "(b) Generators" + ], + "image_footnote": [], + "bbox": [ + 361, + 145, + 500, + 250 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d2158e20fce24fcbc8f3b37c396b41cc8746d5f545a37e29a1ee8f37092f802a.jpg", + "image_caption": [ + "(c) WebData Scale", + "Fig. 5: Analyses on (a) image collection methods, (b) text-to-image generators, (c) data scale of the WebData space, and (d) number of training samples per class." + ], + "image_footnote": [], + "bbox": [ + 504, + 143, + 640, + 250 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/7048f777e6c2586e9ec8052afd78ef9402b4aaa1b59c485911efbbe9fd51aadf.jpg", + "image_caption": [ + "(d) Sample Number" + ], + "image_footnote": [], + "bbox": [ + 647, + 145, + 784, + 250 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "scale WebData can yield better outcomes, as presented in Fig. 5c. Finally, we find in Fig. 5d that the model's performance does not continue to increase with the addition of more training samples. This may be because it is difficult to collect a high-quality training set that is semantically consistent with the text and diverse. With the relatively small size of WebData, it is challenging to ensure semantic consistency, and while the semantic consistency of generated images may be good, their diversity is limited, as analyzed in Fig. 3. How to automatically collect high-quality images for effective model training remains an enduring and challenging problem.", + "bbox": [ + 212, + 342, + 787, + 479 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "More analyses on stage ablation, the impact of text prompts in data collection, data mixing strategies, hyper-parameters in data mixing, different VLMs, visualization of learned prompts, and time complexity are provided in the Supplementary Materials.", + "bbox": [ + 212, + 483, + 787, + 544 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 215, + 587, + 519, + 603 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We developed a novel label-driven automated prompt tuning (LAPT) approach that autonomously configures effective prompts for OOD detection. Requiring only ID class names, our method can automatically mine OOD class names, collect training images per class, and learn effective prompts, thereby outperforming manually crafted prompts with minimal manual intervention. The automatically learned prompts not only enhanced the distinction between ID and OOD samples but also improved the ID classification accuracy and the strengthened generalization robustness to covariate shifts.", + "bbox": [ + 212, + 638, + 787, + 758 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The efficacy of our approach was highly dependent on the quality of collected training samples. Employing more advanced generation techniques to create a diverse set of synthetic images [8] or expanding the retrieval space [43] to obtain real images that match better the class names could yield improved results. These possibilities present exciting avenues for future research.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Beaumont, R.: Clip retrieval: Easily compute clip embeddings and build a clip retrieval system with them. https://github.com/rom1504/clip-retrieval (2022)", + "2. Bitterwolf, J., Mueller, M., Hein, M.: In or out? fixing imagenet out-of-distribution detection evaluation. In: ICML (2023), https://proceedings.mlr.press/v202/bitterwolf23a.html", + "3. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3606-3613 (2014)", + "4. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "5. Djurisic, A., Bozanic, N., Ashok, A., Liu, R.: Extremely simple activation shaping for out-of-distribution detection. arXiv preprint arXiv:2209.09858 (2022)", + "6. Du, X., Gozum, G., Ming, Y., Li, Y.: Siren: Shaping representations for detecting out-of-distribution objects. Advances in Neural Information Processing Systems 35, 20434-20449 (2022)", + "7. Du, X., Wang, X., Gozum, G., Li, Y.: Unknown-aware object detection: Learning what you don't know from videos in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13678-13688 (2022)", + "8. Dunlap, L., Umino, A., Zhang, H., Yang, J., Gonzalez, J.E., Darrell, T.: Diversify your vision datasets with automatic diffusion-based augmentation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=9wrYfqdrwk", + "9. Esmaeilpour, S., Liu, B., Robertson, E., Shu, L.: Zero-shot out-of-distribution detection based on the pre-trained model clip. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 6568-6576 (2022)", + "10. Fort, S., Ren, J., Lakshminarayanan, B.: Exploring the limits of out-of-distribution detection. Advances in Neural Information Processing Systems 34, 7068-7081 (2021)", + "1. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231 (2018)", + "2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "3. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021)", + "4. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019)", + "5. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. arXiv preprint arXiv:1610.02136 (2016)", + "6. Hendrycks, D., Mazeika, M., Dietterich, T.: Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606 (2018)", + "7. Hendrycks, D., Mu, N., Cubuk, E.D., Zoph, B., Gilmer, J., Lakshminarayanan, B.: Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781 (2019)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "LAPT for OOD Detection with Vision-Language Models", + "bbox": [ + 352, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Huang, R., Geng, A., Li, Y.: On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems 34, 677-689 (2021)", + "19. Huang, R., Li, Y.: Mos: Towards scaling out-of-distribution detection for large semantic space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8710-8719 (2021)", + "20. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)", + "21. Jiang, X., Liu, F., Fang, Z., Chen, H., Liu, T., Zheng, F., Han, B.: Negative label guided OOD detection with pretrained vision-language models. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=xU01HXz4an", + "22. Kong, S., Ramanan, D.: Opengan: Open-set recognition via open data generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 813-822 (2021)", + "23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012)", + "24. Lee, K., Lee, K., Lee, H., Shin, J.: A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in neural information processing systems 31 (2018)", + "25. Liang, S., Li, Y., Srikant, R.: Enhancing the reliability of out-of-distribution image detection in neural networks. arXiv preprint arXiv:1706.02690 (2017)", + "26. Liu, W., Wang, X., Owens, J., Li, Y.: Energy-based out-of-distribution detection. Advances in neural information processing systems 33, 21464-21475 (2020)", + "27. Liu, X., Lochman, Y., Zach, C.: Gen: Pushing the limits of softmax-based out-of-distribution detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23946-23955 (2023)", + "28. Lu, F., Zhu, K., Zheng, K., Zhai, W., Cao, Y.: Likelihood-aware semantic alignment for full-spectrum out-of-distribution detection. arXiv preprint arXiv:2312.01732 (2023)", + "29. Ming, Y., Cai, Z., Gu, J., Sun, Y., Li, W., Li, Y.: Delving into out-of-distribution detection with vision-language representations. Advances in Neural Information Processing Systems 35, 35087-35102 (2022)", + "30. Ming, Y., Li, Y.: How does fine-tuning impact out-of-distribution detection for vision-language models? International Journal of Computer Vision 132(2), 596-609 (2024)", + "31. Ming, Y., Sun, Y., Dia, O., Li, Y.: How to exploit hyperspherical embeddings for out-of-distribution detection? arXiv preprint arXiv:2203.04450 (2022)", + "32. Miyai, A., Yu, Q., Irie, G., Aizawa, K.: Locoop: Few-shot out-of-distribution detection via prompt learning. Advances in Neural Information Processing Systems 36 (2024)", + "33. Nguyen, A., Yosinski, J., Clune, J.: Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 427-436 (2015)", + "34. Nie, J., Zhang, Y., Fang, Z., Liu, T., Han, B., Tian, X.: Out-of-distribution detection with negative prompts. In: The Twelfth International Conference on Learning Representations (2024)" + ], + "bbox": [ + 215, + 147, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)", + "36. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "37. Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: Do imagenet classifiers generalize to imagenet? In: International conference on machine learning. pp. 5389-5400. PMLR (2019)", + "38. Ren, J., Fort, S., Liu, J., Roy, A.G., Padhy, S., Lakshminarayanan, B.: A simple fix to mahalanobis distance for improving near-ood detection. arXiv preprint arXiv:2106.09022 (2021)", + "39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "40. Ryu, S., Koo, S., Yu, H., Lee, G.G.: Out-of-domain detection based on generative adversarial network. In: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. pp. 714-718 (2018)", + "41. Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Toward open set recognition. IEEE transactions on pattern analysis and machine intelligence 35(7), 1757-1772 (2012)", + "42. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)", + "43. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)", + "44. Sehwag, V., Chiang, M., Mittal, P.: Ssd: A unified framework for self-supervised outlier detection. arXiv preprint arXiv:2103.12051 (2021)", + "45. Sun, Y., Guo, C., Li, Y.: React: Out-of-distribution detection with rectified activations. Advances in Neural Information Processing Systems 34, 144-157 (2021)", + "46. Sun, Y., Ming, Y., Zhu, X., Li, Y.: Out-of-distribution detection with deep nearest neighbors. In: International Conference on Machine Learning. pp. 20827-20840. PMLR (2022)", + "47. Tack, J., Mo, S., Jeong, J., Shin, J.: Csi: Novelty detection via contrastive learning on distributionally shifted instances. Advances in neural information processing systems 33, 11839-11852 (2020)", + "48. Tao, L., Du, X., Zhu, X., Li, Y.: Non-parametric outlier synthesis. arXiv preprint arXiv:2303.02966 (2023)", + "49. Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist species classification and detection dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8769-8778 (2018)", + "50. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Open-set recognition: A good closed-set classifier is all you need? arXiv preprint arXiv:2110.06207 (2021)", + "51. Verma, V., Lamb, A., Beckham, C., Najafi, A., Mitliagkas, I., Lopez-Paz, D., Bengio, Y.: Manifold mixup: Better representations by interpolating hidden states. In: International conference on machine learning. pp. 6438-6447. PMLR (2019)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "LAPT for OOD Detection with Vision-Language Models", + "bbox": [ + 352, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "52. Wang, H., Li, Z., Feng, L., Zhang, W.: Vim: Out-of-distribution with virtual-logit matching. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4921-4930 (2022)", + "53. Wang, H., Li, Y., Yao, H., Li, X.: Clipn for zero-shot ood detection: Teaching clip to say no. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1802-1812 (2023)", + "54. Wang, Y., Li, B., Che, T., Zhou, K., Liu, Z., Li, D.: Energy-based open-world uncertainty modeling for confidence calibration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9302-9311 (2021)", + "55. Wei, H., Xie, R., Cheng, H., Feng, L., An, B., Li, Y.: Mitigating neural network overconfidence with logit normalization. In: International Conference on Machine Learning. pp. 23631-23644. PMLR (2022)", + "56. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: Sun database: Large-scale scene recognition from abbey to zoo. In: 2010 IEEE computer society conference on computer vision and pattern recognition. pp. 3485-3492. IEEE (2010)", + "57. Xu, K., Chen, R., Franchi, G., Yao, A.: Scaling for training time and post-hoc out-of-distribution detection enhancement. arXiv preprint arXiv:2310.00227 (2023)", + "58. Yang, J., Wang, P., Zou, D., Zhou, Z., Ding, K., Peng, W., Wang, H., Chen, G., Li, B., Sun, Y., Du, X., Zhou, K., Zhang, W., Hendrycks, D., Li, Y., Liu, Z.: Openood: Benchmarking generalized out-of-distribution detection (2022)", + "59. Yang, J., Zhou, K., Li, Y., Liu, Z.: Generalized out-of-distribution detection: A survey. arXiv preprint arXiv:2110.11334 (2021)", + "60. Yang, J., Zhou, K., Liu, Z.: Full-spectrum out-of-distribution detection. International Journal of Computer Vision pp. 1-16 (2023)", + "61. Yang, T., Huang, Y., Xie, Y., Liu, J., Wang, S.: Mixood: Improving out-of-distribution detection with enhanced data mixup. ACM Transactions on Multimedia Computing, Communications and Applications 19(5), 1-18 (2023)", + "62. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)", + "63. Zhang, J., Inkawich, N., Chen, Y., Li, H.: Fine-grained out-of-distribution detection with mixup outlier exposure. CoRR (abs/2106.03917) (2021)", + "64. Zhang, J., Inkawich, N., Linderman, R., Chen, Y., Li, H.: Mixture outlier exposure: Towards out-of-distribution detection in fine-grained environments. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 5531-5540 (January 2023)", + "65. Zhang, J., Yang, J., Wang, P., Wang, H., Lin, Y., Zhang, H., Sun, Y., Du, X., Zhou, K., Zhang, W., Li, Y., Liu, Z., Chen, Y., Li, H.: Openood v1.5: Enhanced benchmark for out-of-distribution detection. arXiv preprint arXiv:2306.09301 (2023)", + "66. Zhang, J., Fu, Q., Chen, X., Du, L., Li, Z., Wang, G., Han, S., Zhang, D., et al.: Out-of-distribution detection based on in-distribution data patterns memorization with modern hopfield energy. In: The Eleventh International Conference on Learning Representations (2022)", + "67. Zhang, R., Fang, R., Zhang, W., Gao, P., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930 (2021)", + "68. Zhang, Y., Deng, B., Jia, K., Zhang, L.: Label propagation with augmented anchors: A simple semi-supervised learning baseline for unsupervised domain adaptation. In: European Conference on Computer Vision. pp. 781-797. Springer (2020)", + "69. Zhang, Y., Deng, B., Tang, H., Zhang, L., Jia, K.: Unsupervised multi-class domain adaptation: Theory, algorithms, and practice. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(5), 2775-2792 (2020)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "70. Zhang, Y., Zhu, W., Tang, H., Ma, Z., Zhou, K., Zhang, L.: Dual memory networks: A versatile adaptation approach for vision-language models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 28718-28728 (2024)", + "71. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence 40(6), 1452-1464 (2017)", + "72. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)", + "73. Zhou, Q., Feng, Z., Gu, Q., Pang, J., Cheng, G., Lu, X., Shi, J., Ma, L.: Context-aware mixup for domain adaptive semantic segmentation. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT) 33(2), 804-817 (2023)", + "74. Zhou, Q., Gu, Q., Pang, J., Lu, X., Ma, L.: Self-adversarial disentangling for specific domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 45(7), 8954-8968 (2023). https://doi.org/10.1109/TPAMI.2023.3238727" + ], + "bbox": [ + 215, + 146, + 787, + 367 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "LAPT for OOD Detection with Vision-Language Models", + "bbox": [ + 352, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/LAPT_ Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models/64854d06-c2d0-416b-8dda-0ced42b5873e_model.json b/2024/LAPT_ Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models/64854d06-c2d0-416b-8dda-0ced42b5873e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..13d29825b41ef0ce13e5f45fb491f7bf11aed284 --- /dev/null +++ b/2024/LAPT_ Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models/64854d06-c2d0-416b-8dda-0ced42b5873e_model.json @@ -0,0 +1,2779 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.219, + 0.141, + 0.784, + 0.187 + ], + "angle": 0, + "content": "LAPT: Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.212, + 0.733, + 0.229 + ], + "angle": 0, + "content": "Yabin Zhang\\(^{1,2}\\), Wenjie Zhu\\(^{1}\\), Chenhang He\\(^{1}\\), and Lei Zhang\\(^{1,2\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.362, + 0.239, + 0.641, + 0.255 + ], + "angle": 0, + "content": "1 The Hong Kong Polytechnic University" + }, + { + "type": "text", + "bbox": [ + 0.407, + 0.255, + 0.596, + 0.267 + ], + "angle": 0, + "content": "2 OPPO Research Institute" + }, + { + "type": "text", + "bbox": [ + 0.362, + 0.268, + 0.642, + 0.283 + ], + "angle": 0, + "content": "{csybzhang,cslzhang}@comp.polyu.edu.hk" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.322, + 0.744, + 0.671 + ], + "angle": 0, + "content": "Abstract. Out-of-distribution (OOD) detection is crucial for model reliability, as it identifies samples from unknown classes and reduces errors due to unexpected inputs. Vision-Language Models (VLMs) such as CLIP are emerging as powerful tools for OOD detection by integrating multi-modal information. However, the practical application of such systems is challenged by manual prompt engineering, which demands domain expertise and is sensitive to linguistic nuances. In this paper, we introduce Label-driven Automated Prompt Tuning (LAPT), a novel approach to OOD detection that reduces the need for manual prompt engineering. We develop distribution-aware prompts with in-distribution (ID) class names and negative labels mined automatically. Training samples linked to these class labels are collected autonomously via image synthesis and retrieval methods, allowing for prompt learning without manual effort. We utilize a simple cross-entropy loss for prompt optimization, with cross-modal and cross-distribution mixing strategies to reduce image noise and explore the intermediate space between distributions, respectively. The LAPT framework operates autonomously, requiring only ID class names as input and eliminating the need for manual intervention. With extensive experiments, LAPT consistently outperforms manually crafted prompts, setting a new standard for OOD detection. Moreover, LAPT not only enhances the distinction between ID and OOD samples, but also improves the ID classification accuracy and strengthens the generalization robustness to covariate shifts, resulting in outstanding performance in challenging full-spectrum OOD detection tasks. Codes are available at https://github.com/YBZh/LAPT." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.682, + 0.744, + 0.71 + ], + "angle": 0, + "content": "Keywords: Out-of-distribution detection \\(\\cdot\\) Vision-language models \\(\\cdot\\) Automated prompt tuning \\(\\cdot\\) Label-driven learning" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.738, + 0.377, + 0.753 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.77, + 0.788, + 0.817 + ], + "angle": 0, + "content": "In real-world applications, artificial intelligence (AI) systems often encounter data of unknown classes, known as out-of-distribution (OOD) data. When tackling such OOD data, AI models may erroneously remain overconfident in their" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.386, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.148, + 0.721, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Fig. 1: VLMs-based OOD detection systems are sensitive to linguistic nuances of text prompts, where results are reported on the OpenOOD dataset with a ViTB/16 image encoder. A lower FPR95 metric denotes better OOD detection performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.784, + 0.447 + ], + "angle": 0, + "content": "predictions [33, 41], resulting in critical errors and posing security risks. Consequently, it is crucial to recognize OOD data for maintaining the reliability and safety of AI systems in open-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Traditional OOD detection methods are primarily focused on image information [15,24,25], neglecting the rich textual knowledge carried by class names. With the rise of vision-language models (VLMs) [20,36], there has been a growing interest in leveraging textual information to facilitate visual tasks [67,70,72], including OOD detection [21,29,30]. For example, Ming et al. [29] compared test images with linguistic descriptions of in-distribution (ID) classes in feature space, leveraging the maximum probability of a scaled softmax function to identify OOD images. While this method achieves notable generalization across various datasets, it engages only a restricted portion of the textual space, thereby not fully exploiting the textual interpretative capabilities of VLMs. Jiang et al. [21] extended it by exploiting larger textual spaces, extracting negative labels from vast text corpora to enhance OOD detection performance significantly. However, the deployment of this system is hindered by the demanding task of prompt engineering, which requires domain-specific knowledge and is sensitive to fine-grained linguistic variations, as illustrated in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To this end, we propose Label-driven Automated Prompt Tuning (LAPT), a novel approach to OOD detection that substantially reduces the need of manual prompt engineering. Specifically, given ID class names and automatically extracted negative labels as in [21], we append distribution-aware prompt tokens to these class labels. These learnable tokens are optimized with minimal human intervention by aggregating training images associated with class labels using pre-trained text-to-image generation models [35, 39] or retrieving real images from web-scale dataset [42, 43]. We employ a simple yet effective cross-entropy loss for prompt optimization, enhanced by cross-modal and cross-distribution data mixing techniques. Cross-modal mixing merges image and text features from the same class, mitigating image noise through multi-modal integration." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.354, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "LAPT for OOD Detection with Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "Cross-distribution mixing stochastically blends ID and negative features along with their respective labels, exploring the intermediate space between ID and negative domains. During testing, we adopt the OOD score introduced in [21] to utilize both ID and negative labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.372 + ], + "angle": 0, + "content": "Extensive experiments demonstrate that our method outperforms the manually engineered text prompts by a large margin and achieves new state-of-the-art on various OOD detection benchmarks. Especially, on the challenging near-OD detection task using the ImageNet-1k as the ID dataset, our approach exceeds the best hand-crafted prompts by \\(10.76\\%\\) in FPR95 and \\(5.71\\%\\) in AUROC, without any manual label annotations or prompt crafting. Moreover, LAPT not only enhances the distinction between ID and OOD samples, but also improves the ID classification accuracy and strengthens the generalization robustness to covariate shifts. These improvements collectively contribute to the superior performance of LAPT in challenging full-spectrum OOD detection tasks [59, 60]. We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.382, + 0.784, + 0.426 + ], + "angle": 0, + "content": "- To circumvent the labor-intensive task of manual prompt engineering for OOD detection with VLMs, we introduce the LAPT methodology, which autonomously learns effective prompts using only ID class names." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.427, + 0.785, + 0.47 + ], + "angle": 0, + "content": "- Utilizing the ID class names and automatically mined negative labels, we devise distribution-aware prompts and autonomously gather the corresponding training images for each class via image synthesis and retrieval." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.472, + 0.785, + 0.531 + ], + "angle": 0, + "content": "- With the collected training data, we employ a simple yet effective cross-entropy objective for prompt optimization, enhanced by cross-modal data mixing to reduce image noise and cross-distribution data mixing to explore the intermediate space between ID and negative regions." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.532, + 0.785, + 0.621 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that our method significantly outperforms the hand-crafted text prompts and achieves new state-of-the-art results across multiple benchmarks. Moreover, LAPT enhances not only the distinction between ID and OOD samples, but also the ID classification accuracy and the generalization robustness to covariate shifts, achieving outstanding results in full-spectrum OOD detection tasks." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.382, + 0.785, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.644, + 0.388, + 0.659 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Traditional OOD detection aims to address the severe issue of deep learning models in producing overconfident predictions on OOD data [33, 59], which poses significant challenges to the deployment of deep models in real-world applications. To tackle this problem, researchers have developed various OOD detection methods, including score-based [15, 19, 24-26, 45, 52, 54, 55], distance-based [6, 31, 44, 46-48] and generative-based [22, 40] approaches. Score-based methods, which arguably garner the most attentions, distinguish ID from OOD samples using carefully designed scoring functions, such as confidence-based [15, 25, 45, 52, 55], discriminator-based [22], energy-based [26, 54], and gradient-based [18] scores. In contrast, distance-based methods detect OOD samples by measuring the distance in feature space between the test data and their nearest" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.283 + ], + "angle": 0, + "content": "ID samples [47] or ID prototypes [48] using metrics like KNN [6,31,46] or Mahalanobis distance [24,44]. While these methods have achieved certain success, they overlook textual information and require training or pre-training on manually labeled ID training data. Moreover, some approaches require manually collecting OOD data for additional training [10,16,63], hindering their real-world applicability. Our method differs from these approaches by effectively leveraging textual information. Utilizing only the ID class names, we mine negative labels, gather training data, and learn effective prompts in an automated fashion, achieving state-of-the-art OOD detection results with minimal human effort." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.789, + 0.603 + ], + "angle": 0, + "content": "OOD detection with VLMs is gaining growing attention as the importance of textual information for OOD image detection becomes increasingly recognized [9, 21, 29, 30, 32, 34, 53]. MCM [29] leverages ID class names for effective zero-shot OOD detection, which is further enhanced by NegLabel [21] through the additional use of mined OOD class names. However, as shown in Fig. 1, NegLabel's performance is sensitive to text prompts, limiting its practical application. ZOC [9] employs VLMs to detect OOD samples by learning a captioner that generates candidate OOD labels. Yet, this captioner struggles to generate effective OOD labels for large-scale ID datasets with numerous ID classes. LoCoOp [32] learns ID prompts using few-shot ID samples, regularizing these ID prompts by mined OOD features from image backgrounds. CLIPN [53] and LSN [34] learn text prompts for OOD classes. Our approach differs significantly from them in two aspects. First, we initialize negative prompts with negative labels instead of initializing negative prompts with 'no' and ID labels. Considering CLIP's difficulty in understanding the meaning of 'no' [34], using OOD labels to initialize negative prompts provides an effective starting point to accelerate convergence. Second, CLIPN and LSN respectively require large-scale multi-modal data and manually collected ID samples, leading to either high training costs or the need of manual annotation. In contrast, our method automatically gathers a small amount of positive/negative training data with ID labels, requiring minimal manual effort yet achieving improved performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.621, + 0.341, + 0.637 + ], + "angle": 0, + "content": "3 Methods" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.651, + 0.39, + 0.667 + ], + "angle": 0, + "content": "3.1 Problem Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Let \\(\\mathcal{X}\\) and \\(\\mathcal{Y} = \\{y_1, \\ldots, y_C\\}\\) be the image space and ID label space, where \\(\\mathcal{Y}\\) is a set of class names, e.g., \\(\\mathcal{Y} = \\{cat, dog, \\ldots, bird\\}\\) and \\(C\\) is the number of ID classes. Given the ID random variable \\(\\pmb{x}^{in} \\in \\mathcal{X}\\) and the OOD random variable \\(\\pmb{x}^{odd} \\in \\mathcal{X}\\), we use \\(\\mathcal{P}_{\\pmb{x}^{in}}\\) and \\(\\mathcal{P}_{\\pmb{x}^{odd}}\\) to denote the ID marginal distribution and OOD marginal distribution, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "OOD detection. In traditional classification tasks, we assume that the test image \\( \\pmb{x} \\) is sampled from the ID distribution and belongs to a specific ID class, i.e., \\( \\pmb{x} \\in \\mathcal{P}_{\\pmb{x}^{in}} \\) and \\( y \\in \\mathcal{V} \\), where \\( y \\) is the label of \\( \\pmb{x} \\). However, in real-world applications, AI systems often encounter samples from unknown classes, i.e., \\( \\pmb{x} \\in \\mathcal{P}_{\\pmb{x}^{ood}} \\) and \\( y \\notin \\mathcal{V} \\). AI models may misclassify these data into known ID classes with high confidence [33,41], posing critical errors and security risks. To tackle" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.353, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "LAPT for OOD Detection with Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "this problem, OOD detection is introduced to classify ID samples into correct ID classes and reject OOD samples simultaneously. The ID class classification is achieved with a \\( C \\)-way classifier following traditional classification [12, 23], while the OOD detection is typically achieved with a score function \\( S \\) [24-26]:" + }, + { + "type": "equation", + "bbox": [ + 0.31, + 0.218, + 0.785, + 0.234 + ], + "angle": 0, + "content": "\\[\nG _ {\\gamma} (\\boldsymbol {x}) = \\operatorname {I D}, \\text {i f} S (\\boldsymbol {x}) \\geq \\gamma ; \\quad \\text {o t h e r w i s e}, G _ {\\gamma} (\\boldsymbol {x}) = \\mathrm {O O D}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.242, + 0.784, + 0.273 + ], + "angle": 0, + "content": "where \\( G_{\\gamma} \\) is the OOD detector with the threshold \\( \\gamma \\). In other words, the test data \\( \\mathbf{x} \\) is detected as ID samples if and only if \\( S(\\mathbf{x}) \\geq \\gamma \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.293, + 0.533, + 0.308 + ], + "angle": 0, + "content": "3.2 Reviews on MCM and NegLabel" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.315, + 0.785, + 0.451 + ], + "angle": 0, + "content": "CLIP-like models. Given the ID test image \\(\\pmb{x}\\) within the label space \\(\\mathcal{V}\\), we extract the image representation \\(\\pmb{v} = f_{img}(\\pmb{x}) \\in \\mathcal{R}^D\\) and the textual representation \\(C = f_{txt}(\\rho(\\mathcal{Y})) \\in \\mathcal{R}^{C \\times D}\\) with the pre-trained CLIP-like encoders, where \\(D\\) denotes the feature dimension. \\(f_{img}(\\cdot)\\) and \\(f_{txt}(\\cdot)\\) denote the image and text encoders, respectively. \\(\\rho(\\cdot)\\) denotes the text prompt function, which is typically manually defined as 'a photo of a \\(<\\text{label}>\\).', where the 'label' token is replaced by the specific class name, e.g., 'cat' or 'dog'. Both \\(\\pmb{v}\\) and \\(C\\) are \\(L_2\\) normalized along the \\(D\\) dimension. Then the zero-shot classification probability can be achieved using \\(C\\) as the classifier:" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.46, + 0.785, + 0.477 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} = \\operatorname {S o f t m a x} \\left(\\boldsymbol {v} \\boldsymbol {C} ^ {T}\\right) \\in \\mathcal {R} ^ {C}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.486, + 0.641, + 0.501 + ], + "angle": 0, + "content": "where the fixed scaling parameter is omitted for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.501, + 0.784, + 0.606 + ], + "angle": 0, + "content": "MCM. Besides the remarkable classification capabilities, MCM [29] reveals that pre-trained CLIP models also exhibit strong zero-shot OOD detection abilities. Specifically, MCM treats textual representations \\( C \\) as ID prototypes and assesses OOD uncertainty based on the scaled cosine distance between the visual features and the nearest ID prototype. Furthermore, it has been validated that applying a softmax function as a post hoc mechanism enlarges the distinction between ID and OOD samples, leading to the following score function:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.615, + 0.785, + 0.653 + ], + "angle": 0, + "content": "\\[\nS _ {m c m} (\\boldsymbol {v}) = \\max _ {i} \\frac {e ^ {\\cos (\\boldsymbol {v} , \\boldsymbol {c} _ {i}) / \\tau}}{\\sum_ {i = 1} ^ {C} e ^ {\\cos (\\boldsymbol {v} , \\boldsymbol {c} _ {i}) / \\tau}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.66, + 0.784, + 0.689 + ], + "angle": 0, + "content": "where \\(\\cos (\\cdot ,\\cdot)\\) measures the cosine similarity, \\(c_{i}\\) is the \\(i\\) -th entry of \\(C\\) , and \\(\\tau >0\\) is the scaling temperature." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.69, + 0.785, + 0.795 + ], + "angle": 0, + "content": "NegLabel. While it is validated in [29] that leveraging the textual information of ID classes enhances the OOD detection capability, the potential contribution of negative labels is overlooked. To fill this gap, NegLabel [21] explores negative class names \\(\\mathcal{Y}^{-} = \\{y_{C + 1},\\dots ,y_{C + M}\\}\\) from extensive text corpora, where \\(\\mathcal{Y}^{-}\\cap \\mathcal{Y} = \\emptyset\\) . Under the assumption that ID samples are more similar to ID labels and less similar to negative labels than OOD samples, NegLabel employs a scoring function to improve OOD detection:" + }, + { + "type": "equation", + "bbox": [ + 0.306, + 0.805, + 0.785, + 0.845 + ], + "angle": 0, + "content": "\\[\nS _ {N e g L a b e l} (\\boldsymbol {v}) = \\frac {\\sum_ {i = 1} ^ {C} e ^ {\\cos (\\boldsymbol {v} , \\boldsymbol {c} _ {i}) / \\tau}}{\\sum_ {i = 1} ^ {C} e ^ {\\cos (\\boldsymbol {v} , \\boldsymbol {c} _ {i}) / \\tau} + \\sum_ {j = C + 1} ^ {C + M} e ^ {\\cos (\\boldsymbol {v} , \\boldsymbol {c} _ {j}) / \\tau}}. \\tag {4}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.59, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.143, + 0.788, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.348, + 0.788, + 0.379 + ], + "angle": 0, + "content": "Fig. 2: The overall framework of our LAPT method, where \\( \\mathbf{v}_{dog} / \\mathbf{v}_{boat} \\), \\( \\mathbf{c}_{dog} / \\mathbf{c}_{boat} \\), and \\( l_{dog} / l_{boat} \\) are image features, textual features, and soft labels of dog/boat samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.785, + 0.44 + ], + "angle": 0, + "content": "NegLabel performs excellently on various OOD detection benchmarks by utilizing both ID and negative text knowledge." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.466, + 0.662, + 0.483 + ], + "angle": 0, + "content": "3.3 LAPT: Label-driven Automated Prompt Tuning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.788, + 0.618 + ], + "angle": 0, + "content": "Though NegLabel has achieved a notable success, its performance is highly sensitive to the configuration of text prompts. As illustrated in Fig. 1, given an ID dataset like ImageNet, varying text prompts leads to considerable differences in OOD detection outcomes. For instance, prompts such as 'The nice \\(<\\) label>\\(^\\prime\\) generally improve performance on far-OOD datasets, while 'a photo of a \\(<\\) label>\\(^\\prime\\) shows enhanced effectiveness for near-OOD scenarios. The choice of prompt can lead to over \\(10\\%\\) fluctuation in the FPR95 metric. Hence, manually setting text prompts for optimal results is a labor-intensive and time-consuming process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.618, + 0.788, + 0.679 + ], + "angle": 0, + "content": "We aim to streamline OOD detection by automating prompt generation with minimal manual intervention. Our method learns continuous prompts using autocollected samples, needing only ID class names. The automated process is very useful in practical applications. Details of LAPT are provided as follows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.808, + 0.767 + ], + "angle": 0, + "content": "Distribution-aware prompts. Given the ID class names, we first follow NegLabel [21] to mine the negative labels \\(\\mathcal{V}^{-}\\) from extensive corpus databases. Then, instead of hand-crafting the optimal text prompts, we design a learnable text prompt function in the following form:" + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.78, + 0.787, + 0.798 + ], + "angle": 0, + "content": "\\[\n\\rho_ {l} (\\mathrm {L A B E L}) = [ \\boldsymbol {V} ] _ {1} [ \\boldsymbol {V} ] _ {2} \\dots [ \\boldsymbol {V} ] _ {N} [ \\mathrm {L A B E L} ], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.809, + 0.787, + 0.84 + ], + "angle": 0, + "content": "where \\([\\mathbf{V}]_n\\in \\mathcal{R}^D (n\\in \\{1,2\\dots ,N\\})\\) represents the learnable context tokens, and \\(N\\) is the total number of context tokens." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.354, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "LAPT for OOD Detection with Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.147, + 0.422, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.147, + 0.593, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.148, + 0.739, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.293, + 0.785, + 0.337 + ], + "angle": 0, + "content": "Fig. 3: Visualization of retrieved images (left), generated images (right), and the statistic of their cosine similarity to the class labels (middle). A larger mean indicates greater image-label consistency, and a larger standard deviation implies more diversity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.785, + 0.484 + ], + "angle": 0, + "content": "We assess three prompt construction variations for OOD detection: unified, class-specific, and distribution-aware prompts. The unified prompt employs identical context tokens across all classes, and the class-specific prompt utilizes unique tokens for each class, which are explored in the ID classification [72]. Here, we propose a distribution-aware prompt strategy, which differentiates tokens for ID and OOD classes, as shown in Fig. 2. Our empirical results suggest that the distribution-aware prompt notably enhances the distinction between ID and OOD distributions, as analyzed in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.785, + 0.64 + ], + "angle": 0, + "content": "Automated sample collection with class labels. Traditional prompt learning methods [32, 72] often optimize the prompts with manually collected image samples, which are typically selected from well-constructed benchmark datasets. Such a manual sampling approach is inefficient for negative label-based OOD detection, primarily because mined negative classes are often not included in existing datasets. Moreover, manually collecting and annotating these data is time-consuming and labor-intensive. To mitigate the manual effort and facilitate practical application, we introduce automated sample collection strategies that require only class names, as detailed subsequently." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.785, + 0.701 + ], + "angle": 0, + "content": "We introduce two automated sample collection methods for enhancing OOD detection: text-to-image generation and text-based image retrieval. In the generation approach, class names are fed into pre-trained text-to-image models [35,39] to generate synthetic images with varied appearances using different seeds:" + }, + { + "type": "equation", + "bbox": [ + 0.424, + 0.71, + 0.785, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\mathcal {X} _ {c o l} ^ {s} = \\mathrm {T 2 I} (\\mathcal {Y} \\cup \\mathcal {Y} ^ {-}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.787, + 0.841 + ], + "angle": 0, + "content": "where \\(\\mathcal{X}_{col}^{s}\\) is the collected set of synthetic images and \\(\\mathrm{T2I}(\\cdot)\\) is the pre-trained generation model. The performance of OOD detection is closely tied to the fidelity of these generated images, depending on the generative power of the underlying model. By harnessing advanced generative models [35, 39], we are able to produce images that are not only visually convincing but also exhibit strong consistency with the input text. Our empirical findings indicate that these synthetic images can significantly enhance the OOD detection capabilities." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.238 + ], + "angle": 0, + "content": "In addition to generating synthetic images with generation models, we also automate the collection of real images through text-based image retrieval. Specifically, we utilize a large-scale real image dataset (e.g., Laion [42, 43]) in conjunction with pre-trained VLMs, such as CLIP, which align images and text within the same feature space. By conducting a search within this space, we retrieve images that are closest to the class names:" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.25, + 0.786, + 0.267 + ], + "angle": 0, + "content": "\\[\n\\mathcal {X} _ {\\text {c o l}} ^ {r} = \\operatorname {C l i p - r e t r i e v a l} (\\mathcal {Y} \\cup \\mathcal {Y} ^ {-}, \\text {W e b D a t a}), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.276, + 0.786, + 0.322 + ], + "angle": 0, + "content": "where \\(\\mathcal{X}_{col}^{r}\\) is the retrieved set of real images, Clip-retrieval is the clip-retrieval tool [1], which enables efficient image retrieval from millions of samples, and Data is the web-scale real image set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.322, + 0.788, + 0.442 + ], + "angle": 0, + "content": "We capitalize on the strengths of both synthetic image generation and real image retrieval. Synthetic images ensure high textual fidelity and control but may lack variety, while real images offer diversity but can vary in text consistency, as shown in Fig. 3. Our strategy prioritizes real images with high text-image alignment, quantified by the cosine similarity between image representation and corresponding text label. If real images fall short in text correlation, indicated by cosine similarity below a threshold, synthetic images are used. This hybrid approach is formalized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.276, + 0.455, + 0.786, + 0.471 + ], + "angle": 0, + "content": "\\[\n\\mathcal {X} _ {c o l} = \\left\\{\\boldsymbol {x}: \\boldsymbol {x} \\in \\mathcal {X} _ {c o l} ^ {r} \\text {i f} \\cos \\left(f _ {i m g} (\\boldsymbol {x}), f _ {t x t} (y)\\right) > \\kappa \\text {e l s e} \\boldsymbol {x} \\in \\mathcal {X} _ {c o l} ^ {s} \\right\\}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.557 + ], + "angle": 0, + "content": "where \\(\\mathcal{X}_{col}\\) represents the final collection of images. The decision to use a real image or a synthetic one is based on whether the cosine similarity exceeds the threshold \\(\\kappa\\), which is set as 0.3 by default following [42]. This hybrid approach optimizes the balance between image diversity and text-image consistency, enhancing the robustness of our method." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.579, + 0.788, + 0.641 + ], + "angle": 0, + "content": "Prompt tuning with cross-modal and cross-distribution mixing. Given the training images collected on a per-class basis, the most straightforward learning approach is to use a classification objective, which is commonly achieved with the following cross-entropy loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.238, + 0.651, + 0.786, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = - \\frac {1}{| \\mathcal {X} _ {c o l} |} \\sum_ {\\boldsymbol {x} \\in \\mathcal {X} _ {c o l}} \\sum_ {i = 1} ^ {C + M} \\boldsymbol {l} _ {i} \\log \\frac {e ^ {\\cos (\\boldsymbol {v} , (f _ {t x t} (\\rho_ {l} (y))) / \\tau}}{\\sum_ {i = 1} ^ {C + M} e ^ {\\cos (\\boldsymbol {v} , \\cos (f _ {t x t} (\\rho_ {l} (y)))) / \\tau}}, \\boldsymbol {v} = f _ {i m g} (\\boldsymbol {x}) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.786, + 0.764 + ], + "angle": 0, + "content": "where \\(y\\) is the class name of \\(\\pmb{x}\\). \\(\\pmb{l}_i\\) is the \\(i\\)-th entry of \\(\\pmb{l} \\in \\mathcal{R}^{C + M}\\), the softlabel of sample \\(\\pmb{x}\\). In Sec. 4.3, we will demonstrate that this training objective aligns effectively with our proposed distribution-aware prompt construction and automated sample collection strategies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Cross-modal mixing. While the simple classification loss has achieved commendable results, we could further enhance it by addressing potential sample noise and exploring broader data space through carefully designed data mixing strategies. Despite the efforts to collect clean images that match their respective class names, noise is inevitable in the collected images. To mitigate the impact" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.353, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "LAPT for OOD Detection with Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "of image noise, we introduce a cross-modal mixing strategy to neutralize potential noise in the images. Specifically, for a given training image \\(\\pmb{x}\\) of class \\(y\\), we introduce the refined training feature \\(\\pmb{v}_{cm}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.202, + 0.785, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {c m} = \\mathrm {L} _ {2} \\left(\\lambda_ {\\alpha} f _ {i m g} (\\boldsymbol {x}) + (1 - \\lambda_ {\\alpha}) f _ {t x t} (\\rho (y))\\right), \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.221, + 0.785, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {l} _ {c m} = \\boldsymbol {l}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.243, + 0.785, + 0.408 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\alpha} \\sim \\mathrm{Beta}(\\alpha, \\alpha) \\in [0,1]\\) with \\(\\alpha \\in (0,\\infty)\\). \\(\\mathrm{L}_2(\\cdot)\\) represents the \\(L_2\\) normalization along feature dimension and \\(\\rho(\\cdot)\\) is the hand-crafted text prompts [36]. The softlabel is unchanged after mixing since the image \\(\\mathbf{x}\\) and its label \\(y\\) belong to the same class. When \\(\\lambda_{\\alpha} = 1\\), we adopt the image features as the training examples, and we use the text features as the training examples when \\(\\lambda_{\\alpha} = 0\\). We denote the cross-entropy loss of Eq. (9) with \\(\\pmb{v}_{cm}\\) and \\(\\pmb{l}_{cm}\\) as \\(\\mathcal{L}_{cm}\\). By mixing textual and visual representations, we aim to create a more robust feature set that can help the model learn to focus on relevant features while disregarding noisy information. This strategy not only reinforces the model's resistance to image noise but also encourages it to leverage complementary information from both modalities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.785, + 0.56 + ], + "angle": 0, + "content": "Cross-distribution mixing. In addition to using cross-modal data mixing to mitigate image noise, we employ the data mixing strategy to explore a broader feature space. In the vanilla NegLabel approach [21], the selected negative classes are typically far from the ID distribution, leaving a substantial intermediate area between ID and negative regions underutilized. To enhance the utilization of these spaces, we introduce a cross-distribution mixing strategy, which combines features and corresponding labels of ID and negative samples to create new training samples. This mixing strategy effectively bridges the distribution gap, generating a continuous spectrum of features that extend from ID to negative areas. The mixing process can be formalized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.57, + 0.785, + 0.586 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {c d} = \\operatorname {L 2} \\left(\\lambda_ {\\beta} f _ {i m g} (\\boldsymbol {x} _ {i d}) + (1 - \\lambda_ {\\beta}) f _ {i m g} (\\boldsymbol {x} _ {o o d})\\right), \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.589, + 0.785, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {l} _ {c d} = \\lambda_ {\\beta} \\boldsymbol {l} _ {i d} + (1 - \\lambda_ {\\beta}) \\boldsymbol {l} _ {o o d}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.785, + 0.717 + ], + "angle": 0, + "content": "where \\(\\pmb{x}_{id}\\) and \\(\\pmb{x}_{ood}\\) are collected positive and negative images, and \\(l_{id}\\) and \\(l_{ood}\\) are soft labels of \\(\\pmb{x}_{id}\\) and \\(\\pmb{x}_{ood}\\), respectively. \\(\\lambda_{\\beta} \\sim \\mathrm{Beta}(\\beta, \\beta) \\in [0,1]\\) with \\(\\beta \\in (0,\\infty)\\). This method not only allows the model to become more aware of the intermediate space between ID and negative regions, but also encourages it to learn more discriminative features that can better generalize to new and unseen OOD samples. We denote the cross-entropy loss of Eq. (9) with \\(\\pmb{v}_{cd}\\) and \\(l_{cd}\\) as \\(\\mathcal{L}_{cd}\\). The final objective is constructed as:" + }, + { + "type": "equation", + "bbox": [ + 0.423, + 0.727, + 0.785, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a l l} = \\mathcal {L} + \\mathcal {L} _ {c m} + \\mathcal {L} _ {c d}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Remarks. Our proposed data mixing strategies are distinct from existing ones [51,61,62,64] in both implementation and objectives. Traditional data mixing strategies typically aim to enhance ID classification performance by mixing randomly selected ID images [51,62]. In contrast, our approach is carefully designed to include multi-modal and multi-distribution data in the mixing process to enhance the OOD detection capability. While there are some methods" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.147, + 0.68, + 0.163 + ], + "angle": 0, + "content": "Algorithm 1 Label-driven Automated Prompt Tuning (LAPT)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.167, + 0.404, + 0.181 + ], + "angle": 0, + "content": "Require: ID label space \\(\\mathcal{V}\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.182, + 0.374, + 0.195 + ], + "angle": 0, + "content": "1: \\(\\mathcal{V}^{-}\\gets \\mathrm{NegMine}(\\mathcal{V})\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.195, + 0.755, + 0.209 + ], + "angle": 0, + "content": "2: Constructing distribution-aware prompts \\(\\rho_{l}\\) with labels \\(\\mathcal{V} \\cup \\mathcal{V}^{-}\\) using Eq. (5)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.209, + 0.71, + 0.222 + ], + "angle": 0, + "content": "3: \\(\\mathcal{X}_{col} \\gets\\) Automated sample collection with labels \\(\\mathcal{Y} \\cup \\mathcal{Y}^{-}\\) using Eq. (8)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.222, + 0.784, + 0.237 + ], + "angle": 0, + "content": "4: Prompt tuning with cross-modal and cross-distribution data mixing using Eq. (14)" + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.182, + 0.784, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.25, + 0.434, + 0.265 + ], + "angle": 0, + "content": "5: return Learned prompts \\(\\rho_l\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.296, + 0.788, + 0.508 + ], + "angle": 0, + "content": "that utilize data mixing for OOD detection, they usually mix ID data to simulate OOD representations [61] or combining manually collected ID and OOD data [64]. These methods, however, do not exploit the textual knowledge and require manually labeled samples. Our cross-modal mixing strategy lowers image noise by blending samples from the same class but across different modalities, creating more robust features. Meanwhile, our cross-distribution mixing scheme goes a step further by mixing automatically collected positive and negative data, allowing us to explore the space that lies between ID and negative regions. This exploration enables our model to better understand and characterize the boundary between ID and OOD, leading to more effective detection of anomalies. Our methods are not just mixing data; they are deliberately designed to align with our label-driven automated prompt learning objective. By addressing noise and exploring intermediate feature space through mixing, we develop a tailored solution that boosts the automated prompt tuning for OOD detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.787, + 0.539 + ], + "angle": 0, + "content": "In the test stage, we adopt the OOD score of Eq. (4) following [21]. The overall pipeline of our automated prompt learning is summarized in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.56, + 0.376, + 0.577 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.591, + 0.312, + 0.606 + ], + "angle": 0, + "content": "4.1 Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Datasets and benchmarks. Our validation primarily uses the ImageNet-1k dataset [4] as ID data, and more evaluations on smaller-scaled ID datasets are provided in the Supplementary Materials. Aligning with standards from prior research [19, 21, 29], we use four diverse datasets of iNaturalist [49], SUN [56], Places [71], and Textures [3] as OOD test sets. Additionally, we assess our method on the OpenOOD benchmark [58, 65], which differentiates OOD datasets into near-OOD (e.g., SSB-hard [50], NINCO [2]) and far-OOD (e.g., iNaturalist [49], Textures [3], OpenImage-O [52]) based on semantic similarity or difficulty. This grouping allows for a comprehensive evaluation of OOD detectors against various OOD sample types." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In addition to conventional OOD detection that solely considers semantic shifts, we consider a full-spectrum OOD detection problem [60], which additionally accounts for non-semantic covariate shift generalization. An optimal system should not only be capable of identifying semantically shifted OOD samples but should also exhibit robustness to non-semantic covariate shifted OOD samples." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.354, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LAPT for OOD Detection with Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 1: OOD detection results with ID data of ImageNet-1k and four OOD datasets by using the VITB/16 CLIP encoder." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.185, + 0.794, + 0.43 + ], + "angle": 0, + "content": "
MethodsOOD datasets
INaturalistSunPlacesTexturesAverage
AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓
Requires Manually Labeled Training Samples
MSP [15]87.4458.3679.7373.7279.6774.4179.6971.9381.6369.61
ODIN [25]94.6530.2287.1754.0485.5455.0687.8551.6788.8047.75
Energy [26]95.3326.1292.6635.9791.4139.8786.7657.6191.5439.89
GradNorm [18]72.5681.5072.8682.0073.7080.4170.2679.3672.3580.82
ViM [52]93.1632.1987.1954.0183.7560.6787.1853.9487.8250.20
KNN [46]94.5229.1792.6735.6291.0239.6185.6764.3590.9742.19
VOS [7]94.6228.9992.5736.8891.2338.3986.3361.0291.1941.32
NPOS [48]96.1916.5890.4443.7789.4445.2788.8046.1291.2237.93
LSN [34]95.8321.5694.3526.3291.2534.4890.4238.5492.9630.22
LoCoOp [32]96.8616.0595.0723.4491.9832.8790.1942.2893.5228.66
Does Not Require Manually Labeled Training Samples
Mahalanobis [24]55.8999.3359.9499.4165.9698.5464.2398.4661.5098.94
Energy [26]85.0981.0884.2479.0283.3875.0865.5693.6579.5782.21
ZOC [9]86.0987.3081.2081.5183.3973.0676.4698.9081.7985.19
MCM [29]94.5932.2092.2538.8090.3146.2086.1258.5090.8243.93
CLIPN [53]95.2723.9493.9326.1792.2833.4590.9340.8393.1031.10
NegLabel [21]99.491.9195.4920.5391.6435.5990.2243.5694.2125.40
LAPT (Ours)99.631.1696.0119.1292.0133.0191.0640.3294.6823.40
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.459, + 0.785, + 0.503 + ], + "angle": 0, + "content": "Following the guidelines in [65], we adopt ImageNet-C [14], ImageNet-R [13], and ImageNet-V2 [37] as the covariate-shifted ID data in the full-spectrum OOD detection evaluation." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.505, + 0.785, + 0.55 + ], + "angle": 0, + "content": "Our evaluation metrics include (1) FPR95, the false positive rate of OOD samples with ID detection rate at \\(95\\%\\); (2) AUROC, the area under the receiver operating characteristic curve; and (3) ID classification accuracy (ID ACC)." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.551, + 0.787, + 0.655 + ], + "angle": 0, + "content": "Implementation details. We adopt the visual encoder of VITB/16 pretrained by CLIP [36]. For automated sample collection, we instantiate the text-to-image generation model with SDXL [35] and employ a subset of the Laion400M dataset [42], approximately 60 million samples, as our WebData. Our model training is conducted using the SGD optimizer with a learning rate schedule that follows cosine annealing, starting from an initial learning rate of 1e-2. We set the batch size to 32 and train the model for 10 epochs in all experiments." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.679, + 0.371, + 0.692 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.705, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Results on four OOD datasets. As illustrated in Tab. 1, our approach consistently surpasses existing methods. Specifically, we re-implemented traditional methods [7, 15, 18, 25, 26, 46, 48, 52] by fine-tuning CLIP-encoders with labeled training samples from ImageNet in accordance with [48]. Without utilizing any manually annotated data, our method significantly outperforms these methods, corroborating the effectiveness of our proposed VLMs adaptation technique. Notably, our approach exceeds the performance of NegLabel [21], confirming the superiority of learnable prompts over hand-crafted prompts. The OOD samples in these datasets have a relatively large semantic difference from the ID data" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 2: OOD detection results on the OpenOOD benchmark, where ImageNet-1k is adopted as ID dataset. Full results are available in the Supplementary Materials." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.185, + 0.763, + 0.362 + ], + "angle": 0, + "content": "
MethodsFPR95 ↓AUROC ↑ACC ↑ ID
Near-OOD Far-OODNear-OOD Far-OOD
Requires Manually Labeled Training Samples
GEN [27]--78.9790.9881.59
AugMix [17] + ReAct [45]--79.9493.7077.63
RMDS [38]--80.0992.6081.14
AugMix [17] + ASH [5]--82.1696.0577.63
Does Not Require Manually Labeled Training Samples
MCM [29]79.0268.5460.1184.7766.28
NegLabel [21]68.1827.3476.9293.3066.82
LAPT (Ours)58.9424.8682.6394.2667.86
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.785, + 0.432 + ], + "angle": 0, + "content": "(e.g., textures vs. animals), making the distinction relatively easy. On more challenging datasets, where the OOD data are semantically closer to the ID data, our advantages are even more pronounced, as shown in Tabs. 2 and 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.787, + 0.584 + ], + "angle": 0, + "content": "Results on OpenOOD benchmark. A more comprehensive evaluation of our method is provided in Tab. 2 on the OpenOOD benchmark [58,65]. Compared to its closest competitor NegLabel, our method not only enhances the OOD detection performance but also improves the ID classification accuracy. In terms of OOD detection, our approach achieves a slight improvement in the relatively easy far-OOD scenarios (e.g., \\(0.96\\%\\) AUROC) and a substantial improvement in the more challenging near-OOD settings (e.g., \\(5.71\\%\\) AUROC), highlighting the outstanding effectiveness of our method in challenging contexts. The competitors with manually labeled training samples are quoted from [58,65], where models are trained with full ImageNet training data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.787, + 0.69 + ], + "angle": 0, + "content": "Full-spectrum results on OpenOOD benchmark. In addition to the capabilities of semantic OOD detection and ID sample recognition, we also analyze the generalization ability of our method under covariate shift [68,69,73,74] on the full-spectrum OOD detection task [60] using the OpenOOD benchmark [58,65]. As illustrated in Tab. 3, our method enhances all the three capacities, achieving improved results across the board. Similar to the findings in Tab. 2, the improvement is more pronounced in the challenging near-OOD scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.711, + 0.473, + 0.726 + ], + "angle": 0, + "content": "4.3 Analyses and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.734, + 0.784, + 0.749 + ], + "angle": 0, + "content": "The analyses are conducted on the OpenOOD dataset with a VITB/16 encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Prompt construction. We conduct comprehensive analyses on prompt construction as demonstrated in Fig. 4. As shown in Fig. 4a, our proposed distribution-aware prompts are well-matched for OOD detection, outperforming both class-specific and unified prompts. Excessively long prompts slightly impair experimental outcomes, as evidenced in Fig. 4b, possibly due to the increased capacity overfitting noise presented in the training images. Consequently," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.354, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "LAPT for OOD Detection with Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 3: Full-spectrum OOD detection results on the OpenOOD benchmark, where ImageNet-1k, ImageNet-C, ImageNet-R, ImageNet-V2 are adopted as ID datasets." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.185, + 0.775, + 0.391 + ], + "angle": 0, + "content": "
MethodsFPR95 ↓AUROC ↑ACC ↑ ID
Near-OOD Far-OODNear-OOD Far-OOD
Requires Manually Labeled Training Samples
DeepAug [13] + SHE [66]83.2668.7068.2778.8557.82
StyAug [11] + GradNorm [18]87.1458.8265.2781.6255.44
AugMix [17] + SHE [66]84.4560.2669.6683.0657.46
ASH [5]93.2759.5660.5286.7554.35
LSA [28]70.5648.0678.2286.8561.74
ISH + SCALE [57]--68.0489.4654.88
Does Not Require Manually Labeled Training Samples
MCM [29]85.3769.8758.9777.1157.69
NegLabel [21]76.2533.3072.7792.0260.11
LAPT (Ours)71.1833.0774.7792.1461.17
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.405, + 0.357, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.404, + 0.5, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.404, + 0.642, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.404, + 0.785, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.536, + 0.785, + 0.551 + ], + "angle": 0, + "content": "Fig. 4: Analyses on prompt (a) types, (b) length, (b) initializations, and (d) positions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.718 + ], + "angle": 0, + "content": "we set the prompt length as \\( N = 2 \\) by default. As indicated in Fig. 4c, initializing prompts with manually crafted text introduces textual bias. For instance, 'The nice' yields better results for far-OOD but proves suboptimal for near-OOD, a bias that is perpetuated in our method when learnable prompts are initialized with 'The nice'. We discover that a random initialization, which does not introduce any prior, performs consistently well across different OOD tasks and is thus adopted as the default practice. Finally, in Fig. 4d, we observe that the placement of label tokens has a minimal impact on the results; hence, we by default place the class label at the end of the prompt, as shown in Eq. (5)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Automated sample collection. As shown in Fig. 5a, the synthetic image generation strategy outperforms realistic image retrieval, possibly because the generator has been exposed to a vast amount of training data (e.g., 5 billion images), which greatly surpasses the scale of our utilized Data (e.g., 60 million images). The best results are obtained by synergistically leveraging both synthetic and real data, confirming their complementary nature. In terms of image generation, employing a stronger image generator can result in better performance, as shown in Fig. 5b. Regarding image retrieval, utilizing a larger-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.357, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.253, + 0.358, + 0.266 + ], + "angle": 0, + "content": "(a) Collection Methods" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.146, + 0.5, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.386, + 0.253, + 0.476, + 0.266 + ], + "angle": 0, + "content": "(b) Generators" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.145, + 0.642, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.253, + 0.63, + 0.266 + ], + "angle": 0, + "content": "(c) WebData Scale" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.146, + 0.785, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.253, + 0.776, + 0.266 + ], + "angle": 0, + "content": "(d) Sample Number" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.277, + 0.785, + 0.307 + ], + "angle": 0, + "content": "Fig. 5: Analyses on (a) image collection methods, (b) text-to-image generators, (c) data scale of the WebData space, and (d) number of training samples per class." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.48 + ], + "angle": 0, + "content": "scale WebData can yield better outcomes, as presented in Fig. 5c. Finally, we find in Fig. 5d that the model's performance does not continue to increase with the addition of more training samples. This may be because it is difficult to collect a high-quality training set that is semantically consistent with the text and diverse. With the relatively small size of WebData, it is challenging to ensure semantic consistency, and while the semantic consistency of generated images may be good, their diversity is limited, as analyzed in Fig. 3. How to automatically collect high-quality images for effective model training remains an enduring and challenging problem." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.484, + 0.788, + 0.545 + ], + "angle": 0, + "content": "More analyses on stage ablation, the impact of text prompts in data collection, data mixing strategies, hyper-parameters in data mixing, different VLMs, visualization of learned prompts, and time complexity are provided in the Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.588, + 0.52, + 0.604 + ], + "angle": 0, + "content": "5 Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.788, + 0.76 + ], + "angle": 0, + "content": "We developed a novel label-driven automated prompt tuning (LAPT) approach that autonomously configures effective prompts for OOD detection. Requiring only ID class names, our method can automatically mine OOD class names, collect training images per class, and learn effective prompts, thereby outperforming manually crafted prompts with minimal manual intervention. The automatically learned prompts not only enhanced the distinction between ID and OOD samples but also improved the ID classification accuracy and the strengthened generalization robustness to covariate shifts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The efficacy of our approach was highly dependent on the quality of collected training samples. Employing more advanced generation techniques to create a diverse set of synthetic images [8] or expanding the retrieval space [43] to obtain real images that match better the class names could yield improved results. These possibilities present exciting avenues for future research." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.354, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LAPT for OOD Detection with Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Beaumont, R.: Clip retrieval: Easily compute clip embeddings and build a clip retrieval system with them. https://github.com/rom1504/clip-retrieval (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.244 + ], + "angle": 0, + "content": "2. Bitterwolf, J., Mueller, M., Hein, M.: In or out? fixing imagenet out-of-distribution detection evaluation. In: ICML (2023), https://proceedings.mlr.press/v202/bitterwolf23a.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "3. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3606-3613 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.328 + ], + "angle": 0, + "content": "4. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.785, + 0.355 + ], + "angle": 0, + "content": "5. Djurisic, A., Bozanic, N., Ashok, A., Liu, R.: Extremely simple activation shaping for out-of-distribution detection. arXiv preprint arXiv:2209.09858 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.357, + 0.785, + 0.397 + ], + "angle": 0, + "content": "6. Du, X., Gozum, G., Ming, Y., Li, Y.: Siren: Shaping representations for detecting out-of-distribution objects. Advances in Neural Information Processing Systems 35, 20434-20449 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.785, + 0.439 + ], + "angle": 0, + "content": "7. Du, X., Wang, X., Gozum, G., Li, Y.: Unknown-aware object detection: Learning what you don't know from videos in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13678-13688 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.441, + 0.785, + 0.493 + ], + "angle": 0, + "content": "8. Dunlap, L., Umino, A., Zhang, H., Yang, J., Gonzalez, J.E., Darrell, T.: Diversify your vision datasets with automatic diffusion-based augmentation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=9wrYfqdrwk" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.785, + 0.536 + ], + "angle": 0, + "content": "9. Esmaeilpour, S., Liu, B., Robertson, E., Shu, L.: Zero-shot out-of-distribution detection based on the pre-trained model clip. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 6568-6576 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.785, + 0.577 + ], + "angle": 0, + "content": "10. Fort, S., Ren, J., Lakshminarayanan, B.: Exploring the limits of out-of-distribution detection. Advances in Neural Information Processing Systems 34, 7068-7081 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.579, + 0.785, + 0.619 + ], + "angle": 0, + "content": "1. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.621, + 0.785, + 0.66 + ], + "angle": 0, + "content": "2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.716 + ], + "angle": 0, + "content": "3. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.743 + ], + "angle": 0, + "content": "4. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.771 + ], + "angle": 0, + "content": "5. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. arXiv preprint arXiv:1610.02136 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "6. Hendrycks, D., Mazeika, M., Dietterich, T.: Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Hendrycks, D., Mu, N., Cubuk, E.D., Zoph, B., Gilmer, J., Lakshminarayanan, B.: Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "18. Huang, R., Geng, A., Li, Y.: On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems 34, 677-689 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "19. Huang, R., Li, Y.: Mos: Towards scaling out-of-distribution detection for large semantic space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8710-8719 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.286 + ], + "angle": 0, + "content": "20. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.342 + ], + "angle": 0, + "content": "21. Jiang, X., Liu, F., Fang, Z., Chen, H., Liu, T., Zheng, F., Han, B.: Negative label guided OOD detection with pretrained vision-language models. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=xU01HXz4an" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.384 + ], + "angle": 0, + "content": "22. Kong, S., Ramanan, D.: Opengan: Open-set recognition via open data generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 813-822 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.467 + ], + "angle": 0, + "content": "24. Lee, K., Lee, K., Lee, H., Shin, J.: A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in neural information processing systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.495 + ], + "angle": 0, + "content": "25. Liang, S., Li, Y., Srikant, R.: Enhancing the reliability of out-of-distribution image detection in neural networks. arXiv preprint arXiv:1706.02690 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "26. Liu, W., Wang, X., Owens, J., Li, Y.: Energy-based out-of-distribution detection. Advances in neural information processing systems 33, 21464-21475 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "27. Liu, X., Lochman, Y., Zach, C.: Gen: Pushing the limits of softmax-based out-of-distribution detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23946-23955 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "28. Lu, F., Zhu, K., Zheng, K., Zhai, W., Cao, Y.: Likelihood-aware semantic alignment for full-spectrum out-of-distribution detection. arXiv preprint arXiv:2312.01732 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.647 + ], + "angle": 0, + "content": "29. Ming, Y., Cai, Z., Gu, J., Sun, Y., Li, W., Li, Y.: Delving into out-of-distribution detection with vision-language representations. Advances in Neural Information Processing Systems 35, 35087-35102 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "30. Ming, Y., Li, Y.: How does fine-tuning impact out-of-distribution detection for vision-language models? International Journal of Computer Vision 132(2), 596-609 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "31. Ming, Y., Sun, Y., Dia, O., Li, Y.: How to exploit hyperspherical embeddings for out-of-distribution detection? arXiv preprint arXiv:2203.04450 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.785, + 0.757 + ], + "angle": 0, + "content": "32. Miyai, A., Yu, Q., Irie, G., Aizawa, K.: Locoop: Few-shot out-of-distribution detection via prompt learning. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "33. Nguyen, A., Yosinski, J., Clune, J.: Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 427-436 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "34. Nie, J., Zhang, Y., Fang, Z., Liu, T., Han, B., Tian, X.: Out-of-distribution detection with negative prompts. In: The Twelfth International Conference on Learning Representations (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.354, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LAPT for OOD Detection with Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "35. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "36. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "37. Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: Do imagenet classifiers generalize to imagenet? In: International conference on machine learning. pp. 5389-5400. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.288, + 0.785, + 0.328 + ], + "angle": 0, + "content": "38. Ren, J., Fort, S., Liu, J., Roy, A.G., Padhy, S., Lakshminarayanan, B.: A simple fix to mahalanobis distance for improving near-ood detection. arXiv preprint arXiv:2106.09022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.412 + ], + "angle": 0, + "content": "40. Ryu, S., Koo, S., Yu, H., Lee, G.G.: Out-of-domain detection based on generative adversarial network. In: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. pp. 714-718 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.453 + ], + "angle": 0, + "content": "41. Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Toward open set recognition. IEEE transactions on pattern analysis and machine intelligence 35(7), 1757-1772 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.508 + ], + "angle": 0, + "content": "42. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "43. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "44. Sehwag, V., Chiang, M., Mittal, P.: Ssd: A unified framework for self-supervised outlier detection. arXiv preprint arXiv:2103.12051 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "45. Sun, Y., Guo, C., Li, Y.: React: Out-of-distribution detection with rectified activations. Advances in Neural Information Processing Systems 34, 144-157 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "46. Sun, Y., Ming, Y., Zhu, X., Li, Y.: Out-of-distribution detection with deep nearest neighbors. In: International Conference on Machine Learning. pp. 20827-20840. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "47. Tack, J., Mo, S., Jeong, J., Shin, J.: Csi: Novelty detection via contrastive learning on distributionally shifted instances. Advances in neural information processing systems 33, 11839-11852 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "48. Tao, L., Du, X., Zhu, X., Li, Y.: Non-parametric outlier synthesis. arXiv preprint arXiv:2303.02966 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.785, + 0.771 + ], + "angle": 0, + "content": "49. Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist species classification and detection dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8769-8778 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "50. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Open-set recognition: A good closed-set classifier is all you need? arXiv preprint arXiv:2110.06207 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "51. Verma, V., Lamb, A., Beckham, C., Najafi, A., Mitliagkas, I., Lopez-Paz, D., Bengio, Y.: Manifold mixup: Better representations by interpolating hidden states. In: International conference on machine learning. pp. 6438-6447. PMLR (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "52. Wang, H., Li, Z., Feng, L., Zhang, W.: Vim: Out-of-distribution with virtual-logit matching. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4921-4930 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "53. Wang, H., Li, Y., Yao, H., Li, X.: Clipn for zero-shot ood detection: Teaching clip to say no. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1802-1812 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.271 + ], + "angle": 0, + "content": "54. Wang, Y., Li, B., Che, T., Zhou, K., Liu, Z., Li, D.: Energy-based open-world uncertainty modeling for confidence calibration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9302-9311 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.312 + ], + "angle": 0, + "content": "55. Wei, H., Xie, R., Cheng, H., Feng, L., An, B., Li, Y.: Mitigating neural network overconfidence with logit normalization. In: International Conference on Machine Learning. pp. 23631-23644. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.312, + 0.785, + 0.353 + ], + "angle": 0, + "content": "56. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: Sun database: Large-scale scene recognition from abbey to zoo. In: 2010 IEEE computer society conference on computer vision and pattern recognition. pp. 3485-3492. IEEE (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.353, + 0.785, + 0.379 + ], + "angle": 0, + "content": "57. Xu, K., Chen, R., Franchi, G., Yao, A.: Scaling for training time and post-hoc out-of-distribution detection enhancement. arXiv preprint arXiv:2310.00227 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.42 + ], + "angle": 0, + "content": "58. Yang, J., Wang, P., Zou, D., Zhou, Z., Ding, K., Peng, W., Wang, H., Chen, G., Li, B., Sun, Y., Du, X., Zhou, K., Zhang, W., Hendrycks, D., Li, Y., Liu, Z.: Openood: Benchmarking generalized out-of-distribution detection (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.447 + ], + "angle": 0, + "content": "59. Yang, J., Zhou, K., Li, Y., Liu, Z.: Generalized out-of-distribution detection: A survey. arXiv preprint arXiv:2110.11334 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.473 + ], + "angle": 0, + "content": "60. Yang, J., Zhou, K., Liu, Z.: Full-spectrum out-of-distribution detection. International Journal of Computer Vision pp. 1-16 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "61. Yang, T., Huang, Y., Xie, Y., Liu, J., Wang, S.: Mixood: Improving out-of-distribution detection with enhanced data mixup. ACM Transactions on Multimedia Computing, Communications and Applications 19(5), 1-18 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.542 + ], + "angle": 0, + "content": "62. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.568 + ], + "angle": 0, + "content": "63. Zhang, J., Inkawich, N., Chen, Y., Li, H.: Fine-grained out-of-distribution detection with mixup outlier exposure. CoRR (abs/2106.03917) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.568, + 0.785, + 0.623 + ], + "angle": 0, + "content": "64. Zhang, J., Inkawich, N., Linderman, R., Chen, Y., Li, H.: Mixture outlier exposure: Towards out-of-distribution detection in fine-grained environments. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 5531-5540 (January 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.785, + 0.664 + ], + "angle": 0, + "content": "65. Zhang, J., Yang, J., Wang, P., Wang, H., Lin, Y., Zhang, H., Sun, Y., Du, X., Zhou, K., Zhang, W., Li, Y., Liu, Z., Chen, Y., Li, H.: Openood v1.5: Enhanced benchmark for out-of-distribution detection. arXiv preprint arXiv:2306.09301 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.718 + ], + "angle": 0, + "content": "66. Zhang, J., Fu, Q., Chen, X., Du, L., Li, Z., Wang, G., Han, S., Zhang, D., et al.: Out-of-distribution detection based on in-distribution data patterns memorization with modern hopfield energy. In: The Eleventh International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "67. Zhang, R., Fang, R., Zhang, W., Gao, P., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "68. Zhang, Y., Deng, B., Jia, K., Zhang, L.: Label propagation with augmented anchors: A simple semi-supervised learning baseline for unsupervised domain adaptation. In: European Conference on Computer Vision. pp. 781-797. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "69. Zhang, Y., Deng, B., Tang, H., Zhang, L., Jia, K.: Unsupervised multi-class domain adaptation: Theory, algorithms, and practice. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(5), 2775-2792 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.354, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LAPT for OOD Detection with Vision-Language Models" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "70. Zhang, Y., Zhu, W., Tang, H., Ma, Z., Zhou, K., Zhang, L.: Dual memory networks: A versatile adaptation approach for vision-language models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 28718-28728 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.788, + 0.245 + ], + "angle": 0, + "content": "71. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence 40(6), 1452-1464 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.788, + 0.273 + ], + "angle": 0, + "content": "72. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.788, + 0.314 + ], + "angle": 0, + "content": "73. Zhou, Q., Feng, Z., Gu, Q., Pang, J., Cheng, G., Lu, X., Shi, J., Ma, L.: Context-aware mixup for domain adaptive semantic segmentation. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT) 33(2), 804-817 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.788, + 0.368 + ], + "angle": 0, + "content": "74. Zhou, Q., Gu, Q., Pang, J., Lu, X., Ma, L.: Self-adversarial disentangling for specific domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 45(7), 8954-8968 (2023). https://doi.org/10.1109/TPAMI.2023.3238727" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.368 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/LAPT_ Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models/64854d06-c2d0-416b-8dda-0ced42b5873e_origin.pdf b/2024/LAPT_ Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models/64854d06-c2d0-416b-8dda-0ced42b5873e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..01ace4ff83ee22552d4bf8885de99eeafc600b65 --- /dev/null +++ b/2024/LAPT_ Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models/64854d06-c2d0-416b-8dda-0ced42b5873e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3845c9153c6ecf60fda08aee0305a0183d0384b2655a52ec13560309326bdf8f +size 952462 diff --git a/2024/LAPT_ Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models/full.md b/2024/LAPT_ Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..058e79b132e2a78e9a656f19efee3593e6b03877 --- /dev/null +++ b/2024/LAPT_ Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models/full.md @@ -0,0 +1,353 @@ +# LAPT: Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models + +Yabin Zhang $^{1,2}$ , Wenjie Zhu $^{1}$ , Chenhang He $^{1}$ , and Lei Zhang $^{1,2\star}$ + +1 The Hong Kong Polytechnic University + +2 OPPO Research Institute + +{csybzhang,cslzhang}@comp.polyu.edu.hk + +Abstract. Out-of-distribution (OOD) detection is crucial for model reliability, as it identifies samples from unknown classes and reduces errors due to unexpected inputs. Vision-Language Models (VLMs) such as CLIP are emerging as powerful tools for OOD detection by integrating multi-modal information. However, the practical application of such systems is challenged by manual prompt engineering, which demands domain expertise and is sensitive to linguistic nuances. In this paper, we introduce Label-driven Automated Prompt Tuning (LAPT), a novel approach to OOD detection that reduces the need for manual prompt engineering. We develop distribution-aware prompts with in-distribution (ID) class names and negative labels mined automatically. Training samples linked to these class labels are collected autonomously via image synthesis and retrieval methods, allowing for prompt learning without manual effort. We utilize a simple cross-entropy loss for prompt optimization, with cross-modal and cross-distribution mixing strategies to reduce image noise and explore the intermediate space between distributions, respectively. The LAPT framework operates autonomously, requiring only ID class names as input and eliminating the need for manual intervention. With extensive experiments, LAPT consistently outperforms manually crafted prompts, setting a new standard for OOD detection. Moreover, LAPT not only enhances the distinction between ID and OOD samples, but also improves the ID classification accuracy and strengthens the generalization robustness to covariate shifts, resulting in outstanding performance in challenging full-spectrum OOD detection tasks. Codes are available at https://github.com/YBZh/LAPT. + +Keywords: Out-of-distribution detection $\cdot$ Vision-language models $\cdot$ Automated prompt tuning $\cdot$ Label-driven learning + +# 1 Introduction + +In real-world applications, artificial intelligence (AI) systems often encounter data of unknown classes, known as out-of-distribution (OOD) data. When tackling such OOD data, AI models may erroneously remain overconfident in their + +![](images/0efbef56b24ae67870d86badb2c358da0c20e7eb962992ce519e2907ec4db2df.jpg) +Fig. 1: VLMs-based OOD detection systems are sensitive to linguistic nuances of text prompts, where results are reported on the OpenOOD dataset with a ViTB/16 image encoder. A lower FPR95 metric denotes better OOD detection performance. + +predictions [33, 41], resulting in critical errors and posing security risks. Consequently, it is crucial to recognize OOD data for maintaining the reliability and safety of AI systems in open-world scenarios. + +Traditional OOD detection methods are primarily focused on image information [15,24,25], neglecting the rich textual knowledge carried by class names. With the rise of vision-language models (VLMs) [20,36], there has been a growing interest in leveraging textual information to facilitate visual tasks [67,70,72], including OOD detection [21,29,30]. For example, Ming et al. [29] compared test images with linguistic descriptions of in-distribution (ID) classes in feature space, leveraging the maximum probability of a scaled softmax function to identify OOD images. While this method achieves notable generalization across various datasets, it engages only a restricted portion of the textual space, thereby not fully exploiting the textual interpretative capabilities of VLMs. Jiang et al. [21] extended it by exploiting larger textual spaces, extracting negative labels from vast text corpora to enhance OOD detection performance significantly. However, the deployment of this system is hindered by the demanding task of prompt engineering, which requires domain-specific knowledge and is sensitive to fine-grained linguistic variations, as illustrated in Fig. 1. + +To this end, we propose Label-driven Automated Prompt Tuning (LAPT), a novel approach to OOD detection that substantially reduces the need of manual prompt engineering. Specifically, given ID class names and automatically extracted negative labels as in [21], we append distribution-aware prompt tokens to these class labels. These learnable tokens are optimized with minimal human intervention by aggregating training images associated with class labels using pre-trained text-to-image generation models [35, 39] or retrieving real images from web-scale dataset [42, 43]. We employ a simple yet effective cross-entropy loss for prompt optimization, enhanced by cross-modal and cross-distribution data mixing techniques. Cross-modal mixing merges image and text features from the same class, mitigating image noise through multi-modal integration. + +Cross-distribution mixing stochastically blends ID and negative features along with their respective labels, exploring the intermediate space between ID and negative domains. During testing, we adopt the OOD score introduced in [21] to utilize both ID and negative labels. + +Extensive experiments demonstrate that our method outperforms the manually engineered text prompts by a large margin and achieves new state-of-the-art on various OOD detection benchmarks. Especially, on the challenging near-OD detection task using the ImageNet-1k as the ID dataset, our approach exceeds the best hand-crafted prompts by $10.76\%$ in FPR95 and $5.71\%$ in AUROC, without any manual label annotations or prompt crafting. Moreover, LAPT not only enhances the distinction between ID and OOD samples, but also improves the ID classification accuracy and strengthens the generalization robustness to covariate shifts. These improvements collectively contribute to the superior performance of LAPT in challenging full-spectrum OOD detection tasks [59, 60]. We summarize our contributions as follows: + +- To circumvent the labor-intensive task of manual prompt engineering for OOD detection with VLMs, we introduce the LAPT methodology, which autonomously learns effective prompts using only ID class names. +- Utilizing the ID class names and automatically mined negative labels, we devise distribution-aware prompts and autonomously gather the corresponding training images for each class via image synthesis and retrieval. +- With the collected training data, we employ a simple yet effective cross-entropy objective for prompt optimization, enhanced by cross-modal data mixing to reduce image noise and cross-distribution data mixing to explore the intermediate space between ID and negative regions. +- Extensive experiments demonstrate that our method significantly outperforms the hand-crafted text prompts and achieves new state-of-the-art results across multiple benchmarks. Moreover, LAPT enhances not only the distinction between ID and OOD samples, but also the ID classification accuracy and the generalization robustness to covariate shifts, achieving outstanding results in full-spectrum OOD detection tasks. + +# 2 Related Work + +Traditional OOD detection aims to address the severe issue of deep learning models in producing overconfident predictions on OOD data [33, 59], which poses significant challenges to the deployment of deep models in real-world applications. To tackle this problem, researchers have developed various OOD detection methods, including score-based [15, 19, 24-26, 45, 52, 54, 55], distance-based [6, 31, 44, 46-48] and generative-based [22, 40] approaches. Score-based methods, which arguably garner the most attentions, distinguish ID from OOD samples using carefully designed scoring functions, such as confidence-based [15, 25, 45, 52, 55], discriminator-based [22], energy-based [26, 54], and gradient-based [18] scores. In contrast, distance-based methods detect OOD samples by measuring the distance in feature space between the test data and their nearest + +ID samples [47] or ID prototypes [48] using metrics like KNN [6,31,46] or Mahalanobis distance [24,44]. While these methods have achieved certain success, they overlook textual information and require training or pre-training on manually labeled ID training data. Moreover, some approaches require manually collecting OOD data for additional training [10,16,63], hindering their real-world applicability. Our method differs from these approaches by effectively leveraging textual information. Utilizing only the ID class names, we mine negative labels, gather training data, and learn effective prompts in an automated fashion, achieving state-of-the-art OOD detection results with minimal human effort. + +OOD detection with VLMs is gaining growing attention as the importance of textual information for OOD image detection becomes increasingly recognized [9, 21, 29, 30, 32, 34, 53]. MCM [29] leverages ID class names for effective zero-shot OOD detection, which is further enhanced by NegLabel [21] through the additional use of mined OOD class names. However, as shown in Fig. 1, NegLabel's performance is sensitive to text prompts, limiting its practical application. ZOC [9] employs VLMs to detect OOD samples by learning a captioner that generates candidate OOD labels. Yet, this captioner struggles to generate effective OOD labels for large-scale ID datasets with numerous ID classes. LoCoOp [32] learns ID prompts using few-shot ID samples, regularizing these ID prompts by mined OOD features from image backgrounds. CLIPN [53] and LSN [34] learn text prompts for OOD classes. Our approach differs significantly from them in two aspects. First, we initialize negative prompts with negative labels instead of initializing negative prompts with 'no' and ID labels. Considering CLIP's difficulty in understanding the meaning of 'no' [34], using OOD labels to initialize negative prompts provides an effective starting point to accelerate convergence. Second, CLIPN and LSN respectively require large-scale multi-modal data and manually collected ID samples, leading to either high training costs or the need of manual annotation. In contrast, our method automatically gathers a small amount of positive/negative training data with ID labels, requiring minimal manual effort yet achieving improved performance. + +# 3 Methods + +# 3.1 Problem Setup + +Let $\mathcal{X}$ and $\mathcal{Y} = \{y_1, \ldots, y_C\}$ be the image space and ID label space, where $\mathcal{Y}$ is a set of class names, e.g., $\mathcal{Y} = \{cat, dog, \ldots, bird\}$ and $C$ is the number of ID classes. Given the ID random variable $\pmb{x}^{in} \in \mathcal{X}$ and the OOD random variable $\pmb{x}^{odd} \in \mathcal{X}$ , we use $\mathcal{P}_{\pmb{x}^{in}}$ and $\mathcal{P}_{\pmb{x}^{odd}}$ to denote the ID marginal distribution and OOD marginal distribution, respectively. + +OOD detection. In traditional classification tasks, we assume that the test image $\pmb{x}$ is sampled from the ID distribution and belongs to a specific ID class, i.e., $\pmb{x} \in \mathcal{P}_{\pmb{x}^{in}}$ and $y \in \mathcal{V}$ , where $y$ is the label of $\pmb{x}$ . However, in real-world applications, AI systems often encounter samples from unknown classes, i.e., $\pmb{x} \in \mathcal{P}_{\pmb{x}^{ood}}$ and $y \notin \mathcal{V}$ . AI models may misclassify these data into known ID classes with high confidence [33,41], posing critical errors and security risks. To tackle + +this problem, OOD detection is introduced to classify ID samples into correct ID classes and reject OOD samples simultaneously. The ID class classification is achieved with a $C$ -way classifier following traditional classification [12, 23], while the OOD detection is typically achieved with a score function $S$ [24-26]: + +$$ +G _ {\gamma} (\boldsymbol {x}) = \operatorname {I D}, \text {i f} S (\boldsymbol {x}) \geq \gamma ; \quad \text {o t h e r w i s e}, G _ {\gamma} (\boldsymbol {x}) = \mathrm {O O D}, \tag {1} +$$ + +where $G_{\gamma}$ is the OOD detector with the threshold $\gamma$ . In other words, the test data $\mathbf{x}$ is detected as ID samples if and only if $S(\mathbf{x}) \geq \gamma$ . + +# 3.2 Reviews on MCM and NegLabel + +CLIP-like models. Given the ID test image $\pmb{x}$ within the label space $\mathcal{V}$ , we extract the image representation $\pmb{v} = f_{img}(\pmb{x}) \in \mathcal{R}^D$ and the textual representation $C = f_{txt}(\rho(\mathcal{Y})) \in \mathcal{R}^{C \times D}$ with the pre-trained CLIP-like encoders, where $D$ denotes the feature dimension. $f_{img}(\cdot)$ and $f_{txt}(\cdot)$ denote the image and text encoders, respectively. $\rho(\cdot)$ denotes the text prompt function, which is typically manually defined as 'a photo of a $<\text{label}>$ .', where the 'label' token is replaced by the specific class name, e.g., 'cat' or 'dog'. Both $\pmb{v}$ and $C$ are $L_2$ normalized along the $D$ dimension. Then the zero-shot classification probability can be achieved using $C$ as the classifier: + +$$ +\boldsymbol {p} = \operatorname {S o f t m a x} \left(\boldsymbol {v} \boldsymbol {C} ^ {T}\right) \in \mathcal {R} ^ {C}, \tag {2} +$$ + +where the fixed scaling parameter is omitted for simplicity. + +MCM. Besides the remarkable classification capabilities, MCM [29] reveals that pre-trained CLIP models also exhibit strong zero-shot OOD detection abilities. Specifically, MCM treats textual representations $C$ as ID prototypes and assesses OOD uncertainty based on the scaled cosine distance between the visual features and the nearest ID prototype. Furthermore, it has been validated that applying a softmax function as a post hoc mechanism enlarges the distinction between ID and OOD samples, leading to the following score function: + +$$ +S _ {m c m} (\boldsymbol {v}) = \max _ {i} \frac {e ^ {\cos (\boldsymbol {v} , \boldsymbol {c} _ {i}) / \tau}}{\sum_ {i = 1} ^ {C} e ^ {\cos (\boldsymbol {v} , \boldsymbol {c} _ {i}) / \tau}}, \tag {3} +$$ + +where $\cos (\cdot ,\cdot)$ measures the cosine similarity, $c_{i}$ is the $i$ -th entry of $C$ , and $\tau >0$ is the scaling temperature. + +NegLabel. While it is validated in [29] that leveraging the textual information of ID classes enhances the OOD detection capability, the potential contribution of negative labels is overlooked. To fill this gap, NegLabel [21] explores negative class names $\mathcal{Y}^{-} = \{y_{C + 1},\dots ,y_{C + M}\}$ from extensive text corpora, where $\mathcal{Y}^{-}\cap \mathcal{Y} = \emptyset$ . Under the assumption that ID samples are more similar to ID labels and less similar to negative labels than OOD samples, NegLabel employs a scoring function to improve OOD detection: + +$$ +S _ {N e g L a b e l} (\boldsymbol {v}) = \frac {\sum_ {i = 1} ^ {C} e ^ {\cos (\boldsymbol {v} , \boldsymbol {c} _ {i}) / \tau}}{\sum_ {i = 1} ^ {C} e ^ {\cos (\boldsymbol {v} , \boldsymbol {c} _ {i}) / \tau} + \sum_ {j = C + 1} ^ {C + M} e ^ {\cos (\boldsymbol {v} , \boldsymbol {c} _ {j}) / \tau}}. \tag {4} +$$ + +![](images/91b5bd354cf89231b19e97f1b85c84e46e1ac22abe9fe8ddbfc2bb010328bfef.jpg) +Fig. 2: The overall framework of our LAPT method, where $\mathbf{v}_{dog} / \mathbf{v}_{boat}$ , $\mathbf{c}_{dog} / \mathbf{c}_{boat}$ , and $l_{dog} / l_{boat}$ are image features, textual features, and soft labels of dog/boat samples. + +![](images/bc3f554681f97bdc63a23f7034badf751fbeb1c41f1a288e910402e571b4f914.jpg) + +NegLabel performs excellently on various OOD detection benchmarks by utilizing both ID and negative text knowledge. + +# 3.3 LAPT: Label-driven Automated Prompt Tuning + +Though NegLabel has achieved a notable success, its performance is highly sensitive to the configuration of text prompts. As illustrated in Fig. 1, given an ID dataset like ImageNet, varying text prompts leads to considerable differences in OOD detection outcomes. For instance, prompts such as 'The nice $<$ label> $^\prime$ generally improve performance on far-OOD datasets, while 'a photo of a $<$ label> $^\prime$ shows enhanced effectiveness for near-OOD scenarios. The choice of prompt can lead to over $10\%$ fluctuation in the FPR95 metric. Hence, manually setting text prompts for optimal results is a labor-intensive and time-consuming process. + +We aim to streamline OOD detection by automating prompt generation with minimal manual intervention. Our method learns continuous prompts using autocollected samples, needing only ID class names. The automated process is very useful in practical applications. Details of LAPT are provided as follows. + +Distribution-aware prompts. Given the ID class names, we first follow NegLabel [21] to mine the negative labels $\mathcal{V}^{-}$ from extensive corpus databases. Then, instead of hand-crafting the optimal text prompts, we design a learnable text prompt function in the following form: + +$$ +\rho_ {l} (\mathrm {L A B E L}) = [ \boldsymbol {V} ] _ {1} [ \boldsymbol {V} ] _ {2} \dots [ \boldsymbol {V} ] _ {N} [ \mathrm {L A B E L} ], \tag {5} +$$ + +where $[\mathbf{V}]_n\in \mathcal{R}^D (n\in \{1,2\dots ,N\})$ represents the learnable context tokens, and $N$ is the total number of context tokens. + +![](images/2fbc81e46f72d0789c3b2c69b77dcf1abda30276c86d7efaf9bdf791056a7959.jpg) +Fig. 3: Visualization of retrieved images (left), generated images (right), and the statistic of their cosine similarity to the class labels (middle). A larger mean indicates greater image-label consistency, and a larger standard deviation implies more diversity. + +![](images/73ee1aca0decd463cbe771ec4897987c38cf8d11bf3a77536958dd3397b08ae7.jpg) + +![](images/563bec17de7b0a7b879641ef7d362a874bdce252380b25f09cb43591ae9b4823.jpg) + +We assess three prompt construction variations for OOD detection: unified, class-specific, and distribution-aware prompts. The unified prompt employs identical context tokens across all classes, and the class-specific prompt utilizes unique tokens for each class, which are explored in the ID classification [72]. Here, we propose a distribution-aware prompt strategy, which differentiates tokens for ID and OOD classes, as shown in Fig. 2. Our empirical results suggest that the distribution-aware prompt notably enhances the distinction between ID and OOD distributions, as analyzed in Sec. 4.3. + +Automated sample collection with class labels. Traditional prompt learning methods [32, 72] often optimize the prompts with manually collected image samples, which are typically selected from well-constructed benchmark datasets. Such a manual sampling approach is inefficient for negative label-based OOD detection, primarily because mined negative classes are often not included in existing datasets. Moreover, manually collecting and annotating these data is time-consuming and labor-intensive. To mitigate the manual effort and facilitate practical application, we introduce automated sample collection strategies that require only class names, as detailed subsequently. + +We introduce two automated sample collection methods for enhancing OOD detection: text-to-image generation and text-based image retrieval. In the generation approach, class names are fed into pre-trained text-to-image models [35,39] to generate synthetic images with varied appearances using different seeds: + +$$ +\mathcal {X} _ {c o l} ^ {s} = \mathrm {T 2 I} (\mathcal {Y} \cup \mathcal {Y} ^ {-}), \tag {6} +$$ + +where $\mathcal{X}_{col}^{s}$ is the collected set of synthetic images and $\mathrm{T2I}(\cdot)$ is the pre-trained generation model. The performance of OOD detection is closely tied to the fidelity of these generated images, depending on the generative power of the underlying model. By harnessing advanced generative models [35, 39], we are able to produce images that are not only visually convincing but also exhibit strong consistency with the input text. Our empirical findings indicate that these synthetic images can significantly enhance the OOD detection capabilities. + +In addition to generating synthetic images with generation models, we also automate the collection of real images through text-based image retrieval. Specifically, we utilize a large-scale real image dataset (e.g., Laion [42, 43]) in conjunction with pre-trained VLMs, such as CLIP, which align images and text within the same feature space. By conducting a search within this space, we retrieve images that are closest to the class names: + +$$ +\mathcal {X} _ {\text {c o l}} ^ {r} = \operatorname {C l i p - r e t r i e v a l} (\mathcal {Y} \cup \mathcal {Y} ^ {-}, \text {W e b D a t a}), \tag {7} +$$ + +where $\mathcal{X}_{col}^{r}$ is the retrieved set of real images, Clip-retrieval is the clip-retrieval tool [1], which enables efficient image retrieval from millions of samples, and Data is the web-scale real image set. + +We capitalize on the strengths of both synthetic image generation and real image retrieval. Synthetic images ensure high textual fidelity and control but may lack variety, while real images offer diversity but can vary in text consistency, as shown in Fig. 3. Our strategy prioritizes real images with high text-image alignment, quantified by the cosine similarity between image representation and corresponding text label. If real images fall short in text correlation, indicated by cosine similarity below a threshold, synthetic images are used. This hybrid approach is formalized as follows: + +$$ +\mathcal {X} _ {c o l} = \left\{\boldsymbol {x}: \boldsymbol {x} \in \mathcal {X} _ {c o l} ^ {r} \text {i f} \cos \left(f _ {i m g} (\boldsymbol {x}), f _ {t x t} (y)\right) > \kappa \text {e l s e} \boldsymbol {x} \in \mathcal {X} _ {c o l} ^ {s} \right\}, \tag {8} +$$ + +where $\mathcal{X}_{col}$ represents the final collection of images. The decision to use a real image or a synthetic one is based on whether the cosine similarity exceeds the threshold $\kappa$ , which is set as 0.3 by default following [42]. This hybrid approach optimizes the balance between image diversity and text-image consistency, enhancing the robustness of our method. + +Prompt tuning with cross-modal and cross-distribution mixing. Given the training images collected on a per-class basis, the most straightforward learning approach is to use a classification objective, which is commonly achieved with the following cross-entropy loss function: + +$$ +\mathcal {L} = - \frac {1}{| \mathcal {X} _ {c o l} |} \sum_ {\boldsymbol {x} \in \mathcal {X} _ {c o l}} \sum_ {i = 1} ^ {C + M} \boldsymbol {l} _ {i} \log \frac {e ^ {\cos (\boldsymbol {v} , (f _ {t x t} (\rho_ {l} (y))) / \tau}}{\sum_ {i = 1} ^ {C + M} e ^ {\cos (\boldsymbol {v} , \cos (f _ {t x t} (\rho_ {l} (y)))) / \tau}}, \boldsymbol {v} = f _ {i m g} (\boldsymbol {x}) \tag {9} +$$ + +where $y$ is the class name of $\pmb{x}$ . $\pmb{l}_i$ is the $i$ -th entry of $\pmb{l} \in \mathcal{R}^{C + M}$ , the softlabel of sample $\pmb{x}$ . In Sec. 4.3, we will demonstrate that this training objective aligns effectively with our proposed distribution-aware prompt construction and automated sample collection strategies. + +Cross-modal mixing. While the simple classification loss has achieved commendable results, we could further enhance it by addressing potential sample noise and exploring broader data space through carefully designed data mixing strategies. Despite the efforts to collect clean images that match their respective class names, noise is inevitable in the collected images. To mitigate the impact + +of image noise, we introduce a cross-modal mixing strategy to neutralize potential noise in the images. Specifically, for a given training image $\pmb{x}$ of class $y$ , we introduce the refined training feature $\pmb{v}_{cm}$ as follows: + +$$ +\boldsymbol {v} _ {c m} = \mathrm {L} _ {2} \left(\lambda_ {\alpha} f _ {i m g} (\boldsymbol {x}) + (1 - \lambda_ {\alpha}) f _ {t x t} (\rho (y))\right), \tag {10} +$$ + +$$ +\boldsymbol {l} _ {c m} = \boldsymbol {l}, \tag {11} +$$ + +where $\lambda_{\alpha} \sim \mathrm{Beta}(\alpha, \alpha) \in [0,1]$ with $\alpha \in (0,\infty)$ . $\mathrm{L}_2(\cdot)$ represents the $L_2$ normalization along feature dimension and $\rho(\cdot)$ is the hand-crafted text prompts [36]. The softlabel is unchanged after mixing since the image $\mathbf{x}$ and its label $y$ belong to the same class. When $\lambda_{\alpha} = 1$ , we adopt the image features as the training examples, and we use the text features as the training examples when $\lambda_{\alpha} = 0$ . We denote the cross-entropy loss of Eq. (9) with $\pmb{v}_{cm}$ and $\pmb{l}_{cm}$ as $\mathcal{L}_{cm}$ . By mixing textual and visual representations, we aim to create a more robust feature set that can help the model learn to focus on relevant features while disregarding noisy information. This strategy not only reinforces the model's resistance to image noise but also encourages it to leverage complementary information from both modalities. + +Cross-distribution mixing. In addition to using cross-modal data mixing to mitigate image noise, we employ the data mixing strategy to explore a broader feature space. In the vanilla NegLabel approach [21], the selected negative classes are typically far from the ID distribution, leaving a substantial intermediate area between ID and negative regions underutilized. To enhance the utilization of these spaces, we introduce a cross-distribution mixing strategy, which combines features and corresponding labels of ID and negative samples to create new training samples. This mixing strategy effectively bridges the distribution gap, generating a continuous spectrum of features that extend from ID to negative areas. The mixing process can be formalized as follows: + +$$ +\boldsymbol {v} _ {c d} = \operatorname {L 2} \left(\lambda_ {\beta} f _ {i m g} (\boldsymbol {x} _ {i d}) + (1 - \lambda_ {\beta}) f _ {i m g} (\boldsymbol {x} _ {o o d})\right), \tag {12} +$$ + +$$ +\boldsymbol {l} _ {c d} = \lambda_ {\beta} \boldsymbol {l} _ {i d} + (1 - \lambda_ {\beta}) \boldsymbol {l} _ {o o d}, \tag {13} +$$ + +where $\pmb{x}_{id}$ and $\pmb{x}_{ood}$ are collected positive and negative images, and $l_{id}$ and $l_{ood}$ are soft labels of $\pmb{x}_{id}$ and $\pmb{x}_{ood}$ , respectively. $\lambda_{\beta} \sim \mathrm{Beta}(\beta, \beta) \in [0,1]$ with $\beta \in (0,\infty)$ . This method not only allows the model to become more aware of the intermediate space between ID and negative regions, but also encourages it to learn more discriminative features that can better generalize to new and unseen OOD samples. We denote the cross-entropy loss of Eq. (9) with $\pmb{v}_{cd}$ and $l_{cd}$ as $\mathcal{L}_{cd}$ . The final objective is constructed as: + +$$ +\mathcal {L} _ {a l l} = \mathcal {L} + \mathcal {L} _ {c m} + \mathcal {L} _ {c d}. \tag {14} +$$ + +Remarks. Our proposed data mixing strategies are distinct from existing ones [51,61,62,64] in both implementation and objectives. Traditional data mixing strategies typically aim to enhance ID classification performance by mixing randomly selected ID images [51,62]. In contrast, our approach is carefully designed to include multi-modal and multi-distribution data in the mixing process to enhance the OOD detection capability. While there are some methods + +# Algorithm 1 Label-driven Automated Prompt Tuning (LAPT) + +Require: ID label space $\mathcal{V}$ + +1: $\mathcal{V}^{-}\gets \mathrm{NegMine}(\mathcal{V})$ +2: Constructing distribution-aware prompts $\rho_{l}$ with labels $\mathcal{V} \cup \mathcal{V}^{-}$ using Eq. (5) +3: $\mathcal{X}_{col} \gets$ Automated sample collection with labels $\mathcal{Y} \cup \mathcal{Y}^{-}$ using Eq. (8) +4: Prompt tuning with cross-modal and cross-distribution data mixing using Eq. (14) + +5: return Learned prompts $\rho_l$ + +that utilize data mixing for OOD detection, they usually mix ID data to simulate OOD representations [61] or combining manually collected ID and OOD data [64]. These methods, however, do not exploit the textual knowledge and require manually labeled samples. Our cross-modal mixing strategy lowers image noise by blending samples from the same class but across different modalities, creating more robust features. Meanwhile, our cross-distribution mixing scheme goes a step further by mixing automatically collected positive and negative data, allowing us to explore the space that lies between ID and negative regions. This exploration enables our model to better understand and characterize the boundary between ID and OOD, leading to more effective detection of anomalies. Our methods are not just mixing data; they are deliberately designed to align with our label-driven automated prompt learning objective. By addressing noise and exploring intermediate feature space through mixing, we develop a tailored solution that boosts the automated prompt tuning for OOD detection. + +In the test stage, we adopt the OOD score of Eq. (4) following [21]. The overall pipeline of our automated prompt learning is summarized in Algorithm 1. + +# 4 Experiments + +# 4.1 Setup + +Datasets and benchmarks. Our validation primarily uses the ImageNet-1k dataset [4] as ID data, and more evaluations on smaller-scaled ID datasets are provided in the Supplementary Materials. Aligning with standards from prior research [19, 21, 29], we use four diverse datasets of iNaturalist [49], SUN [56], Places [71], and Textures [3] as OOD test sets. Additionally, we assess our method on the OpenOOD benchmark [58, 65], which differentiates OOD datasets into near-OOD (e.g., SSB-hard [50], NINCO [2]) and far-OOD (e.g., iNaturalist [49], Textures [3], OpenImage-O [52]) based on semantic similarity or difficulty. This grouping allows for a comprehensive evaluation of OOD detectors against various OOD sample types. + +In addition to conventional OOD detection that solely considers semantic shifts, we consider a full-spectrum OOD detection problem [60], which additionally accounts for non-semantic covariate shift generalization. An optimal system should not only be capable of identifying semantically shifted OOD samples but should also exhibit robustness to non-semantic covariate shifted OOD samples. + +Table 1: OOD detection results with ID data of ImageNet-1k and four OOD datasets by using the VITB/16 CLIP encoder. + +
MethodsOOD datasets
INaturalistSunPlacesTexturesAverage
AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓
Requires Manually Labeled Training Samples
MSP [15]87.4458.3679.7373.7279.6774.4179.6971.9381.6369.61
ODIN [25]94.6530.2287.1754.0485.5455.0687.8551.6788.8047.75
Energy [26]95.3326.1292.6635.9791.4139.8786.7657.6191.5439.89
GradNorm [18]72.5681.5072.8682.0073.7080.4170.2679.3672.3580.82
ViM [52]93.1632.1987.1954.0183.7560.6787.1853.9487.8250.20
KNN [46]94.5229.1792.6735.6291.0239.6185.6764.3590.9742.19
VOS [7]94.6228.9992.5736.8891.2338.3986.3361.0291.1941.32
NPOS [48]96.1916.5890.4443.7789.4445.2788.8046.1291.2237.93
LSN [34]95.8321.5694.3526.3291.2534.4890.4238.5492.9630.22
LoCoOp [32]96.8616.0595.0723.4491.9832.8790.1942.2893.5228.66
Does Not Require Manually Labeled Training Samples
Mahalanobis [24]55.8999.3359.9499.4165.9698.5464.2398.4661.5098.94
Energy [26]85.0981.0884.2479.0283.3875.0865.5693.6579.5782.21
ZOC [9]86.0987.3081.2081.5183.3973.0676.4698.9081.7985.19
MCM [29]94.5932.2092.2538.8090.3146.2086.1258.5090.8243.93
CLIPN [53]95.2723.9493.9326.1792.2833.4590.9340.8393.1031.10
NegLabel [21]99.491.9195.4920.5391.6435.5990.2243.5694.2125.40
LAPT (Ours)99.631.1696.0119.1292.0133.0191.0640.3294.6823.40
+ +Following the guidelines in [65], we adopt ImageNet-C [14], ImageNet-R [13], and ImageNet-V2 [37] as the covariate-shifted ID data in the full-spectrum OOD detection evaluation. + +Our evaluation metrics include (1) FPR95, the false positive rate of OOD samples with ID detection rate at $95\%$ ; (2) AUROC, the area under the receiver operating characteristic curve; and (3) ID classification accuracy (ID ACC). + +Implementation details. We adopt the visual encoder of VITB/16 pretrained by CLIP [36]. For automated sample collection, we instantiate the text-to-image generation model with SDXL [35] and employ a subset of the Laion400M dataset [42], approximately 60 million samples, as our WebData. Our model training is conducted using the SGD optimizer with a learning rate schedule that follows cosine annealing, starting from an initial learning rate of 1e-2. We set the batch size to 32 and train the model for 10 epochs in all experiments. + +# 4.2 Main Results + +Results on four OOD datasets. As illustrated in Tab. 1, our approach consistently surpasses existing methods. Specifically, we re-implemented traditional methods [7, 15, 18, 25, 26, 46, 48, 52] by fine-tuning CLIP-encoders with labeled training samples from ImageNet in accordance with [48]. Without utilizing any manually annotated data, our method significantly outperforms these methods, corroborating the effectiveness of our proposed VLMs adaptation technique. Notably, our approach exceeds the performance of NegLabel [21], confirming the superiority of learnable prompts over hand-crafted prompts. The OOD samples in these datasets have a relatively large semantic difference from the ID data + +Table 2: OOD detection results on the OpenOOD benchmark, where ImageNet-1k is adopted as ID dataset. Full results are available in the Supplementary Materials. + +
MethodsFPR95 ↓AUROC ↑ACC ↑ ID
Near-OOD Far-OODNear-OOD Far-OOD
Requires Manually Labeled Training Samples
GEN [27]--78.9790.9881.59
AugMix [17] + ReAct [45]--79.9493.7077.63
RMDS [38]--80.0992.6081.14
AugMix [17] + ASH [5]--82.1696.0577.63
Does Not Require Manually Labeled Training Samples
MCM [29]79.0268.5460.1184.7766.28
NegLabel [21]68.1827.3476.9293.3066.82
LAPT (Ours)58.9424.8682.6394.2667.86
+ +(e.g., textures vs. animals), making the distinction relatively easy. On more challenging datasets, where the OOD data are semantically closer to the ID data, our advantages are even more pronounced, as shown in Tabs. 2 and 3. + +Results on OpenOOD benchmark. A more comprehensive evaluation of our method is provided in Tab. 2 on the OpenOOD benchmark [58,65]. Compared to its closest competitor NegLabel, our method not only enhances the OOD detection performance but also improves the ID classification accuracy. In terms of OOD detection, our approach achieves a slight improvement in the relatively easy far-OOD scenarios (e.g., $0.96\%$ AUROC) and a substantial improvement in the more challenging near-OOD settings (e.g., $5.71\%$ AUROC), highlighting the outstanding effectiveness of our method in challenging contexts. The competitors with manually labeled training samples are quoted from [58,65], where models are trained with full ImageNet training data. + +Full-spectrum results on OpenOOD benchmark. In addition to the capabilities of semantic OOD detection and ID sample recognition, we also analyze the generalization ability of our method under covariate shift [68,69,73,74] on the full-spectrum OOD detection task [60] using the OpenOOD benchmark [58,65]. As illustrated in Tab. 3, our method enhances all the three capacities, achieving improved results across the board. Similar to the findings in Tab. 2, the improvement is more pronounced in the challenging near-OOD scenarios. + +# 4.3 Analyses and Discussions + +The analyses are conducted on the OpenOOD dataset with a VITB/16 encoder. + +Prompt construction. We conduct comprehensive analyses on prompt construction as demonstrated in Fig. 4. As shown in Fig. 4a, our proposed distribution-aware prompts are well-matched for OOD detection, outperforming both class-specific and unified prompts. Excessively long prompts slightly impair experimental outcomes, as evidenced in Fig. 4b, possibly due to the increased capacity overfitting noise presented in the training images. Consequently, + +Table 3: Full-spectrum OOD detection results on the OpenOOD benchmark, where ImageNet-1k, ImageNet-C, ImageNet-R, ImageNet-V2 are adopted as ID datasets. + +
MethodsFPR95 ↓AUROC ↑ACC ↑ ID
Near-OOD Far-OODNear-OOD Far-OOD
Requires Manually Labeled Training Samples
DeepAug [13] + SHE [66]83.2668.7068.2778.8557.82
StyAug [11] + GradNorm [18]87.1458.8265.2781.6255.44
AugMix [17] + SHE [66]84.4560.2669.6683.0657.46
ASH [5]93.2759.5660.5286.7554.35
LSA [28]70.5648.0678.2286.8561.74
ISH + SCALE [57]--68.0489.4654.88
Does Not Require Manually Labeled Training Samples
MCM [29]85.3769.8758.9777.1157.69
NegLabel [21]76.2533.3072.7792.0260.11
LAPT (Ours)71.1833.0774.7792.1461.17
+ +![](images/eb3482017246b13ee232cc6c9e99ff8e7c953a5f82a7052bca5c7049fac7d883.jpg) +Fig. 4: Analyses on prompt (a) types, (b) length, (b) initializations, and (d) positions. + +![](images/301e8a53e6506831f7ac50114ec0c9471aea431c0b8ecfc08c2d6d26b2462caf.jpg) + +![](images/ba6e84b44e23c9a9cb63d4d65158fa9ab1a8a971bc689e70da6690f52c55b920.jpg) + +![](images/1cc558c13428cc1d4a50181e84c4114f990fcf9ad0a234d4033c8e8beffbce19.jpg) + +we set the prompt length as $N = 2$ by default. As indicated in Fig. 4c, initializing prompts with manually crafted text introduces textual bias. For instance, 'The nice' yields better results for far-OOD but proves suboptimal for near-OOD, a bias that is perpetuated in our method when learnable prompts are initialized with 'The nice'. We discover that a random initialization, which does not introduce any prior, performs consistently well across different OOD tasks and is thus adopted as the default practice. Finally, in Fig. 4d, we observe that the placement of label tokens has a minimal impact on the results; hence, we by default place the class label at the end of the prompt, as shown in Eq. (5) + +Automated sample collection. As shown in Fig. 5a, the synthetic image generation strategy outperforms realistic image retrieval, possibly because the generator has been exposed to a vast amount of training data (e.g., 5 billion images), which greatly surpasses the scale of our utilized Data (e.g., 60 million images). The best results are obtained by synergistically leveraging both synthetic and real data, confirming their complementary nature. In terms of image generation, employing a stronger image generator can result in better performance, as shown in Fig. 5b. Regarding image retrieval, utilizing a larger- + +![](images/c8f2e4f3702b8e0796779848690d227a0d8b9a5702fe75d7251f738fed5ab07d.jpg) +(a) Collection Methods + +![](images/a29cfcf9dea77bc15da4d5e9ce80b54fb424eeddbe71bf1de87e259aaeb76186.jpg) +(b) Generators + +![](images/d2158e20fce24fcbc8f3b37c396b41cc8746d5f545a37e29a1ee8f37092f802a.jpg) +(c) WebData Scale +Fig. 5: Analyses on (a) image collection methods, (b) text-to-image generators, (c) data scale of the WebData space, and (d) number of training samples per class. + +![](images/7048f777e6c2586e9ec8052afd78ef9402b4aaa1b59c485911efbbe9fd51aadf.jpg) +(d) Sample Number + +scale WebData can yield better outcomes, as presented in Fig. 5c. Finally, we find in Fig. 5d that the model's performance does not continue to increase with the addition of more training samples. This may be because it is difficult to collect a high-quality training set that is semantically consistent with the text and diverse. With the relatively small size of WebData, it is challenging to ensure semantic consistency, and while the semantic consistency of generated images may be good, their diversity is limited, as analyzed in Fig. 3. How to automatically collect high-quality images for effective model training remains an enduring and challenging problem. + +More analyses on stage ablation, the impact of text prompts in data collection, data mixing strategies, hyper-parameters in data mixing, different VLMs, visualization of learned prompts, and time complexity are provided in the Supplementary Materials. + +# 5 Conclusion and Limitations + +We developed a novel label-driven automated prompt tuning (LAPT) approach that autonomously configures effective prompts for OOD detection. Requiring only ID class names, our method can automatically mine OOD class names, collect training images per class, and learn effective prompts, thereby outperforming manually crafted prompts with minimal manual intervention. The automatically learned prompts not only enhanced the distinction between ID and OOD samples but also improved the ID classification accuracy and the strengthened generalization robustness to covariate shifts. + +The efficacy of our approach was highly dependent on the quality of collected training samples. Employing more advanced generation techniques to create a diverse set of synthetic images [8] or expanding the retrieval space [43] to obtain real images that match better the class names could yield improved results. These possibilities present exciting avenues for future research. + +# References + +1. Beaumont, R.: Clip retrieval: Easily compute clip embeddings and build a clip retrieval system with them. https://github.com/rom1504/clip-retrieval (2022) +2. Bitterwolf, J., Mueller, M., Hein, M.: In or out? fixing imagenet out-of-distribution detection evaluation. In: ICML (2023), https://proceedings.mlr.press/v202/bitterwolf23a.html +3. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3606-3613 (2014) +4. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +5. Djurisic, A., Bozanic, N., Ashok, A., Liu, R.: Extremely simple activation shaping for out-of-distribution detection. arXiv preprint arXiv:2209.09858 (2022) +6. Du, X., Gozum, G., Ming, Y., Li, Y.: Siren: Shaping representations for detecting out-of-distribution objects. Advances in Neural Information Processing Systems 35, 20434-20449 (2022) +7. Du, X., Wang, X., Gozum, G., Li, Y.: Unknown-aware object detection: Learning what you don't know from videos in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13678-13688 (2022) +8. Dunlap, L., Umino, A., Zhang, H., Yang, J., Gonzalez, J.E., Darrell, T.: Diversify your vision datasets with automatic diffusion-based augmentation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=9wrYfqdrwk +9. Esmaeilpour, S., Liu, B., Robertson, E., Shu, L.: Zero-shot out-of-distribution detection based on the pre-trained model clip. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 6568-6576 (2022) +10. Fort, S., Ren, J., Lakshminarayanan, B.: Exploring the limits of out-of-distribution detection. Advances in Neural Information Processing Systems 34, 7068-7081 (2021) +1. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231 (2018) +2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +3. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021) +4. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019) +5. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. arXiv preprint arXiv:1610.02136 (2016) +6. Hendrycks, D., Mazeika, M., Dietterich, T.: Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606 (2018) +7. Hendrycks, D., Mu, N., Cubuk, E.D., Zoph, B., Gilmer, J., Lakshminarayanan, B.: Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781 (2019) + +18. Huang, R., Geng, A., Li, Y.: On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems 34, 677-689 (2021) +19. Huang, R., Li, Y.: Mos: Towards scaling out-of-distribution detection for large semantic space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8710-8719 (2021) +20. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021) +21. Jiang, X., Liu, F., Fang, Z., Chen, H., Liu, T., Zheng, F., Han, B.: Negative label guided OOD detection with pretrained vision-language models. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=xU01HXz4an +22. Kong, S., Ramanan, D.: Opengan: Open-set recognition via open data generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 813-822 (2021) +23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012) +24. Lee, K., Lee, K., Lee, H., Shin, J.: A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in neural information processing systems 31 (2018) +25. Liang, S., Li, Y., Srikant, R.: Enhancing the reliability of out-of-distribution image detection in neural networks. arXiv preprint arXiv:1706.02690 (2017) +26. Liu, W., Wang, X., Owens, J., Li, Y.: Energy-based out-of-distribution detection. Advances in neural information processing systems 33, 21464-21475 (2020) +27. Liu, X., Lochman, Y., Zach, C.: Gen: Pushing the limits of softmax-based out-of-distribution detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23946-23955 (2023) +28. Lu, F., Zhu, K., Zheng, K., Zhai, W., Cao, Y.: Likelihood-aware semantic alignment for full-spectrum out-of-distribution detection. arXiv preprint arXiv:2312.01732 (2023) +29. Ming, Y., Cai, Z., Gu, J., Sun, Y., Li, W., Li, Y.: Delving into out-of-distribution detection with vision-language representations. Advances in Neural Information Processing Systems 35, 35087-35102 (2022) +30. Ming, Y., Li, Y.: How does fine-tuning impact out-of-distribution detection for vision-language models? International Journal of Computer Vision 132(2), 596-609 (2024) +31. Ming, Y., Sun, Y., Dia, O., Li, Y.: How to exploit hyperspherical embeddings for out-of-distribution detection? arXiv preprint arXiv:2203.04450 (2022) +32. Miyai, A., Yu, Q., Irie, G., Aizawa, K.: Locoop: Few-shot out-of-distribution detection via prompt learning. Advances in Neural Information Processing Systems 36 (2024) +33. Nguyen, A., Yosinski, J., Clune, J.: Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 427-436 (2015) +34. Nie, J., Zhang, Y., Fang, Z., Liu, T., Han, B., Tian, X.: Out-of-distribution detection with negative prompts. In: The Twelfth International Conference on Learning Representations (2024) + +35. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023) +36. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +37. Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: Do imagenet classifiers generalize to imagenet? In: International conference on machine learning. pp. 5389-5400. PMLR (2019) +38. Ren, J., Fort, S., Liu, J., Roy, A.G., Padhy, S., Lakshminarayanan, B.: A simple fix to mahalanobis distance for improving near-ood detection. arXiv preprint arXiv:2106.09022 (2021) +39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +40. Ryu, S., Koo, S., Yu, H., Lee, G.G.: Out-of-domain detection based on generative adversarial network. In: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. pp. 714-718 (2018) +41. Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Toward open set recognition. IEEE transactions on pattern analysis and machine intelligence 35(7), 1757-1772 (2012) +42. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) +43. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021) +44. Sehwag, V., Chiang, M., Mittal, P.: Ssd: A unified framework for self-supervised outlier detection. arXiv preprint arXiv:2103.12051 (2021) +45. Sun, Y., Guo, C., Li, Y.: React: Out-of-distribution detection with rectified activations. Advances in Neural Information Processing Systems 34, 144-157 (2021) +46. Sun, Y., Ming, Y., Zhu, X., Li, Y.: Out-of-distribution detection with deep nearest neighbors. In: International Conference on Machine Learning. pp. 20827-20840. PMLR (2022) +47. Tack, J., Mo, S., Jeong, J., Shin, J.: Csi: Novelty detection via contrastive learning on distributionally shifted instances. Advances in neural information processing systems 33, 11839-11852 (2020) +48. Tao, L., Du, X., Zhu, X., Li, Y.: Non-parametric outlier synthesis. arXiv preprint arXiv:2303.02966 (2023) +49. Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist species classification and detection dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8769-8778 (2018) +50. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Open-set recognition: A good closed-set classifier is all you need? arXiv preprint arXiv:2110.06207 (2021) +51. Verma, V., Lamb, A., Beckham, C., Najafi, A., Mitliagkas, I., Lopez-Paz, D., Bengio, Y.: Manifold mixup: Better representations by interpolating hidden states. In: International conference on machine learning. pp. 6438-6447. PMLR (2019) + +52. Wang, H., Li, Z., Feng, L., Zhang, W.: Vim: Out-of-distribution with virtual-logit matching. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4921-4930 (2022) +53. Wang, H., Li, Y., Yao, H., Li, X.: Clipn for zero-shot ood detection: Teaching clip to say no. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1802-1812 (2023) +54. Wang, Y., Li, B., Che, T., Zhou, K., Liu, Z., Li, D.: Energy-based open-world uncertainty modeling for confidence calibration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9302-9311 (2021) +55. Wei, H., Xie, R., Cheng, H., Feng, L., An, B., Li, Y.: Mitigating neural network overconfidence with logit normalization. In: International Conference on Machine Learning. pp. 23631-23644. PMLR (2022) +56. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: Sun database: Large-scale scene recognition from abbey to zoo. In: 2010 IEEE computer society conference on computer vision and pattern recognition. pp. 3485-3492. IEEE (2010) +57. Xu, K., Chen, R., Franchi, G., Yao, A.: Scaling for training time and post-hoc out-of-distribution detection enhancement. arXiv preprint arXiv:2310.00227 (2023) +58. Yang, J., Wang, P., Zou, D., Zhou, Z., Ding, K., Peng, W., Wang, H., Chen, G., Li, B., Sun, Y., Du, X., Zhou, K., Zhang, W., Hendrycks, D., Li, Y., Liu, Z.: Openood: Benchmarking generalized out-of-distribution detection (2022) +59. Yang, J., Zhou, K., Li, Y., Liu, Z.: Generalized out-of-distribution detection: A survey. arXiv preprint arXiv:2110.11334 (2021) +60. Yang, J., Zhou, K., Liu, Z.: Full-spectrum out-of-distribution detection. International Journal of Computer Vision pp. 1-16 (2023) +61. Yang, T., Huang, Y., Xie, Y., Liu, J., Wang, S.: Mixood: Improving out-of-distribution detection with enhanced data mixup. ACM Transactions on Multimedia Computing, Communications and Applications 19(5), 1-18 (2023) +62. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017) +63. Zhang, J., Inkawich, N., Chen, Y., Li, H.: Fine-grained out-of-distribution detection with mixup outlier exposure. CoRR (abs/2106.03917) (2021) +64. Zhang, J., Inkawich, N., Linderman, R., Chen, Y., Li, H.: Mixture outlier exposure: Towards out-of-distribution detection in fine-grained environments. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 5531-5540 (January 2023) +65. Zhang, J., Yang, J., Wang, P., Wang, H., Lin, Y., Zhang, H., Sun, Y., Du, X., Zhou, K., Zhang, W., Li, Y., Liu, Z., Chen, Y., Li, H.: Openood v1.5: Enhanced benchmark for out-of-distribution detection. arXiv preprint arXiv:2306.09301 (2023) +66. Zhang, J., Fu, Q., Chen, X., Du, L., Li, Z., Wang, G., Han, S., Zhang, D., et al.: Out-of-distribution detection based on in-distribution data patterns memorization with modern hopfield energy. In: The Eleventh International Conference on Learning Representations (2022) +67. Zhang, R., Fang, R., Zhang, W., Gao, P., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930 (2021) +68. Zhang, Y., Deng, B., Jia, K., Zhang, L.: Label propagation with augmented anchors: A simple semi-supervised learning baseline for unsupervised domain adaptation. In: European Conference on Computer Vision. pp. 781-797. Springer (2020) +69. Zhang, Y., Deng, B., Tang, H., Zhang, L., Jia, K.: Unsupervised multi-class domain adaptation: Theory, algorithms, and practice. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(5), 2775-2792 (2020) + +70. Zhang, Y., Zhu, W., Tang, H., Ma, Z., Zhou, K., Zhang, L.: Dual memory networks: A versatile adaptation approach for vision-language models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 28718-28728 (2024) +71. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence 40(6), 1452-1464 (2017) +72. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) +73. Zhou, Q., Feng, Z., Gu, Q., Pang, J., Cheng, G., Lu, X., Shi, J., Ma, L.: Context-aware mixup for domain adaptive semantic segmentation. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT) 33(2), 804-817 (2023) +74. Zhou, Q., Gu, Q., Pang, J., Lu, X., Ma, L.: Self-adversarial disentangling for specific domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 45(7), 8954-8968 (2023). https://doi.org/10.1109/TPAMI.2023.3238727 \ No newline at end of file diff --git a/2024/LAPT_ Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models/images.zip b/2024/LAPT_ Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..940766b902317aeb290237b02363f01070a319cd --- /dev/null +++ b/2024/LAPT_ Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b15a6c9357a71c09771e5770b8a8c4315f066da1c464799307201095aa7d622 +size 525664 diff --git a/2024/LAPT_ Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models/layout.json b/2024/LAPT_ Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..edcdca6407b25111b24a8302cc293756ddecb941 --- /dev/null +++ b/2024/LAPT_ Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models/layout.json @@ -0,0 +1,10720 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "text", + "content": "LAPT: Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 164, + 167, + 448, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 167, + 448, + 181 + ], + "spans": [ + { + "bbox": [ + 164, + 167, + 448, + 181 + ], + "type": "text", + "content": "Yabin Zhang" + }, + { + "bbox": [ + 164, + 167, + 448, + 181 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 164, + 167, + 448, + 181 + ], + "type": "text", + "content": ", Wenjie Zhu" + }, + { + "bbox": [ + 164, + 167, + 448, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 164, + 167, + 448, + 181 + ], + "type": "text", + "content": ", Chenhang He" + }, + { + "bbox": [ + 164, + 167, + 448, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 164, + 167, + 448, + 181 + ], + "type": "text", + "content": ", and Lei Zhang" + }, + { + "bbox": [ + 164, + 167, + 448, + 181 + ], + "type": "inline_equation", + "content": "^{1,2\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 221, + 189, + 392, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 189, + 392, + 201 + ], + "spans": [ + { + "bbox": [ + 221, + 189, + 392, + 201 + ], + "type": "text", + "content": "1 The Hong Kong Polytechnic University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 249, + 201, + 364, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 201, + 364, + 211 + ], + "spans": [ + { + "bbox": [ + 249, + 201, + 364, + 211 + ], + "type": "text", + "content": "2 OPPO Research Institute" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 221, + 212, + 392, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 212, + 392, + 224 + ], + "spans": [ + { + "bbox": [ + 221, + 212, + 392, + 224 + ], + "type": "text", + "content": "{csybzhang,cslzhang}@comp.polyu.edu.hk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 255, + 455, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 255, + 455, + 531 + ], + "spans": [ + { + "bbox": [ + 159, + 255, + 455, + 531 + ], + "type": "text", + "content": "Abstract. Out-of-distribution (OOD) detection is crucial for model reliability, as it identifies samples from unknown classes and reduces errors due to unexpected inputs. Vision-Language Models (VLMs) such as CLIP are emerging as powerful tools for OOD detection by integrating multi-modal information. However, the practical application of such systems is challenged by manual prompt engineering, which demands domain expertise and is sensitive to linguistic nuances. In this paper, we introduce Label-driven Automated Prompt Tuning (LAPT), a novel approach to OOD detection that reduces the need for manual prompt engineering. We develop distribution-aware prompts with in-distribution (ID) class names and negative labels mined automatically. Training samples linked to these class labels are collected autonomously via image synthesis and retrieval methods, allowing for prompt learning without manual effort. We utilize a simple cross-entropy loss for prompt optimization, with cross-modal and cross-distribution mixing strategies to reduce image noise and explore the intermediate space between distributions, respectively. The LAPT framework operates autonomously, requiring only ID class names as input and eliminating the need for manual intervention. With extensive experiments, LAPT consistently outperforms manually crafted prompts, setting a new standard for OOD detection. Moreover, LAPT not only enhances the distinction between ID and OOD samples, but also improves the ID classification accuracy and strengthens the generalization robustness to covariate shifts, resulting in outstanding performance in challenging full-spectrum OOD detection tasks. Codes are available at https://github.com/YBZh/LAPT." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 540, + 455, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 540, + 455, + 562 + ], + "spans": [ + { + "bbox": [ + 159, + 540, + 455, + 562 + ], + "type": "text", + "content": "Keywords: Out-of-distribution detection " + }, + { + "bbox": [ + 159, + 540, + 455, + 562 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 540, + 455, + 562 + ], + "type": "text", + "content": " Vision-language models " + }, + { + "bbox": [ + 159, + 540, + 455, + 562 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 540, + 455, + 562 + ], + "type": "text", + "content": " Automated prompt tuning " + }, + { + "bbox": [ + 159, + 540, + 455, + 562 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 540, + 455, + 562 + ], + "type": "text", + "content": " Label-driven learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 584, + 230, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 584, + 230, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 584, + 230, + 596 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 609, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 482, + 647 + ], + "type": "text", + "content": "In real-world applications, artificial intelligence (AI) systems often encounter data of unknown classes, known as out-of-distribution (OOD) data. When tackling such OOD data, AI models may erroneously remain overconfident in their" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 173, + 117, + 441, + 248 + ], + "blocks": [ + { + "bbox": [ + 173, + 117, + 441, + 248 + ], + "lines": [ + { + "bbox": [ + 173, + 117, + 441, + 248 + ], + "spans": [ + { + "bbox": [ + 173, + 117, + 441, + 248 + ], + "type": "image", + "image_path": "0efbef56b24ae67870d86badb2c358da0c20e7eb962992ce519e2907ec4db2df.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 261, + 482, + 297 + ], + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 297 + ], + "type": "text", + "content": "Fig. 1: VLMs-based OOD detection systems are sensitive to linguistic nuances of text prompts, where results are reported on the OpenOOD dataset with a ViTB/16 image encoder. A lower FPR95 metric denotes better OOD detection performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 318, + 479, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 479, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 479, + 354 + ], + "type": "text", + "content": "predictions [33, 41], resulting in critical errors and posing security risks. Consequently, it is crucial to recognize OOD data for maintaining the reliability and safety of AI systems in open-world scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 533 + ], + "type": "text", + "content": "Traditional OOD detection methods are primarily focused on image information [15,24,25], neglecting the rich textual knowledge carried by class names. With the rise of vision-language models (VLMs) [20,36], there has been a growing interest in leveraging textual information to facilitate visual tasks [67,70,72], including OOD detection [21,29,30]. For example, Ming et al. [29] compared test images with linguistic descriptions of in-distribution (ID) classes in feature space, leveraging the maximum probability of a scaled softmax function to identify OOD images. While this method achieves notable generalization across various datasets, it engages only a restricted portion of the textual space, thereby not fully exploiting the textual interpretative capabilities of VLMs. Jiang et al. [21] extended it by exploiting larger textual spaces, extracting negative labels from vast text corpora to enhance OOD detection performance significantly. However, the deployment of this system is hindered by the demanding task of prompt engineering, which requires domain-specific knowledge and is sensitive to fine-grained linguistic variations, as illustrated in Fig. 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "To this end, we propose Label-driven Automated Prompt Tuning (LAPT), a novel approach to OOD detection that substantially reduces the need of manual prompt engineering. Specifically, given ID class names and automatically extracted negative labels as in [21], we append distribution-aware prompt tokens to these class labels. These learnable tokens are optimized with minimal human intervention by aggregating training images associated with class labels using pre-trained text-to-image generation models [35, 39] or retrieving real images from web-scale dataset [42, 43]. We employ a simple yet effective cross-entropy loss for prompt optimization, enhanced by cross-modal and cross-distribution data mixing techniques. Cross-modal mixing merges image and text features from the same class, mitigating image noise through multi-modal integration." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "Cross-distribution mixing stochastically blends ID and negative features along with their respective labels, exploring the intermediate space between ID and negative domains. During testing, we adopt the OOD score introduced in [21] to utilize both ID and negative labels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 294 + ], + "type": "text", + "content": "Extensive experiments demonstrate that our method outperforms the manually engineered text prompts by a large margin and achieves new state-of-the-art on various OOD detection benchmarks. Especially, on the challenging near-OD detection task using the ImageNet-1k as the ID dataset, our approach exceeds the best hand-crafted prompts by " + }, + { + "bbox": [ + 130, + 164, + 480, + 294 + ], + "type": "inline_equation", + "content": "10.76\\%" + }, + { + "bbox": [ + 130, + 164, + 480, + 294 + ], + "type": "text", + "content": " in FPR95 and " + }, + { + "bbox": [ + 130, + 164, + 480, + 294 + ], + "type": "inline_equation", + "content": "5.71\\%" + }, + { + "bbox": [ + 130, + 164, + 480, + 294 + ], + "type": "text", + "content": " in AUROC, without any manual label annotations or prompt crafting. Moreover, LAPT not only enhances the distinction between ID and OOD samples, but also improves the ID classification accuracy and strengthens the generalization robustness to covariate shifts. These improvements collectively contribute to the superior performance of LAPT in challenging full-spectrum OOD detection tasks [59, 60]. We summarize our contributions as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 302, + 480, + 491 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 302, + 479, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 302, + 479, + 337 + ], + "spans": [ + { + "bbox": [ + 138, + 302, + 479, + 337 + ], + "type": "text", + "content": "- To circumvent the labor-intensive task of manual prompt engineering for OOD detection with VLMs, we introduce the LAPT methodology, which autonomously learns effective prompts using only ID class names." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 338, + 480, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 338, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 138, + 338, + 480, + 372 + ], + "type": "text", + "content": "- Utilizing the ID class names and automatically mined negative labels, we devise distribution-aware prompts and autonomously gather the corresponding training images for each class via image synthesis and retrieval." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 373, + 480, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 373, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 138, + 373, + 480, + 420 + ], + "type": "text", + "content": "- With the collected training data, we employ a simple yet effective cross-entropy objective for prompt optimization, enhanced by cross-modal data mixing to reduce image noise and cross-distribution data mixing to explore the intermediate space between ID and negative regions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 421, + 480, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 421, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 480, + 491 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that our method significantly outperforms the hand-crafted text prompts and achieves new state-of-the-art results across multiple benchmarks. Moreover, LAPT enhances not only the distinction between ID and OOD samples, but also the ID classification accuracy and the generalization robustness to covariate shifts, achieving outstanding results in full-spectrum OOD detection tasks." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 510, + 237, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 237, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 237, + 521 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 480, + 665 + ], + "type": "text", + "content": "Traditional OOD detection aims to address the severe issue of deep learning models in producing overconfident predictions on OOD data [33, 59], which poses significant challenges to the deployment of deep models in real-world applications. To tackle this problem, researchers have developed various OOD detection methods, including score-based [15, 19, 24-26, 45, 52, 54, 55], distance-based [6, 31, 44, 46-48] and generative-based [22, 40] approaches. Score-based methods, which arguably garner the most attentions, distinguish ID from OOD samples using carefully designed scoring functions, such as confidence-based [15, 25, 45, 52, 55], discriminator-based [22], energy-based [26, 54], and gradient-based [18] scores. In contrast, distance-based methods detect OOD samples by measuring the distance in feature space between the test data and their nearest" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 216, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 216, + 91, + 448, + 102 + ], + "type": "text", + "content": "LAPT for OOD Detection with Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": "ID samples [47] or ID prototypes [48] using metrics like KNN [6,31,46] or Mahalanobis distance [24,44]. While these methods have achieved certain success, they overlook textual information and require training or pre-training on manually labeled ID training data. Moreover, some approaches require manually collecting OOD data for additional training [10,16,63], hindering their real-world applicability. Our method differs from these approaches by effectively leveraging textual information. Utilizing only the ID class names, we mine negative labels, gather training data, and learn effective prompts in an automated fashion, achieving state-of-the-art OOD detection results with minimal human effort." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 477 + ], + "type": "text", + "content": "OOD detection with VLMs is gaining growing attention as the importance of textual information for OOD image detection becomes increasingly recognized [9, 21, 29, 30, 32, 34, 53]. MCM [29] leverages ID class names for effective zero-shot OOD detection, which is further enhanced by NegLabel [21] through the additional use of mined OOD class names. However, as shown in Fig. 1, NegLabel's performance is sensitive to text prompts, limiting its practical application. ZOC [9] employs VLMs to detect OOD samples by learning a captioner that generates candidate OOD labels. Yet, this captioner struggles to generate effective OOD labels for large-scale ID datasets with numerous ID classes. LoCoOp [32] learns ID prompts using few-shot ID samples, regularizing these ID prompts by mined OOD features from image backgrounds. CLIPN [53] and LSN [34] learn text prompts for OOD classes. Our approach differs significantly from them in two aspects. First, we initialize negative prompts with negative labels instead of initializing negative prompts with 'no' and ID labels. Considering CLIP's difficulty in understanding the meaning of 'no' [34], using OOD labels to initialize negative prompts provides an effective starting point to accelerate convergence. Second, CLIPN and LSN respectively require large-scale multi-modal data and manually collected ID samples, leading to either high training costs or the need of manual annotation. In contrast, our method automatically gathers a small amount of positive/negative training data with ID labels, requiring minimal manual effort yet achieving improved performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 491, + 208, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 208, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 208, + 504 + ], + "type": "text", + "content": "3 Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 515, + 238, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 238, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 238, + 528 + ], + "type": "text", + "content": "3.1 Problem Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} = \\{y_1, \\ldots, y_C\\}" + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": " be the image space and ID label space, where " + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": " is a set of class names, e.g., " + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} = \\{cat, dog, \\ldots, bird\\}" + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": " is the number of ID classes. Given the ID random variable " + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{in} \\in \\mathcal{X}" + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": " and the OOD random variable " + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{odd} \\in \\mathcal{X}" + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\pmb{x}^{in}}" + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\pmb{x}^{odd}}" + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": " to denote the ID marginal distribution and OOD marginal distribution, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "OOD detection. In traditional classification tasks, we assume that the test image " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " is sampled from the ID distribution and belongs to a specific ID class, i.e., " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathcal{P}_{\\pmb{x}^{in}}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{V}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " is the label of " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". However, in real-world applications, AI systems often encounter samples from unknown classes, i.e., " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathcal{P}_{\\pmb{x}^{ood}}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "y \\notin \\mathcal{V}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". AI models may misclassify these data into known ID classes with high confidence [33,41], posing critical errors and security risks. To tackle" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 163 + ], + "type": "text", + "content": "this problem, OOD detection is introduced to classify ID samples into correct ID classes and reject OOD samples simultaneously. The ID class classification is achieved with a " + }, + { + "bbox": [ + 131, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 131, + 116, + 479, + 163 + ], + "type": "text", + "content": "-way classifier following traditional classification [12, 23], while the OOD detection is typically achieved with a score function " + }, + { + "bbox": [ + 131, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 131, + 116, + 479, + 163 + ], + "type": "text", + "content": " [24-26]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 189, + 172, + 480, + 185 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 172, + 480, + 185 + ], + "spans": [ + { + "bbox": [ + 189, + 172, + 480, + 185 + ], + "type": "interline_equation", + "content": "G _ {\\gamma} (\\boldsymbol {x}) = \\operatorname {I D}, \\text {i f} S (\\boldsymbol {x}) \\geq \\gamma ; \\quad \\text {o t h e r w i s e}, G _ {\\gamma} (\\boldsymbol {x}) = \\mathrm {O O D}, \\tag {1}", + "image_path": "f8188ca6415604129c68859438eb9a83310005b39bcd4d54aa7bb3adb85c0e71.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 191, + 479, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 191, + 479, + 216 + ], + "spans": [ + { + "bbox": [ + 131, + 191, + 479, + 216 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 191, + 479, + 216 + ], + "type": "inline_equation", + "content": "G_{\\gamma}" + }, + { + "bbox": [ + 131, + 191, + 479, + 216 + ], + "type": "text", + "content": " is the OOD detector with the threshold " + }, + { + "bbox": [ + 131, + 191, + 479, + 216 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 191, + 479, + 216 + ], + "type": "text", + "content": ". In other words, the test data " + }, + { + "bbox": [ + 131, + 191, + 479, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 131, + 191, + 479, + 216 + ], + "type": "text", + "content": " is detected as ID samples if and only if " + }, + { + "bbox": [ + 131, + 191, + 479, + 216 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x}) \\geq \\gamma" + }, + { + "bbox": [ + 131, + 191, + 479, + 216 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 232, + 326, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 232, + 326, + 243 + ], + "spans": [ + { + "bbox": [ + 131, + 232, + 326, + 243 + ], + "type": "text", + "content": "3.2 Reviews on MCM and NegLabel" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "text", + "content": "CLIP-like models. Given the ID test image " + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "text", + "content": " within the label space " + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "text", + "content": ", we extract the image representation " + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\pmb{v} = f_{img}(\\pmb{x}) \\in \\mathcal{R}^D" + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "text", + "content": " and the textual representation " + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "inline_equation", + "content": "C = f_{txt}(\\rho(\\mathcal{Y})) \\in \\mathcal{R}^{C \\times D}" + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "text", + "content": " with the pre-trained CLIP-like encoders, where " + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "text", + "content": " denotes the feature dimension. " + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "inline_equation", + "content": "f_{img}(\\cdot)" + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "inline_equation", + "content": "f_{txt}(\\cdot)" + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "text", + "content": " denote the image and text encoders, respectively. " + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\rho(\\cdot)" + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "text", + "content": " denotes the text prompt function, which is typically manually defined as 'a photo of a " + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "inline_equation", + "content": "<\\text{label}>" + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "text", + "content": ".', where the 'label' token is replaced by the specific class name, e.g., 'cat' or 'dog'. Both " + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "text", + "content": " normalized along the " + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "text", + "content": " dimension. Then the zero-shot classification probability can be achieved using " + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 131, + 249, + 480, + 357 + ], + "type": "text", + "content": " as the classifier:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 249, + 364, + 480, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 364, + 480, + 377 + ], + "spans": [ + { + "bbox": [ + 249, + 364, + 480, + 377 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} = \\operatorname {S o f t m a x} \\left(\\boldsymbol {v} \\boldsymbol {C} ^ {T}\\right) \\in \\mathcal {R} ^ {C}, \\tag {2}", + "image_path": "4477dc9c2a03aa174bc622388daec0bc047611852739a22e6bd4b36c97c59cf4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 384, + 392, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 384, + 392, + 396 + ], + "spans": [ + { + "bbox": [ + 131, + 384, + 392, + 396 + ], + "type": "text", + "content": "where the fixed scaling parameter is omitted for simplicity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 396, + 479, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 396, + 479, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 396, + 479, + 479 + ], + "type": "text", + "content": "MCM. Besides the remarkable classification capabilities, MCM [29] reveals that pre-trained CLIP models also exhibit strong zero-shot OOD detection abilities. Specifically, MCM treats textual representations " + }, + { + "bbox": [ + 131, + 396, + 479, + 479 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 131, + 396, + 479, + 479 + ], + "type": "text", + "content": " as ID prototypes and assesses OOD uncertainty based on the scaled cosine distance between the visual features and the nearest ID prototype. Furthermore, it has been validated that applying a softmax function as a post hoc mechanism enlarges the distinction between ID and OOD samples, leading to the following score function:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 233, + 487, + 480, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 487, + 480, + 517 + ], + "spans": [ + { + "bbox": [ + 233, + 487, + 480, + 517 + ], + "type": "interline_equation", + "content": "S _ {m c m} (\\boldsymbol {v}) = \\max _ {i} \\frac {e ^ {\\cos (\\boldsymbol {v} , \\boldsymbol {c} _ {i}) / \\tau}}{\\sum_ {i = 1} ^ {C} e ^ {\\cos (\\boldsymbol {v} , \\boldsymbol {c} _ {i}) / \\tau}}, \\tag {3}", + "image_path": "1ab7bb42425c956defbb9cba23ee62ffe8307da90f35e7f64eb6fbab68114224.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 522, + 479, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 522, + 479, + 545 + ], + "spans": [ + { + "bbox": [ + 131, + 522, + 479, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 522, + 479, + 545 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 131, + 522, + 479, + 545 + ], + "type": "text", + "content": " measures the cosine similarity, " + }, + { + "bbox": [ + 131, + 522, + 479, + 545 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 131, + 522, + 479, + 545 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 131, + 522, + 479, + 545 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 522, + 479, + 545 + ], + "type": "text", + "content": " -th entry of " + }, + { + "bbox": [ + 131, + 522, + 479, + 545 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 131, + 522, + 479, + 545 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 131, + 522, + 479, + 545 + ], + "type": "inline_equation", + "content": "\\tau >0" + }, + { + "bbox": [ + 131, + 522, + 479, + 545 + ], + "type": "text", + "content": " is the scaling temperature." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 546, + 480, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 546, + 480, + 629 + ], + "spans": [ + { + "bbox": [ + 131, + 546, + 480, + 629 + ], + "type": "text", + "content": "NegLabel. While it is validated in [29] that leveraging the textual information of ID classes enhances the OOD detection capability, the potential contribution of negative labels is overlooked. To fill this gap, NegLabel [21] explores negative class names " + }, + { + "bbox": [ + 131, + 546, + 480, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^{-} = \\{y_{C + 1},\\dots ,y_{C + M}\\}" + }, + { + "bbox": [ + 131, + 546, + 480, + 629 + ], + "type": "text", + "content": " from extensive text corpora, where " + }, + { + "bbox": [ + 131, + 546, + 480, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^{-}\\cap \\mathcal{Y} = \\emptyset" + }, + { + "bbox": [ + 131, + 546, + 480, + 629 + ], + "type": "text", + "content": " . Under the assumption that ID samples are more similar to ID labels and less similar to negative labels than OOD samples, NegLabel employs a scoring function to improve OOD detection:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 187, + 637, + 480, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 637, + 480, + 669 + ], + "spans": [ + { + "bbox": [ + 187, + 637, + 480, + 669 + ], + "type": "interline_equation", + "content": "S _ {N e g L a b e l} (\\boldsymbol {v}) = \\frac {\\sum_ {i = 1} ^ {C} e ^ {\\cos (\\boldsymbol {v} , \\boldsymbol {c} _ {i}) / \\tau}}{\\sum_ {i = 1} ^ {C} e ^ {\\cos (\\boldsymbol {v} , \\boldsymbol {c} _ {i}) / \\tau} + \\sum_ {j = C + 1} ^ {C + M} e ^ {\\cos (\\boldsymbol {v} , \\boldsymbol {c} _ {j}) / \\tau}}. \\tag {4}", + "image_path": "af48e37d0f5456fb64a81bf6b74b0abf126c1a495b976488a5c181091efbf164.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 216, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 216, + 91, + 448, + 102 + ], + "type": "text", + "content": "LAPT for OOD Detection with Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 361, + 265 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 361, + 265 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 361, + 265 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 361, + 265 + ], + "type": "image", + "image_path": "91b5bd354cf89231b19e97f1b85c84e46e1ac22abe9fe8ddbfc2bb010328bfef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 275, + 482, + 300 + ], + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 300 + ], + "type": "text", + "content": "Fig. 2: The overall framework of our LAPT method, where " + }, + { + "bbox": [ + 130, + 275, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{dog} / \\mathbf{v}_{boat}" + }, + { + "bbox": [ + 130, + 275, + 482, + 300 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 275, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{dog} / \\mathbf{c}_{boat}" + }, + { + "bbox": [ + 130, + 275, + 482, + 300 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 275, + 482, + 300 + ], + "type": "inline_equation", + "content": "l_{dog} / l_{boat}" + }, + { + "bbox": [ + 130, + 275, + 482, + 300 + ], + "type": "text", + "content": " are image features, textual features, and soft labels of dog/boat samples." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 367, + 113, + 482, + 259 + ], + "blocks": [ + { + "bbox": [ + 367, + 113, + 482, + 259 + ], + "lines": [ + { + "bbox": [ + 367, + 113, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 367, + 113, + 482, + 259 + ], + "type": "image", + "image_path": "bc3f554681f97bdc63a23f7034badf751fbeb1c41f1a288e910402e571b4f914.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 323, + 480, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 480, + 348 + ], + "type": "text", + "content": "NegLabel performs excellently on various OOD detection benchmarks by utilizing both ID and negative text knowledge." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 369, + 405, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 405, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 405, + 382 + ], + "type": "text", + "content": "3.3 LAPT: Label-driven Automated Prompt Tuning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "text", + "content": "Though NegLabel has achieved a notable success, its performance is highly sensitive to the configuration of text prompts. As illustrated in Fig. 1, given an ID dataset like ImageNet, varying text prompts leads to considerable differences in OOD detection outcomes. For instance, prompts such as 'The nice " + }, + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "text", + "content": " label>" + }, + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "inline_equation", + "content": "^\\prime" + }, + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "text", + "content": " generally improve performance on far-OOD datasets, while 'a photo of a " + }, + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "text", + "content": " label>" + }, + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "inline_equation", + "content": "^\\prime" + }, + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "text", + "content": " shows enhanced effectiveness for near-OOD scenarios. The choice of prompt can lead to over " + }, + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "text", + "content": " fluctuation in the FPR95 metric. Hence, manually setting text prompts for optimal results is a labor-intensive and time-consuming process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 489, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 482, + 537 + ], + "type": "text", + "content": "We aim to streamline OOD detection by automating prompt generation with minimal manual intervention. Our method learns continuous prompts using autocollected samples, needing only ID class names. The automated process is very useful in practical applications. Details of LAPT are provided as follows." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 494, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 494, + 607 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 494, + 607 + ], + "type": "text", + "content": "Distribution-aware prompts. Given the ID class names, we first follow NegLabel [21] to mine the negative labels " + }, + { + "bbox": [ + 130, + 558, + 494, + 607 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{-}" + }, + { + "bbox": [ + 130, + 558, + 494, + 607 + ], + "type": "text", + "content": " from extensive corpus databases. Then, instead of hand-crafting the optimal text prompts, we design a learnable text prompt function in the following form:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 214, + 617, + 481, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 617, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 214, + 617, + 481, + 632 + ], + "type": "interline_equation", + "content": "\\rho_ {l} (\\mathrm {L A B E L}) = [ \\boldsymbol {V} ] _ {1} [ \\boldsymbol {V} ] _ {2} \\dots [ \\boldsymbol {V} ] _ {N} [ \\mathrm {L A B E L} ], \\tag {5}", + "image_path": "b57aa0367e7aee8880a2d895c96e1d10ae9cbe99070a870b8d2010081c624524.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "inline_equation", + "content": "[\\mathbf{V}]_n\\in \\mathcal{R}^D (n\\in \\{1,2\\dots ,N\\})" + }, + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "text", + "content": " represents the learnable context tokens, and " + }, + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 640, + 481, + 665 + ], + "type": "text", + "content": " is the total number of context tokens." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 162, + 116, + 258, + 220 + ], + "blocks": [ + { + "bbox": [ + 162, + 116, + 258, + 220 + ], + "lines": [ + { + "bbox": [ + 162, + 116, + 258, + 220 + ], + "spans": [ + { + "bbox": [ + 162, + 116, + 258, + 220 + ], + "type": "image", + "image_path": "2fbc81e46f72d0789c3b2c69b77dcf1abda30276c86d7efaf9bdf791056a7959.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 232, + 480, + 266 + ], + "lines": [ + { + "bbox": [ + 131, + 232, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 232, + 480, + 266 + ], + "type": "text", + "content": "Fig. 3: Visualization of retrieved images (left), generated images (right), and the statistic of their cosine similarity to the class labels (middle). A larger mean indicates greater image-label consistency, and a larger standard deviation implies more diversity." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 259, + 116, + 362, + 220 + ], + "blocks": [ + { + "bbox": [ + 259, + 116, + 362, + 220 + ], + "lines": [ + { + "bbox": [ + 259, + 116, + 362, + 220 + ], + "spans": [ + { + "bbox": [ + 259, + 116, + 362, + 220 + ], + "type": "image", + "image_path": "73ee1aca0decd463cbe771ec4897987c38cf8d11bf3a77536958dd3397b08ae7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 364, + 117, + 452, + 221 + ], + "blocks": [ + { + "bbox": [ + 364, + 117, + 452, + 221 + ], + "lines": [ + { + "bbox": [ + 364, + 117, + 452, + 221 + ], + "spans": [ + { + "bbox": [ + 364, + 117, + 452, + 221 + ], + "type": "image", + "image_path": "563bec17de7b0a7b879641ef7d362a874bdce252380b25f09cb43591ae9b4823.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 287, + 480, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 480, + 383 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 480, + 383 + ], + "type": "text", + "content": "We assess three prompt construction variations for OOD detection: unified, class-specific, and distribution-aware prompts. The unified prompt employs identical context tokens across all classes, and the class-specific prompt utilizes unique tokens for each class, which are explored in the ID classification [72]. Here, we propose a distribution-aware prompt strategy, which differentiates tokens for ID and OOD classes, as shown in Fig. 2. Our empirical results suggest that the distribution-aware prompt notably enhances the distinction between ID and OOD distributions, as analyzed in Sec. 4.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 399, + 480, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 480, + 506 + ], + "type": "text", + "content": "Automated sample collection with class labels. Traditional prompt learning methods [32, 72] often optimize the prompts with manually collected image samples, which are typically selected from well-constructed benchmark datasets. Such a manual sampling approach is inefficient for negative label-based OOD detection, primarily because mined negative classes are often not included in existing datasets. Moreover, manually collecting and annotating these data is time-consuming and labor-intensive. To mitigate the manual effort and facilitate practical application, we introduce automated sample collection strategies that require only class names, as detailed subsequently." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 506, + 480, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 480, + 555 + ], + "type": "text", + "content": "We introduce two automated sample collection methods for enhancing OOD detection: text-to-image generation and text-based image retrieval. In the generation approach, class names are fed into pre-trained text-to-image models [35,39] to generate synthetic images with varied appearances using different seeds:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 259, + 562, + 480, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 562, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 259, + 562, + 480, + 575 + ], + "type": "interline_equation", + "content": "\\mathcal {X} _ {c o l} ^ {s} = \\mathrm {T 2 I} (\\mathcal {Y} \\cup \\mathcal {Y} ^ {-}), \\tag {6}", + "image_path": "4f461dab3249690a338d6caaa0eb4a90a18e7b17519c8072dfd734cdc1e64711.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{col}^{s}" + }, + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "type": "text", + "content": " is the collected set of synthetic images and " + }, + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{T2I}(\\cdot)" + }, + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "type": "text", + "content": " is the pre-trained generation model. The performance of OOD detection is closely tied to the fidelity of these generated images, depending on the generative power of the underlying model. By harnessing advanced generative models [35, 39], we are able to produce images that are not only visually convincing but also exhibit strong consistency with the input text. Our empirical findings indicate that these synthetic images can significantly enhance the OOD detection capabilities." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 216, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 216, + 91, + 448, + 102 + ], + "type": "text", + "content": "LAPT for OOD Detection with Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": "In addition to generating synthetic images with generation models, we also automate the collection of real images through text-based image retrieval. Specifically, we utilize a large-scale real image dataset (e.g., Laion [42, 43]) in conjunction with pre-trained VLMs, such as CLIP, which align images and text within the same feature space. By conducting a search within this space, we retrieve images that are closest to the class names:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 218, + 198, + 481, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 198, + 481, + 211 + ], + "spans": [ + { + "bbox": [ + 218, + 198, + 481, + 211 + ], + "type": "interline_equation", + "content": "\\mathcal {X} _ {\\text {c o l}} ^ {r} = \\operatorname {C l i p - r e t r i e v a l} (\\mathcal {Y} \\cup \\mathcal {Y} ^ {-}, \\text {W e b D a t a}), \\tag {7}", + "image_path": "6362b61ee7976fcac2a260a8fbe251a09c94d1e9eca0305773f5de41c0cb7cf1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 218, + 481, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 218, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 481, + 255 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 218, + 481, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{col}^{r}" + }, + { + "bbox": [ + 130, + 218, + 481, + 255 + ], + "type": "text", + "content": " is the retrieved set of real images, Clip-retrieval is the clip-retrieval tool [1], which enables efficient image retrieval from millions of samples, and Data is the web-scale real image set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 255, + 482, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 350 + ], + "type": "text", + "content": "We capitalize on the strengths of both synthetic image generation and real image retrieval. Synthetic images ensure high textual fidelity and control but may lack variety, while real images offer diversity but can vary in text consistency, as shown in Fig. 3. Our strategy prioritizes real images with high text-image alignment, quantified by the cosine similarity between image representation and corresponding text label. If real images fall short in text correlation, indicated by cosine similarity below a threshold, synthetic images are used. This hybrid approach is formalized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 168, + 360, + 481, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 360, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 168, + 360, + 481, + 373 + ], + "type": "interline_equation", + "content": "\\mathcal {X} _ {c o l} = \\left\\{\\boldsymbol {x}: \\boldsymbol {x} \\in \\mathcal {X} _ {c o l} ^ {r} \\text {i f} \\cos \\left(f _ {i m g} (\\boldsymbol {x}), f _ {t x t} (y)\\right) > \\kappa \\text {e l s e} \\boldsymbol {x} \\in \\mathcal {X} _ {c o l} ^ {s} \\right\\}, \\tag {8}", + "image_path": "80f8c30d845d91dca3424fc34099f234ced5f792759dd801d950db110f42fefc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 380, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 380, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{col}" + }, + { + "bbox": [ + 130, + 380, + 482, + 441 + ], + "type": "text", + "content": " represents the final collection of images. The decision to use a real image or a synthetic one is based on whether the cosine similarity exceeds the threshold " + }, + { + "bbox": [ + 130, + 380, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 130, + 380, + 482, + 441 + ], + "type": "text", + "content": ", which is set as 0.3 by default following [42]. This hybrid approach optimizes the balance between image diversity and text-image consistency, enhancing the robustness of our method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 458, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 458, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 131, + 458, + 482, + 507 + ], + "type": "text", + "content": "Prompt tuning with cross-modal and cross-distribution mixing. Given the training images collected on a per-class basis, the most straightforward learning approach is to use a classification objective, which is commonly achieved with the following cross-entropy loss function:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 145, + 515, + 481, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 515, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 145, + 515, + 481, + 548 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = - \\frac {1}{| \\mathcal {X} _ {c o l} |} \\sum_ {\\boldsymbol {x} \\in \\mathcal {X} _ {c o l}} \\sum_ {i = 1} ^ {C + M} \\boldsymbol {l} _ {i} \\log \\frac {e ^ {\\cos (\\boldsymbol {v} , (f _ {t x t} (\\rho_ {l} (y))) / \\tau}}{\\sum_ {i = 1} ^ {C + M} e ^ {\\cos (\\boldsymbol {v} , \\cos (f _ {t x t} (\\rho_ {l} (y)))) / \\tau}}, \\boldsymbol {v} = f _ {i m g} (\\boldsymbol {x}) \\tag {9}", + "image_path": "8e289e5176fd619193fb99d2e9b59efc7b9250972f5b4a5383d061e6a17169b5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 557, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 557, + 481, + 605 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 557, + 481, + 605 + ], + "type": "text", + "content": " is the class name of " + }, + { + "bbox": [ + 130, + 557, + 481, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 557, + 481, + 605 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 557, + 481, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{l}_i" + }, + { + "bbox": [ + 130, + 557, + 481, + 605 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 557, + 481, + 605 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 557, + 481, + 605 + ], + "type": "text", + "content": "-th entry of " + }, + { + "bbox": [ + 130, + 557, + 481, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{l} \\in \\mathcal{R}^{C + M}" + }, + { + "bbox": [ + 130, + 557, + 481, + 605 + ], + "type": "text", + "content": ", the softlabel of sample " + }, + { + "bbox": [ + 130, + 557, + 481, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 557, + 481, + 605 + ], + "type": "text", + "content": ". In Sec. 4.3, we will demonstrate that this training objective aligns effectively with our proposed distribution-aware prompt construction and automated sample collection strategies." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "Cross-modal mixing. While the simple classification loss has achieved commendable results, we could further enhance it by addressing potential sample noise and exploring broader data space through carefully designed data mixing strategies. Despite the efforts to collect clean images that match their respective class names, noise is inevitable in the collected images. To mitigate the impact" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "of image noise, we introduce a cross-modal mixing strategy to neutralize potential noise in the images. Specifically, for a given training image " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " of class " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ", we introduce the refined training feature " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{cm}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 211, + 159, + 480, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 159, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 211, + 159, + 480, + 171 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {c m} = \\mathrm {L} _ {2} \\left(\\lambda_ {\\alpha} f _ {i m g} (\\boldsymbol {x}) + (1 - \\lambda_ {\\alpha}) f _ {t x t} (\\rho (y))\\right), \\tag {10}", + "image_path": "8f25828ea03a92baed5096be5ad0ed279881d9c1bd63cdb144f2252ef83c2b55.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 212, + 175, + 480, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 175, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 212, + 175, + 480, + 186 + ], + "type": "interline_equation", + "content": "\\boldsymbol {l} _ {c m} = \\boldsymbol {l}, \\tag {11}", + "image_path": "2ec673f90e1f8d54f3c736b14851b796249f3f745d9b5381d74cd9ce77f28815.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "inline_equation", + "content": "\\lambda_{\\alpha} \\sim \\mathrm{Beta}(\\alpha, \\alpha) \\in [0,1]" + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "inline_equation", + "content": "\\alpha \\in (0,\\infty)" + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_2(\\cdot)" + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "text", + "content": " normalization along feature dimension and " + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "inline_equation", + "content": "\\rho(\\cdot)" + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "text", + "content": " is the hand-crafted text prompts [36]. The softlabel is unchanged after mixing since the image " + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "text", + "content": " and its label " + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "text", + "content": " belong to the same class. When " + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "inline_equation", + "content": "\\lambda_{\\alpha} = 1" + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "text", + "content": ", we adopt the image features as the training examples, and we use the text features as the training examples when " + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "inline_equation", + "content": "\\lambda_{\\alpha} = 0" + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "text", + "content": ". We denote the cross-entropy loss of Eq. (9) with " + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{cm}" + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "inline_equation", + "content": "\\pmb{l}_{cm}" + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cm}" + }, + { + "bbox": [ + 130, + 192, + 480, + 323 + ], + "type": "text", + "content": ". By mixing textual and visual representations, we aim to create a more robust feature set that can help the model learn to focus on relevant features while disregarding noisy information. This strategy not only reinforces the model's resistance to image noise but also encourages it to leverage complementary information from both modalities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 324, + 480, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 480, + 443 + ], + "type": "text", + "content": "Cross-distribution mixing. In addition to using cross-modal data mixing to mitigate image noise, we employ the data mixing strategy to explore a broader feature space. In the vanilla NegLabel approach [21], the selected negative classes are typically far from the ID distribution, leaving a substantial intermediate area between ID and negative regions underutilized. To enhance the utilization of these spaces, we introduce a cross-distribution mixing strategy, which combines features and corresponding labels of ID and negative samples to create new training samples. This mixing strategy effectively bridges the distribution gap, generating a continuous spectrum of features that extend from ID to negative areas. The mixing process can be formalized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 207, + 451, + 480, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 451, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 207, + 451, + 480, + 464 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {c d} = \\operatorname {L 2} \\left(\\lambda_ {\\beta} f _ {i m g} (\\boldsymbol {x} _ {i d}) + (1 - \\lambda_ {\\beta}) f _ {i m g} (\\boldsymbol {x} _ {o o d})\\right), \\tag {12}", + "image_path": "da05df4a4995e51e056c646071278054e72af34fb2d4f9be76ebec12b76231f2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 209, + 466, + 480, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 466, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 209, + 466, + 480, + 478 + ], + "type": "interline_equation", + "content": "\\boldsymbol {l} _ {c d} = \\lambda_ {\\beta} \\boldsymbol {l} _ {i d} + (1 - \\lambda_ {\\beta}) \\boldsymbol {l} _ {o o d}, \\tag {13}", + "image_path": "2af6efbe8a5de6ff520ef76a2aeb6ad78818ce1fbca59e74e204f9c472566b62.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{id}" + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{ood}" + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "text", + "content": " are collected positive and negative images, and " + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "inline_equation", + "content": "l_{id}" + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "inline_equation", + "content": "l_{ood}" + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "text", + "content": " are soft labels of " + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{id}" + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{ood}" + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "inline_equation", + "content": "\\lambda_{\\beta} \\sim \\mathrm{Beta}(\\beta, \\beta) \\in [0,1]" + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "inline_equation", + "content": "\\beta \\in (0,\\infty)" + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "text", + "content": ". This method not only allows the model to become more aware of the intermediate space between ID and negative regions, but also encourages it to learn more discriminative features that can better generalize to new and unseen OOD samples. We denote the cross-entropy loss of Eq. (9) with " + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{cd}" + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "inline_equation", + "content": "l_{cd}" + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cd}" + }, + { + "bbox": [ + 130, + 484, + 480, + 567 + ], + "type": "text", + "content": ". The final objective is constructed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 258, + 575, + 480, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 575, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 258, + 575, + 480, + 587 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a l l} = \\mathcal {L} + \\mathcal {L} _ {c m} + \\mathcal {L} _ {c d}. \\tag {14}", + "image_path": "3709cce05155901c2b7a6acf3b8c0543b128198ef31f16b3041333019f4f910e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "content": "Remarks. Our proposed data mixing strategies are distinct from existing ones [51,61,62,64] in both implementation and objectives. Traditional data mixing strategies typically aim to enhance ID classification performance by mixing randomly selected ID images [51,62]. In contrast, our approach is carefully designed to include multi-modal and multi-distribution data in the mixing process to enhance the OOD detection capability. While there are some methods" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 216, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 216, + 91, + 448, + 102 + ], + "type": "text", + "content": "LAPT for OOD Detection with Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 416, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 416, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 416, + 129 + ], + "type": "text", + "content": "Algorithm 1 Label-driven Automated Prompt Tuning (LAPT)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 132, + 247, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 132, + 247, + 143 + ], + "spans": [ + { + "bbox": [ + 133, + 132, + 247, + 143 + ], + "type": "text", + "content": "Require: ID label space " + }, + { + "bbox": [ + 133, + 132, + 247, + 143 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 144, + 479, + 187 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 136, + 144, + 228, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 144, + 228, + 154 + ], + "spans": [ + { + "bbox": [ + 136, + 144, + 228, + 154 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 136, + 144, + 228, + 154 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{-}\\gets \\mathrm{NegMine}(\\mathcal{V})" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 154, + 462, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 154, + 462, + 165 + ], + "spans": [ + { + "bbox": [ + 136, + 154, + 462, + 165 + ], + "type": "text", + "content": "2: Constructing distribution-aware prompts " + }, + { + "bbox": [ + 136, + 154, + 462, + 165 + ], + "type": "inline_equation", + "content": "\\rho_{l}" + }, + { + "bbox": [ + 136, + 154, + 462, + 165 + ], + "type": "text", + "content": " with labels " + }, + { + "bbox": [ + 136, + 154, + 462, + 165 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\cup \\mathcal{V}^{-}" + }, + { + "bbox": [ + 136, + 154, + 462, + 165 + ], + "type": "text", + "content": " using Eq. (5)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 165, + 434, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 165, + 434, + 175 + ], + "spans": [ + { + "bbox": [ + 136, + 165, + 434, + 175 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 136, + 165, + 434, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{col} \\gets" + }, + { + "bbox": [ + 136, + 165, + 434, + 175 + ], + "type": "text", + "content": " Automated sample collection with labels " + }, + { + "bbox": [ + 136, + 165, + 434, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} \\cup \\mathcal{Y}^{-}" + }, + { + "bbox": [ + 136, + 165, + 434, + 175 + ], + "type": "text", + "content": " using Eq. (8)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 175, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 175, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 136, + 175, + 479, + 187 + ], + "type": "text", + "content": "4: Prompt tuning with cross-modal and cross-distribution data mixing using Eq. (14)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 135, + 198, + 265, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 198, + 265, + 209 + ], + "spans": [ + { + "bbox": [ + 135, + 198, + 265, + 209 + ], + "type": "text", + "content": "5: return Learned prompts " + }, + { + "bbox": [ + 135, + 198, + 265, + 209 + ], + "type": "inline_equation", + "content": "\\rho_l" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 234, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 234, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 482, + 402 + ], + "type": "text", + "content": "that utilize data mixing for OOD detection, they usually mix ID data to simulate OOD representations [61] or combining manually collected ID and OOD data [64]. These methods, however, do not exploit the textual knowledge and require manually labeled samples. Our cross-modal mixing strategy lowers image noise by blending samples from the same class but across different modalities, creating more robust features. Meanwhile, our cross-distribution mixing scheme goes a step further by mixing automatically collected positive and negative data, allowing us to explore the space that lies between ID and negative regions. This exploration enables our model to better understand and characterize the boundary between ID and OOD, leading to more effective detection of anomalies. Our methods are not just mixing data; they are deliberately designed to align with our label-driven automated prompt learning objective. By addressing noise and exploring intermediate feature space through mixing, we develop a tailored solution that boosts the automated prompt tuning for OOD detection." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 402, + 481, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 481, + 426 + ], + "type": "text", + "content": "In the test stage, we adopt the OOD score of Eq. (4) following [21]. The overall pipeline of our automated prompt learning is summarized in Algorithm 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 443, + 230, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 443, + 230, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 443, + 230, + 456 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 190, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 190, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 190, + 479 + ], + "type": "text", + "content": "4.1 Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "type": "text", + "content": "Datasets and benchmarks. Our validation primarily uses the ImageNet-1k dataset [4] as ID data, and more evaluations on smaller-scaled ID datasets are provided in the Supplementary Materials. Aligning with standards from prior research [19, 21, 29], we use four diverse datasets of iNaturalist [49], SUN [56], Places [71], and Textures [3] as OOD test sets. Additionally, we assess our method on the OpenOOD benchmark [58, 65], which differentiates OOD datasets into near-OOD (e.g., SSB-hard [50], NINCO [2]) and far-OOD (e.g., iNaturalist [49], Textures [3], OpenImage-O [52]) based on semantic similarity or difficulty. This grouping allows for a comprehensive evaluation of OOD detectors against various OOD sample types." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "In addition to conventional OOD detection that solely considers semantic shifts, we consider a full-spectrum OOD detection problem [60], which additionally accounts for non-semantic covariate shift generalization. An optimal system should not only be capable of identifying semantically shifted OOD samples but should also exhibit robustness to non-semantic covariate shifted OOD samples." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 146, + 485, + 340 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 1: OOD detection results with ID data of ImageNet-1k and four OOD datasets by using the VITB/16 CLIP encoder." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 146, + 485, + 340 + ], + "lines": [ + { + "bbox": [ + 132, + 146, + 485, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 485, + 340 + ], + "type": "table", + "html": "
MethodsOOD datasets
INaturalistSunPlacesTexturesAverage
AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓
Requires Manually Labeled Training Samples
MSP [15]87.4458.3679.7373.7279.6774.4179.6971.9381.6369.61
ODIN [25]94.6530.2287.1754.0485.5455.0687.8551.6788.8047.75
Energy [26]95.3326.1292.6635.9791.4139.8786.7657.6191.5439.89
GradNorm [18]72.5681.5072.8682.0073.7080.4170.2679.3672.3580.82
ViM [52]93.1632.1987.1954.0183.7560.6787.1853.9487.8250.20
KNN [46]94.5229.1792.6735.6291.0239.6185.6764.3590.9742.19
VOS [7]94.6228.9992.5736.8891.2338.3986.3361.0291.1941.32
NPOS [48]96.1916.5890.4443.7789.4445.2788.8046.1291.2237.93
LSN [34]95.8321.5694.3526.3291.2534.4890.4238.5492.9630.22
LoCoOp [32]96.8616.0595.0723.4491.9832.8790.1942.2893.5228.66
Does Not Require Manually Labeled Training Samples
Mahalanobis [24]55.8999.3359.9499.4165.9698.5464.2398.4661.5098.94
Energy [26]85.0981.0884.2479.0283.3875.0865.5693.6579.5782.21
ZOC [9]86.0987.3081.2081.5183.3973.0676.4698.9081.7985.19
MCM [29]94.5932.2092.2538.8090.3146.2086.1258.5090.8243.93
CLIPN [53]95.2723.9493.9326.1792.2833.4590.9340.8393.1031.10
NegLabel [21]99.491.9195.4920.5391.6435.5990.2243.5694.2125.40
LAPT (Ours)99.631.1696.0119.1292.0133.0191.0640.3294.6823.40
", + "image_path": "24bc101485a7f7339b86a906df9be9c43db6298d3cc9a648e4be6ff4477988a0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 363, + 480, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 480, + 398 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 480, + 398 + ], + "type": "text", + "content": "Following the guidelines in [65], we adopt ImageNet-C [14], ImageNet-R [13], and ImageNet-V2 [37] as the covariate-shifted ID data in the full-spectrum OOD detection evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 399, + 480, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 399, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 480, + 435 + ], + "type": "text", + "content": "Our evaluation metrics include (1) FPR95, the false positive rate of OOD samples with ID detection rate at " + }, + { + "bbox": [ + 132, + 399, + 480, + 435 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 132, + 399, + 480, + 435 + ], + "type": "text", + "content": "; (2) AUROC, the area under the receiver operating characteristic curve; and (3) ID classification accuracy (ID ACC)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 436, + 481, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 481, + 518 + ], + "type": "text", + "content": "Implementation details. We adopt the visual encoder of VITB/16 pretrained by CLIP [36]. For automated sample collection, we instantiate the text-to-image generation model with SDXL [35] and employ a subset of the Laion400M dataset [42], approximately 60 million samples, as our WebData. Our model training is conducted using the SGD optimizer with a learning rate schedule that follows cosine annealing, starting from an initial learning rate of 1e-2. We set the batch size to 32 and train the model for 10 epochs in all experiments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 537, + 227, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 537, + 227, + 548 + ], + "spans": [ + { + "bbox": [ + 133, + 537, + 227, + 548 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 558, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 481, + 665 + ], + "type": "text", + "content": "Results on four OOD datasets. As illustrated in Tab. 1, our approach consistently surpasses existing methods. Specifically, we re-implemented traditional methods [7, 15, 18, 25, 26, 46, 48, 52] by fine-tuning CLIP-encoders with labeled training samples from ImageNet in accordance with [48]. Without utilizing any manually annotated data, our method significantly outperforms these methods, corroborating the effectiveness of our proposed VLMs adaptation technique. Notably, our approach exceeds the performance of NegLabel [21], confirming the superiority of learnable prompts over hand-crafted prompts. The OOD samples in these datasets have a relatively large semantic difference from the ID data" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "text", + "content": "LAPT for OOD Detection with Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 146, + 466, + 286 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: OOD detection results on the OpenOOD benchmark, where ImageNet-1k is adopted as ID dataset. Full results are available in the Supplementary Materials." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 146, + 466, + 286 + ], + "lines": [ + { + "bbox": [ + 148, + 146, + 466, + 286 + ], + "spans": [ + { + "bbox": [ + 148, + 146, + 466, + 286 + ], + "type": "table", + "html": "
MethodsFPR95 ↓AUROC ↑ACC ↑ ID
Near-OOD Far-OODNear-OOD Far-OOD
Requires Manually Labeled Training Samples
GEN [27]--78.9790.9881.59
AugMix [17] + ReAct [45]--79.9493.7077.63
RMDS [38]--80.0992.6081.14
AugMix [17] + ASH [5]--82.1696.0577.63
Does Not Require Manually Labeled Training Samples
MCM [29]79.0268.5460.1184.7766.28
NegLabel [21]68.1827.3476.9293.3066.82
LAPT (Ours)58.9424.8682.6394.2667.86
", + "image_path": "02b2a132d0ae1cc7caa6abf8f849d869699e94778488593d033f5b66082dc293.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 307, + 480, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 480, + 342 + ], + "type": "text", + "content": "(e.g., textures vs. animals), making the distinction relatively easy. On more challenging datasets, where the OOD data are semantically closer to the ID data, our advantages are even more pronounced, as shown in Tabs. 2 and 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 481, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 481, + 462 + ], + "type": "text", + "content": "Results on OpenOOD benchmark. A more comprehensive evaluation of our method is provided in Tab. 2 on the OpenOOD benchmark [58,65]. Compared to its closest competitor NegLabel, our method not only enhances the OOD detection performance but also improves the ID classification accuracy. In terms of OOD detection, our approach achieves a slight improvement in the relatively easy far-OOD scenarios (e.g., " + }, + { + "bbox": [ + 130, + 342, + 481, + 462 + ], + "type": "inline_equation", + "content": "0.96\\%" + }, + { + "bbox": [ + 130, + 342, + 481, + 462 + ], + "type": "text", + "content": " AUROC) and a substantial improvement in the more challenging near-OOD settings (e.g., " + }, + { + "bbox": [ + 130, + 342, + 481, + 462 + ], + "type": "inline_equation", + "content": "5.71\\%" + }, + { + "bbox": [ + 130, + 342, + 481, + 462 + ], + "type": "text", + "content": " AUROC), highlighting the outstanding effectiveness of our method in challenging contexts. The competitors with manually labeled training samples are quoted from [58,65], where models are trained with full ImageNet training data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 463, + 481, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 481, + 546 + ], + "type": "text", + "content": "Full-spectrum results on OpenOOD benchmark. In addition to the capabilities of semantic OOD detection and ID sample recognition, we also analyze the generalization ability of our method under covariate shift [68,69,73,74] on the full-spectrum OOD detection task [60] using the OpenOOD benchmark [58,65]. As illustrated in Tab. 3, our method enhances all the three capacities, achieving improved results across the board. Similar to the findings in Tab. 2, the improvement is more pronounced in the challenging near-OOD scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 563, + 289, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 289, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 289, + 574 + ], + "type": "text", + "content": "4.3 Analyses and Discussions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 581, + 479, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 479, + 593 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 479, + 593 + ], + "type": "text", + "content": "The analyses are conducted on the OpenOOD dataset with a VITB/16 encoder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "content": "Prompt construction. We conduct comprehensive analyses on prompt construction as demonstrated in Fig. 4. As shown in Fig. 4a, our proposed distribution-aware prompts are well-matched for OOD detection, outperforming both class-specific and unified prompts. Excessively long prompts slightly impair experimental outcomes, as evidenced in Fig. 4b, possibly due to the increased capacity overfitting noise presented in the training images. Consequently," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 146, + 474, + 309 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 3: Full-spectrum OOD detection results on the OpenOOD benchmark, where ImageNet-1k, ImageNet-C, ImageNet-R, ImageNet-V2 are adopted as ID datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 146, + 474, + 309 + ], + "lines": [ + { + "bbox": [ + 140, + 146, + 474, + 309 + ], + "spans": [ + { + "bbox": [ + 140, + 146, + 474, + 309 + ], + "type": "table", + "html": "
MethodsFPR95 ↓AUROC ↑ACC ↑ ID
Near-OOD Far-OODNear-OOD Far-OOD
Requires Manually Labeled Training Samples
DeepAug [13] + SHE [66]83.2668.7068.2778.8557.82
StyAug [11] + GradNorm [18]87.1458.8265.2781.6255.44
AugMix [17] + SHE [66]84.4560.2669.6683.0657.46
ASH [5]93.2759.5660.5286.7554.35
LSA [28]70.5648.0678.2286.8561.74
ISH + SCALE [57]--68.0489.4654.88
Does Not Require Manually Labeled Training Samples
MCM [29]85.3769.8758.9777.1157.69
NegLabel [21]76.2533.3072.7792.0260.11
LAPT (Ours)71.1833.0774.7792.1461.17
", + "image_path": "f1f53d126a5c627c116cc3826dd4ee419ae329d1fbcd1e9526598ff181881bed.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 320, + 218, + 416 + ], + "blocks": [ + { + "bbox": [ + 133, + 320, + 218, + 416 + ], + "lines": [ + { + "bbox": [ + 133, + 320, + 218, + 416 + ], + "spans": [ + { + "bbox": [ + 133, + 320, + 218, + 416 + ], + "type": "image", + "image_path": "eb3482017246b13ee232cc6c9e99ff8e7c953a5f82a7052bca5c7049fac7d883.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 424, + 480, + 436 + ], + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 436 + ], + "type": "text", + "content": "Fig. 4: Analyses on prompt (a) types, (b) length, (b) initializations, and (d) positions." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 220, + 319, + 306, + 417 + ], + "blocks": [ + { + "bbox": [ + 220, + 319, + 306, + 417 + ], + "lines": [ + { + "bbox": [ + 220, + 319, + 306, + 417 + ], + "spans": [ + { + "bbox": [ + 220, + 319, + 306, + 417 + ], + "type": "image", + "image_path": "301e8a53e6506831f7ac50114ec0c9471aea431c0b8ecfc08c2d6d26b2462caf.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 319, + 392, + 416 + ], + "blocks": [ + { + "bbox": [ + 309, + 319, + 392, + 416 + ], + "lines": [ + { + "bbox": [ + 309, + 319, + 392, + 416 + ], + "spans": [ + { + "bbox": [ + 309, + 319, + 392, + 416 + ], + "type": "image", + "image_path": "ba6e84b44e23c9a9cb63d4d65158fa9ab1a8a971bc689e70da6690f52c55b920.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 395, + 319, + 480, + 416 + ], + "blocks": [ + { + "bbox": [ + 395, + 319, + 480, + 416 + ], + "lines": [ + { + "bbox": [ + 395, + 319, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 395, + 319, + 480, + 416 + ], + "type": "image", + "image_path": "1cc558c13428cc1d4a50181e84c4114f990fcf9ad0a234d4033c8e8beffbce19.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 461, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 568 + ], + "type": "text", + "content": "we set the prompt length as " + }, + { + "bbox": [ + 130, + 461, + 482, + 568 + ], + "type": "inline_equation", + "content": "N = 2" + }, + { + "bbox": [ + 130, + 461, + 482, + 568 + ], + "type": "text", + "content": " by default. As indicated in Fig. 4c, initializing prompts with manually crafted text introduces textual bias. For instance, 'The nice' yields better results for far-OOD but proves suboptimal for near-OOD, a bias that is perpetuated in our method when learnable prompts are initialized with 'The nice'. We discover that a random initialization, which does not introduce any prior, performs consistently well across different OOD tasks and is thus adopted as the default practice. Finally, in Fig. 4d, we observe that the placement of label tokens has a minimal impact on the results; hence, we by default place the class label at the end of the prompt, as shown in Eq. (5)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Automated sample collection. As shown in Fig. 5a, the synthetic image generation strategy outperforms realistic image retrieval, possibly because the generator has been exposed to a vast amount of training data (e.g., 5 billion images), which greatly surpasses the scale of our utilized Data (e.g., 60 million images). The best results are obtained by synergistically leveraging both synthetic and real data, confirming their complementary nature. In terms of image generation, employing a stronger image generator can result in better performance, as shown in Fig. 5b. Regarding image retrieval, utilizing a larger-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 216, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 216, + 91, + 448, + 102 + ], + "type": "text", + "content": "LAPT for OOD Detection with Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 218, + 198 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 218, + 198 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 218, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 218, + 198 + ], + "type": "image", + "image_path": "c8f2e4f3702b8e0796779848690d227a0d8b9a5702fe75d7251f738fed5ab07d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 200, + 219, + 210 + ], + "lines": [ + { + "bbox": [ + 133, + 200, + 219, + 210 + ], + "spans": [ + { + "bbox": [ + 133, + 200, + 219, + 210 + ], + "type": "text", + "content": "(a) Collection Methods" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 221, + 115, + 306, + 198 + ], + "blocks": [ + { + "bbox": [ + 221, + 115, + 306, + 198 + ], + "lines": [ + { + "bbox": [ + 221, + 115, + 306, + 198 + ], + "spans": [ + { + "bbox": [ + 221, + 115, + 306, + 198 + ], + "type": "image", + "image_path": "a29cfcf9dea77bc15da4d5e9ce80b54fb424eeddbe71bf1de87e259aaeb76186.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 200, + 291, + 210 + ], + "lines": [ + { + "bbox": [ + 236, + 200, + 291, + 210 + ], + "spans": [ + { + "bbox": [ + 236, + 200, + 291, + 210 + ], + "type": "text", + "content": "(b) Generators" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 392, + 198 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 392, + 198 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 392, + 198 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 392, + 198 + ], + "type": "image", + "image_path": "d2158e20fce24fcbc8f3b37c396b41cc8746d5f545a37e29a1ee8f37092f802a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 200, + 385, + 210 + ], + "lines": [ + { + "bbox": [ + 316, + 200, + 385, + 210 + ], + "spans": [ + { + "bbox": [ + 316, + 200, + 385, + 210 + ], + "type": "text", + "content": "(c) WebData Scale" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 219, + 480, + 243 + ], + "lines": [ + { + "bbox": [ + 131, + 219, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 131, + 219, + 480, + 243 + ], + "type": "text", + "content": "Fig. 5: Analyses on (a) image collection methods, (b) text-to-image generators, (c) data scale of the WebData space, and (d) number of training samples per class." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 396, + 115, + 480, + 198 + ], + "blocks": [ + { + "bbox": [ + 396, + 115, + 480, + 198 + ], + "lines": [ + { + "bbox": [ + 396, + 115, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 396, + 115, + 480, + 198 + ], + "type": "image", + "image_path": "7048f777e6c2586e9ec8052afd78ef9402b4aaa1b59c485911efbbe9fd51aadf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 200, + 474, + 210 + ], + "lines": [ + { + "bbox": [ + 402, + 200, + 474, + 210 + ], + "spans": [ + { + "bbox": [ + 402, + 200, + 474, + 210 + ], + "type": "text", + "content": "(d) Sample Number" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 271, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 380 + ], + "type": "text", + "content": "scale WebData can yield better outcomes, as presented in Fig. 5c. Finally, we find in Fig. 5d that the model's performance does not continue to increase with the addition of more training samples. This may be because it is difficult to collect a high-quality training set that is semantically consistent with the text and diverse. With the relatively small size of WebData, it is challenging to ensure semantic consistency, and while the semantic consistency of generated images may be good, their diversity is limited, as analyzed in Fig. 3. How to automatically collect high-quality images for effective model training remains an enduring and challenging problem." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 383, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 482, + 431 + ], + "type": "text", + "content": "More analyses on stage ablation, the impact of text prompts in data collection, data mixing strategies, hyper-parameters in data mixing, different VLMs, visualization of learned prompts, and time complexity are provided in the Supplementary Materials." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 465, + 318, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 465, + 318, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 465, + 318, + 478 + ], + "type": "text", + "content": "5 Conclusion and Limitations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 506, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 601 + ], + "type": "text", + "content": "We developed a novel label-driven automated prompt tuning (LAPT) approach that autonomously configures effective prompts for OOD detection. Requiring only ID class names, our method can automatically mine OOD class names, collect training images per class, and learn effective prompts, thereby outperforming manually crafted prompts with minimal manual intervention. The automatically learned prompts not only enhanced the distinction between ID and OOD samples but also improved the ID classification accuracy and the strengthened generalization robustness to covariate shifts." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "The efficacy of our approach was highly dependent on the quality of collected training samples. Employing more advanced generation techniques to create a diverse set of synthetic images [8] or expanding the retrieval space [43] to obtain real images that match better the class names could yield improved results. These possibilities present exciting avenues for future research." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. Beaumont, R.: Clip retrieval: Easily compute clip embeddings and build a clip retrieval system with them. https://github.com/rom1504/clip-retrieval (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 193 + ], + "type": "text", + "content": "2. Bitterwolf, J., Mueller, M., Hein, M.: In or out? fixing imagenet out-of-distribution detection evaluation. In: ICML (2023), https://proceedings.mlr.press/v202/bitterwolf23a.html" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 480, + 226 + ], + "type": "text", + "content": "3. Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., Vedaldi, A.: Describing textures in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3606-3613 (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 259 + ], + "type": "text", + "content": "4. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 261, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 480, + 281 + ], + "type": "text", + "content": "5. Djurisic, A., Bozanic, N., Ashok, A., Liu, R.: Extremely simple activation shaping for out-of-distribution detection. arXiv preprint arXiv:2209.09858 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 282, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 282, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 138, + 282, + 480, + 314 + ], + "type": "text", + "content": "6. Du, X., Gozum, G., Ming, Y., Li, Y.: Siren: Shaping representations for detecting out-of-distribution objects. Advances in Neural Information Processing Systems 35, 20434-20449 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 316, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 480, + 347 + ], + "type": "text", + "content": "7. Du, X., Wang, X., Gozum, G., Li, Y.: Unknown-aware object detection: Learning what you don't know from videos in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13678-13688 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 349, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 349, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 138, + 349, + 480, + 390 + ], + "type": "text", + "content": "8. Dunlap, L., Umino, A., Zhang, H., Yang, J., Gonzalez, J.E., Darrell, T.: Diversify your vision datasets with automatic diffusion-based augmentation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=9wrYfqdrwk" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "type": "text", + "content": "9. Esmaeilpour, S., Liu, B., Robertson, E., Shu, L.: Zero-shot out-of-distribution detection based on the pre-trained model clip. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 6568-6576 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 425, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 456 + ], + "type": "text", + "content": "10. Fort, S., Ren, J., Lakshminarayanan, B.: Exploring the limits of out-of-distribution detection. Advances in Neural Information Processing Systems 34, 7068-7081 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 458, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 480, + 490 + ], + "type": "text", + "content": "1. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "text", + "content": "2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 523, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 567 + ], + "type": "text", + "content": "3. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "text", + "content": "4. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "type": "text", + "content": "5. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. arXiv preprint arXiv:1610.02136 (2016)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "text", + "content": "6. Hendrycks, D., Mazeika, M., Dietterich, T.: Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606 (2018)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "7. Hendrycks, D., Mu, N., Cubuk, E.D., Zoph, B., Gilmer, J., Lakshminarayanan, B.: Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781 (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "text", + "content": "LAPT for OOD Detection with Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "18. Huang, R., Geng, A., Li, Y.: On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems 34, 677-689 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "text", + "content": "19. Huang, R., Li, Y.: Mos: Towards scaling out-of-distribution detection for large semantic space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8710-8719 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 226 + ], + "type": "text", + "content": "20. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "text", + "content": "21. Jiang, X., Liu, F., Fang, Z., Chen, H., Liu, T., Zheng, F., Han, B.: Negative label guided OOD detection with pretrained vision-language models. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=xU01HXz4an" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 304 + ], + "type": "text", + "content": "22. Kong, S., Ramanan, D.: Opengan: Open-set recognition via open data generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 813-822 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "24. Lee, K., Lee, K., Lee, H., Shin, J.: A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in neural information processing systems 31 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 392 + ], + "type": "text", + "content": "25. Liang, S., Li, Y., Srikant, R.: Enhancing the reliability of out-of-distribution image detection in neural networks. arXiv preprint arXiv:1706.02690 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "26. Liu, W., Wang, X., Owens, J., Li, Y.: Energy-based out-of-distribution detection. Advances in neural information processing systems 33, 21464-21475 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "27. Liu, X., Lochman, Y., Zach, C.: Gen: Pushing the limits of softmax-based out-of-distribution detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23946-23955 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "28. Lu, F., Zhu, K., Zheng, K., Zhai, W., Cao, Y.: Likelihood-aware semantic alignment for full-spectrum out-of-distribution detection. arXiv preprint arXiv:2312.01732 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "text", + "content": "29. Ming, Y., Cai, Z., Gu, J., Sun, Y., Li, W., Li, Y.: Delving into out-of-distribution detection with vision-language representations. Advances in Neural Information Processing Systems 35, 35087-35102 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "30. Ming, Y., Li, Y.: How does fine-tuning impact out-of-distribution detection for vision-language models? International Journal of Computer Vision 132(2), 596-609 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "text", + "content": "31. Ming, Y., Sun, Y., Dia, O., Li, Y.: How to exploit hyperspherical embeddings for out-of-distribution detection? arXiv preprint arXiv:2203.04450 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "32. Miyai, A., Yu, Q., Irie, G., Aizawa, K.: Locoop: Few-shot out-of-distribution detection via prompt learning. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "33. Nguyen, A., Yosinski, J., Clune, J.: Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 427-436 (2015)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "34. Nie, J., Zhang, Y., Fang, Z., Liu, T., Han, B., Tian, X.: Out-of-distribution detection with negative prompts. In: The Twelfth International Conference on Learning Representations (2024)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "35. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "type": "text", + "content": "36. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "37. Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: Do imagenet classifiers generalize to imagenet? In: International conference on machine learning. pp. 5389-5400. PMLR (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 480, + 259 + ], + "type": "text", + "content": "38. Ren, J., Fort, S., Liu, J., Roy, A.G., Padhy, S., Lakshminarayanan, B.: A simple fix to mahalanobis distance for improving near-ood detection. arXiv preprint arXiv:2106.09022 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "type": "text", + "content": "40. Ryu, S., Koo, S., Yu, H., Lee, G.G.: Out-of-domain detection based on generative adversarial network. In: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. pp. 714-718 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "type": "text", + "content": "41. Scheirer, W.J., de Rezende Rocha, A., Sapkota, A., Boult, T.E.: Toward open set recognition. IEEE transactions on pattern analysis and machine intelligence 35(7), 1757-1772 (2012)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "type": "text", + "content": "42. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "43. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "44. Sehwag, V., Chiang, M., Mittal, P.: Ssd: A unified framework for self-supervised outlier detection. arXiv preprint arXiv:2103.12051 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "text", + "content": "45. Sun, Y., Guo, C., Li, Y.: React: Out-of-distribution detection with rectified activations. Advances in Neural Information Processing Systems 34, 144-157 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "text", + "content": "46. Sun, Y., Ming, Y., Zhu, X., Li, Y.: Out-of-distribution detection with deep nearest neighbors. In: International Conference on Machine Learning. pp. 20827-20840. PMLR (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "47. Tack, J., Mo, S., Jeong, J., Shin, J.: Csi: Novelty detection via contrastive learning on distributionally shifted instances. Advances in neural information processing systems 33, 11839-11852 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "text", + "content": "48. Tao, L., Du, X., Zhu, X., Li, Y.: Non-parametric outlier synthesis. arXiv preprint arXiv:2303.02966 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 610 + ], + "type": "text", + "content": "49. Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S.: The inaturalist species classification and detection dataset. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8769-8778 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "50. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Open-set recognition: A good closed-set classifier is all you need? arXiv preprint arXiv:2110.06207 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "51. Verma, V., Lamb, A., Beckham, C., Najafi, A., Mitliagkas, I., Lopez-Paz, D., Bengio, Y.: Manifold mixup: Better representations by interpolating hidden states. In: International conference on machine learning. pp. 6438-6447. PMLR (2019)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "text", + "content": "LAPT for OOD Detection with Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "52. Wang, H., Li, Z., Feng, L., Zhang, W.: Vim: Out-of-distribution with virtual-logit matching. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4921-4930 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 182 + ], + "type": "text", + "content": "53. Wang, H., Li, Y., Yao, H., Li, X.: Clipn for zero-shot ood detection: Teaching clip to say no. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1802-1812 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "type": "text", + "content": "54. Wang, Y., Li, B., Che, T., Zhou, K., Liu, Z., Li, D.: Energy-based open-world uncertainty modeling for confidence calibration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9302-9311 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 247 + ], + "type": "text", + "content": "55. Wei, H., Xie, R., Cheng, H., Feng, L., An, B., Li, Y.: Mitigating neural network overconfidence with logit normalization. In: International Conference on Machine Learning. pp. 23631-23644. PMLR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 247, + 480, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 480, + 279 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 480, + 279 + ], + "type": "text", + "content": "56. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: Sun database: Large-scale scene recognition from abbey to zoo. In: 2010 IEEE computer society conference on computer vision and pattern recognition. pp. 3485-3492. IEEE (2010)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 279, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 480, + 300 + ], + "type": "text", + "content": "57. Xu, K., Chen, R., Franchi, G., Yao, A.: Scaling for training time and post-hoc out-of-distribution detection enhancement. arXiv preprint arXiv:2310.00227 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "text", + "content": "58. Yang, J., Wang, P., Zou, D., Zhou, Z., Ding, K., Peng, W., Wang, H., Chen, G., Li, B., Sun, Y., Du, X., Zhou, K., Zhang, W., Hendrycks, D., Li, Y., Liu, Z.: Openood: Benchmarking generalized out-of-distribution detection (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 332, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 354 + ], + "type": "text", + "content": "59. Yang, J., Zhou, K., Li, Y., Liu, Z.: Generalized out-of-distribution detection: A survey. arXiv preprint arXiv:2110.11334 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "text", + "content": "60. Yang, J., Zhou, K., Liu, Z.: Full-spectrum out-of-distribution detection. International Journal of Computer Vision pp. 1-16 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "text", + "content": "61. Yang, T., Huang, Y., Xie, Y., Liu, J., Wang, S.: Mixood: Improving out-of-distribution detection with enhanced data mixup. ACM Transactions on Multimedia Computing, Communications and Applications 19(5), 1-18 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "type": "text", + "content": "62. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 429, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 449 + ], + "type": "text", + "content": "63. Zhang, J., Inkawich, N., Chen, Y., Li, H.: Fine-grained out-of-distribution detection with mixup outlier exposure. CoRR (abs/2106.03917) (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 449, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 480, + 493 + ], + "type": "text", + "content": "64. Zhang, J., Inkawich, N., Linderman, R., Chen, Y., Li, H.: Mixture outlier exposure: Towards out-of-distribution detection in fine-grained environments. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 5531-5540 (January 2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "type": "text", + "content": "65. Zhang, J., Yang, J., Wang, P., Wang, H., Lin, Y., Zhang, H., Sun, Y., Du, X., Zhou, K., Zhang, W., Li, Y., Liu, Z., Chen, Y., Li, H.: Openood v1.5: Enhanced benchmark for out-of-distribution detection. arXiv preprint arXiv:2306.09301 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 525, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 568 + ], + "type": "text", + "content": "66. Zhang, J., Fu, Q., Chen, X., Du, L., Li, Z., Wang, G., Han, S., Zhang, D., et al.: Out-of-distribution detection based on in-distribution data patterns memorization with modern hopfield energy. In: The Eleventh International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "text", + "content": "67. Zhang, R., Fang, R., Zhang, W., Gao, P., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "text", + "content": "68. Zhang, Y., Deng, B., Jia, K., Zhang, L.: Label propagation with augmented anchors: A simple semi-supervised learning baseline for unsupervised domain adaptation. In: European Conference on Computer Vision. pp. 781-797. Springer (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "69. Zhang, Y., Deng, B., Tang, H., Zhang, L., Jia, K.: Unsupervised multi-class domain adaptation: Theory, algorithms, and practice. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(5), 2775-2792 (2020)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 291 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "text", + "content": "70. Zhang, Y., Zhu, W., Tang, H., Ma, Z., Zhou, K., Zhang, L.: Dual memory networks: A versatile adaptation approach for vision-language models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 28718-28728 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 482, + 194 + ], + "type": "text", + "content": "71. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence 40(6), 1452-1464 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 482, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 482, + 216 + ], + "type": "text", + "content": "72. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 482, + 248 + ], + "type": "text", + "content": "73. Zhou, Q., Feng, Z., Gu, Q., Pang, J., Cheng, G., Lu, X., Shi, J., Ma, L.: Context-aware mixup for domain adaptive semantic segmentation. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT) 33(2), 804-817 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 482, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 482, + 291 + ], + "type": "text", + "content": "74. Zhou, Q., Gu, Q., Pang, J., Lu, X., Ma, L.: Self-adversarial disentangling for specific domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 45(7), 8954-8968 (2023). https://doi.org/10.1109/TPAMI.2023.3238727" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 216, + 91, + 447, + 102 + ], + "type": "text", + "content": "LAPT for OOD Detection with Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LASS3D_ Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation/c8ab9b1f-da9d-4178-ac58-4d667a9330b3_content_list.json b/2024/LASS3D_ Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation/c8ab9b1f-da9d-4178-ac58-4d667a9330b3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..be1ce4c02b51d3dc3704d97f22bb154c48c94c3e --- /dev/null +++ b/2024/LASS3D_ Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation/c8ab9b1f-da9d-4178-ac58-4d667a9330b3_content_list.json @@ -0,0 +1,1842 @@ +[ + { + "type": "text", + "text": "LASS3D: Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation", + "text_level": 1, + "bbox": [ + 222, + 143, + 781, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jianan $\\mathrm{Li}^{1,2}$ and Qiulei Dong", + "bbox": [ + 362, + 224, + 640, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Artificial Intelligence, University of Chinese Academy of Sciences", + "bbox": [ + 225, + 251, + 759, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences", + "bbox": [ + 225, + 265, + 772, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Center for Excellence in Brain Science and Intelligence Technology, Chinese Academy of Sciences", + "bbox": [ + 241, + 294, + 759, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lijianan211@mails.ucas.ac.cn, qldong@nlpr.ac.cn", + "bbox": [ + 316, + 323, + 683, + 335 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Precisely annotating large-scale 3D datasets for point cloud segmentation is laborious. To alleviate the annotation burden, several semi-supervised 3D segmentation methods have been proposed in literature. However, two issues remain to be tackled: 1) The utilization of large language-vision models (LVM) in semi-supervised 3D semantic segmentation remains under-explored. 2) The unlabeled points with low-confidence predictions are directly discarded by existing methods. Taking these two issues into consideration, we propose a language-assisted semi-supervised 3D semantic segmentation method named LASS3D, which is built upon the commonly used MeanTeacher framework. In LASS3D, we use two off-the-shelf LVM to generate multi-level captions and leverage the images as the bridge to connect the text data and point clouds. Then, a semantic-aware adaptive fusion module is explored in the student branch, where the semantic information encoded in the embeddings of multi-level captions is injected into 3D features by adaptive fusion and then the semantic information in the text-enhanced 3D features is transferred to the teacher branch by knowledge distillation. In addition, a progressive exploitation strategy is explored for the unreliable points in the teacher branch, which can effectively exploit the information encapsulated in unreliable points via negative learning. Experimental results on both outdoor and indoor datasets demonstrate that LASS3D outperforms the comparative methods in most cases.", + "bbox": [ + 259, + 375, + 743, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D semantic segmentation $\\cdot$ Semi-supervised segmentation", + "bbox": [ + 261, + 694, + 740, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 736, + 375, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Semantic Segmentation [7,22-24,42,43,51] is an important task for perceiving real-world environments, and intensive efforts have been devoted to this task in recent years. However, most of the existing works [8,22,41,42,51,54] in literature", + "bbox": [ + 212, + 768, + 787, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 217, + 824, + 382, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/18270764dc2d370675187c775afcc2039a2ead85285408e22c0ab97be77c74d2.jpg", + "image_caption": [ + "Fig. 1: Illustration of the proportions of reliable and unreliable points predicted by GPC [17] on ScanNet [4]. The horizontal axis denotes the labeled ratio and training iteration. The vertical axis denotes the proportion." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 774, + 270 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "are fully-supervised methods, which require time-consuming and labor-intensive data annotation for training.", + "bbox": [ + 212, + 356, + 782, + 387 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This issue encourages researchers to investigate semi-supervised 3D semantic segmentation, where only a small amount of densely labeled data and a large amount of unlabeled data are provided for training. Existing works [5,17,21,26, 38] on semi-supervised 3D semantic segmentation have shown their effectiveness, but there is ample room for improvement in the following two aspects:", + "bbox": [ + 212, + 387, + 784, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1) Utilization of language-vision models (LVM): Recently, LVM [16,33] trained on web-crawled image-text pairs has been applied in some downstream tasks [2, 31, 39, 50] to boost performance, owing to its strong representation ability. Text embeddings extracted by the LVM generally contain higher-level abstraction of visual concepts with rich semantic information, which is complementary to the 3D features that mainly contain geometric information and thus can facilitate mining the semantic information of unlabeled data. To our best knowledge, LVM has not been applied in semi-supervised 3D semantic segmentation yet.", + "bbox": [ + 212, + 462, + 787, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2) Exploitation of unreliable points (unlabeled points with low-confidence predictions): The proportions of reliable and unreliable points predicted by the state-of-the-art semi-supervised 3D semantic segmentation method GPC [17] on ScanNet [4] dataset are illustrated in Fig. 1. As seen from this figure, a large number of unlabeled points are deemed unreliable, especially at the early training stage. Existing works [17,21] generally adopt a confidence-based filtering strategy to remove the unreliable points for guaranteeing the effectiveness and stability of the training process. However, simply discarding them will inevitably lose some useful information. Thus, a proper learning strategy is expected to be specially designed to exploit these unreliable points effectively.", + "bbox": [ + 212, + 583, + 787, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Taking the above two aspects into consideration, we propose a method named LASS3D for semi-supervised 3D semantic segmentation, where the LVM and unreliable points are effectively exploited. The proposed LASS3D is built upon the MeanTeacher [37] framework, where the weights of the teacher branch are updated by the exponential moving average (EMA) of the student's weights.", + "bbox": [ + 212, + 734, + 787, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the student branch, two LVM (i.e., KOSMOS-2 [32] and GroundedSAM [35]) are used to generate image-level and entity-level captions, and the images", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Li and Q. Dong", + "bbox": [ + 271, + 114, + 395, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "are leveraged as the bridge to connect the text data and point clouds. Then, a semantic-aware adaptive fusion module (SAFM) is designed to assist point cloud learning with the semantic knowledge encoded in text data. Specifically, multi-level point-text pairs are constructed in SAFM to adaptively fuse the 3D features and the text embeddings of the image-level and entity-level captions, which could facilitate capturing both holistic and fine-grained semantic information.", + "bbox": [ + 212, + 146, + 782, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the teacher branch, a fusion adapter is utilized to narrow the distribution gap between the 3D features extracted by the backbone network and the text-enhanced 3D features in the student branch. Combined with the EMA strategy, the semantic information encoded in the text-enhanced 3D features is distilled from the student branch to the teacher branch. Note that we inject semantic information from text embeddings into the 3D features by fusion rather than directly mapping the semantic labels from pixels to points. Because the semantic labels generated by the open-world segmentation models (e.g., SAM [20]) are generally in a one-hot form, which provide hard supervision. However, some generated synonymous labels may not exist in our desired label set (e.g., the generated label is 'couch' but the ground-truth label is 'sofa'), and incorrect hard supervision may adversely affect the point cloud segmentation. Adaptively fusing the text embeddings and 3D features can mitigate this negative effect to some extent. Because the strong LVM can encode synonymous words into similar embeddings and thus can inject similar semantic information into 3D features. To effectively exploit the unlabeled points with unreliable predictions, a progressive exploitation strategy is designed based on the fact that determining the least likely category for an unlabeled sample is generally easier and more accurate than determining its most likely category. In addition, the proposed progressive exploitation strategy can be seamlessly integrated into some existing methods and further boost their performances, which is demonstrated in Sec. 4.4.", + "bbox": [ + 212, + 237, + 787, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, the contributions of this paper are as follows:", + "bbox": [ + 238, + 554, + 661, + 569 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a semantic-aware adaptive fusion module (SAFM) to consolidate point cloud segmentation with the semantic information encoded in text embeddings. The multi-level fusion shows the feasibility and effectiveness of language-assisted 3D semantic segmentation in semi-supervised settings.", + "- We propose a progressive exploitation strategy for the unlabeled unreliable points. This strategy is beneficial to fully exploit the unlabeled data and can be seamlessly embedded into some existing methods.", + "- We propose the LASS3D for semi-supervised 3D semantic segmentation by integrating the above SAFM and progressive exploitation strategy. Its effectiveness is demonstrated by the results on both outdoor-scene and indoor-scene public datasets in Sec. 4." + ], + "bbox": [ + 225, + 575, + 785, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 758, + 387, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Semi-supervised 3D Semantic Segmentation", + "text_level": 1, + "bbox": [ + 214, + 787, + 622, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To alleviate the annotation burden, label-efficient 3D semantic segmentation (including weakly-supervised segmentation [13,28,29,36] and semi-supervised seg", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "LASS3D", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mentation [5,17,21,25,26,38]) has drawn increasing interest among researchers. In this work, we delve into semi-supervised 3D semantic segmentation, which aims at utilizing a small number of densely labeled data and a large number of unlabeled data for model training. The core of semi-supervised 3D segmentation lies in exploiting the unlabeled data to the fullest. For example, Deng et al. [5] proposed to optimize the pseudo labels of the unlabeled data with the superpoints generated by geometry-based and color-based region-growing algorithms. Jiang et al. [17] proposed a label-guided contrastive loss for the unlabeled data, which could enhance the feature representation ability of the model. Kong et al. [21] proposed to leverage the spatial prior of LiDAR point clouds and use a mixing operation to provide supervisory information for the unlabeled data. Li et al. [26] proposed to utilize the reflectivity-prior descriptors to generate high-quality pseudo labels for the unlabeled data. Unal et al. [38] proposed to distill high-level feature information from a synthetically trained 2D network for data-efficient LiDAR semantic segmentation.", + "bbox": [ + 212, + 146, + 787, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, these methods neglect the usage of text data that contains rich semantic information. To our best knowledge, LASS3D is the first work to investigate the language-assisted semi-supervised 3D semantic segmentation, where the powerful representation ability of language-vision models is exploited.", + "bbox": [ + 212, + 373, + 787, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Language-assisted 3D Learning", + "text_level": 1, + "bbox": [ + 214, + 455, + 519, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Large language-vision models (LVM) [16, 33] have drawn growing attention in computer vision recently, attributed to their powerful representation ability. For example, CLIP [33], which is trained on massive paired text-image data via contrastive learning, has prevailed in various downstream visual tasks, such as semantic segmentation [10, 18, 34, 48, 52], object detection [30, 45, 46], and video recognition [27]. Recent advanced methods that focused on language-assisted point cloud learning could be roughly divided into two types: the projection-based methods and the alignment-based methods.", + "bbox": [ + 212, + 479, + 787, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The projection-based methods [9,15,49,53] project the point clouds into the 2D plane so that the image encoder in the LVM can be seamlessly utilized and the text-image alignment of the LVM can be leveraged for downstream tasks.", + "bbox": [ + 212, + 599, + 787, + 646 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6e0b8b61bd268728aeaf448be48b26bfb20157044f2632c27ce243bc71680f1c.jpg", + "image_caption": [ + "(a) Pipeline of PointCLIP." + ], + "image_footnote": [], + "bbox": [ + 225, + 674, + 516, + 777 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/06d6033f83a48b21784513dc72197b314b638be96ac67e9ef2e8839f71dfb0b0.jpg", + "image_caption": [ + "(b) Alignment process of CLIP2Scene.", + "Fig. 2: Illustration of the pipeline of a projection-based method (e.g., PointCLIP [49]) and the alignment process of an alignment-based method (e.g., CLIP2Scene [2])." + ], + "image_footnote": [], + "bbox": [ + 522, + 676, + 785, + 782 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Li and Q. Dong", + "bbox": [ + 271, + 114, + 395, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 2a illustrates the general pipeline of PointCLIP [49], it classifies the point clouds according to the similarities between the text embeddings and features of the projected images. Nevertheless, the projection-based methods are generally designed for object-level 3D classification. In addition, they usually suffer from the loss of geometric information during the projection process, thus resulting in suboptimal performance.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The alignment-based methods [2,6,11,12,14,39,44,47] enforce the consistency among the multi-modal features. Fig. 2b illustrates the alignment process of CLIP2Scene [2], it uses contrastive learning to align the 3D features to their corresponding text embeddings. However, directly aligning multi-modal features may result in suboptimal results, due to the distribution discrepancy between the 3D features and text embeddings.", + "bbox": [ + 212, + 237, + 787, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike existing methods that are either projection-based or alignment-based, we propose a fusion-based method to adaptively fuse the embeddings of multi-level captions and 3D features, which can alleviate the loss of geometric information and mitigate the negative effect brought by the distribution gap.", + "bbox": [ + 212, + 327, + 785, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 214, + 410, + 380, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Architecture", + "text_level": 1, + "bbox": [ + 214, + 441, + 369, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The architecture of the proposed LASS3D is illustrated in Fig. 3. It is built upon the MeanTeacher [37] framework where the teacher branch provides the supervisory signals for the student branch and is updated by the exponential mean average (EMA) of the student's weights:", + "bbox": [ + 212, + 465, + 787, + 527 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {t e a c h e r} ^ {t + 1} = \\alpha \\theta_ {t e a c h e r} ^ {t} + (1 - \\alpha) \\theta_ {s t u d e n t} ^ {t}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 535, + 785, + 556 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c51340eeb77e279aca57a369a8662d9b342a6055cc1a17692bdec13d6d6e408f.jpg", + "image_caption": [ + "Fig. 3: Architecture of the proposed LASS3D. $\\mathcal{C}_i$ and $\\mathcal{C}_e$ stand for the image-level captions and entity-level captions. $\\mathcal{E}_i$ and $\\mathcal{E}_e$ are the embeddings of $\\mathcal{C}_i$ and $\\mathcal{C}_e$ respectively. $/ / /$ denotes the detach operation which is used to stop the gradient." + ], + "image_footnote": [], + "bbox": [ + 218, + 579, + 782, + 781 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "LASS3D", + "bbox": [ + 669, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta_{teacher}^{t}$ and $\\theta_{student}^{t}$ denote the weights of the teacher branch and student branch at time step $t$ , and $\\alpha$ is the update hyperparameter.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The student branch takes the point clouds and their corresponding images as input. The images are fed into the LVM, (i.e., KOSMOS-2 [35] and GroundedSAM [32]) to produce image-level $\\mathcal{C}_i$ and entity-level captions $\\mathcal{C}_e$ . Then, the multi-level captions are tokenized by the text encoder of CLIP [33] to generate their corresponding text embeddings. The text embeddings and 3D features extracted by the 3D backbone network are fused in the semantic-aware adaptive fusion module to obtain the text-enhanced 3D features, which are then used for segmentation in the student branch.", + "bbox": [ + 212, + 178, + 784, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The teacher branch takes the same point clouds as input. The 3D features extracted by the 3D backbone network are aligned with the text-enhanced 3D features through a fusion adapter. Then, the classifier takes the aligned 3D features as input and outputs their corresponding prediction scores. A dual-score separation strategy is adopted to select reliable points and unreliable points from the unlabeled points. The predictions of the reliable points are regarded as the pseudo labels of their corresponding points in the student branch, where the cross-entropy loss $\\mathcal{L}_{CE}^{u}$ is calculated. The predictions of the unreliable points in the student branch are used for the progressive exploitation, where the negative learning loss $\\mathcal{L}_{PE}^{u}$ is calculated. Besides, the predictions of the labeled points in the student branch and their ground-truth labels are used to calculate the cross-entropy loss $\\mathcal{L}_{CE}^{l}$ , and the features output by the fusion adapter are aligned to the text-enhanced 3D features by a Kullback-Leibler divergence loss $\\mathcal{L}_{KL}^{u}$ .", + "bbox": [ + 212, + 299, + 785, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that the text data is only required in the training stage. In the inference stage, only the teacher branch is used and it takes the point clouds as input.", + "bbox": [ + 212, + 496, + 784, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Semantic-aware Adaptive Fusion Module", + "text_level": 1, + "bbox": [ + 214, + 551, + 599, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The architecture of the semantic-aware adaptive fusion module (SAFM) is illustrated in Fig. 4, it takes the embeddings of multi-level captions (i.e., image-level embeddings $\\mathcal{E}_i$ and entity-level embeddings $\\mathcal{E}_e$ ) and 3D features $f_{3D}$ as input and outputs the text-enhanced 3D features $f_{3D}^t$ , which are then fed into the classifier for segmentation and used as the supervision for knowledge distillation.", + "bbox": [ + 212, + 580, + 784, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed SAFM is used to inject the semantic information encoded in multi-level captions into the 3D features. Specifically, the image-level captions $\\mathcal{C}_i$ and entity-level captions $\\mathcal{C}_e$ are generated by KOSMOS-2 [32] and GroundedSAM [35] respectively, and their corresponding embeddings are extracted by the text encoder of CLIP [33]. To guarantee the accuracy of multi-modal fusion, we construct the point-text pairs by using the images as the bridge. Then, given the point-text pairs, their corresponding features are adaptively fused to produce the text-enhanced 3D features. In this subsection, we will introduce the above two key operations of SAFM in detail.", + "bbox": [ + 212, + 657, + 785, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Point-text pair construction. We use the images as the bridge to connect the text data and point cloud data. Specifically, the multi-level pixel-text pairs can be obtained by the language-vision models, i.e., the image-level captions", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Li and Q. Dong", + "bbox": [ + 271, + 114, + 395, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/74abd7d2ade34c9654f4a0b42b8b96c223192a277cf186cd2dd745c81431ba7e.jpg", + "image_caption": [ + "Fig. 4: Architecture of the proposed semantic-aware adaptive fusion module. $f_{3D}$ , $\\hat{f}_{3D}$ , and $f_{3D}^{t}$ denote the extracted 3D features, aligned 3D features, and text-enhanced 3D features. $\\mathcal{E}_i$ and $\\mathcal{E}_e$ are the embeddings of image-level and entity-level captions. $\\oplus$ denotes the add operation." + ], + "image_footnote": [], + "bbox": [ + 246, + 142, + 756, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "correspond to all pixels of the given images, and the entity-level captions correspond to the pixels of the segmented regions. The pixel-point pairs can be easily obtained with the extrinsic and intrinsic parameters of the cameras:", + "bbox": [ + 212, + 363, + 784, + 409 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n[ u, v, 1 ] ^ {\\mathrm {T}} = \\mathbf {K M} [ x, y, z, 1 ] ^ {\\mathrm {T}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 422, + 785, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $[u,v,1]^{\\mathrm{T}}$ represents the homogeneous coordinates of a pixel, $[x,y,z,1]^{\\mathrm{T}}$ represents its corresponding 3D point position in world coordinates, $\\mathbf{K}$ and $\\mathbf{M}$ are the intrinsic and extrinsic parameters of the cameras. With the multi-level pixel-text and pixel-point pairs, the multi-level point-text pairs can be obtained.", + "bbox": [ + 212, + 455, + 784, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adaptive fusion. To effectively utilize the multi-level point-text pairs, an adaptive fusion strategy is proposed. As shown in Fig. 4, the dimension of the 3D features $f_{3D}$ is firstly expanded by an MLP layer to align the embedding space of the text encoder, given that the text encoder in LASS3D is fixed. Since the image-level captions may have different correlations with different regions of the paired points, cross attention is performed between the aligned 3D features $\\hat{f}_{3D}$ and image-level embeddings $\\mathcal{E}_i$ , with $\\hat{f}_{3D}$ serving as the query vectors and $\\mathcal{E}_i$ serving as the key vectors and value vectors. Then, considering that the points related to the same entity-level captions share the same semantic information, we simply add the entity-level embeddings $\\mathcal{E}_e$ with their corresponding 3D features. Finally, an MLP layer is used to map the fused features to their original dimension and output the text-enhanced 3D features $f_{3D}^t$ .", + "bbox": [ + 212, + 518, + 785, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Fig. 3, the text-enhanced 3D features $f_{3D}^{t}$ are not only fed into the classifier of the student branch for segmentation but also serve as the supervisory signals for knowledge distillation in the teacher branch. The features $f_{ada}$ output by the fusion adapter in the teacher branch are aligned to their corresponding $f_{3D}^{t}$ through $\\mathcal{L}_{KL}^{u}$ :", + "bbox": [ + 212, + 700, + 785, + 779 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {K L} ^ {u} = \\mathrm {K L} \\left(f _ {3 D} ^ {t} \\mid \\mid f _ {a d a}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 792, + 785, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{KL}(\\cdot)$ denotes the Kullback-Leibler divergence loss.", + "bbox": [ + 214, + 824, + 632, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "LASS3D", + "bbox": [ + 671, + 113, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/37c560e7b39e5d0cbab11a466d1647bce478e3680bce9ba389ee0e3bd2fb1171.jpg", + "image_caption": [ + "Fig. 5: Illustration of the progressive exploitation for the unreliable point $P_{url}^{i}$ at the first two iterations. $S^{i}$ denotes the prediction confidence of $P_{url}^{i}$ and $\\hat{y}_i$ denotes its negative label. The selected negative class (blue font) is used for negative learning and the used negative class (marked with red dashed box) is discarded in the subsequent iterations. Note that some unreliable points may turn into reliable points during progressive exploitation." + ], + "image_footnote": [], + "bbox": [ + 218, + 146, + 784, + 246 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Progressive Exploitation Strategy", + "text_level": 1, + "bbox": [ + 215, + 369, + 539, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The progressive exploitation strategy is proposed to mine the useful information encapsulated in the unlabeled unreliable points. It contains two key steps: dual-score separation and progressive negative learning. In this subsection, we will introduce the above two steps in detail.", + "bbox": [ + 212, + 393, + 782, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dual-score separation. Unlike GPC [17] and LaserMix [21] which only use the confidence of predictions to measure the reliability, we propose a dual-score separation strategy that additionally introduces the variances among the predictions of different branches into reliability measurement. Specifically, the unlabeled points $P_{u}$ in the teacher branch are separated into the reliable set $P_{rl}$ and unreliable set $P_{url}$ based on both confidence and variances:", + "bbox": [ + 212, + 455, + 784, + 546 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nP _ {r l} = B \\cdot P _ {u}, P _ {u r l} = (1 - B) \\cdot P _ {u},\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 556, + 602, + 571 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nB = \\mathbb {1} \\left[ \\sum_ {c = 1} ^ {C} \\mathbb {1} \\left[ S _ {c} \\geq \\tau_ {r} \\right] \\cdot \\mathbb {1} \\left[ \\sigma_ {c} \\leq \\tau_ {v} \\right] > 0 \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 575, + 784, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $B$ is a binary mask to select $P_{rl}$ and $P_{url}$ , $\\mathbb{1}$ is the indicator function, $C$ denotes the number of classes, $S_{c}$ denotes the prediction confidences of class $c$ in the teacher branch, $\\sigma_{c}$ denotes the prediction variances of class $c$ within the teacher and student branches, $\\tau_{r}$ and $\\tau_{v}$ are two pre-determined thresholds.", + "bbox": [ + 212, + 627, + 784, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The predictions of reliable points are regarded as the pseudo labels of their corresponding points in the student branch, which are used for supervised training, as shown in Eq. (8). The unreliable points are fully exploited via progressive negative learning, which is elaborated below.", + "bbox": [ + 212, + 688, + 784, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Progressive negative learning. Enlightened by the core idea of conventional negative learning [19, 40], we choose to mine information from the classes that the unreliable points do not belong to, rather than struggling to correctly tell which class they belong to. However, conventional methods only choose one negative class (the class with the lowest prediction score) for negative learning, which provides limited information and may cause unstable training due to the", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Li and Q. Dong", + "bbox": [ + 271, + 114, + 395, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dynamic negative labels. Instead, we conduct negative learning in a progressive manner, with more negative classes selected.", + "bbox": [ + 214, + 146, + 782, + 175 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The progressive exploitation at the first two iterations is illustrated in Fig. 5. As shown in this figure, we maintain a list for each unreliable point to record its selected negative classes. In each iteration, one negative class is selected for each unreliable point:", + "bbox": [ + 214, + 176, + 782, + 236 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {i} [ k ] = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} k = \\arg \\min \\left(S ^ {i}\\right) \\text {a n d} S ^ {i} [ k ] \\leq \\tau_ {n} \\\\ 0, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 244, + 782, + 281 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\hat{y}_i[k]$ denotes the $k$ -th dimension of the negative label of $P_{url}^i$ , $S^i$ denotes the prediction confidence of $P_{url}^i$ , and $\\tau_n$ is a pre-determined threshold.", + "bbox": [ + 214, + 290, + 782, + 320 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The selected negative label is used for negative learning:", + "bbox": [ + 238, + 320, + 645, + 335 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {P E} ^ {u} = - \\sum_ {i = 1} ^ {N _ {u}} \\hat {y} _ {i} \\log \\left(1 - O _ {i}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 343, + 782, + 383 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $N_{u}$ is the number of unreliable points with selected negative class and $O_{i}$ is the prediction of $P_{url}^{i}$ . Then, the used negative class is discarded in the subsequent iterations. Simultaneously, the network is optimized via the negative learning loss. Subsequently, a new negative class is selected from the remaining classes for the next iteration. The above process is repeated until no negative class can be selected or the iteration number is reached.", + "bbox": [ + 214, + 388, + 782, + 478 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Moreover, it is worth noting that the proposed progressive exploitation strategy can be seamlessly integrated into methods that discard the unreliable points, and further boost their performances, as shown in Sec. 4.4.", + "bbox": [ + 214, + 479, + 782, + 525 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Total Loss", + "text_level": 1, + "bbox": [ + 215, + 544, + 346, + 556 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The labeled points in the student branch are trained in a supervised manner, with their corresponding ground-truth labels $y^{l}$ being the supervisory signals:", + "bbox": [ + 214, + 566, + 782, + 595 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C E} ^ {l} = \\operatorname {C E} \\left(O ^ {s t u, l}, y ^ {l}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 603, + 782, + 619 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathrm{CE}(\\cdot)$ denotes the cross-entropy loss, $O^{stu,l}$ denotes the predictions of the labeled points in the student branch.", + "bbox": [ + 214, + 626, + 782, + 655 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The predictions of the reliable points in the teacher branch are regarded as the pseudo labels of their corresponding points in the student branch:", + "bbox": [ + 214, + 657, + 782, + 686 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C E} ^ {u} = \\operatorname {C E} \\left(O ^ {s t u, u}, y ^ {u}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 695, + 782, + 710 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $y^{u}$ denotes the reliable predictions output by the teacher branch and $O^{stu,u}$ denotes the predictions of corresponding unlabeled points in the student branch.", + "bbox": [ + 214, + 717, + 782, + 761 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Accordingly, the total loss $\\mathcal{L}_{total}$ is the weighted sum of the two loss terms $\\mathcal{L}_{CE}^{l}$ , $\\mathcal{L}_{CE}^{u}$ in the student branch and the aforementioned $\\mathcal{L}_{PE}^{u}$ , $\\mathcal{L}_{KL}^{u}$ :", + "bbox": [ + 215, + 763, + 782, + 795 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {C E} ^ {l} + \\lambda_ {u} \\mathcal {L} _ {C E} ^ {u} + \\lambda_ {P E} \\mathcal {L} _ {P E} ^ {u} + \\lambda_ {K L} \\mathcal {L} _ {K L} ^ {u}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 801, + 782, + 816 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda_{u},\\lambda_{PE},\\lambda_{KL}$ are the weights of $\\mathcal{L}_{CE}^{u},\\mathcal{L}_{PE}^{u},\\mathcal{L}_{KL}^{u}$ .", + "bbox": [ + 214, + 824, + 612, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "LASS3D", + "bbox": [ + 669, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 143, + 366, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Datasets and Evaluation Metric", + "text_level": 1, + "bbox": [ + 215, + 178, + 524, + 193 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The following outdoor-scene and indoor-scene datasets are used to evaluate the proposed LASS3D:", + "bbox": [ + 215, + 205, + 785, + 234 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- SemanticKITTI [1] is a large-scale 3D outdoor-scene LiDAR dataset consisting of 22 sequences, among which 10 sequences are used for training, 1 sequence is used for validation, and 11 sequences are used for testing. It contains 19 categories after merging and ignoring classes with few points. According to the official splitting, 850 scenes are used for training and validation, and 150 scenes are utilized for testing.", + "- ScanNet V2 [4] is a widely-used 3D indoor-scene dataset consisting of 1613 scans with 20 categories, where the point clouds are reconstructed from multi-view RGB-D images. The training set contains 1201 scans, the validation split contains 312 scans, and the testing set contains 100 scans." + ], + "bbox": [ + 225, + 237, + 785, + 387 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As done in previous works [21, 26, 38], we use the mIoU (mean Intersection over Union) as the evaluation metric.", + "bbox": [ + 215, + 388, + 785, + 417 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 441, + 455, + 458 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We follow the basic experimental settings of LaserMix [21] and GPC [17] for evaluating the proposed LASS3D on the outdoor-scene and indoor-scene datasets respectively. The classifiers and fusion adapter are both multi-layer perceptrons. We follow the settings of IGNet [38], which also adopts the MeanTeacher [37] framework, to set the update hyperparameter $\\alpha$ as 0.999. The iteration number in progressive exploitation is set as 5. The hyperparameters $\\tau_r, \\tau_v, \\tau_n, \\lambda_u, \\lambda_{PE}, \\lambda_{KL}$ are set as 0.9, 0.05, 0.2, 0.1, 0.1, and 0.1 respectively.", + "bbox": [ + 215, + 469, + 785, + 575 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/5f2c1f25f39dc1407cf31d7febd260a2f83f2d4c3b634e051bc4fbac2934da50.jpg", + "table_caption": [ + "Table 1: Comparative results on the SemanticKITTI [1] and ScanNet [4] datasets with varying labeled ratios. $\\mathcal{P}$ denotes the point cloud data, $\\mathcal{I}$ denotes the image data, and $\\mathcal{T}$ denotes the text data. All mIoU scores are given in percentage (\\%). The best results are in bold and the second best results are marked with underlines." + ], + "table_footnote": [], + "table_body": "
MethodModalitySemanticKITTI [1]ScanNet [4]
1%10%20%50%5%10%20%30%40%
MeanTeacher [37]P45.457.159.260.048.757.266.167.369.1
CBST [55]P48.858.359.459.748.557.865.467.369.2
CPS [3]P46.758.759.660.551.758.666.468.070.5
SSS-Net [5]P-----52.455.1--
LaserMix (Range View) [21]P43.458.859.461.4-----
LaserMix (Voxel) [21]P50.660.061.962.3-----
IGNet [38]P, I49.061.363.164.8-----
GPC [17]P54.162.062.562.854.860.566.768.971.3
LiM3D [26]P58.462.263.163.6-----
LASS3DP, I, T58.563.064.164.556.663.167.470.472.0
", + "bbox": [ + 225, + 671, + 782, + 835 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Li and Q. Dong", + "bbox": [ + 271, + 114, + 393, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Comparative Evaluation", + "text_level": 1, + "bbox": [ + 215, + 146, + 460, + 161 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We evaluate the proposed LASS3D on both outdoor-scene (SemanticKITTI [1]) and indoor-scene (ScanNet [4]) datasets in comparison to some classic semi-supervised methods [3, 37, 55] extended from the 2D domain and the state-of-the-art (SOTA) methods [5, 17, 21, 26, 38] that are specially designed for semi-supervised 3D semantic segmentation. Note that IGNet [38], LaserMix [21], and LiM3D [26] utilize some specific characters of the LiDAR point clouds, and thus cannot be applied to the ScanNet dataset. In addition, SSS-Net [5] leverages the color information of the point clouds, and thus cannot be applied to the SemantiKITTI dataset where the color information is unavailable.", + "bbox": [ + 212, + 186, + 787, + 323 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Following the settings in [17, 21, 38], we set the labeled ratio of the outdoor-scene dataset and indoor-scene dataset as $\\{1\\%, 10\\%, 20\\%, 40\\% \\}$ and $\\{5\\%, 10\\%, 20\\%, 30\\%, 50\\% \\}$ respectively for a fair comparison. The comparative results are reported in Tab. 1. As seen from this table, the proposed LASS3D outperforms all the comparative methods that use uni-modal data as input, which demonstrates the effectiveness of LASS3D. In addition, LASS3D outperforms IGNet which takes both point clouds and images as input in most cases, which indicates that the proposed language-assisted method can better boost the 3D segmentation performances and may provide some insights on the exploitation of multi-modal data in semi-supervised 3D semantic segmentation.", + "bbox": [ + 212, + 325, + 787, + 477 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Moreover, the segmentation results of LASS3D and the second-best methods are visualized in Fig. 6. As seen from this figure, LASS3D outperforms its second-best counterparts, demonstrating the effectiveness of LASS3D from the qualitative perspective.", + "bbox": [ + 212, + 479, + 787, + 542 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/aecd1f2ebeff5c4e1d5807621f131766f34d75e2733b7b363547f47ac9537ae4.jpg", + "image_caption": [ + "Fig. 6: Qualitative results on the SemanticKITTI [1] (top) and ScanNet [4] (bottom) datasets by LASS3D and the second-best methods. All models are trained with $10\\%$ labeled data. We use the red boxes to highlight the areas where LASS3D evidently outperforms the comparative methods." + ], + "image_footnote": [], + "bbox": [ + 233, + 579, + 769, + 760 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "LASS3D", + "bbox": [ + 669, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/c5c7013429c761abd06aa0ff5fa8ba965478a3ecb8792aba52cf3d034514eb09.jpg", + "table_caption": [ + "Table 2: Ablation of the loss terms." + ], + "table_footnote": [], + "table_body": "
LCElLCEuLKLLPE1%10%20%50%
✓49.658.860.761.2
✓✓53.760.161.662.8
✓✓✓55.661.462.863.7
✓✓✓✓58.563.064.164.5
", + "bbox": [ + 217, + 171, + 558, + 258 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ad8b546f4dc79c6558bdf26f70e64370029d7781f25a195533f5b8250c5dbd57.jpg", + "table_caption": [ + "Table 3: Ablation of captions." + ], + "table_footnote": [], + "table_body": "
CiCe1%10%20%50%
50.059.861.562.1
✓53.660.561.963.0
✓54.161.062.763.8
✓✓58.563.064.164.5
", + "bbox": [ + 563, + 171, + 787, + 257 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 287, + 388, + 304 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conduct ablation studies on SemanticKITTI [1] to investigate the effect of the key elements in the proposed LASS3D.", + "bbox": [ + 212, + 315, + 784, + 345 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effect of the loss terms. We investigate the effect of the loss terms and report the corresponding results in Tab. 2. As seen from the first two rows of Tab. 2, incorporating the unlabeled data into training can improve the performances. Then, as seen in the third row of Tab. 2, the $\\mathcal{L}_{KL}^u$ is introduced into training and the performances are improved, indicating that the semantic information distilled from the text-enhanced 3D features is helpful for segmentation. Finally, in the last row of Tab. 2, the progressive exploitation is adopted, and the performances are further boosted, demonstrating that the proposed progressive exploitation can make the unreliable points beneficial to the model training.", + "bbox": [ + 212, + 345, + 784, + 482 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effect of different captions. We investigate the effect of different captions and report the corresponding results in Tab. 3. As seen from this table, both the image-level captions $\\mathcal{C}_i$ and entity-level captions $\\mathcal{C}_e$ can improve the segmentation performances, indicating that the semantic information encoded in text embeddings can facilitate the point cloud segmentation. When combining $\\mathcal{C}_i$ and $\\mathcal{C}_e$ , the best performance is achieved. Because the $\\mathcal{C}_i$ provides more holistic semantic information and $\\mathcal{C}_e$ provides more fine-grained semantic information, which are complementary to each other. Thus, we utilize $\\mathcal{C}_i$ and $\\mathcal{C}_e$ in LASS3D.", + "bbox": [ + 212, + 483, + 784, + 603 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effect of the fusion strategy for the image-level embeddings $\\mathcal{E}_i$ . As mentioned in Sec. 3.2, we adopt the cross-attention operation to fuse $\\mathcal{E}_i$ with the 3D features. To investigate the effect of the fusion strategy, we replace the cross-attention operation with add and concatenation operations. The corresponding results are reported in Tab. 4. As seen from this table, the model trained with the cross-attention operation achieves the best performance. Because the image-level captions usually contain holistic semantic information and", + "bbox": [ + 212, + 604, + 784, + 709 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/7f53c3566ec6f090dc50b272c967b7f088858f946bb299c1d6591590bb0b3487.jpg", + "table_caption": [ + "Table 4: Ablation of fusion strategy.", + "Table 5: Ablation of separation strategy." + ], + "table_footnote": [], + "table_body": "
Strategy1%10%20%50%Strategy1%10%20%50%
Add54.061.262.363.5Confidence-based57.762.363.564.1
Concatenation56.262.063.463.9Variance-based57.062.162.863.5
Cross attention58.563.064.164.5Ours58.563.064.164.5
", + "bbox": [ + 218, + 763, + 795, + 835 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Li and Q. Dong", + "bbox": [ + 271, + 114, + 393, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "may have different correlations with different regions of the paired points. Simply adding or concatenating $\\mathcal{E}_i$ to all the paired 3D features leads to the lack of discrimination, while 3D semantic segmentation is a point-level prediction task that requires discriminative representations for each point. Cross attention can adaptively fuse $\\mathcal{E}_i$ with 3D features and produce more discriminative features, which makes it a more suitable fusion strategy for $\\mathcal{E}_i$ .", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of the semantic-aware adaptive fusion module (SAFM). As mentioned in Sec. 2, existing language-assisted point cloud learning methods are either projection-based or alignment-based. Project-based methods are generally designed for object-level 3D classification and are not applicable to scene-level 3D semantic segmentation. To verify the effectiveness of the proposed SAFM, we replace it with an alignment-based method (e.g., the semantic consistency regularization in CLIP2Scene [2]). The corresponding results are reported in Tab. 6. As seen from this table, the model trained with SAFM outperforms the model trained with semantic consistency regularization. Probably because a large distribution discrepancy exists between the 3D features and text embeddings, simply aligning the 3D features to their corresponding text embeddings may result in the loss of some inherent geometric information. SAFM remedies this defect by adaptive fusion, and thus achieves better performance.", + "bbox": [ + 212, + 238, + 792, + 434 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of the separation strategy. As mentioned in Sec. 3.3, we adopt both confidences and variances to measure the reliability of the predictions and separate the unlabeled points into the reliable set and unreliable set. Here we evaluate the effectiveness of our separation strategy. The corresponding results are reported in Tab. 5. As seen from this table, the model trained with our strategy achieves the best performance. Probably because the confidence-based strategy alone is unable to filter out the highly confident wrong predictions, which may confuse the training. The variance-based strategy separates points according to the prediction consistency in different branches, which could be complementary to the confidence-based strategy in measuring reliability. Thus, we combine the above two strategies to measure reliability in a more comprehensive way, which can facilitate the model training.", + "bbox": [ + 212, + 435, + 787, + 617 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of the iteration number. The iteration number of progressive exploitation determines the utilization degree of the unreliable data. Here, we evaluate LASS3D with the iteration number set as $\\{0,1,2,3,5,7,9,11\\}$ , and the corresponding results are shown in Fig. 7. As seen from this figure, the performances are generally improved with the increase of iteration number. Because a larger iteration number means that the exploitation of the unreliable data is more thor-", + "bbox": [ + 212, + 618, + 787, + 708 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/a7cc387d7b14fa082db9e69e60c320d14fd2c4ea991a992212d0e2bfba8a2673.jpg", + "table_caption": [ + "Table 6: Ablation of semantic-aware adap- Table 7: Ablation of the proposed pro- tive fusion module (SAFM). progressive exploitation (PE)." + ], + "table_footnote": [], + "table_body": "
Strategy1%10%20%50%Model1%10%20%50%
Alignment-based53.060.161.862.2GPC [17]54.162.062.562.8
SAFM58.563.064.164.5GPC with PE54.762.863.163.5
", + "bbox": [ + 217, + 777, + 794, + 835 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "LASS3D", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7c22a04ed498dce9ad861404596b94fd1111fb421b03788c12325295e3c01586.jpg", + "image_caption": [ + "Fig. 7: Ablation of the iteration number. Note that the iteration here denotes the negative learning iteration in progressive exploitation, rather than the training iteration." + ], + "image_footnote": [], + "bbox": [ + 380, + 146, + 612, + 252 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "ough and thus brings better performance. However, when the iteration number reaches a certain level, the segmentation performance improves slightly. Probably because many unreliable points turn into reliable points during the progressive exploitation process, with most of their confident negative classes discarded. In addition, more iterations inevitably bring more training costs. Thus, we set the iteration number as 5 for a trade-off between the performance and training cost.", + "bbox": [ + 215, + 324, + 784, + 414 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effectiveness of the progressive exploitation. Considering that the proposed progressive exploitation strategy only requires the prediction confidence and variances, thus it can be seamlessly integrated into some existing methods. We integrate the progressive exploitation strategy into GPC [17] and report the comparative results in Tab. 7. As seen from this table, the proposed progressive exploitation strategy can further boost the performance of GPC, demonstrating the effectiveness of the proposed strategy for exploiting unreliable points.", + "bbox": [ + 215, + 415, + 784, + 518 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.5 Limitation", + "text_level": 1, + "bbox": [ + 217, + 540, + 346, + 554 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The proposed LASS3D has demonstrated its effectiveness in segmentation, however, it still has limitations. For example, LASS3D uses images as the bridge to connect the text data and point clouds, which is a complex operation and may bring accumulated errors in point-text matching. A more direct and accurate method to construct the point-text pairs is in demand.", + "bbox": [ + 215, + 561, + 782, + 637 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 217, + 659, + 356, + 674 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose a language-assisted semi-supervised 3D semantic segmentation method named LASS3D. In LASS3D, a semantic-aware adaptive fusion module is explored to fuse the 3D features with the embeddings of multi-level captions generated by language-vision models, which can inject the semantic information into the 3D features and thus facilitate 3D segmentation. In addition, a progressive exploitation strategy is explored for unlabeled unreliable data, which can further boost performance and can be seamlessly embedded into other methods. Experimental results on outdoor and indoor datasets show that LASS3D outperforms the comparative methods in most cases.", + "bbox": [ + 215, + 688, + 784, + 824 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Li and Q. Dong", + "bbox": [ + 271, + 114, + 393, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (Grant Nos. 61991423, 62376269, 62073199); the Strategic Priority Research Program of the Chinese Academy of Sciences (Grant No. XDA27040811); the Open Project Program of Key Laboratory of Industrial Internet and Big Data, China National Light Industry, Beijing Technology and Business University.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 258, + 325, + 275 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semanticickitti: A dataset for semantic scene understanding of lidar sequences. In: ICCV. pp. 9297-9307 (2019)", + "2. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: Clip2scene: Towards label-efficient 3d scene understanding by clip. In: CVPR. pp. 7020-7030 (2023)", + "3. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: CVPR. pp. 2613-2622 (2021)", + "4. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR. pp. 5828-5839 (2017)", + "5. Deng, S., Dong, Q., Liu, B., Hu, Z.: Superpoint-guided semi-supervised semantic segmentation of 3d point clouds. In: ICRA. pp. 9214-9220 (2022)", + "6. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: CVPR. pp. 7010-7019 (2023)", + "7. Fan, S., Dong, Q., Zhu, F., Lv, Y., Ye, P., Wang, F.: Scf-net: Learning spatial contextual features for large-scale point cloud segmentation. In: CVPR. pp. 14499-14508 (2021)", + "8. Graham, B., Engelcke, M., van der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. CVPR pp. 9224-9232 (2018)", + "9. Guo, Z., Zhang, R., Qiu, L., Ma, X., Miao, X., He, X., Cui, B.: Calip: Zero-shot enhancement of clip with parameter-free attention. In: AAAI. vol. 37, pp. 746-754 (2023)", + "0. He, W., Jamonnak, S., Gou, L., Ren, L.: Clip-s4: Language-guided self-supervised semantic segmentation. In: CVPR. pp. 11207-11216 (2023)", + "1. Hegde, D., Valanarasu, J.M.J., Patel, V.: Clip goes 3d: Leveraging prompt tuning for language grounded 3d recognition. In: ICCVW. pp. 2028-2038 (2023)", + "2. Hess, G., Tondersski, A., Petersson, C., Åström, K., Svensson, L.: Lidarclip or: How i learned to talk to point clouds. In: WACV. pp. 7438-7447 (2024)", + "3. Hu, Q., Yang, B., Fang, G., Guo, Y., Leonardis, A., Trigoni, N., Markham, A.: Sqn: Weakly-supervised semantic segmentation of large-scale 3d point clouds. In: ECCV. pp. 600-619. Springer (2022)", + "4. Huang, R., Pan, X., Zheng, H., Jiang, H., Xie, Z., Wu, C., Song, S., Huang, G.: Joint representation learning for text and 3d point cloud. PR 147, 110086 (2024)", + "5. Huang, T., Dong, B., Yang, Y., Huang, X., Lau, R.W., Ouyang, W., Zuo, W.: Clip2point: Transfer clip to point cloud classification with image-depth pretraining. In: ICCV. pp. 22157-22167 (2023)", + "6. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML. pp. 4904-4916 (2021)" + ], + "bbox": [ + 225, + 287, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "LASS3D", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Jiang, L., Shi, S., Tian, Z., Lai, X., Liu, S., Fu, C.W., Jia, J.: Guided point contrastive learning for semi-supervised point cloud semantic segmentation. In: ICCV. pp. 6423-6432 (2021)", + "18. Jiao, S., Wei, Y., Wang, Y., Zhao, Y., Shi, H.: Learning mask-aware CLIP representations for zero-shot segmentation. In: NeurIPS (2023)", + "19. Kim, Y., Yim, J., Yun, J., Kim, J.: Nlnl: Negative learning for noisy labels. In: ICCV. pp. 101-110 (2019)", + "20. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV. pp. 4015-4026 (2023)", + "21. Kong, L., Ren, J., Pan, L., Liu, Z.: Lasermix for semi-supervised lidar semantic segmentation. In: CVPR. pp. 21705-21715 (2023)", + "22. Lai, X., Chen, Y., Lu, F., Liu, J., Jia, J.: Spherical transformer for lidar-based 3d recognition. In: CVPR. pp. 17545-17555 (2023)", + "23. Lai, X., Liu, J., Jiang, L., Wang, L., Zhao, H., Liu, S., Qi, X., Jia, J.: Stratified transformer for 3d point cloud segmentation. In: CVPR. pp. 8500-8509 (2022)", + "24. Li, J., Dong, Q.: Open-set semantic segmentation for point clouds via adversarial prototype framework. In: CVPR. pp. 9425-9434 (2023)", + "25. Li, J., Dong, Q.: Density-guided semi-supervised 3d semantic segmentation with dual-space hardness sampling. In: CVPR. pp. 3260-3269 (2024)", + "26. Li, L., Shum, H.P., Breckon, T.P.: Less is more: Reducing task and model complexity for 3d point cloud semantic segmentation. In: CVPR. pp. 9361-9371 (2023)", + "27. Lin, Z., Geng, S., Zhang, R., Gao, P., de Melo, G., Wang, X., Dai, J., Qiao, Y., Li, H.: Frozen clip models are efficient video learners. In: ECCV. pp. 388-404. Springer (2022)", + "28. Liu, L., Zhuang, Z., Huang, S., Xiao, X., Xiang, T., Chen, C., Wang, J., Tan, M.: Cpcm: Contextual point cloud modeling for weakly-supervised point cloud semantic segmentation. In: ICCV. pp. 18413-18422 (2023)", + "29. Liu, Z., Qi, X., Fu, C.W.: One thing one click: A self-training approach for weakly supervised 3d semantic segmentation. In: CVPR. pp. 1726-1736 (2021)", + "30. Lu, Y., Xu, C., Wei, X., Xie, X., Tomizuka, M., Keutzer, K., Zhang, S.: Open-vocabulary point-cloud object detection without 3d annotation. In: CVPR. pp. 1190-1199 (2023)", + "31. Ma, W., Li, S., Zhang, J., Liu, C.H., Kang, J., Wang, Y., Huang, G.: Borrowing knowledge from pre-trained language model: A new data-efficient visual learning paradigm. In: ICCV. pp. 18786-18797 (2023)", + "32. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023)", + "33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)", + "34. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Denseclip: Language-guided dense prediction with context-aware prompting. In: CVPR. pp. 18082-18091 (2022)", + "35. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., Zeng, Z., Zhang, H., Li, F., Yang, J., Li, H., Jiang, Q., Zhang, L.: Grounded sam: Assembling open-world models for diverse visual tasks (2024)", + "36. Tang, L., Chen, Z., Zhao, S., Wang, C., Tao, D.: All points matter: entropy-regularized distribution alignment for weakly-supervised 3d segmentation. NeurIPS 36 (2024)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Li and Q. Dong", + "bbox": [ + 271, + 114, + 393, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NIPS 30 (2017)", + "38. Unal, O., Dai, D., Hoyer, L., Can, Y.B., Van Gool, L.: 2d feature distillation for weakly-and semi-supervised 3d semantic segmentation. In: WACV. pp. 7336-7345 (2024)", + "39. Wang, Y., Huang, S., Gao, Y., Wang, Z., Wang, R., Sheng, K., Zhang, B., Liu, S.: Transferring clip's knowledge into zero-shot point cloud semantic segmentation. In: ACM MM. pp. 3745-3754 (2023)", + "40. Wang, Y., Wang, H., Shen, Y., Fei, J., Li, W., Jin, G., Wu, L., Zhao, R., Le, X.: Semi-supervised semantic segmentation using unreliable pseudo-labels. In: CVPR. pp. 4248-4257 (2022)", + "41. Wu, X., Jiang, L., Wang, P.S., Liu, Z., Liu, X., Qiao, Y., Ouyang, W., He, T., Zhao, H.: Point transformer v3: Simpler, faster, stronger. In: CVPR (2024)", + "42. Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point transformer v2: Grouped vector attention and partition-based pooling. In: NeurIPS. vol. 35, pp. 33330-33342 (2022)", + "43. Xiang, P., Wen, X., Liu, Y.S., Zhang, H., Fang, Y., Han, Z.: Retro-fpn: Retrospective feature pyramid network for point cloud semantic segmentation. In: ICCV. pp. 17826-17838 (2023)", + "44. Xue, L., Gao, M., Xing, C., Martin-Martín, R., Wu, J., Xiong, C., Xu, R., Niebles, J.C., Savarese, S.: Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In: CVPR. pp. 1179-1189 (2023)", + "45. Yao, L., Han, J., Wen, Y., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, C., Xu, H.: Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. NeurIPS 35, 9125-9138 (2022)", + "46. Yu, W., Liu, Y., Hua, W., Jiang, D., Ren, B., Bai, X.: Turning a clip model into a scene text detector. In: CVPR. pp. 6978-6988 (2023)", + "47. Zeng, Y., Jiang, C., Mao, J., Han, J., Ye, C., Huang, Q., Yeung, D.Y., Yang, Z., Liang, X., Xu, H.: Clip2: Contrastive language-image-point pretraining from real-world point cloud data. In: CVPR. pp. 15244-15253 (2023)", + "48. Zhang, J., Fan, G., Wang, G., Su, Z., Ma, K., Yi, L.: Language-assisted 3d feature learning for semantic scene understanding. In: AAAI. pp. 3445-3453 (2023)", + "49. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: CVPR. pp. 8552-8562 (2022)", + "50. Zhang, Y., Huo, X., Chen, T., Wu, S., Wong, H.S.: Exploring intra-class variation factors with learnable cluster prompts for semi-supervised image synthesis. In: CVPR. pp. 7392-7401 (2023)", + "51. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV. pp. 16259-16268 (2021)", + "52. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: ECCV. pp. 696-712. Springer (2022)", + "53. Zhu, X., Zhang, R., He, B., Guo, Z., Zeng, Z., Qin, Z., Zhang, S., Gao, P.: Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In: ICCV. pp. 2639-2650 (2023)", + "54. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In: CVPR. pp. 9939-9948 (2021)", + "55. Zou, Y., Yu, Z., Kumar, B., Wang, J.: Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In: ECCV. pp. 289-305 (2018)" + ], + "bbox": [ + 212, + 146, + 787, + 811 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "LASS3D", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/LASS3D_ Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation/c8ab9b1f-da9d-4178-ac58-4d667a9330b3_model.json b/2024/LASS3D_ Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation/c8ab9b1f-da9d-4178-ac58-4d667a9330b3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bb1967c76b2cea904f776c583748fea89257eb78 --- /dev/null +++ b/2024/LASS3D_ Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation/c8ab9b1f-da9d-4178-ac58-4d667a9330b3_model.json @@ -0,0 +1,2544 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.144, + 0.782, + 0.198 + ], + "angle": 0, + "content": "LASS3D: Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation" + }, + { + "type": "text", + "bbox": [ + 0.363, + 0.225, + 0.642, + 0.243 + ], + "angle": 0, + "content": "Jianan \\(\\mathrm{Li}^{1,2}\\) and Qiulei Dong" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.252, + 0.76, + 0.266 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Artificial Intelligence, University of Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.266, + 0.774, + 0.295 + ], + "angle": 0, + "content": "\\(^{2}\\) State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.295, + 0.761, + 0.324 + ], + "angle": 0, + "content": "3 Center for Excellence in Brain Science and Intelligence Technology, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.324, + 0.684, + 0.337 + ], + "angle": 0, + "content": "lijianan211@mails.ucas.ac.cn, qldong@nlpr.ac.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.376, + 0.744, + 0.683 + ], + "angle": 0, + "content": "Abstract. Precisely annotating large-scale 3D datasets for point cloud segmentation is laborious. To alleviate the annotation burden, several semi-supervised 3D segmentation methods have been proposed in literature. However, two issues remain to be tackled: 1) The utilization of large language-vision models (LVM) in semi-supervised 3D semantic segmentation remains under-explored. 2) The unlabeled points with low-confidence predictions are directly discarded by existing methods. Taking these two issues into consideration, we propose a language-assisted semi-supervised 3D semantic segmentation method named LASS3D, which is built upon the commonly used MeanTeacher framework. In LASS3D, we use two off-the-shelf LVM to generate multi-level captions and leverage the images as the bridge to connect the text data and point clouds. Then, a semantic-aware adaptive fusion module is explored in the student branch, where the semantic information encoded in the embeddings of multi-level captions is injected into 3D features by adaptive fusion and then the semantic information in the text-enhanced 3D features is transferred to the teacher branch by knowledge distillation. In addition, a progressive exploitation strategy is explored for the unreliable points in the teacher branch, which can effectively exploit the information encapsulated in unreliable points via negative learning. Experimental results on both outdoor and indoor datasets demonstrate that LASS3D outperforms the comparative methods in most cases." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.695, + 0.741, + 0.71 + ], + "angle": 0, + "content": "Keywords: 3D semantic segmentation \\(\\cdot\\) Semi-supervised segmentation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.737, + 0.377, + 0.753 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.77, + 0.788, + 0.816 + ], + "angle": 0, + "content": "3D Semantic Segmentation [7,22-24,42,43,51] is an important task for perceiving real-world environments, and intensive efforts have been devoted to this task in recent years. However, most of the existing works [8,22,41,42,51,54] in literature" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.383, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.13 + ], + "angle": 0, + "content": "J. Li and Q. Dong" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.776, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.285, + 0.788, + 0.327 + ], + "angle": 0, + "content": "Fig. 1: Illustration of the proportions of reliable and unreliable points predicted by GPC [17] on ScanNet [4]. The horizontal axis denotes the labeled ratio and training iteration. The vertical axis denotes the proportion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.784, + 0.388 + ], + "angle": 0, + "content": "are fully-supervised methods, which require time-consuming and labor-intensive data annotation for training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.785, + 0.463 + ], + "angle": 0, + "content": "This issue encourages researchers to investigate semi-supervised 3D semantic segmentation, where only a small amount of densely labeled data and a large amount of unlabeled data are provided for training. Existing works [5,17,21,26, 38] on semi-supervised 3D semantic segmentation have shown their effectiveness, but there is ample room for improvement in the following two aspects:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.584 + ], + "angle": 0, + "content": "1) Utilization of language-vision models (LVM): Recently, LVM [16,33] trained on web-crawled image-text pairs has been applied in some downstream tasks [2, 31, 39, 50] to boost performance, owing to its strong representation ability. Text embeddings extracted by the LVM generally contain higher-level abstraction of visual concepts with rich semantic information, which is complementary to the 3D features that mainly contain geometric information and thus can facilitate mining the semantic information of unlabeled data. To our best knowledge, LVM has not been applied in semi-supervised 3D semantic segmentation yet." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.735 + ], + "angle": 0, + "content": "2) Exploitation of unreliable points (unlabeled points with low-confidence predictions): The proportions of reliable and unreliable points predicted by the state-of-the-art semi-supervised 3D semantic segmentation method GPC [17] on ScanNet [4] dataset are illustrated in Fig. 1. As seen from this figure, a large number of unlabeled points are deemed unreliable, especially at the early training stage. Existing works [17,21] generally adopt a confidence-based filtering strategy to remove the unreliable points for guaranteeing the effectiveness and stability of the training process. However, simply discarding them will inevitably lose some useful information. Thus, a proper learning strategy is expected to be specially designed to exploit these unreliable points effectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Taking the above two aspects into consideration, we propose a method named LASS3D for semi-supervised 3D semantic segmentation, where the LVM and unreliable points are effectively exploited. The proposed LASS3D is built upon the MeanTeacher [37] framework, where the weights of the teacher branch are updated by the exponential moving average (EMA) of the student's weights." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In the student branch, two LVM (i.e., KOSMOS-2 [32] and GroundedSAM [35]) are used to generate image-level and entity-level captions, and the images" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LASS3D" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.236 + ], + "angle": 0, + "content": "are leveraged as the bridge to connect the text data and point clouds. Then, a semantic-aware adaptive fusion module (SAFM) is designed to assist point cloud learning with the semantic knowledge encoded in text data. Specifically, multi-level point-text pairs are constructed in SAFM to adaptively fuse the 3D features and the text embeddings of the image-level and entity-level captions, which could facilitate capturing both holistic and fine-grained semantic information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.554 + ], + "angle": 0, + "content": "In the teacher branch, a fusion adapter is utilized to narrow the distribution gap between the 3D features extracted by the backbone network and the text-enhanced 3D features in the student branch. Combined with the EMA strategy, the semantic information encoded in the text-enhanced 3D features is distilled from the student branch to the teacher branch. Note that we inject semantic information from text embeddings into the 3D features by fusion rather than directly mapping the semantic labels from pixels to points. Because the semantic labels generated by the open-world segmentation models (e.g., SAM [20]) are generally in a one-hot form, which provide hard supervision. However, some generated synonymous labels may not exist in our desired label set (e.g., the generated label is 'couch' but the ground-truth label is 'sofa'), and incorrect hard supervision may adversely affect the point cloud segmentation. Adaptively fusing the text embeddings and 3D features can mitigate this negative effect to some extent. Because the strong LVM can encode synonymous words into similar embeddings and thus can inject similar semantic information into 3D features. To effectively exploit the unlabeled points with unreliable predictions, a progressive exploitation strategy is designed based on the fact that determining the least likely category for an unlabeled sample is generally easier and more accurate than determining its most likely category. In addition, the proposed progressive exploitation strategy can be seamlessly integrated into some existing methods and further boost their performances, which is demonstrated in Sec. 4.4." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.555, + 0.662, + 0.57 + ], + "angle": 0, + "content": "In summary, the contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.576, + 0.787, + 0.635 + ], + "angle": 0, + "content": "- We propose a semantic-aware adaptive fusion module (SAFM) to consolidate point cloud segmentation with the semantic information encoded in text embeddings. The multi-level fusion shows the feasibility and effectiveness of language-assisted 3D semantic segmentation in semi-supervised settings." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.636, + 0.785, + 0.679 + ], + "angle": 0, + "content": "- We propose a progressive exploitation strategy for the unlabeled unreliable points. This strategy is beneficial to fully exploit the unlabeled data and can be seamlessly embedded into some existing methods." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.68, + 0.785, + 0.739 + ], + "angle": 0, + "content": "- We propose the LASS3D for semi-supervised 3D semantic segmentation by integrating the above SAFM and progressive exploitation strategy. Its effectiveness is demonstrated by the results on both outdoor-scene and indoor-scene public datasets in Sec. 4." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.576, + 0.787, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.759, + 0.388, + 0.775 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.788, + 0.623, + 0.803 + ], + "angle": 0, + "content": "2.1 Semi-supervised 3D Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "To alleviate the annotation burden, label-efficient 3D semantic segmentation (including weakly-supervised segmentation [13,28,29,36] and semi-supervised seg" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.396, + 0.129 + ], + "angle": 0, + "content": "J. Li and Q. Dong" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.373 + ], + "angle": 0, + "content": "mentation [5,17,21,25,26,38]) has drawn increasing interest among researchers. In this work, we delve into semi-supervised 3D semantic segmentation, which aims at utilizing a small number of densely labeled data and a large number of unlabeled data for model training. The core of semi-supervised 3D segmentation lies in exploiting the unlabeled data to the fullest. For example, Deng et al. [5] proposed to optimize the pseudo labels of the unlabeled data with the superpoints generated by geometry-based and color-based region-growing algorithms. Jiang et al. [17] proposed a label-guided contrastive loss for the unlabeled data, which could enhance the feature representation ability of the model. Kong et al. [21] proposed to leverage the spatial prior of LiDAR point clouds and use a mixing operation to provide supervisory information for the unlabeled data. Li et al. [26] proposed to utilize the reflectivity-prior descriptors to generate high-quality pseudo labels for the unlabeled data. Unal et al. [38] proposed to distill high-level feature information from a synthetically trained 2D network for data-efficient LiDAR semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.435 + ], + "angle": 0, + "content": "However, these methods neglect the usage of text data that contains rich semantic information. To our best knowledge, LASS3D is the first work to investigate the language-assisted semi-supervised 3D semantic segmentation, where the powerful representation ability of language-vision models is exploited." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.456, + 0.52, + 0.472 + ], + "angle": 0, + "content": "2.2 Language-assisted 3D Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.6 + ], + "angle": 0, + "content": "Large language-vision models (LVM) [16, 33] have drawn growing attention in computer vision recently, attributed to their powerful representation ability. For example, CLIP [33], which is trained on massive paired text-image data via contrastive learning, has prevailed in various downstream visual tasks, such as semantic segmentation [10, 18, 34, 48, 52], object detection [30, 45, 46], and video recognition [27]. Recent advanced methods that focused on language-assisted point cloud learning could be roughly divided into two types: the projection-based methods and the alignment-based methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.788, + 0.647 + ], + "angle": 0, + "content": "The projection-based methods [9,15,49,53] project the point clouds into the 2D plane so that the image encoder in the LVM can be seamlessly utilized and the text-image alignment of the LVM can be leveraged for downstream tasks." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.675, + 0.517, + 0.778 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.784, + 0.446, + 0.797 + ], + "angle": 0, + "content": "(a) Pipeline of PointCLIP." + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.677, + 0.787, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.784, + 0.766, + 0.797 + ], + "angle": 0, + "content": "(b) Alignment process of CLIP2Scene." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.808, + 0.787, + 0.838 + ], + "angle": 0, + "content": "Fig. 2: Illustration of the pipeline of a projection-based method (e.g., PointCLIP [49]) and the alignment process of an alignment-based method (e.g., CLIP2Scene [2])." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.671, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LASS3D" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "Fig. 2a illustrates the general pipeline of PointCLIP [49], it classifies the point clouds according to the similarities between the text embeddings and features of the projected images. Nevertheless, the projection-based methods are generally designed for object-level 3D classification. In addition, they usually suffer from the loss of geometric information during the projection process, thus resulting in suboptimal performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.328 + ], + "angle": 0, + "content": "The alignment-based methods [2,6,11,12,14,39,44,47] enforce the consistency among the multi-modal features. Fig. 2b illustrates the alignment process of CLIP2Scene [2], it uses contrastive learning to align the 3D features to their corresponding text embeddings. However, directly aligning multi-modal features may result in suboptimal results, due to the distribution discrepancy between the 3D features and text embeddings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.787, + 0.39 + ], + "angle": 0, + "content": "Unlike existing methods that are either projection-based or alignment-based, we propose a fusion-based method to adaptively fuse the embeddings of multi-level captions and 3D features, which can alleviate the loss of geometric information and mitigate the negative effect brought by the distribution gap." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.411, + 0.381, + 0.429 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.442, + 0.37, + 0.456 + ], + "angle": 0, + "content": "3.1 Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.528 + ], + "angle": 0, + "content": "The architecture of the proposed LASS3D is illustrated in Fig. 3. It is built upon the MeanTeacher [37] framework where the teacher branch provides the supervisory signals for the student branch and is updated by the exponential mean average (EMA) of the student's weights:" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.536, + 0.786, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\theta_ {t e a c h e r} ^ {t + 1} = \\alpha \\theta_ {t e a c h e r} ^ {t} + (1 - \\alpha) \\theta_ {s t u d e n t} ^ {t}, \\tag {1}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.58, + 0.784, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.794, + 0.788, + 0.838 + ], + "angle": 0, + "content": "Fig. 3: Architecture of the proposed LASS3D. \\(\\mathcal{C}_i\\) and \\(\\mathcal{C}_e\\) stand for the image-level captions and entity-level captions. \\(\\mathcal{E}_i\\) and \\(\\mathcal{E}_e\\) are the embeddings of \\(\\mathcal{C}_i\\) and \\(\\mathcal{C}_e\\) respectively. \\(/ / /\\) denotes the detach operation which is used to stop the gradient." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.396, + 0.129 + ], + "angle": 0, + "content": "J. Li and Q. Dong" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "where \\(\\theta_{teacher}^{t}\\) and \\(\\theta_{student}^{t}\\) denote the weights of the teacher branch and student branch at time step \\(t\\), and \\(\\alpha\\) is the update hyperparameter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.785, + 0.297 + ], + "angle": 0, + "content": "The student branch takes the point clouds and their corresponding images as input. The images are fed into the LVM, (i.e., KOSMOS-2 [35] and GroundedSAM [32]) to produce image-level \\( \\mathcal{C}_i \\) and entity-level captions \\( \\mathcal{C}_e \\). Then, the multi-level captions are tokenized by the text encoder of CLIP [33] to generate their corresponding text embeddings. The text embeddings and 3D features extracted by the 3D backbone network are fused in the semantic-aware adaptive fusion module to obtain the text-enhanced 3D features, which are then used for segmentation in the student branch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.3, + 0.787, + 0.497 + ], + "angle": 0, + "content": "The teacher branch takes the same point clouds as input. The 3D features extracted by the 3D backbone network are aligned with the text-enhanced 3D features through a fusion adapter. Then, the classifier takes the aligned 3D features as input and outputs their corresponding prediction scores. A dual-score separation strategy is adopted to select reliable points and unreliable points from the unlabeled points. The predictions of the reliable points are regarded as the pseudo labels of their corresponding points in the student branch, where the cross-entropy loss \\(\\mathcal{L}_{CE}^{u}\\) is calculated. The predictions of the unreliable points in the student branch are used for the progressive exploitation, where the negative learning loss \\(\\mathcal{L}_{PE}^{u}\\) is calculated. Besides, the predictions of the labeled points in the student branch and their ground-truth labels are used to calculate the cross-entropy loss \\(\\mathcal{L}_{CE}^{l}\\), and the features output by the fusion adapter are aligned to the text-enhanced 3D features by a Kullback-Leibler divergence loss \\(\\mathcal{L}_{KL}^{u}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.785, + 0.527 + ], + "angle": 0, + "content": "Note that the text data is only required in the training stage. In the inference stage, only the teacher branch is used and it takes the point clouds as input." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.553, + 0.6, + 0.569 + ], + "angle": 0, + "content": "3.2 Semantic-aware Adaptive Fusion Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.785, + 0.657 + ], + "angle": 0, + "content": "The architecture of the semantic-aware adaptive fusion module (SAFM) is illustrated in Fig. 4, it takes the embeddings of multi-level captions (i.e., image-level embeddings \\(\\mathcal{E}_i\\) and entity-level embeddings \\(\\mathcal{E}_e\\)) and 3D features \\(f_{3D}\\) as input and outputs the text-enhanced 3D features \\(f_{3D}^t\\), which are then fed into the classifier for segmentation and used as the supervision for knowledge distillation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.786, + 0.794 + ], + "angle": 0, + "content": "The proposed SAFM is used to inject the semantic information encoded in multi-level captions into the 3D features. Specifically, the image-level captions \\(\\mathcal{C}_i\\) and entity-level captions \\(\\mathcal{C}_e\\) are generated by KOSMOS-2 [32] and GroundedSAM [35] respectively, and their corresponding embeddings are extracted by the text encoder of CLIP [33]. To guarantee the accuracy of multi-modal fusion, we construct the point-text pairs by using the images as the bridge. Then, given the point-text pairs, their corresponding features are adaptively fused to produce the text-enhanced 3D features. In this subsection, we will introduce the above two key operations of SAFM in detail." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Point-text pair construction. We use the images as the bridge to connect the text data and point cloud data. Specifically, the multi-level pixel-text pairs can be obtained by the language-vision models, i.e., the image-level captions" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LASS3D" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.143, + 0.758, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.274, + 0.788, + 0.332 + ], + "angle": 0, + "content": "Fig. 4: Architecture of the proposed semantic-aware adaptive fusion module. \\( f_{3D} \\), \\( \\hat{f}_{3D} \\), and \\( f_{3D}^{t} \\) denote the extracted 3D features, aligned 3D features, and text-enhanced 3D features. \\( \\mathcal{E}_i \\) and \\( \\mathcal{E}_e \\) are the embeddings of image-level and entity-level captions. \\( \\oplus \\) denotes the add operation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.364, + 0.785, + 0.41 + ], + "angle": 0, + "content": "correspond to all pixels of the given images, and the entity-level captions correspond to the pixels of the segmented regions. The pixel-point pairs can be easily obtained with the extrinsic and intrinsic parameters of the cameras:" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.424, + 0.786, + 0.443 + ], + "angle": 0, + "content": "\\[\n[ u, v, 1 ] ^ {\\mathrm {T}} = \\mathbf {K M} [ x, y, z, 1 ] ^ {\\mathrm {T}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.785, + 0.517 + ], + "angle": 0, + "content": "where \\([u,v,1]^{\\mathrm{T}}\\) represents the homogeneous coordinates of a pixel, \\([x,y,z,1]^{\\mathrm{T}}\\) represents its corresponding 3D point position in world coordinates, \\(\\mathbf{K}\\) and \\(\\mathbf{M}\\) are the intrinsic and extrinsic parameters of the cameras. With the multi-level pixel-text and pixel-point pairs, the multi-level point-text pairs can be obtained." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.786, + 0.701 + ], + "angle": 0, + "content": "Adaptive fusion. To effectively utilize the multi-level point-text pairs, an adaptive fusion strategy is proposed. As shown in Fig. 4, the dimension of the 3D features \\( f_{3D} \\) is firstly expanded by an MLP layer to align the embedding space of the text encoder, given that the text encoder in LASS3D is fixed. Since the image-level captions may have different correlations with different regions of the paired points, cross attention is performed between the aligned 3D features \\( \\hat{f}_{3D} \\) and image-level embeddings \\( \\mathcal{E}_i \\), with \\( \\hat{f}_{3D} \\) serving as the query vectors and \\( \\mathcal{E}_i \\) serving as the key vectors and value vectors. Then, considering that the points related to the same entity-level captions share the same semantic information, we simply add the entity-level embeddings \\( \\mathcal{E}_e \\) with their corresponding 3D features. Finally, an MLP layer is used to map the fused features to their original dimension and output the text-enhanced 3D features \\( f_{3D}^t \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.702, + 0.787, + 0.78 + ], + "angle": 0, + "content": "As shown in Fig. 3, the text-enhanced 3D features \\( f_{3D}^{t} \\) are not only fed into the classifier of the student branch for segmentation but also serve as the supervisory signals for knowledge distillation in the teacher branch. The features \\( f_{ada} \\) output by the fusion adapter in the teacher branch are aligned to their corresponding \\( f_{3D}^{t} \\) through \\( \\mathcal{L}_{KL}^{u} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.793, + 0.786, + 0.812 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {K L} ^ {u} = \\mathrm {K L} \\left(f _ {3 D} ^ {t} \\mid \\mid f _ {a d a}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.825, + 0.633, + 0.841 + ], + "angle": 0, + "content": "where \\(\\mathrm{KL}(\\cdot)\\) denotes the Kullback-Leibler divergence loss." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.396, + 0.129 + ], + "angle": 0, + "content": "J. Li and Q. Dong" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.147, + 0.785, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.256, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Fig. 5: Illustration of the progressive exploitation for the unreliable point \\( P_{url}^{i} \\) at the first two iterations. \\( S^{i} \\) denotes the prediction confidence of \\( P_{url}^{i} \\) and \\( \\hat{y}_i \\) denotes its negative label. The selected negative class (blue font) is used for negative learning and the used negative class (marked with red dashed box) is discarded in the subsequent iterations. Note that some unreliable points may turn into reliable points during progressive exploitation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.37, + 0.54, + 0.385 + ], + "angle": 0, + "content": "3.3 Progressive Exploitation Strategy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.784, + 0.455 + ], + "angle": 0, + "content": "The progressive exploitation strategy is proposed to mine the useful information encapsulated in the unlabeled unreliable points. It contains two key steps: dual-score separation and progressive negative learning. In this subsection, we will introduce the above two steps in detail." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.785, + 0.547 + ], + "angle": 0, + "content": "Dual-score separation. Unlike GPC [17] and LaserMix [21] which only use the confidence of predictions to measure the reliability, we propose a dual-score separation strategy that additionally introduces the variances among the predictions of different branches into reliability measurement. Specifically, the unlabeled points \\( P_{u} \\) in the teacher branch are separated into the reliable set \\( P_{rl} \\) and unreliable set \\( P_{url} \\) based on both confidence and variances:" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.557, + 0.604, + 0.573 + ], + "angle": 0, + "content": "\\[\nP _ {r l} = B \\cdot P _ {u}, P _ {u r l} = (1 - B) \\cdot P _ {u},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.577, + 0.785, + 0.618 + ], + "angle": 0, + "content": "\\[\nB = \\mathbb {1} \\left[ \\sum_ {c = 1} ^ {C} \\mathbb {1} \\left[ S _ {c} \\geq \\tau_ {r} \\right] \\cdot \\mathbb {1} \\left[ \\sigma_ {c} \\leq \\tau_ {v} \\right] > 0 \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.785, + 0.688 + ], + "angle": 0, + "content": "where \\( B \\) is a binary mask to select \\( P_{rl} \\) and \\( P_{url} \\), \\( \\mathbb{1} \\) is the indicator function, \\( C \\) denotes the number of classes, \\( S_{c} \\) denotes the prediction confidences of class \\( c \\) in the teacher branch, \\( \\sigma_{c} \\) denotes the prediction variances of class \\( c \\) within the teacher and student branches, \\( \\tau_{r} \\) and \\( \\tau_{v} \\) are two pre-determined thresholds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.785, + 0.749 + ], + "angle": 0, + "content": "The predictions of reliable points are regarded as the pseudo labels of their corresponding points in the student branch, which are used for supervised training, as shown in Eq. (8). The unreliable points are fully exploited via progressive negative learning, which is elaborated below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Progressive negative learning. Enlightened by the core idea of conventional negative learning [19, 40], we choose to mine information from the classes that the unreliable points do not belong to, rather than struggling to correctly tell which class they belong to. However, conventional methods only choose one negative class (the class with the lowest prediction score) for negative learning, which provides limited information and may cause unstable training due to the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.671, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LASS3D" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "dynamic negative labels. Instead, we conduct negative learning in a progressive manner, with more negative classes selected." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.177, + 0.784, + 0.237 + ], + "angle": 0, + "content": "The progressive exploitation at the first two iterations is illustrated in Fig. 5. As shown in this figure, we maintain a list for each unreliable point to record its selected negative classes. In each iteration, one negative class is selected for each unreliable point:" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.245, + 0.784, + 0.282 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {i} [ k ] = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} k = \\arg \\min \\left(S ^ {i}\\right) \\text {a n d} S ^ {i} [ k ] \\leq \\tau_ {n} \\\\ 0, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.291, + 0.784, + 0.321 + ], + "angle": 0, + "content": "where \\(\\hat{y}_i[k]\\) denotes the \\(k\\)-th dimension of the negative label of \\(P_{url}^i\\), \\(S^i\\) denotes the prediction confidence of \\(P_{url}^i\\), and \\(\\tau_n\\) is a pre-determined threshold." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.321, + 0.646, + 0.336 + ], + "angle": 0, + "content": "The selected negative label is used for negative learning:" + }, + { + "type": "equation", + "bbox": [ + 0.404, + 0.344, + 0.784, + 0.385 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {P E} ^ {u} = - \\sum_ {i = 1} ^ {N _ {u}} \\hat {y} _ {i} \\log \\left(1 - O _ {i}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.39, + 0.784, + 0.479 + ], + "angle": 0, + "content": "where \\( N_{u} \\) is the number of unreliable points with selected negative class and \\( O_{i} \\) is the prediction of \\( P_{url}^{i} \\). Then, the used negative class is discarded in the subsequent iterations. Simultaneously, the network is optimized via the negative learning loss. Subsequently, a new negative class is selected from the remaining classes for the next iteration. The above process is repeated until no negative class can be selected or the iteration number is reached." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.481, + 0.784, + 0.526 + ], + "angle": 0, + "content": "Moreover, it is worth noting that the proposed progressive exploitation strategy can be seamlessly integrated into methods that discard the unreliable points, and further boost their performances, as shown in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.545, + 0.347, + 0.558 + ], + "angle": 0, + "content": "3.4 Total Loss" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.567, + 0.784, + 0.597 + ], + "angle": 0, + "content": "The labeled points in the student branch are trained in a supervised manner, with their corresponding ground-truth labels \\( y^{l} \\) being the supervisory signals:" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.604, + 0.784, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C E} ^ {l} = \\operatorname {C E} \\left(O ^ {s t u, l}, y ^ {l}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.627, + 0.784, + 0.656 + ], + "angle": 0, + "content": "where \\(\\mathrm{CE}(\\cdot)\\) denotes the cross-entropy loss, \\(O^{stu,l}\\) denotes the predictions of the labeled points in the student branch." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.658, + 0.784, + 0.688 + ], + "angle": 0, + "content": "The predictions of the reliable points in the teacher branch are regarded as the pseudo labels of their corresponding points in the student branch:" + }, + { + "type": "equation", + "bbox": [ + 0.421, + 0.696, + 0.784, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C E} ^ {u} = \\operatorname {C E} \\left(O ^ {s t u, u}, y ^ {u}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.718, + 0.784, + 0.762 + ], + "angle": 0, + "content": "where \\( y^{u} \\) denotes the reliable predictions output by the teacher branch and \\( O^{stu,u} \\) denotes the predictions of corresponding unlabeled points in the student branch." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.765, + 0.784, + 0.796 + ], + "angle": 0, + "content": "Accordingly, the total loss \\(\\mathcal{L}_{total}\\) is the weighted sum of the two loss terms \\(\\mathcal{L}_{CE}^{l}\\), \\(\\mathcal{L}_{CE}^{u}\\) in the student branch and the aforementioned \\(\\mathcal{L}_{PE}^{u}\\), \\(\\mathcal{L}_{KL}^{u}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.802, + 0.784, + 0.818 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {C E} ^ {l} + \\lambda_ {u} \\mathcal {L} _ {C E} ^ {u} + \\lambda_ {P E} \\mathcal {L} _ {P E} ^ {u} + \\lambda_ {K L} \\mathcal {L} _ {K L} ^ {u}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.825, + 0.614, + 0.842 + ], + "angle": 0, + "content": "where \\(\\lambda_{u},\\lambda_{PE},\\lambda_{KL}\\) are the weights of \\(\\mathcal{L}_{CE}^{u},\\mathcal{L}_{PE}^{u},\\mathcal{L}_{KL}^{u}\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.129 + ], + "angle": 0, + "content": "J. Li and Q. Dong" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.367, + 0.163 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.179, + 0.525, + 0.194 + ], + "angle": 0, + "content": "4.1 Datasets and Evaluation Metric" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.206, + 0.787, + 0.236 + ], + "angle": 0, + "content": "The following outdoor-scene and indoor-scene datasets are used to evaluate the proposed LASS3D:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.238, + 0.786, + 0.327 + ], + "angle": 0, + "content": "- SemanticKITTI [1] is a large-scale 3D outdoor-scene LiDAR dataset consisting of 22 sequences, among which 10 sequences are used for training, 1 sequence is used for validation, and 11 sequences are used for testing. It contains 19 categories after merging and ignoring classes with few points. According to the official splitting, 850 scenes are used for training and validation, and 150 scenes are utilized for testing." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.328, + 0.786, + 0.388 + ], + "angle": 0, + "content": "- ScanNet V2 [4] is a widely-used 3D indoor-scene dataset consisting of 1613 scans with 20 categories, where the point clouds are reconstructed from multi-view RGB-D images. The training set contains 1201 scans, the validation split contains 312 scans, and the testing set contains 100 scans." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.238, + 0.786, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.39, + 0.786, + 0.418 + ], + "angle": 0, + "content": "As done in previous works [21, 26, 38], we use the mIoU (mean Intersection over Union) as the evaluation metric." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.443, + 0.457, + 0.459 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.47, + 0.787, + 0.577 + ], + "angle": 0, + "content": "We follow the basic experimental settings of LaserMix [21] and GPC [17] for evaluating the proposed LASS3D on the outdoor-scene and indoor-scene datasets respectively. The classifiers and fusion adapter are both multi-layer perceptrons. We follow the settings of IGNet [38], which also adopts the MeanTeacher [37] framework, to set the update hyperparameter \\(\\alpha\\) as 0.999. The iteration number in progressive exploitation is set as 5. The hyperparameters \\(\\tau_r, \\tau_v, \\tau_n, \\lambda_u, \\lambda_{PE}, \\lambda_{KL}\\) are set as 0.9, 0.05, 0.2, 0.1, 0.1, and 0.1 respectively." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.603, + 0.787, + 0.66 + ], + "angle": 0, + "content": "Table 1: Comparative results on the SemanticKITTI [1] and ScanNet [4] datasets with varying labeled ratios. \\(\\mathcal{P}\\) denotes the point cloud data, \\(\\mathcal{I}\\) denotes the image data, and \\(\\mathcal{T}\\) denotes the text data. All mIoU scores are given in percentage (\\%). The best results are in bold and the second best results are marked with underlines." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.672, + 0.784, + 0.837 + ], + "angle": 0, + "content": "
MethodModalitySemanticKITTI [1]ScanNet [4]
1%10%20%50%5%10%20%30%40%
MeanTeacher [37]P45.457.159.260.048.757.266.167.369.1
CBST [55]P48.858.359.459.748.557.865.467.369.2
CPS [3]P46.758.759.660.551.758.666.468.070.5
SSS-Net [5]P-----52.455.1--
LaserMix (Range View) [21]P43.458.859.461.4-----
LaserMix (Voxel) [21]P50.660.061.962.3-----
IGNet [38]P, I49.061.363.164.8-----
GPC [17]P54.162.062.562.854.860.566.768.971.3
LiM3D [26]P58.462.263.163.6-----
LASS3DP, I, T58.563.064.164.556.663.167.470.472.0
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.671, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LASS3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.461, + 0.162 + ], + "angle": 0, + "content": "4.3 Comparative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.187, + 0.788, + 0.324 + ], + "angle": 0, + "content": "We evaluate the proposed LASS3D on both outdoor-scene (SemanticKITTI [1]) and indoor-scene (ScanNet [4]) datasets in comparison to some classic semi-supervised methods [3, 37, 55] extended from the 2D domain and the state-of-the-art (SOTA) methods [5, 17, 21, 26, 38] that are specially designed for semi-supervised 3D semantic segmentation. Note that IGNet [38], LaserMix [21], and LiM3D [26] utilize some specific characters of the LiDAR point clouds, and thus cannot be applied to the ScanNet dataset. In addition, SSS-Net [5] leverages the color information of the point clouds, and thus cannot be applied to the SemantiKITTI dataset where the color information is unavailable." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.788, + 0.478 + ], + "angle": 0, + "content": "Following the settings in [17, 21, 38], we set the labeled ratio of the outdoor-scene dataset and indoor-scene dataset as \\(\\{1\\%, 10\\%, 20\\%, 40\\% \\}\\) and \\(\\{5\\%, 10\\%, 20\\%, 30\\%, 50\\% \\}\\) respectively for a fair comparison. The comparative results are reported in Tab. 1. As seen from this table, the proposed LASS3D outperforms all the comparative methods that use uni-modal data as input, which demonstrates the effectiveness of LASS3D. In addition, LASS3D outperforms IGNet which takes both point clouds and images as input in most cases, which indicates that the proposed language-assisted method can better boost the 3D segmentation performances and may provide some insights on the exploitation of multi-modal data in semi-supervised 3D semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.543 + ], + "angle": 0, + "content": "Moreover, the segmentation results of LASS3D and the second-best methods are visualized in Fig. 6. As seen from this figure, LASS3D outperforms its second-best counterparts, demonstrating the effectiveness of LASS3D from the qualitative perspective." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.58, + 0.77, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.837 + ], + "angle": 0, + "content": "Fig. 6: Qualitative results on the SemanticKITTI [1] (top) and ScanNet [4] (bottom) datasets by LASS3D and the second-best methods. All models are trained with \\(10\\%\\) labeled data. We use the red boxes to highlight the areas where LASS3D evidently outperforms the comparative methods." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.129 + ], + "angle": 0, + "content": "J. Li and Q. Dong" + }, + { + "type": "table_caption", + "bbox": [ + 0.266, + 0.145, + 0.51, + 0.158 + ], + "angle": 0, + "content": "Table 2: Ablation of the loss terms." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.172, + 0.56, + 0.259 + ], + "angle": 0, + "content": "
LCElLCEuLKLLPE1%10%20%50%
✓49.658.860.761.2
✓✓53.760.161.662.8
✓✓✓55.661.462.863.7
✓✓✓✓58.563.064.164.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.569, + 0.145, + 0.776, + 0.159 + ], + "angle": 0, + "content": "Table 3: Ablation of captions." + }, + { + "type": "table", + "bbox": [ + 0.564, + 0.172, + 0.789, + 0.258 + ], + "angle": 0, + "content": "
CiCe1%10%20%50%
50.059.861.562.1
✓53.660.561.963.0
✓54.161.062.763.8
✓✓58.563.064.164.5
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.289, + 0.389, + 0.305 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.316, + 0.785, + 0.346 + ], + "angle": 0, + "content": "We conduct ablation studies on SemanticKITTI [1] to investigate the effect of the key elements in the proposed LASS3D." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.347, + 0.785, + 0.483 + ], + "angle": 0, + "content": "Effect of the loss terms. We investigate the effect of the loss terms and report the corresponding results in Tab. 2. As seen from the first two rows of Tab. 2, incorporating the unlabeled data into training can improve the performances. Then, as seen in the third row of Tab. 2, the \\(\\mathcal{L}_{KL}^u\\) is introduced into training and the performances are improved, indicating that the semantic information distilled from the text-enhanced 3D features is helpful for segmentation. Finally, in the last row of Tab. 2, the progressive exploitation is adopted, and the performances are further boosted, demonstrating that the proposed progressive exploitation can make the unreliable points beneficial to the model training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.484, + 0.785, + 0.604 + ], + "angle": 0, + "content": "Effect of different captions. We investigate the effect of different captions and report the corresponding results in Tab. 3. As seen from this table, both the image-level captions \\(\\mathcal{C}_i\\) and entity-level captions \\(\\mathcal{C}_e\\) can improve the segmentation performances, indicating that the semantic information encoded in text embeddings can facilitate the point cloud segmentation. When combining \\(\\mathcal{C}_i\\) and \\(\\mathcal{C}_e\\), the best performance is achieved. Because the \\(\\mathcal{C}_i\\) provides more holistic semantic information and \\(\\mathcal{C}_e\\) provides more fine-grained semantic information, which are complementary to each other. Thus, we utilize \\(\\mathcal{C}_i\\) and \\(\\mathcal{C}_e\\) in LASS3D." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.785, + 0.71 + ], + "angle": 0, + "content": "Effect of the fusion strategy for the image-level embeddings \\(\\mathcal{E}_i\\). As mentioned in Sec. 3.2, we adopt the cross-attention operation to fuse \\(\\mathcal{E}_i\\) with the 3D features. To investigate the effect of the fusion strategy, we replace the cross-attention operation with add and concatenation operations. The corresponding results are reported in Tab. 4. As seen from this table, the model trained with the cross-attention operation achieves the best performance. Because the image-level captions usually contain holistic semantic information and" + }, + { + "type": "table_caption", + "bbox": [ + 0.235, + 0.738, + 0.485, + 0.752 + ], + "angle": 0, + "content": "Table 4: Ablation of fusion strategy." + }, + { + "type": "table_caption", + "bbox": [ + 0.503, + 0.738, + 0.78, + 0.752 + ], + "angle": 0, + "content": "Table 5: Ablation of separation strategy." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.764, + 0.797, + 0.837 + ], + "angle": 0, + "content": "
Strategy1%10%20%50%Strategy1%10%20%50%
Add54.061.262.363.5Confidence-based57.762.363.564.1
Concatenation56.262.063.463.9Variance-based57.062.162.863.5
Cross attention58.563.064.164.5Ours58.563.064.164.5
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LASS3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "may have different correlations with different regions of the paired points. Simply adding or concatenating \\(\\mathcal{E}_i\\) to all the paired 3D features leads to the lack of discrimination, while 3D semantic segmentation is a point-level prediction task that requires discriminative representations for each point. Cross attention can adaptively fuse \\(\\mathcal{E}_i\\) with 3D features and produce more discriminative features, which makes it a more suitable fusion strategy for \\(\\mathcal{E}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.794, + 0.435 + ], + "angle": 0, + "content": "Effectiveness of the semantic-aware adaptive fusion module (SAFM). As mentioned in Sec. 2, existing language-assisted point cloud learning methods are either projection-based or alignment-based. Project-based methods are generally designed for object-level 3D classification and are not applicable to scene-level 3D semantic segmentation. To verify the effectiveness of the proposed SAFM, we replace it with an alignment-based method (e.g., the semantic consistency regularization in CLIP2Scene [2]). The corresponding results are reported in Tab. 6. As seen from this table, the model trained with SAFM outperforms the model trained with semantic consistency regularization. Probably because a large distribution discrepancy exists between the 3D features and text embeddings, simply aligning the 3D features to their corresponding text embeddings may result in the loss of some inherent geometric information. SAFM remedies this defect by adaptive fusion, and thus achieves better performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.788, + 0.618 + ], + "angle": 0, + "content": "Effect of the separation strategy. As mentioned in Sec. 3.3, we adopt both confidences and variances to measure the reliability of the predictions and separate the unlabeled points into the reliable set and unreliable set. Here we evaluate the effectiveness of our separation strategy. The corresponding results are reported in Tab. 5. As seen from this table, the model trained with our strategy achieves the best performance. Probably because the confidence-based strategy alone is unable to filter out the highly confident wrong predictions, which may confuse the training. The variance-based strategy separates points according to the prediction consistency in different branches, which could be complementary to the confidence-based strategy in measuring reliability. Thus, we combine the above two strategies to measure reliability in a more comprehensive way, which can facilitate the model training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.709 + ], + "angle": 0, + "content": "Effect of the iteration number. The iteration number of progressive exploitation determines the utilization degree of the unreliable data. Here, we evaluate LASS3D with the iteration number set as \\(\\{0,1,2,3,5,7,9,11\\}\\), and the corresponding results are shown in Fig. 7. As seen from this figure, the performances are generally improved with the increase of iteration number. Because a larger iteration number means that the exploitation of the unreliable data is more thor-" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.737, + 0.787, + 0.767 + ], + "angle": 0, + "content": "Table 6: Ablation of semantic-aware adap- Table 7: Ablation of the proposed pro- tive fusion module (SAFM). progressive exploitation (PE)." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.778, + 0.795, + 0.837 + ], + "angle": 0, + "content": "
Strategy1%10%20%50%Model1%10%20%50%
Alignment-based53.060.161.862.2GPC [17]54.162.062.562.8
SAFM58.563.064.164.5GPC with PE54.762.863.163.5
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.129 + ], + "angle": 0, + "content": "J. Li and Q. Dong" + }, + { + "type": "image", + "bbox": [ + 0.381, + 0.147, + 0.613, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.269, + 0.784, + 0.297 + ], + "angle": 0, + "content": "Fig. 7: Ablation of the iteration number. Note that the iteration here denotes the negative learning iteration in progressive exploitation, rather than the training iteration." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.415 + ], + "angle": 0, + "content": "ough and thus brings better performance. However, when the iteration number reaches a certain level, the segmentation performance improves slightly. Probably because many unreliable points turn into reliable points during the progressive exploitation process, with most of their confident negative classes discarded. In addition, more iterations inevitably bring more training costs. Thus, we set the iteration number as 5 for a trade-off between the performance and training cost." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.416, + 0.785, + 0.52 + ], + "angle": 0, + "content": "Effectiveness of the progressive exploitation. Considering that the proposed progressive exploitation strategy only requires the prediction confidence and variances, thus it can be seamlessly integrated into some existing methods. We integrate the progressive exploitation strategy into GPC [17] and report the comparative results in Tab. 7. As seen from this table, the proposed progressive exploitation strategy can further boost the performance of GPC, demonstrating the effectiveness of the proposed strategy for exploiting unreliable points." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.541, + 0.348, + 0.555 + ], + "angle": 0, + "content": "4.5 Limitation" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.563, + 0.784, + 0.638 + ], + "angle": 0, + "content": "The proposed LASS3D has demonstrated its effectiveness in segmentation, however, it still has limitations. For example, LASS3D uses images as the bridge to connect the text data and point clouds, which is a complex operation and may bring accumulated errors in point-text matching. A more direct and accurate method to construct the point-text pairs is in demand." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.66, + 0.357, + 0.675 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.825 + ], + "angle": 0, + "content": "In this paper, we propose a language-assisted semi-supervised 3D semantic segmentation method named LASS3D. In LASS3D, a semantic-aware adaptive fusion module is explored to fuse the 3D features with the embeddings of multi-level captions generated by language-vision models, which can inject the semantic information into the 3D features and thus facilitate 3D segmentation. In addition, a progressive exploitation strategy is explored for unlabeled unreliable data, which can further boost performance and can be seamlessly embedded into other methods. Experimental results on outdoor and indoor datasets show that LASS3D outperforms the comparative methods in most cases." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LASS3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (Grant Nos. 61991423, 62376269, 62073199); the Strategic Priority Research Program of the Chinese Academy of Sciences (Grant No. XDA27040811); the Open Project Program of Key Laboratory of Industrial Internet and Big Data, China National Light Industry, Beijing Technology and Business University." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.26, + 0.326, + 0.276 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.331 + ], + "angle": 0, + "content": "1. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semanticickitti: A dataset for semantic scene understanding of lidar sequences. In: ICCV. pp. 9297-9307 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.785, + 0.372 + ], + "angle": 0, + "content": "2. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: Clip2scene: Towards label-efficient 3d scene understanding by clip. In: CVPR. pp. 7020-7030 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.373, + 0.785, + 0.401 + ], + "angle": 0, + "content": "3. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: CVPR. pp. 2613-2622 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.401, + 0.785, + 0.441 + ], + "angle": 0, + "content": "4. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR. pp. 5828-5839 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.442, + 0.785, + 0.47 + ], + "angle": 0, + "content": "5. Deng, S., Dong, Q., Liu, B., Hu, Z.: Superpoint-guided semi-supervised semantic segmentation of 3d point clouds. In: ICRA. pp. 9214-9220 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.47, + 0.785, + 0.497 + ], + "angle": 0, + "content": "6. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: CVPR. pp. 7010-7019 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.497, + 0.785, + 0.538 + ], + "angle": 0, + "content": "7. Fan, S., Dong, Q., Zhu, F., Lv, Y., Ye, P., Wang, F.: Scf-net: Learning spatial contextual features for large-scale point cloud segmentation. In: CVPR. pp. 14499-14508 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.539, + 0.785, + 0.565 + ], + "angle": 0, + "content": "8. Graham, B., Engelcke, M., van der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. CVPR pp. 9224-9232 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.566, + 0.785, + 0.607 + ], + "angle": 0, + "content": "9. Guo, Z., Zhang, R., Qiu, L., Ma, X., Miao, X., He, X., Cui, B.: Calip: Zero-shot enhancement of clip with parameter-free attention. In: AAAI. vol. 37, pp. 746-754 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.608, + 0.785, + 0.635 + ], + "angle": 0, + "content": "0. He, W., Jamonnak, S., Gou, L., Ren, L.: Clip-s4: Language-guided self-supervised semantic segmentation. In: CVPR. pp. 11207-11216 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.635, + 0.785, + 0.662 + ], + "angle": 0, + "content": "1. Hegde, D., Valanarasu, J.M.J., Patel, V.: Clip goes 3d: Leveraging prompt tuning for language grounded 3d recognition. In: ICCVW. pp. 2028-2038 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.662, + 0.785, + 0.689 + ], + "angle": 0, + "content": "2. Hess, G., Tondersski, A., Petersson, C., Åström, K., Svensson, L.: Lidarclip or: How i learned to talk to point clouds. In: WACV. pp. 7438-7447 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.69, + 0.785, + 0.731 + ], + "angle": 0, + "content": "3. Hu, Q., Yang, B., Fang, G., Guo, Y., Leonardis, A., Trigoni, N., Markham, A.: Sqn: Weakly-supervised semantic segmentation of large-scale 3d point clouds. In: ECCV. pp. 600-619. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.785, + 0.758 + ], + "angle": 0, + "content": "4. Huang, R., Pan, X., Zheng, H., Jiang, H., Xie, Z., Wu, C., Song, S., Huang, G.: Joint representation learning for text and 3d point cloud. PR 147, 110086 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "5. Huang, T., Dong, B., Yang, Y., Huang, X., Lau, R.W., Ouyang, W., Zuo, W.: Clip2point: Transfer clip to point cloud classification with image-depth pretraining. In: ICCV. pp. 22157-22167 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML. pp. 4904-4916 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.395, + 0.129 + ], + "angle": 0, + "content": "J. Li and Q. Dong" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "17. Jiang, L., Shi, S., Tian, Z., Lai, X., Liu, S., Fu, C.W., Jia, J.: Guided point contrastive learning for semi-supervised point cloud semantic segmentation. In: ICCV. pp. 6423-6432 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.217 + ], + "angle": 0, + "content": "18. Jiao, S., Wei, Y., Wang, Y., Zhao, Y., Shi, H.: Learning mask-aware CLIP representations for zero-shot segmentation. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.785, + 0.243 + ], + "angle": 0, + "content": "19. Kim, Y., Yim, J., Yun, J., Kim, J.: Nlnl: Negative learning for noisy labels. In: ICCV. pp. 101-110 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.284 + ], + "angle": 0, + "content": "20. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV. pp. 4015-4026 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.311 + ], + "angle": 0, + "content": "21. Kong, L., Ren, J., Pan, L., Liu, Z.: Lasermix for semi-supervised lidar semantic segmentation. In: CVPR. pp. 21705-21715 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.338 + ], + "angle": 0, + "content": "22. Lai, X., Chen, Y., Lu, F., Liu, J., Jia, J.: Spherical transformer for lidar-based 3d recognition. In: CVPR. pp. 17545-17555 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.366 + ], + "angle": 0, + "content": "23. Lai, X., Liu, J., Jiang, L., Wang, L., Zhao, H., Liu, S., Qi, X., Jia, J.: Stratified transformer for 3d point cloud segmentation. In: CVPR. pp. 8500-8509 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.785, + 0.393 + ], + "angle": 0, + "content": "24. Li, J., Dong, Q.: Open-set semantic segmentation for point clouds via adversarial prototype framework. In: CVPR. pp. 9425-9434 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.419 + ], + "angle": 0, + "content": "25. Li, J., Dong, Q.: Density-guided semi-supervised 3d semantic segmentation with dual-space hardness sampling. In: CVPR. pp. 3260-3269 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.419, + 0.785, + 0.446 + ], + "angle": 0, + "content": "26. Li, L., Shum, H.P., Breckon, T.P.: Less is more: Reducing task and model complexity for 3d point cloud semantic segmentation. In: CVPR. pp. 9361-9371 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.785, + 0.487 + ], + "angle": 0, + "content": "27. Lin, Z., Geng, S., Zhang, R., Gao, P., de Melo, G., Wang, X., Dai, J., Qiao, Y., Li, H.: Frozen clip models are efficient video learners. In: ECCV. pp. 388-404. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.785, + 0.528 + ], + "angle": 0, + "content": "28. Liu, L., Zhuang, Z., Huang, S., Xiao, X., Xiang, T., Chen, C., Wang, J., Tan, M.: Cpcm: Contextual point cloud modeling for weakly-supervised point cloud semantic segmentation. In: ICCV. pp. 18413-18422 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.554 + ], + "angle": 0, + "content": "29. Liu, Z., Qi, X., Fu, C.W.: One thing one click: A self-training approach for weakly supervised 3d semantic segmentation. In: CVPR. pp. 1726-1736 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.554, + 0.785, + 0.595 + ], + "angle": 0, + "content": "30. Lu, Y., Xu, C., Wei, X., Xie, X., Tomizuka, M., Keutzer, K., Zhang, S.: Open-vocabulary point-cloud object detection without 3d annotation. In: CVPR. pp. 1190-1199 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.785, + 0.637 + ], + "angle": 0, + "content": "31. Ma, W., Li, S., Zhang, J., Liu, C.H., Kang, J., Wang, Y., Huang, G.: Borrowing knowledge from pre-trained language model: A new data-efficient visual learning paradigm. In: ICCV. pp. 18786-18797 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.676 + ], + "angle": 0, + "content": "32. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.676, + 0.785, + 0.718 + ], + "angle": 0, + "content": "33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "34. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Denseclip: Language-guided dense prediction with context-aware prompting. In: CVPR. pp. 18082-18091 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "35. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., Zeng, Z., Zhang, H., Li, F., Yang, J., Li, H., Jiang, Q., Zhang, L.: Grounded sam: Assembling open-world models for diverse visual tasks (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "36. Tang, L., Chen, Z., Zhao, S., Wang, C., Tao, D.: All points matter: entropy-regularized distribution alignment for weakly-supervised 3d segmentation. NeurIPS 36 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LASS3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "37. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NIPS 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.217 + ], + "angle": 0, + "content": "38. Unal, O., Dai, D., Hoyer, L., Can, Y.B., Van Gool, L.: 2d feature distillation for weakly-and semi-supervised 3d semantic segmentation. In: WACV. pp. 7336-7345 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.787, + 0.258 + ], + "angle": 0, + "content": "39. Wang, Y., Huang, S., Gao, Y., Wang, Z., Wang, R., Sheng, K., Zhang, B., Liu, S.: Transferring clip's knowledge into zero-shot point cloud semantic segmentation. In: ACM MM. pp. 3745-3754 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.3 + ], + "angle": 0, + "content": "40. Wang, Y., Wang, H., Shen, Y., Fei, J., Li, W., Jin, G., Wu, L., Zhao, R., Le, X.: Semi-supervised semantic segmentation using unreliable pseudo-labels. In: CVPR. pp. 4248-4257 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.328 + ], + "angle": 0, + "content": "41. Wu, X., Jiang, L., Wang, P.S., Liu, Z., Liu, X., Qiao, Y., Ouyang, W., He, T., Zhao, H.: Point transformer v3: Simpler, faster, stronger. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.357 + ], + "angle": 0, + "content": "42. Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point transformer v2: Grouped vector attention and partition-based pooling. In: NeurIPS. vol. 35, pp. 33330-33342 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.397 + ], + "angle": 0, + "content": "43. Xiang, P., Wen, X., Liu, Y.S., Zhang, H., Fang, Y., Han, Z.: Retro-fpn: Retrospective feature pyramid network for point cloud semantic segmentation. In: ICCV. pp. 17826-17838 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.439 + ], + "angle": 0, + "content": "44. Xue, L., Gao, M., Xing, C., Martin-Martín, R., Wu, J., Xiong, C., Xu, R., Niebles, J.C., Savarese, S.: Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In: CVPR. pp. 1179-1189 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "45. Yao, L., Han, J., Wen, Y., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, C., Xu, H.: Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. NeurIPS 35, 9125-9138 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.508 + ], + "angle": 0, + "content": "46. Yu, W., Liu, Y., Hua, W., Jiang, D., Ren, B., Bai, X.: Turning a clip model into a scene text detector. In: CVPR. pp. 6978-6988 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "47. Zeng, Y., Jiang, C., Mao, J., Han, J., Ye, C., Huang, Q., Yeung, D.Y., Yang, Z., Liang, X., Xu, H.: Clip2: Contrastive language-image-point pretraining from real-world point cloud data. In: CVPR. pp. 15244-15253 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.577 + ], + "angle": 0, + "content": "48. Zhang, J., Fan, G., Wang, G., Su, Z., Ma, K., Yi, L.: Language-assisted 3d feature learning for semantic scene understanding. In: AAAI. pp. 3445-3453 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.605 + ], + "angle": 0, + "content": "49. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: CVPR. pp. 8552-8562 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "50. Zhang, Y., Huo, X., Chen, T., Wu, S., Wong, H.S.: Exploring intra-class variation factors with learnable cluster prompts for semi-supervised image synthesis. In: CVPR. pp. 7392-7401 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.646, + 0.787, + 0.674 + ], + "angle": 0, + "content": "51. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV. pp. 16259-16268 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.787, + 0.702 + ], + "angle": 0, + "content": "52. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: ECCV. pp. 696-712. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.787, + 0.743 + ], + "angle": 0, + "content": "53. Zhu, X., Zhang, R., He, B., Guo, Z., Zeng, Z., Qin, Z., Zhang, S., Gao, P.: Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In: ICCV. pp. 2639-2650 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.787, + 0.785 + ], + "angle": 0, + "content": "54. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In: CVPR. pp. 9939-9948 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.813 + ], + "angle": 0, + "content": "55. Zou, Y., Yu, Z., Kumar, B., Wang, J.: Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In: ECCV. pp. 289-305 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.813 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/LASS3D_ Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation/c8ab9b1f-da9d-4178-ac58-4d667a9330b3_origin.pdf b/2024/LASS3D_ Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation/c8ab9b1f-da9d-4178-ac58-4d667a9330b3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d0d0f3883cb817a56a6bf4d790ba82821571885a --- /dev/null +++ b/2024/LASS3D_ Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation/c8ab9b1f-da9d-4178-ac58-4d667a9330b3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18fe4aa92e985b30f8d309968f8291dc1bb0493b297a49b97512f76b418a213c +size 1464028 diff --git a/2024/LASS3D_ Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation/full.md b/2024/LASS3D_ Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8a0062ad850e04e45c893fcfe25037f81fe27f8b --- /dev/null +++ b/2024/LASS3D_ Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation/full.md @@ -0,0 +1,334 @@ +# LASS3D: Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation + +Jianan $\mathrm{Li}^{1,2}$ and Qiulei Dong + +$^{1}$ School of Artificial Intelligence, University of Chinese Academy of Sciences + +$^{2}$ State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences + +3 Center for Excellence in Brain Science and Intelligence Technology, Chinese Academy of Sciences + +lijianan211@mails.ucas.ac.cn, qldong@nlpr.ac.cn + +Abstract. Precisely annotating large-scale 3D datasets for point cloud segmentation is laborious. To alleviate the annotation burden, several semi-supervised 3D segmentation methods have been proposed in literature. However, two issues remain to be tackled: 1) The utilization of large language-vision models (LVM) in semi-supervised 3D semantic segmentation remains under-explored. 2) The unlabeled points with low-confidence predictions are directly discarded by existing methods. Taking these two issues into consideration, we propose a language-assisted semi-supervised 3D semantic segmentation method named LASS3D, which is built upon the commonly used MeanTeacher framework. In LASS3D, we use two off-the-shelf LVM to generate multi-level captions and leverage the images as the bridge to connect the text data and point clouds. Then, a semantic-aware adaptive fusion module is explored in the student branch, where the semantic information encoded in the embeddings of multi-level captions is injected into 3D features by adaptive fusion and then the semantic information in the text-enhanced 3D features is transferred to the teacher branch by knowledge distillation. In addition, a progressive exploitation strategy is explored for the unreliable points in the teacher branch, which can effectively exploit the information encapsulated in unreliable points via negative learning. Experimental results on both outdoor and indoor datasets demonstrate that LASS3D outperforms the comparative methods in most cases. + +Keywords: 3D semantic segmentation $\cdot$ Semi-supervised segmentation + +# 1 Introduction + +3D Semantic Segmentation [7,22-24,42,43,51] is an important task for perceiving real-world environments, and intensive efforts have been devoted to this task in recent years. However, most of the existing works [8,22,41,42,51,54] in literature + +![](images/18270764dc2d370675187c775afcc2039a2ead85285408e22c0ab97be77c74d2.jpg) +Fig. 1: Illustration of the proportions of reliable and unreliable points predicted by GPC [17] on ScanNet [4]. The horizontal axis denotes the labeled ratio and training iteration. The vertical axis denotes the proportion. + +are fully-supervised methods, which require time-consuming and labor-intensive data annotation for training. + +This issue encourages researchers to investigate semi-supervised 3D semantic segmentation, where only a small amount of densely labeled data and a large amount of unlabeled data are provided for training. Existing works [5,17,21,26, 38] on semi-supervised 3D semantic segmentation have shown their effectiveness, but there is ample room for improvement in the following two aspects: + +1) Utilization of language-vision models (LVM): Recently, LVM [16,33] trained on web-crawled image-text pairs has been applied in some downstream tasks [2, 31, 39, 50] to boost performance, owing to its strong representation ability. Text embeddings extracted by the LVM generally contain higher-level abstraction of visual concepts with rich semantic information, which is complementary to the 3D features that mainly contain geometric information and thus can facilitate mining the semantic information of unlabeled data. To our best knowledge, LVM has not been applied in semi-supervised 3D semantic segmentation yet. + +2) Exploitation of unreliable points (unlabeled points with low-confidence predictions): The proportions of reliable and unreliable points predicted by the state-of-the-art semi-supervised 3D semantic segmentation method GPC [17] on ScanNet [4] dataset are illustrated in Fig. 1. As seen from this figure, a large number of unlabeled points are deemed unreliable, especially at the early training stage. Existing works [17,21] generally adopt a confidence-based filtering strategy to remove the unreliable points for guaranteeing the effectiveness and stability of the training process. However, simply discarding them will inevitably lose some useful information. Thus, a proper learning strategy is expected to be specially designed to exploit these unreliable points effectively. + +Taking the above two aspects into consideration, we propose a method named LASS3D for semi-supervised 3D semantic segmentation, where the LVM and unreliable points are effectively exploited. The proposed LASS3D is built upon the MeanTeacher [37] framework, where the weights of the teacher branch are updated by the exponential moving average (EMA) of the student's weights. + +In the student branch, two LVM (i.e., KOSMOS-2 [32] and GroundedSAM [35]) are used to generate image-level and entity-level captions, and the images + +are leveraged as the bridge to connect the text data and point clouds. Then, a semantic-aware adaptive fusion module (SAFM) is designed to assist point cloud learning with the semantic knowledge encoded in text data. Specifically, multi-level point-text pairs are constructed in SAFM to adaptively fuse the 3D features and the text embeddings of the image-level and entity-level captions, which could facilitate capturing both holistic and fine-grained semantic information. + +In the teacher branch, a fusion adapter is utilized to narrow the distribution gap between the 3D features extracted by the backbone network and the text-enhanced 3D features in the student branch. Combined with the EMA strategy, the semantic information encoded in the text-enhanced 3D features is distilled from the student branch to the teacher branch. Note that we inject semantic information from text embeddings into the 3D features by fusion rather than directly mapping the semantic labels from pixels to points. Because the semantic labels generated by the open-world segmentation models (e.g., SAM [20]) are generally in a one-hot form, which provide hard supervision. However, some generated synonymous labels may not exist in our desired label set (e.g., the generated label is 'couch' but the ground-truth label is 'sofa'), and incorrect hard supervision may adversely affect the point cloud segmentation. Adaptively fusing the text embeddings and 3D features can mitigate this negative effect to some extent. Because the strong LVM can encode synonymous words into similar embeddings and thus can inject similar semantic information into 3D features. To effectively exploit the unlabeled points with unreliable predictions, a progressive exploitation strategy is designed based on the fact that determining the least likely category for an unlabeled sample is generally easier and more accurate than determining its most likely category. In addition, the proposed progressive exploitation strategy can be seamlessly integrated into some existing methods and further boost their performances, which is demonstrated in Sec. 4.4. + +In summary, the contributions of this paper are as follows: + +- We propose a semantic-aware adaptive fusion module (SAFM) to consolidate point cloud segmentation with the semantic information encoded in text embeddings. The multi-level fusion shows the feasibility and effectiveness of language-assisted 3D semantic segmentation in semi-supervised settings. +- We propose a progressive exploitation strategy for the unlabeled unreliable points. This strategy is beneficial to fully exploit the unlabeled data and can be seamlessly embedded into some existing methods. +- We propose the LASS3D for semi-supervised 3D semantic segmentation by integrating the above SAFM and progressive exploitation strategy. Its effectiveness is demonstrated by the results on both outdoor-scene and indoor-scene public datasets in Sec. 4. + +# 2 Related Work + +# 2.1 Semi-supervised 3D Semantic Segmentation + +To alleviate the annotation burden, label-efficient 3D semantic segmentation (including weakly-supervised segmentation [13,28,29,36] and semi-supervised seg + +mentation [5,17,21,25,26,38]) has drawn increasing interest among researchers. In this work, we delve into semi-supervised 3D semantic segmentation, which aims at utilizing a small number of densely labeled data and a large number of unlabeled data for model training. The core of semi-supervised 3D segmentation lies in exploiting the unlabeled data to the fullest. For example, Deng et al. [5] proposed to optimize the pseudo labels of the unlabeled data with the superpoints generated by geometry-based and color-based region-growing algorithms. Jiang et al. [17] proposed a label-guided contrastive loss for the unlabeled data, which could enhance the feature representation ability of the model. Kong et al. [21] proposed to leverage the spatial prior of LiDAR point clouds and use a mixing operation to provide supervisory information for the unlabeled data. Li et al. [26] proposed to utilize the reflectivity-prior descriptors to generate high-quality pseudo labels for the unlabeled data. Unal et al. [38] proposed to distill high-level feature information from a synthetically trained 2D network for data-efficient LiDAR semantic segmentation. + +However, these methods neglect the usage of text data that contains rich semantic information. To our best knowledge, LASS3D is the first work to investigate the language-assisted semi-supervised 3D semantic segmentation, where the powerful representation ability of language-vision models is exploited. + +# 2.2 Language-assisted 3D Learning + +Large language-vision models (LVM) [16, 33] have drawn growing attention in computer vision recently, attributed to their powerful representation ability. For example, CLIP [33], which is trained on massive paired text-image data via contrastive learning, has prevailed in various downstream visual tasks, such as semantic segmentation [10, 18, 34, 48, 52], object detection [30, 45, 46], and video recognition [27]. Recent advanced methods that focused on language-assisted point cloud learning could be roughly divided into two types: the projection-based methods and the alignment-based methods. + +The projection-based methods [9,15,49,53] project the point clouds into the 2D plane so that the image encoder in the LVM can be seamlessly utilized and the text-image alignment of the LVM can be leveraged for downstream tasks. + +![](images/6e0b8b61bd268728aeaf448be48b26bfb20157044f2632c27ce243bc71680f1c.jpg) +(a) Pipeline of PointCLIP. + +![](images/06d6033f83a48b21784513dc72197b314b638be96ac67e9ef2e8839f71dfb0b0.jpg) +(b) Alignment process of CLIP2Scene. +Fig. 2: Illustration of the pipeline of a projection-based method (e.g., PointCLIP [49]) and the alignment process of an alignment-based method (e.g., CLIP2Scene [2]). + +Fig. 2a illustrates the general pipeline of PointCLIP [49], it classifies the point clouds according to the similarities between the text embeddings and features of the projected images. Nevertheless, the projection-based methods are generally designed for object-level 3D classification. In addition, they usually suffer from the loss of geometric information during the projection process, thus resulting in suboptimal performance. + +The alignment-based methods [2,6,11,12,14,39,44,47] enforce the consistency among the multi-modal features. Fig. 2b illustrates the alignment process of CLIP2Scene [2], it uses contrastive learning to align the 3D features to their corresponding text embeddings. However, directly aligning multi-modal features may result in suboptimal results, due to the distribution discrepancy between the 3D features and text embeddings. + +Unlike existing methods that are either projection-based or alignment-based, we propose a fusion-based method to adaptively fuse the embeddings of multi-level captions and 3D features, which can alleviate the loss of geometric information and mitigate the negative effect brought by the distribution gap. + +# 3 Methodology + +# 3.1 Architecture + +The architecture of the proposed LASS3D is illustrated in Fig. 3. It is built upon the MeanTeacher [37] framework where the teacher branch provides the supervisory signals for the student branch and is updated by the exponential mean average (EMA) of the student's weights: + +$$ +\theta_ {t e a c h e r} ^ {t + 1} = \alpha \theta_ {t e a c h e r} ^ {t} + (1 - \alpha) \theta_ {s t u d e n t} ^ {t}, \tag {1} +$$ + +![](images/c51340eeb77e279aca57a369a8662d9b342a6055cc1a17692bdec13d6d6e408f.jpg) +Fig. 3: Architecture of the proposed LASS3D. $\mathcal{C}_i$ and $\mathcal{C}_e$ stand for the image-level captions and entity-level captions. $\mathcal{E}_i$ and $\mathcal{E}_e$ are the embeddings of $\mathcal{C}_i$ and $\mathcal{C}_e$ respectively. $/ / /$ denotes the detach operation which is used to stop the gradient. + +where $\theta_{teacher}^{t}$ and $\theta_{student}^{t}$ denote the weights of the teacher branch and student branch at time step $t$ , and $\alpha$ is the update hyperparameter. + +The student branch takes the point clouds and their corresponding images as input. The images are fed into the LVM, (i.e., KOSMOS-2 [35] and GroundedSAM [32]) to produce image-level $\mathcal{C}_i$ and entity-level captions $\mathcal{C}_e$ . Then, the multi-level captions are tokenized by the text encoder of CLIP [33] to generate their corresponding text embeddings. The text embeddings and 3D features extracted by the 3D backbone network are fused in the semantic-aware adaptive fusion module to obtain the text-enhanced 3D features, which are then used for segmentation in the student branch. + +The teacher branch takes the same point clouds as input. The 3D features extracted by the 3D backbone network are aligned with the text-enhanced 3D features through a fusion adapter. Then, the classifier takes the aligned 3D features as input and outputs their corresponding prediction scores. A dual-score separation strategy is adopted to select reliable points and unreliable points from the unlabeled points. The predictions of the reliable points are regarded as the pseudo labels of their corresponding points in the student branch, where the cross-entropy loss $\mathcal{L}_{CE}^{u}$ is calculated. The predictions of the unreliable points in the student branch are used for the progressive exploitation, where the negative learning loss $\mathcal{L}_{PE}^{u}$ is calculated. Besides, the predictions of the labeled points in the student branch and their ground-truth labels are used to calculate the cross-entropy loss $\mathcal{L}_{CE}^{l}$ , and the features output by the fusion adapter are aligned to the text-enhanced 3D features by a Kullback-Leibler divergence loss $\mathcal{L}_{KL}^{u}$ . + +Note that the text data is only required in the training stage. In the inference stage, only the teacher branch is used and it takes the point clouds as input. + +# 3.2 Semantic-aware Adaptive Fusion Module + +The architecture of the semantic-aware adaptive fusion module (SAFM) is illustrated in Fig. 4, it takes the embeddings of multi-level captions (i.e., image-level embeddings $\mathcal{E}_i$ and entity-level embeddings $\mathcal{E}_e$ ) and 3D features $f_{3D}$ as input and outputs the text-enhanced 3D features $f_{3D}^t$ , which are then fed into the classifier for segmentation and used as the supervision for knowledge distillation. + +The proposed SAFM is used to inject the semantic information encoded in multi-level captions into the 3D features. Specifically, the image-level captions $\mathcal{C}_i$ and entity-level captions $\mathcal{C}_e$ are generated by KOSMOS-2 [32] and GroundedSAM [35] respectively, and their corresponding embeddings are extracted by the text encoder of CLIP [33]. To guarantee the accuracy of multi-modal fusion, we construct the point-text pairs by using the images as the bridge. Then, given the point-text pairs, their corresponding features are adaptively fused to produce the text-enhanced 3D features. In this subsection, we will introduce the above two key operations of SAFM in detail. + +Point-text pair construction. We use the images as the bridge to connect the text data and point cloud data. Specifically, the multi-level pixel-text pairs can be obtained by the language-vision models, i.e., the image-level captions + +![](images/74abd7d2ade34c9654f4a0b42b8b96c223192a277cf186cd2dd745c81431ba7e.jpg) +Fig. 4: Architecture of the proposed semantic-aware adaptive fusion module. $f_{3D}$ , $\hat{f}_{3D}$ , and $f_{3D}^{t}$ denote the extracted 3D features, aligned 3D features, and text-enhanced 3D features. $\mathcal{E}_i$ and $\mathcal{E}_e$ are the embeddings of image-level and entity-level captions. $\oplus$ denotes the add operation. + +correspond to all pixels of the given images, and the entity-level captions correspond to the pixels of the segmented regions. The pixel-point pairs can be easily obtained with the extrinsic and intrinsic parameters of the cameras: + +$$ +[ u, v, 1 ] ^ {\mathrm {T}} = \mathbf {K M} [ x, y, z, 1 ] ^ {\mathrm {T}}, \tag {2} +$$ + +where $[u,v,1]^{\mathrm{T}}$ represents the homogeneous coordinates of a pixel, $[x,y,z,1]^{\mathrm{T}}$ represents its corresponding 3D point position in world coordinates, $\mathbf{K}$ and $\mathbf{M}$ are the intrinsic and extrinsic parameters of the cameras. With the multi-level pixel-text and pixel-point pairs, the multi-level point-text pairs can be obtained. + +Adaptive fusion. To effectively utilize the multi-level point-text pairs, an adaptive fusion strategy is proposed. As shown in Fig. 4, the dimension of the 3D features $f_{3D}$ is firstly expanded by an MLP layer to align the embedding space of the text encoder, given that the text encoder in LASS3D is fixed. Since the image-level captions may have different correlations with different regions of the paired points, cross attention is performed between the aligned 3D features $\hat{f}_{3D}$ and image-level embeddings $\mathcal{E}_i$ , with $\hat{f}_{3D}$ serving as the query vectors and $\mathcal{E}_i$ serving as the key vectors and value vectors. Then, considering that the points related to the same entity-level captions share the same semantic information, we simply add the entity-level embeddings $\mathcal{E}_e$ with their corresponding 3D features. Finally, an MLP layer is used to map the fused features to their original dimension and output the text-enhanced 3D features $f_{3D}^t$ . + +As shown in Fig. 3, the text-enhanced 3D features $f_{3D}^{t}$ are not only fed into the classifier of the student branch for segmentation but also serve as the supervisory signals for knowledge distillation in the teacher branch. The features $f_{ada}$ output by the fusion adapter in the teacher branch are aligned to their corresponding $f_{3D}^{t}$ through $\mathcal{L}_{KL}^{u}$ : + +$$ +\mathcal {L} _ {K L} ^ {u} = \mathrm {K L} \left(f _ {3 D} ^ {t} \mid \mid f _ {a d a}\right), \tag {3} +$$ + +where $\mathrm{KL}(\cdot)$ denotes the Kullback-Leibler divergence loss. + +![](images/37c560e7b39e5d0cbab11a466d1647bce478e3680bce9ba389ee0e3bd2fb1171.jpg) +Fig. 5: Illustration of the progressive exploitation for the unreliable point $P_{url}^{i}$ at the first two iterations. $S^{i}$ denotes the prediction confidence of $P_{url}^{i}$ and $\hat{y}_i$ denotes its negative label. The selected negative class (blue font) is used for negative learning and the used negative class (marked with red dashed box) is discarded in the subsequent iterations. Note that some unreliable points may turn into reliable points during progressive exploitation. + +# 3.3 Progressive Exploitation Strategy + +The progressive exploitation strategy is proposed to mine the useful information encapsulated in the unlabeled unreliable points. It contains two key steps: dual-score separation and progressive negative learning. In this subsection, we will introduce the above two steps in detail. + +Dual-score separation. Unlike GPC [17] and LaserMix [21] which only use the confidence of predictions to measure the reliability, we propose a dual-score separation strategy that additionally introduces the variances among the predictions of different branches into reliability measurement. Specifically, the unlabeled points $P_{u}$ in the teacher branch are separated into the reliable set $P_{rl}$ and unreliable set $P_{url}$ based on both confidence and variances: + +$$ +P _ {r l} = B \cdot P _ {u}, P _ {u r l} = (1 - B) \cdot P _ {u}, +$$ + +$$ +B = \mathbb {1} \left[ \sum_ {c = 1} ^ {C} \mathbb {1} \left[ S _ {c} \geq \tau_ {r} \right] \cdot \mathbb {1} \left[ \sigma_ {c} \leq \tau_ {v} \right] > 0 \right], \tag {4} +$$ + +where $B$ is a binary mask to select $P_{rl}$ and $P_{url}$ , $\mathbb{1}$ is the indicator function, $C$ denotes the number of classes, $S_{c}$ denotes the prediction confidences of class $c$ in the teacher branch, $\sigma_{c}$ denotes the prediction variances of class $c$ within the teacher and student branches, $\tau_{r}$ and $\tau_{v}$ are two pre-determined thresholds. + +The predictions of reliable points are regarded as the pseudo labels of their corresponding points in the student branch, which are used for supervised training, as shown in Eq. (8). The unreliable points are fully exploited via progressive negative learning, which is elaborated below. + +Progressive negative learning. Enlightened by the core idea of conventional negative learning [19, 40], we choose to mine information from the classes that the unreliable points do not belong to, rather than struggling to correctly tell which class they belong to. However, conventional methods only choose one negative class (the class with the lowest prediction score) for negative learning, which provides limited information and may cause unstable training due to the + +dynamic negative labels. Instead, we conduct negative learning in a progressive manner, with more negative classes selected. + +The progressive exploitation at the first two iterations is illustrated in Fig. 5. As shown in this figure, we maintain a list for each unreliable point to record its selected negative classes. In each iteration, one negative class is selected for each unreliable point: + +$$ +\hat {y} _ {i} [ k ] = \left\{ \begin{array}{l l} 1, & \text {i f} k = \arg \min \left(S ^ {i}\right) \text {a n d} S ^ {i} [ k ] \leq \tau_ {n} \\ 0, & \text {o t h e r w i s e} \end{array} , \right. \tag {5} +$$ + +where $\hat{y}_i[k]$ denotes the $k$ -th dimension of the negative label of $P_{url}^i$ , $S^i$ denotes the prediction confidence of $P_{url}^i$ , and $\tau_n$ is a pre-determined threshold. + +The selected negative label is used for negative learning: + +$$ +\mathcal {L} _ {P E} ^ {u} = - \sum_ {i = 1} ^ {N _ {u}} \hat {y} _ {i} \log \left(1 - O _ {i}\right), \tag {6} +$$ + +where $N_{u}$ is the number of unreliable points with selected negative class and $O_{i}$ is the prediction of $P_{url}^{i}$ . Then, the used negative class is discarded in the subsequent iterations. Simultaneously, the network is optimized via the negative learning loss. Subsequently, a new negative class is selected from the remaining classes for the next iteration. The above process is repeated until no negative class can be selected or the iteration number is reached. + +Moreover, it is worth noting that the proposed progressive exploitation strategy can be seamlessly integrated into methods that discard the unreliable points, and further boost their performances, as shown in Sec. 4.4. + +# 3.4 Total Loss + +The labeled points in the student branch are trained in a supervised manner, with their corresponding ground-truth labels $y^{l}$ being the supervisory signals: + +$$ +\mathcal {L} _ {C E} ^ {l} = \operatorname {C E} \left(O ^ {s t u, l}, y ^ {l}\right), \tag {7} +$$ + +where $\mathrm{CE}(\cdot)$ denotes the cross-entropy loss, $O^{stu,l}$ denotes the predictions of the labeled points in the student branch. + +The predictions of the reliable points in the teacher branch are regarded as the pseudo labels of their corresponding points in the student branch: + +$$ +\mathcal {L} _ {C E} ^ {u} = \operatorname {C E} \left(O ^ {s t u, u}, y ^ {u}\right), \tag {8} +$$ + +where $y^{u}$ denotes the reliable predictions output by the teacher branch and $O^{stu,u}$ denotes the predictions of corresponding unlabeled points in the student branch. + +Accordingly, the total loss $\mathcal{L}_{total}$ is the weighted sum of the two loss terms $\mathcal{L}_{CE}^{l}$ , $\mathcal{L}_{CE}^{u}$ in the student branch and the aforementioned $\mathcal{L}_{PE}^{u}$ , $\mathcal{L}_{KL}^{u}$ : + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {C E} ^ {l} + \lambda_ {u} \mathcal {L} _ {C E} ^ {u} + \lambda_ {P E} \mathcal {L} _ {P E} ^ {u} + \lambda_ {K L} \mathcal {L} _ {K L} ^ {u}, \tag {9} +$$ + +where $\lambda_{u},\lambda_{PE},\lambda_{KL}$ are the weights of $\mathcal{L}_{CE}^{u},\mathcal{L}_{PE}^{u},\mathcal{L}_{KL}^{u}$ . + +# 4 Experiment + +# 4.1 Datasets and Evaluation Metric + +The following outdoor-scene and indoor-scene datasets are used to evaluate the proposed LASS3D: + +- SemanticKITTI [1] is a large-scale 3D outdoor-scene LiDAR dataset consisting of 22 sequences, among which 10 sequences are used for training, 1 sequence is used for validation, and 11 sequences are used for testing. It contains 19 categories after merging and ignoring classes with few points. According to the official splitting, 850 scenes are used for training and validation, and 150 scenes are utilized for testing. +- ScanNet V2 [4] is a widely-used 3D indoor-scene dataset consisting of 1613 scans with 20 categories, where the point clouds are reconstructed from multi-view RGB-D images. The training set contains 1201 scans, the validation split contains 312 scans, and the testing set contains 100 scans. + +As done in previous works [21, 26, 38], we use the mIoU (mean Intersection over Union) as the evaluation metric. + +# 4.2 Implementation Details + +We follow the basic experimental settings of LaserMix [21] and GPC [17] for evaluating the proposed LASS3D on the outdoor-scene and indoor-scene datasets respectively. The classifiers and fusion adapter are both multi-layer perceptrons. We follow the settings of IGNet [38], which also adopts the MeanTeacher [37] framework, to set the update hyperparameter $\alpha$ as 0.999. The iteration number in progressive exploitation is set as 5. The hyperparameters $\tau_r, \tau_v, \tau_n, \lambda_u, \lambda_{PE}, \lambda_{KL}$ are set as 0.9, 0.05, 0.2, 0.1, 0.1, and 0.1 respectively. + +Table 1: Comparative results on the SemanticKITTI [1] and ScanNet [4] datasets with varying labeled ratios. $\mathcal{P}$ denotes the point cloud data, $\mathcal{I}$ denotes the image data, and $\mathcal{T}$ denotes the text data. All mIoU scores are given in percentage (\%). The best results are in bold and the second best results are marked with underlines. + +
MethodModalitySemanticKITTI [1]ScanNet [4]
1%10%20%50%5%10%20%30%40%
MeanTeacher [37]P45.457.159.260.048.757.266.167.369.1
CBST [55]P48.858.359.459.748.557.865.467.369.2
CPS [3]P46.758.759.660.551.758.666.468.070.5
SSS-Net [5]P-----52.455.1--
LaserMix (Range View) [21]P43.458.859.461.4-----
LaserMix (Voxel) [21]P50.660.061.962.3-----
IGNet [38]P, I49.061.363.164.8-----
GPC [17]P54.162.062.562.854.860.566.768.971.3
LiM3D [26]P58.462.263.163.6-----
LASS3DP, I, T58.563.064.164.556.663.167.470.472.0
+ +# 4.3 Comparative Evaluation + +We evaluate the proposed LASS3D on both outdoor-scene (SemanticKITTI [1]) and indoor-scene (ScanNet [4]) datasets in comparison to some classic semi-supervised methods [3, 37, 55] extended from the 2D domain and the state-of-the-art (SOTA) methods [5, 17, 21, 26, 38] that are specially designed for semi-supervised 3D semantic segmentation. Note that IGNet [38], LaserMix [21], and LiM3D [26] utilize some specific characters of the LiDAR point clouds, and thus cannot be applied to the ScanNet dataset. In addition, SSS-Net [5] leverages the color information of the point clouds, and thus cannot be applied to the SemantiKITTI dataset where the color information is unavailable. + +Following the settings in [17, 21, 38], we set the labeled ratio of the outdoor-scene dataset and indoor-scene dataset as $\{1\%, 10\%, 20\%, 40\% \}$ and $\{5\%, 10\%, 20\%, 30\%, 50\% \}$ respectively for a fair comparison. The comparative results are reported in Tab. 1. As seen from this table, the proposed LASS3D outperforms all the comparative methods that use uni-modal data as input, which demonstrates the effectiveness of LASS3D. In addition, LASS3D outperforms IGNet which takes both point clouds and images as input in most cases, which indicates that the proposed language-assisted method can better boost the 3D segmentation performances and may provide some insights on the exploitation of multi-modal data in semi-supervised 3D semantic segmentation. + +Moreover, the segmentation results of LASS3D and the second-best methods are visualized in Fig. 6. As seen from this figure, LASS3D outperforms its second-best counterparts, demonstrating the effectiveness of LASS3D from the qualitative perspective. + +![](images/aecd1f2ebeff5c4e1d5807621f131766f34d75e2733b7b363547f47ac9537ae4.jpg) +Fig. 6: Qualitative results on the SemanticKITTI [1] (top) and ScanNet [4] (bottom) datasets by LASS3D and the second-best methods. All models are trained with $10\%$ labeled data. We use the red boxes to highlight the areas where LASS3D evidently outperforms the comparative methods. + +Table 2: Ablation of the loss terms. + +
LCElLCEuLKLLPE1%10%20%50%
✓49.658.860.761.2
✓✓53.760.161.662.8
✓✓✓55.661.462.863.7
✓✓✓✓58.563.064.164.5
+ +Table 3: Ablation of captions. + +
CiCe1%10%20%50%
50.059.861.562.1
✓53.660.561.963.0
✓54.161.062.763.8
✓✓58.563.064.164.5
+ +# 4.4 Ablation Study + +We conduct ablation studies on SemanticKITTI [1] to investigate the effect of the key elements in the proposed LASS3D. + +Effect of the loss terms. We investigate the effect of the loss terms and report the corresponding results in Tab. 2. As seen from the first two rows of Tab. 2, incorporating the unlabeled data into training can improve the performances. Then, as seen in the third row of Tab. 2, the $\mathcal{L}_{KL}^u$ is introduced into training and the performances are improved, indicating that the semantic information distilled from the text-enhanced 3D features is helpful for segmentation. Finally, in the last row of Tab. 2, the progressive exploitation is adopted, and the performances are further boosted, demonstrating that the proposed progressive exploitation can make the unreliable points beneficial to the model training. + +Effect of different captions. We investigate the effect of different captions and report the corresponding results in Tab. 3. As seen from this table, both the image-level captions $\mathcal{C}_i$ and entity-level captions $\mathcal{C}_e$ can improve the segmentation performances, indicating that the semantic information encoded in text embeddings can facilitate the point cloud segmentation. When combining $\mathcal{C}_i$ and $\mathcal{C}_e$ , the best performance is achieved. Because the $\mathcal{C}_i$ provides more holistic semantic information and $\mathcal{C}_e$ provides more fine-grained semantic information, which are complementary to each other. Thus, we utilize $\mathcal{C}_i$ and $\mathcal{C}_e$ in LASS3D. + +Effect of the fusion strategy for the image-level embeddings $\mathcal{E}_i$ . As mentioned in Sec. 3.2, we adopt the cross-attention operation to fuse $\mathcal{E}_i$ with the 3D features. To investigate the effect of the fusion strategy, we replace the cross-attention operation with add and concatenation operations. The corresponding results are reported in Tab. 4. As seen from this table, the model trained with the cross-attention operation achieves the best performance. Because the image-level captions usually contain holistic semantic information and + +Table 4: Ablation of fusion strategy. +Table 5: Ablation of separation strategy. + +
Strategy1%10%20%50%Strategy1%10%20%50%
Add54.061.262.363.5Confidence-based57.762.363.564.1
Concatenation56.262.063.463.9Variance-based57.062.162.863.5
Cross attention58.563.064.164.5Ours58.563.064.164.5
+ +may have different correlations with different regions of the paired points. Simply adding or concatenating $\mathcal{E}_i$ to all the paired 3D features leads to the lack of discrimination, while 3D semantic segmentation is a point-level prediction task that requires discriminative representations for each point. Cross attention can adaptively fuse $\mathcal{E}_i$ with 3D features and produce more discriminative features, which makes it a more suitable fusion strategy for $\mathcal{E}_i$ . + +Effectiveness of the semantic-aware adaptive fusion module (SAFM). As mentioned in Sec. 2, existing language-assisted point cloud learning methods are either projection-based or alignment-based. Project-based methods are generally designed for object-level 3D classification and are not applicable to scene-level 3D semantic segmentation. To verify the effectiveness of the proposed SAFM, we replace it with an alignment-based method (e.g., the semantic consistency regularization in CLIP2Scene [2]). The corresponding results are reported in Tab. 6. As seen from this table, the model trained with SAFM outperforms the model trained with semantic consistency regularization. Probably because a large distribution discrepancy exists between the 3D features and text embeddings, simply aligning the 3D features to their corresponding text embeddings may result in the loss of some inherent geometric information. SAFM remedies this defect by adaptive fusion, and thus achieves better performance. + +Effect of the separation strategy. As mentioned in Sec. 3.3, we adopt both confidences and variances to measure the reliability of the predictions and separate the unlabeled points into the reliable set and unreliable set. Here we evaluate the effectiveness of our separation strategy. The corresponding results are reported in Tab. 5. As seen from this table, the model trained with our strategy achieves the best performance. Probably because the confidence-based strategy alone is unable to filter out the highly confident wrong predictions, which may confuse the training. The variance-based strategy separates points according to the prediction consistency in different branches, which could be complementary to the confidence-based strategy in measuring reliability. Thus, we combine the above two strategies to measure reliability in a more comprehensive way, which can facilitate the model training. + +Effect of the iteration number. The iteration number of progressive exploitation determines the utilization degree of the unreliable data. Here, we evaluate LASS3D with the iteration number set as $\{0,1,2,3,5,7,9,11\}$ , and the corresponding results are shown in Fig. 7. As seen from this figure, the performances are generally improved with the increase of iteration number. Because a larger iteration number means that the exploitation of the unreliable data is more thor- + +Table 6: Ablation of semantic-aware adap- Table 7: Ablation of the proposed pro- tive fusion module (SAFM). progressive exploitation (PE). + +
Strategy1%10%20%50%Model1%10%20%50%
Alignment-based53.060.161.862.2GPC [17]54.162.062.562.8
SAFM58.563.064.164.5GPC with PE54.762.863.163.5
+ +![](images/7c22a04ed498dce9ad861404596b94fd1111fb421b03788c12325295e3c01586.jpg) +Fig. 7: Ablation of the iteration number. Note that the iteration here denotes the negative learning iteration in progressive exploitation, rather than the training iteration. + +ough and thus brings better performance. However, when the iteration number reaches a certain level, the segmentation performance improves slightly. Probably because many unreliable points turn into reliable points during the progressive exploitation process, with most of their confident negative classes discarded. In addition, more iterations inevitably bring more training costs. Thus, we set the iteration number as 5 for a trade-off between the performance and training cost. + +Effectiveness of the progressive exploitation. Considering that the proposed progressive exploitation strategy only requires the prediction confidence and variances, thus it can be seamlessly integrated into some existing methods. We integrate the progressive exploitation strategy into GPC [17] and report the comparative results in Tab. 7. As seen from this table, the proposed progressive exploitation strategy can further boost the performance of GPC, demonstrating the effectiveness of the proposed strategy for exploiting unreliable points. + +# 4.5 Limitation + +The proposed LASS3D has demonstrated its effectiveness in segmentation, however, it still has limitations. For example, LASS3D uses images as the bridge to connect the text data and point clouds, which is a complex operation and may bring accumulated errors in point-text matching. A more direct and accurate method to construct the point-text pairs is in demand. + +# 5 Conclusion + +In this paper, we propose a language-assisted semi-supervised 3D semantic segmentation method named LASS3D. In LASS3D, a semantic-aware adaptive fusion module is explored to fuse the 3D features with the embeddings of multi-level captions generated by language-vision models, which can inject the semantic information into the 3D features and thus facilitate 3D segmentation. In addition, a progressive exploitation strategy is explored for unlabeled unreliable data, which can further boost performance and can be seamlessly embedded into other methods. Experimental results on outdoor and indoor datasets show that LASS3D outperforms the comparative methods in most cases. + +Acknowledgements. This work was supported by the National Natural Science Foundation of China (Grant Nos. 61991423, 62376269, 62073199); the Strategic Priority Research Program of the Chinese Academy of Sciences (Grant No. XDA27040811); the Open Project Program of Key Laboratory of Industrial Internet and Big Data, China National Light Industry, Beijing Technology and Business University. + +# References + +1. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semanticickitti: A dataset for semantic scene understanding of lidar sequences. In: ICCV. pp. 9297-9307 (2019) +2. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: Clip2scene: Towards label-efficient 3d scene understanding by clip. In: CVPR. pp. 7020-7030 (2023) +3. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: CVPR. pp. 2613-2622 (2021) +4. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR. pp. 5828-5839 (2017) +5. Deng, S., Dong, Q., Liu, B., Hu, Z.: Superpoint-guided semi-supervised semantic segmentation of 3d point clouds. In: ICRA. pp. 9214-9220 (2022) +6. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: CVPR. pp. 7010-7019 (2023) +7. Fan, S., Dong, Q., Zhu, F., Lv, Y., Ye, P., Wang, F.: Scf-net: Learning spatial contextual features for large-scale point cloud segmentation. In: CVPR. pp. 14499-14508 (2021) +8. Graham, B., Engelcke, M., van der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. CVPR pp. 9224-9232 (2018) +9. Guo, Z., Zhang, R., Qiu, L., Ma, X., Miao, X., He, X., Cui, B.: Calip: Zero-shot enhancement of clip with parameter-free attention. In: AAAI. vol. 37, pp. 746-754 (2023) +0. He, W., Jamonnak, S., Gou, L., Ren, L.: Clip-s4: Language-guided self-supervised semantic segmentation. In: CVPR. pp. 11207-11216 (2023) +1. Hegde, D., Valanarasu, J.M.J., Patel, V.: Clip goes 3d: Leveraging prompt tuning for language grounded 3d recognition. In: ICCVW. pp. 2028-2038 (2023) +2. Hess, G., Tondersski, A., Petersson, C., Åström, K., Svensson, L.: Lidarclip or: How i learned to talk to point clouds. In: WACV. pp. 7438-7447 (2024) +3. Hu, Q., Yang, B., Fang, G., Guo, Y., Leonardis, A., Trigoni, N., Markham, A.: Sqn: Weakly-supervised semantic segmentation of large-scale 3d point clouds. In: ECCV. pp. 600-619. Springer (2022) +4. Huang, R., Pan, X., Zheng, H., Jiang, H., Xie, Z., Wu, C., Song, S., Huang, G.: Joint representation learning for text and 3d point cloud. PR 147, 110086 (2024) +5. Huang, T., Dong, B., Yang, Y., Huang, X., Lau, R.W., Ouyang, W., Zuo, W.: Clip2point: Transfer clip to point cloud classification with image-depth pretraining. In: ICCV. pp. 22157-22167 (2023) +6. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML. pp. 4904-4916 (2021) + +17. Jiang, L., Shi, S., Tian, Z., Lai, X., Liu, S., Fu, C.W., Jia, J.: Guided point contrastive learning for semi-supervised point cloud semantic segmentation. In: ICCV. pp. 6423-6432 (2021) +18. Jiao, S., Wei, Y., Wang, Y., Zhao, Y., Shi, H.: Learning mask-aware CLIP representations for zero-shot segmentation. In: NeurIPS (2023) +19. Kim, Y., Yim, J., Yun, J., Kim, J.: Nlnl: Negative learning for noisy labels. In: ICCV. pp. 101-110 (2019) +20. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV. pp. 4015-4026 (2023) +21. Kong, L., Ren, J., Pan, L., Liu, Z.: Lasermix for semi-supervised lidar semantic segmentation. In: CVPR. pp. 21705-21715 (2023) +22. Lai, X., Chen, Y., Lu, F., Liu, J., Jia, J.: Spherical transformer for lidar-based 3d recognition. In: CVPR. pp. 17545-17555 (2023) +23. Lai, X., Liu, J., Jiang, L., Wang, L., Zhao, H., Liu, S., Qi, X., Jia, J.: Stratified transformer for 3d point cloud segmentation. In: CVPR. pp. 8500-8509 (2022) +24. Li, J., Dong, Q.: Open-set semantic segmentation for point clouds via adversarial prototype framework. In: CVPR. pp. 9425-9434 (2023) +25. Li, J., Dong, Q.: Density-guided semi-supervised 3d semantic segmentation with dual-space hardness sampling. In: CVPR. pp. 3260-3269 (2024) +26. Li, L., Shum, H.P., Breckon, T.P.: Less is more: Reducing task and model complexity for 3d point cloud semantic segmentation. In: CVPR. pp. 9361-9371 (2023) +27. Lin, Z., Geng, S., Zhang, R., Gao, P., de Melo, G., Wang, X., Dai, J., Qiao, Y., Li, H.: Frozen clip models are efficient video learners. In: ECCV. pp. 388-404. Springer (2022) +28. Liu, L., Zhuang, Z., Huang, S., Xiao, X., Xiang, T., Chen, C., Wang, J., Tan, M.: Cpcm: Contextual point cloud modeling for weakly-supervised point cloud semantic segmentation. In: ICCV. pp. 18413-18422 (2023) +29. Liu, Z., Qi, X., Fu, C.W.: One thing one click: A self-training approach for weakly supervised 3d semantic segmentation. In: CVPR. pp. 1726-1736 (2021) +30. Lu, Y., Xu, C., Wei, X., Xie, X., Tomizuka, M., Keutzer, K., Zhang, S.: Open-vocabulary point-cloud object detection without 3d annotation. In: CVPR. pp. 1190-1199 (2023) +31. Ma, W., Li, S., Zhang, J., Liu, C.H., Kang, J., Wang, Y., Huang, G.: Borrowing knowledge from pre-trained language model: A new data-efficient visual learning paradigm. In: ICCV. pp. 18786-18797 (2023) +32. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023) +33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021) +34. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Denseclip: Language-guided dense prediction with context-aware prompting. In: CVPR. pp. 18082-18091 (2022) +35. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., Zeng, Z., Zhang, H., Li, F., Yang, J., Li, H., Jiang, Q., Zhang, L.: Grounded sam: Assembling open-world models for diverse visual tasks (2024) +36. Tang, L., Chen, Z., Zhao, S., Wang, C., Tao, D.: All points matter: entropy-regularized distribution alignment for weakly-supervised 3d segmentation. NeurIPS 36 (2024) + +37. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NIPS 30 (2017) +38. Unal, O., Dai, D., Hoyer, L., Can, Y.B., Van Gool, L.: 2d feature distillation for weakly-and semi-supervised 3d semantic segmentation. In: WACV. pp. 7336-7345 (2024) +39. Wang, Y., Huang, S., Gao, Y., Wang, Z., Wang, R., Sheng, K., Zhang, B., Liu, S.: Transferring clip's knowledge into zero-shot point cloud semantic segmentation. In: ACM MM. pp. 3745-3754 (2023) +40. Wang, Y., Wang, H., Shen, Y., Fei, J., Li, W., Jin, G., Wu, L., Zhao, R., Le, X.: Semi-supervised semantic segmentation using unreliable pseudo-labels. In: CVPR. pp. 4248-4257 (2022) +41. Wu, X., Jiang, L., Wang, P.S., Liu, Z., Liu, X., Qiao, Y., Ouyang, W., He, T., Zhao, H.: Point transformer v3: Simpler, faster, stronger. In: CVPR (2024) +42. Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point transformer v2: Grouped vector attention and partition-based pooling. In: NeurIPS. vol. 35, pp. 33330-33342 (2022) +43. Xiang, P., Wen, X., Liu, Y.S., Zhang, H., Fang, Y., Han, Z.: Retro-fpn: Retrospective feature pyramid network for point cloud semantic segmentation. In: ICCV. pp. 17826-17838 (2023) +44. Xue, L., Gao, M., Xing, C., Martin-Martín, R., Wu, J., Xiong, C., Xu, R., Niebles, J.C., Savarese, S.: Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In: CVPR. pp. 1179-1189 (2023) +45. Yao, L., Han, J., Wen, Y., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, C., Xu, H.: Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. NeurIPS 35, 9125-9138 (2022) +46. Yu, W., Liu, Y., Hua, W., Jiang, D., Ren, B., Bai, X.: Turning a clip model into a scene text detector. In: CVPR. pp. 6978-6988 (2023) +47. Zeng, Y., Jiang, C., Mao, J., Han, J., Ye, C., Huang, Q., Yeung, D.Y., Yang, Z., Liang, X., Xu, H.: Clip2: Contrastive language-image-point pretraining from real-world point cloud data. In: CVPR. pp. 15244-15253 (2023) +48. Zhang, J., Fan, G., Wang, G., Su, Z., Ma, K., Yi, L.: Language-assisted 3d feature learning for semantic scene understanding. In: AAAI. pp. 3445-3453 (2023) +49. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: CVPR. pp. 8552-8562 (2022) +50. Zhang, Y., Huo, X., Chen, T., Wu, S., Wong, H.S.: Exploring intra-class variation factors with learnable cluster prompts for semi-supervised image synthesis. In: CVPR. pp. 7392-7401 (2023) +51. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV. pp. 16259-16268 (2021) +52. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: ECCV. pp. 696-712. Springer (2022) +53. Zhu, X., Zhang, R., He, B., Guo, Z., Zeng, Z., Qin, Z., Zhang, S., Gao, P.: Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In: ICCV. pp. 2639-2650 (2023) +54. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In: CVPR. pp. 9939-9948 (2021) +55. Zou, Y., Yu, Z., Kumar, B., Wang, J.: Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In: ECCV. pp. 289-305 (2018) \ No newline at end of file diff --git a/2024/LASS3D_ Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation/images.zip b/2024/LASS3D_ Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f818f7ec724e9ee1d0124236111a5a5996b65ca0 --- /dev/null +++ b/2024/LASS3D_ Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d393a8e8159fb494763166c3c2a1471fdef58d1e454ace02ad1f534a173859c8 +size 396057 diff --git a/2024/LASS3D_ Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation/layout.json b/2024/LASS3D_ Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bbff525f12262c349c6bf275f8f1fd2ef36c0cdd --- /dev/null +++ b/2024/LASS3D_ Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation/layout.json @@ -0,0 +1,9993 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 114, + 478, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 114, + 478, + 156 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 478, + 156 + ], + "type": "text", + "content": "LASS3D: Language-Assisted Semi-Supervised 3D Semantic Segmentation with Progressive Unreliable Data Exploitation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 222, + 178, + 392, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 178, + 392, + 192 + ], + "spans": [ + { + "bbox": [ + 222, + 178, + 392, + 192 + ], + "type": "text", + "content": "Jianan " + }, + { + "bbox": [ + 222, + 178, + 392, + 192 + ], + "type": "inline_equation", + "content": "\\mathrm{Li}^{1,2}" + }, + { + "bbox": [ + 222, + 178, + 392, + 192 + ], + "type": "text", + "content": " and Qiulei Dong" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 199, + 465, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 199, + 465, + 210 + ], + "spans": [ + { + "bbox": [ + 138, + 199, + 465, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 199, + 465, + 210 + ], + "type": "text", + "content": " School of Artificial Intelligence, University of Chinese Academy of Sciences" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 210, + 473, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 210, + 473, + 233 + ], + "spans": [ + { + "bbox": [ + 138, + 210, + 473, + 233 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 210, + 473, + 233 + ], + "type": "text", + "content": " State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 148, + 233, + 465, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 233, + 465, + 256 + ], + "spans": [ + { + "bbox": [ + 148, + 233, + 465, + 256 + ], + "type": "text", + "content": "3 Center for Excellence in Brain Science and Intelligence Technology, Chinese Academy of Sciences" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 194, + 256, + 418, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 256, + 418, + 266 + ], + "spans": [ + { + "bbox": [ + 194, + 256, + 418, + 266 + ], + "type": "text", + "content": "lijianan211@mails.ucas.ac.cn, qldong@nlpr.ac.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 297, + 455, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 297, + 455, + 540 + ], + "spans": [ + { + "bbox": [ + 159, + 297, + 455, + 540 + ], + "type": "text", + "content": "Abstract. Precisely annotating large-scale 3D datasets for point cloud segmentation is laborious. To alleviate the annotation burden, several semi-supervised 3D segmentation methods have been proposed in literature. However, two issues remain to be tackled: 1) The utilization of large language-vision models (LVM) in semi-supervised 3D semantic segmentation remains under-explored. 2) The unlabeled points with low-confidence predictions are directly discarded by existing methods. Taking these two issues into consideration, we propose a language-assisted semi-supervised 3D semantic segmentation method named LASS3D, which is built upon the commonly used MeanTeacher framework. In LASS3D, we use two off-the-shelf LVM to generate multi-level captions and leverage the images as the bridge to connect the text data and point clouds. Then, a semantic-aware adaptive fusion module is explored in the student branch, where the semantic information encoded in the embeddings of multi-level captions is injected into 3D features by adaptive fusion and then the semantic information in the text-enhanced 3D features is transferred to the teacher branch by knowledge distillation. In addition, a progressive exploitation strategy is explored for the unreliable points in the teacher branch, which can effectively exploit the information encapsulated in unreliable points via negative learning. Experimental results on both outdoor and indoor datasets demonstrate that LASS3D outperforms the comparative methods in most cases." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 550, + 453, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 550, + 453, + 562 + ], + "spans": [ + { + "bbox": [ + 160, + 550, + 453, + 562 + ], + "type": "text", + "content": "Keywords: 3D semantic segmentation " + }, + { + "bbox": [ + 160, + 550, + 453, + 562 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 550, + 453, + 562 + ], + "type": "text", + "content": " Semi-supervised segmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 583, + 230, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 583, + 230, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 583, + 230, + 596 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 609, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 482, + 646 + ], + "type": "text", + "content": "3D Semantic Segmentation [7,22-24,42,43,51] is an important task for perceiving real-world environments, and intensive efforts have been devoted to this task in recent years. However, most of the existing works [8,22,41,42,51,54] in literature" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 474, + 214 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 474, + 214 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 474, + 214 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 474, + 214 + ], + "type": "image", + "image_path": "18270764dc2d370675187c775afcc2039a2ead85285408e22c0ab97be77c74d2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 225, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 258 + ], + "type": "text", + "content": "Fig. 1: Illustration of the proportions of reliable and unreliable points predicted by GPC [17] on ScanNet [4]. The horizontal axis denotes the labeled ratio and training iteration. The vertical axis denotes the proportion." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 282, + 479, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 479, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 479, + 307 + ], + "type": "text", + "content": "are fully-supervised methods, which require time-consuming and labor-intensive data annotation for training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 307, + 480, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 480, + 366 + ], + "type": "text", + "content": "This issue encourages researchers to investigate semi-supervised 3D semantic segmentation, where only a small amount of densely labeled data and a large amount of unlabeled data are provided for training. Existing works [5,17,21,26, 38] on semi-supervised 3D semantic segmentation have shown their effectiveness, but there is ample room for improvement in the following two aspects:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 366, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 462 + ], + "type": "text", + "content": "1) Utilization of language-vision models (LVM): Recently, LVM [16,33] trained on web-crawled image-text pairs has been applied in some downstream tasks [2, 31, 39, 50] to boost performance, owing to its strong representation ability. Text embeddings extracted by the LVM generally contain higher-level abstraction of visual concepts with rich semantic information, which is complementary to the 3D features that mainly contain geometric information and thus can facilitate mining the semantic information of unlabeled data. To our best knowledge, LVM has not been applied in semi-supervised 3D semantic segmentation yet." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 582 + ], + "type": "text", + "content": "2) Exploitation of unreliable points (unlabeled points with low-confidence predictions): The proportions of reliable and unreliable points predicted by the state-of-the-art semi-supervised 3D semantic segmentation method GPC [17] on ScanNet [4] dataset are illustrated in Fig. 1. As seen from this figure, a large number of unlabeled points are deemed unreliable, especially at the early training stage. Existing works [17,21] generally adopt a confidence-based filtering strategy to remove the unreliable points for guaranteeing the effectiveness and stability of the training process. However, simply discarding them will inevitably lose some useful information. Thus, a proper learning strategy is expected to be specially designed to exploit these unreliable points effectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 641 + ], + "type": "text", + "content": "Taking the above two aspects into consideration, we propose a method named LASS3D for semi-supervised 3D semantic segmentation, where the LVM and unreliable points are effectively exploited. The proposed LASS3D is built upon the MeanTeacher [37] framework, where the weights of the teacher branch are updated by the exponential moving average (EMA) of the student's weights." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "In the student branch, two LVM (i.e., KOSMOS-2 [32] and GroundedSAM [35]) are used to generate image-level and entity-level captions, and the images" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "text", + "content": "J. Li and Q. Dong" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": "are leveraged as the bridge to connect the text data and point clouds. Then, a semantic-aware adaptive fusion module (SAFM) is designed to assist point cloud learning with the semantic knowledge encoded in text data. Specifically, multi-level point-text pairs are constructed in SAFM to adaptively fuse the 3D features and the text embeddings of the image-level and entity-level captions, which could facilitate capturing both holistic and fine-grained semantic information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 438 + ], + "type": "text", + "content": "In the teacher branch, a fusion adapter is utilized to narrow the distribution gap between the 3D features extracted by the backbone network and the text-enhanced 3D features in the student branch. Combined with the EMA strategy, the semantic information encoded in the text-enhanced 3D features is distilled from the student branch to the teacher branch. Note that we inject semantic information from text embeddings into the 3D features by fusion rather than directly mapping the semantic labels from pixels to points. Because the semantic labels generated by the open-world segmentation models (e.g., SAM [20]) are generally in a one-hot form, which provide hard supervision. However, some generated synonymous labels may not exist in our desired label set (e.g., the generated label is 'couch' but the ground-truth label is 'sofa'), and incorrect hard supervision may adversely affect the point cloud segmentation. Adaptively fusing the text embeddings and 3D features can mitigate this negative effect to some extent. Because the strong LVM can encode synonymous words into similar embeddings and thus can inject similar semantic information into 3D features. To effectively exploit the unlabeled points with unreliable predictions, a progressive exploitation strategy is designed based on the fact that determining the least likely category for an unlabeled sample is generally easier and more accurate than determining its most likely category. In addition, the proposed progressive exploitation strategy can be seamlessly integrated into some existing methods and further boost their performances, which is demonstrated in Sec. 4.4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 439, + 405, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 439, + 405, + 451 + ], + "spans": [ + { + "bbox": [ + 146, + 439, + 405, + 451 + ], + "type": "text", + "content": "In summary, the contributions of this paper are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 456, + 481, + 585 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 456, + 481, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 481, + 502 + ], + "type": "text", + "content": "- We propose a semantic-aware adaptive fusion module (SAFM) to consolidate point cloud segmentation with the semantic information encoded in text embeddings. The multi-level fusion shows the feasibility and effectiveness of language-assisted 3D semantic segmentation in semi-supervised settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 503, + 480, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 480, + 537 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 480, + 537 + ], + "type": "text", + "content": "- We propose a progressive exploitation strategy for the unlabeled unreliable points. This strategy is beneficial to fully exploit the unlabeled data and can be seamlessly embedded into some existing methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 538, + 480, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 538, + 480, + 585 + ], + "spans": [ + { + "bbox": [ + 138, + 538, + 480, + 585 + ], + "type": "text", + "content": "- We propose the LASS3D for semi-supervised 3D semantic segmentation by integrating the above SAFM and progressive exploitation strategy. Its effectiveness is demonstrated by the results on both outdoor-scene and indoor-scene public datasets in Sec. 4." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 601, + 237, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 237, + 613 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 237, + 613 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 624, + 381, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 381, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 381, + 635 + ], + "type": "text", + "content": "2.1 Semi-supervised 3D Semantic Segmentation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "To alleviate the annotation burden, label-efficient 3D semantic segmentation (including weakly-supervised segmentation [13,28,29,36] and semi-supervised seg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "LASS3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "type": "text", + "content": "mentation [5,17,21,25,26,38]) has drawn increasing interest among researchers. In this work, we delve into semi-supervised 3D semantic segmentation, which aims at utilizing a small number of densely labeled data and a large number of unlabeled data for model training. The core of semi-supervised 3D segmentation lies in exploiting the unlabeled data to the fullest. For example, Deng et al. [5] proposed to optimize the pseudo labels of the unlabeled data with the superpoints generated by geometry-based and color-based region-growing algorithms. Jiang et al. [17] proposed a label-guided contrastive loss for the unlabeled data, which could enhance the feature representation ability of the model. Kong et al. [21] proposed to leverage the spatial prior of LiDAR point clouds and use a mixing operation to provide supervisory information for the unlabeled data. Li et al. [26] proposed to utilize the reflectivity-prior descriptors to generate high-quality pseudo labels for the unlabeled data. Unal et al. [38] proposed to distill high-level feature information from a synthetically trained 2D network for data-efficient LiDAR semantic segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 296, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 344 + ], + "type": "text", + "content": "However, these methods neglect the usage of text data that contains rich semantic information. To our best knowledge, LASS3D is the first work to investigate the language-assisted semi-supervised 3D semantic segmentation, where the powerful representation ability of language-vision models is exploited." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 361, + 318, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 361, + 318, + 373 + ], + "spans": [ + { + "bbox": [ + 131, + 361, + 318, + 373 + ], + "type": "text", + "content": "2.2 Language-assisted 3D Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 380, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 475 + ], + "type": "text", + "content": "Large language-vision models (LVM) [16, 33] have drawn growing attention in computer vision recently, attributed to their powerful representation ability. For example, CLIP [33], which is trained on massive paired text-image data via contrastive learning, has prevailed in various downstream visual tasks, such as semantic segmentation [10, 18, 34, 48, 52], object detection [30, 45, 46], and video recognition [27]. Recent advanced methods that focused on language-assisted point cloud learning could be roughly divided into two types: the projection-based methods and the alignment-based methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 475, + 482, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 512 + ], + "type": "text", + "content": "The projection-based methods [9,15,49,53] project the point clouds into the 2D plane so that the image encoder in the LVM can be seamlessly utilized and the text-image alignment of the LVM can be leveraged for downstream tasks." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 138, + 534, + 316, + 616 + ], + "blocks": [ + { + "bbox": [ + 138, + 534, + 316, + 616 + ], + "lines": [ + { + "bbox": [ + 138, + 534, + 316, + 616 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 316, + 616 + ], + "type": "image", + "image_path": "6e0b8b61bd268728aeaf448be48b26bfb20157044f2632c27ce243bc71680f1c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 620, + 272, + 631 + ], + "lines": [ + { + "bbox": [ + 178, + 620, + 272, + 631 + ], + "spans": [ + { + "bbox": [ + 178, + 620, + 272, + 631 + ], + "type": "text", + "content": "(a) Pipeline of PointCLIP." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 320, + 536, + 481, + 620 + ], + "blocks": [ + { + "bbox": [ + 320, + 536, + 481, + 620 + ], + "lines": [ + { + "bbox": [ + 320, + 536, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 320, + 536, + 481, + 620 + ], + "type": "image", + "image_path": "06d6033f83a48b21784513dc72197b314b638be96ac67e9ef2e8839f71dfb0b0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 620, + 468, + 631 + ], + "lines": [ + { + "bbox": [ + 332, + 620, + 468, + 631 + ], + "spans": [ + { + "bbox": [ + 332, + 620, + 468, + 631 + ], + "type": "text", + "content": "(b) Alignment process of CLIP2Scene." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 639, + 481, + 663 + ], + "lines": [ + { + "bbox": [ + 130, + 639, + 481, + 663 + ], + "spans": [ + { + "bbox": [ + 130, + 639, + 481, + 663 + ], + "type": "text", + "content": "Fig. 2: Illustration of the pipeline of a projection-based method (e.g., PointCLIP [49]) and the alignment process of an alignment-based method (e.g., CLIP2Scene [2])." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "text", + "content": "J. Li and Q. Dong" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "Fig. 2a illustrates the general pipeline of PointCLIP [49], it classifies the point clouds according to the similarities between the text embeddings and features of the projected images. Nevertheless, the projection-based methods are generally designed for object-level 3D classification. In addition, they usually suffer from the loss of geometric information during the projection process, thus resulting in suboptimal performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 259 + ], + "type": "text", + "content": "The alignment-based methods [2,6,11,12,14,39,44,47] enforce the consistency among the multi-modal features. Fig. 2b illustrates the alignment process of CLIP2Scene [2], it uses contrastive learning to align the 3D features to their corresponding text embeddings. However, directly aligning multi-modal features may result in suboptimal results, due to the distribution discrepancy between the 3D features and text embeddings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 259, + 481, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 481, + 308 + ], + "type": "text", + "content": "Unlike existing methods that are either projection-based or alignment-based, we propose a fusion-based method to adaptively fuse the embeddings of multi-level captions and 3D features, which can alleviate the loss of geometric information and mitigate the negative effect brought by the distribution gap." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 325, + 233, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 325, + 233, + 339 + ], + "spans": [ + { + "bbox": [ + 131, + 325, + 233, + 339 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 350, + 226, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 350, + 226, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 350, + 226, + 361 + ], + "type": "text", + "content": "3.1 Architecture" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 369, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 418 + ], + "type": "text", + "content": "The architecture of the proposed LASS3D is illustrated in Fig. 3. It is built upon the MeanTeacher [37] framework where the teacher branch provides the supervisory signals for the student branch and is updated by the exponential mean average (EMA) of the student's weights:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 222, + 424, + 481, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 424, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 222, + 424, + 481, + 441 + ], + "type": "interline_equation", + "content": "\\theta_ {t e a c h e r} ^ {t + 1} = \\alpha \\theta_ {t e a c h e r} ^ {t} + (1 - \\alpha) \\theta_ {s t u d e n t} ^ {t}, \\tag {1}", + "image_path": "e4bfc32e8714afd27ff60841fb784eda9c81d4881f778bc9ff6339d0c98b8359.jpg" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 134, + 459, + 479, + 619 + ], + "blocks": [ + { + "bbox": [ + 134, + 459, + 479, + 619 + ], + "lines": [ + { + "bbox": [ + 134, + 459, + 479, + 619 + ], + "spans": [ + { + "bbox": [ + 134, + 459, + 479, + 619 + ], + "type": "image", + "image_path": "c51340eeb77e279aca57a369a8662d9b342a6055cc1a17692bdec13d6d6e408f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 628, + 482, + 663 + ], + "lines": [ + { + "bbox": [ + 130, + 628, + 482, + 663 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 482, + 663 + ], + "type": "text", + "content": "Fig. 3: Architecture of the proposed LASS3D. " + }, + { + "bbox": [ + 130, + 628, + 482, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 130, + 628, + 482, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 628, + 482, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_e" + }, + { + "bbox": [ + 130, + 628, + 482, + 663 + ], + "type": "text", + "content": " stand for the image-level captions and entity-level captions. " + }, + { + "bbox": [ + 130, + 628, + 482, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_i" + }, + { + "bbox": [ + 130, + 628, + 482, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 628, + 482, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_e" + }, + { + "bbox": [ + 130, + 628, + 482, + 663 + ], + "type": "text", + "content": " are the embeddings of " + }, + { + "bbox": [ + 130, + 628, + 482, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 130, + 628, + 482, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 628, + 482, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_e" + }, + { + "bbox": [ + 130, + 628, + 482, + 663 + ], + "type": "text", + "content": " respectively. " + }, + { + "bbox": [ + 130, + 628, + 482, + 663 + ], + "type": "inline_equation", + "content": "/ / /" + }, + { + "bbox": [ + 130, + 628, + 482, + 663 + ], + "type": "text", + "content": " denotes the detach operation which is used to stop the gradient." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 410, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 410, + 91, + 447, + 100 + ], + "type": "text", + "content": "LASS3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\theta_{teacher}^{t}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\theta_{student}^{t}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " denote the weights of the teacher branch and student branch at time step " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " is the update hyperparameter." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 480, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 480, + 235 + ], + "type": "text", + "content": "The student branch takes the point clouds and their corresponding images as input. The images are fed into the LVM, (i.e., KOSMOS-2 [35] and GroundedSAM [32]) to produce image-level " + }, + { + "bbox": [ + 130, + 141, + 480, + 235 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 130, + 141, + 480, + 235 + ], + "type": "text", + "content": " and entity-level captions " + }, + { + "bbox": [ + 130, + 141, + 480, + 235 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_e" + }, + { + "bbox": [ + 130, + 141, + 480, + 235 + ], + "type": "text", + "content": ". Then, the multi-level captions are tokenized by the text encoder of CLIP [33] to generate their corresponding text embeddings. The text embeddings and 3D features extracted by the 3D backbone network are fused in the semantic-aware adaptive fusion module to obtain the text-enhanced 3D features, which are then used for segmentation in the student branch." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 237, + 481, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 481, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 481, + 393 + ], + "type": "text", + "content": "The teacher branch takes the same point clouds as input. The 3D features extracted by the 3D backbone network are aligned with the text-enhanced 3D features through a fusion adapter. Then, the classifier takes the aligned 3D features as input and outputs their corresponding prediction scores. A dual-score separation strategy is adopted to select reliable points and unreliable points from the unlabeled points. The predictions of the reliable points are regarded as the pseudo labels of their corresponding points in the student branch, where the cross-entropy loss " + }, + { + "bbox": [ + 130, + 237, + 481, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}^{u}" + }, + { + "bbox": [ + 130, + 237, + 481, + 393 + ], + "type": "text", + "content": " is calculated. The predictions of the unreliable points in the student branch are used for the progressive exploitation, where the negative learning loss " + }, + { + "bbox": [ + 130, + 237, + 481, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{PE}^{u}" + }, + { + "bbox": [ + 130, + 237, + 481, + 393 + ], + "type": "text", + "content": " is calculated. Besides, the predictions of the labeled points in the student branch and their ground-truth labels are used to calculate the cross-entropy loss " + }, + { + "bbox": [ + 130, + 237, + 481, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}^{l}" + }, + { + "bbox": [ + 130, + 237, + 481, + 393 + ], + "type": "text", + "content": ", and the features output by the fusion adapter are aligned to the text-enhanced 3D features by a Kullback-Leibler divergence loss " + }, + { + "bbox": [ + 130, + 237, + 481, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KL}^{u}" + }, + { + "bbox": [ + 130, + 237, + 481, + 393 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 393, + 480, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 480, + 417 + ], + "type": "text", + "content": "Note that the text data is only required in the training stage. In the inference stage, only the teacher branch is used and it takes the point clouds as input." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 437, + 367, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 437, + 367, + 450 + ], + "spans": [ + { + "bbox": [ + 131, + 437, + 367, + 450 + ], + "type": "text", + "content": "3.2 Semantic-aware Adaptive Fusion Module" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 460, + 480, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 480, + 520 + ], + "type": "text", + "content": "The architecture of the semantic-aware adaptive fusion module (SAFM) is illustrated in Fig. 4, it takes the embeddings of multi-level captions (i.e., image-level embeddings " + }, + { + "bbox": [ + 130, + 460, + 480, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_i" + }, + { + "bbox": [ + 130, + 460, + 480, + 520 + ], + "type": "text", + "content": " and entity-level embeddings " + }, + { + "bbox": [ + 130, + 460, + 480, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_e" + }, + { + "bbox": [ + 130, + 460, + 480, + 520 + ], + "type": "text", + "content": ") and 3D features " + }, + { + "bbox": [ + 130, + 460, + 480, + 520 + ], + "type": "inline_equation", + "content": "f_{3D}" + }, + { + "bbox": [ + 130, + 460, + 480, + 520 + ], + "type": "text", + "content": " as input and outputs the text-enhanced 3D features " + }, + { + "bbox": [ + 130, + 460, + 480, + 520 + ], + "type": "inline_equation", + "content": "f_{3D}^t" + }, + { + "bbox": [ + 130, + 460, + 480, + 520 + ], + "type": "text", + "content": ", which are then fed into the classifier for segmentation and used as the supervision for knowledge distillation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 628 + ], + "type": "text", + "content": "The proposed SAFM is used to inject the semantic information encoded in multi-level captions into the 3D features. Specifically, the image-level captions " + }, + { + "bbox": [ + 130, + 521, + 481, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 130, + 521, + 481, + 628 + ], + "type": "text", + "content": " and entity-level captions " + }, + { + "bbox": [ + 130, + 521, + 481, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_e" + }, + { + "bbox": [ + 130, + 521, + 481, + 628 + ], + "type": "text", + "content": " are generated by KOSMOS-2 [32] and GroundedSAM [35] respectively, and their corresponding embeddings are extracted by the text encoder of CLIP [33]. To guarantee the accuracy of multi-modal fusion, we construct the point-text pairs by using the images as the bridge. Then, given the point-text pairs, their corresponding features are adaptively fused to produce the text-enhanced 3D features. In this subsection, we will introduce the above two key operations of SAFM in detail." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "Point-text pair construction. We use the images as the bridge to connect the text data and point cloud data. Specifically, the multi-level pixel-text pairs can be obtained by the language-vision models, i.e., the image-level captions" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "text", + "content": "J. Li and Q. Dong" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 113, + 463, + 209 + ], + "blocks": [ + { + "bbox": [ + 151, + 113, + 463, + 209 + ], + "lines": [ + { + "bbox": [ + 151, + 113, + 463, + 209 + ], + "spans": [ + { + "bbox": [ + 151, + 113, + 463, + 209 + ], + "type": "image", + "image_path": "74abd7d2ade34c9654f4a0b42b8b96c223192a277cf186cd2dd745c81431ba7e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 217, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 131, + 217, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 131, + 217, + 482, + 262 + ], + "type": "text", + "content": "Fig. 4: Architecture of the proposed semantic-aware adaptive fusion module. " + }, + { + "bbox": [ + 131, + 217, + 482, + 262 + ], + "type": "inline_equation", + "content": "f_{3D}" + }, + { + "bbox": [ + 131, + 217, + 482, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 217, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\hat{f}_{3D}" + }, + { + "bbox": [ + 131, + 217, + 482, + 262 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 217, + 482, + 262 + ], + "type": "inline_equation", + "content": "f_{3D}^{t}" + }, + { + "bbox": [ + 131, + 217, + 482, + 262 + ], + "type": "text", + "content": " denote the extracted 3D features, aligned 3D features, and text-enhanced 3D features. " + }, + { + "bbox": [ + 131, + 217, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_i" + }, + { + "bbox": [ + 131, + 217, + 482, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 217, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_e" + }, + { + "bbox": [ + 131, + 217, + 482, + 262 + ], + "type": "text", + "content": " are the embeddings of image-level and entity-level captions. " + }, + { + "bbox": [ + 131, + 217, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 131, + 217, + 482, + 262 + ], + "type": "text", + "content": " denotes the add operation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 288, + 480, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 480, + 324 + ], + "type": "text", + "content": "correspond to all pixels of the given images, and the entity-level captions correspond to the pixels of the segmented regions. The pixel-point pairs can be easily obtained with the extrinsic and intrinsic parameters of the cameras:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 246, + 335, + 481, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 335, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 246, + 335, + 481, + 350 + ], + "type": "interline_equation", + "content": "[ u, v, 1 ] ^ {\\mathrm {T}} = \\mathbf {K M} [ x, y, z, 1 ] ^ {\\mathrm {T}}, \\tag {2}", + "image_path": "f1ddaabae8af944feda7ca9ed1152dc354e0d8e59885f6bd174123f1c7abc69c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 361, + 480, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 480, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 480, + 409 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 361, + 480, + 409 + ], + "type": "inline_equation", + "content": "[u,v,1]^{\\mathrm{T}}" + }, + { + "bbox": [ + 130, + 361, + 480, + 409 + ], + "type": "text", + "content": " represents the homogeneous coordinates of a pixel, " + }, + { + "bbox": [ + 130, + 361, + 480, + 409 + ], + "type": "inline_equation", + "content": "[x,y,z,1]^{\\mathrm{T}}" + }, + { + "bbox": [ + 130, + 361, + 480, + 409 + ], + "type": "text", + "content": " represents its corresponding 3D point position in world coordinates, " + }, + { + "bbox": [ + 130, + 361, + 480, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 361, + 480, + 409 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 361, + 480, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 361, + 480, + 409 + ], + "type": "text", + "content": " are the intrinsic and extrinsic parameters of the cameras. With the multi-level pixel-text and pixel-point pairs, the multi-level point-text pairs can be obtained." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 411, + 481, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 481, + 555 + ], + "type": "text", + "content": "Adaptive fusion. To effectively utilize the multi-level point-text pairs, an adaptive fusion strategy is proposed. As shown in Fig. 4, the dimension of the 3D features " + }, + { + "bbox": [ + 130, + 411, + 481, + 555 + ], + "type": "inline_equation", + "content": "f_{3D}" + }, + { + "bbox": [ + 130, + 411, + 481, + 555 + ], + "type": "text", + "content": " is firstly expanded by an MLP layer to align the embedding space of the text encoder, given that the text encoder in LASS3D is fixed. Since the image-level captions may have different correlations with different regions of the paired points, cross attention is performed between the aligned 3D features " + }, + { + "bbox": [ + 130, + 411, + 481, + 555 + ], + "type": "inline_equation", + "content": "\\hat{f}_{3D}" + }, + { + "bbox": [ + 130, + 411, + 481, + 555 + ], + "type": "text", + "content": " and image-level embeddings " + }, + { + "bbox": [ + 130, + 411, + 481, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_i" + }, + { + "bbox": [ + 130, + 411, + 481, + 555 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 411, + 481, + 555 + ], + "type": "inline_equation", + "content": "\\hat{f}_{3D}" + }, + { + "bbox": [ + 130, + 411, + 481, + 555 + ], + "type": "text", + "content": " serving as the query vectors and " + }, + { + "bbox": [ + 130, + 411, + 481, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_i" + }, + { + "bbox": [ + 130, + 411, + 481, + 555 + ], + "type": "text", + "content": " serving as the key vectors and value vectors. Then, considering that the points related to the same entity-level captions share the same semantic information, we simply add the entity-level embeddings " + }, + { + "bbox": [ + 130, + 411, + 481, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_e" + }, + { + "bbox": [ + 130, + 411, + 481, + 555 + ], + "type": "text", + "content": " with their corresponding 3D features. Finally, an MLP layer is used to map the fused features to their original dimension and output the text-enhanced 3D features " + }, + { + "bbox": [ + 130, + 411, + 481, + 555 + ], + "type": "inline_equation", + "content": "f_{3D}^t" + }, + { + "bbox": [ + 130, + 411, + 481, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 555, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 481, + 617 + ], + "type": "text", + "content": "As shown in Fig. 3, the text-enhanced 3D features " + }, + { + "bbox": [ + 130, + 555, + 481, + 617 + ], + "type": "inline_equation", + "content": "f_{3D}^{t}" + }, + { + "bbox": [ + 130, + 555, + 481, + 617 + ], + "type": "text", + "content": " are not only fed into the classifier of the student branch for segmentation but also serve as the supervisory signals for knowledge distillation in the teacher branch. The features " + }, + { + "bbox": [ + 130, + 555, + 481, + 617 + ], + "type": "inline_equation", + "content": "f_{ada}" + }, + { + "bbox": [ + 130, + 555, + 481, + 617 + ], + "type": "text", + "content": " output by the fusion adapter in the teacher branch are aligned to their corresponding " + }, + { + "bbox": [ + 130, + 555, + 481, + 617 + ], + "type": "inline_equation", + "content": "f_{3D}^{t}" + }, + { + "bbox": [ + 130, + 555, + 481, + 617 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 130, + 555, + 481, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KL}^{u}" + }, + { + "bbox": [ + 130, + 555, + 481, + 617 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 256, + 628, + 481, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 628, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 256, + 628, + 481, + 643 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {K L} ^ {u} = \\mathrm {K L} \\left(f _ {3 D} ^ {t} \\mid \\mid f _ {a d a}\\right), \\tag {3}", + "image_path": "00cbab5e631f357e7ead72178e3a70e93259990e3381afa1e18bb5bdf1e4b0ab.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 653, + 387, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 653, + 387, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 387, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 653, + 387, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{KL}(\\cdot)" + }, + { + "bbox": [ + 131, + 653, + 387, + 666 + ], + "type": "text", + "content": " denotes the Kullback-Leibler divergence loss." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 90, + 447, + 100 + ], + "type": "text", + "content": "LASS3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 116, + 480, + 195 + ], + "blocks": [ + { + "bbox": [ + 134, + 116, + 480, + 195 + ], + "lines": [ + { + "bbox": [ + 134, + 116, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 480, + 195 + ], + "type": "image", + "image_path": "37c560e7b39e5d0cbab11a466d1647bce478e3680bce9ba389ee0e3bd2fb1171.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "text", + "content": "Fig. 5: Illustration of the progressive exploitation for the unreliable point " + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "inline_equation", + "content": "P_{url}^{i}" + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "text", + "content": " at the first two iterations. " + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "inline_equation", + "content": "S^{i}" + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "text", + "content": " denotes the prediction confidence of " + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "inline_equation", + "content": "P_{url}^{i}" + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "text", + "content": " denotes its negative label. The selected negative class (blue font) is used for negative learning and the used negative class (marked with red dashed box) is discarded in the subsequent iterations. Note that some unreliable points may turn into reliable points during progressive exploitation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 293, + 330, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 330, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 330, + 304 + ], + "type": "text", + "content": "3.3 Progressive Exploitation Strategy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 312, + 479, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 479, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 479, + 360 + ], + "type": "text", + "content": "The progressive exploitation strategy is proposed to mine the useful information encapsulated in the unlabeled unreliable points. It contains two key steps: dual-score separation and progressive negative learning. In this subsection, we will introduce the above two steps in detail." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 361, + 480, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 480, + 433 + ], + "type": "text", + "content": "Dual-score separation. Unlike GPC [17] and LaserMix [21] which only use the confidence of predictions to measure the reliability, we propose a dual-score separation strategy that additionally introduces the variances among the predictions of different branches into reliability measurement. Specifically, the unlabeled points " + }, + { + "bbox": [ + 130, + 361, + 480, + 433 + ], + "type": "inline_equation", + "content": "P_{u}" + }, + { + "bbox": [ + 130, + 361, + 480, + 433 + ], + "type": "text", + "content": " in the teacher branch are separated into the reliable set " + }, + { + "bbox": [ + 130, + 361, + 480, + 433 + ], + "type": "inline_equation", + "content": "P_{rl}" + }, + { + "bbox": [ + 130, + 361, + 480, + 433 + ], + "type": "text", + "content": " and unreliable set " + }, + { + "bbox": [ + 130, + 361, + 480, + 433 + ], + "type": "inline_equation", + "content": "P_{url}" + }, + { + "bbox": [ + 130, + 361, + 480, + 433 + ], + "type": "text", + "content": " based on both confidence and variances:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 219, + 441, + 369, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 441, + 369, + 453 + ], + "spans": [ + { + "bbox": [ + 219, + 441, + 369, + 453 + ], + "type": "interline_equation", + "content": "P _ {r l} = B \\cdot P _ {u}, P _ {u r l} = (1 - B) \\cdot P _ {u},", + "image_path": "8b8bb147839d2a5e2aa47240f2090082334a9234de0128197211efe9bc933ae9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 220, + 456, + 480, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 456, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 220, + 456, + 480, + 489 + ], + "type": "interline_equation", + "content": "B = \\mathbb {1} \\left[ \\sum_ {c = 1} ^ {C} \\mathbb {1} \\left[ S _ {c} \\geq \\tau_ {r} \\right] \\cdot \\mathbb {1} \\left[ \\sigma_ {c} \\leq \\tau_ {v} \\right] > 0 \\right], \\tag {4}", + "image_path": "049b0bbe65ed074fae1e5ad58ccff0f772be68fd81cf18f1972b120f7e8af09c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "text", + "content": " is a binary mask to select " + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "inline_equation", + "content": "P_{rl}" + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "inline_equation", + "content": "P_{url}" + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "inline_equation", + "content": "\\mathbb{1}" + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "text", + "content": " is the indicator function, " + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "text", + "content": " denotes the number of classes, " + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "inline_equation", + "content": "S_{c}" + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "text", + "content": " denotes the prediction confidences of class " + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "text", + "content": " in the teacher branch, " + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "inline_equation", + "content": "\\sigma_{c}" + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "text", + "content": " denotes the prediction variances of class " + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "text", + "content": " within the teacher and student branches, " + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "inline_equation", + "content": "\\tau_{r}" + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "inline_equation", + "content": "\\tau_{v}" + }, + { + "bbox": [ + 130, + 497, + 480, + 544 + ], + "type": "text", + "content": " are two pre-determined thresholds." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 480, + 593 + ], + "type": "text", + "content": "The predictions of reliable points are regarded as the pseudo labels of their corresponding points in the student branch, which are used for supervised training, as shown in Eq. (8). The unreliable points are fully exploited via progressive negative learning, which is elaborated below." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "Progressive negative learning. Enlightened by the core idea of conventional negative learning [19, 40], we choose to mine information from the classes that the unreliable points do not belong to, rather than struggling to correctly tell which class they belong to. However, conventional methods only choose one negative class (the class with the lowest prediction score) for negative learning, which provides limited information and may cause unstable training due to the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "text", + "content": "J. Li and Q. Dong" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 139 + ], + "type": "text", + "content": "dynamic negative labels. Instead, we conduct negative learning in a progressive manner, with more negative classes selected." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 140, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 140, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 131, + 140, + 479, + 187 + ], + "type": "text", + "content": "The progressive exploitation at the first two iterations is illustrated in Fig. 5. As shown in this figure, we maintain a list for each unreliable point to record its selected negative classes. In each iteration, one negative class is selected for each unreliable point:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 202, + 194, + 479, + 223 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 194, + 479, + 223 + ], + "spans": [ + { + "bbox": [ + 202, + 194, + 479, + 223 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {i} [ k ] = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} k = \\arg \\min \\left(S ^ {i}\\right) \\text {a n d} S ^ {i} [ k ] \\leq \\tau_ {n} \\\\ 0, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {5}", + "image_path": "d134168b2d679004f507f418145c373afb49e5932786f976a0e20d8b8efbf38e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "spans": [ + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "inline_equation", + "content": "\\hat{y}_i[k]" + }, + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "text", + "content": "-th dimension of the negative label of " + }, + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "inline_equation", + "content": "P_{url}^i" + }, + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "inline_equation", + "content": "S^i" + }, + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "text", + "content": " denotes the prediction confidence of " + }, + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "inline_equation", + "content": "P_{url}^i" + }, + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "inline_equation", + "content": "\\tau_n" + }, + { + "bbox": [ + 131, + 230, + 479, + 254 + ], + "type": "text", + "content": " is a pre-determined threshold." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 254, + 395, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 254, + 395, + 266 + ], + "spans": [ + { + "bbox": [ + 146, + 254, + 395, + 266 + ], + "type": "text", + "content": "The selected negative label is used for negative learning:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 247, + 272, + 479, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 272, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 247, + 272, + 479, + 304 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {P E} ^ {u} = - \\sum_ {i = 1} ^ {N _ {u}} \\hat {y} _ {i} \\log \\left(1 - O _ {i}\\right), \\tag {6}", + "image_path": "f4140342dd13e98c0bad4a0bbf9b31e91772b4c38a1615460acc58119af8513a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 308, + 479, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 308, + 479, + 379 + ], + "spans": [ + { + "bbox": [ + 131, + 308, + 479, + 379 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 308, + 479, + 379 + ], + "type": "inline_equation", + "content": "N_{u}" + }, + { + "bbox": [ + 131, + 308, + 479, + 379 + ], + "type": "text", + "content": " is the number of unreliable points with selected negative class and " + }, + { + "bbox": [ + 131, + 308, + 479, + 379 + ], + "type": "inline_equation", + "content": "O_{i}" + }, + { + "bbox": [ + 131, + 308, + 479, + 379 + ], + "type": "text", + "content": " is the prediction of " + }, + { + "bbox": [ + 131, + 308, + 479, + 379 + ], + "type": "inline_equation", + "content": "P_{url}^{i}" + }, + { + "bbox": [ + 131, + 308, + 479, + 379 + ], + "type": "text", + "content": ". Then, the used negative class is discarded in the subsequent iterations. Simultaneously, the network is optimized via the negative learning loss. Subsequently, a new negative class is selected from the remaining classes for the next iteration. The above process is repeated until no negative class can be selected or the iteration number is reached." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 380, + 479, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 380, + 479, + 416 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 479, + 416 + ], + "type": "text", + "content": "Moreover, it is worth noting that the proposed progressive exploitation strategy can be seamlessly integrated into methods that discard the unreliable points, and further boost their performances, as shown in Sec. 4.4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 431, + 212, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 212, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 212, + 441 + ], + "type": "text", + "content": "3.4 Total Loss" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 449, + 479, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 449, + 479, + 472 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 479, + 472 + ], + "type": "text", + "content": "The labeled points in the student branch are trained in a supervised manner, with their corresponding ground-truth labels " + }, + { + "bbox": [ + 131, + 449, + 479, + 472 + ], + "type": "inline_equation", + "content": "y^{l}" + }, + { + "bbox": [ + 131, + 449, + 479, + 472 + ], + "type": "text", + "content": " being the supervisory signals:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 260, + 478, + 479, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 478, + 479, + 491 + ], + "spans": [ + { + "bbox": [ + 260, + 478, + 479, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C E} ^ {l} = \\operatorname {C E} \\left(O ^ {s t u, l}, y ^ {l}\\right), \\tag {7}", + "image_path": "8067a83d805e99a3c694214402d6097e15e85316a562d90600bdba09f93f4b41.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 496, + 479, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 496, + 479, + 519 + ], + "spans": [ + { + "bbox": [ + 131, + 496, + 479, + 519 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 496, + 479, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{CE}(\\cdot)" + }, + { + "bbox": [ + 131, + 496, + 479, + 519 + ], + "type": "text", + "content": " denotes the cross-entropy loss, " + }, + { + "bbox": [ + 131, + 496, + 479, + 519 + ], + "type": "inline_equation", + "content": "O^{stu,l}" + }, + { + "bbox": [ + 131, + 496, + 479, + 519 + ], + "type": "text", + "content": " denotes the predictions of the labeled points in the student branch." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 521, + 479, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 479, + 544 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 479, + 544 + ], + "type": "text", + "content": "The predictions of the reliable points in the teacher branch are regarded as the pseudo labels of their corresponding points in the student branch:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 257, + 551, + 479, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 551, + 479, + 563 + ], + "spans": [ + { + "bbox": [ + 257, + 551, + 479, + 563 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C E} ^ {u} = \\operatorname {C E} \\left(O ^ {s t u, u}, y ^ {u}\\right), \\tag {8}", + "image_path": "ce170bb9bbdf7f159bcf40f9c1a0c62edbc3cd4b3bd7f6b44bcb08e64ade1806.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 568, + 479, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 568, + 479, + 603 + ], + "spans": [ + { + "bbox": [ + 131, + 568, + 479, + 603 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 568, + 479, + 603 + ], + "type": "inline_equation", + "content": "y^{u}" + }, + { + "bbox": [ + 131, + 568, + 479, + 603 + ], + "type": "text", + "content": " denotes the reliable predictions output by the teacher branch and " + }, + { + "bbox": [ + 131, + 568, + 479, + 603 + ], + "type": "inline_equation", + "content": "O^{stu,u}" + }, + { + "bbox": [ + 131, + 568, + 479, + 603 + ], + "type": "text", + "content": " denotes the predictions of corresponding unlabeled points in the student branch." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 605, + 479, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 479, + 630 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 479, + 630 + ], + "type": "text", + "content": "Accordingly, the total loss " + }, + { + "bbox": [ + 132, + 605, + 479, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{total}" + }, + { + "bbox": [ + 132, + 605, + 479, + 630 + ], + "type": "text", + "content": " is the weighted sum of the two loss terms " + }, + { + "bbox": [ + 132, + 605, + 479, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}^{l}" + }, + { + "bbox": [ + 132, + 605, + 479, + 630 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 605, + 479, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}^{u}" + }, + { + "bbox": [ + 132, + 605, + 479, + 630 + ], + "type": "text", + "content": " in the student branch and the aforementioned " + }, + { + "bbox": [ + 132, + 605, + 479, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{PE}^{u}" + }, + { + "bbox": [ + 132, + 605, + 479, + 630 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 605, + 479, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KL}^{u}" + }, + { + "bbox": [ + 132, + 605, + 479, + 630 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 205, + 635, + 479, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 635, + 479, + 647 + ], + "spans": [ + { + "bbox": [ + 205, + 635, + 479, + 647 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {C E} ^ {l} + \\lambda_ {u} \\mathcal {L} _ {C E} ^ {u} + \\lambda_ {P E} \\mathcal {L} _ {P E} ^ {u} + \\lambda_ {K L} \\mathcal {L} _ {K L} ^ {u}, \\tag {9}", + "image_path": "1696848523c93a24fd72b2979803bc774c3a2d826fce84604f8ee96dda9807b6.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 131, + 653, + 375, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 653, + 375, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 375, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 653, + 375, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_{u},\\lambda_{PE},\\lambda_{KL}" + }, + { + "bbox": [ + 131, + 653, + 375, + 666 + ], + "type": "text", + "content": " are the weights of " + }, + { + "bbox": [ + 131, + 653, + 375, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}^{u},\\mathcal{L}_{PE}^{u},\\mathcal{L}_{KL}^{u}" + }, + { + "bbox": [ + 131, + 653, + 375, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 410, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 410, + 91, + 447, + 100 + ], + "type": "text", + "content": "LASS3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 224, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 224, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 224, + 129 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 321, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 321, + 153 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 321, + 153 + ], + "type": "text", + "content": "4.1 Datasets and Evaluation Metric" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 163, + 481, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 163, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 132, + 163, + 481, + 186 + ], + "type": "text", + "content": "The following outdoor-scene and indoor-scene datasets are used to evaluate the proposed LASS3D:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 188, + 481, + 307 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 188, + 481, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 188, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 138, + 188, + 481, + 258 + ], + "type": "text", + "content": "- SemanticKITTI [1] is a large-scale 3D outdoor-scene LiDAR dataset consisting of 22 sequences, among which 10 sequences are used for training, 1 sequence is used for validation, and 11 sequences are used for testing. It contains 19 categories after merging and ignoring classes with few points. According to the official splitting, 850 scenes are used for training and validation, and 150 scenes are utilized for testing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 259, + 481, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 259, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 138, + 259, + 481, + 307 + ], + "type": "text", + "content": "- ScanNet V2 [4] is a widely-used 3D indoor-scene dataset consisting of 1613 scans with 20 categories, where the point clouds are reconstructed from multi-view RGB-D images. The training set contains 1201 scans, the validation split contains 312 scans, and the testing set contains 100 scans." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 308, + 481, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 331 + ], + "type": "text", + "content": "As done in previous works [21, 26, 38], we use the mIoU (mean Intersection over Union) as the evaluation metric." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 350, + 279, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 350, + 279, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 279, + 363 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 372, + 481, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 372, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 372, + 481, + 456 + ], + "type": "text", + "content": "We follow the basic experimental settings of LaserMix [21] and GPC [17] for evaluating the proposed LASS3D on the outdoor-scene and indoor-scene datasets respectively. The classifiers and fusion adapter are both multi-layer perceptrons. We follow the settings of IGNet [38], which also adopts the MeanTeacher [37] framework, to set the update hyperparameter " + }, + { + "bbox": [ + 132, + 372, + 481, + 456 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 132, + 372, + 481, + 456 + ], + "type": "text", + "content": " as 0.999. The iteration number in progressive exploitation is set as 5. The hyperparameters " + }, + { + "bbox": [ + 132, + 372, + 481, + 456 + ], + "type": "inline_equation", + "content": "\\tau_r, \\tau_v, \\tau_n, \\lambda_u, \\lambda_{PE}, \\lambda_{KL}" + }, + { + "bbox": [ + 132, + 372, + 481, + 456 + ], + "type": "text", + "content": " are set as 0.9, 0.05, 0.2, 0.1, 0.1, and 0.1 respectively." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 138, + 532, + 479, + 662 + ], + "blocks": [ + { + "bbox": [ + 132, + 477, + 481, + 522 + ], + "lines": [ + { + "bbox": [ + 132, + 477, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 477, + 481, + 522 + ], + "type": "text", + "content": "Table 1: Comparative results on the SemanticKITTI [1] and ScanNet [4] datasets with varying labeled ratios. " + }, + { + "bbox": [ + 132, + 477, + 481, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 132, + 477, + 481, + 522 + ], + "type": "text", + "content": " denotes the point cloud data, " + }, + { + "bbox": [ + 132, + 477, + 481, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 132, + 477, + 481, + 522 + ], + "type": "text", + "content": " denotes the image data, and " + }, + { + "bbox": [ + 132, + 477, + 481, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 132, + 477, + 481, + 522 + ], + "type": "text", + "content": " denotes the text data. All mIoU scores are given in percentage (\\%). The best results are in bold and the second best results are marked with underlines." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 532, + 479, + 662 + ], + "lines": [ + { + "bbox": [ + 138, + 532, + 479, + 662 + ], + "spans": [ + { + "bbox": [ + 138, + 532, + 479, + 662 + ], + "type": "table", + "html": "
MethodModalitySemanticKITTI [1]ScanNet [4]
1%10%20%50%5%10%20%30%40%
MeanTeacher [37]P45.457.159.260.048.757.266.167.369.1
CBST [55]P48.858.359.459.748.557.865.467.369.2
CPS [3]P46.758.759.660.551.758.666.468.070.5
SSS-Net [5]P-----52.455.1--
LaserMix (Range View) [21]P43.458.859.461.4-----
LaserMix (Voxel) [21]P50.660.061.962.3-----
IGNet [38]P, I49.061.363.164.8-----
GPC [17]P54.162.062.562.854.860.566.768.971.3
LiM3D [26]P58.462.263.163.6-----
LASS3DP, I, T58.563.064.164.556.663.167.470.472.0
", + "image_path": "5f2c1f25f39dc1407cf31d7febd260a2f83f2d4c3b634e051bc4fbac2934da50.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 102 + ], + "type": "text", + "content": "J. Li and Q. Dong" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 282, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 282, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 282, + 128 + ], + "type": "text", + "content": "4.3 Comparative Evaluation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 148, + 482, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 148, + 482, + 256 + ], + "spans": [ + { + "bbox": [ + 130, + 148, + 482, + 256 + ], + "type": "text", + "content": "We evaluate the proposed LASS3D on both outdoor-scene (SemanticKITTI [1]) and indoor-scene (ScanNet [4]) datasets in comparison to some classic semi-supervised methods [3, 37, 55] extended from the 2D domain and the state-of-the-art (SOTA) methods [5, 17, 21, 26, 38] that are specially designed for semi-supervised 3D semantic segmentation. Note that IGNet [38], LaserMix [21], and LiM3D [26] utilize some specific characters of the LiDAR point clouds, and thus cannot be applied to the ScanNet dataset. In addition, SSS-Net [5] leverages the color information of the point clouds, and thus cannot be applied to the SemantiKITTI dataset where the color information is unavailable." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 482, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 378 + ], + "type": "text", + "content": "Following the settings in [17, 21, 38], we set the labeled ratio of the outdoor-scene dataset and indoor-scene dataset as " + }, + { + "bbox": [ + 130, + 258, + 482, + 378 + ], + "type": "inline_equation", + "content": "\\{1\\%, 10\\%, 20\\%, 40\\% \\}" + }, + { + "bbox": [ + 130, + 258, + 482, + 378 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 258, + 482, + 378 + ], + "type": "inline_equation", + "content": "\\{5\\%, 10\\%, 20\\%, 30\\%, 50\\% \\}" + }, + { + "bbox": [ + 130, + 258, + 482, + 378 + ], + "type": "text", + "content": " respectively for a fair comparison. The comparative results are reported in Tab. 1. As seen from this table, the proposed LASS3D outperforms all the comparative methods that use uni-modal data as input, which demonstrates the effectiveness of LASS3D. In addition, LASS3D outperforms IGNet which takes both point clouds and images as input in most cases, which indicates that the proposed language-assisted method can better boost the 3D segmentation performances and may provide some insights on the exploitation of multi-modal data in semi-supervised 3D semantic segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "content": "Moreover, the segmentation results of LASS3D and the second-best methods are visualized in Fig. 6. As seen from this figure, LASS3D outperforms its second-best counterparts, demonstrating the effectiveness of LASS3D from the qualitative perspective." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 143, + 459, + 471, + 602 + ], + "blocks": [ + { + "bbox": [ + 143, + 459, + 471, + 602 + ], + "lines": [ + { + "bbox": [ + 143, + 459, + 471, + 602 + ], + "spans": [ + { + "bbox": [ + 143, + 459, + 471, + 602 + ], + "type": "image", + "image_path": "aecd1f2ebeff5c4e1d5807621f131766f34d75e2733b7b363547f47ac9537ae4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 617, + 482, + 662 + ], + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 662 + ], + "type": "text", + "content": "Fig. 6: Qualitative results on the SemanticKITTI [1] (top) and ScanNet [4] (bottom) datasets by LASS3D and the second-best methods. All models are trained with " + }, + { + "bbox": [ + 130, + 617, + 482, + 662 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 617, + 482, + 662 + ], + "type": "text", + "content": " labeled data. We use the red boxes to highlight the areas where LASS3D evidently outperforms the comparative methods." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 410, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 410, + 91, + 447, + 100 + ], + "type": "text", + "content": "LASS3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 136, + 342, + 205 + ], + "blocks": [ + { + "bbox": [ + 162, + 114, + 312, + 125 + ], + "lines": [ + { + "bbox": [ + 162, + 114, + 312, + 125 + ], + "spans": [ + { + "bbox": [ + 162, + 114, + 312, + 125 + ], + "type": "text", + "content": "Table 2: Ablation of the loss terms." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 136, + 342, + 205 + ], + "lines": [ + { + "bbox": [ + 133, + 136, + 342, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 136, + 342, + 205 + ], + "type": "table", + "html": "
LCElLCEuLKLLPE1%10%20%50%
✓49.658.860.761.2
✓✓53.760.161.662.8
✓✓✓55.661.462.863.7
✓✓✓✓58.563.064.164.5
", + "image_path": "c5c7013429c761abd06aa0ff5fa8ba965478a3ecb8792aba52cf3d034514eb09.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 345, + 136, + 482, + 204 + ], + "blocks": [ + { + "bbox": [ + 348, + 114, + 474, + 125 + ], + "lines": [ + { + "bbox": [ + 348, + 114, + 474, + 125 + ], + "spans": [ + { + "bbox": [ + 348, + 114, + 474, + 125 + ], + "type": "text", + "content": "Table 3: Ablation of captions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 345, + 136, + 482, + 204 + ], + "lines": [ + { + "bbox": [ + 345, + 136, + 482, + 204 + ], + "spans": [ + { + "bbox": [ + 345, + 136, + 482, + 204 + ], + "type": "table", + "html": "
CiCe1%10%20%50%
50.059.861.562.1
✓53.660.561.963.0
✓54.161.062.763.8
✓✓58.563.064.164.5
", + "image_path": "ad8b546f4dc79c6558bdf26f70e64370029d7781f25a195533f5b8250c5dbd57.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 228, + 238, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 238, + 241 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 238, + 241 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 250, + 480, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 250, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 480, + 274 + ], + "type": "text", + "content": "We conduct ablation studies on SemanticKITTI [1] to investigate the effect of the key elements in the proposed LASS3D." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 274, + 480, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 480, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 480, + 382 + ], + "type": "text", + "content": "Effect of the loss terms. We investigate the effect of the loss terms and report the corresponding results in Tab. 2. As seen from the first two rows of Tab. 2, incorporating the unlabeled data into training can improve the performances. Then, as seen in the third row of Tab. 2, the " + }, + { + "bbox": [ + 130, + 274, + 480, + 382 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KL}^u" + }, + { + "bbox": [ + 130, + 274, + 480, + 382 + ], + "type": "text", + "content": " is introduced into training and the performances are improved, indicating that the semantic information distilled from the text-enhanced 3D features is helpful for segmentation. Finally, in the last row of Tab. 2, the progressive exploitation is adopted, and the performances are further boosted, demonstrating that the proposed progressive exploitation can make the unreliable points beneficial to the model training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 383, + 480, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 480, + 478 + ], + "type": "text", + "content": "Effect of different captions. We investigate the effect of different captions and report the corresponding results in Tab. 3. As seen from this table, both the image-level captions " + }, + { + "bbox": [ + 130, + 383, + 480, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 130, + 383, + 480, + 478 + ], + "type": "text", + "content": " and entity-level captions " + }, + { + "bbox": [ + 130, + 383, + 480, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_e" + }, + { + "bbox": [ + 130, + 383, + 480, + 478 + ], + "type": "text", + "content": " can improve the segmentation performances, indicating that the semantic information encoded in text embeddings can facilitate the point cloud segmentation. When combining " + }, + { + "bbox": [ + 130, + 383, + 480, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 130, + 383, + 480, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 383, + 480, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_e" + }, + { + "bbox": [ + 130, + 383, + 480, + 478 + ], + "type": "text", + "content": ", the best performance is achieved. Because the " + }, + { + "bbox": [ + 130, + 383, + 480, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 130, + 383, + 480, + 478 + ], + "type": "text", + "content": " provides more holistic semantic information and " + }, + { + "bbox": [ + 130, + 383, + 480, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_e" + }, + { + "bbox": [ + 130, + 383, + 480, + 478 + ], + "type": "text", + "content": " provides more fine-grained semantic information, which are complementary to each other. Thus, we utilize " + }, + { + "bbox": [ + 130, + 383, + 480, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 130, + 383, + 480, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 383, + 480, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_e" + }, + { + "bbox": [ + 130, + 383, + 480, + 478 + ], + "type": "text", + "content": " in LASS3D." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 479, + 480, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 480, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 480, + 562 + ], + "type": "text", + "content": "Effect of the fusion strategy for the image-level embeddings " + }, + { + "bbox": [ + 130, + 479, + 480, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_i" + }, + { + "bbox": [ + 130, + 479, + 480, + 562 + ], + "type": "text", + "content": ". As mentioned in Sec. 3.2, we adopt the cross-attention operation to fuse " + }, + { + "bbox": [ + 130, + 479, + 480, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_i" + }, + { + "bbox": [ + 130, + 479, + 480, + 562 + ], + "type": "text", + "content": " with the 3D features. To investigate the effect of the fusion strategy, we replace the cross-attention operation with add and concatenation operations. The corresponding results are reported in Tab. 4. As seen from this table, the model trained with the cross-attention operation achieves the best performance. Because the image-level captions usually contain holistic semantic information and" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 134, + 605, + 487, + 662 + ], + "blocks": [ + { + "bbox": [ + 143, + 584, + 296, + 595 + ], + "lines": [ + { + "bbox": [ + 143, + 584, + 296, + 595 + ], + "spans": [ + { + "bbox": [ + 143, + 584, + 296, + 595 + ], + "type": "text", + "content": "Table 4: Ablation of fusion strategy." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 584, + 477, + 595 + ], + "lines": [ + { + "bbox": [ + 307, + 584, + 477, + 595 + ], + "spans": [ + { + "bbox": [ + 307, + 584, + 477, + 595 + ], + "type": "text", + "content": "Table 5: Ablation of separation strategy." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 605, + 487, + 662 + ], + "lines": [ + { + "bbox": [ + 134, + 605, + 487, + 662 + ], + "spans": [ + { + "bbox": [ + 134, + 605, + 487, + 662 + ], + "type": "table", + "html": "
Strategy1%10%20%50%Strategy1%10%20%50%
Add54.061.262.363.5Confidence-based57.762.363.564.1
Concatenation56.262.063.463.9Variance-based57.062.162.863.5
Cross attention58.563.064.164.5Ours58.563.064.164.5
", + "image_path": "7f53c3566ec6f090dc50b272c967b7f088858f946bb299c1d6591590bb0b3487.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 102 + ], + "type": "text", + "content": "J. Li and Q. Dong" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "may have different correlations with different regions of the paired points. Simply adding or concatenating " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": " to all the paired 3D features leads to the lack of discrimination, while 3D semantic segmentation is a point-level prediction task that requires discriminative representations for each point. Cross attention can adaptively fuse " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": " with 3D features and produce more discriminative features, which makes it a more suitable fusion strategy for " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 189, + 485, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 485, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 485, + 344 + ], + "type": "text", + "content": "Effectiveness of the semantic-aware adaptive fusion module (SAFM). As mentioned in Sec. 2, existing language-assisted point cloud learning methods are either projection-based or alignment-based. Project-based methods are generally designed for object-level 3D classification and are not applicable to scene-level 3D semantic segmentation. To verify the effectiveness of the proposed SAFM, we replace it with an alignment-based method (e.g., the semantic consistency regularization in CLIP2Scene [2]). The corresponding results are reported in Tab. 6. As seen from this table, the model trained with SAFM outperforms the model trained with semantic consistency regularization. Probably because a large distribution discrepancy exists between the 3D features and text embeddings, simply aligning the 3D features to their corresponding text embeddings may result in the loss of some inherent geometric information. SAFM remedies this defect by adaptive fusion, and thus achieves better performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 345, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 482, + 489 + ], + "type": "text", + "content": "Effect of the separation strategy. As mentioned in Sec. 3.3, we adopt both confidences and variances to measure the reliability of the predictions and separate the unlabeled points into the reliable set and unreliable set. Here we evaluate the effectiveness of our separation strategy. The corresponding results are reported in Tab. 5. As seen from this table, the model trained with our strategy achieves the best performance. Probably because the confidence-based strategy alone is unable to filter out the highly confident wrong predictions, which may confuse the training. The variance-based strategy separates points according to the prediction consistency in different branches, which could be complementary to the confidence-based strategy in measuring reliability. Thus, we combine the above two strategies to measure reliability in a more comprehensive way, which can facilitate the model training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 490, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 561 + ], + "type": "text", + "content": "Effect of the iteration number. The iteration number of progressive exploitation determines the utilization degree of the unreliable data. Here, we evaluate LASS3D with the iteration number set as " + }, + { + "bbox": [ + 130, + 490, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\{0,1,2,3,5,7,9,11\\}" + }, + { + "bbox": [ + 130, + 490, + 482, + 561 + ], + "type": "text", + "content": ", and the corresponding results are shown in Fig. 7. As seen from this figure, the performances are generally improved with the increase of iteration number. Because a larger iteration number means that the exploitation of the unreliable data is more thor-" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 133, + 616, + 486, + 662 + ], + "blocks": [ + { + "bbox": [ + 132, + 583, + 481, + 607 + ], + "lines": [ + { + "bbox": [ + 132, + 583, + 481, + 607 + ], + "spans": [ + { + "bbox": [ + 132, + 583, + 481, + 607 + ], + "type": "text", + "content": "Table 6: Ablation of semantic-aware adap- Table 7: Ablation of the proposed pro- tive fusion module (SAFM). progressive exploitation (PE)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 616, + 486, + 662 + ], + "lines": [ + { + "bbox": [ + 133, + 616, + 486, + 662 + ], + "spans": [ + { + "bbox": [ + 133, + 616, + 486, + 662 + ], + "type": "table", + "html": "
Strategy1%10%20%50%Model1%10%20%50%
Alignment-based53.060.161.862.2GPC [17]54.162.062.562.8
SAFM58.563.064.164.5GPC with PE54.762.863.163.5
", + "image_path": "a7cc387d7b14fa082db9e69e60c320d14fd2c4ea991a992212d0e2bfba8a2673.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "LASS3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 233, + 116, + 375, + 200 + ], + "blocks": [ + { + "bbox": [ + 233, + 116, + 375, + 200 + ], + "lines": [ + { + "bbox": [ + 233, + 116, + 375, + 200 + ], + "spans": [ + { + "bbox": [ + 233, + 116, + 375, + 200 + ], + "type": "image", + "image_path": "7c22a04ed498dce9ad861404596b94fd1111fb421b03788c12325295e3c01586.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 213, + 479, + 235 + ], + "lines": [ + { + "bbox": [ + 132, + 213, + 479, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 479, + 235 + ], + "type": "text", + "content": "Fig. 7: Ablation of the iteration number. Note that the iteration here denotes the negative learning iteration in progressive exploitation, rather than the training iteration." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 257, + 480, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 328 + ], + "type": "text", + "content": "ough and thus brings better performance. However, when the iteration number reaches a certain level, the segmentation performance improves slightly. Probably because many unreliable points turn into reliable points during the progressive exploitation process, with most of their confident negative classes discarded. In addition, more iterations inevitably bring more training costs. Thus, we set the iteration number as 5 for a trade-off between the performance and training cost." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 329, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 329, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 132, + 329, + 480, + 411 + ], + "type": "text", + "content": "Effectiveness of the progressive exploitation. Considering that the proposed progressive exploitation strategy only requires the prediction confidence and variances, thus it can be seamlessly integrated into some existing methods. We integrate the progressive exploitation strategy into GPC [17] and report the comparative results in Tab. 7. As seen from this table, the proposed progressive exploitation strategy can further boost the performance of GPC, demonstrating the effectiveness of the proposed strategy for exploiting unreliable points." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 428, + 212, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 428, + 212, + 439 + ], + "spans": [ + { + "bbox": [ + 133, + 428, + 212, + 439 + ], + "type": "text", + "content": "4.5 Limitation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 445, + 479, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 445, + 479, + 505 + ], + "spans": [ + { + "bbox": [ + 132, + 445, + 479, + 505 + ], + "type": "text", + "content": "The proposed LASS3D has demonstrated its effectiveness in segmentation, however, it still has limitations. For example, LASS3D uses images as the bridge to connect the text data and point clouds, which is a complex operation and may bring accumulated errors in point-text matching. A more direct and accurate method to construct the point-text pairs is in demand." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 522, + 218, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 522, + 218, + 534 + ], + "spans": [ + { + "bbox": [ + 133, + 522, + 218, + 534 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 545, + 480, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 653 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 653 + ], + "type": "text", + "content": "In this paper, we propose a language-assisted semi-supervised 3D semantic segmentation method named LASS3D. In LASS3D, a semantic-aware adaptive fusion module is explored to fuse the 3D features with the embeddings of multi-level captions generated by language-vision models, which can inject the semantic information into the 3D features and thus facilitate 3D segmentation. In addition, a progressive exploitation strategy is explored for unlabeled unreliable data, which can further boost performance and can be seamlessly embedded into other methods. Experimental results on outdoor and indoor datasets show that LASS3D outperforms the comparative methods in most cases." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 102 + ], + "type": "text", + "content": "J. Li and Q. Dong" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (Grant Nos. 61991423, 62376269, 62073199); the Strategic Priority Research Program of the Chinese Academy of Sciences (Grant No. XDA27040811); the Open Project Program of Key Laboratory of Industrial Internet and Big Data, China National Light Industry, Beijing Technology and Business University." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 205, + 199, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 199, + 218 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 199, + 218 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 228, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "type": "text", + "content": "1. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semanticickitti: A dataset for semantic scene understanding of lidar sequences. In: ICCV. pp. 9297-9307 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 262, + 480, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 480, + 294 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 480, + 294 + ], + "type": "text", + "content": "2. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: Clip2scene: Towards label-efficient 3d scene understanding by clip. In: CVPR. pp. 7020-7030 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 295, + 480, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 295, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 138, + 295, + 480, + 317 + ], + "type": "text", + "content": "3. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: CVPR. pp. 2613-2622 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 317, + 480, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 317, + 480, + 349 + ], + "spans": [ + { + "bbox": [ + 138, + 317, + 480, + 349 + ], + "type": "text", + "content": "4. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR. pp. 5828-5839 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 350, + 480, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 480, + 372 + ], + "type": "text", + "content": "5. Deng, S., Dong, Q., Liu, B., Hu, Z.: Superpoint-guided semi-supervised semantic segmentation of 3d point clouds. In: ICRA. pp. 9214-9220 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 372, + 480, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 372, + 480, + 393 + ], + "spans": [ + { + "bbox": [ + 138, + 372, + 480, + 393 + ], + "type": "text", + "content": "6. Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Pla: Language-driven open-vocabulary 3d scene understanding. In: CVPR. pp. 7010-7019 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 393, + 480, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 393, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 138, + 393, + 480, + 426 + ], + "type": "text", + "content": "7. Fan, S., Dong, Q., Zhu, F., Lv, Y., Ye, P., Wang, F.: Scf-net: Learning spatial contextual features for large-scale point cloud segmentation. In: CVPR. pp. 14499-14508 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 426, + 480, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 480, + 447 + ], + "type": "text", + "content": "8. Graham, B., Engelcke, M., van der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. CVPR pp. 9224-9232 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 448, + 480, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 448, + 480, + 480 + ], + "spans": [ + { + "bbox": [ + 138, + 448, + 480, + 480 + ], + "type": "text", + "content": "9. Guo, Z., Zhang, R., Qiu, L., Ma, X., Miao, X., He, X., Cui, B.: Calip: Zero-shot enhancement of clip with parameter-free attention. In: AAAI. vol. 37, pp. 746-754 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 481, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 481, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 481, + 480, + 502 + ], + "type": "text", + "content": "0. He, W., Jamonnak, S., Gou, L., Ren, L.: Clip-s4: Language-guided self-supervised semantic segmentation. In: CVPR. pp. 11207-11216 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 502, + 480, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 524 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 524 + ], + "type": "text", + "content": "1. Hegde, D., Valanarasu, J.M.J., Patel, V.: Clip goes 3d: Leveraging prompt tuning for language grounded 3d recognition. In: ICCVW. pp. 2028-2038 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 524, + 480, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 524, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 524, + 480, + 545 + ], + "type": "text", + "content": "2. Hess, G., Tondersski, A., Petersson, C., Åström, K., Svensson, L.: Lidarclip or: How i learned to talk to point clouds. In: WACV. pp. 7438-7447 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 546, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 480, + 578 + ], + "type": "text", + "content": "3. Hu, Q., Yang, B., Fang, G., Guo, Y., Leonardis, A., Trigoni, N., Markham, A.: Sqn: Weakly-supervised semantic segmentation of large-scale 3d point clouds. In: ECCV. pp. 600-619. Springer (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 578, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 600 + ], + "type": "text", + "content": "4. Huang, R., Pan, X., Zheng, H., Jiang, H., Xie, Z., Wu, C., Song, S., Huang, G.: Joint representation learning for text and 3d point cloud. PR 147, 110086 (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 633 + ], + "type": "text", + "content": "5. Huang, T., Dong, B., Yang, Y., Huang, X., Lau, R.W., Ouyang, W., Zuo, W.: Clip2point: Transfer clip to point cloud classification with image-depth pretraining. In: ICCV. pp. 22157-22167 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "6. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML. pp. 4904-4916 (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "LASS3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "17. Jiang, L., Shi, S., Tian, Z., Lai, X., Liu, S., Fu, C.W., Jia, J.: Guided point contrastive learning for semi-supervised point cloud semantic segmentation. In: ICCV. pp. 6423-6432 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 171 + ], + "type": "text", + "content": "18. Jiao, S., Wei, Y., Wang, Y., Zhao, Y., Shi, H.: Learning mask-aware CLIP representations for zero-shot segmentation. In: NeurIPS (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "type": "text", + "content": "19. Kim, Y., Yim, J., Yun, J., Kim, J.: Nlnl: Negative learning for noisy labels. In: ICCV. pp. 101-110 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "text", + "content": "20. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV. pp. 4015-4026 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 225, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 246 + ], + "type": "text", + "content": "21. Kong, L., Ren, J., Pan, L., Liu, Z.: Lasermix for semi-supervised lidar semantic segmentation. In: CVPR. pp. 21705-21715 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "text", + "content": "22. Lai, X., Chen, Y., Lu, F., Liu, J., Jia, J.: Spherical transformer for lidar-based 3d recognition. In: CVPR. pp. 17545-17555 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "type": "text", + "content": "23. Lai, X., Liu, J., Jiang, L., Wang, L., Zhao, H., Liu, S., Qi, X., Jia, J.: Stratified transformer for 3d point cloud segmentation. In: CVPR. pp. 8500-8509 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "type": "text", + "content": "24. Li, J., Dong, Q.: Open-set semantic segmentation for point clouds via adversarial prototype framework. In: CVPR. pp. 9425-9434 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "type": "text", + "content": "25. Li, J., Dong, Q.: Density-guided semi-supervised 3d semantic segmentation with dual-space hardness sampling. In: CVPR. pp. 3260-3269 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 331, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 353 + ], + "type": "text", + "content": "26. Li, L., Shum, H.P., Breckon, T.P.: Less is more: Reducing task and model complexity for 3d point cloud semantic segmentation. In: CVPR. pp. 9361-9371 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 353, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 480, + 385 + ], + "type": "text", + "content": "27. Lin, Z., Geng, S., Zhang, R., Gao, P., de Melo, G., Wang, X., Dai, J., Qiao, Y., Li, H.: Frozen clip models are efficient video learners. In: ECCV. pp. 388-404. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 385, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 480, + 418 + ], + "type": "text", + "content": "28. Liu, L., Zhuang, Z., Huang, S., Xiao, X., Xiang, T., Chen, C., Wang, J., Tan, M.: Cpcm: Contextual point cloud modeling for weakly-supervised point cloud semantic segmentation. In: ICCV. pp. 18413-18422 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 418, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 438 + ], + "type": "text", + "content": "29. Liu, Z., Qi, X., Fu, C.W.: One thing one click: A self-training approach for weakly supervised 3d semantic segmentation. In: CVPR. pp. 1726-1736 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 438, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 438, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 438, + 480, + 471 + ], + "type": "text", + "content": "30. Lu, Y., Xu, C., Wei, X., Xie, X., Tomizuka, M., Keutzer, K., Zhang, S.: Open-vocabulary point-cloud object detection without 3d annotation. In: CVPR. pp. 1190-1199 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 471, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 480, + 504 + ], + "type": "text", + "content": "31. Ma, W., Li, S., Zhang, J., Liu, C.H., Kang, J., Wang, Y., Huang, G.: Borrowing knowledge from pre-trained language model: A new data-efficient visual learning paradigm. In: ICCV. pp. 18786-18797 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 504, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 535 + ], + "type": "text", + "content": "32. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 535, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 535, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 535, + 480, + 568 + ], + "type": "text", + "content": "33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "text", + "content": "34. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Denseclip: Language-guided dense prediction with context-aware prompting. In: CVPR. pp. 18082-18091 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "text", + "content": "35. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., Zeng, Z., Zhang, H., Li, F., Yang, J., Li, H., Jiang, Q., Zhang, L.: Grounded sam: Assembling open-world models for diverse visual tasks (2024)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "36. Tang, L., Chen, Z., Zhao, S., Wang, C., Tao, D.: All points matter: entropy-regularized distribution alignment for weakly-supervised 3d segmentation. NeurIPS 36 (2024)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 241, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 241, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 241, + 102 + ], + "type": "text", + "content": "J. Li and Q. Dong" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 643 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "37. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NIPS 30 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 171 + ], + "type": "text", + "content": "38. Unal, O., Dai, D., Hoyer, L., Can, Y.B., Van Gool, L.: 2d feature distillation for weakly-and semi-supervised 3d semantic segmentation. In: WACV. pp. 7336-7345 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 204 + ], + "type": "text", + "content": "39. Wang, Y., Huang, S., Gao, Y., Wang, Z., Wang, R., Sheng, K., Zhang, B., Liu, S.: Transferring clip's knowledge into zero-shot point cloud semantic segmentation. In: ACM MM. pp. 3745-3754 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "40. Wang, Y., Wang, H., Shen, Y., Fei, J., Li, W., Jin, G., Wu, L., Zhao, R., Le, X.: Semi-supervised semantic segmentation using unreliable pseudo-labels. In: CVPR. pp. 4248-4257 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "text", + "content": "41. Wu, X., Jiang, L., Wang, P.S., Liu, Z., Liu, X., Qiao, Y., Ouyang, W., He, T., Zhao, H.: Point transformer v3: Simpler, faster, stronger. In: CVPR (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 282 + ], + "type": "text", + "content": "42. Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point transformer v2: Grouped vector attention and partition-based pooling. In: NeurIPS. vol. 35, pp. 33330-33342 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "type": "text", + "content": "43. Xiang, P., Wen, X., Liu, Y.S., Zhang, H., Fang, Y., Han, Z.: Retro-fpn: Retrospective feature pyramid network for point cloud semantic segmentation. In: ICCV. pp. 17826-17838 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "text", + "content": "44. Xue, L., Gao, M., Xing, C., Martin-Martín, R., Wu, J., Xiong, C., Xu, R., Niebles, J.C., Savarese, S.: Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In: CVPR. pp. 1179-1189 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "45. Yao, L., Han, J., Wen, Y., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, C., Xu, H.: Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. NeurIPS 35, 9125-9138 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "text", + "content": "46. Yu, W., Liu, Y., Hua, W., Jiang, D., Ren, B., Bai, X.: Turning a clip model into a scene text detector. In: CVPR. pp. 6978-6988 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "47. Zeng, Y., Jiang, C., Mao, J., Han, J., Ye, C., Huang, Q., Yeung, D.Y., Yang, Z., Liang, X., Xu, H.: Clip2: Contrastive language-image-point pretraining from real-world point cloud data. In: CVPR. pp. 15244-15253 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "text", + "content": "48. Zhang, J., Fan, G., Wang, G., Su, Z., Ma, K., Yi, L.: Language-assisted 3d feature learning for semantic scene understanding. In: AAAI. pp. 3445-3453 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "text", + "content": "49. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: CVPR. pp. 8552-8562 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "50. Zhang, Y., Huo, X., Chen, T., Wu, S., Wong, H.S.: Exploring intra-class variation factors with learnable cluster prompts for semi-supervised image synthesis. In: CVPR. pp. 7392-7401 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 511, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 481, + 533 + ], + "type": "text", + "content": "51. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV. pp. 16259-16268 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 533, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 481, + 555 + ], + "type": "text", + "content": "52. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: ECCV. pp. 696-712. Springer (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "text", + "content": "53. Zhu, X., Zhang, R., He, B., Guo, Z., Zeng, Z., Qin, Z., Zhang, S., Gao, P.: Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In: ICCV. pp. 2639-2650 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 588, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 481, + 621 + ], + "type": "text", + "content": "54. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In: CVPR. pp. 9939-9948 (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 621, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 643 + ], + "type": "text", + "content": "55. Zou, Y., Yu, Z., Kumar, B., Wang, J.: Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In: ECCV. pp. 289-305 (2018)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "LASS3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LATTE3D_ Large-scale Amortized Text-To-Enhanced3D Synthesis/a662796f-9e3f-4f91-946c-d99cfdaedd9c_content_list.json b/2024/LATTE3D_ Large-scale Amortized Text-To-Enhanced3D Synthesis/a662796f-9e3f-4f91-946c-d99cfdaedd9c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dd26e915468fa2578697d6c98d80e01154f3be35 --- /dev/null +++ b/2024/LATTE3D_ Large-scale Amortized Text-To-Enhanced3D Synthesis/a662796f-9e3f-4f91-946c-d99cfdaedd9c_content_list.json @@ -0,0 +1,1836 @@ +[ + { + "type": "text", + "text": "LATTE3D: Large-scale Amortized Text-To-Enhanced3D Synthesis", + "text_level": 1, + "bbox": [ + 305, + 141, + 697, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kevin Xie $^{1,2}$ , Jonathan Lorraine $^{1,2}$ , Tianshi Cao $^{1,2}$ , Jun Gao $^{1,2}$ , James Lucas $^{1}$ , Antonio Torralba $^{1,3}$ , Sanja Fidler $^{1,2}$ , and Xiaohui Zeng $^{1,2}$", + "bbox": [ + 217, + 210, + 784, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Nvidia, $^{2}$ University of Toronto and Vector Institute, $^{3}$ MIT", + "bbox": [ + 300, + 253, + 700, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/044750169467ef79cf0e8c6281f0f18c5e9476fe02b78c8af1c71a3cf4af1dde.jpg", + "image_caption": [ + "Fig.1: Samples generated in $\\sim 400\\mathrm{ms}$ on a single A6000 GPU from text prompts. Objects without prompt labels are generated by our text-to-3D model trained on $\\sim 100\\mathrm{k}$ prompts, while labeled objects are generated by our 3D stylization model trained on $12\\mathrm{k}$ prompts. See our project website** for more." + ], + "image_footnote": [], + "bbox": [ + 246, + 300, + 759, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent text-to-3D generation approaches produce impressive 3D results but require time-consuming optimization that can take up to an hour per prompt. Amortized methods like ATT3D optimize multiple prompts simultaneously to improve efficiency, enabling fast text-to-3D synthesis. However, they cannot capture high-frequency geometry and texture details and struggle to scale to large prompt sets, so they generalize poorly. We introduce LATTE3D, addressing these limitations to achieve fast, high-quality generation on a significantly larger prompt set. Key to our method is 1) building a scalable architecture and 2) leveraging 3D data during optimization through 3D-aware diffusion priors, shape regularization, and model initialization to achieve robustness to diverse and complex training prompts. LATTE3D amortizes both neural field and textured surface generation to produce highly detailed textured meshes in a single forward pass. LATTE3D generates 3D objects in 400ms, and can be further enhanced with fast test-time optimization.", + "bbox": [ + 259, + 590, + 740, + 800 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. **https://research.nvidia.com/labs/toronto-ai/LATTE3D/", + "bbox": [ + 230, + 824, + 761, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 374, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advances in text-to-3D synthesis via image diffusion models mark significant progress in 3D content creation using [3,8,17,25,33,46,47]. However, these methods often involve an expensive and time-consuming optimization that can take up to an hour to generate a single 3D object [17, 33]. We aim to achieve real-time text-to-3D synthesis for diverse text prompts and support fast test-time optimization when a further quality boost is desired.", + "bbox": [ + 212, + 181, + 785, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Seminal work [33] proposed to optimize a neural field using the score distillation sampling (SDS) loss to generate 3D assets. It unlocked open-vocabulary 3D content creation but was (a) lacking high-frequency geometric and texture details, (b) expensive, and (c) prone to failing to generate the desired object. Follow-ups introduced a surface-based fine-tuning stage [3, 17, 47] to generate high-frequency details. While this two-stage pipeline defines most current state-of-the-art text-to-3D approaches, it remains expensive and prone to prompt failure.", + "bbox": [ + 212, + 272, + 785, + 391 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ATT3D [21] was the first to make text-to-3D generation fast by simultaneously training a single model on a set of prompts in a process called amortized optimization. Amortization reduces training time by optimizing a shared text-conditioned model on a set of prompts, enabling the model to generalize to new prompts at inference time. Despite promising results on curated prompts, ATT3D – and the underlying per-prompt methods – remain prone to failing on general prompts. Also, ATT3D only amortizes the first stage of 3D synthesis, producing a neural field representation and thus cannot generate high-frequency details.", + "bbox": [ + 212, + 393, + 787, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we introduce LATTE3D, a Large-scale Amortized Text-to-Enhanced3D synthesis method that can produce high-quality 3D content in real-time. Our work makes the following technical contributions to boost the quality, robustness, scale, and speed of text-to-3D generative models:", + "bbox": [ + 214, + 531, + 785, + 590 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel text-to-3D amortization architecture that can scale to orders of magnitude larger prompt sets.", + "- We leverage 3D data in training to improve quality and robustness, through 3D-aware diffusion priors, regularization loss, and weight initialization through pretraining with 3D reconstruction.", + "- We amortize the surface-based refinement stage, greatly boosting quality." + ], + "bbox": [ + 225, + 604, + 792, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 724, + 387, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early work on 3D object synthesis was typically limited to generating objects from a given class [1,2,7,10,22,24,28,44,48,49,51,53], e.g. cars or chairs. Recent extensions use captioned 3D shapes to train text-to-3D models [13,29] generating shapes from diverse categories, but requiring 3D supervision, restricting them to synthetic datasets of limited size, diversity, and visual fidelity.", + "bbox": [ + 212, + 761, + 785, + 838 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "K. Xie et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The advent of differentiable rendering, both volumetric [26] and surface-based [27], opened the door to inverse image rendering [7, 33], unlocking the use of powerful text-to-image generative models [36-38] in the 3D synthesis process. DreamFusion [33] proposed the SDS loss to optimize a neural field using a text-to-image diffusion model to generate 3D assets. Follow-up work introduced a surface-based refinement stage [17], allowing the synthesis of high-frequency details. This two-stage optimization pipeline now defines most of the state-of-the-art text-to-3D approaches [3,25,46,47]. The recently released large-scale 3D dataset Objaverse-XL [5,6] has spurred researchers to explore the benefits of 3D supervision in the synthesis process. To date, the most successful approaches make", + "bbox": [ + 212, + 146, + 542, + 417 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/18147faf535aeb209090b21e179d8fe9a152c5d764a147c20180725785dbf416.jpg", + "image_caption": [ + "Fig.2: Comparison of SOTA text-to-3D methods on unseen prompts. We plot different methods' user study preference rates compared to LATTE3D. For MVDream, we report varying optimization times." + ], + "image_footnote": [], + "bbox": [ + 555, + 150, + 784, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "text-conditioned image diffusion models 3D aware [15,19,41,42] by fine-tuning 2D diffusion models using rendered multiview images. SDS from multiview diffusion models is then used to obtain 3D shapes [19,41,42].", + "bbox": [ + 212, + 417, + 784, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, these approaches require a lengthy, per-prompt optimization process, hindering applicability to real-world content creation. To address this, two lines of work have been pursued. The first type of method uses a text-to-image-to-3D approach that generates images with a text-to-image diffusion model and trains an image-to-3D lifting network [15, 18, 41, 43]. They offer improved speed but with limited quality details compared to optimization methods, but still take $5 - 20$ seconds per prompt [15, 43] for sampling from the image diffusion model.", + "bbox": [ + 212, + 464, + 785, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In an alternate line of work, ATT3D [21] proposed an amortized framework that optimizes multiple prompts simultaneously with a unified model. Follow-up works expanded the network architecture to achieve better quality [16, 34]. LATTE3D extends this line of work to larger scales by incorporating 3D knowledge to achieve prompt robustness during training, among other design improvements like better architecture. Amortized text-to-3D offers an attractive quality vs. speed trade-off as it uses a single feed-forward architecture without requiring sampling of diffusion models - see Fig. 2.", + "bbox": [ + 212, + 569, + 787, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concurrent Works Concurrently, AToM [34] also amortizes two-stage training with a triplane representation, but they do not scale beyond the smaller sets of prompts from ATT3D. In contrast, we scale to the order of 100000 prompts, using 3D-aware techniques. Another concurrent work of ET3D [4] achieves fast generation by training a GAN model, but it is only trained on a small-scale compositional dataset of 5000 prompts and does not guarantee view consistency as it uses a 2D neural upsampler. Lastly, LGM [43] is a text-to-image-to-3D method that generates Gaussian splats in 5 seconds. We experimentally demonstrate competitive performance to it.", + "bbox": [ + 212, + 703, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "LATTE3D", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 143, + 380, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An established pipeline for high-quality text-to-3D generation consists of two stages, each performing per-prompt optimization using the SDS loss with a text-to-image diffusion model [3, 17, 47]. Stage-1 optimizes a volumetric representation, which is typically a neural radiance field. Meshes afford real-time rendering but are hard to optimize from scratch with image supervision alone. Stage-2 uses the output of the neural field to initialize a signed distance field (SDF) and a texture field, from which a mesh can be derived using differentiable isosurfacing [39, 40]. This surface-based representation is then optimized via differentiable rasterization [27], which is fast even at 1024 rendered image resolution – enabling supervision from higher-resolution text-to-image diffusion models.", + "bbox": [ + 212, + 174, + 787, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ATT3D [21] amortizes optimization of a neural field over a set of prompts with a hypernetwork mapping a text prompt to a neural field, which is trained with the SDS loss over a prompt set, referred to as seen prompts. LATTE3D introduces a new architecture that amortizes both stages of the generation process, aiming to produce high-quality textured meshes in real-time. We scale the approach to a magnitudes larger and more diverse set of prompts by leveraging 3D knowledge in the amortization loop.", + "bbox": [ + 212, + 325, + 787, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach is illustrated in Fig. 3. We initialize LATTE3D with a reconstruction pretraining step to stabilize training (Sec. 3.1). The architecture consists of two networks, one for geometry and the other for texture (Sec. 3.2). Amortized model training with diffusion priors is done through a two-stage pipeline consisting of a volume-based stage-1 and a surface-based stage-2 (Sec. 3.3). To reconcile reconstruction pretraining, which uses 3D geometry as input, with the task of text-to-3D generation, we anneal the network's input during stage-1 to gradually fully replace the dependence on 3D input with a single dummy input. During inference (Sec. 3.4), our model generates a 3D textured mesh from just a text prompt in 400ms and allows an optional lightweight test-time refinement to enhance the quality of geometry and textures (Sec. 3.5).", + "bbox": [ + 212, + 431, + 787, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Pretraining to reconstruct shapes.", + "text_level": 1, + "bbox": [ + 214, + 617, + 542, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although SDS can synthesize text-aligned 3D features, empirically, it is a less efficient source of supervision due to high variance. We find that pretraining the model to first encode-decode 3D shapes subsequently accelerates the early stage of amortized SDS training. We initialize our model $\\mathcal{M}$ by pretraining it on a 3D asset dataset using image reconstruction losses, similar to the reconstruction stage used in [9]. $\\mathcal{M}$ inputs a sampled point cloud and outputs a predicted 3D shape. We then render multiple views of the predicted 3D shape $o$ to compare with the input shape $s$ renderings. We use an $\\ell_2$ loss on the rendered opacity and RGB image:", + "bbox": [ + 212, + 638, + 787, + 776 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c o n}} (o, s, c) = \\left| \\left| \\mathbf {R} _ {\\text {o p a c i t y}} (o, c) - \\mathbf {R} _ {\\text {o p a c i t y}} (s, c) \\right| \\right| _ {2} + \\left| \\left| \\mathbf {R} _ {\\text {R G B}} (o, c) - \\mathbf {R} _ {\\text {R G B}} (s, c) \\right| \\right| _ {2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 785, + 785, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We denote the renderer $\\mathbf{R}$ , the opacity and RGB rendering with $\\mathbf{R}_{\\mathrm{opacity}}$ and $\\mathbf{R}_{\\mathrm{RGB}}$ , respectively, and the camera with $c$ , randomly sampled on a sphere.", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "K. Xie et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/751c8154c5442a8605be520b5c8a401f36b0b3c2f735b70b500fe1c0bdb22ee7.jpg", + "image_caption": [ + "Fig. 3: LATTE3D consists of two networks: a texture network $T$ and geometry network $G$ . When amortizing the first stage, the encoders of both networks share the same set of weights. The training objective includes an SDS gradient from a 3D-aware image prior and a regularization loss that compares the rendered predicted shape's mask with the rendered masks of 3D assets in a library. When amortizing surface-based refinement in stage-2, we freeze the geometry network $G$ and update the texture network $T$ ." + ], + "image_footnote": [], + "bbox": [ + 218, + 146, + 782, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Model Architecture", + "text_level": 1, + "bbox": [ + 215, + 407, + 426, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our model $\\mathcal{M}$ consists of two networks, $G$ and $T$ , for predicting geometry and texture, respectively. After pretraining, we introduce text-conditioning by adding cross-attention layers. We use CLIP [35] embeddings for encoding the text prompt $p$ . In stage-1 training, we tie the geometry and texture networks' weights (except for separate MLP decoder heads), effectively training a single input encoder. In stage-2, we freeze the geometry network $G$ and refine the texture network $T$ with an additional trainable upsampling layer. The final model $\\mathcal{M}$ used in inference is the frozen $G$ from stage-1 and a refined $T$ from stage-2.", + "bbox": [ + 212, + 431, + 787, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The geometry network $G$ consists of a U-Net encoder on a triplane [2,7], followed by a decoder containing another triplane U-Net and a neural volumetric density field. Specifically, a point cloud is fed through a PointNet, then converted to a triplane via average scattering. During training, the point cloud comes from the 3D data, and in inference, a dummy sphere point cloud is used. While this point cloud encoder is inherited from reconstruction pretraining, it becomes unimportant in the final model. The triplane is processed through the encoder and decoder U-Nets. The text embedding is fed into every residual U-Net block in the encoder via cross-attention. For any point in 3D space, we extract triplane features from each plane via bilinear interpolation. An MLP then predicts the density value at that point.", + "bbox": [ + 212, + 568, + 787, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The texture network $T$ and geometry network $G$ share the same encoder and decoder U-Net architecture in pretraining and stage-1. We employ another non-shared MLP neural texture field to predict RGB values at query points. In stage-2, we upsample the triplane features produced by the encoder to a higher resolution to gain extra capacity to generate high-frequency texture details. Specifically, we first bilinearly upsample the latent triplane from 128 resolution to", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "LATTE3D", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8c8399b526d550f0d7c1a5bb92d1daa5b789d24b9b914ca97293aad6992bd54c.jpg", + "image_caption": [ + "Fig. 4: Stylization application: Our model learns to generate diverse stylizations of the same shapes. Left column shows the original shape to be stylized." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 774, + 304 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "480 and add a residual MLP which maps the text-embedding to the residual of the triplane feature, mimicking ATT3D's mapping (hyper)network. The decoder then takes this upsampled triplane as input. Details are in the Appendix.", + "bbox": [ + 214, + 378, + 787, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Rendering. To train the model, we render the generated 3D object into 2D images with different methods in stage-1 and 2. In stage-1, the output 3D geometry is a triplane-based neural field, on which we use volume rendering to get images of 256 resolution using the formulation from VolSDF [50], allowing us to parameterize the density via an approximate signed distance field whose zero-level set defines the object geometry. This choice allows a simple conversion to surface-based rendering. For the geometry network, we observe more stable training with volume rendering with SDS compared to surface rendering. In stage-2, we extract the isosurface from the density field with Marching Cubes [20] and render images via rasterization [14]. Meshes can be rendered at a 1024 image resolution in real-time, helping capture finer details for learning textures.", + "bbox": [ + 214, + 441, + 787, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Amortized Learning", + "text_level": 1, + "bbox": [ + 215, + 631, + 429, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Amortized learning distills knowledge from image diffusion priors into our model. We perform amortized training on volumetric geometry training (stage-1) and surface texture training (stage-2) sequentially. We incorporate 3D information into our training process (a) implicitly via 3D-aware 2D SDS losses and (b) explicitly with regularization.", + "bbox": [ + 214, + 657, + 787, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Curating prompt sets. We create large prompt sets for training using rule-based text generation or ChatGPT [31]. For the former, we start with the categories names from the captions of Objaverse [23], and design rules, like \"object A in style B is doing C\" with a set of styles and activities. Alternatively, we input the captions into ChatGPT and ask it for detailed and diverse prompts describing a similar 3D object. Fig. 3 shows that when generating \"Amigurumi fiddler crab", + "bbox": [ + 214, + 750, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "K. Xie et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "wearing a top hat\", we guide shape synthesis towards resembling a crab by retrieving a related 3D shape used in a regularization loss. Other details of dataset construction are in Sec. 4.1.", + "bbox": [ + 215, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Amortizing Stage-1 Generation In this stage, we train $\\mathcal{M}$ with our datasets. We use the SDS loss with a 3D-aware 2D prior and regularization to the paired 3D shape, which we now explain in detail:", + "bbox": [ + 215, + 210, + 782, + 255 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D-aware SDS loss. During stage-1 training we leverage a 3D aware diffusion prior, which provides a stronger, multiview consistent supervisory signal. Specifically, we use the model from MVDream [42], which was trained on four rendered views of objects from the Objaverse dataset [6] by modifying Stable Diffusion [37] to generate multiview consistent images.", + "bbox": [ + 215, + 267, + 782, + 342 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Regularization loss. While 3D-aware SDS loss with MVDream offers a strong multiview prior, it can still fail to correctly capture the full 3D geometry for some prompts since MVDream only supervises four views and can create enlarged geometry or floaters in the shape. Some failures are shown on the right of Fig. 10.", + "bbox": [ + 215, + 354, + 782, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Combining regularization using 3D shapes with the SDS loss helps training and improves geometry. As shown in Fig. 3, for each training prompt, we regularize the output shape $o$ by comparing its rendered mask to the mask of a shape $s$ retrieved from the 3D dataset with the prompt:", + "bbox": [ + 215, + 415, + 782, + 474 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e g}} (o, s, c) = \\left| \\left| \\mathbf {R} _ {\\text {o p a c i t y}} (o, c) - \\mathbf {R} _ {\\text {o p a c i t y}} (s, c) \\right| \\right| _ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 484, + 782, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{R}_{\\mathrm{opacity}}$ is the opacity mask (volume-rendered density). We balance the regularization $\\mathcal{L}_{\\mathrm{reg}}$ and the SDS loss $\\mathcal{L}_{\\mathrm{SDS}}$ by simply blending them using a convex combination with weight $\\alpha$ . Our final training loss is:", + "bbox": [ + 215, + 506, + 782, + 551 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t r a i n}} = (1 - \\alpha) \\mathcal {L} _ {\\text {S D S}} + \\alpha \\mathcal {L} _ {\\text {r e g}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 561, + 782, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During training, we render five views for each generated shape. We average across four views to compute $\\mathcal{L}_{\\mathrm{SDS}}$ , and one is used to compute $\\mathcal{L}_{\\mathrm{reg}}$ .", + "bbox": [ + 215, + 583, + 782, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Input point cloud annealing To bridge the gap between the real point cloud inputs in training and dummy point cloud in inference, we gradually anneal the input point clouds towards dummy point clouds during stage-1 training. Specifically, we randomly replace a subset of points in the input point cloud with points from the dummy point cloud using a probability that increases linearly from 0 to 1 in the last 5000 training iterations. With point cloud annealing, the model gets better results in inference when the dummy point cloud is used.", + "bbox": [ + 215, + 625, + 782, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Amortizing Stage-2 Generation During stage-2 training, we freeze the network for our geometry and only tune the texture since training both leads to instabilities. We use depth-conditional ControlNet [52] for SDS guidance, allowing a higher, 512 resolution supervision. Depth conditioning encourages the SDS loss to guide the texture to align with the geometry from the geometry network, thereby improving the 3D texture consistency.", + "bbox": [ + 215, + 750, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "LATTE3D", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Inference", + "text_level": 1, + "bbox": [ + 215, + 146, + 339, + 159 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During inference, our model inputs the user's text prompt and a dummy point cloud. Our $\\mathcal{M} = (G,T)$ outputs a textured mesh, where $T$ is the final texture network refined in stage-2, while $G$ is trained in stage-1 only and then frozen. Predicting one shape and rendering the image only takes us 400ms on an A6000 GPU in inference, allowing the generation of user results at interactive speeds. A single A6000 can generate 4 samples simultaneously. Once the user selects a desired sample, textured meshes can be exported from LATTE3D. We use a UV mapping tool to UV parameterize the mesh output by $G$ and use $T$ to predict the albedo colors of each texel in a UV-mapped texture image. The resulting mesh and texture image is compatible with standard rendering software.", + "bbox": [ + 212, + 167, + 787, + 320 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Test Time Optimization", + "text_level": 1, + "bbox": [ + 215, + 339, + 460, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method supports test time optimization if the user wants to boost the quality of a particular prompt. This can be useful if the user's prompt deviates significantly from the seen prompts. As in our stage-2 training, we freeze the geometry and finetune the texture network $T$ with depth-conditional SDS guidance. Our test time optimization is significantly faster to converge than MVDream [42] or other optimization-based text-to-3D methods [17, 33, 47] since we optimize our amortized network instead of a randomly initialized network.", + "bbox": [ + 212, + 361, + 787, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.6 3D Stylization", + "text_level": 1, + "bbox": [ + 215, + 486, + 380, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LATTE3D can also be trained and deployed as a 3D stylization method, allowing users to cheaply create variations from existing 3D assets. During training, we skip the input point cloud annealing step to maintain point cloud reliance and amortize over a dataset where each 3D shape corresponds to many style prompts. Our blend of regularization and SDS guidance drives the model to produce shapes structurally similar to the original shapes yet semantically aligned with the style prompt. Amortization significantly reduces the computational cost per shape-prompt pair compared to optimizing each pair separately. During inference, point clouds from the training set can be combined with novel prompts as input to the model to produce variations of the original shape.", + "bbox": [ + 212, + 508, + 787, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 680, + 375, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We first describe our experimental setup, including datasets, metrics, and baselines for comparison (Sec. 4.1). We then present our quantitative and qualitative results (Sec. 4.2), demonstrating the performance and generalization ability of LATTE3D (Sec. 4.3), followed by design choices ablations (Sec. 4.4).", + "bbox": [ + 212, + 708, + 787, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 787, + 428, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We now discuss dataset construction, model pretraining, and evaluation metrics.", + "bbox": [ + 215, + 810, + 785, + 825 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "K. Xie et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/83912decb4981a65caef63b0e3bb9d6c01c4b7d6c55e50d0f2bff8f0e866a083.jpg", + "image_caption": [ + "MVDream (6min)" + ], + "image_footnote": [], + "bbox": [ + 246, + 146, + 326, + 239 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/79f3f08df33e0a26d89e02401a439b48e10a4c9ff0bc2b96eb1d77250f7c6504.jpg", + "image_caption": [ + "MVDream (36min)" + ], + "image_footnote": [], + "bbox": [ + 331, + 146, + 401, + 239 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b6733d82b10c55e37379f886c89930e565222c7cae08de120e8526350394ccd1.jpg", + "image_caption": [ + "3DTopia (20s)", + "Fig. 5: Qualitative comparison between LATTE3D and ATT3D [21], MVDream [42], Instant3D/3DTopia [15] and LGM [43]. Both rows are unseen prompts: \"a red convertible car with the top down\", \"an ice cream sundae\"." + ], + "image_footnote": [], + "bbox": [ + 413, + 143, + 465, + 239 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/cb1fd2c0efb16cfc2c02bdde6ea2ec5ba30c028ff17e788384425338c6c59069.jpg", + "image_caption": [ + "LGM (5s)" + ], + "image_footnote": [], + "bbox": [ + 475, + 146, + 544, + 239 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9c837622280e3a60d09351465007e50fa6e19f1622a89872d69f8ef07dbe7be8.jpg", + "image_caption": [ + "ATT3D" + ], + "image_footnote": [], + "bbox": [ + 550, + 146, + 619, + 239 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a0399853b212ddab43af57880eb9a0cca5743bb612120d65bbd8f135de172f06.jpg", + "image_caption": [ + "ATT3D+ UNet $^+$ MV" + ], + "image_footnote": [], + "bbox": [ + 625, + 145, + 687, + 237 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d4aa217c847d3243099139360b25baa76411ceb3b5398c4c96dd67d6e86c7412.jpg", + "image_caption": [ + "LATTE3D (400ms)" + ], + "image_footnote": [], + "bbox": [ + 694, + 146, + 756, + 236 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training Set To investigate scaling amortized optimization beyond what has been done in prior works - see Tab. 1 - we construct a new dataset gpt-101k that consists of 101k text prompts and 34k shapes. For the shapes, we use the \"lvis\" subset from Objaverse [6] with 50k human-verified shapes. We filter out poor shapes, e.g. flat images and scans, and retain 34k", + "bbox": [ + 212, + 342, + 486, + 492 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "shapes for our dataset. We use ChatGPT to augment each object's caption to gather 3 prompts per object.", + "bbox": [ + 212, + 493, + 784, + 523 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/4940f1be80cec561eac7b53d1596fa75e89e739d143243fb28c41a1a146da7ca.jpg", + "table_caption": [ + "Table 1: Comparing the sizes of amortization training sets to prior works." + ], + "table_footnote": [], + "table_body": "
name#promptsSource
Animal2400 [21,34]2400Rule
DF415 [34]415DF [33]
animal-style (Ours)12000Rule
gpt-101k (Ours)101608GPT
", + "bbox": [ + 503, + 387, + 785, + 465 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Sets We use two prompt sets for evaluation. To evaluate the benefit of amortized optimization in our method, we evaluate our model on gpt-101k. However, due to the computational cost of running the baselines, we randomly select a 50 prompt subset in gpt-101k to form the seen prompt set for benchmarking. Furthermore, to gauge the generalization abilities of our amortized model, we construct an unseen prompt set by filtering a subset of 67 DreamFusion [33] prompts that are close in distribution to lvis categories.", + "bbox": [ + 212, + 536, + 787, + 642 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines. We compare LATTE3D with the baselines: ATT3D [21], MVDream [42], Instant3D [15] and LGM [43]. We re-implement ATT3D [21] with a hypernetwork. We use the open-source threestudio [8] implementation of MVDream. For MVDream, we train each 3D model for 10000 iterations, with a batch size of 8 on a single A100 GPU with provided default settings. For Instant3D, since the source code has not been released, we use the re-implementation from 3DTopia1. For LGM, we use the official open-source implementation.", + "bbox": [ + 212, + 652, + 790, + 760 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation metrics. We quantitatively compare our model with baselines by evaluating the generated 3D content fidelity and its consistency to the text prompt for the optimization cost. We use three metrics to evaluate fidelity.", + "bbox": [ + 214, + 771, + 785, + 818 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "LATTE3D", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "1 https://github.com/3DTopia/3DTopia", + "bbox": [ + 217, + 824, + 500, + 840 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/35a64fe7d924cc68230f13b68c74e0b24d7a99363960a44c6e0424f47fd592ae.jpg", + "table_caption": [ + "Table 2: Quantitative metrics and average user preference (\\%) of baselines over LATTE3D trained on gpt-101k using seen and unseen prompts. We also report test-time optimization, which takes $10\\mathrm{min}$ ." + ], + "table_footnote": [], + "table_body": "
ModelTimeRender-FID ↓ seenCLIP-Score ↑ unseen(df)Preference % ↑seenUnseen(df)seenUnseen(df)
LATTE3D400ms180.58190.000.26080.2605--
LATTE3D-opt10min171.69178.370.28000.3000--
MVDream6min203.53199.230.22400.26982.512.4
MVDream18min191.77188.840.24400.295026.552.2
MVDream36min158.11143.440.28300.318048.589.3
3DTopia20s217.29222.760.18370.217120.825.6
LGM5s178.67166.400.28400.306326.236.4
", + "bbox": [ + 254, + 198, + 743, + 354 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Render-FID: We compute the FID [11] between the renderings of generated 3D content and a set of images sampled from Stable Diffusion [37] with the same text prompts as input. This metric measures how well the generated shapes align with those from the 2D prior in visual quality.", + "bbox": [ + 212, + 381, + 782, + 439 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "CLIP Score: We compute the average CLIP scores between the text prompt and each rendered image to indicate how well the generated 3D shape aligns with the input text prompt.", + "bbox": [ + 212, + 441, + 782, + 484 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "User Study: We evaluate the overall 3D geometry and texture through Amazon Mechanical Turk user studies. We present videos of 3D shapes generated by our method and a baseline head-to-head, with corresponding text prompts, and ask users for their preference. Each comparison is averaged over three users to obtain preference rate for each baseline relative to LATTE3D.", + "bbox": [ + 212, + 487, + 784, + 560 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Timing: We measure our optimization cost by GPU time per prompt to gauge the total compute cost on each prompt. We divide the total wall-clock time for pretraining, stage-1, and stage-2 by the prompt set's size, then multiply by the number of GPUs used. We measure inference time for Instant3D, LGM, and LATTE3D by the time the model takes from in-taking the text input to outputting the final features before rendering. For MVDream [42], we measure the average time to optimize a single prompt on an A100 GPU.", + "bbox": [ + 212, + 563, + 785, + 667 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 691, + 439, + 707 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We observe robust generalization of LATTE3D by training on the gpt-101k dataset to unseen prompts (df, from DreamFusion [33]), as demonstrated by qualitative (Fig. 5) and quantitative (Fig. 2, Tab. 2) results. We highlight the relative user preference of LATTE3D versus different baselines over inference time in Fig. 2. LATTE3D produce results of competitive user preference to SOTA baselines at a uniquely fast inference time.", + "bbox": [ + 212, + 718, + 785, + 808 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "From the quantitative comparisons in Tab. 2, our performance on both seen and unseen prompts is competitive with baselines while our inference time is", + "bbox": [ + 214, + 810, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "K. Xie et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f4d2dd9e976d1cb01e70ff4f923565d90cfe19bb0c86bd337b931a6867995c01.jpg", + "image_caption": [ + "Fig.6: Results of user study showing the average preference of MVDream at different amounts of optimization time to LATTE3D gpt-101k on seen prompts." + ], + "image_footnote": [], + "bbox": [ + 227, + 150, + 500, + 333 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4c5a66890e75a1005ddeca924f8c637893c2fcbe09b0b69a7595bdd8044ac056.jpg", + "image_caption": [ + "Fig.7: Test time optimization on seen and unseen prompt. The model is trained on gpt-101k. First column: without test time optimization. Other columns: test time optimization with the given time budget denoted. Top: seen prompt \"...animal with a blue and white floral pattern...\" Bottom: unseen prompt \"orchid in clay pot\"." + ], + "image_footnote": [], + "bbox": [ + 519, + 143, + 759, + 276 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "an order of magnitude lower. Image-based metrics of FID and CLIP score may fail to detect generations with 3D geometry errors, less obvious in still images than in the user study, where users are shown videos of objects rotating through multiple views. Interested readers can refer to rendered videos of all methods in the supplementary materials.", + "bbox": [ + 212, + 441, + 787, + 518 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Total optimization cost A benefit of amortized optimization over many prompts is that our total optimization cost on gpt-101k is much lower than optimizing MVDream per-prompt. Fig. 6 illustrates that to achieve user preference on par with LATTE3D, MVDream would require 36 GPU minutes per prompt, whereas LATTE3D spent only 215 GPU seconds per prompt, representing a $10 \\times$ reduction in optimization cost.", + "bbox": [ + 212, + 541, + 787, + 632 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Application", + "text_level": 1, + "bbox": [ + 215, + 655, + 359, + 671 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We illustrate LATTE3D's capacity for enhanced quality through test time optimization and how to adapt LATTE3D for stylizing 3D content.", + "bbox": [ + 212, + 680, + 785, + 712 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Per Prompt Test-time Optimization We show experimentally that LATTE3D quickly adapts to new prompts during test time to improve quality. In Fig. 7, we qualitatively compare LATTE3D with up to 600 steps. Each optimization iteration takes an average of 0.98 seconds on an A100 GPU, giving us a total compute budget of less than 10 GPU minutes per prompt. Even with test-time optimization, our method is still an order of magnitude faster than per-prompt optimization methods such as MVDream, as in Tab. 2. Test-time optimization is", + "bbox": [ + 212, + 733, + 792, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "LATTE3D", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "particularly beneficial on unseen prompts, where FID drops by 11.6 for unseen vs. 8.8 for seen, and the CLIP score gains 0.04 for unseen vs. 0.02 for seen.", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/174459b7cdbcf8a9bbaa5b7529de9f97e75ffa73e2618f419c205ba645881d92.jpg", + "image_caption": [ + "Fig. 8: Results of user study showing the average preference rate for MVDream at different amounts of optimization time to LatTE3D stylization results on animal-style." + ], + "image_footnote": [], + "bbox": [ + 246, + 210, + 754, + 409 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Stylizing 3D content We illustrate how LATTE3D architecture offers us the flexibility to adapt it for targeted 3D content stylization as described in Sec. 3.6. To test this, we manually curate a dataset of 100 animal shapes from Objaverse as a basis and augment the species name of each animal shape name with combinations of activities (e.g. \"riding a skateboard\") and styles (e.g. \"voxel style\") to create our 12000 prompts animal-style dataset. We train a LATTE3D model (stage-1 and stage-2) on animal-style without the point cloud annealing stage. The resulting model adapts to both the text prompt and the point cloud inputs. In Fig. 4, each row contains the outputs of LATTE3D given different prompts while fixing the point cloud input, demonstrating the effectiveness of LATTE3D for text-controlled stylization.", + "bbox": [ + 212, + 520, + 787, + 686 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Fig. 8, we conduct a user study comparing LATTE3D's outputs on animal-style to MVDream and find that LATTE3D is competitive with MVDream with 4000 steps of optimization, representing a $10 \\times$ reduction in optimization cost (compared to running MVDream for each prompt). LATTE3D on animal-style also generalizes to held-out combinations of animal activities and styles with little quality drop, as indicated by its similar relative preference to MVDream. Therefore, LATTE3D enables users to cheaply create variations of their 3D assets via amortized optimization and even explore novel prompt combinations at test time. Additional quantitative and qualitative results can be found in the supplementary materials.", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "K. Xie et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/811428834e0f8ac5324cdda902f0e71d0f93b88e71a560a6e0a316afc816213a.jpg", + "table_caption": [ + "Table 3: Ablation of components in stage-1 training. Trained on gpt-101k data and evaluated on seen and unseen prompts. Preference indicate average user preference of baseline over LATTE3D. LATTE3D is better than all ablated settings in quantitative metrics and is preferred by users on the unseen prompt set." + ], + "table_footnote": [], + "table_body": "
ExpMV Unet Pretrain RegSeenUnseen
Mask-FID↓Render-FID↓Clip-Score↑Preference
ATT3D274.44275.010.209128.324.2
+MV✓243.24236.730.182634.826.9
+MV+UNet✓✓184.80203.330.221448.045.6
+MV+UNet+PT✓✓✓189.09189.510.219151.847.7
LATTE3D (S1)✓✓✓✓176.44186.840.2379--
", + "bbox": [ + 230, + 212, + 772, + 316 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 340, + 401, + 354 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We now analyze the value of each component we introduced in LATTE3D. We include additional results, including ablations on different diffusion priors, in the Appendix.", + "bbox": [ + 212, + 364, + 785, + 409 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Stage-1 ablation. Tab. 3 summarizes our quantitative comparison of the following ablations: $ATT3D$ baseline using a hypernet, $ATT3D + MV$ uses MVDream as diffusion guidance. $ATT3D + UNet + MV$ replaces the hypernet with our model architecture, but is initialized from scratch and given a dummy point cloud as input. $ATT3D + UNet + MV + PT$ is initialized from the reconstruction pretrained (PT) model. We introduce the Mask-FID metric to quantitatively assess geometry adherence to the shape dataset by computing the FID between the rendered binary masks of the generated shapes and those of the 3D dataset.", + "bbox": [ + 212, + 421, + 785, + 542 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We find that each of the components we introduce improves performance. Comparing ATT3D with ATT3D+MV, we see benefits from using the MV-Dream guidance as it mitigates geometric artifacts like Janus faces. Comparing ATT3D+MV with ATT3D+MV+UNet we see a performance boost from our architecture. Further, adding pretraining (ATT3D+UNet+MV+PT) helps, especially in recovering fine geometric and texture details. Finally, adding shape regularization (full LATTE3D) makes the generated geometry adhere better to the 3D data, as evidenced by the drop in Mask-FID. An additional user study is shown in Tab. 3. We show the convergence speed in the Appendix.", + "bbox": [ + 212, + 542, + 787, + 679 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation on using stage-2 refinement. Stage-2 refinement gives large improvements in texture details over stage-1 training. We show the comparison in Fig. 9.", + "bbox": [ + 214, + 691, + 785, + 722 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Regularization We now investigate the design choice of how to weight $\\alpha$ between the SDS loss and the regularization loss as in Eq. 3. In Tab. 3, we show that adding regularization loss improves the quantitative results. We qualitatively contrast training on different fixed weightings in Fig. 10 and the Appendix. We see greater consistency with the input shape by increasing shape regularization with higher blend factors $\\alpha$ .", + "bbox": [ + 212, + 734, + 785, + 824 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "LATTE3D", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/09e8636397efaa3277105f9cb0f25a00774127e50592d968ac3e3d34635ea567.jpg", + "image_caption": [ + "Fig.9: Comparison of stage-1 and stage-2 results. Stage-2 refinement significantly improves the texture quality." + ], + "image_footnote": [], + "bbox": [ + 228, + 152, + 483, + 208 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "\"fun sheep wearing skis in pixel art style, voxel\"", + "bbox": [ + 527, + 142, + 767, + 152 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d5aeebefcb872181cf9a9d91c95685eb82616f8cbed60f161a1354b0edd85845.jpg", + "image_caption": [ + "Fig. 10: Comparison of the impact of blending factor $\\alpha$ , where $(\\alpha = 0)$ is a failure from no regularization." + ], + "image_footnote": [], + "bbox": [ + 527, + 154, + 772, + 213 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/ef9eb354af008ed1c0139c15457c275a9945c8a6c25af83f5c6a5b5417d0c94a.jpg", + "table_caption": [ + "Table 4: An ablation of unseen DreamFusion inputs in inference and annealing in training. Users prefer the annealed model with a dummy input in $51.2\\%$ of cases." + ], + "table_footnote": [], + "table_body": "
Modelwithout annealwith anneal
InputTop-1DummyDummy
CLIP0.26280.25770.2605
FID186.46193.28190.00
", + "bbox": [ + 217, + 396, + 478, + 465 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Annealing in training and different input point clouds in inference. In training, we take retrieved point clouds from the dataset as input and anneal the input to become a fixed dummy point cloud. In inference, we only use the dummy point cloud as input. In Tab. 4, we quantitatively ablate the role of this point cloud annealing process on the behavior of LATTE3D on gpt-101k. Our model without annealing training is somewhat sensitive to the point cloud input. There is a small performance drop when using a", + "bbox": [ + 483, + 308, + 785, + 503 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "dummy as input in inference compared to a retrieved point cloud. However, the performance gap is reduced by introducing point cloud annealing so the model trains with dummy input. Qualitative examples are in the Appendix.", + "bbox": [ + 212, + 505, + 784, + 550 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 215, + 599, + 508, + 616 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We presented a scalable approach to perform amortized text-to-enhanced-3D generation. To successfully scale amortization to larger datasets, we used 3D data through (1) 3D-aware 2D SDS, (2) pretraining, and (3) 3D regularization of the amortization process. We further improve the model architecture to be more scalable. Our model generates high-quality shapes within $400\\mathrm{ms}$ . Moreover, quality and generalization can be improved further via a speedy test time optimization. Our model also has limitations. First, our model uses SDS and thus relies on the understanding of the text-to-image models, which can often fail to respect fine-grained details such as part descriptions. Second, the geometry is frozen in stage-2 and test-time optimization, so the geometry flaws from stage-1 cannot be fixed. Developing amortized training for stable geometry changes in stage-2 is left for future work.", + "bbox": [ + 212, + 657, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "K. Xie et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We thank Matan Atzmon, Or Perel, Clement Fuji Tsang, Masha Shugrina, and her group for helpful feedback. The Python community [30,45] made underlying tools, including PyTorch [32] & matplotlib [12].", + "bbox": [ + 212, + 175, + 787, + 222 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Disclosure of Funding", + "text_level": 1, + "bbox": [ + 215, + 243, + 429, + 261 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "NVIDIA funded this work. Kevin Xie, Jonathan Lorraine, Tianshi Cao, Jun Gao, and Xiaohui Zeng had funding from student scholarships at the University of Toronto and the Vector Institute, which are not in direct support of this work.", + "bbox": [ + 212, + 273, + 785, + 319 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 340, + 323, + 356 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achlioptas, P., Diamanti, O., Mitlagkas, I., Guibas, L.: Learning representations and generative models for 3d point clouds. In: International conference on machine learning. pp. 40-49. PMLR (2018)", + "2. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., Mello, S.D., Gallo, O., Guibas, L., Tremblay, J., Khamis, S., Karras, T., Wetzstein, G.: Efficient geometry-aware 3D generative adversarial networks. In: arXiv (2021)", + "3. Chen, R., Chen, Y., Jiao, N., Jia, K.: Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2023)", + "4. Chen, Y., Li, Z., Liu, P.: ET3D: efficient text-to-3d generation via multi-view distillation. arXiv 2311.15561 (2023)", + "5. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., VanderBilt, E., Kembhavi, A., Vondrick, C., Gkioxari, G., Ehsani, K., Schmidt, L., Farhadi, A.: Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects (2023)", + "6. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3d objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13142-13153 (June 2023)", + "7. Gao, J., Shen, T., Wang, Z., Chen, W., Yin, K., Li, D., Litany, O., Gojcic, Z., Fidler, S.: Get3d: A generative model of high quality 3d textured shapes learned from images. Advances in Neural Information Processing Systems (2022)", + "8. Guo, Y.C., Liu, Y.T., Shao, R., Laforte, C., Voleti, V., Luo, G., Chen, C.H., Zou, Z.X., Wang, C., Cao, Y.P., Zhang, S.H.: threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio (2023)", + "9. Gupta, A., Xiong, W., Nie, Y., Jones, I., Oğuz, B.: 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371 (2023)", + "10. Henzler, P., Mitra, N.J., Ritschel, T.: Escaping plato's cave: 3d shape from adversarial rendering. In: The IEEE International Conference on Computer Vision (ICCV) (October 2019)", + "11. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + ], + "bbox": [ + 217, + 371, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "LATTE3D", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "12. Hunter, J.D.: Matplotlib: A 2D graphics environment. Computing in Science & Engineering (2007)", + "13. Jun, H., Nichol, A.: Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023)", + "14. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics 39(6) (2020)", + "15. Li, J., Tan, H., Zhang, K., Xu, Z., Luan, F., Xu, Y., Hong, Y., Sunkavalli, K., Shakhnarovich, G., Bi, S.: Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214 (2023)", + "16. Li, M., Zhou, P., Liu, J.W., Keppo, J., Lin, M., Yan, S., Xu, X.: Instant3d: Instant text-to-3d generation. arXiv preprint arXiv:2311.08403 (2023)", + "17. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "18. Liu, M., Xu, C., Jin, H., Chen, L., Xu, Z., Su, H., et al.: One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928 (2023)", + "19. Liu, R., Wu, R., Hoorick, B.V., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object (2023)", + "20. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. In: Seminal graphics: pioneering efforts that shaped the field, pp. 347-353 (1998)", + "21. Lorraine, J., Xie, K., Zeng, X., Lin, C.H., Takikawa, T., Sharp, N., Lin, T.Y., Liu, M.Y., Fidler, S., Lucas, J.: Att3d: Amortized text-to-3d object synthesis. In: Proceedings of the International Conference on Computer Vision (ICCV) (2023)", + "22. Lunz, S., Li, Y., Fitzgibbon, A., Kushner, N.: Inverse graphics gan: Learning to generate 3d shapes from unstructured 2d data. arXiv preprint arXiv:2002.12674 (2020)", + "23. Luo, T., Rockwell, C., Lee, H., Johnson, J.: Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279 (2023)", + "24. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4460-4470 (2019)", + "25. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latentnerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600 (2022)", + "26. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "27. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: CVPR. pp. 8280-8290 (2022)", + "28. Nash, C., Ganin, Y., Eslami, S.M.A., Battaglia, P.W.: Polygen: An autoregressive generative model of 3d meshes (2020)", + "29. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751 (2022)", + "30. Oliphant, T.E.: Python for scientific computing. Computing in Science & Engineering (2007)" + ], + "bbox": [ + 217, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "K. Xie et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. OpenAI: Gpt-4 technical report. ArXiv abs/2303.08774 (2023), https://api.sementicscholar.org/CorpusID:257532815", + "32. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in PyTorch. Open-review (2017)", + "33. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)", + "34. Qian, G., Cao, J., Siarohin, A., Kant, Y., Wang, C., Vasilkovsky, M., Lee, H.Y., Fang, Y., Skorokhodov, I., Zhuang, P., et al.: Atom: Amortized text-to-mesh using 2d diffusion. arXiv preprint arXiv:2402.00867 (2024)", + "35. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)", + "36. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)", + "37. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022)", + "38. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)", + "39. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)", + "40. Shen, T., Munkberg, J., Hasselgren, J., Yin, K., Wang, Z., Chen, W., Gojcic, Z., Fidler, S., Sharp, N., Gao, J.: Flexible isosurface extraction for gradient-based mesh optimization. ACM Trans. Graph. 42(4) (jul 2023). https://doi.org/10.1145/3592430, https://doi.org/10.1145/3592430", + "41. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model (2023)", + "42. Shi, Y., Wang, P., Ye, J., Mai, L., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv:2308.16512 (2023)", + "43. Tang, J., Chen, Z., Chen, X., Wang, T., Zeng, G., Liu, Z.: Lgm: Large multiview gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054 (2024)", + "44. Tatarchenko, M., Dosovitskiy, A., Brox, T.: Octree generating networks: Efficient convolutional architectures for high-resolution 3d outputs. In: Proceedings of the IEEE international conference on computer vision. pp. 2088-2096 (2017)", + "45. Van Rossum, G., Drake Jr, F.L.: Python reference manual. Centrum voor Wiskunde en Informatica Amsterdam (1995)", + "46. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774 (2022)", + "47. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213 (2023)" + ], + "bbox": [ + 212, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "LATTE3D", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Wu, J., Zhang, C., Xue, T., Freeman, B., Tenenbaum, J.: Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems 29 (2016)", + "49. Yang, G., Huang, X., Hao, Z., Liu, M.Y., Belongie, S., Hariharan, B.: Pointflow: 3d point cloud generation with continuous normalizing flows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2019)", + "50. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34 (2021)", + "51. Zeng, X., Vahdat, A., Williams, F., Gojcic, Z., Litany, O., Fidler, S., Kreis, K.: Lion: Latent point diffusion models for 3d shape generation. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)", + "52. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models (2023)", + "53. Zhou, L., Du, Y., Wu, J.: 3d shape generation and completion through point-voxel diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5826-5835 (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 369 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "K. Xie et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/LATTE3D_ Large-scale Amortized Text-To-Enhanced3D Synthesis/a662796f-9e3f-4f91-946c-d99cfdaedd9c_model.json b/2024/LATTE3D_ Large-scale Amortized Text-To-Enhanced3D Synthesis/a662796f-9e3f-4f91-946c-d99cfdaedd9c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..54dbe4dd21dffd3311057fd1a7ace207ff9b7ff3 --- /dev/null +++ b/2024/LATTE3D_ Large-scale Amortized Text-To-Enhanced3D Synthesis/a662796f-9e3f-4f91-946c-d99cfdaedd9c_model.json @@ -0,0 +1,2546 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.306, + 0.142, + 0.698, + 0.187 + ], + "angle": 0, + "content": "LATTE3D: Large-scale Amortized Text-To-Enhanced3D Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.212, + 0.785, + 0.245 + ], + "angle": 0, + "content": "Kevin Xie\\(^{1,2}\\), Jonathan Lorraine\\(^{1,2}\\), Tianshi Cao\\(^{1,2}\\), Jun Gao\\(^{1,2}\\), James Lucas\\(^{1}\\), Antonio Torralba\\(^{1,3}\\), Sanja Fidler\\(^{1,2}\\), and Xiaohui Zeng\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.254, + 0.702, + 0.27 + ], + "angle": 0, + "content": "\\(^{1}\\)Nvidia, \\(^{2}\\)University of Toronto and Vector Institute, \\(^{3}\\)MIT" + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.301, + 0.761, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.564 + ], + "angle": 0, + "content": "Fig.1: Samples generated in \\(\\sim 400\\mathrm{ms}\\) on a single A6000 GPU from text prompts. Objects without prompt labels are generated by our text-to-3D model trained on \\(\\sim 100\\mathrm{k}\\) prompts, while labeled objects are generated by our 3D stylization model trained on \\(12\\mathrm{k}\\) prompts. See our project website** for more." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.592, + 0.741, + 0.801 + ], + "angle": 0, + "content": "Abstract. Recent text-to-3D generation approaches produce impressive 3D results but require time-consuming optimization that can take up to an hour per prompt. Amortized methods like ATT3D optimize multiple prompts simultaneously to improve efficiency, enabling fast text-to-3D synthesis. However, they cannot capture high-frequency geometry and texture details and struggle to scale to large prompt sets, so they generalize poorly. We introduce LATTE3D, addressing these limitations to achieve fast, high-quality generation on a significantly larger prompt set. Key to our method is 1) building a scalable architecture and 2) leveraging 3D data during optimization through 3D-aware diffusion priors, shape regularization, and model initialization to achieve robustness to diverse and complex training prompts. LATTE3D amortizes both neural field and textured surface generation to produce highly detailed textured meshes in a single forward pass. LATTE3D generates 3D objects in 400ms, and can be further enhanced with fast test-time optimization." + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.825, + 0.763, + 0.841 + ], + "angle": 0, + "content": "*Equal contribution. **https://research.nvidia.com/labs/toronto-ai/LATTE3D/" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "K. Xie et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.375, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.182, + 0.787, + 0.272 + ], + "angle": 0, + "content": "Recent advances in text-to-3D synthesis via image diffusion models mark significant progress in 3D content creation using [3,8,17,25,33,46,47]. However, these methods often involve an expensive and time-consuming optimization that can take up to an hour to generate a single 3D object [17, 33]. We aim to achieve real-time text-to-3D synthesis for diverse text prompts and support fast test-time optimization when a further quality boost is desired." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.273, + 0.787, + 0.392 + ], + "angle": 0, + "content": "Seminal work [33] proposed to optimize a neural field using the score distillation sampling (SDS) loss to generate 3D assets. It unlocked open-vocabulary 3D content creation but was (a) lacking high-frequency geometric and texture details, (b) expensive, and (c) prone to failing to generate the desired object. Follow-ups introduced a surface-based fine-tuning stage [3, 17, 47] to generate high-frequency details. While this two-stage pipeline defines most current state-of-the-art text-to-3D approaches, it remains expensive and prone to prompt failure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.789, + 0.529 + ], + "angle": 0, + "content": "ATT3D [21] was the first to make text-to-3D generation fast by simultaneously training a single model on a set of prompts in a process called amortized optimization. Amortization reduces training time by optimizing a shared text-conditioned model on a set of prompts, enabling the model to generalize to new prompts at inference time. Despite promising results on curated prompts, ATT3D – and the underlying per-prompt methods – remain prone to failing on general prompts. Also, ATT3D only amortizes the first stage of 3D synthesis, producing a neural field representation and thus cannot generate high-frequency details." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.532, + 0.787, + 0.592 + ], + "angle": 0, + "content": "To address these issues, we introduce LATTE3D, a Large-scale Amortized Text-to-Enhanced3D synthesis method that can produce high-quality 3D content in real-time. Our work makes the following technical contributions to boost the quality, robustness, scale, and speed of text-to-3D generative models:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.635 + ], + "angle": 0, + "content": "- We propose a novel text-to-3D amortization architecture that can scale to orders of magnitude larger prompt sets." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.637, + 0.793, + 0.68 + ], + "angle": 0, + "content": "- We leverage 3D data in training to improve quality and robustness, through 3D-aware diffusion priors, regularization loss, and weight initialization through pretraining with 3D reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.683, + 0.77, + 0.697 + ], + "angle": 0, + "content": "- We amortize the surface-based refinement stage, greatly boosting quality." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.606, + 0.793, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.726, + 0.388, + 0.741 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.762, + 0.787, + 0.839 + ], + "angle": 0, + "content": "Early work on 3D object synthesis was typically limited to generating objects from a given class [1,2,7,10,22,24,28,44,48,49,51,53], e.g. cars or chairs. Recent extensions use captioned 3D shapes to train text-to-3D models [13,29] generating shapes from diverse categories, but requiring 3D supervision, restricting them to synthetic datasets of limited size, diversity, and visual fidelity." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LATTE3D" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.544, + 0.418 + ], + "angle": 0, + "content": "The advent of differentiable rendering, both volumetric [26] and surface-based [27], opened the door to inverse image rendering [7, 33], unlocking the use of powerful text-to-image generative models [36-38] in the 3D synthesis process. DreamFusion [33] proposed the SDS loss to optimize a neural field using a text-to-image diffusion model to generate 3D assets. Follow-up work introduced a surface-based refinement stage [17], allowing the synthesis of high-frequency details. This two-stage optimization pipeline now defines most of the state-of-the-art text-to-3D approaches [3,25,46,47]. The recently released large-scale 3D dataset Objaverse-XL [5,6] has spurred researchers to explore the benefits of 3D supervision in the synthesis process. To date, the most successful approaches make" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.151, + 0.785, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.297, + 0.788, + 0.382 + ], + "angle": 0, + "content": "Fig.2: Comparison of SOTA text-to-3D methods on unseen prompts. We plot different methods' user study preference rates compared to LATTE3D. For MVDream, we report varying optimization times." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.785, + 0.464 + ], + "angle": 0, + "content": "text-conditioned image diffusion models 3D aware [15,19,41,42] by fine-tuning 2D diffusion models using rendered multiview images. SDS from multiview diffusion models is then used to obtain 3D shapes [19,41,42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.787, + 0.569 + ], + "angle": 0, + "content": "However, these approaches require a lengthy, per-prompt optimization process, hindering applicability to real-world content creation. To address this, two lines of work have been pursued. The first type of method uses a text-to-image-to-3D approach that generates images with a text-to-image diffusion model and trains an image-to-3D lifting network [15, 18, 41, 43]. They offer improved speed but with limited quality details compared to optimization methods, but still take \\(5 - 20\\) seconds per prompt [15, 43] for sampling from the image diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.691 + ], + "angle": 0, + "content": "In an alternate line of work, ATT3D [21] proposed an amortized framework that optimizes multiple prompts simultaneously with a unified model. Follow-up works expanded the network architecture to achieve better quality [16, 34]. LATTE3D extends this line of work to larger scales by incorporating 3D knowledge to achieve prompt robustness during training, among other design improvements like better architecture. Amortized text-to-3D offers an attractive quality vs. speed trade-off as it uses a single feed-forward architecture without requiring sampling of diffusion models - see Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Concurrent Works Concurrently, AToM [34] also amortizes two-stage training with a triplane representation, but they do not scale beyond the smaller sets of prompts from ATT3D. In contrast, we scale to the order of 100000 prompts, using 3D-aware techniques. Another concurrent work of ET3D [4] achieves fast generation by training a GAN model, but it is only trained on a small-scale compositional dataset of 5000 prompts and does not guarantee view consistency as it uses a 2D neural upsampler. Lastly, LGM [43] is a text-to-image-to-3D method that generates Gaussian splats in 5 seconds. We experimentally demonstrate competitive performance to it." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "K. Xie et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.381, + 0.163 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.325 + ], + "angle": 0, + "content": "An established pipeline for high-quality text-to-3D generation consists of two stages, each performing per-prompt optimization using the SDS loss with a text-to-image diffusion model [3, 17, 47]. Stage-1 optimizes a volumetric representation, which is typically a neural radiance field. Meshes afford real-time rendering but are hard to optimize from scratch with image supervision alone. Stage-2 uses the output of the neural field to initialize a signed distance field (SDF) and a texture field, from which a mesh can be derived using differentiable isosurfacing [39, 40]. This surface-based representation is then optimized via differentiable rasterization [27], which is fast even at 1024 rendered image resolution – enabling supervision from higher-resolution text-to-image diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.788, + 0.431 + ], + "angle": 0, + "content": "ATT3D [21] amortizes optimization of a neural field over a set of prompts with a hypernetwork mapping a text prompt to a neural field, which is trained with the SDS loss over a prompt set, referred to as seen prompts. LATTE3D introduces a new architecture that amortizes both stages of the generation process, aiming to produce high-quality textured meshes in real-time. We scale the approach to a magnitudes larger and more diverse set of prompts by leveraging 3D knowledge in the amortization loop." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.598 + ], + "angle": 0, + "content": "Our approach is illustrated in Fig. 3. We initialize LATTE3D with a reconstruction pretraining step to stabilize training (Sec. 3.1). The architecture consists of two networks, one for geometry and the other for texture (Sec. 3.2). Amortized model training with diffusion priors is done through a two-stage pipeline consisting of a volume-based stage-1 and a surface-based stage-2 (Sec. 3.3). To reconcile reconstruction pretraining, which uses 3D geometry as input, with the task of text-to-3D generation, we anneal the network's input during stage-1 to gradually fully replace the dependence on 3D input with a single dummy input. During inference (Sec. 3.4), our model generates a 3D textured mesh from just a text prompt in 400ms and allows an optional lightweight test-time refinement to enhance the quality of geometry and textures (Sec. 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.618, + 0.544, + 0.634 + ], + "angle": 0, + "content": "3.1 Pretraining to reconstruct shapes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.788, + 0.777 + ], + "angle": 0, + "content": "Although SDS can synthesize text-aligned 3D features, empirically, it is a less efficient source of supervision due to high variance. We find that pretraining the model to first encode-decode 3D shapes subsequently accelerates the early stage of amortized SDS training. We initialize our model \\(\\mathcal{M}\\) by pretraining it on a 3D asset dataset using image reconstruction losses, similar to the reconstruction stage used in [9]. \\(\\mathcal{M}\\) inputs a sampled point cloud and outputs a predicted 3D shape. We then render multiple views of the predicted 3D shape \\(o\\) to compare with the input shape \\(s\\) renderings. We use an \\(\\ell_2\\) loss on the rendered opacity and RGB image:" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.786, + 0.786, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c o n}} (o, s, c) = \\left| \\left| \\mathbf {R} _ {\\text {o p a c i t y}} (o, c) - \\mathbf {R} _ {\\text {o p a c i t y}} (s, c) \\right| \\right| _ {2} + \\left| \\left| \\mathbf {R} _ {\\text {R G B}} (o, c) - \\mathbf {R} _ {\\text {R G B}} (s, c) \\right| \\right| _ {2} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We denote the renderer \\(\\mathbf{R}\\), the opacity and RGB rendering with \\(\\mathbf{R}_{\\mathrm{opacity}}\\) and \\(\\mathbf{R}_{\\mathrm{RGB}}\\), respectively, and the camera with \\(c\\), randomly sampled on a sphere." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LATTE3D" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.147, + 0.784, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.295, + 0.788, + 0.38 + ], + "angle": 0, + "content": "Fig. 3: LATTE3D consists of two networks: a texture network \\( T \\) and geometry network \\( G \\). When amortizing the first stage, the encoders of both networks share the same set of weights. The training objective includes an SDS gradient from a 3D-aware image prior and a regularization loss that compares the rendered predicted shape's mask with the rendered masks of 3D assets in a library. When amortizing surface-based refinement in stage-2, we freeze the geometry network \\( G \\) and update the texture network \\( T \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.409, + 0.427, + 0.423 + ], + "angle": 0, + "content": "3.2 Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.555 + ], + "angle": 0, + "content": "Our model \\(\\mathcal{M}\\) consists of two networks, \\(G\\) and \\(T\\), for predicting geometry and texture, respectively. After pretraining, we introduce text-conditioning by adding cross-attention layers. We use CLIP [35] embeddings for encoding the text prompt \\(p\\). In stage-1 training, we tie the geometry and texture networks' weights (except for separate MLP decoder heads), effectively training a single input encoder. In stage-2, we freeze the geometry network \\(G\\) and refine the texture network \\(T\\) with an additional trainable upsampling layer. The final model \\(\\mathcal{M}\\) used in inference is the frozen \\(G\\) from stage-1 and a refined \\(T\\) from stage-2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.736 + ], + "angle": 0, + "content": "The geometry network \\( G \\) consists of a U-Net encoder on a triplane [2,7], followed by a decoder containing another triplane U-Net and a neural volumetric density field. Specifically, a point cloud is fed through a PointNet, then converted to a triplane via average scattering. During training, the point cloud comes from the 3D data, and in inference, a dummy sphere point cloud is used. While this point cloud encoder is inherited from reconstruction pretraining, it becomes unimportant in the final model. The triplane is processed through the encoder and decoder U-Nets. The text embedding is fed into every residual U-Net block in the encoder via cross-attention. For any point in 3D space, we extract triplane features from each plane via bilinear interpolation. An MLP then predicts the density value at that point." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The texture network \\( T \\) and geometry network \\( G \\) share the same encoder and decoder U-Net architecture in pretraining and stage-1. We employ another non-shared MLP neural texture field to predict RGB values at query points. In stage-2, we upsample the triplane features produced by the encoder to a higher resolution to gain extra capacity to generate high-frequency texture details. Specifically, we first bilinearly upsample the latent triplane from 128 resolution to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "K. Xie et al." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.145, + 0.775, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.321, + 0.789, + 0.351 + ], + "angle": 0, + "content": "Fig. 4: Stylization application: Our model learns to generate diverse stylizations of the same shapes. Left column shows the original shape to be stylized." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.38, + 0.789, + 0.427 + ], + "angle": 0, + "content": "480 and add a residual MLP which maps the text-embedding to the residual of the triplane feature, mimicking ATT3D's mapping (hyper)network. The decoder then takes this upsampled triplane as input. Details are in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.442, + 0.789, + 0.609 + ], + "angle": 0, + "content": "Rendering. To train the model, we render the generated 3D object into 2D images with different methods in stage-1 and 2. In stage-1, the output 3D geometry is a triplane-based neural field, on which we use volume rendering to get images of 256 resolution using the formulation from VolSDF [50], allowing us to parameterize the density via an approximate signed distance field whose zero-level set defines the object geometry. This choice allows a simple conversion to surface-based rendering. For the geometry network, we observe more stable training with volume rendering with SDS compared to surface rendering. In stage-2, we extract the isosurface from the density field with Marching Cubes [20] and render images via rasterization [14]. Meshes can be rendered at a 1024 image resolution in real-time, helping capture finer details for learning textures." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.632, + 0.431, + 0.648 + ], + "angle": 0, + "content": "3.3 Amortized Learning" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.658, + 0.789, + 0.735 + ], + "angle": 0, + "content": "Amortized learning distills knowledge from image diffusion priors into our model. We perform amortized training on volumetric geometry training (stage-1) and surface texture training (stage-2) sequentially. We incorporate 3D information into our training process (a) implicitly via 3D-aware 2D SDS losses and (b) explicitly with regularization." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.75, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Curating prompt sets. We create large prompt sets for training using rule-based text generation or ChatGPT [31]. For the former, we start with the categories names from the captions of Objaverse [23], and design rules, like \"object A in style B is doing C\" with a set of styles and activities. Alternatively, we input the captions into ChatGPT and ask it for detailed and diverse prompts describing a similar 3D object. Fig. 3 shows that when generating \"Amigurumi fiddler crab" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LATTE3D" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "wearing a top hat\", we guide shape synthesis towards resembling a crab by retrieving a related 3D shape used in a regularization loss. Other details of dataset construction are in Sec. 4.1." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.212, + 0.784, + 0.256 + ], + "angle": 0, + "content": "Amortizing Stage-1 Generation In this stage, we train \\(\\mathcal{M}\\) with our datasets. We use the SDS loss with a 3D-aware 2D prior and regularization to the paired 3D shape, which we now explain in detail:" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.268, + 0.784, + 0.343 + ], + "angle": 0, + "content": "3D-aware SDS loss. During stage-1 training we leverage a 3D aware diffusion prior, which provides a stronger, multiview consistent supervisory signal. Specifically, we use the model from MVDream [42], which was trained on four rendered views of objects from the Objaverse dataset [6] by modifying Stable Diffusion [37] to generate multiview consistent images." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.355, + 0.784, + 0.415 + ], + "angle": 0, + "content": "Regularization loss. While 3D-aware SDS loss with MVDream offers a strong multiview prior, it can still fail to correctly capture the full 3D geometry for some prompts since MVDream only supervises four views and can create enlarged geometry or floaters in the shape. Some failures are shown on the right of Fig. 10." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.416, + 0.784, + 0.476 + ], + "angle": 0, + "content": "Combining regularization using 3D shapes with the SDS loss helps training and improves geometry. As shown in Fig. 3, for each training prompt, we regularize the output shape \\(o\\) by comparing its rendered mask to the mask of a shape \\(s\\) retrieved from the 3D dataset with the prompt:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.485, + 0.784, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e g}} (o, s, c) = \\left| \\left| \\mathbf {R} _ {\\text {o p a c i t y}} (o, c) - \\mathbf {R} _ {\\text {o p a c i t y}} (s, c) \\right| \\right| _ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.507, + 0.784, + 0.552 + ], + "angle": 0, + "content": "where \\(\\mathbf{R}_{\\mathrm{opacity}}\\) is the opacity mask (volume-rendered density). We balance the regularization \\(\\mathcal{L}_{\\mathrm{reg}}\\) and the SDS loss \\(\\mathcal{L}_{\\mathrm{SDS}}\\) by simply blending them using a convex combination with weight \\(\\alpha\\). Our final training loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.562, + 0.784, + 0.576 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t r a i n}} = (1 - \\alpha) \\mathcal {L} _ {\\text {S D S}} + \\alpha \\mathcal {L} _ {\\text {r e g}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.584, + 0.784, + 0.614 + ], + "angle": 0, + "content": "During training, we render five views for each generated shape. We average across four views to compute \\(\\mathcal{L}_{\\mathrm{SDS}}\\), and one is used to compute \\(\\mathcal{L}_{\\mathrm{reg}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.625, + 0.784, + 0.73 + ], + "angle": 0, + "content": "Input point cloud annealing To bridge the gap between the real point cloud inputs in training and dummy point cloud in inference, we gradually anneal the input point clouds towards dummy point clouds during stage-1 training. Specifically, we randomly replace a subset of points in the input point cloud with points from the dummy point cloud using a probability that increases linearly from 0 to 1 in the last 5000 training iterations. With point cloud annealing, the model gets better results in inference when the dummy point cloud is used." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.75, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Amortizing Stage-2 Generation During stage-2 training, we freeze the network for our geometry and only tune the texture since training both leads to instabilities. We use depth-conditional ControlNet [52] for SDS guidance, allowing a higher, 512 resolution supervision. Depth conditioning encourages the SDS loss to guide the texture to align with the geometry from the geometry network, thereby improving the 3D texture consistency." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "K. Xie et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.341, + 0.16 + ], + "angle": 0, + "content": "3.4 Inference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.169, + 0.788, + 0.321 + ], + "angle": 0, + "content": "During inference, our model inputs the user's text prompt and a dummy point cloud. Our \\(\\mathcal{M} = (G,T)\\) outputs a textured mesh, where \\(T\\) is the final texture network refined in stage-2, while \\(G\\) is trained in stage-1 only and then frozen. Predicting one shape and rendering the image only takes us 400ms on an A6000 GPU in inference, allowing the generation of user results at interactive speeds. A single A6000 can generate 4 samples simultaneously. Once the user selects a desired sample, textured meshes can be exported from LATTE3D. We use a UV mapping tool to UV parameterize the mesh output by \\(G\\) and use \\(T\\) to predict the albedo colors of each texel in a UV-mapped texture image. The resulting mesh and texture image is compatible with standard rendering software." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.34, + 0.461, + 0.355 + ], + "angle": 0, + "content": "3.5 Test Time Optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.362, + 0.788, + 0.468 + ], + "angle": 0, + "content": "Our method supports test time optimization if the user wants to boost the quality of a particular prompt. This can be useful if the user's prompt deviates significantly from the seen prompts. As in our stage-2 training, we freeze the geometry and finetune the texture network \\( T \\) with depth-conditional SDS guidance. Our test time optimization is significantly faster to converge than MVDream [42] or other optimization-based text-to-3D methods [17, 33, 47] since we optimize our amortized network instead of a randomly initialized network." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.487, + 0.381, + 0.501 + ], + "angle": 0, + "content": "3.6 3D Stylization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.66 + ], + "angle": 0, + "content": "LATTE3D can also be trained and deployed as a 3D stylization method, allowing users to cheaply create variations from existing 3D assets. During training, we skip the input point cloud annealing step to maintain point cloud reliance and amortize over a dataset where each 3D shape corresponds to many style prompts. Our blend of regularization and SDS guidance drives the model to produce shapes structurally similar to the original shapes yet semantically aligned with the style prompt. Amortization significantly reduces the computational cost per shape-prompt pair compared to optimizing each pair separately. During inference, point clouds from the training set can be combined with novel prompts as input to the model to produce variations of the original shape." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.681, + 0.377, + 0.697 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.709, + 0.788, + 0.77 + ], + "angle": 0, + "content": "We first describe our experimental setup, including datasets, metrics, and baselines for comparison (Sec. 4.1). We then present our quantitative and qualitative results (Sec. 4.2), demonstrating the performance and generalization ability of LATTE3D (Sec. 4.3), followed by design choices ablations (Sec. 4.4)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.789, + 0.429, + 0.804 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.811, + 0.787, + 0.827 + ], + "angle": 0, + "content": "We now discuss dataset construction, model pretraining, and evaluation metrics." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LATTE3D" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.147, + 0.328, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.243, + 0.315, + 0.262 + ], + "angle": 0, + "content": "MVDream (6min)" + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.147, + 0.403, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.243, + 0.393, + 0.262 + ], + "angle": 0, + "content": "MVDream (36min)" + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.145, + 0.466, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.243, + 0.465, + 0.262 + ], + "angle": 0, + "content": "3DTopia (20s)" + }, + { + "type": "image", + "bbox": [ + 0.476, + 0.147, + 0.545, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.243, + 0.517, + 0.262 + ], + "angle": 0, + "content": "LGM (5s)" + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.147, + 0.62, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.242, + 0.596, + 0.252 + ], + "angle": 0, + "content": "ATT3D" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.146, + 0.689, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.242, + 0.68, + 0.261 + ], + "angle": 0, + "content": "ATT3D+ UNet \\(^+\\) MV" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.147, + 0.757, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.242, + 0.747, + 0.262 + ], + "angle": 0, + "content": "LATTE3D (400ms)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.273, + 0.787, + 0.316 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparison between LATTE3D and ATT3D [21], MVDream [42], Instant3D/3DTopia [15] and LGM [43]. Both rows are unseen prompts: \"a red convertible car with the top down\", \"an ice cream sundae\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.487, + 0.493 + ], + "angle": 0, + "content": "Training Set To investigate scaling amortized optimization beyond what has been done in prior works - see Tab. 1 - we construct a new dataset gpt-101k that consists of 101k text prompts and 34k shapes. For the shapes, we use the \"lvis\" subset from Objaverse [6] with 50k human-verified shapes. We filter out poor shapes, e.g. flat images and scans, and retain 34k" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.785, + 0.525 + ], + "angle": 0, + "content": "shapes for our dataset. We use ChatGPT to augment each object's caption to gather 3 prompts per object." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.358, + 0.785, + 0.386 + ], + "angle": 0, + "content": "Table 1: Comparing the sizes of amortization training sets to prior works." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.388, + 0.787, + 0.467 + ], + "angle": 0, + "content": "
name#promptsSource
Animal2400 [21,34]2400Rule
DF415 [34]415DF [33]
animal-style (Ours)12000Rule
gpt-101k (Ours)101608GPT
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.643 + ], + "angle": 0, + "content": "Evaluation Sets We use two prompt sets for evaluation. To evaluate the benefit of amortized optimization in our method, we evaluate our model on gpt-101k. However, due to the computational cost of running the baselines, we randomly select a 50 prompt subset in gpt-101k to form the seen prompt set for benchmarking. Furthermore, to gauge the generalization abilities of our amortized model, we construct an unseen prompt set by filtering a subset of 67 DreamFusion [33] prompts that are close in distribution to lvis categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.654, + 0.792, + 0.761 + ], + "angle": 0, + "content": "Baselines. We compare LATTE3D with the baselines: ATT3D [21], MVDream [42], Instant3D [15] and LGM [43]. We re-implement ATT3D [21] with a hypernetwork. We use the open-source threestudio [8] implementation of MVDream. For MVDream, we train each 3D model for 10000 iterations, with a batch size of 8 on a single A100 GPU with provided default settings. For Instant3D, since the source code has not been released, we use the re-implementation from 3DTopia1. For LGM, we use the official open-source implementation." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.772, + 0.787, + 0.819 + ], + "angle": 0, + "content": "Evaluation metrics. We quantitatively compare our model with baselines by evaluating the generated 3D content fidelity and its consistency to the text prompt for the optimization cost. We use three metrics to evaluate fidelity." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.5, + 0.841 + ], + "angle": 0, + "content": "1 https://github.com/3DTopia/3DTopia" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "K. Xie et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Quantitative metrics and average user preference (\\%) of baselines over LATTE3D trained on gpt-101k using seen and unseen prompts. We also report test-time optimization, which takes \\(10\\mathrm{min}\\)." + }, + { + "type": "table", + "bbox": [ + 0.256, + 0.199, + 0.744, + 0.355 + ], + "angle": 0, + "content": "
ModelTimeRender-FID ↓ seenCLIP-Score ↑ unseen(df)Preference % ↑seenUnseen(df)seenUnseen(df)
LATTE3D400ms180.58190.000.26080.2605--
LATTE3D-opt10min171.69178.370.28000.3000--
MVDream6min203.53199.230.22400.26982.512.4
MVDream18min191.77188.840.24400.295026.552.2
MVDream36min158.11143.440.28300.318048.589.3
3DTopia20s217.29222.760.18370.217120.825.6
LGM5s178.67166.400.28400.306326.236.4
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.784, + 0.44 + ], + "angle": 0, + "content": "Render-FID: We compute the FID [11] between the renderings of generated 3D content and a set of images sampled from Stable Diffusion [37] with the same text prompts as input. This metric measures how well the generated shapes align with those from the 2D prior in visual quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.784, + 0.486 + ], + "angle": 0, + "content": "CLIP Score: We compute the average CLIP scores between the text prompt and each rendered image to indicate how well the generated 3D shape aligns with the input text prompt." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.785, + 0.561 + ], + "angle": 0, + "content": "User Study: We evaluate the overall 3D geometry and texture through Amazon Mechanical Turk user studies. We present videos of 3D shapes generated by our method and a baseline head-to-head, with corresponding text prompts, and ask users for their preference. Each comparison is averaged over three users to obtain preference rate for each baseline relative to LATTE3D." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.786, + 0.669 + ], + "angle": 0, + "content": "Timing: We measure our optimization cost by GPU time per prompt to gauge the total compute cost on each prompt. We divide the total wall-clock time for pretraining, stage-1, and stage-2 by the prompt set's size, then multiply by the number of GPUs used. We measure inference time for Instant3D, LGM, and LATTE3D by the time the model takes from in-taking the text input to outputting the final features before rendering. For MVDream [42], we measure the average time to optimize a single prompt on an A100 GPU." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.693, + 0.441, + 0.708 + ], + "angle": 0, + "content": "4.2 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.787, + 0.809 + ], + "angle": 0, + "content": "We observe robust generalization of LATTE3D by training on the gpt-101k dataset to unseen prompts (df, from DreamFusion [33]), as demonstrated by qualitative (Fig. 5) and quantitative (Fig. 2, Tab. 2) results. We highlight the relative user preference of LATTE3D versus different baselines over inference time in Fig. 2. LATTE3D produce results of competitive user preference to SOTA baselines at a uniquely fast inference time." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "From the quantitative comparisons in Tab. 2, our performance on both seen and unseen prompts is competitive with baselines while our inference time is" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LATTE3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.151, + 0.501, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.346, + 0.51, + 0.404 + ], + "angle": 0, + "content": "Fig.6: Results of user study showing the average preference of MVDream at different amounts of optimization time to LATTE3D gpt-101k on seen prompts." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.145, + 0.761, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.288, + 0.784, + 0.414 + ], + "angle": 0, + "content": "Fig.7: Test time optimization on seen and unseen prompt. The model is trained on gpt-101k. First column: without test time optimization. Other columns: test time optimization with the given time budget denoted. Top: seen prompt \"...animal with a blue and white floral pattern...\" Bottom: unseen prompt \"orchid in clay pot\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.52 + ], + "angle": 0, + "content": "an order of magnitude lower. Image-based metrics of FID and CLIP score may fail to detect generations with 3D geometry errors, less obvious in still images than in the user study, where users are shown videos of objects rotating through multiple views. Interested readers can refer to rendered videos of all methods in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.633 + ], + "angle": 0, + "content": "Total optimization cost A benefit of amortized optimization over many prompts is that our total optimization cost on gpt-101k is much lower than optimizing MVDream per-prompt. Fig. 6 illustrates that to achieve user preference on par with LATTE3D, MVDream would require 36 GPU minutes per prompt, whereas LATTE3D spent only 215 GPU seconds per prompt, representing a \\(10 \\times\\) reduction in optimization cost." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.656, + 0.36, + 0.672 + ], + "angle": 0, + "content": "4.3 Application" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.787, + 0.713 + ], + "angle": 0, + "content": "We illustrate LATTE3D's capacity for enhanced quality through test time optimization and how to adapt LATTE3D for stylizing 3D content." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.794, + 0.842 + ], + "angle": 0, + "content": "Per Prompt Test-time Optimization We show experimentally that LATTE3D quickly adapts to new prompts during test time to improve quality. In Fig. 7, we qualitatively compare LATTE3D with up to 600 steps. Each optimization iteration takes an average of 0.98 seconds on an A100 GPU, giving us a total compute budget of less than 10 GPU minutes per prompt. Even with test-time optimization, our method is still an order of magnitude faster than per-prompt optimization methods such as MVDream, as in Tab. 2. Test-time optimization is" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "K. Xie et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "particularly beneficial on unseen prompts, where FID drops by 11.6 for unseen vs. 8.8 for seen, and the CLIP score gains 0.04 for unseen vs. 0.02 for seen." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.211, + 0.756, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.422, + 0.785, + 0.451 + ], + "angle": 0, + "content": "Fig. 8: Results of user study showing the average preference rate for MVDream at different amounts of optimization time to LatTE3D stylization results on animal-style." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.687 + ], + "angle": 0, + "content": "Stylizing 3D content We illustrate how LATTE3D architecture offers us the flexibility to adapt it for targeted 3D content stylization as described in Sec. 3.6. To test this, we manually curate a dataset of 100 animal shapes from Objaverse as a basis and augment the species name of each animal shape name with combinations of activities (e.g. \"riding a skateboard\") and styles (e.g. \"voxel style\") to create our 12000 prompts animal-style dataset. We train a LATTE3D model (stage-1 and stage-2) on animal-style without the point cloud annealing stage. The resulting model adapts to both the text prompt and the point cloud inputs. In Fig. 4, each row contains the outputs of LATTE3D given different prompts while fixing the point cloud input, demonstrating the effectiveness of LATTE3D for text-controlled stylization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In Fig. 8, we conduct a user study comparing LATTE3D's outputs on animal-style to MVDream and find that LATTE3D is competitive with MVDream with 4000 steps of optimization, representing a \\(10 \\times\\) reduction in optimization cost (compared to running MVDream for each prompt). LATTE3D on animal-style also generalizes to held-out combinations of animal activities and styles with little quality drop, as indicated by its similar relative preference to MVDream. Therefore, LATTE3D enables users to cheaply create variations of their 3D assets via amortized optimization and even explore novel prompt combinations at test time. Additional quantitative and qualitative results can be found in the supplementary materials." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LATTE3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 3: Ablation of components in stage-1 training. Trained on gpt-101k data and evaluated on seen and unseen prompts. Preference indicate average user preference of baseline over LATTE3D. LATTE3D is better than all ablated settings in quantitative metrics and is preferred by users on the unseen prompt set." + }, + { + "type": "table", + "bbox": [ + 0.231, + 0.213, + 0.773, + 0.317 + ], + "angle": 0, + "content": "
ExpMV Unet Pretrain RegSeenUnseen
Mask-FID↓Render-FID↓Clip-Score↑Preference
ATT3D274.44275.010.209128.324.2
+MV✓243.24236.730.182634.826.9
+MV+UNet✓✓184.80203.330.221448.045.6
+MV+UNet+PT✓✓✓189.09189.510.219151.847.7
LATTE3D (S1)✓✓✓✓176.44186.840.2379--
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.342, + 0.403, + 0.356 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.786, + 0.41 + ], + "angle": 0, + "content": "We now analyze the value of each component we introduced in LATTE3D. We include additional results, including ablations on different diffusion priors, in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.787, + 0.543 + ], + "angle": 0, + "content": "Stage-1 ablation. Tab. 3 summarizes our quantitative comparison of the following ablations: \\( ATT3D \\) baseline using a hypernet, \\( ATT3D + MV \\) uses MVDream as diffusion guidance. \\( ATT3D + UNet + MV \\) replaces the hypernet with our model architecture, but is initialized from scratch and given a dummy point cloud as input. \\( ATT3D + UNet + MV + PT \\) is initialized from the reconstruction pretrained (PT) model. We introduce the Mask-FID metric to quantitatively assess geometry adherence to the shape dataset by computing the FID between the rendered binary masks of the generated shapes and those of the 3D dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.788, + 0.68 + ], + "angle": 0, + "content": "We find that each of the components we introduce improves performance. Comparing ATT3D with ATT3D+MV, we see benefits from using the MV-Dream guidance as it mitigates geometric artifacts like Janus faces. Comparing ATT3D+MV with ATT3D+MV+UNet we see a performance boost from our architecture. Further, adding pretraining (ATT3D+UNet+MV+PT) helps, especially in recovering fine geometric and texture details. Finally, adding shape regularization (full LATTE3D) makes the generated geometry adhere better to the 3D data, as evidenced by the drop in Mask-FID. An additional user study is shown in Tab. 3. We show the convergence speed in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.692, + 0.787, + 0.723 + ], + "angle": 0, + "content": "Ablation on using stage-2 refinement. Stage-2 refinement gives large improvements in texture details over stage-1 training. We show the comparison in Fig. 9." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.825 + ], + "angle": 0, + "content": "Regularization We now investigate the design choice of how to weight \\(\\alpha\\) between the SDS loss and the regularization loss as in Eq. 3. In Tab. 3, we show that adding regularization loss improves the quantitative results. We qualitatively contrast training on different fixed weightings in Fig. 10 and the Appendix. We see greater consistency with the input shape by increasing shape regularization with higher blend factors \\(\\alpha\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "K. Xie et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.154, + 0.484, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.218, + 0.484, + 0.26 + ], + "angle": 0, + "content": "Fig.9: Comparison of stage-1 and stage-2 results. Stage-2 refinement significantly improves the texture quality." + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.143, + 0.769, + 0.153 + ], + "angle": 0, + "content": "\"fun sheep wearing skis in pixel art style, voxel\"" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.155, + 0.774, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.226, + 0.779, + 0.268 + ], + "angle": 0, + "content": "Fig. 10: Comparison of the impact of blending factor \\(\\alpha\\), where \\((\\alpha = 0)\\) is a failure from no regularization." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.325, + 0.477, + 0.392 + ], + "angle": 0, + "content": "Table 4: An ablation of unseen DreamFusion inputs in inference and annealing in training. Users prefer the annealed model with a dummy input in \\(51.2\\%\\) of cases." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.397, + 0.479, + 0.466 + ], + "angle": 0, + "content": "
Modelwithout annealwith anneal
InputTop-1DummyDummy
CLIP0.26280.25770.2605
FID186.46193.28190.00
" + }, + { + "type": "text", + "bbox": [ + 0.485, + 0.309, + 0.787, + 0.505 + ], + "angle": 0, + "content": "Annealing in training and different input point clouds in inference. In training, we take retrieved point clouds from the dataset as input and anneal the input to become a fixed dummy point cloud. In inference, we only use the dummy point cloud as input. In Tab. 4, we quantitatively ablate the role of this point cloud annealing process on the behavior of LATTE3D on gpt-101k. Our model without annealing training is somewhat sensitive to the point cloud input. There is a small performance drop when using a" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.785, + 0.551 + ], + "angle": 0, + "content": "dummy as input in inference compared to a retrieved point cloud. However, the performance gap is reduced by introducing point cloud annealing so the model trains with dummy input. Qualitative examples are in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.601, + 0.509, + 0.617 + ], + "angle": 0, + "content": "5 Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We presented a scalable approach to perform amortized text-to-enhanced-3D generation. To successfully scale amortization to larger datasets, we used 3D data through (1) 3D-aware 2D SDS, (2) pretraining, and (3) 3D regularization of the amortization process. We further improve the model architecture to be more scalable. Our model generates high-quality shapes within \\(400\\mathrm{ms}\\). Moreover, quality and generalization can be improved further via a speedy test time optimization. Our model also has limitations. First, our model uses SDS and thus relies on the understanding of the text-to-image models, which can often fail to respect fine-grained details such as part descriptions. Second, the geometry is frozen in stage-2 and test-time optimization, so the geometry flaws from stage-1 cannot be fixed. Developing amortized training for stable geometry changes in stage-2 is left for future work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LATTE3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.223 + ], + "angle": 0, + "content": "We thank Matan Atzmon, Or Perel, Clement Fuji Tsang, Masha Shugrina, and her group for helpful feedback. The Python community [30,45] made underlying tools, including PyTorch [32] & matplotlib [12]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.244, + 0.431, + 0.262 + ], + "angle": 0, + "content": "Disclosure of Funding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.274, + 0.787, + 0.32 + ], + "angle": 0, + "content": "NVIDIA funded this work. Kevin Xie, Jonathan Lorraine, Tianshi Cao, Jun Gao, and Xiaohui Zeng had funding from student scholarships at the University of Toronto and the Vector Institute, which are not in direct support of this work." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.342, + 0.324, + 0.357 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.372, + 0.786, + 0.414 + ], + "angle": 0, + "content": "1. Achlioptas, P., Diamanti, O., Mitlagkas, I., Guibas, L.: Learning representations and generative models for 3d point clouds. In: International conference on machine learning. pp. 40-49. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.414, + 0.786, + 0.455 + ], + "angle": 0, + "content": "2. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., Mello, S.D., Gallo, O., Guibas, L., Tremblay, J., Khamis, S., Karras, T., Wetzstein, G.: Efficient geometry-aware 3D generative adversarial networks. In: arXiv (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.455, + 0.787, + 0.497 + ], + "angle": 0, + "content": "3. Chen, R., Chen, Y., Jiao, N., Jia, K.: Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.497, + 0.787, + 0.524 + ], + "angle": 0, + "content": "4. Chen, Y., Li, Z., Liu, P.: ET3D: efficient text-to-3d generation via multi-view distillation. arXiv 2311.15561 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.524, + 0.787, + 0.579 + ], + "angle": 0, + "content": "5. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., VanderBilt, E., Kembhavi, A., Vondrick, C., Gkioxari, G., Ehsani, K., Schmidt, L., Farhadi, A.: Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.579, + 0.787, + 0.634 + ], + "angle": 0, + "content": "6. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3d objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13142-13153 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.787, + 0.675 + ], + "angle": 0, + "content": "7. Gao, J., Shen, T., Wang, Z., Chen, W., Yin, K., Li, D., Litany, O., Gojcic, Z., Fidler, S.: Get3d: A generative model of high quality 3d textured shapes learned from images. Advances in Neural Information Processing Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.787, + 0.731 + ], + "angle": 0, + "content": "8. Guo, Y.C., Liu, Y.T., Shao, R., Laforte, C., Voleti, V., Luo, G., Chen, C.H., Zou, Z.X., Wang, C., Cao, Y.P., Zhang, S.H.: threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.787, + 0.758 + ], + "angle": 0, + "content": "9. Gupta, A., Xiong, W., Nie, Y., Jones, I., Oğuz, B.: 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.786, + 0.799 + ], + "angle": 0, + "content": "10. Henzler, P., Mitra, N.J., Ritschel, T.: Escaping plato's cave: 3d shape from adversarial rendering. In: The IEEE International Conference on Computer Vision (ICCV) (October 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "11. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.372, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "K. Xie et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "12. Hunter, J.D.: Matplotlib: A 2D graphics environment. Computing in Science & Engineering (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "13. Jun, H., Nichol, A.: Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.785, + 0.242 + ], + "angle": 0, + "content": "14. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics 39(6) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.785, + 0.284 + ], + "angle": 0, + "content": "15. Li, J., Tan, H., Zhang, K., Xu, Z., Luan, F., Xu, Y., Hong, Y., Sunkavalli, K., Shakhnarovich, G., Bi, S.: Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.285, + 0.785, + 0.311 + ], + "angle": 0, + "content": "16. Li, M., Zhou, P., Liu, J.W., Keppo, J., Lin, M., Yan, S., Xu, X.: Instant3d: Instant text-to-3d generation. arXiv preprint arXiv:2311.08403 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.311, + 0.785, + 0.351 + ], + "angle": 0, + "content": "17. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.352, + 0.785, + 0.392 + ], + "angle": 0, + "content": "18. Liu, M., Xu, C., Jin, H., Chen, L., Xu, Z., Su, H., et al.: One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.393, + 0.785, + 0.419 + ], + "angle": 0, + "content": "19. Liu, R., Wu, R., Hoorick, B.V., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.42, + 0.785, + 0.46 + ], + "angle": 0, + "content": "20. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. In: Seminal graphics: pioneering efforts that shaped the field, pp. 347-353 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.461, + 0.785, + 0.5 + ], + "angle": 0, + "content": "21. Lorraine, J., Xie, K., Zeng, X., Lin, C.H., Takikawa, T., Sharp, N., Lin, T.Y., Liu, M.Y., Fidler, S., Lucas, J.: Att3d: Amortized text-to-3d object synthesis. In: Proceedings of the International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.501, + 0.785, + 0.541 + ], + "angle": 0, + "content": "22. Lunz, S., Li, Y., Fitzgibbon, A., Kushner, N.: Inverse graphics gan: Learning to generate 3d shapes from unstructured 2d data. arXiv preprint arXiv:2002.12674 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.542, + 0.785, + 0.568 + ], + "angle": 0, + "content": "23. Luo, T., Rockwell, C., Lee, H., Johnson, J.: Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.569, + 0.785, + 0.622 + ], + "angle": 0, + "content": "24. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4460-4470 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.623, + 0.785, + 0.663 + ], + "angle": 0, + "content": "25. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latentnerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.664, + 0.785, + 0.704 + ], + "angle": 0, + "content": "26. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.705, + 0.785, + 0.745 + ], + "angle": 0, + "content": "27. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: CVPR. pp. 8280-8290 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.746, + 0.785, + 0.772 + ], + "angle": 0, + "content": "28. Nash, C., Ganin, Y., Eslami, S.M.A., Battaglia, P.W.: Polygen: An autoregressive generative model of 3d meshes (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.773, + 0.785, + 0.812 + ], + "angle": 0, + "content": "29. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "30. Oliphant, T.E.: Python for scientific computing. Computing in Science & Engineering (2007)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LATTE3D" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.175 + ], + "angle": 0, + "content": "31. OpenAI: Gpt-4 technical report. ArXiv abs/2303.08774 (2023), https://api.sementicscholar.org/CorpusID:257532815" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.217 + ], + "angle": 0, + "content": "32. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in PyTorch. Open-review (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "33. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.285 + ], + "angle": 0, + "content": "34. Qian, G., Cao, J., Siarohin, A., Kant, Y., Wang, C., Vasilkovsky, M., Lee, H.Y., Fang, Y., Skorokhodov, I., Zhuang, P., et al.: Atom: Amortized text-to-mesh using 2d diffusion. arXiv preprint arXiv:2402.00867 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.342 + ], + "angle": 0, + "content": "35. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "36. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.437 + ], + "angle": 0, + "content": "37. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.493 + ], + "angle": 0, + "content": "38. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.535 + ], + "angle": 0, + "content": "39. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.59 + ], + "angle": 0, + "content": "40. Shen, T., Munkberg, J., Hasselgren, J., Yin, K., Wang, Z., Chen, W., Gojcic, Z., Fidler, S., Sharp, N., Gao, J.: Flexible isosurface extraction for gradient-based mesh optimization. ACM Trans. Graph. 42(4) (jul 2023). https://doi.org/10.1145/3592430, https://doi.org/10.1145/3592430" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.619 + ], + "angle": 0, + "content": "41. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "42. Shi, Y., Wang, P., Ye, J., Mai, L., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv:2308.16512 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "43. Tang, J., Chen, Z., Chen, X., Wang, T., Zeng, G., Liu, Z.: Lgm: Large multiview gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.73 + ], + "angle": 0, + "content": "44. Tatarchenko, M., Dosovitskiy, A., Brox, T.: Octree generating networks: Efficient convolutional architectures for high-resolution 3d outputs. In: Proceedings of the IEEE international conference on computer vision. pp. 2088-2096 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.757 + ], + "angle": 0, + "content": "45. Van Rossum, G., Drake Jr, F.L.: Python reference manual. Centrum voor Wiskunde en Informatica Amsterdam (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.798 + ], + "angle": 0, + "content": "46. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "47. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "K. Xie et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "48. Wu, J., Zhang, C., Xue, T., Freeman, B., Tenenbaum, J.: Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems 29 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.231 + ], + "angle": 0, + "content": "49. Yang, G., Huang, X., Hao, Z., Liu, M.Y., Belongie, S., Hariharan, B.: Pointflow: 3d point cloud generation with continuous normalizing flows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.259 + ], + "angle": 0, + "content": "50. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "51. Zeng, X., Vahdat, A., Williams, F., Gojcic, Z., Litany, O., Fidler, S., Kreis, K.: Lion: Latent point diffusion models for 3d shape generation. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.327 + ], + "angle": 0, + "content": "52. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "53. Zhou, L., Du, Y., Wu, J.: 3d shape generation and completion through point-voxel diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5826-5835 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.37 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/LATTE3D_ Large-scale Amortized Text-To-Enhanced3D Synthesis/a662796f-9e3f-4f91-946c-d99cfdaedd9c_origin.pdf b/2024/LATTE3D_ Large-scale Amortized Text-To-Enhanced3D Synthesis/a662796f-9e3f-4f91-946c-d99cfdaedd9c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..296051b85d82c77c92ae0b50e62cf236c8671e88 --- /dev/null +++ b/2024/LATTE3D_ Large-scale Amortized Text-To-Enhanced3D Synthesis/a662796f-9e3f-4f91-946c-d99cfdaedd9c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7e232bf3c3d481650b57003b59bea2aa56c0dd51cd5cc0b4ac6416888b611f4 +size 5081597 diff --git a/2024/LATTE3D_ Large-scale Amortized Text-To-Enhanced3D Synthesis/full.md b/2024/LATTE3D_ Large-scale Amortized Text-To-Enhanced3D Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..167234eff75e4d06106815d1b1c6ee647817d18a --- /dev/null +++ b/2024/LATTE3D_ Large-scale Amortized Text-To-Enhanced3D Synthesis/full.md @@ -0,0 +1,311 @@ +# LATTE3D: Large-scale Amortized Text-To-Enhanced3D Synthesis + +Kevin Xie $^{1,2}$ , Jonathan Lorraine $^{1,2}$ , Tianshi Cao $^{1,2}$ , Jun Gao $^{1,2}$ , James Lucas $^{1}$ , Antonio Torralba $^{1,3}$ , Sanja Fidler $^{1,2}$ , and Xiaohui Zeng $^{1,2}$ + +$^{1}$ Nvidia, $^{2}$ University of Toronto and Vector Institute, $^{3}$ MIT + +![](images/044750169467ef79cf0e8c6281f0f18c5e9476fe02b78c8af1c71a3cf4af1dde.jpg) +Fig.1: Samples generated in $\sim 400\mathrm{ms}$ on a single A6000 GPU from text prompts. Objects without prompt labels are generated by our text-to-3D model trained on $\sim 100\mathrm{k}$ prompts, while labeled objects are generated by our 3D stylization model trained on $12\mathrm{k}$ prompts. See our project website** for more. + +Abstract. Recent text-to-3D generation approaches produce impressive 3D results but require time-consuming optimization that can take up to an hour per prompt. Amortized methods like ATT3D optimize multiple prompts simultaneously to improve efficiency, enabling fast text-to-3D synthesis. However, they cannot capture high-frequency geometry and texture details and struggle to scale to large prompt sets, so they generalize poorly. We introduce LATTE3D, addressing these limitations to achieve fast, high-quality generation on a significantly larger prompt set. Key to our method is 1) building a scalable architecture and 2) leveraging 3D data during optimization through 3D-aware diffusion priors, shape regularization, and model initialization to achieve robustness to diverse and complex training prompts. LATTE3D amortizes both neural field and textured surface generation to produce highly detailed textured meshes in a single forward pass. LATTE3D generates 3D objects in 400ms, and can be further enhanced with fast test-time optimization. + +# 1 Introduction + +Recent advances in text-to-3D synthesis via image diffusion models mark significant progress in 3D content creation using [3,8,17,25,33,46,47]. However, these methods often involve an expensive and time-consuming optimization that can take up to an hour to generate a single 3D object [17, 33]. We aim to achieve real-time text-to-3D synthesis for diverse text prompts and support fast test-time optimization when a further quality boost is desired. + +Seminal work [33] proposed to optimize a neural field using the score distillation sampling (SDS) loss to generate 3D assets. It unlocked open-vocabulary 3D content creation but was (a) lacking high-frequency geometric and texture details, (b) expensive, and (c) prone to failing to generate the desired object. Follow-ups introduced a surface-based fine-tuning stage [3, 17, 47] to generate high-frequency details. While this two-stage pipeline defines most current state-of-the-art text-to-3D approaches, it remains expensive and prone to prompt failure. + +ATT3D [21] was the first to make text-to-3D generation fast by simultaneously training a single model on a set of prompts in a process called amortized optimization. Amortization reduces training time by optimizing a shared text-conditioned model on a set of prompts, enabling the model to generalize to new prompts at inference time. Despite promising results on curated prompts, ATT3D – and the underlying per-prompt methods – remain prone to failing on general prompts. Also, ATT3D only amortizes the first stage of 3D synthesis, producing a neural field representation and thus cannot generate high-frequency details. + +To address these issues, we introduce LATTE3D, a Large-scale Amortized Text-to-Enhanced3D synthesis method that can produce high-quality 3D content in real-time. Our work makes the following technical contributions to boost the quality, robustness, scale, and speed of text-to-3D generative models: + +- We propose a novel text-to-3D amortization architecture that can scale to orders of magnitude larger prompt sets. +- We leverage 3D data in training to improve quality and robustness, through 3D-aware diffusion priors, regularization loss, and weight initialization through pretraining with 3D reconstruction. +- We amortize the surface-based refinement stage, greatly boosting quality. + +# 2 Related Work + +Early work on 3D object synthesis was typically limited to generating objects from a given class [1,2,7,10,22,24,28,44,48,49,51,53], e.g. cars or chairs. Recent extensions use captioned 3D shapes to train text-to-3D models [13,29] generating shapes from diverse categories, but requiring 3D supervision, restricting them to synthetic datasets of limited size, diversity, and visual fidelity. + +The advent of differentiable rendering, both volumetric [26] and surface-based [27], opened the door to inverse image rendering [7, 33], unlocking the use of powerful text-to-image generative models [36-38] in the 3D synthesis process. DreamFusion [33] proposed the SDS loss to optimize a neural field using a text-to-image diffusion model to generate 3D assets. Follow-up work introduced a surface-based refinement stage [17], allowing the synthesis of high-frequency details. This two-stage optimization pipeline now defines most of the state-of-the-art text-to-3D approaches [3,25,46,47]. The recently released large-scale 3D dataset Objaverse-XL [5,6] has spurred researchers to explore the benefits of 3D supervision in the synthesis process. To date, the most successful approaches make + +![](images/18147faf535aeb209090b21e179d8fe9a152c5d764a147c20180725785dbf416.jpg) +Fig.2: Comparison of SOTA text-to-3D methods on unseen prompts. We plot different methods' user study preference rates compared to LATTE3D. For MVDream, we report varying optimization times. + +text-conditioned image diffusion models 3D aware [15,19,41,42] by fine-tuning 2D diffusion models using rendered multiview images. SDS from multiview diffusion models is then used to obtain 3D shapes [19,41,42]. + +However, these approaches require a lengthy, per-prompt optimization process, hindering applicability to real-world content creation. To address this, two lines of work have been pursued. The first type of method uses a text-to-image-to-3D approach that generates images with a text-to-image diffusion model and trains an image-to-3D lifting network [15, 18, 41, 43]. They offer improved speed but with limited quality details compared to optimization methods, but still take $5 - 20$ seconds per prompt [15, 43] for sampling from the image diffusion model. + +In an alternate line of work, ATT3D [21] proposed an amortized framework that optimizes multiple prompts simultaneously with a unified model. Follow-up works expanded the network architecture to achieve better quality [16, 34]. LATTE3D extends this line of work to larger scales by incorporating 3D knowledge to achieve prompt robustness during training, among other design improvements like better architecture. Amortized text-to-3D offers an attractive quality vs. speed trade-off as it uses a single feed-forward architecture without requiring sampling of diffusion models - see Fig. 2. + +Concurrent Works Concurrently, AToM [34] also amortizes two-stage training with a triplane representation, but they do not scale beyond the smaller sets of prompts from ATT3D. In contrast, we scale to the order of 100000 prompts, using 3D-aware techniques. Another concurrent work of ET3D [4] achieves fast generation by training a GAN model, but it is only trained on a small-scale compositional dataset of 5000 prompts and does not guarantee view consistency as it uses a 2D neural upsampler. Lastly, LGM [43] is a text-to-image-to-3D method that generates Gaussian splats in 5 seconds. We experimentally demonstrate competitive performance to it. + +# 3 Methodology + +An established pipeline for high-quality text-to-3D generation consists of two stages, each performing per-prompt optimization using the SDS loss with a text-to-image diffusion model [3, 17, 47]. Stage-1 optimizes a volumetric representation, which is typically a neural radiance field. Meshes afford real-time rendering but are hard to optimize from scratch with image supervision alone. Stage-2 uses the output of the neural field to initialize a signed distance field (SDF) and a texture field, from which a mesh can be derived using differentiable isosurfacing [39, 40]. This surface-based representation is then optimized via differentiable rasterization [27], which is fast even at 1024 rendered image resolution – enabling supervision from higher-resolution text-to-image diffusion models. + +ATT3D [21] amortizes optimization of a neural field over a set of prompts with a hypernetwork mapping a text prompt to a neural field, which is trained with the SDS loss over a prompt set, referred to as seen prompts. LATTE3D introduces a new architecture that amortizes both stages of the generation process, aiming to produce high-quality textured meshes in real-time. We scale the approach to a magnitudes larger and more diverse set of prompts by leveraging 3D knowledge in the amortization loop. + +Our approach is illustrated in Fig. 3. We initialize LATTE3D with a reconstruction pretraining step to stabilize training (Sec. 3.1). The architecture consists of two networks, one for geometry and the other for texture (Sec. 3.2). Amortized model training with diffusion priors is done through a two-stage pipeline consisting of a volume-based stage-1 and a surface-based stage-2 (Sec. 3.3). To reconcile reconstruction pretraining, which uses 3D geometry as input, with the task of text-to-3D generation, we anneal the network's input during stage-1 to gradually fully replace the dependence on 3D input with a single dummy input. During inference (Sec. 3.4), our model generates a 3D textured mesh from just a text prompt in 400ms and allows an optional lightweight test-time refinement to enhance the quality of geometry and textures (Sec. 3.5). + +# 3.1 Pretraining to reconstruct shapes. + +Although SDS can synthesize text-aligned 3D features, empirically, it is a less efficient source of supervision due to high variance. We find that pretraining the model to first encode-decode 3D shapes subsequently accelerates the early stage of amortized SDS training. We initialize our model $\mathcal{M}$ by pretraining it on a 3D asset dataset using image reconstruction losses, similar to the reconstruction stage used in [9]. $\mathcal{M}$ inputs a sampled point cloud and outputs a predicted 3D shape. We then render multiple views of the predicted 3D shape $o$ to compare with the input shape $s$ renderings. We use an $\ell_2$ loss on the rendered opacity and RGB image: + +$$ +\mathcal {L} _ {\text {r e c o n}} (o, s, c) = \left| \left| \mathbf {R} _ {\text {o p a c i t y}} (o, c) - \mathbf {R} _ {\text {o p a c i t y}} (s, c) \right| \right| _ {2} + \left| \left| \mathbf {R} _ {\text {R G B}} (o, c) - \mathbf {R} _ {\text {R G B}} (s, c) \right| \right| _ {2} \tag {1} +$$ + +We denote the renderer $\mathbf{R}$ , the opacity and RGB rendering with $\mathbf{R}_{\mathrm{opacity}}$ and $\mathbf{R}_{\mathrm{RGB}}$ , respectively, and the camera with $c$ , randomly sampled on a sphere. + +![](images/751c8154c5442a8605be520b5c8a401f36b0b3c2f735b70b500fe1c0bdb22ee7.jpg) +Fig. 3: LATTE3D consists of two networks: a texture network $T$ and geometry network $G$ . When amortizing the first stage, the encoders of both networks share the same set of weights. The training objective includes an SDS gradient from a 3D-aware image prior and a regularization loss that compares the rendered predicted shape's mask with the rendered masks of 3D assets in a library. When amortizing surface-based refinement in stage-2, we freeze the geometry network $G$ and update the texture network $T$ . + +# 3.2 Model Architecture + +Our model $\mathcal{M}$ consists of two networks, $G$ and $T$ , for predicting geometry and texture, respectively. After pretraining, we introduce text-conditioning by adding cross-attention layers. We use CLIP [35] embeddings for encoding the text prompt $p$ . In stage-1 training, we tie the geometry and texture networks' weights (except for separate MLP decoder heads), effectively training a single input encoder. In stage-2, we freeze the geometry network $G$ and refine the texture network $T$ with an additional trainable upsampling layer. The final model $\mathcal{M}$ used in inference is the frozen $G$ from stage-1 and a refined $T$ from stage-2. + +The geometry network $G$ consists of a U-Net encoder on a triplane [2,7], followed by a decoder containing another triplane U-Net and a neural volumetric density field. Specifically, a point cloud is fed through a PointNet, then converted to a triplane via average scattering. During training, the point cloud comes from the 3D data, and in inference, a dummy sphere point cloud is used. While this point cloud encoder is inherited from reconstruction pretraining, it becomes unimportant in the final model. The triplane is processed through the encoder and decoder U-Nets. The text embedding is fed into every residual U-Net block in the encoder via cross-attention. For any point in 3D space, we extract triplane features from each plane via bilinear interpolation. An MLP then predicts the density value at that point. + +The texture network $T$ and geometry network $G$ share the same encoder and decoder U-Net architecture in pretraining and stage-1. We employ another non-shared MLP neural texture field to predict RGB values at query points. In stage-2, we upsample the triplane features produced by the encoder to a higher resolution to gain extra capacity to generate high-frequency texture details. Specifically, we first bilinearly upsample the latent triplane from 128 resolution to + +![](images/8c8399b526d550f0d7c1a5bb92d1daa5b789d24b9b914ca97293aad6992bd54c.jpg) +Fig. 4: Stylization application: Our model learns to generate diverse stylizations of the same shapes. Left column shows the original shape to be stylized. + +480 and add a residual MLP which maps the text-embedding to the residual of the triplane feature, mimicking ATT3D's mapping (hyper)network. The decoder then takes this upsampled triplane as input. Details are in the Appendix. + +Rendering. To train the model, we render the generated 3D object into 2D images with different methods in stage-1 and 2. In stage-1, the output 3D geometry is a triplane-based neural field, on which we use volume rendering to get images of 256 resolution using the formulation from VolSDF [50], allowing us to parameterize the density via an approximate signed distance field whose zero-level set defines the object geometry. This choice allows a simple conversion to surface-based rendering. For the geometry network, we observe more stable training with volume rendering with SDS compared to surface rendering. In stage-2, we extract the isosurface from the density field with Marching Cubes [20] and render images via rasterization [14]. Meshes can be rendered at a 1024 image resolution in real-time, helping capture finer details for learning textures. + +# 3.3 Amortized Learning + +Amortized learning distills knowledge from image diffusion priors into our model. We perform amortized training on volumetric geometry training (stage-1) and surface texture training (stage-2) sequentially. We incorporate 3D information into our training process (a) implicitly via 3D-aware 2D SDS losses and (b) explicitly with regularization. + +Curating prompt sets. We create large prompt sets for training using rule-based text generation or ChatGPT [31]. For the former, we start with the categories names from the captions of Objaverse [23], and design rules, like "object A in style B is doing C" with a set of styles and activities. Alternatively, we input the captions into ChatGPT and ask it for detailed and diverse prompts describing a similar 3D object. Fig. 3 shows that when generating "Amigurumi fiddler crab + +wearing a top hat", we guide shape synthesis towards resembling a crab by retrieving a related 3D shape used in a regularization loss. Other details of dataset construction are in Sec. 4.1. + +Amortizing Stage-1 Generation In this stage, we train $\mathcal{M}$ with our datasets. We use the SDS loss with a 3D-aware 2D prior and regularization to the paired 3D shape, which we now explain in detail: + +3D-aware SDS loss. During stage-1 training we leverage a 3D aware diffusion prior, which provides a stronger, multiview consistent supervisory signal. Specifically, we use the model from MVDream [42], which was trained on four rendered views of objects from the Objaverse dataset [6] by modifying Stable Diffusion [37] to generate multiview consistent images. + +Regularization loss. While 3D-aware SDS loss with MVDream offers a strong multiview prior, it can still fail to correctly capture the full 3D geometry for some prompts since MVDream only supervises four views and can create enlarged geometry or floaters in the shape. Some failures are shown on the right of Fig. 10. + +Combining regularization using 3D shapes with the SDS loss helps training and improves geometry. As shown in Fig. 3, for each training prompt, we regularize the output shape $o$ by comparing its rendered mask to the mask of a shape $s$ retrieved from the 3D dataset with the prompt: + +$$ +\mathcal {L} _ {\text {r e g}} (o, s, c) = \left| \left| \mathbf {R} _ {\text {o p a c i t y}} (o, c) - \mathbf {R} _ {\text {o p a c i t y}} (s, c) \right| \right| _ {2}, \tag {2} +$$ + +where $\mathbf{R}_{\mathrm{opacity}}$ is the opacity mask (volume-rendered density). We balance the regularization $\mathcal{L}_{\mathrm{reg}}$ and the SDS loss $\mathcal{L}_{\mathrm{SDS}}$ by simply blending them using a convex combination with weight $\alpha$ . Our final training loss is: + +$$ +\mathcal {L} _ {\text {t r a i n}} = (1 - \alpha) \mathcal {L} _ {\text {S D S}} + \alpha \mathcal {L} _ {\text {r e g}}. \tag {3} +$$ + +During training, we render five views for each generated shape. We average across four views to compute $\mathcal{L}_{\mathrm{SDS}}$ , and one is used to compute $\mathcal{L}_{\mathrm{reg}}$ . + +Input point cloud annealing To bridge the gap between the real point cloud inputs in training and dummy point cloud in inference, we gradually anneal the input point clouds towards dummy point clouds during stage-1 training. Specifically, we randomly replace a subset of points in the input point cloud with points from the dummy point cloud using a probability that increases linearly from 0 to 1 in the last 5000 training iterations. With point cloud annealing, the model gets better results in inference when the dummy point cloud is used. + +Amortizing Stage-2 Generation During stage-2 training, we freeze the network for our geometry and only tune the texture since training both leads to instabilities. We use depth-conditional ControlNet [52] for SDS guidance, allowing a higher, 512 resolution supervision. Depth conditioning encourages the SDS loss to guide the texture to align with the geometry from the geometry network, thereby improving the 3D texture consistency. + +# 3.4 Inference + +During inference, our model inputs the user's text prompt and a dummy point cloud. Our $\mathcal{M} = (G,T)$ outputs a textured mesh, where $T$ is the final texture network refined in stage-2, while $G$ is trained in stage-1 only and then frozen. Predicting one shape and rendering the image only takes us 400ms on an A6000 GPU in inference, allowing the generation of user results at interactive speeds. A single A6000 can generate 4 samples simultaneously. Once the user selects a desired sample, textured meshes can be exported from LATTE3D. We use a UV mapping tool to UV parameterize the mesh output by $G$ and use $T$ to predict the albedo colors of each texel in a UV-mapped texture image. The resulting mesh and texture image is compatible with standard rendering software. + +# 3.5 Test Time Optimization + +Our method supports test time optimization if the user wants to boost the quality of a particular prompt. This can be useful if the user's prompt deviates significantly from the seen prompts. As in our stage-2 training, we freeze the geometry and finetune the texture network $T$ with depth-conditional SDS guidance. Our test time optimization is significantly faster to converge than MVDream [42] or other optimization-based text-to-3D methods [17, 33, 47] since we optimize our amortized network instead of a randomly initialized network. + +# 3.6 3D Stylization + +LATTE3D can also be trained and deployed as a 3D stylization method, allowing users to cheaply create variations from existing 3D assets. During training, we skip the input point cloud annealing step to maintain point cloud reliance and amortize over a dataset where each 3D shape corresponds to many style prompts. Our blend of regularization and SDS guidance drives the model to produce shapes structurally similar to the original shapes yet semantically aligned with the style prompt. Amortization significantly reduces the computational cost per shape-prompt pair compared to optimizing each pair separately. During inference, point clouds from the training set can be combined with novel prompts as input to the model to produce variations of the original shape. + +# 4 Experiments + +We first describe our experimental setup, including datasets, metrics, and baselines for comparison (Sec. 4.1). We then present our quantitative and qualitative results (Sec. 4.2), demonstrating the performance and generalization ability of LATTE3D (Sec. 4.3), followed by design choices ablations (Sec. 4.4). + +# 4.1 Experimental Setup + +We now discuss dataset construction, model pretraining, and evaluation metrics. + +![](images/83912decb4981a65caef63b0e3bb9d6c01c4b7d6c55e50d0f2bff8f0e866a083.jpg) +MVDream (6min) + +![](images/79f3f08df33e0a26d89e02401a439b48e10a4c9ff0bc2b96eb1d77250f7c6504.jpg) +MVDream (36min) + +![](images/b6733d82b10c55e37379f886c89930e565222c7cae08de120e8526350394ccd1.jpg) +3DTopia (20s) +Fig. 5: Qualitative comparison between LATTE3D and ATT3D [21], MVDream [42], Instant3D/3DTopia [15] and LGM [43]. Both rows are unseen prompts: "a red convertible car with the top down", "an ice cream sundae". + +![](images/cb1fd2c0efb16cfc2c02bdde6ea2ec5ba30c028ff17e788384425338c6c59069.jpg) +LGM (5s) + +![](images/9c837622280e3a60d09351465007e50fa6e19f1622a89872d69f8ef07dbe7be8.jpg) +ATT3D + +![](images/a0399853b212ddab43af57880eb9a0cca5743bb612120d65bbd8f135de172f06.jpg) +ATT3D+ UNet $^+$ MV + +![](images/d4aa217c847d3243099139360b25baa76411ceb3b5398c4c96dd67d6e86c7412.jpg) +LATTE3D (400ms) + +Training Set To investigate scaling amortized optimization beyond what has been done in prior works - see Tab. 1 - we construct a new dataset gpt-101k that consists of 101k text prompts and 34k shapes. For the shapes, we use the "lvis" subset from Objaverse [6] with 50k human-verified shapes. We filter out poor shapes, e.g. flat images and scans, and retain 34k + +shapes for our dataset. We use ChatGPT to augment each object's caption to gather 3 prompts per object. + +Table 1: Comparing the sizes of amortization training sets to prior works. + +
name#promptsSource
Animal2400 [21,34]2400Rule
DF415 [34]415DF [33]
animal-style (Ours)12000Rule
gpt-101k (Ours)101608GPT
+ +Evaluation Sets We use two prompt sets for evaluation. To evaluate the benefit of amortized optimization in our method, we evaluate our model on gpt-101k. However, due to the computational cost of running the baselines, we randomly select a 50 prompt subset in gpt-101k to form the seen prompt set for benchmarking. Furthermore, to gauge the generalization abilities of our amortized model, we construct an unseen prompt set by filtering a subset of 67 DreamFusion [33] prompts that are close in distribution to lvis categories. + +Baselines. We compare LATTE3D with the baselines: ATT3D [21], MVDream [42], Instant3D [15] and LGM [43]. We re-implement ATT3D [21] with a hypernetwork. We use the open-source threestudio [8] implementation of MVDream. For MVDream, we train each 3D model for 10000 iterations, with a batch size of 8 on a single A100 GPU with provided default settings. For Instant3D, since the source code has not been released, we use the re-implementation from 3DTopia1. For LGM, we use the official open-source implementation. + +Evaluation metrics. We quantitatively compare our model with baselines by evaluating the generated 3D content fidelity and its consistency to the text prompt for the optimization cost. We use three metrics to evaluate fidelity. + +Table 2: Quantitative metrics and average user preference (\%) of baselines over LATTE3D trained on gpt-101k using seen and unseen prompts. We also report test-time optimization, which takes $10\mathrm{min}$ . + +
ModelTimeRender-FID ↓ seenCLIP-Score ↑ unseen(df)Preference % ↑seenUnseen(df)seenUnseen(df)
LATTE3D400ms180.58190.000.26080.2605--
LATTE3D-opt10min171.69178.370.28000.3000--
MVDream6min203.53199.230.22400.26982.512.4
MVDream18min191.77188.840.24400.295026.552.2
MVDream36min158.11143.440.28300.318048.589.3
3DTopia20s217.29222.760.18370.217120.825.6
LGM5s178.67166.400.28400.306326.236.4
+ +Render-FID: We compute the FID [11] between the renderings of generated 3D content and a set of images sampled from Stable Diffusion [37] with the same text prompts as input. This metric measures how well the generated shapes align with those from the 2D prior in visual quality. + +CLIP Score: We compute the average CLIP scores between the text prompt and each rendered image to indicate how well the generated 3D shape aligns with the input text prompt. + +User Study: We evaluate the overall 3D geometry and texture through Amazon Mechanical Turk user studies. We present videos of 3D shapes generated by our method and a baseline head-to-head, with corresponding text prompts, and ask users for their preference. Each comparison is averaged over three users to obtain preference rate for each baseline relative to LATTE3D. + +Timing: We measure our optimization cost by GPU time per prompt to gauge the total compute cost on each prompt. We divide the total wall-clock time for pretraining, stage-1, and stage-2 by the prompt set's size, then multiply by the number of GPUs used. We measure inference time for Instant3D, LGM, and LATTE3D by the time the model takes from in-taking the text input to outputting the final features before rendering. For MVDream [42], we measure the average time to optimize a single prompt on an A100 GPU. + +# 4.2 Experimental Results + +We observe robust generalization of LATTE3D by training on the gpt-101k dataset to unseen prompts (df, from DreamFusion [33]), as demonstrated by qualitative (Fig. 5) and quantitative (Fig. 2, Tab. 2) results. We highlight the relative user preference of LATTE3D versus different baselines over inference time in Fig. 2. LATTE3D produce results of competitive user preference to SOTA baselines at a uniquely fast inference time. + +From the quantitative comparisons in Tab. 2, our performance on both seen and unseen prompts is competitive with baselines while our inference time is + +![](images/f4d2dd9e976d1cb01e70ff4f923565d90cfe19bb0c86bd337b931a6867995c01.jpg) +Fig.6: Results of user study showing the average preference of MVDream at different amounts of optimization time to LATTE3D gpt-101k on seen prompts. + +![](images/4c5a66890e75a1005ddeca924f8c637893c2fcbe09b0b69a7595bdd8044ac056.jpg) +Fig.7: Test time optimization on seen and unseen prompt. The model is trained on gpt-101k. First column: without test time optimization. Other columns: test time optimization with the given time budget denoted. Top: seen prompt "...animal with a blue and white floral pattern..." Bottom: unseen prompt "orchid in clay pot". + +an order of magnitude lower. Image-based metrics of FID and CLIP score may fail to detect generations with 3D geometry errors, less obvious in still images than in the user study, where users are shown videos of objects rotating through multiple views. Interested readers can refer to rendered videos of all methods in the supplementary materials. + +Total optimization cost A benefit of amortized optimization over many prompts is that our total optimization cost on gpt-101k is much lower than optimizing MVDream per-prompt. Fig. 6 illustrates that to achieve user preference on par with LATTE3D, MVDream would require 36 GPU minutes per prompt, whereas LATTE3D spent only 215 GPU seconds per prompt, representing a $10 \times$ reduction in optimization cost. + +# 4.3 Application + +We illustrate LATTE3D's capacity for enhanced quality through test time optimization and how to adapt LATTE3D for stylizing 3D content. + +Per Prompt Test-time Optimization We show experimentally that LATTE3D quickly adapts to new prompts during test time to improve quality. In Fig. 7, we qualitatively compare LATTE3D with up to 600 steps. Each optimization iteration takes an average of 0.98 seconds on an A100 GPU, giving us a total compute budget of less than 10 GPU minutes per prompt. Even with test-time optimization, our method is still an order of magnitude faster than per-prompt optimization methods such as MVDream, as in Tab. 2. Test-time optimization is + +particularly beneficial on unseen prompts, where FID drops by 11.6 for unseen vs. 8.8 for seen, and the CLIP score gains 0.04 for unseen vs. 0.02 for seen. + +![](images/174459b7cdbcf8a9bbaa5b7529de9f97e75ffa73e2618f419c205ba645881d92.jpg) +Fig. 8: Results of user study showing the average preference rate for MVDream at different amounts of optimization time to LatTE3D stylization results on animal-style. + +Stylizing 3D content We illustrate how LATTE3D architecture offers us the flexibility to adapt it for targeted 3D content stylization as described in Sec. 3.6. To test this, we manually curate a dataset of 100 animal shapes from Objaverse as a basis and augment the species name of each animal shape name with combinations of activities (e.g. "riding a skateboard") and styles (e.g. "voxel style") to create our 12000 prompts animal-style dataset. We train a LATTE3D model (stage-1 and stage-2) on animal-style without the point cloud annealing stage. The resulting model adapts to both the text prompt and the point cloud inputs. In Fig. 4, each row contains the outputs of LATTE3D given different prompts while fixing the point cloud input, demonstrating the effectiveness of LATTE3D for text-controlled stylization. + +In Fig. 8, we conduct a user study comparing LATTE3D's outputs on animal-style to MVDream and find that LATTE3D is competitive with MVDream with 4000 steps of optimization, representing a $10 \times$ reduction in optimization cost (compared to running MVDream for each prompt). LATTE3D on animal-style also generalizes to held-out combinations of animal activities and styles with little quality drop, as indicated by its similar relative preference to MVDream. Therefore, LATTE3D enables users to cheaply create variations of their 3D assets via amortized optimization and even explore novel prompt combinations at test time. Additional quantitative and qualitative results can be found in the supplementary materials. + +Table 3: Ablation of components in stage-1 training. Trained on gpt-101k data and evaluated on seen and unseen prompts. Preference indicate average user preference of baseline over LATTE3D. LATTE3D is better than all ablated settings in quantitative metrics and is preferred by users on the unseen prompt set. + +
ExpMV Unet Pretrain RegSeenUnseen
Mask-FID↓Render-FID↓Clip-Score↑Preference
ATT3D274.44275.010.209128.324.2
+MV✓243.24236.730.182634.826.9
+MV+UNet✓✓184.80203.330.221448.045.6
+MV+UNet+PT✓✓✓189.09189.510.219151.847.7
LATTE3D (S1)✓✓✓✓176.44186.840.2379--
+ +# 4.4 Ablation Studies + +We now analyze the value of each component we introduced in LATTE3D. We include additional results, including ablations on different diffusion priors, in the Appendix. + +Stage-1 ablation. Tab. 3 summarizes our quantitative comparison of the following ablations: $ATT3D$ baseline using a hypernet, $ATT3D + MV$ uses MVDream as diffusion guidance. $ATT3D + UNet + MV$ replaces the hypernet with our model architecture, but is initialized from scratch and given a dummy point cloud as input. $ATT3D + UNet + MV + PT$ is initialized from the reconstruction pretrained (PT) model. We introduce the Mask-FID metric to quantitatively assess geometry adherence to the shape dataset by computing the FID between the rendered binary masks of the generated shapes and those of the 3D dataset. + +We find that each of the components we introduce improves performance. Comparing ATT3D with ATT3D+MV, we see benefits from using the MV-Dream guidance as it mitigates geometric artifacts like Janus faces. Comparing ATT3D+MV with ATT3D+MV+UNet we see a performance boost from our architecture. Further, adding pretraining (ATT3D+UNet+MV+PT) helps, especially in recovering fine geometric and texture details. Finally, adding shape regularization (full LATTE3D) makes the generated geometry adhere better to the 3D data, as evidenced by the drop in Mask-FID. An additional user study is shown in Tab. 3. We show the convergence speed in the Appendix. + +Ablation on using stage-2 refinement. Stage-2 refinement gives large improvements in texture details over stage-1 training. We show the comparison in Fig. 9. + +Regularization We now investigate the design choice of how to weight $\alpha$ between the SDS loss and the regularization loss as in Eq. 3. In Tab. 3, we show that adding regularization loss improves the quantitative results. We qualitatively contrast training on different fixed weightings in Fig. 10 and the Appendix. We see greater consistency with the input shape by increasing shape regularization with higher blend factors $\alpha$ . + +![](images/09e8636397efaa3277105f9cb0f25a00774127e50592d968ac3e3d34635ea567.jpg) +Fig.9: Comparison of stage-1 and stage-2 results. Stage-2 refinement significantly improves the texture quality. + +"fun sheep wearing skis in pixel art style, voxel" + +![](images/d5aeebefcb872181cf9a9d91c95685eb82616f8cbed60f161a1354b0edd85845.jpg) +Fig. 10: Comparison of the impact of blending factor $\alpha$ , where $(\alpha = 0)$ is a failure from no regularization. + +Table 4: An ablation of unseen DreamFusion inputs in inference and annealing in training. Users prefer the annealed model with a dummy input in $51.2\%$ of cases. + +
Modelwithout annealwith anneal
InputTop-1DummyDummy
CLIP0.26280.25770.2605
FID186.46193.28190.00
+ +Annealing in training and different input point clouds in inference. In training, we take retrieved point clouds from the dataset as input and anneal the input to become a fixed dummy point cloud. In inference, we only use the dummy point cloud as input. In Tab. 4, we quantitatively ablate the role of this point cloud annealing process on the behavior of LATTE3D on gpt-101k. Our model without annealing training is somewhat sensitive to the point cloud input. There is a small performance drop when using a + +dummy as input in inference compared to a retrieved point cloud. However, the performance gap is reduced by introducing point cloud annealing so the model trains with dummy input. Qualitative examples are in the Appendix. + +# 5 Conclusion and Limitation + +We presented a scalable approach to perform amortized text-to-enhanced-3D generation. To successfully scale amortization to larger datasets, we used 3D data through (1) 3D-aware 2D SDS, (2) pretraining, and (3) 3D regularization of the amortization process. We further improve the model architecture to be more scalable. Our model generates high-quality shapes within $400\mathrm{ms}$ . Moreover, quality and generalization can be improved further via a speedy test time optimization. Our model also has limitations. First, our model uses SDS and thus relies on the understanding of the text-to-image models, which can often fail to respect fine-grained details such as part descriptions. Second, the geometry is frozen in stage-2 and test-time optimization, so the geometry flaws from stage-1 cannot be fixed. Developing amortized training for stable geometry changes in stage-2 is left for future work. + +# Acknowledgements + +We thank Matan Atzmon, Or Perel, Clement Fuji Tsang, Masha Shugrina, and her group for helpful feedback. The Python community [30,45] made underlying tools, including PyTorch [32] & matplotlib [12]. + +# Disclosure of Funding + +NVIDIA funded this work. Kevin Xie, Jonathan Lorraine, Tianshi Cao, Jun Gao, and Xiaohui Zeng had funding from student scholarships at the University of Toronto and the Vector Institute, which are not in direct support of this work. + +# References + +1. Achlioptas, P., Diamanti, O., Mitlagkas, I., Guibas, L.: Learning representations and generative models for 3d point clouds. In: International conference on machine learning. pp. 40-49. PMLR (2018) +2. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., Mello, S.D., Gallo, O., Guibas, L., Tremblay, J., Khamis, S., Karras, T., Wetzstein, G.: Efficient geometry-aware 3D generative adversarial networks. In: arXiv (2021) +3. Chen, R., Chen, Y., Jiao, N., Jia, K.: Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2023) +4. Chen, Y., Li, Z., Liu, P.: ET3D: efficient text-to-3d generation via multi-view distillation. arXiv 2311.15561 (2023) +5. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., VanderBilt, E., Kembhavi, A., Vondrick, C., Gkioxari, G., Ehsani, K., Schmidt, L., Farhadi, A.: Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects (2023) +6. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3d objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13142-13153 (June 2023) +7. Gao, J., Shen, T., Wang, Z., Chen, W., Yin, K., Li, D., Litany, O., Gojcic, Z., Fidler, S.: Get3d: A generative model of high quality 3d textured shapes learned from images. Advances in Neural Information Processing Systems (2022) +8. Guo, Y.C., Liu, Y.T., Shao, R., Laforte, C., Voleti, V., Luo, G., Chen, C.H., Zou, Z.X., Wang, C., Cao, Y.P., Zhang, S.H.: threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio (2023) +9. Gupta, A., Xiong, W., Nie, Y., Jones, I., Oğuz, B.: 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371 (2023) +10. Henzler, P., Mitra, N.J., Ritschel, T.: Escaping plato's cave: 3d shape from adversarial rendering. In: The IEEE International Conference on Computer Vision (ICCV) (October 2019) +11. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017) + +12. Hunter, J.D.: Matplotlib: A 2D graphics environment. Computing in Science & Engineering (2007) +13. Jun, H., Nichol, A.: Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023) +14. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics 39(6) (2020) +15. Li, J., Tan, H., Zhang, K., Xu, Z., Luan, F., Xu, Y., Hong, Y., Sunkavalli, K., Shakhnarovich, G., Bi, S.: Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214 (2023) +16. Li, M., Zhou, P., Liu, J.W., Keppo, J., Lin, M., Yan, S., Xu, X.: Instant3d: Instant text-to-3d generation. arXiv preprint arXiv:2311.08403 (2023) +17. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +18. Liu, M., Xu, C., Jin, H., Chen, L., Xu, Z., Su, H., et al.: One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928 (2023) +19. Liu, R., Wu, R., Hoorick, B.V., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object (2023) +20. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. In: Seminal graphics: pioneering efforts that shaped the field, pp. 347-353 (1998) +21. Lorraine, J., Xie, K., Zeng, X., Lin, C.H., Takikawa, T., Sharp, N., Lin, T.Y., Liu, M.Y., Fidler, S., Lucas, J.: Att3d: Amortized text-to-3d object synthesis. In: Proceedings of the International Conference on Computer Vision (ICCV) (2023) +22. Lunz, S., Li, Y., Fitzgibbon, A., Kushner, N.: Inverse graphics gan: Learning to generate 3d shapes from unstructured 2d data. arXiv preprint arXiv:2002.12674 (2020) +23. Luo, T., Rockwell, C., Lee, H., Johnson, J.: Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279 (2023) +24. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4460-4470 (2019) +25. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latentnerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600 (2022) +26. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +27. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: CVPR. pp. 8280-8290 (2022) +28. Nash, C., Ganin, Y., Eslami, S.M.A., Battaglia, P.W.: Polygen: An autoregressive generative model of 3d meshes (2020) +29. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751 (2022) +30. Oliphant, T.E.: Python for scientific computing. Computing in Science & Engineering (2007) + +31. OpenAI: Gpt-4 technical report. ArXiv abs/2303.08774 (2023), https://api.sementicscholar.org/CorpusID:257532815 +32. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in PyTorch. Open-review (2017) +33. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022) +34. Qian, G., Cao, J., Siarohin, A., Kant, Y., Wang, C., Vasilkovsky, M., Lee, H.Y., Fang, Y., Skorokhodov, I., Zhuang, P., et al.: Atom: Amortized text-to-mesh using 2d diffusion. arXiv preprint arXiv:2402.00867 (2024) +35. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021) +36. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021) +37. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022) +38. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) +39. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2021) +40. Shen, T., Munkberg, J., Hasselgren, J., Yin, K., Wang, Z., Chen, W., Gojcic, Z., Fidler, S., Sharp, N., Gao, J.: Flexible isosurface extraction for gradient-based mesh optimization. ACM Trans. Graph. 42(4) (jul 2023). https://doi.org/10.1145/3592430, https://doi.org/10.1145/3592430 +41. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model (2023) +42. Shi, Y., Wang, P., Ye, J., Mai, L., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv:2308.16512 (2023) +43. Tang, J., Chen, Z., Chen, X., Wang, T., Zeng, G., Liu, Z.: Lgm: Large multiview gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054 (2024) +44. Tatarchenko, M., Dosovitskiy, A., Brox, T.: Octree generating networks: Efficient convolutional architectures for high-resolution 3d outputs. In: Proceedings of the IEEE international conference on computer vision. pp. 2088-2096 (2017) +45. Van Rossum, G., Drake Jr, F.L.: Python reference manual. Centrum voor Wiskunde en Informatica Amsterdam (1995) +46. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774 (2022) +47. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213 (2023) + +48. Wu, J., Zhang, C., Xue, T., Freeman, B., Tenenbaum, J.: Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems 29 (2016) +49. Yang, G., Huang, X., Hao, Z., Liu, M.Y., Belongie, S., Hariharan, B.: Pointflow: 3d point cloud generation with continuous normalizing flows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2019) +50. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34 (2021) +51. Zeng, X., Vahdat, A., Williams, F., Gojcic, Z., Litany, O., Fidler, S., Kreis, K.: Lion: Latent point diffusion models for 3d shape generation. In: Advances in Neural Information Processing Systems (NeurIPS) (2022) +52. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models (2023) +53. Zhou, L., Du, Y., Wu, J.: 3d shape generation and completion through point-voxel diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5826-5835 (2021) \ No newline at end of file diff --git a/2024/LATTE3D_ Large-scale Amortized Text-To-Enhanced3D Synthesis/images.zip b/2024/LATTE3D_ Large-scale Amortized Text-To-Enhanced3D Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..484f6f3c9ad193c760bd2dad79406610c7e3855a --- /dev/null +++ b/2024/LATTE3D_ Large-scale Amortized Text-To-Enhanced3D Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:694a1611537b9a7cd6fec4a3ee1678067c6025f20f249a8ee191e71627d9e60c +size 469825 diff --git a/2024/LATTE3D_ Large-scale Amortized Text-To-Enhanced3D Synthesis/layout.json b/2024/LATTE3D_ Large-scale Amortized Text-To-Enhanced3D Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9a3c5ea366291df6c0eb2459d746a35c54876314 --- /dev/null +++ b/2024/LATTE3D_ Large-scale Amortized Text-To-Enhanced3D Synthesis/layout.json @@ -0,0 +1,9260 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 187, + 112, + 427, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 112, + 427, + 148 + ], + "spans": [ + { + "bbox": [ + 187, + 112, + 427, + 148 + ], + "type": "text", + "content": "LATTE3D: Large-scale Amortized Text-To-Enhanced3D Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": "Kevin Xie" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": ", Jonathan Lorraine" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": ", Tianshi Cao" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": ", Jun Gao" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": ", James Lucas" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": ", Antonio Torralba" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": ", Sanja Fidler" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": ", and Xiaohui Zeng" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 184, + 201, + 429, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 201, + 429, + 213 + ], + "spans": [ + { + "bbox": [ + 184, + 201, + 429, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 184, + 201, + 429, + 213 + ], + "type": "text", + "content": "Nvidia, " + }, + { + "bbox": [ + 184, + 201, + 429, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 184, + 201, + 429, + 213 + ], + "type": "text", + "content": "University of Toronto and Vector Institute, " + }, + { + "bbox": [ + 184, + 201, + 429, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 184, + 201, + 429, + 213 + ], + "type": "text", + "content": "MIT" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 151, + 238, + 465, + 393 + ], + "blocks": [ + { + "bbox": [ + 151, + 238, + 465, + 393 + ], + "lines": [ + { + "bbox": [ + 151, + 238, + 465, + 393 + ], + "spans": [ + { + "bbox": [ + 151, + 238, + 465, + 393 + ], + "type": "image", + "image_path": "044750169467ef79cf0e8c6281f0f18c5e9476fe02b78c8af1c71a3cf4af1dde.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 401, + 482, + 446 + ], + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 446 + ], + "type": "text", + "content": "Fig.1: Samples generated in " + }, + { + "bbox": [ + 130, + 401, + 482, + 446 + ], + "type": "inline_equation", + "content": "\\sim 400\\mathrm{ms}" + }, + { + "bbox": [ + 130, + 401, + 482, + 446 + ], + "type": "text", + "content": " on a single A6000 GPU from text prompts. Objects without prompt labels are generated by our text-to-3D model trained on " + }, + { + "bbox": [ + 130, + 401, + 482, + 446 + ], + "type": "inline_equation", + "content": "\\sim 100\\mathrm{k}" + }, + { + "bbox": [ + 130, + 401, + 482, + 446 + ], + "type": "text", + "content": " prompts, while labeled objects are generated by our 3D stylization model trained on " + }, + { + "bbox": [ + 130, + 401, + 482, + 446 + ], + "type": "inline_equation", + "content": "12\\mathrm{k}" + }, + { + "bbox": [ + 130, + 401, + 482, + 446 + ], + "type": "text", + "content": " prompts. See our project website** for more." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 468, + 453, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 468, + 453, + 634 + ], + "spans": [ + { + "bbox": [ + 159, + 468, + 453, + 634 + ], + "type": "text", + "content": "Abstract. Recent text-to-3D generation approaches produce impressive 3D results but require time-consuming optimization that can take up to an hour per prompt. Amortized methods like ATT3D optimize multiple prompts simultaneously to improve efficiency, enabling fast text-to-3D synthesis. However, they cannot capture high-frequency geometry and texture details and struggle to scale to large prompt sets, so they generalize poorly. We introduce LATTE3D, addressing these limitations to achieve fast, high-quality generation on a significantly larger prompt set. Key to our method is 1) building a scalable architecture and 2) leveraging 3D data during optimization through 3D-aware diffusion priors, shape regularization, and model initialization to achieve robustness to diverse and complex training prompts. LATTE3D amortizes both neural field and textured surface generation to produce highly detailed textured meshes in a single forward pass. LATTE3D generates 3D objects in 400ms, and can be further enhanced with fast test-time optimization." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 653, + 466, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 653, + 466, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 653, + 466, + 666 + ], + "type": "text", + "content": "*Equal contribution. **https://research.nvidia.com/labs/toronto-ai/LATTE3D/" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 144, + 481, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 144, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 144, + 481, + 215 + ], + "type": "text", + "content": "Recent advances in text-to-3D synthesis via image diffusion models mark significant progress in 3D content creation using [3,8,17,25,33,46,47]. However, these methods often involve an expensive and time-consuming optimization that can take up to an hour to generate a single 3D object [17, 33]. We aim to achieve real-time text-to-3D synthesis for diverse text prompts and support fast test-time optimization when a further quality boost is desired." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 216, + 481, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 481, + 310 + ], + "type": "text", + "content": "Seminal work [33] proposed to optimize a neural field using the score distillation sampling (SDS) loss to generate 3D assets. It unlocked open-vocabulary 3D content creation but was (a) lacking high-frequency geometric and texture details, (b) expensive, and (c) prone to failing to generate the desired object. Follow-ups introduced a surface-based fine-tuning stage [3, 17, 47] to generate high-frequency details. While this two-stage pipeline defines most current state-of-the-art text-to-3D approaches, it remains expensive and prone to prompt failure." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 312, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 418 + ], + "type": "text", + "content": "ATT3D [21] was the first to make text-to-3D generation fast by simultaneously training a single model on a set of prompts in a process called amortized optimization. Amortization reduces training time by optimizing a shared text-conditioned model on a set of prompts, enabling the model to generalize to new prompts at inference time. Despite promising results on curated prompts, ATT3D – and the underlying per-prompt methods – remain prone to failing on general prompts. Also, ATT3D only amortizes the first stage of 3D synthesis, producing a neural field representation and thus cannot generate high-frequency details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 421, + 481, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 421, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 131, + 421, + 481, + 468 + ], + "type": "text", + "content": "To address these issues, we introduce LATTE3D, a Large-scale Amortized Text-to-Enhanced3D synthesis method that can produce high-quality 3D content in real-time. Our work makes the following technical contributions to boost the quality, robustness, scale, and speed of text-to-3D generative models:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 479, + 485, + 552 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 138, + 479, + 480, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 502 + ], + "type": "text", + "content": "- We propose a novel text-to-3D amortization architecture that can scale to orders of magnitude larger prompt sets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 504, + 485, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 485, + 538 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 485, + 538 + ], + "type": "text", + "content": "- We leverage 3D data in training to improve quality and robustness, through 3D-aware diffusion priors, regularization loss, and weight initialization through pretraining with 3D reconstruction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 540, + 471, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 540, + 471, + 552 + ], + "spans": [ + { + "bbox": [ + 138, + 540, + 471, + 552 + ], + "type": "text", + "content": "- We amortize the surface-based refinement stage, greatly boosting quality." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 574, + 237, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 237, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 237, + 586 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 603, + 481, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 603, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 603, + 481, + 664 + ], + "type": "text", + "content": "Early work on 3D object synthesis was typically limited to generating objects from a given class [1,2,7,10,22,24,28,44,48,49,51,53], e.g. cars or chairs. Recent extensions use captioned 3D shapes to train text-to-3D models [13,29] generating shapes from diverse categories, but requiring 3D supervision, restricting them to synthetic datasets of limited size, diversity, and visual fidelity." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "K. Xie et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 332, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 332, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 332, + 331 + ], + "type": "text", + "content": "The advent of differentiable rendering, both volumetric [26] and surface-based [27], opened the door to inverse image rendering [7, 33], unlocking the use of powerful text-to-image generative models [36-38] in the 3D synthesis process. DreamFusion [33] proposed the SDS loss to optimize a neural field using a text-to-image diffusion model to generate 3D assets. Follow-up work introduced a surface-based refinement stage [17], allowing the synthesis of high-frequency details. This two-stage optimization pipeline now defines most of the state-of-the-art text-to-3D approaches [3,25,46,47]. The recently released large-scale 3D dataset Objaverse-XL [5,6] has spurred researchers to explore the benefits of 3D supervision in the synthesis process. To date, the most successful approaches make" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 340, + 119, + 480, + 225 + ], + "blocks": [ + { + "bbox": [ + 340, + 119, + 480, + 225 + ], + "lines": [ + { + "bbox": [ + 340, + 119, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 340, + 119, + 480, + 225 + ], + "type": "image", + "image_path": "18147faf535aeb209090b21e179d8fe9a152c5d764a147c20180725785dbf416.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 235, + 482, + 302 + ], + "lines": [ + { + "bbox": [ + 339, + 235, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 339, + 235, + 482, + 302 + ], + "type": "text", + "content": "Fig.2: Comparison of SOTA text-to-3D methods on unseen prompts. We plot different methods' user study preference rates compared to LATTE3D. For MVDream, we report varying optimization times." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 331, + 480, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 480, + 367 + ], + "type": "text", + "content": "text-conditioned image diffusion models 3D aware [15,19,41,42] by fine-tuning 2D diffusion models using rendered multiview images. SDS from multiview diffusion models is then used to obtain 3D shapes [19,41,42]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 368, + 481, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 481, + 450 + ], + "type": "text", + "content": "However, these approaches require a lengthy, per-prompt optimization process, hindering applicability to real-world content creation. To address this, two lines of work have been pursued. The first type of method uses a text-to-image-to-3D approach that generates images with a text-to-image diffusion model and trains an image-to-3D lifting network [15, 18, 41, 43]. They offer improved speed but with limited quality details compared to optimization methods, but still take " + }, + { + "bbox": [ + 130, + 368, + 481, + 450 + ], + "type": "inline_equation", + "content": "5 - 20" + }, + { + "bbox": [ + 130, + 368, + 481, + 450 + ], + "type": "text", + "content": " seconds per prompt [15, 43] for sampling from the image diffusion model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 451, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 547 + ], + "type": "text", + "content": "In an alternate line of work, ATT3D [21] proposed an amortized framework that optimizes multiple prompts simultaneously with a unified model. Follow-up works expanded the network architecture to achieve better quality [16, 34]. LATTE3D extends this line of work to larger scales by incorporating 3D knowledge to achieve prompt robustness during training, among other design improvements like better architecture. Amortized text-to-3D offers an attractive quality vs. speed trade-off as it uses a single feed-forward architecture without requiring sampling of diffusion models - see Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": "Concurrent Works Concurrently, AToM [34] also amortizes two-stage training with a triplane representation, but they do not scale beyond the smaller sets of prompts from ATT3D. In contrast, we scale to the order of 100000 prompts, using 3D-aware techniques. Another concurrent work of ET3D [4] achieves fast generation by training a GAN model, but it is only trained on a small-scale compositional dataset of 5000 prompts and does not guarantee view consistency as it uses a 2D neural upsampler. Lastly, LGM [43] is a text-to-image-to-3D method that generates Gaussian splats in 5 seconds. We experimentally demonstrate competitive performance to it." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "LATTE3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 233, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 233, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 233, + 129 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 257 + ], + "type": "text", + "content": "An established pipeline for high-quality text-to-3D generation consists of two stages, each performing per-prompt optimization using the SDS loss with a text-to-image diffusion model [3, 17, 47]. Stage-1 optimizes a volumetric representation, which is typically a neural radiance field. Meshes afford real-time rendering but are hard to optimize from scratch with image supervision alone. Stage-2 uses the output of the neural field to initialize a signed distance field (SDF) and a texture field, from which a mesh can be derived using differentiable isosurfacing [39, 40]. This surface-based representation is then optimized via differentiable rasterization [27], which is fast even at 1024 rendered image resolution – enabling supervision from higher-resolution text-to-image diffusion models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 482, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 341 + ], + "type": "text", + "content": "ATT3D [21] amortizes optimization of a neural field over a set of prompts with a hypernetwork mapping a text prompt to a neural field, which is trained with the SDS loss over a prompt set, referred to as seen prompts. LATTE3D introduces a new architecture that amortizes both stages of the generation process, aiming to produce high-quality textured meshes in real-time. We scale the approach to a magnitudes larger and more diverse set of prompts by leveraging 3D knowledge in the amortization loop." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 473 + ], + "type": "text", + "content": "Our approach is illustrated in Fig. 3. We initialize LATTE3D with a reconstruction pretraining step to stabilize training (Sec. 3.1). The architecture consists of two networks, one for geometry and the other for texture (Sec. 3.2). Amortized model training with diffusion priors is done through a two-stage pipeline consisting of a volume-based stage-1 and a surface-based stage-2 (Sec. 3.3). To reconcile reconstruction pretraining, which uses 3D geometry as input, with the task of text-to-3D generation, we anneal the network's input during stage-1 to gradually fully replace the dependence on 3D input with a single dummy input. During inference (Sec. 3.4), our model generates a 3D textured mesh from just a text prompt in 400ms and allows an optional lightweight test-time refinement to enhance the quality of geometry and textures (Sec. 3.5)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 489, + 332, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 489, + 332, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 489, + 332, + 502 + ], + "type": "text", + "content": "3.1 Pretraining to reconstruct shapes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 506, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 615 + ], + "type": "text", + "content": "Although SDS can synthesize text-aligned 3D features, empirically, it is a less efficient source of supervision due to high variance. We find that pretraining the model to first encode-decode 3D shapes subsequently accelerates the early stage of amortized SDS training. We initialize our model " + }, + { + "bbox": [ + 130, + 506, + 482, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 506, + 482, + 615 + ], + "type": "text", + "content": " by pretraining it on a 3D asset dataset using image reconstruction losses, similar to the reconstruction stage used in [9]. " + }, + { + "bbox": [ + 130, + 506, + 482, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 506, + 482, + 615 + ], + "type": "text", + "content": " inputs a sampled point cloud and outputs a predicted 3D shape. We then render multiple views of the predicted 3D shape " + }, + { + "bbox": [ + 130, + 506, + 482, + 615 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 130, + 506, + 482, + 615 + ], + "type": "text", + "content": " to compare with the input shape " + }, + { + "bbox": [ + 130, + 506, + 482, + 615 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 506, + 482, + 615 + ], + "type": "text", + "content": " renderings. We use an " + }, + { + "bbox": [ + 130, + 506, + 482, + 615 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 506, + 482, + 615 + ], + "type": "text", + "content": " loss on the rendered opacity and RGB image:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 144, + 622, + 481, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 622, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 144, + 622, + 481, + 635 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c o n}} (o, s, c) = \\left| \\left| \\mathbf {R} _ {\\text {o p a c i t y}} (o, c) - \\mathbf {R} _ {\\text {o p a c i t y}} (s, c) \\right| \\right| _ {2} + \\left| \\left| \\mathbf {R} _ {\\text {R G B}} (o, c) - \\mathbf {R} _ {\\text {R G B}} (s, c) \\right| \\right| _ {2} \\tag {1}", + "image_path": "5d7dfb89b425e48b8d4b0104f7661631dc43a65389f7f7b86835a07247cb60b7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "We denote the renderer " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": ", the opacity and RGB rendering with " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\mathrm{opacity}}" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\mathrm{RGB}}" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": ", respectively, and the camera with " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": ", randomly sampled on a sphere." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "K. Xie et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 116, + 479, + 224 + ], + "blocks": [ + { + "bbox": [ + 134, + 116, + 479, + 224 + ], + "lines": [ + { + "bbox": [ + 134, + 116, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 479, + 224 + ], + "type": "image", + "image_path": "751c8154c5442a8605be520b5c8a401f36b0b3c2f735b70b500fe1c0bdb22ee7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 233, + 482, + 300 + ], + "lines": [ + { + "bbox": [ + 130, + 233, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 482, + 300 + ], + "type": "text", + "content": "Fig. 3: LATTE3D consists of two networks: a texture network " + }, + { + "bbox": [ + 130, + 233, + 482, + 300 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 233, + 482, + 300 + ], + "type": "text", + "content": " and geometry network " + }, + { + "bbox": [ + 130, + 233, + 482, + 300 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 233, + 482, + 300 + ], + "type": "text", + "content": ". When amortizing the first stage, the encoders of both networks share the same set of weights. The training objective includes an SDS gradient from a 3D-aware image prior and a regularization loss that compares the rendered predicted shape's mask with the rendered masks of 3D assets in a library. When amortizing surface-based refinement in stage-2, we freeze the geometry network " + }, + { + "bbox": [ + 130, + 233, + 482, + 300 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 233, + 482, + 300 + ], + "type": "text", + "content": " and update the texture network " + }, + { + "bbox": [ + 130, + 233, + 482, + 300 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 233, + 482, + 300 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 323, + 261, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 323, + 261, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 323, + 261, + 335 + ], + "type": "text", + "content": "3.2 Model Architecture" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 439 + ], + "type": "text", + "content": "Our model " + }, + { + "bbox": [ + 130, + 342, + 482, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 342, + 482, + 439 + ], + "type": "text", + "content": " consists of two networks, " + }, + { + "bbox": [ + 130, + 342, + 482, + 439 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 342, + 482, + 439 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 482, + 439 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 342, + 482, + 439 + ], + "type": "text", + "content": ", for predicting geometry and texture, respectively. After pretraining, we introduce text-conditioning by adding cross-attention layers. We use CLIP [35] embeddings for encoding the text prompt " + }, + { + "bbox": [ + 130, + 342, + 482, + 439 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 342, + 482, + 439 + ], + "type": "text", + "content": ". In stage-1 training, we tie the geometry and texture networks' weights (except for separate MLP decoder heads), effectively training a single input encoder. In stage-2, we freeze the geometry network " + }, + { + "bbox": [ + 130, + 342, + 482, + 439 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 342, + 482, + 439 + ], + "type": "text", + "content": " and refine the texture network " + }, + { + "bbox": [ + 130, + 342, + 482, + 439 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 342, + 482, + 439 + ], + "type": "text", + "content": " with an additional trainable upsampling layer. The final model " + }, + { + "bbox": [ + 130, + 342, + 482, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 342, + 482, + 439 + ], + "type": "text", + "content": " used in inference is the frozen " + }, + { + "bbox": [ + 130, + 342, + 482, + 439 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 342, + 482, + 439 + ], + "type": "text", + "content": " from stage-1 and a refined " + }, + { + "bbox": [ + 130, + 342, + 482, + 439 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 342, + 482, + 439 + ], + "type": "text", + "content": " from stage-2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 450, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 582 + ], + "type": "text", + "content": "The geometry network " + }, + { + "bbox": [ + 130, + 450, + 482, + 582 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 450, + 482, + 582 + ], + "type": "text", + "content": " consists of a U-Net encoder on a triplane [2,7], followed by a decoder containing another triplane U-Net and a neural volumetric density field. Specifically, a point cloud is fed through a PointNet, then converted to a triplane via average scattering. During training, the point cloud comes from the 3D data, and in inference, a dummy sphere point cloud is used. While this point cloud encoder is inherited from reconstruction pretraining, it becomes unimportant in the final model. The triplane is processed through the encoder and decoder U-Nets. The text embedding is fed into every residual U-Net block in the encoder via cross-attention. For any point in 3D space, we extract triplane features from each plane via bilinear interpolation. An MLP then predicts the density value at that point." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "The texture network " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and geometry network " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " share the same encoder and decoder U-Net architecture in pretraining and stage-1. We employ another non-shared MLP neural texture field to predict RGB values at query points. In stage-2, we upsample the triplane features produced by the encoder to a higher resolution to gain extra capacity to generate high-frequency texture details. Specifically, we first bilinearly upsample the latent triplane from 128 resolution to" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "LATTE3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 474, + 241 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 474, + 241 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 474, + 241 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 474, + 241 + ], + "type": "image", + "image_path": "8c8399b526d550f0d7c1a5bb92d1daa5b789d24b9b914ca97293aad6992bd54c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 254, + 482, + 277 + ], + "lines": [ + { + "bbox": [ + 132, + 254, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 254, + 482, + 277 + ], + "type": "text", + "content": "Fig. 4: Stylization application: Our model learns to generate diverse stylizations of the same shapes. Left column shows the original shape to be stylized." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 300, + 482, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 300, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 131, + 300, + 482, + 338 + ], + "type": "text", + "content": "480 and add a residual MLP which maps the text-embedding to the residual of the triplane feature, mimicking ATT3D's mapping (hyper)network. The decoder then takes this upsampled triplane as input. Details are in the Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 350, + 482, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 350, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 131, + 350, + 482, + 482 + ], + "type": "text", + "content": "Rendering. To train the model, we render the generated 3D object into 2D images with different methods in stage-1 and 2. In stage-1, the output 3D geometry is a triplane-based neural field, on which we use volume rendering to get images of 256 resolution using the formulation from VolSDF [50], allowing us to parameterize the density via an approximate signed distance field whose zero-level set defines the object geometry. This choice allows a simple conversion to surface-based rendering. For the geometry network, we observe more stable training with volume rendering with SDS compared to surface rendering. In stage-2, we extract the isosurface from the density field with Marching Cubes [20] and render images via rasterization [14]. Meshes can be rendered at a 1024 image resolution in real-time, helping capture finer details for learning textures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 500, + 263, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 263, + 513 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 263, + 513 + ], + "type": "text", + "content": "3.3 Amortized Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 521, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 482, + 582 + ], + "type": "text", + "content": "Amortized learning distills knowledge from image diffusion priors into our model. We perform amortized training on volumetric geometry training (stage-1) and surface texture training (stage-2) sequentially. We incorporate 3D information into our training process (a) implicitly via 3D-aware 2D SDS losses and (b) explicitly with regularization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 482, + 666 + ], + "type": "text", + "content": "Curating prompt sets. We create large prompt sets for training using rule-based text generation or ChatGPT [31]. For the former, we start with the categories names from the captions of Objaverse [23], and design rules, like \"object A in style B is doing C\" with a set of styles and activities. Alternatively, we input the captions into ChatGPT and ask it for detailed and diverse prompts describing a similar 3D object. Fig. 3 shows that when generating \"Amigurumi fiddler crab" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "K. Xie et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 152 + ], + "type": "text", + "content": "wearing a top hat\", we guide shape synthesis towards resembling a crab by retrieving a related 3D shape used in a regularization loss. Other details of dataset construction are in Sec. 4.1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 167, + 479, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 479, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 479, + 202 + ], + "type": "text", + "content": "Amortizing Stage-1 Generation In this stage, we train " + }, + { + "bbox": [ + 132, + 167, + 479, + 202 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 132, + 167, + 479, + 202 + ], + "type": "text", + "content": " with our datasets. We use the SDS loss with a 3D-aware 2D prior and regularization to the paired 3D shape, which we now explain in detail:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 212, + 479, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 212, + 479, + 271 + ], + "spans": [ + { + "bbox": [ + 132, + 212, + 479, + 271 + ], + "type": "text", + "content": "3D-aware SDS loss. During stage-1 training we leverage a 3D aware diffusion prior, which provides a stronger, multiview consistent supervisory signal. Specifically, we use the model from MVDream [42], which was trained on four rendered views of objects from the Objaverse dataset [6] by modifying Stable Diffusion [37] to generate multiview consistent images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 281, + 479, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 479, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 479, + 328 + ], + "type": "text", + "content": "Regularization loss. While 3D-aware SDS loss with MVDream offers a strong multiview prior, it can still fail to correctly capture the full 3D geometry for some prompts since MVDream only supervises four views and can create enlarged geometry or floaters in the shape. Some failures are shown on the right of Fig. 10." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 329, + 479, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 329, + 479, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 329, + 479, + 376 + ], + "type": "text", + "content": "Combining regularization using 3D shapes with the SDS loss helps training and improves geometry. As shown in Fig. 3, for each training prompt, we regularize the output shape " + }, + { + "bbox": [ + 132, + 329, + 479, + 376 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 132, + 329, + 479, + 376 + ], + "type": "text", + "content": " by comparing its rendered mask to the mask of a shape " + }, + { + "bbox": [ + 132, + 329, + 479, + 376 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 132, + 329, + 479, + 376 + ], + "type": "text", + "content": " retrieved from the 3D dataset with the prompt:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 205, + 384, + 479, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 384, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 205, + 384, + 479, + 396 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e g}} (o, s, c) = \\left| \\left| \\mathbf {R} _ {\\text {o p a c i t y}} (o, c) - \\mathbf {R} _ {\\text {o p a c i t y}} (s, c) \\right| \\right| _ {2}, \\tag {2}", + "image_path": "e0617c8d1935a9ea93d64f7027935464f7b1ab6a349d0ec2eb0e25f6082d484e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 401, + 479, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 401, + 479, + 437 + ], + "spans": [ + { + "bbox": [ + 132, + 401, + 479, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 401, + 479, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\mathrm{opacity}}" + }, + { + "bbox": [ + 132, + 401, + 479, + 437 + ], + "type": "text", + "content": " is the opacity mask (volume-rendered density). We balance the regularization " + }, + { + "bbox": [ + 132, + 401, + 479, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 132, + 401, + 479, + 437 + ], + "type": "text", + "content": " and the SDS loss " + }, + { + "bbox": [ + 132, + 401, + 479, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SDS}}" + }, + { + "bbox": [ + 132, + 401, + 479, + 437 + ], + "type": "text", + "content": " by simply blending them using a convex combination with weight " + }, + { + "bbox": [ + 132, + 401, + 479, + 437 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 132, + 401, + 479, + 437 + ], + "type": "text", + "content": ". Our final training loss is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 241, + 445, + 479, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 445, + 479, + 456 + ], + "spans": [ + { + "bbox": [ + 241, + 445, + 479, + 456 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t r a i n}} = (1 - \\alpha) \\mathcal {L} _ {\\text {S D S}} + \\alpha \\mathcal {L} _ {\\text {r e g}}. \\tag {3}", + "image_path": "9dc09520f8abfaf7aca67a98377219242c0f51deb74cdc78e42a4e155bd2def5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 462, + 479, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 462, + 479, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 462, + 479, + 486 + ], + "type": "text", + "content": "During training, we render five views for each generated shape. We average across four views to compute " + }, + { + "bbox": [ + 132, + 462, + 479, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SDS}}" + }, + { + "bbox": [ + 132, + 462, + 479, + 486 + ], + "type": "text", + "content": ", and one is used to compute " + }, + { + "bbox": [ + 132, + 462, + 479, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 132, + 462, + 479, + 486 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 495, + 479, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 479, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 479, + 578 + ], + "type": "text", + "content": "Input point cloud annealing To bridge the gap between the real point cloud inputs in training and dummy point cloud in inference, we gradually anneal the input point clouds towards dummy point clouds during stage-1 training. Specifically, we randomly replace a subset of points in the input point cloud with points from the dummy point cloud using a probability that increases linearly from 0 to 1 in the last 5000 training iterations. With point cloud annealing, the model gets better results in inference when the dummy point cloud is used." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 594, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 479, + 665 + ], + "type": "text", + "content": "Amortizing Stage-2 Generation During stage-2 training, we freeze the network for our geometry and only tune the texture since training both leads to instabilities. We use depth-conditional ControlNet [52] for SDS guidance, allowing a higher, 512 resolution supervision. Depth conditioning encourages the SDS loss to guide the texture to align with the geometry from the geometry network, thereby improving the 3D texture consistency." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "LATTE3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 208, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 208, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 208, + 126 + ], + "type": "text", + "content": "3.4 Inference" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 254 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 254 + ], + "type": "text", + "content": "During inference, our model inputs the user's text prompt and a dummy point cloud. Our " + }, + { + "bbox": [ + 130, + 133, + 482, + 254 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = (G,T)" + }, + { + "bbox": [ + 130, + 133, + 482, + 254 + ], + "type": "text", + "content": " outputs a textured mesh, where " + }, + { + "bbox": [ + 130, + 133, + 482, + 254 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 133, + 482, + 254 + ], + "type": "text", + "content": " is the final texture network refined in stage-2, while " + }, + { + "bbox": [ + 130, + 133, + 482, + 254 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 133, + 482, + 254 + ], + "type": "text", + "content": " is trained in stage-1 only and then frozen. Predicting one shape and rendering the image only takes us 400ms on an A6000 GPU in inference, allowing the generation of user results at interactive speeds. A single A6000 can generate 4 samples simultaneously. Once the user selects a desired sample, textured meshes can be exported from LATTE3D. We use a UV mapping tool to UV parameterize the mesh output by " + }, + { + "bbox": [ + 130, + 133, + 482, + 254 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 133, + 482, + 254 + ], + "type": "text", + "content": " and use " + }, + { + "bbox": [ + 130, + 133, + 482, + 254 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 133, + 482, + 254 + ], + "type": "text", + "content": " to predict the albedo colors of each texel in a UV-mapped texture image. The resulting mesh and texture image is compatible with standard rendering software." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 269, + 282, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 269, + 282, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 269, + 282, + 281 + ], + "type": "text", + "content": "3.5 Test Time Optimization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 286, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 286, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 482, + 370 + ], + "type": "text", + "content": "Our method supports test time optimization if the user wants to boost the quality of a particular prompt. This can be useful if the user's prompt deviates significantly from the seen prompts. As in our stage-2 training, we freeze the geometry and finetune the texture network " + }, + { + "bbox": [ + 130, + 286, + 482, + 370 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 286, + 482, + 370 + ], + "type": "text", + "content": " with depth-conditional SDS guidance. Our test time optimization is significantly faster to converge than MVDream [42] or other optimization-based text-to-3D methods [17, 33, 47] since we optimize our amortized network instead of a randomly initialized network." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 385, + 233, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 233, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 233, + 396 + ], + "type": "text", + "content": "3.6 3D Stylization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 403, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 522 + ], + "type": "text", + "content": "LATTE3D can also be trained and deployed as a 3D stylization method, allowing users to cheaply create variations from existing 3D assets. During training, we skip the input point cloud annealing step to maintain point cloud reliance and amortize over a dataset where each 3D shape corresponds to many style prompts. Our blend of regularization and SDS guidance drives the model to produce shapes structurally similar to the original shapes yet semantically aligned with the style prompt. Amortization significantly reduces the computational cost per shape-prompt pair compared to optimizing each pair separately. During inference, point clouds from the training set can be combined with novel prompts as input to the model to produce variations of the original shape." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 539, + 230, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 539, + 230, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 539, + 230, + 552 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 561, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 561, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 561, + 482, + 609 + ], + "type": "text", + "content": "We first describe our experimental setup, including datasets, metrics, and baselines for comparison (Sec. 4.1). We then present our quantitative and qualitative results (Sec. 4.2), demonstrating the performance and generalization ability of LATTE3D (Sec. 4.3), followed by design choices ablations (Sec. 4.4)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 624, + 262, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 262, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 262, + 636 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 642, + 481, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 642, + 481, + 654 + ], + "spans": [ + { + "bbox": [ + 132, + 642, + 481, + 654 + ], + "type": "text", + "content": "We now discuss dataset construction, model pretraining, and evaluation metrics." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "K. Xie et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 116, + 200, + 190 + ], + "blocks": [ + { + "bbox": [ + 151, + 116, + 200, + 190 + ], + "lines": [ + { + "bbox": [ + 151, + 116, + 200, + 190 + ], + "spans": [ + { + "bbox": [ + 151, + 116, + 200, + 190 + ], + "type": "image", + "image_path": "83912decb4981a65caef63b0e3bb9d6c01c4b7d6c55e50d0f2bff8f0e866a083.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 192, + 192, + 207 + ], + "lines": [ + { + "bbox": [ + 164, + 192, + 192, + 207 + ], + "spans": [ + { + "bbox": [ + 164, + 192, + 192, + 207 + ], + "type": "text", + "content": "MVDream (6min)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 203, + 116, + 246, + 190 + ], + "blocks": [ + { + "bbox": [ + 203, + 116, + 246, + 190 + ], + "lines": [ + { + "bbox": [ + 203, + 116, + 246, + 190 + ], + "spans": [ + { + "bbox": [ + 203, + 116, + 246, + 190 + ], + "type": "image", + "image_path": "79f3f08df33e0a26d89e02401a439b48e10a4c9ff0bc2b96eb1d77250f7c6504.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 192, + 240, + 207 + ], + "lines": [ + { + "bbox": [ + 212, + 192, + 240, + 207 + ], + "spans": [ + { + "bbox": [ + 212, + 192, + 240, + 207 + ], + "type": "text", + "content": "MVDream (36min)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 253, + 114, + 285, + 190 + ], + "blocks": [ + { + "bbox": [ + 253, + 114, + 285, + 190 + ], + "lines": [ + { + "bbox": [ + 253, + 114, + 285, + 190 + ], + "spans": [ + { + "bbox": [ + 253, + 114, + 285, + 190 + ], + "type": "image", + "image_path": "b6733d82b10c55e37379f886c89930e565222c7cae08de120e8526350394ccd1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 192, + 284, + 207 + ], + "lines": [ + { + "bbox": [ + 261, + 192, + 284, + 207 + ], + "spans": [ + { + "bbox": [ + 261, + 192, + 284, + 207 + ], + "type": "text", + "content": "3DTopia (20s)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 216, + 481, + 250 + ], + "lines": [ + { + "bbox": [ + 131, + 216, + 481, + 250 + ], + "spans": [ + { + "bbox": [ + 131, + 216, + 481, + 250 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparison between LATTE3D and ATT3D [21], MVDream [42], Instant3D/3DTopia [15] and LGM [43]. Both rows are unseen prompts: \"a red convertible car with the top down\", \"an ice cream sundae\"." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 291, + 116, + 333, + 190 + ], + "blocks": [ + { + "bbox": [ + 291, + 116, + 333, + 190 + ], + "lines": [ + { + "bbox": [ + 291, + 116, + 333, + 190 + ], + "spans": [ + { + "bbox": [ + 291, + 116, + 333, + 190 + ], + "type": "image", + "image_path": "cb1fd2c0efb16cfc2c02bdde6ea2ec5ba30c028ff17e788384425338c6c59069.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 192, + 316, + 207 + ], + "lines": [ + { + "bbox": [ + 304, + 192, + 316, + 207 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 316, + 207 + ], + "type": "text", + "content": "LGM (5s)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 337, + 116, + 379, + 190 + ], + "blocks": [ + { + "bbox": [ + 337, + 116, + 379, + 190 + ], + "lines": [ + { + "bbox": [ + 337, + 116, + 379, + 190 + ], + "spans": [ + { + "bbox": [ + 337, + 116, + 379, + 190 + ], + "type": "image", + "image_path": "9c837622280e3a60d09351465007e50fa6e19f1622a89872d69f8ef07dbe7be8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 191, + 364, + 199 + ], + "lines": [ + { + "bbox": [ + 346, + 191, + 364, + 199 + ], + "spans": [ + { + "bbox": [ + 346, + 191, + 364, + 199 + ], + "type": "text", + "content": "ATT3D" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 383, + 115, + 421, + 188 + ], + "blocks": [ + { + "bbox": [ + 383, + 115, + 421, + 188 + ], + "lines": [ + { + "bbox": [ + 383, + 115, + 421, + 188 + ], + "spans": [ + { + "bbox": [ + 383, + 115, + 421, + 188 + ], + "type": "image", + "image_path": "a0399853b212ddab43af57880eb9a0cca5743bb612120d65bbd8f135de172f06.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 191, + 416, + 206 + ], + "lines": [ + { + "bbox": [ + 386, + 191, + 416, + 206 + ], + "spans": [ + { + "bbox": [ + 386, + 191, + 416, + 206 + ], + "type": "text", + "content": "ATT3D+ UNet " + }, + { + "bbox": [ + 386, + 191, + 416, + 206 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 386, + 191, + 416, + 206 + ], + "type": "text", + "content": " MV" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 425, + 116, + 463, + 187 + ], + "blocks": [ + { + "bbox": [ + 425, + 116, + 463, + 187 + ], + "lines": [ + { + "bbox": [ + 425, + 116, + 463, + 187 + ], + "spans": [ + { + "bbox": [ + 425, + 116, + 463, + 187 + ], + "type": "image", + "image_path": "d4aa217c847d3243099139360b25baa76411ceb3b5398c4c96dd67d6e86c7412.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 191, + 457, + 207 + ], + "lines": [ + { + "bbox": [ + 432, + 191, + 457, + 207 + ], + "spans": [ + { + "bbox": [ + 432, + 191, + 457, + 207 + ], + "type": "text", + "content": "LATTE3D (400ms)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 271, + 298, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 298, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 298, + 390 + ], + "type": "text", + "content": "Training Set To investigate scaling amortized optimization beyond what has been done in prior works - see Tab. 1 - we construct a new dataset gpt-101k that consists of 101k text prompts and 34k shapes. For the shapes, we use the \"lvis\" subset from Objaverse [6] with 50k human-verified shapes. We filter out poor shapes, e.g. flat images and scans, and retain 34k" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 391, + 480, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 480, + 415 + ], + "type": "text", + "content": "shapes for our dataset. We use ChatGPT to augment each object's caption to gather 3 prompts per object." + } + ] + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 308, + 307, + 481, + 369 + ], + "blocks": [ + { + "bbox": [ + 306, + 283, + 480, + 305 + ], + "lines": [ + { + "bbox": [ + 306, + 283, + 480, + 305 + ], + "spans": [ + { + "bbox": [ + 306, + 283, + 480, + 305 + ], + "type": "text", + "content": "Table 1: Comparing the sizes of amortization training sets to prior works." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 307, + 481, + 369 + ], + "lines": [ + { + "bbox": [ + 308, + 307, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 308, + 307, + 481, + 369 + ], + "type": "table", + "html": "
name#promptsSource
Animal2400 [21,34]2400Rule
DF415 [34]415DF [33]
animal-style (Ours)12000Rule
gpt-101k (Ours)101608GPT
", + "image_path": "4940f1be80cec561eac7b53d1596fa75e89e739d143243fb28c41a1a146da7ca.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 425, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 509 + ], + "type": "text", + "content": "Evaluation Sets We use two prompt sets for evaluation. To evaluate the benefit of amortized optimization in our method, we evaluate our model on gpt-101k. However, due to the computational cost of running the baselines, we randomly select a 50 prompt subset in gpt-101k to form the seen prompt set for benchmarking. Furthermore, to gauge the generalization abilities of our amortized model, we construct an unseen prompt set by filtering a subset of 67 DreamFusion [33] prompts that are close in distribution to lvis categories." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 517, + 484, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 484, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 484, + 602 + ], + "type": "text", + "content": "Baselines. We compare LATTE3D with the baselines: ATT3D [21], MVDream [42], Instant3D [15] and LGM [43]. We re-implement ATT3D [21] with a hypernetwork. We use the open-source threestudio [8] implementation of MVDream. For MVDream, we train each 3D model for 10000 iterations, with a batch size of 8 on a single A100 GPU with provided default settings. For Instant3D, since the source code has not been released, we use the re-implementation from 3DTopia1. For LGM, we use the official open-source implementation." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 131, + 611, + 481, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 611, + 481, + 648 + ], + "spans": [ + { + "bbox": [ + 131, + 611, + 481, + 648 + ], + "type": "text", + "content": "Evaluation metrics. We quantitatively compare our model with baselines by evaluating the generated 3D content fidelity and its consistency to the text prompt for the optimization cost. We use three metrics to evaluate fidelity." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "LATTE3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 306, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 306, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 306, + 666 + ], + "type": "text", + "content": "1 https://github.com/3DTopia/3DTopia" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 156, + 157, + 455, + 281 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Quantitative metrics and average user preference (\\%) of baselines over LATTE3D trained on gpt-101k using seen and unseen prompts. We also report test-time optimization, which takes " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "10\\mathrm{min}" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 156, + 157, + 455, + 281 + ], + "lines": [ + { + "bbox": [ + 156, + 157, + 455, + 281 + ], + "spans": [ + { + "bbox": [ + 156, + 157, + 455, + 281 + ], + "type": "table", + "html": "
ModelTimeRender-FID ↓ seenCLIP-Score ↑ unseen(df)Preference % ↑seenUnseen(df)seenUnseen(df)
LATTE3D400ms180.58190.000.26080.2605--
LATTE3D-opt10min171.69178.370.28000.3000--
MVDream6min203.53199.230.22400.26982.512.4
MVDream18min191.77188.840.24400.295026.552.2
MVDream36min158.11143.440.28300.318048.589.3
3DTopia20s217.29222.760.18370.217120.825.6
LGM5s178.67166.400.28400.306326.236.4
", + "image_path": "35a64fe7d924cc68230f13b68c74e0b24d7a99363960a44c6e0424f47fd592ae.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 302, + 479, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 479, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 479, + 348 + ], + "type": "text", + "content": "Render-FID: We compute the FID [11] between the renderings of generated 3D content and a set of images sampled from Stable Diffusion [37] with the same text prompts as input. This metric measures how well the generated shapes align with those from the 2D prior in visual quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 350, + 479, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 479, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 479, + 384 + ], + "type": "text", + "content": "CLIP Score: We compute the average CLIP scores between the text prompt and each rendered image to indicate how well the generated 3D shape aligns with the input text prompt." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 386, + 480, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 480, + 444 + ], + "type": "text", + "content": "User Study: We evaluate the overall 3D geometry and texture through Amazon Mechanical Turk user studies. We present videos of 3D shapes generated by our method and a baseline head-to-head, with corresponding text prompts, and ask users for their preference. Each comparison is averaged over three users to obtain preference rate for each baseline relative to LATTE3D." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 446, + 481, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 481, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 481, + 529 + ], + "type": "text", + "content": "Timing: We measure our optimization cost by GPU time per prompt to gauge the total compute cost on each prompt. We divide the total wall-clock time for pretraining, stage-1, and stage-2 by the prompt set's size, then multiply by the number of GPUs used. We measure inference time for Instant3D, LGM, and LATTE3D by the time the model takes from in-taking the text input to outputting the final features before rendering. For MVDream [42], we measure the average time to optimize a single prompt on an A100 GPU." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 548, + 269, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 269, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 269, + 560 + ], + "type": "text", + "content": "4.2 Experimental Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 569, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 481, + 640 + ], + "type": "text", + "content": "We observe robust generalization of LATTE3D by training on the gpt-101k dataset to unseen prompts (df, from DreamFusion [33]), as demonstrated by qualitative (Fig. 5) and quantitative (Fig. 2, Tab. 2) results. We highlight the relative user preference of LATTE3D versus different baselines over inference time in Fig. 2. LATTE3D produce results of competitive user preference to SOTA baselines at a uniquely fast inference time." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 642, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 642, + 481, + 665 + ], + "type": "text", + "content": "From the quantitative comparisons in Tab. 2, our performance on both seen and unseen prompts is competitive with baselines while our inference time is" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "K. Xie et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 139, + 119, + 306, + 264 + ], + "blocks": [ + { + "bbox": [ + 139, + 119, + 306, + 264 + ], + "lines": [ + { + "bbox": [ + 139, + 119, + 306, + 264 + ], + "spans": [ + { + "bbox": [ + 139, + 119, + 306, + 264 + ], + "type": "image", + "image_path": "f4d2dd9e976d1cb01e70ff4f923565d90cfe19bb0c86bd337b931a6867995c01.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 274, + 312, + 319 + ], + "lines": [ + { + "bbox": [ + 135, + 274, + 312, + 319 + ], + "spans": [ + { + "bbox": [ + 135, + 274, + 312, + 319 + ], + "type": "text", + "content": "Fig.6: Results of user study showing the average preference of MVDream at different amounts of optimization time to LATTE3D gpt-101k on seen prompts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 318, + 114, + 465, + 219 + ], + "blocks": [ + { + "bbox": [ + 318, + 114, + 465, + 219 + ], + "lines": [ + { + "bbox": [ + 318, + 114, + 465, + 219 + ], + "spans": [ + { + "bbox": [ + 318, + 114, + 465, + 219 + ], + "type": "image", + "image_path": "4c5a66890e75a1005ddeca924f8c637893c2fcbe09b0b69a7595bdd8044ac056.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 228, + 479, + 327 + ], + "lines": [ + { + "bbox": [ + 317, + 228, + 479, + 327 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 479, + 327 + ], + "type": "text", + "content": "Fig.7: Test time optimization on seen and unseen prompt. The model is trained on gpt-101k. First column: without test time optimization. Other columns: test time optimization with the given time budget denoted. Top: seen prompt \"...animal with a blue and white floral pattern...\" Bottom: unseen prompt \"orchid in clay pot\"." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 350, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 411 + ], + "type": "text", + "content": "an order of magnitude lower. Image-based metrics of FID and CLIP score may fail to detect generations with 3D geometry errors, less obvious in still images than in the user study, where users are shown videos of objects rotating through multiple views. Interested readers can refer to rendered videos of all methods in the supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 429, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 501 + ], + "type": "text", + "content": "Total optimization cost A benefit of amortized optimization over many prompts is that our total optimization cost on gpt-101k is much lower than optimizing MVDream per-prompt. Fig. 6 illustrates that to achieve user preference on par with LATTE3D, MVDream would require 36 GPU minutes per prompt, whereas LATTE3D spent only 215 GPU seconds per prompt, representing a " + }, + { + "bbox": [ + 130, + 429, + 482, + 501 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 130, + 429, + 482, + 501 + ], + "type": "text", + "content": " reduction in optimization cost." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 519, + 220, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 519, + 220, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 220, + 532 + ], + "type": "text", + "content": "4.3 Application" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 539, + 481, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 481, + 564 + ], + "type": "text", + "content": "We illustrate LATTE3D's capacity for enhanced quality through test time optimization and how to adapt LATTE3D for stylizing 3D content." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 485, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 485, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 485, + 666 + ], + "type": "text", + "content": "Per Prompt Test-time Optimization We show experimentally that LATTE3D quickly adapts to new prompts during test time to improve quality. In Fig. 7, we qualitatively compare LATTE3D with up to 600 steps. Each optimization iteration takes an average of 0.98 seconds on an A100 GPU, giving us a total compute budget of less than 10 GPU minutes per prompt. Even with test-time optimization, our method is still an order of magnitude faster than per-prompt optimization methods such as MVDream, as in Tab. 2. Test-time optimization is" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "LATTE3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "particularly beneficial on unseen prompts, where FID drops by 11.6 for unseen vs. 8.8 for seen, and the CLIP score gains 0.04 for unseen vs. 0.02 for seen." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 151, + 167, + 462, + 324 + ], + "blocks": [ + { + "bbox": [ + 151, + 167, + 462, + 324 + ], + "lines": [ + { + "bbox": [ + 151, + 167, + 462, + 324 + ], + "spans": [ + { + "bbox": [ + 151, + 167, + 462, + 324 + ], + "type": "image", + "image_path": "174459b7cdbcf8a9bbaa5b7529de9f97e75ffa73e2618f419c205ba645881d92.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 334, + 480, + 357 + ], + "lines": [ + { + "bbox": [ + 131, + 334, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 131, + 334, + 480, + 357 + ], + "type": "text", + "content": "Fig. 8: Results of user study showing the average preference rate for MVDream at different amounts of optimization time to LatTE3D stylization results on animal-style." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 412, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 544 + ], + "type": "text", + "content": "Stylizing 3D content We illustrate how LATTE3D architecture offers us the flexibility to adapt it for targeted 3D content stylization as described in Sec. 3.6. To test this, we manually curate a dataset of 100 animal shapes from Objaverse as a basis and augment the species name of each animal shape name with combinations of activities (e.g. \"riding a skateboard\") and styles (e.g. \"voxel style\") to create our 12000 prompts animal-style dataset. We train a LATTE3D model (stage-1 and stage-2) on animal-style without the point cloud annealing stage. The resulting model adapts to both the text prompt and the point cloud inputs. In Fig. 4, each row contains the outputs of LATTE3D given different prompts while fixing the point cloud input, demonstrating the effectiveness of LATTE3D for text-controlled stylization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "In Fig. 8, we conduct a user study comparing LATTE3D's outputs on animal-style to MVDream and find that LATTE3D is competitive with MVDream with 4000 steps of optimization, representing a " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " reduction in optimization cost (compared to running MVDream for each prompt). LATTE3D on animal-style also generalizes to held-out combinations of animal activities and styles with little quality drop, as indicated by its similar relative preference to MVDream. Therefore, LATTE3D enables users to cheaply create variations of their 3D assets via amortized optimization and even explore novel prompt combinations at test time. Additional quantitative and qualitative results can be found in the supplementary materials." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "K. Xie et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 168, + 473, + 251 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 3: Ablation of components in stage-1 training. Trained on gpt-101k data and evaluated on seen and unseen prompts. Preference indicate average user preference of baseline over LATTE3D. LATTE3D is better than all ablated settings in quantitative metrics and is preferred by users on the unseen prompt set." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 168, + 473, + 251 + ], + "lines": [ + { + "bbox": [ + 141, + 168, + 473, + 251 + ], + "spans": [ + { + "bbox": [ + 141, + 168, + 473, + 251 + ], + "type": "table", + "html": "
ExpMV Unet Pretrain RegSeenUnseen
Mask-FID↓Render-FID↓Clip-Score↑Preference
ATT3D274.44275.010.209128.324.2
+MV✓243.24236.730.182634.826.9
+MV+UNet✓✓184.80203.330.221448.045.6
+MV+UNet+PT✓✓✓189.09189.510.219151.847.7
LATTE3D (S1)✓✓✓✓176.44186.840.2379--
", + "image_path": "811428834e0f8ac5324cdda902f0e71d0f93b88e71a560a6e0a316afc816213a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 270, + 246, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 246, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 246, + 281 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 289, + 481, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 481, + 324 + ], + "type": "text", + "content": "We now analyze the value of each component we introduced in LATTE3D. We include additional results, including ablations on different diffusion priors, in the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 334, + 481, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 481, + 430 + ], + "type": "text", + "content": "Stage-1 ablation. Tab. 3 summarizes our quantitative comparison of the following ablations: " + }, + { + "bbox": [ + 130, + 334, + 481, + 430 + ], + "type": "inline_equation", + "content": "ATT3D" + }, + { + "bbox": [ + 130, + 334, + 481, + 430 + ], + "type": "text", + "content": " baseline using a hypernet, " + }, + { + "bbox": [ + 130, + 334, + 481, + 430 + ], + "type": "inline_equation", + "content": "ATT3D + MV" + }, + { + "bbox": [ + 130, + 334, + 481, + 430 + ], + "type": "text", + "content": " uses MVDream as diffusion guidance. " + }, + { + "bbox": [ + 130, + 334, + 481, + 430 + ], + "type": "inline_equation", + "content": "ATT3D + UNet + MV" + }, + { + "bbox": [ + 130, + 334, + 481, + 430 + ], + "type": "text", + "content": " replaces the hypernet with our model architecture, but is initialized from scratch and given a dummy point cloud as input. " + }, + { + "bbox": [ + 130, + 334, + 481, + 430 + ], + "type": "inline_equation", + "content": "ATT3D + UNet + MV + PT" + }, + { + "bbox": [ + 130, + 334, + 481, + 430 + ], + "type": "text", + "content": " is initialized from the reconstruction pretrained (PT) model. We introduce the Mask-FID metric to quantitatively assess geometry adherence to the shape dataset by computing the FID between the rendered binary masks of the generated shapes and those of the 3D dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 430, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 538 + ], + "type": "text", + "content": "We find that each of the components we introduce improves performance. Comparing ATT3D with ATT3D+MV, we see benefits from using the MV-Dream guidance as it mitigates geometric artifacts like Janus faces. Comparing ATT3D+MV with ATT3D+MV+UNet we see a performance boost from our architecture. Further, adding pretraining (ATT3D+UNet+MV+PT) helps, especially in recovering fine geometric and texture details. Finally, adding shape regularization (full LATTE3D) makes the generated geometry adhere better to the 3D data, as evidenced by the drop in Mask-FID. An additional user study is shown in Tab. 3. We show the convergence speed in the Appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 548, + 481, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 481, + 572 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 481, + 572 + ], + "type": "text", + "content": "Ablation on using stage-2 refinement. Stage-2 refinement gives large improvements in texture details over stage-1 training. We show the comparison in Fig. 9." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 481, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 653 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 653 + ], + "type": "text", + "content": "Regularization We now investigate the design choice of how to weight " + }, + { + "bbox": [ + 130, + 582, + 481, + 653 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 582, + 481, + 653 + ], + "type": "text", + "content": " between the SDS loss and the regularization loss as in Eq. 3. In Tab. 3, we show that adding regularization loss improves the quantitative results. We qualitatively contrast training on different fixed weightings in Fig. 10 and the Appendix. We see greater consistency with the input shape by increasing shape regularization with higher blend factors " + }, + { + "bbox": [ + 130, + 582, + 481, + 653 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 582, + 481, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "LATTE3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 121, + 296, + 165 + ], + "blocks": [ + { + "bbox": [ + 140, + 121, + 296, + 165 + ], + "lines": [ + { + "bbox": [ + 140, + 121, + 296, + 165 + ], + "spans": [ + { + "bbox": [ + 140, + 121, + 296, + 165 + ], + "type": "image", + "image_path": "09e8636397efaa3277105f9cb0f25a00774127e50592d968ac3e3d34635ea567.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 172, + 296, + 205 + ], + "lines": [ + { + "bbox": [ + 137, + 172, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 137, + 172, + 296, + 205 + ], + "type": "text", + "content": "Fig.9: Comparison of stage-1 and stage-2 results. Stage-2 refinement significantly improves the texture quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 323, + 113, + 470, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 113, + 470, + 121 + ], + "spans": [ + { + "bbox": [ + 323, + 113, + 470, + 121 + ], + "type": "text", + "content": "\"fun sheep wearing skis in pixel art style, voxel\"" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 323, + 122, + 473, + 169 + ], + "blocks": [ + { + "bbox": [ + 323, + 122, + 473, + 169 + ], + "lines": [ + { + "bbox": [ + 323, + 122, + 473, + 169 + ], + "spans": [ + { + "bbox": [ + 323, + 122, + 473, + 169 + ], + "type": "image", + "image_path": "d5aeebefcb872181cf9a9d91c95685eb82616f8cbed60f161a1354b0edd85845.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 178, + 476, + 212 + ], + "lines": [ + { + "bbox": [ + 315, + 178, + 476, + 212 + ], + "spans": [ + { + "bbox": [ + 315, + 178, + 476, + 212 + ], + "type": "text", + "content": "Fig. 10: Comparison of the impact of blending factor " + }, + { + "bbox": [ + 315, + 178, + 476, + 212 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 315, + 178, + 476, + 212 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 315, + 178, + 476, + 212 + ], + "type": "inline_equation", + "content": "(\\alpha = 0)" + }, + { + "bbox": [ + 315, + 178, + 476, + 212 + ], + "type": "text", + "content": " is a failure from no regularization." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 133, + 314, + 293, + 369 + ], + "blocks": [ + { + "bbox": [ + 132, + 257, + 291, + 310 + ], + "lines": [ + { + "bbox": [ + 132, + 257, + 291, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 291, + 310 + ], + "type": "text", + "content": "Table 4: An ablation of unseen DreamFusion inputs in inference and annealing in training. Users prefer the annealed model with a dummy input in " + }, + { + "bbox": [ + 132, + 257, + 291, + 310 + ], + "type": "inline_equation", + "content": "51.2\\%" + }, + { + "bbox": [ + 132, + 257, + 291, + 310 + ], + "type": "text", + "content": " of cases." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 314, + 293, + 369 + ], + "lines": [ + { + "bbox": [ + 133, + 314, + 293, + 369 + ], + "spans": [ + { + "bbox": [ + 133, + 314, + 293, + 369 + ], + "type": "table", + "html": "
Modelwithout annealwith anneal
InputTop-1DummyDummy
CLIP0.26280.25770.2605
FID186.46193.28190.00
", + "image_path": "ef9eb354af008ed1c0139c15457c275a9945c8a6c25af83f5c6a5b5417d0c94a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 244, + 481, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 244, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 296, + 244, + 481, + 399 + ], + "type": "text", + "content": "Annealing in training and different input point clouds in inference. In training, we take retrieved point clouds from the dataset as input and anneal the input to become a fixed dummy point cloud. In inference, we only use the dummy point cloud as input. In Tab. 4, we quantitatively ablate the role of this point cloud annealing process on the behavior of LATTE3D on gpt-101k. Our model without annealing training is somewhat sensitive to the point cloud input. There is a small performance drop when using a" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 400, + 480, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 480, + 436 + ], + "type": "text", + "content": "dummy as input in inference compared to a retrieved point cloud. However, the performance gap is reduced by introducing point cloud annealing so the model trains with dummy input. Qualitative examples are in the Appendix." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 475, + 311, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 311, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 311, + 488 + ], + "type": "text", + "content": "5 Conclusion and Limitation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": "We presented a scalable approach to perform amortized text-to-enhanced-3D generation. To successfully scale amortization to larger datasets, we used 3D data through (1) 3D-aware 2D SDS, (2) pretraining, and (3) 3D regularization of the amortization process. We further improve the model architecture to be more scalable. Our model generates high-quality shapes within " + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "inline_equation", + "content": "400\\mathrm{ms}" + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": ". Moreover, quality and generalization can be improved further via a speedy test time optimization. Our model also has limitations. First, our model uses SDS and thus relies on the understanding of the text-to-image models, which can often fail to respect fine-grained details such as part descriptions. Second, the geometry is frozen in stage-2 and test-time optimization, so the geometry flaws from stage-1 cannot be fixed. Developing amortized training for stable geometry changes in stage-2 is left for future work." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "K. Xie et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 176 + ], + "type": "text", + "content": "We thank Matan Atzmon, Or Perel, Clement Fuji Tsang, Masha Shugrina, and her group for helpful feedback. The Python community [30,45] made underlying tools, including PyTorch [32] & matplotlib [12]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 193, + 263, + 207 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 263, + 207 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 263, + 207 + ], + "type": "text", + "content": "Disclosure of Funding" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 217, + 481, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 481, + 253 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 481, + 253 + ], + "type": "text", + "content": "NVIDIA funded this work. Kevin Xie, Jonathan Lorraine, Tianshi Cao, Jun Gao, and Xiaohui Zeng had funding from student scholarships at the University of Toronto and the Vector Institute, which are not in direct support of this work." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 270, + 198, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 198, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 198, + 282 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 294, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 294, + 481, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 294, + 481, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 294, + 481, + 327 + ], + "type": "text", + "content": "1. Achlioptas, P., Diamanti, O., Mitlagkas, I., Guibas, L.: Learning representations and generative models for 3d point clouds. In: International conference on machine learning. pp. 40-49. PMLR (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 327, + 481, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 481, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 481, + 360 + ], + "type": "text", + "content": "2. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., Mello, S.D., Gallo, O., Guibas, L., Tremblay, J., Khamis, S., Karras, T., Wetzstein, G.: Efficient geometry-aware 3D generative adversarial networks. In: arXiv (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 360, + 481, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 360, + 481, + 393 + ], + "spans": [ + { + "bbox": [ + 138, + 360, + 481, + 393 + ], + "type": "text", + "content": "3. Chen, R., Chen, Y., Jiao, N., Jia, K.: Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 393, + 481, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 393, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 138, + 393, + 481, + 415 + ], + "type": "text", + "content": "4. Chen, Y., Li, Z., Liu, P.: ET3D: efficient text-to-3d generation via multi-view distillation. arXiv 2311.15561 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 415, + 481, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 481, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 481, + 458 + ], + "type": "text", + "content": "5. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., VanderBilt, E., Kembhavi, A., Vondrick, C., Gkioxari, G., Ehsani, K., Schmidt, L., Farhadi, A.: Objverse-xl: A universe of " + }, + { + "bbox": [ + 138, + 415, + 481, + 458 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 138, + 415, + 481, + 458 + ], + "type": "text", + "content": " objects (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 458, + 481, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 481, + 502 + ], + "type": "text", + "content": "6. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3d objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13142-13153 (June 2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 502, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 481, + 534 + ], + "type": "text", + "content": "7. Gao, J., Shen, T., Wang, Z., Chen, W., Yin, K., Li, D., Litany, O., Gojcic, Z., Fidler, S.: Get3d: A generative model of high quality 3d textured shapes learned from images. Advances in Neural Information Processing Systems (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 534, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 481, + 578 + ], + "type": "text", + "content": "8. Guo, Y.C., Liu, Y.T., Shao, R., Laforte, C., Voleti, V., Luo, G., Chen, C.H., Zou, Z.X., Wang, C., Cao, Y.P., Zhang, S.H.: threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 578, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 600 + ], + "type": "text", + "content": "9. Gupta, A., Xiong, W., Nie, Y., Jones, I., Oğuz, B.: 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "type": "text", + "content": "10. Henzler, P., Mitra, N.J., Ritschel, T.: Escaping plato's cave: 3d shape from adversarial rendering. In: The IEEE International Conference on Computer Vision (ICCV) (October 2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 481, + 665 + ], + "type": "text", + "content": "11. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "LATTE3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "12. Hunter, J.D.: Matplotlib: A 2D graphics environment. Computing in Science & Engineering (2007)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "text", + "content": "13. Jun, H., Nichol, A.: Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 160, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 480, + 191 + ], + "type": "text", + "content": "14. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics 39(6) (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 480, + 224 + ], + "type": "text", + "content": "15. Li, J., Tan, H., Zhang, K., Xu, Z., Luan, F., Xu, Y., Hong, Y., Sunkavalli, K., Shakhnarovich, G., Bi, S.: Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 225, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 225, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 225, + 480, + 246 + ], + "type": "text", + "content": "16. Li, M., Zhou, P., Liu, J.W., Keppo, J., Lin, M., Yan, S., Xu, X.: Instant3d: Instant text-to-3d generation. arXiv preprint arXiv:2311.08403 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 246, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 246, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 133, + 246, + 480, + 277 + ], + "type": "text", + "content": "17. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 278, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 278, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 133, + 278, + 480, + 310 + ], + "type": "text", + "content": "18. Liu, M., Xu, C., Jin, H., Chen, L., Xu, Z., Su, H., et al.: One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 311, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 311, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 133, + 311, + 480, + 331 + ], + "type": "text", + "content": "19. Liu, R., Wu, R., Hoorick, B.V., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 332, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 332, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 133, + 332, + 480, + 364 + ], + "type": "text", + "content": "20. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. In: Seminal graphics: pioneering efforts that shaped the field, pp. 347-353 (1998)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 365, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 365, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 133, + 365, + 480, + 396 + ], + "type": "text", + "content": "21. Lorraine, J., Xie, K., Zeng, X., Lin, C.H., Takikawa, T., Sharp, N., Lin, T.Y., Liu, M.Y., Fidler, S., Lucas, J.: Att3d: Amortized text-to-3d object synthesis. In: Proceedings of the International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 396, + 480, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 396, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 133, + 396, + 480, + 428 + ], + "type": "text", + "content": "22. Lunz, S., Li, Y., Fitzgibbon, A., Kushner, N.: Inverse graphics gan: Learning to generate 3d shapes from unstructured 2d data. arXiv preprint arXiv:2002.12674 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 429, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 429, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 133, + 429, + 480, + 449 + ], + "type": "text", + "content": "23. Luo, T., Rockwell, C., Lee, H., Johnson, J.: Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 450, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 450, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 133, + 450, + 480, + 492 + ], + "type": "text", + "content": "24. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4460-4470 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 493, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 493, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 133, + 493, + 480, + 525 + ], + "type": "text", + "content": "25. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latentnerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 525, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 525, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 133, + 525, + 480, + 557 + ], + "type": "text", + "content": "26. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 558, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 558, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 133, + 558, + 480, + 590 + ], + "type": "text", + "content": "27. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: CVPR. pp. 8280-8290 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 133, + 590, + 480, + 611 + ], + "type": "text", + "content": "28. Nash, C., Ganin, Y., Eslami, S.M.A., Battaglia, P.W.: Polygen: An autoregressive generative model of 3d meshes (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 612, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 612, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 612, + 480, + 643 + ], + "type": "text", + "content": "29. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "text", + "content": "30. Oliphant, T.E.: Python for scientific computing. Computing in Science & Engineering (2007)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "K. Xie et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 138 + ], + "type": "text", + "content": "31. OpenAI: Gpt-4 technical report. ArXiv abs/2303.08774 (2023), https://api.sementicscholar.org/CorpusID:257532815" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "text", + "content": "32. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in PyTorch. Open-review (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "33. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "text", + "content": "34. Qian, G., Cao, J., Siarohin, A., Kant, Y., Wang, C., Vasilkovsky, M., Lee, H.Y., Fang, Y., Skorokhodov, I., Zhuang, P., et al.: Atom: Amortized text-to-mesh using 2d diffusion. arXiv preprint arXiv:2402.00867 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "text", + "content": "35. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "text", + "content": "36. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 346 + ], + "type": "text", + "content": "37. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 390 + ], + "type": "text", + "content": "38. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "text", + "content": "39. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 480, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 467 + ], + "type": "text", + "content": "40. Shen, T., Munkberg, J., Hasselgren, J., Yin, K., Wang, Z., Chen, W., Gojcic, Z., Fidler, S., Sharp, N., Gao, J.: Flexible isosurface extraction for gradient-based mesh optimization. ACM Trans. Graph. 42(4) (jul 2023). https://doi.org/10.1145/3592430, https://doi.org/10.1145/3592430" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "41. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "text", + "content": "42. Shi, Y., Wang, P., Ye, J., Mai, L., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv:2308.16512 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "43. Tang, J., Chen, Z., Chen, X., Wang, T., Zeng, G., Liu, Z.: Lgm: Large multiview gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "type": "text", + "content": "44. Tatarchenko, M., Dosovitskiy, A., Brox, T.: Octree generating networks: Efficient convolutional architectures for high-resolution 3d outputs. In: Proceedings of the IEEE international conference on computer vision. pp. 2088-2096 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "text", + "content": "45. Van Rossum, G., Drake Jr, F.L.: Python reference manual. Centrum voor Wiskunde en Informatica Amsterdam (1995)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "46. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "47. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "LATTE3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 293 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "48. Wu, J., Zhang, C., Xue, T., Freeman, B., Tenenbaum, J.: Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems 29 (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "type": "text", + "content": "49. Yang, G., Huang, X., Hao, Z., Liu, M.Y., Belongie, S., Hariharan, B.: Pointflow: 3d point cloud generation with continuous normalizing flows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "50. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "51. Zeng, X., Vahdat, A., Williams, F., Gojcic, Z., Litany, O., Fidler, S., Kreis, K.: Lion: Latent point diffusion models for 3d shape generation. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 258 + ], + "type": "text", + "content": "52. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "53. Zhou, L., Du, Y., Wu, J.: 3d shape generation and completion through point-voxel diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5826-5835 (2021)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "K. Xie et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LCM-Lookahead for Encoder-based Text-to-Image Personalization/bbfbe553-d1d9-40ca-841b-fd60b8f4aefe_content_list.json b/2024/LCM-Lookahead for Encoder-based Text-to-Image Personalization/bbfbe553-d1d9-40ca-841b-fd60b8f4aefe_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cc66097882a78af62b3c8fa6ad10b2bccc7a662e --- /dev/null +++ b/2024/LCM-Lookahead for Encoder-based Text-to-Image Personalization/bbfbe553-d1d9-40ca-841b-fd60b8f4aefe_content_list.json @@ -0,0 +1,1715 @@ +[ + { + "type": "text", + "text": "LCM-Lookahead for Encoder-based Text-to-Image Personalization", + "text_level": 1, + "bbox": [ + 294, + 140, + 707, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rinon Gal $^{1,2*}$ , Or Lichter $^{2*}$ , Elad Richardson $^{2*}$ , Or Patashnik $^{2}$ , Amit H. Bermano $^{2}$ , Gal Chechik $^{1}$ , and Daniel Cohen-Or $^{2}$", + "bbox": [ + 238, + 210, + 764, + 244 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a13dc909e404d86b8f75291139bdf566bcf4ad9c0c650dcd3a91e640798894e3.jpg", + "image_caption": [ + "$^{1}$ NVIDIA $^{2}$ Tel Aviv University", + "Fig.1: We introduce a novel LCM-based lookhead mechanism to apply image-space losses to personalization encoder training. These are coupled with consistent data generation and attention sharing techniques to tune existing backbones and improve identity preservation and prompt alignment." + ], + "image_footnote": [], + "bbox": [ + 250, + 282, + 751, + 549 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent advancements in diffusion models have introduced fast sampling methods that can effectively produce high-quality images in just one or a few denoising steps. Interestingly, when these are distilled from existing diffusion models, they often maintain alignment with the original model, retaining similar outputs for similar prompts and seeds. These properties present opportunities to leverage fast sampling methods as a shortcut-mechanism, using them to create a preview of denoised outputs through which we can backpropagate image-space losses. In this work, we explore the potential of using such shortcut-mechanisms to guide the personalization of text-to-image models to specific facial identities. We focus on encoder-based personalization approaches, and demonstrate that by augmenting their training with a lookahead identity loss, we can achieve higher identity fidelity, without sacrificing layout diversity or prompt alignment. Code at https://1cm-lookahead.github.io/.", + "bbox": [ + 259, + 645, + 743, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-image personalization [18,55] methods enable users to tailor pretrained generative models to their own, personal data. Commonly, such methods focus on human data [19,67,70,81,77,56], where users aim to create novel images of specific individuals which were unseen by the pretrained model. Early works proposed to tackle this task by teaching a model new words that describe the user-provided subjects. They do so by optimizing novel word-embeddings [18,68], or by fine-tuning the generative model itself [55,57]. However, such approaches require significant per-subject optimization, leading to lengthy personalization times and large compute requirements. More recent lines of work propose to personalize the model using an encoder - a neural network trained to condition the generative model on user-provided images [19,73,61,79]. While these methods can enable inference-time personalization, they often struggle to maintain a subject's identity, or face difficulties in adapting it to novel styles.", + "bbox": [ + 212, + 175, + 787, + 372 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One manner in which encoder-based methods try to bridge these gaps is by leveraging a pretrained face recognition network as a feature extraction backbone. The intuition here is that the pretraining objective of such networks drives them to encode fine identity details which can later be exploited by the encoder. However, this approach overlooks the training loss itself, which is still driven by only an L2 noise-prediction objective. In the realm of Generative Adversarial Networks (GANs, [21]), it was shown that inversion [1,76] can be significantly improved by incorporating additional losses that better align with human perception [53,45], e.g., an identity loss. Applying such image-space losses to the personalization process could be beneficial here as well. However, the diffusion training process works on noisy image samples from intermediate diffusion time steps, and it is not clear how these should be passed into a perceptual loss which expects clean, realistic images.", + "bbox": [ + 212, + 372, + 787, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here, we present a method for tackling this question by building on recent advancements in fast sampling methods [63,59,22], and specifically latent consistency models (LCM, [41,42]). To do so, we leverage an intriguing property of generative models: fine-tuning alignment, where a child model fine-tuned from a pretrained parent tends to preserve the semantics of the parent's latent space [75,20]. In fig. 2, we show a particular manifestation of this alignment in LCM models, where we compare a single-step LCM output to the DDPM single-step approximations at intermediate steps of the denoising process. The LCM results are not only sharper, but they maintain a high degree of similarity with the final DDPM prediction. This property also holds for personalized models (right, dog). We find this alignment holds particularly well in consistency models, likely by virtue of the consistency training loss itself.", + "bbox": [ + 212, + 568, + 787, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose to leverage this alignment during the training of personalization encoders, where we can create a high-quality preview of the denoiser's final output by doing a single LCM step on the noisy latents, guided by the same personalization encoder. This preview can then be used to calculate image-space losses, such as those derived from an identity detection network. Under this approach, the LCM-model provides a \"shortcut\" through which gradients can", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Gal et al.", + "bbox": [ + 271, + 114, + 336, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/61bf70257639b9b22168d156afcf386ac4b030d48d836a2baf6c8f6a60567de7.jpg", + "image_caption": [ + "Fig.2: LCM output alignment. We first denoise an image partway using DDPM [25] sampling with a baseline SDXL model [47]. We then complete sampling in two manners: (top) By performing a single LCM step. (bottom) By approximating the clean image using DDPM. Even at early steps, LCM outputs provide a good approximation of the final DDPM prediction. This also holds for personalized models (e.g., LoRA trained on the DreamBooth [55] dog, right)." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 792, + 262 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "back-propagate to earlier diffusion timesteps, without relying on low-quality approximations. In practice, we find that naive applications of the shortcut loss can break the alignment between the baseline and LCM models. Hence, we investigate mechanisms for preserving alignment, and show that these can improve downstream performance.", + "bbox": [ + 212, + 393, + 782, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition to the shortcut mechanism, we further explore an additional architectural modification inspired by recent video models and image editing works [74,11,33,50,10,43,24]. There, it has been shown that extending the self-attention mechanism, such that a generated image can also observe self-attention keys and values of a real, source image, can allow for zero-shot appearance transfer from the source to the new image [2,65]. Hence, we propose to augment the encoder with an additional path where the input image is noised, passed through a copy of the diffusion U-Net, and its self-attention keys and values are extracted and appended to the forward pass of the newly generated image.", + "bbox": [ + 212, + 469, + 784, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, we train our encoder by leveraging existing backbones [51,79] and tune it on newly generated data. To create our dataset, we leverage SDXL Turbo [59], a model distilled for single-step sampling from the vanilla SDXL using score distillation sampling [48] and adversarial training [21]. Notably, we observe that this distillation process causes a significant collapse of diversity, such that a sufficiently detailed prompt will generate the same identity regardless of the input seed. We leverage this property to generate a dataset of images with repeated identities and differing styles. Importantly, the use of generated data allows us to avoid the need to collect sensitive personal data, and ensures that our dataset contains proper representation for minority classes that are typically under-represented in real datasets.", + "bbox": [ + 212, + 604, + 785, + 770 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We validate our approach through comparisons to a range of recent baselines, and through a large set of ablation scenarios. These demonstrate that our approach can achieve higher identity fidelity and prompt alignment, and highlight the benefit of integrating image-space losses during model tuning.", + "bbox": [ + 212, + 771, + 785, + 833 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "LCM-Lookahead for Encoder-based Text-to-Image Personalization", + "bbox": [ + 289, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 143, + 382, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text-to-image Personalization. In the task of text-to-image personalization [18,55], the goal is to adapt a pretrained model to better represent a user-given concept which was unseen during the original training. Initial efforts show that this can be achieved by either optimizing a new text-embedding [18] or by further tuning the denoising network itself [55]. This has spurred a variety of different approaches that try to extend the optimized embedding space [4,68,17], restrict the tuned part of the denoising network [57,35,64,8,23], or optimize for specific target prompts [7]. A joint limitation of all these different approaches is that they require a per-subject optimization process which can be time-consuming.", + "bbox": [ + 212, + 176, + 787, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Encoder-Based Personalization. To overcome the limitations of optimization-based approaches, more recent work proposes that pretrained encoders can be used to initialize the optimization process, significantly shortening tuning times [19,56,6,36]. Others tried to completely avoid any optimization, at the cost of accuracy [73,12,61,79,30].", + "bbox": [ + 212, + 325, + 792, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently the problem of face personalization gained focus, with a range of works specifically tackling this domain with the goal of improving identity preservation across prompts [19,67,70,81,77,56]. In PhotoMaker [37] the features of a CLIP-model [51] are modulated using a dedicated ID-oriented dataset. Face0 [67] proposed to replace the CLIP encoder with an identity recognition network. IP-Adapter [79] introduced similar ideas in later variants, with identity network features being used either instead of, or on top of the CLIP image embedding. FaceStudio [78] follows a similar mechanism but also uses a prior model to better adhere to the original prompt. The concurrent InsantID paper builds on IP-Adapter and extends it with a landmark-conditioned ControlNet [82] to preserve pose and simplify identity preservation.", + "bbox": [ + 212, + 402, + 787, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Common to all these methods is that their training signal is based solely on the standard diffusion training loss. This is in contrast with the common practice in GAN-Based encoder methods [53,5,3,66,71,16], which demonstrated improved results by integrating perceptual losses, such as an identity loss. This gap can be attributed to the challenge of applying pixel-based losses during diffusion training. The concurrent PortraitBooth [46] proposes to apply an identity loss only to images sampled with low noise levels, where one can well approximate the clean image. However, this limits the effectiveness of the loss.", + "bbox": [ + 212, + 568, + 787, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our work is a similar tuning-free face-identity encoder, but we propose to integrate an identity loss into the model training by using an LCM-based shortcut mechanism.", + "bbox": [ + 212, + 688, + 787, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fast Diffusion Sampling. Standard diffusion models [25] and their text-to-image variants [54,44,52] are trained to denoise an image over a 1,000 steps. This span can be significantly shortened by leveraging ordinary differential equation solvers to navigate the diffusion flow in fewer steps [62,39,40,38,31].", + "bbox": [ + 212, + 750, + 787, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More recently, several methods were proposed to distill diffusion models into versions that can be sampled from in fewer steps [58]. The idea there is to teach", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Gal et al.", + "bbox": [ + 271, + 114, + 336, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the network to predict points farther along the diffusion flow, either by directly predicting a baseline model's output after multiple steps [58], through consistency modeling techniques [63,42,41,34], by leveraging adversarial training [21] and score sampling [48] methods [59] or by matching distributions between two diffusion models [80]. Of these, LCM-Lora [42] has shown that the consistency modeling approach can be applied through a low-rank adaptation [27] of the model weights.", + "bbox": [ + 212, + 146, + 782, + 251 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our work leverages such models, and LCM-LoRA in particular, in order to introduce image-space losses into personalization encoder training.", + "bbox": [ + 212, + 252, + 784, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 306, + 382, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Text-to-image Personalization", + "text_level": 1, + "bbox": [ + 215, + 337, + 509, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text-to-image personalization methods introduce new, user-provided concepts to a pretrained T2I diffusion model, typically by tuning novel word-embeddings, parameters of the denoising network itself, or mixture of both. Such optimization-based methods use a small (typically 3-5) image set depicting the concept, and tune the relevant parameters using the standard diffusion training loss [25]:", + "bbox": [ + 212, + 362, + 784, + 439 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {D i f f u s i o n} := \\mathbb {E} _ {z, y, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, y) \\| _ {2} ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 450, + 785, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon$ is an unscaled noise sample, $\\epsilon_{\\theta}$ is the denoising network, $t$ is the timestep, $z_{t}$ is an image or latent noised to time $t$ , and $y$ is some conditioning prompt containing a placeholder token which is used to describe the new concept.", + "bbox": [ + 212, + 484, + 782, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To speed up the personalization process, prior work suggested the use of encoders — neural networks trained to take images of the subject, and map them to some conditioning code that can guide the pretrained diffusion network to generate an image of the subject. A common approach [73,30] popularized by IP-Adapter [79] is to do so by extending the diffusion denoiser with additional cross-attention layers, which receive tokens derived from some external encoder module (e.g. a frozen CLIP-backbone [51] with a small projection head). The novel cross-attention heads and the encoder module are then trained using the loss of , where images are drawn from large-scale datasets depicting an array of subjects, commonly from a single domain (e.g. human faces).", + "bbox": [ + 212, + 531, + 787, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this case, the diffusion loss can be rewritten as:", + "bbox": [ + 238, + 681, + 602, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {D i f f u s i o n} := \\mathbb {E} _ {z _ {r}, y, I _ {c}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {r, t}, t, y, E \\left(I _ {c}\\right)\\right) \\| _ {2} ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 292, + 708, + 784, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I_{c}$ is a conditioning image sampled from the training set, $z_{r}$ is the image latent of a reconstruction target showing the same subject as $I_{c}$ , and $E(I_{c})$ is a conditioning code derived from the encoder.", + "bbox": [ + 212, + 743, + 782, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our work builds on a pretrained IP-Adapter, augments it with additional self-attention-based features, and tunes it to improve identity and prompt alignment.", + "bbox": [ + 212, + 787, + 782, + 819 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "LCM-Lookahead for Encoder-based Text-to-Image Personalization", + "bbox": [ + 289, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/305b407c3ef32fe7c83b68c788ff776dab3f5bfaaadf99ca57293a1ae3ffa8ce.jpg", + "image_caption": [ + "Fig. 3: (left) Encoder architecture: Our encoder has two branches: One is the standard IP-Adapter [79] that provides conditions through a new cross-attention head. The second branch is a copy of the SDXL U-net from which we extract self-attention keys and values, which we concatenate with those of the main denoising branch. (right) Training setup: The two encoder paths are provided with a conditioning image (and its noisy latent), and their outputs are used to condition the denoising of a different image of the same subject. We denoise the image with both the baseline SDXL [47] model and an LCM-model [42]. The baseline model's output is used for calculating the standard diffusion loss (eq. (2)). The LCM output is used to calculate the lookhead identity loss (eq. (4)). We portray latents as images for visual clarity." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 504, + 305 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2a925c7743153fab8f6f48b131f0ec76bc9626ef3e7abafb88281b22e5989a9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 143, + 784, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 508, + 330, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our goal is to improve encoder-based methods so that we can concurrently achieve identity preservation, and prompt alignment. For identity preservation, we introduce an LCM-based identity lookahead loss, and a self-attention-sharing based architectural modification for existing encoder pipelines. To improve prompt alignment, we utilize a synthetic dataset which contains consistent characters generated in an array of prompts and styles. fig. 3 provides a high-level outline of our encoder architecture (left) and training process (right). In the following section, we provide additional details on each of these components.", + "bbox": [ + 212, + 537, + 792, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 LCM-Lookahead loss", + "text_level": 1, + "bbox": [ + 215, + 680, + 434, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We begin with the goal of achieving high identity fidelity. Here, we would like to draw on core ideas from existing GAN inversion literature, and particularly the use of identity networks as a loss during encoder training [53,45]. In the GAN-based literature, applying such a loss is trivial since the GAN can produce a clean image in a single forward pass. In the case of diffusion models, the standard training process involves sampling random intermediate diffusion time steps, and performing a single denoising step. These single-step predictions are typically noisy or blurry, and feeding them into a downstream image embedding network has previously been shown to be sub-optimal [15,69]. These prior", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Gal et al.", + "bbox": [ + 271, + 114, + 336, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "works showed that one can improve the quality of guidance by training feature extractors on noisy images, or by performing multiple forward steps to create a clean image, and back-propagating through the entire diffusion chain. However, the first approach is expensive, and the second is impractical in training scenarios: both because such training is typically already memory constrained, and because end-to-end sampling increases iteration times by a factor of $\\sim 100$ , making training infeasible. Instead, we propose to utilize a pretrained LCM-LoRA model [42], tuned from the same baseline SDXL [47] backbone, to create higher-quality previews of the fully-denoised images using a single diffusion step. This preview can be fed into the downstream feature extractor, in our case a face recognition network [14], and gradients can be backpropagated to the encoder through this LCM-path. We focus on the LoRA variant of the LCM model as it best preserves the alignment with the baseline model.", + "bbox": [ + 212, + 146, + 782, + 342 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "More concretely, let $(I_c, I_r)$ be a conditioning image and reconstruction-target pair, let $z_{r,t}$ be the image latent $z_r$ , noised to an intermediary time step $t$ , and let $\\epsilon_{LCM}$ be the LCM denoising network. The preview image is given by:", + "bbox": [ + 212, + 342, + 782, + 388 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {r, 0} = \\frac {1}{\\sqrt {\\tilde {\\alpha} _ {t}}} \\left(z _ {r, t} - \\sqrt {\\tilde {\\beta} _ {t}} \\cdot \\epsilon_ {L C M} \\left(z _ {r, t}, y, t, E \\left(I _ {c}\\right)\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 398, + 784, + 430 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $y$ and $E(I_c)$ are the prompt and encoder conditioning codes respectively, $\\alpha_{t}$ and $\\beta_{t}$ are defined by the diffusion schedule. The lookahead loss is then:", + "bbox": [ + 214, + 439, + 782, + 469 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathscr {L} _ {L H} = \\mathscr {D} \\left(D _ {V A E} \\left(\\hat {z} _ {r, 0}\\right), I _ {c}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 481, + 784, + 497 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{D}$ is some image-space distance metric (e.g. an identity loss) and $D_{VAE}$ is the VAE decoder which maps the latents back to image space.", + "bbox": [ + 212, + 507, + 782, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Maintaining alignment", + "text_level": 1, + "bbox": [ + 214, + 559, + 447, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In initial experiments, we found that applying the lookahead loss of eq. (4) can improve identities over short-training runs. However, over time, this loss causes the LCM pathway to focus solely on the loss-metric at the expense of its prior output, breaking the alignment with the baseline model. We investigated a series of options for improving alignment preservation, including the use of existing distribution matching options like Score Distillation Sampling [48], or appending the standard Consistency Model loss [62] to the LCM-path. Full details and results from this investigation are presented in the supplementary. In practice, the best downstream performance was achieved using a model interpolation approach. There, for half of our training iterations, we randomly re-scale the LoRA component of the LCM-LoRA using $\\alpha_{LoRA} \\in [0.1, 1.0]$ . We hypothesize that applying the loss through the continuously interpolated model makes it more difficult for the encoder to converge to a solution which works differently for the LCM- and non-LCM paths. Moreover, this serves as a form of augmentation which makes adversarial solutions less likely. Lower values of $\\alpha_{LoRA}$ can still act as a preview for intermediate (but not overly noisy) outputs, which can still provide a signal through the image-space model.", + "bbox": [ + 212, + 582, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "LCM-Lookahead for Encoder-based Text-to-Image Personalization", + "bbox": [ + 289, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, similarly to Wallace et al. [69], we find it useful to focus on early (noisy) diffusion time steps. Here, we do so by applying annealing to the training time step sampling. We use the importance-weighting function of Huang et al. [29]: $f(t) = \\frac{1}{T} (1 - \\alpha \\cos \\frac{\\pi t}{T})$ , where $f(t)$ is the probability to sample time step $t$ , $T$ is the total number of diffusion time steps (i.e. 1,000) and $\\alpha$ is a hyperparameter which we empirically set to 0.2.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Extended self-attention features", + "text_level": 1, + "bbox": [ + 215, + 258, + 526, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As a second component for improving identity fidelity, we propose to leverage recent ideas in video-based modeling [74,11] and appearance transfer [2,24,10]. In these works, it was shown that expanding the self-attention mechanism such that a generated image can attend to the keys and values derived from a source image, can lead to a significant increase in visual similarity between the generated image and the source. Here, we use a similar idea to transfer identity features from the conditioning image to the generated output. These are applied on top of the baseline encoder from which we start.", + "bbox": [ + 212, + 281, + 787, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, we create a copy of the denoising U-Net which we call a \"KV encoder\". We pass a noisy version of our conditioning image through this U-Net, and cache the self-attention keys and values derived from this pass. Then, when performing the diffusion denoising pass, we append these keys and values to those derived from the denoised image at each self-attention layer: $K^{l} \\coloneqq K_{z_{r,t}}^{l} \\odot K_{z_{c,t}}^{l}$ , $V_{z_{r,t}}^{\\prime l} \\coloneqq V^{l} \\odot V_{z_{c,t}}^{l}$ , where $l$ is the layer index and $z_{r,t}, z_{c,t}$ subscripts denote attention features coming from the reconstruction target latent and the conditioning image latent respectively. This mechanism is illustrated in fig. 3.", + "bbox": [ + 214, + 402, + 787, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Directly applying this approach with a pretrained U-net as our encoder can lead to excessive appearance transfer and loss of editing. Hence, we do not keep the encoder frozen, but rather tune it using LoRA [27]. As our training set (detailed below) contains target images which differ in style from the source, this draws the network towards discarding appearance properties that are related to the style and not to the content, greatly improving prompt alignment. We note that a similar attention-expansion idea was used for personalization in the concurrent work of Purushwalkam et al. [49]. However, their target images are all photo-realistic, and their results do indeed suffer from reduced editability.", + "bbox": [ + 212, + 523, + 787, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, this pathway requires us to slightly modify the standard classifier-free guidance [26] equation. See the supplementary for more details.", + "bbox": [ + 212, + 659, + 785, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4 Consistent data generation", + "text_level": 1, + "bbox": [ + 215, + 710, + 485, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Having introduced components to improve identity preservation, we now turn to improving prompt alignment. Here, we hypothesize that the limited editability in current encoder-based methods is largely grounded in their training set, which typically focuses on reconstructing real images. Moreover, existing large-scale sets are either closed and proprietary or use the withdrawn LAION dataset [60]. They also commonly contain biases, which may result in models whose performance deteriorates on minority classes.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Gal et al.", + "bbox": [ + 271, + 114, + 336, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1bb1abab6a433fae4cb2410f361a143506d50fcf7b4c71cec2e96ae7a6b12cd8.jpg", + "image_caption": [ + "portrait photo" + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 320, + 210 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/7bcf26da4b1608e1b871bc46f43345f2e69a9b8e20e4d4a4b5bcae8ebaf88bdc.jpg", + "image_caption": [ + "drinking coffee" + ], + "image_footnote": [], + "bbox": [ + 321, + 143, + 410, + 212 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c366debd8e08ac0bf82180864e735afe7f806bf69e33375760349f32085e0a1f.jpg", + "image_caption": [ + "oil painting", + "Fig.4: Consistent Data. Consistent data generated using SDXL-Turbo with the description \"old man with curly hair and a moustache\" incorporated into different prompts (e.g. \"as an oil painting\", \"as a wanted poster\")" + ], + "image_footnote": [], + "bbox": [ + 413, + 143, + 500, + 212 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/05a28b5cabcbabceccb72cc75ca540c04c9743613c983d21e64f40598055fcc4.jpg", + "image_caption": [ + "wanted poster" + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 589, + 212 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8d6687628468c72a5e22554ede6d150edd4cd5947ac7b06f08b3aeac85cd04ca.jpg", + "image_caption": [ + "pencil drawing" + ], + "image_footnote": [], + "bbox": [ + 591, + 143, + 679, + 212 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/68deec775c59c5608fbf62bbd966dbfa36b9318f4dcf12d7815e28c90edb4958.jpg", + "image_caption": [ + "comic book" + ], + "image_footnote": [], + "bbox": [ + 681, + 143, + 767, + 212 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Instead of relying on such data, we propose to generate a novel consistent dataset in which we generated the same synthetic subjects across a wide range of prompts. These can then serve as training data for our encoder, where the encoder itself is provided with one image of a given identity, and the denoising goal considers another image portraying the same subject. Such cross-image training has already been shown to be beneficial with real data [30,37]. Here, the use of the generated data allows us to take the idea a step further, and ensure our reconstruction targets also contain stylized images. These can in turn prevent the encoder from focusing on photo-realism.", + "bbox": [ + 212, + 321, + 784, + 458 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To create our data, we investigate multiple approaches for consistent generation, including: (1) generating celebrities, which the model is already familiar with and can consistently generate across various prompts, (2) ConsiStory [65], a recently introduced approach that aligns identities through attention feature sharing, and (3) using SDXL-Turbo [59].", + "bbox": [ + 212, + 458, + 784, + 534 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While SDXL-Turbo is designed for fast sampling and not for consistent image generation, its adversarial training leads to mode-collapse. We find that, as a consequence, conditioning it on sufficiently detailed subject prompts will lead to a fixed identity across seeds and styles, as shown in fig. 4. We find this approach to achieve the best trade-off between generation time, identity consistency, and ability to change styles across the generated images. Hence, we use it to generate $500k$ images spanning roughly $100k$ identities. In practice, we trained for less than an epoch ( $40k$ identities in total). Further experiments and comparisons on consistent data generation can be found in the supplementary.", + "bbox": [ + 212, + 534, + 785, + 670 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.5 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 691, + 455, + 707 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We initialize our IP-Adapter backbone using the CLIP-based Face-ID model, and tune the same parameters as the original IP Adapter using a learning rate of $1e - 5$ . The encoder and denoiser U-Nets are initialized from a pretrained SDXL model. We tune the encoder U-Net using LoRA with rank 4 and a learning rate of $5e - 6$ . The decoder is kept frozen. For our LCM-Lookahead we use TinyVAE [9] to decode the latents. This lighter model reduces our memory consumption and also improves gradient flow during backpropagation. We tune the models over 5,000 iterations with a total batch size of 8 split across 2 NVIDIA A100 GPUs.", + "bbox": [ + 212, + 717, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "LCM-Lookahead for Encoder-based Text-to-Image Personalization", + "bbox": [ + 289, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3d58957a2c5e50c7c36d1925291bd566dac8e581fd2df77a3ca27eddf58cefd3.jpg", + "image_caption": [ + "Fig. 5: LCM-Lookahead Guidance. Results of classifier guidance when using different classifiers on top of our LCM-Lookahead, or standard $\\hat{x}_0$ approximation. Each classifier preserves different attributes of the guiding image. $\\hat{x}_0$ guidance may result in reduced quality or visible artifacts. Identity similarity values ( $\\uparrow$ , measured using [28]) are shown at the bottom." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 375 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 483, + 375, + 501 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Classifier guidance with LCM-Lookahead", + "text_level": 1, + "bbox": [ + 215, + 513, + 594, + 529 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We begin with an exploration of the LCM-Lookahead mechanism using a toy experiment, where we investigate its application to classifier guidance. Although this training-free method is less effective compared to encoder-based approaches, it serves as a simple use-case in which we can analyze the lookahead mechanism and discern its potential. Specifically, we follow [69] and apply repeated guidance iterations on an early diffusion time step ( $t = 44$ out of 50 DDIM [62] steps). At each iteration, we denoise the current latents using SDXL-LCM, decode them to an image, apply our pixel-space loss, and backpropagate to modify the latents. Final results are generated by continuing unguided-DDIM sampling for the remaining diffusion steps. Fig. 5 shows several outputs of such guidance, using a perceptual LPIPS loss [83], CLIP loss [51] and an identity loss [14]. Each loss preserves different attributes of the guiding image. LPIPS preserves the semantic structure of the image, CLIP preserves semantic attributes such as facial hair, and the identity loss explicitly improves facial similarity. When applying the same guidance to the single-step DDIM-approximations ( $\\hat{x}_0$ in fig. 5) we observe artifacts or reduced performance, attributed to the fact that $\\hat{x}_0$ is typically blurry and discolored in early timesteps. The identity loss is particularly robust to blur, but its performance still improves by using a lookhead loss.", + "bbox": [ + 212, + 537, + 787, + 809 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our encoder focuses on the lookahead identity loss to improve identity preservation during training.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Gal et al.", + "bbox": [ + 271, + 114, + 336, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/5dc27c536686dcbe86cdfce26f91d64f7f608fb58f146e20fd24f5c6a59086d1.jpg", + "image_caption": [ + "Fig. 6: Qualitative results. Our method personalizes a model to specific face identities at inference time, enabling both photo realistic and stylized prompts." + ], + "image_footnote": [], + "bbox": [ + 230, + 143, + 772, + 506 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Encoder evaluation", + "text_level": 1, + "bbox": [ + 215, + 571, + 419, + 585 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "fig. 6 shows a set of images synthesized using our encoder, across a range of identities and prompts. Our method can produce both photo-realistic results and stylized outputs, while largely preserving the subject's identity.", + "bbox": [ + 212, + 597, + 785, + 641 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To better gauge the quality of our results, we evaluate our method against a set of prior and concurrent works on personalized face generation. Specifically, we consider the three leading tuning-free approaches which have public SDXL implementations: IP-Adapter [79], InstantID [70], and PhotoMaker [37]. For IP-Adapter, we use the FaceID model that serves as our backbone, and evaluate it using two adapter-scale settings: The \"official\" value (1.0), and the one commonly used by the community (0.5) which significantly improves alignment with the textual prompts but harms identity.", + "bbox": [ + 212, + 643, + 787, + 763 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative Comparison. We first conduct a qualitative comparison. Prior art often shows results on well-known celebrities. However, we find that it is excessively easy to overfit on such identities (indeed, SDXL already contains tokens describing them, relegating an encoder's job to simply finding these tokens).", + "bbox": [ + 212, + 777, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "LCM-Lookahead for Encoder-based Text-to-Image Personalization", + "bbox": [ + 289, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9395643aac0f46f9ec3e2642ba35abb4a932511f0f6e8fd45575dc98c91a5d0e.jpg", + "image_caption": [ + "Fig. 7: Comparisons against prior and concurrent face-personalization encoders. IP-A (1.0) and (0.5) represent the IP-Adapter results with a scale of 1.0 and 0.5, respectively. Notably, IP-A (1.0) serves as the backbone which we fine-tune." + ], + "image_footnote": [], + "bbox": [ + 238, + 143, + 764, + 484 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Moreover, some recent papers test on identities contained in their training set (e.g. LAION-Faces contains images of famous researchers such as Yann Lecun). To paint a full picture, we provide such comparisons in the supplementary. Here, we instead collect a small set of 50 images with permissive licenses which were uploaded to https://unsplash.com/ over the period of Feb 19th - March 4th (2024). Our assumption is that these portray novel individuals which are less likely to exist in any prior training set, and thus offer the 'cleanest' benchmark.", + "bbox": [ + 212, + 563, + 787, + 667 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison results are shown in fig. 7. Notably, our method outperforms our IP-Adapter backbone in editability, while matching or exceeding it on identity preservation. InstantID achieves comparable results. However, it heavily restricts the target's pose to that of the conditioning image. Moreover, InstantID was trained on significantly more data and compute (60 million images and 48 GPUs over an undisclosed time-frame). We are hopeful that our approach could also be applied on top of an InstantID backbone and lead to improved results.", + "bbox": [ + 212, + 669, + 787, + 773 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In the supplementary, we further investigate the extent of diverse styles that can be achieved by our approach. While it does not yet reach the expressiveness offered by recent optimization based approaches, we observe that our method can handle varied styles including non photo-realistic rendering or mosaicing.", + "bbox": [ + 212, + 773, + 787, + 835 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Gal et al.", + "bbox": [ + 271, + 114, + 336, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Quantitative Comparison. Next, we move to quantitative evaluations. Here, we follow prior work [19] and compare the baselines across two metrics - identity similarity and prompt alignment. Our prompts include photo-realistic reconstructions, but also stylization and material change which prior art often struggles with. To measure identity similarity we use CurricularFace [28], which differs from both our loss network and from the backbones used to extract features for the baselines. Text alignment is measured using CLIP similarity (using the ViT-B/16 version which differs from the IP-Adapter backbone). We report both metrics using two sets: (1) 5,000 identities randomly sampled from the FFHQ dataset [32] and (2) The 50 unsplash identities. The results are shown in table 1. Our method is situated on the pareto front, providing good identity preservation and high editability. Note in particular that we outperform the backbone IP-Adapter in its typically used setup ( $\\alpha = 0.5$ ) on both metrics.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Since identity metrics are highly sensitive to both the success of editing (i.e. stylized images will have lower identity scores than failed edits), and to poses (which InstantID copies), we also verify our results using a user-study. There, for each question we showed users a reference image and a prompt pair, and the outputs of two models conditioned on this pair (ours, and a random baseline). We asked users to select the image that better preserves the reference identity and better aligns with the prompt. In total, we collected 460 responses from 43 different users. In table 3 we report the percentage of users that preferred our method over each baseline. The results largely align with the automatic metrics, showing that our approach is preferred to the backbone on which we build. InstantID still outperforms all methods, primarily on account of its much improved editability, but the margin is less severe.", + "bbox": [ + 212, + 343, + 787, + 523 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1911bf6b20b63281a0328262340335a5718a8303aeb60e41bf588bf00fb93154.jpg", + "table_caption": [ + "Table 1: Quantitative comparisons" + ], + "table_footnote": [ + "Table 3: User study results. For each matchup, we report the fraction of users who preferred our method, and the fraction that preferred the baseline." + ], + "table_body": "
FFHQ-5000Unsplash-50
ID ↑CLIP-T ↑ID ↑CLIP-T ↑
Ours0.34526.330.30826.79
IP-A (0.5)0.26825.820.25026.36
IP-A (1.0)0.36821.390.38722.06
PhotoMaker0.34426.690.21827.19
InstantID0.63128.580.61229.06
", + "bbox": [ + 215, + 574, + 486, + 660 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/3ca319455f3fbf59c05932bdea501d4fb25d2e78c884abb4310bde7fb4326e73.jpg", + "table_caption": [ + "Table 2: Ablation study" + ], + "table_footnote": [], + "table_body": "
FFHQ-5000Unsplash-50
ID ↑CLIP-T ↑ID ↑CLIP-T ↑
Baseline IPA0.36821.390.38722.06
+ Our Data0.22028.140.20528.25
w/ ID-Loss x00.28227.500.27227.74
w/ ID-Loss LCM0.30127.310.28127.74
+ KV Injection0.34526.330.30826.79
", + "bbox": [ + 500, + 574, + 797, + 660 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b206cd69673496c908b758e09e30835b9c2910fde10d0045bc5c41f0f926bf18.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PhotoMakerIP-A (1.0)IP-A (0.5)InstantID
Ours71.18%82.25%57.32%44.06%
Baseline28.82%17.75%42.68%55.94%
", + "bbox": [ + 269, + 700, + 727, + 739 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Ablation", + "text_level": 1, + "bbox": [ + 215, + 787, + 334, + 801 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We conduct an ablation study to evaluate the contribution of our suggested components, using the following setups: (1) The baseline IP-Adapter tuned on", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "LCM-Lookahead for Encoder-based Text-to-Image Personalization", + "bbox": [ + 289, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "our data, without extended attention or identity losses. (2) The setup of $(1) +$ an identity loss derived through standard $x0$ approximations. (3) The setup of $(1) +$ an identity loss derived through an LCM shortcut. (4) Our full model (the setup of $(3) +$ attention injection branch). For comparison, we again provide the backbone IP Adapter results.Quantitative results are provided in table 2. The use of our novel dataset greatly improves prompt alignment, at the cost of some identity fidelity. We attribute this in part to our smaller training scale, and to the fact that the identity score is impacted by stylization. Adding an identity loss significantly increases identity preservation, while passing this loss through the LCM shortcut provides a noticeably larger increase. Finally, injecting attention features leads to even better identity alignment, but at the cost of editability.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our results demonstrate that appending losses through an LCM-shortcut mechanism can provide improvements over the direct approximation approach. All-in-all, the combination of our components leads to a high level of both identity preservation and prompt alignment.", + "bbox": [ + 212, + 313, + 787, + 375 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Limitations and ethnic concerns", + "text_level": 1, + "bbox": [ + 214, + 396, + 552, + 412 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "While our model can improve existing encoders, it is not free of limitations. First, like prior tuning-free encoders, it falls short of the quality of optimization-based methods. This is particularly noticeable when working with inputs that are noticeably out-of-domain compared to standard face imagery (see supplementary).", + "bbox": [ + 212, + 428, + 785, + 488 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Secondly, our model may still suffer from biases inherent in both the backbone that we built on, as well as the diffusion model itself. Hence, it may serve to amplify social biases. Moreover, facial editing and generation software can be used to spread disinformation or defame individuals. Existing detection tools can help mitigate such risks [72,13], and we hope that these continue to improve.", + "bbox": [ + 212, + 488, + 787, + 566 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusions", + "text_level": 1, + "bbox": [ + 214, + 587, + 370, + 604 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we presented LCM-Lookahead, a novel mechanism for applying image-space losses to diffusion training using a fast-sampling-based shortcut mechanism. We then build on top of this mechanism to provide better identity signal to a personalization encoder, leading to improved identity preservation. Our work further explores the shortcomings of current personalization encoders and proposes two additional techniques to further improve their results. First, we show that consistent data generation methods can greatly impact prompt-alignment quality, and that SDXL-Turbo can serve to create such data. Indeed, fine-tuning on such data can even restore editability to encoders which have overfit on the photo-realistic domain. Finally, we show that the common self-attention key-value injection mechanism can also be applied to encoder-based personalization, improving the faithfulness of the generated results. We hope that both our LCM-Lookahead and our improved training scheme will serve to further push the boundaries of text-to-image personalization.", + "bbox": [ + 212, + 619, + 787, + 832 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Gal et al.", + "bbox": [ + 271, + 114, + 336, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 146, + 375, + 161 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was partially supported by ISF (grants 1337/22, 2492/20 and 3441/21).", + "bbox": [ + 215, + 171, + 787, + 186 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 210, + 323, + 224 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. State-of-the-art in the architecture, methods and applications of stylegan. In: Computer Graphics Forum. vol. 41, pp. 591-611. Wiley Online Library (2022)", + "2. Alaluf, Y., Garibi, D., Patashnik, O., Averbuch-Elor, H., Cohen-Or, D.: Cross-image attention for zero-shot appearance transfer (2023)", + "3. Alaluf, Y., Patashnik, O., Cohen-Or, D.: Restyle: A residual-based stylegan encoder via iterative refinement. arXiv preprint arXiv:2104.02699 (2021)", + "4. Alaluf, Y., Richardson, E., Metzer, G., Cohen-Or, D.: A neural space-time representation for text-to-image personalization. ACM Transactions on Graphics (TOG) 42(6), 1-10 (2023)", + "5. Alaluf, Y., Tov, O., Mokady, R., Gal, R., Bermano, A.H.: Hyperstyle: Stylegan inversion with hypernetworks for real image editing (2021)", + "6. Arar, M., Gal, R., Atzmon, Y., Chechik, G., Cohen-Or, D., Shamir, A., H. Bermano, A.: Domain-agnostic tuning-encoder for fast personalization of text-to-image models. In: SIGGRAPH Asia 2023 Conference Papers. pp. 1–10 (2023)", + "7. Arar, M., Voynov, A., Hertz, A., Avrahami, O., Fruchter, S., Pritch, Y., Cohen-Or, D., Shamir, A.: Palp: Prompt aligned personalization of text-to-image models. arXiv preprint arXiv:2401.06105 (2024)", + "8. Avrahami, O., Aberman, K., Fried, O., Cohen-Or, D., Lischinski, D.: Break-ascene: Extracting multiple concepts from a single image. In: SIGGRAPH Asia 2023 Conference Papers. SA '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3610548.3618154, https://doi.org/10.1145/3610548.3618154", + "9. Bohan, O.B.: Tiny autoencoder for stable diffusion. https://github.com/madebyollin/taesd (2023)", + "10. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 22560-22570 (October 2023)", + "11. Ceylan, D., Huang, C.H.P., Mitra, N.J.: Pix2video: Video editing using image diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23206-23217 (2023)", + "12. Chen, W., Hu, H., LI, Y., Ruiz, N., Jia, X., Chang, M.W., Cohen, W.W.: Subject-driven text-to-image generation via apprenticeship learning. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=wv3bHyQbX7", + "13. Corvi, R., Cozzolino, D., Zingarini, G., Poggi, G., Nagano, K., Verdoliva, L.: On the detection of synthetic images generated by diffusion models. In: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5 (2023). https://doi.org/10.1109/ICASSP49357.2023.10095167", + "14. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition (2019)", + "15. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021)" + ], + "bbox": [ + 218, + 241, + 784, + 837 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "LCM-Lookahead for Encoder-based Text-to-Image Personalization", + "bbox": [ + 289, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Dinh, T.M., Tran, A.T., Nguyen, R., Hua, B.S.: Hyperinverter: Improving stylegan inversion via hypernetwork. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11389-11398 (2022)", + "17. Dong, Z., Wei, P., Lin, L.: Dreamartist: Towards controllable one-shot text-to-image generation via contrastive prompt-tuning. arXiv preprint arXiv:2211.11337 (2022)", + "18. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion (2022). https://doi.org/10.48550/ARXIV.2208.01618, https://arxiv.org/abs/2208.01618", + "19. Gal, R., Arar, M., Atzmon, Y., Bermano, A.H., Chechik, G., Cohen-Or, D.: Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG) 42(4), 1-13 (2023)", + "20. Gal, R., Patashnik, O., Maron, H., Chechik, G., Cohen-Or, D.: Stylegan-nada: Clipguided domain adaptation of image generators. arXiv preprint arXiv:2108.00946 (2021)", + "21. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)", + "22. Gu, J., Zhai, S., Zhang, Y., Liu, L., Susskind, J.M.: Boot: Data-free distillation of denoising diffusion models with bootstrapping. In: ICML 2023 Workshop on Structured Probabilistic Inference $\\{\\backslash\\&\\}$ Generative Modeling (2023)", + "23. Han, L., Li, Y., Zhang, H., Milanfar, P., Metaxas, D., Yang, F.: Svdiff: Compact parameter space for diffusion fine-tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7323-7334 (October 2023)", + "24. Hertz, A., Voynov, A., Fruchter, S., Cohen-Or, D.: Style aligned image generation via shared attention (2023)", + "25. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020)", + "26. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications (2021)", + "27. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Chen, W.: Lora: Low-rank adaptation of large language models. ArXiv abs/2106.09685 (2021)", + "28. Huang, Y., Wang, Y., Tai, Y., Liu, X., Shen, P., Li, S., Li, J., Huang, F.: Curricularface: adaptive curriculum learning loss for deep face recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5901-5910 (2020)", + "29. Huang, Z., Wu, T., Jiang, Y., Chan, K.C., Liu, Z.: Reversion: Diffusion-based relation inversion from images. arXiv preprint arXiv:2303.13495 (2023)", + "30. Jia, X., Zhao, Y., Chan, K.C., Li, Y., Zhang, H., Gong, B., Hou, T., Wang, H., Su, Y.C.: Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642 (2023)", + "31. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems 35, 26565-26577 (2022)", + "32. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4401-4410 (2019)", + "33. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Gal et al.", + "bbox": [ + 271, + 114, + 336, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "shot video generators. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15954-15964 (October 2023)", + "34. Kim, D., Lai, C.H., Liao, W.H., Murata, N., Takida, Y., Uesaka, T., He, Y., Mitsufuji, Y., Ermon, S.: Consistency trajectory models: Learning probability flow ODE trajectory of diffusion. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=ymjI8feDTD", + "35. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. arXiv (2022)", + "36. Li, D., Li, J., Hoi, S.C.: Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720 (2023)", + "37. Li, Z., Cao, M., Wang, X., Qi, Z., Cheng, M.M., Shan, Y.: Photomaker: Customizing realistic human photos via stacked id embedding. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2024)", + "38. Liu, L., Ren, Y., Lin, Z., Zhao, Z.: Pseudo numerical methods for diffusion models on manifolds. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=PlKWd2yBkY", + "39. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: DPM-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=2uAaGwlp_V", + "40. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: DPM-solver++: Fast solver for guided sampling of diffusion probabilistic models (2023), https://openreview.net/forum?id=4vGwQqviud5", + "41. Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthesizing high-resolution images with few-step inference (2023)", + "42. Luo, S., Tan, Y., Patil, S., Gu, D., von Platen, P., Passos, A., Huang, L., Li, J., Zhao, H.: Lcm-lora: A universal stable-diffusion acceleration module (2023)", + "43. Mou, C., Wang, X., Song, J., Shan, Y., Zhang, J.: Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421 (2023)", + "44. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)", + "45. Nitzan, Y., Bermano, A., Li, Y., Cohen-Or, D.: Face identity disentanglement via latent space mapping. ACM Transactions on Graphics (TOG) 39(6), 1-14 (2020)", + "46. Peng, X., Zhu, J., Jiang, B., Tai, Y., Luo, D., Zhang, J., Lin, W., Jin, T., Wang, C., Ji, R.: Portrait booth: A versatile portrait model for fast identity-preserved personalization. arXiv preprint arXiv:2312.06354 (2023)", + "47. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: SDXL: Improving latent diffusion models for high-resolution image synthesis. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=di52zR8xgf", + "48. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=FjNys5c7VvY", + "49. Purushwalkam, S., Gokul, A., Joty, S., Naik, N.: Bootpig: Bootstrapping zero-shot personalized image generation capabilities in pretrained diffusion models (2024)", + "50. QI, C., Cun, X., Zhang, Y., Lei, C., Wang, X., Shan, Y., Chen, Q.: Fatezero: Fusing attentions for zero-shot text-based video editing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15932-15942 (October 2023)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "LCM-Lookahead for Encoder-based Text-to-Image Personalization", + "bbox": [ + 289, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "51. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)", + "52. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)", + "53. Richardson, E., Alaluf, Y., Patashnik, O., Nitzan, Y., Azar, Y., Shapiro, S., Cohen-Or, D.: Encoding in style: a stylegan encoder for image-to-image translation. arXiv preprint arXiv:2008.00951 (2020)", + "54. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021)", + "55. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation (2022)", + "56. Ruiz, N., Li, Y., Jampani, V., Wei, W., Hou, T., Pritch, Y., Wadhwa, N., Rubinstein, M., Aberman, K.: Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models (2023)", + "57. Ryu, S.: Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora (2023)", + "58. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=TIdIXIpzhoI", + "59. Sauer, A., Lorenz, D., Blattmann, A., Rombach, R.: Adversarial diffusion distillation (2023)", + "60. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)", + "61. Shi, J., Xiong, W., Lin, Z., Jung, H.J.: Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411 (2023)", + "62. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: International Conference on Learning Representations (2020)", + "63. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models (2023)", + "64. Tewel, Y., Gal, R., Chechik, G., Atzmon, Y.: Key-locked rank one editing for text-to-image personalization. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-11 (2023)", + "65. Tewel, Y., Kaduri, O., Gal, R., Kasten, Y., Wolf, L., Chechik, G., Atzmon, Y.: Training-free consistent text-to-image generation (2024)", + "66. Tov, O., Alaluf, Y., Nitzan, Y., Patashnik, O., Cohen-Or, D.: Designing an encoder for stylegan image manipulation. arXiv preprint arXiv:2102.02766 (2021)", + "67. Valevski, D., Lumen, D., Matias, Y., Leviathan, Y.: Face0: Instantaneously conditioning a text-to-image model on a face. In: SIGGRAPH Asia 2023 Conference Papers. SA '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3610548.3618249, https://doi.org/10.1145/3610548.3618249", + "68. Voynov, A., Chu, Q., Cohen-Or, D., Aberman, K.: $p+$ : Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522 (2023)", + "69. Wallace, B., Gokul, A., Ermon, S., Naik, N.: End-to-end diffusion latent optimization improves classifier guidance. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7280-7290 (October 2023)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Gal et al.", + "bbox": [ + 271, + 114, + 336, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "70. Wang, Q., Bai, X., Wang, H., Qin, Z., Chen, A.: Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519 (2024)", + "71. Wang, T., Zhang, Y., Fan, Y., Wang, J., Chen, Q.: High-fidelity gan inversion for image attribute editing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "72. Wang, Z., Bao, J., Zhou, W., Wang, W., Hu, H., Chen, H., Li, H.: Dire for diffusion-generated image detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 22445-22455 (October 2023)", + "73. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15943-15953 (October 2023)", + "74. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7623-7633 (2023)", + "75. Wu, Z., Nitzan, Y., Shechtman, E., Lischinski, D.: Stylealign: Analysis and applications of aligned stylegan models (2021)", + "76. Xia, W., Zhang, Y., Yang, Y., Xue, J.H., Zhou, B., Yang, M.H.: Gan inversion: A survey (2021)", + "77. Xiao, G., Yin, T., Freeman, W.T., Durand, F., Han, S.: Fastcomposer: Tuning-free multi-subject image generation with localized attention. arXiv (2023)", + "78. Yan, Y., Zhang, C., Wang, R., Cheng, P., Yu, G., Fu, B.: Facestudio: Put your face everywhere in seconds (2023)", + "79. Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721 (2023)", + "80. Yin, T., Gharbi, M., Zhang, R., Shechtman, E., Durand, F., Freeman, W.T., Park, T.: One-step diffusion with distribution matching distillation. CVPR (2024)", + "81. Yuan, G., Cun, X., Zhang, Y., Li, M., Qi, C., Wang, X., Shan, Y., Zheng, H.: Inserting anybody in diffusion models via celeb basis. arXiv preprint arXiv:2306.00926 (2023)", + "82. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023)", + "83. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + ], + "bbox": [ + 212, + 146, + 787, + 660 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "LCM-Lookahead for Encoder-based Text-to-Image Personalization", + "bbox": [ + 289, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/LCM-Lookahead for Encoder-based Text-to-Image Personalization/bbfbe553-d1d9-40ca-841b-fd60b8f4aefe_model.json b/2024/LCM-Lookahead for Encoder-based Text-to-Image Personalization/bbfbe553-d1d9-40ca-841b-fd60b8f4aefe_model.json new file mode 100644 index 0000000000000000000000000000000000000000..41ede883f011942aa49a5769ce416bdbf9577b3b --- /dev/null +++ b/2024/LCM-Lookahead for Encoder-based Text-to-Image Personalization/bbfbe553-d1d9-40ca-841b-fd60b8f4aefe_model.json @@ -0,0 +1,2680 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.295, + 0.141, + 0.709, + 0.187 + ], + "angle": 0, + "content": "LCM-Lookahead for Encoder-based Text-to-Image Personalization" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.212, + 0.765, + 0.245 + ], + "angle": 0, + "content": "Rinon Gal\\(^{1,2*}\\), Or Lichter\\(^{2*}\\), Elad Richardson\\(^{2*}\\), Or Patashnik\\(^{2}\\), Amit H. Bermano\\(^{2}\\), Gal Chechik\\(^{1}\\), and Daniel Cohen-Or\\(^{2}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.397, + 0.255, + 0.607, + 0.271 + ], + "angle": 0, + "content": "\\(^{1}\\)NVIDIA \\(^{2}\\)Tel Aviv University" + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.283, + 0.753, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.563, + 0.788, + 0.625 + ], + "angle": 0, + "content": "Fig.1: We introduce a novel LCM-based lookhead mechanism to apply image-space losses to personalization encoder training. These are coupled with consistent data generation and attention sharing techniques to tune existing backbones and improve identity preservation and prompt alignment." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.646, + 0.744, + 0.842 + ], + "angle": 0, + "content": "Abstract. Recent advancements in diffusion models have introduced fast sampling methods that can effectively produce high-quality images in just one or a few denoising steps. Interestingly, when these are distilled from existing diffusion models, they often maintain alignment with the original model, retaining similar outputs for similar prompts and seeds. These properties present opportunities to leverage fast sampling methods as a shortcut-mechanism, using them to create a preview of denoised outputs through which we can backpropagate image-space losses. In this work, we explore the potential of using such shortcut-mechanisms to guide the personalization of text-to-image models to specific facial identities. We focus on encoder-based personalization approaches, and demonstrate that by augmenting their training with a lookahead identity loss, we can achieve higher identity fidelity, without sacrificing layout diversity or prompt alignment. Code at https://1cm-lookahead.github.io/." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.337, + 0.128 + ], + "angle": 0, + "content": "Gal et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.376, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.373 + ], + "angle": 0, + "content": "Text-to-image personalization [18,55] methods enable users to tailor pretrained generative models to their own, personal data. Commonly, such methods focus on human data [19,67,70,81,77,56], where users aim to create novel images of specific individuals which were unseen by the pretrained model. Early works proposed to tackle this task by teaching a model new words that describe the user-provided subjects. They do so by optimizing novel word-embeddings [18,68], or by fine-tuning the generative model itself [55,57]. However, such approaches require significant per-subject optimization, leading to lengthy personalization times and large compute requirements. More recent lines of work propose to personalize the model using an encoder - a neural network trained to condition the generative model on user-provided images [19,73,61,79]. While these methods can enable inference-time personalization, they often struggle to maintain a subject's identity, or face difficulties in adapting it to novel styles." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.788, + 0.569 + ], + "angle": 0, + "content": "One manner in which encoder-based methods try to bridge these gaps is by leveraging a pretrained face recognition network as a feature extraction backbone. The intuition here is that the pretraining objective of such networks drives them to encode fine identity details which can later be exploited by the encoder. However, this approach overlooks the training loss itself, which is still driven by only an L2 noise-prediction objective. In the realm of Generative Adversarial Networks (GANs, [21]), it was shown that inversion [1,76] can be significantly improved by incorporating additional losses that better align with human perception [53,45], e.g., an identity loss. Applying such image-space losses to the personalization process could be beneficial here as well. However, the diffusion training process works on noisy image samples from intermediate diffusion time steps, and it is not clear how these should be passed into a perceptual loss which expects clean, realistic images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Here, we present a method for tackling this question by building on recent advancements in fast sampling methods [63,59,22], and specifically latent consistency models (LCM, [41,42]). To do so, we leverage an intriguing property of generative models: fine-tuning alignment, where a child model fine-tuned from a pretrained parent tends to preserve the semantics of the parent's latent space [75,20]. In fig. 2, we show a particular manifestation of this alignment in LCM models, where we compare a single-step LCM output to the DDPM single-step approximations at intermediate steps of the denoising process. The LCM results are not only sharper, but they maintain a high degree of similarity with the final DDPM prediction. This property also holds for personalized models (right, dog). We find this alignment holds particularly well in consistency models, likely by virtue of the consistency training loss itself." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We propose to leverage this alignment during the training of personalization encoders, where we can create a high-quality preview of the denoiser's final output by doing a single LCM step on the noisy latents, guided by the same personalization encoder. This preview can then be used to calculate image-space losses, such as those derived from an identity detection network. Under this approach, the LCM-model provides a \"shortcut\" through which gradients can" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.29, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "LCM-Lookahead for Encoder-based Text-to-Image Personalization" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.794, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.366 + ], + "angle": 0, + "content": "Fig.2: LCM output alignment. We first denoise an image partway using DDPM [25] sampling with a baseline SDXL model [47]. We then complete sampling in two manners: (top) By performing a single LCM step. (bottom) By approximating the clean image using DDPM. Even at early steps, LCM outputs provide a good approximation of the final DDPM prediction. This also holds for personalized models (e.g., LoRA trained on the DreamBooth [55] dog, right)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.784, + 0.469 + ], + "angle": 0, + "content": "back-propagate to earlier diffusion timesteps, without relying on low-quality approximations. In practice, we find that naive applications of the shortcut loss can break the alignment between the baseline and LCM models. Hence, we investigate mechanisms for preserving alignment, and show that these can improve downstream performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.785, + 0.606 + ], + "angle": 0, + "content": "In addition to the shortcut mechanism, we further explore an additional architectural modification inspired by recent video models and image editing works [74,11,33,50,10,43,24]. There, it has been shown that extending the self-attention mechanism, such that a generated image can also observe self-attention keys and values of a real, source image, can allow for zero-shot appearance transfer from the source to the new image [2,65]. Hence, we propose to augment the encoder with an additional path where the input image is noised, passed through a copy of the diffusion U-Net, and its self-attention keys and values are extracted and appended to the forward pass of the newly generated image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.606, + 0.786, + 0.771 + ], + "angle": 0, + "content": "Finally, we train our encoder by leveraging existing backbones [51,79] and tune it on newly generated data. To create our dataset, we leverage SDXL Turbo [59], a model distilled for single-step sampling from the vanilla SDXL using score distillation sampling [48] and adversarial training [21]. Notably, we observe that this distillation process causes a significant collapse of diversity, such that a sufficiently detailed prompt will generate the same identity regardless of the input seed. We leverage this property to generate a dataset of images with repeated identities and differing styles. Importantly, the use of generated data allows us to avoid the need to collect sensitive personal data, and ensures that our dataset contains proper representation for minority classes that are typically under-represented in real datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.772, + 0.786, + 0.834 + ], + "angle": 0, + "content": "We validate our approach through comparisons to a range of recent baselines, and through a large set of ablation scenarios. These demonstrate that our approach can achieve higher identity fidelity and prompt alignment, and highlight the benefit of integrating image-space losses during model tuning." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.337, + 0.128 + ], + "angle": 0, + "content": "Gal et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.383, + 0.161 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Text-to-image Personalization. In the task of text-to-image personalization [18,55], the goal is to adapt a pretrained model to better represent a user-given concept which was unseen during the original training. Initial efforts show that this can be achieved by either optimizing a new text-embedding [18] or by further tuning the denoising network itself [55]. This has spurred a variety of different approaches that try to extend the optimized embedding space [4,68,17], restrict the tuned part of the denoising network [57,35,64,8,23], or optimize for specific target prompts [7]. A joint limitation of all these different approaches is that they require a per-subject optimization process which can be time-consuming." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.794, + 0.403 + ], + "angle": 0, + "content": "Encoder-Based Personalization. To overcome the limitations of optimization-based approaches, more recent work proposes that pretrained encoders can be used to initialize the optimization process, significantly shortening tuning times [19,56,6,36]. Others tried to completely avoid any optimization, at the cost of accuracy [73,12,61,79,30]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.788, + 0.569 + ], + "angle": 0, + "content": "Recently the problem of face personalization gained focus, with a range of works specifically tackling this domain with the goal of improving identity preservation across prompts [19,67,70,81,77,56]. In PhotoMaker [37] the features of a CLIP-model [51] are modulated using a dedicated ID-oriented dataset. Face0 [67] proposed to replace the CLIP encoder with an identity recognition network. IP-Adapter [79] introduced similar ideas in later variants, with identity network features being used either instead of, or on top of the CLIP image embedding. FaceStudio [78] follows a similar mechanism but also uses a prior model to better adhere to the original prompt. The concurrent InsantID paper builds on IP-Adapter and extends it with a landmark-conditioned ControlNet [82] to preserve pose and simplify identity preservation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Common to all these methods is that their training signal is based solely on the standard diffusion training loss. This is in contrast with the common practice in GAN-Based encoder methods [53,5,3,66,71,16], which demonstrated improved results by integrating perceptual losses, such as an identity loss. This gap can be attributed to the challenge of applying pixel-based losses during diffusion training. The concurrent PortraitBooth [46] proposes to apply an identity loss only to images sampled with low noise levels, where one can well approximate the clean image. However, this limits the effectiveness of the loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Our work is a similar tuning-free face-identity encoder, but we propose to integrate an identity loss into the model training by using an LCM-based shortcut mechanism." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Fast Diffusion Sampling. Standard diffusion models [25] and their text-to-image variants [54,44,52] are trained to denoise an image over a 1,000 steps. This span can be significantly shortened by leveraging ordinary differential equation solvers to navigate the diffusion flow in fewer steps [62,39,40,38,31]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "More recently, several methods were proposed to distill diffusion models into versions that can be sampled from in fewer steps [58]. The idea there is to teach" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.29, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "LCM-Lookahead for Encoder-based Text-to-Image Personalization" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.252 + ], + "angle": 0, + "content": "the network to predict points farther along the diffusion flow, either by directly predicting a baseline model's output after multiple steps [58], through consistency modeling techniques [63,42,41,34], by leveraging adversarial training [21] and score sampling [48] methods [59] or by matching distributions between two diffusion models [80]. Of these, LCM-Lora [42] has shown that the consistency modeling approach can be applied through a low-rank adaptation [27] of the model weights." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.285 + ], + "angle": 0, + "content": "Our work leverages such models, and LCM-LoRA in particular, in order to introduce image-space losses into personalization encoder training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.307, + 0.383, + 0.323 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.338, + 0.51, + 0.354 + ], + "angle": 0, + "content": "3.1 Text-to-image Personalization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.785, + 0.44 + ], + "angle": 0, + "content": "Text-to-image personalization methods introduce new, user-provided concepts to a pretrained T2I diffusion model, typically by tuning novel word-embeddings, parameters of the denoising network itself, or mixture of both. Such optimization-based methods use a small (typically 3-5) image set depicting the concept, and tune the relevant parameters using the standard diffusion training loss [25]:" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.451, + 0.786, + 0.477 + ], + "angle": 0, + "content": "\\[\nL _ {D i f f u s i o n} := \\mathbb {E} _ {z, y, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, y) \\| _ {2} ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.784, + 0.531 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is an unscaled noise sample, \\(\\epsilon_{\\theta}\\) is the denoising network, \\(t\\) is the timestep, \\(z_{t}\\) is an image or latent noised to time \\(t\\), and \\(y\\) is some conditioning prompt containing a placeholder token which is used to describe the new concept." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.532, + 0.788, + 0.682 + ], + "angle": 0, + "content": "To speed up the personalization process, prior work suggested the use of encoders — neural networks trained to take images of the subject, and map them to some conditioning code that can guide the pretrained diffusion network to generate an image of the subject. A common approach [73,30] popularized by IP-Adapter [79] is to do so by extending the diffusion denoiser with additional cross-attention layers, which receive tokens derived from some external encoder module (e.g. a frozen CLIP-backbone [51] with a small projection head). The novel cross-attention heads and the encoder module are then trained using the loss of , where images are drawn from large-scale datasets depicting an array of subjects, commonly from a single domain (e.g. human faces)." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.683, + 0.604, + 0.697 + ], + "angle": 0, + "content": "In this case, the diffusion loss can be rewritten as:" + }, + { + "type": "equation", + "bbox": [ + 0.293, + 0.709, + 0.785, + 0.734 + ], + "angle": 0, + "content": "\\[\nL _ {D i f f u s i o n} := \\mathbb {E} _ {z _ {r}, y, I _ {c}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {r, t}, t, y, E \\left(I _ {c}\\right)\\right) \\| _ {2} ^ {2} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.744, + 0.784, + 0.789 + ], + "angle": 0, + "content": "where \\( I_{c} \\) is a conditioning image sampled from the training set, \\( z_{r} \\) is the image latent of a reconstruction target showing the same subject as \\( I_{c} \\), and \\( E(I_{c}) \\) is a conditioning code derived from the encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.789, + 0.784, + 0.82 + ], + "angle": 0, + "content": "Our work builds on a pretrained IP-Adapter, augments it with additional self-attention-based features, and tunes it to improve identity and prompt alignment." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.128 + ], + "angle": 0, + "content": "Gal et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.505, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.144, + 0.785, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.316, + 0.79, + 0.483 + ], + "angle": 0, + "content": "Fig. 3: (left) Encoder architecture: Our encoder has two branches: One is the standard IP-Adapter [79] that provides conditions through a new cross-attention head. The second branch is a copy of the SDXL U-net from which we extract self-attention keys and values, which we concatenate with those of the main denoising branch. (right) Training setup: The two encoder paths are provided with a conditioning image (and its noisy latent), and their outputs are used to condition the denoising of a different image of the same subject. We denoise the image with both the baseline SDXL [47] model and an LCM-model [42]. The baseline model's output is used for calculating the standard diffusion loss (eq. (2)). The LCM output is used to calculate the lookhead identity loss (eq. (4)). We portray latents as images for visual clarity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.509, + 0.331, + 0.525 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.794, + 0.661 + ], + "angle": 0, + "content": "Our goal is to improve encoder-based methods so that we can concurrently achieve identity preservation, and prompt alignment. For identity preservation, we introduce an LCM-based identity lookahead loss, and a self-attention-sharing based architectural modification for existing encoder pipelines. To improve prompt alignment, we utilize a synthetic dataset which contains consistent characters generated in an array of prompts and styles. fig. 3 provides a high-level outline of our encoder architecture (left) and training process (right). In the following section, we provide additional details on each of these components." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.681, + 0.436, + 0.696 + ], + "angle": 0, + "content": "4.1 LCM-Lookahead loss" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.842 + ], + "angle": 0, + "content": "We begin with the goal of achieving high identity fidelity. Here, we would like to draw on core ideas from existing GAN inversion literature, and particularly the use of identity networks as a loss during encoder training [53,45]. In the GAN-based literature, applying such a loss is trivial since the GAN can produce a clean image in a single forward pass. In the case of diffusion models, the standard training process involves sampling random intermediate diffusion time steps, and performing a single denoising step. These single-step predictions are typically noisy or blurry, and feeding them into a downstream image embedding network has previously been shown to be sub-optimal [15,69]. These prior" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.29, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "LCM-Lookahead for Encoder-based Text-to-Image Personalization" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.343 + ], + "angle": 0, + "content": "works showed that one can improve the quality of guidance by training feature extractors on noisy images, or by performing multiple forward steps to create a clean image, and back-propagating through the entire diffusion chain. However, the first approach is expensive, and the second is impractical in training scenarios: both because such training is typically already memory constrained, and because end-to-end sampling increases iteration times by a factor of \\(\\sim 100\\), making training infeasible. Instead, we propose to utilize a pretrained LCM-LoRA model [42], tuned from the same baseline SDXL [47] backbone, to create higher-quality previews of the fully-denoised images using a single diffusion step. This preview can be fed into the downstream feature extractor, in our case a face recognition network [14], and gradients can be backpropagated to the encoder through this LCM-path. We focus on the LoRA variant of the LCM model as it best preserves the alignment with the baseline model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.784, + 0.389 + ], + "angle": 0, + "content": "More concretely, let \\((I_c, I_r)\\) be a conditioning image and reconstruction-target pair, let \\(z_{r,t}\\) be the image latent \\(z_r\\), noised to an intermediary time step \\(t\\), and let \\(\\epsilon_{LCM}\\) be the LCM denoising network. The preview image is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.399, + 0.785, + 0.431 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {r, 0} = \\frac {1}{\\sqrt {\\tilde {\\alpha} _ {t}}} \\left(z _ {r, t} - \\sqrt {\\tilde {\\beta} _ {t}} \\cdot \\epsilon_ {L C M} \\left(z _ {r, t}, y, t, E \\left(I _ {c}\\right)\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.44, + 0.784, + 0.47 + ], + "angle": 0, + "content": "where \\(y\\) and \\(E(I_c)\\) are the prompt and encoder conditioning codes respectively, \\(\\alpha_{t}\\) and \\(\\beta_{t}\\) are defined by the diffusion schedule. The lookahead loss is then:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.482, + 0.785, + 0.498 + ], + "angle": 0, + "content": "\\[\n\\mathscr {L} _ {L H} = \\mathscr {D} \\left(D _ {V A E} \\left(\\hat {z} _ {r, 0}\\right), I _ {c}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.784, + 0.539 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}\\) is some image-space distance metric (e.g. an identity loss) and \\(D_{VAE}\\) is the VAE decoder which maps the latents back to image space." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.56, + 0.449, + 0.575 + ], + "angle": 0, + "content": "4.2 Maintaining alignment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In initial experiments, we found that applying the lookahead loss of eq. (4) can improve identities over short-training runs. However, over time, this loss causes the LCM pathway to focus solely on the loss-metric at the expense of its prior output, breaking the alignment with the baseline model. We investigated a series of options for improving alignment preservation, including the use of existing distribution matching options like Score Distillation Sampling [48], or appending the standard Consistency Model loss [62] to the LCM-path. Full details and results from this investigation are presented in the supplementary. In practice, the best downstream performance was achieved using a model interpolation approach. There, for half of our training iterations, we randomly re-scale the LoRA component of the LCM-LoRA using \\(\\alpha_{LoRA} \\in [0.1, 1.0]\\). We hypothesize that applying the loss through the continuously interpolated model makes it more difficult for the encoder to converge to a solution which works differently for the LCM- and non-LCM paths. Moreover, this serves as a form of augmentation which makes adversarial solutions less likely. Lower values of \\(\\alpha_{LoRA}\\) can still act as a preview for intermediate (but not overly noisy) outputs, which can still provide a signal through the image-space model." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.337, + 0.128 + ], + "angle": 0, + "content": "Gal et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Finally, similarly to Wallace et al. [69], we find it useful to focus on early (noisy) diffusion time steps. Here, we do so by applying annealing to the training time step sampling. We use the importance-weighting function of Huang et al. [29]: \\( f(t) = \\frac{1}{T} (1 - \\alpha \\cos \\frac{\\pi t}{T}) \\), where \\( f(t) \\) is the probability to sample time step \\( t \\), \\( T \\) is the total number of diffusion time steps (i.e. 1,000) and \\( \\alpha \\) is a hyperparameter which we empirically set to 0.2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.259, + 0.527, + 0.273 + ], + "angle": 0, + "content": "4.3 Extended self-attention features" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.282, + 0.788, + 0.403 + ], + "angle": 0, + "content": "As a second component for improving identity fidelity, we propose to leverage recent ideas in video-based modeling [74,11] and appearance transfer [2,24,10]. In these works, it was shown that expanding the self-attention mechanism such that a generated image can attend to the keys and values derived from a source image, can lead to a significant increase in visual similarity between the generated image and the source. Here, we use a similar idea to transfer identity features from the conditioning image to the generated output. These are applied on top of the baseline encoder from which we start." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.403, + 0.788, + 0.524 + ], + "angle": 0, + "content": "Specifically, we create a copy of the denoising U-Net which we call a \"KV encoder\". We pass a noisy version of our conditioning image through this U-Net, and cache the self-attention keys and values derived from this pass. Then, when performing the diffusion denoising pass, we append these keys and values to those derived from the denoised image at each self-attention layer: \\( K^{l} \\coloneqq K_{z_{r,t}}^{l} \\odot K_{z_{c,t}}^{l} \\), \\( V_{z_{r,t}}^{\\prime l} \\coloneqq V^{l} \\odot V_{z_{c,t}}^{l} \\), where \\( l \\) is the layer index and \\( z_{r,t}, z_{c,t} \\) subscripts denote attention features coming from the reconstruction target latent and the conditioning image latent respectively. This mechanism is illustrated in fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.659 + ], + "angle": 0, + "content": "Directly applying this approach with a pretrained U-net as our encoder can lead to excessive appearance transfer and loss of editing. Hence, we do not keep the encoder frozen, but rather tune it using LoRA [27]. As our training set (detailed below) contains target images which differ in style from the source, this draws the network towards discarding appearance properties that are related to the style and not to the content, greatly improving prompt alignment. We note that a similar attention-expansion idea was used for personalization in the concurrent work of Purushwalkam et al. [49]. However, their target images are all photo-realistic, and their results do indeed suffer from reduced editability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.787, + 0.691 + ], + "angle": 0, + "content": "Finally, this pathway requires us to slightly modify the standard classifier-free guidance [26] equation. See the supplementary for more details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.711, + 0.486, + 0.726 + ], + "angle": 0, + "content": "4.4 Consistent data generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Having introduced components to improve identity preservation, we now turn to improving prompt alignment. Here, we hypothesize that the limited editability in current encoder-based methods is largely grounded in their training set, which typically focuses on reconstructing real images. Moreover, existing large-scale sets are either closed and proprietary or use the withdrawn LAION dataset [60]. They also commonly contain biases, which may result in models whose performance deteriorates on minority classes." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.29, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "LCM-Lookahead for Encoder-based Text-to-Image Personalization" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.144, + 0.321, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.215, + 0.301, + 0.236 + ], + "angle": 0, + "content": "portrait photo" + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.145, + 0.411, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.215, + 0.392, + 0.235 + ], + "angle": 0, + "content": "drinking coffee" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.145, + 0.501, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.433, + 0.215, + 0.482, + 0.236 + ], + "angle": 0, + "content": "oil painting" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.145, + 0.59, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.215, + 0.567, + 0.235 + ], + "angle": 0, + "content": "wanted poster" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.145, + 0.68, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.215, + 0.661, + 0.235 + ], + "angle": 0, + "content": "pencil drawing" + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.144, + 0.768, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.709, + 0.215, + 0.743, + 0.235 + ], + "angle": 0, + "content": "comic book" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.247, + 0.785, + 0.294 + ], + "angle": 0, + "content": "Fig.4: Consistent Data. Consistent data generated using SDXL-Turbo with the description \"old man with curly hair and a moustache\" incorporated into different prompts (e.g. \"as an oil painting\", \"as a wanted poster\")" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.785, + 0.459 + ], + "angle": 0, + "content": "Instead of relying on such data, we propose to generate a novel consistent dataset in which we generated the same synthetic subjects across a wide range of prompts. These can then serve as training data for our encoder, where the encoder itself is provided with one image of a given identity, and the denoising goal considers another image portraying the same subject. Such cross-image training has already been shown to be beneficial with real data [30,37]. Here, the use of the generated data allows us to take the idea a step further, and ensure our reconstruction targets also contain stylized images. These can in turn prevent the encoder from focusing on photo-realism." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.785, + 0.535 + ], + "angle": 0, + "content": "To create our data, we investigate multiple approaches for consistent generation, including: (1) generating celebrities, which the model is already familiar with and can consistently generate across various prompts, (2) ConsiStory [65], a recently introduced approach that aligns identities through attention feature sharing, and (3) using SDXL-Turbo [59]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.787, + 0.671 + ], + "angle": 0, + "content": "While SDXL-Turbo is designed for fast sampling and not for consistent image generation, its adversarial training leads to mode-collapse. We find that, as a consequence, conditioning it on sufficiently detailed subject prompts will lead to a fixed identity across seeds and styles, as shown in fig. 4. We find this approach to achieve the best trade-off between generation time, identity consistency, and ability to change styles across the generated images. Hence, we use it to generate \\(500k\\) images spanning roughly \\(100k\\) identities. In practice, we trained for less than an epoch (\\(40k\\) identities in total). Further experiments and comparisons on consistent data generation can be found in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.693, + 0.457, + 0.708 + ], + "angle": 0, + "content": "4.5 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We initialize our IP-Adapter backbone using the CLIP-based Face-ID model, and tune the same parameters as the original IP Adapter using a learning rate of \\( 1e - 5 \\). The encoder and denoiser U-Nets are initialized from a pretrained SDXL model. We tune the encoder U-Net using LoRA with rank 4 and a learning rate of \\( 5e - 6 \\). The decoder is kept frozen. For our LCM-Lookahead we use TinyVAE [9] to decode the latents. This lighter model reduces our memory consumption and also improves gradient flow during backpropagation. We tune the models over 5,000 iterations with a total batch size of 8 split across 2 NVIDIA A100 GPUs." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.127 + ], + "angle": 0, + "content": "Gal et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.389, + 0.789, + 0.466 + ], + "angle": 0, + "content": "Fig. 5: LCM-Lookahead Guidance. Results of classifier guidance when using different classifiers on top of our LCM-Lookahead, or standard \\(\\hat{x}_0\\) approximation. Each classifier preserves different attributes of the guiding image. \\(\\hat{x}_0\\) guidance may result in reduced quality or visible artifacts. Identity similarity values (\\(\\uparrow\\), measured using [28]) are shown at the bottom." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.484, + 0.377, + 0.502 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.514, + 0.596, + 0.53 + ], + "angle": 0, + "content": "5.1 Classifier guidance with LCM-Lookahead" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.81 + ], + "angle": 0, + "content": "We begin with an exploration of the LCM-Lookahead mechanism using a toy experiment, where we investigate its application to classifier guidance. Although this training-free method is less effective compared to encoder-based approaches, it serves as a simple use-case in which we can analyze the lookahead mechanism and discern its potential. Specifically, we follow [69] and apply repeated guidance iterations on an early diffusion time step (\\(t = 44\\) out of 50 DDIM [62] steps). At each iteration, we denoise the current latents using SDXL-LCM, decode them to an image, apply our pixel-space loss, and backpropagate to modify the latents. Final results are generated by continuing unguided-DDIM sampling for the remaining diffusion steps. Fig. 5 shows several outputs of such guidance, using a perceptual LPIPS loss [83], CLIP loss [51] and an identity loss [14]. Each loss preserves different attributes of the guiding image. LPIPS preserves the semantic structure of the image, CLIP preserves semantic attributes such as facial hair, and the identity loss explicitly improves facial similarity. When applying the same guidance to the single-step DDIM-approximations (\\(\\hat{x}_0\\) in fig. 5) we observe artifacts or reduced performance, attributed to the fact that \\(\\hat{x}_0\\) is typically blurry and discolored in early timesteps. The identity loss is particularly robust to blur, but its performance still improves by using a lookhead loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Our encoder focuses on the lookahead identity loss to improve identity preservation during training." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.29, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "LCM-Lookahead for Encoder-based Text-to-Image Personalization" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.144, + 0.774, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.519, + 0.785, + 0.551 + ], + "angle": 0, + "content": "Fig. 6: Qualitative results. Our method personalizes a model to specific face identities at inference time, enabling both photo realistic and stylized prompts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.573, + 0.421, + 0.587 + ], + "angle": 0, + "content": "5.2 Encoder evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.787, + 0.642 + ], + "angle": 0, + "content": "fig. 6 shows a set of images synthesized using our encoder, across a range of identities and prompts. Our method can produce both photo-realistic results and stylized outputs, while largely preserving the subject's identity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.765 + ], + "angle": 0, + "content": "To better gauge the quality of our results, we evaluate our method against a set of prior and concurrent works on personalized face generation. Specifically, we consider the three leading tuning-free approaches which have public SDXL implementations: IP-Adapter [79], InstantID [70], and PhotoMaker [37]. For IP-Adapter, we use the FaceID model that serves as our backbone, and evaluate it using two adapter-scale settings: The \"official\" value (1.0), and the one commonly used by the community (0.5) which significantly improves alignment with the textual prompts but harms identity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.779, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Qualitative Comparison. We first conduct a qualitative comparison. Prior art often shows results on well-known celebrities. However, we find that it is excessively easy to overfit on such identities (indeed, SDXL already contains tokens describing them, relegating an encoder's job to simply finding these tokens)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.127 + ], + "angle": 0, + "content": "Gal et al." + }, + { + "type": "image", + "bbox": [ + 0.239, + 0.145, + 0.765, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.498, + 0.788, + 0.545 + ], + "angle": 0, + "content": "Fig. 7: Comparisons against prior and concurrent face-personalization encoders. IP-A (1.0) and (0.5) represent the IP-Adapter results with a scale of 1.0 and 0.5, respectively. Notably, IP-A (1.0) serves as the backbone which we fine-tune." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.669 + ], + "angle": 0, + "content": "Moreover, some recent papers test on identities contained in their training set (e.g. LAION-Faces contains images of famous researchers such as Yann Lecun). To paint a full picture, we provide such comparisons in the supplementary. Here, we instead collect a small set of 50 images with permissive licenses which were uploaded to https://unsplash.com/ over the period of Feb 19th - March 4th (2024). Our assumption is that these portray novel individuals which are less likely to exist in any prior training set, and thus offer the 'cleanest' benchmark." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.67, + 0.788, + 0.775 + ], + "angle": 0, + "content": "Comparison results are shown in fig. 7. Notably, our method outperforms our IP-Adapter backbone in editability, while matching or exceeding it on identity preservation. InstantID achieves comparable results. However, it heavily restricts the target's pose to that of the conditioning image. Moreover, InstantID was trained on significantly more data and compute (60 million images and 48 GPUs over an undisclosed time-frame). We are hopeful that our approach could also be applied on top of an InstantID backbone and lead to improved results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.775, + 0.788, + 0.837 + ], + "angle": 0, + "content": "In the supplementary, we further investigate the extent of diverse styles that can be achieved by our approach. While it does not yet reach the expressiveness offered by recent optimization based approaches, we observe that our method can handle varied styles including non photo-realistic rendering or mosaicing." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.29, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LCM-Lookahead for Encoder-based Text-to-Image Personalization" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "Quantitative Comparison. Next, we move to quantitative evaluations. Here, we follow prior work [19] and compare the baselines across two metrics - identity similarity and prompt alignment. Our prompts include photo-realistic reconstructions, but also stylization and material change which prior art often struggles with. To measure identity similarity we use CurricularFace [28], which differs from both our loss network and from the backbones used to extract features for the baselines. Text alignment is measured using CLIP similarity (using the ViT-B/16 version which differs from the IP-Adapter backbone). We report both metrics using two sets: (1) 5,000 identities randomly sampled from the FFHQ dataset [32] and (2) The 50 unsplash identities. The results are shown in table 1. Our method is situated on the pareto front, providing good identity preservation and high editability. Note in particular that we outperform the backbone IP-Adapter in its typically used setup (\\(\\alpha = 0.5\\)) on both metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.525 + ], + "angle": 0, + "content": "Since identity metrics are highly sensitive to both the success of editing (i.e. stylized images will have lower identity scores than failed edits), and to poses (which InstantID copies), we also verify our results using a user-study. There, for each question we showed users a reference image and a prompt pair, and the outputs of two models conditioned on this pair (ours, and a random baseline). We asked users to select the image that better preserves the reference identity and better aligns with the prompt. In total, we collected 460 responses from 43 different users. In table 3 we report the percentage of users that preferred our method over each baseline. The results largely align with the automatic metrics, showing that our approach is preferred to the backbone on which we build. InstantID still outperforms all methods, primarily on account of its much improved editability, but the margin is less severe." + }, + { + "type": "table_caption", + "bbox": [ + 0.227, + 0.55, + 0.477, + 0.565 + ], + "angle": 0, + "content": "Table 1: Quantitative comparisons" + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.575, + 0.487, + 0.661 + ], + "angle": 0, + "content": "
FFHQ-5000Unsplash-50
ID ↑CLIP-T ↑ID ↑CLIP-T ↑
Ours0.34526.330.30826.79
IP-A (0.5)0.26825.820.25026.36
IP-A (1.0)0.36821.390.38722.06
PhotoMaker0.34426.690.21827.19
InstantID0.63128.580.61229.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.556, + 0.55, + 0.731, + 0.565 + ], + "angle": 0, + "content": "Table 2: Ablation study" + }, + { + "type": "table", + "bbox": [ + 0.5, + 0.575, + 0.798, + 0.661 + ], + "angle": 0, + "content": "
FFHQ-5000Unsplash-50
ID ↑CLIP-T ↑ID ↑CLIP-T ↑
Baseline IPA0.36821.390.38722.06
+ Our Data0.22028.140.20528.25
w/ ID-Loss x00.28227.500.27227.74
w/ ID-Loss LCM0.30127.310.28127.74
+ KV Injection0.34526.330.30826.79
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.216, + 0.662, + 0.787, + 0.692 + ], + "angle": 0, + "content": "Table 3: User study results. For each matchup, we report the fraction of users who preferred our method, and the fraction that preferred the baseline." + }, + { + "type": "table", + "bbox": [ + 0.271, + 0.702, + 0.728, + 0.74 + ], + "angle": 0, + "content": "
PhotoMakerIP-A (1.0)IP-A (0.5)InstantID
Ours71.18%82.25%57.32%44.06%
Baseline28.82%17.75%42.68%55.94%
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.335, + 0.802 + ], + "angle": 0, + "content": "5.3 Ablation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We conduct an ablation study to evaluate the contribution of our suggested components, using the following setups: (1) The baseline IP-Adapter tuned on" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.129 + ], + "angle": 0, + "content": "Gal et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "our data, without extended attention or identity losses. (2) The setup of \\((1) +\\) an identity loss derived through standard \\(x0\\) approximations. (3) The setup of \\((1) +\\) an identity loss derived through an LCM shortcut. (4) Our full model (the setup of \\((3) +\\) attention injection branch). For comparison, we again provide the backbone IP Adapter results.Quantitative results are provided in table 2. The use of our novel dataset greatly improves prompt alignment, at the cost of some identity fidelity. We attribute this in part to our smaller training scale, and to the fact that the identity score is impacted by stylization. Adding an identity loss significantly increases identity preservation, while passing this loss through the LCM shortcut provides a noticeably larger increase. Finally, injecting attention features leads to even better identity alignment, but at the cost of editability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Our results demonstrate that appending losses through an LCM-shortcut mechanism can provide improvements over the direct approximation approach. All-in-all, the combination of our components leads to a high level of both identity preservation and prompt alignment." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.397, + 0.553, + 0.414 + ], + "angle": 0, + "content": "6 Limitations and ethnic concerns" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.787, + 0.489 + ], + "angle": 0, + "content": "While our model can improve existing encoders, it is not free of limitations. First, like prior tuning-free encoders, it falls short of the quality of optimization-based methods. This is particularly noticeable when working with inputs that are noticeably out-of-domain compared to standard face imagery (see supplementary)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.567 + ], + "angle": 0, + "content": "Secondly, our model may still suffer from biases inherent in both the backbone that we built on, as well as the diffusion model itself. Hence, it may serve to amplify social biases. Moreover, facial editing and generation software can be used to spread disinformation or defame individuals. Existing detection tools can help mitigate such risks [72,13], and we hope that these continue to improve." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.588, + 0.371, + 0.605 + ], + "angle": 0, + "content": "7 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.788, + 0.833 + ], + "angle": 0, + "content": "In this work, we presented LCM-Lookahead, a novel mechanism for applying image-space losses to diffusion training using a fast-sampling-based shortcut mechanism. We then build on top of this mechanism to provide better identity signal to a personalization encoder, leading to improved identity preservation. Our work further explores the shortcomings of current personalization encoders and proposes two additional techniques to further improve their results. First, we show that consistent data generation methods can greatly impact prompt-alignment quality, and that SDXL-Turbo can serve to create such data. Indeed, fine-tuning on such data can even restore editability to encoders which have overfit on the photo-realistic domain. Finally, we show that the common self-attention key-value injection mechanism can also be applied to encoder-based personalization, improving the faithfulness of the generated results. We hope that both our LCM-Lookahead and our improved training scheme will serve to further push the boundaries of text-to-image personalization." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.29, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LCM-Lookahead for Encoder-based Text-to-Image Personalization" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.147, + 0.377, + 0.162 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.172, + 0.788, + 0.188 + ], + "angle": 0, + "content": "This work was partially supported by ISF (grants 1337/22, 2492/20 and 3441/21)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.211, + 0.324, + 0.226 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.242, + 0.785, + 0.27 + ], + "angle": 0, + "content": "1. State-of-the-art in the architecture, methods and applications of stylegan. In: Computer Graphics Forum. vol. 41, pp. 591-611. Wiley Online Library (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.271, + 0.785, + 0.297 + ], + "angle": 0, + "content": "2. Alaluf, Y., Garibi, D., Patashnik, O., Averbuch-Elor, H., Cohen-Or, D.: Cross-image attention for zero-shot appearance transfer (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.298, + 0.785, + 0.325 + ], + "angle": 0, + "content": "3. Alaluf, Y., Patashnik, O., Cohen-Or, D.: Restyle: A residual-based stylegan encoder via iterative refinement. arXiv preprint arXiv:2104.02699 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.327, + 0.785, + 0.367 + ], + "angle": 0, + "content": "4. Alaluf, Y., Richardson, E., Metzer, G., Cohen-Or, D.: A neural space-time representation for text-to-image personalization. ACM Transactions on Graphics (TOG) 42(6), 1-10 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.368, + 0.785, + 0.394 + ], + "angle": 0, + "content": "5. Alaluf, Y., Tov, O., Mokady, R., Gal, R., Bermano, A.H.: Hyperstyle: Stylegan inversion with hypernetworks for real image editing (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.395, + 0.785, + 0.436 + ], + "angle": 0, + "content": "6. Arar, M., Gal, R., Atzmon, Y., Chechik, G., Cohen-Or, D., Shamir, A., H. Bermano, A.: Domain-agnostic tuning-encoder for fast personalization of text-to-image models. In: SIGGRAPH Asia 2023 Conference Papers. pp. 1–10 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.437, + 0.785, + 0.477 + ], + "angle": 0, + "content": "7. Arar, M., Voynov, A., Hertz, A., Avrahami, O., Fruchter, S., Pritch, Y., Cohen-Or, D., Shamir, A.: Palp: Prompt aligned personalization of text-to-image models. arXiv preprint arXiv:2401.06105 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.478, + 0.785, + 0.545 + ], + "angle": 0, + "content": "8. Avrahami, O., Aberman, K., Fried, O., Cohen-Or, D., Lischinski, D.: Break-ascene: Extracting multiple concepts from a single image. In: SIGGRAPH Asia 2023 Conference Papers. SA '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3610548.3618154, https://doi.org/10.1145/3610548.3618154" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.547, + 0.785, + 0.574 + ], + "angle": 0, + "content": "9. Bohan, O.B.: Tiny autoencoder for stable diffusion. https://github.com/madebyollin/taesd (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.576, + 0.785, + 0.629 + ], + "angle": 0, + "content": "10. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 22560-22570 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.63, + 0.785, + 0.671 + ], + "angle": 0, + "content": "11. Ceylan, D., Huang, C.H.P., Mitra, N.J.: Pix2video: Video editing using image diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23206-23217 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.672, + 0.785, + 0.727 + ], + "angle": 0, + "content": "12. Chen, W., Hu, H., LI, Y., Ruiz, N., Jia, X., Chang, M.W., Cohen, W.W.: Subject-driven text-to-image generation via apprenticeship learning. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=wv3bHyQbX7" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.728, + 0.785, + 0.782 + ], + "angle": 0, + "content": "13. Corvi, R., Cozzolino, D., Zingarini, G., Poggi, G., Nagano, K., Verdoliva, L.: On the detection of synthetic images generated by diffusion models. In: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5 (2023). https://doi.org/10.1109/ICASSP49357.2023.10095167" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.783, + 0.785, + 0.809 + ], + "angle": 0, + "content": "14. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.81, + 0.785, + 0.838 + ], + "angle": 0, + "content": "15. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.242, + 0.785, + 0.838 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.128 + ], + "angle": 0, + "content": "Gal et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "16. Dinh, T.M., Tran, A.T., Nguyen, R., Hua, B.S.: Hyperinverter: Improving stylegan inversion via hypernetwork. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11389-11398 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "17. Dong, Z., Wei, P., Lin, L.: Dreamartist: Towards controllable one-shot text-to-image generation via contrastive prompt-tuning. arXiv preprint arXiv:2211.11337 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.286 + ], + "angle": 0, + "content": "18. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion (2022). https://doi.org/10.48550/ARXIV.2208.01618, https://arxiv.org/abs/2208.01618" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.328 + ], + "angle": 0, + "content": "19. Gal, R., Arar, M., Atzmon, Y., Bermano, A.H., Chechik, G., Cohen-Or, D.: Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG) 42(4), 1-13 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "20. Gal, R., Patashnik, O., Maron, H., Chechik, G., Cohen-Or, D.: Stylegan-nada: Clipguided domain adaptation of image generators. arXiv preprint arXiv:2108.00946 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.411 + ], + "angle": 0, + "content": "21. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.787, + 0.453 + ], + "angle": 0, + "content": "22. Gu, J., Zhai, S., Zhang, Y., Liu, L., Susskind, J.M.: Boot: Data-free distillation of denoising diffusion models with bootstrapping. In: ICML 2023 Workshop on Structured Probabilistic Inference \\(\\{\\backslash\\&\\}\\) Generative Modeling (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.495 + ], + "angle": 0, + "content": "23. Han, L., Li, Y., Zhang, H., Milanfar, P., Metaxas, D., Yang, F.: Svdiff: Compact parameter space for diffusion fine-tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7323-7334 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.522 + ], + "angle": 0, + "content": "24. Hertz, A., Voynov, A., Fruchter, S., Cohen-Or, D.: Style aligned image generation via shared attention (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.55 + ], + "angle": 0, + "content": "25. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.577 + ], + "angle": 0, + "content": "26. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.605 + ], + "angle": 0, + "content": "27. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Chen, W.: Lora: Low-rank adaptation of large language models. ArXiv abs/2106.09685 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.66 + ], + "angle": 0, + "content": "28. Huang, Y., Wang, Y., Tai, Y., Liu, X., Shen, P., Li, S., Li, J., Huang, F.: Curricularface: adaptive curriculum learning loss for deep face recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5901-5910 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.787, + 0.688 + ], + "angle": 0, + "content": "29. Huang, Z., Wu, T., Jiang, Y., Chan, K.C., Liu, Z.: Reversion: Diffusion-based relation inversion from images. arXiv preprint arXiv:2303.13495 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.73 + ], + "angle": 0, + "content": "30. Jia, X., Zhao, Y., Chan, K.C., Li, Y., Zhang, H., Gong, B., Hou, T., Wang, H., Su, Y.C.: Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.787, + 0.77 + ], + "angle": 0, + "content": "31. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems 35, 26565-26577 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.787, + 0.813 + ], + "angle": 0, + "content": "32. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4401-4410 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "33. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.29, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LCM-Lookahead for Encoder-based Text-to-Image Personalization" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.148, + 0.786, + 0.175 + ], + "angle": 0, + "content": "shot video generators. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15954-15964 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.787, + 0.23 + ], + "angle": 0, + "content": "34. Kim, D., Lai, C.H., Liao, W.H., Murata, N., Takida, Y., Uesaka, T., He, Y., Mitsufuji, Y., Ermon, S.: Consistency trajectory models: Learning probability flow ODE trajectory of diffusion. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=ymjI8feDTD" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.786, + 0.256 + ], + "angle": 0, + "content": "35. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. arXiv (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.786, + 0.297 + ], + "angle": 0, + "content": "36. Li, D., Li, J., Hoi, S.C.: Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.297, + 0.786, + 0.338 + ], + "angle": 0, + "content": "37. Li, Z., Cao, M., Wang, X., Qi, Z., Cheng, M.M., Shan, Y.: Photomaker: Customizing realistic human photos via stacked id embedding. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.786, + 0.379 + ], + "angle": 0, + "content": "38. Liu, L., Ren, Y., Lin, Z., Zhao, Z.: Pseudo numerical methods for diffusion models on manifolds. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=PlKWd2yBkY" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.786, + 0.433 + ], + "angle": 0, + "content": "39. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: DPM-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=2uAaGwlp_V" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.786, + 0.473 + ], + "angle": 0, + "content": "40. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: DPM-solver++: Fast solver for guided sampling of diffusion probabilistic models (2023), https://openreview.net/forum?id=4vGwQqviud5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.786, + 0.501 + ], + "angle": 0, + "content": "41. Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthesizing high-resolution images with few-step inference (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.786, + 0.528 + ], + "angle": 0, + "content": "42. Luo, S., Tan, Y., Patil, S., Gu, D., von Platen, P., Passos, A., Huang, L., Li, J., Zhao, H.: Lcm-lora: A universal stable-diffusion acceleration module (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.786, + 0.555 + ], + "angle": 0, + "content": "43. Mou, C., Wang, X., Song, J., Shan, Y., Zhang, J.: Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.786, + 0.595 + ], + "angle": 0, + "content": "44. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.786, + 0.623 + ], + "angle": 0, + "content": "45. Nitzan, Y., Bermano, A., Li, Y., Cohen-Or, D.: Face identity disentanglement via latent space mapping. ACM Transactions on Graphics (TOG) 39(6), 1-14 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.786, + 0.663 + ], + "angle": 0, + "content": "46. Peng, X., Zhu, J., Jiang, B., Tai, Y., Luo, D., Zhang, J., Lin, W., Jin, T., Wang, C., Ji, R.: Portrait booth: A versatile portrait model for fast identity-preserved personalization. arXiv preprint arXiv:2312.06354 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.786, + 0.718 + ], + "angle": 0, + "content": "47. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: SDXL: Improving latent diffusion models for high-resolution image synthesis. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=di52zR8xgf" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.786, + 0.759 + ], + "angle": 0, + "content": "48. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=FjNys5c7VvY" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.786, + 0.786 + ], + "angle": 0, + "content": "49. Purushwalkam, S., Gokul, A., Joty, S., Naik, N.: Bootpig: Bootstrapping zero-shot personalized image generation capabilities in pretrained diffusion models (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.786, + 0.84 + ], + "angle": 0, + "content": "50. QI, C., Cun, X., Zhang, Y., Lei, C., Wang, X., Shan, Y., Chen, Q.: Fatezero: Fusing attentions for zero-shot text-based video editing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15932-15942 (October 2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.128 + ], + "angle": 0, + "content": "Gal et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.204 + ], + "angle": 0, + "content": "51. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.786, + 0.245 + ], + "angle": 0, + "content": "52. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "53. Richardson, E., Alaluf, Y., Patashnik, O., Nitzan, Y., Azar, Y., Shapiro, S., Cohen-Or, D.: Encoding in style: a stylegan encoder for image-to-image translation. arXiv preprint arXiv:2008.00951 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.786, + 0.314 + ], + "angle": 0, + "content": "54. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.355 + ], + "angle": 0, + "content": "55. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.786, + 0.397 + ], + "angle": 0, + "content": "56. Ruiz, N., Li, Y., Jampani, V., Wei, W., Hou, T., Pritch, Y., Wadhwa, N., Rubinstein, M., Aberman, K.: Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.786, + 0.424 + ], + "angle": 0, + "content": "57. Ryu, S.: Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.786, + 0.467 + ], + "angle": 0, + "content": "58. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=TIdIXIpzhoI" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.786, + 0.494 + ], + "angle": 0, + "content": "59. Sauer, A., Lorenz, D., Blattmann, A., Rombach, R.: Adversarial diffusion distillation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.786, + 0.535 + ], + "angle": 0, + "content": "60. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.786, + 0.563 + ], + "angle": 0, + "content": "61. Shi, J., Xiong, W., Lin, Z., Jung, H.J.: Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.786, + 0.591 + ], + "angle": 0, + "content": "62. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.743, + 0.606 + ], + "angle": 0, + "content": "63. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.607, + 0.786, + 0.646 + ], + "angle": 0, + "content": "64. Tewel, Y., Gal, R., Chechik, G., Atzmon, Y.: Key-locked rank one editing for text-to-image personalization. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-11 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.786, + 0.674 + ], + "angle": 0, + "content": "65. Tewel, Y., Kaduri, O., Gal, R., Kasten, Y., Wolf, L., Chechik, G., Atzmon, Y.: Training-free consistent text-to-image generation (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.786, + 0.702 + ], + "angle": 0, + "content": "66. Tov, O., Alaluf, Y., Nitzan, Y., Patashnik, O., Cohen-Or, D.: Designing an encoder for stylegan image manipulation. arXiv preprint arXiv:2102.02766 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.786, + 0.769 + ], + "angle": 0, + "content": "67. Valevski, D., Lumen, D., Matias, Y., Leviathan, Y.: Face0: Instantaneously conditioning a text-to-image model on a face. In: SIGGRAPH Asia 2023 Conference Papers. SA '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3610548.3618249, https://doi.org/10.1145/3610548.3618249" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.786, + 0.799 + ], + "angle": 0, + "content": "68. Voynov, A., Chu, Q., Cohen-Or, D., Aberman, K.: \\( p+ \\): Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.786, + 0.841 + ], + "angle": 0, + "content": "69. Wallace, B., Gokul, A., Ermon, S., Naik, N.: End-to-end diffusion latent optimization improves classifier guidance. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7280-7290 (October 2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.29, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LCM-Lookahead for Encoder-based Text-to-Image Personalization" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "70. Wang, Q., Bai, X., Wang, H., Qin, Z., Chen, A.: Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.217 + ], + "angle": 0, + "content": "71. Wang, T., Zhang, Y., Fan, Y., Wang, J., Chen, Q.: High-fidelity gan inversion for image attribute editing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.218, + 0.788, + 0.259 + ], + "angle": 0, + "content": "72. Wang, Z., Bao, J., Zhou, W., Wang, W., Hu, H., Chen, H., Li, H.: Dire for diffusion-generated image detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 22445-22455 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.313 + ], + "angle": 0, + "content": "73. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15943-15953 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.37 + ], + "angle": 0, + "content": "74. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7623-7633 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.397 + ], + "angle": 0, + "content": "75. Wu, Z., Nitzan, Y., Shechtman, E., Lischinski, D.: Stylealign: Analysis and applications of aligned stylegan models (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.425 + ], + "angle": 0, + "content": "76. Xia, W., Zhang, Y., Yang, Y., Xue, J.H., Zhou, B., Yang, M.H.: Gan inversion: A survey (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.452 + ], + "angle": 0, + "content": "77. Xiao, G., Yin, T., Freeman, W.T., Durand, F., Han, S.: Fastcomposer: Tuning-free multi-subject image generation with localized attention. arXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.479 + ], + "angle": 0, + "content": "78. Yan, Y., Zhang, C., Wang, R., Cheng, P., Yu, G., Fu, B.: Facestudio: Put your face everywhere in seconds (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.522 + ], + "angle": 0, + "content": "79. Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.55 + ], + "angle": 0, + "content": "80. Yin, T., Gharbi, M., Zhang, R., Shechtman, E., Durand, F., Freeman, W.T., Park, T.: One-step diffusion with distribution matching distillation. CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.591 + ], + "angle": 0, + "content": "81. Yuan, G., Cun, X., Zhang, Y., Li, M., Qi, C., Wang, X., Shan, Y., Zheng, H.: Inserting anybody in diffusion models via celeb basis. arXiv preprint arXiv:2306.00926 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.591, + 0.788, + 0.632 + ], + "angle": 0, + "content": "82. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.632, + 0.788, + 0.661 + ], + "angle": 0, + "content": "83. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.661 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/LCM-Lookahead for Encoder-based Text-to-Image Personalization/bbfbe553-d1d9-40ca-841b-fd60b8f4aefe_origin.pdf b/2024/LCM-Lookahead for Encoder-based Text-to-Image Personalization/bbfbe553-d1d9-40ca-841b-fd60b8f4aefe_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7c60e53d555a60fa09e9cf6e2be5708099af38c4 --- /dev/null +++ b/2024/LCM-Lookahead for Encoder-based Text-to-Image Personalization/bbfbe553-d1d9-40ca-841b-fd60b8f4aefe_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:726dab9b5ee2fdca1f4f0663083d39c67d5bd8bc66102d19b06f0fff7804a167 +size 2423558 diff --git a/2024/LCM-Lookahead for Encoder-based Text-to-Image Personalization/full.md b/2024/LCM-Lookahead for Encoder-based Text-to-Image Personalization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..93dc8de32584bc0e15d115c6e198181e8aedbca7 --- /dev/null +++ b/2024/LCM-Lookahead for Encoder-based Text-to-Image Personalization/full.md @@ -0,0 +1,312 @@ +# LCM-Lookahead for Encoder-based Text-to-Image Personalization + +Rinon Gal $^{1,2*}$ , Or Lichter $^{2*}$ , Elad Richardson $^{2*}$ , Or Patashnik $^{2}$ , Amit H. Bermano $^{2}$ , Gal Chechik $^{1}$ , and Daniel Cohen-Or $^{2}$ + +![](images/a13dc909e404d86b8f75291139bdf566bcf4ad9c0c650dcd3a91e640798894e3.jpg) +$^{1}$ NVIDIA $^{2}$ Tel Aviv University +Fig.1: We introduce a novel LCM-based lookhead mechanism to apply image-space losses to personalization encoder training. These are coupled with consistent data generation and attention sharing techniques to tune existing backbones and improve identity preservation and prompt alignment. + +Abstract. Recent advancements in diffusion models have introduced fast sampling methods that can effectively produce high-quality images in just one or a few denoising steps. Interestingly, when these are distilled from existing diffusion models, they often maintain alignment with the original model, retaining similar outputs for similar prompts and seeds. These properties present opportunities to leverage fast sampling methods as a shortcut-mechanism, using them to create a preview of denoised outputs through which we can backpropagate image-space losses. In this work, we explore the potential of using such shortcut-mechanisms to guide the personalization of text-to-image models to specific facial identities. We focus on encoder-based personalization approaches, and demonstrate that by augmenting their training with a lookahead identity loss, we can achieve higher identity fidelity, without sacrificing layout diversity or prompt alignment. Code at https://1cm-lookahead.github.io/. + +# 1 Introduction + +Text-to-image personalization [18,55] methods enable users to tailor pretrained generative models to their own, personal data. Commonly, such methods focus on human data [19,67,70,81,77,56], where users aim to create novel images of specific individuals which were unseen by the pretrained model. Early works proposed to tackle this task by teaching a model new words that describe the user-provided subjects. They do so by optimizing novel word-embeddings [18,68], or by fine-tuning the generative model itself [55,57]. However, such approaches require significant per-subject optimization, leading to lengthy personalization times and large compute requirements. More recent lines of work propose to personalize the model using an encoder - a neural network trained to condition the generative model on user-provided images [19,73,61,79]. While these methods can enable inference-time personalization, they often struggle to maintain a subject's identity, or face difficulties in adapting it to novel styles. + +One manner in which encoder-based methods try to bridge these gaps is by leveraging a pretrained face recognition network as a feature extraction backbone. The intuition here is that the pretraining objective of such networks drives them to encode fine identity details which can later be exploited by the encoder. However, this approach overlooks the training loss itself, which is still driven by only an L2 noise-prediction objective. In the realm of Generative Adversarial Networks (GANs, [21]), it was shown that inversion [1,76] can be significantly improved by incorporating additional losses that better align with human perception [53,45], e.g., an identity loss. Applying such image-space losses to the personalization process could be beneficial here as well. However, the diffusion training process works on noisy image samples from intermediate diffusion time steps, and it is not clear how these should be passed into a perceptual loss which expects clean, realistic images. + +Here, we present a method for tackling this question by building on recent advancements in fast sampling methods [63,59,22], and specifically latent consistency models (LCM, [41,42]). To do so, we leverage an intriguing property of generative models: fine-tuning alignment, where a child model fine-tuned from a pretrained parent tends to preserve the semantics of the parent's latent space [75,20]. In fig. 2, we show a particular manifestation of this alignment in LCM models, where we compare a single-step LCM output to the DDPM single-step approximations at intermediate steps of the denoising process. The LCM results are not only sharper, but they maintain a high degree of similarity with the final DDPM prediction. This property also holds for personalized models (right, dog). We find this alignment holds particularly well in consistency models, likely by virtue of the consistency training loss itself. + +We propose to leverage this alignment during the training of personalization encoders, where we can create a high-quality preview of the denoiser's final output by doing a single LCM step on the noisy latents, guided by the same personalization encoder. This preview can then be used to calculate image-space losses, such as those derived from an identity detection network. Under this approach, the LCM-model provides a "shortcut" through which gradients can + +![](images/61bf70257639b9b22168d156afcf386ac4b030d48d836a2baf6c8f6a60567de7.jpg) +Fig.2: LCM output alignment. We first denoise an image partway using DDPM [25] sampling with a baseline SDXL model [47]. We then complete sampling in two manners: (top) By performing a single LCM step. (bottom) By approximating the clean image using DDPM. Even at early steps, LCM outputs provide a good approximation of the final DDPM prediction. This also holds for personalized models (e.g., LoRA trained on the DreamBooth [55] dog, right). + +back-propagate to earlier diffusion timesteps, without relying on low-quality approximations. In practice, we find that naive applications of the shortcut loss can break the alignment between the baseline and LCM models. Hence, we investigate mechanisms for preserving alignment, and show that these can improve downstream performance. + +In addition to the shortcut mechanism, we further explore an additional architectural modification inspired by recent video models and image editing works [74,11,33,50,10,43,24]. There, it has been shown that extending the self-attention mechanism, such that a generated image can also observe self-attention keys and values of a real, source image, can allow for zero-shot appearance transfer from the source to the new image [2,65]. Hence, we propose to augment the encoder with an additional path where the input image is noised, passed through a copy of the diffusion U-Net, and its self-attention keys and values are extracted and appended to the forward pass of the newly generated image. + +Finally, we train our encoder by leveraging existing backbones [51,79] and tune it on newly generated data. To create our dataset, we leverage SDXL Turbo [59], a model distilled for single-step sampling from the vanilla SDXL using score distillation sampling [48] and adversarial training [21]. Notably, we observe that this distillation process causes a significant collapse of diversity, such that a sufficiently detailed prompt will generate the same identity regardless of the input seed. We leverage this property to generate a dataset of images with repeated identities and differing styles. Importantly, the use of generated data allows us to avoid the need to collect sensitive personal data, and ensures that our dataset contains proper representation for minority classes that are typically under-represented in real datasets. + +We validate our approach through comparisons to a range of recent baselines, and through a large set of ablation scenarios. These demonstrate that our approach can achieve higher identity fidelity and prompt alignment, and highlight the benefit of integrating image-space losses during model tuning. + +# 2 Related work + +Text-to-image Personalization. In the task of text-to-image personalization [18,55], the goal is to adapt a pretrained model to better represent a user-given concept which was unseen during the original training. Initial efforts show that this can be achieved by either optimizing a new text-embedding [18] or by further tuning the denoising network itself [55]. This has spurred a variety of different approaches that try to extend the optimized embedding space [4,68,17], restrict the tuned part of the denoising network [57,35,64,8,23], or optimize for specific target prompts [7]. A joint limitation of all these different approaches is that they require a per-subject optimization process which can be time-consuming. + +Encoder-Based Personalization. To overcome the limitations of optimization-based approaches, more recent work proposes that pretrained encoders can be used to initialize the optimization process, significantly shortening tuning times [19,56,6,36]. Others tried to completely avoid any optimization, at the cost of accuracy [73,12,61,79,30]. + +Recently the problem of face personalization gained focus, with a range of works specifically tackling this domain with the goal of improving identity preservation across prompts [19,67,70,81,77,56]. In PhotoMaker [37] the features of a CLIP-model [51] are modulated using a dedicated ID-oriented dataset. Face0 [67] proposed to replace the CLIP encoder with an identity recognition network. IP-Adapter [79] introduced similar ideas in later variants, with identity network features being used either instead of, or on top of the CLIP image embedding. FaceStudio [78] follows a similar mechanism but also uses a prior model to better adhere to the original prompt. The concurrent InsantID paper builds on IP-Adapter and extends it with a landmark-conditioned ControlNet [82] to preserve pose and simplify identity preservation. + +Common to all these methods is that their training signal is based solely on the standard diffusion training loss. This is in contrast with the common practice in GAN-Based encoder methods [53,5,3,66,71,16], which demonstrated improved results by integrating perceptual losses, such as an identity loss. This gap can be attributed to the challenge of applying pixel-based losses during diffusion training. The concurrent PortraitBooth [46] proposes to apply an identity loss only to images sampled with low noise levels, where one can well approximate the clean image. However, this limits the effectiveness of the loss. + +Our work is a similar tuning-free face-identity encoder, but we propose to integrate an identity loss into the model training by using an LCM-based shortcut mechanism. + +Fast Diffusion Sampling. Standard diffusion models [25] and their text-to-image variants [54,44,52] are trained to denoise an image over a 1,000 steps. This span can be significantly shortened by leveraging ordinary differential equation solvers to navigate the diffusion flow in fewer steps [62,39,40,38,31]. + +More recently, several methods were proposed to distill diffusion models into versions that can be sampled from in fewer steps [58]. The idea there is to teach + +the network to predict points farther along the diffusion flow, either by directly predicting a baseline model's output after multiple steps [58], through consistency modeling techniques [63,42,41,34], by leveraging adversarial training [21] and score sampling [48] methods [59] or by matching distributions between two diffusion models [80]. Of these, LCM-Lora [42] has shown that the consistency modeling approach can be applied through a low-rank adaptation [27] of the model weights. + +Our work leverages such models, and LCM-LoRA in particular, in order to introduce image-space losses into personalization encoder training. + +# 3 Preliminaries + +# 3.1 Text-to-image Personalization + +Text-to-image personalization methods introduce new, user-provided concepts to a pretrained T2I diffusion model, typically by tuning novel word-embeddings, parameters of the denoising network itself, or mixture of both. Such optimization-based methods use a small (typically 3-5) image set depicting the concept, and tune the relevant parameters using the standard diffusion training loss [25]: + +$$ +L _ {D i f f u s i o n} := \mathbb {E} _ {z, y, \epsilon \sim \mathcal {N} (0, 1), t} \left[ \| \epsilon - \epsilon_ {\theta} (z _ {t}, t, y) \| _ {2} ^ {2} \right], \tag {1} +$$ + +where $\epsilon$ is an unscaled noise sample, $\epsilon_{\theta}$ is the denoising network, $t$ is the timestep, $z_{t}$ is an image or latent noised to time $t$ , and $y$ is some conditioning prompt containing a placeholder token which is used to describe the new concept. + +To speed up the personalization process, prior work suggested the use of encoders — neural networks trained to take images of the subject, and map them to some conditioning code that can guide the pretrained diffusion network to generate an image of the subject. A common approach [73,30] popularized by IP-Adapter [79] is to do so by extending the diffusion denoiser with additional cross-attention layers, which receive tokens derived from some external encoder module (e.g. a frozen CLIP-backbone [51] with a small projection head). The novel cross-attention heads and the encoder module are then trained using the loss of , where images are drawn from large-scale datasets depicting an array of subjects, commonly from a single domain (e.g. human faces). + +In this case, the diffusion loss can be rewritten as: + +$$ +L _ {D i f f u s i o n} := \mathbb {E} _ {z _ {r}, y, I _ {c}, \epsilon \sim \mathcal {N} (0, 1), t} \left[ \| \epsilon - \epsilon_ {\theta} \left(z _ {r, t}, t, y, E \left(I _ {c}\right)\right) \| _ {2} ^ {2} \right], \tag {2} +$$ + +where $I_{c}$ is a conditioning image sampled from the training set, $z_{r}$ is the image latent of a reconstruction target showing the same subject as $I_{c}$ , and $E(I_{c})$ is a conditioning code derived from the encoder. + +Our work builds on a pretrained IP-Adapter, augments it with additional self-attention-based features, and tunes it to improve identity and prompt alignment. + +![](images/305b407c3ef32fe7c83b68c788ff776dab3f5bfaaadf99ca57293a1ae3ffa8ce.jpg) +Fig. 3: (left) Encoder architecture: Our encoder has two branches: One is the standard IP-Adapter [79] that provides conditions through a new cross-attention head. The second branch is a copy of the SDXL U-net from which we extract self-attention keys and values, which we concatenate with those of the main denoising branch. (right) Training setup: The two encoder paths are provided with a conditioning image (and its noisy latent), and their outputs are used to condition the denoising of a different image of the same subject. We denoise the image with both the baseline SDXL [47] model and an LCM-model [42]. The baseline model's output is used for calculating the standard diffusion loss (eq. (2)). The LCM output is used to calculate the lookhead identity loss (eq. (4)). We portray latents as images for visual clarity. + +![](images/2a925c7743153fab8f6f48b131f0ec76bc9626ef3e7abafb88281b22e5989a9e.jpg) + +# 4 Method + +Our goal is to improve encoder-based methods so that we can concurrently achieve identity preservation, and prompt alignment. For identity preservation, we introduce an LCM-based identity lookahead loss, and a self-attention-sharing based architectural modification for existing encoder pipelines. To improve prompt alignment, we utilize a synthetic dataset which contains consistent characters generated in an array of prompts and styles. fig. 3 provides a high-level outline of our encoder architecture (left) and training process (right). In the following section, we provide additional details on each of these components. + +# 4.1 LCM-Lookahead loss + +We begin with the goal of achieving high identity fidelity. Here, we would like to draw on core ideas from existing GAN inversion literature, and particularly the use of identity networks as a loss during encoder training [53,45]. In the GAN-based literature, applying such a loss is trivial since the GAN can produce a clean image in a single forward pass. In the case of diffusion models, the standard training process involves sampling random intermediate diffusion time steps, and performing a single denoising step. These single-step predictions are typically noisy or blurry, and feeding them into a downstream image embedding network has previously been shown to be sub-optimal [15,69]. These prior + +works showed that one can improve the quality of guidance by training feature extractors on noisy images, or by performing multiple forward steps to create a clean image, and back-propagating through the entire diffusion chain. However, the first approach is expensive, and the second is impractical in training scenarios: both because such training is typically already memory constrained, and because end-to-end sampling increases iteration times by a factor of $\sim 100$ , making training infeasible. Instead, we propose to utilize a pretrained LCM-LoRA model [42], tuned from the same baseline SDXL [47] backbone, to create higher-quality previews of the fully-denoised images using a single diffusion step. This preview can be fed into the downstream feature extractor, in our case a face recognition network [14], and gradients can be backpropagated to the encoder through this LCM-path. We focus on the LoRA variant of the LCM model as it best preserves the alignment with the baseline model. + +More concretely, let $(I_c, I_r)$ be a conditioning image and reconstruction-target pair, let $z_{r,t}$ be the image latent $z_r$ , noised to an intermediary time step $t$ , and let $\epsilon_{LCM}$ be the LCM denoising network. The preview image is given by: + +$$ +\hat {z} _ {r, 0} = \frac {1}{\sqrt {\tilde {\alpha} _ {t}}} \left(z _ {r, t} - \sqrt {\tilde {\beta} _ {t}} \cdot \epsilon_ {L C M} \left(z _ {r, t}, y, t, E \left(I _ {c}\right)\right)\right), \tag {3} +$$ + +where $y$ and $E(I_c)$ are the prompt and encoder conditioning codes respectively, $\alpha_{t}$ and $\beta_{t}$ are defined by the diffusion schedule. The lookahead loss is then: + +$$ +\mathscr {L} _ {L H} = \mathscr {D} \left(D _ {V A E} \left(\hat {z} _ {r, 0}\right), I _ {c}\right), \tag {4} +$$ + +where $\mathcal{D}$ is some image-space distance metric (e.g. an identity loss) and $D_{VAE}$ is the VAE decoder which maps the latents back to image space. + +# 4.2 Maintaining alignment + +In initial experiments, we found that applying the lookahead loss of eq. (4) can improve identities over short-training runs. However, over time, this loss causes the LCM pathway to focus solely on the loss-metric at the expense of its prior output, breaking the alignment with the baseline model. We investigated a series of options for improving alignment preservation, including the use of existing distribution matching options like Score Distillation Sampling [48], or appending the standard Consistency Model loss [62] to the LCM-path. Full details and results from this investigation are presented in the supplementary. In practice, the best downstream performance was achieved using a model interpolation approach. There, for half of our training iterations, we randomly re-scale the LoRA component of the LCM-LoRA using $\alpha_{LoRA} \in [0.1, 1.0]$ . We hypothesize that applying the loss through the continuously interpolated model makes it more difficult for the encoder to converge to a solution which works differently for the LCM- and non-LCM paths. Moreover, this serves as a form of augmentation which makes adversarial solutions less likely. Lower values of $\alpha_{LoRA}$ can still act as a preview for intermediate (but not overly noisy) outputs, which can still provide a signal through the image-space model. + +Finally, similarly to Wallace et al. [69], we find it useful to focus on early (noisy) diffusion time steps. Here, we do so by applying annealing to the training time step sampling. We use the importance-weighting function of Huang et al. [29]: $f(t) = \frac{1}{T} (1 - \alpha \cos \frac{\pi t}{T})$ , where $f(t)$ is the probability to sample time step $t$ , $T$ is the total number of diffusion time steps (i.e. 1,000) and $\alpha$ is a hyperparameter which we empirically set to 0.2. + +# 4.3 Extended self-attention features + +As a second component for improving identity fidelity, we propose to leverage recent ideas in video-based modeling [74,11] and appearance transfer [2,24,10]. In these works, it was shown that expanding the self-attention mechanism such that a generated image can attend to the keys and values derived from a source image, can lead to a significant increase in visual similarity between the generated image and the source. Here, we use a similar idea to transfer identity features from the conditioning image to the generated output. These are applied on top of the baseline encoder from which we start. + +Specifically, we create a copy of the denoising U-Net which we call a "KV encoder". We pass a noisy version of our conditioning image through this U-Net, and cache the self-attention keys and values derived from this pass. Then, when performing the diffusion denoising pass, we append these keys and values to those derived from the denoised image at each self-attention layer: $K^{l} \coloneqq K_{z_{r,t}}^{l} \odot K_{z_{c,t}}^{l}$ , $V_{z_{r,t}}^{\prime l} \coloneqq V^{l} \odot V_{z_{c,t}}^{l}$ , where $l$ is the layer index and $z_{r,t}, z_{c,t}$ subscripts denote attention features coming from the reconstruction target latent and the conditioning image latent respectively. This mechanism is illustrated in fig. 3. + +Directly applying this approach with a pretrained U-net as our encoder can lead to excessive appearance transfer and loss of editing. Hence, we do not keep the encoder frozen, but rather tune it using LoRA [27]. As our training set (detailed below) contains target images which differ in style from the source, this draws the network towards discarding appearance properties that are related to the style and not to the content, greatly improving prompt alignment. We note that a similar attention-expansion idea was used for personalization in the concurrent work of Purushwalkam et al. [49]. However, their target images are all photo-realistic, and their results do indeed suffer from reduced editability. + +Finally, this pathway requires us to slightly modify the standard classifier-free guidance [26] equation. See the supplementary for more details. + +# 4.4 Consistent data generation + +Having introduced components to improve identity preservation, we now turn to improving prompt alignment. Here, we hypothesize that the limited editability in current encoder-based methods is largely grounded in their training set, which typically focuses on reconstructing real images. Moreover, existing large-scale sets are either closed and proprietary or use the withdrawn LAION dataset [60]. They also commonly contain biases, which may result in models whose performance deteriorates on minority classes. + +![](images/1bb1abab6a433fae4cb2410f361a143506d50fcf7b4c71cec2e96ae7a6b12cd8.jpg) +portrait photo + +![](images/7bcf26da4b1608e1b871bc46f43345f2e69a9b8e20e4d4a4b5bcae8ebaf88bdc.jpg) +drinking coffee + +![](images/c366debd8e08ac0bf82180864e735afe7f806bf69e33375760349f32085e0a1f.jpg) +oil painting +Fig.4: Consistent Data. Consistent data generated using SDXL-Turbo with the description "old man with curly hair and a moustache" incorporated into different prompts (e.g. "as an oil painting", "as a wanted poster") + +![](images/05a28b5cabcbabceccb72cc75ca540c04c9743613c983d21e64f40598055fcc4.jpg) +wanted poster + +![](images/8d6687628468c72a5e22554ede6d150edd4cd5947ac7b06f08b3aeac85cd04ca.jpg) +pencil drawing + +![](images/68deec775c59c5608fbf62bbd966dbfa36b9318f4dcf12d7815e28c90edb4958.jpg) +comic book + +Instead of relying on such data, we propose to generate a novel consistent dataset in which we generated the same synthetic subjects across a wide range of prompts. These can then serve as training data for our encoder, where the encoder itself is provided with one image of a given identity, and the denoising goal considers another image portraying the same subject. Such cross-image training has already been shown to be beneficial with real data [30,37]. Here, the use of the generated data allows us to take the idea a step further, and ensure our reconstruction targets also contain stylized images. These can in turn prevent the encoder from focusing on photo-realism. + +To create our data, we investigate multiple approaches for consistent generation, including: (1) generating celebrities, which the model is already familiar with and can consistently generate across various prompts, (2) ConsiStory [65], a recently introduced approach that aligns identities through attention feature sharing, and (3) using SDXL-Turbo [59]. + +While SDXL-Turbo is designed for fast sampling and not for consistent image generation, its adversarial training leads to mode-collapse. We find that, as a consequence, conditioning it on sufficiently detailed subject prompts will lead to a fixed identity across seeds and styles, as shown in fig. 4. We find this approach to achieve the best trade-off between generation time, identity consistency, and ability to change styles across the generated images. Hence, we use it to generate $500k$ images spanning roughly $100k$ identities. In practice, we trained for less than an epoch ( $40k$ identities in total). Further experiments and comparisons on consistent data generation can be found in the supplementary. + +# 4.5 Implementation Details + +We initialize our IP-Adapter backbone using the CLIP-based Face-ID model, and tune the same parameters as the original IP Adapter using a learning rate of $1e - 5$ . The encoder and denoiser U-Nets are initialized from a pretrained SDXL model. We tune the encoder U-Net using LoRA with rank 4 and a learning rate of $5e - 6$ . The decoder is kept frozen. For our LCM-Lookahead we use TinyVAE [9] to decode the latents. This lighter model reduces our memory consumption and also improves gradient flow during backpropagation. We tune the models over 5,000 iterations with a total batch size of 8 split across 2 NVIDIA A100 GPUs. + +![](images/3d58957a2c5e50c7c36d1925291bd566dac8e581fd2df77a3ca27eddf58cefd3.jpg) +Fig. 5: LCM-Lookahead Guidance. Results of classifier guidance when using different classifiers on top of our LCM-Lookahead, or standard $\hat{x}_0$ approximation. Each classifier preserves different attributes of the guiding image. $\hat{x}_0$ guidance may result in reduced quality or visible artifacts. Identity similarity values ( $\uparrow$ , measured using [28]) are shown at the bottom. + +# 5 Experiments + +# 5.1 Classifier guidance with LCM-Lookahead + +We begin with an exploration of the LCM-Lookahead mechanism using a toy experiment, where we investigate its application to classifier guidance. Although this training-free method is less effective compared to encoder-based approaches, it serves as a simple use-case in which we can analyze the lookahead mechanism and discern its potential. Specifically, we follow [69] and apply repeated guidance iterations on an early diffusion time step ( $t = 44$ out of 50 DDIM [62] steps). At each iteration, we denoise the current latents using SDXL-LCM, decode them to an image, apply our pixel-space loss, and backpropagate to modify the latents. Final results are generated by continuing unguided-DDIM sampling for the remaining diffusion steps. Fig. 5 shows several outputs of such guidance, using a perceptual LPIPS loss [83], CLIP loss [51] and an identity loss [14]. Each loss preserves different attributes of the guiding image. LPIPS preserves the semantic structure of the image, CLIP preserves semantic attributes such as facial hair, and the identity loss explicitly improves facial similarity. When applying the same guidance to the single-step DDIM-approximations ( $\hat{x}_0$ in fig. 5) we observe artifacts or reduced performance, attributed to the fact that $\hat{x}_0$ is typically blurry and discolored in early timesteps. The identity loss is particularly robust to blur, but its performance still improves by using a lookhead loss. + +Our encoder focuses on the lookahead identity loss to improve identity preservation during training. + +![](images/5dc27c536686dcbe86cdfce26f91d64f7f608fb58f146e20fd24f5c6a59086d1.jpg) +Fig. 6: Qualitative results. Our method personalizes a model to specific face identities at inference time, enabling both photo realistic and stylized prompts. + +# 5.2 Encoder evaluation + +fig. 6 shows a set of images synthesized using our encoder, across a range of identities and prompts. Our method can produce both photo-realistic results and stylized outputs, while largely preserving the subject's identity. + +To better gauge the quality of our results, we evaluate our method against a set of prior and concurrent works on personalized face generation. Specifically, we consider the three leading tuning-free approaches which have public SDXL implementations: IP-Adapter [79], InstantID [70], and PhotoMaker [37]. For IP-Adapter, we use the FaceID model that serves as our backbone, and evaluate it using two adapter-scale settings: The "official" value (1.0), and the one commonly used by the community (0.5) which significantly improves alignment with the textual prompts but harms identity. + +Qualitative Comparison. We first conduct a qualitative comparison. Prior art often shows results on well-known celebrities. However, we find that it is excessively easy to overfit on such identities (indeed, SDXL already contains tokens describing them, relegating an encoder's job to simply finding these tokens). + +![](images/9395643aac0f46f9ec3e2642ba35abb4a932511f0f6e8fd45575dc98c91a5d0e.jpg) +Fig. 7: Comparisons against prior and concurrent face-personalization encoders. IP-A (1.0) and (0.5) represent the IP-Adapter results with a scale of 1.0 and 0.5, respectively. Notably, IP-A (1.0) serves as the backbone which we fine-tune. + +Moreover, some recent papers test on identities contained in their training set (e.g. LAION-Faces contains images of famous researchers such as Yann Lecun). To paint a full picture, we provide such comparisons in the supplementary. Here, we instead collect a small set of 50 images with permissive licenses which were uploaded to https://unsplash.com/ over the period of Feb 19th - March 4th (2024). Our assumption is that these portray novel individuals which are less likely to exist in any prior training set, and thus offer the 'cleanest' benchmark. + +Comparison results are shown in fig. 7. Notably, our method outperforms our IP-Adapter backbone in editability, while matching or exceeding it on identity preservation. InstantID achieves comparable results. However, it heavily restricts the target's pose to that of the conditioning image. Moreover, InstantID was trained on significantly more data and compute (60 million images and 48 GPUs over an undisclosed time-frame). We are hopeful that our approach could also be applied on top of an InstantID backbone and lead to improved results. + +In the supplementary, we further investigate the extent of diverse styles that can be achieved by our approach. While it does not yet reach the expressiveness offered by recent optimization based approaches, we observe that our method can handle varied styles including non photo-realistic rendering or mosaicing. + +Quantitative Comparison. Next, we move to quantitative evaluations. Here, we follow prior work [19] and compare the baselines across two metrics - identity similarity and prompt alignment. Our prompts include photo-realistic reconstructions, but also stylization and material change which prior art often struggles with. To measure identity similarity we use CurricularFace [28], which differs from both our loss network and from the backbones used to extract features for the baselines. Text alignment is measured using CLIP similarity (using the ViT-B/16 version which differs from the IP-Adapter backbone). We report both metrics using two sets: (1) 5,000 identities randomly sampled from the FFHQ dataset [32] and (2) The 50 unsplash identities. The results are shown in table 1. Our method is situated on the pareto front, providing good identity preservation and high editability. Note in particular that we outperform the backbone IP-Adapter in its typically used setup ( $\alpha = 0.5$ ) on both metrics. + +Since identity metrics are highly sensitive to both the success of editing (i.e. stylized images will have lower identity scores than failed edits), and to poses (which InstantID copies), we also verify our results using a user-study. There, for each question we showed users a reference image and a prompt pair, and the outputs of two models conditioned on this pair (ours, and a random baseline). We asked users to select the image that better preserves the reference identity and better aligns with the prompt. In total, we collected 460 responses from 43 different users. In table 3 we report the percentage of users that preferred our method over each baseline. The results largely align with the automatic metrics, showing that our approach is preferred to the backbone on which we build. InstantID still outperforms all methods, primarily on account of its much improved editability, but the margin is less severe. + +Table 1: Quantitative comparisons + +
FFHQ-5000Unsplash-50
ID ↑CLIP-T ↑ID ↑CLIP-T ↑
Ours0.34526.330.30826.79
IP-A (0.5)0.26825.820.25026.36
IP-A (1.0)0.36821.390.38722.06
PhotoMaker0.34426.690.21827.19
InstantID0.63128.580.61229.06
+ +Table 3: User study results. For each matchup, we report the fraction of users who preferred our method, and the fraction that preferred the baseline. + +Table 2: Ablation study + +
FFHQ-5000Unsplash-50
ID ↑CLIP-T ↑ID ↑CLIP-T ↑
Baseline IPA0.36821.390.38722.06
+ Our Data0.22028.140.20528.25
w/ ID-Loss x00.28227.500.27227.74
w/ ID-Loss LCM0.30127.310.28127.74
+ KV Injection0.34526.330.30826.79
+ +
PhotoMakerIP-A (1.0)IP-A (0.5)InstantID
Ours71.18%82.25%57.32%44.06%
Baseline28.82%17.75%42.68%55.94%
+ +# 5.3 Ablation + +We conduct an ablation study to evaluate the contribution of our suggested components, using the following setups: (1) The baseline IP-Adapter tuned on + +our data, without extended attention or identity losses. (2) The setup of $(1) +$ an identity loss derived through standard $x0$ approximations. (3) The setup of $(1) +$ an identity loss derived through an LCM shortcut. (4) Our full model (the setup of $(3) +$ attention injection branch). For comparison, we again provide the backbone IP Adapter results.Quantitative results are provided in table 2. The use of our novel dataset greatly improves prompt alignment, at the cost of some identity fidelity. We attribute this in part to our smaller training scale, and to the fact that the identity score is impacted by stylization. Adding an identity loss significantly increases identity preservation, while passing this loss through the LCM shortcut provides a noticeably larger increase. Finally, injecting attention features leads to even better identity alignment, but at the cost of editability. + +Our results demonstrate that appending losses through an LCM-shortcut mechanism can provide improvements over the direct approximation approach. All-in-all, the combination of our components leads to a high level of both identity preservation and prompt alignment. + +# 6 Limitations and ethnic concerns + +While our model can improve existing encoders, it is not free of limitations. First, like prior tuning-free encoders, it falls short of the quality of optimization-based methods. This is particularly noticeable when working with inputs that are noticeably out-of-domain compared to standard face imagery (see supplementary). + +Secondly, our model may still suffer from biases inherent in both the backbone that we built on, as well as the diffusion model itself. Hence, it may serve to amplify social biases. Moreover, facial editing and generation software can be used to spread disinformation or defame individuals. Existing detection tools can help mitigate such risks [72,13], and we hope that these continue to improve. + +# 7 Conclusions + +In this work, we presented LCM-Lookahead, a novel mechanism for applying image-space losses to diffusion training using a fast-sampling-based shortcut mechanism. We then build on top of this mechanism to provide better identity signal to a personalization encoder, leading to improved identity preservation. Our work further explores the shortcomings of current personalization encoders and proposes two additional techniques to further improve their results. First, we show that consistent data generation methods can greatly impact prompt-alignment quality, and that SDXL-Turbo can serve to create such data. Indeed, fine-tuning on such data can even restore editability to encoders which have overfit on the photo-realistic domain. Finally, we show that the common self-attention key-value injection mechanism can also be applied to encoder-based personalization, improving the faithfulness of the generated results. We hope that both our LCM-Lookahead and our improved training scheme will serve to further push the boundaries of text-to-image personalization. + +# Acknowledgements + +This work was partially supported by ISF (grants 1337/22, 2492/20 and 3441/21). + +# References + +1. State-of-the-art in the architecture, methods and applications of stylegan. In: Computer Graphics Forum. vol. 41, pp. 591-611. Wiley Online Library (2022) +2. Alaluf, Y., Garibi, D., Patashnik, O., Averbuch-Elor, H., Cohen-Or, D.: Cross-image attention for zero-shot appearance transfer (2023) +3. Alaluf, Y., Patashnik, O., Cohen-Or, D.: Restyle: A residual-based stylegan encoder via iterative refinement. arXiv preprint arXiv:2104.02699 (2021) +4. Alaluf, Y., Richardson, E., Metzer, G., Cohen-Or, D.: A neural space-time representation for text-to-image personalization. ACM Transactions on Graphics (TOG) 42(6), 1-10 (2023) +5. Alaluf, Y., Tov, O., Mokady, R., Gal, R., Bermano, A.H.: Hyperstyle: Stylegan inversion with hypernetworks for real image editing (2021) +6. Arar, M., Gal, R., Atzmon, Y., Chechik, G., Cohen-Or, D., Shamir, A., H. Bermano, A.: Domain-agnostic tuning-encoder for fast personalization of text-to-image models. In: SIGGRAPH Asia 2023 Conference Papers. pp. 1–10 (2023) +7. Arar, M., Voynov, A., Hertz, A., Avrahami, O., Fruchter, S., Pritch, Y., Cohen-Or, D., Shamir, A.: Palp: Prompt aligned personalization of text-to-image models. arXiv preprint arXiv:2401.06105 (2024) +8. Avrahami, O., Aberman, K., Fried, O., Cohen-Or, D., Lischinski, D.: Break-ascene: Extracting multiple concepts from a single image. In: SIGGRAPH Asia 2023 Conference Papers. SA '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3610548.3618154, https://doi.org/10.1145/3610548.3618154 +9. Bohan, O.B.: Tiny autoencoder for stable diffusion. https://github.com/madebyollin/taesd (2023) +10. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 22560-22570 (October 2023) +11. Ceylan, D., Huang, C.H.P., Mitra, N.J.: Pix2video: Video editing using image diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23206-23217 (2023) +12. Chen, W., Hu, H., LI, Y., Ruiz, N., Jia, X., Chang, M.W., Cohen, W.W.: Subject-driven text-to-image generation via apprenticeship learning. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=wv3bHyQbX7 +13. Corvi, R., Cozzolino, D., Zingarini, G., Poggi, G., Nagano, K., Verdoliva, L.: On the detection of synthetic images generated by diffusion models. In: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5 (2023). https://doi.org/10.1109/ICASSP49357.2023.10095167 +14. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition (2019) +15. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021) + +16. Dinh, T.M., Tran, A.T., Nguyen, R., Hua, B.S.: Hyperinverter: Improving stylegan inversion via hypernetwork. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11389-11398 (2022) +17. Dong, Z., Wei, P., Lin, L.: Dreamartist: Towards controllable one-shot text-to-image generation via contrastive prompt-tuning. arXiv preprint arXiv:2211.11337 (2022) +18. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion (2022). https://doi.org/10.48550/ARXIV.2208.01618, https://arxiv.org/abs/2208.01618 +19. Gal, R., Arar, M., Atzmon, Y., Bermano, A.H., Chechik, G., Cohen-Or, D.: Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG) 42(4), 1-13 (2023) +20. Gal, R., Patashnik, O., Maron, H., Chechik, G., Cohen-Or, D.: Stylegan-nada: Clipguided domain adaptation of image generators. arXiv preprint arXiv:2108.00946 (2021) +21. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014) +22. Gu, J., Zhai, S., Zhang, Y., Liu, L., Susskind, J.M.: Boot: Data-free distillation of denoising diffusion models with bootstrapping. In: ICML 2023 Workshop on Structured Probabilistic Inference $\{\backslash\&\}$ Generative Modeling (2023) +23. Han, L., Li, Y., Zhang, H., Milanfar, P., Metaxas, D., Yang, F.: Svdiff: Compact parameter space for diffusion fine-tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7323-7334 (October 2023) +24. Hertz, A., Voynov, A., Fruchter, S., Cohen-Or, D.: Style aligned image generation via shared attention (2023) +25. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020) +26. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications (2021) +27. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Chen, W.: Lora: Low-rank adaptation of large language models. ArXiv abs/2106.09685 (2021) +28. Huang, Y., Wang, Y., Tai, Y., Liu, X., Shen, P., Li, S., Li, J., Huang, F.: Curricularface: adaptive curriculum learning loss for deep face recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5901-5910 (2020) +29. Huang, Z., Wu, T., Jiang, Y., Chan, K.C., Liu, Z.: Reversion: Diffusion-based relation inversion from images. arXiv preprint arXiv:2303.13495 (2023) +30. Jia, X., Zhao, Y., Chan, K.C., Li, Y., Zhang, H., Gong, B., Hou, T., Wang, H., Su, Y.C.: Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642 (2023) +31. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems 35, 26565-26577 (2022) +32. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4401-4410 (2019) +33. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero- + +shot video generators. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15954-15964 (October 2023) +34. Kim, D., Lai, C.H., Liao, W.H., Murata, N., Takida, Y., Uesaka, T., He, Y., Mitsufuji, Y., Ermon, S.: Consistency trajectory models: Learning probability flow ODE trajectory of diffusion. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=ymjI8feDTD +35. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. arXiv (2022) +36. Li, D., Li, J., Hoi, S.C.: Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720 (2023) +37. Li, Z., Cao, M., Wang, X., Qi, Z., Cheng, M.M., Shan, Y.: Photomaker: Customizing realistic human photos via stacked id embedding. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2024) +38. Liu, L., Ren, Y., Lin, Z., Zhao, Z.: Pseudo numerical methods for diffusion models on manifolds. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=PlKWd2yBkY +39. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: DPM-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=2uAaGwlp_V +40. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: DPM-solver++: Fast solver for guided sampling of diffusion probabilistic models (2023), https://openreview.net/forum?id=4vGwQqviud5 +41. Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthesizing high-resolution images with few-step inference (2023) +42. Luo, S., Tan, Y., Patil, S., Gu, D., von Platen, P., Passos, A., Huang, L., Li, J., Zhao, H.: Lcm-lora: A universal stable-diffusion acceleration module (2023) +43. Mou, C., Wang, X., Song, J., Shan, Y., Zhang, J.: Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421 (2023) +44. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021) +45. Nitzan, Y., Bermano, A., Li, Y., Cohen-Or, D.: Face identity disentanglement via latent space mapping. ACM Transactions on Graphics (TOG) 39(6), 1-14 (2020) +46. Peng, X., Zhu, J., Jiang, B., Tai, Y., Luo, D., Zhang, J., Lin, W., Jin, T., Wang, C., Ji, R.: Portrait booth: A versatile portrait model for fast identity-preserved personalization. arXiv preprint arXiv:2312.06354 (2023) +47. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: SDXL: Improving latent diffusion models for high-resolution image synthesis. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=di52zR8xgf +48. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=FjNys5c7VvY +49. Purushwalkam, S., Gokul, A., Joty, S., Naik, N.: Bootpig: Bootstrapping zero-shot personalized image generation capabilities in pretrained diffusion models (2024) +50. QI, C., Cun, X., Zhang, Y., Lei, C., Wang, X., Shan, Y., Chen, Q.: Fatezero: Fusing attentions for zero-shot text-based video editing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15932-15942 (October 2023) + +51. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021) +52. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) +53. Richardson, E., Alaluf, Y., Patashnik, O., Nitzan, Y., Azar, Y., Shapiro, S., Cohen-Or, D.: Encoding in style: a stylegan encoder for image-to-image translation. arXiv preprint arXiv:2008.00951 (2020) +54. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021) +55. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation (2022) +56. Ruiz, N., Li, Y., Jampani, V., Wei, W., Hou, T., Pritch, Y., Wadhwa, N., Rubinstein, M., Aberman, K.: Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models (2023) +57. Ryu, S.: Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora (2023) +58. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=TIdIXIpzhoI +59. Sauer, A., Lorenz, D., Blattmann, A., Rombach, R.: Adversarial diffusion distillation (2023) +60. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021) +61. Shi, J., Xiong, W., Lin, Z., Jung, H.J.: Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411 (2023) +62. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: International Conference on Learning Representations (2020) +63. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models (2023) +64. Tewel, Y., Gal, R., Chechik, G., Atzmon, Y.: Key-locked rank one editing for text-to-image personalization. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-11 (2023) +65. Tewel, Y., Kaduri, O., Gal, R., Kasten, Y., Wolf, L., Chechik, G., Atzmon, Y.: Training-free consistent text-to-image generation (2024) +66. Tov, O., Alaluf, Y., Nitzan, Y., Patashnik, O., Cohen-Or, D.: Designing an encoder for stylegan image manipulation. arXiv preprint arXiv:2102.02766 (2021) +67. Valevski, D., Lumen, D., Matias, Y., Leviathan, Y.: Face0: Instantaneously conditioning a text-to-image model on a face. In: SIGGRAPH Asia 2023 Conference Papers. SA '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3610548.3618249, https://doi.org/10.1145/3610548.3618249 +68. Voynov, A., Chu, Q., Cohen-Or, D., Aberman, K.: $p+$ : Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522 (2023) +69. Wallace, B., Gokul, A., Ermon, S., Naik, N.: End-to-end diffusion latent optimization improves classifier guidance. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7280-7290 (October 2023) + +70. Wang, Q., Bai, X., Wang, H., Qin, Z., Chen, A.: Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519 (2024) +71. Wang, T., Zhang, Y., Fan, Y., Wang, J., Chen, Q.: High-fidelity gan inversion for image attribute editing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +72. Wang, Z., Bao, J., Zhou, W., Wang, W., Hu, H., Chen, H., Li, H.: Dire for diffusion-generated image detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 22445-22455 (October 2023) +73. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15943-15953 (October 2023) +74. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7623-7633 (2023) +75. Wu, Z., Nitzan, Y., Shechtman, E., Lischinski, D.: Stylealign: Analysis and applications of aligned stylegan models (2021) +76. Xia, W., Zhang, Y., Yang, Y., Xue, J.H., Zhou, B., Yang, M.H.: Gan inversion: A survey (2021) +77. Xiao, G., Yin, T., Freeman, W.T., Durand, F., Han, S.: Fastcomposer: Tuning-free multi-subject image generation with localized attention. arXiv (2023) +78. Yan, Y., Zhang, C., Wang, R., Cheng, P., Yu, G., Fu, B.: Facestudio: Put your face everywhere in seconds (2023) +79. Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721 (2023) +80. Yin, T., Gharbi, M., Zhang, R., Shechtman, E., Durand, F., Freeman, W.T., Park, T.: One-step diffusion with distribution matching distillation. CVPR (2024) +81. Yuan, G., Cun, X., Zhang, Y., Li, M., Qi, C., Wang, X., Shan, Y., Zheng, H.: Inserting anybody in diffusion models via celeb basis. arXiv preprint arXiv:2306.00926 (2023) +82. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023) +83. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) \ No newline at end of file diff --git a/2024/LCM-Lookahead for Encoder-based Text-to-Image Personalization/images.zip b/2024/LCM-Lookahead for Encoder-based Text-to-Image Personalization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..818d635c28e0edd923d00e42da674bc72101ac19 --- /dev/null +++ b/2024/LCM-Lookahead for Encoder-based Text-to-Image Personalization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fafec0264f9cdca8ba4d9d8fb6700003ee52650e3ea98118ad2155f89e4a5058 +size 705026 diff --git a/2024/LCM-Lookahead for Encoder-based Text-to-Image Personalization/layout.json b/2024/LCM-Lookahead for Encoder-based Text-to-Image Personalization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b56b8f0fb3531f6d6afbb78c5ba9453fd5dbf0b0 --- /dev/null +++ b/2024/LCM-Lookahead for Encoder-based Text-to-Image Personalization/layout.json @@ -0,0 +1,9418 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 180, + 111, + 433, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 111, + 433, + 148 + ], + "spans": [ + { + "bbox": [ + 180, + 111, + 433, + 148 + ], + "type": "text", + "content": "LCM-Lookahead for Encoder-based Text-to-Image Personalization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "spans": [ + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": "Rinon Gal" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Or Lichter" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Elad Richardson" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Or Patashnik" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Amit H. Bermano" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Gal Chechik" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", and Daniel Cohen-Or" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 153, + 224, + 460, + 435 + ], + "blocks": [ + { + "bbox": [ + 242, + 201, + 371, + 214 + ], + "lines": [ + { + "bbox": [ + 242, + 201, + 371, + 214 + ], + "spans": [ + { + "bbox": [ + 242, + 201, + 371, + 214 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 242, + 201, + 371, + 214 + ], + "type": "text", + "content": "NVIDIA " + }, + { + "bbox": [ + 242, + 201, + 371, + 214 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 242, + 201, + 371, + 214 + ], + "type": "text", + "content": "Tel Aviv University" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 153, + 224, + 460, + 435 + ], + "lines": [ + { + "bbox": [ + 153, + 224, + 460, + 435 + ], + "spans": [ + { + "bbox": [ + 153, + 224, + 460, + 435 + ], + "type": "image", + "image_path": "a13dc909e404d86b8f75291139bdf566bcf4ad9c0c650dcd3a91e640798894e3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 445, + 482, + 495 + ], + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 495 + ], + "type": "text", + "content": "Fig.1: We introduce a novel LCM-based lookhead mechanism to apply image-space losses to personalization encoder training. These are coupled with consistent data generation and attention sharing techniques to tune existing backbones and improve identity preservation and prompt alignment." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 511, + 455, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 511, + 455, + 666 + ], + "spans": [ + { + "bbox": [ + 159, + 511, + 455, + 666 + ], + "type": "text", + "content": "Abstract. Recent advancements in diffusion models have introduced fast sampling methods that can effectively produce high-quality images in just one or a few denoising steps. Interestingly, when these are distilled from existing diffusion models, they often maintain alignment with the original model, retaining similar outputs for similar prompts and seeds. These properties present opportunities to leverage fast sampling methods as a shortcut-mechanism, using them to create a preview of denoised outputs through which we can backpropagate image-space losses. In this work, we explore the potential of using such shortcut-mechanisms to guide the personalization of text-to-image models to specific facial identities. We focus on encoder-based personalization approaches, and demonstrate that by augmenting their training with a lookahead identity loss, we can achieve higher identity fidelity, without sacrificing layout diversity or prompt alignment. Code at https://1cm-lookahead.github.io/." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 295 + ], + "type": "text", + "content": "Text-to-image personalization [18,55] methods enable users to tailor pretrained generative models to their own, personal data. Commonly, such methods focus on human data [19,67,70,81,77,56], where users aim to create novel images of specific individuals which were unseen by the pretrained model. Early works proposed to tackle this task by teaching a model new words that describe the user-provided subjects. They do so by optimizing novel word-embeddings [18,68], or by fine-tuning the generative model itself [55,57]. However, such approaches require significant per-subject optimization, leading to lengthy personalization times and large compute requirements. More recent lines of work propose to personalize the model using an encoder - a neural network trained to condition the generative model on user-provided images [19,73,61,79]. While these methods can enable inference-time personalization, they often struggle to maintain a subject's identity, or face difficulties in adapting it to novel styles." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 295, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 482, + 450 + ], + "type": "text", + "content": "One manner in which encoder-based methods try to bridge these gaps is by leveraging a pretrained face recognition network as a feature extraction backbone. The intuition here is that the pretraining objective of such networks drives them to encode fine identity details which can later be exploited by the encoder. However, this approach overlooks the training loss itself, which is still driven by only an L2 noise-prediction objective. In the realm of Generative Adversarial Networks (GANs, [21]), it was shown that inversion [1,76] can be significantly improved by incorporating additional losses that better align with human perception [53,45], e.g., an identity loss. Applying such image-space losses to the personalization process could be beneficial here as well. However, the diffusion training process works on noisy image samples from intermediate diffusion time steps, and it is not clear how these should be passed into a perceptual loss which expects clean, realistic images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 450, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 594 + ], + "type": "text", + "content": "Here, we present a method for tackling this question by building on recent advancements in fast sampling methods [63,59,22], and specifically latent consistency models (LCM, [41,42]). To do so, we leverage an intriguing property of generative models: fine-tuning alignment, where a child model fine-tuned from a pretrained parent tends to preserve the semantics of the parent's latent space [75,20]. In fig. 2, we show a particular manifestation of this alignment in LCM models, where we compare a single-step LCM output to the DDPM single-step approximations at intermediate steps of the denoising process. The LCM results are not only sharper, but they maintain a high degree of similarity with the final DDPM prediction. This property also holds for personalized models (right, dog). We find this alignment holds particularly well in consistency models, likely by virtue of the consistency training loss itself." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We propose to leverage this alignment during the training of personalization encoders, where we can create a high-quality preview of the denoiser's final output by doing a single LCM step on the noisy latents, guided by the same personalization encoder. This preview can then be used to calculate image-space losses, such as those derived from an identity detection network. Under this approach, the LCM-model provides a \"shortcut\" through which gradients can" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "text", + "content": "Gal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 485, + 208 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 485, + 208 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 485, + 208 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 485, + 208 + ], + "type": "image", + "image_path": "61bf70257639b9b22168d156afcf386ac4b030d48d836a2baf6c8f6a60567de7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 216, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 289 + ], + "type": "text", + "content": "Fig.2: LCM output alignment. We first denoise an image partway using DDPM [25] sampling with a baseline SDXL model [47]. We then complete sampling in two manners: (top) By performing a single LCM step. (bottom) By approximating the clean image using DDPM. Even at early steps, LCM outputs provide a good approximation of the final DDPM prediction. This also holds for personalized models (e.g., LoRA trained on the DreamBooth [55] dog, right)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 312, + 479, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 479, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 479, + 371 + ], + "type": "text", + "content": "back-propagate to earlier diffusion timesteps, without relying on low-quality approximations. In practice, we find that naive applications of the shortcut loss can break the alignment between the baseline and LCM models. Hence, we investigate mechanisms for preserving alignment, and show that these can improve downstream performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 372, + 480, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 480, + 479 + ], + "type": "text", + "content": "In addition to the shortcut mechanism, we further explore an additional architectural modification inspired by recent video models and image editing works [74,11,33,50,10,43,24]. There, it has been shown that extending the self-attention mechanism, such that a generated image can also observe self-attention keys and values of a real, source image, can allow for zero-shot appearance transfer from the source to the new image [2,65]. Hence, we propose to augment the encoder with an additional path where the input image is noised, passed through a copy of the diffusion U-Net, and its self-attention keys and values are extracted and appended to the forward pass of the newly generated image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 479, + 481, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 481, + 610 + ], + "type": "text", + "content": "Finally, we train our encoder by leveraging existing backbones [51,79] and tune it on newly generated data. To create our dataset, we leverage SDXL Turbo [59], a model distilled for single-step sampling from the vanilla SDXL using score distillation sampling [48] and adversarial training [21]. Notably, we observe that this distillation process causes a significant collapse of diversity, such that a sufficiently detailed prompt will generate the same identity regardless of the input seed. We leverage this property to generate a dataset of images with repeated identities and differing styles. Importantly, the use of generated data allows us to avoid the need to collect sensitive personal data, and ensures that our dataset contains proper representation for minority classes that are typically under-represented in real datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 611, + 481, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 481, + 660 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 481, + 660 + ], + "type": "text", + "content": "We validate our approach through comparisons to a range of recent baselines, and through a large set of ablation scenarios. These demonstrate that our approach can achieve higher identity fidelity and prompt alignment, and highlight the benefit of integrating image-space losses during model tuning." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "type": "text", + "content": "LCM-Lookahead for Encoder-based Text-to-Image Personalization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": "Text-to-image Personalization. In the task of text-to-image personalization [18,55], the goal is to adapt a pretrained model to better represent a user-given concept which was unseen during the original training. Initial efforts show that this can be achieved by either optimizing a new text-embedding [18] or by further tuning the denoising network itself [55]. This has spurred a variety of different approaches that try to extend the optimized embedding space [4,68,17], restrict the tuned part of the denoising network [57,35,64,8,23], or optimize for specific target prompts [7]. A joint limitation of all these different approaches is that they require a per-subject optimization process which can be time-consuming." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 485, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 485, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 485, + 319 + ], + "type": "text", + "content": "Encoder-Based Personalization. To overcome the limitations of optimization-based approaches, more recent work proposes that pretrained encoders can be used to initialize the optimization process, significantly shortening tuning times [19,56,6,36]. Others tried to completely avoid any optimization, at the cost of accuracy [73,12,61,79,30]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 450 + ], + "type": "text", + "content": "Recently the problem of face personalization gained focus, with a range of works specifically tackling this domain with the goal of improving identity preservation across prompts [19,67,70,81,77,56]. In PhotoMaker [37] the features of a CLIP-model [51] are modulated using a dedicated ID-oriented dataset. Face0 [67] proposed to replace the CLIP encoder with an identity recognition network. IP-Adapter [79] introduced similar ideas in later variants, with identity network features being used either instead of, or on top of the CLIP image embedding. FaceStudio [78] follows a similar mechanism but also uses a prior model to better adhere to the original prompt. The concurrent InsantID paper builds on IP-Adapter and extends it with a landmark-conditioned ControlNet [82] to preserve pose and simplify identity preservation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 450, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 545 + ], + "type": "text", + "content": "Common to all these methods is that their training signal is based solely on the standard diffusion training loss. This is in contrast with the common practice in GAN-Based encoder methods [53,5,3,66,71,16], which demonstrated improved results by integrating perceptual losses, such as an identity loss. This gap can be attributed to the challenge of applying pixel-based losses during diffusion training. The concurrent PortraitBooth [46] proposes to apply an identity loss only to images sampled with low noise levels, where one can well approximate the clean image. However, this limits the effectiveness of the loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 582 + ], + "type": "text", + "content": "Our work is a similar tuning-free face-identity encoder, but we propose to integrate an identity loss into the model training by using an LCM-based shortcut mechanism." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "text", + "content": "Fast Diffusion Sampling. Standard diffusion models [25] and their text-to-image variants [54,44,52] are trained to denoise an image over a 1,000 steps. This span can be significantly shortened by leveraging ordinary differential equation solvers to navigate the diffusion flow in fewer steps [62,39,40,38,31]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "More recently, several methods were proposed to distill diffusion models into versions that can be sampled from in fewer steps [58]. The idea there is to teach" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "text", + "content": "Gal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": "the network to predict points farther along the diffusion flow, either by directly predicting a baseline model's output after multiple steps [58], through consistency modeling techniques [63,42,41,34], by leveraging adversarial training [21] and score sampling [48] methods [59] or by matching distributions between two diffusion models [80]. Of these, LCM-Lora [42] has shown that the consistency modeling approach can be applied through a low-rank adaptation [27] of the model weights." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 480, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 225 + ], + "type": "text", + "content": "Our work leverages such models, and LCM-LoRA in particular, in order to introduce image-space losses into personalization encoder training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 243, + 234, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 243, + 234, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 234, + 255 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 267, + 312, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 312, + 280 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 312, + 280 + ], + "type": "text", + "content": "3.1 Text-to-image Personalization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 287, + 480, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 480, + 348 + ], + "type": "text", + "content": "Text-to-image personalization methods introduce new, user-provided concepts to a pretrained T2I diffusion model, typically by tuning novel word-embeddings, parameters of the denoising network itself, or mixture of both. Such optimization-based methods use a small (typically 3-5) image set depicting the concept, and tune the relevant parameters using the standard diffusion training loss [25]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 202, + 357, + 481, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 357, + 481, + 377 + ], + "spans": [ + { + "bbox": [ + 202, + 357, + 481, + 377 + ], + "type": "interline_equation", + "content": "L _ {D i f f u s i o n} := \\mathbb {E} _ {z, y, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, y) \\| _ {2} ^ {2} \\right], \\tag {1}", + "image_path": "048569e59dcfdf09ef7b6e3ad96a20bfd5adbc91bcbea00c377cfec28eed06b1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 384, + 479, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 479, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 479, + 420 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 384, + 479, + 420 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 384, + 479, + 420 + ], + "type": "text", + "content": " is an unscaled noise sample, " + }, + { + "bbox": [ + 130, + 384, + 479, + 420 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 384, + 479, + 420 + ], + "type": "text", + "content": " is the denoising network, " + }, + { + "bbox": [ + 130, + 384, + 479, + 420 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 384, + 479, + 420 + ], + "type": "text", + "content": " is the timestep, " + }, + { + "bbox": [ + 130, + 384, + 479, + 420 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 384, + 479, + 420 + ], + "type": "text", + "content": " is an image or latent noised to time " + }, + { + "bbox": [ + 130, + 384, + 479, + 420 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 384, + 479, + 420 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 384, + 479, + 420 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 384, + 479, + 420 + ], + "type": "text", + "content": " is some conditioning prompt containing a placeholder token which is used to describe the new concept." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 421, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 482, + 540 + ], + "type": "text", + "content": "To speed up the personalization process, prior work suggested the use of encoders — neural networks trained to take images of the subject, and map them to some conditioning code that can guide the pretrained diffusion network to generate an image of the subject. A common approach [73,30] popularized by IP-Adapter [79] is to do so by extending the diffusion denoiser with additional cross-attention layers, which receive tokens derived from some external encoder module (e.g. a frozen CLIP-backbone [51] with a small projection head). The novel cross-attention heads and the encoder module are then trained using the loss of , where images are drawn from large-scale datasets depicting an array of subjects, commonly from a single domain (e.g. human faces)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 146, + 540, + 369, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 540, + 369, + 552 + ], + "spans": [ + { + "bbox": [ + 146, + 540, + 369, + 552 + ], + "type": "text", + "content": "In this case, the diffusion loss can be rewritten as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 179, + 561, + 480, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 561, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 179, + 561, + 480, + 581 + ], + "type": "interline_equation", + "content": "L _ {D i f f u s i o n} := \\mathbb {E} _ {z _ {r}, y, I _ {c}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {r, t}, t, y, E \\left(I _ {c}\\right)\\right) \\| _ {2} ^ {2} \\right], \\tag {2}", + "image_path": "b75364f69964325f31dc4ed39d526f0bd47e1988d4645ff8e7ede2f565f028f7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 589, + 479, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 589, + 479, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 589, + 479, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 589, + 479, + 624 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 130, + 589, + 479, + 624 + ], + "type": "text", + "content": " is a conditioning image sampled from the training set, " + }, + { + "bbox": [ + 130, + 589, + 479, + 624 + ], + "type": "inline_equation", + "content": "z_{r}" + }, + { + "bbox": [ + 130, + 589, + 479, + 624 + ], + "type": "text", + "content": " is the image latent of a reconstruction target showing the same subject as " + }, + { + "bbox": [ + 130, + 589, + 479, + 624 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 130, + 589, + 479, + 624 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 589, + 479, + 624 + ], + "type": "inline_equation", + "content": "E(I_{c})" + }, + { + "bbox": [ + 130, + 589, + 479, + 624 + ], + "type": "text", + "content": " is a conditioning code derived from the encoder." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 624, + 479, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 479, + 649 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 479, + 649 + ], + "type": "text", + "content": "Our work builds on a pretrained IP-Adapter, augments it with additional self-attention-based features, and tunes it to improve identity and prompt alignment." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 177, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 177, + 91, + 448, + 103 + ], + "type": "text", + "content": "LCM-Lookahead for Encoder-based Text-to-Image Personalization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 309, + 242 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 309, + 242 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 309, + 242 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 309, + 242 + ], + "type": "image", + "image_path": "305b407c3ef32fe7c83b68c788ff776dab3f5bfaaadf99ca57293a1ae3ffa8ce.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 250, + 483, + 382 + ], + "lines": [ + { + "bbox": [ + 130, + 250, + 483, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 483, + 382 + ], + "type": "text", + "content": "Fig. 3: (left) Encoder architecture: Our encoder has two branches: One is the standard IP-Adapter [79] that provides conditions through a new cross-attention head. The second branch is a copy of the SDXL U-net from which we extract self-attention keys and values, which we concatenate with those of the main denoising branch. (right) Training setup: The two encoder paths are provided with a conditioning image (and its noisy latent), and their outputs are used to condition the denoising of a different image of the same subject. We denoise the image with both the baseline SDXL [47] model and an LCM-model [42]. The baseline model's output is used for calculating the standard diffusion loss (eq. (2)). The LCM output is used to calculate the lookhead identity loss (eq. (4)). We portray latents as images for visual clarity." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 480, + 242 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 480, + 242 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 480, + 242 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 480, + 242 + ], + "type": "image", + "image_path": "2a925c7743153fab8f6f48b131f0ec76bc9626ef3e7abafb88281b22e5989a9e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 403, + 202, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 202, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 202, + 415 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 485, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 485, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 485, + 523 + ], + "type": "text", + "content": "Our goal is to improve encoder-based methods so that we can concurrently achieve identity preservation, and prompt alignment. For identity preservation, we introduce an LCM-based identity lookahead loss, and a self-attention-sharing based architectural modification for existing encoder pipelines. To improve prompt alignment, we utilize a synthetic dataset which contains consistent characters generated in an array of prompts and styles. fig. 3 provides a high-level outline of our encoder architecture (left) and training process (right). In the following section, we provide additional details on each of these components." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 539, + 266, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 539, + 266, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 539, + 266, + 551 + ], + "type": "text", + "content": "4.1 LCM-Lookahead loss" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "We begin with the goal of achieving high identity fidelity. Here, we would like to draw on core ideas from existing GAN inversion literature, and particularly the use of identity networks as a loss during encoder training [53,45]. In the GAN-based literature, applying such a loss is trivial since the GAN can produce a clean image in a single forward pass. In the case of diffusion models, the standard training process involves sampling random intermediate diffusion time steps, and performing a single denoising step. These single-step predictions are typically noisy or blurry, and feeding them into a downstream image embedding network has previously been shown to be sub-optimal [15,69]. These prior" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "text", + "content": "Gal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 271 + ], + "type": "text", + "content": "works showed that one can improve the quality of guidance by training feature extractors on noisy images, or by performing multiple forward steps to create a clean image, and back-propagating through the entire diffusion chain. However, the first approach is expensive, and the second is impractical in training scenarios: both because such training is typically already memory constrained, and because end-to-end sampling increases iteration times by a factor of " + }, + { + "bbox": [ + 130, + 116, + 479, + 271 + ], + "type": "inline_equation", + "content": "\\sim 100" + }, + { + "bbox": [ + 130, + 116, + 479, + 271 + ], + "type": "text", + "content": ", making training infeasible. Instead, we propose to utilize a pretrained LCM-LoRA model [42], tuned from the same baseline SDXL [47] backbone, to create higher-quality previews of the fully-denoised images using a single diffusion step. This preview can be fed into the downstream feature extractor, in our case a face recognition network [14], and gradients can be backpropagated to the encoder through this LCM-path. We focus on the LoRA variant of the LCM model as it best preserves the alignment with the baseline model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 271, + 479, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 479, + 308 + ], + "type": "text", + "content": "More concretely, let " + }, + { + "bbox": [ + 130, + 271, + 479, + 308 + ], + "type": "inline_equation", + "content": "(I_c, I_r)" + }, + { + "bbox": [ + 130, + 271, + 479, + 308 + ], + "type": "text", + "content": " be a conditioning image and reconstruction-target pair, let " + }, + { + "bbox": [ + 130, + 271, + 479, + 308 + ], + "type": "inline_equation", + "content": "z_{r,t}" + }, + { + "bbox": [ + 130, + 271, + 479, + 308 + ], + "type": "text", + "content": " be the image latent " + }, + { + "bbox": [ + 130, + 271, + 479, + 308 + ], + "type": "inline_equation", + "content": "z_r" + }, + { + "bbox": [ + 130, + 271, + 479, + 308 + ], + "type": "text", + "content": ", noised to an intermediary time step " + }, + { + "bbox": [ + 130, + 271, + 479, + 308 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 271, + 479, + 308 + ], + "type": "text", + "content": ", and let " + }, + { + "bbox": [ + 130, + 271, + 479, + 308 + ], + "type": "inline_equation", + "content": "\\epsilon_{LCM}" + }, + { + "bbox": [ + 130, + 271, + 479, + 308 + ], + "type": "text", + "content": " be the LCM denoising network. The preview image is given by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 200, + 316, + 480, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 316, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 200, + 316, + 480, + 341 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {r, 0} = \\frac {1}{\\sqrt {\\tilde {\\alpha} _ {t}}} \\left(z _ {r, t} - \\sqrt {\\tilde {\\beta} _ {t}} \\cdot \\epsilon_ {L C M} \\left(z _ {r, t}, y, t, E \\left(I _ {c}\\right)\\right)\\right), \\tag {3}", + "image_path": "dd9768e1eb8b9fda8674259999bf080e5738e6aedd562cee879867c5e0aa0de9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 348, + 479, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 348, + 479, + 372 + ], + "spans": [ + { + "bbox": [ + 131, + 348, + 479, + 372 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 348, + 479, + 372 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 131, + 348, + 479, + 372 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 348, + 479, + 372 + ], + "type": "inline_equation", + "content": "E(I_c)" + }, + { + "bbox": [ + 131, + 348, + 479, + 372 + ], + "type": "text", + "content": " are the prompt and encoder conditioning codes respectively, " + }, + { + "bbox": [ + 131, + 348, + 479, + 372 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 131, + 348, + 479, + 372 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 348, + 479, + 372 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 131, + 348, + 479, + 372 + ], + "type": "text", + "content": " are defined by the diffusion schedule. The lookahead loss is then:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 247, + 381, + 480, + 394 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 381, + 480, + 394 + ], + "spans": [ + { + "bbox": [ + 247, + 381, + 480, + 394 + ], + "type": "interline_equation", + "content": "\\mathscr {L} _ {L H} = \\mathscr {D} \\left(D _ {V A E} \\left(\\hat {z} _ {r, 0}\\right), I _ {c}\\right), \\tag {4}", + "image_path": "e63a6079c0c03754764f8b81916c0f52185e04551a4b091aefa92de638a4f60d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "text", + "content": " is some image-space distance metric (e.g. an identity loss) and " + }, + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "inline_equation", + "content": "D_{VAE}" + }, + { + "bbox": [ + 130, + 402, + 479, + 426 + ], + "type": "text", + "content": " is the VAE decoder which maps the latents back to image space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 443, + 274, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 443, + 274, + 455 + ], + "spans": [ + { + "bbox": [ + 131, + 443, + 274, + 455 + ], + "type": "text", + "content": "4.2 Maintaining alignment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 461, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 481, + 666 + ], + "type": "text", + "content": "In initial experiments, we found that applying the lookahead loss of eq. (4) can improve identities over short-training runs. However, over time, this loss causes the LCM pathway to focus solely on the loss-metric at the expense of its prior output, breaking the alignment with the baseline model. We investigated a series of options for improving alignment preservation, including the use of existing distribution matching options like Score Distillation Sampling [48], or appending the standard Consistency Model loss [62] to the LCM-path. Full details and results from this investigation are presented in the supplementary. In practice, the best downstream performance was achieved using a model interpolation approach. There, for half of our training iterations, we randomly re-scale the LoRA component of the LCM-LoRA using " + }, + { + "bbox": [ + 130, + 461, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\alpha_{LoRA} \\in [0.1, 1.0]" + }, + { + "bbox": [ + 130, + 461, + 481, + 666 + ], + "type": "text", + "content": ". We hypothesize that applying the loss through the continuously interpolated model makes it more difficult for the encoder to converge to a solution which works differently for the LCM- and non-LCM paths. Moreover, this serves as a form of augmentation which makes adversarial solutions less likely. Lower values of " + }, + { + "bbox": [ + 130, + 461, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\alpha_{LoRA}" + }, + { + "bbox": [ + 130, + 461, + 481, + 666 + ], + "type": "text", + "content": " can still act as a preview for intermediate (but not overly noisy) outputs, which can still provide a signal through the image-space model." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "type": "text", + "content": "LCM-Lookahead for Encoder-based Text-to-Image Personalization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Finally, similarly to Wallace et al. [69], we find it useful to focus on early (noisy) diffusion time steps. Here, we do so by applying annealing to the training time step sampling. We use the importance-weighting function of Huang et al. [29]: " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "f(t) = \\frac{1}{T} (1 - \\alpha \\cos \\frac{\\pi t}{T})" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "f(t)" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " is the probability to sample time step " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " is the total number of diffusion time steps (i.e. 1,000) and " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " is a hyperparameter which we empirically set to 0.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 205, + 322, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 322, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 322, + 216 + ], + "type": "text", + "content": "4.3 Extended self-attention features" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 223, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 482, + 319 + ], + "type": "text", + "content": "As a second component for improving identity fidelity, we propose to leverage recent ideas in video-based modeling [74,11] and appearance transfer [2,24,10]. In these works, it was shown that expanding the self-attention mechanism such that a generated image can attend to the keys and values derived from a source image, can lead to a significant increase in visual similarity between the generated image and the source. Here, we use a similar idea to transfer identity features from the conditioning image to the generated output. These are applied on top of the baseline encoder from which we start." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 319, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 319, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 131, + 319, + 482, + 415 + ], + "type": "text", + "content": "Specifically, we create a copy of the denoising U-Net which we call a \"KV encoder\". We pass a noisy version of our conditioning image through this U-Net, and cache the self-attention keys and values derived from this pass. Then, when performing the diffusion denoising pass, we append these keys and values to those derived from the denoised image at each self-attention layer: " + }, + { + "bbox": [ + 131, + 319, + 482, + 415 + ], + "type": "inline_equation", + "content": "K^{l} \\coloneqq K_{z_{r,t}}^{l} \\odot K_{z_{c,t}}^{l}" + }, + { + "bbox": [ + 131, + 319, + 482, + 415 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 319, + 482, + 415 + ], + "type": "inline_equation", + "content": "V_{z_{r,t}}^{\\prime l} \\coloneqq V^{l} \\odot V_{z_{c,t}}^{l}" + }, + { + "bbox": [ + 131, + 319, + 482, + 415 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 319, + 482, + 415 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 131, + 319, + 482, + 415 + ], + "type": "text", + "content": " is the layer index and " + }, + { + "bbox": [ + 131, + 319, + 482, + 415 + ], + "type": "inline_equation", + "content": "z_{r,t}, z_{c,t}" + }, + { + "bbox": [ + 131, + 319, + 482, + 415 + ], + "type": "text", + "content": " subscripts denote attention features coming from the reconstruction target latent and the conditioning image latent respectively. This mechanism is illustrated in fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 521 + ], + "type": "text", + "content": "Directly applying this approach with a pretrained U-net as our encoder can lead to excessive appearance transfer and loss of editing. Hence, we do not keep the encoder frozen, but rather tune it using LoRA [27]. As our training set (detailed below) contains target images which differ in style from the source, this draws the network towards discarding appearance properties that are related to the style and not to the content, greatly improving prompt alignment. We note that a similar attention-expansion idea was used for personalization in the concurrent work of Purushwalkam et al. [49]. However, their target images are all photo-realistic, and their results do indeed suffer from reduced editability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 522, + 481, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 481, + 547 + ], + "type": "text", + "content": "Finally, this pathway requires us to slightly modify the standard classifier-free guidance [26] equation. See the supplementary for more details." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 563, + 297, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 297, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 297, + 574 + ], + "type": "text", + "content": "4.4 Consistent data generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Having introduced components to improve identity preservation, we now turn to improving prompt alignment. Here, we hypothesize that the limited editability in current encoder-based methods is largely grounded in their training set, which typically focuses on reconstructing real images. Moreover, existing large-scale sets are either closed and proprietary or use the withdrawn LAION dataset [60]. They also commonly contain biases, which may result in models whose performance deteriorates on minority classes." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "text", + "content": "Gal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 196, + 167 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 196, + 167 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 196, + 167 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 196, + 167 + ], + "type": "image", + "image_path": "1bb1abab6a433fae4cb2410f361a143506d50fcf7b4c71cec2e96ae7a6b12cd8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 170, + 184, + 186 + ], + "lines": [ + { + "bbox": [ + 156, + 170, + 184, + 186 + ], + "spans": [ + { + "bbox": [ + 156, + 170, + 184, + 186 + ], + "type": "text", + "content": "portrait photo" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 197, + 114, + 251, + 168 + ], + "blocks": [ + { + "bbox": [ + 197, + 114, + 251, + 168 + ], + "lines": [ + { + "bbox": [ + 197, + 114, + 251, + 168 + ], + "spans": [ + { + "bbox": [ + 197, + 114, + 251, + 168 + ], + "type": "image", + "image_path": "7bcf26da4b1608e1b871bc46f43345f2e69a9b8e20e4d4a4b5bcae8ebaf88bdc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 170, + 239, + 186 + ], + "lines": [ + { + "bbox": [ + 209, + 170, + 239, + 186 + ], + "spans": [ + { + "bbox": [ + 209, + 170, + 239, + 186 + ], + "type": "text", + "content": "drinking coffee" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 253, + 114, + 306, + 168 + ], + "blocks": [ + { + "bbox": [ + 253, + 114, + 306, + 168 + ], + "lines": [ + { + "bbox": [ + 253, + 114, + 306, + 168 + ], + "spans": [ + { + "bbox": [ + 253, + 114, + 306, + 168 + ], + "type": "image", + "image_path": "c366debd8e08ac0bf82180864e735afe7f806bf69e33375760349f32085e0a1f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 264, + 170, + 294, + 186 + ], + "lines": [ + { + "bbox": [ + 264, + 170, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 264, + 170, + 294, + 186 + ], + "type": "text", + "content": "oil painting" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 195, + 480, + 232 + ], + "lines": [ + { + "bbox": [ + 131, + 195, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 131, + 195, + 480, + 232 + ], + "type": "text", + "content": "Fig.4: Consistent Data. Consistent data generated using SDXL-Turbo with the description \"old man with curly hair and a moustache\" incorporated into different prompts (e.g. \"as an oil painting\", \"as a wanted poster\")" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 361, + 168 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 361, + 168 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 361, + 168 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 361, + 168 + ], + "type": "image", + "image_path": "05a28b5cabcbabceccb72cc75ca540c04c9743613c983d21e64f40598055fcc4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 170, + 347, + 186 + ], + "lines": [ + { + "bbox": [ + 321, + 170, + 347, + 186 + ], + "spans": [ + { + "bbox": [ + 321, + 170, + 347, + 186 + ], + "type": "text", + "content": "wanted poster" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 362, + 114, + 416, + 168 + ], + "blocks": [ + { + "bbox": [ + 362, + 114, + 416, + 168 + ], + "lines": [ + { + "bbox": [ + 362, + 114, + 416, + 168 + ], + "spans": [ + { + "bbox": [ + 362, + 114, + 416, + 168 + ], + "type": "image", + "image_path": "8d6687628468c72a5e22554ede6d150edd4cd5947ac7b06f08b3aeac85cd04ca.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 170, + 404, + 186 + ], + "lines": [ + { + "bbox": [ + 374, + 170, + 404, + 186 + ], + "spans": [ + { + "bbox": [ + 374, + 170, + 404, + 186 + ], + "type": "text", + "content": "pencil drawing" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 417, + 114, + 470, + 168 + ], + "blocks": [ + { + "bbox": [ + 417, + 114, + 470, + 168 + ], + "lines": [ + { + "bbox": [ + 417, + 114, + 470, + 168 + ], + "spans": [ + { + "bbox": [ + 417, + 114, + 470, + 168 + ], + "type": "image", + "image_path": "68deec775c59c5608fbf62bbd966dbfa36b9318f4dcf12d7815e28c90edb4958.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 170, + 454, + 186 + ], + "lines": [ + { + "bbox": [ + 433, + 170, + 454, + 186 + ], + "spans": [ + { + "bbox": [ + 433, + 170, + 454, + 186 + ], + "type": "text", + "content": "comic book" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 255, + 480, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 480, + 363 + ], + "type": "text", + "content": "Instead of relying on such data, we propose to generate a novel consistent dataset in which we generated the same synthetic subjects across a wide range of prompts. These can then serve as training data for our encoder, where the encoder itself is provided with one image of a given identity, and the denoising goal considers another image portraying the same subject. Such cross-image training has already been shown to be beneficial with real data [30,37]. Here, the use of the generated data allows us to take the idea a step further, and ensure our reconstruction targets also contain stylized images. These can in turn prevent the encoder from focusing on photo-realism." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 363, + 480, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 480, + 423 + ], + "type": "text", + "content": "To create our data, we investigate multiple approaches for consistent generation, including: (1) generating celebrities, which the model is already familiar with and can consistently generate across various prompts, (2) ConsiStory [65], a recently introduced approach that aligns identities through attention feature sharing, and (3) using SDXL-Turbo [59]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 423, + 481, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 481, + 531 + ], + "type": "text", + "content": "While SDXL-Turbo is designed for fast sampling and not for consistent image generation, its adversarial training leads to mode-collapse. We find that, as a consequence, conditioning it on sufficiently detailed subject prompts will lead to a fixed identity across seeds and styles, as shown in fig. 4. We find this approach to achieve the best trade-off between generation time, identity consistency, and ability to change styles across the generated images. Hence, we use it to generate " + }, + { + "bbox": [ + 130, + 423, + 481, + 531 + ], + "type": "inline_equation", + "content": "500k" + }, + { + "bbox": [ + 130, + 423, + 481, + 531 + ], + "type": "text", + "content": " images spanning roughly " + }, + { + "bbox": [ + 130, + 423, + 481, + 531 + ], + "type": "inline_equation", + "content": "100k" + }, + { + "bbox": [ + 130, + 423, + 481, + 531 + ], + "type": "text", + "content": " identities. In practice, we trained for less than an epoch (" + }, + { + "bbox": [ + 130, + 423, + 481, + 531 + ], + "type": "inline_equation", + "content": "40k" + }, + { + "bbox": [ + 130, + 423, + 481, + 531 + ], + "type": "text", + "content": " identities in total). Further experiments and comparisons on consistent data generation can be found in the supplementary." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 131, + 548, + 279, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 279, + 560 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 279, + 560 + ], + "type": "text", + "content": "4.5 Implementation Details" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 568, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 665 + ], + "type": "text", + "content": "We initialize our IP-Adapter backbone using the CLIP-based Face-ID model, and tune the same parameters as the original IP Adapter using a learning rate of " + }, + { + "bbox": [ + 130, + 568, + 482, + 665 + ], + "type": "inline_equation", + "content": "1e - 5" + }, + { + "bbox": [ + 130, + 568, + 482, + 665 + ], + "type": "text", + "content": ". The encoder and denoiser U-Nets are initialized from a pretrained SDXL model. We tune the encoder U-Net using LoRA with rank 4 and a learning rate of " + }, + { + "bbox": [ + 130, + 568, + 482, + 665 + ], + "type": "inline_equation", + "content": "5e - 6" + }, + { + "bbox": [ + 130, + 568, + 482, + 665 + ], + "type": "text", + "content": ". The decoder is kept frozen. For our LCM-Lookahead we use TinyVAE [9] to decode the latents. This lighter model reduces our memory consumption and also improves gradient flow during backpropagation. We tune the models over 5,000 iterations with a total batch size of 8 split across 2 NVIDIA A100 GPUs." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 177, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 177, + 91, + 448, + 102 + ], + "type": "text", + "content": "LCM-Lookahead for Encoder-based Text-to-Image Personalization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 297 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 297 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 297 + ], + "type": "image", + "image_path": "3d58957a2c5e50c7c36d1925291bd566dac8e581fd2df77a3ca27eddf58cefd3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "type": "text", + "content": "Fig. 5: LCM-Lookahead Guidance. Results of classifier guidance when using different classifiers on top of our LCM-Lookahead, or standard " + }, + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "type": "text", + "content": " approximation. Each classifier preserves different attributes of the guiding image. " + }, + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "type": "text", + "content": " guidance may result in reduced quality or visible artifacts. Identity similarity values (" + }, + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 130, + 308, + 482, + 369 + ], + "type": "text", + "content": ", measured using [28]) are shown at the bottom." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 383, + 230, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 383, + 230, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 383, + 230, + 397 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 407, + 364, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 364, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 364, + 419 + ], + "type": "text", + "content": "5.1 Classifier guidance with LCM-Lookahead" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 641 + ], + "type": "text", + "content": "We begin with an exploration of the LCM-Lookahead mechanism using a toy experiment, where we investigate its application to classifier guidance. Although this training-free method is less effective compared to encoder-based approaches, it serves as a simple use-case in which we can analyze the lookahead mechanism and discern its potential. Specifically, we follow [69] and apply repeated guidance iterations on an early diffusion time step (" + }, + { + "bbox": [ + 130, + 426, + 482, + 641 + ], + "type": "inline_equation", + "content": "t = 44" + }, + { + "bbox": [ + 130, + 426, + 482, + 641 + ], + "type": "text", + "content": " out of 50 DDIM [62] steps). At each iteration, we denoise the current latents using SDXL-LCM, decode them to an image, apply our pixel-space loss, and backpropagate to modify the latents. Final results are generated by continuing unguided-DDIM sampling for the remaining diffusion steps. Fig. 5 shows several outputs of such guidance, using a perceptual LPIPS loss [83], CLIP loss [51] and an identity loss [14]. Each loss preserves different attributes of the guiding image. LPIPS preserves the semantic structure of the image, CLIP preserves semantic attributes such as facial hair, and the identity loss explicitly improves facial similarity. When applying the same guidance to the single-step DDIM-approximations (" + }, + { + "bbox": [ + 130, + 426, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 130, + 426, + 482, + 641 + ], + "type": "text", + "content": " in fig. 5) we observe artifacts or reduced performance, attributed to the fact that " + }, + { + "bbox": [ + 130, + 426, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 130, + 426, + 482, + 641 + ], + "type": "text", + "content": " is typically blurry and discolored in early timesteps. The identity loss is particularly robust to blur, but its performance still improves by using a lookhead loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Our encoder focuses on the lookahead identity loss to improve identity preservation during training." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "text", + "content": "Gal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 114, + 473, + 401 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 473, + 401 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 473, + 401 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 473, + 401 + ], + "type": "image", + "image_path": "5dc27c536686dcbe86cdfce26f91d64f7f608fb58f146e20fd24f5c6a59086d1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 411, + 480, + 436 + ], + "lines": [ + { + "bbox": [ + 130, + 411, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 480, + 436 + ], + "type": "text", + "content": "Fig. 6: Qualitative results. Our method personalizes a model to specific face identities at inference time, enabling both photo realistic and stylized prompts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 453, + 257, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 257, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 257, + 464 + ], + "type": "text", + "content": "5.2 Encoder evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 473, + 481, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 481, + 508 + ], + "type": "text", + "content": "fig. 6 shows a set of images synthesized using our encoder, across a range of identities and prompts. Our method can produce both photo-realistic results and stylized outputs, while largely preserving the subject's identity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 605 + ], + "type": "text", + "content": "To better gauge the quality of our results, we evaluate our method against a set of prior and concurrent works on personalized face generation. Specifically, we consider the three leading tuning-free approaches which have public SDXL implementations: IP-Adapter [79], InstantID [70], and PhotoMaker [37]. For IP-Adapter, we use the FaceID model that serves as our backbone, and evaluate it using two adapter-scale settings: The \"official\" value (1.0), and the one commonly used by the community (0.5) which significantly improves alignment with the textual prompts but harms identity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": "Qualitative Comparison. We first conduct a qualitative comparison. Prior art often shows results on well-known celebrities. However, we find that it is excessively easy to overfit on such identities (indeed, SDXL already contains tokens describing them, relegating an encoder's job to simply finding these tokens)." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "type": "text", + "content": "LCM-Lookahead for Encoder-based Text-to-Image Personalization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 146, + 114, + 468, + 384 + ], + "blocks": [ + { + "bbox": [ + 146, + 114, + 468, + 384 + ], + "lines": [ + { + "bbox": [ + 146, + 114, + 468, + 384 + ], + "spans": [ + { + "bbox": [ + 146, + 114, + 468, + 384 + ], + "type": "image", + "image_path": "9395643aac0f46f9ec3e2642ba35abb4a932511f0f6e8fd45575dc98c91a5d0e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 394, + 482, + 431 + ], + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 431 + ], + "type": "text", + "content": "Fig. 7: Comparisons against prior and concurrent face-personalization encoders. IP-A (1.0) and (0.5) represent the IP-Adapter results with a scale of 1.0 and 0.5, respectively. Notably, IP-A (1.0) serves as the backbone which we fine-tune." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 446, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 529 + ], + "type": "text", + "content": "Moreover, some recent papers test on identities contained in their training set (e.g. LAION-Faces contains images of famous researchers such as Yann Lecun). To paint a full picture, we provide such comparisons in the supplementary. Here, we instead collect a small set of 50 images with permissive licenses which were uploaded to https://unsplash.com/ over the period of Feb 19th - March 4th (2024). Our assumption is that these portray novel individuals which are less likely to exist in any prior training set, and thus offer the 'cleanest' benchmark." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 530, + 482, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 530, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 530, + 482, + 613 + ], + "type": "text", + "content": "Comparison results are shown in fig. 7. Notably, our method outperforms our IP-Adapter backbone in editability, while matching or exceeding it on identity preservation. InstantID achieves comparable results. However, it heavily restricts the target's pose to that of the conditioning image. Moreover, InstantID was trained on significantly more data and compute (60 million images and 48 GPUs over an undisclosed time-frame). We are hopeful that our approach could also be applied on top of an InstantID backbone and lead to improved results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 613, + 482, + 662 + ], + "type": "text", + "content": "In the supplementary, we further investigate the extent of diverse styles that can be achieved by our approach. While it does not yet reach the expressiveness offered by recent optimization based approaches, we observe that our method can handle varied styles including non photo-realistic rendering or mosaicing." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 100 + ], + "type": "text", + "content": "Gal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "Quantitative Comparison. Next, we move to quantitative evaluations. Here, we follow prior work [19] and compare the baselines across two metrics - identity similarity and prompt alignment. Our prompts include photo-realistic reconstructions, but also stylization and material change which prior art often struggles with. To measure identity similarity we use CurricularFace [28], which differs from both our loss network and from the backbones used to extract features for the baselines. Text alignment is measured using CLIP similarity (using the ViT-B/16 version which differs from the IP-Adapter backbone). We report both metrics using two sets: (1) 5,000 identities randomly sampled from the FFHQ dataset [32] and (2) The 50 unsplash identities. The results are shown in table 1. Our method is situated on the pareto front, providing good identity preservation and high editability. Note in particular that we outperform the backbone IP-Adapter in its typically used setup (" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": ") on both metrics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 272, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 415 + ], + "type": "text", + "content": "Since identity metrics are highly sensitive to both the success of editing (i.e. stylized images will have lower identity scores than failed edits), and to poses (which InstantID copies), we also verify our results using a user-study. There, for each question we showed users a reference image and a prompt pair, and the outputs of two models conditioned on this pair (ours, and a random baseline). We asked users to select the image that better preserves the reference identity and better aligns with the prompt. In total, we collected 460 responses from 43 different users. In table 3 we report the percentage of users that preferred our method over each baseline. The results largely align with the automatic metrics, showing that our approach is preferred to the backbone on which we build. InstantID still outperforms all methods, primarily on account of its much improved editability, but the margin is less severe." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 132, + 455, + 298, + 523 + ], + "blocks": [ + { + "bbox": [ + 138, + 435, + 291, + 447 + ], + "lines": [ + { + "bbox": [ + 138, + 435, + 291, + 447 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 291, + 447 + ], + "type": "text", + "content": "Table 1: Quantitative comparisons" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 455, + 298, + 523 + ], + "lines": [ + { + "bbox": [ + 132, + 455, + 298, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 455, + 298, + 523 + ], + "type": "table", + "html": "
FFHQ-5000Unsplash-50
ID ↑CLIP-T ↑ID ↑CLIP-T ↑
Ours0.34526.330.30826.79
IP-A (0.5)0.26825.820.25026.36
IP-A (1.0)0.36821.390.38722.06
PhotoMaker0.34426.690.21827.19
InstantID0.63128.580.61229.06
", + "image_path": "1911bf6b20b63281a0328262340335a5718a8303aeb60e41bf588bf00fb93154.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 132, + 524, + 481, + 548 + ], + "lines": [ + { + "bbox": [ + 132, + 524, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 524, + 481, + 548 + ], + "type": "text", + "content": "Table 3: User study results. For each matchup, we report the fraction of users who preferred our method, and the fraction that preferred the baseline." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 306, + 455, + 488, + 523 + ], + "blocks": [ + { + "bbox": [ + 340, + 435, + 447, + 447 + ], + "lines": [ + { + "bbox": [ + 340, + 435, + 447, + 447 + ], + "spans": [ + { + "bbox": [ + 340, + 435, + 447, + 447 + ], + "type": "text", + "content": "Table 2: Ablation study" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 455, + 488, + 523 + ], + "lines": [ + { + "bbox": [ + 306, + 455, + 488, + 523 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 488, + 523 + ], + "type": "table", + "html": "
FFHQ-5000Unsplash-50
ID ↑CLIP-T ↑ID ↑CLIP-T ↑
Baseline IPA0.36821.390.38722.06
+ Our Data0.22028.140.20528.25
w/ ID-Loss x00.28227.500.27227.74
w/ ID-Loss LCM0.30127.310.28127.74
+ KV Injection0.34526.330.30826.79
", + "image_path": "3ca319455f3fbf59c05932bdea501d4fb25d2e78c884abb4310bde7fb4326e73.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 165, + 555, + 445, + 586 + ], + "blocks": [ + { + "bbox": [ + 165, + 555, + 445, + 586 + ], + "lines": [ + { + "bbox": [ + 165, + 555, + 445, + 586 + ], + "spans": [ + { + "bbox": [ + 165, + 555, + 445, + 586 + ], + "type": "table", + "html": "
PhotoMakerIP-A (1.0)IP-A (0.5)InstantID
Ours71.18%82.25%57.32%44.06%
Baseline28.82%17.75%42.68%55.94%
", + "image_path": "b206cd69673496c908b758e09e30835b9c2910fde10d0045bc5c41f0f926bf18.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 624, + 205, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 205, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 205, + 635 + ], + "type": "text", + "content": "5.3 Ablation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "We conduct an ablation study to evaluate the contribution of our suggested components, using the following setups: (1) The baseline IP-Adapter tuned on" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "type": "text", + "content": "LCM-Lookahead for Encoder-based Text-to-Image Personalization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "our data, without extended attention or identity losses. (2) The setup of " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "(1) +" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": " an identity loss derived through standard " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "x0" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": " approximations. (3) The setup of " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "(1) +" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": " an identity loss derived through an LCM shortcut. (4) Our full model (the setup of " + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "(3) +" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": " attention injection branch). For comparison, we again provide the backbone IP Adapter results.Quantitative results are provided in table 2. The use of our novel dataset greatly improves prompt alignment, at the cost of some identity fidelity. We attribute this in part to our smaller training scale, and to the fact that the identity score is impacted by stylization. Adding an identity loss significantly increases identity preservation, while passing this loss through the LCM shortcut provides a noticeably larger increase. Finally, injecting attention features leads to even better identity alignment, but at the cost of editability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 297 + ], + "type": "text", + "content": "Our results demonstrate that appending losses through an LCM-shortcut mechanism can provide improvements over the direct approximation approach. All-in-all, the combination of our components leads to a high level of both identity preservation and prompt alignment." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 314, + 338, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 314, + 338, + 327 + ], + "spans": [ + { + "bbox": [ + 131, + 314, + 338, + 327 + ], + "type": "text", + "content": "6 Limitations and ethnic concerns" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 339, + 481, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 481, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 481, + 387 + ], + "type": "text", + "content": "While our model can improve existing encoders, it is not free of limitations. First, like prior tuning-free encoders, it falls short of the quality of optimization-based methods. This is particularly noticeable when working with inputs that are noticeably out-of-domain compared to standard face imagery (see supplementary)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 387, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 449 + ], + "type": "text", + "content": "Secondly, our model may still suffer from biases inherent in both the backbone that we built on, as well as the diffusion model itself. Hence, it may serve to amplify social biases. Moreover, facial editing and generation software can be used to spread disinformation or defame individuals. Existing detection tools can help mitigate such risks [72,13], and we hope that these continue to improve." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 465, + 227, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 465, + 227, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 465, + 227, + 479 + ], + "type": "text", + "content": "7 Conclusions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 491, + 482, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 659 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 659 + ], + "type": "text", + "content": "In this work, we presented LCM-Lookahead, a novel mechanism for applying image-space losses to diffusion training using a fast-sampling-based shortcut mechanism. We then build on top of this mechanism to provide better identity signal to a personalization encoder, leading to improved identity preservation. Our work further explores the shortcomings of current personalization encoders and proposes two additional techniques to further improve their results. First, we show that consistent data generation methods can greatly impact prompt-alignment quality, and that SDXL-Turbo can serve to create such data. Indeed, fine-tuning on such data can even restore editability to encoders which have overfit on the photo-realistic domain. Finally, we show that the common self-attention key-value injection mechanism can also be applied to encoder-based personalization, improving the faithfulness of the generated results. We hope that both our LCM-Lookahead and our improved training scheme will serve to further push the boundaries of text-to-image personalization." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "type": "text", + "content": "Gal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 230, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 230, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 230, + 128 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 136, + 482, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 136, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 136, + 482, + 148 + ], + "type": "text", + "content": "This work was partially supported by ISF (grants 1337/22, 2492/20 and 3441/21)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 167, + 198, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 167, + 198, + 178 + ], + "spans": [ + { + "bbox": [ + 133, + 167, + 198, + 178 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 191, + 480, + 663 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 191, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 191, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 138, + 191, + 480, + 213 + ], + "type": "text", + "content": "1. State-of-the-art in the architecture, methods and applications of stylegan. In: Computer Graphics Forum. vol. 41, pp. 591-611. Wiley Online Library (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 214, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 214, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 138, + 214, + 480, + 235 + ], + "type": "text", + "content": "2. Alaluf, Y., Garibi, D., Patashnik, O., Averbuch-Elor, H., Cohen-Or, D.: Cross-image attention for zero-shot appearance transfer (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 236, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 236, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 138, + 236, + 480, + 257 + ], + "type": "text", + "content": "3. Alaluf, Y., Patashnik, O., Cohen-Or, D.: Restyle: A residual-based stylegan encoder via iterative refinement. arXiv preprint arXiv:2104.02699 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 258, + 480, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 258, + 480, + 290 + ], + "spans": [ + { + "bbox": [ + 138, + 258, + 480, + 290 + ], + "type": "text", + "content": "4. Alaluf, Y., Richardson, E., Metzer, G., Cohen-Or, D.: A neural space-time representation for text-to-image personalization. ACM Transactions on Graphics (TOG) 42(6), 1-10 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 291, + 480, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 291, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 291, + 480, + 312 + ], + "type": "text", + "content": "5. Alaluf, Y., Tov, O., Mokady, R., Gal, R., Bermano, A.H.: Hyperstyle: Stylegan inversion with hypernetworks for real image editing (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 312, + 480, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 480, + 345 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 480, + 345 + ], + "type": "text", + "content": "6. Arar, M., Gal, R., Atzmon, Y., Chechik, G., Cohen-Or, D., Shamir, A., H. Bermano, A.: Domain-agnostic tuning-encoder for fast personalization of text-to-image models. In: SIGGRAPH Asia 2023 Conference Papers. pp. 1–10 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 346, + 480, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 377 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 377 + ], + "type": "text", + "content": "7. Arar, M., Voynov, A., Hertz, A., Avrahami, O., Fruchter, S., Pritch, Y., Cohen-Or, D., Shamir, A.: Palp: Prompt aligned personalization of text-to-image models. arXiv preprint arXiv:2401.06105 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 378, + 480, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 378, + 480, + 431 + ], + "spans": [ + { + "bbox": [ + 138, + 378, + 480, + 431 + ], + "type": "text", + "content": "8. Avrahami, O., Aberman, K., Fried, O., Cohen-Or, D., Lischinski, D.: Break-ascene: Extracting multiple concepts from a single image. In: SIGGRAPH Asia 2023 Conference Papers. SA '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3610548.3618154, https://doi.org/10.1145/3610548.3618154" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 433, + 480, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 433, + 480, + 454 + ], + "spans": [ + { + "bbox": [ + 138, + 433, + 480, + 454 + ], + "type": "text", + "content": "9. Bohan, O.B.: Tiny autoencoder for stable diffusion. https://github.com/madebyollin/taesd (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 456, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 456, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 134, + 456, + 480, + 498 + ], + "type": "text", + "content": "10. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 22560-22570 (October 2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 498, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 498, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 134, + 498, + 480, + 531 + ], + "type": "text", + "content": "11. Ceylan, D., Huang, C.H.P., Mitra, N.J.: Pix2video: Video editing using image diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23206-23217 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 532, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 532, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 134, + 532, + 480, + 575 + ], + "type": "text", + "content": "12. Chen, W., Hu, H., LI, Y., Ruiz, N., Jia, X., Chang, M.W., Cohen, W.W.: Subject-driven text-to-image generation via apprenticeship learning. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=wv3bHyQbX7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 576, + 480, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 576, + 480, + 619 + ], + "spans": [ + { + "bbox": [ + 134, + 576, + 480, + 619 + ], + "type": "text", + "content": "13. Corvi, R., Cozzolino, D., Zingarini, G., Poggi, G., Nagano, K., Verdoliva, L.: On the detection of synthetic images generated by diffusion models. In: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5 (2023). https://doi.org/10.1109/ICASSP49357.2023.10095167" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 620, + 480, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 620, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 134, + 620, + 480, + 640 + ], + "type": "text", + "content": "14. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 641, + 480, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 641, + 480, + 663 + ], + "spans": [ + { + "bbox": [ + 134, + 641, + 480, + 663 + ], + "type": "text", + "content": "15. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "type": "text", + "content": "LCM-Lookahead for Encoder-based Text-to-Image Personalization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "16. Dinh, T.M., Tran, A.T., Nguyen, R., Hua, B.S.: Hyperinverter: Improving stylegan inversion via hypernetwork. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11389-11398 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "17. Dong, Z., Wei, P., Lin, L.: Dreamartist: Towards controllable one-shot text-to-image generation via contrastive prompt-tuning. arXiv preprint arXiv:2211.11337 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "text", + "content": "18. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion (2022). https://doi.org/10.48550/ARXIV.2208.01618, https://arxiv.org/abs/2208.01618" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "text", + "content": "19. Gal, R., Arar, M., Atzmon, Y., Bermano, A.H., Chechik, G., Cohen-Or, D.: Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG) 42(4), 1-13 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "20. Gal, R., Patashnik, O., Maron, H., Chechik, G., Cohen-Or, D.: Stylegan-nada: Clipguided domain adaptation of image generators. arXiv preprint arXiv:2108.00946 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "text", + "content": "21. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "text", + "content": "22. Gu, J., Zhai, S., Zhang, Y., Liu, L., Susskind, J.M.: Boot: Data-free distillation of denoising diffusion models with bootstrapping. In: ICML 2023 Workshop on Structured Probabilistic Inference " + }, + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "inline_equation", + "content": "\\{\\backslash\\&\\}" + }, + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "text", + "content": " Generative Modeling (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "text", + "content": "23. Han, L., Li, Y., Zhang, H., Milanfar, P., Metaxas, D., Yang, F.: Svdiff: Compact parameter space for diffusion fine-tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7323-7334 (October 2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "text", + "content": "24. Hertz, A., Voynov, A., Fruchter, S., Cohen-Or, D.: Style aligned image generation via shared attention (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "text", + "content": "25. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "text", + "content": "26. Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "text", + "content": "27. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Chen, W.: Lora: Low-rank adaptation of large language models. ArXiv abs/2106.09685 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 522 + ], + "type": "text", + "content": "28. Huang, Y., Wang, Y., Tai, Y., Liu, X., Shen, P., Li, S., Li, J., Huang, F.: Curricularface: adaptive curriculum learning loss for deep face recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5901-5910 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 522, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 544 + ], + "type": "text", + "content": "29. Huang, Z., Wu, T., Jiang, Y., Chan, K.C., Liu, Z.: Reversion: Diffusion-based relation inversion from images. arXiv preprint arXiv:2303.13495 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "text", + "content": "30. Jia, X., Zhao, Y., Chan, K.C., Li, Y., Zhang, H., Gong, B., Hou, T., Wang, H., Su, Y.C.: Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 578, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 609 + ], + "type": "text", + "content": "31. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems 35, 26565-26577 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "text", + "content": "32. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4401-4410 (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "33. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "text", + "content": "Gal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 149, + 117, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 117, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 117, + 481, + 138 + ], + "type": "text", + "content": "shot video generators. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15954-15964 (October 2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "type": "text", + "content": "34. Kim, D., Lai, C.H., Liao, W.H., Murata, N., Takida, Y., Uesaka, T., He, Y., Mitsufuji, Y., Ermon, S.: Consistency trajectory models: Learning probability flow ODE trajectory of diffusion. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=ymjI8feDTD" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 202 + ], + "type": "text", + "content": "35. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. arXiv (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 481, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 481, + 235 + ], + "type": "text", + "content": "36. Li, D., Li, J., Hoi, S.C.: Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. arXiv preprint arXiv:2305.14720 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 235, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 481, + 267 + ], + "type": "text", + "content": "37. Li, Z., Cao, M., Wang, X., Qi, Z., Cheng, M.M., Shan, Y.: Photomaker: Customizing realistic human photos via stacked id embedding. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "type": "text", + "content": "38. Liu, L., Ren, Y., Lin, Z., Zhao, Z.: Pseudo numerical methods for diffusion models on manifolds. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=PlKWd2yBkY" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 342 + ], + "type": "text", + "content": "39. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: DPM-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=2uAaGwlp_V" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 343, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 481, + 374 + ], + "type": "text", + "content": "40. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: DPM-solver++: Fast solver for guided sampling of diffusion probabilistic models (2023), https://openreview.net/forum?id=4vGwQqviud5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 375, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 396 + ], + "type": "text", + "content": "41. Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthesizing high-resolution images with few-step inference (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "type": "text", + "content": "42. Luo, S., Tan, Y., Patil, S., Gu, D., von Platen, P., Passos, A., Huang, L., Li, J., Zhao, H.: Lcm-lora: A universal stable-diffusion acceleration module (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 418, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 439 + ], + "type": "text", + "content": "43. Mou, C., Wang, X., Song, J., Shan, Y., Zhang, J.: Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 439, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 471 + ], + "type": "text", + "content": "44. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 471, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 481, + 493 + ], + "type": "text", + "content": "45. Nitzan, Y., Bermano, A., Li, Y., Cohen-Or, D.: Face identity disentanglement via latent space mapping. ACM Transactions on Graphics (TOG) 39(6), 1-14 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "text", + "content": "46. Peng, X., Zhu, J., Jiang, B., Tai, Y., Luo, D., Zhang, J., Lin, W., Jin, T., Wang, C., Ji, R.: Portrait booth: A versatile portrait model for fast identity-preserved personalization. arXiv preprint arXiv:2312.06354 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 525, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 568 + ], + "type": "text", + "content": "47. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: SDXL: Improving latent diffusion models for high-resolution image synthesis. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=di52zR8xgf" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "48. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=FjNys5c7VvY" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "type": "text", + "content": "49. Purushwalkam, S., Gokul, A., Joty, S., Naik, N.: Bootpig: Bootstrapping zero-shot personalized image generation capabilities in pretrained diffusion models (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "text", + "content": "50. QI, C., Cun, X., Zhang, Y., Lei, C., Wang, X., Shan, Y., Chen, Q.: Fatezero: Fusing attentions for zero-shot text-based video editing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15932-15942 (October 2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "type": "text", + "content": "LCM-Lookahead for Encoder-based Text-to-Image Personalization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 161 + ], + "type": "text", + "content": "51. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "type": "text", + "content": "52. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "53. Richardson, E., Alaluf, Y., Patashnik, O., Nitzan, Y., Azar, Y., Shapiro, S., Cohen-Or, D.: Encoding in style: a stylegan encoder for image-to-image translation. arXiv preprint arXiv:2008.00951 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "text", + "content": "54. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "55. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "text", + "content": "56. Ruiz, N., Li, Y., Jampani, V., Wei, W., Hou, T., Pritch, Y., Wadhwa, N., Rubinstein, M., Aberman, K.: Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "text", + "content": "57. Ryu, S.: Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "58. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=TIdIXIpzhoI" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 370, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 391 + ], + "type": "text", + "content": "59. Sauer, A., Lorenz, D., Blattmann, A., Rombach, R.: Adversarial diffusion distillation (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "text", + "content": "60. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 445 + ], + "type": "text", + "content": "61. Shi, J., Xiong, W., Lin, Z., Jung, H.J.: Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "text", + "content": "62. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 454, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 454, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 454, + 479 + ], + "type": "text", + "content": "63. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 480, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 480, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 480, + 481, + 511 + ], + "type": "text", + "content": "64. Tewel, Y., Gal, R., Chechik, G., Atzmon, Y.: Key-locked rank one editing for text-to-image personalization. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 1-11 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "65. Tewel, Y., Kaduri, O., Gal, R., Kasten, Y., Wolf, L., Chechik, G., Atzmon, Y.: Training-free consistent text-to-image generation (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "text", + "content": "66. Tov, O., Alaluf, Y., Nitzan, Y., Patashnik, O., Cohen-Or, D.: Designing an encoder for stylegan image manipulation. arXiv preprint arXiv:2102.02766 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 556, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 481, + 609 + ], + "type": "text", + "content": "67. Valevski, D., Lumen, D., Matias, Y., Leviathan, Y.: Face0: Instantaneously conditioning a text-to-image model on a face. In: SIGGRAPH Asia 2023 Conference Papers. SA '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3610548.3618249, https://doi.org/10.1145/3610548.3618249" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "text", + "content": "68. Voynov, A., Chu, Q., Cohen-Or, D., Aberman, K.: " + }, + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "inline_equation", + "content": "p+" + }, + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "text", + "content": ": Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "69. Wallace, B., Gokul, A., Ermon, S., Naik, N.: End-to-end diffusion latent optimization improves classifier guidance. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7280-7290 (October 2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 101 + ], + "type": "text", + "content": "Gal et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 523 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "70. Wang, Q., Bai, X., Wang, H., Qin, Z., Chen, A.: Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "text", + "content": "71. Wang, T., Zhang, Y., Fan, Y., Wang, J., Chen, Q.: High-fidelity gan inversion for image attribute editing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 172, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 482, + 205 + ], + "type": "text", + "content": "72. Wang, Z., Bao, J., Zhou, W., Wang, W., Hu, H., Chen, H., Li, H.: Dire for diffusion-generated image detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 22445-22455 (October 2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 247 + ], + "type": "text", + "content": "73. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15943-15953 (October 2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 248, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 293 + ], + "type": "text", + "content": "74. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7623-7633 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 314 + ], + "type": "text", + "content": "75. Wu, Z., Nitzan, Y., Shechtman, E., Lischinski, D.: Stylealign: Analysis and applications of aligned stylegan models (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 314, + 482, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 336 + ], + "type": "text", + "content": "76. Xia, W., Zhang, Y., Yang, Y., Xue, J.H., Zhou, B., Yang, M.H.: Gan inversion: A survey (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 336, + 482, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 357 + ], + "type": "text", + "content": "77. Xiao, G., Yin, T., Freeman, W.T., Durand, F., Han, S.: Fastcomposer: Tuning-free multi-subject image generation with localized attention. arXiv (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 358, + 482, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 379 + ], + "type": "text", + "content": "78. Yan, Y., Zhang, C., Wang, R., Cheng, P., Yu, G., Fu, B.: Facestudio: Put your face everywhere in seconds (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "text", + "content": "79. Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 413, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 435 + ], + "type": "text", + "content": "80. Yin, T., Gharbi, M., Zhang, R., Shechtman, E., Durand, F., Freeman, W.T., Park, T.: One-step diffusion with distribution matching distillation. CVPR (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "type": "text", + "content": "81. Yuan, G., Cun, X., Zhang, Y., Li, M., Qi, C., Wang, X., Shan, Y., Zheng, H.: Inserting anybody in diffusion models via celeb basis. arXiv preprint arXiv:2306.00926 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 468, + 482, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 500 + ], + "type": "text", + "content": "82. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 500, + 482, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 482, + 523 + ], + "type": "text", + "content": "83. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 177, + 91, + 447, + 102 + ], + "type": "text", + "content": "LCM-Lookahead for Encoder-based Text-to-Image Personalization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LEGO_ Learning EGOcentric Action Frame Generation via Visual Instruction Tuning/13a5a5c4-078d-4d85-be9d-7966d6d6437e_content_list.json b/2024/LEGO_ Learning EGOcentric Action Frame Generation via Visual Instruction Tuning/13a5a5c4-078d-4d85-be9d-7966d6d6437e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4c15eb92198f05b88deaa3520624c7aa8e326a42 --- /dev/null +++ b/2024/LEGO_ Learning EGOcentric Action Frame Generation via Visual Instruction Tuning/13a5a5c4-078d-4d85-be9d-7966d6d6437e_content_list.json @@ -0,0 +1,2104 @@ +[ + { + "type": "text", + "text": "LEGO: Learning EGOcentric Action Frame Generation via Visual Instruction Tuning", + "text_level": 1, + "bbox": [ + 254, + 140, + 750, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bolin Lai $^{1,2,\\dagger}$ , Xiaoliang Dai $^{1}$ , Lawrence Chen $^{1}$ , Guan Pang $^{1}$ , James M. Rehg $^{3}$ , and Miao Liu $^{1,*}$", + "bbox": [ + 274, + 210, + 727, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 GenAI, Meta", + "bbox": [ + 449, + 253, + 553, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Georgia Institute of Technology", + "bbox": [ + 387, + 268, + 614, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 University of Illinois Urbana-Champaign", + "bbox": [ + 357, + 282, + 643, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "bolin.lai@gatech.edu {xiaoliangdai,lawrencechen,gpang,miaoliu}@meta.com jrehg@illinois.edu", + "bbox": [ + 222, + 297, + 779, + 325 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bbb63d72c539697167526da88405b46948a15f31e23638eca417d8266b74387d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 383, + 264, + 407 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "User Prompt", + "text_level": 1, + "bbox": [ + 294, + 353, + 372, + 364 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"Can you provide instructions on how to wash the trouser with the brush in my current situation?\"", + "bbox": [ + 281, + 369, + 388, + 419 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6a7072f59eb18a44aaafacb5ed0a0cc5fd2b54ea97b19cf1a992fa6095a0ef5f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 351, + 517, + 424 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Check the Care Label: Check the care label on your trousers for any specific washing instructions.", + "bbox": [ + 230, + 434, + 344, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2. Prep the Area: Find a clean and flat surface to work on, such as a table or countertop.", + "bbox": [ + 230, + 494, + 339, + 535 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3. Spot Check Stains:\" :", + "bbox": [ + 230, + 545, + 336, + 568 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(a) LLM Response", + "bbox": [ + 230, + 579, + 339, + 590 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/86fe4e4caa66be29e7ca78900cdab603d56c5686b89b32eb4a4e8c9b43c83415.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 356, + 549, + 387, + 573 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. You should submerge the trouser in the water.", + "bbox": [ + 410, + 436, + 522, + 457 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2. Use the brush to scrub the trouser; focusing on any stains or areas that may require extra attention.", + "bbox": [ + 408, + 467, + 522, + 517 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3. Once the trouser is clean, you should rinse it.\"", + "bbox": [ + 410, + 527, + 514, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "中", + "bbox": [ + 462, + 560, + 470, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(b) Visual LLM Response", + "bbox": [ + 403, + 579, + 555, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Actions in Existing Egocentric Datasets", + "text_level": 1, + "bbox": [ + 539, + 349, + 779, + 362 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tackle open the tap \npuncture the sandal \nrub oil into dough place mug \nturn the mothendose sew the cloth doll \nwash trouser with brush \nsqueeze ball of dough \nclean a sieve shake the spray adhesive \nclean rice adjust gasattach the wire \nplace spice \nopen fridge pour the topsoil \nclean place ball", + "bbox": [ + 555, + 364, + 743, + 438 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/be2629e345d61ded979010eb6eee5b270c97ba36f8e7287d4a4e852ba4a40d16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 550, + 565, + 574 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/acbd7ecd6c89a1bea1d9fd71317a8766d6c4895ff647dce857a47ca74f11d78b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 584, + 460, + 751, + 556 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bb45d21632f838a2a5dcfc1602d35bb71675fb6991faa874ec6addc94934d2c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 542, + 784, + 573 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(c) Our model (LEGO) Response", + "bbox": [ + 584, + 579, + 779, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fig. 1: When a user performing a complex task asks a large language model (LLM) for instructions (a) on how to complete the steps, she receives a generic answer and has to translate the guidance into her specific situation. If she is wearing a camera, then the prompt can be augmented with an egocentric view of the scene and passed to a Visual LLM (b), and the description is now contextualized to her situation. But she still faces the challenge of parsing a written description. When she uses our novel LEGO model (c), however, the combined image and prompt are used to automatically generate an image that provides visual guidance tailored to her exact situation from the egocentric viewpoint. Now she can complete her task seamlessly!", + "bbox": [ + 212, + 604, + 787, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Generating instructional images of human daily actions from an egocentric viewpoint serves as a key step towards efficient skill transfer. In this paper, we introduce a novel problem - egocentric action", + "bbox": [ + 261, + 762, + 740, + 803 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was done when the first author was an intern at GenAI, Meta.", + "bbox": [ + 230, + 810, + 725, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 233, + 825, + 390, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "frame generation. The goal is to synthesize an image depicting an action in the user's context (i.e., action frame) by conditioning on a user prompt and an input egocentric image. Notably, existing egocentric action datasets lack the detailed annotations that describe the execution of actions. Additionally, existing diffusion-based image manipulation models are sub-optimal in controlling the state transition of an action in egocentric image pixel space because of the domain gap. To this end, we propose to Learn EGOcentric (LEGO) action frame generation via visual instruction tuning. First, we introduce a prompt enhancement scheme to generate enriched action descriptions from a visual large language model (VLLM) by visual instruction tuning. Then we propose a novel method to leverage image and text embeddings from the VLLM as additional conditioning to improve the performance of a diffusion model. We validate our model on two egocentric datasets - Ego4D and Epic-Kitchens. Our experiments show substantial improvement over prior image manipulation models in both quantitative and qualitative evaluation. We also conduct detailed ablation studies and analysis to provide insights in our method. More details of the dataset and code are available on the website (https://bolinlai.github.io/Lego_EgoActGen/).", + "bbox": [ + 261, + 147, + 743, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: Egocentric Vision $\\cdot$ Instruction Tuning $\\cdot$ Diffusion Model", + "bbox": [ + 261, + 422, + 730, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 462, + 375, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The emergence of Large Language Models (LLMs) [6, 10, 64, 99], such as ChatGPT, has revolutionized the transfer of knowledge. However, an LLM alone is not a sufficient tool for human skill transfer. Consider the question answering example in Fig. 1(a). The LLM can summarize general world knowledge, but its response may not be directly applicable to the user's current circumstances. To bridge this gap, egocentric visual perception provides a novel means to capture the actions and intentions as well as the surrounding context of the camera wearer. As shown in Fig. 1(b), recent Visual Large Language Models (VLLMs) [1, 12, 46, 52, 102] can generate instructions based on the egocentric visual input. However, such verbose textual instructions are not the optimal medium for enabling efficient human skill transfer (e.g., via AR devices). Neuroscience studies have revealed that the human brain processes text much more slowly than images [4], and that humans can interpret an action from a single static image [22]. Motivated by these discoveries, we seek to design an image generation architecture that can synthesize an image which not only vividly depicts how an action should be conducted, but also seamlessly aligns with the user's visual perspective.", + "bbox": [ + 212, + 492, + 787, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Formally, we introduce the novel problem of egocentric action frame generation as depicted in Fig. 1(c). Given a user query about how to perform a specific action and an egocentric image capturing the moment before the action begins, the goal is to synthesize an egocentric image illustrating the execution of the action in the same egocentric context. We address this problem by harnessing diffusion models [27,67], which have been shown to be powerful tools for image", + "bbox": [ + 212, + 750, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "B. Lai et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "manipulation [5,25,44,56,100]. There are two major challenges in using diffusion models to generate action frames from an egocentric perspective. First, the action annotations of the existing egocentric datasets [13, 21] are simply composed of a verb and nouns (see word cloud in Fig. 1), and thus lack the necessary details for diffusion models to learn the action state transition and to associate the action with correct objects and body movements. Second, the existing diffusion models are pre-trained primarily on exocentric (third-person-view) images, and have limited ability to represent complicated human daily activities from an egocentric perspective. In contrast, our proposed problem requires image generation in the egocentric view, conditioned on the user prompt of actions. The resulting domain gap impedes existing methods from synthesizing faithful and consistent egocentric action frames.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these challenges, we propose to Learn EGOcentric (LEGO) action frame generation with visual instruction tuning. First, we introduce a prompt enhancement scheme to generate enriched action descriptions at scale from an instruction-tuned VLLM. Second, we incorporate the image and text embeddings from finetuned VLLM as additional conditioning into the diffusion model to narrow the domain gap and improve the controllability of action frame generation. Our experimental results suggest that the enriched action descriptions and our innovative utilization of VLLM embeddings both improve the image generation performance. Our model is able to provide a generated key action frame together with a detailed action descriptions to facilitate human skill transfer from the egocentric perspective. Overall, our contributions can be summarized as follows:", + "bbox": [ + 212, + 329, + 787, + 508 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the novel problem of egocentric action frame generation to facilitate the process of skill transfer and address the challenges of missing action details in prompts and domain gap in existing image diffusion models.", + "- We propose a prompt enhancement strategy based on visual instruction tuning to enrich egocentric action descriptions, and demonstrate how the enriched descriptions can help the diffusion model understand the action state transition from the egocentric perspective.", + "- We propose a novel approach to incorporate the text and visual embeddings from the VLLM into the latent diffusion model to bridge the domain gap and improve the performance for egocentric action frame generation.", + "- We conduct thorough experiments on the Ego4D and Epic-Kitchens datasets to validate the superiority of our model over prior approaches. We also showcase the contribution of each component of our model design through ablation studies. We further provide analysis on how the visual instruction-tuned embeddings benefit model performance." + ], + "bbox": [ + 225, + 523, + 784, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 775, + 387, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-Guided Image Manipulation. The recent emergence of diffusion models enables text-guided image manipulation including image restoration [33], style", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Learning Egocentric Action Frame Generation", + "bbox": [ + 421, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "transfer [75], personalized image synthesis [68,71,87], pose generation [28,36] and generic image editing [15,18,34,38,44,57,60,79,82,82-84,94,98,101]. SDEdit [56] converts the image to the latent space by adding noise through a stochastic differential equation and then denoises the representation for image editing. Rombach et al. [67] further expand SDEdit from the original stroke-based editing to text-based editing. Null-text inversion (NTI) [58] inverts a real image by DDIM [35] to yield the diffusion process and then reconstructs the image. The image can then be edited following the same strategies as Prompt-to-Prompt [25]. NTI relies on accurate inversion process which can be improved by using coupled transformations [80] or proximal guidance [24] and accelerated by a blended guidance strategy [61]. To associate the nouns with correct objects, DIFFEDIT [11] generates a mask to localize the manipulated regions. However, most inversion-based methods require accurate image captions, which are largely unavailable in the existing egocentric dataset. Recently, InstructPix2Pix [5] demonstrates the potential to edit a real image without the inversion process and original captions. However, how to leverage the diffusion model to control the state transition of an action within the egocentric image plane remains unexplored.", + "bbox": [ + 212, + 145, + 787, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Large Language Model for Image Generation. LLMs [6, 10, 64, 76, 78, 99] and VLLMs [1, 12, 23, 46, 52, 73, 95, 102] have shown their strong capability of understanding complex human instructions. Recently, LLMs and VLLMs are used to guide image generation [2, 7-9, 51, 93, 102]. Wen et al. [88] use a pretrained VLLM to pinpoint the misalignment of the text prompt and the synthetic image and then correct it using the diffusion model. Lian et al. [49] propose to generate a layout map using an LLM to improve the understanding of prompts with spatial and numerical reasoning. InstructEdit [83] uses BLIP-2 [46] to infer the objects that will be edited and then generates a mask with SAM [39] for object grounding. A pre-trained LLM can also be used as a controller to connect with various foundation models [70, 89, 90]. GILL [40] learns text-relevant image embeddings from VLLM in a few additional tokens for image generation. Importantly, all previous efforts apply the off-the-shelf foundational models directly to their problems without finetuning. In contrast, our method uses visual instruction tuning to improve the image and text embeddings from the VLLM, which narrows the domain gap and thereby facilitates the action frame generation from the egocentric point of view.", + "bbox": [ + 212, + 409, + 787, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Egocentric Vision. Recent efforts seek to understand human actions and perceptual attention [16, 29, 30, 37, 41-43, 48, 69, 74, 86], model hand-object interactions [20, 53, 54, 65, 91], and estimate human body poses [45, 55, 77, 81] from the egocentric perspective. Here, we mainly discuss the most relevant works on egocentric visual-language models and egocentric visual content generation. Lin et al. [50] propose the first egocentric video-language pre-training model - EgoVLP. Pramanick et al. [62] further improve it by incorporating multi-modal fusion directly into the video and language backbone. Ashutosh et al. [3] propose to learn a hierarchical video-language embedding for long egocentric videos. Ramakrishnan et al. [66] introduce NaQ, which is a data augmentation strategy to train models for long egocentric video search with natural language queries. In terms", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "B. Lai et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8994e926f2e10d991bc7e4bc4892d5ec209c1ae0ca2897970c19ed210bb51dbd.jpg", + "image_caption": [ + "Example for In-context Learning", + "Input Query" + ], + "image_footnote": [ + "Action Label: \"chop end of onion\" Bounding Boxes: \"left hand-[0.203, 0.368, 0.499, 1.000], right hand-[0.530, 0.527, 0.797, 1.000], chopping board-[0.248, 0.156, 0.644, 0.750], onion-[0.462, 0.377, 0.509, 0.506], knife-[0.484, 0.343, 0.546, 0.586]\"" + ], + "bbox": [ + 287, + 162, + 464, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/692f9f9ea4a6532001bf520451bdb128cea9028caef34ef5a6d582f10aeb9147.jpg", + "image_caption": [ + "Query and Response", + "Input Query", + "Fig. 2: Examples for data curation using GPT-3.5. We provide detailed action descriptions of several example images as well as their action labels and bounding boxes for in-context learning. In addition, we input the action label and bounding boxes of another action as a query. GPT-3.5 is able to generate descriptions with enriched information (highlighted by underlines) in the response." + ], + "image_footnote": [ + "Action Description: \"The person presses the onion on the chopping board with the left hand and then cuts off the end of the onion with a knife in the right hand.\"", + "Action Label: \"squidge into lunch box\" Bounding Boxes: \"left hand-[0.427, 0.799, 0.591, 1.00], right hand-[0.654, 0.949, 0.707, 1.00], lunch box-[0.390, 0.512, 0.696, 0.993], sink-[0.00, 0.802, 0.315, 1.00], container lid-[0.403, 0.454, 0.513, 0.635], fork-[0.589, 0.798, 0.672, 1.00]\"", + "GPT-3.5 Response: \"The person uses their left hand to hold a lunch box, while their right hand uses a fork to squidge food into the lunch box, which is placed on the sink with the container lid nearby.\"" + ], + "bbox": [ + 539, + 162, + 717, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of egocentric visual generation, Jia et al. [32] leverage GANs [19] to generate future head motion in hand forecasting task. Zhang et al. [96] leverage GANs to facilitate future gaze anticipation. Ye et al. [92] propose the affordance diffusion model that takes in the image of an object and generates possible hand-object interactions in the egocentric view. In this paper, we present the first work that investigates how to leverage VLLMs and diffusion models to generate action state transition on the egocentric image plane.", + "bbox": [ + 212, + 479, + 787, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 609, + 328, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The problem setting of egocentric action frame generation is illustrated in Fig. 1(c). Given an egocentric image frame $\\mathcal{X}$ that captures the user's current visual context as well as a user query prompt $\\mathcal{P}$ regarding how to perform an action, our goal is to synthesize the action frame $\\mathcal{Y}$ that visually depicts how the action should be conducted in the same situation (i.e., keep a consistent background).", + "bbox": [ + 212, + 642, + 795, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The key insight of our proposed LEGO model is leveraging the strong capability of a VLLM to enhance the diffusion model for egocentric action frame generation. The annotations of existing egocentric datasets do not describe the details of how actions are conducted. As a remedy, we leverage visual instruction tuning to finetune a VLLM that enriches action descriptions based on the egocentric visual prompt. In addition, the existing diffusion-based image manipulation models are limited in understanding egocentric action state transition, due to the domain gap between the exocentric pre-training dataset and the ego-", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Learning Egocentric Action Frame Generation", + "bbox": [ + 421, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f8e36e3ed9cec7943cda7be08b211f7d4835e082ce1346fbfdcb6504d546c0f5.jpg", + "image_caption": [ + "Fig. 3: Overview of our proposed LEGO model. We first finetune a visual large language model (VLLM) to generate the enriched action description with visual instruction tuning. We then project image and text embeddings from the finetuned VLLM to the feature space of the latent diffusion model (LDM). Finally, we train the LDM to synthesize the egocentric action frame conditioning on the input frame, enriched action description, as well as the VLLM image and text embeddings." + ], + "image_footnote": [], + "bbox": [ + 222, + 145, + 782, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "centric action dataset for our problem. To bridge this gap, we propose a novel approach that leverages VLLM embeddings to control the state transition of actions and to generate action frames accordingly. We detail the VLLM-based data enrichment pipeline and our model design in the following sections.", + "bbox": [ + 212, + 535, + 787, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Egocentric Visual Instruction Tuning", + "text_level": 1, + "bbox": [ + 214, + 625, + 570, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data Curation for Visual Instruction Tuning. As shown in Fig. 2, we use GPT-3.5 to generate detailed action descriptions based on an input query composed of a short action label and object bounding boxes that are provided in the existing datasets. First, we prepare several examples of possible inputs along with their expected output descriptions for GPT-3.5 to perform in-context learning. These examples cover a diverse set of scenes in the egocentric action dataset. Each example is composed of an action label, a manually annotated action description, and relative spatial information of hands and objects-of-interests represented by bounding box coordinates. GPT-3.5 can learn from the given examples and generate similar detailed action descriptions based on a new input query. The resulting GPT-3.5 curated data is then further used for visual instruction tuning. More details of the prompt are provided in the supplementary.", + "bbox": [ + 212, + 657, + 787, + 843 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "B. Lai et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual Instruction Tuning. We follow the finetuning strategy in prior work [52] as shown in Fig. 3(a). Specifically, we use the pre-trained CLIP visual encoder [63] $\\phi$ to encode the visual representation and then apply a linear projection layer $\\tau$ to map the CLIP visual features into the semantic space of the LLM, i.e., $\\mathcal{H}_i = \\tau (\\phi (\\mathcal{X}))$ . To construct the user prompt $\\mathcal{P}$ , we insert the action label annotation into a prompt template to create a coherent query that is aligned with the instruction-following nature of an LLM. We then tokenize $\\mathcal{P}$ , and feed both the prompt text tokens and image tokens $\\mathcal{H}_i$ as inputs into the LLM. Finally, the LLM is trained to generate enriched action description (denoted as $\\mathcal{R}$ ) based on the user prompt and image input.", + "bbox": [ + 212, + 146, + 789, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Prompt Enrichment at Scale. Note that the visual instruction tuned VLLM doesn't rely on any object bounding boxes as input. Therefore, we can generate enriched action descriptions for all egocentric action data at scale.", + "bbox": [ + 212, + 297, + 787, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Egocentric Action Frame Generation", + "text_level": 1, + "bbox": [ + 214, + 364, + 565, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We leverage a latent diffusion model (LDM) [67] to synthesize the action frame conditioning on the input frame and the detailed action description $\\mathcal{R}$ generated by our finetuned VLLM (see Fig. 3(b)). Following the regular steps in LDMs [5, 67], the input image is first encoded into a latent space using a pre-trained autoencoder $\\mathcal{E}$ . Then the input to the denoising UNet is a concatenation of the latent input representation $\\mathcal{E}(\\mathcal{X})$ and a Gaussian noise $\\mathcal{G}$ .", + "bbox": [ + 212, + 388, + 787, + 478 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our key innovation is to design the U-Net conditioning component so that the diffusion model can interpret the egocentric actions correctly. To start, we follow [5] and adopt the conventional pre-trained CLIP text encoder $\\psi$ to extract a text representation of $\\mathcal{R}$ , i.e., $\\psi(\\mathcal{R}) \\in \\mathbb{R}^{N \\times D}$ where $N$ is the maximum number of text tokens and $D$ is the number of feature channels. We further leverage the image and text embeddings from the visual instruction tuned VLLM as additional LDM conditioning to alleviate the domain gap issue.", + "bbox": [ + 212, + 479, + 787, + 584 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, we feed the VLLM image embedding $\\mathcal{H}_i$ into an extra linear layer $\\sigma$ to map it to LDM feature space, i.e., $\\sigma(\\mathcal{H}_i) \\in \\mathbb{R}^{M \\times D}$ , where $M$ is the number of image tokens. Note that $\\mathcal{H}_i$ is already projected to the semantic space during visual instruction tuning, and therefore differs from the image embedding $\\mathcal{E}(\\mathcal{X})$ from the auto-encoder. Moreover, we also extract the text embedding $\\mathcal{H}_t$ before the last linear layer of LLM. We adopt a fixed token number $N$ and enforce padding or truncation behavior, as in the CLIP text encoder. The text embedding is then fed to a projection layer $\\mu$ . In LLM decoder, the response is generated iteratively and each word embedding only conditions on the context ahead of it. To extract the holistic semantic meaning of $\\mathcal{H}_t$ in LDM feature space, we further add self-attention layers $\\pi$ after the projection, i.e., $\\pi(\\mu(\\mathcal{H}_t)) \\in \\mathbb{R}^{N \\times D}$ . Thus, the U-Net conditioning can be formulated as:", + "bbox": [ + 212, + 584, + 787, + 765 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} = \\left[ \\psi (\\mathcal {R}), \\sigma \\left(\\mathcal {H} _ {i}\\right), \\pi \\left(\\mu \\left(\\mathcal {H} _ {t}\\right)\\right) \\right] \\in \\mathbb {R} ^ {(2 N + M) \\times D}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 776, + 785, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The conditioning $\\mathcal{C}$ is fed into the denoising UNet at multiple layers via the cross-attention mechanism [67]. We assume the intermediate feature of a specific", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Learning Egocentric Action Frame Generation", + "bbox": [ + 421, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "UNet layer is $\\mathcal{U}$ , which is learned from the UNet input (i.e., the concatenation of input frame representation $\\mathcal{E}(\\mathcal{X})$ and Gaussian noise $\\mathcal{G}$ ). The cross-attention at this UNet layer can be formulated as:", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nC r o s s A t t (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {D}}\\right) \\cdot V, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 203, + 785, + 237 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $Q = W_{Q} \\cdot \\mathcal{U}$ , $K = W_{K} \\cdot \\mathcal{C}$ and $V = W_{V} \\cdot \\mathcal{C}$ . Note that $W_{Q}, W_{K}$ and $W_{V}$ are learnable matrices. We also adopt the classifier-free guidance following [5] (see supplementary for details). Finally, the UNet output is converted to the image domain by a pre-trained decoder.", + "bbox": [ + 212, + 246, + 787, + 308 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 330, + 457, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "All parameters of the VLLM are initialized from the pre-trained LLaVA [52] weights. During training, we freeze the CLIP image encoder and finetune the projection layer and LLM with cross-entropy loss for 3 epochs. To improve the diversity of the text prompts, we randomly select the question template from 10 candidates in each iteration. For LDM training, the text encoder, UNet and autoencoder are initialized with pre-trained weights [67]. The projection layers $\\sigma$ and $\\mu$ and the self-attention layers $\\pi$ are initialized using the Xavier algorithm [17]. The text encoder is frozen and the remaining weights are finetuned with L2 regression loss between the predicted noise and real noise for 20,000 iterations. All input and target images are resized to a resolution of $256\\times 256$ . Please refer to the supplementary for more details of training and inference.", + "bbox": [ + 212, + 354, + 787, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 545, + 375, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Data and Metrics", + "text_level": 1, + "bbox": [ + 214, + 577, + 408, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. We conduct our experiments on two well-established egocentric action datasets - Ego4D [21] and Epic-Kitchens-100 [13]. Both datasets were densely annotated with action starting time $t$ and ending time $\\hat{t}$ . In our problem setting, we select an egocentric image frame $\\delta_{i}$ seconds before the action begins as the input $\\mathcal{X}$ , and an image $\\delta_{o}$ seconds after the action begins as the target frame $\\mathcal{Y}$ . On the Ego4D dataset, based on the annotations of Pre-Condition-15 time (PRE-15) $t_{pre}$ , and Point-of-No-Return time (PNR) $t_{pnr}$ , we set $\\delta_{i} = t - t_{pre}$ and $\\delta_{o} = t_{pnr} - t$ . For Epic-Kitchens, PNR and PRE-15 annotations are not available. Instead, we empirically select $\\delta_{i} = 0.25$ seconds and $\\delta_{o} = t + \\lambda * (\\hat{t} - t)$ , where $\\lambda = 0.6$ , for our experiments. More details of data preparation and improvement are elaborated in the supplementary.", + "bbox": [ + 212, + 603, + 787, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Metrics. We adopt image-to-image similarity, image-to-text similarity, and user study as metrics in our experiments.", + "bbox": [ + 212, + 768, + 787, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Image-to-Image Metrics. We implement six metrics to evaluate image-to-image similarity. To begin with, we adopt three contrastive learning based", + "bbox": [ + 225, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "B. Lai et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "metrics including image-to-image (1) EgoVLP score [50], (2) EgoVLP $^+$ score [50] and (3) CLIP score [63]. EgoVLP is a contrastive video-language pretraining model trained with egocentric videos. Since EgoVLP takes multiple frames as input, we consider two types of inputs: duplicating the output frame as a static input (i.e., EgoVLP score) and combining the input frame with the output frame (i.e., EgoVLP $^+$ score). As a result, EgoVLP $^+$ can effectively measure whether the generated frame can depict the state transition of an action. Importantly, given that EgoVLP is pre-trained on egocentric data and action labels, we consider EgoVLP and EgoVLP $^+$ score as the primary automatic metrics. In addition, we also report (4) Fréchet Inception Distance (FID) [26], (5) Peak Signal-to-Noise Ratio (PSNR) and (6) Learned Perceptual Image Patch Similarity (LPIPS) [97] (with SqueezeNet [31] as the encoder) to make a thorough evaluation. Note that instead of measuring similarity of input and output frames as in prior works [12,80], in our problem setting, we measure the similarity between the generated action frame and ground truth, which better reflects whether the generation results can illustrate the execution of an action.", + "bbox": [ + 241, + 145, + 787, + 402 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "- Image-to-Text Metrics. We find the widely-used image-to-text CLIP score can not align actions with egocentric images due to the domain gap. Similar misalignment problem is also observed in [14, 59, 72, 85]. In our experiments, we utilize BLIP [47] to generate image captions of the output images and then calculate text-to-text similarity using CLIP text encoder (following [34]). We implement this metric with two BLIP structures: BLIP-B and BLIP-L. Though this solution may still suffer from the same domain gap issue, it is a more appropriate evaluation metric in our problem setting. See more evidence and discussions in the supplementary.", + "bbox": [ + 227, + 405, + 787, + 542 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "- **User Study.** We also conduct a user study on a subset of test data to further validate the advantage of our model based on human preference. We sample 60 examples from each dataset and present the generated frames from our model as well as the baseline models to raters on Amazon Mechanical Turk (AMT). We also provide the input frames and the corresponding action labels during evaluation. For each instance, we ask the raters to select the image that best aligns with the provided action label while preserving the most contextual information from the input frame. To minimize potential bias, we hire 5 AMT raters to annotate each example thus resulting in 300 samples for each dataset. User study interface is shown in supplementary.", + "bbox": [ + 227, + 542, + 787, + 695 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Comparison with Prior Approaches", + "text_level": 1, + "bbox": [ + 215, + 734, + 553, + 750 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We compare our proposed model with previous state-of-the-art methods for text-guided image manipulation, including ProxEdit [24], SDEdit [56] and InstructPix2Pix (IP2P) [5]. For a fair comparison, we finetune these baseline methods with the same data used in our experiments. Specifically, we train IP2P in an end-to-end way on the two datasets with existing egocentric action labels as", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Learning Egocentric Action Frame Generation", + "bbox": [ + 421, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/19d75c7ae80be6358c559b982c27b84ced2d11471693e2064330a3193ac15429.jpg", + "table_caption": [ + "Table 1: Comparison with prior image manipulation approaches in image-to-image metrics. $\\downarrow$ means a lower score in this metric suggests a better performance. The best results are highlighted with **boldface**. The orange row refers to our LEGO model." + ], + "table_footnote": [], + "table_body": "
MethodsEgoVLPEgoVLP+CLIPFID ↓PSNRLPIPS ↓
Ego4DProxEdit [24]44.5172.6868.1733.0111.8840.90
SDEdit [56]50.0772.9073.3533.3511.8141.60
IP2P [5]62.1978.8478.7524.7312.1637.16
LEGO65.6580.4480.6123.8312.2936.43
EK-100ProxEdit [24]32.2752.7765.8051.3511.0646.35
SDEdit [56]33.8456.8074.7627.4111.3043.33
IP2P [5]42.9761.0677.0320.6411.2340.82
LEGO45.8962.6678.6321.5711.3340.36
", + "bbox": [ + 233, + 198, + 767, + 348 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/8644cec964d8578ff76f5b9720226022b5bbe3b385d63f00de313a16cb1e94e1.jpg", + "table_caption": [ + "Table 2: Image-to-text metrics of our model and baselines. The best results are highlighted with boldface. The orange row refers to our LEGO model performance." + ], + "table_footnote": [], + "table_body": "
MethodsEgo4DEpic-Kitchens
BLIP-BBLIP-LBLIP-BBLIP-L
ProxEdit [24]17.7317.3523.6523.39
SDEdit [56]19.8019.7421.5121.30
IP2P [5]20.0020.5625.3726.36
LEGO20.3820.7026.9827.41
", + "bbox": [ + 233, + 440, + 511, + 523 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3d54ba3028473f279513c0b8fc8c34ecbb6c6d9a903308e97aa675438364ae3e.jpg", + "image_caption": [ + "Fig.4: User study of our model and baselines. Win rate is the percentage of each model being picked as the best." + ], + "image_footnote": [], + "bbox": [ + 519, + 362, + 766, + 482 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "text conditioning. ProxEdit and SDEdit rely on the off-the-shelf latent diffusion model to synthesize edited images and thus can not be trained end to end. Therefore, we first finetune the latent diffusion model with egocentric images and action labels and then use the finetuned latent diffusion model parameters to implement ProxEdit and SDEdit approaches. Please refer to the supplementary for more baseline implementation details and a thorough comparison with prior methods that leverage LLMs for image generation.", + "bbox": [ + 212, + 566, + 787, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In terms of image-to-image metrics in Tab. 1, both ProxEdit and SDEdit perform poorly on this novel problem, suggesting the challenging nature of generating egocentric action frames. IP2P performs much better in almost all metrics by end-to-end training and serves as the strongest baseline model in our experiments. However, our LEGO model consistently outperforms IP2P by a large margin (3.46%, 1.60%, 1.86%, 0.90, 0.13 and 0.73% respectively) in all six metrics on Ego4D. LEGO also exceeds IP2P notably (2.92%, 1.60%, 1.60%, 0.10 and 0.46% respectively) in five metrics on Epic-Kitchens. Though, LEGO slightly lags behind IP2P in FID score, it still achieves the second best performance.", + "bbox": [ + 212, + 672, + 787, + 808 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "With regard to image-to-text metrics in Tab. 2, LEGO outperforms the strongest baseline (IP2P) by $0.38\\%$ and $0.14\\%$ on Ego4D and by $1.61\\%$ and", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "B. Lai et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/90f1142b2076eb8b9cd07b80ab0d511b9e20304e25902745a3b4208f97c799d7.jpg", + "image_caption": [ + "Fig. 5: Visualization of the proposed LEGO model and all baselines on Ego4D (the first three rows) and Epic-Kitchens (the last three rows). The action frames generated by LEGO align with the user prompt better and preserve more contexts in input frames." + ], + "image_footnote": [], + "bbox": [ + 240, + 143, + 763, + 556 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "$1.05\\%$ on Epic-Kitchens. The result suggests the synthetic frames from our model can better align with the action descriptions. However, the performance gain is rather limited. We emphasize that this is because the domain gap still exists for the BLIP model. Besides baseline models, we also measure the image-to-text similarity of input frames and action prompts to validate models' capability of learning action state transition in semantics. The BLIP-B/BLIP-L scores are $15.44\\% / 15.49\\%$ on Ego4D and $20.11\\% / 20.52\\%$ on Epic-Kitchens, lagging behind all baseline models. The result evidences diffusion models are able to understand actions and edit the input frame semantically towards the action prompt.", + "bbox": [ + 212, + 642, + 787, + 777 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Due to the limitation of automatic metrics, we additionally implement user study as a complement. We shuffle the order of the results from our model and the baselines while presenting them to the raters. We define the win rate as the percentage of each model being picked as the best out of the total 300", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Learning Egocentric Action Frame Generation", + "bbox": [ + 421, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b816b15b296273af7f7290d15dac4041a1a449993d575d8632bb0ed54e0a83ed.jpg", + "image_caption": [ + "Input Frame" + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 259, + 172 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9f974032032c25ffc259cd1a20f7a7784ebcfa6b9a999132fead39093da75d10.jpg", + "image_caption": [ + "LEGO" + ], + "image_footnote": [], + "bbox": [ + 225, + 203, + 253, + 229 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/51390262476acd0907430caf3bce304be445d5e58f5034221078d342413828d6.jpg", + "image_caption": [ + "cut a portion of clay brush a wood mix with both hands with a brush" + ], + "image_footnote": [], + "bbox": [ + 264, + 145, + 437, + 246 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/69dd1a6c4d01d0466fabc1e3a5dc2d909b01917a994ef16efa11f9be47f9c6f1.jpg", + "image_caption": [ + "\"put tray in oven\"", + "Fig. 6: Additional visualization of our proposed model. LEGO is able to synthesize faithful action frames as well as preserve the contexts in various scenarios." + ], + "image_footnote": [], + "bbox": [ + 439, + 145, + 553, + 246 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e23e14c041747c1193cec14305b1150f3bb23b4654e6c3d8a3ad236a0e9e1560.jpg", + "image_caption": [ + "\"close container\"" + ], + "image_footnote": [], + "bbox": [ + 553, + 146, + 666, + 246 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f0a5be6566924e6471da1f7f68c3bf797b24ddb3f2952e81c61bb4f554e091ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 146, + 782, + 195 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/92b7729b3181df5a05b4551156b81e7bf8296346f53fc9d1ceb8bbf68b7b6d6e.jpg", + "image_caption": [ + "\"take soy milk\"" + ], + "image_footnote": [], + "bbox": [ + 669, + 196, + 782, + 246 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "samples. Results are illustrated in Fig. 4. Our model surpasses the best baseline by $44.00\\%$ and $38.34\\%$ respectively on Ego4D and Epic-Kitchens. The prominent gap further validates the superiority of our model.", + "bbox": [ + 212, + 340, + 784, + 386 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Visualization of Generated Frames", + "text_level": 1, + "bbox": [ + 214, + 407, + 545, + 422 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We additionally showcase examples of generated image from LEGO and baseline models in Fig. 5. ProxEdit and SDEdit fail to understand the user prompt and thus generate frames of irrelevant actions (e.g., row3). They may also easily synthesize the action in a different environment (e.g., row2). InstructPix2Pix is able to preserve more contexts but still fails to semantically align with the action in user prompt (e.g., row1). In contrast, LEGO can synthesize action frames that better align with the user prompts and retain more contexts in the background. More examples of LEGO output are presented in Fig. 6 and supplementary.", + "bbox": [ + 212, + 431, + 784, + 551 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 574, + 388, + 589 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We present comprehensive ablation studies to investigate the contribution of enriched action descriptions, the VLLM image embedding and the VLLM text embedding separately. Results are demonstrated in Tab. 3. Given the limitation of automatic metrics, we also provide user study as the primary metric. Notably, conditioning on enriched action descriptions can moderately improve the model performance, supporting our claim that expanding the action description can facilitate the learning of the state transition during an egocentric action. Moreover, utilizing the image embedding and text embedding from VLLM as additional conditions both improve the model performance by a notable margin because VLLM embeddings can effectively narrow the domain gap. Interestingly, the image embedding leads to larger performance boost on both datasets. These results suggest the VLLM image embedding $\\mathcal{H}_i$ incorporates important high-level semantic meanings that are not captured in the auto-encoder image embedding $\\mathcal{E}(\\mathcal{X})$ or VLLM text emebdding $\\mathcal{H}_t$ . Finally, our full LEGO model uses both VLLM image and text embeddings (Desc.+Joint Embed.) and thus achieves the best performance on both datasets in all metrics.", + "bbox": [ + 212, + 598, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "B. Lai et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e461f5dae280bc574483fbb083e5ed11567abb13d0383636704d4690699a12f5.jpg", + "table_caption": [ + "Table 3: Analysis of egocentric action frame generation performance with different conditionings. Joint Embed. refers to incorporating both VLLM image and text embeddings. Similar to Fig. 4, we present win rate as the user study result, i.e., the percentage of each model picked as the best (\\% is omitted for simplicity). The best results are highlighted with **boldface**. The orange rows refer to our full LEGO model." + ], + "table_footnote": [], + "table_body": "
ConditioningEgo4DEpic-Kitchens
User StudyEgoVLPEgoVLP+CLIPUser StudyEgoVLPEgoVLP+CLIP
Actions Labels5.3362.1978.8478.757.0842.9761.0677.03
Descriptions13.0062.9179.0979.1812.5043.7261.4677.47
Desc. + Img Embed.26.0065.3580.1380.5724.1745.8262.2978.60
Desc. + Txt Embed.21.3363.2979.4079.2122.0844.6862.0277.74
Desc. + Joint Embed.34.3465.6580.4480.6134.1745.8962.6678.63
", + "bbox": [ + 215, + 226, + 792, + 321 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/32649c8a74a3cc65dea9385aaefc6800e87309d0a49fe775f110c3d19bfc0af8.jpg", + "table_caption": [ + "Table 4: Performance of LEGO without finetuning (w/o FT). The best results are highlighted with **boldface**. The orange rows refer to our full LEGO model." + ], + "table_footnote": [], + "table_body": "
ConditioningEgo4DEpic-Kitchens
EgoVLPEgoVLP+CLIPEgoVLPEgoVLP+CLIP
Descriptions62.9179.0979.1843.7261.4677.47
Desc.+Joint Embed.(w/o FT)64.5779.7280.2344.7461.8778.35
Desc.+Joint Embed.(w/ FT)65.6580.4480.6145.8962.6678.63
", + "bbox": [ + 215, + 378, + 790, + 470 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Analysis of Visual Instruction Tuning", + "text_level": 1, + "bbox": [ + 214, + 503, + 571, + 520 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "First, we assess the quality of enriched action descriptions from the visual instruction tuned VLLM by user study. For each sample in user study, we ask the raters to select whether the description aligns with the input and target frames (see supplementary for more details of user study setting and interface). The percentage of samples with aligned frames and action descriptions is $87\\%$ on Ego4D and $84\\%$ on Epic-Kitchens. The high alignment percentage suggests the visual instruction tuned VLLM can effectively expand action labels with details captured in the input frame. We additionally implement the same user study for off-the-shelf VLLM (i.e., without finetuning). The percentage of alignment is just $27\\%$ on Ego4D and $30\\%$ on Epic-Kitchens with hallucination existing in $92\\%$ of unaligned instances. The notable drop supports our argument that visual instruction tuning is a critical step for high-quality prompt enrichment.", + "bbox": [ + 212, + 536, + 787, + 717 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In addition, we further investigate whether the visual instruction tuned embeddings can contribute more to the diffusion model than the off-the-self counterpart, which has not been well studied in prior work. As shown in Tab. 4, without any finetuning, the image and text embeddings from VLLM can still improve the baseline model (Descriptions). However finetuned embeddings yield much larger improvement (e.g., $1.08\\%$ and $1.15\\%$ in EgoVLP score) on both datasets. The result suggests that visual instruction tuning is a necessary step to learn semantic image and text embeddings that are more aligned with the", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Learning Egocentric Action Frame Generation", + "bbox": [ + 421, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/771ec954a02730affa2393565be317566054377595ac6d16570e1c534346b433.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 162, + 379, + 215 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/bffe851e6fc13a86dc8e98805295c967d1203ada550d9ff9fc4f59412a8907e1.jpg", + "image_caption": [ + "Fig. 7: Visualization of generating various actions from the same input frame. The first action (\"open drawer\") is the existing label for this example in the dataset. We fill another five actions into the user query and synthesize the action frames using our model. All generated frames align well with the actions and preserve most contexts." + ], + "image_footnote": [], + "bbox": [ + 217, + 244, + 254, + 268 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "\"Can you provide instructions on how to {action} in my current situation?\"", + "bbox": [ + 264, + 217, + 375, + 267 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ed431d60297d8bcc73609181e5df9d3afcb5f6ca1ed5ccbcd5b922028eee54ab.jpg", + "image_caption": [ + "\"open drawer\"" + ], + "image_footnote": [], + "bbox": [ + 395, + 148, + 511, + 200 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/cc4e8453b5e774f08fca5106185f04ba980f670673820f7b3fd466669a0414a0.jpg", + "image_caption": [ + "\"dry hands\"" + ], + "image_footnote": [], + "bbox": [ + 514, + 148, + 627, + 200 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d994be76e84a8483fc6dbd67883afd705ed53feff5fed79fa396627070db72d6.jpg", + "image_caption": [ + "\"cut cucumber\"" + ], + "image_footnote": [], + "bbox": [ + 630, + 148, + 745, + 200 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/deb5960858bafc06089cc1a8dad5944d634f12dd52a0aee0a0ac7a770d5f5b53.jpg", + "image_caption": [ + "\"open microwave\"" + ], + "image_footnote": [], + "bbox": [ + 397, + 214, + 511, + 266 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/159d8b797a447c7c0b0a3320727e86d8ae99283a87d17b12c2921e2a07ce4dbb.jpg", + "image_caption": [ + "\"pick up bowl\"" + ], + "image_footnote": [], + "bbox": [ + 514, + 214, + 629, + 266 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/007d64c34566d8aed4e12ee1ff1f71551abcb90da2698d560bd335db0cb30f87.jpg", + "image_caption": [ + "\"take knife\"" + ], + "image_footnote": [], + "bbox": [ + 630, + 214, + 745, + 266 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0ddb6f1ecf13a62554b791b575a2f3902e670bd68daf08d1b82238226dc3309f.jpg", + "image_caption": [ + "LEGO" + ], + "image_footnote": [], + "bbox": [ + 751, + 239, + 779, + 265 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "egocentric input frame and action prompt, which thus narrows the domain gap and greatly boosts the performance of the latent diffusion model.", + "bbox": [ + 212, + 385, + 784, + 415 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.6 Generation of Various Actions with the Same Contexts", + "text_level": 1, + "bbox": [ + 214, + 441, + 717, + 455 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In addition to generating action frames based on the pre-defined action labels in our datasets, we validate the generative nature of our model i.e., whether LEGO is able to synthesize the correct action frames conditioning on the same contexts yet different actions (novel image-action pairs). Results are illustrated in Fig. 7. We feed different actions with the same input frame to our model. The synthesized frames correctly depict the execution of the actions in the same environment. This result further indicates that our model can understand the action state transition and generalize to different user queries.", + "bbox": [ + 212, + 472, + 787, + 593 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 622, + 359, + 637 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce the novel problem of egocentric action frame generation. We also propose a novel model — LEGO that leverages visual instruction tuning and a diffusion model to address this problem. Our key intuition is to use visual instruction tuning to generate informative responses that depict the execution of the egocentric action, and then design the conditioning for the denoising U-Net to exploit the image and text embeddings from a visual instruction tuned VLLM. Our experiments on two large-scale egocentric action datasets validate the advantage of the proposed approach as well as the contribution of each model component. We believe our work provides an important step in understanding the action state transition and controllability of diffusion models, and suggests future research directions in egocentric AI systems, action state transition, image generation and human skill transfer.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "B. Lai et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Portions of this work were supported in part by a gift from Meta. We thank Sangmin Lee for the valuable discussion and suggestions.", + "bbox": [ + 212, + 172, + 785, + 204 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 224, + 323, + 239 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)", + "2. Anonymity: Making multimodal generation easier: When diffusion models meet llms. openreview (2023)", + "3. Ashutosh, K., Girdhar, R., Torresani, L., Grauman, K.: Hiervl: Learning hierarchical video-language embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23066-23078 (2023)", + "4. Baskin, J.H., Edersheim, J.G., Price, B.H.: Is a picture worth a thousand words? neuroimaging in the courtroom. American Journal of Law & Medicine 33(2-3), 239-269 (2007)", + "5. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)", + "6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)", + "7. Chakrabarty, T., Singh, K., Saakyan, A., Muresan, S.: Learning to follow object-centric image editing instructions faithfully. arXiv preprint arXiv:2310.19145 (2023)", + "8. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., et al.: Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426 (2023)", + "9. Chen, W.G., Spiridonova, I., Yang, J., Gao, J., Li, C.: Llava-interactive: An all-in-one demo for image chat, segmentation, generation and editing. arXiv preprint arXiv:2311.00571 (2023)", + "10. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022)", + "11. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427 (2022)", + "12. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. Advances in Neural Information Processing Systems (2023)", + "13. Damen, D., Doughty, H., Farinella, G.M., Furnari, A., Kazakos, E., Ma, J., Moltisanti, D., Munro, J., Perrett, T., Price, W., et al.: Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision pp. 1-23 (2022)", + "14. Du, Y., Yang, S., Dai, B., Dai, H., Nachum, O., Tenenbaum, J.B., Schuurmans, D., Abbeel, P.: Learning universal policies via text-guided video generation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + ], + "bbox": [ + 225, + 253, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Learning Egocentric Action Frame Generation", + "bbox": [ + 421, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Epstein, D., Jabri, A., Poole, B., Efros, A.A., Holynski, A.: Diffusion self-guidance for controllable image generation. In: Advances in Neural Information Processing Systems (2023)", + "16. Girdhar, R., Grauman, K.: Anticipative video transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13505-13515 (2021)", + "17. Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedforward neural networks. In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. pp. 249-256. JMLR Workshop and Conference Proceedings (2010)", + "18. Goel, V., Peruzzo, E., Jiang, Y., Xu, D., Sebe, N., Darrell, T., Wang, Z., Shi, H.: Pair-diffusion: Object-level image editing with structure-and-appearance paired diffusion models. arXiv preprint arXiv:2303.17546 (2023)", + "19. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139–144 (2020)", + "20. Goyal, M., Modi, S., Goyal, R., Gupta, S.: Human hands as probes for interactive object understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3293-3303 (2022)", + "21. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18995-19012 (2022)", + "22. Hafri, A., Trueswell, J.C., Epstein, R.A.: Neural representations of observed actions generalize across static and dynamic visual input. Journal of Neuroscience 37(11), 3056-3071 (2017)", + "23. Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., et al.: Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905 (2023)", + "24. Han, L., Wen, S., Chen, Q., Zhang, Z., Song, K., Ren, M., Gao, R., Stathopoulos, A., He, X., Chen, Y., et al.: Proxedit: Improving tuning-free real image editing with proximal guidance. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4291-4301 (2024)", + "25. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022)", + "26. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)", + "27. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)", + "28. Huang, J., Liu, Y., Qin, J., Chen, S.: Kv inversion: Kv embeddings learning for text-conditioned real image action editing. arXiv preprint arXiv:2309.16608 (2023)", + "29. Huang, Y., Cai, M., Li, Z., Lu, F., Sato, Y.: Mutual context network for jointly estimating egocentric gaze and action. IEEE Transactions on Image Processing 29, 7795-7806 (2020)", + "30. Huang, Y., Cai, M., Li, Z., Sato, Y.: Predicting gaze in egocentric video by learning task-dependent attention transition. In: Proceedings of the European conference on computer vision (ECCV). pp. 754-769 (2018)" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "B. Lai et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Iandola, F.N., Han, S., Moskewicz, M.W., Ashraf, K., Dally, W.J., Keutzer, K.: SqueezeNet: Alexnet-level accuracy with 50x fewer parameters and $< 0.5$ mb model size. arXiv preprint arXiv:1602.07360 (2016)", + "32. Jia, W., Liu, M., Rehg, J.M.: Generative adversarial network for future hand segmentation from egocentric video. In: European Conference on Computer Vision. pp. 639-656. Springer (2022)", + "33. Jiang, Y., Zhang, Z., Xue, T., Gu, J.: Autodir: Automatic all-in-one image restoration with latent diffusion. arXiv preprint arXiv:2310.10123 (2023)", + "34. Joseph, K., Udhayanan, P., Shukla, T., Agarwal, A., Karanam, S., Goswami, K., Srinivasan, B.V.: Iterative multi-granular image editing using diffusion models. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2024)", + "35. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. In: Advances in Neural Information Processing Systems (2022)", + "36. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)", + "37. Kazakos, E., Nagrani, A., Zisserman, A., Damen, D.: Epic-fusion: Audio-visual temporal binding for egocentric action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5492-5501 (2019)", + "38. Kim, S., Jang, W., Kim, H., Kim, J., Choi, Y., Kim, S., Lee, G.: User-friendly image editing with minimal text input: Leveraging captioning and injection techniques. arXiv preprint arXiv:2306.02717 (2023)", + "39. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)", + "40. Koh, J.Y., Fried, D., Salakhutdinov, R.R.: Generating images with multimodal language models. Advances in Neural Information Processing Systems 36 (2024)", + "41. Lai, B., Liu, M., Ryan, F., Rehg, J.M.: In the eye of transformer: Global-local correlation for egocentric gaze estimation. British Machine Vision Conference (2022)", + "42. Lai, B., Liu, M., Ryan, F., Rehg, J.M.: In the eye of transformer: Global-local correlation for egocentric gaze estimation and beyond. International Journal of Computer Vision 132(3), 854-871 (2024)", + "43. Lai, B., Ryan, F., Jia, W., Liu, M., Rehg, J.M.: Listen to look into the future: Audio-visual egocentric gaze anticipation. arXiv preprint arXiv:2305.03907 (2023)", + "44. Li, D., Li, J., Hoi, S.C.: Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems (2023)", + "45. Li, J., Liu, K., Wu, J.: Ego-body pose estimation via ego-head pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17142-17151 (2023)", + "46. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. International Conference on Machine Learning (2023)", + "47. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Learning Egocentric Action Frame Generation", + "bbox": [ + 421, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Li, Y., Liu, M., Rehg, J.M.: In the eye of beholder: Joint learning of gaze and actions in first person video. In: Proceedings of the European conference on computer vision (ECCV). pp. 619-635 (2018)", + "49. Lian, L., Li, B., Yala, A., Darrell, T.: Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655 (2023)", + "50. Lin, K.Q., Wang, J., Soldan, M., Wray, M., Yan, R., XU, E.Z., Gao, D., Tu, R.C., Zhao, W., Kong, W., et al.: Egocentric video-language pretraining. Advances in Neural Information Processing Systems 35, 7575-7586 (2022)", + "51. Liu, B., Zhang, H., Liu, J., Wang, Q.: Acigs: An automated large-scale crops image generation system based on large visual language multi-modal models. In: 2023 20th Annual IEEE International Conference on Sensing, Communication, and Networking (SECON). pp. 7-13. IEEE (2023)", + "52. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems (2023)", + "53. Liu, M., Tang, S., Li, Y., Rehg, J.M.: Forecasting human-object interaction: joint prediction of motor attention and actions in first person video. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 704-721. Springer (2020)", + "54. Liu, S., Tripathi, S., Majumdar, S., Wang, X.: Joint hand motion and interaction hotspots prediction from egocentric videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3282-3292 (2022)", + "55. Luo, Z., Hachiuma, R., Yuan, Y., Kitani, K.: Dynamics-regulated kinematic policy for egocentric pose estimation. Advances in Neural Information Processing Systems 34, 25019-25032 (2021)", + "56. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: International Conference on Learning Representations (2022)", + "57. Mirzaei, A., Aumentado-Armstrong, T., Brubaker, M.A., Kelly, J., Levinshtein, A., Derpanis, K.G., Gilitschenski, I.: Watch your steps: Local image and scene editing by text instructions. arXiv preprint arXiv:2308.08947 (2023)", + "58. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6038-6047 (2023)", + "59. Molad, E., Horwitz, E., Valevski, D., Acha, A.R., Matias, Y., Pritch, Y., Leviathan, Y., Hoshen, Y.: Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329 (2023)", + "60. Orgad, H., Kawar, B., Belinkov, Y.: Editing implicit assumptions in text-to-image diffusion models. arXiv preprint arXiv:2303.08084 (2023)", + "61. Pan, Z., Gherardi, R., Xie, X., Huang, S.: Effective real image editing with accelerated iterative diffusion inversion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15912-15921 (2023)", + "62. Pramanick, S., Song, Y., Nag, S., Lin, K.Q., Shah, H., Shou, M.Z., Chellappa, R., Zhang, P.: Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5285-5297 (2023)", + "63. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "B. Lai et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)", + "65. Ragusa, F., Farinella, G.M., Furnari, A.: Stillfast: An end-to-end approach for short-term object interaction anticipation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3635-3644 (2023)", + "66. Ramakrishnan, S.K., Al-Halah, Z., Grauman, K.: Naq: Leveraging narrations as queries to supervise episodic memory. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6694-6703 (2023)", + "67. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "68. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)", + "69. Ryan, F., Jiang, H., Shukla, A., Rehg, J.M., Ithapu, V.K.: Egocentric auditory attention localization in conversations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14663-14674 (2023)", + "70. Shen, Y., Song, K., Tan, X., Li, D., Lu, W., Zhuang, Y.: Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580 (2023)", + "71. Shi, J., Xiong, W., Lin, Z., Jung, H.J.: Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411 (2023)", + "72. Stein, G., Cresswell, J.C., Hosseinzadeh, R., Sui, Y., Ross, B.L., Villecloze, V., Liu, Z., Caterini, A.L., Taylor, J.E.T., Loaiza-Ganem, G.: Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. arXiv preprint arXiv:2306.04675 (2023)", + "73. Su, Y., Lan, T., Li, H., Xu, J., Wang, Y., Cai, D.: Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355 (2023)", + "74. Sudhakaran, S., Escalera, S., Lanz, O.: Lsta: Long short-term attention for egocentric action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9954-9963 (2019)", + "75. Sun, Z., Zhou, Y., He, H., Mok, P.: Sgdiff: A style guided diffusion model for fashion synthesis. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 8433-8442 (2023)", + "76. Thoppilan, R., De Freitas, D., Hall, J., Shazeer, N., Kulshreshtha, A., Cheng, H.T., Jin, A., Bos, T., Baker, L., Du, Y., et al.: Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239 (2022)", + "77. Tome, D., Alldieck, T., Peluse, P., Pons-Moll, G., Agapito, L., Badino, H., De la Torre, F.: Selfpose: 3d egocentric pose estimation from a headset mounted camera. IEEE Transactions on Pattern Analysis and Machine Intelligence (2020)", + "78. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)", + "79. Tsaban, L., Passos, A.: Ledits: Real image editing with ddpm inversion and semantic guidance. arXiv preprint arXiv:2307.00522 (2023)", + "80. Wallace, B., Gokul, A., Naik, N.: Edict: Exact diffusion inversion via coupled transformations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22532-22541 (2023)" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Learning Egocentric Action Frame Generation", + "bbox": [ + 421, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "81. Wang, J., Luvizon, D., Xu, W., Liu, L., Sarkar, K., Theobalt, C.: Scene-aware egocentric 3d human pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13031-13040 (2023)", + "82. Wang, K., Yang, F., Yang, S., Butt, M.A., van de Weijer, J.: Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. arXiv preprint arXiv:2309.15664 (2023)", + "83. Wang, Q., Zhang, B., Birsak, M., Wonka, P.: Instructedit: Improving automatic masks for diffusion-based image editing with user instructions. arXiv preprint arXiv:2305.18047 (2023)", + "84. Wang, Q., Zhang, B., Birsak, M., Wonka, P.: Mdp: A generalized framework for text-guided image editing by manipulating the diffusion path. arXiv preprint arXiv:2303.16765 (2023)", + "85. Wang, W., Xie, K., Liu, Z., Chen, H., Cao, Y., Wang, X., Shen, C.: Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599 (2023)", + "86. Wang, X., Zhu, L., Wang, H., Yang, Y.: Interactive prototype learning for egocentric action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8168-8177 (2021)", + "87. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848 (2023)", + "88. Wen, S., Fang, G., Zhang, R., Gao, P., Dong, H., Metaxas, D.: Improving compositional text-to-image generation with large vision-language models. arXiv preprint arXiv:2310.06311 (2023)", + "89. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023)", + "90. Wu, S., Fei, H., Qu, L., Ji, W., Chua, T.S.: Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519 (2023)", + "91. Xu, Y., Li, Y.L., Huang, Z., Liu, M.X., Lu, C., Tai, Y.W., Tang, C.K.: Egopca: A new framework for egocentric hand-object interaction understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5273-5284 (2023)", + "92. Ye, Y., Li, X., Gupta, A., De Mello, S., Birchfield, S., Song, J., Tulsiani, S., Liu, S.: Affordance diffusion: Synthesizing hand-object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22479-22489 (2023)", + "93. Yu, Q., Li, J., Ye, W., Tang, S., Zhuang, Y.: Interactive data synthesis for systematic vision adaptation via llms-aigcs collaboration. arXiv preprint arXiv:2305.12799 (2023)", + "94. Yu, Z., Li, H., Fu, F., Miao, X., Cui, B.: Fisted: Accelerating text-to-image editing via cache-enabled sparse diffusion inference. arXiv preprint arXiv:2305.17423 (2023)", + "95. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858 (2023)", + "96. Zhang, M., Teck Ma, K., Hwee Lim, J., Zhao, Q., Feng, J.: Deep future gaze: Gaze anticipation on egocentric videos using adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4372-4381 (2017)" + ], + "bbox": [ + 222, + 146, + 785, + 839 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "B. Lai et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 19 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "97. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)", + "98. Zhang, S., Yang, X., Feng, Y., Qin, C., Chen, C.C., Yu, N., Chen, Z., Wang, H., Savarese, S., Ermon, S., et al.: Hive: Harnessing human feedback for instructional visual editing. arXiv preprint arXiv:2303.09618 (2023)", + "99. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)", + "100. Zhang, Y., Huang, N., Tang, F., Huang, H., Ma, C., Dong, W., Xu, C.: Inversion-based style transfer with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10146-10156 (2023)", + "101. Zhang, Z., Han, L., Ghosh, A., Metaxas, D.N., Ren, J.: Sine: Single image editing with text-to-image diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6027-6037 (2023)", + "102. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + ], + "bbox": [ + 217, + 146, + 787, + 397 + ], + "page_idx": 20 + }, + { + "type": "header", + "text": "Learning Egocentric Action Frame Generation", + "bbox": [ + 421, + 114, + 730, + 128 + ], + "page_idx": 20 + }, + { + "type": "page_number", + "text": "21", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 20 + } +] \ No newline at end of file diff --git a/2024/LEGO_ Learning EGOcentric Action Frame Generation via Visual Instruction Tuning/13a5a5c4-078d-4d85-be9d-7966d6d6437e_model.json b/2024/LEGO_ Learning EGOcentric Action Frame Generation via Visual Instruction Tuning/13a5a5c4-078d-4d85-be9d-7966d6d6437e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f5c9c2a6075569b6d8339469e41507b3d1bc8310 --- /dev/null +++ b/2024/LEGO_ Learning EGOcentric Action Frame Generation via Visual Instruction Tuning/13a5a5c4-078d-4d85-be9d-7966d6d6437e_model.json @@ -0,0 +1,3399 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.255, + 0.141, + 0.75, + 0.187 + ], + "angle": 0, + "content": "LEGO: Learning EGOcentric Action Frame Generation via Visual Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.212, + 0.728, + 0.244 + ], + "angle": 0, + "content": "Bolin Lai\\(^{1,2,\\dagger}\\), Xiaoliang Dai\\(^{1}\\), Lawrence Chen\\(^{1}\\), Guan Pang\\(^{1}\\), James M. Rehg\\(^{3}\\), and Miao Liu\\(^{1,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.45, + 0.255, + 0.554, + 0.268 + ], + "angle": 0, + "content": "1 GenAI, Meta" + }, + { + "type": "text", + "bbox": [ + 0.388, + 0.269, + 0.615, + 0.284 + ], + "angle": 0, + "content": "2 Georgia Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.359, + 0.283, + 0.644, + 0.298 + ], + "angle": 0, + "content": "3 University of Illinois Urbana-Champaign" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.298, + 0.78, + 0.326 + ], + "angle": 0, + "content": "bolin.lai@gatech.edu {xiaoliangdai,lawrencechen,gpang,miaoliu}@meta.com jrehg@illinois.edu" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.384, + 0.266, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.295, + 0.354, + 0.373, + 0.366 + ], + "angle": 0, + "content": "User Prompt" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.371, + 0.389, + 0.42 + ], + "angle": 0, + "content": "\"Can you provide instructions on how to wash the trouser with the brush in my current situation?\"" + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.352, + 0.519, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.435, + 0.346, + 0.486 + ], + "angle": 0, + "content": "1. Check the Care Label: Check the care label on your trousers for any specific washing instructions." + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.496, + 0.341, + 0.536 + ], + "angle": 0, + "content": "2. Prep the Area: Find a clean and flat surface to work on, such as a table or countertop." + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.546, + 0.337, + 0.569 + ], + "angle": 0, + "content": "3. Spot Check Stains:\" :" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.58, + 0.34, + 0.592 + ], + "angle": 0, + "content": "(a) LLM Response" + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.55, + 0.388, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.438, + 0.523, + 0.458 + ], + "angle": 0, + "content": "1. You should submerge the trouser in the water." + }, + { + "type": "text", + "bbox": [ + 0.41, + 0.468, + 0.524, + 0.518 + ], + "angle": 0, + "content": "2. Use the brush to scrub the trouser; focusing on any stains or areas that may require extra attention." + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.529, + 0.515, + 0.558 + ], + "angle": 0, + "content": "3. Once the trouser is clean, you should rinse it.\"" + }, + { + "type": "text", + "bbox": [ + 0.464, + 0.561, + 0.471, + 0.571 + ], + "angle": 0, + "content": "中" + }, + { + "type": "text", + "bbox": [ + 0.405, + 0.58, + 0.556, + 0.592 + ], + "angle": 0, + "content": "(b) Visual LLM Response" + }, + { + "type": "title", + "bbox": [ + 0.54, + 0.35, + 0.781, + 0.363 + ], + "angle": 0, + "content": "Actions in Existing Egocentric Datasets" + }, + { + "type": "text", + "bbox": [ + 0.556, + 0.366, + 0.744, + 0.439 + ], + "angle": 0, + "content": "tackle open the tap \npuncture the sandal \nrub oil into dough place mug \nturn the mothendose sew the cloth doll \nwash trouser with brush \nsqueeze ball of dough \nclean a sieve shake the spray adhesive \nclean rice adjust gasattach the wire \nplace spice \nopen fridge pour the topsoil \nclean place ball" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.551, + 0.566, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.586, + 0.462, + 0.752, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.543, + 0.785, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.586, + 0.58, + 0.781, + 0.592 + ], + "angle": 0, + "content": "(c) Our model (LEGO) Response" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.731 + ], + "angle": 0, + "content": "Fig. 1: When a user performing a complex task asks a large language model (LLM) for instructions (a) on how to complete the steps, she receives a generic answer and has to translate the guidance into her specific situation. If she is wearing a camera, then the prompt can be augmented with an egocentric view of the scene and passed to a Visual LLM (b), and the description is now contextualized to her situation. But she still faces the challenge of parsing a written description. When she uses our novel LEGO model (c), however, the combined image and prompt are used to automatically generate an image that provides visual guidance tailored to her exact situation from the egocentric viewpoint. Now she can complete her task seamlessly!" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.763, + 0.741, + 0.804 + ], + "angle": 0, + "content": "Abstract. Generating instructional images of human daily actions from an egocentric viewpoint serves as a key step towards efficient skill transfer. In this paper, we introduce a novel problem - egocentric action" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.811, + 0.727, + 0.826 + ], + "angle": 0, + "content": "This work was done when the first author was an intern at GenAI, Meta." + }, + { + "type": "page_footnote", + "bbox": [ + 0.235, + 0.827, + 0.392, + 0.84 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.232, + 0.811, + 0.727, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "B. Lai et al." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.148, + 0.744, + 0.414 + ], + "angle": 0, + "content": "frame generation. The goal is to synthesize an image depicting an action in the user's context (i.e., action frame) by conditioning on a user prompt and an input egocentric image. Notably, existing egocentric action datasets lack the detailed annotations that describe the execution of actions. Additionally, existing diffusion-based image manipulation models are sub-optimal in controlling the state transition of an action in egocentric image pixel space because of the domain gap. To this end, we propose to Learn EGOcentric (LEGO) action frame generation via visual instruction tuning. First, we introduce a prompt enhancement scheme to generate enriched action descriptions from a visual large language model (VLLM) by visual instruction tuning. Then we propose a novel method to leverage image and text embeddings from the VLLM as additional conditioning to improve the performance of a diffusion model. We validate our model on two egocentric datasets - Ego4D and Epic-Kitchens. Our experiments show substantial improvement over prior image manipulation models in both quantitative and qualitative evaluation. We also conduct detailed ablation studies and analysis to provide insights in our method. More details of the dataset and code are available on the website (https://bolinlai.github.io/Lego_EgoActGen/)." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.424, + 0.731, + 0.438 + ], + "angle": 0, + "content": "Keywords: Egocentric Vision \\(\\cdot\\) Instruction Tuning \\(\\cdot\\) Diffusion Model" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.463, + 0.377, + 0.48 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.789, + 0.75 + ], + "angle": 0, + "content": "The emergence of Large Language Models (LLMs) [6, 10, 64, 99], such as ChatGPT, has revolutionized the transfer of knowledge. However, an LLM alone is not a sufficient tool for human skill transfer. Consider the question answering example in Fig. 1(a). The LLM can summarize general world knowledge, but its response may not be directly applicable to the user's current circumstances. To bridge this gap, egocentric visual perception provides a novel means to capture the actions and intentions as well as the surrounding context of the camera wearer. As shown in Fig. 1(b), recent Visual Large Language Models (VLLMs) [1, 12, 46, 52, 102] can generate instructions based on the egocentric visual input. However, such verbose textual instructions are not the optimal medium for enabling efficient human skill transfer (e.g., via AR devices). Neuroscience studies have revealed that the human brain processes text much more slowly than images [4], and that humans can interpret an action from a single static image [22]. Motivated by these discoveries, we seek to design an image generation architecture that can synthesize an image which not only vividly depicts how an action should be conducted, but also seamlessly aligns with the user's visual perspective." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.79, + 0.841 + ], + "angle": 0, + "content": "Formally, we introduce the novel problem of egocentric action frame generation as depicted in Fig. 1(c). Given a user query about how to perform a specific action and an egocentric image capturing the moment before the action begins, the goal is to synthesize an egocentric image illustrating the execution of the action in the same egocentric context. We address this problem by harnessing diffusion models [27,67], which have been shown to be powerful tools for image" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.422, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Learning Egocentric Action Frame Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "manipulation [5,25,44,56,100]. There are two major challenges in using diffusion models to generate action frames from an egocentric perspective. First, the action annotations of the existing egocentric datasets [13, 21] are simply composed of a verb and nouns (see word cloud in Fig. 1), and thus lack the necessary details for diffusion models to learn the action state transition and to associate the action with correct objects and body movements. Second, the existing diffusion models are pre-trained primarily on exocentric (third-person-view) images, and have limited ability to represent complicated human daily activities from an egocentric perspective. In contrast, our proposed problem requires image generation in the egocentric view, conditioned on the user prompt of actions. The resulting domain gap impedes existing methods from synthesizing faithful and consistent egocentric action frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.51 + ], + "angle": 0, + "content": "To address these challenges, we propose to Learn EGOcentric (LEGO) action frame generation with visual instruction tuning. First, we introduce a prompt enhancement scheme to generate enriched action descriptions at scale from an instruction-tuned VLLM. Second, we incorporate the image and text embeddings from finetuned VLLM as additional conditioning into the diffusion model to narrow the domain gap and improve the controllability of action frame generation. Our experimental results suggest that the enriched action descriptions and our innovative utilization of VLLM embeddings both improve the image generation performance. Our model is able to provide a generated key action frame together with a detailed action descriptions to facilitate human skill transfer from the egocentric perspective. Overall, our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.524, + 0.784, + 0.568 + ], + "angle": 0, + "content": "- We introduce the novel problem of egocentric action frame generation to facilitate the process of skill transfer and address the challenges of missing action details in prompts and domain gap in existing image diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.57, + 0.785, + 0.629 + ], + "angle": 0, + "content": "- We propose a prompt enhancement strategy based on visual instruction tuning to enrich egocentric action descriptions, and demonstrate how the enriched descriptions can help the diffusion model understand the action state transition from the egocentric perspective." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.631, + 0.784, + 0.674 + ], + "angle": 0, + "content": "- We propose a novel approach to incorporate the text and visual embeddings from the VLLM into the latent diffusion model to bridge the domain gap and improve the performance for egocentric action frame generation." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.676, + 0.785, + 0.75 + ], + "angle": 0, + "content": "- We conduct thorough experiments on the Ego4D and Epic-Kitchens datasets to validate the superiority of our model over prior approaches. We also showcase the contribution of each component of our model design through ablation studies. We further provide analysis on how the visual instruction-tuned embeddings benefit model performance." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.524, + 0.785, + 0.75 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.776, + 0.388, + 0.792 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Text-Guided Image Manipulation. The recent emergence of diffusion models enables text-guided image manipulation including image restoration [33], style" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "B. Lai et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.405 + ], + "angle": 0, + "content": "transfer [75], personalized image synthesis [68,71,87], pose generation [28,36] and generic image editing [15,18,34,38,44,57,60,79,82,82-84,94,98,101]. SDEdit [56] converts the image to the latent space by adding noise through a stochastic differential equation and then denoises the representation for image editing. Rombach et al. [67] further expand SDEdit from the original stroke-based editing to text-based editing. Null-text inversion (NTI) [58] inverts a real image by DDIM [35] to yield the diffusion process and then reconstructs the image. The image can then be edited following the same strategies as Prompt-to-Prompt [25]. NTI relies on accurate inversion process which can be improved by using coupled transformations [80] or proximal guidance [24] and accelerated by a blended guidance strategy [61]. To associate the nouns with correct objects, DIFFEDIT [11] generates a mask to localize the manipulated regions. However, most inversion-based methods require accurate image captions, which are largely unavailable in the existing egocentric dataset. Recently, InstructPix2Pix [5] demonstrates the potential to edit a real image without the inversion process and original captions. However, how to leverage the diffusion model to control the state transition of an action within the egocentric image plane remains unexplored." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.788, + 0.669 + ], + "angle": 0, + "content": "Large Language Model for Image Generation. LLMs [6, 10, 64, 76, 78, 99] and VLLMs [1, 12, 23, 46, 52, 73, 95, 102] have shown their strong capability of understanding complex human instructions. Recently, LLMs and VLLMs are used to guide image generation [2, 7-9, 51, 93, 102]. Wen et al. [88] use a pretrained VLLM to pinpoint the misalignment of the text prompt and the synthetic image and then correct it using the diffusion model. Lian et al. [49] propose to generate a layout map using an LLM to improve the understanding of prompts with spatial and numerical reasoning. InstructEdit [83] uses BLIP-2 [46] to infer the objects that will be edited and then generates a mask with SAM [39] for object grounding. A pre-trained LLM can also be used as a controller to connect with various foundation models [70, 89, 90]. GILL [40] learns text-relevant image embeddings from VLLM in a few additional tokens for image generation. Importantly, all previous efforts apply the off-the-shelf foundational models directly to their problems without finetuning. In contrast, our method uses visual instruction tuning to improve the image and text embeddings from the VLLM, which narrows the domain gap and thereby facilitates the action frame generation from the egocentric point of view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Egocentric Vision. Recent efforts seek to understand human actions and perceptual attention [16, 29, 30, 37, 41-43, 48, 69, 74, 86], model hand-object interactions [20, 53, 54, 65, 91], and estimate human body poses [45, 55, 77, 81] from the egocentric perspective. Here, we mainly discuss the most relevant works on egocentric visual-language models and egocentric visual content generation. Lin et al. [50] propose the first egocentric video-language pre-training model - EgoVLP. Pramanick et al. [62] further improve it by incorporating multi-modal fusion directly into the video and language backbone. Ashutosh et al. [3] propose to learn a hierarchical video-language embedding for long egocentric videos. Ramakrishnan et al. [66] introduce NaQ, which is a data augmentation strategy to train models for long egocentric video search with natural language queries. In terms" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.422, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Learning Egocentric Action Frame Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.152, + 0.464, + 0.163 + ], + "angle": 0, + "content": "Example for In-context Learning" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.164, + 0.465, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.245, + 0.327, + 0.256 + ], + "angle": 0, + "content": "Input Query" + }, + { + "type": "image_footnote", + "bbox": [ + 0.263, + 0.257, + 0.486, + 0.316 + ], + "angle": 0, + "content": "Action Label: \"chop end of onion\" Bounding Boxes: \"left hand-[0.203, 0.368, 0.499, 1.000], right hand-[0.530, 0.527, 0.797, 1.000], chopping board-[0.248, 0.156, 0.644, 0.750], onion-[0.462, 0.377, 0.509, 0.506], knife-[0.484, 0.343, 0.546, 0.586]\"" + }, + { + "type": "image_footnote", + "bbox": [ + 0.261, + 0.321, + 0.487, + 0.362 + ], + "angle": 0, + "content": "Action Description: \"The person presses the onion on the chopping board with the left hand and then cuts off the end of the onion with a knife in the right hand.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.573, + 0.152, + 0.684, + 0.163 + ], + "angle": 0, + "content": "Query and Response" + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.164, + 0.718, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.245, + 0.583, + 0.256 + ], + "angle": 0, + "content": "Input Query" + }, + { + "type": "image_footnote", + "bbox": [ + 0.518, + 0.257, + 0.741, + 0.317 + ], + "angle": 0, + "content": "Action Label: \"squidge into lunch box\" Bounding Boxes: \"left hand-[0.427, 0.799, 0.591, 1.00], right hand-[0.654, 0.949, 0.707, 1.00], lunch box-[0.390, 0.512, 0.696, 0.993], sink-[0.00, 0.802, 0.315, 1.00], container lid-[0.403, 0.454, 0.513, 0.635], fork-[0.589, 0.798, 0.672, 1.00]\"" + }, + { + "type": "image_footnote", + "bbox": [ + 0.515, + 0.321, + 0.748, + 0.362 + ], + "angle": 0, + "content": "GPT-3.5 Response: \"The person uses their left hand to hold a lunch box, while their right hand uses a fork to squidge food into the lunch box, which is placed on the sink with the container lid nearby.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.38, + 0.789, + 0.45 + ], + "angle": 0, + "content": "Fig. 2: Examples for data curation using GPT-3.5. We provide detailed action descriptions of several example images as well as their action labels and bounding boxes for in-context learning. In addition, we input the action label and bounding boxes of another action as a query. GPT-3.5 is able to generate descriptions with enriched information (highlighted by underlines) in the response." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.789, + 0.587 + ], + "angle": 0, + "content": "of egocentric visual generation, Jia et al. [32] leverage GANs [19] to generate future head motion in hand forecasting task. Zhang et al. [96] leverage GANs to facilitate future gaze anticipation. Ye et al. [92] propose the affordance diffusion model that takes in the image of an object and generates possible hand-object interactions in the egocentric view. In this paper, we present the first work that investigates how to leverage VLLMs and diffusion models to generate action state transition on the egocentric image plane." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.61, + 0.33, + 0.626 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.796, + 0.719 + ], + "angle": 0, + "content": "The problem setting of egocentric action frame generation is illustrated in Fig. 1(c). Given an egocentric image frame \\(\\mathcal{X}\\) that captures the user's current visual context as well as a user query prompt \\(\\mathcal{P}\\) regarding how to perform an action, our goal is to synthesize the action frame \\(\\mathcal{Y}\\) that visually depicts how the action should be conducted in the same situation (i.e., keep a consistent background)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "The key insight of our proposed LEGO model is leveraging the strong capability of a VLLM to enhance the diffusion model for egocentric action frame generation. The annotations of existing egocentric datasets do not describe the details of how actions are conducted. As a remedy, we leverage visual instruction tuning to finetune a VLLM that enriches action descriptions based on the egocentric visual prompt. In addition, the existing diffusion-based image manipulation models are limited in understanding egocentric action state transition, due to the domain gap between the exocentric pre-training dataset and the ego-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "B. Lai et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.146, + 0.784, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.504 + ], + "angle": 0, + "content": "Fig. 3: Overview of our proposed LEGO model. We first finetune a visual large language model (VLLM) to generate the enriched action description with visual instruction tuning. We then project image and text embeddings from the finetuned VLLM to the feature space of the latent diffusion model (LDM). Finally, we train the LDM to synthesize the egocentric action frame conditioning on the input frame, enriched action description, as well as the VLLM image and text embeddings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.598 + ], + "angle": 0, + "content": "centric action dataset for our problem. To bridge this gap, we propose a novel approach that leverages VLLM embeddings to control the state transition of actions and to generate action frames accordingly. We detail the VLLM-based data enrichment pipeline and our model design in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.626, + 0.571, + 0.643 + ], + "angle": 0, + "content": "3.1 Egocentric Visual Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Data Curation for Visual Instruction Tuning. As shown in Fig. 2, we use GPT-3.5 to generate detailed action descriptions based on an input query composed of a short action label and object bounding boxes that are provided in the existing datasets. First, we prepare several examples of possible inputs along with their expected output descriptions for GPT-3.5 to perform in-context learning. These examples cover a diverse set of scenes in the egocentric action dataset. Each example is composed of an action label, a manually annotated action description, and relative spatial information of hands and objects-of-interests represented by bounding box coordinates. GPT-3.5 can learn from the given examples and generate similar detailed action descriptions based on a new input query. The resulting GPT-3.5 curated data is then further used for visual instruction tuning. More details of the prompt are provided in the supplementary." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.422, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Learning Egocentric Action Frame Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.79, + 0.298 + ], + "angle": 0, + "content": "Visual Instruction Tuning. We follow the finetuning strategy in prior work [52] as shown in Fig. 3(a). Specifically, we use the pre-trained CLIP visual encoder [63] \\(\\phi\\) to encode the visual representation and then apply a linear projection layer \\(\\tau\\) to map the CLIP visual features into the semantic space of the LLM, i.e., \\(\\mathcal{H}_i = \\tau (\\phi (\\mathcal{X}))\\). To construct the user prompt \\(\\mathcal{P}\\), we insert the action label annotation into a prompt template to create a coherent query that is aligned with the instruction-following nature of an LLM. We then tokenize \\(\\mathcal{P}\\), and feed both the prompt text tokens and image tokens \\(\\mathcal{H}_i\\) as inputs into the LLM. Finally, the LLM is trained to generate enriched action description (denoted as \\(\\mathcal{R}\\)) based on the user prompt and image input." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.346 + ], + "angle": 0, + "content": "User Prompt Enrichment at Scale. Note that the visual instruction tuned VLLM doesn't rely on any object bounding boxes as input. Therefore, we can generate enriched action descriptions for all egocentric action data at scale." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.365, + 0.566, + 0.38 + ], + "angle": 0, + "content": "3.2 Egocentric Action Frame Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.479 + ], + "angle": 0, + "content": "We leverage a latent diffusion model (LDM) [67] to synthesize the action frame conditioning on the input frame and the detailed action description \\(\\mathcal{R}\\) generated by our finetuned VLLM (see Fig. 3(b)). Following the regular steps in LDMs [5, 67], the input image is first encoded into a latent space using a pre-trained autoencoder \\(\\mathcal{E}\\). Then the input to the denoising UNet is a concatenation of the latent input representation \\(\\mathcal{E}(\\mathcal{X})\\) and a Gaussian noise \\(\\mathcal{G}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.585 + ], + "angle": 0, + "content": "Our key innovation is to design the U-Net conditioning component so that the diffusion model can interpret the egocentric actions correctly. To start, we follow [5] and adopt the conventional pre-trained CLIP text encoder \\(\\psi\\) to extract a text representation of \\(\\mathcal{R}\\), i.e., \\(\\psi(\\mathcal{R}) \\in \\mathbb{R}^{N \\times D}\\) where \\(N\\) is the maximum number of text tokens and \\(D\\) is the number of feature channels. We further leverage the image and text embeddings from the visual instruction tuned VLLM as additional LDM conditioning to alleviate the domain gap issue." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.788, + 0.766 + ], + "angle": 0, + "content": "Specifically, we feed the VLLM image embedding \\(\\mathcal{H}_i\\) into an extra linear layer \\(\\sigma\\) to map it to LDM feature space, i.e., \\(\\sigma(\\mathcal{H}_i) \\in \\mathbb{R}^{M \\times D}\\), where \\(M\\) is the number of image tokens. Note that \\(\\mathcal{H}_i\\) is already projected to the semantic space during visual instruction tuning, and therefore differs from the image embedding \\(\\mathcal{E}(\\mathcal{X})\\) from the auto-encoder. Moreover, we also extract the text embedding \\(\\mathcal{H}_t\\) before the last linear layer of LLM. We adopt a fixed token number \\(N\\) and enforce padding or truncation behavior, as in the CLIP text encoder. The text embedding is then fed to a projection layer \\(\\mu\\). In LLM decoder, the response is generated iteratively and each word embedding only conditions on the context ahead of it. To extract the holistic semantic meaning of \\(\\mathcal{H}_t\\) in LDM feature space, we further add self-attention layers \\(\\pi\\) after the projection, i.e., \\(\\pi(\\mu(\\mathcal{H}_t)) \\in \\mathbb{R}^{N \\times D}\\). Thus, the U-Net conditioning can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.777, + 0.787, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} = \\left[ \\psi (\\mathcal {R}), \\sigma \\left(\\mathcal {H} _ {i}\\right), \\pi \\left(\\mu \\left(\\mathcal {H} _ {t}\\right)\\right) \\right] \\in \\mathbb {R} ^ {(2 N + M) \\times D}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "The conditioning \\(\\mathcal{C}\\) is fed into the denoising UNet at multiple layers via the cross-attention mechanism [67]. We assume the intermediate feature of a specific" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "B. Lai et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "UNet layer is \\(\\mathcal{U}\\), which is learned from the UNet input (i.e., the concatenation of input frame representation \\(\\mathcal{E}(\\mathcal{X})\\) and Gaussian noise \\(\\mathcal{G}\\)). The cross-attention at this UNet layer can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.204, + 0.786, + 0.238 + ], + "angle": 0, + "content": "\\[\nC r o s s A t t (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {D}}\\right) \\cdot V, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.247, + 0.788, + 0.309 + ], + "angle": 0, + "content": "where \\(Q = W_{Q} \\cdot \\mathcal{U}\\), \\(K = W_{K} \\cdot \\mathcal{C}\\) and \\(V = W_{V} \\cdot \\mathcal{C}\\). Note that \\(W_{Q}, W_{K}\\) and \\(W_{V}\\) are learnable matrices. We also adopt the classifier-free guidance following [5] (see supplementary for details). Finally, the UNet output is converted to the image domain by a pre-trained decoder." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.331, + 0.458, + 0.346 + ], + "angle": 0, + "content": "3.3 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.789, + 0.522 + ], + "angle": 0, + "content": "All parameters of the VLLM are initialized from the pre-trained LLaVA [52] weights. During training, we freeze the CLIP image encoder and finetune the projection layer and LLM with cross-entropy loss for 3 epochs. To improve the diversity of the text prompts, we randomly select the question template from 10 candidates in each iteration. For LDM training, the text encoder, UNet and autoencoder are initialized with pre-trained weights [67]. The projection layers \\(\\sigma\\) and \\(\\mu\\) and the self-attention layers \\(\\pi\\) are initialized using the Xavier algorithm [17]. The text encoder is frozen and the remaining weights are finetuned with L2 regression loss between the predicted noise and real noise for 20,000 iterations. All input and target images are resized to a resolution of \\(256\\times 256\\). Please refer to the supplementary for more details of training and inference." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.546, + 0.377, + 0.563 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.578, + 0.41, + 0.592 + ], + "angle": 0, + "content": "4.1 Data and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.604, + 0.789, + 0.77 + ], + "angle": 0, + "content": "Datasets. We conduct our experiments on two well-established egocentric action datasets - Ego4D [21] and Epic-Kitchens-100 [13]. Both datasets were densely annotated with action starting time \\( t \\) and ending time \\( \\hat{t} \\). In our problem setting, we select an egocentric image frame \\( \\delta_{i} \\) seconds before the action begins as the input \\( \\mathcal{X} \\), and an image \\( \\delta_{o} \\) seconds after the action begins as the target frame \\( \\mathcal{Y} \\). On the Ego4D dataset, based on the annotations of Pre-Condition-15 time (PRE-15) \\( t_{pre} \\), and Point-of-No-Return time (PNR) \\( t_{pnr} \\), we set \\( \\delta_{i} = t - t_{pre} \\) and \\( \\delta_{o} = t_{pnr} - t \\). For Epic-Kitchens, PNR and PRE-15 annotations are not available. Instead, we empirically select \\( \\delta_{i} = 0.25 \\) seconds and \\( \\delta_{o} = t + \\lambda * (\\hat{t} - t) \\), where \\( \\lambda = 0.6 \\), for our experiments. More details of data preparation and improvement are elaborated in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.77, + 0.788, + 0.801 + ], + "angle": 0, + "content": "Metrics. We adopt image-to-image similarity, image-to-text similarity, and user study as metrics in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "- Image-to-Image Metrics. We implement six metrics to evaluate image-to-image similarity. To begin with, we adopt three contrastive learning based" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.422, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Learning Egocentric Action Frame Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.146, + 0.788, + 0.403 + ], + "angle": 0, + "content": "metrics including image-to-image (1) EgoVLP score [50], (2) EgoVLP\\(^+\\) score [50] and (3) CLIP score [63]. EgoVLP is a contrastive video-language pretraining model trained with egocentric videos. Since EgoVLP takes multiple frames as input, we consider two types of inputs: duplicating the output frame as a static input (i.e., EgoVLP score) and combining the input frame with the output frame (i.e., EgoVLP\\(^+\\) score). As a result, EgoVLP\\(^+\\) can effectively measure whether the generated frame can depict the state transition of an action. Importantly, given that EgoVLP is pre-trained on egocentric data and action labels, we consider EgoVLP and EgoVLP\\(^+\\) score as the primary automatic metrics. In addition, we also report (4) Fréchet Inception Distance (FID) [26], (5) Peak Signal-to-Noise Ratio (PSNR) and (6) Learned Perceptual Image Patch Similarity (LPIPS) [97] (with SqueezeNet [31] as the encoder) to make a thorough evaluation. Note that instead of measuring similarity of input and output frames as in prior works [12,80], in our problem setting, we measure the similarity between the generated action frame and ground truth, which better reflects whether the generation results can illustrate the execution of an action." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.406, + 0.788, + 0.543 + ], + "angle": 0, + "content": "- Image-to-Text Metrics. We find the widely-used image-to-text CLIP score can not align actions with egocentric images due to the domain gap. Similar misalignment problem is also observed in [14, 59, 72, 85]. In our experiments, we utilize BLIP [47] to generate image captions of the output images and then calculate text-to-text similarity using CLIP text encoder (following [34]). We implement this metric with two BLIP structures: BLIP-B and BLIP-L. Though this solution may still suffer from the same domain gap issue, it is a more appropriate evaluation metric in our problem setting. See more evidence and discussions in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.544, + 0.788, + 0.696 + ], + "angle": 0, + "content": "- **User Study.** We also conduct a user study on a subset of test data to further validate the advantage of our model based on human preference. We sample 60 examples from each dataset and present the generated frames from our model as well as the baseline models to raters on Amazon Mechanical Turk (AMT). We also provide the input frames and the corresponding action labels during evaluation. For each instance, we ask the raters to select the image that best aligns with the provided action label while preserving the most contextual information from the input frame. To minimize potential bias, we hire 5 AMT raters to annotate each example thus resulting in 300 samples for each dataset. User study interface is shown in supplementary." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.735, + 0.555, + 0.751 + ], + "angle": 0, + "content": "4.2 Comparison with Prior Approaches" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We compare our proposed model with previous state-of-the-art methods for text-guided image manipulation, including ProxEdit [24], SDEdit [56] and InstructPix2Pix (IP2P) [5]. For a fair comparison, we finetune these baseline methods with the same data used in our experiments. Specifically, we train IP2P in an end-to-end way on the two datasets with existing egocentric action labels as" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "B. Lai et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 1: Comparison with prior image manipulation approaches in image-to-image metrics. \\(\\downarrow\\) means a lower score in this metric suggests a better performance. The best results are highlighted with **boldface**. The orange row refers to our LEGO model." + }, + { + "type": "table", + "bbox": [ + 0.234, + 0.199, + 0.768, + 0.349 + ], + "angle": 0, + "content": "
MethodsEgoVLPEgoVLP+CLIPFID ↓PSNRLPIPS ↓
Ego4DProxEdit [24]44.5172.6868.1733.0111.8840.90
SDEdit [56]50.0772.9073.3533.3511.8141.60
IP2P [5]62.1978.8478.7524.7312.1637.16
LEGO65.6580.4480.6123.8312.2936.43
EK-100ProxEdit [24]32.2752.7765.8051.3511.0646.35
SDEdit [56]33.8456.8074.7627.4111.3043.33
IP2P [5]42.9761.0677.0320.6411.2340.82
LEGO45.8962.6678.6321.5711.3340.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.23, + 0.374, + 0.517, + 0.43 + ], + "angle": 0, + "content": "Table 2: Image-to-text metrics of our model and baselines. The best results are highlighted with boldface. The orange row refers to our LEGO model performance." + }, + { + "type": "table", + "bbox": [ + 0.234, + 0.441, + 0.512, + 0.525 + ], + "angle": 0, + "content": "
MethodsEgo4DEpic-Kitchens
BLIP-BBLIP-LBLIP-BBLIP-L
ProxEdit [24]17.7317.3523.6523.39
SDEdit [56]19.8019.7421.5121.30
IP2P [5]20.0020.5625.3726.36
LEGO20.3820.7026.9827.41
" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.363, + 0.767, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.493, + 0.774, + 0.536 + ], + "angle": 0, + "content": "Fig.4: User study of our model and baselines. Win rate is the percentage of each model being picked as the best." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.673 + ], + "angle": 0, + "content": "text conditioning. ProxEdit and SDEdit rely on the off-the-shelf latent diffusion model to synthesize edited images and thus can not be trained end to end. Therefore, we first finetune the latent diffusion model with egocentric images and action labels and then use the finetuned latent diffusion model parameters to implement ProxEdit and SDEdit approaches. Please refer to the supplementary for more baseline implementation details and a thorough comparison with prior methods that leverage LLMs for image generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.809 + ], + "angle": 0, + "content": "In terms of image-to-image metrics in Tab. 1, both ProxEdit and SDEdit perform poorly on this novel problem, suggesting the challenging nature of generating egocentric action frames. IP2P performs much better in almost all metrics by end-to-end training and serves as the strongest baseline model in our experiments. However, our LEGO model consistently outperforms IP2P by a large margin (3.46%, 1.60%, 1.86%, 0.90, 0.13 and 0.73% respectively) in all six metrics on Ego4D. LEGO also exceeds IP2P notably (2.92%, 1.60%, 1.60%, 0.10 and 0.46% respectively) in five metrics on Epic-Kitchens. Though, LEGO slightly lags behind IP2P in FID score, it still achieves the second best performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "With regard to image-to-text metrics in Tab. 2, LEGO outperforms the strongest baseline (IP2P) by \\(0.38\\%\\) and \\(0.14\\%\\) on Ego4D and by \\(1.61\\%\\) and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.423, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Learning Egocentric Action Frame Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.144, + 0.764, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.568, + 0.788, + 0.611 + ], + "angle": 0, + "content": "Fig. 5: Visualization of the proposed LEGO model and all baselines on Ego4D (the first three rows) and Epic-Kitchens (the last three rows). The action frames generated by LEGO align with the user prompt better and preserve more contexts in input frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.779 + ], + "angle": 0, + "content": "\\(1.05\\%\\) on Epic-Kitchens. The result suggests the synthetic frames from our model can better align with the action descriptions. However, the performance gain is rather limited. We emphasize that this is because the domain gap still exists for the BLIP model. Besides baseline models, we also measure the image-to-text similarity of input frames and action prompts to validate models' capability of learning action state transition in semantics. The BLIP-B/BLIP-L scores are \\(15.44\\% / 15.49\\%\\) on Ego4D and \\(20.11\\% / 20.52\\%\\) on Epic-Kitchens, lagging behind all baseline models. The result evidences diffusion models are able to understand actions and edit the input frame semantically towards the action prompt." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Due to the limitation of automatic metrics, we additionally implement user study as a complement. We shuffle the order of the results from our model and the baselines while presenting them to the raters. We define the win rate as the percentage of each model being picked as the best out of the total 300" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "B. Lai et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.26, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.175, + 0.256, + 0.195 + ], + "angle": 0, + "content": "Input Frame" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.204, + 0.254, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.235, + 0.256, + 0.245 + ], + "angle": 0, + "content": "LEGO" + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.146, + 0.439, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.25, + 0.436, + 0.272 + ], + "angle": 0, + "content": "cut a portion of clay brush a wood mix with both hands with a brush" + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.146, + 0.554, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.25, + 0.549, + 0.262 + ], + "angle": 0, + "content": "\"put tray in oven\"" + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.147, + 0.668, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.25, + 0.664, + 0.262 + ], + "angle": 0, + "content": "\"close container\"" + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.147, + 0.784, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.197, + 0.784, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.25, + 0.772, + 0.262 + ], + "angle": 0, + "content": "\"take soy milk\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.283, + 0.787, + 0.312 + ], + "angle": 0, + "content": "Fig. 6: Additional visualization of our proposed model. LEGO is able to synthesize faithful action frames as well as preserve the contexts in various scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.785, + 0.387 + ], + "angle": 0, + "content": "samples. Results are illustrated in Fig. 4. Our model surpasses the best baseline by \\(44.00\\%\\) and \\(38.34\\%\\) respectively on Ego4D and Epic-Kitchens. The prominent gap further validates the superiority of our model." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.408, + 0.547, + 0.423 + ], + "angle": 0, + "content": "4.3 Visualization of Generated Frames" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.785, + 0.553 + ], + "angle": 0, + "content": "We additionally showcase examples of generated image from LEGO and baseline models in Fig. 5. ProxEdit and SDEdit fail to understand the user prompt and thus generate frames of irrelevant actions (e.g., row3). They may also easily synthesize the action in a different environment (e.g., row2). InstructPix2Pix is able to preserve more contexts but still fails to semantically align with the action in user prompt (e.g., row1). In contrast, LEGO can synthesize action frames that better align with the user prompts and retain more contexts in the background. More examples of LEGO output are presented in Fig. 6 and supplementary." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.575, + 0.389, + 0.59 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We present comprehensive ablation studies to investigate the contribution of enriched action descriptions, the VLLM image embedding and the VLLM text embedding separately. Results are demonstrated in Tab. 3. Given the limitation of automatic metrics, we also provide user study as the primary metric. Notably, conditioning on enriched action descriptions can moderately improve the model performance, supporting our claim that expanding the action description can facilitate the learning of the state transition during an egocentric action. Moreover, utilizing the image embedding and text embedding from VLLM as additional conditions both improve the model performance by a notable margin because VLLM embeddings can effectively narrow the domain gap. Interestingly, the image embedding leads to larger performance boost on both datasets. These results suggest the VLLM image embedding \\(\\mathcal{H}_i\\) incorporates important high-level semantic meanings that are not captured in the auto-encoder image embedding \\(\\mathcal{E}(\\mathcal{X})\\) or VLLM text emebdding \\(\\mathcal{H}_t\\). Finally, our full LEGO model uses both VLLM image and text embeddings (Desc.+Joint Embed.) and thus achieves the best performance on both datasets in all metrics." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.422, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Learning Egocentric Action Frame Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 3: Analysis of egocentric action frame generation performance with different conditionings. Joint Embed. refers to incorporating both VLLM image and text embeddings. Similar to Fig. 4, we present win rate as the user study result, i.e., the percentage of each model picked as the best (\\% is omitted for simplicity). The best results are highlighted with **boldface**. The orange rows refer to our full LEGO model." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.227, + 0.794, + 0.322 + ], + "angle": 0, + "content": "
ConditioningEgo4DEpic-Kitchens
User StudyEgoVLPEgoVLP+CLIPUser StudyEgoVLPEgoVLP+CLIP
Actions Labels5.3362.1978.8478.757.0842.9761.0677.03
Descriptions13.0062.9179.0979.1812.5043.7261.4677.47
Desc. + Img Embed.26.0065.3580.1380.5724.1745.8262.2978.60
Desc. + Txt Embed.21.3363.2979.4079.2122.0844.6862.0277.74
Desc. + Joint Embed.34.3465.6580.4480.6134.1745.8962.6678.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.338, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Table 4: Performance of LEGO without finetuning (w/o FT). The best results are highlighted with **boldface**. The orange rows refer to our full LEGO model." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.379, + 0.792, + 0.472 + ], + "angle": 0, + "content": "
ConditioningEgo4DEpic-Kitchens
EgoVLPEgoVLP+CLIPEgoVLPEgoVLP+CLIP
Descriptions62.9179.0979.1843.7261.4677.47
Desc.+Joint Embed.(w/o FT)64.5779.7280.2344.7461.8778.35
Desc.+Joint Embed.(w/ FT)65.6580.4480.6145.8962.6678.63
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.504, + 0.573, + 0.521 + ], + "angle": 0, + "content": "4.5 Analysis of Visual Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.718 + ], + "angle": 0, + "content": "First, we assess the quality of enriched action descriptions from the visual instruction tuned VLLM by user study. For each sample in user study, we ask the raters to select whether the description aligns with the input and target frames (see supplementary for more details of user study setting and interface). The percentage of samples with aligned frames and action descriptions is \\(87\\%\\) on Ego4D and \\(84\\%\\) on Epic-Kitchens. The high alignment percentage suggests the visual instruction tuned VLLM can effectively expand action labels with details captured in the input frame. We additionally implement the same user study for off-the-shelf VLLM (i.e., without finetuning). The percentage of alignment is just \\(27\\%\\) on Ego4D and \\(30\\%\\) on Epic-Kitchens with hallucination existing in \\(92\\%\\) of unaligned instances. The notable drop supports our argument that visual instruction tuning is a critical step for high-quality prompt enrichment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In addition, we further investigate whether the visual instruction tuned embeddings can contribute more to the diffusion model than the off-the-self counterpart, which has not been well studied in prior work. As shown in Tab. 4, without any finetuning, the image and text embeddings from VLLM can still improve the baseline model (Descriptions). However finetuned embeddings yield much larger improvement (e.g., \\(1.08\\%\\) and \\(1.15\\%\\) in EgoVLP score) on both datasets. The result suggests that visual instruction tuning is a necessary step to learn semantic image and text embeddings that are more aligned with the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "B. Lai et al." + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.164, + 0.38, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.246, + 0.256, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.218, + 0.377, + 0.268 + ], + "angle": 0, + "content": "\"Can you provide instructions on how to {action} in my current situation?\"" + }, + { + "type": "image", + "bbox": [ + 0.397, + 0.149, + 0.512, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.202, + 0.5, + 0.213 + ], + "angle": 0, + "content": "\"open drawer\"" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.15, + 0.629, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.202, + 0.607, + 0.213 + ], + "angle": 0, + "content": "\"dry hands\"" + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.15, + 0.746, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.202, + 0.733, + 0.213 + ], + "angle": 0, + "content": "\"cut cucumber\"" + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.215, + 0.512, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.401, + 0.268, + 0.508, + 0.279 + ], + "angle": 0, + "content": "\"open microwave\"" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.215, + 0.63, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.268, + 0.613, + 0.28 + ], + "angle": 0, + "content": "\"pick up bowl\"" + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.215, + 0.746, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.268, + 0.723, + 0.28 + ], + "angle": 0, + "content": "\"take knife\"" + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.24, + 0.78, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.752, + 0.268, + 0.786, + 0.28 + ], + "angle": 0, + "content": "LEGO" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.298, + 0.787, + 0.354 + ], + "angle": 0, + "content": "Fig. 7: Visualization of generating various actions from the same input frame. The first action (\"open drawer\") is the existing label for this example in the dataset. We fill another five actions into the user query and synthesize the action frames using our model. All generated frames align well with the actions and preserve most contexts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.785, + 0.416 + ], + "angle": 0, + "content": "egocentric input frame and action prompt, which thus narrows the domain gap and greatly boosts the performance of the latent diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.443, + 0.718, + 0.457 + ], + "angle": 0, + "content": "4.6 Generation of Various Actions with the Same Contexts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.594 + ], + "angle": 0, + "content": "In addition to generating action frames based on the pre-defined action labels in our datasets, we validate the generative nature of our model i.e., whether LEGO is able to synthesize the correct action frames conditioning on the same contexts yet different actions (novel image-action pairs). Results are illustrated in Fig. 7. We feed different actions with the same input frame to our model. The synthesized frames correctly depict the execution of the actions in the same environment. This result further indicates that our model can understand the action state transition and generalize to different user queries." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.623, + 0.36, + 0.638 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we introduce the novel problem of egocentric action frame generation. We also propose a novel model — LEGO that leverages visual instruction tuning and a diffusion model to address this problem. Our key intuition is to use visual instruction tuning to generate informative responses that depict the execution of the egocentric action, and then design the conditioning for the denoising U-Net to exploit the image and text embeddings from a visual instruction tuned VLLM. Our experiments on two large-scale egocentric action datasets validate the advantage of the proposed approach as well as the contribution of each model component. We believe our work provides an important step in understanding the action state transition and controllability of diffusion models, and suggests future research directions in egocentric AI systems, action state transition, image generation and human skill transfer." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.422, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Learning Egocentric Action Frame Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.786, + 0.205 + ], + "angle": 0, + "content": "Portions of this work were supported in part by a gift from Meta. We thank Sangmin Lee for the valuable discussion and suggestions." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.225, + 0.324, + 0.241 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.254, + 0.785, + 0.308 + ], + "angle": 0, + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.31, + 0.785, + 0.336 + ], + "angle": 0, + "content": "2. Anonymity: Making multimodal generation easier: When diffusion models meet llms. openreview (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.337, + 0.785, + 0.377 + ], + "angle": 0, + "content": "3. Ashutosh, K., Girdhar, R., Torresani, L., Grauman, K.: Hiervl: Learning hierarchical video-language embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23066-23078 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.378, + 0.785, + 0.417 + ], + "angle": 0, + "content": "4. Baskin, J.H., Edersheim, J.G., Price, B.H.: Is a picture worth a thousand words? neuroimaging in the courtroom. American Journal of Law & Medicine 33(2-3), 239-269 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.418, + 0.785, + 0.459 + ], + "angle": 0, + "content": "5. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.459, + 0.785, + 0.514 + ], + "angle": 0, + "content": "6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.515, + 0.785, + 0.554 + ], + "angle": 0, + "content": "7. Chakrabarty, T., Singh, K., Saakyan, A., Muresan, S.: Learning to follow object-centric image editing instructions faithfully. arXiv preprint arXiv:2310.19145 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.555, + 0.785, + 0.595 + ], + "angle": 0, + "content": "8. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., et al.: Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.596, + 0.785, + 0.636 + ], + "angle": 0, + "content": "9. Chen, W.G., Spiridonova, I., Yang, J., Gao, J., Li, C.: Llava-interactive: An all-in-one demo for image chat, segmentation, generation and editing. arXiv preprint arXiv:2311.00571 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.637, + 0.785, + 0.677 + ], + "angle": 0, + "content": "10. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.678, + 0.785, + 0.704 + ], + "angle": 0, + "content": "11. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.705, + 0.785, + 0.744 + ], + "angle": 0, + "content": "12. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. Advances in Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.745, + 0.785, + 0.799 + ], + "angle": 0, + "content": "13. Damen, D., Doughty, H., Farinella, G.M., Furnari, A., Kazakos, E., Ma, J., Moltisanti, D., Munro, J., Perrett, T., Price, W., et al.: Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision pp. 1-23 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "14. Du, Y., Yang, S., Dai, B., Dai, H., Nachum, O., Tenenbaum, J.B., Schuurmans, D., Abbeel, P.: Learning universal policies via text-guided video generation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.254, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.354, + 0.127 + ], + "angle": 0, + "content": "B. Lai et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "15. Epstein, D., Jabri, A., Poole, B., Efros, A.A., Holynski, A.: Diffusion self-guidance for controllable image generation. In: Advances in Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.191, + 0.785, + 0.218 + ], + "angle": 0, + "content": "16. Girdhar, R., Grauman, K.: Anticipative video transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13505-13515 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.22, + 0.785, + 0.274 + ], + "angle": 0, + "content": "17. Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedforward neural networks. In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. pp. 249-256. JMLR Workshop and Conference Proceedings (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.276, + 0.785, + 0.317 + ], + "angle": 0, + "content": "18. Goel, V., Peruzzo, E., Jiang, Y., Xu, D., Sebe, N., Darrell, T., Wang, Z., Shi, H.: Pair-diffusion: Object-level image editing with structure-and-appearance paired diffusion models. arXiv preprint arXiv:2303.17546 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.359 + ], + "angle": 0, + "content": "19. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139–144 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.361, + 0.785, + 0.402 + ], + "angle": 0, + "content": "20. Goyal, M., Modi, S., Goyal, R., Gupta, S.: Human hands as probes for interactive object understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3293-3303 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.403, + 0.785, + 0.459 + ], + "angle": 0, + "content": "21. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18995-19012 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "22. Hafri, A., Trueswell, J.C., Epstein, R.A.: Neural representations of observed actions generalize across static and dynamic visual input. Journal of Neuroscience 37(11), 3056-3071 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.502, + 0.785, + 0.543 + ], + "angle": 0, + "content": "23. Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., et al.: Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.545, + 0.785, + 0.599 + ], + "angle": 0, + "content": "24. Han, L., Wen, S., Chen, Q., Zhang, Z., Song, K., Ren, M., Gao, R., Stathopoulos, A., He, X., Chen, Y., et al.: Proxedit: Improving tuning-free real image editing with proximal guidance. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4291-4301 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.601, + 0.785, + 0.641 + ], + "angle": 0, + "content": "25. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.643, + 0.785, + 0.684 + ], + "angle": 0, + "content": "26. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.686, + 0.785, + 0.712 + ], + "angle": 0, + "content": "27. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.714, + 0.785, + 0.755 + ], + "angle": 0, + "content": "28. Huang, J., Liu, Y., Qin, J., Chen, S.: Kv inversion: Kv embeddings learning for text-conditioned real image action editing. arXiv preprint arXiv:2309.16608 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "29. Huang, Y., Cai, M., Li, Z., Lu, F., Sato, Y.: Mutual context network for jointly estimating egocentric gaze and action. IEEE Transactions on Image Processing 29, 7795-7806 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "30. Huang, Y., Cai, M., Li, Z., Sato, Y.: Predicting gaze in egocentric video by learning task-dependent attention transition. In: Proceedings of the European conference on computer vision (ECCV). pp. 754-769 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.422, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Learning Egocentric Action Frame Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "31. Iandola, F.N., Han, S., Moskewicz, M.W., Ashraf, K., Dally, W.J., Keutzer, K.: SqueezeNet: Alexnet-level accuracy with 50x fewer parameters and \\(< 0.5\\) mb model size. arXiv preprint arXiv:1602.07360 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.191, + 0.785, + 0.233 + ], + "angle": 0, + "content": "32. Jia, W., Liu, M., Rehg, J.M.: Generative adversarial network for future hand segmentation from egocentric video. In: European Conference on Computer Vision. pp. 639-656. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.234, + 0.785, + 0.26 + ], + "angle": 0, + "content": "33. Jiang, Y., Zhang, Z., Xue, T., Gu, J.: Autodir: Automatic all-in-one image restoration with latent diffusion. arXiv preprint arXiv:2310.10123 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.262, + 0.785, + 0.316 + ], + "angle": 0, + "content": "34. Joseph, K., Udhayanan, P., Shukla, T., Agarwal, A., Karanam, S., Goswami, K., Srinivasan, B.V.: Iterative multi-granular image editing using diffusion models. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.345 + ], + "angle": 0, + "content": "35. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. In: Advances in Neural Information Processing Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.347, + 0.785, + 0.401 + ], + "angle": 0, + "content": "36. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.403, + 0.785, + 0.444 + ], + "angle": 0, + "content": "37. Kazakos, E., Nagrani, A., Zisserman, A., Damen, D.: Epic-fusion: Audio-visual temporal binding for egocentric action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5492-5501 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.785, + 0.486 + ], + "angle": 0, + "content": "38. Kim, S., Jang, W., Kim, H., Kim, J., Choi, Y., Kim, S., Lee, G.: User-friendly image editing with minimal text input: Leveraging captioning and injection techniques. arXiv preprint arXiv:2306.02717 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.488, + 0.785, + 0.528 + ], + "angle": 0, + "content": "39. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.531, + 0.785, + 0.557 + ], + "angle": 0, + "content": "40. Koh, J.Y., Fried, D., Salakhutdinov, R.R.: Generating images with multimodal language models. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.559, + 0.785, + 0.599 + ], + "angle": 0, + "content": "41. Lai, B., Liu, M., Ryan, F., Rehg, J.M.: In the eye of transformer: Global-local correlation for egocentric gaze estimation. British Machine Vision Conference (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.601, + 0.785, + 0.642 + ], + "angle": 0, + "content": "42. Lai, B., Liu, M., Ryan, F., Rehg, J.M.: In the eye of transformer: Global-local correlation for egocentric gaze estimation and beyond. International Journal of Computer Vision 132(3), 854-871 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.644, + 0.785, + 0.671 + ], + "angle": 0, + "content": "43. Lai, B., Ryan, F., Jia, W., Liu, M., Rehg, J.M.: Listen to look into the future: Audio-visual egocentric gaze anticipation. arXiv preprint arXiv:2305.03907 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.785, + 0.713 + ], + "angle": 0, + "content": "44. Li, D., Li, J., Hoi, S.C.: Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.785, + 0.755 + ], + "angle": 0, + "content": "45. Li, J., Liu, K., Wu, J.: Ego-body pose estimation via ego-head pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17142-17151 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "46. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. International Conference on Machine Learning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "47. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "B. Lai et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "48. Li, Y., Liu, M., Rehg, J.M.: In the eye of beholder: Joint learning of gaze and actions in first person video. In: Proceedings of the European conference on computer vision (ECCV). pp. 619-635 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "49. Lian, L., Li, B., Yala, A., Darrell, T.: Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.23, + 0.785, + 0.271 + ], + "angle": 0, + "content": "50. Lin, K.Q., Wang, J., Soldan, M., Wray, M., Yan, R., XU, E.Z., Gao, D., Tu, R.C., Zhao, W., Kong, W., et al.: Egocentric video-language pretraining. Advances in Neural Information Processing Systems 35, 7575-7586 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.271, + 0.785, + 0.325 + ], + "angle": 0, + "content": "51. Liu, B., Zhang, H., Liu, J., Wang, Q.: Acigs: An automated large-scale crops image generation system based on large visual language multi-modal models. In: 2023 20th Annual IEEE International Conference on Sensing, Communication, and Networking (SECON). pp. 7-13. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.325, + 0.785, + 0.352 + ], + "angle": 0, + "content": "52. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.352, + 0.785, + 0.407 + ], + "angle": 0, + "content": "53. Liu, M., Tang, S., Li, Y., Rehg, J.M.: Forecasting human-object interaction: joint prediction of motor attention and actions in first person video. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 704-721. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.407, + 0.785, + 0.447 + ], + "angle": 0, + "content": "54. Liu, S., Tripathi, S., Majumdar, S., Wang, X.: Joint hand motion and interaction hotspots prediction from egocentric videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3282-3292 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.447, + 0.785, + 0.488 + ], + "angle": 0, + "content": "55. Luo, Z., Hachiuma, R., Yuan, Y., Kitani, K.: Dynamics-regulated kinematic policy for egocentric pose estimation. Advances in Neural Information Processing Systems 34, 25019-25032 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.488, + 0.785, + 0.528 + ], + "angle": 0, + "content": "56. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.785, + 0.569 + ], + "angle": 0, + "content": "57. Mirzaei, A., Aumentado-Armstrong, T., Brubaker, M.A., Kelly, J., Levinshtein, A., Derpanis, K.G., Gilitschenski, I.: Watch your steps: Local image and scene editing by text instructions. arXiv preprint arXiv:2308.08947 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.569, + 0.785, + 0.623 + ], + "angle": 0, + "content": "58. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6038-6047 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.785, + 0.664 + ], + "angle": 0, + "content": "59. Molad, E., Horwitz, E., Valevski, D., Acha, A.R., Matias, Y., Pritch, Y., Leviathan, Y., Hoshen, Y.: Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.664, + 0.785, + 0.691 + ], + "angle": 0, + "content": "60. Orgad, H., Kawar, B., Belinkov, Y.: Editing implicit assumptions in text-to-image diffusion models. arXiv preprint arXiv:2303.08084 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.785, + 0.732 + ], + "angle": 0, + "content": "61. Pan, Z., Gherardi, R., Xie, X., Huang, S.: Effective real image editing with accelerated iterative diffusion inversion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15912-15921 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.785, + 0.786 + ], + "angle": 0, + "content": "62. Pramanick, S., Song, Y., Nag, S., Lin, K.Q., Shah, H., Shou, M.Z., Chellappa, R., Zhang, P.: Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5285-5297 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "63. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.422, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Learning Egocentric Action Frame Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.203 + ], + "angle": 0, + "content": "64. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.203, + 0.785, + 0.243 + ], + "angle": 0, + "content": "65. Ragusa, F., Farinella, G.M., Furnari, A.: Stillfast: An end-to-end approach for short-term object interaction anticipation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3635-3644 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.243, + 0.785, + 0.285 + ], + "angle": 0, + "content": "66. Ramakrishnan, S.K., Al-Halah, Z., Grauman, K.: Naq: Leveraging narrations as queries to supervise episodic memory. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6694-6703 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.285, + 0.785, + 0.325 + ], + "angle": 0, + "content": "67. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.325, + 0.785, + 0.379 + ], + "angle": 0, + "content": "68. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.785, + 0.42 + ], + "angle": 0, + "content": "69. Ryan, F., Jiang, H., Shukla, A., Rehg, J.M., Ithapu, V.K.: Egocentric auditory attention localization in conversations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14663-14674 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.785, + 0.461 + ], + "angle": 0, + "content": "70. Shen, Y., Song, K., Tan, X., Li, D., Lu, W., Zhuang, Y.: Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.785, + 0.488 + ], + "angle": 0, + "content": "71. Shi, J., Xiong, W., Lin, Z., Jung, H.J.: Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.488, + 0.785, + 0.542 + ], + "angle": 0, + "content": "72. Stein, G., Cresswell, J.C., Hosseinzadeh, R., Sui, Y., Ross, B.L., Villecloze, V., Liu, Z., Caterini, A.L., Taylor, J.E.T., Loaiza-Ganem, G.: Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. arXiv preprint arXiv:2306.04675 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.785, + 0.569 + ], + "angle": 0, + "content": "73. Su, Y., Lan, T., Li, H., Xu, J., Wang, Y., Cai, D.: Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.569, + 0.785, + 0.609 + ], + "angle": 0, + "content": "74. Sudhakaran, S., Escalera, S., Lanz, O.: Lsta: Long short-term attention for egocentric action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9954-9963 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.609, + 0.785, + 0.651 + ], + "angle": 0, + "content": "75. Sun, Z., Zhou, Y., He, H., Mok, P.: Sgdiff: A style guided diffusion model for fashion synthesis. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 8433-8442 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.651, + 0.785, + 0.692 + ], + "angle": 0, + "content": "76. Thoppilan, R., De Freitas, D., Hall, J., Shazeer, N., Kulshreshtha, A., Cheng, H.T., Jin, A., Bos, T., Baker, L., Du, Y., et al.: Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.692, + 0.785, + 0.732 + ], + "angle": 0, + "content": "77. Tome, D., Alldieck, T., Peluse, P., Pons-Moll, G., Agapito, L., Badino, H., De la Torre, F.: Selfpose: 3d egocentric pose estimation from a headset mounted camera. IEEE Transactions on Pattern Analysis and Machine Intelligence (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.785, + 0.772 + ], + "angle": 0, + "content": "78. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "79. Tsaban, L., Passos, A.: Ledits: Real image editing with ddpm inversion and semantic guidance. arXiv preprint arXiv:2307.00522 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "80. Wallace, B., Gokul, A., Naik, N.: Edict: Exact diffusion inversion via coupled transformations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22532-22541 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.354, + 0.127 + ], + "angle": 0, + "content": "B. Lai et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.147, + 0.787, + 0.19 + ], + "angle": 0, + "content": "81. Wang, J., Luvizon, D., Xu, W., Liu, L., Sarkar, K., Theobalt, C.: Scene-aware egocentric 3d human pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13031-13040 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.19, + 0.786, + 0.233 + ], + "angle": 0, + "content": "82. Wang, K., Yang, F., Yang, S., Butt, M.A., van de Weijer, J.: Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. arXiv preprint arXiv:2309.15664 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.787, + 0.274 + ], + "angle": 0, + "content": "83. Wang, Q., Zhang, B., Birsak, M., Wonka, P.: Instructedit: Improving automatic masks for diffusion-based image editing with user instructions. arXiv preprint arXiv:2305.18047 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.276, + 0.787, + 0.317 + ], + "angle": 0, + "content": "84. Wang, Q., Zhang, B., Birsak, M., Wonka, P.: Mdp: A generalized framework for text-guided image editing by manipulating the diffusion path. arXiv preprint arXiv:2303.16765 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.787, + 0.359 + ], + "angle": 0, + "content": "85. Wang, W., Xie, K., Liu, Z., Chen, H., Cao, Y., Wang, X., Shen, C.: Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.36, + 0.787, + 0.402 + ], + "angle": 0, + "content": "86. Wang, X., Zhu, L., Wang, H., Yang, Y.: Interactive prototype learning for egocentric action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8168-8177 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.402, + 0.787, + 0.445 + ], + "angle": 0, + "content": "87. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.445, + 0.787, + 0.487 + ], + "angle": 0, + "content": "88. Wen, S., Fang, G., Zhang, R., Gao, P., Dong, H., Metaxas, D.: Improving compositional text-to-image generation with large vision-language models. arXiv preprint arXiv:2310.06311 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.488, + 0.787, + 0.529 + ], + "angle": 0, + "content": "89. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.53, + 0.787, + 0.558 + ], + "angle": 0, + "content": "90. Wu, S., Fei, H., Qu, L., Ji, W., Chua, T.S.: Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.559, + 0.787, + 0.613 + ], + "angle": 0, + "content": "91. Xu, Y., Li, Y.L., Huang, Z., Liu, M.X., Lu, C., Tai, Y.W., Tang, C.K.: Egopca: A new framework for egocentric hand-object interaction understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5273-5284 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.615, + 0.787, + 0.671 + ], + "angle": 0, + "content": "92. Ye, Y., Li, X., Gupta, A., De Mello, S., Birchfield, S., Song, J., Tulsiani, S., Liu, S.: Affordance diffusion: Synthesizing hand-object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22479-22489 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.671, + 0.787, + 0.713 + ], + "angle": 0, + "content": "93. Yu, Q., Li, J., Ye, W., Tang, S., Zhuang, Y.: Interactive data synthesis for systematic vision adaptation via llms-aigcs collaboration. arXiv preprint arXiv:2305.12799 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.714, + 0.787, + 0.755 + ], + "angle": 0, + "content": "94. Yu, Z., Li, H., Fu, F., Miao, X., Cui, B.: Fisted: Accelerating text-to-image editing via cache-enabled sparse diffusion inference. arXiv preprint arXiv:2305.17423 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.787, + 0.785 + ], + "angle": 0, + "content": "95. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "96. Zhang, M., Teck Ma, K., Hwee Lim, J., Zhao, Q., Feng, J.: Deep future gaze: Gaze anticipation on egocentric videos using adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4372-4381 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.422, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Learning Egocentric Action Frame Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "21" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "97. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.19, + 0.787, + 0.231 + ], + "angle": 0, + "content": "98. Zhang, S., Yang, X., Feng, Y., Qin, C., Chen, C.C., Yu, N., Chen, Z., Wang, H., Savarese, S., Ermon, S., et al.: Hive: Harnessing human feedback for instructional visual editing. arXiv preprint arXiv:2303.09618 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "99. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "100. Zhang, Y., Huang, N., Tang, F., Huang, H., Ma, C., Dong, W., Xu, C.: Inversion-based style transfer with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10146-10156 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.315, + 0.785, + 0.357 + ], + "angle": 0, + "content": "101. Zhang, Z., Han, L., Ghosh, A., Metaxas, D.N., Ren, J.: Sine: Single image editing with text-to-image diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6027-6037 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.357, + 0.785, + 0.398 + ], + "angle": 0, + "content": "102. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.788, + 0.398 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/LEGO_ Learning EGOcentric Action Frame Generation via Visual Instruction Tuning/13a5a5c4-078d-4d85-be9d-7966d6d6437e_origin.pdf b/2024/LEGO_ Learning EGOcentric Action Frame Generation via Visual Instruction Tuning/13a5a5c4-078d-4d85-be9d-7966d6d6437e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4acbf0394892ac052c8047a8aac958507e1ab8d1 --- /dev/null +++ b/2024/LEGO_ Learning EGOcentric Action Frame Generation via Visual Instruction Tuning/13a5a5c4-078d-4d85-be9d-7966d6d6437e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2fdbb1581cb71604f46c0ad1b30f45df6673e6311f51f6c4deaa6efefedd76c3 +size 2486162 diff --git a/2024/LEGO_ Learning EGOcentric Action Frame Generation via Visual Instruction Tuning/full.md b/2024/LEGO_ Learning EGOcentric Action Frame Generation via Visual Instruction Tuning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b9b9334607f20977ec2657c2b8c70c4f6cadf7c4 --- /dev/null +++ b/2024/LEGO_ Learning EGOcentric Action Frame Generation via Visual Instruction Tuning/full.md @@ -0,0 +1,396 @@ +# LEGO: Learning EGOcentric Action Frame Generation via Visual Instruction Tuning + +Bolin Lai $^{1,2,\dagger}$ , Xiaoliang Dai $^{1}$ , Lawrence Chen $^{1}$ , Guan Pang $^{1}$ , James M. Rehg $^{3}$ , and Miao Liu $^{1,*}$ + +1 GenAI, Meta + +2 Georgia Institute of Technology + +3 University of Illinois Urbana-Champaign + +bolin.lai@gatech.edu {xiaoliangdai,lawrencechen,gpang,miaoliu}@meta.com jrehg@illinois.edu + +![](images/bbb63d72c539697167526da88405b46948a15f31e23638eca417d8266b74387d.jpg) + +# User Prompt + +"Can you provide instructions on how to wash the trouser with the brush in my current situation?" + +![](images/6a7072f59eb18a44aaafacb5ed0a0cc5fd2b54ea97b19cf1a992fa6095a0ef5f.jpg) + +1. Check the Care Label: Check the care label on your trousers for any specific washing instructions. + +2. Prep the Area: Find a clean and flat surface to work on, such as a table or countertop. + +3. Spot Check Stains:" : + +(a) LLM Response + +![](images/86fe4e4caa66be29e7ca78900cdab603d56c5686b89b32eb4a4e8c9b43c83415.jpg) + +1. You should submerge the trouser in the water. + +2. Use the brush to scrub the trouser; focusing on any stains or areas that may require extra attention. + +3. Once the trouser is clean, you should rinse it." + +中 + +(b) Visual LLM Response + +# Actions in Existing Egocentric Datasets + +tackle open the tap +puncture the sandal +rub oil into dough place mug +turn the mothendose sew the cloth doll +wash trouser with brush +squeeze ball of dough +clean a sieve shake the spray adhesive +clean rice adjust gasattach the wire +place spice +open fridge pour the topsoil +clean place ball + +![](images/be2629e345d61ded979010eb6eee5b270c97ba36f8e7287d4a4e852ba4a40d16.jpg) + +![](images/acbd7ecd6c89a1bea1d9fd71317a8766d6c4895ff647dce857a47ca74f11d78b.jpg) + +![](images/bb45d21632f838a2a5dcfc1602d35bb71675fb6991faa874ec6addc94934d2c7.jpg) + +(c) Our model (LEGO) Response + +Fig. 1: When a user performing a complex task asks a large language model (LLM) for instructions (a) on how to complete the steps, she receives a generic answer and has to translate the guidance into her specific situation. If she is wearing a camera, then the prompt can be augmented with an egocentric view of the scene and passed to a Visual LLM (b), and the description is now contextualized to her situation. But she still faces the challenge of parsing a written description. When she uses our novel LEGO model (c), however, the combined image and prompt are used to automatically generate an image that provides visual guidance tailored to her exact situation from the egocentric viewpoint. Now she can complete her task seamlessly! + +Abstract. Generating instructional images of human daily actions from an egocentric viewpoint serves as a key step towards efficient skill transfer. In this paper, we introduce a novel problem - egocentric action + +frame generation. The goal is to synthesize an image depicting an action in the user's context (i.e., action frame) by conditioning on a user prompt and an input egocentric image. Notably, existing egocentric action datasets lack the detailed annotations that describe the execution of actions. Additionally, existing diffusion-based image manipulation models are sub-optimal in controlling the state transition of an action in egocentric image pixel space because of the domain gap. To this end, we propose to Learn EGOcentric (LEGO) action frame generation via visual instruction tuning. First, we introduce a prompt enhancement scheme to generate enriched action descriptions from a visual large language model (VLLM) by visual instruction tuning. Then we propose a novel method to leverage image and text embeddings from the VLLM as additional conditioning to improve the performance of a diffusion model. We validate our model on two egocentric datasets - Ego4D and Epic-Kitchens. Our experiments show substantial improvement over prior image manipulation models in both quantitative and qualitative evaluation. We also conduct detailed ablation studies and analysis to provide insights in our method. More details of the dataset and code are available on the website (https://bolinlai.github.io/Lego_EgoActGen/). + +Keywords: Egocentric Vision $\cdot$ Instruction Tuning $\cdot$ Diffusion Model + +# 1 Introduction + +The emergence of Large Language Models (LLMs) [6, 10, 64, 99], such as ChatGPT, has revolutionized the transfer of knowledge. However, an LLM alone is not a sufficient tool for human skill transfer. Consider the question answering example in Fig. 1(a). The LLM can summarize general world knowledge, but its response may not be directly applicable to the user's current circumstances. To bridge this gap, egocentric visual perception provides a novel means to capture the actions and intentions as well as the surrounding context of the camera wearer. As shown in Fig. 1(b), recent Visual Large Language Models (VLLMs) [1, 12, 46, 52, 102] can generate instructions based on the egocentric visual input. However, such verbose textual instructions are not the optimal medium for enabling efficient human skill transfer (e.g., via AR devices). Neuroscience studies have revealed that the human brain processes text much more slowly than images [4], and that humans can interpret an action from a single static image [22]. Motivated by these discoveries, we seek to design an image generation architecture that can synthesize an image which not only vividly depicts how an action should be conducted, but also seamlessly aligns with the user's visual perspective. + +Formally, we introduce the novel problem of egocentric action frame generation as depicted in Fig. 1(c). Given a user query about how to perform a specific action and an egocentric image capturing the moment before the action begins, the goal is to synthesize an egocentric image illustrating the execution of the action in the same egocentric context. We address this problem by harnessing diffusion models [27,67], which have been shown to be powerful tools for image + +manipulation [5,25,44,56,100]. There are two major challenges in using diffusion models to generate action frames from an egocentric perspective. First, the action annotations of the existing egocentric datasets [13, 21] are simply composed of a verb and nouns (see word cloud in Fig. 1), and thus lack the necessary details for diffusion models to learn the action state transition and to associate the action with correct objects and body movements. Second, the existing diffusion models are pre-trained primarily on exocentric (third-person-view) images, and have limited ability to represent complicated human daily activities from an egocentric perspective. In contrast, our proposed problem requires image generation in the egocentric view, conditioned on the user prompt of actions. The resulting domain gap impedes existing methods from synthesizing faithful and consistent egocentric action frames. + +To address these challenges, we propose to Learn EGOcentric (LEGO) action frame generation with visual instruction tuning. First, we introduce a prompt enhancement scheme to generate enriched action descriptions at scale from an instruction-tuned VLLM. Second, we incorporate the image and text embeddings from finetuned VLLM as additional conditioning into the diffusion model to narrow the domain gap and improve the controllability of action frame generation. Our experimental results suggest that the enriched action descriptions and our innovative utilization of VLLM embeddings both improve the image generation performance. Our model is able to provide a generated key action frame together with a detailed action descriptions to facilitate human skill transfer from the egocentric perspective. Overall, our contributions can be summarized as follows: + +- We introduce the novel problem of egocentric action frame generation to facilitate the process of skill transfer and address the challenges of missing action details in prompts and domain gap in existing image diffusion models. +- We propose a prompt enhancement strategy based on visual instruction tuning to enrich egocentric action descriptions, and demonstrate how the enriched descriptions can help the diffusion model understand the action state transition from the egocentric perspective. +- We propose a novel approach to incorporate the text and visual embeddings from the VLLM into the latent diffusion model to bridge the domain gap and improve the performance for egocentric action frame generation. +- We conduct thorough experiments on the Ego4D and Epic-Kitchens datasets to validate the superiority of our model over prior approaches. We also showcase the contribution of each component of our model design through ablation studies. We further provide analysis on how the visual instruction-tuned embeddings benefit model performance. + +# 2 Related Work + +Text-Guided Image Manipulation. The recent emergence of diffusion models enables text-guided image manipulation including image restoration [33], style + +transfer [75], personalized image synthesis [68,71,87], pose generation [28,36] and generic image editing [15,18,34,38,44,57,60,79,82,82-84,94,98,101]. SDEdit [56] converts the image to the latent space by adding noise through a stochastic differential equation and then denoises the representation for image editing. Rombach et al. [67] further expand SDEdit from the original stroke-based editing to text-based editing. Null-text inversion (NTI) [58] inverts a real image by DDIM [35] to yield the diffusion process and then reconstructs the image. The image can then be edited following the same strategies as Prompt-to-Prompt [25]. NTI relies on accurate inversion process which can be improved by using coupled transformations [80] or proximal guidance [24] and accelerated by a blended guidance strategy [61]. To associate the nouns with correct objects, DIFFEDIT [11] generates a mask to localize the manipulated regions. However, most inversion-based methods require accurate image captions, which are largely unavailable in the existing egocentric dataset. Recently, InstructPix2Pix [5] demonstrates the potential to edit a real image without the inversion process and original captions. However, how to leverage the diffusion model to control the state transition of an action within the egocentric image plane remains unexplored. + +Large Language Model for Image Generation. LLMs [6, 10, 64, 76, 78, 99] and VLLMs [1, 12, 23, 46, 52, 73, 95, 102] have shown their strong capability of understanding complex human instructions. Recently, LLMs and VLLMs are used to guide image generation [2, 7-9, 51, 93, 102]. Wen et al. [88] use a pretrained VLLM to pinpoint the misalignment of the text prompt and the synthetic image and then correct it using the diffusion model. Lian et al. [49] propose to generate a layout map using an LLM to improve the understanding of prompts with spatial and numerical reasoning. InstructEdit [83] uses BLIP-2 [46] to infer the objects that will be edited and then generates a mask with SAM [39] for object grounding. A pre-trained LLM can also be used as a controller to connect with various foundation models [70, 89, 90]. GILL [40] learns text-relevant image embeddings from VLLM in a few additional tokens for image generation. Importantly, all previous efforts apply the off-the-shelf foundational models directly to their problems without finetuning. In contrast, our method uses visual instruction tuning to improve the image and text embeddings from the VLLM, which narrows the domain gap and thereby facilitates the action frame generation from the egocentric point of view. + +Egocentric Vision. Recent efforts seek to understand human actions and perceptual attention [16, 29, 30, 37, 41-43, 48, 69, 74, 86], model hand-object interactions [20, 53, 54, 65, 91], and estimate human body poses [45, 55, 77, 81] from the egocentric perspective. Here, we mainly discuss the most relevant works on egocentric visual-language models and egocentric visual content generation. Lin et al. [50] propose the first egocentric video-language pre-training model - EgoVLP. Pramanick et al. [62] further improve it by incorporating multi-modal fusion directly into the video and language backbone. Ashutosh et al. [3] propose to learn a hierarchical video-language embedding for long egocentric videos. Ramakrishnan et al. [66] introduce NaQ, which is a data augmentation strategy to train models for long egocentric video search with natural language queries. In terms + +Example for In-context Learning +Input Query +![](images/8994e926f2e10d991bc7e4bc4892d5ec209c1ae0ca2897970c19ed210bb51dbd.jpg) +Action Label: "chop end of onion" Bounding Boxes: "left hand-[0.203, 0.368, 0.499, 1.000], right hand-[0.530, 0.527, 0.797, 1.000], chopping board-[0.248, 0.156, 0.644, 0.750], onion-[0.462, 0.377, 0.509, 0.506], knife-[0.484, 0.343, 0.546, 0.586]" + +Query and Response +Input Query +Fig. 2: Examples for data curation using GPT-3.5. We provide detailed action descriptions of several example images as well as their action labels and bounding boxes for in-context learning. In addition, we input the action label and bounding boxes of another action as a query. GPT-3.5 is able to generate descriptions with enriched information (highlighted by underlines) in the response. +![](images/692f9f9ea4a6532001bf520451bdb128cea9028caef34ef5a6d582f10aeb9147.jpg) +Action Description: "The person presses the onion on the chopping board with the left hand and then cuts off the end of the onion with a knife in the right hand." +Action Label: "squidge into lunch box" Bounding Boxes: "left hand-[0.427, 0.799, 0.591, 1.00], right hand-[0.654, 0.949, 0.707, 1.00], lunch box-[0.390, 0.512, 0.696, 0.993], sink-[0.00, 0.802, 0.315, 1.00], container lid-[0.403, 0.454, 0.513, 0.635], fork-[0.589, 0.798, 0.672, 1.00]" +GPT-3.5 Response: "The person uses their left hand to hold a lunch box, while their right hand uses a fork to squidge food into the lunch box, which is placed on the sink with the container lid nearby." + +of egocentric visual generation, Jia et al. [32] leverage GANs [19] to generate future head motion in hand forecasting task. Zhang et al. [96] leverage GANs to facilitate future gaze anticipation. Ye et al. [92] propose the affordance diffusion model that takes in the image of an object and generates possible hand-object interactions in the egocentric view. In this paper, we present the first work that investigates how to leverage VLLMs and diffusion models to generate action state transition on the egocentric image plane. + +# 3 Method + +The problem setting of egocentric action frame generation is illustrated in Fig. 1(c). Given an egocentric image frame $\mathcal{X}$ that captures the user's current visual context as well as a user query prompt $\mathcal{P}$ regarding how to perform an action, our goal is to synthesize the action frame $\mathcal{Y}$ that visually depicts how the action should be conducted in the same situation (i.e., keep a consistent background). + +The key insight of our proposed LEGO model is leveraging the strong capability of a VLLM to enhance the diffusion model for egocentric action frame generation. The annotations of existing egocentric datasets do not describe the details of how actions are conducted. As a remedy, we leverage visual instruction tuning to finetune a VLLM that enriches action descriptions based on the egocentric visual prompt. In addition, the existing diffusion-based image manipulation models are limited in understanding egocentric action state transition, due to the domain gap between the exocentric pre-training dataset and the ego- + +![](images/f8e36e3ed9cec7943cda7be08b211f7d4835e082ce1346fbfdcb6504d546c0f5.jpg) +Fig. 3: Overview of our proposed LEGO model. We first finetune a visual large language model (VLLM) to generate the enriched action description with visual instruction tuning. We then project image and text embeddings from the finetuned VLLM to the feature space of the latent diffusion model (LDM). Finally, we train the LDM to synthesize the egocentric action frame conditioning on the input frame, enriched action description, as well as the VLLM image and text embeddings. + +centric action dataset for our problem. To bridge this gap, we propose a novel approach that leverages VLLM embeddings to control the state transition of actions and to generate action frames accordingly. We detail the VLLM-based data enrichment pipeline and our model design in the following sections. + +# 3.1 Egocentric Visual Instruction Tuning + +Data Curation for Visual Instruction Tuning. As shown in Fig. 2, we use GPT-3.5 to generate detailed action descriptions based on an input query composed of a short action label and object bounding boxes that are provided in the existing datasets. First, we prepare several examples of possible inputs along with their expected output descriptions for GPT-3.5 to perform in-context learning. These examples cover a diverse set of scenes in the egocentric action dataset. Each example is composed of an action label, a manually annotated action description, and relative spatial information of hands and objects-of-interests represented by bounding box coordinates. GPT-3.5 can learn from the given examples and generate similar detailed action descriptions based on a new input query. The resulting GPT-3.5 curated data is then further used for visual instruction tuning. More details of the prompt are provided in the supplementary. + +Visual Instruction Tuning. We follow the finetuning strategy in prior work [52] as shown in Fig. 3(a). Specifically, we use the pre-trained CLIP visual encoder [63] $\phi$ to encode the visual representation and then apply a linear projection layer $\tau$ to map the CLIP visual features into the semantic space of the LLM, i.e., $\mathcal{H}_i = \tau (\phi (\mathcal{X}))$ . To construct the user prompt $\mathcal{P}$ , we insert the action label annotation into a prompt template to create a coherent query that is aligned with the instruction-following nature of an LLM. We then tokenize $\mathcal{P}$ , and feed both the prompt text tokens and image tokens $\mathcal{H}_i$ as inputs into the LLM. Finally, the LLM is trained to generate enriched action description (denoted as $\mathcal{R}$ ) based on the user prompt and image input. + +User Prompt Enrichment at Scale. Note that the visual instruction tuned VLLM doesn't rely on any object bounding boxes as input. Therefore, we can generate enriched action descriptions for all egocentric action data at scale. + +# 3.2 Egocentric Action Frame Generation + +We leverage a latent diffusion model (LDM) [67] to synthesize the action frame conditioning on the input frame and the detailed action description $\mathcal{R}$ generated by our finetuned VLLM (see Fig. 3(b)). Following the regular steps in LDMs [5, 67], the input image is first encoded into a latent space using a pre-trained autoencoder $\mathcal{E}$ . Then the input to the denoising UNet is a concatenation of the latent input representation $\mathcal{E}(\mathcal{X})$ and a Gaussian noise $\mathcal{G}$ . + +Our key innovation is to design the U-Net conditioning component so that the diffusion model can interpret the egocentric actions correctly. To start, we follow [5] and adopt the conventional pre-trained CLIP text encoder $\psi$ to extract a text representation of $\mathcal{R}$ , i.e., $\psi(\mathcal{R}) \in \mathbb{R}^{N \times D}$ where $N$ is the maximum number of text tokens and $D$ is the number of feature channels. We further leverage the image and text embeddings from the visual instruction tuned VLLM as additional LDM conditioning to alleviate the domain gap issue. + +Specifically, we feed the VLLM image embedding $\mathcal{H}_i$ into an extra linear layer $\sigma$ to map it to LDM feature space, i.e., $\sigma(\mathcal{H}_i) \in \mathbb{R}^{M \times D}$ , where $M$ is the number of image tokens. Note that $\mathcal{H}_i$ is already projected to the semantic space during visual instruction tuning, and therefore differs from the image embedding $\mathcal{E}(\mathcal{X})$ from the auto-encoder. Moreover, we also extract the text embedding $\mathcal{H}_t$ before the last linear layer of LLM. We adopt a fixed token number $N$ and enforce padding or truncation behavior, as in the CLIP text encoder. The text embedding is then fed to a projection layer $\mu$ . In LLM decoder, the response is generated iteratively and each word embedding only conditions on the context ahead of it. To extract the holistic semantic meaning of $\mathcal{H}_t$ in LDM feature space, we further add self-attention layers $\pi$ after the projection, i.e., $\pi(\mu(\mathcal{H}_t)) \in \mathbb{R}^{N \times D}$ . Thus, the U-Net conditioning can be formulated as: + +$$ +\mathcal {C} = \left[ \psi (\mathcal {R}), \sigma \left(\mathcal {H} _ {i}\right), \pi \left(\mu \left(\mathcal {H} _ {t}\right)\right) \right] \in \mathbb {R} ^ {(2 N + M) \times D}. \tag {1} +$$ + +The conditioning $\mathcal{C}$ is fed into the denoising UNet at multiple layers via the cross-attention mechanism [67]. We assume the intermediate feature of a specific + +UNet layer is $\mathcal{U}$ , which is learned from the UNet input (i.e., the concatenation of input frame representation $\mathcal{E}(\mathcal{X})$ and Gaussian noise $\mathcal{G}$ ). The cross-attention at this UNet layer can be formulated as: + +$$ +C r o s s A t t (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {D}}\right) \cdot V, \tag {2} +$$ + +where $Q = W_{Q} \cdot \mathcal{U}$ , $K = W_{K} \cdot \mathcal{C}$ and $V = W_{V} \cdot \mathcal{C}$ . Note that $W_{Q}, W_{K}$ and $W_{V}$ are learnable matrices. We also adopt the classifier-free guidance following [5] (see supplementary for details). Finally, the UNet output is converted to the image domain by a pre-trained decoder. + +# 3.3 Implementation Details + +All parameters of the VLLM are initialized from the pre-trained LLaVA [52] weights. During training, we freeze the CLIP image encoder and finetune the projection layer and LLM with cross-entropy loss for 3 epochs. To improve the diversity of the text prompts, we randomly select the question template from 10 candidates in each iteration. For LDM training, the text encoder, UNet and autoencoder are initialized with pre-trained weights [67]. The projection layers $\sigma$ and $\mu$ and the self-attention layers $\pi$ are initialized using the Xavier algorithm [17]. The text encoder is frozen and the remaining weights are finetuned with L2 regression loss between the predicted noise and real noise for 20,000 iterations. All input and target images are resized to a resolution of $256\times 256$ . Please refer to the supplementary for more details of training and inference. + +# 4 Experiments + +# 4.1 Data and Metrics + +Datasets. We conduct our experiments on two well-established egocentric action datasets - Ego4D [21] and Epic-Kitchens-100 [13]. Both datasets were densely annotated with action starting time $t$ and ending time $\hat{t}$ . In our problem setting, we select an egocentric image frame $\delta_{i}$ seconds before the action begins as the input $\mathcal{X}$ , and an image $\delta_{o}$ seconds after the action begins as the target frame $\mathcal{Y}$ . On the Ego4D dataset, based on the annotations of Pre-Condition-15 time (PRE-15) $t_{pre}$ , and Point-of-No-Return time (PNR) $t_{pnr}$ , we set $\delta_{i} = t - t_{pre}$ and $\delta_{o} = t_{pnr} - t$ . For Epic-Kitchens, PNR and PRE-15 annotations are not available. Instead, we empirically select $\delta_{i} = 0.25$ seconds and $\delta_{o} = t + \lambda * (\hat{t} - t)$ , where $\lambda = 0.6$ , for our experiments. More details of data preparation and improvement are elaborated in the supplementary. + +Metrics. We adopt image-to-image similarity, image-to-text similarity, and user study as metrics in our experiments. + +- Image-to-Image Metrics. We implement six metrics to evaluate image-to-image similarity. To begin with, we adopt three contrastive learning based + +metrics including image-to-image (1) EgoVLP score [50], (2) EgoVLP $^+$ score [50] and (3) CLIP score [63]. EgoVLP is a contrastive video-language pretraining model trained with egocentric videos. Since EgoVLP takes multiple frames as input, we consider two types of inputs: duplicating the output frame as a static input (i.e., EgoVLP score) and combining the input frame with the output frame (i.e., EgoVLP $^+$ score). As a result, EgoVLP $^+$ can effectively measure whether the generated frame can depict the state transition of an action. Importantly, given that EgoVLP is pre-trained on egocentric data and action labels, we consider EgoVLP and EgoVLP $^+$ score as the primary automatic metrics. In addition, we also report (4) Fréchet Inception Distance (FID) [26], (5) Peak Signal-to-Noise Ratio (PSNR) and (6) Learned Perceptual Image Patch Similarity (LPIPS) [97] (with SqueezeNet [31] as the encoder) to make a thorough evaluation. Note that instead of measuring similarity of input and output frames as in prior works [12,80], in our problem setting, we measure the similarity between the generated action frame and ground truth, which better reflects whether the generation results can illustrate the execution of an action. + +- Image-to-Text Metrics. We find the widely-used image-to-text CLIP score can not align actions with egocentric images due to the domain gap. Similar misalignment problem is also observed in [14, 59, 72, 85]. In our experiments, we utilize BLIP [47] to generate image captions of the output images and then calculate text-to-text similarity using CLIP text encoder (following [34]). We implement this metric with two BLIP structures: BLIP-B and BLIP-L. Though this solution may still suffer from the same domain gap issue, it is a more appropriate evaluation metric in our problem setting. See more evidence and discussions in the supplementary. + +- **User Study.** We also conduct a user study on a subset of test data to further validate the advantage of our model based on human preference. We sample 60 examples from each dataset and present the generated frames from our model as well as the baseline models to raters on Amazon Mechanical Turk (AMT). We also provide the input frames and the corresponding action labels during evaluation. For each instance, we ask the raters to select the image that best aligns with the provided action label while preserving the most contextual information from the input frame. To minimize potential bias, we hire 5 AMT raters to annotate each example thus resulting in 300 samples for each dataset. User study interface is shown in supplementary. + +# 4.2 Comparison with Prior Approaches + +We compare our proposed model with previous state-of-the-art methods for text-guided image manipulation, including ProxEdit [24], SDEdit [56] and InstructPix2Pix (IP2P) [5]. For a fair comparison, we finetune these baseline methods with the same data used in our experiments. Specifically, we train IP2P in an end-to-end way on the two datasets with existing egocentric action labels as + +Table 1: Comparison with prior image manipulation approaches in image-to-image metrics. $\downarrow$ means a lower score in this metric suggests a better performance. The best results are highlighted with **boldface**. The orange row refers to our LEGO model. + +
MethodsEgoVLPEgoVLP+CLIPFID ↓PSNRLPIPS ↓
Ego4DProxEdit [24]44.5172.6868.1733.0111.8840.90
SDEdit [56]50.0772.9073.3533.3511.8141.60
IP2P [5]62.1978.8478.7524.7312.1637.16
LEGO65.6580.4480.6123.8312.2936.43
EK-100ProxEdit [24]32.2752.7765.8051.3511.0646.35
SDEdit [56]33.8456.8074.7627.4111.3043.33
IP2P [5]42.9761.0677.0320.6411.2340.82
LEGO45.8962.6678.6321.5711.3340.36
+ +Table 2: Image-to-text metrics of our model and baselines. The best results are highlighted with boldface. The orange row refers to our LEGO model performance. + +
MethodsEgo4DEpic-Kitchens
BLIP-BBLIP-LBLIP-BBLIP-L
ProxEdit [24]17.7317.3523.6523.39
SDEdit [56]19.8019.7421.5121.30
IP2P [5]20.0020.5625.3726.36
LEGO20.3820.7026.9827.41
+ +![](images/3d54ba3028473f279513c0b8fc8c34ecbb6c6d9a903308e97aa675438364ae3e.jpg) +Fig.4: User study of our model and baselines. Win rate is the percentage of each model being picked as the best. + +text conditioning. ProxEdit and SDEdit rely on the off-the-shelf latent diffusion model to synthesize edited images and thus can not be trained end to end. Therefore, we first finetune the latent diffusion model with egocentric images and action labels and then use the finetuned latent diffusion model parameters to implement ProxEdit and SDEdit approaches. Please refer to the supplementary for more baseline implementation details and a thorough comparison with prior methods that leverage LLMs for image generation. + +In terms of image-to-image metrics in Tab. 1, both ProxEdit and SDEdit perform poorly on this novel problem, suggesting the challenging nature of generating egocentric action frames. IP2P performs much better in almost all metrics by end-to-end training and serves as the strongest baseline model in our experiments. However, our LEGO model consistently outperforms IP2P by a large margin (3.46%, 1.60%, 1.86%, 0.90, 0.13 and 0.73% respectively) in all six metrics on Ego4D. LEGO also exceeds IP2P notably (2.92%, 1.60%, 1.60%, 0.10 and 0.46% respectively) in five metrics on Epic-Kitchens. Though, LEGO slightly lags behind IP2P in FID score, it still achieves the second best performance. + +With regard to image-to-text metrics in Tab. 2, LEGO outperforms the strongest baseline (IP2P) by $0.38\%$ and $0.14\%$ on Ego4D and by $1.61\%$ and + +![](images/90f1142b2076eb8b9cd07b80ab0d511b9e20304e25902745a3b4208f97c799d7.jpg) +Fig. 5: Visualization of the proposed LEGO model and all baselines on Ego4D (the first three rows) and Epic-Kitchens (the last three rows). The action frames generated by LEGO align with the user prompt better and preserve more contexts in input frames. + +$1.05\%$ on Epic-Kitchens. The result suggests the synthetic frames from our model can better align with the action descriptions. However, the performance gain is rather limited. We emphasize that this is because the domain gap still exists for the BLIP model. Besides baseline models, we also measure the image-to-text similarity of input frames and action prompts to validate models' capability of learning action state transition in semantics. The BLIP-B/BLIP-L scores are $15.44\% / 15.49\%$ on Ego4D and $20.11\% / 20.52\%$ on Epic-Kitchens, lagging behind all baseline models. The result evidences diffusion models are able to understand actions and edit the input frame semantically towards the action prompt. + +Due to the limitation of automatic metrics, we additionally implement user study as a complement. We shuffle the order of the results from our model and the baselines while presenting them to the raters. We define the win rate as the percentage of each model being picked as the best out of the total 300 + +![](images/b816b15b296273af7f7290d15dac4041a1a449993d575d8632bb0ed54e0a83ed.jpg) +Input Frame + +![](images/9f974032032c25ffc259cd1a20f7a7784ebcfa6b9a999132fead39093da75d10.jpg) +LEGO + +![](images/51390262476acd0907430caf3bce304be445d5e58f5034221078d342413828d6.jpg) +cut a portion of clay brush a wood mix with both hands with a brush + +![](images/69dd1a6c4d01d0466fabc1e3a5dc2d909b01917a994ef16efa11f9be47f9c6f1.jpg) +"put tray in oven" +Fig. 6: Additional visualization of our proposed model. LEGO is able to synthesize faithful action frames as well as preserve the contexts in various scenarios. + +![](images/e23e14c041747c1193cec14305b1150f3bb23b4654e6c3d8a3ad236a0e9e1560.jpg) +"close container" + +![](images/f0a5be6566924e6471da1f7f68c3bf797b24ddb3f2952e81c61bb4f554e091ed.jpg) + +![](images/92b7729b3181df5a05b4551156b81e7bf8296346f53fc9d1ceb8bbf68b7b6d6e.jpg) +"take soy milk" + +samples. Results are illustrated in Fig. 4. Our model surpasses the best baseline by $44.00\%$ and $38.34\%$ respectively on Ego4D and Epic-Kitchens. The prominent gap further validates the superiority of our model. + +# 4.3 Visualization of Generated Frames + +We additionally showcase examples of generated image from LEGO and baseline models in Fig. 5. ProxEdit and SDEdit fail to understand the user prompt and thus generate frames of irrelevant actions (e.g., row3). They may also easily synthesize the action in a different environment (e.g., row2). InstructPix2Pix is able to preserve more contexts but still fails to semantically align with the action in user prompt (e.g., row1). In contrast, LEGO can synthesize action frames that better align with the user prompts and retain more contexts in the background. More examples of LEGO output are presented in Fig. 6 and supplementary. + +# 4.4 Ablation Study + +We present comprehensive ablation studies to investigate the contribution of enriched action descriptions, the VLLM image embedding and the VLLM text embedding separately. Results are demonstrated in Tab. 3. Given the limitation of automatic metrics, we also provide user study as the primary metric. Notably, conditioning on enriched action descriptions can moderately improve the model performance, supporting our claim that expanding the action description can facilitate the learning of the state transition during an egocentric action. Moreover, utilizing the image embedding and text embedding from VLLM as additional conditions both improve the model performance by a notable margin because VLLM embeddings can effectively narrow the domain gap. Interestingly, the image embedding leads to larger performance boost on both datasets. These results suggest the VLLM image embedding $\mathcal{H}_i$ incorporates important high-level semantic meanings that are not captured in the auto-encoder image embedding $\mathcal{E}(\mathcal{X})$ or VLLM text emebdding $\mathcal{H}_t$ . Finally, our full LEGO model uses both VLLM image and text embeddings (Desc.+Joint Embed.) and thus achieves the best performance on both datasets in all metrics. + +Table 3: Analysis of egocentric action frame generation performance with different conditionings. Joint Embed. refers to incorporating both VLLM image and text embeddings. Similar to Fig. 4, we present win rate as the user study result, i.e., the percentage of each model picked as the best (\% is omitted for simplicity). The best results are highlighted with **boldface**. The orange rows refer to our full LEGO model. + +
ConditioningEgo4DEpic-Kitchens
User StudyEgoVLPEgoVLP+CLIPUser StudyEgoVLPEgoVLP+CLIP
Actions Labels5.3362.1978.8478.757.0842.9761.0677.03
Descriptions13.0062.9179.0979.1812.5043.7261.4677.47
Desc. + Img Embed.26.0065.3580.1380.5724.1745.8262.2978.60
Desc. + Txt Embed.21.3363.2979.4079.2122.0844.6862.0277.74
Desc. + Joint Embed.34.3465.6580.4480.6134.1745.8962.6678.63
+ +Table 4: Performance of LEGO without finetuning (w/o FT). The best results are highlighted with **boldface**. The orange rows refer to our full LEGO model. + +
ConditioningEgo4DEpic-Kitchens
EgoVLPEgoVLP+CLIPEgoVLPEgoVLP+CLIP
Descriptions62.9179.0979.1843.7261.4677.47
Desc.+Joint Embed.(w/o FT)64.5779.7280.2344.7461.8778.35
Desc.+Joint Embed.(w/ FT)65.6580.4480.6145.8962.6678.63
+ +# 4.5 Analysis of Visual Instruction Tuning + +First, we assess the quality of enriched action descriptions from the visual instruction tuned VLLM by user study. For each sample in user study, we ask the raters to select whether the description aligns with the input and target frames (see supplementary for more details of user study setting and interface). The percentage of samples with aligned frames and action descriptions is $87\%$ on Ego4D and $84\%$ on Epic-Kitchens. The high alignment percentage suggests the visual instruction tuned VLLM can effectively expand action labels with details captured in the input frame. We additionally implement the same user study for off-the-shelf VLLM (i.e., without finetuning). The percentage of alignment is just $27\%$ on Ego4D and $30\%$ on Epic-Kitchens with hallucination existing in $92\%$ of unaligned instances. The notable drop supports our argument that visual instruction tuning is a critical step for high-quality prompt enrichment. + +In addition, we further investigate whether the visual instruction tuned embeddings can contribute more to the diffusion model than the off-the-self counterpart, which has not been well studied in prior work. As shown in Tab. 4, without any finetuning, the image and text embeddings from VLLM can still improve the baseline model (Descriptions). However finetuned embeddings yield much larger improvement (e.g., $1.08\%$ and $1.15\%$ in EgoVLP score) on both datasets. The result suggests that visual instruction tuning is a necessary step to learn semantic image and text embeddings that are more aligned with the + +![](images/771ec954a02730affa2393565be317566054377595ac6d16570e1c534346b433.jpg) + +![](images/bffe851e6fc13a86dc8e98805295c967d1203ada550d9ff9fc4f59412a8907e1.jpg) +Fig. 7: Visualization of generating various actions from the same input frame. The first action ("open drawer") is the existing label for this example in the dataset. We fill another five actions into the user query and synthesize the action frames using our model. All generated frames align well with the actions and preserve most contexts. + +"Can you provide instructions on how to {action} in my current situation?" + +![](images/ed431d60297d8bcc73609181e5df9d3afcb5f6ca1ed5ccbcd5b922028eee54ab.jpg) +"open drawer" + +![](images/cc4e8453b5e774f08fca5106185f04ba980f670673820f7b3fd466669a0414a0.jpg) +"dry hands" + +![](images/d994be76e84a8483fc6dbd67883afd705ed53feff5fed79fa396627070db72d6.jpg) +"cut cucumber" + +![](images/deb5960858bafc06089cc1a8dad5944d634f12dd52a0aee0a0ac7a770d5f5b53.jpg) +"open microwave" + +![](images/159d8b797a447c7c0b0a3320727e86d8ae99283a87d17b12c2921e2a07ce4dbb.jpg) +"pick up bowl" + +![](images/007d64c34566d8aed4e12ee1ff1f71551abcb90da2698d560bd335db0cb30f87.jpg) +"take knife" + +![](images/0ddb6f1ecf13a62554b791b575a2f3902e670bd68daf08d1b82238226dc3309f.jpg) +LEGO + +egocentric input frame and action prompt, which thus narrows the domain gap and greatly boosts the performance of the latent diffusion model. + +# 4.6 Generation of Various Actions with the Same Contexts + +In addition to generating action frames based on the pre-defined action labels in our datasets, we validate the generative nature of our model i.e., whether LEGO is able to synthesize the correct action frames conditioning on the same contexts yet different actions (novel image-action pairs). Results are illustrated in Fig. 7. We feed different actions with the same input frame to our model. The synthesized frames correctly depict the execution of the actions in the same environment. This result further indicates that our model can understand the action state transition and generalize to different user queries. + +# 5 Conclusion + +In this paper, we introduce the novel problem of egocentric action frame generation. We also propose a novel model — LEGO that leverages visual instruction tuning and a diffusion model to address this problem. Our key intuition is to use visual instruction tuning to generate informative responses that depict the execution of the egocentric action, and then design the conditioning for the denoising U-Net to exploit the image and text embeddings from a visual instruction tuned VLLM. Our experiments on two large-scale egocentric action datasets validate the advantage of the proposed approach as well as the contribution of each model component. We believe our work provides an important step in understanding the action state transition and controllability of diffusion models, and suggests future research directions in egocentric AI systems, action state transition, image generation and human skill transfer. + +# Acknowledgements + +Portions of this work were supported in part by a gift from Meta. We thank Sangmin Lee for the valuable discussion and suggestions. + +# References + +1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022) +2. Anonymity: Making multimodal generation easier: When diffusion models meet llms. openreview (2023) +3. Ashutosh, K., Girdhar, R., Torresani, L., Grauman, K.: Hiervl: Learning hierarchical video-language embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23066-23078 (2023) +4. Baskin, J.H., Edersheim, J.G., Price, B.H.: Is a picture worth a thousand words? neuroimaging in the courtroom. American Journal of Law & Medicine 33(2-3), 239-269 (2007) +5. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023) +6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) +7. Chakrabarty, T., Singh, K., Saakyan, A., Muresan, S.: Learning to follow object-centric image editing instructions faithfully. arXiv preprint arXiv:2310.19145 (2023) +8. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., et al.: Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426 (2023) +9. Chen, W.G., Spiridonova, I., Yang, J., Gao, J., Li, C.: Llava-interactive: An all-in-one demo for image chat, segmentation, generation and editing. arXiv preprint arXiv:2311.00571 (2023) +10. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022) +11. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427 (2022) +12. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. Advances in Neural Information Processing Systems (2023) +13. Damen, D., Doughty, H., Farinella, G.M., Furnari, A., Kazakos, E., Ma, J., Moltisanti, D., Munro, J., Perrett, T., Price, W., et al.: Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision pp. 1-23 (2022) +14. Du, Y., Yang, S., Dai, B., Dai, H., Nachum, O., Tenenbaum, J.B., Schuurmans, D., Abbeel, P.: Learning universal policies via text-guided video generation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) + +15. Epstein, D., Jabri, A., Poole, B., Efros, A.A., Holynski, A.: Diffusion self-guidance for controllable image generation. In: Advances in Neural Information Processing Systems (2023) +16. Girdhar, R., Grauman, K.: Anticipative video transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13505-13515 (2021) +17. Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedforward neural networks. In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. pp. 249-256. JMLR Workshop and Conference Proceedings (2010) +18. Goel, V., Peruzzo, E., Jiang, Y., Xu, D., Sebe, N., Darrell, T., Wang, Z., Shi, H.: Pair-diffusion: Object-level image editing with structure-and-appearance paired diffusion models. arXiv preprint arXiv:2303.17546 (2023) +19. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139–144 (2020) +20. Goyal, M., Modi, S., Goyal, R., Gupta, S.: Human hands as probes for interactive object understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3293-3303 (2022) +21. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18995-19012 (2022) +22. Hafri, A., Trueswell, J.C., Epstein, R.A.: Neural representations of observed actions generalize across static and dynamic visual input. Journal of Neuroscience 37(11), 3056-3071 (2017) +23. Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., et al.: Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905 (2023) +24. Han, L., Wen, S., Chen, Q., Zhang, Z., Song, K., Ren, M., Gao, R., Stathopoulos, A., He, X., Chen, Y., et al.: Proxedit: Improving tuning-free real image editing with proximal guidance. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4291-4301 (2024) +25. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022) +26. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017) +27. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) +28. Huang, J., Liu, Y., Qin, J., Chen, S.: Kv inversion: Kv embeddings learning for text-conditioned real image action editing. arXiv preprint arXiv:2309.16608 (2023) +29. Huang, Y., Cai, M., Li, Z., Lu, F., Sato, Y.: Mutual context network for jointly estimating egocentric gaze and action. IEEE Transactions on Image Processing 29, 7795-7806 (2020) +30. Huang, Y., Cai, M., Li, Z., Sato, Y.: Predicting gaze in egocentric video by learning task-dependent attention transition. In: Proceedings of the European conference on computer vision (ECCV). pp. 754-769 (2018) + +31. Iandola, F.N., Han, S., Moskewicz, M.W., Ashraf, K., Dally, W.J., Keutzer, K.: SqueezeNet: Alexnet-level accuracy with 50x fewer parameters and $< 0.5$ mb model size. arXiv preprint arXiv:1602.07360 (2016) +32. Jia, W., Liu, M., Rehg, J.M.: Generative adversarial network for future hand segmentation from egocentric video. In: European Conference on Computer Vision. pp. 639-656. Springer (2022) +33. Jiang, Y., Zhang, Z., Xue, T., Gu, J.: Autodir: Automatic all-in-one image restoration with latent diffusion. arXiv preprint arXiv:2310.10123 (2023) +34. Joseph, K., Udhayanan, P., Shukla, T., Agarwal, A., Karanam, S., Goswami, K., Srinivasan, B.V.: Iterative multi-granular image editing using diffusion models. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2024) +35. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. In: Advances in Neural Information Processing Systems (2022) +36. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023) +37. Kazakos, E., Nagrani, A., Zisserman, A., Damen, D.: Epic-fusion: Audio-visual temporal binding for egocentric action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5492-5501 (2019) +38. Kim, S., Jang, W., Kim, H., Kim, J., Choi, Y., Kim, S., Lee, G.: User-friendly image editing with minimal text input: Leveraging captioning and injection techniques. arXiv preprint arXiv:2306.02717 (2023) +39. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) +40. Koh, J.Y., Fried, D., Salakhutdinov, R.R.: Generating images with multimodal language models. Advances in Neural Information Processing Systems 36 (2024) +41. Lai, B., Liu, M., Ryan, F., Rehg, J.M.: In the eye of transformer: Global-local correlation for egocentric gaze estimation. British Machine Vision Conference (2022) +42. Lai, B., Liu, M., Ryan, F., Rehg, J.M.: In the eye of transformer: Global-local correlation for egocentric gaze estimation and beyond. International Journal of Computer Vision 132(3), 854-871 (2024) +43. Lai, B., Ryan, F., Jia, W., Liu, M., Rehg, J.M.: Listen to look into the future: Audio-visual egocentric gaze anticipation. arXiv preprint arXiv:2305.03907 (2023) +44. Li, D., Li, J., Hoi, S.C.: Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems (2023) +45. Li, J., Liu, K., Wu, J.: Ego-body pose estimation via ego-head pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17142-17151 (2023) +46. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. International Conference on Machine Learning (2023) +47. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) + +48. Li, Y., Liu, M., Rehg, J.M.: In the eye of beholder: Joint learning of gaze and actions in first person video. In: Proceedings of the European conference on computer vision (ECCV). pp. 619-635 (2018) +49. Lian, L., Li, B., Yala, A., Darrell, T.: Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655 (2023) +50. Lin, K.Q., Wang, J., Soldan, M., Wray, M., Yan, R., XU, E.Z., Gao, D., Tu, R.C., Zhao, W., Kong, W., et al.: Egocentric video-language pretraining. Advances in Neural Information Processing Systems 35, 7575-7586 (2022) +51. Liu, B., Zhang, H., Liu, J., Wang, Q.: Acigs: An automated large-scale crops image generation system based on large visual language multi-modal models. In: 2023 20th Annual IEEE International Conference on Sensing, Communication, and Networking (SECON). pp. 7-13. IEEE (2023) +52. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems (2023) +53. Liu, M., Tang, S., Li, Y., Rehg, J.M.: Forecasting human-object interaction: joint prediction of motor attention and actions in first person video. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 704-721. Springer (2020) +54. Liu, S., Tripathi, S., Majumdar, S., Wang, X.: Joint hand motion and interaction hotspots prediction from egocentric videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3282-3292 (2022) +55. Luo, Z., Hachiuma, R., Yuan, Y., Kitani, K.: Dynamics-regulated kinematic policy for egocentric pose estimation. Advances in Neural Information Processing Systems 34, 25019-25032 (2021) +56. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: International Conference on Learning Representations (2022) +57. Mirzaei, A., Aumentado-Armstrong, T., Brubaker, M.A., Kelly, J., Levinshtein, A., Derpanis, K.G., Gilitschenski, I.: Watch your steps: Local image and scene editing by text instructions. arXiv preprint arXiv:2308.08947 (2023) +58. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6038-6047 (2023) +59. Molad, E., Horwitz, E., Valevski, D., Acha, A.R., Matias, Y., Pritch, Y., Leviathan, Y., Hoshen, Y.: Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329 (2023) +60. Orgad, H., Kawar, B., Belinkov, Y.: Editing implicit assumptions in text-to-image diffusion models. arXiv preprint arXiv:2303.08084 (2023) +61. Pan, Z., Gherardi, R., Xie, X., Huang, S.: Effective real image editing with accelerated iterative diffusion inversion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15912-15921 (2023) +62. Pramanick, S., Song, Y., Nag, S., Lin, K.Q., Shah, H., Shou, M.Z., Chellappa, R., Zhang, P.: Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5285-5297 (2023) +63. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) + +64. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) +65. Ragusa, F., Farinella, G.M., Furnari, A.: Stillfast: An end-to-end approach for short-term object interaction anticipation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3635-3644 (2023) +66. Ramakrishnan, S.K., Al-Halah, Z., Grauman, K.: Naq: Leveraging narrations as queries to supervise episodic memory. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6694-6703 (2023) +67. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +68. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023) +69. Ryan, F., Jiang, H., Shukla, A., Rehg, J.M., Ithapu, V.K.: Egocentric auditory attention localization in conversations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14663-14674 (2023) +70. Shen, Y., Song, K., Tan, X., Li, D., Lu, W., Zhuang, Y.: Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580 (2023) +71. Shi, J., Xiong, W., Lin, Z., Jung, H.J.: Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411 (2023) +72. Stein, G., Cresswell, J.C., Hosseinzadeh, R., Sui, Y., Ross, B.L., Villecloze, V., Liu, Z., Caterini, A.L., Taylor, J.E.T., Loaiza-Ganem, G.: Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. arXiv preprint arXiv:2306.04675 (2023) +73. Su, Y., Lan, T., Li, H., Xu, J., Wang, Y., Cai, D.: Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355 (2023) +74. Sudhakaran, S., Escalera, S., Lanz, O.: Lsta: Long short-term attention for egocentric action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9954-9963 (2019) +75. Sun, Z., Zhou, Y., He, H., Mok, P.: Sgdiff: A style guided diffusion model for fashion synthesis. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 8433-8442 (2023) +76. Thoppilan, R., De Freitas, D., Hall, J., Shazeer, N., Kulshreshtha, A., Cheng, H.T., Jin, A., Bos, T., Baker, L., Du, Y., et al.: Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239 (2022) +77. Tome, D., Alldieck, T., Peluse, P., Pons-Moll, G., Agapito, L., Badino, H., De la Torre, F.: Selfpose: 3d egocentric pose estimation from a headset mounted camera. IEEE Transactions on Pattern Analysis and Machine Intelligence (2020) +78. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) +79. Tsaban, L., Passos, A.: Ledits: Real image editing with ddpm inversion and semantic guidance. arXiv preprint arXiv:2307.00522 (2023) +80. Wallace, B., Gokul, A., Naik, N.: Edict: Exact diffusion inversion via coupled transformations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22532-22541 (2023) + +81. Wang, J., Luvizon, D., Xu, W., Liu, L., Sarkar, K., Theobalt, C.: Scene-aware egocentric 3d human pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13031-13040 (2023) +82. Wang, K., Yang, F., Yang, S., Butt, M.A., van de Weijer, J.: Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. arXiv preprint arXiv:2309.15664 (2023) +83. Wang, Q., Zhang, B., Birsak, M., Wonka, P.: Instructedit: Improving automatic masks for diffusion-based image editing with user instructions. arXiv preprint arXiv:2305.18047 (2023) +84. Wang, Q., Zhang, B., Birsak, M., Wonka, P.: Mdp: A generalized framework for text-guided image editing by manipulating the diffusion path. arXiv preprint arXiv:2303.16765 (2023) +85. Wang, W., Xie, K., Liu, Z., Chen, H., Cao, Y., Wang, X., Shen, C.: Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599 (2023) +86. Wang, X., Zhu, L., Wang, H., Yang, Y.: Interactive prototype learning for egocentric action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8168-8177 (2021) +87. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848 (2023) +88. Wen, S., Fang, G., Zhang, R., Gao, P., Dong, H., Metaxas, D.: Improving compositional text-to-image generation with large vision-language models. arXiv preprint arXiv:2310.06311 (2023) +89. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023) +90. Wu, S., Fei, H., Qu, L., Ji, W., Chua, T.S.: Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519 (2023) +91. Xu, Y., Li, Y.L., Huang, Z., Liu, M.X., Lu, C., Tai, Y.W., Tang, C.K.: Egopca: A new framework for egocentric hand-object interaction understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5273-5284 (2023) +92. Ye, Y., Li, X., Gupta, A., De Mello, S., Birchfield, S., Song, J., Tulsiani, S., Liu, S.: Affordance diffusion: Synthesizing hand-object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22479-22489 (2023) +93. Yu, Q., Li, J., Ye, W., Tang, S., Zhuang, Y.: Interactive data synthesis for systematic vision adaptation via llms-aigcs collaboration. arXiv preprint arXiv:2305.12799 (2023) +94. Yu, Z., Li, H., Fu, F., Miao, X., Cui, B.: Fisted: Accelerating text-to-image editing via cache-enabled sparse diffusion inference. arXiv preprint arXiv:2305.17423 (2023) +95. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858 (2023) +96. Zhang, M., Teck Ma, K., Hwee Lim, J., Zhao, Q., Feng, J.: Deep future gaze: Gaze anticipation on egocentric videos using adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4372-4381 (2017) + +97. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018) +98. Zhang, S., Yang, X., Feng, Y., Qin, C., Chen, C.C., Yu, N., Chen, Z., Wang, H., Savarese, S., Ermon, S., et al.: Hive: Harnessing human feedback for instructional visual editing. arXiv preprint arXiv:2303.09618 (2023) +99. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022) +100. Zhang, Y., Huang, N., Tang, F., Huang, H., Ma, C., Dong, W., Xu, C.: Inversion-based style transfer with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10146-10156 (2023) +101. Zhang, Z., Han, L., Ghosh, A., Metaxas, D.N., Ren, J.: Sine: Single image editing with text-to-image diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6027-6037 (2023) +102. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) \ No newline at end of file diff --git a/2024/LEGO_ Learning EGOcentric Action Frame Generation via Visual Instruction Tuning/images.zip b/2024/LEGO_ Learning EGOcentric Action Frame Generation via Visual Instruction Tuning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1a70dd2514ace145ef9869af9142510e35f48c9d --- /dev/null +++ b/2024/LEGO_ Learning EGOcentric Action Frame Generation via Visual Instruction Tuning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:130858f63de1dcc8ee8a826e1744bf57c3611eb398c77c74fa3d523c1416aaf5 +size 572116 diff --git a/2024/LEGO_ Learning EGOcentric Action Frame Generation via Visual Instruction Tuning/layout.json b/2024/LEGO_ Learning EGOcentric Action Frame Generation via Visual Instruction Tuning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4a25e77093a9b2ae861f11cc22f0e9c2397a9fb5 --- /dev/null +++ b/2024/LEGO_ Learning EGOcentric Action Frame Generation via Visual Instruction Tuning/layout.json @@ -0,0 +1,12255 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "spans": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "type": "text", + "content": "LEGO: Learning EGOcentric Action Frame Generation via Visual Instruction Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "spans": [ + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "text", + "content": "Bolin Lai" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "inline_equation", + "content": "^{1,2,\\dagger}" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "text", + "content": ", Xiaoliang Dai" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "text", + "content": ", Lawrence Chen" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "text", + "content": ", Guan Pang" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "text", + "content": ", James M. Rehg" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "text", + "content": ", and Miao Liu" + }, + { + "bbox": [ + 168, + 167, + 445, + 193 + ], + "type": "inline_equation", + "content": "^{1,*}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 275, + 201, + 339, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 201, + 339, + 212 + ], + "spans": [ + { + "bbox": [ + 275, + 201, + 339, + 212 + ], + "type": "text", + "content": "1 GenAI, Meta" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 237, + 213, + 376, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 213, + 376, + 224 + ], + "spans": [ + { + "bbox": [ + 237, + 213, + 376, + 224 + ], + "type": "text", + "content": "2 Georgia Institute of Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 219, + 224, + 394, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 224, + 394, + 236 + ], + "spans": [ + { + "bbox": [ + 219, + 224, + 394, + 236 + ], + "type": "text", + "content": "3 University of Illinois Urbana-Champaign" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 236, + 477, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 236, + 477, + 258 + ], + "spans": [ + { + "bbox": [ + 136, + 236, + 477, + 258 + ], + "type": "text", + "content": "bolin.lai@gatech.edu {xiaoliangdai,lawrencechen,gpang,miaoliu}@meta.com jrehg@illinois.edu" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 138, + 304, + 162, + 323 + ], + "blocks": [ + { + "bbox": [ + 138, + 304, + 162, + 323 + ], + "lines": [ + { + "bbox": [ + 138, + 304, + 162, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 162, + 323 + ], + "type": "image", + "image_path": "bbb63d72c539697167526da88405b46948a15f31e23638eca417d8266b74387d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 180, + 280, + 228, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 280, + 228, + 289 + ], + "spans": [ + { + "bbox": [ + 180, + 280, + 228, + 289 + ], + "type": "text", + "content": "User Prompt" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 172, + 293, + 238, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 293, + 238, + 332 + ], + "spans": [ + { + "bbox": [ + 172, + 293, + 238, + 332 + ], + "type": "text", + "content": "\"Can you provide instructions on how to wash the trouser with the brush in my current situation?\"" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 242, + 278, + 317, + 336 + ], + "blocks": [ + { + "bbox": [ + 242, + 278, + 317, + 336 + ], + "lines": [ + { + "bbox": [ + 242, + 278, + 317, + 336 + ], + "spans": [ + { + "bbox": [ + 242, + 278, + 317, + 336 + ], + "type": "image", + "image_path": "6a7072f59eb18a44aaafacb5ed0a0cc5fd2b54ea97b19cf1a992fa6095a0ef5f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 344, + 211, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 344, + 211, + 384 + ], + "spans": [ + { + "bbox": [ + 141, + 344, + 211, + 384 + ], + "type": "text", + "content": "1. Check the Care Label: Check the care label on your trousers for any specific washing instructions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 392, + 208, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 392, + 208, + 424 + ], + "spans": [ + { + "bbox": [ + 141, + 392, + 208, + 424 + ], + "type": "text", + "content": "2. Prep the Area: Find a clean and flat surface to work on, such as a table or countertop." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 432, + 206, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 432, + 206, + 450 + ], + "spans": [ + { + "bbox": [ + 141, + 432, + 206, + 450 + ], + "type": "text", + "content": "3. Spot Check Stains:\" :" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 459, + 208, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 459, + 208, + 468 + ], + "spans": [ + { + "bbox": [ + 141, + 459, + 208, + 468 + ], + "type": "text", + "content": "(a) LLM Response" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 218, + 435, + 237, + 454 + ], + "blocks": [ + { + "bbox": [ + 218, + 435, + 237, + 454 + ], + "lines": [ + { + "bbox": [ + 218, + 435, + 237, + 454 + ], + "spans": [ + { + "bbox": [ + 218, + 435, + 237, + 454 + ], + "type": "image", + "image_path": "86fe4e4caa66be29e7ca78900cdab603d56c5686b89b32eb4a4e8c9b43c83415.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 251, + 346, + 320, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 346, + 320, + 362 + ], + "spans": [ + { + "bbox": [ + 251, + 346, + 320, + 362 + ], + "type": "text", + "content": "1. You should submerge the trouser in the water." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 250, + 370, + 320, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 370, + 320, + 410 + ], + "spans": [ + { + "bbox": [ + 250, + 370, + 320, + 410 + ], + "type": "text", + "content": "2. Use the brush to scrub the trouser; focusing on any stains or areas that may require extra attention." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 251, + 418, + 315, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 418, + 315, + 441 + ], + "spans": [ + { + "bbox": [ + 251, + 418, + 315, + 441 + ], + "type": "text", + "content": "3. Once the trouser is clean, you should rinse it.\"" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 283, + 444, + 288, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 283, + 444, + 288, + 452 + ], + "spans": [ + { + "bbox": [ + 283, + 444, + 288, + 452 + ], + "type": "text", + "content": "中" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 247, + 459, + 340, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 459, + 340, + 468 + ], + "spans": [ + { + "bbox": [ + 247, + 459, + 340, + 468 + ], + "type": "text", + "content": "(b) Visual LLM Response" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 330, + 277, + 477, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 277, + 477, + 287 + ], + "spans": [ + { + "bbox": [ + 330, + 277, + 477, + 287 + ], + "type": "text", + "content": "Actions in Existing Egocentric Datasets" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 340, + 289, + 455, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 289, + 455, + 347 + ], + "spans": [ + { + "bbox": [ + 340, + 289, + 455, + 347 + ], + "type": "text", + "content": "tackle open the tap \npuncture the sandal \nrub oil into dough place mug \nturn the mothendose sew the cloth doll \nwash trouser with brush \nsqueeze ball of dough \nclean a sieve shake the spray adhesive \nclean rice adjust gasattach the wire \nplace spice \nopen fridge pour the topsoil \nclean place ball" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 328, + 436, + 346, + 455 + ], + "blocks": [ + { + "bbox": [ + 328, + 436, + 346, + 455 + ], + "lines": [ + { + "bbox": [ + 328, + 436, + 346, + 455 + ], + "spans": [ + { + "bbox": [ + 328, + 436, + 346, + 455 + ], + "type": "image", + "image_path": "be2629e345d61ded979010eb6eee5b270c97ba36f8e7287d4a4e852ba4a40d16.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 358, + 365, + 460, + 441 + ], + "blocks": [ + { + "bbox": [ + 358, + 365, + 460, + 441 + ], + "lines": [ + { + "bbox": [ + 358, + 365, + 460, + 441 + ], + "spans": [ + { + "bbox": [ + 358, + 365, + 460, + 441 + ], + "type": "image", + "image_path": "acbd7ecd6c89a1bea1d9fd71317a8766d6c4895ff647dce857a47ca74f11d78b.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 463, + 430, + 480, + 454 + ], + "blocks": [ + { + "bbox": [ + 463, + 430, + 480, + 454 + ], + "lines": [ + { + "bbox": [ + 463, + 430, + 480, + 454 + ], + "spans": [ + { + "bbox": [ + 463, + 430, + 480, + 454 + ], + "type": "image", + "image_path": "bb45d21632f838a2a5dcfc1602d35bb71675fb6991faa874ec6addc94934d2c7.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 358, + 459, + 477, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 459, + 477, + 468 + ], + "spans": [ + { + "bbox": [ + 358, + 459, + 477, + 468 + ], + "type": "text", + "content": "(c) Our model (LEGO) Response" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 130, + 479, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 578 + ], + "type": "text", + "content": "Fig. 1: When a user performing a complex task asks a large language model (LLM) for instructions (a) on how to complete the steps, she receives a generic answer and has to translate the guidance into her specific situation. If she is wearing a camera, then the prompt can be augmented with an egocentric view of the scene and passed to a Visual LLM (b), and the description is now contextualized to her situation. But she still faces the challenge of parsing a written description. When she uses our novel LEGO model (c), however, the combined image and prompt are used to automatically generate an image that provides visual guidance tailored to her exact situation from the egocentric viewpoint. Now she can complete her task seamlessly!" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 160, + 604, + 453, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 604, + 453, + 636 + ], + "spans": [ + { + "bbox": [ + 160, + 604, + 453, + 636 + ], + "type": "text", + "content": "Abstract. Generating instructional images of human daily actions from an egocentric viewpoint serves as a key step towards efficient skill transfer. In this paper, we introduce a novel problem - egocentric action" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 642, + 444, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 642, + 444, + 654 + ], + "spans": [ + { + "bbox": [ + 141, + 642, + 444, + 654 + ], + "type": "text", + "content": "This work was done when the first author was an intern at GenAI, Meta." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 143, + 654, + 239, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 654, + 239, + 665 + ], + "spans": [ + { + "bbox": [ + 143, + 654, + 239, + 665 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 160, + 117, + 455, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 117, + 455, + 327 + ], + "spans": [ + { + "bbox": [ + 160, + 117, + 455, + 327 + ], + "type": "text", + "content": "frame generation. The goal is to synthesize an image depicting an action in the user's context (i.e., action frame) by conditioning on a user prompt and an input egocentric image. Notably, existing egocentric action datasets lack the detailed annotations that describe the execution of actions. Additionally, existing diffusion-based image manipulation models are sub-optimal in controlling the state transition of an action in egocentric image pixel space because of the domain gap. To this end, we propose to Learn EGOcentric (LEGO) action frame generation via visual instruction tuning. First, we introduce a prompt enhancement scheme to generate enriched action descriptions from a visual large language model (VLLM) by visual instruction tuning. Then we propose a novel method to leverage image and text embeddings from the VLLM as additional conditioning to improve the performance of a diffusion model. We validate our model on two egocentric datasets - Ego4D and Epic-Kitchens. Our experiments show substantial improvement over prior image manipulation models in both quantitative and qualitative evaluation. We also conduct detailed ablation studies and analysis to provide insights in our method. More details of the dataset and code are available on the website (https://bolinlai.github.io/Lego_EgoActGen/)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 335, + 447, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 335, + 447, + 346 + ], + "spans": [ + { + "bbox": [ + 160, + 335, + 447, + 346 + ], + "type": "text", + "content": "Keywords: Egocentric Vision " + }, + { + "bbox": [ + 160, + 335, + 447, + 346 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 335, + 447, + 346 + ], + "type": "text", + "content": " Instruction Tuning " + }, + { + "bbox": [ + 160, + 335, + 447, + 346 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 335, + 447, + 346 + ], + "type": "text", + "content": " Diffusion Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 366, + 230, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 366, + 230, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 366, + 230, + 380 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 390, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 594 + ], + "type": "text", + "content": "The emergence of Large Language Models (LLMs) [6, 10, 64, 99], such as ChatGPT, has revolutionized the transfer of knowledge. However, an LLM alone is not a sufficient tool for human skill transfer. Consider the question answering example in Fig. 1(a). The LLM can summarize general world knowledge, but its response may not be directly applicable to the user's current circumstances. To bridge this gap, egocentric visual perception provides a novel means to capture the actions and intentions as well as the surrounding context of the camera wearer. As shown in Fig. 1(b), recent Visual Large Language Models (VLLMs) [1, 12, 46, 52, 102] can generate instructions based on the egocentric visual input. However, such verbose textual instructions are not the optimal medium for enabling efficient human skill transfer (e.g., via AR devices). Neuroscience studies have revealed that the human brain processes text much more slowly than images [4], and that humans can interpret an action from a single static image [22]. Motivated by these discoveries, we seek to design an image generation architecture that can synthesize an image which not only vividly depicts how an action should be conducted, but also seamlessly aligns with the user's visual perspective." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "type": "text", + "content": "Formally, we introduce the novel problem of egocentric action frame generation as depicted in Fig. 1(c). Given a user query about how to perform a specific action and an egocentric image capturing the moment before the action begins, the goal is to synthesize an egocentric image illustrating the execution of the action in the same egocentric context. We address this problem by harnessing diffusion models [27,67], which have been shown to be powerful tools for image" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "B. Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "manipulation [5,25,44,56,100]. There are two major challenges in using diffusion models to generate action frames from an egocentric perspective. First, the action annotations of the existing egocentric datasets [13, 21] are simply composed of a verb and nouns (see word cloud in Fig. 1), and thus lack the necessary details for diffusion models to learn the action state transition and to associate the action with correct objects and body movements. Second, the existing diffusion models are pre-trained primarily on exocentric (third-person-view) images, and have limited ability to represent complicated human daily activities from an egocentric perspective. In contrast, our proposed problem requires image generation in the egocentric view, conditioned on the user prompt of actions. The resulting domain gap impedes existing methods from synthesizing faithful and consistent egocentric action frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 261, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 403 + ], + "type": "text", + "content": "To address these challenges, we propose to Learn EGOcentric (LEGO) action frame generation with visual instruction tuning. First, we introduce a prompt enhancement scheme to generate enriched action descriptions at scale from an instruction-tuned VLLM. Second, we incorporate the image and text embeddings from finetuned VLLM as additional conditioning into the diffusion model to narrow the domain gap and improve the controllability of action frame generation. Our experimental results suggest that the enriched action descriptions and our innovative utilization of VLLM embeddings both improve the image generation performance. Our model is able to provide a generated key action frame together with a detailed action descriptions to facilitate human skill transfer from the egocentric perspective. Overall, our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 415, + 480, + 594 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 415, + 479, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 479, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 479, + 449 + ], + "type": "text", + "content": "- We introduce the novel problem of egocentric action frame generation to facilitate the process of skill transfer and address the challenges of missing action details in prompts and domain gap in existing image diffusion models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 451, + 480, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 451, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 139, + 451, + 480, + 498 + ], + "type": "text", + "content": "- We propose a prompt enhancement strategy based on visual instruction tuning to enrich egocentric action descriptions, and demonstrate how the enriched descriptions can help the diffusion model understand the action state transition from the egocentric perspective." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 499, + 479, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 499, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 499, + 479, + 533 + ], + "type": "text", + "content": "- We propose a novel approach to incorporate the text and visual embeddings from the VLLM into the latent diffusion model to bridge the domain gap and improve the performance for egocentric action frame generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 139, + 535, + 480, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 535, + 480, + 594 + ], + "spans": [ + { + "bbox": [ + 139, + 535, + 480, + 594 + ], + "type": "text", + "content": "- We conduct thorough experiments on the Ego4D and Epic-Kitchens datasets to validate the superiority of our model over prior approaches. We also showcase the contribution of each component of our model design through ablation studies. We further provide analysis on how the visual instruction-tuned embeddings benefit model performance." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 614, + 237, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 614, + 237, + 627 + ], + "spans": [ + { + "bbox": [ + 132, + 614, + 237, + 627 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Text-Guided Image Manipulation. The recent emergence of diffusion models enables text-guided image manipulation including image restoration [33], style" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "text", + "content": "Learning Egocentric Action Frame Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 320 + ], + "type": "text", + "content": "transfer [75], personalized image synthesis [68,71,87], pose generation [28,36] and generic image editing [15,18,34,38,44,57,60,79,82,82-84,94,98,101]. SDEdit [56] converts the image to the latent space by adding noise through a stochastic differential equation and then denoises the representation for image editing. Rombach et al. [67] further expand SDEdit from the original stroke-based editing to text-based editing. Null-text inversion (NTI) [58] inverts a real image by DDIM [35] to yield the diffusion process and then reconstructs the image. The image can then be edited following the same strategies as Prompt-to-Prompt [25]. NTI relies on accurate inversion process which can be improved by using coupled transformations [80] or proximal guidance [24] and accelerated by a blended guidance strategy [61]. To associate the nouns with correct objects, DIFFEDIT [11] generates a mask to localize the manipulated regions. However, most inversion-based methods require accurate image captions, which are largely unavailable in the existing egocentric dataset. Recently, InstructPix2Pix [5] demonstrates the potential to edit a real image without the inversion process and original captions. However, how to leverage the diffusion model to control the state transition of an action within the egocentric image plane remains unexplored." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 324, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 529 + ], + "type": "text", + "content": "Large Language Model for Image Generation. LLMs [6, 10, 64, 76, 78, 99] and VLLMs [1, 12, 23, 46, 52, 73, 95, 102] have shown their strong capability of understanding complex human instructions. Recently, LLMs and VLLMs are used to guide image generation [2, 7-9, 51, 93, 102]. Wen et al. [88] use a pretrained VLLM to pinpoint the misalignment of the text prompt and the synthetic image and then correct it using the diffusion model. Lian et al. [49] propose to generate a layout map using an LLM to improve the understanding of prompts with spatial and numerical reasoning. InstructEdit [83] uses BLIP-2 [46] to infer the objects that will be edited and then generates a mask with SAM [39] for object grounding. A pre-trained LLM can also be used as a controller to connect with various foundation models [70, 89, 90]. GILL [40] learns text-relevant image embeddings from VLLM in a few additional tokens for image generation. Importantly, all previous efforts apply the off-the-shelf foundational models directly to their problems without finetuning. In contrast, our method uses visual instruction tuning to improve the image and text embeddings from the VLLM, which narrows the domain gap and thereby facilitates the action frame generation from the egocentric point of view." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Egocentric Vision. Recent efforts seek to understand human actions and perceptual attention [16, 29, 30, 37, 41-43, 48, 69, 74, 86], model hand-object interactions [20, 53, 54, 65, 91], and estimate human body poses [45, 55, 77, 81] from the egocentric perspective. Here, we mainly discuss the most relevant works on egocentric visual-language models and egocentric visual content generation. Lin et al. [50] propose the first egocentric video-language pre-training model - EgoVLP. Pramanick et al. [62] further improve it by incorporating multi-modal fusion directly into the video and language backbone. Ashutosh et al. [3] propose to learn a hierarchical video-language embedding for long egocentric videos. Ramakrishnan et al. [66] introduce NaQ, which is a data augmentation strategy to train models for long egocentric video search with natural language queries. In terms" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "B. Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 176, + 129, + 284, + 190 + ], + "blocks": [ + { + "bbox": [ + 176, + 120, + 283, + 129 + ], + "lines": [ + { + "bbox": [ + 176, + 120, + 283, + 129 + ], + "spans": [ + { + "bbox": [ + 176, + 120, + 283, + 129 + ], + "type": "text", + "content": "Example for In-context Learning" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 176, + 129, + 284, + 190 + ], + "lines": [ + { + "bbox": [ + 176, + 129, + 284, + 190 + ], + "spans": [ + { + "bbox": [ + 176, + 129, + 284, + 190 + ], + "type": "image", + "image_path": "8994e926f2e10d991bc7e4bc4892d5ec209c1ae0ca2897970c19ed210bb51dbd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 194, + 200, + 202 + ], + "lines": [ + { + "bbox": [ + 161, + 194, + 200, + 202 + ], + "spans": [ + { + "bbox": [ + 161, + 194, + 200, + 202 + ], + "type": "text", + "content": "Input Query" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 160, + 203, + 297, + 250 + ], + "lines": [ + { + "bbox": [ + 160, + 203, + 297, + 250 + ], + "spans": [ + { + "bbox": [ + 160, + 203, + 297, + 250 + ], + "type": "text", + "content": "Action Label: \"chop end of onion\" Bounding Boxes: \"left hand-[0.203, 0.368, 0.499, 1.000], right hand-[0.530, 0.527, 0.797, 1.000], chopping board-[0.248, 0.156, 0.644, 0.750], onion-[0.462, 0.377, 0.509, 0.506], knife-[0.484, 0.343, 0.546, 0.586]\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 330, + 129, + 439, + 191 + ], + "blocks": [ + { + "bbox": [ + 159, + 254, + 298, + 286 + ], + "lines": [ + { + "bbox": [ + 159, + 254, + 298, + 286 + ], + "spans": [ + { + "bbox": [ + 159, + 254, + 298, + 286 + ], + "type": "text", + "content": "Action Description: \"The person presses the onion on the chopping board with the left hand and then cuts off the end of the onion with a knife in the right hand.\"" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 350, + 120, + 418, + 129 + ], + "lines": [ + { + "bbox": [ + 350, + 120, + 418, + 129 + ], + "spans": [ + { + "bbox": [ + 350, + 120, + 418, + 129 + ], + "type": "text", + "content": "Query and Response" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 330, + 129, + 439, + 191 + ], + "lines": [ + { + "bbox": [ + 330, + 129, + 439, + 191 + ], + "spans": [ + { + "bbox": [ + 330, + 129, + 439, + 191 + ], + "type": "image", + "image_path": "692f9f9ea4a6532001bf520451bdb128cea9028caef34ef5a6d582f10aeb9147.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 194, + 356, + 202 + ], + "lines": [ + { + "bbox": [ + 317, + 194, + 356, + 202 + ], + "spans": [ + { + "bbox": [ + 317, + 194, + 356, + 202 + ], + "type": "text", + "content": "Input Query" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 203, + 453, + 251 + ], + "lines": [ + { + "bbox": [ + 317, + 203, + 453, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 203, + 453, + 251 + ], + "type": "text", + "content": "Action Label: \"squidge into lunch box\" Bounding Boxes: \"left hand-[0.427, 0.799, 0.591, 1.00], right hand-[0.654, 0.949, 0.707, 1.00], lunch box-[0.390, 0.512, 0.696, 0.993], sink-[0.00, 0.802, 0.315, 1.00], container lid-[0.403, 0.454, 0.513, 0.635], fork-[0.589, 0.798, 0.672, 1.00]\"" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 315, + 254, + 457, + 286 + ], + "lines": [ + { + "bbox": [ + 315, + 254, + 457, + 286 + ], + "spans": [ + { + "bbox": [ + 315, + 254, + 457, + 286 + ], + "type": "text", + "content": "GPT-3.5 Response: \"The person uses their left hand to hold a lunch box, while their right hand uses a fork to squidge food into the lunch box, which is placed on the sink with the container lid nearby.\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 131, + 300, + 482, + 356 + ], + "lines": [ + { + "bbox": [ + 131, + 300, + 482, + 356 + ], + "spans": [ + { + "bbox": [ + 131, + 300, + 482, + 356 + ], + "type": "text", + "content": "Fig. 2: Examples for data curation using GPT-3.5. We provide detailed action descriptions of several example images as well as their action labels and bounding boxes for in-context learning. In addition, we input the action label and bounding boxes of another action as a query. GPT-3.5 is able to generate descriptions with enriched information (highlighted by underlines) in the response." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 380, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 464 + ], + "type": "text", + "content": "of egocentric visual generation, Jia et al. [32] leverage GANs [19] to generate future head motion in hand forecasting task. Zhang et al. [96] leverage GANs to facilitate future gaze anticipation. Ye et al. [92] propose the affordance diffusion model that takes in the image of an object and generates possible hand-object interactions in the egocentric view. In this paper, we present the first work that investigates how to leverage VLLMs and diffusion models to generate action state transition on the egocentric image plane." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 483, + 201, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 201, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 201, + 495 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 509, + 487, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 487, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 487, + 569 + ], + "type": "text", + "content": "The problem setting of egocentric action frame generation is illustrated in Fig. 1(c). Given an egocentric image frame " + }, + { + "bbox": [ + 130, + 509, + 487, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 509, + 487, + 569 + ], + "type": "text", + "content": " that captures the user's current visual context as well as a user query prompt " + }, + { + "bbox": [ + 130, + 509, + 487, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 509, + 487, + 569 + ], + "type": "text", + "content": " regarding how to perform an action, our goal is to synthesize the action frame " + }, + { + "bbox": [ + 130, + 509, + 487, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 130, + 509, + 487, + 569 + ], + "type": "text", + "content": " that visually depicts how the action should be conducted in the same situation (i.e., keep a consistent background)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "The key insight of our proposed LEGO model is leveraging the strong capability of a VLLM to enhance the diffusion model for egocentric action frame generation. The annotations of existing egocentric datasets do not describe the details of how actions are conducted. As a remedy, we leverage visual instruction tuning to finetune a VLLM that enriches action descriptions based on the egocentric visual prompt. In addition, the existing diffusion-based image manipulation models are limited in understanding egocentric action state transition, due to the domain gap between the exocentric pre-training dataset and the ego-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "text", + "content": "Learning Egocentric Action Frame Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 115, + 479, + 323 + ], + "blocks": [ + { + "bbox": [ + 136, + 115, + 479, + 323 + ], + "lines": [ + { + "bbox": [ + 136, + 115, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 479, + 323 + ], + "type": "image", + "image_path": "f8e36e3ed9cec7943cda7be08b211f7d4835e082ce1346fbfdcb6504d546c0f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 331, + 482, + 399 + ], + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 399 + ], + "type": "text", + "content": "Fig. 3: Overview of our proposed LEGO model. We first finetune a visual large language model (VLLM) to generate the enriched action description with visual instruction tuning. We then project image and text embeddings from the finetuned VLLM to the feature space of the latent diffusion model (LDM). Finally, we train the LDM to synthesize the egocentric action frame conditioning on the input frame, enriched action description, as well as the VLLM image and text embeddings." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 424, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 473 + ], + "type": "text", + "content": "centric action dataset for our problem. To bridge this gap, we propose a novel approach that leverages VLLM embeddings to control the state transition of actions and to generate action frames accordingly. We detail the VLLM-based data enrichment pipeline and our model design in the following sections." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 495, + 349, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 495, + 349, + 509 + ], + "spans": [ + { + "bbox": [ + 131, + 495, + 349, + 509 + ], + "type": "text", + "content": "3.1 Egocentric Visual Instruction Tuning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": "Data Curation for Visual Instruction Tuning. As shown in Fig. 2, we use GPT-3.5 to generate detailed action descriptions based on an input query composed of a short action label and object bounding boxes that are provided in the existing datasets. First, we prepare several examples of possible inputs along with their expected output descriptions for GPT-3.5 to perform in-context learning. These examples cover a diverse set of scenes in the egocentric action dataset. Each example is composed of an action label, a manually annotated action description, and relative spatial information of hands and objects-of-interests represented by bounding box coordinates. GPT-3.5 can learn from the given examples and generate similar detailed action descriptions based on a new input query. The resulting GPT-3.5 curated data is then further used for visual instruction tuning. More details of the prompt are provided in the supplementary." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "B. Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 483, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 483, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 483, + 236 + ], + "type": "text", + "content": "Visual Instruction Tuning. We follow the finetuning strategy in prior work [52] as shown in Fig. 3(a). Specifically, we use the pre-trained CLIP visual encoder [63] " + }, + { + "bbox": [ + 130, + 116, + 483, + 236 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 116, + 483, + 236 + ], + "type": "text", + "content": " to encode the visual representation and then apply a linear projection layer " + }, + { + "bbox": [ + 130, + 116, + 483, + 236 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 116, + 483, + 236 + ], + "type": "text", + "content": " to map the CLIP visual features into the semantic space of the LLM, i.e., " + }, + { + "bbox": [ + 130, + 116, + 483, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_i = \\tau (\\phi (\\mathcal{X}))" + }, + { + "bbox": [ + 130, + 116, + 483, + 236 + ], + "type": "text", + "content": ". To construct the user prompt " + }, + { + "bbox": [ + 130, + 116, + 483, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 116, + 483, + 236 + ], + "type": "text", + "content": ", we insert the action label annotation into a prompt template to create a coherent query that is aligned with the instruction-following nature of an LLM. We then tokenize " + }, + { + "bbox": [ + 130, + 116, + 483, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 116, + 483, + 236 + ], + "type": "text", + "content": ", and feed both the prompt text tokens and image tokens " + }, + { + "bbox": [ + 130, + 116, + 483, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_i" + }, + { + "bbox": [ + 130, + 116, + 483, + 236 + ], + "type": "text", + "content": " as inputs into the LLM. Finally, the LLM is trained to generate enriched action description (denoted as " + }, + { + "bbox": [ + 130, + 116, + 483, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 116, + 483, + 236 + ], + "type": "text", + "content": ") based on the user prompt and image input." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 274 + ], + "type": "text", + "content": "User Prompt Enrichment at Scale. Note that the visual instruction tuned VLLM doesn't rely on any object bounding boxes as input. Therefore, we can generate enriched action descriptions for all egocentric action data at scale." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 289, + 346, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 346, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 346, + 300 + ], + "type": "text", + "content": "3.2 Egocentric Action Frame Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 308, + 482, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 379 + ], + "type": "text", + "content": "We leverage a latent diffusion model (LDM) [67] to synthesize the action frame conditioning on the input frame and the detailed action description " + }, + { + "bbox": [ + 130, + 308, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 308, + 482, + 379 + ], + "type": "text", + "content": " generated by our finetuned VLLM (see Fig. 3(b)). Following the regular steps in LDMs [5, 67], the input image is first encoded into a latent space using a pre-trained autoencoder " + }, + { + "bbox": [ + 130, + 308, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 308, + 482, + 379 + ], + "type": "text", + "content": ". Then the input to the denoising UNet is a concatenation of the latent input representation " + }, + { + "bbox": [ + 130, + 308, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\mathcal{X})" + }, + { + "bbox": [ + 130, + 308, + 482, + 379 + ], + "type": "text", + "content": " and a Gaussian noise " + }, + { + "bbox": [ + 130, + 308, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 308, + 482, + 379 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 380, + 482, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 463 + ], + "type": "text", + "content": "Our key innovation is to design the U-Net conditioning component so that the diffusion model can interpret the egocentric actions correctly. To start, we follow [5] and adopt the conventional pre-trained CLIP text encoder " + }, + { + "bbox": [ + 130, + 380, + 482, + 463 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 130, + 380, + 482, + 463 + ], + "type": "text", + "content": " to extract a text representation of " + }, + { + "bbox": [ + 130, + 380, + 482, + 463 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 380, + 482, + 463 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 380, + 482, + 463 + ], + "type": "inline_equation", + "content": "\\psi(\\mathcal{R}) \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 130, + 380, + 482, + 463 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 380, + 482, + 463 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 380, + 482, + 463 + ], + "type": "text", + "content": " is the maximum number of text tokens and " + }, + { + "bbox": [ + 130, + 380, + 482, + 463 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 380, + 482, + 463 + ], + "type": "text", + "content": " is the number of feature channels. We further leverage the image and text embeddings from the visual instruction tuned VLLM as additional LDM conditioning to alleviate the domain gap issue." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "text", + "content": "Specifically, we feed the VLLM image embedding " + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_i" + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "text", + "content": " into an extra linear layer " + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "text", + "content": " to map it to LDM feature space, i.e., " + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\sigma(\\mathcal{H}_i) \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "text", + "content": " is the number of image tokens. Note that " + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_i" + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "text", + "content": " is already projected to the semantic space during visual instruction tuning, and therefore differs from the image embedding " + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\mathcal{X})" + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "text", + "content": " from the auto-encoder. Moreover, we also extract the text embedding " + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_t" + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "text", + "content": " before the last linear layer of LLM. We adopt a fixed token number " + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "text", + "content": " and enforce padding or truncation behavior, as in the CLIP text encoder. The text embedding is then fed to a projection layer " + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "text", + "content": ". In LLM decoder, the response is generated iteratively and each word embedding only conditions on the context ahead of it. To extract the holistic semantic meaning of " + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_t" + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "text", + "content": " in LDM feature space, we further add self-attention layers " + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "text", + "content": " after the projection, i.e., " + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\pi(\\mu(\\mathcal{H}_t)) \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 130, + 463, + 482, + 606 + ], + "type": "text", + "content": ". Thus, the U-Net conditioning can be formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 210, + 615, + 481, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 615, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 210, + 615, + 481, + 635 + ], + "type": "interline_equation", + "content": "\\mathcal {C} = \\left[ \\psi (\\mathcal {R}), \\sigma \\left(\\mathcal {H} _ {i}\\right), \\pi \\left(\\mu \\left(\\mathcal {H} _ {t}\\right)\\right) \\right] \\in \\mathbb {R} ^ {(2 N + M) \\times D}. \\tag {1}", + "image_path": "287fbea14a2756d8c28493107b167fcf53e78946798ac6a9e7fa71728f17a1b5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "The conditioning " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " is fed into the denoising UNet at multiple layers via the cross-attention mechanism [67]. We assume the intermediate feature of a specific" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 258, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 258, + 91, + 448, + 103 + ], + "type": "text", + "content": "Learning Egocentric Action Frame Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "UNet layer is " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": ", which is learned from the UNet input (i.e., the concatenation of input frame representation " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\mathcal{X})" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " and Gaussian noise " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "). The cross-attention at this UNet layer can be formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 211, + 161, + 481, + 188 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 161, + 481, + 188 + ], + "spans": [ + { + "bbox": [ + 211, + 161, + 481, + 188 + ], + "type": "interline_equation", + "content": "C r o s s A t t (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {D}}\\right) \\cdot V, \\tag {2}", + "image_path": "07cb4fe5388b07403e2208538d33aa5c406039c366ad66c4e7fd7bf804d05225.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 195, + 482, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 195, + 482, + 244 + ], + "spans": [ + { + "bbox": [ + 130, + 195, + 482, + 244 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 195, + 482, + 244 + ], + "type": "inline_equation", + "content": "Q = W_{Q} \\cdot \\mathcal{U}" + }, + { + "bbox": [ + 130, + 195, + 482, + 244 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 195, + 482, + 244 + ], + "type": "inline_equation", + "content": "K = W_{K} \\cdot \\mathcal{C}" + }, + { + "bbox": [ + 130, + 195, + 482, + 244 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 195, + 482, + 244 + ], + "type": "inline_equation", + "content": "V = W_{V} \\cdot \\mathcal{C}" + }, + { + "bbox": [ + 130, + 195, + 482, + 244 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 130, + 195, + 482, + 244 + ], + "type": "inline_equation", + "content": "W_{Q}, W_{K}" + }, + { + "bbox": [ + 130, + 195, + 482, + 244 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 195, + 482, + 244 + ], + "type": "inline_equation", + "content": "W_{V}" + }, + { + "bbox": [ + 130, + 195, + 482, + 244 + ], + "type": "text", + "content": " are learnable matrices. We also adopt the classifier-free guidance following [5] (see supplementary for details). Finally, the UNet output is converted to the image domain by a pre-trained decoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 262, + 280, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 262, + 280, + 274 + ], + "spans": [ + { + "bbox": [ + 131, + 262, + 280, + 274 + ], + "type": "text", + "content": "3.3 Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 281, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 413 + ], + "type": "text", + "content": "All parameters of the VLLM are initialized from the pre-trained LLaVA [52] weights. During training, we freeze the CLIP image encoder and finetune the projection layer and LLM with cross-entropy loss for 3 epochs. To improve the diversity of the text prompts, we randomly select the question template from 10 candidates in each iteration. For LDM training, the text encoder, UNet and autoencoder are initialized with pre-trained weights [67]. The projection layers " + }, + { + "bbox": [ + 130, + 281, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 281, + 482, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 281, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 281, + 482, + 413 + ], + "type": "text", + "content": " and the self-attention layers " + }, + { + "bbox": [ + 130, + 281, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 281, + 482, + 413 + ], + "type": "text", + "content": " are initialized using the Xavier algorithm [17]. The text encoder is frozen and the remaining weights are finetuned with L2 regression loss between the predicted noise and real noise for 20,000 iterations. All input and target images are resized to a resolution of " + }, + { + "bbox": [ + 130, + 281, + 482, + 413 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 130, + 281, + 482, + 413 + ], + "type": "text", + "content": ". Please refer to the supplementary for more details of training and inference." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 432, + 230, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 432, + 230, + 445 + ], + "spans": [ + { + "bbox": [ + 131, + 432, + 230, + 445 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 457, + 250, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 457, + 250, + 468 + ], + "spans": [ + { + "bbox": [ + 131, + 457, + 250, + 468 + ], + "type": "text", + "content": "4.1 Data and Metrics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "text", + "content": "Datasets. We conduct our experiments on two well-established egocentric action datasets - Ego4D [21] and Epic-Kitchens-100 [13]. Both datasets were densely annotated with action starting time " + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "text", + "content": " and ending time " + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\hat{t}" + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "text", + "content": ". In our problem setting, we select an egocentric image frame " + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "text", + "content": " seconds before the action begins as the input " + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "text", + "content": ", and an image " + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\delta_{o}" + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "text", + "content": " seconds after the action begins as the target frame " + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "text", + "content": ". On the Ego4D dataset, based on the annotations of Pre-Condition-15 time (PRE-15) " + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "inline_equation", + "content": "t_{pre}" + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "text", + "content": ", and Point-of-No-Return time (PNR) " + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "inline_equation", + "content": "t_{pnr}" + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "text", + "content": ", we set " + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\delta_{i} = t - t_{pre}" + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\delta_{o} = t_{pnr} - t" + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "text", + "content": ". For Epic-Kitchens, PNR and PRE-15 annotations are not available. Instead, we empirically select " + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\delta_{i} = 0.25" + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "text", + "content": " seconds and " + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\delta_{o} = t + \\lambda * (\\hat{t} - t)" + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\lambda = 0.6" + }, + { + "bbox": [ + 130, + 478, + 482, + 609 + ], + "type": "text", + "content": ", for our experiments. More details of data preparation and improvement are elaborated in the supplementary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 609, + 482, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 482, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 482, + 634 + ], + "type": "text", + "content": "Metrics. We adopt image-to-image similarity, image-to-text similarity, and user study as metrics in our experiments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 641, + 482, + 666 + ], + "type": "text", + "content": "- Image-to-Image Metrics. We implement six metrics to evaluate image-to-image similarity. To begin with, we adopt three contrastive learning based" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "B. Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 148, + 115, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 115, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 148, + 115, + 482, + 319 + ], + "type": "text", + "content": "metrics including image-to-image (1) EgoVLP score [50], (2) EgoVLP" + }, + { + "bbox": [ + 148, + 115, + 482, + 319 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 148, + 115, + 482, + 319 + ], + "type": "text", + "content": " score [50] and (3) CLIP score [63]. EgoVLP is a contrastive video-language pretraining model trained with egocentric videos. Since EgoVLP takes multiple frames as input, we consider two types of inputs: duplicating the output frame as a static input (i.e., EgoVLP score) and combining the input frame with the output frame (i.e., EgoVLP" + }, + { + "bbox": [ + 148, + 115, + 482, + 319 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 148, + 115, + 482, + 319 + ], + "type": "text", + "content": " score). As a result, EgoVLP" + }, + { + "bbox": [ + 148, + 115, + 482, + 319 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 148, + 115, + 482, + 319 + ], + "type": "text", + "content": " can effectively measure whether the generated frame can depict the state transition of an action. Importantly, given that EgoVLP is pre-trained on egocentric data and action labels, we consider EgoVLP and EgoVLP" + }, + { + "bbox": [ + 148, + 115, + 482, + 319 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 148, + 115, + 482, + 319 + ], + "type": "text", + "content": " score as the primary automatic metrics. In addition, we also report (4) Fréchet Inception Distance (FID) [26], (5) Peak Signal-to-Noise Ratio (PSNR) and (6) Learned Perceptual Image Patch Similarity (LPIPS) [97] (with SqueezeNet [31] as the encoder) to make a thorough evaluation. Note that instead of measuring similarity of input and output frames as in prior works [12,80], in our problem setting, we measure the similarity between the generated action frame and ground truth, which better reflects whether the generation results can illustrate the execution of an action." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 139, + 321, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 321, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 139, + 321, + 482, + 430 + ], + "type": "text", + "content": "- Image-to-Text Metrics. We find the widely-used image-to-text CLIP score can not align actions with egocentric images due to the domain gap. Similar misalignment problem is also observed in [14, 59, 72, 85]. In our experiments, we utilize BLIP [47] to generate image captions of the output images and then calculate text-to-text similarity using CLIP text encoder (following [34]). We implement this metric with two BLIP structures: BLIP-B and BLIP-L. Though this solution may still suffer from the same domain gap issue, it is a more appropriate evaluation metric in our problem setting. See more evidence and discussions in the supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 430, + 482, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 430, + 482, + 551 + ], + "spans": [ + { + "bbox": [ + 139, + 430, + 482, + 551 + ], + "type": "text", + "content": "- **User Study.** We also conduct a user study on a subset of test data to further validate the advantage of our model based on human preference. We sample 60 examples from each dataset and present the generated frames from our model as well as the baseline models to raters on Amazon Mechanical Turk (AMT). We also provide the input frames and the corresponding action labels during evaluation. For each instance, we ask the raters to select the image that best aligns with the provided action label while preserving the most contextual information from the input frame. To minimize potential bias, we hire 5 AMT raters to annotate each example thus resulting in 300 samples for each dataset. User study interface is shown in supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 582, + 339, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 339, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 339, + 594 + ], + "type": "text", + "content": "4.2 Comparison with Prior Approaches" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "We compare our proposed model with previous state-of-the-art methods for text-guided image manipulation, including ProxEdit [24], SDEdit [56] and InstructPix2Pix (IP2P) [5]. For a fair comparison, we finetune these baseline methods with the same data used in our experiments. Specifically, we train IP2P in an end-to-end way on the two datasets with existing egocentric action labels as" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "text", + "content": "Learning Egocentric Action Frame Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 157, + 470, + 276 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 1: Comparison with prior image manipulation approaches in image-to-image metrics. " + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": " means a lower score in this metric suggests a better performance. The best results are highlighted with **boldface**. The orange row refers to our LEGO model." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 157, + 470, + 276 + ], + "lines": [ + { + "bbox": [ + 143, + 157, + 470, + 276 + ], + "spans": [ + { + "bbox": [ + 143, + 157, + 470, + 276 + ], + "type": "table", + "html": "
MethodsEgoVLPEgoVLP+CLIPFID ↓PSNRLPIPS ↓
Ego4DProxEdit [24]44.5172.6868.1733.0111.8840.90
SDEdit [56]50.0772.9073.3533.3511.8141.60
IP2P [5]62.1978.8478.7524.7312.1637.16
LEGO65.6580.4480.6123.8312.2936.43
EK-100ProxEdit [24]32.2752.7765.8051.3511.0646.35
SDEdit [56]33.8456.8074.7627.4111.3043.33
IP2P [5]42.9761.0677.0320.6411.2340.82
LEGO45.8962.6678.6321.5711.3340.36
", + "image_path": "19d75c7ae80be6358c559b982c27b84ced2d11471693e2064330a3193ac15429.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 143, + 349, + 313, + 415 + ], + "blocks": [ + { + "bbox": [ + 140, + 296, + 316, + 340 + ], + "lines": [ + { + "bbox": [ + 140, + 296, + 316, + 340 + ], + "spans": [ + { + "bbox": [ + 140, + 296, + 316, + 340 + ], + "type": "text", + "content": "Table 2: Image-to-text metrics of our model and baselines. The best results are highlighted with boldface. The orange row refers to our LEGO model performance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 349, + 313, + 415 + ], + "lines": [ + { + "bbox": [ + 143, + 349, + 313, + 415 + ], + "spans": [ + { + "bbox": [ + 143, + 349, + 313, + 415 + ], + "type": "table", + "html": "
MethodsEgo4DEpic-Kitchens
BLIP-BBLIP-LBLIP-BBLIP-L
ProxEdit [24]17.7317.3523.6523.39
SDEdit [56]19.8019.7421.5121.30
IP2P [5]20.0020.5625.3726.36
LEGO20.3820.7026.9827.41
", + "image_path": "8644cec964d8578ff76f5b9720226022b5bbe3b385d63f00de313a16cb1e94e1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 318, + 287, + 469, + 382 + ], + "blocks": [ + { + "bbox": [ + 318, + 287, + 469, + 382 + ], + "lines": [ + { + "bbox": [ + 318, + 287, + 469, + 382 + ], + "spans": [ + { + "bbox": [ + 318, + 287, + 469, + 382 + ], + "type": "image", + "image_path": "3d54ba3028473f279513c0b8fc8c34ecbb6c6d9a903308e97aa675438364ae3e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 390, + 473, + 424 + ], + "lines": [ + { + "bbox": [ + 313, + 390, + 473, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 473, + 424 + ], + "type": "text", + "content": "Fig.4: User study of our model and baselines. Win rate is the percentage of each model being picked as the best." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 533 + ], + "type": "text", + "content": "text conditioning. ProxEdit and SDEdit rely on the off-the-shelf latent diffusion model to synthesize edited images and thus can not be trained end to end. Therefore, we first finetune the latent diffusion model with egocentric images and action labels and then use the finetuned latent diffusion model parameters to implement ProxEdit and SDEdit approaches. Please refer to the supplementary for more baseline implementation details and a thorough comparison with prior methods that leverage LLMs for image generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 640 + ], + "type": "text", + "content": "In terms of image-to-image metrics in Tab. 1, both ProxEdit and SDEdit perform poorly on this novel problem, suggesting the challenging nature of generating egocentric action frames. IP2P performs much better in almost all metrics by end-to-end training and serves as the strongest baseline model in our experiments. However, our LEGO model consistently outperforms IP2P by a large margin (3.46%, 1.60%, 1.86%, 0.90, 0.13 and 0.73% respectively) in all six metrics on Ego4D. LEGO also exceeds IP2P notably (2.92%, 1.60%, 1.60%, 0.10 and 0.46% respectively) in five metrics on Epic-Kitchens. Though, LEGO slightly lags behind IP2P in FID score, it still achieves the second best performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "With regard to image-to-text metrics in Tab. 2, LEGO outperforms the strongest baseline (IP2P) by " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.38\\%" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.14\\%" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " on Ego4D and by " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.61\\%" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "B. Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 147, + 114, + 467, + 441 + ], + "blocks": [ + { + "bbox": [ + 147, + 114, + 467, + 441 + ], + "lines": [ + { + "bbox": [ + 147, + 114, + 467, + 441 + ], + "spans": [ + { + "bbox": [ + 147, + 114, + 467, + 441 + ], + "type": "image", + "image_path": "90f1142b2076eb8b9cd07b80ab0d511b9e20304e25902745a3b4208f97c799d7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 449, + 482, + 483 + ], + "lines": [ + { + "bbox": [ + 132, + 449, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 482, + 483 + ], + "type": "text", + "content": "Fig. 5: Visualization of the proposed LEGO model and all baselines on Ego4D (the first three rows) and Epic-Kitchens (the last three rows). The action frames generated by LEGO align with the user prompt better and preserve more contexts in input frames." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 509, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 616 + ], + "type": "inline_equation", + "content": "1.05\\%" + }, + { + "bbox": [ + 130, + 509, + 482, + 616 + ], + "type": "text", + "content": " on Epic-Kitchens. The result suggests the synthetic frames from our model can better align with the action descriptions. However, the performance gain is rather limited. We emphasize that this is because the domain gap still exists for the BLIP model. Besides baseline models, we also measure the image-to-text similarity of input frames and action prompts to validate models' capability of learning action state transition in semantics. The BLIP-B/BLIP-L scores are " + }, + { + "bbox": [ + 130, + 509, + 482, + 616 + ], + "type": "inline_equation", + "content": "15.44\\% / 15.49\\%" + }, + { + "bbox": [ + 130, + 509, + 482, + 616 + ], + "type": "text", + "content": " on Ego4D and " + }, + { + "bbox": [ + 130, + 509, + 482, + 616 + ], + "type": "inline_equation", + "content": "20.11\\% / 20.52\\%" + }, + { + "bbox": [ + 130, + 509, + 482, + 616 + ], + "type": "text", + "content": " on Epic-Kitchens, lagging behind all baseline models. The result evidences diffusion models are able to understand actions and edit the input frame semantically towards the action prompt." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Due to the limitation of automatic metrics, we additionally implement user study as a complement. We shuffle the order of the results from our model and the baselines while presenting them to the raters. We define the win rate as the percentage of each model being picked as the best out of the total 300" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "text", + "content": "Learning Egocentric Action Frame Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 159, + 137 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 159, + 137 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 159, + 137 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 159, + 137 + ], + "type": "image", + "image_path": "b816b15b296273af7f7290d15dac4041a1a449993d575d8632bb0ed54e0a83ed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 138, + 156, + 154 + ], + "lines": [ + { + "bbox": [ + 136, + 138, + 156, + 154 + ], + "spans": [ + { + "bbox": [ + 136, + 138, + 156, + 154 + ], + "type": "text", + "content": "Input Frame" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 161, + 155, + 182 + ], + "blocks": [ + { + "bbox": [ + 138, + 161, + 155, + 182 + ], + "lines": [ + { + "bbox": [ + 138, + 161, + 155, + 182 + ], + "spans": [ + { + "bbox": [ + 138, + 161, + 155, + 182 + ], + "type": "image", + "image_path": "9f974032032c25ffc259cd1a20f7a7784ebcfa6b9a999132fead39093da75d10.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 186, + 156, + 194 + ], + "lines": [ + { + "bbox": [ + 134, + 186, + 156, + 194 + ], + "spans": [ + { + "bbox": [ + 134, + 186, + 156, + 194 + ], + "type": "text", + "content": "LEGO" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 162, + 115, + 268, + 195 + ], + "blocks": [ + { + "bbox": [ + 162, + 115, + 268, + 195 + ], + "lines": [ + { + "bbox": [ + 162, + 115, + 268, + 195 + ], + "spans": [ + { + "bbox": [ + 162, + 115, + 268, + 195 + ], + "type": "image", + "image_path": "51390262476acd0907430caf3bce304be445d5e58f5034221078d342413828d6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 198, + 266, + 215 + ], + "lines": [ + { + "bbox": [ + 148, + 198, + 266, + 215 + ], + "spans": [ + { + "bbox": [ + 148, + 198, + 266, + 215 + ], + "type": "text", + "content": "cut a portion of clay brush a wood mix with both hands with a brush" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 269, + 115, + 339, + 195 + ], + "blocks": [ + { + "bbox": [ + 269, + 115, + 339, + 195 + ], + "lines": [ + { + "bbox": [ + 269, + 115, + 339, + 195 + ], + "spans": [ + { + "bbox": [ + 269, + 115, + 339, + 195 + ], + "type": "image", + "image_path": "69dd1a6c4d01d0466fabc1e3a5dc2d909b01917a994ef16efa11f9be47f9c6f1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 198, + 335, + 207 + ], + "lines": [ + { + "bbox": [ + 271, + 198, + 335, + 207 + ], + "spans": [ + { + "bbox": [ + 271, + 198, + 335, + 207 + ], + "type": "text", + "content": "\"put tray in oven\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 224, + 481, + 247 + ], + "lines": [ + { + "bbox": [ + 131, + 224, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 131, + 224, + 481, + 247 + ], + "type": "text", + "content": "Fig. 6: Additional visualization of our proposed model. LEGO is able to synthesize faithful action frames as well as preserve the contexts in various scenarios." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 339, + 116, + 408, + 195 + ], + "blocks": [ + { + "bbox": [ + 339, + 116, + 408, + 195 + ], + "lines": [ + { + "bbox": [ + 339, + 116, + 408, + 195 + ], + "spans": [ + { + "bbox": [ + 339, + 116, + 408, + 195 + ], + "type": "image", + "image_path": "e23e14c041747c1193cec14305b1150f3bb23b4654e6c3d8a3ad236a0e9e1560.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 198, + 406, + 207 + ], + "lines": [ + { + "bbox": [ + 345, + 198, + 406, + 207 + ], + "spans": [ + { + "bbox": [ + 345, + 198, + 406, + 207 + ], + "type": "text", + "content": "\"close container\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 410, + 116, + 479, + 155 + ], + "blocks": [ + { + "bbox": [ + 410, + 116, + 479, + 155 + ], + "lines": [ + { + "bbox": [ + 410, + 116, + 479, + 155 + ], + "spans": [ + { + "bbox": [ + 410, + 116, + 479, + 155 + ], + "type": "image", + "image_path": "f0a5be6566924e6471da1f7f68c3bf797b24ddb3f2952e81c61bb4f554e091ed.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 410, + 156, + 479, + 195 + ], + "blocks": [ + { + "bbox": [ + 410, + 156, + 479, + 195 + ], + "lines": [ + { + "bbox": [ + 410, + 156, + 479, + 195 + ], + "spans": [ + { + "bbox": [ + 410, + 156, + 479, + 195 + ], + "type": "image", + "image_path": "92b7729b3181df5a05b4551156b81e7bf8296346f53fc9d1ceb8bbf68b7b6d6e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 198, + 472, + 207 + ], + "lines": [ + { + "bbox": [ + 417, + 198, + 472, + 207 + ], + "spans": [ + { + "bbox": [ + 417, + 198, + 472, + 207 + ], + "type": "text", + "content": "\"take soy milk\"" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 270, + 480, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 480, + 306 + ], + "type": "text", + "content": "samples. Results are illustrated in Fig. 4. Our model surpasses the best baseline by " + }, + { + "bbox": [ + 130, + 270, + 480, + 306 + ], + "type": "inline_equation", + "content": "44.00\\%" + }, + { + "bbox": [ + 130, + 270, + 480, + 306 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 270, + 480, + 306 + ], + "type": "inline_equation", + "content": "38.34\\%" + }, + { + "bbox": [ + 130, + 270, + 480, + 306 + ], + "type": "text", + "content": " respectively on Ego4D and Epic-Kitchens. The prominent gap further validates the superiority of our model." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 323, + 334, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 323, + 334, + 335 + ], + "spans": [ + { + "bbox": [ + 131, + 323, + 334, + 335 + ], + "type": "text", + "content": "4.3 Visualization of Generated Frames" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 342, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 480, + 437 + ], + "type": "text", + "content": "We additionally showcase examples of generated image from LEGO and baseline models in Fig. 5. ProxEdit and SDEdit fail to understand the user prompt and thus generate frames of irrelevant actions (e.g., row3). They may also easily synthesize the action in a different environment (e.g., row2). InstructPix2Pix is able to preserve more contexts but still fails to semantically align with the action in user prompt (e.g., row1). In contrast, LEGO can synthesize action frames that better align with the user prompts and retain more contexts in the background. More examples of LEGO output are presented in Fig. 6 and supplementary." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 131, + 455, + 238, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 455, + 238, + 467 + ], + "spans": [ + { + "bbox": [ + 131, + 455, + 238, + 467 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "text", + "content": "We present comprehensive ablation studies to investigate the contribution of enriched action descriptions, the VLLM image embedding and the VLLM text embedding separately. Results are demonstrated in Tab. 3. Given the limitation of automatic metrics, we also provide user study as the primary metric. Notably, conditioning on enriched action descriptions can moderately improve the model performance, supporting our claim that expanding the action description can facilitate the learning of the state transition during an egocentric action. Moreover, utilizing the image embedding and text embedding from VLLM as additional conditions both improve the model performance by a notable margin because VLLM embeddings can effectively narrow the domain gap. Interestingly, the image embedding leads to larger performance boost on both datasets. These results suggest the VLLM image embedding " + }, + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_i" + }, + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "text", + "content": " incorporates important high-level semantic meanings that are not captured in the auto-encoder image embedding " + }, + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\mathcal{X})" + }, + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "text", + "content": " or VLLM text emebdding " + }, + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_t" + }, + { + "bbox": [ + 130, + 474, + 481, + 665 + ], + "type": "text", + "content": ". Finally, our full LEGO model uses both VLLM image and text embeddings (Desc.+Joint Embed.) and thus achieves the best performance on both datasets in all metrics." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "B. Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 179, + 485, + 255 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 3: Analysis of egocentric action frame generation performance with different conditionings. Joint Embed. refers to incorporating both VLLM image and text embeddings. Similar to Fig. 4, we present win rate as the user study result, i.e., the percentage of each model picked as the best (\\% is omitted for simplicity). The best results are highlighted with **boldface**. The orange rows refer to our full LEGO model." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 179, + 485, + 255 + ], + "lines": [ + { + "bbox": [ + 132, + 179, + 485, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 179, + 485, + 255 + ], + "type": "table", + "html": "
ConditioningEgo4DEpic-Kitchens
User StudyEgoVLPEgoVLP+CLIPUser StudyEgoVLPEgoVLP+CLIP
Actions Labels5.3362.1978.8478.757.0842.9761.0677.03
Descriptions13.0062.9179.0979.1812.5043.7261.4677.47
Desc. + Img Embed.26.0065.3580.1380.5724.1745.8262.2978.60
Desc. + Txt Embed.21.3363.2979.4079.2122.0844.6862.0277.74
Desc. + Joint Embed.34.3465.6580.4480.6134.1745.8962.6678.63
", + "image_path": "e461f5dae280bc574483fbb083e5ed11567abb13d0383636704d4690699a12f5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 132, + 300, + 484, + 373 + ], + "blocks": [ + { + "bbox": [ + 130, + 267, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 291 + ], + "type": "text", + "content": "Table 4: Performance of LEGO without finetuning (w/o FT). The best results are highlighted with **boldface**. The orange rows refer to our full LEGO model." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 300, + 484, + 373 + ], + "lines": [ + { + "bbox": [ + 132, + 300, + 484, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 484, + 373 + ], + "type": "table", + "html": "
ConditioningEgo4DEpic-Kitchens
EgoVLPEgoVLP+CLIPEgoVLPEgoVLP+CLIP
Descriptions62.9179.0979.1843.7261.4677.47
Desc.+Joint Embed.(w/o FT)64.5779.7280.2344.7461.8778.35
Desc.+Joint Embed.(w/ FT)65.6580.4480.6145.8962.6678.63
", + "image_path": "32649c8a74a3cc65dea9385aaefc6800e87309d0a49fe775f110c3d19bfc0af8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 399, + 350, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 399, + 350, + 412 + ], + "spans": [ + { + "bbox": [ + 131, + 399, + 350, + 412 + ], + "type": "text", + "content": "4.5 Analysis of Visual Instruction Tuning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "type": "text", + "content": "First, we assess the quality of enriched action descriptions from the visual instruction tuned VLLM by user study. For each sample in user study, we ask the raters to select whether the description aligns with the input and target frames (see supplementary for more details of user study setting and interface). The percentage of samples with aligned frames and action descriptions is " + }, + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "type": "inline_equation", + "content": "87\\%" + }, + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "type": "text", + "content": " on Ego4D and " + }, + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "type": "inline_equation", + "content": "84\\%" + }, + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "type": "text", + "content": " on Epic-Kitchens. The high alignment percentage suggests the visual instruction tuned VLLM can effectively expand action labels with details captured in the input frame. We additionally implement the same user study for off-the-shelf VLLM (i.e., without finetuning). The percentage of alignment is just " + }, + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "type": "text", + "content": " on Ego4D and " + }, + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "type": "text", + "content": " on Epic-Kitchens with hallucination existing in " + }, + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "type": "inline_equation", + "content": "92\\%" + }, + { + "bbox": [ + 130, + 425, + 482, + 568 + ], + "type": "text", + "content": " of unaligned instances. The notable drop supports our argument that visual instruction tuning is a critical step for high-quality prompt enrichment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "In addition, we further investigate whether the visual instruction tuned embeddings can contribute more to the diffusion model than the off-the-self counterpart, which has not been well studied in prior work. As shown in Tab. 4, without any finetuning, the image and text embeddings from VLLM can still improve the baseline model (Descriptions). However finetuned embeddings yield much larger improvement (e.g., " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "1.08\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "1.15\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " in EgoVLP score) on both datasets. The result suggests that visual instruction tuning is a necessary step to learn semantic image and text embeddings that are more aligned with the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "text", + "content": "Learning Egocentric Action Frame Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 129, + 232, + 171 + ], + "blocks": [ + { + "bbox": [ + 160, + 129, + 232, + 171 + ], + "lines": [ + { + "bbox": [ + 160, + 129, + 232, + 171 + ], + "spans": [ + { + "bbox": [ + 160, + 129, + 232, + 171 + ], + "type": "image", + "image_path": "771ec954a02730affa2393565be317566054377595ac6d16570e1c534346b433.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 194, + 156, + 213 + ], + "blocks": [ + { + "bbox": [ + 133, + 194, + 156, + 213 + ], + "lines": [ + { + "bbox": [ + 133, + 194, + 156, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 156, + 213 + ], + "type": "image", + "image_path": "bffe851e6fc13a86dc8e98805295c967d1203ada550d9ff9fc4f59412a8907e1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 236, + 481, + 280 + ], + "lines": [ + { + "bbox": [ + 131, + 236, + 481, + 280 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 481, + 280 + ], + "type": "text", + "content": "Fig. 7: Visualization of generating various actions from the same input frame. The first action (\"open drawer\") is the existing label for this example in the dataset. We fill another five actions into the user query and synthesize the action frames using our model. All generated frames align well with the actions and preserve most contexts." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 162, + 172, + 230, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 172, + 230, + 212 + ], + "spans": [ + { + "bbox": [ + 162, + 172, + 230, + 212 + ], + "type": "text", + "content": "\"Can you provide instructions on how to {action} in my current situation?\"" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 242, + 118, + 313, + 159 + ], + "blocks": [ + { + "bbox": [ + 242, + 118, + 313, + 159 + ], + "lines": [ + { + "bbox": [ + 242, + 118, + 313, + 159 + ], + "spans": [ + { + "bbox": [ + 242, + 118, + 313, + 159 + ], + "type": "image", + "image_path": "ed431d60297d8bcc73609181e5df9d3afcb5f6ca1ed5ccbcd5b922028eee54ab.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 159, + 306, + 168 + ], + "lines": [ + { + "bbox": [ + 251, + 159, + 306, + 168 + ], + "spans": [ + { + "bbox": [ + 251, + 159, + 306, + 168 + ], + "type": "text", + "content": "\"open drawer\"" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 315, + 118, + 384, + 159 + ], + "blocks": [ + { + "bbox": [ + 315, + 118, + 384, + 159 + ], + "lines": [ + { + "bbox": [ + 315, + 118, + 384, + 159 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 384, + 159 + ], + "type": "image", + "image_path": "cc4e8453b5e774f08fca5106185f04ba980f670673820f7b3fd466669a0414a0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 159, + 371, + 168 + ], + "lines": [ + { + "bbox": [ + 327, + 159, + 371, + 168 + ], + "spans": [ + { + "bbox": [ + 327, + 159, + 371, + 168 + ], + "type": "text", + "content": "\"dry hands\"" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 386, + 118, + 456, + 159 + ], + "blocks": [ + { + "bbox": [ + 386, + 118, + 456, + 159 + ], + "lines": [ + { + "bbox": [ + 386, + 118, + 456, + 159 + ], + "spans": [ + { + "bbox": [ + 386, + 118, + 456, + 159 + ], + "type": "image", + "image_path": "d994be76e84a8483fc6dbd67883afd705ed53feff5fed79fa396627070db72d6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 159, + 448, + 168 + ], + "lines": [ + { + "bbox": [ + 391, + 159, + 448, + 168 + ], + "spans": [ + { + "bbox": [ + 391, + 159, + 448, + 168 + ], + "type": "text", + "content": "\"cut cucumber\"" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 243, + 170, + 313, + 211 + ], + "blocks": [ + { + "bbox": [ + 243, + 170, + 313, + 211 + ], + "lines": [ + { + "bbox": [ + 243, + 170, + 313, + 211 + ], + "spans": [ + { + "bbox": [ + 243, + 170, + 313, + 211 + ], + "type": "image", + "image_path": "deb5960858bafc06089cc1a8dad5944d634f12dd52a0aee0a0ac7a770d5f5b53.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 245, + 212, + 310, + 220 + ], + "lines": [ + { + "bbox": [ + 245, + 212, + 310, + 220 + ], + "spans": [ + { + "bbox": [ + 245, + 212, + 310, + 220 + ], + "type": "text", + "content": "\"open microwave\"" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 315, + 170, + 385, + 211 + ], + "blocks": [ + { + "bbox": [ + 315, + 170, + 385, + 211 + ], + "lines": [ + { + "bbox": [ + 315, + 170, + 385, + 211 + ], + "spans": [ + { + "bbox": [ + 315, + 170, + 385, + 211 + ], + "type": "image", + "image_path": "159d8b797a447c7c0b0a3320727e86d8ae99283a87d17b12c2921e2a07ce4dbb.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 212, + 375, + 221 + ], + "lines": [ + { + "bbox": [ + 323, + 212, + 375, + 221 + ], + "spans": [ + { + "bbox": [ + 323, + 212, + 375, + 221 + ], + "type": "text", + "content": "\"pick up bowl\"" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 386, + 170, + 456, + 211 + ], + "blocks": [ + { + "bbox": [ + 386, + 170, + 456, + 211 + ], + "lines": [ + { + "bbox": [ + 386, + 170, + 456, + 211 + ], + "spans": [ + { + "bbox": [ + 386, + 170, + 456, + 211 + ], + "type": "image", + "image_path": "007d64c34566d8aed4e12ee1ff1f71551abcb90da2698d560bd335db0cb30f87.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 212, + 442, + 221 + ], + "lines": [ + { + "bbox": [ + 398, + 212, + 442, + 221 + ], + "spans": [ + { + "bbox": [ + 398, + 212, + 442, + 221 + ], + "type": "text", + "content": "\"take knife\"" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 460, + 190, + 477, + 210 + ], + "blocks": [ + { + "bbox": [ + 460, + 190, + 477, + 210 + ], + "lines": [ + { + "bbox": [ + 460, + 190, + 477, + 210 + ], + "spans": [ + { + "bbox": [ + 460, + 190, + 477, + 210 + ], + "type": "image", + "image_path": "0ddb6f1ecf13a62554b791b575a2f3902e670bd68daf08d1b82238226dc3309f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 212, + 481, + 221 + ], + "lines": [ + { + "bbox": [ + 460, + 212, + 481, + 221 + ], + "spans": [ + { + "bbox": [ + 460, + 212, + 481, + 221 + ], + "type": "text", + "content": "LEGO" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 305, + 480, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 480, + 329 + ], + "type": "text", + "content": "egocentric input frame and action prompt, which thus narrows the domain gap and greatly boosts the performance of the latent diffusion model." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 131, + 350, + 439, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 350, + 439, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 350, + 439, + 361 + ], + "type": "text", + "content": "4.6 Generation of Various Actions with the Same Contexts" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 374, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 470 + ], + "type": "text", + "content": "In addition to generating action frames based on the pre-defined action labels in our datasets, we validate the generative nature of our model i.e., whether LEGO is able to synthesize the correct action frames conditioning on the same contexts yet different actions (novel image-action pairs). Results are illustrated in Fig. 7. We feed different actions with the same input frame to our model. The synthesized frames correctly depict the execution of the actions in the same environment. This result further indicates that our model can understand the action state transition and generalize to different user queries." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 493, + 220, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 220, + 505 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 220, + 505 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduce the novel problem of egocentric action frame generation. We also propose a novel model — LEGO that leverages visual instruction tuning and a diffusion model to address this problem. Our key intuition is to use visual instruction tuning to generate informative responses that depict the execution of the egocentric action, and then design the conditioning for the denoising U-Net to exploit the image and text embeddings from a visual instruction tuned VLLM. Our experiments on two large-scale egocentric action datasets validate the advantage of the proposed approach as well as the contribution of each model component. We believe our work provides an important step in understanding the action state transition and controllability of diffusion models, and suggests future research directions in egocentric AI systems, action state transition, image generation and human skill transfer." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "B. Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 481, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 481, + 162 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 481, + 162 + ], + "type": "text", + "content": "Portions of this work were supported in part by a gift from Meta. We thank Sangmin Lee for the valuable discussion and suggestions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 178, + 198, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 178, + 198, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 198, + 190 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 201, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 142, + 201, + 480, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 201, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 142, + 201, + 480, + 243 + ], + "type": "text", + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 142, + 245, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 245, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 142, + 245, + 480, + 266 + ], + "type": "text", + "content": "2. Anonymity: Making multimodal generation easier: When diffusion models meet llms. openreview (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 266, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 266, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 142, + 266, + 480, + 298 + ], + "type": "text", + "content": "3. Ashutosh, K., Girdhar, R., Torresani, L., Grauman, K.: Hiervl: Learning hierarchical video-language embeddings. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23066-23078 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 299, + 480, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 299, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 142, + 299, + 480, + 330 + ], + "type": "text", + "content": "4. Baskin, J.H., Edersheim, J.G., Price, B.H.: Is a picture worth a thousand words? neuroimaging in the courtroom. American Journal of Law & Medicine 33(2-3), 239-269 (2007)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 331, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 331, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 142, + 331, + 480, + 363 + ], + "type": "text", + "content": "5. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 363, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 363, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 142, + 363, + 480, + 407 + ], + "type": "text", + "content": "6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 407, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 407, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 142, + 407, + 480, + 438 + ], + "type": "text", + "content": "7. Chakrabarty, T., Singh, K., Saakyan, A., Muresan, S.: Learning to follow object-centric image editing instructions faithfully. arXiv preprint arXiv:2310.19145 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 439, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 439, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 142, + 439, + 480, + 471 + ], + "type": "text", + "content": "8. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., et al.: Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 472, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 472, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 142, + 472, + 480, + 503 + ], + "type": "text", + "content": "9. Chen, W.G., Spiridonova, I., Yang, J., Gao, J., Li, C.: Llava-interactive: An all-in-one demo for image chat, segmentation, generation and editing. arXiv preprint arXiv:2311.00571 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 480, + 536 + ], + "type": "text", + "content": "10. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 536, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 557 + ], + "type": "text", + "content": "11. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 558, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 558, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 558, + 480, + 589 + ], + "type": "text", + "content": "12. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. Advances in Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "type": "text", + "content": "13. Damen, D., Doughty, H., Farinella, G.M., Furnari, A., Kazakos, E., Ma, J., Moltisanti, D., Munro, J., Perrett, T., Price, W., et al.: Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision pp. 1-23 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "14. Du, Y., Yang, S., Dai, B., Dai, H., Nachum, O., Tenenbaum, J.B., Schuurmans, D., Abbeel, P.: Learning universal policies via text-guided video generation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "text", + "content": "Learning Egocentric Action Frame Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "text", + "content": "15. Epstein, D., Jabri, A., Poole, B., Efros, A.A., Holynski, A.: Diffusion self-guidance for controllable image generation. In: Advances in Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 151, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 151, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 151, + 480, + 172 + ], + "type": "text", + "content": "16. Girdhar, R., Grauman, K.: Anticipative video transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13505-13515 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 174, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 174, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 174, + 480, + 217 + ], + "type": "text", + "content": "17. Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedforward neural networks. In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. pp. 249-256. JMLR Workshop and Conference Proceedings (2010)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 218, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 218, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 218, + 480, + 251 + ], + "type": "text", + "content": "18. Goel, V., Peruzzo, E., Jiang, Y., Xu, D., Sebe, N., Darrell, T., Wang, Z., Shi, H.: Pair-diffusion: Object-level image editing with structure-and-appearance paired diffusion models. arXiv preprint arXiv:2303.17546 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 480, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 284 + ], + "type": "text", + "content": "19. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139–144 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 285, + 480, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 480, + 318 + ], + "type": "text", + "content": "20. Goyal, M., Modi, S., Goyal, R., Gupta, S.: Human hands as probes for interactive object understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3293-3303 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 319, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 319, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 319, + 480, + 363 + ], + "type": "text", + "content": "21. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18995-19012 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 480, + 396 + ], + "type": "text", + "content": "22. Hafri, A., Trueswell, J.C., Epstein, R.A.: Neural representations of observed actions generalize across static and dynamic visual input. Journal of Neuroscience 37(11), 3056-3071 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 397, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 397, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 480, + 430 + ], + "type": "text", + "content": "23. Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., et al.: Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 431, + 480, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 431, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 138, + 431, + 480, + 474 + ], + "type": "text", + "content": "24. Han, L., Wen, S., Chen, Q., Zhang, Z., Song, K., Ren, M., Gao, R., Stathopoulos, A., He, X., Chen, Y., et al.: Proxedit: Improving tuning-free real image editing with proximal guidance. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4291-4301 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 475, + 480, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 475, + 480, + 507 + ], + "spans": [ + { + "bbox": [ + 138, + 475, + 480, + 507 + ], + "type": "text", + "content": "25. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 509, + 480, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 509, + 480, + 541 + ], + "spans": [ + { + "bbox": [ + 138, + 509, + 480, + 541 + ], + "type": "text", + "content": "26. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 543, + 480, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 543, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 138, + 543, + 480, + 563 + ], + "type": "text", + "content": "27. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 565, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 565, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 565, + 480, + 597 + ], + "type": "text", + "content": "28. Huang, J., Liu, Y., Qin, J., Chen, S.: Kv inversion: Kv embeddings learning for text-conditioned real image action editing. arXiv preprint arXiv:2309.16608 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "type": "text", + "content": "29. Huang, Y., Cai, M., Li, Z., Lu, F., Sato, Y.: Mutual context network for jointly estimating egocentric gaze and action. IEEE Transactions on Image Processing 29, 7795-7806 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "30. Huang, Y., Cai, M., Li, Z., Sato, Y.: Predicting gaze in egocentric video by learning task-dependent attention transition. In: Proceedings of the European conference on computer vision (ECCV). pp. 754-769 (2018)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "B. Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "text", + "content": "31. Iandola, F.N., Han, S., Moskewicz, M.W., Ashraf, K., Dally, W.J., Keutzer, K.: SqueezeNet: Alexnet-level accuracy with 50x fewer parameters and " + }, + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "inline_equation", + "content": "< 0.5" + }, + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "text", + "content": " mb model size. arXiv preprint arXiv:1602.07360 (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 151, + 480, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 151, + 480, + 184 + ], + "spans": [ + { + "bbox": [ + 138, + 151, + 480, + 184 + ], + "type": "text", + "content": "32. Jia, W., Liu, M., Rehg, J.M.: Generative adversarial network for future hand segmentation from egocentric video. In: European Conference on Computer Vision. pp. 639-656. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 185, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 185, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 185, + 480, + 205 + ], + "type": "text", + "content": "33. Jiang, Y., Zhang, Z., Xue, T., Gu, J.: Autodir: Automatic all-in-one image restoration with latent diffusion. arXiv preprint arXiv:2310.10123 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 207, + 480, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 207, + 480, + 250 + ], + "spans": [ + { + "bbox": [ + 138, + 207, + 480, + 250 + ], + "type": "text", + "content": "34. Joseph, K., Udhayanan, P., Shukla, T., Agarwal, A., Karanam, S., Goswami, K., Srinivasan, B.V.: Iterative multi-granular image editing using diffusion models. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 480, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 273 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 273 + ], + "type": "text", + "content": "35. Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. In: Advances in Neural Information Processing Systems (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 274, + 480, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 480, + 317 + ], + "type": "text", + "content": "36. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 319, + 480, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 319, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 138, + 319, + 480, + 351 + ], + "type": "text", + "content": "37. Kazakos, E., Nagrani, A., Zisserman, A., Damen, D.: Epic-fusion: Audio-visual temporal binding for egocentric action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5492-5501 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 353, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 480, + 384 + ], + "type": "text", + "content": "38. Kim, S., Jang, W., Kim, H., Kim, J., Choi, Y., Kim, S., Lee, G.: User-friendly image editing with minimal text input: Leveraging captioning and injection techniques. arXiv preprint arXiv:2306.02717 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 480, + 418 + ], + "type": "text", + "content": "39. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 420, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 420, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 138, + 420, + 480, + 441 + ], + "type": "text", + "content": "40. Koh, J.Y., Fried, D., Salakhutdinov, R.R.: Generating images with multimodal language models. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 442, + 480, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 442, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 138, + 442, + 480, + 474 + ], + "type": "text", + "content": "41. Lai, B., Liu, M., Ryan, F., Rehg, J.M.: In the eye of transformer: Global-local correlation for egocentric gaze estimation. British Machine Vision Conference (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 475, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 475, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 138, + 475, + 480, + 508 + ], + "type": "text", + "content": "42. Lai, B., Liu, M., Ryan, F., Rehg, J.M.: In the eye of transformer: Global-local correlation for egocentric gaze estimation and beyond. International Journal of Computer Vision 132(3), 854-871 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 510, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 480, + 531 + ], + "type": "text", + "content": "43. Lai, B., Ryan, F., Jia, W., Liu, M., Rehg, J.M.: Listen to look into the future: Audio-visual egocentric gaze anticipation. arXiv preprint arXiv:2305.03907 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 533, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 564 + ], + "type": "text", + "content": "44. Li, D., Li, J., Hoi, S.C.: Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 566, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 480, + 597 + ], + "type": "text", + "content": "45. Li, J., Liu, K., Wu, J.: Ego-body pose estimation via ego-head pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17142-17151 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "type": "text", + "content": "46. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. International Conference on Machine Learning (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "47. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "text", + "content": "Learning Egocentric Action Frame Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "text", + "content": "48. Li, Y., Liu, M., Rehg, J.M.: In the eye of beholder: Joint learning of gaze and actions in first person video. In: Proceedings of the European conference on computer vision (ECCV). pp. 619-635 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 138, + 149, + 480, + 182 + ], + "type": "text", + "content": "49. Lian, L., Li, B., Yala, A., Darrell, T.: Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 182, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 182, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 138, + 182, + 480, + 214 + ], + "type": "text", + "content": "50. Lin, K.Q., Wang, J., Soldan, M., Wray, M., Yan, R., XU, E.Z., Gao, D., Tu, R.C., Zhao, W., Kong, W., et al.: Egocentric video-language pretraining. Advances in Neural Information Processing Systems 35, 7575-7586 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 214, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 214, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 138, + 214, + 480, + 257 + ], + "type": "text", + "content": "51. Liu, B., Zhang, H., Liu, J., Wang, Q.: Acigs: An automated large-scale crops image generation system based on large visual language multi-modal models. In: 2023 20th Annual IEEE International Conference on Sensing, Communication, and Networking (SECON). pp. 7-13. IEEE (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 257, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 257, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 138, + 257, + 480, + 278 + ], + "type": "text", + "content": "52. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 278, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 278, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 480, + 322 + ], + "type": "text", + "content": "53. Liu, M., Tang, S., Li, Y., Rehg, J.M.: Forecasting human-object interaction: joint prediction of motor attention and actions in first person video. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 704-721. Springer (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 322, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 322, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 322, + 480, + 354 + ], + "type": "text", + "content": "54. Liu, S., Tripathi, S., Majumdar, S., Wang, X.: Joint hand motion and interaction hotspots prediction from egocentric videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3282-3292 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 354, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 386 + ], + "type": "text", + "content": "55. Luo, Z., Hachiuma, R., Yuan, Y., Kitani, K.: Dynamics-regulated kinematic policy for egocentric pose estimation. Advances in Neural Information Processing Systems 34, 25019-25032 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 480, + 418 + ], + "type": "text", + "content": "56. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: Guided image synthesis and editing with stochastic differential equations. In: International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 418, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 480, + 450 + ], + "type": "text", + "content": "57. Mirzaei, A., Aumentado-Armstrong, T., Brubaker, M.A., Kelly, J., Levinshtein, A., Derpanis, K.G., Gilitschenski, I.: Watch your steps: Local image and scene editing by text instructions. arXiv preprint arXiv:2308.08947 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 450, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 450, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 450, + 480, + 493 + ], + "type": "text", + "content": "58. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6038-6047 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 493, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 480, + 525 + ], + "type": "text", + "content": "59. Molad, E., Horwitz, E., Valevski, D., Acha, A.R., Matias, Y., Pritch, Y., Leviathan, Y., Hoshen, Y.: Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 525, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 547 + ], + "type": "text", + "content": "60. Orgad, H., Kawar, B., Belinkov, Y.: Editing implicit assumptions in text-to-image diffusion models. arXiv preprint arXiv:2303.08084 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 547, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 480, + 579 + ], + "type": "text", + "content": "61. Pan, Z., Gherardi, R., Xie, X., Huang, S.: Effective real image editing with accelerated iterative diffusion inversion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15912-15921 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 579, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 480, + 622 + ], + "type": "text", + "content": "62. Pramanick, S., Song, Y., Nag, S., Lin, K.Q., Shah, H., Shou, M.Z., Chellappa, R., Zhang, P.: Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5285-5297 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "text", + "content": "63. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "B. Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 160 + ], + "type": "text", + "content": "64. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 160, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 160, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 138, + 160, + 480, + 192 + ], + "type": "text", + "content": "65. Ragusa, F., Farinella, G.M., Furnari, A.: Stillfast: An end-to-end approach for short-term object interaction anticipation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3635-3644 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 192, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 192, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 138, + 192, + 480, + 225 + ], + "type": "text", + "content": "66. Ramakrishnan, S.K., Al-Halah, Z., Grauman, K.: Naq: Leveraging narrations as queries to supervise episodic memory. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6694-6703 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 225, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 225, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 138, + 225, + 480, + 257 + ], + "type": "text", + "content": "67. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 257, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 257, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 257, + 480, + 300 + ], + "type": "text", + "content": "68. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 300, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 480, + 332 + ], + "type": "text", + "content": "69. Ryan, F., Jiang, H., Shukla, A., Rehg, J.M., Ithapu, V.K.: Egocentric auditory attention localization in conversations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14663-14674 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 332, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 480, + 365 + ], + "type": "text", + "content": "70. Shen, Y., Song, K., Tan, X., Li, D., Lu, W., Zhuang, Y.: Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 365, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 386 + ], + "type": "text", + "content": "71. Shi, J., Xiong, W., Lin, Z., Jung, H.J.: Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 386, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 480, + 429 + ], + "type": "text", + "content": "72. Stein, G., Cresswell, J.C., Hosseinzadeh, R., Sui, Y., Ross, B.L., Villecloze, V., Liu, Z., Caterini, A.L., Taylor, J.E.T., Loaiza-Ganem, G.: Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. arXiv preprint arXiv:2306.04675 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 429, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 480, + 450 + ], + "type": "text", + "content": "73. Su, Y., Lan, T., Li, H., Xu, J., Wang, Y., Cai, D.: Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 450, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 450, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 450, + 480, + 482 + ], + "type": "text", + "content": "74. Sudhakaran, S., Escalera, S., Lanz, O.: Lsta: Long short-term attention for egocentric action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9954-9963 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 482, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 482, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 138, + 482, + 480, + 515 + ], + "type": "text", + "content": "75. Sun, Z., Zhou, Y., He, H., Mok, P.: Sgdiff: A style guided diffusion model for fashion synthesis. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 8433-8442 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 515, + 480, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 515, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 138, + 515, + 480, + 548 + ], + "type": "text", + "content": "76. Thoppilan, R., De Freitas, D., Hall, J., Shazeer, N., Kulshreshtha, A., Cheng, H.T., Jin, A., Bos, T., Baker, L., Du, Y., et al.: Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 548, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 548, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 548, + 480, + 579 + ], + "type": "text", + "content": "77. Tome, D., Alldieck, T., Peluse, P., Pons-Moll, G., Agapito, L., Badino, H., De la Torre, F.: Selfpose: 3d egocentric pose estimation from a headset mounted camera. IEEE Transactions on Pattern Analysis and Machine Intelligence (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 579, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 480, + 611 + ], + "type": "text", + "content": "78. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "text", + "content": "79. Tsaban, L., Passos, A.: Ledits: Real image editing with ddpm inversion and semantic guidance. arXiv preprint arXiv:2307.00522 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "80. Wallace, B., Gokul, A., Naik, N.: Edict: Exact diffusion inversion via coupled transformations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22532-22541 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "text", + "content": "Learning Egocentric Action Frame Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 136, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 136, + 116, + 481, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 481, + 150 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 481, + 150 + ], + "type": "text", + "content": "81. Wang, J., Luvizon, D., Xu, W., Liu, L., Sarkar, K., Theobalt, C.: Scene-aware egocentric 3d human pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13031-13040 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 150, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 150, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 138, + 150, + 481, + 184 + ], + "type": "text", + "content": "82. Wang, K., Yang, F., Yang, S., Butt, M.A., van de Weijer, J.: Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. arXiv preprint arXiv:2309.15664 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 184, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 481, + 217 + ], + "type": "text", + "content": "83. Wang, Q., Zhang, B., Birsak, M., Wonka, P.: Instructedit: Improving automatic masks for diffusion-based image editing with user instructions. arXiv preprint arXiv:2305.18047 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 218, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 218, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 218, + 481, + 251 + ], + "type": "text", + "content": "84. Wang, Q., Zhang, B., Birsak, M., Wonka, P.: Mdp: A generalized framework for text-guided image editing by manipulating the diffusion path. arXiv preprint arXiv:2303.16765 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 481, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 481, + 284 + ], + "type": "text", + "content": "85. Wang, W., Xie, K., Liu, Z., Chen, H., Cao, Y., Wang, X., Shen, C.: Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 285, + 481, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 481, + 318 + ], + "type": "text", + "content": "86. Wang, X., Zhu, L., Wang, H., Yang, Y.: Interactive prototype learning for egocentric action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8168-8177 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 318, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 318, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 318, + 481, + 352 + ], + "type": "text", + "content": "87. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 352, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 352, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 352, + 481, + 385 + ], + "type": "text", + "content": "88. Wen, S., Fang, G., Zhang, R., Gao, P., Dong, H., Metaxas, D.: Improving compositional text-to-image generation with large vision-language models. arXiv preprint arXiv:2310.06311 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 481, + 418 + ], + "type": "text", + "content": "89. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 419, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 419, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 138, + 419, + 481, + 441 + ], + "type": "text", + "content": "90. Wu, S., Fei, H., Qu, L., Ji, W., Chua, T.S.: Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 442, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 442, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 138, + 442, + 481, + 485 + ], + "type": "text", + "content": "91. Xu, Y., Li, Y.L., Huang, Z., Liu, M.X., Lu, C., Tai, Y.W., Tang, C.K.: Egopca: A new framework for egocentric hand-object interaction understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5273-5284 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 487, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 481, + 531 + ], + "type": "text", + "content": "92. Ye, Y., Li, X., Gupta, A., De Mello, S., Birchfield, S., Song, J., Tulsiani, S., Liu, S.: Affordance diffusion: Synthesizing hand-object interactions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22479-22489 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 531, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 531, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 531, + 481, + 564 + ], + "type": "text", + "content": "93. Yu, Q., Li, J., Ye, W., Tang, S., Zhuang, Y.: Interactive data synthesis for systematic vision adaptation via llms-aigcs collaboration. arXiv preprint arXiv:2305.12799 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 565, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 565, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 565, + 481, + 597 + ], + "type": "text", + "content": "94. Yu, Z., Li, H., Fu, F., Miao, X., Cui, B.: Fisted: Accelerating text-to-image editing via cache-enabled sparse diffusion inference. arXiv preprint arXiv:2305.17423 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 599, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 481, + 621 + ], + "type": "text", + "content": "95. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "text", + "content": "96. Zhang, M., Teck Ma, K., Hwee Lim, J., Zhao, Q., Feng, J.: Deep future gaze: Gaze anticipation on egocentric videos using adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4372-4381 (2017)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "B. Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 482, + 315 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 136, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 482, + 149 + ], + "type": "text", + "content": "97. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 136, + 150, + 481, + 182 + ], + "type": "text", + "content": "98. Zhang, S., Yang, X., Feng, Y., Qin, C., Chen, C.C., Yu, N., Chen, Z., Wang, H., Savarese, S., Ermon, S., et al.: Hive: Harnessing human feedback for instructional visual editing. arXiv preprint arXiv:2303.09618 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 137, + 183, + 481, + 215 + ], + "type": "text", + "content": "99. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 134, + 216, + 480, + 248 + ], + "type": "text", + "content": "100. Zhang, Y., Huang, N., Tang, F., Huang, H., Ma, C., Dong, W., Xu, C.: Inversion-based style transfer with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10146-10156 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 249, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 249, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 133, + 249, + 480, + 282 + ], + "type": "text", + "content": "101. Zhang, Z., Han, L., Ghosh, A., Metaxas, D.N., Ren, J.: Sine: Single image editing with text-to-image diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6027-6037 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 282, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 282, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 133, + 282, + 480, + 315 + ], + "type": "text", + "content": "102. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 258, + 91, + 447, + 102 + ], + "type": "text", + "content": "Learning Egocentric Action Frame Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "21" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 20 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LEIA_ Latent View-invariant Embeddings for Implicit 3D Articulation/0e9336bc-d9b2-471b-94f5-34342577a604_content_list.json b/2024/LEIA_ Latent View-invariant Embeddings for Implicit 3D Articulation/0e9336bc-d9b2-471b-94f5-34342577a604_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ebd293b2dce8acd218398cece510ad82cd5090a1 --- /dev/null +++ b/2024/LEIA_ Latent View-invariant Embeddings for Implicit 3D Articulation/0e9336bc-d9b2-471b-94f5-34342577a604_content_list.json @@ -0,0 +1,1555 @@ +[ + { + "type": "text", + "text": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation", + "text_level": 1, + "bbox": [ + 243, + 140, + 756, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Archana Swaminathan1, Anubhav Gupta1, Kamal Gupta1, Shishira R. Maiya1, Vatsal Agarwal1, and Abhinav Shrivastava1", + "bbox": [ + 233, + 212, + 769, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Maryland, College Park MD 20742, USA https://archana1998.github.io/leia/", + "bbox": [ + 320, + 255, + 681, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Neural Radiance Fields (NeRFs) have revolutionized the reconstruction of static scenes and objects in 3D, offering unprecedented quality. However, extending NeRFs to model dynamic objects or object articulations remains a challenging problem. Previous works have tackled this issue by focusing on part-level reconstruction and motion estimation for objects, but they often rely on heuristics regarding the number of moving parts or object categories, which can limit their practical use. In this work, we introduce LEIA, a novel approach for representing dynamic 3D objects. Our method involves observing the object at distinct time steps or \"states\" and conditioning a hypernetwork on the current state, using this to parameterize our NeRF. This approach allows us to learn a view-invariant latent representation for each state. We further demonstrate that by interpolating between these states, we can generate novel articulation configurations in 3D space that were previously unseen. Our experimental results highlight the effectiveness of our method in articulating objects in a manner that is independent of the viewing angle and joint configuration. Notably, our approach outperforms previous methods that rely on motion information for articulation registration.", + "bbox": [ + 261, + 321, + 738, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Articulated Objects $\\cdot$ Neural Radiance Fields $\\cdot$ 3D Vision", + "bbox": [ + 263, + 585, + 732, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 626, + 374, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our world is full of dynamic objects moving and interacting in space and time. Humans develop this (rather impressive) understanding of how these everyday objects move and interact in three-dimensional space at a very early stage of the brain development [31]. The task involves understanding not only the static geometry but also the dynamic movements and spatial relationships between parts of an object, often referred to as articulations of the object. Understanding and representing object articulations from images and/or videos is also pivotal in enabling machines to perceive and navigate the physical world with finesse. In this work, we propose a novel method to model the object articulations by learning view-invariant latent embeddings of the 3D object from multiview images. Existing works in modeling object articulation in the three dimensional space typically use priors in the form of pre-trained large scale models [51], videos [27],", + "bbox": [ + 217, + 657, + 785, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f83dcbc494267fd326451141bc3f048acbce5aef71827d86cbb77962b0e072dd.jpg", + "image_caption": [ + "Fig. 1: Our method LEIA, takes in multi-view images of an object in four articulation states and is able to learn a view-invariant latent embedding for the state. We show that we can interpolate between the latents to generate any number of intermediate unseen states for the object using LEIA, given the camera position." + ], + "image_footnote": [], + "bbox": [ + 225, + 148, + 769, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "or assumptions regarding rigidity/shape of the object [17]. These approaches often fail to generalize for the long-tail of objects, especially in the case when a video of the object with articulation is unavailable, or in presence of large and multiple articulations in the object. In our work, we postulate that multiview images of an object in different states can provide enough signal to model its 3D shape and articulation, even without priors from large-scale foundation models or information from videos. We achieve this by learning a generalizable, view-invariant latent embedding of the object for different states. We train these embeddings jointly with a hypernetwork that predicts weights of a NeRF [19] parameterizing the object state. The hypernetwork can be trained using multiview images of discretized states of the object, each state representing a different object articulation. In Figure 1, we show three objects from our dataset in two different articulation states. The latent embeddings representing each state or articulation can be interpolated during test time to generate the weights of a NeRF that is able to reconstruct a never seen before state of the object.", + "bbox": [ + 212, + 405, + 787, + 632 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key contributions can be summarized as following:", + "bbox": [ + 238, + 633, + 640, + 650 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce an end-to-end method LEIA for generating novel states for articulated objects solely with multiview images captured at multiple states.", + "- We demonstrate that interpolating between embeddings can generate states of articulations of object not seen during training. The embedding space becomes interpolable with a manifold loss that encourages the latents to follow a structure that establishes a linear relationship between them, by minimizing the distance between the nearest neighbours in the latent space.", + "- Remarkably, LEIA achieves this without the need for any ground-truth 3D supervision, motion information, or articulation codes, establishing its versatility and effectiveness in capturing complex articulations.", + "- Our analyses demonstrate LEIA's robustness to single and multiple articulations, as well as combinations of motions. We can disentangle articulations" + ], + "bbox": [ + 223, + 652, + 784, + 837 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "A. Swaminathan et al.", + "bbox": [ + 271, + 114, + 423, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in different object parts if multiple are movable, making it scalable without constraints on the number of parts or motion types, unlike prior work.", + "bbox": [ + 246, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 217, + 200, + 385, + 217 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Radiance Fields. Neural Radiance Fields [1, 19, 21, 42] have proven to be a massive success in modeling 3D scenes, due to the high fidelity of 3D reconstruction and novel view synthesis for static scenes. For dynamic scenes, Neural Volumes [18] utilizes an encoder-decoder voxel-based representation, complemented by an implicit voxel warp field. Occupancy Flow [24] tackled non-rigid geometry by assigning a motion vector to each point in both space and time, but requires full 3D ground truth supervision. Some of the first dynamic NeRF approaches [26, 28, 39] optimize an underlying volumetric deformable function and [15] conditions the NeRF on time. [16, 26, 45] followed this work, by learning a 5D spatiotemporal neural field. However, the approaches above usually require a video as input and do not handle well the case of large articulations in everyday objects given multiple states of objects as input.", + "bbox": [ + 217, + 234, + 785, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Representations for Articulation. Due to the fine-grained nature of the task, deep neural network models for representing articulation require conditioning or refining parts of the architecture to suit the task. Category-specific reconstruction of deformable objects from images [7, 11-13], sparked interest in identifying and recovering the deformation in the 3D space [13, 22, 52]. Several works focus on shape reconstruction from videos [38, 47-49] that estimate a shape template for humans and animals. Other works that learn articulation from videos include Qian et al. [29] that detects and segments articulation planes from in-the-wild videos and [44, 53] which decouple the static and dynamic parts of videos. These works focus on modeling humans and animals where a large amount of data is available. Preliminary work in using images and shape started with A-SDF [28], which generates unseen articulations using Signed Distance Function [25], by providing an input shape and articulation code. CaDeX [14] is a unified representation for shape and motion, obtained from a point cloud input. Ditto [10] is a similar work that uses implicit representations for joint geometry and articulation modeling, with ground truth point clouds. CLA-NeRF [40] learns unseen articulated states from observing multi-view image input along with articulation information. Wei et. al [43] obtains an SDF-based articulable representation of common objects by feeding in images of articulated states across multiple categories, and CARTO [9] uses stereo images as input and uses a stereo encoder to infer the 3D shape, 6D pose etc. of multiple unknown objects. [4] focuses on manipulating object shapes to deform according to specified articulation commands. Moving away from 3D supervision and articulation annotations, Jiayi et al. [17] proposed PARIS, a method that is able to obtain unseen articulated states of an object, given multi-view images of just the start and end state of articulation. PARIS employs a composite rendering based approach, by decoupling the object into a static and moving part and then separately estimating and compositing the static and mobile neural radiance fields. In our work,", + "bbox": [ + 217, + 417, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation", + "bbox": [ + 263, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/12d9798901219f9d0c3dbf3f1a86aba94019b9210b2c0b21528da6f3959d954f.jpg", + "table_caption": [ + "Table 1: Comparison of LEIA with existing methods. We show that LEIA is the first approach that does not use or learn any explicit prior along with not having any articulation input. This gives us flexibility in scaling to modeling articulations of objects with more than one part, and can thus handle a wide range of motion. We also have one universal model that can learn to represent both prismatic and revolute motion, unlike PARIS that has two separate models. 3D Sup. refers to \"3D Supervision\"." + ], + "table_footnote": [], + "table_body": "
MethodImage/Shape InputDatasetArticulation Input3D Sup.?Prior# States# Parts
A-SDF [28]Shape CodeShape2MotionArticulation Code✓Articulation≥4≥1
CLA-NeRF [40]Multiview ImagesPartNet-MobilityArticulated PoseXArticulation≥41
NASAM [43]Multiview Images++PartNet-MobilityN/AXCategory≥4≥1
CARTO [9]Stereo ImagesCustomJoint Code✓Articulation≥41
Ditto [10]Point CloudsShape2MotionAnnotations✓Articulation21
PARIS [17]Multiview ImagesPartNet-MobilityN/AXMotion (learned)21
LEIAMultiview ImagesPartNet-MobilityN/A✓No4≥1
", + "bbox": [ + 217, + 239, + 787, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we start LEIA with a similar input setting, by preparing multi-view renderings of objects in different articulation states. Unlike PARIS, we do not decouple the object into static and moveable parts, we rather learn to predict the weights of a single NeRF for any unseen articulated state using a state-modulated HyperNetwork. We aren't limited by the decoupling and the learned motion prior in the case of PARIS, which enables us to scale and learn any amount and kind of motion an object can possibly have, including combinations. We show a summed comparison of LEIA with prior work in Table 1.", + "bbox": [ + 212, + 362, + 787, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hypernetworks with Implicit Neural Representations. Hypernetworks, a specialized class of networks designed to predict parameters for another network, aim to achieve generalization across novel tasks [8]. In the realm of stylizing 3D scenes, [3] employed a hypernetwork for applying diverse styles, while scene reconstructions from limited data points were achieved by [35, 36]. Despite the promise shown in representation, these hypernetworks operate on input data points, necessitating test-time optimizations and rendering them unsuitable for compression tasks. Rather than using the provided data (image/video) as input for the hypernetwork, an alternative approach involves employing an auto-decoder framework. In this framework, a learnable latent, without the need for an encoder, represents a data point. This technique, applied by [33] to represent a dataset of videos, assigns each latent to a distinct video. While this method yields a representation for each set of frames, the lack of decoupling in spatial-temporal coordinates limits its scalability to real-world frames. On a related note, [34] extended a similar approach to 3D shapes and scenes, effectively acquiring a latent representation suitable for various downstream tasks. We use a similar framework for learning the latents corresponding to each articulated state in our method.", + "bbox": [ + 212, + 483, + 787, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 777, + 331, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Background.** Our architecture is based on neural radiance fields, or NeRF [19], which parameterizes the radiance and volume density at a 3D location of a", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "A. Swaminathan et al.", + "bbox": [ + 271, + 114, + 423, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "scene as observed from a camera placed in a particular position using a neural network. $F_{\\mathrm{NeRF}}:(\\mathbf{x},\\mathbf{d}) \\to (\\mathbf{c},\\sigma)$ , where $\\mathbf{c} = (r,g,b)$ , and $\\sigma$ represent the radiance and volume density, and $\\mathbf{x}$ and $\\mathbf{d}$ represent the 3D location and viewing direction respectively. To render a pixel, the radiance $C(\\mathbf{r})$ of a camera ray $\\mathbf{r}(s) = \\mathbf{o} + s \\cdot \\mathbf{d}$ is integrated from near to far bounds $s_n$ and $s_f$ such that, $C(\\mathbf{r}) = \\int_{s_n}^{s_f} T(s)\\sigma(\\mathbf{r}(s))\\mathbf{c}(\\mathbf{r}(s),\\mathbf{d})ds$ , where the function $T(s)$ denotes the accumulated transmittance from $s_n$ to $s$ . To optimize the parameters of the MLP, a loss function is used that measures the discrepancy between the ground-truth and rendered images. Traditionally, the L2 loss is used for this purpose. However, to make the training more robust to outliers and to improve convergence, the Smooth L1 Loss can be employed as an alternative. The Smooth L1 Loss is a combination of L1 and L2 losses, behaving like L1 loss for large errors and like L2 loss for small errors. The loss function for the NeRF model using the Smooth L1 Loss is defined as:", + "bbox": [ + 212, + 146, + 787, + 356 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S m o o t h L 1 L o s s} (x) = \\left\\{ \\begin{array}{l l} 0. 5 \\cdot x ^ {2} & \\text {i f} | x | < 1 \\\\ | x | - 0. 5 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 361, + 785, + 400 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {S m o o t h L 1 - N e R F}} = \\sum_ {r \\in R} \\text {S m o o t h L 1 L o s s} \\left(\\hat {C} (\\mathbf {r}) - C (\\mathbf {r})\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 407, + 784, + 440 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $R$ is the set of rays used for sampling, and $\\hat{C}(\\mathbf{r})$ and $C(\\mathbf{r})$ represent the ground-truth and rendered colors, respectively. By minimizing this loss function, the NeRF model learns to accurately reproduce the radiance of the scene, leading to high-quality image synthesis from novel viewpoints. While this formulation works well for generating novel scene viewpoints for a static scene, it cannot handle dynamic scenes. [45] tried to address this shortcoming by making the MLP learn a spatiotemporal radiance field and used time, $t$ , as an additional input. This works in principle but it becomes expensive to scale to lengthier videos due to constantly increasing sampling space for NeRF. We use the $L_{\\mathrm{SmoothL1 - NeRF}}$ loss along with AdamW optimization to train LEIA, along with a $L_{\\mathrm{mask}}$ that is the BCE loss between the predicted opacity and ground truth foreground loss.", + "bbox": [ + 212, + 446, + 787, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Approach", + "text_level": 1, + "bbox": [ + 214, + 630, + 344, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hypernetworks, or hypernets for short, are neural networks that generate weights for another neural network, known as the target network. Hypernets can be conditioned on various domains and can predict the weights for multiple neural networks simultaneously, if trained appropriately.", + "bbox": [ + 212, + 650, + 785, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this work, we use a hypernet to modulate a Neural Radiance Field (NeRF) based on the state of articulation, thereby learning a parametrization for each state. We employ a learnable latent embedding, $Z$ , as input to this hypernet $h_l, l \\in L$ where $L$ is the number of layers of the hypernet. This approach allows us to not only modulate the NeRF but also create useful representations for each state. Our system can thus be represented as follows:", + "bbox": [ + 212, + 710, + 787, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\theta_ {t}} (\\mathbf {x}, \\mathbf {d}) = \\left(\\mathbf {c} _ {t}, \\sigma_ {t}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 808, + 785, + 833 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {t} = h (z _ {t}), z _ {t} \\in Z\n$$\n", + "text_format": "latex", + "bbox": [ + 464, + 828, + 589, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation", + "bbox": [ + 261, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a2588fda4b5e7f167ce848216164e9c33e2ebada7f090c2a663771e2ec4e6663.jpg", + "image_caption": [ + "Fig. 2: Overview of our method. We take multi-view images in different states as input. A learnable latent dictionary based off an autoencoder learns an embedding per state id. The latent embedding is used as an input to the hypernet, that modulates and generates weights of the NeRF to reconstruct the state that is fed in. At inference time, we do a weighed interpolation of the learnt latents to obtain a corresponding newly generated intermediate state." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 789, + 308 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The latents $Z$ , give us an additional ability to interpolate across seen states and generate novel states of the object, as we show in our work. The set of latents, $Z$ , in Eq. (3) can be understood as a learnable dictionary where keys are the state ids. Formally, we represent this as", + "bbox": [ + 212, + 430, + 787, + 491 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nZ = \\left\\{t: z _ {t} \\mid t \\in [ 0, 1, 2 \\dots T ] \\right\\} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 502, + 785, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where, $T$ is the total number of discretized states sampled for the object. Each of these states, $z_{t}$ , is used as an input to the hypernet to get the parameterization for NeRF as shown in Figure 2. Once parameterized, the NeRF is trained using multi-view images of the object taken from various camera angles, providing a comprehensive view of the object's articulation in the current state. We sample only one state per batch.", + "bbox": [ + 212, + 530, + 787, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Directly predicting the weights $\\theta$ of the base network $f_{\\theta}$ , using the hypernet $h_l$ , is expensive, parameter-heavy, and unsuitable for compression. Hence, we follow [32, 37] and instead predict low-rank matrices, which are then applied to the base network weights. This type of modulation acts as a form of subnetwork selection, analogous to systems proposed in [6, 30]. For a base network $f_{\\theta}$ with $L$ layers, our formulation now looks like", + "bbox": [ + 212, + 621, + 787, + 712 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta} \\left(\\left(\\mathbf {x}, \\mathbf {d}\\right) \\mid \\theta_ {t} ^ {l _ {1}}, \\theta_ {t} ^ {l _ {2}} \\dots \\theta_ {t} ^ {l _ {L}}\\right) = \\mathbf {c} _ {t}, \\sigma_ {t}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 720, + 575, + 739 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {t} ^ {l} = \\eta \\left(P ^ {l} \\times Q ^ {l}\\right) \\circ \\theta^ {l} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 496, + 742, + 784, + 760 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nh _ {l} (z _ {t}) = [ P ^ {l}, Q ^ {l} ]\n$$\n", + "text_format": "latex", + "bbox": [ + 470, + 763, + 589, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\theta^l$ represents the weights of the $l$ -th layer and $\\theta_t^l$ denotes the modulated weights for frame $t$ . Here, $\\eta$ signifies an activation function on the matrix-product of low rank matrices $\\mathbf{P}^l \\in \\mathbb{R}^{K \\times r}$ , $\\mathbf{Q}^l \\in \\mathbb{R}^{r \\times K}$ , where $K$ is the width of the", + "bbox": [ + 212, + 792, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "A. Swaminathan et al.", + "bbox": [ + 271, + 114, + 423, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "base network $f_{\\theta}$ and rank $r \\ll K$ . These matrices are responsible for adjusting the weights $\\theta_{l}$ as dictated by the corresponding hypernetwork $h_l$ . Note that all hypernetworks use the same latent $z_{t} \\in \\mathbb{R}^{D}$ as input. The rank $r$ acts as a hyperparameter that controls the compression-performance trade-off. We further elaborate on details about model architecture and design choices for the hypernet, NeRF and the learnable latent dictionary in the supplementary.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Interpolation", + "text_level": 1, + "bbox": [ + 214, + 258, + 372, + 273 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given an object with states $t_1$ and $t_2$ , and a scale $\\alpha$ , the task of state interpolation involves creating $\\alpha - 1$ coherent states between the given states. In order to achieve this, we do a linear interpolation on the state latents $z_1$ and $z_2$ and pass the resulting latent through the hypernetwork. This gives us the weight modulation required in the NeRF, and the updated base network is used to obtain different viewpoints of the intermediate state.", + "bbox": [ + 212, + 284, + 787, + 376 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nz _ {\\text {i n t e r}} = \\left(1 - \\beta_ {i}\\right) \\cdot z _ {t} + \\beta_ {i} \\cdot z _ {t - 1} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 388, + 785, + 410 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {\\text {i n t e r}}, \\sigma_ {\\text {i n t e r}} = f _ {\\theta_ {\\text {i n t e r}}} (\\mathbf {x}, \\mathbf {d}; h (z _ {\\text {i n t e r}}))\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 407, + 617, + 421 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e}, \\beta_ {i} \\in \\left[ \\frac {1}{\\alpha}, \\frac {2}{\\alpha}, \\dots , \\frac {\\alpha - 1}{\\alpha} \\right] \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 435, + 784, + 468 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "essentially generating $\\alpha - 1$ states between any two given states. In our experiments, we quantitatively evaluate the result of the interpolation with unseen ground truth obtained from our dataset of all states, by linearly averaging the two extreme states present in the training, to obtain a latent for a state id.", + "bbox": [ + 212, + 473, + 787, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Model Architecture", + "text_level": 1, + "bbox": [ + 214, + 556, + 426, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our experiments, both the base network $f_{\\theta}$ and hypernetworks $h_l$ are simple MLPs that take in a coordinate input, and an id depicting the state of articulation. A network $l_n$ is used to learn the latent dictionary for the states, which is a linear layer learnt in conjunction with the hypernet, and uses the torch.nn.Embedding class as a lookup table for the learnable latent embeddings, depicted by $z_t$ for $t \\in T$ . This latent is fed to the hypernetwork $h_l$ , that modulates the weights of the base network $f_{\\theta}$ for reconstructing the output of the corresponding state $\\mathrm{id}$ $t$ . The base network $f_{\\theta}$ is based off the NeRF architecture of Instant-NGP [21]. This architecture has separate fully-connected blocks for the geometry and texture of the scene, and we use separate hypernets to modulate each layer of these two FC blocks.", + "bbox": [ + 212, + 580, + 787, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Optimization: Loss functions and regularizers", + "text_level": 1, + "bbox": [ + 214, + 768, + 635, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Latent Manifold Loss. We use the latent manifold loss function that enforces a structured latent space by encouraging local consistency among learnt latents. For each latent vector, the loss computes the average Euclidean distance between", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation", + "bbox": [ + 261, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the vector and its $K$ nearest-neighbours on the manifold. This process enforces a smooth and continuous latent manifold, which is beneficial for models that rely on meaningful linear interpolations between the points on the manifold, and for tasks where the geometry of the latent space is crucial. Mathematically, we represent this loss for a particular state-id $i$ as:", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a n i f o l d}} \\left(\\mathbf {l} _ {i}\\right) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\| \\mathbf {l} _ {i} - \\mathbf {n} _ {k} \\| _ {2} ^ {2} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 237, + 785, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{n}_k$ are the $K$ nearest neighbors of $\\mathbf{l}_i$ in the latent space, and $\\| \\cdot \\|_2$ denotes the Euclidean (L2) norm. The loss is averaged over the selected latents and their nearest neighbors to ensure local uniformity in the manifold's geometry, where $K$ is a hyperparameter that determines the number of neighbors considered.", + "bbox": [ + 214, + 290, + 787, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Depth and Occlusion Regularization. Our depth and occlusion regularizations are designed to refine the clarity of rendered images by addressing occlusion and depth smoothness [23, 50]. The occlusion regularization loss, $L_{\\mathrm{occ}}$ , aims to mitigate the obscuring of objects located beyond a specified depth threshold during rendering. This is accomplished by generating a binary mask, $m_k$ , where the mask elements are set to 1 up to a certain index $M$ reflecting the regularization range, and 0 thereafter. The occlusion loss is then articulated as the normalized sum of the product of the mask and the sampled density values $\\sigma_k$ along a ray, as given by", + "bbox": [ + 214, + 352, + 789, + 488 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {o c c}} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\sigma_ {k} \\cdot m _ {k}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 491, + 785, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $K$ is the total number of sampled points on the ray. This formulation drives the model to prefer representations that reduce occlusions close to the camera, ensuring objects further away are not improperly concealed.", + "bbox": [ + 212, + 541, + 787, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For depth continuity, the depth smoothness regularization loss, $L_{\\mathrm{DS}}(\\theta ,\\mathbf{R})$ enforces the gradual transition of depth values among neighboring pixels, reducing sharp depth disparities that cause visual inconsistencies. If a ray $r$ intersects with a single depth value, this value is evaluated directly. In contrast, for multiple depth values, the loss is the aggregate of squared differences between adjacent depth estimates. Formally, for depth predictions $\\hat{d}_0(r)$ , the loss is quantified as", + "bbox": [ + 212, + 587, + 787, + 679 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {D S}} (\\theta , \\mathbf {R}) = \\sum_ {r \\in \\mathbf {R}} \\sum_ {i, j = 1} ^ {\\text {P a t c h} - 1} \\left(\\hat {d} _ {0} \\left(r _ {i j}\\right) - \\hat {d} _ {0} \\left(r _ {(i + 1) j}\\right)\\right) ^ {2} + \\left(\\hat {d} _ {0} \\left(r _ {i j}\\right) - \\hat {d} _ {0} \\left(r _ {i (j + 1)}\\right)\\right) ^ {2}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 694, + 785, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This measure is averaged across the rays to yield a measure of the depth map's smoothness. By integrating these regularizers into our training regimen, we significantly dampen the disturbances introduced by overlapping and obscured pixels, resulting in a more consistent interpolation between states. These benefits are substantiated through extensive ablation studies and quantitative analyses presented in the experiments.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "A. Swaminathan et al.", + "bbox": [ + 271, + 114, + 423, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Positional Encoding of the Latent. We incorporate positional encoding to capture the order of input elements, crucial for understanding articulation. Positional encoding injects the sequence with its inherent order, a key factor in articulation semantics. We employ the scheme from [41], which uses sine and cosine functions parameterized to encode varying frequencies. The computed positional encodings are added to the latent vector $z_{t}$ , enriching it with semantic information that more accurately encodes the state of articulation. We analyze the effects of positional encoding on our model in the experimental section.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 297, + 375, + 314 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Setup", + "text_level": 1, + "bbox": [ + 214, + 335, + 312, + 351 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. In this work, we use the PartNet-Mobility dataset [2, 20, 46], a large-scale synthetic collection of articulated objects in over 40 categories. The dataset has a variety of articulations defined, with objects comprising of single and multi-joint parts. Two types of motion, revolute (rotational) joint and prismatic (translation) articulation are represented in the dataset. To make up our training dataset, we choose 100 camera views that are arranged in a dome-like setup, capturing the upper hemisphere of the object. This is similar to the setup used by PARIS. We use the SAPIEN [46] library that the PartNet-Mobility dataset was released with, to render our RGB images at linearly spaced intervals of articulation to make up frames of a video that depicts the range of motion and obtain the camera parameters accordingly, converting it into the Blender [5] coordinate system to fit in our codebase. We use instances from 8 different common household item categories storage, microwave, laptop, oven, washer, dishwasher, sunglasses, box and choose 1-4 objects per category, bringing our total number of objects to 12. We also show the efficacy of our method on 68 images from a real-world scene of a chest of drawers, which are captured with a mobile phone and post-processed to remove background. Throughout LEIA, we train with a total of four states and interpolate between the two extreme states.", + "bbox": [ + 212, + 367, + 787, + 638 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines. The closest prior work with a setup similar to ours is PARIS, which takes in multi-view images of objects in two states and disentangles the static and moving part of the object. While we run experiments with four input states, we don't focus on learning a defined articulation for the object, instead we are able to recognize any arbitrary motion and combinations of motions given relevant states. We emphasize on the following differences:", + "bbox": [ + 212, + 641, + 787, + 731 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. PARIS uses separate models to train objects that have rotation and translation motion, respectively. While they have a method to estimate the type of motion, once determined, it is necessary to train on the appropriate model.", + "2. As PARIS also learns motion parameters for the articulation, given their pipeline of disentanglement of the static and moving parts of the object, this restricts the learning of motion parameters if there are multiple parts in the object, moving differently." + ], + "bbox": [ + 220, + 732, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation", + "bbox": [ + 261, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/6048f10e8cf9c094aee6378a3f22bb957f3fd86ce2b24b61ac8a29d1f1e06efb.jpg", + "table_caption": [ + "Table 2: Quantitative Results for Interpolated State Reconstruction. We compared our method with the PARIS baseline, trained on selected objects from the SAPIEN dataset. The results from three experiments are summarized below. The VanillaInt experiment involves simple interpolation of the latents. Our best-performing method, LEIA, introduces structure to the latents with a manifold loss and regularizers. Although PARIS learns a motion prior and LEIA implicitly performs state interpolation, both methods perform similarly for single-part objects. However, our approach excels with multi-part objects, outperforming PARIS significantly due to its flexibility in handling various motion types and articulations without constraints." + ], + "table_footnote": [], + "table_body": "
MetricsMethodsSingle-Part ArticulationMulti-Part ArticulationAverage
45135 Storage17128 Microwave10211 Laptop101917 Oven103778 Washer12085 Dishwasher44781 Storage245427 Storage345575 Storage4101297 Sunglasses7187 Oven102377 Box
PSNR↑PARIS28.6625.9424.9728.1334.4627.2426.3524.4125.7332.3329.3026.2027.81
VanillaInt24.2022.0022.6324.5336.0324.1630.9629.7829.1032.1130.3527.8527.81
LEIA26.6926.3825.0428.7136.1426.4931.0729.7829.8035.6030.8028.0529.55
SSIM↑PARIS0.990.970.970.970.980.960.950.950.940.960.960.930.96
VanillaInt0.950.910.900.930.990.920.950.950.930.970.950.930.94
LEIA0.970.950.950.980.990.950.960.950.950.970.960.950.96
LPIPS↓PARIS0.020.060.190.030.030.060.050.050.040.060.050.080.06
VanillaInt0.050.100.170.030.020.080.030.030.050.090.050.130.07
LEIA0.030.070.100.020.020.070.030.030.040.090.050.120.06
CD↓PARIS0.180.080.120.750.060.450.620.420.600.200.990.910.45
VanillaInt0.240.070.060.100.060.200.480.470.460.970.500.830.37
LEIA0.290.060.360.100.050.270.520.380.410.960.350.620.36
", + "bbox": [ + 217, + 281, + 787, + 435 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3. LEIA is motion-prior free, and we have a universal architecture that implicitly learns intermediate states, so we are able to train with multiple types of motion occurring in the same object, with no change in the code.", + "bbox": [ + 220, + 464, + 785, + 508 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For comparing with PARIS, we run authors' original source code with our dataset, using the hyperparameters provided by the authors and training until convergence. We test our work and PARIS with the appearance quality of the reconstructed image of the interpolated state. We also set up a simple baseline that does just vanilla interpolation between the learnt latent embeddings, without enforcing any structure or constraints on them. We call this baseline VanillaInt and compare it with LEIA, which has been finetuned to do interpolation with the addition of the manifold loss and the depth and occlusion regularization for denoising the resulting output.", + "bbox": [ + 212, + 511, + 787, + 647 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Quantitative Metrics. We use three appearance quality metrics, PSNR (Peak Signal-to-Noise Ratio), SSIM (Structural Similarity Index), and LPIPS (Learned Perceptual Image Patch Similarity). These are computed between reconstructed images of the interpolated state, and the unseen ground truth of the state from our dataset. We also report the Chamfer Distance to measure quality of 3D reconstruction by doing point sampling.", + "bbox": [ + 212, + 647, + 787, + 739 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 215, + 765, + 372, + 779 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Novel State Synthesis. We show the qualitative and quantitative results of interpolating our learned latent embeddings corresponding to the start and end states, in Table 2 and Figure 3. Our results show that the latent embeddings can", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "A. Swaminathan et al.", + "bbox": [ + 271, + 114, + 421, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/4ad50401da0834031c7f891aec74172bd3b21281ab7b9416271fb3d760158546.jpg", + "image_caption": [ + "Fig. 3: Qualitative Results. We show results of PARIS and LEIA for reconstructing the unseen intermediate state, for both single and multiple articulations. We see that PARIS especially fails when there are two parts of the object moving differently, as the motion parameters are not registered correctly. LEIA handles this case successfully as it is not dependent on part disentanglement to identify and register articulation. LEIA also performs comparable to PARIS for single-part articulation, despite not having a dedicated model for the motion or part disentanglement." + ], + "image_footnote": [], + "bbox": [ + 251, + 143, + 733, + 659 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "be linearly interpolated, and have a structure in the latent space. We can generate any number of states between the start and the end state by doing a weighted", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation", + "bbox": [ + 263, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/715c8b2973ffca7878e9bc50c48b5382415e843156db77caaf2097ba0500265c.jpg", + "image_caption": [ + "Fig. 4: Real World Results. LEIA is able to faithfully interpolate and reconstruct between two states of images from our real world data, proving its ability to generalize and work in an in-the-wild setting." + ], + "image_footnote": [], + "bbox": [ + 246, + 157, + 754, + 228 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "combination of the latents representing the start and the end states. For training, we choose four states, using the two additional states compared to PARIS to enforce structure in the latent space with the latent manifold loss, thereby making them amenable to interpolation. This helps us extend our formulation to multiple articulations without restriction. We also show results of our baseline VanillaInt and show that the numbers are boosted with the addition of the latent manifold loss and regularizers. For single-part objects, we perform comparably to PARIS despite having no part disentanglement and no dedicated model to represent the exact type of motion. For multiple-part articulations, we beat PARIS across all objects, as shown in Figure 3. Our comparable performance in 3D reconstruction shows that LEIA's learnt view-invariant latent embeddings are powerful in preserving spatio-temporal consistency. We also show that our method can work in the real world setting, in Figure 4. Ablations are shown for number of states, latent manifold loss, and depth and occlusion regularizations in Table 3 and Figure 6.", + "bbox": [ + 212, + 316, + 787, + 542 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Analysis on Disentanglement of Joints. To check separability of the latent space, we train LEIA with multi-view images across multiple states, captured from a video that showed three parts of an object Storage1, move one after the other. We do a t-SNE based dimensionality reduction on the embeddings and plot them, shown in Figure 5. The latents exhibit clear separability, with the motion of each joint represented in a smooth trajectory, moving outward as the respective drawer in the object moves away from the starting \"closed\" state. This shows that the separated latent space is an indicator of multiple joints, positioning us to be capable of learning representations of multiple articulations.", + "bbox": [ + 212, + 545, + 787, + 681 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Analysis", + "text_level": 1, + "bbox": [ + 215, + 705, + 408, + 720 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We perform ablations to tune various design choices, with results shown in Table 3 and Figure 6. The first experiment we performed was testing out the impact of the latent manifold loss. We notice qualitatively that without the manifold loss, our latent space lacked structure and the linear relationship between the embeddings was not captured. As shown in Figure 6, the intermediate state overfits to the extreme state without the latent manifold while the addition of the loss enables it to accurately capture the intermediate state. This figure also", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "A. Swaminathan et al.", + "bbox": [ + 271, + 114, + 421, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/bbd16fba78450d42a9ec7c3bf80b66b2c60b3fccff5869e70baaf030834d920a.jpg", + "image_caption": [ + "Fig. 5: t-SNE plot. After dimensionality reduction on jointly-learned state embeddings of an object with different moving parts. Our learned representations are separated and follow a smooth trajectory for each of the moving parts." + ], + "image_footnote": [], + "bbox": [ + 220, + 150, + 452, + 330 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/11557f08838939f2233e75026b499749108d281efb43efc012e2102fa5273a25.jpg", + "table_caption": [ + "Table 3: Ablations. Our ablations reveal that latent manifold loss, depth, and occlusion regularization enhance LEIA's visual metrics. Opting for four states improved latent structure with the manifold loss." + ], + "table_footnote": [], + "table_body": "
Latent Manifold LossPSNRSSIMLPIPS
with29.400.950.05
without28.540.940.06
Depth RegularizationPSNRSSIMLPIPS
with29.630.960.05
without26.930.930.07
Occlusion RegularizationPSNRSSIMLPIPS
with29.640.950.05
without28.640.950.06
Positional encodingPSNRSSIMLPIPS
with27.110.940.06
without28.480.950.06
Number of StatesPSNRSSIMLPIPS
228.040.950.06
429.690.960.05
", + "bbox": [ + 504, + 229, + 781, + 426 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a68404074e403e14993805059219a8ad50f682219ca07963196a48066b0a9038.jpg", + "image_caption": [ + "Fig. 6: Qualitative Results of Ablations. We show some qualitative results of using latent manifold loss, positional encoding and varying the number of states. Using the manifold loss prevents us from overfitting to the extreme states, and using four states help us interpolate with a lot of clarity. Positional encoding helps add missing information for thin parts, but for large objects it doesn't help much." + ], + "image_footnote": [], + "bbox": [ + 218, + 452, + 767, + 691 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "shows the effect of positional encoding which adds detail when the latent embedding fails to capture it, as evidenced in the results of sunglasses. This can", + "bbox": [ + 215, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation", + "bbox": [ + 263, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/70fd1f6b3209cee7e9da6a876b8f5bef52c1c2dbe1874431545a28352300f853.jpg", + "image_caption": [ + "Fig. 7: Failure cases. Our model fails at reconstructing the geometry correctly at camera positions where the two states we interpolate between have a change in the visible shape, like these examples where the microwave and oven being closed shows a deformation in the figure as compared to the open state, as the relationship between the motion is not easily captured in the latent due to the structure change." + ], + "image_footnote": [], + "bbox": [ + 228, + 170, + 763, + 219 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "happen for thin parts of the object that may not be well-captured across camera views. Notably, the inverse happens in the case of large objects, where the positional encoding adds extra noise to the reconstruction, resulting in oversmoothing. We also show how adding just two more intermediate states in LEIA makes a huge difference in reconstruction results for both single and multi-part objects, enabling us to scale and be flexible. Quantitative numbers for depth and occlusion regularization ablation are shown in Table 3. However, they didn't universally help in the cases where the latent embedding already learned the state representation well enough but reduced noise when it appeared.", + "bbox": [ + 212, + 332, + 787, + 470 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Limitations", + "text_level": 1, + "bbox": [ + 215, + 488, + 362, + 503 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "While LEIA works in achieving good quality reconstruction of intermediate states, our latent embeddings do not yet ensure accurate physical consistency in the motion of the intermediates, which is a tradeoff we chose while opting for not decoupling the object and learning the motion parameters for the moveable part. This allowed us to scale our approach to interpolate between multiple moving parts of the object, which would be more representative of the real world where common objects can be articulated in multiple ways. LEIA also struggles when there is severe occlusion, as referenced by Figure 7, as the occlusion occurs at gaps in the learnt embeddings between two states.", + "bbox": [ + 212, + 518, + 787, + 654 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 674, + 359, + 690 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we present LEIA, a method capable of successfully interpolating between two discrete articulation states of a deformable object with moving parts. LEIA handles multiple joints, including cases where moving parts are separated by static regions. It outperforms existing methods, especially when multiple parts move independently. The learned latent embeddings are view-invariant and separable, demonstrating LEIA's scalability and flexibility. We conducted comprehensive evaluations on synthetic and real data to investigate inherent challenges. Despite advancements, significant occlusion scenarios remain challenging. We hope future research builds upon our work.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "A. Swaminathan et al.", + "bbox": [ + 271, + 114, + 421, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was partially supported by NSF CAREER Award (#2238769) to Abhinav Shrivastava, and IARPA via Department of Interior/Interior Business Center (DOI/IBC) contract number 140D0423C0076. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of NSF, IARPA, DOI/IBC, or the U.S. Government.", + "bbox": [ + 212, + 176, + 787, + 297 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 321, + 321, + 338 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5470-5479 (June 2022)", + "2. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)", + "3. Chiang, P.Z., Tsai, M.S., Tseng, H.Y., Lai, W.S., Chiu, W.C.: Stylizing 3d scene via implicit representation and hypernetwork. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1475-1484 (2022)", + "4. Chu, R., Liu, Z., Ye, X., Tan, X., Qi, X., Fu, C.W., Jia, J.: Command-driven articulated object understanding and manipulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8813-8823 (June 2023)", + "5. Community, B.O.: Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam (2018), http://www.blender.org", + "6. Frankle, J., Carbin, M.: The lottery ticket hypothesis: Finding sparse, trainable neural networks. arXiv preprint arXiv:1803.03635 (2018)", + "7. Goel, S., Kanazawa, A., Malik, J.: Shape and viewpoints without keypoints. In: ECCV (2020)", + "8. Ha, D., Dai, A.M., Le, Q.V.: Hypernetworks. In: International Conference on Learning Representations (2017), https://openreview.net/forum?id=rkpAce1lx", + "9. Heppert, N., Irshad, M.Z., Zakharov, S., Liu, K., Ambrus, R.A., Bohg, J., Valada, A., Kollar, T.: Carto: Category and joint agnostic reconstruction of articulated objects. In: Proc. IEEE Conf. Comput. Vis. Pattern Recog. pp. 21201-21210 (2023)", + "10. Jiang, Z., Hsu, C.C., Zhu, Y.: Ditto: Building digital twins of articulated objects from interaction. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "11. Kanazawa, A., Tulsiani, S., Efros, A.A., Malik, J.: Learning category-specific mesh reconstruction from image collections. In: ECCV (2018)", + "12. Kokkinos, F., Kokkinos, I.: To the point: Correspondence-driven monocular 3d category reconstruction. In: Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems (2021), https://openreview.net/forum?id=AWMU04iXQ08" + ], + "bbox": [ + 217, + 353, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation", + "bbox": [ + 261, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Kulkarni, N., Gupta, A., Fouhey, D.F., Tulsiani, S.: Articulation-aware canonical surface mapping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 452-461 (2020)", + "14. Lei, J., Daniilidis, K.: Cadex: Learning canonical deformation coordinate space for dynamic surface representation via neural homeomorphism. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6624-6634 (June 2022)", + "15. Li, T., Slavcheva, M., Zollhoefer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., Lv, Z.: Neural 3d video synthesis from multi-view video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)", + "16. Lin, H., Peng, S., Xu, Z., Yan, Y., Shuai, Q., Bao, H., Zhou, X.: Efficient neural radiance fields for interactive free-viewpoint video. In: SIGGRAPH Asia Conference Proceedings (2022)", + "17. Liu, J., Mahdavi-Amiri, A., Savva, M.: PARIS: Part-level reconstruction and motion analysis for articulated objects. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2023)", + "18. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. ACM Trans. Graph. 38(4), 65:1-65:14 (Jul 2019)", + "19. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "20. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)", + "21. Müller, T., Evans, Å., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022). https://doi.org/10.1145/3528223.3530127, https://doi.org/10.1145/3528223.3530127", + "22. Neverova, N., Novotny, D., Khalidov, V., Szafraniec, M., Labatut, P., Vedaldi, A.: Continuous surface embeddings (2020)", + "23. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S.M., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2022)", + "24. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Occupancy flow: 4d reconstruction by learning particle dynamics. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)", + "25. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)", + "26. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. ICCV (2021)", + "27. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021)", + "28. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "A. Swaminathan et al.", + "bbox": [ + 271, + 114, + 421, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Qian, S., Jin, L., Rockwell, C., Chen, S., Fouhey, D.F.: Understanding 3d object articulation in internet videos. In: CVPR (2022)", + "30. Ramanujan, V., Wortsman, M., Kembhavi, A., Farhadi, A., Rastegari, M.: What's hidden in a randomly weighted neural network? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11893-11902 (2020)", + "31. Saffran, J.R., Aslin, R.N., Newport, E.L.: Statistical learning by 8-month-old infants. Science 274(5294), 1926-1928 (1996)", + "32. Schwarz, J.R., Tack, J., Teh, Y.W., Lee, J., Shin, J.: Modality-agnostic variational compression of implicit neural representations. arXiv preprint arXiv:2301.09479 (2023)", + "33. Sen, B., Agarwal, A., Namboodiri, V.P., Jawahar, C.: Inr-v: A continuous representation space for video-based generative tasks. arXiv preprint arXiv:2210.16579 (2022)", + "34. Sen, B., Singh, G., Agarwal, A., Agaram, R., Krishna, K.M., Sridhar, S.: Hyp-nerf: Learning improved nef priors using a hypernetwork. arXiv preprint arXiv:2306.06093 (2023)", + "35. Sitzmann, V., Rezchikov, S., Freeman, B., Tenenbaum, J., Durand, F.: Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems 34, 19313-19325 (2021)", + "36. Sitzmann, V., Zollhöfer, M., Wetzstein, G.: Scene representation networks: Continuous 3d-structure-aware neural scene representations. Advances in Neural Information Processing Systems 32 (2019)", + "37. Skorokhodov, I., Ignatyev, S., Elhoseiny, M.: Adversarial generation of continuous images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10753-10764 (2021)", + "38. Tan, J., Yang, G., Ramanan, D.: Distilling neural fields for real-time articulated shape reconstruction. In: CVPR (2023)", + "39. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In: IEEE International Conference on Computer Vision (ICCV). IEEE (2021)", + "40. Tseng, W.C., Liao, H.J., Lin, Y.C., Sun, M.: Cla-nerf: Category-level articulated neural radiance field. In: ICRA (2022)", + "41. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017), https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf", + "42. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS (2021)", + "43. Wei, F., Chabra, R., Ma, L., Lassner, C., Zollhoefer, M., Rusinkiewicz, S., Sweeney, C., Newcombe, R., Slavcheva, M.: Self-supervised neural articulated shape and appearance models. In: Proceedings IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "44. Wu, T., Zhong, F., Tagliasacchi, A., Cole, F., Oztireli, C.: D\\~nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 32653-32666. Curran Associates," + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation", + "bbox": [ + 261, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Inc. (2022), https://proceedings.neurips.cc/paper_files/paper/2022/file/d2cc447db9e56c13b993c11b45956281-Paper-Conference.pdf", + "45. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time neural irradiance fields for free-viewpoint video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9421-9431 (2021)", + "46. Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: SAPIEN: A simulated part-based interactive environment. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)", + "47. Yang, G., Sun, D., Jampani, V., Vlasic, D., Cole, F., Chang, H., Ramanan, D., Freeman, W.T., Liu, C.: Lasr: Learning articulated shape reconstruction from a monocular video. In: CVPR (2021)", + "48. Yang, G., Sun, D., Jampani, V., Vlasic, D., Cole, F., Liu, C., Ramanan, D.: Viser: Video-specific surface embeddings for articulated 3d shape reconstruction. In: NeurIPS (2021)", + "49. Yang, G., Vo, M., Neverova, N., Ramanan, D., Vedaldi, A., Joo, H.: Banmo: Building animatable 3d neural models from many casual videos. In: CVPR (2022)", + "50. Yang, J., Pavone, M., Wang, Y.: Freenerf: Improving few-shot neural rendering with free frequency regularization (2023)", + "51. Yao, C.H., Hung, W.C., Li, Y., Rubinstein, M., Yang, M.H., Jampani, V.: Lassie: Learning articulated shape from sparse image ensemble via 3d part discovery. In: NeurIPS (2022)", + "52. Yao, C.H., Hung, W.C., Li, Y., Rubinstein, M., Yang, M.H., Jampani, V.: LASSIE: Learning Articulated Shape from Sparse Image Ensemble via 3d part discovery. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)", + "53. Yuan, W., Lv, Z., Schmidt, T., Lovegrove, S.: Star: Self-supervised tracking and reconstruction of rigid objects in motion with neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13144-13152 (2021)" + ], + "bbox": [ + 215, + 147, + 785, + 547 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "A. Swaminathan et al.", + "bbox": [ + 271, + 114, + 421, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/LEIA_ Latent View-invariant Embeddings for Implicit 3D Articulation/0e9336bc-d9b2-471b-94f5-34342577a604_model.json b/2024/LEIA_ Latent View-invariant Embeddings for Implicit 3D Articulation/0e9336bc-d9b2-471b-94f5-34342577a604_model.json new file mode 100644 index 0000000000000000000000000000000000000000..29e904f276933f5e45de90c7db26f455ffb2d27e --- /dev/null +++ b/2024/LEIA_ Latent View-invariant Embeddings for Implicit 3D Articulation/0e9336bc-d9b2-471b-94f5-34342577a604_model.json @@ -0,0 +1,2216 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.245, + 0.141, + 0.758, + 0.186 + ], + "angle": 0, + "content": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.213, + 0.77, + 0.244 + ], + "angle": 0, + "content": "Archana Swaminathan1, Anubhav Gupta1, Kamal Gupta1, Shishira R. Maiya1, Vatsal Agarwal1, and Abhinav Shrivastava1" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.256, + 0.682, + 0.283 + ], + "angle": 0, + "content": "University of Maryland, College Park MD 20742, USA https://archana1998.github.io/leia/" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.322, + 0.74, + 0.571 + ], + "angle": 0, + "content": "Abstract. Neural Radiance Fields (NeRFs) have revolutionized the reconstruction of static scenes and objects in 3D, offering unprecedented quality. However, extending NeRFs to model dynamic objects or object articulations remains a challenging problem. Previous works have tackled this issue by focusing on part-level reconstruction and motion estimation for objects, but they often rely on heuristics regarding the number of moving parts or object categories, which can limit their practical use. In this work, we introduce LEIA, a novel approach for representing dynamic 3D objects. Our method involves observing the object at distinct time steps or \"states\" and conditioning a hypernetwork on the current state, using this to parameterize our NeRF. This approach allows us to learn a view-invariant latent representation for each state. We further demonstrate that by interpolating between these states, we can generate novel articulation configurations in 3D space that were previously unseen. Our experimental results highlight the effectiveness of our method in articulating objects in a manner that is independent of the viewing angle and joint configuration. Notably, our approach outperforms previous methods that rely on motion information for articulation registration." + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.586, + 0.733, + 0.599 + ], + "angle": 0, + "content": "Keywords: Articulated Objects \\(\\cdot\\) Neural Radiance Fields \\(\\cdot\\) 3D Vision" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.627, + 0.375, + 0.642 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.659, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Our world is full of dynamic objects moving and interacting in space and time. Humans develop this (rather impressive) understanding of how these everyday objects move and interact in three-dimensional space at a very early stage of the brain development [31]. The task involves understanding not only the static geometry but also the dynamic movements and spatial relationships between parts of an object, often referred to as articulations of the object. Understanding and representing object articulations from images and/or videos is also pivotal in enabling machines to perceive and navigate the physical world with finesse. In this work, we propose a novel method to model the object articulations by learning view-invariant latent embeddings of the 3D object from multiview images. Existing works in modeling object articulation in the three dimensional space typically use priors in the form of pre-trained large scale models [51], videos [27]," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.424, + 0.128 + ], + "angle": 0, + "content": "A. Swaminathan et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.149, + 0.77, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.316, + 0.789, + 0.375 + ], + "angle": 0, + "content": "Fig. 1: Our method LEIA, takes in multi-view images of an object in four articulation states and is able to learn a view-invariant latent embedding for the state. We show that we can interpolate between the latents to generate any number of intermediate unseen states for the object using LEIA, given the camera position." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.789, + 0.633 + ], + "angle": 0, + "content": "or assumptions regarding rigidity/shape of the object [17]. These approaches often fail to generalize for the long-tail of objects, especially in the case when a video of the object with articulation is unavailable, or in presence of large and multiple articulations in the object. In our work, we postulate that multiview images of an object in different states can provide enough signal to model its 3D shape and articulation, even without priors from large-scale foundation models or information from videos. We achieve this by learning a generalizable, view-invariant latent embedding of the object for different states. We train these embeddings jointly with a hypernetwork that predicts weights of a NeRF [19] parameterizing the object state. The hypernetwork can be trained using multiview images of discretized states of the object, each state representing a different object articulation. In Figure 1, we show three objects from our dataset in two different articulation states. The latent embeddings representing each state or articulation can be interpolated during test time to generate the weights of a NeRF that is able to reconstruct a never seen before state of the object." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.635, + 0.641, + 0.651 + ], + "angle": 0, + "content": "Our key contributions can be summarized as following:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.653, + 0.785, + 0.681 + ], + "angle": 0, + "content": "- We introduce an end-to-end method LEIA for generating novel states for articulated objects solely with multiview images captured at multiple states." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.686, + 0.785, + 0.758 + ], + "angle": 0, + "content": "- We demonstrate that interpolating between embeddings can generate states of articulations of object not seen during training. The embedding space becomes interpolable with a manifold loss that encourages the latents to follow a structure that establishes a linear relationship between them, by minimizing the distance between the nearest neighbours in the latent space." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.763, + 0.784, + 0.805 + ], + "angle": 0, + "content": "- Remarkably, LEIA achieves this without the need for any ground-truth 3D supervision, motion information, or articulation codes, establishing its versatility and effectiveness in capturing complex articulations." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.81, + 0.784, + 0.838 + ], + "angle": 0, + "content": "- Our analyses demonstrate LEIA's robustness to single and multiple articulations, as well as combinations of motions. We can disentangle articulations" + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.653, + 0.785, + 0.838 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.264, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "in different object parts if multiple are movable, making it scalable without constraints on the number of parts or motion types, unlike prior work." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.202, + 0.387, + 0.218 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.236, + 0.786, + 0.417 + ], + "angle": 0, + "content": "Neural Radiance Fields. Neural Radiance Fields [1, 19, 21, 42] have proven to be a massive success in modeling 3D scenes, due to the high fidelity of 3D reconstruction and novel view synthesis for static scenes. For dynamic scenes, Neural Volumes [18] utilizes an encoder-decoder voxel-based representation, complemented by an implicit voxel warp field. Occupancy Flow [24] tackled non-rigid geometry by assigning a motion vector to each point in both space and time, but requires full 3D ground truth supervision. Some of the first dynamic NeRF approaches [26, 28, 39] optimize an underlying volumetric deformable function and [15] conditions the NeRF on time. [16, 26, 45] followed this work, by learning a 5D spatiotemporal neural field. However, the approaches above usually require a video as input and do not handle well the case of large articulations in everyday objects given multiple states of objects as input." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.418, + 0.786, + 0.84 + ], + "angle": 0, + "content": "3D Representations for Articulation. Due to the fine-grained nature of the task, deep neural network models for representing articulation require conditioning or refining parts of the architecture to suit the task. Category-specific reconstruction of deformable objects from images [7, 11-13], sparked interest in identifying and recovering the deformation in the 3D space [13, 22, 52]. Several works focus on shape reconstruction from videos [38, 47-49] that estimate a shape template for humans and animals. Other works that learn articulation from videos include Qian et al. [29] that detects and segments articulation planes from in-the-wild videos and [44, 53] which decouple the static and dynamic parts of videos. These works focus on modeling humans and animals where a large amount of data is available. Preliminary work in using images and shape started with A-SDF [28], which generates unseen articulations using Signed Distance Function [25], by providing an input shape and articulation code. CaDeX [14] is a unified representation for shape and motion, obtained from a point cloud input. Ditto [10] is a similar work that uses implicit representations for joint geometry and articulation modeling, with ground truth point clouds. CLA-NeRF [40] learns unseen articulated states from observing multi-view image input along with articulation information. Wei et. al [43] obtains an SDF-based articulable representation of common objects by feeding in images of articulated states across multiple categories, and CARTO [9] uses stereo images as input and uses a stereo encoder to infer the 3D shape, 6D pose etc. of multiple unknown objects. [4] focuses on manipulating object shapes to deform according to specified articulation commands. Moving away from 3D supervision and articulation annotations, Jiayi et al. [17] proposed PARIS, a method that is able to obtain unseen articulated states of an object, given multi-view images of just the start and end state of articulation. PARIS employs a composite rendering based approach, by decoupling the object into a static and moving part and then separately estimating and compositing the static and mobile neural radiance fields. In our work," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.424, + 0.129 + ], + "angle": 0, + "content": "A. Swaminathan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.229 + ], + "angle": 0, + "content": "Table 1: Comparison of LEIA with existing methods. We show that LEIA is the first approach that does not use or learn any explicit prior along with not having any articulation input. This gives us flexibility in scaling to modeling articulations of objects with more than one part, and can thus handle a wide range of motion. We also have one universal model that can learn to represent both prismatic and revolute motion, unlike PARIS that has two separate models. 3D Sup. refers to \"3D Supervision\"." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.24, + 0.789, + 0.334 + ], + "angle": 0, + "content": "
MethodImage/Shape InputDatasetArticulation Input3D Sup.?Prior# States# Parts
A-SDF [28]Shape CodeShape2MotionArticulation Code✓Articulation≥4≥1
CLA-NeRF [40]Multiview ImagesPartNet-MobilityArticulated PoseXArticulation≥41
NASAM [43]Multiview Images++PartNet-MobilityN/AXCategory≥4≥1
CARTO [9]Stereo ImagesCustomJoint Code✓Articulation≥41
Ditto [10]Point CloudsShape2MotionAnnotations✓Articulation21
PARIS [17]Multiview ImagesPartNet-MobilityN/AXMotion (learned)21
LEIAMultiview ImagesPartNet-MobilityN/A✓No4≥1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.789, + 0.483 + ], + "angle": 0, + "content": "we start LEIA with a similar input setting, by preparing multi-view renderings of objects in different articulation states. Unlike PARIS, we do not decouple the object into static and moveable parts, we rather learn to predict the weights of a single NeRF for any unseen articulated state using a state-modulated HyperNetwork. We aren't limited by the decoupling and the learned motion prior in the case of PARIS, which enables us to scale and learn any amount and kind of motion an object can possibly have, including combinations. We show a summed comparison of LEIA with prior work in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.484, + 0.789, + 0.755 + ], + "angle": 0, + "content": "Hypernetworks with Implicit Neural Representations. Hypernetworks, a specialized class of networks designed to predict parameters for another network, aim to achieve generalization across novel tasks [8]. In the realm of stylizing 3D scenes, [3] employed a hypernetwork for applying diverse styles, while scene reconstructions from limited data points were achieved by [35, 36]. Despite the promise shown in representation, these hypernetworks operate on input data points, necessitating test-time optimizations and rendering them unsuitable for compression tasks. Rather than using the provided data (image/video) as input for the hypernetwork, an alternative approach involves employing an auto-decoder framework. In this framework, a learnable latent, without the need for an encoder, represents a data point. This technique, applied by [33] to represent a dataset of videos, assigns each latent to a distinct video. While this method yields a representation for each set of frames, the lack of decoupling in spatial-temporal coordinates limits its scalability to real-world frames. On a related note, [34] extended a similar approach to 3D shapes and scenes, effectively acquiring a latent representation suitable for various downstream tasks. We use a similar framework for learning the latents corresponding to each articulated state in our method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.778, + 0.333, + 0.794 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "**Background.** Our architecture is based on neural radiance fields, or NeRF [19], which parameterizes the radiance and volume density at a 3D location of a" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.263, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.357 + ], + "angle": 0, + "content": "scene as observed from a camera placed in a particular position using a neural network. \\( F_{\\mathrm{NeRF}}:(\\mathbf{x},\\mathbf{d}) \\to (\\mathbf{c},\\sigma) \\), where \\( \\mathbf{c} = (r,g,b) \\), and \\( \\sigma \\) represent the radiance and volume density, and \\( \\mathbf{x} \\) and \\( \\mathbf{d} \\) represent the 3D location and viewing direction respectively. To render a pixel, the radiance \\( C(\\mathbf{r}) \\) of a camera ray \\( \\mathbf{r}(s) = \\mathbf{o} + s \\cdot \\mathbf{d} \\) is integrated from near to far bounds \\( s_n \\) and \\( s_f \\) such that, \\( C(\\mathbf{r}) = \\int_{s_n}^{s_f} T(s)\\sigma(\\mathbf{r}(s))\\mathbf{c}(\\mathbf{r}(s),\\mathbf{d})ds \\), where the function \\( T(s) \\) denotes the accumulated transmittance from \\( s_n \\) to \\( s \\). To optimize the parameters of the MLP, a loss function is used that measures the discrepancy between the ground-truth and rendered images. Traditionally, the L2 loss is used for this purpose. However, to make the training more robust to outliers and to improve convergence, the Smooth L1 Loss can be employed as an alternative. The Smooth L1 Loss is a combination of L1 and L2 losses, behaving like L1 loss for large errors and like L2 loss for small errors. The loss function for the NeRF model using the Smooth L1 Loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.362, + 0.786, + 0.401 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S m o o t h L 1 L o s s} (x) = \\left\\{ \\begin{array}{l l} 0. 5 \\cdot x ^ {2} & \\text {i f} | x | < 1 \\\\ | x | - 0. 5 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.409, + 0.785, + 0.441 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {S m o o t h L 1 - N e R F}} = \\sum_ {r \\in R} \\text {S m o o t h L 1 L o s s} \\left(\\hat {C} (\\mathbf {r}) - C (\\mathbf {r})\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.614 + ], + "angle": 0, + "content": "where \\( R \\) is the set of rays used for sampling, and \\( \\hat{C}(\\mathbf{r}) \\) and \\( C(\\mathbf{r}) \\) represent the ground-truth and rendered colors, respectively. By minimizing this loss function, the NeRF model learns to accurately reproduce the radiance of the scene, leading to high-quality image synthesis from novel viewpoints. While this formulation works well for generating novel scene viewpoints for a static scene, it cannot handle dynamic scenes. [45] tried to address this shortcoming by making the MLP learn a spatiotemporal radiance field and used time, \\( t \\), as an additional input. This works in principle but it becomes expensive to scale to lengthier videos due to constantly increasing sampling space for NeRF. We use the \\( L_{\\mathrm{SmoothL1 - NeRF}} \\) loss along with AdamW optimization to train LEIA, along with a \\( L_{\\mathrm{mask}} \\) that is the BCE loss between the predicted opacity and ground truth foreground loss." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.631, + 0.345, + 0.646 + ], + "angle": 0, + "content": "3.1 Approach" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.651, + 0.787, + 0.711 + ], + "angle": 0, + "content": "Hypernetworks, or hypernets for short, are neural networks that generate weights for another neural network, known as the target network. Hypernets can be conditioned on various domains and can predict the weights for multiple neural networks simultaneously, if trained appropriately." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.788, + 0.803 + ], + "angle": 0, + "content": "In this work, we use a hypernet to modulate a Neural Radiance Field (NeRF) based on the state of articulation, thereby learning a parametrization for each state. We employ a learnable latent embedding, \\( Z \\), as input to this hypernet \\( h_l, l \\in L \\) where \\( L \\) is the number of layers of the hypernet. This approach allows us to not only modulate the NeRF but also create useful representations for each state. Our system can thus be represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.809, + 0.786, + 0.834 + ], + "angle": 0, + "content": "\\[\nF _ {\\theta_ {t}} (\\mathbf {x}, \\mathbf {d}) = \\left(\\mathbf {c} _ {t}, \\sigma_ {t}\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.465, + 0.829, + 0.59, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\theta_ {t} = h (z _ {t}), z _ {t} \\in Z\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.424, + 0.128 + ], + "angle": 0, + "content": "A. Swaminathan et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.79, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.317, + 0.789, + 0.402 + ], + "angle": 0, + "content": "Fig. 2: Overview of our method. We take multi-view images in different states as input. A learnable latent dictionary based off an autoencoder learns an embedding per state id. The latent embedding is used as an input to the hypernet, that modulates and generates weights of the NeRF to reconstruct the state that is fed in. At inference time, we do a weighed interpolation of the learnt latents to obtain a corresponding newly generated intermediate state." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.789, + 0.492 + ], + "angle": 0, + "content": "The latents \\( Z \\), give us an additional ability to interpolate across seen states and generate novel states of the object, as we show in our work. The set of latents, \\( Z \\), in Eq. (3) can be understood as a learnable dictionary where keys are the state ids. Formally, we represent this as" + }, + { + "type": "equation", + "bbox": [ + 0.396, + 0.503, + 0.787, + 0.521 + ], + "angle": 0, + "content": "\\[\nZ = \\left\\{t: z _ {t} \\mid t \\in [ 0, 1, 2 \\dots T ] \\right\\} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.789, + 0.621 + ], + "angle": 0, + "content": "where, \\( T \\) is the total number of discretized states sampled for the object. Each of these states, \\( z_{t} \\), is used as an input to the hypernet to get the parameterization for NeRF as shown in Figure 2. Once parameterized, the NeRF is trained using multi-view images of the object taken from various camera angles, providing a comprehensive view of the object's articulation in the current state. We sample only one state per batch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.789, + 0.713 + ], + "angle": 0, + "content": "Directly predicting the weights \\(\\theta\\) of the base network \\(f_{\\theta}\\), using the hypernet \\(h_l\\), is expensive, parameter-heavy, and unsuitable for compression. Hence, we follow [32, 37] and instead predict low-rank matrices, which are then applied to the base network weights. This type of modulation acts as a form of subnetwork selection, analogous to systems proposed in [6, 30]. For a base network \\(f_{\\theta}\\) with \\(L\\) layers, our formulation now looks like" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.722, + 0.576, + 0.741 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta} \\left(\\left(\\mathbf {x}, \\mathbf {d}\\right) \\mid \\theta_ {t} ^ {l _ {1}}, \\theta_ {t} ^ {l _ {2}} \\dots \\theta_ {t} ^ {l _ {L}}\\right) = \\mathbf {c} _ {t}, \\sigma_ {t}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.498, + 0.743, + 0.785, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\theta_ {t} ^ {l} = \\eta \\left(P ^ {l} \\times Q ^ {l}\\right) \\circ \\theta^ {l} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.472, + 0.764, + 0.591, + 0.783 + ], + "angle": 0, + "content": "\\[\nh _ {l} (z _ {t}) = [ P ^ {l}, Q ^ {l} ]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.794, + 0.789, + 0.842 + ], + "angle": 0, + "content": "where \\(\\theta^l\\) represents the weights of the \\(l\\)-th layer and \\(\\theta_t^l\\) denotes the modulated weights for frame \\(t\\). Here, \\(\\eta\\) signifies an activation function on the matrix-product of low rank matrices \\(\\mathbf{P}^l \\in \\mathbb{R}^{K \\times r}\\), \\(\\mathbf{Q}^l \\in \\mathbb{R}^{r \\times K}\\), where \\(K\\) is the width of the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.263, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "base network \\( f_{\\theta} \\) and rank \\( r \\ll K \\). These matrices are responsible for adjusting the weights \\( \\theta_{l} \\) as dictated by the corresponding hypernetwork \\( h_l \\). Note that all hypernetworks use the same latent \\( z_{t} \\in \\mathbb{R}^{D} \\) as input. The rank \\( r \\) acts as a hyperparameter that controls the compression-performance trade-off. We further elaborate on details about model architecture and design choices for the hypernet, NeRF and the learnable latent dictionary in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.26, + 0.373, + 0.275 + ], + "angle": 0, + "content": "3.2 Interpolation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.285, + 0.789, + 0.377 + ], + "angle": 0, + "content": "Given an object with states \\( t_1 \\) and \\( t_2 \\), and a scale \\( \\alpha \\), the task of state interpolation involves creating \\( \\alpha - 1 \\) coherent states between the given states. In order to achieve this, we do a linear interpolation on the state latents \\( z_1 \\) and \\( z_2 \\) and pass the resulting latent through the hypernetwork. This gives us the weight modulation required in the NeRF, and the updated base network is used to obtain different viewpoints of the intermediate state." + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.389, + 0.786, + 0.411 + ], + "angle": 0, + "content": "\\[\nz _ {\\text {i n t e r}} = \\left(1 - \\beta_ {i}\\right) \\cdot z _ {t} + \\beta_ {i} \\cdot z _ {t - 1} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.408, + 0.619, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {\\text {i n t e r}}, \\sigma_ {\\text {i n t e r}} = f _ {\\theta_ {\\text {i n t e r}}} (\\mathbf {x}, \\mathbf {d}; h (z _ {\\text {i n t e r}}))\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.396, + 0.436, + 0.785, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e}, \\beta_ {i} \\in \\left[ \\frac {1}{\\alpha}, \\frac {2}{\\alpha}, \\dots , \\frac {\\alpha - 1}{\\alpha} \\right] \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.788, + 0.534 + ], + "angle": 0, + "content": "essentially generating \\(\\alpha - 1\\) states between any two given states. In our experiments, we quantitatively evaluate the result of the interpolation with unseen ground truth obtained from our dataset of all states, by linearly averaging the two extreme states present in the training, to obtain a latent for a state id." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.557, + 0.427, + 0.57 + ], + "angle": 0, + "content": "3.3 Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.789, + 0.747 + ], + "angle": 0, + "content": "In our experiments, both the base network \\( f_{\\theta} \\) and hypernetworks \\( h_l \\) are simple MLPs that take in a coordinate input, and an id depicting the state of articulation. A network \\( l_n \\) is used to learn the latent dictionary for the states, which is a linear layer learnt in conjunction with the hypernet, and uses the torch.nn.Embedding class as a lookup table for the learnable latent embeddings, depicted by \\( z_t \\) for \\( t \\in T \\). This latent is fed to the hypernetwork \\( h_l \\), that modulates the weights of the base network \\( f_{\\theta} \\) for reconstructing the output of the corresponding state \\( \\mathrm{id} \\) \\( t \\). The base network \\( f_{\\theta} \\) is based off the NeRF architecture of Instant-NGP [21]. This architecture has separate fully-connected blocks for the geometry and texture of the scene, and we use separate hypernets to modulate each layer of these two FC blocks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.77, + 0.637, + 0.785 + ], + "angle": 0, + "content": "3.4 Optimization: Loss functions and regularizers" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Latent Manifold Loss. We use the latent manifold loss function that enforces a structured latent space by encouraging local consistency among learnt latents. For each latent vector, the loss computes the average Euclidean distance between" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.424, + 0.128 + ], + "angle": 0, + "content": "A. Swaminathan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "the vector and its \\( K \\) nearest-neighbours on the manifold. This process enforces a smooth and continuous latent manifold, which is beneficial for models that rely on meaningful linear interpolations between the points on the manifold, and for tasks where the geometry of the latent space is crucial. Mathematically, we represent this loss for a particular state-id \\( i \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.238, + 0.786, + 0.278 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a n i f o l d}} \\left(\\mathbf {l} _ {i}\\right) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\| \\mathbf {l} _ {i} - \\mathbf {n} _ {k} \\| _ {2} ^ {2} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.291, + 0.789, + 0.351 + ], + "angle": 0, + "content": "where \\(\\mathbf{n}_k\\) are the \\(K\\) nearest neighbors of \\(\\mathbf{l}_i\\) in the latent space, and \\(\\| \\cdot \\|_2\\) denotes the Euclidean (L2) norm. The loss is averaged over the selected latents and their nearest neighbors to ensure local uniformity in the manifold's geometry, where \\(K\\) is a hyperparameter that determines the number of neighbors considered." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.353, + 0.79, + 0.489 + ], + "angle": 0, + "content": "Depth and Occlusion Regularization. Our depth and occlusion regularizations are designed to refine the clarity of rendered images by addressing occlusion and depth smoothness [23, 50]. The occlusion regularization loss, \\( L_{\\mathrm{occ}} \\), aims to mitigate the obscuring of objects located beyond a specified depth threshold during rendering. This is accomplished by generating a binary mask, \\( m_k \\), where the mask elements are set to 1 up to a certain index \\( M \\) reflecting the regularization range, and 0 thereafter. The occlusion loss is then articulated as the normalized sum of the product of the mask and the sampled density values \\( \\sigma_k \\) along a ray, as given by" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.492, + 0.787, + 0.533 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {o c c}} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\sigma_ {k} \\cdot m _ {k}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.587 + ], + "angle": 0, + "content": "where \\(K\\) is the total number of sampled points on the ray. This formulation drives the model to prefer representations that reduce occlusions close to the camera, ensuring objects further away are not improperly concealed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.588, + 0.789, + 0.68 + ], + "angle": 0, + "content": "For depth continuity, the depth smoothness regularization loss, \\(L_{\\mathrm{DS}}(\\theta ,\\mathbf{R})\\) enforces the gradual transition of depth values among neighboring pixels, reducing sharp depth disparities that cause visual inconsistencies. If a ray \\(r\\) intersects with a single depth value, this value is evaluated directly. In contrast, for multiple depth values, the loss is the aggregate of squared differences between adjacent depth estimates. Formally, for depth predictions \\(\\hat{d}_0(r)\\), the loss is quantified as" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.695, + 0.786, + 0.75 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {D S}} (\\theta , \\mathbf {R}) = \\sum_ {r \\in \\mathbf {R}} \\sum_ {i, j = 1} ^ {\\text {P a t c h} - 1} \\left(\\hat {d} _ {0} \\left(r _ {i j}\\right) - \\hat {d} _ {0} \\left(r _ {(i + 1) j}\\right)\\right) ^ {2} + \\left(\\hat {d} _ {0} \\left(r _ {i j}\\right) - \\hat {d} _ {0} \\left(r _ {i (j + 1)}\\right)\\right) ^ {2}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This measure is averaged across the rays to yield a measure of the depth map's smoothness. By integrating these regularizers into our training regimen, we significantly dampen the disturbances introduced by overlapping and obscured pixels, resulting in a more consistent interpolation between states. These benefits are substantiated through extensive ablation studies and quantitative analyses presented in the experiments." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.263, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "Positional Encoding of the Latent. We incorporate positional encoding to capture the order of input elements, crucial for understanding articulation. Positional encoding injects the sequence with its inherent order, a key factor in articulation semantics. We employ the scheme from [41], which uses sine and cosine functions parameterized to encode varying frequencies. The computed positional encodings are added to the latent vector \\( z_{t} \\), enriching it with semantic information that more accurately encodes the state of articulation. We analyze the effects of positional encoding on our model in the experimental section." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.298, + 0.377, + 0.315 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.336, + 0.313, + 0.352 + ], + "angle": 0, + "content": "4.1 Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.789, + 0.64 + ], + "angle": 0, + "content": "Datasets. In this work, we use the PartNet-Mobility dataset [2, 20, 46], a large-scale synthetic collection of articulated objects in over 40 categories. The dataset has a variety of articulations defined, with objects comprising of single and multi-joint parts. Two types of motion, revolute (rotational) joint and prismatic (translation) articulation are represented in the dataset. To make up our training dataset, we choose 100 camera views that are arranged in a dome-like setup, capturing the upper hemisphere of the object. This is similar to the setup used by PARIS. We use the SAPIEN [46] library that the PartNet-Mobility dataset was released with, to render our RGB images at linearly spaced intervals of articulation to make up frames of a video that depicts the range of motion and obtain the camera parameters accordingly, converting it into the Blender [5] coordinate system to fit in our codebase. We use instances from 8 different common household item categories storage, microwave, laptop, oven, washer, dishwasher, sunglasses, box and choose 1-4 objects per category, bringing our total number of objects to 12. We also show the efficacy of our method on 68 images from a real-world scene of a chest of drawers, which are captured with a mobile phone and post-processed to remove background. Throughout LEIA, we train with a total of four states and interpolate between the two extreme states." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.789, + 0.732 + ], + "angle": 0, + "content": "Baselines. The closest prior work with a setup similar to ours is PARIS, which takes in multi-view images of objects in two states and disentangles the static and moving part of the object. While we run experiments with four input states, we don't focus on learning a defined articulation for the object, instead we are able to recognize any arbitrary motion and combinations of motions given relevant states. We emphasize on the following differences:" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.733, + 0.786, + 0.778 + ], + "angle": 0, + "content": "1. PARIS uses separate models to train objects that have rotation and translation motion, respectively. While they have a method to estimate the type of motion, once determined, it is necessary to train on the appropriate model." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.78, + 0.786, + 0.84 + ], + "angle": 0, + "content": "2. As PARIS also learns motion parameters for the articulation, given their pipeline of disentanglement of the static and moving parts of the object, this restricts the learning of motion parameters if there are multiple parts in the object, moving differently." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.733, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.128 + ], + "angle": 0, + "content": "A. Swaminathan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.271 + ], + "angle": 0, + "content": "Table 2: Quantitative Results for Interpolated State Reconstruction. We compared our method with the PARIS baseline, trained on selected objects from the SAPIEN dataset. The results from three experiments are summarized below. The VanillaInt experiment involves simple interpolation of the latents. Our best-performing method, LEIA, introduces structure to the latents with a manifold loss and regularizers. Although PARIS learns a motion prior and LEIA implicitly performs state interpolation, both methods perform similarly for single-part objects. However, our approach excels with multi-part objects, outperforming PARIS significantly due to its flexibility in handling various motion types and articulations without constraints." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.282, + 0.789, + 0.436 + ], + "angle": 0, + "content": "
MetricsMethodsSingle-Part ArticulationMulti-Part ArticulationAverage
45135 Storage17128 Microwave10211 Laptop101917 Oven103778 Washer12085 Dishwasher44781 Storage245427 Storage345575 Storage4101297 Sunglasses7187 Oven102377 Box
PSNR↑PARIS28.6625.9424.9728.1334.4627.2426.3524.4125.7332.3329.3026.2027.81
VanillaInt24.2022.0022.6324.5336.0324.1630.9629.7829.1032.1130.3527.8527.81
LEIA26.6926.3825.0428.7136.1426.4931.0729.7829.8035.6030.8028.0529.55
SSIM↑PARIS0.990.970.970.970.980.960.950.950.940.960.960.930.96
VanillaInt0.950.910.900.930.990.920.950.950.930.970.950.930.94
LEIA0.970.950.950.980.990.950.960.950.950.970.960.950.96
LPIPS↓PARIS0.020.060.190.030.030.060.050.050.040.060.050.080.06
VanillaInt0.050.100.170.030.020.080.030.030.050.090.050.130.07
LEIA0.030.070.100.020.020.070.030.030.040.090.050.120.06
CD↓PARIS0.180.080.120.750.060.450.620.420.600.200.990.910.45
VanillaInt0.240.070.060.100.060.200.480.470.460.970.500.830.37
LEIA0.290.060.360.100.050.270.520.380.410.960.350.620.36
" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.465, + 0.787, + 0.51 + ], + "angle": 0, + "content": "3. LEIA is motion-prior free, and we have a universal architecture that implicitly learns intermediate states, so we are able to train with multiple types of motion occurring in the same object, with no change in the code." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.789, + 0.648 + ], + "angle": 0, + "content": "For comparing with PARIS, we run authors' original source code with our dataset, using the hyperparameters provided by the authors and training until convergence. We test our work and PARIS with the appearance quality of the reconstructed image of the interpolated state. We also set up a simple baseline that does just vanilla interpolation between the learnt latent embeddings, without enforcing any structure or constraints on them. We call this baseline VanillaInt and compare it with LEIA, which has been finetuned to do interpolation with the addition of the manifold loss and the depth and occlusion regularization for denoising the resulting output." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.648, + 0.789, + 0.74 + ], + "angle": 0, + "content": "Quantitative Metrics. We use three appearance quality metrics, PSNR (Peak Signal-to-Noise Ratio), SSIM (Structural Similarity Index), and LPIPS (Learned Perceptual Image Patch Similarity). These are computed between reconstructed images of the interpolated state, and the unseen ground truth of the state from our dataset. We also report the Chamfer Distance to measure quality of 3D reconstruction by doing point sampling." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.766, + 0.373, + 0.78 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Novel State Synthesis. We show the qualitative and quantitative results of interpolating our learned latent embeddings corresponding to the start and end states, in Table 2 and Figure 3. Our results show that the latent embeddings can" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.264, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.144, + 0.734, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.682, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Fig. 3: Qualitative Results. We show results of PARIS and LEIA for reconstructing the unseen intermediate state, for both single and multiple articulations. We see that PARIS especially fails when there are two parts of the object moving differently, as the motion parameters are not registered correctly. LEIA handles this case successfully as it is not dependent on part disentanglement to identify and register articulation. LEIA also performs comparable to PARIS for single-part articulation, despite not having a dedicated model for the motion or part disentanglement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "be linearly interpolated, and have a structure in the latent space. We can generate any number of states between the start and the end state by doing a weighted" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.128 + ], + "angle": 0, + "content": "A. Swaminathan et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.158, + 0.756, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.244, + 0.788, + 0.288 + ], + "angle": 0, + "content": "Fig. 4: Real World Results. LEIA is able to faithfully interpolate and reconstruct between two states of images from our real world data, proving its ability to generalize and work in an in-the-wild setting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.318, + 0.789, + 0.544 + ], + "angle": 0, + "content": "combination of the latents representing the start and the end states. For training, we choose four states, using the two additional states compared to PARIS to enforce structure in the latent space with the latent manifold loss, thereby making them amenable to interpolation. This helps us extend our formulation to multiple articulations without restriction. We also show results of our baseline VanillaInt and show that the numbers are boosted with the addition of the latent manifold loss and regularizers. For single-part objects, we perform comparably to PARIS despite having no part disentanglement and no dedicated model to represent the exact type of motion. For multiple-part articulations, we beat PARIS across all objects, as shown in Figure 3. Our comparable performance in 3D reconstruction shows that LEIA's learnt view-invariant latent embeddings are powerful in preserving spatio-temporal consistency. We also show that our method can work in the real world setting, in Figure 4. Ablations are shown for number of states, latent manifold loss, and depth and occlusion regularizations in Table 3 and Figure 6." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.789, + 0.683 + ], + "angle": 0, + "content": "Analysis on Disentanglement of Joints. To check separability of the latent space, we train LEIA with multi-view images across multiple states, captured from a video that showed three parts of an object Storage1, move one after the other. We do a t-SNE based dimensionality reduction on the embeddings and plot them, shown in Figure 5. The latents exhibit clear separability, with the motion of each joint represented in a smooth trajectory, moving outward as the respective drawer in the object moves away from the starting \"closed\" state. This shows that the separated latent space is an indicator of multiple joints, positioning us to be capable of learning representations of multiple articulations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.707, + 0.41, + 0.722 + ], + "angle": 0, + "content": "4.3 Ablation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We perform ablations to tune various design choices, with results shown in Table 3 and Figure 6. The first experiment we performed was testing out the impact of the latent manifold loss. We notice qualitatively that without the manifold loss, our latent space lacked structure and the linear relationship between the embeddings was not captured. As shown in Figure 6, the intermediate state overfits to the extreme state without the latent manifold while the addition of the loss enables it to accurately capture the intermediate state. This figure also" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.264, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.151, + 0.454, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.334, + 0.476, + 0.417 + ], + "angle": 0, + "content": "Fig. 5: t-SNE plot. After dimensionality reduction on jointly-learned state embeddings of an object with different moving parts. Our learned representations are separated and follow a smooth trajectory for each of the moving parts." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.148, + 0.788, + 0.217 + ], + "angle": 0, + "content": "Table 3: Ablations. Our ablations reveal that latent manifold loss, depth, and occlusion regularization enhance LEIA's visual metrics. Opting for four states improved latent structure with the manifold loss." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.23, + 0.782, + 0.427 + ], + "angle": 0, + "content": "
Latent Manifold LossPSNRSSIMLPIPS
with29.400.950.05
without28.540.940.06
Depth RegularizationPSNRSSIMLPIPS
with29.630.960.05
without26.930.930.07
Occlusion RegularizationPSNRSSIMLPIPS
with29.640.950.05
without28.640.950.06
Positional encodingPSNRSSIMLPIPS
with27.110.940.06
without28.480.950.06
Number of StatesPSNRSSIMLPIPS
228.040.950.06
429.690.960.05
" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.453, + 0.769, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.704, + 0.788, + 0.773 + ], + "angle": 0, + "content": "Fig. 6: Qualitative Results of Ablations. We show some qualitative results of using latent manifold loss, positional encoding and varying the number of states. Using the manifold loss prevents us from overfitting to the extreme states, and using four states help us interpolate with a lot of clarity. Positional encoding helps add missing information for thin parts, but for large objects it doesn't help much." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "shows the effect of positional encoding which adds detail when the latent embedding fails to capture it, as evidenced in the results of sunglasses. This can" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.423, + 0.128 + ], + "angle": 0, + "content": "A. Swaminathan et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.171, + 0.764, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.235, + 0.788, + 0.306 + ], + "angle": 0, + "content": "Fig. 7: Failure cases. Our model fails at reconstructing the geometry correctly at camera positions where the two states we interpolate between have a change in the visible shape, like these examples where the microwave and oven being closed shows a deformation in the figure as compared to the open state, as the relationship between the motion is not easily captured in the latent due to the structure change." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.788, + 0.471 + ], + "angle": 0, + "content": "happen for thin parts of the object that may not be well-captured across camera views. Notably, the inverse happens in the case of large objects, where the positional encoding adds extra noise to the reconstruction, resulting in oversmoothing. We also show how adding just two more intermediate states in LEIA makes a huge difference in reconstruction results for both single and multi-part objects, enabling us to scale and be flexible. Quantitative numbers for depth and occlusion regularization ablation are shown in Table 3. However, they didn't universally help in the cases where the latent embedding already learned the state representation well enough but reduced noise when it appeared." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.489, + 0.364, + 0.505 + ], + "angle": 0, + "content": "5 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.788, + 0.655 + ], + "angle": 0, + "content": "While LEIA works in achieving good quality reconstruction of intermediate states, our latent embeddings do not yet ensure accurate physical consistency in the motion of the intermediates, which is a tradeoff we chose while opting for not decoupling the object and learning the motion parameters for the moveable part. This allowed us to scale our approach to interpolate between multiple moving parts of the object, which would be more representative of the real world where common objects can be articulated in multiple ways. LEIA also struggles when there is severe occlusion, as referenced by Figure 7, as the occlusion occurs at gaps in the learnt embeddings between two states." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.675, + 0.36, + 0.691 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this work, we present LEIA, a method capable of successfully interpolating between two discrete articulation states of a deformable object with moving parts. LEIA handles multiple joints, including cases where moving parts are separated by static regions. It outperforms existing methods, especially when multiple parts move independently. The learned latent embeddings are view-invariant and separable, demonstrating LEIA's scalability and flexibility. We conducted comprehensive evaluations on synthetic and real data to investigate inherent challenges. Despite advancements, significant occlusion scenarios remain challenging. We hope future research builds upon our work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.263, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.298 + ], + "angle": 0, + "content": "This work was partially supported by NSF CAREER Award (#2238769) to Abhinav Shrivastava, and IARPA via Department of Interior/Interior Business Center (DOI/IBC) contract number 140D0423C0076. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of NSF, IARPA, DOI/IBC, or the U.S. Government." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.322, + 0.323, + 0.339 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.354, + 0.787, + 0.41 + ], + "angle": 0, + "content": "1. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5470-5479 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.41, + 0.787, + 0.451 + ], + "angle": 0, + "content": "2. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.452, + 0.787, + 0.493 + ], + "angle": 0, + "content": "3. Chiang, P.Z., Tsai, M.S., Tseng, H.Y., Lai, W.S., Chiu, W.C.: Stylizing 3d scene via implicit representation and hypernetwork. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1475-1484 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.493, + 0.787, + 0.548 + ], + "angle": 0, + "content": "4. Chu, R., Liu, Z., Ye, X., Tan, X., Qi, X., Fu, C.W., Jia, J.: Command-driven articulated object understanding and manipulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8813-8823 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.549, + 0.787, + 0.59 + ], + "angle": 0, + "content": "5. Community, B.O.: Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam (2018), http://www.blender.org" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.591, + 0.787, + 0.618 + ], + "angle": 0, + "content": "6. Frankle, J., Carbin, M.: The lottery ticket hypothesis: Finding sparse, trainable neural networks. arXiv preprint arXiv:1803.03635 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.619, + 0.787, + 0.646 + ], + "angle": 0, + "content": "7. Goel, S., Kanazawa, A., Malik, J.: Shape and viewpoints without keypoints. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.646, + 0.787, + 0.673 + ], + "angle": 0, + "content": "8. Ha, D., Dai, A.M., Le, Q.V.: Hypernetworks. In: International Conference on Learning Representations (2017), https://openreview.net/forum?id=rkpAce1lx" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.787, + 0.715 + ], + "angle": 0, + "content": "9. Heppert, N., Irshad, M.Z., Zakharov, S., Liu, K., Ambrus, R.A., Bohg, J., Valada, A., Kollar, T.: Carto: Category and joint agnostic reconstruction of articulated objects. In: Proc. IEEE Conf. Comput. Vis. Pattern Recog. pp. 21201-21210 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.715, + 0.787, + 0.757 + ], + "angle": 0, + "content": "10. Jiang, Z., Hsu, C.C., Zhu, Y.: Ditto: Building digital twins of articulated objects from interaction. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.758, + 0.787, + 0.784 + ], + "angle": 0, + "content": "11. Kanazawa, A., Tulsiani, S., Efros, A.A., Malik, J.: Learning category-specific mesh reconstruction from image collections. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "12. Kokkinos, F., Kokkinos, I.: To the point: Correspondence-driven monocular 3d category reconstruction. In: Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems (2021), https://openreview.net/forum?id=AWMU04iXQ08" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.354, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.127 + ], + "angle": 0, + "content": "A. Swaminathan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "13. Kulkarni, N., Gupta, A., Fouhey, D.F., Tulsiani, S.: Articulation-aware canonical surface mapping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 452-461 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.244 + ], + "angle": 0, + "content": "14. Lei, J., Daniilidis, K.: Cadex: Learning canonical deformation coordinate space for dynamic surface representation via neural homeomorphism. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6624-6634 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.244, + 0.785, + 0.298 + ], + "angle": 0, + "content": "15. Li, T., Slavcheva, M., Zollhoefer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., Lv, Z.: Neural 3d video synthesis from multi-view video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.298, + 0.785, + 0.338 + ], + "angle": 0, + "content": "16. Lin, H., Peng, S., Xu, Z., Yan, Y., Shuai, Q., Bao, H., Zhou, X.: Efficient neural radiance fields for interactive free-viewpoint video. In: SIGGRAPH Asia Conference Proceedings (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.339, + 0.785, + 0.379 + ], + "angle": 0, + "content": "17. Liu, J., Mahdavi-Amiri, A., Savva, M.: PARIS: Part-level reconstruction and motion analysis for articulated objects. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.379, + 0.785, + 0.42 + ], + "angle": 0, + "content": "18. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. ACM Trans. Graph. 38(4), 65:1-65:14 (Jul 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.42, + 0.785, + 0.461 + ], + "angle": 0, + "content": "19. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.514 + ], + "angle": 0, + "content": "20. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.568 + ], + "angle": 0, + "content": "21. Müller, T., Evans, Å., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022). https://doi.org/10.1145/3528223.3530127, https://doi.org/10.1145/3528223.3530127" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.568, + 0.785, + 0.597 + ], + "angle": 0, + "content": "22. Neverova, N., Novotny, D., Khalidov, V., Szafraniec, M., Labatut, P., Vedaldi, A.: Continuous surface embeddings (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.597, + 0.785, + 0.651 + ], + "angle": 0, + "content": "23. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S.M., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.651, + 0.785, + 0.692 + ], + "angle": 0, + "content": "24. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Occupancy flow: 4d reconstruction by learning particle dynamics. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.692, + 0.785, + 0.732 + ], + "angle": 0, + "content": "25. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.732, + 0.785, + 0.759 + ], + "angle": 0, + "content": "26. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "27. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "28. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.263, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "29. Qian, S., Jin, L., Rockwell, C., Chen, S., Fouhey, D.F.: Understanding 3d object articulation in internet videos. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.217 + ], + "angle": 0, + "content": "30. Ramanujan, V., Wortsman, M., Kembhavi, A., Farhadi, A., Rastegari, M.: What's hidden in a randomly weighted neural network? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11893-11902 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "31. Saffran, J.R., Aslin, R.N., Newport, E.L.: Statistical learning by 8-month-old infants. Science 274(5294), 1926-1928 (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.285 + ], + "angle": 0, + "content": "32. Schwarz, J.R., Tack, J., Teh, Y.W., Lee, J., Shin, J.: Modality-agnostic variational compression of implicit neural representations. arXiv preprint arXiv:2301.09479 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "33. Sen, B., Agarwal, A., Namboodiri, V.P., Jawahar, C.: Inr-v: A continuous representation space for video-based generative tasks. arXiv preprint arXiv:2210.16579 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.369 + ], + "angle": 0, + "content": "34. Sen, B., Singh, G., Agarwal, A., Agaram, R., Krishna, K.M., Sridhar, S.: Hyp-nerf: Learning improved nef priors using a hypernetwork. arXiv preprint arXiv:2306.06093 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.411 + ], + "angle": 0, + "content": "35. Sitzmann, V., Rezchikov, S., Freeman, B., Tenenbaum, J., Durand, F.: Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems 34, 19313-19325 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.452 + ], + "angle": 0, + "content": "36. Sitzmann, V., Zollhöfer, M., Wetzstein, G.: Scene representation networks: Continuous 3d-structure-aware neural scene representations. Advances in Neural Information Processing Systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.494 + ], + "angle": 0, + "content": "37. Skorokhodov, I., Ignatyev, S., Elhoseiny, M.: Adversarial generation of continuous images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10753-10764 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "38. Tan, J., Yang, G., Ramanan, D.: Distilling neural fields for real-time articulated shape reconstruction. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.577 + ], + "angle": 0, + "content": "39. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In: IEEE International Conference on Computer Vision (ICCV). IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.604 + ], + "angle": 0, + "content": "40. Tseng, W.C., Liao, H.J., Lin, Y.C., Sun, M.: Cla-nerf: Category-level articulated neural radiance field. In: ICRA (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.785, + 0.688 + ], + "angle": 0, + "content": "41. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017), https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.729 + ], + "angle": 0, + "content": "42. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.785, + 0.784 + ], + "angle": 0, + "content": "43. Wei, F., Chabra, R., Ma, L., Lassner, C., Zollhoefer, M., Rusinkiewicz, S., Sweeney, C., Newcombe, R., Slavcheva, M.: Self-supervised neural articulated shape and appearance models. In: Proceedings IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "44. Wu, T., Zhong, F., Tagliasacchi, A., Cole, F., Oztireli, C.: D\\~nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 32653-32666. Curran Associates," + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.127 + ], + "angle": 0, + "content": "A. Swaminathan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "Inc. (2022), https://proceedings.neurips.cc/paper_files/paper/2022/file/d2cc447db9e56c13b993c11b45956281-Paper-Conference.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.786, + 0.217 + ], + "angle": 0, + "content": "45. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time neural irradiance fields for free-viewpoint video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9421-9431 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.786, + 0.272 + ], + "angle": 0, + "content": "46. Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: SAPIEN: A simulated part-based interactive environment. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.786, + 0.313 + ], + "angle": 0, + "content": "47. Yang, G., Sun, D., Jampani, V., Vlasic, D., Cole, F., Chang, H., Ramanan, D., Freeman, W.T., Liu, C.: Lasr: Learning articulated shape reconstruction from a monocular video. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.355 + ], + "angle": 0, + "content": "48. Yang, G., Sun, D., Jampani, V., Vlasic, D., Cole, F., Liu, C., Ramanan, D.: Viser: Video-specific surface embeddings for articulated 3d shape reconstruction. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.786, + 0.384 + ], + "angle": 0, + "content": "49. Yang, G., Vo, M., Neverova, N., Ramanan, D., Vedaldi, A., Joo, H.: Banmo: Building animatable 3d neural models from many casual videos. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.385, + 0.786, + 0.41 + ], + "angle": 0, + "content": "50. Yang, J., Pavone, M., Wang, Y.: Freenerf: Improving few-shot neural rendering with free frequency regularization (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.786, + 0.451 + ], + "angle": 0, + "content": "51. Yao, C.H., Hung, W.C., Li, Y., Rubinstein, M., Yang, M.H., Jampani, V.: Lassie: Learning articulated shape from sparse image ensemble via 3d part discovery. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.493 + ], + "angle": 0, + "content": "52. Yao, C.H., Hung, W.C., Li, Y., Rubinstein, M., Yang, M.H., Jampani, V.: LASSIE: Learning Articulated Shape from Sparse Image Ensemble via 3d part discovery. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.786, + 0.549 + ], + "angle": 0, + "content": "53. Yuan, W., Lv, Z., Schmidt, T., Lovegrove, S.: Star: Self-supervised tracking and reconstruction of rigid objects in motion with neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13144-13152 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.786, + 0.549 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/LEIA_ Latent View-invariant Embeddings for Implicit 3D Articulation/0e9336bc-d9b2-471b-94f5-34342577a604_origin.pdf b/2024/LEIA_ Latent View-invariant Embeddings for Implicit 3D Articulation/0e9336bc-d9b2-471b-94f5-34342577a604_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d90e84d7465b2a8680b3a9d291974dc4214ed69c --- /dev/null +++ b/2024/LEIA_ Latent View-invariant Embeddings for Implicit 3D Articulation/0e9336bc-d9b2-471b-94f5-34342577a604_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600a868f4faddfdf21d2f4fc8ddc3a759241db5cc14b851ff7a0638202b9fd18 +size 7517773 diff --git a/2024/LEIA_ Latent View-invariant Embeddings for Implicit 3D Articulation/full.md b/2024/LEIA_ Latent View-invariant Embeddings for Implicit 3D Articulation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d8a36a6610ec5197969a6b254a5a111c00e88c32 --- /dev/null +++ b/2024/LEIA_ Latent View-invariant Embeddings for Implicit 3D Articulation/full.md @@ -0,0 +1,282 @@ +# LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation + +Archana Swaminathan1, Anubhav Gupta1, Kamal Gupta1, Shishira R. Maiya1, Vatsal Agarwal1, and Abhinav Shrivastava1 + +University of Maryland, College Park MD 20742, USA https://archana1998.github.io/leia/ + +Abstract. Neural Radiance Fields (NeRFs) have revolutionized the reconstruction of static scenes and objects in 3D, offering unprecedented quality. However, extending NeRFs to model dynamic objects or object articulations remains a challenging problem. Previous works have tackled this issue by focusing on part-level reconstruction and motion estimation for objects, but they often rely on heuristics regarding the number of moving parts or object categories, which can limit their practical use. In this work, we introduce LEIA, a novel approach for representing dynamic 3D objects. Our method involves observing the object at distinct time steps or "states" and conditioning a hypernetwork on the current state, using this to parameterize our NeRF. This approach allows us to learn a view-invariant latent representation for each state. We further demonstrate that by interpolating between these states, we can generate novel articulation configurations in 3D space that were previously unseen. Our experimental results highlight the effectiveness of our method in articulating objects in a manner that is independent of the viewing angle and joint configuration. Notably, our approach outperforms previous methods that rely on motion information for articulation registration. + +Keywords: Articulated Objects $\cdot$ Neural Radiance Fields $\cdot$ 3D Vision + +# 1 Introduction + +Our world is full of dynamic objects moving and interacting in space and time. Humans develop this (rather impressive) understanding of how these everyday objects move and interact in three-dimensional space at a very early stage of the brain development [31]. The task involves understanding not only the static geometry but also the dynamic movements and spatial relationships between parts of an object, often referred to as articulations of the object. Understanding and representing object articulations from images and/or videos is also pivotal in enabling machines to perceive and navigate the physical world with finesse. In this work, we propose a novel method to model the object articulations by learning view-invariant latent embeddings of the 3D object from multiview images. Existing works in modeling object articulation in the three dimensional space typically use priors in the form of pre-trained large scale models [51], videos [27], + +![](images/f83dcbc494267fd326451141bc3f048acbce5aef71827d86cbb77962b0e072dd.jpg) +Fig. 1: Our method LEIA, takes in multi-view images of an object in four articulation states and is able to learn a view-invariant latent embedding for the state. We show that we can interpolate between the latents to generate any number of intermediate unseen states for the object using LEIA, given the camera position. + +or assumptions regarding rigidity/shape of the object [17]. These approaches often fail to generalize for the long-tail of objects, especially in the case when a video of the object with articulation is unavailable, or in presence of large and multiple articulations in the object. In our work, we postulate that multiview images of an object in different states can provide enough signal to model its 3D shape and articulation, even without priors from large-scale foundation models or information from videos. We achieve this by learning a generalizable, view-invariant latent embedding of the object for different states. We train these embeddings jointly with a hypernetwork that predicts weights of a NeRF [19] parameterizing the object state. The hypernetwork can be trained using multiview images of discretized states of the object, each state representing a different object articulation. In Figure 1, we show three objects from our dataset in two different articulation states. The latent embeddings representing each state or articulation can be interpolated during test time to generate the weights of a NeRF that is able to reconstruct a never seen before state of the object. + +Our key contributions can be summarized as following: + +- We introduce an end-to-end method LEIA for generating novel states for articulated objects solely with multiview images captured at multiple states. +- We demonstrate that interpolating between embeddings can generate states of articulations of object not seen during training. The embedding space becomes interpolable with a manifold loss that encourages the latents to follow a structure that establishes a linear relationship between them, by minimizing the distance between the nearest neighbours in the latent space. +- Remarkably, LEIA achieves this without the need for any ground-truth 3D supervision, motion information, or articulation codes, establishing its versatility and effectiveness in capturing complex articulations. +- Our analyses demonstrate LEIA's robustness to single and multiple articulations, as well as combinations of motions. We can disentangle articulations + +in different object parts if multiple are movable, making it scalable without constraints on the number of parts or motion types, unlike prior work. + +# 2 Related Work + +Neural Radiance Fields. Neural Radiance Fields [1, 19, 21, 42] have proven to be a massive success in modeling 3D scenes, due to the high fidelity of 3D reconstruction and novel view synthesis for static scenes. For dynamic scenes, Neural Volumes [18] utilizes an encoder-decoder voxel-based representation, complemented by an implicit voxel warp field. Occupancy Flow [24] tackled non-rigid geometry by assigning a motion vector to each point in both space and time, but requires full 3D ground truth supervision. Some of the first dynamic NeRF approaches [26, 28, 39] optimize an underlying volumetric deformable function and [15] conditions the NeRF on time. [16, 26, 45] followed this work, by learning a 5D spatiotemporal neural field. However, the approaches above usually require a video as input and do not handle well the case of large articulations in everyday objects given multiple states of objects as input. + +3D Representations for Articulation. Due to the fine-grained nature of the task, deep neural network models for representing articulation require conditioning or refining parts of the architecture to suit the task. Category-specific reconstruction of deformable objects from images [7, 11-13], sparked interest in identifying and recovering the deformation in the 3D space [13, 22, 52]. Several works focus on shape reconstruction from videos [38, 47-49] that estimate a shape template for humans and animals. Other works that learn articulation from videos include Qian et al. [29] that detects and segments articulation planes from in-the-wild videos and [44, 53] which decouple the static and dynamic parts of videos. These works focus on modeling humans and animals where a large amount of data is available. Preliminary work in using images and shape started with A-SDF [28], which generates unseen articulations using Signed Distance Function [25], by providing an input shape and articulation code. CaDeX [14] is a unified representation for shape and motion, obtained from a point cloud input. Ditto [10] is a similar work that uses implicit representations for joint geometry and articulation modeling, with ground truth point clouds. CLA-NeRF [40] learns unseen articulated states from observing multi-view image input along with articulation information. Wei et. al [43] obtains an SDF-based articulable representation of common objects by feeding in images of articulated states across multiple categories, and CARTO [9] uses stereo images as input and uses a stereo encoder to infer the 3D shape, 6D pose etc. of multiple unknown objects. [4] focuses on manipulating object shapes to deform according to specified articulation commands. Moving away from 3D supervision and articulation annotations, Jiayi et al. [17] proposed PARIS, a method that is able to obtain unseen articulated states of an object, given multi-view images of just the start and end state of articulation. PARIS employs a composite rendering based approach, by decoupling the object into a static and moving part and then separately estimating and compositing the static and mobile neural radiance fields. In our work, + +Table 1: Comparison of LEIA with existing methods. We show that LEIA is the first approach that does not use or learn any explicit prior along with not having any articulation input. This gives us flexibility in scaling to modeling articulations of objects with more than one part, and can thus handle a wide range of motion. We also have one universal model that can learn to represent both prismatic and revolute motion, unlike PARIS that has two separate models. 3D Sup. refers to "3D Supervision". + +
MethodImage/Shape InputDatasetArticulation Input3D Sup.?Prior# States# Parts
A-SDF [28]Shape CodeShape2MotionArticulation Code✓Articulation≥4≥1
CLA-NeRF [40]Multiview ImagesPartNet-MobilityArticulated PoseXArticulation≥41
NASAM [43]Multiview Images++PartNet-MobilityN/AXCategory≥4≥1
CARTO [9]Stereo ImagesCustomJoint Code✓Articulation≥41
Ditto [10]Point CloudsShape2MotionAnnotations✓Articulation21
PARIS [17]Multiview ImagesPartNet-MobilityN/AXMotion (learned)21
LEIAMultiview ImagesPartNet-MobilityN/A✓No4≥1
+ +we start LEIA with a similar input setting, by preparing multi-view renderings of objects in different articulation states. Unlike PARIS, we do not decouple the object into static and moveable parts, we rather learn to predict the weights of a single NeRF for any unseen articulated state using a state-modulated HyperNetwork. We aren't limited by the decoupling and the learned motion prior in the case of PARIS, which enables us to scale and learn any amount and kind of motion an object can possibly have, including combinations. We show a summed comparison of LEIA with prior work in Table 1. + +Hypernetworks with Implicit Neural Representations. Hypernetworks, a specialized class of networks designed to predict parameters for another network, aim to achieve generalization across novel tasks [8]. In the realm of stylizing 3D scenes, [3] employed a hypernetwork for applying diverse styles, while scene reconstructions from limited data points were achieved by [35, 36]. Despite the promise shown in representation, these hypernetworks operate on input data points, necessitating test-time optimizations and rendering them unsuitable for compression tasks. Rather than using the provided data (image/video) as input for the hypernetwork, an alternative approach involves employing an auto-decoder framework. In this framework, a learnable latent, without the need for an encoder, represents a data point. This technique, applied by [33] to represent a dataset of videos, assigns each latent to a distinct video. While this method yields a representation for each set of frames, the lack of decoupling in spatial-temporal coordinates limits its scalability to real-world frames. On a related note, [34] extended a similar approach to 3D shapes and scenes, effectively acquiring a latent representation suitable for various downstream tasks. We use a similar framework for learning the latents corresponding to each articulated state in our method. + +# 3 Method + +**Background.** Our architecture is based on neural radiance fields, or NeRF [19], which parameterizes the radiance and volume density at a 3D location of a + +scene as observed from a camera placed in a particular position using a neural network. $F_{\mathrm{NeRF}}:(\mathbf{x},\mathbf{d}) \to (\mathbf{c},\sigma)$ , where $\mathbf{c} = (r,g,b)$ , and $\sigma$ represent the radiance and volume density, and $\mathbf{x}$ and $\mathbf{d}$ represent the 3D location and viewing direction respectively. To render a pixel, the radiance $C(\mathbf{r})$ of a camera ray $\mathbf{r}(s) = \mathbf{o} + s \cdot \mathbf{d}$ is integrated from near to far bounds $s_n$ and $s_f$ such that, $C(\mathbf{r}) = \int_{s_n}^{s_f} T(s)\sigma(\mathbf{r}(s))\mathbf{c}(\mathbf{r}(s),\mathbf{d})ds$ , where the function $T(s)$ denotes the accumulated transmittance from $s_n$ to $s$ . To optimize the parameters of the MLP, a loss function is used that measures the discrepancy between the ground-truth and rendered images. Traditionally, the L2 loss is used for this purpose. However, to make the training more robust to outliers and to improve convergence, the Smooth L1 Loss can be employed as an alternative. The Smooth L1 Loss is a combination of L1 and L2 losses, behaving like L1 loss for large errors and like L2 loss for small errors. The loss function for the NeRF model using the Smooth L1 Loss is defined as: + +$$ +\operatorname {S m o o t h L 1 L o s s} (x) = \left\{ \begin{array}{l l} 0. 5 \cdot x ^ {2} & \text {i f} | x | < 1 \\ | x | - 0. 5 & \text {o t h e r w i s e} \end{array} \right. \tag {1} +$$ + +$$ +L _ {\text {S m o o t h L 1 - N e R F}} = \sum_ {r \in R} \text {S m o o t h L 1 L o s s} \left(\hat {C} (\mathbf {r}) - C (\mathbf {r})\right) \tag {2} +$$ + +where $R$ is the set of rays used for sampling, and $\hat{C}(\mathbf{r})$ and $C(\mathbf{r})$ represent the ground-truth and rendered colors, respectively. By minimizing this loss function, the NeRF model learns to accurately reproduce the radiance of the scene, leading to high-quality image synthesis from novel viewpoints. While this formulation works well for generating novel scene viewpoints for a static scene, it cannot handle dynamic scenes. [45] tried to address this shortcoming by making the MLP learn a spatiotemporal radiance field and used time, $t$ , as an additional input. This works in principle but it becomes expensive to scale to lengthier videos due to constantly increasing sampling space for NeRF. We use the $L_{\mathrm{SmoothL1 - NeRF}}$ loss along with AdamW optimization to train LEIA, along with a $L_{\mathrm{mask}}$ that is the BCE loss between the predicted opacity and ground truth foreground loss. + +# 3.1 Approach + +Hypernetworks, or hypernets for short, are neural networks that generate weights for another neural network, known as the target network. Hypernets can be conditioned on various domains and can predict the weights for multiple neural networks simultaneously, if trained appropriately. + +In this work, we use a hypernet to modulate a Neural Radiance Field (NeRF) based on the state of articulation, thereby learning a parametrization for each state. We employ a learnable latent embedding, $Z$ , as input to this hypernet $h_l, l \in L$ where $L$ is the number of layers of the hypernet. This approach allows us to not only modulate the NeRF but also create useful representations for each state. Our system can thus be represented as follows: + +$$ +F _ {\theta_ {t}} (\mathbf {x}, \mathbf {d}) = \left(\mathbf {c} _ {t}, \sigma_ {t}\right) \tag {3} +$$ + +$$ +\theta_ {t} = h (z _ {t}), z _ {t} \in Z +$$ + +![](images/a2588fda4b5e7f167ce848216164e9c33e2ebada7f090c2a663771e2ec4e6663.jpg) +Fig. 2: Overview of our method. We take multi-view images in different states as input. A learnable latent dictionary based off an autoencoder learns an embedding per state id. The latent embedding is used as an input to the hypernet, that modulates and generates weights of the NeRF to reconstruct the state that is fed in. At inference time, we do a weighed interpolation of the learnt latents to obtain a corresponding newly generated intermediate state. + +The latents $Z$ , give us an additional ability to interpolate across seen states and generate novel states of the object, as we show in our work. The set of latents, $Z$ , in Eq. (3) can be understood as a learnable dictionary where keys are the state ids. Formally, we represent this as + +$$ +Z = \left\{t: z _ {t} \mid t \in [ 0, 1, 2 \dots T ] \right\} \tag {4} +$$ + +where, $T$ is the total number of discretized states sampled for the object. Each of these states, $z_{t}$ , is used as an input to the hypernet to get the parameterization for NeRF as shown in Figure 2. Once parameterized, the NeRF is trained using multi-view images of the object taken from various camera angles, providing a comprehensive view of the object's articulation in the current state. We sample only one state per batch. + +Directly predicting the weights $\theta$ of the base network $f_{\theta}$ , using the hypernet $h_l$ , is expensive, parameter-heavy, and unsuitable for compression. Hence, we follow [32, 37] and instead predict low-rank matrices, which are then applied to the base network weights. This type of modulation acts as a form of subnetwork selection, analogous to systems proposed in [6, 30]. For a base network $f_{\theta}$ with $L$ layers, our formulation now looks like + +$$ +f _ {\theta} \left(\left(\mathbf {x}, \mathbf {d}\right) \mid \theta_ {t} ^ {l _ {1}}, \theta_ {t} ^ {l _ {2}} \dots \theta_ {t} ^ {l _ {L}}\right) = \mathbf {c} _ {t}, \sigma_ {t} +$$ + +$$ +\theta_ {t} ^ {l} = \eta \left(P ^ {l} \times Q ^ {l}\right) \circ \theta^ {l} \tag {5} +$$ + +$$ +h _ {l} (z _ {t}) = [ P ^ {l}, Q ^ {l} ] +$$ + +where $\theta^l$ represents the weights of the $l$ -th layer and $\theta_t^l$ denotes the modulated weights for frame $t$ . Here, $\eta$ signifies an activation function on the matrix-product of low rank matrices $\mathbf{P}^l \in \mathbb{R}^{K \times r}$ , $\mathbf{Q}^l \in \mathbb{R}^{r \times K}$ , where $K$ is the width of the + +base network $f_{\theta}$ and rank $r \ll K$ . These matrices are responsible for adjusting the weights $\theta_{l}$ as dictated by the corresponding hypernetwork $h_l$ . Note that all hypernetworks use the same latent $z_{t} \in \mathbb{R}^{D}$ as input. The rank $r$ acts as a hyperparameter that controls the compression-performance trade-off. We further elaborate on details about model architecture and design choices for the hypernet, NeRF and the learnable latent dictionary in the supplementary. + +# 3.2 Interpolation + +Given an object with states $t_1$ and $t_2$ , and a scale $\alpha$ , the task of state interpolation involves creating $\alpha - 1$ coherent states between the given states. In order to achieve this, we do a linear interpolation on the state latents $z_1$ and $z_2$ and pass the resulting latent through the hypernetwork. This gives us the weight modulation required in the NeRF, and the updated base network is used to obtain different viewpoints of the intermediate state. + +$$ +z _ {\text {i n t e r}} = \left(1 - \beta_ {i}\right) \cdot z _ {t} + \beta_ {i} \cdot z _ {t - 1} \tag {6} +$$ + +$$ +\mathbf {c} _ {\text {i n t e r}}, \sigma_ {\text {i n t e r}} = f _ {\theta_ {\text {i n t e r}}} (\mathbf {x}, \mathbf {d}; h (z _ {\text {i n t e r}})) +$$ + +$$ +\text {w h e r e}, \beta_ {i} \in \left[ \frac {1}{\alpha}, \frac {2}{\alpha}, \dots , \frac {\alpha - 1}{\alpha} \right] \tag {7} +$$ + +essentially generating $\alpha - 1$ states between any two given states. In our experiments, we quantitatively evaluate the result of the interpolation with unseen ground truth obtained from our dataset of all states, by linearly averaging the two extreme states present in the training, to obtain a latent for a state id. + +# 3.3 Model Architecture + +In our experiments, both the base network $f_{\theta}$ and hypernetworks $h_l$ are simple MLPs that take in a coordinate input, and an id depicting the state of articulation. A network $l_n$ is used to learn the latent dictionary for the states, which is a linear layer learnt in conjunction with the hypernet, and uses the torch.nn.Embedding class as a lookup table for the learnable latent embeddings, depicted by $z_t$ for $t \in T$ . This latent is fed to the hypernetwork $h_l$ , that modulates the weights of the base network $f_{\theta}$ for reconstructing the output of the corresponding state $\mathrm{id}$ $t$ . The base network $f_{\theta}$ is based off the NeRF architecture of Instant-NGP [21]. This architecture has separate fully-connected blocks for the geometry and texture of the scene, and we use separate hypernets to modulate each layer of these two FC blocks. + +# 3.4 Optimization: Loss functions and regularizers + +Latent Manifold Loss. We use the latent manifold loss function that enforces a structured latent space by encouraging local consistency among learnt latents. For each latent vector, the loss computes the average Euclidean distance between + +the vector and its $K$ nearest-neighbours on the manifold. This process enforces a smooth and continuous latent manifold, which is beneficial for models that rely on meaningful linear interpolations between the points on the manifold, and for tasks where the geometry of the latent space is crucial. Mathematically, we represent this loss for a particular state-id $i$ as: + +$$ +\mathcal {L} _ {\text {m a n i f o l d}} \left(\mathbf {l} _ {i}\right) = \frac {1}{K} \sum_ {k = 1} ^ {K} \| \mathbf {l} _ {i} - \mathbf {n} _ {k} \| _ {2} ^ {2} \tag {8} +$$ + +where $\mathbf{n}_k$ are the $K$ nearest neighbors of $\mathbf{l}_i$ in the latent space, and $\| \cdot \|_2$ denotes the Euclidean (L2) norm. The loss is averaged over the selected latents and their nearest neighbors to ensure local uniformity in the manifold's geometry, where $K$ is a hyperparameter that determines the number of neighbors considered. + +Depth and Occlusion Regularization. Our depth and occlusion regularizations are designed to refine the clarity of rendered images by addressing occlusion and depth smoothness [23, 50]. The occlusion regularization loss, $L_{\mathrm{occ}}$ , aims to mitigate the obscuring of objects located beyond a specified depth threshold during rendering. This is accomplished by generating a binary mask, $m_k$ , where the mask elements are set to 1 up to a certain index $M$ reflecting the regularization range, and 0 thereafter. The occlusion loss is then articulated as the normalized sum of the product of the mask and the sampled density values $\sigma_k$ along a ray, as given by + +$$ +L _ {\text {o c c}} = \frac {1}{K} \sum_ {k = 1} ^ {K} \sigma_ {k} \cdot m _ {k}, \tag {9} +$$ + +where $K$ is the total number of sampled points on the ray. This formulation drives the model to prefer representations that reduce occlusions close to the camera, ensuring objects further away are not improperly concealed. + +For depth continuity, the depth smoothness regularization loss, $L_{\mathrm{DS}}(\theta ,\mathbf{R})$ enforces the gradual transition of depth values among neighboring pixels, reducing sharp depth disparities that cause visual inconsistencies. If a ray $r$ intersects with a single depth value, this value is evaluated directly. In contrast, for multiple depth values, the loss is the aggregate of squared differences between adjacent depth estimates. Formally, for depth predictions $\hat{d}_0(r)$ , the loss is quantified as + +$$ +L _ {\mathrm {D S}} (\theta , \mathbf {R}) = \sum_ {r \in \mathbf {R}} \sum_ {i, j = 1} ^ {\text {P a t c h} - 1} \left(\hat {d} _ {0} \left(r _ {i j}\right) - \hat {d} _ {0} \left(r _ {(i + 1) j}\right)\right) ^ {2} + \left(\hat {d} _ {0} \left(r _ {i j}\right) - \hat {d} _ {0} \left(r _ {i (j + 1)}\right)\right) ^ {2}. \tag {10} +$$ + +This measure is averaged across the rays to yield a measure of the depth map's smoothness. By integrating these regularizers into our training regimen, we significantly dampen the disturbances introduced by overlapping and obscured pixels, resulting in a more consistent interpolation between states. These benefits are substantiated through extensive ablation studies and quantitative analyses presented in the experiments. + +Positional Encoding of the Latent. We incorporate positional encoding to capture the order of input elements, crucial for understanding articulation. Positional encoding injects the sequence with its inherent order, a key factor in articulation semantics. We employ the scheme from [41], which uses sine and cosine functions parameterized to encode varying frequencies. The computed positional encodings are added to the latent vector $z_{t}$ , enriching it with semantic information that more accurately encodes the state of articulation. We analyze the effects of positional encoding on our model in the experimental section. + +# 4 Experiments + +# 4.1 Setup + +Datasets. In this work, we use the PartNet-Mobility dataset [2, 20, 46], a large-scale synthetic collection of articulated objects in over 40 categories. The dataset has a variety of articulations defined, with objects comprising of single and multi-joint parts. Two types of motion, revolute (rotational) joint and prismatic (translation) articulation are represented in the dataset. To make up our training dataset, we choose 100 camera views that are arranged in a dome-like setup, capturing the upper hemisphere of the object. This is similar to the setup used by PARIS. We use the SAPIEN [46] library that the PartNet-Mobility dataset was released with, to render our RGB images at linearly spaced intervals of articulation to make up frames of a video that depicts the range of motion and obtain the camera parameters accordingly, converting it into the Blender [5] coordinate system to fit in our codebase. We use instances from 8 different common household item categories storage, microwave, laptop, oven, washer, dishwasher, sunglasses, box and choose 1-4 objects per category, bringing our total number of objects to 12. We also show the efficacy of our method on 68 images from a real-world scene of a chest of drawers, which are captured with a mobile phone and post-processed to remove background. Throughout LEIA, we train with a total of four states and interpolate between the two extreme states. + +Baselines. The closest prior work with a setup similar to ours is PARIS, which takes in multi-view images of objects in two states and disentangles the static and moving part of the object. While we run experiments with four input states, we don't focus on learning a defined articulation for the object, instead we are able to recognize any arbitrary motion and combinations of motions given relevant states. We emphasize on the following differences: + +1. PARIS uses separate models to train objects that have rotation and translation motion, respectively. While they have a method to estimate the type of motion, once determined, it is necessary to train on the appropriate model. +2. As PARIS also learns motion parameters for the articulation, given their pipeline of disentanglement of the static and moving parts of the object, this restricts the learning of motion parameters if there are multiple parts in the object, moving differently. + +Table 2: Quantitative Results for Interpolated State Reconstruction. We compared our method with the PARIS baseline, trained on selected objects from the SAPIEN dataset. The results from three experiments are summarized below. The VanillaInt experiment involves simple interpolation of the latents. Our best-performing method, LEIA, introduces structure to the latents with a manifold loss and regularizers. Although PARIS learns a motion prior and LEIA implicitly performs state interpolation, both methods perform similarly for single-part objects. However, our approach excels with multi-part objects, outperforming PARIS significantly due to its flexibility in handling various motion types and articulations without constraints. + +
MetricsMethodsSingle-Part ArticulationMulti-Part ArticulationAverage
45135 Storage17128 Microwave10211 Laptop101917 Oven103778 Washer12085 Dishwasher44781 Storage245427 Storage345575 Storage4101297 Sunglasses7187 Oven102377 Box
PSNR↑PARIS28.6625.9424.9728.1334.4627.2426.3524.4125.7332.3329.3026.2027.81
VanillaInt24.2022.0022.6324.5336.0324.1630.9629.7829.1032.1130.3527.8527.81
LEIA26.6926.3825.0428.7136.1426.4931.0729.7829.8035.6030.8028.0529.55
SSIM↑PARIS0.990.970.970.970.980.960.950.950.940.960.960.930.96
VanillaInt0.950.910.900.930.990.920.950.950.930.970.950.930.94
LEIA0.970.950.950.980.990.950.960.950.950.970.960.950.96
LPIPS↓PARIS0.020.060.190.030.030.060.050.050.040.060.050.080.06
VanillaInt0.050.100.170.030.020.080.030.030.050.090.050.130.07
LEIA0.030.070.100.020.020.070.030.030.040.090.050.120.06
CD↓PARIS0.180.080.120.750.060.450.620.420.600.200.990.910.45
VanillaInt0.240.070.060.100.060.200.480.470.460.970.500.830.37
LEIA0.290.060.360.100.050.270.520.380.410.960.350.620.36
+ +3. LEIA is motion-prior free, and we have a universal architecture that implicitly learns intermediate states, so we are able to train with multiple types of motion occurring in the same object, with no change in the code. + +For comparing with PARIS, we run authors' original source code with our dataset, using the hyperparameters provided by the authors and training until convergence. We test our work and PARIS with the appearance quality of the reconstructed image of the interpolated state. We also set up a simple baseline that does just vanilla interpolation between the learnt latent embeddings, without enforcing any structure or constraints on them. We call this baseline VanillaInt and compare it with LEIA, which has been finetuned to do interpolation with the addition of the manifold loss and the depth and occlusion regularization for denoising the resulting output. + +Quantitative Metrics. We use three appearance quality metrics, PSNR (Peak Signal-to-Noise Ratio), SSIM (Structural Similarity Index), and LPIPS (Learned Perceptual Image Patch Similarity). These are computed between reconstructed images of the interpolated state, and the unseen ground truth of the state from our dataset. We also report the Chamfer Distance to measure quality of 3D reconstruction by doing point sampling. + +# 4.2 Main Results + +Novel State Synthesis. We show the qualitative and quantitative results of interpolating our learned latent embeddings corresponding to the start and end states, in Table 2 and Figure 3. Our results show that the latent embeddings can + +![](images/4ad50401da0834031c7f891aec74172bd3b21281ab7b9416271fb3d760158546.jpg) +Fig. 3: Qualitative Results. We show results of PARIS and LEIA for reconstructing the unseen intermediate state, for both single and multiple articulations. We see that PARIS especially fails when there are two parts of the object moving differently, as the motion parameters are not registered correctly. LEIA handles this case successfully as it is not dependent on part disentanglement to identify and register articulation. LEIA also performs comparable to PARIS for single-part articulation, despite not having a dedicated model for the motion or part disentanglement. + +be linearly interpolated, and have a structure in the latent space. We can generate any number of states between the start and the end state by doing a weighted + +![](images/715c8b2973ffca7878e9bc50c48b5382415e843156db77caaf2097ba0500265c.jpg) +Fig. 4: Real World Results. LEIA is able to faithfully interpolate and reconstruct between two states of images from our real world data, proving its ability to generalize and work in an in-the-wild setting. + +combination of the latents representing the start and the end states. For training, we choose four states, using the two additional states compared to PARIS to enforce structure in the latent space with the latent manifold loss, thereby making them amenable to interpolation. This helps us extend our formulation to multiple articulations without restriction. We also show results of our baseline VanillaInt and show that the numbers are boosted with the addition of the latent manifold loss and regularizers. For single-part objects, we perform comparably to PARIS despite having no part disentanglement and no dedicated model to represent the exact type of motion. For multiple-part articulations, we beat PARIS across all objects, as shown in Figure 3. Our comparable performance in 3D reconstruction shows that LEIA's learnt view-invariant latent embeddings are powerful in preserving spatio-temporal consistency. We also show that our method can work in the real world setting, in Figure 4. Ablations are shown for number of states, latent manifold loss, and depth and occlusion regularizations in Table 3 and Figure 6. + +Analysis on Disentanglement of Joints. To check separability of the latent space, we train LEIA with multi-view images across multiple states, captured from a video that showed three parts of an object Storage1, move one after the other. We do a t-SNE based dimensionality reduction on the embeddings and plot them, shown in Figure 5. The latents exhibit clear separability, with the motion of each joint represented in a smooth trajectory, moving outward as the respective drawer in the object moves away from the starting "closed" state. This shows that the separated latent space is an indicator of multiple joints, positioning us to be capable of learning representations of multiple articulations. + +# 4.3 Ablation Analysis + +We perform ablations to tune various design choices, with results shown in Table 3 and Figure 6. The first experiment we performed was testing out the impact of the latent manifold loss. We notice qualitatively that without the manifold loss, our latent space lacked structure and the linear relationship between the embeddings was not captured. As shown in Figure 6, the intermediate state overfits to the extreme state without the latent manifold while the addition of the loss enables it to accurately capture the intermediate state. This figure also + +![](images/bbd16fba78450d42a9ec7c3bf80b66b2c60b3fccff5869e70baaf030834d920a.jpg) +Fig. 5: t-SNE plot. After dimensionality reduction on jointly-learned state embeddings of an object with different moving parts. Our learned representations are separated and follow a smooth trajectory for each of the moving parts. + +Table 3: Ablations. Our ablations reveal that latent manifold loss, depth, and occlusion regularization enhance LEIA's visual metrics. Opting for four states improved latent structure with the manifold loss. + +
Latent Manifold LossPSNRSSIMLPIPS
with29.400.950.05
without28.540.940.06
Depth RegularizationPSNRSSIMLPIPS
with29.630.960.05
without26.930.930.07
Occlusion RegularizationPSNRSSIMLPIPS
with29.640.950.05
without28.640.950.06
Positional encodingPSNRSSIMLPIPS
with27.110.940.06
without28.480.950.06
Number of StatesPSNRSSIMLPIPS
228.040.950.06
429.690.960.05
+ +![](images/a68404074e403e14993805059219a8ad50f682219ca07963196a48066b0a9038.jpg) +Fig. 6: Qualitative Results of Ablations. We show some qualitative results of using latent manifold loss, positional encoding and varying the number of states. Using the manifold loss prevents us from overfitting to the extreme states, and using four states help us interpolate with a lot of clarity. Positional encoding helps add missing information for thin parts, but for large objects it doesn't help much. + +shows the effect of positional encoding which adds detail when the latent embedding fails to capture it, as evidenced in the results of sunglasses. This can + +![](images/70fd1f6b3209cee7e9da6a876b8f5bef52c1c2dbe1874431545a28352300f853.jpg) +Fig. 7: Failure cases. Our model fails at reconstructing the geometry correctly at camera positions where the two states we interpolate between have a change in the visible shape, like these examples where the microwave and oven being closed shows a deformation in the figure as compared to the open state, as the relationship between the motion is not easily captured in the latent due to the structure change. + +happen for thin parts of the object that may not be well-captured across camera views. Notably, the inverse happens in the case of large objects, where the positional encoding adds extra noise to the reconstruction, resulting in oversmoothing. We also show how adding just two more intermediate states in LEIA makes a huge difference in reconstruction results for both single and multi-part objects, enabling us to scale and be flexible. Quantitative numbers for depth and occlusion regularization ablation are shown in Table 3. However, they didn't universally help in the cases where the latent embedding already learned the state representation well enough but reduced noise when it appeared. + +# 5 Limitations + +While LEIA works in achieving good quality reconstruction of intermediate states, our latent embeddings do not yet ensure accurate physical consistency in the motion of the intermediates, which is a tradeoff we chose while opting for not decoupling the object and learning the motion parameters for the moveable part. This allowed us to scale our approach to interpolate between multiple moving parts of the object, which would be more representative of the real world where common objects can be articulated in multiple ways. LEIA also struggles when there is severe occlusion, as referenced by Figure 7, as the occlusion occurs at gaps in the learnt embeddings between two states. + +# 6 Conclusion + +In this work, we present LEIA, a method capable of successfully interpolating between two discrete articulation states of a deformable object with moving parts. LEIA handles multiple joints, including cases where moving parts are separated by static regions. It outperforms existing methods, especially when multiple parts move independently. The learned latent embeddings are view-invariant and separable, demonstrating LEIA's scalability and flexibility. We conducted comprehensive evaluations on synthetic and real data to investigate inherent challenges. Despite advancements, significant occlusion scenarios remain challenging. We hope future research builds upon our work. + +# Acknowledgements + +This work was partially supported by NSF CAREER Award (#2238769) to Abhinav Shrivastava, and IARPA via Department of Interior/Interior Business Center (DOI/IBC) contract number 140D0423C0076. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of NSF, IARPA, DOI/IBC, or the U.S. Government. + +# References + +1. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5470-5479 (June 2022) +2. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015) +3. Chiang, P.Z., Tsai, M.S., Tseng, H.Y., Lai, W.S., Chiu, W.C.: Stylizing 3d scene via implicit representation and hypernetwork. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1475-1484 (2022) +4. Chu, R., Liu, Z., Ye, X., Tan, X., Qi, X., Fu, C.W., Jia, J.: Command-driven articulated object understanding and manipulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8813-8823 (June 2023) +5. Community, B.O.: Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam (2018), http://www.blender.org +6. Frankle, J., Carbin, M.: The lottery ticket hypothesis: Finding sparse, trainable neural networks. arXiv preprint arXiv:1803.03635 (2018) +7. Goel, S., Kanazawa, A., Malik, J.: Shape and viewpoints without keypoints. In: ECCV (2020) +8. Ha, D., Dai, A.M., Le, Q.V.: Hypernetworks. In: International Conference on Learning Representations (2017), https://openreview.net/forum?id=rkpAce1lx +9. Heppert, N., Irshad, M.Z., Zakharov, S., Liu, K., Ambrus, R.A., Bohg, J., Valada, A., Kollar, T.: Carto: Category and joint agnostic reconstruction of articulated objects. In: Proc. IEEE Conf. Comput. Vis. Pattern Recog. pp. 21201-21210 (2023) +10. Jiang, Z., Hsu, C.C., Zhu, Y.: Ditto: Building digital twins of articulated objects from interaction. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +11. Kanazawa, A., Tulsiani, S., Efros, A.A., Malik, J.: Learning category-specific mesh reconstruction from image collections. In: ECCV (2018) +12. Kokkinos, F., Kokkinos, I.: To the point: Correspondence-driven monocular 3d category reconstruction. In: Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems (2021), https://openreview.net/forum?id=AWMU04iXQ08 + +13. Kulkarni, N., Gupta, A., Fouhey, D.F., Tulsiani, S.: Articulation-aware canonical surface mapping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 452-461 (2020) +14. Lei, J., Daniilidis, K.: Cadex: Learning canonical deformation coordinate space for dynamic surface representation via neural homeomorphism. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6624-6634 (June 2022) +15. Li, T., Slavcheva, M., Zollhoefer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., Lv, Z.: Neural 3d video synthesis from multi-view video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022) +16. Lin, H., Peng, S., Xu, Z., Yan, Y., Shuai, Q., Bao, H., Zhou, X.: Efficient neural radiance fields for interactive free-viewpoint video. In: SIGGRAPH Asia Conference Proceedings (2022) +17. Liu, J., Mahdavi-Amiri, A., Savva, M.: PARIS: Part-level reconstruction and motion analysis for articulated objects. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2023) +18. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. ACM Trans. Graph. 38(4), 65:1-65:14 (Jul 2019) +19. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +20. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019) +21. Müller, T., Evans, Å., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022). https://doi.org/10.1145/3528223.3530127, https://doi.org/10.1145/3528223.3530127 +22. Neverova, N., Novotny, D., Khalidov, V., Szafraniec, M., Labatut, P., Vedaldi, A.: Continuous surface embeddings (2020) +23. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S.M., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2022) +24. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Occupancy flow: 4d reconstruction by learning particle dynamics. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019) +25. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019) +26. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. ICCV (2021) +27. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021) +28. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020) + +29. Qian, S., Jin, L., Rockwell, C., Chen, S., Fouhey, D.F.: Understanding 3d object articulation in internet videos. In: CVPR (2022) +30. Ramanujan, V., Wortsman, M., Kembhavi, A., Farhadi, A., Rastegari, M.: What's hidden in a randomly weighted neural network? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11893-11902 (2020) +31. Saffran, J.R., Aslin, R.N., Newport, E.L.: Statistical learning by 8-month-old infants. Science 274(5294), 1926-1928 (1996) +32. Schwarz, J.R., Tack, J., Teh, Y.W., Lee, J., Shin, J.: Modality-agnostic variational compression of implicit neural representations. arXiv preprint arXiv:2301.09479 (2023) +33. Sen, B., Agarwal, A., Namboodiri, V.P., Jawahar, C.: Inr-v: A continuous representation space for video-based generative tasks. arXiv preprint arXiv:2210.16579 (2022) +34. Sen, B., Singh, G., Agarwal, A., Agaram, R., Krishna, K.M., Sridhar, S.: Hyp-nerf: Learning improved nef priors using a hypernetwork. arXiv preprint arXiv:2306.06093 (2023) +35. Sitzmann, V., Rezchikov, S., Freeman, B., Tenenbaum, J., Durand, F.: Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems 34, 19313-19325 (2021) +36. Sitzmann, V., Zollhöfer, M., Wetzstein, G.: Scene representation networks: Continuous 3d-structure-aware neural scene representations. Advances in Neural Information Processing Systems 32 (2019) +37. Skorokhodov, I., Ignatyev, S., Elhoseiny, M.: Adversarial generation of continuous images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10753-10764 (2021) +38. Tan, J., Yang, G., Ramanan, D.: Distilling neural fields for real-time articulated shape reconstruction. In: CVPR (2023) +39. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In: IEEE International Conference on Computer Vision (ICCV). IEEE (2021) +40. Tseng, W.C., Liao, H.J., Lin, Y.C., Sun, M.: Cla-nerf: Category-level articulated neural radiance field. In: ICRA (2022) +41. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017), https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf +42. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS (2021) +43. Wei, F., Chabra, R., Ma, L., Lassner, C., Zollhoefer, M., Rusinkiewicz, S., Sweeney, C., Newcombe, R., Slavcheva, M.: Self-supervised neural articulated shape and appearance models. In: Proceedings IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +44. Wu, T., Zhong, F., Tagliasacchi, A., Cole, F., Oztireli, C.: D\~nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 32653-32666. Curran Associates, + +Inc. (2022), https://proceedings.neurips.cc/paper_files/paper/2022/file/d2cc447db9e56c13b993c11b45956281-Paper-Conference.pdf +45. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time neural irradiance fields for free-viewpoint video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9421-9431 (2021) +46. Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: SAPIEN: A simulated part-based interactive environment. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) +47. Yang, G., Sun, D., Jampani, V., Vlasic, D., Cole, F., Chang, H., Ramanan, D., Freeman, W.T., Liu, C.: Lasr: Learning articulated shape reconstruction from a monocular video. In: CVPR (2021) +48. Yang, G., Sun, D., Jampani, V., Vlasic, D., Cole, F., Liu, C., Ramanan, D.: Viser: Video-specific surface embeddings for articulated 3d shape reconstruction. In: NeurIPS (2021) +49. Yang, G., Vo, M., Neverova, N., Ramanan, D., Vedaldi, A., Joo, H.: Banmo: Building animatable 3d neural models from many casual videos. In: CVPR (2022) +50. Yang, J., Pavone, M., Wang, Y.: Freenerf: Improving few-shot neural rendering with free frequency regularization (2023) +51. Yao, C.H., Hung, W.C., Li, Y., Rubinstein, M., Yang, M.H., Jampani, V.: Lassie: Learning articulated shape from sparse image ensemble via 3d part discovery. In: NeurIPS (2022) +52. Yao, C.H., Hung, W.C., Li, Y., Rubinstein, M., Yang, M.H., Jampani, V.: LASSIE: Learning Articulated Shape from Sparse Image Ensemble via 3d part discovery. In: Advances in Neural Information Processing Systems (NeurIPS) (2022) +53. Yuan, W., Lv, Z., Schmidt, T., Lovegrove, S.: Star: Self-supervised tracking and reconstruction of rigid objects in motion with neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13144-13152 (2021) \ No newline at end of file diff --git a/2024/LEIA_ Latent View-invariant Embeddings for Implicit 3D Articulation/images.zip b/2024/LEIA_ Latent View-invariant Embeddings for Implicit 3D Articulation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a997b72431a4eaa87013dd49b8a39d88e7ad3c25 --- /dev/null +++ b/2024/LEIA_ Latent View-invariant Embeddings for Implicit 3D Articulation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ad8570a0032b10207eb553cfa16a14e04c2c8942de7a81a6e66ce6b0b539868 +size 504110 diff --git a/2024/LEIA_ Latent View-invariant Embeddings for Implicit 3D Articulation/layout.json b/2024/LEIA_ Latent View-invariant Embeddings for Implicit 3D Articulation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..94628a656f133034207663223b207c8bdef6ef3d --- /dev/null +++ b/2024/LEIA_ Latent View-invariant Embeddings for Implicit 3D Articulation/layout.json @@ -0,0 +1,8350 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 463, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 463, + 147 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 463, + 147 + ], + "type": "text", + "content": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 143, + 168, + 471, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 168, + 471, + 193 + ], + "spans": [ + { + "bbox": [ + 143, + 168, + 471, + 193 + ], + "type": "text", + "content": "Archana Swaminathan1, Anubhav Gupta1, Kamal Gupta1, Shishira R. Maiya1, Vatsal Agarwal1, and Abhinav Shrivastava1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 196, + 202, + 417, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 202, + 417, + 224 + ], + "spans": [ + { + "bbox": [ + 196, + 202, + 417, + 224 + ], + "type": "text", + "content": "University of Maryland, College Park MD 20742, USA https://archana1998.github.io/leia/" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 255, + 452, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 255, + 452, + 452 + ], + "spans": [ + { + "bbox": [ + 160, + 255, + 452, + 452 + ], + "type": "text", + "content": "Abstract. Neural Radiance Fields (NeRFs) have revolutionized the reconstruction of static scenes and objects in 3D, offering unprecedented quality. However, extending NeRFs to model dynamic objects or object articulations remains a challenging problem. Previous works have tackled this issue by focusing on part-level reconstruction and motion estimation for objects, but they often rely on heuristics regarding the number of moving parts or object categories, which can limit their practical use. In this work, we introduce LEIA, a novel approach for representing dynamic 3D objects. Our method involves observing the object at distinct time steps or \"states\" and conditioning a hypernetwork on the current state, using this to parameterize our NeRF. This approach allows us to learn a view-invariant latent representation for each state. We further demonstrate that by interpolating between these states, we can generate novel articulation configurations in 3D space that were previously unseen. Our experimental results highlight the effectiveness of our method in articulating objects in a manner that is independent of the viewing angle and joint configuration. Notably, our approach outperforms previous methods that rely on motion information for articulation registration." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 161, + 464, + 448, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 464, + 448, + 474 + ], + "spans": [ + { + "bbox": [ + 161, + 464, + 448, + 474 + ], + "type": "text", + "content": "Keywords: Articulated Objects " + }, + { + "bbox": [ + 161, + 464, + 448, + 474 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 161, + 464, + 448, + 474 + ], + "type": "text", + "content": " Neural Radiance Fields " + }, + { + "bbox": [ + 161, + 464, + 448, + 474 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 161, + 464, + 448, + 474 + ], + "type": "text", + "content": " 3D Vision" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 496, + 229, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 496, + 229, + 508 + ], + "spans": [ + { + "bbox": [ + 133, + 496, + 229, + 508 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 521, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 521, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 521, + 481, + 666 + ], + "type": "text", + "content": "Our world is full of dynamic objects moving and interacting in space and time. Humans develop this (rather impressive) understanding of how these everyday objects move and interact in three-dimensional space at a very early stage of the brain development [31]. The task involves understanding not only the static geometry but also the dynamic movements and spatial relationships between parts of an object, often referred to as articulations of the object. Understanding and representing object articulations from images and/or videos is also pivotal in enabling machines to perceive and navigate the physical world with finesse. In this work, we propose a novel method to model the object articulations by learning view-invariant latent embeddings of the 3D object from multiview images. Existing works in modeling object articulation in the three dimensional space typically use priors in the form of pre-trained large scale models [51], videos [27]," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 118, + 471, + 242 + ], + "blocks": [ + { + "bbox": [ + 138, + 118, + 471, + 242 + ], + "lines": [ + { + "bbox": [ + 138, + 118, + 471, + 242 + ], + "spans": [ + { + "bbox": [ + 138, + 118, + 471, + 242 + ], + "type": "image", + "image_path": "f83dcbc494267fd326451141bc3f048acbce5aef71827d86cbb77962b0e072dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 250, + 482, + 297 + ], + "lines": [ + { + "bbox": [ + 130, + 250, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 482, + 297 + ], + "type": "text", + "content": "Fig. 1: Our method LEIA, takes in multi-view images of an object in four articulation states and is able to learn a view-invariant latent embedding for the state. We show that we can interpolate between the latents to generate any number of intermediate unseen states for the object using LEIA, given the camera position." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 321, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 501 + ], + "type": "text", + "content": "or assumptions regarding rigidity/shape of the object [17]. These approaches often fail to generalize for the long-tail of objects, especially in the case when a video of the object with articulation is unavailable, or in presence of large and multiple articulations in the object. In our work, we postulate that multiview images of an object in different states can provide enough signal to model its 3D shape and articulation, even without priors from large-scale foundation models or information from videos. We achieve this by learning a generalizable, view-invariant latent embedding of the object for different states. We train these embeddings jointly with a hypernetwork that predicts weights of a NeRF [19] parameterizing the object state. The hypernetwork can be trained using multiview images of discretized states of the object, each state representing a different object articulation. In Figure 1, we show three objects from our dataset in two different articulation states. The latent embeddings representing each state or articulation can be interpolated during test time to generate the weights of a NeRF that is able to reconstruct a never seen before state of the object." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 502, + 392, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 502, + 392, + 515 + ], + "spans": [ + { + "bbox": [ + 146, + 502, + 392, + 515 + ], + "type": "text", + "content": "Our key contributions can be summarized as following:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 517, + 480, + 663 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 138, + 517, + 480, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 517, + 480, + 539 + ], + "spans": [ + { + "bbox": [ + 138, + 517, + 480, + 539 + ], + "type": "text", + "content": "- We introduce an end-to-end method LEIA for generating novel states for articulated objects solely with multiview images captured at multiple states." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 543, + 480, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 543, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 137, + 543, + 480, + 600 + ], + "type": "text", + "content": "- We demonstrate that interpolating between embeddings can generate states of articulations of object not seen during training. The embedding space becomes interpolable with a manifold loss that encourages the latents to follow a structure that establishes a linear relationship between them, by minimizing the distance between the nearest neighbours in the latent space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 604, + 479, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 604, + 479, + 637 + ], + "spans": [ + { + "bbox": [ + 137, + 604, + 479, + 637 + ], + "type": "text", + "content": "- Remarkably, LEIA achieves this without the need for any ground-truth 3D supervision, motion information, or articulation codes, establishing its versatility and effectiveness in capturing complex articulations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 641, + 479, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 641, + 479, + 663 + ], + "spans": [ + { + "bbox": [ + 137, + 641, + 479, + 663 + ], + "type": "text", + "content": "- Our analyses demonstrate LEIA's robustness to single and multiple articulations, as well as combinations of motions. We can disentangle articulations" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 259, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 259, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 259, + 101 + ], + "type": "text", + "content": "A. Swaminathan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 151, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 151, + 116, + 479, + 140 + ], + "type": "text", + "content": "in different object parts if multiple are movable, making it scalable without constraints on the number of parts or motion types, unlike prior work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 159, + 236, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 159, + 236, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 159, + 236, + 172 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 186, + 481, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 186, + 481, + 330 + ], + "spans": [ + { + "bbox": [ + 133, + 186, + 481, + 330 + ], + "type": "text", + "content": "Neural Radiance Fields. Neural Radiance Fields [1, 19, 21, 42] have proven to be a massive success in modeling 3D scenes, due to the high fidelity of 3D reconstruction and novel view synthesis for static scenes. For dynamic scenes, Neural Volumes [18] utilizes an encoder-decoder voxel-based representation, complemented by an implicit voxel warp field. Occupancy Flow [24] tackled non-rigid geometry by assigning a motion vector to each point in both space and time, but requires full 3D ground truth supervision. Some of the first dynamic NeRF approaches [26, 28, 39] optimize an underlying volumetric deformable function and [15] conditions the NeRF on time. [16, 26, 45] followed this work, by learning a 5D spatiotemporal neural field. However, the approaches above usually require a video as input and do not handle well the case of large articulations in everyday objects given multiple states of objects as input." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 331, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 331, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 331, + 481, + 665 + ], + "type": "text", + "content": "3D Representations for Articulation. Due to the fine-grained nature of the task, deep neural network models for representing articulation require conditioning or refining parts of the architecture to suit the task. Category-specific reconstruction of deformable objects from images [7, 11-13], sparked interest in identifying and recovering the deformation in the 3D space [13, 22, 52]. Several works focus on shape reconstruction from videos [38, 47-49] that estimate a shape template for humans and animals. Other works that learn articulation from videos include Qian et al. [29] that detects and segments articulation planes from in-the-wild videos and [44, 53] which decouple the static and dynamic parts of videos. These works focus on modeling humans and animals where a large amount of data is available. Preliminary work in using images and shape started with A-SDF [28], which generates unseen articulations using Signed Distance Function [25], by providing an input shape and articulation code. CaDeX [14] is a unified representation for shape and motion, obtained from a point cloud input. Ditto [10] is a similar work that uses implicit representations for joint geometry and articulation modeling, with ground truth point clouds. CLA-NeRF [40] learns unseen articulated states from observing multi-view image input along with articulation information. Wei et. al [43] obtains an SDF-based articulable representation of common objects by feeding in images of articulated states across multiple categories, and CARTO [9] uses stereo images as input and uses a stereo encoder to infer the 3D shape, 6D pose etc. of multiple unknown objects. [4] focuses on manipulating object shapes to deform according to specified articulation commands. Moving away from 3D supervision and articulation annotations, Jiayi et al. [17] proposed PARIS, a method that is able to obtain unseen articulated states of an object, given multi-view images of just the start and end state of articulation. PARIS employs a composite rendering based approach, by decoupling the object into a static and moving part and then separately estimating and compositing the static and mobile neural radiance fields. In our work," + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "type": "text", + "content": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 190, + 482, + 264 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "text", + "content": "Table 1: Comparison of LEIA with existing methods. We show that LEIA is the first approach that does not use or learn any explicit prior along with not having any articulation input. This gives us flexibility in scaling to modeling articulations of objects with more than one part, and can thus handle a wide range of motion. We also have one universal model that can learn to represent both prismatic and revolute motion, unlike PARIS that has two separate models. 3D Sup. refers to \"3D Supervision\"." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 190, + 482, + 264 + ], + "lines": [ + { + "bbox": [ + 133, + 190, + 482, + 264 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 482, + 264 + ], + "type": "table", + "html": "
MethodImage/Shape InputDatasetArticulation Input3D Sup.?Prior# States# Parts
A-SDF [28]Shape CodeShape2MotionArticulation Code✓Articulation≥4≥1
CLA-NeRF [40]Multiview ImagesPartNet-MobilityArticulated PoseXArticulation≥41
NASAM [43]Multiview Images++PartNet-MobilityN/AXCategory≥4≥1
CARTO [9]Stereo ImagesCustomJoint Code✓Articulation≥41
Ditto [10]Point CloudsShape2MotionAnnotations✓Articulation21
PARIS [17]Multiview ImagesPartNet-MobilityN/AXMotion (learned)21
LEIAMultiview ImagesPartNet-MobilityN/A✓No4≥1
", + "image_path": "12d9798901219f9d0c3dbf3f1a86aba94019b9210b2c0b21528da6f3959d954f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 287, + 482, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 482, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 482, + 382 + ], + "type": "text", + "content": "we start LEIA with a similar input setting, by preparing multi-view renderings of objects in different articulation states. Unlike PARIS, we do not decouple the object into static and moveable parts, we rather learn to predict the weights of a single NeRF for any unseen articulated state using a state-modulated HyperNetwork. We aren't limited by the decoupling and the learned motion prior in the case of PARIS, which enables us to scale and learn any amount and kind of motion an object can possibly have, including combinations. We show a summed comparison of LEIA with prior work in Table 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 383, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 482, + 597 + ], + "type": "text", + "content": "Hypernetworks with Implicit Neural Representations. Hypernetworks, a specialized class of networks designed to predict parameters for another network, aim to achieve generalization across novel tasks [8]. In the realm of stylizing 3D scenes, [3] employed a hypernetwork for applying diverse styles, while scene reconstructions from limited data points were achieved by [35, 36]. Despite the promise shown in representation, these hypernetworks operate on input data points, necessitating test-time optimizations and rendering them unsuitable for compression tasks. Rather than using the provided data (image/video) as input for the hypernetwork, an alternative approach involves employing an auto-decoder framework. In this framework, a learnable latent, without the need for an encoder, represents a data point. This technique, applied by [33] to represent a dataset of videos, assigns each latent to a distinct video. While this method yields a representation for each set of frames, the lack of decoupling in spatial-temporal coordinates limits its scalability to real-world frames. On a related note, [34] extended a similar approach to 3D shapes and scenes, effectively acquiring a latent representation suitable for various downstream tasks. We use a similar framework for learning the latents corresponding to each articulated state in our method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 616, + 203, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 203, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 203, + 628 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "**Background.** Our architecture is based on neural radiance fields, or NeRF [19], which parameterizes the radiance and volume density at a 3D location of a" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 259, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 259, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 259, + 102 + ], + "type": "text", + "content": "A. Swaminathan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": "scene as observed from a camera placed in a particular position using a neural network. " + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{NeRF}}:(\\mathbf{x},\\mathbf{d}) \\to (\\mathbf{c},\\sigma)" + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "inline_equation", + "content": "\\mathbf{c} = (r,g,b)" + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": " represent the radiance and volume density, and " + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": " represent the 3D location and viewing direction respectively. To render a pixel, the radiance " + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "inline_equation", + "content": "C(\\mathbf{r})" + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": " of a camera ray " + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(s) = \\mathbf{o} + s \\cdot \\mathbf{d}" + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": " is integrated from near to far bounds " + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "inline_equation", + "content": "s_n" + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "inline_equation", + "content": "s_f" + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": " such that, " + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "inline_equation", + "content": "C(\\mathbf{r}) = \\int_{s_n}^{s_f} T(s)\\sigma(\\mathbf{r}(s))\\mathbf{c}(\\mathbf{r}(s),\\mathbf{d})ds" + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": ", where the function " + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "inline_equation", + "content": "T(s)" + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": " denotes the accumulated transmittance from " + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "inline_equation", + "content": "s_n" + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": ". To optimize the parameters of the MLP, a loss function is used that measures the discrepancy between the ground-truth and rendered images. Traditionally, the L2 loss is used for this purpose. However, to make the training more robust to outliers and to improve convergence, the Smooth L1 Loss can be employed as an alternative. The Smooth L1 Loss is a combination of L1 and L2 losses, behaving like L1 loss for large errors and like L2 loss for small errors. The loss function for the NeRF model using the Smooth L1 Loss is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 213, + 286, + 481, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 286, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 213, + 286, + 481, + 317 + ], + "type": "interline_equation", + "content": "\\operatorname {S m o o t h L 1 L o s s} (x) = \\left\\{ \\begin{array}{l l} 0. 5 \\cdot x ^ {2} & \\text {i f} | x | < 1 \\\\ | x | - 0. 5 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {1}", + "image_path": "11711539b42167f8a1646cfc8430a0874eead45ca056f59f433912641feeabda.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 194, + 323, + 480, + 349 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 323, + 480, + 349 + ], + "spans": [ + { + "bbox": [ + 194, + 323, + 480, + 349 + ], + "type": "interline_equation", + "content": "L _ {\\text {S m o o t h L 1 - N e R F}} = \\sum_ {r \\in R} \\text {S m o o t h L 1 L o s s} \\left(\\hat {C} (\\mathbf {r}) - C (\\mathbf {r})\\right) \\tag {2}", + "image_path": "659f3eece9ee4696b9b878e64c088c5ebc4e706f91fc91e827296432e36a324a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 354, + 482, + 486 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 354, + 482, + 486 + ], + "type": "text", + "content": " is the set of rays used for sampling, and " + }, + { + "bbox": [ + 130, + 354, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\hat{C}(\\mathbf{r})" + }, + { + "bbox": [ + 130, + 354, + 482, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 354, + 482, + 486 + ], + "type": "inline_equation", + "content": "C(\\mathbf{r})" + }, + { + "bbox": [ + 130, + 354, + 482, + 486 + ], + "type": "text", + "content": " represent the ground-truth and rendered colors, respectively. By minimizing this loss function, the NeRF model learns to accurately reproduce the radiance of the scene, leading to high-quality image synthesis from novel viewpoints. While this formulation works well for generating novel scene viewpoints for a static scene, it cannot handle dynamic scenes. [45] tried to address this shortcoming by making the MLP learn a spatiotemporal radiance field and used time, " + }, + { + "bbox": [ + 130, + 354, + 482, + 486 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 354, + 482, + 486 + ], + "type": "text", + "content": ", as an additional input. This works in principle but it becomes expensive to scale to lengthier videos due to constantly increasing sampling space for NeRF. We use the " + }, + { + "bbox": [ + 130, + 354, + 482, + 486 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{SmoothL1 - NeRF}}" + }, + { + "bbox": [ + 130, + 354, + 482, + 486 + ], + "type": "text", + "content": " loss along with AdamW optimization to train LEIA, along with a " + }, + { + "bbox": [ + 130, + 354, + 482, + 486 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{mask}}" + }, + { + "bbox": [ + 130, + 354, + 482, + 486 + ], + "type": "text", + "content": " that is the BCE loss between the predicted opacity and ground truth foreground loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 499, + 211, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 499, + 211, + 511 + ], + "spans": [ + { + "bbox": [ + 131, + 499, + 211, + 511 + ], + "type": "text", + "content": "3.1 Approach" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 515, + 481, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 481, + 563 + ], + "type": "text", + "content": "Hypernetworks, or hypernets for short, are neural networks that generate weights for another neural network, known as the target network. Hypernets can be conditioned on various domains and can predict the weights for multiple neural networks simultaneously, if trained appropriately." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "text", + "content": "In this work, we use a hypernet to modulate a Neural Radiance Field (NeRF) based on the state of articulation, thereby learning a parametrization for each state. We employ a learnable latent embedding, " + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "text", + "content": ", as input to this hypernet " + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "inline_equation", + "content": "h_l, l \\in L" + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "text", + "content": " is the number of layers of the hypernet. This approach allows us to not only modulate the NeRF but also create useful representations for each state. Our system can thus be represented as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 254, + 640, + 481, + 660 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 640, + 481, + 660 + ], + "spans": [ + { + "bbox": [ + 254, + 640, + 481, + 660 + ], + "type": "interline_equation", + "content": "F _ {\\theta_ {t}} (\\mathbf {x}, \\mathbf {d}) = \\left(\\mathbf {c} _ {t}, \\sigma_ {t}\\right) \\tag {3}", + "image_path": "bc4ae9c9561100147d1743e9dc8e4ac69fb9df3c507ff013222eb55aee731085.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 284, + 656, + 361, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 656, + 361, + 666 + ], + "spans": [ + { + "bbox": [ + 284, + 656, + 361, + 666 + ], + "type": "interline_equation", + "content": "\\theta_ {t} = h (z _ {t}), z _ {t} \\in Z", + "image_path": "6037b65e0be39a7b9289e896cce509bad4de1e526986f8e55a780d31204ab2f3.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 160, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 160, + 91, + 448, + 102 + ], + "type": "text", + "content": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 483, + 244 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 483, + 244 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 483, + 244 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 483, + 244 + ], + "type": "image", + "image_path": "a2588fda4b5e7f167ce848216164e9c33e2ebada7f090c2a663771e2ec4e6663.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 251, + 482, + 318 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 318 + ], + "type": "text", + "content": "Fig. 2: Overview of our method. We take multi-view images in different states as input. A learnable latent dictionary based off an autoencoder learns an embedding per state id. The latent embedding is used as an input to the hypernet, that modulates and generates weights of the NeRF to reconstruct the state that is fed in. At inference time, we do a weighed interpolation of the learnt latents to obtain a corresponding newly generated intermediate state." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 341, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 389 + ], + "type": "text", + "content": "The latents " + }, + { + "bbox": [ + 130, + 341, + 482, + 389 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 341, + 482, + 389 + ], + "type": "text", + "content": ", give us an additional ability to interpolate across seen states and generate novel states of the object, as we show in our work. The set of latents, " + }, + { + "bbox": [ + 130, + 341, + 482, + 389 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 341, + 482, + 389 + ], + "type": "text", + "content": ", in Eq. (3) can be understood as a learnable dictionary where keys are the state ids. Formally, we represent this as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 242, + 398, + 481, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 398, + 481, + 412 + ], + "spans": [ + { + "bbox": [ + 242, + 398, + 481, + 412 + ], + "type": "interline_equation", + "content": "Z = \\left\\{t: z _ {t} \\mid t \\in [ 0, 1, 2 \\dots T ] \\right\\} \\tag {4}", + "image_path": "60c4f48e939bb2bc025601886114bd9448314242ffe9c6537e76afa38d530b61.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 420, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 482, + 491 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 130, + 420, + 482, + 491 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 420, + 482, + 491 + ], + "type": "text", + "content": " is the total number of discretized states sampled for the object. Each of these states, " + }, + { + "bbox": [ + 130, + 420, + 482, + 491 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 420, + 482, + 491 + ], + "type": "text", + "content": ", is used as an input to the hypernet to get the parameterization for NeRF as shown in Figure 2. Once parameterized, the NeRF is trained using multi-view images of the object taken from various camera angles, providing a comprehensive view of the object's articulation in the current state. We sample only one state per batch." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 492, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 482, + 564 + ], + "type": "text", + "content": "Directly predicting the weights " + }, + { + "bbox": [ + 130, + 492, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 492, + 482, + 564 + ], + "type": "text", + "content": " of the base network " + }, + { + "bbox": [ + 130, + 492, + 482, + 564 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 130, + 492, + 482, + 564 + ], + "type": "text", + "content": ", using the hypernet " + }, + { + "bbox": [ + 130, + 492, + 482, + 564 + ], + "type": "inline_equation", + "content": "h_l" + }, + { + "bbox": [ + 130, + 492, + 482, + 564 + ], + "type": "text", + "content": ", is expensive, parameter-heavy, and unsuitable for compression. Hence, we follow [32, 37] and instead predict low-rank matrices, which are then applied to the base network weights. This type of modulation acts as a form of subnetwork selection, analogous to systems proposed in [6, 30]. For a base network " + }, + { + "bbox": [ + 130, + 492, + 482, + 564 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 130, + 492, + 482, + 564 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 492, + 482, + 564 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 492, + 482, + 564 + ], + "type": "text", + "content": " layers, our formulation now looks like" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 218, + 571, + 352, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 571, + 352, + 586 + ], + "spans": [ + { + "bbox": [ + 218, + 571, + 352, + 586 + ], + "type": "interline_equation", + "content": "f _ {\\theta} \\left(\\left(\\mathbf {x}, \\mathbf {d}\\right) \\mid \\theta_ {t} ^ {l _ {1}}, \\theta_ {t} ^ {l _ {2}} \\dots \\theta_ {t} ^ {l _ {L}}\\right) = \\mathbf {c} _ {t}, \\sigma_ {t}", + "image_path": "12b41b6da50ac7d16dc625be5f9799cbefe2d92550b0a921530012e81235cc09.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 588, + 480, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 588, + 480, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 480, + 602 + ], + "type": "interline_equation", + "content": "\\theta_ {t} ^ {l} = \\eta \\left(P ^ {l} \\times Q ^ {l}\\right) \\circ \\theta^ {l} \\tag {5}", + "image_path": "6752c185c1b2b6ab9dcfad6fcdbc5c117e61c2b85733e07be6438066dc87cd23.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 288, + 605, + 361, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 605, + 361, + 620 + ], + "spans": [ + { + "bbox": [ + 288, + 605, + 361, + 620 + ], + "type": "interline_equation", + "content": "h _ {l} (z _ {t}) = [ P ^ {l}, Q ^ {l} ]", + "image_path": "b0c0604d821b4c4fe994532f101c1e8f87ffd9de38fad42e1677efd5bf6ac053.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\theta^l" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": " represents the weights of the " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": "-th layer and " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\theta_t^l" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": " denotes the modulated weights for frame " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": " signifies an activation function on the matrix-product of low rank matrices " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^l \\in \\mathbb{R}^{K \\times r}" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^l \\in \\mathbb{R}^{r \\times K}" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 628, + 482, + 666 + ], + "type": "text", + "content": " is the width of the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 259, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 259, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 259, + 101 + ], + "type": "text", + "content": "A. Swaminathan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "base network " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " and rank " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "r \\ll K" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ". These matrices are responsible for adjusting the weights " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\theta_{l}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " as dictated by the corresponding hypernetwork " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "h_l" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ". Note that all hypernetworks use the same latent " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "z_{t} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " as input. The rank " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " acts as a hyperparameter that controls the compression-performance trade-off. We further elaborate on details about model architecture and design choices for the hypernet, NeRF and the learnable latent dictionary in the supplementary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 205, + 228, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 205, + 228, + 217 + ], + "spans": [ + { + "bbox": [ + 131, + 205, + 228, + 217 + ], + "type": "text", + "content": "3.2 Interpolation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 225, + 482, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 298 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 298 + ], + "type": "text", + "content": "Given an object with states " + }, + { + "bbox": [ + 130, + 225, + 482, + 298 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 130, + 225, + 482, + 298 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 225, + 482, + 298 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 130, + 225, + 482, + 298 + ], + "type": "text", + "content": ", and a scale " + }, + { + "bbox": [ + 130, + 225, + 482, + 298 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 225, + 482, + 298 + ], + "type": "text", + "content": ", the task of state interpolation involves creating " + }, + { + "bbox": [ + 130, + 225, + 482, + 298 + ], + "type": "inline_equation", + "content": "\\alpha - 1" + }, + { + "bbox": [ + 130, + 225, + 482, + 298 + ], + "type": "text", + "content": " coherent states between the given states. In order to achieve this, we do a linear interpolation on the state latents " + }, + { + "bbox": [ + 130, + 225, + 482, + 298 + ], + "type": "inline_equation", + "content": "z_1" + }, + { + "bbox": [ + 130, + 225, + 482, + 298 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 225, + 482, + 298 + ], + "type": "inline_equation", + "content": "z_2" + }, + { + "bbox": [ + 130, + 225, + 482, + 298 + ], + "type": "text", + "content": " and pass the resulting latent through the hypernetwork. This gives us the weight modulation required in the NeRF, and the updated base network is used to obtain different viewpoints of the intermediate state." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 255, + 308, + 481, + 325 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 308, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 255, + 308, + 481, + 325 + ], + "type": "interline_equation", + "content": "z _ {\\text {i n t e r}} = \\left(1 - \\beta_ {i}\\right) \\cdot z _ {t} + \\beta_ {i} \\cdot z _ {t - 1} \\tag {6}", + "image_path": "11cb2c84d6206f93415cc838cc947fd1c30dceb8f654286e3da094b6e3a09b7d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 228, + 323, + 378, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 323, + 378, + 334 + ], + "spans": [ + { + "bbox": [ + 228, + 323, + 378, + 334 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {\\text {i n t e r}}, \\sigma_ {\\text {i n t e r}} = f _ {\\theta_ {\\text {i n t e r}}} (\\mathbf {x}, \\mathbf {d}; h (z _ {\\text {i n t e r}}))", + "image_path": "14b81ef9af353f6df0a5c82f5f33efb9b57b7bc8f4ad85e12be8e940f7fc2db0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 242, + 345, + 480, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 345, + 480, + 371 + ], + "spans": [ + { + "bbox": [ + 242, + 345, + 480, + 371 + ], + "type": "interline_equation", + "content": "\\text {w h e r e}, \\beta_ {i} \\in \\left[ \\frac {1}{\\alpha}, \\frac {2}{\\alpha}, \\dots , \\frac {\\alpha - 1}{\\alpha} \\right] \\tag {7}", + "image_path": "ec9cf3575cdc0b80951e67f6f0b317d1d7b51d8816e3ed5b42542ca292afde17.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 375, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 422 + ], + "type": "text", + "content": "essentially generating " + }, + { + "bbox": [ + 130, + 375, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\alpha - 1" + }, + { + "bbox": [ + 130, + 375, + 482, + 422 + ], + "type": "text", + "content": " states between any two given states. In our experiments, we quantitatively evaluate the result of the interpolation with unseen ground truth obtained from our dataset of all states, by linearly averaging the two extreme states present in the training, to obtain a latent for a state id." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 441, + 261, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 441, + 261, + 451 + ], + "spans": [ + { + "bbox": [ + 131, + 441, + 261, + 451 + ], + "type": "text", + "content": "3.3 Model Architecture" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "type": "text", + "content": "In our experiments, both the base network " + }, + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "type": "text", + "content": " and hypernetworks " + }, + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "type": "inline_equation", + "content": "h_l" + }, + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "type": "text", + "content": " are simple MLPs that take in a coordinate input, and an id depicting the state of articulation. A network " + }, + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "type": "inline_equation", + "content": "l_n" + }, + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "type": "text", + "content": " is used to learn the latent dictionary for the states, which is a linear layer learnt in conjunction with the hypernet, and uses the torch.nn.Embedding class as a lookup table for the learnable latent embeddings, depicted by " + }, + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "type": "inline_equation", + "content": "t \\in T" + }, + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "type": "text", + "content": ". This latent is fed to the hypernetwork " + }, + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "type": "inline_equation", + "content": "h_l" + }, + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "type": "text", + "content": ", that modulates the weights of the base network " + }, + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "type": "text", + "content": " for reconstructing the output of the corresponding state " + }, + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "type": "inline_equation", + "content": "\\mathrm{id}" + }, + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "type": "text", + "content": ". The base network " + }, + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 130, + 460, + 482, + 591 + ], + "type": "text", + "content": " is based off the NeRF architecture of Instant-NGP [21]. This architecture has separate fully-connected blocks for the geometry and texture of the scene, and we use separate hypernets to modulate each layer of these two FC blocks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 609, + 389, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 389, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 389, + 621 + ], + "type": "text", + "content": "3.4 Optimization: Loss functions and regularizers" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Latent Manifold Loss. We use the latent manifold loss function that enforces a structured latent space by encouraging local consistency among learnt latents. For each latent vector, the loss computes the average Euclidean distance between" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 160, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 160, + 91, + 448, + 103 + ], + "type": "text", + "content": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "the vector and its " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " nearest-neighbours on the manifold. This process enforces a smooth and continuous latent manifold, which is beneficial for models that rely on meaningful linear interpolations between the points on the manifold, and for tasks where the geometry of the latent space is crucial. Mathematically, we represent this loss for a particular state-id " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 238, + 188, + 481, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 188, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 238, + 188, + 481, + 220 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a n i f o l d}} \\left(\\mathbf {l} _ {i}\\right) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\| \\mathbf {l} _ {i} - \\mathbf {n} _ {k} \\| _ {2} ^ {2} \\tag {8}", + "image_path": "320abfd5e015770940a2e0f1ce3f778c8b4c0d23d3cfda7d37f8e742d363fa43.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 230, + 482, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 230, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 230, + 482, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 230, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_k" + }, + { + "bbox": [ + 131, + 230, + 482, + 277 + ], + "type": "text", + "content": " are the " + }, + { + "bbox": [ + 131, + 230, + 482, + 277 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 230, + 482, + 277 + ], + "type": "text", + "content": " nearest neighbors of " + }, + { + "bbox": [ + 131, + 230, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_i" + }, + { + "bbox": [ + 131, + 230, + 482, + 277 + ], + "type": "text", + "content": " in the latent space, and " + }, + { + "bbox": [ + 131, + 230, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_2" + }, + { + "bbox": [ + 131, + 230, + 482, + 277 + ], + "type": "text", + "content": " denotes the Euclidean (L2) norm. The loss is averaged over the selected latents and their nearest neighbors to ensure local uniformity in the manifold's geometry, where " + }, + { + "bbox": [ + 131, + 230, + 482, + 277 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 230, + 482, + 277 + ], + "type": "text", + "content": " is a hyperparameter that determines the number of neighbors considered." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 279, + 483, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 279, + 483, + 387 + ], + "spans": [ + { + "bbox": [ + 131, + 279, + 483, + 387 + ], + "type": "text", + "content": "Depth and Occlusion Regularization. Our depth and occlusion regularizations are designed to refine the clarity of rendered images by addressing occlusion and depth smoothness [23, 50]. The occlusion regularization loss, " + }, + { + "bbox": [ + 131, + 279, + 483, + 387 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{occ}}" + }, + { + "bbox": [ + 131, + 279, + 483, + 387 + ], + "type": "text", + "content": ", aims to mitigate the obscuring of objects located beyond a specified depth threshold during rendering. This is accomplished by generating a binary mask, " + }, + { + "bbox": [ + 131, + 279, + 483, + 387 + ], + "type": "inline_equation", + "content": "m_k" + }, + { + "bbox": [ + 131, + 279, + 483, + 387 + ], + "type": "text", + "content": ", where the mask elements are set to 1 up to a certain index " + }, + { + "bbox": [ + 131, + 279, + 483, + 387 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 279, + 483, + 387 + ], + "type": "text", + "content": " reflecting the regularization range, and 0 thereafter. The occlusion loss is then articulated as the normalized sum of the product of the mask and the sampled density values " + }, + { + "bbox": [ + 131, + 279, + 483, + 387 + ], + "type": "inline_equation", + "content": "\\sigma_k" + }, + { + "bbox": [ + 131, + 279, + 483, + 387 + ], + "type": "text", + "content": " along a ray, as given by" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 260, + 389, + 481, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 389, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 260, + 389, + 481, + 422 + ], + "type": "interline_equation", + "content": "L _ {\\text {o c c}} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\sigma_ {k} \\cdot m _ {k}, \\tag {9}", + "image_path": "6e90872714def90045e182a2cd858f0112e7c3d75c6e2b9828a6cc54b9ad09d7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 429, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 429, + 482, + 464 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 429, + 482, + 464 + ], + "type": "text", + "content": " is the total number of sampled points on the ray. This formulation drives the model to prefer representations that reduce occlusions close to the camera, ensuring objects further away are not improperly concealed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 465, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 465, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 465, + 482, + 538 + ], + "type": "text", + "content": "For depth continuity, the depth smoothness regularization loss, " + }, + { + "bbox": [ + 130, + 465, + 482, + 538 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{DS}}(\\theta ,\\mathbf{R})" + }, + { + "bbox": [ + 130, + 465, + 482, + 538 + ], + "type": "text", + "content": " enforces the gradual transition of depth values among neighboring pixels, reducing sharp depth disparities that cause visual inconsistencies. If a ray " + }, + { + "bbox": [ + 130, + 465, + 482, + 538 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 465, + 482, + 538 + ], + "type": "text", + "content": " intersects with a single depth value, this value is evaluated directly. In contrast, for multiple depth values, the loss is the aggregate of squared differences between adjacent depth estimates. Formally, for depth predictions " + }, + { + "bbox": [ + 130, + 465, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\hat{d}_0(r)" + }, + { + "bbox": [ + 130, + 465, + 482, + 538 + ], + "type": "text", + "content": ", the loss is quantified as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 550, + 481, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 550, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 143, + 550, + 481, + 594 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {D S}} (\\theta , \\mathbf {R}) = \\sum_ {r \\in \\mathbf {R}} \\sum_ {i, j = 1} ^ {\\text {P a t c h} - 1} \\left(\\hat {d} _ {0} \\left(r _ {i j}\\right) - \\hat {d} _ {0} \\left(r _ {(i + 1) j}\\right)\\right) ^ {2} + \\left(\\hat {d} _ {0} \\left(r _ {i j}\\right) - \\hat {d} _ {0} \\left(r _ {i (j + 1)}\\right)\\right) ^ {2}. \\tag {10}", + "image_path": "afbd7cc973429ddef9bd955bea908a44241f5f55a29ae10db80ea19bd2d2417b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "This measure is averaged across the rays to yield a measure of the depth map's smoothness. By integrating these regularizers into our training regimen, we significantly dampen the disturbances introduced by overlapping and obscured pixels, resulting in a more consistent interpolation between states. These benefits are substantiated through extensive ablation studies and quantitative analyses presented in the experiments." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 259, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 259, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 259, + 101 + ], + "type": "text", + "content": "A. Swaminathan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "Positional Encoding of the Latent. We incorporate positional encoding to capture the order of input elements, crucial for understanding articulation. Positional encoding injects the sequence with its inherent order, a key factor in articulation semantics. We employ the scheme from [41], which uses sine and cosine functions parameterized to encode varying frequencies. The computed positional encodings are added to the latent vector " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ", enriching it with semantic information that more accurately encodes the state of articulation. We analyze the effects of positional encoding on our model in the experimental section." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 236, + 230, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 236, + 230, + 249 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 230, + 249 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 266, + 191, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 266, + 191, + 278 + ], + "spans": [ + { + "bbox": [ + 131, + 266, + 191, + 278 + ], + "type": "text", + "content": "4.1 Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 291, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 482, + 506 + ], + "type": "text", + "content": "Datasets. In this work, we use the PartNet-Mobility dataset [2, 20, 46], a large-scale synthetic collection of articulated objects in over 40 categories. The dataset has a variety of articulations defined, with objects comprising of single and multi-joint parts. Two types of motion, revolute (rotational) joint and prismatic (translation) articulation are represented in the dataset. To make up our training dataset, we choose 100 camera views that are arranged in a dome-like setup, capturing the upper hemisphere of the object. This is similar to the setup used by PARIS. We use the SAPIEN [46] library that the PartNet-Mobility dataset was released with, to render our RGB images at linearly spaced intervals of articulation to make up frames of a video that depicts the range of motion and obtain the camera parameters accordingly, converting it into the Blender [5] coordinate system to fit in our codebase. We use instances from 8 different common household item categories storage, microwave, laptop, oven, washer, dishwasher, sunglasses, box and choose 1-4 objects per category, bringing our total number of objects to 12. We also show the efficacy of our method on 68 images from a real-world scene of a chest of drawers, which are captured with a mobile phone and post-processed to remove background. Throughout LEIA, we train with a total of four states and interpolate between the two extreme states." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 508, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 579 + ], + "type": "text", + "content": "Baselines. The closest prior work with a setup similar to ours is PARIS, which takes in multi-view images of objects in two states and disentangles the static and moving part of the object. While we run experiments with four input states, we don't focus on learning a defined articulation for the object, instead we are able to recognize any arbitrary motion and combinations of motions given relevant states. We emphasize on the following differences:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 580, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 135, + 580, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 580, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 135, + 580, + 481, + 616 + ], + "type": "text", + "content": "1. PARIS uses separate models to train objects that have rotation and translation motion, respectively. While they have a method to estimate the type of motion, once determined, it is necessary to train on the appropriate model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 135, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 617, + 481, + 665 + ], + "type": "text", + "content": "2. As PARIS also learns motion parameters for the articulation, given their pipeline of disentanglement of the static and moving parts of the object, this restricts the learning of motion parameters if there are multiple parts in the object, moving differently." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "type": "text", + "content": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 223, + 482, + 345 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "text", + "content": "Table 2: Quantitative Results for Interpolated State Reconstruction. We compared our method with the PARIS baseline, trained on selected objects from the SAPIEN dataset. The results from three experiments are summarized below. The VanillaInt experiment involves simple interpolation of the latents. Our best-performing method, LEIA, introduces structure to the latents with a manifold loss and regularizers. Although PARIS learns a motion prior and LEIA implicitly performs state interpolation, both methods perform similarly for single-part objects. However, our approach excels with multi-part objects, outperforming PARIS significantly due to its flexibility in handling various motion types and articulations without constraints." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 223, + 482, + 345 + ], + "lines": [ + { + "bbox": [ + 133, + 223, + 482, + 345 + ], + "spans": [ + { + "bbox": [ + 133, + 223, + 482, + 345 + ], + "type": "table", + "html": "
MetricsMethodsSingle-Part ArticulationMulti-Part ArticulationAverage
45135 Storage17128 Microwave10211 Laptop101917 Oven103778 Washer12085 Dishwasher44781 Storage245427 Storage345575 Storage4101297 Sunglasses7187 Oven102377 Box
PSNR↑PARIS28.6625.9424.9728.1334.4627.2426.3524.4125.7332.3329.3026.2027.81
VanillaInt24.2022.0022.6324.5336.0324.1630.9629.7829.1032.1130.3527.8527.81
LEIA26.6926.3825.0428.7136.1426.4931.0729.7829.8035.6030.8028.0529.55
SSIM↑PARIS0.990.970.970.970.980.960.950.950.940.960.960.930.96
VanillaInt0.950.910.900.930.990.920.950.950.930.970.950.930.94
LEIA0.970.950.950.980.990.950.960.950.950.970.960.950.96
LPIPS↓PARIS0.020.060.190.030.030.060.050.050.040.060.050.080.06
VanillaInt0.050.100.170.030.020.080.030.030.050.090.050.130.07
LEIA0.030.070.100.020.020.070.030.030.040.090.050.120.06
CD↓PARIS0.180.080.120.750.060.450.620.420.600.200.990.910.45
VanillaInt0.240.070.060.100.060.200.480.470.460.970.500.830.37
LEIA0.290.060.360.100.050.270.520.380.410.960.350.620.36
", + "image_path": "6048f10e8cf9c094aee6378a3f22bb957f3fd86ce2b24b61ac8a29d1f1e06efb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 368, + 481, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 368, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 135, + 368, + 481, + 403 + ], + "type": "text", + "content": "3. LEIA is motion-prior free, and we have a universal architecture that implicitly learns intermediate states, so we are able to train with multiple types of motion occurring in the same object, with no change in the code." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 405, + 482, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 482, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 482, + 513 + ], + "type": "text", + "content": "For comparing with PARIS, we run authors' original source code with our dataset, using the hyperparameters provided by the authors and training until convergence. We test our work and PARIS with the appearance quality of the reconstructed image of the interpolated state. We also set up a simple baseline that does just vanilla interpolation between the learnt latent embeddings, without enforcing any structure or constraints on them. We call this baseline VanillaInt and compare it with LEIA, which has been finetuned to do interpolation with the addition of the manifold loss and the depth and occlusion regularization for denoising the resulting output." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 513, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 513, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 513, + 482, + 586 + ], + "type": "text", + "content": "Quantitative Metrics. We use three appearance quality metrics, PSNR (Peak Signal-to-Noise Ratio), SSIM (Structural Similarity Index), and LPIPS (Learned Perceptual Image Patch Similarity). These are computed between reconstructed images of the interpolated state, and the unseen ground truth of the state from our dataset. We also report the Chamfer Distance to measure quality of 3D reconstruction by doing point sampling." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 606, + 228, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 606, + 228, + 617 + ], + "spans": [ + { + "bbox": [ + 132, + 606, + 228, + 617 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Novel State Synthesis. We show the qualitative and quantitative results of interpolating our learned latent embeddings corresponding to the start and end states, in Table 2 and Figure 3. Our results show that the latent embeddings can" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "type": "text", + "content": "A. Swaminathan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 154, + 114, + 449, + 522 + ], + "blocks": [ + { + "bbox": [ + 154, + 114, + 449, + 522 + ], + "lines": [ + { + "bbox": [ + 154, + 114, + 449, + 522 + ], + "spans": [ + { + "bbox": [ + 154, + 114, + 449, + 522 + ], + "type": "image", + "image_path": "4ad50401da0834031c7f891aec74172bd3b21281ab7b9416271fb3d760158546.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 540, + 482, + 617 + ], + "lines": [ + { + "bbox": [ + 130, + 540, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 482, + 617 + ], + "type": "text", + "content": "Fig. 3: Qualitative Results. We show results of PARIS and LEIA for reconstructing the unseen intermediate state, for both single and multiple articulations. We see that PARIS especially fails when there are two parts of the object moving differently, as the motion parameters are not registered correctly. LEIA handles this case successfully as it is not dependent on part disentanglement to identify and register articulation. LEIA also performs comparable to PARIS for single-part articulation, despite not having a dedicated model for the motion or part disentanglement." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "be linearly interpolated, and have a structure in the latent space. We can generate any number of states between the start and the end state by doing a weighted" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "type": "text", + "content": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 125, + 462, + 181 + ], + "blocks": [ + { + "bbox": [ + 151, + 125, + 462, + 181 + ], + "lines": [ + { + "bbox": [ + 151, + 125, + 462, + 181 + ], + "spans": [ + { + "bbox": [ + 151, + 125, + 462, + 181 + ], + "type": "image", + "image_path": "715c8b2973ffca7878e9bc50c48b5382415e843156db77caaf2097ba0500265c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 193, + 482, + 228 + ], + "lines": [ + { + "bbox": [ + 131, + 193, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 131, + 193, + 482, + 228 + ], + "type": "text", + "content": "Fig. 4: Real World Results. LEIA is able to faithfully interpolate and reconstruct between two states of images from our real world data, proving its ability to generalize and work in an in-the-wild setting." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 251, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 430 + ], + "type": "text", + "content": "combination of the latents representing the start and the end states. For training, we choose four states, using the two additional states compared to PARIS to enforce structure in the latent space with the latent manifold loss, thereby making them amenable to interpolation. This helps us extend our formulation to multiple articulations without restriction. We also show results of our baseline VanillaInt and show that the numbers are boosted with the addition of the latent manifold loss and regularizers. For single-part objects, we perform comparably to PARIS despite having no part disentanglement and no dedicated model to represent the exact type of motion. For multiple-part articulations, we beat PARIS across all objects, as shown in Figure 3. Our comparable performance in 3D reconstruction shows that LEIA's learnt view-invariant latent embeddings are powerful in preserving spatio-temporal consistency. We also show that our method can work in the real world setting, in Figure 4. Ablations are shown for number of states, latent manifold loss, and depth and occlusion regularizations in Table 3 and Figure 6." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "text", + "content": "Analysis on Disentanglement of Joints. To check separability of the latent space, we train LEIA with multi-view images across multiple states, captured from a video that showed three parts of an object Storage1, move one after the other. We do a t-SNE based dimensionality reduction on the embeddings and plot them, shown in Figure 5. The latents exhibit clear separability, with the motion of each joint represented in a smooth trajectory, moving outward as the respective drawer in the object moves away from the starting \"closed\" state. This shows that the separated latent space is an indicator of multiple joints, positioning us to be capable of learning representations of multiple articulations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 559, + 250, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 250, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 250, + 571 + ], + "type": "text", + "content": "4.3 Ablation Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "We perform ablations to tune various design choices, with results shown in Table 3 and Figure 6. The first experiment we performed was testing out the impact of the latent manifold loss. We notice qualitatively that without the manifold loss, our latent space lacked structure and the linear relationship between the embeddings was not captured. As shown in Figure 6, the intermediate state overfits to the extreme state without the latent manifold while the addition of the loss enables it to accurately capture the intermediate state. This figure also" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "type": "text", + "content": "A. Swaminathan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 119, + 277, + 262 + ], + "blocks": [ + { + "bbox": [ + 135, + 119, + 277, + 262 + ], + "lines": [ + { + "bbox": [ + 135, + 119, + 277, + 262 + ], + "spans": [ + { + "bbox": [ + 135, + 119, + 277, + 262 + ], + "type": "image", + "image_path": "bbd16fba78450d42a9ec7c3bf80b66b2c60b3fccff5869e70baaf030834d920a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 264, + 291, + 330 + ], + "lines": [ + { + "bbox": [ + 132, + 264, + 291, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 264, + 291, + 330 + ], + "type": "text", + "content": "Fig. 5: t-SNE plot. After dimensionality reduction on jointly-learned state embeddings of an object with different moving parts. Our learned representations are separated and follow a smooth trajectory for each of the moving parts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 309, + 182, + 478, + 338 + ], + "blocks": [ + { + "bbox": [ + 304, + 117, + 482, + 171 + ], + "lines": [ + { + "bbox": [ + 304, + 117, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 304, + 117, + 482, + 171 + ], + "type": "text", + "content": "Table 3: Ablations. Our ablations reveal that latent manifold loss, depth, and occlusion regularization enhance LEIA's visual metrics. Opting for four states improved latent structure with the manifold loss." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 182, + 478, + 338 + ], + "lines": [ + { + "bbox": [ + 309, + 182, + 478, + 338 + ], + "spans": [ + { + "bbox": [ + 309, + 182, + 478, + 338 + ], + "type": "table", + "html": "
Latent Manifold LossPSNRSSIMLPIPS
with29.400.950.05
without28.540.940.06
Depth RegularizationPSNRSSIMLPIPS
with29.630.960.05
without26.930.930.07
Occlusion RegularizationPSNRSSIMLPIPS
with29.640.950.05
without28.640.950.06
Positional encodingPSNRSSIMLPIPS
with27.110.940.06
without28.480.950.06
Number of StatesPSNRSSIMLPIPS
228.040.950.06
429.690.960.05
", + "image_path": "11557f08838939f2233e75026b499749108d281efb43efc012e2102fa5273a25.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 134, + 358, + 470, + 548 + ], + "blocks": [ + { + "bbox": [ + 134, + 358, + 470, + 548 + ], + "lines": [ + { + "bbox": [ + 134, + 358, + 470, + 548 + ], + "spans": [ + { + "bbox": [ + 134, + 358, + 470, + 548 + ], + "type": "image", + "image_path": "a68404074e403e14993805059219a8ad50f682219ca07963196a48066b0a9038.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 557, + 482, + 612 + ], + "lines": [ + { + "bbox": [ + 132, + 557, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 482, + 612 + ], + "type": "text", + "content": "Fig. 6: Qualitative Results of Ablations. We show some qualitative results of using latent manifold loss, positional encoding and varying the number of states. Using the manifold loss prevents us from overfitting to the extreme states, and using four states help us interpolate with a lot of clarity. Positional encoding helps add missing information for thin parts, but for large objects it doesn't help much." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": "shows the effect of positional encoding which adds detail when the latent embedding fails to capture it, as evidenced in the results of sunglasses. This can" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 161, + 91, + 447, + 102 + ], + "type": "text", + "content": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 135, + 467, + 174 + ], + "blocks": [ + { + "bbox": [ + 140, + 135, + 467, + 174 + ], + "lines": [ + { + "bbox": [ + 140, + 135, + 467, + 174 + ], + "spans": [ + { + "bbox": [ + 140, + 135, + 467, + 174 + ], + "type": "image", + "image_path": "70fd1f6b3209cee7e9da6a876b8f5bef52c1c2dbe1874431545a28352300f853.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 186, + 482, + 242 + ], + "lines": [ + { + "bbox": [ + 130, + 186, + 482, + 242 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 482, + 242 + ], + "type": "text", + "content": "Fig. 7: Failure cases. Our model fails at reconstructing the geometry correctly at camera positions where the two states we interpolate between have a change in the visible shape, like these examples where the microwave and oven being closed shows a deformation in the figure as compared to the open state, as the relationship between the motion is not easily captured in the latent due to the structure change." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 263, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 482, + 373 + ], + "type": "text", + "content": "happen for thin parts of the object that may not be well-captured across camera views. Notably, the inverse happens in the case of large objects, where the positional encoding adds extra noise to the reconstruction, resulting in oversmoothing. We also show how adding just two more intermediate states in LEIA makes a huge difference in reconstruction results for both single and multi-part objects, enabling us to scale and be flexible. Quantitative numbers for depth and occlusion regularization ablation are shown in Table 3. However, they didn't universally help in the cases where the latent embedding already learned the state representation well enough but reduced noise when it appeared." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 387, + 222, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 222, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 222, + 399 + ], + "type": "text", + "content": "5 Limitations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 411, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 518 + ], + "type": "text", + "content": "While LEIA works in achieving good quality reconstruction of intermediate states, our latent embeddings do not yet ensure accurate physical consistency in the motion of the intermediates, which is a tradeoff we chose while opting for not decoupling the object and learning the motion parameters for the moveable part. This allowed us to scale our approach to interpolate between multiple moving parts of the object, which would be more representative of the real world where common objects can be articulated in multiple ways. LEIA also struggles when there is severe occlusion, as referenced by Figure 7, as the occlusion occurs at gaps in the learnt embeddings between two states." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 534, + 220, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 220, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 220, + 547 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "In this work, we present LEIA, a method capable of successfully interpolating between two discrete articulation states of a deformable object with moving parts. LEIA handles multiple joints, including cases where moving parts are separated by static regions. It outperforms existing methods, especially when multiple parts move independently. The learned latent embeddings are view-invariant and separable, demonstrating LEIA's scalability and flexibility. We conducted comprehensive evaluations on synthetic and real data to investigate inherent challenges. Despite advancements, significant occlusion scenarios remain challenging. We hope future research builds upon our work." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 101 + ], + "type": "text", + "content": "A. Swaminathan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "content": "This work was partially supported by NSF CAREER Award (#2238769) to Abhinav Shrivastava, and IARPA via Department of Interior/Interior Business Center (DOI/IBC) contract number 140D0423C0076. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of NSF, IARPA, DOI/IBC, or the U.S. Government." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 255, + 197, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 197, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 197, + 268 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 280, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 280, + 481, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 280, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 138, + 280, + 481, + 324 + ], + "type": "text", + "content": "1. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5470-5479 (June 2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 324, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 324, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 324, + 481, + 357 + ], + "type": "text", + "content": "2. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 357, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 481, + 390 + ], + "type": "text", + "content": "3. Chiang, P.Z., Tsai, M.S., Tseng, H.Y., Lai, W.S., Chiu, W.C.: Stylizing 3d scene via implicit representation and hypernetwork. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1475-1484 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 390, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 481, + 434 + ], + "type": "text", + "content": "4. Chu, R., Liu, Z., Ye, X., Tan, X., Qi, X., Fu, C.W., Jia, J.: Command-driven articulated object understanding and manipulation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8813-8823 (June 2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 434, + 481, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 481, + 467 + ], + "type": "text", + "content": "5. Community, B.O.: Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam (2018), http://www.blender.org" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 468, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 481, + 489 + ], + "type": "text", + "content": "6. Frankle, J., Carbin, M.: The lottery ticket hypothesis: Finding sparse, trainable neural networks. arXiv preprint arXiv:1803.03635 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 490, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 481, + 511 + ], + "type": "text", + "content": "7. Goel, S., Kanazawa, A., Malik, J.: Shape and viewpoints without keypoints. In: ECCV (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 511, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 511, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 511, + 481, + 533 + ], + "type": "text", + "content": "8. Ha, D., Dai, A.M., Le, Q.V.: Hypernetworks. In: International Conference on Learning Representations (2017), https://openreview.net/forum?id=rkpAce1lx" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 533, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 481, + 566 + ], + "type": "text", + "content": "9. Heppert, N., Irshad, M.Z., Zakharov, S., Liu, K., Ambrus, R.A., Bohg, J., Valada, A., Kollar, T.: Carto: Category and joint agnostic reconstruction of articulated objects. In: Proc. IEEE Conf. Comput. Vis. Pattern Recog. pp. 21201-21210 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 566, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 566, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 133, + 566, + 481, + 599 + ], + "type": "text", + "content": "10. Jiang, Z., Hsu, C.C., Zhu, Y.: Ditto: Building digital twins of articulated objects from interaction. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 600, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 600, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 133, + 600, + 481, + 620 + ], + "type": "text", + "content": "11. Kanazawa, A., Tulsiani, S., Efros, A.A., Malik, J.: Learning category-specific mesh reconstruction from image collections. In: ECCV (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 621, + 481, + 665 + ], + "type": "text", + "content": "12. Kokkinos, F., Kokkinos, I.: To the point: Correspondence-driven monocular 3d category reconstruction. In: Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems (2021), https://openreview.net/forum?id=AWMU04iXQ08" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "type": "text", + "content": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "13. Kulkarni, N., Gupta, A., Fouhey, D.F., Tulsiani, S.: Articulation-aware canonical surface mapping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 452-461 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 193 + ], + "type": "text", + "content": "14. Lei, J., Daniilidis, K.: Cadex: Learning canonical deformation coordinate space for dynamic surface representation via neural homeomorphism. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6624-6634 (June 2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 480, + 236 + ], + "type": "text", + "content": "15. Li, T., Slavcheva, M., Zollhoefer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., Lv, Z.: Neural 3d video synthesis from multi-view video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 236, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 480, + 267 + ], + "type": "text", + "content": "16. Lin, H., Peng, S., Xu, Z., Yan, Y., Shuai, Q., Bao, H., Zhou, X.: Efficient neural radiance fields for interactive free-viewpoint video. In: SIGGRAPH Asia Conference Proceedings (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 268, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 268, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 268, + 480, + 300 + ], + "type": "text", + "content": "17. Liu, J., Mahdavi-Amiri, A., Savva, M.: PARIS: Part-level reconstruction and motion analysis for articulated objects. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 300, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 300, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 133, + 300, + 480, + 332 + ], + "type": "text", + "content": "18. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. ACM Trans. Graph. 38(4), 65:1-65:14 (Jul 2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 332, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 332, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 133, + 332, + 480, + 365 + ], + "type": "text", + "content": "19. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 365, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 407 + ], + "type": "text", + "content": "20. Mo, K., Zhu, S., Chang, A.X., Yi, L., Tripathi, S., Guibas, L.J., Su, H.: PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 407, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 449 + ], + "type": "text", + "content": "21. Müller, T., Evans, Å., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022). https://doi.org/10.1145/3528223.3530127, https://doi.org/10.1145/3528223.3530127" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 449, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 480, + 472 + ], + "type": "text", + "content": "22. Neverova, N., Novotny, D., Khalidov, V., Szafraniec, M., Labatut, P., Vedaldi, A.: Continuous surface embeddings (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 472, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 472, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 133, + 472, + 480, + 515 + ], + "type": "text", + "content": "23. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S.M., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 515, + 480, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 515, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 133, + 515, + 480, + 548 + ], + "type": "text", + "content": "24. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Occupancy flow: 4d reconstruction by learning particle dynamics. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 548, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 548, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 133, + 548, + 480, + 579 + ], + "type": "text", + "content": "25. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 579, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 579, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 133, + 579, + 480, + 601 + ], + "type": "text", + "content": "26. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. ICCV (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 133, + 601, + 480, + 633 + ], + "type": "text", + "content": "27. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "text", + "content": "28. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 100 + ], + "type": "text", + "content": "A. Swaminathan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "29. Qian, S., Jin, L., Rockwell, C., Chen, S., Fouhey, D.F.: Understanding 3d object articulation in internet videos. In: CVPR (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "text", + "content": "30. Ramanujan, V., Wortsman, M., Kembhavi, A., Farhadi, A., Rastegari, M.: What's hidden in a randomly weighted neural network? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11893-11902 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "31. Saffran, J.R., Aslin, R.N., Newport, E.L.: Statistical learning by 8-month-old infants. Science 274(5294), 1926-1928 (1996)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "text", + "content": "32. Schwarz, J.R., Tack, J., Teh, Y.W., Lee, J., Shin, J.: Modality-agnostic variational compression of implicit neural representations. arXiv preprint arXiv:2301.09479 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "33. Sen, B., Agarwal, A., Namboodiri, V.P., Jawahar, C.: Inr-v: A continuous representation space for video-based generative tasks. arXiv preprint arXiv:2210.16579 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "text", + "content": "34. Sen, B., Singh, G., Agarwal, A., Agaram, R., Krishna, K.M., Sridhar, S.: Hyp-nerf: Learning improved nef priors using a hypernetwork. arXiv preprint arXiv:2306.06093 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "35. Sitzmann, V., Rezchikov, S., Freeman, B., Tenenbaum, J., Durand, F.: Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems 34, 19313-19325 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "type": "text", + "content": "36. Sitzmann, V., Zollhöfer, M., Wetzstein, G.: Scene representation networks: Continuous 3d-structure-aware neural scene representations. Advances in Neural Information Processing Systems 32 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "text", + "content": "37. Skorokhodov, I., Ignatyev, S., Elhoseiny, M.: Adversarial generation of continuous images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10753-10764 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "38. Tan, J., Yang, G., Ramanan, D.: Distilling neural fields for real-time articulated shape reconstruction. In: CVPR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 413, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 456 + ], + "type": "text", + "content": "39. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In: IEEE International Conference on Computer Vision (ICCV). IEEE (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "type": "text", + "content": "40. Tseng, W.C., Liao, H.J., Lin, Y.C., Sun, M.: Cla-nerf: Category-level articulated neural radiance field. In: ICRA (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 544 + ], + "type": "text", + "content": "41. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017), https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "type": "text", + "content": "42. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 577, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 480, + 620 + ], + "type": "text", + "content": "43. Wei, F., Chabra, R., Ma, L., Lassner, C., Zollhoefer, M., Rusinkiewicz, S., Sweeney, C., Newcombe, R., Slavcheva, M.: Self-supervised neural articulated shape and appearance models. In: Proceedings IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "type": "text", + "content": "44. Wu, T., Zhong, F., Tagliasacchi, A., Cole, F., Oztireli, C.: D\\~nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 32653-32666. Curran Associates," + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 160, + 91, + 447, + 102 + ], + "type": "text", + "content": "LEIA: Latent View-invariant Embeddings for Implicit 3D Articulation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 434 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 147, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 117, + 480, + 138 + ], + "type": "text", + "content": "Inc. (2022), https://proceedings.neurips.cc/paper_files/paper/2022/file/d2cc447db9e56c13b993c11b45956281-Paper-Conference.pdf" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "text", + "content": "45. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time neural irradiance fields for free-viewpoint video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9421-9431 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 215 + ], + "type": "text", + "content": "46. Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., Yi, L., Chang, A.X., Guibas, L.J., Su, H.: SAPIEN: A simulated part-based interactive environment. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 216, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 481, + 247 + ], + "type": "text", + "content": "47. Yang, G., Sun, D., Jampani, V., Vlasic, D., Cole, F., Chang, H., Ramanan, D., Freeman, W.T., Liu, C.: Lasr: Learning articulated shape reconstruction from a monocular video. In: CVPR (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "48. Yang, G., Sun, D., Jampani, V., Vlasic, D., Cole, F., Liu, C., Ramanan, D.: Viser: Video-specific surface embeddings for articulated 3d shape reconstruction. In: NeurIPS (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "text", + "content": "49. Yang, G., Vo, M., Neverova, N., Ramanan, D., Vedaldi, A., Joo, H.: Banmo: Building animatable 3d neural models from many casual videos. In: CVPR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 324 + ], + "type": "text", + "content": "50. Yang, J., Pavone, M., Wang, Y.: Freenerf: Improving few-shot neural rendering with free frequency regularization (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "text", + "content": "51. Yao, C.H., Hung, W.C., Li, Y., Rubinstein, M., Yang, M.H., Jampani, V.: Lassie: Learning articulated shape from sparse image ensemble via 3d part discovery. In: NeurIPS (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "text", + "content": "52. Yao, C.H., Hung, W.C., Li, Y., Rubinstein, M., Yang, M.H., Jampani, V.: LASSIE: Learning Articulated Shape from Sparse Image Ensemble via 3d part discovery. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 434 + ], + "type": "text", + "content": "53. Yuan, W., Lv, Z., Schmidt, T., Lovegrove, S.: Star: Self-supervised tracking and reconstruction of rigid objects in motion with neural rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13144-13152 (2021)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 100 + ], + "type": "text", + "content": "A. Swaminathan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LEROjD_ Lidar Extended Radar-Only Object Detection/e0f1cd56-0b49-499c-b5fc-daef469190ce_content_list.json b/2024/LEROjD_ Lidar Extended Radar-Only Object Detection/e0f1cd56-0b49-499c-b5fc-daef469190ce_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2b2c2a480a38f04f41b3dcf28b4798f63a5069f3 --- /dev/null +++ b/2024/LEROjD_ Lidar Extended Radar-Only Object Detection/e0f1cd56-0b49-499c-b5fc-daef469190ce_content_list.json @@ -0,0 +1,1587 @@ +[ + { + "type": "text", + "text": "LEROjD: Lidar Extended Radar-Only Object Detection", + "text_level": 1, + "bbox": [ + 241, + 140, + 759, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Patrick Palmer1, Martin Krüger1, Stefan Schütte1, Richard Altendorfer2, Ganesh Adam2, and Torsten Bertram1", + "bbox": [ + 215, + 210, + 785, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Control Theory and Systems Engineering, TU Dortmund University {patrick.palmer, martin2.krueger, stefan.schuette, torsten.bertram}@tu-dortmund.de", + "bbox": [ + 225, + 253, + 772, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ ZF Group, {richard.altendorfer, ganeshadam}@zf.com", + "bbox": [ + 295, + 296, + 705, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Accurate 3D object detection is vital for automated driving. While lidar sensors are well suited for this task, they are expensive and have limitations in adverse weather conditions. $3 + 1\\mathrm{D}$ imaging radar sensors offer a cost-effective, robust alternative but face challenges due to their low resolution and high measurement noise. Existing $3 + 1\\mathrm{D}$ imaging radar datasets include radar and lidar data, enabling cross-modal model improvements. Although lidar should not be used during inference, it can aid the training of radar-only object detectors. We explore two strategies to transfer knowledge from the lidar to the radar domain and radar-only object detectors: 1. multi-stage training with sequential lidar point cloud thin-out, and 2. cross-modal knowledge distillation. In the multi-stage process, three thin-out methods are examined. Our results show significant performance gains of up to 4.2 percentage points in mean Average Precision with multi-stage training and up to 3.9 percentage points with knowledge distillation by initializing the student with the teacher's weights. The main benefit of these approaches is their applicability to other 3D object detection networks without altering their architecture, as we show by analyzing it on two different object detectors. Our code is available at https://github.com/rst-tu-dortmund/lerojd.", + "bbox": [ + 259, + 344, + 738, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D Object Detection $\\cdot$ 3+1D Imaging Radar $\\cdot$ Cross-Modal Object Detection", + "bbox": [ + 259, + 619, + 740, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 672, + 375, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Environment perception is the first module in each automated driving stack. Multiple sensor modalities, like cameras, lidars, and radars are utilized for this task. Radar sensors are of unique interest in perception due to their robustness against poor lighting, challenging weather conditions like rain or snow, and cost-effectiveness [3, 36, 47]. One exclusive advantage is the ability to measure the relative radial velocity of reflections directly due to the Doppler effect.", + "bbox": [ + 212, + 703, + 787, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While a precise localization of objects is possible with traditional radar sensors without elevation angle measurements [9], it is inherently limited to the horizontal plane. Additionally, it is hard to predict the objects' extent due to", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7d1e10ae07ead6a221ae33a784919070263ff3eb3ad9df5150d61440c4f24bd0.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 232, + 143, + 485, + 296 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1ab6dd17d4634553e796d6f4a621a3c1ef026a3fc8aec8b8d1ca16ce5311a81e.jpg", + "image_caption": [ + "(b)", + "Fig. 1: Architecture overview of (a) a knowledge distillation-based method and (b) a multi-stage training method (MSTM) for utilizing lidar data in the training of radar-only object detectors. The ground truth (GT) label is the same for both methods. The diagrams above the dotted line represent the training process, while the diagrams below the dotted line represent inference." + ], + "image_footnote": [], + "bbox": [ + 513, + 146, + 771, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the reflections' sparsity. The introduction of $3 + 1\\mathrm{D}$ high-resolution imaging radar sensors has recently mitigated these limitations, at least partially. In addition to measuring the elevation angle of reflections, the density of measurements is increased [14, 20]. Therefore, approaches that only utilize radar sensors for perception of the environment are of particular interest.", + "bbox": [ + 212, + 424, + 782, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite improvements in radar-based object detection, the performance still lags behind other sensor modalities like lidar [36]. One persistent major limitation of 2+1D classic radar and 3+1D imaging radar sensors is the relative sparsity of the point cloud, which limits the detection performance.", + "bbox": [ + 212, + 502, + 784, + 563 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lidar sensors, on the other hand, are well suited for object detection and are therefore frequently employed as a reference for evaluating the performance of different sensor modalities due to their ability to produce an accurate and dense understanding of the scene. Their effectiveness is particularly pronounced in detecting nearby traffic participants without occlusion [53].", + "bbox": [ + 212, + 565, + 784, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "All currently available datasets that contain $3 + 1\\mathrm{D}$ imaging radar data additionally accommodate lidar sensor data [4,8,32,35,36,40,63,65,69]. The lidar sensor data of these datasets is currently either utilized for labeling, combining multiple sensor modalities for accurate object detection, or comparing the performance of radar-only techniques to another sensor modality. While the majority of series production vehicles may not include lidar sensors due to cost and vehicle design constraints, they will still be available in the training process of learning-based methods. Extending radar-only methods with lidar sensor data in training has been shown to be a viable method for estimating point flow on the imaging radar point cloud [12]. These observations lead to the following research question: Can lidar sensor data be utilized in the training process of imaging radar-based 3D object detectors to improve the object detection performance on radar-only data during inference?", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "P. Palmer et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To use different sensor modalities in training, transfer learning and knowledge distillation (KD) principles can be utilized. While KD is commonly employed across sensor modalities such as camera images and lidar point clouds, its application between $3 + 1\\mathrm{D}$ imaging radar and lidar sensors remains unexplored. Since lidar and imaging radar share a structurally similar data representation as point clouds, an identical base network with different input modalities can be utilized to transfer knowledge between different input modalities. We investigate two approaches to transfer knowledge from lidar-based to radar-based object detectors: a KD-based and a multi-stage training approach. The main principles of the two methods are visualized in Figure 1.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The contributions of this work are summarized as follows:", + "bbox": [ + 238, + 297, + 658, + 311 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We investigate a combination of lidar and radar sensors in the training stage of object detectors to improve radar-only object detection at inference.", + "- We investigate three thin-out strategies for lidar point clouds to transfer knowledge from dense lidar to sparse lidar and radar-only object detectors.", + "- We propose a multi-stage training procedure to transfer knowledge from dense lidar to sparse lidar and, finally, to radar-only object detection.", + "- We modify and analyze several knowledge distillation-based approaches to transfer knowledge from lidar to radar-only object detectors." + ], + "bbox": [ + 225, + 320, + 782, + 440 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 462, + 387, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Imaging radar sensors commonly utilize the point cloud as a data representation format instead of the radar tensor, due to its higher computational efficiency. This format is similar to the one used for lidars, enabling the application of object detection methods developed for lidars to radars. Object detection on lidars can be split into two main categories. Point-based methods, like PointNet++ [39], downsample the original point cloud, encode it using a backbone, and finally apply a detection head. Voxel-based methods discretize the point cloud into a 3D grid and apply 3D convolutions to the grid before finally applying a detection head [6,10,71]. The main drawback of voxel-based methods is the high memory consumption and the loss of spatial information due to the discretization of the point cloud. To overcome the limitation of high memory consumption, [23,27] have proposed the PointPillars network. Pillars are a specific form of voxels that span over the full height of the scene and represent the point cloud in a 2D grid.", + "bbox": [ + 212, + 492, + 787, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Methods from the lidar domain have been shown to perform reasonably well on radar data [36,37]. However, they are limited due to the sparsity of the radar point cloud. SMURF [30] considers two representations of the radar point cloud to address sparsity. Using kernel density estimation, it utilizes pillarization and density features derived from a multi-dimensional Gaussian mixture distribution. RPFA-Net [57] is a PointPillars [23] based network, which introduces a selfattention mechanism to extract global context information from the radar point cloud. RadarMFNet [48] utilizes a multi-frame radar point cloud representation to address the sparsity of the radar point cloud in conjunction with an anchor-based detector and temporal pooling layers.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "LEROjD: Lidar Extended Radar-Only Object Detection", + "bbox": [ + 357, + 114, + 730, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One way of improving 3D object detection on radar data is the fusion with additional sensor modalities like camera [56, 68, 70], lidar [50] or camera and lidar [5, 13, 55] at the cost of introducing additional sensors at inference.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The concept of knowledge distillation was first introduced by [18]. It consists of two networks, which are labeled as teacher and student. The teacher is a large and complex model, while the student is less complex and more computationally efficient. The student is trained to mimic the teacher network's performance by utilizing the teacher's predictions and the ground truth labels. This has been utilized in the context of 3D object detection by [29,59,64] to construct computation time-efficient models that have similar performance as large models. [26] and [22] have extended upon this concept by utilizing KD to extract knowledge from a teacher trained on lidar data to a student who utilizes camera images.", + "bbox": [ + 212, + 191, + 785, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To the best of our knowledge, few studies have investigated the effect of KD considering radar data. A transfer of knowledge from an image-based teacher network to a radar-based student has been shown to improve the performance for the task of people counting [44]. For the task of 3D object detection, [22] has shown that a transfer learning-based approach from a lidar and image-based network to the radar domain results in improved object detection performance for classical $2 + 1\\mathrm{D}$ radar sensors. The main drawback of [22] is that in addition to 3D object labels, instance segmentation labels for the image domain are required. Additionally, [22] introduces multiple sub-networks to derive the KD losses, which makes the model more complex, while we aim for a simpler approach. HiddenGems [12] utilizes lidar point clouds to derive point flow information and train a network to predict the point flow on radar point clouds. See Beyond Seeing [11] utilizes lidar point cloud for feature hallucination on radar point clouds. Both methods require the lidar point cloud only in the training process but require major modifications to the network architecture.", + "bbox": [ + 212, + 330, + 787, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sampling of point clouds is a broadly explored topic to reduce the complexity of computing large point clouds. Most commonly, farthest point sampling (FPS) [39,42,60], voxelized FPS [41] or random sampling [71] are utilized in the context of 3D object detection. Random sampling has been shown to be beneficial when used for semantic segmentation of point clouds [19] and to outperform farthest point sampling at this task [25]. One limitation of random sampling is that while points at a close distance are kept, points at a far distance are more likely to be removed. Therefore, [7] proposes a polar cylinder balanced random sampling to keep a more balanced distribution of points across the range.", + "bbox": [ + 212, + 558, + 787, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 724, + 330, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Two methods of transferring knowledge from lidar-based to radar-based object detectors are investigated: Knowledge Distillation (KD) (Section 3.1), which is modified for the task of transfer learning and our proposed multi-stage training procedure with sequential point cloud thin-out (Section 3.2). Additionally, the utilized thin-out strategies for lidar point cloud (Section 3.3) are described.", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "P. Palmer et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Knowledge Distillation", + "text_level": 1, + "bbox": [ + 215, + 146, + 450, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "KD is commonly used for two tasks. First, designing computationally efficient models by transferring knowledge from a larger teacher network to a smaller student network [18, 49, 51]. Second, to transfer knowledge across sensor modalities by utilizing different model architectures for the teacher and student [1, 16, 67]. Lidar and radar-based point cloud object detection can utilize the same model structure but with different input modalities. This enables the utilization of KD methods first described for designing computationally efficient models for cross-modal knowledge transfer from lidar- to radar-based object detectors. In this case, the teacher is trained on the full lidar point cloud, while the student is trained on the radar point cloud. Three different loss terms, as described by [59], are employed:", + "bbox": [ + 212, + 171, + 787, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Logit-KD is the first, classic type of distilling knowledge described by [18]. For 3D object detection, the logit-KD loss $\\mathcal{L}_{\\mathrm{logit}}$ is split into a classification $\\mathcal{L}_{\\mathrm{l -cls}}$ and regression loss $\\mathcal{L}_{\\mathrm{l - reg}}$ . These losses are calculated by comparing the student's and teacher's predictions utilizing the 3d detectors' regression loss and bi-linear interpolation between student and teacher output classes.", + "bbox": [ + 212, + 337, + 787, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature-KD is widely utilized in 2D object detection [28, 51]. It utilizes a loss term that forces the student network to mimic the teacher's intermediate feature map (feat). A feature mimicking the last layer of the bird's eye view feature encoder, similar to [59], is utilized in this work.", + "bbox": [ + 212, + 412, + 787, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Label-KD is a recent distillation approach that leans on the concept of the Logit KD but simplifies and generalizes it. It is first described by [33]. The teacher predictions are filtered by their scores using a score threshold, and an adapted ground truth set is constructed by combining the filtered predictions and the ground truth set. This adapted set is utilized in student training. The loss is split into a classification $\\mathcal{L}_{\\mathrm{cls}}$ and regression loss $\\mathcal{L}_{\\mathrm{reg}}$ . It replaces $\\mathcal{L}_{\\mathrm{label}}$ usually calculated on the ground truth set.", + "bbox": [ + 212, + 474, + 787, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The three KD loss terms are combined into a joint loss weighted with $\\lambda$", + "bbox": [ + 238, + 579, + 759, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {j o i n t}} = \\underbrace {\\lambda_ {\\mathrm {l - r e g}} \\mathcal {L} _ {\\mathrm {l - r e g}} + \\lambda_ {\\mathrm {l - c l s}} \\mathcal {L} _ {\\mathrm {l - c l s .}}} _ {\\mathcal {L} _ {\\text {l o g i t}}} + \\underbrace {\\lambda_ {\\text {f e a t}} \\mathcal {L} _ {\\text {f e a t}}} _ {\\mathcal {L} _ {\\text {f e a t u r e}}} + \\underbrace {\\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}} + \\lambda_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}}} _ {\\mathcal {L} _ {\\text {l a b e l}} ^ {*}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 269, + 604, + 785, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Multi-Stage Lidar Thin-Out Training Procedure", + "text_level": 1, + "bbox": [ + 214, + 664, + 660, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Utilizing a different input data modality for either pre-training a network on a large dataset or utilizing simulated data in the training process of a point cloud-based network has been shown to improve the object detection performance [54]. The multi-stage training method (MSTM) proposed in this work extends upon this by utilizing a Curriculum learning [2] based training procedure by which the network is trained on iteratively sparsified lidar point clouds, similar to [52], and fine-tuned on the radar point cloud. Figure 2 visualizes our multi-stage training procedure. The network is first trained on the full lidar point cloud until convergence. In the following steps, the lidar point cloud is thinned-out iteratively by a factor of 2 and utilized for training a network whose weights are", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "LEROjD: Lidar Extended Radar-Only Object Detection", + "bbox": [ + 356, + 114, + 730, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4c1e883322e8a05bc2812e572cfa34f1d9933ec98787b0de36e580ef8d1e9d00.jpg", + "image_caption": [ + "Fig. 2: MSTM pipeline. A 3D object detection network is iteratively trained on increasingly sparse lidar data. Three different thin-out strategies are utilized for sparsification. The lidar point cloud is mixed with the radar point cloud in the second to last step. In the last step, the network is only trained on radar data. At inference, only the orange-shaded part is executed. Small points represent lidar points, large points radar points. The color of points corresponds to the distance from the ego-vehicle. The camera image is not used as an input but only aids the visualization." + ], + "image_footnote": [], + "bbox": [ + 232, + 143, + 774, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "initialized using the previously trained model. This forces the network to learn features for a good object detection performance on increasingly sparser point clouds. In the second to last step, the lidar point cloud is mixed with the radar points so that the network can translate from the lidar to the radar domain. In the last step, the model is trained only on radar points. A training without multiple stages is called single-stage training method (SSTM) in this work.", + "bbox": [ + 212, + 594, + 787, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to the training where only lidar points are utilized in the first training stages, we investigate the utilization of the radar point cloud in conjunction with the lidar point cloud in all stages. The thin-out of the lidar points remains the same and is mixed with the radar point cloud in each step. This conditions the model on radar from the first step in order to prioritize features in the lidar point cloud that relate to a good object detection on radar-only data.", + "bbox": [ + 212, + 686, + 787, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For mixing lidar and radar point clouds, the voxel or pillar feature encoder is modified to prioritize the radar point cloud in the random sampling process as done by [34]. Otherwise, the vastly higher number of lidar points, even when thinned-out, might lead to the complete exclusion of radar points.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "P. Palmer et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Lidar Thin-Out Strategies", + "text_level": 1, + "bbox": [ + 215, + 146, + 477, + 161 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Three different methods for sub-sampling the lidar point cloud are investigated. Figure 2 shows examples for each thin-out stage.", + "bbox": [ + 212, + 191, + 782, + 222 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Random sampling is the simplest method of sparsifying a point cloud. It neglects the structure and inherent limitations of the point cloud representation, especially for objects far away or with a high degree of occlusion. Neglecting the structure can lead to the complete loss of information for objects represented only by few points.", + "bbox": [ + 212, + 224, + 784, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "K-nearest neighbor sampling approximates the reflection density distribution of the radar point cloud with the lidar point cloud by only keeping lidar reflections close to radar reflections. This algorithm is described in Algorithm 1. Objects not detected by the radar sensor are, therefore, also not represented by the k-nearest neighbor thinned-out lidar point cloud.", + "bbox": [ + 212, + 305, + 784, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Voxel-based sampling aims to reduce the number of points in each area of the point cloud while keeping the general distribution of the point cloud. This is motivated by the fact that radar sensors do not suffer as much from loss of resolution with distance as lidar sensors. The approach is described in Algorithm 2, and executed iteratively for a sequence of sparsification steps.", + "bbox": [ + 212, + 383, + 785, + 460 + ], + "page_idx": 6 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: K-nearest neighbor sampling" + ], + "code_body": "Input : Lidar point cloud $L\\in \\mathbb{R}^{N\\times 3}$ , radar point cloud $R\\in \\mathbb{R}^{M\\times 3}$ Output: Sub-sampled lidar point cloud $L_{s}\\in \\mathbb{R}^{K\\times 3}$ \n1 Calculate the euclidean distance $c_{i}$ between each lidar point $l_{i}\\in L$ and its nearest radar point from $R$ . \n2 Select a share of $K$ lidar points with the smallest distances c; \n3 return Share of Points $L_{s}\\in \\mathbb{R}^{K\\times 3}$ \nAlgorithm 2: Voxel-based sampling. \nInput : Lidar point cloud $L\\in \\mathbb{R}^{N\\times 3}$ Output: Sub-sampled lidar point cloud $L_{s}\\in \\mathbb{R}^{K\\times 3}$ \n1 Initialize $L_{s}$ with $L$ . \n2 Voxelize the lidar point cloud into v voxels; \n3 Calculate the number of points in each voxel $P_{v}$ . \n4 Calculate the minimum number of points per voxel $p_{\\mathrm{min}}$ , so that at least $0.75N$ points are in voxels with more than $p_{\\mathrm{min}}$ points; \n5 From each voxel with more than $p_{\\mathrm{min}}$ points, choose $p_{\\mathrm{min}}$ random points to keep and add remaining points to $L_{p}$ . \n6 Randomly select O.5 N points from $L_{p}$ and remove them from $L_{s}$ . \n7 return Share of Points $L_{s}\\in \\mathbb{R}^{K\\times 3}$", + "bbox": [ + 215, + 484, + 785, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Another common sparsification method utilized for imitating low-resolution lidar sensors is layer-based sampling [52, 61]. This approach is not investigated because radar sensors do not capture the environment on a layer basis.", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "LEROjD: Lidar Extended Radar-Only Object Detection", + "bbox": [ + 356, + 113, + 730, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experimental Evaluation", + "text_level": 1, + "bbox": [ + 215, + 143, + 493, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 175, + 426, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset: All experiments are conducted utilizing the View-of-Delft (VoD) dataset [36]. It contains synchronized data of multiple sensor modalities. The 64-layer lidar sensor and the imaging radar are utilized in this work. A point cloud accumulated over 5 frames [36], which has been shown to improve object detection performance compared to no accumulation, is used for radar data [36, 38]. We detected a duplication of identical points in the lidar point cloud, which could adversely affect all sampling methods; thus, we eliminated the duplicated points from the point cloud. Although the VoD dataset is among the best currently available datasets for imaging radar-based object detection, it is limited by its size compared to other automotive datasets without imaging-radar, like [31, 45]. Given the absence of publicly available labels and limited online evaluation for the test dataset, we repurpose the validation dataset as a test set. Consequently, we partition the original training set into a new training set $(80\\%)$ and a dedicated validation set $(20\\%)$ to ensure robust model training.", + "bbox": [ + 212, + 198, + 792, + 410 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation Metrics: The primary performance metric utilized to compare the results is the mean average precision (mAP), as used by [15] [37]. Similar to the evaluation of the Waymo data set [46] in [62], we split the results into two distance bins: short-range (SR): $0 - 30\\mathrm{m}$ and mid-range (MR): $30 - 50\\mathrm{m}$ . All experiments are conducted utilizing three different random network initializations that are averaged.", + "bbox": [ + 212, + 410, + 789, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training: Most experiments use the PointPillars model [23] as an object detector with the same configuration as utilized by [36]. For imaging radar data, PointPillars has been shown to perform among the best out of multiple state-of-the-art 3D object detection methods while still performing adequately on lidar data [37]. Furthermore PointPillars is a relevant baseline for radar-specific object detection methods [57,58]. To show that the proposed MSTM and KD apply to various object detectors, the most promising approaches from the evaluations on PointPillars are evaluated on DSVT-P [17], as an example for a transformer-based model. All SSTM trainings are conducted with an early stopping policy for a maximum of 125 epochs. For the MSTM the initial training on the full lidar point cloud is conducted for 125 epochs, while each refinement step is trained for 30 epochs. All trainings utilize the Adam optimizer [21] and an adapted learning rate scheduler that reaches its maximum learning rate earlier and has a faster descent than the scheduler described by [43]. This improves the object detection performance on radar data.", + "bbox": [ + 212, + 501, + 789, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Notation: To distinguish between methods, the following notation is used:", + "bbox": [ + 238, + 727, + 782, + 742 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {\\mathcal {L S} / \\mathcal {T O}} ^ {\\mathcal {T M}} \\rightarrow \\mathcal {T} ^ {\\mathcal {K D}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 767, + 573, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The notation is split into two parts. The part left of the arrow represents the data set utilized for pre-training, while the right part represents the data and training method utilized for the last (fine-tuning) training stage of the model.", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "P. Palmer et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1a636084813b97cf9700c424ac466e9931042e32798443ffb23b1e9139e87326.jpg", + "table_caption": [ + "Table 1: Examples of training steps and evaluation data corresponding to the notation, emplying voxel-based sub-sampling. Voxel sampling is shortened to v for legibility." + ], + "table_footnote": [], + "table_body": "
NotationTrainingEvaluation
L1-1/8/vL1/v → L1/2/v → L1/4/v → L1/8/vL1/8/v
RLMSTM1-1/8/vRL1/v → RL1/2/v → RL1/4/v → RL1/8/vRL1/8/v
L1MSTM1-1/16/v → RL1/v → L1/2/v → L1/4/v → L1/8/v → RL1/16/v → RR
RLMSTM1-1/16/v → RRL1/v → RL1/2/v → RL1/4/v → RL1/8/v → RL1/16/v → RR
RLSSTM1/4/v → RRL1/4/v → RR
", + "bbox": [ + 225, + 184, + 772, + 303 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This part is omitted if it is the same as the training dataset. The training data $\\mathcal{T}$ can either be lidar (L), radar (R), or mixed radar + lidar data (RL). The training method $\\mathcal{T}\\mathcal{M}$ can denote either MSTM or SSTM. The lidar share $\\mathcal{L}\\mathcal{S}$ is the fraction of the original lidar point cloud that is utilized. Radar-only training always uses the full radar point cloud, therefore $\\mathcal{L}\\mathcal{S}$ is omitted. For the MSTM, this is represented by a range of fractions that are iterated in training. The thin-out method $\\mathcal{T}\\mathcal{O}$ can either be random (rand), k-nearest neighbor (knn) or voxel-based (vox). $\\mathcal{KD}$ represents the KD method; this can either be label (lab), logit (log), feature (feat), or a combination of those (joint). $\\mathcal{KD}$ is omitted if just an initialization and fine-tuning is utilized. Examples for the training corresponding to specific notations are listed in Table 1.", + "bbox": [ + 212, + 328, + 787, + 494 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Evaluation of MSTM on Lidar-only and Mixed Radar + Lidar", + "text_level": 1, + "bbox": [ + 215, + 513, + 782, + 530 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To evaluate the applicability of our proposed MSTM to the thinned-out lidar point cloud, we evaluate the training without the last two steps involving radar data. Training is conducted for just the lidar point cloud and the mixed radar + lidar point cloud in all stages. The multi-stage trained network is evaluated after each thin-out stage on the thinned-out lidar (or mixed radar + lidar) point cloud. Thin-out stages up to $1/256$ of the original lidar point cloud are considered due to the lidar point cloud containing fewer points than the radar point cloud at $1/256$ of the original lidar point cloud. The MSTM is compared to the SSTM, which is trained only on the thinned-out point cloud. The results are shown in Fig. 3, complete quantitative results are given in the supplementary.", + "bbox": [ + 212, + 537, + 787, + 688 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "All thin-out strategies result in an increasing degradation of the detection performance. This is most pronounced when considering a random thin-out, which drops approximately linearly with each thin-out step. K-nearest neighbor and voxel-based sampling consistently perform better than random sampling due to keeping a higher point density in areas around objects, which supports object detection. For k-nearest neighbor sampling, the performance only drops by 0.6 percentage points between $\\mathrm{L}_{1 / \\mathrm{knn}}^{\\mathrm{SSTM}}$ and $\\mathrm{L}_{1 / 2 / \\mathrm{knn}}^{\\mathrm{SSTM}}$ due to mostly ground points being removed in the first thin-out stage. For voxel-based sampling, after a sharp performance drop in the first thin-out stage, only a relatively slight performance drop is observable until the $^{1/16}$ thin-out stage. In the first stages,", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "LEROjD: Lidar Extended Radar-Only Object Detection", + "bbox": [ + 356, + 114, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9ea428973687d550449b45cf72c04b8e0f264792b40cbc1352e6158db55bb224.jpg", + "image_caption": [ + "Fig. 3: Comparison between SSTM and MSTM for different lidar sampling strategies." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 781, + 299 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "the performance drop can mainly be attributed to the pedestrian class, which drops by 12.8 percentage points in the first thin-out stage. In comparison, the detection performance of the car and cyclist classes only decreased by 0.7 and 3 percentage points, respectively. This can be explained by the voxel size of $1\\mathrm{m}$ in each dimension utilized in the voxel-based sampling. A single pedestrian only occupies a small number of these voxels. When the point cloud gets thinned out, the entire object's structure gets lost, making detection difficult. In contrast, the object's structure and physical appearance can be adequately represented when it occupies more voxels, as observed in the car and cyclist classes. In the final thin-out stage of $\\mathrm{L}^{\\mathrm{SSTM}}$ and $\\mathrm{L}^{\\mathrm{MSTM}}$ , voxel-based sampling performs worse than k-nearest neighbor sampling because many voxels only consist of ground points or points of surrounding background objects.", + "bbox": [ + 212, + 354, + 787, + 536 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The multi-stage training does not result in a useful knowledge transfer and, therefore, significant performance benefit for all considered thin-out strategies.", + "bbox": [ + 214, + 537, + 785, + 566 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A contrary behavior is observed on the combined radar + lidar point cloud. The MSTM consistently outperforms the SSTM when using k-nearest neighbor or voxel-based sampling. Knowledge from the dense point cloud can be transferred to the thin radar + lidar point cloud. Additionally, performance is consistently higher than training just on lidar, especially for voxel-based subsampling, which performs best at lower thin-out stages. At small thin-out stages, the voxel-based sub-sampling can still represent the whole object space and give meaningful environmental information. At the same time, the radar point cloud is sufficiently dense for object detection.", + "bbox": [ + 212, + 566, + 787, + 703 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Evaluation of MSTM with Last Radar-Only Step", + "text_level": 1, + "bbox": [ + 215, + 724, + 666, + 739 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "This chapter analyses the performance of the MSTM when applied to radar data, as described in Section 3.2. The MSTM is evaluated for two different procedures. Utilizing only the lidar point cloud and utilizing the mixed radar + lidar point cloud in the first stages. The results of MSTM are shown in Table 2. The lidar thin-out stages up to $1/16$ of the original lidar point cloud are considered due to the performance of $\\mathrm{L}_{1-1/32}^{\\mathrm{MSTM}}$ dropping below $\\mathrm{R}^{\\mathrm{SSTM}}$ .", + "bbox": [ + 212, + 750, + 787, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "P. Palmer et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/52ce632e4cc3e482cfd399a01e8e480481225ff362d67794e5de607c9ee90335.jpg", + "table_caption": [ + "Table 2: 3D object detection results on the VoD radar dataset trained using MSTM. All MSTM methods are trained on radar-only data in the last step. The best and second best results are marked in bold and underlined, respectively." + ], + "table_footnote": [], + "table_body": "
Training methodmAPCarPedestrianCyclist
SRMRSRMRSRMRSRMR
RSSTM (baseline)36.711.945.218.117.17.447.710.2
L1-1/16 / rand → R36.416.144.920.017.59.846.818.9
L1-1/16 / knn → R35.714.045.623.216.77.144.811.7
L1-1/16 / vox → R34.114.347.023.015.39.140.010.7
RLMSTM 1-1/16 / rand → R35.614.944.120.217.87.244.917.3
RLMSTM 1-1/16 / knn → R38.214.745.523.918.88.250.311.9
RLMSTM 1-1/16 / vox → R39.715.445.922.518.49.754.713.9
", + "bbox": [ + 233, + 198, + 764, + 363 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "When considering the pre-training using just lidar data, it is observable that the overall performance drops in the SR bin. In contrast, the MR bin is improved for all thin-out methods. The best performance is achieved by utilizing the random thin-out strategy. It performs especially well for pedestrians and cyclists in the mid-range, due to the comparably small size of pedestrians and cyclists resulting in a worse radar reflection characteristic. In contrast these objects are detected well by the lidar sensor. Knowledge about the representation of objects can be transferred from the lidar point cloud to the radar point cloud. K-nearest neighbor and voxel-based thin-out do not improve the performance in the SR bin for the pedestrian and cyclist class but perform better in the short and mid-range for the car class which is explained by the large size of cars, resulting in a better representation of cars in the thin point cloud.", + "bbox": [ + 217, + 388, + 785, + 570 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "When considering the pre-training with the mixed radar + lidar point cloud, a contrary observation is made compared to only using the lidar in pre-training. The overall performance for the random thinned-out lidar point cloud is worse with the mixed point cloud than when only considering the lidar point cloud. One consistent aspect is the good performance of the cyclists in the mid-range, which still surpasses all other methods, excluding $\\mathrm{L}_{1 - 1/16}^{\\mathrm{MSTM}}$ . However, the k-nearest neighbor and voxel-based thin-out strategies perform better with the mixed point cloud. The voxel-based thin-out strategy achieves the best performance. It performs best on objects in SR, mainly due to its outstanding performance in detecting cyclists, but is limited in the mid-range, getting surpassed by the MSTM with random thin-out. The MSTM with voxel-based thin-out can increase the object detection performance on the radar point cloud by 3 percentage points in the SR and 3.5 percentage points in the MR.", + "bbox": [ + 217, + 571, + 785, + 767 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Evaluation of Cross-Modality KD", + "text_level": 1, + "bbox": [ + 215, + 787, + 539, + 801 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The configuration of the teacher's training data is of particular interest, as the teacher's performance directly influences the student's. The simplest choice is", + "bbox": [ + 215, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "LEROjD: Lidar Extended Radar-Only Object Detection", + "bbox": [ + 357, + 114, + 730, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/bdca4fec737c4512b93ef9a3955cc2d392ca4937e682c572d362de297a4d2f83.jpg", + "table_caption": [ + "Table 3: 3D object detection results of the teacher with different training sets. Each teacher is evaluated on the test set of the same data configuration as the training set. The best and second best results are marked in bold and underlined, respectively." + ], + "table_footnote": [], + "table_body": "
DatamAPCarPedestrianCyclist
SRMRSRMRSRMRSRMR
L1SSTM56.530.060.742.440.119.568.528.1
RLSSTM61.645.161.946.544.434.078.554.8
RLSSTM1/4 / rand56.931.157.232.437.221.376.439.5
RLSSTM1/4 / knn58.936.954.431.943.030.979.547.9
RLSSTM1/4 / Vox56.139.360.141.632.828.875.547.5
", + "bbox": [ + 250, + 198, + 746, + 321 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "to train the teacher solely on the lidar point cloud. Section 4.2 and Section 4.3 show that mixing the radar and lidar point clouds can benefit radar-only object detection. Therefore, teachers trained on a mixed point clouds, in addition to the ones solely trained on lidar, are compared. Mixed point clouds containing $1/4$ of the original lidar point cloud are considered, as a closer representation of the teachers to the student's data can lead to a better performance. Thin out of $1/4$ is chosen as the radar + lidar detection performance with radar and $1/8$ of the lidar points is worse than the SSTM on only lidar data. Table 3 shows the performance of all utilized teachers.", + "bbox": [ + 212, + 351, + 787, + 487 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 3 shows the results of the teacher network for different training data configurations. It can be observed that, as expected, the mixed radar and lidar point cloud performs the best, only being surpassed by $\\mathrm{RL}_{1/4}^{\\mathrm{SSTM}}$ on the cyclist class in short range. No model can be considered the overall second best among the other training sets. The performance varies between vehicle classes.", + "bbox": [ + 212, + 487, + 787, + 563 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The KD is evaluated individually for each KD method and teacher training set. A joint KD is also considered comprised of all three KD losses. All student networks are initialized (Init) with the weights of the teacher network, as it has been shown to improve the student's performance [59]. Additionally, for the transfer learning between datasets, the pre-training utilizing MSTM in Section 4.3 has been shown to improve the performance on the radar dataset.", + "bbox": [ + 212, + 564, + 787, + 654 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The results utilizing the KD are shown in Table 4. Just the initialization of the student network with the teacher's weights already results in a performance benefit in the SR with $\\mathrm{RL}_{1}^{\\mathrm{SSTM}}$ and $\\mathrm{RL}_{1/4}^{\\mathrm{SSTM}}$ as teachers and in the MR with all teachers. This is overall only surpassed by the $\\mathrm{RL}_{1-1/16}^{\\mathrm{MSTM}} / \\mathrm{vox} \\to \\mathrm{R}$ showing that the MSTM is substantially better than a simple initialization.", + "bbox": [ + 212, + 654, + 787, + 734 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For label-KD, the best-performing models are the ones where the teacher performs the best. Specifically $\\mathrm{L}_1^{\\mathrm{SSTM}}\\rightarrow \\mathrm{R}^{\\mathrm{lab}}$ , which is only surpassed by $\\mathrm{RL}_{1/4/knn}^{\\mathrm{SSTM}}\\rightarrow \\mathrm{R}^{\\mathrm{lab}}$ in the MR. For worse-performing teachers, using label-KD loss does not result in a performance benefit due to it replacing the ground truth label loss. Feature-KD requires a teacher who learns features from radar data. This is observable in the SR performance of models, where the teacher is trained on mixed radar + lidar data. Logit-KD works well on teacher datasets", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "P. Palmer et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/80ac5cdede6324392a3a815a49aacf10558dc0280fd26e24e3eabc29a3c68c92.jpg", + "table_caption": [ + "Table 4: 3D object detection results on the VoD radar dataset utilizing the different KDs and teachers trained on different data configurations. Only the mAP over all classes is specified. For each distillation method and range bin, the best and second best results are marked in bold and underlined, respectively. The best result for each teacher are marked in cyan for the SR bin and magenta for the MR bin." + ], + "table_footnote": [], + "table_body": "
Teacher dataInit only (→R)Logit-KD (→Rlog)Feature-KD (→Rfeat)Label-KD (→Rlab)Joint-KD (→RJoint)
SRMRSRMRSRMRSRMRSRMR
R1SSTM36.711.9--------
L1SSTM34.913.532.812.634.813.636.613.736.013.5
RLSSTM39.014.839.213.539.113.838.813.738.912.7
RLSSTM1/4 / rand39.014.238.911.639.411.834.012.236.412.3
RLSSTM1/4 / knn34.814.637.214.736.613.033.810.932.912.9
RLSSTM1/4 / Vox35.215.335.613.337.115.835.314.332.99.7
", + "bbox": [ + 215, + 226, + 787, + 385 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "which closely resemble the radar point cloud and perform well on the teacher set. This results in good performance of $\\mathrm{RL}_1^{\\mathrm{SSTM}}\\rightarrow \\mathrm{R}^{\\log}$ . Besides the KD method the thin-out method utilized in the teacher's training influences the student's performance. Random sampling results in good performance in the SR, while voxel-based sampling results in good MR performance. This is somewhat contrary to what has been observed in the MSTM, where random sampling performs the best in the MR and can be explained by the different thin-out stages utilized in the MSTM and KD. Further qualitative results, as well as detailed quantitative results are given in the supplementary. Overall, initializing the student with the teacher's parameters yields good performance, with further enhancements primarily achievable through feature-KD with a teacher trained on mixed radar + lidar data.", + "bbox": [ + 212, + 412, + 787, + 593 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Evaluation on DSVT as a Transformer-Based Object Detector", + "text_level": 1, + "bbox": [ + 214, + 616, + 774, + 633 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 5 shows the results of the MSTM and KD on DSVT-P [17] for selected methods. Similar effects, as observed for PointPillars, are seen on DSVT-P. Initializing the student with the teacher's weights and the MSTM led to a performance benefit. However, contrary to PointPillars, the KD approach does not contribute to any improvements for DSVT-P.", + "bbox": [ + 212, + 641, + 787, + 717 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.6 Limitations", + "text_level": 1, + "bbox": [ + 214, + 739, + 357, + 753 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The best-performing methods in this work apply only to detectors that share an architecture with the target model, as a direct transfer of weights is performed for the best performance. Different models may require different levels and steps in the thinning process. These choices are additional tuning parameters that must be selected appropriately to maximize the benefit of transferring knowledge.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "LEROjD: Lidar Extended Radar-Only Object Detection", + "bbox": [ + 356, + 114, + 730, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1579f1bc2c008b97721a3f3a172b912531f2452f3254eafd5c1678eb649cdea0.jpg", + "table_caption": [ + "Table 5: Radar-only detection performance utilizing DSVT-P as an object detector. The best and second best results are marked in bold and underlined, respectively." + ], + "table_footnote": [], + "table_body": "
Training methodmAPCarPedestrianCyclist
SRMRSRMRSRMRSRMR
RSSTM38.313.042.520.221.912.350.56.6
RLMSTM1-1/16 / vox→R41.515.647.823.624.511.052.312.4
RLSSTMM1/4 / vox→R38.813.244.918.823.311.248.29.7
RLSSTMM1/4 / vox→Rfeat34.711.344.08.820.52.839.712.7
", + "bbox": [ + 233, + 184, + 763, + 306 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 330, + 357, + 347 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we investigated two methods to transfer knowledge from lidar-based object detectors to radar-only object detection. First, MSTM with sequential sub-sampling of the lidar point cloud, and second, a KD-based approach. For the MSTM, we have investigated three thin-out strategies for the lidar point cloud. These thin-out strategies are also analyzed for the training of the KD teacher network. Both methods can substantially improve the object detection performance of a radar-only object detector. The MSTM with voxel-based thinout performs the best overall and can improve detection performance by up to 3.5 percentage points. For the KD methods, it is shown that initializing the student with the teacher's parameters, especially a teacher trained on mixed lidar and radar data, can improve the object detection performance on radar-only data, with further enhancement primarily achieved by utilizing feature-KD.", + "bbox": [ + 212, + 362, + 785, + 542 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In future work, the applicability to further 3D object detection networks and the behavior with more advanced knowledge distillation like the ones utilized by [22] could be investigated. Due to different effects being noticed with the MSTM and the KD methods, combining both methods could be investigated by utilizing the MSTM as a teacher. To overcome the limitations of choosing a strict thin-out strategy, a learnable point cloud thin-out method [24,66] can be used.", + "bbox": [ + 212, + 545, + 785, + 648 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 671, + 323, + 686 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Afouras, T., Chung, J.S., Zisserman, A.: Asr is all you need: Cross-modal distillation for lip reading. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 2143-2147. IEEE (2020)", + "2. Bengio, Y., Louradour, J., Collobert, R., Weston, J.: Curriculum learning. In: Proceedings of the 26th annual international conference on machine learning. pp. 41-48 (2009)", + "3. Briske, S., Ruf, F., Höhne, F.: Recent evolution of automotive imaging radar and its information content. IET Radar, Sonar & Navigation 12(10), 1078-1081 (2018)", + "4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous" + ], + "bbox": [ + 222, + 700, + 784, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "P. Palmer et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)", + "5. Chen, X., Zhang, T., Wang, Y., Wang, Y., Zhao, H.: Futr3d: A unified sensor fusion framework for 3d detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 172-181 (2023)", + "6. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21674-21683 (June 2023)", + "7. Cheng, H., Han, X., Jiang, H., He, D., Xiao, G.: Pcb-random sampling for lidar semantic segmentation in autonomous driving scene. arXiv preprint arXiv:2209.13797 (2022)", + "8. Choi, M., Yang, S., Han, S., Lee, Y., Lee, M., Choi, K.H., Kim, K.S.: Msc-rad4r: Ros-based automotive dataset with 4d radar. IEEE Robotics and Automation Letters 8(11), 7194-7201 (2023)", + "9. Danzer, A., Griebel, T., Bach, M., Dietmayer, K.: 2d car detection in radar data with pointnets. In: 2019 IEEE Intelligent Transportation Systems Conference (ITSC). pp. 61-66 (2019)", + "10. Deng, J., Shi, S., Li, P., Zhou, W., Zhang, Y., Li, H.: Voxel r-cnn: Towards high performance voxel-based 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1201-1209 (2021)", + "11. Deng, J., Chan, G., Zhong, H., Lu, C.X.: See beyond seeing: Robust 3d object detection from point clouds via cross-modal feature augmentation. arXiv preprint arXiv:2309.17336 (2023)", + "12. Ding, F., Palffy, A., Gavrila, D.M., Lu, C.X.: Hidden gems: 4d radar scene flow learning using cross-modal supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9340-9349 (June 2023)", + "13. Drews, F., Feng, D., Faion, F., Rosenbaum, L., Ulrich, M., Glaser, C.: Deepfusion: A robust and modular 3d object detector for lidars, cameras and radars. In: 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 560-567. IEEE (2022)", + "14. Engels, F., Heidenreich, P., Winternantel, M., Stäcker, L., Al Kadi, M., Zoubir, A.M.: Automotive radar signal processing: Research directions and practical challenges. IEEE Journal of Selected Topics in Signal Processing 15(4), 865-878 (2021)", + "15. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012)", + "16. Gupta, S., Hoffman, J., Malik, J.: Cross modal distillation for supervision transfer. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2827-2836 (2016)", + "17. Haiyang Wang, Chen Shi, S.S.M.L.S.W.D.H.B.S., Wang, L.: Dsvt: Dynamic sparse voxel transformer with rotated sets. In: CVPR (2023)", + "18. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. Deep Learning Workshop at NIPS (2014)", + "19. Hu, Q., Yang, B., Xie, L., Rosa, S., Guo, Y., Wang, Z., Trigoni, N., Markham, A.: Randla-net: Efficient semantic segmentation of large-scale point clouds. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11108-11117 (2020)", + "20. Jiang, M., Xu, G., Pei, H., Feng, Z., Ma, S., Zhang, H., Hong, W.: 4d high-resolution imagery of point clouds for automotive mmwave radar. IEEE Transactions on Intelligent Transportation Systems pp. 1-15 (2023)" + ], + "bbox": [ + 217, + 146, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "LEROjD: Lidar Extended Radar-Only Object Detection", + "bbox": [ + 357, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "21. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. International Conference for Learning Representations (2015)", + "22. Klingner, M., Borse, S., Kumar, V.R., Rezaei, B., Narayanan, V., Yogamani, S., Porikli, F.: X3kd: Knowledge distillation across modalities, tasks and stages for multi-camera 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13343-13353 (June 2023)", + "23. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)", + "24. Lang, I., Manor, A., Avidan, S.: Samplenet: Differentiable point cloud sampling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7578-7588 (2020)", + "25. Li, D., Wei, Y., Zhu, R.: A comparative study on point cloud down-sampling strategies for deep learning-based crop organ segmentation. Plant Methods 19(1), 124 (2023)", + "26. Li, J., Lu, M., Liu, J., Guo, Y., Du, Y., Du, L., Zhang, S.: Bev-lgkd: A unified lidarguided knowledge distillation framework for multi-view bev 3d object detection. IEEE Transactions on Intelligent Vehicles pp. 1-10 (2023)", + "27. Li, J., Luo, C., Yang, X.: Pillarnext: Rethinking network designs for 3d object detection in lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17567-17576 (June 2023)", + "28. Li, Q., Jin, S., Yan, J.: Mimicking very efficient network for object detection. In: Proceedings of the iee conference on computer vision and pattern recognition. pp. 6356-6364 (2017)", + "29. Li, Z., Li, Y., Wang, Y., Xie, G., Qu, H., Lyu, Z.: A lightweight model for 3d point cloud object detection. Applied Sciences 13(11) (2023)", + "30. Liu, J., Zhao, Q., Xiong, W., Huang, T., Han, Q.L., Zhu, B.: Smurf: Spatial multi-representation fusion for 3d object detection with 4d imaging radar. IEEE Transactions on Intelligent Vehicles pp. 1-14 (2023)", + "31. Mao, J., Minzhe, N., Jiang, C., liang, h., Chen, J., Liang, X., Li, Y., Ye, C., Zhang, W., Li, Z., Yu, J., Xu, C., Xu, H.: One million scenes for autonomous driving: Once dataset. In: Vanschoren, J., Yeung, S. (eds.) Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks. vol. 1 (2021)", + "32. Meyer, M., Kuschk, G.: Automotive radar dataset for deep learning based 3d object detection. In: 2019 16th european radar conference (EuRAD). pp. 129-132. IEEE (2019)", + "33. Nguyen, C.H., Nguyen, T.C., Tang, T.N., Phan, N.L.: Improving object detection by label assignment distillation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1005-1014 (January 2022)", + "34. Nobis, F., Shafiei, E., Karle, P., Betz, J., Lienkamp, M.: Radar voxel fusion for 3d object detection. Applied Sciences 11(12) (2021)", + "35. Paek, D.H., Kong, S.H., Wijaya, K.T.: K-radar: 4d radar object detection for autonomous driving in various weather conditions. In: Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2022)", + "36. Palffy, A., Pool, E., Baratam, S., Kooij, J.F.P., Gavrila, D.M.: Multi-class road user detection with 3+1d radar in the view-of-delft dataset. IEEE Robotics and Automation Letters 7(2), 4961-4968 (2022)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "P. Palmer et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Palmer, P., Krueger, M., Altendorfer, R., Adam, G., Bertram, T.: Reviewing 3d object detectors in the context of high-resolution $3 + 1$ d radar. In: Workshop on 3D Vision and Robotics at the Conference on Computer Vision and Pattern Recognition 2023 (2023)", + "38. Palmer, P., Krueger, M., Altendorfer, R., Bertram, T.: Ego-motion estimation and dynamic motion separation from 3d point clouds for accumulating data and improving 3d object detection. In: AmE 2023 - Automotive meets Electronics; 14. GMM Symposium. pp. 86-91 (2023)", + "39. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)", + "40. Rebut, J., Ouaknine, A., Malik, W., Pérez, P.: Raw high-definition radar for multitask learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17021-17030 (June 2022)", + "41. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10529–10538 (2020)", + "42. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-779 (2019)", + "43. Smith, L.N., Topin, N.: Super-convergence: Very fast training of neural networks using large learning rates. In: Artificial intelligence and machine learning for multidomain operations applications. vol. 11006, pp. 369-386. SPIE (2019)", + "44. Stephan, M., Hazra, S., Santra, A., Weigel, R., Fischer, G.: People counting solution using an fmcw radar with knowledge distillation from camera data. In: 2021 IEEE Sensors. pp. 1-4 (2021)", + "45. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)", + "46. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020)", + "47. Sun, S., Zhang, Y.D.: 4d automotive radar sensing for autonomous vehicles: A sparsity-oriented approach. IEEE Journal of Selected Topics in Signal Processing 15(4), 879-891 (2021)", + "48. Tan, B., Ma, Z., Zhu, X., Li, S., Zheng, L., Chen, S., Huang, L., Bai, J.: 3d object detection for multi-frame 4d automotive millimeter-wave radar point cloud. IEEE Sensors Journal (2022)", + "49. Tung, F., Mori, G.: Similarity-preserving knowledge distillation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1365-1374 (2019)", + "50. Wang, L., Zhang, X., Li, J., Xv, B., Fu, R., Chen, H., Yang, L., Jin, D., Zhao, L.: Multi-modal and multi-scale fusion 3d object detection of 4d radar and lidar for autonomous driving. IEEE Transactions on Vehicular Technology 72(5), 5628-5641 (2023)", + "51. Wang, T., Yuan, L., Zhang, X., Feng, J.: Distilling object detectors with fine-grained feature imitation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4933-4942 (2019)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "LEROjD: Lidar Extended Radar-Only Object Detection", + "bbox": [ + 357, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "52. Wei, Y., Wei, Z., Rao, Y., Li, J., Zhou, J., Lu, J.: Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. In: European Conference on Computer Vision. pp. 179-195. Springer (2022)", + "53. Wu, H., Wen, C., Shi, S., Li, X., Wang, C.: Virtual sparse convolution for multimodal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21653-21662 (June 2023)", + "54. Xiao, A., Huang, J., Guan, D., Zhan, F., Lu, S.: Transfer learning from synthetic to real lidar point cloud for semantic segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2795-2803 (2022)", + "55. Xiao, Y., Liu, Y., Luan, K., Cheng, Y., Chen, X., Lu, H.: Deep lidar-radar-visual fusion for object detection in urban environments. Remote Sensing 15(18) (2023)", + "56. Xiong, W., Liu, J., Huang, T., Han, Q.L., Xia, Y., Zhu, B.: Lxl: Lidar excluded lean 3d object detection with 4d imaging radar and camera fusion. IEEE Transactions on Intelligent Vehicles pp. 1-14 (2023)", + "57. Xu, B., Zhang, X., Wang, L., Hu, X., Li, Z., Pan, S., Li, J., Deng, Y.: Rpfa-net: a 4d radar pillar feature attention network for 3d object detection. In: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). pp. 3061-3066 (2021)", + "58. Yan, Q., Wang, Y.: Mvfan: Multi-view feature assisted network for 4d radar object detection. In: International Conference on Neural Information Processing. pp. 493-511. Springer (2023)", + "59. Yang, J., Shi, S., Ding, R., Wang, Z., Qi, X.: Towards efficient 3d object detection with knowledge distillation. In: Advances in Neural Information Processing Systems. vol. 35, pp. 21300-21313. Curran Associates, Inc. (2022)", + "60. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3dssd: Point-based 3d single stage object detector. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11040-11048 (2020)", + "61. You, Y., Wang, Y., Chao, W.L., Garg, D., Pleiss, G., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-lidar++: Accurate depth for 3d object detection in autonomous driving. In: ICLR (2020)", + "62. Zamanakos, G., Tsochatzidis, L., Amanatiadis, A., Pratikakis, I.: A comprehensive survey of lidar-based 3d object detection methods with deep learning for autonomous driving. Computers & Graphics 99, 153-181 (2021)", + "63. Zhang, J., Zhuge, H., Liu, Y., Peng, G., Wu, Z., Zhang, H., Lyu, Q., Li, H., Zhao, C., Kircali, D., Mharolkar, S., Yang, X., Yi, S., Wang, Y., Wang, D.: Ntu4dradlm: 4d radar-centric multi-modal dataset for localization and mapping. In: 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC) (2023)", + "64. Zhang, L., Dong, R., Tai, H.S., Ma, K.: Pointdistiller: Structured knowledge distillation towards efficient and compact 3d detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21791-21801 (June 2023)", + "65. Zhang, X., Wang, L., Chen, J., Fang, C., Yang, L., Song, Z., Yang, G., Wang, Y., Zhang, X., Yang, Q., Li, J.: Dual radar: A multi-modal dataset with dual 4d radar for autonomous driving. arXiv preprint arXiv:2310.07602 (2023)", + "66. Zhang, Y., Hu, Q., Xu, G., Ma, Y., Wan, J., Guo, Y.: Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18953-18962 (June 2022)", + "67. Zhao, M., Li, T., Abu Alsheikh, M., Tian, Y., Zhao, H., Torralba, A., Katabi, D.: Through-wall human pose estimation using radio signals. In: Proceedings of" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "P. Palmer et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE conference on computer vision and pattern recognition. pp. 7356-7365 (2018)", + "68. Zheng, L., Li, S., Tan, B., Yang, L., Chen, S., Huang, L., Bai, J., Zhu, X., Ma, Z.: Rcfusion: Fusing 4-d radar and camera with bird's-eye view features for 3-d object detection. IEEE Transactions on Instrumentation and Measurement 72, 1-14 (2023)", + "69. Zheng, L., Ma, Z., Zhu, X., Tan, B., Li, S., Long, K., Sun, W., Chen, S., Zhang, L., Wan, M., Huang, L., Bai, J.: Tj4dradset: A 4d radar dataset for autonomous driving. In: 2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC). pp. 493-498 (2022)", + "70. Zhou, T., Chen, J., Shi, Y., Jiang, K., Yang, M., Yang, D.: Bridging the view disparity between radar and camera features for multi-modal fusion 3d object detection. IEEE Transactions on Intelligent Vehicles 8(2), 1523-1535 (2023)", + "71. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4490-4499 (2018)" + ], + "bbox": [ + 215, + 146, + 785, + 369 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "LEROjD: Lidar Extended Radar-Only Object Detection", + "bbox": [ + 357, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/LEROjD_ Lidar Extended Radar-Only Object Detection/e0f1cd56-0b49-499c-b5fc-daef469190ce_model.json b/2024/LEROjD_ Lidar Extended Radar-Only Object Detection/e0f1cd56-0b49-499c-b5fc-daef469190ce_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c1ada1352d3c41df055e42d3f25bc5aff6231501 --- /dev/null +++ b/2024/LEROjD_ Lidar Extended Radar-Only Object Detection/e0f1cd56-0b49-499c-b5fc-daef469190ce_model.json @@ -0,0 +1,2438 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.243, + 0.141, + 0.761, + 0.187 + ], + "angle": 0, + "content": "LEROjD: Lidar Extended Radar-Only Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.212, + 0.786, + 0.244 + ], + "angle": 0, + "content": "Patrick Palmer1, Martin Krüger1, Stefan Schütte1, Richard Altendorfer2, Ganesh Adam2, and Torsten Bertram1" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.254, + 0.774, + 0.296 + ], + "angle": 0, + "content": "\\(^{1}\\) Institute of Control Theory and Systems Engineering, TU Dortmund University {patrick.palmer, martin2.krueger, stefan.schuette, torsten.bertram}@tu-dortmund.de" + }, + { + "type": "text", + "bbox": [ + 0.296, + 0.297, + 0.706, + 0.311 + ], + "angle": 0, + "content": "\\(^{2}\\) ZF Group, {richard.altendorfer, ganeshadam}@zf.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.345, + 0.74, + 0.61 + ], + "angle": 0, + "content": "Abstract. Accurate 3D object detection is vital for automated driving. While lidar sensors are well suited for this task, they are expensive and have limitations in adverse weather conditions. \\(3 + 1\\mathrm{D}\\) imaging radar sensors offer a cost-effective, robust alternative but face challenges due to their low resolution and high measurement noise. Existing \\(3 + 1\\mathrm{D}\\) imaging radar datasets include radar and lidar data, enabling cross-modal model improvements. Although lidar should not be used during inference, it can aid the training of radar-only object detectors. We explore two strategies to transfer knowledge from the lidar to the radar domain and radar-only object detectors: 1. multi-stage training with sequential lidar point cloud thin-out, and 2. cross-modal knowledge distillation. In the multi-stage process, three thin-out methods are examined. Our results show significant performance gains of up to 4.2 percentage points in mean Average Precision with multi-stage training and up to 3.9 percentage points with knowledge distillation by initializing the student with the teacher's weights. The main benefit of these approaches is their applicability to other 3D object detection networks without altering their architecture, as we show by analyzing it on two different object detectors. Our code is available at https://github.com/rst-tu-dortmund/lerojd." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.621, + 0.741, + 0.65 + ], + "angle": 0, + "content": "Keywords: 3D Object Detection \\(\\cdot\\) 3+1D Imaging Radar \\(\\cdot\\) Cross-Modal Object Detection" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.674, + 0.377, + 0.691 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Environment perception is the first module in each automated driving stack. Multiple sensor modalities, like cameras, lidars, and radars are utilized for this task. Radar sensors are of unique interest in perception due to their robustness against poor lighting, challenging weather conditions like rain or snow, and cost-effectiveness [3, 36, 47]. One exclusive advantage is the ability to measure the relative radial velocity of reflections directly due to the Doppler effect." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "While a precise localization of objects is possible with traditional radar sensors without elevation angle measurements [9], it is inherently limited to the horizontal plane. Additionally, it is hard to predict the objects' extent due to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "P. Palmer et al." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.144, + 0.486, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.299, + 0.369, + 0.309 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.147, + 0.772, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.3, + 0.654, + 0.309 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.321, + 0.788, + 0.391 + ], + "angle": 0, + "content": "Fig. 1: Architecture overview of (a) a knowledge distillation-based method and (b) a multi-stage training method (MSTM) for utilizing lidar data in the training of radar-only object detectors. The ground truth (GT) label is the same for both methods. The diagrams above the dotted line represent the training process, while the diagrams below the dotted line represent inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.784, + 0.501 + ], + "angle": 0, + "content": "the reflections' sparsity. The introduction of \\(3 + 1\\mathrm{D}\\) high-resolution imaging radar sensors has recently mitigated these limitations, at least partially. In addition to measuring the elevation angle of reflections, the density of measurements is increased [14, 20]. Therefore, approaches that only utilize radar sensors for perception of the environment are of particular interest." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.785, + 0.564 + ], + "angle": 0, + "content": "Despite improvements in radar-based object detection, the performance still lags behind other sensor modalities like lidar [36]. One persistent major limitation of 2+1D classic radar and 3+1D imaging radar sensors is the relative sparsity of the point cloud, which limits the detection performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.785, + 0.642 + ], + "angle": 0, + "content": "Lidar sensors, on the other hand, are well suited for object detection and are therefore frequently employed as a reference for evaluating the performance of different sensor modalities due to their ability to produce an accurate and dense understanding of the scene. Their effectiveness is particularly pronounced in detecting nearby traffic participants without occlusion [53]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "All currently available datasets that contain \\(3 + 1\\mathrm{D}\\) imaging radar data additionally accommodate lidar sensor data [4,8,32,35,36,40,63,65,69]. The lidar sensor data of these datasets is currently either utilized for labeling, combining multiple sensor modalities for accurate object detection, or comparing the performance of radar-only techniques to another sensor modality. While the majority of series production vehicles may not include lidar sensors due to cost and vehicle design constraints, they will still be available in the training process of learning-based methods. Extending radar-only methods with lidar sensor data in training has been shown to be a viable method for estimating point flow on the imaging radar point cloud [12]. These observations lead to the following research question: Can lidar sensor data be utilized in the training process of imaging radar-based 3D object detectors to improve the object detection performance on radar-only data during inference?" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.358, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "LEROjD: Lidar Extended Radar-Only Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "To use different sensor modalities in training, transfer learning and knowledge distillation (KD) principles can be utilized. While KD is commonly employed across sensor modalities such as camera images and lidar point clouds, its application between \\(3 + 1\\mathrm{D}\\) imaging radar and lidar sensors remains unexplored. Since lidar and imaging radar share a structurally similar data representation as point clouds, an identical base network with different input modalities can be utilized to transfer knowledge between different input modalities. We investigate two approaches to transfer knowledge from lidar-based to radar-based object detectors: a KD-based and a multi-stage training approach. The main principles of the two methods are visualized in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.299, + 0.659, + 0.312 + ], + "angle": 0, + "content": "The contributions of this work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.321, + 0.784, + 0.351 + ], + "angle": 0, + "content": "- We investigate a combination of lidar and radar sensors in the training stage of object detectors to improve radar-only object detection at inference." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.351, + 0.784, + 0.381 + ], + "angle": 0, + "content": "- We investigate three thin-out strategies for lidar point clouds to transfer knowledge from dense lidar to sparse lidar and radar-only object detectors." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.381, + 0.784, + 0.411 + ], + "angle": 0, + "content": "- We propose a multi-stage training procedure to transfer knowledge from dense lidar to sparse lidar and, finally, to radar-only object detection." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.411, + 0.784, + 0.441 + ], + "angle": 0, + "content": "- We modify and analyze several knowledge distillation-based approaches to transfer knowledge from lidar to radar-only object detectors." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.321, + 0.784, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.463, + 0.388, + 0.479 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Imaging radar sensors commonly utilize the point cloud as a data representation format instead of the radar tensor, due to its higher computational efficiency. This format is similar to the one used for lidars, enabling the application of object detection methods developed for lidars to radars. Object detection on lidars can be split into two main categories. Point-based methods, like PointNet++ [39], downsample the original point cloud, encode it using a backbone, and finally apply a detection head. Voxel-based methods discretize the point cloud into a 3D grid and apply 3D convolutions to the grid before finally applying a detection head [6,10,71]. The main drawback of voxel-based methods is the high memory consumption and the loss of spatial information due to the discretization of the point cloud. To overcome the limitation of high memory consumption, [23,27] have proposed the PointPillars network. Pillars are a specific form of voxels that span over the full height of the scene and represent the point cloud in a 2D grid." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Methods from the lidar domain have been shown to perform reasonably well on radar data [36,37]. However, they are limited due to the sparsity of the radar point cloud. SMURF [30] considers two representations of the radar point cloud to address sparsity. Using kernel density estimation, it utilizes pillarization and density features derived from a multi-dimensional Gaussian mixture distribution. RPFA-Net [57] is a PointPillars [23] based network, which introduces a selfattention mechanism to extract global context information from the radar point cloud. RadarMFNet [48] utilizes a multi-frame radar point cloud representation to address the sparsity of the radar point cloud in conjunction with an anchor-based detector and temporal pooling layers." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "P. Palmer et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "One way of improving 3D object detection on radar data is the fusion with additional sensor modalities like camera [56, 68, 70], lidar [50] or camera and lidar [5, 13, 55] at the cost of introducing additional sensors at inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.787, + 0.33 + ], + "angle": 0, + "content": "The concept of knowledge distillation was first introduced by [18]. It consists of two networks, which are labeled as teacher and student. The teacher is a large and complex model, while the student is less complex and more computationally efficient. The student is trained to mimic the teacher network's performance by utilizing the teacher's predictions and the ground truth labels. This has been utilized in the context of 3D object detection by [29,59,64] to construct computation time-efficient models that have similar performance as large models. [26] and [22] have extended upon this concept by utilizing KD to extract knowledge from a teacher trained on lidar data to a student who utilizes camera images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.788, + 0.558 + ], + "angle": 0, + "content": "To the best of our knowledge, few studies have investigated the effect of KD considering radar data. A transfer of knowledge from an image-based teacher network to a radar-based student has been shown to improve the performance for the task of people counting [44]. For the task of 3D object detection, [22] has shown that a transfer learning-based approach from a lidar and image-based network to the radar domain results in improved object detection performance for classical \\(2 + 1\\mathrm{D}\\) radar sensors. The main drawback of [22] is that in addition to 3D object labels, instance segmentation labels for the image domain are required. Additionally, [22] introduces multiple sub-networks to derive the KD losses, which makes the model more complex, while we aim for a simpler approach. HiddenGems [12] utilizes lidar point clouds to derive point flow information and train a network to predict the point flow on radar point clouds. See Beyond Seeing [11] utilizes lidar point cloud for feature hallucination on radar point clouds. Both methods require the lidar point cloud only in the training process but require major modifications to the network architecture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.559, + 0.788, + 0.696 + ], + "angle": 0, + "content": "Sampling of point clouds is a broadly explored topic to reduce the complexity of computing large point clouds. Most commonly, farthest point sampling (FPS) [39,42,60], voxelized FPS [41] or random sampling [71] are utilized in the context of 3D object detection. Random sampling has been shown to be beneficial when used for semantic segmentation of point clouds [19] and to outperform farthest point sampling at this task [25]. One limitation of random sampling is that while points at a close distance are kept, points at a far distance are more likely to be removed. Therefore, [7] proposes a polar cylinder balanced random sampling to keep a more balanced distribution of points across the range." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.726, + 0.331, + 0.74 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Two methods of transferring knowledge from lidar-based to radar-based object detectors are investigated: Knowledge Distillation (KD) (Section 3.1), which is modified for the task of transfer learning and our proposed multi-stage training procedure with sequential point cloud thin-out (Section 3.2). Additionally, the utilized thin-out strategies for lidar point cloud (Section 3.3) are described." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.357, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "LEROjD: Lidar Extended Radar-Only Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.452, + 0.162 + ], + "angle": 0, + "content": "3.1 Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.337 + ], + "angle": 0, + "content": "KD is commonly used for two tasks. First, designing computationally efficient models by transferring knowledge from a larger teacher network to a smaller student network [18, 49, 51]. Second, to transfer knowledge across sensor modalities by utilizing different model architectures for the teacher and student [1, 16, 67]. Lidar and radar-based point cloud object detection can utilize the same model structure but with different input modalities. This enables the utilization of KD methods first described for designing computationally efficient models for cross-modal knowledge transfer from lidar- to radar-based object detectors. In this case, the teacher is trained on the full lidar point cloud, while the student is trained on the radar point cloud. Three different loss terms, as described by [59], are employed:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.788, + 0.413 + ], + "angle": 0, + "content": "Logit-KD is the first, classic type of distilling knowledge described by [18]. For 3D object detection, the logit-KD loss \\(\\mathcal{L}_{\\mathrm{logit}}\\) is split into a classification \\(\\mathcal{L}_{\\mathrm{l -cls}}\\) and regression loss \\(\\mathcal{L}_{\\mathrm{l - reg}}\\). These losses are calculated by comparing the student's and teacher's predictions utilizing the 3d detectors' regression loss and bi-linear interpolation between student and teacher output classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.788, + 0.473 + ], + "angle": 0, + "content": "Feature-KD is widely utilized in 2D object detection [28, 51]. It utilizes a loss term that forces the student network to mimic the teacher's intermediate feature map (feat). A feature mimicking the last layer of the bird's eye view feature encoder, similar to [59], is utilized in this work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.788, + 0.579 + ], + "angle": 0, + "content": "Label-KD is a recent distillation approach that leans on the concept of the Logit KD but simplifies and generalizes it. It is first described by [33]. The teacher predictions are filtered by their scores using a score threshold, and an adapted ground truth set is constructed by combining the filtered predictions and the ground truth set. This adapted set is utilized in student training. The loss is split into a classification \\(\\mathcal{L}_{\\mathrm{cls}}\\) and regression loss \\(\\mathcal{L}_{\\mathrm{reg}}\\). It replaces \\(\\mathcal{L}_{\\mathrm{label}}\\) usually calculated on the ground truth set." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.58, + 0.76, + 0.596 + ], + "angle": 0, + "content": "The three KD loss terms are combined into a joint loss weighted with \\(\\lambda\\)" + }, + { + "type": "equation", + "bbox": [ + 0.271, + 0.606, + 0.786, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {j o i n t}} = \\underbrace {\\lambda_ {\\mathrm {l - r e g}} \\mathcal {L} _ {\\mathrm {l - r e g}} + \\lambda_ {\\mathrm {l - c l s}} \\mathcal {L} _ {\\mathrm {l - c l s .}}} _ {\\mathcal {L} _ {\\text {l o g i t}}} + \\underbrace {\\lambda_ {\\text {f e a t}} \\mathcal {L} _ {\\text {f e a t}}} _ {\\mathcal {L} _ {\\text {f e a t u r e}}} + \\underbrace {\\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}} + \\lambda_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}}} _ {\\mathcal {L} _ {\\text {l a b e l}} ^ {*}} \\tag {1}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.665, + 0.661, + 0.68 + ], + "angle": 0, + "content": "3.2 Multi-Stage Lidar Thin-Out Training Procedure" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Utilizing a different input data modality for either pre-training a network on a large dataset or utilizing simulated data in the training process of a point cloud-based network has been shown to improve the object detection performance [54]. The multi-stage training method (MSTM) proposed in this work extends upon this by utilizing a Curriculum learning [2] based training procedure by which the network is trained on iteratively sparsified lidar point clouds, similar to [52], and fine-tuned on the radar point cloud. Figure 2 visualizes our multi-stage training procedure. The network is first trained on the full lidar point cloud until convergence. In the following steps, the lidar point cloud is thinned-out iteratively by a factor of 2 and utilized for training a network whose weights are" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "P. Palmer et al." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.144, + 0.775, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.561 + ], + "angle": 0, + "content": "Fig. 2: MSTM pipeline. A 3D object detection network is iteratively trained on increasingly sparse lidar data. Three different thin-out strategies are utilized for sparsification. The lidar point cloud is mixed with the radar point cloud in the second to last step. In the last step, the network is only trained on radar data. At inference, only the orange-shaded part is executed. Small points represent lidar points, large points radar points. The color of points corresponds to the distance from the ego-vehicle. The camera image is not used as an input but only aids the visualization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.686 + ], + "angle": 0, + "content": "initialized using the previously trained model. This forces the network to learn features for a good object detection performance on increasingly sparser point clouds. In the second to last step, the lidar point cloud is mixed with the radar points so that the network can translate from the lidar to the radar domain. In the last step, the model is trained only on radar points. A training without multiple stages is called single-stage training method (SSTM) in this work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.779 + ], + "angle": 0, + "content": "In addition to the training where only lidar points are utilized in the first training stages, we investigate the utilization of the radar point cloud in conjunction with the lidar point cloud in all stages. The thin-out of the lidar points remains the same and is mixed with the radar point cloud in each step. This conditions the model on radar from the first step in order to prioritize features in the lidar point cloud that relate to a good object detection on radar-only data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "For mixing lidar and radar point clouds, the voxel or pillar feature encoder is modified to prioritize the radar point cloud in the random sampling process as done by [34]. Otherwise, the vastly higher number of lidar points, even when thinned-out, might lead to the complete exclusion of radar points." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.357, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "LEROjD: Lidar Extended Radar-Only Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.478, + 0.162 + ], + "angle": 0, + "content": "3.3 Lidar Thin-Out Strategies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.784, + 0.223 + ], + "angle": 0, + "content": "Three different methods for sub-sampling the lidar point cloud are investigated. Figure 2 shows examples for each thin-out stage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.226, + 0.785, + 0.302 + ], + "angle": 0, + "content": "Random sampling is the simplest method of sparsifying a point cloud. It neglects the structure and inherent limitations of the point cloud representation, especially for objects far away or with a high degree of occlusion. Neglecting the structure can lead to the complete loss of information for objects represented only by few points." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.306, + 0.785, + 0.382 + ], + "angle": 0, + "content": "K-nearest neighbor sampling approximates the reflection density distribution of the radar point cloud with the lidar point cloud by only keeping lidar reflections close to radar reflections. This algorithm is described in Algorithm 1. Objects not detected by the radar sensor are, therefore, also not represented by the k-nearest neighbor thinned-out lidar point cloud." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.787, + 0.461 + ], + "angle": 0, + "content": "Voxel-based sampling aims to reduce the number of points in each area of the point cloud while keeping the general distribution of the point cloud. This is motivated by the fact that radar sensors do not suffer as much from loss of resolution with distance as lidar sensors. The approach is described in Algorithm 2, and executed iteratively for a sequence of sparsification steps." + }, + { + "type": "code_caption", + "bbox": [ + 0.229, + 0.469, + 0.55, + 0.484 + ], + "angle": 0, + "content": "Algorithm 1: K-nearest neighbor sampling" + }, + { + "type": "algorithm", + "bbox": [ + 0.216, + 0.485, + 0.787, + 0.791 + ], + "angle": 0, + "content": "Input : Lidar point cloud \\(L\\in \\mathbb{R}^{N\\times 3}\\) , radar point cloud \\(R\\in \\mathbb{R}^{M\\times 3}\\) Output: Sub-sampled lidar point cloud \\(L_{s}\\in \\mathbb{R}^{K\\times 3}\\) \n1 Calculate the euclidean distance \\(c_{i}\\) between each lidar point \\(l_{i}\\in L\\) and its nearest radar point from \\(R\\) . \n2 Select a share of \\(K\\) lidar points with the smallest distances c; \n3 return Share of Points \\(L_{s}\\in \\mathbb{R}^{K\\times 3}\\) \nAlgorithm 2: Voxel-based sampling. \nInput : Lidar point cloud \\(L\\in \\mathbb{R}^{N\\times 3}\\) Output: Sub-sampled lidar point cloud \\(L_{s}\\in \\mathbb{R}^{K\\times 3}\\) \n1 Initialize \\(L_{s}\\) with \\(L\\) . \n2 Voxelize the lidar point cloud into v voxels; \n3 Calculate the number of points in each voxel \\(P_{v}\\) . \n4 Calculate the minimum number of points per voxel \\(p_{\\mathrm{min}}\\) , so that at least \\(0.75N\\) points are in voxels with more than \\(p_{\\mathrm{min}}\\) points; \n5 From each voxel with more than \\(p_{\\mathrm{min}}\\) points, choose \\(p_{\\mathrm{min}}\\) random points to keep and add remaining points to \\(L_{p}\\) . \n6 Randomly select O.5 N points from \\(L_{p}\\) and remove them from \\(L_{s}\\) . \n7 return Share of Points \\(L_{s}\\in \\mathbb{R}^{K\\times 3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Another common sparsification method utilized for imitating low-resolution lidar sensors is layer-based sampling [52, 61]. This approach is not investigated because radar sensors do not capture the environment on a layer basis." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "P. Palmer et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.494, + 0.163 + ], + "angle": 0, + "content": "4 Experimental Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.176, + 0.428, + 0.192 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.199, + 0.794, + 0.411 + ], + "angle": 0, + "content": "Dataset: All experiments are conducted utilizing the View-of-Delft (VoD) dataset [36]. It contains synchronized data of multiple sensor modalities. The 64-layer lidar sensor and the imaging radar are utilized in this work. A point cloud accumulated over 5 frames [36], which has been shown to improve object detection performance compared to no accumulation, is used for radar data [36, 38]. We detected a duplication of identical points in the lidar point cloud, which could adversely affect all sampling methods; thus, we eliminated the duplicated points from the point cloud. Although the VoD dataset is among the best currently available datasets for imaging radar-based object detection, it is limited by its size compared to other automotive datasets without imaging-radar, like [31, 45]. Given the absence of publicly available labels and limited online evaluation for the test dataset, we repurpose the validation dataset as a test set. Consequently, we partition the original training set into a new training set \\((80\\%)\\) and a dedicated validation set \\((20\\%)\\) to ensure robust model training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.79, + 0.501 + ], + "angle": 0, + "content": "Evaluation Metrics: The primary performance metric utilized to compare the results is the mean average precision (mAP), as used by [15] [37]. Similar to the evaluation of the Waymo data set [46] in [62], we split the results into two distance bins: short-range (SR): \\(0 - 30\\mathrm{m}\\) and mid-range (MR): \\(30 - 50\\mathrm{m}\\). All experiments are conducted utilizing three different random network initializations that are averaged." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.79, + 0.728 + ], + "angle": 0, + "content": "Training: Most experiments use the PointPillars model [23] as an object detector with the same configuration as utilized by [36]. For imaging radar data, PointPillars has been shown to perform among the best out of multiple state-of-the-art 3D object detection methods while still performing adequately on lidar data [37]. Furthermore PointPillars is a relevant baseline for radar-specific object detection methods [57,58]. To show that the proposed MSTM and KD apply to various object detectors, the most promising approaches from the evaluations on PointPillars are evaluated on DSVT-P [17], as an example for a transformer-based model. All SSTM trainings are conducted with an early stopping policy for a maximum of 125 epochs. For the MSTM the initial training on the full lidar point cloud is conducted for 125 epochs, while each refinement step is trained for 30 epochs. All trainings utilize the Adam optimizer [21] and an adapted learning rate scheduler that reaches its maximum learning rate earlier and has a faster descent than the scheduler described by [43]. This improves the object detection performance on radar data." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.728, + 0.783, + 0.743 + ], + "angle": 0, + "content": "Notation: To distinguish between methods, the following notation is used:" + }, + { + "type": "equation", + "bbox": [ + 0.433, + 0.768, + 0.574, + 0.787 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {\\mathcal {L S} / \\mathcal {T O}} ^ {\\mathcal {T M}} \\rightarrow \\mathcal {T} ^ {\\mathcal {K D}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The notation is split into two parts. The part left of the arrow represents the data set utilized for pre-training, while the right part represents the data and training method utilized for the last (fine-tuning) training stage of the model." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.357, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "LEROjD: Lidar Extended Radar-Only Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 1: Examples of training steps and evaluation data corresponding to the notation, emplying voxel-based sub-sampling. Voxel sampling is shortened to v for legibility." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.185, + 0.773, + 0.304 + ], + "angle": 0, + "content": "
NotationTrainingEvaluation
L1-1/8/vL1/v → L1/2/v → L1/4/v → L1/8/vL1/8/v
RLMSTM1-1/8/vRL1/v → RL1/2/v → RL1/4/v → RL1/8/vRL1/8/v
L1MSTM1-1/16/v → RL1/v → L1/2/v → L1/4/v → L1/8/v → RL1/16/v → RR
RLMSTM1-1/16/v → RRL1/v → RL1/2/v → RL1/4/v → RL1/8/v → RL1/16/v → RR
RLSSTM1/4/v → RRL1/4/v → RR
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.495 + ], + "angle": 0, + "content": "This part is omitted if it is the same as the training dataset. The training data \\(\\mathcal{T}\\) can either be lidar (L), radar (R), or mixed radar + lidar data (RL). The training method \\(\\mathcal{T}\\mathcal{M}\\) can denote either MSTM or SSTM. The lidar share \\(\\mathcal{L}\\mathcal{S}\\) is the fraction of the original lidar point cloud that is utilized. Radar-only training always uses the full radar point cloud, therefore \\(\\mathcal{L}\\mathcal{S}\\) is omitted. For the MSTM, this is represented by a range of fractions that are iterated in training. The thin-out method \\(\\mathcal{T}\\mathcal{O}\\) can either be random (rand), k-nearest neighbor (knn) or voxel-based (vox). \\(\\mathcal{KD}\\) represents the KD method; this can either be label (lab), logit (log), feature (feat), or a combination of those (joint). \\(\\mathcal{KD}\\) is omitted if just an initialization and fine-tuning is utilized. Examples for the training corresponding to specific notations are listed in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.515, + 0.784, + 0.531 + ], + "angle": 0, + "content": "4.2 Evaluation of MSTM on Lidar-only and Mixed Radar + Lidar" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.689 + ], + "angle": 0, + "content": "To evaluate the applicability of our proposed MSTM to the thinned-out lidar point cloud, we evaluate the training without the last two steps involving radar data. Training is conducted for just the lidar point cloud and the mixed radar + lidar point cloud in all stages. The multi-stage trained network is evaluated after each thin-out stage on the thinned-out lidar (or mixed radar + lidar) point cloud. Thin-out stages up to \\(1/256\\) of the original lidar point cloud are considered due to the lidar point cloud containing fewer points than the radar point cloud at \\(1/256\\) of the original lidar point cloud. The MSTM is compared to the SSTM, which is trained only on the thinned-out point cloud. The results are shown in Fig. 3, complete quantitative results are given in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "All thin-out strategies result in an increasing degradation of the detection performance. This is most pronounced when considering a random thin-out, which drops approximately linearly with each thin-out step. K-nearest neighbor and voxel-based sampling consistently perform better than random sampling due to keeping a higher point density in areas around objects, which supports object detection. For k-nearest neighbor sampling, the performance only drops by 0.6 percentage points between \\(\\mathrm{L}_{1 / \\mathrm{knn}}^{\\mathrm{SSTM}}\\) and \\(\\mathrm{L}_{1 / 2 / \\mathrm{knn}}^{\\mathrm{SSTM}}\\) due to mostly ground points being removed in the first thin-out stage. For voxel-based sampling, after a sharp performance drop in the first thin-out stage, only a relatively slight performance drop is observable until the \\(^{1/16}\\) thin-out stage. In the first stages," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "P. Palmer et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.782, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.312, + 0.785, + 0.327 + ], + "angle": 0, + "content": "Fig. 3: Comparison between SSTM and MSTM for different lidar sampling strategies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.537 + ], + "angle": 0, + "content": "the performance drop can mainly be attributed to the pedestrian class, which drops by 12.8 percentage points in the first thin-out stage. In comparison, the detection performance of the car and cyclist classes only decreased by 0.7 and 3 percentage points, respectively. This can be explained by the voxel size of \\(1\\mathrm{m}\\) in each dimension utilized in the voxel-based sampling. A single pedestrian only occupies a small number of these voxels. When the point cloud gets thinned out, the entire object's structure gets lost, making detection difficult. In contrast, the object's structure and physical appearance can be adequately represented when it occupies more voxels, as observed in the car and cyclist classes. In the final thin-out stage of \\(\\mathrm{L}^{\\mathrm{SSTM}}\\) and \\(\\mathrm{L}^{\\mathrm{MSTM}}\\), voxel-based sampling performs worse than k-nearest neighbor sampling because many voxels only consist of ground points or points of surrounding background objects." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.538, + 0.786, + 0.567 + ], + "angle": 0, + "content": "The multi-stage training does not result in a useful knowledge transfer and, therefore, significant performance benefit for all considered thin-out strategies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.704 + ], + "angle": 0, + "content": "A contrary behavior is observed on the combined radar + lidar point cloud. The MSTM consistently outperforms the SSTM when using k-nearest neighbor or voxel-based sampling. Knowledge from the dense point cloud can be transferred to the thin radar + lidar point cloud. Additionally, performance is consistently higher than training just on lidar, especially for voxel-based subsampling, which performs best at lower thin-out stages. At small thin-out stages, the voxel-based sub-sampling can still represent the whole object space and give meaningful environmental information. At the same time, the radar point cloud is sufficiently dense for object detection." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.725, + 0.667, + 0.74 + ], + "angle": 0, + "content": "4.3 Evaluation of MSTM with Last Radar-Only Step" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.843 + ], + "angle": 0, + "content": "This chapter analyses the performance of the MSTM when applied to radar data, as described in Section 3.2. The MSTM is evaluated for two different procedures. Utilizing only the lidar point cloud and utilizing the mixed radar + lidar point cloud in the first stages. The results of MSTM are shown in Table 2. The lidar thin-out stages up to \\(1/16\\) of the original lidar point cloud are considered due to the performance of \\(\\mathrm{L}_{1-1/32}^{\\mathrm{MSTM}}\\) dropping below \\(\\mathrm{R}^{\\mathrm{SSTM}}\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.358, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "LEROjD: Lidar Extended Radar-Only Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: 3D object detection results on the VoD radar dataset trained using MSTM. All MSTM methods are trained on radar-only data in the last step. The best and second best results are marked in bold and underlined, respectively." + }, + { + "type": "table", + "bbox": [ + 0.234, + 0.199, + 0.766, + 0.364 + ], + "angle": 0, + "content": "
Training methodmAPCarPedestrianCyclist
SRMRSRMRSRMRSRMR
RSSTM (baseline)36.711.945.218.117.17.447.710.2
L1-1/16 / rand → R36.416.144.920.017.59.846.818.9
L1-1/16 / knn → R35.714.045.623.216.77.144.811.7
L1-1/16 / vox → R34.114.347.023.015.39.140.010.7
RLMSTM 1-1/16 / rand → R35.614.944.120.217.87.244.917.3
RLMSTM 1-1/16 / knn → R38.214.745.523.918.88.250.311.9
RLMSTM 1-1/16 / vox → R39.715.445.922.518.49.754.713.9
" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.39, + 0.786, + 0.571 + ], + "angle": 0, + "content": "When considering the pre-training using just lidar data, it is observable that the overall performance drops in the SR bin. In contrast, the MR bin is improved for all thin-out methods. The best performance is achieved by utilizing the random thin-out strategy. It performs especially well for pedestrians and cyclists in the mid-range, due to the comparably small size of pedestrians and cyclists resulting in a worse radar reflection characteristic. In contrast these objects are detected well by the lidar sensor. Knowledge about the representation of objects can be transferred from the lidar point cloud to the radar point cloud. K-nearest neighbor and voxel-based thin-out do not improve the performance in the SR bin for the pedestrian and cyclist class but perform better in the short and mid-range for the car class which is explained by the large size of cars, resulting in a better representation of cars in the thin point cloud." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.572, + 0.787, + 0.768 + ], + "angle": 0, + "content": "When considering the pre-training with the mixed radar + lidar point cloud, a contrary observation is made compared to only using the lidar in pre-training. The overall performance for the random thinned-out lidar point cloud is worse with the mixed point cloud than when only considering the lidar point cloud. One consistent aspect is the good performance of the cyclists in the mid-range, which still surpasses all other methods, excluding \\( \\mathrm{L}_{1 - 1/16}^{\\mathrm{MSTM}} \\). However, the k-nearest neighbor and voxel-based thin-out strategies perform better with the mixed point cloud. The voxel-based thin-out strategy achieves the best performance. It performs best on objects in SR, mainly due to its outstanding performance in detecting cyclists, but is limited in the mid-range, getting surpassed by the MSTM with random thin-out. The MSTM with voxel-based thin-out can increase the object detection performance on the radar point cloud by 3 percentage points in the SR and 3.5 percentage points in the MR." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.788, + 0.54, + 0.803 + ], + "angle": 0, + "content": "4.4 Evaluation of Cross-Modality KD" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The configuration of the teacher's training data is of particular interest, as the teacher's performance directly influences the student's. The simplest choice is" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "P. Palmer et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.187 + ], + "angle": 0, + "content": "Table 3: 3D object detection results of the teacher with different training sets. Each teacher is evaluated on the test set of the same data configuration as the training set. The best and second best results are marked in bold and underlined, respectively." + }, + { + "type": "table", + "bbox": [ + 0.251, + 0.199, + 0.747, + 0.322 + ], + "angle": 0, + "content": "
DatamAPCarPedestrianCyclist
SRMRSRMRSRMRSRMR
L1SSTM56.530.060.742.440.119.568.528.1
RLSSTM61.645.161.946.544.434.078.554.8
RLSSTM1/4 / rand56.931.157.232.437.221.376.439.5
RLSSTM1/4 / knn58.936.954.431.943.030.979.547.9
RLSSTM1/4 / Vox56.139.360.141.632.828.875.547.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.488 + ], + "angle": 0, + "content": "to train the teacher solely on the lidar point cloud. Section 4.2 and Section 4.3 show that mixing the radar and lidar point clouds can benefit radar-only object detection. Therefore, teachers trained on a mixed point clouds, in addition to the ones solely trained on lidar, are compared. Mixed point clouds containing \\(1/4\\) of the original lidar point cloud are considered, as a closer representation of the teachers to the student's data can lead to a better performance. Thin out of \\(1/4\\) is chosen as the radar + lidar detection performance with radar and \\(1/8\\) of the lidar points is worse than the SSTM on only lidar data. Table 3 shows the performance of all utilized teachers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.564 + ], + "angle": 0, + "content": "Table 3 shows the results of the teacher network for different training data configurations. It can be observed that, as expected, the mixed radar and lidar point cloud performs the best, only being surpassed by \\(\\mathrm{RL}_{1/4}^{\\mathrm{SSTM}}\\) on the cyclist class in short range. No model can be considered the overall second best among the other training sets. The performance varies between vehicle classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.655 + ], + "angle": 0, + "content": "The KD is evaluated individually for each KD method and teacher training set. A joint KD is also considered comprised of all three KD losses. All student networks are initialized (Init) with the weights of the teacher network, as it has been shown to improve the student's performance [59]. Additionally, for the transfer learning between datasets, the pre-training utilizing MSTM in Section 4.3 has been shown to improve the performance on the radar dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.788, + 0.735 + ], + "angle": 0, + "content": "The results utilizing the KD are shown in Table 4. Just the initialization of the student network with the teacher's weights already results in a performance benefit in the SR with \\(\\mathrm{RL}_{1}^{\\mathrm{SSTM}}\\) and \\(\\mathrm{RL}_{1/4}^{\\mathrm{SSTM}}\\) as teachers and in the MR with all teachers. This is overall only surpassed by the \\(\\mathrm{RL}_{1-1/16}^{\\mathrm{MSTM}} / \\mathrm{vox} \\to \\mathrm{R}\\) showing that the MSTM is substantially better than a simple initialization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "For label-KD, the best-performing models are the ones where the teacher performs the best. Specifically \\(\\mathrm{L}_1^{\\mathrm{SSTM}}\\rightarrow \\mathrm{R}^{\\mathrm{lab}}\\), which is only surpassed by \\(\\mathrm{RL}_{1/4/knn}^{\\mathrm{SSTM}}\\rightarrow \\mathrm{R}^{\\mathrm{lab}}\\) in the MR. For worse-performing teachers, using label-KD loss does not result in a performance benefit due to it replacing the ground truth label loss. Feature-KD requires a teacher who learns features from radar data. This is observable in the SR performance of models, where the teacher is trained on mixed radar + lidar data. Logit-KD works well on teacher datasets" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.357, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "LEROjD: Lidar Extended Radar-Only Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 4: 3D object detection results on the VoD radar dataset utilizing the different KDs and teachers trained on different data configurations. Only the mAP over all classes is specified. For each distillation method and range bin, the best and second best results are marked in bold and underlined, respectively. The best result for each teacher are marked in cyan for the SR bin and magenta for the MR bin." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.227, + 0.789, + 0.386 + ], + "angle": 0, + "content": "
Teacher dataInit only (→R)Logit-KD (→Rlog)Feature-KD (→Rfeat)Label-KD (→Rlab)Joint-KD (→RJoint)
SRMRSRMRSRMRSRMRSRMR
R1SSTM36.711.9--------
L1SSTM34.913.532.812.634.813.636.613.736.013.5
RLSSTM39.014.839.213.539.113.838.813.738.912.7
RLSSTM1/4 / rand39.014.238.911.639.411.834.012.236.412.3
RLSSTM1/4 / knn34.814.637.214.736.613.033.810.932.912.9
RLSSTM1/4 / Vox35.215.335.613.337.115.835.314.332.99.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.789, + 0.594 + ], + "angle": 0, + "content": "which closely resemble the radar point cloud and perform well on the teacher set. This results in good performance of \\(\\mathrm{RL}_1^{\\mathrm{SSTM}}\\rightarrow \\mathrm{R}^{\\log}\\). Besides the KD method the thin-out method utilized in the teacher's training influences the student's performance. Random sampling results in good performance in the SR, while voxel-based sampling results in good MR performance. This is somewhat contrary to what has been observed in the MSTM, where random sampling performs the best in the MR and can be explained by the different thin-out stages utilized in the MSTM and KD. Further qualitative results, as well as detailed quantitative results are given in the supplementary. Overall, initializing the student with the teacher's parameters yields good performance, with further enhancements primarily achievable through feature-KD with a teacher trained on mixed radar + lidar data." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.617, + 0.776, + 0.634 + ], + "angle": 0, + "content": "4.5 Evaluation on DSVT as a Transformer-Based Object Detector" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Table 5 shows the results of the MSTM and KD on DSVT-P [17] for selected methods. Similar effects, as observed for PointPillars, are seen on DSVT-P. Initializing the student with the teacher's weights and the MSTM led to a performance benefit. However, contrary to PointPillars, the KD approach does not contribute to any improvements for DSVT-P." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.74, + 0.359, + 0.755 + ], + "angle": 0, + "content": "4.6 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The best-performing methods in this work apply only to detectors that share an architecture with the target model, as a direct transfer of weights is performed for the best performance. Different models may require different levels and steps in the thinning process. These choices are additional tuning parameters that must be selected appropriately to maximize the benefit of transferring knowledge." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "P. Palmer et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 5: Radar-only detection performance utilizing DSVT-P as an object detector. The best and second best results are marked in bold and underlined, respectively." + }, + { + "type": "table", + "bbox": [ + 0.234, + 0.185, + 0.764, + 0.307 + ], + "angle": 0, + "content": "
Training methodmAPCarPedestrianCyclist
SRMRSRMRSRMRSRMR
RSSTM38.313.042.520.221.912.350.56.6
RLMSTM1-1/16 / vox→R41.515.647.823.624.511.052.312.4
RLSSTMM1/4 / vox→R38.813.244.918.823.311.248.29.7
RLSSTMM1/4 / vox→Rfeat34.711.344.08.820.52.839.712.7
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.332, + 0.359, + 0.348 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.787, + 0.544 + ], + "angle": 0, + "content": "In this paper, we investigated two methods to transfer knowledge from lidar-based object detectors to radar-only object detection. First, MSTM with sequential sub-sampling of the lidar point cloud, and second, a KD-based approach. For the MSTM, we have investigated three thin-out strategies for the lidar point cloud. These thin-out strategies are also analyzed for the training of the KD teacher network. Both methods can substantially improve the object detection performance of a radar-only object detector. The MSTM with voxel-based thinout performs the best overall and can improve detection performance by up to 3.5 percentage points. For the KD methods, it is shown that initializing the student with the teacher's parameters, especially a teacher trained on mixed lidar and radar data, can improve the object detection performance on radar-only data, with further enhancement primarily achieved by utilizing feature-KD." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.787, + 0.65 + ], + "angle": 0, + "content": "In future work, the applicability to further 3D object detection networks and the behavior with more advanced knowledge distillation like the ones utilized by [22] could be investigated. Due to different effects being noticed with the MSTM and the KD methods, combining both methods could be investigated by utilizing the MSTM as a teacher. To overcome the limitations of choosing a strict thin-out strategy, a learnable point cloud thin-out method [24,66] can be used." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.672, + 0.325, + 0.687 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.702, + 0.785, + 0.743 + ], + "angle": 0, + "content": "1. Afouras, T., Chung, J.S., Zisserman, A.: Asr is all you need: Cross-modal distillation for lip reading. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 2143-2147. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.745, + 0.785, + 0.785 + ], + "angle": 0, + "content": "2. Bengio, Y., Louradour, J., Collobert, R., Weston, J.: Curriculum learning. In: Proceedings of the 26th annual international conference on machine learning. pp. 41-48 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.786, + 0.785, + 0.812 + ], + "angle": 0, + "content": "3. Briske, S., Ruf, F., Höhne, F.: Recent evolution of automotive imaging radar and its information content. IET Radar, Sonar & Navigation 12(10), 1078-1081 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous" + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.702, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.358, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "LEROjD: Lidar Extended Radar-Only Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.175, + 0.787, + 0.217 + ], + "angle": 0, + "content": "5. Chen, X., Zhang, T., Wang, Y., Wang, Y., Zhao, H.: Futr3d: A unified sensor fusion framework for 3d detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 172-181 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.217, + 0.787, + 0.257 + ], + "angle": 0, + "content": "6. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21674-21683 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.257, + 0.787, + 0.298 + ], + "angle": 0, + "content": "7. Cheng, H., Han, X., Jiang, H., He, D., Xiao, G.: Pcb-random sampling for lidar semantic segmentation in autonomous driving scene. arXiv preprint arXiv:2209.13797 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.298, + 0.787, + 0.338 + ], + "angle": 0, + "content": "8. Choi, M., Yang, S., Han, S., Lee, Y., Lee, M., Choi, K.H., Kim, K.S.: Msc-rad4r: Ros-based automotive dataset with 4d radar. IEEE Robotics and Automation Letters 8(11), 7194-7201 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.338, + 0.787, + 0.379 + ], + "angle": 0, + "content": "9. Danzer, A., Griebel, T., Bach, M., Dietmayer, K.: 2d car detection in radar data with pointnets. In: 2019 IEEE Intelligent Transportation Systems Conference (ITSC). pp. 61-66 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.379, + 0.787, + 0.42 + ], + "angle": 0, + "content": "10. Deng, J., Shi, S., Li, P., Zhou, W., Zhang, Y., Li, H.: Voxel r-cnn: Towards high performance voxel-based 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1201-1209 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.42, + 0.787, + 0.46 + ], + "angle": 0, + "content": "11. Deng, J., Chan, G., Zhong, H., Lu, C.X.: See beyond seeing: Robust 3d object detection from point clouds via cross-modal feature augmentation. arXiv preprint arXiv:2309.17336 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.46, + 0.787, + 0.514 + ], + "angle": 0, + "content": "12. Ding, F., Palffy, A., Gavrila, D.M., Lu, C.X.: Hidden gems: 4d radar scene flow learning using cross-modal supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9340-9349 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.514, + 0.787, + 0.569 + ], + "angle": 0, + "content": "13. Drews, F., Feng, D., Faion, F., Rosenbaum, L., Ulrich, M., Glaser, C.: Deepfusion: A robust and modular 3d object detector for lidars, cameras and radars. In: 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 560-567. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.569, + 0.787, + 0.61 + ], + "angle": 0, + "content": "14. Engels, F., Heidenreich, P., Winternantel, M., Stäcker, L., Al Kadi, M., Zoubir, A.M.: Automotive radar signal processing: Research directions and practical challenges. IEEE Journal of Selected Topics in Signal Processing 15(4), 865-878 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.61, + 0.787, + 0.651 + ], + "angle": 0, + "content": "15. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.651, + 0.787, + 0.692 + ], + "angle": 0, + "content": "16. Gupta, S., Hoffman, J., Malik, J.: Cross modal distillation for supervision transfer. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2827-2836 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.692, + 0.787, + 0.718 + ], + "angle": 0, + "content": "17. Haiyang Wang, Chen Shi, S.S.M.L.S.W.D.H.B.S., Wang, L.: Dsvt: Dynamic sparse voxel transformer with rotated sets. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.718, + 0.787, + 0.744 + ], + "angle": 0, + "content": "18. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. Deep Learning Workshop at NIPS (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.744, + 0.787, + 0.8 + ], + "angle": 0, + "content": "19. Hu, Q., Yang, B., Xie, L., Rosa, S., Guo, Y., Wang, Z., Trigoni, N., Markham, A.: Randla-net: Efficient semantic segmentation of large-scale point clouds. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11108-11117 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "20. Jiang, M., Xu, G., Pei, H., Feng, Z., Ma, S., Zhang, H., Hong, W.: 4d high-resolution imagery of point clouds for automotive mmwave radar. IEEE Transactions on Intelligent Transportation Systems pp. 1-15 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "P. Palmer et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "21. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. International Conference for Learning Representations (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.787, + 0.245 + ], + "angle": 0, + "content": "22. Klingner, M., Borse, S., Kumar, V.R., Rezaei, B., Narayanan, V., Yogamani, S., Porikli, F.: X3kd: Knowledge distillation across modalities, tasks and stages for multi-camera 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13343-13353 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.785, + 0.288 + ], + "angle": 0, + "content": "23. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.785, + 0.33 + ], + "angle": 0, + "content": "24. Lang, I., Manor, A., Avidan, S.: Samplenet: Differentiable point cloud sampling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7578-7588 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.785, + 0.373 + ], + "angle": 0, + "content": "25. Li, D., Wei, Y., Zhu, R.: A comparative study on point cloud down-sampling strategies for deep learning-based crop organ segmentation. Plant Methods 19(1), 124 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.785, + 0.415 + ], + "angle": 0, + "content": "26. Li, J., Lu, M., Liu, J., Guo, Y., Du, Y., Du, L., Zhang, S.: Bev-lgkd: A unified lidarguided knowledge distillation framework for multi-view bev 3d object detection. IEEE Transactions on Intelligent Vehicles pp. 1-10 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.785, + 0.459 + ], + "angle": 0, + "content": "27. Li, J., Luo, C., Yang, X.: Pillarnext: Rethinking network designs for 3d object detection in lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17567-17576 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "28. Li, Q., Jin, S., Yan, J.: Mimicking very efficient network for object detection. In: Proceedings of the iee conference on computer vision and pattern recognition. pp. 6356-6364 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.529 + ], + "angle": 0, + "content": "29. Li, Z., Li, Y., Wang, Y., Xie, G., Qu, H., Lyu, Z.: A lightweight model for 3d point cloud object detection. Applied Sciences 13(11) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.531, + 0.785, + 0.572 + ], + "angle": 0, + "content": "30. Liu, J., Zhao, Q., Xiong, W., Huang, T., Han, Q.L., Zhu, B.: Smurf: Spatial multi-representation fusion for 3d object detection with 4d imaging radar. IEEE Transactions on Intelligent Vehicles pp. 1-14 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.785, + 0.628 + ], + "angle": 0, + "content": "31. Mao, J., Minzhe, N., Jiang, C., liang, h., Chen, J., Liang, X., Li, Y., Ye, C., Zhang, W., Li, Z., Yu, J., Xu, C., Xu, H.: One million scenes for autonomous driving: Once dataset. In: Vanschoren, J., Yeung, S. (eds.) Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks. vol. 1 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.63, + 0.785, + 0.67 + ], + "angle": 0, + "content": "32. Meyer, M., Kuschk, G.: Automotive radar dataset for deep learning based 3d object detection. In: 2019 16th european radar conference (EuRAD). pp. 129-132. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.785, + 0.727 + ], + "angle": 0, + "content": "33. Nguyen, C.H., Nguyen, T.C., Tang, T.N., Phan, N.L.: Improving object detection by label assignment distillation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1005-1014 (January 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.785, + 0.755 + ], + "angle": 0, + "content": "34. Nobis, F., Shafiei, E., Karle, P., Betz, J., Lienkamp, M.: Radar voxel fusion for 3d object detection. Applied Sciences 11(12) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "35. Paek, D.H., Kong, S.H., Wijaya, K.T.: K-radar: 4d radar object detection for autonomous driving in various weather conditions. In: Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "36. Palffy, A., Pool, E., Baratam, S., Kooij, J.F.P., Gavrila, D.M.: Multi-class road user detection with 3+1d radar in the view-of-delft dataset. IEEE Robotics and Automation Letters 7(2), 4961-4968 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.358, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "LEROjD: Lidar Extended Radar-Only Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "37. Palmer, P., Krueger, M., Altendorfer, R., Adam, G., Bertram, T.: Reviewing 3d object detectors in the context of high-resolution \\(3 + 1\\)d radar. In: Workshop on 3D Vision and Robotics at the Conference on Computer Vision and Pattern Recognition 2023 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.788, + 0.257 + ], + "angle": 0, + "content": "38. Palmer, P., Krueger, M., Altendorfer, R., Bertram, T.: Ego-motion estimation and dynamic motion separation from 3d point clouds for accumulating data and improving 3d object detection. In: AmE 2023 - Automotive meets Electronics; 14. GMM Symposium. pp. 86-91 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.257, + 0.788, + 0.298 + ], + "angle": 0, + "content": "39. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.338 + ], + "angle": 0, + "content": "40. Rebut, J., Ouaknine, A., Malik, W., Pérez, P.: Raw high-definition radar for multitask learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17021-17030 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.338, + 0.788, + 0.379 + ], + "angle": 0, + "content": "41. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10529–10538 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.42 + ], + "angle": 0, + "content": "42. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-779 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.42, + 0.788, + 0.46 + ], + "angle": 0, + "content": "43. Smith, L.N., Topin, N.: Super-convergence: Very fast training of neural networks using large learning rates. In: Artificial intelligence and machine learning for multidomain operations applications. vol. 11006, pp. 369-386. SPIE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.46, + 0.788, + 0.5 + ], + "angle": 0, + "content": "44. Stephan, M., Hazra, S., Santra, A., Weigel, R., Fischer, G.: People counting solution using an fmcw radar with knowledge distillation from camera data. In: 2021 IEEE Sensors. pp. 1-4 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.501, + 0.788, + 0.582 + ], + "angle": 0, + "content": "45. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.582, + 0.788, + 0.637 + ], + "angle": 0, + "content": "46. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.678 + ], + "angle": 0, + "content": "47. Sun, S., Zhang, Y.D.: 4d automotive radar sensing for autonomous vehicles: A sparsity-oriented approach. IEEE Journal of Selected Topics in Signal Processing 15(4), 879-891 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.678, + 0.788, + 0.718 + ], + "angle": 0, + "content": "48. Tan, B., Ma, Z., Zhu, X., Li, S., Zheng, L., Chen, S., Huang, L., Bai, J.: 3d object detection for multi-frame 4d automotive millimeter-wave radar point cloud. IEEE Sensors Journal (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.744 + ], + "angle": 0, + "content": "49. Tung, F., Mori, G.: Similarity-preserving knowledge distillation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1365-1374 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.744, + 0.788, + 0.799 + ], + "angle": 0, + "content": "50. Wang, L., Zhang, X., Li, J., Xv, B., Fu, R., Chen, H., Yang, L., Jin, D., Zhao, L.: Multi-modal and multi-scale fusion 3d object detection of 4d radar and lidar for autonomous driving. IEEE Transactions on Vehicular Technology 72(5), 5628-5641 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.84 + ], + "angle": 0, + "content": "51. Wang, T., Yuan, L., Zhang, X., Feng, J.: Distilling object detectors with fine-grained feature imitation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4933-4942 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "P. Palmer et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "52. Wei, Y., Wei, Z., Rao, Y., Li, J., Zhou, J., Lu, J.: Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. In: European Conference on Computer Vision. pp. 179-195. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "53. Wu, H., Wen, C., Shi, S., Li, X., Wang, C.: Virtual sparse convolution for multimodal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21653-21662 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "54. Xiao, A., Huang, J., Guan, D., Zhan, F., Lu, S.: Transfer learning from synthetic to real lidar point cloud for semantic segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2795-2803 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "55. Xiao, Y., Liu, Y., Luan, K., Cheng, Y., Chen, X., Lu, H.: Deep lidar-radar-visual fusion for object detection in urban environments. Remote Sensing 15(18) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "56. Xiong, W., Liu, J., Huang, T., Han, Q.L., Xia, Y., Zhu, B.: Lxl: Lidar excluded lean 3d object detection with 4d imaging radar and camera fusion. IEEE Transactions on Intelligent Vehicles pp. 1-14 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.397 + ], + "angle": 0, + "content": "57. Xu, B., Zhang, X., Wang, L., Hu, X., Li, Z., Pan, S., Li, J., Deng, Y.: Rpfa-net: a 4d radar pillar feature attention network for 3d object detection. In: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). pp. 3061-3066 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "58. Yan, Q., Wang, Y.: Mvfan: Multi-view feature assisted network for 4d radar object detection. In: International Conference on Neural Information Processing. pp. 493-511. Springer (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "59. Yang, J., Shi, S., Ding, R., Wang, Z., Qi, X.: Towards efficient 3d object detection with knowledge distillation. In: Advances in Neural Information Processing Systems. vol. 35, pp. 21300-21313. Curran Associates, Inc. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "60. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3dssd: Point-based 3d single stage object detector. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11040-11048 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "61. You, Y., Wang, Y., Chao, W.L., Garg, D., Pleiss, G., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-lidar++: Accurate depth for 3d object detection in autonomous driving. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "62. Zamanakos, G., Tsochatzidis, L., Amanatiadis, A., Pratikakis, I.: A comprehensive survey of lidar-based 3d object detection methods with deep learning for autonomous driving. Computers & Graphics 99, 153-181 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.66 + ], + "angle": 0, + "content": "63. Zhang, J., Zhuge, H., Liu, Y., Peng, G., Wu, Z., Zhang, H., Lyu, Q., Li, H., Zhao, C., Kircali, D., Mharolkar, S., Yang, X., Yi, S., Wang, Y., Wang, D.: Ntu4dradlm: 4d radar-centric multi-modal dataset for localization and mapping. In: 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.716 + ], + "angle": 0, + "content": "64. Zhang, L., Dong, R., Tai, H.S., Ma, K.: Pointdistiller: Structured knowledge distillation towards efficient and compact 3d detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21791-21801 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.758 + ], + "angle": 0, + "content": "65. Zhang, X., Wang, L., Chen, J., Fang, C., Yang, L., Song, Z., Yang, G., Wang, Y., Zhang, X., Yang, Q., Li, J.: Dual radar: A multi-modal dataset with dual 4d radar for autonomous driving. arXiv preprint arXiv:2310.07602 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.813 + ], + "angle": 0, + "content": "66. Zhang, Y., Hu, Q., Xu, G., Ma, Y., Wan, J., Guo, Y.: Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18953-18962 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "67. Zhao, M., Li, T., Abu Alsheikh, M., Tian, Y., Zhao, H., Torralba, A., Katabi, D.: Through-wall human pose estimation using radio signals. In: Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.358, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "LEROjD: Lidar Extended Radar-Only Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "the IEEE conference on computer vision and pattern recognition. pp. 7356-7365 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.787, + 0.231 + ], + "angle": 0, + "content": "68. Zheng, L., Li, S., Tan, B., Yang, L., Chen, S., Huang, L., Bai, J., Zhu, X., Ma, Z.: Rcfusion: Fusing 4-d radar and camera with bird's-eye view features for 3-d object detection. IEEE Transactions on Instrumentation and Measurement 72, 1-14 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.286 + ], + "angle": 0, + "content": "69. Zheng, L., Ma, Z., Zhu, X., Tan, B., Li, S., Long, K., Sun, W., Chen, S., Zhang, L., Wan, M., Huang, L., Bai, J.: Tj4dradset: A 4d radar dataset for autonomous driving. In: 2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC). pp. 493-498 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.288, + 0.787, + 0.327 + ], + "angle": 0, + "content": "70. Zhou, T., Chen, J., Shi, Y., Jiang, K., Yang, M., Yang, D.: Bridging the view disparity between radar and camera features for multi-modal fusion 3d object detection. IEEE Transactions on Intelligent Vehicles 8(2), 1523-1535 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "71. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4490-4499 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.37 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/LEROjD_ Lidar Extended Radar-Only Object Detection/e0f1cd56-0b49-499c-b5fc-daef469190ce_origin.pdf b/2024/LEROjD_ Lidar Extended Radar-Only Object Detection/e0f1cd56-0b49-499c-b5fc-daef469190ce_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d3cb98c651a6df7ea231ed4b9fe0b42e9f4a3e5d --- /dev/null +++ b/2024/LEROjD_ Lidar Extended Radar-Only Object Detection/e0f1cd56-0b49-499c-b5fc-daef469190ce_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7ca0941dcb6c3895eee2536307d296a97d4ce7725e26882501ab99bba7ada49 +size 1782316 diff --git a/2024/LEROjD_ Lidar Extended Radar-Only Object Detection/full.md b/2024/LEROjD_ Lidar Extended Radar-Only Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..aa12d031097c4149ffe0ad80bb6063ab07ba7d33 --- /dev/null +++ b/2024/LEROjD_ Lidar Extended Radar-Only Object Detection/full.md @@ -0,0 +1,289 @@ +# LEROjD: Lidar Extended Radar-Only Object Detection + +Patrick Palmer1, Martin Krüger1, Stefan Schütte1, Richard Altendorfer2, Ganesh Adam2, and Torsten Bertram1 + +$^{1}$ Institute of Control Theory and Systems Engineering, TU Dortmund University {patrick.palmer, martin2.krueger, stefan.schuette, torsten.bertram}@tu-dortmund.de + +$^{2}$ ZF Group, {richard.altendorfer, ganeshadam}@zf.com + +Abstract. Accurate 3D object detection is vital for automated driving. While lidar sensors are well suited for this task, they are expensive and have limitations in adverse weather conditions. $3 + 1\mathrm{D}$ imaging radar sensors offer a cost-effective, robust alternative but face challenges due to their low resolution and high measurement noise. Existing $3 + 1\mathrm{D}$ imaging radar datasets include radar and lidar data, enabling cross-modal model improvements. Although lidar should not be used during inference, it can aid the training of radar-only object detectors. We explore two strategies to transfer knowledge from the lidar to the radar domain and radar-only object detectors: 1. multi-stage training with sequential lidar point cloud thin-out, and 2. cross-modal knowledge distillation. In the multi-stage process, three thin-out methods are examined. Our results show significant performance gains of up to 4.2 percentage points in mean Average Precision with multi-stage training and up to 3.9 percentage points with knowledge distillation by initializing the student with the teacher's weights. The main benefit of these approaches is their applicability to other 3D object detection networks without altering their architecture, as we show by analyzing it on two different object detectors. Our code is available at https://github.com/rst-tu-dortmund/lerojd. + +Keywords: 3D Object Detection $\cdot$ 3+1D Imaging Radar $\cdot$ Cross-Modal Object Detection + +# 1 Introduction + +Environment perception is the first module in each automated driving stack. Multiple sensor modalities, like cameras, lidars, and radars are utilized for this task. Radar sensors are of unique interest in perception due to their robustness against poor lighting, challenging weather conditions like rain or snow, and cost-effectiveness [3, 36, 47]. One exclusive advantage is the ability to measure the relative radial velocity of reflections directly due to the Doppler effect. + +While a precise localization of objects is possible with traditional radar sensors without elevation angle measurements [9], it is inherently limited to the horizontal plane. Additionally, it is hard to predict the objects' extent due to + +![](images/7d1e10ae07ead6a221ae33a784919070263ff3eb3ad9df5150d61440c4f24bd0.jpg) +(a) + +![](images/1ab6dd17d4634553e796d6f4a621a3c1ef026a3fc8aec8b8d1ca16ce5311a81e.jpg) +(b) +Fig. 1: Architecture overview of (a) a knowledge distillation-based method and (b) a multi-stage training method (MSTM) for utilizing lidar data in the training of radar-only object detectors. The ground truth (GT) label is the same for both methods. The diagrams above the dotted line represent the training process, while the diagrams below the dotted line represent inference. + +the reflections' sparsity. The introduction of $3 + 1\mathrm{D}$ high-resolution imaging radar sensors has recently mitigated these limitations, at least partially. In addition to measuring the elevation angle of reflections, the density of measurements is increased [14, 20]. Therefore, approaches that only utilize radar sensors for perception of the environment are of particular interest. + +Despite improvements in radar-based object detection, the performance still lags behind other sensor modalities like lidar [36]. One persistent major limitation of 2+1D classic radar and 3+1D imaging radar sensors is the relative sparsity of the point cloud, which limits the detection performance. + +Lidar sensors, on the other hand, are well suited for object detection and are therefore frequently employed as a reference for evaluating the performance of different sensor modalities due to their ability to produce an accurate and dense understanding of the scene. Their effectiveness is particularly pronounced in detecting nearby traffic participants without occlusion [53]. + +All currently available datasets that contain $3 + 1\mathrm{D}$ imaging radar data additionally accommodate lidar sensor data [4,8,32,35,36,40,63,65,69]. The lidar sensor data of these datasets is currently either utilized for labeling, combining multiple sensor modalities for accurate object detection, or comparing the performance of radar-only techniques to another sensor modality. While the majority of series production vehicles may not include lidar sensors due to cost and vehicle design constraints, they will still be available in the training process of learning-based methods. Extending radar-only methods with lidar sensor data in training has been shown to be a viable method for estimating point flow on the imaging radar point cloud [12]. These observations lead to the following research question: Can lidar sensor data be utilized in the training process of imaging radar-based 3D object detectors to improve the object detection performance on radar-only data during inference? + +To use different sensor modalities in training, transfer learning and knowledge distillation (KD) principles can be utilized. While KD is commonly employed across sensor modalities such as camera images and lidar point clouds, its application between $3 + 1\mathrm{D}$ imaging radar and lidar sensors remains unexplored. Since lidar and imaging radar share a structurally similar data representation as point clouds, an identical base network with different input modalities can be utilized to transfer knowledge between different input modalities. We investigate two approaches to transfer knowledge from lidar-based to radar-based object detectors: a KD-based and a multi-stage training approach. The main principles of the two methods are visualized in Figure 1. + +The contributions of this work are summarized as follows: + +- We investigate a combination of lidar and radar sensors in the training stage of object detectors to improve radar-only object detection at inference. +- We investigate three thin-out strategies for lidar point clouds to transfer knowledge from dense lidar to sparse lidar and radar-only object detectors. +- We propose a multi-stage training procedure to transfer knowledge from dense lidar to sparse lidar and, finally, to radar-only object detection. +- We modify and analyze several knowledge distillation-based approaches to transfer knowledge from lidar to radar-only object detectors. + +# 2 Related Work + +Imaging radar sensors commonly utilize the point cloud as a data representation format instead of the radar tensor, due to its higher computational efficiency. This format is similar to the one used for lidars, enabling the application of object detection methods developed for lidars to radars. Object detection on lidars can be split into two main categories. Point-based methods, like PointNet++ [39], downsample the original point cloud, encode it using a backbone, and finally apply a detection head. Voxel-based methods discretize the point cloud into a 3D grid and apply 3D convolutions to the grid before finally applying a detection head [6,10,71]. The main drawback of voxel-based methods is the high memory consumption and the loss of spatial information due to the discretization of the point cloud. To overcome the limitation of high memory consumption, [23,27] have proposed the PointPillars network. Pillars are a specific form of voxels that span over the full height of the scene and represent the point cloud in a 2D grid. + +Methods from the lidar domain have been shown to perform reasonably well on radar data [36,37]. However, they are limited due to the sparsity of the radar point cloud. SMURF [30] considers two representations of the radar point cloud to address sparsity. Using kernel density estimation, it utilizes pillarization and density features derived from a multi-dimensional Gaussian mixture distribution. RPFA-Net [57] is a PointPillars [23] based network, which introduces a selfattention mechanism to extract global context information from the radar point cloud. RadarMFNet [48] utilizes a multi-frame radar point cloud representation to address the sparsity of the radar point cloud in conjunction with an anchor-based detector and temporal pooling layers. + +One way of improving 3D object detection on radar data is the fusion with additional sensor modalities like camera [56, 68, 70], lidar [50] or camera and lidar [5, 13, 55] at the cost of introducing additional sensors at inference. + +The concept of knowledge distillation was first introduced by [18]. It consists of two networks, which are labeled as teacher and student. The teacher is a large and complex model, while the student is less complex and more computationally efficient. The student is trained to mimic the teacher network's performance by utilizing the teacher's predictions and the ground truth labels. This has been utilized in the context of 3D object detection by [29,59,64] to construct computation time-efficient models that have similar performance as large models. [26] and [22] have extended upon this concept by utilizing KD to extract knowledge from a teacher trained on lidar data to a student who utilizes camera images. + +To the best of our knowledge, few studies have investigated the effect of KD considering radar data. A transfer of knowledge from an image-based teacher network to a radar-based student has been shown to improve the performance for the task of people counting [44]. For the task of 3D object detection, [22] has shown that a transfer learning-based approach from a lidar and image-based network to the radar domain results in improved object detection performance for classical $2 + 1\mathrm{D}$ radar sensors. The main drawback of [22] is that in addition to 3D object labels, instance segmentation labels for the image domain are required. Additionally, [22] introduces multiple sub-networks to derive the KD losses, which makes the model more complex, while we aim for a simpler approach. HiddenGems [12] utilizes lidar point clouds to derive point flow information and train a network to predict the point flow on radar point clouds. See Beyond Seeing [11] utilizes lidar point cloud for feature hallucination on radar point clouds. Both methods require the lidar point cloud only in the training process but require major modifications to the network architecture. + +Sampling of point clouds is a broadly explored topic to reduce the complexity of computing large point clouds. Most commonly, farthest point sampling (FPS) [39,42,60], voxelized FPS [41] or random sampling [71] are utilized in the context of 3D object detection. Random sampling has been shown to be beneficial when used for semantic segmentation of point clouds [19] and to outperform farthest point sampling at this task [25]. One limitation of random sampling is that while points at a close distance are kept, points at a far distance are more likely to be removed. Therefore, [7] proposes a polar cylinder balanced random sampling to keep a more balanced distribution of points across the range. + +# 3 Method + +Two methods of transferring knowledge from lidar-based to radar-based object detectors are investigated: Knowledge Distillation (KD) (Section 3.1), which is modified for the task of transfer learning and our proposed multi-stage training procedure with sequential point cloud thin-out (Section 3.2). Additionally, the utilized thin-out strategies for lidar point cloud (Section 3.3) are described. + +# 3.1 Knowledge Distillation + +KD is commonly used for two tasks. First, designing computationally efficient models by transferring knowledge from a larger teacher network to a smaller student network [18, 49, 51]. Second, to transfer knowledge across sensor modalities by utilizing different model architectures for the teacher and student [1, 16, 67]. Lidar and radar-based point cloud object detection can utilize the same model structure but with different input modalities. This enables the utilization of KD methods first described for designing computationally efficient models for cross-modal knowledge transfer from lidar- to radar-based object detectors. In this case, the teacher is trained on the full lidar point cloud, while the student is trained on the radar point cloud. Three different loss terms, as described by [59], are employed: + +Logit-KD is the first, classic type of distilling knowledge described by [18]. For 3D object detection, the logit-KD loss $\mathcal{L}_{\mathrm{logit}}$ is split into a classification $\mathcal{L}_{\mathrm{l -cls}}$ and regression loss $\mathcal{L}_{\mathrm{l - reg}}$ . These losses are calculated by comparing the student's and teacher's predictions utilizing the 3d detectors' regression loss and bi-linear interpolation between student and teacher output classes. + +Feature-KD is widely utilized in 2D object detection [28, 51]. It utilizes a loss term that forces the student network to mimic the teacher's intermediate feature map (feat). A feature mimicking the last layer of the bird's eye view feature encoder, similar to [59], is utilized in this work. + +Label-KD is a recent distillation approach that leans on the concept of the Logit KD but simplifies and generalizes it. It is first described by [33]. The teacher predictions are filtered by their scores using a score threshold, and an adapted ground truth set is constructed by combining the filtered predictions and the ground truth set. This adapted set is utilized in student training. The loss is split into a classification $\mathcal{L}_{\mathrm{cls}}$ and regression loss $\mathcal{L}_{\mathrm{reg}}$ . It replaces $\mathcal{L}_{\mathrm{label}}$ usually calculated on the ground truth set. + +The three KD loss terms are combined into a joint loss weighted with $\lambda$ + +$$ +\mathcal {L} _ {\text {j o i n t}} = \underbrace {\lambda_ {\mathrm {l - r e g}} \mathcal {L} _ {\mathrm {l - r e g}} + \lambda_ {\mathrm {l - c l s}} \mathcal {L} _ {\mathrm {l - c l s .}}} _ {\mathcal {L} _ {\text {l o g i t}}} + \underbrace {\lambda_ {\text {f e a t}} \mathcal {L} _ {\text {f e a t}}} _ {\mathcal {L} _ {\text {f e a t u r e}}} + \underbrace {\lambda_ {\text {r e g}} \mathcal {L} _ {\text {r e g}} + \lambda_ {\text {c l s}} \mathcal {L} _ {\text {c l s}}} _ {\mathcal {L} _ {\text {l a b e l}} ^ {*}} \tag {1} +$$ + +# 3.2 Multi-Stage Lidar Thin-Out Training Procedure + +Utilizing a different input data modality for either pre-training a network on a large dataset or utilizing simulated data in the training process of a point cloud-based network has been shown to improve the object detection performance [54]. The multi-stage training method (MSTM) proposed in this work extends upon this by utilizing a Curriculum learning [2] based training procedure by which the network is trained on iteratively sparsified lidar point clouds, similar to [52], and fine-tuned on the radar point cloud. Figure 2 visualizes our multi-stage training procedure. The network is first trained on the full lidar point cloud until convergence. In the following steps, the lidar point cloud is thinned-out iteratively by a factor of 2 and utilized for training a network whose weights are + +![](images/4c1e883322e8a05bc2812e572cfa34f1d9933ec98787b0de36e580ef8d1e9d00.jpg) +Fig. 2: MSTM pipeline. A 3D object detection network is iteratively trained on increasingly sparse lidar data. Three different thin-out strategies are utilized for sparsification. The lidar point cloud is mixed with the radar point cloud in the second to last step. In the last step, the network is only trained on radar data. At inference, only the orange-shaded part is executed. Small points represent lidar points, large points radar points. The color of points corresponds to the distance from the ego-vehicle. The camera image is not used as an input but only aids the visualization. + +initialized using the previously trained model. This forces the network to learn features for a good object detection performance on increasingly sparser point clouds. In the second to last step, the lidar point cloud is mixed with the radar points so that the network can translate from the lidar to the radar domain. In the last step, the model is trained only on radar points. A training without multiple stages is called single-stage training method (SSTM) in this work. + +In addition to the training where only lidar points are utilized in the first training stages, we investigate the utilization of the radar point cloud in conjunction with the lidar point cloud in all stages. The thin-out of the lidar points remains the same and is mixed with the radar point cloud in each step. This conditions the model on radar from the first step in order to prioritize features in the lidar point cloud that relate to a good object detection on radar-only data. + +For mixing lidar and radar point clouds, the voxel or pillar feature encoder is modified to prioritize the radar point cloud in the random sampling process as done by [34]. Otherwise, the vastly higher number of lidar points, even when thinned-out, might lead to the complete exclusion of radar points. + +# 3.3 Lidar Thin-Out Strategies + +Three different methods for sub-sampling the lidar point cloud are investigated. Figure 2 shows examples for each thin-out stage. + +Random sampling is the simplest method of sparsifying a point cloud. It neglects the structure and inherent limitations of the point cloud representation, especially for objects far away or with a high degree of occlusion. Neglecting the structure can lead to the complete loss of information for objects represented only by few points. + +K-nearest neighbor sampling approximates the reflection density distribution of the radar point cloud with the lidar point cloud by only keeping lidar reflections close to radar reflections. This algorithm is described in Algorithm 1. Objects not detected by the radar sensor are, therefore, also not represented by the k-nearest neighbor thinned-out lidar point cloud. + +Voxel-based sampling aims to reduce the number of points in each area of the point cloud while keeping the general distribution of the point cloud. This is motivated by the fact that radar sensors do not suffer as much from loss of resolution with distance as lidar sensors. The approach is described in Algorithm 2, and executed iteratively for a sequence of sparsification steps. + +Algorithm 1: K-nearest neighbor sampling +Input : Lidar point cloud $L\in \mathbb{R}^{N\times 3}$ , radar point cloud $R\in \mathbb{R}^{M\times 3}$ Output: Sub-sampled lidar point cloud $L_{s}\in \mathbb{R}^{K\times 3}$ +1 Calculate the euclidean distance $c_{i}$ between each lidar point $l_{i}\in L$ and its nearest radar point from $R$ . +2 Select a share of $K$ lidar points with the smallest distances c; +3 return Share of Points $L_{s}\in \mathbb{R}^{K\times 3}$ +Algorithm 2: Voxel-based sampling. +Input : Lidar point cloud $L\in \mathbb{R}^{N\times 3}$ Output: Sub-sampled lidar point cloud $L_{s}\in \mathbb{R}^{K\times 3}$ +1 Initialize $L_{s}$ with $L$ . +2 Voxelize the lidar point cloud into v voxels; +3 Calculate the number of points in each voxel $P_{v}$ . +4 Calculate the minimum number of points per voxel $p_{\mathrm{min}}$ , so that at least $0.75N$ points are in voxels with more than $p_{\mathrm{min}}$ points; +5 From each voxel with more than $p_{\mathrm{min}}$ points, choose $p_{\mathrm{min}}$ random points to keep and add remaining points to $L_{p}$ . +6 Randomly select O.5 N points from $L_{p}$ and remove them from $L_{s}$ . +7 return Share of Points $L_{s}\in \mathbb{R}^{K\times 3}$ + +Another common sparsification method utilized for imitating low-resolution lidar sensors is layer-based sampling [52, 61]. This approach is not investigated because radar sensors do not capture the environment on a layer basis. + +# 4 Experimental Evaluation + +# 4.1 Experimental Setup + +Dataset: All experiments are conducted utilizing the View-of-Delft (VoD) dataset [36]. It contains synchronized data of multiple sensor modalities. The 64-layer lidar sensor and the imaging radar are utilized in this work. A point cloud accumulated over 5 frames [36], which has been shown to improve object detection performance compared to no accumulation, is used for radar data [36, 38]. We detected a duplication of identical points in the lidar point cloud, which could adversely affect all sampling methods; thus, we eliminated the duplicated points from the point cloud. Although the VoD dataset is among the best currently available datasets for imaging radar-based object detection, it is limited by its size compared to other automotive datasets without imaging-radar, like [31, 45]. Given the absence of publicly available labels and limited online evaluation for the test dataset, we repurpose the validation dataset as a test set. Consequently, we partition the original training set into a new training set $(80\%)$ and a dedicated validation set $(20\%)$ to ensure robust model training. + +Evaluation Metrics: The primary performance metric utilized to compare the results is the mean average precision (mAP), as used by [15] [37]. Similar to the evaluation of the Waymo data set [46] in [62], we split the results into two distance bins: short-range (SR): $0 - 30\mathrm{m}$ and mid-range (MR): $30 - 50\mathrm{m}$ . All experiments are conducted utilizing three different random network initializations that are averaged. + +Training: Most experiments use the PointPillars model [23] as an object detector with the same configuration as utilized by [36]. For imaging radar data, PointPillars has been shown to perform among the best out of multiple state-of-the-art 3D object detection methods while still performing adequately on lidar data [37]. Furthermore PointPillars is a relevant baseline for radar-specific object detection methods [57,58]. To show that the proposed MSTM and KD apply to various object detectors, the most promising approaches from the evaluations on PointPillars are evaluated on DSVT-P [17], as an example for a transformer-based model. All SSTM trainings are conducted with an early stopping policy for a maximum of 125 epochs. For the MSTM the initial training on the full lidar point cloud is conducted for 125 epochs, while each refinement step is trained for 30 epochs. All trainings utilize the Adam optimizer [21] and an adapted learning rate scheduler that reaches its maximum learning rate earlier and has a faster descent than the scheduler described by [43]. This improves the object detection performance on radar data. + +Notation: To distinguish between methods, the following notation is used: + +$$ +\mathcal {T} _ {\mathcal {L S} / \mathcal {T O}} ^ {\mathcal {T M}} \rightarrow \mathcal {T} ^ {\mathcal {K D}}. +$$ + +The notation is split into two parts. The part left of the arrow represents the data set utilized for pre-training, while the right part represents the data and training method utilized for the last (fine-tuning) training stage of the model. + +Table 1: Examples of training steps and evaluation data corresponding to the notation, emplying voxel-based sub-sampling. Voxel sampling is shortened to v for legibility. + +
NotationTrainingEvaluation
L1-1/8/vL1/v → L1/2/v → L1/4/v → L1/8/vL1/8/v
RLMSTM1-1/8/vRL1/v → RL1/2/v → RL1/4/v → RL1/8/vRL1/8/v
L1MSTM1-1/16/v → RL1/v → L1/2/v → L1/4/v → L1/8/v → RL1/16/v → RR
RLMSTM1-1/16/v → RRL1/v → RL1/2/v → RL1/4/v → RL1/8/v → RL1/16/v → RR
RLSSTM1/4/v → RRL1/4/v → RR
+ +This part is omitted if it is the same as the training dataset. The training data $\mathcal{T}$ can either be lidar (L), radar (R), or mixed radar + lidar data (RL). The training method $\mathcal{T}\mathcal{M}$ can denote either MSTM or SSTM. The lidar share $\mathcal{L}\mathcal{S}$ is the fraction of the original lidar point cloud that is utilized. Radar-only training always uses the full radar point cloud, therefore $\mathcal{L}\mathcal{S}$ is omitted. For the MSTM, this is represented by a range of fractions that are iterated in training. The thin-out method $\mathcal{T}\mathcal{O}$ can either be random (rand), k-nearest neighbor (knn) or voxel-based (vox). $\mathcal{KD}$ represents the KD method; this can either be label (lab), logit (log), feature (feat), or a combination of those (joint). $\mathcal{KD}$ is omitted if just an initialization and fine-tuning is utilized. Examples for the training corresponding to specific notations are listed in Table 1. + +# 4.2 Evaluation of MSTM on Lidar-only and Mixed Radar + Lidar + +To evaluate the applicability of our proposed MSTM to the thinned-out lidar point cloud, we evaluate the training without the last two steps involving radar data. Training is conducted for just the lidar point cloud and the mixed radar + lidar point cloud in all stages. The multi-stage trained network is evaluated after each thin-out stage on the thinned-out lidar (or mixed radar + lidar) point cloud. Thin-out stages up to $1/256$ of the original lidar point cloud are considered due to the lidar point cloud containing fewer points than the radar point cloud at $1/256$ of the original lidar point cloud. The MSTM is compared to the SSTM, which is trained only on the thinned-out point cloud. The results are shown in Fig. 3, complete quantitative results are given in the supplementary. + +All thin-out strategies result in an increasing degradation of the detection performance. This is most pronounced when considering a random thin-out, which drops approximately linearly with each thin-out step. K-nearest neighbor and voxel-based sampling consistently perform better than random sampling due to keeping a higher point density in areas around objects, which supports object detection. For k-nearest neighbor sampling, the performance only drops by 0.6 percentage points between $\mathrm{L}_{1 / \mathrm{knn}}^{\mathrm{SSTM}}$ and $\mathrm{L}_{1 / 2 / \mathrm{knn}}^{\mathrm{SSTM}}$ due to mostly ground points being removed in the first thin-out stage. For voxel-based sampling, after a sharp performance drop in the first thin-out stage, only a relatively slight performance drop is observable until the $^{1/16}$ thin-out stage. In the first stages, + +![](images/9ea428973687d550449b45cf72c04b8e0f264792b40cbc1352e6158db55bb224.jpg) +Fig. 3: Comparison between SSTM and MSTM for different lidar sampling strategies. + +the performance drop can mainly be attributed to the pedestrian class, which drops by 12.8 percentage points in the first thin-out stage. In comparison, the detection performance of the car and cyclist classes only decreased by 0.7 and 3 percentage points, respectively. This can be explained by the voxel size of $1\mathrm{m}$ in each dimension utilized in the voxel-based sampling. A single pedestrian only occupies a small number of these voxels. When the point cloud gets thinned out, the entire object's structure gets lost, making detection difficult. In contrast, the object's structure and physical appearance can be adequately represented when it occupies more voxels, as observed in the car and cyclist classes. In the final thin-out stage of $\mathrm{L}^{\mathrm{SSTM}}$ and $\mathrm{L}^{\mathrm{MSTM}}$ , voxel-based sampling performs worse than k-nearest neighbor sampling because many voxels only consist of ground points or points of surrounding background objects. + +The multi-stage training does not result in a useful knowledge transfer and, therefore, significant performance benefit for all considered thin-out strategies. + +A contrary behavior is observed on the combined radar + lidar point cloud. The MSTM consistently outperforms the SSTM when using k-nearest neighbor or voxel-based sampling. Knowledge from the dense point cloud can be transferred to the thin radar + lidar point cloud. Additionally, performance is consistently higher than training just on lidar, especially for voxel-based subsampling, which performs best at lower thin-out stages. At small thin-out stages, the voxel-based sub-sampling can still represent the whole object space and give meaningful environmental information. At the same time, the radar point cloud is sufficiently dense for object detection. + +# 4.3 Evaluation of MSTM with Last Radar-Only Step + +This chapter analyses the performance of the MSTM when applied to radar data, as described in Section 3.2. The MSTM is evaluated for two different procedures. Utilizing only the lidar point cloud and utilizing the mixed radar + lidar point cloud in the first stages. The results of MSTM are shown in Table 2. The lidar thin-out stages up to $1/16$ of the original lidar point cloud are considered due to the performance of $\mathrm{L}_{1-1/32}^{\mathrm{MSTM}}$ dropping below $\mathrm{R}^{\mathrm{SSTM}}$ . + +Table 2: 3D object detection results on the VoD radar dataset trained using MSTM. All MSTM methods are trained on radar-only data in the last step. The best and second best results are marked in bold and underlined, respectively. + +
Training methodmAPCarPedestrianCyclist
SRMRSRMRSRMRSRMR
RSSTM (baseline)36.711.945.218.117.17.447.710.2
L1-1/16 / rand → R36.416.144.920.017.59.846.818.9
L1-1/16 / knn → R35.714.045.623.216.77.144.811.7
L1-1/16 / vox → R34.114.347.023.015.39.140.010.7
RLMSTM 1-1/16 / rand → R35.614.944.120.217.87.244.917.3
RLMSTM 1-1/16 / knn → R38.214.745.523.918.88.250.311.9
RLMSTM 1-1/16 / vox → R39.715.445.922.518.49.754.713.9
+ +When considering the pre-training using just lidar data, it is observable that the overall performance drops in the SR bin. In contrast, the MR bin is improved for all thin-out methods. The best performance is achieved by utilizing the random thin-out strategy. It performs especially well for pedestrians and cyclists in the mid-range, due to the comparably small size of pedestrians and cyclists resulting in a worse radar reflection characteristic. In contrast these objects are detected well by the lidar sensor. Knowledge about the representation of objects can be transferred from the lidar point cloud to the radar point cloud. K-nearest neighbor and voxel-based thin-out do not improve the performance in the SR bin for the pedestrian and cyclist class but perform better in the short and mid-range for the car class which is explained by the large size of cars, resulting in a better representation of cars in the thin point cloud. + +When considering the pre-training with the mixed radar + lidar point cloud, a contrary observation is made compared to only using the lidar in pre-training. The overall performance for the random thinned-out lidar point cloud is worse with the mixed point cloud than when only considering the lidar point cloud. One consistent aspect is the good performance of the cyclists in the mid-range, which still surpasses all other methods, excluding $\mathrm{L}_{1 - 1/16}^{\mathrm{MSTM}}$ . However, the k-nearest neighbor and voxel-based thin-out strategies perform better with the mixed point cloud. The voxel-based thin-out strategy achieves the best performance. It performs best on objects in SR, mainly due to its outstanding performance in detecting cyclists, but is limited in the mid-range, getting surpassed by the MSTM with random thin-out. The MSTM with voxel-based thin-out can increase the object detection performance on the radar point cloud by 3 percentage points in the SR and 3.5 percentage points in the MR. + +# 4.4 Evaluation of Cross-Modality KD + +The configuration of the teacher's training data is of particular interest, as the teacher's performance directly influences the student's. The simplest choice is + +Table 3: 3D object detection results of the teacher with different training sets. Each teacher is evaluated on the test set of the same data configuration as the training set. The best and second best results are marked in bold and underlined, respectively. + +
DatamAPCarPedestrianCyclist
SRMRSRMRSRMRSRMR
L1SSTM56.530.060.742.440.119.568.528.1
RLSSTM61.645.161.946.544.434.078.554.8
RLSSTM1/4 / rand56.931.157.232.437.221.376.439.5
RLSSTM1/4 / knn58.936.954.431.943.030.979.547.9
RLSSTM1/4 / Vox56.139.360.141.632.828.875.547.5
+ +to train the teacher solely on the lidar point cloud. Section 4.2 and Section 4.3 show that mixing the radar and lidar point clouds can benefit radar-only object detection. Therefore, teachers trained on a mixed point clouds, in addition to the ones solely trained on lidar, are compared. Mixed point clouds containing $1/4$ of the original lidar point cloud are considered, as a closer representation of the teachers to the student's data can lead to a better performance. Thin out of $1/4$ is chosen as the radar + lidar detection performance with radar and $1/8$ of the lidar points is worse than the SSTM on only lidar data. Table 3 shows the performance of all utilized teachers. + +Table 3 shows the results of the teacher network for different training data configurations. It can be observed that, as expected, the mixed radar and lidar point cloud performs the best, only being surpassed by $\mathrm{RL}_{1/4}^{\mathrm{SSTM}}$ on the cyclist class in short range. No model can be considered the overall second best among the other training sets. The performance varies between vehicle classes. + +The KD is evaluated individually for each KD method and teacher training set. A joint KD is also considered comprised of all three KD losses. All student networks are initialized (Init) with the weights of the teacher network, as it has been shown to improve the student's performance [59]. Additionally, for the transfer learning between datasets, the pre-training utilizing MSTM in Section 4.3 has been shown to improve the performance on the radar dataset. + +The results utilizing the KD are shown in Table 4. Just the initialization of the student network with the teacher's weights already results in a performance benefit in the SR with $\mathrm{RL}_{1}^{\mathrm{SSTM}}$ and $\mathrm{RL}_{1/4}^{\mathrm{SSTM}}$ as teachers and in the MR with all teachers. This is overall only surpassed by the $\mathrm{RL}_{1-1/16}^{\mathrm{MSTM}} / \mathrm{vox} \to \mathrm{R}$ showing that the MSTM is substantially better than a simple initialization. + +For label-KD, the best-performing models are the ones where the teacher performs the best. Specifically $\mathrm{L}_1^{\mathrm{SSTM}}\rightarrow \mathrm{R}^{\mathrm{lab}}$ , which is only surpassed by $\mathrm{RL}_{1/4/knn}^{\mathrm{SSTM}}\rightarrow \mathrm{R}^{\mathrm{lab}}$ in the MR. For worse-performing teachers, using label-KD loss does not result in a performance benefit due to it replacing the ground truth label loss. Feature-KD requires a teacher who learns features from radar data. This is observable in the SR performance of models, where the teacher is trained on mixed radar + lidar data. Logit-KD works well on teacher datasets + +Table 4: 3D object detection results on the VoD radar dataset utilizing the different KDs and teachers trained on different data configurations. Only the mAP over all classes is specified. For each distillation method and range bin, the best and second best results are marked in bold and underlined, respectively. The best result for each teacher are marked in cyan for the SR bin and magenta for the MR bin. + +
Teacher dataInit only (→R)Logit-KD (→Rlog)Feature-KD (→Rfeat)Label-KD (→Rlab)Joint-KD (→RJoint)
SRMRSRMRSRMRSRMRSRMR
R1SSTM36.711.9--------
L1SSTM34.913.532.812.634.813.636.613.736.013.5
RLSSTM39.014.839.213.539.113.838.813.738.912.7
RLSSTM1/4 / rand39.014.238.911.639.411.834.012.236.412.3
RLSSTM1/4 / knn34.814.637.214.736.613.033.810.932.912.9
RLSSTM1/4 / Vox35.215.335.613.337.115.835.314.332.99.7
+ +which closely resemble the radar point cloud and perform well on the teacher set. This results in good performance of $\mathrm{RL}_1^{\mathrm{SSTM}}\rightarrow \mathrm{R}^{\log}$ . Besides the KD method the thin-out method utilized in the teacher's training influences the student's performance. Random sampling results in good performance in the SR, while voxel-based sampling results in good MR performance. This is somewhat contrary to what has been observed in the MSTM, where random sampling performs the best in the MR and can be explained by the different thin-out stages utilized in the MSTM and KD. Further qualitative results, as well as detailed quantitative results are given in the supplementary. Overall, initializing the student with the teacher's parameters yields good performance, with further enhancements primarily achievable through feature-KD with a teacher trained on mixed radar + lidar data. + +# 4.5 Evaluation on DSVT as a Transformer-Based Object Detector + +Table 5 shows the results of the MSTM and KD on DSVT-P [17] for selected methods. Similar effects, as observed for PointPillars, are seen on DSVT-P. Initializing the student with the teacher's weights and the MSTM led to a performance benefit. However, contrary to PointPillars, the KD approach does not contribute to any improvements for DSVT-P. + +# 4.6 Limitations + +The best-performing methods in this work apply only to detectors that share an architecture with the target model, as a direct transfer of weights is performed for the best performance. Different models may require different levels and steps in the thinning process. These choices are additional tuning parameters that must be selected appropriately to maximize the benefit of transferring knowledge. + +Table 5: Radar-only detection performance utilizing DSVT-P as an object detector. The best and second best results are marked in bold and underlined, respectively. + +
Training methodmAPCarPedestrianCyclist
SRMRSRMRSRMRSRMR
RSSTM38.313.042.520.221.912.350.56.6
RLMSTM1-1/16 / vox→R41.515.647.823.624.511.052.312.4
RLSSTMM1/4 / vox→R38.813.244.918.823.311.248.29.7
RLSSTMM1/4 / vox→Rfeat34.711.344.08.820.52.839.712.7
+ +# 5 Conclusion + +In this paper, we investigated two methods to transfer knowledge from lidar-based object detectors to radar-only object detection. First, MSTM with sequential sub-sampling of the lidar point cloud, and second, a KD-based approach. For the MSTM, we have investigated three thin-out strategies for the lidar point cloud. These thin-out strategies are also analyzed for the training of the KD teacher network. Both methods can substantially improve the object detection performance of a radar-only object detector. The MSTM with voxel-based thinout performs the best overall and can improve detection performance by up to 3.5 percentage points. For the KD methods, it is shown that initializing the student with the teacher's parameters, especially a teacher trained on mixed lidar and radar data, can improve the object detection performance on radar-only data, with further enhancement primarily achieved by utilizing feature-KD. + +In future work, the applicability to further 3D object detection networks and the behavior with more advanced knowledge distillation like the ones utilized by [22] could be investigated. Due to different effects being noticed with the MSTM and the KD methods, combining both methods could be investigated by utilizing the MSTM as a teacher. To overcome the limitations of choosing a strict thin-out strategy, a learnable point cloud thin-out method [24,66] can be used. + +# References + +1. Afouras, T., Chung, J.S., Zisserman, A.: Asr is all you need: Cross-modal distillation for lip reading. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 2143-2147. IEEE (2020) +2. Bengio, Y., Louradour, J., Collobert, R., Weston, J.: Curriculum learning. In: Proceedings of the 26th annual international conference on machine learning. pp. 41-48 (2009) +3. Briske, S., Ruf, F., Höhne, F.: Recent evolution of automotive imaging radar and its information content. IET Radar, Sonar & Navigation 12(10), 1078-1081 (2018) +4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous + +driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020) +5. Chen, X., Zhang, T., Wang, Y., Wang, Y., Zhao, H.: Futr3d: A unified sensor fusion framework for 3d detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 172-181 (2023) +6. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21674-21683 (June 2023) +7. Cheng, H., Han, X., Jiang, H., He, D., Xiao, G.: Pcb-random sampling for lidar semantic segmentation in autonomous driving scene. arXiv preprint arXiv:2209.13797 (2022) +8. Choi, M., Yang, S., Han, S., Lee, Y., Lee, M., Choi, K.H., Kim, K.S.: Msc-rad4r: Ros-based automotive dataset with 4d radar. IEEE Robotics and Automation Letters 8(11), 7194-7201 (2023) +9. Danzer, A., Griebel, T., Bach, M., Dietmayer, K.: 2d car detection in radar data with pointnets. In: 2019 IEEE Intelligent Transportation Systems Conference (ITSC). pp. 61-66 (2019) +10. Deng, J., Shi, S., Li, P., Zhou, W., Zhang, Y., Li, H.: Voxel r-cnn: Towards high performance voxel-based 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1201-1209 (2021) +11. Deng, J., Chan, G., Zhong, H., Lu, C.X.: See beyond seeing: Robust 3d object detection from point clouds via cross-modal feature augmentation. arXiv preprint arXiv:2309.17336 (2023) +12. Ding, F., Palffy, A., Gavrila, D.M., Lu, C.X.: Hidden gems: 4d radar scene flow learning using cross-modal supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9340-9349 (June 2023) +13. Drews, F., Feng, D., Faion, F., Rosenbaum, L., Ulrich, M., Glaser, C.: Deepfusion: A robust and modular 3d object detector for lidars, cameras and radars. In: 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 560-567. IEEE (2022) +14. Engels, F., Heidenreich, P., Winternantel, M., Stäcker, L., Al Kadi, M., Zoubir, A.M.: Automotive radar signal processing: Research directions and practical challenges. IEEE Journal of Selected Topics in Signal Processing 15(4), 865-878 (2021) +15. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012) +16. Gupta, S., Hoffman, J., Malik, J.: Cross modal distillation for supervision transfer. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2827-2836 (2016) +17. Haiyang Wang, Chen Shi, S.S.M.L.S.W.D.H.B.S., Wang, L.: Dsvt: Dynamic sparse voxel transformer with rotated sets. In: CVPR (2023) +18. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. Deep Learning Workshop at NIPS (2014) +19. Hu, Q., Yang, B., Xie, L., Rosa, S., Guo, Y., Wang, Z., Trigoni, N., Markham, A.: Randla-net: Efficient semantic segmentation of large-scale point clouds. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11108-11117 (2020) +20. Jiang, M., Xu, G., Pei, H., Feng, Z., Ma, S., Zhang, H., Hong, W.: 4d high-resolution imagery of point clouds for automotive mmwave radar. IEEE Transactions on Intelligent Transportation Systems pp. 1-15 (2023) + +21. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. International Conference for Learning Representations (2015) +22. Klingner, M., Borse, S., Kumar, V.R., Rezaei, B., Narayanan, V., Yogamani, S., Porikli, F.: X3kd: Knowledge distillation across modalities, tasks and stages for multi-camera 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13343-13353 (June 2023) +23. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019) +24. Lang, I., Manor, A., Avidan, S.: Samplenet: Differentiable point cloud sampling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7578-7588 (2020) +25. Li, D., Wei, Y., Zhu, R.: A comparative study on point cloud down-sampling strategies for deep learning-based crop organ segmentation. Plant Methods 19(1), 124 (2023) +26. Li, J., Lu, M., Liu, J., Guo, Y., Du, Y., Du, L., Zhang, S.: Bev-lgkd: A unified lidarguided knowledge distillation framework for multi-view bev 3d object detection. IEEE Transactions on Intelligent Vehicles pp. 1-10 (2023) +27. Li, J., Luo, C., Yang, X.: Pillarnext: Rethinking network designs for 3d object detection in lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17567-17576 (June 2023) +28. Li, Q., Jin, S., Yan, J.: Mimicking very efficient network for object detection. In: Proceedings of the iee conference on computer vision and pattern recognition. pp. 6356-6364 (2017) +29. Li, Z., Li, Y., Wang, Y., Xie, G., Qu, H., Lyu, Z.: A lightweight model for 3d point cloud object detection. Applied Sciences 13(11) (2023) +30. Liu, J., Zhao, Q., Xiong, W., Huang, T., Han, Q.L., Zhu, B.: Smurf: Spatial multi-representation fusion for 3d object detection with 4d imaging radar. IEEE Transactions on Intelligent Vehicles pp. 1-14 (2023) +31. Mao, J., Minzhe, N., Jiang, C., liang, h., Chen, J., Liang, X., Li, Y., Ye, C., Zhang, W., Li, Z., Yu, J., Xu, C., Xu, H.: One million scenes for autonomous driving: Once dataset. In: Vanschoren, J., Yeung, S. (eds.) Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks. vol. 1 (2021) +32. Meyer, M., Kuschk, G.: Automotive radar dataset for deep learning based 3d object detection. In: 2019 16th european radar conference (EuRAD). pp. 129-132. IEEE (2019) +33. Nguyen, C.H., Nguyen, T.C., Tang, T.N., Phan, N.L.: Improving object detection by label assignment distillation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1005-1014 (January 2022) +34. Nobis, F., Shafiei, E., Karle, P., Betz, J., Lienkamp, M.: Radar voxel fusion for 3d object detection. Applied Sciences 11(12) (2021) +35. Paek, D.H., Kong, S.H., Wijaya, K.T.: K-radar: 4d radar object detection for autonomous driving in various weather conditions. In: Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2022) +36. Palffy, A., Pool, E., Baratam, S., Kooij, J.F.P., Gavrila, D.M.: Multi-class road user detection with 3+1d radar in the view-of-delft dataset. IEEE Robotics and Automation Letters 7(2), 4961-4968 (2022) + +37. Palmer, P., Krueger, M., Altendorfer, R., Adam, G., Bertram, T.: Reviewing 3d object detectors in the context of high-resolution $3 + 1$ d radar. In: Workshop on 3D Vision and Robotics at the Conference on Computer Vision and Pattern Recognition 2023 (2023) +38. Palmer, P., Krueger, M., Altendorfer, R., Bertram, T.: Ego-motion estimation and dynamic motion separation from 3d point clouds for accumulating data and improving 3d object detection. In: AmE 2023 - Automotive meets Electronics; 14. GMM Symposium. pp. 86-91 (2023) +39. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017) +40. Rebut, J., Ouaknine, A., Malik, W., Pérez, P.: Raw high-definition radar for multitask learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17021-17030 (June 2022) +41. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10529–10538 (2020) +42. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-779 (2019) +43. Smith, L.N., Topin, N.: Super-convergence: Very fast training of neural networks using large learning rates. In: Artificial intelligence and machine learning for multidomain operations applications. vol. 11006, pp. 369-386. SPIE (2019) +44. Stephan, M., Hazra, S., Santra, A., Weigel, R., Fischer, G.: People counting solution using an fmcw radar with knowledge distillation from camera data. In: 2021 IEEE Sensors. pp. 1-4 (2021) +45. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) +46. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020) +47. Sun, S., Zhang, Y.D.: 4d automotive radar sensing for autonomous vehicles: A sparsity-oriented approach. IEEE Journal of Selected Topics in Signal Processing 15(4), 879-891 (2021) +48. Tan, B., Ma, Z., Zhu, X., Li, S., Zheng, L., Chen, S., Huang, L., Bai, J.: 3d object detection for multi-frame 4d automotive millimeter-wave radar point cloud. IEEE Sensors Journal (2022) +49. Tung, F., Mori, G.: Similarity-preserving knowledge distillation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1365-1374 (2019) +50. Wang, L., Zhang, X., Li, J., Xv, B., Fu, R., Chen, H., Yang, L., Jin, D., Zhao, L.: Multi-modal and multi-scale fusion 3d object detection of 4d radar and lidar for autonomous driving. IEEE Transactions on Vehicular Technology 72(5), 5628-5641 (2023) +51. Wang, T., Yuan, L., Zhang, X., Feng, J.: Distilling object detectors with fine-grained feature imitation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4933-4942 (2019) + +52. Wei, Y., Wei, Z., Rao, Y., Li, J., Zhou, J., Lu, J.: Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. In: European Conference on Computer Vision. pp. 179-195. Springer (2022) +53. Wu, H., Wen, C., Shi, S., Li, X., Wang, C.: Virtual sparse convolution for multimodal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21653-21662 (June 2023) +54. Xiao, A., Huang, J., Guan, D., Zhan, F., Lu, S.: Transfer learning from synthetic to real lidar point cloud for semantic segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2795-2803 (2022) +55. Xiao, Y., Liu, Y., Luan, K., Cheng, Y., Chen, X., Lu, H.: Deep lidar-radar-visual fusion for object detection in urban environments. Remote Sensing 15(18) (2023) +56. Xiong, W., Liu, J., Huang, T., Han, Q.L., Xia, Y., Zhu, B.: Lxl: Lidar excluded lean 3d object detection with 4d imaging radar and camera fusion. IEEE Transactions on Intelligent Vehicles pp. 1-14 (2023) +57. Xu, B., Zhang, X., Wang, L., Hu, X., Li, Z., Pan, S., Li, J., Deng, Y.: Rpfa-net: a 4d radar pillar feature attention network for 3d object detection. In: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). pp. 3061-3066 (2021) +58. Yan, Q., Wang, Y.: Mvfan: Multi-view feature assisted network for 4d radar object detection. In: International Conference on Neural Information Processing. pp. 493-511. Springer (2023) +59. Yang, J., Shi, S., Ding, R., Wang, Z., Qi, X.: Towards efficient 3d object detection with knowledge distillation. In: Advances in Neural Information Processing Systems. vol. 35, pp. 21300-21313. Curran Associates, Inc. (2022) +60. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3dssd: Point-based 3d single stage object detector. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11040-11048 (2020) +61. You, Y., Wang, Y., Chao, W.L., Garg, D., Pleiss, G., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-lidar++: Accurate depth for 3d object detection in autonomous driving. In: ICLR (2020) +62. Zamanakos, G., Tsochatzidis, L., Amanatiadis, A., Pratikakis, I.: A comprehensive survey of lidar-based 3d object detection methods with deep learning for autonomous driving. Computers & Graphics 99, 153-181 (2021) +63. Zhang, J., Zhuge, H., Liu, Y., Peng, G., Wu, Z., Zhang, H., Lyu, Q., Li, H., Zhao, C., Kircali, D., Mharolkar, S., Yang, X., Yi, S., Wang, Y., Wang, D.: Ntu4dradlm: 4d radar-centric multi-modal dataset for localization and mapping. In: 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC) (2023) +64. Zhang, L., Dong, R., Tai, H.S., Ma, K.: Pointdistiller: Structured knowledge distillation towards efficient and compact 3d detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21791-21801 (June 2023) +65. Zhang, X., Wang, L., Chen, J., Fang, C., Yang, L., Song, Z., Yang, G., Wang, Y., Zhang, X., Yang, Q., Li, J.: Dual radar: A multi-modal dataset with dual 4d radar for autonomous driving. arXiv preprint arXiv:2310.07602 (2023) +66. Zhang, Y., Hu, Q., Xu, G., Ma, Y., Wan, J., Guo, Y.: Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18953-18962 (June 2022) +67. Zhao, M., Li, T., Abu Alsheikh, M., Tian, Y., Zhao, H., Torralba, A., Katabi, D.: Through-wall human pose estimation using radio signals. In: Proceedings of + +the IEEE conference on computer vision and pattern recognition. pp. 7356-7365 (2018) +68. Zheng, L., Li, S., Tan, B., Yang, L., Chen, S., Huang, L., Bai, J., Zhu, X., Ma, Z.: Rcfusion: Fusing 4-d radar and camera with bird's-eye view features for 3-d object detection. IEEE Transactions on Instrumentation and Measurement 72, 1-14 (2023) +69. Zheng, L., Ma, Z., Zhu, X., Tan, B., Li, S., Long, K., Sun, W., Chen, S., Zhang, L., Wan, M., Huang, L., Bai, J.: Tj4dradset: A 4d radar dataset for autonomous driving. In: 2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC). pp. 493-498 (2022) +70. Zhou, T., Chen, J., Shi, Y., Jiang, K., Yang, M., Yang, D.: Bridging the view disparity between radar and camera features for multi-modal fusion 3d object detection. IEEE Transactions on Intelligent Vehicles 8(2), 1523-1535 (2023) +71. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4490-4499 (2018) \ No newline at end of file diff --git a/2024/LEROjD_ Lidar Extended Radar-Only Object Detection/images.zip b/2024/LEROjD_ Lidar Extended Radar-Only Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e5b5fd6e4bfee8919a3dd6de7b1dabc73353f356 --- /dev/null +++ b/2024/LEROjD_ Lidar Extended Radar-Only Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37b82298ea5cc36a3f4c33b965c05a71c9093acb71827d36761f40f9fc3912cd +size 379288 diff --git a/2024/LEROjD_ Lidar Extended Radar-Only Object Detection/layout.json b/2024/LEROjD_ Lidar Extended Radar-Only Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..63771c2d0cfe4989424dded568b7bc72ba909ba4 --- /dev/null +++ b/2024/LEROjD_ Lidar Extended Radar-Only Object Detection/layout.json @@ -0,0 +1,8823 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 111, + 465, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 111, + 465, + 148 + ], + "spans": [ + { + "bbox": [ + 148, + 111, + 465, + 148 + ], + "type": "text", + "content": "LEROjD: Lidar Extended Radar-Only Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "text", + "content": "Patrick Palmer1, Martin Krüger1, Stefan Schütte1, Richard Altendorfer2, Ganesh Adam2, and Torsten Bertram1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 201, + 473, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 201, + 473, + 234 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 473, + 234 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 201, + 473, + 234 + ], + "type": "text", + "content": " Institute of Control Theory and Systems Engineering, TU Dortmund University {patrick.palmer, martin2.krueger, stefan.schuette, torsten.bertram}@tu-dortmund.de" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 181, + 235, + 432, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 235, + 432, + 246 + ], + "spans": [ + { + "bbox": [ + 181, + 235, + 432, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 181, + 235, + 432, + 246 + ], + "type": "text", + "content": " ZF Group, {richard.altendorfer, ganeshadam}@zf.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 273, + 452, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 273, + 452, + 483 + ], + "spans": [ + { + "bbox": [ + 159, + 273, + 452, + 483 + ], + "type": "text", + "content": "Abstract. Accurate 3D object detection is vital for automated driving. While lidar sensors are well suited for this task, they are expensive and have limitations in adverse weather conditions. " + }, + { + "bbox": [ + 159, + 273, + 452, + 483 + ], + "type": "inline_equation", + "content": "3 + 1\\mathrm{D}" + }, + { + "bbox": [ + 159, + 273, + 452, + 483 + ], + "type": "text", + "content": " imaging radar sensors offer a cost-effective, robust alternative but face challenges due to their low resolution and high measurement noise. Existing " + }, + { + "bbox": [ + 159, + 273, + 452, + 483 + ], + "type": "inline_equation", + "content": "3 + 1\\mathrm{D}" + }, + { + "bbox": [ + 159, + 273, + 452, + 483 + ], + "type": "text", + "content": " imaging radar datasets include radar and lidar data, enabling cross-modal model improvements. Although lidar should not be used during inference, it can aid the training of radar-only object detectors. We explore two strategies to transfer knowledge from the lidar to the radar domain and radar-only object detectors: 1. multi-stage training with sequential lidar point cloud thin-out, and 2. cross-modal knowledge distillation. In the multi-stage process, three thin-out methods are examined. Our results show significant performance gains of up to 4.2 percentage points in mean Average Precision with multi-stage training and up to 3.9 percentage points with knowledge distillation by initializing the student with the teacher's weights. The main benefit of these approaches is their applicability to other 3D object detection networks without altering their architecture, as we show by analyzing it on two different object detectors. Our code is available at https://github.com/rst-tu-dortmund/lerojd." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 491, + 453, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 491, + 453, + 514 + ], + "spans": [ + { + "bbox": [ + 159, + 491, + 453, + 514 + ], + "type": "text", + "content": "Keywords: 3D Object Detection " + }, + { + "bbox": [ + 159, + 491, + 453, + 514 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 491, + 453, + 514 + ], + "type": "text", + "content": " 3+1D Imaging Radar " + }, + { + "bbox": [ + 159, + 491, + 453, + 514 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 491, + 453, + 514 + ], + "type": "text", + "content": " Cross-Modal Object Detection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "content": "Environment perception is the first module in each automated driving stack. Multiple sensor modalities, like cameras, lidars, and radars are utilized for this task. Radar sensors are of unique interest in perception due to their robustness against poor lighting, challenging weather conditions like rain or snow, and cost-effectiveness [3, 36, 47]. One exclusive advantage is the ability to measure the relative radial velocity of reflections directly due to the Doppler effect." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "While a precise localization of objects is possible with traditional radar sensors without elevation angle measurements [9], it is inherently limited to the horizontal plane. Additionally, it is hard to predict the objects' extent due to" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 114, + 297, + 235 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 297, + 235 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 297, + 235 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 297, + 235 + ], + "type": "image", + "image_path": "7d1e10ae07ead6a221ae33a784919070263ff3eb3ad9df5150d61440c4f24bd0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 236, + 225, + 244 + ], + "lines": [ + { + "bbox": [ + 216, + 236, + 225, + 244 + ], + "spans": [ + { + "bbox": [ + 216, + 236, + 225, + 244 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 314, + 116, + 472, + 236 + ], + "blocks": [ + { + "bbox": [ + 314, + 116, + 472, + 236 + ], + "lines": [ + { + "bbox": [ + 314, + 116, + 472, + 236 + ], + "spans": [ + { + "bbox": [ + 314, + 116, + 472, + 236 + ], + "type": "image", + "image_path": "1ab6dd17d4634553e796d6f4a621a3c1ef026a3fc8aec8b8d1ca16ce5311a81e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 237, + 400, + 244 + ], + "lines": [ + { + "bbox": [ + 391, + 237, + 400, + 244 + ], + "spans": [ + { + "bbox": [ + 391, + 237, + 400, + 244 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 254, + 482, + 309 + ], + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 309 + ], + "type": "text", + "content": "Fig. 1: Architecture overview of (a) a knowledge distillation-based method and (b) a multi-stage training method (MSTM) for utilizing lidar data in the training of radar-only object detectors. The ground truth (GT) label is the same for both methods. The diagrams above the dotted line represent the training process, while the diagrams below the dotted line represent inference." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 336, + 479, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 479, + 396 + ], + "type": "text", + "content": "the reflections' sparsity. The introduction of " + }, + { + "bbox": [ + 130, + 336, + 479, + 396 + ], + "type": "inline_equation", + "content": "3 + 1\\mathrm{D}" + }, + { + "bbox": [ + 130, + 336, + 479, + 396 + ], + "type": "text", + "content": " high-resolution imaging radar sensors has recently mitigated these limitations, at least partially. In addition to measuring the elevation angle of reflections, the density of measurements is increased [14, 20]. Therefore, approaches that only utilize radar sensors for perception of the environment are of particular interest." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 398, + 480, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 480, + 446 + ], + "type": "text", + "content": "Despite improvements in radar-based object detection, the performance still lags behind other sensor modalities like lidar [36]. One persistent major limitation of 2+1D classic radar and 3+1D imaging radar sensors is the relative sparsity of the point cloud, which limits the detection performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 448, + 480, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 480, + 508 + ], + "type": "text", + "content": "Lidar sensors, on the other hand, are well suited for object detection and are therefore frequently employed as a reference for evaluating the performance of different sensor modalities due to their ability to produce an accurate and dense understanding of the scene. Their effectiveness is particularly pronounced in detecting nearby traffic participants without occlusion [53]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "All currently available datasets that contain " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "3 + 1\\mathrm{D}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " imaging radar data additionally accommodate lidar sensor data [4,8,32,35,36,40,63,65,69]. The lidar sensor data of these datasets is currently either utilized for labeling, combining multiple sensor modalities for accurate object detection, or comparing the performance of radar-only techniques to another sensor modality. While the majority of series production vehicles may not include lidar sensors due to cost and vehicle design constraints, they will still be available in the training process of learning-based methods. Extending radar-only methods with lidar sensor data in training has been shown to be a viable method for estimating point flow on the imaging radar point cloud [12]. These observations lead to the following research question: Can lidar sensor data be utilized in the training process of imaging radar-based 3D object detectors to improve the object detection performance on radar-only data during inference?" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "P. Palmer et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "To use different sensor modalities in training, transfer learning and knowledge distillation (KD) principles can be utilized. While KD is commonly employed across sensor modalities such as camera images and lidar point clouds, its application between " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "3 + 1\\mathrm{D}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " imaging radar and lidar sensors remains unexplored. Since lidar and imaging radar share a structurally similar data representation as point clouds, an identical base network with different input modalities can be utilized to transfer knowledge between different input modalities. We investigate two approaches to transfer knowledge from lidar-based to radar-based object detectors: a KD-based and a multi-stage training approach. The main principles of the two methods are visualized in Figure 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 236, + 403, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 236, + 403, + 247 + ], + "spans": [ + { + "bbox": [ + 146, + 236, + 403, + 247 + ], + "type": "text", + "content": "The contributions of this work are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 254, + 479, + 349 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 254, + 479, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 479, + 277 + ], + "type": "text", + "content": "- We investigate a combination of lidar and radar sensors in the training stage of object detectors to improve radar-only object detection at inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 277, + 479, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 479, + 301 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 479, + 301 + ], + "type": "text", + "content": "- We investigate three thin-out strategies for lidar point clouds to transfer knowledge from dense lidar to sparse lidar and radar-only object detectors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 301, + 479, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 479, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 479, + 325 + ], + "type": "text", + "content": "- We propose a multi-stage training procedure to transfer knowledge from dense lidar to sparse lidar and, finally, to radar-only object detection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 325, + 479, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 325, + 479, + 349 + ], + "spans": [ + { + "bbox": [ + 138, + 325, + 479, + 349 + ], + "type": "text", + "content": "- We modify and analyze several knowledge distillation-based approaches to transfer knowledge from lidar to radar-only object detectors." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 366, + 237, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 366, + 237, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 366, + 237, + 379 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 390, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 545 + ], + "type": "text", + "content": "Imaging radar sensors commonly utilize the point cloud as a data representation format instead of the radar tensor, due to its higher computational efficiency. This format is similar to the one used for lidars, enabling the application of object detection methods developed for lidars to radars. Object detection on lidars can be split into two main categories. Point-based methods, like PointNet++ [39], downsample the original point cloud, encode it using a backbone, and finally apply a detection head. Voxel-based methods discretize the point cloud into a 3D grid and apply 3D convolutions to the grid before finally applying a detection head [6,10,71]. The main drawback of voxel-based methods is the high memory consumption and the loss of spatial information due to the discretization of the point cloud. To overcome the limitation of high memory consumption, [23,27] have proposed the PointPillars network. Pillars are a specific form of voxels that span over the full height of the scene and represent the point cloud in a 2D grid." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Methods from the lidar domain have been shown to perform reasonably well on radar data [36,37]. However, they are limited due to the sparsity of the radar point cloud. SMURF [30] considers two representations of the radar point cloud to address sparsity. Using kernel density estimation, it utilizes pillarization and density features derived from a multi-dimensional Gaussian mixture distribution. RPFA-Net [57] is a PointPillars [23] based network, which introduces a selfattention mechanism to extract global context information from the radar point cloud. RadarMFNet [48] utilizes a multi-frame radar point cloud representation to address the sparsity of the radar point cloud in conjunction with an anchor-based detector and temporal pooling layers." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 219, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 219, + 91, + 447, + 103 + ], + "type": "text", + "content": "LEROjD: Lidar Extended Radar-Only Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "One way of improving 3D object detection on radar data is the fusion with additional sensor modalities like camera [56, 68, 70], lidar [50] or camera and lidar [5, 13, 55] at the cost of introducing additional sensors at inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 481, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 481, + 261 + ], + "type": "text", + "content": "The concept of knowledge distillation was first introduced by [18]. It consists of two networks, which are labeled as teacher and student. The teacher is a large and complex model, while the student is less complex and more computationally efficient. The student is trained to mimic the teacher network's performance by utilizing the teacher's predictions and the ground truth labels. This has been utilized in the context of 3D object detection by [29,59,64] to construct computation time-efficient models that have similar performance as large models. [26] and [22] have extended upon this concept by utilizing KD to extract knowledge from a teacher trained on lidar data to a student who utilizes camera images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 262, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 441 + ], + "type": "text", + "content": "To the best of our knowledge, few studies have investigated the effect of KD considering radar data. A transfer of knowledge from an image-based teacher network to a radar-based student has been shown to improve the performance for the task of people counting [44]. For the task of 3D object detection, [22] has shown that a transfer learning-based approach from a lidar and image-based network to the radar domain results in improved object detection performance for classical " + }, + { + "bbox": [ + 130, + 262, + 482, + 441 + ], + "type": "inline_equation", + "content": "2 + 1\\mathrm{D}" + }, + { + "bbox": [ + 130, + 262, + 482, + 441 + ], + "type": "text", + "content": " radar sensors. The main drawback of [22] is that in addition to 3D object labels, instance segmentation labels for the image domain are required. Additionally, [22] introduces multiple sub-networks to derive the KD losses, which makes the model more complex, while we aim for a simpler approach. HiddenGems [12] utilizes lidar point clouds to derive point flow information and train a network to predict the point flow on radar point clouds. See Beyond Seeing [11] utilizes lidar point cloud for feature hallucination on radar point clouds. Both methods require the lidar point cloud only in the training process but require major modifications to the network architecture." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 442, + 482, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 442, + 482, + 551 + ], + "spans": [ + { + "bbox": [ + 130, + 442, + 482, + 551 + ], + "type": "text", + "content": "Sampling of point clouds is a broadly explored topic to reduce the complexity of computing large point clouds. Most commonly, farthest point sampling (FPS) [39,42,60], voxelized FPS [41] or random sampling [71] are utilized in the context of 3D object detection. Random sampling has been shown to be beneficial when used for semantic segmentation of point clouds [19] and to outperform farthest point sampling at this task [25]. One limitation of random sampling is that while points at a close distance are kept, points at a far distance are more likely to be removed. Therefore, [7] proposes a polar cylinder balanced random sampling to keep a more balanced distribution of points across the range." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 574, + 202, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 202, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 202, + 586 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "Two methods of transferring knowledge from lidar-based to radar-based object detectors are investigated: Knowledge Distillation (KD) (Section 3.1), which is modified for the task of transfer learning and our proposed multi-stage training procedure with sequential point cloud thin-out (Section 3.2). Additionally, the utilized thin-out strategies for lidar point cloud (Section 3.3) are described." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "P. Palmer et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 276, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 276, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 276, + 128 + ], + "type": "text", + "content": "3.1 Knowledge Distillation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 266 + ], + "type": "text", + "content": "KD is commonly used for two tasks. First, designing computationally efficient models by transferring knowledge from a larger teacher network to a smaller student network [18, 49, 51]. Second, to transfer knowledge across sensor modalities by utilizing different model architectures for the teacher and student [1, 16, 67]. Lidar and radar-based point cloud object detection can utilize the same model structure but with different input modalities. This enables the utilization of KD methods first described for designing computationally efficient models for cross-modal knowledge transfer from lidar- to radar-based object detectors. In this case, the teacher is trained on the full lidar point cloud, while the student is trained on the radar point cloud. Three different loss terms, as described by [59], are employed:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 267, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 327 + ], + "type": "text", + "content": "Logit-KD is the first, classic type of distilling knowledge described by [18]. For 3D object detection, the logit-KD loss " + }, + { + "bbox": [ + 130, + 267, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{logit}}" + }, + { + "bbox": [ + 130, + 267, + 482, + 327 + ], + "type": "text", + "content": " is split into a classification " + }, + { + "bbox": [ + 130, + 267, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{l -cls}}" + }, + { + "bbox": [ + 130, + 267, + 482, + 327 + ], + "type": "text", + "content": " and regression loss " + }, + { + "bbox": [ + 130, + 267, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{l - reg}}" + }, + { + "bbox": [ + 130, + 267, + 482, + 327 + ], + "type": "text", + "content": ". These losses are calculated by comparing the student's and teacher's predictions utilizing the 3d detectors' regression loss and bi-linear interpolation between student and teacher output classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 327, + 482, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 374 + ], + "type": "text", + "content": "Feature-KD is widely utilized in 2D object detection [28, 51]. It utilizes a loss term that forces the student network to mimic the teacher's intermediate feature map (feat). A feature mimicking the last layer of the bird's eye view feature encoder, similar to [59], is utilized in this work." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 376, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 458 + ], + "type": "text", + "content": "Label-KD is a recent distillation approach that leans on the concept of the Logit KD but simplifies and generalizes it. It is first described by [33]. The teacher predictions are filtered by their scores using a score threshold, and an adapted ground truth set is constructed by combining the filtered predictions and the ground truth set. This adapted set is utilized in student training. The loss is split into a classification " + }, + { + "bbox": [ + 130, + 376, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 130, + 376, + 482, + 458 + ], + "type": "text", + "content": " and regression loss " + }, + { + "bbox": [ + 130, + 376, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 130, + 376, + 482, + 458 + ], + "type": "text", + "content": ". It replaces " + }, + { + "bbox": [ + 130, + 376, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{label}}" + }, + { + "bbox": [ + 130, + 376, + 482, + 458 + ], + "type": "text", + "content": " usually calculated on the ground truth set." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 146, + 459, + 465, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 459, + 465, + 472 + ], + "spans": [ + { + "bbox": [ + 146, + 459, + 465, + 472 + ], + "type": "text", + "content": "The three KD loss terms are combined into a joint loss weighted with " + }, + { + "bbox": [ + 146, + 459, + 465, + 472 + ], + "type": "inline_equation", + "content": "\\lambda" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 165, + 479, + 481, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 479, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 165, + 479, + 481, + 510 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {j o i n t}} = \\underbrace {\\lambda_ {\\mathrm {l - r e g}} \\mathcal {L} _ {\\mathrm {l - r e g}} + \\lambda_ {\\mathrm {l - c l s}} \\mathcal {L} _ {\\mathrm {l - c l s .}}} _ {\\mathcal {L} _ {\\text {l o g i t}}} + \\underbrace {\\lambda_ {\\text {f e a t}} \\mathcal {L} _ {\\text {f e a t}}} _ {\\mathcal {L} _ {\\text {f e a t u r e}}} + \\underbrace {\\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}} + \\lambda_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}}} _ {\\mathcal {L} _ {\\text {l a b e l}} ^ {*}} \\tag {1}", + "image_path": "64f55bd8c968a27fc99405dcd23ffd8676c81eb6a1f51b9fcca27a9aeeecc9d4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 526, + 404, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 526, + 404, + 538 + ], + "spans": [ + { + "bbox": [ + 131, + 526, + 404, + 538 + ], + "type": "text", + "content": "3.2 Multi-Stage Lidar Thin-Out Training Procedure" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "Utilizing a different input data modality for either pre-training a network on a large dataset or utilizing simulated data in the training process of a point cloud-based network has been shown to improve the object detection performance [54]. The multi-stage training method (MSTM) proposed in this work extends upon this by utilizing a Curriculum learning [2] based training procedure by which the network is trained on iteratively sparsified lidar point clouds, similar to [52], and fine-tuned on the radar point cloud. Figure 2 visualizes our multi-stage training procedure. The network is first trained on the full lidar point cloud until convergence. In the following steps, the lidar point cloud is thinned-out iteratively by a factor of 2 and utilized for training a network whose weights are" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 218, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 218, + 91, + 447, + 103 + ], + "type": "text", + "content": "LEROjD: Lidar Extended Radar-Only Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 114, + 474, + 358 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 474, + 358 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 474, + 358 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 474, + 358 + ], + "type": "image", + "image_path": "4c1e883322e8a05bc2812e572cfa34f1d9933ec98787b0de36e580ef8d1e9d00.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 366, + 482, + 444 + ], + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 444 + ], + "type": "text", + "content": "Fig. 2: MSTM pipeline. A 3D object detection network is iteratively trained on increasingly sparse lidar data. Three different thin-out strategies are utilized for sparsification. The lidar point cloud is mixed with the radar point cloud in the second to last step. In the last step, the network is only trained on radar data. At inference, only the orange-shaded part is executed. Small points represent lidar points, large points radar points. The color of points corresponds to the distance from the ego-vehicle. The camera image is not used as an input but only aids the visualization." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 471, + 482, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 543 + ], + "type": "text", + "content": "initialized using the previously trained model. This forces the network to learn features for a good object detection performance on increasingly sparser point clouds. In the second to last step, the lidar point cloud is mixed with the radar points so that the network can translate from the lidar to the radar domain. In the last step, the model is trained only on radar points. A training without multiple stages is called single-stage training method (SSTM) in this work." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "text", + "content": "In addition to the training where only lidar points are utilized in the first training stages, we investigate the utilization of the radar point cloud in conjunction with the lidar point cloud in all stages. The thin-out of the lidar points remains the same and is mixed with the radar point cloud in each step. This conditions the model on radar from the first step in order to prioritize features in the lidar point cloud that relate to a good object detection on radar-only data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "For mixing lidar and radar point clouds, the voxel or pillar feature encoder is modified to prioritize the radar point cloud in the random sampling process as done by [34]. Otherwise, the vastly higher number of lidar points, even when thinned-out, might lead to the complete exclusion of radar points." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "P. Palmer et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 292, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 292, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 292, + 128 + ], + "type": "text", + "content": "3.3 Lidar Thin-Out Strategies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 479, + 176 + ], + "type": "text", + "content": "Three different methods for sub-sampling the lidar point cloud are investigated. Figure 2 shows examples for each thin-out stage." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 178, + 480, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 178, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 480, + 239 + ], + "type": "text", + "content": "Random sampling is the simplest method of sparsifying a point cloud. It neglects the structure and inherent limitations of the point cloud representation, especially for objects far away or with a high degree of occlusion. Neglecting the structure can lead to the complete loss of information for objects represented only by few points." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 242, + 480, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 242, + 480, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 480, + 302 + ], + "type": "text", + "content": "K-nearest neighbor sampling approximates the reflection density distribution of the radar point cloud with the lidar point cloud by only keeping lidar reflections close to radar reflections. This algorithm is described in Algorithm 1. Objects not detected by the radar sensor are, therefore, also not represented by the k-nearest neighbor thinned-out lidar point cloud." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 304, + 481, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 481, + 365 + ], + "type": "text", + "content": "Voxel-based sampling aims to reduce the number of points in each area of the point cloud while keeping the general distribution of the point cloud. This is motivated by the fact that radar sensors do not suffer as much from loss of resolution with distance as lidar sensors. The approach is described in Algorithm 2, and executed iteratively for a sequence of sparsification steps." + } + ] + } + ], + "index": 6 + }, + { + "type": "code", + "bbox": [ + 132, + 384, + 481, + 626 + ], + "blocks": [ + { + "bbox": [ + 140, + 371, + 336, + 383 + ], + "lines": [ + { + "bbox": [ + 140, + 371, + 336, + 383 + ], + "spans": [ + { + "bbox": [ + 140, + 371, + 336, + 383 + ], + "type": "text", + "content": "Algorithm 1: K-nearest neighbor sampling" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "lines": [ + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": "Input : Lidar point cloud " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "L\\in \\mathbb{R}^{N\\times 3}" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": " , radar point cloud " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "R\\in \\mathbb{R}^{M\\times 3}" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": " Output: Sub-sampled lidar point cloud " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "L_{s}\\in \\mathbb{R}^{K\\times 3}" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": " \n1 Calculate the euclidean distance " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": " between each lidar point " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "l_{i}\\in L" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": " and its nearest radar point from " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": " . \n2 Select a share of " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": " lidar points with the smallest distances c; \n3 return Share of Points " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "L_{s}\\in \\mathbb{R}^{K\\times 3}" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": " \nAlgorithm 2: Voxel-based sampling. \nInput : Lidar point cloud " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "L\\in \\mathbb{R}^{N\\times 3}" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": " Output: Sub-sampled lidar point cloud " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "L_{s}\\in \\mathbb{R}^{K\\times 3}" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": " \n1 Initialize " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": " . \n2 Voxelize the lidar point cloud into v voxels; \n3 Calculate the number of points in each voxel " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "P_{v}" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": " . \n4 Calculate the minimum number of points per voxel " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{min}}" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": " , so that at least " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "0.75N" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": " points are in voxels with more than " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{min}}" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": " points; \n5 From each voxel with more than " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{min}}" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": " points, choose " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{min}}" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": " random points to keep and add remaining points to " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "L_{p}" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": " . \n6 Randomly select O.5 N points from " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "L_{p}" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": " and remove them from " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "text", + "content": " . \n7 return Share of Points " + }, + { + "bbox": [ + 132, + 384, + 481, + 626 + ], + "type": "inline_equation", + "content": "L_{s}\\in \\mathbb{R}^{K\\times 3}" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_body" + } + ], + "index": 8, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Another common sparsification method utilized for imitating low-resolution lidar sensors is layer-based sampling [52, 61]. This approach is not investigated because radar sensors do not capture the environment on a layer basis." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 218, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 218, + 90, + 447, + 103 + ], + "type": "text", + "content": "LEROjD: Lidar Extended Radar-Only Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 302, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 302, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 302, + 129 + ], + "type": "text", + "content": "4 Experimental Evaluation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 261, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 261, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 261, + 152 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 157, + 485, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 157, + 485, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 157, + 485, + 325 + ], + "type": "text", + "content": "Dataset: All experiments are conducted utilizing the View-of-Delft (VoD) dataset [36]. It contains synchronized data of multiple sensor modalities. The 64-layer lidar sensor and the imaging radar are utilized in this work. A point cloud accumulated over 5 frames [36], which has been shown to improve object detection performance compared to no accumulation, is used for radar data [36, 38]. We detected a duplication of identical points in the lidar point cloud, which could adversely affect all sampling methods; thus, we eliminated the duplicated points from the point cloud. Although the VoD dataset is among the best currently available datasets for imaging radar-based object detection, it is limited by its size compared to other automotive datasets without imaging-radar, like [31, 45]. Given the absence of publicly available labels and limited online evaluation for the test dataset, we repurpose the validation dataset as a test set. Consequently, we partition the original training set into a new training set " + }, + { + "bbox": [ + 130, + 157, + 485, + 325 + ], + "type": "inline_equation", + "content": "(80\\%)" + }, + { + "bbox": [ + 130, + 157, + 485, + 325 + ], + "type": "text", + "content": " and a dedicated validation set " + }, + { + "bbox": [ + 130, + 157, + 485, + 325 + ], + "type": "inline_equation", + "content": "(20\\%)" + }, + { + "bbox": [ + 130, + 157, + 485, + 325 + ], + "type": "text", + "content": " to ensure robust model training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 325, + 483, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 483, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 483, + 396 + ], + "type": "text", + "content": "Evaluation Metrics: The primary performance metric utilized to compare the results is the mean average precision (mAP), as used by [15] [37]. Similar to the evaluation of the Waymo data set [46] in [62], we split the results into two distance bins: short-range (SR): " + }, + { + "bbox": [ + 130, + 325, + 483, + 396 + ], + "type": "inline_equation", + "content": "0 - 30\\mathrm{m}" + }, + { + "bbox": [ + 130, + 325, + 483, + 396 + ], + "type": "text", + "content": " and mid-range (MR): " + }, + { + "bbox": [ + 130, + 325, + 483, + 396 + ], + "type": "inline_equation", + "content": "30 - 50\\mathrm{m}" + }, + { + "bbox": [ + 130, + 325, + 483, + 396 + ], + "type": "text", + "content": ". All experiments are conducted utilizing three different random network initializations that are averaged." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 397, + 483, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 483, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 483, + 576 + ], + "type": "text", + "content": "Training: Most experiments use the PointPillars model [23] as an object detector with the same configuration as utilized by [36]. For imaging radar data, PointPillars has been shown to perform among the best out of multiple state-of-the-art 3D object detection methods while still performing adequately on lidar data [37]. Furthermore PointPillars is a relevant baseline for radar-specific object detection methods [57,58]. To show that the proposed MSTM and KD apply to various object detectors, the most promising approaches from the evaluations on PointPillars are evaluated on DSVT-P [17], as an example for a transformer-based model. All SSTM trainings are conducted with an early stopping policy for a maximum of 125 epochs. For the MSTM the initial training on the full lidar point cloud is conducted for 125 epochs, while each refinement step is trained for 30 epochs. All trainings utilize the Adam optimizer [21] and an adapted learning rate scheduler that reaches its maximum learning rate earlier and has a faster descent than the scheduler described by [43]. This improves the object detection performance on radar data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 146, + 576, + 479, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 576, + 479, + 588 + ], + "spans": [ + { + "bbox": [ + 146, + 576, + 479, + 588 + ], + "type": "text", + "content": "Notation: To distinguish between methods, the following notation is used:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 264, + 608, + 351, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 608, + 351, + 623 + ], + "spans": [ + { + "bbox": [ + 264, + 608, + 351, + 623 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {\\mathcal {L S} / \\mathcal {T O}} ^ {\\mathcal {T M}} \\rightarrow \\mathcal {T} ^ {\\mathcal {K D}}.", + "image_path": "0726643649ccdf8a80de6b999c9b4b475396d1d2a8418a5874aa644401118586.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "The notation is split into two parts. The part left of the arrow represents the data set utilized for pre-training, while the right part represents the data and training method utilized for the last (fine-tuning) training stage of the model." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "P. Palmer et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 146, + 473, + 240 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 1: Examples of training steps and evaluation data corresponding to the notation, emplying voxel-based sub-sampling. Voxel sampling is shortened to v for legibility." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 146, + 473, + 240 + ], + "lines": [ + { + "bbox": [ + 138, + 146, + 473, + 240 + ], + "spans": [ + { + "bbox": [ + 138, + 146, + 473, + 240 + ], + "type": "table", + "html": "
NotationTrainingEvaluation
L1-1/8/vL1/v → L1/2/v → L1/4/v → L1/8/vL1/8/v
RLMSTM1-1/8/vRL1/v → RL1/2/v → RL1/4/v → RL1/8/vRL1/8/v
L1MSTM1-1/16/v → RL1/v → L1/2/v → L1/4/v → L1/8/v → RL1/16/v → RR
RLMSTM1-1/16/v → RRL1/v → RL1/2/v → RL1/4/v → RL1/8/v → RL1/16/v → RR
RLSSTM1/4/v → RRL1/4/v → RR
", + "image_path": "1a636084813b97cf9700c424ac466e9931042e32798443ffb23b1e9139e87326.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 260, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 392 + ], + "type": "text", + "content": "This part is omitted if it is the same as the training dataset. The training data " + }, + { + "bbox": [ + 130, + 260, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 260, + 482, + 392 + ], + "type": "text", + "content": " can either be lidar (L), radar (R), or mixed radar + lidar data (RL). The training method " + }, + { + "bbox": [ + 130, + 260, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{T}\\mathcal{M}" + }, + { + "bbox": [ + 130, + 260, + 482, + 392 + ], + "type": "text", + "content": " can denote either MSTM or SSTM. The lidar share " + }, + { + "bbox": [ + 130, + 260, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{L}\\mathcal{S}" + }, + { + "bbox": [ + 130, + 260, + 482, + 392 + ], + "type": "text", + "content": " is the fraction of the original lidar point cloud that is utilized. Radar-only training always uses the full radar point cloud, therefore " + }, + { + "bbox": [ + 130, + 260, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{L}\\mathcal{S}" + }, + { + "bbox": [ + 130, + 260, + 482, + 392 + ], + "type": "text", + "content": " is omitted. For the MSTM, this is represented by a range of fractions that are iterated in training. The thin-out method " + }, + { + "bbox": [ + 130, + 260, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{T}\\mathcal{O}" + }, + { + "bbox": [ + 130, + 260, + 482, + 392 + ], + "type": "text", + "content": " can either be random (rand), k-nearest neighbor (knn) or voxel-based (vox). " + }, + { + "bbox": [ + 130, + 260, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{KD}" + }, + { + "bbox": [ + 130, + 260, + 482, + 392 + ], + "type": "text", + "content": " represents the KD method; this can either be label (lab), logit (log), feature (feat), or a combination of those (joint). " + }, + { + "bbox": [ + 130, + 260, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{KD}" + }, + { + "bbox": [ + 130, + 260, + 482, + 392 + ], + "type": "text", + "content": " is omitted if just an initialization and fine-tuning is utilized. Examples for the training corresponding to specific notations are listed in Table 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 407, + 479, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 479, + 420 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 479, + 420 + ], + "type": "text", + "content": "4.2 Evaluation of MSTM on Lidar-only and Mixed Radar + Lidar" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": "To evaluate the applicability of our proposed MSTM to the thinned-out lidar point cloud, we evaluate the training without the last two steps involving radar data. Training is conducted for just the lidar point cloud and the mixed radar + lidar point cloud in all stages. The multi-stage trained network is evaluated after each thin-out stage on the thinned-out lidar (or mixed radar + lidar) point cloud. Thin-out stages up to " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "1/256" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " of the original lidar point cloud are considered due to the lidar point cloud containing fewer points than the radar point cloud at " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "1/256" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " of the original lidar point cloud. The MSTM is compared to the SSTM, which is trained only on the thinned-out point cloud. The results are shown in Fig. 3, complete quantitative results are given in the supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "All thin-out strategies result in an increasing degradation of the detection performance. This is most pronounced when considering a random thin-out, which drops approximately linearly with each thin-out step. K-nearest neighbor and voxel-based sampling consistently perform better than random sampling due to keeping a higher point density in areas around objects, which supports object detection. For k-nearest neighbor sampling, the performance only drops by 0.6 percentage points between " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_{1 / \\mathrm{knn}}^{\\mathrm{SSTM}}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_{1 / 2 / \\mathrm{knn}}^{\\mathrm{SSTM}}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " due to mostly ground points being removed in the first thin-out stage. For voxel-based sampling, after a sharp performance drop in the first thin-out stage, only a relatively slight performance drop is observable until the " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "^{1/16}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " thin-out stage. In the first stages," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 218, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 218, + 91, + 447, + 103 + ], + "type": "text", + "content": "LEROjD: Lidar Extended Radar-Only Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 478, + 237 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 478, + 237 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 478, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 478, + 237 + ], + "type": "image", + "image_path": "9ea428973687d550449b45cf72c04b8e0f264792b40cbc1352e6158db55bb224.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 247, + 480, + 258 + ], + "lines": [ + { + "bbox": [ + 132, + 247, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 480, + 258 + ], + "type": "text", + "content": "Fig. 3: Comparison between SSTM and MSTM for different lidar sampling strategies." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 281, + 482, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 425 + ], + "type": "text", + "content": "the performance drop can mainly be attributed to the pedestrian class, which drops by 12.8 percentage points in the first thin-out stage. In comparison, the detection performance of the car and cyclist classes only decreased by 0.7 and 3 percentage points, respectively. This can be explained by the voxel size of " + }, + { + "bbox": [ + 130, + 281, + 482, + 425 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}" + }, + { + "bbox": [ + 130, + 281, + 482, + 425 + ], + "type": "text", + "content": " in each dimension utilized in the voxel-based sampling. A single pedestrian only occupies a small number of these voxels. When the point cloud gets thinned out, the entire object's structure gets lost, making detection difficult. In contrast, the object's structure and physical appearance can be adequately represented when it occupies more voxels, as observed in the car and cyclist classes. In the final thin-out stage of " + }, + { + "bbox": [ + 130, + 281, + 482, + 425 + ], + "type": "inline_equation", + "content": "\\mathrm{L}^{\\mathrm{SSTM}}" + }, + { + "bbox": [ + 130, + 281, + 482, + 425 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 281, + 482, + 425 + ], + "type": "inline_equation", + "content": "\\mathrm{L}^{\\mathrm{MSTM}}" + }, + { + "bbox": [ + 130, + 281, + 482, + 425 + ], + "type": "text", + "content": ", voxel-based sampling performs worse than k-nearest neighbor sampling because many voxels only consist of ground points or points of surrounding background objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 426, + 481, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 426, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 481, + 449 + ], + "type": "text", + "content": "The multi-stage training does not result in a useful knowledge transfer and, therefore, significant performance benefit for all considered thin-out strategies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": "A contrary behavior is observed on the combined radar + lidar point cloud. The MSTM consistently outperforms the SSTM when using k-nearest neighbor or voxel-based sampling. Knowledge from the dense point cloud can be transferred to the thin radar + lidar point cloud. Additionally, performance is consistently higher than training just on lidar, especially for voxel-based subsampling, which performs best at lower thin-out stages. At small thin-out stages, the voxel-based sub-sampling can still represent the whole object space and give meaningful environmental information. At the same time, the radar point cloud is sufficiently dense for object detection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 574, + 408, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 408, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 408, + 586 + ], + "type": "text", + "content": "4.3 Evaluation of MSTM with Last Radar-Only Step" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "This chapter analyses the performance of the MSTM when applied to radar data, as described in Section 3.2. The MSTM is evaluated for two different procedures. Utilizing only the lidar point cloud and utilizing the mixed radar + lidar point cloud in the first stages. The results of MSTM are shown in Table 2. The lidar thin-out stages up to " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "1/16" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": " of the original lidar point cloud are considered due to the performance of " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_{1-1/32}^{\\mathrm{MSTM}}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": " dropping below " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{R}^{\\mathrm{SSTM}}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "P. Palmer et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 157, + 468, + 288 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: 3D object detection results on the VoD radar dataset trained using MSTM. All MSTM methods are trained on radar-only data in the last step. The best and second best results are marked in bold and underlined, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 157, + 468, + 288 + ], + "lines": [ + { + "bbox": [ + 143, + 157, + 468, + 288 + ], + "spans": [ + { + "bbox": [ + 143, + 157, + 468, + 288 + ], + "type": "table", + "html": "
Training methodmAPCarPedestrianCyclist
SRMRSRMRSRMRSRMR
RSSTM (baseline)36.711.945.218.117.17.447.710.2
L1-1/16 / rand → R36.416.144.920.017.59.846.818.9
L1-1/16 / knn → R35.714.045.623.216.77.144.811.7
L1-1/16 / vox → R34.114.347.023.015.39.140.010.7
RLMSTM 1-1/16 / rand → R35.614.944.120.217.87.244.917.3
RLMSTM 1-1/16 / knn → R38.214.745.523.918.88.250.311.9
RLMSTM 1-1/16 / vox → R39.715.445.922.518.49.754.713.9
", + "image_path": "52ce632e4cc3e482cfd399a01e8e480481225ff362d67794e5de607c9ee90335.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 308, + 481, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 308, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 133, + 308, + 481, + 452 + ], + "type": "text", + "content": "When considering the pre-training using just lidar data, it is observable that the overall performance drops in the SR bin. In contrast, the MR bin is improved for all thin-out methods. The best performance is achieved by utilizing the random thin-out strategy. It performs especially well for pedestrians and cyclists in the mid-range, due to the comparably small size of pedestrians and cyclists resulting in a worse radar reflection characteristic. In contrast these objects are detected well by the lidar sensor. Knowledge about the representation of objects can be transferred from the lidar point cloud to the radar point cloud. K-nearest neighbor and voxel-based thin-out do not improve the performance in the SR bin for the pedestrian and cyclist class but perform better in the short and mid-range for the car class which is explained by the large size of cars, resulting in a better representation of cars in the thin point cloud." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 453, + 481, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 453, + 481, + 608 + ], + "spans": [ + { + "bbox": [ + 133, + 453, + 481, + 608 + ], + "type": "text", + "content": "When considering the pre-training with the mixed radar + lidar point cloud, a contrary observation is made compared to only using the lidar in pre-training. The overall performance for the random thinned-out lidar point cloud is worse with the mixed point cloud than when only considering the lidar point cloud. One consistent aspect is the good performance of the cyclists in the mid-range, which still surpasses all other methods, excluding " + }, + { + "bbox": [ + 133, + 453, + 481, + 608 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_{1 - 1/16}^{\\mathrm{MSTM}}" + }, + { + "bbox": [ + 133, + 453, + 481, + 608 + ], + "type": "text", + "content": ". However, the k-nearest neighbor and voxel-based thin-out strategies perform better with the mixed point cloud. The voxel-based thin-out strategy achieves the best performance. It performs best on objects in SR, mainly due to its outstanding performance in detecting cyclists, but is limited in the mid-range, getting surpassed by the MSTM with random thin-out. The MSTM with voxel-based thin-out can increase the object detection performance on the radar point cloud by 3 percentage points in the SR and 3.5 percentage points in the MR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 624, + 330, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 330, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 330, + 635 + ], + "type": "text", + "content": "4.4 Evaluation of Cross-Modality KD" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": "The configuration of the teacher's training data is of particular interest, as the teacher's performance directly influences the student's. The simplest choice is" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 219, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 219, + 91, + 447, + 103 + ], + "type": "text", + "content": "LEROjD: Lidar Extended Radar-Only Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 153, + 157, + 457, + 255 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "text", + "content": "Table 3: 3D object detection results of the teacher with different training sets. Each teacher is evaluated on the test set of the same data configuration as the training set. The best and second best results are marked in bold and underlined, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 157, + 457, + 255 + ], + "lines": [ + { + "bbox": [ + 153, + 157, + 457, + 255 + ], + "spans": [ + { + "bbox": [ + 153, + 157, + 457, + 255 + ], + "type": "table", + "html": "
DatamAPCarPedestrianCyclist
SRMRSRMRSRMRSRMR
L1SSTM56.530.060.742.440.119.568.528.1
RLSSTM61.645.161.946.544.434.078.554.8
RLSSTM1/4 / rand56.931.157.232.437.221.376.439.5
RLSSTM1/4 / knn58.936.954.431.943.030.979.547.9
RLSSTM1/4 / Vox56.139.360.141.632.828.875.547.5
", + "image_path": "bdca4fec737c4512b93ef9a3955cc2d392ca4937e682c572d362de297a4d2f83.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "type": "text", + "content": "to train the teacher solely on the lidar point cloud. Section 4.2 and Section 4.3 show that mixing the radar and lidar point clouds can benefit radar-only object detection. Therefore, teachers trained on a mixed point clouds, in addition to the ones solely trained on lidar, are compared. Mixed point clouds containing " + }, + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "type": "text", + "content": " of the original lidar point cloud are considered, as a closer representation of the teachers to the student's data can lead to a better performance. Thin out of " + }, + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "type": "text", + "content": " is chosen as the radar + lidar detection performance with radar and " + }, + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 130, + 278, + 482, + 386 + ], + "type": "text", + "content": " of the lidar points is worse than the SSTM on only lidar data. Table 3 shows the performance of all utilized teachers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 386, + 482, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 446 + ], + "type": "text", + "content": "Table 3 shows the results of the teacher network for different training data configurations. It can be observed that, as expected, the mixed radar and lidar point cloud performs the best, only being surpassed by " + }, + { + "bbox": [ + 130, + 386, + 482, + 446 + ], + "type": "inline_equation", + "content": "\\mathrm{RL}_{1/4}^{\\mathrm{SSTM}}" + }, + { + "bbox": [ + 130, + 386, + 482, + 446 + ], + "type": "text", + "content": " on the cyclist class in short range. No model can be considered the overall second best among the other training sets. The performance varies between vehicle classes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 447, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 518 + ], + "type": "text", + "content": "The KD is evaluated individually for each KD method and teacher training set. A joint KD is also considered comprised of all three KD losses. All student networks are initialized (Init) with the weights of the teacher network, as it has been shown to improve the student's performance [59]. Additionally, for the transfer learning between datasets, the pre-training utilizing MSTM in Section 4.3 has been shown to improve the performance on the radar dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 518, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 482, + 582 + ], + "type": "text", + "content": "The results utilizing the KD are shown in Table 4. Just the initialization of the student network with the teacher's weights already results in a performance benefit in the SR with " + }, + { + "bbox": [ + 130, + 518, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathrm{RL}_{1}^{\\mathrm{SSTM}}" + }, + { + "bbox": [ + 130, + 518, + 482, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 518, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathrm{RL}_{1/4}^{\\mathrm{SSTM}}" + }, + { + "bbox": [ + 130, + 518, + 482, + 582 + ], + "type": "text", + "content": " as teachers and in the MR with all teachers. This is overall only surpassed by the " + }, + { + "bbox": [ + 130, + 518, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathrm{RL}_{1-1/16}^{\\mathrm{MSTM}} / \\mathrm{vox} \\to \\mathrm{R}" + }, + { + "bbox": [ + 130, + 518, + 482, + 582 + ], + "type": "text", + "content": " showing that the MSTM is substantially better than a simple initialization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "For label-KD, the best-performing models are the ones where the teacher performs the best. Specifically " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_1^{\\mathrm{SSTM}}\\rightarrow \\mathrm{R}^{\\mathrm{lab}}" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": ", which is only surpassed by " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{RL}_{1/4/knn}^{\\mathrm{SSTM}}\\rightarrow \\mathrm{R}^{\\mathrm{lab}}" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " in the MR. For worse-performing teachers, using label-KD loss does not result in a performance benefit due to it replacing the ground truth label loss. Feature-KD requires a teacher who learns features from radar data. This is observable in the SR performance of models, where the teacher is trained on mixed radar + lidar data. Logit-KD works well on teacher datasets" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "P. Palmer et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 179, + 482, + 305 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 4: 3D object detection results on the VoD radar dataset utilizing the different KDs and teachers trained on different data configurations. Only the mAP over all classes is specified. For each distillation method and range bin, the best and second best results are marked in bold and underlined, respectively. The best result for each teacher are marked in cyan for the SR bin and magenta for the MR bin." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 179, + 482, + 305 + ], + "lines": [ + { + "bbox": [ + 132, + 179, + 482, + 305 + ], + "spans": [ + { + "bbox": [ + 132, + 179, + 482, + 305 + ], + "type": "table", + "html": "
Teacher dataInit only (→R)Logit-KD (→Rlog)Feature-KD (→Rfeat)Label-KD (→Rlab)Joint-KD (→RJoint)
SRMRSRMRSRMRSRMRSRMR
R1SSTM36.711.9--------
L1SSTM34.913.532.812.634.813.636.613.736.013.5
RLSSTM39.014.839.213.539.113.838.813.738.912.7
RLSSTM1/4 / rand39.014.238.911.639.411.834.012.236.412.3
RLSSTM1/4 / knn34.814.637.214.736.613.033.810.932.912.9
RLSSTM1/4 / Vox35.215.335.613.337.115.835.314.332.99.7
", + "image_path": "80ac5cdede6324392a3a815a49aacf10558dc0280fd26e24e3eabc29a3c68c92.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 327, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 470 + ], + "type": "text", + "content": "which closely resemble the radar point cloud and perform well on the teacher set. This results in good performance of " + }, + { + "bbox": [ + 130, + 327, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\mathrm{RL}_1^{\\mathrm{SSTM}}\\rightarrow \\mathrm{R}^{\\log}" + }, + { + "bbox": [ + 130, + 327, + 482, + 470 + ], + "type": "text", + "content": ". Besides the KD method the thin-out method utilized in the teacher's training influences the student's performance. Random sampling results in good performance in the SR, while voxel-based sampling results in good MR performance. This is somewhat contrary to what has been observed in the MSTM, where random sampling performs the best in the MR and can be explained by the different thin-out stages utilized in the MSTM and KD. Further qualitative results, as well as detailed quantitative results are given in the supplementary. Overall, initializing the student with the teacher's parameters yields good performance, with further enhancements primarily achievable through feature-KD with a teacher trained on mixed radar + lidar data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 488, + 474, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 488, + 474, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 488, + 474, + 502 + ], + "type": "text", + "content": "4.5 Evaluation on DSVT as a Transformer-Based Object Detector" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 508, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 568 + ], + "type": "text", + "content": "Table 5 shows the results of the MSTM and KD on DSVT-P [17] for selected methods. Similar effects, as observed for PointPillars, are seen on DSVT-P. Initializing the student with the teacher's weights and the MSTM led to a performance benefit. However, contrary to PointPillars, the KD approach does not contribute to any improvements for DSVT-P." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 586, + 219, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 219, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 219, + 597 + ], + "type": "text", + "content": "4.6 Limitations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "The best-performing methods in this work apply only to detectors that share an architecture with the target model, as a direct transfer of weights is performed for the best performance. Different models may require different levels and steps in the thinning process. These choices are additional tuning parameters that must be selected appropriately to maximize the benefit of transferring knowledge." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 218, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 218, + 91, + 447, + 103 + ], + "type": "text", + "content": "LEROjD: Lidar Extended Radar-Only Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 146, + 467, + 243 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 5: Radar-only detection performance utilizing DSVT-P as an object detector. The best and second best results are marked in bold and underlined, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 146, + 467, + 243 + ], + "lines": [ + { + "bbox": [ + 143, + 146, + 467, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 146, + 467, + 243 + ], + "type": "table", + "html": "
Training methodmAPCarPedestrianCyclist
SRMRSRMRSRMRSRMR
RSSTM38.313.042.520.221.912.350.56.6
RLMSTM1-1/16 / vox→R41.515.647.823.624.511.052.312.4
RLSSTMM1/4 / vox→R38.813.244.918.823.311.248.29.7
RLSSTMM1/4 / vox→Rfeat34.711.344.08.820.52.839.712.7
", + "image_path": "1579f1bc2c008b97721a3f3a172b912531f2452f3254eafd5c1678eb649cdea0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 262, + 219, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 219, + 275 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 219, + 275 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 287, + 481, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 481, + 430 + ], + "type": "text", + "content": "In this paper, we investigated two methods to transfer knowledge from lidar-based object detectors to radar-only object detection. First, MSTM with sequential sub-sampling of the lidar point cloud, and second, a KD-based approach. For the MSTM, we have investigated three thin-out strategies for the lidar point cloud. These thin-out strategies are also analyzed for the training of the KD teacher network. Both methods can substantially improve the object detection performance of a radar-only object detector. The MSTM with voxel-based thinout performs the best overall and can improve detection performance by up to 3.5 percentage points. For the KD methods, it is shown that initializing the student with the teacher's parameters, especially a teacher trained on mixed lidar and radar data, can improve the object detection performance on radar-only data, with further enhancement primarily achieved by utilizing feature-KD." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 432, + 481, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 481, + 514 + ], + "type": "text", + "content": "In future work, the applicability to further 3D object detection networks and the behavior with more advanced knowledge distillation like the ones utilized by [22] could be investigated. Due to different effects being noticed with the MSTM and the KD methods, combining both methods could be investigated by utilizing the MSTM as a teacher. To overcome the limitations of choosing a strict thin-out strategy, a learnable point cloud thin-out method [24,66] can be used." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 532, + 198, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 198, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 198, + 544 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 555, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 136, + 555, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 555, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 136, + 555, + 480, + 588 + ], + "type": "text", + "content": "1. Afouras, T., Chung, J.S., Zisserman, A.: Asr is all you need: Cross-modal distillation for lip reading. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 2143-2147. IEEE (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 590, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 590, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 136, + 590, + 480, + 621 + ], + "type": "text", + "content": "2. Bengio, Y., Louradour, J., Collobert, R., Weston, J.: Curriculum learning. In: Proceedings of the 26th annual international conference on machine learning. pp. 41-48 (2009)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 136, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 136, + 622, + 480, + 643 + ], + "type": "text", + "content": "3. Briske, S., Ruf, F., Höhne, F.: Recent evolution of automotive imaging radar and its information content. IET Radar, Sonar & Navigation 12(10), 1078-1081 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 136, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 643, + 480, + 665 + ], + "type": "text", + "content": "4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "P. Palmer et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "text", + "content": "driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 138, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 138, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 137, + 138, + 481, + 171 + ], + "type": "text", + "content": "5. Chen, X., Zhang, T., Wang, Y., Wang, Y., Zhao, H.: Futr3d: A unified sensor fusion framework for 3d detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 172-181 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 171, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 171, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 137, + 171, + 481, + 203 + ], + "type": "text", + "content": "6. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21674-21683 (June 2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 203, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 203, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 137, + 203, + 481, + 236 + ], + "type": "text", + "content": "7. Cheng, H., Han, X., Jiang, H., He, D., Xiao, G.: Pcb-random sampling for lidar semantic segmentation in autonomous driving scene. arXiv preprint arXiv:2209.13797 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 236, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 137, + 236, + 481, + 267 + ], + "type": "text", + "content": "8. Choi, M., Yang, S., Han, S., Lee, Y., Lee, M., Choi, K.H., Kim, K.S.: Msc-rad4r: Ros-based automotive dataset with 4d radar. IEEE Robotics and Automation Letters 8(11), 7194-7201 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 267, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 267, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 137, + 267, + 481, + 300 + ], + "type": "text", + "content": "9. Danzer, A., Griebel, T., Bach, M., Dietmayer, K.: 2d car detection in radar data with pointnets. In: 2019 IEEE Intelligent Transportation Systems Conference (ITSC). pp. 61-66 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 134, + 300, + 481, + 332 + ], + "type": "text", + "content": "10. Deng, J., Shi, S., Li, P., Zhou, W., Zhang, Y., Li, H.: Voxel r-cnn: Towards high performance voxel-based 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1201-1209 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 332, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 332, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 133, + 332, + 481, + 364 + ], + "type": "text", + "content": "11. Deng, J., Chan, G., Zhong, H., Lu, C.X.: See beyond seeing: Robust 3d object detection from point clouds via cross-modal feature augmentation. arXiv preprint arXiv:2309.17336 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 364, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 364, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 364, + 481, + 407 + ], + "type": "text", + "content": "12. Ding, F., Palffy, A., Gavrila, D.M., Lu, C.X.: Hidden gems: 4d radar scene flow learning using cross-modal supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9340-9349 (June 2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 407, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 407, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 133, + 407, + 481, + 450 + ], + "type": "text", + "content": "13. Drews, F., Feng, D., Faion, F., Rosenbaum, L., Ulrich, M., Glaser, C.: Deepfusion: A robust and modular 3d object detector for lidars, cameras and radars. In: 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 560-567. IEEE (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 450, + 481, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 450, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 133, + 450, + 481, + 483 + ], + "type": "text", + "content": "14. Engels, F., Heidenreich, P., Winternantel, M., Stäcker, L., Al Kadi, M., Zoubir, A.M.: Automotive radar signal processing: Research directions and practical challenges. IEEE Journal of Selected Topics in Signal Processing 15(4), 865-878 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 483, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 483, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 133, + 483, + 481, + 515 + ], + "type": "text", + "content": "15. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 515, + 481, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 515, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 133, + 515, + 481, + 548 + ], + "type": "text", + "content": "16. Gupta, S., Hoffman, J., Malik, J.: Cross modal distillation for supervision transfer. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2827-2836 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 548, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 548, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 133, + 548, + 481, + 568 + ], + "type": "text", + "content": "17. Haiyang Wang, Chen Shi, S.S.M.L.S.W.D.H.B.S., Wang, L.: Dsvt: Dynamic sparse voxel transformer with rotated sets. In: CVPR (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 568, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 568, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 133, + 568, + 481, + 589 + ], + "type": "text", + "content": "18. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. Deep Learning Workshop at NIPS (2014)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 589, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 589, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 133, + 589, + 481, + 633 + ], + "type": "text", + "content": "19. Hu, Q., Yang, B., Xie, L., Rosa, S., Guo, Y., Wang, Z., Trigoni, N., Markham, A.: Randla-net: Efficient semantic segmentation of large-scale point clouds. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11108-11117 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 481, + 665 + ], + "type": "text", + "content": "20. Jiang, M., Xu, G., Pei, H., Feng, Z., Ma, S., Zhang, H., Hong, W.: 4d high-resolution imagery of point clouds for automotive mmwave radar. IEEE Transactions on Intelligent Transportation Systems pp. 1-15 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 219, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 219, + 91, + 447, + 102 + ], + "type": "text", + "content": "LEROjD: Lidar Extended Radar-Only Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "21. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. International Conference for Learning Representations (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 194 + ], + "type": "text", + "content": "22. Klingner, M., Borse, S., Kumar, V.R., Rezaei, B., Narayanan, V., Yogamani, S., Porikli, F.: X3kd: Knowledge distillation across modalities, tasks and stages for multi-camera 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13343-13353 (June 2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 195, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 480, + 228 + ], + "type": "text", + "content": "23. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 229, + 480, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 480, + 261 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 480, + 261 + ], + "type": "text", + "content": "24. Lang, I., Manor, A., Avidan, S.: Samplenet: Differentiable point cloud sampling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7578-7588 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 480, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 480, + 295 + ], + "type": "text", + "content": "25. Li, D., Wei, Y., Zhu, R.: A comparative study on point cloud down-sampling strategies for deep learning-based crop organ segmentation. Plant Methods 19(1), 124 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 480, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 328 + ], + "type": "text", + "content": "26. Li, J., Lu, M., Liu, J., Guo, Y., Du, Y., Du, L., Zhang, S.: Bev-lgkd: A unified lidarguided knowledge distillation framework for multi-view bev 3d object detection. IEEE Transactions on Intelligent Vehicles pp. 1-10 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 330, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 480, + 363 + ], + "type": "text", + "content": "27. Li, J., Luo, C., Yang, X.: Pillarnext: Rethinking network designs for 3d object detection in lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17567-17576 (June 2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "text", + "content": "28. Li, Q., Jin, S., Yan, J.: Mimicking very efficient network for object detection. In: Proceedings of the iee conference on computer vision and pattern recognition. pp. 6356-6364 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "text", + "content": "29. Li, Z., Li, Y., Wang, Y., Xie, G., Qu, H., Lyu, Z.: A lightweight model for 3d point cloud object detection. Applied Sciences 13(11) (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 420, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 480, + 453 + ], + "type": "text", + "content": "30. Liu, J., Zhao, Q., Xiong, W., Huang, T., Han, Q.L., Zhu, B.: Smurf: Spatial multi-representation fusion for 3d object detection with 4d imaging radar. IEEE Transactions on Intelligent Vehicles pp. 1-14 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 453, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 480, + 497 + ], + "type": "text", + "content": "31. Mao, J., Minzhe, N., Jiang, C., liang, h., Chen, J., Liang, X., Li, Y., Ye, C., Zhang, W., Li, Z., Yu, J., Xu, C., Xu, H.: One million scenes for autonomous driving: Once dataset. In: Vanschoren, J., Yeung, S. (eds.) Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks. vol. 1 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "type": "text", + "content": "32. Meyer, M., Kuschk, G.: Automotive radar dataset for deep learning based 3d object detection. In: 2019 16th european radar conference (EuRAD). pp. 129-132. IEEE (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 532, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 480, + 575 + ], + "type": "text", + "content": "33. Nguyen, C.H., Nguyen, T.C., Tang, T.N., Phan, N.L.: Improving object detection by label assignment distillation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1005-1014 (January 2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 576, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 480, + 597 + ], + "type": "text", + "content": "34. Nobis, F., Shafiei, E., Karle, P., Betz, J., Lienkamp, M.: Radar voxel fusion for 3d object detection. Applied Sciences 11(12) (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "text", + "content": "35. Paek, D.H., Kong, S.H., Wijaya, K.T.: K-radar: 4d radar object detection for autonomous driving in various weather conditions. In: Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "36. Palffy, A., Pool, E., Baratam, S., Kooij, J.F.P., Gavrila, D.M.: Multi-class road user detection with 3+1d radar in the view-of-delft dataset. IEEE Robotics and Automation Letters 7(2), 4961-4968 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "P. Palmer et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "37. Palmer, P., Krueger, M., Altendorfer, R., Adam, G., Bertram, T.: Reviewing 3d object detectors in the context of high-resolution " + }, + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "inline_equation", + "content": "3 + 1" + }, + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "d radar. In: Workshop on 3D Vision and Robotics at the Conference on Computer Vision and Pattern Recognition 2023 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 482, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 203 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 203 + ], + "type": "text", + "content": "38. Palmer, P., Krueger, M., Altendorfer, R., Bertram, T.: Ego-motion estimation and dynamic motion separation from 3d point clouds for accumulating data and improving 3d object detection. In: AmE 2023 - Automotive meets Electronics; 14. GMM Symposium. pp. 86-91 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 203, + 482, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 203, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 203, + 482, + 236 + ], + "type": "text", + "content": "39. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 236, + 482, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 267 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 267 + ], + "type": "text", + "content": "40. Rebut, J., Ouaknine, A., Malik, W., Pérez, P.: Raw high-definition radar for multitask learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17021-17030 (June 2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 267, + 482, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 300 + ], + "type": "text", + "content": "41. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10529–10538 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 300, + 482, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 332 + ], + "type": "text", + "content": "42. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-779 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 332, + 482, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 364 + ], + "type": "text", + "content": "43. Smith, L.N., Topin, N.: Super-convergence: Very fast training of neural networks using large learning rates. In: Artificial intelligence and machine learning for multidomain operations applications. vol. 11006, pp. 369-386. SPIE (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 364, + 482, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 396 + ], + "type": "text", + "content": "44. Stephan, M., Hazra, S., Santra, A., Weigel, R., Fischer, G.: People counting solution using an fmcw radar with knowledge distillation from camera data. In: 2021 IEEE Sensors. pp. 1-4 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 396, + 482, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 460 + ], + "type": "text", + "content": "45. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 460, + 482, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 504 + ], + "type": "text", + "content": "46. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2446-2454 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 504, + 482, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 536 + ], + "type": "text", + "content": "47. Sun, S., Zhang, Y.D.: 4d automotive radar sensing for autonomous vehicles: A sparsity-oriented approach. IEEE Journal of Selected Topics in Signal Processing 15(4), 879-891 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "type": "text", + "content": "48. Tan, B., Ma, Z., Zhu, X., Li, S., Zheng, L., Chen, S., Huang, L., Bai, J.: 3d object detection for multi-frame 4d automotive millimeter-wave radar point cloud. IEEE Sensors Journal (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 568, + 482, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 589 + ], + "type": "text", + "content": "49. Tung, F., Mori, G.: Similarity-preserving knowledge distillation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1365-1374 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 589, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 589, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 589, + 482, + 632 + ], + "type": "text", + "content": "50. Wang, L., Zhang, X., Li, J., Xv, B., Fu, R., Chen, H., Yang, L., Jin, D., Zhao, L.: Multi-modal and multi-scale fusion 3d object detection of 4d radar and lidar for autonomous driving. IEEE Transactions on Vehicular Technology 72(5), 5628-5641 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "type": "text", + "content": "51. Wang, T., Yuan, L., Zhang, X., Feng, J.: Distilling object detectors with fine-grained feature imitation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4933-4942 (2019)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 219, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 219, + 91, + 447, + 102 + ], + "type": "text", + "content": "LEROjD: Lidar Extended Radar-Only Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "52. Wei, Y., Wei, Z., Rao, Y., Li, J., Zhou, J., Lu, J.: Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. In: European Conference on Computer Vision. pp. 179-195. Springer (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "53. Wu, H., Wen, C., Shi, S., Li, X., Wang, C.: Virtual sparse convolution for multimodal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21653-21662 (June 2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "54. Xiao, A., Huang, J., Guan, D., Zhan, F., Lu, S.: Transfer learning from synthetic to real lidar point cloud for semantic segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 2795-2803 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "55. Xiao, Y., Liu, Y., Luan, K., Cheng, Y., Chen, X., Lu, H.: Deep lidar-radar-visual fusion for object detection in urban environments. Remote Sensing 15(18) (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "56. Xiong, W., Liu, J., Huang, T., Han, Q.L., Xia, Y., Zhu, B.: Lxl: Lidar excluded lean 3d object detection with 4d imaging radar and camera fusion. IEEE Transactions on Intelligent Vehicles pp. 1-14 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 314 + ], + "type": "text", + "content": "57. Xu, B., Zhang, X., Wang, L., Hu, X., Li, Z., Pan, S., Li, J., Deng, Y.: Rpfa-net: a 4d radar pillar feature attention network for 3d object detection. In: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). pp. 3061-3066 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "58. Yan, Q., Wang, Y.: Mvfan: Multi-view feature assisted network for 4d radar object detection. In: International Conference on Neural Information Processing. pp. 493-511. Springer (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "59. Yang, J., Shi, S., Ding, R., Wang, Z., Qi, X.: Towards efficient 3d object detection with knowledge distillation. In: Advances in Neural Information Processing Systems. vol. 35, pp. 21300-21313. Curran Associates, Inc. (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "60. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3dssd: Point-based 3d single stage object detector. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11040-11048 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "61. You, Y., Wang, Y., Chao, W.L., Garg, D., Pleiss, G., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-lidar++: Accurate depth for 3d object detection in autonomous driving. In: ICLR (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "62. Zamanakos, G., Tsochatzidis, L., Amanatiadis, A., Pratikakis, I.: A comprehensive survey of lidar-based 3d object detection methods with deep learning for autonomous driving. Computers & Graphics 99, 153-181 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "type": "text", + "content": "63. Zhang, J., Zhuge, H., Liu, Y., Peng, G., Wu, Z., Zhang, H., Lyu, Q., Li, H., Zhao, C., Kircali, D., Mharolkar, S., Yang, X., Yi, S., Wang, Y., Wang, D.: Ntu4dradlm: 4d radar-centric multi-modal dataset for localization and mapping. In: 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC) (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 522, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 567 + ], + "type": "text", + "content": "64. Zhang, L., Dong, R., Tai, H.S., Ma, K.: Pointdistiller: Structured knowledge distillation towards efficient and compact 3d detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21791-21801 (June 2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "text", + "content": "65. Zhang, X., Wang, L., Chen, J., Fang, C., Yang, L., Song, Z., Yang, G., Wang, Y., Zhang, X., Yang, Q., Li, J.: Dual radar: A multi-modal dataset with dual 4d radar for autonomous driving. arXiv preprint arXiv:2310.07602 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 643 + ], + "type": "text", + "content": "66. Zhang, Y., Hu, Q., Xu, G., Ma, Y., Wan, J., Guo, Y.: Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18953-18962 (June 2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "67. Zhao, M., Li, T., Abu Alsheikh, M., Tian, Y., Zhao, H., Torralba, A., Katabi, D.: Through-wall human pose estimation using radio signals. In: Proceedings of" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "P. Palmer et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 293 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "text", + "content": "the IEEE conference on computer vision and pattern recognition. pp. 7356-7365 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "type": "text", + "content": "68. Zheng, L., Li, S., Tan, B., Yang, L., Chen, S., Huang, L., Bai, J., Zhu, X., Ma, Z.: Rcfusion: Fusing 4-d radar and camera with bird's-eye view features for 3-d object detection. IEEE Transactions on Instrumentation and Measurement 72, 1-14 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "text", + "content": "69. Zheng, L., Ma, Z., Zhu, X., Tan, B., Li, S., Long, K., Sun, W., Chen, S., Zhang, L., Wan, M., Huang, L., Bai, J.: Tj4dradset: A 4d radar dataset for autonomous driving. In: 2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC). pp. 493-498 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 258 + ], + "type": "text", + "content": "70. Zhou, T., Chen, J., Shi, Y., Jiang, K., Yang, M., Yang, D.: Bridging the view disparity between radar and camera features for multi-modal fusion 3d object detection. IEEE Transactions on Intelligent Vehicles 8(2), 1523-1535 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "71. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4490-4499 (2018)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 219, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 219, + 91, + 447, + 102 + ], + "type": "text", + "content": "LEROjD: Lidar Extended Radar-Only Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LG-Gaze_ Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation/82982acf-4d4b-4f4d-a9b4-ee45c99a1bec_content_list.json b/2024/LG-Gaze_ Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation/82982acf-4d4b-4f4d-a9b4-ee45c99a1bec_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7569c20f59ef975b6218c25d5a8d471bad4a922f --- /dev/null +++ b/2024/LG-Gaze_ Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation/82982acf-4d4b-4f4d-a9b4-ee45c99a1bec_content_list.json @@ -0,0 +1,1695 @@ +[ + { + "type": "text", + "text": "LG-Gaze: Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation", + "text_level": 1, + "bbox": [ + 225, + 141, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pengwei Yin*, Jingjing Wang*, Guanzhong Zeng, Di Xie, and Jiang Zhu", + "bbox": [ + 228, + 212, + 774, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hikvision Research Institute, Hangzhou, China {yinpengwei, wangjingjing9, zengguanzhong, xiedi, zhujiang.hri}@hikvision.com", + "bbox": [ + 313, + 239, + 687, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The ability of gaze estimation models to generalize is often significantly hindered by various factors unrelated to gaze, especially when the training dataset is limited. Current strategies aim to address this challenge through different domain generalization techniques, yet they have had limited success due to the risk of overfitting when solely relying on value labels for regression. Recent progress in pre-trained vision-language models has motivated us to capitalize on the abundant semantic information available. We propose a novel approach in this paper, reframing the gaze estimation task as a vision-language alignment issue. Our proposed framework, named Language-Guided Gaze Estimation (LG-Gaze), learns continuous and geometry-sensitive features for gaze estimation benefit from the rich prior knowledge of vision-language models. Specifically, LG-Gaze aligns gaze features with continuous linguistic features through our proposed multimodal contrastive regression loss, which customizes adaptive weights for different negative samples. Furthermore, to better adapt to the labels for gaze estimation task, we propose a geometry-aware interpolation method to obtain more precise gaze embeddings. Through extensive experiments, we validate the efficacy of our framework in four different cross-domain evaluation tasks.", + "bbox": [ + 261, + 327, + 743, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Gaze Estimation $\\cdot$ Vision-Language Model $\\cdot$ Continuous Regression Task $\\cdot$ Contrastive Learning", + "bbox": [ + 259, + 603, + 740, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 654, + 375, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The gaze estimation is crucial for understanding human behavior. Precise gaze estimation offers significant assistance for many applications, such as human-computer interaction [1], augmented reality [21], and driver monitoring systems [20]. The deep learning has led to significant improvements in gaze estimation based on appearance. While these methods show impressive results in evaluations within the same domain, they tend to experience a decline in performance when evaluated on different domains. This decline is primarily attributed to overfitting rather than learning robust features from the original domain.", + "bbox": [ + 212, + 683, + 787, + 805 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally.", + "bbox": [ + 215, + 811, + 457, + 825 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cd66d2d4a0db9ec2ad0105ec28b765a233834089de243d74d429afdb2c248896.jpg", + "image_caption": [ + "(a) Traditional Methods: Supervised by Value Label", + "(b) LG-Gaze: Guided by Language Model", + "Fig. 1: (a) The traditional method of gaze generalization involves overseeing model training by means of numerical label regression. (b) In our study, we introduce a text models to steer the development of robust features in visual models." + ], + "image_footnote": [], + "bbox": [ + 225, + 148, + 781, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gaze data encompass diverse elements, such as appearance, wearables, environments, and image clarity [31, 34], yet gaze annotations predominantly focus on eye direction, marginalizing other variables as noise. Such irrelevant factors exacerbate the domain disparity, hindering gaze models' adaptability. Additionally, supervisory numeric labels might induce overfitting, influenced by these extraneous gaze components.", + "bbox": [ + 212, + 316, + 782, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Figure 1(a), training a gaze model typically involves utilizing a vision encoder to extract features and then employing a regressor for prediction. Traditional methods train a gaze model by using data disturbing [31], feature enhancement [30], and adversarial learning [6] to improve the generalization ability of the gaze model. However, these regularization techniques have limited effectiveness as they rely solely on numerical labels for supervision, making them susceptible to overfitting due to irrelevant factors. The above may indicate that it is difficult to guide the model to learn ideal robust features through these indirect regularization terms.", + "bbox": [ + 212, + 411, + 784, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To counter overfitting in gaze representation learning, we advocate for multimodal integration. Language, rich in semantics and knowledge, complements visual data. Each gaze label is described textually, \"the yaw/pitch degree angle of the person is {yaw/pitch}.\". Leveraging visual-language models, such as CLIP, which excel in generalized latent space learning from vast image-text pairs, we align image features with a language space. This distillation of CLIP's language expertise acts as regularization, enhancing generalization. However, these methods are only applicable to a limited set of discrete scalars and are not suitable for continuous tasks such as gaze estimation. Aligning gaze features to discrete text features can result in inaccurate feature alignment, which in turn affects the performance of continuous regression tasks.", + "bbox": [ + 212, + 550, + 787, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we propose a novel training method named LG-Gaze, which is designed to learn continuous and geometry-sensitive features for gaze estimation, as depicted in Figure 1(b). This framework aligns gaze features with continuous linguistic features extracted by a powerful language model, which not only prevents model overfitting but also enhances the generalization capabilities of gaze model. Primarily, unlike common discrete multimodal contrastive learning methods [15, 22], we introduce a loss function for continuous multimodal contrastive learning. Our Multimodal Contrastive Regression loss", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Pengwei et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "function (MCR) customizes adaptive weights for different negative samples according to label distance, facilitating more refined feature alignment in the feature space and benefiting the model in learning more continuous features. Next, to better adapt to the vectorial property of labels for gaze estimation task, we propose a geometry-aware interpolation method to obtain more precise gaze embeddings. The geometry-aware interpolation method combines spherical [26] and bilinear interpolation techniques to resolve the issue of creating semantic labels for gaze, ensuring accurate gaze prompts. Furthermore, MCR also resolves the limitations of previous contrastive learning functions by utilizing a more uniform distribution of global negative samples for contrast. Unlike most methods that are limited to intra-batch comparisons, our loss function benefits from extensive global contrast, resulting in a reasonable and robust feature distribution.", + "bbox": [ + 212, + 146, + 787, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main contributions can be summarized as follows:", + "bbox": [ + 240, + 328, + 632, + 340 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- In this paper, we propose a novel training framework called LG-Gaze for gaze estimation. LG-Gaze guides the gaze model to learn generalized representations by leveraging textual features extracted from language models, resulting in more robust gaze representations. By incorporating rich language knowledge, this approach achieves robust gaze features, ultimately enhancing the cross-domain generalization capability of the gaze model.", + "- To learn continuous feature for regression tasks, we propose a new multimodal contrastive learning loss function named MCR. The method employs adaptive weights for different negative sample, helping learn more refined features. Moreover, MCR leverages ample global negative samples for contrast, leading to a reasonable feature distribution.", + "- We introduce a geometry-aware interpolation method that employs spherical interpolation techniques to compute accurate gaze embeddings, thereby ensuring the accuracy of semantic labels.", + "- Experimental results and visualizations demonstrate that LG-Gaze not only achieves remarkable performance compared to the baseline but also surpasses state-of-the-art domain generalization methods for gaze estimation." + ], + "bbox": [ + 225, + 352, + 782, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 632, + 395, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Gaze Estimation", + "text_level": 1, + "bbox": [ + 215, + 664, + 398, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Appearance-based gaze estimation has garnered attention in recent years [7,8, 13,38]. However, it faces challenges in cross-domain evaluation due to the domain gap resulting from various gaze-irrelevant factors. Common approaches often rely on diverse gaze datasets [12,37] for training, aiming to equip models with robust generalization capabilities. Unfortunately, collecting gaze data remains costly, and the diversity of available data remains limited. To address this issue, we propose enhancing gaze estimation models through domain generalization (DG) methods. These methods enable models to generalize to unseen distributions and enhance cross-domain performance. Most existing DG techniques are tailored for classification tasks, leaving a gap in the context of gaze estimation.", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "LG-Gaze: Language-Guided Gaze Estimation", + "bbox": [ + 429, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PureGaze [6] debuts a self-adversarial schema, discarding irrelevant gaze cues, enhancing pertinent ones. Xu et al. [31] and NeRF-Gaze [33] similarly counteract gaze distractions via adversarial data modifications and augmentation. However, residual gaze-confounding factors persist, hindering estimation precision. CLIP-Gaze [34] is designed to segregate gaze features from predefined textual gaze-irrelevant features, thereby enhancing the generalization capability and robustness of gaze features. Our LG-Gaze, by directly aligning with language features, also exhibits resilience against gaze-irrelevant factors. More critically, the text features in LG-Gaze possess favorable rank properties, attributed to the geometry-aware continuous prompts learning.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Vision Language Model", + "text_level": 1, + "bbox": [ + 215, + 318, + 457, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In recent research, several studies have utilized the text manipulation and visual alignment capabilities of CLIP (Contrastive Language-Image Pretraining) [22] to improve opendetection and generalization performance in specific tasks. Notable examples include DetCLIP [32], CLIP-Gap [28], and CLIP-Cluster [25]. To further enhance the performance of vision-language models on downstream tasks, an effective strategy involves learning more contextually appropriate prompts through text prompt tuning [40, 41]. Additionally, DenseCLIP [24] explores the application of semantic knowledge into monocular depth estimation tasks. By matching visual features with textual features, this method achieves zero-shot monocular depth estimation. \"Teaching CLIP to Count to Ten\" and CrowdCLIP propose visual language models for understanding quantities, enabling object counting and crowd estimation. While these methods enhance models' understanding of quantities, they do not address regression tasks. OrdinalCLIP [15] and L2RCLIP [29] are capable of addressing ordinal regression tasks. However, they are not as effective for continuous regression tasks. In this paper, our method aligns gaze features with continuous linguistic features through our proposed multimodal contrastive regression loss (MCR) and the geometry-aware interpolation method.", + "bbox": [ + 212, + 342, + 787, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Contrastive Learning", + "text_level": 1, + "bbox": [ + 215, + 633, + 437, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In computer vision, SimCLR [4] and CLIP [22] pioneer contrastive loss, exceeding in image classification and retrieval. Yet, these techniques aren't suited for regression, where prediction targets are continuous, not categorical. Rank-N-Contrast (RNC) [35] employs sample rankings to enhance continuity understanding. Nonetheless, RNC presumes infinite distance between negative and positive pairs, neglecting the correlation between label and feature proximities. Contrastive Domain Generalization (CDG) [30] leverages contrastive loss to encourage the clustering of features corresponding to similar gaze directions while separating those associated with significantly different gaze directions. However, CDG relies on empirical thresholding to define positive and negative sample pairs, which limits its applicability for continuous regression tasks. In our research, we introduce a novel contrastive loss function specifically designed for", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Pengwei et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "continuous regression tasks. Our approach prioritizes sample ordering while considering label distances. Unlike previous contrastive learning methods that are confined to constructing sample pairs within batches, our method demonstrates a more targeted enhancement. To elaborate further, we draw inspiration from the Momentum Contrast (MoCo) framework [11], which utilizes a queue to store negative sample features. However, these negative samples are obtained from n steps back in time and exhibit inherent uncontrollable randomness. Unlike delayed updates, our approach synchronously updates these samples, ensuring a controlled distribution instead of relying on randomness.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 304, + 330, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem Statement", + "text_level": 1, + "bbox": [ + 214, + 334, + 423, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the context of gaze estimation tasks, which fundamentally belong to regression tasks, we define the data for the source domain as $\\mathcal{D}_s = (\\pmb {I}_i,\\pmb {g}_i)_{i = 1}^M$ , where $(\\pmb {I}_i,\\pmb {g}_i)$ represents the $i$ -th pair of image $\\pmb {I}_i\\in \\mathbb{R}^{224\\times 224\\times 3}$ and their corresponding gaze direction $\\pmb {g}_i\\in \\mathbb{R}^3$ . Here, $M$ denotes the total number of pairs in $\\mathcal{D}_s$ . Our objective in gaze estimation is to train a neural network composed of an image encoder $\\mathcal{E}_v(\\cdot):\\mathbb{R}^{224\\times 224\\times 3}\\to \\mathbb{R}^d$ to extract gaze representations $\\pmb{f}_i^g$ based on input image $\\mathcal{I}_i$ and an MLP regressor $\\mathcal{R}_g(\\cdot):\\mathbb{R}^d\\to \\mathbb{R}^3$ to predict gaze directions $\\pmb {g}_i$ . Typically, once we have trained a robust gaze representation, training $\\mathcal{R}_g$ becomes relatively straightforward using regression functions (e.g., $L_{1}$ or AngleLoss).", + "bbox": [ + 212, + 357, + 787, + 493 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, gaze estimation research has encountered a persistent challenge that hinders its advancement: How can we design robust gaze representations $\\{f_i^g\\}_{i=1}^M$ that are continuous, semantic, and sensitive to geometric factors for gaze estimation?", + "bbox": [ + 214, + 494, + 787, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although many previous methods explore additional technologies to enhance feature representations, such as adversarial training [6], contrastive learning as an auxiliary task [30] and mitigating the impact of gaze-irrelevant factors [31], they often focus solely on finding regularization terms to improve model performance without explicitly aiming for robust and well-distributed features. Consequently, existing methods often face overfitting issues and struggle to achieve the desired performance. In this paper, we propose a novel language-guided gaze representation learning framework that leverages rich prior knowledge from a natural language model. Our approach aims to learn robust and well-distributed gaze representations, addressing the aforementioned challenges.", + "bbox": [ + 212, + 555, + 787, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 LG-Gaze Framework", + "text_level": 1, + "bbox": [ + 214, + 726, + 433, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To harness the full power of language, we utilize the text encoder from CLIP as our language model. CLIP adopts a vision-language pre-training framework and learns representations from image-text pairs, thereby constructing a joint vision-language latent space. Given the strong performance of the CLIP model in downstream tasks, we implemented our method using the original CLIP text encoder from the referenced paper.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "LG-Gaze: Language-Guided Gaze Estimation", + "bbox": [ + 429, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a7a4efe40ad6232923f5142fce103ccd87432d55edecb724f1e4d76b100d68ce.jpg", + "image_caption": [ + "Fig. 2: We reformulate the task as an image-language matching problem, which mainly consists of a trainable prompt, a frozen text encoder, a trainable image encoder." + ], + "image_footnote": [], + "bbox": [ + 225, + 146, + 781, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Suppose that each training step is provided with a data batch $(I_i, g_i)_{i=1}^B$ , where $g_i$ represents the gaze label of the image $I_i$ . Our goal is to construct semantic labels and use a language model to extract text features. This process will help the gaze model learn robust features by aligning them with the extracted features. To leverage the prior knowledge within text features, we begin by constructing a learnable prompt $\\{\\pmb{v}_1, \\pmb{v}_2, \\dots, \\pmb{v}_{L-1}\\} \\in \\mathbb{R}^{512 \\times (L-1)}$ as the context for describing gaze estimation tasks instead of handcrafting the prompt context. Simultaneously, we treat gaze labels as individual words. For each sample, we concatenate the learnable prompt with the embedding $\\pmb{v}_i^g \\in \\mathbb{R}^{512}$ to form a sequence $\\pmb{T}_i = \\{\\pmb{v}_1, \\pmb{v}_2, \\dots, \\pmb{v}_{L-1}, \\pmb{v}_i^g\\}$ as a semantic label. Specifically, we utilize the gaze embedding as the final token in the sequence $\\pmb{T}_i \\in \\mathbb{R}^{512 \\times L}$ .", + "bbox": [ + 212, + 315, + 787, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To ensure the continuity, ordinality, and geometric properties of gaze label embeddings, we introduce $N$ learnable embedding anchors, denoted as $\\{\\pmb{A}_j\\}_{j=1}^N$ . Each anchor corresponds to a unique gaze vector, uniformly distributed across the entire label space. We utilize a nearest-neighbor interpolation method to represent gaze label embeddings associated with sample relationships. This approach yields a continuous and geometry-aware representation of gaze embeddings, as detailed in Section 3.3.", + "bbox": [ + 212, + 484, + 787, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our research, we send image $\\pmb{I}_i$ to an image encoder $\\mathcal{E}_v$ to extract the image feature $\\pmb{f}_i^g$ . Consequently, the features of a batch can be represented as $\\{\\pmb{f}_i^g\\}_{i=1}^B$ . Simultaneously, we utilize the language model $\\mathcal{E}_t$ to obtain corresponding text features $\\{\\pmb{f}_i^t\\}_{i=1}^B$ from $\\{\\pmb{T}_i\\}_{i=1}^B$ . While the parameters of the language model remain frozen, we train the entire image encoder to align image features with the language feature space. Subsequently, we utilize an image-text contrastive loss to optimize the network, which includes both an image-to-text loss and a text-to-image loss, following the methodology of CLIP. Detailed information about our proposed contrastive learning method will be provided in Section 3.4.", + "bbox": [ + 212, + 589, + 787, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After aligning the features, we obtain robust gaze representations. Finally, the image features are fed into a regressor denoted as $\\mathcal{R}_g$ , yielding the predicted gaze direction $\\hat{g}_i \\in \\mathbb{R}^3$ . Specifically, we predict through supervised learning, which is defined as:", + "bbox": [ + 212, + 724, + 787, + 785 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {G a z e} \\left(\\hat {\\boldsymbol {g}} _ {i}, \\boldsymbol {g} _ {i}\\right) = \\operatorname {a r c c o s} \\left(\\frac {\\hat {\\boldsymbol {g}} _ {i} \\cdot \\boldsymbol {g} _ {i}}{\\| \\hat {\\boldsymbol {g}} _ {i} \\| \\| \\boldsymbol {g} _ {i} \\|}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 796, + 785, + 830 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Pengwei et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/033aec55813a9be3dcbeefff1f5884736d6f21eb0a75c9499efee0d931465aed.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 250, + 143, + 385, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e9f7f381ffcd0d676f276b5cf23cdb942d4f2526e59ae15c03c38aff14e99528.jpg", + "image_caption": [ + "(b)", + "Fig. 3: These four subfigures contain all steps of our proposed interpolation method." + ], + "image_footnote": [], + "bbox": [ + 388, + 148, + 506, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/905df1fe32afa36bb6b3a60891a104df9ba4790a79720c2356de9b9845d9fe85.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 511, + 146, + 630, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6c28bcfb26141a4c535ef31a63ca03f62e0693a6cded3393a3622e3df3d46d40.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 635, + 146, + 754, + 237 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Geometry-aware Continuous Gaze Prompts", + "text_level": 1, + "bbox": [ + 214, + 310, + 622, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, to better represent gaze labels in the embedding space, we construct a gaze embedding for each label. This embedding should have the capability for continuous representation and geometric attributes. To address the conflict between the finite vocabulary in natural language processing and the infinite nature of gaze labels, we interpolate each gaze embedding $\\pmb{v}_i^g$ through $N$ gaze anchor embeddings $\\{\\pmb{A_j}\\}_{j = 1}^N$ .", + "bbox": [ + 212, + 332, + 785, + 424 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Common approaches, such as those described in [15], employ global linear interpolation to represent $\\pmb{v}_i^g$ , which interpolates the target embedding using all anchor embeddings. The weight $w_{i,j}$ for each anchor embedding is computed by Equation 2, determined by the similarity between its corresponding gaze vector $\\pmb{g}_j$ and the target $\\pmb{g}_i$ . However, global interpolation cannot accurately represent $\\pmb{v}_i^g$ because anchors that are not proximate fail to provide effective information.", + "bbox": [ + 214, + 424, + 787, + 516 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i, j} = \\frac {\\cos \\left(\\boldsymbol {g} _ {i} , \\boldsymbol {g} _ {j}\\right)}{\\sum_ {j = 1} ^ {N} \\cos \\left(\\boldsymbol {g} _ {i} , \\boldsymbol {g} _ {j}\\right)} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 523, + 785, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further enhance interpolation, we attempted to employ a bilinear interpolation method. However, as it still utilizes a linear representation, this method remains insufficiently precise in expressing the geometric relationships between gaze vectors.", + "bbox": [ + 212, + 566, + 785, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To address this issue, we decided to adopt a spherical interpolation method [26], which can accurately calculate the interpolation weights between vectors. As illustrated in Figure 3(a), a flat spherical interpolation method is depicted. For the vector $\\pmb{g}_i$ positioned between $\\pmb{g}_1$ and $\\pmb{g}_2$ , we initially compute $\\theta = \\arccos(\\langle \\pmb{g}_1, \\pmb{g}_2 \\rangle)$ and $\\theta_1 = \\arccos(\\langle \\pmb{g}_1, \\pmb{g}_i \\rangle)$ . Subsequently, we compute the scalar $t$ by $t = \\theta_1 / \\theta$ , then we apply the spherical interpolation Equation 3 to determine the corresponding weights for $\\pmb{g}_1$ and $\\pmb{g}_2$ . Finally, we can represent the target vector $\\pmb{g}_i$ through the summation of these weights. Ultimately, we confirm that this method of weight calculation is accurate.", + "bbox": [ + 214, + 628, + 787, + 763 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {g} _ {i} = \\frac {\\sin ((1 - t) \\theta)}{\\sin (\\theta)} \\boldsymbol {g} _ {1} + \\frac {\\sin (t \\theta)}{\\sin (\\theta)} \\boldsymbol {g} _ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 773, + 785, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Building upon the aforementioned spherical interpolation method, we further integrate the concept of bilinear interpolation to derive the interpolation", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "LG-Gaze: Language-Guided Gaze Estimation", + "bbox": [ + 428, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "weights for each gaze label on a stereographic sphere. As depicted in Figure 3(b), we initially identify the four nearest anchor embeddings $A_{1-4}$ based on their proximity. Subsequently, as shown in Figure 3(c), we calculate the weights $w_{a,1}$ and $w_{a,2}$ for $A_1$ and $A_2$ using Equation 3. Similarly, we can determine $w_{b,3}$ and $w_{b,4}$ . In Figure 3(d), we proceed to obtain $w_{i,a}$ and $w_{i,b}$ using a similar method. Ultimately, we acquire the associated weights for $A_{1-4}$ to interpolate $\\pmb{v}_i^g$ , which can be expressed as follows:", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {i} ^ {g} = \\left(w _ {i, a} * w _ {a, 1}\\right) \\cdot \\boldsymbol {A} _ {1} + \\left(w _ {i, a} * w _ {a, 2}\\right) \\cdot \\boldsymbol {A} _ {2} + \\left(w _ {i, b} * w _ {b, 3}\\right) \\cdot \\boldsymbol {A} _ {3} + \\left(w _ {i, b} * w _ {b, 4}\\right) \\cdot \\boldsymbol {A} _ {4} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 265, + 785, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, to enhance the geometric relationships between gaze embeddings, we employ a simple loss function to constrain anchor embeddings $\\{A_j\\}_{j=1}^N$ . As shown in Equation 5, the similarity between anchor embeddings should be consistent with their relationships in the label space.", + "bbox": [ + 212, + 294, + 787, + 354 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {G e o}} = \\frac {1}{M * M} \\sum_ {i = 1} ^ {M} \\sum_ {j = 1} ^ {M} L _ {1} \\left(\\cos \\left(\\boldsymbol {A} _ {i}, \\boldsymbol {A} _ {j}\\right), \\cos \\left(\\boldsymbol {g} _ {i}, \\boldsymbol {g} _ {j}\\right)\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 367, + 785, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Continuous Multimodal Contrastive Regression Loss Function", + "text_level": 1, + "bbox": [ + 214, + 431, + 772, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mathematically, from the aforementioned steps, we obtain image features $\\{\\pmb{f}_i^g\\}_{i=1}^B$ and their corresponding text features $\\{\\pmb{f}_i^t\\}_{i=1}^B$ . Both CLIP and OrdinalCLIP utilize the InfoNCE loss to align the multimodal features. To delve into specifics, consider an image-text feature pair $(\\pmb{f}_i^g, \\pmb{f}_i^t)$ . Here, $\\pmb{f}_i^g / \\pmb{f}_i^t$ serves as the positive sample for $\\pmb{f}_i^t / \\pmb{f}_i^g$ , as they are associated with the same gaze label. Meanwhile, all other image/text features in a mini-batch are treated as negative samples and are consequently pushed away from $\\pmb{f}_i^t / \\pmb{f}_i^g$ . However, this training objective poses challenges for regression tasks because it overlooks the finer-grained distance ordering relationships between samples. Consequently, failing to differentiate the relative distances between image/text and their corresponding negative samples will inevitably weaken the learning effectiveness, especially in cross-modal representation learning for regression tasks.", + "bbox": [ + 212, + 455, + 787, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Considering the proximity relationships between data, for the $i$ -th sample, we introduce a weight parameter $\\{w(i,j)\\}_{j=1}^{B}$ for each negative sample. This parameter is computed based on the distances between labels of samples. Our proposed multimodal contrastive regression loss (MCR) effectively controls the feature distances among samples, preventing negative samples from being excessively contracted. As a result, we learn a semantically meaningful representation space tailored for regression tasks, ensuring consistency between sample feature relationships and the label space. The proposed text-to-image MCR loss is formulated as follows.", + "bbox": [ + 212, + 638, + 787, + 773 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {t 2 i} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\left(\\frac {\\exp \\left(\\sin \\left(\\boldsymbol {f} _ {i} ^ {t} , \\boldsymbol {f} _ {i} ^ {g}\\right)\\right)}{\\exp \\left(\\sin \\left(\\boldsymbol {f} _ {i} ^ {t} , \\boldsymbol {f} _ {i} ^ {g}\\right)\\right) + \\sum_ {j = 1} ^ {B} w (i , j) \\cdot \\exp \\left(\\sin \\left(\\boldsymbol {f} _ {i} ^ {t} , \\boldsymbol {f} _ {j} ^ {g}\\right)\\right)}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 785, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Pengwei et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\text{sim}(\\pmb{f}_i^t, \\pmb{f}_i^g) = \\cos(\\pmb{f}_i^t, \\pmb{f}_i^g)$ and $w(i,j) = \\cos(\\pmb{g}_i, \\pmb{g}_j)$ denotes the contrastive weight of the $j$ -th negative image sample with respect to the $i$ -th text feature in our LG-Gaze framework. The weight parameter should be directly proportional to the label distance between negative samples. As the distance between labels increases, the penalty on negative sample weights should also increase. This encourages feature embeddings to adhere to real-world relationships.", + "bbox": [ + 212, + 145, + 782, + 251 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The MCR loss for image-to-text can be formulated in a similar manner. However, what sets it apart is our deliberate expansion of the number of negative samples to facilitate the learning of a robust representation. To achieve this, we draw inspiration from MoCo, which utilizes a dynamic queue to maintain a diverse set of negative samples, thereby enhancing the effectiveness of contrastive learning. Nevertheless, this approach still faces some challenges. For instance, if the queue update rate is too slow, the negative samples stored in the queue may become outdated and fail to align with the current model. Additionally, due to the inherent randomness and limited diversity of negative samples, there could be discrepancies between the queue data and the entire data distribution. To address this issue, we utilize the flexibility of language labels by creating $K$ global text features $\\{f_i^t\\}_{i=1}^K$ as negative samples (where $K$ is significantly larger than $B$ ). These text labels correspond to gaze vectors that show a uniformly dense distribution in the label space. Furthermore, these $K$ text labels are interpolated based on $N$ anchor $\\{A_j\\}_{j=1}^N$ . As a result, the proposed image-to-text MCR loss can be formulated as follows:", + "bbox": [ + 212, + 252, + 784, + 492 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {i 2 t} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\left(\\frac {\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {f} _ {i} ^ {g} , \\boldsymbol {f} _ {i} ^ {t}\\right)\\right)}{\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {f} _ {i} ^ {g} , \\boldsymbol {f} _ {i} ^ {t}\\right)\\right) + \\sum_ {j = 1} ^ {B + K} w (i , j) \\cdot \\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {f} _ {i} ^ {g} , \\boldsymbol {f} _ {j} ^ {t}\\right)\\right)}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 503, + 784, + 559 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Obviously, we can observe that the difference between $\\ell_{i2t}$ and $\\ell_{t2i}$ lies in the number of negative samples. We concatenate $K$ global negative samples with the $B$ negative samples from mini-batches to form the complete set of negative samples. Through this operation, the global negative samples remain relatively fixed, mitigating issues related to randomness. Additionally, their uniform distribution helps alleviate discrepancies between the negative samples and the entire data distribution. Furthermore, by computing features based on the current model, we address potential outdated information. Finally, our total MCR loss can be denoted as:", + "bbox": [ + 212, + 560, + 784, + 694 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {M C R} = \\ell_ {i 2 t} + \\ell_ {t 2 i} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 436, + 698, + 784, + 714 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Overall Objective", + "text_level": 1, + "bbox": [ + 215, + 737, + 406, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In summary, the overall objective implemented in our framework is:", + "bbox": [ + 214, + 766, + 702, + 781 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {1} \\ell_ {G e o} + \\lambda_ {2} \\ell_ {M C R} + \\lambda_ {3} \\ell_ {G a z e}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 796, + 617, + 810 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "$\\lambda_1, \\lambda_2, \\lambda_3$ are hyperparameters, and we empirically set $\\lambda_1 = \\lambda_2 = \\lambda_3 = 1.0$ .", + "bbox": [ + 238, + 824, + 776, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "LG-Gaze: Language-Guided Gaze Estimation", + "bbox": [ + 428, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experiment Details", + "text_level": 1, + "bbox": [ + 215, + 176, + 423, + 193 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Data Preparation We evaluate gaze estimation method on four cross-domain tasks, utilizing ETH-XGaze [37] and Gaze360 [12] as the training datasets, and MPIIFaceGaze [39] and Eye-Diap [9] as the test datasets. We denote them as $\\mathcal{D}_{\\mathrm{E}}$ (ETH-XGaze) $\\rightarrow \\mathcal{D}_{\\mathrm{M}}$ (MPIIFaceGaze), $\\mathcal{D}_{\\mathrm{E}} \\rightarrow \\mathcal{D}_{\\mathrm{D}}$ (EyeDiap), $\\mathcal{D}_{\\mathrm{G}}$ (Gaze360) $\\rightarrow \\mathcal{D}_{\\mathrm{M}}$ , $\\mathcal{D}_{\\mathrm{G}} \\rightarrow \\mathcal{D}_{\\mathrm{D}}$ . See the supplementary material for more details.", + "bbox": [ + 215, + 204, + 785, + 279 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Vision Model Implementation Details We use ResNet-18 as our gaze feature extractor $\\mathcal{E}_v$ , and a fully connected layer to regress a 3-dimensional gaze vector. We resize and normalize all the images to $224\\times 224$ pixels and scale the pixel values between 0 and 1. We set the batch size to 64. We maintain the same setup for training the source domain models on $\\mathcal{D}_E$ and $\\mathcal{D}_G$ .", + "bbox": [ + 215, + 303, + 785, + 380 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Language Model Implementation Details The architecture of the text encoder is adapted from the Transformer [27], incorporating modifications outlined by Radford [23]. The dimensions for both text and gaze features have been standardized to 1024. Our models are constructed using the foundational open-source code of CLIP. The class token is strategically placed at the terminal position of the sequence, with the context tokens quantity designated as $L = 10$ . Additionally, we set the yaw angle range for the anchors to be $[-180^{\\circ}, +179^{\\circ}]$ and the pitch range to be $[-90^{\\circ}, +90^{\\circ}]$ with a step size of $30^{\\circ}$ , resulting in a total of 91 anchors. We set the number of negative samples to 256, which are uniformly distributed in the gaze label space using the Fibonacci sphere algorithm [10].", + "bbox": [ + 215, + 402, + 785, + 555 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training Implementation Details We conducted the experiments on a single Tesla V100 GPU. Specifically, we use the SGD optimizer with Nesterov momentum, a learning rate (LR) of $5 \\times 10^{-2}$ and a weight decay of $1 \\times 10^{-5}$ for the parameters. We train for 30 epochs using a Cosineannealing LR scheduler [18] with a 3-epoch warm-up. We apply a data augmentation technique involving a random color field and grayscale, as described in [5].", + "bbox": [ + 215, + 577, + 785, + 669 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Performance Comparison with SOTA Gaze Estimation Methods", + "text_level": 1, + "bbox": [ + 215, + 691, + 785, + 708 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparison of Cross-domain Evaluation. Table 1 presents the quantitative results of four cross-domain tests. The second row compares our LG-Gaze with the SOTA domain DG methods for gaze estimation. The CNN baseline method refers to the one that relies solely on the vision encoder to extract the single visual modality feature and uses $\\ell_{gaze}$ for supervision. In summary, the LG-Gaze achieves the best overall performance and demonstrates state-of-the-art results on three cross-domain evaluation tasks. It also achieves the second-best performance for $\\mathcal{D}_{\\mathrm{E}} \\rightarrow \\mathcal{D}_{\\mathrm{M}}$ , highlighting the effectiveness of our method.", + "bbox": [ + 215, + 718, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Pengwei et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/88e9b38758c892e5fc8df110bd261b1dd1525e2e3d94db24a1e67580b18b1acf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskMethods| Dt|DE→DMDE→DdDG→DMDG→DdAvg
DGCNN Baseline08.479.327.548.938.57
PureGaze [6]07.087.489.289.328.29
CDG [30]‡06.737.957.037.277.25
Xu et al. [31]06.507.447.559.037.63
CLIP-Gaze [34]06.417.516.897.066.97
Our LG-Gaze06.457.226.836.866.84
UDAPnP-GA * [17]105.535.876.187.926.38
RUDA [2]1005.706.296.205.866.01
CRGA [30]>05.485.665.896.495.88
LatentGaze [14]1005.217.81--6.51
Liu et al. [16]1005.356.627.188.616.94
UnReGA [3]1005.115.705.425.805.51
", + "bbox": [ + 222, + 143, + 779, + 364 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 1: Comparison with state-of-the-art methods. We report the results by angular error in degrees, and use bold and underline to indicate the best and the second best result in each column for a specific task. $\\ddagger$ means the model uses ResNet-50 as backbone, * shows that the experimental settings are different.", + "bbox": [ + 215, + 364, + 782, + 419 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/dfde9f82953867fdeba0ccd0616215f08bb787a8a8d4f6480a6bd00e1f0f9d64.jpg", + "image_caption": [ + "(a) CNN Baseline" + ], + "image_footnote": [], + "bbox": [ + 276, + 450, + 380, + 531 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6902da96e17fd5701284d8e5e5c233712fa637d786bb190822914e6fadf1f03f.jpg", + "image_caption": [ + "(b) PureGaze", + "Fig. 4: Illustration of the feature distribution. Various colors indicate different gaze directions and similar gaze directions have similar colors. (Recommend viewing in color)." + ], + "image_footnote": [], + "bbox": [ + 392, + 450, + 496, + 531 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4e04de218db6f8d6f6a3ed8d34d44f77d9a314296542b33d1f9fb892e692f6ed.jpg", + "image_caption": [ + "(c) CDG" + ], + "image_footnote": [], + "bbox": [ + 508, + 450, + 611, + 530 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/783a01144f579992f4f82be927a111de510155cb42ed23d131d2c6bb09634d92.jpg", + "image_caption": [ + "(d) Ours" + ], + "image_footnote": [], + "bbox": [ + 620, + 450, + 723, + 529 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In addition, we present a comparison with SOTA unsupervised domain adaptation (UDA) methods in the third row of Table 1. Note that UDA methods require the use of a small batch of unlabeled target domain samples, which is more costly than DG methods. It can be clearly seen from the table that our LG-Gaze achieves comparable performance without extra target domain data. Especially on the two tasks of $\\mathcal{D}_{\\mathrm{G}} \\to \\mathcal{D}_{\\mathrm{M}}$ and $\\mathcal{D}_{\\mathrm{G}} \\to \\mathcal{D}_{\\mathrm{D}}$ , it is close to UDA methods. Therefore, this verifies the powerful performance of our proposed method, as it does not rely on any prior information from the target domain.", + "bbox": [ + 212, + 614, + 784, + 734 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison of Feature Visualization. To evaluate and analyze the gaze features $f^g$ extracted by different models, we follow the same scheme as CDG [30], and visualize the gaze features of the cross-domain task $\\mathcal{D}_{\\mathrm{G}} \\to \\mathcal{D}_{\\mathrm{D}}$ using t-SNE [19] to compare different methods. Figure 4 shows the visualization results from all DG methods in the second row of Table 1, where feature points with", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "LG-Gaze: Language-Guided Gaze Estimation", + "bbox": [ + 429, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/1d1340db530c76f92acad24e14d16967ab73804af8e50a775ca080aa14ee4742.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2: Comparisons to SOTA contrastive learning methods. Bold indicates the best results in each column, and underline denote the second best result in each column." + ], + "table_body": "
MethodsDE→DMDE→DdDG→DMDG→DdAvg
CNN Baseline8.359.667.589.018.65
Vanilla CLIP [22]82.4379.6482.4379.6481.04
CoOp [41]9.2311.0613.1510.8511.07
OridinalCLIP [15]7.8810.119.129.749.21
Our LG-Gaze6.457.226.836.866.84
", + "bbox": [ + 222, + 143, + 785, + 250 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "similar gaze labels have similar colors. For the baseline model, the features of different gaze labels are mixed together, which is unreasonable for the regression task as shown in Figure 4 (a). The ideal feature distribution should be continuous and gradual with labels for the regression task [36]. Figure 4 (b,c) represents several classic DG methods that have been enhanced compared to the baseline, but they exhibit unreasonable feature distributions. In general, LG-Gaze has the most reasonable feature distribution, and the visualization result is shown in Figure 4 (d). The strong correlation between gaze direction and features indicates that guiding visual feature learning through language is effective.", + "bbox": [ + 212, + 321, + 787, + 460 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Performance Comparison with SOTA VLM Regression Methods", + "text_level": 1, + "bbox": [ + 215, + 491, + 785, + 507 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2 presents the performance of various methods that employ VLM models. To investigate the applicability of ordinal regression methods to gaze estimation, we discretize the gaze labels into integers. This mapping allows the yaw and pitch in the gaze label to be categorized into a finite number of categories. During testing, we predict yaw and pitch separately, but the accuracy of the gaze vector, which is composed of yaw and pitch, is evaluated.", + "bbox": [ + 212, + 525, + 787, + 617 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b5197bdd03a284cf665c291e27132f65a8f5ad9f71f232f9a0ced0ff056a5475.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3: Comparison of different text prompt tuning methods for gaze model cross-domain evaluation in four tasks. The rest of the settings are consistent for all methods, except for the contrastive loss function. Bold indicates the best results in each column, and underline denote the second best result results in each column." + ], + "table_body": "
MethodsDE→DMDE→DdDG→DMDG→DdAvg
CNN Baseline8.359.667.589.018.65
KL-Loss6.777.366.977.087.05
CDG [30]6.987.457.147.187.19
RNC [35]6.637.286.996.966.97
MoCo [11]6.527.336.957.046.96
MCR6.457.226.836.866.84
", + "bbox": [ + 233, + 645, + 774, + 768 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Pengwei et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For Vanilla CLIP, there is no need to use source data for model training. Instead, a zero-shot method is used for prediction, resulting in only two outcomes. First, we construct the text prompt \"gaze estimation: the yaw/pitch degree angle of the person is {yaw/pitch}.\" Then, we obtain the model's prediction of yaw and pitch by maximizing the alignment between text features and image features, which is essentially in line with the original image-text matching. However, we can observe from Table 2 that Vanilla CLIP exhibits poor performance, possibly due to its struggle in distinguishing ordinal concepts. For CoOp, the gaze prediction performance can be greatly enhanced through text prompt tuning, but it is weaker than the CNN baseline. This could be due to its inability to learn the ordinal relationship between samples. For OrdinalCLIP, although it can learn the ordinal relationship, the discretization of labels may lead to errors. Additionally, following the methodology of OrdinalCLIP, we have created a visualization of the semantic labels constructed by different VLM methods in the supplementary material.", + "bbox": [ + 217, + 146, + 787, + 372 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Performance Comparison with SOTA Contrastive Regression Loss Functions", + "text_level": 1, + "bbox": [ + 215, + 393, + 766, + 422 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 3 presents various methods for feature alignment or contrastive learning in gaze estimation. For KL-Loss, it primarily constrains the similarity matrix between gaze features and text features, as well as the similarity matrix composed of gaze labels. We follow the ideas of CDG [30], RNC [35], and MoCo [11], and transform them into image-to-text and text-to-image contrastive loss functions similar to LG-Gaze. From Table 3, we can see that our MCR loss function outperforms other methods significantly, attributed to the inclusion of constructed global negative samples and the weight coefficient for negative samples.", + "bbox": [ + 215, + 433, + 787, + 555 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 575, + 388, + 590 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of Our Framework. Table 4 shows the performance of different combinations of loss functions. Keeping only $\\ell_{Gaze}$ is equivalent to CNN baseline, and combining all the losses is equivalent to full LG-Gaze. Compared with CNN baseline, using MCR loss to learn robust features with language guidance can significantly improve the performance. Further adding $\\ell_{Geo}$ enhances the geometric relevance between prompts, and thus improves the model performance.", + "bbox": [ + 215, + 599, + 787, + 691 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of Interpolation Methods. To compare the performance of global linear interpolation, bilinear interpolation and our proposed geometry-aware spherical interpolation. Our method demonstrates superior performance due to its consideration of geometric accuracy, as shown in Table 5.", + "bbox": [ + 215, + 712, + 787, + 772 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of MCR Loss. To delve into the efficacy of our MCR loss, we setted varying numbers of negative samples to compare their cross-domain performance, as illustrated in Table 6. When the number of negative samples is zero,", + "bbox": [ + 215, + 792, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "LG-Gaze: Language-Guided Gaze Estimation", + "bbox": [ + 429, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/2e470c842f5aa3d4d4bd90fda8ac9a70b9f28d5fb7dfd57c4ed0d16f9edeefe4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
lGeolMCRlGazeD_E → D_MD_E → D_DD_G → D_MD_G → D_DAvg
××√8.359.667.589.018.65
×√√6.767.356.936.987.01
√√√6.457.226.836.866.84
", + "bbox": [ + 236, + 143, + 767, + 213 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/ca58fdd25fefa340af5a2b7cce2d15cdf204be81bdd2070104db80c395ee255d.jpg", + "table_caption": [ + "Table 4: Ablation study on loss functions. Results are reported by in degrees." + ], + "table_footnote": [], + "table_body": "
MethodsDE→DMDE→DdDG→DMDG→DdAvg
Global Li-Inter7.197.547.297.337.34
Bilinear-Inter6.787.316.926.96.98
Ours6.457.226.836.866.84
", + "bbox": [ + 235, + 253, + 772, + 320 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/becc11b0132f690d5549212001e67ea6fdf2730f6e25272bc0b0bf85e68863b2.jpg", + "table_caption": [ + "Table 5: The performance of different interpolation methods." + ], + "table_footnote": [], + "table_body": "
Number\\( \\mathcal{D}_{\\mathrm{E}}\\rightarrow \\mathcal{D}_{\\mathrm{M}} \\)\\( \\mathcal{D}_{\\mathrm{E}}\\rightarrow \\mathcal{D}_{\\mathrm{D}} \\)\\( \\mathcal{D}_{\\mathrm{G}}\\rightarrow \\mathcal{D}_{\\mathrm{M}} \\)\\( \\mathcal{D}_{\\mathrm{G}}\\rightarrow \\mathcal{D}_{\\mathrm{D}} \\)Avg
07.007.447.137.267.21
646.877.317.087.027.07
1286.667.216.987.056.98
2566.457.226.836.866.84
5126.427.196.856.816.82
10246.427.186.836.796.81
", + "bbox": [ + 250, + 359, + 759, + 468 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 6: The Comparison for different number of negative samples.", + "bbox": [ + 272, + 469, + 727, + 482 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "it equates to a common contrastive regression loss, which is observed to achieve less than remarkable performance. As the number of negative samples increases from 64 to 256, there is a marked improvement in performance, indicating that a higher quantity of negative samples in contrastive learning contributes to performance gains. However, when the number of negative samples is further increased to 1024, the performance nearly reaches convergence, with minimal gains.", + "bbox": [ + 212, + 525, + 784, + 616 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Moreover, we conducted an experiment on loss weights to explore the effectiveness of our method. For more details, please see the supplementary material.", + "bbox": [ + 212, + 616, + 784, + 646 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 670, + 359, + 686 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose a novel training framework named LG-Gaze for gaze estimation. Leveraging textual features extracted from language models, LG-Gaze guides the learning of gaze features. This approach enables the gaze model to acquire robust gaze representations and achieve a reasonable feature distribution, ultimately enhancing the cross-domain generalization capability of gaze model. Our proposed LG-Gaze achieves state-of-the-art performance on domain generalization for gaze estimation task.", + "bbox": [ + 212, + 703, + 784, + 808 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This work was sponsored by National Key R&D Program of China (2023YFE0204200).", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Pengwei et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Andrist, S., Tan, X.Z., Gleicher, M., Mutlu, B.: Conversational gaze aversion for humanlike robots. In: 2014 9th ACM/IEEE International Conference on Human-Robot Interaction (HRI). pp. 25-32 (2014) 1", + "2. Bao, Y., Liu, Y., Wang, H., Lu, F.: Generalizing gaze estimation with rotation consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4207-4216 (2022) 11", + "3. Cai, X., Zeng, J., Shan, S., Chen, X.: Source-free adaptive gaze estimation by uncertainty reduction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22035-22045 (2023) 11", + "4. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020) 4", + "5. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020) 10", + "6. Cheng, Y., Bao, Y.: Puregaze: Purifying gaze feature for generalizable gaze estimation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 436-443 (2022) 2, 4, 5, 11", + "7. Cheng, Y., Lu, F., Zhang, X.: Appearance-based gaze estimation via evaluation-guided asymmetric regression. In: Proceedings of the European Conference on Computer Vision (ECCV) (September 2018) 3", + "8. Cheng, Y., Zhang, X., Lu, F., Sato, Y.: Gaze estimation by exploring two-eye asymmetry. IEEE Transactions on Image Processing 29, 5259-5272 (2020) 3", + "9. Funes Mora, K.A., Monay, F., Odobez, J.M.: Eyediap: A database for the development and evaluation of gaze estimation algorithms from rgb and rgb-d cameras. In: Proceedings of the symposium on eye tracking research and applications. pp. 255-258 (2014) 10", + "0. González, Á.: Measurement of areas on a sphere using fibonacci and latitude-longitude lattices. Mathematical Geosciences 42, 49-64 (2010) 10", + "1. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) 5, 12, 13", + "2. Kellnhofer, P., Recasens, A., Stent, S., Matusik, W., Torralba, A.: Gaze360: Physically unconstrained gaze estimation in the wild. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6912-6921 (2019) 3, 10", + "3. Krafka, K., Khosla, A., Kellnhofer, P., Kannan, H., Bhandarkar, S.M., Matusik, W., Torralba, A.: Eye tracking for everyone. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 2176-2184 (2016) 3", + "4. Lee, I., Yun, J.S., Kim, H.H., Na, Y., Yoo, S.B.: Latentgaze: Cross-domain gaze estimation through gaze-aware analytic latent code manipulation. In: Proceedings of the Asian Conference on Computer Vision. pp. 3379-3395 (2022) 11", + "5. Li, W., Huang, X., Zhu, Z., Tang, Y., Li, X., Zhou, J., Lu, J.: Ordinalclip: Learning rank prompts for language-guided ordinal regression. Advances in Neural Information Processing Systems 35, 35313-35325 (2022) 2, 4, 7, 12", + "6. Liu, R., Bao, Y., Xu, M., Wang, H., Liu, Y., Lu, F.: Jitter does matter: Adapting gaze estimation to new domains. arXiv preprint arXiv:2210.02082 (2022) 11", + "7. Liu, Y., Liu, R., Wang, H., Lu, F.: Generalizing gaze estimation with outlier-guided collaborative adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2021) 11" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "LG-Gaze: Language-Guided Gaze Estimation", + "bbox": [ + 429, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) 10", + "19. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) 11", + "20. Mavely, A.G., Judith, J.E., Sahal, P.A., Kuruvilla, S.A.: Eye gaze tracking based driver monitoring system. In: 2017 IEEE International Conference on Circuits and Systems (ICCS). pp. 364-367 (2017). https://doi.org/10.1109/ICCS1.2017.8326022 1", + "21. Padmanaban, N., Konrad, R., Cooper, E.A., Wetzstein, G.: Optimizing vr for all users through adaptive focus displays. In: ACM SIGGRAPH 2017 Talks. SIGGRAPH '17, Association for Computing Machinery, New York, NY, USA (2017). https://doi.org/10.1145/3084363.3085029, https://doi.org/10.1145/3084363.3085029 1", + "22. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 2, 4, 12", + "23. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog p. 9 (2019) 10", + "24. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Denseclip: Language-guided dense prediction with context-aware prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18082-18091 (2022) 4", + "25. Shen, S., Li, W., Wang, X., Zhang, D., Jin, Z., Zhou, J., Lu, J.: Clip-cluster: Clip-guided attribute hallucination for face clustering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 20786-20795 (2023) 4", + "26. Shoemake, K.: Animating rotation with quaternion curves. In: Proceedings of the 12th annual conference on Computer graphics and interactive techniques. pp. 245-254 (1985) 3, 7", + "27. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems (2017) 10", + "28. Vidit, V., Engilberge, M., Salzmann, M.: Clip the gap: A single domain generalization approach for object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3219-3229 (2023) 4", + "29. Wang, R., Li, P., Huang, H., Cao, C., He, R., He, Z.: Learning-to-rank meets language: Boosting language-driven ordering alignment for ordinal classification. Advances in Neural Information Processing Systems 36 (2024) 4", + "30. Wang, Y., Jiang, Y., Li, J., Ni, B., Dai, W., Li, C., Xiong, H., Li, T.: Contrastive regression for domain adaptation on gaze estimation. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19354-19363 (2022). https://doi.org/10.1109/CVPR52688.2022.01877 2, 4, 5, 11, 12, 13", + "31. Xu, M., Wang, H., Lu, F.: Learning a generalized gaze estimator from gaze-consistent feature. Proceedings of the AAAI Conference on Artificial Intelligence 37(3), 3027-3035 (Jun 2023) 2, 4, 5, 11", + "32. Yao, L., Han, J., Wen, Y., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, C., Xu, H.: Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 9125-9138. Curran Associates, Inc. (2022), https://proceedings.neurips.cc/paper_files/" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Pengwei et al.", + "bbox": [ + 271, + 114, + 387, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "paper/2022/file/3ba960559212691be13fa81d9e5e0047-Paper-Conference.pdf 4", + "33. Yin, P., Wang, J., Dai, J., Wu, X.: Nerf-gaze: A head-eye redirection parametric model for gaze estimation. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2024) 4", + "34. Yin, P., Zeng, G., Wang, J., Xie, D.: Clip-gaze: Towards general gaze estimation via visual-linguistic model. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 6729–6737 (2024) 2, 4, 11", + "35. Zha, K., Cao, P., Son, J., Yang, Y., Katabi, D.: Rank-n-contrast: Learning continuous representations for regression. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) 4, 12, 13", + "36. Zha, K., Cao, P., Son, J., Yang, Y., Katabi, D.: Rank-n-contrast: Learning continuous representations for regression. Advances in Neural Information Processing Systems 36 (2024) 12", + "37. Zhang, X., Park, S., Beeler, T., Bradley, D., Tang, S., Hilliges, O.: Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze variation. In: European Conference on Computer Vision. pp. 365-381. Springer (2020) 3, 10", + "38. Zhang, X., Sugano, Y., Fritz, M., Bulling, A.: Appearance-based gaze estimation in the wild. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 4511-4520 (2015) 3", + "39. Zhang, X., Sugano, Y., Fritz, M., Bulling, A.: Mpiigaze: Real-world dataset and deep appearance-based gaze estimation. IEEE transactions on pattern analysis and machine intelligence 41(1), 162-175 (2017) 10", + "40. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022) 4", + "41. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) 4, 12" + ], + "bbox": [ + 215, + 147, + 784, + 534 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "LG-Gaze: Language-Guided Gaze Estimation", + "bbox": [ + 429, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/LG-Gaze_ Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation/82982acf-4d4b-4f4d-a9b4-ee45c99a1bec_model.json b/2024/LG-Gaze_ Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation/82982acf-4d4b-4f4d-a9b4-ee45c99a1bec_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7abf8caf8e076b8f36c8d601146b447b21902c1b --- /dev/null +++ b/2024/LG-Gaze_ Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation/82982acf-4d4b-4f4d-a9b4-ee45c99a1bec_model.json @@ -0,0 +1,2280 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.227, + 0.142, + 0.78, + 0.187 + ], + "angle": 0, + "content": "LG-Gaze: Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.213, + 0.775, + 0.229 + ], + "angle": 0, + "content": "Pengwei Yin*, Jingjing Wang*, Guanzhong Zeng, Di Xie, and Jiang Zhu" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.24, + 0.688, + 0.283 + ], + "angle": 0, + "content": "Hikvision Research Institute, Hangzhou, China {yinpengwei, wangjingjing9, zengguanzhong, xiedi, zhujiang.hri}@hikvision.com" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.328, + 0.744, + 0.593 + ], + "angle": 0, + "content": "Abstract. The ability of gaze estimation models to generalize is often significantly hindered by various factors unrelated to gaze, especially when the training dataset is limited. Current strategies aim to address this challenge through different domain generalization techniques, yet they have had limited success due to the risk of overfitting when solely relying on value labels for regression. Recent progress in pre-trained vision-language models has motivated us to capitalize on the abundant semantic information available. We propose a novel approach in this paper, reframing the gaze estimation task as a vision-language alignment issue. Our proposed framework, named Language-Guided Gaze Estimation (LG-Gaze), learns continuous and geometry-sensitive features for gaze estimation benefit from the rich prior knowledge of vision-language models. Specifically, LG-Gaze aligns gaze features with continuous linguistic features through our proposed multimodal contrastive regression loss, which customizes adaptive weights for different negative samples. Furthermore, to better adapt to the labels for gaze estimation task, we propose a geometry-aware interpolation method to obtain more precise gaze embeddings. Through extensive experiments, we validate the efficacy of our framework in four different cross-domain evaluation tasks." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.604, + 0.741, + 0.633 + ], + "angle": 0, + "content": "Keywords: Gaze Estimation \\(\\cdot\\) Vision-Language Model \\(\\cdot\\) Continuous Regression Task \\(\\cdot\\) Contrastive Learning" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.655, + 0.377, + 0.67 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.684, + 0.788, + 0.806 + ], + "angle": 0, + "content": "The gaze estimation is crucial for understanding human behavior. Precise gaze estimation offers significant assistance for many applications, such as human-computer interaction [1], augmented reality [21], and driver monitoring systems [20]. The deep learning has led to significant improvements in gaze estimation based on appearance. While these methods show impressive results in evaluations within the same domain, they tend to experience a decline in performance when evaluated on different domains. This decline is primarily attributed to overfitting rather than learning robust features from the original domain." + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.812, + 0.458, + 0.827 + ], + "angle": 0, + "content": "*These authors contributed equally." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "Y. Pengwei et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.149, + 0.782, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.214, + 0.45, + 0.226 + ], + "angle": 0, + "content": "(a) Traditional Methods: Supervised by Value Label" + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.214, + 0.695, + 0.226 + ], + "angle": 0, + "content": "(b) LG-Gaze: Guided by Language Model" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.241, + 0.788, + 0.283 + ], + "angle": 0, + "content": "Fig. 1: (a) The traditional method of gaze generalization involves overseeing model training by means of numerical label regression. (b) In our study, we introduce a text models to steer the development of robust features in visual models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.318, + 0.784, + 0.409 + ], + "angle": 0, + "content": "Gaze data encompass diverse elements, such as appearance, wearables, environments, and image clarity [31, 34], yet gaze annotations predominantly focus on eye direction, marginalizing other variables as noise. Such irrelevant factors exacerbate the domain disparity, hindering gaze models' adaptability. Additionally, supervisory numeric labels might induce overfitting, influenced by these extraneous gaze components." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.785, + 0.547 + ], + "angle": 0, + "content": "As shown in Figure 1(a), training a gaze model typically involves utilizing a vision encoder to extract features and then employing a regressor for prediction. Traditional methods train a gaze model by using data disturbing [31], feature enhancement [30], and adversarial learning [6] to improve the generalization ability of the gaze model. However, these regularization techniques have limited effectiveness as they rely solely on numerical labels for supervision, making them susceptible to overfitting due to irrelevant factors. The above may indicate that it is difficult to guide the model to learn ideal robust features through these indirect regularization terms." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.717 + ], + "angle": 0, + "content": "To counter overfitting in gaze representation learning, we advocate for multimodal integration. Language, rich in semantics and knowledge, complements visual data. Each gaze label is described textually, \"the yaw/pitch degree angle of the person is {yaw/pitch}.\". Leveraging visual-language models, such as CLIP, which excel in generalized latent space learning from vast image-text pairs, we align image features with a language space. This distillation of CLIP's language expertise acts as regularization, enhancing generalization. However, these methods are only applicable to a limited set of discrete scalars and are not suitable for continuous tasks such as gaze estimation. Aligning gaze features to discrete text features can result in inaccurate feature alignment, which in turn affects the performance of continuous regression tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To address these issues, we propose a novel training method named LG-Gaze, which is designed to learn continuous and geometry-sensitive features for gaze estimation, as depicted in Figure 1(b). This framework aligns gaze features with continuous linguistic features extracted by a powerful language model, which not only prevents model overfitting but also enhances the generalization capabilities of gaze model. Primarily, unlike common discrete multimodal contrastive learning methods [15, 22], we introduce a loss function for continuous multimodal contrastive learning. Our Multimodal Contrastive Regression loss" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.43, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LG-Gaze: Language-Guided Gaze Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.327 + ], + "angle": 0, + "content": "function (MCR) customizes adaptive weights for different negative samples according to label distance, facilitating more refined feature alignment in the feature space and benefiting the model in learning more continuous features. Next, to better adapt to the vectorial property of labels for gaze estimation task, we propose a geometry-aware interpolation method to obtain more precise gaze embeddings. The geometry-aware interpolation method combines spherical [26] and bilinear interpolation techniques to resolve the issue of creating semantic labels for gaze, ensuring accurate gaze prompts. Furthermore, MCR also resolves the limitations of previous contrastive learning functions by utilizing a more uniform distribution of global negative samples for contrast. Unlike most methods that are limited to intra-batch comparisons, our loss function benefits from extensive global contrast, resulting in a reasonable and robust feature distribution." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.329, + 0.633, + 0.342 + ], + "angle": 0, + "content": "The main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.353, + 0.784, + 0.44 + ], + "angle": 0, + "content": "- In this paper, we propose a novel training framework called LG-Gaze for gaze estimation. LG-Gaze guides the gaze model to learn generalized representations by leveraging textual features extracted from language models, resulting in more robust gaze representations. By incorporating rich language knowledge, this approach achieves robust gaze features, ultimately enhancing the cross-domain generalization capability of the gaze model." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.444, + 0.784, + 0.516 + ], + "angle": 0, + "content": "- To learn continuous feature for regression tasks, we propose a new multimodal contrastive learning loss function named MCR. The method employs adaptive weights for different negative sample, helping learn more refined features. Moreover, MCR leverages ample global negative samples for contrast, leading to a reasonable feature distribution." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.519, + 0.784, + 0.561 + ], + "angle": 0, + "content": "- We introduce a geometry-aware interpolation method that employs spherical interpolation techniques to compute accurate gaze embeddings, thereby ensuring the accuracy of semantic labels." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.564, + 0.784, + 0.606 + ], + "angle": 0, + "content": "- Experimental results and visualizations demonstrate that LG-Gaze not only achieves remarkable performance compared to the baseline but also surpasses state-of-the-art domain generalization methods for gaze estimation." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.353, + 0.784, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.633, + 0.396, + 0.649 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.665, + 0.4, + 0.679 + ], + "angle": 0, + "content": "2.1 Gaze Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Appearance-based gaze estimation has garnered attention in recent years [7,8, 13,38]. However, it faces challenges in cross-domain evaluation due to the domain gap resulting from various gaze-irrelevant factors. Common approaches often rely on diverse gaze datasets [12,37] for training, aiming to equip models with robust generalization capabilities. Unfortunately, collecting gaze data remains costly, and the diversity of available data remains limited. To address this issue, we propose enhancing gaze estimation models through domain generalization (DG) methods. These methods enable models to generalize to unseen distributions and enhance cross-domain performance. Most existing DG techniques are tailored for classification tasks, leaving a gap in the context of gaze estimation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "Y. Pengwei et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "PureGaze [6] debuts a self-adversarial schema, discarding irrelevant gaze cues, enhancing pertinent ones. Xu et al. [31] and NeRF-Gaze [33] similarly counteract gaze distractions via adversarial data modifications and augmentation. However, residual gaze-confounding factors persist, hindering estimation precision. CLIP-Gaze [34] is designed to segregate gaze features from predefined textual gaze-irrelevant features, thereby enhancing the generalization capability and robustness of gaze features. Our LG-Gaze, by directly aligning with language features, also exhibits resilience against gaze-irrelevant factors. More critically, the text features in LG-Gaze possess favorable rank properties, attributed to the geometry-aware continuous prompts learning." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.319, + 0.458, + 0.334 + ], + "angle": 0, + "content": "2.2 Vision Language Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.789, + 0.614 + ], + "angle": 0, + "content": "In recent research, several studies have utilized the text manipulation and visual alignment capabilities of CLIP (Contrastive Language-Image Pretraining) [22] to improve opendetection and generalization performance in specific tasks. Notable examples include DetCLIP [32], CLIP-Gap [28], and CLIP-Cluster [25]. To further enhance the performance of vision-language models on downstream tasks, an effective strategy involves learning more contextually appropriate prompts through text prompt tuning [40, 41]. Additionally, DenseCLIP [24] explores the application of semantic knowledge into monocular depth estimation tasks. By matching visual features with textual features, this method achieves zero-shot monocular depth estimation. \"Teaching CLIP to Count to Ten\" and CrowdCLIP propose visual language models for understanding quantities, enabling object counting and crowd estimation. While these methods enhance models' understanding of quantities, they do not address regression tasks. OrdinalCLIP [15] and L2RCLIP [29] are capable of addressing ordinal regression tasks. However, they are not as effective for continuous regression tasks. In this paper, our method aligns gaze features with continuous linguistic features through our proposed multimodal contrastive regression loss (MCR) and the geometry-aware interpolation method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.635, + 0.439, + 0.65 + ], + "angle": 0, + "content": "2.3 Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In computer vision, SimCLR [4] and CLIP [22] pioneer contrastive loss, exceeding in image classification and retrieval. Yet, these techniques aren't suited for regression, where prediction targets are continuous, not categorical. Rank-N-Contrast (RNC) [35] employs sample rankings to enhance continuity understanding. Nonetheless, RNC presumes infinite distance between negative and positive pairs, neglecting the correlation between label and feature proximities. Contrastive Domain Generalization (CDG) [30] leverages contrastive loss to encourage the clustering of features corresponding to similar gaze directions while separating those associated with significantly different gaze directions. However, CDG relies on empirical thresholding to define positive and negative sample pairs, which limits its applicability for continuous regression tasks. In our research, we introduce a novel contrastive loss function specifically designed for" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.43, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LG-Gaze: Language-Guided Gaze Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "continuous regression tasks. Our approach prioritizes sample ordering while considering label distances. Unlike previous contrastive learning methods that are confined to constructing sample pairs within batches, our method demonstrates a more targeted enhancement. To elaborate further, we draw inspiration from the Momentum Contrast (MoCo) framework [11], which utilizes a queue to store negative sample features. However, these negative samples are obtained from n steps back in time and exhibit inherent uncontrollable randomness. Unlike delayed updates, our approach synchronously updates these samples, ensuring a controlled distribution instead of relying on randomness." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.305, + 0.331, + 0.321 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.335, + 0.424, + 0.349 + ], + "angle": 0, + "content": "3.1 Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.494 + ], + "angle": 0, + "content": "In the context of gaze estimation tasks, which fundamentally belong to regression tasks, we define the data for the source domain as \\(\\mathcal{D}_s = (\\pmb {I}_i,\\pmb {g}_i)_{i = 1}^M\\), where \\((\\pmb {I}_i,\\pmb {g}_i)\\) represents the \\(i\\)-th pair of image \\(\\pmb {I}_i\\in \\mathbb{R}^{224\\times 224\\times 3}\\) and their corresponding gaze direction \\(\\pmb {g}_i\\in \\mathbb{R}^3\\). Here, \\(M\\) denotes the total number of pairs in \\(\\mathcal{D}_s\\). Our objective in gaze estimation is to train a neural network composed of an image encoder \\(\\mathcal{E}_v(\\cdot):\\mathbb{R}^{224\\times 224\\times 3}\\to \\mathbb{R}^d\\) to extract gaze representations \\(\\pmb{f}_i^g\\) based on input image \\(\\mathcal{I}_i\\) and an MLP regressor \\(\\mathcal{R}_g(\\cdot):\\mathbb{R}^d\\to \\mathbb{R}^3\\) to predict gaze directions \\(\\pmb {g}_i\\). Typically, once we have trained a robust gaze representation, training \\(\\mathcal{R}_g\\) becomes relatively straightforward using regression functions (e.g., \\(L_{1}\\) or AngleLoss)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.495, + 0.788, + 0.554 + ], + "angle": 0, + "content": "However, gaze estimation research has encountered a persistent challenge that hinders its advancement: How can we design robust gaze representations \\(\\{f_i^g\\}_{i=1}^M\\) that are continuous, semantic, and sensitive to geometric factors for gaze estimation?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.788, + 0.706 + ], + "angle": 0, + "content": "Although many previous methods explore additional technologies to enhance feature representations, such as adversarial training [6], contrastive learning as an auxiliary task [30] and mitigating the impact of gaze-irrelevant factors [31], they often focus solely on finding regularization terms to improve model performance without explicitly aiming for robust and well-distributed features. Consequently, existing methods often face overfitting issues and struggle to achieve the desired performance. In this paper, we propose a novel language-guided gaze representation learning framework that leverages rich prior knowledge from a natural language model. Our approach aims to learn robust and well-distributed gaze representations, addressing the aforementioned challenges." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.727, + 0.434, + 0.74 + ], + "angle": 0, + "content": "3.2 LG-Gaze Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To harness the full power of language, we utilize the text encoder from CLIP as our language model. CLIP adopts a vision-language pre-training framework and learns representations from image-text pairs, thereby constructing a joint vision-language latent space. Given the strong performance of the CLIP model in downstream tasks, we implemented our method using the original CLIP text encoder from the referenced paper." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "Y. Pengwei et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.147, + 0.782, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.259, + 0.788, + 0.289 + ], + "angle": 0, + "content": "Fig. 2: We reformulate the task as an image-language matching problem, which mainly consists of a trainable prompt, a frozen text encoder, a trainable image encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.316, + 0.789, + 0.484 + ], + "angle": 0, + "content": "Suppose that each training step is provided with a data batch \\((I_i, g_i)_{i=1}^B\\), where \\(g_i\\) represents the gaze label of the image \\(I_i\\). Our goal is to construct semantic labels and use a language model to extract text features. This process will help the gaze model learn robust features by aligning them with the extracted features. To leverage the prior knowledge within text features, we begin by constructing a learnable prompt \\(\\{\\pmb{v}_1, \\pmb{v}_2, \\dots, \\pmb{v}_{L-1}\\} \\in \\mathbb{R}^{512 \\times (L-1)}\\) as the context for describing gaze estimation tasks instead of handcrafting the prompt context. Simultaneously, we treat gaze labels as individual words. For each sample, we concatenate the learnable prompt with the embedding \\(\\pmb{v}_i^g \\in \\mathbb{R}^{512}\\) to form a sequence \\(\\pmb{T}_i = \\{\\pmb{v}_1, \\pmb{v}_2, \\dots, \\pmb{v}_{L-1}, \\pmb{v}_i^g\\}\\) as a semantic label. Specifically, we utilize the gaze embedding as the final token in the sequence \\(\\pmb{T}_i \\in \\mathbb{R}^{512 \\times L}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.789, + 0.589 + ], + "angle": 0, + "content": "To ensure the continuity, ordinality, and geometric properties of gaze label embeddings, we introduce \\(N\\) learnable embedding anchors, denoted as \\(\\{\\pmb{A}_j\\}_{j=1}^N\\). Each anchor corresponds to a unique gaze vector, uniformly distributed across the entire label space. We utilize a nearest-neighbor interpolation method to represent gaze label embeddings associated with sample relationships. This approach yields a continuous and geometry-aware representation of gaze embeddings, as detailed in Section 3.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.789, + 0.726 + ], + "angle": 0, + "content": "In our research, we send image \\( \\pmb{I}_i \\) to an image encoder \\( \\mathcal{E}_v \\) to extract the image feature \\( \\pmb{f}_i^g \\). Consequently, the features of a batch can be represented as \\( \\{\\pmb{f}_i^g\\}_{i=1}^B \\). Simultaneously, we utilize the language model \\( \\mathcal{E}_t \\) to obtain corresponding text features \\( \\{\\pmb{f}_i^t\\}_{i=1}^B \\) from \\( \\{\\pmb{T}_i\\}_{i=1}^B \\). While the parameters of the language model remain frozen, we train the entire image encoder to align image features with the language feature space. Subsequently, we utilize an image-text contrastive loss to optimize the network, which includes both an image-to-text loss and a text-to-image loss, following the methodology of CLIP. Detailed information about our proposed contrastive learning method will be provided in Section 3.4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.726, + 0.789, + 0.786 + ], + "angle": 0, + "content": "After aligning the features, we obtain robust gaze representations. Finally, the image features are fed into a regressor denoted as \\(\\mathcal{R}_g\\), yielding the predicted gaze direction \\(\\hat{g}_i \\in \\mathbb{R}^3\\). Specifically, we predict through supervised learning, which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.797, + 0.787, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {G a z e} \\left(\\hat {\\boldsymbol {g}} _ {i}, \\boldsymbol {g} _ {i}\\right) = \\operatorname {a r c c o s} \\left(\\frac {\\hat {\\boldsymbol {g}} _ {i} \\cdot \\boldsymbol {g} _ {i}}{\\| \\hat {\\boldsymbol {g}} _ {i} \\| \\| \\boldsymbol {g} _ {i} \\|}\\right) \\tag {1}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.429, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "LG-Gaze: Language-Guided Gaze Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.145, + 0.387, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.3, + 0.248, + 0.318, + 0.258 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.149, + 0.508, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.436, + 0.248, + 0.454, + 0.259 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.147, + 0.631, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.248, + 0.582, + 0.258 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.147, + 0.756, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.689, + 0.246, + 0.705, + 0.256 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.269, + 0.782, + 0.284 + ], + "angle": 0, + "content": "Fig. 3: These four subfigures contain all steps of our proposed interpolation method." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.311, + 0.623, + 0.327 + ], + "angle": 0, + "content": "3.3 Geometry-aware Continuous Gaze Prompts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.787, + 0.425 + ], + "angle": 0, + "content": "In this section, to better represent gaze labels in the embedding space, we construct a gaze embedding for each label. This embedding should have the capability for continuous representation and geometric attributes. To address the conflict between the finite vocabulary in natural language processing and the infinite nature of gaze labels, we interpolate each gaze embedding \\(\\pmb{v}_i^g\\) through \\(N\\) gaze anchor embeddings \\(\\{\\pmb{A_j}\\}_{j = 1}^N\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.425, + 0.788, + 0.517 + ], + "angle": 0, + "content": "Common approaches, such as those described in [15], employ global linear interpolation to represent \\( \\pmb{v}_i^g \\), which interpolates the target embedding using all anchor embeddings. The weight \\( w_{i,j} \\) for each anchor embedding is computed by Equation 2, determined by the similarity between its corresponding gaze vector \\( \\pmb{g}_j \\) and the target \\( \\pmb{g}_i \\). However, global interpolation cannot accurately represent \\( \\pmb{v}_i^g \\) because anchors that are not proximate fail to provide effective information." + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.525, + 0.786, + 0.565 + ], + "angle": 0, + "content": "\\[\nw _ {i, j} = \\frac {\\cos \\left(\\boldsymbol {g} _ {i} , \\boldsymbol {g} _ {j}\\right)}{\\sum_ {j = 1} ^ {N} \\cos \\left(\\boldsymbol {g} _ {i} , \\boldsymbol {g} _ {j}\\right)} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.787, + 0.627 + ], + "angle": 0, + "content": "To further enhance interpolation, we attempted to employ a bilinear interpolation method. However, as it still utilizes a linear representation, this method remains insufficiently precise in expressing the geometric relationships between gaze vectors." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.629, + 0.788, + 0.765 + ], + "angle": 0, + "content": "To address this issue, we decided to adopt a spherical interpolation method [26], which can accurately calculate the interpolation weights between vectors. As illustrated in Figure 3(a), a flat spherical interpolation method is depicted. For the vector \\( \\pmb{g}_i \\) positioned between \\( \\pmb{g}_1 \\) and \\( \\pmb{g}_2 \\), we initially compute \\( \\theta = \\arccos(\\langle \\pmb{g}_1, \\pmb{g}_2 \\rangle) \\) and \\( \\theta_1 = \\arccos(\\langle \\pmb{g}_1, \\pmb{g}_i \\rangle) \\). Subsequently, we compute the scalar \\( t \\) by \\( t = \\theta_1 / \\theta \\), then we apply the spherical interpolation Equation 3 to determine the corresponding weights for \\( \\pmb{g}_1 \\) and \\( \\pmb{g}_2 \\). Finally, we can represent the target vector \\( \\pmb{g}_i \\) through the summation of these weights. Ultimately, we confirm that this method of weight calculation is accurate." + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.774, + 0.787, + 0.808 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {g} _ {i} = \\frac {\\sin ((1 - t) \\theta)}{\\sin (\\theta)} \\boldsymbol {g} _ {1} + \\frac {\\sin (t \\theta)}{\\sin (\\theta)} \\boldsymbol {g} _ {2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Building upon the aforementioned spherical interpolation method, we further integrate the concept of bilinear interpolation to derive the interpolation" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "Y. Pengwei et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "weights for each gaze label on a stereographic sphere. As depicted in Figure 3(b), we initially identify the four nearest anchor embeddings \\( A_{1-4} \\) based on their proximity. Subsequently, as shown in Figure 3(c), we calculate the weights \\( w_{a,1} \\) and \\( w_{a,2} \\) for \\( A_1 \\) and \\( A_2 \\) using Equation 3. Similarly, we can determine \\( w_{b,3} \\) and \\( w_{b,4} \\). In Figure 3(d), we proceed to obtain \\( w_{i,a} \\) and \\( w_{i,b} \\) using a similar method. Ultimately, we acquire the associated weights for \\( A_{1-4} \\) to interpolate \\( \\pmb{v}_i^g \\), which can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.266, + 0.786, + 0.283 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {i} ^ {g} = \\left(w _ {i, a} * w _ {a, 1}\\right) \\cdot \\boldsymbol {A} _ {1} + \\left(w _ {i, a} * w _ {a, 2}\\right) \\cdot \\boldsymbol {A} _ {2} + \\left(w _ {i, b} * w _ {b, 3}\\right) \\cdot \\boldsymbol {A} _ {3} + \\left(w _ {i, b} * w _ {b, 4}\\right) \\cdot \\boldsymbol {A} _ {4} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.788, + 0.356 + ], + "angle": 0, + "content": "Additionally, to enhance the geometric relationships between gaze embeddings, we employ a simple loss function to constrain anchor embeddings \\(\\{A_j\\}_{j=1}^N\\). As shown in Equation 5, the similarity between anchor embeddings should be consistent with their relationships in the label space." + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.368, + 0.786, + 0.41 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {G e o}} = \\frac {1}{M * M} \\sum_ {i = 1} ^ {M} \\sum_ {j = 1} ^ {M} L _ {1} \\left(\\cos \\left(\\boldsymbol {A} _ {i}, \\boldsymbol {A} _ {j}\\right), \\cos \\left(\\boldsymbol {g} _ {i}, \\boldsymbol {g} _ {j}\\right)\\right) \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.432, + 0.774, + 0.447 + ], + "angle": 0, + "content": "3.4 Continuous Multimodal Contrastive Regression Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.789, + 0.639 + ], + "angle": 0, + "content": "Mathematically, from the aforementioned steps, we obtain image features \\(\\{\\pmb{f}_i^g\\}_{i=1}^B\\) and their corresponding text features \\(\\{\\pmb{f}_i^t\\}_{i=1}^B\\). Both CLIP and OrdinalCLIP utilize the InfoNCE loss to align the multimodal features. To delve into specifics, consider an image-text feature pair \\((\\pmb{f}_i^g, \\pmb{f}_i^t)\\). Here, \\(\\pmb{f}_i^g / \\pmb{f}_i^t\\) serves as the positive sample for \\(\\pmb{f}_i^t / \\pmb{f}_i^g\\), as they are associated with the same gaze label. Meanwhile, all other image/text features in a mini-batch are treated as negative samples and are consequently pushed away from \\(\\pmb{f}_i^t / \\pmb{f}_i^g\\). However, this training objective poses challenges for regression tasks because it overlooks the finer-grained distance ordering relationships between samples. Consequently, failing to differentiate the relative distances between image/text and their corresponding negative samples will inevitably weaken the learning effectiveness, especially in cross-modal representation learning for regression tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.788, + 0.775 + ], + "angle": 0, + "content": "Considering the proximity relationships between data, for the \\(i\\)-th sample, we introduce a weight parameter \\(\\{w(i,j)\\}_{j=1}^{B}\\) for each negative sample. This parameter is computed based on the distances between labels of samples. Our proposed multimodal contrastive regression loss (MCR) effectively controls the feature distances among samples, preventing negative samples from being excessively contracted. As a result, we learn a semantically meaningful representation space tailored for regression tasks, ensuring consistency between sample feature relationships and the label space. The proposed text-to-image MCR loss is formulated as follows." + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.786, + 0.786, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\ell_ {t 2 i} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\left(\\frac {\\exp \\left(\\sin \\left(\\boldsymbol {f} _ {i} ^ {t} , \\boldsymbol {f} _ {i} ^ {g}\\right)\\right)}{\\exp \\left(\\sin \\left(\\boldsymbol {f} _ {i} ^ {t} , \\boldsymbol {f} _ {i} ^ {g}\\right)\\right) + \\sum_ {j = 1} ^ {B} w (i , j) \\cdot \\exp \\left(\\sin \\left(\\boldsymbol {f} _ {i} ^ {t} , \\boldsymbol {f} _ {j} ^ {g}\\right)\\right)}\\right) \\tag {6}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.429, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "LG-Gaze: Language-Guided Gaze Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.252 + ], + "angle": 0, + "content": "where \\( \\text{sim}(\\pmb{f}_i^t, \\pmb{f}_i^g) = \\cos(\\pmb{f}_i^t, \\pmb{f}_i^g) \\) and \\( w(i,j) = \\cos(\\pmb{g}_i, \\pmb{g}_j) \\) denotes the contrastive weight of the \\( j \\)-th negative image sample with respect to the \\( i \\)-th text feature in our LG-Gaze framework. The weight parameter should be directly proportional to the label distance between negative samples. As the distance between labels increases, the penalty on negative sample weights should also increase. This encourages feature embeddings to adhere to real-world relationships." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.493 + ], + "angle": 0, + "content": "The MCR loss for image-to-text can be formulated in a similar manner. However, what sets it apart is our deliberate expansion of the number of negative samples to facilitate the learning of a robust representation. To achieve this, we draw inspiration from MoCo, which utilizes a dynamic queue to maintain a diverse set of negative samples, thereby enhancing the effectiveness of contrastive learning. Nevertheless, this approach still faces some challenges. For instance, if the queue update rate is too slow, the negative samples stored in the queue may become outdated and fail to align with the current model. Additionally, due to the inherent randomness and limited diversity of negative samples, there could be discrepancies between the queue data and the entire data distribution. To address this issue, we utilize the flexibility of language labels by creating \\(K\\) global text features \\(\\{f_i^t\\}_{i=1}^K\\) as negative samples (where \\(K\\) is significantly larger than \\(B\\)). These text labels correspond to gaze vectors that show a uniformly dense distribution in the label space. Furthermore, these \\(K\\) text labels are interpolated based on \\(N\\) anchor \\(\\{A_j\\}_{j=1}^N\\). As a result, the proposed image-to-text MCR loss can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.505, + 0.785, + 0.56 + ], + "angle": 0, + "content": "\\[\n\\ell_ {i 2 t} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\left(\\frac {\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {f} _ {i} ^ {g} , \\boldsymbol {f} _ {i} ^ {t}\\right)\\right)}{\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {f} _ {i} ^ {g} , \\boldsymbol {f} _ {i} ^ {t}\\right)\\right) + \\sum_ {j = 1} ^ {B + K} w (i , j) \\cdot \\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {f} _ {i} ^ {g} , \\boldsymbol {f} _ {j} ^ {t}\\right)\\right)}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.785, + 0.695 + ], + "angle": 0, + "content": "Obviously, we can observe that the difference between \\(\\ell_{i2t}\\) and \\(\\ell_{t2i}\\) lies in the number of negative samples. We concatenate \\(K\\) global negative samples with the \\(B\\) negative samples from mini-batches to form the complete set of negative samples. Through this operation, the global negative samples remain relatively fixed, mitigating issues related to randomness. Additionally, their uniform distribution helps alleviate discrepancies between the negative samples and the entire data distribution. Furthermore, by computing features based on the current model, we address potential outdated information. Finally, our total MCR loss can be denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.437, + 0.699, + 0.785, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\ell_ {M C R} = \\ell_ {i 2 t} + \\ell_ {t 2 i} \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.738, + 0.408, + 0.754 + ], + "angle": 0, + "content": "3.5 Overall Objective" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.767, + 0.704, + 0.782 + ], + "angle": 0, + "content": "In summary, the overall objective implemented in our framework is:" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.797, + 0.619, + 0.811 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {1} \\ell_ {G e o} + \\lambda_ {2} \\ell_ {M C R} + \\lambda_ {3} \\ell_ {G a z e}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.825, + 0.777, + 0.841 + ], + "angle": 0, + "content": "\\(\\lambda_1, \\lambda_2, \\lambda_3\\) are hyperparameters, and we empirically set \\(\\lambda_1 = \\lambda_2 = \\lambda_3 = 1.0\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "Y. Pengwei et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.377, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.178, + 0.424, + 0.194 + ], + "angle": 0, + "content": "4.1 Experiment Details" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.205, + 0.787, + 0.28 + ], + "angle": 0, + "content": "Data Preparation We evaluate gaze estimation method on four cross-domain tasks, utilizing ETH-XGaze [37] and Gaze360 [12] as the training datasets, and MPIIFaceGaze [39] and Eye-Diap [9] as the test datasets. We denote them as \\(\\mathcal{D}_{\\mathrm{E}}\\) (ETH-XGaze) \\(\\rightarrow \\mathcal{D}_{\\mathrm{M}}\\) (MPIIFaceGaze), \\(\\mathcal{D}_{\\mathrm{E}} \\rightarrow \\mathcal{D}_{\\mathrm{D}}\\) (EyeDiap), \\(\\mathcal{D}_{\\mathrm{G}}\\) (Gaze360) \\(\\rightarrow \\mathcal{D}_{\\mathrm{M}}\\), \\(\\mathcal{D}_{\\mathrm{G}} \\rightarrow \\mathcal{D}_{\\mathrm{D}}\\). See the supplementary material for more details." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.304, + 0.787, + 0.381 + ], + "angle": 0, + "content": "Vision Model Implementation Details We use ResNet-18 as our gaze feature extractor \\(\\mathcal{E}_v\\), and a fully connected layer to regress a 3-dimensional gaze vector. We resize and normalize all the images to \\(224\\times 224\\) pixels and scale the pixel values between 0 and 1. We set the batch size to 64. We maintain the same setup for training the source domain models on \\(\\mathcal{D}_E\\) and \\(\\mathcal{D}_G\\)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.403, + 0.787, + 0.556 + ], + "angle": 0, + "content": "Language Model Implementation Details The architecture of the text encoder is adapted from the Transformer [27], incorporating modifications outlined by Radford [23]. The dimensions for both text and gaze features have been standardized to 1024. Our models are constructed using the foundational open-source code of CLIP. The class token is strategically placed at the terminal position of the sequence, with the context tokens quantity designated as \\( L = 10 \\). Additionally, we set the yaw angle range for the anchors to be \\( [-180^{\\circ}, +179^{\\circ}] \\) and the pitch range to be \\( [-90^{\\circ}, +90^{\\circ}] \\) with a step size of \\( 30^{\\circ} \\), resulting in a total of 91 anchors. We set the number of negative samples to 256, which are uniformly distributed in the gaze label space using the Fibonacci sphere algorithm [10]." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.578, + 0.787, + 0.67 + ], + "angle": 0, + "content": "Training Implementation Details We conducted the experiments on a single Tesla V100 GPU. Specifically, we use the SGD optimizer with Nesterov momentum, a learning rate (LR) of \\(5 \\times 10^{-2}\\) and a weight decay of \\(1 \\times 10^{-5}\\) for the parameters. We train for 30 epochs using a Cosineannealing LR scheduler [18] with a 3-epoch warm-up. We apply a data augmentation technique involving a random color field and grayscale, as described in [5]." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.693, + 0.787, + 0.709 + ], + "angle": 0, + "content": "4.2 Performance Comparison with SOTA Gaze Estimation Methods" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.719, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Comparison of Cross-domain Evaluation. Table 1 presents the quantitative results of four cross-domain tests. The second row compares our LG-Gaze with the SOTA domain DG methods for gaze estimation. The CNN baseline method refers to the one that relies solely on the vision encoder to extract the single visual modality feature and uses \\(\\ell_{gaze}\\) for supervision. In summary, the LG-Gaze achieves the best overall performance and demonstrates state-of-the-art results on three cross-domain evaluation tasks. It also achieves the second-best performance for \\(\\mathcal{D}_{\\mathrm{E}} \\rightarrow \\mathcal{D}_{\\mathrm{M}}\\), highlighting the effectiveness of our method." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.43, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LG-Gaze: Language-Guided Gaze Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.144, + 0.781, + 0.365 + ], + "angle": 0, + "content": "
TaskMethods| Dt|DE→DMDE→DdDG→DMDG→DdAvg
DGCNN Baseline08.479.327.548.938.57
PureGaze [6]07.087.489.289.328.29
CDG [30]‡06.737.957.037.277.25
Xu et al. [31]06.507.447.559.037.63
CLIP-Gaze [34]06.417.516.897.066.97
Our LG-Gaze06.457.226.836.866.84
UDAPnP-GA * [17]105.535.876.187.926.38
RUDA [2]1005.706.296.205.866.01
CRGA [30]>05.485.665.896.495.88
LatentGaze [14]1005.217.81--6.51
Liu et al. [16]1005.356.627.188.616.94
UnReGA [3]1005.115.705.425.805.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.366, + 0.784, + 0.42 + ], + "angle": 0, + "content": "Table 1: Comparison with state-of-the-art methods. We report the results by angular error in degrees, and use bold and underline to indicate the best and the second best result in each column for a specific task. \\(\\ddagger\\) means the model uses ResNet-50 as backbone, * shows that the experimental settings are different." + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.451, + 0.382, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.294, + 0.533, + 0.371, + 0.542 + ], + "angle": 0, + "content": "(a) CNN Baseline" + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.452, + 0.498, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.533, + 0.474, + 0.542 + ], + "angle": 0, + "content": "(b) PureGaze" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.452, + 0.612, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.533, + 0.579, + 0.542 + ], + "angle": 0, + "content": "(c) CDG" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.451, + 0.725, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.533, + 0.69, + 0.542 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.555, + 0.784, + 0.584 + ], + "angle": 0, + "content": "Fig. 4: Illustration of the feature distribution. Various colors indicate different gaze directions and similar gaze directions have similar colors. (Recommend viewing in color)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.785, + 0.736 + ], + "angle": 0, + "content": "In addition, we present a comparison with SOTA unsupervised domain adaptation (UDA) methods in the third row of Table 1. Note that UDA methods require the use of a small batch of unlabeled target domain samples, which is more costly than DG methods. It can be clearly seen from the table that our LG-Gaze achieves comparable performance without extra target domain data. Especially on the two tasks of \\(\\mathcal{D}_{\\mathrm{G}} \\to \\mathcal{D}_{\\mathrm{M}}\\) and \\(\\mathcal{D}_{\\mathrm{G}} \\to \\mathcal{D}_{\\mathrm{D}}\\), it is close to UDA methods. Therefore, this verifies the powerful performance of our proposed method, as it does not rely on any prior information from the target domain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Comparison of Feature Visualization. To evaluate and analyze the gaze features \\( f^g \\) extracted by different models, we follow the same scheme as CDG [30], and visualize the gaze features of the cross-domain task \\( \\mathcal{D}_{\\mathrm{G}} \\to \\mathcal{D}_{\\mathrm{D}} \\) using t-SNE [19] to compare different methods. Figure 4 shows the visualization results from all DG methods in the second row of Table 1, where feature points with" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "Y. Pengwei et al." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.144, + 0.787, + 0.25 + ], + "angle": 0, + "content": "
MethodsDE→DMDE→DdDG→DMDG→DdAvg
CNN Baseline8.359.667.589.018.65
Vanilla CLIP [22]82.4379.6482.4379.6481.04
CoOp [41]9.2311.0613.1510.8511.07
OridinalCLIP [15]7.8810.119.129.749.21
Our LG-Gaze6.457.226.836.866.84
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.216, + 0.251, + 0.786, + 0.277 + ], + "angle": 0, + "content": "Table 2: Comparisons to SOTA contrastive learning methods. Bold indicates the best results in each column, and underline denote the second best result in each column." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.788, + 0.461 + ], + "angle": 0, + "content": "similar gaze labels have similar colors. For the baseline model, the features of different gaze labels are mixed together, which is unreasonable for the regression task as shown in Figure 4 (a). The ideal feature distribution should be continuous and gradual with labels for the regression task [36]. Figure 4 (b,c) represents several classic DG methods that have been enhanced compared to the baseline, but they exhibit unreasonable feature distributions. In general, LG-Gaze has the most reasonable feature distribution, and the visualization result is shown in Figure 4 (d). The strong correlation between gaze direction and features indicates that guiding visual feature learning through language is effective." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.492, + 0.787, + 0.508 + ], + "angle": 0, + "content": "4.3 Performance Comparison with SOTA VLM Regression Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.788, + 0.618 + ], + "angle": 0, + "content": "Table 2 presents the performance of various methods that employ VLM models. To investigate the applicability of ordinal regression methods to gaze estimation, we discretize the gaze labels into integers. This mapping allows the yaw and pitch in the gaze label to be categorized into a finite number of categories. During testing, we predict yaw and pitch separately, but the accuracy of the gaze vector, which is composed of yaw and pitch, is evaluated." + }, + { + "type": "table", + "bbox": [ + 0.234, + 0.646, + 0.776, + 0.769 + ], + "angle": 0, + "content": "
MethodsDE→DMDE→DdDG→DMDG→DdAvg
CNN Baseline8.359.667.589.018.65
KL-Loss6.777.366.977.087.05
CDG [30]6.987.457.147.187.19
RNC [35]6.637.286.996.966.97
MoCo [11]6.527.336.957.046.96
MCR6.457.226.836.866.84
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.214, + 0.769, + 0.787, + 0.824 + ], + "angle": 0, + "content": "Table 3: Comparison of different text prompt tuning methods for gaze model cross-domain evaluation in four tasks. The rest of the settings are consistent for all methods, except for the contrastive loss function. Bold indicates the best results in each column, and underline denote the second best result results in each column." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.43, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LG-Gaze: Language-Guided Gaze Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.788, + 0.373 + ], + "angle": 0, + "content": "For Vanilla CLIP, there is no need to use source data for model training. Instead, a zero-shot method is used for prediction, resulting in only two outcomes. First, we construct the text prompt \"gaze estimation: the yaw/pitch degree angle of the person is {yaw/pitch}.\" Then, we obtain the model's prediction of yaw and pitch by maximizing the alignment between text features and image features, which is essentially in line with the original image-text matching. However, we can observe from Table 2 that Vanilla CLIP exhibits poor performance, possibly due to its struggle in distinguishing ordinal concepts. For CoOp, the gaze prediction performance can be greatly enhanced through text prompt tuning, but it is weaker than the CNN baseline. This could be due to its inability to learn the ordinal relationship between samples. For OrdinalCLIP, although it can learn the ordinal relationship, the discretization of labels may lead to errors. Additionally, following the methodology of OrdinalCLIP, we have created a visualization of the semantic labels constructed by different VLM methods in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.395, + 0.767, + 0.424 + ], + "angle": 0, + "content": "4.4 Performance Comparison with SOTA Contrastive Regression Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.434, + 0.788, + 0.556 + ], + "angle": 0, + "content": "Table 3 presents various methods for feature alignment or contrastive learning in gaze estimation. For KL-Loss, it primarily constrains the similarity matrix between gaze features and text features, as well as the similarity matrix composed of gaze labels. We follow the ideas of CDG [30], RNC [35], and MoCo [11], and transform them into image-to-text and text-to-image contrastive loss functions similar to LG-Gaze. From Table 3, we can see that our MCR loss function outperforms other methods significantly, attributed to the inclusion of constructed global negative samples and the weight coefficient for negative samples." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.577, + 0.389, + 0.592 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.601, + 0.788, + 0.692 + ], + "angle": 0, + "content": "Effectiveness of Our Framework. Table 4 shows the performance of different combinations of loss functions. Keeping only \\(\\ell_{Gaze}\\) is equivalent to CNN baseline, and combining all the losses is equivalent to full LG-Gaze. Compared with CNN baseline, using MCR loss to learn robust features with language guidance can significantly improve the performance. Further adding \\(\\ell_{Geo}\\) enhances the geometric relevance between prompts, and thus improves the model performance." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.713, + 0.788, + 0.773 + ], + "angle": 0, + "content": "Effectiveness of Interpolation Methods. To compare the performance of global linear interpolation, bilinear interpolation and our proposed geometry-aware spherical interpolation. Our method demonstrates superior performance due to its consideration of geometric accuracy, as shown in Table 5." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.794, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Effectiveness of MCR Loss. To delve into the efficacy of our MCR loss, we setted varying numbers of negative samples to compare their cross-domain performance, as illustrated in Table 6. When the number of negative samples is zero," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "Y. Pengwei et al." + }, + { + "type": "table", + "bbox": [ + 0.238, + 0.144, + 0.768, + 0.214 + ], + "angle": 0, + "content": "
lGeolMCRlGazeD_E → D_MD_E → D_DD_G → D_MD_G → D_DAvg
××√8.359.667.589.018.65
×√√6.767.356.936.987.01
√√√6.457.226.836.866.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.239, + 0.214, + 0.761, + 0.227 + ], + "angle": 0, + "content": "Table 4: Ablation study on loss functions. Results are reported by in degrees." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.254, + 0.774, + 0.321 + ], + "angle": 0, + "content": "
MethodsDE→DMDE→DdDG→DMDG→DdAvg
Global Li-Inter7.197.547.297.337.34
Bilinear-Inter6.787.316.926.96.98
Ours6.457.226.836.866.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.295, + 0.321, + 0.707, + 0.334 + ], + "angle": 0, + "content": "Table 5: The performance of different interpolation methods." + }, + { + "type": "table", + "bbox": [ + 0.25, + 0.361, + 0.761, + 0.469 + ], + "angle": 0, + "content": "
Number\\( \\mathcal{D}_{\\mathrm{E}}\\rightarrow \\mathcal{D}_{\\mathrm{M}} \\)\\( \\mathcal{D}_{\\mathrm{E}}\\rightarrow \\mathcal{D}_{\\mathrm{D}} \\)\\( \\mathcal{D}_{\\mathrm{G}}\\rightarrow \\mathcal{D}_{\\mathrm{M}} \\)\\( \\mathcal{D}_{\\mathrm{G}}\\rightarrow \\mathcal{D}_{\\mathrm{D}} \\)Avg
07.007.447.137.267.21
646.877.317.087.027.07
1286.667.216.987.056.98
2566.457.226.836.866.84
5126.427.196.856.816.82
10246.427.186.836.796.81
" + }, + { + "type": "table_caption", + "bbox": [ + 0.273, + 0.47, + 0.728, + 0.483 + ], + "angle": 0, + "content": "Table 6: The Comparison for different number of negative samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.785, + 0.617 + ], + "angle": 0, + "content": "it equates to a common contrastive regression loss, which is observed to achieve less than remarkable performance. As the number of negative samples increases from 64 to 256, there is a marked improvement in performance, indicating that a higher quantity of negative samples in contrastive learning contributes to performance gains. However, when the number of negative samples is further increased to 1024, the performance nearly reaches convergence, with minimal gains." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.785, + 0.647 + ], + "angle": 0, + "content": "Moreover, we conducted an experiment on loss weights to explore the effectiveness of our method. For more details, please see the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.671, + 0.36, + 0.687 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.785, + 0.809 + ], + "angle": 0, + "content": "In this paper, we propose a novel training framework named LG-Gaze for gaze estimation. Leveraging textual features extracted from language models, LG-Gaze guides the learning of gaze features. This approach enables the gaze model to acquire robust gaze representations and achieve a reasonable feature distribution, ultimately enhancing the cross-domain generalization capability of gaze model. Our proposed LG-Gaze achieves state-of-the-art performance on domain generalization for gaze estimation task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Acknowledgements. This work was sponsored by National Key R&D Program of China (2023YFE0204200)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.43, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LG-Gaze: Language-Guided Gaze Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Andrist, S., Tan, X.Z., Gleicher, M., Mutlu, B.: Conversational gaze aversion for humanlike robots. In: 2014 9th ACM/IEEE International Conference on Human-Robot Interaction (HRI). pp. 25-32 (2014) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.26 + ], + "angle": 0, + "content": "2. Bao, Y., Liu, Y., Wang, H., Lu, F.: Generalizing gaze estimation with rotation consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4207-4216 (2022) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.301 + ], + "angle": 0, + "content": "3. Cai, X., Zeng, J., Shan, S., Chen, X.: Source-free adaptive gaze estimation by uncertainty reduction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22035-22045 (2023) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.302, + 0.785, + 0.342 + ], + "angle": 0, + "content": "4. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.371 + ], + "angle": 0, + "content": "5. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.412 + ], + "angle": 0, + "content": "6. Cheng, Y., Bao, Y.: Puregaze: Purifying gaze feature for generalizable gaze estimation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 436-443 (2022) 2, 4, 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.453 + ], + "angle": 0, + "content": "7. Cheng, Y., Lu, F., Zhang, X.: Appearance-based gaze estimation via evaluation-guided asymmetric regression. In: Proceedings of the European Conference on Computer Vision (ECCV) (September 2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.454, + 0.785, + 0.481 + ], + "angle": 0, + "content": "8. Cheng, Y., Zhang, X., Lu, F., Sato, Y.: Gaze estimation by exploring two-eye asymmetry. IEEE Transactions on Image Processing 29, 5259-5272 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.785, + 0.536 + ], + "angle": 0, + "content": "9. Funes Mora, K.A., Monay, F., Odobez, J.M.: Eyediap: A database for the development and evaluation of gaze estimation algorithms from rgb and rgb-d cameras. In: Proceedings of the symposium on eye tracking research and applications. pp. 255-258 (2014) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.785, + 0.564 + ], + "angle": 0, + "content": "0. González, Á.: Measurement of areas on a sphere using fibonacci and latitude-longitude lattices. Mathematical Geosciences 42, 49-64 (2010) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "1. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) 5, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.647 + ], + "angle": 0, + "content": "2. Kellnhofer, P., Recasens, A., Stent, S., Matusik, W., Torralba, A.: Gaze360: Physically unconstrained gaze estimation in the wild. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6912-6921 (2019) 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "3. Krafka, K., Khosla, A., Kellnhofer, P., Kannan, H., Bhandarkar, S.M., Matusik, W., Torralba, A.: Eye tracking for everyone. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 2176-2184 (2016) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.73 + ], + "angle": 0, + "content": "4. Lee, I., Yun, J.S., Kim, H.H., Na, Y., Yoo, S.B.: Latentgaze: Cross-domain gaze estimation through gaze-aware analytic latent code manipulation. In: Proceedings of the Asian Conference on Computer Vision. pp. 3379-3395 (2022) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.785, + 0.771 + ], + "angle": 0, + "content": "5. Li, W., Huang, X., Zhu, Z., Tang, Y., Li, X., Zhou, J., Lu, J.: Ordinalclip: Learning rank prompts for language-guided ordinal regression. Advances in Neural Information Processing Systems 35, 35313-35325 (2022) 2, 4, 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "6. Liu, R., Bao, Y., Xu, M., Wang, H., Liu, Y., Lu, F.: Jitter does matter: Adapting gaze estimation to new domains. arXiv preprint arXiv:2210.02082 (2022) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Liu, Y., Liu, R., Wang, H., Lu, F.: Generalizing gaze estimation with outlier-guided collaborative adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2021) 11" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.129 + ], + "angle": 0, + "content": "Y. Pengwei et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "19. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.257 + ], + "angle": 0, + "content": "20. Mavely, A.G., Judith, J.E., Sahal, P.A., Kuruvilla, S.A.: Eye gaze tracking based driver monitoring system. In: 2017 IEEE International Conference on Circuits and Systems (ICCS). pp. 364-367 (2017). https://doi.org/10.1109/ICCS1.2017.8326022 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.327 + ], + "angle": 0, + "content": "21. Padmanaban, N., Konrad, R., Cooper, E.A., Wetzstein, G.: Optimizing vr for all users through adaptive focus displays. In: ACM SIGGRAPH 2017 Talks. SIGGRAPH '17, Association for Computing Machinery, New York, NY, USA (2017). https://doi.org/10.1145/3084363.3085029, https://doi.org/10.1145/3084363.3085029 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.383 + ], + "angle": 0, + "content": "22. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 2, 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "23. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog p. 9 (2019) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.466 + ], + "angle": 0, + "content": "24. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Denseclip: Language-guided dense prediction with context-aware prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18082-18091 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "25. Shen, S., Li, W., Wang, X., Zhang, D., Jin, Z., Zhou, J., Lu, J.: Clip-cluster: Clip-guided attribute hallucination for face clustering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 20786-20795 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.549 + ], + "angle": 0, + "content": "26. Shoemake, K.: Animating rotation with quaternion curves. In: Proceedings of the 12th annual conference on Computer graphics and interactive techniques. pp. 245-254 (1985) 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "27. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems (2017) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.633 + ], + "angle": 0, + "content": "28. Vidit, V., Engilberge, M., Salzmann, M.: Clip the gap: A single domain generalization approach for object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3219-3229 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.785, + 0.674 + ], + "angle": 0, + "content": "29. Wang, R., Li, P., Huang, H., Cao, C., He, R., He, Z.: Learning-to-rank meets language: Boosting language-driven ordering alignment for ordinal classification. Advances in Neural Information Processing Systems 36 (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.73 + ], + "angle": 0, + "content": "30. Wang, Y., Jiang, Y., Li, J., Ni, B., Dai, W., Li, C., Xiong, H., Li, T.: Contrastive regression for domain adaptation on gaze estimation. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19354-19363 (2022). https://doi.org/10.1109/CVPR52688.2022.01877 2, 4, 5, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.785, + 0.771 + ], + "angle": 0, + "content": "31. Xu, M., Wang, H., Lu, F.: Learning a generalized gaze estimator from gaze-consistent feature. Proceedings of the AAAI Conference on Artificial Intelligence 37(3), 3027-3035 (Jun 2023) 2, 4, 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. Yao, L., Han, J., Wen, Y., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, C., Xu, H.: Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 9125-9138. Curran Associates, Inc. (2022), https://proceedings.neurips.cc/paper_files/" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.43, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LG-Gaze: Language-Guided Gaze Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.148, + 0.785, + 0.174 + ], + "angle": 0, + "content": "paper/2022/file/3ba960559212691be13fa81d9e5e0047-Paper-Conference.pdf 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.217 + ], + "angle": 0, + "content": "33. Yin, P., Wang, J., Dai, J., Wu, X.: Nerf-gaze: A head-eye redirection parametric model for gaze estimation. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "34. Yin, P., Zeng, G., Wang, J., Xie, D.: Clip-gaze: Towards general gaze estimation via visual-linguistic model. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 6729–6737 (2024) 2, 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "35. Zha, K., Cao, P., Son, J., Yang, Y., Katabi, D.: Rank-n-contrast: Learning continuous representations for regression. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) 4, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.341 + ], + "angle": 0, + "content": "36. Zha, K., Cao, P., Son, J., Yang, Y., Katabi, D.: Rank-n-contrast: Learning continuous representations for regression. Advances in Neural Information Processing Systems 36 (2024) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.384 + ], + "angle": 0, + "content": "37. Zhang, X., Park, S., Beeler, T., Bradley, D., Tang, S., Hilliges, O.: Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze variation. In: European Conference on Computer Vision. pp. 365-381. Springer (2020) 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "38. Zhang, X., Sugano, Y., Fritz, M., Bulling, A.: Appearance-based gaze estimation in the wild. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 4511-4520 (2015) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "39. Zhang, X., Sugano, Y., Fritz, M., Bulling, A.: Mpiigaze: Real-world dataset and deep appearance-based gaze estimation. IEEE transactions on pattern analysis and machine intelligence 41(1), 162-175 (2017) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "40. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.535 + ], + "angle": 0, + "content": "41. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) 4, 12" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.535 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/LG-Gaze_ Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation/82982acf-4d4b-4f4d-a9b4-ee45c99a1bec_origin.pdf b/2024/LG-Gaze_ Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation/82982acf-4d4b-4f4d-a9b4-ee45c99a1bec_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..39e364c765b3bb7db2498151fa53db1163741a01 --- /dev/null +++ b/2024/LG-Gaze_ Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation/82982acf-4d4b-4f4d-a9b4-ee45c99a1bec_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a18ede80e3d834ca9fe222e77c9687796147aaa38251d27142b01d0746065321 +size 1405732 diff --git a/2024/LG-Gaze_ Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation/full.md b/2024/LG-Gaze_ Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..de4d4e788314ddb64ff7cf36c1a9e34726c44a63 --- /dev/null +++ b/2024/LG-Gaze_ Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation/full.md @@ -0,0 +1,295 @@ +# LG-Gaze: Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation + +Pengwei Yin*, Jingjing Wang*, Guanzhong Zeng, Di Xie, and Jiang Zhu + +Hikvision Research Institute, Hangzhou, China {yinpengwei, wangjingjing9, zengguanzhong, xiedi, zhujiang.hri}@hikvision.com + +Abstract. The ability of gaze estimation models to generalize is often significantly hindered by various factors unrelated to gaze, especially when the training dataset is limited. Current strategies aim to address this challenge through different domain generalization techniques, yet they have had limited success due to the risk of overfitting when solely relying on value labels for regression. Recent progress in pre-trained vision-language models has motivated us to capitalize on the abundant semantic information available. We propose a novel approach in this paper, reframing the gaze estimation task as a vision-language alignment issue. Our proposed framework, named Language-Guided Gaze Estimation (LG-Gaze), learns continuous and geometry-sensitive features for gaze estimation benefit from the rich prior knowledge of vision-language models. Specifically, LG-Gaze aligns gaze features with continuous linguistic features through our proposed multimodal contrastive regression loss, which customizes adaptive weights for different negative samples. Furthermore, to better adapt to the labels for gaze estimation task, we propose a geometry-aware interpolation method to obtain more precise gaze embeddings. Through extensive experiments, we validate the efficacy of our framework in four different cross-domain evaluation tasks. + +Keywords: Gaze Estimation $\cdot$ Vision-Language Model $\cdot$ Continuous Regression Task $\cdot$ Contrastive Learning + +# 1 Introduction + +The gaze estimation is crucial for understanding human behavior. Precise gaze estimation offers significant assistance for many applications, such as human-computer interaction [1], augmented reality [21], and driver monitoring systems [20]. The deep learning has led to significant improvements in gaze estimation based on appearance. While these methods show impressive results in evaluations within the same domain, they tend to experience a decline in performance when evaluated on different domains. This decline is primarily attributed to overfitting rather than learning robust features from the original domain. + +![](images/cd66d2d4a0db9ec2ad0105ec28b765a233834089de243d74d429afdb2c248896.jpg) +(a) Traditional Methods: Supervised by Value Label +(b) LG-Gaze: Guided by Language Model +Fig. 1: (a) The traditional method of gaze generalization involves overseeing model training by means of numerical label regression. (b) In our study, we introduce a text models to steer the development of robust features in visual models. + +Gaze data encompass diverse elements, such as appearance, wearables, environments, and image clarity [31, 34], yet gaze annotations predominantly focus on eye direction, marginalizing other variables as noise. Such irrelevant factors exacerbate the domain disparity, hindering gaze models' adaptability. Additionally, supervisory numeric labels might induce overfitting, influenced by these extraneous gaze components. + +As shown in Figure 1(a), training a gaze model typically involves utilizing a vision encoder to extract features and then employing a regressor for prediction. Traditional methods train a gaze model by using data disturbing [31], feature enhancement [30], and adversarial learning [6] to improve the generalization ability of the gaze model. However, these regularization techniques have limited effectiveness as they rely solely on numerical labels for supervision, making them susceptible to overfitting due to irrelevant factors. The above may indicate that it is difficult to guide the model to learn ideal robust features through these indirect regularization terms. + +To counter overfitting in gaze representation learning, we advocate for multimodal integration. Language, rich in semantics and knowledge, complements visual data. Each gaze label is described textually, "the yaw/pitch degree angle of the person is {yaw/pitch}.". Leveraging visual-language models, such as CLIP, which excel in generalized latent space learning from vast image-text pairs, we align image features with a language space. This distillation of CLIP's language expertise acts as regularization, enhancing generalization. However, these methods are only applicable to a limited set of discrete scalars and are not suitable for continuous tasks such as gaze estimation. Aligning gaze features to discrete text features can result in inaccurate feature alignment, which in turn affects the performance of continuous regression tasks. + +To address these issues, we propose a novel training method named LG-Gaze, which is designed to learn continuous and geometry-sensitive features for gaze estimation, as depicted in Figure 1(b). This framework aligns gaze features with continuous linguistic features extracted by a powerful language model, which not only prevents model overfitting but also enhances the generalization capabilities of gaze model. Primarily, unlike common discrete multimodal contrastive learning methods [15, 22], we introduce a loss function for continuous multimodal contrastive learning. Our Multimodal Contrastive Regression loss + +function (MCR) customizes adaptive weights for different negative samples according to label distance, facilitating more refined feature alignment in the feature space and benefiting the model in learning more continuous features. Next, to better adapt to the vectorial property of labels for gaze estimation task, we propose a geometry-aware interpolation method to obtain more precise gaze embeddings. The geometry-aware interpolation method combines spherical [26] and bilinear interpolation techniques to resolve the issue of creating semantic labels for gaze, ensuring accurate gaze prompts. Furthermore, MCR also resolves the limitations of previous contrastive learning functions by utilizing a more uniform distribution of global negative samples for contrast. Unlike most methods that are limited to intra-batch comparisons, our loss function benefits from extensive global contrast, resulting in a reasonable and robust feature distribution. + +The main contributions can be summarized as follows: + +- In this paper, we propose a novel training framework called LG-Gaze for gaze estimation. LG-Gaze guides the gaze model to learn generalized representations by leveraging textual features extracted from language models, resulting in more robust gaze representations. By incorporating rich language knowledge, this approach achieves robust gaze features, ultimately enhancing the cross-domain generalization capability of the gaze model. +- To learn continuous feature for regression tasks, we propose a new multimodal contrastive learning loss function named MCR. The method employs adaptive weights for different negative sample, helping learn more refined features. Moreover, MCR leverages ample global negative samples for contrast, leading to a reasonable feature distribution. +- We introduce a geometry-aware interpolation method that employs spherical interpolation techniques to compute accurate gaze embeddings, thereby ensuring the accuracy of semantic labels. +- Experimental results and visualizations demonstrate that LG-Gaze not only achieves remarkable performance compared to the baseline but also surpasses state-of-the-art domain generalization methods for gaze estimation. + +# 2 Related Works + +# 2.1 Gaze Estimation + +Appearance-based gaze estimation has garnered attention in recent years [7,8, 13,38]. However, it faces challenges in cross-domain evaluation due to the domain gap resulting from various gaze-irrelevant factors. Common approaches often rely on diverse gaze datasets [12,37] for training, aiming to equip models with robust generalization capabilities. Unfortunately, collecting gaze data remains costly, and the diversity of available data remains limited. To address this issue, we propose enhancing gaze estimation models through domain generalization (DG) methods. These methods enable models to generalize to unseen distributions and enhance cross-domain performance. Most existing DG techniques are tailored for classification tasks, leaving a gap in the context of gaze estimation. + +PureGaze [6] debuts a self-adversarial schema, discarding irrelevant gaze cues, enhancing pertinent ones. Xu et al. [31] and NeRF-Gaze [33] similarly counteract gaze distractions via adversarial data modifications and augmentation. However, residual gaze-confounding factors persist, hindering estimation precision. CLIP-Gaze [34] is designed to segregate gaze features from predefined textual gaze-irrelevant features, thereby enhancing the generalization capability and robustness of gaze features. Our LG-Gaze, by directly aligning with language features, also exhibits resilience against gaze-irrelevant factors. More critically, the text features in LG-Gaze possess favorable rank properties, attributed to the geometry-aware continuous prompts learning. + +# 2.2 Vision Language Model + +In recent research, several studies have utilized the text manipulation and visual alignment capabilities of CLIP (Contrastive Language-Image Pretraining) [22] to improve opendetection and generalization performance in specific tasks. Notable examples include DetCLIP [32], CLIP-Gap [28], and CLIP-Cluster [25]. To further enhance the performance of vision-language models on downstream tasks, an effective strategy involves learning more contextually appropriate prompts through text prompt tuning [40, 41]. Additionally, DenseCLIP [24] explores the application of semantic knowledge into monocular depth estimation tasks. By matching visual features with textual features, this method achieves zero-shot monocular depth estimation. "Teaching CLIP to Count to Ten" and CrowdCLIP propose visual language models for understanding quantities, enabling object counting and crowd estimation. While these methods enhance models' understanding of quantities, they do not address regression tasks. OrdinalCLIP [15] and L2RCLIP [29] are capable of addressing ordinal regression tasks. However, they are not as effective for continuous regression tasks. In this paper, our method aligns gaze features with continuous linguistic features through our proposed multimodal contrastive regression loss (MCR) and the geometry-aware interpolation method. + +# 2.3 Contrastive Learning + +In computer vision, SimCLR [4] and CLIP [22] pioneer contrastive loss, exceeding in image classification and retrieval. Yet, these techniques aren't suited for regression, where prediction targets are continuous, not categorical. Rank-N-Contrast (RNC) [35] employs sample rankings to enhance continuity understanding. Nonetheless, RNC presumes infinite distance between negative and positive pairs, neglecting the correlation between label and feature proximities. Contrastive Domain Generalization (CDG) [30] leverages contrastive loss to encourage the clustering of features corresponding to similar gaze directions while separating those associated with significantly different gaze directions. However, CDG relies on empirical thresholding to define positive and negative sample pairs, which limits its applicability for continuous regression tasks. In our research, we introduce a novel contrastive loss function specifically designed for + +continuous regression tasks. Our approach prioritizes sample ordering while considering label distances. Unlike previous contrastive learning methods that are confined to constructing sample pairs within batches, our method demonstrates a more targeted enhancement. To elaborate further, we draw inspiration from the Momentum Contrast (MoCo) framework [11], which utilizes a queue to store negative sample features. However, these negative samples are obtained from n steps back in time and exhibit inherent uncontrollable randomness. Unlike delayed updates, our approach synchronously updates these samples, ensuring a controlled distribution instead of relying on randomness. + +# 3 Method + +# 3.1 Problem Statement + +In the context of gaze estimation tasks, which fundamentally belong to regression tasks, we define the data for the source domain as $\mathcal{D}_s = (\pmb {I}_i,\pmb {g}_i)_{i = 1}^M$ , where $(\pmb {I}_i,\pmb {g}_i)$ represents the $i$ -th pair of image $\pmb {I}_i\in \mathbb{R}^{224\times 224\times 3}$ and their corresponding gaze direction $\pmb {g}_i\in \mathbb{R}^3$ . Here, $M$ denotes the total number of pairs in $\mathcal{D}_s$ . Our objective in gaze estimation is to train a neural network composed of an image encoder $\mathcal{E}_v(\cdot):\mathbb{R}^{224\times 224\times 3}\to \mathbb{R}^d$ to extract gaze representations $\pmb{f}_i^g$ based on input image $\mathcal{I}_i$ and an MLP regressor $\mathcal{R}_g(\cdot):\mathbb{R}^d\to \mathbb{R}^3$ to predict gaze directions $\pmb {g}_i$ . Typically, once we have trained a robust gaze representation, training $\mathcal{R}_g$ becomes relatively straightforward using regression functions (e.g., $L_{1}$ or AngleLoss). + +However, gaze estimation research has encountered a persistent challenge that hinders its advancement: How can we design robust gaze representations $\{f_i^g\}_{i=1}^M$ that are continuous, semantic, and sensitive to geometric factors for gaze estimation? + +Although many previous methods explore additional technologies to enhance feature representations, such as adversarial training [6], contrastive learning as an auxiliary task [30] and mitigating the impact of gaze-irrelevant factors [31], they often focus solely on finding regularization terms to improve model performance without explicitly aiming for robust and well-distributed features. Consequently, existing methods often face overfitting issues and struggle to achieve the desired performance. In this paper, we propose a novel language-guided gaze representation learning framework that leverages rich prior knowledge from a natural language model. Our approach aims to learn robust and well-distributed gaze representations, addressing the aforementioned challenges. + +# 3.2 LG-Gaze Framework + +To harness the full power of language, we utilize the text encoder from CLIP as our language model. CLIP adopts a vision-language pre-training framework and learns representations from image-text pairs, thereby constructing a joint vision-language latent space. Given the strong performance of the CLIP model in downstream tasks, we implemented our method using the original CLIP text encoder from the referenced paper. + +![](images/a7a4efe40ad6232923f5142fce103ccd87432d55edecb724f1e4d76b100d68ce.jpg) +Fig. 2: We reformulate the task as an image-language matching problem, which mainly consists of a trainable prompt, a frozen text encoder, a trainable image encoder. + +Suppose that each training step is provided with a data batch $(I_i, g_i)_{i=1}^B$ , where $g_i$ represents the gaze label of the image $I_i$ . Our goal is to construct semantic labels and use a language model to extract text features. This process will help the gaze model learn robust features by aligning them with the extracted features. To leverage the prior knowledge within text features, we begin by constructing a learnable prompt $\{\pmb{v}_1, \pmb{v}_2, \dots, \pmb{v}_{L-1}\} \in \mathbb{R}^{512 \times (L-1)}$ as the context for describing gaze estimation tasks instead of handcrafting the prompt context. Simultaneously, we treat gaze labels as individual words. For each sample, we concatenate the learnable prompt with the embedding $\pmb{v}_i^g \in \mathbb{R}^{512}$ to form a sequence $\pmb{T}_i = \{\pmb{v}_1, \pmb{v}_2, \dots, \pmb{v}_{L-1}, \pmb{v}_i^g\}$ as a semantic label. Specifically, we utilize the gaze embedding as the final token in the sequence $\pmb{T}_i \in \mathbb{R}^{512 \times L}$ . + +To ensure the continuity, ordinality, and geometric properties of gaze label embeddings, we introduce $N$ learnable embedding anchors, denoted as $\{\pmb{A}_j\}_{j=1}^N$ . Each anchor corresponds to a unique gaze vector, uniformly distributed across the entire label space. We utilize a nearest-neighbor interpolation method to represent gaze label embeddings associated with sample relationships. This approach yields a continuous and geometry-aware representation of gaze embeddings, as detailed in Section 3.3. + +In our research, we send image $\pmb{I}_i$ to an image encoder $\mathcal{E}_v$ to extract the image feature $\pmb{f}_i^g$ . Consequently, the features of a batch can be represented as $\{\pmb{f}_i^g\}_{i=1}^B$ . Simultaneously, we utilize the language model $\mathcal{E}_t$ to obtain corresponding text features $\{\pmb{f}_i^t\}_{i=1}^B$ from $\{\pmb{T}_i\}_{i=1}^B$ . While the parameters of the language model remain frozen, we train the entire image encoder to align image features with the language feature space. Subsequently, we utilize an image-text contrastive loss to optimize the network, which includes both an image-to-text loss and a text-to-image loss, following the methodology of CLIP. Detailed information about our proposed contrastive learning method will be provided in Section 3.4. + +After aligning the features, we obtain robust gaze representations. Finally, the image features are fed into a regressor denoted as $\mathcal{R}_g$ , yielding the predicted gaze direction $\hat{g}_i \in \mathbb{R}^3$ . Specifically, we predict through supervised learning, which is defined as: + +$$ +\mathcal {L} _ {G a z e} \left(\hat {\boldsymbol {g}} _ {i}, \boldsymbol {g} _ {i}\right) = \operatorname {a r c c o s} \left(\frac {\hat {\boldsymbol {g}} _ {i} \cdot \boldsymbol {g} _ {i}}{\| \hat {\boldsymbol {g}} _ {i} \| \| \boldsymbol {g} _ {i} \|}\right) \tag {1} +$$ + +![](images/033aec55813a9be3dcbeefff1f5884736d6f21eb0a75c9499efee0d931465aed.jpg) +(a) + +![](images/e9f7f381ffcd0d676f276b5cf23cdb942d4f2526e59ae15c03c38aff14e99528.jpg) +(b) +Fig. 3: These four subfigures contain all steps of our proposed interpolation method. + +![](images/905df1fe32afa36bb6b3a60891a104df9ba4790a79720c2356de9b9845d9fe85.jpg) +(c) + +![](images/6c28bcfb26141a4c535ef31a63ca03f62e0693a6cded3393a3622e3df3d46d40.jpg) +(d) + +# 3.3 Geometry-aware Continuous Gaze Prompts + +In this section, to better represent gaze labels in the embedding space, we construct a gaze embedding for each label. This embedding should have the capability for continuous representation and geometric attributes. To address the conflict between the finite vocabulary in natural language processing and the infinite nature of gaze labels, we interpolate each gaze embedding $\pmb{v}_i^g$ through $N$ gaze anchor embeddings $\{\pmb{A_j}\}_{j = 1}^N$ . + +Common approaches, such as those described in [15], employ global linear interpolation to represent $\pmb{v}_i^g$ , which interpolates the target embedding using all anchor embeddings. The weight $w_{i,j}$ for each anchor embedding is computed by Equation 2, determined by the similarity between its corresponding gaze vector $\pmb{g}_j$ and the target $\pmb{g}_i$ . However, global interpolation cannot accurately represent $\pmb{v}_i^g$ because anchors that are not proximate fail to provide effective information. + +$$ +w _ {i, j} = \frac {\cos \left(\boldsymbol {g} _ {i} , \boldsymbol {g} _ {j}\right)}{\sum_ {j = 1} ^ {N} \cos \left(\boldsymbol {g} _ {i} , \boldsymbol {g} _ {j}\right)} \tag {2} +$$ + +To further enhance interpolation, we attempted to employ a bilinear interpolation method. However, as it still utilizes a linear representation, this method remains insufficiently precise in expressing the geometric relationships between gaze vectors. + +To address this issue, we decided to adopt a spherical interpolation method [26], which can accurately calculate the interpolation weights between vectors. As illustrated in Figure 3(a), a flat spherical interpolation method is depicted. For the vector $\pmb{g}_i$ positioned between $\pmb{g}_1$ and $\pmb{g}_2$ , we initially compute $\theta = \arccos(\langle \pmb{g}_1, \pmb{g}_2 \rangle)$ and $\theta_1 = \arccos(\langle \pmb{g}_1, \pmb{g}_i \rangle)$ . Subsequently, we compute the scalar $t$ by $t = \theta_1 / \theta$ , then we apply the spherical interpolation Equation 3 to determine the corresponding weights for $\pmb{g}_1$ and $\pmb{g}_2$ . Finally, we can represent the target vector $\pmb{g}_i$ through the summation of these weights. Ultimately, we confirm that this method of weight calculation is accurate. + +$$ +\boldsymbol {g} _ {i} = \frac {\sin ((1 - t) \theta)}{\sin (\theta)} \boldsymbol {g} _ {1} + \frac {\sin (t \theta)}{\sin (\theta)} \boldsymbol {g} _ {2} \tag {3} +$$ + +Building upon the aforementioned spherical interpolation method, we further integrate the concept of bilinear interpolation to derive the interpolation + +weights for each gaze label on a stereographic sphere. As depicted in Figure 3(b), we initially identify the four nearest anchor embeddings $A_{1-4}$ based on their proximity. Subsequently, as shown in Figure 3(c), we calculate the weights $w_{a,1}$ and $w_{a,2}$ for $A_1$ and $A_2$ using Equation 3. Similarly, we can determine $w_{b,3}$ and $w_{b,4}$ . In Figure 3(d), we proceed to obtain $w_{i,a}$ and $w_{i,b}$ using a similar method. Ultimately, we acquire the associated weights for $A_{1-4}$ to interpolate $\pmb{v}_i^g$ , which can be expressed as follows: + +$$ +\boldsymbol {v} _ {i} ^ {g} = \left(w _ {i, a} * w _ {a, 1}\right) \cdot \boldsymbol {A} _ {1} + \left(w _ {i, a} * w _ {a, 2}\right) \cdot \boldsymbol {A} _ {2} + \left(w _ {i, b} * w _ {b, 3}\right) \cdot \boldsymbol {A} _ {3} + \left(w _ {i, b} * w _ {b, 4}\right) \cdot \boldsymbol {A} _ {4} \tag {4} +$$ + +Additionally, to enhance the geometric relationships between gaze embeddings, we employ a simple loss function to constrain anchor embeddings $\{A_j\}_{j=1}^N$ . As shown in Equation 5, the similarity between anchor embeddings should be consistent with their relationships in the label space. + +$$ +\ell_ {\text {G e o}} = \frac {1}{M * M} \sum_ {i = 1} ^ {M} \sum_ {j = 1} ^ {M} L _ {1} \left(\cos \left(\boldsymbol {A} _ {i}, \boldsymbol {A} _ {j}\right), \cos \left(\boldsymbol {g} _ {i}, \boldsymbol {g} _ {j}\right)\right) \tag {5} +$$ + +# 3.4 Continuous Multimodal Contrastive Regression Loss Function + +Mathematically, from the aforementioned steps, we obtain image features $\{\pmb{f}_i^g\}_{i=1}^B$ and their corresponding text features $\{\pmb{f}_i^t\}_{i=1}^B$ . Both CLIP and OrdinalCLIP utilize the InfoNCE loss to align the multimodal features. To delve into specifics, consider an image-text feature pair $(\pmb{f}_i^g, \pmb{f}_i^t)$ . Here, $\pmb{f}_i^g / \pmb{f}_i^t$ serves as the positive sample for $\pmb{f}_i^t / \pmb{f}_i^g$ , as they are associated with the same gaze label. Meanwhile, all other image/text features in a mini-batch are treated as negative samples and are consequently pushed away from $\pmb{f}_i^t / \pmb{f}_i^g$ . However, this training objective poses challenges for regression tasks because it overlooks the finer-grained distance ordering relationships between samples. Consequently, failing to differentiate the relative distances between image/text and their corresponding negative samples will inevitably weaken the learning effectiveness, especially in cross-modal representation learning for regression tasks. + +Considering the proximity relationships between data, for the $i$ -th sample, we introduce a weight parameter $\{w(i,j)\}_{j=1}^{B}$ for each negative sample. This parameter is computed based on the distances between labels of samples. Our proposed multimodal contrastive regression loss (MCR) effectively controls the feature distances among samples, preventing negative samples from being excessively contracted. As a result, we learn a semantically meaningful representation space tailored for regression tasks, ensuring consistency between sample feature relationships and the label space. The proposed text-to-image MCR loss is formulated as follows. + +$$ +\ell_ {t 2 i} = - \frac {1}{B} \sum_ {i = 1} ^ {B} \log \left(\frac {\exp \left(\sin \left(\boldsymbol {f} _ {i} ^ {t} , \boldsymbol {f} _ {i} ^ {g}\right)\right)}{\exp \left(\sin \left(\boldsymbol {f} _ {i} ^ {t} , \boldsymbol {f} _ {i} ^ {g}\right)\right) + \sum_ {j = 1} ^ {B} w (i , j) \cdot \exp \left(\sin \left(\boldsymbol {f} _ {i} ^ {t} , \boldsymbol {f} _ {j} ^ {g}\right)\right)}\right) \tag {6} +$$ + +where $\text{sim}(\pmb{f}_i^t, \pmb{f}_i^g) = \cos(\pmb{f}_i^t, \pmb{f}_i^g)$ and $w(i,j) = \cos(\pmb{g}_i, \pmb{g}_j)$ denotes the contrastive weight of the $j$ -th negative image sample with respect to the $i$ -th text feature in our LG-Gaze framework. The weight parameter should be directly proportional to the label distance between negative samples. As the distance between labels increases, the penalty on negative sample weights should also increase. This encourages feature embeddings to adhere to real-world relationships. + +The MCR loss for image-to-text can be formulated in a similar manner. However, what sets it apart is our deliberate expansion of the number of negative samples to facilitate the learning of a robust representation. To achieve this, we draw inspiration from MoCo, which utilizes a dynamic queue to maintain a diverse set of negative samples, thereby enhancing the effectiveness of contrastive learning. Nevertheless, this approach still faces some challenges. For instance, if the queue update rate is too slow, the negative samples stored in the queue may become outdated and fail to align with the current model. Additionally, due to the inherent randomness and limited diversity of negative samples, there could be discrepancies between the queue data and the entire data distribution. To address this issue, we utilize the flexibility of language labels by creating $K$ global text features $\{f_i^t\}_{i=1}^K$ as negative samples (where $K$ is significantly larger than $B$ ). These text labels correspond to gaze vectors that show a uniformly dense distribution in the label space. Furthermore, these $K$ text labels are interpolated based on $N$ anchor $\{A_j\}_{j=1}^N$ . As a result, the proposed image-to-text MCR loss can be formulated as follows: + +$$ +\ell_ {i 2 t} = - \frac {1}{B} \sum_ {i = 1} ^ {B} \log \left(\frac {\exp \left(\operatorname {s i m} \left(\boldsymbol {f} _ {i} ^ {g} , \boldsymbol {f} _ {i} ^ {t}\right)\right)}{\exp \left(\operatorname {s i m} \left(\boldsymbol {f} _ {i} ^ {g} , \boldsymbol {f} _ {i} ^ {t}\right)\right) + \sum_ {j = 1} ^ {B + K} w (i , j) \cdot \exp \left(\operatorname {s i m} \left(\boldsymbol {f} _ {i} ^ {g} , \boldsymbol {f} _ {j} ^ {t}\right)\right)}\right) \tag {7} +$$ + +Obviously, we can observe that the difference between $\ell_{i2t}$ and $\ell_{t2i}$ lies in the number of negative samples. We concatenate $K$ global negative samples with the $B$ negative samples from mini-batches to form the complete set of negative samples. Through this operation, the global negative samples remain relatively fixed, mitigating issues related to randomness. Additionally, their uniform distribution helps alleviate discrepancies between the negative samples and the entire data distribution. Furthermore, by computing features based on the current model, we address potential outdated information. Finally, our total MCR loss can be denoted as: + +$$ +\ell_ {M C R} = \ell_ {i 2 t} + \ell_ {t 2 i} \tag {8} +$$ + +# 3.5 Overall Objective + +In summary, the overall objective implemented in our framework is: + +$$ +\mathcal {L} = \lambda_ {1} \ell_ {G e o} + \lambda_ {2} \ell_ {M C R} + \lambda_ {3} \ell_ {G a z e} +$$ + +$\lambda_1, \lambda_2, \lambda_3$ are hyperparameters, and we empirically set $\lambda_1 = \lambda_2 = \lambda_3 = 1.0$ . + +# 4 Experiments + +# 4.1 Experiment Details + +Data Preparation We evaluate gaze estimation method on four cross-domain tasks, utilizing ETH-XGaze [37] and Gaze360 [12] as the training datasets, and MPIIFaceGaze [39] and Eye-Diap [9] as the test datasets. We denote them as $\mathcal{D}_{\mathrm{E}}$ (ETH-XGaze) $\rightarrow \mathcal{D}_{\mathrm{M}}$ (MPIIFaceGaze), $\mathcal{D}_{\mathrm{E}} \rightarrow \mathcal{D}_{\mathrm{D}}$ (EyeDiap), $\mathcal{D}_{\mathrm{G}}$ (Gaze360) $\rightarrow \mathcal{D}_{\mathrm{M}}$ , $\mathcal{D}_{\mathrm{G}} \rightarrow \mathcal{D}_{\mathrm{D}}$ . See the supplementary material for more details. + +Vision Model Implementation Details We use ResNet-18 as our gaze feature extractor $\mathcal{E}_v$ , and a fully connected layer to regress a 3-dimensional gaze vector. We resize and normalize all the images to $224\times 224$ pixels and scale the pixel values between 0 and 1. We set the batch size to 64. We maintain the same setup for training the source domain models on $\mathcal{D}_E$ and $\mathcal{D}_G$ . + +Language Model Implementation Details The architecture of the text encoder is adapted from the Transformer [27], incorporating modifications outlined by Radford [23]. The dimensions for both text and gaze features have been standardized to 1024. Our models are constructed using the foundational open-source code of CLIP. The class token is strategically placed at the terminal position of the sequence, with the context tokens quantity designated as $L = 10$ . Additionally, we set the yaw angle range for the anchors to be $[-180^{\circ}, +179^{\circ}]$ and the pitch range to be $[-90^{\circ}, +90^{\circ}]$ with a step size of $30^{\circ}$ , resulting in a total of 91 anchors. We set the number of negative samples to 256, which are uniformly distributed in the gaze label space using the Fibonacci sphere algorithm [10]. + +Training Implementation Details We conducted the experiments on a single Tesla V100 GPU. Specifically, we use the SGD optimizer with Nesterov momentum, a learning rate (LR) of $5 \times 10^{-2}$ and a weight decay of $1 \times 10^{-5}$ for the parameters. We train for 30 epochs using a Cosineannealing LR scheduler [18] with a 3-epoch warm-up. We apply a data augmentation technique involving a random color field and grayscale, as described in [5]. + +# 4.2 Performance Comparison with SOTA Gaze Estimation Methods + +Comparison of Cross-domain Evaluation. Table 1 presents the quantitative results of four cross-domain tests. The second row compares our LG-Gaze with the SOTA domain DG methods for gaze estimation. The CNN baseline method refers to the one that relies solely on the vision encoder to extract the single visual modality feature and uses $\ell_{gaze}$ for supervision. In summary, the LG-Gaze achieves the best overall performance and demonstrates state-of-the-art results on three cross-domain evaluation tasks. It also achieves the second-best performance for $\mathcal{D}_{\mathrm{E}} \rightarrow \mathcal{D}_{\mathrm{M}}$ , highlighting the effectiveness of our method. + +
TaskMethods| Dt|DE→DMDE→DdDG→DMDG→DdAvg
DGCNN Baseline08.479.327.548.938.57
PureGaze [6]07.087.489.289.328.29
CDG [30]‡06.737.957.037.277.25
Xu et al. [31]06.507.447.559.037.63
CLIP-Gaze [34]06.417.516.897.066.97
Our LG-Gaze06.457.226.836.866.84
UDAPnP-GA * [17]105.535.876.187.926.38
RUDA [2]1005.706.296.205.866.01
CRGA [30]>05.485.665.896.495.88
LatentGaze [14]1005.217.81--6.51
Liu et al. [16]1005.356.627.188.616.94
UnReGA [3]1005.115.705.425.805.51
+ +Table 1: Comparison with state-of-the-art methods. We report the results by angular error in degrees, and use bold and underline to indicate the best and the second best result in each column for a specific task. $\ddagger$ means the model uses ResNet-50 as backbone, * shows that the experimental settings are different. + +![](images/dfde9f82953867fdeba0ccd0616215f08bb787a8a8d4f6480a6bd00e1f0f9d64.jpg) +(a) CNN Baseline + +![](images/6902da96e17fd5701284d8e5e5c233712fa637d786bb190822914e6fadf1f03f.jpg) +(b) PureGaze +Fig. 4: Illustration of the feature distribution. Various colors indicate different gaze directions and similar gaze directions have similar colors. (Recommend viewing in color). + +![](images/4e04de218db6f8d6f6a3ed8d34d44f77d9a314296542b33d1f9fb892e692f6ed.jpg) +(c) CDG + +![](images/783a01144f579992f4f82be927a111de510155cb42ed23d131d2c6bb09634d92.jpg) +(d) Ours + +In addition, we present a comparison with SOTA unsupervised domain adaptation (UDA) methods in the third row of Table 1. Note that UDA methods require the use of a small batch of unlabeled target domain samples, which is more costly than DG methods. It can be clearly seen from the table that our LG-Gaze achieves comparable performance without extra target domain data. Especially on the two tasks of $\mathcal{D}_{\mathrm{G}} \to \mathcal{D}_{\mathrm{M}}$ and $\mathcal{D}_{\mathrm{G}} \to \mathcal{D}_{\mathrm{D}}$ , it is close to UDA methods. Therefore, this verifies the powerful performance of our proposed method, as it does not rely on any prior information from the target domain. + +Comparison of Feature Visualization. To evaluate and analyze the gaze features $f^g$ extracted by different models, we follow the same scheme as CDG [30], and visualize the gaze features of the cross-domain task $\mathcal{D}_{\mathrm{G}} \to \mathcal{D}_{\mathrm{D}}$ using t-SNE [19] to compare different methods. Figure 4 shows the visualization results from all DG methods in the second row of Table 1, where feature points with + +
MethodsDE→DMDE→DdDG→DMDG→DdAvg
CNN Baseline8.359.667.589.018.65
Vanilla CLIP [22]82.4379.6482.4379.6481.04
CoOp [41]9.2311.0613.1510.8511.07
OridinalCLIP [15]7.8810.119.129.749.21
Our LG-Gaze6.457.226.836.866.84
+ +Table 2: Comparisons to SOTA contrastive learning methods. Bold indicates the best results in each column, and underline denote the second best result in each column. + +similar gaze labels have similar colors. For the baseline model, the features of different gaze labels are mixed together, which is unreasonable for the regression task as shown in Figure 4 (a). The ideal feature distribution should be continuous and gradual with labels for the regression task [36]. Figure 4 (b,c) represents several classic DG methods that have been enhanced compared to the baseline, but they exhibit unreasonable feature distributions. In general, LG-Gaze has the most reasonable feature distribution, and the visualization result is shown in Figure 4 (d). The strong correlation between gaze direction and features indicates that guiding visual feature learning through language is effective. + +# 4.3 Performance Comparison with SOTA VLM Regression Methods + +Table 2 presents the performance of various methods that employ VLM models. To investigate the applicability of ordinal regression methods to gaze estimation, we discretize the gaze labels into integers. This mapping allows the yaw and pitch in the gaze label to be categorized into a finite number of categories. During testing, we predict yaw and pitch separately, but the accuracy of the gaze vector, which is composed of yaw and pitch, is evaluated. + +
MethodsDE→DMDE→DdDG→DMDG→DdAvg
CNN Baseline8.359.667.589.018.65
KL-Loss6.777.366.977.087.05
CDG [30]6.987.457.147.187.19
RNC [35]6.637.286.996.966.97
MoCo [11]6.527.336.957.046.96
MCR6.457.226.836.866.84
+ +Table 3: Comparison of different text prompt tuning methods for gaze model cross-domain evaluation in four tasks. The rest of the settings are consistent for all methods, except for the contrastive loss function. Bold indicates the best results in each column, and underline denote the second best result results in each column. + +For Vanilla CLIP, there is no need to use source data for model training. Instead, a zero-shot method is used for prediction, resulting in only two outcomes. First, we construct the text prompt "gaze estimation: the yaw/pitch degree angle of the person is {yaw/pitch}." Then, we obtain the model's prediction of yaw and pitch by maximizing the alignment between text features and image features, which is essentially in line with the original image-text matching. However, we can observe from Table 2 that Vanilla CLIP exhibits poor performance, possibly due to its struggle in distinguishing ordinal concepts. For CoOp, the gaze prediction performance can be greatly enhanced through text prompt tuning, but it is weaker than the CNN baseline. This could be due to its inability to learn the ordinal relationship between samples. For OrdinalCLIP, although it can learn the ordinal relationship, the discretization of labels may lead to errors. Additionally, following the methodology of OrdinalCLIP, we have created a visualization of the semantic labels constructed by different VLM methods in the supplementary material. + +# 4.4 Performance Comparison with SOTA Contrastive Regression Loss Functions + +Table 3 presents various methods for feature alignment or contrastive learning in gaze estimation. For KL-Loss, it primarily constrains the similarity matrix between gaze features and text features, as well as the similarity matrix composed of gaze labels. We follow the ideas of CDG [30], RNC [35], and MoCo [11], and transform them into image-to-text and text-to-image contrastive loss functions similar to LG-Gaze. From Table 3, we can see that our MCR loss function outperforms other methods significantly, attributed to the inclusion of constructed global negative samples and the weight coefficient for negative samples. + +# 4.5 Ablation Study + +Effectiveness of Our Framework. Table 4 shows the performance of different combinations of loss functions. Keeping only $\ell_{Gaze}$ is equivalent to CNN baseline, and combining all the losses is equivalent to full LG-Gaze. Compared with CNN baseline, using MCR loss to learn robust features with language guidance can significantly improve the performance. Further adding $\ell_{Geo}$ enhances the geometric relevance between prompts, and thus improves the model performance. + +Effectiveness of Interpolation Methods. To compare the performance of global linear interpolation, bilinear interpolation and our proposed geometry-aware spherical interpolation. Our method demonstrates superior performance due to its consideration of geometric accuracy, as shown in Table 5. + +Effectiveness of MCR Loss. To delve into the efficacy of our MCR loss, we setted varying numbers of negative samples to compare their cross-domain performance, as illustrated in Table 6. When the number of negative samples is zero, + +
lGeolMCRlGazeD_E → D_MD_E → D_DD_G → D_MD_G → D_DAvg
××√8.359.667.589.018.65
×√√6.767.356.936.987.01
√√√6.457.226.836.866.84
+ +Table 4: Ablation study on loss functions. Results are reported by in degrees. + +
MethodsDE→DMDE→DdDG→DMDG→DdAvg
Global Li-Inter7.197.547.297.337.34
Bilinear-Inter6.787.316.926.96.98
Ours6.457.226.836.866.84
+ +Table 5: The performance of different interpolation methods. + +
Number\( \mathcal{D}_{\mathrm{E}}\rightarrow \mathcal{D}_{\mathrm{M}} \)\( \mathcal{D}_{\mathrm{E}}\rightarrow \mathcal{D}_{\mathrm{D}} \)\( \mathcal{D}_{\mathrm{G}}\rightarrow \mathcal{D}_{\mathrm{M}} \)\( \mathcal{D}_{\mathrm{G}}\rightarrow \mathcal{D}_{\mathrm{D}} \)Avg
07.007.447.137.267.21
646.877.317.087.027.07
1286.667.216.987.056.98
2566.457.226.836.866.84
5126.427.196.856.816.82
10246.427.186.836.796.81
+ +Table 6: The Comparison for different number of negative samples. + +it equates to a common contrastive regression loss, which is observed to achieve less than remarkable performance. As the number of negative samples increases from 64 to 256, there is a marked improvement in performance, indicating that a higher quantity of negative samples in contrastive learning contributes to performance gains. However, when the number of negative samples is further increased to 1024, the performance nearly reaches convergence, with minimal gains. + +Moreover, we conducted an experiment on loss weights to explore the effectiveness of our method. For more details, please see the supplementary material. + +# 5 Conclusion + +In this paper, we propose a novel training framework named LG-Gaze for gaze estimation. Leveraging textual features extracted from language models, LG-Gaze guides the learning of gaze features. This approach enables the gaze model to acquire robust gaze representations and achieve a reasonable feature distribution, ultimately enhancing the cross-domain generalization capability of gaze model. Our proposed LG-Gaze achieves state-of-the-art performance on domain generalization for gaze estimation task. + +Acknowledgements. This work was sponsored by National Key R&D Program of China (2023YFE0204200). + +# References + +1. Andrist, S., Tan, X.Z., Gleicher, M., Mutlu, B.: Conversational gaze aversion for humanlike robots. In: 2014 9th ACM/IEEE International Conference on Human-Robot Interaction (HRI). pp. 25-32 (2014) 1 +2. Bao, Y., Liu, Y., Wang, H., Lu, F.: Generalizing gaze estimation with rotation consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4207-4216 (2022) 11 +3. Cai, X., Zeng, J., Shan, S., Chen, X.: Source-free adaptive gaze estimation by uncertainty reduction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22035-22045 (2023) 11 +4. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020) 4 +5. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020) 10 +6. Cheng, Y., Bao, Y.: Puregaze: Purifying gaze feature for generalizable gaze estimation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 436-443 (2022) 2, 4, 5, 11 +7. Cheng, Y., Lu, F., Zhang, X.: Appearance-based gaze estimation via evaluation-guided asymmetric regression. In: Proceedings of the European Conference on Computer Vision (ECCV) (September 2018) 3 +8. Cheng, Y., Zhang, X., Lu, F., Sato, Y.: Gaze estimation by exploring two-eye asymmetry. IEEE Transactions on Image Processing 29, 5259-5272 (2020) 3 +9. Funes Mora, K.A., Monay, F., Odobez, J.M.: Eyediap: A database for the development and evaluation of gaze estimation algorithms from rgb and rgb-d cameras. In: Proceedings of the symposium on eye tracking research and applications. pp. 255-258 (2014) 10 +0. González, Á.: Measurement of areas on a sphere using fibonacci and latitude-longitude lattices. Mathematical Geosciences 42, 49-64 (2010) 10 +1. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) 5, 12, 13 +2. Kellnhofer, P., Recasens, A., Stent, S., Matusik, W., Torralba, A.: Gaze360: Physically unconstrained gaze estimation in the wild. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6912-6921 (2019) 3, 10 +3. Krafka, K., Khosla, A., Kellnhofer, P., Kannan, H., Bhandarkar, S.M., Matusik, W., Torralba, A.: Eye tracking for everyone. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 2176-2184 (2016) 3 +4. Lee, I., Yun, J.S., Kim, H.H., Na, Y., Yoo, S.B.: Latentgaze: Cross-domain gaze estimation through gaze-aware analytic latent code manipulation. In: Proceedings of the Asian Conference on Computer Vision. pp. 3379-3395 (2022) 11 +5. Li, W., Huang, X., Zhu, Z., Tang, Y., Li, X., Zhou, J., Lu, J.: Ordinalclip: Learning rank prompts for language-guided ordinal regression. Advances in Neural Information Processing Systems 35, 35313-35325 (2022) 2, 4, 7, 12 +6. Liu, R., Bao, Y., Xu, M., Wang, H., Liu, Y., Lu, F.: Jitter does matter: Adapting gaze estimation to new domains. arXiv preprint arXiv:2210.02082 (2022) 11 +7. Liu, Y., Liu, R., Wang, H., Lu, F.: Generalizing gaze estimation with outlier-guided collaborative adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2021) 11 + +18. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) 10 +19. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) 11 +20. Mavely, A.G., Judith, J.E., Sahal, P.A., Kuruvilla, S.A.: Eye gaze tracking based driver monitoring system. In: 2017 IEEE International Conference on Circuits and Systems (ICCS). pp. 364-367 (2017). https://doi.org/10.1109/ICCS1.2017.8326022 1 +21. Padmanaban, N., Konrad, R., Cooper, E.A., Wetzstein, G.: Optimizing vr for all users through adaptive focus displays. In: ACM SIGGRAPH 2017 Talks. SIGGRAPH '17, Association for Computing Machinery, New York, NY, USA (2017). https://doi.org/10.1145/3084363.3085029, https://doi.org/10.1145/3084363.3085029 1 +22. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 2, 4, 12 +23. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog p. 9 (2019) 10 +24. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Denseclip: Language-guided dense prediction with context-aware prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18082-18091 (2022) 4 +25. Shen, S., Li, W., Wang, X., Zhang, D., Jin, Z., Zhou, J., Lu, J.: Clip-cluster: Clip-guided attribute hallucination for face clustering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 20786-20795 (2023) 4 +26. Shoemake, K.: Animating rotation with quaternion curves. In: Proceedings of the 12th annual conference on Computer graphics and interactive techniques. pp. 245-254 (1985) 3, 7 +27. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems (2017) 10 +28. Vidit, V., Engilberge, M., Salzmann, M.: Clip the gap: A single domain generalization approach for object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3219-3229 (2023) 4 +29. Wang, R., Li, P., Huang, H., Cao, C., He, R., He, Z.: Learning-to-rank meets language: Boosting language-driven ordering alignment for ordinal classification. Advances in Neural Information Processing Systems 36 (2024) 4 +30. Wang, Y., Jiang, Y., Li, J., Ni, B., Dai, W., Li, C., Xiong, H., Li, T.: Contrastive regression for domain adaptation on gaze estimation. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19354-19363 (2022). https://doi.org/10.1109/CVPR52688.2022.01877 2, 4, 5, 11, 12, 13 +31. Xu, M., Wang, H., Lu, F.: Learning a generalized gaze estimator from gaze-consistent feature. Proceedings of the AAAI Conference on Artificial Intelligence 37(3), 3027-3035 (Jun 2023) 2, 4, 5, 11 +32. Yao, L., Han, J., Wen, Y., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, C., Xu, H.: Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 9125-9138. Curran Associates, Inc. (2022), https://proceedings.neurips.cc/paper_files/ + +paper/2022/file/3ba960559212691be13fa81d9e5e0047-Paper-Conference.pdf 4 +33. Yin, P., Wang, J., Dai, J., Wu, X.: Nerf-gaze: A head-eye redirection parametric model for gaze estimation. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2024) 4 +34. Yin, P., Zeng, G., Wang, J., Xie, D.: Clip-gaze: Towards general gaze estimation via visual-linguistic model. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 6729–6737 (2024) 2, 4, 11 +35. Zha, K., Cao, P., Son, J., Yang, Y., Katabi, D.: Rank-n-contrast: Learning continuous representations for regression. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) 4, 12, 13 +36. Zha, K., Cao, P., Son, J., Yang, Y., Katabi, D.: Rank-n-contrast: Learning continuous representations for regression. Advances in Neural Information Processing Systems 36 (2024) 12 +37. Zhang, X., Park, S., Beeler, T., Bradley, D., Tang, S., Hilliges, O.: Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze variation. In: European Conference on Computer Vision. pp. 365-381. Springer (2020) 3, 10 +38. Zhang, X., Sugano, Y., Fritz, M., Bulling, A.: Appearance-based gaze estimation in the wild. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 4511-4520 (2015) 3 +39. Zhang, X., Sugano, Y., Fritz, M., Bulling, A.: Mpiigaze: Real-world dataset and deep appearance-based gaze estimation. IEEE transactions on pattern analysis and machine intelligence 41(1), 162-175 (2017) 10 +40. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022) 4 +41. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) 4, 12 \ No newline at end of file diff --git a/2024/LG-Gaze_ Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation/images.zip b/2024/LG-Gaze_ Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3b53ede4fc8b16aeb930d9ef322c2395768e8af2 --- /dev/null +++ b/2024/LG-Gaze_ Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3153fc89e06633a6c8429449ac9a936f1556b686cef1373e1de183af830d130 +size 426208 diff --git a/2024/LG-Gaze_ Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation/layout.json b/2024/LG-Gaze_ Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9757c9ddff368a7d067b39ab23cb8f51c2879582 --- /dev/null +++ b/2024/LG-Gaze_ Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation/layout.json @@ -0,0 +1,9366 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 112, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 112, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 138, + 112, + 477, + 148 + ], + "type": "text", + "content": "LG-Gaze: Learning Geometry-aware Continuous Prompts for Language-Guided Gaze Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 168, + 474, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 168, + 474, + 181 + ], + "spans": [ + { + "bbox": [ + 140, + 168, + 474, + 181 + ], + "type": "text", + "content": "Pengwei Yin*, Jingjing Wang*, Guanzhong Zeng, Di Xie, and Jiang Zhu" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 192, + 190, + 421, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 190, + 421, + 224 + ], + "spans": [ + { + "bbox": [ + 192, + 190, + 421, + 224 + ], + "type": "text", + "content": "Hikvision Research Institute, Hangzhou, China {yinpengwei, wangjingjing9, zengguanzhong, xiedi, zhujiang.hri}@hikvision.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 259, + 455, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 259, + 455, + 469 + ], + "spans": [ + { + "bbox": [ + 160, + 259, + 455, + 469 + ], + "type": "text", + "content": "Abstract. The ability of gaze estimation models to generalize is often significantly hindered by various factors unrelated to gaze, especially when the training dataset is limited. Current strategies aim to address this challenge through different domain generalization techniques, yet they have had limited success due to the risk of overfitting when solely relying on value labels for regression. Recent progress in pre-trained vision-language models has motivated us to capitalize on the abundant semantic information available. We propose a novel approach in this paper, reframing the gaze estimation task as a vision-language alignment issue. Our proposed framework, named Language-Guided Gaze Estimation (LG-Gaze), learns continuous and geometry-sensitive features for gaze estimation benefit from the rich prior knowledge of vision-language models. Specifically, LG-Gaze aligns gaze features with continuous linguistic features through our proposed multimodal contrastive regression loss, which customizes adaptive weights for different negative samples. Furthermore, to better adapt to the labels for gaze estimation task, we propose a geometry-aware interpolation method to obtain more precise gaze embeddings. Through extensive experiments, we validate the efficacy of our framework in four different cross-domain evaluation tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 478, + 453, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 478, + 453, + 501 + ], + "spans": [ + { + "bbox": [ + 159, + 478, + 453, + 501 + ], + "type": "text", + "content": "Keywords: Gaze Estimation " + }, + { + "bbox": [ + 159, + 478, + 453, + 501 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 478, + 453, + 501 + ], + "type": "text", + "content": " Vision-Language Model " + }, + { + "bbox": [ + 159, + 478, + 453, + 501 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 478, + 453, + 501 + ], + "type": "text", + "content": " Continuous Regression Task " + }, + { + "bbox": [ + 159, + 478, + 453, + 501 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 478, + 453, + 501 + ], + "type": "text", + "content": " Contrastive Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 518, + 230, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 518, + 230, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 518, + 230, + 530 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 541, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 541, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 541, + 482, + 638 + ], + "type": "text", + "content": "The gaze estimation is crucial for understanding human behavior. Precise gaze estimation offers significant assistance for many applications, such as human-computer interaction [1], augmented reality [21], and driver monitoring systems [20]. The deep learning has led to significant improvements in gaze estimation based on appearance. While these methods show impressive results in evaluations within the same domain, they tend to experience a decline in performance when evaluated on different domains. This decline is primarily attributed to overfitting rather than learning robust features from the original domain." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 643, + 280, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 280, + 654 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 280, + 654 + ], + "type": "text", + "content": "*These authors contributed equally." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 118, + 478, + 167 + ], + "blocks": [ + { + "bbox": [ + 138, + 118, + 478, + 167 + ], + "lines": [ + { + "bbox": [ + 138, + 118, + 478, + 167 + ], + "spans": [ + { + "bbox": [ + 138, + 118, + 478, + 167 + ], + "type": "image", + "image_path": "cd66d2d4a0db9ec2ad0105ec28b765a233834089de243d74d429afdb2c248896.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 169, + 275, + 178 + ], + "lines": [ + { + "bbox": [ + 136, + 169, + 275, + 178 + ], + "spans": [ + { + "bbox": [ + 136, + 169, + 275, + 178 + ], + "type": "text", + "content": "(a) Traditional Methods: Supervised by Value Label" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 169, + 425, + 178 + ], + "lines": [ + { + "bbox": [ + 312, + 169, + 425, + 178 + ], + "spans": [ + { + "bbox": [ + 312, + 169, + 425, + 178 + ], + "type": "text", + "content": "(b) LG-Gaze: Guided by Language Model" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 190, + 482, + 224 + ], + "lines": [ + { + "bbox": [ + 132, + 190, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 190, + 482, + 224 + ], + "type": "text", + "content": "Fig. 1: (a) The traditional method of gaze generalization involves overseeing model training by means of numerical label regression. (b) In our study, we introduce a text models to steer the development of robust features in visual models." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 251, + 479, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 479, + 323 + ], + "type": "text", + "content": "Gaze data encompass diverse elements, such as appearance, wearables, environments, and image clarity [31, 34], yet gaze annotations predominantly focus on eye direction, marginalizing other variables as noise. Such irrelevant factors exacerbate the domain disparity, hindering gaze models' adaptability. Additionally, supervisory numeric labels might induce overfitting, influenced by these extraneous gaze components." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 326, + 480, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 480, + 433 + ], + "type": "text", + "content": "As shown in Figure 1(a), training a gaze model typically involves utilizing a vision encoder to extract features and then employing a regressor for prediction. Traditional methods train a gaze model by using data disturbing [31], feature enhancement [30], and adversarial learning [6] to improve the generalization ability of the gaze model. However, these regularization techniques have limited effectiveness as they rely solely on numerical labels for supervision, making them susceptible to overfitting due to irrelevant factors. The above may indicate that it is difficult to guide the model to learn ideal robust features through these indirect regularization terms." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 436, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 567 + ], + "type": "text", + "content": "To counter overfitting in gaze representation learning, we advocate for multimodal integration. Language, rich in semantics and knowledge, complements visual data. Each gaze label is described textually, \"the yaw/pitch degree angle of the person is {yaw/pitch}.\". Leveraging visual-language models, such as CLIP, which excel in generalized latent space learning from vast image-text pairs, we align image features with a language space. This distillation of CLIP's language expertise acts as regularization, enhancing generalization. However, these methods are only applicable to a limited set of discrete scalars and are not suitable for continuous tasks such as gaze estimation. Aligning gaze features to discrete text features can result in inaccurate feature alignment, which in turn affects the performance of continuous regression tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "To address these issues, we propose a novel training method named LG-Gaze, which is designed to learn continuous and geometry-sensitive features for gaze estimation, as depicted in Figure 1(b). This framework aligns gaze features with continuous linguistic features extracted by a powerful language model, which not only prevents model overfitting but also enhances the generalization capabilities of gaze model. Primarily, unlike common discrete multimodal contrastive learning methods [15, 22], we introduce a loss function for continuous multimodal contrastive learning. Our Multimodal Contrastive Regression loss" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "Y. Pengwei et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "text", + "content": "function (MCR) customizes adaptive weights for different negative samples according to label distance, facilitating more refined feature alignment in the feature space and benefiting the model in learning more continuous features. Next, to better adapt to the vectorial property of labels for gaze estimation task, we propose a geometry-aware interpolation method to obtain more precise gaze embeddings. The geometry-aware interpolation method combines spherical [26] and bilinear interpolation techniques to resolve the issue of creating semantic labels for gaze, ensuring accurate gaze prompts. Furthermore, MCR also resolves the limitations of previous contrastive learning functions by utilizing a more uniform distribution of global negative samples for contrast. Unlike most methods that are limited to intra-batch comparisons, our loss function benefits from extensive global contrast, resulting in a reasonable and robust feature distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 260, + 387, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 260, + 387, + 270 + ], + "spans": [ + { + "bbox": [ + 147, + 260, + 387, + 270 + ], + "type": "text", + "content": "The main contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 279, + 479, + 479 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 279, + 479, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 279, + 479, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 279, + 479, + 348 + ], + "type": "text", + "content": "- In this paper, we propose a novel training framework called LG-Gaze for gaze estimation. LG-Gaze guides the gaze model to learn generalized representations by leveraging textual features extracted from language models, resulting in more robust gaze representations. By incorporating rich language knowledge, this approach achieves robust gaze features, ultimately enhancing the cross-domain generalization capability of the gaze model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 351, + 479, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 351, + 479, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 351, + 479, + 408 + ], + "type": "text", + "content": "- To learn continuous feature for regression tasks, we propose a new multimodal contrastive learning loss function named MCR. The method employs adaptive weights for different negative sample, helping learn more refined features. Moreover, MCR leverages ample global negative samples for contrast, leading to a reasonable feature distribution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 411, + 479, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 479, + 444 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 479, + 444 + ], + "type": "text", + "content": "- We introduce a geometry-aware interpolation method that employs spherical interpolation techniques to compute accurate gaze embeddings, thereby ensuring the accuracy of semantic labels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 446, + 479, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 446, + 479, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 446, + 479, + 479 + ], + "type": "text", + "content": "- Experimental results and visualizations demonstrate that LG-Gaze not only achieves remarkable performance compared to the baseline but also surpasses state-of-the-art domain generalization methods for gaze estimation." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 501, + 242, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 242, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 242, + 514 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 526, + 244, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 244, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 244, + 537 + ], + "type": "text", + "content": "2.1 Gaze Estimation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "Appearance-based gaze estimation has garnered attention in recent years [7,8, 13,38]. However, it faces challenges in cross-domain evaluation due to the domain gap resulting from various gaze-irrelevant factors. Common approaches often rely on diverse gaze datasets [12,37] for training, aiming to equip models with robust generalization capabilities. Unfortunately, collecting gaze data remains costly, and the diversity of available data remains limited. To address this issue, we propose enhancing gaze estimation models through domain generalization (DG) methods. These methods enable models to generalize to unseen distributions and enhance cross-domain performance. Most existing DG techniques are tailored for classification tasks, leaving a gap in the context of gaze estimation." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 263, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 263, + 91, + 447, + 102 + ], + "type": "text", + "content": "LG-Gaze: Language-Guided Gaze Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "PureGaze [6] debuts a self-adversarial schema, discarding irrelevant gaze cues, enhancing pertinent ones. Xu et al. [31] and NeRF-Gaze [33] similarly counteract gaze distractions via adversarial data modifications and augmentation. However, residual gaze-confounding factors persist, hindering estimation precision. CLIP-Gaze [34] is designed to segregate gaze features from predefined textual gaze-irrelevant features, thereby enhancing the generalization capability and robustness of gaze features. Our LG-Gaze, by directly aligning with language features, also exhibits resilience against gaze-irrelevant factors. More critically, the text features in LG-Gaze possess favorable rank properties, attributed to the geometry-aware continuous prompts learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 252, + 280, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 252, + 280, + 264 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 280, + 264 + ], + "type": "text", + "content": "2.2 Vision Language Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 271, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 486 + ], + "type": "text", + "content": "In recent research, several studies have utilized the text manipulation and visual alignment capabilities of CLIP (Contrastive Language-Image Pretraining) [22] to improve opendetection and generalization performance in specific tasks. Notable examples include DetCLIP [32], CLIP-Gap [28], and CLIP-Cluster [25]. To further enhance the performance of vision-language models on downstream tasks, an effective strategy involves learning more contextually appropriate prompts through text prompt tuning [40, 41]. Additionally, DenseCLIP [24] explores the application of semantic knowledge into monocular depth estimation tasks. By matching visual features with textual features, this method achieves zero-shot monocular depth estimation. \"Teaching CLIP to Count to Ten\" and CrowdCLIP propose visual language models for understanding quantities, enabling object counting and crowd estimation. While these methods enhance models' understanding of quantities, they do not address regression tasks. OrdinalCLIP [15] and L2RCLIP [29] are capable of addressing ordinal regression tasks. However, they are not as effective for continuous regression tasks. In this paper, our method aligns gaze features with continuous linguistic features through our proposed multimodal contrastive regression loss (MCR) and the geometry-aware interpolation method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 502, + 268, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 268, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 268, + 514 + ], + "type": "text", + "content": "2.3 Contrastive Learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "In computer vision, SimCLR [4] and CLIP [22] pioneer contrastive loss, exceeding in image classification and retrieval. Yet, these techniques aren't suited for regression, where prediction targets are continuous, not categorical. Rank-N-Contrast (RNC) [35] employs sample rankings to enhance continuity understanding. Nonetheless, RNC presumes infinite distance between negative and positive pairs, neglecting the correlation between label and feature proximities. Contrastive Domain Generalization (CDG) [30] leverages contrastive loss to encourage the clustering of features corresponding to similar gaze directions while separating those associated with significantly different gaze directions. However, CDG relies on empirical thresholding to define positive and negative sample pairs, which limits its applicability for continuous regression tasks. In our research, we introduce a novel contrastive loss function specifically designed for" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "Y. Pengwei et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "continuous regression tasks. Our approach prioritizes sample ordering while considering label distances. Unlike previous contrastive learning methods that are confined to constructing sample pairs within batches, our method demonstrates a more targeted enhancement. To elaborate further, we draw inspiration from the Momentum Contrast (MoCo) framework [11], which utilizes a queue to store negative sample features. However, these negative samples are obtained from n steps back in time and exhibit inherent uncontrollable randomness. Unlike delayed updates, our approach synchronously updates these samples, ensuring a controlled distribution instead of relying on randomness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 241, + 202, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 241, + 202, + 254 + ], + "spans": [ + { + "bbox": [ + 131, + 241, + 202, + 254 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 265, + 259, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 265, + 259, + 276 + ], + "spans": [ + { + "bbox": [ + 131, + 265, + 259, + 276 + ], + "type": "text", + "content": "3.1 Problem Statement" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "text", + "content": "In the context of gaze estimation tasks, which fundamentally belong to regression tasks, we define the data for the source domain as " + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s = (\\pmb {I}_i,\\pmb {g}_i)_{i = 1}^M" + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "inline_equation", + "content": "(\\pmb {I}_i,\\pmb {g}_i)" + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "text", + "content": "-th pair of image " + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "inline_equation", + "content": "\\pmb {I}_i\\in \\mathbb{R}^{224\\times 224\\times 3}" + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "text", + "content": " and their corresponding gaze direction " + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "inline_equation", + "content": "\\pmb {g}_i\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "text", + "content": " denotes the total number of pairs in " + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "text", + "content": ". Our objective in gaze estimation is to train a neural network composed of an image encoder " + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_v(\\cdot):\\mathbb{R}^{224\\times 224\\times 3}\\to \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "text", + "content": " to extract gaze representations " + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^g" + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "text", + "content": " based on input image " + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_i" + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "text", + "content": " and an MLP regressor " + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_g(\\cdot):\\mathbb{R}^d\\to \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "text", + "content": " to predict gaze directions " + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "inline_equation", + "content": "\\pmb {g}_i" + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "text", + "content": ". Typically, once we have trained a robust gaze representation, training " + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_g" + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "text", + "content": " becomes relatively straightforward using regression functions (e.g., " + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 130, + 283, + 482, + 391 + ], + "type": "text", + "content": " or AngleLoss)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 392, + 482, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 392, + 482, + 438 + ], + "spans": [ + { + "bbox": [ + 131, + 392, + 482, + 438 + ], + "type": "text", + "content": "However, gaze estimation research has encountered a persistent challenge that hinders its advancement: How can we design robust gaze representations " + }, + { + "bbox": [ + 131, + 392, + 482, + 438 + ], + "type": "inline_equation", + "content": "\\{f_i^g\\}_{i=1}^M" + }, + { + "bbox": [ + 131, + 392, + 482, + 438 + ], + "type": "text", + "content": " that are continuous, semantic, and sensitive to geometric factors for gaze estimation?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 440, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 559 + ], + "type": "text", + "content": "Although many previous methods explore additional technologies to enhance feature representations, such as adversarial training [6], contrastive learning as an auxiliary task [30] and mitigating the impact of gaze-irrelevant factors [31], they often focus solely on finding regularization terms to improve model performance without explicitly aiming for robust and well-distributed features. Consequently, existing methods often face overfitting issues and struggle to achieve the desired performance. In this paper, we propose a novel language-guided gaze representation learning framework that leverages rich prior knowledge from a natural language model. Our approach aims to learn robust and well-distributed gaze representations, addressing the aforementioned challenges." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 575, + 265, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 575, + 265, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 575, + 265, + 586 + ], + "type": "text", + "content": "3.2 LG-Gaze Framework" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "To harness the full power of language, we utilize the text encoder from CLIP as our language model. CLIP adopts a vision-language pre-training framework and learns representations from image-text pairs, thereby constructing a joint vision-language latent space. Given the strong performance of the CLIP model in downstream tasks, we implemented our method using the original CLIP text encoder from the referenced paper." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 263, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 263, + 91, + 447, + 102 + ], + "type": "text", + "content": "LG-Gaze: Language-Guided Gaze Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 116, + 478, + 196 + ], + "blocks": [ + { + "bbox": [ + 138, + 116, + 478, + 196 + ], + "lines": [ + { + "bbox": [ + 138, + 116, + 478, + 196 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 478, + 196 + ], + "type": "image", + "image_path": "a7a4efe40ad6232923f5142fce103ccd87432d55edecb724f1e4d76b100d68ce.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 205, + 482, + 228 + ], + "lines": [ + { + "bbox": [ + 131, + 205, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 131, + 205, + 482, + 228 + ], + "type": "text", + "content": "Fig. 2: We reformulate the task as an image-language matching problem, which mainly consists of a trainable prompt, a frozen text encoder, a trainable image encoder." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 250, + 482, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 250, + 482, + 383 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 482, + 383 + ], + "type": "text", + "content": "Suppose that each training step is provided with a data batch " + }, + { + "bbox": [ + 130, + 250, + 482, + 383 + ], + "type": "inline_equation", + "content": "(I_i, g_i)_{i=1}^B" + }, + { + "bbox": [ + 130, + 250, + 482, + 383 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 250, + 482, + 383 + ], + "type": "inline_equation", + "content": "g_i" + }, + { + "bbox": [ + 130, + 250, + 482, + 383 + ], + "type": "text", + "content": " represents the gaze label of the image " + }, + { + "bbox": [ + 130, + 250, + 482, + 383 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 130, + 250, + 482, + 383 + ], + "type": "text", + "content": ". Our goal is to construct semantic labels and use a language model to extract text features. This process will help the gaze model learn robust features by aligning them with the extracted features. To leverage the prior knowledge within text features, we begin by constructing a learnable prompt " + }, + { + "bbox": [ + 130, + 250, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\{\\pmb{v}_1, \\pmb{v}_2, \\dots, \\pmb{v}_{L-1}\\} \\in \\mathbb{R}^{512 \\times (L-1)}" + }, + { + "bbox": [ + 130, + 250, + 482, + 383 + ], + "type": "text", + "content": " as the context for describing gaze estimation tasks instead of handcrafting the prompt context. Simultaneously, we treat gaze labels as individual words. For each sample, we concatenate the learnable prompt with the embedding " + }, + { + "bbox": [ + 130, + 250, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\pmb{v}_i^g \\in \\mathbb{R}^{512}" + }, + { + "bbox": [ + 130, + 250, + 482, + 383 + ], + "type": "text", + "content": " to form a sequence " + }, + { + "bbox": [ + 130, + 250, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\pmb{T}_i = \\{\\pmb{v}_1, \\pmb{v}_2, \\dots, \\pmb{v}_{L-1}, \\pmb{v}_i^g\\}" + }, + { + "bbox": [ + 130, + 250, + 482, + 383 + ], + "type": "text", + "content": " as a semantic label. Specifically, we utilize the gaze embedding as the final token in the sequence " + }, + { + "bbox": [ + 130, + 250, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\pmb{T}_i \\in \\mathbb{R}^{512 \\times L}" + }, + { + "bbox": [ + 130, + 250, + 482, + 383 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 384, + 482, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 466 + ], + "type": "text", + "content": "To ensure the continuity, ordinality, and geometric properties of gaze label embeddings, we introduce " + }, + { + "bbox": [ + 130, + 384, + 482, + 466 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 384, + 482, + 466 + ], + "type": "text", + "content": " learnable embedding anchors, denoted as " + }, + { + "bbox": [ + 130, + 384, + 482, + 466 + ], + "type": "inline_equation", + "content": "\\{\\pmb{A}_j\\}_{j=1}^N" + }, + { + "bbox": [ + 130, + 384, + 482, + 466 + ], + "type": "text", + "content": ". Each anchor corresponds to a unique gaze vector, uniformly distributed across the entire label space. We utilize a nearest-neighbor interpolation method to represent gaze label embeddings associated with sample relationships. This approach yields a continuous and geometry-aware representation of gaze embeddings, as detailed in Section 3.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 467, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 482, + 574 + ], + "type": "text", + "content": "In our research, we send image " + }, + { + "bbox": [ + 130, + 467, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\pmb{I}_i" + }, + { + "bbox": [ + 130, + 467, + 482, + 574 + ], + "type": "text", + "content": " to an image encoder " + }, + { + "bbox": [ + 130, + 467, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_v" + }, + { + "bbox": [ + 130, + 467, + 482, + 574 + ], + "type": "text", + "content": " to extract the image feature " + }, + { + "bbox": [ + 130, + 467, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^g" + }, + { + "bbox": [ + 130, + 467, + 482, + 574 + ], + "type": "text", + "content": ". Consequently, the features of a batch can be represented as " + }, + { + "bbox": [ + 130, + 467, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\{\\pmb{f}_i^g\\}_{i=1}^B" + }, + { + "bbox": [ + 130, + 467, + 482, + 574 + ], + "type": "text", + "content": ". Simultaneously, we utilize the language model " + }, + { + "bbox": [ + 130, + 467, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_t" + }, + { + "bbox": [ + 130, + 467, + 482, + 574 + ], + "type": "text", + "content": " to obtain corresponding text features " + }, + { + "bbox": [ + 130, + 467, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\{\\pmb{f}_i^t\\}_{i=1}^B" + }, + { + "bbox": [ + 130, + 467, + 482, + 574 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 467, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\{\\pmb{T}_i\\}_{i=1}^B" + }, + { + "bbox": [ + 130, + 467, + 482, + 574 + ], + "type": "text", + "content": ". While the parameters of the language model remain frozen, we train the entire image encoder to align image features with the language feature space. Subsequently, we utilize an image-text contrastive loss to optimize the network, which includes both an image-to-text loss and a text-to-image loss, following the methodology of CLIP. Detailed information about our proposed contrastive learning method will be provided in Section 3.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 574, + 482, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 622 + ], + "type": "text", + "content": "After aligning the features, we obtain robust gaze representations. Finally, the image features are fed into a regressor denoted as " + }, + { + "bbox": [ + 130, + 574, + 482, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_g" + }, + { + "bbox": [ + 130, + 574, + 482, + 622 + ], + "type": "text", + "content": ", yielding the predicted gaze direction " + }, + { + "bbox": [ + 130, + 574, + 482, + 622 + ], + "type": "inline_equation", + "content": "\\hat{g}_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 574, + 482, + 622 + ], + "type": "text", + "content": ". Specifically, we predict through supervised learning, which is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 227, + 631, + 481, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 631, + 481, + 658 + ], + "spans": [ + { + "bbox": [ + 227, + 631, + 481, + 658 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {G a z e} \\left(\\hat {\\boldsymbol {g}} _ {i}, \\boldsymbol {g} _ {i}\\right) = \\operatorname {a r c c o s} \\left(\\frac {\\hat {\\boldsymbol {g}} _ {i} \\cdot \\boldsymbol {g} _ {i}}{\\| \\hat {\\boldsymbol {g}} _ {i} \\| \\| \\boldsymbol {g} _ {i} \\|}\\right) \\tag {1}", + "image_path": "72edab9bfdb5bd19c2fcbd57de0c2f6ca04eec277679a7616db7766e126e19da.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "Y. Pengwei et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 114, + 236, + 192 + ], + "blocks": [ + { + "bbox": [ + 153, + 114, + 236, + 192 + ], + "lines": [ + { + "bbox": [ + 153, + 114, + 236, + 192 + ], + "spans": [ + { + "bbox": [ + 153, + 114, + 236, + 192 + ], + "type": "image", + "image_path": "033aec55813a9be3dcbeefff1f5884736d6f21eb0a75c9499efee0d931465aed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 183, + 196, + 194, + 204 + ], + "lines": [ + { + "bbox": [ + 183, + 196, + 194, + 204 + ], + "spans": [ + { + "bbox": [ + 183, + 196, + 194, + 204 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 238, + 118, + 310, + 190 + ], + "blocks": [ + { + "bbox": [ + 238, + 118, + 310, + 190 + ], + "lines": [ + { + "bbox": [ + 238, + 118, + 310, + 190 + ], + "spans": [ + { + "bbox": [ + 238, + 118, + 310, + 190 + ], + "type": "image", + "image_path": "e9f7f381ffcd0d676f276b5cf23cdb942d4f2526e59ae15c03c38aff14e99528.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 266, + 196, + 277, + 205 + ], + "lines": [ + { + "bbox": [ + 266, + 196, + 277, + 205 + ], + "spans": [ + { + "bbox": [ + 266, + 196, + 277, + 205 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 213, + 478, + 224 + ], + "lines": [ + { + "bbox": [ + 133, + 213, + 478, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 213, + 478, + 224 + ], + "type": "text", + "content": "Fig. 3: These four subfigures contain all steps of our proposed interpolation method." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 313, + 116, + 386, + 189 + ], + "blocks": [ + { + "bbox": [ + 313, + 116, + 386, + 189 + ], + "lines": [ + { + "bbox": [ + 313, + 116, + 386, + 189 + ], + "spans": [ + { + "bbox": [ + 313, + 116, + 386, + 189 + ], + "type": "image", + "image_path": "905df1fe32afa36bb6b3a60891a104df9ba4790a79720c2356de9b9845d9fe85.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 196, + 356, + 204 + ], + "lines": [ + { + "bbox": [ + 345, + 196, + 356, + 204 + ], + "spans": [ + { + "bbox": [ + 345, + 196, + 356, + 204 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 389, + 116, + 462, + 188 + ], + "blocks": [ + { + "bbox": [ + 389, + 116, + 462, + 188 + ], + "lines": [ + { + "bbox": [ + 389, + 116, + 462, + 188 + ], + "spans": [ + { + "bbox": [ + 389, + 116, + 462, + 188 + ], + "type": "image", + "image_path": "6c28bcfb26141a4c535ef31a63ca03f62e0693a6cded3393a3622e3df3d46d40.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 194, + 431, + 202 + ], + "lines": [ + { + "bbox": [ + 421, + 194, + 431, + 202 + ], + "spans": [ + { + "bbox": [ + 421, + 194, + 431, + 202 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 246, + 381, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 246, + 381, + 258 + ], + "spans": [ + { + "bbox": [ + 131, + 246, + 381, + 258 + ], + "type": "text", + "content": "3.3 Geometry-aware Continuous Gaze Prompts" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 263, + 481, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 481, + 336 + ], + "type": "text", + "content": "In this section, to better represent gaze labels in the embedding space, we construct a gaze embedding for each label. This embedding should have the capability for continuous representation and geometric attributes. To address the conflict between the finite vocabulary in natural language processing and the infinite nature of gaze labels, we interpolate each gaze embedding " + }, + { + "bbox": [ + 130, + 263, + 481, + 336 + ], + "type": "inline_equation", + "content": "\\pmb{v}_i^g" + }, + { + "bbox": [ + 130, + 263, + 481, + 336 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 130, + 263, + 481, + 336 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 263, + 481, + 336 + ], + "type": "text", + "content": " gaze anchor embeddings " + }, + { + "bbox": [ + 130, + 263, + 481, + 336 + ], + "type": "inline_equation", + "content": "\\{\\pmb{A_j}\\}_{j = 1}^N" + }, + { + "bbox": [ + 130, + 263, + 481, + 336 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 336, + 482, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 336, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 131, + 336, + 482, + 409 + ], + "type": "text", + "content": "Common approaches, such as those described in [15], employ global linear interpolation to represent " + }, + { + "bbox": [ + 131, + 336, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\pmb{v}_i^g" + }, + { + "bbox": [ + 131, + 336, + 482, + 409 + ], + "type": "text", + "content": ", which interpolates the target embedding using all anchor embeddings. The weight " + }, + { + "bbox": [ + 131, + 336, + 482, + 409 + ], + "type": "inline_equation", + "content": "w_{i,j}" + }, + { + "bbox": [ + 131, + 336, + 482, + 409 + ], + "type": "text", + "content": " for each anchor embedding is computed by Equation 2, determined by the similarity between its corresponding gaze vector " + }, + { + "bbox": [ + 131, + 336, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\pmb{g}_j" + }, + { + "bbox": [ + 131, + 336, + 482, + 409 + ], + "type": "text", + "content": " and the target " + }, + { + "bbox": [ + 131, + 336, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\pmb{g}_i" + }, + { + "bbox": [ + 131, + 336, + 482, + 409 + ], + "type": "text", + "content": ". However, global interpolation cannot accurately represent " + }, + { + "bbox": [ + 131, + 336, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\pmb{v}_i^g" + }, + { + "bbox": [ + 131, + 336, + 482, + 409 + ], + "type": "text", + "content": " because anchors that are not proximate fail to provide effective information." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 251, + 415, + 481, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 415, + 481, + 447 + ], + "spans": [ + { + "bbox": [ + 251, + 415, + 481, + 447 + ], + "type": "interline_equation", + "content": "w _ {i, j} = \\frac {\\cos \\left(\\boldsymbol {g} _ {i} , \\boldsymbol {g} _ {j}\\right)}{\\sum_ {j = 1} ^ {N} \\cos \\left(\\boldsymbol {g} _ {i} , \\boldsymbol {g} _ {j}\\right)} \\tag {2}", + "image_path": "38f298a6f853d4922ab7d84c7ccd03849bb0f35efde2f9bfe02e5eccec76d2fd.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 449, + 481, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 481, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 481, + 496 + ], + "type": "text", + "content": "To further enhance interpolation, we attempted to employ a bilinear interpolation method. However, as it still utilizes a linear representation, this method remains insufficiently precise in expressing the geometric relationships between gaze vectors." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "text", + "content": "To address this issue, we decided to adopt a spherical interpolation method [26], which can accurately calculate the interpolation weights between vectors. As illustrated in Figure 3(a), a flat spherical interpolation method is depicted. For the vector " + }, + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{g}_i" + }, + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "text", + "content": " positioned between " + }, + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{g}_1" + }, + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{g}_2" + }, + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "text", + "content": ", we initially compute " + }, + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\theta = \\arccos(\\langle \\pmb{g}_1, \\pmb{g}_2 \\rangle)" + }, + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\theta_1 = \\arccos(\\langle \\pmb{g}_1, \\pmb{g}_i \\rangle)" + }, + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "text", + "content": ". Subsequently, we compute the scalar " + }, + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "t = \\theta_1 / \\theta" + }, + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "text", + "content": ", then we apply the spherical interpolation Equation 3 to determine the corresponding weights for " + }, + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{g}_1" + }, + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{g}_2" + }, + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "text", + "content": ". Finally, we can represent the target vector " + }, + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{g}_i" + }, + { + "bbox": [ + 131, + 498, + 482, + 605 + ], + "type": "text", + "content": " through the summation of these weights. Ultimately, we confirm that this method of weight calculation is accurate." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 233, + 613, + 481, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 613, + 481, + 639 + ], + "spans": [ + { + "bbox": [ + 233, + 613, + 481, + 639 + ], + "type": "interline_equation", + "content": "\\boldsymbol {g} _ {i} = \\frac {\\sin ((1 - t) \\theta)}{\\sin (\\theta)} \\boldsymbol {g} _ {1} + \\frac {\\sin (t \\theta)}{\\sin (\\theta)} \\boldsymbol {g} _ {2} \\tag {3}", + "image_path": "c0656e2009c27dae7b3fc11d565f59cbcb8f2efce12f2de36df1d127068d63fc.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Building upon the aforementioned spherical interpolation method, we further integrate the concept of bilinear interpolation to derive the interpolation" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 262, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 262, + 91, + 448, + 102 + ], + "type": "text", + "content": "LG-Gaze: Language-Guided Gaze Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "weights for each gaze label on a stereographic sphere. As depicted in Figure 3(b), we initially identify the four nearest anchor embeddings " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "A_{1-4}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " based on their proximity. Subsequently, as shown in Figure 3(c), we calculate the weights " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "w_{a,1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "w_{a,2}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "A_1" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "A_2" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " using Equation 3. Similarly, we can determine " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "w_{b,3}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "w_{b,4}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": ". In Figure 3(d), we proceed to obtain " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "w_{i,a}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "w_{i,b}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " using a similar method. Ultimately, we acquire the associated weights for " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "A_{1-4}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " to interpolate " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\pmb{v}_i^g" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": ", which can be expressed as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 210, + 481, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 210, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 210, + 481, + 224 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {i} ^ {g} = \\left(w _ {i, a} * w _ {a, 1}\\right) \\cdot \\boldsymbol {A} _ {1} + \\left(w _ {i, a} * w _ {a, 2}\\right) \\cdot \\boldsymbol {A} _ {2} + \\left(w _ {i, b} * w _ {b, 3}\\right) \\cdot \\boldsymbol {A} _ {3} + \\left(w _ {i, b} * w _ {b, 4}\\right) \\cdot \\boldsymbol {A} _ {4} \\tag {4}", + "image_path": "7cf5f275db586cf26dd4015c8b1d1fc6e0549c4fa243847181c63ccd1eea1a85.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 233, + 482, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 482, + 281 + ], + "type": "text", + "content": "Additionally, to enhance the geometric relationships between gaze embeddings, we employ a simple loss function to constrain anchor embeddings " + }, + { + "bbox": [ + 130, + 233, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\{A_j\\}_{j=1}^N" + }, + { + "bbox": [ + 130, + 233, + 482, + 281 + ], + "type": "text", + "content": ". As shown in Equation 5, the similarity between anchor embeddings should be consistent with their relationships in the label space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 194, + 291, + 481, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 291, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 194, + 291, + 481, + 324 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {G e o}} = \\frac {1}{M * M} \\sum_ {i = 1} ^ {M} \\sum_ {j = 1} ^ {M} L _ {1} \\left(\\cos \\left(\\boldsymbol {A} _ {i}, \\boldsymbol {A} _ {j}\\right), \\cos \\left(\\boldsymbol {g} _ {i}, \\boldsymbol {g} _ {j}\\right)\\right) \\tag {5}", + "image_path": "24abcdcf33e3aff84f5ebee04624f0f92b7167dac6321168aec59359cfcfae15.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 342, + 473, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 342, + 473, + 354 + ], + "spans": [ + { + "bbox": [ + 131, + 342, + 473, + 354 + ], + "type": "text", + "content": "3.4 Continuous Multimodal Contrastive Regression Loss Function" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 361, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 506 + ], + "type": "text", + "content": "Mathematically, from the aforementioned steps, we obtain image features " + }, + { + "bbox": [ + 130, + 361, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\{\\pmb{f}_i^g\\}_{i=1}^B" + }, + { + "bbox": [ + 130, + 361, + 482, + 506 + ], + "type": "text", + "content": " and their corresponding text features " + }, + { + "bbox": [ + 130, + 361, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\{\\pmb{f}_i^t\\}_{i=1}^B" + }, + { + "bbox": [ + 130, + 361, + 482, + 506 + ], + "type": "text", + "content": ". Both CLIP and OrdinalCLIP utilize the InfoNCE loss to align the multimodal features. To delve into specifics, consider an image-text feature pair " + }, + { + "bbox": [ + 130, + 361, + 482, + 506 + ], + "type": "inline_equation", + "content": "(\\pmb{f}_i^g, \\pmb{f}_i^t)" + }, + { + "bbox": [ + 130, + 361, + 482, + 506 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 361, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^g / \\pmb{f}_i^t" + }, + { + "bbox": [ + 130, + 361, + 482, + 506 + ], + "type": "text", + "content": " serves as the positive sample for " + }, + { + "bbox": [ + 130, + 361, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^t / \\pmb{f}_i^g" + }, + { + "bbox": [ + 130, + 361, + 482, + 506 + ], + "type": "text", + "content": ", as they are associated with the same gaze label. Meanwhile, all other image/text features in a mini-batch are treated as negative samples and are consequently pushed away from " + }, + { + "bbox": [ + 130, + 361, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^t / \\pmb{f}_i^g" + }, + { + "bbox": [ + 130, + 361, + 482, + 506 + ], + "type": "text", + "content": ". However, this training objective poses challenges for regression tasks because it overlooks the finer-grained distance ordering relationships between samples. Consequently, failing to differentiate the relative distances between image/text and their corresponding negative samples will inevitably weaken the learning effectiveness, especially in cross-modal representation learning for regression tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 506, + 482, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 613 + ], + "type": "text", + "content": "Considering the proximity relationships between data, for the " + }, + { + "bbox": [ + 130, + 506, + 482, + 613 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 506, + 482, + 613 + ], + "type": "text", + "content": "-th sample, we introduce a weight parameter " + }, + { + "bbox": [ + 130, + 506, + 482, + 613 + ], + "type": "inline_equation", + "content": "\\{w(i,j)\\}_{j=1}^{B}" + }, + { + "bbox": [ + 130, + 506, + 482, + 613 + ], + "type": "text", + "content": " for each negative sample. This parameter is computed based on the distances between labels of samples. Our proposed multimodal contrastive regression loss (MCR) effectively controls the feature distances among samples, preventing negative samples from being excessively contracted. As a result, we learn a semantically meaningful representation space tailored for regression tasks, ensuring consistency between sample feature relationships and the label space. The proposed text-to-image MCR loss is formulated as follows." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 144, + 622, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 622, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 144, + 622, + 481, + 666 + ], + "type": "interline_equation", + "content": "\\ell_ {t 2 i} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\left(\\frac {\\exp \\left(\\sin \\left(\\boldsymbol {f} _ {i} ^ {t} , \\boldsymbol {f} _ {i} ^ {g}\\right)\\right)}{\\exp \\left(\\sin \\left(\\boldsymbol {f} _ {i} ^ {t} , \\boldsymbol {f} _ {i} ^ {g}\\right)\\right) + \\sum_ {j = 1} ^ {B} w (i , j) \\cdot \\exp \\left(\\sin \\left(\\boldsymbol {f} _ {i} ^ {t} , \\boldsymbol {f} _ {j} ^ {g}\\right)\\right)}\\right) \\tag {6}", + "image_path": "9b116838e6a488ecdc33dc26b25030ced69c65b0c7cca0a024b28e240873251b.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "Y. Pengwei et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 199 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 115, + 479, + 199 + ], + "type": "inline_equation", + "content": "\\text{sim}(\\pmb{f}_i^t, \\pmb{f}_i^g) = \\cos(\\pmb{f}_i^t, \\pmb{f}_i^g)" + }, + { + "bbox": [ + 130, + 115, + 479, + 199 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 479, + 199 + ], + "type": "inline_equation", + "content": "w(i,j) = \\cos(\\pmb{g}_i, \\pmb{g}_j)" + }, + { + "bbox": [ + 130, + 115, + 479, + 199 + ], + "type": "text", + "content": " denotes the contrastive weight of the " + }, + { + "bbox": [ + 130, + 115, + 479, + 199 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 115, + 479, + 199 + ], + "type": "text", + "content": "-th negative image sample with respect to the " + }, + { + "bbox": [ + 130, + 115, + 479, + 199 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 115, + 479, + 199 + ], + "type": "text", + "content": "-th text feature in our LG-Gaze framework. The weight parameter should be directly proportional to the label distance between negative samples. As the distance between labels increases, the penalty on negative sample weights should also increase. This encourages feature embeddings to adhere to real-world relationships." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 480, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 390 + ], + "type": "text", + "content": "The MCR loss for image-to-text can be formulated in a similar manner. However, what sets it apart is our deliberate expansion of the number of negative samples to facilitate the learning of a robust representation. To achieve this, we draw inspiration from MoCo, which utilizes a dynamic queue to maintain a diverse set of negative samples, thereby enhancing the effectiveness of contrastive learning. Nevertheless, this approach still faces some challenges. For instance, if the queue update rate is too slow, the negative samples stored in the queue may become outdated and fail to align with the current model. Additionally, due to the inherent randomness and limited diversity of negative samples, there could be discrepancies between the queue data and the entire data distribution. To address this issue, we utilize the flexibility of language labels by creating " + }, + { + "bbox": [ + 130, + 200, + 480, + 390 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 200, + 480, + 390 + ], + "type": "text", + "content": " global text features " + }, + { + "bbox": [ + 130, + 200, + 480, + 390 + ], + "type": "inline_equation", + "content": "\\{f_i^t\\}_{i=1}^K" + }, + { + "bbox": [ + 130, + 200, + 480, + 390 + ], + "type": "text", + "content": " as negative samples (where " + }, + { + "bbox": [ + 130, + 200, + 480, + 390 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 200, + 480, + 390 + ], + "type": "text", + "content": " is significantly larger than " + }, + { + "bbox": [ + 130, + 200, + 480, + 390 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 200, + 480, + 390 + ], + "type": "text", + "content": "). These text labels correspond to gaze vectors that show a uniformly dense distribution in the label space. Furthermore, these " + }, + { + "bbox": [ + 130, + 200, + 480, + 390 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 200, + 480, + 390 + ], + "type": "text", + "content": " text labels are interpolated based on " + }, + { + "bbox": [ + 130, + 200, + 480, + 390 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 200, + 480, + 390 + ], + "type": "text", + "content": " anchor " + }, + { + "bbox": [ + 130, + 200, + 480, + 390 + ], + "type": "inline_equation", + "content": "\\{A_j\\}_{j=1}^N" + }, + { + "bbox": [ + 130, + 200, + 480, + 390 + ], + "type": "text", + "content": ". As a result, the proposed image-to-text MCR loss can be formulated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 399, + 480, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 399, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 140, + 399, + 480, + 443 + ], + "type": "interline_equation", + "content": "\\ell_ {i 2 t} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\left(\\frac {\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {f} _ {i} ^ {g} , \\boldsymbol {f} _ {i} ^ {t}\\right)\\right)}{\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {f} _ {i} ^ {g} , \\boldsymbol {f} _ {i} ^ {t}\\right)\\right) + \\sum_ {j = 1} ^ {B + K} w (i , j) \\cdot \\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {f} _ {i} ^ {g} , \\boldsymbol {f} _ {j} ^ {t}\\right)\\right)}\\right) \\tag {7}", + "image_path": "6d5e10ec97b13a2ca66b4da3ecc7e0c69c694ec1ea3317bbc42d2116db76fcec.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 444, + 480, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 480, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 480, + 550 + ], + "type": "text", + "content": "Obviously, we can observe that the difference between " + }, + { + "bbox": [ + 130, + 444, + 480, + 550 + ], + "type": "inline_equation", + "content": "\\ell_{i2t}" + }, + { + "bbox": [ + 130, + 444, + 480, + 550 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 444, + 480, + 550 + ], + "type": "inline_equation", + "content": "\\ell_{t2i}" + }, + { + "bbox": [ + 130, + 444, + 480, + 550 + ], + "type": "text", + "content": " lies in the number of negative samples. We concatenate " + }, + { + "bbox": [ + 130, + 444, + 480, + 550 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 444, + 480, + 550 + ], + "type": "text", + "content": " global negative samples with the " + }, + { + "bbox": [ + 130, + 444, + 480, + 550 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 444, + 480, + 550 + ], + "type": "text", + "content": " negative samples from mini-batches to form the complete set of negative samples. Through this operation, the global negative samples remain relatively fixed, mitigating issues related to randomness. Additionally, their uniform distribution helps alleviate discrepancies between the negative samples and the entire data distribution. Furthermore, by computing features based on the current model, we address potential outdated information. Finally, our total MCR loss can be denoted as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 267, + 553, + 480, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 553, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 267, + 553, + 480, + 566 + ], + "type": "interline_equation", + "content": "\\ell_ {M C R} = \\ell_ {i 2 t} + \\ell_ {t 2 i} \\tag {8}", + "image_path": "340fb0f7ff7b1e7a2f1010675a7a941d9be073ea0ff3060e9a1389944e957b9f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 584, + 249, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 584, + 249, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 584, + 249, + 597 + ], + "type": "text", + "content": "3.5 Overall Objective" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 607, + 430, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 607, + 430, + 619 + ], + "spans": [ + { + "bbox": [ + 131, + 607, + 430, + 619 + ], + "type": "text", + "content": "In summary, the overall objective implemented in our framework is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 236, + 631, + 378, + 642 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 631, + 378, + 642 + ], + "spans": [ + { + "bbox": [ + 236, + 631, + 378, + 642 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {1} \\ell_ {G e o} + \\lambda_ {2} \\ell_ {M C R} + \\lambda_ {3} \\ell_ {G a z e}", + "image_path": "921d967de74e2932581f3d47d2324d660ed0640cb8c62cac6957e89e97304869.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 146, + 653, + 475, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 653, + 475, + 666 + ], + "spans": [ + { + "bbox": [ + 146, + 653, + 475, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2, \\lambda_3" + }, + { + "bbox": [ + 146, + 653, + 475, + 666 + ], + "type": "text", + "content": " are hyperparameters, and we empirically set " + }, + { + "bbox": [ + 146, + 653, + 475, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_1 = \\lambda_2 = \\lambda_3 = 1.0" + }, + { + "bbox": [ + 146, + 653, + 475, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 262, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 262, + 91, + 447, + 102 + ], + "type": "text", + "content": "LG-Gaze: Language-Guided Gaze Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 259, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 259, + 153 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 259, + 153 + ], + "type": "text", + "content": "4.1 Experiment Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 162, + 481, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 221 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 221 + ], + "type": "text", + "content": "Data Preparation We evaluate gaze estimation method on four cross-domain tasks, utilizing ETH-XGaze [37] and Gaze360 [12] as the training datasets, and MPIIFaceGaze [39] and Eye-Diap [9] as the test datasets. We denote them as " + }, + { + "bbox": [ + 132, + 162, + 481, + 221 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{E}}" + }, + { + "bbox": [ + 132, + 162, + 481, + 221 + ], + "type": "text", + "content": " (ETH-XGaze) " + }, + { + "bbox": [ + 132, + 162, + 481, + 221 + ], + "type": "inline_equation", + "content": "\\rightarrow \\mathcal{D}_{\\mathrm{M}}" + }, + { + "bbox": [ + 132, + 162, + 481, + 221 + ], + "type": "text", + "content": " (MPIIFaceGaze), " + }, + { + "bbox": [ + 132, + 162, + 481, + 221 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{E}} \\rightarrow \\mathcal{D}_{\\mathrm{D}}" + }, + { + "bbox": [ + 132, + 162, + 481, + 221 + ], + "type": "text", + "content": " (EyeDiap), " + }, + { + "bbox": [ + 132, + 162, + 481, + 221 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{G}}" + }, + { + "bbox": [ + 132, + 162, + 481, + 221 + ], + "type": "text", + "content": " (Gaze360) " + }, + { + "bbox": [ + 132, + 162, + 481, + 221 + ], + "type": "inline_equation", + "content": "\\rightarrow \\mathcal{D}_{\\mathrm{M}}" + }, + { + "bbox": [ + 132, + 162, + 481, + 221 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 162, + 481, + 221 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{G}} \\rightarrow \\mathcal{D}_{\\mathrm{D}}" + }, + { + "bbox": [ + 132, + 162, + 481, + 221 + ], + "type": "text", + "content": ". See the supplementary material for more details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 240, + 481, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 301 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 301 + ], + "type": "text", + "content": "Vision Model Implementation Details We use ResNet-18 as our gaze feature extractor " + }, + { + "bbox": [ + 132, + 240, + 481, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_v" + }, + { + "bbox": [ + 132, + 240, + 481, + 301 + ], + "type": "text", + "content": ", and a fully connected layer to regress a 3-dimensional gaze vector. We resize and normalize all the images to " + }, + { + "bbox": [ + 132, + 240, + 481, + 301 + ], + "type": "inline_equation", + "content": "224\\times 224" + }, + { + "bbox": [ + 132, + 240, + 481, + 301 + ], + "type": "text", + "content": " pixels and scale the pixel values between 0 and 1. We set the batch size to 64. We maintain the same setup for training the source domain models on " + }, + { + "bbox": [ + 132, + 240, + 481, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_E" + }, + { + "bbox": [ + 132, + 240, + 481, + 301 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 240, + 481, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_G" + }, + { + "bbox": [ + 132, + 240, + 481, + 301 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 319, + 481, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 440 + ], + "type": "text", + "content": "Language Model Implementation Details The architecture of the text encoder is adapted from the Transformer [27], incorporating modifications outlined by Radford [23]. The dimensions for both text and gaze features have been standardized to 1024. Our models are constructed using the foundational open-source code of CLIP. The class token is strategically placed at the terminal position of the sequence, with the context tokens quantity designated as " + }, + { + "bbox": [ + 132, + 319, + 481, + 440 + ], + "type": "inline_equation", + "content": "L = 10" + }, + { + "bbox": [ + 132, + 319, + 481, + 440 + ], + "type": "text", + "content": ". Additionally, we set the yaw angle range for the anchors to be " + }, + { + "bbox": [ + 132, + 319, + 481, + 440 + ], + "type": "inline_equation", + "content": "[-180^{\\circ}, +179^{\\circ}]" + }, + { + "bbox": [ + 132, + 319, + 481, + 440 + ], + "type": "text", + "content": " and the pitch range to be " + }, + { + "bbox": [ + 132, + 319, + 481, + 440 + ], + "type": "inline_equation", + "content": "[-90^{\\circ}, +90^{\\circ}]" + }, + { + "bbox": [ + 132, + 319, + 481, + 440 + ], + "type": "text", + "content": " with a step size of " + }, + { + "bbox": [ + 132, + 319, + 481, + 440 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 132, + 319, + 481, + 440 + ], + "type": "text", + "content": ", resulting in a total of 91 anchors. We set the number of negative samples to 256, which are uniformly distributed in the gaze label space using the Fibonacci sphere algorithm [10]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 457, + 481, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 530 + ], + "type": "text", + "content": "Training Implementation Details We conducted the experiments on a single Tesla V100 GPU. Specifically, we use the SGD optimizer with Nesterov momentum, a learning rate (LR) of " + }, + { + "bbox": [ + 132, + 457, + 481, + 530 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-2}" + }, + { + "bbox": [ + 132, + 457, + 481, + 530 + ], + "type": "text", + "content": " and a weight decay of " + }, + { + "bbox": [ + 132, + 457, + 481, + 530 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 132, + 457, + 481, + 530 + ], + "type": "text", + "content": " for the parameters. We train for 30 epochs using a Cosineannealing LR scheduler [18] with a 3-epoch warm-up. We apply a data augmentation technique involving a random color field and grayscale, as described in [5]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 548, + 481, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 481, + 561 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 481, + 561 + ], + "type": "text", + "content": "4.2 Performance Comparison with SOTA Gaze Estimation Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 569, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 481, + 666 + ], + "type": "text", + "content": "Comparison of Cross-domain Evaluation. Table 1 presents the quantitative results of four cross-domain tests. The second row compares our LG-Gaze with the SOTA domain DG methods for gaze estimation. The CNN baseline method refers to the one that relies solely on the vision encoder to extract the single visual modality feature and uses " + }, + { + "bbox": [ + 132, + 569, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\ell_{gaze}" + }, + { + "bbox": [ + 132, + 569, + 481, + 666 + ], + "type": "text", + "content": " for supervision. In summary, the LG-Gaze achieves the best overall performance and demonstrates state-of-the-art results on three cross-domain evaluation tasks. It also achieves the second-best performance for " + }, + { + "bbox": [ + 132, + 569, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{E}} \\rightarrow \\mathcal{D}_{\\mathrm{M}}" + }, + { + "bbox": [ + 132, + 569, + 481, + 666 + ], + "type": "text", + "content": ", highlighting the effectiveness of our method." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "Y. Pengwei et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 114, + 477, + 289 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 477, + 289 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 477, + 289 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 477, + 289 + ], + "type": "table", + "html": "
TaskMethods| Dt|DE→DMDE→DdDG→DMDG→DdAvg
DGCNN Baseline08.479.327.548.938.57
PureGaze [6]07.087.489.289.328.29
CDG [30]‡06.737.957.037.277.25
Xu et al. [31]06.507.447.559.037.63
CLIP-Gaze [34]06.417.516.897.066.97
Our LG-Gaze06.457.226.836.866.84
UDAPnP-GA * [17]105.535.876.187.926.38
RUDA [2]1005.706.296.205.866.01
CRGA [30]>05.485.665.896.495.88
LatentGaze [14]1005.217.81--6.51
Liu et al. [16]1005.356.627.188.616.94
UnReGA [3]1005.115.705.425.805.51
", + "image_path": "88e9b38758c892e5fc8df110bd261b1dd1525e2e3d94db24a1e67580b18b1acf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 289, + 479, + 332 + ], + "lines": [ + { + "bbox": [ + 132, + 289, + 479, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 479, + 332 + ], + "type": "text", + "content": "Table 1: Comparison with state-of-the-art methods. We report the results by angular error in degrees, and use bold and underline to indicate the best and the second best result in each column for a specific task. " + }, + { + "bbox": [ + 132, + 289, + 479, + 332 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 132, + 289, + 479, + 332 + ], + "type": "text", + "content": " means the model uses ResNet-50 as backbone, * shows that the experimental settings are different." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 169, + 357, + 233, + 421 + ], + "blocks": [ + { + "bbox": [ + 169, + 357, + 233, + 421 + ], + "lines": [ + { + "bbox": [ + 169, + 357, + 233, + 421 + ], + "spans": [ + { + "bbox": [ + 169, + 357, + 233, + 421 + ], + "type": "image", + "image_path": "dfde9f82953867fdeba0ccd0616215f08bb787a8a8d4f6480a6bd00e1f0f9d64.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 422, + 227, + 429 + ], + "lines": [ + { + "bbox": [ + 179, + 422, + 227, + 429 + ], + "spans": [ + { + "bbox": [ + 179, + 422, + 227, + 429 + ], + "type": "text", + "content": "(a) CNN Baseline" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 240, + 357, + 304, + 421 + ], + "blocks": [ + { + "bbox": [ + 240, + 357, + 304, + 421 + ], + "lines": [ + { + "bbox": [ + 240, + 357, + 304, + 421 + ], + "spans": [ + { + "bbox": [ + 240, + 357, + 304, + 421 + ], + "type": "image", + "image_path": "6902da96e17fd5701284d8e5e5c233712fa637d786bb190822914e6fadf1f03f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 422, + 290, + 429 + ], + "lines": [ + { + "bbox": [ + 255, + 422, + 290, + 429 + ], + "spans": [ + { + "bbox": [ + 255, + 422, + 290, + 429 + ], + "type": "text", + "content": "(b) PureGaze" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 439, + 479, + 462 + ], + "lines": [ + { + "bbox": [ + 132, + 439, + 479, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 479, + 462 + ], + "type": "text", + "content": "Fig. 4: Illustration of the feature distribution. Various colors indicate different gaze directions and similar gaze directions have similar colors. (Recommend viewing in color)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 311, + 357, + 374, + 420 + ], + "blocks": [ + { + "bbox": [ + 311, + 357, + 374, + 420 + ], + "lines": [ + { + "bbox": [ + 311, + 357, + 374, + 420 + ], + "spans": [ + { + "bbox": [ + 311, + 357, + 374, + 420 + ], + "type": "image", + "image_path": "4e04de218db6f8d6f6a3ed8d34d44f77d9a314296542b33d1f9fb892e692f6ed.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 422, + 354, + 429 + ], + "lines": [ + { + "bbox": [ + 331, + 422, + 354, + 429 + ], + "spans": [ + { + "bbox": [ + 331, + 422, + 354, + 429 + ], + "type": "text", + "content": "(c) CDG" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 380, + 357, + 443, + 419 + ], + "blocks": [ + { + "bbox": [ + 380, + 357, + 443, + 419 + ], + "lines": [ + { + "bbox": [ + 380, + 357, + 443, + 419 + ], + "spans": [ + { + "bbox": [ + 380, + 357, + 443, + 419 + ], + "type": "image", + "image_path": "783a01144f579992f4f82be927a111de510155cb42ed23d131d2c6bb09634d92.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 422, + 422, + 429 + ], + "lines": [ + { + "bbox": [ + 400, + 422, + 422, + 429 + ], + "spans": [ + { + "bbox": [ + 400, + 422, + 422, + 429 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 487, + 480, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 480, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 480, + 582 + ], + "type": "text", + "content": "In addition, we present a comparison with SOTA unsupervised domain adaptation (UDA) methods in the third row of Table 1. Note that UDA methods require the use of a small batch of unlabeled target domain samples, which is more costly than DG methods. It can be clearly seen from the table that our LG-Gaze achieves comparable performance without extra target domain data. Especially on the two tasks of " + }, + { + "bbox": [ + 130, + 487, + 480, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{G}} \\to \\mathcal{D}_{\\mathrm{M}}" + }, + { + "bbox": [ + 130, + 487, + 480, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 487, + 480, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{G}} \\to \\mathcal{D}_{\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 487, + 480, + 582 + ], + "type": "text", + "content": ", it is close to UDA methods. Therefore, this verifies the powerful performance of our proposed method, as it does not rely on any prior information from the target domain." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Comparison of Feature Visualization. To evaluate and analyze the gaze features " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "f^g" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " extracted by different models, we follow the same scheme as CDG [30], and visualize the gaze features of the cross-domain task " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{G}} \\to \\mathcal{D}_{\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " using t-SNE [19] to compare different methods. Figure 4 shows the visualization results from all DG methods in the second row of Table 1, where feature points with" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 263, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 263, + 91, + 447, + 102 + ], + "type": "text", + "content": "LG-Gaze: Language-Guided Gaze Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 114, + 481, + 198 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 481, + 198 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 481, + 198 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 481, + 198 + ], + "type": "table", + "html": "
MethodsDE→DMDE→DdDG→DMDG→DdAvg
CNN Baseline8.359.667.589.018.65
Vanilla CLIP [22]82.4379.6482.4379.6481.04
CoOp [41]9.2311.0613.1510.8511.07
OridinalCLIP [15]7.8810.119.129.749.21
Our LG-Gaze6.457.226.836.866.84
", + "image_path": "1d1340db530c76f92acad24e14d16967ab73804af8e50a775ca080aa14ee4742.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 132, + 198, + 481, + 219 + ], + "lines": [ + { + "bbox": [ + 132, + 198, + 481, + 219 + ], + "spans": [ + { + "bbox": [ + 132, + 198, + 481, + 219 + ], + "type": "text", + "content": "Table 2: Comparisons to SOTA contrastive learning methods. Bold indicates the best results in each column, and underline denote the second best result in each column." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 255, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 365 + ], + "type": "text", + "content": "similar gaze labels have similar colors. For the baseline model, the features of different gaze labels are mixed together, which is unreasonable for the regression task as shown in Figure 4 (a). The ideal feature distribution should be continuous and gradual with labels for the regression task [36]. Figure 4 (b,c) represents several classic DG methods that have been enhanced compared to the baseline, but they exhibit unreasonable feature distributions. In general, LG-Gaze has the most reasonable feature distribution, and the visualization result is shown in Figure 4 (d). The strong correlation between gaze direction and features indicates that guiding visual feature learning through language is effective." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 389, + 481, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 389, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 389, + 481, + 402 + ], + "type": "text", + "content": "4.3 Performance Comparison with SOTA VLM Regression Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 416, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 482, + 489 + ], + "type": "text", + "content": "Table 2 presents the performance of various methods that employ VLM models. To investigate the applicability of ordinal regression methods to gaze estimation, we discretize the gaze labels into integers. This mapping allows the yaw and pitch in the gaze label to be categorized into a finite number of categories. During testing, we predict yaw and pitch separately, but the accuracy of the gaze vector, which is composed of yaw and pitch, is evaluated." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 143, + 511, + 474, + 609 + ], + "blocks": [ + { + "bbox": [ + 143, + 511, + 474, + 609 + ], + "lines": [ + { + "bbox": [ + 143, + 511, + 474, + 609 + ], + "spans": [ + { + "bbox": [ + 143, + 511, + 474, + 609 + ], + "type": "table", + "html": "
MethodsDE→DMDE→DdDG→DMDG→DdAvg
CNN Baseline8.359.667.589.018.65
KL-Loss6.777.366.977.087.05
CDG [30]6.987.457.147.187.19
RNC [35]6.637.286.996.966.97
MoCo [11]6.527.336.957.046.96
MCR6.457.226.836.866.84
", + "image_path": "b5197bdd03a284cf665c291e27132f65a8f5ad9f71f232f9a0ced0ff056a5475.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 130, + 609, + 481, + 652 + ], + "lines": [ + { + "bbox": [ + 130, + 609, + 481, + 652 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 481, + 652 + ], + "type": "text", + "content": "Table 3: Comparison of different text prompt tuning methods for gaze model cross-domain evaluation in four tasks. The rest of the settings are consistent for all methods, except for the contrastive loss function. Bold indicates the best results in each column, and underline denote the second best result results in each column." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "Y. Pengwei et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 482, + 295 + ], + "type": "text", + "content": "For Vanilla CLIP, there is no need to use source data for model training. Instead, a zero-shot method is used for prediction, resulting in only two outcomes. First, we construct the text prompt \"gaze estimation: the yaw/pitch degree angle of the person is {yaw/pitch}.\" Then, we obtain the model's prediction of yaw and pitch by maximizing the alignment between text features and image features, which is essentially in line with the original image-text matching. However, we can observe from Table 2 that Vanilla CLIP exhibits poor performance, possibly due to its struggle in distinguishing ordinal concepts. For CoOp, the gaze prediction performance can be greatly enhanced through text prompt tuning, but it is weaker than the CNN baseline. This could be due to its inability to learn the ordinal relationship between samples. For OrdinalCLIP, although it can learn the ordinal relationship, the discretization of labels may lead to errors. Additionally, following the methodology of OrdinalCLIP, we have created a visualization of the semantic labels constructed by different VLM methods in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 312, + 469, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 469, + 335 + ], + "type": "text", + "content": "4.4 Performance Comparison with SOTA Contrastive Regression Loss Functions" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 343, + 482, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 482, + 440 + ], + "type": "text", + "content": "Table 3 presents various methods for feature alignment or contrastive learning in gaze estimation. For KL-Loss, it primarily constrains the similarity matrix between gaze features and text features, as well as the similarity matrix composed of gaze labels. We follow the ideas of CDG [30], RNC [35], and MoCo [11], and transform them into image-to-text and text-to-image contrastive loss functions similar to LG-Gaze. From Table 3, we can see that our MCR loss function outperforms other methods significantly, attributed to the inclusion of constructed global negative samples and the weight coefficient for negative samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 456, + 238, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 238, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 238, + 468 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 475, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 482, + 548 + ], + "type": "text", + "content": "Effectiveness of Our Framework. Table 4 shows the performance of different combinations of loss functions. Keeping only " + }, + { + "bbox": [ + 132, + 475, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\ell_{Gaze}" + }, + { + "bbox": [ + 132, + 475, + 482, + 548 + ], + "type": "text", + "content": " is equivalent to CNN baseline, and combining all the losses is equivalent to full LG-Gaze. Compared with CNN baseline, using MCR loss to learn robust features with language guidance can significantly improve the performance. Further adding " + }, + { + "bbox": [ + 132, + 475, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\ell_{Geo}" + }, + { + "bbox": [ + 132, + 475, + 482, + 548 + ], + "type": "text", + "content": " enhances the geometric relevance between prompts, and thus improves the model performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 564, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 564, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 564, + 482, + 612 + ], + "type": "text", + "content": "Effectiveness of Interpolation Methods. To compare the performance of global linear interpolation, bilinear interpolation and our proposed geometry-aware spherical interpolation. Our method demonstrates superior performance due to its consideration of geometric accuracy, as shown in Table 5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 628, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 628, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 628, + 482, + 665 + ], + "type": "text", + "content": "Effectiveness of MCR Loss. To delve into the efficacy of our MCR loss, we setted varying numbers of negative samples to compare their cross-domain performance, as illustrated in Table 6. When the number of negative samples is zero," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 263, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 263, + 91, + 447, + 102 + ], + "type": "text", + "content": "LG-Gaze: Language-Guided Gaze Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 145, + 114, + 470, + 169 + ], + "blocks": [ + { + "bbox": [ + 145, + 114, + 470, + 169 + ], + "lines": [ + { + "bbox": [ + 145, + 114, + 470, + 169 + ], + "spans": [ + { + "bbox": [ + 145, + 114, + 470, + 169 + ], + "type": "table", + "html": "
lGeolMCRlGazeD_E → D_MD_E → D_DD_G → D_MD_G → D_DAvg
××√8.359.667.589.018.65
×√√6.767.356.936.987.01
√√√6.457.226.836.866.84
", + "image_path": "2e470c842f5aa3d4d4bd90fda8ac9a70b9f28d5fb7dfd57c4ed0d16f9edeefe4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 144, + 201, + 473, + 254 + ], + "blocks": [ + { + "bbox": [ + 146, + 169, + 465, + 179 + ], + "lines": [ + { + "bbox": [ + 146, + 169, + 465, + 179 + ], + "spans": [ + { + "bbox": [ + 146, + 169, + 465, + 179 + ], + "type": "text", + "content": "Table 4: Ablation study on loss functions. Results are reported by in degrees." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 201, + 473, + 254 + ], + "lines": [ + { + "bbox": [ + 144, + 201, + 473, + 254 + ], + "spans": [ + { + "bbox": [ + 144, + 201, + 473, + 254 + ], + "type": "table", + "html": "
MethodsDE→DMDE→DdDG→DMDG→DdAvg
Global Li-Inter7.197.547.297.337.34
Bilinear-Inter6.787.316.926.96.98
Ours6.457.226.836.866.84
", + "image_path": "ca58fdd25fefa340af5a2b7cce2d15cdf204be81bdd2070104db80c395ee255d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 153, + 285, + 465, + 371 + ], + "blocks": [ + { + "bbox": [ + 180, + 254, + 432, + 264 + ], + "lines": [ + { + "bbox": [ + 180, + 254, + 432, + 264 + ], + "spans": [ + { + "bbox": [ + 180, + 254, + 432, + 264 + ], + "type": "text", + "content": "Table 5: The performance of different interpolation methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 285, + 465, + 371 + ], + "lines": [ + { + "bbox": [ + 153, + 285, + 465, + 371 + ], + "spans": [ + { + "bbox": [ + 153, + 285, + 465, + 371 + ], + "type": "table", + "html": "
Number\\( \\mathcal{D}_{\\mathrm{E}}\\rightarrow \\mathcal{D}_{\\mathrm{M}} \\)\\( \\mathcal{D}_{\\mathrm{E}}\\rightarrow \\mathcal{D}_{\\mathrm{D}} \\)\\( \\mathcal{D}_{\\mathrm{G}}\\rightarrow \\mathcal{D}_{\\mathrm{M}} \\)\\( \\mathcal{D}_{\\mathrm{G}}\\rightarrow \\mathcal{D}_{\\mathrm{D}} \\)Avg
07.007.447.137.267.21
646.877.317.087.027.07
1286.667.216.987.056.98
2566.457.226.836.866.84
5126.427.196.856.816.82
10246.427.186.836.796.81
", + "image_path": "becc11b0132f690d5549212001e67ea6fdf2730f6e25272bc0b0bf85e68863b2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 167, + 372, + 445, + 382 + ], + "lines": [ + { + "bbox": [ + 167, + 372, + 445, + 382 + ], + "spans": [ + { + "bbox": [ + 167, + 372, + 445, + 382 + ], + "type": "text", + "content": "Table 6: The Comparison for different number of negative samples." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 416, + 480, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 480, + 488 + ], + "type": "text", + "content": "it equates to a common contrastive regression loss, which is observed to achieve less than remarkable performance. As the number of negative samples increases from 64 to 256, there is a marked improvement in performance, indicating that a higher quantity of negative samples in contrastive learning contributes to performance gains. However, when the number of negative samples is further increased to 1024, the performance nearly reaches convergence, with minimal gains." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 488, + 480, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 480, + 512 + ], + "type": "text", + "content": "Moreover, we conducted an experiment on loss weights to explore the effectiveness of our method. For more details, please see the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 531, + 220, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 220, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 220, + 544 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "text", + "content": "In this paper, we propose a novel training framework named LG-Gaze for gaze estimation. Leveraging textual features extracted from language models, LG-Gaze guides the learning of gaze features. This approach enables the gaze model to acquire robust gaze representations and achieve a reasonable feature distribution, ultimately enhancing the cross-domain generalization capability of gaze model. Our proposed LG-Gaze achieves state-of-the-art performance on domain generalization for gaze estimation task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "Acknowledgements. This work was sponsored by National Key R&D Program of China (2023YFE0204200)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "Y. Pengwei et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "text", + "content": "1. Andrist, S., Tan, X.Z., Gleicher, M., Mutlu, B.: Conversational gaze aversion for humanlike robots. In: 2014 9th ACM/IEEE International Conference on Human-Robot Interaction (HRI). pp. 25-32 (2014) 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "text", + "content": "2. Bao, Y., Liu, Y., Wang, H., Lu, F.: Generalizing gaze estimation with rotation consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4207-4216 (2022) 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 206, + 480, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 238 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 238 + ], + "type": "text", + "content": "3. Cai, X., Zeng, J., Shan, S., Chen, X.: Source-free adaptive gaze estimation by uncertainty reduction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22035-22045 (2023) 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 239, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 270 + ], + "type": "text", + "content": "4. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "text", + "content": "5. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020) 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 293, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 326 + ], + "type": "text", + "content": "6. Cheng, Y., Bao, Y.: Puregaze: Purifying gaze feature for generalizable gaze estimation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 436-443 (2022) 2, 4, 5, 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "type": "text", + "content": "7. Cheng, Y., Lu, F., Zhang, X.: Appearance-based gaze estimation via evaluation-guided asymmetric regression. In: Proceedings of the European Conference on Computer Vision (ECCV) (September 2018) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 359, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 359, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 359, + 480, + 380 + ], + "type": "text", + "content": "8. Cheng, Y., Zhang, X., Lu, F., Sato, Y.: Gaze estimation by exploring two-eye asymmetry. IEEE Transactions on Image Processing 29, 5259-5272 (2020) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 381, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 480, + 424 + ], + "type": "text", + "content": "9. Funes Mora, K.A., Monay, F., Odobez, J.M.: Eyediap: A database for the development and evaluation of gaze estimation algorithms from rgb and rgb-d cameras. In: Proceedings of the symposium on eye tracking research and applications. pp. 255-258 (2014) 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 425, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 446 + ], + "type": "text", + "content": "0. González, Á.: Measurement of areas on a sphere using fibonacci and latitude-longitude lattices. Mathematical Geosciences 42, 49-64 (2010) 10" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "text", + "content": "1. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) 5, 12, 13" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 479, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 512 + ], + "type": "text", + "content": "2. Kellnhofer, P., Recasens, A., Stent, S., Matusik, W., Torralba, A.: Gaze360: Physically unconstrained gaze estimation in the wild. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6912-6921 (2019) 3, 10" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "text", + "content": "3. Krafka, K., Khosla, A., Kellnhofer, P., Kannan, H., Bhandarkar, S.M., Matusik, W., Torralba, A.: Eye tracking for everyone. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 2176-2184 (2016) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "type": "text", + "content": "4. Lee, I., Yun, J.S., Kim, H.H., Na, Y., Yoo, S.B.: Latentgaze: Cross-domain gaze estimation through gaze-aware analytic latent code manipulation. In: Proceedings of the Asian Conference on Computer Vision. pp. 3379-3395 (2022) 11" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 610 + ], + "type": "text", + "content": "5. Li, W., Huang, X., Zhu, Z., Tang, Y., Li, X., Zhou, J., Lu, J.: Ordinalclip: Learning rank prompts for language-guided ordinal regression. Advances in Neural Information Processing Systems 35, 35313-35325 (2022) 2, 4, 7, 12" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "text", + "content": "6. Liu, R., Bao, Y., Xu, M., Wang, H., Liu, Y., Lu, F.: Jitter does matter: Adapting gaze estimation to new domains. arXiv preprint arXiv:2210.02082 (2022) 11" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "7. Liu, Y., Liu, R., Wang, H., Lu, F.: Generalizing gaze estimation with outlier-guided collaborative adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2021) 11" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 263, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 263, + 91, + 447, + 102 + ], + "type": "text", + "content": "LG-Gaze: Language-Guided Gaze Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "18. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "19. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 203 + ], + "type": "text", + "content": "20. Mavely, A.G., Judith, J.E., Sahal, P.A., Kuruvilla, S.A.: Eye gaze tracking based driver monitoring system. In: 2017 IEEE International Conference on Circuits and Systems (ICCS). pp. 364-367 (2017). https://doi.org/10.1109/ICCS1.2017.8326022 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 258 + ], + "type": "text", + "content": "21. Padmanaban, N., Konrad, R., Cooper, E.A., Wetzstein, G.: Optimizing vr for all users through adaptive focus displays. In: ACM SIGGRAPH 2017 Talks. SIGGRAPH '17, Association for Computing Machinery, New York, NY, USA (2017). https://doi.org/10.1145/3084363.3085029, https://doi.org/10.1145/3084363.3085029 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 303 + ], + "type": "text", + "content": "22. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 2, 4, 12" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "23. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog p. 9 (2019) 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 369 + ], + "type": "text", + "content": "24. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Denseclip: Language-guided dense prediction with context-aware prompting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18082-18091 (2022) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "25. Shen, S., Li, W., Wang, X., Zhang, D., Jin, Z., Zhou, J., Lu, J.: Clip-cluster: Clip-guided attribute hallucination for face clustering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 20786-20795 (2023) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "type": "text", + "content": "26. Shoemake, K.: Animating rotation with quaternion curves. In: Proceedings of the 12th annual conference on Computer graphics and interactive techniques. pp. 245-254 (1985) 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "27. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems (2017) 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "text", + "content": "28. Vidit, V., Engilberge, M., Salzmann, M.: Clip the gap: A single domain generalization approach for object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3219-3229 (2023) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "text", + "content": "29. Wang, R., Li, P., Huang, H., Cao, C., He, R., He, Z.: Learning-to-rank meets language: Boosting language-driven ordering alignment for ordinal classification. Advances in Neural Information Processing Systems 36 (2024) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 578 + ], + "type": "text", + "content": "30. Wang, Y., Jiang, Y., Li, J., Ni, B., Dai, W., Li, C., Xiong, H., Li, T.: Contrastive regression for domain adaptation on gaze estimation. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19354-19363 (2022). https://doi.org/10.1109/CVPR52688.2022.01877 2, 4, 5, 11, 12, 13" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "text", + "content": "31. Xu, M., Wang, H., Lu, F.: Learning a generalized gaze estimator from gaze-consistent feature. Proceedings of the AAAI Conference on Artificial Intelligence 37(3), 3027-3035 (Jun 2023) 2, 4, 5, 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 611, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 665 + ], + "type": "text", + "content": "32. Yao, L., Han, J., Wen, Y., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, C., Xu, H.: Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 9125-9138. Curran Associates, Inc. (2022), https://proceedings.neurips.cc/paper_files/" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 102 + ], + "type": "text", + "content": "Y. Pengwei et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 423 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 149, + 117, + 480, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 117, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 149, + 117, + 480, + 137 + ], + "type": "text", + "content": "paper/2022/file/3ba960559212691be13fa81d9e5e0047-Paper-Conference.pdf 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "text", + "content": "33. Yin, P., Wang, J., Dai, J., Wu, X.: Nerf-gaze: A head-eye redirection parametric model for gaze estimation. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2024) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "34. Yin, P., Zeng, G., Wang, J., Xie, D.: Clip-gaze: Towards general gaze estimation via visual-linguistic model. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 6729–6737 (2024) 2, 4, 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "35. Zha, K., Cao, P., Son, J., Yang, Y., Katabi, D.: Rank-n-contrast: Learning continuous representations for regression. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) 4, 12, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 270 + ], + "type": "text", + "content": "36. Zha, K., Cao, P., Son, J., Yang, Y., Katabi, D.: Rank-n-contrast: Learning continuous representations for regression. Advances in Neural Information Processing Systems 36 (2024) 12" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "type": "text", + "content": "37. Zhang, X., Park, S., Beeler, T., Bradley, D., Tang, S., Hilliges, O.: Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze variation. In: European Conference on Computer Vision. pp. 365-381. Springer (2020) 3, 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "38. Zhang, X., Sugano, Y., Fritz, M., Bulling, A.: Appearance-based gaze estimation in the wild. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 4511-4520 (2015) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "39. Zhang, X., Sugano, Y., Fritz, M., Bulling, A.: Mpiigaze: Real-world dataset and deep appearance-based gaze estimation. IEEE transactions on pattern analysis and machine intelligence 41(1), 162-175 (2017) 10" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "40. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 423 + ], + "type": "text", + "content": "41. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) 4, 12" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 263, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 263, + 91, + 447, + 102 + ], + "type": "text", + "content": "LG-Gaze: Language-Guided Gaze Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LGM_ Large Multi-View Gaussian Model for High-Resolution 3D Content Creation/269a8871-4800-4336-9c13-fbe2b1d2c82b_content_list.json b/2024/LGM_ Large Multi-View Gaussian Model for High-Resolution 3D Content Creation/269a8871-4800-4336-9c13-fbe2b1d2c82b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..49f5a32075dbdfbb8f0ac3041221f18644882b32 --- /dev/null +++ b/2024/LGM_ Large Multi-View Gaussian Model for High-Resolution 3D Content Creation/269a8871-4800-4336-9c13-fbe2b1d2c82b_content_list.json @@ -0,0 +1,1918 @@ +[ + { + "type": "text", + "text": "LGM: Large Multi-View Gaussian Model for High-Resolution 3D Content Creation", + "text_level": 1, + "bbox": [ + 245, + 140, + 756, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiaxiang Tang $^{1\\star}$ , Zhaoxi Chen $^{2}$ , Xiaokang Chen $^{1}$ , Tengfei Wang $^{3}$ , Gang Zeng $^{1}$ , and Ziwei Liu $^{2}$", + "bbox": [ + 217, + 210, + 784, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National Key Lab of General AI, Peking University $^{2}$ S-Lab, Nanyang Technological University $^{3}$ Shanghai AI Lab", + "bbox": [ + 321, + 253, + 679, + 296 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/08258ec94e30e51b7ad9b35f5165fdf0d0690d7324dfacd3f0514be4d512bd2b.jpg", + "image_caption": [ + "Fig.1: Our method generates high-resolution 3D Gaussians in 5 seconds from single-view images or texts." + ], + "image_footnote": [], + "bbox": [ + 225, + 314, + 781, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. 3D content creation has achieved significant progress in terms of both quality and speed. Although current feed-forward models can produce 3D objects in seconds, their resolution is constrained by the intensive computation required during training. In this paper, we introduce Large Multi-View Gaussian Model (LGM), a novel framework designed to generate high-resolution 3D models from text prompts or single-view images. Our key insights are two-fold: 1) 3D Representation: We propose multi-view Gaussian features as an efficient yet powerful representation, which can then be fused together for differentiable rendering. 2) 3D Backbone: We present an asymmetric U-Net as a high-throughput backbone operating on multi-view images, which can be produced from text or single-view image input by leveraging multi-view", + "bbox": [ + 259, + 648, + 743, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work done while visiting S-Lab, Nanyang Technological University.", + "bbox": [ + 217, + 824, + 679, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "diffusion models. Extensive experiments demonstrate the high fidelity and efficiency of our approach. Notably, we maintain the fast speed to generate 3D objects within 5 seconds while boosting the training resolution to 512, thereby achieving high-resolution 3D content generation. Our project page is available at https://me.kiui.moe/lgm/.", + "bbox": [ + 259, + 147, + 738, + 218 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: 3D Generation $\\cdot$ Gaussian Splitting $\\cdot$ High Resolution", + "bbox": [ + 261, + 229, + 710, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 267, + 375, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Automatic 3D content creation has great potential in numerous fields such as digital games, virtual reality, and films. The fundamental techniques, like imaged-to-3D and text-to-3D, provide significant benefits by remarkably decreasing the requirement for manual labor among professional 3D artists, enabling those without expertise to participate in 3D asset creation.", + "bbox": [ + 212, + 296, + 785, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous research on 3D generation has predominantly focused on score distillation sampling (SDS) [22,24,36,48] to lift 2D diffusion priors into 3D. These optimization-based methods can create highly detailed 3D objects from text or single-view image inputs, but they often face issues such as slow generation speed and limited diversity. Recent advancements have significantly decreased the time required to generate 3D objects using large reconstruction models from single-view or few-shot images [15,19,53,56,58]. These methods utilize transformers to directly regress triplane-based [2] neural radiance fields (NeRF) [32]. However, these methods cannot produce detailed textures and complicated geometry due to the low-resolution training. We argue that their bottlenecks are 1) inefficient 3D representation, and 2) heavily parameterized 3D backbone. For instance, given a fixed compute budget, the triplane representation of LRM [15] is limited to the resolution of 32, while the resolution of the rendered image is capped at 128 due to the online volume rendering. Despite this, these methods suffer from the computationally intensive transformer-based backbone, which also leads to a limited training resolution.", + "bbox": [ + 212, + 371, + 785, + 612 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we present a novel method to train a few-shot 3D reconstruction model without relying on triplane-based volume rendering or transformers [15]. Instead, our approach employs 3D Gaussian splatting [17] of which features are predicted by an asymmetric U-Net as a high-throughput backbone [41,47]. The motivation of this design is to achieve high-resolution 3D generation, which necessitates an expressive 3D representation and the ability to train at high resolutions. Gaussian splatting stands out for 1) the expressiveness of compactly representing a scene compared with a single triplane, and 2) rendering efficiency compared with heavy volume rendering, which facilitates high-resolution training. However, it requires a sufficient number of 3D Gaussians to accurately represent detailed 3D information. Inspired by splatter image [47], we found that U-Net is effective in generating a sufficient number of Gaussians from multiview pixels, which maintains the capacity for high-resolution training at the same time. Note that, compared to previous methods [15,63], our default model is capable of generating 3D models with up to 65,536 Gaussians and can be", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Tang et al.", + "bbox": [ + 271, + 114, + 344, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "trained at a resolution of 512, while still maintaining the rapid generation speed of feed-forward regression models. As shown in Figure 1, our model supports both image-to-3D and text-to-3D tasks, capable of producing high-resolution, richly detailed 3D Gaussians in approximately 5 seconds.", + "bbox": [ + 212, + 146, + 784, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method adopts a multi-view reconstruction setting similar to Instant3D [19]. In this process, the image and camera embedding from each input view are transformed into a feature map, which can be decoded and fused as a set of Gaussians. Differentiable rendering is applied to render novel views from the fused 3D Gaussians, allowing end-to-end image-level supervision in high resolution. To enhance information sharing across all input views, attention blocks are integrated into the deeper layers of the U-Net. This enables us to train our network on multi-view image datasets [12] using only regressing objectives. During inference, our method leverages existing image or text to multi-view diffusion models [27,44,45,52] to produce multi-view images as inputs for our Gaussian fusion network. To overcome the domain gap between multi-view images rendered from actual 3D objects and synthesized using diffusion models, we further propose two proper data augmentations for robust training. Finally, considering the preference for polygonal meshes in downstream tasks, we design a general algorithm to convert generated 3D Gaussians to smooth and textured meshes.", + "bbox": [ + 212, + 207, + 805, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our contributions are:", + "bbox": [ + 238, + 434, + 493, + 448 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose a novel framework to generate high-resolution 3D Gaussians by fusing information from multi-view images, which can be generated from text prompts or single-view images.", + "2. We design an asymmetric U-Net based architecture for efficient end-to-end training with significantly higher resolution, investigate data augmentation techniques for robust training, and propose a general mesh extraction approach from 3D Gaussians.", + "3. Extensive experiments demonstrate the superior quality, resolution, and efficiency of our method in both text-to-3D and image-to-3D tasks." + ], + "bbox": [ + 220, + 455, + 782, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 613, + 387, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "High-Resolution 3D Generation. Current approaches for generating high-fidelity 3D models mostly rely on SDS-based optimization techniques. It requires both an expressive 3D representation and high-resolution supervision to effectively distill detailed information from 2D diffusion models into 3D. Due to the significant memory consumption associated with high-resolution rendering of NeRF, Magic3D [22] first converts NeRF to DMTet [43] and subsequently trains a second stage for finer resolution refinement. The hybrid representation of DMTet geometry and hash grid [34] textures enables the capture of high-quality 3D information, which can be efficiently rendered using differentiable rasterization [18]. Fantasia3D [6] explores to directly train DMTet with disentangled geometry and appearance generation. Subsequent studies [8,20,21,37,48,50,55] also employ a similar mesh-based stage, enabling high-resolution supervision for enhanced detail. Another promising 3D representation is Gaussian splatting [17]", + "bbox": [ + 212, + 643, + 787, + 843 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Large Multi-View Gaussian Model", + "bbox": [ + 500, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for its expressiveness and efficient rendering capabilities. Nonetheless, achieving rich details with this method necessitates appropriate initialization and careful densification during optimization [10, 60]. In contrast, our work investigates a feed-forward approach to directly generate a sufficient number of 3D Gaussians. Efficient 3D Generation. In contrast to SDS-based optimization methods, feed-forward 3D native methods are able to generate 3D assets within seconds after training on large-scale 3D datasets [11, 12]. Some works attempt to train text-conditioned diffusion models on 3D representations such as point clouds and volumes [1, 5, 9, 16, 26, 33, 35, 54, 59, 62]. However, these methods either cannot generalize well to large datasets or only produce low-quality 3D assets with simple textures. Recently, LRM [15] first shows that a regression model can be trained to robustly predict NeRF from a single-view image in just 5 seconds, which can be further exported to meshes. Instant3D [19] trains a text to multiview images diffusion model and a multi-view LRM to perform fast and diverse text-to-3D generation. The following works extend LRM to predict poses given multi-view images [53], combine with diffusion [58], and specialize on human data [56]. These feed-forward models can be trained with simple regression objectives and significantly accelerate the speed of 3D object generation. However, their triplane NeRF-based representation is restricted to a relatively low resolution and limits the final generation fidelity. Our model instead seeks to train a high-fidelity feed-forward model using Gaussian splatting and U-Net.", + "bbox": [ + 212, + 142, + 785, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gaussian Splatting for Generation. We specifically discuss recent methods in generation tasks using Gaussian splatting [4,7,23,39,57]. DreamGaussian [48] first combines 3D Gaussians with SDS-based optimization approaches to decrease generation time. GSGen [10] and GaussianDreamer [60] explore various densification and initialization strategies for text to 3D Gaussians generation. Despite the acceleration achieved, generating high-fidelity 3D Gaussians using these optimization-based methods still requires several minutes. TriplaneGaussian [63] introduces Gaussian splatting into the framework of LRM. This method starts by predicting Gaussian centers as point clouds and then projects them onto a triplane for other features. Nonetheless, the number of Gaussians and the resolution of the triplane are still limited, affecting the quality of the generated Gaussians. Splatter image [47] proposes to predict 3D Gaussians as pixels on the output feature map using U-Net from single-view images. This approach mainly focuses on single-view or two-view scenarios, limiting its generalization to large-scale datasets. Similarly, PixelSplat [3] predicts Gaussian parameters for each pixel of two posed images from scene datasets. We design a 4-view reconstruction model combined with existing multi-view diffusion models for general text or image to high-fidelity 3D object generation.", + "bbox": [ + 212, + 463, + 787, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Large Multi-View Gaussian Model", + "text_level": 1, + "bbox": [ + 215, + 760, + 586, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first provide the background information on Gaussian splatting and multiview diffusion models (Section 3.1). Then we introduce our high-resolution 3D content generation framework (Section 3.2), where the core part is an asymmet", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Tang et al.", + "bbox": [ + 271, + 114, + 344, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2b2dd69bcfb4a2fc87443d3049e3c0080251648237d0506cf8412a84c787dd5d.jpg", + "image_caption": [ + "Fig. 2: Pipeline. Our model is trained to reconstruct 3D Gaussians from multi-view images, which can be synthesized by off-the-shelf models [45,52] at inference time from only text, or only image, or both input. Polygonal meshes can be extracted optionally." + ], + "image_footnote": [], + "bbox": [ + 218, + 148, + 782, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ric U-Net backbone to predict and fuse 3D Gaussians from multi-view images (Section 3.3). We design careful data augmentation and training pipeline to enhance robustness and stability (Section 3.4). Finally, we describe an effective method for smooth textured mesh extraction from the generated 3D Gaussians (Section 3.5).", + "bbox": [ + 212, + 324, + 782, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 214, + 424, + 372, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gaussian Splatting. As introduced in [17], Gaussian splatting employs a collection of 3D Gaussians to represent 3D data. Specifically, each Gaussian is defined by a center $\\mathbf{x} \\in \\mathbb{R}^3$ , a scaling factor $\\mathbf{s} \\in \\mathbb{R}^3$ , and a rotation quaternion $\\mathbf{q} \\in \\mathbb{R}^4$ . Additionally, an opacity value $\\alpha \\in \\mathbb{R}$ and a color feature $\\mathbf{c} \\in \\mathbb{R}^C$ are maintained for rendering, where spherical harmonics can be used to model view-dependent effects. These parameters can be collectively denoted by $\\Theta$ , with $\\Theta_i = \\{\\mathbf{x}_i, \\mathbf{s}_i, \\mathbf{q}_i, \\alpha_i, \\mathbf{c}_i\\}$ representing the parameters for the $i$ -th Gaussian. Rendering of the 3D Gaussians involves projecting them onto the image plane as 2D Gaussians and performing alpha composition for each pixel in front-to-back depth order, thereby determining the final color and alpha.", + "bbox": [ + 212, + 449, + 787, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-View Diffusion Models. Original 2D diffusion models [40,42] primarily focus on generating single-view images and do not support 3D viewpoint manipulation. Recently, several methods [20,27,44,45,52] propose to fine-tune multi-view diffusion models on 3D datasets to incorporate camera poses as an additional input. These approaches enable the creation of multi-view images of the same object, either from a text prompt or a single-view image. However, due to the absence of an actual 3D model, inconsistencies may still occur across the generated views.", + "bbox": [ + 212, + 623, + 787, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Overall Framework", + "text_level": 1, + "bbox": [ + 214, + 768, + 419, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Figure 2, we adopt a two-step 3D generation pipeline at inference. Firstly, we take advantage of off-the-shelf text or image to multi-view diffusion models to generate multi-view images. Specifically, we adopt MVDream [45]", + "bbox": [ + 212, + 794, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Large Multi-View Gaussian Model", + "bbox": [ + 500, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/99cf58200fd2f09fa6de9c08debfde0f750783dead6723aed4515a9158f7bc81.jpg", + "image_caption": [ + "Fig. 3: Architecture of LGM. Our network adopts an asymmetric U-Net with cross-view self-attention. We take four images with camera ray embeddings as the input, and output four feature maps of 14 channels (visualized using the RGB and opacity channels) which are interpreted as 3D Gaussians. The fused Gaussians are then rendered at novel views and supervised with ground truth images." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 781, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for text input and ImageDream [52] for image (and optionally text) input. Both models are designed to generate multi-view images at four orthogonal azimuths and a fixed elevation. In the second step, we use a U-Net based model to predict 3D Gaussians from these sparse view images. Specifically, our model is trained to take four images with camera pose embeddings as input and predict four sets of Gaussians, which are fused to form the final 3D Gaussians. The generated Gaussians can be optionally converted to polygonal meshes using an extra conversion step, which is more suitable for downstream tasks.", + "bbox": [ + 212, + 455, + 787, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Asymmetric U-Net for 3D Gaussians", + "text_level": 1, + "bbox": [ + 215, + 598, + 568, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At the core of our framework is an asymmetric U-Net to predict and fuse Gaussians from multi-view images. The network architecture is shown in Figure 3. We take four images and corresponding camera poses as the input. Following previous works [58], we use the Plücker ray embedding to densely encode the camera poses. The RGB value and ray embedding are concatenated into a 9-channel feature map as the input to the first layer:", + "bbox": [ + 212, + 622, + 787, + 713 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {i} = \\left\\{\\mathbf {c} _ {i}, \\mathbf {o} _ {i} \\times \\mathbf {d} _ {i}, \\mathbf {d} _ {i} \\right\\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 723, + 785, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_i$ is the input feature for pixel $i$ , $\\mathbf{c}_i$ is the RGB value, $\\mathbf{d}_i$ is the ray direction, and $\\mathbf{o}_i$ is the ray origin.", + "bbox": [ + 212, + 750, + 784, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The U-Net is built with residual layers [13] and self-attention layers [51] similar to previous works [14, 31, 47]. We only add self-attention at deeper layers where the feature map resolution is down-sampled to save memory. To propagate information across multiple views, we flatten the four image features and", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Tang et al.", + "bbox": [ + 271, + 114, + 346, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "concatenate them before applying self-attention, similar to previous multi-view diffusion models [45, 52].", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Each pixel of the output feature map is treated as a 3D Gaussian inspired by splatter image [47]. Differently, our U-Net is designed to be asymmetric with a smaller output resolution compared to input, which allows us to use higher resolution input images and limit the number of output Gaussians. We discard the depth prediction required by explicit ray-wise camera projection in [47]. The output feature map contains 14 channels corresponding to the original attributes of each Gaussian $\\Theta_{i}$ . To stabilize the training, we choose some different activation functions compared to the original Gaussian Splatting [17]. We clamp the predicted positions $\\mathbf{x}_i$ into $[-1,1]^3$ , and multiply the softplus-activated scales $\\mathbf{s}_i$ with 0.1, such that the generated Gaussians at the beginning of training is close to the scene center. For each input view, the output feature map is transformed into a set of Gaussians. We simply concatenate these Gaussians from all four views as the final 3D Gaussians, which are used to render images at novel views for supervision.", + "bbox": [ + 212, + 176, + 787, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Robust Training", + "text_level": 1, + "bbox": [ + 215, + 406, + 398, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data Augmentation. We use multi-view images rendered from the Objaverse [12] dataset for training. However, at inference, we use synthesized multiview images by diffusion models [45, 52]. To mitigate the domain gap, we design two types of data augmentation for more robust training.", + "bbox": [ + 212, + 428, + 782, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Grid Distortion. Synthesizing 3D consistent multi-view images using 2D diffusion models has been explored by many works [25,44,45,52]. However, since there is no underlying 3D representation, the generated multi-view images often suffer from subtle inconsistency across different views. We try to simulate such inconsistency using grid distortion. Except for the first input view, the other three input views are randomly distorted with a random grid during training. This makes the model more robust to inconsistent multi-view input images.", + "bbox": [ + 212, + 488, + 787, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Orbital Camera Jitter. Another problem is that the synthesized multi-view images may not accurately follow the given camera poses. Following [15], we always normalized the camera poses at each training step such that the first view's camera pose is fixed. We therefore apply camera jitter to the last three input views during training.", + "bbox": [ + 212, + 594, + 787, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Loss Function. To supervise the concatenated Gaussians, we use the differentiable renderer implementation from [17] to render them. At each training step, we render the RGB image and alpha image of eight views, including four input views and four novel views. Following [15], we apply mean square error loss and VGG-based LPIPS loss [61] to the RGB image:", + "bbox": [ + 212, + 688, + 787, + 763 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r g b}} = \\mathcal {L} _ {\\mathrm {M S E}} \\left(I _ {\\mathrm {r g b}}, I _ {\\mathrm {r g b}} ^ {\\mathrm {G T}}\\right) + \\mathcal {L} _ {\\mathrm {L P I P S}} \\left(I _ {\\mathrm {r g b}}, I _ {\\mathrm {r g b}} ^ {\\mathrm {G T}}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 768, + 785, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further apply mean square error loss on the alpha image for faster convergence of the shape:", + "bbox": [ + 212, + 794, + 785, + 824 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\alpha} = \\mathcal {L} _ {\\mathrm {M S E}} (I _ {\\alpha}, I _ {\\alpha} ^ {\\mathrm {G T}}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 823, + 785, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Large Multi-View Gaussian Model", + "bbox": [ + 500, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/48f05a3506576e0012668129ebe86e37c3e442abd332d231c543ae7d1f7838c3.jpg", + "image_caption": [ + "Fig. 4: Mesh Extraction Pipeline. We implement an efficient pipeline to convert the 3D Gaussians into smooth and textured meshes." + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 781, + 373 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Mesh Extraction", + "text_level": 1, + "bbox": [ + 215, + 441, + 401, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since polygonal meshes are still the most widely used 3D representation in downstream tasks, we hope to further extract meshes from our generated Gaussians. Previous works [48] have tried to directly convert the opacity value of 3D Gaussians into an occupancy field for mesh extraction. However, we find this method dependent on aggressive densification during the optimization of 3D Gaussians to produce smooth occupancy field. On the contrary, the generated Gaussians in our method are usually sparse and cannot produce a suitable occupancy field, leading to an unsatisfactory surface with visible holes.", + "bbox": [ + 212, + 464, + 787, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Instead, we propose a more general mesh extraction pipeline from 3D Gaussians as illustrated in Figure 4. We first train an efficient NeRF [34] using the rendered images from 3D Gaussians on-the-fly, and then convert the NeRF to polygonal meshes [49]. Marching Cubes [28] is applied to extract a coarse mesh, which is then iteratively refined together with the appearance hash grid using differentiable rendering. Finally, we bake the appearance field onto the refined mesh to extract texture images. For more details, please refer to the supplementary materials and NeRF2Mesh [49]. With adequately optimized implementation, it takes only about 1 minute to perform this Gaussians to NeRF to mesh conversion.", + "bbox": [ + 212, + 585, + 787, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 757, + 375, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 786, + 457, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. We use a filtered subset of the Objaverse [12] dataset to train our model. Since there are many low-quality 3D models (e.g., partial scans, missing", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Tang et al.", + "bbox": [ + 271, + 114, + 346, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "textures) in the original Objaverse dataset, we filter the dataset by two empirical rules: (1) We examine the captions from Cap3D [30], and curate a list of words that usually appears in bad models (e.g., 'resembling', 'debris', 'frame'), which is used to filter all models whose caption includes any of these words. (2) We discard models with mostly white color after rendering, which usually indicates missing texture. These lead to a final set of around 80K 3D objects. We render the RGBA image from 100 camera views at the resolution of $512 \\times 512$ for training and validation.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Network Architecture. Our asymmetric U-Net model consists of 6 down blocks, 1 middle block, and 5 up blocks, with the input image at $256 \\times 256$ and output Gaussian feature map at $128 \\times 128$ . We use 4 input views, so the number of output Gaussians is $128 \\times 128 \\times 4 = 65, 536$ . The feature channels for all blocks are [64, 128, 256, 512, 1024, 1024], [1024] and [1024, 1024, 512, 256, 128] respectively. Each block contains a series of residual layers and an optional down-sample or up-sample layer. For the last 3 of down blocks, the middle block, and the first 3 up blocks, we also insert cross-view self-attention layers after the residual layers. The final feature maps are processed by a $1 \\times 1$ convolution layer to 14-channel pixel-wise Gaussian features. Following previous works [40, 47], we adopt Silu activation and group normalization for the U-Net.", + "bbox": [ + 212, + 287, + 789, + 454 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training. We train our model on 32 NVIDIA A100 (80G) GPUs for about 4 days. A batch size of 8 for each GPU is used, leading to an effective batch size of 256. For each batch, we randomly sample 8 camera views, with the first 4 views as the input, and all 8 views as the output for supervision. Similar to LRM [15], we transform the cameras of each batch such that the first input view is always the front view with an identity rotation matrix and fixed translation. The input images are assumed to have a white background. The output 3D Gaussians are rendered at $512 \\times 512$ resolution for mean square error loss. We resize the images to $256 \\times 256$ for LPIPS loss to save memory. The AdamW [29] optimizer is adopted with the learning rate of $4 \\times 10^{-4}$ , weight decay of 0.05, and betas of (0.9, 0.95). The learning rate is cosine annealed to 0 during the training. We clip the gradient with a maximum norm of 1.0. The probability for grid distortion and camera jitter is set to $50\\%$ .", + "bbox": [ + 212, + 472, + 789, + 670 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Inference. Our whole pipeline, including two multi-view diffusion models, takes only about 10 GB of GPU memory for inference, which is friendly for deployment. For the multi-view diffusion models, we use a guidance scale of 5 for ImageDream [52] and 7.5 for MVDream [45] following the original paper. The number of diffusion steps is set to 30 using the DDIM [46] scheduler. The camera elevation is fixed to 0, and azimuths to $[0,90,180,270]$ degree for the four generated views. For ImageDream [52], the text prompt is always left empty so the only input is a single-view image. Since the images generated by MVDream may contain various backgrounds, we apply background removal [38] and use white background.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Large Multi-View Gaussian Model", + "bbox": [ + 500, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/207f8e44de316b9bd214edeee9e026a5f962a23589db49ab0e0c67bb91b78f66.jpg", + "image_caption": [ + "Fig. 5: Comparisons of generated 3D Gaussians for image-to-3D. Our method generates Gaussian splatting with better visual quality on various challenging images." + ], + "image_footnote": [], + "bbox": [ + 230, + 148, + 776, + 751 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Tang et al.", + "bbox": [ + 271, + 114, + 346, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d1a5a59111478c47a7582feaea089a4c299d1fc6a03331f6f80a0b2793675188.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 148, + 277, + 210 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0a2ff86c817deb658e55878be19c8b1300536f611ea8073fa77fec0d2ee85679.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 218, + 284, + 271 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8120ff055e1af89ec5fb3f6e8d22a600fabbccaa8a294ed187b7234af669dcbf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 275, + 279, + 303 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/172e67b0c384be0bc723ccf57a5146e0546808a021ff9fce7bb9d631a5d67640.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 238, + 305, + 272, + 333 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8d308a89c4d341fafabaf56c50d170f6b917929d857ae779220bfc21e809dbe5.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 233, + 344, + 272, + 398 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/34500a389efec3eddb7c0d37b0eb49dc067b78bb0608ad305a1861c7f3db09e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 148, + 370, + 209 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/49786530e9791b6747beb2e90f9f81e32eab23c404d0ef3dc7b14e5d0f679339.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 217, + 370, + 272 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6c8d87cf5511f934967af10bf0691b7dc1069f6460d0a7a24db56db0cffcc40f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 275, + 362, + 305 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b5ffcc6fe9e94a6a95139e15479fb8894a236ba87e41593259fc23380c694424.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 320, + 305, + 356, + 335 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f5d12265e95ac72291c29a99d7d87c74ee1fe72ef5092916ce39b697008c65fd.jpg", + "image_caption": [ + "Fig. 6: Comparisons with LRM for image-to-3D. We compare our method with available results from LRM [15]." + ], + "image_footnote": [], + "bbox": [ + 316, + 345, + 356, + 401 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d569a9ac3d51fb110440c6f3a4f323f09464a3c1f7971e909fd53d8f5b9cf863.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 148, + 442, + 209 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/eba24e0c95905d0172d6a0cebef90d4154c9815d8f6ba2bb43cc128cc1401f57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 214, + 447, + 272 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/68a3b88d4fce57a20c0f08bde490a9b59a4542c9fcc1cedbfd322667f1461b22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 397, + 275, + 442, + 305 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b7606e5432514ac4b08dbfe921f9d29582c4cb7ac260b6af5787a62605611fa7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 305, + 442, + 337 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c5a747d6f0ba8f4a8706fb9b0d5f9ff9a6c5e5b7f3f78c5fb287d630fdcb3604.jpg", + "image_caption": [ + "LRM" + ], + "image_footnote": [], + "bbox": [ + 397, + 345, + 447, + 402 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b3b1dc7a22505225cb513c359b3753bf132f38c7a6aff07ebe3093032b5628a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 150, + 539, + 209 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d281c646bdf23ac47d7b71ed34a75d62d4c13b56480f0480a67e73497ab48670.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 214, + 537, + 272 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7dd866c9daba5c23ae05ff3ececabcdc63dfe02e6118866d8e6be92d1fad3e84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 275, + 526, + 305 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ffa30161d78479890d788e5d14844be60c36da7f5e316f67f935dd976382ada5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 305, + 519, + 339 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9ea7f3454f4a43859f7f421247b3a60f614e2f0a9e7e5e17564c7896dd155c06.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 344, + 521, + 404 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3850f4fe566f5646d7e33940b17086dfffa16d1b8b2a44129dd3cc0f93ebb7d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 150, + 612, + 210 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3299a8beeddc28f0f1051b468917db8f5be6ed3471609a790f56959702858aab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 552, + 215, + 617, + 272 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6d50475a10baece5c0b09efe8a28b9965a6d3369be14adc1b98eaaa512fac180.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 277, + 607, + 308 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/da5aa86a658340a30e311e8ac41dd2c9a90a96e034cc51841fb2f30bfe2eb83e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 309, + 602, + 338 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/cb169e6f4ac1b45fe1b4b0ab8d9202bf96b236ca817d2309db26ba3b1b44b1ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 343, + 601, + 402 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/270fc7c81c638b4904dd517e3bed84c734b3271508e1d94942e80b1c45a6a0e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 148, + 692, + 212 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1c4c31b97244152477f02bef8b2489349fe63aba58ff42165946f94a7800d80e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 214, + 697, + 273 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b1b889c3e523dfd4f43248f889470b7cf04c42295e242e290b4e6bf82fc30235.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 277, + 692, + 308 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/27a727f39bea55e923af42be8bdf524c0c4c5266fefb7aee6c890112b5510073.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 655, + 309, + 683, + 339 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9c2cbaac60d6606d3a874059b71925d6804ceec8ba4d8c63973dd6280ece94d8.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 651, + 344, + 681, + 402 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/14da764935bab1b1c326c5e1a0536a23785d1ed620b6ce3363f97b8719143daf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 150, + 771, + 212 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/502ec699069d8914328e01b84e27c90c800710c7165e319492a702f4b83aa293.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 215, + 782, + 273 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b3d6f0b8dfe2806fd8a26cde56e477a30ce6c0baf071bfa9475d67b728b455ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 277, + 774, + 308 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/540344f95e609d526187b72090741e3495e55624b64e642c95238aaa2c9a16df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 727, + 309, + 764, + 339 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c0a61109034a2414f0ce8210e78f0d6a3e931b6949c29051e6562b8dbc275bd7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 344, + 766, + 404 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Qualitative Comparisons", + "text_level": 1, + "bbox": [ + 215, + 494, + 465, + 510 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Image-to-3D. We first compare against recent methods [48,63] that are capable of generating 3D Gaussians. Figure 5 shows images rendered from the generated 3D Gaussians for comparison. The 3D Gaussians produced by our method have better visual quality and effectively preserve the content from the input view. Our high-resolution 3D Gaussians can be transformed into smooth textured meshes with minimal loss of quality in most cases. We also compare our results against LRM [15] using the available videos from their website in Figure 6. Specifically, our multi-view setting successfully mitigates the issue of blurry back views and flat geometry, resulting in enhanced detail even in unseen views.", + "bbox": [ + 212, + 526, + 787, + 662 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Text-to-3D. We then compare with recent methods [16,48] on text-to-3D tasks. We observe an enhanced quality and efficiency in our method, generating more realistic 3D objects, as illustrated in Figure 7. Due to the multi-view diffusion models, our model is also free from multi-face problems.", + "bbox": [ + 212, + 690, + 787, + 752 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Diversity. Notably, our pipeline exhibits high diversity in 3D generation, owing to the capability of multi-view diffusion model [45, 52]. As shown in Figure 8, with different random seeds, we can generate a variety of feasible objects from the same ambiguous text prompt or single-view image.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Large Multi-View Gaussian Model", + "bbox": [ + 500, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0b47b6ea8fd9aeefcf32c5fea4a4776389ead96885f02efcde5ebf4f9942680e.jpg", + "image_caption": [ + "Fig. 7: Comparisons of generated 3D models for text-to-3D. Our method achieves better text alignment and visual quality." + ], + "image_footnote": [], + "bbox": [ + 220, + 148, + 781, + 465 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/861673706a4a5e8038b8fcc538c065b8690f3480c2b0fb276bf5802dec9b717c.jpg", + "image_caption": [ + "Fig. 8: Diversity of our 3D generation. We can generate diverse 3D models given an ambiguous text description or single-view image." + ], + "image_footnote": [], + "bbox": [ + 225, + 518, + 781, + 712 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Quantitative Comparisons", + "text_level": 1, + "bbox": [ + 215, + 782, + 478, + 799 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We majorly conduct a user study to quantitatively evaluate our image-to-3D Gaussians generation performance. For a collection of 30 images, we render 360-", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Tang et al.", + "bbox": [ + 271, + 114, + 346, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/31123417aec3179d34cd435d3371485b08b28edbab960c2910038d3c29036ebe.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Image Consistency ↑Overall Quality ↑
DreamGaussian [48]2.301.98
TriplaneGaussian [63]3.022.67
LGM (Ours)4.183.95
", + "bbox": [ + 228, + 143, + 774, + 203 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 1: User Study on the quality of generated 3D Gaussians for image-to-3D tasks. The rating is of scale 1-5, the higher the better.", + "bbox": [ + 215, + 213, + 784, + 242 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d398eff655e1caec482ce2711b3ce1f011d5dece702d3ac5a9a270c41466a381.jpg", + "image_caption": [ + "Fig. 9: Ablation Study. We carry out ablation study on designs of our method." + ], + "image_footnote": [], + "bbox": [ + 225, + 271, + 509, + 506 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/949dfc8040fab91d32a1e3e9eda93b980e9832427876bf53e48a12590563ce37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 271, + 784, + 505 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "degree rotating videos of the 3D Gaussians generated from DreamGaussain [48] (only the first stage), TriplaneGaussian [63], and ours. There are in total 90 videos for evaluation in our user study. Each volunteer is shown 30 samples from mixed random methods, and asked to rate in two aspects: image consistency and overall model quality. We collect results from 20 volunteers and get 600 valid scores in total. As shown in Table 1, our method is preferred as it aligns with the original image content and shows better overall quality.", + "bbox": [ + 212, + 563, + 787, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 691, + 390, + 709 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Number of Views. We train an image-to-3D model with only one input views similar to splatter image [47], i.e., without the multi-view generation step. The U-Net takes the single input view as input with self-attention, and outputs Gaussian features as in our multi-view model. To compensate the number of Gaussians, we predict two Gaussians for each pixel of the output feature maps, leading to $128 \\times 128 \\times 2 = 32$ , 768 Gaussians. As illustrated in the top-left part of Figure 9, the single-view model can reconstruct faithful front-view, but fails to distinguish the back view and results in blurriness. This is as expected since the regressive", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Large Multi-View Gaussian Model", + "bbox": [ + 500, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "U-Net is more suitable for reconstruction tasks, and it's hard to generalize to large datasets in our experiments.", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Data Augmentation. We train a smaller model with or without applying data augmentation to validate its effectiveness. Although we observe a lower training loss for the model without data augmentation, the domain gap during inference leads to more floaters and worse geometry as shown in the bottom-left part of Figure 9. The model with data augmentation strategy can better correct the 3D inconsistency and inaccurate camera poses in the generated multi-view images, leading to fewer floaters.", + "bbox": [ + 215, + 207, + 785, + 311 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Training Resolution. Lastly, we train a model with a fewer number of Gaussians and smaller rendering resolution as in the right part of Figure 9. We remove the last up block of the U-Net so the number of output Gaussians is $64 \\times 64 \\times 4 = 16,384$ , and we render it at $256 \\times 256$ for supervision. The model can still converge and successfully reconstruct 3D Gaussians, but the details are worse compared to the $256 \\times 256$ input multi-view images. In contrast, our large resolution model at $512 \\times 512$ can capture better details and generate Gaussians with higher resolution.", + "bbox": [ + 215, + 342, + 785, + 463 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.5 Limitations", + "text_level": 1, + "bbox": [ + 215, + 494, + 354, + 508 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Since our model is essentially a multi-view reconstruction model, the 3D generation quality highly depends on the quality of four input views. However, current multi-view diffusion models [45, 52] are far from perfect: (1) There can be 3D inconsistency which misleads the reconstruction model to generate floaters. (2) The resolution of synthesized multi-view images is restricted to $256 \\times 256$ . (3) ImageDream [52] also fails to handle input image with a large elevation angle. We expect these limitations can be mitigated with better multi-view diffusion models such as Zero-1-to-3++ [44] in future works.", + "bbox": [ + 215, + 526, + 785, + 648 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 678, + 357, + 695 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we present a large multi-view Gaussian model for high-resolution 3D content generation. Our model, distinct from previous methods reliant on NeRF and transformers, employs Gaussian splatting and U-Net to address the challenges of high memory requirements and low-resolution training. Additionally, we explore data augmentation for better robustness, and introduce a mesh extraction algorithm for the generated 3D Gaussians. Our approach achieves both high-resolution and high-efficiency for 3D objects generation, proving its versatility and applicability in various contexts.", + "bbox": [ + 215, + 718, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Tang et al.", + "bbox": [ + 271, + 114, + 344, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported by the Sichuan Science and Technology Program (2023YFSY0008), National Natural Science Foundation of China (61632003, 61375022, 61403005), Grant SCITLAB-20017 of Intelligent Terminal Key Laboratory of SiChuan Province, Beijing Advanced Innovation Center for Intelligent Robots and Systems (2018IRS11), and PEK-SenseTime Joint Laboratory of Machine Vision. This project is also funded by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOE-T2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 321, + 321, + 335 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Cao, Z., Hong, F., Wu, T., Pan, L., Liu, Z.: Large-vocabulary 3d diffusion model with transformer. arXiv preprint arXiv:2309.07920 (2023)", + "2. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., Mello, S.D., Gallo, O., Guibas, L., Tremblay, J., Khamis, S., Karras, T., Wetzstein, G.: Efficient geometry-aware 3D generative adversarial networks. In: CVPR (2022)", + "3. Charatan, D., Li, S., Tagliasacchi, A., Sitzmann, V.: pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. arXiv preprint arXiv:2312.12337 (2023)", + "4. Chen, G., Wang, W.: A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890 (2024)", + "5. Chen, H., Gu, J., Chen, A., Tian, W., Tu, Z., Liu, L., Su, H.: Single-stage diffusion nef: A unified approach to 3d generation and reconstruction. arXiv preprint arXiv:2304.06714 (2023)", + "6. Chen, R., Chen, Y., Jiao, N., Jia, K.: *Fantasia3d: Disentangling geometry* and appearance for high-quality text-to-3d content creation. arXiv preprint arXiv:2303.13873 (2023)", + "7. Chen, Y., Chen, Z., Zhang, C., Wang, F., Yang, X., Wang, Y., Cai, Z., Yang, L., Liu, H., Lin, G.: Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. arXiv preprint arXiv:2311.14521 (2023)", + "8. Chen, Y., Zhang, C., Yang, X., Cai, Z., Yu, G., Yang, L., Lin, G.: It3d: Improved text-to-3d generation with explicit view synthesis. arXiv preprint arXiv:2308.11473 (2023)", + "9. Chen, Z., Hong, F., Mei, H., Wang, G., Yang, L., Liu, Z.: Primdiffusion: Volumetric primitives diffusion for 3d human generation. arXiv preprint arXiv:2312.04559 (2023)", + "10. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023)", + "11. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. arXiv preprint arXiv:2307.05663 (2023)", + "12. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., Vanderbilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objverse: A universe of annotated 3d objects. In: CVPR. pp. 13142-13153 (2023)", + "13. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + ], + "bbox": [ + 218, + 353, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Large Multi-View Gaussian Model", + "bbox": [ + 500, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS 33, 6840-6851 (2020)", + "15. Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400 (2023)", + "16. Jun, H., Nichol, A.: Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023)", + "17. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ToG 42(4), 1-14 (2023)", + "18. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ToG 39(6) (2020)", + "19. Li, J., Tan, H., Zhang, K., Xu, Z., Luan, F., Xu, Y., Hong, Y., Sunkavalli, K., Shakhnarovich, G., Bi, S.: Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214 (2023)", + "20. Li, W., Chen, R., Chen, X., Tan, P.: Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596 (2023)", + "21. Li, Y., Dou, Y., Shi, Y., Lei, Y., Chen, X., Zhang, Y., Zhou, P., Ni, B.: Focaldreamer: Text-driven 3d editing via focal-fusion assembly. arXiv preprint arXiv:2308.10608 (2023)", + "22. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: CVPR. pp. 300–309 (2023)", + "23. Ling, H., Kim, S.W., Torralba, A., Fidler, S., Kreis, K.: Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763 (2023)", + "24. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328 (2023)", + "25. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023)", + "26. Liu, Z., Feng, Y., Black, M.J., Nowrouzezahrai, D., Paull, L., Liu, W.: Meshdiffusion: Score-based generative 3d mesh modeling. arXiv preprint arXiv:2303.08133 (2023)", + "27. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008 (2023)", + "28. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. In: Seminal graphics: pioneering efforts that shaped the field, pp. 347-353 (1998)", + "29. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "30. Luo, T., Rockwell, C., Lee, H., Johnson, J.: Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279 (2023)", + "31. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latentnerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600 (2022)", + "32. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "33. Müller, N., Siddiqui, Y., Porzi, L., Bulo, S.R., Kontschieder, P., Nießner, M.: Diffrf: Rendering-guided 3d radiance field diffusion. In: CVPR. pp. 4328-4338 (2023)" + ], + "bbox": [ + 217, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Tang et al.", + "bbox": [ + 271, + 114, + 344, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG (2022)", + "35. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751 (2022)", + "36. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)", + "37. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., et al.: Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843 (2023)", + "38. Qin, X., Zhang, Z., Huang, C., Dehghan, M., Zaiane, O.R., Jagersand, M.: U2-net: Going deeper with nested u-structure for salient object detection. Pattern recognition 106, 107404 (2020)", + "39. Ren, J., Pan, L., Tang, J., Zhang, C., Cao, A., Zeng, G., Liu, Z.: Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142 (2023)", + "40. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)", + "41. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)", + "42. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS 35, 36479-36494 (2022)", + "43. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In: NeurIPS (2021)", + "44. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model (2023)", + "45. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023)", + "46. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)", + "47. Szymanowicz, S., Rupprecht, C., Vedaldi, A.: Splatter image: Ultra-fast single-view 3d reconstruction. In: arXiv (2023)", + "48. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)", + "49. Tang, J., Zhou, H., Chen, X., Hu, T., Ding, E., Wang, J., Zeng, G.: Delicate textured mesh recovery from nef via adaptive surface refinement. arXiv preprint arXiv:2303.02091 (2022)", + "50. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023)", + "51. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)", + "52. Wang, P., Shi, Y.: Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201 (2023)", + "53. Wang, P., Tan, H., Bi, S., Xu, Y., Luan, F., Sunkavalli, K., Wang, W., Xu, Z., Zhang, K.: Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. arXiv preprint arXiv:2311.12024 (2023)" + ], + "bbox": [ + 212, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Large Multi-View Gaussian Model", + "bbox": [ + 500, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "54. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. In: CVPR. pp. 4563-4573 (2023)", + "55. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213 (2023)", + "56. Weng, Z., Liu, J., Tan, H., Xu, Z., Zhou, Y., Yeung-Levy, S., Yang, J.: Single-view 3d human digitalization with large reconstruction models. arXiv preprint arXiv:2401.12175 (2024)", + "57. Xu, D., Yuan, Y., Mardani, M., Liu, S., Song, J., Wang, Z., Vahdat, A.: Agg: Amortized generative 3d gaussians for single image to 3d. arXiv preprint arXiv:2401.04099 (2024)", + "58. Xu, Y., Tan, H., Luan, F., Bi, S., Wang, P., Li, J., Shi, Z., Sunkavalli, K., Wetzstein, G., Xu, Z., et al.: Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. arXiv preprint arXiv:2311.09217 (2023)", + "59. Yariv, L., Puny, O., Neverova, N., Gafni, O., Lipman, Y.: Mosaic-sdf for 3d generative models. arXiv preprint arXiv:2312.09222 (2023)", + "60. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023)", + "61. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)", + "62. Zhao, Z., Liu, W., Chen, X., Zeng, X., Wang, R., Cheng, P., Fu, B., Chen, T., Yu, G., Gao, S.: Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. arXiv preprint arXiv:2306.17115 (2023)", + "63. Zou, Z.X., Yu, Z., Guo, Y.C., Li, Y., Liang, D., Cao, Y.P., Zhang, S.H.: Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. arXiv preprint arXiv:2312.09147 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 534 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Tang et al.", + "bbox": [ + 271, + 114, + 344, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/LGM_ Large Multi-View Gaussian Model for High-Resolution 3D Content Creation/269a8871-4800-4336-9c13-fbe2b1d2c82b_model.json b/2024/LGM_ Large Multi-View Gaussian Model for High-Resolution 3D Content Creation/269a8871-4800-4336-9c13-fbe2b1d2c82b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e2bc93d793be4bb8afe13c423d761fa92161e127 --- /dev/null +++ b/2024/LGM_ Large Multi-View Gaussian Model for High-Resolution 3D Content Creation/269a8871-4800-4336-9c13-fbe2b1d2c82b_model.json @@ -0,0 +1,2601 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.246, + 0.141, + 0.758, + 0.187 + ], + "angle": 0, + "content": "LGM: Large Multi-View Gaussian Model for High-Resolution 3D Content Creation" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.212, + 0.785, + 0.243 + ], + "angle": 0, + "content": "Jiaxiang Tang\\(^{1\\star}\\), Zhaoxi Chen\\(^{2}\\), Xiaokang Chen\\(^{1}\\), Tengfei Wang\\(^{3}\\), Gang Zeng\\(^{1}\\), and Ziwei Liu\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.322, + 0.255, + 0.68, + 0.297 + ], + "angle": 0, + "content": "\\(^{1}\\) National Key Lab of General AI, Peking University \\(^{2}\\) S-Lab, Nanyang Technological University \\(^{3}\\) Shanghai AI Lab" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.315, + 0.782, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.625 + ], + "angle": 0, + "content": "Fig.1: Our method generates high-resolution 3D Gaussians in 5 seconds from single-view images or texts." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.65, + 0.744, + 0.817 + ], + "angle": 0, + "content": "Abstract. 3D content creation has achieved significant progress in terms of both quality and speed. Although current feed-forward models can produce 3D objects in seconds, their resolution is constrained by the intensive computation required during training. In this paper, we introduce Large Multi-View Gaussian Model (LGM), a novel framework designed to generate high-resolution 3D models from text prompts or single-view images. Our key insights are two-fold: 1) 3D Representation: We propose multi-view Gaussian features as an efficient yet powerful representation, which can then be fused together for differentiable rendering. 2) 3D Backbone: We present an asymmetric U-Net as a high-throughput backbone operating on multi-view images, which can be produced from text or single-view image input by leveraging multi-view" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.681, + 0.841 + ], + "angle": 0, + "content": "* Work done while visiting S-Lab, Nanyang Technological University." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.129 + ], + "angle": 0, + "content": "Tang et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.74, + 0.219 + ], + "angle": 0, + "content": "diffusion models. Extensive experiments demonstrate the high fidelity and efficiency of our approach. Notably, we maintain the fast speed to generate 3D objects within 5 seconds while boosting the training resolution to 512, thereby achieving high-resolution 3D content generation. Our project page is available at https://me.kiui.moe/lgm/." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.23, + 0.712, + 0.244 + ], + "angle": 0, + "content": "Keywords: 3D Generation \\(\\cdot\\) Gaussian Splitting \\(\\cdot\\) High Resolution" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.268, + 0.376, + 0.284 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.786, + 0.371 + ], + "angle": 0, + "content": "Automatic 3D content creation has great potential in numerous fields such as digital games, virtual reality, and films. The fundamental techniques, like imaged-to-3D and text-to-3D, provide significant benefits by remarkably decreasing the requirement for manual labor among professional 3D artists, enabling those without expertise to participate in 3D asset creation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.787, + 0.613 + ], + "angle": 0, + "content": "Previous research on 3D generation has predominantly focused on score distillation sampling (SDS) [22,24,36,48] to lift 2D diffusion priors into 3D. These optimization-based methods can create highly detailed 3D objects from text or single-view image inputs, but they often face issues such as slow generation speed and limited diversity. Recent advancements have significantly decreased the time required to generate 3D objects using large reconstruction models from single-view or few-shot images [15,19,53,56,58]. These methods utilize transformers to directly regress triplane-based [2] neural radiance fields (NeRF) [32]. However, these methods cannot produce detailed textures and complicated geometry due to the low-resolution training. We argue that their bottlenecks are 1) inefficient 3D representation, and 2) heavily parameterized 3D backbone. For instance, given a fixed compute budget, the triplane representation of LRM [15] is limited to the resolution of 32, while the resolution of the rendered image is capped at 128 due to the online volume rendering. Despite this, these methods suffer from the computationally intensive transformer-based backbone, which also leads to a limited training resolution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To address these challenges, we present a novel method to train a few-shot 3D reconstruction model without relying on triplane-based volume rendering or transformers [15]. Instead, our approach employs 3D Gaussian splatting [17] of which features are predicted by an asymmetric U-Net as a high-throughput backbone [41,47]. The motivation of this design is to achieve high-resolution 3D generation, which necessitates an expressive 3D representation and the ability to train at high resolutions. Gaussian splatting stands out for 1) the expressiveness of compactly representing a scene compared with a single triplane, and 2) rendering efficiency compared with heavy volume rendering, which facilitates high-resolution training. However, it requires a sufficient number of 3D Gaussians to accurately represent detailed 3D information. Inspired by splatter image [47], we found that U-Net is effective in generating a sufficient number of Gaussians from multiview pixels, which maintains the capacity for high-resolution training at the same time. Note that, compared to previous methods [15,63], our default model is capable of generating 3D models with up to 65,536 Gaussians and can be" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.5, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Large Multi-View Gaussian Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.207 + ], + "angle": 0, + "content": "trained at a resolution of 512, while still maintaining the rapid generation speed of feed-forward regression models. As shown in Figure 1, our model supports both image-to-3D and text-to-3D tasks, capable of producing high-resolution, richly detailed 3D Gaussians in approximately 5 seconds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.806, + 0.433 + ], + "angle": 0, + "content": "Our method adopts a multi-view reconstruction setting similar to Instant3D [19]. In this process, the image and camera embedding from each input view are transformed into a feature map, which can be decoded and fused as a set of Gaussians. Differentiable rendering is applied to render novel views from the fused 3D Gaussians, allowing end-to-end image-level supervision in high resolution. To enhance information sharing across all input views, attention blocks are integrated into the deeper layers of the U-Net. This enables us to train our network on multi-view image datasets [12] using only regressing objectives. During inference, our method leverages existing image or text to multi-view diffusion models [27,44,45,52] to produce multi-view images as inputs for our Gaussian fusion network. To overcome the domain gap between multi-view images rendered from actual 3D objects and synthesized using diffusion models, we further propose two proper data augmentations for robust training. Finally, considering the preference for polygonal meshes in downstream tasks, we design a general algorithm to convert generated 3D Gaussians to smooth and textured meshes." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.435, + 0.495, + 0.449 + ], + "angle": 0, + "content": "In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.457, + 0.784, + 0.5 + ], + "angle": 0, + "content": "1. We propose a novel framework to generate high-resolution 3D Gaussians by fusing information from multi-view images, which can be generated from text prompts or single-view images." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.502, + 0.784, + 0.56 + ], + "angle": 0, + "content": "2. We design an asymmetric U-Net based architecture for efficient end-to-end training with significantly higher resolution, investigate data augmentation techniques for robust training, and propose a general mesh extraction approach from 3D Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.562, + 0.784, + 0.591 + ], + "angle": 0, + "content": "3. Extensive experiments demonstrate the superior quality, resolution, and efficiency of our method in both text-to-3D and image-to-3D tasks." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.457, + 0.784, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.614, + 0.388, + 0.629 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.844 + ], + "angle": 0, + "content": "High-Resolution 3D Generation. Current approaches for generating high-fidelity 3D models mostly rely on SDS-based optimization techniques. It requires both an expressive 3D representation and high-resolution supervision to effectively distill detailed information from 2D diffusion models into 3D. Due to the significant memory consumption associated with high-resolution rendering of NeRF, Magic3D [22] first converts NeRF to DMTet [43] and subsequently trains a second stage for finer resolution refinement. The hybrid representation of DMTet geometry and hash grid [34] textures enables the capture of high-quality 3D information, which can be efficiently rendered using differentiable rasterization [18]. Fantasia3D [6] explores to directly train DMTet with disentangled geometry and appearance generation. Subsequent studies [8,20,21,37,48,50,55] also employ a similar mesh-based stage, enabling high-resolution supervision for enhanced detail. Another promising 3D representation is Gaussian splatting [17]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.129 + ], + "angle": 0, + "content": "Tang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.143, + 0.786, + 0.462 + ], + "angle": 0, + "content": "for its expressiveness and efficient rendering capabilities. Nonetheless, achieving rich details with this method necessitates appropriate initialization and careful densification during optimization [10, 60]. In contrast, our work investigates a feed-forward approach to directly generate a sufficient number of 3D Gaussians. Efficient 3D Generation. In contrast to SDS-based optimization methods, feed-forward 3D native methods are able to generate 3D assets within seconds after training on large-scale 3D datasets [11, 12]. Some works attempt to train text-conditioned diffusion models on 3D representations such as point clouds and volumes [1, 5, 9, 16, 26, 33, 35, 54, 59, 62]. However, these methods either cannot generalize well to large datasets or only produce low-quality 3D assets with simple textures. Recently, LRM [15] first shows that a regression model can be trained to robustly predict NeRF from a single-view image in just 5 seconds, which can be further exported to meshes. Instant3D [19] trains a text to multiview images diffusion model and a multi-view LRM to perform fast and diverse text-to-3D generation. The following works extend LRM to predict poses given multi-view images [53], combine with diffusion [58], and specialize on human data [56]. These feed-forward models can be trained with simple regression objectives and significantly accelerate the speed of 3D object generation. However, their triplane NeRF-based representation is restricted to a relatively low resolution and limits the final generation fidelity. Our model instead seeks to train a high-fidelity feed-forward model using Gaussian splatting and U-Net." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.788, + 0.737 + ], + "angle": 0, + "content": "Gaussian Splatting for Generation. We specifically discuss recent methods in generation tasks using Gaussian splatting [4,7,23,39,57]. DreamGaussian [48] first combines 3D Gaussians with SDS-based optimization approaches to decrease generation time. GSGen [10] and GaussianDreamer [60] explore various densification and initialization strategies for text to 3D Gaussians generation. Despite the acceleration achieved, generating high-fidelity 3D Gaussians using these optimization-based methods still requires several minutes. TriplaneGaussian [63] introduces Gaussian splatting into the framework of LRM. This method starts by predicting Gaussian centers as point clouds and then projects them onto a triplane for other features. Nonetheless, the number of Gaussians and the resolution of the triplane are still limited, affecting the quality of the generated Gaussians. Splatter image [47] proposes to predict 3D Gaussians as pixels on the output feature map using U-Net from single-view images. This approach mainly focuses on single-view or two-view scenarios, limiting its generalization to large-scale datasets. Similarly, PixelSplat [3] predicts Gaussian parameters for each pixel of two posed images from scene datasets. We design a 4-view reconstruction model combined with existing multi-view diffusion models for general text or image to high-fidelity 3D object generation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.761, + 0.588, + 0.779 + ], + "angle": 0, + "content": "3 Large Multi-View Gaussian Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We first provide the background information on Gaussian splatting and multiview diffusion models (Section 3.1). Then we introduce our high-resolution 3D content generation framework (Section 3.2), where the core part is an asymmet" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.5, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Large Multi-View Gaussian Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.149, + 0.784, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.294 + ], + "angle": 0, + "content": "Fig. 2: Pipeline. Our model is trained to reconstruct 3D Gaussians from multi-view images, which can be synthesized by off-the-shelf models [45,52] at inference time from only text, or only image, or both input. Polygonal meshes can be extracted optionally." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.325, + 0.784, + 0.401 + ], + "angle": 0, + "content": "ric U-Net backbone to predict and fuse 3D Gaussians from multi-view images (Section 3.3). We design careful data augmentation and training pipeline to enhance robustness and stability (Section 3.4). Finally, we describe an effective method for smooth textured mesh extraction from the generated 3D Gaussians (Section 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.425, + 0.373, + 0.438 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.602 + ], + "angle": 0, + "content": "Gaussian Splatting. As introduced in [17], Gaussian splatting employs a collection of 3D Gaussians to represent 3D data. Specifically, each Gaussian is defined by a center \\(\\mathbf{x} \\in \\mathbb{R}^3\\), a scaling factor \\(\\mathbf{s} \\in \\mathbb{R}^3\\), and a rotation quaternion \\(\\mathbf{q} \\in \\mathbb{R}^4\\). Additionally, an opacity value \\(\\alpha \\in \\mathbb{R}\\) and a color feature \\(\\mathbf{c} \\in \\mathbb{R}^C\\) are maintained for rendering, where spherical harmonics can be used to model view-dependent effects. These parameters can be collectively denoted by \\(\\Theta\\), with \\(\\Theta_i = \\{\\mathbf{x}_i, \\mathbf{s}_i, \\mathbf{q}_i, \\alpha_i, \\mathbf{c}_i\\}\\) representing the parameters for the \\(i\\)-th Gaussian. Rendering of the 3D Gaussians involves projecting them onto the image plane as 2D Gaussians and performing alpha composition for each pixel in front-to-back depth order, thereby determining the final color and alpha." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.788, + 0.746 + ], + "angle": 0, + "content": "Multi-View Diffusion Models. Original 2D diffusion models [40,42] primarily focus on generating single-view images and do not support 3D viewpoint manipulation. Recently, several methods [20,27,44,45,52] propose to fine-tune multi-view diffusion models on 3D datasets to incorporate camera poses as an additional input. These approaches enable the creation of multi-view images of the same object, either from a text prompt or a single-view image. However, due to the absence of an actual 3D model, inconsistencies may still occur across the generated views." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.769, + 0.421, + 0.783 + ], + "angle": 0, + "content": "3.2 Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.842 + ], + "angle": 0, + "content": "As illustrated in Figure 2, we adopt a two-step 3D generation pipeline at inference. Firstly, we take advantage of off-the-shelf text or image to multi-view diffusion models to generate multi-view images. Specifically, we adopt MVDream [45]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.129 + ], + "angle": 0, + "content": "Tang et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.145, + 0.782, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.43 + ], + "angle": 0, + "content": "Fig. 3: Architecture of LGM. Our network adopts an asymmetric U-Net with cross-view self-attention. We take four images with camera ray embeddings as the input, and output four feature maps of 14 channels (visualized using the RGB and opacity channels) which are interpreted as 3D Gaussians. The fused Gaussians are then rendered at novel views and supervised with ground truth images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.579 + ], + "angle": 0, + "content": "for text input and ImageDream [52] for image (and optionally text) input. Both models are designed to generate multi-view images at four orthogonal azimuths and a fixed elevation. In the second step, we use a U-Net based model to predict 3D Gaussians from these sparse view images. Specifically, our model is trained to take four images with camera pose embeddings as input and predict four sets of Gaussians, which are fused to form the final 3D Gaussians. The generated Gaussians can be optionally converted to polygonal meshes using an extra conversion step, which is more suitable for downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.599, + 0.569, + 0.615 + ], + "angle": 0, + "content": "3.3 Asymmetric U-Net for 3D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.623, + 0.788, + 0.714 + ], + "angle": 0, + "content": "At the core of our framework is an asymmetric U-Net to predict and fuse Gaussians from multi-view images. The network architecture is shown in Figure 3. We take four images and corresponding camera poses as the input. Following previous works [58], we use the Plücker ray embedding to densely encode the camera poses. The RGB value and ray embedding are concatenated into a 9-channel feature map as the input to the first layer:" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.724, + 0.787, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {i} = \\left\\{\\mathbf {c} _ {i}, \\mathbf {o} _ {i} \\times \\mathbf {d} _ {i}, \\mathbf {d} _ {i} \\right\\} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.779 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_i\\) is the input feature for pixel \\(i\\), \\(\\mathbf{c}_i\\) is the RGB value, \\(\\mathbf{d}_i\\) is the ray direction, and \\(\\mathbf{o}_i\\) is the ray origin." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The U-Net is built with residual layers [13] and self-attention layers [51] similar to previous works [14, 31, 47]. We only add self-attention at deeper layers where the feature map resolution is down-sampled to save memory. To propagate information across multiple views, we flatten the four image features and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.5, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Large Multi-View Gaussian Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "concatenate them before applying self-attention, similar to previous multi-view diffusion models [45, 52]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.389 + ], + "angle": 0, + "content": "Each pixel of the output feature map is treated as a 3D Gaussian inspired by splatter image [47]. Differently, our U-Net is designed to be asymmetric with a smaller output resolution compared to input, which allows us to use higher resolution input images and limit the number of output Gaussians. We discard the depth prediction required by explicit ray-wise camera projection in [47]. The output feature map contains 14 channels corresponding to the original attributes of each Gaussian \\(\\Theta_{i}\\). To stabilize the training, we choose some different activation functions compared to the original Gaussian Splatting [17]. We clamp the predicted positions \\(\\mathbf{x}_i\\) into \\([-1,1]^3\\), and multiply the softplus-activated scales \\(\\mathbf{s}_i\\) with 0.1, such that the generated Gaussians at the beginning of training is close to the scene center. For each input view, the output feature map is transformed into a set of Gaussians. We simply concatenate these Gaussians from all four views as the final 3D Gaussians, which are used to render images at novel views for supervision." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.407, + 0.4, + 0.423 + ], + "angle": 0, + "content": "3.4 Robust Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.784, + 0.489 + ], + "angle": 0, + "content": "Data Augmentation. We use multi-view images rendered from the Objaverse [12] dataset for training. However, at inference, we use synthesized multiview images by diffusion models [45, 52]. To mitigate the domain gap, we design two types of data augmentation for more robust training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.594 + ], + "angle": 0, + "content": "Grid Distortion. Synthesizing 3D consistent multi-view images using 2D diffusion models has been explored by many works [25,44,45,52]. However, since there is no underlying 3D representation, the generated multi-view images often suffer from subtle inconsistency across different views. We try to simulate such inconsistency using grid distortion. Except for the first input view, the other three input views are randomly distorted with a random grid during training. This makes the model more robust to inconsistent multi-view input images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.671 + ], + "angle": 0, + "content": "Orbital Camera Jitter. Another problem is that the synthesized multi-view images may not accurately follow the given camera poses. Following [15], we always normalized the camera poses at each training step such that the first view's camera pose is fixed. We therefore apply camera jitter to the last three input views during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Loss Function. To supervise the concatenated Gaussians, we use the differentiable renderer implementation from [17] to render them. At each training step, we render the RGB image and alpha image of eight views, including four input views and four novel views. Following [15], we apply mean square error loss and VGG-based LPIPS loss [61] to the RGB image:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.77, + 0.787, + 0.791 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r g b}} = \\mathcal {L} _ {\\mathrm {M S E}} \\left(I _ {\\mathrm {r g b}}, I _ {\\mathrm {r g b}} ^ {\\mathrm {G T}}\\right) + \\mathcal {L} _ {\\mathrm {L P I P S}} \\left(I _ {\\mathrm {r g b}}, I _ {\\mathrm {r g b}} ^ {\\mathrm {G T}}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.825 + ], + "angle": 0, + "content": "We further apply mean square error loss on the alpha image for faster convergence of the shape:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.824, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\alpha} = \\mathcal {L} _ {\\mathrm {M S E}} (I _ {\\alpha}, I _ {\\alpha} ^ {\\mathrm {G T}}) \\tag {3}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.129 + ], + "angle": 0, + "content": "Tang et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.145, + 0.782, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.386, + 0.788, + 0.414 + ], + "angle": 0, + "content": "Fig. 4: Mesh Extraction Pipeline. We implement an efficient pipeline to convert the 3D Gaussians into smooth and textured meshes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.442, + 0.403, + 0.456 + ], + "angle": 0, + "content": "3.5 Mesh Extraction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.585 + ], + "angle": 0, + "content": "Since polygonal meshes are still the most widely used 3D representation in downstream tasks, we hope to further extract meshes from our generated Gaussians. Previous works [48] have tried to directly convert the opacity value of 3D Gaussians into an occupancy field for mesh extraction. However, we find this method dependent on aggressive densification during the optimization of 3D Gaussians to produce smooth occupancy field. On the contrary, the generated Gaussians in our method are usually sparse and cannot produce a suitable occupancy field, leading to an unsatisfactory surface with visible holes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.586, + 0.789, + 0.737 + ], + "angle": 0, + "content": "Instead, we propose a more general mesh extraction pipeline from 3D Gaussians as illustrated in Figure 4. We first train an efficient NeRF [34] using the rendered images from 3D Gaussians on-the-fly, and then convert the NeRF to polygonal meshes [49]. Marching Cubes [28] is applied to extract a coarse mesh, which is then iteratively refined together with the appearance hash grid using differentiable rendering. Finally, we bake the appearance field onto the refined mesh to extract texture images. For more details, please refer to the supplementary materials and NeRF2Mesh [49]. With adequately optimized implementation, it takes only about 1 minute to perform this Gaussians to NeRF to mesh conversion." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.758, + 0.376, + 0.775 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.787, + 0.458, + 0.803 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Datasets. We use a filtered subset of the Objaverse [12] dataset to train our model. Since there are many low-quality 3D models (e.g., partial scans, missing" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.5, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Large Multi-View Gaussian Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "textures) in the original Objaverse dataset, we filter the dataset by two empirical rules: (1) We examine the captions from Cap3D [30], and curate a list of words that usually appears in bad models (e.g., 'resembling', 'debris', 'frame'), which is used to filter all models whose caption includes any of these words. (2) We discard models with mostly white color after rendering, which usually indicates missing texture. These lead to a final set of around 80K 3D objects. We render the RGBA image from 100 camera views at the resolution of \\(512 \\times 512\\) for training and validation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.288, + 0.79, + 0.455 + ], + "angle": 0, + "content": "Network Architecture. Our asymmetric U-Net model consists of 6 down blocks, 1 middle block, and 5 up blocks, with the input image at \\(256 \\times 256\\) and output Gaussian feature map at \\(128 \\times 128\\). We use 4 input views, so the number of output Gaussians is \\(128 \\times 128 \\times 4 = 65, 536\\). The feature channels for all blocks are [64, 128, 256, 512, 1024, 1024], [1024] and [1024, 1024, 512, 256, 128] respectively. Each block contains a series of residual layers and an optional down-sample or up-sample layer. For the last 3 of down blocks, the middle block, and the first 3 up blocks, we also insert cross-view self-attention layers after the residual layers. The final feature maps are processed by a \\(1 \\times 1\\) convolution layer to 14-channel pixel-wise Gaussian features. Following previous works [40, 47], we adopt Silu activation and group normalization for the U-Net." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.79, + 0.671 + ], + "angle": 0, + "content": "Training. We train our model on 32 NVIDIA A100 (80G) GPUs for about 4 days. A batch size of 8 for each GPU is used, leading to an effective batch size of 256. For each batch, we randomly sample 8 camera views, with the first 4 views as the input, and all 8 views as the output for supervision. Similar to LRM [15], we transform the cameras of each batch such that the first input view is always the front view with an identity rotation matrix and fixed translation. The input images are assumed to have a white background. The output 3D Gaussians are rendered at \\(512 \\times 512\\) resolution for mean square error loss. We resize the images to \\(256 \\times 256\\) for LPIPS loss to save memory. The AdamW [29] optimizer is adopted with the learning rate of \\(4 \\times 10^{-4}\\), weight decay of 0.05, and betas of (0.9, 0.95). The learning rate is cosine annealed to 0 during the training. We clip the gradient with a maximum norm of 1.0. The probability for grid distortion and camera jitter is set to \\(50\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Inference. Our whole pipeline, including two multi-view diffusion models, takes only about 10 GB of GPU memory for inference, which is friendly for deployment. For the multi-view diffusion models, we use a guidance scale of 5 for ImageDream [52] and 7.5 for MVDream [45] following the original paper. The number of diffusion steps is set to 30 using the DDIM [46] scheduler. The camera elevation is fixed to 0, and azimuths to \\([0,90,180,270]\\) degree for the four generated views. For ImageDream [52], the text prompt is always left empty so the only input is a single-view image. Since the images generated by MVDream may contain various backgrounds, we apply background removal [38] and use white background." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.129 + ], + "angle": 0, + "content": "Tang et al." + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.149, + 0.777, + 0.752 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.796 + ], + "angle": 0, + "content": "Fig. 5: Comparisons of generated 3D Gaussians for image-to-3D. Our method generates Gaussian splatting with better visual quality on various challenging images." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.5, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Large Multi-View Gaussian Model" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.149, + 0.278, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.219, + 0.285, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.276, + 0.28, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.239, + 0.306, + 0.273, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.345, + 0.274, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.409, + 0.272, + 0.422 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.149, + 0.371, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.218, + 0.372, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.276, + 0.363, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.306, + 0.357, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.346, + 0.357, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.15, + 0.444, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.215, + 0.449, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.276, + 0.444, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.306, + 0.443, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.346, + 0.449, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.41, + 0.41, + 0.437, + 0.421 + ], + "angle": 0, + "content": "LRM" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.151, + 0.54, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.215, + 0.539, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.276, + 0.527, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.306, + 0.521, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.345, + 0.522, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.151, + 0.614, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.216, + 0.618, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.278, + 0.608, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.31, + 0.603, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.344, + 0.602, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.15, + 0.694, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.215, + 0.698, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.278, + 0.693, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.31, + 0.684, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.345, + 0.682, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.41, + 0.684, + 0.42 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.151, + 0.772, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.217, + 0.784, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.722, + 0.278, + 0.776, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.31, + 0.766, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.345, + 0.767, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.435, + 0.785, + 0.465 + ], + "angle": 0, + "content": "Fig. 6: Comparisons with LRM for image-to-3D. We compare our method with available results from LRM [15]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.496, + 0.466, + 0.511 + ], + "angle": 0, + "content": "4.2 Qualitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.788, + 0.664 + ], + "angle": 0, + "content": "Image-to-3D. We first compare against recent methods [48,63] that are capable of generating 3D Gaussians. Figure 5 shows images rendered from the generated 3D Gaussians for comparison. The 3D Gaussians produced by our method have better visual quality and effectively preserve the content from the input view. Our high-resolution 3D Gaussians can be transformed into smooth textured meshes with minimal loss of quality in most cases. We also compare our results against LRM [15] using the available videos from their website in Figure 6. Specifically, our multi-view setting successfully mitigates the issue of blurry back views and flat geometry, resulting in enhanced detail even in unseen views." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.691, + 0.788, + 0.753 + ], + "angle": 0, + "content": "Text-to-3D. We then compare with recent methods [16,48] on text-to-3D tasks. We observe an enhanced quality and efficiency in our method, generating more realistic 3D objects, as illustrated in Figure 7. Due to the multi-view diffusion models, our model is also free from multi-face problems." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Diversity. Notably, our pipeline exhibits high diversity in 3D generation, owing to the capability of multi-view diffusion model [45, 52]. As shown in Figure 8, with different random seeds, we can generate a variety of feasible objects from the same ambiguous text prompt or single-view image." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.129 + ], + "angle": 0, + "content": "Tang et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.15, + 0.782, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.477, + 0.789, + 0.507 + ], + "angle": 0, + "content": "Fig. 7: Comparisons of generated 3D models for text-to-3D. Our method achieves better text alignment and visual quality." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.52, + 0.782, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.725, + 0.788, + 0.755 + ], + "angle": 0, + "content": "Fig. 8: Diversity of our 3D generation. We can generate diverse 3D models given an ambiguous text description or single-view image." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.784, + 0.48, + 0.8 + ], + "angle": 0, + "content": "4.3 Quantitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "We majorly conduct a user study to quantitatively evaluate our image-to-3D Gaussians generation performance. For a collection of 30 images, we render 360-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.5, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Large Multi-View Gaussian Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.23, + 0.144, + 0.775, + 0.204 + ], + "angle": 0, + "content": "
Image Consistency ↑Overall Quality ↑
DreamGaussian [48]2.301.98
TriplaneGaussian [63]3.022.67
LGM (Ours)4.183.95
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.214, + 0.785, + 0.243 + ], + "angle": 0, + "content": "Table 1: User Study on the quality of generated 3D Gaussians for image-to-3D tasks. The rating is of scale 1-5, the higher the better." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.272, + 0.51, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.272, + 0.785, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.519, + 0.773, + 0.535 + ], + "angle": 0, + "content": "Fig. 9: Ablation Study. We carry out ablation study on designs of our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.673 + ], + "angle": 0, + "content": "degree rotating videos of the 3D Gaussians generated from DreamGaussain [48] (only the first stage), TriplaneGaussian [63], and ours. There are in total 90 videos for evaluation in our user study. Each volunteer is shown 30 samples from mixed random methods, and asked to rate in two aspects: image consistency and overall model quality. We collect results from 20 volunteers and get 600 valid scores in total. As shown in Table 1, our method is preferred as it aligns with the original image content and shows better overall quality." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.693, + 0.391, + 0.71 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Number of Views. We train an image-to-3D model with only one input views similar to splatter image [47], i.e., without the multi-view generation step. The U-Net takes the single input view as input with self-attention, and outputs Gaussian features as in our multi-view model. To compensate the number of Gaussians, we predict two Gaussians for each pixel of the output feature maps, leading to \\(128 \\times 128 \\times 2 = 32\\), 768 Gaussians. As illustrated in the top-left part of Figure 9, the single-view model can reconstruct faithful front-view, but fails to distinguish the back view and results in blurriness. This is as expected since the regressive" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.129 + ], + "angle": 0, + "content": "Tang et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "U-Net is more suitable for reconstruction tasks, and it's hard to generalize to large datasets in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.208, + 0.787, + 0.313 + ], + "angle": 0, + "content": "Data Augmentation. We train a smaller model with or without applying data augmentation to validate its effectiveness. Although we observe a lower training loss for the model without data augmentation, the domain gap during inference leads to more floaters and worse geometry as shown in the bottom-left part of Figure 9. The model with data augmentation strategy can better correct the 3D inconsistency and inaccurate camera poses in the generated multi-view images, leading to fewer floaters." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.464 + ], + "angle": 0, + "content": "Training Resolution. Lastly, we train a model with a fewer number of Gaussians and smaller rendering resolution as in the right part of Figure 9. We remove the last up block of the U-Net so the number of output Gaussians is \\(64 \\times 64 \\times 4 = 16,384\\), and we render it at \\(256 \\times 256\\) for supervision. The model can still converge and successfully reconstruct 3D Gaussians, but the details are worse compared to the \\(256 \\times 256\\) input multi-view images. In contrast, our large resolution model at \\(512 \\times 512\\) can capture better details and generate Gaussians with higher resolution." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.495, + 0.356, + 0.509 + ], + "angle": 0, + "content": "4.5 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.527, + 0.787, + 0.65 + ], + "angle": 0, + "content": "Since our model is essentially a multi-view reconstruction model, the 3D generation quality highly depends on the quality of four input views. However, current multi-view diffusion models [45, 52] are far from perfect: (1) There can be 3D inconsistency which misleads the reconstruction model to generate floaters. (2) The resolution of synthesized multi-view images is restricted to \\(256 \\times 256\\). (3) ImageDream [52] also fails to handle input image with a large elevation angle. We expect these limitations can be mitigated with better multi-view diffusion models such as Zero-1-to-3++ [44] in future works." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.679, + 0.358, + 0.696 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.719, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this work, we present a large multi-view Gaussian model for high-resolution 3D content generation. Our model, distinct from previous methods reliant on NeRF and transformers, employs Gaussian splatting and U-Net to address the challenges of high memory requirements and low-resolution training. Additionally, we explore data augmentation for better robustness, and introduce a mesh extraction algorithm for the generated 3D Gaussians. Our approach achieves both high-resolution and high-efficiency for 3D objects generation, proving its versatility and applicability in various contexts." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.5, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Large Multi-View Gaussian Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported by the Sichuan Science and Technology Program (2023YFSY0008), National Natural Science Foundation of China (61632003, 61375022, 61403005), Grant SCITLAB-20017 of Intelligent Terminal Key Laboratory of SiChuan Province, Beijing Advanced Innovation Center for Intelligent Robots and Systems (2018IRS11), and PEK-SenseTime Joint Laboratory of Machine Vision. This project is also funded by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOE-T2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.322, + 0.323, + 0.337 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.354, + 0.785, + 0.381 + ], + "angle": 0, + "content": "1. Cao, Z., Hong, F., Wu, T., Pan, L., Liu, Z.: Large-vocabulary 3d diffusion model with transformer. arXiv preprint arXiv:2309.07920 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.383, + 0.785, + 0.422 + ], + "angle": 0, + "content": "2. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., Mello, S.D., Gallo, O., Guibas, L., Tremblay, J., Khamis, S., Karras, T., Wetzstein, G.: Efficient geometry-aware 3D generative adversarial networks. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.424, + 0.785, + 0.465 + ], + "angle": 0, + "content": "3. Charatan, D., Li, S., Tagliasacchi, A., Sitzmann, V.: pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. arXiv preprint arXiv:2312.12337 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.467, + 0.785, + 0.492 + ], + "angle": 0, + "content": "4. Chen, G., Wang, W.: A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.494, + 0.785, + 0.534 + ], + "angle": 0, + "content": "5. Chen, H., Gu, J., Chen, A., Tian, W., Tu, Z., Liu, L., Su, H.: Single-stage diffusion nef: A unified approach to 3d generation and reconstruction. arXiv preprint arXiv:2304.06714 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.785, + 0.576 + ], + "angle": 0, + "content": "6. Chen, R., Chen, Y., Jiao, N., Jia, K.: *Fantasia3d: Disentangling geometry* and appearance for high-quality text-to-3d content creation. arXiv preprint arXiv:2303.13873 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.785, + 0.617 + ], + "angle": 0, + "content": "7. Chen, Y., Chen, Z., Zhang, C., Wang, F., Yang, X., Wang, Y., Cai, Z., Yang, L., Liu, H., Lin, G.: Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. arXiv preprint arXiv:2311.14521 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.618, + 0.785, + 0.659 + ], + "angle": 0, + "content": "8. Chen, Y., Zhang, C., Yang, X., Cai, Z., Yu, G., Yang, L., Lin, G.: It3d: Improved text-to-3d generation with explicit view synthesis. arXiv preprint arXiv:2308.11473 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.785, + 0.7 + ], + "angle": 0, + "content": "9. Chen, Z., Hong, F., Mei, H., Wang, G., Yang, L., Liu, Z.: Primdiffusion: Volumetric primitives diffusion for 3d human generation. arXiv preprint arXiv:2312.04559 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.728 + ], + "angle": 0, + "content": "10. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.729, + 0.785, + 0.77 + ], + "angle": 0, + "content": "11. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. arXiv preprint arXiv:2307.05663 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "12. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., Vanderbilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objverse: A universe of annotated 3d objects. In: CVPR. pp. 13142-13153 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.354, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.129 + ], + "angle": 0, + "content": "Tang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.785, + 0.217 + ], + "angle": 0, + "content": "15. Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.785, + 0.243 + ], + "angle": 0, + "content": "16. Jun, H., Nichol, A.: Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.785, + 0.27 + ], + "angle": 0, + "content": "17. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ToG 42(4), 1-14 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.27, + 0.785, + 0.298 + ], + "angle": 0, + "content": "18. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ToG 39(6) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.298, + 0.785, + 0.338 + ], + "angle": 0, + "content": "19. Li, J., Tan, H., Zhang, K., Xu, Z., Luan, F., Xu, Y., Hong, Y., Sunkavalli, K., Shakhnarovich, G., Bi, S.: Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.338, + 0.785, + 0.365 + ], + "angle": 0, + "content": "20. Li, W., Chen, R., Chen, X., Tan, P.: Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.365, + 0.785, + 0.405 + ], + "angle": 0, + "content": "21. Li, Y., Dou, Y., Shi, Y., Lei, Y., Chen, X., Zhang, Y., Zhou, P., Ni, B.: Focaldreamer: Text-driven 3d editing via focal-fusion assembly. arXiv preprint arXiv:2308.10608 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.405, + 0.785, + 0.447 + ], + "angle": 0, + "content": "22. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: CVPR. pp. 300–309 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.447, + 0.785, + 0.488 + ], + "angle": 0, + "content": "23. Ling, H., Kim, S.W., Torralba, A., Fidler, S., Kreis, K.: Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.488, + 0.785, + 0.514 + ], + "angle": 0, + "content": "24. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.514, + 0.785, + 0.555 + ], + "angle": 0, + "content": "25. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.555, + 0.785, + 0.596 + ], + "angle": 0, + "content": "26. Liu, Z., Feng, Y., Black, M.J., Nowrouzezahrai, D., Paull, L., Liu, W.: Meshdiffusion: Score-based generative 3d mesh modeling. arXiv preprint arXiv:2303.08133 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.596, + 0.785, + 0.637 + ], + "angle": 0, + "content": "27. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.637, + 0.785, + 0.678 + ], + "angle": 0, + "content": "28. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. In: Seminal graphics: pioneering efforts that shaped the field, pp. 347-353 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.678, + 0.785, + 0.704 + ], + "angle": 0, + "content": "29. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.704, + 0.785, + 0.732 + ], + "angle": 0, + "content": "30. Luo, T., Rockwell, C., Lee, H., Johnson, J.: Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.732, + 0.785, + 0.772 + ], + "angle": 0, + "content": "31. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latentnerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "32. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.813, + 0.785, + 0.841 + ], + "angle": 0, + "content": "33. Müller, N., Siddiqui, Y., Porzi, L., Bulo, S.R., Kontschieder, P., Nießner, M.: Diffrf: Rendering-guided 3d radiance field diffusion. In: CVPR. pp. 4328-4338 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.5, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Large Multi-View Gaussian Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "34. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.215 + ], + "angle": 0, + "content": "35. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.216, + 0.785, + 0.242 + ], + "angle": 0, + "content": "36. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.297 + ], + "angle": 0, + "content": "37. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., et al.: Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.338 + ], + "angle": 0, + "content": "38. Qin, X., Zhang, Z., Huang, C., Dehghan, M., Zaiane, O.R., Jagersand, M.: U2-net: Going deeper with nested u-structure for salient object detection. Pattern recognition 106, 107404 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.365 + ], + "angle": 0, + "content": "39. Ren, J., Pan, L., Tang, J., Zhang, C., Cao, A., Zeng, G., Liu, Z.: Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.367, + 0.785, + 0.392 + ], + "angle": 0, + "content": "40. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.447 + ], + "angle": 0, + "content": "41. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.448, + 0.785, + 0.501 + ], + "angle": 0, + "content": "42. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS 35, 36479-36494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.528 + ], + "angle": 0, + "content": "43. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.785, + 0.555 + ], + "angle": 0, + "content": "44. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.785, + 0.582 + ], + "angle": 0, + "content": "45. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.609 + ], + "angle": 0, + "content": "46. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.636 + ], + "angle": 0, + "content": "47. Szymanowicz, S., Rupprecht, C., Vedaldi, A.: Splatter image: Ultra-fast single-view 3d reconstruction. In: arXiv (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.663 + ], + "angle": 0, + "content": "48. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.703 + ], + "angle": 0, + "content": "49. Tang, J., Zhou, H., Chen, X., Hu, T., Ding, E., Wang, J., Zeng, G.: Delicate textured mesh recovery from nef via adaptive surface refinement. arXiv preprint arXiv:2303.02091 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.744 + ], + "angle": 0, + "content": "50. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.771 + ], + "angle": 0, + "content": "51. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.798 + ], + "angle": 0, + "content": "52. Wang, P., Shi, Y.: Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "53. Wang, P., Tan, H., Bi, S., Xu, Y., Luan, F., Sunkavalli, K., Wang, W., Xu, Z., Zhang, K.: Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. arXiv preprint arXiv:2311.12024 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.129 + ], + "angle": 0, + "content": "Tang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "54. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. In: CVPR. pp. 4563-4573 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "55. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "56. Weng, Z., Liu, J., Tan, H., Xu, Z., Zhou, Y., Yeung-Levy, S., Yang, J.: Single-view 3d human digitalization with large reconstruction models. arXiv preprint arXiv:2401.12175 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.313 + ], + "angle": 0, + "content": "57. Xu, D., Yuan, Y., Mardani, M., Liu, S., Song, J., Wang, Z., Vahdat, A.: Agg: Amortized generative 3d gaussians for single image to 3d. arXiv preprint arXiv:2401.04099 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "58. Xu, Y., Tan, H., Luan, F., Bi, S., Wang, P., Li, J., Shi, Z., Sunkavalli, K., Wetzstein, G., Xu, Z., et al.: Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. arXiv preprint arXiv:2311.09217 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.384 + ], + "angle": 0, + "content": "59. Yariv, L., Puny, O., Neverova, N., Gafni, O., Lipman, Y.: Mosaic-sdf for 3d generative models. arXiv preprint arXiv:2312.09222 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.385, + 0.785, + 0.424 + ], + "angle": 0, + "content": "60. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.785, + 0.452 + ], + "angle": 0, + "content": "61. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "62. Zhao, Z., Liu, W., Chen, X., Zeng, X., Wang, R., Cheng, P., Fu, B., Chen, T., Yu, G., Gao, S.: Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. arXiv preprint arXiv:2306.17115 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.535 + ], + "angle": 0, + "content": "63. Zou, Z.X., Yu, Z., Guo, Y.C., Li, Y., Liang, D., Cao, Y.P., Zhang, S.H.: Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. arXiv preprint arXiv:2312.09147 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.535 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/LGM_ Large Multi-View Gaussian Model for High-Resolution 3D Content Creation/269a8871-4800-4336-9c13-fbe2b1d2c82b_origin.pdf b/2024/LGM_ Large Multi-View Gaussian Model for High-Resolution 3D Content Creation/269a8871-4800-4336-9c13-fbe2b1d2c82b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..197ba6b9491e51f2e701d26f165d3e6a2a5a24ec --- /dev/null +++ b/2024/LGM_ Large Multi-View Gaussian Model for High-Resolution 3D Content Creation/269a8871-4800-4336-9c13-fbe2b1d2c82b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0d6647aa13ce30354ee81e24c867463e2b57b06da716a709f7454a8b4ffd19f +size 5651094 diff --git a/2024/LGM_ Large Multi-View Gaussian Model for High-Resolution 3D Content Creation/full.md b/2024/LGM_ Large Multi-View Gaussian Model for High-Resolution 3D Content Creation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..06af194d0e5afc020831a1a61211d44a4cb87bf3 --- /dev/null +++ b/2024/LGM_ Large Multi-View Gaussian Model for High-Resolution 3D Content Creation/full.md @@ -0,0 +1,318 @@ +# LGM: Large Multi-View Gaussian Model for High-Resolution 3D Content Creation + +Jiaxiang Tang $^{1\star}$ , Zhaoxi Chen $^{2}$ , Xiaokang Chen $^{1}$ , Tengfei Wang $^{3}$ , Gang Zeng $^{1}$ , and Ziwei Liu $^{2}$ + +$^{1}$ National Key Lab of General AI, Peking University $^{2}$ S-Lab, Nanyang Technological University $^{3}$ Shanghai AI Lab + +![](images/08258ec94e30e51b7ad9b35f5165fdf0d0690d7324dfacd3f0514be4d512bd2b.jpg) +Fig.1: Our method generates high-resolution 3D Gaussians in 5 seconds from single-view images or texts. + +Abstract. 3D content creation has achieved significant progress in terms of both quality and speed. Although current feed-forward models can produce 3D objects in seconds, their resolution is constrained by the intensive computation required during training. In this paper, we introduce Large Multi-View Gaussian Model (LGM), a novel framework designed to generate high-resolution 3D models from text prompts or single-view images. Our key insights are two-fold: 1) 3D Representation: We propose multi-view Gaussian features as an efficient yet powerful representation, which can then be fused together for differentiable rendering. 2) 3D Backbone: We present an asymmetric U-Net as a high-throughput backbone operating on multi-view images, which can be produced from text or single-view image input by leveraging multi-view + +diffusion models. Extensive experiments demonstrate the high fidelity and efficiency of our approach. Notably, we maintain the fast speed to generate 3D objects within 5 seconds while boosting the training resolution to 512, thereby achieving high-resolution 3D content generation. Our project page is available at https://me.kiui.moe/lgm/. + +Keywords: 3D Generation $\cdot$ Gaussian Splitting $\cdot$ High Resolution + +# 1 Introduction + +Automatic 3D content creation has great potential in numerous fields such as digital games, virtual reality, and films. The fundamental techniques, like imaged-to-3D and text-to-3D, provide significant benefits by remarkably decreasing the requirement for manual labor among professional 3D artists, enabling those without expertise to participate in 3D asset creation. + +Previous research on 3D generation has predominantly focused on score distillation sampling (SDS) [22,24,36,48] to lift 2D diffusion priors into 3D. These optimization-based methods can create highly detailed 3D objects from text or single-view image inputs, but they often face issues such as slow generation speed and limited diversity. Recent advancements have significantly decreased the time required to generate 3D objects using large reconstruction models from single-view or few-shot images [15,19,53,56,58]. These methods utilize transformers to directly regress triplane-based [2] neural radiance fields (NeRF) [32]. However, these methods cannot produce detailed textures and complicated geometry due to the low-resolution training. We argue that their bottlenecks are 1) inefficient 3D representation, and 2) heavily parameterized 3D backbone. For instance, given a fixed compute budget, the triplane representation of LRM [15] is limited to the resolution of 32, while the resolution of the rendered image is capped at 128 due to the online volume rendering. Despite this, these methods suffer from the computationally intensive transformer-based backbone, which also leads to a limited training resolution. + +To address these challenges, we present a novel method to train a few-shot 3D reconstruction model without relying on triplane-based volume rendering or transformers [15]. Instead, our approach employs 3D Gaussian splatting [17] of which features are predicted by an asymmetric U-Net as a high-throughput backbone [41,47]. The motivation of this design is to achieve high-resolution 3D generation, which necessitates an expressive 3D representation and the ability to train at high resolutions. Gaussian splatting stands out for 1) the expressiveness of compactly representing a scene compared with a single triplane, and 2) rendering efficiency compared with heavy volume rendering, which facilitates high-resolution training. However, it requires a sufficient number of 3D Gaussians to accurately represent detailed 3D information. Inspired by splatter image [47], we found that U-Net is effective in generating a sufficient number of Gaussians from multiview pixels, which maintains the capacity for high-resolution training at the same time. Note that, compared to previous methods [15,63], our default model is capable of generating 3D models with up to 65,536 Gaussians and can be + +trained at a resolution of 512, while still maintaining the rapid generation speed of feed-forward regression models. As shown in Figure 1, our model supports both image-to-3D and text-to-3D tasks, capable of producing high-resolution, richly detailed 3D Gaussians in approximately 5 seconds. + +Our method adopts a multi-view reconstruction setting similar to Instant3D [19]. In this process, the image and camera embedding from each input view are transformed into a feature map, which can be decoded and fused as a set of Gaussians. Differentiable rendering is applied to render novel views from the fused 3D Gaussians, allowing end-to-end image-level supervision in high resolution. To enhance information sharing across all input views, attention blocks are integrated into the deeper layers of the U-Net. This enables us to train our network on multi-view image datasets [12] using only regressing objectives. During inference, our method leverages existing image or text to multi-view diffusion models [27,44,45,52] to produce multi-view images as inputs for our Gaussian fusion network. To overcome the domain gap between multi-view images rendered from actual 3D objects and synthesized using diffusion models, we further propose two proper data augmentations for robust training. Finally, considering the preference for polygonal meshes in downstream tasks, we design a general algorithm to convert generated 3D Gaussians to smooth and textured meshes. + +In summary, our contributions are: + +1. We propose a novel framework to generate high-resolution 3D Gaussians by fusing information from multi-view images, which can be generated from text prompts or single-view images. +2. We design an asymmetric U-Net based architecture for efficient end-to-end training with significantly higher resolution, investigate data augmentation techniques for robust training, and propose a general mesh extraction approach from 3D Gaussians. +3. Extensive experiments demonstrate the superior quality, resolution, and efficiency of our method in both text-to-3D and image-to-3D tasks. + +# 2 Related Work + +High-Resolution 3D Generation. Current approaches for generating high-fidelity 3D models mostly rely on SDS-based optimization techniques. It requires both an expressive 3D representation and high-resolution supervision to effectively distill detailed information from 2D diffusion models into 3D. Due to the significant memory consumption associated with high-resolution rendering of NeRF, Magic3D [22] first converts NeRF to DMTet [43] and subsequently trains a second stage for finer resolution refinement. The hybrid representation of DMTet geometry and hash grid [34] textures enables the capture of high-quality 3D information, which can be efficiently rendered using differentiable rasterization [18]. Fantasia3D [6] explores to directly train DMTet with disentangled geometry and appearance generation. Subsequent studies [8,20,21,37,48,50,55] also employ a similar mesh-based stage, enabling high-resolution supervision for enhanced detail. Another promising 3D representation is Gaussian splatting [17] + +for its expressiveness and efficient rendering capabilities. Nonetheless, achieving rich details with this method necessitates appropriate initialization and careful densification during optimization [10, 60]. In contrast, our work investigates a feed-forward approach to directly generate a sufficient number of 3D Gaussians. Efficient 3D Generation. In contrast to SDS-based optimization methods, feed-forward 3D native methods are able to generate 3D assets within seconds after training on large-scale 3D datasets [11, 12]. Some works attempt to train text-conditioned diffusion models on 3D representations such as point clouds and volumes [1, 5, 9, 16, 26, 33, 35, 54, 59, 62]. However, these methods either cannot generalize well to large datasets or only produce low-quality 3D assets with simple textures. Recently, LRM [15] first shows that a regression model can be trained to robustly predict NeRF from a single-view image in just 5 seconds, which can be further exported to meshes. Instant3D [19] trains a text to multiview images diffusion model and a multi-view LRM to perform fast and diverse text-to-3D generation. The following works extend LRM to predict poses given multi-view images [53], combine with diffusion [58], and specialize on human data [56]. These feed-forward models can be trained with simple regression objectives and significantly accelerate the speed of 3D object generation. However, their triplane NeRF-based representation is restricted to a relatively low resolution and limits the final generation fidelity. Our model instead seeks to train a high-fidelity feed-forward model using Gaussian splatting and U-Net. + +Gaussian Splatting for Generation. We specifically discuss recent methods in generation tasks using Gaussian splatting [4,7,23,39,57]. DreamGaussian [48] first combines 3D Gaussians with SDS-based optimization approaches to decrease generation time. GSGen [10] and GaussianDreamer [60] explore various densification and initialization strategies for text to 3D Gaussians generation. Despite the acceleration achieved, generating high-fidelity 3D Gaussians using these optimization-based methods still requires several minutes. TriplaneGaussian [63] introduces Gaussian splatting into the framework of LRM. This method starts by predicting Gaussian centers as point clouds and then projects them onto a triplane for other features. Nonetheless, the number of Gaussians and the resolution of the triplane are still limited, affecting the quality of the generated Gaussians. Splatter image [47] proposes to predict 3D Gaussians as pixels on the output feature map using U-Net from single-view images. This approach mainly focuses on single-view or two-view scenarios, limiting its generalization to large-scale datasets. Similarly, PixelSplat [3] predicts Gaussian parameters for each pixel of two posed images from scene datasets. We design a 4-view reconstruction model combined with existing multi-view diffusion models for general text or image to high-fidelity 3D object generation. + +# 3 Large Multi-View Gaussian Model + +We first provide the background information on Gaussian splatting and multiview diffusion models (Section 3.1). Then we introduce our high-resolution 3D content generation framework (Section 3.2), where the core part is an asymmet + +![](images/2b2dd69bcfb4a2fc87443d3049e3c0080251648237d0506cf8412a84c787dd5d.jpg) +Fig. 2: Pipeline. Our model is trained to reconstruct 3D Gaussians from multi-view images, which can be synthesized by off-the-shelf models [45,52] at inference time from only text, or only image, or both input. Polygonal meshes can be extracted optionally. + +ric U-Net backbone to predict and fuse 3D Gaussians from multi-view images (Section 3.3). We design careful data augmentation and training pipeline to enhance robustness and stability (Section 3.4). Finally, we describe an effective method for smooth textured mesh extraction from the generated 3D Gaussians (Section 3.5). + +# 3.1 Preliminaries + +Gaussian Splatting. As introduced in [17], Gaussian splatting employs a collection of 3D Gaussians to represent 3D data. Specifically, each Gaussian is defined by a center $\mathbf{x} \in \mathbb{R}^3$ , a scaling factor $\mathbf{s} \in \mathbb{R}^3$ , and a rotation quaternion $\mathbf{q} \in \mathbb{R}^4$ . Additionally, an opacity value $\alpha \in \mathbb{R}$ and a color feature $\mathbf{c} \in \mathbb{R}^C$ are maintained for rendering, where spherical harmonics can be used to model view-dependent effects. These parameters can be collectively denoted by $\Theta$ , with $\Theta_i = \{\mathbf{x}_i, \mathbf{s}_i, \mathbf{q}_i, \alpha_i, \mathbf{c}_i\}$ representing the parameters for the $i$ -th Gaussian. Rendering of the 3D Gaussians involves projecting them onto the image plane as 2D Gaussians and performing alpha composition for each pixel in front-to-back depth order, thereby determining the final color and alpha. + +Multi-View Diffusion Models. Original 2D diffusion models [40,42] primarily focus on generating single-view images and do not support 3D viewpoint manipulation. Recently, several methods [20,27,44,45,52] propose to fine-tune multi-view diffusion models on 3D datasets to incorporate camera poses as an additional input. These approaches enable the creation of multi-view images of the same object, either from a text prompt or a single-view image. However, due to the absence of an actual 3D model, inconsistencies may still occur across the generated views. + +# 3.2 Overall Framework + +As illustrated in Figure 2, we adopt a two-step 3D generation pipeline at inference. Firstly, we take advantage of off-the-shelf text or image to multi-view diffusion models to generate multi-view images. Specifically, we adopt MVDream [45] + +![](images/99cf58200fd2f09fa6de9c08debfde0f750783dead6723aed4515a9158f7bc81.jpg) +Fig. 3: Architecture of LGM. Our network adopts an asymmetric U-Net with cross-view self-attention. We take four images with camera ray embeddings as the input, and output four feature maps of 14 channels (visualized using the RGB and opacity channels) which are interpreted as 3D Gaussians. The fused Gaussians are then rendered at novel views and supervised with ground truth images. + +for text input and ImageDream [52] for image (and optionally text) input. Both models are designed to generate multi-view images at four orthogonal azimuths and a fixed elevation. In the second step, we use a U-Net based model to predict 3D Gaussians from these sparse view images. Specifically, our model is trained to take four images with camera pose embeddings as input and predict four sets of Gaussians, which are fused to form the final 3D Gaussians. The generated Gaussians can be optionally converted to polygonal meshes using an extra conversion step, which is more suitable for downstream tasks. + +# 3.3 Asymmetric U-Net for 3D Gaussians + +At the core of our framework is an asymmetric U-Net to predict and fuse Gaussians from multi-view images. The network architecture is shown in Figure 3. We take four images and corresponding camera poses as the input. Following previous works [58], we use the Plücker ray embedding to densely encode the camera poses. The RGB value and ray embedding are concatenated into a 9-channel feature map as the input to the first layer: + +$$ +\mathbf {f} _ {i} = \left\{\mathbf {c} _ {i}, \mathbf {o} _ {i} \times \mathbf {d} _ {i}, \mathbf {d} _ {i} \right\} \tag {1} +$$ + +where $\mathbf{f}_i$ is the input feature for pixel $i$ , $\mathbf{c}_i$ is the RGB value, $\mathbf{d}_i$ is the ray direction, and $\mathbf{o}_i$ is the ray origin. + +The U-Net is built with residual layers [13] and self-attention layers [51] similar to previous works [14, 31, 47]. We only add self-attention at deeper layers where the feature map resolution is down-sampled to save memory. To propagate information across multiple views, we flatten the four image features and + +concatenate them before applying self-attention, similar to previous multi-view diffusion models [45, 52]. + +Each pixel of the output feature map is treated as a 3D Gaussian inspired by splatter image [47]. Differently, our U-Net is designed to be asymmetric with a smaller output resolution compared to input, which allows us to use higher resolution input images and limit the number of output Gaussians. We discard the depth prediction required by explicit ray-wise camera projection in [47]. The output feature map contains 14 channels corresponding to the original attributes of each Gaussian $\Theta_{i}$ . To stabilize the training, we choose some different activation functions compared to the original Gaussian Splatting [17]. We clamp the predicted positions $\mathbf{x}_i$ into $[-1,1]^3$ , and multiply the softplus-activated scales $\mathbf{s}_i$ with 0.1, such that the generated Gaussians at the beginning of training is close to the scene center. For each input view, the output feature map is transformed into a set of Gaussians. We simply concatenate these Gaussians from all four views as the final 3D Gaussians, which are used to render images at novel views for supervision. + +# 3.4 Robust Training + +Data Augmentation. We use multi-view images rendered from the Objaverse [12] dataset for training. However, at inference, we use synthesized multiview images by diffusion models [45, 52]. To mitigate the domain gap, we design two types of data augmentation for more robust training. + +Grid Distortion. Synthesizing 3D consistent multi-view images using 2D diffusion models has been explored by many works [25,44,45,52]. However, since there is no underlying 3D representation, the generated multi-view images often suffer from subtle inconsistency across different views. We try to simulate such inconsistency using grid distortion. Except for the first input view, the other three input views are randomly distorted with a random grid during training. This makes the model more robust to inconsistent multi-view input images. + +Orbital Camera Jitter. Another problem is that the synthesized multi-view images may not accurately follow the given camera poses. Following [15], we always normalized the camera poses at each training step such that the first view's camera pose is fixed. We therefore apply camera jitter to the last three input views during training. + +Loss Function. To supervise the concatenated Gaussians, we use the differentiable renderer implementation from [17] to render them. At each training step, we render the RGB image and alpha image of eight views, including four input views and four novel views. Following [15], we apply mean square error loss and VGG-based LPIPS loss [61] to the RGB image: + +$$ +\mathcal {L} _ {\mathrm {r g b}} = \mathcal {L} _ {\mathrm {M S E}} \left(I _ {\mathrm {r g b}}, I _ {\mathrm {r g b}} ^ {\mathrm {G T}}\right) + \mathcal {L} _ {\mathrm {L P I P S}} \left(I _ {\mathrm {r g b}}, I _ {\mathrm {r g b}} ^ {\mathrm {G T}}\right) \tag {2} +$$ + +We further apply mean square error loss on the alpha image for faster convergence of the shape: + +$$ +\mathcal {L} _ {\alpha} = \mathcal {L} _ {\mathrm {M S E}} (I _ {\alpha}, I _ {\alpha} ^ {\mathrm {G T}}) \tag {3} +$$ + +![](images/48f05a3506576e0012668129ebe86e37c3e442abd332d231c543ae7d1f7838c3.jpg) +Fig. 4: Mesh Extraction Pipeline. We implement an efficient pipeline to convert the 3D Gaussians into smooth and textured meshes. + +# 3.5 Mesh Extraction + +Since polygonal meshes are still the most widely used 3D representation in downstream tasks, we hope to further extract meshes from our generated Gaussians. Previous works [48] have tried to directly convert the opacity value of 3D Gaussians into an occupancy field for mesh extraction. However, we find this method dependent on aggressive densification during the optimization of 3D Gaussians to produce smooth occupancy field. On the contrary, the generated Gaussians in our method are usually sparse and cannot produce a suitable occupancy field, leading to an unsatisfactory surface with visible holes. + +Instead, we propose a more general mesh extraction pipeline from 3D Gaussians as illustrated in Figure 4. We first train an efficient NeRF [34] using the rendered images from 3D Gaussians on-the-fly, and then convert the NeRF to polygonal meshes [49]. Marching Cubes [28] is applied to extract a coarse mesh, which is then iteratively refined together with the appearance hash grid using differentiable rendering. Finally, we bake the appearance field onto the refined mesh to extract texture images. For more details, please refer to the supplementary materials and NeRF2Mesh [49]. With adequately optimized implementation, it takes only about 1 minute to perform this Gaussians to NeRF to mesh conversion. + +# 4 Experiments + +# 4.1 Implementation Details + +Datasets. We use a filtered subset of the Objaverse [12] dataset to train our model. Since there are many low-quality 3D models (e.g., partial scans, missing + +textures) in the original Objaverse dataset, we filter the dataset by two empirical rules: (1) We examine the captions from Cap3D [30], and curate a list of words that usually appears in bad models (e.g., 'resembling', 'debris', 'frame'), which is used to filter all models whose caption includes any of these words. (2) We discard models with mostly white color after rendering, which usually indicates missing texture. These lead to a final set of around 80K 3D objects. We render the RGBA image from 100 camera views at the resolution of $512 \times 512$ for training and validation. + +Network Architecture. Our asymmetric U-Net model consists of 6 down blocks, 1 middle block, and 5 up blocks, with the input image at $256 \times 256$ and output Gaussian feature map at $128 \times 128$ . We use 4 input views, so the number of output Gaussians is $128 \times 128 \times 4 = 65, 536$ . The feature channels for all blocks are [64, 128, 256, 512, 1024, 1024], [1024] and [1024, 1024, 512, 256, 128] respectively. Each block contains a series of residual layers and an optional down-sample or up-sample layer. For the last 3 of down blocks, the middle block, and the first 3 up blocks, we also insert cross-view self-attention layers after the residual layers. The final feature maps are processed by a $1 \times 1$ convolution layer to 14-channel pixel-wise Gaussian features. Following previous works [40, 47], we adopt Silu activation and group normalization for the U-Net. + +Training. We train our model on 32 NVIDIA A100 (80G) GPUs for about 4 days. A batch size of 8 for each GPU is used, leading to an effective batch size of 256. For each batch, we randomly sample 8 camera views, with the first 4 views as the input, and all 8 views as the output for supervision. Similar to LRM [15], we transform the cameras of each batch such that the first input view is always the front view with an identity rotation matrix and fixed translation. The input images are assumed to have a white background. The output 3D Gaussians are rendered at $512 \times 512$ resolution for mean square error loss. We resize the images to $256 \times 256$ for LPIPS loss to save memory. The AdamW [29] optimizer is adopted with the learning rate of $4 \times 10^{-4}$ , weight decay of 0.05, and betas of (0.9, 0.95). The learning rate is cosine annealed to 0 during the training. We clip the gradient with a maximum norm of 1.0. The probability for grid distortion and camera jitter is set to $50\%$ . + +Inference. Our whole pipeline, including two multi-view diffusion models, takes only about 10 GB of GPU memory for inference, which is friendly for deployment. For the multi-view diffusion models, we use a guidance scale of 5 for ImageDream [52] and 7.5 for MVDream [45] following the original paper. The number of diffusion steps is set to 30 using the DDIM [46] scheduler. The camera elevation is fixed to 0, and azimuths to $[0,90,180,270]$ degree for the four generated views. For ImageDream [52], the text prompt is always left empty so the only input is a single-view image. Since the images generated by MVDream may contain various backgrounds, we apply background removal [38] and use white background. + +![](images/207f8e44de316b9bd214edeee9e026a5f962a23589db49ab0e0c67bb91b78f66.jpg) +Fig. 5: Comparisons of generated 3D Gaussians for image-to-3D. Our method generates Gaussian splatting with better visual quality on various challenging images. + +![](images/d1a5a59111478c47a7582feaea089a4c299d1fc6a03331f6f80a0b2793675188.jpg) + +![](images/0a2ff86c817deb658e55878be19c8b1300536f611ea8073fa77fec0d2ee85679.jpg) + +![](images/8120ff055e1af89ec5fb3f6e8d22a600fabbccaa8a294ed187b7234af669dcbf.jpg) + +![](images/172e67b0c384be0bc723ccf57a5146e0546808a021ff9fce7bb9d631a5d67640.jpg) + +![](images/8d308a89c4d341fafabaf56c50d170f6b917929d857ae779220bfc21e809dbe5.jpg) +Input + +![](images/34500a389efec3eddb7c0d37b0eb49dc067b78bb0608ad305a1861c7f3db09e3.jpg) + +![](images/49786530e9791b6747beb2e90f9f81e32eab23c404d0ef3dc7b14e5d0f679339.jpg) + +![](images/6c8d87cf5511f934967af10bf0691b7dc1069f6460d0a7a24db56db0cffcc40f.jpg) + +![](images/b5ffcc6fe9e94a6a95139e15479fb8894a236ba87e41593259fc23380c694424.jpg) + +![](images/f5d12265e95ac72291c29a99d7d87c74ee1fe72ef5092916ce39b697008c65fd.jpg) +Fig. 6: Comparisons with LRM for image-to-3D. We compare our method with available results from LRM [15]. + +![](images/d569a9ac3d51fb110440c6f3a4f323f09464a3c1f7971e909fd53d8f5b9cf863.jpg) + +![](images/eba24e0c95905d0172d6a0cebef90d4154c9815d8f6ba2bb43cc128cc1401f57.jpg) + +![](images/68a3b88d4fce57a20c0f08bde490a9b59a4542c9fcc1cedbfd322667f1461b22.jpg) + +![](images/b7606e5432514ac4b08dbfe921f9d29582c4cb7ac260b6af5787a62605611fa7.jpg) + +![](images/c5a747d6f0ba8f4a8706fb9b0d5f9ff9a6c5e5b7f3f78c5fb287d630fdcb3604.jpg) +LRM + +![](images/b3b1dc7a22505225cb513c359b3753bf132f38c7a6aff07ebe3093032b5628a0.jpg) + +![](images/d281c646bdf23ac47d7b71ed34a75d62d4c13b56480f0480a67e73497ab48670.jpg) + +![](images/7dd866c9daba5c23ae05ff3ececabcdc63dfe02e6118866d8e6be92d1fad3e84.jpg) + +![](images/ffa30161d78479890d788e5d14844be60c36da7f5e316f67f935dd976382ada5.jpg) + +![](images/9ea7f3454f4a43859f7f421247b3a60f614e2f0a9e7e5e17564c7896dd155c06.jpg) + +![](images/3850f4fe566f5646d7e33940b17086dfffa16d1b8b2a44129dd3cc0f93ebb7d0.jpg) + +![](images/3299a8beeddc28f0f1051b468917db8f5be6ed3471609a790f56959702858aab.jpg) + +![](images/6d50475a10baece5c0b09efe8a28b9965a6d3369be14adc1b98eaaa512fac180.jpg) + +![](images/da5aa86a658340a30e311e8ac41dd2c9a90a96e034cc51841fb2f30bfe2eb83e.jpg) + +![](images/cb169e6f4ac1b45fe1b4b0ab8d9202bf96b236ca817d2309db26ba3b1b44b1ae.jpg) + +![](images/270fc7c81c638b4904dd517e3bed84c734b3271508e1d94942e80b1c45a6a0e3.jpg) + +![](images/1c4c31b97244152477f02bef8b2489349fe63aba58ff42165946f94a7800d80e.jpg) + +![](images/b1b889c3e523dfd4f43248f889470b7cf04c42295e242e290b4e6bf82fc30235.jpg) + +![](images/27a727f39bea55e923af42be8bdf524c0c4c5266fefb7aee6c890112b5510073.jpg) + +![](images/9c2cbaac60d6606d3a874059b71925d6804ceec8ba4d8c63973dd6280ece94d8.jpg) +Ours + +![](images/14da764935bab1b1c326c5e1a0536a23785d1ed620b6ce3363f97b8719143daf.jpg) + +![](images/502ec699069d8914328e01b84e27c90c800710c7165e319492a702f4b83aa293.jpg) + +![](images/b3d6f0b8dfe2806fd8a26cde56e477a30ce6c0baf071bfa9475d67b728b455ab.jpg) + +![](images/540344f95e609d526187b72090741e3495e55624b64e642c95238aaa2c9a16df.jpg) + +![](images/c0a61109034a2414f0ce8210e78f0d6a3e931b6949c29051e6562b8dbc275bd7.jpg) + +# 4.2 Qualitative Comparisons + +Image-to-3D. We first compare against recent methods [48,63] that are capable of generating 3D Gaussians. Figure 5 shows images rendered from the generated 3D Gaussians for comparison. The 3D Gaussians produced by our method have better visual quality and effectively preserve the content from the input view. Our high-resolution 3D Gaussians can be transformed into smooth textured meshes with minimal loss of quality in most cases. We also compare our results against LRM [15] using the available videos from their website in Figure 6. Specifically, our multi-view setting successfully mitigates the issue of blurry back views and flat geometry, resulting in enhanced detail even in unseen views. + +Text-to-3D. We then compare with recent methods [16,48] on text-to-3D tasks. We observe an enhanced quality and efficiency in our method, generating more realistic 3D objects, as illustrated in Figure 7. Due to the multi-view diffusion models, our model is also free from multi-face problems. + +Diversity. Notably, our pipeline exhibits high diversity in 3D generation, owing to the capability of multi-view diffusion model [45, 52]. As shown in Figure 8, with different random seeds, we can generate a variety of feasible objects from the same ambiguous text prompt or single-view image. + +![](images/0b47b6ea8fd9aeefcf32c5fea4a4776389ead96885f02efcde5ebf4f9942680e.jpg) +Fig. 7: Comparisons of generated 3D models for text-to-3D. Our method achieves better text alignment and visual quality. + +![](images/861673706a4a5e8038b8fcc538c065b8690f3480c2b0fb276bf5802dec9b717c.jpg) +Fig. 8: Diversity of our 3D generation. We can generate diverse 3D models given an ambiguous text description or single-view image. + +# 4.3 Quantitative Comparisons + +We majorly conduct a user study to quantitatively evaluate our image-to-3D Gaussians generation performance. For a collection of 30 images, we render 360- + +
Image Consistency ↑Overall Quality ↑
DreamGaussian [48]2.301.98
TriplaneGaussian [63]3.022.67
LGM (Ours)4.183.95
+ +Table 1: User Study on the quality of generated 3D Gaussians for image-to-3D tasks. The rating is of scale 1-5, the higher the better. + +![](images/d398eff655e1caec482ce2711b3ce1f011d5dece702d3ac5a9a270c41466a381.jpg) +Fig. 9: Ablation Study. We carry out ablation study on designs of our method. + +![](images/949dfc8040fab91d32a1e3e9eda93b980e9832427876bf53e48a12590563ce37.jpg) + +degree rotating videos of the 3D Gaussians generated from DreamGaussain [48] (only the first stage), TriplaneGaussian [63], and ours. There are in total 90 videos for evaluation in our user study. Each volunteer is shown 30 samples from mixed random methods, and asked to rate in two aspects: image consistency and overall model quality. We collect results from 20 volunteers and get 600 valid scores in total. As shown in Table 1, our method is preferred as it aligns with the original image content and shows better overall quality. + +# 4.4 Ablation Study + +Number of Views. We train an image-to-3D model with only one input views similar to splatter image [47], i.e., without the multi-view generation step. The U-Net takes the single input view as input with self-attention, and outputs Gaussian features as in our multi-view model. To compensate the number of Gaussians, we predict two Gaussians for each pixel of the output feature maps, leading to $128 \times 128 \times 2 = 32$ , 768 Gaussians. As illustrated in the top-left part of Figure 9, the single-view model can reconstruct faithful front-view, but fails to distinguish the back view and results in blurriness. This is as expected since the regressive + +U-Net is more suitable for reconstruction tasks, and it's hard to generalize to large datasets in our experiments. + +Data Augmentation. We train a smaller model with or without applying data augmentation to validate its effectiveness. Although we observe a lower training loss for the model without data augmentation, the domain gap during inference leads to more floaters and worse geometry as shown in the bottom-left part of Figure 9. The model with data augmentation strategy can better correct the 3D inconsistency and inaccurate camera poses in the generated multi-view images, leading to fewer floaters. + +Training Resolution. Lastly, we train a model with a fewer number of Gaussians and smaller rendering resolution as in the right part of Figure 9. We remove the last up block of the U-Net so the number of output Gaussians is $64 \times 64 \times 4 = 16,384$ , and we render it at $256 \times 256$ for supervision. The model can still converge and successfully reconstruct 3D Gaussians, but the details are worse compared to the $256 \times 256$ input multi-view images. In contrast, our large resolution model at $512 \times 512$ can capture better details and generate Gaussians with higher resolution. + +# 4.5 Limitations + +Since our model is essentially a multi-view reconstruction model, the 3D generation quality highly depends on the quality of four input views. However, current multi-view diffusion models [45, 52] are far from perfect: (1) There can be 3D inconsistency which misleads the reconstruction model to generate floaters. (2) The resolution of synthesized multi-view images is restricted to $256 \times 256$ . (3) ImageDream [52] also fails to handle input image with a large elevation angle. We expect these limitations can be mitigated with better multi-view diffusion models such as Zero-1-to-3++ [44] in future works. + +# 5 Conclusion + +In this work, we present a large multi-view Gaussian model for high-resolution 3D content generation. Our model, distinct from previous methods reliant on NeRF and transformers, employs Gaussian splatting and U-Net to address the challenges of high memory requirements and low-resolution training. Additionally, we explore data augmentation for better robustness, and introduce a mesh extraction algorithm for the generated 3D Gaussians. Our approach achieves both high-resolution and high-efficiency for 3D objects generation, proving its versatility and applicability in various contexts. + +Acknowledgements. This work is supported by the Sichuan Science and Technology Program (2023YFSY0008), National Natural Science Foundation of China (61632003, 61375022, 61403005), Grant SCITLAB-20017 of Intelligent Terminal Key Laboratory of SiChuan Province, Beijing Advanced Innovation Center for Intelligent Robots and Systems (2018IRS11), and PEK-SenseTime Joint Laboratory of Machine Vision. This project is also funded by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOE-T2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). + +# References + +1. Cao, Z., Hong, F., Wu, T., Pan, L., Liu, Z.: Large-vocabulary 3d diffusion model with transformer. arXiv preprint arXiv:2309.07920 (2023) +2. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., Mello, S.D., Gallo, O., Guibas, L., Tremblay, J., Khamis, S., Karras, T., Wetzstein, G.: Efficient geometry-aware 3D generative adversarial networks. In: CVPR (2022) +3. Charatan, D., Li, S., Tagliasacchi, A., Sitzmann, V.: pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. arXiv preprint arXiv:2312.12337 (2023) +4. Chen, G., Wang, W.: A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890 (2024) +5. Chen, H., Gu, J., Chen, A., Tian, W., Tu, Z., Liu, L., Su, H.: Single-stage diffusion nef: A unified approach to 3d generation and reconstruction. arXiv preprint arXiv:2304.06714 (2023) +6. Chen, R., Chen, Y., Jiao, N., Jia, K.: *Fantasia3d: Disentangling geometry* and appearance for high-quality text-to-3d content creation. arXiv preprint arXiv:2303.13873 (2023) +7. Chen, Y., Chen, Z., Zhang, C., Wang, F., Yang, X., Wang, Y., Cai, Z., Yang, L., Liu, H., Lin, G.: Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. arXiv preprint arXiv:2311.14521 (2023) +8. Chen, Y., Zhang, C., Yang, X., Cai, Z., Yu, G., Yang, L., Lin, G.: It3d: Improved text-to-3d generation with explicit view synthesis. arXiv preprint arXiv:2308.11473 (2023) +9. Chen, Z., Hong, F., Mei, H., Wang, G., Yang, L., Liu, Z.: Primdiffusion: Volumetric primitives diffusion for 3d human generation. arXiv preprint arXiv:2312.04559 (2023) +10. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023) +11. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. arXiv preprint arXiv:2307.05663 (2023) +12. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., Vanderbilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objverse: A universe of annotated 3d objects. In: CVPR. pp. 13142-13153 (2023) +13. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016) + +14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS 33, 6840-6851 (2020) +15. Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400 (2023) +16. Jun, H., Nichol, A.: Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023) +17. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ToG 42(4), 1-14 (2023) +18. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ToG 39(6) (2020) +19. Li, J., Tan, H., Zhang, K., Xu, Z., Luan, F., Xu, Y., Hong, Y., Sunkavalli, K., Shakhnarovich, G., Bi, S.: Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214 (2023) +20. Li, W., Chen, R., Chen, X., Tan, P.: Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596 (2023) +21. Li, Y., Dou, Y., Shi, Y., Lei, Y., Chen, X., Zhang, Y., Zhou, P., Ni, B.: Focaldreamer: Text-driven 3d editing via focal-fusion assembly. arXiv preprint arXiv:2308.10608 (2023) +22. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: CVPR. pp. 300–309 (2023) +23. Ling, H., Kim, S.W., Torralba, A., Fidler, S., Kreis, K.: Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763 (2023) +24. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328 (2023) +25. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023) +26. Liu, Z., Feng, Y., Black, M.J., Nowrouzezahrai, D., Paull, L., Liu, W.: Meshdiffusion: Score-based generative 3d mesh modeling. arXiv preprint arXiv:2303.08133 (2023) +27. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008 (2023) +28. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. In: Seminal graphics: pioneering efforts that shaped the field, pp. 347-353 (1998) +29. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +30. Luo, T., Rockwell, C., Lee, H., Johnson, J.: Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279 (2023) +31. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latentnerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600 (2022) +32. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +33. Müller, N., Siddiqui, Y., Porzi, L., Bulo, S.R., Kontschieder, P., Nießner, M.: Diffrf: Rendering-guided 3d radiance field diffusion. In: CVPR. pp. 4328-4338 (2023) + +34. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG (2022) +35. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751 (2022) +36. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022) +37. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., et al.: Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843 (2023) +38. Qin, X., Zhang, Z., Huang, C., Dehghan, M., Zaiane, O.R., Jagersand, M.: U2-net: Going deeper with nested u-structure for salient object detection. Pattern recognition 106, 107404 (2020) +39. Ren, J., Pan, L., Tang, J., Zhang, C., Cao, A., Zeng, G., Liu, Z.: Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142 (2023) +40. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022) +41. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) +42. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS 35, 36479-36494 (2022) +43. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In: NeurIPS (2021) +44. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model (2023) +45. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023) +46. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) +47. Szymanowicz, S., Rupprecht, C., Vedaldi, A.: Splatter image: Ultra-fast single-view 3d reconstruction. In: arXiv (2023) +48. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023) +49. Tang, J., Zhou, H., Chen, X., Hu, T., Ding, E., Wang, J., Zeng, G.: Delicate textured mesh recovery from nef via adaptive surface refinement. arXiv preprint arXiv:2303.02091 (2022) +50. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023) +51. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017) +52. Wang, P., Shi, Y.: Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201 (2023) +53. Wang, P., Tan, H., Bi, S., Xu, Y., Luan, F., Sunkavalli, K., Wang, W., Xu, Z., Zhang, K.: Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. arXiv preprint arXiv:2311.12024 (2023) + +54. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. In: CVPR. pp. 4563-4573 (2023) +55. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213 (2023) +56. Weng, Z., Liu, J., Tan, H., Xu, Z., Zhou, Y., Yeung-Levy, S., Yang, J.: Single-view 3d human digitalization with large reconstruction models. arXiv preprint arXiv:2401.12175 (2024) +57. Xu, D., Yuan, Y., Mardani, M., Liu, S., Song, J., Wang, Z., Vahdat, A.: Agg: Amortized generative 3d gaussians for single image to 3d. arXiv preprint arXiv:2401.04099 (2024) +58. Xu, Y., Tan, H., Luan, F., Bi, S., Wang, P., Li, J., Shi, Z., Sunkavalli, K., Wetzstein, G., Xu, Z., et al.: Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. arXiv preprint arXiv:2311.09217 (2023) +59. Yariv, L., Puny, O., Neverova, N., Gafni, O., Lipman, Y.: Mosaic-sdf for 3d generative models. arXiv preprint arXiv:2312.09222 (2023) +60. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023) +61. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) +62. Zhao, Z., Liu, W., Chen, X., Zeng, X., Wang, R., Cheng, P., Fu, B., Chen, T., Yu, G., Gao, S.: Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. arXiv preprint arXiv:2306.17115 (2023) +63. Zou, Z.X., Yu, Z., Guo, Y.C., Li, Y., Liang, D., Cao, Y.P., Zhang, S.H.: Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. arXiv preprint arXiv:2312.09147 (2023) \ No newline at end of file diff --git a/2024/LGM_ Large Multi-View Gaussian Model for High-Resolution 3D Content Creation/images.zip b/2024/LGM_ Large Multi-View Gaussian Model for High-Resolution 3D Content Creation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6f7197b50c732d698da1d41ba34ac473580aa11c --- /dev/null +++ b/2024/LGM_ Large Multi-View Gaussian Model for High-Resolution 3D Content Creation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e27351d0a4db7e1d00a4bd19038bfb6c6eb00c9fc046cab2c7d60abcd303f96 +size 673375 diff --git a/2024/LGM_ Large Multi-View Gaussian Model for High-Resolution 3D Content Creation/layout.json b/2024/LGM_ Large Multi-View Gaussian Model for High-Resolution 3D Content Creation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ba3fe4df1c796b32a7d52d69fd27cacf5483aa91 --- /dev/null +++ b/2024/LGM_ Large Multi-View Gaussian Model for High-Resolution 3D Content Creation/layout.json @@ -0,0 +1,9222 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 150, + 111, + 463, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 111, + 463, + 148 + ], + "spans": [ + { + "bbox": [ + 150, + 111, + 463, + 148 + ], + "type": "text", + "content": "LGM: Large Multi-View Gaussian Model for High-Resolution 3D Content Creation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 167, + 480, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 167, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 167, + 480, + 192 + ], + "type": "text", + "content": "Jiaxiang Tang" + }, + { + "bbox": [ + 133, + 167, + 480, + 192 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 133, + 167, + 480, + 192 + ], + "type": "text", + "content": ", Zhaoxi Chen" + }, + { + "bbox": [ + 133, + 167, + 480, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 167, + 480, + 192 + ], + "type": "text", + "content": ", Xiaokang Chen" + }, + { + "bbox": [ + 133, + 167, + 480, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 167, + 480, + 192 + ], + "type": "text", + "content": ", Tengfei Wang" + }, + { + "bbox": [ + 133, + 167, + 480, + 192 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 167, + 480, + 192 + ], + "type": "text", + "content": ", Gang Zeng" + }, + { + "bbox": [ + 133, + 167, + 480, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 167, + 480, + 192 + ], + "type": "text", + "content": ", and Ziwei Liu" + }, + { + "bbox": [ + 133, + 167, + 480, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 197, + 201, + 416, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 201, + 416, + 235 + ], + "spans": [ + { + "bbox": [ + 197, + 201, + 416, + 235 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 197, + 201, + 416, + 235 + ], + "type": "text", + "content": " National Key Lab of General AI, Peking University " + }, + { + "bbox": [ + 197, + 201, + 416, + 235 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 197, + 201, + 416, + 235 + ], + "type": "text", + "content": " S-Lab, Nanyang Technological University " + }, + { + "bbox": [ + 197, + 201, + 416, + 235 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 197, + 201, + 416, + 235 + ], + "type": "text", + "content": " Shanghai AI Lab" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 249, + 478, + 460 + ], + "blocks": [ + { + "bbox": [ + 138, + 249, + 478, + 460 + ], + "lines": [ + { + "bbox": [ + 138, + 249, + 478, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 478, + 460 + ], + "type": "image", + "image_path": "08258ec94e30e51b7ad9b35f5165fdf0d0690d7324dfacd3f0514be4d512bd2b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 471, + 482, + 495 + ], + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 495 + ], + "type": "text", + "content": "Fig.1: Our method generates high-resolution 3D Gaussians in 5 seconds from single-view images or texts." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 514, + 455, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 514, + 455, + 647 + ], + "spans": [ + { + "bbox": [ + 159, + 514, + 455, + 647 + ], + "type": "text", + "content": "Abstract. 3D content creation has achieved significant progress in terms of both quality and speed. Although current feed-forward models can produce 3D objects in seconds, their resolution is constrained by the intensive computation required during training. In this paper, we introduce Large Multi-View Gaussian Model (LGM), a novel framework designed to generate high-resolution 3D models from text prompts or single-view images. Our key insights are two-fold: 1) 3D Representation: We propose multi-view Gaussian features as an efficient yet powerful representation, which can then be fused together for differentiable rendering. 2) 3D Backbone: We present an asymmetric U-Net as a high-throughput backbone operating on multi-view images, which can be produced from text or single-view image input by leveraging multi-view" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 416, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 416, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 416, + 666 + ], + "type": "text", + "content": "* Work done while visiting S-Lab, Nanyang Technological University." + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 452, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 452, + 173 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 452, + 173 + ], + "type": "text", + "content": "diffusion models. Extensive experiments demonstrate the high fidelity and efficiency of our approach. Notably, we maintain the fast speed to generate 3D objects within 5 seconds while boosting the training resolution to 512, thereby achieving high-resolution 3D content generation. Our project page is available at https://me.kiui.moe/lgm/." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 182, + 435, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 182, + 435, + 193 + ], + "spans": [ + { + "bbox": [ + 160, + 182, + 435, + 193 + ], + "type": "text", + "content": "Keywords: 3D Generation " + }, + { + "bbox": [ + 160, + 182, + 435, + 193 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 182, + 435, + 193 + ], + "type": "text", + "content": " Gaussian Splitting " + }, + { + "bbox": [ + 160, + 182, + 435, + 193 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 182, + 435, + 193 + ], + "type": "text", + "content": " High Resolution" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 212, + 230, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 212, + 230, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 212, + 230, + 224 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 235, + 481, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 481, + 293 + ], + "type": "text", + "content": "Automatic 3D content creation has great potential in numerous fields such as digital games, virtual reality, and films. The fundamental techniques, like imaged-to-3D and text-to-3D, provide significant benefits by remarkably decreasing the requirement for manual labor among professional 3D artists, enabling those without expertise to participate in 3D asset creation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 294, + 481, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 481, + 485 + ], + "type": "text", + "content": "Previous research on 3D generation has predominantly focused on score distillation sampling (SDS) [22,24,36,48] to lift 2D diffusion priors into 3D. These optimization-based methods can create highly detailed 3D objects from text or single-view image inputs, but they often face issues such as slow generation speed and limited diversity. Recent advancements have significantly decreased the time required to generate 3D objects using large reconstruction models from single-view or few-shot images [15,19,53,56,58]. These methods utilize transformers to directly regress triplane-based [2] neural radiance fields (NeRF) [32]. However, these methods cannot produce detailed textures and complicated geometry due to the low-resolution training. We argue that their bottlenecks are 1) inefficient 3D representation, and 2) heavily parameterized 3D backbone. For instance, given a fixed compute budget, the triplane representation of LRM [15] is limited to the resolution of 32, while the resolution of the rendered image is capped at 128 due to the online volume rendering. Despite this, these methods suffer from the computationally intensive transformer-based backbone, which also leads to a limited training resolution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "To address these challenges, we present a novel method to train a few-shot 3D reconstruction model without relying on triplane-based volume rendering or transformers [15]. Instead, our approach employs 3D Gaussian splatting [17] of which features are predicted by an asymmetric U-Net as a high-throughput backbone [41,47]. The motivation of this design is to achieve high-resolution 3D generation, which necessitates an expressive 3D representation and the ability to train at high resolutions. Gaussian splatting stands out for 1) the expressiveness of compactly representing a scene compared with a single triplane, and 2) rendering efficiency compared with heavy volume rendering, which facilitates high-resolution training. However, it requires a sufficient number of 3D Gaussians to accurately represent detailed 3D information. Inspired by splatter image [47], we found that U-Net is effective in generating a sufficient number of Gaussians from multiview pixels, which maintains the capacity for high-resolution training at the same time. Note that, compared to previous methods [15,63], our default model is capable of generating 3D models with up to 65,536 Gaussians and can be" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "text", + "content": "Tang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": "trained at a resolution of 512, while still maintaining the rapid generation speed of feed-forward regression models. As shown in Figure 1, our model supports both image-to-3D and text-to-3D tasks, capable of producing high-resolution, richly detailed 3D Gaussians in approximately 5 seconds." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 493, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 493, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 493, + 342 + ], + "type": "text", + "content": "Our method adopts a multi-view reconstruction setting similar to Instant3D [19]. In this process, the image and camera embedding from each input view are transformed into a feature map, which can be decoded and fused as a set of Gaussians. Differentiable rendering is applied to render novel views from the fused 3D Gaussians, allowing end-to-end image-level supervision in high resolution. To enhance information sharing across all input views, attention blocks are integrated into the deeper layers of the U-Net. This enables us to train our network on multi-view image datasets [12] using only regressing objectives. During inference, our method leverages existing image or text to multi-view diffusion models [27,44,45,52] to produce multi-view images as inputs for our Gaussian fusion network. To overcome the domain gap between multi-view images rendered from actual 3D objects and synthesized using diffusion models, we further propose two proper data augmentations for robust training. Finally, considering the preference for polygonal meshes in downstream tasks, we design a general algorithm to convert generated 3D Gaussians to smooth and textured meshes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 344, + 302, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 344, + 302, + 355 + ], + "spans": [ + { + "bbox": [ + 146, + 344, + 302, + 355 + ], + "type": "text", + "content": "In summary, our contributions are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 361, + 479, + 468 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 136, + 361, + 479, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 361, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 136, + 361, + 479, + 396 + ], + "type": "text", + "content": "1. We propose a novel framework to generate high-resolution 3D Gaussians by fusing information from multi-view images, which can be generated from text prompts or single-view images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 397, + 479, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 397, + 479, + 443 + ], + "spans": [ + { + "bbox": [ + 136, + 397, + 479, + 443 + ], + "type": "text", + "content": "2. We design an asymmetric U-Net based architecture for efficient end-to-end training with significantly higher resolution, investigate data augmentation techniques for robust training, and propose a general mesh extraction approach from 3D Gaussians." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 445, + 479, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 445, + 479, + 468 + ], + "spans": [ + { + "bbox": [ + 135, + 445, + 479, + 468 + ], + "type": "text", + "content": "3. Extensive experiments demonstrate the superior quality, resolution, and efficiency of our method in both text-to-3D and image-to-3D tasks." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 486, + 237, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 237, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 237, + 498 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "text", + "content": "High-Resolution 3D Generation. Current approaches for generating high-fidelity 3D models mostly rely on SDS-based optimization techniques. It requires both an expressive 3D representation and high-resolution supervision to effectively distill detailed information from 2D diffusion models into 3D. Due to the significant memory consumption associated with high-resolution rendering of NeRF, Magic3D [22] first converts NeRF to DMTet [43] and subsequently trains a second stage for finer resolution refinement. The hybrid representation of DMTet geometry and hash grid [34] textures enables the capture of high-quality 3D information, which can be efficiently rendered using differentiable rasterization [18]. Fantasia3D [6] explores to directly train DMTet with disentangled geometry and appearance generation. Subsequent studies [8,20,21,37,48,50,55] also employ a similar mesh-based stage, enabling high-resolution supervision for enhanced detail. Another promising 3D representation is Gaussian splatting [17]" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 447, + 102 + ], + "type": "text", + "content": "Large Multi-View Gaussian Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 113, + 481, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 113, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 113, + 481, + 365 + ], + "type": "text", + "content": "for its expressiveness and efficient rendering capabilities. Nonetheless, achieving rich details with this method necessitates appropriate initialization and careful densification during optimization [10, 60]. In contrast, our work investigates a feed-forward approach to directly generate a sufficient number of 3D Gaussians. Efficient 3D Generation. In contrast to SDS-based optimization methods, feed-forward 3D native methods are able to generate 3D assets within seconds after training on large-scale 3D datasets [11, 12]. Some works attempt to train text-conditioned diffusion models on 3D representations such as point clouds and volumes [1, 5, 9, 16, 26, 33, 35, 54, 59, 62]. However, these methods either cannot generalize well to large datasets or only produce low-quality 3D assets with simple textures. Recently, LRM [15] first shows that a regression model can be trained to robustly predict NeRF from a single-view image in just 5 seconds, which can be further exported to meshes. Instant3D [19] trains a text to multiview images diffusion model and a multi-view LRM to perform fast and diverse text-to-3D generation. The following works extend LRM to predict poses given multi-view images [53], combine with diffusion [58], and specialize on human data [56]. These feed-forward models can be trained with simple regression objectives and significantly accelerate the speed of 3D object generation. However, their triplane NeRF-based representation is restricted to a relatively low resolution and limits the final generation fidelity. Our model instead seeks to train a high-fidelity feed-forward model using Gaussian splatting and U-Net." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 367, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 583 + ], + "type": "text", + "content": "Gaussian Splatting for Generation. We specifically discuss recent methods in generation tasks using Gaussian splatting [4,7,23,39,57]. DreamGaussian [48] first combines 3D Gaussians with SDS-based optimization approaches to decrease generation time. GSGen [10] and GaussianDreamer [60] explore various densification and initialization strategies for text to 3D Gaussians generation. Despite the acceleration achieved, generating high-fidelity 3D Gaussians using these optimization-based methods still requires several minutes. TriplaneGaussian [63] introduces Gaussian splatting into the framework of LRM. This method starts by predicting Gaussian centers as point clouds and then projects them onto a triplane for other features. Nonetheless, the number of Gaussians and the resolution of the triplane are still limited, affecting the quality of the generated Gaussians. Splatter image [47] proposes to predict 3D Gaussians as pixels on the output feature map using U-Net from single-view images. This approach mainly focuses on single-view or two-view scenarios, limiting its generalization to large-scale datasets. Similarly, PixelSplat [3] predicts Gaussian parameters for each pixel of two posed images from scene datasets. We design a 4-view reconstruction model combined with existing multi-view diffusion models for general text or image to high-fidelity 3D object generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 602, + 359, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 602, + 359, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 602, + 359, + 616 + ], + "type": "text", + "content": "3 Large Multi-View Gaussian Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "We first provide the background information on Gaussian splatting and multiview diffusion models (Section 3.1). Then we introduce our high-resolution 3D content generation framework (Section 3.2), where the core part is an asymmet" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "text", + "content": "Tang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 118, + 479, + 190 + ], + "blocks": [ + { + "bbox": [ + 134, + 118, + 479, + 190 + ], + "lines": [ + { + "bbox": [ + 134, + 118, + 479, + 190 + ], + "spans": [ + { + "bbox": [ + 134, + 118, + 479, + 190 + ], + "type": "image", + "image_path": "2b2dd69bcfb4a2fc87443d3049e3c0080251648237d0506cf8412a84c787dd5d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 200, + 480, + 232 + ], + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 232 + ], + "type": "text", + "content": "Fig. 2: Pipeline. Our model is trained to reconstruct 3D Gaussians from multi-view images, which can be synthesized by off-the-shelf models [45,52] at inference time from only text, or only image, or both input. Polygonal meshes can be extracted optionally." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 257, + 479, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 479, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 479, + 317 + ], + "type": "text", + "content": "ric U-Net backbone to predict and fuse 3D Gaussians from multi-view images (Section 3.3). We design careful data augmentation and training pipeline to enhance robustness and stability (Section 3.4). Finally, we describe an effective method for smooth textured mesh extraction from the generated 3D Gaussians (Section 3.5)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 336, + 228, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 336, + 228, + 346 + ], + "spans": [ + { + "bbox": [ + 131, + 336, + 228, + 346 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "type": "text", + "content": "Gaussian Splatting. As introduced in [17], Gaussian splatting employs a collection of 3D Gaussians to represent 3D data. Specifically, each Gaussian is defined by a center " + }, + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "type": "text", + "content": ", a scaling factor " + }, + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "type": "text", + "content": ", and a rotation quaternion " + }, + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "type": "text", + "content": ". Additionally, an opacity value " + }, + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "type": "text", + "content": " and a color feature " + }, + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "type": "text", + "content": " are maintained for rendering, where spherical harmonics can be used to model view-dependent effects. These parameters can be collectively denoted by " + }, + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "type": "inline_equation", + "content": "\\Theta_i = \\{\\mathbf{x}_i, \\mathbf{s}_i, \\mathbf{q}_i, \\alpha_i, \\mathbf{c}_i\\}" + }, + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "type": "text", + "content": " representing the parameters for the " + }, + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "type": "text", + "content": "-th Gaussian. Rendering of the 3D Gaussians involves projecting them onto the image plane as 2D Gaussians and performing alpha composition for each pixel in front-to-back depth order, thereby determining the final color and alpha." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 494, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 590 + ], + "type": "text", + "content": "Multi-View Diffusion Models. Original 2D diffusion models [40,42] primarily focus on generating single-view images and do not support 3D viewpoint manipulation. Recently, several methods [20,27,44,45,52] propose to fine-tune multi-view diffusion models on 3D datasets to incorporate camera poses as an additional input. These approaches enable the creation of multi-view images of the same object, either from a text prompt or a single-view image. However, due to the absence of an actual 3D model, inconsistencies may still occur across the generated views." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 609, + 257, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 257, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 257, + 620 + ], + "type": "text", + "content": "3.2 Overall Framework" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 666 + ], + "type": "text", + "content": "As illustrated in Figure 2, we adopt a two-step 3D generation pipeline at inference. Firstly, we take advantage of off-the-shelf text or image to multi-view diffusion models to generate multi-view images. Specifically, we adopt MVDream [45]" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 447, + 102 + ], + "type": "text", + "content": "Large Multi-View Gaussian Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 478, + 276 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 478, + 276 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 478, + 276 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 478, + 276 + ], + "type": "image", + "image_path": "99cf58200fd2f09fa6de9c08debfde0f750783dead6723aed4515a9158f7bc81.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 284, + 482, + 340 + ], + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 340 + ], + "type": "text", + "content": "Fig. 3: Architecture of LGM. Our network adopts an asymmetric U-Net with cross-view self-attention. We take four images with camera ray embeddings as the input, and output four feature maps of 14 channels (visualized using the RGB and opacity channels) which are interpreted as 3D Gaussians. The fused Gaussians are then rendered at novel views and supervised with ground truth images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 361, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 458 + ], + "type": "text", + "content": "for text input and ImageDream [52] for image (and optionally text) input. Both models are designed to generate multi-view images at four orthogonal azimuths and a fixed elevation. In the second step, we use a U-Net based model to predict 3D Gaussians from these sparse view images. Specifically, our model is trained to take four images with camera pose embeddings as input and predict four sets of Gaussians, which are fused to form the final 3D Gaussians. The generated Gaussians can be optionally converted to polygonal meshes using an extra conversion step, which is more suitable for downstream tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 474, + 348, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 474, + 348, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 474, + 348, + 487 + ], + "type": "text", + "content": "3.3 Asymmetric U-Net for 3D Gaussians" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 493, + 482, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 482, + 565 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 482, + 565 + ], + "type": "text", + "content": "At the core of our framework is an asymmetric U-Net to predict and fuse Gaussians from multi-view images. The network architecture is shown in Figure 3. We take four images and corresponding camera poses as the input. Following previous works [58], we use the Plücker ray embedding to densely encode the camera poses. The RGB value and ray embedding are concatenated into a 9-channel feature map as the input to the first layer:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 260, + 573, + 481, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 573, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 260, + 573, + 481, + 586 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {i} = \\left\\{\\mathbf {c} _ {i}, \\mathbf {o} _ {i} \\times \\mathbf {d} _ {i}, \\mathbf {d} _ {i} \\right\\} \\tag {1}", + "image_path": "6127fead47e97397f783ae2f9239edc94e763ef53c27793ae647f429e90c8711.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "text", + "content": " is the input feature for pixel " + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "text", + "content": " is the RGB value, " + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_i" + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "text", + "content": " is the ray direction, and " + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_i" + }, + { + "bbox": [ + 130, + 594, + 480, + 616 + ], + "type": "text", + "content": " is the ray origin." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "The U-Net is built with residual layers [13] and self-attention layers [51] similar to previous works [14, 31, 47]. We only add self-attention at deeper layers where the feature map resolution is down-sampled to save memory. To propagate information across multiple views, we flatten the four image features and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "text", + "content": "Tang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "concatenate them before applying self-attention, similar to previous multi-view diffusion models [45, 52]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "content": "Each pixel of the output feature map is treated as a 3D Gaussian inspired by splatter image [47]. Differently, our U-Net is designed to be asymmetric with a smaller output resolution compared to input, which allows us to use higher resolution input images and limit the number of output Gaussians. We discard the depth prediction required by explicit ray-wise camera projection in [47]. The output feature map contains 14 channels corresponding to the original attributes of each Gaussian " + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\Theta_{i}" + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "content": ". To stabilize the training, we choose some different activation functions compared to the original Gaussian Splatting [17]. We clamp the predicted positions " + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "inline_equation", + "content": "[-1,1]^3" + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "content": ", and multiply the softplus-activated scales " + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 130, + 140, + 482, + 308 + ], + "type": "text", + "content": " with 0.1, such that the generated Gaussians at the beginning of training is close to the scene center. For each input view, the output feature map is transformed into a set of Gaussians. We simply concatenate these Gaussians from all four views as the final 3D Gaussians, which are used to render images at novel views for supervision." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 322, + 244, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 244, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 244, + 335 + ], + "type": "text", + "content": "3.4 Robust Training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 339, + 479, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 479, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 479, + 387 + ], + "type": "text", + "content": "Data Augmentation. We use multi-view images rendered from the Objaverse [12] dataset for training. However, at inference, we use synthesized multiview images by diffusion models [45, 52]. To mitigate the domain gap, we design two types of data augmentation for more robust training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 387, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 470 + ], + "type": "text", + "content": "Grid Distortion. Synthesizing 3D consistent multi-view images using 2D diffusion models has been explored by many works [25,44,45,52]. However, since there is no underlying 3D representation, the generated multi-view images often suffer from subtle inconsistency across different views. We try to simulate such inconsistency using grid distortion. Except for the first input view, the other three input views are randomly distorted with a random grid during training. This makes the model more robust to inconsistent multi-view input images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 471, + 482, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 531 + ], + "type": "text", + "content": "Orbital Camera Jitter. Another problem is that the synthesized multi-view images may not accurately follow the given camera poses. Following [15], we always normalized the camera poses at each training step such that the first view's camera pose is fixed. We therefore apply camera jitter to the last three input views during training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "text", + "content": "Loss Function. To supervise the concatenated Gaussians, we use the differentiable renderer implementation from [17] to render them. At each training step, we render the RGB image and alpha image of eight views, including four input views and four novel views. Following [15], we apply mean square error loss and VGG-based LPIPS loss [61] to the RGB image:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 212, + 609, + 481, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 609, + 481, + 626 + ], + "spans": [ + { + "bbox": [ + 212, + 609, + 481, + 626 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r g b}} = \\mathcal {L} _ {\\mathrm {M S E}} \\left(I _ {\\mathrm {r g b}}, I _ {\\mathrm {r g b}} ^ {\\mathrm {G T}}\\right) + \\mathcal {L} _ {\\mathrm {L P I P S}} \\left(I _ {\\mathrm {r g b}}, I _ {\\mathrm {r g b}} ^ {\\mathrm {G T}}\\right) \\tag {2}", + "image_path": "0d5094ae212a3c56afd8d1c7f364cb50153026dbde4b810165d09bb1b6687624.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 481, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 653 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 653 + ], + "type": "text", + "content": "We further apply mean square error loss on the alpha image for faster convergence of the shape:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 261, + 652, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 652, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 261, + 652, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\alpha} = \\mathcal {L} _ {\\mathrm {M S E}} (I _ {\\alpha}, I _ {\\alpha} ^ {\\mathrm {G T}}) \\tag {3}", + "image_path": "f37761d19ac4d1ce854402130745c5091a05f7cfc080436858dc26318e06255e.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 447, + 102 + ], + "type": "text", + "content": "Large Multi-View Gaussian Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 478, + 296 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 478, + 296 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 478, + 296 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 478, + 296 + ], + "type": "image", + "image_path": "48f05a3506576e0012668129ebe86e37c3e442abd332d231c543ae7d1f7838c3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 305, + 482, + 327 + ], + "lines": [ + { + "bbox": [ + 131, + 305, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 131, + 305, + 482, + 327 + ], + "type": "text", + "content": "Fig. 4: Mesh Extraction Pipeline. We implement an efficient pipeline to convert the 3D Gaussians into smooth and textured meshes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 350, + 246, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 350, + 246, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 246, + 361 + ], + "type": "text", + "content": "3.5 Mesh Extraction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 368, + 482, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 463 + ], + "type": "text", + "content": "Since polygonal meshes are still the most widely used 3D representation in downstream tasks, we hope to further extract meshes from our generated Gaussians. Previous works [48] have tried to directly convert the opacity value of 3D Gaussians into an occupancy field for mesh extraction. However, we find this method dependent on aggressive densification during the optimization of 3D Gaussians to produce smooth occupancy field. On the contrary, the generated Gaussians in our method are usually sparse and cannot produce a suitable occupancy field, leading to an unsatisfactory surface with visible holes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 464, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 583 + ], + "type": "text", + "content": "Instead, we propose a more general mesh extraction pipeline from 3D Gaussians as illustrated in Figure 4. We first train an efficient NeRF [34] using the rendered images from 3D Gaussians on-the-fly, and then convert the NeRF to polygonal meshes [49]. Marching Cubes [28] is applied to extract a coarse mesh, which is then iteratively refined together with the appearance hash grid using differentiable rendering. Finally, we bake the appearance field onto the refined mesh to extract texture images. For more details, please refer to the supplementary materials and NeRF2Mesh [49]. With adequately optimized implementation, it takes only about 1 minute to perform this Gaussians to NeRF to mesh conversion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 600, + 230, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 230, + 613 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 230, + 613 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 623, + 280, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 280, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 280, + 635 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "Datasets. We use a filtered subset of the Objaverse [12] dataset to train our model. Since there are many low-quality 3D models (e.g., partial scans, missing" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "text", + "content": "Tang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "textures) in the original Objaverse dataset, we filter the dataset by two empirical rules: (1) We examine the captions from Cap3D [30], and curate a list of words that usually appears in bad models (e.g., 'resembling', 'debris', 'frame'), which is used to filter all models whose caption includes any of these words. (2) We discard models with mostly white color after rendering, which usually indicates missing texture. These lead to a final set of around 80K 3D objects. We render the RGBA image from 100 camera views at the resolution of " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " for training and validation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 228, + 483, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 483, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 483, + 360 + ], + "type": "text", + "content": "Network Architecture. Our asymmetric U-Net model consists of 6 down blocks, 1 middle block, and 5 up blocks, with the input image at " + }, + { + "bbox": [ + 130, + 228, + 483, + 360 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 228, + 483, + 360 + ], + "type": "text", + "content": " and output Gaussian feature map at " + }, + { + "bbox": [ + 130, + 228, + 483, + 360 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 130, + 228, + 483, + 360 + ], + "type": "text", + "content": ". We use 4 input views, so the number of output Gaussians is " + }, + { + "bbox": [ + 130, + 228, + 483, + 360 + ], + "type": "inline_equation", + "content": "128 \\times 128 \\times 4 = 65, 536" + }, + { + "bbox": [ + 130, + 228, + 483, + 360 + ], + "type": "text", + "content": ". The feature channels for all blocks are [64, 128, 256, 512, 1024, 1024], [1024] and [1024, 1024, 512, 256, 128] respectively. Each block contains a series of residual layers and an optional down-sample or up-sample layer. For the last 3 of down blocks, the middle block, and the first 3 up blocks, we also insert cross-view self-attention layers after the residual layers. The final feature maps are processed by a " + }, + { + "bbox": [ + 130, + 228, + 483, + 360 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 228, + 483, + 360 + ], + "type": "text", + "content": " convolution layer to 14-channel pixel-wise Gaussian features. Following previous works [40, 47], we adopt Silu activation and group normalization for the U-Net." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 374, + 483, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 483, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 483, + 531 + ], + "type": "text", + "content": "Training. We train our model on 32 NVIDIA A100 (80G) GPUs for about 4 days. A batch size of 8 for each GPU is used, leading to an effective batch size of 256. For each batch, we randomly sample 8 camera views, with the first 4 views as the input, and all 8 views as the output for supervision. Similar to LRM [15], we transform the cameras of each batch such that the first input view is always the front view with an identity rotation matrix and fixed translation. The input images are assumed to have a white background. The output 3D Gaussians are rendered at " + }, + { + "bbox": [ + 130, + 374, + 483, + 531 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 374, + 483, + 531 + ], + "type": "text", + "content": " resolution for mean square error loss. We resize the images to " + }, + { + "bbox": [ + 130, + 374, + 483, + 531 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 374, + 483, + 531 + ], + "type": "text", + "content": " for LPIPS loss to save memory. The AdamW [29] optimizer is adopted with the learning rate of " + }, + { + "bbox": [ + 130, + 374, + 483, + 531 + ], + "type": "inline_equation", + "content": "4 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 374, + 483, + 531 + ], + "type": "text", + "content": ", weight decay of 0.05, and betas of (0.9, 0.95). The learning rate is cosine annealed to 0 during the training. We clip the gradient with a maximum norm of 1.0. The probability for grid distortion and camera jitter is set to " + }, + { + "bbox": [ + 130, + 374, + 483, + 531 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 374, + 483, + 531 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Inference. Our whole pipeline, including two multi-view diffusion models, takes only about 10 GB of GPU memory for inference, which is friendly for deployment. For the multi-view diffusion models, we use a guidance scale of 5 for ImageDream [52] and 7.5 for MVDream [45] following the original paper. The number of diffusion steps is set to 30 using the DDIM [46] scheduler. The camera elevation is fixed to 0, and azimuths to " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "[0,90,180,270]" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " degree for the four generated views. For ImageDream [52], the text prompt is always left empty so the only input is a single-view image. Since the images generated by MVDream may contain various backgrounds, we apply background removal [38] and use white background." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 447, + 102 + ], + "type": "text", + "content": "Large Multi-View Gaussian Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 118, + 475, + 595 + ], + "blocks": [ + { + "bbox": [ + 141, + 118, + 475, + 595 + ], + "lines": [ + { + "bbox": [ + 141, + 118, + 475, + 595 + ], + "spans": [ + { + "bbox": [ + 141, + 118, + 475, + 595 + ], + "type": "image", + "image_path": "207f8e44de316b9bd214edeee9e026a5f962a23589db49ab0e0c67bb91b78f66.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 606, + 482, + 630 + ], + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 630 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 630 + ], + "type": "text", + "content": "Fig. 5: Comparisons of generated 3D Gaussians for image-to-3D. Our method generates Gaussian splatting with better visual quality on various challenging images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "text", + "content": "Tang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 118, + 170, + 167 + ], + "blocks": [ + { + "bbox": [ + 137, + 118, + 170, + 167 + ], + "lines": [ + { + "bbox": [ + 137, + 118, + 170, + 167 + ], + "spans": [ + { + "bbox": [ + 137, + 118, + 170, + 167 + ], + "type": "image", + "image_path": "d1a5a59111478c47a7582feaea089a4c299d1fc6a03331f6f80a0b2793675188.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 136, + 173, + 174, + 215 + ], + "blocks": [ + { + "bbox": [ + 136, + 173, + 174, + 215 + ], + "lines": [ + { + "bbox": [ + 136, + 173, + 174, + 215 + ], + "spans": [ + { + "bbox": [ + 136, + 173, + 174, + 215 + ], + "type": "image", + "image_path": "0a2ff86c817deb658e55878be19c8b1300536f611ea8073fa77fec0d2ee85679.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 140, + 218, + 171, + 240 + ], + "blocks": [ + { + "bbox": [ + 140, + 218, + 171, + 240 + ], + "lines": [ + { + "bbox": [ + 140, + 218, + 171, + 240 + ], + "spans": [ + { + "bbox": [ + 140, + 218, + 171, + 240 + ], + "type": "image", + "image_path": "8120ff055e1af89ec5fb3f6e8d22a600fabbccaa8a294ed187b7234af669dcbf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 146, + 242, + 167, + 264 + ], + "blocks": [ + { + "bbox": [ + 146, + 242, + 167, + 264 + ], + "lines": [ + { + "bbox": [ + 146, + 242, + 167, + 264 + ], + "spans": [ + { + "bbox": [ + 146, + 242, + 167, + 264 + ], + "type": "image", + "image_path": "172e67b0c384be0bc723ccf57a5146e0546808a021ff9fce7bb9d631a5d67640.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 143, + 273, + 167, + 316 + ], + "blocks": [ + { + "bbox": [ + 143, + 273, + 167, + 316 + ], + "lines": [ + { + "bbox": [ + 143, + 273, + 167, + 316 + ], + "spans": [ + { + "bbox": [ + 143, + 273, + 167, + 316 + ], + "type": "image", + "image_path": "8d308a89c4d341fafabaf56c50d170f6b917929d857ae779220bfc21e809dbe5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 323, + 166, + 334 + ], + "lines": [ + { + "bbox": [ + 146, + 323, + 166, + 334 + ], + "spans": [ + { + "bbox": [ + 146, + 323, + 166, + 334 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 189, + 118, + 227, + 166 + ], + "blocks": [ + { + "bbox": [ + 189, + 118, + 227, + 166 + ], + "lines": [ + { + "bbox": [ + 189, + 118, + 227, + 166 + ], + "spans": [ + { + "bbox": [ + 189, + 118, + 227, + 166 + ], + "type": "image", + "image_path": "34500a389efec3eddb7c0d37b0eb49dc067b78bb0608ad305a1861c7f3db09e3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 186, + 172, + 227, + 216 + ], + "blocks": [ + { + "bbox": [ + 186, + 172, + 227, + 216 + ], + "lines": [ + { + "bbox": [ + 186, + 172, + 227, + 216 + ], + "spans": [ + { + "bbox": [ + 186, + 172, + 227, + 216 + ], + "type": "image", + "image_path": "49786530e9791b6747beb2e90f9f81e32eab23c404d0ef3dc7b14e5d0f679339.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 190, + 218, + 222, + 242 + ], + "blocks": [ + { + "bbox": [ + 190, + 218, + 222, + 242 + ], + "lines": [ + { + "bbox": [ + 190, + 218, + 222, + 242 + ], + "spans": [ + { + "bbox": [ + 190, + 218, + 222, + 242 + ], + "type": "image", + "image_path": "6c8d87cf5511f934967af10bf0691b7dc1069f6460d0a7a24db56db0cffcc40f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 196, + 242, + 218, + 266 + ], + "blocks": [ + { + "bbox": [ + 196, + 242, + 218, + 266 + ], + "lines": [ + { + "bbox": [ + 196, + 242, + 218, + 266 + ], + "spans": [ + { + "bbox": [ + 196, + 242, + 218, + 266 + ], + "type": "image", + "image_path": "b5ffcc6fe9e94a6a95139e15479fb8894a236ba87e41593259fc23380c694424.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 194, + 274, + 218, + 318 + ], + "blocks": [ + { + "bbox": [ + 194, + 274, + 218, + 318 + ], + "lines": [ + { + "bbox": [ + 194, + 274, + 218, + 318 + ], + "spans": [ + { + "bbox": [ + 194, + 274, + 218, + 318 + ], + "type": "image", + "image_path": "f5d12265e95ac72291c29a99d7d87c74ee1fe72ef5092916ce39b697008c65fd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 344, + 480, + 368 + ], + "lines": [ + { + "bbox": [ + 131, + 344, + 480, + 368 + ], + "spans": [ + { + "bbox": [ + 131, + 344, + 480, + 368 + ], + "type": "text", + "content": "Fig. 6: Comparisons with LRM for image-to-3D. We compare our method with available results from LRM [15]." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 249, + 118, + 271, + 166 + ], + "blocks": [ + { + "bbox": [ + 249, + 118, + 271, + 166 + ], + "lines": [ + { + "bbox": [ + 249, + 118, + 271, + 166 + ], + "spans": [ + { + "bbox": [ + 249, + 118, + 271, + 166 + ], + "type": "image", + "image_path": "d569a9ac3d51fb110440c6f3a4f323f09464a3c1f7971e909fd53d8f5b9cf863.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 244, + 170, + 274, + 216 + ], + "blocks": [ + { + "bbox": [ + 244, + 170, + 274, + 216 + ], + "lines": [ + { + "bbox": [ + 244, + 170, + 274, + 216 + ], + "spans": [ + { + "bbox": [ + 244, + 170, + 274, + 216 + ], + "type": "image", + "image_path": "eba24e0c95905d0172d6a0cebef90d4154c9815d8f6ba2bb43cc128cc1401f57.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 243, + 218, + 271, + 242 + ], + "blocks": [ + { + "bbox": [ + 243, + 218, + 271, + 242 + ], + "lines": [ + { + "bbox": [ + 243, + 218, + 271, + 242 + ], + "spans": [ + { + "bbox": [ + 243, + 218, + 271, + 242 + ], + "type": "image", + "image_path": "68a3b88d4fce57a20c0f08bde490a9b59a4542c9fcc1cedbfd322667f1461b22.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 252, + 242, + 271, + 267 + ], + "blocks": [ + { + "bbox": [ + 252, + 242, + 271, + 267 + ], + "lines": [ + { + "bbox": [ + 252, + 242, + 271, + 267 + ], + "spans": [ + { + "bbox": [ + 252, + 242, + 271, + 267 + ], + "type": "image", + "image_path": "b7606e5432514ac4b08dbfe921f9d29582c4cb7ac260b6af5787a62605611fa7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 243, + 274, + 274, + 319 + ], + "blocks": [ + { + "bbox": [ + 243, + 274, + 274, + 319 + ], + "lines": [ + { + "bbox": [ + 243, + 274, + 274, + 319 + ], + "spans": [ + { + "bbox": [ + 243, + 274, + 274, + 319 + ], + "type": "image", + "image_path": "c5a747d6f0ba8f4a8706fb9b0d5f9ff9a6c5e5b7f3f78c5fb287d630fdcb3604.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 324, + 267, + 333 + ], + "lines": [ + { + "bbox": [ + 250, + 324, + 267, + 333 + ], + "spans": [ + { + "bbox": [ + 250, + 324, + 267, + 333 + ], + "type": "text", + "content": "LRM" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 286, + 119, + 330, + 166 + ], + "blocks": [ + { + "bbox": [ + 286, + 119, + 330, + 166 + ], + "lines": [ + { + "bbox": [ + 286, + 119, + 330, + 166 + ], + "spans": [ + { + "bbox": [ + 286, + 119, + 330, + 166 + ], + "type": "image", + "image_path": "b3b1dc7a22505225cb513c359b3753bf132f38c7a6aff07ebe3093032b5628a0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 286, + 170, + 329, + 216 + ], + "blocks": [ + { + "bbox": [ + 286, + 170, + 329, + 216 + ], + "lines": [ + { + "bbox": [ + 286, + 170, + 329, + 216 + ], + "spans": [ + { + "bbox": [ + 286, + 170, + 329, + 216 + ], + "type": "image", + "image_path": "d281c646bdf23ac47d7b71ed34a75d62d4c13b56480f0480a67e73497ab48670.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 294, + 218, + 322, + 242 + ], + "blocks": [ + { + "bbox": [ + 294, + 218, + 322, + 242 + ], + "lines": [ + { + "bbox": [ + 294, + 218, + 322, + 242 + ], + "spans": [ + { + "bbox": [ + 294, + 218, + 322, + 242 + ], + "type": "image", + "image_path": "7dd866c9daba5c23ae05ff3ececabcdc63dfe02e6118866d8e6be92d1fad3e84.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 299, + 242, + 318, + 269 + ], + "blocks": [ + { + "bbox": [ + 299, + 242, + 318, + 269 + ], + "lines": [ + { + "bbox": [ + 299, + 242, + 318, + 269 + ], + "spans": [ + { + "bbox": [ + 299, + 242, + 318, + 269 + ], + "type": "image", + "image_path": "ffa30161d78479890d788e5d14844be60c36da7f5e316f67f935dd976382ada5.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 294, + 273, + 319, + 320 + ], + "blocks": [ + { + "bbox": [ + 294, + 273, + 319, + 320 + ], + "lines": [ + { + "bbox": [ + 294, + 273, + 319, + 320 + ], + "spans": [ + { + "bbox": [ + 294, + 273, + 319, + 320 + ], + "type": "image", + "image_path": "9ea7f3454f4a43859f7f421247b3a60f614e2f0a9e7e5e17564c7896dd155c06.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 340, + 119, + 375, + 167 + ], + "blocks": [ + { + "bbox": [ + 340, + 119, + 375, + 167 + ], + "lines": [ + { + "bbox": [ + 340, + 119, + 375, + 167 + ], + "spans": [ + { + "bbox": [ + 340, + 119, + 375, + 167 + ], + "type": "image", + "image_path": "3850f4fe566f5646d7e33940b17086dfffa16d1b8b2a44129dd3cc0f93ebb7d0.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 338, + 171, + 378, + 216 + ], + "blocks": [ + { + "bbox": [ + 338, + 171, + 378, + 216 + ], + "lines": [ + { + "bbox": [ + 338, + 171, + 378, + 216 + ], + "spans": [ + { + "bbox": [ + 338, + 171, + 378, + 216 + ], + "type": "image", + "image_path": "3299a8beeddc28f0f1051b468917db8f5be6ed3471609a790f56959702858aab.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 341, + 220, + 372, + 244 + ], + "blocks": [ + { + "bbox": [ + 341, + 220, + 372, + 244 + ], + "lines": [ + { + "bbox": [ + 341, + 220, + 372, + 244 + ], + "spans": [ + { + "bbox": [ + 341, + 220, + 372, + 244 + ], + "type": "image", + "image_path": "6d50475a10baece5c0b09efe8a28b9965a6d3369be14adc1b98eaaa512fac180.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 347, + 245, + 369, + 268 + ], + "blocks": [ + { + "bbox": [ + 347, + 245, + 369, + 268 + ], + "lines": [ + { + "bbox": [ + 347, + 245, + 369, + 268 + ], + "spans": [ + { + "bbox": [ + 347, + 245, + 369, + 268 + ], + "type": "image", + "image_path": "da5aa86a658340a30e311e8ac41dd2c9a90a96e034cc51841fb2f30bfe2eb83e.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 342, + 272, + 368, + 319 + ], + "blocks": [ + { + "bbox": [ + 342, + 272, + 368, + 319 + ], + "lines": [ + { + "bbox": [ + 342, + 272, + 368, + 319 + ], + "spans": [ + { + "bbox": [ + 342, + 272, + 368, + 319 + ], + "type": "image", + "image_path": "cb169e6f4ac1b45fe1b4b0ab8d9202bf96b236ca817d2309db26ba3b1b44b1ae.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 391, + 118, + 424, + 168 + ], + "blocks": [ + { + "bbox": [ + 391, + 118, + 424, + 168 + ], + "lines": [ + { + "bbox": [ + 391, + 118, + 424, + 168 + ], + "spans": [ + { + "bbox": [ + 391, + 118, + 424, + 168 + ], + "type": "image", + "image_path": "270fc7c81c638b4904dd517e3bed84c734b3271508e1d94942e80b1c45a6a0e3.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 391, + 170, + 427, + 217 + ], + "blocks": [ + { + "bbox": [ + 391, + 170, + 427, + 217 + ], + "lines": [ + { + "bbox": [ + 391, + 170, + 427, + 217 + ], + "spans": [ + { + "bbox": [ + 391, + 170, + 427, + 217 + ], + "type": "image", + "image_path": "1c4c31b97244152477f02bef8b2489349fe63aba58ff42165946f94a7800d80e.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 389, + 220, + 424, + 244 + ], + "blocks": [ + { + "bbox": [ + 389, + 220, + 424, + 244 + ], + "lines": [ + { + "bbox": [ + 389, + 220, + 424, + 244 + ], + "spans": [ + { + "bbox": [ + 389, + 220, + 424, + 244 + ], + "type": "image", + "image_path": "b1b889c3e523dfd4f43248f889470b7cf04c42295e242e290b4e6bf82fc30235.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 401, + 245, + 418, + 269 + ], + "blocks": [ + { + "bbox": [ + 401, + 245, + 418, + 269 + ], + "lines": [ + { + "bbox": [ + 401, + 245, + 418, + 269 + ], + "spans": [ + { + "bbox": [ + 401, + 245, + 418, + 269 + ], + "type": "image", + "image_path": "27a727f39bea55e923af42be8bdf524c0c4c5266fefb7aee6c890112b5510073.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 399, + 273, + 417, + 319 + ], + "blocks": [ + { + "bbox": [ + 399, + 273, + 417, + 319 + ], + "lines": [ + { + "bbox": [ + 399, + 273, + 417, + 319 + ], + "spans": [ + { + "bbox": [ + 399, + 273, + 417, + 319 + ], + "type": "image", + "image_path": "9c2cbaac60d6606d3a874059b71925d6804ceec8ba4d8c63973dd6280ece94d8.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 324, + 418, + 332 + ], + "lines": [ + { + "bbox": [ + 400, + 324, + 418, + 332 + ], + "spans": [ + { + "bbox": [ + 400, + 324, + 418, + 332 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 439, + 119, + 472, + 168 + ], + "blocks": [ + { + "bbox": [ + 439, + 119, + 472, + 168 + ], + "lines": [ + { + "bbox": [ + 439, + 119, + 472, + 168 + ], + "spans": [ + { + "bbox": [ + 439, + 119, + 472, + 168 + ], + "type": "image", + "image_path": "14da764935bab1b1c326c5e1a0536a23785d1ed620b6ce3363f97b8719143daf.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 438, + 171, + 479, + 217 + ], + "blocks": [ + { + "bbox": [ + 438, + 171, + 479, + 217 + ], + "lines": [ + { + "bbox": [ + 438, + 171, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 438, + 171, + 479, + 217 + ], + "type": "image", + "image_path": "502ec699069d8914328e01b84e27c90c800710c7165e319492a702f4b83aa293.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 441, + 220, + 474, + 244 + ], + "blocks": [ + { + "bbox": [ + 441, + 220, + 474, + 244 + ], + "lines": [ + { + "bbox": [ + 441, + 220, + 474, + 244 + ], + "spans": [ + { + "bbox": [ + 441, + 220, + 474, + 244 + ], + "type": "image", + "image_path": "b3d6f0b8dfe2806fd8a26cde56e477a30ce6c0baf071bfa9475d67b728b455ab.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 445, + 245, + 468, + 269 + ], + "blocks": [ + { + "bbox": [ + 445, + 245, + 468, + 269 + ], + "lines": [ + { + "bbox": [ + 445, + 245, + 468, + 269 + ], + "spans": [ + { + "bbox": [ + 445, + 245, + 468, + 269 + ], + "type": "image", + "image_path": "540344f95e609d526187b72090741e3495e55624b64e642c95238aaa2c9a16df.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 443, + 273, + 469, + 320 + ], + "blocks": [ + { + "bbox": [ + 443, + 273, + 469, + 320 + ], + "lines": [ + { + "bbox": [ + 443, + 273, + 469, + 320 + ], + "spans": [ + { + "bbox": [ + 443, + 273, + 469, + 320 + ], + "type": "image", + "image_path": "c0a61109034a2414f0ce8210e78f0d6a3e931b6949c29051e6562b8dbc275bd7.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 132, + 392, + 285, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 285, + 404 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 285, + 404 + ], + "type": "text", + "content": "4.2 Qualitative Comparisons" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 525 + ], + "type": "text", + "content": "Image-to-3D. We first compare against recent methods [48,63] that are capable of generating 3D Gaussians. Figure 5 shows images rendered from the generated 3D Gaussians for comparison. The 3D Gaussians produced by our method have better visual quality and effectively preserve the content from the input view. Our high-resolution 3D Gaussians can be transformed into smooth textured meshes with minimal loss of quality in most cases. We also compare our results against LRM [15] using the available videos from their website in Figure 6. Specifically, our multi-view setting successfully mitigates the issue of blurry back views and flat geometry, resulting in enhanced detail even in unseen views." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 130, + 547, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 547, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 547, + 482, + 596 + ], + "type": "text", + "content": "Text-to-3D. We then compare with recent methods [16,48] on text-to-3D tasks. We observe an enhanced quality and efficiency in our method, generating more realistic 3D objects, as illustrated in Figure 7. Due to the multi-view diffusion models, our model is also free from multi-face problems." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Diversity. Notably, our pipeline exhibits high diversity in 3D generation, owing to the capability of multi-view diffusion model [45, 52]. As shown in Figure 8, with different random seeds, we can generate a variety of feasible objects from the same ambiguous text prompt or single-view image." + } + ] + } + ], + "index": 44 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 447, + 102 + ], + "type": "text", + "content": "Large Multi-View Gaussian Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 118, + 478, + 369 + ], + "blocks": [ + { + "bbox": [ + 135, + 118, + 478, + 369 + ], + "lines": [ + { + "bbox": [ + 135, + 118, + 478, + 369 + ], + "spans": [ + { + "bbox": [ + 135, + 118, + 478, + 369 + ], + "type": "image", + "image_path": "0b47b6ea8fd9aeefcf32c5fea4a4776389ead96885f02efcde5ebf4f9942680e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 377, + 482, + 401 + ], + "lines": [ + { + "bbox": [ + 132, + 377, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 132, + 377, + 482, + 401 + ], + "type": "text", + "content": "Fig. 7: Comparisons of generated 3D models for text-to-3D. Our method achieves better text alignment and visual quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 411, + 478, + 564 + ], + "blocks": [ + { + "bbox": [ + 138, + 411, + 478, + 564 + ], + "lines": [ + { + "bbox": [ + 138, + 411, + 478, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 478, + 564 + ], + "type": "image", + "image_path": "861673706a4a5e8038b8fcc538c065b8690f3480c2b0fb276bf5802dec9b717c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 574, + 482, + 597 + ], + "lines": [ + { + "bbox": [ + 132, + 574, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 482, + 597 + ], + "type": "text", + "content": "Fig. 8: Diversity of our 3D generation. We can generate diverse 3D models given an ambiguous text description or single-view image." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 620, + 293, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 293, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 293, + 633 + ], + "type": "text", + "content": "4.3 Quantitative Comparisons" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "We majorly conduct a user study to quantitatively evaluate our image-to-3D Gaussians generation performance. For a collection of 30 images, we render 360-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "text", + "content": "Tang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 114, + 474, + 161 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 474, + 161 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 474, + 161 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 474, + 161 + ], + "type": "table", + "html": "
Image Consistency ↑Overall Quality ↑
DreamGaussian [48]2.301.98
TriplaneGaussian [63]3.022.67
LGM (Ours)4.183.95
", + "image_path": "31123417aec3179d34cd435d3371485b08b28edbab960c2910038d3c29036ebe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 169, + 480, + 192 + ], + "lines": [ + { + "bbox": [ + 132, + 169, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 169, + 480, + 192 + ], + "type": "text", + "content": "Table 1: User Study on the quality of generated 3D Gaussians for image-to-3D tasks. The rating is of scale 1-5, the higher the better." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 138, + 215, + 312, + 401 + ], + "blocks": [ + { + "bbox": [ + 138, + 215, + 312, + 401 + ], + "lines": [ + { + "bbox": [ + 138, + 215, + 312, + 401 + ], + "spans": [ + { + "bbox": [ + 138, + 215, + 312, + 401 + ], + "type": "image", + "image_path": "d398eff655e1caec482ce2711b3ce1f011d5dece702d3ac5a9a270c41466a381.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 139, + 411, + 473, + 423 + ], + "lines": [ + { + "bbox": [ + 139, + 411, + 473, + 423 + ], + "spans": [ + { + "bbox": [ + 139, + 411, + 473, + 423 + ], + "type": "text", + "content": "Fig. 9: Ablation Study. We carry out ablation study on designs of our method." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 215, + 480, + 400 + ], + "blocks": [ + { + "bbox": [ + 317, + 215, + 480, + 400 + ], + "lines": [ + { + "bbox": [ + 317, + 215, + 480, + 400 + ], + "spans": [ + { + "bbox": [ + 317, + 215, + 480, + 400 + ], + "type": "image", + "image_path": "949dfc8040fab91d32a1e3e9eda93b980e9832427876bf53e48a12590563ce37.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 446, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 533 + ], + "type": "text", + "content": "degree rotating videos of the 3D Gaussians generated from DreamGaussain [48] (only the first stage), TriplaneGaussian [63], and ours. There are in total 90 videos for evaluation in our user study. Each volunteer is shown 30 samples from mixed random methods, and asked to rate in two aspects: image consistency and overall model quality. We collect results from 20 volunteers and get 600 valid scores in total. As shown in Table 1, our method is preferred as it aligns with the original image content and shows better overall quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 548, + 239, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 239, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 239, + 562 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Number of Views. We train an image-to-3D model with only one input views similar to splatter image [47], i.e., without the multi-view generation step. The U-Net takes the single input view as input with self-attention, and outputs Gaussian features as in our multi-view model. To compensate the number of Gaussians, we predict two Gaussians for each pixel of the output feature maps, leading to " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "128 \\times 128 \\times 2 = 32" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", 768 Gaussians. As illustrated in the top-left part of Figure 9, the single-view model can reconstruct faithful front-view, but fails to distinguish the back view and results in blurriness. This is as expected since the regressive" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 447, + 102 + ], + "type": "text", + "content": "Large Multi-View Gaussian Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "U-Net is more suitable for reconstruction tasks, and it's hard to generalize to large datasets in our experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 164, + 481, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 164, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 164, + 481, + 247 + ], + "type": "text", + "content": "Data Augmentation. We train a smaller model with or without applying data augmentation to validate its effectiveness. Although we observe a lower training loss for the model without data augmentation, the domain gap during inference leads to more floaters and worse geometry as shown in the bottom-left part of Figure 9. The model with data augmentation strategy can better correct the 3D inconsistency and inaccurate camera poses in the generated multi-view images, leading to fewer floaters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 271, + 481, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 367 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 367 + ], + "type": "text", + "content": "Training Resolution. Lastly, we train a model with a fewer number of Gaussians and smaller rendering resolution as in the right part of Figure 9. We remove the last up block of the U-Net so the number of output Gaussians is " + }, + { + "bbox": [ + 132, + 271, + 481, + 367 + ], + "type": "inline_equation", + "content": "64 \\times 64 \\times 4 = 16,384" + }, + { + "bbox": [ + 132, + 271, + 481, + 367 + ], + "type": "text", + "content": ", and we render it at " + }, + { + "bbox": [ + 132, + 271, + 481, + 367 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 132, + 271, + 481, + 367 + ], + "type": "text", + "content": " for supervision. The model can still converge and successfully reconstruct 3D Gaussians, but the details are worse compared to the " + }, + { + "bbox": [ + 132, + 271, + 481, + 367 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 132, + 271, + 481, + 367 + ], + "type": "text", + "content": " input multi-view images. In contrast, our large resolution model at " + }, + { + "bbox": [ + 132, + 271, + 481, + 367 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 132, + 271, + 481, + 367 + ], + "type": "text", + "content": " can capture better details and generate Gaussians with higher resolution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 392, + 217, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 217, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 217, + 403 + ], + "type": "text", + "content": "4.5 Limitations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 417, + 481, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 417, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 417, + 481, + 514 + ], + "type": "text", + "content": "Since our model is essentially a multi-view reconstruction model, the 3D generation quality highly depends on the quality of four input views. However, current multi-view diffusion models [45, 52] are far from perfect: (1) There can be 3D inconsistency which misleads the reconstruction model to generate floaters. (2) The resolution of synthesized multi-view images is restricted to " + }, + { + "bbox": [ + 132, + 417, + 481, + 514 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 132, + 417, + 481, + 514 + ], + "type": "text", + "content": ". (3) ImageDream [52] also fails to handle input image with a large elevation angle. We expect these limitations can be mitigated with better multi-view diffusion models such as Zero-1-to-3++ [44] in future works." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 537, + 219, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 219, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 219, + 551 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 569, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 481, + 665 + ], + "type": "text", + "content": "In this work, we present a large multi-view Gaussian model for high-resolution 3D content generation. Our model, distinct from previous methods reliant on NeRF and transformers, employs Gaussian splatting and U-Net to address the challenges of high memory requirements and low-resolution training. Additionally, we explore data augmentation for better robustness, and introduce a mesh extraction algorithm for the generated 3D Gaussians. Our approach achieves both high-resolution and high-efficiency for 3D objects generation, proving its versatility and applicability in various contexts." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "text", + "content": "Tang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "Acknowledgements. This work is supported by the Sichuan Science and Technology Program (2023YFSY0008), National Natural Science Foundation of China (61632003, 61375022, 61403005), Grant SCITLAB-20017 of Intelligent Terminal Key Laboratory of SiChuan Province, Beijing Advanced Innovation Center for Intelligent Robots and Systems (2018IRS11), and PEK-SenseTime Joint Laboratory of Machine Vision. This project is also funded by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOE-T2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 255, + 197, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 255, + 197, + 266 + ], + "spans": [ + { + "bbox": [ + 133, + 255, + 197, + 266 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 280, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 280, + 480, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 280, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 138, + 280, + 480, + 301 + ], + "type": "text", + "content": "1. Cao, Z., Hong, F., Wu, T., Pan, L., Liu, Z.: Large-vocabulary 3d diffusion model with transformer. arXiv preprint arXiv:2309.07920 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 303, + 480, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 303, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 303, + 480, + 334 + ], + "type": "text", + "content": "2. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., Mello, S.D., Gallo, O., Guibas, L., Tremblay, J., Khamis, S., Karras, T., Wetzstein, G.: Efficient geometry-aware 3D generative adversarial networks. In: CVPR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 335, + 480, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 480, + 368 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 480, + 368 + ], + "type": "text", + "content": "3. Charatan, D., Li, S., Tagliasacchi, A., Sitzmann, V.: pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. arXiv preprint arXiv:2312.12337 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 369, + 480, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 369, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 138, + 369, + 480, + 389 + ], + "type": "text", + "content": "4. Chen, G., Wang, W.: A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 391, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 391, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 391, + 480, + 422 + ], + "type": "text", + "content": "5. Chen, H., Gu, J., Chen, A., Tian, W., Tu, Z., Liu, L., Su, H.: Single-stage diffusion nef: A unified approach to 3d generation and reconstruction. arXiv preprint arXiv:2304.06714 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 423, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 480, + 456 + ], + "type": "text", + "content": "6. Chen, R., Chen, Y., Jiao, N., Jia, K.: *Fantasia3d: Disentangling geometry* and appearance for high-quality text-to-3d content creation. arXiv preprint arXiv:2303.13873 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 456, + 480, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 488 + ], + "type": "text", + "content": "7. Chen, Y., Chen, Z., Zhang, C., Wang, F., Yang, X., Wang, Y., Cai, Z., Yang, L., Liu, H., Lin, G.: Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. arXiv preprint arXiv:2311.14521 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 489, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 489, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 489, + 480, + 521 + ], + "type": "text", + "content": "8. Chen, Y., Zhang, C., Yang, X., Cai, Z., Yu, G., Yang, L., Lin, G.: It3d: Improved text-to-3d generation with explicit view synthesis. arXiv preprint arXiv:2308.11473 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 522, + 480, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 480, + 554 + ], + "type": "text", + "content": "9. Chen, Z., Hong, F., Mei, H., Wang, G., Yang, L., Liu, Z.: Primdiffusion: Volumetric primitives diffusion for 3d human generation. arXiv preprint arXiv:2312.04559 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 555, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 576 + ], + "type": "text", + "content": "10. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 577, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 577, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 134, + 577, + 480, + 609 + ], + "type": "text", + "content": "11. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objverse-xl: A universe of " + }, + { + "bbox": [ + 134, + 577, + 480, + 609 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 134, + 577, + 480, + 609 + ], + "type": "text", + "content": " objects. arXiv preprint arXiv:2307.05663 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 610, + 480, + 643 + ], + "type": "text", + "content": "12. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., Vanderbilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objverse: A universe of annotated 3d objects. In: CVPR. pp. 13142-13153 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "text", + "content": "13. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 447, + 102 + ], + "type": "text", + "content": "Large Multi-View Gaussian Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "14. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 480, + 171 + ], + "type": "text", + "content": "15. Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "type": "text", + "content": "16. Jun, H., Nichol, A.: Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 192, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 480, + 213 + ], + "type": "text", + "content": "17. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ToG 42(4), 1-14 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 213, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 213, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 213, + 480, + 236 + ], + "type": "text", + "content": "18. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ToG 39(6) (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 236, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 480, + 267 + ], + "type": "text", + "content": "19. Li, J., Tan, H., Zhang, K., Xu, Z., Luan, F., Xu, Y., Hong, Y., Sunkavalli, K., Shakhnarovich, G., Bi, S.: Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 267, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 267, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 133, + 267, + 480, + 289 + ], + "type": "text", + "content": "20. Li, W., Chen, R., Chen, X., Tan, P.: Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 289, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 289, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 133, + 289, + 480, + 320 + ], + "type": "text", + "content": "21. Li, Y., Dou, Y., Shi, Y., Lei, Y., Chen, X., Zhang, Y., Zhou, P., Ni, B.: Focaldreamer: Text-driven 3d editing via focal-fusion assembly. arXiv preprint arXiv:2308.10608 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 320, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 320, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 133, + 320, + 480, + 354 + ], + "type": "text", + "content": "22. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: CVPR. pp. 300–309 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 354, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 354, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 133, + 354, + 480, + 386 + ], + "type": "text", + "content": "23. Ling, H., Kim, S.W., Torralba, A., Fidler, S., Kreis, K.: Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 386, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 386, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 386, + 480, + 407 + ], + "type": "text", + "content": "24. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 407, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 407, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 133, + 407, + 480, + 439 + ], + "type": "text", + "content": "25. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 439, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 439, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 133, + 439, + 480, + 472 + ], + "type": "text", + "content": "26. Liu, Z., Feng, Y., Black, M.J., Nowrouzezahrai, D., Paull, L., Liu, W.: Meshdiffusion: Score-based generative 3d mesh modeling. arXiv preprint arXiv:2303.08133 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 133, + 472, + 480, + 504 + ], + "type": "text", + "content": "27. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 133, + 504, + 480, + 536 + ], + "type": "text", + "content": "28. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. In: Seminal graphics: pioneering efforts that shaped the field, pp. 347-353 (1998)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 536, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 536, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 133, + 536, + 480, + 557 + ], + "type": "text", + "content": "29. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 557, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 557, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 133, + 557, + 480, + 579 + ], + "type": "text", + "content": "30. Luo, T., Rockwell, C., Lee, H., Johnson, J.: Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 579, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 579, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 133, + 579, + 480, + 611 + ], + "type": "text", + "content": "31. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latentnerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "type": "text", + "content": "32. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 133, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 480, + 666 + ], + "type": "text", + "content": "33. Müller, N., Siddiqui, Y., Porzi, L., Bulo, S.R., Kontschieder, P., Nießner, M.: Diffrf: Rendering-guided 3d radiance field diffusion. In: CVPR. pp. 4328-4338 (2023)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "text", + "content": "Tang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "34. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 170 + ], + "type": "text", + "content": "35. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 191 + ], + "type": "text", + "content": "36. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 235 + ], + "type": "text", + "content": "37. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., et al.: Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "type": "text", + "content": "38. Qin, X., Zhang, Z., Huang, C., Dehghan, M., Zaiane, O.R., Jagersand, M.: U2-net: Going deeper with nested u-structure for salient object detection. Pattern recognition 106, 107404 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "text", + "content": "39. Ren, J., Pan, L., Tang, J., Zhang, C., Cao, A., Zeng, G., Liu, Z.: Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 290, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 290, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 480, + 310 + ], + "type": "text", + "content": "40. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 311, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 354 + ], + "type": "text", + "content": "41. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 354, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 396 + ], + "type": "text", + "content": "42. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS 35, 36479-36494 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "text", + "content": "43. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In: NeurIPS (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 439 + ], + "type": "text", + "content": "44. Shi, R., Chen, H., Zhang, Z., Liu, M., Xu, C., Wei, X., Chen, L., Zeng, C., Su, H.: Zero123++: a single image to consistent multi-view diffusion base model (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "type": "text", + "content": "45. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "text", + "content": "46. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 483, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 503 + ], + "type": "text", + "content": "47. Szymanowicz, S., Rupprecht, C., Vedaldi, A.: Splatter image: Ultra-fast single-view 3d reconstruction. In: arXiv (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "text", + "content": "48. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 525, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 556 + ], + "type": "text", + "content": "49. Tang, J., Zhou, H., Chen, X., Hu, T., Ding, E., Wang, J., Zeng, G.: Delicate textured mesh recovery from nef via adaptive surface refinement. arXiv preprint arXiv:2303.02091 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 557, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 589 + ], + "type": "text", + "content": "50. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 590, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 610 + ], + "type": "text", + "content": "51. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "52. Wang, P., Shi, Y.: Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201 (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "53. Wang, P., Tan, H., Bi, S., Xu, Y., Luan, F., Sunkavalli, K., Wang, W., Xu, Z., Zhang, K.: Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. arXiv preprint arXiv:2311.12024 (2023)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 447, + 101 + ], + "type": "text", + "content": "Large Multi-View Gaussian Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 423 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "54. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. In: CVPR. pp. 4563-4573 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "55. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "56. Weng, Z., Liu, J., Tan, H., Xu, Z., Zhou, Y., Yeung-Levy, S., Yang, J.: Single-view 3d human digitalization with large reconstruction models. arXiv preprint arXiv:2401.12175 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "text", + "content": "57. Xu, D., Yuan, Y., Mardani, M., Liu, S., Song, J., Wang, Z., Vahdat, A.: Agg: Amortized generative 3d gaussians for single image to 3d. arXiv preprint arXiv:2401.04099 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "58. Xu, Y., Tan, H., Luan, F., Bi, S., Wang, P., Li, J., Shi, Z., Sunkavalli, K., Wetzstein, G., Xu, Z., et al.: Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. arXiv preprint arXiv:2311.09217 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "type": "text", + "content": "59. Yariv, L., Puny, O., Neverova, N., Gafni, O., Lipman, Y.: Mosaic-sdf for 3d generative models. arXiv preprint arXiv:2312.09222 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "60. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 337, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 480, + 357 + ], + "type": "text", + "content": "61. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "text", + "content": "62. Zhao, Z., Liu, W., Chen, X., Zeng, X., Wang, R., Cheng, P., Fu, B., Chen, T., Yu, G., Gao, S.: Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. arXiv preprint arXiv:2306.17115 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "text", + "content": "63. Zou, Z.X., Yu, Z., Guo, Y.C., Li, Y., Liang, D., Cao, Y.P., Zhang, S.H.: Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. arXiv preprint arXiv:2312.09147 (2023)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "text", + "content": "Tang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LHRS-Bot_ Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model/2d33d60a-efea-43ed-9423-c305740d4cd0_content_list.json b/2024/LHRS-Bot_ Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model/2d33d60a-efea-43ed-9423-c305740d4cd0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e9840626ee4330dbaaa68f9cdc6a91f31608607d --- /dev/null +++ b/2024/LHRS-Bot_ Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model/2d33d60a-efea-43ed-9423-c305740d4cd0_content_list.json @@ -0,0 +1,1698 @@ +[ + { + "type": "text", + "text": "LHRS-Bot: Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model", + "text_level": 1, + "bbox": [ + 238, + 140, + 763, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dilxat Muhtar†, Zhenshi Li†, Feng Gu, Xueliang Zhang*, and Pengfeng Xiao", + "bbox": [ + 223, + 234, + 774, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nanjing University \nLzhenshi@outlook.com, {dmuhtar, gufeng}@smail.nju.edu.com \n{zxl, xiaopf}@nju.edu.cn", + "bbox": [ + 279, + 263, + 720, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The revolutionary capabilities of large language models (LLMs) have paved the way for multimodal large language models (MLLMs) and fostered diverse applications across various specialized domains. In the remote sensing (RS) field, however, the diverse geographical landscapes and varied objects in RS imagery are not adequately considered in recent MLLM endeavors. To bridge this gap, we construct a large-scale RS image-text dataset, LHRs $^1$ -Align, and an informative RS-specific instruction dataset, LHRs-Instruct, leveraging the extensive volunteered geographic information (VGI) and globally available RS images. Building on this foundation, we introduce LHRs-Bot, an MLLM tailored for RS image understanding through a novel multi-level vision-language alignment strategy and a curriculum learning method. Additionally, we introduce LHRs-Bench, a benchmark for thoroughly evaluating MLLMs' abilities in RS image understanding. Comprehensive experiments demonstrate that LHRs-Bot exhibits a profound understanding of RS images and the ability to perform nuanced reasoning within the RS domain $^2$ .", + "bbox": [ + 261, + 349, + 753, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Multimodal $\\cdot$ Large language model $\\cdot$ Remote sensing", + "bbox": [ + 261, + 585, + 705, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 628, + 374, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recently advanced large language models (LLMs) have presented remarkable capabilities in engaging in informative conversations and solving complex problems [2,8,11,23,46-48]. To extend the perceptual capabilities of LLMs, multimodal large language models (MLLMs) bring eyes to LLMs. These models, by aligning with visual representations and visual instruction tuning, have demonstrated impressive multimodal instruction following capabilities and can act as a general-purpose interface for a broad range of tasks [1, 3, 22, 28, 30, 33, 64].", + "bbox": [ + 217, + 665, + 785, + 771 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal contribution, listed in random order.", + "bbox": [ + 238, + 782, + 532, + 797 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 240, + 797, + 395, + 811 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "LHRS stands for 'Language Helps Remote Sensing'.", + "bbox": [ + 240, + 811, + 589, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "2Data, Code and model are available at https://github.com/NJU-LHRS/LHRS-Bot", + "bbox": [ + 240, + 825, + 784, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This success in general domain has served as a catalyst for the development of domain specialized MLLMs, with notable applications in biomedicine [29], autonomous driving [36], and robotics [15].", + "bbox": [ + 212, + 146, + 460, + 252 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Among various specialized domains, remote sensing (RS) image understanding is particularly important, as it directly involves the monitoring and analysis of the Earth's surface and environment [5, 41, 51, 55]. However, RS images present significant challenges for holistic image understanding due to the intricate complexity of different objects across diverse landscapes world", + "bbox": [ + 212, + 253, + 460, + 434 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/106257a820d7b236c17b5b0d3e5f88fe33f6417123e81c6f4d4ed96e396fe0f6.jpg", + "image_caption": [ + "Fig.1: Overall performance comparison between LHRS-Bot and existing MLLMs on different RS image understanding tasks." + ], + "image_footnote": [], + "bbox": [ + 478, + 143, + 790, + 368 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "wide, coupled with the variability in visual scales [19, 37, 42]. MLLMs bring together vision and language, offering possibilities for a unified and interactive understanding of RS images [51]. In this direction, several RS-specific MLLMs have been proposed, such as RSGPT [21], GeoChat [25], and SkyEyeGPT [60]. However, they are struggling to fit existing public RS datasets, due to the oversight of exploiting the extensive and worldwide RS features to infuse comprehensive RS visual knowledge into LLMs. Furthermore, these contemporary approaches of bridging vision and language domains tend to focus on the incorporation of high-level visual semantics. This strategy neglects the fact that different levels of visual information are crucial to fully align the linguistic and visual domains, which could facilitate a unified RS image understanding at different granularity [37, 49, 52]. Detailed related works are available in supplementary material.", + "bbox": [ + 212, + 434, + 787, + 616 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To bridge the gap, we present LHRS-Bot, a specialized MLLM for RS, enhanced by globally available volunteered geographical information (VGI) and worldwide RS images. Specifically, we construct a large-scale, semantically rich, and feature-diverse dataset, LHRS-Align, by geographically pairing RS images with plentiful attributed information from the OpenStreetMap VGI database3, and consequently generating image captions using LLMs. Our rigorous data cleaning process, which includes dedduplication, pruning, and semantic balancing, ensures that LHRS-Align comprises 1.15 million meaningful and high-quality RS image-text pairs, laying a solid foundation for embedding rich RS visual knowledge into MLLMs. Furthermore, we construct our instruction dataset, LHRS-Instruct, by reorganizing a broad range of open-source RS datasets into multitask instructional datasets and employing GPT-4 [38] to create complex instructional data from the high-quality subset of LHRS-Align. Building on these", + "bbox": [ + 212, + 617, + 787, + 813 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "D. Muhtar et al.", + "bbox": [ + 271, + 114, + 382, + 127 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "3https://www.openstreetmap.org/", + "bbox": [ + 236, + 824, + 480, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "datasets, we propose a novel RS-specific MLLM, LHRS-Bot, by implementing a novel bridging strategy for efficiently summarizing multi-level visual representations, alongside a curriculum learning methodology. LHRS-Bot achieves state-of-the-art performance across a variety of RS image understanding tasks (Fig. 1) and demonstrates remarkable abilities to detect intricate objects, engage with human conversation, and provide insights from visual information within RS images (Fig. 5).", + "bbox": [ + 212, + 146, + 782, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, we manually construct a high-quality benchmark, LHRS-Bench, to facilitate the RS community in evaluating the RS-domain MLLMs across diverse evaluation dimensions. LHRS-Bench includes 690 single-choice questions, spanning 5 top-level evaluation dimensions including 11 fine-grained categories to facilitate a comprehensive, objective, and quantitative RS-specific evaluation.", + "bbox": [ + 212, + 252, + 784, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main contributions of our work are summarized as follows:", + "bbox": [ + 240, + 329, + 694, + 342 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce LHRS-Align, a large-scale RS image-text dataset, created through a meticulous data generation pipeline using open-source global geographic data. LHRS-Align showcases a broad spectrum of RS semantic visual information, establishing a solid foundation for RS-specific MLLMs.", + "2. We create LHRS-Instruct, a multimodal instruction-following dataset tailored for RS image understanding. LHRS-Instruct not only contains instruction data to serve for various RS image understanding tasks but also contains complex visual reasoning data generated by GPT-4.", + "3. We propose LHRS-Bot, an MLLM specifically designed for the RS domain. LHRS-Bot employs a novel bridging strategy and leverages a curriculum learning approach to fully exploit the inherent knowledge of the proposed datasets. LHRS-Bot demonstrates superior performance across various RS image understanding tasks, along with the ability to follow complex instructions.", + "4. We establish LHRS-Bench, a benchmark aimed at providing a comprehensive and systematic evaluation of RS MLLMs. LHRS-Bench comprises 690 meticulously crafted single-choice questions, encompassing a broad spectrum of 11 fine-grained dimensions." + ], + "bbox": [ + 222, + 353, + 782, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Dataset", + "text_level": 1, + "bbox": [ + 215, + 650, + 328, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Large-Scale RS Alignment Dataset—LHRS-Align", + "text_level": 1, + "bbox": [ + 214, + 683, + 687, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LHRS-Align is developed by initially geo-aligning orthorectified RS images with geographic vector features from the free spatial database, OSM. The attributes of these features are then utilized to generate image captions employing an LLM. The detailed pipeline is illustrated in Fig. 2 and introduced as below.", + "bbox": [ + 212, + 709, + 782, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data Preparation. The RS images of LHRS-Align are sourced from Google Earth imagery4, encompassing a worldwide coverage with a spatial resolution of 1 meter. It should be noted that our proposed pipeline is applicable to any RS", + "bbox": [ + 214, + 770, + 784, + 816 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "LHRS-Bot", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "4https://www.google.com/earth/", + "bbox": [ + 236, + 824, + 472, + 839 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ea02bf74919b00b4eb6e8c1423a9dfe77370e4852127a9c0bb4ddf01593faaf4.jpg", + "image_caption": [ + "Fig. 2: Pipeline of constructing LHRs-Align." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 787, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "imagery with geographic coordinates. The geographic features are gathered from OSM, a crowdsourced VGI database containing abundant geographical objects. Each object (feature) in the OSM database is annotated as the digitization of the object geometry (e.g., polygons, lines, and points) as well as the semantic tags. A tag is a key-value pair that provides information about a geographic object5, for example, \"landuse=residential\". Thanks to the free access to the globally encompassing data sources, our pipeline can produce a dataset that offers a vast and comprehensive variety on a worldwide scale. For the current version of LHRS-Align, we collect RS images and OSM features within a scope of $0.3^{\\circ}$ longitude $\\times 0.28^{\\circ}$ latitude for each of the selected cities from all over the world. We customize our own OSM database to streamline the subsequent feature selection process. More information about our data source and database is introduced in supplementary material.", + "bbox": [ + 212, + 354, + 787, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Geo-alignment between Images and Geographic Features. We iterate through our customized database to identify qualified features serving as \"anchors\". Based on the geographical extent of each anchor feature, the corresponding RS image is acquired, which is a square bounding box centered on the feature. These images subsequently undergo a stream of postprocessing and pruning procedures. Following this, due to the possibility of multiple features being present within a single image extent, all the features within each image are further identified. At this point, we have established a raw data pool of around 4 million pairs, in which one RS image corresponds to multiple geographical features. Further details for this processes are provided in supplementary material.", + "bbox": [ + 212, + 551, + 787, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attribute Pruning and Semantic Balancing. In the OSM database, all the features are assigned with their semantic attributes, that is, multiple key-value tags. However, not all the keys are favorable information for visual identification, such as website and postal code. Hence, we filter out irrelevant keys through a pipeline as detailed in supplementary material. Fig. 2 provides an example showcasing an image along with its geographical features and their corresponding attributes. Furthermore, the OSM database exhibits considerable fea", + "bbox": [ + 212, + 705, + 787, + 811 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "D. Muhtar et al.", + "bbox": [ + 271, + 114, + 382, + 127 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "5https://wiki.openstreetmap.org/wiki/Map_features", + "bbox": [ + 236, + 825, + 619, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ture imbalance, notably dominated by geographical features such as residential or farmland. Given the importance of balanced semantic information for effective vision-language alignment [40,54], we implement a data balancing method, as detailed in supplementary material, to create a more semantically balanced dataset. As a result, this balancing process retains 1.15 million images remaining out of the original 4 million.", + "bbox": [ + 212, + 146, + 782, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Caption Generation with LLM. The use of image-caption pairs to incorporate visual knowledge into LLMs has been validated as an effective approach [3, 14, 29, 56, 64]. However, using plain image key-value labels can be problematic, as they lack grammatical structure, potentially disrupting the generative and comprehension abilities of LLMs. Similarly, rule-based caption generation strategies may introduce significant redundancy. Therefore, we opt to use open-source LLMs to generate image captions by summarising the key-value tags of the corresponding images. We experiment with various in-context learning examples and LLMs, ranging in size from 7B to 70B. After comprehensive testing by a domain expert, we select Vicuna-v1.5-13B [10] for caption generation as it achieves the best trade-offs in terms of generation efficiency and quality. Detailed information about the in-context examples and the specific deployment of Vicuna-v1.5-13B [10] can be found in supplementary material. Our final dataset, referred as LHRS-Align, comprises approximately 1.15 million image-caption pairs. Each pair is accompanied with additional metadata, such as image resolution and geographical location (country and city). Several examples of LHRS-Align are also presented in supplementary material.", + "bbox": [ + 212, + 237, + 787, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.2 RS-Specific Instruction-Following Dataset—LHRS-Instruct", + "text_level": 1, + "bbox": [ + 214, + 513, + 764, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We craft a multimodal instruction-following dataset LHRS-Instruct by utilizing two public image-caption datasets and our LHRS-Align dataset.", + "bbox": [ + 212, + 537, + 782, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instruction-Following Dataset from Public RS Caption Dataset. We start by meticulously evaluating and selecting all image_caption datasets within the RS domain. Our selection process focuses on the quality of captions and involves a thorough human evaluation. This requires each image to be accompanied by five distinct captions, with each caption uniquely detailing different aspects of the image. Based on this criterion, we identify RSITMD [58] and NWPU [9] as the datasets with superior caption quality. Subsequent to selection, we design a rigorous data cleaning process, which involves deduplication, computation of caption lengths, and assessment of relevance through CLIP-score calculations. This results in the extraction of 3.4K and 30K high-quality image_caption pairs from RSITMD and NWPU, respectively. We then use several in-context examples along with the five captions per image for prompting Vicuna-v1.5-13B [10] to generate a conversation that specifically focuses on the content of the objects in the image. Finally, we filter out the most relevant conversations through pre-defined rules and evaluations from three domain experts. This process results in 2.9K and 25K instruction data from these two datasets, respectively. For more detailed information on the in-context examples used and the data filtering criteria, please refer to the supplementary material.", + "bbox": [ + 212, + 568, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "LHRS-Bot", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instruction-Following Dataset from LHRS-Align. Considering the limitations of public image-caption datasets, which typically lack bounding box annotations, and the constrained capabilities of Vicuna-v1.5-13B [10], the generated conversations might omit crucial details such as the locations and quantity of objects in the images, along with visual reasoning information. To mitigate this issue, we meticulously select 15K representative RS images that are rich in attribute information from our LHRS-Align dataset. We then proceed to calculate the bounding boxes for various objects within these images, utilizing the built-in spatial coordinates of different attributes in the OSM database. Leveraging this enriched data, we prompt GPT-4 to craft complex instruction data that incorporates visual reasoning, detailed image descriptions, and conversations about the location and number of objects. The final step involves a manual cleaning of the generated results. This process yields a dataset comprising 0.9K samples about visual reasoning, 4K samples about detailed image description, and 7K samples about conversation. For more details, please refer to supplementary material.", + "bbox": [ + 212, + 146, + 787, + 373 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2.3 RS-specific MLLM Evaluation Benchmark—LHRS-Bench", + "text_level": 1, + "bbox": [ + 215, + 407, + 754, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We construct LHRS-Bench to evaluate the performance of MLLMs in the RS domain. LHRS-Bench employs hierarchical taxonomies to comprehensively evaluate MLLMs across various dimensions (Tab. 1). To prevent data leakage, we abstain from using public RS datasets and instead meticulously collect images from Google Earth. In terms of question format, we utilize single-choice questions", + "bbox": [ + 212, + 444, + 787, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[16, 26, 27, 34] with 2 to 4 candidate answers, enabling a more quantitative and objective evaluation compared to open-ended questions [4, 6, 33, 57]. Every question and answer within LHRS-Bench are manually crafted to guarantee precision and reliability. Totally, LHRS-Bench comprises 108 images and 690 questions. Some examples of LHRS-Bench are shown in supplementary material.", + "bbox": [ + 212, + 521, + 486, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The evaluation dimensions of LHRs-Bench encompass 5 top-level dimensions, further delineated into 11 fine-grained sub-dimensions, as shown in Tab. 1. These top-level dimensions include recognition, imagery, spatial-awareness, quantity, and reasoning. For detailed explanations about each dimensions, please refer to supplementary material.", + "bbox": [ + 212, + 688, + 488, + 839 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/967fd325dec05d8af141afe85f16d102134a5245da9ceca656af534c0620d9a1.jpg", + "table_caption": [ + "Table 1: Overview of the dataset distribution. LHRS-Bench contains 108 RS images with 690 question-answer pairs, covering five major evaluation dimensions and 11 sub-dimensions of questions. Each sample is a human-annotated single-choice question that may involve one or more evaluation dimensions simultaneously." + ], + "table_footnote": [], + "table_body": "
DimensionSub-dimensionCount
RecognitionIdentity634
Color113
Orientation39
Shape37
Area75
ImageryResolution21
Modality23
Spatial-awarenessLocation204
Distance22
Quantity-137
Reasoning-46
", + "bbox": [ + 511, + 647, + 769, + 824 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "D. Muhtar et al.", + "bbox": [ + 271, + 114, + 382, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/49dbacddee4fb62b00e112d482402d203ab36359fad30843dafe281939bbf15a.jpg", + "image_caption": [ + "Fig. 3: Architecture of the proposed LHRS-Bot. LHRS-Bot employs learnable queries in conjunction with a vision perceiver to summarize multi-level visual representations. These representations are subsequently concatenated with language token embeddings for processing within the LLM." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 462, + 328, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present LHRS-Bot to harness the full potential of LLMs and general visual encoders for enhanced RS images understanding and reasoning. This section delves into each components of LHRS-Bot, detailing how we achieve fully alignment between vision and language models across multiple levels. Subsequently, we explain the curriculum training strategy for training LHRS-Bot.", + "bbox": [ + 212, + 496, + 787, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.1 Model Architecture", + "text_level": 1, + "bbox": [ + 215, + 599, + 426, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The overall architecture of LHRS-Bot is presented in Fig. 3. LHRS-Bot is primarily composed of three components: a universal vision encoder, a vision perceiver, and a foundational LLM.", + "bbox": [ + 212, + 627, + 785, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Vision Encoder. LHRS-Bot utilizes the pre-trained CLIP visual encoder ViT-L/14 [40] with input resolution of $224 \\times 224$ pixel for encoding images. The prior works [33,56,64] usually utilize the final hidden feature for abstracting the semantic content of an image. However, several studies have highlighted that features at different levels of a deep learning model embed distinct types of information [17,31,39]. Therefore, we argue that previous image encoding methods fail to sufficiently capture the semantic content of images, leading to a reliance on extensive image-text pairs for vision-language alignment. Consequently, we retain multiple hidden features from various layers of the image encoder. This approach offers a more thorough representation of the underlying image, facilitating a thorough and efficient domain alignment.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "LHRS-Bot", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/23b5a95e198ddfb8378944814b1c364b70a84c8720f0ee3803d005bf9817c263.jpg", + "image_caption": [ + "Fig. 4: Curriculum learning strategy. We progressively unlock various parameters and introduce data of escalating complexity at different stages, enabling the model to gradually assimilate visual knowledge with increasing difficulty." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 784, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Vision Perceiver. Incorporating multi-level visual features brings extended image tokens, increasing computational overhead and potentially overwhelming language information. To mitigate this issue, we introduce a set of learnable queries for each level of image features. These queries are designed to summarize the semantic information of each level of visual features through a series of stacked cross-attention and MLP layers. Once summarized, these image features are then fed into the LLM. Furthermore, we implement a descending query allocation strategy inspired by the observation that image feature redundancy tends to increase with network depth [7,24]. This strategy allocates more queries to the shallower image feature levels while gradually reducing the number of queries for deeper levels.", + "bbox": [ + 212, + 324, + 787, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Large Language Model. We employ the LLaMA2-7B [48] as the central expert for interpreting visual and language information to output the desired results following the previous practices [25, 29, 43].", + "bbox": [ + 212, + 489, + 787, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.2 Curriculum Learning Strategy", + "text_level": 1, + "bbox": [ + 214, + 546, + 511, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We implement a three-stage curriculum learning strategy by progressively increasing the difficulty of training tasks (Fig. 4). This strategy incrementally aligns the vision encoder with the LLM, ultimately enabling the LLM to solve multiple visual tasks and follow complex vision-language instructions.", + "bbox": [ + 212, + 570, + 787, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Stage 1: Pre-training. In the first stage, we use the large-scale, weakly labelled LHRS-Align dataset (Sec. 2.1) to train the vision perceiver to incorporate a wide range of RS visual knowledge into the LLM. Specifically, for each sample in the LHRS-Align dataset, the output visual embedding tokens $X_{v}$ obtained from the vision perceiver are concatenated with the language embedding tokens $X_{l}$ from the corresponding image caption. The concatenated result is then fed into the LLM. The objective of this stage is to maximise the conditional probability $P(X_{l}|X_{v};\\theta)$ of generating the caption tokens $X_{l}$ given the prefix visual tokens $X_{v}$ :", + "bbox": [ + 212, + 631, + 787, + 767 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\log P (X _ {l} | X _ {v}; \\theta) \\\\ = \\sum_ {i = 1} ^ {L} \\log P \\left(x _ {i} \\mid X _ {v}, X _ {l, < i}; \\theta\\right), \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 777, + 784, + 835 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "D. Muhtar et al.", + "bbox": [ + 271, + 114, + 383, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $L$ is the length of caption tokens, $\\theta$ represents the trainable parameters, and $X_{l, < i}$ denotes the generated tokens prior to the current prediction token $x_i$ .", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Stage 2: Multi-task Pre-training. In the second stage, we further train the vision perceiver and efficiently fine-tune the LLM using LoRA [20] on multitask instruction datasets to improve multimodal task solving capabilities of LLM. The construction of the multi-task instruction datasets begins with the collection of diverse public datasets in the RS domain, encompassing tasks such as classification, VQA, and visual grounding. These collected datasets are then meticulously transformed into instruction-based formats using manually designed templates. More detailed information about these multi-task datasets and the templates used can be found in the supplementary material. Moreover, we integrate the instruction data from NWPU, RSITMD and the 4K detailed descriptions data from LHRS-Instruct (Sec. 2.2) to further enrich the LLM with fine-grained RS visual knowledge. The objective for this stage remains similar to Eq. (1) from stage 1, albeit with an adaptation that the generated tokens are now additionally conditioned on the instruction tokens $X_{\\text{instruct}}$ :", + "bbox": [ + 212, + 178, + 787, + 388 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\log P (X _ {l} | X _ {v}, X _ {\\text {i n s t r u c t}}; \\theta) \\\\ = \\sum_ {i = 1} ^ {L} \\log P \\left(x _ {i} \\mid X _ {v}, X _ {\\text {i n s t r u c t}}, X _ {l, < i}; \\theta\\right). \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 402, + 784, + 460 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Stage 3: Supervised Fine-tuning. In the final stage, our focus shifts towards enhancing the conversational and reasoning capabilities of our model by incorporating more complex instruction data. We utilize all the instruction data from the LHRS-Instruct dataset (Sec. 2.2) and randomly select 20K data from the LLaVA complex reasoning dataset [29]. Additionally, we mix the above datasets with the multi-task datasets from stage 2 using a lower data sampling ratio to further unleash the multi-task problem-solving ability of LHRS-Bot.", + "bbox": [ + 212, + 474, + 787, + 582 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 597, + 366, + 614 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conduct comprehensive experiments across image classification, VQA, and visual grounding tasks within the RS domain. These experiments are designed to rigorously validate the superior multi-task solving capabilities of LHRS-Bot. Furthermore, we evaluate the performance of various LLMs in the RS domain using the proposed LHRS-Bench benchmark. We present both quantitative and quality results.", + "bbox": [ + 212, + 633, + 787, + 726 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 750, + 428, + 766 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset. We select diverse RS datasets for providing a comprehensive assessment across different tasks within the RS field: seven RS classification datasets (namely AID [53], WHU-RS19 [13], NWPU [9], SIRI-WHU [65], EuroSAT [18], METER-ML [63], and fMoW [12]), two VQA datasets (LR and HR subsets", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "LHRS-Bot", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/42423866970876809e604dbc3f8175a9ab3eb2cb2935978df25024421f4da5d4.jpg", + "table_caption": [ + "Table 2: Classification accuracy of different models on each dataset." + ], + "table_footnote": [], + "table_body": "
MethodAIDWHU-RS19NWPUSIRI-WHUEuroSATMETER-MLfMoWAvg.
LLaVA-1.531.1054.5534.9617.7126.1221.7311.4328.23
MiniGPTv232.9664.8028.1535.4638.5614.295.2031.35
InstructBLIP29.5036.7634.0118.2020.2514.426.7122.84
mPLUG-OWL248.7972.6646.5854.8333.2936.2717.8544.32
QWen-VL-Chat55.3072.2542.7354.5826.4238.776.8942.42
InternLM-XComposer51.6172.8947.5146.8339.7040.2111.2844.29
LHRS-Bot91.2693.1783.9462.6651.4069.8156.5671.83
", + "bbox": [ + 215, + 162, + 784, + 275 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "of RSVQA [35]), and two visual grounding datasets (RSVG [45] and DIOR-RSVG [59]). We exclude any data from the test sets that overlap with the training set to avoid data leakage. More details about each test dataset are presented in supplementary material.", + "bbox": [ + 212, + 292, + 784, + 354 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baseline. We evaluate LHRS-Bot and various powerful open-source MLLMs, including LLaVA-v1.5 [32], MiniGPT-v2 [64], InstructBLIP [14], mPLUG-Owl2 [56], QWen-VL-Chat [3], and InternLM-Xcomposer [61], on different tasks. Furthermore, we also include the results of RSGPT [21], GeoChat [25] and SkyEyeGPT [60] from the original papers for comparisons whenever possible. Unless otherwise noted, we utilized 7B variants for all our experiments.", + "bbox": [ + 212, + 362, + 807, + 455 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation. We extract hidden states at layers $\\{i = L / 3, j = 2L / 3, k = L - 1\\}$ for summarizing through vision perceiver, where $L$ denotes the number of layers in the vision encoder. Our vision perceiver implementation comprises six layers of cross-attention and MLP, with each hidden state assigned $\\{64, 48, 32\\}$ learnable queries. We apply LoRA module to every linear layer of the LLM, setting the rank $r$ and the scale factor $\\alpha$ for LoRA to 128 and 256, respectively. Moreover, we add task identifier [CLS], [VQA], and [VG] for classification, VQA, and visual grounding tasks, respectively, for enabling LHRS-Bot to better distinguish different tasks [64]. For additional details on the hyperparameters used at each stage, please refer to supplementary material.", + "bbox": [ + 212, + 463, + 787, + 614 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 214, + 633, + 459, + 648 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We evaluate different MLLMs using greedy decoding with the same prompt for each task and report the average result of two independent trials. Details about the evaluation prompts are provided in supplementary material.", + "bbox": [ + 212, + 657, + 785, + 705 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Classification. The classification results across different datasets are presented in Tab. 2. These results show that LHRS-Bot outperforms all competing models in all seven classification datasets, with a notable margin of $27.51\\%$ higher average accuracy than the second ranked method. It is worth mentioning that, except for NWPU, METER-ML, and fMoW, the remaining datasets are completely absent from the multi-task training dataset6, demonstrating the strong", + "bbox": [ + 212, + 712, + 787, + 806 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "D. Muhtar et al.", + "bbox": [ + 271, + 114, + 382, + 127 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "However, we have no means of confirming whether these data appeared in the LHRs-Align dataset, considering that images in LHRs-Align, along with several of the", + "bbox": [ + 214, + 810, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/70df29b5b48afce876174b97e55390e0030cf6a4362a9d8a1c619190d5bd24a3.jpg", + "table_caption": [ + "Table 3: Accuracy of different models on each visual question answering dataset. Following the practice of prior works [21, 25, 60], we omit area and count questions during evaluation." + ], + "table_footnote": [], + "table_body": "
MethodRSVQA-LRRSVQA-HR
Rural/UrbanPresenceCompareAvg.PresenceCompareAvg.
LLaVA-1.559.2273.1665.1965.8648.9659.0253.99
MiniGPTv260.0251.6467.6459.7768.3464.7166.53
InstructBLIP62.6248.8365.9259.1262.6362.9062.77
mPLUG-Owl257.9974.0463.6965.2447.6058.4753.04
QWen-VL-Chat62.0047.6566.5458.7361.7565.9863.87
InternLM-XCompose59.0066.7452.9159.5567.7966.6267.21
RSGPT94.0091.1791.7092.2990.9290.0290.47
GeoChat91.0990.3394.0091.8158.4583.1970.82
SkyEyeGPT88.9388.6375.0084.1680.0080.1382.56
LHRS-Bot89.0788.5190.0089.1992.5792.5392.55
", + "bbox": [ + 243, + 191, + 754, + 342 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "generalization ability of LHRS-Bot. In addition, in the context of the fine-grained RS classification dataset fMoW, our method significantly surpasses other models, leading by an impressive $40\\%$ on average. This highlights the critical role of continue pre-training in the RS domain and the effectiveness of our alignment dataset LHRS-Align.", + "bbox": [ + 215, + 366, + 784, + 441 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Visual Question Answering. We report the results of different models on RSVQA-LR and RSVQA-HR datasets in Tab. 3. LHRS-Bot demonstrates performance on par with the previously leading RS domain-specific models on the VQA task. Notably, LHRS-Bot achieves comparable results to RSGPT, despite RSGPT being fine-tuned for five epochs on the RSVQA datasets. Furthermore, LHRS-Bot sets a new record in the RSVQA-HR task with high-resolution RS images, showcasing its superior visual recognition abilities with high-resolution RS images.", + "bbox": [ + 215, + 478, + 784, + 599 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Visual Grounding. The results for the visual grounding task are shown in Tab. 4, where LHRS-Bot outperforms other counterparts. Notably, despite SkyEyeGPT being trained on a more extensive object detection dataset, LHRS-Bot demonstrates comparable performance. This suggests that LHRS-Bot has effectively acquired fine-grained visual information throughout the training process.", + "bbox": [ + 215, + 625, + 485, + 791 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/084468b90b955fdb307fb7d7059f85fc122b7c88d2a720f6d7a50fb40d1bada1.jpg", + "table_caption": [ + "Table 4: Comparison of visual grouting tasks. Accuracy is calculated based on whether the predicted bounding box overlaps with the ground-truth box by more than 0.5 IoU." + ], + "table_footnote": [], + "table_body": "
MethodRSVGDIOR-RSVGAvg.
QWen-VL-Chat44.7680.6562.71
MiniGPTv246.6485.9966.32
SkyEyeGPT70.5088.5979.55
LHRS-Bot73.4588.1080.78
", + "bbox": [ + 514, + 696, + 764, + 772 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "LHRS-Bot", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_footnote", + "text": "classification datasets (AID, WHU-RS19, and SIRI-WHU), are collected from Google Earth.", + "bbox": [ + 215, + 811, + 782, + 837 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/325e327ea26c956f8fa30f7c62c6f5dcd117078860e86e35ca2a986dda1c8787.jpg", + "table_caption": [ + "Table 5: Comparison results on the LHRS-Bench dataset with different open-source and closed-source MLLMs." + ], + "table_footnote": [], + "table_body": "
MethodIdentityColorOrientationShapeAreaResolutionModalityLocationDistanceQuantityReasoningAvg.
LLaVA-1.542.9030.9715.3843.2444.0023.8021.7435.7827.2721.8954.3537.45
MiniGPTv220.1915.932.568.1113.3314.294.3515.2013.649.4936.9616.88
InstructBLIP20.3522.125.1324.3222.6738.1000.0013.7322.8513.1430.4318.87
mPLUG-OWL241.4845.1312.8240.5432.0033.3300.0030.8813.6432.1254.3437.00
QWen-VL-Chat31.3832.747.6932.4329.3319.054.3524.0221.4521.1756.5228.65
InternLM-XComposer34.3830.977.6937.8426.6700.008.7033.3322.7318.2565.2231.09
GPT-4-Turbo50.3243.3610.2667.5744.0042.8639.1342.1622.7321.8971.7444.56
Claude-3-Opus41.0136.2817.9564.8641.3366.6739.1334.8045.4526.2854.3539.08
LHRS-Bot44.6442.4823.0867.5725.3342.8632.3556.5222.7318.2573.9139.38
", + "bbox": [ + 215, + 184, + 785, + 287 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Evaluation on LHRS-Bench", + "text_level": 1, + "bbox": [ + 215, + 306, + 488, + 320 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We evaluate LHRS-Bot with different open-source and closed-source MLLMs on our LHRS-Bench benchmark. For each question within the benchmark, choices are shuffled, and the generation process is repeated four times. An answer is considered incorrect if any attempt yields an incorrect response. Same prompts are used for evaluation, as detailed in supplementary material. We consider the model answers correctly if it outputs the correct choice or the answer corresponding to the correct choice. We calculate accuracy as our evaluation metric.", + "bbox": [ + 212, + 339, + 785, + 445 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The results presented in Tab. 5 indicate that LHRS-Bot outperforms all other open-source models. It is worth noting that, LHRS-Bot greatly outperforms other open-source models in terms of the evaluation dimensions of spatial resolution and image modality, even though these aspects were not explicitly trained for in the models, highlighting the effectiveness of RS-specific MLLM training. Additionally, LHRS-Bot outperforms the strong closed-source model Claude and even matches the performance of GPT-4V. However, there is still significant room for improvement in MLLM capabilities within the RS domain. Examples of evaluation results are available on supplementary material.", + "bbox": [ + 212, + 446, + 785, + 583 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 215, + 606, + 447, + 621 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conduct qualitative demonstrations to illustrate the extensive capabilities of LHRS-Bot in interpreting free-form RS images, as showcased in Fig. 5. The results demonstrate that LHRS-Bot not only comprehends the overall content of an image but also exhibits a strong ability for discerning fine-grained details. Moreover, LHRS-Bot possesses a perceptive understanding of the orientation, position, and quantity of objects within an image. Furthermore, LHRS-Bot is capable of making reasoned inferences from the images and acknowledges its limitations by indicating uncertainty when it cannot provide a definitive answer.", + "bbox": [ + 212, + 638, + 785, + 760 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Ablation Analysis", + "text_level": 1, + "bbox": [ + 215, + 787, + 410, + 803 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this subsection, we carry out extensive ablation experiments to demonstrate the effectiveness of our dataset as well as the model design.", + "bbox": [ + 215, + 821, + 785, + 852 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "D. Muhtar et al.", + "bbox": [ + 271, + 114, + 382, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a6287d5176ccce0e788920bc7dfc760d7e45e5a1beea22f7606e382ab9ab466e.jpg", + "image_caption": [ + "Fig. 5: Qualitative examples between real users and LHRS-Bot." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 343 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/64cbb8a8c919864d12d22fea3a7b2ee9d32861cf06c0c4bded6f348161ef8eba.jpg", + "table_caption": [ + "Table 6: Performance with different alignment dataset." + ], + "table_footnote": [], + "table_body": "
AIDWHU-RS19ClassificationVQAVisual Grounding
NWPUSIRI-WHUEuroSATMETER-MLfMoWRSVQA-LRRSVQA-HRRSVGDIOR-RSVG
RS5M86.3196.7582.8060.2054.9264.7742.6878.3691.6570.5887.12
SkyScript88.2190.2678.0151.4343.6650.0231.2788.7991.2471.2585.05
LHRS-Align91.2693.1783.9462.6651.4069.8156.5689.1992.5573.4588.10
", + "bbox": [ + 215, + 402, + 787, + 454 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of LHRS-Align dataset. There are already many large-scale image-caption datasets in the field of RS [50,62]. To evaluate how different alignment datasets impact model performance, we substitute our LHRS-Align dataset with various image-caption datasets and repeat the entire training process. This process is then evaluated across three distinct tasks: classification, VQA, and visual grounding. The experimental setup remains consistent with Sec. 4.1, with the exception of the dataset used in the first stage.", + "bbox": [ + 212, + 482, + 787, + 587 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The results are presented in Tab. 6. Notably, despite possessing the smallest dataset (RS5M: 5M, SkyScript: 1.5M, LHRS-Align : 1.15M), LHRS-Align outperforms other settings. It is worth mentioning that both SkyScript and LHRS-Align, which leverage RS images combined with the OSM database, achieve results comparable to or better than the even larger RS5M, which is derived from publicly available natural image datasets, such as LAION [44]. This highlights the importance of domain-specific data collection. Additionally, superior performance of LHRS-Align over SkyScript suggests that enriched captions and a meticulous data cleaning pipeline can substantially enhance the effectiveness of MLLMs.", + "bbox": [ + 212, + 588, + 787, + 738 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of Model Design. We evaluate the effectiveness of designs involving multi-level visual hidden representations as well as descending query allocation strategies. Specifically, our evaluation focuses on comparing with the single level visual hidden state, projecting it through a two-layer MLP or a vision perceiver. We also perform comparisons between multi-level visual hidden", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "LHRS-Bot", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/09e8ddafa2e97a06121eda032b6efd4852dc31821b56cd6e3bfb2744617ab91b.jpg", + "table_caption": [ + "Table 7: Performance comparison across various architectural designs." + ], + "table_footnote": [], + "table_body": "
MethodMulti Level Number of Query ClassificationVQA Visual Grounding
MLP✘-64.4690.06
Vision Perceiver✘6463.6983.30
Vision Perceiver✓[64,64,64]63.7487.88
Vision Perceiver✓[32,48,64]60.6887.11
Token Merging✓6449.5181.49
Vision Perceiver✓[64,48,32]65.4390.14
", + "bbox": [ + 248, + 171, + 753, + 272 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "representations, employing either identical queries for each level or incremental query allocation strategies. Building on the concept introduced by the token merging [7], we explore the potential of integrating this strategy with a vision perceiver to efficiently leverage multi-level representations. To minimize experimental costs, our training process comprises just two stages: firstly, we employ LHRS-Align to train the bridge layer, which aligns the vision and language domains. Subsequently, we train the model using the LHRS-Instruct and multi-task instruction datasets.", + "bbox": [ + 212, + 285, + 785, + 404 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We report the average result on each task in Tab. 7. It is evident that the approach employed by LHRS-Bot achieves superior performance. The results further illustrate that, although the extraction of a singular visual representation performs better in tasks such as classification and VQA tasks, which depends on high-level visual semantics, it falls short in visual grounding tasks that require detailed visual information. Furthermore, employing the same number of queries to summarise different levels of visual information does not lead to better results. In contrast, adopting a diminishing allocation strategy yields superior results. This further demonstrates the large redundancy among the deep visual tokens [7]. Lastly, the token merging strategy proves ineffective, possibly due to the loss of crucial information during the merging process.", + "bbox": [ + 212, + 407, + 787, + 574 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 603, + 359, + 619 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We propose LHRS-Bot, an MLLM for the RS domain. To unleash the potential of LLMs for RS image understanding, we curate a large-scale dataset, LHRS-Align, for RS-specific alignment, and LHRS-Instruct, a multimodal instruction-following dataset to enhance LHRS-Bot's instruction-following capabilities. Additionally, we introduce LHRS-Bench, a benchmark providing a systematic evaluation framework for RS MLLMs.", + "bbox": [ + 212, + 642, + 785, + 732 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In comprehensive comparison experiments, LHRS-Bot demonstrates exceptional RS domain knowledge, surpassing both general-purpose and RS-specific MLLMs. Despite its impressive performance, LHRS-Bot shares limitations common to MLLMs, such as susceptibility to hallucinations. We believe that developing higher-quality alignment and instructional datasets, along with improved training strategies, can further enhance the performance of MLLMs for interpreting RS images.", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "D. Muhtar et al.", + "bbox": [ + 271, + 114, + 382, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research was supported by the National Natural Science Foundation of China under Grant 42071297, and in part by the AI & AI for Science Project of Nanjing University under Grant 02091480605203. We are grateful to High Performance Computing Center of Nanjing University for their help on GPU resources. We also would like to thank the anonymous reviewers for their constructive comments.", + "bbox": [ + 212, + 180, + 787, + 270 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 297, + 321, + 313 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)", + "2. Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., et al.: Qwen technical report. arXiv preprint arXiv:2309.16609 (2023)", + "3. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023)", + "4. Bai, S., Yang, S., Bai, J., Wang, P., Zhang, X., Lin, J., Wang, X., Zhou, C., Zhou, J.: Touchstone: Evaluating vision-language models by language models. arXiv preprint arXiv:2308.16890 (2023)", + "5. Bashmal, L., Bazi, Y., Melgani, F., Al Rahhal, M.M., Al Zuair, M.A.: Language integration in remote sensing: Tasks, datasets, and future directions. IEEE Geoscience and Remote Sensing Magazine (2023)", + "6. Bitton, Y., Bansal, H., Hessel, J., Shao, R., Zhu, W., Awadalla, A., Gardner, J., Taori, R., Schimdt, L.: Visit-bench: A benchmark for vision-language instruction following inspired by real-world use. arXiv preprint arXiv:2308.06595 (2023)", + "7. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461 (2022)", + "8. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)", + "9. Cheng, G., Han, J., Lu, X.: Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE 105(10), 1865-1883 (2017)", + "10. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)", + "11. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. Journal of Machine Learning Research 24(240), 1-113 (2023)", + "12. Christie, G., Fendley, N., Wilson, J., Mukherjee, R.: Functional map of the world. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6172-6180 (2018)" + ], + "bbox": [ + 218, + 332, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "LHRS-Bot", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Dai, D., Yang, W.: Satellite image classification via two-layer sparse coding with biased image representation. IEEE Geoscience and Remote Sensing Letters 8(1), 173-176 (2011)", + "14. Dai, W., Li, J., Li, D., Tiong, A., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arxiv 2023. arXiv preprint arXiv:2305.06500", + "15. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023)", + "16. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)", + "17. Ghiasi, A., Kazemi, H., Borgnia, E., Reich, S., Shu, M., Goldblum, M., Wilson, A.G., Goldstein, T.: What do vision transformers learn? a visual exploration. arXiv preprint arXiv:2212.06727 (2022)", + "18. Helber, P., Bischke, B., Dengel, A., Borth, D.: Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 12(7), 2217-2226 (2019)", + "19. Hossain, M.D., Chen, D.: Segmentation for object-based image analysis (obia): A review of algorithms and challenges from remote sensing perspective. ISPRS Journal of Photogrammetry and Remote Sensing 150, 115-134 (2019)", + "20. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)", + "21. Hu, Y., Yuan, J., Wen, C., Lu, X., Li, X.: Rsgpt: A remote sensing vision language model and benchmark. arXiv preprint arXiv:2307.15266 (2023)", + "22. Huang, S., Dong, L., Wang, W., Hao, Y., Singhal, S., Ma, S., Lv, T., Cui, L., Mohammed, O.K., Liu, Q., et al.: Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045 (2023)", + "23. Jiang, A.Q., Sablayrolles, A., Roux, A., Mensch, A., Savary, B., Bamford, C., Chaplot, D.S., Casas, D.d.l., Hanna, E.B., Bressand, F., et al.: Mixtral of experts. arXiv preprint arXiv:2401.04088 (2024)", + "24. Ju, C., Wang, H., Li, Z., Chen, X., Zhai, Z., Huang, W., Xiao, S.: Turbo: Informativity-driven acceleration plug-in for vision-language models. arXiv preprint arXiv:2312.07408 (2023)", + "25. Kuckreja, K., Danish, M.S., Naseer, M., Das, A., Khan, S., Khan, F.S.: Geochat: Grounded large vision-language model for remote sensing. arXiv preprint arXiv:2311.15826 (2023)", + "26. Li, B., Ge, Y., Ge, Y., Wang, G., Wang, R., Zhang, R., Shan, Y.: Seed-bench-2: Benchmarking multimodal large language models. arXiv preprint arXiv:2311.17092 (2023)", + "27. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125 (2023)", + "28. Li, C., Gan, Z., Yang, Z., Yang, J., Li, L., Wang, L., Gao, J.: Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020 (2023)", + "29. Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H., Gao, J.: Llava-med: Training a large language-and-vision assistant for biomedicine in one day. arXiv preprint arXiv:2306.00890 (2023)" + ], + "bbox": [ + 217, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "D. Muhtar et al.", + "bbox": [ + 271, + 114, + 382, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)", + "31. Li, M., Wang, S., Zhang, Q.: Visualizing the emergence of intermediate visual patterns in dnns. Advances in Neural Information Processing Systems 34, 6594-6607 (2021)", + "32. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)", + "33. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)", + "34. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023)", + "35. Lobry, S., Marcos, D., Murray, J., Tuia, D.: Rsvqa: Visual question answering for remote sensing data. IEEE Transactions on Geoscience and Remote Sensing 58(12), 8555-8566 (2020)", + "36. Ma, Y., Cao, Y., Sun, J., Pavone, M., Xiao, C.: Multimodal language model for driving. arXiv preprint arXiv:2312.00438 (2023)", + "37. Muhtar, D., Zhang, X., Xiao, P., Li, Z., Gu, F.: Cmid: A unified self-supervised learning framework for remote sensing image understanding. IEEE Transactions on Geoscience and Remote Sensing (2023)", + "38. OpenAI: Gpt-4 technical report (2023)", + "39. Park, N., Kim, W., Heo, B., Kim, T., Yun, S.: What do self-supervised vision transformers learn? arXiv preprint arXiv:2305.00729 (2023)", + "40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "41. Ratledge, N., Cadamuro, G., de la Cuesta, B., Stigler, M., Burke, M.: Using machine learning to assess the livelihood impact of electricity access. Nature 611(7936), 491-495 (2022)", + "42. Reed, C.J., Gupta, R., Li, S., Brockman, S., Funk, C., Clipp, B., Keutzer, K., Candido, S., Uytendaele, M., Darrell, T.: Scale-MAE: A scale-aware masked autoencoder for multiscale geospatial representation learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4088-4099 (2023)", + "43. Saygin Seyfioglu, M., Ikezogwo, W.O., Ghezloo, F., Krishna, R., Shapiro, L.: Quilt-lava: Visual instruction tuning by extracting localized narratives from open-source histopathology videos. arXiv e-prints pp. arXiv-2312 (2023)", + "44. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)", + "45. Sun, Y., Feng, S., Li, X., Ye, Y., Kang, J., Huang, X.: Visual grounding in remote sensing images. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 404-412 (2022)", + "46. Team, M.N.: Introducing mpt-7b: A new standard for open-source, commercially usable llms (2023), www.mosaicml.com/blog/mpt-7b, accessed: 2023-05-05", + "47. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "LHRS-Bot", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)", + "49. Wang, D., Zhang, J., Du, B., Xu, M., Liu, L., Tao, D., Zhang, L.: Samrs: Scaling-up remote sensing segmentation dataset with segment anything model. In: Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2023)", + "50. Wang, Z., Prabha, R., Huang, T., Wu, J., Rajagopal, R.: Skyscript: A large and semantically diverse vision-language dataset for remote sensing. arXiv preprint arXiv:2312.12856 (2023)", + "51. Wen, C., Hu, Y., Li, X., Yuan, Z., Zhu, X.X.: Vision-language models in remote sensing: Current progress and future trends. arXiv preprint arXiv:2305.05726 (2023)", + "52. Xia, G.S., Bai, X., Ding, J., Zhu, Z., Belongie, S., Luo, J., Datcu, M., Pelillo, M., Zhang, L.: DOTA: A large-scale dataset for object detection in aerial images. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3974-3983 (2018)", + "53. Xia, G.S., Hu, J., Hu, F., Shi, B., Bai, X., Zhong, Y., Zhang, L., Lu, X.: AID: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing 55(7), 3965-3981 (2017)", + "54. Xu, H., Xie, S., Tan, X.E., Huang, P.Y., Howes, R., Sharma, V., Li, S.W., Ghosh, G., Zettlemoyer, L., Feichtenhofer, C.: Demystifying clip data. arXiv preprint arXiv:2309.16671 (2023)", + "55. Yang, J., Gong, P., Fu, R., Zhang, M., Chen, J., Liang, S., Xu, B., Shi, J., Dickinson, R.: The role of satellite remote sensing in climate change studies. Nature climate change 3(10), 875-883 (2013)", + "56. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)", + "57. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mmvet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023)", + "58. Yuan, Z., Zhang, W., Fu, K., Li, X., Deng, C., Wang, H., Sun, X.: Exploring a fine-grained multiscale method for cross-modal remote sensing image retrieval. arXiv preprint arXiv:2204.09868 (2022)", + "59. Zhan, Y., Xiong, Z., Yuan, Y.: Rsvg: Exploring data and models for visual grounding on remote sensing data. IEEE Transactions on Geoscience and Remote Sensing 61, 1-13 (2023)", + "60. Zhan, Y., Xiong, Z., Yuan, Y.: Skyeyegpt: Unifying remote sensing vision-language tasks via instruction tuning with large language model. arXiv preprint arXiv:2401.09712 (2024)", + "61. Zhang, P., Wang, X.D.B., Cao, Y., Xu, C., Ouyang, L., Zhao, Z., Ding, S., Zhang, S., Duan, H., Yan, H., et al.: Internl m-xcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112 (2023)", + "62. Zhang, Z., Zhao, T., Guo, Y., Yin, J.: Rs5m: A large scale vision-language dataset for remote sensing vision-language foundation model. arXiv preprint arXiv:2306.11300 (2023)", + "63. Zhu, B., Lui, N., Irvin, J., Le, J., Tadwalkar, S., Wang, C., Ouyang, Z., Liu, F.Y., Ng, A.Y., Jackson, R.B.: Meter-ml: A multi-sensor earth observation benchmark for automated methane source mapping. arXiv preprint arXiv:2207.11166 (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "D. Muhtar et al.", + "bbox": [ + 271, + 114, + 382, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)", + "65. Zhu, Q., Zhong, Y., Zhao, B., Xia, G.S., Zhang, L.: Bag-of-visual-words scene classifier with local and global features for high spatial resolution remote sensing imagery. IEEE Geoscience and Remote Sensing Letters 13(6), 747-751 (2016)" + ], + "bbox": [ + 215, + 146, + 787, + 231 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "LHRS-Bot", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/LHRS-Bot_ Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model/2d33d60a-efea-43ed-9423-c305740d4cd0_model.json b/2024/LHRS-Bot_ Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model/2d33d60a-efea-43ed-9423-c305740d4cd0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0dd23a1a88296266f9f6280750e97e87bc65bff7 --- /dev/null +++ b/2024/LHRS-Bot_ Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model/2d33d60a-efea-43ed-9423-c305740d4cd0_model.json @@ -0,0 +1,2482 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.24, + 0.141, + 0.764, + 0.207 + ], + "angle": 0, + "content": "LHRS-Bot: Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.236, + 0.776, + 0.253 + ], + "angle": 0, + "content": "Dilxat Muhtar†, Zhenshi Li†, Feng Gu, Xueliang Zhang*, and Pengfeng Xiao" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.264, + 0.722, + 0.306 + ], + "angle": 0, + "content": "Nanjing University \nLzhenshi@outlook.com, {dmuhtar, gufeng}@smail.nju.edu.com \n{zxl, xiaopf}@nju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.35, + 0.754, + 0.571 + ], + "angle": 0, + "content": "Abstract. The revolutionary capabilities of large language models (LLMs) have paved the way for multimodal large language models (MLLMs) and fostered diverse applications across various specialized domains. In the remote sensing (RS) field, however, the diverse geographical landscapes and varied objects in RS imagery are not adequately considered in recent MLLM endeavors. To bridge this gap, we construct a large-scale RS image-text dataset, LHRs\\(^1\\)-Align, and an informative RS-specific instruction dataset, LHRs-Instruct, leveraging the extensive volunteered geographic information (VGI) and globally available RS images. Building on this foundation, we introduce LHRs-Bot, an MLLM tailored for RS image understanding through a novel multi-level vision-language alignment strategy and a curriculum learning method. Additionally, we introduce LHRs-Bench, a benchmark for thoroughly evaluating MLLMs' abilities in RS image understanding. Comprehensive experiments demonstrate that LHRs-Bot exhibits a profound understanding of RS images and the ability to perform nuanced reasoning within the RS domain\\(^2\\)." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.587, + 0.707, + 0.6 + ], + "angle": 0, + "content": "Keywords: Multimodal \\(\\cdot\\) Large language model \\(\\cdot\\) Remote sensing" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.63, + 0.375, + 0.644 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.666, + 0.786, + 0.772 + ], + "angle": 0, + "content": "The recently advanced large language models (LLMs) have presented remarkable capabilities in engaging in informative conversations and solving complex problems [2,8,11,23,46-48]. To extend the perceptual capabilities of LLMs, multimodal large language models (MLLMs) bring eyes to LLMs. These models, by aligning with visual representations and visual instruction tuning, have demonstrated impressive multimodal instruction following capabilities and can act as a general-purpose interface for a broad range of tasks [1, 3, 22, 28, 30, 33, 64]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.24, + 0.784, + 0.533, + 0.798 + ], + "angle": 0, + "content": "\\(\\dagger\\) Equal contribution, listed in random order." + }, + { + "type": "page_footnote", + "bbox": [ + 0.241, + 0.799, + 0.397, + 0.812 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.241, + 0.813, + 0.59, + 0.826 + ], + "angle": 0, + "content": "LHRS stands for 'Language Helps Remote Sensing'." + }, + { + "type": "page_footnote", + "bbox": [ + 0.241, + 0.826, + 0.785, + 0.84 + ], + "angle": 0, + "content": "2Data, Code and model are available at https://github.com/NJU-LHRS/LHRS-Bot" + }, + { + "type": "list", + "bbox": [ + 0.24, + 0.784, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.128 + ], + "angle": 0, + "content": "D. Muhtar et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.461, + 0.253 + ], + "angle": 0, + "content": "This success in general domain has served as a catalyst for the development of domain specialized MLLMs, with notable applications in biomedicine [29], autonomous driving [36], and robotics [15]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.254, + 0.461, + 0.435 + ], + "angle": 0, + "content": "Among various specialized domains, remote sensing (RS) image understanding is particularly important, as it directly involves the monitoring and analysis of the Earth's surface and environment [5, 41, 51, 55]. However, RS images present significant challenges for holistic image understanding due to the intricate complexity of different objects across diverse landscapes world" + }, + { + "type": "image", + "bbox": [ + 0.479, + 0.144, + 0.791, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.469, + 0.383, + 0.788, + 0.426 + ], + "angle": 0, + "content": "Fig.1: Overall performance comparison between LHRS-Bot and existing MLLMs on different RS image understanding tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.617 + ], + "angle": 0, + "content": "wide, coupled with the variability in visual scales [19, 37, 42]. MLLMs bring together vision and language, offering possibilities for a unified and interactive understanding of RS images [51]. In this direction, several RS-specific MLLMs have been proposed, such as RSGPT [21], GeoChat [25], and SkyEyeGPT [60]. However, they are struggling to fit existing public RS datasets, due to the oversight of exploiting the extensive and worldwide RS features to infuse comprehensive RS visual knowledge into LLMs. Furthermore, these contemporary approaches of bridging vision and language domains tend to focus on the incorporation of high-level visual semantics. This strategy neglects the fact that different levels of visual information are crucial to fully align the linguistic and visual domains, which could facilitate a unified RS image understanding at different granularity [37, 49, 52]. Detailed related works are available in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.618, + 0.788, + 0.814 + ], + "angle": 0, + "content": "To bridge the gap, we present LHRS-Bot, a specialized MLLM for RS, enhanced by globally available volunteered geographical information (VGI) and worldwide RS images. Specifically, we construct a large-scale, semantically rich, and feature-diverse dataset, LHRS-Align, by geographically pairing RS images with plentiful attributed information from the OpenStreetMap VGI database3, and consequently generating image captions using LLMs. Our rigorous data cleaning process, which includes dedduplication, pruning, and semantic balancing, ensures that LHRS-Align comprises 1.15 million meaningful and high-quality RS image-text pairs, laying a solid foundation for embedding rich RS visual knowledge into MLLMs. Furthermore, we construct our instruction dataset, LHRS-Instruct, by reorganizing a broad range of open-source RS datasets into multitask instructional datasets and employing GPT-4 [38] to create complex instructional data from the high-quality subset of LHRS-Align. Building on these" + }, + { + "type": "page_footnote", + "bbox": [ + 0.238, + 0.825, + 0.482, + 0.841 + ], + "angle": 0, + "content": "3https://www.openstreetmap.org/" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LHRS-Bot" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "datasets, we propose a novel RS-specific MLLM, LHRS-Bot, by implementing a novel bridging strategy for efficiently summarizing multi-level visual representations, alongside a curriculum learning methodology. LHRS-Bot achieves state-of-the-art performance across a variety of RS image understanding tasks (Fig. 1) and demonstrates remarkable abilities to detect intricate objects, engage with human conversation, and provide insights from visual information within RS images (Fig. 5)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.328 + ], + "angle": 0, + "content": "Finally, we manually construct a high-quality benchmark, LHRS-Bench, to facilitate the RS community in evaluating the RS-domain MLLMs across diverse evaluation dimensions. LHRS-Bench includes 690 single-choice questions, spanning 5 top-level evaluation dimensions including 11 fine-grained categories to facilitate a comprehensive, objective, and quantitative RS-specific evaluation." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.33, + 0.695, + 0.343 + ], + "angle": 0, + "content": "The main contributions of our work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.354, + 0.784, + 0.413 + ], + "angle": 0, + "content": "1. We introduce LHRS-Align, a large-scale RS image-text dataset, created through a meticulous data generation pipeline using open-source global geographic data. LHRS-Align showcases a broad spectrum of RS semantic visual information, establishing a solid foundation for RS-specific MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.416, + 0.784, + 0.473 + ], + "angle": 0, + "content": "2. We create LHRS-Instruct, a multimodal instruction-following dataset tailored for RS image understanding. LHRS-Instruct not only contains instruction data to serve for various RS image understanding tasks but also contains complex visual reasoning data generated by GPT-4." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.476, + 0.784, + 0.562 + ], + "angle": 0, + "content": "3. We propose LHRS-Bot, an MLLM specifically designed for the RS domain. LHRS-Bot employs a novel bridging strategy and leverages a curriculum learning approach to fully exploit the inherent knowledge of the proposed datasets. LHRS-Bot demonstrates superior performance across various RS image understanding tasks, along with the ability to follow complex instructions." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.566, + 0.784, + 0.624 + ], + "angle": 0, + "content": "4. We establish LHRS-Bench, a benchmark aimed at providing a comprehensive and systematic evaluation of RS MLLMs. LHRS-Bench comprises 690 meticulously crafted single-choice questions, encompassing a broad spectrum of 11 fine-grained dimensions." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.354, + 0.784, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.651, + 0.329, + 0.667 + ], + "angle": 0, + "content": "2 Dataset" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.684, + 0.688, + 0.699 + ], + "angle": 0, + "content": "2.1 Large-Scale RS Alignment Dataset—LHRS-Align" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.71, + 0.784, + 0.77 + ], + "angle": 0, + "content": "LHRS-Align is developed by initially geo-aligning orthorectified RS images with geographic vector features from the free spatial database, OSM. The attributes of these features are then utilized to generate image captions employing an LLM. The detailed pipeline is illustrated in Fig. 2 and introduced as below." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.771, + 0.785, + 0.817 + ], + "angle": 0, + "content": "Data Preparation. The RS images of LHRS-Align are sourced from Google Earth imagery4, encompassing a worldwide coverage with a spatial resolution of 1 meter. It should be noted that our proposed pipeline is applicable to any RS" + }, + { + "type": "page_footnote", + "bbox": [ + 0.238, + 0.825, + 0.473, + 0.84 + ], + "angle": 0, + "content": "4https://www.google.com/earth/" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.128 + ], + "angle": 0, + "content": "D. Muhtar et al." + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.145, + 0.788, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.325, + 0.653, + 0.34 + ], + "angle": 0, + "content": "Fig. 2: Pipeline of constructing LHRs-Align." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.788, + 0.55 + ], + "angle": 0, + "content": "imagery with geographic coordinates. The geographic features are gathered from OSM, a crowdsourced VGI database containing abundant geographical objects. Each object (feature) in the OSM database is annotated as the digitization of the object geometry (e.g., polygons, lines, and points) as well as the semantic tags. A tag is a key-value pair that provides information about a geographic object5, for example, \"landuse=residential\". Thanks to the free access to the globally encompassing data sources, our pipeline can produce a dataset that offers a vast and comprehensive variety on a worldwide scale. For the current version of LHRS-Align, we collect RS images and OSM features within a scope of \\(0.3^{\\circ}\\) longitude \\(\\times 0.28^{\\circ}\\) latitude for each of the selected cities from all over the world. We customize our own OSM database to streamline the subsequent feature selection process. More information about our data source and database is introduced in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Geo-alignment between Images and Geographic Features. We iterate through our customized database to identify qualified features serving as \"anchors\". Based on the geographical extent of each anchor feature, the corresponding RS image is acquired, which is a square bounding box centered on the feature. These images subsequently undergo a stream of postprocessing and pruning procedures. Following this, due to the possibility of multiple features being present within a single image extent, all the features within each image are further identified. At this point, we have established a raw data pool of around 4 million pairs, in which one RS image corresponds to multiple geographical features. Further details for this processes are provided in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.706, + 0.788, + 0.812 + ], + "angle": 0, + "content": "Attribute Pruning and Semantic Balancing. In the OSM database, all the features are assigned with their semantic attributes, that is, multiple key-value tags. However, not all the keys are favorable information for visual identification, such as website and postal code. Hence, we filter out irrelevant keys through a pipeline as detailed in supplementary material. Fig. 2 provides an example showcasing an image along with its geographical features and their corresponding attributes. Furthermore, the OSM database exhibits considerable fea" + }, + { + "type": "page_footnote", + "bbox": [ + 0.238, + 0.826, + 0.62, + 0.84 + ], + "angle": 0, + "content": "5https://wiki.openstreetmap.org/wiki/Map_features" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LHRS-Bot" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.236 + ], + "angle": 0, + "content": "ture imbalance, notably dominated by geographical features such as residential or farmland. Given the importance of balanced semantic information for effective vision-language alignment [40,54], we implement a data balancing method, as detailed in supplementary material, to create a more semantically balanced dataset. As a result, this balancing process retains 1.15 million images remaining out of the original 4 million." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.495 + ], + "angle": 0, + "content": "Caption Generation with LLM. The use of image-caption pairs to incorporate visual knowledge into LLMs has been validated as an effective approach [3, 14, 29, 56, 64]. However, using plain image key-value labels can be problematic, as they lack grammatical structure, potentially disrupting the generative and comprehension abilities of LLMs. Similarly, rule-based caption generation strategies may introduce significant redundancy. Therefore, we opt to use open-source LLMs to generate image captions by summarising the key-value tags of the corresponding images. We experiment with various in-context learning examples and LLMs, ranging in size from 7B to 70B. After comprehensive testing by a domain expert, we select Vicuna-v1.5-13B [10] for caption generation as it achieves the best trade-offs in terms of generation efficiency and quality. Detailed information about the in-context examples and the specific deployment of Vicuna-v1.5-13B [10] can be found in supplementary material. Our final dataset, referred as LHRS-Align, comprises approximately 1.15 million image-caption pairs. Each pair is accompanied with additional metadata, such as image resolution and geographical location (country and city). Several examples of LHRS-Align are also presented in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.515, + 0.766, + 0.53 + ], + "angle": 0, + "content": "2.2 RS-Specific Instruction-Following Dataset—LHRS-Instruct" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.784, + 0.568 + ], + "angle": 0, + "content": "We craft a multimodal instruction-following dataset LHRS-Instruct by utilizing two public image-caption datasets and our LHRS-Align dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Instruction-Following Dataset from Public RS Caption Dataset. We start by meticulously evaluating and selecting all image_caption datasets within the RS domain. Our selection process focuses on the quality of captions and involves a thorough human evaluation. This requires each image to be accompanied by five distinct captions, with each caption uniquely detailing different aspects of the image. Based on this criterion, we identify RSITMD [58] and NWPU [9] as the datasets with superior caption quality. Subsequent to selection, we design a rigorous data cleaning process, which involves deduplication, computation of caption lengths, and assessment of relevance through CLIP-score calculations. This results in the extraction of 3.4K and 30K high-quality image_caption pairs from RSITMD and NWPU, respectively. We then use several in-context examples along with the five captions per image for prompting Vicuna-v1.5-13B [10] to generate a conversation that specifically focuses on the content of the objects in the image. Finally, we filter out the most relevant conversations through pre-defined rules and evaluations from three domain experts. This process results in 2.9K and 25K instruction data from these two datasets, respectively. For more detailed information on the in-context examples used and the data filtering criteria, please refer to the supplementary material." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.128 + ], + "angle": 0, + "content": "D. Muhtar et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.374 + ], + "angle": 0, + "content": "Instruction-Following Dataset from LHRS-Align. Considering the limitations of public image-caption datasets, which typically lack bounding box annotations, and the constrained capabilities of Vicuna-v1.5-13B [10], the generated conversations might omit crucial details such as the locations and quantity of objects in the images, along with visual reasoning information. To mitigate this issue, we meticulously select 15K representative RS images that are rich in attribute information from our LHRS-Align dataset. We then proceed to calculate the bounding boxes for various objects within these images, utilizing the built-in spatial coordinates of different attributes in the OSM database. Leveraging this enriched data, we prompt GPT-4 to craft complex instruction data that incorporates visual reasoning, detailed image descriptions, and conversations about the location and number of objects. The final step involves a manual cleaning of the generated results. This process yields a dataset comprising 0.9K samples about visual reasoning, 4K samples about detailed image description, and 7K samples about conversation. For more details, please refer to supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.408, + 0.755, + 0.423 + ], + "angle": 0, + "content": "2.3 RS-specific MLLM Evaluation Benchmark—LHRS-Bench" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.521 + ], + "angle": 0, + "content": "We construct LHRS-Bench to evaluate the performance of MLLMs in the RS domain. LHRS-Bench employs hierarchical taxonomies to comprehensively evaluate MLLMs across various dimensions (Tab. 1). To prevent data leakage, we abstain from using public RS datasets and instead meticulously collect images from Google Earth. In terms of question format, we utilize single-choice questions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.488, + 0.687 + ], + "angle": 0, + "content": "[16, 26, 27, 34] with 2 to 4 candidate answers, enabling a more quantitative and objective evaluation compared to open-ended questions [4, 6, 33, 57]. Every question and answer within LHRS-Bench are manually crafted to guarantee precision and reliability. Totally, LHRS-Bench comprises 108 images and 690 questions. Some examples of LHRS-Bench are shown in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.49, + 0.84 + ], + "angle": 0, + "content": "The evaluation dimensions of LHRs-Bench encompass 5 top-level dimensions, further delineated into 11 fine-grained sub-dimensions, as shown in Tab. 1. These top-level dimensions include recognition, imagery, spatial-awareness, quantity, and reasoning. For detailed explanations about each dimensions, please refer to supplementary material." + }, + { + "type": "table_caption", + "bbox": [ + 0.497, + 0.536, + 0.788, + 0.647 + ], + "angle": 0, + "content": "Table 1: Overview of the dataset distribution. LHRS-Bench contains 108 RS images with 690 question-answer pairs, covering five major evaluation dimensions and 11 sub-dimensions of questions. Each sample is a human-annotated single-choice question that may involve one or more evaluation dimensions simultaneously." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.648, + 0.77, + 0.825 + ], + "angle": 0, + "content": "
DimensionSub-dimensionCount
RecognitionIdentity634
Color113
Orientation39
Shape37
Area75
ImageryResolution21
Modality23
Spatial-awarenessLocation204
Distance22
Quantity-137
Reasoning-46
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LHRS-Bot" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.394, + 0.788, + 0.451 + ], + "angle": 0, + "content": "Fig. 3: Architecture of the proposed LHRS-Bot. LHRS-Bot employs learnable queries in conjunction with a vision perceiver to summarize multi-level visual representations. These representations are subsequently concatenated with language token embeddings for processing within the LLM." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.463, + 0.33, + 0.478 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.788, + 0.575 + ], + "angle": 0, + "content": "We present LHRS-Bot to harness the full potential of LLMs and general visual encoders for enhanced RS images understanding and reasoning. This section delves into each components of LHRS-Bot, detailing how we achieve fully alignment between vision and language models across multiple levels. Subsequently, we explain the curriculum training strategy for training LHRS-Bot." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.6, + 0.427, + 0.613 + ], + "angle": 0, + "content": "3.1 Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.787, + 0.672 + ], + "angle": 0, + "content": "The overall architecture of LHRS-Bot is presented in Fig. 3. LHRS-Bot is primarily composed of three components: a universal vision encoder, a vision perceiver, and a foundational LLM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Vision Encoder. LHRS-Bot utilizes the pre-trained CLIP visual encoder ViT-L/14 [40] with input resolution of \\(224 \\times 224\\) pixel for encoding images. The prior works [33,56,64] usually utilize the final hidden feature for abstracting the semantic content of an image. However, several studies have highlighted that features at different levels of a deep learning model embed distinct types of information [17,31,39]. Therefore, we argue that previous image encoding methods fail to sufficiently capture the semantic content of images, leading to a reliance on extensive image-text pairs for vision-language alignment. Consequently, we retain multiple hidden features from various layers of the image encoder. This approach offers a more thorough representation of the underlying image, facilitating a thorough and efficient domain alignment." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.384, + 0.128 + ], + "angle": 0, + "content": "D. Muhtar et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.785, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.318 + ], + "angle": 0, + "content": "Fig. 4: Curriculum learning strategy. We progressively unlock various parameters and introduce data of escalating complexity at different stages, enabling the model to gradually assimilate visual knowledge with increasing difficulty." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.325, + 0.789, + 0.491 + ], + "angle": 0, + "content": "Vision Perceiver. Incorporating multi-level visual features brings extended image tokens, increasing computational overhead and potentially overwhelming language information. To mitigate this issue, we introduce a set of learnable queries for each level of image features. These queries are designed to summarize the semantic information of each level of visual features through a series of stacked cross-attention and MLP layers. Once summarized, these image features are then fed into the LLM. Furthermore, we implement a descending query allocation strategy inspired by the observation that image feature redundancy tends to increase with network depth [7,24]. This strategy allocates more queries to the shallower image feature levels while gradually reducing the number of queries for deeper levels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.789, + 0.539 + ], + "angle": 0, + "content": "Large Language Model. We employ the LLaMA2-7B [48] as the central expert for interpreting visual and language information to output the desired results following the previous practices [25, 29, 43]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.547, + 0.513, + 0.563 + ], + "angle": 0, + "content": "3.2 Curriculum Learning Strategy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.571, + 0.789, + 0.632 + ], + "angle": 0, + "content": "We implement a three-stage curriculum learning strategy by progressively increasing the difficulty of training tasks (Fig. 4). This strategy incrementally aligns the vision encoder with the LLM, ultimately enabling the LLM to solve multiple visual tasks and follow complex vision-language instructions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.632, + 0.789, + 0.768 + ], + "angle": 0, + "content": "Stage 1: Pre-training. In the first stage, we use the large-scale, weakly labelled LHRS-Align dataset (Sec. 2.1) to train the vision perceiver to incorporate a wide range of RS visual knowledge into the LLM. Specifically, for each sample in the LHRS-Align dataset, the output visual embedding tokens \\( X_{v} \\) obtained from the vision perceiver are concatenated with the language embedding tokens \\( X_{l} \\) from the corresponding image caption. The concatenated result is then fed into the LLM. The objective of this stage is to maximise the conditional probability \\( P(X_{l}|X_{v};\\theta) \\) of generating the caption tokens \\( X_{l} \\) given the prefix visual tokens \\( X_{v} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.778, + 0.785, + 0.837 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\log P (X _ {l} | X _ {v}; \\theta) \\\\ = \\sum_ {i = 1} ^ {L} \\log P \\left(x _ {i} \\mid X _ {v}, X _ {l, < i}; \\theta\\right), \\tag {1} \\\\ \\end{array}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LHRS-Bot" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "where \\( L \\) is the length of caption tokens, \\( \\theta \\) represents the trainable parameters, and \\( X_{l, < i} \\) denotes the generated tokens prior to the current prediction token \\( x_i \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.39 + ], + "angle": 0, + "content": "Stage 2: Multi-task Pre-training. In the second stage, we further train the vision perceiver and efficiently fine-tune the LLM using LoRA [20] on multitask instruction datasets to improve multimodal task solving capabilities of LLM. The construction of the multi-task instruction datasets begins with the collection of diverse public datasets in the RS domain, encompassing tasks such as classification, VQA, and visual grounding. These collected datasets are then meticulously transformed into instruction-based formats using manually designed templates. More detailed information about these multi-task datasets and the templates used can be found in the supplementary material. Moreover, we integrate the instruction data from NWPU, RSITMD and the 4K detailed descriptions data from LHRS-Instruct (Sec. 2.2) to further enrich the LLM with fine-grained RS visual knowledge. The objective for this stage remains similar to Eq. (1) from stage 1, albeit with an adaptation that the generated tokens are now additionally conditioned on the instruction tokens \\( X_{\\text{instruct}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.403, + 0.785, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\log P (X _ {l} | X _ {v}, X _ {\\text {i n s t r u c t}}; \\theta) \\\\ = \\sum_ {i = 1} ^ {L} \\log P \\left(x _ {i} \\mid X _ {v}, X _ {\\text {i n s t r u c t}}, X _ {l, < i}; \\theta\\right). \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.583 + ], + "angle": 0, + "content": "Stage 3: Supervised Fine-tuning. In the final stage, our focus shifts towards enhancing the conversational and reasoning capabilities of our model by incorporating more complex instruction data. We utilize all the instruction data from the LHRS-Instruct dataset (Sec. 2.2) and randomly select 20K data from the LLaVA complex reasoning dataset [29]. Additionally, we mix the above datasets with the multi-task datasets from stage 2 using a lower data sampling ratio to further unleash the multi-task problem-solving ability of LHRS-Bot." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.598, + 0.367, + 0.615 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.634, + 0.788, + 0.727 + ], + "angle": 0, + "content": "We conduct comprehensive experiments across image classification, VQA, and visual grounding tasks within the RS domain. These experiments are designed to rigorously validate the superior multi-task solving capabilities of LHRS-Bot. Furthermore, we evaluate the performance of various LLMs in the RS domain using the proposed LHRS-Bench benchmark. We present both quantitative and quality results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.751, + 0.429, + 0.767 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Dataset. We select diverse RS datasets for providing a comprehensive assessment across different tasks within the RS field: seven RS classification datasets (namely AID [53], WHU-RS19 [13], NWPU [9], SIRI-WHU [65], EuroSAT [18], METER-ML [63], and fMoW [12]), two VQA datasets (LR and HR subsets" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.128 + ], + "angle": 0, + "content": "D. Muhtar et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.271, + 0.145, + 0.731, + 0.16 + ], + "angle": 0, + "content": "Table 2: Classification accuracy of different models on each dataset." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.164, + 0.785, + 0.276 + ], + "angle": 0, + "content": "
MethodAIDWHU-RS19NWPUSIRI-WHUEuroSATMETER-MLfMoWAvg.
LLaVA-1.531.1054.5534.9617.7126.1221.7311.4328.23
MiniGPTv232.9664.8028.1535.4638.5614.295.2031.35
InstructBLIP29.5036.7634.0118.2020.2514.426.7122.84
mPLUG-OWL248.7972.6646.5854.8333.2936.2717.8544.32
QWen-VL-Chat55.3072.2542.7354.5826.4238.776.8942.42
InternLM-XComposer51.6172.8947.5146.8339.7040.2111.2844.29
LHRS-Bot91.2693.1783.9462.6651.4069.8156.5671.83
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.293, + 0.785, + 0.355 + ], + "angle": 0, + "content": "of RSVQA [35]), and two visual grounding datasets (RSVG [45] and DIOR-RSVG [59]). We exclude any data from the test sets that overlap with the training set to avoid data leakage. More details about each test dataset are presented in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.808, + 0.456 + ], + "angle": 0, + "content": "Baseline. We evaluate LHRS-Bot and various powerful open-source MLLMs, including LLaVA-v1.5 [32], MiniGPT-v2 [64], InstructBLIP [14], mPLUG-Owl2 [56], QWen-VL-Chat [3], and InternLM-Xcomposer [61], on different tasks. Furthermore, we also include the results of RSGPT [21], GeoChat [25] and SkyEyeGPT [60] from the original papers for comparisons whenever possible. Unless otherwise noted, we utilized 7B variants for all our experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.788, + 0.616 + ], + "angle": 0, + "content": "Implementation. We extract hidden states at layers \\(\\{i = L / 3, j = 2L / 3, k = L - 1\\}\\) for summarizing through vision perceiver, where \\(L\\) denotes the number of layers in the vision encoder. Our vision perceiver implementation comprises six layers of cross-attention and MLP, with each hidden state assigned \\(\\{64, 48, 32\\}\\) learnable queries. We apply LoRA module to every linear layer of the LLM, setting the rank \\(r\\) and the scale factor \\(\\alpha\\) for LoRA to 128 and 256, respectively. Moreover, we add task identifier [CLS], [VQA], and [VG] for classification, VQA, and visual grounding tasks, respectively, for enabling LHRS-Bot to better distinguish different tasks [64]. For additional details on the hyperparameters used at each stage, please refer to supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.635, + 0.46, + 0.65 + ], + "angle": 0, + "content": "4.2 Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.706 + ], + "angle": 0, + "content": "We evaluate different MLLMs using greedy decoding with the same prompt for each task and report the average result of two independent trials. Details about the evaluation prompts are provided in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.713, + 0.788, + 0.807 + ], + "angle": 0, + "content": "Classification. The classification results across different datasets are presented in Tab. 2. These results show that LHRS-Bot outperforms all competing models in all seven classification datasets, with a notable margin of \\(27.51\\%\\) higher average accuracy than the second ranked method. It is worth mentioning that, except for NWPU, METER-ML, and fMoW, the remaining datasets are completely absent from the multi-task training dataset6, demonstrating the strong" + }, + { + "type": "page_footnote", + "bbox": [ + 0.215, + 0.811, + 0.788, + 0.842 + ], + "angle": 0, + "content": "However, we have no means of confirming whether these data appeared in the LHRs-Align dataset, considering that images in LHRs-Align, along with several of the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LHRS-Bot" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.186 + ], + "angle": 0, + "content": "Table 3: Accuracy of different models on each visual question answering dataset. Following the practice of prior works [21, 25, 60], we omit area and count questions during evaluation." + }, + { + "type": "table", + "bbox": [ + 0.244, + 0.192, + 0.755, + 0.343 + ], + "angle": 0, + "content": "
MethodRSVQA-LRRSVQA-HR
Rural/UrbanPresenceCompareAvg.PresenceCompareAvg.
LLaVA-1.559.2273.1665.1965.8648.9659.0253.99
MiniGPTv260.0251.6467.6459.7768.3464.7166.53
InstructBLIP62.6248.8365.9259.1262.6362.9062.77
mPLUG-Owl257.9974.0463.6965.2447.6058.4753.04
QWen-VL-Chat62.0047.6566.5458.7361.7565.9863.87
InternLM-XCompose59.0066.7452.9159.5567.7966.6267.21
RSGPT94.0091.1791.7092.2990.9290.0290.47
GeoChat91.0990.3394.0091.8158.4583.1970.82
SkyEyeGPT88.9388.6375.0084.1680.0080.1382.56
LHRS-Bot89.0788.5190.0089.1992.5792.5392.55
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.367, + 0.785, + 0.442 + ], + "angle": 0, + "content": "generalization ability of LHRS-Bot. In addition, in the context of the fine-grained RS classification dataset fMoW, our method significantly surpasses other models, leading by an impressive \\(40\\%\\) on average. This highlights the critical role of continue pre-training in the RS domain and the effectiveness of our alignment dataset LHRS-Align." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.479, + 0.785, + 0.6 + ], + "angle": 0, + "content": "Visual Question Answering. We report the results of different models on RSVQA-LR and RSVQA-HR datasets in Tab. 3. LHRS-Bot demonstrates performance on par with the previously leading RS domain-specific models on the VQA task. Notably, LHRS-Bot achieves comparable results to RSGPT, despite RSGPT being fine-tuned for five epochs on the RSVQA datasets. Furthermore, LHRS-Bot sets a new record in the RSVQA-HR task with high-resolution RS images, showcasing its superior visual recognition abilities with high-resolution RS images." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.625, + 0.486, + 0.792 + ], + "angle": 0, + "content": "Visual Grounding. The results for the visual grounding task are shown in Tab. 4, where LHRS-Bot outperforms other counterparts. Notably, despite SkyEyeGPT being trained on a more extensive object detection dataset, LHRS-Bot demonstrates comparable performance. This suggests that LHRS-Bot has effectively acquired fine-grained visual information throughout the training process." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.626, + 0.784, + 0.695 + ], + "angle": 0, + "content": "Table 4: Comparison of visual grouting tasks. Accuracy is calculated based on whether the predicted bounding box overlaps with the ground-truth box by more than 0.5 IoU." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.697, + 0.765, + 0.773 + ], + "angle": 0, + "content": "
MethodRSVGDIOR-RSVGAvg.
QWen-VL-Chat44.7680.6562.71
MiniGPTv246.6485.9966.32
SkyEyeGPT70.5088.5979.55
LHRS-Bot73.4588.1080.78
" + }, + { + "type": "page_footnote", + "bbox": [ + 0.217, + 0.812, + 0.784, + 0.838 + ], + "angle": 0, + "content": "classification datasets (AID, WHU-RS19, and SIRI-WHU), are collected from Google Earth." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.128 + ], + "angle": 0, + "content": "D. Muhtar et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.172 + ], + "angle": 0, + "content": "Table 5: Comparison results on the LHRS-Bench dataset with different open-source and closed-source MLLMs." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.185, + 0.787, + 0.289 + ], + "angle": 0, + "content": "
MethodIdentityColorOrientationShapeAreaResolutionModalityLocationDistanceQuantityReasoningAvg.
LLaVA-1.542.9030.9715.3843.2444.0023.8021.7435.7827.2721.8954.3537.45
MiniGPTv220.1915.932.568.1113.3314.294.3515.2013.649.4936.9616.88
InstructBLIP20.3522.125.1324.3222.6738.1000.0013.7322.8513.1430.4318.87
mPLUG-OWL241.4845.1312.8240.5432.0033.3300.0030.8813.6432.1254.3437.00
QWen-VL-Chat31.3832.747.6932.4329.3319.054.3524.0221.4521.1756.5228.65
InternLM-XComposer34.3830.977.6937.8426.6700.008.7033.3322.7318.2565.2231.09
GPT-4-Turbo50.3243.3610.2667.5744.0042.8639.1342.1622.7321.8971.7444.56
Claude-3-Opus41.0136.2817.9564.8641.3366.6739.1334.8045.4526.2854.3539.08
LHRS-Bot44.6442.4823.0867.5725.3342.8632.3556.5222.7318.2573.9139.38
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.307, + 0.49, + 0.321 + ], + "angle": 0, + "content": "4.3 Evaluation on LHRS-Bench" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.787, + 0.446 + ], + "angle": 0, + "content": "We evaluate LHRS-Bot with different open-source and closed-source MLLMs on our LHRS-Bench benchmark. For each question within the benchmark, choices are shuffled, and the generation process is repeated four times. An answer is considered incorrect if any attempt yields an incorrect response. Same prompts are used for evaluation, as detailed in supplementary material. We consider the model answers correctly if it outputs the correct choice or the answer corresponding to the correct choice. We calculate accuracy as our evaluation metric." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.787, + 0.584 + ], + "angle": 0, + "content": "The results presented in Tab. 5 indicate that LHRS-Bot outperforms all other open-source models. It is worth noting that, LHRS-Bot greatly outperforms other open-source models in terms of the evaluation dimensions of spatial resolution and image modality, even though these aspects were not explicitly trained for in the models, highlighting the effectiveness of RS-specific MLLM training. Additionally, LHRS-Bot outperforms the strong closed-source model Claude and even matches the performance of GPT-4V. However, there is still significant room for improvement in MLLM capabilities within the RS domain. Examples of evaluation results are available on supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.607, + 0.449, + 0.622 + ], + "angle": 0, + "content": "4.4 Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.787, + 0.761 + ], + "angle": 0, + "content": "We conduct qualitative demonstrations to illustrate the extensive capabilities of LHRS-Bot in interpreting free-form RS images, as showcased in Fig. 5. The results demonstrate that LHRS-Bot not only comprehends the overall content of an image but also exhibits a strong ability for discerning fine-grained details. Moreover, LHRS-Bot possesses a perceptive understanding of the orientation, position, and quantity of objects within an image. Furthermore, LHRS-Bot is capable of making reasoned inferences from the images and acknowledges its limitations by indicating uncertainty when it cannot provide a definitive answer." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.789, + 0.411, + 0.804 + ], + "angle": 0, + "content": "4.5 Ablation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.822, + 0.787, + 0.853 + ], + "angle": 0, + "content": "In this subsection, we carry out extensive ablation experiments to demonstrate the effectiveness of our dataset as well as the model design." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LHRS-Bot" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.788, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.287, + 0.354, + 0.714, + 0.368 + ], + "angle": 0, + "content": "Fig. 5: Qualitative examples between real users and LHRS-Bot." + }, + { + "type": "table_caption", + "bbox": [ + 0.314, + 0.377, + 0.687, + 0.391 + ], + "angle": 0, + "content": "Table 6: Performance with different alignment dataset." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.403, + 0.789, + 0.455 + ], + "angle": 0, + "content": "
AIDWHU-RS19ClassificationVQAVisual Grounding
NWPUSIRI-WHUEuroSATMETER-MLfMoWRSVQA-LRRSVQA-HRRSVGDIOR-RSVG
RS5M86.3196.7582.8060.2054.9264.7742.6878.3691.6570.5887.12
SkyScript88.2190.2678.0151.4343.6650.0231.2788.7991.2471.2585.05
LHRS-Align91.2693.1783.9462.6651.4069.8156.5689.1992.5573.4588.10
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.788, + 0.588 + ], + "angle": 0, + "content": "Effectiveness of LHRS-Align dataset. There are already many large-scale image-caption datasets in the field of RS [50,62]. To evaluate how different alignment datasets impact model performance, we substitute our LHRS-Align dataset with various image-caption datasets and repeat the entire training process. This process is then evaluated across three distinct tasks: classification, VQA, and visual grounding. The experimental setup remains consistent with Sec. 4.1, with the exception of the dataset used in the first stage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.589, + 0.789, + 0.739 + ], + "angle": 0, + "content": "The results are presented in Tab. 6. Notably, despite possessing the smallest dataset (RS5M: 5M, SkyScript: 1.5M, LHRS-Align : 1.15M), LHRS-Align outperforms other settings. It is worth mentioning that both SkyScript and LHRS-Align, which leverage RS images combined with the OSM database, achieve results comparable to or better than the even larger RS5M, which is derived from publicly available natural image datasets, such as LAION [44]. This highlights the importance of domain-specific data collection. Additionally, superior performance of LHRS-Align over SkyScript suggests that enriched captions and a meticulous data cleaning pipeline can substantially enhance the effectiveness of MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Effectiveness of Model Design. We evaluate the effectiveness of designs involving multi-level visual hidden representations as well as descending query allocation strategies. Specifically, our evaluation focuses on comparing with the single level visual hidden state, projecting it through a two-layer MLP or a vision perceiver. We also perform comparisons between multi-level visual hidden" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.128 + ], + "angle": 0, + "content": "D. Muhtar et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.264, + 0.145, + 0.737, + 0.16 + ], + "angle": 0, + "content": "Table 7: Performance comparison across various architectural designs." + }, + { + "type": "table", + "bbox": [ + 0.249, + 0.172, + 0.754, + 0.273 + ], + "angle": 0, + "content": "
MethodMulti Level Number of Query ClassificationVQA Visual Grounding
MLP✘-64.4690.06
Vision Perceiver✘6463.6983.30
Vision Perceiver✓[64,64,64]63.7487.88
Vision Perceiver✓[32,48,64]60.6887.11
Token Merging✓6449.5181.49
Vision Perceiver✓[64,48,32]65.4390.14
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.286, + 0.787, + 0.405 + ], + "angle": 0, + "content": "representations, employing either identical queries for each level or incremental query allocation strategies. Building on the concept introduced by the token merging [7], we explore the potential of integrating this strategy with a vision perceiver to efficiently leverage multi-level representations. To minimize experimental costs, our training process comprises just two stages: firstly, we employ LHRS-Align to train the bridge layer, which aligns the vision and language domains. Subsequently, we train the model using the LHRS-Instruct and multi-task instruction datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.788, + 0.575 + ], + "angle": 0, + "content": "We report the average result on each task in Tab. 7. It is evident that the approach employed by LHRS-Bot achieves superior performance. The results further illustrate that, although the extraction of a singular visual representation performs better in tasks such as classification and VQA tasks, which depends on high-level visual semantics, it falls short in visual grounding tasks that require detailed visual information. Furthermore, employing the same number of queries to summarise different levels of visual information does not lead to better results. In contrast, adopting a diminishing allocation strategy yields superior results. This further demonstrates the large redundancy among the deep visual tokens [7]. Lastly, the token merging strategy proves ineffective, possibly due to the loss of crucial information during the merging process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.604, + 0.36, + 0.62 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.787, + 0.733 + ], + "angle": 0, + "content": "We propose LHRS-Bot, an MLLM for the RS domain. To unleash the potential of LLMs for RS image understanding, we curate a large-scale dataset, LHRS-Align, for RS-specific alignment, and LHRS-Instruct, a multimodal instruction-following dataset to enhance LHRS-Bot's instruction-following capabilities. Additionally, we introduce LHRS-Bench, a benchmark providing a systematic evaluation framework for RS MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In comprehensive comparison experiments, LHRS-Bot demonstrates exceptional RS domain knowledge, surpassing both general-purpose and RS-specific MLLMs. Despite its impressive performance, LHRS-Bot shares limitations common to MLLMs, such as susceptibility to hallucinations. We believe that developing higher-quality alignment and instructional datasets, along with improved training strategies, can further enhance the performance of MLLMs for interpreting RS images." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LHRS-Bot" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.271 + ], + "angle": 0, + "content": "This research was supported by the National Natural Science Foundation of China under Grant 42071297, and in part by the AI & AI for Science Project of Nanjing University under Grant 02091480605203. We are grateful to High Performance Computing Center of Nanjing University for their help on GPU resources. We also would like to thank the anonymous reviewers for their constructive comments." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.298, + 0.323, + 0.314 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.333, + 0.787, + 0.389 + ], + "angle": 0, + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.787, + 0.418 + ], + "angle": 0, + "content": "2. Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., et al.: Qwen technical report. arXiv preprint arXiv:2309.16609 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.787, + 0.46 + ], + "angle": 0, + "content": "3. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.462, + 0.787, + 0.502 + ], + "angle": 0, + "content": "4. Bai, S., Yang, S., Bai, J., Wang, P., Zhang, X., Lin, J., Wang, X., Zhou, C., Zhou, J.: Touchstone: Evaluating vision-language models by language models. arXiv preprint arXiv:2308.16890 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.504, + 0.787, + 0.544 + ], + "angle": 0, + "content": "5. Bashmal, L., Bazi, Y., Melgani, F., Al Rahhal, M.M., Al Zuair, M.A.: Language integration in remote sensing: Tasks, datasets, and future directions. IEEE Geoscience and Remote Sensing Magazine (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.546, + 0.787, + 0.587 + ], + "angle": 0, + "content": "6. Bitton, Y., Bansal, H., Hessel, J., Shao, R., Zhu, W., Awadalla, A., Gardner, J., Taori, R., Schimdt, L.: Visit-bench: A benchmark for vision-language instruction following inspired by real-world use. arXiv preprint arXiv:2308.06595 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.589, + 0.787, + 0.615 + ], + "angle": 0, + "content": "7. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.617, + 0.787, + 0.657 + ], + "angle": 0, + "content": "8. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.787, + 0.686 + ], + "angle": 0, + "content": "9. Cheng, G., Han, J., Lu, X.: Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE 105(10), 1865-1883 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.688, + 0.787, + 0.741 + ], + "angle": 0, + "content": "10. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.744, + 0.787, + 0.799 + ], + "angle": 0, + "content": "11. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. Journal of Machine Learning Research 24(240), 1-113 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "12. Christie, G., Fendley, N., Wilson, J., Mukherjee, R.: Functional map of the world. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6172-6180 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.219, + 0.333, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.128 + ], + "angle": 0, + "content": "D. Muhtar et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "13. Dai, D., Yang, W.: Satellite image classification via two-layer sparse coding with biased image representation. IEEE Geoscience and Remote Sensing Letters 8(1), 173-176 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "14. Dai, W., Li, J., Li, D., Tiong, A., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arxiv 2023. arXiv preprint arXiv:2305.06500" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.787, + 0.271 + ], + "angle": 0, + "content": "15. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.787, + 0.311 + ], + "angle": 0, + "content": "16. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.311, + 0.787, + 0.352 + ], + "angle": 0, + "content": "17. Ghiasi, A., Kazemi, H., Borgnia, E., Reich, S., Shu, M., Goldblum, M., Wilson, A.G., Goldstein, T.: What do vision transformers learn? a visual exploration. arXiv preprint arXiv:2212.06727 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.352, + 0.787, + 0.407 + ], + "angle": 0, + "content": "18. Helber, P., Bischke, B., Dengel, A., Borth, D.: Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 12(7), 2217-2226 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.407, + 0.787, + 0.448 + ], + "angle": 0, + "content": "19. Hossain, M.D., Chen, D.: Segmentation for object-based image analysis (obia): A review of algorithms and challenges from remote sensing perspective. ISPRS Journal of Photogrammetry and Remote Sensing 150, 115-134 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.448, + 0.787, + 0.488 + ], + "angle": 0, + "content": "20. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.488, + 0.787, + 0.515 + ], + "angle": 0, + "content": "21. Hu, Y., Yuan, J., Wen, C., Lu, X., Li, X.: Rsgpt: A remote sensing vision language model and benchmark. arXiv preprint arXiv:2307.15266 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.515, + 0.787, + 0.556 + ], + "angle": 0, + "content": "22. Huang, S., Dong, L., Wang, W., Hao, Y., Singhal, S., Ma, S., Lv, T., Cui, L., Mohammed, O.K., Liu, Q., et al.: Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.556, + 0.787, + 0.597 + ], + "angle": 0, + "content": "23. Jiang, A.Q., Sablayrolles, A., Roux, A., Mensch, A., Savary, B., Bamford, C., Chaplot, D.S., Casas, D.d.l., Hanna, E.B., Bressand, F., et al.: Mixtral of experts. arXiv preprint arXiv:2401.04088 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.597, + 0.787, + 0.638 + ], + "angle": 0, + "content": "24. Ju, C., Wang, H., Li, Z., Chen, X., Zhai, Z., Huang, W., Xiao, S.: Turbo: Informativity-driven acceleration plug-in for vision-language models. arXiv preprint arXiv:2312.07408 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.638, + 0.787, + 0.678 + ], + "angle": 0, + "content": "25. Kuckreja, K., Danish, M.S., Naseer, M., Das, A., Khan, S., Khan, F.S.: Geochat: Grounded large vision-language model for remote sensing. arXiv preprint arXiv:2311.15826 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.678, + 0.787, + 0.718 + ], + "angle": 0, + "content": "26. Li, B., Ge, Y., Ge, Y., Wang, G., Wang, R., Zhang, R., Shan, Y.: Seed-bench-2: Benchmarking multimodal large language models. arXiv preprint arXiv:2311.17092 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "27. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "28. Li, C., Gan, Z., Yang, Z., Yang, J., Li, L., Wang, L., Gao, J.: Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "29. Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H., Gao, J.: Llava-med: Training a large language-and-vision assistant for biomedicine in one day. arXiv preprint arXiv:2306.00890 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LHRS-Bot" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "30. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "31. Li, M., Wang, S., Zhang, Q.: Visualizing the emergence of intermediate visual patterns in dnns. Advances in Neural Information Processing Systems 34, 6594-6607 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.232, + 0.786, + 0.259 + ], + "angle": 0, + "content": "32. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.26, + 0.787, + 0.285 + ], + "angle": 0, + "content": "33. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "34. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "35. Lobry, S., Marcos, D., Murray, J., Tuia, D.: Rsvqa: Visual question answering for remote sensing data. IEEE Transactions on Geoscience and Remote Sensing 58(12), 8555-8566 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.371, + 0.787, + 0.397 + ], + "angle": 0, + "content": "36. Ma, Y., Cao, Y., Sun, J., Pavone, M., Xiao, C.: Multimodal language model for driving. arXiv preprint arXiv:2312.00438 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.398, + 0.787, + 0.439 + ], + "angle": 0, + "content": "37. Muhtar, D., Zhang, X., Xiao, P., Li, Z., Gu, F.: Cmid: A unified self-supervised learning framework for remote sensing image understanding. IEEE Transactions on Geoscience and Remote Sensing (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.439, + 0.503, + 0.453 + ], + "angle": 0, + "content": "38. OpenAI: Gpt-4 technical report (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.453, + 0.787, + 0.481 + ], + "angle": 0, + "content": "39. Park, N., Kim, W., Heo, B., Kim, T., Yun, S.: What do self-supervised vision transformers learn? arXiv preprint arXiv:2305.00729 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.481, + 0.787, + 0.535 + ], + "angle": 0, + "content": "40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.536, + 0.787, + 0.577 + ], + "angle": 0, + "content": "41. Ratledge, N., Cadamuro, G., de la Cuesta, B., Stigler, M., Burke, M.: Using machine learning to assess the livelihood impact of electricity access. Nature 611(7936), 491-495 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.578, + 0.787, + 0.632 + ], + "angle": 0, + "content": "42. Reed, C.J., Gupta, R., Li, S., Brockman, S., Funk, C., Clipp, B., Keutzer, K., Candido, S., Uytendaele, M., Darrell, T.: Scale-MAE: A scale-aware masked autoencoder for multiscale geospatial representation learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4088-4099 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.633, + 0.787, + 0.674 + ], + "angle": 0, + "content": "43. Saygin Seyfioglu, M., Ikezogwo, W.O., Ghezloo, F., Krishna, R., Shapiro, L.: Quilt-lava: Visual instruction tuning by extracting localized narratives from open-source histopathology videos. arXiv e-prints pp. arXiv-2312 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.675, + 0.787, + 0.73 + ], + "angle": 0, + "content": "44. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.73, + 0.787, + 0.771 + ], + "angle": 0, + "content": "45. Sun, Y., Feng, S., Li, X., Ye, Y., Kang, J., Huang, X.: Visual grounding in remote sensing images. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 404-412 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "46. Team, M.N.: Introducing mpt-7b: A new standard for open-source, commercially usable llms (2023), www.mosaicml.com/blog/mpt-7b, accessed: 2023-05-05" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "47. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.128 + ], + "angle": 0, + "content": "D. Muhtar et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "48. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.243 + ], + "angle": 0, + "content": "49. Wang, D., Zhang, J., Du, B., Xu, M., Liu, L., Tao, D., Zhang, L.: Samrs: Scaling-up remote sensing segmentation dataset with segment anything model. In: Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.284 + ], + "angle": 0, + "content": "50. Wang, Z., Prabha, R., Huang, T., Wu, J., Rajagopal, R.: Skyscript: A large and semantically diverse vision-language dataset for remote sensing. arXiv preprint arXiv:2312.12856 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.325 + ], + "angle": 0, + "content": "51. Wen, C., Hu, Y., Li, X., Yuan, Z., Zhu, X.X.: Vision-language models in remote sensing: Current progress and future trends. arXiv preprint arXiv:2305.05726 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.379 + ], + "angle": 0, + "content": "52. Xia, G.S., Bai, X., Ding, J., Zhu, Z., Belongie, S., Luo, J., Datcu, M., Pelillo, M., Zhang, L.: DOTA: A large-scale dataset for object detection in aerial images. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3974-3983 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.42 + ], + "angle": 0, + "content": "53. Xia, G.S., Hu, J., Hu, F., Shi, B., Bai, X., Zhong, Y., Zhang, L., Lu, X.: AID: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing 55(7), 3965-3981 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.461 + ], + "angle": 0, + "content": "54. Xu, H., Xie, S., Tan, X.E., Huang, P.Y., Howes, R., Sharma, V., Li, S.W., Ghosh, G., Zettlemoyer, L., Feichtenhofer, C.: Demystifying clip data. arXiv preprint arXiv:2309.16671 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.5 + ], + "angle": 0, + "content": "55. Yang, J., Gong, P., Fu, R., Zhang, M., Chen, J., Liang, S., Xu, B., Shi, J., Dickinson, R.: The role of satellite remote sensing in climate change studies. Nature climate change 3(10), 875-883 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.542 + ], + "angle": 0, + "content": "56. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.582 + ], + "angle": 0, + "content": "57. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mmvet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.785, + 0.623 + ], + "angle": 0, + "content": "58. Yuan, Z., Zhang, W., Fu, K., Li, X., Deng, C., Wang, H., Sun, X.: Exploring a fine-grained multiscale method for cross-modal remote sensing image retrieval. arXiv preprint arXiv:2204.09868 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.785, + 0.663 + ], + "angle": 0, + "content": "59. Zhan, Y., Xiong, Z., Yuan, Y.: Rsvg: Exploring data and models for visual grounding on remote sensing data. IEEE Transactions on Geoscience and Remote Sensing 61, 1-13 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.785, + 0.704 + ], + "angle": 0, + "content": "60. Zhan, Y., Xiong, Z., Yuan, Y.: Skyeyegpt: Unifying remote sensing vision-language tasks via instruction tuning with large language model. arXiv preprint arXiv:2401.09712 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.758 + ], + "angle": 0, + "content": "61. Zhang, P., Wang, X.D.B., Cao, Y., Xu, C., Ouyang, L., Zhao, Z., Ding, S., Zhang, S., Duan, H., Yan, H., et al.: Internl m-xcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "62. Zhang, Z., Zhao, T., Guo, Y., Yin, J.: Rs5m: A large scale vision-language dataset for remote sensing vision-language foundation model. arXiv preprint arXiv:2306.11300 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "63. Zhu, B., Lui, N., Irvin, J., Le, J., Tadwalkar, S., Wang, C., Ouyang, Z., Liu, F.Y., Ng, A.Y., Jackson, R.B.: Meter-ml: A multi-sensor earth observation benchmark for automated methane source mapping. arXiv preprint arXiv:2207.11166 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LHRS-Bot" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "64. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "65. Zhu, Q., Zhong, Y., Zhao, B., Xia, G.S., Zhang, L.: Bag-of-visual-words scene classifier with local and global features for high spatial resolution remote sensing imagery. IEEE Geoscience and Remote Sensing Letters 13(6), 747-751 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.232 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/LHRS-Bot_ Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model/2d33d60a-efea-43ed-9423-c305740d4cd0_origin.pdf b/2024/LHRS-Bot_ Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model/2d33d60a-efea-43ed-9423-c305740d4cd0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..16138ba616dff82ce357b1b3bb0fe12798bc73fb --- /dev/null +++ b/2024/LHRS-Bot_ Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model/2d33d60a-efea-43ed-9423-c305740d4cd0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3ef5455a7bea1dd10e1bf23eb75fd5ae41c45781fd1d4be53b6ab3ddbc6d6e2 +size 10987866 diff --git a/2024/LHRS-Bot_ Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model/full.md b/2024/LHRS-Bot_ Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..25ea6192260bab23d01366e2a517236fe1d1f0cd --- /dev/null +++ b/2024/LHRS-Bot_ Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model/full.md @@ -0,0 +1,277 @@ +# LHRS-Bot: Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model + +Dilxat Muhtar†, Zhenshi Li†, Feng Gu, Xueliang Zhang*, and Pengfeng Xiao + +Nanjing University +Lzhenshi@outlook.com, {dmuhtar, gufeng}@smail.nju.edu.com +{zxl, xiaopf}@nju.edu.cn + +Abstract. The revolutionary capabilities of large language models (LLMs) have paved the way for multimodal large language models (MLLMs) and fostered diverse applications across various specialized domains. In the remote sensing (RS) field, however, the diverse geographical landscapes and varied objects in RS imagery are not adequately considered in recent MLLM endeavors. To bridge this gap, we construct a large-scale RS image-text dataset, LHRs $^1$ -Align, and an informative RS-specific instruction dataset, LHRs-Instruct, leveraging the extensive volunteered geographic information (VGI) and globally available RS images. Building on this foundation, we introduce LHRs-Bot, an MLLM tailored for RS image understanding through a novel multi-level vision-language alignment strategy and a curriculum learning method. Additionally, we introduce LHRs-Bench, a benchmark for thoroughly evaluating MLLMs' abilities in RS image understanding. Comprehensive experiments demonstrate that LHRs-Bot exhibits a profound understanding of RS images and the ability to perform nuanced reasoning within the RS domain $^2$ . + +Keywords: Multimodal $\cdot$ Large language model $\cdot$ Remote sensing + +# 1 Introduction + +The recently advanced large language models (LLMs) have presented remarkable capabilities in engaging in informative conversations and solving complex problems [2,8,11,23,46-48]. To extend the perceptual capabilities of LLMs, multimodal large language models (MLLMs) bring eyes to LLMs. These models, by aligning with visual representations and visual instruction tuning, have demonstrated impressive multimodal instruction following capabilities and can act as a general-purpose interface for a broad range of tasks [1, 3, 22, 28, 30, 33, 64]. + +This success in general domain has served as a catalyst for the development of domain specialized MLLMs, with notable applications in biomedicine [29], autonomous driving [36], and robotics [15]. + +Among various specialized domains, remote sensing (RS) image understanding is particularly important, as it directly involves the monitoring and analysis of the Earth's surface and environment [5, 41, 51, 55]. However, RS images present significant challenges for holistic image understanding due to the intricate complexity of different objects across diverse landscapes world + +![](images/106257a820d7b236c17b5b0d3e5f88fe33f6417123e81c6f4d4ed96e396fe0f6.jpg) +Fig.1: Overall performance comparison between LHRS-Bot and existing MLLMs on different RS image understanding tasks. + +wide, coupled with the variability in visual scales [19, 37, 42]. MLLMs bring together vision and language, offering possibilities for a unified and interactive understanding of RS images [51]. In this direction, several RS-specific MLLMs have been proposed, such as RSGPT [21], GeoChat [25], and SkyEyeGPT [60]. However, they are struggling to fit existing public RS datasets, due to the oversight of exploiting the extensive and worldwide RS features to infuse comprehensive RS visual knowledge into LLMs. Furthermore, these contemporary approaches of bridging vision and language domains tend to focus on the incorporation of high-level visual semantics. This strategy neglects the fact that different levels of visual information are crucial to fully align the linguistic and visual domains, which could facilitate a unified RS image understanding at different granularity [37, 49, 52]. Detailed related works are available in supplementary material. + +To bridge the gap, we present LHRS-Bot, a specialized MLLM for RS, enhanced by globally available volunteered geographical information (VGI) and worldwide RS images. Specifically, we construct a large-scale, semantically rich, and feature-diverse dataset, LHRS-Align, by geographically pairing RS images with plentiful attributed information from the OpenStreetMap VGI database3, and consequently generating image captions using LLMs. Our rigorous data cleaning process, which includes dedduplication, pruning, and semantic balancing, ensures that LHRS-Align comprises 1.15 million meaningful and high-quality RS image-text pairs, laying a solid foundation for embedding rich RS visual knowledge into MLLMs. Furthermore, we construct our instruction dataset, LHRS-Instruct, by reorganizing a broad range of open-source RS datasets into multitask instructional datasets and employing GPT-4 [38] to create complex instructional data from the high-quality subset of LHRS-Align. Building on these + +datasets, we propose a novel RS-specific MLLM, LHRS-Bot, by implementing a novel bridging strategy for efficiently summarizing multi-level visual representations, alongside a curriculum learning methodology. LHRS-Bot achieves state-of-the-art performance across a variety of RS image understanding tasks (Fig. 1) and demonstrates remarkable abilities to detect intricate objects, engage with human conversation, and provide insights from visual information within RS images (Fig. 5). + +Finally, we manually construct a high-quality benchmark, LHRS-Bench, to facilitate the RS community in evaluating the RS-domain MLLMs across diverse evaluation dimensions. LHRS-Bench includes 690 single-choice questions, spanning 5 top-level evaluation dimensions including 11 fine-grained categories to facilitate a comprehensive, objective, and quantitative RS-specific evaluation. + +The main contributions of our work are summarized as follows: + +1. We introduce LHRS-Align, a large-scale RS image-text dataset, created through a meticulous data generation pipeline using open-source global geographic data. LHRS-Align showcases a broad spectrum of RS semantic visual information, establishing a solid foundation for RS-specific MLLMs. +2. We create LHRS-Instruct, a multimodal instruction-following dataset tailored for RS image understanding. LHRS-Instruct not only contains instruction data to serve for various RS image understanding tasks but also contains complex visual reasoning data generated by GPT-4. +3. We propose LHRS-Bot, an MLLM specifically designed for the RS domain. LHRS-Bot employs a novel bridging strategy and leverages a curriculum learning approach to fully exploit the inherent knowledge of the proposed datasets. LHRS-Bot demonstrates superior performance across various RS image understanding tasks, along with the ability to follow complex instructions. +4. We establish LHRS-Bench, a benchmark aimed at providing a comprehensive and systematic evaluation of RS MLLMs. LHRS-Bench comprises 690 meticulously crafted single-choice questions, encompassing a broad spectrum of 11 fine-grained dimensions. + +# 2 Dataset + +# 2.1 Large-Scale RS Alignment Dataset—LHRS-Align + +LHRS-Align is developed by initially geo-aligning orthorectified RS images with geographic vector features from the free spatial database, OSM. The attributes of these features are then utilized to generate image captions employing an LLM. The detailed pipeline is illustrated in Fig. 2 and introduced as below. + +Data Preparation. The RS images of LHRS-Align are sourced from Google Earth imagery4, encompassing a worldwide coverage with a spatial resolution of 1 meter. It should be noted that our proposed pipeline is applicable to any RS + +![](images/ea02bf74919b00b4eb6e8c1423a9dfe77370e4852127a9c0bb4ddf01593faaf4.jpg) +Fig. 2: Pipeline of constructing LHRs-Align. + +imagery with geographic coordinates. The geographic features are gathered from OSM, a crowdsourced VGI database containing abundant geographical objects. Each object (feature) in the OSM database is annotated as the digitization of the object geometry (e.g., polygons, lines, and points) as well as the semantic tags. A tag is a key-value pair that provides information about a geographic object5, for example, "landuse=residential". Thanks to the free access to the globally encompassing data sources, our pipeline can produce a dataset that offers a vast and comprehensive variety on a worldwide scale. For the current version of LHRS-Align, we collect RS images and OSM features within a scope of $0.3^{\circ}$ longitude $\times 0.28^{\circ}$ latitude for each of the selected cities from all over the world. We customize our own OSM database to streamline the subsequent feature selection process. More information about our data source and database is introduced in supplementary material. + +Geo-alignment between Images and Geographic Features. We iterate through our customized database to identify qualified features serving as "anchors". Based on the geographical extent of each anchor feature, the corresponding RS image is acquired, which is a square bounding box centered on the feature. These images subsequently undergo a stream of postprocessing and pruning procedures. Following this, due to the possibility of multiple features being present within a single image extent, all the features within each image are further identified. At this point, we have established a raw data pool of around 4 million pairs, in which one RS image corresponds to multiple geographical features. Further details for this processes are provided in supplementary material. + +Attribute Pruning and Semantic Balancing. In the OSM database, all the features are assigned with their semantic attributes, that is, multiple key-value tags. However, not all the keys are favorable information for visual identification, such as website and postal code. Hence, we filter out irrelevant keys through a pipeline as detailed in supplementary material. Fig. 2 provides an example showcasing an image along with its geographical features and their corresponding attributes. Furthermore, the OSM database exhibits considerable fea + +ture imbalance, notably dominated by geographical features such as residential or farmland. Given the importance of balanced semantic information for effective vision-language alignment [40,54], we implement a data balancing method, as detailed in supplementary material, to create a more semantically balanced dataset. As a result, this balancing process retains 1.15 million images remaining out of the original 4 million. + +Caption Generation with LLM. The use of image-caption pairs to incorporate visual knowledge into LLMs has been validated as an effective approach [3, 14, 29, 56, 64]. However, using plain image key-value labels can be problematic, as they lack grammatical structure, potentially disrupting the generative and comprehension abilities of LLMs. Similarly, rule-based caption generation strategies may introduce significant redundancy. Therefore, we opt to use open-source LLMs to generate image captions by summarising the key-value tags of the corresponding images. We experiment with various in-context learning examples and LLMs, ranging in size from 7B to 70B. After comprehensive testing by a domain expert, we select Vicuna-v1.5-13B [10] for caption generation as it achieves the best trade-offs in terms of generation efficiency and quality. Detailed information about the in-context examples and the specific deployment of Vicuna-v1.5-13B [10] can be found in supplementary material. Our final dataset, referred as LHRS-Align, comprises approximately 1.15 million image-caption pairs. Each pair is accompanied with additional metadata, such as image resolution and geographical location (country and city). Several examples of LHRS-Align are also presented in supplementary material. + +# 2.2 RS-Specific Instruction-Following Dataset—LHRS-Instruct + +We craft a multimodal instruction-following dataset LHRS-Instruct by utilizing two public image-caption datasets and our LHRS-Align dataset. + +Instruction-Following Dataset from Public RS Caption Dataset. We start by meticulously evaluating and selecting all image_caption datasets within the RS domain. Our selection process focuses on the quality of captions and involves a thorough human evaluation. This requires each image to be accompanied by five distinct captions, with each caption uniquely detailing different aspects of the image. Based on this criterion, we identify RSITMD [58] and NWPU [9] as the datasets with superior caption quality. Subsequent to selection, we design a rigorous data cleaning process, which involves deduplication, computation of caption lengths, and assessment of relevance through CLIP-score calculations. This results in the extraction of 3.4K and 30K high-quality image_caption pairs from RSITMD and NWPU, respectively. We then use several in-context examples along with the five captions per image for prompting Vicuna-v1.5-13B [10] to generate a conversation that specifically focuses on the content of the objects in the image. Finally, we filter out the most relevant conversations through pre-defined rules and evaluations from three domain experts. This process results in 2.9K and 25K instruction data from these two datasets, respectively. For more detailed information on the in-context examples used and the data filtering criteria, please refer to the supplementary material. + +Instruction-Following Dataset from LHRS-Align. Considering the limitations of public image-caption datasets, which typically lack bounding box annotations, and the constrained capabilities of Vicuna-v1.5-13B [10], the generated conversations might omit crucial details such as the locations and quantity of objects in the images, along with visual reasoning information. To mitigate this issue, we meticulously select 15K representative RS images that are rich in attribute information from our LHRS-Align dataset. We then proceed to calculate the bounding boxes for various objects within these images, utilizing the built-in spatial coordinates of different attributes in the OSM database. Leveraging this enriched data, we prompt GPT-4 to craft complex instruction data that incorporates visual reasoning, detailed image descriptions, and conversations about the location and number of objects. The final step involves a manual cleaning of the generated results. This process yields a dataset comprising 0.9K samples about visual reasoning, 4K samples about detailed image description, and 7K samples about conversation. For more details, please refer to supplementary material. + +# 2.3 RS-specific MLLM Evaluation Benchmark—LHRS-Bench + +We construct LHRS-Bench to evaluate the performance of MLLMs in the RS domain. LHRS-Bench employs hierarchical taxonomies to comprehensively evaluate MLLMs across various dimensions (Tab. 1). To prevent data leakage, we abstain from using public RS datasets and instead meticulously collect images from Google Earth. In terms of question format, we utilize single-choice questions + +[16, 26, 27, 34] with 2 to 4 candidate answers, enabling a more quantitative and objective evaluation compared to open-ended questions [4, 6, 33, 57]. Every question and answer within LHRS-Bench are manually crafted to guarantee precision and reliability. Totally, LHRS-Bench comprises 108 images and 690 questions. Some examples of LHRS-Bench are shown in supplementary material. + +The evaluation dimensions of LHRs-Bench encompass 5 top-level dimensions, further delineated into 11 fine-grained sub-dimensions, as shown in Tab. 1. These top-level dimensions include recognition, imagery, spatial-awareness, quantity, and reasoning. For detailed explanations about each dimensions, please refer to supplementary material. + +Table 1: Overview of the dataset distribution. LHRS-Bench contains 108 RS images with 690 question-answer pairs, covering five major evaluation dimensions and 11 sub-dimensions of questions. Each sample is a human-annotated single-choice question that may involve one or more evaluation dimensions simultaneously. + +
DimensionSub-dimensionCount
RecognitionIdentity634
Color113
Orientation39
Shape37
Area75
ImageryResolution21
Modality23
Spatial-awarenessLocation204
Distance22
Quantity-137
Reasoning-46
+ +![](images/49dbacddee4fb62b00e112d482402d203ab36359fad30843dafe281939bbf15a.jpg) +Fig. 3: Architecture of the proposed LHRS-Bot. LHRS-Bot employs learnable queries in conjunction with a vision perceiver to summarize multi-level visual representations. These representations are subsequently concatenated with language token embeddings for processing within the LLM. + +# 3 Method + +We present LHRS-Bot to harness the full potential of LLMs and general visual encoders for enhanced RS images understanding and reasoning. This section delves into each components of LHRS-Bot, detailing how we achieve fully alignment between vision and language models across multiple levels. Subsequently, we explain the curriculum training strategy for training LHRS-Bot. + +# 3.1 Model Architecture + +The overall architecture of LHRS-Bot is presented in Fig. 3. LHRS-Bot is primarily composed of three components: a universal vision encoder, a vision perceiver, and a foundational LLM. + +Vision Encoder. LHRS-Bot utilizes the pre-trained CLIP visual encoder ViT-L/14 [40] with input resolution of $224 \times 224$ pixel for encoding images. The prior works [33,56,64] usually utilize the final hidden feature for abstracting the semantic content of an image. However, several studies have highlighted that features at different levels of a deep learning model embed distinct types of information [17,31,39]. Therefore, we argue that previous image encoding methods fail to sufficiently capture the semantic content of images, leading to a reliance on extensive image-text pairs for vision-language alignment. Consequently, we retain multiple hidden features from various layers of the image encoder. This approach offers a more thorough representation of the underlying image, facilitating a thorough and efficient domain alignment. + +![](images/23b5a95e198ddfb8378944814b1c364b70a84c8720f0ee3803d005bf9817c263.jpg) +Fig. 4: Curriculum learning strategy. We progressively unlock various parameters and introduce data of escalating complexity at different stages, enabling the model to gradually assimilate visual knowledge with increasing difficulty. + +Vision Perceiver. Incorporating multi-level visual features brings extended image tokens, increasing computational overhead and potentially overwhelming language information. To mitigate this issue, we introduce a set of learnable queries for each level of image features. These queries are designed to summarize the semantic information of each level of visual features through a series of stacked cross-attention and MLP layers. Once summarized, these image features are then fed into the LLM. Furthermore, we implement a descending query allocation strategy inspired by the observation that image feature redundancy tends to increase with network depth [7,24]. This strategy allocates more queries to the shallower image feature levels while gradually reducing the number of queries for deeper levels. + +Large Language Model. We employ the LLaMA2-7B [48] as the central expert for interpreting visual and language information to output the desired results following the previous practices [25, 29, 43]. + +# 3.2 Curriculum Learning Strategy + +We implement a three-stage curriculum learning strategy by progressively increasing the difficulty of training tasks (Fig. 4). This strategy incrementally aligns the vision encoder with the LLM, ultimately enabling the LLM to solve multiple visual tasks and follow complex vision-language instructions. + +Stage 1: Pre-training. In the first stage, we use the large-scale, weakly labelled LHRS-Align dataset (Sec. 2.1) to train the vision perceiver to incorporate a wide range of RS visual knowledge into the LLM. Specifically, for each sample in the LHRS-Align dataset, the output visual embedding tokens $X_{v}$ obtained from the vision perceiver are concatenated with the language embedding tokens $X_{l}$ from the corresponding image caption. The concatenated result is then fed into the LLM. The objective of this stage is to maximise the conditional probability $P(X_{l}|X_{v};\theta)$ of generating the caption tokens $X_{l}$ given the prefix visual tokens $X_{v}$ : + +$$ +\begin{array}{l} \mathcal {L} = \log P (X _ {l} | X _ {v}; \theta) \\ = \sum_ {i = 1} ^ {L} \log P \left(x _ {i} \mid X _ {v}, X _ {l, < i}; \theta\right), \tag {1} \\ \end{array} +$$ + +where $L$ is the length of caption tokens, $\theta$ represents the trainable parameters, and $X_{l, < i}$ denotes the generated tokens prior to the current prediction token $x_i$ . + +Stage 2: Multi-task Pre-training. In the second stage, we further train the vision perceiver and efficiently fine-tune the LLM using LoRA [20] on multitask instruction datasets to improve multimodal task solving capabilities of LLM. The construction of the multi-task instruction datasets begins with the collection of diverse public datasets in the RS domain, encompassing tasks such as classification, VQA, and visual grounding. These collected datasets are then meticulously transformed into instruction-based formats using manually designed templates. More detailed information about these multi-task datasets and the templates used can be found in the supplementary material. Moreover, we integrate the instruction data from NWPU, RSITMD and the 4K detailed descriptions data from LHRS-Instruct (Sec. 2.2) to further enrich the LLM with fine-grained RS visual knowledge. The objective for this stage remains similar to Eq. (1) from stage 1, albeit with an adaptation that the generated tokens are now additionally conditioned on the instruction tokens $X_{\text{instruct}}$ : + +$$ +\begin{array}{l} \mathcal {L} = \log P (X _ {l} | X _ {v}, X _ {\text {i n s t r u c t}}; \theta) \\ = \sum_ {i = 1} ^ {L} \log P \left(x _ {i} \mid X _ {v}, X _ {\text {i n s t r u c t}}, X _ {l, < i}; \theta\right). \tag {2} \\ \end{array} +$$ + +Stage 3: Supervised Fine-tuning. In the final stage, our focus shifts towards enhancing the conversational and reasoning capabilities of our model by incorporating more complex instruction data. We utilize all the instruction data from the LHRS-Instruct dataset (Sec. 2.2) and randomly select 20K data from the LLaVA complex reasoning dataset [29]. Additionally, we mix the above datasets with the multi-task datasets from stage 2 using a lower data sampling ratio to further unleash the multi-task problem-solving ability of LHRS-Bot. + +# 4 Experiment + +We conduct comprehensive experiments across image classification, VQA, and visual grounding tasks within the RS domain. These experiments are designed to rigorously validate the superior multi-task solving capabilities of LHRS-Bot. Furthermore, we evaluate the performance of various LLMs in the RS domain using the proposed LHRS-Bench benchmark. We present both quantitative and quality results. + +# 4.1 Experimental Setup + +Dataset. We select diverse RS datasets for providing a comprehensive assessment across different tasks within the RS field: seven RS classification datasets (namely AID [53], WHU-RS19 [13], NWPU [9], SIRI-WHU [65], EuroSAT [18], METER-ML [63], and fMoW [12]), two VQA datasets (LR and HR subsets + +Table 2: Classification accuracy of different models on each dataset. + +
MethodAIDWHU-RS19NWPUSIRI-WHUEuroSATMETER-MLfMoWAvg.
LLaVA-1.531.1054.5534.9617.7126.1221.7311.4328.23
MiniGPTv232.9664.8028.1535.4638.5614.295.2031.35
InstructBLIP29.5036.7634.0118.2020.2514.426.7122.84
mPLUG-OWL248.7972.6646.5854.8333.2936.2717.8544.32
QWen-VL-Chat55.3072.2542.7354.5826.4238.776.8942.42
InternLM-XComposer51.6172.8947.5146.8339.7040.2111.2844.29
LHRS-Bot91.2693.1783.9462.6651.4069.8156.5671.83
+ +of RSVQA [35]), and two visual grounding datasets (RSVG [45] and DIOR-RSVG [59]). We exclude any data from the test sets that overlap with the training set to avoid data leakage. More details about each test dataset are presented in supplementary material. + +Baseline. We evaluate LHRS-Bot and various powerful open-source MLLMs, including LLaVA-v1.5 [32], MiniGPT-v2 [64], InstructBLIP [14], mPLUG-Owl2 [56], QWen-VL-Chat [3], and InternLM-Xcomposer [61], on different tasks. Furthermore, we also include the results of RSGPT [21], GeoChat [25] and SkyEyeGPT [60] from the original papers for comparisons whenever possible. Unless otherwise noted, we utilized 7B variants for all our experiments. + +Implementation. We extract hidden states at layers $\{i = L / 3, j = 2L / 3, k = L - 1\}$ for summarizing through vision perceiver, where $L$ denotes the number of layers in the vision encoder. Our vision perceiver implementation comprises six layers of cross-attention and MLP, with each hidden state assigned $\{64, 48, 32\}$ learnable queries. We apply LoRA module to every linear layer of the LLM, setting the rank $r$ and the scale factor $\alpha$ for LoRA to 128 and 256, respectively. Moreover, we add task identifier [CLS], [VQA], and [VG] for classification, VQA, and visual grounding tasks, respectively, for enabling LHRS-Bot to better distinguish different tasks [64]. For additional details on the hyperparameters used at each stage, please refer to supplementary material. + +# 4.2 Quantitative Evaluation + +We evaluate different MLLMs using greedy decoding with the same prompt for each task and report the average result of two independent trials. Details about the evaluation prompts are provided in supplementary material. + +Classification. The classification results across different datasets are presented in Tab. 2. These results show that LHRS-Bot outperforms all competing models in all seven classification datasets, with a notable margin of $27.51\%$ higher average accuracy than the second ranked method. It is worth mentioning that, except for NWPU, METER-ML, and fMoW, the remaining datasets are completely absent from the multi-task training dataset6, demonstrating the strong + +Table 3: Accuracy of different models on each visual question answering dataset. Following the practice of prior works [21, 25, 60], we omit area and count questions during evaluation. + +
MethodRSVQA-LRRSVQA-HR
Rural/UrbanPresenceCompareAvg.PresenceCompareAvg.
LLaVA-1.559.2273.1665.1965.8648.9659.0253.99
MiniGPTv260.0251.6467.6459.7768.3464.7166.53
InstructBLIP62.6248.8365.9259.1262.6362.9062.77
mPLUG-Owl257.9974.0463.6965.2447.6058.4753.04
QWen-VL-Chat62.0047.6566.5458.7361.7565.9863.87
InternLM-XCompose59.0066.7452.9159.5567.7966.6267.21
RSGPT94.0091.1791.7092.2990.9290.0290.47
GeoChat91.0990.3394.0091.8158.4583.1970.82
SkyEyeGPT88.9388.6375.0084.1680.0080.1382.56
LHRS-Bot89.0788.5190.0089.1992.5792.5392.55
+ +generalization ability of LHRS-Bot. In addition, in the context of the fine-grained RS classification dataset fMoW, our method significantly surpasses other models, leading by an impressive $40\%$ on average. This highlights the critical role of continue pre-training in the RS domain and the effectiveness of our alignment dataset LHRS-Align. + +Visual Question Answering. We report the results of different models on RSVQA-LR and RSVQA-HR datasets in Tab. 3. LHRS-Bot demonstrates performance on par with the previously leading RS domain-specific models on the VQA task. Notably, LHRS-Bot achieves comparable results to RSGPT, despite RSGPT being fine-tuned for five epochs on the RSVQA datasets. Furthermore, LHRS-Bot sets a new record in the RSVQA-HR task with high-resolution RS images, showcasing its superior visual recognition abilities with high-resolution RS images. + +Visual Grounding. The results for the visual grounding task are shown in Tab. 4, where LHRS-Bot outperforms other counterparts. Notably, despite SkyEyeGPT being trained on a more extensive object detection dataset, LHRS-Bot demonstrates comparable performance. This suggests that LHRS-Bot has effectively acquired fine-grained visual information throughout the training process. + +Table 4: Comparison of visual grouting tasks. Accuracy is calculated based on whether the predicted bounding box overlaps with the ground-truth box by more than 0.5 IoU. + +
MethodRSVGDIOR-RSVGAvg.
QWen-VL-Chat44.7680.6562.71
MiniGPTv246.6485.9966.32
SkyEyeGPT70.5088.5979.55
LHRS-Bot73.4588.1080.78
+ +Table 5: Comparison results on the LHRS-Bench dataset with different open-source and closed-source MLLMs. + +
MethodIdentityColorOrientationShapeAreaResolutionModalityLocationDistanceQuantityReasoningAvg.
LLaVA-1.542.9030.9715.3843.2444.0023.8021.7435.7827.2721.8954.3537.45
MiniGPTv220.1915.932.568.1113.3314.294.3515.2013.649.4936.9616.88
InstructBLIP20.3522.125.1324.3222.6738.1000.0013.7322.8513.1430.4318.87
mPLUG-OWL241.4845.1312.8240.5432.0033.3300.0030.8813.6432.1254.3437.00
QWen-VL-Chat31.3832.747.6932.4329.3319.054.3524.0221.4521.1756.5228.65
InternLM-XComposer34.3830.977.6937.8426.6700.008.7033.3322.7318.2565.2231.09
GPT-4-Turbo50.3243.3610.2667.5744.0042.8639.1342.1622.7321.8971.7444.56
Claude-3-Opus41.0136.2817.9564.8641.3366.6739.1334.8045.4526.2854.3539.08
LHRS-Bot44.6442.4823.0867.5725.3342.8632.3556.5222.7318.2573.9139.38
+ +# 4.3 Evaluation on LHRS-Bench + +We evaluate LHRS-Bot with different open-source and closed-source MLLMs on our LHRS-Bench benchmark. For each question within the benchmark, choices are shuffled, and the generation process is repeated four times. An answer is considered incorrect if any attempt yields an incorrect response. Same prompts are used for evaluation, as detailed in supplementary material. We consider the model answers correctly if it outputs the correct choice or the answer corresponding to the correct choice. We calculate accuracy as our evaluation metric. + +The results presented in Tab. 5 indicate that LHRS-Bot outperforms all other open-source models. It is worth noting that, LHRS-Bot greatly outperforms other open-source models in terms of the evaluation dimensions of spatial resolution and image modality, even though these aspects were not explicitly trained for in the models, highlighting the effectiveness of RS-specific MLLM training. Additionally, LHRS-Bot outperforms the strong closed-source model Claude and even matches the performance of GPT-4V. However, there is still significant room for improvement in MLLM capabilities within the RS domain. Examples of evaluation results are available on supplementary material. + +# 4.4 Qualitative Evaluation + +We conduct qualitative demonstrations to illustrate the extensive capabilities of LHRS-Bot in interpreting free-form RS images, as showcased in Fig. 5. The results demonstrate that LHRS-Bot not only comprehends the overall content of an image but also exhibits a strong ability for discerning fine-grained details. Moreover, LHRS-Bot possesses a perceptive understanding of the orientation, position, and quantity of objects within an image. Furthermore, LHRS-Bot is capable of making reasoned inferences from the images and acknowledges its limitations by indicating uncertainty when it cannot provide a definitive answer. + +# 4.5 Ablation Analysis + +In this subsection, we carry out extensive ablation experiments to demonstrate the effectiveness of our dataset as well as the model design. + +![](images/a6287d5176ccce0e788920bc7dfc760d7e45e5a1beea22f7606e382ab9ab466e.jpg) +Fig. 5: Qualitative examples between real users and LHRS-Bot. + +Table 6: Performance with different alignment dataset. + +
AIDWHU-RS19ClassificationVQAVisual Grounding
NWPUSIRI-WHUEuroSATMETER-MLfMoWRSVQA-LRRSVQA-HRRSVGDIOR-RSVG
RS5M86.3196.7582.8060.2054.9264.7742.6878.3691.6570.5887.12
SkyScript88.2190.2678.0151.4343.6650.0231.2788.7991.2471.2585.05
LHRS-Align91.2693.1783.9462.6651.4069.8156.5689.1992.5573.4588.10
+ +Effectiveness of LHRS-Align dataset. There are already many large-scale image-caption datasets in the field of RS [50,62]. To evaluate how different alignment datasets impact model performance, we substitute our LHRS-Align dataset with various image-caption datasets and repeat the entire training process. This process is then evaluated across three distinct tasks: classification, VQA, and visual grounding. The experimental setup remains consistent with Sec. 4.1, with the exception of the dataset used in the first stage. + +The results are presented in Tab. 6. Notably, despite possessing the smallest dataset (RS5M: 5M, SkyScript: 1.5M, LHRS-Align : 1.15M), LHRS-Align outperforms other settings. It is worth mentioning that both SkyScript and LHRS-Align, which leverage RS images combined with the OSM database, achieve results comparable to or better than the even larger RS5M, which is derived from publicly available natural image datasets, such as LAION [44]. This highlights the importance of domain-specific data collection. Additionally, superior performance of LHRS-Align over SkyScript suggests that enriched captions and a meticulous data cleaning pipeline can substantially enhance the effectiveness of MLLMs. + +Effectiveness of Model Design. We evaluate the effectiveness of designs involving multi-level visual hidden representations as well as descending query allocation strategies. Specifically, our evaluation focuses on comparing with the single level visual hidden state, projecting it through a two-layer MLP or a vision perceiver. We also perform comparisons between multi-level visual hidden + +Table 7: Performance comparison across various architectural designs. + +
MethodMulti Level Number of Query ClassificationVQA Visual Grounding
MLP✘-64.4690.06
Vision Perceiver✘6463.6983.30
Vision Perceiver✓[64,64,64]63.7487.88
Vision Perceiver✓[32,48,64]60.6887.11
Token Merging✓6449.5181.49
Vision Perceiver✓[64,48,32]65.4390.14
+ +representations, employing either identical queries for each level or incremental query allocation strategies. Building on the concept introduced by the token merging [7], we explore the potential of integrating this strategy with a vision perceiver to efficiently leverage multi-level representations. To minimize experimental costs, our training process comprises just two stages: firstly, we employ LHRS-Align to train the bridge layer, which aligns the vision and language domains. Subsequently, we train the model using the LHRS-Instruct and multi-task instruction datasets. + +We report the average result on each task in Tab. 7. It is evident that the approach employed by LHRS-Bot achieves superior performance. The results further illustrate that, although the extraction of a singular visual representation performs better in tasks such as classification and VQA tasks, which depends on high-level visual semantics, it falls short in visual grounding tasks that require detailed visual information. Furthermore, employing the same number of queries to summarise different levels of visual information does not lead to better results. In contrast, adopting a diminishing allocation strategy yields superior results. This further demonstrates the large redundancy among the deep visual tokens [7]. Lastly, the token merging strategy proves ineffective, possibly due to the loss of crucial information during the merging process. + +# 5 Conclusion + +We propose LHRS-Bot, an MLLM for the RS domain. To unleash the potential of LLMs for RS image understanding, we curate a large-scale dataset, LHRS-Align, for RS-specific alignment, and LHRS-Instruct, a multimodal instruction-following dataset to enhance LHRS-Bot's instruction-following capabilities. Additionally, we introduce LHRS-Bench, a benchmark providing a systematic evaluation framework for RS MLLMs. + +In comprehensive comparison experiments, LHRS-Bot demonstrates exceptional RS domain knowledge, surpassing both general-purpose and RS-specific MLLMs. Despite its impressive performance, LHRS-Bot shares limitations common to MLLMs, such as susceptibility to hallucinations. We believe that developing higher-quality alignment and instructional datasets, along with improved training strategies, can further enhance the performance of MLLMs for interpreting RS images. + +# Acknowledgement + +This research was supported by the National Natural Science Foundation of China under Grant 42071297, and in part by the AI & AI for Science Project of Nanjing University under Grant 02091480605203. We are grateful to High Performance Computing Center of Nanjing University for their help on GPU resources. We also would like to thank the anonymous reviewers for their constructive comments. + +# References + +1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022) +2. Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., et al.: Qwen technical report. arXiv preprint arXiv:2309.16609 (2023) +3. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023) +4. Bai, S., Yang, S., Bai, J., Wang, P., Zhang, X., Lin, J., Wang, X., Zhou, C., Zhou, J.: Touchstone: Evaluating vision-language models by language models. arXiv preprint arXiv:2308.16890 (2023) +5. Bashmal, L., Bazi, Y., Melgani, F., Al Rahhal, M.M., Al Zuair, M.A.: Language integration in remote sensing: Tasks, datasets, and future directions. IEEE Geoscience and Remote Sensing Magazine (2023) +6. Bitton, Y., Bansal, H., Hessel, J., Shao, R., Zhu, W., Awadalla, A., Gardner, J., Taori, R., Schimdt, L.: Visit-bench: A benchmark for vision-language instruction following inspired by real-world use. arXiv preprint arXiv:2308.06595 (2023) +7. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461 (2022) +8. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) +9. Cheng, G., Han, J., Lu, X.: Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE 105(10), 1865-1883 (2017) +10. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023) +11. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. Journal of Machine Learning Research 24(240), 1-113 (2023) +12. Christie, G., Fendley, N., Wilson, J., Mukherjee, R.: Functional map of the world. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6172-6180 (2018) + +13. Dai, D., Yang, W.: Satellite image classification via two-layer sparse coding with biased image representation. IEEE Geoscience and Remote Sensing Letters 8(1), 173-176 (2011) +14. Dai, W., Li, J., Li, D., Tiong, A., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arxiv 2023. arXiv preprint arXiv:2305.06500 +15. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023) +16. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023) +17. Ghiasi, A., Kazemi, H., Borgnia, E., Reich, S., Shu, M., Goldblum, M., Wilson, A.G., Goldstein, T.: What do vision transformers learn? a visual exploration. arXiv preprint arXiv:2212.06727 (2022) +18. Helber, P., Bischke, B., Dengel, A., Borth, D.: Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 12(7), 2217-2226 (2019) +19. Hossain, M.D., Chen, D.: Segmentation for object-based image analysis (obia): A review of algorithms and challenges from remote sensing perspective. ISPRS Journal of Photogrammetry and Remote Sensing 150, 115-134 (2019) +20. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) +21. Hu, Y., Yuan, J., Wen, C., Lu, X., Li, X.: Rsgpt: A remote sensing vision language model and benchmark. arXiv preprint arXiv:2307.15266 (2023) +22. Huang, S., Dong, L., Wang, W., Hao, Y., Singhal, S., Ma, S., Lv, T., Cui, L., Mohammed, O.K., Liu, Q., et al.: Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045 (2023) +23. Jiang, A.Q., Sablayrolles, A., Roux, A., Mensch, A., Savary, B., Bamford, C., Chaplot, D.S., Casas, D.d.l., Hanna, E.B., Bressand, F., et al.: Mixtral of experts. arXiv preprint arXiv:2401.04088 (2024) +24. Ju, C., Wang, H., Li, Z., Chen, X., Zhai, Z., Huang, W., Xiao, S.: Turbo: Informativity-driven acceleration plug-in for vision-language models. arXiv preprint arXiv:2312.07408 (2023) +25. Kuckreja, K., Danish, M.S., Naseer, M., Das, A., Khan, S., Khan, F.S.: Geochat: Grounded large vision-language model for remote sensing. arXiv preprint arXiv:2311.15826 (2023) +26. Li, B., Ge, Y., Ge, Y., Wang, G., Wang, R., Zhang, R., Shan, Y.: Seed-bench-2: Benchmarking multimodal large language models. arXiv preprint arXiv:2311.17092 (2023) +27. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125 (2023) +28. Li, C., Gan, Z., Yang, Z., Yang, J., Li, L., Wang, L., Gao, J.: Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020 (2023) +29. Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H., Gao, J.: Llava-med: Training a large language-and-vision assistant for biomedicine in one day. arXiv preprint arXiv:2306.00890 (2023) + +30. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) +31. Li, M., Wang, S., Zhang, Q.: Visualizing the emergence of intermediate visual patterns in dnns. Advances in Neural Information Processing Systems 34, 6594-6607 (2021) +32. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023) +33. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023) +34. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023) +35. Lobry, S., Marcos, D., Murray, J., Tuia, D.: Rsvqa: Visual question answering for remote sensing data. IEEE Transactions on Geoscience and Remote Sensing 58(12), 8555-8566 (2020) +36. Ma, Y., Cao, Y., Sun, J., Pavone, M., Xiao, C.: Multimodal language model for driving. arXiv preprint arXiv:2312.00438 (2023) +37. Muhtar, D., Zhang, X., Xiao, P., Li, Z., Gu, F.: Cmid: A unified self-supervised learning framework for remote sensing image understanding. IEEE Transactions on Geoscience and Remote Sensing (2023) +38. OpenAI: Gpt-4 technical report (2023) +39. Park, N., Kim, W., Heo, B., Kim, T., Yun, S.: What do self-supervised vision transformers learn? arXiv preprint arXiv:2305.00729 (2023) +40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +41. Ratledge, N., Cadamuro, G., de la Cuesta, B., Stigler, M., Burke, M.: Using machine learning to assess the livelihood impact of electricity access. Nature 611(7936), 491-495 (2022) +42. Reed, C.J., Gupta, R., Li, S., Brockman, S., Funk, C., Clipp, B., Keutzer, K., Candido, S., Uytendaele, M., Darrell, T.: Scale-MAE: A scale-aware masked autoencoder for multiscale geospatial representation learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4088-4099 (2023) +43. Saygin Seyfioglu, M., Ikezogwo, W.O., Ghezloo, F., Krishna, R., Shapiro, L.: Quilt-lava: Visual instruction tuning by extracting localized narratives from open-source histopathology videos. arXiv e-prints pp. arXiv-2312 (2023) +44. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) +45. Sun, Y., Feng, S., Li, X., Ye, Y., Kang, J., Huang, X.: Visual grounding in remote sensing images. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 404-412 (2022) +46. Team, M.N.: Introducing mpt-7b: A new standard for open-source, commercially usable llms (2023), www.mosaicml.com/blog/mpt-7b, accessed: 2023-05-05 +47. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) + +48. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) +49. Wang, D., Zhang, J., Du, B., Xu, M., Liu, L., Tao, D., Zhang, L.: Samrs: Scaling-up remote sensing segmentation dataset with segment anything model. In: Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2023) +50. Wang, Z., Prabha, R., Huang, T., Wu, J., Rajagopal, R.: Skyscript: A large and semantically diverse vision-language dataset for remote sensing. arXiv preprint arXiv:2312.12856 (2023) +51. Wen, C., Hu, Y., Li, X., Yuan, Z., Zhu, X.X.: Vision-language models in remote sensing: Current progress and future trends. arXiv preprint arXiv:2305.05726 (2023) +52. Xia, G.S., Bai, X., Ding, J., Zhu, Z., Belongie, S., Luo, J., Datcu, M., Pelillo, M., Zhang, L.: DOTA: A large-scale dataset for object detection in aerial images. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3974-3983 (2018) +53. Xia, G.S., Hu, J., Hu, F., Shi, B., Bai, X., Zhong, Y., Zhang, L., Lu, X.: AID: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing 55(7), 3965-3981 (2017) +54. Xu, H., Xie, S., Tan, X.E., Huang, P.Y., Howes, R., Sharma, V., Li, S.W., Ghosh, G., Zettlemoyer, L., Feichtenhofer, C.: Demystifying clip data. arXiv preprint arXiv:2309.16671 (2023) +55. Yang, J., Gong, P., Fu, R., Zhang, M., Chen, J., Liang, S., Xu, B., Shi, J., Dickinson, R.: The role of satellite remote sensing in climate change studies. Nature climate change 3(10), 875-883 (2013) +56. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023) +57. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mmvet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023) +58. Yuan, Z., Zhang, W., Fu, K., Li, X., Deng, C., Wang, H., Sun, X.: Exploring a fine-grained multiscale method for cross-modal remote sensing image retrieval. arXiv preprint arXiv:2204.09868 (2022) +59. Zhan, Y., Xiong, Z., Yuan, Y.: Rsvg: Exploring data and models for visual grounding on remote sensing data. IEEE Transactions on Geoscience and Remote Sensing 61, 1-13 (2023) +60. Zhan, Y., Xiong, Z., Yuan, Y.: Skyeyegpt: Unifying remote sensing vision-language tasks via instruction tuning with large language model. arXiv preprint arXiv:2401.09712 (2024) +61. Zhang, P., Wang, X.D.B., Cao, Y., Xu, C., Ouyang, L., Zhao, Z., Ding, S., Zhang, S., Duan, H., Yan, H., et al.: Internl m-xcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112 (2023) +62. Zhang, Z., Zhao, T., Guo, Y., Yin, J.: Rs5m: A large scale vision-language dataset for remote sensing vision-language foundation model. arXiv preprint arXiv:2306.11300 (2023) +63. Zhu, B., Lui, N., Irvin, J., Le, J., Tadwalkar, S., Wang, C., Ouyang, Z., Liu, F.Y., Ng, A.Y., Jackson, R.B.: Meter-ml: A multi-sensor earth observation benchmark for automated methane source mapping. arXiv preprint arXiv:2207.11166 (2022) + +64. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) +65. Zhu, Q., Zhong, Y., Zhao, B., Xia, G.S., Zhang, L.: Bag-of-visual-words scene classifier with local and global features for high spatial resolution remote sensing imagery. IEEE Geoscience and Remote Sensing Letters 13(6), 747-751 (2016) \ No newline at end of file diff --git a/2024/LHRS-Bot_ Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model/images.zip b/2024/LHRS-Bot_ Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fb9518b976cbb9cd09eb80f69ab7710e7f78c852 --- /dev/null +++ b/2024/LHRS-Bot_ Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81036f5a2549c14568e5cc4cb56e89eced07e125bff62604d27b2960078e6dc9 +size 599477 diff --git a/2024/LHRS-Bot_ Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model/layout.json b/2024/LHRS-Bot_ Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..079c9a82d7831d1366c969946661250d0dc56950 --- /dev/null +++ b/2024/LHRS-Bot_ Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model/layout.json @@ -0,0 +1,8057 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 111, + 467, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 111, + 467, + 163 + ], + "spans": [ + { + "bbox": [ + 146, + 111, + 467, + 163 + ], + "type": "text", + "content": "LHRS-Bot: Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 137, + 186, + 474, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 186, + 474, + 200 + ], + "spans": [ + { + "bbox": [ + 137, + 186, + 474, + 200 + ], + "type": "text", + "content": "Dilxat Muhtar†, Zhenshi Li†, Feng Gu, Xueliang Zhang*, and Pengfeng Xiao" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 171, + 209, + 441, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 209, + 441, + 242 + ], + "spans": [ + { + "bbox": [ + 171, + 209, + 441, + 242 + ], + "type": "text", + "content": "Nanjing University \nLzhenshi@outlook.com, {dmuhtar, gufeng}@smail.nju.edu.com \n{zxl, xiaopf}@nju.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 277, + 461, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 277, + 461, + 452 + ], + "spans": [ + { + "bbox": [ + 160, + 277, + 461, + 452 + ], + "type": "text", + "content": "Abstract. The revolutionary capabilities of large language models (LLMs) have paved the way for multimodal large language models (MLLMs) and fostered diverse applications across various specialized domains. In the remote sensing (RS) field, however, the diverse geographical landscapes and varied objects in RS imagery are not adequately considered in recent MLLM endeavors. To bridge this gap, we construct a large-scale RS image-text dataset, LHRs" + }, + { + "bbox": [ + 160, + 277, + 461, + 452 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 160, + 277, + 461, + 452 + ], + "type": "text", + "content": "-Align, and an informative RS-specific instruction dataset, LHRs-Instruct, leveraging the extensive volunteered geographic information (VGI) and globally available RS images. Building on this foundation, we introduce LHRs-Bot, an MLLM tailored for RS image understanding through a novel multi-level vision-language alignment strategy and a curriculum learning method. Additionally, we introduce LHRs-Bench, a benchmark for thoroughly evaluating MLLMs' abilities in RS image understanding. Comprehensive experiments demonstrate that LHRs-Bot exhibits a profound understanding of RS images and the ability to perform nuanced reasoning within the RS domain" + }, + { + "bbox": [ + 160, + 277, + 461, + 452 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 160, + 277, + 461, + 452 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 464, + 432, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 464, + 432, + 475 + ], + "spans": [ + { + "bbox": [ + 160, + 464, + 432, + 475 + ], + "type": "text", + "content": "Keywords: Multimodal " + }, + { + "bbox": [ + 160, + 464, + 432, + 475 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 464, + 432, + 475 + ], + "type": "text", + "content": " Large language model " + }, + { + "bbox": [ + 160, + 464, + 432, + 475 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 464, + 432, + 475 + ], + "type": "text", + "content": " Remote sensing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 498, + 229, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 498, + 229, + 510 + ], + "spans": [ + { + "bbox": [ + 133, + 498, + 229, + 510 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 527, + 481, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 527, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 133, + 527, + 481, + 611 + ], + "type": "text", + "content": "The recently advanced large language models (LLMs) have presented remarkable capabilities in engaging in informative conversations and solving complex problems [2,8,11,23,46-48]. To extend the perceptual capabilities of LLMs, multimodal large language models (MLLMs) bring eyes to LLMs. These models, by aligning with visual representations and visual instruction tuning, have demonstrated impressive multimodal instruction following capabilities and can act as a general-purpose interface for a broad range of tasks [1, 3, 22, 28, 30, 33, 64]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 620, + 326, + 632 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 620, + 326, + 632 + ], + "spans": [ + { + "bbox": [ + 146, + 620, + 326, + 632 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 146, + 620, + 326, + 632 + ], + "type": "text", + "content": " Equal contribution, listed in random order." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 147, + 632, + 242, + 643 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 632, + 242, + 643 + ], + "spans": [ + { + "bbox": [ + 147, + 632, + 242, + 643 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 147, + 643, + 361, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 643, + 361, + 654 + ], + "spans": [ + { + "bbox": [ + 147, + 643, + 361, + 654 + ], + "type": "text", + "content": "LHRS stands for 'Language Helps Remote Sensing'." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 147, + 654, + 480, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 654, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 147, + 654, + 480, + 665 + ], + "type": "text", + "content": "2Data, Code and model are available at https://github.com/NJU-LHRS/LHRS-Bot" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 282, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 282, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 282, + 200 + ], + "type": "text", + "content": "This success in general domain has served as a catalyst for the development of domain specialized MLLMs, with notable applications in biomedicine [29], autonomous driving [36], and robotics [15]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 201, + 282, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 282, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 282, + 344 + ], + "type": "text", + "content": "Among various specialized domains, remote sensing (RS) image understanding is particularly important, as it directly involves the monitoring and analysis of the Earth's surface and environment [5, 41, 51, 55]. However, RS images present significant challenges for holistic image understanding due to the intricate complexity of different objects across diverse landscapes world" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 293, + 114, + 484, + 292 + ], + "blocks": [ + { + "bbox": [ + 293, + 114, + 484, + 292 + ], + "lines": [ + { + "bbox": [ + 293, + 114, + 484, + 292 + ], + "spans": [ + { + "bbox": [ + 293, + 114, + 484, + 292 + ], + "type": "image", + "image_path": "106257a820d7b236c17b5b0d3e5f88fe33f6417123e81c6f4d4ed96e396fe0f6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 303, + 482, + 337 + ], + "lines": [ + { + "bbox": [ + 287, + 303, + 482, + 337 + ], + "spans": [ + { + "bbox": [ + 287, + 303, + 482, + 337 + ], + "type": "text", + "content": "Fig.1: Overall performance comparison between LHRS-Bot and existing MLLMs on different RS image understanding tasks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 344, + 482, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 488 + ], + "type": "text", + "content": "wide, coupled with the variability in visual scales [19, 37, 42]. MLLMs bring together vision and language, offering possibilities for a unified and interactive understanding of RS images [51]. In this direction, several RS-specific MLLMs have been proposed, such as RSGPT [21], GeoChat [25], and SkyEyeGPT [60]. However, they are struggling to fit existing public RS datasets, due to the oversight of exploiting the extensive and worldwide RS features to infuse comprehensive RS visual knowledge into LLMs. Furthermore, these contemporary approaches of bridging vision and language domains tend to focus on the incorporation of high-level visual semantics. This strategy neglects the fact that different levels of visual information are crucial to fully align the linguistic and visual domains, which could facilitate a unified RS image understanding at different granularity [37, 49, 52]. Detailed related works are available in supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 489, + 482, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 482, + 644 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 482, + 644 + ], + "type": "text", + "content": "To bridge the gap, we present LHRS-Bot, a specialized MLLM for RS, enhanced by globally available volunteered geographical information (VGI) and worldwide RS images. Specifically, we construct a large-scale, semantically rich, and feature-diverse dataset, LHRS-Align, by geographically pairing RS images with plentiful attributed information from the OpenStreetMap VGI database3, and consequently generating image captions using LLMs. Our rigorous data cleaning process, which includes dedduplication, pruning, and semantic balancing, ensures that LHRS-Align comprises 1.15 million meaningful and high-quality RS image-text pairs, laying a solid foundation for embedding rich RS visual knowledge into MLLMs. Furthermore, we construct our instruction dataset, LHRS-Instruct, by reorganizing a broad range of open-source RS datasets into multitask instructional datasets and employing GPT-4 [38] to create complex instructional data from the high-quality subset of LHRS-Align. Building on these" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "text", + "content": "D. Muhtar et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 653, + 294, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 653, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 145, + 653, + 294, + 666 + ], + "type": "text", + "content": "3https://www.openstreetmap.org/" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "datasets, we propose a novel RS-specific MLLM, LHRS-Bot, by implementing a novel bridging strategy for efficiently summarizing multi-level visual representations, alongside a curriculum learning methodology. LHRS-Bot achieves state-of-the-art performance across a variety of RS image understanding tasks (Fig. 1) and demonstrates remarkable abilities to detect intricate objects, engage with human conversation, and provide insights from visual information within RS images (Fig. 5)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 480, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 259 + ], + "type": "text", + "content": "Finally, we manually construct a high-quality benchmark, LHRS-Bench, to facilitate the RS community in evaluating the RS-domain MLLMs across diverse evaluation dimensions. LHRS-Bench includes 690 single-choice questions, spanning 5 top-level evaluation dimensions including 11 fine-grained categories to facilitate a comprehensive, objective, and quantitative RS-specific evaluation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 261, + 425, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 261, + 425, + 271 + ], + "spans": [ + { + "bbox": [ + 147, + 261, + 425, + 271 + ], + "type": "text", + "content": "The main contributions of our work are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 280, + 479, + 494 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 136, + 280, + 479, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 280, + 479, + 327 + ], + "spans": [ + { + "bbox": [ + 136, + 280, + 479, + 327 + ], + "type": "text", + "content": "1. We introduce LHRS-Align, a large-scale RS image-text dataset, created through a meticulous data generation pipeline using open-source global geographic data. LHRS-Align showcases a broad spectrum of RS semantic visual information, establishing a solid foundation for RS-specific MLLMs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 329, + 479, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 329, + 479, + 374 + ], + "spans": [ + { + "bbox": [ + 136, + 329, + 479, + 374 + ], + "type": "text", + "content": "2. We create LHRS-Instruct, a multimodal instruction-following dataset tailored for RS image understanding. LHRS-Instruct not only contains instruction data to serve for various RS image understanding tasks but also contains complex visual reasoning data generated by GPT-4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 376, + 479, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 376, + 479, + 445 + ], + "spans": [ + { + "bbox": [ + 136, + 376, + 479, + 445 + ], + "type": "text", + "content": "3. We propose LHRS-Bot, an MLLM specifically designed for the RS domain. LHRS-Bot employs a novel bridging strategy and leverages a curriculum learning approach to fully exploit the inherent knowledge of the proposed datasets. LHRS-Bot demonstrates superior performance across various RS image understanding tasks, along with the ability to follow complex instructions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 448, + 479, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 448, + 479, + 494 + ], + "spans": [ + { + "bbox": [ + 136, + 448, + 479, + 494 + ], + "type": "text", + "content": "4. We establish LHRS-Bench, a benchmark aimed at providing a comprehensive and systematic evaluation of RS MLLMs. LHRS-Bench comprises 690 meticulously crafted single-choice questions, encompassing a broad spectrum of 11 fine-grained dimensions." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 515, + 201, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 201, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 201, + 528 + ], + "type": "text", + "content": "2 Dataset" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 541, + 421, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 541, + 421, + 553 + ], + "spans": [ + { + "bbox": [ + 131, + 541, + 421, + 553 + ], + "type": "text", + "content": "2.1 Large-Scale RS Alignment Dataset—LHRS-Align" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 562, + 479, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 562, + 479, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 562, + 479, + 609 + ], + "type": "text", + "content": "LHRS-Align is developed by initially geo-aligning orthorectified RS images with geographic vector features from the free spatial database, OSM. The attributes of these features are then utilized to generate image captions employing an LLM. The detailed pipeline is illustrated in Fig. 2 and introduced as below." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 610, + 480, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 610, + 480, + 647 + ], + "spans": [ + { + "bbox": [ + 131, + 610, + 480, + 647 + ], + "type": "text", + "content": "Data Preparation. The RS images of LHRS-Align are sourced from Google Earth imagery4, encompassing a worldwide coverage with a spatial resolution of 1 meter. It should be noted that our proposed pipeline is applicable to any RS" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "LHRS-Bot" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 653, + 289, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 653, + 289, + 665 + ], + "spans": [ + { + "bbox": [ + 145, + 653, + 289, + 665 + ], + "type": "text", + "content": "4https://www.google.com/earth/" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 482, + 248 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 248 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 248 + ], + "type": "image", + "image_path": "ea02bf74919b00b4eb6e8c1423a9dfe77370e4852127a9c0bb4ddf01593faaf4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 257, + 399, + 269 + ], + "lines": [ + { + "bbox": [ + 214, + 257, + 399, + 269 + ], + "spans": [ + { + "bbox": [ + 214, + 257, + 399, + 269 + ], + "type": "text", + "content": "Fig. 2: Pipeline of constructing LHRs-Align." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 281, + 482, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 435 + ], + "type": "text", + "content": "imagery with geographic coordinates. The geographic features are gathered from OSM, a crowdsourced VGI database containing abundant geographical objects. Each object (feature) in the OSM database is annotated as the digitization of the object geometry (e.g., polygons, lines, and points) as well as the semantic tags. A tag is a key-value pair that provides information about a geographic object5, for example, \"landuse=residential\". Thanks to the free access to the globally encompassing data sources, our pipeline can produce a dataset that offers a vast and comprehensive variety on a worldwide scale. For the current version of LHRS-Align, we collect RS images and OSM features within a scope of " + }, + { + "bbox": [ + 130, + 281, + 482, + 435 + ], + "type": "inline_equation", + "content": "0.3^{\\circ}" + }, + { + "bbox": [ + 130, + 281, + 482, + 435 + ], + "type": "text", + "content": " longitude " + }, + { + "bbox": [ + 130, + 281, + 482, + 435 + ], + "type": "inline_equation", + "content": "\\times 0.28^{\\circ}" + }, + { + "bbox": [ + 130, + 281, + 482, + 435 + ], + "type": "text", + "content": " latitude for each of the selected cities from all over the world. We customize our own OSM database to streamline the subsequent feature selection process. More information about our data source and database is introduced in supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": "Geo-alignment between Images and Geographic Features. We iterate through our customized database to identify qualified features serving as \"anchors\". Based on the geographical extent of each anchor feature, the corresponding RS image is acquired, which is a square bounding box centered on the feature. These images subsequently undergo a stream of postprocessing and pruning procedures. Following this, due to the possibility of multiple features being present within a single image extent, all the features within each image are further identified. At this point, we have established a raw data pool of around 4 million pairs, in which one RS image corresponds to multiple geographical features. Further details for this processes are provided in supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": "Attribute Pruning and Semantic Balancing. In the OSM database, all the features are assigned with their semantic attributes, that is, multiple key-value tags. However, not all the keys are favorable information for visual identification, such as website and postal code. Hence, we filter out irrelevant keys through a pipeline as detailed in supplementary material. Fig. 2 provides an example showcasing an image along with its geographical features and their corresponding attributes. Furthermore, the OSM database exhibits considerable fea" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "text", + "content": "D. Muhtar et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 654, + 379, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 654, + 379, + 665 + ], + "spans": [ + { + "bbox": [ + 145, + 654, + 379, + 665 + ], + "type": "text", + "content": "5https://wiki.openstreetmap.org/wiki/Map_features" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": "ture imbalance, notably dominated by geographical features such as residential or farmland. Given the importance of balanced semantic information for effective vision-language alignment [40,54], we implement a data balancing method, as detailed in supplementary material, to create a more semantically balanced dataset. As a result, this balancing process retains 1.15 million images remaining out of the original 4 million." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 392 + ], + "type": "text", + "content": "Caption Generation with LLM. The use of image-caption pairs to incorporate visual knowledge into LLMs has been validated as an effective approach [3, 14, 29, 56, 64]. However, using plain image key-value labels can be problematic, as they lack grammatical structure, potentially disrupting the generative and comprehension abilities of LLMs. Similarly, rule-based caption generation strategies may introduce significant redundancy. Therefore, we opt to use open-source LLMs to generate image captions by summarising the key-value tags of the corresponding images. We experiment with various in-context learning examples and LLMs, ranging in size from 7B to 70B. After comprehensive testing by a domain expert, we select Vicuna-v1.5-13B [10] for caption generation as it achieves the best trade-offs in terms of generation efficiency and quality. Detailed information about the in-context examples and the specific deployment of Vicuna-v1.5-13B [10] can be found in supplementary material. Our final dataset, referred as LHRS-Align, comprises approximately 1.15 million image-caption pairs. Each pair is accompanied with additional metadata, such as image resolution and geographical location (country and city). Several examples of LHRS-Align are also presented in supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 407, + 468, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 407, + 468, + 419 + ], + "spans": [ + { + "bbox": [ + 131, + 407, + 468, + 419 + ], + "type": "text", + "content": "2.2 RS-Specific Instruction-Following Dataset—LHRS-Instruct" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 426, + 479, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 479, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 479, + 449 + ], + "type": "text", + "content": "We craft a multimodal instruction-following dataset LHRS-Instruct by utilizing two public image-caption datasets and our LHRS-Align dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": "Instruction-Following Dataset from Public RS Caption Dataset. We start by meticulously evaluating and selecting all image_caption datasets within the RS domain. Our selection process focuses on the quality of captions and involves a thorough human evaluation. This requires each image to be accompanied by five distinct captions, with each caption uniquely detailing different aspects of the image. Based on this criterion, we identify RSITMD [58] and NWPU [9] as the datasets with superior caption quality. Subsequent to selection, we design a rigorous data cleaning process, which involves deduplication, computation of caption lengths, and assessment of relevance through CLIP-score calculations. This results in the extraction of 3.4K and 30K high-quality image_caption pairs from RSITMD and NWPU, respectively. We then use several in-context examples along with the five captions per image for prompting Vicuna-v1.5-13B [10] to generate a conversation that specifically focuses on the content of the objects in the image. Finally, we filter out the most relevant conversations through pre-defined rules and evaluations from three domain experts. This process results in 2.9K and 25K instruction data from these two datasets, respectively. For more detailed information on the in-context examples used and the data filtering criteria, please refer to the supplementary material." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "LHRS-Bot" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "content": "Instruction-Following Dataset from LHRS-Align. Considering the limitations of public image-caption datasets, which typically lack bounding box annotations, and the constrained capabilities of Vicuna-v1.5-13B [10], the generated conversations might omit crucial details such as the locations and quantity of objects in the images, along with visual reasoning information. To mitigate this issue, we meticulously select 15K representative RS images that are rich in attribute information from our LHRS-Align dataset. We then proceed to calculate the bounding boxes for various objects within these images, utilizing the built-in spatial coordinates of different attributes in the OSM database. Leveraging this enriched data, we prompt GPT-4 to craft complex instruction data that incorporates visual reasoning, detailed image descriptions, and conversations about the location and number of objects. The final step involves a manual cleaning of the generated results. This process yields a dataset comprising 0.9K samples about visual reasoning, 4K samples about detailed image description, and 7K samples about conversation. For more details, please refer to supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 323, + 462, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 323, + 462, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 323, + 462, + 335 + ], + "type": "text", + "content": "2.3 RS-specific MLLM Evaluation Benchmark—LHRS-Bench" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 352, + 482, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 412 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 412 + ], + "type": "text", + "content": "We construct LHRS-Bench to evaluate the performance of MLLMs in the RS domain. LHRS-Bench employs hierarchical taxonomies to comprehensively evaluate MLLMs across various dimensions (Tab. 1). To prevent data leakage, we abstain from using public RS datasets and instead meticulously collect images from Google Earth. In terms of question format, we utilize single-choice questions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 413, + 298, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 298, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 298, + 544 + ], + "type": "text", + "content": "[16, 26, 27, 34] with 2 to 4 candidate answers, enabling a more quantitative and objective evaluation compared to open-ended questions [4, 6, 33, 57]. Every question and answer within LHRS-Bench are manually crafted to guarantee precision and reliability. Totally, LHRS-Bench comprises 108 images and 690 questions. Some examples of LHRS-Bench are shown in supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 299, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 299, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 299, + 665 + ], + "type": "text", + "content": "The evaluation dimensions of LHRs-Bench encompass 5 top-level dimensions, further delineated into 11 fine-grained sub-dimensions, as shown in Tab. 1. These top-level dimensions include recognition, imagery, spatial-awareness, quantity, and reasoning. For detailed explanations about each dimensions, please refer to supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 313, + 513, + 471, + 653 + ], + "blocks": [ + { + "bbox": [ + 304, + 424, + 482, + 512 + ], + "lines": [ + { + "bbox": [ + 304, + 424, + 482, + 512 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 482, + 512 + ], + "type": "text", + "content": "Table 1: Overview of the dataset distribution. LHRS-Bench contains 108 RS images with 690 question-answer pairs, covering five major evaluation dimensions and 11 sub-dimensions of questions. Each sample is a human-annotated single-choice question that may involve one or more evaluation dimensions simultaneously." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 513, + 471, + 653 + ], + "lines": [ + { + "bbox": [ + 313, + 513, + 471, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 471, + 653 + ], + "type": "table", + "html": "
DimensionSub-dimensionCount
RecognitionIdentity634
Color113
Orientation39
Shape37
Area75
ImageryResolution21
Modality23
Spatial-awarenessLocation204
Distance22
Quantity-137
Reasoning-46
", + "image_path": "967fd325dec05d8af141afe85f16d102134a5245da9ceca656af534c0620d9a1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "text", + "content": "D. Muhtar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 304 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 304 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 304 + ], + "type": "image", + "image_path": "49dbacddee4fb62b00e112d482402d203ab36359fad30843dafe281939bbf15a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 357 + ], + "type": "text", + "content": "Fig. 3: Architecture of the proposed LHRS-Bot. LHRS-Bot employs learnable queries in conjunction with a vision perceiver to summarize multi-level visual representations. These representations are subsequently concatenated with language token embeddings for processing within the LLM." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 366, + 201, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 366, + 201, + 378 + ], + "spans": [ + { + "bbox": [ + 132, + 366, + 201, + 378 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 393, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 482, + 455 + ], + "type": "text", + "content": "We present LHRS-Bot to harness the full potential of LLMs and general visual encoders for enhanced RS images understanding and reasoning. This section delves into each components of LHRS-Bot, detailing how we achieve fully alignment between vision and language models across multiple levels. Subsequently, we explain the curriculum training strategy for training LHRS-Bot." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 475, + 261, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 261, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 261, + 485 + ], + "type": "text", + "content": "3.1 Model Architecture" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 497, + 481, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 481, + 532 + ], + "type": "text", + "content": "The overall architecture of LHRS-Bot is presented in Fig. 3. LHRS-Bot is primarily composed of three components: a universal vision encoder, a vision perceiver, and a foundational LLM." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Vision Encoder. LHRS-Bot utilizes the pre-trained CLIP visual encoder ViT-L/14 [40] with input resolution of " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " pixel for encoding images. The prior works [33,56,64] usually utilize the final hidden feature for abstracting the semantic content of an image. However, several studies have highlighted that features at different levels of a deep learning model embed distinct types of information [17,31,39]. Therefore, we argue that previous image encoding methods fail to sufficiently capture the semantic content of images, leading to a reliance on extensive image-text pairs for vision-language alignment. Consequently, we retain multiple hidden features from various layers of the image encoder. This approach offers a more thorough representation of the underlying image, facilitating a thorough and efficient domain alignment." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "LHRS-Bot" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 480, + 208 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 208 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 208 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 208 + ], + "type": "image", + "image_path": "23b5a95e198ddfb8378944814b1c364b70a84c8720f0ee3803d005bf9817c263.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 216, + 482, + 251 + ], + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 251 + ], + "type": "text", + "content": "Fig. 4: Curriculum learning strategy. We progressively unlock various parameters and introduce data of escalating complexity at different stages, enabling the model to gradually assimilate visual knowledge with increasing difficulty." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 257, + 482, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 388 + ], + "type": "text", + "content": "Vision Perceiver. Incorporating multi-level visual features brings extended image tokens, increasing computational overhead and potentially overwhelming language information. To mitigate this issue, we introduce a set of learnable queries for each level of image features. These queries are designed to summarize the semantic information of each level of visual features through a series of stacked cross-attention and MLP layers. Once summarized, these image features are then fed into the LLM. Furthermore, we implement a descending query allocation strategy inspired by the observation that image feature redundancy tends to increase with network depth [7,24]. This strategy allocates more queries to the shallower image feature levels while gradually reducing the number of queries for deeper levels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 426 + ], + "type": "text", + "content": "Large Language Model. We employ the LLaMA2-7B [48] as the central expert for interpreting visual and language information to output the desired results following the previous practices [25, 29, 43]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 433, + 313, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 433, + 313, + 445 + ], + "spans": [ + { + "bbox": [ + 131, + 433, + 313, + 445 + ], + "type": "text", + "content": "3.2 Curriculum Learning Strategy" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 452, + 482, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 452, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 482, + 500 + ], + "type": "text", + "content": "We implement a three-stage curriculum learning strategy by progressively increasing the difficulty of training tasks (Fig. 4). This strategy incrementally aligns the vision encoder with the LLM, ultimately enabling the LLM to solve multiple visual tasks and follow complex vision-language instructions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 500, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 482, + 608 + ], + "type": "text", + "content": "Stage 1: Pre-training. In the first stage, we use the large-scale, weakly labelled LHRS-Align dataset (Sec. 2.1) to train the vision perceiver to incorporate a wide range of RS visual knowledge into the LLM. Specifically, for each sample in the LHRS-Align dataset, the output visual embedding tokens " + }, + { + "bbox": [ + 130, + 500, + 482, + 608 + ], + "type": "inline_equation", + "content": "X_{v}" + }, + { + "bbox": [ + 130, + 500, + 482, + 608 + ], + "type": "text", + "content": " obtained from the vision perceiver are concatenated with the language embedding tokens " + }, + { + "bbox": [ + 130, + 500, + 482, + 608 + ], + "type": "inline_equation", + "content": "X_{l}" + }, + { + "bbox": [ + 130, + 500, + 482, + 608 + ], + "type": "text", + "content": " from the corresponding image caption. The concatenated result is then fed into the LLM. The objective of this stage is to maximise the conditional probability " + }, + { + "bbox": [ + 130, + 500, + 482, + 608 + ], + "type": "inline_equation", + "content": "P(X_{l}|X_{v};\\theta)" + }, + { + "bbox": [ + 130, + 500, + 482, + 608 + ], + "type": "text", + "content": " of generating the caption tokens " + }, + { + "bbox": [ + 130, + 500, + 482, + 608 + ], + "type": "inline_equation", + "content": "X_{l}" + }, + { + "bbox": [ + 130, + 500, + 482, + 608 + ], + "type": "text", + "content": " given the prefix visual tokens " + }, + { + "bbox": [ + 130, + 500, + 482, + 608 + ], + "type": "inline_equation", + "content": "X_{v}" + }, + { + "bbox": [ + 130, + 500, + 482, + 608 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 242, + 616, + 480, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 616, + 480, + 662 + ], + "spans": [ + { + "bbox": [ + 242, + 616, + 480, + 662 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\log P (X _ {l} | X _ {v}; \\theta) \\\\ = \\sum_ {i = 1} ^ {L} \\log P \\left(x _ {i} \\mid X _ {v}, X _ {l, < i}; \\theta\\right), \\tag {1} \\\\ \\end{array}", + "image_path": "9143b5ac411737df129535b8503f45787f8b3e2e6ebe37c89dcd5c2ce3f992c1.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "text", + "content": "D. Muhtar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " is the length of caption tokens, " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " represents the trainable parameters, and " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "X_{l, < i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " denotes the generated tokens prior to the current prediction token " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 308 + ], + "type": "text", + "content": "Stage 2: Multi-task Pre-training. In the second stage, we further train the vision perceiver and efficiently fine-tune the LLM using LoRA [20] on multitask instruction datasets to improve multimodal task solving capabilities of LLM. The construction of the multi-task instruction datasets begins with the collection of diverse public datasets in the RS domain, encompassing tasks such as classification, VQA, and visual grounding. These collected datasets are then meticulously transformed into instruction-based formats using manually designed templates. More detailed information about these multi-task datasets and the templates used can be found in the supplementary material. Moreover, we integrate the instruction data from NWPU, RSITMD and the 4K detailed descriptions data from LHRS-Instruct (Sec. 2.2) to further enrich the LLM with fine-grained RS visual knowledge. The objective for this stage remains similar to Eq. (1) from stage 1, albeit with an adaptation that the generated tokens are now additionally conditioned on the instruction tokens " + }, + { + "bbox": [ + 130, + 141, + 482, + 308 + ], + "type": "inline_equation", + "content": "X_{\\text{instruct}}" + }, + { + "bbox": [ + 130, + 141, + 482, + 308 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 220, + 319, + 480, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 319, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 220, + 319, + 480, + 365 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\log P (X _ {l} | X _ {v}, X _ {\\text {i n s t r u c t}}; \\theta) \\\\ = \\sum_ {i = 1} ^ {L} \\log P \\left(x _ {i} \\mid X _ {v}, X _ {\\text {i n s t r u c t}}, X _ {l, < i}; \\theta\\right). \\tag {2} \\\\ \\end{array}", + "image_path": "7f07c7805293954d66c1b59a8d7e5646267cbc62bb28b8b2b420ace408a7fc7a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 461 + ], + "type": "text", + "content": "Stage 3: Supervised Fine-tuning. In the final stage, our focus shifts towards enhancing the conversational and reasoning capabilities of our model by incorporating more complex instruction data. We utilize all the instruction data from the LHRS-Instruct dataset (Sec. 2.2) and randomly select 20K data from the LLaVA complex reasoning dataset [29]. Additionally, we mix the above datasets with the multi-task datasets from stage 2 using a lower data sampling ratio to further unleash the multi-task problem-solving ability of LHRS-Bot." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 473, + 224, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 473, + 224, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 473, + 224, + 487 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 502, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 575 + ], + "type": "text", + "content": "We conduct comprehensive experiments across image classification, VQA, and visual grounding tasks within the RS domain. These experiments are designed to rigorously validate the superior multi-task solving capabilities of LHRS-Bot. Furthermore, we evaluate the performance of various LLMs in the RS domain using the proposed LHRS-Bench benchmark. We present both quantitative and quality results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 594, + 262, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 262, + 607 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 262, + 607 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Dataset. We select diverse RS datasets for providing a comprehensive assessment across different tasks within the RS field: seven RS classification datasets (namely AID [53], WHU-RS19 [13], NWPU [9], SIRI-WHU [65], EuroSAT [18], METER-ML [63], and fMoW [12]), two VQA datasets (LR and HR subsets" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "LHRS-Bot" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 129, + 480, + 218 + ], + "blocks": [ + { + "bbox": [ + 165, + 114, + 447, + 126 + ], + "lines": [ + { + "bbox": [ + 165, + 114, + 447, + 126 + ], + "spans": [ + { + "bbox": [ + 165, + 114, + 447, + 126 + ], + "type": "text", + "content": "Table 2: Classification accuracy of different models on each dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 129, + 480, + 218 + ], + "lines": [ + { + "bbox": [ + 132, + 129, + 480, + 218 + ], + "spans": [ + { + "bbox": [ + 132, + 129, + 480, + 218 + ], + "type": "table", + "html": "
MethodAIDWHU-RS19NWPUSIRI-WHUEuroSATMETER-MLfMoWAvg.
LLaVA-1.531.1054.5534.9617.7126.1221.7311.4328.23
MiniGPTv232.9664.8028.1535.4638.5614.295.2031.35
InstructBLIP29.5036.7634.0118.2020.2514.426.7122.84
mPLUG-OWL248.7972.6646.5854.8333.2936.2717.8544.32
QWen-VL-Chat55.3072.2542.7354.5826.4238.776.8942.42
InternLM-XComposer51.6172.8947.5146.8339.7040.2111.2844.29
LHRS-Bot91.2693.1783.9462.6651.4069.8156.5671.83
", + "image_path": "42423866970876809e604dbc3f8175a9ab3eb2cb2935978df25024421f4da5d4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 232, + 480, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 480, + 281 + ], + "type": "text", + "content": "of RSVQA [35]), and two visual grounding datasets (RSVG [45] and DIOR-RSVG [59]). We exclude any data from the test sets that overlap with the training set to avoid data leakage. More details about each test dataset are presented in supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 287, + 494, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 494, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 494, + 361 + ], + "type": "text", + "content": "Baseline. We evaluate LHRS-Bot and various powerful open-source MLLMs, including LLaVA-v1.5 [32], MiniGPT-v2 [64], InstructBLIP [14], mPLUG-Owl2 [56], QWen-VL-Chat [3], and InternLM-Xcomposer [61], on different tasks. Furthermore, we also include the results of RSGPT [21], GeoChat [25] and SkyEyeGPT [60] from the original papers for comparisons whenever possible. Unless otherwise noted, we utilized 7B variants for all our experiments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "type": "text", + "content": "Implementation. We extract hidden states at layers " + }, + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\{i = L / 3, j = 2L / 3, k = L - 1\\}" + }, + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "type": "text", + "content": " for summarizing through vision perceiver, where " + }, + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "type": "text", + "content": " denotes the number of layers in the vision encoder. Our vision perceiver implementation comprises six layers of cross-attention and MLP, with each hidden state assigned " + }, + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\{64, 48, 32\\}" + }, + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "type": "text", + "content": " learnable queries. We apply LoRA module to every linear layer of the LLM, setting the rank " + }, + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "type": "text", + "content": " and the scale factor " + }, + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 367, + 482, + 487 + ], + "type": "text", + "content": " for LoRA to 128 and 256, respectively. Moreover, we add task identifier [CLS], [VQA], and [VG] for classification, VQA, and visual grounding tasks, respectively, for enabling LHRS-Bot to better distinguish different tasks [64]. For additional details on the hyperparameters used at each stage, please refer to supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 502, + 281, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 281, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 281, + 514 + ], + "type": "text", + "content": "4.2 Quantitative Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 481, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 559 + ], + "type": "text", + "content": "We evaluate different MLLMs using greedy decoding with the same prompt for each task and report the average result of two independent trials. Details about the evaluation prompts are provided in supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 564, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 482, + 639 + ], + "type": "text", + "content": "Classification. The classification results across different datasets are presented in Tab. 2. These results show that LHRS-Bot outperforms all competing models in all seven classification datasets, with a notable margin of " + }, + { + "bbox": [ + 130, + 564, + 482, + 639 + ], + "type": "inline_equation", + "content": "27.51\\%" + }, + { + "bbox": [ + 130, + 564, + 482, + 639 + ], + "type": "text", + "content": " higher average accuracy than the second ranked method. It is worth mentioning that, except for NWPU, METER-ML, and fMoW, the remaining datasets are completely absent from the multi-task training dataset6, demonstrating the strong" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "text", + "content": "D. Muhtar et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 131, + 642, + 482, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 642, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 642, + 482, + 666 + ], + "type": "text", + "content": "However, we have no means of confirming whether these data appeared in the LHRs-Align dataset, considering that images in LHRs-Align, along with several of the" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 149, + 152, + 462, + 271 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 147 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 147 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 147 + ], + "type": "text", + "content": "Table 3: Accuracy of different models on each visual question answering dataset. Following the practice of prior works [21, 25, 60], we omit area and count questions during evaluation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 152, + 462, + 271 + ], + "lines": [ + { + "bbox": [ + 149, + 152, + 462, + 271 + ], + "spans": [ + { + "bbox": [ + 149, + 152, + 462, + 271 + ], + "type": "table", + "html": "
MethodRSVQA-LRRSVQA-HR
Rural/UrbanPresenceCompareAvg.PresenceCompareAvg.
LLaVA-1.559.2273.1665.1965.8648.9659.0253.99
MiniGPTv260.0251.6467.6459.7768.3464.7166.53
InstructBLIP62.6248.8365.9259.1262.6362.9062.77
mPLUG-Owl257.9974.0463.6965.2447.6058.4753.04
QWen-VL-Chat62.0047.6566.5458.7361.7565.9863.87
InternLM-XCompose59.0066.7452.9159.5567.7966.6267.21
RSGPT94.0091.1791.7092.2990.9290.0290.47
GeoChat91.0990.3394.0091.8158.4583.1970.82
SkyEyeGPT88.9388.6375.0084.1680.0080.1382.56
LHRS-Bot89.0788.5190.0089.1992.5792.5392.55
", + "image_path": "70df29b5b48afce876174b97e55390e0030cf6a4362a9d8a1c619190d5bd24a3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 290, + 480, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 290, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 480, + 350 + ], + "type": "text", + "content": "generalization ability of LHRS-Bot. In addition, in the context of the fine-grained RS classification dataset fMoW, our method significantly surpasses other models, leading by an impressive " + }, + { + "bbox": [ + 132, + 290, + 480, + 350 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 132, + 290, + 480, + 350 + ], + "type": "text", + "content": " on average. This highlights the critical role of continue pre-training in the RS domain and the effectiveness of our alignment dataset LHRS-Align." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 379, + 480, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 379, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 379, + 480, + 475 + ], + "type": "text", + "content": "Visual Question Answering. We report the results of different models on RSVQA-LR and RSVQA-HR datasets in Tab. 3. LHRS-Bot demonstrates performance on par with the previously leading RS domain-specific models on the VQA task. Notably, LHRS-Bot achieves comparable results to RSGPT, despite RSGPT being fine-tuned for five epochs on the RSVQA datasets. Furthermore, LHRS-Bot sets a new record in the RSVQA-HR task with high-resolution RS images, showcasing its superior visual recognition abilities with high-resolution RS images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 495, + 297, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 297, + 627 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 297, + 627 + ], + "type": "text", + "content": "Visual Grounding. The results for the visual grounding task are shown in Tab. 4, where LHRS-Bot outperforms other counterparts. Notably, despite SkyEyeGPT being trained on a more extensive object detection dataset, LHRS-Bot demonstrates comparable performance. This suggests that LHRS-Bot has effectively acquired fine-grained visual information throughout the training process." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 552, + 468, + 612 + ], + "blocks": [ + { + "bbox": [ + 306, + 495, + 479, + 550 + ], + "lines": [ + { + "bbox": [ + 306, + 495, + 479, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 479, + 550 + ], + "type": "text", + "content": "Table 4: Comparison of visual grouting tasks. Accuracy is calculated based on whether the predicted bounding box overlaps with the ground-truth box by more than 0.5 IoU." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 552, + 468, + 612 + ], + "lines": [ + { + "bbox": [ + 315, + 552, + 468, + 612 + ], + "spans": [ + { + "bbox": [ + 315, + 552, + 468, + 612 + ], + "type": "table", + "html": "
MethodRSVGDIOR-RSVGAvg.
QWen-VL-Chat44.7680.6562.71
MiniGPTv246.6485.9966.32
SkyEyeGPT70.5088.5979.55
LHRS-Bot73.4588.1080.78
", + "image_path": "084468b90b955fdb307fb7d7059f85fc122b7c88d2a720f6d7a50fb40d1bada1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "LHRS-Bot" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 643, + 479, + 663 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 479, + 663 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 479, + 663 + ], + "type": "text", + "content": "classification datasets (AID, WHU-RS19, and SIRI-WHU), are collected from Google Earth." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 146, + 481, + 228 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 136 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 136 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 136 + ], + "type": "text", + "content": "Table 5: Comparison results on the LHRS-Bench dataset with different open-source and closed-source MLLMs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 146, + 481, + 228 + ], + "lines": [ + { + "bbox": [ + 132, + 146, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 481, + 228 + ], + "type": "table", + "html": "
MethodIdentityColorOrientationShapeAreaResolutionModalityLocationDistanceQuantityReasoningAvg.
LLaVA-1.542.9030.9715.3843.2444.0023.8021.7435.7827.2721.8954.3537.45
MiniGPTv220.1915.932.568.1113.3314.294.3515.2013.649.4936.9616.88
InstructBLIP20.3522.125.1324.3222.6738.1000.0013.7322.8513.1430.4318.87
mPLUG-OWL241.4845.1312.8240.5432.0033.3300.0030.8813.6432.1254.3437.00
QWen-VL-Chat31.3832.747.6932.4329.3319.054.3524.0221.4521.1756.5228.65
InternLM-XComposer34.3830.977.6937.8426.6700.008.7033.3322.7318.2565.2231.09
GPT-4-Turbo50.3243.3610.2667.5744.0042.8639.1342.1622.7321.8971.7444.56
Claude-3-Opus41.0136.2817.9564.8641.3366.6739.1334.8045.4526.2854.3539.08
LHRS-Bot44.6442.4823.0867.5725.3342.8632.3556.5222.7318.2573.9139.38
", + "image_path": "325e327ea26c956f8fa30f7c62c6f5dcd117078860e86e35ca2a986dda1c8787.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 243, + 299, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 243, + 299, + 254 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 299, + 254 + ], + "type": "text", + "content": "4.3 Evaluation on LHRS-Bench" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 269, + 481, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 481, + 353 + ], + "type": "text", + "content": "We evaluate LHRS-Bot with different open-source and closed-source MLLMs on our LHRS-Bench benchmark. For each question within the benchmark, choices are shuffled, and the generation process is repeated four times. An answer is considered incorrect if any attempt yields an incorrect response. Same prompts are used for evaluation, as detailed in supplementary material. We consider the model answers correctly if it outputs the correct choice or the answer corresponding to the correct choice. We calculate accuracy as our evaluation metric." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 354, + 481, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 481, + 462 + ], + "type": "text", + "content": "The results presented in Tab. 5 indicate that LHRS-Bot outperforms all other open-source models. It is worth noting that, LHRS-Bot greatly outperforms other open-source models in terms of the evaluation dimensions of spatial resolution and image modality, even though these aspects were not explicitly trained for in the models, highlighting the effectiveness of RS-specific MLLM training. Additionally, LHRS-Bot outperforms the strong closed-source model Claude and even matches the performance of GPT-4V. However, there is still significant room for improvement in MLLM capabilities within the RS domain. Examples of evaluation results are available on supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 480, + 274, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 480, + 274, + 492 + ], + "spans": [ + { + "bbox": [ + 132, + 480, + 274, + 492 + ], + "type": "text", + "content": "4.4 Qualitative Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 506, + 481, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 481, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 481, + 602 + ], + "type": "text", + "content": "We conduct qualitative demonstrations to illustrate the extensive capabilities of LHRS-Bot in interpreting free-form RS images, as showcased in Fig. 5. The results demonstrate that LHRS-Bot not only comprehends the overall content of an image but also exhibits a strong ability for discerning fine-grained details. Moreover, LHRS-Bot possesses a perceptive understanding of the orientation, position, and quantity of objects within an image. Furthermore, LHRS-Bot is capable of making reasoned inferences from the images and acknowledges its limitations by indicating uncertainty when it cannot provide a definitive answer." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 624, + 251, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 251, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 251, + 636 + ], + "type": "text", + "content": "4.5 Ablation Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 651, + 481, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 651, + 481, + 675 + ], + "spans": [ + { + "bbox": [ + 132, + 651, + 481, + 675 + ], + "type": "text", + "content": "In this subsection, we carry out extensive ablation experiments to demonstrate the effectiveness of our dataset as well as the model design." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "text", + "content": "D. Muhtar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 272 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 272 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 272 + ], + "type": "image", + "image_path": "a6287d5176ccce0e788920bc7dfc760d7e45e5a1beea22f7606e382ab9ab466e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 280, + 436, + 291 + ], + "lines": [ + { + "bbox": [ + 175, + 280, + 436, + 291 + ], + "spans": [ + { + "bbox": [ + 175, + 280, + 436, + 291 + ], + "type": "text", + "content": "Fig. 5: Qualitative examples between real users and LHRS-Bot." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 132, + 319, + 482, + 360 + ], + "blocks": [ + { + "bbox": [ + 192, + 298, + 420, + 309 + ], + "lines": [ + { + "bbox": [ + 192, + 298, + 420, + 309 + ], + "spans": [ + { + "bbox": [ + 192, + 298, + 420, + 309 + ], + "type": "text", + "content": "Table 6: Performance with different alignment dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 319, + 482, + 360 + ], + "lines": [ + { + "bbox": [ + 132, + 319, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 482, + 360 + ], + "type": "table", + "html": "
AIDWHU-RS19ClassificationVQAVisual Grounding
NWPUSIRI-WHUEuroSATMETER-MLfMoWRSVQA-LRRSVQA-HRRSVGDIOR-RSVG
RS5M86.3196.7582.8060.2054.9264.7742.6878.3691.6570.5887.12
SkyScript88.2190.2678.0151.4343.6650.0231.2788.7991.2471.2585.05
LHRS-Align91.2693.1783.9462.6651.4069.8156.5689.1992.5573.4588.10
", + "image_path": "64cbb8a8c919864d12d22fea3a7b2ee9d32861cf06c0c4bded6f348161ef8eba.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 382, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 482, + 465 + ], + "type": "text", + "content": "Effectiveness of LHRS-Align dataset. There are already many large-scale image-caption datasets in the field of RS [50,62]. To evaluate how different alignment datasets impact model performance, we substitute our LHRS-Align dataset with various image-caption datasets and repeat the entire training process. This process is then evaluated across three distinct tasks: classification, VQA, and visual grounding. The experimental setup remains consistent with Sec. 4.1, with the exception of the dataset used in the first stage." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 466, + 482, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 466, + 482, + 585 + ], + "spans": [ + { + "bbox": [ + 130, + 466, + 482, + 585 + ], + "type": "text", + "content": "The results are presented in Tab. 6. Notably, despite possessing the smallest dataset (RS5M: 5M, SkyScript: 1.5M, LHRS-Align : 1.15M), LHRS-Align outperforms other settings. It is worth mentioning that both SkyScript and LHRS-Align, which leverage RS images combined with the OSM database, achieve results comparable to or better than the even larger RS5M, which is derived from publicly available natural image datasets, such as LAION [44]. This highlights the importance of domain-specific data collection. Additionally, superior performance of LHRS-Align over SkyScript suggests that enriched captions and a meticulous data cleaning pipeline can substantially enhance the effectiveness of MLLMs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Effectiveness of Model Design. We evaluate the effectiveness of designs involving multi-level visual hidden representations as well as descending query allocation strategies. Specifically, our evaluation focuses on comparing with the single level visual hidden state, projecting it through a two-layer MLP or a vision perceiver. We also perform comparisons between multi-level visual hidden" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "LHRS-Bot" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 152, + 136, + 461, + 216 + ], + "blocks": [ + { + "bbox": [ + 161, + 114, + 451, + 126 + ], + "lines": [ + { + "bbox": [ + 161, + 114, + 451, + 126 + ], + "spans": [ + { + "bbox": [ + 161, + 114, + 451, + 126 + ], + "type": "text", + "content": "Table 7: Performance comparison across various architectural designs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 152, + 136, + 461, + 216 + ], + "lines": [ + { + "bbox": [ + 152, + 136, + 461, + 216 + ], + "spans": [ + { + "bbox": [ + 152, + 136, + 461, + 216 + ], + "type": "table", + "html": "
MethodMulti Level Number of Query ClassificationVQA Visual Grounding
MLP✘-64.4690.06
Vision Perceiver✘6463.6983.30
Vision Perceiver✓[64,64,64]63.7487.88
Vision Perceiver✓[32,48,64]60.6887.11
Token Merging✓6449.5181.49
Vision Perceiver✓[64,48,32]65.4390.14
", + "image_path": "09e8ddafa2e97a06121eda032b6efd4852dc31821b56cd6e3bfb2744617ab91b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 226, + 481, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 481, + 320 + ], + "type": "text", + "content": "representations, employing either identical queries for each level or incremental query allocation strategies. Building on the concept introduced by the token merging [7], we explore the potential of integrating this strategy with a vision perceiver to efficiently leverage multi-level representations. To minimize experimental costs, our training process comprises just two stages: firstly, we employ LHRS-Align to train the bridge layer, which aligns the vision and language domains. Subsequently, we train the model using the LHRS-Instruct and multi-task instruction datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 323, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 455 + ], + "type": "text", + "content": "We report the average result on each task in Tab. 7. It is evident that the approach employed by LHRS-Bot achieves superior performance. The results further illustrate that, although the extraction of a singular visual representation performs better in tasks such as classification and VQA tasks, which depends on high-level visual semantics, it falls short in visual grounding tasks that require detailed visual information. Furthermore, employing the same number of queries to summarise different levels of visual information does not lead to better results. In contrast, adopting a diminishing allocation strategy yields superior results. This further demonstrates the large redundancy among the deep visual tokens [7]. Lastly, the token merging strategy proves ineffective, possibly due to the loss of crucial information during the merging process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 478, + 220, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 478, + 220, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 478, + 220, + 491 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 509, + 481, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 481, + 580 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 481, + 580 + ], + "type": "text", + "content": "We propose LHRS-Bot, an MLLM for the RS domain. To unleash the potential of LLMs for RS image understanding, we curate a large-scale dataset, LHRS-Align, for RS-specific alignment, and LHRS-Instruct, a multimodal instruction-following dataset to enhance LHRS-Bot's instruction-following capabilities. Additionally, we introduce LHRS-Bench, a benchmark providing a systematic evaluation framework for RS MLLMs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "In comprehensive comparison experiments, LHRS-Bot demonstrates exceptional RS domain knowledge, surpassing both general-purpose and RS-specific MLLMs. Despite its impressive performance, LHRS-Bot shares limitations common to MLLMs, such as susceptibility to hallucinations. We believe that developing higher-quality alignment and instructional datasets, along with improved training strategies, can further enhance the performance of MLLMs for interpreting RS images." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "text", + "content": "D. Muhtar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 214 + ], + "type": "text", + "content": "This research was supported by the National Natural Science Foundation of China under Grant 42071297, and in part by the AI & AI for Science Project of Nanjing University under Grant 02091480605203. We are grateful to High Performance Computing Center of Nanjing University for their help on GPU resources. We also would like to thank the anonymous reviewers for their constructive comments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 236, + 197, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 197, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 197, + 248 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 263, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 263, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 263, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 263, + 481, + 308 + ], + "type": "text", + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 309, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 481, + 331 + ], + "type": "text", + "content": "2. Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., et al.: Qwen technical report. arXiv preprint arXiv:2309.16609 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 332, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 481, + 364 + ], + "type": "text", + "content": "3. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 365, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 481, + 397 + ], + "type": "text", + "content": "4. Bai, S., Yang, S., Bai, J., Wang, P., Zhang, X., Lin, J., Wang, X., Zhou, C., Zhou, J.: Touchstone: Evaluating vision-language models by language models. arXiv preprint arXiv:2308.16890 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 399, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 481, + 430 + ], + "type": "text", + "content": "5. Bashmal, L., Bazi, Y., Melgani, F., Al Rahhal, M.M., Al Zuair, M.A.: Language integration in remote sensing: Tasks, datasets, and future directions. IEEE Geoscience and Remote Sensing Magazine (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 432, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 481, + 464 + ], + "type": "text", + "content": "6. Bitton, Y., Bansal, H., Hessel, J., Shao, R., Zhu, W., Awadalla, A., Gardner, J., Taori, R., Schimdt, L.: Visit-bench: A benchmark for vision-language instruction following inspired by real-world use. arXiv preprint arXiv:2308.06595 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 466, + 481, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 466, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 466, + 481, + 487 + ], + "type": "text", + "content": "7. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 488, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 481, + 520 + ], + "type": "text", + "content": "8. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 522, + 481, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 481, + 543 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 481, + 543 + ], + "type": "text", + "content": "9. Cheng, G., Han, J., Lu, X.: Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE 105(10), 1865-1883 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 544, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 544, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 134, + 544, + 481, + 586 + ], + "type": "text", + "content": "10. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 134, + 544, + 481, + 586 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 134, + 544, + 481, + 586 + ], + "type": "text", + "content": " * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 589, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 589, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 589, + 481, + 632 + ], + "type": "text", + "content": "11. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. Journal of Machine Learning Research 24(240), 1-113 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "text", + "content": "12. Christie, G., Fendley, N., Wilson, J., Mukherjee, R.: Functional map of the world. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6172-6180 (2018)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "LHRS-Bot" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "13. Dai, D., Yang, W.: Satellite image classification via two-layer sparse coding with biased image representation. IEEE Geoscience and Remote Sensing Letters 8(1), 173-176 (2011)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "text", + "content": "14. Dai, W., Li, J., Li, D., Tiong, A., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arxiv 2023. arXiv preprint arXiv:2305.06500" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "type": "text", + "content": "15. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 481, + 246 + ], + "type": "text", + "content": "16. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 246, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 246, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 133, + 246, + 481, + 278 + ], + "type": "text", + "content": "17. Ghiasi, A., Kazemi, H., Borgnia, E., Reich, S., Shu, M., Goldblum, M., Wilson, A.G., Goldstein, T.: What do vision transformers learn? a visual exploration. arXiv preprint arXiv:2212.06727 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 278, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 278, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 133, + 278, + 481, + 322 + ], + "type": "text", + "content": "18. Helber, P., Bischke, B., Dengel, A., Borth, D.: Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 12(7), 2217-2226 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 322, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 322, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 133, + 322, + 481, + 354 + ], + "type": "text", + "content": "19. Hossain, M.D., Chen, D.: Segmentation for object-based image analysis (obia): A review of algorithms and challenges from remote sensing perspective. ISPRS Journal of Photogrammetry and Remote Sensing 150, 115-134 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 133, + 354, + 481, + 386 + ], + "type": "text", + "content": "20. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 386, + 481, + 407 + ], + "type": "text", + "content": "21. Hu, Y., Yuan, J., Wen, C., Lu, X., Li, X.: Rsgpt: A remote sensing vision language model and benchmark. arXiv preprint arXiv:2307.15266 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 407, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 407, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 133, + 407, + 481, + 440 + ], + "type": "text", + "content": "22. Huang, S., Dong, L., Wang, W., Hao, Y., Singhal, S., Ma, S., Lv, T., Cui, L., Mohammed, O.K., Liu, Q., et al.: Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 440, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 440, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 133, + 440, + 481, + 472 + ], + "type": "text", + "content": "23. Jiang, A.Q., Sablayrolles, A., Roux, A., Mensch, A., Savary, B., Bamford, C., Chaplot, D.S., Casas, D.d.l., Hanna, E.B., Bressand, F., et al.: Mixtral of experts. arXiv preprint arXiv:2401.04088 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 472, + 481, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 472, + 481, + 505 + ], + "spans": [ + { + "bbox": [ + 133, + 472, + 481, + 505 + ], + "type": "text", + "content": "24. Ju, C., Wang, H., Li, Z., Chen, X., Zhai, Z., Huang, W., Xiao, S.: Turbo: Informativity-driven acceleration plug-in for vision-language models. arXiv preprint arXiv:2312.07408 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 505, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 505, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 133, + 505, + 481, + 536 + ], + "type": "text", + "content": "25. Kuckreja, K., Danish, M.S., Naseer, M., Das, A., Khan, S., Khan, F.S.: Geochat: Grounded large vision-language model for remote sensing. arXiv preprint arXiv:2311.15826 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 536, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 536, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 133, + 536, + 481, + 568 + ], + "type": "text", + "content": "26. Li, B., Ge, Y., Ge, Y., Wang, G., Wang, R., Zhang, R., Shan, Y.: Seed-bench-2: Benchmarking multimodal large language models. arXiv preprint arXiv:2311.17092 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 133, + 568, + 481, + 601 + ], + "type": "text", + "content": "27. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 133, + 601, + 481, + 633 + ], + "type": "text", + "content": "28. Li, C., Gan, Z., Yang, Z., Yang, J., Li, L., Wang, L., Gao, J.: Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 481, + 665 + ], + "type": "text", + "content": "29. Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H., Gao, J.: Llava-med: Training a large language-and-vision assistant for biomedicine in one day. arXiv preprint arXiv:2306.00890 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "text", + "content": "D. Muhtar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "30. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "31. Li, M., Wang, S., Zhang, Q.: Visualizing the emergence of intermediate visual patterns in dnns. Advances in Neural Information Processing Systems 34, 6594-6607 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "32. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 225 + ], + "type": "text", + "content": "33. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "34. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "35. Lobry, S., Marcos, D., Murray, J., Tuia, D.: Rsvqa: Visual question answering for remote sensing data. IEEE Transactions on Geoscience and Remote Sensing 58(12), 8555-8566 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "text", + "content": "36. Ma, Y., Cao, Y., Sun, J., Pavone, M., Xiao, C.: Multimodal language model for driving. arXiv preprint arXiv:2312.00438 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "text", + "content": "37. Muhtar, D., Zhang, X., Xiao, P., Li, Z., Gu, F.: Cmid: A unified self-supervised learning framework for remote sensing image understanding. IEEE Transactions on Geoscience and Remote Sensing (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 307, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 307, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 307, + 358 + ], + "type": "text", + "content": "38. OpenAI: Gpt-4 technical report (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "type": "text", + "content": "39. Park, N., Kim, W., Heo, B., Kim, T., Yun, S.: What do self-supervised vision transformers learn? arXiv preprint arXiv:2305.00729 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 423 + ], + "type": "text", + "content": "40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "text", + "content": "41. Ratledge, N., Cadamuro, G., de la Cuesta, B., Stigler, M., Burke, M.: Using machine learning to assess the livelihood impact of electricity access. Nature 611(7936), 491-495 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 500 + ], + "type": "text", + "content": "42. Reed, C.J., Gupta, R., Li, S., Brockman, S., Funk, C., Clipp, B., Keutzer, K., Candido, S., Uytendaele, M., Darrell, T.: Scale-MAE: A scale-aware masked autoencoder for multiscale geospatial representation learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4088-4099 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "text", + "content": "43. Saygin Seyfioglu, M., Ikezogwo, W.O., Ghezloo, F., Krishna, R., Shapiro, L.: Quilt-lava: Visual instruction tuning by extracting localized narratives from open-source histopathology videos. arXiv e-prints pp. arXiv-2312 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 578 + ], + "type": "text", + "content": "44. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "text", + "content": "45. Sun, Y., Feng, S., Li, X., Ye, Y., Kang, J., Huang, X.: Visual grounding in remote sensing images. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 404-412 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "text", + "content": "46. Team, M.N.: Introducing mpt-7b: A new standard for open-source, commercially usable llms (2023), www.mosaicml.com/blog/mpt-7b, accessed: 2023-05-05" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "47. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "LHRS-Bot" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "48. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 192 + ], + "type": "text", + "content": "49. Wang, D., Zhang, J., Du, B., Xu, M., Liu, L., Tao, D., Zhang, L.: Samrs: Scaling-up remote sensing segmentation dataset with segment anything model. In: Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "text", + "content": "50. Wang, Z., Prabha, R., Huang, T., Wu, J., Rajagopal, R.: Skyscript: A large and semantically diverse vision-language dataset for remote sensing. arXiv preprint arXiv:2312.12856 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 224, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 257 + ], + "type": "text", + "content": "51. Wen, C., Hu, Y., Li, X., Yuan, Z., Zhu, X.X.: Vision-language models in remote sensing: Current progress and future trends. arXiv preprint arXiv:2305.05726 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 257, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 300 + ], + "type": "text", + "content": "52. Xia, G.S., Bai, X., Ding, J., Zhu, Z., Belongie, S., Luo, J., Datcu, M., Pelillo, M., Zhang, L.: DOTA: A large-scale dataset for object detection in aerial images. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3974-3983 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "text", + "content": "53. Xia, G.S., Hu, J., Hu, F., Shi, B., Bai, X., Zhong, Y., Zhang, L., Lu, X.: AID: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing 55(7), 3965-3981 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "type": "text", + "content": "54. Xu, H., Xie, S., Tan, X.E., Huang, P.Y., Howes, R., Sharma, V., Li, S.W., Ghosh, G., Zettlemoyer, L., Feichtenhofer, C.: Demystifying clip data. arXiv preprint arXiv:2309.16671 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 365, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 396 + ], + "type": "text", + "content": "55. Yang, J., Gong, P., Fu, R., Zhang, M., Chen, J., Liang, S., Xu, B., Shi, J., Dickinson, R.: The role of satellite remote sensing in climate change studies. Nature climate change 3(10), 875-883 (2013)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 396, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 429 + ], + "type": "text", + "content": "56. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "type": "text", + "content": "57. Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mmvet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 460, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 480, + 493 + ], + "type": "text", + "content": "58. Yuan, Z., Zhang, W., Fu, K., Li, X., Deng, C., Wang, H., Sun, X.: Exploring a fine-grained multiscale method for cross-modal remote sensing image retrieval. arXiv preprint arXiv:2204.09868 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "type": "text", + "content": "59. Zhan, Y., Xiong, Z., Yuan, Y.: Rsvg: Exploring data and models for visual grounding on remote sensing data. IEEE Transactions on Geoscience and Remote Sensing 61, 1-13 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 557 + ], + "type": "text", + "content": "60. Zhan, Y., Xiong, Z., Yuan, Y.: Skyeyegpt: Unifying remote sensing vision-language tasks via instruction tuning with large language model. arXiv preprint arXiv:2401.09712 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 557, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 600 + ], + "type": "text", + "content": "61. Zhang, P., Wang, X.D.B., Cao, Y., Xu, C., Ouyang, L., Zhao, Z., Ding, S., Zhang, S., Duan, H., Yan, H., et al.: Internl m-xcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "text", + "content": "62. Zhang, Z., Zhao, T., Guo, Y., Yin, J.: Rs5m: A large scale vision-language dataset for remote sensing vision-language foundation model. arXiv preprint arXiv:2306.11300 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "63. Zhu, B., Lui, N., Irvin, J., Le, J., Tadwalkar, S., Wang, C., Ouyang, Z., Liu, F.Y., Ng, A.Y., Jackson, R.B.: Meter-ml: A multi-sensor earth observation benchmark for automated methane source mapping. arXiv preprint arXiv:2207.11166 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "text", + "content": "D. Muhtar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 183 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "64. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "text", + "content": "65. Zhu, Q., Zhong, Y., Zhao, B., Xia, G.S., Zhang, L.: Bag-of-visual-words scene classifier with local and global features for high spatial resolution remote sensing imagery. IEEE Geoscience and Remote Sensing Letters 13(6), 747-751 (2016)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "LHRS-Bot" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LISO_ Lidar-only Self-Supervised 3D Object Detection/d6e4374b-831f-4119-9bbf-2ac849c4dff4_content_list.json b/2024/LISO_ Lidar-only Self-Supervised 3D Object Detection/d6e4374b-831f-4119-9bbf-2ac849c4dff4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..89961bae5b91c27f3cbe58db27b87825d2b69417 --- /dev/null +++ b/2024/LISO_ Lidar-only Self-Supervised 3D Object Detection/d6e4374b-831f-4119-9bbf-2ac849c4dff4_content_list.json @@ -0,0 +1,1830 @@ +[ + { + "type": "text", + "text": "LISO: Lidar-only Self-Supervised 3D Object Detection", + "text_level": 1, + "bbox": [ + 250, + 140, + 751, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stefan Andreas Baur $^{1,2}$ , Frank Moosmann $^{1}$ , and Andreas Geiger $^{2,3}$", + "bbox": [ + 246, + 210, + 754, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Mercedes-Benz, Germany", + "bbox": [ + 406, + 239, + 594, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 University of Tübingen", + "bbox": [ + 416, + 253, + 584, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Tübingen AI Center", + "bbox": [ + 424, + 267, + 576, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. 3D object detection is one of the most important components in any Self-Driving stack, but current state-of-the-art (SOTA) lidar object detectors require costly & slow manual annotation of 3D bounding boxes to perform well. Recently, several methods emerged to generate pseudo ground truth without human supervision, however, all of these methods have various drawbacks: Some methods require sensor rigs with full camera coverage and accurate calibration, partly supplemented by an auxiliary optical flow engine. Others require expensive high-precision localization to find objects that disappeared over multiple drives.", + "bbox": [ + 261, + 321, + 740, + 446 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce a novel self-supervised method to train SOTA lidar object detection networks, requiring only unlabeled sequences of lidar point clouds. We call this trajectory-regularized self-training. It utilizes a SOTA self-supervised lidar scene flow network under the hood to generate, track, and iteratively refine pseudo ground truth. We demonstrate the effectiveness of our approach for multiple SOTA object detection networks across multiple real-world datasets. Code will be released1.", + "bbox": [ + 259, + 446, + 740, + 544 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Self-Supervised $\\cdot$ LiDAR $\\cdot$ Object Detection", + "bbox": [ + 261, + 558, + 643, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 599, + 375, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human developmental research reveals that infants less than one year of age are able to categorize animate and inanimate objects based on observed motion cues and can generalize this categorization to previously unseen objects [27]. Yet, lidar object detectors with SOTA performance are trained using manually selected and categorized annotations. These annotations are very expensive to obtain and become outdated quickly: new lidar sensors are coming to market regularly, and trained SOTA object detectors are sensitive to changes of sensor characteristics, mounting position and other domain gaps [5,8,18,26,32,34], even varying object sizes across different countries [29], making existing annotations difficult to transfer, resulting in high re-labeling efforts with each change.", + "bbox": [ + 212, + 633, + 787, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we aim at bridging this gap by distilling motion cues observed in self-supervised lidar scene flow into SOTA single-frame lidar object detectors.", + "bbox": [ + 212, + 785, + 785, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1 https://github.com/baurst/liso", + "bbox": [ + 217, + 824, + 467, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e2bc39fdbf5cf91e3443405a50bfc7572b857682d65f6605438fa6c1ee13a9a6.jpg", + "image_caption": [ + "Fig. 1: Objects predicted by our method using no manual annotations. Red boxes are ground truth boxes, yellow boxes are predicted by our network." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 498, + 305 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7719cef71cd1bfe6ca48e6dbb9bafb55519594dd4b3a49aaf3650b086a01579c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 143, + 785, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce a novel and robust self-supervised, easy-to-use method, operating solely on lidar point cloud sequences for easy-to-obtain $360^{\\circ}$ annotations. No human-annotated bounding boxes, cameras, costly high-precision GPS, or tedious sensor rig calibration is required, which could be especially useful in robotics research where data is often limited (even KITTI [11] only has forward facing cameras) or maintaining a multi-sensor-setup including its calibration is too cumbersome and expensive. Our method trains and runs a self-supervised lidar flow estimator [3] under the hood in order to create motion cues. We show in our experiments that these motion cues are a key factor to the success of our method. Based on the estimated lidar flow we bootstrap initial pseudo ground truth using simple clustering and track optimization. With these mined bounding boxes of moving objects we initialize a self-supervised, trajectory-regularized variant of self-training [1] (which is semi-supervised in its original form): We train a first version of a SOTA object detector, then iteratively re-generate and trajectory-regularize the pseudo ground truth and retrain the detector. Since the single-frame object detector has no concept of motion, it generalizes to detect any movable object on the way. Exemplary output of the trained detector, 3D boxes in single-frame point clouds, is depicted in Fig. 1.", + "bbox": [ + 212, + 382, + 787, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contribution is a novel self-supervised trajectory-regularized self-training framework for single-frame 3D object detection with the following properties:", + "bbox": [ + 212, + 657, + 787, + 688 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- It is based entirely on lidar, i.e. without the limitations of prior works: no cameras, no calibration, no high-precision GPS, no manual annotations.", + "- It is agnostic w.r.t. the sensor model, mounting pose, and detectors architecture and works with the same set of hyper parameters: We demonstrate this across four different datasets and different SOTA detector networks.", + "- It is able to generalize from moving objects (motion cues) to movable objects (final detection results) and significantly outperforms SOTA methods while being simpler." + ], + "bbox": [ + 223, + 710, + 782, + 835 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Baur et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We show that using motion cues together with trajectory-regularized self-training is key to this success. The code of this approach will be published for easy use and comparison by other researchers.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The paper is organized as follows: Sec. 2 discusses related work before the proposed method is described in detail in Sec. 3. Extensive experiments in Sec. 4 show the performance of our method before we conclude in Sec. 5.", + "bbox": [ + 212, + 191, + 784, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 260, + 387, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Single Frame Lidar 3D Object Detection", + "text_level": 1, + "bbox": [ + 214, + 292, + 596, + 308 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object detectors operating on 3D point clouds are an active research field. The currently best-performing ones are using deep neural networks trained via supervised learning and can be categorized by their internal representation: Some networks operate on points directly like PointRCNN [21], 3DSSD [35], and IASSD [39]. Others project the points either to a virtual range image [13,14,25] or into a voxel representation like VoxelNet [40], PointPillars, CenterPoint [36], and Transfusion-L [2]. However, all aforementioned methods require large human-annotated datasets in order to perform well and obtaining such annotations is very expensive. We address this problem in this paper, enabling training of SOTA object detectors using pseudo ground truth.", + "bbox": [ + 212, + 319, + 784, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Object Distilling from Motion Cues", + "text_level": 1, + "bbox": [ + 214, + 493, + 555, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multiple approaches have been suggested which leverage motion cues from lidar frame pairs in order to detect moving objects. Using the assumption of local geometric constancy, they decompose dynamic scenes into separate moving entities by applying as-rigid-as-possible optimization. Examples of these are the works by Dewan et al. [7], RSF [6] and OGC [23]. The two former methods are optimization-based whereas OGC uses these constraints as loss-function to self-supervise a segmentation network. Although these methods avoid the need for expensive labels, in contrast to our method they tend not to work well in low-resolution areas, they are typically slow, and can only detect moving but not movable (i.e. static but potentially moving) objects.", + "bbox": [ + 212, + 518, + 784, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3 Pseudo Ground Truth for Object Detection", + "text_level": 1, + "bbox": [ + 214, + 691, + 620, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different approaches have been proposed to mine pseudo ground truth for training object detectors: Najibi et al. [15] and Seidenschwarz et al. [19] use motion cues similar to section 2.2 in order to distill moving objects as pseudo ground truth and use it to train an object detector. [15] runs optimization for each frame pair to obtain lidar scene flow, clusters points, fits boxes, tracks them using a Kalman Filter and finally refines them on ICP-registered point clouds. [19] optimizes a clustering algorithm on motion cues through message passing which they then apply to segment point clouds into a set of instances and subsequently", + "bbox": [ + 212, + 719, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "LISO: Lidar-only Self-Supervised 3D Object Detection", + "bbox": [ + 367, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "fit boxes. As shown in [19], and in contrast to our approach, both methods suffer from a large performance gap between moving and movable objects. We demonstrate that our method does not suffer from this gap.", + "bbox": [ + 212, + 146, + 784, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[12], [22] were the first to iteratively apply a detection, tracking, retraining paradigm to autonomous driving (AD) data. They build upon consistency constraints between object detectors which they train in the lidar and camera domain, making use of an optical flow network which is trained using supervision. The approach requires a calibrated sensor rig with possibly large camera coverage, IMU, and precision GPS. Similarly, [30] uses video sequences together with lidar scene flow to jointly train a camera and a lidar object detector. MODEST [33, 37] does not require coverage by calibrated cameras, but instead adds the additional requirement to having multiple lidar recordings of the same location in order to identify objects that vanished over time. With such demanding requirements, these approaches are not easy to use and are partially unsuited for popular AD datasets.", + "bbox": [ + 212, + 191, + 787, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Oyster [38] is the approach most similar to ours. It uses DBSCAN [9] on lidar point clouds to initialize pseudo ground truth. Clusters are then tracked using forward-and-reverse tracking in sensor coordinates (i.e. without consideration for ego motion), using a complex policy for confidence-based track retention. After training an object detector on close range data, they employ zero-shot generalization to the far range data, track again and iteratively retrain the detector. In their experiments they use different hyperparameters for different datasets. Our method, in contrast, explicitly considers sensor motion, produces much cleaner initial proposals by leveraging a self-supervised lidar scene flow network, does not require zero-shot-generalization from near-to-far-field, and works with current SOTA object detectors. We show on multiple datasets that our method outperforms [38], that it works robustly with the same set of hyper-parameters and also generalizes well to detect movable objects. Additionally, we will release our code to facilitate further research on this topic.", + "bbox": [ + 212, + 373, + 787, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 604, + 330, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A general overview of our method is sketched in Fig. 2 and some steps are illustrated in Fig. 3. As input, we take raw (unlabeled) point cloud sequences and undergo three stages, all of which are detailed in the following: Preprocessing of point clouds and lidar scene flow computation (Sec. 3.1), initial pseudo ground truth generation (Sec. 3.2) and repeated training with pseudo ground truth refinement (Sec. 3.3). The final output of the method is a trained object detector which can detect movable objects in raw single-frame point clouds.", + "bbox": [ + 212, + 632, + 787, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preprocessing", + "text_level": 1, + "bbox": [ + 215, + 757, + 377, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We preprocess the raw input point clouds as follows:", + "bbox": [ + 212, + 779, + 596, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ground Removal: First, we remove distracting ground points from each single point cloud using JCP [20], which is a simple, robust, yet effective algorithm to remove ground points using changes in observed height above ground.", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Baur et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a2419607cf4bada2dba4040260d1595d7e5c5598b8a839ccf64da75de44c50fc.jpg", + "image_caption": [ + "Fig. 2: Overview of the proposed method. Point cloud sequences are preprocessed (blue, Sec 3.1), initial pseudo ground truth is created (orange, Sec. 3.2) and the object detector is iteratively trained and pseudo ground truth regenerated (red, Sec. 3.2)." + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 772, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ego Motion Estimation: Second, we compute ego-motion between neighboring frame pairs using KISS-ICP [28], which is based on a robust version of ICP. The output is a cm-level accurate transformation $\\mathbf{T}_{\\mathrm{ego}}^{t\\rightarrow t + 1}\\in \\mathbb{R}^{4\\times 4}$ , describing the ego vehicle position at time $t + 1$ represented in the ego frame from time $t$ .", + "bbox": [ + 212, + 347, + 782, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lidar Scene Flow Estimation: Third, we compute lidar scene flow between neighboring frame pairs resulting in a flow vector $\\mathbf{f}_i = (dx, dy, dz)$ for every point $i$ in the first point cloud $\\mathcal{P}^t$ . We chose to use SLIM [3] as its code is readily available, it is easy to use, features fast inference, and produces SOTA results. The network is trained self-supervised on raw point cloud sequences, minimizing a k-nearest-neighbor loss between forward and time-reversed point clouds.", + "bbox": [ + 212, + 407, + 782, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The components of our preprocessing steps (Ground Removal, Ego Motion Estimation, Lidar Scene Flow Estimation) have been selected for robustness and are all used with their default parameters from their respective publications.", + "bbox": [ + 212, + 500, + 784, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Initial Pseudo Ground Truth Generation", + "text_level": 1, + "bbox": [ + 214, + 570, + 599, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The aim of our method is to mine pseudo ground truth for training a 3D lidar object detector. For a single point cloud $\\mathcal{P}^t$ (consisting of $n \\in \\mathbb{N}$ points $\\mathbf{p}_i$ , $\\mathbf{p}_i \\in \\mathcal{P}^t$ ) this ground truth is a set of 3D bounding boxes $\\mathcal{B}$ with confidences representing the objects at time $t$ :", + "bbox": [ + 212, + 599, + 784, + 660 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} ^ {t} = \\left\\{\\mathrm {B} _ {j} ^ {t}, j \\in \\mathbb {N} \\right\\} = \\left\\{(x, y, z, l, w, h, \\theta , c) _ {j} \\right\\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 672, + 785, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $(x,y,z) = \\mathbf{x}$ define the center position, $(l,w,h)$ the length, width and height, $\\theta$ the heading (orientation around up axis), and $c$ the confidence for a single box.", + "bbox": [ + 212, + 702, + 782, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The key success factor of our method is to focus on a high precision (and potentially low recall) of the initial set of bounding boxes in order to avoid \"wrong\" objects to negatively influence the object detector. We achieve this by leveraging shape, density, and especially motion cues to robustly identify moving objects solely (see also top-right of Fig. 3). Sec. 3.3 targets at generalizing these to movable objects later on.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "LISO: Lidar-only Self-Supervised 3D Object Detection", + "bbox": [ + 367, + 114, + 730, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f3c4d1e935ad2a4bde8085c20d1a5f696b3250cf82c8b899d4479722a9be7957.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 143, + 500, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3493f65229214257f38198882a27355dc4737ed0976b4ef54c51808fa6e87b19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 143, + 781, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/97756f71ce6ff3ab58addacbf7064edebf7a1d303c00ab3b9ffe011287d42bbb.jpg", + "image_caption": [ + "Fig. 3: Overview over preprocessing, initial pseudo ground truth generation and training with examples. Top left: In the first step, self-supervised lidar scene flow is computed and corrected for vehicle ego-motion. Points are colored by flow direction and magnitude. Top right: In the second step, the scene flow is clustered and bounding boxes are fitted (to the moving objects). Bottom left: In the third step, the network is trained on the pseudo ground truth and is generalizing to static objects also, since it does not have the motion information as input signal. Points are thus colored by laser intensity. Bottom right: Ground truth, for reference." + ], + "image_footnote": [], + "bbox": [ + 220, + 252, + 500, + 359 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5b50df209bfd86b02f7e70be74b96d432d91fc41491b6203b968d88069d886be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 252, + 781, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Flow Clustering: The points $\\mathbf{p}_i \\in \\mathcal{P}^t$ in each preprocessed point cloud are clustered based on geometry and motion: All stationary points in a scene should have a flow $\\mathbf{f}_i$ similar to the vehicle's ego-motion, i.e. $\\mathbf{f}_i \\approx \\mathbf{f}_{i,\\mathrm{sta}} = ((\\mathbf{T}_{\\mathrm{ego}}^{t \\rightarrow t + 1})^{-1} - \\mathbf{I}_4) \\cdot \\mathbf{p}_i$ . The residual flow must then be caused by motion of other actors: $\\mathbf{f}_{i,\\mathrm{dyn}} = \\mathbf{f}_i - \\mathbf{f}_{i,\\mathrm{sta}}$ . We filter all static points by applying a threshold of $1\\mathrm{m/s}$ to the residual flow and cluster the remaining points based on their point location and flow vector in 6D using DBSCAN [9] (with parameters $\\varepsilon = 1.0$ , $\\min\\mathrm{Pts} = 5$ ).", + "bbox": [ + 212, + 515, + 787, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We fit a 3D bounding box $\\mathrm{B}_j^t$ to each resulting cluster following [37] and discard boxes with $l / w > 4.0$ , $lw < 0.35\\mathrm{m}^2$ , or $lwh < 0.5\\mathrm{m}^3$ . The heading $\\theta$ is set to match the \"forward-axis\" of the motion, i.e. we orient the boxes along the direction of the residual flow $\\mathbf{f}_{i,\\mathrm{dyn}}$ . The confidence $c$ is initialized to 1 during clustering, and later replaced by detectors' confidence $cf$ . Section 3.3.", + "bbox": [ + 212, + 637, + 787, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tracking: We run a simple flow based tracker: Since we have accurate egomotion available, we can track accurately in a fixed coordinate system w.r.t. the world. First, using the residual flow, we can compute a rigid body transform for each box at $t$ that transforms the proposed box $\\mathbf{B}_i^t$ forward in time towards its suspected location at $t + 1$ , $\\hat{\\mathbf{B}}_i^{t + 1}$ . The propagated boxes $\\hat{\\mathcal{B}}^t + 1$ are matched greedily against the actual boxes $\\mathcal{B}^{t + 1}$ based on box-center distance to the new detections. Unmatched boxes in $\\mathcal{B}^{t + 1}$ which are further away than $1.5\\mathrm{m}$ from propagated boxes spawn new tracklets. Unmatched tracklets from $\\mathcal{B}^t$ are prop", + "bbox": [ + 212, + 715, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Baur et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "agated according to the last observed box motion for up to one time step, after that unmatched tracklets are terminated. We run tracking forward and reverse in time, connecting tracklets from forward and reverse tracking to tracks.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The resulting set of tracks is post-filtered: We discard tracks that are shorter than 4 time steps or that have a median box confidence $c$ lower than the threshold 0.3. This retains only stable and high-confident tracks and avoids false positives to enter the pseudo ground truth.", + "bbox": [ + 212, + 191, + 784, + 251 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Track Optimization: We reduce positional noise of the tracks by minimizing translational jerk on all tracks longer than $3\\mathrm{m}$ . Let $\\mathcal{X}_{\\mathrm{obs}}$ be the sequence of (noisy) observed box center positions $\\mathbf{x}_i$ for consecutive time steps $i\\in \\{1,\\dots,T\\}$ of a track and their derivative $\\frac{d\\mathbf{x}_i}{dt}\\approx \\frac{\\mathbf{x}_{i + 1} - \\mathbf{x}_i}{\\Delta t}$ . We compute smoothed track positions $\\mathcal{X}_{\\mathrm{smooth}}$ by initializing them to $\\mathcal{X}_{\\mathrm{obs}}$ and minimizing the following loss w.r.t. $\\mathbf{x}_{\\mathrm{smooth}}$ :", + "bbox": [ + 212, + 252, + 785, + 344 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL = \\sum_ {i = 1} ^ {T} \\left\\| \\frac {d ^ {3} \\mathbf {x} _ {i , \\text {s m o o t h}}}{d t ^ {3}} \\right\\| _ {2} ^ {2} + \\beta \\left\\| \\mathbf {x} _ {i, \\text {s m o o t h}} - \\mathbf {x} _ {i, \\text {o b s}} \\right\\| _ {2} ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 356, + 785, + 395 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "I.e. we minimize the jerk $\\frac{d^3\\mathbf{x}}{dt^3}$ and use a quadratic regularizer term $(\\beta = 3)$ on the positions. This simple optimization goal is better suited for our method than more sophisticated ones like an unrolled bicycle model, as it is both faster and also class-agnostic, i.e. more universal: Using a bicycle model is problematic when optimizing pedestrian tracks, since our method does not allow to distinguish different classes.", + "bbox": [ + 212, + 409, + 782, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We subsequently align the orientation $\\theta$ of each detection in a track to the direction of the smoothed track at its position. Box dimensions $\\{l, w, h\\}$ of all boxes in a track are adapted to the 90 percentile of observed box dimensions in a track. All hyperparameters related to clustering, tracking and track optimization have been tuned visually on two sample snippets from nuScenes [4].", + "bbox": [ + 212, + 501, + 784, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Trajectory-Regularized Self-Training", + "text_level": 1, + "bbox": [ + 214, + 601, + 565, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After having mined initial pseudo ground truth of moving objects with a high precision and low recall, we now aim at iteratively improving our pseudo ground truth by training and using an object detector so that the pseudo ground truth generalizes to movable objects. We achieve this by executing iterative trajectory-regularized self-training, which is composed of the following two steps:", + "bbox": [ + 212, + 628, + 782, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training: We train the target object detection network using the current pseudo ground truth in a supervised training setup. Any single-frame object detection network can be plugged into our pipeline. In our experiments we do not deviate from the basic training setup of our object detection networks. Like any SOTA object detection method, we apply standard augmentation techniques to a point cloud during network training: Random rotation, scaling $(\\pm 5\\%)$ , and random translation up to $5\\mathrm{m}$ around the origin. Furthermore, we randomly pick 1 to 15 objects from the pseudo ground truth database and insert a random subset of their points at random locations into the scene.", + "bbox": [ + 212, + 704, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "LISO: Lidar-only Self-Supervised 3D Object Detection", + "bbox": [ + 367, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pseudo Ground Truth Regeneration: We halt training after a set number of training steps $s$ ( $s = 30k$ in our experiments) and create new, improved pseudo ground truth by running the trained detector in inference mode over all sequences in the training dataset, using the detectors' confidence as box confidence $c$ . We regularize these detections by running the flow-based tracker and track smoothing exactly like we do for the initial pseudo ground truth generation. Every 2nd iteration (i.e. every 60k steps in the experiments), we discard the network weights after regenerating the pseudo ground truth.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 4 illustrates the effect of our self-training: We see that the pseudo ground truth improves with each re-generation and that dropping network weights allows the network to re-focus on the generalized pseudo ground truth. Our strict requirements towards boxes being added to the pseudo ground truth leads to lower recall compared to the network, which as a result scores higher overall.", + "bbox": [ + 212, + 267, + 787, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Two aspects are key to making our method perform well:", + "bbox": [ + 238, + 343, + 653, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Being restrictive when composing the pseudo ground truth (i.e. using plausible tracks of high-confidence network detections or clusters with significant flow only) avoids adding false positives into the pseudo ground truth and hence avoids that our network increasingly hallucinates with each iteration.", + "bbox": [ + 225, + 367, + 782, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Not using flow but only single-frame point clouds as input for the detector allows the detector to focus on appearance of objects solely.", + "bbox": [ + 223, + 428, + 782, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These allow our method to generalize from initially mined moving objects to movable objects in the scene.", + "bbox": [ + 215, + 467, + 787, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Evaluation", + "text_level": 1, + "bbox": [ + 215, + 520, + 356, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate our method on multiple datasets across multiple SOTA networks. SLIM [3], the self-supervised flow network we use, is trained and inferred as in the published version, but Bird's-Eye-View (BEV) range is extended from $70 \\times 70\\mathrm{m}$ , $640 \\times 640$ pixels to $120 \\times 120\\mathrm{m}$ and $920 \\times 920$ pixels.", + "bbox": [ + 212, + 551, + 787, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Datasets and Metrics", + "text_level": 1, + "bbox": [ + 215, + 633, + 439, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate our method on four different AD datasets. For a fair comparison, we compute metrics for movable objects by mapping all animate objects (Cars, Trucks, Trailers, Motorcycles, Cyclists, Pedestrians and other Vehicles) to a single category and discard all inanimate objects (Barrier, Traffic Cone,...) since our method does not predict any class attributes. In Table 2 and Table 3 we also give class-based results for completeness. Class labels for true positives are taken from ground truth. For false positives, the predicted class label is assigned randomly according to the label distribution in the dataset.", + "bbox": [ + 212, + 657, + 787, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Waymo Open Dataset (WOD): [24] is a large, geographically diverse dataset recorded with a proprietary high quality lidar. We evaluate using the protocol of [15, 19], using an area of whole $100\\mathrm{m} \\times 40\\mathrm{m}$ BEV grid around the ego vehicle, artificially cropping the predictions of our method down to this reduced area.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Baur et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2aa3199686d64cc933f1da60d195c9e087c79b9baeac1b3243da51ce6ca977e0.jpg", + "image_caption": [ + "(a) Effect of self-supervised self-learning" + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 500, + 276 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/cb4fb497bafe7662ab370a0bb8aa987f7ca827223f5cfb44096feb9b7233fc70.jpg", + "image_caption": [ + "(b) Evaluation of Self-Training Hyperparameters", + "Fig. 4: Left: Effect of self-supervised self-learning: At the beginning of the training, only moving objects are contained within the pseudo ground truth, which thus scores 0 on static objects. Thanks to the self-training, the network generalizes to movable objects and the score of the pseudo ground truth on static objects starts to increase with every regeneration at each \"X\" in the graph. The pseudo ground truth's performance is measured on a subset of the train set of WOD dataset since our tracking is looking ahead in time, which would be unfair w.r.t. other methods. The network (here: Centerpoint) is evaluated on a small subset of the validation set. Right: Evaluation of Self-Training Hyperparameters: Network (Centerpoint) performance over the course of a training on WOD. $s$ is the number of training steps between pseudo ground truth regenerations, $r$ is the number of regenerations after which network weights are dropped and re-initialized." + ], + "image_footnote": [], + "bbox": [ + 501, + 142, + 782, + 276 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "KITTI: [10, 11] is recorded using a Velodyne HDL64 lidar sensor, where some parts have been annotated with 3D object boxes and tracking information in the forward facing camera field of view. A large portion of the published data is only available in raw format without annotations, which our method is able to leverage, due to not requiring any ground truth annotations for training.", + "bbox": [ + 212, + 494, + 784, + 570 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Argoverse 2: AV2 [31] is recorded using two stacked Velodyne VLP32 lidar sensors and annotated with 3D object boxes. On AV2 and KITTI, we evaluate both 2D (in BEV space) and 3D box IoU at IoU thresholds of 0.3 and 0.5 in the area of $100 \\times 100\\mathrm{m}$ around the ego vehicle.", + "bbox": [ + 212, + 570, + 784, + 630 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "nuScenes: [4] is recorded using a Velodyne VLP32 lidar sensor. We evaluate using the nuScenes protocol.", + "bbox": [ + 212, + 631, + 784, + 660 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Networks", + "text_level": 1, + "bbox": [ + 214, + 681, + 339, + 695 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate our proposed method with two SOTA lidar object detection networks of different architecture: Centerpoint [36] and Transfusion-L [2]. We do not modify the published network architectures or losses and train both networks with a batch size of four.", + "bbox": [ + 212, + 703, + 784, + 762 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "CenterPoint (CP): [36] is based on 2D convolutions in BEV space. Object centers are represented as heatmap in the BEV, which are then reduced to a sensible amount of boxes using non-maximum-suppression.", + "bbox": [ + 212, + 763, + 784, + 808 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Transfusion-L (TF): [2] is a Transformer based architecture, which is applied after initial encoding of the LiDAR point cloud into a 2D BEV feature tensor.", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "LISO: Lidar-only Self-Supervised 3D Object Detection", + "bbox": [ + 367, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/21d57fce2a0e068601eea715db077e70138226f286444f7a3fed5f102b7c33f2.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1: Evaluation on AV2 [31] and nuScenes [4]: CP, TF: network architecture, in the first two rows trained supervised for comparison. †: Metrics as reported in [38]. SF: self-supervised lidar scene flow by SLIM, GF: ground truth lidar scene flow. Note that nuScenes uses a minimum precision and recall threshold of 0.1, and since the recall of GT flow clustering is lower than 0.1, all results are clipped away. For full nuScenes scores see supplementary material." + ], + "table_body": "
AV2nuScenes
BEV-IOU3D-IOUAP↑AOE↓
AP@0.3↑AP@0.5↑AP@0.3↑AP@0.5↑
Sup.
CP [36]0.7780.7780.7360.5170.4840.560
TF [2]0.7830.6540.7670.6010.6270.501
Unsupervised
DBSCAN [9]0.0540.0260.0200.0030.0080.171
DBSCAN(SF)0.0700.0260.0650.0150.0030.082
DBSCAN(GF)0.1050.0410.0880.0240.0001.000
RSF [6]0.0740.0260.0550.0140.0191.003
Self Train
Oyster [38] †0.3540.245----
Oyster-CP [38]0.3810.2660.1500.0020.0911.514
Oyster-TF [38]0.3400.1980.1820.0160.0931.564
LISO-CP0.4480.3350.3670.1880.1091.062
LISO-TF0.4170.2940.3170.1760.1340.938
", + "bbox": [ + 267, + 143, + 736, + 367 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3Baselines", + "text_level": 1, + "bbox": [ + 215, + 497, + 339, + 512 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Besides the obvious baseline of training the object detection networks supervised from scratch on ground truth from the dataset, we also compare to multiple strong unsupervised baselines. For details, see Sec. 2.", + "bbox": [ + 212, + 531, + 785, + 575 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "RSF: We include RSF [6] in our evaluation as strong representative of methods doing object distilling from motion cues. We ran experiments using the published code of the authors.", + "bbox": [ + 212, + 578, + 785, + 622 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "DBSCAN: This algorithm [9] clusters points in the point cloud with similar 3D locations into objects. We additionally evaluate extending the cluster space to 6D by either using SLIM scene flow (SF) or ground-truth scene flow (GT). Thus, the \"DBSCAN(SF)\" baseline essentially corresponds to the raw detections used for our initial pseudo ground truth generation. We use the implementation from [17] with parameter values 1.0 for epsilon and 5 for the minimum number of points for a valid cluster.", + "bbox": [ + 212, + 625, + 787, + 729 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "SeMoLi: [19] is the most recent baseline. As it showed to outperform [15] as self-supervised object detection method we include only published results of [19] in our evaluation.", + "bbox": [ + 212, + 732, + 785, + 776 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Oyster: We implemented [38] using PyTorch [16], and apply the proposed framework to the CenterPoint (Oyster-CP) and Transfusion-L (Oyster-TF) to be as close as possible to our method and evaluation. We verify our reimplementation, by reproducing the reported metrics by [38] on AV2, see Table 1.", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Baur et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/b10176f0f853634da26cb8a984a5863514efc834bc28741f62934a47c391cc48.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AP@0.4 ↑AP@0.5 ↑AP@0.7 ↑
MovableMovingStillVehi.Ped.Cycl.MovableMovable
BEV3DBEV3DBEV3DBEVBEVBEVBEVBEV3DBEV3D
Sup.CP [36]0.7650.6840.7210.6240.7350.6570.9120.5130.1340.7190.5780.5100.202
TF [2]0.7460.7230.7140.6680.7330.7100.9180.4570.2160.7110.6690.5320.385
UnsupervisedDBSCAN [9]0.0270.0080.0090.0000.0270.0060.1840.0020.0010.0190.0020.0000.000
DBSCAN(SF)0.0260.0100.0640.0410.0000.0000.0730.0100.0090.0110.0010.0000.000
DBSCAN(GF)0.1140.0710.3180.1200.0000.0000.1130.1110.2400.0520.0230.0020.000
RSF [6]0.0300.0200.0800.0550.0000.0000.1090.0000.0020.0250.0120.0110.000
SeMoLi [19] †-0.195-0.575--------0.018
LISO-CP0.2920.2110.2720.2040.2080.1400.6070.0290.0100.2390.1460.1100.019
Self TrainOyster-CP [38]0.2170.0840.1510.0620.1760.0560.5620.0000.0000.1760.0230.0670.000
Oyster-TF [38]0.1210.0150.0510.0070.0980.0100.4750.0000.0000.0920.0030.0310.000
LISO-CP0.3800.3080.3500.2960.3220.2550.6950.0550.0220.3430.2190.1820.021
LISO-TF0.3270.2080.3490.2450.2330.1260.6690.0240.0120.2770.1360.1160.009
", + "bbox": [ + 215, + 143, + 785, + 343 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 2: Evaluation on WOD dataset: We evaluate using the protocol of [15, 19], using an area of whole $100\\mathrm{m}\\times 40\\mathrm{m}$ BEV grid around the ego vehicle, considering objects that move faster than $1\\mathrm{m/s}$ to be moving (difficulty level L2). CP, TF: network architecture, in the first two lines trained supervised for comparison. $\\dagger$ : Results taken from [19]. SF: lidar scene flow by SLIM, GF: ground truth lidar scene flow. For class-specific 3D detection scores see supplementary material.", + "bbox": [ + 212, + 343, + 787, + 428 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Results", + "text_level": 1, + "bbox": [ + 215, + 472, + 323, + 486 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative Results: On all four datasets we see that our method consistently outperforms all self-supervised baselines in all metrics for movable objects, see Table 1, Table 2, and Table 3. Only SeMoLi [19] beats LISO on moving objects (see Table 2) but significantly drops below LISO's performance on movable objects, hinting to difficulties with generalization. Our method does not suffer from this performance gap and performs nearly equally well on movable and moving objects. We also observe that, despite mostly better performance in the supervised case, Transfusion-L responds less favorable than Centerpoint to the self-supervised training approaches. We suspect that Centerpoint, due to its convolutional architecture and its lack of positional encoding compared to Transfusion-L, is less susceptible to overfitting to where moving objects have been observed and where not to expect them.", + "bbox": [ + 212, + 503, + 787, + 684 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We also observe that the nuScenes dataset seems to be the most challenging for the self-supervised methods, leading to the biggest gap between supervised and unsupervised object detection performance. We attribute this to the sparseness of the lidar sensor with only 32 layers. All self-supervised methods have difficulties estimating the orientation of objects (AOE on nuScenes, $2\\pi$ periodic, see Table 1), which the IoU-based metrics cannot reveal.", + "bbox": [ + 212, + 686, + 787, + 776 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative Results:", + "text_level": 1, + "bbox": [ + 215, + 777, + 383, + 792 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Fig. 5 we compare our method to the best-performing baseline, Oyster, using Centerpoint as network architecture. It can be seen, that our method has higher accuracy and predicts box orientations more accurately.", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "LISO: Lidar-only Self-Supervised 3D Object Detection", + "bbox": [ + 367, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/1154efe53c541cfa515a6b0a6f782e57b5ae296d62897ac637c33f7759bf0641.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Movable (Moving & Still)CarPedestrianCyclist
BEV-IOUAP@0.3↑3D-IOUAP@0.5↑AP@0.3↑AP@0.5↑BEV-IOUAP@0.3↑AP@0.5↑BEV-IOUAP@0.3↑AP@0.5↑BEV-IOUAP@0.3↑AP@0.5↑
CP [36]0.7550.6900.7360.6010.8140.7940.3700.1280.4090.157
TF [2]0.7470.6650.7290.5820.8200.7760.3110.0960.2630.032
DBSCAN [9]0.0230.0020.0100.0000.0260.0050.0000.0000.0640.007
RSF [6]0.0290.0190.0290.0110.0660.0490.0000.0000.1920.043
Oyster-CP [38]0.2350.0980.1140.0000.3270.1350.0000.0000.0000.000
Oyster-TF [38]0.2730.0880.1280.0000.3640.1210.0000.0000.0190.000
LISO-CP0.4460.3300.4190.1590.5200.4110.0970.0190.4450.053
LISO-TF0.3610.2070.2940.0360.4250.2970.0840.0140.3480.003
", + "bbox": [ + 215, + 143, + 782, + 285 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 3: Evaluation on KITTI dataset: We evaluate on the forward facing field of view where GT annotations are available, but run inference on the whole $100 \\times 100\\mathrm{m}$ BEV grid. Also note that flow annotations are not available for KITTI Object. CP, TF: network architecture, in the first two lines trained supervised for comparison.", + "bbox": [ + 215, + 286, + 785, + 340 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/933af890fd25eaf0af9a63097052fdb38d181555c5caab7463bbc734aecb6bf8.jpg", + "image_caption": [ + "Fig. 5: Qualitative Results on WOD. Red boxes are ground truth boxes, yellow boxes are predictions. Left: Oyster-CP Right: LISO-CP" + ], + "image_footnote": [], + "bbox": [ + 220, + 367, + 500, + 592 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/45ee8413268cef33df008a415851f006e812015a85c8c8752cb91ee27de6f2f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 367, + 782, + 590 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Fig. 6, in contrast, showcases an important failure-case and general problem of all unsupervised methods: ground truth, especially in the WOD dataset, is also annotated in very sparse and distant regions. It is very challenging to generate pseudo ground truth for such regions whereas supervised trainings do not have to generalize from moving to such movable (but very sparse) objects. This may be the main reason for a still noticeable gap in evaluation results between supervised and unsupervised methods.", + "bbox": [ + 212, + 657, + 787, + 765 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 785, + 390, + 801 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We investigated the influence of different choices for the components in our pipeline, see Table 4 and cf. Fig. 2.", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Baur et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0b029a796ed8a84455295746dea26f96686598721ca138d91a8e0a4ee81c26f2.jpg", + "image_caption": [ + "Fig. 6: Missed ground truth boxes in the WOD dataset." + ], + "image_footnote": [], + "bbox": [ + 370, + 143, + 633, + 292 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Motion Cues as Clustering Input: When adding motion cues to box clustering for Oyster we observe that detection performance already increases noticeably (comparing #3 to #4, Table 4). This demonstrates that self-supervised motion cues are a key ingredient to get high-quality initial pseudo ground truth as opposed to just using clustered 3D points.", + "bbox": [ + 212, + 345, + 787, + 422 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/248c3dcb24860bbab8605b2742807fc7488424cbad3160d034efd7226754dfbe.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TagClusterInputMethodSelf TrainTrackOptim.AP@0.4(BEV)AP@0.4(3D)
#1P, SFno tracker0.1770.086
#2P, SFno tracker✓0.2010.131
#3POyster✓0.2170.084
#4P, SFOyster✓0.2550.104
#5P, SFLISO(K, SF)0.2790.209
#6P, SFLISO(K, SF)✓0.3600.290
#7P, SFLISO(K, SF)✓0.2920.211
#8P, SFLISO(K, SF)✓✓0.3800.308
#9P, SFLISO(G, SF)✓✓0.4110.339
#10P, GFLISO(G, GF)✓✓0.4230.349
", + "bbox": [ + 215, + 446, + 787, + 619 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 4: Ablation study of different parts of our pipeline on WOD with Centerpoint [36]. Abbreviations: P: points, SF: self-supervised lidar scene flow (SLIM), GF: ground truth lidar scene flow. K: KISS-ICP ego-motion, G: ground truth ego-motion.", + "bbox": [ + 212, + 619, + 787, + 662 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Motion Cues for Tracking: Our scene-flow-based tracker is the biggest contributing factor to the success of our method: the ablation reveals that leveraging motion cues greatly improves tracking (comparing #4 to #6) and thus ultimately improves self-training, i.e. more accurate tracking allows for much more strictness when matching and filtering tracklets. The Oyster tracker uses a matching threshold of $5\\mathrm{m}$ , our threshold is only $1.5\\mathrm{m}$ . This strictness leads to higher quality pseudo ground truth.", + "bbox": [ + 212, + 703, + 787, + 809 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of Self Training: Even without using any tracker to filter tracklets self training has a benefit (comparing #1 to #2). This is due to the network", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "LISO: Lidar-only Self-Supervised 3D Object Detection", + "bbox": [ + 367, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "generalizing to new instances and confidence thresholding the detections. However, having an additional way to discard implausible detections (by tracking) amplifies the positive effect of self training (comparing #5 to #6 and #7 to #8), as it can better prevent false positives from entering the pseudo ground truth.", + "bbox": [ + 212, + 146, + 785, + 205 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Track Optimization: The added benefit of track optimization is more independent of Self Training (comparing going from \\#5 to \\#7 with going from \\#6 to \\#8). Finally we notice, that with the combination of a strict tracker, track optimization and self training, the performance is very close to using ground truth flow and ego-motion (comparing \\#8 to \\#9 and \\#10).", + "bbox": [ + 212, + 207, + 785, + 281 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Hyperparameters for Iterative Training: Fig. 4 demonstrates that the performance of our proposed method is relatively consistent across a range of hyperparameters. However, using too few steps between regeneration of pseudo ground truth ( $s = 20k$ ) leads to degrading performance, because the network does not have enough time to generalize and stabilize using the pseudo ground truth, which then has negative effects during regeneration of the pseudo ground truth. The experiment reveals that doing multiple iterations of self-learning and dropping network weights to allow the network to adjust is beneficial for a good performance. We selected $r = 2$ and $s = 30k$ as a good compromise between performance and speed for all other experiments.", + "bbox": [ + 212, + 282, + 787, + 434 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 455, + 359, + 472 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have proposed a novel framework for self-supervised 3D lidar object detection based on self-supervised lidar scene flow. Using simple clustering on lidar scene flow in combination with a novel flow-based tracker, we were able to generate pseudo ground truth with moving objects at high precision. This enabled us to bootstrap SOTA 3D lidar object detectors, without using any human labels. The detector generalizes from detecting moving to detecting movable objects over multiple self-training iterations. Experiments revealed that our trajectory-regularized self-learning, based on our scene-flow-based tracker, is key to the success of our method. We have demonstrated the effectiveness of our approach on two SOTA lidar object detectors as well as four AD datasets. All experiments were conducted with the same set of hyperparameters, demonstrating the robustness and wide applicability of our method. Our method achieves a significant improvement in the state-of-the-art of self-supervised lidar object detection on all four datasets. The biggest limitation of our approach is that our method is not able to distinguish different classes. Future work could be around generating pseudo ground truth for class labels, e.g. based on motion or size characteristics.", + "bbox": [ + 212, + 488, + 787, + 729 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 753, + 323, + 768 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Amini, M., Feofanov, V., Pauletto, L., Devijver, E., Maximov, Y.: Self-training: A survey. CoRR abs/2202.12040 (2022), https://arxiv.org/abs/2202.12040", + "2. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In:" + ], + "bbox": [ + 222, + 782, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Baur et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 1080-1089. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.00116, https://doi.org/10.1109/CVPR52688.2022.00116", + "3. Baur, S.A., Emmerichs, D.J., Moosmann, F., Pinggera, P., Ommer, B., Geiger, A.: SLIM: self-supervised lidar scene flow and motion segmentation. In: 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021. pp. 13106-13116. IEEE (2021). https://doi.org/10.1109/ICCV48922.2021.01288, https://doi.org/10.1109/ICCV48922.2021.01288", + "4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. pp. 11618-11628. IEEE (2020). https://doi.org/10.1109/CVPR42600.2020.01164, https://doi.org/10.1109/CVPR42600.2020.01164", + "5. Chen, Z., Luo, Y., Wang, Z., Baktashmotlagh, M., Huang, Z.: Revisiting domain-adaptive 3d object detection by reliable, diverse and class-balanced pseudolabeling. In: IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023. pp. 3691-3703. IEEE (2023). https://doi.org/10.1109/ICCV51070.2023.00344, https://doi.org/10.1109/ICCV51070.2023.00344", + "6. Deng, D., Zakhor, A.: Rsf: Optimizing rigid scene flow from 3d point clouds without labels. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1277-1286 (January 2023)", + "7. Dewan, A., Caselitz, T., Tipaldi, G.D., Burgard, W.: Motion-based detection and tracking in 3d lidar scans. In: Kragic, D., Bicchi, A., Luca, A.D. (eds.) 2016 IEEE International Conference on Robotics and Automation, ICRA 2016, Stockholm, Sweden, May 16-21, 2016. pp. 4508-4513. IEEE (2016). https://doi.org/10.1109/ICRA.2016.7487649, https://doi.org/10.1109/ICRA.2016.7487649", + "8. Eskandar, G.: An empirical study of the generalization ability of lidar 3d object detectors to unseen domains. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23815-23825 (June 2024)", + "9. Ester, M., Kriegel, H., Sander, J., Xu, X.: A density-based algorithm for discovering clusters in large spatial databases with noise. In: Simoudis, E., Han, J., Fayyad, U.M. (eds.) Proceedings of the Second International Conference on Knowledge Discovery and Data Mining (KDD-96), Portland, Oregon, USA. pp. 226-231. AAAI Press (1996), http://www.aaai.org/Library/KDD/1996/kdd96-037.php", + "10. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The KITTI dataset. Int. J. Robotics Res. 32(11), 1231-1237 (2013). https://doi.org/10.1177/0278364913491297, https://doi.org/10.1177/0278364913491297", + "11. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the KITTI vision benchmark suite. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition, Providence, RI, USA, June 16-21, 2012. pp. 3354-3361. IEEE Computer Society (2012). https://doi.org/10.1109/CVPR.2012.6248074, https://doi.org/10.1109/CVPR.2012.6248074", + "12. Harley, A.W., Zuo, Y., Wen, J., Mangal, A., Potdar, S., Chaudhry, R., Fragkiadaki, K.: Track, check, repeat: An EM approach to unsupervised tracking. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 16581-16591. Computer Vision Foundation / IEEE (2021). https:" + ], + "bbox": [ + 217, + 146, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "LISO: Lidar-only Self-Supervised 3D Object Detection", + "bbox": [ + 367, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "//doi.org/10.1109/CVPR46437.2021.01631, https://openaccess.thecvf.com/content/CVPR2021/html/Harley_Track_Check_Repeat_An_EM_Approach_to_Unsupervised_Tracking_CVPR_2021_paper.html", + "13. Liang, Z., Zhang, Z., Zhang, M., Zhao, X., Pu, S.: Rangeioudet: Range image based real-time 3d object detector optimized by intersection over union. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 7140-7149. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.00706, https://openaccess.thecvf.com/content/CVPR2021/html/Liang_RangeIoUDet_Range_Issue-Based.Real-Time_3D_Object_Detector_Optimized_by_CVPR_2021_paper.html", + "14. Meyer, G.P., Laddha, A., Kee, E., Vallespi-Gonzalez, C., Wellington, C.K.: Lasernet: An efficient probabilistic 3d object detector for autonomous driving. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019. pp. 12677-12686. Computer Vision Foundation / IEEE (2019). https://doi.org/10.1109/CVPR.2019.01296, http://openaccess.thecvf.com/content_CVPR_2019/html/Meyer_LaserNet_An_Efficient_Probabilistic_3D_Object_Detector_for_Autonomous_Driving_CVPR_2019_paper.html", + "15. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Motion inspired unsupervised perception and prediction in autonomous driving. In: Avidan, S., Brostow, G.J., Cissé, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVIII. Lecture Notes in Computer Science, vol. 13698, pp. 424-443. Springer (2022). https://doi.org/10.1007/978-3-031-19839-7_25, https://doi.org/10.1007/978-3-031-19839-7_25", + "16. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Köpf, A., Yang, E.Z., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Wallach, H.M., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E.B., Garnett, R. (eds.) Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada. pp. 8024-8035 (2019), https://proceedings.neurips.cc/paper/2019/hash/bdbca288fee7f92f2bfa9f7012727740-AAbstract.html", + "17. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine learning in Python. Journal of Machine Learning Research 12, 2825–2830 (2011)", + "18. Rist, C.B., Enzweiler, M., Gavrila, D.M.: Cross-sensor deep domain adaptation for lidar detection and segmentation. In: 2019 IEEE Intelligent Vehicles Symposium, IV 2019, Paris, France, June 9-12, 2019. pp. 1535-1542. IEEE (2019). https://doi.org/10.1109/IVS.2019.8814047, https://doi.org/10.1109/IVS.2019.8814047", + "19. Seidenschwarz, J., Osep, A., Ferroni, F., Lucey, S., Leal-Taixe, L.: Semoli: What moves together belongs together (2024), https://arxiv.org/abs/2402.19463", + "20. Shen, Z., Liang, H., Lin, L., Wang, Z., Huang, W., Yu, J.: Fast ground segmentation for 3d lidar point cloud based on jump-convolution-process. Remote. Sens. 13(16), 3239 (2021). https://doi.org/10.3390/rs13163239, https://doi.org/10.3390/rs13163239" + ], + "bbox": [ + 217, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Baur et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "21. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019. pp. 770-779. Computer Vision Foundation / IEEE (2019). https://doi.org/10.1109/CVPR.2019.00086, http://openaccess.thecvf.com/content_CVPR_2019/html/Shi_PointRCNN_3D Object_Proposal_Generation_and_Detection_From_Point_Cloud_CVPR_2019_paper.html", + "22. Shin, S., Golodetz, S., Vankadari, M., Zhou, K., Markham, A., Trigoni, N.: Sample, crop, track: Self-supervised mobile 3d object detection for urban driving lidar. CoRR abs/2209.10471 (2022). https://doi.org/10.48550/arXiv.2209.10471, https://doi.org/10.48550/arXiv.2209.10471", + "23. Song, Z., Yang, B.: OGC: Unsupervised 3d object segmentation from rigid dynamics of point clouds. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=ecNbE00tqBU", + "24. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. pp. 2443-2451. IEEE (2020). https://doi.org/10.1109/CVPR42600.2020.00252, https://doi.org/10.1109/CVPR42600.2020.00252", + "25. Sun, P., Wang, W., Chai, Y., Elsayed, G., Bewley, A., Zhang, X., Sminchisescu, C., Anguelov, D.: RSN: range sparse net for efficient, accurate lidar 3d object detection. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 5725-5734. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.00567, https://openaccess.thecvf.com/content/CVPR2021/html/Sun_RSN_Range_Sparse_Net_for_Efficient_Accurate_LiDAR_3D_Object_CVPR_2021_paper.html", + "26. Théodose, R., Denis, D., Chateau, T., Frémont, V., Checchin, P.: A deep learning approach for lidar resolution-agnostic object detection. IEEE Trans. Intell. Transp. Syst. 23(9), 14582-14593 (2022). https://doi.org/10.1109/TITS.2021.3130487, https://doi.org/10.1109/TITS.2021.3130487", + "27. Trauble, B., Pauen, S., Poulin-Dubois, D.: Speed and direction changes induce the perception of animacy in 7-month-old infants. Frontiers in Psychology 5 (2014). https://doi.org/10.3389/fpsyg.2014.01141, https://www.frontiersin.org/articles/10.3389/fpsyg.2014.01141", + "28. Vizzo, I., Guadagnino, T., Mersch, B., Wiesmann, L., Behley, J., Stachniss, C.: KISS-ICP: in defense of point-to-point ICP - simple, accurate, and robust registration if done the right way. IEEE Robotics Autom. Lett. 8(2), 1029–1036 (2023). https://doi.org/10.1109/LRA.2023.3236571, https://doi.org/10.1109/LRA.2023.3236571", + "29. Wang, Y., Chen, X., You, Y., Li, L.E., Hariharan, B., Campbell, M.E., Weinberger, K.Q., Chao, W.: Train in germany, test in the USA: making 3d object detectors generalize. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. pp. 11710-11720. Computer Vision Foundation / IEEE (2020). https://doi.org/10.1109/CVPR42600.2020.01173, https://openaccess.thecvf.com/content_CVPR_2020/html/Wang_" + ], + "bbox": [ + 215, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "LISO: Lidar-only Self-Supervised 3D Object Detection", + "bbox": [ + 367, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Train_in_Germany_Test_in_the_USA_Making_3D_Object_CVPR_2020_paper.html", + "30. Wang, Y., Chen, Y., Zhang, Z.: 4d unsupervised object discovery. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022 (2022), http://papers.nips.cc/paper_files/paper/2022/ hash/e7407ab5e89c405d28ff6807ffec594a-Abstraction-Conference.html", + "31. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., Ramanan, D., Carr, P., Hays, J.: Argov-verse 2: Next generation datasets for self-driving perception and forecasting. In: Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021) (2021)", + "32. Wozniak, M.K., Hansson, M., Thiel, M., Jensfelt, P.: UADA3D: unsupervised adversarial domain adaptation for 3d object detection with sparse lidar and large domain gaps. CoRR abs/2403.17633 (2024). https://doi.org/10.48550/ARXIV.2403.17633, https://doi.org/10.48550/arXiv.2403.17633", + "33. Xu, J., Waslander, S.L.: Hypermodest: Self-supervised 3d object detection with confidence score filtering. CoRR abs/2304.14446 (2023). https://doi.org/10.48550/arXiv.2304.14446, https://doi.org/10.48550/arXiv.2304.14446", + "34. Xu, Q., Zhou, Y., Wang, W., Qi, C.R., Anguelov, D.: SPG: unsupervised domain adaptation for 3d object detection via semantic point generation. In: 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021. pp. 15426-15436. IEEE (2021). https://doi.org/10.1109/ICCV48922.2021.01516, https://doi.org/10.1109/ICCV48922.2021.01516", + "35. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3dssd: Point-based 3d single stage object detector. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. pp. 11037-11045. Computer Vision Foundation / IEEE (2020). https://doi.org/10.1109/CVPR42600.2020.01105, https://openaccess.thecvf.com/content_CVPR_2020/html/Yang_3DSSD_Point-Based_3D_Single_SStage_Object_Detector_CVPR_2020_paper.html", + "36. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 11784-11793. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.01161, https://openaccess.thecvf.com/content/CVPR2021/html/Yin_Center-Based_3D_Object_Detection_and_Tracking_CVPR_2021_paper.html", + "37. You, Y., Luo, K., Phoo, C.P., Chao, W., Sun, W., Hariharan, B., Campbell, M.E., Weinberger, K.Q.: Learning to detect mobile objects from lidar scans without labels. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 1120-1130. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.00120, https://doi.org/10.1109/CVPR52688.2022.00120", + "38. Zhang, L., Yang, A.J., Xiong, Y., Casas, S., Yang, B., Ren, M., Urtasun, R.: Towards unsupervised object detection from lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9317-9328 (June 2023)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "S. Baur et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Zhang, Y., Hu, Q., Xu, G., Ma, Y., Wan, J., Guo, Y.: Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 18931-18940. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.01838, https://doi.org/10.1109/CVPR52688.2022.01838", + "40. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018. pp. 4490-4499. IEEE Computer Society (2018). https://doi.org/10.1109/CVPR.2018.00472, http://openaccess.thecvf.com/content_cvpr_2018/html/Zhou_VoxelNet_End-to-End_Learning_CVPR_2018_paper.html" + ], + "bbox": [ + 215, + 146, + 785, + 314 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "LISO: Lidar-only Self-Supervised 3D Object Detection", + "bbox": [ + 367, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/LISO_ Lidar-only Self-Supervised 3D Object Detection/d6e4374b-831f-4119-9bbf-2ac849c4dff4_model.json b/2024/LISO_ Lidar-only Self-Supervised 3D Object Detection/d6e4374b-831f-4119-9bbf-2ac849c4dff4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..272e180320dfaac2154835db9b2d831fb1237959 --- /dev/null +++ b/2024/LISO_ Lidar-only Self-Supervised 3D Object Detection/d6e4374b-831f-4119-9bbf-2ac849c4dff4_model.json @@ -0,0 +1,2339 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.25, + 0.141, + 0.753, + 0.185 + ], + "angle": 0, + "content": "LISO: Lidar-only Self-Supervised 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.212, + 0.756, + 0.229 + ], + "angle": 0, + "content": "Stefan Andreas Baur\\(^{1,2}\\), Frank Moosmann\\(^{1}\\), and Andreas Geiger\\(^{2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.408, + 0.24, + 0.596, + 0.254 + ], + "angle": 0, + "content": "1 Mercedes-Benz, Germany" + }, + { + "type": "text", + "bbox": [ + 0.417, + 0.254, + 0.586, + 0.268 + ], + "angle": 0, + "content": "2 University of Tübingen" + }, + { + "type": "text", + "bbox": [ + 0.425, + 0.268, + 0.577, + 0.282 + ], + "angle": 0, + "content": "3 Tübingen AI Center" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.323, + 0.741, + 0.447 + ], + "angle": 0, + "content": "Abstract. 3D object detection is one of the most important components in any Self-Driving stack, but current state-of-the-art (SOTA) lidar object detectors require costly & slow manual annotation of 3D bounding boxes to perform well. Recently, several methods emerged to generate pseudo ground truth without human supervision, however, all of these methods have various drawbacks: Some methods require sensor rigs with full camera coverage and accurate calibration, partly supplemented by an auxiliary optical flow engine. Others require expensive high-precision localization to find objects that disappeared over multiple drives." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.448, + 0.741, + 0.545 + ], + "angle": 0, + "content": "We introduce a novel self-supervised method to train SOTA lidar object detection networks, requiring only unlabeled sequences of lidar point clouds. We call this trajectory-regularized self-training. It utilizes a SOTA self-supervised lidar scene flow network under the hood to generate, track, and iteratively refine pseudo ground truth. We demonstrate the effectiveness of our approach for multiple SOTA object detection networks across multiple real-world datasets. Code will be released1." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.559, + 0.644, + 0.573 + ], + "angle": 0, + "content": "Keywords: Self-Supervised \\(\\cdot\\) LiDAR \\(\\cdot\\) Object Detection" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.601, + 0.377, + 0.617 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.634, + 0.788, + 0.785 + ], + "angle": 0, + "content": "Human developmental research reveals that infants less than one year of age are able to categorize animate and inanimate objects based on observed motion cues and can generalize this categorization to previously unseen objects [27]. Yet, lidar object detectors with SOTA performance are trained using manually selected and categorized annotations. These annotations are very expensive to obtain and become outdated quickly: new lidar sensors are coming to market regularly, and trained SOTA object detectors are sensitive to changes of sensor characteristics, mounting position and other domain gaps [5,8,18,26,32,34], even varying object sizes across different countries [29], making existing annotations difficult to transfer, resulting in high re-labeling efforts with each change." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.786, + 0.787, + 0.817 + ], + "angle": 0, + "content": "In this paper, we aim at bridging this gap by distilling motion cues observed in self-supervised lidar scene flow into SOTA single-frame lidar object detectors." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.468, + 0.84 + ], + "angle": 0, + "content": "1 https://github.com/baurst/liso" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "S. Baur et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.499, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.145, + 0.786, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.317, + 0.788, + 0.346 + ], + "angle": 0, + "content": "Fig. 1: Objects predicted by our method using no manual annotations. Red boxes are ground truth boxes, yellow boxes are predicted by our network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.655 + ], + "angle": 0, + "content": "We introduce a novel and robust self-supervised, easy-to-use method, operating solely on lidar point cloud sequences for easy-to-obtain \\(360^{\\circ}\\) annotations. No human-annotated bounding boxes, cameras, costly high-precision GPS, or tedious sensor rig calibration is required, which could be especially useful in robotics research where data is often limited (even KITTI [11] only has forward facing cameras) or maintaining a multi-sensor-setup including its calibration is too cumbersome and expensive. Our method trains and runs a self-supervised lidar flow estimator [3] under the hood in order to create motion cues. We show in our experiments that these motion cues are a key factor to the success of our method. Based on the estimated lidar flow we bootstrap initial pseudo ground truth using simple clustering and track optimization. With these mined bounding boxes of moving objects we initialize a self-supervised, trajectory-regularized variant of self-training [1] (which is semi-supervised in its original form): We train a first version of a SOTA object detector, then iteratively re-generate and trajectory-regularize the pseudo ground truth and retrain the detector. Since the single-frame object detector has no concept of motion, it generalizes to detect any movable object on the way. Exemplary output of the trained detector, 3D boxes in single-frame point clouds, is depicted in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Our contribution is a novel self-supervised trajectory-regularized self-training framework for single-frame 3D object detection with the following properties:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.712, + 0.784, + 0.742 + ], + "angle": 0, + "content": "- It is based entirely on lidar, i.e. without the limitations of prior works: no cameras, no calibration, no high-precision GPS, no manual annotations." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.747, + 0.784, + 0.788 + ], + "angle": 0, + "content": "- It is agnostic w.r.t. the sensor model, mounting pose, and detectors architecture and works with the same set of hyper parameters: We demonstrate this across four different datasets and different SOTA detector networks." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.795, + 0.784, + 0.837 + ], + "angle": 0, + "content": "- It is able to generalize from moving objects (motion cues) to movable objects (final detection results) and significantly outperforms SOTA methods while being simpler." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.712, + 0.784, + 0.837 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.368, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LISO: Lidar-only Self-Supervised 3D Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "We show that using motion cues together with trajectory-regularized self-training is key to this success. The code of this approach will be published for easy use and comparison by other researchers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.785, + 0.238 + ], + "angle": 0, + "content": "The paper is organized as follows: Sec. 2 discusses related work before the proposed method is described in detail in Sec. 3. Extensive experiments in Sec. 4 show the performance of our method before we conclude in Sec. 5." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.261, + 0.388, + 0.277 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.294, + 0.597, + 0.309 + ], + "angle": 0, + "content": "2.1 Single Frame Lidar 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.32, + 0.785, + 0.471 + ], + "angle": 0, + "content": "Object detectors operating on 3D point clouds are an active research field. The currently best-performing ones are using deep neural networks trained via supervised learning and can be categorized by their internal representation: Some networks operate on points directly like PointRCNN [21], 3DSSD [35], and IASSD [39]. Others project the points either to a virtual range image [13,14,25] or into a voxel representation like VoxelNet [40], PointPillars, CenterPoint [36], and Transfusion-L [2]. However, all aforementioned methods require large human-annotated datasets in order to perform well and obtaining such annotations is very expensive. We address this problem in this paper, enabling training of SOTA object detectors using pseudo ground truth." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.494, + 0.557, + 0.51 + ], + "angle": 0, + "content": "2.2 Object Distilling from Motion Cues" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.785, + 0.67 + ], + "angle": 0, + "content": "Multiple approaches have been suggested which leverage motion cues from lidar frame pairs in order to detect moving objects. Using the assumption of local geometric constancy, they decompose dynamic scenes into separate moving entities by applying as-rigid-as-possible optimization. Examples of these are the works by Dewan et al. [7], RSF [6] and OGC [23]. The two former methods are optimization-based whereas OGC uses these constraints as loss-function to self-supervise a segmentation network. Although these methods avoid the need for expensive labels, in contrast to our method they tend not to work well in low-resolution areas, they are typically slow, and can only detect moving but not movable (i.e. static but potentially moving) objects." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.693, + 0.621, + 0.709 + ], + "angle": 0, + "content": "2.3 Pseudo Ground Truth for Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Different approaches have been proposed to mine pseudo ground truth for training object detectors: Najibi et al. [15] and Seidenschwarz et al. [19] use motion cues similar to section 2.2 in order to distill moving objects as pseudo ground truth and use it to train an object detector. [15] runs optimization for each frame pair to obtain lidar scene flow, clusters points, fits boxes, tracks them using a Kalman Filter and finally refines them on ICP-registered point clouds. [19] optimizes a clustering algorithm on motion cues through message passing which they then apply to segment point clouds into a set of instances and subsequently" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "S. Baur et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.192 + ], + "angle": 0, + "content": "fit boxes. As shown in [19], and in contrast to our approach, both methods suffer from a large performance gap between moving and movable objects. We demonstrate that our method does not suffer from this gap." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.788, + 0.373 + ], + "angle": 0, + "content": "[12], [22] were the first to iteratively apply a detection, tracking, retraining paradigm to autonomous driving (AD) data. They build upon consistency constraints between object detectors which they train in the lidar and camera domain, making use of an optical flow network which is trained using supervision. The approach requires a calibrated sensor rig with possibly large camera coverage, IMU, and precision GPS. Similarly, [30] uses video sequences together with lidar scene flow to jointly train a camera and a lidar object detector. MODEST [33, 37] does not require coverage by calibrated cameras, but instead adds the additional requirement to having multiple lidar recordings of the same location in order to identify objects that vanished over time. With such demanding requirements, these approaches are not easy to use and are partially unsuited for popular AD datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.586 + ], + "angle": 0, + "content": "Oyster [38] is the approach most similar to ours. It uses DBSCAN [9] on lidar point clouds to initialize pseudo ground truth. Clusters are then tracked using forward-and-reverse tracking in sensor coordinates (i.e. without consideration for ego motion), using a complex policy for confidence-based track retention. After training an object detector on close range data, they employ zero-shot generalization to the far range data, track again and iteratively retrain the detector. In their experiments they use different hyperparameters for different datasets. Our method, in contrast, explicitly considers sensor motion, produces much cleaner initial proposals by leveraging a self-supervised lidar scene flow network, does not require zero-shot-generalization from near-to-far-field, and works with current SOTA object detectors. We show on multiple datasets that our method outperforms [38], that it works robustly with the same set of hyper-parameters and also generalizes well to detect movable objects. Additionally, we will release our code to facilitate further research on this topic." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.605, + 0.331, + 0.621 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.74 + ], + "angle": 0, + "content": "A general overview of our method is sketched in Fig. 2 and some steps are illustrated in Fig. 3. As input, we take raw (unlabeled) point cloud sequences and undergo three stages, all of which are detailed in the following: Preprocessing of point clouds and lidar scene flow computation (Sec. 3.1), initial pseudo ground truth generation (Sec. 3.2) and repeated training with pseudo ground truth refinement (Sec. 3.3). The final output of the method is a trained object detector which can detect movable objects in raw single-frame point clouds." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.758, + 0.379, + 0.774 + ], + "angle": 0, + "content": "3.1 Preprocessing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.597, + 0.794 + ], + "angle": 0, + "content": "We preprocess the raw input point clouds as follows:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Ground Removal: First, we remove distracting ground points from each single point cloud using JCP [20], which is a simple, robust, yet effective algorithm to remove ground points using changes in observed height above ground." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.368, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "LISO: Lidar-only Self-Supervised 3D Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.144, + 0.773, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.274, + 0.788, + 0.318 + ], + "angle": 0, + "content": "Fig. 2: Overview of the proposed method. Point cloud sequences are preprocessed (blue, Sec 3.1), initial pseudo ground truth is created (orange, Sec. 3.2) and the object detector is iteratively trained and pseudo ground truth regenerated (red, Sec. 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.784, + 0.407 + ], + "angle": 0, + "content": "Ego Motion Estimation: Second, we compute ego-motion between neighboring frame pairs using KISS-ICP [28], which is based on a robust version of ICP. The output is a cm-level accurate transformation \\(\\mathbf{T}_{\\mathrm{ego}}^{t\\rightarrow t + 1}\\in \\mathbb{R}^{4\\times 4}\\), describing the ego vehicle position at time \\(t + 1\\) represented in the ego frame from time \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.409, + 0.784, + 0.499 + ], + "angle": 0, + "content": "Lidar Scene Flow Estimation: Third, we compute lidar scene flow between neighboring frame pairs resulting in a flow vector \\(\\mathbf{f}_i = (dx, dy, dz)\\) for every point \\(i\\) in the first point cloud \\(\\mathcal{P}^t\\). We chose to use SLIM [3] as its code is readily available, it is easy to use, features fast inference, and produces SOTA results. The network is trained self-supervised on raw point cloud sequences, minimizing a k-nearest-neighbor loss between forward and time-reversed point clouds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.785, + 0.545 + ], + "angle": 0, + "content": "The components of our preprocessing steps (Ground Removal, Ego Motion Estimation, Lidar Scene Flow Estimation) have been selected for robustness and are all used with their default parameters from their respective publications." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.571, + 0.6, + 0.587 + ], + "angle": 0, + "content": "3.2 Initial Pseudo Ground Truth Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.785, + 0.661 + ], + "angle": 0, + "content": "The aim of our method is to mine pseudo ground truth for training a 3D lidar object detector. For a single point cloud \\(\\mathcal{P}^t\\) (consisting of \\(n \\in \\mathbb{N}\\) points \\(\\mathbf{p}_i\\), \\(\\mathbf{p}_i \\in \\mathcal{P}^t\\)) this ground truth is a set of 3D bounding boxes \\(\\mathcal{B}\\) with confidences representing the objects at time \\(t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.673, + 0.787, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} ^ {t} = \\left\\{\\mathrm {B} _ {j} ^ {t}, j \\in \\mathbb {N} \\right\\} = \\left\\{(x, y, z, l, w, h, \\theta , c) _ {j} \\right\\} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.784, + 0.748 + ], + "angle": 0, + "content": "Here, \\((x,y,z) = \\mathbf{x}\\) define the center position, \\((l,w,h)\\) the length, width and height, \\(\\theta\\) the heading (orientation around up axis), and \\(c\\) the confidence for a single box." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The key success factor of our method is to focus on a high precision (and potentially low recall) of the initial set of bounding boxes in order to avoid \"wrong\" objects to negatively influence the object detector. We achieve this by leveraging shape, density, and especially motion cues to robustly identify moving objects solely (see also top-right of Fig. 3). Sec. 3.3 targets at generalizing these to movable objects later on." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "S. Baur et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.501, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.782, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.253, + 0.501, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.253, + 0.782, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.484 + ], + "angle": 0, + "content": "Fig. 3: Overview over preprocessing, initial pseudo ground truth generation and training with examples. Top left: In the first step, self-supervised lidar scene flow is computed and corrected for vehicle ego-motion. Points are colored by flow direction and magnitude. Top right: In the second step, the scene flow is clustered and bounding boxes are fitted (to the moving objects). Bottom left: In the third step, the network is trained on the pseudo ground truth and is generalizing to static objects also, since it does not have the motion information as input signal. Points are thus colored by laser intensity. Bottom right: Ground truth, for reference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.637 + ], + "angle": 0, + "content": "Flow Clustering: The points \\(\\mathbf{p}_i \\in \\mathcal{P}^t\\) in each preprocessed point cloud are clustered based on geometry and motion: All stationary points in a scene should have a flow \\(\\mathbf{f}_i\\) similar to the vehicle's ego-motion, i.e. \\(\\mathbf{f}_i \\approx \\mathbf{f}_{i,\\mathrm{sta}} = ((\\mathbf{T}_{\\mathrm{ego}}^{t \\rightarrow t + 1})^{-1} - \\mathbf{I}_4) \\cdot \\mathbf{p}_i\\). The residual flow must then be caused by motion of other actors: \\(\\mathbf{f}_{i,\\mathrm{dyn}} = \\mathbf{f}_i - \\mathbf{f}_{i,\\mathrm{sta}}\\). We filter all static points by applying a threshold of \\(1\\mathrm{m/s}\\) to the residual flow and cluster the remaining points based on their point location and flow vector in 6D using DBSCAN [9] (with parameters \\(\\varepsilon = 1.0\\), \\(\\min\\mathrm{Pts} = 5\\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.638, + 0.788, + 0.715 + ], + "angle": 0, + "content": "We fit a 3D bounding box \\(\\mathrm{B}_j^t\\) to each resulting cluster following [37] and discard boxes with \\(l / w > 4.0\\), \\(lw < 0.35\\mathrm{m}^2\\), or \\(lwh < 0.5\\mathrm{m}^3\\). The heading \\(\\theta\\) is set to match the \"forward-axis\" of the motion, i.e. we orient the boxes along the direction of the residual flow \\(\\mathbf{f}_{i,\\mathrm{dyn}}\\). The confidence \\(c\\) is initialized to 1 during clustering, and later replaced by detectors' confidence \\(cf\\). Section 3.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.716, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Tracking: We run a simple flow based tracker: Since we have accurate egomotion available, we can track accurately in a fixed coordinate system w.r.t. the world. First, using the residual flow, we can compute a rigid body transform for each box at \\( t \\) that transforms the proposed box \\( \\mathbf{B}_i^t \\) forward in time towards its suspected location at \\( t + 1 \\), \\( \\hat{\\mathbf{B}}_i^{t + 1} \\). The propagated boxes \\( \\hat{\\mathcal{B}}^t + 1 \\) are matched greedily against the actual boxes \\( \\mathcal{B}^{t + 1} \\) based on box-center distance to the new detections. Unmatched boxes in \\( \\mathcal{B}^{t + 1} \\) which are further away than \\( 1.5\\mathrm{m} \\) from propagated boxes spawn new tracklets. Unmatched tracklets from \\( \\mathcal{B}^t \\) are prop" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.368, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "LISO: Lidar-only Self-Supervised 3D Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "agated according to the last observed box motion for up to one time step, after that unmatched tracklets are terminated. We run tracking forward and reverse in time, connecting tracklets from forward and reverse tracking to tracks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.252 + ], + "angle": 0, + "content": "The resulting set of tracks is post-filtered: We discard tracks that are shorter than 4 time steps or that have a median box confidence \\( c \\) lower than the threshold 0.3. This retains only stable and high-confident tracks and avoids false positives to enter the pseudo ground truth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.787, + 0.345 + ], + "angle": 0, + "content": "Track Optimization: We reduce positional noise of the tracks by minimizing translational jerk on all tracks longer than \\(3\\mathrm{m}\\). Let \\(\\mathcal{X}_{\\mathrm{obs}}\\) be the sequence of (noisy) observed box center positions \\(\\mathbf{x}_i\\) for consecutive time steps \\(i\\in \\{1,\\dots,T\\}\\) of a track and their derivative \\(\\frac{d\\mathbf{x}_i}{dt}\\approx \\frac{\\mathbf{x}_{i + 1} - \\mathbf{x}_i}{\\Delta t}\\). We compute smoothed track positions \\(\\mathcal{X}_{\\mathrm{smooth}}\\) by initializing them to \\(\\mathcal{X}_{\\mathrm{obs}}\\) and minimizing the following loss w.r.t. \\(\\mathbf{x}_{\\mathrm{smooth}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.357, + 0.786, + 0.396 + ], + "angle": 0, + "content": "\\[\nL = \\sum_ {i = 1} ^ {T} \\left\\| \\frac {d ^ {3} \\mathbf {x} _ {i , \\text {s m o o t h}}}{d t ^ {3}} \\right\\| _ {2} ^ {2} + \\beta \\left\\| \\mathbf {x} _ {i, \\text {s m o o t h}} - \\mathbf {x} _ {i, \\text {o b s}} \\right\\| _ {2} ^ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.784, + 0.501 + ], + "angle": 0, + "content": "I.e. we minimize the jerk \\(\\frac{d^3\\mathbf{x}}{dt^3}\\) and use a quadratic regularizer term \\((\\beta = 3)\\) on the positions. This simple optimization goal is better suited for our method than more sophisticated ones like an unrolled bicycle model, as it is both faster and also class-agnostic, i.e. more universal: Using a bicycle model is problematic when optimizing pedestrian tracks, since our method does not allow to distinguish different classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.785, + 0.579 + ], + "angle": 0, + "content": "We subsequently align the orientation \\(\\theta\\) of each detection in a track to the direction of the smoothed track at its position. Box dimensions \\(\\{l, w, h\\}\\) of all boxes in a track are adapted to the 90 percentile of observed box dimensions in a track. All hyperparameters related to clustering, tracking and track optimization have been tuned visually on two sample snippets from nuScenes [4]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.602, + 0.566, + 0.618 + ], + "angle": 0, + "content": "3.3 Trajectory-Regularized Self-Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.784, + 0.704 + ], + "angle": 0, + "content": "After having mined initial pseudo ground truth of moving objects with a high precision and low recall, we now aim at iteratively improving our pseudo ground truth by training and using an object detector so that the pseudo ground truth generalizes to movable objects. We achieve this by executing iterative trajectory-regularized self-training, which is composed of the following two steps:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Training: We train the target object detection network using the current pseudo ground truth in a supervised training setup. Any single-frame object detection network can be plugged into our pipeline. In our experiments we do not deviate from the basic training setup of our object detection networks. Like any SOTA object detection method, we apply standard augmentation techniques to a point cloud during network training: Random rotation, scaling \\((\\pm 5\\%)\\), and random translation up to \\(5\\mathrm{m}\\) around the origin. Furthermore, we randomly pick 1 to 15 objects from the pseudo ground truth database and insert a random subset of their points at random locations into the scene." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "S. Baur et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "Pseudo Ground Truth Regeneration: We halt training after a set number of training steps \\( s \\) (\\( s = 30k \\) in our experiments) and create new, improved pseudo ground truth by running the trained detector in inference mode over all sequences in the training dataset, using the detectors' confidence as box confidence \\( c \\). We regularize these detections by running the flow-based tracker and track smoothing exactly like we do for the initial pseudo ground truth generation. Every 2nd iteration (i.e. every 60k steps in the experiments), we discard the network weights after regenerating the pseudo ground truth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.344 + ], + "angle": 0, + "content": "Fig. 4 illustrates the effect of our self-training: We see that the pseudo ground truth improves with each re-generation and that dropping network weights allows the network to re-focus on the generalized pseudo ground truth. Our strict requirements towards boxes being added to the pseudo ground truth leads to lower recall compared to the network, which as a result scores higher overall." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.344, + 0.655, + 0.358 + ], + "angle": 0, + "content": "Two aspects are key to making our method perform well:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.368, + 0.784, + 0.426 + ], + "angle": 0, + "content": "- Being restrictive when composing the pseudo ground truth (i.e. using plausible tracks of high-confidence network detections or clusters with significant flow only) avoids adding false positives into the pseudo ground truth and hence avoids that our network increasingly hallucinates with each iteration." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.429, + 0.784, + 0.456 + ], + "angle": 0, + "content": "- Not using flow but only single-frame point clouds as input for the detector allows the detector to focus on appearance of objects solely." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.468, + 0.788, + 0.499 + ], + "angle": 0, + "content": "These allow our method to generalize from initially mined moving objects to movable objects in the scene." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.521, + 0.357, + 0.537 + ], + "angle": 0, + "content": "4 Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.613 + ], + "angle": 0, + "content": "We evaluate our method on multiple datasets across multiple SOTA networks. SLIM [3], the self-supervised flow network we use, is trained and inferred as in the published version, but Bird's-Eye-View (BEV) range is extended from \\(70 \\times 70\\mathrm{m}\\), \\(640 \\times 640\\) pixels to \\(120 \\times 120\\mathrm{m}\\) and \\(920 \\times 920\\) pixels." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.635, + 0.441, + 0.649 + ], + "angle": 0, + "content": "4.1 Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.78 + ], + "angle": 0, + "content": "We evaluate our method on four different AD datasets. For a fair comparison, we compute metrics for movable objects by mapping all animate objects (Cars, Trucks, Trailers, Motorcycles, Cyclists, Pedestrians and other Vehicles) to a single category and discard all inanimate objects (Barrier, Traffic Cone,...) since our method does not predict any class attributes. In Table 2 and Table 3 we also give class-based results for completeness. Class labels for true positives are taken from ground truth. For false positives, the predicted class label is assigned randomly according to the label distribution in the dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Waymo Open Dataset (WOD): [24] is a large, geographically diverse dataset recorded with a proprietary high quality lidar. We evaluate using the protocol of [15, 19], using an area of whole \\(100\\mathrm{m} \\times 40\\mathrm{m}\\) BEV grid around the ego vehicle, artificially cropping the predictions of our method down to this reduced area." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.368, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "LISO: Lidar-only Self-Supervised 3D Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.501, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.279, + 0.478, + 0.291 + ], + "angle": 0, + "content": "(a) Effect of self-supervised self-learning" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.143, + 0.783, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.279, + 0.785, + 0.291 + ], + "angle": 0, + "content": "(b) Evaluation of Self-Training Hyperparameters" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.302, + 0.785, + 0.468 + ], + "angle": 0, + "content": "Fig. 4: Left: Effect of self-supervised self-learning: At the beginning of the training, only moving objects are contained within the pseudo ground truth, which thus scores 0 on static objects. Thanks to the self-training, the network generalizes to movable objects and the score of the pseudo ground truth on static objects starts to increase with every regeneration at each \"X\" in the graph. The pseudo ground truth's performance is measured on a subset of the train set of WOD dataset since our tracking is looking ahead in time, which would be unfair w.r.t. other methods. The network (here: Centerpoint) is evaluated on a small subset of the validation set. Right: Evaluation of Self-Training Hyperparameters: Network (Centerpoint) performance over the course of a training on WOD. \\(s\\) is the number of training steps between pseudo ground truth regenerations, \\(r\\) is the number of regenerations after which network weights are dropped and re-initialized." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.785, + 0.571 + ], + "angle": 0, + "content": "KITTI: [10, 11] is recorded using a Velodyne HDL64 lidar sensor, where some parts have been annotated with 3D object boxes and tracking information in the forward facing camera field of view. A large portion of the published data is only available in raw format without annotations, which our method is able to leverage, due to not requiring any ground truth annotations for training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.571, + 0.785, + 0.631 + ], + "angle": 0, + "content": "Argoverse 2: AV2 [31] is recorded using two stacked Velodyne VLP32 lidar sensors and annotated with 3D object boxes. On AV2 and KITTI, we evaluate both 2D (in BEV space) and 3D box IoU at IoU thresholds of 0.3 and 0.5 in the area of \\(100 \\times 100\\mathrm{m}\\) around the ego vehicle." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.632, + 0.785, + 0.661 + ], + "angle": 0, + "content": "nuScenes: [4] is recorded using a Velodyne VLP32 lidar sensor. We evaluate using the nuScenes protocol." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.682, + 0.341, + 0.696 + ], + "angle": 0, + "content": "4.2 Networks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.785, + 0.763 + ], + "angle": 0, + "content": "We evaluate our proposed method with two SOTA lidar object detection networks of different architecture: Centerpoint [36] and Transfusion-L [2]. We do not modify the published network architectures or losses and train both networks with a batch size of four." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.809 + ], + "angle": 0, + "content": "CenterPoint (CP): [36] is based on 2D convolutions in BEV space. Object centers are represented as heatmap in the BEV, which are then reduced to a sensible amount of boxes using non-maximum-suppression." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Transfusion-L (TF): [2] is a Transformer based architecture, which is applied after initial encoding of the LiDAR point cloud into a 2D BEV feature tensor." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "S. Baur et al." + }, + { + "type": "table", + "bbox": [ + 0.269, + 0.144, + 0.737, + 0.368 + ], + "angle": 0, + "content": "
AV2nuScenes
BEV-IOU3D-IOUAP↑AOE↓
AP@0.3↑AP@0.5↑AP@0.3↑AP@0.5↑
Sup.
CP [36]0.7780.7780.7360.5170.4840.560
TF [2]0.7830.6540.7670.6010.6270.501
Unsupervised
DBSCAN [9]0.0540.0260.0200.0030.0080.171
DBSCAN(SF)0.0700.0260.0650.0150.0030.082
DBSCAN(GF)0.1050.0410.0880.0240.0001.000
RSF [6]0.0740.0260.0550.0140.0191.003
Self Train
Oyster [38] †0.3540.245----
Oyster-CP [38]0.3810.2660.1500.0020.0911.514
Oyster-TF [38]0.3400.1980.1820.0160.0931.564
LISO-CP0.4480.3350.3670.1880.1091.062
LISO-TF0.4170.2940.3170.1760.1340.938
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.214, + 0.369, + 0.788, + 0.454 + ], + "angle": 0, + "content": "Table 1: Evaluation on AV2 [31] and nuScenes [4]: CP, TF: network architecture, in the first two rows trained supervised for comparison. †: Metrics as reported in [38]. SF: self-supervised lidar scene flow by SLIM, GF: ground truth lidar scene flow. Note that nuScenes uses a minimum precision and recall threshold of 0.1, and since the recall of GT flow clustering is lower than 0.1, all results are clipped away. For full nuScenes scores see supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.498, + 0.341, + 0.513 + ], + "angle": 0, + "content": "4.3Baselines" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.532, + 0.787, + 0.577 + ], + "angle": 0, + "content": "Besides the obvious baseline of training the object detection networks supervised from scratch on ground truth from the dataset, we also compare to multiple strong unsupervised baselines. For details, see Sec. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.787, + 0.623 + ], + "angle": 0, + "content": "RSF: We include RSF [6] in our evaluation as strong representative of methods doing object distilling from motion cues. We ran experiments using the published code of the authors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.788, + 0.731 + ], + "angle": 0, + "content": "DBSCAN: This algorithm [9] clusters points in the point cloud with similar 3D locations into objects. We additionally evaluate extending the cluster space to 6D by either using SLIM scene flow (SF) or ground-truth scene flow (GT). Thus, the \"DBSCAN(SF)\" baseline essentially corresponds to the raw detections used for our initial pseudo ground truth generation. We use the implementation from [17] with parameter values 1.0 for epsilon and 5 for the minimum number of points for a valid cluster." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.733, + 0.787, + 0.777 + ], + "angle": 0, + "content": "SeMoLi: [19] is the most recent baseline. As it showed to outperform [15] as self-supervised object detection method we include only published results of [19] in our evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Oyster: We implemented [38] using PyTorch [16], and apply the proposed framework to the CenterPoint (Oyster-CP) and Transfusion-L (Oyster-TF) to be as close as possible to our method and evaluation. We verify our reimplementation, by reproducing the reported metrics by [38] on AV2, see Table 1." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.368, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LISO: Lidar-only Self-Supervised 3D Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.144, + 0.787, + 0.344 + ], + "angle": 0, + "content": "
AP@0.4 ↑AP@0.5 ↑AP@0.7 ↑
MovableMovingStillVehi.Ped.Cycl.MovableMovable
BEV3DBEV3DBEV3DBEVBEVBEVBEVBEV3DBEV3D
Sup.CP [36]0.7650.6840.7210.6240.7350.6570.9120.5130.1340.7190.5780.5100.202
TF [2]0.7460.7230.7140.6680.7330.7100.9180.4570.2160.7110.6690.5320.385
UnsupervisedDBSCAN [9]0.0270.0080.0090.0000.0270.0060.1840.0020.0010.0190.0020.0000.000
DBSCAN(SF)0.0260.0100.0640.0410.0000.0000.0730.0100.0090.0110.0010.0000.000
DBSCAN(GF)0.1140.0710.3180.1200.0000.0000.1130.1110.2400.0520.0230.0020.000
RSF [6]0.0300.0200.0800.0550.0000.0000.1090.0000.0020.0250.0120.0110.000
SeMoLi [19] †-0.195-0.575--------0.018
LISO-CP0.2920.2110.2720.2040.2080.1400.6070.0290.0100.2390.1460.1100.019
Self TrainOyster-CP [38]0.2170.0840.1510.0620.1760.0560.5620.0000.0000.1760.0230.0670.000
Oyster-TF [38]0.1210.0150.0510.0070.0980.0100.4750.0000.0000.0920.0030.0310.000
LISO-CP0.3800.3080.3500.2960.3220.2550.6950.0550.0220.3430.2190.1820.021
LISO-TF0.3270.2080.3490.2450.2330.1260.6690.0240.0120.2770.1360.1160.009
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.429 + ], + "angle": 0, + "content": "Table 2: Evaluation on WOD dataset: We evaluate using the protocol of [15, 19], using an area of whole \\(100\\mathrm{m}\\times 40\\mathrm{m}\\) BEV grid around the ego vehicle, considering objects that move faster than \\(1\\mathrm{m/s}\\) to be moving (difficulty level L2). CP, TF: network architecture, in the first two lines trained supervised for comparison. \\(\\dagger\\): Results taken from [19]. SF: lidar scene flow by SLIM, GF: ground truth lidar scene flow. For class-specific 3D detection scores see supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.473, + 0.325, + 0.487 + ], + "angle": 0, + "content": "4.4 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.788, + 0.685 + ], + "angle": 0, + "content": "Quantitative Results: On all four datasets we see that our method consistently outperforms all self-supervised baselines in all metrics for movable objects, see Table 1, Table 2, and Table 3. Only SeMoLi [19] beats LISO on moving objects (see Table 2) but significantly drops below LISO's performance on movable objects, hinting to difficulties with generalization. Our method does not suffer from this performance gap and performs nearly equally well on movable and moving objects. We also observe that, despite mostly better performance in the supervised case, Transfusion-L responds less favorable than Centerpoint to the self-supervised training approaches. We suspect that Centerpoint, due to its convolutional architecture and its lack of positional encoding compared to Transfusion-L, is less susceptible to overfitting to where moving objects have been observed and where not to expect them." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.777 + ], + "angle": 0, + "content": "We also observe that the nuScenes dataset seems to be the most challenging for the self-supervised methods, leading to the biggest gap between supervised and unsupervised object detection performance. We attribute this to the sparseness of the lidar sensor with only 32 layers. All self-supervised methods have difficulties estimating the orientation of objects (AOE on nuScenes, \\(2\\pi\\) periodic, see Table 1), which the IoU-based metrics cannot reveal." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.779, + 0.384, + 0.793 + ], + "angle": 0, + "content": "Qualitative Results:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In Fig. 5 we compare our method to the best-performing baseline, Oyster, using Centerpoint as network architecture. It can be seen, that our method has higher accuracy and predicts box orientations more accurately." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "S. Baur et al." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.144, + 0.784, + 0.286 + ], + "angle": 0, + "content": "
Movable (Moving & Still)CarPedestrianCyclist
BEV-IOUAP@0.3↑3D-IOUAP@0.5↑AP@0.3↑AP@0.5↑BEV-IOUAP@0.3↑AP@0.5↑BEV-IOUAP@0.3↑AP@0.5↑BEV-IOUAP@0.3↑AP@0.5↑
CP [36]0.7550.6900.7360.6010.8140.7940.3700.1280.4090.157
TF [2]0.7470.6650.7290.5820.8200.7760.3110.0960.2630.032
DBSCAN [9]0.0230.0020.0100.0000.0260.0050.0000.0000.0640.007
RSF [6]0.0290.0190.0290.0110.0660.0490.0000.0000.1920.043
Oyster-CP [38]0.2350.0980.1140.0000.3270.1350.0000.0000.0000.000
Oyster-TF [38]0.2730.0880.1280.0000.3640.1210.0000.0000.0190.000
LISO-CP0.4460.3300.4190.1590.5200.4110.0970.0190.4450.053
LISO-TF0.3610.2070.2940.0360.4250.2970.0840.0140.3480.003
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.287, + 0.787, + 0.342 + ], + "angle": 0, + "content": "Table 3: Evaluation on KITTI dataset: We evaluate on the forward facing field of view where GT annotations are available, but run inference on the whole \\(100 \\times 100\\mathrm{m}\\) BEV grid. Also note that flow annotations are not available for KITTI Object. CP, TF: network architecture, in the first two lines trained supervised for comparison." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.368, + 0.501, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.368, + 0.783, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.601, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Fig. 5: Qualitative Results on WOD. Red boxes are ground truth boxes, yellow boxes are predictions. Left: Oyster-CP Right: LISO-CP" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.766 + ], + "angle": 0, + "content": "Fig. 6, in contrast, showcases an important failure-case and general problem of all unsupervised methods: ground truth, especially in the WOD dataset, is also annotated in very sparse and distant regions. It is very challenging to generate pseudo ground truth for such regions whereas supervised trainings do not have to generalize from moving to such movable (but very sparse) objects. This may be the main reason for a still noticeable gap in evaluation results between supervised and unsupervised methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.786, + 0.391, + 0.803 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We investigated the influence of different choices for the components in our pipeline, see Table 4 and cf. Fig. 2." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.368, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LISO: Lidar-only Self-Supervised 3D Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.144, + 0.635, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.301, + 0.69, + 0.317 + ], + "angle": 0, + "content": "Fig. 6: Missed ground truth boxes in the WOD dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.424 + ], + "angle": 0, + "content": "Motion Cues as Clustering Input: When adding motion cues to box clustering for Oyster we observe that detection performance already increases noticeably (comparing #3 to #4, Table 4). This demonstrates that self-supervised motion cues are a key ingredient to get high-quality initial pseudo ground truth as opposed to just using clustered 3D points." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.447, + 0.789, + 0.62 + ], + "angle": 0, + "content": "
TagClusterInputMethodSelf TrainTrackOptim.AP@0.4(BEV)AP@0.4(3D)
#1P, SFno tracker0.1770.086
#2P, SFno tracker✓0.2010.131
#3POyster✓0.2170.084
#4P, SFOyster✓0.2550.104
#5P, SFLISO(K, SF)0.2790.209
#6P, SFLISO(K, SF)✓0.3600.290
#7P, SFLISO(K, SF)✓0.2920.211
#8P, SFLISO(K, SF)✓✓0.3800.308
#9P, SFLISO(G, SF)✓✓0.4110.339
#10P, GFLISO(G, GF)✓✓0.4230.349
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.62, + 0.788, + 0.663 + ], + "angle": 0, + "content": "Table 4: Ablation study of different parts of our pipeline on WOD with Centerpoint [36]. Abbreviations: P: points, SF: self-supervised lidar scene flow (SLIM), GF: ground truth lidar scene flow. K: KISS-ICP ego-motion, G: ground truth ego-motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Motion Cues for Tracking: Our scene-flow-based tracker is the biggest contributing factor to the success of our method: the ablation reveals that leveraging motion cues greatly improves tracking (comparing #4 to #6) and thus ultimately improves self-training, i.e. more accurate tracking allows for much more strictness when matching and filtering tracklets. The Oyster tracker uses a matching threshold of \\(5\\mathrm{m}\\), our threshold is only \\(1.5\\mathrm{m}\\). This strictness leads to higher quality pseudo ground truth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Effect of Self Training: Even without using any tracker to filter tracklets self training has a benefit (comparing #1 to #2). This is due to the network" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "S. Baur et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.786, + 0.207 + ], + "angle": 0, + "content": "generalizing to new instances and confidence thresholding the detections. However, having an additional way to discard implausible detections (by tracking) amplifies the positive effect of self training (comparing #5 to #6 and #7 to #8), as it can better prevent false positives from entering the pseudo ground truth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.787, + 0.282 + ], + "angle": 0, + "content": "Track Optimization: The added benefit of track optimization is more independent of Self Training (comparing going from \\#5 to \\#7 with going from \\#6 to \\#8). Finally we notice, that with the combination of a strict tracker, track optimization and self training, the performance is very close to using ground truth flow and ego-motion (comparing \\#8 to \\#9 and \\#10)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Hyperparameters for Iterative Training: Fig. 4 demonstrates that the performance of our proposed method is relatively consistent across a range of hyperparameters. However, using too few steps between regeneration of pseudo ground truth (\\(s = 20k\\)) leads to degrading performance, because the network does not have enough time to generalize and stabilize using the pseudo ground truth, which then has negative effects during regeneration of the pseudo ground truth. The experiment reveals that doing multiple iterations of self-learning and dropping network weights to allow the network to adjust is beneficial for a good performance. We selected \\(r = 2\\) and \\(s = 30k\\) as a good compromise between performance and speed for all other experiments." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.457, + 0.36, + 0.473 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.73 + ], + "angle": 0, + "content": "We have proposed a novel framework for self-supervised 3D lidar object detection based on self-supervised lidar scene flow. Using simple clustering on lidar scene flow in combination with a novel flow-based tracker, we were able to generate pseudo ground truth with moving objects at high precision. This enabled us to bootstrap SOTA 3D lidar object detectors, without using any human labels. The detector generalizes from detecting moving to detecting movable objects over multiple self-training iterations. Experiments revealed that our trajectory-regularized self-learning, based on our scene-flow-based tracker, is key to the success of our method. We have demonstrated the effectiveness of our approach on two SOTA lidar object detectors as well as four AD datasets. All experiments were conducted with the same set of hyperparameters, demonstrating the robustness and wide applicability of our method. Our method achieves a significant improvement in the state-of-the-art of self-supervised lidar object detection on all four datasets. The biggest limitation of our approach is that our method is not able to distinguish different classes. Future work could be around generating pseudo ground truth for class labels, e.g. based on motion or size characteristics." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.754, + 0.325, + 0.769 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.784, + 0.786, + 0.812 + ], + "angle": 0, + "content": "1. Amini, M., Feofanov, V., Pauletto, L., Devijver, E., Maximov, Y.: Self-training: A survey. CoRR abs/2202.12040 (2022), https://arxiv.org/abs/2202.12040" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "2. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In:" + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.784, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.368, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LISO: Lidar-only Self-Supervised 3D Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.786, + 0.202 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 1080-1089. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.00116, https://doi.org/10.1109/CVPR52688.2022.00116" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.204, + 0.787, + 0.285 + ], + "angle": 0, + "content": "3. Baur, S.A., Emmerichs, D.J., Moosmann, F., Pinggera, P., Ommer, B., Geiger, A.: SLIM: self-supervised lidar scene flow and motion segmentation. In: 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021. pp. 13106-13116. IEEE (2021). https://doi.org/10.1109/ICCV48922.2021.01288, https://doi.org/10.1109/ICCV48922.2021.01288" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.287, + 0.786, + 0.369 + ], + "angle": 0, + "content": "4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. pp. 11618-11628. IEEE (2020). https://doi.org/10.1109/CVPR42600.2020.01164, https://doi.org/10.1109/CVPR42600.2020.01164" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.787, + 0.45 + ], + "angle": 0, + "content": "5. Chen, Z., Luo, Y., Wang, Z., Baktashmotlagh, M., Huang, Z.: Revisiting domain-adaptive 3d object detection by reliable, diverse and class-balanced pseudolabeling. In: IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023. pp. 3691-3703. IEEE (2023). https://doi.org/10.1109/ICCV51070.2023.00344, https://doi.org/10.1109/ICCV51070.2023.00344" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.453, + 0.787, + 0.493 + ], + "angle": 0, + "content": "6. Deng, D., Zakhor, A.: Rsf: Optimizing rigid scene flow from 3d point clouds without labels. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1277-1286 (January 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.495, + 0.787, + 0.563 + ], + "angle": 0, + "content": "7. Dewan, A., Caselitz, T., Tipaldi, G.D., Burgard, W.: Motion-based detection and tracking in 3d lidar scans. In: Kragic, D., Bicchi, A., Luca, A.D. (eds.) 2016 IEEE International Conference on Robotics and Automation, ICRA 2016, Stockholm, Sweden, May 16-21, 2016. pp. 4508-4513. IEEE (2016). https://doi.org/10.1109/ICRA.2016.7487649, https://doi.org/10.1109/ICRA.2016.7487649" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.787, + 0.605 + ], + "angle": 0, + "content": "8. Eskandar, G.: An empirical study of the generalization ability of lidar 3d object detectors to unseen domains. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23815-23825 (June 2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.607, + 0.787, + 0.674 + ], + "angle": 0, + "content": "9. Ester, M., Kriegel, H., Sander, J., Xu, X.: A density-based algorithm for discovering clusters in large spatial databases with noise. In: Simoudis, E., Han, J., Fayyad, U.M. (eds.) Proceedings of the Second International Conference on Knowledge Discovery and Data Mining (KDD-96), Portland, Oregon, USA. pp. 226-231. AAAI Press (1996), http://www.aaai.org/Library/KDD/1996/kdd96-037.php" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.676, + 0.785, + 0.715 + ], + "angle": 0, + "content": "10. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The KITTI dataset. Int. J. Robotics Res. 32(11), 1231-1237 (2013). https://doi.org/10.1177/0278364913491297, https://doi.org/10.1177/0278364913491297" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.717, + 0.786, + 0.784 + ], + "angle": 0, + "content": "11. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the KITTI vision benchmark suite. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition, Providence, RI, USA, June 16-21, 2012. pp. 3354-3361. IEEE Computer Society (2012). https://doi.org/10.1109/CVPR.2012.6248074, https://doi.org/10.1109/CVPR.2012.6248074" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.786, + 0.786, + 0.84 + ], + "angle": 0, + "content": "12. Harley, A.W., Zuo, Y., Wen, J., Mangal, A., Potdar, S., Chaudhry, R., Fragkiadaki, K.: Track, check, repeat: An EM approach to unsupervised tracking. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 16581-16591. Computer Vision Foundation / IEEE (2021). https:" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "S. Baur et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "//doi.org/10.1109/CVPR46437.2021.01631, https://openaccess.thecvf.com/content/CVPR2021/html/Harley_Track_Check_Repeat_An_EM_Approach_to_Unsupervised_Tracking_CVPR_2021_paper.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.287 + ], + "angle": 0, + "content": "13. Liang, Z., Zhang, Z., Zhang, M., Zhao, X., Pu, S.: Rangeioudet: Range image based real-time 3d object detector optimized by intersection over union. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 7140-7149. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.00706, https://openaccess.thecvf.com/content/CVPR2021/html/Liang_RangeIoUDet_Range_Issue-Based.Real-Time_3D_Object_Detector_Optimized_by_CVPR_2021_paper.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.287, + 0.787, + 0.398 + ], + "angle": 0, + "content": "14. Meyer, G.P., Laddha, A., Kee, E., Vallespi-Gonzalez, C., Wellington, C.K.: Lasernet: An efficient probabilistic 3d object detector for autonomous driving. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019. pp. 12677-12686. Computer Vision Foundation / IEEE (2019). https://doi.org/10.1109/CVPR.2019.01296, http://openaccess.thecvf.com/content_CVPR_2019/html/Meyer_LaserNet_An_Efficient_Probabilistic_3D_Object_Detector_for_Autonomous_Driving_CVPR_2019_paper.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.398, + 0.787, + 0.495 + ], + "angle": 0, + "content": "15. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Motion inspired unsupervised perception and prediction in autonomous driving. In: Avidan, S., Brostow, G.J., Cissé, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVIII. Lecture Notes in Computer Science, vol. 13698, pp. 424-443. Springer (2022). https://doi.org/10.1007/978-3-031-19839-7_25, https://doi.org/10.1007/978-3-031-19839-7_25" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.495, + 0.787, + 0.632 + ], + "angle": 0, + "content": "16. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Köpf, A., Yang, E.Z., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Wallach, H.M., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E.B., Garnett, R. (eds.) Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada. pp. 8024-8035 (2019), https://proceedings.neurips.cc/paper/2019/hash/bdbca288fee7f92f2bfa9f7012727740-AAbstract.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.633, + 0.787, + 0.688 + ], + "angle": 0, + "content": "17. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine learning in Python. Journal of Machine Learning Research 12, 2825–2830 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.689, + 0.787, + 0.757 + ], + "angle": 0, + "content": "18. Rist, C.B., Enzweiler, M., Gavrila, D.M.: Cross-sensor deep domain adaptation for lidar detection and segmentation. In: 2019 IEEE Intelligent Vehicles Symposium, IV 2019, Paris, France, June 9-12, 2019. pp. 1535-1542. IEEE (2019). https://doi.org/10.1109/IVS.2019.8814047, https://doi.org/10.1109/IVS.2019.8814047" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.758, + 0.787, + 0.786 + ], + "angle": 0, + "content": "19. Seidenschwarz, J., Osep, A., Ferroni, F., Lucey, S., Leal-Taixe, L.: Semoli: What moves together belongs together (2024), https://arxiv.org/abs/2402.19463" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "20. Shen, Z., Liang, H., Lin, L., Wang, Z., Huang, W., Yu, J.: Fast ground segmentation for 3d lidar point cloud based on jump-convolution-process. Remote. Sens. 13(16), 3239 (2021). https://doi.org/10.3390/rs13163239, https://doi.org/10.3390/rs13163239" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.368, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LISO: Lidar-only Self-Supervised 3D Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.244 + ], + "angle": 0, + "content": "21. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019. pp. 770-779. Computer Vision Foundation / IEEE (2019). https://doi.org/10.1109/CVPR.2019.00086, http://openaccess.thecvf.com/content_CVPR_2019/html/Shi_PointRCNN_3D Object_Proposal_Generation_and_Detection_From_Point_Cloud_CVPR_2019_paper.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.787, + 0.302 + ], + "angle": 0, + "content": "22. Shin, S., Golodetz, S., Vankadari, M., Zhou, K., Markham, A., Trigoni, N.: Sample, crop, track: Self-supervised mobile 3d object detection for urban driving lidar. CoRR abs/2209.10471 (2022). https://doi.org/10.48550/arXiv.2209.10471, https://doi.org/10.48550/arXiv.2209.10471" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.787, + 0.359 + ], + "angle": 0, + "content": "23. Song, Z., Yang, B.: OGC: Unsupervised 3d object segmentation from rigid dynamics of point clouds. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=ecNbE00tqBU" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.787, + 0.471 + ], + "angle": 0, + "content": "24. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. pp. 2443-2451. IEEE (2020). https://doi.org/10.1109/CVPR42600.2020.00252, https://doi.org/10.1109/CVPR42600.2020.00252" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.57 + ], + "angle": 0, + "content": "25. Sun, P., Wang, W., Chai, Y., Elsayed, G., Bewley, A., Zhang, X., Sminchisescu, C., Anguelov, D.: RSN: range sparse net for efficient, accurate lidar 3d object detection. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 5725-5734. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.00567, https://openaccess.thecvf.com/content/CVPR2021/html/Sun_RSN_Range_Sparse_Net_for_Efficient_Accurate_LiDAR_3D_Object_CVPR_2021_paper.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.572, + 0.787, + 0.627 + ], + "angle": 0, + "content": "26. Théodose, R., Denis, D., Chateau, T., Frémont, V., Checchin, P.: A deep learning approach for lidar resolution-agnostic object detection. IEEE Trans. Intell. Transp. Syst. 23(9), 14582-14593 (2022). https://doi.org/10.1109/TITS.2021.3130487, https://doi.org/10.1109/TITS.2021.3130487" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.787, + 0.685 + ], + "angle": 0, + "content": "27. Trauble, B., Pauen, S., Poulin-Dubois, D.: Speed and direction changes induce the perception of animacy in 7-month-old infants. Frontiers in Psychology 5 (2014). https://doi.org/10.3389/fpsyg.2014.01141, https://www.frontiersin.org/articles/10.3389/fpsyg.2014.01141" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.687, + 0.787, + 0.754 + ], + "angle": 0, + "content": "28. Vizzo, I., Guadagnino, T., Mersch, B., Wiesmann, L., Behley, J., Stachniss, C.: KISS-ICP: in defense of point-to-point ICP - simple, accurate, and robust registration if done the right way. IEEE Robotics Autom. Lett. 8(2), 1029–1036 (2023). https://doi.org/10.1109/LRA.2023.3236571, https://doi.org/10.1109/LRA.2023.3236571" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.841 + ], + "angle": 0, + "content": "29. Wang, Y., Chen, X., You, Y., Li, L.E., Hariharan, B., Campbell, M.E., Weinberger, K.Q., Chao, W.: Train in germany, test in the USA: making 3d object detectors generalize. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. pp. 11710-11720. Computer Vision Foundation / IEEE (2020). https://doi.org/10.1109/CVPR42600.2020.01173, https://openaccess.thecvf.com/content_CVPR_2020/html/Wang_" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "S. Baur et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.242, + 0.148, + 0.783, + 0.174 + ], + "angle": 0, + "content": "Train_in_Germany_Test_in_the_USA_Making_3D_Object_CVPR_2020_paper.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.178, + 0.787, + 0.259 + ], + "angle": 0, + "content": "30. Wang, Y., Chen, Y., Zhang, Z.: 4d unsupervised object discovery. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022 (2022), http://papers.nips.cc/paper_files/paper/2022/ hash/e7407ab5e89c405d28ff6807ffec594a-Abstraction-Conference.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.787, + 0.331 + ], + "angle": 0, + "content": "31. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., Ramanan, D., Carr, P., Hays, J.: Argov-verse 2: Next generation datasets for self-driving perception and forecasting. In: Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.333, + 0.786, + 0.387 + ], + "angle": 0, + "content": "32. Wozniak, M.K., Hansson, M., Thiel, M., Jensfelt, P.: UADA3D: unsupervised adversarial domain adaptation for 3d object detection with sparse lidar and large domain gaps. CoRR abs/2403.17633 (2024). https://doi.org/10.48550/ARXIV.2403.17633, https://doi.org/10.48550/arXiv.2403.17633" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.39, + 0.786, + 0.431 + ], + "angle": 0, + "content": "33. Xu, J., Waslander, S.L.: Hypermodest: Self-supervised 3d object detection with confidence score filtering. CoRR abs/2304.14446 (2023). https://doi.org/10.48550/arXiv.2304.14446, https://doi.org/10.48550/arXiv.2304.14446" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.786, + 0.514 + ], + "angle": 0, + "content": "34. Xu, Q., Zhou, Y., Wang, W., Qi, C.R., Anguelov, D.: SPG: unsupervised domain adaptation for 3d object detection via semantic point generation. In: 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021. pp. 15426-15436. IEEE (2021). https://doi.org/10.1109/ICCV48922.2021.01516, https://doi.org/10.1109/ICCV48922.2021.01516" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.517, + 0.787, + 0.611 + ], + "angle": 0, + "content": "35. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3dssd: Point-based 3d single stage object detector. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. pp. 11037-11045. Computer Vision Foundation / IEEE (2020). https://doi.org/10.1109/CVPR42600.2020.01105, https://openaccess.thecvf.com/content_CVPR_2020/html/Yang_3DSSD_Point-Based_3D_Single_SStage_Object_Detector_CVPR_2020_paper.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.616, + 0.787, + 0.698 + ], + "angle": 0, + "content": "36. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 11784-11793. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.01161, https://openaccess.thecvf.com/content/CVPR2021/html/Yin_Center-Based_3D_Object_Detection_and_Tracking_CVPR_2021_paper.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.787, + 0.782 + ], + "angle": 0, + "content": "37. You, Y., Luo, K., Phoo, C.P., Chao, W., Sun, W., Hariharan, B., Campbell, M.E., Weinberger, K.Q.: Learning to detect mobile objects from lidar scans without labels. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 1120-1130. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.00120, https://doi.org/10.1109/CVPR52688.2022.00120" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "38. Zhang, L., Yang, A.J., Xiong, Y., Casas, S., Yang, B., Ren, M., Urtasun, R.: Towards unsupervised object detection from lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9317-9328 (June 2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.368, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LISO: Lidar-only Self-Supervised 3D Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.23 + ], + "angle": 0, + "content": "39. Zhang, Y., Hu, Q., Xu, G., Ma, Y., Wan, J., Guo, Y.: Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 18931-18940. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.01838, https://doi.org/10.1109/CVPR52688.2022.01838" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.787, + 0.315 + ], + "angle": 0, + "content": "40. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018. pp. 4490-4499. IEEE Computer Society (2018). https://doi.org/10.1109/CVPR.2018.00472, http://openaccess.thecvf.com/content_cvpr_2018/html/Zhou_VoxelNet_End-to-End_Learning_CVPR_2018_paper.html" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.315 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/LISO_ Lidar-only Self-Supervised 3D Object Detection/d6e4374b-831f-4119-9bbf-2ac849c4dff4_origin.pdf b/2024/LISO_ Lidar-only Self-Supervised 3D Object Detection/d6e4374b-831f-4119-9bbf-2ac849c4dff4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..370c2e36fb2d37fc94dbe8a1822251f68ae49399 --- /dev/null +++ b/2024/LISO_ Lidar-only Self-Supervised 3D Object Detection/d6e4374b-831f-4119-9bbf-2ac849c4dff4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5efeca5c07bf994f2345ac99e1bc6c252889d204b1547672fea81e7d07fe8412 +size 8494128 diff --git a/2024/LISO_ Lidar-only Self-Supervised 3D Object Detection/full.md b/2024/LISO_ Lidar-only Self-Supervised 3D Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..040134cf7592d07a234a0931bfc8eb7e240951f0 --- /dev/null +++ b/2024/LISO_ Lidar-only Self-Supervised 3D Object Detection/full.md @@ -0,0 +1,286 @@ +# LISO: Lidar-only Self-Supervised 3D Object Detection + +Stefan Andreas Baur $^{1,2}$ , Frank Moosmann $^{1}$ , and Andreas Geiger $^{2,3}$ + +1 Mercedes-Benz, Germany + +2 University of Tübingen + +3 Tübingen AI Center + +Abstract. 3D object detection is one of the most important components in any Self-Driving stack, but current state-of-the-art (SOTA) lidar object detectors require costly & slow manual annotation of 3D bounding boxes to perform well. Recently, several methods emerged to generate pseudo ground truth without human supervision, however, all of these methods have various drawbacks: Some methods require sensor rigs with full camera coverage and accurate calibration, partly supplemented by an auxiliary optical flow engine. Others require expensive high-precision localization to find objects that disappeared over multiple drives. + +We introduce a novel self-supervised method to train SOTA lidar object detection networks, requiring only unlabeled sequences of lidar point clouds. We call this trajectory-regularized self-training. It utilizes a SOTA self-supervised lidar scene flow network under the hood to generate, track, and iteratively refine pseudo ground truth. We demonstrate the effectiveness of our approach for multiple SOTA object detection networks across multiple real-world datasets. Code will be released1. + +Keywords: Self-Supervised $\cdot$ LiDAR $\cdot$ Object Detection + +# 1 Introduction + +Human developmental research reveals that infants less than one year of age are able to categorize animate and inanimate objects based on observed motion cues and can generalize this categorization to previously unseen objects [27]. Yet, lidar object detectors with SOTA performance are trained using manually selected and categorized annotations. These annotations are very expensive to obtain and become outdated quickly: new lidar sensors are coming to market regularly, and trained SOTA object detectors are sensitive to changes of sensor characteristics, mounting position and other domain gaps [5,8,18,26,32,34], even varying object sizes across different countries [29], making existing annotations difficult to transfer, resulting in high re-labeling efforts with each change. + +In this paper, we aim at bridging this gap by distilling motion cues observed in self-supervised lidar scene flow into SOTA single-frame lidar object detectors. + +![](images/e2bc39fdbf5cf91e3443405a50bfc7572b857682d65f6605438fa6c1ee13a9a6.jpg) +Fig. 1: Objects predicted by our method using no manual annotations. Red boxes are ground truth boxes, yellow boxes are predicted by our network. + +![](images/7719cef71cd1bfe6ca48e6dbb9bafb55519594dd4b3a49aaf3650b086a01579c.jpg) + +We introduce a novel and robust self-supervised, easy-to-use method, operating solely on lidar point cloud sequences for easy-to-obtain $360^{\circ}$ annotations. No human-annotated bounding boxes, cameras, costly high-precision GPS, or tedious sensor rig calibration is required, which could be especially useful in robotics research where data is often limited (even KITTI [11] only has forward facing cameras) or maintaining a multi-sensor-setup including its calibration is too cumbersome and expensive. Our method trains and runs a self-supervised lidar flow estimator [3] under the hood in order to create motion cues. We show in our experiments that these motion cues are a key factor to the success of our method. Based on the estimated lidar flow we bootstrap initial pseudo ground truth using simple clustering and track optimization. With these mined bounding boxes of moving objects we initialize a self-supervised, trajectory-regularized variant of self-training [1] (which is semi-supervised in its original form): We train a first version of a SOTA object detector, then iteratively re-generate and trajectory-regularize the pseudo ground truth and retrain the detector. Since the single-frame object detector has no concept of motion, it generalizes to detect any movable object on the way. Exemplary output of the trained detector, 3D boxes in single-frame point clouds, is depicted in Fig. 1. + +Our contribution is a novel self-supervised trajectory-regularized self-training framework for single-frame 3D object detection with the following properties: + +- It is based entirely on lidar, i.e. without the limitations of prior works: no cameras, no calibration, no high-precision GPS, no manual annotations. +- It is agnostic w.r.t. the sensor model, mounting pose, and detectors architecture and works with the same set of hyper parameters: We demonstrate this across four different datasets and different SOTA detector networks. +- It is able to generalize from moving objects (motion cues) to movable objects (final detection results) and significantly outperforms SOTA methods while being simpler. + +We show that using motion cues together with trajectory-regularized self-training is key to this success. The code of this approach will be published for easy use and comparison by other researchers. + +The paper is organized as follows: Sec. 2 discusses related work before the proposed method is described in detail in Sec. 3. Extensive experiments in Sec. 4 show the performance of our method before we conclude in Sec. 5. + +# 2 Related Work + +# 2.1 Single Frame Lidar 3D Object Detection + +Object detectors operating on 3D point clouds are an active research field. The currently best-performing ones are using deep neural networks trained via supervised learning and can be categorized by their internal representation: Some networks operate on points directly like PointRCNN [21], 3DSSD [35], and IASSD [39]. Others project the points either to a virtual range image [13,14,25] or into a voxel representation like VoxelNet [40], PointPillars, CenterPoint [36], and Transfusion-L [2]. However, all aforementioned methods require large human-annotated datasets in order to perform well and obtaining such annotations is very expensive. We address this problem in this paper, enabling training of SOTA object detectors using pseudo ground truth. + +# 2.2 Object Distilling from Motion Cues + +Multiple approaches have been suggested which leverage motion cues from lidar frame pairs in order to detect moving objects. Using the assumption of local geometric constancy, they decompose dynamic scenes into separate moving entities by applying as-rigid-as-possible optimization. Examples of these are the works by Dewan et al. [7], RSF [6] and OGC [23]. The two former methods are optimization-based whereas OGC uses these constraints as loss-function to self-supervise a segmentation network. Although these methods avoid the need for expensive labels, in contrast to our method they tend not to work well in low-resolution areas, they are typically slow, and can only detect moving but not movable (i.e. static but potentially moving) objects. + +# 2.3 Pseudo Ground Truth for Object Detection + +Different approaches have been proposed to mine pseudo ground truth for training object detectors: Najibi et al. [15] and Seidenschwarz et al. [19] use motion cues similar to section 2.2 in order to distill moving objects as pseudo ground truth and use it to train an object detector. [15] runs optimization for each frame pair to obtain lidar scene flow, clusters points, fits boxes, tracks them using a Kalman Filter and finally refines them on ICP-registered point clouds. [19] optimizes a clustering algorithm on motion cues through message passing which they then apply to segment point clouds into a set of instances and subsequently + +fit boxes. As shown in [19], and in contrast to our approach, both methods suffer from a large performance gap between moving and movable objects. We demonstrate that our method does not suffer from this gap. + +[12], [22] were the first to iteratively apply a detection, tracking, retraining paradigm to autonomous driving (AD) data. They build upon consistency constraints between object detectors which they train in the lidar and camera domain, making use of an optical flow network which is trained using supervision. The approach requires a calibrated sensor rig with possibly large camera coverage, IMU, and precision GPS. Similarly, [30] uses video sequences together with lidar scene flow to jointly train a camera and a lidar object detector. MODEST [33, 37] does not require coverage by calibrated cameras, but instead adds the additional requirement to having multiple lidar recordings of the same location in order to identify objects that vanished over time. With such demanding requirements, these approaches are not easy to use and are partially unsuited for popular AD datasets. + +Oyster [38] is the approach most similar to ours. It uses DBSCAN [9] on lidar point clouds to initialize pseudo ground truth. Clusters are then tracked using forward-and-reverse tracking in sensor coordinates (i.e. without consideration for ego motion), using a complex policy for confidence-based track retention. After training an object detector on close range data, they employ zero-shot generalization to the far range data, track again and iteratively retrain the detector. In their experiments they use different hyperparameters for different datasets. Our method, in contrast, explicitly considers sensor motion, produces much cleaner initial proposals by leveraging a self-supervised lidar scene flow network, does not require zero-shot-generalization from near-to-far-field, and works with current SOTA object detectors. We show on multiple datasets that our method outperforms [38], that it works robustly with the same set of hyper-parameters and also generalizes well to detect movable objects. Additionally, we will release our code to facilitate further research on this topic. + +# 3 Method + +A general overview of our method is sketched in Fig. 2 and some steps are illustrated in Fig. 3. As input, we take raw (unlabeled) point cloud sequences and undergo three stages, all of which are detailed in the following: Preprocessing of point clouds and lidar scene flow computation (Sec. 3.1), initial pseudo ground truth generation (Sec. 3.2) and repeated training with pseudo ground truth refinement (Sec. 3.3). The final output of the method is a trained object detector which can detect movable objects in raw single-frame point clouds. + +# 3.1 Preprocessing + +We preprocess the raw input point clouds as follows: + +Ground Removal: First, we remove distracting ground points from each single point cloud using JCP [20], which is a simple, robust, yet effective algorithm to remove ground points using changes in observed height above ground. + +![](images/a2419607cf4bada2dba4040260d1595d7e5c5598b8a839ccf64da75de44c50fc.jpg) +Fig. 2: Overview of the proposed method. Point cloud sequences are preprocessed (blue, Sec 3.1), initial pseudo ground truth is created (orange, Sec. 3.2) and the object detector is iteratively trained and pseudo ground truth regenerated (red, Sec. 3.2). + +Ego Motion Estimation: Second, we compute ego-motion between neighboring frame pairs using KISS-ICP [28], which is based on a robust version of ICP. The output is a cm-level accurate transformation $\mathbf{T}_{\mathrm{ego}}^{t\rightarrow t + 1}\in \mathbb{R}^{4\times 4}$ , describing the ego vehicle position at time $t + 1$ represented in the ego frame from time $t$ . + +Lidar Scene Flow Estimation: Third, we compute lidar scene flow between neighboring frame pairs resulting in a flow vector $\mathbf{f}_i = (dx, dy, dz)$ for every point $i$ in the first point cloud $\mathcal{P}^t$ . We chose to use SLIM [3] as its code is readily available, it is easy to use, features fast inference, and produces SOTA results. The network is trained self-supervised on raw point cloud sequences, minimizing a k-nearest-neighbor loss between forward and time-reversed point clouds. + +The components of our preprocessing steps (Ground Removal, Ego Motion Estimation, Lidar Scene Flow Estimation) have been selected for robustness and are all used with their default parameters from their respective publications. + +# 3.2 Initial Pseudo Ground Truth Generation + +The aim of our method is to mine pseudo ground truth for training a 3D lidar object detector. For a single point cloud $\mathcal{P}^t$ (consisting of $n \in \mathbb{N}$ points $\mathbf{p}_i$ , $\mathbf{p}_i \in \mathcal{P}^t$ ) this ground truth is a set of 3D bounding boxes $\mathcal{B}$ with confidences representing the objects at time $t$ : + +$$ +\mathcal {B} ^ {t} = \left\{\mathrm {B} _ {j} ^ {t}, j \in \mathbb {N} \right\} = \left\{(x, y, z, l, w, h, \theta , c) _ {j} \right\} \tag {1} +$$ + +Here, $(x,y,z) = \mathbf{x}$ define the center position, $(l,w,h)$ the length, width and height, $\theta$ the heading (orientation around up axis), and $c$ the confidence for a single box. + +The key success factor of our method is to focus on a high precision (and potentially low recall) of the initial set of bounding boxes in order to avoid "wrong" objects to negatively influence the object detector. We achieve this by leveraging shape, density, and especially motion cues to robustly identify moving objects solely (see also top-right of Fig. 3). Sec. 3.3 targets at generalizing these to movable objects later on. + +![](images/f3c4d1e935ad2a4bde8085c20d1a5f696b3250cf82c8b899d4479722a9be7957.jpg) + +![](images/3493f65229214257f38198882a27355dc4737ed0976b4ef54c51808fa6e87b19.jpg) + +![](images/97756f71ce6ff3ab58addacbf7064edebf7a1d303c00ab3b9ffe011287d42bbb.jpg) +Fig. 3: Overview over preprocessing, initial pseudo ground truth generation and training with examples. Top left: In the first step, self-supervised lidar scene flow is computed and corrected for vehicle ego-motion. Points are colored by flow direction and magnitude. Top right: In the second step, the scene flow is clustered and bounding boxes are fitted (to the moving objects). Bottom left: In the third step, the network is trained on the pseudo ground truth and is generalizing to static objects also, since it does not have the motion information as input signal. Points are thus colored by laser intensity. Bottom right: Ground truth, for reference. + +![](images/5b50df209bfd86b02f7e70be74b96d432d91fc41491b6203b968d88069d886be.jpg) + +Flow Clustering: The points $\mathbf{p}_i \in \mathcal{P}^t$ in each preprocessed point cloud are clustered based on geometry and motion: All stationary points in a scene should have a flow $\mathbf{f}_i$ similar to the vehicle's ego-motion, i.e. $\mathbf{f}_i \approx \mathbf{f}_{i,\mathrm{sta}} = ((\mathbf{T}_{\mathrm{ego}}^{t \rightarrow t + 1})^{-1} - \mathbf{I}_4) \cdot \mathbf{p}_i$ . The residual flow must then be caused by motion of other actors: $\mathbf{f}_{i,\mathrm{dyn}} = \mathbf{f}_i - \mathbf{f}_{i,\mathrm{sta}}$ . We filter all static points by applying a threshold of $1\mathrm{m/s}$ to the residual flow and cluster the remaining points based on their point location and flow vector in 6D using DBSCAN [9] (with parameters $\varepsilon = 1.0$ , $\min\mathrm{Pts} = 5$ ). + +We fit a 3D bounding box $\mathrm{B}_j^t$ to each resulting cluster following [37] and discard boxes with $l / w > 4.0$ , $lw < 0.35\mathrm{m}^2$ , or $lwh < 0.5\mathrm{m}^3$ . The heading $\theta$ is set to match the "forward-axis" of the motion, i.e. we orient the boxes along the direction of the residual flow $\mathbf{f}_{i,\mathrm{dyn}}$ . The confidence $c$ is initialized to 1 during clustering, and later replaced by detectors' confidence $cf$ . Section 3.3. + +Tracking: We run a simple flow based tracker: Since we have accurate egomotion available, we can track accurately in a fixed coordinate system w.r.t. the world. First, using the residual flow, we can compute a rigid body transform for each box at $t$ that transforms the proposed box $\mathbf{B}_i^t$ forward in time towards its suspected location at $t + 1$ , $\hat{\mathbf{B}}_i^{t + 1}$ . The propagated boxes $\hat{\mathcal{B}}^t + 1$ are matched greedily against the actual boxes $\mathcal{B}^{t + 1}$ based on box-center distance to the new detections. Unmatched boxes in $\mathcal{B}^{t + 1}$ which are further away than $1.5\mathrm{m}$ from propagated boxes spawn new tracklets. Unmatched tracklets from $\mathcal{B}^t$ are prop + +agated according to the last observed box motion for up to one time step, after that unmatched tracklets are terminated. We run tracking forward and reverse in time, connecting tracklets from forward and reverse tracking to tracks. + +The resulting set of tracks is post-filtered: We discard tracks that are shorter than 4 time steps or that have a median box confidence $c$ lower than the threshold 0.3. This retains only stable and high-confident tracks and avoids false positives to enter the pseudo ground truth. + +Track Optimization: We reduce positional noise of the tracks by minimizing translational jerk on all tracks longer than $3\mathrm{m}$ . Let $\mathcal{X}_{\mathrm{obs}}$ be the sequence of (noisy) observed box center positions $\mathbf{x}_i$ for consecutive time steps $i\in \{1,\dots,T\}$ of a track and their derivative $\frac{d\mathbf{x}_i}{dt}\approx \frac{\mathbf{x}_{i + 1} - \mathbf{x}_i}{\Delta t}$ . We compute smoothed track positions $\mathcal{X}_{\mathrm{smooth}}$ by initializing them to $\mathcal{X}_{\mathrm{obs}}$ and minimizing the following loss w.r.t. $\mathbf{x}_{\mathrm{smooth}}$ : + +$$ +L = \sum_ {i = 1} ^ {T} \left\| \frac {d ^ {3} \mathbf {x} _ {i , \text {s m o o t h}}}{d t ^ {3}} \right\| _ {2} ^ {2} + \beta \left\| \mathbf {x} _ {i, \text {s m o o t h}} - \mathbf {x} _ {i, \text {o b s}} \right\| _ {2} ^ {2} \tag {2} +$$ + +I.e. we minimize the jerk $\frac{d^3\mathbf{x}}{dt^3}$ and use a quadratic regularizer term $(\beta = 3)$ on the positions. This simple optimization goal is better suited for our method than more sophisticated ones like an unrolled bicycle model, as it is both faster and also class-agnostic, i.e. more universal: Using a bicycle model is problematic when optimizing pedestrian tracks, since our method does not allow to distinguish different classes. + +We subsequently align the orientation $\theta$ of each detection in a track to the direction of the smoothed track at its position. Box dimensions $\{l, w, h\}$ of all boxes in a track are adapted to the 90 percentile of observed box dimensions in a track. All hyperparameters related to clustering, tracking and track optimization have been tuned visually on two sample snippets from nuScenes [4]. + +# 3.3 Trajectory-Regularized Self-Training + +After having mined initial pseudo ground truth of moving objects with a high precision and low recall, we now aim at iteratively improving our pseudo ground truth by training and using an object detector so that the pseudo ground truth generalizes to movable objects. We achieve this by executing iterative trajectory-regularized self-training, which is composed of the following two steps: + +Training: We train the target object detection network using the current pseudo ground truth in a supervised training setup. Any single-frame object detection network can be plugged into our pipeline. In our experiments we do not deviate from the basic training setup of our object detection networks. Like any SOTA object detection method, we apply standard augmentation techniques to a point cloud during network training: Random rotation, scaling $(\pm 5\%)$ , and random translation up to $5\mathrm{m}$ around the origin. Furthermore, we randomly pick 1 to 15 objects from the pseudo ground truth database and insert a random subset of their points at random locations into the scene. + +Pseudo Ground Truth Regeneration: We halt training after a set number of training steps $s$ ( $s = 30k$ in our experiments) and create new, improved pseudo ground truth by running the trained detector in inference mode over all sequences in the training dataset, using the detectors' confidence as box confidence $c$ . We regularize these detections by running the flow-based tracker and track smoothing exactly like we do for the initial pseudo ground truth generation. Every 2nd iteration (i.e. every 60k steps in the experiments), we discard the network weights after regenerating the pseudo ground truth. + +Fig. 4 illustrates the effect of our self-training: We see that the pseudo ground truth improves with each re-generation and that dropping network weights allows the network to re-focus on the generalized pseudo ground truth. Our strict requirements towards boxes being added to the pseudo ground truth leads to lower recall compared to the network, which as a result scores higher overall. + +Two aspects are key to making our method perform well: + +- Being restrictive when composing the pseudo ground truth (i.e. using plausible tracks of high-confidence network detections or clusters with significant flow only) avoids adding false positives into the pseudo ground truth and hence avoids that our network increasingly hallucinates with each iteration. + +- Not using flow but only single-frame point clouds as input for the detector allows the detector to focus on appearance of objects solely. + +These allow our method to generalize from initially mined moving objects to movable objects in the scene. + +# 4 Evaluation + +We evaluate our method on multiple datasets across multiple SOTA networks. SLIM [3], the self-supervised flow network we use, is trained and inferred as in the published version, but Bird's-Eye-View (BEV) range is extended from $70 \times 70\mathrm{m}$ , $640 \times 640$ pixels to $120 \times 120\mathrm{m}$ and $920 \times 920$ pixels. + +# 4.1 Datasets and Metrics + +We evaluate our method on four different AD datasets. For a fair comparison, we compute metrics for movable objects by mapping all animate objects (Cars, Trucks, Trailers, Motorcycles, Cyclists, Pedestrians and other Vehicles) to a single category and discard all inanimate objects (Barrier, Traffic Cone,...) since our method does not predict any class attributes. In Table 2 and Table 3 we also give class-based results for completeness. Class labels for true positives are taken from ground truth. For false positives, the predicted class label is assigned randomly according to the label distribution in the dataset. + +Waymo Open Dataset (WOD): [24] is a large, geographically diverse dataset recorded with a proprietary high quality lidar. We evaluate using the protocol of [15, 19], using an area of whole $100\mathrm{m} \times 40\mathrm{m}$ BEV grid around the ego vehicle, artificially cropping the predictions of our method down to this reduced area. + +![](images/2aa3199686d64cc933f1da60d195c9e087c79b9baeac1b3243da51ce6ca977e0.jpg) +(a) Effect of self-supervised self-learning + +![](images/cb4fb497bafe7662ab370a0bb8aa987f7ca827223f5cfb44096feb9b7233fc70.jpg) +(b) Evaluation of Self-Training Hyperparameters +Fig. 4: Left: Effect of self-supervised self-learning: At the beginning of the training, only moving objects are contained within the pseudo ground truth, which thus scores 0 on static objects. Thanks to the self-training, the network generalizes to movable objects and the score of the pseudo ground truth on static objects starts to increase with every regeneration at each "X" in the graph. The pseudo ground truth's performance is measured on a subset of the train set of WOD dataset since our tracking is looking ahead in time, which would be unfair w.r.t. other methods. The network (here: Centerpoint) is evaluated on a small subset of the validation set. Right: Evaluation of Self-Training Hyperparameters: Network (Centerpoint) performance over the course of a training on WOD. $s$ is the number of training steps between pseudo ground truth regenerations, $r$ is the number of regenerations after which network weights are dropped and re-initialized. + +KITTI: [10, 11] is recorded using a Velodyne HDL64 lidar sensor, where some parts have been annotated with 3D object boxes and tracking information in the forward facing camera field of view. A large portion of the published data is only available in raw format without annotations, which our method is able to leverage, due to not requiring any ground truth annotations for training. + +Argoverse 2: AV2 [31] is recorded using two stacked Velodyne VLP32 lidar sensors and annotated with 3D object boxes. On AV2 and KITTI, we evaluate both 2D (in BEV space) and 3D box IoU at IoU thresholds of 0.3 and 0.5 in the area of $100 \times 100\mathrm{m}$ around the ego vehicle. + +nuScenes: [4] is recorded using a Velodyne VLP32 lidar sensor. We evaluate using the nuScenes protocol. + +# 4.2 Networks + +We evaluate our proposed method with two SOTA lidar object detection networks of different architecture: Centerpoint [36] and Transfusion-L [2]. We do not modify the published network architectures or losses and train both networks with a batch size of four. + +CenterPoint (CP): [36] is based on 2D convolutions in BEV space. Object centers are represented as heatmap in the BEV, which are then reduced to a sensible amount of boxes using non-maximum-suppression. + +Transfusion-L (TF): [2] is a Transformer based architecture, which is applied after initial encoding of the LiDAR point cloud into a 2D BEV feature tensor. + +
AV2nuScenes
BEV-IOU3D-IOUAP↑AOE↓
AP@0.3↑AP@0.5↑AP@0.3↑AP@0.5↑
Sup.
CP [36]0.7780.7780.7360.5170.4840.560
TF [2]0.7830.6540.7670.6010.6270.501
Unsupervised
DBSCAN [9]0.0540.0260.0200.0030.0080.171
DBSCAN(SF)0.0700.0260.0650.0150.0030.082
DBSCAN(GF)0.1050.0410.0880.0240.0001.000
RSF [6]0.0740.0260.0550.0140.0191.003
Self Train
Oyster [38] †0.3540.245----
Oyster-CP [38]0.3810.2660.1500.0020.0911.514
Oyster-TF [38]0.3400.1980.1820.0160.0931.564
LISO-CP0.4480.3350.3670.1880.1091.062
LISO-TF0.4170.2940.3170.1760.1340.938
+ +Table 1: Evaluation on AV2 [31] and nuScenes [4]: CP, TF: network architecture, in the first two rows trained supervised for comparison. †: Metrics as reported in [38]. SF: self-supervised lidar scene flow by SLIM, GF: ground truth lidar scene flow. Note that nuScenes uses a minimum precision and recall threshold of 0.1, and since the recall of GT flow clustering is lower than 0.1, all results are clipped away. For full nuScenes scores see supplementary material. + +# 4.3Baselines + +Besides the obvious baseline of training the object detection networks supervised from scratch on ground truth from the dataset, we also compare to multiple strong unsupervised baselines. For details, see Sec. 2. + +RSF: We include RSF [6] in our evaluation as strong representative of methods doing object distilling from motion cues. We ran experiments using the published code of the authors. + +DBSCAN: This algorithm [9] clusters points in the point cloud with similar 3D locations into objects. We additionally evaluate extending the cluster space to 6D by either using SLIM scene flow (SF) or ground-truth scene flow (GT). Thus, the "DBSCAN(SF)" baseline essentially corresponds to the raw detections used for our initial pseudo ground truth generation. We use the implementation from [17] with parameter values 1.0 for epsilon and 5 for the minimum number of points for a valid cluster. + +SeMoLi: [19] is the most recent baseline. As it showed to outperform [15] as self-supervised object detection method we include only published results of [19] in our evaluation. + +Oyster: We implemented [38] using PyTorch [16], and apply the proposed framework to the CenterPoint (Oyster-CP) and Transfusion-L (Oyster-TF) to be as close as possible to our method and evaluation. We verify our reimplementation, by reproducing the reported metrics by [38] on AV2, see Table 1. + +
AP@0.4 ↑AP@0.5 ↑AP@0.7 ↑
MovableMovingStillVehi.Ped.Cycl.MovableMovable
BEV3DBEV3DBEV3DBEVBEVBEVBEVBEV3DBEV3D
Sup.CP [36]0.7650.6840.7210.6240.7350.6570.9120.5130.1340.7190.5780.5100.202
TF [2]0.7460.7230.7140.6680.7330.7100.9180.4570.2160.7110.6690.5320.385
UnsupervisedDBSCAN [9]0.0270.0080.0090.0000.0270.0060.1840.0020.0010.0190.0020.0000.000
DBSCAN(SF)0.0260.0100.0640.0410.0000.0000.0730.0100.0090.0110.0010.0000.000
DBSCAN(GF)0.1140.0710.3180.1200.0000.0000.1130.1110.2400.0520.0230.0020.000
RSF [6]0.0300.0200.0800.0550.0000.0000.1090.0000.0020.0250.0120.0110.000
SeMoLi [19] †-0.195-0.575--------0.018
LISO-CP0.2920.2110.2720.2040.2080.1400.6070.0290.0100.2390.1460.1100.019
Self TrainOyster-CP [38]0.2170.0840.1510.0620.1760.0560.5620.0000.0000.1760.0230.0670.000
Oyster-TF [38]0.1210.0150.0510.0070.0980.0100.4750.0000.0000.0920.0030.0310.000
LISO-CP0.3800.3080.3500.2960.3220.2550.6950.0550.0220.3430.2190.1820.021
LISO-TF0.3270.2080.3490.2450.2330.1260.6690.0240.0120.2770.1360.1160.009
+ +Table 2: Evaluation on WOD dataset: We evaluate using the protocol of [15, 19], using an area of whole $100\mathrm{m}\times 40\mathrm{m}$ BEV grid around the ego vehicle, considering objects that move faster than $1\mathrm{m/s}$ to be moving (difficulty level L2). CP, TF: network architecture, in the first two lines trained supervised for comparison. $\dagger$ : Results taken from [19]. SF: lidar scene flow by SLIM, GF: ground truth lidar scene flow. For class-specific 3D detection scores see supplementary material. + +# 4.4 Results + +Quantitative Results: On all four datasets we see that our method consistently outperforms all self-supervised baselines in all metrics for movable objects, see Table 1, Table 2, and Table 3. Only SeMoLi [19] beats LISO on moving objects (see Table 2) but significantly drops below LISO's performance on movable objects, hinting to difficulties with generalization. Our method does not suffer from this performance gap and performs nearly equally well on movable and moving objects. We also observe that, despite mostly better performance in the supervised case, Transfusion-L responds less favorable than Centerpoint to the self-supervised training approaches. We suspect that Centerpoint, due to its convolutional architecture and its lack of positional encoding compared to Transfusion-L, is less susceptible to overfitting to where moving objects have been observed and where not to expect them. + +We also observe that the nuScenes dataset seems to be the most challenging for the self-supervised methods, leading to the biggest gap between supervised and unsupervised object detection performance. We attribute this to the sparseness of the lidar sensor with only 32 layers. All self-supervised methods have difficulties estimating the orientation of objects (AOE on nuScenes, $2\pi$ periodic, see Table 1), which the IoU-based metrics cannot reveal. + +# Qualitative Results: + +In Fig. 5 we compare our method to the best-performing baseline, Oyster, using Centerpoint as network architecture. It can be seen, that our method has higher accuracy and predicts box orientations more accurately. + +
Movable (Moving & Still)CarPedestrianCyclist
BEV-IOUAP@0.3↑3D-IOUAP@0.5↑AP@0.3↑AP@0.5↑BEV-IOUAP@0.3↑AP@0.5↑BEV-IOUAP@0.3↑AP@0.5↑BEV-IOUAP@0.3↑AP@0.5↑
CP [36]0.7550.6900.7360.6010.8140.7940.3700.1280.4090.157
TF [2]0.7470.6650.7290.5820.8200.7760.3110.0960.2630.032
DBSCAN [9]0.0230.0020.0100.0000.0260.0050.0000.0000.0640.007
RSF [6]0.0290.0190.0290.0110.0660.0490.0000.0000.1920.043
Oyster-CP [38]0.2350.0980.1140.0000.3270.1350.0000.0000.0000.000
Oyster-TF [38]0.2730.0880.1280.0000.3640.1210.0000.0000.0190.000
LISO-CP0.4460.3300.4190.1590.5200.4110.0970.0190.4450.053
LISO-TF0.3610.2070.2940.0360.4250.2970.0840.0140.3480.003
+ +Table 3: Evaluation on KITTI dataset: We evaluate on the forward facing field of view where GT annotations are available, but run inference on the whole $100 \times 100\mathrm{m}$ BEV grid. Also note that flow annotations are not available for KITTI Object. CP, TF: network architecture, in the first two lines trained supervised for comparison. + +![](images/933af890fd25eaf0af9a63097052fdb38d181555c5caab7463bbc734aecb6bf8.jpg) +Fig. 5: Qualitative Results on WOD. Red boxes are ground truth boxes, yellow boxes are predictions. Left: Oyster-CP Right: LISO-CP + +![](images/45ee8413268cef33df008a415851f006e812015a85c8c8752cb91ee27de6f2f3.jpg) + +Fig. 6, in contrast, showcases an important failure-case and general problem of all unsupervised methods: ground truth, especially in the WOD dataset, is also annotated in very sparse and distant regions. It is very challenging to generate pseudo ground truth for such regions whereas supervised trainings do not have to generalize from moving to such movable (but very sparse) objects. This may be the main reason for a still noticeable gap in evaluation results between supervised and unsupervised methods. + +# 4.5 Ablation Study + +We investigated the influence of different choices for the components in our pipeline, see Table 4 and cf. Fig. 2. + +![](images/0b029a796ed8a84455295746dea26f96686598721ca138d91a8e0a4ee81c26f2.jpg) +Fig. 6: Missed ground truth boxes in the WOD dataset. + +Motion Cues as Clustering Input: When adding motion cues to box clustering for Oyster we observe that detection performance already increases noticeably (comparing #3 to #4, Table 4). This demonstrates that self-supervised motion cues are a key ingredient to get high-quality initial pseudo ground truth as opposed to just using clustered 3D points. + +
TagClusterInputMethodSelf TrainTrackOptim.AP@0.4(BEV)AP@0.4(3D)
#1P, SFno tracker0.1770.086
#2P, SFno tracker✓0.2010.131
#3POyster✓0.2170.084
#4P, SFOyster✓0.2550.104
#5P, SFLISO(K, SF)0.2790.209
#6P, SFLISO(K, SF)✓0.3600.290
#7P, SFLISO(K, SF)✓0.2920.211
#8P, SFLISO(K, SF)✓✓0.3800.308
#9P, SFLISO(G, SF)✓✓0.4110.339
#10P, GFLISO(G, GF)✓✓0.4230.349
+ +Table 4: Ablation study of different parts of our pipeline on WOD with Centerpoint [36]. Abbreviations: P: points, SF: self-supervised lidar scene flow (SLIM), GF: ground truth lidar scene flow. K: KISS-ICP ego-motion, G: ground truth ego-motion. + +Motion Cues for Tracking: Our scene-flow-based tracker is the biggest contributing factor to the success of our method: the ablation reveals that leveraging motion cues greatly improves tracking (comparing #4 to #6) and thus ultimately improves self-training, i.e. more accurate tracking allows for much more strictness when matching and filtering tracklets. The Oyster tracker uses a matching threshold of $5\mathrm{m}$ , our threshold is only $1.5\mathrm{m}$ . This strictness leads to higher quality pseudo ground truth. + +Effect of Self Training: Even without using any tracker to filter tracklets self training has a benefit (comparing #1 to #2). This is due to the network + +generalizing to new instances and confidence thresholding the detections. However, having an additional way to discard implausible detections (by tracking) amplifies the positive effect of self training (comparing #5 to #6 and #7 to #8), as it can better prevent false positives from entering the pseudo ground truth. + +Track Optimization: The added benefit of track optimization is more independent of Self Training (comparing going from \#5 to \#7 with going from \#6 to \#8). Finally we notice, that with the combination of a strict tracker, track optimization and self training, the performance is very close to using ground truth flow and ego-motion (comparing \#8 to \#9 and \#10). + +Hyperparameters for Iterative Training: Fig. 4 demonstrates that the performance of our proposed method is relatively consistent across a range of hyperparameters. However, using too few steps between regeneration of pseudo ground truth ( $s = 20k$ ) leads to degrading performance, because the network does not have enough time to generalize and stabilize using the pseudo ground truth, which then has negative effects during regeneration of the pseudo ground truth. The experiment reveals that doing multiple iterations of self-learning and dropping network weights to allow the network to adjust is beneficial for a good performance. We selected $r = 2$ and $s = 30k$ as a good compromise between performance and speed for all other experiments. + +# 5 Conclusion + +We have proposed a novel framework for self-supervised 3D lidar object detection based on self-supervised lidar scene flow. Using simple clustering on lidar scene flow in combination with a novel flow-based tracker, we were able to generate pseudo ground truth with moving objects at high precision. This enabled us to bootstrap SOTA 3D lidar object detectors, without using any human labels. The detector generalizes from detecting moving to detecting movable objects over multiple self-training iterations. Experiments revealed that our trajectory-regularized self-learning, based on our scene-flow-based tracker, is key to the success of our method. We have demonstrated the effectiveness of our approach on two SOTA lidar object detectors as well as four AD datasets. All experiments were conducted with the same set of hyperparameters, demonstrating the robustness and wide applicability of our method. Our method achieves a significant improvement in the state-of-the-art of self-supervised lidar object detection on all four datasets. The biggest limitation of our approach is that our method is not able to distinguish different classes. Future work could be around generating pseudo ground truth for class labels, e.g. based on motion or size characteristics. + +# References + +1. Amini, M., Feofanov, V., Pauletto, L., Devijver, E., Maximov, Y.: Self-training: A survey. CoRR abs/2202.12040 (2022), https://arxiv.org/abs/2202.12040 +2. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: + +IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 1080-1089. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.00116, https://doi.org/10.1109/CVPR52688.2022.00116 +3. Baur, S.A., Emmerichs, D.J., Moosmann, F., Pinggera, P., Ommer, B., Geiger, A.: SLIM: self-supervised lidar scene flow and motion segmentation. In: 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021. pp. 13106-13116. IEEE (2021). https://doi.org/10.1109/ICCV48922.2021.01288, https://doi.org/10.1109/ICCV48922.2021.01288 +4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. pp. 11618-11628. IEEE (2020). https://doi.org/10.1109/CVPR42600.2020.01164, https://doi.org/10.1109/CVPR42600.2020.01164 +5. Chen, Z., Luo, Y., Wang, Z., Baktashmotlagh, M., Huang, Z.: Revisiting domain-adaptive 3d object detection by reliable, diverse and class-balanced pseudolabeling. In: IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023. pp. 3691-3703. IEEE (2023). https://doi.org/10.1109/ICCV51070.2023.00344, https://doi.org/10.1109/ICCV51070.2023.00344 +6. Deng, D., Zakhor, A.: Rsf: Optimizing rigid scene flow from 3d point clouds without labels. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1277-1286 (January 2023) +7. Dewan, A., Caselitz, T., Tipaldi, G.D., Burgard, W.: Motion-based detection and tracking in 3d lidar scans. In: Kragic, D., Bicchi, A., Luca, A.D. (eds.) 2016 IEEE International Conference on Robotics and Automation, ICRA 2016, Stockholm, Sweden, May 16-21, 2016. pp. 4508-4513. IEEE (2016). https://doi.org/10.1109/ICRA.2016.7487649, https://doi.org/10.1109/ICRA.2016.7487649 +8. Eskandar, G.: An empirical study of the generalization ability of lidar 3d object detectors to unseen domains. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23815-23825 (June 2024) +9. Ester, M., Kriegel, H., Sander, J., Xu, X.: A density-based algorithm for discovering clusters in large spatial databases with noise. In: Simoudis, E., Han, J., Fayyad, U.M. (eds.) Proceedings of the Second International Conference on Knowledge Discovery and Data Mining (KDD-96), Portland, Oregon, USA. pp. 226-231. AAAI Press (1996), http://www.aaai.org/Library/KDD/1996/kdd96-037.php +10. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The KITTI dataset. Int. J. Robotics Res. 32(11), 1231-1237 (2013). https://doi.org/10.1177/0278364913491297, https://doi.org/10.1177/0278364913491297 +11. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the KITTI vision benchmark suite. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition, Providence, RI, USA, June 16-21, 2012. pp. 3354-3361. IEEE Computer Society (2012). https://doi.org/10.1109/CVPR.2012.6248074, https://doi.org/10.1109/CVPR.2012.6248074 +12. Harley, A.W., Zuo, Y., Wen, J., Mangal, A., Potdar, S., Chaudhry, R., Fragkiadaki, K.: Track, check, repeat: An EM approach to unsupervised tracking. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 16581-16591. Computer Vision Foundation / IEEE (2021). https: + +//doi.org/10.1109/CVPR46437.2021.01631, https://openaccess.thecvf.com/content/CVPR2021/html/Harley_Track_Check_Repeat_An_EM_Approach_to_Unsupervised_Tracking_CVPR_2021_paper.html +13. Liang, Z., Zhang, Z., Zhang, M., Zhao, X., Pu, S.: Rangeioudet: Range image based real-time 3d object detector optimized by intersection over union. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 7140-7149. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.00706, https://openaccess.thecvf.com/content/CVPR2021/html/Liang_RangeIoUDet_Range_Issue-Based.Real-Time_3D_Object_Detector_Optimized_by_CVPR_2021_paper.html +14. Meyer, G.P., Laddha, A., Kee, E., Vallespi-Gonzalez, C., Wellington, C.K.: Lasernet: An efficient probabilistic 3d object detector for autonomous driving. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019. pp. 12677-12686. Computer Vision Foundation / IEEE (2019). https://doi.org/10.1109/CVPR.2019.01296, http://openaccess.thecvf.com/content_CVPR_2019/html/Meyer_LaserNet_An_Efficient_Probabilistic_3D_Object_Detector_for_Autonomous_Driving_CVPR_2019_paper.html +15. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Motion inspired unsupervised perception and prediction in autonomous driving. In: Avidan, S., Brostow, G.J., Cissé, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVIII. Lecture Notes in Computer Science, vol. 13698, pp. 424-443. Springer (2022). https://doi.org/10.1007/978-3-031-19839-7_25, https://doi.org/10.1007/978-3-031-19839-7_25 +16. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Köpf, A., Yang, E.Z., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Wallach, H.M., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E.B., Garnett, R. (eds.) Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada. pp. 8024-8035 (2019), https://proceedings.neurips.cc/paper/2019/hash/bdbca288fee7f92f2bfa9f7012727740-AAbstract.html +17. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine learning in Python. Journal of Machine Learning Research 12, 2825–2830 (2011) +18. Rist, C.B., Enzweiler, M., Gavrila, D.M.: Cross-sensor deep domain adaptation for lidar detection and segmentation. In: 2019 IEEE Intelligent Vehicles Symposium, IV 2019, Paris, France, June 9-12, 2019. pp. 1535-1542. IEEE (2019). https://doi.org/10.1109/IVS.2019.8814047, https://doi.org/10.1109/IVS.2019.8814047 +19. Seidenschwarz, J., Osep, A., Ferroni, F., Lucey, S., Leal-Taixe, L.: Semoli: What moves together belongs together (2024), https://arxiv.org/abs/2402.19463 +20. Shen, Z., Liang, H., Lin, L., Wang, Z., Huang, W., Yu, J.: Fast ground segmentation for 3d lidar point cloud based on jump-convolution-process. Remote. Sens. 13(16), 3239 (2021). https://doi.org/10.3390/rs13163239, https://doi.org/10.3390/rs13163239 + +21. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019. pp. 770-779. Computer Vision Foundation / IEEE (2019). https://doi.org/10.1109/CVPR.2019.00086, http://openaccess.thecvf.com/content_CVPR_2019/html/Shi_PointRCNN_3D Object_Proposal_Generation_and_Detection_From_Point_Cloud_CVPR_2019_paper.html +22. Shin, S., Golodetz, S., Vankadari, M., Zhou, K., Markham, A., Trigoni, N.: Sample, crop, track: Self-supervised mobile 3d object detection for urban driving lidar. CoRR abs/2209.10471 (2022). https://doi.org/10.48550/arXiv.2209.10471, https://doi.org/10.48550/arXiv.2209.10471 +23. Song, Z., Yang, B.: OGC: Unsupervised 3d object segmentation from rigid dynamics of point clouds. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=ecNbE00tqBU +24. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. pp. 2443-2451. IEEE (2020). https://doi.org/10.1109/CVPR42600.2020.00252, https://doi.org/10.1109/CVPR42600.2020.00252 +25. Sun, P., Wang, W., Chai, Y., Elsayed, G., Bewley, A., Zhang, X., Sminchisescu, C., Anguelov, D.: RSN: range sparse net for efficient, accurate lidar 3d object detection. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 5725-5734. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.00567, https://openaccess.thecvf.com/content/CVPR2021/html/Sun_RSN_Range_Sparse_Net_for_Efficient_Accurate_LiDAR_3D_Object_CVPR_2021_paper.html +26. Théodose, R., Denis, D., Chateau, T., Frémont, V., Checchin, P.: A deep learning approach for lidar resolution-agnostic object detection. IEEE Trans. Intell. Transp. Syst. 23(9), 14582-14593 (2022). https://doi.org/10.1109/TITS.2021.3130487, https://doi.org/10.1109/TITS.2021.3130487 +27. Trauble, B., Pauen, S., Poulin-Dubois, D.: Speed and direction changes induce the perception of animacy in 7-month-old infants. Frontiers in Psychology 5 (2014). https://doi.org/10.3389/fpsyg.2014.01141, https://www.frontiersin.org/articles/10.3389/fpsyg.2014.01141 +28. Vizzo, I., Guadagnino, T., Mersch, B., Wiesmann, L., Behley, J., Stachniss, C.: KISS-ICP: in defense of point-to-point ICP - simple, accurate, and robust registration if done the right way. IEEE Robotics Autom. Lett. 8(2), 1029–1036 (2023). https://doi.org/10.1109/LRA.2023.3236571, https://doi.org/10.1109/LRA.2023.3236571 +29. Wang, Y., Chen, X., You, Y., Li, L.E., Hariharan, B., Campbell, M.E., Weinberger, K.Q., Chao, W.: Train in germany, test in the USA: making 3d object detectors generalize. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. pp. 11710-11720. Computer Vision Foundation / IEEE (2020). https://doi.org/10.1109/CVPR42600.2020.01173, https://openaccess.thecvf.com/content_CVPR_2020/html/Wang_ + +Train_in_Germany_Test_in_the_USA_Making_3D_Object_CVPR_2020_paper.html +30. Wang, Y., Chen, Y., Zhang, Z.: 4d unsupervised object discovery. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022 (2022), http://papers.nips.cc/paper_files/paper/2022/ hash/e7407ab5e89c405d28ff6807ffec594a-Abstraction-Conference.html +31. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., Ramanan, D., Carr, P., Hays, J.: Argov-verse 2: Next generation datasets for self-driving perception and forecasting. In: Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021) (2021) +32. Wozniak, M.K., Hansson, M., Thiel, M., Jensfelt, P.: UADA3D: unsupervised adversarial domain adaptation for 3d object detection with sparse lidar and large domain gaps. CoRR abs/2403.17633 (2024). https://doi.org/10.48550/ARXIV.2403.17633, https://doi.org/10.48550/arXiv.2403.17633 +33. Xu, J., Waslander, S.L.: Hypermodest: Self-supervised 3d object detection with confidence score filtering. CoRR abs/2304.14446 (2023). https://doi.org/10.48550/arXiv.2304.14446, https://doi.org/10.48550/arXiv.2304.14446 +34. Xu, Q., Zhou, Y., Wang, W., Qi, C.R., Anguelov, D.: SPG: unsupervised domain adaptation for 3d object detection via semantic point generation. In: 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021. pp. 15426-15436. IEEE (2021). https://doi.org/10.1109/ICCV48922.2021.01516, https://doi.org/10.1109/ICCV48922.2021.01516 +35. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3dssd: Point-based 3d single stage object detector. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. pp. 11037-11045. Computer Vision Foundation / IEEE (2020). https://doi.org/10.1109/CVPR42600.2020.01105, https://openaccess.thecvf.com/content_CVPR_2020/html/Yang_3DSSD_Point-Based_3D_Single_SStage_Object_Detector_CVPR_2020_paper.html +36. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 11784-11793. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.01161, https://openaccess.thecvf.com/content/CVPR2021/html/Yin_Center-Based_3D_Object_Detection_and_Tracking_CVPR_2021_paper.html +37. You, Y., Luo, K., Phoo, C.P., Chao, W., Sun, W., Hariharan, B., Campbell, M.E., Weinberger, K.Q.: Learning to detect mobile objects from lidar scans without labels. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 1120-1130. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.00120, https://doi.org/10.1109/CVPR52688.2022.00120 +38. Zhang, L., Yang, A.J., Xiong, Y., Casas, S., Yang, B., Ren, M., Urtasun, R.: Towards unsupervised object detection from lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9317-9328 (June 2023) + +39. Zhang, Y., Hu, Q., Xu, G., Ma, Y., Wan, J., Guo, Y.: Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 18931-18940. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.01838, https://doi.org/10.1109/CVPR52688.2022.01838 +40. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018. pp. 4490-4499. IEEE Computer Society (2018). https://doi.org/10.1109/CVPR.2018.00472, http://openaccess.thecvf.com/content_cvpr_2018/html/Zhou_VoxelNet_End-to-End_Learning_CVPR_2018_paper.html \ No newline at end of file diff --git a/2024/LISO_ Lidar-only Self-Supervised 3D Object Detection/images.zip b/2024/LISO_ Lidar-only Self-Supervised 3D Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4e83e60bda2939adc948567f4fdc21c4b9420110 --- /dev/null +++ b/2024/LISO_ Lidar-only Self-Supervised 3D Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e863160659c32ab4b75f5ce056d5c0510bbd98559b23b90e992029d0e2d4937 +size 746301 diff --git a/2024/LISO_ Lidar-only Self-Supervised 3D Object Detection/layout.json b/2024/LISO_ Lidar-only Self-Supervised 3D Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..abf61de21be5aafa4198ab8c2283b75b8202afcd --- /dev/null +++ b/2024/LISO_ Lidar-only Self-Supervised 3D Object Detection/layout.json @@ -0,0 +1,8807 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 111, + 460, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 111, + 460, + 146 + ], + "spans": [ + { + "bbox": [ + 153, + 111, + 460, + 146 + ], + "type": "text", + "content": "LISO: Lidar-only Self-Supervised 3D Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 151, + 167, + 462, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 167, + 462, + 181 + ], + "spans": [ + { + "bbox": [ + 151, + 167, + 462, + 181 + ], + "type": "text", + "content": "Stefan Andreas Baur" + }, + { + "bbox": [ + 151, + 167, + 462, + 181 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 151, + 167, + 462, + 181 + ], + "type": "text", + "content": ", Frank Moosmann" + }, + { + "bbox": [ + 151, + 167, + 462, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 151, + 167, + 462, + 181 + ], + "type": "text", + "content": ", and Andreas Geiger" + }, + { + "bbox": [ + 151, + 167, + 462, + 181 + ], + "type": "inline_equation", + "content": "^{2,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 249, + 190, + 364, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 190, + 364, + 201 + ], + "spans": [ + { + "bbox": [ + 249, + 190, + 364, + 201 + ], + "type": "text", + "content": "1 Mercedes-Benz, Germany" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 255, + 201, + 358, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 201, + 358, + 212 + ], + "spans": [ + { + "bbox": [ + 255, + 201, + 358, + 212 + ], + "type": "text", + "content": "2 University of Tübingen" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 260, + 212, + 353, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 212, + 353, + 223 + ], + "spans": [ + { + "bbox": [ + 260, + 212, + 353, + 223 + ], + "type": "text", + "content": "3 Tübingen AI Center" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 255, + 453, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 255, + 453, + 354 + ], + "spans": [ + { + "bbox": [ + 160, + 255, + 453, + 354 + ], + "type": "text", + "content": "Abstract. 3D object detection is one of the most important components in any Self-Driving stack, but current state-of-the-art (SOTA) lidar object detectors require costly & slow manual annotation of 3D bounding boxes to perform well. Recently, several methods emerged to generate pseudo ground truth without human supervision, however, all of these methods have various drawbacks: Some methods require sensor rigs with full camera coverage and accurate calibration, partly supplemented by an auxiliary optical flow engine. Others require expensive high-precision localization to find objects that disappeared over multiple drives." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 354, + 453, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 354, + 453, + 431 + ], + "spans": [ + { + "bbox": [ + 159, + 354, + 453, + 431 + ], + "type": "text", + "content": "We introduce a novel self-supervised method to train SOTA lidar object detection networks, requiring only unlabeled sequences of lidar point clouds. We call this trajectory-regularized self-training. It utilizes a SOTA self-supervised lidar scene flow network under the hood to generate, track, and iteratively refine pseudo ground truth. We demonstrate the effectiveness of our approach for multiple SOTA object detection networks across multiple real-world datasets. Code will be released1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 442, + 394, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 442, + 394, + 453 + ], + "spans": [ + { + "bbox": [ + 160, + 442, + 394, + 453 + ], + "type": "text", + "content": "Keywords: Self-Supervised " + }, + { + "bbox": [ + 160, + 442, + 394, + 453 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 442, + 394, + 453 + ], + "type": "text", + "content": " LiDAR " + }, + { + "bbox": [ + 160, + 442, + 394, + 453 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 442, + 394, + 453 + ], + "type": "text", + "content": " Object Detection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 475, + 230, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 230, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 230, + 488 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 502, + 482, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 621 + ], + "type": "text", + "content": "Human developmental research reveals that infants less than one year of age are able to categorize animate and inanimate objects based on observed motion cues and can generalize this categorization to previously unseen objects [27]. Yet, lidar object detectors with SOTA performance are trained using manually selected and categorized annotations. These annotations are very expensive to obtain and become outdated quickly: new lidar sensors are coming to market regularly, and trained SOTA object detectors are sensitive to changes of sensor characteristics, mounting position and other domain gaps [5,8,18,26,32,34], even varying object sizes across different countries [29], making existing annotations difficult to transfer, resulting in high re-labeling efforts with each change." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 622, + 481, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 481, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 481, + 647 + ], + "type": "text", + "content": "In this paper, we aim at bridging this gap by distilling motion cues observed in self-supervised lidar scene flow into SOTA single-frame lidar object detectors." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 286, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 286, + 665 + ], + "type": "text", + "content": "1 https://github.com/baurst/liso" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 305, + 242 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 305, + 242 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 305, + 242 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 305, + 242 + ], + "type": "image", + "image_path": "e2bc39fdbf5cf91e3443405a50bfc7572b857682d65f6605438fa6c1ee13a9a6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 251, + 482, + 274 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 274 + ], + "type": "text", + "content": "Fig. 1: Objects predicted by our method using no manual annotations. Red boxes are ground truth boxes, yellow boxes are predicted by our network." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 481, + 242 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 481, + 242 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 481, + 242 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 481, + 242 + ], + "type": "image", + "image_path": "7719cef71cd1bfe6ca48e6dbb9bafb55519594dd4b3a49aaf3650b086a01579c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 303, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 518 + ], + "type": "text", + "content": "We introduce a novel and robust self-supervised, easy-to-use method, operating solely on lidar point cloud sequences for easy-to-obtain " + }, + { + "bbox": [ + 130, + 303, + 482, + 518 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 130, + 303, + 482, + 518 + ], + "type": "text", + "content": " annotations. No human-annotated bounding boxes, cameras, costly high-precision GPS, or tedious sensor rig calibration is required, which could be especially useful in robotics research where data is often limited (even KITTI [11] only has forward facing cameras) or maintaining a multi-sensor-setup including its calibration is too cumbersome and expensive. Our method trains and runs a self-supervised lidar flow estimator [3] under the hood in order to create motion cues. We show in our experiments that these motion cues are a key factor to the success of our method. Based on the estimated lidar flow we bootstrap initial pseudo ground truth using simple clustering and track optimization. With these mined bounding boxes of moving objects we initialize a self-supervised, trajectory-regularized variant of self-training [1] (which is semi-supervised in its original form): We train a first version of a SOTA object detector, then iteratively re-generate and trajectory-regularize the pseudo ground truth and retrain the detector. Since the single-frame object detector has no concept of motion, it generalizes to detect any movable object on the way. Exemplary output of the trained detector, 3D boxes in single-frame point clouds, is depicted in Fig. 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 545 + ], + "type": "text", + "content": "Our contribution is a novel self-supervised trajectory-regularized self-training framework for single-frame 3D object detection with the following properties:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 563, + 479, + 662 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 137, + 563, + 479, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 563, + 479, + 587 + ], + "spans": [ + { + "bbox": [ + 137, + 563, + 479, + 587 + ], + "type": "text", + "content": "- It is based entirely on lidar, i.e. without the limitations of prior works: no cameras, no calibration, no high-precision GPS, no manual annotations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 591, + 479, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 591, + 479, + 624 + ], + "spans": [ + { + "bbox": [ + 137, + 591, + 479, + 624 + ], + "type": "text", + "content": "- It is agnostic w.r.t. the sensor model, mounting pose, and detectors architecture and works with the same set of hyper parameters: We demonstrate this across four different datasets and different SOTA detector networks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 629, + 479, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 629, + 479, + 662 + ], + "spans": [ + { + "bbox": [ + 137, + 629, + 479, + 662 + ], + "type": "text", + "content": "- It is able to generalize from moving objects (motion cues) to movable objects (final detection results) and significantly outperforms SOTA methods while being simpler." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "S. Baur et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "We show that using motion cues together with trajectory-regularized self-training is key to this success. The code of this approach will be published for easy use and comparison by other researchers." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 188 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 188 + ], + "type": "text", + "content": "The paper is organized as follows: Sec. 2 discusses related work before the proposed method is described in detail in Sec. 3. Extensive experiments in Sec. 4 show the performance of our method before we conclude in Sec. 5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 206, + 237, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 206, + 237, + 219 + ], + "spans": [ + { + "bbox": [ + 131, + 206, + 237, + 219 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 232, + 365, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 232, + 365, + 244 + ], + "spans": [ + { + "bbox": [ + 131, + 232, + 365, + 244 + ], + "type": "text", + "content": "2.1 Single Frame Lidar 3D Object Detection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 253, + 480, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 253, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 480, + 373 + ], + "type": "text", + "content": "Object detectors operating on 3D point clouds are an active research field. The currently best-performing ones are using deep neural networks trained via supervised learning and can be categorized by their internal representation: Some networks operate on points directly like PointRCNN [21], 3DSSD [35], and IASSD [39]. Others project the points either to a virtual range image [13,14,25] or into a voxel representation like VoxelNet [40], PointPillars, CenterPoint [36], and Transfusion-L [2]. However, all aforementioned methods require large human-annotated datasets in order to perform well and obtaining such annotations is very expensive. We address this problem in this paper, enabling training of SOTA object detectors using pseudo ground truth." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 391, + 340, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 391, + 340, + 403 + ], + "spans": [ + { + "bbox": [ + 131, + 391, + 340, + 403 + ], + "type": "text", + "content": "2.2 Object Distilling from Motion Cues" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 411, + 480, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 480, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 480, + 530 + ], + "type": "text", + "content": "Multiple approaches have been suggested which leverage motion cues from lidar frame pairs in order to detect moving objects. Using the assumption of local geometric constancy, they decompose dynamic scenes into separate moving entities by applying as-rigid-as-possible optimization. Examples of these are the works by Dewan et al. [7], RSF [6] and OGC [23]. The two former methods are optimization-based whereas OGC uses these constraints as loss-function to self-supervise a segmentation network. Although these methods avoid the need for expensive labels, in contrast to our method they tend not to work well in low-resolution areas, they are typically slow, and can only detect moving but not movable (i.e. static but potentially moving) objects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 548, + 380, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 380, + 561 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 380, + 561 + ], + "type": "text", + "content": "2.3 Pseudo Ground Truth for Object Detection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": "Different approaches have been proposed to mine pseudo ground truth for training object detectors: Najibi et al. [15] and Seidenschwarz et al. [19] use motion cues similar to section 2.2 in order to distill moving objects as pseudo ground truth and use it to train an object detector. [15] runs optimization for each frame pair to obtain lidar scene flow, clusters points, fits boxes, tracks them using a Kalman Filter and finally refines them on ICP-registered point clouds. [19] optimizes a clustering algorithm on motion cues through message passing which they then apply to segment point clouds into a set of instances and subsequently" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "text", + "content": "LISO: Lidar-only Self-Supervised 3D Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": "fit boxes. As shown in [19], and in contrast to our approach, both methods suffer from a large performance gap between moving and movable objects. We demonstrate that our method does not suffer from this gap." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 295 + ], + "type": "text", + "content": "[12], [22] were the first to iteratively apply a detection, tracking, retraining paradigm to autonomous driving (AD) data. They build upon consistency constraints between object detectors which they train in the lidar and camera domain, making use of an optical flow network which is trained using supervision. The approach requires a calibrated sensor rig with possibly large camera coverage, IMU, and precision GPS. Similarly, [30] uses video sequences together with lidar scene flow to jointly train a camera and a lidar object detector. MODEST [33, 37] does not require coverage by calibrated cameras, but instead adds the additional requirement to having multiple lidar recordings of the same location in order to identify objects that vanished over time. With such demanding requirements, these approaches are not easy to use and are partially unsuited for popular AD datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 464 + ], + "type": "text", + "content": "Oyster [38] is the approach most similar to ours. It uses DBSCAN [9] on lidar point clouds to initialize pseudo ground truth. Clusters are then tracked using forward-and-reverse tracking in sensor coordinates (i.e. without consideration for ego motion), using a complex policy for confidence-based track retention. After training an object detector on close range data, they employ zero-shot generalization to the far range data, track again and iteratively retrain the detector. In their experiments they use different hyperparameters for different datasets. Our method, in contrast, explicitly considers sensor motion, produces much cleaner initial proposals by leveraging a self-supervised lidar scene flow network, does not require zero-shot-generalization from near-to-far-field, and works with current SOTA object detectors. We show on multiple datasets that our method outperforms [38], that it works robustly with the same set of hyper-parameters and also generalizes well to detect movable objects. Additionally, we will release our code to facilitate further research on this topic." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 479, + 202, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 202, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 202, + 491 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 501, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 586 + ], + "type": "text", + "content": "A general overview of our method is sketched in Fig. 2 and some steps are illustrated in Fig. 3. As input, we take raw (unlabeled) point cloud sequences and undergo three stages, all of which are detailed in the following: Preprocessing of point clouds and lidar scene flow computation (Sec. 3.1), initial pseudo ground truth generation (Sec. 3.2) and repeated training with pseudo ground truth refinement (Sec. 3.3). The final output of the method is a trained object detector which can detect movable objects in raw single-frame point clouds." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 600, + 231, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 231, + 613 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 231, + 613 + ], + "type": "text", + "content": "3.1 Preprocessing" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 365, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 365, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 365, + 628 + ], + "type": "text", + "content": "We preprocess the raw input point clouds as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Ground Removal: First, we remove distracting ground points from each single point cloud using JCP [20], which is a simple, robust, yet effective algorithm to remove ground points using changes in observed height above ground." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "S. Baur et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 473, + 205 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 473, + 205 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 473, + 205 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 473, + 205 + ], + "type": "image", + "image_path": "a2419607cf4bada2dba4040260d1595d7e5c5598b8a839ccf64da75de44c50fc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 217, + 482, + 251 + ], + "lines": [ + { + "bbox": [ + 131, + 217, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 131, + 217, + 482, + 251 + ], + "type": "text", + "content": "Fig. 2: Overview of the proposed method. Point cloud sequences are preprocessed (blue, Sec 3.1), initial pseudo ground truth is created (orange, Sec. 3.2) and the object detector is iteratively trained and pseudo ground truth regenerated (red, Sec. 3.2)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 275, + 479, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 479, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 479, + 322 + ], + "type": "text", + "content": "Ego Motion Estimation: Second, we compute ego-motion between neighboring frame pairs using KISS-ICP [28], which is based on a robust version of ICP. The output is a cm-level accurate transformation " + }, + { + "bbox": [ + 130, + 275, + 479, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{ego}}^{t\\rightarrow t + 1}\\in \\mathbb{R}^{4\\times 4}" + }, + { + "bbox": [ + 130, + 275, + 479, + 322 + ], + "type": "text", + "content": ", describing the ego vehicle position at time " + }, + { + "bbox": [ + 130, + 275, + 479, + 322 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 130, + 275, + 479, + 322 + ], + "type": "text", + "content": " represented in the ego frame from time " + }, + { + "bbox": [ + 130, + 275, + 479, + 322 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 275, + 479, + 322 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 323, + 479, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 479, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 479, + 395 + ], + "type": "text", + "content": "Lidar Scene Flow Estimation: Third, we compute lidar scene flow between neighboring frame pairs resulting in a flow vector " + }, + { + "bbox": [ + 130, + 323, + 479, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i = (dx, dy, dz)" + }, + { + "bbox": [ + 130, + 323, + 479, + 395 + ], + "type": "text", + "content": " for every point " + }, + { + "bbox": [ + 130, + 323, + 479, + 395 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 323, + 479, + 395 + ], + "type": "text", + "content": " in the first point cloud " + }, + { + "bbox": [ + 130, + 323, + 479, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^t" + }, + { + "bbox": [ + 130, + 323, + 479, + 395 + ], + "type": "text", + "content": ". We chose to use SLIM [3] as its code is readily available, it is easy to use, features fast inference, and produces SOTA results. The network is trained self-supervised on raw point cloud sequences, minimizing a k-nearest-neighbor loss between forward and time-reversed point clouds." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 396, + 480, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 480, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 480, + 431 + ], + "type": "text", + "content": "The components of our preprocessing steps (Ground Removal, Ego Motion Estimation, Lidar Scene Flow Estimation) have been selected for robustness and are all used with their default parameters from their respective publications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 452, + 367, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 452, + 367, + 464 + ], + "spans": [ + { + "bbox": [ + 131, + 452, + 367, + 464 + ], + "type": "text", + "content": "3.2 Initial Pseudo Ground Truth Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 475, + 480, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 480, + 523 + ], + "type": "text", + "content": "The aim of our method is to mine pseudo ground truth for training a 3D lidar object detector. For a single point cloud " + }, + { + "bbox": [ + 130, + 475, + 480, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^t" + }, + { + "bbox": [ + 130, + 475, + 480, + 523 + ], + "type": "text", + "content": " (consisting of " + }, + { + "bbox": [ + 130, + 475, + 480, + 523 + ], + "type": "inline_equation", + "content": "n \\in \\mathbb{N}" + }, + { + "bbox": [ + 130, + 475, + 480, + 523 + ], + "type": "text", + "content": " points " + }, + { + "bbox": [ + 130, + 475, + 480, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 130, + 475, + 480, + 523 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 475, + 480, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i \\in \\mathcal{P}^t" + }, + { + "bbox": [ + 130, + 475, + 480, + 523 + ], + "type": "text", + "content": ") this ground truth is a set of 3D bounding boxes " + }, + { + "bbox": [ + 130, + 475, + 480, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 475, + 480, + 523 + ], + "type": "text", + "content": " with confidences representing the objects at time " + }, + { + "bbox": [ + 130, + 475, + 480, + 523 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 475, + 480, + 523 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 214, + 533, + 481, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 533, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 214, + 533, + 481, + 548 + ], + "type": "interline_equation", + "content": "\\mathcal {B} ^ {t} = \\left\\{\\mathrm {B} _ {j} ^ {t}, j \\in \\mathbb {N} \\right\\} = \\left\\{(x, y, z, l, w, h, \\theta , c) _ {j} \\right\\} \\tag {1}", + "image_path": "537f42a737bb4ce9e04c4375c4e417f9d94c6aad3e30079159cf28e3a35ee026.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 556, + 479, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 479, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 479, + 592 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 556, + 479, + 592 + ], + "type": "inline_equation", + "content": "(x,y,z) = \\mathbf{x}" + }, + { + "bbox": [ + 130, + 556, + 479, + 592 + ], + "type": "text", + "content": " define the center position, " + }, + { + "bbox": [ + 130, + 556, + 479, + 592 + ], + "type": "inline_equation", + "content": "(l,w,h)" + }, + { + "bbox": [ + 130, + 556, + 479, + 592 + ], + "type": "text", + "content": " the length, width and height, " + }, + { + "bbox": [ + 130, + 556, + 479, + 592 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 556, + 479, + 592 + ], + "type": "text", + "content": " the heading (orientation around up axis), and " + }, + { + "bbox": [ + 130, + 556, + 479, + 592 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 556, + 479, + 592 + ], + "type": "text", + "content": " the confidence for a single box." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "The key success factor of our method is to focus on a high precision (and potentially low recall) of the initial set of bounding boxes in order to avoid \"wrong\" objects to negatively influence the object detector. We achieve this by leveraging shape, density, and especially motion cues to robustly identify moving objects solely (see also top-right of Fig. 3). Sec. 3.3 targets at generalizing these to movable objects later on." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 447, + 103 + ], + "type": "text", + "content": "LISO: Lidar-only Self-Supervised 3D Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 306, + 199 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 306, + 199 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 306, + 199 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 306, + 199 + ], + "type": "image", + "image_path": "f3c4d1e935ad2a4bde8085c20d1a5f696b3250cf82c8b899d4479722a9be7957.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 478, + 199 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 478, + 199 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 478, + 199 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 478, + 199 + ], + "type": "image", + "image_path": "3493f65229214257f38198882a27355dc4737ed0976b4ef54c51808fa6e87b19.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 135, + 200, + 306, + 285 + ], + "blocks": [ + { + "bbox": [ + 135, + 200, + 306, + 285 + ], + "lines": [ + { + "bbox": [ + 135, + 200, + 306, + 285 + ], + "spans": [ + { + "bbox": [ + 135, + 200, + 306, + 285 + ], + "type": "image", + "image_path": "97756f71ce6ff3ab58addacbf7064edebf7a1d303c00ab3b9ffe011287d42bbb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 294, + 482, + 383 + ], + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 383 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 383 + ], + "type": "text", + "content": "Fig. 3: Overview over preprocessing, initial pseudo ground truth generation and training with examples. Top left: In the first step, self-supervised lidar scene flow is computed and corrected for vehicle ego-motion. Points are colored by flow direction and magnitude. Top right: In the second step, the scene flow is clustered and bounding boxes are fitted (to the moving objects). Bottom left: In the third step, the network is trained on the pseudo ground truth and is generalizing to static objects also, since it does not have the motion information as input signal. Points are thus colored by laser intensity. Bottom right: Ground truth, for reference." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 200, + 478, + 285 + ], + "blocks": [ + { + "bbox": [ + 307, + 200, + 478, + 285 + ], + "lines": [ + { + "bbox": [ + 307, + 200, + 478, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 200, + 478, + 285 + ], + "type": "image", + "image_path": "5b50df209bfd86b02f7e70be74b96d432d91fc41491b6203b968d88069d886be.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 408, + 482, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 504 + ], + "type": "text", + "content": "Flow Clustering: The points " + }, + { + "bbox": [ + 130, + 408, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i \\in \\mathcal{P}^t" + }, + { + "bbox": [ + 130, + 408, + 482, + 504 + ], + "type": "text", + "content": " in each preprocessed point cloud are clustered based on geometry and motion: All stationary points in a scene should have a flow " + }, + { + "bbox": [ + 130, + 408, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 130, + 408, + 482, + 504 + ], + "type": "text", + "content": " similar to the vehicle's ego-motion, i.e. " + }, + { + "bbox": [ + 130, + 408, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i \\approx \\mathbf{f}_{i,\\mathrm{sta}} = ((\\mathbf{T}_{\\mathrm{ego}}^{t \\rightarrow t + 1})^{-1} - \\mathbf{I}_4) \\cdot \\mathbf{p}_i" + }, + { + "bbox": [ + 130, + 408, + 482, + 504 + ], + "type": "text", + "content": ". The residual flow must then be caused by motion of other actors: " + }, + { + "bbox": [ + 130, + 408, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{i,\\mathrm{dyn}} = \\mathbf{f}_i - \\mathbf{f}_{i,\\mathrm{sta}}" + }, + { + "bbox": [ + 130, + 408, + 482, + 504 + ], + "type": "text", + "content": ". We filter all static points by applying a threshold of " + }, + { + "bbox": [ + 130, + 408, + 482, + 504 + ], + "type": "inline_equation", + "content": "1\\mathrm{m/s}" + }, + { + "bbox": [ + 130, + 408, + 482, + 504 + ], + "type": "text", + "content": " to the residual flow and cluster the remaining points based on their point location and flow vector in 6D using DBSCAN [9] (with parameters " + }, + { + "bbox": [ + 130, + 408, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\varepsilon = 1.0" + }, + { + "bbox": [ + 130, + 408, + 482, + 504 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 408, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\min\\mathrm{Pts} = 5" + }, + { + "bbox": [ + 130, + 408, + 482, + 504 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 505, + 482, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 505, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 505, + 482, + 566 + ], + "type": "text", + "content": "We fit a 3D bounding box " + }, + { + "bbox": [ + 130, + 505, + 482, + 566 + ], + "type": "inline_equation", + "content": "\\mathrm{B}_j^t" + }, + { + "bbox": [ + 130, + 505, + 482, + 566 + ], + "type": "text", + "content": " to each resulting cluster following [37] and discard boxes with " + }, + { + "bbox": [ + 130, + 505, + 482, + 566 + ], + "type": "inline_equation", + "content": "l / w > 4.0" + }, + { + "bbox": [ + 130, + 505, + 482, + 566 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 505, + 482, + 566 + ], + "type": "inline_equation", + "content": "lw < 0.35\\mathrm{m}^2" + }, + { + "bbox": [ + 130, + 505, + 482, + 566 + ], + "type": "text", + "content": ", or " + }, + { + "bbox": [ + 130, + 505, + 482, + 566 + ], + "type": "inline_equation", + "content": "lwh < 0.5\\mathrm{m}^3" + }, + { + "bbox": [ + 130, + 505, + 482, + 566 + ], + "type": "text", + "content": ". The heading " + }, + { + "bbox": [ + 130, + 505, + 482, + 566 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 505, + 482, + 566 + ], + "type": "text", + "content": " is set to match the \"forward-axis\" of the motion, i.e. we orient the boxes along the direction of the residual flow " + }, + { + "bbox": [ + 130, + 505, + 482, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{i,\\mathrm{dyn}}" + }, + { + "bbox": [ + 130, + 505, + 482, + 566 + ], + "type": "text", + "content": ". The confidence " + }, + { + "bbox": [ + 130, + 505, + 482, + 566 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 505, + 482, + 566 + ], + "type": "text", + "content": " is initialized to 1 during clustering, and later replaced by detectors' confidence " + }, + { + "bbox": [ + 130, + 505, + 482, + 566 + ], + "type": "inline_equation", + "content": "cf" + }, + { + "bbox": [ + 130, + 505, + 482, + 566 + ], + "type": "text", + "content": ". Section 3.3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "content": "Tracking: We run a simple flow based tracker: Since we have accurate egomotion available, we can track accurately in a fixed coordinate system w.r.t. the world. First, using the residual flow, we can compute a rigid body transform for each box at " + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "content": " that transforms the proposed box " + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_i^t" + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "content": " forward in time towards its suspected location at " + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{B}}_i^{t + 1}" + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "content": ". The propagated boxes " + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{B}}^t + 1" + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "content": " are matched greedily against the actual boxes " + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^{t + 1}" + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "content": " based on box-center distance to the new detections. Unmatched boxes in " + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^{t + 1}" + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "content": " which are further away than " + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{m}" + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "content": " from propagated boxes spawn new tracklets. Unmatched tracklets from " + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^t" + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "content": " are prop" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "S. Baur et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "agated according to the last observed box motion for up to one time step, after that unmatched tracklets are terminated. We run tracking forward and reverse in time, connecting tracklets from forward and reverse tracking to tracks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 199 + ], + "type": "text", + "content": "The resulting set of tracks is post-filtered: We discard tracks that are shorter than 4 time steps or that have a median box confidence " + }, + { + "bbox": [ + 130, + 152, + 480, + 199 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 152, + 480, + 199 + ], + "type": "text", + "content": " lower than the threshold 0.3. This retains only stable and high-confident tracks and avoids false positives to enter the pseudo ground truth." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 200, + 481, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 481, + 273 + ], + "type": "text", + "content": "Track Optimization: We reduce positional noise of the tracks by minimizing translational jerk on all tracks longer than " + }, + { + "bbox": [ + 130, + 200, + 481, + 273 + ], + "type": "inline_equation", + "content": "3\\mathrm{m}" + }, + { + "bbox": [ + 130, + 200, + 481, + 273 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 130, + 200, + 481, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{obs}}" + }, + { + "bbox": [ + 130, + 200, + 481, + 273 + ], + "type": "text", + "content": " be the sequence of (noisy) observed box center positions " + }, + { + "bbox": [ + 130, + 200, + 481, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 200, + 481, + 273 + ], + "type": "text", + "content": " for consecutive time steps " + }, + { + "bbox": [ + 130, + 200, + 481, + 273 + ], + "type": "inline_equation", + "content": "i\\in \\{1,\\dots,T\\}" + }, + { + "bbox": [ + 130, + 200, + 481, + 273 + ], + "type": "text", + "content": " of a track and their derivative " + }, + { + "bbox": [ + 130, + 200, + 481, + 273 + ], + "type": "inline_equation", + "content": "\\frac{d\\mathbf{x}_i}{dt}\\approx \\frac{\\mathbf{x}_{i + 1} - \\mathbf{x}_i}{\\Delta t}" + }, + { + "bbox": [ + 130, + 200, + 481, + 273 + ], + "type": "text", + "content": ". We compute smoothed track positions " + }, + { + "bbox": [ + 130, + 200, + 481, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{smooth}}" + }, + { + "bbox": [ + 130, + 200, + 481, + 273 + ], + "type": "text", + "content": " by initializing them to " + }, + { + "bbox": [ + 130, + 200, + 481, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{obs}}" + }, + { + "bbox": [ + 130, + 200, + 481, + 273 + ], + "type": "text", + "content": " and minimizing the following loss w.r.t. " + }, + { + "bbox": [ + 130, + 200, + 481, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{smooth}}" + }, + { + "bbox": [ + 130, + 200, + 481, + 273 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 203, + 282, + 481, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 282, + 481, + 313 + ], + "spans": [ + { + "bbox": [ + 203, + 282, + 481, + 313 + ], + "type": "interline_equation", + "content": "L = \\sum_ {i = 1} ^ {T} \\left\\| \\frac {d ^ {3} \\mathbf {x} _ {i , \\text {s m o o t h}}}{d t ^ {3}} \\right\\| _ {2} ^ {2} + \\beta \\left\\| \\mathbf {x} _ {i, \\text {s m o o t h}} - \\mathbf {x} _ {i, \\text {o b s}} \\right\\| _ {2} ^ {2} \\tag {2}", + "image_path": "1b4b59daf125466ec4cb87bc2308028ad877598e6387f16225559a5bc77b8fd9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 324, + 479, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 479, + 396 + ], + "type": "text", + "content": "I.e. we minimize the jerk " + }, + { + "bbox": [ + 130, + 324, + 479, + 396 + ], + "type": "inline_equation", + "content": "\\frac{d^3\\mathbf{x}}{dt^3}" + }, + { + "bbox": [ + 130, + 324, + 479, + 396 + ], + "type": "text", + "content": " and use a quadratic regularizer term " + }, + { + "bbox": [ + 130, + 324, + 479, + 396 + ], + "type": "inline_equation", + "content": "(\\beta = 3)" + }, + { + "bbox": [ + 130, + 324, + 479, + 396 + ], + "type": "text", + "content": " on the positions. This simple optimization goal is better suited for our method than more sophisticated ones like an unrolled bicycle model, as it is both faster and also class-agnostic, i.e. more universal: Using a bicycle model is problematic when optimizing pedestrian tracks, since our method does not allow to distinguish different classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 397, + 480, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 480, + 458 + ], + "type": "text", + "content": "We subsequently align the orientation " + }, + { + "bbox": [ + 130, + 397, + 480, + 458 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 397, + 480, + 458 + ], + "type": "text", + "content": " of each detection in a track to the direction of the smoothed track at its position. Box dimensions " + }, + { + "bbox": [ + 130, + 397, + 480, + 458 + ], + "type": "inline_equation", + "content": "\\{l, w, h\\}" + }, + { + "bbox": [ + 130, + 397, + 480, + 458 + ], + "type": "text", + "content": " of all boxes in a track are adapted to the 90 percentile of observed box dimensions in a track. All hyperparameters related to clustering, tracking and track optimization have been tuned visually on two sample snippets from nuScenes [4]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 476, + 346, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 476, + 346, + 489 + ], + "spans": [ + { + "bbox": [ + 131, + 476, + 346, + 489 + ], + "type": "text", + "content": "3.3 Trajectory-Regularized Self-Training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 479, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 479, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 479, + 557 + ], + "type": "text", + "content": "After having mined initial pseudo ground truth of moving objects with a high precision and low recall, we now aim at iteratively improving our pseudo ground truth by training and using an object detector so that the pseudo ground truth generalizes to movable objects. We achieve this by executing iterative trajectory-regularized self-training, which is composed of the following two steps:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": "Training: We train the target object detection network using the current pseudo ground truth in a supervised training setup. Any single-frame object detection network can be plugged into our pipeline. In our experiments we do not deviate from the basic training setup of our object detection networks. Like any SOTA object detection method, we apply standard augmentation techniques to a point cloud during network training: Random rotation, scaling " + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "inline_equation", + "content": "(\\pm 5\\%)" + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": ", and random translation up to " + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "inline_equation", + "content": "5\\mathrm{m}" + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": " around the origin. Furthermore, we randomly pick 1 to 15 objects from the pseudo ground truth database and insert a random subset of their points at random locations into the scene." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "text", + "content": "LISO: Lidar-only Self-Supervised 3D Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "Pseudo Ground Truth Regeneration: We halt training after a set number of training steps " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "s = 30k" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " in our experiments) and create new, improved pseudo ground truth by running the trained detector in inference mode over all sequences in the training dataset, using the detectors' confidence as box confidence " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": ". We regularize these detections by running the flow-based tracker and track smoothing exactly like we do for the initial pseudo ground truth generation. Every 2nd iteration (i.e. every 60k steps in the experiments), we discard the network weights after regenerating the pseudo ground truth." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 272 + ], + "type": "text", + "content": "Fig. 4 illustrates the effect of our self-training: We see that the pseudo ground truth improves with each re-generation and that dropping network weights allows the network to re-focus on the generalized pseudo ground truth. Our strict requirements towards boxes being added to the pseudo ground truth leads to lower recall compared to the network, which as a result scores higher overall." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 272, + 400, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 272, + 400, + 283 + ], + "spans": [ + { + "bbox": [ + 146, + 272, + 400, + 283 + ], + "type": "text", + "content": "Two aspects are key to making our method perform well:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 291, + 479, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 291, + 479, + 337 + ], + "spans": [ + { + "bbox": [ + 138, + 291, + 479, + 337 + ], + "type": "text", + "content": "- Being restrictive when composing the pseudo ground truth (i.e. using plausible tracks of high-confidence network detections or clusters with significant flow only) avoids adding false positives into the pseudo ground truth and hence avoids that our network increasingly hallucinates with each iteration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 339, + 479, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 339, + 479, + 361 + ], + "spans": [ + { + "bbox": [ + 137, + 339, + 479, + 361 + ], + "type": "text", + "content": "- Not using flow but only single-frame point clouds as input for the detector allows the detector to focus on appearance of objects solely." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 370, + 482, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 482, + 395 + ], + "type": "text", + "content": "These allow our method to generalize from initially mined moving objects to movable objects in the scene." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 412, + 218, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 412, + 218, + 425 + ], + "spans": [ + { + "bbox": [ + 132, + 412, + 218, + 425 + ], + "type": "text", + "content": "4 Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 437, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 485 + ], + "type": "text", + "content": "We evaluate our method on multiple datasets across multiple SOTA networks. SLIM [3], the self-supervised flow network we use, is trained and inferred as in the published version, but Bird's-Eye-View (BEV) range is extended from " + }, + { + "bbox": [ + 130, + 437, + 482, + 485 + ], + "type": "inline_equation", + "content": "70 \\times 70\\mathrm{m}" + }, + { + "bbox": [ + 130, + 437, + 482, + 485 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 437, + 482, + 485 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 130, + 437, + 482, + 485 + ], + "type": "text", + "content": " pixels to " + }, + { + "bbox": [ + 130, + 437, + 482, + 485 + ], + "type": "inline_equation", + "content": "120 \\times 120\\mathrm{m}" + }, + { + "bbox": [ + 130, + 437, + 482, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 437, + 482, + 485 + ], + "type": "inline_equation", + "content": "920 \\times 920" + }, + { + "bbox": [ + 130, + 437, + 482, + 485 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 502, + 269, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 269, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 269, + 514 + ], + "type": "text", + "content": "4.1 Datasets and Metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 617 + ], + "type": "text", + "content": "We evaluate our method on four different AD datasets. For a fair comparison, we compute metrics for movable objects by mapping all animate objects (Cars, Trucks, Trailers, Motorcycles, Cyclists, Pedestrians and other Vehicles) to a single category and discard all inanimate objects (Barrier, Traffic Cone,...) since our method does not predict any class attributes. In Table 2 and Table 3 we also give class-based results for completeness. Class labels for true positives are taken from ground truth. For false positives, the predicted class label is assigned randomly according to the label distribution in the dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Waymo Open Dataset (WOD): [24] is a large, geographically diverse dataset recorded with a proprietary high quality lidar. We evaluate using the protocol of [15, 19], using an area of whole " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "100\\mathrm{m} \\times 40\\mathrm{m}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " BEV grid around the ego vehicle, artificially cropping the predictions of our method down to this reduced area." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "S. Baur et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 306, + 219 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 306, + 219 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 306, + 219 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 306, + 219 + ], + "type": "image", + "image_path": "2aa3199686d64cc933f1da60d195c9e087c79b9baeac1b3243da51ce6ca977e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 220, + 292, + 230 + ], + "lines": [ + { + "bbox": [ + 149, + 220, + 292, + 230 + ], + "spans": [ + { + "bbox": [ + 149, + 220, + 292, + 230 + ], + "type": "text", + "content": "(a) Effect of self-supervised self-learning" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 113, + 479, + 219 + ], + "blocks": [ + { + "bbox": [ + 307, + 113, + 479, + 219 + ], + "lines": [ + { + "bbox": [ + 307, + 113, + 479, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 113, + 479, + 219 + ], + "type": "image", + "image_path": "cb4fb497bafe7662ab370a0bb8aa987f7ca827223f5cfb44096feb9b7233fc70.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 220, + 480, + 230 + ], + "lines": [ + { + "bbox": [ + 308, + 220, + 480, + 230 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 480, + 230 + ], + "type": "text", + "content": "(b) Evaluation of Self-Training Hyperparameters" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 239, + 480, + 370 + ], + "lines": [ + { + "bbox": [ + 130, + 239, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 480, + 370 + ], + "type": "text", + "content": "Fig. 4: Left: Effect of self-supervised self-learning: At the beginning of the training, only moving objects are contained within the pseudo ground truth, which thus scores 0 on static objects. Thanks to the self-training, the network generalizes to movable objects and the score of the pseudo ground truth on static objects starts to increase with every regeneration at each \"X\" in the graph. The pseudo ground truth's performance is measured on a subset of the train set of WOD dataset since our tracking is looking ahead in time, which would be unfair w.r.t. other methods. The network (here: Centerpoint) is evaluated on a small subset of the validation set. Right: Evaluation of Self-Training Hyperparameters: Network (Centerpoint) performance over the course of a training on WOD. " + }, + { + "bbox": [ + 130, + 239, + 480, + 370 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 239, + 480, + 370 + ], + "type": "text", + "content": " is the number of training steps between pseudo ground truth regenerations, " + }, + { + "bbox": [ + 130, + 239, + 480, + 370 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 239, + 480, + 370 + ], + "type": "text", + "content": " is the number of regenerations after which network weights are dropped and re-initialized." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 392, + 480, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 480, + 452 + ], + "type": "text", + "content": "KITTI: [10, 11] is recorded using a Velodyne HDL64 lidar sensor, where some parts have been annotated with 3D object boxes and tracking information in the forward facing camera field of view. A large portion of the published data is only available in raw format without annotations, which our method is able to leverage, due to not requiring any ground truth annotations for training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 452, + 480, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 452, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 480, + 499 + ], + "type": "text", + "content": "Argoverse 2: AV2 [31] is recorded using two stacked Velodyne VLP32 lidar sensors and annotated with 3D object boxes. On AV2 and KITTI, we evaluate both 2D (in BEV space) and 3D box IoU at IoU thresholds of 0.3 and 0.5 in the area of " + }, + { + "bbox": [ + 130, + 452, + 480, + 499 + ], + "type": "inline_equation", + "content": "100 \\times 100\\mathrm{m}" + }, + { + "bbox": [ + 130, + 452, + 480, + 499 + ], + "type": "text", + "content": " around the ego vehicle." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 500, + 480, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 480, + 523 + ], + "type": "text", + "content": "nuScenes: [4] is recorded using a Velodyne VLP32 lidar sensor. We evaluate using the nuScenes protocol." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 540, + 208, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 540, + 208, + 551 + ], + "spans": [ + { + "bbox": [ + 131, + 540, + 208, + 551 + ], + "type": "text", + "content": "4.2 Networks" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 557, + 480, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 480, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 480, + 604 + ], + "type": "text", + "content": "We evaluate our proposed method with two SOTA lidar object detection networks of different architecture: Centerpoint [36] and Transfusion-L [2]. We do not modify the published network architectures or losses and train both networks with a batch size of four." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 640 + ], + "type": "text", + "content": "CenterPoint (CP): [36] is based on 2D convolutions in BEV space. Object centers are represented as heatmap in the BEV, which are then reduced to a sensible amount of boxes using non-maximum-suppression." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "Transfusion-L (TF): [2] is a Transformer based architecture, which is applied after initial encoding of the LiDAR point cloud into a 2D BEV feature tensor." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "text", + "content": "LISO: Lidar-only Self-Supervised 3D Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 164, + 114, + 451, + 291 + ], + "blocks": [ + { + "bbox": [ + 164, + 114, + 451, + 291 + ], + "lines": [ + { + "bbox": [ + 164, + 114, + 451, + 291 + ], + "spans": [ + { + "bbox": [ + 164, + 114, + 451, + 291 + ], + "type": "table", + "html": "
AV2nuScenes
BEV-IOU3D-IOUAP↑AOE↓
AP@0.3↑AP@0.5↑AP@0.3↑AP@0.5↑
Sup.
CP [36]0.7780.7780.7360.5170.4840.560
TF [2]0.7830.6540.7670.6010.6270.501
Unsupervised
DBSCAN [9]0.0540.0260.0200.0030.0080.171
DBSCAN(SF)0.0700.0260.0650.0150.0030.082
DBSCAN(GF)0.1050.0410.0880.0240.0001.000
RSF [6]0.0740.0260.0550.0140.0191.003
Self Train
Oyster [38] †0.3540.245----
Oyster-CP [38]0.3810.2660.1500.0020.0911.514
Oyster-TF [38]0.3400.1980.1820.0160.0931.564
LISO-CP0.4480.3350.3670.1880.1091.062
LISO-TF0.4170.2940.3170.1760.1340.938
", + "image_path": "21d57fce2a0e068601eea715db077e70138226f286444f7a3fed5f102b7c33f2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 130, + 292, + 482, + 359 + ], + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 359 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 359 + ], + "type": "text", + "content": "Table 1: Evaluation on AV2 [31] and nuScenes [4]: CP, TF: network architecture, in the first two rows trained supervised for comparison. †: Metrics as reported in [38]. SF: self-supervised lidar scene flow by SLIM, GF: ground truth lidar scene flow. Note that nuScenes uses a minimum precision and recall threshold of 0.1, and since the recall of GT flow clustering is lower than 0.1, all results are clipped away. For full nuScenes scores see supplementary material." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 394, + 208, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 394, + 208, + 406 + ], + "spans": [ + { + "bbox": [ + 132, + 394, + 208, + 406 + ], + "type": "text", + "content": "4.3Baselines" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 421, + 481, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 481, + 456 + ], + "type": "text", + "content": "Besides the obvious baseline of training the object detection networks supervised from scratch on ground truth from the dataset, we also compare to multiple strong unsupervised baselines. For details, see Sec. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 458, + 481, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 481, + 493 + ], + "type": "text", + "content": "RSF: We include RSF [6] in our evaluation as strong representative of methods doing object distilling from motion cues. We ran experiments using the published code of the authors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 495, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 578 + ], + "type": "text", + "content": "DBSCAN: This algorithm [9] clusters points in the point cloud with similar 3D locations into objects. We additionally evaluate extending the cluster space to 6D by either using SLIM scene flow (SF) or ground-truth scene flow (GT). Thus, the \"DBSCAN(SF)\" baseline essentially corresponds to the raw detections used for our initial pseudo ground truth generation. We use the implementation from [17] with parameter values 1.0 for epsilon and 5 for the minimum number of points for a valid cluster." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 580, + 481, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 580, + 481, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 580, + 481, + 615 + ], + "type": "text", + "content": "SeMoLi: [19] is the most recent baseline. As it showed to outperform [15] as self-supervised object detection method we include only published results of [19] in our evaluation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Oyster: We implemented [38] using PyTorch [16], and apply the proposed framework to the CenterPoint (Oyster-CP) and Transfusion-L (Oyster-TF) to be as close as possible to our method and evaluation. We verify our reimplementation, by reproducing the reported metrics by [38] on AV2, see Table 1." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "S. Baur et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 481, + 272 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 272 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 272 + ], + "type": "table", + "html": "
AP@0.4 ↑AP@0.5 ↑AP@0.7 ↑
MovableMovingStillVehi.Ped.Cycl.MovableMovable
BEV3DBEV3DBEV3DBEVBEVBEVBEVBEV3DBEV3D
Sup.CP [36]0.7650.6840.7210.6240.7350.6570.9120.5130.1340.7190.5780.5100.202
TF [2]0.7460.7230.7140.6680.7330.7100.9180.4570.2160.7110.6690.5320.385
UnsupervisedDBSCAN [9]0.0270.0080.0090.0000.0270.0060.1840.0020.0010.0190.0020.0000.000
DBSCAN(SF)0.0260.0100.0640.0410.0000.0000.0730.0100.0090.0110.0010.0000.000
DBSCAN(GF)0.1140.0710.3180.1200.0000.0000.1130.1110.2400.0520.0230.0020.000
RSF [6]0.0300.0200.0800.0550.0000.0000.1090.0000.0020.0250.0120.0110.000
SeMoLi [19] †-0.195-0.575--------0.018
LISO-CP0.2920.2110.2720.2040.2080.1400.6070.0290.0100.2390.1460.1100.019
Self TrainOyster-CP [38]0.2170.0840.1510.0620.1760.0560.5620.0000.0000.1760.0230.0670.000
Oyster-TF [38]0.1210.0150.0510.0070.0980.0100.4750.0000.0000.0920.0030.0310.000
LISO-CP0.3800.3080.3500.2960.3220.2550.6950.0550.0220.3430.2190.1820.021
LISO-TF0.3270.2080.3490.2450.2330.1260.6690.0240.0120.2770.1360.1160.009
", + "image_path": "b10176f0f853634da26cb8a984a5863514efc834bc28741f62934a47c391cc48.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 272, + 482, + 339 + ], + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 339 + ], + "type": "text", + "content": "Table 2: Evaluation on WOD dataset: We evaluate using the protocol of [15, 19], using an area of whole " + }, + { + "bbox": [ + 130, + 272, + 482, + 339 + ], + "type": "inline_equation", + "content": "100\\mathrm{m}\\times 40\\mathrm{m}" + }, + { + "bbox": [ + 130, + 272, + 482, + 339 + ], + "type": "text", + "content": " BEV grid around the ego vehicle, considering objects that move faster than " + }, + { + "bbox": [ + 130, + 272, + 482, + 339 + ], + "type": "inline_equation", + "content": "1\\mathrm{m/s}" + }, + { + "bbox": [ + 130, + 272, + 482, + 339 + ], + "type": "text", + "content": " to be moving (difficulty level L2). CP, TF: network architecture, in the first two lines trained supervised for comparison. " + }, + { + "bbox": [ + 130, + 272, + 482, + 339 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 272, + 482, + 339 + ], + "type": "text", + "content": ": Results taken from [19]. SF: lidar scene flow by SLIM, GF: ground truth lidar scene flow. For class-specific 3D detection scores see supplementary material." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 374, + 198, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 198, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 198, + 385 + ], + "type": "text", + "content": "4.4 Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 399, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 542 + ], + "type": "text", + "content": "Quantitative Results: On all four datasets we see that our method consistently outperforms all self-supervised baselines in all metrics for movable objects, see Table 1, Table 2, and Table 3. Only SeMoLi [19] beats LISO on moving objects (see Table 2) but significantly drops below LISO's performance on movable objects, hinting to difficulties with generalization. Our method does not suffer from this performance gap and performs nearly equally well on movable and moving objects. We also observe that, despite mostly better performance in the supervised case, Transfusion-L responds less favorable than Centerpoint to the self-supervised training approaches. We suspect that Centerpoint, due to its convolutional architecture and its lack of positional encoding compared to Transfusion-L, is less susceptible to overfitting to where moving objects have been observed and where not to expect them." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 544, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 615 + ], + "type": "text", + "content": "We also observe that the nuScenes dataset seems to be the most challenging for the self-supervised methods, leading to the biggest gap between supervised and unsupervised object detection performance. We attribute this to the sparseness of the lidar sensor with only 32 layers. All self-supervised methods have difficulties estimating the orientation of objects (AOE on nuScenes, " + }, + { + "bbox": [ + 130, + 544, + 482, + 615 + ], + "type": "inline_equation", + "content": "2\\pi" + }, + { + "bbox": [ + 130, + 544, + 482, + 615 + ], + "type": "text", + "content": " periodic, see Table 1), which the IoU-based metrics cannot reveal." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 616, + 235, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 235, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 235, + 628 + ], + "type": "text", + "content": "Qualitative Results:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "In Fig. 5 we compare our method to the best-performing baseline, Oyster, using Centerpoint as network architecture. It can be seen, that our method has higher accuracy and predicts box orientations more accurately." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "text", + "content": "LISO: Lidar-only Self-Supervised 3D Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 479, + 226 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 226 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 226 + ], + "type": "table", + "html": "
Movable (Moving & Still)CarPedestrianCyclist
BEV-IOUAP@0.3↑3D-IOUAP@0.5↑AP@0.3↑AP@0.5↑BEV-IOUAP@0.3↑AP@0.5↑BEV-IOUAP@0.3↑AP@0.5↑BEV-IOUAP@0.3↑AP@0.5↑
CP [36]0.7550.6900.7360.6010.8140.7940.3700.1280.4090.157
TF [2]0.7470.6650.7290.5820.8200.7760.3110.0960.2630.032
DBSCAN [9]0.0230.0020.0100.0000.0260.0050.0000.0000.0640.007
RSF [6]0.0290.0190.0290.0110.0660.0490.0000.0000.1920.043
Oyster-CP [38]0.2350.0980.1140.0000.3270.1350.0000.0000.0000.000
Oyster-TF [38]0.2730.0880.1280.0000.3640.1210.0000.0000.0190.000
LISO-CP0.4460.3300.4190.1590.5200.4110.0970.0190.4450.053
LISO-TF0.3610.2070.2940.0360.4250.2970.0840.0140.3480.003
", + "image_path": "1154efe53c541cfa515a6b0a6f782e57b5ae296d62897ac637c33f7759bf0641.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "text", + "content": "Table 3: Evaluation on KITTI dataset: We evaluate on the forward facing field of view where GT annotations are available, but run inference on the whole " + }, + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "inline_equation", + "content": "100 \\times 100\\mathrm{m}" + }, + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "text", + "content": " BEV grid. Also note that flow annotations are not available for KITTI Object. CP, TF: network architecture, in the first two lines trained supervised for comparison." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 135, + 291, + 306, + 469 + ], + "blocks": [ + { + "bbox": [ + 135, + 291, + 306, + 469 + ], + "lines": [ + { + "bbox": [ + 135, + 291, + 306, + 469 + ], + "spans": [ + { + "bbox": [ + 135, + 291, + 306, + 469 + ], + "type": "image", + "image_path": "933af890fd25eaf0af9a63097052fdb38d181555c5caab7463bbc734aecb6bf8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 475, + 482, + 499 + ], + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 499 + ], + "type": "text", + "content": "Fig. 5: Qualitative Results on WOD. Red boxes are ground truth boxes, yellow boxes are predictions. Left: Oyster-CP Right: LISO-CP" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 291, + 479, + 468 + ], + "blocks": [ + { + "bbox": [ + 307, + 291, + 479, + 468 + ], + "lines": [ + { + "bbox": [ + 307, + 291, + 479, + 468 + ], + "spans": [ + { + "bbox": [ + 307, + 291, + 479, + 468 + ], + "type": "image", + "image_path": "45ee8413268cef33df008a415851f006e812015a85c8c8752cb91ee27de6f2f3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 606 + ], + "type": "text", + "content": "Fig. 6, in contrast, showcases an important failure-case and general problem of all unsupervised methods: ground truth, especially in the WOD dataset, is also annotated in very sparse and distant regions. It is very challenging to generate pseudo ground truth for such regions whereas supervised trainings do not have to generalize from moving to such movable (but very sparse) objects. This may be the main reason for a still noticeable gap in evaluation results between supervised and unsupervised methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 622, + 239, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 239, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 239, + 635 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "We investigated the influence of different choices for the components in our pipeline, see Table 4 and cf. Fig. 2." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "S. Baur et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 227, + 114, + 388, + 232 + ], + "blocks": [ + { + "bbox": [ + 227, + 114, + 388, + 232 + ], + "lines": [ + { + "bbox": [ + 227, + 114, + 388, + 232 + ], + "spans": [ + { + "bbox": [ + 227, + 114, + 388, + 232 + ], + "type": "image", + "image_path": "0b029a796ed8a84455295746dea26f96686598721ca138d91a8e0a4ee81c26f2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 238, + 422, + 251 + ], + "lines": [ + { + "bbox": [ + 190, + 238, + 422, + 251 + ], + "spans": [ + { + "bbox": [ + 190, + 238, + 422, + 251 + ], + "type": "text", + "content": "Fig. 6: Missed ground truth boxes in the WOD dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 274, + 482, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 335 + ], + "type": "text", + "content": "Motion Cues as Clustering Input: When adding motion cues to box clustering for Oyster we observe that detection performance already increases noticeably (comparing #3 to #4, Table 4). This demonstrates that self-supervised motion cues are a key ingredient to get high-quality initial pseudo ground truth as opposed to just using clustered 3D points." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 132, + 354, + 482, + 491 + ], + "blocks": [ + { + "bbox": [ + 132, + 354, + 482, + 491 + ], + "lines": [ + { + "bbox": [ + 132, + 354, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 482, + 491 + ], + "type": "table", + "html": "
TagClusterInputMethodSelf TrainTrackOptim.AP@0.4(BEV)AP@0.4(3D)
#1P, SFno tracker0.1770.086
#2P, SFno tracker✓0.2010.131
#3POyster✓0.2170.084
#4P, SFOyster✓0.2550.104
#5P, SFLISO(K, SF)0.2790.209
#6P, SFLISO(K, SF)✓0.3600.290
#7P, SFLISO(K, SF)✓0.2920.211
#8P, SFLISO(K, SF)✓✓0.3800.308
#9P, SFLISO(G, SF)✓✓0.4110.339
#10P, GFLISO(G, GF)✓✓0.4230.349
", + "image_path": "248c3dcb24860bbab8605b2742807fc7488424cbad3160d034efd7226754dfbe.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 491, + 482, + 525 + ], + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 525 + ], + "type": "text", + "content": "Table 4: Ablation study of different parts of our pipeline on WOD with Centerpoint [36]. Abbreviations: P: points, SF: self-supervised lidar scene flow (SLIM), GF: ground truth lidar scene flow. K: KISS-ICP ego-motion, G: ground truth ego-motion." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": "Motion Cues for Tracking: Our scene-flow-based tracker is the biggest contributing factor to the success of our method: the ablation reveals that leveraging motion cues greatly improves tracking (comparing #4 to #6) and thus ultimately improves self-training, i.e. more accurate tracking allows for much more strictness when matching and filtering tracklets. The Oyster tracker uses a matching threshold of " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "5\\mathrm{m}" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": ", our threshold is only " + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{m}" + }, + { + "bbox": [ + 130, + 557, + 482, + 641 + ], + "type": "text", + "content": ". This strictness leads to higher quality pseudo ground truth." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Effect of Self Training: Even without using any tracker to filter tracklets self training has a benefit (comparing #1 to #2). This is due to the network" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 225, + 90, + 447, + 102 + ], + "type": "text", + "content": "LISO: Lidar-only Self-Supervised 3D Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 481, + 163 + ], + "type": "text", + "content": "generalizing to new instances and confidence thresholding the detections. However, having an additional way to discard implausible detections (by tracking) amplifies the positive effect of self training (comparing #5 to #6 and #7 to #8), as it can better prevent false positives from entering the pseudo ground truth." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 481, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 481, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 481, + 223 + ], + "type": "text", + "content": "Track Optimization: The added benefit of track optimization is more independent of Self Training (comparing going from \\#5 to \\#7 with going from \\#6 to \\#8). Finally we notice, that with the combination of a strict tracker, track optimization and self training, the performance is very close to using ground truth flow and ego-motion (comparing \\#8 to \\#9 and \\#10)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 344 + ], + "type": "text", + "content": "Hyperparameters for Iterative Training: Fig. 4 demonstrates that the performance of our proposed method is relatively consistent across a range of hyperparameters. However, using too few steps between regeneration of pseudo ground truth (" + }, + { + "bbox": [ + 130, + 224, + 482, + 344 + ], + "type": "inline_equation", + "content": "s = 20k" + }, + { + "bbox": [ + 130, + 224, + 482, + 344 + ], + "type": "text", + "content": ") leads to degrading performance, because the network does not have enough time to generalize and stabilize using the pseudo ground truth, which then has negative effects during regeneration of the pseudo ground truth. The experiment reveals that doing multiple iterations of self-learning and dropping network weights to allow the network to adjust is beneficial for a good performance. We selected " + }, + { + "bbox": [ + 130, + 224, + 482, + 344 + ], + "type": "inline_equation", + "content": "r = 2" + }, + { + "bbox": [ + 130, + 224, + 482, + 344 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 224, + 482, + 344 + ], + "type": "inline_equation", + "content": "s = 30k" + }, + { + "bbox": [ + 130, + 224, + 482, + 344 + ], + "type": "text", + "content": " as a good compromise between performance and speed for all other experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 361, + 220, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 220, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 220, + 374 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 387, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 578 + ], + "type": "text", + "content": "We have proposed a novel framework for self-supervised 3D lidar object detection based on self-supervised lidar scene flow. Using simple clustering on lidar scene flow in combination with a novel flow-based tracker, we were able to generate pseudo ground truth with moving objects at high precision. This enabled us to bootstrap SOTA 3D lidar object detectors, without using any human labels. The detector generalizes from detecting moving to detecting movable objects over multiple self-training iterations. Experiments revealed that our trajectory-regularized self-learning, based on our scene-flow-based tracker, is key to the success of our method. We have demonstrated the effectiveness of our approach on two SOTA lidar object detectors as well as four AD datasets. All experiments were conducted with the same set of hyperparameters, demonstrating the robustness and wide applicability of our method. Our method achieves a significant improvement in the state-of-the-art of self-supervised lidar object detection on all four datasets. The biggest limitation of our approach is that our method is not able to distinguish different classes. Future work could be around generating pseudo ground truth for class labels, e.g. based on motion or size characteristics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 597, + 198, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 198, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 198, + 609 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 620, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 136, + 620, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 620, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 136, + 620, + 481, + 643 + ], + "type": "text", + "content": "1. Amini, M., Feofanov, V., Pauletto, L., Devijver, E., Maximov, Y.: Self-training: A survey. CoRR abs/2202.12040 (2022), https://arxiv.org/abs/2202.12040" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 643, + 481, + 665 + ], + "type": "text", + "content": "2. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In:" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "S. Baur et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 149, + 116, + 481, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 481, + 159 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 1080-1089. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.00116, https://doi.org/10.1109/CVPR52688.2022.00116" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 161, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 161, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 137, + 161, + 481, + 225 + ], + "type": "text", + "content": "3. Baur, S.A., Emmerichs, D.J., Moosmann, F., Pinggera, P., Ommer, B., Geiger, A.: SLIM: self-supervised lidar scene flow and motion segmentation. In: 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021. pp. 13106-13116. IEEE (2021). https://doi.org/10.1109/ICCV48922.2021.01288, https://doi.org/10.1109/ICCV48922.2021.01288" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 227, + 481, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 227, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 138, + 227, + 481, + 292 + ], + "type": "text", + "content": "4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. pp. 11618-11628. IEEE (2020). https://doi.org/10.1109/CVPR42600.2020.01164, https://doi.org/10.1109/CVPR42600.2020.01164" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 293, + 481, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 481, + 356 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 481, + 356 + ], + "type": "text", + "content": "5. Chen, Z., Luo, Y., Wang, Z., Baktashmotlagh, M., Huang, Z.: Revisiting domain-adaptive 3d object detection by reliable, diverse and class-balanced pseudolabeling. In: IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023. pp. 3691-3703. IEEE (2023). https://doi.org/10.1109/ICCV51070.2023.00344, https://doi.org/10.1109/ICCV51070.2023.00344" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 358, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 138, + 358, + 481, + 390 + ], + "type": "text", + "content": "6. Deng, D., Zakhor, A.: Rsf: Optimizing rigid scene flow from 3d point clouds without labels. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1277-1286 (January 2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 392, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 481, + 445 + ], + "type": "text", + "content": "7. Dewan, A., Caselitz, T., Tipaldi, G.D., Burgard, W.: Motion-based detection and tracking in 3d lidar scans. In: Kragic, D., Bicchi, A., Luca, A.D. (eds.) 2016 IEEE International Conference on Robotics and Automation, ICRA 2016, Stockholm, Sweden, May 16-21, 2016. pp. 4508-4513. IEEE (2016). https://doi.org/10.1109/ICRA.2016.7487649, https://doi.org/10.1109/ICRA.2016.7487649" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 447, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 481, + 479 + ], + "type": "text", + "content": "8. Eskandar, G.: An empirical study of the generalization ability of lidar 3d object detectors to unseen domains. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23815-23825 (June 2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 480, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 480, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 480, + 481, + 533 + ], + "type": "text", + "content": "9. Ester, M., Kriegel, H., Sander, J., Xu, X.: A density-based algorithm for discovering clusters in large spatial databases with noise. In: Simoudis, E., Han, J., Fayyad, U.M. (eds.) Proceedings of the Second International Conference on Knowledge Discovery and Data Mining (KDD-96), Portland, Oregon, USA. pp. 226-231. AAAI Press (1996), http://www.aaai.org/Library/KDD/1996/kdd96-037.php" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 535, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 535, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 134, + 535, + 480, + 566 + ], + "type": "text", + "content": "10. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The KITTI dataset. Int. J. Robotics Res. 32(11), 1231-1237 (2013). https://doi.org/10.1177/0278364913491297, https://doi.org/10.1177/0278364913491297" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 567, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 567, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 133, + 567, + 481, + 620 + ], + "type": "text", + "content": "11. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the KITTI vision benchmark suite. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition, Providence, RI, USA, June 16-21, 2012. pp. 3354-3361. IEEE Computer Society (2012). https://doi.org/10.1109/CVPR.2012.6248074, https://doi.org/10.1109/CVPR.2012.6248074" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 622, + 481, + 665 + ], + "type": "text", + "content": "12. Harley, A.W., Zuo, Y., Wen, J., Mangal, A., Potdar, S., Chaudhry, R., Fragkiadaki, K.: Track, check, repeat: An EM approach to unsupervised tracking. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 16581-16591. Computer Vision Foundation / IEEE (2021). https:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "text", + "content": "LISO: Lidar-only Self-Supervised 3D Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 149, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 149, + 117, + 480, + 149 + ], + "type": "text", + "content": "//doi.org/10.1109/CVPR46437.2021.01631, https://openaccess.thecvf.com/content/CVPR2021/html/Harley_Track_Check_Repeat_An_EM_Approach_to_Unsupervised_Tracking_CVPR_2021_paper.html" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 227 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 227 + ], + "type": "text", + "content": "13. Liang, Z., Zhang, Z., Zhang, M., Zhao, X., Pu, S.: Rangeioudet: Range image based real-time 3d object detector optimized by intersection over union. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 7140-7149. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.00706, https://openaccess.thecvf.com/content/CVPR2021/html/Liang_RangeIoUDet_Range_Issue-Based.Real-Time_3D_Object_Detector_Optimized_by_CVPR_2021_paper.html" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 227, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 227, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 133, + 227, + 481, + 315 + ], + "type": "text", + "content": "14. Meyer, G.P., Laddha, A., Kee, E., Vallespi-Gonzalez, C., Wellington, C.K.: Lasernet: An efficient probabilistic 3d object detector for autonomous driving. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019. pp. 12677-12686. Computer Vision Foundation / IEEE (2019). https://doi.org/10.1109/CVPR.2019.01296, http://openaccess.thecvf.com/content_CVPR_2019/html/Meyer_LaserNet_An_Efficient_Probabilistic_3D_Object_Detector_for_Autonomous_Driving_CVPR_2019_paper.html" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 315, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 315, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 133, + 315, + 481, + 392 + ], + "type": "text", + "content": "15. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Motion inspired unsupervised perception and prediction in autonomous driving. In: Avidan, S., Brostow, G.J., Cissé, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVIII. Lecture Notes in Computer Science, vol. 13698, pp. 424-443. Springer (2022). https://doi.org/10.1007/978-3-031-19839-7_25, https://doi.org/10.1007/978-3-031-19839-7_25" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 392, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 392, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 133, + 392, + 481, + 500 + ], + "type": "text", + "content": "16. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Köpf, A., Yang, E.Z., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Wallach, H.M., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E.B., Garnett, R. (eds.) Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada. pp. 8024-8035 (2019), https://proceedings.neurips.cc/paper/2019/hash/bdbca288fee7f92f2bfa9f7012727740-AAbstract.html" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 501, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 501, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 133, + 501, + 481, + 544 + ], + "type": "text", + "content": "17. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine learning in Python. Journal of Machine Learning Research 12, 2825–2830 (2011)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 545, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 545, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 133, + 545, + 481, + 599 + ], + "type": "text", + "content": "18. Rist, C.B., Enzweiler, M., Gavrila, D.M.: Cross-sensor deep domain adaptation for lidar detection and segmentation. In: 2019 IEEE Intelligent Vehicles Symposium, IV 2019, Paris, France, June 9-12, 2019. pp. 1535-1542. IEEE (2019). https://doi.org/10.1109/IVS.2019.8814047, https://doi.org/10.1109/IVS.2019.8814047" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 600, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 600, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 133, + 600, + 481, + 622 + ], + "type": "text", + "content": "19. Seidenschwarz, J., Osep, A., Ferroni, F., Lucey, S., Leal-Taixe, L.: Semoli: What moves together belongs together (2024), https://arxiv.org/abs/2402.19463" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 622, + 481, + 665 + ], + "type": "text", + "content": "20. Shen, Z., Liang, H., Lin, L., Wang, Z., Huang, W., Yu, J.: Fast ground segmentation for 3d lidar point cloud based on jump-convolution-process. Remote. Sens. 13(16), 3239 (2021). https://doi.org/10.3390/rs13163239, https://doi.org/10.3390/rs13163239" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "S. Baur et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 193 + ], + "type": "text", + "content": "21. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019. pp. 770-779. Computer Vision Foundation / IEEE (2019). https://doi.org/10.1109/CVPR.2019.00086, http://openaccess.thecvf.com/content_CVPR_2019/html/Shi_PointRCNN_3D Object_Proposal_Generation_and_Detection_From_Point_Cloud_CVPR_2019_paper.html" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 195, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 481, + 239 + ], + "type": "text", + "content": "22. Shin, S., Golodetz, S., Vankadari, M., Zhou, K., Markham, A., Trigoni, N.: Sample, crop, track: Self-supervised mobile 3d object detection for urban driving lidar. CoRR abs/2209.10471 (2022). https://doi.org/10.48550/arXiv.2209.10471, https://doi.org/10.48550/arXiv.2209.10471" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 240, + 481, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 284 + ], + "type": "text", + "content": "23. Song, Z., Yang, B.: OGC: Unsupervised 3d object segmentation from rigid dynamics of point clouds. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=ecNbE00tqBU" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 285, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 373 + ], + "type": "text", + "content": "24. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. pp. 2443-2451. IEEE (2020). https://doi.org/10.1109/CVPR42600.2020.00252, https://doi.org/10.1109/CVPR42600.2020.00252" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 374, + 481, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 451 + ], + "type": "text", + "content": "25. Sun, P., Wang, W., Chai, Y., Elsayed, G., Bewley, A., Zhang, X., Sminchisescu, C., Anguelov, D.: RSN: range sparse net for efficient, accurate lidar 3d object detection. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 5725-5734. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.00567, https://openaccess.thecvf.com/content/CVPR2021/html/Sun_RSN_Range_Sparse_Net_for_Efficient_Accurate_LiDAR_3D_Object_CVPR_2021_paper.html" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 453, + 481, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 496 + ], + "type": "text", + "content": "26. Théodose, R., Denis, D., Chateau, T., Frémont, V., Checchin, P.: A deep learning approach for lidar resolution-agnostic object detection. IEEE Trans. Intell. Transp. Syst. 23(9), 14582-14593 (2022). https://doi.org/10.1109/TITS.2021.3130487, https://doi.org/10.1109/TITS.2021.3130487" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 498, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 542 + ], + "type": "text", + "content": "27. Trauble, B., Pauen, S., Poulin-Dubois, D.: Speed and direction changes induce the perception of animacy in 7-month-old infants. Frontiers in Psychology 5 (2014). https://doi.org/10.3389/fpsyg.2014.01141, https://www.frontiersin.org/articles/10.3389/fpsyg.2014.01141" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 544, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 597 + ], + "type": "text", + "content": "28. Vizzo, I., Guadagnino, T., Mersch, B., Wiesmann, L., Behley, J., Stachniss, C.: KISS-ICP: in defense of point-to-point ICP - simple, accurate, and robust registration if done the right way. IEEE Robotics Autom. Lett. 8(2), 1029–1036 (2023). https://doi.org/10.1109/LRA.2023.3236571, https://doi.org/10.1109/LRA.2023.3236571" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 599, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 666 + ], + "type": "text", + "content": "29. Wang, Y., Chen, X., You, Y., Li, L.E., Hariharan, B., Campbell, M.E., Weinberger, K.Q., Chao, W.: Train in germany, test in the USA: making 3d object detectors generalize. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. pp. 11710-11720. Computer Vision Foundation / IEEE (2020). https://doi.org/10.1109/CVPR42600.2020.01173, https://openaccess.thecvf.com/content_CVPR_2020/html/Wang_" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "text", + "content": "LISO: Lidar-only Self-Supervised 3D Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 148, + 117, + 479, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 117, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 148, + 117, + 479, + 137 + ], + "type": "text", + "content": "Train_in_Germany_Test_in_the_USA_Making_3D_Object_CVPR_2020_paper.html" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 205 + ], + "type": "text", + "content": "30. Wang, Y., Chen, Y., Zhang, Z.: 4d unsupervised object discovery. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022 (2022), http://papers.nips.cc/paper_files/paper/2022/ hash/e7407ab5e89c405d28ff6807ffec594a-Abstraction-Conference.html" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 207, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 481, + 262 + ], + "type": "text", + "content": "31. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., Ramanan, D., Carr, P., Hays, J.: Argov-verse 2: Next generation datasets for self-driving perception and forecasting. In: Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021) (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 263, + 481, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 481, + 306 + ], + "type": "text", + "content": "32. Wozniak, M.K., Hansson, M., Thiel, M., Jensfelt, P.: UADA3D: unsupervised adversarial domain adaptation for 3d object detection with sparse lidar and large domain gaps. CoRR abs/2403.17633 (2024). https://doi.org/10.48550/ARXIV.2403.17633, https://doi.org/10.48550/arXiv.2403.17633" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "text", + "content": "33. Xu, J., Waslander, S.L.: Hypermodest: Self-supervised 3d object detection with confidence score filtering. CoRR abs/2304.14446 (2023). https://doi.org/10.48550/arXiv.2304.14446, https://doi.org/10.48550/arXiv.2304.14446" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 342, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 407 + ], + "type": "text", + "content": "34. Xu, Q., Zhou, Y., Wang, W., Qi, C.R., Anguelov, D.: SPG: unsupervised domain adaptation for 3d object detection via semantic point generation. In: 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021. pp. 15426-15436. IEEE (2021). https://doi.org/10.1109/ICCV48922.2021.01516, https://doi.org/10.1109/ICCV48922.2021.01516" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 409, + 481, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 481, + 483 + ], + "type": "text", + "content": "35. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3dssd: Point-based 3d single stage object detector. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. pp. 11037-11045. Computer Vision Foundation / IEEE (2020). https://doi.org/10.1109/CVPR42600.2020.01105, https://openaccess.thecvf.com/content_CVPR_2020/html/Yang_3DSSD_Point-Based_3D_Single_SStage_Object_Detector_CVPR_2020_paper.html" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 487, + 481, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 552 + ], + "type": "text", + "content": "36. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021. pp. 11784-11793. Computer Vision Foundation / IEEE (2021). https://doi.org/10.1109/CVPR46437.2021.01161, https://openaccess.thecvf.com/content/CVPR2021/html/Yin_Center-Based_3D_Object_Detection_and_Tracking_CVPR_2021_paper.html" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 554, + 481, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 619 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 619 + ], + "type": "text", + "content": "37. You, Y., Luo, K., Phoo, C.P., Chao, W., Sun, W., Hariharan, B., Campbell, M.E., Weinberger, K.Q.: Learning to detect mobile objects from lidar scans without labels. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 1120-1130. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.00120, https://doi.org/10.1109/CVPR52688.2022.00120" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "38. Zhang, L., Yang, A.J., Xiong, Y., Casas, S., Yang, B., Ren, M., Urtasun, R.: Towards unsupervised object detection from lidar point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9317-9328 (June 2023)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "S. Baur et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 249 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 182 + ], + "type": "text", + "content": "39. Zhang, Y., Hu, Q., Xu, G., Ma, Y., Wan, J., Guo, Y.: Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 18931-18940. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.01838, https://doi.org/10.1109/CVPR52688.2022.01838" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 182, + 481, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 249 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 249 + ], + "type": "text", + "content": "40. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018. pp. 4490-4499. IEEE Computer Society (2018). https://doi.org/10.1109/CVPR.2018.00472, http://openaccess.thecvf.com/content_cvpr_2018/html/Zhou_VoxelNet_End-to-End_Learning_CVPR_2018_paper.html" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 225, + 91, + 447, + 102 + ], + "type": "text", + "content": "LISO: Lidar-only Self-Supervised 3D Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LITA_ Language Instructed Temporal-Localization Assistant/94963324-666f-4292-a883-31e4dd7adc61_content_list.json b/2024/LITA_ Language Instructed Temporal-Localization Assistant/94963324-666f-4292-a883-31e4dd7adc61_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..71d736302e2a7c44ffc9273a57a57804b7d47960 --- /dev/null +++ b/2024/LITA_ Language Instructed Temporal-Localization Assistant/94963324-666f-4292-a883-31e4dd7adc61_content_list.json @@ -0,0 +1,1605 @@ +[ + { + "type": "text", + "text": "LITA: Language Instructed Temporal-Localization Assistant", + "text_level": 1, + "bbox": [ + 316, + 141, + 686, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz", + "bbox": [ + 254, + 212, + 750, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "NVIDIA", + "bbox": [ + 470, + 255, + 529, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. There has been tremendous progress in multimodal Large Language Models (LLMs). Recent works have extended these models to video input with promising instruction following capabilities. However, an important missing piece is temporal localization. These models cannot accurately answer the \"When?\" questions. We identify three key aspects that limit their temporal localization capabilities: (i) time representation, (ii) architecture, and (iii) data. We address these shortcomings by proposing Language Instructed Temporal-Localization Assistant (LITA) with the following features: (1) We introduce time tokens that encode timestamps relative to the video length to better represent time in videos. (2) We introduce SlowFast tokens in the architecture to capture temporal information at fine temporal resolution. (3) We emphasize temporal localization data for LITA. In addition to leveraging existing video datasets with timestamps, we propose a new task, Reasoning Temporal Localization (RTL), along with the dataset, ActivityNet-RTL, for learning and evaluating this task. Reasoning temporal localization requires both the reasoning and temporal localization of Video LLMs. LITA demonstrates strong performance on this challenging task, nearly doubling the temporal mean intersection-over-union (mIoU) of baselines. In addition, we show that our emphasis on temporal localization also substantially improves video-based text generation compared to existing Video LLMs, including a $36\\%$ relative improvement of Temporal Understanding. Code is available at: https://github.com/NVlabs/LITA", + "bbox": [ + 261, + 303, + 738, + 622 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Large Language Models $\\cdot$ Video Understanding $\\cdot$ Temporal Localization", + "bbox": [ + 259, + 635, + 740, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 688, + 375, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large language models (LLMs) [3,7,11,28,33,34] have demonstrated impressive instruction following capabilities, and shown that language can be a universal interface for various tasks [7,28]. These models can be further extended to multimodal LLMs to process language and other modalities, such as image, video, and audio [1,24,43].", + "bbox": [ + 212, + 719, + 785, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While most multimodal LLMs focus on images for visual content, several recent works introduce models that specialize in processing videos [20,25,27,39]. These Video LLMs preserve the instruction following capabilities of LLMs and", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ab78a62d5276cf97160c0ecb4374363d597abe3a15cb26e164e018518969e32a.jpg", + "image_caption": [ + "Fig. 1: Example to illustrate our proposed Reasoning Temporal Localization (RTL). Instead of directly querying about an event, questions in RTL require further reasoning to answer. Here, the model needs to compare all activities in the video to find the timestamps of the most energetic activity (i.e., handspring)." + ], + "image_footnote": [ + "A: Her dance is the most energetic between 7.37s and 12.81s when she does a handspring, which is more energetic than standing and lying on the floor." + ], + "bbox": [ + 220, + 143, + 500, + 209 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "allow users to ask various questions about a given video. However, one important missing piece in these Video LLMs is temporal localization. When prompted with the \"When?\" questions, these models cannot accurately localize time periods, and often hallucinate irrelevant information [41]. Temporal localization is an important component that differentiates videos from images, and has been widely studied outside the context of instruction following LLMs [4,10,15,31]. It is thus crucial for Video LLMs to have temporal localization capabilities.", + "bbox": [ + 212, + 310, + 782, + 415 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We identify three key aspects that limit the temporal localization capabilities of existing Video LLMs: time representation, architecture, and data. First, existing models often represent timestamps as plain text (e.g. 01:22 or 142sec). However, given a set of frames, the correct timestamp still depends on the frame rate, which the model does not have access to. This makes learning temporal localization harder. Second, the architecture of existing Video LLMs might not have sufficient temporal resolution to interpolate time information accurately. For example, Video-LLaMA [39] only uniformly samples 8 frames from the entire video, which is insufficient for accurate temporal localization. Finally, temporal localization is largely ignored in the data used by existing Video LLMs. Data with timestamps are only a small subset of video instruction tuning data, and the accuracy of these timestamps is also not verified.", + "bbox": [ + 212, + 416, + 784, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our Approach. We address the aforementioned shortcomings of existing Video LLMs, and propose Language Instructed Temporal-Localization Assistant (LITA): (1) Time Representation: We introduce time tokens to represent relative timestamps and allow Video LLMs to better communicate about time than using plain text. (2) Architecture: We introduce SlowFast tokens to capture temporal information at fine temporal resolution to enable accurate temporal localization. (3) Data: We emphasize temporal localization data for LITA. We propose a new task, Reasoning Temporal Localization (RTL), along with the dataset, ActivityNet-RTL, for learning this task.", + "bbox": [ + 212, + 598, + 784, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The first important design of LITA is to use relative representation for time (e.g. first $10\\%$ of the video) instead of the absolute time representation with plain text (e.g. 01:22). We divide a given video into $T$ equal length chunks, and introduce $T$ time tokens $\\langle 1\\rangle$ to $\\langle \\mathbf{T}\\rangle$ to represent the relative time location in the video. During training and inference, these time tokens can be easily encoded and decoded from plain text timestamps given the length of the video. The start and end timestamps are well-defined by the time tokens given only the input", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "D.-A. Huang et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "video. This is in contrast to plain text timestamps. Without the frame rate, the correct absolute timestamp is ill-defined given just the video frames.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The second important design of LITA is to use densely sampled input frames from videos. It is unlikely to achieve accurate temporal localization with only sparsely sampled frames. The challenge is that the LLM module inside Video LLMs cannot naively process large numbers of frames simultaneously due to context length limitation. Take LLaVA [24] as an example. Each image is converted to 256 tokens, which are then fed into its LLM module as input. If we directly feed 100 frames to the LLM module, then that is $256 \\times 100 = 25600$ tokens, which is already over the max context length for some LLMs [6,34]. Inspired by the SlowFast architecture for videos [12], we instead consider two types of tokens, fast tokens and slow tokens, to address this efficiency issue. We generate fast tokens at a high temporal resolution to provide the temporal information, while keeping the tokens per frame at a low number. On the other hand, we generate slow tokens at a low temporal resolution, which allows us to use a higher number of tokens per frame to provide the spatial information.", + "bbox": [ + 212, + 176, + 785, + 387 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, We emphasize temporal localization data for LITA. We include dense video captioning [16] and event localization [37] in instruction tuning of LITA. These tasks include human annotated timestamps to promote accurate temporal localization. In addition to leveraging existing data and tasks, we further propose a new task, Reasoning Temporal Localization (RTL), along with the dataset, ActivityNet-RTL, for training and evaluating this task. Answers to RTL questions can only be derived by utilizing world knowledge and temporal reasoning. Figure 1 shows an example. To answer the question: \"When does the woman's dance become the most energetic?\" the model needs to first recognize the woman's dance moves in the video, then reason about the most active part, and finally temporally localize the event (i.e. handspring). In addition to the predicted timestamps, we further consider the explanation provided by the model. Thus our new task not only assesses temporal understanding, but also requires strong reasoning capabilities that are unique to LLMs.", + "bbox": [ + 212, + 388, + 785, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the challenging RTL task, LITA doubles baseline's performance for temporal metrics (mIOU, Precision@0.5), while providing much better explanations. In addition to enabling accurate temporal localization, we show that our emphasis on temporal understanding also improves LITA's core Video LLM capabilities. LITA substantially improves all scores on a benchmark for video-based question answering [27]. This includes a $22\\%$ relative improvement for Correctness of Information, and a $36\\%$ relative improvement for Temporal Understanding compared to existing Video LLMs.", + "bbox": [ + 212, + 599, + 785, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 744, + 387, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal Large Language Models. Large language models (LLMs) [7,28] inspire recent works to address multimodal tasks by leveraging LLMs [35]. Some approaches add additional parameters inside LLMs, such as gated cross-attention layers [1,2,19] or adapter layers [40], to adapt it to process multimodal inputs.", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "LITA: Language Instructed Temporal-Localization Assistant", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several works, on the other hand, only use modules, such as projection layers or Q-Formers, to project outputs of visual encoders to the input space of LLMs [8, 24,43]. Recent works further expand multimodal LLM to visual grounding tasks, such as detection [5,23,29] and segmentation [18]. The most related to ours is LISA [18], which extends referring segmentation to reasoning segmentation. We share the same spirit and propose Reasoning Temporal Localization to jointly evaluate reasoning and temporal understanding.", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video Large Language Models. Building on the success of multimodal LLMs, several works extend image LLMs to Video LLMs [20, 25, 27, 39]. These works mainly use the approach of projecting visual tokens to LLMs' input space using projection layers [25, 27] or Q-Formers [20, 39]. While these models show promise in descriptive questions and instructions, they still lack temporal localization capabilities. LITA is designed to address this shortcoming, while also improving downstream video tasks. Concurrent works [13, 21, 22] include improvements like better image-video joint training. The most related concurrent work to ours is VTimeLLM [13], which also aims to address temporal localization of Video LLMs. We further introduce the reasoning aspect to temporal localization.", + "bbox": [ + 212, + 255, + 787, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal Localization in Videos. The goal of temporal localization is to pinpoint activities within untrimmed video sequences on a temporal scale [37]. The target activities can be predefined action classes [9, 15] or events described by natural language [4, 42]. Our goal of video temporal understanding is also related to various video tasks, such as dense video captioning [14, 16, 38] and action segmentation [17, 26, 32]. Models for these temporal tasks can have quite different design. Instruction following Video LLMs like LITA provide a way to unify these frameworks.", + "bbox": [ + 212, + 410, + 787, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Language Instructed Temporal-Localization", + "text_level": 1, + "bbox": [ + 214, + 553, + 671, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LITA enables temporal localization for Video LLMs by: (1) relative time representation with the time tokens, (2) SlowFast tokens to capture temporal information at fine temporal resolution, (3) multi-task training that includes accurate timestamps. We will first introduce the overall architecture and discuss further details of individual components.", + "bbox": [ + 212, + 583, + 787, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Architecture", + "text_level": 1, + "bbox": [ + 214, + 680, + 369, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An overview of LITA is shown in Figure 2. We build on Image LLMs. In particular, we select LLaVA due to its simplicity and effectiveness [24]. Note that LITA does not depend on the specific underlying Image LLM architecture and can be easily adapted to other base architectures.", + "bbox": [ + 212, + 703, + 785, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a video, we first uniformly select $T$ frames and encode each frame into $M$ tokens. $T$ should be large enough to support the desired granularity of temporal localization. $T \\times M$ is a large number that usually cannot be directly processed by the LLM module. Thus, we then use SlowFast pooling to reduce the $T \\times M$ tokens to $T + M$ tokens.", + "bbox": [ + 212, + 765, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "D.-A. Huang et al.", + "bbox": [ + 271, + 114, + 395, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A1: She is dancing from $<2>$ to $<3>$ . A2: Her cloth is black. A3: $<1><2>$ She is standing. $<2><3>$ .", + "bbox": [ + 220, + 148, + 771, + 161 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ff7a7f18c17206da75583469eada7ab33bc207ab2aaf74ea4df304eb4721f20d.jpg", + "image_caption": [ + "Fig. 2: Overview of LITA. The input video frames are first encoded into visual tokens (numbered by frame), which are further processed by two pathways. The Fast Token pathway averages all the tokens in a frame to maintain a high temporal resolution. The Slow Token pathway sparsely samples frames to maintain a larger number of tokens per frame to provide spatial information. Timestamps are converted to time tokens $<1>$ to $$ . This is important for better temporal localization learning. Various video tasks on the right can be converted to natural language visual question answering (Q1-3 and A1-3) to jointly optimize LITA." + ], + "image_footnote": [], + "bbox": [ + 222, + 164, + 772, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The slow and fast tokens are projected by a linear layer and concatenated with the text tokens to use as input to the LLM module. The text tokens (prompt) are processed to convert any referenced timestamps to specialized time tokens ( $<1>$ to $<\\mathrm{T}>$ ). All the input tokens are then jointly processed by the LLM module sequentially. We fine-tune the entire model with our reasoning temporal localization data (Section 4) along with other video tasks, such as dense video captioning and event localization. LITA learns to use time tokens instead of absolute timestamps. For temporal localization, we can then directly ask LITA the \"When\" questions (e.g. \"When is she dancing?\"). LITA would respond with time tokens (e.g. \"She is dancing from $<2>$ to $<3>$ ). which can then be converted to timestamps given the video length.", + "bbox": [ + 212, + 446, + 789, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Time Tokens", + "text_level": 1, + "bbox": [ + 215, + 642, + 370, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use a relative time representation instead of absolute timestamps in LITA. As shown in Figure 2, the LLM module can only see the visual tokens (slow and fast) and the language tokens (text prompt). There is not enough information in this input space for the LLM module to infer the absolute timestamp because the frame rate is not known to the model in advance. A better way is to represent timestamps relative to the video length, thus removing the dependency on the frame rate. We divide the video into $T$ chunks and use $T$ specialized time tokens $<1>$ to $<\\mathsf{T}>$ for timestamps. Given a continuous timestamp $\\tau$ and video length $L$ , $\\tau$ can be easily converted to time token $<\\mathsf{t}>$ , where $t = \\mathrm{round}(\\tau(T-1)/L) + 1$ , and conversely $<\\mathsf{t}>$ can be converted back to $\\tau = L(t-1)/(T-1)$ . While this does introduce discretization error, it greatly simplifies the time representation", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "LITA: Language Instructed Temporal-Localization Assistant", + "bbox": [ + 328, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with LLMs. Relative timestamp is also used in other temporally heavy video tasks, such as dense video captioning [38].", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given this time representation, many video tasks related to temporal localization can be transformed into language instructions and answers. For example, dense video captioning can be achieved by prompting the model with \"Describe the video. Each sentence begins with start and end timestamps.\" (Q3 and A3 in Fig. 2). Standard event localization is also transformed to \"When does X happen?\" (Q1 and A1 in Fig. 2). We also incorporate standard video question answering (Q2 and A2 in Fig. 2). More details are discussed in Section 3.4.", + "bbox": [ + 212, + 176, + 787, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 SlowFast Visual Tokens", + "text_level": 1, + "bbox": [ + 214, + 306, + 455, + 320 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We have discussed how we discretize time in videos into $T$ steps in order to make Video LLMs better at reasoning about time. Still, the visual input should match the temporal resolution $T$ in order to achieve effective temporal processing. Ideally, one would need at least $T$ frames to temporally localize events with the resolution $T$ . However, naively feeding all $T$ frames into the LLM module could be computationally prohibitive. In our experiment, we use $T = 100$ and $M = 256$ (CLIP ViT-L-14 [30]). This leads to 25600 tokens per video.", + "bbox": [ + 212, + 330, + 787, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inspired by SlowFast models [12], we consider two pathways to pool the $T \\times M$ tokens for $T$ frames. The first is densely sampled fast tokens to provide temporal information. We obtain $T$ fast tokens from $T$ frames by averaging all the tokens belonging to the same frame. The second is the sparsely sampled slow tokens to maintain better spatial information. We select a spatial downsampling ratio of $s$ , and uniformly select $s^2$ frames from the video. For each selected frame, we perform a $s \\times s$ spatial average pooling to the $M$ tokens, which lead to $\\frac{M}{s^2}$ slow tokens per frame. This leads to a total $M = \\frac{M}{s^2} \\times s^2$ slow tokens. We use $s = 2$ in our experiments. This leads to a total of $T + M$ tokens to represent a video instead of $T \\times M$ tokens.", + "bbox": [ + 212, + 438, + 787, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 Training Tasks", + "text_level": 1, + "bbox": [ + 214, + 613, + 383, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to architecture, training tasks and data also play an important role for LLMs. We emphasize temporal localization data and train LITA with the following five tasks: (1) dense video captioning [16], (2) event localization [37], (3) video question answering [36], (4) natural language visual question answering [24], and (5) our proposed reasoning temporal localization. Temporal localization is a crucial component for three out of the five tasks (1, 2, and 5).", + "bbox": [ + 212, + 638, + 787, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We now introduce each task in order. The first three tasks are standard video tasks and equip LITA with basic video understanding:", + "bbox": [ + 212, + 729, + 787, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dense Video Captioning. In dense video captioning [16], each video is described by a set of sentences, and each sentence comes with the start and end timestamps of the event. Each sentence in dense video captioning can thus be represented as: SENTENCE. We then sort all sentences by its start time, and directly concatenate all sentences and timestamps. One", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "D.-A. Huang et al.", + "bbox": [ + 271, + 114, + 395, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "example prompt to the model for this task is: \"Provide a detailed description of the given video. Each sentence should begin with the start and end timestamps.\" Other prompts are included in the supplementary materials.", + "bbox": [ + 212, + 146, + 785, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Event Localization. In event localization, the goal is to temporally localize the event described by a sentence. We use a simple answer format: . One example prompt for this task is: \"When does \"SENTENCE\" happen in the video? Answer the question only using start and end timestamps.\"", + "bbox": [ + 212, + 195, + 785, + 257 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video Question Answering. The question answering task is already represented as language instructions. However, answers in existing question answering datasets often consist of a single word or phrase because models for this task might not be able to generate longer text. We follow Liu et al. [23] and append the following prompt to the question: \"Answer the question using a single word or phrase.\" The goal is to provide the context for short answers so that it affects the model's text generation less.", + "bbox": [ + 212, + 258, + 785, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Natural Language Visual Question Answering. Training with the above three tasks gives LITA video understanding capabilities. However, we observe that models trained with only these tasks often provide short answers and lack natural language conversation capabilities. We thus also train LITA with natural language visual question answering or visual instruction tuning datasets [24]. The goal is to improve the natural language conversation of LITA. We find that mixing instruction tuning datasets [24] with standard video tasks improves LITA's conversation quality while maintaining good video understanding.", + "bbox": [ + 212, + 369, + 785, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reasoning Temporal Localization. Finally, we also train LITA with our reasoning temporal localization task (details in Section 4). The answer to a reasoning temporal localization question consists of two parts: timestamps and explanation. We find it challenging for models to simultaneously output both of them without any example. Nevertheless, with some training data, LITA quickly pick up reasoning and temporal localization, and provide both the timestamps and explanation of its reasoning in answers.", + "bbox": [ + 212, + 493, + 785, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Reasoning Temporal Localization", + "text_level": 1, + "bbox": [ + 214, + 623, + 571, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We now discuss further details of the proposed Reasoning Temporal Localization (RTL) task. Standard temporal localization does not fully leverage the potential of Video LLMs. One impressive aspect of LLMs is its reasoning capabilities. LLMs can even answer complex questions that involve multi-step reasoning. Therefore, we propose the RTL task to utilize both of Video LLMs' temporal understanding and reasoning capabilities.", + "bbox": [ + 212, + 655, + 785, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Problem Definition", + "text_level": 1, + "bbox": [ + 214, + 768, + 419, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In reasoning temporal localization, the query is still a \"when\" question that asks about the start and end timestamps of an event. The key difference compared to the standard temporal localization task is that the target event is not directly", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "LITA: Language Instructed Temporal-Localization Assistant", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b5a6e94588981e6b0cd9296e5371fb8db5061a42b43885e657f5144cb7992af1.jpg", + "image_caption": [ + "(a) Temporal Reasoning" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 401, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a395746009371616fd09e92cecf11e28819650b81d679f742a3dc2af14efdbb3.jpg", + "image_caption": [ + "(b) World Knowledge", + "Fig. 3: Examples from our ActivityNet-RTL dataset. RTL questions ask about events that are not explicitly described. The model needs to utilize reasoning or its world knowledge to answer. This is in contrast to standard temporal localization, which directly asks about the event of interest. For example in (c), standard temporal localization might directly ask: \"when does the man sit on the floor?\"" + ], + "image_footnote": [], + "bbox": [ + 408, + 143, + 589, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ec4c2a7996f42e3f5bab8ece02e17acfe4a72fb0c095be728221009526a24d1a.jpg", + "image_caption": [ + "(c) Complex Reasoning" + ], + "image_footnote": [], + "bbox": [ + 596, + 143, + 779, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "described in the question, and can only be inferred by reasoning and using world knowledge of the model. The answer to such a question thus consists of two parts: (1) the start and end timestamps of the target event, and (2) an explanation of the reasoning process the model goes through to derive the timestamps.", + "bbox": [ + 212, + 351, + 787, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Some examples are shown in Figure 3 to better illustrate the idea. The answer format is: [start end] Explanation. In Figure 3(a), the model not only has to localize \"adjust their position to avoid obstacles,\" but also needs to temporally reason about which instance happened earlier in the video. In Figure 3(b), instead of directly asking about \"one-leg row,\" it asks about the workout targeting balance and stability. The model thus needs to utilize its knowledge of what kind of exercises are good for balance and stability. Finally, there are questions that require multi-step reasoning. In Figure 3(c), the question asks about \"the most atypical ways of jump roping,\" which requires the model to understand what is typical and atypical for jump roping, and then temporally find the most atypical time period. A standard temporal localization task, in contrast, would just ask when the man is sitting on the floor.", + "bbox": [ + 212, + 412, + 787, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 ActivityNet-RTL Dataset", + "text_level": 1, + "bbox": [ + 215, + 618, + 472, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The above examples are from ActivityNet-RTL, a dataset curated by us for the Reasoning Temporal Localization (RTL) task. We build our dataset from the ActivityNet Captions dataset [16], which annotates multiple events described by sentences in a video, and all the events are temporally localized with start and end timestamps. Consider the following toy example:", + "bbox": [ + 212, + 643, + 787, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "[00:00-00:10] A woman is standing.", + "bbox": [ + 215, + 726, + 480, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "[00:12-00:30] The woman is dancing.", + "bbox": [ + 218, + 741, + 486, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "[00:32-00:36] The woman is sleeping.", + "bbox": [ + 218, + 755, + 493, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We then use this as context and ask GPT-4 to generate temporal localization questions that require further reasoning to answer. We also ask GPT-4 to simultaneously generate the answer that includes the queried start and end timestamps, along with the explanation about the reasoning process.", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "D.-A. Huang et al.", + "bbox": [ + 271, + 114, + 395, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Using the above toy example, by seeing that the woman has done three activities, one possible reasoning temporal localization question is to ask \"When is she the least active?\" Since sleeping is the least active out of the three activities, the target time period is 00:32-00:36, the period when she is sleeping. This illustrates how GPT-4 can still generate interesting questions without seeing the actual video. We annotate few-shot examples for GPT-4 as in previous works to improve the generation quality [24]. All of our prompts are included in the supplementary materials.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training Set Generation. For our training set, we directly use the results generated by GPT-4 with 10,009 videos from the training split of ActivityNet-Captions. This leads to 33,557 Reasoning Temporal Localization question-answer pairs. By inspecting the GPT generated results, we find that most of the questions are valid temporal localization questions given the context. The main shortcoming is that not all question-answer pairs require reasoning. Sometimes GPT-4 generates questions that directly ask about events that are already described by the dense video captions. However, we do hope that LITA can also answer these standard temporal localization questions correctly using natural language. We thus leave these questions in the training set.", + "bbox": [ + 212, + 273, + 787, + 426 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Set Curation. On the other hand, the evaluation set requires manual efforts otherwise we would end up with many non-reasoning questions. We start from the GPT-4 generated questions using a subset of the ActivityNet-Captions validation set, and manually remove non-reasoning questions. We also double check the timestamps and explanations in the answers. This leads to a total of 229 question-answer pairs for 160 videos.", + "bbox": [ + 212, + 430, + 787, + 521 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Metrics", + "text_level": 1, + "bbox": [ + 215, + 556, + 326, + 571 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We consider three metrics for ActivityNet-RTL: mIOU, Precision@0.5, and GPT-4 Relative Scores. The first two metrics are for temporal localization, and the third metric is to evaluate the explanation capability. mIOU averages the intersection-over-union (IOU) between predicted and groundtruth start and end timestamps. Precision@0.5 measures the percentage of predictions that have over 0.5 IOU. We first average these two metrics per video, and then average over all videos in the evaluation set. This avoids over-weighting videos and time periods with more questions, as some time periods in a video have multiple questions.", + "bbox": [ + 212, + 594, + 815, + 717 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To evaluate the quality of the explanation, we follow the evaluation pipeline of LLaVA [24] and leverage GPT-4 for evaluation. GPT-4 is asked to evaluate the helpfulness, relevance, accuracy, and level of details of the explanations, and give a score from 1 to 10. We ask GPT-4 to evaluate both the predicted and groundtruth explanations, and normalize the score for the prediction by the score of the groundtruth. For this metric, we directly average over all question-answer pairs as the explanations could be quite different even for questions about the same time period in the same video.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "LITA: Language Instructed Temporal-Localization Assistant", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We evaluate LITA with both temporal localization and video tasks that do not involve temporal localization because most existing Video LLMs cannot handle temporal localization. In addition to our proposed Reasoning Temporal Localization, we further evaluate LITA on Video-based Text Generation Performance Benchmarking proposed by Maaz et al. [27]. This provides a holistic evaluation of LITA as a Video LLM and not just for temporal localization.", + "bbox": [ + 212, + 176, + 787, + 268 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 287, + 455, + 304 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Architecture. We uniformly sample 100 frames from a video, and use 100 time tokens $<1>$ to $<100>$ to represent timestamps. We use CLIP-L-14 [30] as the visual encoder, and Vicuna [6] as the LLM module. We follow LLaVA's architecture and train a single linear layer for projection [24]. We use 4 frames for slow tokens and use average pool window $s = 2$ . With 1 fast token per frame, this leads to a total of $100 + \\frac{256}{4} \\times 4 = 356$ tokens per video.", + "bbox": [ + 212, + 313, + 785, + 405 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training Datasets. We discussed training tasks in Section 3.4. We now discuss the training datasets for each task. For dense video captioning and event localization, we use the training splits of ActivityNet-Captions [16] and YouCook2 [42], which combine to around 11k videos. The event localization dataset can be generated from the dense video captioning dataset by using the caption as query and the timestamps as target. For video question answering, we use NExT-QA [36] as it contains more complex questions. For image instruction tuning, we use LLaVA-150K [24]. For reasoning temporal localization, we use the training split of our ActivityNet-RTL, which builds on the training split of ActivityNet-Captions.", + "bbox": [ + 212, + 407, + 785, + 544 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training Setup. For each of the five tasks, we randomly select 100K samples with replacement (total 500K). We then use batch size 128 and learning rate 2e-5 to train for 4k iterations. The training takes around 13 hours for 13B and 9 hours for 7B models using 8 A100 GPUs. The linear projection is initialized with the LLaVA pre-trained weights.", + "bbox": [ + 212, + 547, + 785, + 625 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Reasoning Temporal Localization Evaluation", + "text_level": 1, + "bbox": [ + 215, + 645, + 630, + 660 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We first evaluate on the newly proposed ActivityNet-RTL for reasoning temporal localization. Please refer to Section 4 for dataset details and metrics. We use \"P@0.5\" for Precision@0.5 and \"Score\" for the GPT evaluation score for explanations. Other than variations of our model, we include Video-LLaMA-v2 [39] and Video-ChatGPT [27] for comparison. We observe that most of their outputs on ActivityNet-RTL omit any timestamps and thus mIOU and Precision@0.5 become absolute. Therefore, for these methods we only evaluate the GPT-Score. In addition, we ablate the following variations for LITA, all of which are trained with the same five training tasks as LITA:", + "bbox": [ + 212, + 670, + 785, + 805 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "- \"Slow Tokens Only\" samples 4 frames from the video, and computes 64 tokens per frame. It does not use the fast tokens, and it also does not use time tokens", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "D.-A. Huang et al.", + "bbox": [ + 271, + 114, + 395, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/671dafffd172072992a73a62858e7d26c5e3e56edbebacb250a44d6a6ef89a0d.jpg", + "image_caption": [ + "Fig. 4: Qualitative results on ActivityNet-RTL. Overall, LITA not only more accurately localize events in the video, but also provides sensible explanations with more details. In the first example, LITA correctly identifies the second arm wrestling. In the second example, LITA provides further details that they are roasting marshmallows. In the third example, LITA impressively recognizes that the girl \"falls off the beam but gets back\" and explains that this shows resilience. These are in contrast to the more generic answers by Video-LLaMA-v2." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 405, + 250 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/78817c65ea273a07790a5faa48d6a4ae01e26f5478c71e1d54a228012f2ce620.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 143, + 589, + 250 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9cdde68de66a7f83d7c090c040babb37c083211a4c5360c40b6eb435760b811e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 143, + 784, + 250 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b82d890e9e958b3d4e8f6e16d9f59f76e3fd519c9189aeab918ab434815a0046.jpg", + "table_caption": [ + "Table 1: Results on ActivityNet-RTL. LITA substantially outperforms all baselines for all metrics. This shows the importance of our design choices. Interestingly the temporal localization accuracy also improves as we scale the model from 7B to 13B." + ], + "table_footnote": [], + "table_body": "
ModelSizemIOUP@0.5Score
LITA-7B7B24.121.244.0
Video-LLaMA-v2 [39]13B--32.1
Video-ChatGPT [27]13B--38.8
Slow Tokens Only13B14.611.832.2
SlowFast Tokens13B17.514.534.1
LITA-13B13B28.625.946.3
", + "bbox": [ + 344, + 435, + 660, + 555 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "for relative timestamps. This can be seen as naively implementing a Video LLM via the LLaVA architecture.", + "bbox": [ + 212, + 590, + 785, + 621 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "- \"SlowFast Tokens\" additionally includes fast tokens (i.e. 1 token per frame) compared to \"Slow Tokens Only\". This improves the architectural design for representing video, and should allow better temporal processing.", + "bbox": [ + 212, + 628, + 785, + 675 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "- \"LITA\" is our full model that further includes time tokens to better represent timestamps compared to \"SlowFast Tokens.\" We consider two model sizes, 7B and 13B, for LITA to understand the effect of different model sizes.", + "bbox": [ + 212, + 681, + 785, + 726 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Importance of Our Model Components. Results on ActivityNet-RTL are shown in Table 1. All metrics are averaged over three trials. LITA substantially outperforms all baselines for all metrics, and almost double mIOU and P@0.5 when compared to \"Slow Tokens Only\", which is considered as a naive extension of Image LLMs to Video LLMs. \"Score\" is assisted by GPT-4 to evaluate the quality of the explanation provided by the models. While we prompt GPT-4 to ignore the timestamps mentioned in the explanations, we observe that the scores", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "LITA: Language Instructed Temporal-Localization Assistant", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ad2fc7048d7557098aa757fbd615e4741d000bc623e3543b7b4108d63e895f2d.jpg", + "table_caption": [ + "Table 2: Video-based Text Generation Benchmarking results. LITA significantly outperforms existing Video LLMs including Video-LLaMA-v2 [39] and VideoChatGPT [27] on all evaluated aspects. This shows that LITA not only enables accurate temporal localization, but also generally improves video understanding for Video LLMs." + ], + "table_footnote": [], + "table_body": "
ModelCorrectnessDetailContextTemporalConsistencyAverage
LLaMA-Adapter2.032.322.301.982.152.16
Video-LLaMA1.962.182.161.821.791.98
Video-LLaMA-v22.362.422.741.832.122.29
VideoChat2.232.502.531.942.242.29
Video-ChatGPT2.402.522.621.982.372.38
LITA2.942.983.432.683.193.04
", + "bbox": [ + 269, + 212, + 733, + 327 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "are still slightly affected by the timestamps. Nevertheless, LITA provides better explanations for its reasoning due to an overall better video understanding.", + "bbox": [ + 212, + 352, + 782, + 383 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "LITA Gives Detailed and Accurate Explanations. Qualitative results are shown in Figure 4. In the first example, LITA correctly localizes when the second arm wrestling happens. On the other hand, Video-LLaMA-v2 does not recognize that there is arm wrestling in the video. In the second example, while both models identify that the correct activity is cooking with fire, LITA provides much more accurate details, including the fact that they are roasting marshmallows, and the start and end time for that. Finally, in the third example, LITA impressively recognizes the event where the girl \"falls off the beam but gets back\" and correctly responds that this shows resilience in her performance. In contrast, Video-LLaMA-v2 gives a generic answer given the \"resilience\" prompt.", + "bbox": [ + 212, + 386, + 784, + 537 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Temporal Localization Scales with LITA Size. One interesting observation is that the temporal localization quality of LITA also improves as we scale up the model from 7B to 13B (Table 1). One might think that scaling the model only improves language understanding, but our result shows that this could also improve temporal reasoning and understanding for Video LLMs.", + "bbox": [ + 212, + 540, + 785, + 617 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Video-Based Generation Evaluation", + "text_level": 1, + "bbox": [ + 214, + 636, + 555, + 651 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In addition to our proposed reasoning temporal localization, we further evaluate LITA on standard evaluation for Video LLMs to better compare with existing Video LLMs. We use the \"Video-based Text Generation Performance Benchmarking\" proposed by Maaz et al. [27]. This benchmark selects videos from ActivityNet [4] and annotates question-answer pairs with natural language. This is in contrast to existing video question answering benchmarks, where the answers are often limited to a single word or phrase. In this benchmark, there are specific questions like: \"What kind of tricks is the man performing while skating?\" or generic questions like \"Can you describe the video in detail?\" For evaluation, the benchmark uses GPT to measure the following aspects of Video LLMs' answers: Correctness of Information, Detail Orientation, Contextual Understanding, Temporal Understanding, and Consistency.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "D.-A. Huang et al.", + "bbox": [ + 271, + 114, + 395, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e28e1f90eaf249d3b5c07ffe849740ad7068884de8fd2342ae5f13f2e64b8f58.jpg", + "table_caption": [ + "Table 3: Analysis of LITA's training tasks on ActivityNet-RTL. \"RTL\" is needed to predict both timestamps and explanations. \"Video\" includes standard video tasks to improve video understanding. \"NLVQA\" further improves reasoning and natural language generation capabilities of LITA." + ], + "table_footnote": [], + "table_body": "
ModelRTL VideoNLVQAmIOUP@0.5Score
RTL Only✓✘✘26.620.9
No NLVQA✓✓✘26.923.5
LITA✓✓✓28.625.9
", + "bbox": [ + 323, + 212, + 679, + 287 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The results are shown in Table 2. We compare LITA with LLaMA-Adapter [40], Video-LLaMA [39], VideoChat [20], and Video-ChatGPT [27]. Video-ChatGPT slightly outperforms other baselines including Video-LLaMA-v2. LITA significantly outperforms these two existing Video LLMs from all aspects. In particular, LITA achieves a $22\\%$ improvement for Correctness of Information (2.94 vs. 2.40) and a $36\\%$ relative improvement for Temporal Understanding (2.68 vs. 1.98). This shows that our emphasis of temporal understanding in training not only enables accurate temporal localization, but also improves the video understanding of LITA. We hypothesize that temporal localization enables the model to learn more details about videos, leading to improved video understanding. A similar observation was made for Image LLMs [5], where joint training with grounding tasks also improved non-grounding text generation.", + "bbox": [ + 212, + 314, + 792, + 496 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Evaluating the Effects of Training Tasks", + "text_level": 1, + "bbox": [ + 214, + 521, + 591, + 537 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We have analyze the effect of our model components in Section 5.2, where we use all of our training tasks. Now we further analyze the effect of these training tasks for our model. We split the five tasks into three groups: RTL, Video, and NLVQA. \"RTL\" only includes the proposed reasoning temporal localization. Without training on our ActivityNet-RTL, the model would not output timestamps for us to evaluate temporal localization in many cases. The second group \"Video\" includes all the standard video tasks: dense video captioning, event localization, and video question answering. Using these video tasks, the model should learn better video understanding. Finally, \"NLVQA\" refers to the natural language visual question answering task to improve LITA's natural language conversation. We refer to training with just RTL as \"RTL Only,\" and training with both RTL and Video but without NLVQA as \"No NLVQA.\" Training with all three and thus all tasks is our proposed LITA.", + "bbox": [ + 212, + 547, + 787, + 744 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results. The results on ActivityNet-RTL are shown in Table 3 and qualitative comparisons are shown in Figure 5. By only training on RTL, \"RTL Only\" does not have enough supervision to learn the task. This is reflected in both timestamp accuracy (P@0.5) and explanation quality (Score). In addition, for non-temporal questions (Q1 and Q2) in Figure 5, the model cannot properly answer and always use the answer format for reasoning temporal localization. By", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "LITA: Language Instructed Temporal-Localization Assistant", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ec5d9a1307ac584ff42a3d6d64e111fddadff4af2fbd406011f09951818c4a94.jpg", + "image_caption": [ + "Fig. 5: Qualitative results to analyze the effect of training tasks. $\\tau$ is the continuous time in seconds. The questions are not from ActivityNet-RTL as we are also interested in questions other than reasoning temporal localization. \"RTL Only\" answers in the format for reasoning temporal localization, as this is the only training data it has. \"No NLVQA\" further includes standard video tasks and can correctly answer non-temporal questions (Q1 and Q2). However, the answers are short due to the short answers in existing video question answering datasets. LITA improves both reasoning and natural language conversation by utilizing NLVQA datasets." + ], + "image_footnote": [], + "bbox": [ + 218, + 152, + 781, + 281 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "adding standard video tasks in Video to training, \"No NLVQA\" improves all metrics compared to \"RTL Only\". Qualitatively in Figure 5, it can also answer Q1 and Q2 correctly. However, this capability mainly comes from the inclusion of video question answering datasets, which leads to short answers. LITA further improves by including NLVQA, which contains complex reasoning questions to improve its reasoning capabilities. More importantly, as shown in Figure 5, LITA now answers questions with natural language instead of short phrases.", + "bbox": [ + 212, + 438, + 787, + 544 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 574, + 359, + 590 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We propose Language Instructed Temporal-Localization Assistant (LITA), which enables accurate temporal localization using Video LLMs. LITA demonstrates promising capabilities to answer complex temporal localization questions. At the same time, LITA substantially improves video-based text generation compared to existing Video LLMs even for non-temporal questions. This is the result of both our model design and data strategy. For model design, we propose time tokens to better represent the time and SlowFast tokens to efficiently process video inputs. Our experiments show the importance of these model components. For data strategy, we emphasize temporal localization data in training LITA. To achieve this, we propose the Reasoning Temporal Localization task and curate the ActivityNet-RTL dataset. Our results show that the inclusion of temporal localization data not only enables temporal localization for Video LLMs, but also improves the general video understanding capabilities. We further analyze our training data and show the benefits of incorporating standard video tasks and image instruction tuning data to train Video LLMs.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "D.-A. Huang et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: NeurIPS (2022)", + "2. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortsman, M., Schmidt, L.: OpenFlamingo: An open-source framework for training large autoregressive vision-language models. arXiv:2308.01390 (2023)", + "3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS (2020)", + "4. Caba Heilbron, F., Escorcia, V., Ghanem, B., Carlos Niebles, J.: Activitynet: A large-scale video benchmark for human activity understanding. In: CVPR (2015)", + "5. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: MiniGPT-v2: large language model as a unified interface for vision-language multi-task learning. arXiv:2310.09478 (2023)", + "6. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality (2023), https://lmsys.org/blog/2023-03-30-vicuna/", + "7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: PaLM: Scaling language modeling with pathways. arXiv:2204.02311 (2022)", + "8. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructblip: Towards general-purpose vision-language models with instruction tuning (2023)", + "9. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: The EPIC-KITCHENS Dataset: Collection, challenges and baselines. IEEE Trans. PAMI 43(11), 4125-4141 (2021)", + "10. Damen, D., Doughty, H., Farinella, G.M., Furnari, A., Ma, J., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. IJCV 130, 33-55 (2022)", + "11. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: NAACL (2019)", + "12. Feichtenhofer, C., Fan, H., Malik, J., He, K.: SlowFast networks for video recognition. In: ICCV (2019)", + "13. Huang, B., Wang, X., Chen, H., Song, Z., Zhu, W.: Vtimellm: Empower llm to grasp video moments. arXiv preprint arXiv:2311.18445 (2023)", + "14. Huang, G., Pang, B., Zhu, Z., Rivera, C., Soricut, R.: Multimodal pretraining for dense video captioning. In: AACL-IJCNLP (2020)", + "15. Jiang, Y.G., Liu, J., Zamir, A.R., Toderici, G., Laptev, I., Shah, M., Sukthankar, R.: Thumos challenge: Action recognition with a large number of classes (2014), https://www.crcv.ucf.edu/THUMOS14/home.html", + "16. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Niebles, J.C.: Dense-captioning events in videos. In: ICCV (2017)", + "17. Kuehne, H., Arslan, A.B., Serre, T.: The language of actions: Recovering the syntax and semantics of goal-directed human activities. In: CVPR (2014)" + ], + "bbox": [ + 220, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "LITA: Language Instructed Temporal-Localization Assistant", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: LISA: Reasoning segmentation via large language model. arXiv:2308.00692 (2023)", + "19. Li, B., Zhang, Y., Chen, L., Wang, J., Pu, F., Yang, J., Li, C., Liu, Z.: MIMIC-IT: Multi-modal in-context instruction tuning. arXiv:2306.05425 (2023)", + "20. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: VideoChat: Chat-centric video understanding. arXiv:2305.06355 (2023)", + "21. Li, Y., Wang, C., Jia, J.: Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043 (2023)", + "22. Lin, B., Zhu, B., Ye, Y., Ning, M., Jin, P., Yuan, L.: Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122 (2023)", + "23. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv:2310.03744 (2023)", + "24. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023)", + "25. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability (2023)", + "26. Miech, A., Zhukov, D., Alayrac, J.B., Tapaswi, M., Laptev, I., Sivic, J.: HowTo100M: Learning a text-video embedding by watching hundred million narrated video clips. In: ICCV (2019)", + "27. Muhammad Maaz, Hanoona Rasheed, S.K., Khan, F.: Video-ChatGPT: Towards detailed video understanding via large vision and language models. ArXiv 2306.05424 (2023)", + "28. OpenAI: GPT-4 technical report. arXiv:2303.08774 (2023)", + "29. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv:2306.14824 (2023)", + "30. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)", + "31. Sigurdsson, G.A., Varol, G., Wang, X., Farhadi, A., Laptev, I., Gupta, A.: Hollywood in homes: Crowdsourcing data collection for activity understanding. In: ECCV (2016)", + "32. Tang, Y., Ding, D., Rao, Y., Zheng, Y., Zhang, D., Zhao, L., Lu, J., Zhou, J.: COIN: A large-scale dataset for comprehensive instructional video analysis. In: CVPR (2019)", + "33. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: LLaMA: Open and efficient foundation language models. arXiv:2302.13971 (2023)", + "34. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: LLaMA 2: Open foundation and fine-tuned chat models. arXiv:2307.09288 (2023)", + "35. Wu, S., Fei, H., Qu, L., Ji, W., Chua, T.S.: NExT-GPT: Any-to-any multimodal LLM. arXiv:2309.05519 (2023)", + "36. Xiao, J., Shang, X., Yao, A., Chua, T.S.: NExT-QA: Next phase of question-answering to explaining temporal actions. In: CVPR (2021)", + "37. Yan, S., Xiong, X., Nagrani, A., Arnab, A., Wang, Z., Ge, W., Ross, D., Schmid, C.: UnLoc: A unified framework for video localization tasks. In: ICCV (2023)", + "38. Yang, A., Nagrani, A., Seo, P.H., Miech, A., Pont-Tuset, J., Laptev, I., Sivic, J., Schmid, C.: Vid2Seq: Large-scale pretraining of a visual language model for dense video captioning. In: CVPR (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "D.-A. Huang et al.", + "bbox": [ + 271, + 114, + 395, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Zhang, H., Li, X., Bing, L.: Video-LLaMA: An instruction-tuned audio-visual language model for video understanding. arXiv:2306.02858 (2023)", + "40. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: LLaMA-Adapter: Efficient fine-tuning of language models with zero-init attention. arXiv:2303.16199 (2023)", + "41. Zhang, Y., Li, Y., Cui, L., Cai, D., Liu, L., Fu, T., Huang, X., Zhao, E., Zhang, Y., Chen, Y., Wang, L., Luu, A.T., Bi, W., Shi, F., Shi, S.: Siren's song in the ai ocean: A survey on hallucination in large language models. arXiv:2309.01219 (2023)", + "42. Zhou, L., Xu, C., Corso, J.J.: Towards automatic learning of procedures from web instructional videos. In: AAAI (2018)", + "43. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: MiniGPT-4: Enhancing vision-language understanding with advanced large language models. arXiv:2304.10592 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 340 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "LITA: Language Instructed Temporal-Localization Assistant", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/LITA_ Language Instructed Temporal-Localization Assistant/94963324-666f-4292-a883-31e4dd7adc61_model.json b/2024/LITA_ Language Instructed Temporal-Localization Assistant/94963324-666f-4292-a883-31e4dd7adc61_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0fdfbac8e5ed905f516c1d227724f4b7ceca9041 --- /dev/null +++ b/2024/LITA_ Language Instructed Temporal-Localization Assistant/94963324-666f-4292-a883-31e4dd7adc61_model.json @@ -0,0 +1,2126 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.318, + 0.142, + 0.687, + 0.187 + ], + "angle": 0, + "content": "LITA: Language Instructed Temporal-Localization Assistant" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.213, + 0.75, + 0.244 + ], + "angle": 0, + "content": "De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz" + }, + { + "type": "text", + "bbox": [ + 0.471, + 0.256, + 0.531, + 0.269 + ], + "angle": 0, + "content": "NVIDIA" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.304, + 0.74, + 0.623 + ], + "angle": 0, + "content": "Abstract. There has been tremendous progress in multimodal Large Language Models (LLMs). Recent works have extended these models to video input with promising instruction following capabilities. However, an important missing piece is temporal localization. These models cannot accurately answer the \"When?\" questions. We identify three key aspects that limit their temporal localization capabilities: (i) time representation, (ii) architecture, and (iii) data. We address these shortcomings by proposing Language Instructed Temporal-Localization Assistant (LITA) with the following features: (1) We introduce time tokens that encode timestamps relative to the video length to better represent time in videos. (2) We introduce SlowFast tokens in the architecture to capture temporal information at fine temporal resolution. (3) We emphasize temporal localization data for LITA. In addition to leveraging existing video datasets with timestamps, we propose a new task, Reasoning Temporal Localization (RTL), along with the dataset, ActivityNet-RTL, for learning and evaluating this task. Reasoning temporal localization requires both the reasoning and temporal localization of Video LLMs. LITA demonstrates strong performance on this challenging task, nearly doubling the temporal mean intersection-over-union (mIoU) of baselines. In addition, we show that our emphasis on temporal localization also substantially improves video-based text generation compared to existing Video LLMs, including a \\(36\\%\\) relative improvement of Temporal Understanding. Code is available at: https://github.com/NVlabs/LITA" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.636, + 0.741, + 0.663 + ], + "angle": 0, + "content": "Keywords: Large Language Models \\(\\cdot\\) Video Understanding \\(\\cdot\\) Temporal Localization" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.689, + 0.377, + 0.705 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.786, + 0.795 + ], + "angle": 0, + "content": "Large language models (LLMs) [3,7,11,28,33,34] have demonstrated impressive instruction following capabilities, and shown that language can be a universal interface for various tasks [7,28]. These models can be further extended to multimodal LLMs to process language and other modalities, such as image, video, and audio [1,24,43]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "While most multimodal LLMs focus on images for visual content, several recent works introduce models that specialize in processing videos [20,25,27,39]. These Video LLMs preserve the instruction following capabilities of LLMs and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.128 + ], + "angle": 0, + "content": "D.-A. Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.145, + 0.501, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.508, + 0.158, + 0.77, + 0.2 + ], + "angle": 0, + "content": "A: Her dance is the most energetic between 7.37s and 12.81s when she does a handspring, which is more energetic than standing and lying on the floor." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.224, + 0.788, + 0.281 + ], + "angle": 0, + "content": "Fig. 1: Example to illustrate our proposed Reasoning Temporal Localization (RTL). Instead of directly querying about an event, questions in RTL require further reasoning to answer. Here, the model needs to compare all activities in the video to find the timestamps of the most energetic activity (i.e., handspring)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.784, + 0.416 + ], + "angle": 0, + "content": "allow users to ask various questions about a given video. However, one important missing piece in these Video LLMs is temporal localization. When prompted with the \"When?\" questions, these models cannot accurately localize time periods, and often hallucinate irrelevant information [41]. Temporal localization is an important component that differentiates videos from images, and has been widely studied outside the context of instruction following LLMs [4,10,15,31]. It is thus crucial for Video LLMs to have temporal localization capabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.785, + 0.597 + ], + "angle": 0, + "content": "We identify three key aspects that limit the temporal localization capabilities of existing Video LLMs: time representation, architecture, and data. First, existing models often represent timestamps as plain text (e.g. 01:22 or 142sec). However, given a set of frames, the correct timestamp still depends on the frame rate, which the model does not have access to. This makes learning temporal localization harder. Second, the architecture of existing Video LLMs might not have sufficient temporal resolution to interpolate time information accurately. For example, Video-LLaMA [39] only uniformly samples 8 frames from the entire video, which is insufficient for accurate temporal localization. Finally, temporal localization is largely ignored in the data used by existing Video LLMs. Data with timestamps are only a small subset of video instruction tuning data, and the accuracy of these timestamps is also not verified." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.785, + 0.734 + ], + "angle": 0, + "content": "Our Approach. We address the aforementioned shortcomings of existing Video LLMs, and propose Language Instructed Temporal-Localization Assistant (LITA): (1) Time Representation: We introduce time tokens to represent relative timestamps and allow Video LLMs to better communicate about time than using plain text. (2) Architecture: We introduce SlowFast tokens to capture temporal information at fine temporal resolution to enable accurate temporal localization. (3) Data: We emphasize temporal localization data for LITA. We propose a new task, Reasoning Temporal Localization (RTL), along with the dataset, ActivityNet-RTL, for learning this task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The first important design of LITA is to use relative representation for time (e.g. first \\(10\\%\\) of the video) instead of the absolute time representation with plain text (e.g. 01:22). We divide a given video into \\(T\\) equal length chunks, and introduce \\(T\\) time tokens \\(\\langle 1\\rangle\\) to \\(\\langle \\mathbf{T}\\rangle\\) to represent the relative time location in the video. During training and inference, these time tokens can be easily encoded and decoded from plain text timestamps given the length of the video. The start and end timestamps are well-defined by the time tokens given only the input" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.329, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LITA: Language Instructed Temporal-Localization Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "video. This is in contrast to plain text timestamps. Without the frame rate, the correct absolute timestamp is ill-defined given just the video frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.388 + ], + "angle": 0, + "content": "The second important design of LITA is to use densely sampled input frames from videos. It is unlikely to achieve accurate temporal localization with only sparsely sampled frames. The challenge is that the LLM module inside Video LLMs cannot naively process large numbers of frames simultaneously due to context length limitation. Take LLaVA [24] as an example. Each image is converted to 256 tokens, which are then fed into its LLM module as input. If we directly feed 100 frames to the LLM module, then that is \\(256 \\times 100 = 25600\\) tokens, which is already over the max context length for some LLMs [6,34]. Inspired by the SlowFast architecture for videos [12], we instead consider two types of tokens, fast tokens and slow tokens, to address this efficiency issue. We generate fast tokens at a high temporal resolution to provide the temporal information, while keeping the tokens per frame at a low number. On the other hand, we generate slow tokens at a low temporal resolution, which allows us to use a higher number of tokens per frame to provide the spatial information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.787, + 0.6 + ], + "angle": 0, + "content": "Finally, We emphasize temporal localization data for LITA. We include dense video captioning [16] and event localization [37] in instruction tuning of LITA. These tasks include human annotated timestamps to promote accurate temporal localization. In addition to leveraging existing data and tasks, we further propose a new task, Reasoning Temporal Localization (RTL), along with the dataset, ActivityNet-RTL, for training and evaluating this task. Answers to RTL questions can only be derived by utilizing world knowledge and temporal reasoning. Figure 1 shows an example. To answer the question: \"When does the woman's dance become the most energetic?\" the model needs to first recognize the woman's dance moves in the video, then reason about the most active part, and finally temporally localize the event (i.e. handspring). In addition to the predicted timestamps, we further consider the explanation provided by the model. Thus our new task not only assesses temporal understanding, but also requires strong reasoning capabilities that are unique to LLMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.787, + 0.722 + ], + "angle": 0, + "content": "For the challenging RTL task, LITA doubles baseline's performance for temporal metrics (mIOU, Precision@0.5), while providing much better explanations. In addition to enabling accurate temporal localization, we show that our emphasis on temporal understanding also improves LITA's core Video LLM capabilities. LITA substantially improves all scores on a benchmark for video-based question answering [27]. This includes a \\(22\\%\\) relative improvement for Correctness of Information, and a \\(36\\%\\) relative improvement for Temporal Understanding compared to existing Video LLMs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.746, + 0.388, + 0.763 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Multimodal Large Language Models. Large language models (LLMs) [7,28] inspire recent works to address multimodal tasks by leveraging LLMs [35]. Some approaches add additional parameters inside LLMs, such as gated cross-attention layers [1,2,19] or adapter layers [40], to adapt it to process multimodal inputs." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.129 + ], + "angle": 0, + "content": "D.-A. Huang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "Several works, on the other hand, only use modules, such as projection layers or Q-Formers, to project outputs of visual encoders to the input space of LLMs [8, 24,43]. Recent works further expand multimodal LLM to visual grounding tasks, such as detection [5,23,29] and segmentation [18]. The most related to ours is LISA [18], which extends referring segmentation to reasoning segmentation. We share the same spirit and propose Reasoning Temporal Localization to jointly evaluate reasoning and temporal understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.256, + 0.788, + 0.408 + ], + "angle": 0, + "content": "Video Large Language Models. Building on the success of multimodal LLMs, several works extend image LLMs to Video LLMs [20, 25, 27, 39]. These works mainly use the approach of projecting visual tokens to LLMs' input space using projection layers [25, 27] or Q-Formers [20, 39]. While these models show promise in descriptive questions and instructions, they still lack temporal localization capabilities. LITA is designed to address this shortcoming, while also improving downstream video tasks. Concurrent works [13, 21, 22] include improvements like better image-video joint training. The most related concurrent work to ours is VTimeLLM [13], which also aims to address temporal localization of Video LLMs. We further introduce the reasoning aspect to temporal localization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.788, + 0.532 + ], + "angle": 0, + "content": "Temporal Localization in Videos. The goal of temporal localization is to pinpoint activities within untrimmed video sequences on a temporal scale [37]. The target activities can be predefined action classes [9, 15] or events described by natural language [4, 42]. Our goal of video temporal understanding is also related to various video tasks, such as dense video captioning [14, 16, 38] and action segmentation [17, 26, 32]. Models for these temporal tasks can have quite different design. Instruction following Video LLMs like LITA provide a way to unify these frameworks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.554, + 0.673, + 0.572 + ], + "angle": 0, + "content": "3 Language Instructed Temporal-Localization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.66 + ], + "angle": 0, + "content": "LITA enables temporal localization for Video LLMs by: (1) relative time representation with the time tokens, (2) SlowFast tokens to capture temporal information at fine temporal resolution, (3) multi-task training that includes accurate timestamps. We will first introduce the overall architecture and discuss further details of individual components." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.681, + 0.37, + 0.696 + ], + "angle": 0, + "content": "3.1 Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.764 + ], + "angle": 0, + "content": "An overview of LITA is shown in Figure 2. We build on Image LLMs. In particular, we select LLaVA due to its simplicity and effectiveness [24]. Note that LITA does not depend on the specific underlying Image LLM architecture and can be easily adapted to other base architectures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Given a video, we first uniformly select \\( T \\) frames and encode each frame into \\( M \\) tokens. \\( T \\) should be large enough to support the desired granularity of temporal localization. \\( T \\times M \\) is a large number that usually cannot be directly processed by the LLM module. Thus, we then use SlowFast pooling to reduce the \\( T \\times M \\) tokens to \\( T + M \\) tokens." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.329, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "LITA: Language Instructed Temporal-Localization Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.15, + 0.772, + 0.162 + ], + "angle": 0, + "content": "A1: She is dancing from \\(<2>\\) to \\(<3>\\). A2: Her cloth is black. A3: \\(<1><2>\\) She is standing. \\(<2><3>\\)." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.165, + 0.774, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.304, + 0.789, + 0.415 + ], + "angle": 0, + "content": "Fig. 2: Overview of LITA. The input video frames are first encoded into visual tokens (numbered by frame), which are further processed by two pathways. The Fast Token pathway averages all the tokens in a frame to maintain a high temporal resolution. The Slow Token pathway sparsely samples frames to maintain a larger number of tokens per frame to provide spatial information. Timestamps are converted to time tokens \\(<1>\\) to \\(\\). This is important for better temporal localization learning. Various video tasks on the right can be converted to natural language visual question answering (Q1-3 and A1-3) to jointly optimize LITA." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.79, + 0.613 + ], + "angle": 0, + "content": "The slow and fast tokens are projected by a linear layer and concatenated with the text tokens to use as input to the LLM module. The text tokens (prompt) are processed to convert any referenced timestamps to specialized time tokens (\\(<1>\\) to \\(<\\mathrm{T}>\\)). All the input tokens are then jointly processed by the LLM module sequentially. We fine-tune the entire model with our reasoning temporal localization data (Section 4) along with other video tasks, such as dense video captioning and event localization. LITA learns to use time tokens instead of absolute timestamps. For temporal localization, we can then directly ask LITA the \"When\" questions (e.g. \"When is she dancing?\"). LITA would respond with time tokens (e.g. \"She is dancing from \\(<2>\\) to \\(<3>\\)). which can then be converted to timestamps given the video length." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.643, + 0.371, + 0.657 + ], + "angle": 0, + "content": "3.2 Time Tokens" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We use a relative time representation instead of absolute timestamps in LITA. As shown in Figure 2, the LLM module can only see the visual tokens (slow and fast) and the language tokens (text prompt). There is not enough information in this input space for the LLM module to infer the absolute timestamp because the frame rate is not known to the model in advance. A better way is to represent timestamps relative to the video length, thus removing the dependency on the frame rate. We divide the video into \\( T \\) chunks and use \\( T \\) specialized time tokens \\( <1> \\) to \\( <\\mathsf{T}> \\) for timestamps. Given a continuous timestamp \\( \\tau \\) and video length \\( L \\), \\( \\tau \\) can be easily converted to time token \\( <\\mathsf{t}> \\), where \\( t = \\mathrm{round}(\\tau(T-1)/L) + 1 \\), and conversely \\( <\\mathsf{t}> \\) can be converted back to \\( \\tau = L(t-1)/(T-1) \\). While this does introduce discretization error, it greatly simplifies the time representation" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.129 + ], + "angle": 0, + "content": "D.-A. Huang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "with LLMs. Relative timestamp is also used in other temporally heavy video tasks, such as dense video captioning [38]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.284 + ], + "angle": 0, + "content": "Given this time representation, many video tasks related to temporal localization can be transformed into language instructions and answers. For example, dense video captioning can be achieved by prompting the model with \"Describe the video. Each sentence begins with start and end timestamps.\" (Q3 and A3 in Fig. 2). Standard event localization is also transformed to \"When does X happen?\" (Q1 and A1 in Fig. 2). We also incorporate standard video question answering (Q2 and A2 in Fig. 2). More details are discussed in Section 3.4." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.307, + 0.457, + 0.321 + ], + "angle": 0, + "content": "3.3 SlowFast Visual Tokens" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.438 + ], + "angle": 0, + "content": "We have discussed how we discretize time in videos into \\( T \\) steps in order to make Video LLMs better at reasoning about time. Still, the visual input should match the temporal resolution \\( T \\) in order to achieve effective temporal processing. Ideally, one would need at least \\( T \\) frames to temporally localize events with the resolution \\( T \\). However, naively feeding all \\( T \\) frames into the LLM module could be computationally prohibitive. In our experiment, we use \\( T = 100 \\) and \\( M = 256 \\) (CLIP ViT-L-14 [30]). This leads to 25600 tokens per video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.591 + ], + "angle": 0, + "content": "Inspired by SlowFast models [12], we consider two pathways to pool the \\( T \\times M \\) tokens for \\( T \\) frames. The first is densely sampled fast tokens to provide temporal information. We obtain \\( T \\) fast tokens from \\( T \\) frames by averaging all the tokens belonging to the same frame. The second is the sparsely sampled slow tokens to maintain better spatial information. We select a spatial downsampling ratio of \\( s \\), and uniformly select \\( s^2 \\) frames from the video. For each selected frame, we perform a \\( s \\times s \\) spatial average pooling to the \\( M \\) tokens, which lead to \\( \\frac{M}{s^2} \\) slow tokens per frame. This leads to a total \\( M = \\frac{M}{s^2} \\times s^2 \\) slow tokens. We use \\( s = 2 \\) in our experiments. This leads to a total of \\( T + M \\) tokens to represent a video instead of \\( T \\times M \\) tokens." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.614, + 0.384, + 0.629 + ], + "angle": 0, + "content": "3.4 Training Tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.788, + 0.731 + ], + "angle": 0, + "content": "In addition to architecture, training tasks and data also play an important role for LLMs. We emphasize temporal localization data and train LITA with the following five tasks: (1) dense video captioning [16], (2) event localization [37], (3) video question answering [36], (4) natural language visual question answering [24], and (5) our proposed reasoning temporal localization. Temporal localization is a crucial component for three out of the five tasks (1, 2, and 5)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.731, + 0.788, + 0.762 + ], + "angle": 0, + "content": "We now introduce each task in order. The first three tasks are standard video tasks and equip LITA with basic video understanding:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Dense Video Captioning. In dense video captioning [16], each video is described by a set of sentences, and each sentence comes with the start and end timestamps of the event. Each sentence in dense video captioning can thus be represented as: SENTENCE. We then sort all sentences by its start time, and directly concatenate all sentences and timestamps. One" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LITA: Language Instructed Temporal-Localization Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.786, + 0.193 + ], + "angle": 0, + "content": "example prompt to the model for this task is: \"Provide a detailed description of the given video. Each sentence should begin with the start and end timestamps.\" Other prompts are included in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.196, + 0.786, + 0.258 + ], + "angle": 0, + "content": "Event Localization. In event localization, the goal is to temporally localize the event described by a sentence. We use a simple answer format: . One example prompt for this task is: \"When does \"SENTENCE\" happen in the video? Answer the question only using start and end timestamps.\"" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.26, + 0.787, + 0.366 + ], + "angle": 0, + "content": "Video Question Answering. The question answering task is already represented as language instructions. However, answers in existing question answering datasets often consist of a single word or phrase because models for this task might not be able to generate longer text. We follow Liu et al. [23] and append the following prompt to the question: \"Answer the question using a single word or phrase.\" The goal is to provide the context for short answers so that it affects the model's text generation less." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.787, + 0.491 + ], + "angle": 0, + "content": "Natural Language Visual Question Answering. Training with the above three tasks gives LITA video understanding capabilities. However, we observe that models trained with only these tasks often provide short answers and lack natural language conversation capabilities. We thus also train LITA with natural language visual question answering or visual instruction tuning datasets [24]. The goal is to improve the natural language conversation of LITA. We find that mixing instruction tuning datasets [24] with standard video tasks improves LITA's conversation quality while maintaining good video understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.787, + 0.601 + ], + "angle": 0, + "content": "Reasoning Temporal Localization. Finally, we also train LITA with our reasoning temporal localization task (details in Section 4). The answer to a reasoning temporal localization question consists of two parts: timestamps and explanation. We find it challenging for models to simultaneously output both of them without any example. Nevertheless, with some training data, LITA quickly pick up reasoning and temporal localization, and provide both the timestamps and explanation of its reasoning in answers." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.624, + 0.573, + 0.642 + ], + "angle": 0, + "content": "4 Reasoning Temporal Localization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.786, + 0.747 + ], + "angle": 0, + "content": "We now discuss further details of the proposed Reasoning Temporal Localization (RTL) task. Standard temporal localization does not fully leverage the potential of Video LLMs. One impressive aspect of LLMs is its reasoning capabilities. LLMs can even answer complex questions that involve multi-step reasoning. Therefore, we propose the RTL task to utilize both of Video LLMs' temporal understanding and reasoning capabilities." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.77, + 0.421, + 0.785 + ], + "angle": 0, + "content": "4.1 Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.786, + 0.841 + ], + "angle": 0, + "content": "In reasoning temporal localization, the query is still a \"when\" question that asks about the start and end timestamps of an event. The key difference compared to the standard temporal localization task is that the target event is not directly" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.129 + ], + "angle": 0, + "content": "D.-A. Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.145, + 0.403, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.228, + 0.357, + 0.237 + ], + "angle": 0, + "content": "(a) Temporal Reasoning" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.145, + 0.591, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.228, + 0.541, + 0.237 + ], + "angle": 0, + "content": "(b) World Knowledge" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.145, + 0.781, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.228, + 0.733, + 0.237 + ], + "angle": 0, + "content": "(c) Complex Reasoning" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.252, + 0.788, + 0.322 + ], + "angle": 0, + "content": "Fig. 3: Examples from our ActivityNet-RTL dataset. RTL questions ask about events that are not explicitly described. The model needs to utilize reasoning or its world knowledge to answer. This is in contrast to standard temporal localization, which directly asks about the event of interest. For example in (c), standard temporal localization might directly ask: \"when does the man sit on the floor?\"" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.413 + ], + "angle": 0, + "content": "described in the question, and can only be inferred by reasoning and using world knowledge of the model. The answer to such a question thus consists of two parts: (1) the start and end timestamps of the target event, and (2) an explanation of the reasoning process the model goes through to derive the timestamps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.789, + 0.595 + ], + "angle": 0, + "content": "Some examples are shown in Figure 3 to better illustrate the idea. The answer format is: [start end] Explanation. In Figure 3(a), the model not only has to localize \"adjust their position to avoid obstacles,\" but also needs to temporally reason about which instance happened earlier in the video. In Figure 3(b), instead of directly asking about \"one-leg row,\" it asks about the workout targeting balance and stability. The model thus needs to utilize its knowledge of what kind of exercises are good for balance and stability. Finally, there are questions that require multi-step reasoning. In Figure 3(c), the question asks about \"the most atypical ways of jump roping,\" which requires the model to understand what is typical and atypical for jump roping, and then temporally find the most atypical time period. A standard temporal localization task, in contrast, would just ask when the man is sitting on the floor." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.619, + 0.473, + 0.634 + ], + "angle": 0, + "content": "4.2 ActivityNet-RTL Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.716 + ], + "angle": 0, + "content": "The above examples are from ActivityNet-RTL, a dataset curated by us for the Reasoning Temporal Localization (RTL) task. We build our dataset from the ActivityNet Captions dataset [16], which annotates multiple events described by sentences in a video, and all the events are temporally localized with start and end timestamps. Consider the following toy example:" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.727, + 0.481, + 0.741 + ], + "angle": 0, + "content": "[00:00-00:10] A woman is standing." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.742, + 0.488, + 0.754 + ], + "angle": 0, + "content": "[00:12-00:30] The woman is dancing." + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.756, + 0.495, + 0.769 + ], + "angle": 0, + "content": "[00:32-00:36] The woman is sleeping." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "We then use this as context and ask GPT-4 to generate temporal localization questions that require further reasoning to answer. We also ask GPT-4 to simultaneously generate the answer that includes the queried start and end timestamps, along with the explanation about the reasoning process." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.329, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LITA: Language Instructed Temporal-Localization Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "Using the above toy example, by seeing that the woman has done three activities, one possible reasoning temporal localization question is to ask \"When is she the least active?\" Since sleeping is the least active out of the three activities, the target time period is 00:32-00:36, the period when she is sleeping. This illustrates how GPT-4 can still generate interesting questions without seeing the actual video. We annotate few-shot examples for GPT-4 as in previous works to improve the generation quality [24]. All of our prompts are included in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.274, + 0.788, + 0.427 + ], + "angle": 0, + "content": "Training Set Generation. For our training set, we directly use the results generated by GPT-4 with 10,009 videos from the training split of ActivityNet-Captions. This leads to 33,557 Reasoning Temporal Localization question-answer pairs. By inspecting the GPT generated results, we find that most of the questions are valid temporal localization questions given the context. The main shortcoming is that not all question-answer pairs require reasoning. Sometimes GPT-4 generates questions that directly ask about events that are already described by the dense video captions. However, we do hope that LITA can also answer these standard temporal localization questions correctly using natural language. We thus leave these questions in the training set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Evaluation Set Curation. On the other hand, the evaluation set requires manual efforts otherwise we would end up with many non-reasoning questions. We start from the GPT-4 generated questions using a subset of the ActivityNet-Captions validation set, and manually remove non-reasoning questions. We also double check the timestamps and explanations in the answers. This leads to a total of 229 question-answer pairs for 160 videos." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.558, + 0.327, + 0.572 + ], + "angle": 0, + "content": "4.3 Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.816, + 0.718 + ], + "angle": 0, + "content": "We consider three metrics for ActivityNet-RTL: mIOU, Precision@0.5, and GPT-4 Relative Scores. The first two metrics are for temporal localization, and the third metric is to evaluate the explanation capability. mIOU averages the intersection-over-union (IOU) between predicted and groundtruth start and end timestamps. Precision@0.5 measures the percentage of predictions that have over 0.5 IOU. We first average these two metrics per video, and then average over all videos in the evaluation set. This avoids over-weighting videos and time periods with more questions, as some time periods in a video have multiple questions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To evaluate the quality of the explanation, we follow the evaluation pipeline of LLaVA [24] and leverage GPT-4 for evaluation. GPT-4 is asked to evaluate the helpfulness, relevance, accuracy, and level of details of the explanations, and give a score from 1 to 10. We ask GPT-4 to evaluate both the predicted and groundtruth explanations, and normalize the score for the prediction by the score of the groundtruth. For this metric, we directly average over all question-answer pairs as the explanations could be quite different even for questions about the same time period in the same video." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.129 + ], + "angle": 0, + "content": "D.-A. Huang et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.377, + 0.163 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.269 + ], + "angle": 0, + "content": "We evaluate LITA with both temporal localization and video tasks that do not involve temporal localization because most existing Video LLMs cannot handle temporal localization. In addition to our proposed Reasoning Temporal Localization, we further evaluate LITA on Video-based Text Generation Performance Benchmarking proposed by Maaz et al. [27]. This provides a holistic evaluation of LITA as a Video LLM and not just for temporal localization." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.289, + 0.457, + 0.305 + ], + "angle": 0, + "content": "5.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.787, + 0.406 + ], + "angle": 0, + "content": "Architecture. We uniformly sample 100 frames from a video, and use 100 time tokens \\(<1>\\) to \\(<100>\\) to represent timestamps. We use CLIP-L-14 [30] as the visual encoder, and Vicuna [6] as the LLM module. We follow LLaVA's architecture and train a single linear layer for projection [24]. We use 4 frames for slow tokens and use average pool window \\(s = 2\\). With 1 fast token per frame, this leads to a total of \\(100 + \\frac{256}{4} \\times 4 = 356\\) tokens per video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.409, + 0.787, + 0.545 + ], + "angle": 0, + "content": "Training Datasets. We discussed training tasks in Section 3.4. We now discuss the training datasets for each task. For dense video captioning and event localization, we use the training splits of ActivityNet-Captions [16] and YouCook2 [42], which combine to around 11k videos. The event localization dataset can be generated from the dense video captioning dataset by using the caption as query and the timestamps as target. For video question answering, we use NExT-QA [36] as it contains more complex questions. For image instruction tuning, we use LLaVA-150K [24]. For reasoning temporal localization, we use the training split of our ActivityNet-RTL, which builds on the training split of ActivityNet-Captions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.787, + 0.625 + ], + "angle": 0, + "content": "Training Setup. For each of the five tasks, we randomly select 100K samples with replacement (total 500K). We then use batch size 128 and learning rate 2e-5 to train for 4k iterations. The training takes around 13 hours for 13B and 9 hours for 7B models using 8 A100 GPUs. The linear projection is initialized with the LLaVA pre-trained weights." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.646, + 0.631, + 0.661 + ], + "angle": 0, + "content": "5.2 Reasoning Temporal Localization Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.787, + 0.806 + ], + "angle": 0, + "content": "We first evaluate on the newly proposed ActivityNet-RTL for reasoning temporal localization. Please refer to Section 4 for dataset details and metrics. We use \"P@0.5\" for Precision@0.5 and \"Score\" for the GPT evaluation score for explanations. Other than variations of our model, we include Video-LLaMA-v2 [39] and Video-ChatGPT [27] for comparison. We observe that most of their outputs on ActivityNet-RTL omit any timestamps and thus mIOU and Precision@0.5 become absolute. Therefore, for these methods we only evaluate the GPT-Score. In addition, we ablate the following variations for LITA, all of which are trained with the same five training tasks as LITA:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- \"Slow Tokens Only\" samples 4 frames from the video, and computes 64 tokens per frame. It does not use the fast tokens, and it also does not use time tokens" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.329, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LITA: Language Instructed Temporal-Localization Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.406, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.144, + 0.591, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.144, + 0.785, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.263, + 0.788, + 0.36 + ], + "angle": 0, + "content": "Fig. 4: Qualitative results on ActivityNet-RTL. Overall, LITA not only more accurately localize events in the video, but also provides sensible explanations with more details. In the first example, LITA correctly identifies the second arm wrestling. In the second example, LITA provides further details that they are roasting marshmallows. In the third example, LITA impressively recognizes that the girl \"falls off the beam but gets back\" and explains that this shows resilience. These are in contrast to the more generic answers by Video-LLaMA-v2." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.381, + 0.788, + 0.424 + ], + "angle": 0, + "content": "Table 1: Results on ActivityNet-RTL. LITA substantially outperforms all baselines for all metrics. This shows the importance of our design choices. Interestingly the temporal localization accuracy also improves as we scale the model from 7B to 13B." + }, + { + "type": "table", + "bbox": [ + 0.345, + 0.436, + 0.661, + 0.556 + ], + "angle": 0, + "content": "
ModelSizemIOUP@0.5Score
LITA-7B7B24.121.244.0
Video-LLaMA-v2 [39]13B--32.1
Video-ChatGPT [27]13B--38.8
Slow Tokens Only13B14.611.832.2
SlowFast Tokens13B17.514.534.1
LITA-13B13B28.625.946.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.787, + 0.622 + ], + "angle": 0, + "content": "for relative timestamps. This can be seen as naively implementing a Video LLM via the LLaVA architecture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.787, + 0.676 + ], + "angle": 0, + "content": "- \"SlowFast Tokens\" additionally includes fast tokens (i.e. 1 token per frame) compared to \"Slow Tokens Only\". This improves the architectural design for representing video, and should allow better temporal processing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.682, + 0.787, + 0.727 + ], + "angle": 0, + "content": "- \"LITA\" is our full model that further includes time tokens to better represent timestamps compared to \"SlowFast Tokens.\" We consider two model sizes, 7B and 13B, for LITA to understand the effect of different model sizes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Importance of Our Model Components. Results on ActivityNet-RTL are shown in Table 1. All metrics are averaged over three trials. LITA substantially outperforms all baselines for all metrics, and almost double mIOU and P@0.5 when compared to \"Slow Tokens Only\", which is considered as a naive extension of Image LLMs to Video LLMs. \"Score\" is assisted by GPT-4 to evaluate the quality of the explanation provided by the models. While we prompt GPT-4 to ignore the timestamps mentioned in the explanations, we observe that the scores" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.129 + ], + "angle": 0, + "content": "D.-A. Huang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 2: Video-based Text Generation Benchmarking results. LITA significantly outperforms existing Video LLMs including Video-LLaMA-v2 [39] and VideoChatGPT [27] on all evaluated aspects. This shows that LITA not only enables accurate temporal localization, but also generally improves video understanding for Video LLMs." + }, + { + "type": "table", + "bbox": [ + 0.27, + 0.213, + 0.735, + 0.328 + ], + "angle": 0, + "content": "
ModelCorrectnessDetailContextTemporalConsistencyAverage
LLaMA-Adapter2.032.322.301.982.152.16
Video-LLaMA1.962.182.161.821.791.98
Video-LLaMA-v22.362.422.741.832.122.29
VideoChat2.232.502.531.942.242.29
Video-ChatGPT2.402.522.621.982.372.38
LITA2.942.983.432.683.193.04
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.784, + 0.384 + ], + "angle": 0, + "content": "are still slightly affected by the timestamps. Nevertheless, LITA provides better explanations for its reasoning due to an overall better video understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.785, + 0.538 + ], + "angle": 0, + "content": "LITA Gives Detailed and Accurate Explanations. Qualitative results are shown in Figure 4. In the first example, LITA correctly localizes when the second arm wrestling happens. On the other hand, Video-LLaMA-v2 does not recognize that there is arm wrestling in the video. In the second example, while both models identify that the correct activity is cooking with fire, LITA provides much more accurate details, including the fact that they are roasting marshmallows, and the start and end time for that. Finally, in the third example, LITA impressively recognizes the event where the girl \"falls off the beam but gets back\" and correctly responds that this shows resilience in her performance. In contrast, Video-LLaMA-v2 gives a generic answer given the \"resilience\" prompt." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.541, + 0.787, + 0.618 + ], + "angle": 0, + "content": "Temporal Localization Scales with LITA Size. One interesting observation is that the temporal localization quality of LITA also improves as we scale up the model from 7B to 13B (Table 1). One might think that scaling the model only improves language understanding, but our result shows that this could also improve temporal reasoning and understanding for Video LLMs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.637, + 0.557, + 0.652 + ], + "angle": 0, + "content": "5.3 Video-Based Generation Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In addition to our proposed reasoning temporal localization, we further evaluate LITA on standard evaluation for Video LLMs to better compare with existing Video LLMs. We use the \"Video-based Text Generation Performance Benchmarking\" proposed by Maaz et al. [27]. This benchmark selects videos from ActivityNet [4] and annotates question-answer pairs with natural language. This is in contrast to existing video question answering benchmarks, where the answers are often limited to a single word or phrase. In this benchmark, there are specific questions like: \"What kind of tricks is the man performing while skating?\" or generic questions like \"Can you describe the video in detail?\" For evaluation, the benchmark uses GPT to measure the following aspects of Video LLMs' answers: Correctness of Information, Detail Orientation, Contextual Understanding, Temporal Understanding, and Consistency." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.329, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LITA: Language Instructed Temporal-Localization Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 3: Analysis of LITA's training tasks on ActivityNet-RTL. \"RTL\" is needed to predict both timestamps and explanations. \"Video\" includes standard video tasks to improve video understanding. \"NLVQA\" further improves reasoning and natural language generation capabilities of LITA." + }, + { + "type": "table", + "bbox": [ + 0.325, + 0.213, + 0.681, + 0.288 + ], + "angle": 0, + "content": "
ModelRTL VideoNLVQAmIOUP@0.5Score
RTL Only✓✘✘26.620.9
No NLVQA✓✓✘26.923.5
LITA✓✓✓28.625.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.794, + 0.497 + ], + "angle": 0, + "content": "The results are shown in Table 2. We compare LITA with LLaMA-Adapter [40], Video-LLaMA [39], VideoChat [20], and Video-ChatGPT [27]. Video-ChatGPT slightly outperforms other baselines including Video-LLaMA-v2. LITA significantly outperforms these two existing Video LLMs from all aspects. In particular, LITA achieves a \\(22\\%\\) improvement for Correctness of Information (2.94 vs. 2.40) and a \\(36\\%\\) relative improvement for Temporal Understanding (2.68 vs. 1.98). This shows that our emphasis of temporal understanding in training not only enables accurate temporal localization, but also improves the video understanding of LITA. We hypothesize that temporal localization enables the model to learn more details about videos, leading to improved video understanding. A similar observation was made for Image LLMs [5], where joint training with grounding tasks also improved non-grounding text generation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.522, + 0.593, + 0.538 + ], + "angle": 0, + "content": "5.4 Evaluating the Effects of Training Tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.746 + ], + "angle": 0, + "content": "We have analyze the effect of our model components in Section 5.2, where we use all of our training tasks. Now we further analyze the effect of these training tasks for our model. We split the five tasks into three groups: RTL, Video, and NLVQA. \"RTL\" only includes the proposed reasoning temporal localization. Without training on our ActivityNet-RTL, the model would not output timestamps for us to evaluate temporal localization in many cases. The second group \"Video\" includes all the standard video tasks: dense video captioning, event localization, and video question answering. Using these video tasks, the model should learn better video understanding. Finally, \"NLVQA\" refers to the natural language visual question answering task to improve LITA's natural language conversation. We refer to training with just RTL as \"RTL Only,\" and training with both RTL and Video but without NLVQA as \"No NLVQA.\" Training with all three and thus all tasks is our proposed LITA." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Results. The results on ActivityNet-RTL are shown in Table 3 and qualitative comparisons are shown in Figure 5. By only training on RTL, \"RTL Only\" does not have enough supervision to learn the task. This is reflected in both timestamp accuracy (P@0.5) and explanation quality (Score). In addition, for non-temporal questions (Q1 and Q2) in Figure 5, the model cannot properly answer and always use the answer format for reasoning temporal localization. By" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.397, + 0.128 + ], + "angle": 0, + "content": "D.-A. Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.153, + 0.782, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.295, + 0.788, + 0.407 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results to analyze the effect of training tasks. \\(\\tau\\) is the continuous time in seconds. The questions are not from ActivityNet-RTL as we are also interested in questions other than reasoning temporal localization. \"RTL Only\" answers in the format for reasoning temporal localization, as this is the only training data it has. \"No NLVQA\" further includes standard video tasks and can correctly answer non-temporal questions (Q1 and Q2). However, the answers are short due to the short answers in existing video question answering datasets. LITA improves both reasoning and natural language conversation by utilizing NLVQA datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.545 + ], + "angle": 0, + "content": "adding standard video tasks in Video to training, \"No NLVQA\" improves all metrics compared to \"RTL Only\". Qualitatively in Figure 5, it can also answer Q1 and Q2 correctly. However, this capability mainly comes from the inclusion of video question answering datasets, which leads to short answers. LITA further improves by including NLVQA, which contains complex reasoning questions to improve its reasoning capabilities. More importantly, as shown in Figure 5, LITA now answers questions with natural language instead of short phrases." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.575, + 0.36, + 0.591 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We propose Language Instructed Temporal-Localization Assistant (LITA), which enables accurate temporal localization using Video LLMs. LITA demonstrates promising capabilities to answer complex temporal localization questions. At the same time, LITA substantially improves video-based text generation compared to existing Video LLMs even for non-temporal questions. This is the result of both our model design and data strategy. For model design, we propose time tokens to better represent the time and SlowFast tokens to efficiently process video inputs. Our experiments show the importance of these model components. For data strategy, we emphasize temporal localization data in training LITA. To achieve this, we propose the Reasoning Temporal Localization task and curate the ActivityNet-RTL dataset. Our results show that the inclusion of temporal localization data not only enables temporal localization for Video LLMs, but also improves the general video understanding capabilities. We further analyze our training data and show the benefits of incorporating standard video tasks and image instruction tuning data to train Video LLMs." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.329, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LITA: Language Instructed Temporal-Localization Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.221 + ], + "angle": 0, + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.223, + 0.785, + 0.277 + ], + "angle": 0, + "content": "2. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortsman, M., Schmidt, L.: OpenFlamingo: An open-source framework for training large autoregressive vision-language models. arXiv:2308.01390 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.279, + 0.785, + 0.319 + ], + "angle": 0, + "content": "3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.321, + 0.785, + 0.348 + ], + "angle": 0, + "content": "4. Caba Heilbron, F., Escorcia, V., Ghanem, B., Carlos Niebles, J.: Activitynet: A large-scale video benchmark for human activity understanding. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.35, + 0.785, + 0.39 + ], + "angle": 0, + "content": "5. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: MiniGPT-v2: large language model as a unified interface for vision-language multi-task learning. arXiv:2310.09478 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.785, + 0.445 + ], + "angle": 0, + "content": "6. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality (2023), https://lmsys.org/blog/2023-03-30-vicuna/" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.447, + 0.785, + 0.488 + ], + "angle": 0, + "content": "7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: PaLM: Scaling language modeling with pathways. arXiv:2204.02311 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.49, + 0.785, + 0.53 + ], + "angle": 0, + "content": "8. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructblip: Towards general-purpose vision-language models with instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.532, + 0.785, + 0.585 + ], + "angle": 0, + "content": "9. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: The EPIC-KITCHENS Dataset: Collection, challenges and baselines. IEEE Trans. PAMI 43(11), 4125-4141 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.588, + 0.785, + 0.628 + ], + "angle": 0, + "content": "10. Damen, D., Doughty, H., Farinella, G.M., Furnari, A., Ma, J., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. IJCV 130, 33-55 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.63, + 0.785, + 0.656 + ], + "angle": 0, + "content": "11. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: NAACL (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.659, + 0.785, + 0.684 + ], + "angle": 0, + "content": "12. Feichtenhofer, C., Fan, H., Malik, J., He, K.: SlowFast networks for video recognition. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.687, + 0.785, + 0.713 + ], + "angle": 0, + "content": "13. Huang, B., Wang, X., Chen, H., Song, Z., Zhu, W.: Vtimellm: Empower llm to grasp video moments. arXiv preprint arXiv:2311.18445 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.715, + 0.785, + 0.741 + ], + "angle": 0, + "content": "14. Huang, G., Pang, B., Zhu, Z., Rivera, C., Soricut, R.: Multimodal pretraining for dense video captioning. In: AACL-IJCNLP (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.743, + 0.785, + 0.783 + ], + "angle": 0, + "content": "15. Jiang, Y.G., Liu, J., Zamir, A.R., Toderici, G., Laptev, I., Shah, M., Sukthankar, R.: Thumos challenge: Action recognition with a large number of classes (2014), https://www.crcv.ucf.edu/THUMOS14/home.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.786, + 0.785, + 0.811 + ], + "angle": 0, + "content": "16. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Niebles, J.C.: Dense-captioning events in videos. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "17. Kuehne, H., Arslan, A.B., Serre, T.: The language of actions: Recovering the syntax and semantics of goal-directed human activities. In: CVPR (2014)" + }, + { + "type": "list", + "bbox": [ + 0.221, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.396, + 0.128 + ], + "angle": 0, + "content": "D.-A. Huang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: LISA: Reasoning segmentation via large language model. arXiv:2308.00692 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "19. Li, B., Zhang, Y., Chen, L., Wang, J., Pu, F., Yang, J., Li, C., Liu, Z.: MIMIC-IT: Multi-modal in-context instruction tuning. arXiv:2306.05425 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "20. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: VideoChat: Chat-centric video understanding. arXiv:2305.06355 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "21. Li, Y., Wang, C., Jia, J.: Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "22. Lin, B., Zhu, B., Ye, Y., Ning, M., Jin, P., Yuan, L.: Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.327 + ], + "angle": 0, + "content": "23. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv:2310.03744 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.342 + ], + "angle": 0, + "content": "24. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "25. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "26. Miech, A., Zhukov, D., Alayrac, J.B., Tapaswi, M., Laptev, I., Sivic, J.: HowTo100M: Learning a text-video embedding by watching hundred million narrated video clips. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "27. Muhammad Maaz, Hanoona Rasheed, S.K., Khan, F.: Video-ChatGPT: Towards detailed video understanding via large vision and language models. ArXiv 2306.05424 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.632, + 0.466 + ], + "angle": 0, + "content": "28. OpenAI: GPT-4 technical report. arXiv:2303.08774 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "29. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv:2306.14824 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "30. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "31. Sigurdsson, G.A., Varol, G., Wang, X., Farhadi, A., Laptev, I., Gupta, A.: Hollywood in homes: Crowdsourcing data collection for activity understanding. In: ECCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.632 + ], + "angle": 0, + "content": "32. Tang, Y., Ding, D., Rao, Y., Zheng, Y., Zhang, D., Zhao, L., Lu, J., Zhou, J.: COIN: A large-scale dataset for comprehensive instructional video analysis. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "33. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: LLaMA: Open and efficient foundation language models. arXiv:2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "34. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: LLaMA 2: Open foundation and fine-tuned chat models. arXiv:2307.09288 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.743 + ], + "angle": 0, + "content": "35. Wu, S., Fei, H., Qu, L., Ji, W., Chua, T.S.: NExT-GPT: Any-to-any multimodal LLM. arXiv:2309.05519 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "36. Xiao, J., Shang, X., Yao, A., Chua, T.S.: NExT-QA: Next phase of question-answering to explaining temporal actions. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "37. Yan, S., Xiong, X., Nagrani, A., Arnab, A., Wang, Z., Ge, W., Ross, D., Schmid, C.: UnLoc: A unified framework for video localization tasks. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "38. Yang, A., Nagrani, A., Seo, P.H., Miech, A., Pont-Tuset, J., Laptev, I., Sivic, J., Schmid, C.: Vid2Seq: Large-scale pretraining of a visual language model for dense video captioning. In: CVPR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.329, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LITA: Language Instructed Temporal-Localization Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "39. Zhang, H., Li, X., Bing, L.: Video-LLaMA: An instruction-tuned audio-visual language model for video understanding. arXiv:2306.02858 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.217 + ], + "angle": 0, + "content": "40. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: LLaMA-Adapter: Efficient fine-tuning of language models with zero-init attention. arXiv:2303.16199 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.217, + 0.788, + 0.272 + ], + "angle": 0, + "content": "41. Zhang, Y., Li, Y., Cui, L., Cai, D., Liu, L., Fu, T., Huang, X., Zhao, E., Zhang, Y., Chen, Y., Wang, L., Luu, A.T., Bi, W., Shi, F., Shi, S.: Siren's song in the ai ocean: A survey on hallucination in large language models. arXiv:2309.01219 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.3 + ], + "angle": 0, + "content": "42. Zhou, L., Xu, C., Corso, J.J.: Towards automatic learning of procedures from web instructional videos. In: AAAI (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.342 + ], + "angle": 0, + "content": "43. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: MiniGPT-4: Enhancing vision-language understanding with advanced large language models. arXiv:2304.10592 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.342 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/LITA_ Language Instructed Temporal-Localization Assistant/94963324-666f-4292-a883-31e4dd7adc61_origin.pdf b/2024/LITA_ Language Instructed Temporal-Localization Assistant/94963324-666f-4292-a883-31e4dd7adc61_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8c7ff38abc3f75d85dd4d9d8ba8712a54020fa2b --- /dev/null +++ b/2024/LITA_ Language Instructed Temporal-Localization Assistant/94963324-666f-4292-a883-31e4dd7adc61_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36a89d7d6a7a7b40f77266f889be1a64a34efcbf3d17a047d6a3820618a3705f +size 3766196 diff --git a/2024/LITA_ Language Instructed Temporal-Localization Assistant/full.md b/2024/LITA_ Language Instructed Temporal-Localization Assistant/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cd53d88c87edec5caa9fb45f68f219f7b519c91a --- /dev/null +++ b/2024/LITA_ Language Instructed Temporal-Localization Assistant/full.md @@ -0,0 +1,260 @@ +# LITA: Language Instructed Temporal-Localization Assistant + +De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz + +NVIDIA + +Abstract. There has been tremendous progress in multimodal Large Language Models (LLMs). Recent works have extended these models to video input with promising instruction following capabilities. However, an important missing piece is temporal localization. These models cannot accurately answer the "When?" questions. We identify three key aspects that limit their temporal localization capabilities: (i) time representation, (ii) architecture, and (iii) data. We address these shortcomings by proposing Language Instructed Temporal-Localization Assistant (LITA) with the following features: (1) We introduce time tokens that encode timestamps relative to the video length to better represent time in videos. (2) We introduce SlowFast tokens in the architecture to capture temporal information at fine temporal resolution. (3) We emphasize temporal localization data for LITA. In addition to leveraging existing video datasets with timestamps, we propose a new task, Reasoning Temporal Localization (RTL), along with the dataset, ActivityNet-RTL, for learning and evaluating this task. Reasoning temporal localization requires both the reasoning and temporal localization of Video LLMs. LITA demonstrates strong performance on this challenging task, nearly doubling the temporal mean intersection-over-union (mIoU) of baselines. In addition, we show that our emphasis on temporal localization also substantially improves video-based text generation compared to existing Video LLMs, including a $36\%$ relative improvement of Temporal Understanding. Code is available at: https://github.com/NVlabs/LITA + +Keywords: Large Language Models $\cdot$ Video Understanding $\cdot$ Temporal Localization + +# 1 Introduction + +Large language models (LLMs) [3,7,11,28,33,34] have demonstrated impressive instruction following capabilities, and shown that language can be a universal interface for various tasks [7,28]. These models can be further extended to multimodal LLMs to process language and other modalities, such as image, video, and audio [1,24,43]. + +While most multimodal LLMs focus on images for visual content, several recent works introduce models that specialize in processing videos [20,25,27,39]. These Video LLMs preserve the instruction following capabilities of LLMs and + +Fig. 1: Example to illustrate our proposed Reasoning Temporal Localization (RTL). Instead of directly querying about an event, questions in RTL require further reasoning to answer. Here, the model needs to compare all activities in the video to find the timestamps of the most energetic activity (i.e., handspring). +![](images/ab78a62d5276cf97160c0ecb4374363d597abe3a15cb26e164e018518969e32a.jpg) +A: Her dance is the most energetic between 7.37s and 12.81s when she does a handspring, which is more energetic than standing and lying on the floor. + +allow users to ask various questions about a given video. However, one important missing piece in these Video LLMs is temporal localization. When prompted with the "When?" questions, these models cannot accurately localize time periods, and often hallucinate irrelevant information [41]. Temporal localization is an important component that differentiates videos from images, and has been widely studied outside the context of instruction following LLMs [4,10,15,31]. It is thus crucial for Video LLMs to have temporal localization capabilities. + +We identify three key aspects that limit the temporal localization capabilities of existing Video LLMs: time representation, architecture, and data. First, existing models often represent timestamps as plain text (e.g. 01:22 or 142sec). However, given a set of frames, the correct timestamp still depends on the frame rate, which the model does not have access to. This makes learning temporal localization harder. Second, the architecture of existing Video LLMs might not have sufficient temporal resolution to interpolate time information accurately. For example, Video-LLaMA [39] only uniformly samples 8 frames from the entire video, which is insufficient for accurate temporal localization. Finally, temporal localization is largely ignored in the data used by existing Video LLMs. Data with timestamps are only a small subset of video instruction tuning data, and the accuracy of these timestamps is also not verified. + +Our Approach. We address the aforementioned shortcomings of existing Video LLMs, and propose Language Instructed Temporal-Localization Assistant (LITA): (1) Time Representation: We introduce time tokens to represent relative timestamps and allow Video LLMs to better communicate about time than using plain text. (2) Architecture: We introduce SlowFast tokens to capture temporal information at fine temporal resolution to enable accurate temporal localization. (3) Data: We emphasize temporal localization data for LITA. We propose a new task, Reasoning Temporal Localization (RTL), along with the dataset, ActivityNet-RTL, for learning this task. + +The first important design of LITA is to use relative representation for time (e.g. first $10\%$ of the video) instead of the absolute time representation with plain text (e.g. 01:22). We divide a given video into $T$ equal length chunks, and introduce $T$ time tokens $\langle 1\rangle$ to $\langle \mathbf{T}\rangle$ to represent the relative time location in the video. During training and inference, these time tokens can be easily encoded and decoded from plain text timestamps given the length of the video. The start and end timestamps are well-defined by the time tokens given only the input + +video. This is in contrast to plain text timestamps. Without the frame rate, the correct absolute timestamp is ill-defined given just the video frames. + +The second important design of LITA is to use densely sampled input frames from videos. It is unlikely to achieve accurate temporal localization with only sparsely sampled frames. The challenge is that the LLM module inside Video LLMs cannot naively process large numbers of frames simultaneously due to context length limitation. Take LLaVA [24] as an example. Each image is converted to 256 tokens, which are then fed into its LLM module as input. If we directly feed 100 frames to the LLM module, then that is $256 \times 100 = 25600$ tokens, which is already over the max context length for some LLMs [6,34]. Inspired by the SlowFast architecture for videos [12], we instead consider two types of tokens, fast tokens and slow tokens, to address this efficiency issue. We generate fast tokens at a high temporal resolution to provide the temporal information, while keeping the tokens per frame at a low number. On the other hand, we generate slow tokens at a low temporal resolution, which allows us to use a higher number of tokens per frame to provide the spatial information. + +Finally, We emphasize temporal localization data for LITA. We include dense video captioning [16] and event localization [37] in instruction tuning of LITA. These tasks include human annotated timestamps to promote accurate temporal localization. In addition to leveraging existing data and tasks, we further propose a new task, Reasoning Temporal Localization (RTL), along with the dataset, ActivityNet-RTL, for training and evaluating this task. Answers to RTL questions can only be derived by utilizing world knowledge and temporal reasoning. Figure 1 shows an example. To answer the question: "When does the woman's dance become the most energetic?" the model needs to first recognize the woman's dance moves in the video, then reason about the most active part, and finally temporally localize the event (i.e. handspring). In addition to the predicted timestamps, we further consider the explanation provided by the model. Thus our new task not only assesses temporal understanding, but also requires strong reasoning capabilities that are unique to LLMs. + +For the challenging RTL task, LITA doubles baseline's performance for temporal metrics (mIOU, Precision@0.5), while providing much better explanations. In addition to enabling accurate temporal localization, we show that our emphasis on temporal understanding also improves LITA's core Video LLM capabilities. LITA substantially improves all scores on a benchmark for video-based question answering [27]. This includes a $22\%$ relative improvement for Correctness of Information, and a $36\%$ relative improvement for Temporal Understanding compared to existing Video LLMs. + +# 2 Related Work + +Multimodal Large Language Models. Large language models (LLMs) [7,28] inspire recent works to address multimodal tasks by leveraging LLMs [35]. Some approaches add additional parameters inside LLMs, such as gated cross-attention layers [1,2,19] or adapter layers [40], to adapt it to process multimodal inputs. + +Several works, on the other hand, only use modules, such as projection layers or Q-Formers, to project outputs of visual encoders to the input space of LLMs [8, 24,43]. Recent works further expand multimodal LLM to visual grounding tasks, such as detection [5,23,29] and segmentation [18]. The most related to ours is LISA [18], which extends referring segmentation to reasoning segmentation. We share the same spirit and propose Reasoning Temporal Localization to jointly evaluate reasoning and temporal understanding. + +Video Large Language Models. Building on the success of multimodal LLMs, several works extend image LLMs to Video LLMs [20, 25, 27, 39]. These works mainly use the approach of projecting visual tokens to LLMs' input space using projection layers [25, 27] or Q-Formers [20, 39]. While these models show promise in descriptive questions and instructions, they still lack temporal localization capabilities. LITA is designed to address this shortcoming, while also improving downstream video tasks. Concurrent works [13, 21, 22] include improvements like better image-video joint training. The most related concurrent work to ours is VTimeLLM [13], which also aims to address temporal localization of Video LLMs. We further introduce the reasoning aspect to temporal localization. + +Temporal Localization in Videos. The goal of temporal localization is to pinpoint activities within untrimmed video sequences on a temporal scale [37]. The target activities can be predefined action classes [9, 15] or events described by natural language [4, 42]. Our goal of video temporal understanding is also related to various video tasks, such as dense video captioning [14, 16, 38] and action segmentation [17, 26, 32]. Models for these temporal tasks can have quite different design. Instruction following Video LLMs like LITA provide a way to unify these frameworks. + +# 3 Language Instructed Temporal-Localization + +LITA enables temporal localization for Video LLMs by: (1) relative time representation with the time tokens, (2) SlowFast tokens to capture temporal information at fine temporal resolution, (3) multi-task training that includes accurate timestamps. We will first introduce the overall architecture and discuss further details of individual components. + +# 3.1 Architecture + +An overview of LITA is shown in Figure 2. We build on Image LLMs. In particular, we select LLaVA due to its simplicity and effectiveness [24]. Note that LITA does not depend on the specific underlying Image LLM architecture and can be easily adapted to other base architectures. + +Given a video, we first uniformly select $T$ frames and encode each frame into $M$ tokens. $T$ should be large enough to support the desired granularity of temporal localization. $T \times M$ is a large number that usually cannot be directly processed by the LLM module. Thus, we then use SlowFast pooling to reduce the $T \times M$ tokens to $T + M$ tokens. + +A1: She is dancing from $<2>$ to $<3>$ . A2: Her cloth is black. A3: $<1><2>$ She is standing. $<2><3>$ . + +![](images/ff7a7f18c17206da75583469eada7ab33bc207ab2aaf74ea4df304eb4721f20d.jpg) +Fig. 2: Overview of LITA. The input video frames are first encoded into visual tokens (numbered by frame), which are further processed by two pathways. The Fast Token pathway averages all the tokens in a frame to maintain a high temporal resolution. The Slow Token pathway sparsely samples frames to maintain a larger number of tokens per frame to provide spatial information. Timestamps are converted to time tokens $<1>$ to $$ . This is important for better temporal localization learning. Various video tasks on the right can be converted to natural language visual question answering (Q1-3 and A1-3) to jointly optimize LITA. + +The slow and fast tokens are projected by a linear layer and concatenated with the text tokens to use as input to the LLM module. The text tokens (prompt) are processed to convert any referenced timestamps to specialized time tokens ( $<1>$ to $<\mathrm{T}>$ ). All the input tokens are then jointly processed by the LLM module sequentially. We fine-tune the entire model with our reasoning temporal localization data (Section 4) along with other video tasks, such as dense video captioning and event localization. LITA learns to use time tokens instead of absolute timestamps. For temporal localization, we can then directly ask LITA the "When" questions (e.g. "When is she dancing?"). LITA would respond with time tokens (e.g. "She is dancing from $<2>$ to $<3>$ ). which can then be converted to timestamps given the video length. + +# 3.2 Time Tokens + +We use a relative time representation instead of absolute timestamps in LITA. As shown in Figure 2, the LLM module can only see the visual tokens (slow and fast) and the language tokens (text prompt). There is not enough information in this input space for the LLM module to infer the absolute timestamp because the frame rate is not known to the model in advance. A better way is to represent timestamps relative to the video length, thus removing the dependency on the frame rate. We divide the video into $T$ chunks and use $T$ specialized time tokens $<1>$ to $<\mathsf{T}>$ for timestamps. Given a continuous timestamp $\tau$ and video length $L$ , $\tau$ can be easily converted to time token $<\mathsf{t}>$ , where $t = \mathrm{round}(\tau(T-1)/L) + 1$ , and conversely $<\mathsf{t}>$ can be converted back to $\tau = L(t-1)/(T-1)$ . While this does introduce discretization error, it greatly simplifies the time representation + +with LLMs. Relative timestamp is also used in other temporally heavy video tasks, such as dense video captioning [38]. + +Given this time representation, many video tasks related to temporal localization can be transformed into language instructions and answers. For example, dense video captioning can be achieved by prompting the model with "Describe the video. Each sentence begins with start and end timestamps." (Q3 and A3 in Fig. 2). Standard event localization is also transformed to "When does X happen?" (Q1 and A1 in Fig. 2). We also incorporate standard video question answering (Q2 and A2 in Fig. 2). More details are discussed in Section 3.4. + +# 3.3 SlowFast Visual Tokens + +We have discussed how we discretize time in videos into $T$ steps in order to make Video LLMs better at reasoning about time. Still, the visual input should match the temporal resolution $T$ in order to achieve effective temporal processing. Ideally, one would need at least $T$ frames to temporally localize events with the resolution $T$ . However, naively feeding all $T$ frames into the LLM module could be computationally prohibitive. In our experiment, we use $T = 100$ and $M = 256$ (CLIP ViT-L-14 [30]). This leads to 25600 tokens per video. + +Inspired by SlowFast models [12], we consider two pathways to pool the $T \times M$ tokens for $T$ frames. The first is densely sampled fast tokens to provide temporal information. We obtain $T$ fast tokens from $T$ frames by averaging all the tokens belonging to the same frame. The second is the sparsely sampled slow tokens to maintain better spatial information. We select a spatial downsampling ratio of $s$ , and uniformly select $s^2$ frames from the video. For each selected frame, we perform a $s \times s$ spatial average pooling to the $M$ tokens, which lead to $\frac{M}{s^2}$ slow tokens per frame. This leads to a total $M = \frac{M}{s^2} \times s^2$ slow tokens. We use $s = 2$ in our experiments. This leads to a total of $T + M$ tokens to represent a video instead of $T \times M$ tokens. + +# 3.4 Training Tasks + +In addition to architecture, training tasks and data also play an important role for LLMs. We emphasize temporal localization data and train LITA with the following five tasks: (1) dense video captioning [16], (2) event localization [37], (3) video question answering [36], (4) natural language visual question answering [24], and (5) our proposed reasoning temporal localization. Temporal localization is a crucial component for three out of the five tasks (1, 2, and 5). + +We now introduce each task in order. The first three tasks are standard video tasks and equip LITA with basic video understanding: + +Dense Video Captioning. In dense video captioning [16], each video is described by a set of sentences, and each sentence comes with the start and end timestamps of the event. Each sentence in dense video captioning can thus be represented as: SENTENCE. We then sort all sentences by its start time, and directly concatenate all sentences and timestamps. One + +example prompt to the model for this task is: "Provide a detailed description of the given video. Each sentence should begin with the start and end timestamps." Other prompts are included in the supplementary materials. + +Event Localization. In event localization, the goal is to temporally localize the event described by a sentence. We use a simple answer format: . One example prompt for this task is: "When does "SENTENCE" happen in the video? Answer the question only using start and end timestamps." + +Video Question Answering. The question answering task is already represented as language instructions. However, answers in existing question answering datasets often consist of a single word or phrase because models for this task might not be able to generate longer text. We follow Liu et al. [23] and append the following prompt to the question: "Answer the question using a single word or phrase." The goal is to provide the context for short answers so that it affects the model's text generation less. + +Natural Language Visual Question Answering. Training with the above three tasks gives LITA video understanding capabilities. However, we observe that models trained with only these tasks often provide short answers and lack natural language conversation capabilities. We thus also train LITA with natural language visual question answering or visual instruction tuning datasets [24]. The goal is to improve the natural language conversation of LITA. We find that mixing instruction tuning datasets [24] with standard video tasks improves LITA's conversation quality while maintaining good video understanding. + +Reasoning Temporal Localization. Finally, we also train LITA with our reasoning temporal localization task (details in Section 4). The answer to a reasoning temporal localization question consists of two parts: timestamps and explanation. We find it challenging for models to simultaneously output both of them without any example. Nevertheless, with some training data, LITA quickly pick up reasoning and temporal localization, and provide both the timestamps and explanation of its reasoning in answers. + +# 4 Reasoning Temporal Localization + +We now discuss further details of the proposed Reasoning Temporal Localization (RTL) task. Standard temporal localization does not fully leverage the potential of Video LLMs. One impressive aspect of LLMs is its reasoning capabilities. LLMs can even answer complex questions that involve multi-step reasoning. Therefore, we propose the RTL task to utilize both of Video LLMs' temporal understanding and reasoning capabilities. + +# 4.1 Problem Definition + +In reasoning temporal localization, the query is still a "when" question that asks about the start and end timestamps of an event. The key difference compared to the standard temporal localization task is that the target event is not directly + +![](images/b5a6e94588981e6b0cd9296e5371fb8db5061a42b43885e657f5144cb7992af1.jpg) +(a) Temporal Reasoning + +![](images/a395746009371616fd09e92cecf11e28819650b81d679f742a3dc2af14efdbb3.jpg) +(b) World Knowledge +Fig. 3: Examples from our ActivityNet-RTL dataset. RTL questions ask about events that are not explicitly described. The model needs to utilize reasoning or its world knowledge to answer. This is in contrast to standard temporal localization, which directly asks about the event of interest. For example in (c), standard temporal localization might directly ask: "when does the man sit on the floor?" + +![](images/ec4c2a7996f42e3f5bab8ece02e17acfe4a72fb0c095be728221009526a24d1a.jpg) +(c) Complex Reasoning + +described in the question, and can only be inferred by reasoning and using world knowledge of the model. The answer to such a question thus consists of two parts: (1) the start and end timestamps of the target event, and (2) an explanation of the reasoning process the model goes through to derive the timestamps. + +Some examples are shown in Figure 3 to better illustrate the idea. The answer format is: [start end] Explanation. In Figure 3(a), the model not only has to localize "adjust their position to avoid obstacles," but also needs to temporally reason about which instance happened earlier in the video. In Figure 3(b), instead of directly asking about "one-leg row," it asks about the workout targeting balance and stability. The model thus needs to utilize its knowledge of what kind of exercises are good for balance and stability. Finally, there are questions that require multi-step reasoning. In Figure 3(c), the question asks about "the most atypical ways of jump roping," which requires the model to understand what is typical and atypical for jump roping, and then temporally find the most atypical time period. A standard temporal localization task, in contrast, would just ask when the man is sitting on the floor. + +# 4.2 ActivityNet-RTL Dataset + +The above examples are from ActivityNet-RTL, a dataset curated by us for the Reasoning Temporal Localization (RTL) task. We build our dataset from the ActivityNet Captions dataset [16], which annotates multiple events described by sentences in a video, and all the events are temporally localized with start and end timestamps. Consider the following toy example: + +[00:00-00:10] A woman is standing. + +[00:12-00:30] The woman is dancing. + +[00:32-00:36] The woman is sleeping. + +We then use this as context and ask GPT-4 to generate temporal localization questions that require further reasoning to answer. We also ask GPT-4 to simultaneously generate the answer that includes the queried start and end timestamps, along with the explanation about the reasoning process. + +Using the above toy example, by seeing that the woman has done three activities, one possible reasoning temporal localization question is to ask "When is she the least active?" Since sleeping is the least active out of the three activities, the target time period is 00:32-00:36, the period when she is sleeping. This illustrates how GPT-4 can still generate interesting questions without seeing the actual video. We annotate few-shot examples for GPT-4 as in previous works to improve the generation quality [24]. All of our prompts are included in the supplementary materials. + +Training Set Generation. For our training set, we directly use the results generated by GPT-4 with 10,009 videos from the training split of ActivityNet-Captions. This leads to 33,557 Reasoning Temporal Localization question-answer pairs. By inspecting the GPT generated results, we find that most of the questions are valid temporal localization questions given the context. The main shortcoming is that not all question-answer pairs require reasoning. Sometimes GPT-4 generates questions that directly ask about events that are already described by the dense video captions. However, we do hope that LITA can also answer these standard temporal localization questions correctly using natural language. We thus leave these questions in the training set. + +Evaluation Set Curation. On the other hand, the evaluation set requires manual efforts otherwise we would end up with many non-reasoning questions. We start from the GPT-4 generated questions using a subset of the ActivityNet-Captions validation set, and manually remove non-reasoning questions. We also double check the timestamps and explanations in the answers. This leads to a total of 229 question-answer pairs for 160 videos. + +# 4.3 Metrics + +We consider three metrics for ActivityNet-RTL: mIOU, Precision@0.5, and GPT-4 Relative Scores. The first two metrics are for temporal localization, and the third metric is to evaluate the explanation capability. mIOU averages the intersection-over-union (IOU) between predicted and groundtruth start and end timestamps. Precision@0.5 measures the percentage of predictions that have over 0.5 IOU. We first average these two metrics per video, and then average over all videos in the evaluation set. This avoids over-weighting videos and time periods with more questions, as some time periods in a video have multiple questions. + +To evaluate the quality of the explanation, we follow the evaluation pipeline of LLaVA [24] and leverage GPT-4 for evaluation. GPT-4 is asked to evaluate the helpfulness, relevance, accuracy, and level of details of the explanations, and give a score from 1 to 10. We ask GPT-4 to evaluate both the predicted and groundtruth explanations, and normalize the score for the prediction by the score of the groundtruth. For this metric, we directly average over all question-answer pairs as the explanations could be quite different even for questions about the same time period in the same video. + +# 5 Experiments + +We evaluate LITA with both temporal localization and video tasks that do not involve temporal localization because most existing Video LLMs cannot handle temporal localization. In addition to our proposed Reasoning Temporal Localization, we further evaluate LITA on Video-based Text Generation Performance Benchmarking proposed by Maaz et al. [27]. This provides a holistic evaluation of LITA as a Video LLM and not just for temporal localization. + +# 5.1 Implementation Details + +Architecture. We uniformly sample 100 frames from a video, and use 100 time tokens $<1>$ to $<100>$ to represent timestamps. We use CLIP-L-14 [30] as the visual encoder, and Vicuna [6] as the LLM module. We follow LLaVA's architecture and train a single linear layer for projection [24]. We use 4 frames for slow tokens and use average pool window $s = 2$ . With 1 fast token per frame, this leads to a total of $100 + \frac{256}{4} \times 4 = 356$ tokens per video. + +Training Datasets. We discussed training tasks in Section 3.4. We now discuss the training datasets for each task. For dense video captioning and event localization, we use the training splits of ActivityNet-Captions [16] and YouCook2 [42], which combine to around 11k videos. The event localization dataset can be generated from the dense video captioning dataset by using the caption as query and the timestamps as target. For video question answering, we use NExT-QA [36] as it contains more complex questions. For image instruction tuning, we use LLaVA-150K [24]. For reasoning temporal localization, we use the training split of our ActivityNet-RTL, which builds on the training split of ActivityNet-Captions. + +Training Setup. For each of the five tasks, we randomly select 100K samples with replacement (total 500K). We then use batch size 128 and learning rate 2e-5 to train for 4k iterations. The training takes around 13 hours for 13B and 9 hours for 7B models using 8 A100 GPUs. The linear projection is initialized with the LLaVA pre-trained weights. + +# 5.2 Reasoning Temporal Localization Evaluation + +We first evaluate on the newly proposed ActivityNet-RTL for reasoning temporal localization. Please refer to Section 4 for dataset details and metrics. We use "P@0.5" for Precision@0.5 and "Score" for the GPT evaluation score for explanations. Other than variations of our model, we include Video-LLaMA-v2 [39] and Video-ChatGPT [27] for comparison. We observe that most of their outputs on ActivityNet-RTL omit any timestamps and thus mIOU and Precision@0.5 become absolute. Therefore, for these methods we only evaluate the GPT-Score. In addition, we ablate the following variations for LITA, all of which are trained with the same five training tasks as LITA: + +- "Slow Tokens Only" samples 4 frames from the video, and computes 64 tokens per frame. It does not use the fast tokens, and it also does not use time tokens + +![](images/671dafffd172072992a73a62858e7d26c5e3e56edbebacb250a44d6a6ef89a0d.jpg) +Fig. 4: Qualitative results on ActivityNet-RTL. Overall, LITA not only more accurately localize events in the video, but also provides sensible explanations with more details. In the first example, LITA correctly identifies the second arm wrestling. In the second example, LITA provides further details that they are roasting marshmallows. In the third example, LITA impressively recognizes that the girl "falls off the beam but gets back" and explains that this shows resilience. These are in contrast to the more generic answers by Video-LLaMA-v2. + +![](images/78817c65ea273a07790a5faa48d6a4ae01e26f5478c71e1d54a228012f2ce620.jpg) + +![](images/9cdde68de66a7f83d7c090c040babb37c083211a4c5360c40b6eb435760b811e.jpg) + +Table 1: Results on ActivityNet-RTL. LITA substantially outperforms all baselines for all metrics. This shows the importance of our design choices. Interestingly the temporal localization accuracy also improves as we scale the model from 7B to 13B. + +
ModelSizemIOUP@0.5Score
LITA-7B7B24.121.244.0
Video-LLaMA-v2 [39]13B--32.1
Video-ChatGPT [27]13B--38.8
Slow Tokens Only13B14.611.832.2
SlowFast Tokens13B17.514.534.1
LITA-13B13B28.625.946.3
+ +for relative timestamps. This can be seen as naively implementing a Video LLM via the LLaVA architecture. + +- "SlowFast Tokens" additionally includes fast tokens (i.e. 1 token per frame) compared to "Slow Tokens Only". This improves the architectural design for representing video, and should allow better temporal processing. + +- "LITA" is our full model that further includes time tokens to better represent timestamps compared to "SlowFast Tokens." We consider two model sizes, 7B and 13B, for LITA to understand the effect of different model sizes. + +Importance of Our Model Components. Results on ActivityNet-RTL are shown in Table 1. All metrics are averaged over three trials. LITA substantially outperforms all baselines for all metrics, and almost double mIOU and P@0.5 when compared to "Slow Tokens Only", which is considered as a naive extension of Image LLMs to Video LLMs. "Score" is assisted by GPT-4 to evaluate the quality of the explanation provided by the models. While we prompt GPT-4 to ignore the timestamps mentioned in the explanations, we observe that the scores + +Table 2: Video-based Text Generation Benchmarking results. LITA significantly outperforms existing Video LLMs including Video-LLaMA-v2 [39] and VideoChatGPT [27] on all evaluated aspects. This shows that LITA not only enables accurate temporal localization, but also generally improves video understanding for Video LLMs. + +
ModelCorrectnessDetailContextTemporalConsistencyAverage
LLaMA-Adapter2.032.322.301.982.152.16
Video-LLaMA1.962.182.161.821.791.98
Video-LLaMA-v22.362.422.741.832.122.29
VideoChat2.232.502.531.942.242.29
Video-ChatGPT2.402.522.621.982.372.38
LITA2.942.983.432.683.193.04
+ +are still slightly affected by the timestamps. Nevertheless, LITA provides better explanations for its reasoning due to an overall better video understanding. + +LITA Gives Detailed and Accurate Explanations. Qualitative results are shown in Figure 4. In the first example, LITA correctly localizes when the second arm wrestling happens. On the other hand, Video-LLaMA-v2 does not recognize that there is arm wrestling in the video. In the second example, while both models identify that the correct activity is cooking with fire, LITA provides much more accurate details, including the fact that they are roasting marshmallows, and the start and end time for that. Finally, in the third example, LITA impressively recognizes the event where the girl "falls off the beam but gets back" and correctly responds that this shows resilience in her performance. In contrast, Video-LLaMA-v2 gives a generic answer given the "resilience" prompt. + +Temporal Localization Scales with LITA Size. One interesting observation is that the temporal localization quality of LITA also improves as we scale up the model from 7B to 13B (Table 1). One might think that scaling the model only improves language understanding, but our result shows that this could also improve temporal reasoning and understanding for Video LLMs. + +# 5.3 Video-Based Generation Evaluation + +In addition to our proposed reasoning temporal localization, we further evaluate LITA on standard evaluation for Video LLMs to better compare with existing Video LLMs. We use the "Video-based Text Generation Performance Benchmarking" proposed by Maaz et al. [27]. This benchmark selects videos from ActivityNet [4] and annotates question-answer pairs with natural language. This is in contrast to existing video question answering benchmarks, where the answers are often limited to a single word or phrase. In this benchmark, there are specific questions like: "What kind of tricks is the man performing while skating?" or generic questions like "Can you describe the video in detail?" For evaluation, the benchmark uses GPT to measure the following aspects of Video LLMs' answers: Correctness of Information, Detail Orientation, Contextual Understanding, Temporal Understanding, and Consistency. + +Table 3: Analysis of LITA's training tasks on ActivityNet-RTL. "RTL" is needed to predict both timestamps and explanations. "Video" includes standard video tasks to improve video understanding. "NLVQA" further improves reasoning and natural language generation capabilities of LITA. + +
ModelRTL VideoNLVQAmIOUP@0.5Score
RTL Only✓✘✘26.620.9
No NLVQA✓✓✘26.923.5
LITA✓✓✓28.625.9
+ +The results are shown in Table 2. We compare LITA with LLaMA-Adapter [40], Video-LLaMA [39], VideoChat [20], and Video-ChatGPT [27]. Video-ChatGPT slightly outperforms other baselines including Video-LLaMA-v2. LITA significantly outperforms these two existing Video LLMs from all aspects. In particular, LITA achieves a $22\%$ improvement for Correctness of Information (2.94 vs. 2.40) and a $36\%$ relative improvement for Temporal Understanding (2.68 vs. 1.98). This shows that our emphasis of temporal understanding in training not only enables accurate temporal localization, but also improves the video understanding of LITA. We hypothesize that temporal localization enables the model to learn more details about videos, leading to improved video understanding. A similar observation was made for Image LLMs [5], where joint training with grounding tasks also improved non-grounding text generation. + +# 5.4 Evaluating the Effects of Training Tasks + +We have analyze the effect of our model components in Section 5.2, where we use all of our training tasks. Now we further analyze the effect of these training tasks for our model. We split the five tasks into three groups: RTL, Video, and NLVQA. "RTL" only includes the proposed reasoning temporal localization. Without training on our ActivityNet-RTL, the model would not output timestamps for us to evaluate temporal localization in many cases. The second group "Video" includes all the standard video tasks: dense video captioning, event localization, and video question answering. Using these video tasks, the model should learn better video understanding. Finally, "NLVQA" refers to the natural language visual question answering task to improve LITA's natural language conversation. We refer to training with just RTL as "RTL Only," and training with both RTL and Video but without NLVQA as "No NLVQA." Training with all three and thus all tasks is our proposed LITA. + +Results. The results on ActivityNet-RTL are shown in Table 3 and qualitative comparisons are shown in Figure 5. By only training on RTL, "RTL Only" does not have enough supervision to learn the task. This is reflected in both timestamp accuracy (P@0.5) and explanation quality (Score). In addition, for non-temporal questions (Q1 and Q2) in Figure 5, the model cannot properly answer and always use the answer format for reasoning temporal localization. By + +![](images/ec5d9a1307ac584ff42a3d6d64e111fddadff4af2fbd406011f09951818c4a94.jpg) +Fig. 5: Qualitative results to analyze the effect of training tasks. $\tau$ is the continuous time in seconds. The questions are not from ActivityNet-RTL as we are also interested in questions other than reasoning temporal localization. "RTL Only" answers in the format for reasoning temporal localization, as this is the only training data it has. "No NLVQA" further includes standard video tasks and can correctly answer non-temporal questions (Q1 and Q2). However, the answers are short due to the short answers in existing video question answering datasets. LITA improves both reasoning and natural language conversation by utilizing NLVQA datasets. + +adding standard video tasks in Video to training, "No NLVQA" improves all metrics compared to "RTL Only". Qualitatively in Figure 5, it can also answer Q1 and Q2 correctly. However, this capability mainly comes from the inclusion of video question answering datasets, which leads to short answers. LITA further improves by including NLVQA, which contains complex reasoning questions to improve its reasoning capabilities. More importantly, as shown in Figure 5, LITA now answers questions with natural language instead of short phrases. + +# 6 Conclusion + +We propose Language Instructed Temporal-Localization Assistant (LITA), which enables accurate temporal localization using Video LLMs. LITA demonstrates promising capabilities to answer complex temporal localization questions. At the same time, LITA substantially improves video-based text generation compared to existing Video LLMs even for non-temporal questions. This is the result of both our model design and data strategy. For model design, we propose time tokens to better represent the time and SlowFast tokens to efficiently process video inputs. Our experiments show the importance of these model components. For data strategy, we emphasize temporal localization data in training LITA. To achieve this, we propose the Reasoning Temporal Localization task and curate the ActivityNet-RTL dataset. Our results show that the inclusion of temporal localization data not only enables temporal localization for Video LLMs, but also improves the general video understanding capabilities. We further analyze our training data and show the benefits of incorporating standard video tasks and image instruction tuning data to train Video LLMs. + +# References + +1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: NeurIPS (2022) +2. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortsman, M., Schmidt, L.: OpenFlamingo: An open-source framework for training large autoregressive vision-language models. arXiv:2308.01390 (2023) +3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS (2020) +4. Caba Heilbron, F., Escorcia, V., Ghanem, B., Carlos Niebles, J.: Activitynet: A large-scale video benchmark for human activity understanding. In: CVPR (2015) +5. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: MiniGPT-v2: large language model as a unified interface for vision-language multi-task learning. arXiv:2310.09478 (2023) +6. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality (2023), https://lmsys.org/blog/2023-03-30-vicuna/ +7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: PaLM: Scaling language modeling with pathways. arXiv:2204.02311 (2022) +8. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructblip: Towards general-purpose vision-language models with instruction tuning (2023) +9. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: The EPIC-KITCHENS Dataset: Collection, challenges and baselines. IEEE Trans. PAMI 43(11), 4125-4141 (2021) +10. Damen, D., Doughty, H., Farinella, G.M., Furnari, A., Ma, J., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. IJCV 130, 33-55 (2022) +11. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: NAACL (2019) +12. Feichtenhofer, C., Fan, H., Malik, J., He, K.: SlowFast networks for video recognition. In: ICCV (2019) +13. Huang, B., Wang, X., Chen, H., Song, Z., Zhu, W.: Vtimellm: Empower llm to grasp video moments. arXiv preprint arXiv:2311.18445 (2023) +14. Huang, G., Pang, B., Zhu, Z., Rivera, C., Soricut, R.: Multimodal pretraining for dense video captioning. In: AACL-IJCNLP (2020) +15. Jiang, Y.G., Liu, J., Zamir, A.R., Toderici, G., Laptev, I., Shah, M., Sukthankar, R.: Thumos challenge: Action recognition with a large number of classes (2014), https://www.crcv.ucf.edu/THUMOS14/home.html +16. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Niebles, J.C.: Dense-captioning events in videos. In: ICCV (2017) +17. Kuehne, H., Arslan, A.B., Serre, T.: The language of actions: Recovering the syntax and semantics of goal-directed human activities. In: CVPR (2014) + +18. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: LISA: Reasoning segmentation via large language model. arXiv:2308.00692 (2023) +19. Li, B., Zhang, Y., Chen, L., Wang, J., Pu, F., Yang, J., Li, C., Liu, Z.: MIMIC-IT: Multi-modal in-context instruction tuning. arXiv:2306.05425 (2023) +20. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: VideoChat: Chat-centric video understanding. arXiv:2305.06355 (2023) +21. Li, Y., Wang, C., Jia, J.: Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043 (2023) +22. Lin, B., Zhu, B., Ye, Y., Ning, M., Jin, P., Yuan, L.: Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122 (2023) +23. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv:2310.03744 (2023) +24. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023) +25. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability (2023) +26. Miech, A., Zhukov, D., Alayrac, J.B., Tapaswi, M., Laptev, I., Sivic, J.: HowTo100M: Learning a text-video embedding by watching hundred million narrated video clips. In: ICCV (2019) +27. Muhammad Maaz, Hanoona Rasheed, S.K., Khan, F.: Video-ChatGPT: Towards detailed video understanding via large vision and language models. ArXiv 2306.05424 (2023) +28. OpenAI: GPT-4 technical report. arXiv:2303.08774 (2023) +29. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv:2306.14824 (2023) +30. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) +31. Sigurdsson, G.A., Varol, G., Wang, X., Farhadi, A., Laptev, I., Gupta, A.: Hollywood in homes: Crowdsourcing data collection for activity understanding. In: ECCV (2016) +32. Tang, Y., Ding, D., Rao, Y., Zheng, Y., Zhang, D., Zhao, L., Lu, J., Zhou, J.: COIN: A large-scale dataset for comprehensive instructional video analysis. In: CVPR (2019) +33. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: LLaMA: Open and efficient foundation language models. arXiv:2302.13971 (2023) +34. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: LLaMA 2: Open foundation and fine-tuned chat models. arXiv:2307.09288 (2023) +35. Wu, S., Fei, H., Qu, L., Ji, W., Chua, T.S.: NExT-GPT: Any-to-any multimodal LLM. arXiv:2309.05519 (2023) +36. Xiao, J., Shang, X., Yao, A., Chua, T.S.: NExT-QA: Next phase of question-answering to explaining temporal actions. In: CVPR (2021) +37. Yan, S., Xiong, X., Nagrani, A., Arnab, A., Wang, Z., Ge, W., Ross, D., Schmid, C.: UnLoc: A unified framework for video localization tasks. In: ICCV (2023) +38. Yang, A., Nagrani, A., Seo, P.H., Miech, A., Pont-Tuset, J., Laptev, I., Sivic, J., Schmid, C.: Vid2Seq: Large-scale pretraining of a visual language model for dense video captioning. In: CVPR (2023) + +39. Zhang, H., Li, X., Bing, L.: Video-LLaMA: An instruction-tuned audio-visual language model for video understanding. arXiv:2306.02858 (2023) +40. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: LLaMA-Adapter: Efficient fine-tuning of language models with zero-init attention. arXiv:2303.16199 (2023) +41. Zhang, Y., Li, Y., Cui, L., Cai, D., Liu, L., Fu, T., Huang, X., Zhao, E., Zhang, Y., Chen, Y., Wang, L., Luu, A.T., Bi, W., Shi, F., Shi, S.: Siren's song in the ai ocean: A survey on hallucination in large language models. arXiv:2309.01219 (2023) +42. Zhou, L., Xu, C., Corso, J.J.: Towards automatic learning of procedures from web instructional videos. In: AAAI (2018) +43. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: MiniGPT-4: Enhancing vision-language understanding with advanced large language models. arXiv:2304.10592 (2023) \ No newline at end of file diff --git a/2024/LITA_ Language Instructed Temporal-Localization Assistant/images.zip b/2024/LITA_ Language Instructed Temporal-Localization Assistant/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..25965ff51890bdfd8016336c605fac2982dc2ae3 --- /dev/null +++ b/2024/LITA_ Language Instructed Temporal-Localization Assistant/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aaccad2c4aac295d0bfc92823a0cef895443973fedc3b4bdfee92d7283f017e6 +size 274662 diff --git a/2024/LITA_ Language Instructed Temporal-Localization Assistant/layout.json b/2024/LITA_ Language Instructed Temporal-Localization Assistant/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ce2d64b7698a6ac3cc3ae7116693ad16504d9faf --- /dev/null +++ b/2024/LITA_ Language Instructed Temporal-Localization Assistant/layout.json @@ -0,0 +1,7872 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 194, + 112, + 420, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 112, + 420, + 148 + ], + "spans": [ + { + "bbox": [ + 194, + 112, + 420, + 148 + ], + "type": "text", + "content": "LITA: Language Instructed Temporal-Localization Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 156, + 168, + 459, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 168, + 459, + 193 + ], + "spans": [ + { + "bbox": [ + 156, + 168, + 459, + 193 + ], + "type": "text", + "content": "De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 288, + 202, + 324, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 202, + 324, + 213 + ], + "spans": [ + { + "bbox": [ + 288, + 202, + 324, + 213 + ], + "type": "text", + "content": "NVIDIA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 240, + 452, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 240, + 452, + 493 + ], + "spans": [ + { + "bbox": [ + 160, + 240, + 452, + 493 + ], + "type": "text", + "content": "Abstract. There has been tremendous progress in multimodal Large Language Models (LLMs). Recent works have extended these models to video input with promising instruction following capabilities. However, an important missing piece is temporal localization. These models cannot accurately answer the \"When?\" questions. We identify three key aspects that limit their temporal localization capabilities: (i) time representation, (ii) architecture, and (iii) data. We address these shortcomings by proposing Language Instructed Temporal-Localization Assistant (LITA) with the following features: (1) We introduce time tokens that encode timestamps relative to the video length to better represent time in videos. (2) We introduce SlowFast tokens in the architecture to capture temporal information at fine temporal resolution. (3) We emphasize temporal localization data for LITA. In addition to leveraging existing video datasets with timestamps, we propose a new task, Reasoning Temporal Localization (RTL), along with the dataset, ActivityNet-RTL, for learning and evaluating this task. Reasoning temporal localization requires both the reasoning and temporal localization of Video LLMs. LITA demonstrates strong performance on this challenging task, nearly doubling the temporal mean intersection-over-union (mIoU) of baselines. In addition, we show that our emphasis on temporal localization also substantially improves video-based text generation compared to existing Video LLMs, including a " + }, + { + "bbox": [ + 160, + 240, + 452, + 493 + ], + "type": "inline_equation", + "content": "36\\%" + }, + { + "bbox": [ + 160, + 240, + 452, + 493 + ], + "type": "text", + "content": " relative improvement of Temporal Understanding. Code is available at: https://github.com/NVlabs/LITA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 503, + 453, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 503, + 453, + 525 + ], + "spans": [ + { + "bbox": [ + 159, + 503, + 453, + 525 + ], + "type": "text", + "content": "Keywords: Large Language Models " + }, + { + "bbox": [ + 159, + 503, + 453, + 525 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 503, + 453, + 525 + ], + "type": "text", + "content": " Video Understanding " + }, + { + "bbox": [ + 159, + 503, + 453, + 525 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 503, + 453, + 525 + ], + "type": "text", + "content": " Temporal Localization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 545, + 230, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 230, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 230, + 558 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 629 + ], + "type": "text", + "content": "Large language models (LLMs) [3,7,11,28,33,34] have demonstrated impressive instruction following capabilities, and shown that language can be a universal interface for various tasks [7,28]. These models can be further extended to multimodal LLMs to process language and other modalities, such as image, video, and audio [1,24,43]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "While most multimodal LLMs focus on images for visual content, several recent works introduce models that specialize in processing videos [20,25,27,39]. These Video LLMs preserve the instruction following capabilities of LLMs and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 306, + 166 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 306, + 166 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 306, + 166 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 306, + 166 + ], + "type": "image", + "image_path": "ab78a62d5276cf97160c0ecb4374363d597abe3a15cb26e164e018518969e32a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 125, + 471, + 158 + ], + "lines": [ + { + "bbox": [ + 310, + 125, + 471, + 158 + ], + "spans": [ + { + "bbox": [ + 310, + 125, + 471, + 158 + ], + "type": "text", + "content": "A: Her dance is the most energetic between 7.37s and 12.81s when she does a handspring, which is more energetic than standing and lying on the floor." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 130, + 177, + 482, + 222 + ], + "lines": [ + { + "bbox": [ + 130, + 177, + 482, + 222 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 482, + 222 + ], + "type": "text", + "content": "Fig. 1: Example to illustrate our proposed Reasoning Temporal Localization (RTL). Instead of directly querying about an event, questions in RTL require further reasoning to answer. Here, the model needs to compare all activities in the video to find the timestamps of the most energetic activity (i.e., handspring)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 246, + 479, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 479, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 479, + 329 + ], + "type": "text", + "content": "allow users to ask various questions about a given video. However, one important missing piece in these Video LLMs is temporal localization. When prompted with the \"When?\" questions, these models cannot accurately localize time periods, and often hallucinate irrelevant information [41]. Temporal localization is an important component that differentiates videos from images, and has been widely studied outside the context of instruction following LLMs [4,10,15,31]. It is thus crucial for Video LLMs to have temporal localization capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 330, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 480, + 472 + ], + "type": "text", + "content": "We identify three key aspects that limit the temporal localization capabilities of existing Video LLMs: time representation, architecture, and data. First, existing models often represent timestamps as plain text (e.g. 01:22 or 142sec). However, given a set of frames, the correct timestamp still depends on the frame rate, which the model does not have access to. This makes learning temporal localization harder. Second, the architecture of existing Video LLMs might not have sufficient temporal resolution to interpolate time information accurately. For example, Video-LLaMA [39] only uniformly samples 8 frames from the entire video, which is insufficient for accurate temporal localization. Finally, temporal localization is largely ignored in the data used by existing Video LLMs. Data with timestamps are only a small subset of video instruction tuning data, and the accuracy of these timestamps is also not verified." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 480, + 581 + ], + "type": "text", + "content": "Our Approach. We address the aforementioned shortcomings of existing Video LLMs, and propose Language Instructed Temporal-Localization Assistant (LITA): (1) Time Representation: We introduce time tokens to represent relative timestamps and allow Video LLMs to better communicate about time than using plain text. (2) Architecture: We introduce SlowFast tokens to capture temporal information at fine temporal resolution to enable accurate temporal localization. (3) Data: We emphasize temporal localization data for LITA. We propose a new task, Reasoning Temporal Localization (RTL), along with the dataset, ActivityNet-RTL, for learning this task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "The first important design of LITA is to use relative representation for time (e.g. first " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " of the video) instead of the absolute time representation with plain text (e.g. 01:22). We divide a given video into " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " equal length chunks, and introduce " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " time tokens " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\langle 1\\rangle" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\langle \\mathbf{T}\\rangle" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " to represent the relative time location in the video. During training and inference, these time tokens can be easily encoded and decoded from plain text timestamps given the length of the video. The start and end timestamps are well-defined by the time tokens given only the input" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "D.-A. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "video. This is in contrast to plain text timestamps. Without the frame rate, the correct absolute timestamp is ill-defined given just the video frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 307 + ], + "type": "text", + "content": "The second important design of LITA is to use densely sampled input frames from videos. It is unlikely to achieve accurate temporal localization with only sparsely sampled frames. The challenge is that the LLM module inside Video LLMs cannot naively process large numbers of frames simultaneously due to context length limitation. Take LLaVA [24] as an example. Each image is converted to 256 tokens, which are then fed into its LLM module as input. If we directly feed 100 frames to the LLM module, then that is " + }, + { + "bbox": [ + 130, + 140, + 481, + 307 + ], + "type": "inline_equation", + "content": "256 \\times 100 = 25600" + }, + { + "bbox": [ + 130, + 140, + 481, + 307 + ], + "type": "text", + "content": " tokens, which is already over the max context length for some LLMs [6,34]. Inspired by the SlowFast architecture for videos [12], we instead consider two types of tokens, fast tokens and slow tokens, to address this efficiency issue. We generate fast tokens at a high temporal resolution to provide the temporal information, while keeping the tokens per frame at a low number. On the other hand, we generate slow tokens at a low temporal resolution, which allows us to use a higher number of tokens per frame to provide the spatial information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 481, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 481, + 475 + ], + "type": "text", + "content": "Finally, We emphasize temporal localization data for LITA. We include dense video captioning [16] and event localization [37] in instruction tuning of LITA. These tasks include human annotated timestamps to promote accurate temporal localization. In addition to leveraging existing data and tasks, we further propose a new task, Reasoning Temporal Localization (RTL), along with the dataset, ActivityNet-RTL, for training and evaluating this task. Answers to RTL questions can only be derived by utilizing world knowledge and temporal reasoning. Figure 1 shows an example. To answer the question: \"When does the woman's dance become the most energetic?\" the model needs to first recognize the woman's dance moves in the video, then reason about the most active part, and finally temporally localize the event (i.e. handspring). In addition to the predicted timestamps, we further consider the explanation provided by the model. Thus our new task not only assesses temporal understanding, but also requires strong reasoning capabilities that are unique to LLMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 475, + 481, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 481, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 481, + 571 + ], + "type": "text", + "content": "For the challenging RTL task, LITA doubles baseline's performance for temporal metrics (mIOU, Precision@0.5), while providing much better explanations. In addition to enabling accurate temporal localization, we show that our emphasis on temporal understanding also improves LITA's core Video LLM capabilities. LITA substantially improves all scores on a benchmark for video-based question answering [27]. This includes a " + }, + { + "bbox": [ + 130, + 475, + 481, + 571 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 130, + 475, + 481, + 571 + ], + "type": "text", + "content": " relative improvement for Correctness of Information, and a " + }, + { + "bbox": [ + 130, + 475, + 481, + 571 + ], + "type": "inline_equation", + "content": "36\\%" + }, + { + "bbox": [ + 130, + 475, + 481, + 571 + ], + "type": "text", + "content": " relative improvement for Temporal Understanding compared to existing Video LLMs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 590, + 237, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 237, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 237, + 604 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Multimodal Large Language Models. Large language models (LLMs) [7,28] inspire recent works to address multimodal tasks by leveraging LLMs [35]. Some approaches add additional parameters inside LLMs, such as gated cross-attention layers [1,2,19] or adapter layers [40], to adapt it to process multimodal inputs." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "LITA: Language Instructed Temporal-Localization Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "Several works, on the other hand, only use modules, such as projection layers or Q-Formers, to project outputs of visual encoders to the input space of LLMs [8, 24,43]. Recent works further expand multimodal LLM to visual grounding tasks, such as detection [5,23,29] and segmentation [18]. The most related to ours is LISA [18], which extends referring segmentation to reasoning segmentation. We share the same spirit and propose Reasoning Temporal Localization to jointly evaluate reasoning and temporal understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 202, + 482, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 202, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 482, + 323 + ], + "type": "text", + "content": "Video Large Language Models. Building on the success of multimodal LLMs, several works extend image LLMs to Video LLMs [20, 25, 27, 39]. These works mainly use the approach of projecting visual tokens to LLMs' input space using projection layers [25, 27] or Q-Formers [20, 39]. While these models show promise in descriptive questions and instructions, they still lack temporal localization capabilities. LITA is designed to address this shortcoming, while also improving downstream video tasks. Concurrent works [13, 21, 22] include improvements like better image-video joint training. The most related concurrent work to ours is VTimeLLM [13], which also aims to address temporal localization of Video LLMs. We further introduce the reasoning aspect to temporal localization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 325, + 482, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 421 + ], + "type": "text", + "content": "Temporal Localization in Videos. The goal of temporal localization is to pinpoint activities within untrimmed video sequences on a temporal scale [37]. The target activities can be predefined action classes [9, 15] or events described by natural language [4, 42]. Our goal of video temporal understanding is also related to various video tasks, such as dense video captioning [14, 16, 38] and action segmentation [17, 26, 32]. Models for these temporal tasks can have quite different design. Instruction following Video LLMs like LITA provide a way to unify these frameworks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 438, + 411, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 438, + 411, + 453 + ], + "spans": [ + { + "bbox": [ + 131, + 438, + 411, + 453 + ], + "type": "text", + "content": "3 Language Instructed Temporal-Localization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 522 + ], + "type": "text", + "content": "LITA enables temporal localization for Video LLMs by: (1) relative time representation with the time tokens, (2) SlowFast tokens to capture temporal information at fine temporal resolution, (3) multi-task training that includes accurate timestamps. We will first introduce the overall architecture and discuss further details of individual components." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 539, + 226, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 539, + 226, + 551 + ], + "spans": [ + { + "bbox": [ + 131, + 539, + 226, + 551 + ], + "type": "text", + "content": "3.1 Architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 605 + ], + "type": "text", + "content": "An overview of LITA is shown in Figure 2. We build on Image LLMs. In particular, we select LLaVA due to its simplicity and effectiveness [24]. Note that LITA does not depend on the specific underlying Image LLM architecture and can be easily adapted to other base architectures." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": "Given a video, we first uniformly select " + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": " frames and encode each frame into " + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": " tokens. " + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": " should be large enough to support the desired granularity of temporal localization. " + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "inline_equation", + "content": "T \\times M" + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": " is a large number that usually cannot be directly processed by the LLM module. Thus, we then use SlowFast pooling to reduce the " + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "inline_equation", + "content": "T \\times M" + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": " tokens to " + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "inline_equation", + "content": "T + M" + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": " tokens." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "text", + "content": "D.-A. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 135, + 118, + 472, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 118, + 472, + 128 + ], + "spans": [ + { + "bbox": [ + 135, + 118, + 472, + 128 + ], + "type": "text", + "content": "A1: She is dancing from " + }, + { + "bbox": [ + 135, + 118, + 472, + 128 + ], + "type": "inline_equation", + "content": "<2>" + }, + { + "bbox": [ + 135, + 118, + 472, + 128 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 135, + 118, + 472, + 128 + ], + "type": "inline_equation", + "content": "<3>" + }, + { + "bbox": [ + 135, + 118, + 472, + 128 + ], + "type": "text", + "content": ". A2: Her cloth is black. A3: " + }, + { + "bbox": [ + 135, + 118, + 472, + 128 + ], + "type": "inline_equation", + "content": "<1><2>" + }, + { + "bbox": [ + 135, + 118, + 472, + 128 + ], + "type": "text", + "content": " She is standing. " + }, + { + "bbox": [ + 135, + 118, + 472, + 128 + ], + "type": "inline_equation", + "content": "<2><3>" + }, + { + "bbox": [ + 135, + 118, + 472, + 128 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 136, + 130, + 473, + 232 + ], + "blocks": [ + { + "bbox": [ + 136, + 130, + 473, + 232 + ], + "lines": [ + { + "bbox": [ + 136, + 130, + 473, + 232 + ], + "spans": [ + { + "bbox": [ + 136, + 130, + 473, + 232 + ], + "type": "image", + "image_path": "ff7a7f18c17206da75583469eada7ab33bc207ab2aaf74ea4df304eb4721f20d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 240, + 482, + 328 + ], + "lines": [ + { + "bbox": [ + 130, + 240, + 482, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 482, + 328 + ], + "type": "text", + "content": "Fig. 2: Overview of LITA. The input video frames are first encoded into visual tokens (numbered by frame), which are further processed by two pathways. The Fast Token pathway averages all the tokens in a frame to maintain a high temporal resolution. The Slow Token pathway sparsely samples frames to maintain a larger number of tokens per frame to provide spatial information. Timestamps are converted to time tokens " + }, + { + "bbox": [ + 130, + 240, + 482, + 328 + ], + "type": "inline_equation", + "content": "<1>" + }, + { + "bbox": [ + 130, + 240, + 482, + 328 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 240, + 482, + 328 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 130, + 240, + 482, + 328 + ], + "type": "text", + "content": ". This is important for better temporal localization learning. Various video tasks on the right can be converted to natural language visual question answering (Q1-3 and A1-3) to jointly optimize LITA." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 354, + 483, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 483, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 483, + 485 + ], + "type": "text", + "content": "The slow and fast tokens are projected by a linear layer and concatenated with the text tokens to use as input to the LLM module. The text tokens (prompt) are processed to convert any referenced timestamps to specialized time tokens (" + }, + { + "bbox": [ + 130, + 354, + 483, + 485 + ], + "type": "inline_equation", + "content": "<1>" + }, + { + "bbox": [ + 130, + 354, + 483, + 485 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 354, + 483, + 485 + ], + "type": "inline_equation", + "content": "<\\mathrm{T}>" + }, + { + "bbox": [ + 130, + 354, + 483, + 485 + ], + "type": "text", + "content": "). All the input tokens are then jointly processed by the LLM module sequentially. We fine-tune the entire model with our reasoning temporal localization data (Section 4) along with other video tasks, such as dense video captioning and event localization. LITA learns to use time tokens instead of absolute timestamps. For temporal localization, we can then directly ask LITA the \"When\" questions (e.g. \"When is she dancing?\"). LITA would respond with time tokens (e.g. \"She is dancing from " + }, + { + "bbox": [ + 130, + 354, + 483, + 485 + ], + "type": "inline_equation", + "content": "<2>" + }, + { + "bbox": [ + 130, + 354, + 483, + 485 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 354, + 483, + 485 + ], + "type": "inline_equation", + "content": "<3>" + }, + { + "bbox": [ + 130, + 354, + 483, + 485 + ], + "type": "text", + "content": "). which can then be converted to timestamps given the video length." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 509, + 227, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 227, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 227, + 520 + ], + "type": "text", + "content": "3.2 Time Tokens" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "We use a relative time representation instead of absolute timestamps in LITA. As shown in Figure 2, the LLM module can only see the visual tokens (slow and fast) and the language tokens (text prompt). There is not enough information in this input space for the LLM module to infer the absolute timestamp because the frame rate is not known to the model in advance. A better way is to represent timestamps relative to the video length, thus removing the dependency on the frame rate. We divide the video into " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " chunks and use " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " specialized time tokens " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "<1>" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "<\\mathsf{T}>" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " for timestamps. Given a continuous timestamp " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " and video length " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " can be easily converted to time token " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "<\\mathsf{t}>" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "t = \\mathrm{round}(\\tau(T-1)/L) + 1" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", and conversely " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "<\\mathsf{t}>" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " can be converted back to " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tau = L(t-1)/(T-1)" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ". While this does introduce discretization error, it greatly simplifies the time representation" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "text", + "content": "LITA: Language Instructed Temporal-Localization Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "with LLMs. Relative timestamp is also used in other temporally heavy video tasks, such as dense video captioning [38]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": "Given this time representation, many video tasks related to temporal localization can be transformed into language instructions and answers. For example, dense video captioning can be achieved by prompting the model with \"Describe the video. Each sentence begins with start and end timestamps.\" (Q3 and A3 in Fig. 2). Standard event localization is also transformed to \"When does X happen?\" (Q1 and A1 in Fig. 2). We also incorporate standard video question answering (Q2 and A2 in Fig. 2). More details are discussed in Section 3.4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 243, + 279, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 243, + 279, + 254 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 279, + 254 + ], + "type": "text", + "content": "3.3 SlowFast Visual Tokens" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 262, + 482, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 346 + ], + "type": "text", + "content": "We have discussed how we discretize time in videos into " + }, + { + "bbox": [ + 130, + 262, + 482, + 346 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 262, + 482, + 346 + ], + "type": "text", + "content": " steps in order to make Video LLMs better at reasoning about time. Still, the visual input should match the temporal resolution " + }, + { + "bbox": [ + 130, + 262, + 482, + 346 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 262, + 482, + 346 + ], + "type": "text", + "content": " in order to achieve effective temporal processing. Ideally, one would need at least " + }, + { + "bbox": [ + 130, + 262, + 482, + 346 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 262, + 482, + 346 + ], + "type": "text", + "content": " frames to temporally localize events with the resolution " + }, + { + "bbox": [ + 130, + 262, + 482, + 346 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 262, + 482, + 346 + ], + "type": "text", + "content": ". However, naively feeding all " + }, + { + "bbox": [ + 130, + 262, + 482, + 346 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 262, + 482, + 346 + ], + "type": "text", + "content": " frames into the LLM module could be computationally prohibitive. In our experiment, we use " + }, + { + "bbox": [ + 130, + 262, + 482, + 346 + ], + "type": "inline_equation", + "content": "T = 100" + }, + { + "bbox": [ + 130, + 262, + 482, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 262, + 482, + 346 + ], + "type": "inline_equation", + "content": "M = 256" + }, + { + "bbox": [ + 130, + 262, + 482, + 346 + ], + "type": "text", + "content": " (CLIP ViT-L-14 [30]). This leads to 25600 tokens per video." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": "Inspired by SlowFast models [12], we consider two pathways to pool the " + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "T \\times M" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": " tokens for " + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": " frames. The first is densely sampled fast tokens to provide temporal information. We obtain " + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": " fast tokens from " + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": " frames by averaging all the tokens belonging to the same frame. The second is the sparsely sampled slow tokens to maintain better spatial information. We select a spatial downsampling ratio of " + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": ", and uniformly select " + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "s^2" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": " frames from the video. For each selected frame, we perform a " + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "s \\times s" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": " spatial average pooling to the " + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": " tokens, which lead to " + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\frac{M}{s^2}" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": " slow tokens per frame. This leads to a total " + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "M = \\frac{M}{s^2} \\times s^2" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": " slow tokens. We use " + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "s = 2" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": " in our experiments. This leads to a total of " + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "T + M" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": " tokens to represent a video instead of " + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "inline_equation", + "content": "T \\times M" + }, + { + "bbox": [ + 130, + 347, + 482, + 468 + ], + "type": "text", + "content": " tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 486, + 235, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 486, + 235, + 498 + ], + "spans": [ + { + "bbox": [ + 131, + 486, + 235, + 498 + ], + "type": "text", + "content": "3.4 Training Tasks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 506, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 578 + ], + "type": "text", + "content": "In addition to architecture, training tasks and data also play an important role for LLMs. We emphasize temporal localization data and train LITA with the following five tasks: (1) dense video captioning [16], (2) event localization [37], (3) video question answering [36], (4) natural language visual question answering [24], and (5) our proposed reasoning temporal localization. Temporal localization is a crucial component for three out of the five tasks (1, 2, and 5)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 578, + 482, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 482, + 603 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 482, + 603 + ], + "type": "text", + "content": "We now introduce each task in order. The first three tasks are standard video tasks and equip LITA with basic video understanding:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Dense Video Captioning. In dense video captioning [16], each video is described by a set of sentences, and each sentence comes with the start and end timestamps of the event. Each sentence in dense video captioning can thus be represented as: SENTENCE. We then sort all sentences by its start time, and directly concatenate all sentences and timestamps. One" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "text", + "content": "D.-A. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "content": "example prompt to the model for this task is: \"Provide a detailed description of the given video. Each sentence should begin with the start and end timestamps.\" Other prompts are included in the supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 155, + 481, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 155, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 155, + 481, + 204 + ], + "type": "text", + "content": "Event Localization. In event localization, the goal is to temporally localize the event described by a sentence. We use a simple answer format: . One example prompt for this task is: \"When does \"SENTENCE\" happen in the video? Answer the question only using start and end timestamps.\"" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 481, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 481, + 289 + ], + "type": "text", + "content": "Video Question Answering. The question answering task is already represented as language instructions. However, answers in existing question answering datasets often consist of a single word or phrase because models for this task might not be able to generate longer text. We follow Liu et al. [23] and append the following prompt to the question: \"Answer the question using a single word or phrase.\" The goal is to provide the context for short answers so that it affects the model's text generation less." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 293, + 481, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 481, + 388 + ], + "type": "text", + "content": "Natural Language Visual Question Answering. Training with the above three tasks gives LITA video understanding capabilities. However, we observe that models trained with only these tasks often provide short answers and lack natural language conversation capabilities. We thus also train LITA with natural language visual question answering or visual instruction tuning datasets [24]. The goal is to improve the natural language conversation of LITA. We find that mixing instruction tuning datasets [24] with standard video tasks improves LITA's conversation quality while maintaining good video understanding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 391, + 481, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 481, + 475 + ], + "type": "text", + "content": "Reasoning Temporal Localization. Finally, we also train LITA with our reasoning temporal localization task (details in Section 4). The answer to a reasoning temporal localization question consists of two parts: timestamps and explanation. We find it challenging for models to simultaneously output both of them without any example. Nevertheless, with some training data, LITA quickly pick up reasoning and temporal localization, and provide both the timestamps and explanation of its reasoning in answers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 494, + 350, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 494, + 350, + 508 + ], + "spans": [ + { + "bbox": [ + 131, + 494, + 350, + 508 + ], + "type": "text", + "content": "4 Reasoning Temporal Localization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 519, + 481, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 481, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 481, + 591 + ], + "type": "text", + "content": "We now discuss further details of the proposed Reasoning Temporal Localization (RTL) task. Standard temporal localization does not fully leverage the potential of Video LLMs. One impressive aspect of LLMs is its reasoning capabilities. LLMs can even answer complex questions that involve multi-step reasoning. Therefore, we propose the RTL task to utilize both of Video LLMs' temporal understanding and reasoning capabilities." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 609, + 257, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 257, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 257, + 621 + ], + "type": "text", + "content": "4.1 Problem Definition" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "In reasoning temporal localization, the query is still a \"when\" question that asks about the start and end timestamps of an event. The key difference compared to the standard temporal localization task is that the target event is not directly" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "LITA: Language Instructed Temporal-Localization Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 246, + 178 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 246, + 178 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 246, + 178 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 246, + 178 + ], + "type": "image", + "image_path": "b5a6e94588981e6b0cd9296e5371fb8db5061a42b43885e657f5144cb7992af1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 180, + 218, + 187 + ], + "lines": [ + { + "bbox": [ + 162, + 180, + 218, + 187 + ], + "spans": [ + { + "bbox": [ + 162, + 180, + 218, + 187 + ], + "type": "text", + "content": "(a) Temporal Reasoning" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 114, + 361, + 178 + ], + "blocks": [ + { + "bbox": [ + 250, + 114, + 361, + 178 + ], + "lines": [ + { + "bbox": [ + 250, + 114, + 361, + 178 + ], + "spans": [ + { + "bbox": [ + 250, + 114, + 361, + 178 + ], + "type": "image", + "image_path": "a395746009371616fd09e92cecf11e28819650b81d679f742a3dc2af14efdbb3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 180, + 331, + 187 + ], + "lines": [ + { + "bbox": [ + 280, + 180, + 331, + 187 + ], + "spans": [ + { + "bbox": [ + 280, + 180, + 331, + 187 + ], + "type": "text", + "content": "(b) World Knowledge" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 199, + 482, + 255 + ], + "lines": [ + { + "bbox": [ + 130, + 199, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 199, + 482, + 255 + ], + "type": "text", + "content": "Fig. 3: Examples from our ActivityNet-RTL dataset. RTL questions ask about events that are not explicitly described. The model needs to utilize reasoning or its world knowledge to answer. This is in contrast to standard temporal localization, which directly asks about the event of interest. For example in (c), standard temporal localization might directly ask: \"when does the man sit on the floor?\"" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 365, + 114, + 477, + 178 + ], + "blocks": [ + { + "bbox": [ + 365, + 114, + 477, + 178 + ], + "lines": [ + { + "bbox": [ + 365, + 114, + 477, + 178 + ], + "spans": [ + { + "bbox": [ + 365, + 114, + 477, + 178 + ], + "type": "image", + "image_path": "ec4c2a7996f42e3f5bab8ece02e17acfe4a72fb0c095be728221009526a24d1a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 180, + 448, + 187 + ], + "lines": [ + { + "bbox": [ + 394, + 180, + 448, + 187 + ], + "spans": [ + { + "bbox": [ + 394, + 180, + 448, + 187 + ], + "type": "text", + "content": "(c) Complex Reasoning" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 278, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 327 + ], + "type": "text", + "content": "described in the question, and can only be inferred by reasoning and using world knowledge of the model. The answer to such a question thus consists of two parts: (1) the start and end timestamps of the target event, and (2) an explanation of the reasoning process the model goes through to derive the timestamps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 327, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 471 + ], + "type": "text", + "content": "Some examples are shown in Figure 3 to better illustrate the idea. The answer format is: [start end] Explanation. In Figure 3(a), the model not only has to localize \"adjust their position to avoid obstacles,\" but also needs to temporally reason about which instance happened earlier in the video. In Figure 3(b), instead of directly asking about \"one-leg row,\" it asks about the workout targeting balance and stability. The model thus needs to utilize its knowledge of what kind of exercises are good for balance and stability. Finally, there are questions that require multi-step reasoning. In Figure 3(c), the question asks about \"the most atypical ways of jump roping,\" which requires the model to understand what is typical and atypical for jump roping, and then temporally find the most atypical time period. A standard temporal localization task, in contrast, would just ask when the man is sitting on the floor." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 490, + 289, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 289, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 289, + 502 + ], + "type": "text", + "content": "4.2 ActivityNet-RTL Dataset" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 510, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 567 + ], + "type": "text", + "content": "The above examples are from ActivityNet-RTL, a dataset curated by us for the Reasoning Temporal Localization (RTL) task. We build our dataset from the ActivityNet Captions dataset [16], which annotates multiple events described by sentences in a video, and all the events are temporally localized with start and end timestamps. Consider the following toy example:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 575, + 294, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 575, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 575, + 294, + 586 + ], + "type": "text", + "content": "[00:00-00:10] A woman is standing." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 587, + 298, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 587, + 298, + 597 + ], + "spans": [ + { + "bbox": [ + 134, + 587, + 298, + 597 + ], + "type": "text", + "content": "[00:12-00:30] The woman is dancing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 598, + 302, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 598, + 302, + 609 + ], + "spans": [ + { + "bbox": [ + 134, + 598, + 302, + 609 + ], + "type": "text", + "content": "[00:32-00:36] The woman is sleeping." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "We then use this as context and ask GPT-4 to generate temporal localization questions that require further reasoning to answer. We also ask GPT-4 to simultaneously generate the answer that includes the queried start and end timestamps, along with the explanation about the reasoning process." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "text", + "content": "D.-A. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "Using the above toy example, by seeing that the woman has done three activities, one possible reasoning temporal localization question is to ask \"When is she the least active?\" Since sleeping is the least active out of the three activities, the target time period is 00:32-00:36, the period when she is sleeping. This illustrates how GPT-4 can still generate interesting questions without seeing the actual video. We annotate few-shot examples for GPT-4 as in previous works to improve the generation quality [24]. All of our prompts are included in the supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 217, + 482, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 338 + ], + "type": "text", + "content": "Training Set Generation. For our training set, we directly use the results generated by GPT-4 with 10,009 videos from the training split of ActivityNet-Captions. This leads to 33,557 Reasoning Temporal Localization question-answer pairs. By inspecting the GPT generated results, we find that most of the questions are valid temporal localization questions given the context. The main shortcoming is that not all question-answer pairs require reasoning. Sometimes GPT-4 generates questions that directly ask about events that are already described by the dense video captions. However, we do hope that LITA can also answer these standard temporal localization questions correctly using natural language. We thus leave these questions in the training set." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 341, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 413 + ], + "type": "text", + "content": "Evaluation Set Curation. On the other hand, the evaluation set requires manual efforts otherwise we would end up with many non-reasoning questions. We start from the GPT-4 generated questions using a subset of the ActivityNet-Captions validation set, and manually remove non-reasoning questions. We also double check the timestamps and explanations in the answers. This leads to a total of 229 question-answer pairs for 160 videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 441, + 200, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 200, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 200, + 453 + ], + "type": "text", + "content": "4.3 Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 471, + 499, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 499, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 499, + 568 + ], + "type": "text", + "content": "We consider three metrics for ActivityNet-RTL: mIOU, Precision@0.5, and GPT-4 Relative Scores. The first two metrics are for temporal localization, and the third metric is to evaluate the explanation capability. mIOU averages the intersection-over-union (IOU) between predicted and groundtruth start and end timestamps. Precision@0.5 measures the percentage of predictions that have over 0.5 IOU. We first average these two metrics per video, and then average over all videos in the evaluation set. This avoids over-weighting videos and time periods with more questions, as some time periods in a video have multiple questions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "To evaluate the quality of the explanation, we follow the evaluation pipeline of LLaVA [24] and leverage GPT-4 for evaluation. GPT-4 is asked to evaluate the helpfulness, relevance, accuracy, and level of details of the explanations, and give a score from 1 to 10. We ask GPT-4 to evaluate both the predicted and groundtruth explanations, and normalize the score for the prediction by the score of the groundtruth. For this metric, we directly average over all question-answer pairs as the explanations could be quite different even for questions about the same time period in the same video." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "LITA: Language Instructed Temporal-Localization Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "content": "We evaluate LITA with both temporal localization and video tasks that do not involve temporal localization because most existing Video LLMs cannot handle temporal localization. In addition to our proposed Reasoning Temporal Localization, we further evaluate LITA on Video-based Text Generation Performance Benchmarking proposed by Maaz et al. [27]. This provides a holistic evaluation of LITA as a Video LLM and not just for temporal localization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 228, + 279, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 279, + 241 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 279, + 241 + ], + "type": "text", + "content": "5.1 Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 248, + 481, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 481, + 321 + ], + "type": "text", + "content": "Architecture. We uniformly sample 100 frames from a video, and use 100 time tokens " + }, + { + "bbox": [ + 130, + 248, + 481, + 321 + ], + "type": "inline_equation", + "content": "<1>" + }, + { + "bbox": [ + 130, + 248, + 481, + 321 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 248, + 481, + 321 + ], + "type": "inline_equation", + "content": "<100>" + }, + { + "bbox": [ + 130, + 248, + 481, + 321 + ], + "type": "text", + "content": " to represent timestamps. We use CLIP-L-14 [30] as the visual encoder, and Vicuna [6] as the LLM module. We follow LLaVA's architecture and train a single linear layer for projection [24]. We use 4 frames for slow tokens and use average pool window " + }, + { + "bbox": [ + 130, + 248, + 481, + 321 + ], + "type": "inline_equation", + "content": "s = 2" + }, + { + "bbox": [ + 130, + 248, + 481, + 321 + ], + "type": "text", + "content": ". With 1 fast token per frame, this leads to a total of " + }, + { + "bbox": [ + 130, + 248, + 481, + 321 + ], + "type": "inline_equation", + "content": "100 + \\frac{256}{4} \\times 4 = 356" + }, + { + "bbox": [ + 130, + 248, + 481, + 321 + ], + "type": "text", + "content": " tokens per video." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 323, + 481, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 481, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 481, + 431 + ], + "type": "text", + "content": "Training Datasets. We discussed training tasks in Section 3.4. We now discuss the training datasets for each task. For dense video captioning and event localization, we use the training splits of ActivityNet-Captions [16] and YouCook2 [42], which combine to around 11k videos. The event localization dataset can be generated from the dense video captioning dataset by using the caption as query and the timestamps as target. For video question answering, we use NExT-QA [36] as it contains more complex questions. For image instruction tuning, we use LLaVA-150K [24]. For reasoning temporal localization, we use the training split of our ActivityNet-RTL, which builds on the training split of ActivityNet-Captions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 434, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 481, + 495 + ], + "type": "text", + "content": "Training Setup. For each of the five tasks, we randomly select 100K samples with replacement (total 500K). We then use batch size 128 and learning rate 2e-5 to train for 4k iterations. The training takes around 13 hours for 13B and 9 hours for 7B models using 8 A100 GPUs. The linear projection is initialized with the LLaVA pre-trained weights." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 511, + 386, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 386, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 386, + 523 + ], + "type": "text", + "content": "5.2 Reasoning Temporal Localization Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 531, + 481, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 481, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 481, + 638 + ], + "type": "text", + "content": "We first evaluate on the newly proposed ActivityNet-RTL for reasoning temporal localization. Please refer to Section 4 for dataset details and metrics. We use \"P@0.5\" for Precision@0.5 and \"Score\" for the GPT evaluation score for explanations. Other than variations of our model, we include Video-LLaMA-v2 [39] and Video-ChatGPT [27] for comparison. We observe that most of their outputs on ActivityNet-RTL omit any timestamps and thus mIOU and Precision@0.5 become absolute. Therefore, for these methods we only evaluate the GPT-Score. In addition, we ablate the following variations for LITA, all of which are trained with the same five training tasks as LITA:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "- \"Slow Tokens Only\" samples 4 frames from the video, and computes 64 tokens per frame. It does not use the fast tokens, and it also does not use time tokens" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "text", + "content": "D.-A. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 248, + 198 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 248, + 198 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 248, + 198 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 248, + 198 + ], + "type": "image", + "image_path": "671dafffd172072992a73a62858e7d26c5e3e56edbebacb250a44d6a6ef89a0d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 208, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 130, + 208, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 208, + 482, + 285 + ], + "type": "text", + "content": "Fig. 4: Qualitative results on ActivityNet-RTL. Overall, LITA not only more accurately localize events in the video, but also provides sensible explanations with more details. In the first example, LITA correctly identifies the second arm wrestling. In the second example, LITA provides further details that they are roasting marshmallows. In the third example, LITA impressively recognizes that the girl \"falls off the beam but gets back\" and explains that this shows resilience. These are in contrast to the more generic answers by Video-LLaMA-v2." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 114, + 361, + 198 + ], + "blocks": [ + { + "bbox": [ + 250, + 114, + 361, + 198 + ], + "lines": [ + { + "bbox": [ + 250, + 114, + 361, + 198 + ], + "spans": [ + { + "bbox": [ + 250, + 114, + 361, + 198 + ], + "type": "image", + "image_path": "78817c65ea273a07790a5faa48d6a4ae01e26f5478c71e1d54a228012f2ce620.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 365, + 114, + 480, + 198 + ], + "blocks": [ + { + "bbox": [ + 365, + 114, + 480, + 198 + ], + "lines": [ + { + "bbox": [ + 365, + 114, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 365, + 114, + 480, + 198 + ], + "type": "image", + "image_path": "9cdde68de66a7f83d7c090c040babb37c083211a4c5360c40b6eb435760b811e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 211, + 345, + 404, + 440 + ], + "blocks": [ + { + "bbox": [ + 130, + 301, + 482, + 335 + ], + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 335 + ], + "type": "text", + "content": "Table 1: Results on ActivityNet-RTL. LITA substantially outperforms all baselines for all metrics. This shows the importance of our design choices. Interestingly the temporal localization accuracy also improves as we scale the model from 7B to 13B." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 211, + 345, + 404, + 440 + ], + "lines": [ + { + "bbox": [ + 211, + 345, + 404, + 440 + ], + "spans": [ + { + "bbox": [ + 211, + 345, + 404, + 440 + ], + "type": "table", + "html": "
ModelSizemIOUP@0.5Score
LITA-7B7B24.121.244.0
Video-LLaMA-v2 [39]13B--32.1
Video-ChatGPT [27]13B--38.8
Slow Tokens Only13B14.611.832.2
SlowFast Tokens13B17.514.534.1
LITA-13B13B28.625.946.3
", + "image_path": "b82d890e9e958b3d4e8f6e16d9f59f76e3fd519c9189aeab918ab434815a0046.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 468, + 481, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 481, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 481, + 492 + ], + "type": "text", + "content": "for relative timestamps. This can be seen as naively implementing a Video LLM via the LLaVA architecture." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 481, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 535 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 535 + ], + "type": "text", + "content": "- \"SlowFast Tokens\" additionally includes fast tokens (i.e. 1 token per frame) compared to \"Slow Tokens Only\". This improves the architectural design for representing video, and should allow better temporal processing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 540, + 481, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 481, + 575 + ], + "type": "text", + "content": "- \"LITA\" is our full model that further includes time tokens to better represent timestamps compared to \"SlowFast Tokens.\" We consider two model sizes, 7B and 13B, for LITA to understand the effect of different model sizes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Importance of Our Model Components. Results on ActivityNet-RTL are shown in Table 1. All metrics are averaged over three trials. LITA substantially outperforms all baselines for all metrics, and almost double mIOU and P@0.5 when compared to \"Slow Tokens Only\", which is considered as a naive extension of Image LLMs to Video LLMs. \"Score\" is assisted by GPT-4 to evaluate the quality of the explanation provided by the models. While we prompt GPT-4 to ignore the timestamps mentioned in the explanations, we observe that the scores" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "LITA: Language Instructed Temporal-Localization Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 165, + 168, + 449, + 259 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 2: Video-based Text Generation Benchmarking results. LITA significantly outperforms existing Video LLMs including Video-LLaMA-v2 [39] and VideoChatGPT [27] on all evaluated aspects. This shows that LITA not only enables accurate temporal localization, but also generally improves video understanding for Video LLMs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 165, + 168, + 449, + 259 + ], + "lines": [ + { + "bbox": [ + 165, + 168, + 449, + 259 + ], + "spans": [ + { + "bbox": [ + 165, + 168, + 449, + 259 + ], + "type": "table", + "html": "
ModelCorrectnessDetailContextTemporalConsistencyAverage
LLaMA-Adapter2.032.322.301.982.152.16
Video-LLaMA1.962.182.161.821.791.98
Video-LLaMA-v22.362.422.741.832.122.29
VideoChat2.232.502.531.942.242.29
Video-ChatGPT2.402.522.621.982.372.38
LITA2.942.983.432.683.193.04
", + "image_path": "ad2fc7048d7557098aa757fbd615e4741d000bc623e3543b7b4108d63e895f2d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 279, + 479, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 479, + 304 + ], + "type": "text", + "content": "are still slightly affected by the timestamps. Nevertheless, LITA provides better explanations for its reasoning due to an overall better video understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 306, + 480, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 480, + 426 + ], + "type": "text", + "content": "LITA Gives Detailed and Accurate Explanations. Qualitative results are shown in Figure 4. In the first example, LITA correctly localizes when the second arm wrestling happens. On the other hand, Video-LLaMA-v2 does not recognize that there is arm wrestling in the video. In the second example, while both models identify that the correct activity is cooking with fire, LITA provides much more accurate details, including the fact that they are roasting marshmallows, and the start and end time for that. Finally, in the third example, LITA impressively recognizes the event where the girl \"falls off the beam but gets back\" and correctly responds that this shows resilience in her performance. In contrast, Video-LLaMA-v2 gives a generic answer given the \"resilience\" prompt." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 428, + 481, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 428, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 428, + 481, + 489 + ], + "type": "text", + "content": "Temporal Localization Scales with LITA Size. One interesting observation is that the temporal localization quality of LITA also improves as we scale up the model from 7B to 13B (Table 1). One might think that scaling the model only improves language understanding, but our result shows that this could also improve temporal reasoning and understanding for Video LLMs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 504, + 340, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 504, + 340, + 516 + ], + "spans": [ + { + "bbox": [ + 131, + 504, + 340, + 516 + ], + "type": "text", + "content": "5.3 Video-Based Generation Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "In addition to our proposed reasoning temporal localization, we further evaluate LITA on standard evaluation for Video LLMs to better compare with existing Video LLMs. We use the \"Video-based Text Generation Performance Benchmarking\" proposed by Maaz et al. [27]. This benchmark selects videos from ActivityNet [4] and annotates question-answer pairs with natural language. This is in contrast to existing video question answering benchmarks, where the answers are often limited to a single word or phrase. In this benchmark, there are specific questions like: \"What kind of tricks is the man performing while skating?\" or generic questions like \"Can you describe the video in detail?\" For evaluation, the benchmark uses GPT to measure the following aspects of Video LLMs' answers: Correctness of Information, Detail Orientation, Contextual Understanding, Temporal Understanding, and Consistency." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 102 + ], + "type": "text", + "content": "D.-A. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 198, + 168, + 416, + 228 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 3: Analysis of LITA's training tasks on ActivityNet-RTL. \"RTL\" is needed to predict both timestamps and explanations. \"Video\" includes standard video tasks to improve video understanding. \"NLVQA\" further improves reasoning and natural language generation capabilities of LITA." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 198, + 168, + 416, + 228 + ], + "lines": [ + { + "bbox": [ + 198, + 168, + 416, + 228 + ], + "spans": [ + { + "bbox": [ + 198, + 168, + 416, + 228 + ], + "type": "table", + "html": "
ModelRTL VideoNLVQAmIOUP@0.5Score
RTL Only✓✘✘26.620.9
No NLVQA✓✓✘26.923.5
LITA✓✓✓28.625.9
", + "image_path": "e28e1f90eaf249d3b5c07ffe849740ad7068884de8fd2342ae5f13f2e64b8f58.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 249, + 485, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 485, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 485, + 393 + ], + "type": "text", + "content": "The results are shown in Table 2. We compare LITA with LLaMA-Adapter [40], Video-LLaMA [39], VideoChat [20], and Video-ChatGPT [27]. Video-ChatGPT slightly outperforms other baselines including Video-LLaMA-v2. LITA significantly outperforms these two existing Video LLMs from all aspects. In particular, LITA achieves a " + }, + { + "bbox": [ + 130, + 249, + 485, + 393 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 130, + 249, + 485, + 393 + ], + "type": "text", + "content": " improvement for Correctness of Information (2.94 vs. 2.40) and a " + }, + { + "bbox": [ + 130, + 249, + 485, + 393 + ], + "type": "inline_equation", + "content": "36\\%" + }, + { + "bbox": [ + 130, + 249, + 485, + 393 + ], + "type": "text", + "content": " relative improvement for Temporal Understanding (2.68 vs. 1.98). This shows that our emphasis of temporal understanding in training not only enables accurate temporal localization, but also improves the video understanding of LITA. We hypothesize that temporal localization enables the model to learn more details about videos, leading to improved video understanding. A similar observation was made for Image LLMs [5], where joint training with grounding tasks also improved non-grounding text generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 413, + 362, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 413, + 362, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 413, + 362, + 426 + ], + "type": "text", + "content": "5.4 Evaluating the Effects of Training Tasks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 434, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 590 + ], + "type": "text", + "content": "We have analyze the effect of our model components in Section 5.2, where we use all of our training tasks. Now we further analyze the effect of these training tasks for our model. We split the five tasks into three groups: RTL, Video, and NLVQA. \"RTL\" only includes the proposed reasoning temporal localization. Without training on our ActivityNet-RTL, the model would not output timestamps for us to evaluate temporal localization in many cases. The second group \"Video\" includes all the standard video tasks: dense video captioning, event localization, and video question answering. Using these video tasks, the model should learn better video understanding. Finally, \"NLVQA\" refers to the natural language visual question answering task to improve LITA's natural language conversation. We refer to training with just RTL as \"RTL Only,\" and training with both RTL and Video but without NLVQA as \"No NLVQA.\" Training with all three and thus all tasks is our proposed LITA." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "Results. The results on ActivityNet-RTL are shown in Table 3 and qualitative comparisons are shown in Figure 5. By only training on RTL, \"RTL Only\" does not have enough supervision to learn the task. This is reflected in both timestamp accuracy (P@0.5) and explanation quality (Score). In addition, for non-temporal questions (Q1 and Q2) in Figure 5, the model cannot properly answer and always use the answer format for reasoning temporal localization. By" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "LITA: Language Instructed Temporal-Localization Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 121, + 478, + 223 + ], + "blocks": [ + { + "bbox": [ + 134, + 121, + 478, + 223 + ], + "lines": [ + { + "bbox": [ + 134, + 121, + 478, + 223 + ], + "spans": [ + { + "bbox": [ + 134, + 121, + 478, + 223 + ], + "type": "image", + "image_path": "ec5d9a1307ac584ff42a3d6d64e111fddadff4af2fbd406011f09951818c4a94.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 233, + 482, + 322 + ], + "lines": [ + { + "bbox": [ + 130, + 233, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 482, + 322 + ], + "type": "text", + "content": "Fig. 5: Qualitative results to analyze the effect of training tasks. " + }, + { + "bbox": [ + 130, + 233, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 233, + 482, + 322 + ], + "type": "text", + "content": " is the continuous time in seconds. The questions are not from ActivityNet-RTL as we are also interested in questions other than reasoning temporal localization. \"RTL Only\" answers in the format for reasoning temporal localization, as this is the only training data it has. \"No NLVQA\" further includes standard video tasks and can correctly answer non-temporal questions (Q1 and Q2). However, the answers are short due to the short answers in existing video question answering datasets. LITA improves both reasoning and natural language conversation by utilizing NLVQA datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 347, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 431 + ], + "type": "text", + "content": "adding standard video tasks in Video to training, \"No NLVQA\" improves all metrics compared to \"RTL Only\". Qualitatively in Figure 5, it can also answer Q1 and Q2 correctly. However, this capability mainly comes from the inclusion of video question answering datasets, which leads to short answers. LITA further improves by including NLVQA, which contains complex reasoning questions to improve its reasoning capabilities. More importantly, as shown in Figure 5, LITA now answers questions with natural language instead of short phrases." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 455, + 220, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 455, + 220, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 455, + 220, + 468 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "We propose Language Instructed Temporal-Localization Assistant (LITA), which enables accurate temporal localization using Video LLMs. LITA demonstrates promising capabilities to answer complex temporal localization questions. At the same time, LITA substantially improves video-based text generation compared to existing Video LLMs even for non-temporal questions. This is the result of both our model design and data strategy. For model design, we propose time tokens to better represent the time and SlowFast tokens to efficiently process video inputs. Our experiments show the importance of these model components. For data strategy, we emphasize temporal localization data in training LITA. To achieve this, we propose the Reasoning Temporal Localization task and curate the ActivityNet-RTL dataset. Our results show that the inclusion of temporal localization data not only enables temporal localization for Video LLMs, but also improves the general video understanding capabilities. We further analyze our training data and show the benefits of incorporating standard video tasks and image instruction tuning data to train Video LLMs." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "D.-A. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "type": "text", + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: NeurIPS (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 480, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 480, + 219 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 480, + 219 + ], + "type": "text", + "content": "2. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., Jitsev, J., Kornblith, S., Koh, P.W., Ilharco, G., Wortsman, M., Schmidt, L.: OpenFlamingo: An open-source framework for training large autoregressive vision-language models. arXiv:2308.01390 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 220, + 480, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 480, + 252 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 480, + 252 + ], + "type": "text", + "content": "3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 254, + 480, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 480, + 275 + ], + "type": "text", + "content": "4. Caba Heilbron, F., Escorcia, V., Ghanem, B., Carlos Niebles, J.: Activitynet: A large-scale video benchmark for human activity understanding. In: CVPR (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 277, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 480, + 308 + ], + "type": "text", + "content": "5. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: MiniGPT-v2: large language model as a unified interface for vision-language multi-task learning. arXiv:2310.09478 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 310, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 480, + 352 + ], + "type": "text", + "content": "6. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 138, + 310, + 480, + 352 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 138, + 310, + 480, + 352 + ], + "type": "text", + "content": " * chatgpt quality (2023), https://lmsys.org/blog/2023-03-30-vicuna/" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 354, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 386 + ], + "type": "text", + "content": "7. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: PaLM: Scaling language modeling with pathways. arXiv:2204.02311 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 388, + 480, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 419 + ], + "type": "text", + "content": "8. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructblip: Towards general-purpose vision-language models with instruction tuning (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 421, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 421, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 480, + 463 + ], + "type": "text", + "content": "9. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: The EPIC-KITCHENS Dataset: Collection, challenges and baselines. IEEE Trans. PAMI 43(11), 4125-4141 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 135, + 465, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 465, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 135, + 465, + 480, + 497 + ], + "type": "text", + "content": "10. Damen, D., Doughty, H., Farinella, G.M., Furnari, A., Ma, J., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. IJCV 130, 33-55 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 135, + 498, + 480, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 498, + 480, + 519 + ], + "spans": [ + { + "bbox": [ + 135, + 498, + 480, + 519 + ], + "type": "text", + "content": "11. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: NAACL (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 135, + 521, + 480, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 521, + 480, + 541 + ], + "spans": [ + { + "bbox": [ + 135, + 521, + 480, + 541 + ], + "type": "text", + "content": "12. Feichtenhofer, C., Fan, H., Malik, J., He, K.: SlowFast networks for video recognition. In: ICCV (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 135, + 544, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 544, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 135, + 544, + 480, + 564 + ], + "type": "text", + "content": "13. Huang, B., Wang, X., Chen, H., Song, Z., Zhu, W.: Vtimellm: Empower llm to grasp video moments. arXiv preprint arXiv:2311.18445 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 135, + 566, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 566, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 135, + 566, + 480, + 586 + ], + "type": "text", + "content": "14. Huang, G., Pang, B., Zhu, Z., Rivera, C., Soricut, R.: Multimodal pretraining for dense video captioning. In: AACL-IJCNLP (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 135, + 588, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 588, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 135, + 588, + 480, + 620 + ], + "type": "text", + "content": "15. Jiang, Y.G., Liu, J., Zamir, A.R., Toderici, G., Laptev, I., Shah, M., Sukthankar, R.: Thumos challenge: Action recognition with a large number of classes (2014), https://www.crcv.ucf.edu/THUMOS14/home.html" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 135, + 622, + 480, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 622, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 135, + 622, + 480, + 642 + ], + "type": "text", + "content": "16. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Niebles, J.C.: Dense-captioning events in videos. In: ICCV (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 135, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 643, + 480, + 665 + ], + "type": "text", + "content": "17. Kuehne, H., Arslan, A.B., Serre, T.: The language of actions: Recovering the syntax and semantics of goal-directed human activities. In: CVPR (2014)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "LITA: Language Instructed Temporal-Localization Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "18. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: LISA: Reasoning segmentation via large language model. arXiv:2308.00692 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 160 + ], + "type": "text", + "content": "19. Li, B., Zhang, Y., Chen, L., Wang, J., Pu, F., Yang, J., Li, C., Liu, Z.: MIMIC-IT: Multi-modal in-context instruction tuning. arXiv:2306.05425 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "text", + "content": "20. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: VideoChat: Chat-centric video understanding. arXiv:2305.06355 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "21. Li, Y., Wang, C., Jia, J.: Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "22. Lin, B., Zhu, B., Ye, Y., Ning, M., Jin, P., Yuan, L.: Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 238, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 258 + ], + "type": "text", + "content": "23. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv:2310.03744 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 270 + ], + "type": "text", + "content": "24. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "text", + "content": "25. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "26. Miech, A., Zhukov, D., Alayrac, J.B., Tapaswi, M., Laptev, I., Sivic, J.: HowTo100M: Learning a text-video embedding by watching hundred million narrated video clips. In: ICCV (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "27. Muhammad Maaz, Hanoona Rasheed, S.K., Khan, F.: Video-ChatGPT: Towards detailed video understanding via large vision and language models. ArXiv 2306.05424 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 358, + 386, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 386, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 386, + 369 + ], + "type": "text", + "content": "28. OpenAI: GPT-4 technical report. arXiv:2303.08774 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "29. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv:2306.14824 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "30. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "31. Sigurdsson, G.A., Varol, G., Wang, X., Farhadi, A., Laptev, I., Gupta, A.: Hollywood in homes: Crowdsourcing data collection for activity understanding. In: ECCV (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "32. Tang, Y., Ding, D., Rao, Y., Zheng, Y., Zhang, D., Zhao, L., Lu, J., Zhou, J.: COIN: A large-scale dataset for comprehensive instructional video analysis. In: CVPR (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "33. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: LLaMA: Open and efficient foundation language models. arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "34. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: LLaMA 2: Open foundation and fine-tuned chat models. arXiv:2307.09288 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "35. Wu, S., Fei, H., Qu, L., Ji, W., Chua, T.S.: NExT-GPT: Any-to-any multimodal LLM. arXiv:2309.05519 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "text", + "content": "36. Xiao, J., Shang, X., Yao, A., Chua, T.S.: NExT-QA: Next phase of question-answering to explaining temporal actions. In: CVPR (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "37. Yan, S., Xiong, X., Nagrani, A., Arnab, A., Wang, Z., Ge, W., Ross, D., Schmid, C.: UnLoc: A unified framework for video localization tasks. In: ICCV (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "38. Yang, A., Nagrani, A., Seo, P.H., Miech, A., Pont-Tuset, J., Laptev, I., Sivic, J., Schmid, C.: Vid2Seq: Large-scale pretraining of a visual language model for dense video captioning. In: CVPR (2023)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 242, + 101 + ], + "type": "text", + "content": "D.-A. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 270 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "39. Zhang, H., Li, X., Bing, L.: Video-LLaMA: An instruction-tuned audio-visual language model for video understanding. arXiv:2306.02858 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "text", + "content": "40. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: LLaMA-Adapter: Efficient fine-tuning of language models with zero-init attention. arXiv:2303.16199 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 171, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 171, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 482, + 215 + ], + "type": "text", + "content": "41. Zhang, Y., Li, Y., Cui, L., Cai, D., Liu, L., Fu, T., Huang, X., Zhao, E., Zhang, Y., Chen, Y., Wang, L., Luu, A.T., Bi, W., Shi, F., Shi, S.: Siren's song in the ai ocean: A survey on hallucination in large language models. arXiv:2309.01219 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 215, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 237 + ], + "type": "text", + "content": "42. Zhou, L., Xu, C., Corso, J.J.: Towards automatic learning of procedures from web instructional videos. In: AAAI (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "text", + "content": "43. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: MiniGPT-4: Enhancing vision-language understanding with advanced large language models. arXiv:2304.10592 (2023)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "LITA: Language Instructed Temporal-Localization Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LLM as Copilot for Coarse-grained Vision-and-Language Navigation/5a4eb16e-327f-48aa-9238-afeb5ed269ac_content_list.json b/2024/LLM as Copilot for Coarse-grained Vision-and-Language Navigation/5a4eb16e-327f-48aa-9238-afeb5ed269ac_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9070b518df5b615053da7d44877597df683dfccd --- /dev/null +++ b/2024/LLM as Copilot for Coarse-grained Vision-and-Language Navigation/5a4eb16e-327f-48aa-9238-afeb5ed269ac_content_list.json @@ -0,0 +1,1966 @@ +[ + { + "type": "text", + "text": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation", + "text_level": 1, + "bbox": [ + 321, + 141, + 679, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yanyuan Qiao $^{1}$ , Qianyi Liu $^{2,3}$ , Jiajun Liu $^{4,5}$ , Jing Liu $^{2,3}$ , and Qi Wu $^{1*}$", + "bbox": [ + 240, + 210, + 759, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Australian Institute for Machine Learning, The University of Adelaide", + "bbox": [ + 285, + 238, + 715, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Institute of Automation, Chinese Academy of Sciences", + "bbox": [ + 330, + 253, + 671, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 School of Artificial Intelligence, University of Chinese Academy of Sciences", + "bbox": [ + 266, + 267, + 733, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 CSIRO Data61", + "bbox": [ + 449, + 281, + 552, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5 The University of Queensland", + "bbox": [ + 403, + 294, + 598, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yanyuan.qiao,qi.wu01}@adelaide.edu.au, liuqianyi2022@ia.ac.cn,", + "bbox": [ + 225, + 310, + 776, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jiajun.liu@csiro.au, jliu@nlpr.ia.ac.cn", + "bbox": [ + 325, + 324, + 668, + 335 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Vision-and-Language Navigation (VLN) involves guiding an agent through indoor environments using human-provided textual instructions. Coarse-grained VLN, with short and high-level instructions, has gained popularity as it closely mirrors real-world scenarios. However, a significant challenge is these instructions are often too concise for agents to comprehend and act upon. Previous studies have explored allowing agents to seek assistance during navigation, but typically offer rigid support from pre-existing datasets or simulators. The advent of Large Language Models (LLMs) presents a novel avenue for aiding VLN agents. This paper introduces VLN-Copilot, a framework enabling agents to actively seek assistance when encountering confusion, with the LLM serving as a copilot to facilitate navigation. Our approach includes the introduction of a confusion score, quantifying the level of uncertainty in an agent's action decisions, while the LLM offers real-time detailed guidance for navigation. Experimental results on two coarse-grained VLN datasets show the efficacy of our method.", + "bbox": [ + 259, + 372, + 740, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Vision-and-Language $\\cdot$ Navigation $\\cdot$ Large Language Models", + "bbox": [ + 261, + 579, + 697, + 594 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 619, + 356, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Vision-and-Language Navigation (VLN) tasks have attracted great attention from researchers, which enables an agent to navigate to a specified destination indoors based on the provided textual instructions and visual observations. The tasks of VLN show great potential in real-world applications such as home assistant robots. Most VLN tasks such as R2R [3] and RxR [17] etc. provide fine-grained step-by-step instructions, such as \"Go out the door in front of you across the room. Once in the hallway turn Left. Walk forward and then turn left into the sitting area. Stop in right in front of the fireplace.\" However, though these detailed instructions could improve the agents' ability of instructions following, they impede the further application of VLN in the real world since human beings would not like to give such long instructions to a robot in daily life. Contrastive to these tasks, the coarse-grained instruction task represented by the Remote", + "bbox": [ + 212, + 650, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 220, + 825, + 364, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e55f9e2f288adaacf22334edf36d690e10d38829f245995fcb2b513033a9042e.jpg", + "image_caption": [ + "Fig. 1: Our proposed VLN-Copilot. When an agent is confused about making an action decision, it will proactively seek help from the LLM, and then the LLM will assist the agent by analyzing the situations faced by the agent and providing appropriate guidance." + ], + "image_footnote": [], + "bbox": [ + 316, + 143, + 687, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Embodied Referring Expression (REVERIE) task [29] is more expected to enhance the practical utility of VLN because it involves instructions that are closer to those encountered in real-world scenarios, such as \"Go clean the mirror in the bathroom.\" Giving such coarse-grained instructions to the agent would bring more challenges to the agent's navigation than fine-grained instructions.", + "bbox": [ + 212, + 454, + 787, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VLN Agents navigating in environments can benefit from seeking external assistance [34]. Early works [27, 37] tried to let agents seek help from oracles or simulators, but the assistance content obtained was fixed and limited. Since this assistance information is based on pre-built ground truth information in datasets or simulators, which are not consistent with real-life scenarios. Inspired by the development of Large Language Models (LLMs) with rich internalized world knowledge, many works have been made to exploit their capabilities to aid navigation tasks. One recent method named MiC [31] makes the first step to use LLM in the coarse-grained VLN task of REVERIE, which queries the LLM to generate more detailed instructions by prompting the LLM with extra planning examples. The promising performance of MiC shows the great potential to use LLM in coarse-grained VLN tasks. However, MiC has some weaknesses in applying LLM for VLN. First, the way the agent seeks help from the LLM is a passive way that depends on the change of the room where the agent is located. This approach may not help the agent when it faces difficulties in making predictions while still in the same room. Second, MiC generates fine-grained planning no matter what challenges the agent faces. This practice lacks a thorough analysis of the difficulties the agent may have and fails to give the agent diverse guidance for different scenarios.", + "bbox": [ + 212, + 534, + 787, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we propose a novel method named VLN-Copilot that makes the LLM a competent copilot to assist the agent in navigation. As shown in Figure 1, before making a movement, the agent calculates a confusion score to assess the level of", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Qiao et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "uncertainty associated with its current action choice. If this score surpasses a predefined threshold, it triggers the agent to seek guidance. This practice is more proactive than relying on the change of scenes which may be wrong. When the agent decides to seek assistance, it consults the LLM and then the LLM conducts a comprehensive evaluation of the situation, identifies the type of guidance required, and generates the necessary guidance to aid the agent in navigating effectively. The types of guidance information are selected from the in-depth pre-analysis of prevailing challenges the agent may face in the sampled environment using LLM.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We conduct experiments on the REVERIE and CVDN-target dataset and our model surpasses the previous methods. Specifically, our VLN-Copilot obtains $42.32\\%$ on the main navigation metric SPL and $26.55\\%$ on the main object grounding metric RGSPL on the REVERIE dataset, and 4.47 in GP on the CVDN-target dataset, respectively.", + "bbox": [ + 212, + 267, + 787, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 238, + 328, + 540, + 343 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel method named VLN-Copilot which utilizes LLM to analyze the difficulties the agent may meet and give diverse guidance for different situations, rather than to generate planning by prompting LLM with extra planning examples.", + "- We introduce the confusion score to help the agent proactively ask the LLM for help when confused to make predictions rather than passively trigger conversations.", + "- We utilize LLM with elaborately designed prompts to conduct a detailed challenge analysis of the agent's different situations, enabling the LLM to provide better guidance on the fly.", + "- Extensive experiments on two VLN benchmarks show the effectiveness of our proposed method." + ], + "bbox": [ + 225, + 352, + 782, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 526, + 367, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Vision-and-Language Navigation", + "text_level": 1, + "bbox": [ + 214, + 558, + 486, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In recent years, Vision-and-Language Navigation (VLN) has attracted widespread attention due to its promising applications in home assistant robots and related fields [9, 12, 18, 20, 28, 40]. Many different tasks have emerged in the field of VLN [4, 14, 16, 25, 27, 29, 37]. In terms of the navigation space accessible to the agent, these cover indoor [3, 16, 17], outdoor [4, 35], and aerial scenarios [8, 25].", + "bbox": [ + 212, + 583, + 787, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early VLN task R2R [3] started from indoor environments requiring the agent to reach the destination according to the given instructions, and there were some subsequent tasks (e.g., RxR [17] and VLN-CE [16]). These tasks all provide detailed natural language instructions, such as \"Walk down the stairs at the bottom of the stairs take a left. Walk straight to get to the kitchen inside and wait just inside the door.\" These detailed instructions guide the agent through complex step-by-step processes and are more focused on testing the agent's ability to understand and follow instructions. Some other tasks adopt an object-oriented perspective, such as REVERIE [29] and SOON [44]. These tasks require not only reaching a designated destination but also discovering specific target objects. Different from SOON providing detailed instructions and object recognition prompts, such as \"The cabinet is square, closed and wooden. The cabinet is placed below the sink and the water-tap. There are some stools and table nearby.\"", + "bbox": [ + 212, + 659, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation", + "bbox": [ + 318, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "REVERIE's instructions appear more concise and natural, such as \"Bring me the swivel chair from the office.\" These coarse-grained instructions, akin to instructions encountered in real-world scenarios, enhance the potential for innovative research within the domain of coarse-grained instruction VLN. However, the task remains challenging due to the limited information these instructions provide.", + "bbox": [ + 217, + 146, + 785, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 VLN Agents Asking for Help", + "text_level": 1, + "bbox": [ + 217, + 244, + 457, + 261 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Agents can gain advantages by seeking external assistance to assist navigation, and there have been studies promoting navigation agents to seek help during navigation tasks [27, 34]. Thomason et al. [37] introduce the Navigation from Dialog History (NDH) task, which contains dialog history between agents and oracles during navigation. The agent follows guide instructions and asks questions to the oracle when needed. The task provides a valuable resource for investigating the navigation interactions between agent and oracle. Similarly, HANNA [27] builds a simulator where navigation agents can request language-and-vision assistance when they face difficulties during navigation. However, the agent's signals asking for help are fixed and predefined in the simulator. Additionally, the assistance provided is based on ground truth information derived from the R2R dataset, which may not fully represent real-world navigation scenarios due to its reliance on pre-built datasets. Similarly, Zhu et al. proposed SCoA [45], which enables navigation agents to ask questions from a set of pre-constructed candidate question-answer pairs to aid in navigation. Recently, LLM-Planner [36] and MiC [31] have attempted to leverage the knowledge of large language models (LLM) to assist agents in navigation. Agents can query the LLM to generate detailed step-by-step instructions to aid navigation. However, the ways the agents ask for help from both methods are passive, either based on fixed time steps or when the room the agent is in changes. This practice may miss the opportunity to help the agent when it has difficulties in prediction actions. Besides, these two methods concentrate on generating planning by prompting LLM with extra planning data. Different from these methods, we propose a method named VLN-Copilot that allows agents to proactively seek help when they are not sure where to go in this work. Moreover, our VLN-Copilot pays more attention to analyzing the difficulties the agent may meet and can dynamically give appropriate types of guidance in diverse situations for the agent.", + "bbox": [ + 217, + 272, + 785, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 217, + 674, + 313, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 217, + 707, + 349, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The main goal for the agent in the VLN task is to navigate to a specific destination following the high-level concise instructions. Specifically, the agent is asked to navigate in the simulated environment, which can be described as an undirected graph $\\mathcal{G} = \\{\\mathcal{V},\\mathcal{E}\\}$ , where $\\mathcal{V}$ represents the navigable nodes and $\\mathcal{E}$ represents the connectivity edges. Given the natural language instruction $L$ , the agent starts from the initial node $V_{0}$ and perceives a panoramic view $\\mathcal{R}_t$ of the current node $V_{t}$ to make the action $a_{t}$ at each time step $t$ . The panoramic view $\\mathcal{R}_t$ has $K$ single-view images $r_i$ and can be represented as", + "bbox": [ + 217, + 734, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Qiao et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/48d79f507645ff044e946dc24f4b66da236e91bc512b62ddd0f1eca1bbfe105d.jpg", + "image_caption": [ + "Fig.2: Overview of the VLN-Copilot. At each time step, the agent predicts the logits over the candidate's view of actions. After that, the agent will calculate the confusion score to decide whether to ask LLM for help (Sec. 3.3). When the confusion score surpasses threshold $\\tau$ , the agent queries the LLM for help with the prompt that helps LLM perceive the environment (Sec. 3.4). Then the LLM generates useful guidance information, which will be returned to the agent to assist in decision-making (Sec. 3.6)." + ], + "image_footnote": [], + "bbox": [ + 253, + 152, + 751, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{R}_t = \\{r_i\\}_{i=1}^K$ , of which $r_i$ is represented by an image feature vector with the corresponding orientation feature vector. Besides, for the sub-task of object grounding, the object features $\\mathcal{O}_t = \\{o_i\\}_{i=1}^M$ are extracted from the panoramic view $\\mathcal{R}_t$ with the annotated object bounding boxes or by object detectors. With these observations of the environment, the agent takes a series of actions $\\langle a_0, \\dots, a_N \\rangle$ to navigate to the target location, where each action $a_t$ selects one navigable node from the candidates. When the target object is grounded or the pre-defined maximum navigation steps have been achieved, the navigation ends.", + "bbox": [ + 212, + 420, + 787, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Overview", + "text_level": 1, + "bbox": [ + 215, + 569, + 323, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As is shown in Figure 2 and Algorithm 1, given the coarse-grained instruction, the agent first perceives the environment to obtain the current visual observation and starts to predict a probability over candidate actions. Then, a confusion score is computed according to the probability distribution, which determines whether the agent is confused about making predictions in the current situation and whether the agent should ask the LLM for help. If the confusion score exceeds a predefined threshold $\\tau$ (obtained via grid search) which means the agent is confused about making decisions, the agent initiates a conversation with the LLM to seek their help with an elaborately designed prompt. Specifically, the prompt contains coarse-grained instruction, fine-grained scene descriptions of candidate views, historical descriptions of trajectory scenes, and the requirements for the LLM. Then, the LLM analyzes the agent's situation and gives advice about which type of guidance information is useful as well as the details about that guidance information. The types of guidance information are refined from the results of challenge pre-analysis by LLM. Finally, the response about the guidance from the LLM is appended after the coarse-grained instruction to guide the agent's action prediction collaboratively. This paradigm aims to optimize agents' decision-making processes by", + "bbox": [ + 212, + 598, + 787, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation", + "bbox": [ + 318, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "combining instruction, observation, and LLM-based world knowledge via the conversations between the agent and LLM.", + "bbox": [ + 215, + 146, + 785, + 176 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "Algorithm 1 VLN-Copilot \nNotation Summary: I: high-level instruction $\\mathcal{R}_t$ : visual observation at timestep t $S_{t}$ : confusion score at timestep t $W_{t}^{F}$ : fine-grained scene descriptions of all candidate views at timestep t $W_{t}^{\\mathrm{Tr}}$ : historical trajectory description at timestep t $T_{\\mathrm{guide}}$ : prompt template for LLM to generate guidance information LLM: large language model $\\tau$ : pre-defined threshold for confusion score \n $t\\gets 0$ $\\rightharpoondown$ Initial timestep while $t < \\max$ step and $\\hat{a}_{t}\\neq$ \"stop\" do if $S_{t} > = \\tau$ then $P_{\\mathrm{guide}}\\leftarrow T_{\\mathrm{guide}}(I,W_t^F,W_t^{\\mathrm{Tr}})$ $R_{\\mathrm{guide}}\\gets \\mathrm{LLM}(P_{\\mathrm{guide}})$ $I_{t}^{\\mathrm{Inter}}\\gets \\mathrm{CONCAT}(I,R_{\\mathrm{guide}})$ $\\hat{a}_{t}\\gets \\mathrm{Agent}(I_{t}^{\\mathrm{Inter}},\\mathcal{R}_{t})$ else $\\hat{a}_{t}\\gets \\mathrm{Agent}(I,\\mathcal{R}_{t})$ end if $t\\gets t + 1$ \nend while", + "guess_lang": "txt", + "bbox": [ + 287, + 210, + 718, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 When to Ask LLM for Help", + "text_level": 1, + "bbox": [ + 215, + 478, + 452, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Different from previous methods [31] that set fixed timestep for interactions with LLM or trigger the interactions when the room scenes change, we select a more active practice for the agent by using the confusion score to decide whether to ask for help. Specifically, at each time step $t$ , when the agent is going to predict the logits over the candidate nodes for the next action $a_{t}$ , we calculate the entropy of the predicted probability distribution over each candidate navigable node $v_{i}$ as the confusion score $S_{t}$ :", + "bbox": [ + 212, + 505, + 787, + 595 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {t} \\left(v _ {i}\\right) = \\operatorname {S o f t m a x} \\left(s _ {t} \\left(v _ {i}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 607, + 785, + 625 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {t} = - \\sum_ {i} p _ {t} \\left(v _ {i}\\right) \\log_ {2} \\left(p _ {t} \\left(v _ {i}\\right) + \\epsilon\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 627, + 785, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $s_t(v_i)$ is the predicted logits assigned by the agent to the $i$ -th navigable node $v_i$ at time step $t$ , and $\\epsilon$ is a minimal number added to avoid numerical instability.", + "bbox": [ + 212, + 669, + 784, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, if the computed confusion score $S_{t}$ surpasses a predefined threshold $\\tau$ , which means the agent is confused about making a decision among some candidate nodes, it will send a message to LLM asking for help.", + "bbox": [ + 212, + 700, + 785, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 How LLM Perceives the Environment", + "text_level": 1, + "bbox": [ + 215, + 768, + 521, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LLMs would give inaccurate advice for the agent's navigation since they are not able to directly acquire visual content from the environment. Both current visual observations from the agent and previously observed scenes are essential for the LLM to analyze the", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Qiao et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "navigation and supply important information. Thus, enhancing the LLM's environmental awareness as far as possible is a critical first step to making it a competent copilot. To solve this problem, we translate the dynamic observations and trajectories into multi-grained spatiotemporal descriptions as the eye for the LLM, using the out-of-the-box vision-and-language model of BLIP [19] model. Detailedly, the visually-grounded description mainly consists of three parts: the coarse-grained scene description of the current panoramic view $\\mathcal{R}_t$ , the fine-grained scene description of each view $r_i$ of selected candidate navigable nodes at the current time step $t$ , and the trajectory description of previously observed scenes before the current time step $t$ .", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, we obtain both the panoramic view image $\\mathcal{R}_t$ of the agent's observation and the single view image $r_i$ of each candidate navigable node at the time step $t$ . We first feed $\\mathcal{R}_t$ to the BLIP model to generate the coarse-grained scene description $W_t^C$ for the agent's current location. For the fine-grained scene description, the agent is provided with $k$ candidate views for selection. Then, we feed each single-view image $r_i$ of the $k$ selected candidate navigable nodes to BLIP to generate the fine-grained scene description $W_t^{F_i}$ :", + "bbox": [ + 212, + 282, + 787, + 388 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nW _ {t} ^ {\\mathrm {C}} = \\operatorname {B L I P} \\left(\\mathcal {R} _ {t}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 400, + 785, + 416 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nW _ {t} ^ {F _ {i}} = \\operatorname {B L I P} \\left(r _ {i}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 421, + 784, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Then we concatenate all previous coarse-grained scene descriptions $W^{\\mathrm{C}}$ of each time step before the current time $t$ as the historical trajectory description $W_{t}^{\\mathrm{Tr}}$ such as \"I passed a bedroom, and then I passed ...\" This process can be formulated as follows:", + "bbox": [ + 214, + 449, + 787, + 494 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nW _ {t} ^ {\\mathrm {T r}} = \\operatorname {C O N C A T} \\left(W _ {1} ^ {\\mathrm {C}}, \\dots , W _ {t - 1} ^ {\\mathrm {C}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 506, + 785, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that we only generate the fine-grained scene descriptions and trajectory descriptions when the agent is confused to make the action prediction and asks the LLM for help, while we generate coarse-grained scene descriptions at each time step. In addition, for better comprehension of LLM, we added the candidate view id $i$ as the prefix to each corresponding native fine-grained scene description $W_{t}^{F_{i}}$ . For example, we converted the generated scene description of \"A room with a tilted ceiling, a bathroom mirror on the wall, and a skylight above the window.\" into the candidate view description with the following format: \"I see candidate 0 is a room with a tilted ceiling ...\"", + "bbox": [ + 212, + 535, + 787, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5 Difficulty and Useful Guidance-Type Pre-Analysis", + "text_level": 1, + "bbox": [ + 214, + 678, + 606, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Before asking the agent to query an LLM to generate guidance for assistance, it would be beneficial to explore the difficulties the agent may encounter in making decisions and what guidance information the LLM should provide to be useful for the agent's navigation. However, manual analysis by humans is costly and impractical. Thus, we turn to LLM for analysis. Specifically, we first collect some navigation samples. When the agent finds it difficult to make decisions during the navigation process, we convert the situation the agent encountered into a text description and save it as a navigation sample. Then, we use an elaborate prompt to query the LLM to analyze the situation encountered by the agent in the current navigation example, such as what is the reason", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation", + "bbox": [ + 318, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "System: You are a helpful assistant for navigation. There is an agent navigating indoors according to task instructions and current observation. At each step, the agent needs to choose one of the candidate directions to go. I will tell you which direction is correct.", + "bbox": [ + 222, + 150, + 772, + 170 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "You need to first analyze the current situation, such as what difficulties the agent encountered during navigation that made it difficult to choose a direction and what types of guidance information may be useful.", + "bbox": [ + 223, + 174, + 771, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Task instruction: Turn off the white table lamp in the bedroom", + "bbox": [ + 223, + 196, + 455, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Observation:", + "bbox": [ + 225, + 205, + 277, + 213 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "I see candidate 1 is a white door leads to a bedroom with a bed, electric fan and table", + "bbox": [ + 225, + 213, + 535, + 222 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "I see candidate 2 is a bathroom has a sink and the door leads to the shower room", + "bbox": [ + 225, + 222, + 524, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "I see candidate 3 is a hallway with white walls, stairs, and a door on the second floor", + "bbox": [ + 225, + 229, + 531, + 237 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "History trajectory: I have passed a balcony with rocking chairs and a table, then I go to a hallway with a picture of a bedroom.", + "bbox": [ + 225, + 237, + 681, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Correct answer: candidate 1", + "bbox": [ + 225, + 244, + 334, + 251 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Assistant:", + "bbox": [ + 225, + 257, + 281, + 265 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Difficulty: The agent encountered difficulties in choosing a direction because there are multiple doors at the same time. This makes it challenging for", + "bbox": [ + 225, + 265, + 769, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the agent to determine which direction could find the lamp.", + "bbox": [ + 225, + 273, + 436, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Please list possible type of guidance may be useful.", + "bbox": [ + 225, + 287, + 429, + 296 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Assistant:", + "bbox": [ + 225, + 303, + 279, + 310 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Guidance Type: The agent needs to the destination information to make an informed decision.", + "bbox": [ + 225, + 310, + 568, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 3: Prompt template for LLM to pre-analyze the difficulties the agent encounters and what type of guidance information maybe useful. The color of red, blue represent multi-turn responses generated from LLM.", + "bbox": [ + 212, + 352, + 787, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for the difficulty in making a decision, and the helpful types of guidance. Finally, we collect the guidance information generated by LLM and cluster it to obtain the final guidance categories.", + "bbox": [ + 212, + 426, + 787, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Navigation Samples Collection We first let the agent navigate around the environment with confusion scores computed. If the confusion score surpasses the pre-defined threshold at the time step $t$ , which means the agent comes across difficulties in making the action prediction, then we collect the tuple $(I, W_{t}^{F}, W_{t}^{Tr}, A_{t})$ of the high-level task instruction $I$ , fine-grained candidate scene descriptions $W_{t}^{F}$ , the trajectory description $W_{t}^{Tr}$ and the ground-truth action $A_{t}$ as a sample data $D_{\\text{sample}}$ for analysis. Note that $W_{t}^{F}$ includes $k$ fine-grained scene descriptions of the selected candidate views. In total, we randomly collect 5,000 samples.", + "bbox": [ + 212, + 497, + 787, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Propmt Design To make the LLM better analyze the situation that the agent meets, we elaborately designed a prompt template for multi-turn query-and-response. We first set the unified system prompt $P_{\\text{system}}$ that requires the LLM as a helpful assistant for navigation. Then, we transform each sample data $D_{\\text{sample}}$ into sample prompt $P_{\\text{sample}}$ by respectively adding the prefix of \"Task instruction:\", \"Observation:\", \"History trajectory:\" and \"Correct answer:\" into $I, W_t^F, W_t^{Tr}$ and $A_t$ . Next, we decompose the task of analyzing the challenge into two sub-tasks, the first of analyzing what difficulties the agent may meet in making decisions and the second of analyzing what types of guidance may be useful for navigation in an iterative way. Correspondingly, we use \"########Assistant: Difficulty:\" and \"########Assistant: Guidance Type:\" as the sub-task prompts $P_{\\text{diff}}$ and $P_{\\text{guide}}$ . We concatenate the prompts of $P_{\\text{system}}$ , $P_{\\text{sample}}$ and $P_{\\text{diff}}$ as $P_{\\text{task1}}$ and feed $P_{\\text{task1}}$ to LLM, which will generate the response $R_{\\text{diff}}$ about the difficulty the agent may meet. We further feed $P_{\\text{task2}}$ to LLM, the concatenation", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Qiao et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "System: You are a helpful assistant for navigation. There is an agent navigating indoors according to task instructions and current observation. At each step, the agent needs to choose one of the candidate directions to go.", + "bbox": [ + 225, + 148, + 769, + 167 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "You need to analyze what additional guidance information the agent need to make the choice. There are four types of information: location, prioritization, destination description, previous observations. You can only give one answer from these types.", + "bbox": [ + 227, + 172, + 767, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Task instruction: Turn off the white table lamp in the bedroom", + "bbox": [ + 227, + 195, + 455, + 204 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Observation:", + "bbox": [ + 227, + 204, + 281, + 210 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "I see candidate 1 is a white door leads to a bedroom with a bed, electric fan and table", + "bbox": [ + 227, + 210, + 535, + 218 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "I see candidate 2 is a bathroom has a sink and the door leads to the shower room", + "bbox": [ + 227, + 218, + 522, + 224 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "I see candidate 3 is a hallway with white walls, stairs, and a door on the second floor", + "bbox": [ + 227, + 224, + 532, + 233 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "History trajectory: I have passed a balcony with rocking chairs and a table, then I go to a hallway with a picture of a bedroom.", + "bbox": [ + 227, + 233, + 679, + 241 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "####Assistant:", + "bbox": [ + 227, + 247, + 282, + 255 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Guidance Type: Destination description", + "bbox": [ + 227, + 255, + 375, + 263 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Please give description of the destination.", + "bbox": [ + 227, + 270, + 392, + 279 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Assistant:", + "bbox": [ + 227, + 285, + 282, + 292 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The bedroom is a cozy and comfortable space, typically with a queen-sized bed in the center of the room. The bed is dressed with crisp, white linens and fluffy pillows.", + "bbox": [ + 227, + 292, + 767, + 308 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fig. 4: Prompt template for LLM to generate guidance information.", + "bbox": [ + 299, + 340, + 700, + 354 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "of $P_{task1}$ , $R_{diff}$ and $P_{guide}$ and we will get the response $R_{guide}$ about what types of guidance may be useful for navigation. The prompt template is depicted in Figure 3.", + "bbox": [ + 212, + 388, + 787, + 421 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Guidance-Type Clustering After the above-mentioned operations, we obtain a collection of the guidance response $R_{guide}$ . We first use Sentence-BERT [33] to embed each $R_{guide}$ and get the corresponding embedding $E_{guide}$ . Then we employ the silhouette coefficient [15] to determine the optimal number of clusters for K-Means clustering. The silhouette coefficient is an indicator to evaluate the quality of clustering. It combines the two factors of clustering cohesion and separation, calculates a silhouette coefficient value for each sample, and then calculates a silhouette coefficient value through all samples. By calculating the average silhouette coefficient under different cluster numbers $k$ , the clustering quality under each cluster number can be evaluated. In general, the number of clusters with the highest average silhouette coefficient is considered the optimal number of clusters because it provides the best balance of cohesion and separation. And clustering result shows that the best number of clusters is seven. In addition, after analysis based on the practical situation, we found that some of the guidance content is similar, and some guidance is not executable, such as \"Room Layout\", which cannot be obtained from the environment. Thus, we manually merge these duplicate types, remove infeasible types, and finally reserve four applicable categories:", + "bbox": [ + 212, + 452, + 787, + 696 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Location: information about the agent's current position and the destination.", + "- Prioritization: information about the urgent task that needs to be completed based on an assessment of agent progress.", + "- Destination Description: detailed information about the target location, providing more cues to reach its destination.", + "- Previous Observations: information about the agent's navigation trajectory, lets the agent remember history information." + ], + "bbox": [ + 225, + 712, + 784, + 823 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation", + "bbox": [ + 318, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.6 LLM Guides the Agent On-the-Fly", + "text_level": 1, + "bbox": [ + 215, + 146, + 500, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "After the pre-analysis of the agent's navigation, we employ the analysis result and a new prompt to ask LLM to aid the agent on the fly when it is confused to make decisions. We first illustrate the modification of the prompt for the on-the-fly guidance. Then, we show how LLM guides the agent with different types of guidance information.", + "bbox": [ + 215, + 174, + 785, + 234 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Prompt Design We also designed the prompt template to ask the LLM to analyze the agent's current situation and decide which guidance type from the selected four categories might be useful in that situation. This prompt template is similar to the pre-analysis prompt template, but with some modifications to guide agent navigation. The full prompt template is shown in Figure 4. First, we modify the system prompt $P_{system}$ by asking the LLM to choose which guidance information is useful from the four pre-selected categories instead of analyzing the difficulties and generating unlimited guidance information. Second, we delete the ground truth action in the sample prompt $P_{sample}$ . At last, the task prompt is revised to \"########Assistant: Guidance Type:\".", + "bbox": [ + 215, + 258, + 785, + 396 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Integrating Guidance with Instruction During the navigation of the agent, when the confusion score $S_{t}$ surpasses the pre-defined threshold $\\tau$ at time step $t$ , it will ask the LLM for guidance. With the help of the above-mentioned prompts, the LLM can perceive the environment and trajectory of the agent, analyze the situation, and select the guidance type from the pre-selected four categories. For different types of guidance, the LLM will generate different guidance information $I_{t}^{G}$ to assist the navigation. Specifically, the LLM will generate detailed content about the type of \"Location\", \"Prioritization\" and \"Destination Description\" according to its internalized world knowledge. For these three types of guidance, the generation process is also interactive, which first generates the type of information and then generates the details according to the type of information. Such a two-step process ensures the quality of generated guidance. For the type of \"Previous Observations\", we re-use the historical trajectory as the guidance information. Some examples of generated guidance information can be found in Figure 7. Finally, we append the guidance information $I_{t}^{G}$ to the high-level instruction $I$ , and then the agent will navigate under the guidance of the integrated instruction $I_{t}^{Inter}$ .", + "bbox": [ + 215, + 420, + 785, + 647 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 672, + 349, + 689 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Settings", + "text_level": 1, + "bbox": [ + 215, + 705, + 313, + 722 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets We evaluate our method on two VLN tasks: REVERIE and CVDN-target, which are all based upon the MP3D indoor environments. Compared to Room-to-Room (R2R) task, these instructions are high-level and concise, such as \"Bring me the swivel chair from the office\". CVDN-target is our proposed VLN setup similar to REVERIE. The instruction contains a target and conversation history. Here we remove the conversation history and only keep the information containing the target as the instruction, such as \"To find a book in the room.\"", + "bbox": [ + 215, + 733, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Qiao et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/7c72e6d02efef929d3d7376ff8df33aa004c8aa3cbb174279b50452aedd9fe1f.jpg", + "table_caption": [ + "Table 1: Comparison with the state-of-the-art methods on REVERIE." + ], + "table_footnote": [], + "table_body": "
MethodsVal UnseenTest Unseen
NavigationGroundingNavigationGrounding
TLOSR↑SR↑SPL↑RGS↑RGSPL↑TLOSR↑SR↑SPL↑RGS↑RGSPL↑
Human------21.1886.8381.5153.6677.8451.44
RecBERT [13]16.7835.0230.6724.9018.7715.2715.8632.9129.6123.9916.5013.51
Airbert [11]18.7134.5127.8921.8818.2314.1817.9134.2030.2823.6116.8313.28
HAMT [5]14.0836.8432.9530.2018.9217.2813.6233.4130.4026.6714.8813.08
VLN-PETL [32]14.4737.0331.8127.6718.2615.9614.0036.0630.8326.7315.1313.03
LANA [39]16.2838.5434.0029.2619.0316.1816.7536.4133.5026.8917.5314.25
HOP+ [30]14.5740.0436.0731.1322.4919.3315.1735.8133.8228.2420.2016.86
DUET [7]22.1151.0746.9833.7332.1523.0321.3056.9152.5136.0631.8822.06
Lily [22]21.8753.7148.1134.4332.1523.4321.9460.5154.3237.3432.0221.94
BEVBert [1]-56.4051.7836.3734.7124.44-57.2652.8136.4132.0622.09
AutoVLN [6]-62.1455.8940.8536.5826.76-62.3055.1738.8832.2322.68
AZHP [10]22.3253.6548.3136.6334.0025.79-55.3151.5735.8532.2522.44
BSG [24]24.7158.0552.1235.5935.3624.2422.9062.8356.4538.7033.1522.34
KERM [20]21.8555.2150.4435.3834.5124.4517.3257.5852.4339.2132.3923.64
ScaleVLN [42]--56.9741.8435.7626.05--56.1339.5232.5322.78
GRidMM [41]23.2057.4851.3736.4734.5724.5619.9759.5553.1336.6034.8723.45
MiC [31]20.6462.3756.9743.6037.5228.7218.1162.4055.7441.9735.2526.17
VLN-Copilot21.8962.6257.4043.6338.8829.7521.7263.2657.8142.3236.5626.55
", + "bbox": [ + 233, + 171, + 769, + 452 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation Metrics We utilize the standard evaluation metrics to evaluate our method, we report Success Rate (SR), Success weighted by Path Length (SPL) [2], and Goal Progress (GP) as the main navigation metrics, Remote Grounding Success rate (RGS) and RGS weighted by Path Length (RGSPL) as the main object grounding metrics.", + "bbox": [ + 212, + 479, + 787, + 542 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details We implement our proposed method using PyTorch. The model is trained on one single NVIDIA RTX 4090 GPU. For REVERIE, we fine-tuned the pre-trained model from [6] with a batch size of 4 for 100k iterations. For the CVDN-target, we fine-tuned the pre-trained model from [7] with a batch size of 4 for 50k iterations. AdamW [26] optimizer is adopted and the learning rate is set to $1 \\times 10^{-5}$ . The threshold $\\tau$ is set as 1. For the baseline agent, we use the recent state-of-the-art model of AutoVLN [6]. When collecting data for pre-analysis, we use two agents based on pre-trained AutoVLN [6] and HAMT [5] to avoid agent bias. LLM is integrated into the training process, we use the open-source Vicuna-7B-1.5 model [43], which is fine-tuned from Llama 2 [38].", + "bbox": [ + 212, + 564, + 787, + 715 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 214, + 738, + 558, + 753 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "REVERIE As shown in Table 1, our method achieves competitive performance on the REVERIE benchmark compared to previous methods, such as AutoVLN [6], Mic [31] and ScaleVLN [42]. Particularly, compared with the SoTA method MiC [31], our method outperforms Mic in terms of the object grounding metric RGSPL with $1.36\\%$ and RGSPL with $1.03\\%$ on the Val Unseen split. It is worth noting that compared with", + "bbox": [ + 212, + 763, + 790, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation", + "bbox": [ + 320, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 769, + 116, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/30f9fadd9fda3c00cad1028e05be5c7d76da48ad65e38599db0b448f8a8eb543.jpg", + "table_caption": [ + "Table 2: Comparison on CVDN-target Validation Unseen." + ], + "table_footnote": [], + "table_body": "
MethodsTLGP↑
HAMT† [5]37.042.74
DUET† [7]78.533.70
AutoVLN† [6]61.973.98
VLN-Copilot40.384.47
", + "bbox": [ + 246, + 184, + 392, + 250 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 4: Ablation of different threshold $\\tau$ .", + "bbox": [ + 215, + 272, + 418, + 299 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/dddaa0e6f839e5a2f0bb53541b9dc1b15727d1ad7a0d7f17b161c8c3752f1bb4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
τ0.811.21.4
SPL↑42.7843.6343.6842.55
RGSPL↑28.3529.7528.1728.42
", + "bbox": [ + 225, + 313, + 413, + 354 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/0db18d108cf36fa8d6491d5e79f271fb56cb06f78a8db04a6b98b83a6a0a8d0a.jpg", + "table_caption": [ + "Table 3: Ablation of LLM Analyzer." + ], + "table_footnote": [], + "table_body": "
MethodNavigationGrounding
TLOSR↑SR↑SPL↑RGS↑RGSPL↑
No-Selection (Baseline)21.9657.8252.7740.1534.7626.37
All-Inclusive20.7059.7655.5041.3636.4727.33
Random-Selection20.1558.0253.7942.1635.3027.08
LLM-Selection (Ours)21.8962.6257.4043.6338.8829.75
", + "bbox": [ + 434, + 178, + 785, + 244 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/39376dcb91efb1728639d8a486844658893c41f478321b010ea6419e56f4f530.jpg", + "table_caption": [ + "Table 5: Ablation study of the guidance information." + ], + "table_footnote": [], + "table_body": "
MethodNavigationGrounding
TLOSR↑SR↑SPL↑RGS↑RGSPL↑
Baseline21.9657.8252.7740.1534.7626.37
+ Location20.9762.2056.8042.9737.7428.46
+ Prioritization20.4859.4754.7041.7435.8627.40
+ Destination Description19.2159.5953.9941.9134.9927.11
+ Previous Observations19.2257.9653.7342.8335.6427.12
", + "bbox": [ + 434, + 290, + 785, + 363 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "ScaleVLN [42], which collects 4.9M instruction-trajectory pairs to train the model, our method still achieves better performance (2.8% on SPL and 3.77% on RGSPL).", + "bbox": [ + 215, + 390, + 785, + 421 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "CVDN-target As shown in Table 2, we compare VLN-Copilot with three main-stream VLN models HAMT [5], DUET [7], and AutoVLN [6] on the CVDN-target dataset. For a fair comparison, both models are trained on the CVDN-target dataset. $\\dagger$ represents the reproduction results of these methods. It highlights the effectiveness of VLN-Copilot, especially in terms of GP metrics, where it outperforms all other methods. Although TL is slightly higher than HAMT, its significantly improved GP emphasizes its overall advantage in handling unseen validation data.", + "bbox": [ + 212, + 441, + 787, + 547 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 568, + 362, + 583 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effect of LLM Analyzer To evaluate the ability of LLM for guidance type selection, we conduct an ablation study, there are four settings, as follows: (1) No-Selection: our baseline, which means the agent does not use any guidance information; (2) All-Inclusive: the agent lets the LLM generate all types of guidance and integrates this information; (3) Random-Selection: the agent randomly selects one of four guidance types and lets the LLM generate corresponding guidance information; (4) LLM-Selection: our method, the LLM outputs guidance type and generates the information. As illustrated in Table 3, we can see that the LLM-Selection method achieved the highest navigation main metric spl with a value of $43.63\\%$ , followed by the Random-Selection method with a score of $42.16\\%$ . This suggests that blindly combining all types of guidance information is not the best option, probably due to the redundancy of excessively detailed information. Compared to the Baseline, the LLM-Selection method improves the SR by $4.63\\%$ , SPL by $3.48\\%$ , RGS by $4.12\\%$ , and $3.38\\%$ .", + "bbox": [ + 212, + 592, + 787, + 789 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Different Threshold of Confusion Score As shown in Table 4, we conduct an ablation study on threshold $\\tau$ in the confusion score. Although when $\\tau = 1$ , SPL is slightly lower", + "bbox": [ + 215, + 809, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Qiao et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5805db1901ab93fa70dd35ef66d2b0c4962f877a8db9c4fd3546dd4f67de78a3.jpg", + "image_caption": [ + "Fig.5:Confusion Score" + ], + "image_footnote": [], + "bbox": [ + 285, + 143, + 500, + 270 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/befe241a5f3078077a5b31fed9fbf0dc67967eb321a00f9072657552949ca5ee.jpg", + "image_caption": [ + "Fig. 6: Learning Curve" + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 715, + 268 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "than $\\tau = 1.2$ , while RGSPL reaches a higher value, so considering the two indicators of SPL and RGSPL, we set $\\tau$ to 1.", + "bbox": [ + 212, + 328, + 782, + 357 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of Guidance To validate the effect of different types of guidance information, we conducted an ablation study that only adds one type of information. As shown in Table 5, adding each type of the four categories of guidance information contributes to performance improvement. As shown in Line 1, adding location information contributes mostly, which shows that letting the agent know its own location and target location is of great help. As shown in Line 4, adding previous observations information, the agent's SPL increased significantly $(2.68\\uparrow)$ , but the SR increased slightly $(0.96\\uparrow)$ . This may be attributed to the agent's ability to reference historical trajectories, thereby gaining a better understanding of the navigation environment. This, in turn, helps to avoid redundant observations and improve navigation efficiency. Destination Description and Previous Observations reduce TL from 21.96 to 19.21 and 19.22, which shows such two guidance can help the agent reach the destination faster.", + "bbox": [ + 212, + 381, + 787, + 561 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Trends of Confusion Score We plot the agent's cumulative confusion score at each time step every 5k training iterations. As shown in Figure 5, the trend of confusion score is decreasing, especially in the model using VLN-Copilot, the curve is steeper, illustrating that our model could reduce the agent's confusion degree in navigation.", + "bbox": [ + 212, + 585, + 787, + 647 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Learning Curves We compared the learning curves between methods with and without VLN-Copilot during training. As shown in Figure 6, our VLN-Copilot converges faster than the model without Copilot and can achieve better performance (i.e. higher SPL value) in fewer iterations.", + "bbox": [ + 212, + 670, + 787, + 729 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Qualitative Results", + "text_level": 1, + "bbox": [ + 214, + 753, + 390, + 768 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We present some examples to see the qualitative results of the LLM-generated guidance. As shown in Figure 7, for the example on the left, LLM selects \"Destination Description\" as the guidance type. Since the guidance information contains a description of the destination and matches candidate direction 2, it is easier to select the answer. For", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation", + "bbox": [ + 320, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/67dc9fb7d2011b9cc71af352c796d92b4d2d5bab7f9c01fc5ce20d4f37063bb4.jpg", + "image_caption": [ + "Fig. 7: Examples of the LLM-generated Guidance." + ], + "image_footnote": [], + "bbox": [ + 236, + 143, + 496, + 364 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4fd149b0527292061a4b4d3d7d387c8e2481ef87bbaccfcbabce32cf2a1c673a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 143, + 767, + 364 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "the example on the right, since Candidate 2 and Candidate 3 are both in the kitchen, and the task is to go to the kitchen, it looks like each candidate is similar, making the choice difficult. While LLM predicts that the guidance type is \"location\" and provides the agent's current location and destination, we find that the agent has reached the target location and therefore should stop rather than select a direction to go. More results can be found in supplementary material.", + "bbox": [ + 212, + 431, + 787, + 523 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Limitations and Future Work", + "text_level": 1, + "bbox": [ + 214, + 545, + 498, + 561 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we utilize Large Language Models (LLMs) to analyze the navigation environment by transforming the dynamic environment into textual descriptions. This practice may be sub-optimal for environmental perception. Our future work will focus on how to effectively apply Multimodal Large Language Models (MLLMs) such as LLaVA [23] and Video-LLaVA [21] to perceive the navigation environment and provide guidance information.", + "bbox": [ + 212, + 575, + 787, + 667 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 688, + 344, + 704 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we propose a method named VLN-Copilot that enables the agent to proactively ask for help and lets the LLM act as a co-pilot to assist the agent in completing the coarse-grained VLN task. We first introduce a confusion score calculation mechanism that allows the agent to actively query the LLM when it feels uncertain about the direction of choice. Then, we use LLM to analyze the challenges faced by the agent, and then generate appropriate guidance information for different situations that the agent is in. We conduct extensive experiments on two coarse-grained VLN tasks, including quantitative and qualitative experiments, demonstrating our method's effectiveness.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Qiao et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 145, + 382, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We thank Mr. Xiangyu Shi for the help in supplying 2 extra NVIDIA RTX 4090 GPUs for experiments.", + "bbox": [ + 212, + 178, + 785, + 209 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 233, + 310, + 250 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. An, D., Qi, Y., Li, Y., Huang, Y., Wang, L., Tan, T., Shao, J.: Bevbert: Multimodal map pre-training for language-guided navigation. In: ICCV. pp. 2737-2748 (2023)", + "2. Anderson, P., Chang, A.X., Chaplot, D.S., Dosovitskiy, A., Gupta, S., Koltun, V., Kosecka, J., Malik, J., Mottaghi, R., Savva, M., Zamir, A.R.: On evaluation of embodied navigation agents. CoRR abs/1807.06757 (2018)", + "3. Anderson, P., Wu, Q., Teney, D., Bruce, J., Johnson, M., Sünderhauf, N., Reid, I.D., Gould, S., van den Hengel, A.: Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In: CVPR. pp. 3674-3683 (2018)", + "4. Chen, H., Suhr, A., Misra, D., Snavely, N., Artzi, Y.: TOUCHDOWN: natural language navigation and spatial reasoning in visual street environments. In: CVPR. pp. 12538-12547 (2019)", + "5. Chen, S., Guhur, P., Schmid, C., Laptev, I.: History aware multimodal transformer for vision-and-language navigation. In: NeurIPS. pp. 5834-5847 (2021)", + "6. Chen, S., Guhur, P.L., Tapaswi, M., Schmid, C., Laptev, I.: Learning from unlabeled 3d environments for vision-and-language navigation. In: ECCV (2022)", + "7. Chen, S., Guhur, P.L., Tapaswi, M., Schmid, C., Laptev, I.: Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In: CVPR (2022)", + "8. Fan, Y., Chen, W., Jiang, T., Zhou, C., Zhang, Y., Wang, X.E.: Aerial vision-and-dialog navigation. In: Findings of the Association for Computational Linguistics: ACL 2023 (2023)", + "9. Feng, W., Fu, T.J., Lu, Y., Wang, W.Y.: ULN: Towards underspecified vision-and-language navigation. In: EMNLP. pp. 6394-6412 (2022)", + "0. Gao, C., Peng, X., Yan, M., Wang, H., Yang, L., Ren, H., Li, H., Liu, S.: Adaptive zone-aware hierarchical planner for vision-language navigation. In: CVPR. pp. 14911-14920 (2023)", + "1. Guhur, P., Tapaswi, M., Chen, S., Laptev, I., Schmid, C.: Airbert: In-domain pretraining for vision-and-language navigation. In: ICCV. pp. 1634-1643 (2021)", + "2. Hong, Y., Opazo, C.R., Wu, Q., Gould, S.: Sub-instruction aware vision-and-language navigation. In: Webber, B., Cohn, T., He, Y., Liu, Y. (eds.) EMNLP. pp. 3360-3376 (2020)", + "3. Hong, Y., Wu, Q., Qi, Y., Opazo, C.R., Gould, S.: A recurrent vision-and-language BERT for navigation. In: CVPR. pp. 1643-1653 (2021)", + "4. Jain, V., Magalhães, G., Ku, A., Vaswani, A., Ie, E., Baldridge, J.: Stay on the path: Instruction fidelity in vision-and-language navigation. pp. 1862-1872 (2019)", + "5. Kaufman, L., Rousseeuw, P.J.: Finding groups in data: an introduction to cluster analysis. John Wiley & Sons (2009)", + "6. Krantz, J., Wijmans, E., Majumdar, A., Batra, D., Lee, S.: Beyond the nav-graph: Vision-and-language navigation in continuous environments. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J. (eds.) ECCV. pp. 104-120 (2020)", + "7. Ku, A., Anderson, P., Patel, R., Ie, E., Baldridge, J.: Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In: EMNLP. pp. 4392-4412 (2020)", + "8. Li, J., Tan, H., Bansal, M.: Envideit: Environment editing for vision-and-language navigation. In: CVPR. pp. 15386-15396 (2022)" + ], + "bbox": [ + 225, + 266, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation", + "bbox": [ + 318, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 769, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Li, J., Li, D., Xiong, C., Hoi, S.C.H.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML (2022), https://api.sementicscholar.org/CorpusID:246411402", + "20. Li, X., Wang, Z., Yang, J., Wang, Y., Jiang, S.: Kerm: Knowledge enhanced reasoning for vision-and-language navigation. In: CVPR. pp. 2583-2592 (2023)", + "21. Lin, B., Zhu, B., Ye, Y., Ning, M., Jin, P., Yuan, L.: Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122 (2023)", + "22. Lin, K., Chen, P., Huang, D., Li, T.H., Tan, M., Gan, C.: Learning vision-and-language navigation from youtube videos. In: ICCV. pp. 8317-8326 (2023)", + "23. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)", + "24. Liu, R., Wang, X., Wang, W., Yang, Y.: Bird's-eye-view scene graph for vision-language navigation. In: ICCV. pp. 10968-10980 (2023)", + "25. Liu, S., Zhang, H., Qi, Y., Wang, P., Zhang, Y., Wu, Q.: Aerialvln: Vision-and-language navigation for uavs. In: ICCV. pp. 15384-15394 (October 2023)", + "26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR. OpenReview.net (2019)", + "27. Nguyen, K., III, H.D.: Help, anna! visual navigation with natural multimodal assistance via retrospective curiosity-encouraging imitation learning. In: Inui, K., Jiang, J., Ng, V., Wan, X. (eds.) EMNLP. pp. 684–695. Association for Computational Linguistics (2019)", + "28. Qi, Y., Pan, Z., Hong, Y., Yang, M., van den Hengel, A., Wu, Q.: The road to know-where: An object-and-room informed sequential bert for indoor vision-language navigation. In: ICCV. pp. 1655–1664 (2021)", + "29. Qi, Y., Wu, Q., Anderson, P., Wang, X., Wang, W.Y., Shen, C., van den Hengel, A.: REVERIE: remote embodied visual referring expression in real indoor environments. In: CVPR. pp. 9979-9988 (2020)", + "30. Qiao, Y., Qi, Y., Hong, Y., Yu, Z., Wang, P., Wu, Q.: Hop+: History-enhanced and order-aware pre-training for vision-and-language navigation. IEEE TPAMI (2023)", + "31. Qiao, Y., Qi, Y., Yu, Z., Liu, J., Wu, Q.: March in chat: Interactive prompting for remote embodied referring expression. In: ICCV. pp. 15758-15767 (2023)", + "32. Qiao, Y., Yu, Z., Wu, Q.: Vln-petl: Parameter-efficient transfer learning for vision-and-language navigation. In: ICCV. pp. 15443-15452 (2023)", + "33. Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert-networks. In: Inui, K., Jiang, J., Ng, V., Wan, X. (eds.) EMNLP-IJCNLP. pp. 3980-3990 (2019)", + "34. Roman, H.R., Bisk, Y., Thomason, J., Celikyilmaz, A., Gao, J.: Rmm: A recursive mental model for dialog navigation. arXiv preprint arXiv:2005.00728 (2020)", + "35. Schumann, R., Zhu, W., Feng, W., Fu, T.J., Riezler, S., Wang, W.Y.: Velma: Verbalization embodiment of llm agents for vision and language navigation in street view. In: AAAI (2023)", + "36. Song, C.H., Wu, J., Washington, C., Sadler, B.M., Chao, W.L., Su, Y.: Llm-planner: Few-shot grounded planning for embodied agents with large language models. arXiv preprint arXiv:2212.04088 (2022)", + "37. Thomason, J., Murray, M., Cakmak, M., Zettlemoyer, L.: Vision-and-dialog navigation. In: CoRL. pp. 394-406 (2019)", + "38. Touvron, H., Martin, L., Stone, K.R., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D.M., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A.S., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I.M., Korenev, A.V., Koura, P.S., Lachaux, M.A., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A.," + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Qiao et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. ArXiv abs/2307.09288 (2023), https://api.sementicscholar.org/CorpusID:259950998", + "39. Wang, X., Wang, W., Shao, J., Yang, Y.: Lana: A language-capable navigator for instruction following and generation. In: CVPR. pp. 19048-19058 (2023)", + "40. Wang, X., Huang, Q., Celikyilmaz, A., Gao, J., Shen, D., Wang, Y., Wang, W.Y., Zhang, L.: Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In: CVPR. pp. 6629-6638 (2019)", + "41. Wang, Z., Li, X., Yang, J., Liu, Y., Jiang, S.: Gridmm: Grid memory map for vision-and-language navigation. In: ICCV. pp. 15625-15636 (2023)", + "42. Wang, Z., Li, J., Hong, Y., Wang, Y., Wu, Q., Bansal, M., Gould, S., Tan, H., Qiao, Y.: Scaling data generation in vision-and-language navigation. In: ICCV. pp. 12009-12020 (2023)", + "43. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685 (2023)", + "44. Zhu, F., Liang, X., Zhu, Y., Yu, Q., Chang, X., Liang, X.: SOON: scenario oriented object navigation with graph-based exploration. In: CVPR. pp. 12689-12699 (2021)", + "45. Zhu*, Y., Weng*, Y., Zhu, F., Liang, X., Ye, Q., Lu, Y., jiao, J.: Self-motivated communication agent for real-world vision-dialog navigation. In: ICCV (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 438 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation", + "bbox": [ + 318, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/LLM as Copilot for Coarse-grained Vision-and-Language Navigation/5a4eb16e-327f-48aa-9238-afeb5ed269ac_model.json b/2024/LLM as Copilot for Coarse-grained Vision-and-Language Navigation/5a4eb16e-327f-48aa-9238-afeb5ed269ac_model.json new file mode 100644 index 0000000000000000000000000000000000000000..28a55aa33cdad411ac21e852c936d1385d4831f6 --- /dev/null +++ b/2024/LLM as Copilot for Coarse-grained Vision-and-Language Navigation/5a4eb16e-327f-48aa-9238-afeb5ed269ac_model.json @@ -0,0 +1,2555 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.322, + 0.142, + 0.68, + 0.186 + ], + "angle": 0, + "content": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.212, + 0.761, + 0.229 + ], + "angle": 0, + "content": "Yanyuan Qiao\\(^{1}\\), Qianyi Liu\\(^{2,3}\\), Jiajun Liu\\(^{4,5}\\), Jing Liu\\(^{2,3}\\), and Qi Wu\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.239, + 0.717, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\) Australian Institute for Machine Learning, The University of Adelaide" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.254, + 0.672, + 0.268 + ], + "angle": 0, + "content": "\\(^{2}\\) Institute of Automation, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.267, + 0.268, + 0.735, + 0.281 + ], + "angle": 0, + "content": "3 School of Artificial Intelligence, University of Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.45, + 0.282, + 0.553, + 0.294 + ], + "angle": 0, + "content": "4 CSIRO Data61" + }, + { + "type": "text", + "bbox": [ + 0.405, + 0.295, + 0.599, + 0.309 + ], + "angle": 0, + "content": "5 The University of Queensland" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.311, + 0.777, + 0.324 + ], + "angle": 0, + "content": "{yanyuan.qiao,qi.wu01}@adelaide.edu.au, liuqianyi2022@ia.ac.cn," + }, + { + "type": "text", + "bbox": [ + 0.326, + 0.325, + 0.669, + 0.337 + ], + "angle": 0, + "content": "jiajun.liu@csiro.au, jliu@nlpr.ia.ac.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.373, + 0.741, + 0.568 + ], + "angle": 0, + "content": "Abstract. Vision-and-Language Navigation (VLN) involves guiding an agent through indoor environments using human-provided textual instructions. Coarse-grained VLN, with short and high-level instructions, has gained popularity as it closely mirrors real-world scenarios. However, a significant challenge is these instructions are often too concise for agents to comprehend and act upon. Previous studies have explored allowing agents to seek assistance during navigation, but typically offer rigid support from pre-existing datasets or simulators. The advent of Large Language Models (LLMs) presents a novel avenue for aiding VLN agents. This paper introduces VLN-Copilot, a framework enabling agents to actively seek assistance when encountering confusion, with the LLM serving as a copilot to facilitate navigation. Our approach includes the introduction of a confusion score, quantifying the level of uncertainty in an agent's action decisions, while the LLM offers real-time detailed guidance for navigation. Experimental results on two coarse-grained VLN datasets show the efficacy of our method." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.58, + 0.698, + 0.595 + ], + "angle": 0, + "content": "Keywords: Vision-and-Language \\(\\cdot\\) Navigation \\(\\cdot\\) Large Language Models" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.62, + 0.357, + 0.635 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.817 + ], + "angle": 0, + "content": "The Vision-and-Language Navigation (VLN) tasks have attracted great attention from researchers, which enables an agent to navigate to a specified destination indoors based on the provided textual instructions and visual observations. The tasks of VLN show great potential in real-world applications such as home assistant robots. Most VLN tasks such as R2R [3] and RxR [17] etc. provide fine-grained step-by-step instructions, such as \"Go out the door in front of you across the room. Once in the hallway turn Left. Walk forward and then turn left into the sitting area. Stop in right in front of the fireplace.\" However, though these detailed instructions could improve the agents' ability of instructions following, they impede the further application of VLN in the real world since human beings would not like to give such long instructions to a robot in daily life. Contrastive to these tasks, the coarse-grained instruction task represented by the Remote" + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.826, + 0.366, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Y. Qiao et al." + }, + { + "type": "image", + "bbox": [ + 0.317, + 0.145, + 0.688, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.376, + 0.788, + 0.419 + ], + "angle": 0, + "content": "Fig. 1: Our proposed VLN-Copilot. When an agent is confused about making an action decision, it will proactively seek help from the LLM, and then the LLM will assist the agent by analyzing the situations faced by the agent and providing appropriate guidance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.788, + 0.532 + ], + "angle": 0, + "content": "Embodied Referring Expression (REVERIE) task [29] is more expected to enhance the practical utility of VLN because it involves instructions that are closer to those encountered in real-world scenarios, such as \"Go clean the mirror in the bathroom.\" Giving such coarse-grained instructions to the agent would bring more challenges to the agent's navigation than fine-grained instructions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.789, + 0.793 + ], + "angle": 0, + "content": "VLN Agents navigating in environments can benefit from seeking external assistance [34]. Early works [27, 37] tried to let agents seek help from oracles or simulators, but the assistance content obtained was fixed and limited. Since this assistance information is based on pre-built ground truth information in datasets or simulators, which are not consistent with real-life scenarios. Inspired by the development of Large Language Models (LLMs) with rich internalized world knowledge, many works have been made to exploit their capabilities to aid navigation tasks. One recent method named MiC [31] makes the first step to use LLM in the coarse-grained VLN task of REVERIE, which queries the LLM to generate more detailed instructions by prompting the LLM with extra planning examples. The promising performance of MiC shows the great potential to use LLM in coarse-grained VLN tasks. However, MiC has some weaknesses in applying LLM for VLN. First, the way the agent seeks help from the LLM is a passive way that depends on the change of the room where the agent is located. This approach may not help the agent when it faces difficulties in making predictions while still in the same room. Second, MiC generates fine-grained planning no matter what challenges the agent faces. This practice lacks a thorough analysis of the difficulties the agent may have and fails to give the agent diverse guidance for different scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.84 + ], + "angle": 0, + "content": "To address these issues, we propose a novel method named VLN-Copilot that makes the LLM a competent copilot to assist the agent in navigation. As shown in Figure 1, before making a movement, the agent calculates a confusion score to assess the level of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.32, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "uncertainty associated with its current action choice. If this score surpasses a predefined threshold, it triggers the agent to seek guidance. This practice is more proactive than relying on the change of scenes which may be wrong. When the agent decides to seek assistance, it consults the LLM and then the LLM conducts a comprehensive evaluation of the situation, identifies the type of guidance required, and generates the necessary guidance to aid the agent in navigating effectively. The types of guidance information are selected from the in-depth pre-analysis of prevailing challenges the agent may face in the sampled environment using LLM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.328 + ], + "angle": 0, + "content": "We conduct experiments on the REVERIE and CVDN-target dataset and our model surpasses the previous methods. Specifically, our VLN-Copilot obtains \\(42.32\\%\\) on the main navigation metric SPL and \\(26.55\\%\\) on the main object grounding metric RGSPL on the REVERIE dataset, and 4.47 in GP on the CVDN-target dataset, respectively." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.329, + 0.542, + 0.344 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.353, + 0.784, + 0.395 + ], + "angle": 0, + "content": "- We propose a novel method named VLN-Copilot which utilizes LLM to analyze the difficulties the agent may meet and give diverse guidance for different situations, rather than to generate planning by prompting LLM with extra planning examples." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.399, + 0.784, + 0.425 + ], + "angle": 0, + "content": "- We introduce the confusion score to help the agent proactively ask the LLM for help when confused to make predictions rather than passively trigger conversations." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.429, + 0.784, + 0.47 + ], + "angle": 0, + "content": "- We utilize LLM with elaborately designed prompts to conduct a detailed challenge analysis of the agent's different situations, enabling the LLM to provide better guidance on the fly." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.474, + 0.784, + 0.501 + ], + "angle": 0, + "content": "- Extensive experiments on two VLN benchmarks show the effectiveness of our proposed method." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.353, + 0.784, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.527, + 0.368, + 0.543 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.559, + 0.488, + 0.575 + ], + "angle": 0, + "content": "2.1 Vision-and-Language Navigation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.659 + ], + "angle": 0, + "content": "In recent years, Vision-and-Language Navigation (VLN) has attracted widespread attention due to its promising applications in home assistant robots and related fields [9, 12, 18, 20, 28, 40]. Many different tasks have emerged in the field of VLN [4, 14, 16, 25, 27, 29, 37]. In terms of the navigation space accessible to the agent, these cover indoor [3, 16, 17], outdoor [4, 35], and aerial scenarios [8, 25]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Early VLN task R2R [3] started from indoor environments requiring the agent to reach the destination according to the given instructions, and there were some subsequent tasks (e.g., RxR [17] and VLN-CE [16]). These tasks all provide detailed natural language instructions, such as \"Walk down the stairs at the bottom of the stairs take a left. Walk straight to get to the kitchen inside and wait just inside the door.\" These detailed instructions guide the agent through complex step-by-step processes and are more focused on testing the agent's ability to understand and follow instructions. Some other tasks adopt an object-oriented perspective, such as REVERIE [29] and SOON [44]. These tasks require not only reaching a designated destination but also discovering specific target objects. Different from SOON providing detailed instructions and object recognition prompts, such as \"The cabinet is square, closed and wooden. The cabinet is placed below the sink and the water-tap. There are some stools and table nearby.\"" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Y. Qiao et al." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.222 + ], + "angle": 0, + "content": "REVERIE's instructions appear more concise and natural, such as \"Bring me the swivel chair from the office.\" These coarse-grained instructions, akin to instructions encountered in real-world scenarios, enhance the potential for innovative research within the domain of coarse-grained instruction VLN. However, the task remains challenging due to the limited information these instructions provide." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.246, + 0.458, + 0.262 + ], + "angle": 0, + "content": "2.2 VLN Agents Asking for Help" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.273, + 0.787, + 0.65 + ], + "angle": 0, + "content": "Agents can gain advantages by seeking external assistance to assist navigation, and there have been studies promoting navigation agents to seek help during navigation tasks [27, 34]. Thomason et al. [37] introduce the Navigation from Dialog History (NDH) task, which contains dialog history between agents and oracles during navigation. The agent follows guide instructions and asks questions to the oracle when needed. The task provides a valuable resource for investigating the navigation interactions between agent and oracle. Similarly, HANNA [27] builds a simulator where navigation agents can request language-and-vision assistance when they face difficulties during navigation. However, the agent's signals asking for help are fixed and predefined in the simulator. Additionally, the assistance provided is based on ground truth information derived from the R2R dataset, which may not fully represent real-world navigation scenarios due to its reliance on pre-built datasets. Similarly, Zhu et al. proposed SCoA [45], which enables navigation agents to ask questions from a set of pre-constructed candidate question-answer pairs to aid in navigation. Recently, LLM-Planner [36] and MiC [31] have attempted to leverage the knowledge of large language models (LLM) to assist agents in navigation. Agents can query the LLM to generate detailed step-by-step instructions to aid navigation. However, the ways the agents ask for help from both methods are passive, either based on fixed time steps or when the room the agent is in changes. This practice may miss the opportunity to help the agent when it has difficulties in prediction actions. Besides, these two methods concentrate on generating planning by prompting LLM with extra planning data. Different from these methods, we propose a method named VLN-Copilot that allows agents to proactively seek help when they are not sure where to go in this work. Moreover, our VLN-Copilot pays more attention to analyzing the difficulties the agent may meet and can dynamically give appropriate types of guidance in diverse situations for the agent." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.675, + 0.315, + 0.691 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.708, + 0.35, + 0.723 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The main goal for the agent in the VLN task is to navigate to a specific destination following the high-level concise instructions. Specifically, the agent is asked to navigate in the simulated environment, which can be described as an undirected graph \\(\\mathcal{G} = \\{\\mathcal{V},\\mathcal{E}\\}\\), where \\(\\mathcal{V}\\) represents the navigable nodes and \\(\\mathcal{E}\\) represents the connectivity edges. Given the natural language instruction \\(L\\), the agent starts from the initial node \\(V_{0}\\) and perceives a panoramic view \\(\\mathcal{R}_t\\) of the current node \\(V_{t}\\) to make the action \\(a_{t}\\) at each time step \\(t\\). The panoramic view \\(\\mathcal{R}_t\\) has \\(K\\) single-view images \\(r_i\\) and can be represented as" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.32, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.154, + 0.752, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.308, + 0.788, + 0.392 + ], + "angle": 0, + "content": "Fig.2: Overview of the VLN-Copilot. At each time step, the agent predicts the logits over the candidate's view of actions. After that, the agent will calculate the confusion score to decide whether to ask LLM for help (Sec. 3.3). When the confusion score surpasses threshold \\(\\tau\\), the agent queries the LLM for help with the prompt that helps LLM perceive the environment (Sec. 3.4). Then the LLM generates useful guidance information, which will be returned to the agent to assist in decision-making (Sec. 3.6)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.788, + 0.543 + ], + "angle": 0, + "content": "\\(\\mathcal{R}_t = \\{r_i\\}_{i=1}^K\\), of which \\(r_i\\) is represented by an image feature vector with the corresponding orientation feature vector. Besides, for the sub-task of object grounding, the object features \\(\\mathcal{O}_t = \\{o_i\\}_{i=1}^M\\) are extracted from the panoramic view \\(\\mathcal{R}_t\\) with the annotated object bounding boxes or by object detectors. With these observations of the environment, the agent takes a series of actions \\(\\langle a_0, \\dots, a_N \\rangle\\) to navigate to the target location, where each action \\(a_t\\) selects one navigable node from the candidates. When the target object is grounded or the pre-defined maximum navigation steps have been achieved, the navigation ends." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.57, + 0.325, + 0.584 + ], + "angle": 0, + "content": "3.2 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.844 + ], + "angle": 0, + "content": "As is shown in Figure 2 and Algorithm 1, given the coarse-grained instruction, the agent first perceives the environment to obtain the current visual observation and starts to predict a probability over candidate actions. Then, a confusion score is computed according to the probability distribution, which determines whether the agent is confused about making predictions in the current situation and whether the agent should ask the LLM for help. If the confusion score exceeds a predefined threshold \\(\\tau\\) (obtained via grid search) which means the agent is confused about making decisions, the agent initiates a conversation with the LLM to seek their help with an elaborately designed prompt. Specifically, the prompt contains coarse-grained instruction, fine-grained scene descriptions of candidate views, historical descriptions of trajectory scenes, and the requirements for the LLM. Then, the LLM analyzes the agent's situation and gives advice about which type of guidance information is useful as well as the details about that guidance information. The types of guidance information are refined from the results of challenge pre-analysis by LLM. Finally, the response about the guidance from the LLM is appended after the coarse-grained instruction to guide the agent's action prediction collaboratively. This paradigm aims to optimize agents' decision-making processes by" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Y. Qiao et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.787, + 0.178 + ], + "angle": 0, + "content": "combining instruction, observation, and LLM-based world knowledge via the conversations between the agent and LLM." + }, + { + "type": "code", + "bbox": [ + 0.289, + 0.211, + 0.72, + 0.46 + ], + "angle": 0, + "content": "Algorithm 1 VLN-Copilot \nNotation Summary: I: high-level instruction \\(\\mathcal{R}_t\\) : visual observation at timestep t \\(S_{t}\\) : confusion score at timestep t \\(W_{t}^{F}\\) : fine-grained scene descriptions of all candidate views at timestep t \\(W_{t}^{\\mathrm{Tr}}\\) : historical trajectory description at timestep t \\(T_{\\mathrm{guide}}\\) : prompt template for LLM to generate guidance information LLM: large language model \\(\\tau\\) : pre-defined threshold for confusion score \n\\(t\\gets 0\\) \\(\\rightharpoondown\\) Initial timestep while \\(t < \\max\\) step and \\(\\hat{a}_{t}\\neq\\) \"stop\" do if \\(S_{t} > = \\tau\\) then \\(P_{\\mathrm{guide}}\\leftarrow T_{\\mathrm{guide}}(I,W_t^F,W_t^{\\mathrm{Tr}})\\) \\(R_{\\mathrm{guide}}\\gets \\mathrm{LLM}(P_{\\mathrm{guide}})\\) \\(I_{t}^{\\mathrm{Inter}}\\gets \\mathrm{CONCAT}(I,R_{\\mathrm{guide}})\\) \\(\\hat{a}_{t}\\gets \\mathrm{Agent}(I_{t}^{\\mathrm{Inter}},\\mathcal{R}_{t})\\) else \\(\\hat{a}_{t}\\gets \\mathrm{Agent}(I,\\mathcal{R}_{t})\\) end if \\(t\\gets t + 1\\) \nend while" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.479, + 0.454, + 0.496 + ], + "angle": 0, + "content": "3.3 When to Ask LLM for Help" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.789, + 0.597 + ], + "angle": 0, + "content": "Different from previous methods [31] that set fixed timestep for interactions with LLM or trigger the interactions when the room scenes change, we select a more active practice for the agent by using the confusion score to decide whether to ask for help. Specifically, at each time step \\( t \\), when the agent is going to predict the logits over the candidate nodes for the next action \\( a_{t} \\), we calculate the entropy of the predicted probability distribution over each candidate navigable node \\( v_{i} \\) as the confusion score \\( S_{t} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.608, + 0.787, + 0.626 + ], + "angle": 0, + "content": "\\[\np _ {t} \\left(v _ {i}\\right) = \\operatorname {S o f t m a x} \\left(s _ {t} \\left(v _ {i}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.628, + 0.787, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {t} = - \\sum_ {i} p _ {t} \\left(v _ {i}\\right) \\log_ {2} \\left(p _ {t} \\left(v _ {i}\\right) + \\epsilon\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.67, + 0.785, + 0.699 + ], + "angle": 0, + "content": "where \\( s_t(v_i) \\) is the predicted logits assigned by the agent to the \\( i \\)-th navigable node \\( v_i \\) at time step \\( t \\), and \\( \\epsilon \\) is a minimal number added to avoid numerical instability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.787, + 0.746 + ], + "angle": 0, + "content": "Then, if the computed confusion score \\( S_{t} \\) surpasses a predefined threshold \\( \\tau \\), which means the agent is confused about making a decision among some candidate nodes, it will send a message to LLM asking for help." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.769, + 0.522, + 0.784 + ], + "angle": 0, + "content": "3.4 How LLM Perceives the Environment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "LLMs would give inaccurate advice for the agent's navigation since they are not able to directly acquire visual content from the environment. Both current visual observations from the agent and previously observed scenes are essential for the LLM to analyze the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.32, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "navigation and supply important information. Thus, enhancing the LLM's environmental awareness as far as possible is a critical first step to making it a competent copilot. To solve this problem, we translate the dynamic observations and trajectories into multi-grained spatiotemporal descriptions as the eye for the LLM, using the out-of-the-box vision-and-language model of BLIP [19] model. Detailedly, the visually-grounded description mainly consists of three parts: the coarse-grained scene description of the current panoramic view \\(\\mathcal{R}_t\\), the fine-grained scene description of each view \\(r_i\\) of selected candidate navigable nodes at the current time step \\(t\\), and the trajectory description of previously observed scenes before the current time step \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.389 + ], + "angle": 0, + "content": "Specifically, we obtain both the panoramic view image \\(\\mathcal{R}_t\\) of the agent's observation and the single view image \\(r_i\\) of each candidate navigable node at the time step \\(t\\). We first feed \\(\\mathcal{R}_t\\) to the BLIP model to generate the coarse-grained scene description \\(W_t^C\\) for the agent's current location. For the fine-grained scene description, the agent is provided with \\(k\\) candidate views for selection. Then, we feed each single-view image \\(r_i\\) of the \\(k\\) selected candidate navigable nodes to BLIP to generate the fine-grained scene description \\(W_t^{F_i}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.445, + 0.401, + 0.786, + 0.417 + ], + "angle": 0, + "content": "\\[\nW _ {t} ^ {\\mathrm {C}} = \\operatorname {B L I P} \\left(\\mathcal {R} _ {t}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.439, + 0.422, + 0.785, + 0.438 + ], + "angle": 0, + "content": "\\[\nW _ {t} ^ {F _ {i}} = \\operatorname {B L I P} \\left(r _ {i}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.45, + 0.788, + 0.496 + ], + "angle": 0, + "content": "Then we concatenate all previous coarse-grained scene descriptions \\( W^{\\mathrm{C}} \\) of each time step before the current time \\( t \\) as the historical trajectory description \\( W_{t}^{\\mathrm{Tr}} \\) such as \"I passed a bedroom, and then I passed ...\" This process can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.507, + 0.786, + 0.524 + ], + "angle": 0, + "content": "\\[\nW _ {t} ^ {\\mathrm {T r}} = \\operatorname {C O N C A T} \\left(W _ {1} ^ {\\mathrm {C}}, \\dots , W _ {t - 1} ^ {\\mathrm {C}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.657 + ], + "angle": 0, + "content": "Note that we only generate the fine-grained scene descriptions and trajectory descriptions when the agent is confused to make the action prediction and asks the LLM for help, while we generate coarse-grained scene descriptions at each time step. In addition, for better comprehension of LLM, we added the candidate view id \\( i \\) as the prefix to each corresponding native fine-grained scene description \\( W_{t}^{F_{i}} \\). For example, we converted the generated scene description of \"A room with a tilted ceiling, a bathroom mirror on the wall, and a skylight above the window.\" into the candidate view description with the following format: \"I see candidate 0 is a room with a tilted ceiling ...\"" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.679, + 0.607, + 0.695 + ], + "angle": 0, + "content": "3.5 Difficulty and Useful Guidance-Type Pre-Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Before asking the agent to query an LLM to generate guidance for assistance, it would be beneficial to explore the difficulties the agent may encounter in making decisions and what guidance information the LLM should provide to be useful for the agent's navigation. However, manual analysis by humans is costly and impractical. Thus, we turn to LLM for analysis. Specifically, we first collect some navigation samples. When the agent finds it difficult to make decisions during the navigation process, we convert the situation the agent encountered into a text description and save it as a navigation sample. Then, we use an elaborate prompt to query the LLM to analyze the situation encountered by the agent in the current navigation example, such as what is the reason" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Y. Qiao et al." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.151, + 0.773, + 0.171 + ], + "angle": 0, + "content": "System: You are a helpful assistant for navigation. There is an agent navigating indoors according to task instructions and current observation. At each step, the agent needs to choose one of the candidate directions to go. I will tell you which direction is correct." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.175, + 0.772, + 0.193 + ], + "angle": 0, + "content": "You need to first analyze the current situation, such as what difficulties the agent encountered during navigation that made it difficult to choose a direction and what types of guidance information may be useful." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.198, + 0.457, + 0.207 + ], + "angle": 0, + "content": "Task instruction: Turn off the white table lamp in the bedroom" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.206, + 0.279, + 0.214 + ], + "angle": 0, + "content": "Observation:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.214, + 0.536, + 0.223 + ], + "angle": 0, + "content": "I see candidate 1 is a white door leads to a bedroom with a bed, electric fan and table" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.223, + 0.525, + 0.231 + ], + "angle": 0, + "content": "I see candidate 2 is a bathroom has a sink and the door leads to the shower room" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.231, + 0.532, + 0.238 + ], + "angle": 0, + "content": "I see candidate 3 is a hallway with white walls, stairs, and a door on the second floor" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.238, + 0.682, + 0.246 + ], + "angle": 0, + "content": "History trajectory: I have passed a balcony with rocking chairs and a table, then I go to a hallway with a picture of a bedroom." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.245, + 0.335, + 0.252 + ], + "angle": 0, + "content": "Correct answer: candidate 1" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.258, + 0.282, + 0.266 + ], + "angle": 0, + "content": "Assistant:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.266, + 0.771, + 0.274 + ], + "angle": 0, + "content": "Difficulty: The agent encountered difficulties in choosing a direction because there are multiple doors at the same time. This makes it challenging for" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.274, + 0.437, + 0.282 + ], + "angle": 0, + "content": "the agent to determine which direction could find the lamp." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.288, + 0.43, + 0.297 + ], + "angle": 0, + "content": "Please list possible type of guidance may be useful." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.304, + 0.281, + 0.311 + ], + "angle": 0, + "content": "Assistant:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.311, + 0.569, + 0.32 + ], + "angle": 0, + "content": "Guidance Type: The agent needs to the destination information to make an informed decision." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.353, + 0.788, + 0.395 + ], + "angle": 0, + "content": "Fig. 3: Prompt template for LLM to pre-analyze the difficulties the agent encounters and what type of guidance information maybe useful. The color of red, blue represent multi-turn responses generated from LLM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.788, + 0.473 + ], + "angle": 0, + "content": "for the difficulty in making a decision, and the helpful types of guidance. Finally, we collect the guidance information generated by LLM and cluster it to obtain the final guidance categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.789, + 0.619 + ], + "angle": 0, + "content": "Navigation Samples Collection We first let the agent navigate around the environment with confusion scores computed. If the confusion score surpasses the pre-defined threshold at the time step \\( t \\), which means the agent comes across difficulties in making the action prediction, then we collect the tuple \\( (I, W_{t}^{F}, W_{t}^{Tr}, A_{t}) \\) of the high-level task instruction \\( I \\), fine-grained candidate scene descriptions \\( W_{t}^{F} \\), the trajectory description \\( W_{t}^{Tr} \\) and the ground-truth action \\( A_{t} \\) as a sample data \\( D_{\\text{sample}} \\) for analysis. Note that \\( W_{t}^{F} \\) includes \\( k \\) fine-grained scene descriptions of the selected candidate views. In total, we randomly collect 5,000 samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Propmt Design To make the LLM better analyze the situation that the agent meets, we elaborately designed a prompt template for multi-turn query-and-response. We first set the unified system prompt \\( P_{\\text{system}} \\) that requires the LLM as a helpful assistant for navigation. Then, we transform each sample data \\( D_{\\text{sample}} \\) into sample prompt \\( P_{\\text{sample}} \\) by respectively adding the prefix of \"Task instruction:\", \"Observation:\", \"History trajectory:\" and \"Correct answer:\" into \\( I, W_t^F, W_t^{Tr} \\) and \\( A_t \\). Next, we decompose the task of analyzing the challenge into two sub-tasks, the first of analyzing what difficulties the agent may meet in making decisions and the second of analyzing what types of guidance may be useful for navigation in an iterative way. Correspondingly, we use \"########Assistant: Difficulty:\" and \"########Assistant: Guidance Type:\" as the sub-task prompts \\( P_{\\text{diff}} \\) and \\( P_{\\text{guide}} \\). We concatenate the prompts of \\( P_{\\text{system}} \\), \\( P_{\\text{sample}} \\) and \\( P_{\\text{diff}} \\) as \\( P_{\\text{task1}} \\) and feed \\( P_{\\text{task1}} \\) to LLM, which will generate the response \\( R_{\\text{diff}} \\) about the difficulty the agent may meet. We further feed \\( P_{\\text{task2}} \\) to LLM, the concatenation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.32, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.15, + 0.77, + 0.169 + ], + "angle": 0, + "content": "System: You are a helpful assistant for navigation. There is an agent navigating indoors according to task instructions and current observation. At each step, the agent needs to choose one of the candidate directions to go." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.174, + 0.769, + 0.192 + ], + "angle": 0, + "content": "You need to analyze what additional guidance information the agent need to make the choice. There are four types of information: location, prioritization, destination description, previous observations. You can only give one answer from these types." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.196, + 0.457, + 0.205 + ], + "angle": 0, + "content": "Task instruction: Turn off the white table lamp in the bedroom" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.205, + 0.282, + 0.212 + ], + "angle": 0, + "content": "Observation:" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.212, + 0.536, + 0.219 + ], + "angle": 0, + "content": "I see candidate 1 is a white door leads to a bedroom with a bed, electric fan and table" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.219, + 0.523, + 0.226 + ], + "angle": 0, + "content": "I see candidate 2 is a bathroom has a sink and the door leads to the shower room" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.226, + 0.533, + 0.234 + ], + "angle": 0, + "content": "I see candidate 3 is a hallway with white walls, stairs, and a door on the second floor" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.234, + 0.681, + 0.242 + ], + "angle": 0, + "content": "History trajectory: I have passed a balcony with rocking chairs and a table, then I go to a hallway with a picture of a bedroom." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.248, + 0.284, + 0.256 + ], + "angle": 0, + "content": "####Assistant:" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.256, + 0.376, + 0.265 + ], + "angle": 0, + "content": "Guidance Type: Destination description" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.271, + 0.393, + 0.28 + ], + "angle": 0, + "content": "Please give description of the destination." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.286, + 0.284, + 0.293 + ], + "angle": 0, + "content": "Assistant:" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.293, + 0.768, + 0.309 + ], + "angle": 0, + "content": "The bedroom is a cozy and comfortable space, typically with a queen-sized bed in the center of the room. The bed is dressed with crisp, white linens and fluffy pillows." + }, + { + "type": "image_caption", + "bbox": [ + 0.3, + 0.342, + 0.701, + 0.356 + ], + "angle": 0, + "content": "Fig. 4: Prompt template for LLM to generate guidance information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.789, + 0.422 + ], + "angle": 0, + "content": "of \\( P_{task1} \\), \\( R_{diff} \\) and \\( P_{guide} \\) and we will get the response \\( R_{guide} \\) about what types of guidance may be useful for navigation. The prompt template is depicted in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.789, + 0.697 + ], + "angle": 0, + "content": "Guidance-Type Clustering After the above-mentioned operations, we obtain a collection of the guidance response \\( R_{guide} \\). We first use Sentence-BERT [33] to embed each \\( R_{guide} \\) and get the corresponding embedding \\( E_{guide} \\). Then we employ the silhouette coefficient [15] to determine the optimal number of clusters for K-Means clustering. The silhouette coefficient is an indicator to evaluate the quality of clustering. It combines the two factors of clustering cohesion and separation, calculates a silhouette coefficient value for each sample, and then calculates a silhouette coefficient value through all samples. By calculating the average silhouette coefficient under different cluster numbers \\( k \\), the clustering quality under each cluster number can be evaluated. In general, the number of clusters with the highest average silhouette coefficient is considered the optimal number of clusters because it provides the best balance of cohesion and separation. And clustering result shows that the best number of clusters is seven. In addition, after analysis based on the practical situation, we found that some of the guidance content is similar, and some guidance is not executable, such as \"Room Layout\", which cannot be obtained from the environment. Thus, we manually merge these duplicate types, remove infeasible types, and finally reserve four applicable categories:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.713, + 0.747, + 0.727 + ], + "angle": 0, + "content": "- Location: information about the agent's current position and the destination." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.73, + 0.785, + 0.759 + ], + "angle": 0, + "content": "- Prioritization: information about the urgent task that needs to be completed based on an assessment of agent progress." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.762, + 0.785, + 0.79 + ], + "angle": 0, + "content": "- Destination Description: detailed information about the target location, providing more cues to reach its destination." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.794, + 0.785, + 0.824 + ], + "angle": 0, + "content": "- Previous Observations: information about the agent's navigation trajectory, lets the agent remember history information." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.713, + 0.785, + 0.824 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Y. Qiao et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.5, + 0.163 + ], + "angle": 0, + "content": "3.6 LLM Guides the Agent On-the-Fly" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.175, + 0.787, + 0.235 + ], + "angle": 0, + "content": "After the pre-analysis of the agent's navigation, we employ the analysis result and a new prompt to ask LLM to aid the agent on the fly when it is confused to make decisions. We first illustrate the modification of the prompt for the on-the-fly guidance. Then, we show how LLM guides the agent with different types of guidance information." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.397 + ], + "angle": 0, + "content": "Prompt Design We also designed the prompt template to ask the LLM to analyze the agent's current situation and decide which guidance type from the selected four categories might be useful in that situation. This prompt template is similar to the pre-analysis prompt template, but with some modifications to guide agent navigation. The full prompt template is shown in Figure 4. First, we modify the system prompt \\( P_{system} \\) by asking the LLM to choose which guidance information is useful from the four pre-selected categories instead of analyzing the difficulties and generating unlimited guidance information. Second, we delete the ground truth action in the sample prompt \\( P_{sample} \\). At last, the task prompt is revised to \"########Assistant: Guidance Type:\"." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.421, + 0.787, + 0.648 + ], + "angle": 0, + "content": "Integrating Guidance with Instruction During the navigation of the agent, when the confusion score \\( S_{t} \\) surpasses the pre-defined threshold \\( \\tau \\) at time step \\( t \\), it will ask the LLM for guidance. With the help of the above-mentioned prompts, the LLM can perceive the environment and trajectory of the agent, analyze the situation, and select the guidance type from the pre-selected four categories. For different types of guidance, the LLM will generate different guidance information \\( I_{t}^{G} \\) to assist the navigation. Specifically, the LLM will generate detailed content about the type of \"Location\", \"Prioritization\" and \"Destination Description\" according to its internalized world knowledge. For these three types of guidance, the generation process is also interactive, which first generates the type of information and then generates the details according to the type of information. Such a two-step process ensures the quality of generated guidance. For the type of \"Previous Observations\", we re-use the historical trajectory as the guidance information. Some examples of generated guidance information can be found in Figure 7. Finally, we append the guidance information \\( I_{t}^{G} \\) to the high-level instruction \\( I \\), and then the agent will navigate under the guidance of the integrated instruction \\( I_{t}^{Inter} \\)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.673, + 0.35, + 0.69 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.707, + 0.314, + 0.723 + ], + "angle": 0, + "content": "4.1 Settings" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.734, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Datasets We evaluate our method on two VLN tasks: REVERIE and CVDN-target, which are all based upon the MP3D indoor environments. Compared to Room-to-Room (R2R) task, these instructions are high-level and concise, such as \"Bring me the swivel chair from the office\". CVDN-target is our proposed VLN setup similar to REVERIE. The instruction contains a target and conversation history. Here we remove the conversation history and only keep the information containing the target as the instruction, such as \"To find a book in the room.\"" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.321, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.295, + 0.145, + 0.707, + 0.159 + ], + "angle": 0, + "content": "Table 1: Comparison with the state-of-the-art methods on REVERIE." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.172, + 0.77, + 0.453 + ], + "angle": 0, + "content": "
MethodsVal UnseenTest Unseen
NavigationGroundingNavigationGrounding
TLOSR↑SR↑SPL↑RGS↑RGSPL↑TLOSR↑SR↑SPL↑RGS↑RGSPL↑
Human------21.1886.8381.5153.6677.8451.44
RecBERT [13]16.7835.0230.6724.9018.7715.2715.8632.9129.6123.9916.5013.51
Airbert [11]18.7134.5127.8921.8818.2314.1817.9134.2030.2823.6116.8313.28
HAMT [5]14.0836.8432.9530.2018.9217.2813.6233.4130.4026.6714.8813.08
VLN-PETL [32]14.4737.0331.8127.6718.2615.9614.0036.0630.8326.7315.1313.03
LANA [39]16.2838.5434.0029.2619.0316.1816.7536.4133.5026.8917.5314.25
HOP+ [30]14.5740.0436.0731.1322.4919.3315.1735.8133.8228.2420.2016.86
DUET [7]22.1151.0746.9833.7332.1523.0321.3056.9152.5136.0631.8822.06
Lily [22]21.8753.7148.1134.4332.1523.4321.9460.5154.3237.3432.0221.94
BEVBert [1]-56.4051.7836.3734.7124.44-57.2652.8136.4132.0622.09
AutoVLN [6]-62.1455.8940.8536.5826.76-62.3055.1738.8832.2322.68
AZHP [10]22.3253.6548.3136.6334.0025.79-55.3151.5735.8532.2522.44
BSG [24]24.7158.0552.1235.5935.3624.2422.9062.8356.4538.7033.1522.34
KERM [20]21.8555.2150.4435.3834.5124.4517.3257.5852.4339.2132.3923.64
ScaleVLN [42]--56.9741.8435.7626.05--56.1339.5232.5322.78
GRidMM [41]23.2057.4851.3736.4734.5724.5619.9759.5553.1336.6034.8723.45
MiC [31]20.6462.3756.9743.6037.5228.7218.1162.4055.7441.9735.2526.17
VLN-Copilot21.8962.6257.4043.6338.8829.7521.7263.2657.8142.3236.5626.55
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.543 + ], + "angle": 0, + "content": "Evaluation Metrics We utilize the standard evaluation metrics to evaluate our method, we report Success Rate (SR), Success weighted by Path Length (SPL) [2], and Goal Progress (GP) as the main navigation metrics, Remote Grounding Success rate (RGS) and RGS weighted by Path Length (RGSPL) as the main object grounding metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.716 + ], + "angle": 0, + "content": "Implementation Details We implement our proposed method using PyTorch. The model is trained on one single NVIDIA RTX 4090 GPU. For REVERIE, we fine-tuned the pre-trained model from [6] with a batch size of 4 for 100k iterations. For the CVDN-target, we fine-tuned the pre-trained model from [7] with a batch size of 4 for 50k iterations. AdamW [26] optimizer is adopted and the learning rate is set to \\(1 \\times 10^{-5}\\). The threshold \\(\\tau\\) is set as 1. For the baseline agent, we use the recent state-of-the-art model of AutoVLN [6]. When collecting data for pre-analysis, we use two agents based on pre-trained AutoVLN [6] and HAMT [5] to avoid agent bias. LLM is integrated into the training process, we use the open-source Vicuna-7B-1.5 model [43], which is fine-tuned from Llama 2 [38]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.739, + 0.56, + 0.755 + ], + "angle": 0, + "content": "4.2 Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.792, + 0.842 + ], + "angle": 0, + "content": "REVERIE As shown in Table 1, our method achieves competitive performance on the REVERIE benchmark compared to previous methods, such as AutoVLN [6], Mic [31] and ScaleVLN [42]. Particularly, compared with the SoTA method MiC [31], our method outperforms Mic in terms of the object grounding metric RGSPL with \\(1.36\\%\\) and RGSPL with \\(1.03\\%\\) on the Val Unseen split. It is worth noting that compared with" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Y. Qiao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.419, + 0.173 + ], + "angle": 0, + "content": "Table 2: Comparison on CVDN-target Validation Unseen." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.185, + 0.393, + 0.251 + ], + "angle": 0, + "content": "
MethodsTLGP↑
HAMT† [5]37.042.74
DUET† [7]78.533.70
AutoVLN† [6]61.973.98
VLN-Copilot40.384.47
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.273, + 0.419, + 0.3 + ], + "angle": 0, + "content": "Table 4: Ablation of different threshold \\( \\tau \\) ." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.314, + 0.414, + 0.355 + ], + "angle": 0, + "content": "
τ0.811.21.4
SPL↑42.7843.6343.6842.55
RGSPL↑28.3529.7528.1728.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.152, + 0.719, + 0.166 + ], + "angle": 0, + "content": "Table 3: Ablation of LLM Analyzer." + }, + { + "type": "table", + "bbox": [ + 0.435, + 0.179, + 0.787, + 0.245 + ], + "angle": 0, + "content": "
MethodNavigationGrounding
TLOSR↑SR↑SPL↑RGS↑RGSPL↑
No-Selection (Baseline)21.9657.8252.7740.1534.7626.37
All-Inclusive20.7059.7655.5041.3636.4727.33
Random-Selection20.1558.0253.7942.1635.3027.08
LLM-Selection (Ours)21.8962.6257.4043.6338.8829.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.451, + 0.264, + 0.766, + 0.278 + ], + "angle": 0, + "content": "Table 5: Ablation study of the guidance information." + }, + { + "type": "table", + "bbox": [ + 0.435, + 0.291, + 0.787, + 0.364 + ], + "angle": 0, + "content": "
MethodNavigationGrounding
TLOSR↑SR↑SPL↑RGS↑RGSPL↑
Baseline21.9657.8252.7740.1534.7626.37
+ Location20.9762.2056.8042.9737.7428.46
+ Prioritization20.4859.4754.7041.7435.8627.40
+ Destination Description19.2159.5953.9941.9134.9927.11
+ Previous Observations19.2257.9653.7342.8335.6427.12
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.391, + 0.787, + 0.422 + ], + "angle": 0, + "content": "ScaleVLN [42], which collects 4.9M instruction-trajectory pairs to train the model, our method still achieves better performance (2.8% on SPL and 3.77% on RGSPL)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.788, + 0.549 + ], + "angle": 0, + "content": "CVDN-target As shown in Table 2, we compare VLN-Copilot with three main-stream VLN models HAMT [5], DUET [7], and AutoVLN [6] on the CVDN-target dataset. For a fair comparison, both models are trained on the CVDN-target dataset. \\(\\dagger\\) represents the reproduction results of these methods. It highlights the effectiveness of VLN-Copilot, especially in terms of GP metrics, where it outperforms all other methods. Although TL is slightly higher than HAMT, its significantly improved GP emphasizes its overall advantage in handling unseen validation data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.569, + 0.364, + 0.584 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.593, + 0.789, + 0.79 + ], + "angle": 0, + "content": "Effect of LLM Analyzer To evaluate the ability of LLM for guidance type selection, we conduct an ablation study, there are four settings, as follows: (1) No-Selection: our baseline, which means the agent does not use any guidance information; (2) All-Inclusive: the agent lets the LLM generate all types of guidance and integrates this information; (3) Random-Selection: the agent randomly selects one of four guidance types and lets the LLM generate corresponding guidance information; (4) LLM-Selection: our method, the LLM outputs guidance type and generates the information. As illustrated in Table 3, we can see that the LLM-Selection method achieved the highest navigation main metric spl with a value of \\(43.63\\%\\), followed by the Random-Selection method with a score of \\(42.16\\%\\). This suggests that blindly combining all types of guidance information is not the best option, probably due to the redundancy of excessively detailed information. Compared to the Baseline, the LLM-Selection method improves the SR by \\(4.63\\%\\), SPL by \\(3.48\\%\\), RGS by \\(4.12\\%\\), and \\(3.38\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Different Threshold of Confusion Score As shown in Table 4, we conduct an ablation study on threshold \\(\\tau\\) in the confusion score. Although when \\(\\tau = 1\\), SPL is slightly lower" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.321, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.145, + 0.501, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.284, + 0.466, + 0.298 + ], + "angle": 0, + "content": "Fig.5:Confusion Score" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.145, + 0.717, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.54, + 0.284, + 0.678, + 0.298 + ], + "angle": 0, + "content": "Fig. 6: Learning Curve" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.784, + 0.358 + ], + "angle": 0, + "content": "than \\(\\tau = 1.2\\), while RGSPL reaches a higher value, so considering the two indicators of SPL and RGSPL, we set \\(\\tau\\) to 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.788, + 0.563 + ], + "angle": 0, + "content": "Effect of Guidance To validate the effect of different types of guidance information, we conducted an ablation study that only adds one type of information. As shown in Table 5, adding each type of the four categories of guidance information contributes to performance improvement. As shown in Line 1, adding location information contributes mostly, which shows that letting the agent know its own location and target location is of great help. As shown in Line 4, adding previous observations information, the agent's SPL increased significantly \\((2.68\\uparrow)\\), but the SR increased slightly \\((0.96\\uparrow)\\). This may be attributed to the agent's ability to reference historical trajectories, thereby gaining a better understanding of the navigation environment. This, in turn, helps to avoid redundant observations and improve navigation efficiency. Destination Description and Previous Observations reduce TL from 21.96 to 19.21 and 19.22, which shows such two guidance can help the agent reach the destination faster." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.788, + 0.648 + ], + "angle": 0, + "content": "Trends of Confusion Score We plot the agent's cumulative confusion score at each time step every 5k training iterations. As shown in Figure 5, the trend of confusion score is decreasing, especially in the model using VLN-Copilot, the curve is steeper, illustrating that our model could reduce the agent's confusion degree in navigation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.731 + ], + "angle": 0, + "content": "Learning Curves We compared the learning curves between methods with and without VLN-Copilot during training. As shown in Figure 6, our VLN-Copilot converges faster than the model without Copilot and can achieve better performance (i.e. higher SPL value) in fewer iterations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.754, + 0.391, + 0.769 + ], + "angle": 0, + "content": "4.4 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We present some examples to see the qualitative results of the LLM-generated guidance. As shown in Figure 7, for the example on the left, LLM selects \"Destination Description\" as the guidance type. Since the guidance information contains a description of the destination and matches candidate direction 2, it is easier to select the answer. For" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Y. Qiao et al." + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.145, + 0.498, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.145, + 0.768, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.39, + 0.653, + 0.404 + ], + "angle": 0, + "content": "Fig. 7: Examples of the LLM-generated Guidance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.525 + ], + "angle": 0, + "content": "the example on the right, since Candidate 2 and Candidate 3 are both in the kitchen, and the task is to go to the kitchen, it looks like each candidate is similar, making the choice difficult. While LLM predicts that the guidance type is \"location\" and provides the agent's current location and destination, we find that the agent has reached the target location and therefore should stop rather than select a direction to go. More results can be found in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.546, + 0.499, + 0.562 + ], + "angle": 0, + "content": "5 Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.576, + 0.789, + 0.668 + ], + "angle": 0, + "content": "In this work, we utilize Large Language Models (LLMs) to analyze the navigation environment by transforming the dynamic environment into textual descriptions. This practice may be sub-optimal for environmental perception. Our future work will focus on how to effectively apply Multimodal Large Language Models (MLLMs) such as LLaVA [23] and Video-LLaVA [21] to perceive the navigation environment and provide guidance information." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.689, + 0.345, + 0.705 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In this work, we propose a method named VLN-Copilot that enables the agent to proactively ask for help and lets the LLM act as a co-pilot to assist the agent in completing the coarse-grained VLN task. We first introduce a confusion score calculation mechanism that allows the agent to actively query the LLM when it feels uncertain about the direction of choice. Then, we use LLM to analyze the challenges faced by the agent, and then generate appropriate guidance information for different situations that the agent is in. We conduct extensive experiments on two coarse-grained VLN tasks, including quantitative and qualitative experiments, demonstrating our method's effectiveness." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.32, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.146, + 0.383, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.786, + 0.21 + ], + "angle": 0, + "content": "We thank Mr. Xiangyu Shi for the help in supplying 2 extra NVIDIA RTX 4090 GPUs for experiments." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.234, + 0.312, + 0.25 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.786, + 0.295 + ], + "angle": 0, + "content": "1. An, D., Qi, Y., Li, Y., Huang, Y., Wang, L., Tan, T., Shao, J.: Bevbert: Multimodal map pre-training for language-guided navigation. In: ICCV. pp. 2737-2748 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.786, + 0.336 + ], + "angle": 0, + "content": "2. Anderson, P., Chang, A.X., Chaplot, D.S., Dosovitskiy, A., Gupta, S., Koltun, V., Kosecka, J., Malik, J., Mottaghi, R., Savva, M., Zamir, A.R.: On evaluation of embodied navigation agents. CoRR abs/1807.06757 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.786, + 0.379 + ], + "angle": 0, + "content": "3. Anderson, P., Wu, Q., Teney, D., Bruce, J., Johnson, M., Sünderhauf, N., Reid, I.D., Gould, S., van den Hengel, A.: Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In: CVPR. pp. 3674-3683 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.38, + 0.786, + 0.42 + ], + "angle": 0, + "content": "4. Chen, H., Suhr, A., Misra, D., Snavely, N., Artzi, Y.: TOUCHDOWN: natural language navigation and spatial reasoning in visual street environments. In: CVPR. pp. 12538-12547 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.422, + 0.786, + 0.449 + ], + "angle": 0, + "content": "5. Chen, S., Guhur, P., Schmid, C., Laptev, I.: History aware multimodal transformer for vision-and-language navigation. In: NeurIPS. pp. 5834-5847 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.45, + 0.786, + 0.477 + ], + "angle": 0, + "content": "6. Chen, S., Guhur, P.L., Tapaswi, M., Schmid, C., Laptev, I.: Learning from unlabeled 3d environments for vision-and-language navigation. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.478, + 0.786, + 0.505 + ], + "angle": 0, + "content": "7. Chen, S., Guhur, P.L., Tapaswi, M., Schmid, C., Laptev, I.: Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.506, + 0.786, + 0.533 + ], + "angle": 0, + "content": "8. Fan, Y., Chen, W., Jiang, T., Zhou, C., Zhang, Y., Wang, X.E.: Aerial vision-and-dialog navigation. In: Findings of the Association for Computational Linguistics: ACL 2023 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.534, + 0.786, + 0.56 + ], + "angle": 0, + "content": "9. Feng, W., Fu, T.J., Lu, Y., Wang, W.Y.: ULN: Towards underspecified vision-and-language navigation. In: EMNLP. pp. 6394-6412 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.562, + 0.786, + 0.589 + ], + "angle": 0, + "content": "0. Gao, C., Peng, X., Yan, M., Wang, H., Yang, L., Ren, H., Li, H., Liu, S.: Adaptive zone-aware hierarchical planner for vision-language navigation. In: CVPR. pp. 14911-14920 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.59, + 0.786, + 0.617 + ], + "angle": 0, + "content": "1. Guhur, P., Tapaswi, M., Chen, S., Laptev, I., Schmid, C.: Airbert: In-domain pretraining for vision-and-language navigation. In: ICCV. pp. 1634-1643 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.618, + 0.786, + 0.645 + ], + "angle": 0, + "content": "2. Hong, Y., Opazo, C.R., Wu, Q., Gould, S.: Sub-instruction aware vision-and-language navigation. In: Webber, B., Cohn, T., He, Y., Liu, Y. (eds.) EMNLP. pp. 3360-3376 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.646, + 0.786, + 0.673 + ], + "angle": 0, + "content": "3. Hong, Y., Wu, Q., Qi, Y., Opazo, C.R., Gould, S.: A recurrent vision-and-language BERT for navigation. In: CVPR. pp. 1643-1653 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.674, + 0.786, + 0.701 + ], + "angle": 0, + "content": "4. Jain, V., Magalhães, G., Ku, A., Vaswani, A., Ie, E., Baldridge, J.: Stay on the path: Instruction fidelity in vision-and-language navigation. pp. 1862-1872 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.786, + 0.728 + ], + "angle": 0, + "content": "5. Kaufman, L., Rousseeuw, P.J.: Finding groups in data: an introduction to cluster analysis. John Wiley & Sons (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.786, + 0.77 + ], + "angle": 0, + "content": "6. Krantz, J., Wijmans, E., Majumdar, A., Batra, D., Lee, S.: Beyond the nav-graph: Vision-and-language navigation in continuous environments. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J. (eds.) ECCV. pp. 104-120 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.786, + 0.812 + ], + "angle": 0, + "content": "7. Ku, A., Anderson, P., Patel, R., Ie, E., Baldridge, J.: Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In: EMNLP. pp. 4392-4412 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "8. Li, J., Tan, H., Bansal, M.: Envideit: Environment editing for vision-and-language navigation. In: CVPR. pp. 15386-15396 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.267, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Y. Qiao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.189 + ], + "angle": 0, + "content": "19. Li, J., Li, D., Xiong, C., Hoi, S.C.H.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML (2022), https://api.sementicscholar.org/CorpusID:246411402" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.786, + 0.218 + ], + "angle": 0, + "content": "20. Li, X., Wang, Z., Yang, J., Wang, Y., Jiang, S.: Kerm: Knowledge enhanced reasoning for vision-and-language navigation. In: CVPR. pp. 2583-2592 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.245 + ], + "angle": 0, + "content": "21. Lin, B., Zhu, B., Ye, Y., Ning, M., Jin, P., Yuan, L.: Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.786, + 0.273 + ], + "angle": 0, + "content": "22. Lin, K., Chen, P., Huang, D., Li, T.H., Tan, M., Gan, C.: Learning vision-and-language navigation from youtube videos. In: ICCV. pp. 8317-8326 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.786, + 0.299 + ], + "angle": 0, + "content": "23. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.786, + 0.328 + ], + "angle": 0, + "content": "24. Liu, R., Wang, X., Wang, W., Yang, Y.: Bird's-eye-view scene graph for vision-language navigation. In: ICCV. pp. 10968-10980 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.356 + ], + "angle": 0, + "content": "25. Liu, S., Zhang, H., Qi, Y., Wang, P., Zhang, Y., Wu, Q.: Aerialvln: Vision-and-language navigation for uavs. In: ICCV. pp. 15384-15394 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.786, + 0.383 + ], + "angle": 0, + "content": "26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR. OpenReview.net (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.786, + 0.425 + ], + "angle": 0, + "content": "27. Nguyen, K., III, H.D.: Help, anna! visual navigation with natural multimodal assistance via retrospective curiosity-encouraging imitation learning. In: Inui, K., Jiang, J., Ng, V., Wan, X. (eds.) EMNLP. pp. 684–695. Association for Computational Linguistics (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.786, + 0.467 + ], + "angle": 0, + "content": "28. Qi, Y., Pan, Z., Hong, Y., Yang, M., van den Hengel, A., Wu, Q.: The road to know-where: An object-and-room informed sequential bert for indoor vision-language navigation. In: ICCV. pp. 1655–1664 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.786, + 0.508 + ], + "angle": 0, + "content": "29. Qi, Y., Wu, Q., Anderson, P., Wang, X., Wang, W.Y., Shen, C., van den Hengel, A.: REVERIE: remote embodied visual referring expression in real indoor environments. In: CVPR. pp. 9979-9988 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.786, + 0.536 + ], + "angle": 0, + "content": "30. Qiao, Y., Qi, Y., Hong, Y., Yu, Z., Wang, P., Wu, Q.: Hop+: History-enhanced and order-aware pre-training for vision-and-language navigation. IEEE TPAMI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.786, + 0.564 + ], + "angle": 0, + "content": "31. Qiao, Y., Qi, Y., Yu, Z., Liu, J., Wu, Q.: March in chat: Interactive prompting for remote embodied referring expression. In: ICCV. pp. 15758-15767 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.786, + 0.592 + ], + "angle": 0, + "content": "32. Qiao, Y., Yu, Z., Wu, Q.: Vln-petl: Parameter-efficient transfer learning for vision-and-language navigation. In: ICCV. pp. 15443-15452 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.786, + 0.632 + ], + "angle": 0, + "content": "33. Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert-networks. In: Inui, K., Jiang, J., Ng, V., Wan, X. (eds.) EMNLP-IJCNLP. pp. 3980-3990 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.632, + 0.786, + 0.66 + ], + "angle": 0, + "content": "34. Roman, H.R., Bisk, Y., Thomason, J., Celikyilmaz, A., Gao, J.: Rmm: A recursive mental model for dialog navigation. arXiv preprint arXiv:2005.00728 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.786, + 0.688 + ], + "angle": 0, + "content": "35. Schumann, R., Zhu, W., Feng, W., Fu, T.J., Riezler, S., Wang, W.Y.: Velma: Verbalization embodiment of llm agents for vision and language navigation in street view. In: AAAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.786, + 0.729 + ], + "angle": 0, + "content": "36. Song, C.H., Wu, J., Washington, C., Sadler, B.M., Chao, W.L., Su, Y.: Llm-planner: Few-shot grounded planning for embodied agents with large language models. arXiv preprint arXiv:2212.04088 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.786, + 0.757 + ], + "angle": 0, + "content": "37. Thomason, J., Murray, M., Cakmak, M., Zettlemoyer, L.: Vision-and-dialog navigation. In: CoRL. pp. 394-406 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.786, + 0.84 + ], + "angle": 0, + "content": "38. Touvron, H., Martin, L., Stone, K.R., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D.M., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A.S., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I.M., Korenev, A.V., Koura, P.S., Lachaux, M.A., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A.," + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.32, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.216 + ], + "angle": 0, + "content": "Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. ArXiv abs/2307.09288 (2023), https://api.sementicscholar.org/CorpusID:259950998" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.787, + 0.245 + ], + "angle": 0, + "content": "39. Wang, X., Wang, W., Shao, J., Yang, Y.: Lana: A language-capable navigator for instruction following and generation. In: CVPR. pp. 19048-19058 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.286 + ], + "angle": 0, + "content": "40. Wang, X., Huang, Q., Celikyilmaz, A., Gao, J., Shen, D., Wang, Y., Wang, W.Y., Zhang, L.: Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In: CVPR. pp. 6629-6638 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.314 + ], + "angle": 0, + "content": "41. Wang, Z., Li, X., Yang, J., Liu, Y., Jiang, S.: Gridmm: Grid memory map for vision-and-language navigation. In: ICCV. pp. 15625-15636 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.342 + ], + "angle": 0, + "content": "42. Wang, Z., Li, J., Hong, Y., Wang, Y., Wu, Q., Bansal, M., Gould, S., Tan, H., Qiao, Y.: Scaling data generation in vision-and-language navigation. In: ICCV. pp. 12009-12020 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.382 + ], + "angle": 0, + "content": "43. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.787, + 0.411 + ], + "angle": 0, + "content": "44. Zhu, F., Liang, X., Zhu, Y., Yu, Q., Chang, X., Liang, X.: SOON: scenario oriented object navigation with graph-based exploration. In: CVPR. pp. 12689-12699 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.439 + ], + "angle": 0, + "content": "45. Zhu*, Y., Weng*, Y., Zhu, F., Liang, X., Ye, Q., Lu, Y., jiao, J.: Self-motivated communication agent for real-world vision-dialog navigation. In: ICCV (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.439 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/LLM as Copilot for Coarse-grained Vision-and-Language Navigation/5a4eb16e-327f-48aa-9238-afeb5ed269ac_origin.pdf b/2024/LLM as Copilot for Coarse-grained Vision-and-Language Navigation/5a4eb16e-327f-48aa-9238-afeb5ed269ac_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f1ad7334440150bce151cfb46302a498e96e2159 --- /dev/null +++ b/2024/LLM as Copilot for Coarse-grained Vision-and-Language Navigation/5a4eb16e-327f-48aa-9238-afeb5ed269ac_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e97d02fbbe4d33e4fc280a6df713ca056ec62d8a416ac5935d0a6d564d46ce93 +size 7648223 diff --git a/2024/LLM as Copilot for Coarse-grained Vision-and-Language Navigation/full.md b/2024/LLM as Copilot for Coarse-grained Vision-and-Language Navigation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5cebcab1b58fc6d7267dc65d0d9235f7682cbc0e --- /dev/null +++ b/2024/LLM as Copilot for Coarse-grained Vision-and-Language Navigation/full.md @@ -0,0 +1,342 @@ +# LLM as Copilot for Coarse-grained Vision-and-Language Navigation + +Yanyuan Qiao $^{1}$ , Qianyi Liu $^{2,3}$ , Jiajun Liu $^{4,5}$ , Jing Liu $^{2,3}$ , and Qi Wu $^{1*}$ + +$^{1}$ Australian Institute for Machine Learning, The University of Adelaide + +$^{2}$ Institute of Automation, Chinese Academy of Sciences + +3 School of Artificial Intelligence, University of Chinese Academy of Sciences + +4 CSIRO Data61 + +5 The University of Queensland + +{yanyuan.qiao,qi.wu01}@adelaide.edu.au, liuqianyi2022@ia.ac.cn, + +jiajun.liu@csiro.au, jliu@nlpr.ia.ac.cn + +Abstract. Vision-and-Language Navigation (VLN) involves guiding an agent through indoor environments using human-provided textual instructions. Coarse-grained VLN, with short and high-level instructions, has gained popularity as it closely mirrors real-world scenarios. However, a significant challenge is these instructions are often too concise for agents to comprehend and act upon. Previous studies have explored allowing agents to seek assistance during navigation, but typically offer rigid support from pre-existing datasets or simulators. The advent of Large Language Models (LLMs) presents a novel avenue for aiding VLN agents. This paper introduces VLN-Copilot, a framework enabling agents to actively seek assistance when encountering confusion, with the LLM serving as a copilot to facilitate navigation. Our approach includes the introduction of a confusion score, quantifying the level of uncertainty in an agent's action decisions, while the LLM offers real-time detailed guidance for navigation. Experimental results on two coarse-grained VLN datasets show the efficacy of our method. + +Keywords: Vision-and-Language $\cdot$ Navigation $\cdot$ Large Language Models + +# 1 Introduction + +The Vision-and-Language Navigation (VLN) tasks have attracted great attention from researchers, which enables an agent to navigate to a specified destination indoors based on the provided textual instructions and visual observations. The tasks of VLN show great potential in real-world applications such as home assistant robots. Most VLN tasks such as R2R [3] and RxR [17] etc. provide fine-grained step-by-step instructions, such as "Go out the door in front of you across the room. Once in the hallway turn Left. Walk forward and then turn left into the sitting area. Stop in right in front of the fireplace." However, though these detailed instructions could improve the agents' ability of instructions following, they impede the further application of VLN in the real world since human beings would not like to give such long instructions to a robot in daily life. Contrastive to these tasks, the coarse-grained instruction task represented by the Remote + +![](images/e55f9e2f288adaacf22334edf36d690e10d38829f245995fcb2b513033a9042e.jpg) +Fig. 1: Our proposed VLN-Copilot. When an agent is confused about making an action decision, it will proactively seek help from the LLM, and then the LLM will assist the agent by analyzing the situations faced by the agent and providing appropriate guidance. + +Embodied Referring Expression (REVERIE) task [29] is more expected to enhance the practical utility of VLN because it involves instructions that are closer to those encountered in real-world scenarios, such as "Go clean the mirror in the bathroom." Giving such coarse-grained instructions to the agent would bring more challenges to the agent's navigation than fine-grained instructions. + +VLN Agents navigating in environments can benefit from seeking external assistance [34]. Early works [27, 37] tried to let agents seek help from oracles or simulators, but the assistance content obtained was fixed and limited. Since this assistance information is based on pre-built ground truth information in datasets or simulators, which are not consistent with real-life scenarios. Inspired by the development of Large Language Models (LLMs) with rich internalized world knowledge, many works have been made to exploit their capabilities to aid navigation tasks. One recent method named MiC [31] makes the first step to use LLM in the coarse-grained VLN task of REVERIE, which queries the LLM to generate more detailed instructions by prompting the LLM with extra planning examples. The promising performance of MiC shows the great potential to use LLM in coarse-grained VLN tasks. However, MiC has some weaknesses in applying LLM for VLN. First, the way the agent seeks help from the LLM is a passive way that depends on the change of the room where the agent is located. This approach may not help the agent when it faces difficulties in making predictions while still in the same room. Second, MiC generates fine-grained planning no matter what challenges the agent faces. This practice lacks a thorough analysis of the difficulties the agent may have and fails to give the agent diverse guidance for different scenarios. + +To address these issues, we propose a novel method named VLN-Copilot that makes the LLM a competent copilot to assist the agent in navigation. As shown in Figure 1, before making a movement, the agent calculates a confusion score to assess the level of + +uncertainty associated with its current action choice. If this score surpasses a predefined threshold, it triggers the agent to seek guidance. This practice is more proactive than relying on the change of scenes which may be wrong. When the agent decides to seek assistance, it consults the LLM and then the LLM conducts a comprehensive evaluation of the situation, identifies the type of guidance required, and generates the necessary guidance to aid the agent in navigating effectively. The types of guidance information are selected from the in-depth pre-analysis of prevailing challenges the agent may face in the sampled environment using LLM. + +We conduct experiments on the REVERIE and CVDN-target dataset and our model surpasses the previous methods. Specifically, our VLN-Copilot obtains $42.32\%$ on the main navigation metric SPL and $26.55\%$ on the main object grounding metric RGSPL on the REVERIE dataset, and 4.47 in GP on the CVDN-target dataset, respectively. + +In summary, our contributions are as follows: + +- We propose a novel method named VLN-Copilot which utilizes LLM to analyze the difficulties the agent may meet and give diverse guidance for different situations, rather than to generate planning by prompting LLM with extra planning examples. +- We introduce the confusion score to help the agent proactively ask the LLM for help when confused to make predictions rather than passively trigger conversations. +- We utilize LLM with elaborately designed prompts to conduct a detailed challenge analysis of the agent's different situations, enabling the LLM to provide better guidance on the fly. +- Extensive experiments on two VLN benchmarks show the effectiveness of our proposed method. + +# 2 Related Work + +# 2.1 Vision-and-Language Navigation + +In recent years, Vision-and-Language Navigation (VLN) has attracted widespread attention due to its promising applications in home assistant robots and related fields [9, 12, 18, 20, 28, 40]. Many different tasks have emerged in the field of VLN [4, 14, 16, 25, 27, 29, 37]. In terms of the navigation space accessible to the agent, these cover indoor [3, 16, 17], outdoor [4, 35], and aerial scenarios [8, 25]. + +Early VLN task R2R [3] started from indoor environments requiring the agent to reach the destination according to the given instructions, and there were some subsequent tasks (e.g., RxR [17] and VLN-CE [16]). These tasks all provide detailed natural language instructions, such as "Walk down the stairs at the bottom of the stairs take a left. Walk straight to get to the kitchen inside and wait just inside the door." These detailed instructions guide the agent through complex step-by-step processes and are more focused on testing the agent's ability to understand and follow instructions. Some other tasks adopt an object-oriented perspective, such as REVERIE [29] and SOON [44]. These tasks require not only reaching a designated destination but also discovering specific target objects. Different from SOON providing detailed instructions and object recognition prompts, such as "The cabinet is square, closed and wooden. The cabinet is placed below the sink and the water-tap. There are some stools and table nearby." + +REVERIE's instructions appear more concise and natural, such as "Bring me the swivel chair from the office." These coarse-grained instructions, akin to instructions encountered in real-world scenarios, enhance the potential for innovative research within the domain of coarse-grained instruction VLN. However, the task remains challenging due to the limited information these instructions provide. + +# 2.2 VLN Agents Asking for Help + +Agents can gain advantages by seeking external assistance to assist navigation, and there have been studies promoting navigation agents to seek help during navigation tasks [27, 34]. Thomason et al. [37] introduce the Navigation from Dialog History (NDH) task, which contains dialog history between agents and oracles during navigation. The agent follows guide instructions and asks questions to the oracle when needed. The task provides a valuable resource for investigating the navigation interactions between agent and oracle. Similarly, HANNA [27] builds a simulator where navigation agents can request language-and-vision assistance when they face difficulties during navigation. However, the agent's signals asking for help are fixed and predefined in the simulator. Additionally, the assistance provided is based on ground truth information derived from the R2R dataset, which may not fully represent real-world navigation scenarios due to its reliance on pre-built datasets. Similarly, Zhu et al. proposed SCoA [45], which enables navigation agents to ask questions from a set of pre-constructed candidate question-answer pairs to aid in navigation. Recently, LLM-Planner [36] and MiC [31] have attempted to leverage the knowledge of large language models (LLM) to assist agents in navigation. Agents can query the LLM to generate detailed step-by-step instructions to aid navigation. However, the ways the agents ask for help from both methods are passive, either based on fixed time steps or when the room the agent is in changes. This practice may miss the opportunity to help the agent when it has difficulties in prediction actions. Besides, these two methods concentrate on generating planning by prompting LLM with extra planning data. Different from these methods, we propose a method named VLN-Copilot that allows agents to proactively seek help when they are not sure where to go in this work. Moreover, our VLN-Copilot pays more attention to analyzing the difficulties the agent may meet and can dynamically give appropriate types of guidance in diverse situations for the agent. + +# 3 Method + +# 3.1 Preliminaries + +The main goal for the agent in the VLN task is to navigate to a specific destination following the high-level concise instructions. Specifically, the agent is asked to navigate in the simulated environment, which can be described as an undirected graph $\mathcal{G} = \{\mathcal{V},\mathcal{E}\}$ , where $\mathcal{V}$ represents the navigable nodes and $\mathcal{E}$ represents the connectivity edges. Given the natural language instruction $L$ , the agent starts from the initial node $V_{0}$ and perceives a panoramic view $\mathcal{R}_t$ of the current node $V_{t}$ to make the action $a_{t}$ at each time step $t$ . The panoramic view $\mathcal{R}_t$ has $K$ single-view images $r_i$ and can be represented as + +![](images/48d79f507645ff044e946dc24f4b66da236e91bc512b62ddd0f1eca1bbfe105d.jpg) +Fig.2: Overview of the VLN-Copilot. At each time step, the agent predicts the logits over the candidate's view of actions. After that, the agent will calculate the confusion score to decide whether to ask LLM for help (Sec. 3.3). When the confusion score surpasses threshold $\tau$ , the agent queries the LLM for help with the prompt that helps LLM perceive the environment (Sec. 3.4). Then the LLM generates useful guidance information, which will be returned to the agent to assist in decision-making (Sec. 3.6). + +$\mathcal{R}_t = \{r_i\}_{i=1}^K$ , of which $r_i$ is represented by an image feature vector with the corresponding orientation feature vector. Besides, for the sub-task of object grounding, the object features $\mathcal{O}_t = \{o_i\}_{i=1}^M$ are extracted from the panoramic view $\mathcal{R}_t$ with the annotated object bounding boxes or by object detectors. With these observations of the environment, the agent takes a series of actions $\langle a_0, \dots, a_N \rangle$ to navigate to the target location, where each action $a_t$ selects one navigable node from the candidates. When the target object is grounded or the pre-defined maximum navigation steps have been achieved, the navigation ends. + +# 3.2 Overview + +As is shown in Figure 2 and Algorithm 1, given the coarse-grained instruction, the agent first perceives the environment to obtain the current visual observation and starts to predict a probability over candidate actions. Then, a confusion score is computed according to the probability distribution, which determines whether the agent is confused about making predictions in the current situation and whether the agent should ask the LLM for help. If the confusion score exceeds a predefined threshold $\tau$ (obtained via grid search) which means the agent is confused about making decisions, the agent initiates a conversation with the LLM to seek their help with an elaborately designed prompt. Specifically, the prompt contains coarse-grained instruction, fine-grained scene descriptions of candidate views, historical descriptions of trajectory scenes, and the requirements for the LLM. Then, the LLM analyzes the agent's situation and gives advice about which type of guidance information is useful as well as the details about that guidance information. The types of guidance information are refined from the results of challenge pre-analysis by LLM. Finally, the response about the guidance from the LLM is appended after the coarse-grained instruction to guide the agent's action prediction collaboratively. This paradigm aims to optimize agents' decision-making processes by + +combining instruction, observation, and LLM-based world knowledge via the conversations between the agent and LLM. + +```txt +Algorithm 1 VLN-Copilot +Notation Summary: I: high-level instruction $\mathcal{R}_t$ : visual observation at timestep t $S_{t}$ : confusion score at timestep t $W_{t}^{F}$ : fine-grained scene descriptions of all candidate views at timestep t $W_{t}^{\mathrm{Tr}}$ : historical trajectory description at timestep t $T_{\mathrm{guide}}$ : prompt template for LLM to generate guidance information LLM: large language model $\tau$ : pre-defined threshold for confusion score + $t\gets 0$ $\rightharpoondown$ Initial timestep while $t < \max$ step and $\hat{a}_{t}\neq$ "stop" do if $S_{t} > = \tau$ then $P_{\mathrm{guide}}\leftarrow T_{\mathrm{guide}}(I,W_t^F,W_t^{\mathrm{Tr}})$ $R_{\mathrm{guide}}\gets \mathrm{LLM}(P_{\mathrm{guide}})$ $I_{t}^{\mathrm{Inter}}\gets \mathrm{CONCAT}(I,R_{\mathrm{guide}})$ $\hat{a}_{t}\gets \mathrm{Agent}(I_{t}^{\mathrm{Inter}},\mathcal{R}_{t})$ else $\hat{a}_{t}\gets \mathrm{Agent}(I,\mathcal{R}_{t})$ end if $t\gets t + 1$ +end while +``` + +# 3.3 When to Ask LLM for Help + +Different from previous methods [31] that set fixed timestep for interactions with LLM or trigger the interactions when the room scenes change, we select a more active practice for the agent by using the confusion score to decide whether to ask for help. Specifically, at each time step $t$ , when the agent is going to predict the logits over the candidate nodes for the next action $a_{t}$ , we calculate the entropy of the predicted probability distribution over each candidate navigable node $v_{i}$ as the confusion score $S_{t}$ : + +$$ +p _ {t} \left(v _ {i}\right) = \operatorname {S o f t m a x} \left(s _ {t} \left(v _ {i}\right)\right), \tag {1} +$$ + +$$ +\mathbf {S} _ {t} = - \sum_ {i} p _ {t} \left(v _ {i}\right) \log_ {2} \left(p _ {t} \left(v _ {i}\right) + \epsilon\right), \tag {2} +$$ + +where $s_t(v_i)$ is the predicted logits assigned by the agent to the $i$ -th navigable node $v_i$ at time step $t$ , and $\epsilon$ is a minimal number added to avoid numerical instability. + +Then, if the computed confusion score $S_{t}$ surpasses a predefined threshold $\tau$ , which means the agent is confused about making a decision among some candidate nodes, it will send a message to LLM asking for help. + +# 3.4 How LLM Perceives the Environment + +LLMs would give inaccurate advice for the agent's navigation since they are not able to directly acquire visual content from the environment. Both current visual observations from the agent and previously observed scenes are essential for the LLM to analyze the + +navigation and supply important information. Thus, enhancing the LLM's environmental awareness as far as possible is a critical first step to making it a competent copilot. To solve this problem, we translate the dynamic observations and trajectories into multi-grained spatiotemporal descriptions as the eye for the LLM, using the out-of-the-box vision-and-language model of BLIP [19] model. Detailedly, the visually-grounded description mainly consists of three parts: the coarse-grained scene description of the current panoramic view $\mathcal{R}_t$ , the fine-grained scene description of each view $r_i$ of selected candidate navigable nodes at the current time step $t$ , and the trajectory description of previously observed scenes before the current time step $t$ . + +Specifically, we obtain both the panoramic view image $\mathcal{R}_t$ of the agent's observation and the single view image $r_i$ of each candidate navigable node at the time step $t$ . We first feed $\mathcal{R}_t$ to the BLIP model to generate the coarse-grained scene description $W_t^C$ for the agent's current location. For the fine-grained scene description, the agent is provided with $k$ candidate views for selection. Then, we feed each single-view image $r_i$ of the $k$ selected candidate navigable nodes to BLIP to generate the fine-grained scene description $W_t^{F_i}$ : + +$$ +W _ {t} ^ {\mathrm {C}} = \operatorname {B L I P} \left(\mathcal {R} _ {t}\right), \tag {3} +$$ + +$$ +W _ {t} ^ {F _ {i}} = \operatorname {B L I P} \left(r _ {i}\right), \tag {4} +$$ + +Then we concatenate all previous coarse-grained scene descriptions $W^{\mathrm{C}}$ of each time step before the current time $t$ as the historical trajectory description $W_{t}^{\mathrm{Tr}}$ such as "I passed a bedroom, and then I passed ..." This process can be formulated as follows: + +$$ +W _ {t} ^ {\mathrm {T r}} = \operatorname {C O N C A T} \left(W _ {1} ^ {\mathrm {C}}, \dots , W _ {t - 1} ^ {\mathrm {C}}\right), \tag {5} +$$ + +Note that we only generate the fine-grained scene descriptions and trajectory descriptions when the agent is confused to make the action prediction and asks the LLM for help, while we generate coarse-grained scene descriptions at each time step. In addition, for better comprehension of LLM, we added the candidate view id $i$ as the prefix to each corresponding native fine-grained scene description $W_{t}^{F_{i}}$ . For example, we converted the generated scene description of "A room with a tilted ceiling, a bathroom mirror on the wall, and a skylight above the window." into the candidate view description with the following format: "I see candidate 0 is a room with a tilted ceiling ..." + +# 3.5 Difficulty and Useful Guidance-Type Pre-Analysis + +Before asking the agent to query an LLM to generate guidance for assistance, it would be beneficial to explore the difficulties the agent may encounter in making decisions and what guidance information the LLM should provide to be useful for the agent's navigation. However, manual analysis by humans is costly and impractical. Thus, we turn to LLM for analysis. Specifically, we first collect some navigation samples. When the agent finds it difficult to make decisions during the navigation process, we convert the situation the agent encountered into a text description and save it as a navigation sample. Then, we use an elaborate prompt to query the LLM to analyze the situation encountered by the agent in the current navigation example, such as what is the reason + +System: You are a helpful assistant for navigation. There is an agent navigating indoors according to task instructions and current observation. At each step, the agent needs to choose one of the candidate directions to go. I will tell you which direction is correct. + +You need to first analyze the current situation, such as what difficulties the agent encountered during navigation that made it difficult to choose a direction and what types of guidance information may be useful. + +Task instruction: Turn off the white table lamp in the bedroom + +Observation: + +I see candidate 1 is a white door leads to a bedroom with a bed, electric fan and table + +I see candidate 2 is a bathroom has a sink and the door leads to the shower room + +I see candidate 3 is a hallway with white walls, stairs, and a door on the second floor + +History trajectory: I have passed a balcony with rocking chairs and a table, then I go to a hallway with a picture of a bedroom. + +Correct answer: candidate 1 + +Assistant: + +Difficulty: The agent encountered difficulties in choosing a direction because there are multiple doors at the same time. This makes it challenging for + +the agent to determine which direction could find the lamp. + +Please list possible type of guidance may be useful. + +Assistant: + +Guidance Type: The agent needs to the destination information to make an informed decision. + +Fig. 3: Prompt template for LLM to pre-analyze the difficulties the agent encounters and what type of guidance information maybe useful. The color of red, blue represent multi-turn responses generated from LLM. + +for the difficulty in making a decision, and the helpful types of guidance. Finally, we collect the guidance information generated by LLM and cluster it to obtain the final guidance categories. + +Navigation Samples Collection We first let the agent navigate around the environment with confusion scores computed. If the confusion score surpasses the pre-defined threshold at the time step $t$ , which means the agent comes across difficulties in making the action prediction, then we collect the tuple $(I, W_{t}^{F}, W_{t}^{Tr}, A_{t})$ of the high-level task instruction $I$ , fine-grained candidate scene descriptions $W_{t}^{F}$ , the trajectory description $W_{t}^{Tr}$ and the ground-truth action $A_{t}$ as a sample data $D_{\text{sample}}$ for analysis. Note that $W_{t}^{F}$ includes $k$ fine-grained scene descriptions of the selected candidate views. In total, we randomly collect 5,000 samples. + +Propmt Design To make the LLM better analyze the situation that the agent meets, we elaborately designed a prompt template for multi-turn query-and-response. We first set the unified system prompt $P_{\text{system}}$ that requires the LLM as a helpful assistant for navigation. Then, we transform each sample data $D_{\text{sample}}$ into sample prompt $P_{\text{sample}}$ by respectively adding the prefix of "Task instruction:", "Observation:", "History trajectory:" and "Correct answer:" into $I, W_t^F, W_t^{Tr}$ and $A_t$ . Next, we decompose the task of analyzing the challenge into two sub-tasks, the first of analyzing what difficulties the agent may meet in making decisions and the second of analyzing what types of guidance may be useful for navigation in an iterative way. Correspondingly, we use "########Assistant: Difficulty:" and "########Assistant: Guidance Type:" as the sub-task prompts $P_{\text{diff}}$ and $P_{\text{guide}}$ . We concatenate the prompts of $P_{\text{system}}$ , $P_{\text{sample}}$ and $P_{\text{diff}}$ as $P_{\text{task1}}$ and feed $P_{\text{task1}}$ to LLM, which will generate the response $R_{\text{diff}}$ about the difficulty the agent may meet. We further feed $P_{\text{task2}}$ to LLM, the concatenation + +System: You are a helpful assistant for navigation. There is an agent navigating indoors according to task instructions and current observation. At each step, the agent needs to choose one of the candidate directions to go. + +You need to analyze what additional guidance information the agent need to make the choice. There are four types of information: location, prioritization, destination description, previous observations. You can only give one answer from these types. + +Task instruction: Turn off the white table lamp in the bedroom + +Observation: + +I see candidate 1 is a white door leads to a bedroom with a bed, electric fan and table + +I see candidate 2 is a bathroom has a sink and the door leads to the shower room + +I see candidate 3 is a hallway with white walls, stairs, and a door on the second floor + +History trajectory: I have passed a balcony with rocking chairs and a table, then I go to a hallway with a picture of a bedroom. + +####Assistant: + +Guidance Type: Destination description + +Please give description of the destination. + +Assistant: + +The bedroom is a cozy and comfortable space, typically with a queen-sized bed in the center of the room. The bed is dressed with crisp, white linens and fluffy pillows. + +Fig. 4: Prompt template for LLM to generate guidance information. + +of $P_{task1}$ , $R_{diff}$ and $P_{guide}$ and we will get the response $R_{guide}$ about what types of guidance may be useful for navigation. The prompt template is depicted in Figure 3. + +Guidance-Type Clustering After the above-mentioned operations, we obtain a collection of the guidance response $R_{guide}$ . We first use Sentence-BERT [33] to embed each $R_{guide}$ and get the corresponding embedding $E_{guide}$ . Then we employ the silhouette coefficient [15] to determine the optimal number of clusters for K-Means clustering. The silhouette coefficient is an indicator to evaluate the quality of clustering. It combines the two factors of clustering cohesion and separation, calculates a silhouette coefficient value for each sample, and then calculates a silhouette coefficient value through all samples. By calculating the average silhouette coefficient under different cluster numbers $k$ , the clustering quality under each cluster number can be evaluated. In general, the number of clusters with the highest average silhouette coefficient is considered the optimal number of clusters because it provides the best balance of cohesion and separation. And clustering result shows that the best number of clusters is seven. In addition, after analysis based on the practical situation, we found that some of the guidance content is similar, and some guidance is not executable, such as "Room Layout", which cannot be obtained from the environment. Thus, we manually merge these duplicate types, remove infeasible types, and finally reserve four applicable categories: + +- Location: information about the agent's current position and the destination. +- Prioritization: information about the urgent task that needs to be completed based on an assessment of agent progress. +- Destination Description: detailed information about the target location, providing more cues to reach its destination. +- Previous Observations: information about the agent's navigation trajectory, lets the agent remember history information. + +# 3.6 LLM Guides the Agent On-the-Fly + +After the pre-analysis of the agent's navigation, we employ the analysis result and a new prompt to ask LLM to aid the agent on the fly when it is confused to make decisions. We first illustrate the modification of the prompt for the on-the-fly guidance. Then, we show how LLM guides the agent with different types of guidance information. + +Prompt Design We also designed the prompt template to ask the LLM to analyze the agent's current situation and decide which guidance type from the selected four categories might be useful in that situation. This prompt template is similar to the pre-analysis prompt template, but with some modifications to guide agent navigation. The full prompt template is shown in Figure 4. First, we modify the system prompt $P_{system}$ by asking the LLM to choose which guidance information is useful from the four pre-selected categories instead of analyzing the difficulties and generating unlimited guidance information. Second, we delete the ground truth action in the sample prompt $P_{sample}$ . At last, the task prompt is revised to "########Assistant: Guidance Type:". + +Integrating Guidance with Instruction During the navigation of the agent, when the confusion score $S_{t}$ surpasses the pre-defined threshold $\tau$ at time step $t$ , it will ask the LLM for guidance. With the help of the above-mentioned prompts, the LLM can perceive the environment and trajectory of the agent, analyze the situation, and select the guidance type from the pre-selected four categories. For different types of guidance, the LLM will generate different guidance information $I_{t}^{G}$ to assist the navigation. Specifically, the LLM will generate detailed content about the type of "Location", "Prioritization" and "Destination Description" according to its internalized world knowledge. For these three types of guidance, the generation process is also interactive, which first generates the type of information and then generates the details according to the type of information. Such a two-step process ensures the quality of generated guidance. For the type of "Previous Observations", we re-use the historical trajectory as the guidance information. Some examples of generated guidance information can be found in Figure 7. Finally, we append the guidance information $I_{t}^{G}$ to the high-level instruction $I$ , and then the agent will navigate under the guidance of the integrated instruction $I_{t}^{Inter}$ . + +# 4 Experiment + +# 4.1 Settings + +Datasets We evaluate our method on two VLN tasks: REVERIE and CVDN-target, which are all based upon the MP3D indoor environments. Compared to Room-to-Room (R2R) task, these instructions are high-level and concise, such as "Bring me the swivel chair from the office". CVDN-target is our proposed VLN setup similar to REVERIE. The instruction contains a target and conversation history. Here we remove the conversation history and only keep the information containing the target as the instruction, such as "To find a book in the room." + +Table 1: Comparison with the state-of-the-art methods on REVERIE. + +
MethodsVal UnseenTest Unseen
NavigationGroundingNavigationGrounding
TLOSR↑SR↑SPL↑RGS↑RGSPL↑TLOSR↑SR↑SPL↑RGS↑RGSPL↑
Human------21.1886.8381.5153.6677.8451.44
RecBERT [13]16.7835.0230.6724.9018.7715.2715.8632.9129.6123.9916.5013.51
Airbert [11]18.7134.5127.8921.8818.2314.1817.9134.2030.2823.6116.8313.28
HAMT [5]14.0836.8432.9530.2018.9217.2813.6233.4130.4026.6714.8813.08
VLN-PETL [32]14.4737.0331.8127.6718.2615.9614.0036.0630.8326.7315.1313.03
LANA [39]16.2838.5434.0029.2619.0316.1816.7536.4133.5026.8917.5314.25
HOP+ [30]14.5740.0436.0731.1322.4919.3315.1735.8133.8228.2420.2016.86
DUET [7]22.1151.0746.9833.7332.1523.0321.3056.9152.5136.0631.8822.06
Lily [22]21.8753.7148.1134.4332.1523.4321.9460.5154.3237.3432.0221.94
BEVBert [1]-56.4051.7836.3734.7124.44-57.2652.8136.4132.0622.09
AutoVLN [6]-62.1455.8940.8536.5826.76-62.3055.1738.8832.2322.68
AZHP [10]22.3253.6548.3136.6334.0025.79-55.3151.5735.8532.2522.44
BSG [24]24.7158.0552.1235.5935.3624.2422.9062.8356.4538.7033.1522.34
KERM [20]21.8555.2150.4435.3834.5124.4517.3257.5852.4339.2132.3923.64
ScaleVLN [42]--56.9741.8435.7626.05--56.1339.5232.5322.78
GRidMM [41]23.2057.4851.3736.4734.5724.5619.9759.5553.1336.6034.8723.45
MiC [31]20.6462.3756.9743.6037.5228.7218.1162.4055.7441.9735.2526.17
VLN-Copilot21.8962.6257.4043.6338.8829.7521.7263.2657.8142.3236.5626.55
+ +Evaluation Metrics We utilize the standard evaluation metrics to evaluate our method, we report Success Rate (SR), Success weighted by Path Length (SPL) [2], and Goal Progress (GP) as the main navigation metrics, Remote Grounding Success rate (RGS) and RGS weighted by Path Length (RGSPL) as the main object grounding metrics. + +Implementation Details We implement our proposed method using PyTorch. The model is trained on one single NVIDIA RTX 4090 GPU. For REVERIE, we fine-tuned the pre-trained model from [6] with a batch size of 4 for 100k iterations. For the CVDN-target, we fine-tuned the pre-trained model from [7] with a batch size of 4 for 50k iterations. AdamW [26] optimizer is adopted and the learning rate is set to $1 \times 10^{-5}$ . The threshold $\tau$ is set as 1. For the baseline agent, we use the recent state-of-the-art model of AutoVLN [6]. When collecting data for pre-analysis, we use two agents based on pre-trained AutoVLN [6] and HAMT [5] to avoid agent bias. LLM is integrated into the training process, we use the open-source Vicuna-7B-1.5 model [43], which is fine-tuned from Llama 2 [38]. + +# 4.2 Comparison with State-of-the-Art Methods + +REVERIE As shown in Table 1, our method achieves competitive performance on the REVERIE benchmark compared to previous methods, such as AutoVLN [6], Mic [31] and ScaleVLN [42]. Particularly, compared with the SoTA method MiC [31], our method outperforms Mic in terms of the object grounding metric RGSPL with $1.36\%$ and RGSPL with $1.03\%$ on the Val Unseen split. It is worth noting that compared with + +Table 2: Comparison on CVDN-target Validation Unseen. + +
MethodsTLGP↑
HAMT† [5]37.042.74
DUET† [7]78.533.70
AutoVLN† [6]61.973.98
VLN-Copilot40.384.47
+ +Table 4: Ablation of different threshold $\tau$ . + +
τ0.811.21.4
SPL↑42.7843.6343.6842.55
RGSPL↑28.3529.7528.1728.42
+ +Table 3: Ablation of LLM Analyzer. + +
MethodNavigationGrounding
TLOSR↑SR↑SPL↑RGS↑RGSPL↑
No-Selection (Baseline)21.9657.8252.7740.1534.7626.37
All-Inclusive20.7059.7655.5041.3636.4727.33
Random-Selection20.1558.0253.7942.1635.3027.08
LLM-Selection (Ours)21.8962.6257.4043.6338.8829.75
+ +Table 5: Ablation study of the guidance information. + +
MethodNavigationGrounding
TLOSR↑SR↑SPL↑RGS↑RGSPL↑
Baseline21.9657.8252.7740.1534.7626.37
+ Location20.9762.2056.8042.9737.7428.46
+ Prioritization20.4859.4754.7041.7435.8627.40
+ Destination Description19.2159.5953.9941.9134.9927.11
+ Previous Observations19.2257.9653.7342.8335.6427.12
+ +ScaleVLN [42], which collects 4.9M instruction-trajectory pairs to train the model, our method still achieves better performance (2.8% on SPL and 3.77% on RGSPL). + +CVDN-target As shown in Table 2, we compare VLN-Copilot with three main-stream VLN models HAMT [5], DUET [7], and AutoVLN [6] on the CVDN-target dataset. For a fair comparison, both models are trained on the CVDN-target dataset. $\dagger$ represents the reproduction results of these methods. It highlights the effectiveness of VLN-Copilot, especially in terms of GP metrics, where it outperforms all other methods. Although TL is slightly higher than HAMT, its significantly improved GP emphasizes its overall advantage in handling unseen validation data. + +# 4.3 Ablation Study + +Effect of LLM Analyzer To evaluate the ability of LLM for guidance type selection, we conduct an ablation study, there are four settings, as follows: (1) No-Selection: our baseline, which means the agent does not use any guidance information; (2) All-Inclusive: the agent lets the LLM generate all types of guidance and integrates this information; (3) Random-Selection: the agent randomly selects one of four guidance types and lets the LLM generate corresponding guidance information; (4) LLM-Selection: our method, the LLM outputs guidance type and generates the information. As illustrated in Table 3, we can see that the LLM-Selection method achieved the highest navigation main metric spl with a value of $43.63\%$ , followed by the Random-Selection method with a score of $42.16\%$ . This suggests that blindly combining all types of guidance information is not the best option, probably due to the redundancy of excessively detailed information. Compared to the Baseline, the LLM-Selection method improves the SR by $4.63\%$ , SPL by $3.48\%$ , RGS by $4.12\%$ , and $3.38\%$ . + +Different Threshold of Confusion Score As shown in Table 4, we conduct an ablation study on threshold $\tau$ in the confusion score. Although when $\tau = 1$ , SPL is slightly lower + +![](images/5805db1901ab93fa70dd35ef66d2b0c4962f877a8db9c4fd3546dd4f67de78a3.jpg) +Fig.5:Confusion Score + +![](images/befe241a5f3078077a5b31fed9fbf0dc67967eb321a00f9072657552949ca5ee.jpg) +Fig. 6: Learning Curve + +than $\tau = 1.2$ , while RGSPL reaches a higher value, so considering the two indicators of SPL and RGSPL, we set $\tau$ to 1. + +Effect of Guidance To validate the effect of different types of guidance information, we conducted an ablation study that only adds one type of information. As shown in Table 5, adding each type of the four categories of guidance information contributes to performance improvement. As shown in Line 1, adding location information contributes mostly, which shows that letting the agent know its own location and target location is of great help. As shown in Line 4, adding previous observations information, the agent's SPL increased significantly $(2.68\uparrow)$ , but the SR increased slightly $(0.96\uparrow)$ . This may be attributed to the agent's ability to reference historical trajectories, thereby gaining a better understanding of the navigation environment. This, in turn, helps to avoid redundant observations and improve navigation efficiency. Destination Description and Previous Observations reduce TL from 21.96 to 19.21 and 19.22, which shows such two guidance can help the agent reach the destination faster. + +Trends of Confusion Score We plot the agent's cumulative confusion score at each time step every 5k training iterations. As shown in Figure 5, the trend of confusion score is decreasing, especially in the model using VLN-Copilot, the curve is steeper, illustrating that our model could reduce the agent's confusion degree in navigation. + +Learning Curves We compared the learning curves between methods with and without VLN-Copilot during training. As shown in Figure 6, our VLN-Copilot converges faster than the model without Copilot and can achieve better performance (i.e. higher SPL value) in fewer iterations. + +# 4.4 Qualitative Results + +We present some examples to see the qualitative results of the LLM-generated guidance. As shown in Figure 7, for the example on the left, LLM selects "Destination Description" as the guidance type. Since the guidance information contains a description of the destination and matches candidate direction 2, it is easier to select the answer. For + +![](images/67dc9fb7d2011b9cc71af352c796d92b4d2d5bab7f9c01fc5ce20d4f37063bb4.jpg) +Fig. 7: Examples of the LLM-generated Guidance. + +![](images/4fd149b0527292061a4b4d3d7d387c8e2481ef87bbaccfcbabce32cf2a1c673a.jpg) + +the example on the right, since Candidate 2 and Candidate 3 are both in the kitchen, and the task is to go to the kitchen, it looks like each candidate is similar, making the choice difficult. While LLM predicts that the guidance type is "location" and provides the agent's current location and destination, we find that the agent has reached the target location and therefore should stop rather than select a direction to go. More results can be found in supplementary material. + +# 5 Limitations and Future Work + +In this work, we utilize Large Language Models (LLMs) to analyze the navigation environment by transforming the dynamic environment into textual descriptions. This practice may be sub-optimal for environmental perception. Our future work will focus on how to effectively apply Multimodal Large Language Models (MLLMs) such as LLaVA [23] and Video-LLaVA [21] to perceive the navigation environment and provide guidance information. + +# 6 Conclusion + +In this work, we propose a method named VLN-Copilot that enables the agent to proactively ask for help and lets the LLM act as a co-pilot to assist the agent in completing the coarse-grained VLN task. We first introduce a confusion score calculation mechanism that allows the agent to actively query the LLM when it feels uncertain about the direction of choice. Then, we use LLM to analyze the challenges faced by the agent, and then generate appropriate guidance information for different situations that the agent is in. We conduct extensive experiments on two coarse-grained VLN tasks, including quantitative and qualitative experiments, demonstrating our method's effectiveness. + +# Acknowledgements + +We thank Mr. Xiangyu Shi for the help in supplying 2 extra NVIDIA RTX 4090 GPUs for experiments. + +# References + +1. An, D., Qi, Y., Li, Y., Huang, Y., Wang, L., Tan, T., Shao, J.: Bevbert: Multimodal map pre-training for language-guided navigation. In: ICCV. pp. 2737-2748 (2023) +2. Anderson, P., Chang, A.X., Chaplot, D.S., Dosovitskiy, A., Gupta, S., Koltun, V., Kosecka, J., Malik, J., Mottaghi, R., Savva, M., Zamir, A.R.: On evaluation of embodied navigation agents. CoRR abs/1807.06757 (2018) +3. Anderson, P., Wu, Q., Teney, D., Bruce, J., Johnson, M., Sünderhauf, N., Reid, I.D., Gould, S., van den Hengel, A.: Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In: CVPR. pp. 3674-3683 (2018) +4. Chen, H., Suhr, A., Misra, D., Snavely, N., Artzi, Y.: TOUCHDOWN: natural language navigation and spatial reasoning in visual street environments. In: CVPR. pp. 12538-12547 (2019) +5. Chen, S., Guhur, P., Schmid, C., Laptev, I.: History aware multimodal transformer for vision-and-language navigation. In: NeurIPS. pp. 5834-5847 (2021) +6. Chen, S., Guhur, P.L., Tapaswi, M., Schmid, C., Laptev, I.: Learning from unlabeled 3d environments for vision-and-language navigation. In: ECCV (2022) +7. Chen, S., Guhur, P.L., Tapaswi, M., Schmid, C., Laptev, I.: Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In: CVPR (2022) +8. Fan, Y., Chen, W., Jiang, T., Zhou, C., Zhang, Y., Wang, X.E.: Aerial vision-and-dialog navigation. In: Findings of the Association for Computational Linguistics: ACL 2023 (2023) +9. Feng, W., Fu, T.J., Lu, Y., Wang, W.Y.: ULN: Towards underspecified vision-and-language navigation. In: EMNLP. pp. 6394-6412 (2022) +0. Gao, C., Peng, X., Yan, M., Wang, H., Yang, L., Ren, H., Li, H., Liu, S.: Adaptive zone-aware hierarchical planner for vision-language navigation. In: CVPR. pp. 14911-14920 (2023) +1. Guhur, P., Tapaswi, M., Chen, S., Laptev, I., Schmid, C.: Airbert: In-domain pretraining for vision-and-language navigation. In: ICCV. pp. 1634-1643 (2021) +2. Hong, Y., Opazo, C.R., Wu, Q., Gould, S.: Sub-instruction aware vision-and-language navigation. In: Webber, B., Cohn, T., He, Y., Liu, Y. (eds.) EMNLP. pp. 3360-3376 (2020) +3. Hong, Y., Wu, Q., Qi, Y., Opazo, C.R., Gould, S.: A recurrent vision-and-language BERT for navigation. In: CVPR. pp. 1643-1653 (2021) +4. Jain, V., Magalhães, G., Ku, A., Vaswani, A., Ie, E., Baldridge, J.: Stay on the path: Instruction fidelity in vision-and-language navigation. pp. 1862-1872 (2019) +5. Kaufman, L., Rousseeuw, P.J.: Finding groups in data: an introduction to cluster analysis. John Wiley & Sons (2009) +6. Krantz, J., Wijmans, E., Majumdar, A., Batra, D., Lee, S.: Beyond the nav-graph: Vision-and-language navigation in continuous environments. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J. (eds.) ECCV. pp. 104-120 (2020) +7. Ku, A., Anderson, P., Patel, R., Ie, E., Baldridge, J.: Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In: EMNLP. pp. 4392-4412 (2020) +8. Li, J., Tan, H., Bansal, M.: Envideit: Environment editing for vision-and-language navigation. In: CVPR. pp. 15386-15396 (2022) + +19. Li, J., Li, D., Xiong, C., Hoi, S.C.H.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML (2022), https://api.sementicscholar.org/CorpusID:246411402 +20. Li, X., Wang, Z., Yang, J., Wang, Y., Jiang, S.: Kerm: Knowledge enhanced reasoning for vision-and-language navigation. In: CVPR. pp. 2583-2592 (2023) +21. Lin, B., Zhu, B., Ye, Y., Ning, M., Jin, P., Yuan, L.: Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122 (2023) +22. Lin, K., Chen, P., Huang, D., Li, T.H., Tan, M., Gan, C.: Learning vision-and-language navigation from youtube videos. In: ICCV. pp. 8317-8326 (2023) +23. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023) +24. Liu, R., Wang, X., Wang, W., Yang, Y.: Bird's-eye-view scene graph for vision-language navigation. In: ICCV. pp. 10968-10980 (2023) +25. Liu, S., Zhang, H., Qi, Y., Wang, P., Zhang, Y., Wu, Q.: Aerialvln: Vision-and-language navigation for uavs. In: ICCV. pp. 15384-15394 (October 2023) +26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR. OpenReview.net (2019) +27. Nguyen, K., III, H.D.: Help, anna! visual navigation with natural multimodal assistance via retrospective curiosity-encouraging imitation learning. In: Inui, K., Jiang, J., Ng, V., Wan, X. (eds.) EMNLP. pp. 684–695. Association for Computational Linguistics (2019) +28. Qi, Y., Pan, Z., Hong, Y., Yang, M., van den Hengel, A., Wu, Q.: The road to know-where: An object-and-room informed sequential bert for indoor vision-language navigation. In: ICCV. pp. 1655–1664 (2021) +29. Qi, Y., Wu, Q., Anderson, P., Wang, X., Wang, W.Y., Shen, C., van den Hengel, A.: REVERIE: remote embodied visual referring expression in real indoor environments. In: CVPR. pp. 9979-9988 (2020) +30. Qiao, Y., Qi, Y., Hong, Y., Yu, Z., Wang, P., Wu, Q.: Hop+: History-enhanced and order-aware pre-training for vision-and-language navigation. IEEE TPAMI (2023) +31. Qiao, Y., Qi, Y., Yu, Z., Liu, J., Wu, Q.: March in chat: Interactive prompting for remote embodied referring expression. In: ICCV. pp. 15758-15767 (2023) +32. Qiao, Y., Yu, Z., Wu, Q.: Vln-petl: Parameter-efficient transfer learning for vision-and-language navigation. In: ICCV. pp. 15443-15452 (2023) +33. Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert-networks. In: Inui, K., Jiang, J., Ng, V., Wan, X. (eds.) EMNLP-IJCNLP. pp. 3980-3990 (2019) +34. Roman, H.R., Bisk, Y., Thomason, J., Celikyilmaz, A., Gao, J.: Rmm: A recursive mental model for dialog navigation. arXiv preprint arXiv:2005.00728 (2020) +35. Schumann, R., Zhu, W., Feng, W., Fu, T.J., Riezler, S., Wang, W.Y.: Velma: Verbalization embodiment of llm agents for vision and language navigation in street view. In: AAAI (2023) +36. Song, C.H., Wu, J., Washington, C., Sadler, B.M., Chao, W.L., Su, Y.: Llm-planner: Few-shot grounded planning for embodied agents with large language models. arXiv preprint arXiv:2212.04088 (2022) +37. Thomason, J., Murray, M., Cakmak, M., Zettlemoyer, L.: Vision-and-dialog navigation. In: CoRL. pp. 394-406 (2019) +38. Touvron, H., Martin, L., Stone, K.R., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D.M., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A.S., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I.M., Korenev, A.V., Koura, P.S., Lachaux, M.A., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., + +Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. ArXiv abs/2307.09288 (2023), https://api.sementicscholar.org/CorpusID:259950998 +39. Wang, X., Wang, W., Shao, J., Yang, Y.: Lana: A language-capable navigator for instruction following and generation. In: CVPR. pp. 19048-19058 (2023) +40. Wang, X., Huang, Q., Celikyilmaz, A., Gao, J., Shen, D., Wang, Y., Wang, W.Y., Zhang, L.: Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In: CVPR. pp. 6629-6638 (2019) +41. Wang, Z., Li, X., Yang, J., Liu, Y., Jiang, S.: Gridmm: Grid memory map for vision-and-language navigation. In: ICCV. pp. 15625-15636 (2023) +42. Wang, Z., Li, J., Hong, Y., Wang, Y., Wu, Q., Bansal, M., Gould, S., Tan, H., Qiao, Y.: Scaling data generation in vision-and-language navigation. In: ICCV. pp. 12009-12020 (2023) +43. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685 (2023) +44. Zhu, F., Liang, X., Zhu, Y., Yu, Q., Chang, X., Liang, X.: SOON: scenario oriented object navigation with graph-based exploration. In: CVPR. pp. 12689-12699 (2021) +45. Zhu*, Y., Weng*, Y., Zhu, F., Liang, X., Ye, Q., Lu, Y., jiao, J.: Self-motivated communication agent for real-world vision-dialog navigation. In: ICCV (2021) \ No newline at end of file diff --git a/2024/LLM as Copilot for Coarse-grained Vision-and-Language Navigation/images.zip b/2024/LLM as Copilot for Coarse-grained Vision-and-Language Navigation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4ff237b9ceef0c52a60e656f17bba48dc458c16d --- /dev/null +++ b/2024/LLM as Copilot for Coarse-grained Vision-and-Language Navigation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63f3ca15ac9f968619a0ffcbbed67966bc9294fb372f2b45fee777d4a6f16331 +size 418409 diff --git a/2024/LLM as Copilot for Coarse-grained Vision-and-Language Navigation/layout.json b/2024/LLM as Copilot for Coarse-grained Vision-and-Language Navigation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ce10988d6557ec966af316b7959504414de9d2fb --- /dev/null +++ b/2024/LLM as Copilot for Coarse-grained Vision-and-Language Navigation/layout.json @@ -0,0 +1,10431 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 197, + 112, + 416, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 112, + 416, + 147 + ], + "spans": [ + { + "bbox": [ + 197, + 112, + 416, + 147 + ], + "type": "text", + "content": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 147, + 167, + 465, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 167, + 465, + 181 + ], + "spans": [ + { + "bbox": [ + 147, + 167, + 465, + 181 + ], + "type": "text", + "content": "Yanyuan Qiao" + }, + { + "bbox": [ + 147, + 167, + 465, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 167, + 465, + 181 + ], + "type": "text", + "content": ", Qianyi Liu" + }, + { + "bbox": [ + 147, + 167, + 465, + 181 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 147, + 167, + 465, + 181 + ], + "type": "text", + "content": ", Jiajun Liu" + }, + { + "bbox": [ + 147, + 167, + 465, + 181 + ], + "type": "inline_equation", + "content": "^{4,5}" + }, + { + "bbox": [ + 147, + 167, + 465, + 181 + ], + "type": "text", + "content": ", Jing Liu" + }, + { + "bbox": [ + 147, + 167, + 465, + 181 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 147, + 167, + 465, + 181 + ], + "type": "text", + "content": ", and Qi Wu" + }, + { + "bbox": [ + 147, + 167, + 465, + 181 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 175, + 189, + 438, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 189, + 438, + 200 + ], + "spans": [ + { + "bbox": [ + 175, + 189, + 438, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 175, + 189, + 438, + 200 + ], + "type": "text", + "content": " Australian Institute for Machine Learning, The University of Adelaide" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 202, + 201, + 411, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 201, + 411, + 212 + ], + "spans": [ + { + "bbox": [ + 202, + 201, + 411, + 212 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 202, + 201, + 411, + 212 + ], + "type": "text", + "content": " Institute of Automation, Chinese Academy of Sciences" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 163, + 212, + 449, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 163, + 212, + 449, + 222 + ], + "spans": [ + { + "bbox": [ + 163, + 212, + 449, + 222 + ], + "type": "text", + "content": "3 School of Artificial Intelligence, University of Chinese Academy of Sciences" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 275, + 223, + 338, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 223, + 338, + 232 + ], + "spans": [ + { + "bbox": [ + 275, + 223, + 338, + 232 + ], + "type": "text", + "content": "4 CSIRO Data61" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 247, + 233, + 366, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 233, + 366, + 244 + ], + "spans": [ + { + "bbox": [ + 247, + 233, + 366, + 244 + ], + "type": "text", + "content": "5 The University of Queensland" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 246, + 475, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 475, + 256 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 475, + 256 + ], + "type": "text", + "content": "{yanyuan.qiao,qi.wu01}@adelaide.edu.au, liuqianyi2022@ia.ac.cn," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 199, + 257, + 409, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 257, + 409, + 266 + ], + "spans": [ + { + "bbox": [ + 199, + 257, + 409, + 266 + ], + "type": "text", + "content": "jiajun.liu@csiro.au, jliu@nlpr.ia.ac.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 159, + 295, + 453, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 295, + 453, + 449 + ], + "spans": [ + { + "bbox": [ + 159, + 295, + 453, + 449 + ], + "type": "text", + "content": "Abstract. Vision-and-Language Navigation (VLN) involves guiding an agent through indoor environments using human-provided textual instructions. Coarse-grained VLN, with short and high-level instructions, has gained popularity as it closely mirrors real-world scenarios. However, a significant challenge is these instructions are often too concise for agents to comprehend and act upon. Previous studies have explored allowing agents to seek assistance during navigation, but typically offer rigid support from pre-existing datasets or simulators. The advent of Large Language Models (LLMs) presents a novel avenue for aiding VLN agents. This paper introduces VLN-Copilot, a framework enabling agents to actively seek assistance when encountering confusion, with the LLM serving as a copilot to facilitate navigation. Our approach includes the introduction of a confusion score, quantifying the level of uncertainty in an agent's action decisions, while the LLM offers real-time detailed guidance for navigation. Experimental results on two coarse-grained VLN datasets show the efficacy of our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 160, + 459, + 427, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 459, + 427, + 471 + ], + "spans": [ + { + "bbox": [ + 160, + 459, + 427, + 471 + ], + "type": "text", + "content": "Keywords: Vision-and-Language " + }, + { + "bbox": [ + 160, + 459, + 427, + 471 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 459, + 427, + 471 + ], + "type": "text", + "content": " Navigation " + }, + { + "bbox": [ + 160, + 459, + 427, + 471 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 459, + 427, + 471 + ], + "type": "text", + "content": " Large Language Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 491, + 218, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 218, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 218, + 502 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 515, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 647 + ], + "type": "text", + "content": "The Vision-and-Language Navigation (VLN) tasks have attracted great attention from researchers, which enables an agent to navigate to a specified destination indoors based on the provided textual instructions and visual observations. The tasks of VLN show great potential in real-world applications such as home assistant robots. Most VLN tasks such as R2R [3] and RxR [17] etc. provide fine-grained step-by-step instructions, such as \"Go out the door in front of you across the room. Once in the hallway turn Left. Walk forward and then turn left into the sitting area. Stop in right in front of the fireplace.\" However, though these detailed instructions could improve the agents' ability of instructions following, they impede the further application of VLN in the real world since human beings would not like to give such long instructions to a robot in daily life. Contrastive to these tasks, the coarse-grained instruction task represented by the Remote" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 654, + 223, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 654, + 223, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 654, + 223, + 665 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 194, + 114, + 421, + 281 + ], + "blocks": [ + { + "bbox": [ + 194, + 114, + 421, + 281 + ], + "lines": [ + { + "bbox": [ + 194, + 114, + 421, + 281 + ], + "spans": [ + { + "bbox": [ + 194, + 114, + 421, + 281 + ], + "type": "image", + "image_path": "e55f9e2f288adaacf22334edf36d690e10d38829f245995fcb2b513033a9042e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 297, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 131, + 297, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 131, + 297, + 482, + 331 + ], + "type": "text", + "content": "Fig. 1: Our proposed VLN-Copilot. When an agent is confused about making an action decision, it will proactively seek help from the LLM, and then the LLM will assist the agent by analyzing the situations faced by the agent and providing appropriate guidance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 360, + 482, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 421 + ], + "type": "text", + "content": "Embodied Referring Expression (REVERIE) task [29] is more expected to enhance the practical utility of VLN because it involves instructions that are closer to those encountered in real-world scenarios, such as \"Go clean the mirror in the bathroom.\" Giving such coarse-grained instructions to the agent would bring more challenges to the agent's navigation than fine-grained instructions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 423, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 628 + ], + "type": "text", + "content": "VLN Agents navigating in environments can benefit from seeking external assistance [34]. Early works [27, 37] tried to let agents seek help from oracles or simulators, but the assistance content obtained was fixed and limited. Since this assistance information is based on pre-built ground truth information in datasets or simulators, which are not consistent with real-life scenarios. Inspired by the development of Large Language Models (LLMs) with rich internalized world knowledge, many works have been made to exploit their capabilities to aid navigation tasks. One recent method named MiC [31] makes the first step to use LLM in the coarse-grained VLN task of REVERIE, which queries the LLM to generate more detailed instructions by prompting the LLM with extra planning examples. The promising performance of MiC shows the great potential to use LLM in coarse-grained VLN tasks. However, MiC has some weaknesses in applying LLM for VLN. First, the way the agent seeks help from the LLM is a passive way that depends on the change of the room where the agent is located. This approach may not help the agent when it faces difficulties in making predictions while still in the same room. Second, MiC generates fine-grained planning no matter what challenges the agent faces. This practice lacks a thorough analysis of the difficulties the agent may have and fails to give the agent diverse guidance for different scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "To address these issues, we propose a novel method named VLN-Copilot that makes the LLM a competent copilot to assist the agent in navigation. As shown in Figure 1, before making a movement, the agent calculates a confusion score to assess the level of" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Y. Qiao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "uncertainty associated with its current action choice. If this score surpasses a predefined threshold, it triggers the agent to seek guidance. This practice is more proactive than relying on the change of scenes which may be wrong. When the agent decides to seek assistance, it consults the LLM and then the LLM conducts a comprehensive evaluation of the situation, identifies the type of guidance required, and generates the necessary guidance to aid the agent in navigating effectively. The types of guidance information are selected from the in-depth pre-analysis of prevailing challenges the agent may face in the sampled environment using LLM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 259 + ], + "type": "text", + "content": "We conduct experiments on the REVERIE and CVDN-target dataset and our model surpasses the previous methods. Specifically, our VLN-Copilot obtains " + }, + { + "bbox": [ + 130, + 212, + 482, + 259 + ], + "type": "inline_equation", + "content": "42.32\\%" + }, + { + "bbox": [ + 130, + 212, + 482, + 259 + ], + "type": "text", + "content": " on the main navigation metric SPL and " + }, + { + "bbox": [ + 130, + 212, + 482, + 259 + ], + "type": "inline_equation", + "content": "26.55\\%" + }, + { + "bbox": [ + 130, + 212, + 482, + 259 + ], + "type": "text", + "content": " on the main object grounding metric RGSPL on the REVERIE dataset, and 4.47 in GP on the CVDN-target dataset, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 260, + 331, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 260, + 331, + 272 + ], + "spans": [ + { + "bbox": [ + 146, + 260, + 331, + 272 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 279, + 479, + 396 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 279, + 479, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 279, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 279, + 479, + 312 + ], + "type": "text", + "content": "- We propose a novel method named VLN-Copilot which utilizes LLM to analyze the difficulties the agent may meet and give diverse guidance for different situations, rather than to generate planning by prompting LLM with extra planning examples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 316, + 479, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 479, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 479, + 336 + ], + "type": "text", + "content": "- We introduce the confusion score to help the agent proactively ask the LLM for help when confused to make predictions rather than passively trigger conversations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 339, + 479, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 339, + 479, + 372 + ], + "spans": [ + { + "bbox": [ + 138, + 339, + 479, + 372 + ], + "type": "text", + "content": "- We utilize LLM with elaborately designed prompts to conduct a detailed challenge analysis of the agent's different situations, enabling the LLM to provide better guidance on the fly." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 375, + 479, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 375, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 375, + 479, + 396 + ], + "type": "text", + "content": "- Extensive experiments on two VLN benchmarks show the effectiveness of our proposed method." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 417, + 225, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 417, + 225, + 430 + ], + "spans": [ + { + "bbox": [ + 131, + 417, + 225, + 430 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 442, + 298, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 442, + 298, + 455 + ], + "spans": [ + { + "bbox": [ + 131, + 442, + 298, + 455 + ], + "type": "text", + "content": "2.1 Vision-and-Language Navigation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 462, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 521 + ], + "type": "text", + "content": "In recent years, Vision-and-Language Navigation (VLN) has attracted widespread attention due to its promising applications in home assistant robots and related fields [9, 12, 18, 20, 28, 40]. Many different tasks have emerged in the field of VLN [4, 14, 16, 25, 27, 29, 37]. In terms of the navigation space accessible to the agent, these cover indoor [3, 16, 17], outdoor [4, 35], and aerial scenarios [8, 25]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": "Early VLN task R2R [3] started from indoor environments requiring the agent to reach the destination according to the given instructions, and there were some subsequent tasks (e.g., RxR [17] and VLN-CE [16]). These tasks all provide detailed natural language instructions, such as \"Walk down the stairs at the bottom of the stairs take a left. Walk straight to get to the kitchen inside and wait just inside the door.\" These detailed instructions guide the agent through complex step-by-step processes and are more focused on testing the agent's ability to understand and follow instructions. Some other tasks adopt an object-oriented perspective, such as REVERIE [29] and SOON [44]. These tasks require not only reaching a designated destination but also discovering specific target objects. Different from SOON providing detailed instructions and object recognition prompts, such as \"The cabinet is square, closed and wooden. The cabinet is placed below the sink and the water-tap. There are some stools and table nearby.\"" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 195, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 195, + 91, + 447, + 102 + ], + "type": "text", + "content": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 175 + ], + "type": "text", + "content": "REVERIE's instructions appear more concise and natural, such as \"Bring me the swivel chair from the office.\" These coarse-grained instructions, akin to instructions encountered in real-world scenarios, enhance the potential for innovative research within the domain of coarse-grained instruction VLN. However, the task remains challenging due to the limited information these instructions provide." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 194, + 280, + 207 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 280, + 207 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 280, + 207 + ], + "type": "text", + "content": "2.2 VLN Agents Asking for Help" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 216, + 481, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 481, + 514 + ], + "type": "text", + "content": "Agents can gain advantages by seeking external assistance to assist navigation, and there have been studies promoting navigation agents to seek help during navigation tasks [27, 34]. Thomason et al. [37] introduce the Navigation from Dialog History (NDH) task, which contains dialog history between agents and oracles during navigation. The agent follows guide instructions and asks questions to the oracle when needed. The task provides a valuable resource for investigating the navigation interactions between agent and oracle. Similarly, HANNA [27] builds a simulator where navigation agents can request language-and-vision assistance when they face difficulties during navigation. However, the agent's signals asking for help are fixed and predefined in the simulator. Additionally, the assistance provided is based on ground truth information derived from the R2R dataset, which may not fully represent real-world navigation scenarios due to its reliance on pre-built datasets. Similarly, Zhu et al. proposed SCoA [45], which enables navigation agents to ask questions from a set of pre-constructed candidate question-answer pairs to aid in navigation. Recently, LLM-Planner [36] and MiC [31] have attempted to leverage the knowledge of large language models (LLM) to assist agents in navigation. Agents can query the LLM to generate detailed step-by-step instructions to aid navigation. However, the ways the agents ask for help from both methods are passive, either based on fixed time steps or when the room the agent is in changes. This practice may miss the opportunity to help the agent when it has difficulties in prediction actions. Besides, these two methods concentrate on generating planning by prompting LLM with extra planning data. Different from these methods, we propose a method named VLN-Copilot that allows agents to proactively seek help when they are not sure where to go in this work. Moreover, our VLN-Copilot pays more attention to analyzing the difficulties the agent may meet and can dynamically give appropriate types of guidance in diverse situations for the agent." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 534, + 192, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 534, + 192, + 547 + ], + "spans": [ + { + "bbox": [ + 133, + 534, + 192, + 547 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 560, + 214, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 560, + 214, + 572 + ], + "spans": [ + { + "bbox": [ + 133, + 560, + 214, + 572 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "text", + "content": "The main goal for the agent in the VLN task is to navigate to a specific destination following the high-level concise instructions. Specifically, the agent is asked to navigate in the simulated environment, which can be described as an undirected graph " + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\{\\mathcal{V},\\mathcal{E}\\}" + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "text", + "content": " represents the navigable nodes and " + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "text", + "content": " represents the connectivity edges. Given the natural language instruction " + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "text", + "content": ", the agent starts from the initial node " + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "V_{0}" + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "text", + "content": " and perceives a panoramic view " + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t" + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "text", + "content": " of the current node " + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "text", + "content": " to make the action " + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "text", + "content": " at each time step " + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "text", + "content": ". The panoramic view " + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t" + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "text", + "content": " has " + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "text", + "content": " single-view images " + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 133, + 582, + 481, + 665 + ], + "type": "text", + "content": " and can be represented as" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Y. Qiao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 155, + 121, + 460, + 219 + ], + "blocks": [ + { + "bbox": [ + 155, + 121, + 460, + 219 + ], + "lines": [ + { + "bbox": [ + 155, + 121, + 460, + 219 + ], + "spans": [ + { + "bbox": [ + 155, + 121, + 460, + 219 + ], + "type": "image", + "image_path": "48d79f507645ff044e946dc24f4b66da236e91bc512b62ddd0f1eca1bbfe105d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "type": "text", + "content": "Fig.2: Overview of the VLN-Copilot. At each time step, the agent predicts the logits over the candidate's view of actions. After that, the agent will calculate the confusion score to decide whether to ask LLM for help (Sec. 3.3). When the confusion score surpasses threshold " + }, + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "type": "text", + "content": ", the agent queries the LLM for help with the prompt that helps LLM perceive the environment (Sec. 3.4). Then the LLM generates useful guidance information, which will be returned to the agent to assist in decision-making (Sec. 3.6)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 333, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t = \\{r_i\\}_{i=1}^K" + }, + { + "bbox": [ + 130, + 333, + 482, + 430 + ], + "type": "text", + "content": ", of which " + }, + { + "bbox": [ + 130, + 333, + 482, + 430 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 130, + 333, + 482, + 430 + ], + "type": "text", + "content": " is represented by an image feature vector with the corresponding orientation feature vector. Besides, for the sub-task of object grounding, the object features " + }, + { + "bbox": [ + 130, + 333, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_t = \\{o_i\\}_{i=1}^M" + }, + { + "bbox": [ + 130, + 333, + 482, + 430 + ], + "type": "text", + "content": " are extracted from the panoramic view " + }, + { + "bbox": [ + 130, + 333, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t" + }, + { + "bbox": [ + 130, + 333, + 482, + 430 + ], + "type": "text", + "content": " with the annotated object bounding boxes or by object detectors. With these observations of the environment, the agent takes a series of actions " + }, + { + "bbox": [ + 130, + 333, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\langle a_0, \\dots, a_N \\rangle" + }, + { + "bbox": [ + 130, + 333, + 482, + 430 + ], + "type": "text", + "content": " to navigate to the target location, where each action " + }, + { + "bbox": [ + 130, + 333, + 482, + 430 + ], + "type": "inline_equation", + "content": "a_t" + }, + { + "bbox": [ + 130, + 333, + 482, + 430 + ], + "type": "text", + "content": " selects one navigable node from the candidates. When the target object is grounded or the pre-defined maximum navigation steps have been achieved, the navigation ends." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 451, + 198, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 451, + 198, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 198, + 462 + ], + "type": "text", + "content": "3.2 Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": "As is shown in Figure 2 and Algorithm 1, given the coarse-grained instruction, the agent first perceives the environment to obtain the current visual observation and starts to predict a probability over candidate actions. Then, a confusion score is computed according to the probability distribution, which determines whether the agent is confused about making predictions in the current situation and whether the agent should ask the LLM for help. If the confusion score exceeds a predefined threshold " + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": " (obtained via grid search) which means the agent is confused about making decisions, the agent initiates a conversation with the LLM to seek their help with an elaborately designed prompt. Specifically, the prompt contains coarse-grained instruction, fine-grained scene descriptions of candidate views, historical descriptions of trajectory scenes, and the requirements for the LLM. Then, the LLM analyzes the agent's situation and gives advice about which type of guidance information is useful as well as the details about that guidance information. The types of guidance information are refined from the results of challenge pre-analysis by LLM. Finally, the response about the guidance from the LLM is appended after the coarse-grained instruction to guide the agent's action prediction collaboratively. This paradigm aims to optimize agents' decision-making processes by" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 195, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 195, + 91, + 448, + 102 + ], + "type": "text", + "content": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 140 + ], + "type": "text", + "content": "combining instruction, observation, and LLM-based world knowledge via the conversations between the agent and LLM." + } + ] + } + ], + "index": 2 + }, + { + "type": "code", + "bbox": [ + 176, + 167, + 440, + 364 + ], + "blocks": [ + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "lines": [ + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "spans": [ + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "text", + "content": "Algorithm 1 VLN-Copilot \nNotation Summary: I: high-level instruction " + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t" + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "text", + "content": " : visual observation at timestep t " + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "text", + "content": " : confusion score at timestep t " + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "inline_equation", + "content": "W_{t}^{F}" + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "text", + "content": " : fine-grained scene descriptions of all candidate views at timestep t " + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "inline_equation", + "content": "W_{t}^{\\mathrm{Tr}}" + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "text", + "content": " : historical trajectory description at timestep t " + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{guide}}" + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "text", + "content": " : prompt template for LLM to generate guidance information LLM: large language model " + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "text", + "content": " : pre-defined threshold for confusion score \n" + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "inline_equation", + "content": "t\\gets 0" + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "inline_equation", + "content": "\\rightharpoondown" + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "text", + "content": " Initial timestep while " + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "inline_equation", + "content": "t < \\max" + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "text", + "content": " step and " + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "inline_equation", + "content": "\\hat{a}_{t}\\neq" + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "text", + "content": " \"stop\" do if " + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "inline_equation", + "content": "S_{t} > = \\tau" + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "text", + "content": " then " + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{guide}}\\leftarrow T_{\\mathrm{guide}}(I,W_t^F,W_t^{\\mathrm{Tr}})" + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{guide}}\\gets \\mathrm{LLM}(P_{\\mathrm{guide}})" + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "inline_equation", + "content": "I_{t}^{\\mathrm{Inter}}\\gets \\mathrm{CONCAT}(I,R_{\\mathrm{guide}})" + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "inline_equation", + "content": "\\hat{a}_{t}\\gets \\mathrm{Agent}(I_{t}^{\\mathrm{Inter}},\\mathcal{R}_{t})" + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "text", + "content": " else " + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "inline_equation", + "content": "\\hat{a}_{t}\\gets \\mathrm{Agent}(I,\\mathcal{R}_{t})" + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "text", + "content": " end if " + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "inline_equation", + "content": "t\\gets t + 1" + }, + { + "bbox": [ + 176, + 167, + 440, + 364 + ], + "type": "text", + "content": " \nend while" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "bbox": [ + 132, + 379, + 277, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 379, + 277, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 379, + 277, + 392 + ], + "type": "text", + "content": "3.3 When to Ask LLM for Help" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "type": "text", + "content": "Different from previous methods [31] that set fixed timestep for interactions with LLM or trigger the interactions when the room scenes change, we select a more active practice for the agent by using the confusion score to decide whether to ask for help. Specifically, at each time step " + }, + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "type": "text", + "content": ", when the agent is going to predict the logits over the candidate nodes for the next action " + }, + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "type": "text", + "content": ", we calculate the entropy of the predicted probability distribution over each candidate navigable node " + }, + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "type": "text", + "content": " as the confusion score " + }, + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 130, + 400, + 482, + 472 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 227, + 481, + 481, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 481, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 227, + 481, + 481, + 495 + ], + "type": "interline_equation", + "content": "p _ {t} \\left(v _ {i}\\right) = \\operatorname {S o f t m a x} \\left(s _ {t} \\left(v _ {i}\\right)\\right), \\tag {1}", + "image_path": "272759b30967bca360ea3f5662c86d14344ba56f96057af7d0718f386dd226fc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 242, + 497, + 481, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 497, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 242, + 497, + 481, + 520 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {t} = - \\sum_ {i} p _ {t} \\left(v _ {i}\\right) \\log_ {2} \\left(p _ {t} \\left(v _ {i}\\right) + \\epsilon\\right), \\tag {2}", + "image_path": "8bb9f047055a6fab9e2525a841b5505f00c39395c98761f46dcc24c1749f8f76.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 530, + 480, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 530, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 530, + 480, + 553 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 530, + 480, + 553 + ], + "type": "inline_equation", + "content": "s_t(v_i)" + }, + { + "bbox": [ + 130, + 530, + 480, + 553 + ], + "type": "text", + "content": " is the predicted logits assigned by the agent to the " + }, + { + "bbox": [ + 130, + 530, + 480, + 553 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 530, + 480, + 553 + ], + "type": "text", + "content": "-th navigable node " + }, + { + "bbox": [ + 130, + 530, + 480, + 553 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 130, + 530, + 480, + 553 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 130, + 530, + 480, + 553 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 530, + 480, + 553 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 530, + 480, + 553 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 530, + 480, + 553 + ], + "type": "text", + "content": " is a minimal number added to avoid numerical instability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 555, + 481, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 481, + 590 + ], + "type": "text", + "content": "Then, if the computed confusion score " + }, + { + "bbox": [ + 130, + 555, + 481, + 590 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 130, + 555, + 481, + 590 + ], + "type": "text", + "content": " surpasses a predefined threshold " + }, + { + "bbox": [ + 130, + 555, + 481, + 590 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 555, + 481, + 590 + ], + "type": "text", + "content": ", which means the agent is confused about making a decision among some candidate nodes, it will send a message to LLM asking for help." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 609, + 319, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 319, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 319, + 620 + ], + "type": "text", + "content": "3.4 How LLM Perceives the Environment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "LLMs would give inaccurate advice for the agent's navigation since they are not able to directly acquire visual content from the environment. Both current visual observations from the agent and previously observed scenes are essential for the LLM to analyze the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Y. Qiao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "navigation and supply important information. Thus, enhancing the LLM's environmental awareness as far as possible is a critical first step to making it a competent copilot. To solve this problem, we translate the dynamic observations and trajectories into multi-grained spatiotemporal descriptions as the eye for the LLM, using the out-of-the-box vision-and-language model of BLIP [19] model. Detailedly, the visually-grounded description mainly consists of three parts: the coarse-grained scene description of the current panoramic view " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ", the fine-grained scene description of each view " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " of selected candidate navigable nodes at the current time step " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ", and the trajectory description of previously observed scenes before the current time step " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": "Specifically, we obtain both the panoramic view image " + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t" + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": " of the agent's observation and the single view image " + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": " of each candidate navigable node at the time step " + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": ". We first feed " + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t" + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": " to the BLIP model to generate the coarse-grained scene description " + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "inline_equation", + "content": "W_t^C" + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": " for the agent's current location. For the fine-grained scene description, the agent is provided with " + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": " candidate views for selection. Then, we feed each single-view image " + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": " selected candidate navigable nodes to BLIP to generate the fine-grained scene description " + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "inline_equation", + "content": "W_t^{F_i}" + }, + { + "bbox": [ + 130, + 224, + 482, + 308 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 272, + 317, + 481, + 330 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 317, + 481, + 330 + ], + "spans": [ + { + "bbox": [ + 272, + 317, + 481, + 330 + ], + "type": "interline_equation", + "content": "W _ {t} ^ {\\mathrm {C}} = \\operatorname {B L I P} \\left(\\mathcal {R} _ {t}\\right), \\tag {3}", + "image_path": "517c8248861b787c2d52d020966ec0add47fc265d100cb43057f65a977c6705f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 268, + 334, + 480, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 334, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 268, + 334, + 480, + 346 + ], + "type": "interline_equation", + "content": "W _ {t} ^ {F _ {i}} = \\operatorname {B L I P} \\left(r _ {i}\\right), \\tag {4}", + "image_path": "19317f81adc48049751a978cc672900c41db87951d800a2cf10a6bcdbe4a7fdf.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 356, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 356, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 131, + 356, + 482, + 392 + ], + "type": "text", + "content": "Then we concatenate all previous coarse-grained scene descriptions " + }, + { + "bbox": [ + 131, + 356, + 482, + 392 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{C}}" + }, + { + "bbox": [ + 131, + 356, + 482, + 392 + ], + "type": "text", + "content": " of each time step before the current time " + }, + { + "bbox": [ + 131, + 356, + 482, + 392 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 356, + 482, + 392 + ], + "type": "text", + "content": " as the historical trajectory description " + }, + { + "bbox": [ + 131, + 356, + 482, + 392 + ], + "type": "inline_equation", + "content": "W_{t}^{\\mathrm{Tr}}" + }, + { + "bbox": [ + 131, + 356, + 482, + 392 + ], + "type": "text", + "content": " such as \"I passed a bedroom, and then I passed ...\" This process can be formulated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 236, + 401, + 481, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 401, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 236, + 401, + 481, + 415 + ], + "type": "interline_equation", + "content": "W _ {t} ^ {\\mathrm {T r}} = \\operatorname {C O N C A T} \\left(W _ {1} ^ {\\mathrm {C}}, \\dots , W _ {t - 1} ^ {\\mathrm {C}}\\right), \\tag {5}", + "image_path": "e6d2c364ff6e3f341683eab41c4b382a4edae829950403f7b75acad36292b708.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 424, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 520 + ], + "type": "text", + "content": "Note that we only generate the fine-grained scene descriptions and trajectory descriptions when the agent is confused to make the action prediction and asks the LLM for help, while we generate coarse-grained scene descriptions at each time step. In addition, for better comprehension of LLM, we added the candidate view id " + }, + { + "bbox": [ + 130, + 424, + 482, + 520 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 424, + 482, + 520 + ], + "type": "text", + "content": " as the prefix to each corresponding native fine-grained scene description " + }, + { + "bbox": [ + 130, + 424, + 482, + 520 + ], + "type": "inline_equation", + "content": "W_{t}^{F_{i}}" + }, + { + "bbox": [ + 130, + 424, + 482, + 520 + ], + "type": "text", + "content": ". For example, we converted the generated scene description of \"A room with a tilted ceiling, a bathroom mirror on the wall, and a skylight above the window.\" into the candidate view description with the following format: \"I see candidate 0 is a room with a tilted ceiling ...\"" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 537, + 371, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 537, + 371, + 550 + ], + "spans": [ + { + "bbox": [ + 131, + 537, + 371, + 550 + ], + "type": "text", + "content": "3.5 Difficulty and Useful Guidance-Type Pre-Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Before asking the agent to query an LLM to generate guidance for assistance, it would be beneficial to explore the difficulties the agent may encounter in making decisions and what guidance information the LLM should provide to be useful for the agent's navigation. However, manual analysis by humans is costly and impractical. Thus, we turn to LLM for analysis. Specifically, we first collect some navigation samples. When the agent finds it difficult to make decisions during the navigation process, we convert the situation the agent encountered into a text description and save it as a navigation sample. Then, we use an elaborate prompt to query the LLM to analyze the situation encountered by the agent in the current navigation example, such as what is the reason" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 195, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 195, + 91, + 447, + 102 + ], + "type": "text", + "content": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 136, + 119, + 473, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 119, + 473, + 135 + ], + "spans": [ + { + "bbox": [ + 136, + 119, + 473, + 135 + ], + "type": "text", + "content": "System: You are a helpful assistant for navigation. There is an agent navigating indoors according to task instructions and current observation. At each step, the agent needs to choose one of the candidate directions to go. I will tell you which direction is correct." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 138, + 472, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 138, + 472, + 152 + ], + "spans": [ + { + "bbox": [ + 137, + 138, + 472, + 152 + ], + "type": "text", + "content": "You need to first analyze the current situation, such as what difficulties the agent encountered during navigation that made it difficult to choose a direction and what types of guidance information may be useful." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 156, + 279, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 156, + 279, + 163 + ], + "spans": [ + { + "bbox": [ + 137, + 156, + 279, + 163 + ], + "type": "text", + "content": "Task instruction: Turn off the white table lamp in the bedroom" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 163, + 170, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 163, + 170, + 169 + ], + "spans": [ + { + "bbox": [ + 138, + 163, + 170, + 169 + ], + "type": "text", + "content": "Observation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 169, + 328, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 169, + 328, + 176 + ], + "spans": [ + { + "bbox": [ + 138, + 169, + 328, + 176 + ], + "type": "text", + "content": "I see candidate 1 is a white door leads to a bedroom with a bed, electric fan and table" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 176, + 321, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 321, + 182 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 321, + 182 + ], + "type": "text", + "content": "I see candidate 2 is a bathroom has a sink and the door leads to the shower room" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 182, + 325, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 182, + 325, + 188 + ], + "spans": [ + { + "bbox": [ + 138, + 182, + 325, + 188 + ], + "type": "text", + "content": "I see candidate 3 is a hallway with white walls, stairs, and a door on the second floor" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 188, + 417, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 188, + 417, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 188, + 417, + 194 + ], + "type": "text", + "content": "History trajectory: I have passed a balcony with rocking chairs and a table, then I go to a hallway with a picture of a bedroom." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 194, + 205, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 205, + 199 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 205, + 199 + ], + "type": "text", + "content": "Correct answer: candidate 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 204, + 172, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 204, + 172, + 210 + ], + "spans": [ + { + "bbox": [ + 138, + 204, + 172, + 210 + ], + "type": "text", + "content": "Assistant:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 210, + 471, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 210, + 471, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 210, + 471, + 217 + ], + "type": "text", + "content": "Difficulty: The agent encountered difficulties in choosing a direction because there are multiple doors at the same time. This makes it challenging for" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 217, + 267, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 267, + 223 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 267, + 223 + ], + "type": "text", + "content": "the agent to determine which direction could find the lamp." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 228, + 263, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 263, + 235 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 263, + 235 + ], + "type": "text", + "content": "Please list possible type of guidance may be useful." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 240, + 171, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 240, + 171, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 240, + 171, + 246 + ], + "type": "text", + "content": "Assistant:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 246, + 348, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 348, + 253 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 348, + 253 + ], + "type": "text", + "content": "Guidance Type: The agent needs to the destination information to make an informed decision." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 279, + 482, + 312 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 482, + 312 + ], + "type": "text", + "content": "Fig. 3: Prompt template for LLM to pre-analyze the difficulties the agent encounters and what type of guidance information maybe useful. The color of red, blue represent multi-turn responses generated from LLM." + } + ] + } + ], + "index": 17, + "type": "text" + }, + { + "bbox": [ + 130, + 338, + 482, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 374 + ], + "type": "text", + "content": "for the difficulty in making a decision, and the helpful types of guidance. Finally, we collect the guidance information generated by LLM and cluster it to obtain the final guidance categories." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "text", + "content": "Navigation Samples Collection We first let the agent navigate around the environment with confusion scores computed. If the confusion score surpasses the pre-defined threshold at the time step " + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "text", + "content": ", which means the agent comes across difficulties in making the action prediction, then we collect the tuple " + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "inline_equation", + "content": "(I, W_{t}^{F}, W_{t}^{Tr}, A_{t})" + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "text", + "content": " of the high-level task instruction " + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "text", + "content": ", fine-grained candidate scene descriptions " + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "inline_equation", + "content": "W_{t}^{F}" + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "text", + "content": ", the trajectory description " + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "inline_equation", + "content": "W_{t}^{Tr}" + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "text", + "content": " and the ground-truth action " + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "inline_equation", + "content": "A_{t}" + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "text", + "content": " as a sample data " + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "inline_equation", + "content": "D_{\\text{sample}}" + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "text", + "content": " for analysis. Note that " + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "inline_equation", + "content": "W_{t}^{F}" + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "text", + "content": " includes " + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "text", + "content": " fine-grained scene descriptions of the selected candidate views. In total, we randomly collect 5,000 samples." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Propmt Design To make the LLM better analyze the situation that the agent meets, we elaborately designed a prompt template for multi-turn query-and-response. We first set the unified system prompt " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "P_{\\text{system}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " that requires the LLM as a helpful assistant for navigation. Then, we transform each sample data " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "D_{\\text{sample}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " into sample prompt " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "P_{\\text{sample}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " by respectively adding the prefix of \"Task instruction:\", \"Observation:\", \"History trajectory:\" and \"Correct answer:\" into " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "I, W_t^F, W_t^{Tr}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "A_t" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ". Next, we decompose the task of analyzing the challenge into two sub-tasks, the first of analyzing what difficulties the agent may meet in making decisions and the second of analyzing what types of guidance may be useful for navigation in an iterative way. Correspondingly, we use \"########Assistant: Difficulty:\" and \"########Assistant: Guidance Type:\" as the sub-task prompts " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "P_{\\text{diff}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "P_{\\text{guide}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ". We concatenate the prompts of " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "P_{\\text{system}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "P_{\\text{sample}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "P_{\\text{diff}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "P_{\\text{task1}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " and feed " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "P_{\\text{task1}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " to LLM, which will generate the response " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "R_{\\text{diff}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " about the difficulty the agent may meet. We further feed " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "P_{\\text{task2}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " to LLM, the concatenation" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Y. Qiao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 118, + 471, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 118, + 471, + 133 + ], + "spans": [ + { + "bbox": [ + 138, + 118, + 471, + 133 + ], + "type": "text", + "content": "System: You are a helpful assistant for navigation. There is an agent navigating indoors according to task instructions and current observation. At each step, the agent needs to choose one of the candidate directions to go." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 139, + 137, + 470, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 137, + 470, + 152 + ], + "spans": [ + { + "bbox": [ + 139, + 137, + 470, + 152 + ], + "type": "text", + "content": "You need to analyze what additional guidance information the agent need to make the choice. There are four types of information: location, prioritization, destination description, previous observations. You can only give one answer from these types." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 155, + 279, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 155, + 279, + 162 + ], + "spans": [ + { + "bbox": [ + 139, + 155, + 279, + 162 + ], + "type": "text", + "content": "Task instruction: Turn off the white table lamp in the bedroom" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 162, + 172, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 162, + 172, + 167 + ], + "spans": [ + { + "bbox": [ + 139, + 162, + 172, + 167 + ], + "type": "text", + "content": "Observation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 139, + 167, + 328, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 167, + 328, + 173 + ], + "spans": [ + { + "bbox": [ + 139, + 167, + 328, + 173 + ], + "type": "text", + "content": "I see candidate 1 is a white door leads to a bedroom with a bed, electric fan and table" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 139, + 173, + 320, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 173, + 320, + 178 + ], + "spans": [ + { + "bbox": [ + 139, + 173, + 320, + 178 + ], + "type": "text", + "content": "I see candidate 2 is a bathroom has a sink and the door leads to the shower room" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 139, + 178, + 326, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 178, + 326, + 185 + ], + "spans": [ + { + "bbox": [ + 139, + 178, + 326, + 185 + ], + "type": "text", + "content": "I see candidate 3 is a hallway with white walls, stairs, and a door on the second floor" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 139, + 185, + 416, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 185, + 416, + 191 + ], + "spans": [ + { + "bbox": [ + 139, + 185, + 416, + 191 + ], + "type": "text", + "content": "History trajectory: I have passed a balcony with rocking chairs and a table, then I go to a hallway with a picture of a bedroom." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 139, + 196, + 173, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 196, + 173, + 202 + ], + "spans": [ + { + "bbox": [ + 139, + 196, + 173, + 202 + ], + "type": "text", + "content": "####Assistant:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 139, + 202, + 230, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 202, + 230, + 209 + ], + "spans": [ + { + "bbox": [ + 139, + 202, + 230, + 209 + ], + "type": "text", + "content": "Guidance Type: Destination description" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 139, + 214, + 240, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 214, + 240, + 221 + ], + "spans": [ + { + "bbox": [ + 139, + 214, + 240, + 221 + ], + "type": "text", + "content": "Please give description of the destination." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 139, + 226, + 173, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 226, + 173, + 232 + ], + "spans": [ + { + "bbox": [ + 139, + 226, + 173, + 232 + ], + "type": "text", + "content": "Assistant:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 139, + 232, + 470, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 232, + 470, + 244 + ], + "spans": [ + { + "bbox": [ + 139, + 232, + 470, + 244 + ], + "type": "text", + "content": "The bedroom is a cozy and comfortable space, typically with a queen-sized bed in the center of the room. The bed is dressed with crisp, white linens and fluffy pillows." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 183, + 270, + 429, + 281 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 270, + 429, + 281 + ], + "spans": [ + { + "bbox": [ + 183, + 270, + 429, + 281 + ], + "type": "text", + "content": "Fig. 4: Prompt template for LLM to generate guidance information." + } + ] + } + ], + "index": 15, + "type": "text" + }, + { + "bbox": [ + 130, + 308, + 482, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 334 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 130, + 308, + 482, + 334 + ], + "type": "inline_equation", + "content": "P_{task1}" + }, + { + "bbox": [ + 130, + 308, + 482, + 334 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 308, + 482, + 334 + ], + "type": "inline_equation", + "content": "R_{diff}" + }, + { + "bbox": [ + 130, + 308, + 482, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 308, + 482, + 334 + ], + "type": "inline_equation", + "content": "P_{guide}" + }, + { + "bbox": [ + 130, + 308, + 482, + 334 + ], + "type": "text", + "content": " and we will get the response " + }, + { + "bbox": [ + 130, + 308, + 482, + 334 + ], + "type": "inline_equation", + "content": "R_{guide}" + }, + { + "bbox": [ + 130, + 308, + 482, + 334 + ], + "type": "text", + "content": " about what types of guidance may be useful for navigation. The prompt template is depicted in Figure 3." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 358, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 552 + ], + "type": "text", + "content": "Guidance-Type Clustering After the above-mentioned operations, we obtain a collection of the guidance response " + }, + { + "bbox": [ + 130, + 358, + 482, + 552 + ], + "type": "inline_equation", + "content": "R_{guide}" + }, + { + "bbox": [ + 130, + 358, + 482, + 552 + ], + "type": "text", + "content": ". We first use Sentence-BERT [33] to embed each " + }, + { + "bbox": [ + 130, + 358, + 482, + 552 + ], + "type": "inline_equation", + "content": "R_{guide}" + }, + { + "bbox": [ + 130, + 358, + 482, + 552 + ], + "type": "text", + "content": " and get the corresponding embedding " + }, + { + "bbox": [ + 130, + 358, + 482, + 552 + ], + "type": "inline_equation", + "content": "E_{guide}" + }, + { + "bbox": [ + 130, + 358, + 482, + 552 + ], + "type": "text", + "content": ". Then we employ the silhouette coefficient [15] to determine the optimal number of clusters for K-Means clustering. The silhouette coefficient is an indicator to evaluate the quality of clustering. It combines the two factors of clustering cohesion and separation, calculates a silhouette coefficient value for each sample, and then calculates a silhouette coefficient value through all samples. By calculating the average silhouette coefficient under different cluster numbers " + }, + { + "bbox": [ + 130, + 358, + 482, + 552 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 358, + 482, + 552 + ], + "type": "text", + "content": ", the clustering quality under each cluster number can be evaluated. In general, the number of clusters with the highest average silhouette coefficient is considered the optimal number of clusters because it provides the best balance of cohesion and separation. And clustering result shows that the best number of clusters is seven. In addition, after analysis based on the practical situation, we found that some of the guidance content is similar, and some guidance is not executable, such as \"Room Layout\", which cannot be obtained from the environment. Thus, we manually merge these duplicate types, remove infeasible types, and finally reserve four applicable categories:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 564, + 480, + 652 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 564, + 457, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 564, + 457, + 575 + ], + "spans": [ + { + "bbox": [ + 138, + 564, + 457, + 575 + ], + "type": "text", + "content": "- Location: information about the agent's current position and the destination." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 578, + 480, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 601 + ], + "type": "text", + "content": "- Prioritization: information about the urgent task that needs to be completed based on an assessment of agent progress." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 603, + 480, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 603, + 480, + 625 + ], + "spans": [ + { + "bbox": [ + 138, + 603, + 480, + 625 + ], + "type": "text", + "content": "- Destination Description: detailed information about the target location, providing more cues to reach its destination." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 628, + 480, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 628, + 480, + 652 + ], + "spans": [ + { + "bbox": [ + 138, + 628, + 480, + 652 + ], + "type": "text", + "content": "- Previous Observations: information about the agent's navigation trajectory, lets the agent remember history information." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 195, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 195, + 91, + 447, + 102 + ], + "type": "text", + "content": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 306, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 306, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 306, + 129 + ], + "type": "text", + "content": "3.6 LLM Guides the Agent On-the-Fly" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 481, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 481, + 186 + ], + "type": "text", + "content": "After the pre-analysis of the agent's navigation, we employ the analysis result and a new prompt to ask LLM to aid the agent on the fly when it is confused to make decisions. We first illustrate the modification of the prompt for the on-the-fly guidance. Then, we show how LLM guides the agent with different types of guidance information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 314 + ], + "type": "text", + "content": "Prompt Design We also designed the prompt template to ask the LLM to analyze the agent's current situation and decide which guidance type from the selected four categories might be useful in that situation. This prompt template is similar to the pre-analysis prompt template, but with some modifications to guide agent navigation. The full prompt template is shown in Figure 4. First, we modify the system prompt " + }, + { + "bbox": [ + 132, + 205, + 481, + 314 + ], + "type": "inline_equation", + "content": "P_{system}" + }, + { + "bbox": [ + 132, + 205, + 481, + 314 + ], + "type": "text", + "content": " by asking the LLM to choose which guidance information is useful from the four pre-selected categories instead of analyzing the difficulties and generating unlimited guidance information. Second, we delete the ground truth action in the sample prompt " + }, + { + "bbox": [ + 132, + 205, + 481, + 314 + ], + "type": "inline_equation", + "content": "P_{sample}" + }, + { + "bbox": [ + 132, + 205, + 481, + 314 + ], + "type": "text", + "content": ". At last, the task prompt is revised to \"########Assistant: Guidance Type:\"." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 333, + 481, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 481, + 513 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 481, + 513 + ], + "type": "text", + "content": "Integrating Guidance with Instruction During the navigation of the agent, when the confusion score " + }, + { + "bbox": [ + 132, + 333, + 481, + 513 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 132, + 333, + 481, + 513 + ], + "type": "text", + "content": " surpasses the pre-defined threshold " + }, + { + "bbox": [ + 132, + 333, + 481, + 513 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 132, + 333, + 481, + 513 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 132, + 333, + 481, + 513 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 333, + 481, + 513 + ], + "type": "text", + "content": ", it will ask the LLM for guidance. With the help of the above-mentioned prompts, the LLM can perceive the environment and trajectory of the agent, analyze the situation, and select the guidance type from the pre-selected four categories. For different types of guidance, the LLM will generate different guidance information " + }, + { + "bbox": [ + 132, + 333, + 481, + 513 + ], + "type": "inline_equation", + "content": "I_{t}^{G}" + }, + { + "bbox": [ + 132, + 333, + 481, + 513 + ], + "type": "text", + "content": " to assist the navigation. Specifically, the LLM will generate detailed content about the type of \"Location\", \"Prioritization\" and \"Destination Description\" according to its internalized world knowledge. For these three types of guidance, the generation process is also interactive, which first generates the type of information and then generates the details according to the type of information. Such a two-step process ensures the quality of generated guidance. For the type of \"Previous Observations\", we re-use the historical trajectory as the guidance information. Some examples of generated guidance information can be found in Figure 7. Finally, we append the guidance information " + }, + { + "bbox": [ + 132, + 333, + 481, + 513 + ], + "type": "inline_equation", + "content": "I_{t}^{G}" + }, + { + "bbox": [ + 132, + 333, + 481, + 513 + ], + "type": "text", + "content": " to the high-level instruction " + }, + { + "bbox": [ + 132, + 333, + 481, + 513 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 132, + 333, + 481, + 513 + ], + "type": "text", + "content": ", and then the agent will navigate under the guidance of the integrated instruction " + }, + { + "bbox": [ + 132, + 333, + 481, + 513 + ], + "type": "inline_equation", + "content": "I_{t}^{Inter}" + }, + { + "bbox": [ + 132, + 333, + 481, + 513 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 533, + 214, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 214, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 214, + 546 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 559, + 192, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 192, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 192, + 572 + ], + "type": "text", + "content": "4.1 Settings" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 581, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 481, + 665 + ], + "type": "text", + "content": "Datasets We evaluate our method on two VLN tasks: REVERIE and CVDN-target, which are all based upon the MP3D indoor environments. Compared to Room-to-Room (R2R) task, these instructions are high-level and concise, such as \"Bring me the swivel chair from the office\". CVDN-target is our proposed VLN setup similar to REVERIE. The instruction contains a target and conversation history. Here we remove the conversation history and only keep the information containing the target as the instruction, such as \"To find a book in the room.\"" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Y. Qiao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 136, + 471, + 358 + ], + "blocks": [ + { + "bbox": [ + 180, + 114, + 432, + 125 + ], + "lines": [ + { + "bbox": [ + 180, + 114, + 432, + 125 + ], + "spans": [ + { + "bbox": [ + 180, + 114, + 432, + 125 + ], + "type": "text", + "content": "Table 1: Comparison with the state-of-the-art methods on REVERIE." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 136, + 471, + 358 + ], + "lines": [ + { + "bbox": [ + 143, + 136, + 471, + 358 + ], + "spans": [ + { + "bbox": [ + 143, + 136, + 471, + 358 + ], + "type": "table", + "html": "
MethodsVal UnseenTest Unseen
NavigationGroundingNavigationGrounding
TLOSR↑SR↑SPL↑RGS↑RGSPL↑TLOSR↑SR↑SPL↑RGS↑RGSPL↑
Human------21.1886.8381.5153.6677.8451.44
RecBERT [13]16.7835.0230.6724.9018.7715.2715.8632.9129.6123.9916.5013.51
Airbert [11]18.7134.5127.8921.8818.2314.1817.9134.2030.2823.6116.8313.28
HAMT [5]14.0836.8432.9530.2018.9217.2813.6233.4130.4026.6714.8813.08
VLN-PETL [32]14.4737.0331.8127.6718.2615.9614.0036.0630.8326.7315.1313.03
LANA [39]16.2838.5434.0029.2619.0316.1816.7536.4133.5026.8917.5314.25
HOP+ [30]14.5740.0436.0731.1322.4919.3315.1735.8133.8228.2420.2016.86
DUET [7]22.1151.0746.9833.7332.1523.0321.3056.9152.5136.0631.8822.06
Lily [22]21.8753.7148.1134.4332.1523.4321.9460.5154.3237.3432.0221.94
BEVBert [1]-56.4051.7836.3734.7124.44-57.2652.8136.4132.0622.09
AutoVLN [6]-62.1455.8940.8536.5826.76-62.3055.1738.8832.2322.68
AZHP [10]22.3253.6548.3136.6334.0025.79-55.3151.5735.8532.2522.44
BSG [24]24.7158.0552.1235.5935.3624.2422.9062.8356.4538.7033.1522.34
KERM [20]21.8555.2150.4435.3834.5124.4517.3257.5852.4339.2132.3923.64
ScaleVLN [42]--56.9741.8435.7626.05--56.1339.5232.5322.78
GRidMM [41]23.2057.4851.3736.4734.5724.5619.9759.5553.1336.6034.8723.45
MiC [31]20.6462.3756.9743.6037.5228.7218.1162.4055.7441.9735.2526.17
VLN-Copilot21.8962.6257.4043.6338.8829.7521.7263.2657.8142.3236.5626.55
", + "image_path": "7c72e6d02efef929d3d7376ff8df33aa004c8aa3cbb174279b50452aedd9fe1f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "content": "Evaluation Metrics We utilize the standard evaluation metrics to evaluate our method, we report Success Rate (SR), Success weighted by Path Length (SPL) [2], and Goal Progress (GP) as the main navigation metrics, Remote Grounding Success rate (RGS) and RGS weighted by Path Length (RGSPL) as the main object grounding metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 447, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 567 + ], + "type": "text", + "content": "Implementation Details We implement our proposed method using PyTorch. The model is trained on one single NVIDIA RTX 4090 GPU. For REVERIE, we fine-tuned the pre-trained model from [6] with a batch size of 4 for 100k iterations. For the CVDN-target, we fine-tuned the pre-trained model from [7] with a batch size of 4 for 50k iterations. AdamW [26] optimizer is adopted and the learning rate is set to " + }, + { + "bbox": [ + 130, + 447, + 482, + 567 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 130, + 447, + 482, + 567 + ], + "type": "text", + "content": ". The threshold " + }, + { + "bbox": [ + 130, + 447, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 447, + 482, + 567 + ], + "type": "text", + "content": " is set as 1. For the baseline agent, we use the recent state-of-the-art model of AutoVLN [6]. When collecting data for pre-analysis, we use two agents based on pre-trained AutoVLN [6] and HAMT [5] to avoid agent bias. LLM is integrated into the training process, we use the open-source Vicuna-7B-1.5 model [43], which is fine-tuned from Llama 2 [38]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 585, + 342, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 585, + 342, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 585, + 342, + 597 + ], + "type": "text", + "content": "4.2 Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 484, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 484, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 484, + 666 + ], + "type": "text", + "content": "REVERIE As shown in Table 1, our method achieves competitive performance on the REVERIE benchmark compared to previous methods, such as AutoVLN [6], Mic [31] and ScaleVLN [42]. Particularly, compared with the SoTA method MiC [31], our method outperforms Mic in terms of the object grounding metric RGSPL with " + }, + { + "bbox": [ + 130, + 605, + 484, + 666 + ], + "type": "inline_equation", + "content": "1.36\\%" + }, + { + "bbox": [ + 130, + 605, + 484, + 666 + ], + "type": "text", + "content": " and RGSPL with " + }, + { + "bbox": [ + 130, + 605, + 484, + 666 + ], + "type": "inline_equation", + "content": "1.03\\%" + }, + { + "bbox": [ + 130, + 605, + 484, + 666 + ], + "type": "text", + "content": " on the Val Unseen split. It is worth noting that compared with" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 196, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 196, + 91, + 447, + 102 + ], + "type": "text", + "content": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 146, + 240, + 198 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 256, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 256, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 256, + 137 + ], + "type": "text", + "content": "Table 2: Comparison on CVDN-target Validation Unseen." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 146, + 240, + 198 + ], + "lines": [ + { + "bbox": [ + 151, + 146, + 240, + 198 + ], + "spans": [ + { + "bbox": [ + 151, + 146, + 240, + 198 + ], + "type": "table", + "html": "
MethodsTLGP↑
HAMT† [5]37.042.74
DUET† [7]78.533.70
AutoVLN† [6]61.973.98
VLN-Copilot40.384.47
", + "image_path": "30f9fadd9fda3c00cad1028e05be5c7d76da48ad65e38599db0b448f8a8eb543.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 216, + 256, + 237 + ], + "lines": [ + { + "bbox": [ + 132, + 216, + 256, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 256, + 237 + ], + "type": "text", + "content": "Table 4: Ablation of different threshold " + }, + { + "bbox": [ + 132, + 216, + 256, + 237 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 132, + 216, + 256, + 237 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 138, + 248, + 253, + 281 + ], + "blocks": [ + { + "bbox": [ + 138, + 248, + 253, + 281 + ], + "lines": [ + { + "bbox": [ + 138, + 248, + 253, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 248, + 253, + 281 + ], + "type": "table", + "html": "
τ0.811.21.4
SPL↑42.7843.6343.6842.55
RGSPL↑28.3529.7528.1728.42
", + "image_path": "dddaa0e6f839e5a2f0bb53541b9dc1b15727d1ad7a0d7f17b161c8c3752f1bb4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 266, + 141, + 481, + 194 + ], + "blocks": [ + { + "bbox": [ + 305, + 120, + 440, + 131 + ], + "lines": [ + { + "bbox": [ + 305, + 120, + 440, + 131 + ], + "spans": [ + { + "bbox": [ + 305, + 120, + 440, + 131 + ], + "type": "text", + "content": "Table 3: Ablation of LLM Analyzer." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 266, + 141, + 481, + 194 + ], + "lines": [ + { + "bbox": [ + 266, + 141, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 266, + 141, + 481, + 194 + ], + "type": "table", + "html": "
MethodNavigationGrounding
TLOSR↑SR↑SPL↑RGS↑RGSPL↑
No-Selection (Baseline)21.9657.8252.7740.1534.7626.37
All-Inclusive20.7059.7655.5041.3636.4727.33
Random-Selection20.1558.0253.7942.1635.3027.08
LLM-Selection (Ours)21.8962.6257.4043.6338.8829.75
", + "image_path": "0db18d108cf36fa8d6491d5e79f271fb56cb06f78a8db04a6b98b83a6a0a8d0a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 266, + 230, + 481, + 288 + ], + "blocks": [ + { + "bbox": [ + 276, + 209, + 468, + 220 + ], + "lines": [ + { + "bbox": [ + 276, + 209, + 468, + 220 + ], + "spans": [ + { + "bbox": [ + 276, + 209, + 468, + 220 + ], + "type": "text", + "content": "Table 5: Ablation study of the guidance information." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 266, + 230, + 481, + 288 + ], + "lines": [ + { + "bbox": [ + 266, + 230, + 481, + 288 + ], + "spans": [ + { + "bbox": [ + 266, + 230, + 481, + 288 + ], + "type": "table", + "html": "
MethodNavigationGrounding
TLOSR↑SR↑SPL↑RGS↑RGSPL↑
Baseline21.9657.8252.7740.1534.7626.37
+ Location20.9762.2056.8042.9737.7428.46
+ Prioritization20.4859.4754.7041.7435.8627.40
+ Destination Description19.2159.5953.9941.9134.9927.11
+ Previous Observations19.2257.9653.7342.8335.6427.12
", + "image_path": "39376dcb91efb1728639d8a486844658893c41f478321b010ea6419e56f4f530.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 309, + 481, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 309, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 132, + 309, + 481, + 334 + ], + "type": "text", + "content": "ScaleVLN [42], which collects 4.9M instruction-trajectory pairs to train the model, our method still achieves better performance (2.8% on SPL and 3.77% on RGSPL)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 350, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 434 + ], + "type": "text", + "content": "CVDN-target As shown in Table 2, we compare VLN-Copilot with three main-stream VLN models HAMT [5], DUET [7], and AutoVLN [6] on the CVDN-target dataset. For a fair comparison, both models are trained on the CVDN-target dataset. " + }, + { + "bbox": [ + 130, + 350, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 350, + 482, + 434 + ], + "type": "text", + "content": " represents the reproduction results of these methods. It highlights the effectiveness of VLN-Copilot, especially in terms of GP metrics, where it outperforms all other methods. Although TL is slightly higher than HAMT, its significantly improved GP emphasizes its overall advantage in handling unseen validation data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 450, + 222, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 222, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 222, + 462 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 469, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 469, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 469, + 482, + 625 + ], + "type": "text", + "content": "Effect of LLM Analyzer To evaluate the ability of LLM for guidance type selection, we conduct an ablation study, there are four settings, as follows: (1) No-Selection: our baseline, which means the agent does not use any guidance information; (2) All-Inclusive: the agent lets the LLM generate all types of guidance and integrates this information; (3) Random-Selection: the agent randomly selects one of four guidance types and lets the LLM generate corresponding guidance information; (4) LLM-Selection: our method, the LLM outputs guidance type and generates the information. As illustrated in Table 3, we can see that the LLM-Selection method achieved the highest navigation main metric spl with a value of " + }, + { + "bbox": [ + 130, + 469, + 482, + 625 + ], + "type": "inline_equation", + "content": "43.63\\%" + }, + { + "bbox": [ + 130, + 469, + 482, + 625 + ], + "type": "text", + "content": ", followed by the Random-Selection method with a score of " + }, + { + "bbox": [ + 130, + 469, + 482, + 625 + ], + "type": "inline_equation", + "content": "42.16\\%" + }, + { + "bbox": [ + 130, + 469, + 482, + 625 + ], + "type": "text", + "content": ". This suggests that blindly combining all types of guidance information is not the best option, probably due to the redundancy of excessively detailed information. Compared to the Baseline, the LLM-Selection method improves the SR by " + }, + { + "bbox": [ + 130, + 469, + 482, + 625 + ], + "type": "inline_equation", + "content": "4.63\\%" + }, + { + "bbox": [ + 130, + 469, + 482, + 625 + ], + "type": "text", + "content": ", SPL by " + }, + { + "bbox": [ + 130, + 469, + 482, + 625 + ], + "type": "inline_equation", + "content": "3.48\\%" + }, + { + "bbox": [ + 130, + 469, + 482, + 625 + ], + "type": "text", + "content": ", RGS by " + }, + { + "bbox": [ + 130, + 469, + 482, + 625 + ], + "type": "inline_equation", + "content": "4.12\\%" + }, + { + "bbox": [ + 130, + 469, + 482, + 625 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 469, + 482, + 625 + ], + "type": "inline_equation", + "content": "3.38\\%" + }, + { + "bbox": [ + 130, + 469, + 482, + 625 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": "Different Threshold of Confusion Score As shown in Table 4, we conduct an ablation study on threshold " + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": " in the confusion score. Although when " + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\tau = 1" + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": ", SPL is slightly lower" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Y. Qiao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 175, + 114, + 306, + 214 + ], + "blocks": [ + { + "bbox": [ + 175, + 114, + 306, + 214 + ], + "lines": [ + { + "bbox": [ + 175, + 114, + 306, + 214 + ], + "spans": [ + { + "bbox": [ + 175, + 114, + 306, + 214 + ], + "type": "image", + "image_path": "5805db1901ab93fa70dd35ef66d2b0c4962f877a8db9c4fd3546dd4f67de78a3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 224, + 285, + 236 + ], + "lines": [ + { + "bbox": [ + 196, + 224, + 285, + 236 + ], + "spans": [ + { + "bbox": [ + 196, + 224, + 285, + 236 + ], + "type": "text", + "content": "Fig.5:Confusion Score" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 438, + 213 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 438, + 213 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 438, + 213 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 438, + 213 + ], + "type": "image", + "image_path": "befe241a5f3078077a5b31fed9fbf0dc67967eb321a00f9072657552949ca5ee.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 330, + 224, + 414, + 236 + ], + "lines": [ + { + "bbox": [ + 330, + 224, + 414, + 236 + ], + "spans": [ + { + "bbox": [ + 330, + 224, + 414, + 236 + ], + "type": "text", + "content": "Fig. 6: Learning Curve" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 260, + 479, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 479, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 479, + 283 + ], + "type": "text", + "content": "than " + }, + { + "bbox": [ + 130, + 260, + 479, + 283 + ], + "type": "inline_equation", + "content": "\\tau = 1.2" + }, + { + "bbox": [ + 130, + 260, + 479, + 283 + ], + "type": "text", + "content": ", while RGSPL reaches a higher value, so considering the two indicators of SPL and RGSPL, we set " + }, + { + "bbox": [ + 130, + 260, + 479, + 283 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 260, + 479, + 283 + ], + "type": "text", + "content": " to 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 302, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 482, + 445 + ], + "type": "text", + "content": "Effect of Guidance To validate the effect of different types of guidance information, we conducted an ablation study that only adds one type of information. As shown in Table 5, adding each type of the four categories of guidance information contributes to performance improvement. As shown in Line 1, adding location information contributes mostly, which shows that letting the agent know its own location and target location is of great help. As shown in Line 4, adding previous observations information, the agent's SPL increased significantly " + }, + { + "bbox": [ + 130, + 302, + 482, + 445 + ], + "type": "inline_equation", + "content": "(2.68\\uparrow)" + }, + { + "bbox": [ + 130, + 302, + 482, + 445 + ], + "type": "text", + "content": ", but the SR increased slightly " + }, + { + "bbox": [ + 130, + 302, + 482, + 445 + ], + "type": "inline_equation", + "content": "(0.96\\uparrow)" + }, + { + "bbox": [ + 130, + 302, + 482, + 445 + ], + "type": "text", + "content": ". This may be attributed to the agent's ability to reference historical trajectories, thereby gaining a better understanding of the navigation environment. This, in turn, helps to avoid redundant observations and improve navigation efficiency. Destination Description and Previous Observations reduce TL from 21.96 to 19.21 and 19.22, which shows such two guidance can help the agent reach the destination faster." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 464, + 482, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 513 + ], + "type": "text", + "content": "Trends of Confusion Score We plot the agent's cumulative confusion score at each time step every 5k training iterations. As shown in Figure 5, the trend of confusion score is decreasing, especially in the model using VLN-Copilot, the curve is steeper, illustrating that our model could reduce the agent's confusion degree in navigation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 531, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 578 + ], + "type": "text", + "content": "Learning Curves We compared the learning curves between methods with and without VLN-Copilot during training. As shown in Figure 6, our VLN-Copilot converges faster than the model without Copilot and can achieve better performance (i.e. higher SPL value) in fewer iterations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 597, + 239, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 239, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 239, + 609 + ], + "type": "text", + "content": "4.4 Qualitative Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "We present some examples to see the qualitative results of the LLM-generated guidance. As shown in Figure 7, for the example on the left, LLM selects \"Destination Description\" as the guidance type. Since the guidance information contains a description of the destination and matches candidate direction 2, it is easier to select the answer. For" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 196, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 196, + 91, + 447, + 102 + ], + "type": "text", + "content": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 114, + 304, + 289 + ], + "blocks": [ + { + "bbox": [ + 145, + 114, + 304, + 289 + ], + "lines": [ + { + "bbox": [ + 145, + 114, + 304, + 289 + ], + "spans": [ + { + "bbox": [ + 145, + 114, + 304, + 289 + ], + "type": "image", + "image_path": "67dc9fb7d2011b9cc71af352c796d92b4d2d5bab7f9c01fc5ce20d4f37063bb4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 308, + 399, + 319 + ], + "lines": [ + { + "bbox": [ + 214, + 308, + 399, + 319 + ], + "spans": [ + { + "bbox": [ + 214, + 308, + 399, + 319 + ], + "type": "text", + "content": "Fig. 7: Examples of the LLM-generated Guidance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 114, + 470, + 289 + ], + "blocks": [ + { + "bbox": [ + 306, + 114, + 470, + 289 + ], + "lines": [ + { + "bbox": [ + 306, + 114, + 470, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 470, + 289 + ], + "type": "image", + "image_path": "4fd149b0527292061a4b4d3d7d387c8e2481ef87bbaccfcbabce32cf2a1c673a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 342, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 415 + ], + "type": "text", + "content": "the example on the right, since Candidate 2 and Candidate 3 are both in the kitchen, and the task is to go to the kitchen, it looks like each candidate is similar, making the choice difficult. While LLM predicts that the guidance type is \"location\" and provides the agent's current location and destination, we find that the agent has reached the target location and therefore should stop rather than select a direction to go. More results can be found in supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 432, + 305, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 432, + 305, + 445 + ], + "spans": [ + { + "bbox": [ + 131, + 432, + 305, + 445 + ], + "type": "text", + "content": "5 Limitations and Future Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 456, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 529 + ], + "type": "text", + "content": "In this work, we utilize Large Language Models (LLMs) to analyze the navigation environment by transforming the dynamic environment into textual descriptions. This practice may be sub-optimal for environmental perception. Our future work will focus on how to effectively apply Multimodal Large Language Models (MLLMs) such as LLaVA [23] and Video-LLaVA [21] to perceive the navigation environment and provide guidance information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 545, + 211, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 545, + 211, + 558 + ], + "spans": [ + { + "bbox": [ + 131, + 545, + 211, + 558 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "In this work, we propose a method named VLN-Copilot that enables the agent to proactively ask for help and lets the LLM act as a co-pilot to assist the agent in completing the coarse-grained VLN task. We first introduce a confusion score calculation mechanism that allows the agent to actively query the LLM when it feels uncertain about the direction of choice. Then, we use LLM to analyze the challenges faced by the agent, and then generate appropriate guidance information for different situations that the agent is in. We conduct extensive experiments on two coarse-grained VLN tasks, including quantitative and qualitative experiments, demonstrating our method's effectiveness." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Y. Qiao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 234, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 234, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 234, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 481, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 481, + 166 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 481, + 166 + ], + "type": "text", + "content": "We thank Mr. Xiangyu Shi for the help in supplying 2 extra NVIDIA RTX 4090 GPUs for experiments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 190, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 190, + 198 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 190, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 211, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 211, + 481, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 481, + 233 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 481, + 233 + ], + "type": "text", + "content": "1. An, D., Qi, Y., Li, Y., Huang, Y., Wang, L., Tan, T., Shao, J.: Bevbert: Multimodal map pre-training for language-guided navigation. In: ICCV. pp. 2737-2748 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 234, + 481, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 481, + 266 + ], + "type": "text", + "content": "2. Anderson, P., Chang, A.X., Chaplot, D.S., Dosovitskiy, A., Gupta, S., Koltun, V., Kosecka, J., Malik, J., Mottaghi, R., Savva, M., Zamir, A.R.: On evaluation of embodied navigation agents. CoRR abs/1807.06757 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 267, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 481, + 300 + ], + "type": "text", + "content": "3. Anderson, P., Wu, Q., Teney, D., Bruce, J., Johnson, M., Sünderhauf, N., Reid, I.D., Gould, S., van den Hengel, A.: Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In: CVPR. pp. 3674-3683 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 481, + 332 + ], + "type": "text", + "content": "4. Chen, H., Suhr, A., Misra, D., Snavely, N., Artzi, Y.: TOUCHDOWN: natural language navigation and spatial reasoning in visual street environments. In: CVPR. pp. 12538-12547 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 334, + 481, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 334, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 334, + 481, + 355 + ], + "type": "text", + "content": "5. Chen, S., Guhur, P., Schmid, C., Laptev, I.: History aware multimodal transformer for vision-and-language navigation. In: NeurIPS. pp. 5834-5847 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 356, + 481, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 356, + 481, + 377 + ], + "spans": [ + { + "bbox": [ + 138, + 356, + 481, + 377 + ], + "type": "text", + "content": "6. Chen, S., Guhur, P.L., Tapaswi, M., Schmid, C., Laptev, I.: Learning from unlabeled 3d environments for vision-and-language navigation. In: ECCV (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 378, + 481, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 378, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 378, + 481, + 399 + ], + "type": "text", + "content": "7. Chen, S., Guhur, P.L., Tapaswi, M., Schmid, C., Laptev, I.: Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In: CVPR (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 400, + 481, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 400, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 400, + 481, + 422 + ], + "type": "text", + "content": "8. Fan, Y., Chen, W., Jiang, T., Zhou, C., Zhang, Y., Wang, X.E.: Aerial vision-and-dialog navigation. In: Findings of the Association for Computational Linguistics: ACL 2023 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 422, + 481, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 481, + 443 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 481, + 443 + ], + "type": "text", + "content": "9. Feng, W., Fu, T.J., Lu, Y., Wang, W.Y.: ULN: Towards underspecified vision-and-language navigation. In: EMNLP. pp. 6394-6412 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 445, + 481, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 481, + 466 + ], + "type": "text", + "content": "0. Gao, C., Peng, X., Yan, M., Wang, H., Yang, L., Ren, H., Li, H., Liu, S.: Adaptive zone-aware hierarchical planner for vision-language navigation. In: CVPR. pp. 14911-14920 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 467, + 481, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 481, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 481, + 488 + ], + "type": "text", + "content": "1. Guhur, P., Tapaswi, M., Chen, S., Laptev, I., Schmid, C.: Airbert: In-domain pretraining for vision-and-language navigation. In: ICCV. pp. 1634-1643 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 489, + 481, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 489, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 489, + 481, + 510 + ], + "type": "text", + "content": "2. Hong, Y., Opazo, C.R., Wu, Q., Gould, S.: Sub-instruction aware vision-and-language navigation. In: Webber, B., Cohn, T., He, Y., Liu, Y. (eds.) EMNLP. pp. 3360-3376 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 511, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 511, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 511, + 481, + 533 + ], + "type": "text", + "content": "3. Hong, Y., Wu, Q., Qi, Y., Opazo, C.R., Gould, S.: A recurrent vision-and-language BERT for navigation. In: CVPR. pp. 1643-1653 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 533, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 481, + 555 + ], + "type": "text", + "content": "4. Jain, V., Magalhães, G., Ku, A., Vaswani, A., Ie, E., Baldridge, J.: Stay on the path: Instruction fidelity in vision-and-language navigation. pp. 1862-1872 (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 555, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 576 + ], + "type": "text", + "content": "5. Kaufman, L., Rousseeuw, P.J.: Finding groups in data: an introduction to cluster analysis. John Wiley & Sons (2009)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 578, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 609 + ], + "type": "text", + "content": "6. Krantz, J., Wijmans, E., Majumdar, A., Batra, D., Lee, S.: Beyond the nav-graph: Vision-and-language navigation in continuous environments. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J. (eds.) ECCV. pp. 104-120 (2020)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "type": "text", + "content": "7. Ku, A., Anderson, P., Patel, R., Ie, E., Baldridge, J.: Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In: EMNLP. pp. 4392-4412 (2020)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "8. Li, J., Tan, H., Bansal, M.: Envideit: Environment editing for vision-and-language navigation. In: CVPR. pp. 15386-15396 (2022)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 195, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 195, + 91, + 447, + 102 + ], + "type": "text", + "content": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "text", + "content": "19. Li, J., Li, D., Xiong, C., Hoi, S.C.H.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML (2022), https://api.sementicscholar.org/CorpusID:246411402" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "text", + "content": "20. Li, X., Wang, Z., Yang, J., Wang, Y., Jiang, S.: Kerm: Knowledge enhanced reasoning for vision-and-language navigation. In: CVPR. pp. 2583-2592 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "text", + "content": "21. Lin, B., Zhu, B., Ye, Y., Ning, M., Jin, P., Yuan, L.: Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "22. Lin, K., Chen, P., Huang, D., Li, T.H., Tan, M., Gan, C.: Learning vision-and-language navigation from youtube videos. In: ICCV. pp. 8317-8326 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 236 + ], + "type": "text", + "content": "23. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "text", + "content": "24. Liu, R., Wang, X., Wang, W., Yang, Y.: Bird's-eye-view scene graph for vision-language navigation. In: ICCV. pp. 10968-10980 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "25. Liu, S., Zhang, H., Qi, Y., Wang, P., Zhang, Y., Wu, Q.: Aerialvln: Vision-and-language navigation for uavs. In: ICCV. pp. 15384-15394 (October 2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 281, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 303 + ], + "type": "text", + "content": "26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR. OpenReview.net (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 303, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 481, + 336 + ], + "type": "text", + "content": "27. Nguyen, K., III, H.D.: Help, anna! visual navigation with natural multimodal assistance via retrospective curiosity-encouraging imitation learning. In: Inui, K., Jiang, J., Ng, V., Wan, X. (eds.) EMNLP. pp. 684–695. Association for Computational Linguistics (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "28. Qi, Y., Pan, Z., Hong, Y., Yang, M., van den Hengel, A., Wu, Q.: The road to know-where: An object-and-room informed sequential bert for indoor vision-language navigation. In: ICCV. pp. 1655–1664 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "29. Qi, Y., Wu, Q., Anderson, P., Wang, X., Wang, W.Y., Shen, C., van den Hengel, A.: REVERIE: remote embodied visual referring expression in real indoor environments. In: CVPR. pp. 9979-9988 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 402, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 424 + ], + "type": "text", + "content": "30. Qiao, Y., Qi, Y., Hong, Y., Yu, Z., Wang, P., Wu, Q.: Hop+: History-enhanced and order-aware pre-training for vision-and-language navigation. IEEE TPAMI (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 424, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 446 + ], + "type": "text", + "content": "31. Qiao, Y., Qi, Y., Yu, Z., Liu, J., Wu, Q.: March in chat: Interactive prompting for remote embodied referring expression. In: ICCV. pp. 15758-15767 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 446, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 468 + ], + "type": "text", + "content": "32. Qiao, Y., Yu, Z., Wu, Q.: Vln-petl: Parameter-efficient transfer learning for vision-and-language navigation. In: ICCV. pp. 15443-15452 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "text", + "content": "33. Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert-networks. In: Inui, K., Jiang, J., Ng, V., Wan, X. (eds.) EMNLP-IJCNLP. pp. 3980-3990 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 500, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 481, + 522 + ], + "type": "text", + "content": "34. Roman, H.R., Bisk, Y., Thomason, J., Celikyilmaz, A., Gao, J.: Rmm: A recursive mental model for dialog navigation. arXiv preprint arXiv:2005.00728 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 522, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 544 + ], + "type": "text", + "content": "35. Schumann, R., Zhu, W., Feng, W., Fu, T.J., Riezler, S., Wang, W.Y.: Velma: Verbalization embodiment of llm agents for vision and language navigation in street view. In: AAAI (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 544, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 577 + ], + "type": "text", + "content": "36. Song, C.H., Wu, J., Washington, C., Sadler, B.M., Chao, W.L., Su, Y.: Llm-planner: Few-shot grounded planning for embodied agents with large language models. arXiv preprint arXiv:2212.04088 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 577, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 599 + ], + "type": "text", + "content": "37. Thomason, J., Murray, M., Cakmak, M., Zettlemoyer, L.: Vision-and-dialog navigation. In: CoRL. pp. 394-406 (2019)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 599, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 665 + ], + "type": "text", + "content": "38. Touvron, H., Martin, L., Stone, K.R., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D.M., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A.S., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I.M., Korenev, A.V., Koura, P.S., Lachaux, M.A., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A.," + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Y. Qiao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 347 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 171 + ], + "type": "text", + "content": "Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. ArXiv abs/2307.09288 (2023), https://api.sementicscholar.org/CorpusID:259950998" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 194 + ], + "type": "text", + "content": "39. Wang, X., Wang, W., Shao, J., Yang, Y.: Lana: A language-capable navigator for instruction following and generation. In: CVPR. pp. 19048-19058 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "40. Wang, X., Huang, Q., Celikyilmaz, A., Gao, J., Shen, D., Wang, Y., Wang, W.Y., Zhang, L.: Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In: CVPR. pp. 6629-6638 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "text", + "content": "41. Wang, Z., Li, X., Yang, J., Liu, Y., Jiang, S.: Gridmm: Grid memory map for vision-and-language navigation. In: ICCV. pp. 15625-15636 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "text", + "content": "42. Wang, Z., Li, J., Hong, Y., Wang, Y., Wu, Q., Bansal, M., Gould, S., Tan, H., Qiao, Y.: Scaling data generation in vision-and-language navigation. In: ICCV. pp. 12009-12020 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 302 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 302 + ], + "type": "text", + "content": "43. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 303, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 481, + 325 + ], + "type": "text", + "content": "44. Zhu, F., Liang, X., Zhu, Y., Yu, Q., Chang, X., Liang, X.: SOON: scenario oriented object navigation with graph-based exploration. In: CVPR. pp. 12689-12699 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "text", + "content": "45. Zhu*, Y., Weng*, Y., Zhu, F., Liang, X., Ye, Q., Lu, Y., jiao, J.: Self-motivated communication agent for real-world vision-dialog navigation. In: ICCV (2021)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 195, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 195, + 91, + 447, + 102 + ], + "type": "text", + "content": "LLM as Copilot for Coarse-grained Vision-and-Language Navigation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LLM as Dataset Analyst_ Subpopulation Structure Discovery with Large Language Model/cfcf5c50-f6b7-4af0-a369-8bdd05a30c95_content_list.json b/2024/LLM as Dataset Analyst_ Subpopulation Structure Discovery with Large Language Model/cfcf5c50-f6b7-4af0-a369-8bdd05a30c95_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2205be3f48ab922e80647ce4b2dba51beb0cf4bf --- /dev/null +++ b/2024/LLM as Dataset Analyst_ Subpopulation Structure Discovery with Large Language Model/cfcf5c50-f6b7-4af0-a369-8bdd05a30c95_content_list.json @@ -0,0 +1,1746 @@ +[ + { + "type": "text", + "text": "LLM as Dataset Analyst: Subpopulation Structure Discovery with Large Language Model", + "text_level": 1, + "bbox": [ + 222, + 141, + 781, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yulin Luo $^{1*}$ , Ruichuan An $^{1,2*}$ , Bocheng Zou $^{1,3}$", + "bbox": [ + 300, + 215, + 673, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiming Tang $^{1,4}$ , Jiaming Liu $^{1}$ , and Shanghang Zhang $^{1\\dagger}$", + "bbox": [ + 300, + 229, + 699, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Laboratory of Multimedia Information Processing,", + "bbox": [ + 292, + 258, + 705, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Computer Science, Peking University", + "bbox": [ + 344, + 273, + 656, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Xi'an Jiaotong University $^{3}$ University of Wisconsin-Madison", + "bbox": [ + 297, + 286, + 704, + 300 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ National University of Singapore", + "bbox": [ + 388, + 300, + 614, + 314 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yulin@stu.pku.edu.cn", + "bbox": [ + 421, + 315, + 578, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The distribution of subpopulations is an important property hidden within a dataset. Uncovering and analyzing the subpopulation distribution within datasets provides a comprehensive understanding of the datasets, standing as a powerful tool beneficial to various downstream tasks, including Dataset Subpopulation Organization, Subpopulation Shift, and Slice Discovery. Despite its importance, there has been no work that systematically explores the subpopulation distribution of datasets to our knowledge. To address the limitation and solve all the mentioned tasks in a unified way, we introduce a novel concept of subpopulation structures to represent, analyze, and utilize subpopulation distributions within datasets. To characterize the structures in an interpretable manner, we propose the Subpopulation Structure Discovery with Large Language Models (SSD-LLM) framework, which employs world knowledge and instruction-following capabilities of Large Language Models (LLMs) to linguistically analyze informative image captions and summarize the structures. Furthermore, we propose complete workflows to address downstream tasks, named Task-specific Tuning, showcasing the application of the discovered structure to a spectrum of subpopulation-related tasks, including dataset subpopulation organization, subpopulation shift, and slice discovery. With the help of SSD-LLM, we can structuralize the datasets into subpopulation-level automatically, achieve average $+3.3\\%$ worst group accuracy gain compared to previous methods on subpopulation shift benchmark Waterbirds, Metashift and $\\mathrm{Nico + + }$ , and also identify more consistent slice topics with a higher model error rate of $3.95\\%$ on slice discovery task for ImageNet. The code will be available at https://llm-as-dataset-analyst.github.io/.", + "bbox": [ + 259, + 397, + 743, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Subpopulation Structure Discovery $\\cdot$ Large Language Model", + "bbox": [ + 261, + 773, + 741, + 789 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contribution.", + "bbox": [ + 218, + 810, + 370, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Author.", + "bbox": [ + 220, + 824, + 388, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2ea742e3c8f49a8bfce083b93f20818f43d97ec73e2616ebdf66973d3c795aa7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 145, + 772, + 231 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b46704d9a56276f8d906e8d11358b16c9e541043ff9eaa82c87a802589239293.jpg", + "image_caption": [ + "Fig. 1: (A) The Workflow of Subpopulation Structure Discovery with Large Language Models (SSD-LLM). SSD-LLM can further support several downstream tasks including: (B) Dataset Subpopulation Organization; (C) Subpopulation Shift; (D) Slice discovery." + ], + "image_footnote": [], + "bbox": [ + 225, + 232, + 772, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 474, + 375, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Subpopulation, defined by a set of data points that share common characteristics, is an important concept in machine learning [48]. Many tasks are subpopulation-related. For example, image clustering conditioned on text criteria [18] is to partition an image dataset into different subpopulations based on user-specified criteria, studying subpopulation shift [22, 48, 52] is to mitigate the negative impact of imbalanced subpopulation distributions in the training set on the model, slice discovery [2, 9] is aimed at identifying subpopulations model underperform.", + "bbox": [ + 212, + 507, + 782, + 612 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Summarizing the commonalities of these tasks, we find that analyzing the subpopulation distribution is the key to solving all these problems. If the subpopulation distribution can be characterized, image clustering results under different criteria are naturally obtained [18], additional images can be supplemented to rare subgroups to balance the whole dataset [8], and slices can be easily discovered by statistics error rate on validation set [2]. Despite its importance, existing work [48] lacks systematic exploration of subpopulation distribution. To adjust the issue, for the first time, we propose the concept of subpopulation structure to represent, analyze, and utilize subpopulation distributions within datasets. By definition, a subpopulation structure is a set of hierarchical relations among several subpopulations determined by certain criteria.", + "bbox": [ + 212, + 613, + 784, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Former works like Metashift [22] and $\\mathrm{NICO}++$ [52] have constructed image datasets including the subpopulation information, which organizes the images with respect to some extra attributes, and can be viewed as a class-, attribute-, subpopulation-layer structure. The problem of such a structure is ignoring", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Luo et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b14f2039cd2e15854989983efd1152ef0c4722147efc9f7e9efc03ae79b0ba20.jpg", + "image_caption": [ + "Fig. 2: Metashift has the same-level attributes Surfboard, Water, and Grass for class Dog, which is irrational due to the possible overlap. As an improvement, we take dimensions into consideration. The class Dog has dimensions including Action, Co-occurrence Object, Location, etc., and in dimension Location, it includes various attributes like Water, Grass, etc, which offers a more appropriate assignment for the samples." + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 397, + 291 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c75575809299d06f39836d5a109e823090a7b0f6b6ecfecac583a8556a3b9ea6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 143, + 781, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the category of attributes (or Dimension), leading to attribute inconsistency and confusion. To solve this issue, we introduce a class-, dimension-, attribute-, and subpopulation-layer structure. The comparison of the two structures can be seen in Figure. 2. By articulating the classification dimensions, this improved structure provides more nuanced attribute assignments.", + "bbox": [ + 212, + 411, + 784, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Identifying descriptive information within a dataset often requires large amounts of human manufacture [22, 52], which urges the need of automatic workflows to complete the subpopulation structure discovery. However, automatically identifying subpopulation structures within image datasets presents a significant challenge. The approach must be capable of extracting key information from images and summarizing essential content from extensive texts. Furthermore, it necessitates comprehensive world knowledge, enabling a broad understanding of various aspects of the datasets, including diverse categories, common attributes, and the relationships between dimensions and attributes.", + "bbox": [ + 212, + 489, + 782, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, Large Language Model (LLM) [17, 27, 46] and Multimodal Large Language Model (MLLM) [12, 25] have attracted wide attention due to their superior capacities. LLM has shown extensive world knowledge and remarkable abilities in summarization, instruction following [27], etc. MLLM extends the capabilities of LLM to handle visual inputs. By visual instruction tuning [25], MLLM can verbalize the rich information of images. Motivated by these, we propose a novel framework Subpopulation Structure Discovery with Large Language Model(SSD-LLM), illustrated in Figure. 1, to automatically uncover the structure. The core idea is to generate informative captions from images with MLLM, followed by analyzing and summarizing the subpopulation structure of datasets with LLM. Specifically, we design two elaborate prompt engineering components, Criteria Initialization and Criteria Refinement. The former utilizes a generate-and-select paradigm to summarize dimensions and attributes sequentially. The latter employs self-consistency as an indicator", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Subpopulation Structure Discovery with Large Language Model", + "bbox": [ + 305, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to evaluate and refine the criteria. After obtaining complete criteria, each image is assigned to corresponding attributes according to its caption. The final subpopulation structures can be leveraged to finish various downstream tasks with the help of our proposed Task-specific Tuning. In this work, we focus on three application scenarios, i.e. dataset subpopulation organization, subpopulation shift, and slice discovery. We validate the effectiveness of SSD-LLM on these subpopulation-related tasks. For subpopulation shift, we achieve an improvement of $+3.3\\%$ in worst group accuracy across three datasets compared to SOTA methods, and for slice discovery, we can identify more consistent slice topics with a higher model error rate of $3.95\\%$ .", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 238, + 297, + 566, + 311 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the concept of subpopulation structure to characterize subpopulation distribution in an interpretable manner for the first time.", + "- We propose class-dimension-attribute-subpopulation structure, reducing the attribute confusion of the current class-attribute-subpopulation structure.", + "- We propose Subpopulation Structure Discovery with Large Language Models (SSD-LLM) framework to uncover the underlying subpopulation structure of datasets automatically, with two elaborate prompt engineering components. Criteria Initialization and Criteria Refinement.", + "- We provide methods for Task-specific Tuning, enabling the application of the structures across a spectrum of subpopulation-related tasks, including dataset subpopulation organization, subpopulation shift, and slice discovery." + ], + "bbox": [ + 225, + 321, + 784, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 510, + 395, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Hierarchical Structure of Image Datasets", + "text_level": 1, + "bbox": [ + 215, + 541, + 599, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent research has emphasized the need to organize datasets into hierarchical structures allowing for benchmarking various downstream tasks [22, 37, 43, 52]. Metashift [22] builds a collection of 12,868 sets of images related to 410 main subjects and their contexts. NICO++ [52], Waterbirds [43], and ImageNetBG [37] also propose methods for constructing various types of hierarchical datasets. However, the construction of these hierarchical datasets often requires manual annotation, hindering automatic construction. These approaches focus on just a single dimension, such as object context in Metashift, background in Waterbirds, and ImageNetBG, while more practical scenarios may involve multiple dimensions hidden within the comprehensive visual information.", + "bbox": [ + 212, + 566, + 787, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Extract Information from Image Captions", + "text_level": 1, + "bbox": [ + 215, + 739, + 607, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent works such as ALIA [8], VeCAF [51], Bias2Text [15], and ICTC [19] explore utilizing caption models to obtain information from datasets. ALIA provides a method to augment datasets by generating variations of existing images through captioning and text-to-image models. While ALIA [8] supports dataset improvement, it lacks knowledge about attribute types, bias, or subpopulation", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Luo et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "shift existence. Bias2Text [15] detects dataset bias by transforming images into descriptive captions and keywords. However, without large language model participation, Bias2Text fails to support classification dimension selection and can only differentiate images with basic keywords. More recently, ICTC [19] enables conditional image clustering using an LLM in a straightforward manner. Although ICTC clusters images when given the criterion, it requires human-assigned text prompts. Compared to ICTC, our prompt engineering paradigm supports scalable automatic subpopulation structure dataset organization without human criteria assignment and can generate comprehensive criteria tailored to datasets for various downstream tasks.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3 LLM Prompt Engineering", + "text_level": 1, + "bbox": [ + 214, + 320, + 478, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As the popularity of LLMs has surged, prompt engineering, the process of crafting and refining prompts to guide LLMs towards desired outputs [17,46] has also shown more and more importance. Various prompt engineering methods [28,35] and principles have emerged, and researchers or engineers have explored their applications in a diverse range of downstream tasks [4,10,27,33,34]. In particular, in-context learning has emerged as a pivotal technique, validated both experimentally [7, 21] and theoretically [6, 30, 47]. This approach involves providing the LLM with context information relevant to the task at hand, enabling it to generate more accurate and relevant responses. Least-to-most prompting [53] breaks down complex tasks into smaller, more manageable steps, enhancing LLM's reasoning skill by querying the LLM with more simplified sub-questions. Self-consistency [45] proposes to ensemble multiple responses to the LLM given the same prompt to get enhanced results, suggesting that consistent responses as an indicator for correct problem solving. Self-refining [31] demonstrates that we can use LLMs to refine their outputs by themselves with careful designing of prompts. In this work, we leverage a combination of prompt engineering techniques, including in-context learning, chain-of-thought, self-consistency, and self-refining to tackle subpopulation structure discovery effectively.", + "bbox": [ + 212, + 345, + 787, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 641, + 330, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce our proposed method, subpopulation structure discovery with large language models(SSD-LLM). We describe the overall pipeline of SSD-LLM in Section 3.1, outlining how the paradigm automatically discovers the latent subpopulation structures inside the dataset. The process begins with captioning the images in the dataset with an MLLM, detailed in Section 3.2, and proceeds with Criteria Initialization with an LLM in Section 3.3. The paradigm then refines the initialized criteria through a recursive self-refinement procedure, detailed in Section 3.4. Finally, images are assigned to attributes, completing the subpopulation structure discovery process, as elaborated in Section 3.5. The section concludes with a discussion of how to apply the method to various downstream tasks, as presented in Section 3.6.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Subpopulation Structure Discovery with Large Language Model", + "bbox": [ + 305, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7e2bfa10079f1eb7e656919e3e4e6611b6d2b8f58572c9918abc14a3f134fbc4.jpg", + "image_caption": [ + "Fig. 3: Subpopulation Structure Discovery with Large Language Model (SSD-LLM). (Step 1) Multimodality Large Language Model (MLLM) extracts informative captions from images. (Step 2) LLM initializes the criteria with a sample-based generate-and-select paradigm. (Step 3) LLM refines the criteria using self-consistency as an indicator. (Step 4) LLM assigns each caption with specific attributes according to the refined criteria, uncovering the intrinsic subpopulation structures hidden in the dataset. The resulting criteria and subpopulations are used in several downstream tasks." + ], + "image_footnote": [], + "bbox": [ + 223, + 146, + 774, + 308 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 449, + 339, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To automatically discover the subpopulation structures, we propose a novel prompt engineering paradigm that effectively leverages the capabilities of both multimodal large language models (MLLMs) and large language models (LLMs). Our proposed method comprises four key steps (Figure. 3). First, we transform the images into information-rich captions that capture the main information in the images using MLLMs. Second, we employ a novel sample-based prompt engineering method to guide an LLM to produce criteria consisting of dimensions and corresponding attributes organizing the dataset. Third, we prompt the LLM to refine this generated criteria. Last, we assign all the images in the dataset to specific attributes accordingly, uncovering the intrinsic subpopulation structures in the dataset, and paving the way for further analysis about the dataset. Detailed descriptions of each step are provided below. For notations, consistent with various former works [30,46], we denote the operation of getting responses from the language models as $LLM$ and $MLLM$ , and use [,] to represent the concatenation operation of two pacts of texts.", + "bbox": [ + 212, + 474, + 787, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Caption Extraction", + "text_level": 1, + "bbox": [ + 215, + 723, + 423, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To begin our approach, we leverage the powerful image captioning capabilities of the MLLM to transform the images into informative and detailed captions. Instead of briefly describing the images, we prompt the MLLM to generate more detailed captions centered around the main subject CLS. The prompt we used in this step is stated as follow:", + "bbox": [ + 212, + 750, + 787, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$P_{1} =$ \"Describe the image of the subject CLS in detail.\"", + "bbox": [ + 240, + 824, + 645, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Luo et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/42f35fcf09f848c5e2546c225603f45735886993dac04ea59ab019fbada94f0a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Step 1 Caption ExtractionStep 2-1 Dimension Generation
Input: Dataset: Dimg, MLLMInput: Captions: C, LLM
Output: Image Captions: COutput: dimensions: Dims
1: for i in range(NumOfIterations)1: for i in range(NumOfIterations)
2: img = Dimg.sample()2: c = C/sample(NumOfSamples)
3: c = MLLM(img, P1)3: S.append(LLM([P2^1, c]))
4: C.append(c)4: end for
5: end for5: Dims = LLMSummary(S)
", + "bbox": [ + 241, + 172, + 764, + 296 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Criteria Initialization", + "text_level": 1, + "bbox": [ + 215, + 319, + 437, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To discover the hidden subpopulation structures within the dataset, we employ an LLM to delve into the information-rich captions generated in the previous step. Our objective is to identify certain criteria that effectively partition the images into several distinct subgroups. Beyond simply dividing the dataset into subgroups, we articulate the classification dimension for the partition and record all the resulting attributes generated from the classification process. Along with the class information and the resulting subpopulations, this criteria naturally form a four-layer structure, class-, dimension-, attribute-, and subpopulation-. Noticing criteria encompass multiple dimensions and their corresponding attributes, we adopt a generate-and-select paradigm with the LLM to discover the dimensions and the attributes sequentially.", + "bbox": [ + 212, + 347, + 787, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To determine the dimensions and attributes, we employ an iterative sampling approach, repeatedly prompting the LLM to propose dimensions and attributes based on batches of image captions. In each iteration, the LLM generates candidate dimensions and attributes, which are subsequently processed through an LLM summarization process. This sample-and-summarize approach effectively addresses the challenges when processing large datasets. Since the number of dimensions that can differentiate images in a dataset is relatively small, and these dimensions have an appearance in numerous images, our sample-based approach effectively identifies relevant dimensions. The prompts we used in this step are stated as follow, omitting Chain-of-thought examples for simplicity.", + "bbox": [ + 212, + 513, + 787, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$P_2^1 = \\text{\"Suggest some dimensions that can differentiate the following image captions.}$", + "bbox": [ + 215, + 665, + 784, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$P_2^2 = \\text{\"Suggest a complete criterion to differentiate the following image captions by the given dimension.}\"$", + "bbox": [ + 215, + 695, + 784, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Criteria Refinement", + "text_level": 1, + "bbox": [ + 215, + 751, + 426, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further refine the criteria and ensure its effectiveness in classifying image captions across the dataset, we implement a recursive refining process. This approach proposes a novel method for identifying image captions requiring further refinement utilizing the self-consistency of LLM responses as an indicator [45].", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Subpopulation Structure Discovery with Large Language Model", + "bbox": [ + 305, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Step 2-2 Attribute Generation", + "bbox": [ + 215, + 146, + 444, + 161 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Input: Dimensions: Dims, Captions: $C$ , Large language model: LLM", + "bbox": [ + 215, + 164, + 683, + 178 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Output: Initialized criteria: Criteria", + "bbox": [ + 215, + 179, + 472, + 191 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: for dim in Dims do", + "2: for i in range(NumOfIterations) do", + "3: $c = C$ .sample(NumOfSamples)", + "4: $S.\\mathrm{append}(\\mathrm{LLM}([P_2^2, dim, c]))$", + "5: end for", + "6: Attributes $=$ LLMSumarry(S) *list of attributes", + "7: Criteria[dim] = Attributes", + "8: $S.\\mathrm{reset}()$", + "9: end for" + ], + "bbox": [ + 223, + 193, + 581, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This choice stems from our empirical observation that if an image can be accurately classified according to a particular dimension, it should consistently be classified into the same attribute multiple times. Inconsistent responses, however, suggest that the current criteria require further refinement, either to merge redundant attributes or to include new attributes. The prompts we used in this step are stated as follow and the pseudocode is included in the appendix:", + "bbox": [ + 212, + 352, + 785, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$P_3^1 = \\text{\"Classify the caption by the criteria listed below.\"}$", + "bbox": [ + 240, + 443, + 640, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$P_{3}^{2} = \\text{\"We are unable to classify the following image caption using the provided criteria due to attributes redundancy or disappearance. If redundancy, please prune the redundant attributes. If missing, please suggest an additional attribute that would match the image caption.\"}$", + "bbox": [ + 212, + 459, + 787, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Subpopulation Assignment", + "text_level": 1, + "bbox": [ + 214, + 542, + 483, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Equipped with the comprehensive criteria, we proceed to systematically assign each image to the specific attributes of each dimension. Images assigned to the same attributes across all dimensions form distinct subgroups within the dataset, revealing the intrinsic subpopulation structures hidden within the data. These subpopulation structures can then be leveraged to perform various downstream tasks, completing our overall pipeline for employing an LLM to analyze the image dataset. The prompt we used in this step is stated as follow:", + "bbox": [ + 212, + 568, + 785, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$P_{4} =$ \"Please assign following caption to one attribute of given dimension.\"", + "bbox": [ + 238, + 674, + 781, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Step 4 Subpopulation Assignment", + "bbox": [ + 215, + 717, + 468, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Input: Captions: $C$ , Large language model: LLM, criteria: Criteria", + "bbox": [ + 215, + 736, + 673, + 751 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Output: Further assignments for each caption $c$ .", + "bbox": [ + 215, + 751, + 547, + 765 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: for $c$ in $C$ do", + "2: for [dim, Attributes] in Criteria do", + "3: c.Assign(LLM([P4,c, dim, Attributes]))", + "4: end for", + "5: end for" + ], + "bbox": [ + 223, + 766, + 519, + 833 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Luo et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f3520e9641c319c381fc8d660d62f50d3451f8644963803942cfc7492cb72cd9.jpg", + "image_caption": [ + "Fig. 4: A visualization of organised subpopulations in a dataset of cats." + ], + "image_footnote": [], + "bbox": [ + 223, + 147, + 766, + 333 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.6 Task-specific Tuning for Downstream Tasks", + "text_level": 1, + "bbox": [ + 215, + 390, + 619, + 405 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Once we have identified the criteria and subpopulations within the dataset, we can leverage this information to tackle several downstream tasks effectively. This includes organizing the subpopulations, which can reveal valuable insights into the data, such as potential subpopulation biases and the presence of long-tail attributes. In practice, the subpopulation structures can be used to improve model performance on various tasks if combined with extra operations, including handling subpopulation shifts and slice discovery.", + "bbox": [ + 212, + 416, + 787, + 521 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset Subpopulation Organization Organizing subpopulations within a dataset can provide beneficial information about the numbers of data points in each subpopulation and can reflect dataset biases and help us identify some long-tail subgroups. We conduct experiments on the task of image clustering conditioned on human-specified criteria to evaluate the quality of subpopulation organization. In specific, when organising the subpopulations of a given image dataset, we first select out the relevant dimensions and then attach attributes assigned by SSD-LLM directly to the images accordingly. Our method, SSD-LLM, automates subpopulation organization within datasets, and thus holds the potential to revolutionize the way hierarchical datasets are constructed.", + "bbox": [ + 212, + 523, + 787, + 672 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Subpopulation Shift Subpopulation shift [48] stands as a common challenge in machine learning, occurring when the proportion of some subpopulations between training and deployment changes, and is shown to be of significant influence to model performances [39, 48]. SSD-LLM, combined with image generation, offers a solution to better handle the scenarios of subpopulation shifts. In our experiments, after we apply SSD-LLM to the datasets, we collect statistics the number of images contained in each subpopulation and utilize diffusion model to generate images for underrepresented subpopulations. Subsequently, we sample attributes from the subpopulation structure for each underrepresented subpopulation and employ LLM to make complete sentences based on these words as the input prompt of a diffusion model. The diffusion model generates images augmented", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Subpopulation Structure Discovery with Large Language Model", + "bbox": [ + 305, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/6224993018e615fe55668b739eeea2e4c53df0945387539b58b6622607faa8f3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetCriterionSCAN*IC|TCOurs
Stanford 40 ActionAction0.3460.7470.817
Location0.3570.6710.705
Mood0.2760.7460.768
Place365Place0.332-0.696
PPMIMusical Instruction0.5980.9340.955
Cifar10Object0.8390.9110.921
STL10Object0.7980.9860.988
", + "bbox": [ + 313, + 143, + 689, + 271 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1: Quantitative results of Dataset Subpopulation Organization. Method labeled with * is evaluated by having a human provided ground truth labels, cause the method itself is an unsupervised learning paradigm.", + "bbox": [ + 212, + 272, + 784, + 313 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "to the image dataset which further helps to achieve balanced classes and attributes. Moreover, we propose to harness an LLM to suggest extra dimensions and attributes based on the current sets in this task for enriched subpopulation structure, generating more diverse images.", + "bbox": [ + 212, + 356, + 782, + 415 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Slice Discovery Slice discovery is a task aiming at uncovering subpopulations within a dataset where a machine learning model consistently exhibits poor performances. These subpopulations with underperformances, or slices, provide valuable insights into the model's limitations, potential biases [14], and how to improve the performances. SSD-LLM conducts slice discovery for an image dataset with the help of the assigned attributes. In detail, we first calculate the error rates on all subpopulations discovered with the SSD-LLM. Then we identify out the subpopulations with the highest error rate and use the LLM to summarize out discriptions based on the attributes of the subpopulations in the form of texts representing the slice topics, completing the task of slice discovery.", + "bbox": [ + 212, + 417, + 784, + 568 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 590, + 375, + 609 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We now present experimental results demonstrating the effectiveness of SSD-LLM. In particular, we present the main settings and results in this section and defer extra details, including various visualization result, to appendix. In our experiments, we mainly use LLaVA1.5 [26] for MLLM and GPT-4 [1] for LLM.", + "bbox": [ + 212, + 625, + 782, + 684 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We conduct experiments on Dataset Subpopulation Organization, Subpopulation Shift, and Slice Discovery. Our superior performance underscores the method's efficacy in identifying and analyzing subgroups, further affirming its utility in addressing related challenges. Moreover, it illustrates that the unified paradigm can effectively address a variety of downstream tasks.", + "bbox": [ + 212, + 685, + 782, + 760 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Evaluation on Dataset Subpopulation Organization", + "text_level": 1, + "bbox": [ + 214, + 782, + 683, + 799 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Setup The data assignment process facilitated by SSD-LLM, can be considered a form of clustering. To evaluate the quality of these identified subpopulations,", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Luo et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/551c3d42a03159f814d7fd0af7def7c604ca79065a3794096dfdf962a01662a0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeMethodAverage AccuracyWorst Group Accuracy
WaterbirdsMetashiftNico++AverageWaterbirdsMetashiftNico++Average
VanillaERM84.191.276.383.769.182.117.856.3
Subgroup Robust MethodsGroupDRO86.991.574.084.173.183.112.256.1
JTT88.991.277.585.971.282.615.656.5
LfF86.680.477.581.575.072.315.654.3
LISA89.291.475.085.277.079.018.958.3
Imbalanced LearningResample86.292.277.385.270.081.016.755.9
Reweight86.291.573.883.871.983.112.255.7
Focal89.391.673.184.771.681.016.756.4
CBLoss86.891.476.384.874.483.112.256.6
BSoftmax88.491.374.284.674.182.616.757.8
Traditional Data AugmentationMixup89.291.473.084.577.579.014.457.0
RandAug86.390.972.083.171.480.916.756.3
DiffusionClass Prompt85.991.578.085.171.382.718.557.5
Class-Attribute89.191.478.686.473.583.818.858.7
CiP88.091.178.385.873.582.419.358.4
LLM+DiffusionSSD-LLM (Ours)90.593.080.488.079.184.822.162.0
", + "bbox": [ + 222, + 143, + 777, + 319 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 2: Comparison of methods for image classification with subpopulation shifts.", + "bbox": [ + 222, + 320, + 777, + 332 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "we assess the clustering accuracy by comparing images against secondary labels that reflect subgroup attributes derived via SSD-LLM. Full information for datasets, text criterion, and model selection can be found in appendix.", + "bbox": [ + 212, + 377, + 782, + 421 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison Methods SCAN [41] is a two-stage clustering method that decouples feature learning and clustering. IC|TC [19] is a new paradigm for image clustering that supports human interaction. It utilizes the given Text Criteria to accurately control the quality of the clustering results.", + "bbox": [ + 212, + 422, + 784, + 482 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results and Analysis In Table. 1, we report the average accuracy achieved by each method based on the predefined textual criteria. When compared to IC|TC, SSD-LLM demonstrates competitive performance. It is important to note that IC|TC incorporate artificial judgment in its process, which leads to poor scalability when handling large datasets. In contrast, our approach is fully automated. We conduct visualization of our organized subpopulations, as shown in Figure. 4. We can visually observe that the pictures and dimensions are indeed consistent, indicating the effectiveness of our mining and assigning process.", + "bbox": [ + 212, + 484, + 784, + 604 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Evaluation on Subpopulation Shift", + "text_level": 1, + "bbox": [ + 214, + 628, + 545, + 645 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Setup We evaluate subpopulation shifts on three commonly used image datasets, Metashift (Cats vs Dogs), Waterbirds (Landbirds vs Waterbirds), and NICO++. We choose Average Accuracy and Worst Group Accuracy as evaluation metrics. To ensure a fair comparison, following [48], we conduct a random search of 16 trials over a joint distribution of all hyperparameters. We then use the validation set to select the best hyperparameters for each algorithm, fix them, and rerun the experiments under five different random seeds to report the final average results. To make the evaluation more realistic, we consider the model selection setting Attributes are unknown in both training and validation.", + "bbox": [ + 212, + 657, + 784, + 792 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison Methods Following recent benchmarking efforts [48], we compare SSD-LLM with several types of methods: (1) vanilla: ERM [42], (2) Subgroup Robust Methods: GroupDRO [38], LfF [32], JTT [24], LISA [49], (3) Imbalanced", + "bbox": [ + 212, + 795, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Subpopulation Structure Discovery with Large Language Model", + "bbox": [ + 305, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a6244cd2546454c418c04d2c2a6cf3dcbb68363509120b03f209e0b960ee9913.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method|CategoriesBoatBirdCarCatDogTruckTopic Error Rate
ImageNet4.330.8111.3311.140.6911.716.72
General Prompt47.8212.1143.5514.2210.1912.6523.42
GPT-Suggest57.5512.8743.5912.7116.3428.1228.53
Domino(Bert)76.2642.2654.2133.8924.5029.5443.44
B2T77.6230.0458.1736.3619.8033.4742.58
SSD-LLM (Ours)79.3145.6760.3432.9726.4839.5747.39
", + "bbox": [ + 241, + 143, + 763, + 258 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 3: Results of slice discovery on Imagenet-1K with various SDMs.", + "bbox": [ + 261, + 258, + 738, + 270 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Learning: ReSample [13], ReWeight [13], Focal [23], CBLoss [5], Bsoftmax [36], (4) Traditional Data Augmentation: Mixup [50], RandAug [3], (5) Diffusion: Class Prompt [40], Class-Attribute Prompt [40], CiP [20].", + "bbox": [ + 212, + 311, + 785, + 357 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results and Analysis From Table 2, SSD-LLM surpasses previous methods, with a $+1.6\\%$ improvement in average accuracy and $+3.3\\%$ in worst group accuracy across three datasets. The analysis is as follows: (1) Despite being based on conventional ERM, data-based approaches show competitive performance compared to model-based algorithms, highlighting their potential. (2) For diffusion-based methods, class-attribute prompts outperform class prompts, underscoring the importance of understanding dataset imbalanced attributes for effective image generation. However, the need for pre-identifying these attributes emphasizes the value of SSD-LLM, which automates attribute discovery and provides detailed annotations, enhancing performance. (3) The superior performance of CiP over Class Prompt underscores the significance of diverse text prompts. (4) Our method's superior performance results from a comprehensive analysis of subpopulation imbalances within the dataset. The strategic text diversity achieved through attribute sampling and LLM sentence-making effectively addresses subpopulation shifts.", + "bbox": [ + 212, + 357, + 785, + 583 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Evaluation on Slice Discovery", + "text_level": 1, + "bbox": [ + 215, + 604, + 504, + 619 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Setup In contrast to typical slice discovery tasks, we redefine evaluation pipeline following [11]. In this study, we evaluate bugs found of ImageNet models. Specifically, a classification is deemed incorrect when an image containing target object is erroneously identified by the model as containing an unrelated object.", + "bbox": [ + 212, + 628, + 785, + 688 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison Methods Domino [9] represents a state-of-the-art method in slice discovery, which effectively clusters errors identified in the validation set and characterizes them through captions generated automatically. B2T [16] is a recently proposed framework which identifies and interprets visual biases in vision models using keyword extraction from captions of mispredicted images.", + "bbox": [ + 212, + 688, + 785, + 763 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results and Analysis We evaluate the effectiveness of our method in slice discovery on 6 representative superclasses in Imagenet. As shown in Table. 3, our SSD-LLM overcomes all other SDMs, including Domino [9], by a significant margin. We find topics given by Domino tend to encounter two unsatisfactory cases: loss of semantics, and missing class information. These cases are also discussed", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Luo et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/38e3b2cfb85ac22a1de5e9de4972769cc02ab21cec3af3f9ccf6c9386a378871.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsIndexCriteria InitializationCriteria RefinementAccuracy
DimensionAttributeActionLocationMood
SSD LLMA10 × 2020 × 20×75.667.371.6
B5 × 2020 × 20×65.260.066.2
C20 × 2020 × 20×75.967.171.3
D10 × 2010 × 40×70.368.969.5
E10 × 2040 × 10×80.270.074.4
F10 × 2020 × 20✓81.770.576.8
", + "bbox": [ + 235, + 143, + 767, + 256 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 4: Ablation study on the number of samples (NUM of rounds × NUM per round), and Criteria Refinement (with or without).", + "bbox": [ + 215, + 257, + 784, + 284 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "in [11], where they reasoned this phenomenon into the inherent difficulties of automatic SDMs. However, evidence suggests that our SSD-LLM can handle these problems, while keeping a high error rate and maintaining automation. Specifically, SSD-LLM achieves an average error rate of $47.39\\%$ , surpassing Domino [9] by $3.95\\%$ . Furthermore, when we trace back to origin dataset, the discovered slice is also very consistent (detailed visualizations included in appendix). Interestingly, we find the data mining process of SSD-LLM is just the same as human data scientists, who take up hypotheses and improve model performance by viewing batches of bad subpopulations. Experiments show that our method overcome the inherent difficulties while maintaining automaton, paving the way for data-centric methods.", + "bbox": [ + 212, + 325, + 787, + 491 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 513, + 388, + 529 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Hyperparameters of Criteria Initialization/Refinement The $\\mathrm{N} \\times \\mathrm{M}$ in Table. 4 represents the rounds and samples of suggestions. Setting A serves as our baseline. For A, B, and C, results show that insufficient suggested dimensions lead to decreased performance, while enough dimension samples lead to stable performance, as the ICTC task requires only the most appropriate match. Too few suggestions may fail to find suitable dimensions, resulting in mismatched attribute generation. For A, D, and E, results highlight the importance of the number of captions in summarizing attributes. Too many captions can cause some to be overlooked, reducing total identified attributes and performance. The improvement from A to F demonstrates that Criteria Refinement enhances attribute comprehensiveness and final performance.", + "bbox": [ + 212, + 537, + 787, + 703 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Diffusion Generation Strategy From Table. 5, we analyze key components for employing diffusion models to address subpopulation shift. For $\\mathrm{A}\\rightarrow \\mathrm{B}$ , we confirm that managing imbalance attributes within datasets helps solve the task. For $\\mathrm{B}\\rightarrow \\mathrm{D}$ , we illustrate that a comprehensive subpopulation structure benefits the task. For $\\mathrm{D}\\rightarrow \\mathrm{E}$ , balanced subpopulation sampling improves data quality and model training. For $\\mathrm{E}\\rightarrow \\mathrm{F}$ , LLM-generated prompts produce more reasonable images, enhancing results. For $\\mathrm{F}\\rightarrow \\mathrm{G}$ , LLM suggests additional attributes to enrich subpopulation structures, generating more diverse images and improving model generalization. For $\\mathrm{F}\\rightarrow \\mathrm{H}$ , we verify the scaling capability of our SSD-LLM.", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Subpopulation Structure Discovery with Large Language Model", + "bbox": [ + 305, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e9d152beedeead21b134e13c9db385deb11ab1f49f7c4ba1cbb23019501e2fb2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsIndexAttribute ModeSample ModeSD Prompt ModeNumberAverage AccuracyWorst Group Accuracy
ERM————————×184.169.1
SSD-LLMAGT Attribute UnknownrandomDirect×185.971.3
BGT Attribute KnownrandomDirect×189.173.5
CGT Attribute KnownweightedDirect×189.373.8
DSSD-LLM AttributerandomDirect×189.576.2
ESSD-LLM AttributeweightedDirect×189.877.4
FSSD-LLM AttributeweightedLLM Sentence×190.178.3
GSSD-LLM & LLM SuggestweightedLLM Sentence×190.579.1
HSSD-LLM AttributeweightedLLM Sentence×291.179.2
", + "bbox": [ + 220, + 143, + 784, + 261 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 5: Ablation study on the sample mode and SD prompt mode.", + "bbox": [ + 269, + 262, + 730, + 273 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Discussion", + "text_level": 1, + "bbox": [ + 215, + 311, + 354, + 328 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Although our method, SSD-LLM, has already shown effectiveness in various settings, the current exploration of the algorithm is limited to image datasets. Besides, this approach may bear the potential bias from the MLLMs and LLMs. However, it may be reduced from extra human-in-the-loop guidances if available. For future works, we suggest the following promising directions:", + "bbox": [ + 212, + 340, + 782, + 416 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Structure Format The four-layer subpopulation structure can be expanded to more suitable structures according to specific task requirements.", + "- Downstream Tasks SSD-LLM can have more applications in various computer vision and multimodality tasks, e.g. object detection and VQA.", + "- Dataset Construction The subpopulation structure obtained from SSD-LLM holds the potential to guide dataset construction with better fairness [44] or further supporting the construction of unbiased datasets [29].", + "- Technical Extensions The core procedures of SSD-LLM, using LLM to conduct group-level summarizations, can be extended to more types of contents including patterns of model hallucinations." + ], + "bbox": [ + 223, + 424, + 782, + 571 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 594, + 359, + 609 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we present the first systematic exploration of subpopulation structure discovery. We provide a precise definition of subpopulation structure and introduce a fine-grained criteria to determine the structures. We propose SSD-LLM for automatic subpopulation structure discovery incorporating elaborate prompt engineering techniques. SSD-LLM can be combined with subsequent operations to better tackle several downstream tasks, including dataset subpopulation organization, subpopulation shift and slice discovery, with minor Task-specific Tuning. We conduct extensive experiments to verify our proposed methods, demonstrating the remarkable effectiveness and and generality of SSD-LLM.", + "bbox": [ + 212, + 623, + 784, + 758 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 780, + 403, + 797 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Shanghai Zhang is supported by the National Science and Technology Major Project of China (No. 2022ZD0117801).", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Luo et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)", + "2. Chen, M., Li, Y., Xu, Q.: Hibug: On human-interpretable model debug. In: NeurIPS (2023)", + "3. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 702-703 (2020)", + "4. Cui, J., Li, Z., Yan, Y., Chen, B., Yuan, L.: Chatlaw: Open-source legal large language model with integrated external knowledge bases. arXiv preprint arXiv:2306.16092 (2023)", + "5. Cui, Y., Jia, M., Lin, T.Y., Song, Y., Belongie, S.: Class-balanced loss based on effective number of samples. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9268-9277 (2019)", + "6. Dai, D., Sun, Y., Dong, L., Hao, Y., Ma, S., Sui, Z., Wei, F.: Why can gpt learn in-context? language models implicitly perform gradient descent as meta-optimizers. In: ICLR 2023 Workshop on Mathematical and Empirical Understanding of Foundation Models (2023)", + "7. Dong, Q., Li, L., Dai, D., Zheng, C., Wu, Z., Chang, B., Sun, X., Xu, J., Sui, Z.: A survey for in-context learning. arXiv preprint arXiv:2301.00234 (2022)", + "8. Dunlap, L., Umino, A., Zhang, H., Yang, J., Gonzalez, J.E., Darrell, T.: Diversify your vision datasets with automatic diffusion-based augmentation. arXiv preprint arXiv:2305.16289 (2023)", + "9. Eyuboglu, S., Varma, M., Saab, K., Delbrouck, J.B., Lee-Messer, C., Dunnmon, J., Zou, J., Ré, C.: Domino: Discovering systematic errors with cross-modal embeddings. arXiv preprint arXiv:2203.14960 (2022)", + "0. Fu, Y., Peng, H., Khot, T., Lapata, M.: Improving language model negotiation with self-play and in-context learning from ai feedback. arXiv preprint arXiv:2305.10142 (2023)", + "1. Gao, I., Ilharco, G., Lundberg, S., Ribeiro, M.T.: Adaptive testing of computer vision models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4003-4014 (2023)", + "2. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023)", + "3. Japkowicz, N.: The class imbalance problem: Significance and strategies. In: Proc. of the Int'l Conf. on artificial intelligence. vol. 56, pp. 111-117 (2000)", + "4. Johnson, N., Cabrera, A., Plumb, G., Talwalkar, A.: Where does my model underperform? a human evaluation of slice discovery algorithms. Proceedings of the AAAI Conference on Human Computation and Crowdsourcing 11(1), 65-76 (Nov 2023). https://doi.org/10.1609/hcomp.v11i1.27548, http://dx.doi.org/10.1609/hcomp.v11i1.27548", + "5. Kim, Y., Mo, S., Kim, M., Lee, K., Lee, J., Shin, J.: Bias-to-text: Debi-aising unknown visual biases through language interpretation. arXiv preprint arXiv:2301.11104 (2023)", + "6. Kim, Y., Mo, S., Kim, M., Lee, K., Lee, J., Shin, J.: Bias-to-text: Debiased unknown visual biases through language interpretation. arXiv preprint arXiv:2301.11104 (2023)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Subpopulation Structure Discovery with Large Language Model", + "bbox": [ + 305, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Kojima, T., Gu, S.S., Reid, M., Matsuo, Y., Iwasawa, Y.: Large language models are zero-shot reasoners. NeurIPS 35, 22199-22213 (2022)", + "18. Kwon, S., Park, J., Kim, M., Cho, J., Ryu, E.K., Lee, K.: Image clustering conditioned on text criteria. In: ICLR (2023)", + "19. Kwon, S., Park, J., Kim, M., Cho, J., Ryu, E.K., Lee, K.: Image clustering conditioned on text criteria. arXiv preprint arXiv:2310.18297 (2023)", + "20. Lei, S., Chen, H., Zhang, S., Zhao, B., Tao, D.: Image captions are natural prompts for text-to-image models. arXiv preprint arXiv:2307.08526 (2023)", + "21. Li, X., Qiu, X.: Mot: Pre-thinking and recalling enable chatgpt to self-improve with memory-of-thoughts. arXiv preprint arXiv:2305.05181 (2023)", + "22. Liang, W., Zou, J.: Metashift: A dataset of datasets for evaluating contextual distribution shifts and training conflicts. arXiv preprint arXiv:2202.06523 (2022)", + "23. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)", + "24. Liu, E.Z., Haghgoo, B., Chen, A.S., Raghunathan, A., Koh, P.W., Sagawa, S., Liang, P., Finn, C.: Just train twice: Improving group robustness without training group information. In: International Conference on Machine Learning. pp. 6781-6792. PMLR (2021)", + "25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)", + "26. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)", + "27. Liu, J., Hu, T., Zhang, Y., Gai, X., Feng, Y., Liu, Z.: A chatgpt aided explainable framework for zero-shot medical image diagnosis. arXiv preprint arXiv:2307.01981 (2023)", + "28. Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., Neubig, G.: Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys 55(9), 1-35 (2023)", + "29. Liu, Z., He, K.: A decade's battle on dataset bias: Are we there yet? (2024), https://arxiv.org/abs/2403.08632", + "30. Luo, Y., Tang, Y., Shen, C., Zhou, Z., Dong, B.: Prompt engineering through the lens of optimal control. arXiv preprint arXiv:2310.14201 (2023)", + "31. Madaan, A., Tandon, N., Gupta, P., Hallinan, S., Gao, L., Wiegreffe, S., Alon, U., Dziri, N., Prabhumoye, S., Yang, Y., et al.: Self-refine: Iterative refinement with self-feedback. arXiv preprint arXiv:2303.17651 (2023)", + "32. Nam, J., Cha, H., Ahn, S., Lee, J., Shin, J.: Learning from failure: De-biasing classifier from biased classifier. Advances in Neural Information Processing Systems 33, 20673-20684 (2020)", + "33. Park, J.S., O'Brien, J.C., Cai, C.J., Morris, M.R., Liang, P., Bernstein, M.S.: Generative agents: Interactive simulacra of human behavior. arXiv preprint arXiv:2304.03442 (2023)", + "34. Qian, C., Cong, X., Yang, C., Chen, W., Su, Y., Xu, J., Liu, Z., Sun, M.: Communicative agents for software development. arXiv preprint arXiv:2307.07924 (2023)", + "35. Qiao, S., Ou, Y., Zhang, N., Chen, X., Yao, Y., Deng, S., Tan, C., Huang, F., Chen, H.: Reasoning with language model prompting: A survey. arXiv preprint arXiv:2212.09597 (2022)", + "36. Ren, J., Yu, C., Ma, X., Zhao, H., Yi, S., et al.: Balanced meta-softmax for long-tailed visual recognition. Advances in neural information processing systems 33, 4175-4186 (2020)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Luo et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115, 211-252 (2015)", + "38. Sagawa, S., Koh, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731 (2019)", + "39. Santurkar, S., Tsipras, D., Madry, A.: Breeds: Benchmarks for subpopulation shift (2020)", + "40. Shipard, J., Wiliem, A., Thanh, K.N., Xiang, W., Fookes, C.: Diversity is definitely needed: Improving model-agnostic zero-shot classification via stable diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 769-778 (2023)", + "41. Van Gansbeke, W., Vandenhende, S., Georgoulis, S., Proesmans, M., Van Gool, L.: Scan: Learning to classify images without labels. In: European conference on computer vision. pp. 268-285. Springer (2020)", + "42. Vapnik, V.N.: An overview of statistical learning theory. IEEE transactions on neural networks 10(5), 988-999 (1999)", + "43. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)", + "44. Wang, A., Liu, A., Zhang, R., Kleiman, A., Kim, L., Zhao, D., Shirai, I., Narayanan, A., Russakovsky, O.: Revise: A tool for measuring and mitigating bias in visual datasets. International Journal of Computer Vision 130(7), 1790-1810 (2022)", + "45. Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A., Zhou, D.: Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171 (2022)", + "46. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)", + "47. Xie, S.M., Raghunathan, A., Liang, P., Ma, T.: An explanation of in-context learning as implicit bayesian inference. arXiv preprint arXiv:2111.02080 (2021)", + "48. Yang, Y., Zhang, H., Katabi, D., Ghassemi, M.: Change is hard: A closer look at subpopulation shift (2023)", + "49. Yao, H., Wang, Y., Li, S., Zhang, L., Liang, W., Zou, J., Finn, C.: Improving out-of-distribution robustness via selective augmentation. In: International Conference on Machine Learning. pp. 25407-25437. PMLR (2022)", + "50. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)", + "51. Zhang, R., Cai, Z., Yang, H., Liu, Z., Gudovskiy, D., Okuno, T., Nakata, Y., Keutzer, K., Chang, B., Du, Y., et al.: Vecaf: Vlm-empowered collaborative active finetuning with training objective awareness. arXiv preprint arXiv:2401.07853 (2024)", + "52. Zhang, X., He, Y., Xu, R., Yu, H., Shen, Z., Cui, P.: Nico++: Towards better benchmarking for domain generalization (2022)", + "53. Zhou, D., Schärli, N., Hou, L., Wei, J., Scales, N., Wang, X., Schuurmans, D., Cui, C., Bousquet, O., Le, Q., et al.: Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 784 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Subpopulation Structure Discovery with Large Language Model", + "bbox": [ + 305, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/LLM as Dataset Analyst_ Subpopulation Structure Discovery with Large Language Model/cfcf5c50-f6b7-4af0-a369-8bdd05a30c95_model.json b/2024/LLM as Dataset Analyst_ Subpopulation Structure Discovery with Large Language Model/cfcf5c50-f6b7-4af0-a369-8bdd05a30c95_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c4f1190e3d7fb67222de58582722a593733d9061 --- /dev/null +++ b/2024/LLM as Dataset Analyst_ Subpopulation Structure Discovery with Large Language Model/cfcf5c50-f6b7-4af0-a369-8bdd05a30c95_model.json @@ -0,0 +1,2500 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.142, + 0.782, + 0.187 + ], + "angle": 0, + "content": "LLM as Dataset Analyst: Subpopulation Structure Discovery with Large Language Model" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.216, + 0.674, + 0.231 + ], + "angle": 0, + "content": "Yulin Luo\\(^{1*}\\), Ruichuan An\\(^{1,2*}\\), Bocheng Zou\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.231, + 0.7, + 0.248 + ], + "angle": 0, + "content": "Yiming Tang\\(^{1,4}\\), Jiaming Liu\\(^{1}\\), and Shanghang Zhang\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.259, + 0.707, + 0.273 + ], + "angle": 0, + "content": "\\(^{1}\\)State Key Laboratory of Multimedia Information Processing," + }, + { + "type": "text", + "bbox": [ + 0.345, + 0.274, + 0.658, + 0.287 + ], + "angle": 0, + "content": "School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.287, + 0.705, + 0.301 + ], + "angle": 0, + "content": "\\(^{2}\\)Xi'an Jiaotong University \\(^{3}\\)University of Wisconsin-Madison" + }, + { + "type": "text", + "bbox": [ + 0.39, + 0.301, + 0.615, + 0.315 + ], + "angle": 0, + "content": "\\(^{4}\\)National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.423, + 0.316, + 0.58, + 0.329 + ], + "angle": 0, + "content": "yulin@stu.pku.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.398, + 0.744, + 0.76 + ], + "angle": 0, + "content": "Abstract. The distribution of subpopulations is an important property hidden within a dataset. Uncovering and analyzing the subpopulation distribution within datasets provides a comprehensive understanding of the datasets, standing as a powerful tool beneficial to various downstream tasks, including Dataset Subpopulation Organization, Subpopulation Shift, and Slice Discovery. Despite its importance, there has been no work that systematically explores the subpopulation distribution of datasets to our knowledge. To address the limitation and solve all the mentioned tasks in a unified way, we introduce a novel concept of subpopulation structures to represent, analyze, and utilize subpopulation distributions within datasets. To characterize the structures in an interpretable manner, we propose the Subpopulation Structure Discovery with Large Language Models (SSD-LLM) framework, which employs world knowledge and instruction-following capabilities of Large Language Models (LLMs) to linguistically analyze informative image captions and summarize the structures. Furthermore, we propose complete workflows to address downstream tasks, named Task-specific Tuning, showcasing the application of the discovered structure to a spectrum of subpopulation-related tasks, including dataset subpopulation organization, subpopulation shift, and slice discovery. With the help of SSD-LLM, we can structuralize the datasets into subpopulation-level automatically, achieve average \\(+3.3\\%\\) worst group accuracy gain compared to previous methods on subpopulation shift benchmark Waterbirds, Metashift and \\(\\mathrm{Nico + + }\\), and also identify more consistent slice topics with a higher model error rate of \\(3.95\\%\\) on slice discovery task for ImageNet. The code will be available at https://llm-as-dataset-analyst.github.io/." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.775, + 0.743, + 0.79 + ], + "angle": 0, + "content": "Keywords: Subpopulation Structure Discovery \\(\\cdot\\) Large Language Model" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.811, + 0.371, + 0.825 + ], + "angle": 0, + "content": "* Equal Contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.825, + 0.389, + 0.84 + ], + "angle": 0, + "content": "† Corresponding Author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Y. Luo et al." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.146, + 0.773, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.233, + 0.774, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.405, + 0.788, + 0.449 + ], + "angle": 0, + "content": "Fig. 1: (A) The Workflow of Subpopulation Structure Discovery with Large Language Models (SSD-LLM). SSD-LLM can further support several downstream tasks including: (B) Dataset Subpopulation Organization; (C) Subpopulation Shift; (D) Slice discovery." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.476, + 0.376, + 0.492 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.784, + 0.613 + ], + "angle": 0, + "content": "Subpopulation, defined by a set of data points that share common characteristics, is an important concept in machine learning [48]. Many tasks are subpopulation-related. For example, image clustering conditioned on text criteria [18] is to partition an image dataset into different subpopulations based on user-specified criteria, studying subpopulation shift [22, 48, 52] is to mitigate the negative impact of imbalanced subpopulation distributions in the training set on the model, slice discovery [2, 9] is aimed at identifying subpopulations model underperform." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.785, + 0.78 + ], + "angle": 0, + "content": "Summarizing the commonalities of these tasks, we find that analyzing the subpopulation distribution is the key to solving all these problems. If the subpopulation distribution can be characterized, image clustering results under different criteria are naturally obtained [18], additional images can be supplemented to rare subgroups to balance the whole dataset [8], and slices can be easily discovered by statistics error rate on validation set [2]. Despite its importance, existing work [48] lacks systematic exploration of subpopulation distribution. To adjust the issue, for the first time, we propose the concept of subpopulation structure to represent, analyze, and utilize subpopulation distributions within datasets. By definition, a subpopulation structure is a set of hierarchical relations among several subpopulations determined by certain criteria." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Former works like Metashift [22] and \\(\\mathrm{NICO}++\\) [52] have constructed image datasets including the subpopulation information, which organizes the images with respect to some extra attributes, and can be viewed as a class-, attribute-, subpopulation-layer structure. The problem of such a structure is ignoring" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.306, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Subpopulation Structure Discovery with Large Language Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.145, + 0.398, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.145, + 0.782, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.305, + 0.788, + 0.377 + ], + "angle": 0, + "content": "Fig. 2: Metashift has the same-level attributes Surfboard, Water, and Grass for class Dog, which is irrational due to the possible overlap. As an improvement, we take dimensions into consideration. The class Dog has dimensions including Action, Co-occurrence Object, Location, etc., and in dimension Location, it includes various attributes like Water, Grass, etc, which offers a more appropriate assignment for the samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.785, + 0.488 + ], + "angle": 0, + "content": "the category of attributes (or Dimension), leading to attribute inconsistency and confusion. To solve this issue, we introduce a class-, dimension-, attribute-, and subpopulation-layer structure. The comparison of the two structures can be seen in Figure. 2. By articulating the classification dimensions, this improved structure provides more nuanced attribute assignments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.784, + 0.626 + ], + "angle": 0, + "content": "Identifying descriptive information within a dataset often requires large amounts of human manufacture [22, 52], which urges the need of automatic workflows to complete the subpopulation structure discovery. However, automatically identifying subpopulation structures within image datasets presents a significant challenge. The approach must be capable of extracting key information from images and summarizing essential content from extensive texts. Furthermore, it necessitates comprehensive world knowledge, enabling a broad understanding of various aspects of the datasets, including diverse categories, common attributes, and the relationships between dimensions and attributes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Recently, Large Language Model (LLM) [17, 27, 46] and Multimodal Large Language Model (MLLM) [12, 25] have attracted wide attention due to their superior capacities. LLM has shown extensive world knowledge and remarkable abilities in summarization, instruction following [27], etc. MLLM extends the capabilities of LLM to handle visual inputs. By visual instruction tuning [25], MLLM can verbalize the rich information of images. Motivated by these, we propose a novel framework Subpopulation Structure Discovery with Large Language Model(SSD-LLM), illustrated in Figure. 1, to automatically uncover the structure. The core idea is to generate informative captions from images with MLLM, followed by analyzing and summarizing the subpopulation structure of datasets with LLM. Specifically, we design two elaborate prompt engineering components, Criteria Initialization and Criteria Refinement. The former utilizes a generate-and-select paradigm to summarize dimensions and attributes sequentially. The latter employs self-consistency as an indicator" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Y. Luo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "to evaluate and refine the criteria. After obtaining complete criteria, each image is assigned to corresponding attributes according to its caption. The final subpopulation structures can be leveraged to finish various downstream tasks with the help of our proposed Task-specific Tuning. In this work, we focus on three application scenarios, i.e. dataset subpopulation organization, subpopulation shift, and slice discovery. We validate the effectiveness of SSD-LLM on these subpopulation-related tasks. For subpopulation shift, we achieve an improvement of \\(+3.3\\%\\) in worst group accuracy across three datasets compared to SOTA methods, and for slice discovery, we can identify more consistent slice topics with a higher model error rate of \\(3.95\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.299, + 0.568, + 0.312 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.322, + 0.784, + 0.351 + ], + "angle": 0, + "content": "- We introduce the concept of subpopulation structure to characterize subpopulation distribution in an interpretable manner for the first time." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.353, + 0.785, + 0.381 + ], + "angle": 0, + "content": "- We propose class-dimension-attribute-subpopulation structure, reducing the attribute confusion of the current class-attribute-subpopulation structure." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.383, + 0.785, + 0.439 + ], + "angle": 0, + "content": "- We propose Subpopulation Structure Discovery with Large Language Models (SSD-LLM) framework to uncover the underlying subpopulation structure of datasets automatically, with two elaborate prompt engineering components. Criteria Initialization and Criteria Refinement." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.443, + 0.785, + 0.486 + ], + "angle": 0, + "content": "- We provide methods for Task-specific Tuning, enabling the application of the structures across a spectrum of subpopulation-related tasks, including dataset subpopulation organization, subpopulation shift, and slice discovery." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.322, + 0.785, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.511, + 0.396, + 0.526 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.542, + 0.6, + 0.557 + ], + "angle": 0, + "content": "2.1 Hierarchical Structure of Image Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Recent research has emphasized the need to organize datasets into hierarchical structures allowing for benchmarking various downstream tasks [22, 37, 43, 52]. Metashift [22] builds a collection of 12,868 sets of images related to 410 main subjects and their contexts. NICO++ [52], Waterbirds [43], and ImageNetBG [37] also propose methods for constructing various types of hierarchical datasets. However, the construction of these hierarchical datasets often requires manual annotation, hindering automatic construction. These approaches focus on just a single dimension, such as object context in Metashift, background in Waterbirds, and ImageNetBG, while more practical scenarios may involve multiple dimensions hidden within the comprehensive visual information." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.74, + 0.609, + 0.755 + ], + "angle": 0, + "content": "2.2 Extract Information from Image Captions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Recent works such as ALIA [8], VeCAF [51], Bias2Text [15], and ICTC [19] explore utilizing caption models to obtain information from datasets. ALIA provides a method to augment datasets by generating variations of existing images through captioning and text-to-image models. While ALIA [8] supports dataset improvement, it lacks knowledge about attribute types, bias, or subpopulation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.306, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Subpopulation Structure Discovery with Large Language Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "shift existence. Bias2Text [15] detects dataset bias by transforming images into descriptive captions and keywords. However, without large language model participation, Bias2Text fails to support classification dimension selection and can only differentiate images with basic keywords. More recently, ICTC [19] enables conditional image clustering using an LLM in a straightforward manner. Although ICTC clusters images when given the criterion, it requires human-assigned text prompts. Compared to ICTC, our prompt engineering paradigm supports scalable automatic subpopulation structure dataset organization without human criteria assignment and can generate comprehensive criteria tailored to datasets for various downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.321, + 0.479, + 0.337 + ], + "angle": 0, + "content": "2.3 LLM Prompt Engineering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.789, + 0.619 + ], + "angle": 0, + "content": "As the popularity of LLMs has surged, prompt engineering, the process of crafting and refining prompts to guide LLMs towards desired outputs [17,46] has also shown more and more importance. Various prompt engineering methods [28,35] and principles have emerged, and researchers or engineers have explored their applications in a diverse range of downstream tasks [4,10,27,33,34]. In particular, in-context learning has emerged as a pivotal technique, validated both experimentally [7, 21] and theoretically [6, 30, 47]. This approach involves providing the LLM with context information relevant to the task at hand, enabling it to generate more accurate and relevant responses. Least-to-most prompting [53] breaks down complex tasks into smaller, more manageable steps, enhancing LLM's reasoning skill by querying the LLM with more simplified sub-questions. Self-consistency [45] proposes to ensemble multiple responses to the LLM given the same prompt to get enhanced results, suggesting that consistent responses as an indicator for correct problem solving. Self-refining [31] demonstrates that we can use LLMs to refine their outputs by themselves with careful designing of prompts. In this work, we leverage a combination of prompt engineering techniques, including in-context learning, chain-of-thought, self-consistency, and self-refining to tackle subpopulation structure discovery effectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.642, + 0.331, + 0.658 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this section, we introduce our proposed method, subpopulation structure discovery with large language models(SSD-LLM). We describe the overall pipeline of SSD-LLM in Section 3.1, outlining how the paradigm automatically discovers the latent subpopulation structures inside the dataset. The process begins with captioning the images in the dataset with an MLLM, detailed in Section 3.2, and proceeds with Criteria Initialization with an LLM in Section 3.3. The paradigm then refines the initialized criteria through a recursive self-refinement procedure, detailed in Section 3.4. Finally, images are assigned to attributes, completing the subpopulation structure discovery process, as elaborated in Section 3.5. The section concludes with a discussion of how to apply the method to various downstream tasks, as presented in Section 3.6." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Y. Luo et al." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.147, + 0.775, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.321, + 0.788, + 0.419 + ], + "angle": 0, + "content": "Fig. 3: Subpopulation Structure Discovery with Large Language Model (SSD-LLM). (Step 1) Multimodality Large Language Model (MLLM) extracts informative captions from images. (Step 2) LLM initializes the criteria with a sample-based generate-and-select paradigm. (Step 3) LLM refines the criteria using self-consistency as an indicator. (Step 4) LLM assigns each caption with specific attributes according to the refined criteria, uncovering the intrinsic subpopulation structures hidden in the dataset. The resulting criteria and subpopulations are used in several downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.45, + 0.341, + 0.463 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.789, + 0.703 + ], + "angle": 0, + "content": "To automatically discover the subpopulation structures, we propose a novel prompt engineering paradigm that effectively leverages the capabilities of both multimodal large language models (MLLMs) and large language models (LLMs). Our proposed method comprises four key steps (Figure. 3). First, we transform the images into information-rich captions that capture the main information in the images using MLLMs. Second, we employ a novel sample-based prompt engineering method to guide an LLM to produce criteria consisting of dimensions and corresponding attributes organizing the dataset. Third, we prompt the LLM to refine this generated criteria. Last, we assign all the images in the dataset to specific attributes accordingly, uncovering the intrinsic subpopulation structures in the dataset, and paving the way for further analysis about the dataset. Detailed descriptions of each step are provided below. For notations, consistent with various former works [30,46], we denote the operation of getting responses from the language models as \\( LLM \\) and \\( MLLM \\), and use [,] to represent the concatenation operation of two pacts of texts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.724, + 0.424, + 0.74 + ], + "angle": 0, + "content": "3.2 Caption Extraction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.825 + ], + "angle": 0, + "content": "To begin our approach, we leverage the powerful image captioning capabilities of the MLLM to transform the images into informative and detailed captions. Instead of briefly describing the images, we prompt the MLLM to generate more detailed captions centered around the main subject CLS. The prompt we used in this step is stated as follow:" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.825, + 0.646, + 0.841 + ], + "angle": 0, + "content": "\\(P_{1} =\\) \"Describe the image of the subject CLS in detail.\"" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.306, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Subpopulation Structure Discovery with Large Language Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.173, + 0.766, + 0.297 + ], + "angle": 0, + "content": "
Step 1 Caption ExtractionStep 2-1 Dimension Generation
Input: Dataset: Dimg, MLLMInput: Captions: C, LLM
Output: Image Captions: COutput: dimensions: Dims
1: for i in range(NumOfIterations)1: for i in range(NumOfIterations)
2: img = Dimg.sample()2: c = C/sample(NumOfSamples)
3: c = MLLM(img, P1)3: S.append(LLM([P2^1, c]))
4: C.append(c)4: end for
5: end for5: Dims = LLMSummary(S)
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.32, + 0.438, + 0.334 + ], + "angle": 0, + "content": "3.3 Criteria Initialization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.788, + 0.514 + ], + "angle": 0, + "content": "To discover the hidden subpopulation structures within the dataset, we employ an LLM to delve into the information-rich captions generated in the previous step. Our objective is to identify certain criteria that effectively partition the images into several distinct subgroups. Beyond simply dividing the dataset into subgroups, we articulate the classification dimension for the partition and record all the resulting attributes generated from the classification process. Along with the class information and the resulting subpopulations, this criteria naturally form a four-layer structure, class-, dimension-, attribute-, and subpopulation-. Noticing criteria encompass multiple dimensions and their corresponding attributes, we adopt a generate-and-select paradigm with the LLM to discover the dimensions and the attributes sequentially." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.665 + ], + "angle": 0, + "content": "To determine the dimensions and attributes, we employ an iterative sampling approach, repeatedly prompting the LLM to propose dimensions and attributes based on batches of image captions. In each iteration, the LLM generates candidate dimensions and attributes, which are subsequently processed through an LLM summarization process. This sample-and-summarize approach effectively addresses the challenges when processing large datasets. Since the number of dimensions that can differentiate images in a dataset is relatively small, and these dimensions have an appearance in numerous images, our sample-based approach effectively identifies relevant dimensions. The prompts we used in this step are stated as follow, omitting Chain-of-thought examples for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.666, + 0.785, + 0.695 + ], + "angle": 0, + "content": "\\(P_2^1 = \\text{\"Suggest some dimensions that can differentiate the following image captions.}\\)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.696, + 0.785, + 0.726 + ], + "angle": 0, + "content": "\\(P_2^2 = \\text{\"Suggest a complete criterion to differentiate the following image captions by the given dimension.}\"\\)" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.752, + 0.427, + 0.766 + ], + "angle": 0, + "content": "3.4 Criteria Refinement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To further refine the criteria and ensure its effectiveness in classifying image captions across the dataset, we implement a recursive refining process. This approach proposes a novel method for identifying image captions requiring further refinement utilizing the self-consistency of LLM responses as an indicator [45]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Y. Luo et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.217, + 0.147, + 0.446, + 0.162 + ], + "angle": 0, + "content": "Step 2-2 Attribute Generation" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.165, + 0.684, + 0.179 + ], + "angle": 0, + "content": "Input: Dimensions: Dims, Captions: \\(C\\), Large language model: LLM" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.18, + 0.473, + 0.193 + ], + "angle": 0, + "content": "Output: Initialized criteria: Criteria" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.194, + 0.386, + 0.206 + ], + "angle": 0, + "content": "1: for dim in Dims do" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.207, + 0.506, + 0.221 + ], + "angle": 0, + "content": "2: for i in range(NumOfIterations) do" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.222, + 0.479, + 0.235 + ], + "angle": 0, + "content": "3: \\( c = C \\).sample(NumOfSamples)" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.235, + 0.473, + 0.25 + ], + "angle": 0, + "content": "4: \\(S.\\mathrm{append}(\\mathrm{LLM}([P_2^2, dim, c]))\\)" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.25, + 0.316, + 0.261 + ], + "angle": 0, + "content": "5: end for" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.263, + 0.583, + 0.277 + ], + "angle": 0, + "content": "6: Attributes \\( = \\) LLMSumarry(S) *list of attributes" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.277, + 0.442, + 0.291 + ], + "angle": 0, + "content": "7: Criteria[dim] = Attributes" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.291, + 0.317, + 0.305 + ], + "angle": 0, + "content": "8: \\(S.\\mathrm{reset}()\\)" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.306, + 0.3, + 0.317 + ], + "angle": 0, + "content": "9: end for" + }, + { + "type": "list", + "bbox": [ + 0.224, + 0.194, + 0.583, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.787, + 0.443 + ], + "angle": 0, + "content": "This choice stems from our empirical observation that if an image can be accurately classified according to a particular dimension, it should consistently be classified into the same attribute multiple times. Inconsistent responses, however, suggest that the current criteria require further refinement, either to merge redundant attributes or to include new attributes. The prompts we used in this step are stated as follow and the pseudocode is included in the appendix:" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.444, + 0.642, + 0.459 + ], + "angle": 0, + "content": "\\(P_3^1 = \\text{\"Classify the caption by the criteria listed below.\"}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.788, + 0.52 + ], + "angle": 0, + "content": "\\(P_{3}^{2} = \\text{\"We are unable to classify the following image caption using the provided criteria due to attributes redundancy or disappearance. If redundancy, please prune the redundant attributes. If missing, please suggest an additional attribute that would match the image caption.\"}\\)" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.543, + 0.484, + 0.559 + ], + "angle": 0, + "content": "3.5 Subpopulation Assignment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.787, + 0.674 + ], + "angle": 0, + "content": "Equipped with the comprehensive criteria, we proceed to systematically assign each image to the specific attributes of each dimension. Images assigned to the same attributes across all dimensions form distinct subgroups within the dataset, revealing the intrinsic subpopulation structures hidden within the data. These subpopulation structures can then be leveraged to perform various downstream tasks, completing our overall pipeline for employing an LLM to analyze the image dataset. The prompt we used in this step is stated as follow:" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.675, + 0.782, + 0.69 + ], + "angle": 0, + "content": "\\(P_{4} =\\) \"Please assign following caption to one attribute of given dimension.\"" + }, + { + "type": "code_caption", + "bbox": [ + 0.217, + 0.718, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Step 4 Subpopulation Assignment" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.737, + 0.674, + 0.752 + ], + "angle": 0, + "content": "Input: Captions: \\(C\\), Large language model: LLM, criteria: Criteria" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.752, + 0.548, + 0.766 + ], + "angle": 0, + "content": "Output: Further assignments for each caption \\( c \\)." + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.767, + 0.339, + 0.778 + ], + "angle": 0, + "content": "1: for \\(c\\) in \\(C\\) do" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.779, + 0.506, + 0.793 + ], + "angle": 0, + "content": "2: for [dim, Attributes] in Criteria do" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.794, + 0.52, + 0.808 + ], + "angle": 0, + "content": "3: c.Assign(LLM([P4,c, dim, Attributes]))" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.809, + 0.316, + 0.82 + ], + "angle": 0, + "content": "4: end for" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.822, + 0.301, + 0.834 + ], + "angle": 0, + "content": "5: end for" + }, + { + "type": "list", + "bbox": [ + 0.224, + 0.767, + 0.52, + 0.834 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.306, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Subpopulation Structure Discovery with Large Language Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.148, + 0.767, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.346, + 0.741, + 0.361 + ], + "angle": 0, + "content": "Fig. 4: A visualization of organised subpopulations in a dataset of cats." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.391, + 0.62, + 0.406 + ], + "angle": 0, + "content": "3.6 Task-specific Tuning for Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Once we have identified the criteria and subpopulations within the dataset, we can leverage this information to tackle several downstream tasks effectively. This includes organizing the subpopulations, which can reveal valuable insights into the data, such as potential subpopulation biases and the presence of long-tail attributes. In practice, the subpopulation structures can be used to improve model performance on various tasks if combined with extra operations, including handling subpopulation shifts and slice discovery." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Dataset Subpopulation Organization Organizing subpopulations within a dataset can provide beneficial information about the numbers of data points in each subpopulation and can reflect dataset biases and help us identify some long-tail subgroups. We conduct experiments on the task of image clustering conditioned on human-specified criteria to evaluate the quality of subpopulation organization. In specific, when organising the subpopulations of a given image dataset, we first select out the relevant dimensions and then attach attributes assigned by SSD-LLM directly to the images accordingly. Our method, SSD-LLM, automates subpopulation organization within datasets, and thus holds the potential to revolutionize the way hierarchical datasets are constructed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Subpopulation Shift Subpopulation shift [48] stands as a common challenge in machine learning, occurring when the proportion of some subpopulations between training and deployment changes, and is shown to be of significant influence to model performances [39, 48]. SSD-LLM, combined with image generation, offers a solution to better handle the scenarios of subpopulation shifts. In our experiments, after we apply SSD-LLM to the datasets, we collect statistics the number of images contained in each subpopulation and utilize diffusion model to generate images for underrepresented subpopulations. Subsequently, we sample attributes from the subpopulation structure for each underrepresented subpopulation and employ LLM to make complete sentences based on these words as the input prompt of a diffusion model. The diffusion model generates images augmented" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "Y. Luo et al." + }, + { + "type": "table", + "bbox": [ + 0.314, + 0.144, + 0.691, + 0.272 + ], + "angle": 0, + "content": "
DatasetCriterionSCAN*IC|TCOurs
Stanford 40 ActionAction0.3460.7470.817
Location0.3570.6710.705
Mood0.2760.7460.768
Place365Place0.332-0.696
PPMIMusical Instruction0.5980.9340.955
Cifar10Object0.8390.9110.921
STL10Object0.7980.9860.988
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "Table 1: Quantitative results of Dataset Subpopulation Organization. Method labeled with * is evaluated by having a human provided ground truth labels, cause the method itself is an unsupervised learning paradigm." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.784, + 0.416 + ], + "angle": 0, + "content": "to the image dataset which further helps to achieve balanced classes and attributes. Moreover, we propose to harness an LLM to suggest extra dimensions and attributes based on the current sets in this task for enriched subpopulation structure, generating more diverse images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.785, + 0.569 + ], + "angle": 0, + "content": "Slice Discovery Slice discovery is a task aiming at uncovering subpopulations within a dataset where a machine learning model consistently exhibits poor performances. These subpopulations with underperformances, or slices, provide valuable insights into the model's limitations, potential biases [14], and how to improve the performances. SSD-LLM conducts slice discovery for an image dataset with the help of the assigned attributes. In detail, we first calculate the error rates on all subpopulations discovered with the SSD-LLM. Then we identify out the subpopulations with the highest error rate and use the LLM to summarize out discriptions based on the attributes of the subpopulations in the form of texts representing the slice topics, completing the task of slice discovery." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.592, + 0.377, + 0.61 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.784, + 0.685 + ], + "angle": 0, + "content": "We now present experimental results demonstrating the effectiveness of SSD-LLM. In particular, we present the main settings and results in this section and defer extra details, including various visualization result, to appendix. In our experiments, we mainly use LLaVA1.5 [26] for MLLM and GPT-4 [1] for LLM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.686, + 0.784, + 0.761 + ], + "angle": 0, + "content": "We conduct experiments on Dataset Subpopulation Organization, Subpopulation Shift, and Slice Discovery. Our superior performance underscores the method's efficacy in identifying and analyzing subgroups, further affirming its utility in addressing related challenges. Moreover, it illustrates that the unified paradigm can effectively address a variety of downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.784, + 0.684, + 0.8 + ], + "angle": 0, + "content": "4.1 Evaluation on Dataset Subpopulation Organization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Setup The data assignment process facilitated by SSD-LLM, can be considered a form of clustering. To evaluate the quality of these identified subpopulations," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.306, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Subpopulation Structure Discovery with Large Language Model" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.144, + 0.778, + 0.32 + ], + "angle": 0, + "content": "
TypeMethodAverage AccuracyWorst Group Accuracy
WaterbirdsMetashiftNico++AverageWaterbirdsMetashiftNico++Average
VanillaERM84.191.276.383.769.182.117.856.3
Subgroup Robust MethodsGroupDRO86.991.574.084.173.183.112.256.1
JTT88.991.277.585.971.282.615.656.5
LfF86.680.477.581.575.072.315.654.3
LISA89.291.475.085.277.079.018.958.3
Imbalanced LearningResample86.292.277.385.270.081.016.755.9
Reweight86.291.573.883.871.983.112.255.7
Focal89.391.673.184.771.681.016.756.4
CBLoss86.891.476.384.874.483.112.256.6
BSoftmax88.491.374.284.674.182.616.757.8
Traditional Data AugmentationMixup89.291.473.084.577.579.014.457.0
RandAug86.390.972.083.171.480.916.756.3
DiffusionClass Prompt85.991.578.085.171.382.718.557.5
Class-Attribute89.191.478.686.473.583.818.858.7
CiP88.091.178.385.873.582.419.358.4
LLM+DiffusionSSD-LLM (Ours)90.593.080.488.079.184.822.162.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.223, + 0.321, + 0.779, + 0.333 + ], + "angle": 0, + "content": "Table 2: Comparison of methods for image classification with subpopulation shifts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.784, + 0.422 + ], + "angle": 0, + "content": "we assess the clustering accuracy by comparing images against secondary labels that reflect subgroup attributes derived via SSD-LLM. Full information for datasets, text criterion, and model selection can be found in appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.785, + 0.483 + ], + "angle": 0, + "content": "Comparison Methods SCAN [41] is a two-stage clustering method that decouples feature learning and clustering. IC|TC [19] is a new paradigm for image clustering that supports human interaction. It utilizes the given Text Criteria to accurately control the quality of the clustering results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.785, + 0.605 + ], + "angle": 0, + "content": "Results and Analysis In Table. 1, we report the average accuracy achieved by each method based on the predefined textual criteria. When compared to IC|TC, SSD-LLM demonstrates competitive performance. It is important to note that IC|TC incorporate artificial judgment in its process, which leads to poor scalability when handling large datasets. In contrast, our approach is fully automated. We conduct visualization of our organized subpopulations, as shown in Figure. 4. We can visually observe that the pictures and dimensions are indeed consistent, indicating the effectiveness of our mining and assigning process." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.63, + 0.547, + 0.646 + ], + "angle": 0, + "content": "4.2 Evaluation on Subpopulation Shift" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.785, + 0.794 + ], + "angle": 0, + "content": "Setup We evaluate subpopulation shifts on three commonly used image datasets, Metashift (Cats vs Dogs), Waterbirds (Landbirds vs Waterbirds), and NICO++. We choose Average Accuracy and Worst Group Accuracy as evaluation metrics. To ensure a fair comparison, following [48], we conduct a random search of 16 trials over a joint distribution of all hyperparameters. We then use the validation set to select the best hyperparameters for each algorithm, fix them, and rerun the experiments under five different random seeds to report the final average results. To make the evaluation more realistic, we consider the model selection setting Attributes are unknown in both training and validation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Comparison Methods Following recent benchmarking efforts [48], we compare SSD-LLM with several types of methods: (1) vanilla: ERM [42], (2) Subgroup Robust Methods: GroupDRO [38], LfF [32], JTT [24], LISA [49], (3) Imbalanced" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "Y. Luo et al." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.144, + 0.764, + 0.259 + ], + "angle": 0, + "content": "
Method|CategoriesBoatBirdCarCatDogTruckTopic Error Rate
ImageNet4.330.8111.3311.140.6911.716.72
General Prompt47.8212.1143.5514.2210.1912.6523.42
GPT-Suggest57.5512.8743.5912.7116.3428.1228.53
Domino(Bert)76.2642.2654.2133.8924.5029.5443.44
B2T77.6230.0458.1736.3619.8033.4742.58
SSD-LLM (Ours)79.3145.6760.3432.9726.4839.5747.39
" + }, + { + "type": "table_caption", + "bbox": [ + 0.262, + 0.259, + 0.74, + 0.271 + ], + "angle": 0, + "content": "Table 3: Results of slice discovery on Imagenet-1K with various SDMs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.786, + 0.358 + ], + "angle": 0, + "content": "Learning: ReSample [13], ReWeight [13], Focal [23], CBLoss [5], Bsoftmax [36], (4) Traditional Data Augmentation: Mixup [50], RandAug [3], (5) Diffusion: Class Prompt [40], Class-Attribute Prompt [40], CiP [20]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.787, + 0.584 + ], + "angle": 0, + "content": "Results and Analysis From Table 2, SSD-LLM surpasses previous methods, with a \\(+1.6\\%\\) improvement in average accuracy and \\(+3.3\\%\\) in worst group accuracy across three datasets. The analysis is as follows: (1) Despite being based on conventional ERM, data-based approaches show competitive performance compared to model-based algorithms, highlighting their potential. (2) For diffusion-based methods, class-attribute prompts outperform class prompts, underscoring the importance of understanding dataset imbalanced attributes for effective image generation. However, the need for pre-identifying these attributes emphasizes the value of SSD-LLM, which automates attribute discovery and provides detailed annotations, enhancing performance. (3) The superior performance of CiP over Class Prompt underscores the significance of diverse text prompts. (4) Our method's superior performance results from a comprehensive analysis of subpopulation imbalances within the dataset. The strategic text diversity achieved through attribute sampling and LLM sentence-making effectively addresses subpopulation shifts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.606, + 0.506, + 0.621 + ], + "angle": 0, + "content": "4.3 Evaluation on Slice Discovery" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.786, + 0.689 + ], + "angle": 0, + "content": "Setup In contrast to typical slice discovery tasks, we redefine evaluation pipeline following [11]. In this study, we evaluate bugs found of ImageNet models. Specifically, a classification is deemed incorrect when an image containing target object is erroneously identified by the model as containing an unrelated object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.764 + ], + "angle": 0, + "content": "Comparison Methods Domino [9] represents a state-of-the-art method in slice discovery, which effectively clusters errors identified in the validation set and characterizes them through captions generated automatically. B2T [16] is a recently proposed framework which identifies and interprets visual biases in vision models using keyword extraction from captions of mispredicted images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Results and Analysis We evaluate the effectiveness of our method in slice discovery on 6 representative superclasses in Imagenet. As shown in Table. 3, our SSD-LLM overcomes all other SDMs, including Domino [9], by a significant margin. We find topics given by Domino tend to encounter two unsatisfactory cases: loss of semantics, and missing class information. These cases are also discussed" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.306, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Subpopulation Structure Discovery with Large Language Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.144, + 0.768, + 0.257 + ], + "angle": 0, + "content": "
MethodsIndexCriteria InitializationCriteria RefinementAccuracy
DimensionAttributeActionLocationMood
SSD LLMA10 × 2020 × 20×75.667.371.6
B5 × 2020 × 20×65.260.066.2
C20 × 2020 × 20×75.967.171.3
D10 × 2010 × 40×70.368.969.5
E10 × 2040 × 10×80.270.074.4
F10 × 2020 × 20✓81.770.576.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.258, + 0.785, + 0.285 + ], + "angle": 0, + "content": "Table 4: Ablation study on the number of samples (NUM of rounds × NUM per round), and Criteria Refinement (with or without)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.788, + 0.492 + ], + "angle": 0, + "content": "in [11], where they reasoned this phenomenon into the inherent difficulties of automatic SDMs. However, evidence suggests that our SSD-LLM can handle these problems, while keeping a high error rate and maintaining automation. Specifically, SSD-LLM achieves an average error rate of \\(47.39\\%\\), surpassing Domino [9] by \\(3.95\\%\\). Furthermore, when we trace back to origin dataset, the discovered slice is also very consistent (detailed visualizations included in appendix). Interestingly, we find the data mining process of SSD-LLM is just the same as human data scientists, who take up hypotheses and improve model performance by viewing batches of bad subpopulations. Experiments show that our method overcome the inherent difficulties while maintaining automaton, paving the way for data-centric methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.514, + 0.389, + 0.53 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Hyperparameters of Criteria Initialization/Refinement The \\(\\mathrm{N} \\times \\mathrm{M}\\) in Table. 4 represents the rounds and samples of suggestions. Setting A serves as our baseline. For A, B, and C, results show that insufficient suggested dimensions lead to decreased performance, while enough dimension samples lead to stable performance, as the ICTC task requires only the most appropriate match. Too few suggestions may fail to find suitable dimensions, resulting in mismatched attribute generation. For A, D, and E, results highlight the importance of the number of captions in summarizing attributes. Too many captions can cause some to be overlooked, reducing total identified attributes and performance. The improvement from A to F demonstrates that Criteria Refinement enhances attribute comprehensiveness and final performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Diffusion Generation Strategy From Table. 5, we analyze key components for employing diffusion models to address subpopulation shift. For \\(\\mathrm{A}\\rightarrow \\mathrm{B}\\), we confirm that managing imbalance attributes within datasets helps solve the task. For \\(\\mathrm{B}\\rightarrow \\mathrm{D}\\), we illustrate that a comprehensive subpopulation structure benefits the task. For \\(\\mathrm{D}\\rightarrow \\mathrm{E}\\), balanced subpopulation sampling improves data quality and model training. For \\(\\mathrm{E}\\rightarrow \\mathrm{F}\\), LLM-generated prompts produce more reasonable images, enhancing results. For \\(\\mathrm{F}\\rightarrow \\mathrm{G}\\), LLM suggests additional attributes to enrich subpopulation structures, generating more diverse images and improving model generalization. For \\(\\mathrm{F}\\rightarrow \\mathrm{H}\\), we verify the scaling capability of our SSD-LLM." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "Y. Luo et al." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.144, + 0.785, + 0.262 + ], + "angle": 0, + "content": "
MethodsIndexAttribute ModeSample ModeSD Prompt ModeNumberAverage AccuracyWorst Group Accuracy
ERM————————×184.169.1
SSD-LLMAGT Attribute UnknownrandomDirect×185.971.3
BGT Attribute KnownrandomDirect×189.173.5
CGT Attribute KnownweightedDirect×189.373.8
DSSD-LLM AttributerandomDirect×189.576.2
ESSD-LLM AttributeweightedDirect×189.877.4
FSSD-LLM AttributeweightedLLM Sentence×190.178.3
GSSD-LLM & LLM SuggestweightedLLM Sentence×190.579.1
HSSD-LLM AttributeweightedLLM Sentence×291.179.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.271, + 0.263, + 0.731, + 0.275 + ], + "angle": 0, + "content": "Table 5: Ablation study on the sample mode and SD prompt mode." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.313, + 0.355, + 0.329 + ], + "angle": 0, + "content": "5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.783, + 0.417 + ], + "angle": 0, + "content": "Although our method, SSD-LLM, has already shown effectiveness in various settings, the current exploration of the algorithm is limited to image datasets. Besides, this approach may bear the potential bias from the MLLMs and LLMs. However, it may be reduced from extra human-in-the-loop guidances if available. For future works, we suggest the following promising directions:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.425, + 0.784, + 0.453 + ], + "angle": 0, + "content": "- Structure Format The four-layer subpopulation structure can be expanded to more suitable structures according to specific task requirements." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.455, + 0.784, + 0.483 + ], + "angle": 0, + "content": "- Downstream Tasks SSD-LLM can have more applications in various computer vision and multimodality tasks, e.g. object detection and VQA." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.484, + 0.784, + 0.527 + ], + "angle": 0, + "content": "- Dataset Construction The subpopulation structure obtained from SSD-LLM holds the potential to guide dataset construction with better fairness [44] or further supporting the construction of unbiased datasets [29]." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.529, + 0.784, + 0.572 + ], + "angle": 0, + "content": "- Technical Extensions The core procedures of SSD-LLM, using LLM to conduct group-level summarizations, can be extended to more types of contents including patterns of model hallucinations." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.425, + 0.784, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.595, + 0.36, + 0.61 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.785, + 0.76 + ], + "angle": 0, + "content": "In this work, we present the first systematic exploration of subpopulation structure discovery. We provide a precise definition of subpopulation structure and introduce a fine-grained criteria to determine the structures. We propose SSD-LLM for automatic subpopulation structure discovery incorporating elaborate prompt engineering techniques. SSD-LLM can be combined with subsequent operations to better tackle several downstream tasks, including dataset subpopulation organization, subpopulation shift and slice discovery, with minor Task-specific Tuning. We conduct extensive experiments to verify our proposed methods, demonstrating the remarkable effectiveness and and generality of SSD-LLM." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.781, + 0.405, + 0.798 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Shanghai Zhang is supported by the National Science and Technology Major Project of China (No. 2022ZD0117801)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.306, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Subpopulation Structure Discovery with Large Language Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.214 + ], + "angle": 0, + "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.215, + 0.785, + 0.241 + ], + "angle": 0, + "content": "2. Chen, M., Li, Y., Xu, Q.: Hibug: On human-interpretable model debug. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.242, + 0.785, + 0.295 + ], + "angle": 0, + "content": "3. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 702-703 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.337 + ], + "angle": 0, + "content": "4. Cui, J., Li, Z., Yan, Y., Chen, B., Yuan, L.: Chatlaw: Open-source legal large language model with integrated external knowledge bases. arXiv preprint arXiv:2306.16092 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.785, + 0.378 + ], + "angle": 0, + "content": "5. Cui, Y., Jia, M., Lin, T.Y., Song, Y., Belongie, S.: Class-balanced loss based on effective number of samples. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9268-9277 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.785, + 0.432 + ], + "angle": 0, + "content": "6. Dai, D., Sun, Y., Dong, L., Hao, Y., Ma, S., Sui, Z., Wei, F.: Why can gpt learn in-context? language models implicitly perform gradient descent as meta-optimizers. In: ICLR 2023 Workshop on Mathematical and Empirical Understanding of Foundation Models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.785, + 0.459 + ], + "angle": 0, + "content": "7. Dong, Q., Li, L., Dai, D., Zheng, C., Wu, Z., Chang, B., Sun, X., Xu, J., Sui, Z.: A survey for in-context learning. arXiv preprint arXiv:2301.00234 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.785, + 0.499 + ], + "angle": 0, + "content": "8. Dunlap, L., Umino, A., Zhang, H., Yang, J., Gonzalez, J.E., Darrell, T.: Diversify your vision datasets with automatic diffusion-based augmentation. arXiv preprint arXiv:2305.16289 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.5, + 0.785, + 0.541 + ], + "angle": 0, + "content": "9. Eyuboglu, S., Varma, M., Saab, K., Delbrouck, J.B., Lee-Messer, C., Dunnmon, J., Zou, J., Ré, C.: Domino: Discovering systematic errors with cross-modal embeddings. arXiv preprint arXiv:2203.14960 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.785, + 0.582 + ], + "angle": 0, + "content": "0. Fu, Y., Peng, H., Khot, T., Lapata, M.: Improving language model negotiation with self-play and in-context learning from ai feedback. arXiv preprint arXiv:2305.10142 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.583, + 0.785, + 0.622 + ], + "angle": 0, + "content": "1. Gao, I., Ilharco, G., Lundberg, S., Ribeiro, M.T.: Adaptive testing of computer vision models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4003-4014 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.785, + 0.663 + ], + "angle": 0, + "content": "2. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.664, + 0.785, + 0.69 + ], + "angle": 0, + "content": "3. Japkowicz, N.: The class imbalance problem: Significance and strategies. In: Proc. of the Int'l Conf. on artificial intelligence. vol. 56, pp. 111-117 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.785, + 0.758 + ], + "angle": 0, + "content": "4. Johnson, N., Cabrera, A., Plumb, G., Talwalkar, A.: Where does my model underperform? a human evaluation of slice discovery algorithms. Proceedings of the AAAI Conference on Human Computation and Crowdsourcing 11(1), 65-76 (Nov 2023). https://doi.org/10.1609/hcomp.v11i1.27548, http://dx.doi.org/10.1609/hcomp.v11i1.27548" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "5. Kim, Y., Mo, S., Kim, M., Lee, K., Lee, J., Shin, J.: Bias-to-text: Debi-aising unknown visual biases through language interpretation. arXiv preprint arXiv:2301.11104 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Kim, Y., Mo, S., Kim, M., Lee, K., Lee, J., Shin, J.: Bias-to-text: Debiased unknown visual biases through language interpretation. arXiv preprint arXiv:2301.11104 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Y. Luo et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "17. Kojima, T., Gu, S.S., Reid, M., Matsuo, Y., Iwasawa, Y.: Large language models are zero-shot reasoners. NeurIPS 35, 22199-22213 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "18. Kwon, S., Park, J., Kim, M., Cho, J., Ryu, E.K., Lee, K.: Image clustering conditioned on text criteria. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "19. Kwon, S., Park, J., Kim, M., Cho, J., Ryu, E.K., Lee, K.: Image clustering conditioned on text criteria. arXiv preprint arXiv:2310.18297 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "20. Lei, S., Chen, H., Zhang, S., Zhao, B., Tao, D.: Image captions are natural prompts for text-to-image models. arXiv preprint arXiv:2307.08526 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "21. Li, X., Qiu, X.: Mot: Pre-thinking and recalling enable chatgpt to self-improve with memory-of-thoughts. arXiv preprint arXiv:2305.05181 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.288, + 0.785, + 0.314 + ], + "angle": 0, + "content": "22. Liang, W., Zou, J.: Metashift: A dataset of datasets for evaluating contextual distribution shifts and training conflicts. arXiv preprint arXiv:2202.06523 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.356 + ], + "angle": 0, + "content": "23. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.41 + ], + "angle": 0, + "content": "24. Liu, E.Z., Haghgoo, B., Chen, A.S., Raghunathan, A., Koh, P.W., Sagawa, S., Liang, P., Finn, C.: Just train twice: Improving group robustness without training group information. In: International Conference on Machine Learning. pp. 6781-6792. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.438 + ], + "angle": 0, + "content": "25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "26. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "27. Liu, J., Hu, T., Zhang, Y., Gai, X., Feng, Y., Liu, Z.: A chatgpt aided explainable framework for zero-shot medical image diagnosis. arXiv preprint arXiv:2307.01981 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "28. Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., Neubig, G.: Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys 55(9), 1-35 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.551, + 0.785, + 0.577 + ], + "angle": 0, + "content": "29. Liu, Z., He, K.: A decade's battle on dataset bias: Are we there yet? (2024), https://arxiv.org/abs/2403.08632" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "30. Luo, Y., Tang, Y., Shen, C., Zhou, Z., Dong, B.: Prompt engineering through the lens of optimal control. arXiv preprint arXiv:2310.14201 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "31. Madaan, A., Tandon, N., Gupta, P., Hallinan, S., Gao, L., Wiegreffe, S., Alon, U., Dziri, N., Prabhumoye, S., Yang, Y., et al.: Self-refine: Iterative refinement with self-feedback. arXiv preprint arXiv:2303.17651 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "32. Nam, J., Cha, H., Ahn, S., Lee, J., Shin, J.: Learning from failure: De-biasing classifier from biased classifier. Advances in Neural Information Processing Systems 33, 20673-20684 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.729 + ], + "angle": 0, + "content": "33. Park, J.S., O'Brien, J.C., Cai, C.J., Morris, M.R., Liang, P., Bernstein, M.S.: Generative agents: Interactive simulacra of human behavior. arXiv preprint arXiv:2304.03442 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.785, + 0.757 + ], + "angle": 0, + "content": "34. Qian, C., Cong, X., Yang, C., Chen, W., Su, Y., Xu, J., Liu, Z., Sun, M.: Communicative agents for software development. arXiv preprint arXiv:2307.07924 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "35. Qiao, S., Ou, Y., Zhang, N., Chen, X., Yao, Y., Deng, S., Tan, C., Huang, F., Chen, H.: Reasoning with language model prompting: A survey. arXiv preprint arXiv:2212.09597 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "36. Ren, J., Yu, C., Ma, X., Zhao, H., Yi, S., et al.: Balanced meta-softmax for long-tailed visual recognition. Advances in neural information processing systems 33, 4175-4186 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.306, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Subpopulation Structure Discovery with Large Language Model" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "37. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115, 211-252 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "38. Sagawa, S., Koh, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.259 + ], + "angle": 0, + "content": "39. Santurkar, S., Tsipras, D., Madry, A.: Breeds: Benchmarks for subpopulation shift (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.313 + ], + "angle": 0, + "content": "40. Shipard, J., Wiliem, A., Thanh, K.N., Xiang, W., Fookes, C.: Diversity is definitely needed: Improving model-agnostic zero-shot classification via stable diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 769-778 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.357 + ], + "angle": 0, + "content": "41. Van Gansbeke, W., Vandenhende, S., Georgoulis, S., Proesmans, M., Van Gool, L.: Scan: Learning to classify images without labels. In: European conference on computer vision. pp. 268-285. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.384 + ], + "angle": 0, + "content": "42. Vapnik, V.N.: An overview of statistical learning theory. IEEE transactions on neural networks 10(5), 988-999 (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.411 + ], + "angle": 0, + "content": "43. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.411, + 0.788, + 0.453 + ], + "angle": 0, + "content": "44. Wang, A., Liu, A., Zhang, R., Kleiman, A., Kim, L., Zhao, D., Shirai, I., Narayanan, A., Russakovsky, O.: Revise: A tool for measuring and mitigating bias in visual datasets. International Journal of Computer Vision 130(7), 1790-1810 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.493 + ], + "angle": 0, + "content": "45. Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A., Zhou, D.: Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.536 + ], + "angle": 0, + "content": "46. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.564 + ], + "angle": 0, + "content": "47. Xie, S.M., Raghunathan, A., Liang, P., Ma, T.: An explanation of in-context learning as implicit bayesian inference. arXiv preprint arXiv:2111.02080 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.592 + ], + "angle": 0, + "content": "48. Yang, Y., Zhang, H., Katabi, D., Ghassemi, M.: Change is hard: A closer look at subpopulation shift (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.632 + ], + "angle": 0, + "content": "49. Yao, H., Wang, Y., Li, S., Zhang, L., Liang, W., Zou, J., Finn, C.: Improving out-of-distribution robustness via selective augmentation. In: International Conference on Machine Learning. pp. 25407-25437. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.632, + 0.788, + 0.66 + ], + "angle": 0, + "content": "50. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.715 + ], + "angle": 0, + "content": "51. Zhang, R., Cai, Z., Yang, H., Liu, Z., Gudovskiy, D., Okuno, T., Nakata, Y., Keutzer, K., Chang, B., Du, Y., et al.: Vecaf: Vlm-empowered collaborative active finetuning with training objective awareness. arXiv preprint arXiv:2401.07853 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.715, + 0.788, + 0.743 + ], + "angle": 0, + "content": "52. Zhang, X., He, Y., Xu, R., Yu, H., Shen, Z., Cui, P.: Nico++: Towards better benchmarking for domain generalization (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.743, + 0.788, + 0.785 + ], + "angle": 0, + "content": "53. Zhou, D., Schärli, N., Hou, L., Wei, J., Scales, N., Wang, X., Schuurmans, D., Cui, C., Bousquet, O., Le, Q., et al.: Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.785 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/LLM as Dataset Analyst_ Subpopulation Structure Discovery with Large Language Model/cfcf5c50-f6b7-4af0-a369-8bdd05a30c95_origin.pdf b/2024/LLM as Dataset Analyst_ Subpopulation Structure Discovery with Large Language Model/cfcf5c50-f6b7-4af0-a369-8bdd05a30c95_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9d4f42a3625f8f2db28352338395f0a5c85c686e --- /dev/null +++ b/2024/LLM as Dataset Analyst_ Subpopulation Structure Discovery with Large Language Model/cfcf5c50-f6b7-4af0-a369-8bdd05a30c95_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:beb58d32a4ca33a2ab0f2184e9ccae7c0f1e9e6849134645f588cf7fef870a4a +size 1163356 diff --git a/2024/LLM as Dataset Analyst_ Subpopulation Structure Discovery with Large Language Model/full.md b/2024/LLM as Dataset Analyst_ Subpopulation Structure Discovery with Large Language Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d5f6c83fbd8d08ea201fe94f9ebf50cbcc680023 --- /dev/null +++ b/2024/LLM as Dataset Analyst_ Subpopulation Structure Discovery with Large Language Model/full.md @@ -0,0 +1,293 @@ +# LLM as Dataset Analyst: Subpopulation Structure Discovery with Large Language Model + +Yulin Luo $^{1*}$ , Ruichuan An $^{1,2*}$ , Bocheng Zou $^{1,3}$ + +Yiming Tang $^{1,4}$ , Jiaming Liu $^{1}$ , and Shanghang Zhang $^{1\dagger}$ + +$^{1}$ State Key Laboratory of Multimedia Information Processing, + +School of Computer Science, Peking University + +$^{2}$ Xi'an Jiaotong University $^{3}$ University of Wisconsin-Madison + +$^{4}$ National University of Singapore + +yulin@stu.pku.edu.cn + +Abstract. The distribution of subpopulations is an important property hidden within a dataset. Uncovering and analyzing the subpopulation distribution within datasets provides a comprehensive understanding of the datasets, standing as a powerful tool beneficial to various downstream tasks, including Dataset Subpopulation Organization, Subpopulation Shift, and Slice Discovery. Despite its importance, there has been no work that systematically explores the subpopulation distribution of datasets to our knowledge. To address the limitation and solve all the mentioned tasks in a unified way, we introduce a novel concept of subpopulation structures to represent, analyze, and utilize subpopulation distributions within datasets. To characterize the structures in an interpretable manner, we propose the Subpopulation Structure Discovery with Large Language Models (SSD-LLM) framework, which employs world knowledge and instruction-following capabilities of Large Language Models (LLMs) to linguistically analyze informative image captions and summarize the structures. Furthermore, we propose complete workflows to address downstream tasks, named Task-specific Tuning, showcasing the application of the discovered structure to a spectrum of subpopulation-related tasks, including dataset subpopulation organization, subpopulation shift, and slice discovery. With the help of SSD-LLM, we can structuralize the datasets into subpopulation-level automatically, achieve average $+3.3\%$ worst group accuracy gain compared to previous methods on subpopulation shift benchmark Waterbirds, Metashift and $\mathrm{Nico + + }$ , and also identify more consistent slice topics with a higher model error rate of $3.95\%$ on slice discovery task for ImageNet. The code will be available at https://llm-as-dataset-analyst.github.io/. + +Keywords: Subpopulation Structure Discovery $\cdot$ Large Language Model + +![](images/2ea742e3c8f49a8bfce083b93f20818f43d97ec73e2616ebdf66973d3c795aa7.jpg) + +![](images/b46704d9a56276f8d906e8d11358b16c9e541043ff9eaa82c87a802589239293.jpg) +Fig. 1: (A) The Workflow of Subpopulation Structure Discovery with Large Language Models (SSD-LLM). SSD-LLM can further support several downstream tasks including: (B) Dataset Subpopulation Organization; (C) Subpopulation Shift; (D) Slice discovery. + +# 1 Introduction + +Subpopulation, defined by a set of data points that share common characteristics, is an important concept in machine learning [48]. Many tasks are subpopulation-related. For example, image clustering conditioned on text criteria [18] is to partition an image dataset into different subpopulations based on user-specified criteria, studying subpopulation shift [22, 48, 52] is to mitigate the negative impact of imbalanced subpopulation distributions in the training set on the model, slice discovery [2, 9] is aimed at identifying subpopulations model underperform. + +Summarizing the commonalities of these tasks, we find that analyzing the subpopulation distribution is the key to solving all these problems. If the subpopulation distribution can be characterized, image clustering results under different criteria are naturally obtained [18], additional images can be supplemented to rare subgroups to balance the whole dataset [8], and slices can be easily discovered by statistics error rate on validation set [2]. Despite its importance, existing work [48] lacks systematic exploration of subpopulation distribution. To adjust the issue, for the first time, we propose the concept of subpopulation structure to represent, analyze, and utilize subpopulation distributions within datasets. By definition, a subpopulation structure is a set of hierarchical relations among several subpopulations determined by certain criteria. + +Former works like Metashift [22] and $\mathrm{NICO}++$ [52] have constructed image datasets including the subpopulation information, which organizes the images with respect to some extra attributes, and can be viewed as a class-, attribute-, subpopulation-layer structure. The problem of such a structure is ignoring + +![](images/b14f2039cd2e15854989983efd1152ef0c4722147efc9f7e9efc03ae79b0ba20.jpg) +Fig. 2: Metashift has the same-level attributes Surfboard, Water, and Grass for class Dog, which is irrational due to the possible overlap. As an improvement, we take dimensions into consideration. The class Dog has dimensions including Action, Co-occurrence Object, Location, etc., and in dimension Location, it includes various attributes like Water, Grass, etc, which offers a more appropriate assignment for the samples. + +![](images/c75575809299d06f39836d5a109e823090a7b0f6b6ecfecac583a8556a3b9ea6.jpg) + +the category of attributes (or Dimension), leading to attribute inconsistency and confusion. To solve this issue, we introduce a class-, dimension-, attribute-, and subpopulation-layer structure. The comparison of the two structures can be seen in Figure. 2. By articulating the classification dimensions, this improved structure provides more nuanced attribute assignments. + +Identifying descriptive information within a dataset often requires large amounts of human manufacture [22, 52], which urges the need of automatic workflows to complete the subpopulation structure discovery. However, automatically identifying subpopulation structures within image datasets presents a significant challenge. The approach must be capable of extracting key information from images and summarizing essential content from extensive texts. Furthermore, it necessitates comprehensive world knowledge, enabling a broad understanding of various aspects of the datasets, including diverse categories, common attributes, and the relationships between dimensions and attributes. + +Recently, Large Language Model (LLM) [17, 27, 46] and Multimodal Large Language Model (MLLM) [12, 25] have attracted wide attention due to their superior capacities. LLM has shown extensive world knowledge and remarkable abilities in summarization, instruction following [27], etc. MLLM extends the capabilities of LLM to handle visual inputs. By visual instruction tuning [25], MLLM can verbalize the rich information of images. Motivated by these, we propose a novel framework Subpopulation Structure Discovery with Large Language Model(SSD-LLM), illustrated in Figure. 1, to automatically uncover the structure. The core idea is to generate informative captions from images with MLLM, followed by analyzing and summarizing the subpopulation structure of datasets with LLM. Specifically, we design two elaborate prompt engineering components, Criteria Initialization and Criteria Refinement. The former utilizes a generate-and-select paradigm to summarize dimensions and attributes sequentially. The latter employs self-consistency as an indicator + +to evaluate and refine the criteria. After obtaining complete criteria, each image is assigned to corresponding attributes according to its caption. The final subpopulation structures can be leveraged to finish various downstream tasks with the help of our proposed Task-specific Tuning. In this work, we focus on three application scenarios, i.e. dataset subpopulation organization, subpopulation shift, and slice discovery. We validate the effectiveness of SSD-LLM on these subpopulation-related tasks. For subpopulation shift, we achieve an improvement of $+3.3\%$ in worst group accuracy across three datasets compared to SOTA methods, and for slice discovery, we can identify more consistent slice topics with a higher model error rate of $3.95\%$ . + +Our contributions are summarized as follows: + +- We introduce the concept of subpopulation structure to characterize subpopulation distribution in an interpretable manner for the first time. +- We propose class-dimension-attribute-subpopulation structure, reducing the attribute confusion of the current class-attribute-subpopulation structure. +- We propose Subpopulation Structure Discovery with Large Language Models (SSD-LLM) framework to uncover the underlying subpopulation structure of datasets automatically, with two elaborate prompt engineering components. Criteria Initialization and Criteria Refinement. +- We provide methods for Task-specific Tuning, enabling the application of the structures across a spectrum of subpopulation-related tasks, including dataset subpopulation organization, subpopulation shift, and slice discovery. + +# 2 Related Works + +# 2.1 Hierarchical Structure of Image Datasets + +Recent research has emphasized the need to organize datasets into hierarchical structures allowing for benchmarking various downstream tasks [22, 37, 43, 52]. Metashift [22] builds a collection of 12,868 sets of images related to 410 main subjects and their contexts. NICO++ [52], Waterbirds [43], and ImageNetBG [37] also propose methods for constructing various types of hierarchical datasets. However, the construction of these hierarchical datasets often requires manual annotation, hindering automatic construction. These approaches focus on just a single dimension, such as object context in Metashift, background in Waterbirds, and ImageNetBG, while more practical scenarios may involve multiple dimensions hidden within the comprehensive visual information. + +# 2.2 Extract Information from Image Captions + +Recent works such as ALIA [8], VeCAF [51], Bias2Text [15], and ICTC [19] explore utilizing caption models to obtain information from datasets. ALIA provides a method to augment datasets by generating variations of existing images through captioning and text-to-image models. While ALIA [8] supports dataset improvement, it lacks knowledge about attribute types, bias, or subpopulation + +shift existence. Bias2Text [15] detects dataset bias by transforming images into descriptive captions and keywords. However, without large language model participation, Bias2Text fails to support classification dimension selection and can only differentiate images with basic keywords. More recently, ICTC [19] enables conditional image clustering using an LLM in a straightforward manner. Although ICTC clusters images when given the criterion, it requires human-assigned text prompts. Compared to ICTC, our prompt engineering paradigm supports scalable automatic subpopulation structure dataset organization without human criteria assignment and can generate comprehensive criteria tailored to datasets for various downstream tasks. + +# 2.3 LLM Prompt Engineering + +As the popularity of LLMs has surged, prompt engineering, the process of crafting and refining prompts to guide LLMs towards desired outputs [17,46] has also shown more and more importance. Various prompt engineering methods [28,35] and principles have emerged, and researchers or engineers have explored their applications in a diverse range of downstream tasks [4,10,27,33,34]. In particular, in-context learning has emerged as a pivotal technique, validated both experimentally [7, 21] and theoretically [6, 30, 47]. This approach involves providing the LLM with context information relevant to the task at hand, enabling it to generate more accurate and relevant responses. Least-to-most prompting [53] breaks down complex tasks into smaller, more manageable steps, enhancing LLM's reasoning skill by querying the LLM with more simplified sub-questions. Self-consistency [45] proposes to ensemble multiple responses to the LLM given the same prompt to get enhanced results, suggesting that consistent responses as an indicator for correct problem solving. Self-refining [31] demonstrates that we can use LLMs to refine their outputs by themselves with careful designing of prompts. In this work, we leverage a combination of prompt engineering techniques, including in-context learning, chain-of-thought, self-consistency, and self-refining to tackle subpopulation structure discovery effectively. + +# 3 Method + +In this section, we introduce our proposed method, subpopulation structure discovery with large language models(SSD-LLM). We describe the overall pipeline of SSD-LLM in Section 3.1, outlining how the paradigm automatically discovers the latent subpopulation structures inside the dataset. The process begins with captioning the images in the dataset with an MLLM, detailed in Section 3.2, and proceeds with Criteria Initialization with an LLM in Section 3.3. The paradigm then refines the initialized criteria through a recursive self-refinement procedure, detailed in Section 3.4. Finally, images are assigned to attributes, completing the subpopulation structure discovery process, as elaborated in Section 3.5. The section concludes with a discussion of how to apply the method to various downstream tasks, as presented in Section 3.6. + +![](images/7e2bfa10079f1eb7e656919e3e4e6611b6d2b8f58572c9918abc14a3f134fbc4.jpg) +Fig. 3: Subpopulation Structure Discovery with Large Language Model (SSD-LLM). (Step 1) Multimodality Large Language Model (MLLM) extracts informative captions from images. (Step 2) LLM initializes the criteria with a sample-based generate-and-select paradigm. (Step 3) LLM refines the criteria using self-consistency as an indicator. (Step 4) LLM assigns each caption with specific attributes according to the refined criteria, uncovering the intrinsic subpopulation structures hidden in the dataset. The resulting criteria and subpopulations are used in several downstream tasks. + +# 3.1 Overview + +To automatically discover the subpopulation structures, we propose a novel prompt engineering paradigm that effectively leverages the capabilities of both multimodal large language models (MLLMs) and large language models (LLMs). Our proposed method comprises four key steps (Figure. 3). First, we transform the images into information-rich captions that capture the main information in the images using MLLMs. Second, we employ a novel sample-based prompt engineering method to guide an LLM to produce criteria consisting of dimensions and corresponding attributes organizing the dataset. Third, we prompt the LLM to refine this generated criteria. Last, we assign all the images in the dataset to specific attributes accordingly, uncovering the intrinsic subpopulation structures in the dataset, and paving the way for further analysis about the dataset. Detailed descriptions of each step are provided below. For notations, consistent with various former works [30,46], we denote the operation of getting responses from the language models as $LLM$ and $MLLM$ , and use [,] to represent the concatenation operation of two pacts of texts. + +# 3.2 Caption Extraction + +To begin our approach, we leverage the powerful image captioning capabilities of the MLLM to transform the images into informative and detailed captions. Instead of briefly describing the images, we prompt the MLLM to generate more detailed captions centered around the main subject CLS. The prompt we used in this step is stated as follow: + +$P_{1} =$ "Describe the image of the subject CLS in detail." + +
Step 1 Caption ExtractionStep 2-1 Dimension Generation
Input: Dataset: Dimg, MLLMInput: Captions: C, LLM
Output: Image Captions: COutput: dimensions: Dims
1: for i in range(NumOfIterations)1: for i in range(NumOfIterations)
2: img = Dimg.sample()2: c = C/sample(NumOfSamples)
3: c = MLLM(img, P1)3: S.append(LLM([P2^1, c]))
4: C.append(c)4: end for
5: end for5: Dims = LLMSummary(S)
+ +# 3.3 Criteria Initialization + +To discover the hidden subpopulation structures within the dataset, we employ an LLM to delve into the information-rich captions generated in the previous step. Our objective is to identify certain criteria that effectively partition the images into several distinct subgroups. Beyond simply dividing the dataset into subgroups, we articulate the classification dimension for the partition and record all the resulting attributes generated from the classification process. Along with the class information and the resulting subpopulations, this criteria naturally form a four-layer structure, class-, dimension-, attribute-, and subpopulation-. Noticing criteria encompass multiple dimensions and their corresponding attributes, we adopt a generate-and-select paradigm with the LLM to discover the dimensions and the attributes sequentially. + +To determine the dimensions and attributes, we employ an iterative sampling approach, repeatedly prompting the LLM to propose dimensions and attributes based on batches of image captions. In each iteration, the LLM generates candidate dimensions and attributes, which are subsequently processed through an LLM summarization process. This sample-and-summarize approach effectively addresses the challenges when processing large datasets. Since the number of dimensions that can differentiate images in a dataset is relatively small, and these dimensions have an appearance in numerous images, our sample-based approach effectively identifies relevant dimensions. The prompts we used in this step are stated as follow, omitting Chain-of-thought examples for simplicity. + +$P_2^1 = \text{"Suggest some dimensions that can differentiate the following image captions.}$ + +$P_2^2 = \text{"Suggest a complete criterion to differentiate the following image captions by the given dimension.}"$ + +# 3.4 Criteria Refinement + +To further refine the criteria and ensure its effectiveness in classifying image captions across the dataset, we implement a recursive refining process. This approach proposes a novel method for identifying image captions requiring further refinement utilizing the self-consistency of LLM responses as an indicator [45]. + +Step 2-2 Attribute Generation + +Input: Dimensions: Dims, Captions: $C$ , Large language model: LLM + +Output: Initialized criteria: Criteria + +1: for dim in Dims do +2: for i in range(NumOfIterations) do +3: $c = C$ .sample(NumOfSamples) +4: $S.\mathrm{append}(\mathrm{LLM}([P_2^2, dim, c]))$ +5: end for +6: Attributes $=$ LLMSumarry(S) *list of attributes +7: Criteria[dim] = Attributes +8: $S.\mathrm{reset}()$ +9: end for + +This choice stems from our empirical observation that if an image can be accurately classified according to a particular dimension, it should consistently be classified into the same attribute multiple times. Inconsistent responses, however, suggest that the current criteria require further refinement, either to merge redundant attributes or to include new attributes. The prompts we used in this step are stated as follow and the pseudocode is included in the appendix: + +$P_3^1 = \text{"Classify the caption by the criteria listed below."}$ + +$P_{3}^{2} = \text{"We are unable to classify the following image caption using the provided criteria due to attributes redundancy or disappearance. If redundancy, please prune the redundant attributes. If missing, please suggest an additional attribute that would match the image caption."}$ + +# 3.5 Subpopulation Assignment + +Equipped with the comprehensive criteria, we proceed to systematically assign each image to the specific attributes of each dimension. Images assigned to the same attributes across all dimensions form distinct subgroups within the dataset, revealing the intrinsic subpopulation structures hidden within the data. These subpopulation structures can then be leveraged to perform various downstream tasks, completing our overall pipeline for employing an LLM to analyze the image dataset. The prompt we used in this step is stated as follow: + +$P_{4} =$ "Please assign following caption to one attribute of given dimension." + +Step 4 Subpopulation Assignment + +Input: Captions: $C$ , Large language model: LLM, criteria: Criteria + +Output: Further assignments for each caption $c$ . + +1: for $c$ in $C$ do +2: for [dim, Attributes] in Criteria do +3: c.Assign(LLM([P4,c, dim, Attributes])) +4: end for +5: end for + +![](images/f3520e9641c319c381fc8d660d62f50d3451f8644963803942cfc7492cb72cd9.jpg) +Fig. 4: A visualization of organised subpopulations in a dataset of cats. + +# 3.6 Task-specific Tuning for Downstream Tasks + +Once we have identified the criteria and subpopulations within the dataset, we can leverage this information to tackle several downstream tasks effectively. This includes organizing the subpopulations, which can reveal valuable insights into the data, such as potential subpopulation biases and the presence of long-tail attributes. In practice, the subpopulation structures can be used to improve model performance on various tasks if combined with extra operations, including handling subpopulation shifts and slice discovery. + +Dataset Subpopulation Organization Organizing subpopulations within a dataset can provide beneficial information about the numbers of data points in each subpopulation and can reflect dataset biases and help us identify some long-tail subgroups. We conduct experiments on the task of image clustering conditioned on human-specified criteria to evaluate the quality of subpopulation organization. In specific, when organising the subpopulations of a given image dataset, we first select out the relevant dimensions and then attach attributes assigned by SSD-LLM directly to the images accordingly. Our method, SSD-LLM, automates subpopulation organization within datasets, and thus holds the potential to revolutionize the way hierarchical datasets are constructed. + +Subpopulation Shift Subpopulation shift [48] stands as a common challenge in machine learning, occurring when the proportion of some subpopulations between training and deployment changes, and is shown to be of significant influence to model performances [39, 48]. SSD-LLM, combined with image generation, offers a solution to better handle the scenarios of subpopulation shifts. In our experiments, after we apply SSD-LLM to the datasets, we collect statistics the number of images contained in each subpopulation and utilize diffusion model to generate images for underrepresented subpopulations. Subsequently, we sample attributes from the subpopulation structure for each underrepresented subpopulation and employ LLM to make complete sentences based on these words as the input prompt of a diffusion model. The diffusion model generates images augmented + +
DatasetCriterionSCAN*IC|TCOurs
Stanford 40 ActionAction0.3460.7470.817
Location0.3570.6710.705
Mood0.2760.7460.768
Place365Place0.332-0.696
PPMIMusical Instruction0.5980.9340.955
Cifar10Object0.8390.9110.921
STL10Object0.7980.9860.988
+ +Table 1: Quantitative results of Dataset Subpopulation Organization. Method labeled with * is evaluated by having a human provided ground truth labels, cause the method itself is an unsupervised learning paradigm. + +to the image dataset which further helps to achieve balanced classes and attributes. Moreover, we propose to harness an LLM to suggest extra dimensions and attributes based on the current sets in this task for enriched subpopulation structure, generating more diverse images. + +Slice Discovery Slice discovery is a task aiming at uncovering subpopulations within a dataset where a machine learning model consistently exhibits poor performances. These subpopulations with underperformances, or slices, provide valuable insights into the model's limitations, potential biases [14], and how to improve the performances. SSD-LLM conducts slice discovery for an image dataset with the help of the assigned attributes. In detail, we first calculate the error rates on all subpopulations discovered with the SSD-LLM. Then we identify out the subpopulations with the highest error rate and use the LLM to summarize out discriptions based on the attributes of the subpopulations in the form of texts representing the slice topics, completing the task of slice discovery. + +# 4 Experiments + +We now present experimental results demonstrating the effectiveness of SSD-LLM. In particular, we present the main settings and results in this section and defer extra details, including various visualization result, to appendix. In our experiments, we mainly use LLaVA1.5 [26] for MLLM and GPT-4 [1] for LLM. + +We conduct experiments on Dataset Subpopulation Organization, Subpopulation Shift, and Slice Discovery. Our superior performance underscores the method's efficacy in identifying and analyzing subgroups, further affirming its utility in addressing related challenges. Moreover, it illustrates that the unified paradigm can effectively address a variety of downstream tasks. + +# 4.1 Evaluation on Dataset Subpopulation Organization + +Setup The data assignment process facilitated by SSD-LLM, can be considered a form of clustering. To evaluate the quality of these identified subpopulations, + +
TypeMethodAverage AccuracyWorst Group Accuracy
WaterbirdsMetashiftNico++AverageWaterbirdsMetashiftNico++Average
VanillaERM84.191.276.383.769.182.117.856.3
Subgroup Robust MethodsGroupDRO86.991.574.084.173.183.112.256.1
JTT88.991.277.585.971.282.615.656.5
LfF86.680.477.581.575.072.315.654.3
LISA89.291.475.085.277.079.018.958.3
Imbalanced LearningResample86.292.277.385.270.081.016.755.9
Reweight86.291.573.883.871.983.112.255.7
Focal89.391.673.184.771.681.016.756.4
CBLoss86.891.476.384.874.483.112.256.6
BSoftmax88.491.374.284.674.182.616.757.8
Traditional Data AugmentationMixup89.291.473.084.577.579.014.457.0
RandAug86.390.972.083.171.480.916.756.3
DiffusionClass Prompt85.991.578.085.171.382.718.557.5
Class-Attribute89.191.478.686.473.583.818.858.7
CiP88.091.178.385.873.582.419.358.4
LLM+DiffusionSSD-LLM (Ours)90.593.080.488.079.184.822.162.0
+ +Table 2: Comparison of methods for image classification with subpopulation shifts. + +we assess the clustering accuracy by comparing images against secondary labels that reflect subgroup attributes derived via SSD-LLM. Full information for datasets, text criterion, and model selection can be found in appendix. + +Comparison Methods SCAN [41] is a two-stage clustering method that decouples feature learning and clustering. IC|TC [19] is a new paradigm for image clustering that supports human interaction. It utilizes the given Text Criteria to accurately control the quality of the clustering results. + +Results and Analysis In Table. 1, we report the average accuracy achieved by each method based on the predefined textual criteria. When compared to IC|TC, SSD-LLM demonstrates competitive performance. It is important to note that IC|TC incorporate artificial judgment in its process, which leads to poor scalability when handling large datasets. In contrast, our approach is fully automated. We conduct visualization of our organized subpopulations, as shown in Figure. 4. We can visually observe that the pictures and dimensions are indeed consistent, indicating the effectiveness of our mining and assigning process. + +# 4.2 Evaluation on Subpopulation Shift + +Setup We evaluate subpopulation shifts on three commonly used image datasets, Metashift (Cats vs Dogs), Waterbirds (Landbirds vs Waterbirds), and NICO++. We choose Average Accuracy and Worst Group Accuracy as evaluation metrics. To ensure a fair comparison, following [48], we conduct a random search of 16 trials over a joint distribution of all hyperparameters. We then use the validation set to select the best hyperparameters for each algorithm, fix them, and rerun the experiments under five different random seeds to report the final average results. To make the evaluation more realistic, we consider the model selection setting Attributes are unknown in both training and validation. + +Comparison Methods Following recent benchmarking efforts [48], we compare SSD-LLM with several types of methods: (1) vanilla: ERM [42], (2) Subgroup Robust Methods: GroupDRO [38], LfF [32], JTT [24], LISA [49], (3) Imbalanced + +
Method|CategoriesBoatBirdCarCatDogTruckTopic Error Rate
ImageNet4.330.8111.3311.140.6911.716.72
General Prompt47.8212.1143.5514.2210.1912.6523.42
GPT-Suggest57.5512.8743.5912.7116.3428.1228.53
Domino(Bert)76.2642.2654.2133.8924.5029.5443.44
B2T77.6230.0458.1736.3619.8033.4742.58
SSD-LLM (Ours)79.3145.6760.3432.9726.4839.5747.39
+ +Table 3: Results of slice discovery on Imagenet-1K with various SDMs. + +Learning: ReSample [13], ReWeight [13], Focal [23], CBLoss [5], Bsoftmax [36], (4) Traditional Data Augmentation: Mixup [50], RandAug [3], (5) Diffusion: Class Prompt [40], Class-Attribute Prompt [40], CiP [20]. + +Results and Analysis From Table 2, SSD-LLM surpasses previous methods, with a $+1.6\%$ improvement in average accuracy and $+3.3\%$ in worst group accuracy across three datasets. The analysis is as follows: (1) Despite being based on conventional ERM, data-based approaches show competitive performance compared to model-based algorithms, highlighting their potential. (2) For diffusion-based methods, class-attribute prompts outperform class prompts, underscoring the importance of understanding dataset imbalanced attributes for effective image generation. However, the need for pre-identifying these attributes emphasizes the value of SSD-LLM, which automates attribute discovery and provides detailed annotations, enhancing performance. (3) The superior performance of CiP over Class Prompt underscores the significance of diverse text prompts. (4) Our method's superior performance results from a comprehensive analysis of subpopulation imbalances within the dataset. The strategic text diversity achieved through attribute sampling and LLM sentence-making effectively addresses subpopulation shifts. + +# 4.3 Evaluation on Slice Discovery + +Setup In contrast to typical slice discovery tasks, we redefine evaluation pipeline following [11]. In this study, we evaluate bugs found of ImageNet models. Specifically, a classification is deemed incorrect when an image containing target object is erroneously identified by the model as containing an unrelated object. + +Comparison Methods Domino [9] represents a state-of-the-art method in slice discovery, which effectively clusters errors identified in the validation set and characterizes them through captions generated automatically. B2T [16] is a recently proposed framework which identifies and interprets visual biases in vision models using keyword extraction from captions of mispredicted images. + +Results and Analysis We evaluate the effectiveness of our method in slice discovery on 6 representative superclasses in Imagenet. As shown in Table. 3, our SSD-LLM overcomes all other SDMs, including Domino [9], by a significant margin. We find topics given by Domino tend to encounter two unsatisfactory cases: loss of semantics, and missing class information. These cases are also discussed + +
MethodsIndexCriteria InitializationCriteria RefinementAccuracy
DimensionAttributeActionLocationMood
SSD LLMA10 × 2020 × 20×75.667.371.6
B5 × 2020 × 20×65.260.066.2
C20 × 2020 × 20×75.967.171.3
D10 × 2010 × 40×70.368.969.5
E10 × 2040 × 10×80.270.074.4
F10 × 2020 × 20✓81.770.576.8
+ +Table 4: Ablation study on the number of samples (NUM of rounds × NUM per round), and Criteria Refinement (with or without). + +in [11], where they reasoned this phenomenon into the inherent difficulties of automatic SDMs. However, evidence suggests that our SSD-LLM can handle these problems, while keeping a high error rate and maintaining automation. Specifically, SSD-LLM achieves an average error rate of $47.39\%$ , surpassing Domino [9] by $3.95\%$ . Furthermore, when we trace back to origin dataset, the discovered slice is also very consistent (detailed visualizations included in appendix). Interestingly, we find the data mining process of SSD-LLM is just the same as human data scientists, who take up hypotheses and improve model performance by viewing batches of bad subpopulations. Experiments show that our method overcome the inherent difficulties while maintaining automaton, paving the way for data-centric methods. + +# 4.4 Ablation Study + +Hyperparameters of Criteria Initialization/Refinement The $\mathrm{N} \times \mathrm{M}$ in Table. 4 represents the rounds and samples of suggestions. Setting A serves as our baseline. For A, B, and C, results show that insufficient suggested dimensions lead to decreased performance, while enough dimension samples lead to stable performance, as the ICTC task requires only the most appropriate match. Too few suggestions may fail to find suitable dimensions, resulting in mismatched attribute generation. For A, D, and E, results highlight the importance of the number of captions in summarizing attributes. Too many captions can cause some to be overlooked, reducing total identified attributes and performance. The improvement from A to F demonstrates that Criteria Refinement enhances attribute comprehensiveness and final performance. + +Diffusion Generation Strategy From Table. 5, we analyze key components for employing diffusion models to address subpopulation shift. For $\mathrm{A}\rightarrow \mathrm{B}$ , we confirm that managing imbalance attributes within datasets helps solve the task. For $\mathrm{B}\rightarrow \mathrm{D}$ , we illustrate that a comprehensive subpopulation structure benefits the task. For $\mathrm{D}\rightarrow \mathrm{E}$ , balanced subpopulation sampling improves data quality and model training. For $\mathrm{E}\rightarrow \mathrm{F}$ , LLM-generated prompts produce more reasonable images, enhancing results. For $\mathrm{F}\rightarrow \mathrm{G}$ , LLM suggests additional attributes to enrich subpopulation structures, generating more diverse images and improving model generalization. For $\mathrm{F}\rightarrow \mathrm{H}$ , we verify the scaling capability of our SSD-LLM. + +
MethodsIndexAttribute ModeSample ModeSD Prompt ModeNumberAverage AccuracyWorst Group Accuracy
ERM————————×184.169.1
SSD-LLMAGT Attribute UnknownrandomDirect×185.971.3
BGT Attribute KnownrandomDirect×189.173.5
CGT Attribute KnownweightedDirect×189.373.8
DSSD-LLM AttributerandomDirect×189.576.2
ESSD-LLM AttributeweightedDirect×189.877.4
FSSD-LLM AttributeweightedLLM Sentence×190.178.3
GSSD-LLM & LLM SuggestweightedLLM Sentence×190.579.1
HSSD-LLM AttributeweightedLLM Sentence×291.179.2
+ +Table 5: Ablation study on the sample mode and SD prompt mode. + +# 5 Discussion + +Although our method, SSD-LLM, has already shown effectiveness in various settings, the current exploration of the algorithm is limited to image datasets. Besides, this approach may bear the potential bias from the MLLMs and LLMs. However, it may be reduced from extra human-in-the-loop guidances if available. For future works, we suggest the following promising directions: + +- Structure Format The four-layer subpopulation structure can be expanded to more suitable structures according to specific task requirements. +- Downstream Tasks SSD-LLM can have more applications in various computer vision and multimodality tasks, e.g. object detection and VQA. +- Dataset Construction The subpopulation structure obtained from SSD-LLM holds the potential to guide dataset construction with better fairness [44] or further supporting the construction of unbiased datasets [29]. +- Technical Extensions The core procedures of SSD-LLM, using LLM to conduct group-level summarizations, can be extended to more types of contents including patterns of model hallucinations. + +# 6 Conclusion + +In this work, we present the first systematic exploration of subpopulation structure discovery. We provide a precise definition of subpopulation structure and introduce a fine-grained criteria to determine the structures. We propose SSD-LLM for automatic subpopulation structure discovery incorporating elaborate prompt engineering techniques. SSD-LLM can be combined with subsequent operations to better tackle several downstream tasks, including dataset subpopulation organization, subpopulation shift and slice discovery, with minor Task-specific Tuning. We conduct extensive experiments to verify our proposed methods, demonstrating the remarkable effectiveness and and generality of SSD-LLM. + +# Acknowledgements + +Shanghai Zhang is supported by the National Science and Technology Major Project of China (No. 2022ZD0117801). + +# References + +1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) +2. Chen, M., Li, Y., Xu, Q.: Hibug: On human-interpretable model debug. In: NeurIPS (2023) +3. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 702-703 (2020) +4. Cui, J., Li, Z., Yan, Y., Chen, B., Yuan, L.: Chatlaw: Open-source legal large language model with integrated external knowledge bases. arXiv preprint arXiv:2306.16092 (2023) +5. Cui, Y., Jia, M., Lin, T.Y., Song, Y., Belongie, S.: Class-balanced loss based on effective number of samples. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9268-9277 (2019) +6. Dai, D., Sun, Y., Dong, L., Hao, Y., Ma, S., Sui, Z., Wei, F.: Why can gpt learn in-context? language models implicitly perform gradient descent as meta-optimizers. In: ICLR 2023 Workshop on Mathematical and Empirical Understanding of Foundation Models (2023) +7. Dong, Q., Li, L., Dai, D., Zheng, C., Wu, Z., Chang, B., Sun, X., Xu, J., Sui, Z.: A survey for in-context learning. arXiv preprint arXiv:2301.00234 (2022) +8. Dunlap, L., Umino, A., Zhang, H., Yang, J., Gonzalez, J.E., Darrell, T.: Diversify your vision datasets with automatic diffusion-based augmentation. arXiv preprint arXiv:2305.16289 (2023) +9. Eyuboglu, S., Varma, M., Saab, K., Delbrouck, J.B., Lee-Messer, C., Dunnmon, J., Zou, J., Ré, C.: Domino: Discovering systematic errors with cross-modal embeddings. arXiv preprint arXiv:2203.14960 (2022) +0. Fu, Y., Peng, H., Khot, T., Lapata, M.: Improving language model negotiation with self-play and in-context learning from ai feedback. arXiv preprint arXiv:2305.10142 (2023) +1. Gao, I., Ilharco, G., Lundberg, S., Ribeiro, M.T.: Adaptive testing of computer vision models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4003-4014 (2023) +2. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023) +3. Japkowicz, N.: The class imbalance problem: Significance and strategies. In: Proc. of the Int'l Conf. on artificial intelligence. vol. 56, pp. 111-117 (2000) +4. Johnson, N., Cabrera, A., Plumb, G., Talwalkar, A.: Where does my model underperform? a human evaluation of slice discovery algorithms. Proceedings of the AAAI Conference on Human Computation and Crowdsourcing 11(1), 65-76 (Nov 2023). https://doi.org/10.1609/hcomp.v11i1.27548, http://dx.doi.org/10.1609/hcomp.v11i1.27548 +5. Kim, Y., Mo, S., Kim, M., Lee, K., Lee, J., Shin, J.: Bias-to-text: Debi-aising unknown visual biases through language interpretation. arXiv preprint arXiv:2301.11104 (2023) +6. Kim, Y., Mo, S., Kim, M., Lee, K., Lee, J., Shin, J.: Bias-to-text: Debiased unknown visual biases through language interpretation. arXiv preprint arXiv:2301.11104 (2023) + +17. Kojima, T., Gu, S.S., Reid, M., Matsuo, Y., Iwasawa, Y.: Large language models are zero-shot reasoners. NeurIPS 35, 22199-22213 (2022) +18. Kwon, S., Park, J., Kim, M., Cho, J., Ryu, E.K., Lee, K.: Image clustering conditioned on text criteria. In: ICLR (2023) +19. Kwon, S., Park, J., Kim, M., Cho, J., Ryu, E.K., Lee, K.: Image clustering conditioned on text criteria. arXiv preprint arXiv:2310.18297 (2023) +20. Lei, S., Chen, H., Zhang, S., Zhao, B., Tao, D.: Image captions are natural prompts for text-to-image models. arXiv preprint arXiv:2307.08526 (2023) +21. Li, X., Qiu, X.: Mot: Pre-thinking and recalling enable chatgpt to self-improve with memory-of-thoughts. arXiv preprint arXiv:2305.05181 (2023) +22. Liang, W., Zou, J.: Metashift: A dataset of datasets for evaluating contextual distribution shifts and training conflicts. arXiv preprint arXiv:2202.06523 (2022) +23. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017) +24. Liu, E.Z., Haghgoo, B., Chen, A.S., Raghunathan, A., Koh, P.W., Sagawa, S., Liang, P., Finn, C.: Just train twice: Improving group robustness without training group information. In: International Conference on Machine Learning. pp. 6781-6792. PMLR (2021) +25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023) +26. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024) +27. Liu, J., Hu, T., Zhang, Y., Gai, X., Feng, Y., Liu, Z.: A chatgpt aided explainable framework for zero-shot medical image diagnosis. arXiv preprint arXiv:2307.01981 (2023) +28. Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., Neubig, G.: Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys 55(9), 1-35 (2023) +29. Liu, Z., He, K.: A decade's battle on dataset bias: Are we there yet? (2024), https://arxiv.org/abs/2403.08632 +30. Luo, Y., Tang, Y., Shen, C., Zhou, Z., Dong, B.: Prompt engineering through the lens of optimal control. arXiv preprint arXiv:2310.14201 (2023) +31. Madaan, A., Tandon, N., Gupta, P., Hallinan, S., Gao, L., Wiegreffe, S., Alon, U., Dziri, N., Prabhumoye, S., Yang, Y., et al.: Self-refine: Iterative refinement with self-feedback. arXiv preprint arXiv:2303.17651 (2023) +32. Nam, J., Cha, H., Ahn, S., Lee, J., Shin, J.: Learning from failure: De-biasing classifier from biased classifier. Advances in Neural Information Processing Systems 33, 20673-20684 (2020) +33. Park, J.S., O'Brien, J.C., Cai, C.J., Morris, M.R., Liang, P., Bernstein, M.S.: Generative agents: Interactive simulacra of human behavior. arXiv preprint arXiv:2304.03442 (2023) +34. Qian, C., Cong, X., Yang, C., Chen, W., Su, Y., Xu, J., Liu, Z., Sun, M.: Communicative agents for software development. arXiv preprint arXiv:2307.07924 (2023) +35. Qiao, S., Ou, Y., Zhang, N., Chen, X., Yao, Y., Deng, S., Tan, C., Huang, F., Chen, H.: Reasoning with language model prompting: A survey. arXiv preprint arXiv:2212.09597 (2022) +36. Ren, J., Yu, C., Ma, X., Zhao, H., Yi, S., et al.: Balanced meta-softmax for long-tailed visual recognition. Advances in neural information processing systems 33, 4175-4186 (2020) + +37. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115, 211-252 (2015) +38. Sagawa, S., Koh, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731 (2019) +39. Santurkar, S., Tsipras, D., Madry, A.: Breeds: Benchmarks for subpopulation shift (2020) +40. Shipard, J., Wiliem, A., Thanh, K.N., Xiang, W., Fookes, C.: Diversity is definitely needed: Improving model-agnostic zero-shot classification via stable diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 769-778 (2023) +41. Van Gansbeke, W., Vandenhende, S., Georgoulis, S., Proesmans, M., Van Gool, L.: Scan: Learning to classify images without labels. In: European conference on computer vision. pp. 268-285. Springer (2020) +42. Vapnik, V.N.: An overview of statistical learning theory. IEEE transactions on neural networks 10(5), 988-999 (1999) +43. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011) +44. Wang, A., Liu, A., Zhang, R., Kleiman, A., Kim, L., Zhao, D., Shirai, I., Narayanan, A., Russakovsky, O.: Revise: A tool for measuring and mitigating bias in visual datasets. International Journal of Computer Vision 130(7), 1790-1810 (2022) +45. Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A., Zhou, D.: Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171 (2022) +46. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022) +47. Xie, S.M., Raghunathan, A., Liang, P., Ma, T.: An explanation of in-context learning as implicit bayesian inference. arXiv preprint arXiv:2111.02080 (2021) +48. Yang, Y., Zhang, H., Katabi, D., Ghassemi, M.: Change is hard: A closer look at subpopulation shift (2023) +49. Yao, H., Wang, Y., Li, S., Zhang, L., Liang, W., Zou, J., Finn, C.: Improving out-of-distribution robustness via selective augmentation. In: International Conference on Machine Learning. pp. 25407-25437. PMLR (2022) +50. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017) +51. Zhang, R., Cai, Z., Yang, H., Liu, Z., Gudovskiy, D., Okuno, T., Nakata, Y., Keutzer, K., Chang, B., Du, Y., et al.: Vecaf: Vlm-empowered collaborative active finetuning with training objective awareness. arXiv preprint arXiv:2401.07853 (2024) +52. Zhang, X., He, Y., Xu, R., Yu, H., Shen, Z., Cui, P.: Nico++: Towards better benchmarking for domain generalization (2022) +53. Zhou, D., Schärli, N., Hou, L., Wei, J., Scales, N., Wang, X., Schuurmans, D., Cui, C., Bousquet, O., Le, Q., et al.: Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625 (2022) \ No newline at end of file diff --git a/2024/LLM as Dataset Analyst_ Subpopulation Structure Discovery with Large Language Model/images.zip b/2024/LLM as Dataset Analyst_ Subpopulation Structure Discovery with Large Language Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1c8817fae732ac0c874b73b14acac026f5a39b3c --- /dev/null +++ b/2024/LLM as Dataset Analyst_ Subpopulation Structure Discovery with Large Language Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef6e048d163e7a368e40cfa3a0b502b3fe6232353755570f6338fbed9e658f04 +size 562115 diff --git a/2024/LLM as Dataset Analyst_ Subpopulation Structure Discovery with Large Language Model/layout.json b/2024/LLM as Dataset Analyst_ Subpopulation Structure Discovery with Large Language Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..38a3028462668eab7165e9f6abcbffd9653ec0e8 --- /dev/null +++ b/2024/LLM as Dataset Analyst_ Subpopulation Structure Discovery with Large Language Model/layout.json @@ -0,0 +1,8332 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 112, + 478, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 112, + 478, + 148 + ], + "spans": [ + { + "bbox": [ + 136, + 112, + 478, + 148 + ], + "type": "text", + "content": "LLM as Dataset Analyst: Subpopulation Structure Discovery with Large Language Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 184, + 171, + 412, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 171, + 412, + 182 + ], + "spans": [ + { + "bbox": [ + 184, + 171, + 412, + 182 + ], + "type": "text", + "content": "Yulin Luo" + }, + { + "bbox": [ + 184, + 171, + 412, + 182 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 184, + 171, + 412, + 182 + ], + "type": "text", + "content": ", Ruichuan An" + }, + { + "bbox": [ + 184, + 171, + 412, + 182 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 184, + 171, + 412, + 182 + ], + "type": "text", + "content": ", Bocheng Zou" + }, + { + "bbox": [ + 184, + 171, + 412, + 182 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 184, + 182, + 428, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 182, + 428, + 196 + ], + "spans": [ + { + "bbox": [ + 184, + 182, + 428, + 196 + ], + "type": "text", + "content": "Yiming Tang" + }, + { + "bbox": [ + 184, + 182, + 428, + 196 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 184, + 182, + 428, + 196 + ], + "type": "text", + "content": ", Jiaming Liu" + }, + { + "bbox": [ + 184, + 182, + 428, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 184, + 182, + 428, + 196 + ], + "type": "text", + "content": ", and Shanghang Zhang" + }, + { + "bbox": [ + 184, + 182, + 428, + 196 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 205, + 432, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 205, + 432, + 216 + ], + "spans": [ + { + "bbox": [ + 179, + 205, + 432, + 216 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 179, + 205, + 432, + 216 + ], + "type": "text", + "content": "State Key Laboratory of Multimedia Information Processing," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 211, + 217, + 402, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 217, + 402, + 227 + ], + "spans": [ + { + "bbox": [ + 211, + 217, + 402, + 227 + ], + "type": "text", + "content": "School of Computer Science, Peking University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 182, + 227, + 431, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 227, + 431, + 238 + ], + "spans": [ + { + "bbox": [ + 182, + 227, + 431, + 238 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 182, + 227, + 431, + 238 + ], + "type": "text", + "content": "Xi'an Jiaotong University " + }, + { + "bbox": [ + 182, + 227, + 431, + 238 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 182, + 227, + 431, + 238 + ], + "type": "text", + "content": "University of Wisconsin-Madison" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 238, + 238, + 376, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 238, + 376, + 249 + ], + "spans": [ + { + "bbox": [ + 238, + 238, + 376, + 249 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 238, + 238, + 376, + 249 + ], + "type": "text", + "content": "National University of Singapore" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 258, + 250, + 354, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 250, + 354, + 260 + ], + "spans": [ + { + "bbox": [ + 258, + 250, + 354, + 260 + ], + "type": "text", + "content": "yulin@stu.pku.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 315, + 455, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 315, + 455, + 601 + ], + "spans": [ + { + "bbox": [ + 159, + 315, + 455, + 601 + ], + "type": "text", + "content": "Abstract. The distribution of subpopulations is an important property hidden within a dataset. Uncovering and analyzing the subpopulation distribution within datasets provides a comprehensive understanding of the datasets, standing as a powerful tool beneficial to various downstream tasks, including Dataset Subpopulation Organization, Subpopulation Shift, and Slice Discovery. Despite its importance, there has been no work that systematically explores the subpopulation distribution of datasets to our knowledge. To address the limitation and solve all the mentioned tasks in a unified way, we introduce a novel concept of subpopulation structures to represent, analyze, and utilize subpopulation distributions within datasets. To characterize the structures in an interpretable manner, we propose the Subpopulation Structure Discovery with Large Language Models (SSD-LLM) framework, which employs world knowledge and instruction-following capabilities of Large Language Models (LLMs) to linguistically analyze informative image captions and summarize the structures. Furthermore, we propose complete workflows to address downstream tasks, named Task-specific Tuning, showcasing the application of the discovered structure to a spectrum of subpopulation-related tasks, including dataset subpopulation organization, subpopulation shift, and slice discovery. With the help of SSD-LLM, we can structuralize the datasets into subpopulation-level automatically, achieve average " + }, + { + "bbox": [ + 159, + 315, + 455, + 601 + ], + "type": "inline_equation", + "content": "+3.3\\%" + }, + { + "bbox": [ + 159, + 315, + 455, + 601 + ], + "type": "text", + "content": " worst group accuracy gain compared to previous methods on subpopulation shift benchmark Waterbirds, Metashift and " + }, + { + "bbox": [ + 159, + 315, + 455, + 601 + ], + "type": "inline_equation", + "content": "\\mathrm{Nico + + }" + }, + { + "bbox": [ + 159, + 315, + 455, + 601 + ], + "type": "text", + "content": ", and also identify more consistent slice topics with a higher model error rate of " + }, + { + "bbox": [ + 159, + 315, + 455, + 601 + ], + "type": "inline_equation", + "content": "3.95\\%" + }, + { + "bbox": [ + 159, + 315, + 455, + 601 + ], + "type": "text", + "content": " on slice discovery task for ImageNet. The code will be available at https://llm-as-dataset-analyst.github.io/." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 160, + 613, + 454, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 613, + 454, + 625 + ], + "spans": [ + { + "bbox": [ + 160, + 613, + 454, + 625 + ], + "type": "text", + "content": "Keywords: Subpopulation Structure Discovery " + }, + { + "bbox": [ + 160, + 613, + 454, + 625 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 613, + 454, + 625 + ], + "type": "text", + "content": " Large Language Model" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 642, + 227, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 642, + 227, + 653 + ], + "spans": [ + { + "bbox": [ + 134, + 642, + 227, + 653 + ], + "type": "text", + "content": "* Equal Contribution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 135, + 653, + 238, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 653, + 238, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 653, + 238, + 665 + ], + "type": "text", + "content": "† Corresponding Author." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 115, + 473, + 183 + ], + "blocks": [ + { + "bbox": [ + 138, + 115, + 473, + 183 + ], + "lines": [ + { + "bbox": [ + 138, + 115, + 473, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 473, + 183 + ], + "type": "image", + "image_path": "2ea742e3c8f49a8bfce083b93f20818f43d97ec73e2616ebdf66973d3c795aa7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 184, + 473, + 311 + ], + "blocks": [ + { + "bbox": [ + 138, + 184, + 473, + 311 + ], + "lines": [ + { + "bbox": [ + 138, + 184, + 473, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 473, + 311 + ], + "type": "image", + "image_path": "b46704d9a56276f8d906e8d11358b16c9e541043ff9eaa82c87a802589239293.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 320, + 482, + 355 + ], + "lines": [ + { + "bbox": [ + 132, + 320, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 482, + 355 + ], + "type": "text", + "content": "Fig. 1: (A) The Workflow of Subpopulation Structure Discovery with Large Language Models (SSD-LLM). SSD-LLM can further support several downstream tasks including: (B) Dataset Subpopulation Organization; (C) Subpopulation Shift; (D) Slice discovery." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 376, + 230, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 230, + 389 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 230, + 389 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 479, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 479, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 479, + 485 + ], + "type": "text", + "content": "Subpopulation, defined by a set of data points that share common characteristics, is an important concept in machine learning [48]. Many tasks are subpopulation-related. For example, image clustering conditioned on text criteria [18] is to partition an image dataset into different subpopulations based on user-specified criteria, studying subpopulation shift [22, 48, 52] is to mitigate the negative impact of imbalanced subpopulation distributions in the training set on the model, slice discovery [2, 9] is aimed at identifying subpopulations model underperform." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 480, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 480, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 480, + 617 + ], + "type": "text", + "content": "Summarizing the commonalities of these tasks, we find that analyzing the subpopulation distribution is the key to solving all these problems. If the subpopulation distribution can be characterized, image clustering results under different criteria are naturally obtained [18], additional images can be supplemented to rare subgroups to balance the whole dataset [8], and slices can be easily discovered by statistics error rate on validation set [2]. Despite its importance, existing work [48] lacks systematic exploration of subpopulation distribution. To adjust the issue, for the first time, we propose the concept of subpopulation structure to represent, analyze, and utilize subpopulation distributions within datasets. By definition, a subpopulation structure is a set of hierarchical relations among several subpopulations determined by certain criteria." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Former works like Metashift [22] and " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{NICO}++" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " [52] have constructed image datasets including the subpopulation information, which organizes the images with respect to some extra attributes, and can be viewed as a class-, attribute-, subpopulation-layer structure. The problem of such a structure is ignoring" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Y. Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 243, + 231 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 243, + 231 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 243, + 231 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 243, + 231 + ], + "type": "image", + "image_path": "b14f2039cd2e15854989983efd1152ef0c4722147efc9f7e9efc03ae79b0ba20.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 241, + 482, + 298 + ], + "lines": [ + { + "bbox": [ + 132, + 241, + 482, + 298 + ], + "spans": [ + { + "bbox": [ + 132, + 241, + 482, + 298 + ], + "type": "text", + "content": "Fig. 2: Metashift has the same-level attributes Surfboard, Water, and Grass for class Dog, which is irrational due to the possible overlap. As an improvement, we take dimensions into consideration. The class Dog has dimensions including Action, Co-occurrence Object, Location, etc., and in dimension Location, it includes various attributes like Water, Grass, etc, which offers a more appropriate assignment for the samples." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 248, + 114, + 478, + 232 + ], + "blocks": [ + { + "bbox": [ + 248, + 114, + 478, + 232 + ], + "lines": [ + { + "bbox": [ + 248, + 114, + 478, + 232 + ], + "spans": [ + { + "bbox": [ + 248, + 114, + 478, + 232 + ], + "type": "image", + "image_path": "c75575809299d06f39836d5a109e823090a7b0f6b6ecfecac583a8556a3b9ea6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 326, + 480, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 480, + 386 + ], + "type": "text", + "content": "the category of attributes (or Dimension), leading to attribute inconsistency and confusion. To solve this issue, we introduce a class-, dimension-, attribute-, and subpopulation-layer structure. The comparison of the two structures can be seen in Figure. 2. By articulating the classification dimensions, this improved structure provides more nuanced attribute assignments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 388, + 479, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 479, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 479, + 495 + ], + "type": "text", + "content": "Identifying descriptive information within a dataset often requires large amounts of human manufacture [22, 52], which urges the need of automatic workflows to complete the subpopulation structure discovery. However, automatically identifying subpopulation structures within image datasets presents a significant challenge. The approach must be capable of extracting key information from images and summarizing essential content from extensive texts. Furthermore, it necessitates comprehensive world knowledge, enabling a broad understanding of various aspects of the datasets, including diverse categories, common attributes, and the relationships between dimensions and attributes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Recently, Large Language Model (LLM) [17, 27, 46] and Multimodal Large Language Model (MLLM) [12, 25] have attracted wide attention due to their superior capacities. LLM has shown extensive world knowledge and remarkable abilities in summarization, instruction following [27], etc. MLLM extends the capabilities of LLM to handle visual inputs. By visual instruction tuning [25], MLLM can verbalize the rich information of images. Motivated by these, we propose a novel framework Subpopulation Structure Discovery with Large Language Model(SSD-LLM), illustrated in Figure. 1, to automatically uncover the structure. The core idea is to generate informative captions from images with MLLM, followed by analyzing and summarizing the subpopulation structure of datasets with LLM. Specifically, we design two elaborate prompt engineering components, Criteria Initialization and Criteria Refinement. The former utilizes a generate-and-select paradigm to summarize dimensions and attributes sequentially. The latter employs self-consistency as an indicator" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "type": "text", + "content": "Subpopulation Structure Discovery with Large Language Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "to evaluate and refine the criteria. After obtaining complete criteria, each image is assigned to corresponding attributes according to its caption. The final subpopulation structures can be leveraged to finish various downstream tasks with the help of our proposed Task-specific Tuning. In this work, we focus on three application scenarios, i.e. dataset subpopulation organization, subpopulation shift, and slice discovery. We validate the effectiveness of SSD-LLM on these subpopulation-related tasks. For subpopulation shift, we achieve an improvement of " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "+3.3\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " in worst group accuracy across three datasets compared to SOTA methods, and for slice discovery, we can identify more consistent slice topics with a higher model error rate of " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "3.95\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 236, + 347, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 236, + 347, + 247 + ], + "spans": [ + { + "bbox": [ + 146, + 236, + 347, + 247 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 255, + 480, + 384 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 255, + 479, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 255, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 479, + 277 + ], + "type": "text", + "content": "- We introduce the concept of subpopulation structure to characterize subpopulation distribution in an interpretable manner for the first time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 279, + 480, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 279, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 138, + 279, + 480, + 301 + ], + "type": "text", + "content": "- We propose class-dimension-attribute-subpopulation structure, reducing the attribute confusion of the current class-attribute-subpopulation structure." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 303, + 480, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 303, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 303, + 480, + 347 + ], + "type": "text", + "content": "- We propose Subpopulation Structure Discovery with Large Language Models (SSD-LLM) framework to uncover the underlying subpopulation structure of datasets automatically, with two elaborate prompt engineering components. Criteria Initialization and Criteria Refinement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 350, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 480, + 384 + ], + "type": "text", + "content": "- We provide methods for Task-specific Tuning, enabling the application of the structures across a spectrum of subpopulation-related tasks, including dataset subpopulation organization, subpopulation shift, and slice discovery." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 404, + 242, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 404, + 242, + 416 + ], + "spans": [ + { + "bbox": [ + 132, + 404, + 242, + 416 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 429, + 367, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 367, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 367, + 441 + ], + "type": "text", + "content": "2.1 Hierarchical Structure of Image Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 569 + ], + "type": "text", + "content": "Recent research has emphasized the need to organize datasets into hierarchical structures allowing for benchmarking various downstream tasks [22, 37, 43, 52]. Metashift [22] builds a collection of 12,868 sets of images related to 410 main subjects and their contexts. NICO++ [52], Waterbirds [43], and ImageNetBG [37] also propose methods for constructing various types of hierarchical datasets. However, the construction of these hierarchical datasets often requires manual annotation, hindering automatic construction. These approaches focus on just a single dimension, such as object context in Metashift, background in Waterbirds, and ImageNetBG, while more practical scenarios may involve multiple dimensions hidden within the comprehensive visual information." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 586, + 372, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 372, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 372, + 597 + ], + "type": "text", + "content": "2.2 Extract Information from Image Captions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "Recent works such as ALIA [8], VeCAF [51], Bias2Text [15], and ICTC [19] explore utilizing caption models to obtain information from datasets. ALIA provides a method to augment datasets by generating variations of existing images through captioning and text-to-image models. While ALIA [8] supports dataset improvement, it lacks knowledge about attribute types, bias, or subpopulation" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Y. Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "shift existence. Bias2Text [15] detects dataset bias by transforming images into descriptive captions and keywords. However, without large language model participation, Bias2Text fails to support classification dimension selection and can only differentiate images with basic keywords. More recently, ICTC [19] enables conditional image clustering using an LLM in a straightforward manner. Although ICTC clusters images when given the criterion, it requires human-assigned text prompts. Compared to ICTC, our prompt engineering paradigm supports scalable automatic subpopulation structure dataset organization without human criteria assignment and can generate comprehensive criteria tailored to datasets for various downstream tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 254, + 293, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 254, + 293, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 254, + 293, + 266 + ], + "type": "text", + "content": "2.3 LLM Prompt Engineering" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 274, + 482, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 490 + ], + "type": "text", + "content": "As the popularity of LLMs has surged, prompt engineering, the process of crafting and refining prompts to guide LLMs towards desired outputs [17,46] has also shown more and more importance. Various prompt engineering methods [28,35] and principles have emerged, and researchers or engineers have explored their applications in a diverse range of downstream tasks [4,10,27,33,34]. In particular, in-context learning has emerged as a pivotal technique, validated both experimentally [7, 21] and theoretically [6, 30, 47]. This approach involves providing the LLM with context information relevant to the task at hand, enabling it to generate more accurate and relevant responses. Least-to-most prompting [53] breaks down complex tasks into smaller, more manageable steps, enhancing LLM's reasoning skill by querying the LLM with more simplified sub-questions. Self-consistency [45] proposes to ensemble multiple responses to the LLM given the same prompt to get enhanced results, suggesting that consistent responses as an indicator for correct problem solving. Self-refining [31] demonstrates that we can use LLMs to refine their outputs by themselves with careful designing of prompts. In this work, we leverage a combination of prompt engineering techniques, including in-context learning, chain-of-thought, self-consistency, and self-refining to tackle subpopulation structure discovery effectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 508, + 202, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 508, + 202, + 521 + ], + "spans": [ + { + "bbox": [ + 131, + 508, + 202, + 521 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "In this section, we introduce our proposed method, subpopulation structure discovery with large language models(SSD-LLM). We describe the overall pipeline of SSD-LLM in Section 3.1, outlining how the paradigm automatically discovers the latent subpopulation structures inside the dataset. The process begins with captioning the images in the dataset with an MLLM, detailed in Section 3.2, and proceeds with Criteria Initialization with an LLM in Section 3.3. The paradigm then refines the initialized criteria through a recursive self-refinement procedure, detailed in Section 3.4. Finally, images are assigned to attributes, completing the subpopulation structure discovery process, as elaborated in Section 3.5. The section concludes with a discussion of how to apply the method to various downstream tasks, as presented in Section 3.6." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "type": "text", + "content": "Subpopulation Structure Discovery with Large Language Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 116, + 474, + 244 + ], + "blocks": [ + { + "bbox": [ + 137, + 116, + 474, + 244 + ], + "lines": [ + { + "bbox": [ + 137, + 116, + 474, + 244 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 474, + 244 + ], + "type": "image", + "image_path": "7e2bfa10079f1eb7e656919e3e4e6611b6d2b8f58572c9918abc14a3f134fbc4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 254, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 331 + ], + "type": "text", + "content": "Fig. 3: Subpopulation Structure Discovery with Large Language Model (SSD-LLM). (Step 1) Multimodality Large Language Model (MLLM) extracts informative captions from images. (Step 2) LLM initializes the criteria with a sample-based generate-and-select paradigm. (Step 3) LLM refines the criteria using self-consistency as an indicator. (Step 4) LLM assigns each caption with specific attributes according to the refined criteria, uncovering the intrinsic subpopulation structures hidden in the dataset. The resulting criteria and subpopulations are used in several downstream tasks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 356, + 208, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 356, + 208, + 366 + ], + "spans": [ + { + "bbox": [ + 132, + 356, + 208, + 366 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 556 + ], + "type": "text", + "content": "To automatically discover the subpopulation structures, we propose a novel prompt engineering paradigm that effectively leverages the capabilities of both multimodal large language models (MLLMs) and large language models (LLMs). Our proposed method comprises four key steps (Figure. 3). First, we transform the images into information-rich captions that capture the main information in the images using MLLMs. Second, we employ a novel sample-based prompt engineering method to guide an LLM to produce criteria consisting of dimensions and corresponding attributes organizing the dataset. Third, we prompt the LLM to refine this generated criteria. Last, we assign all the images in the dataset to specific attributes accordingly, uncovering the intrinsic subpopulation structures in the dataset, and paving the way for further analysis about the dataset. Detailed descriptions of each step are provided below. For notations, consistent with various former works [30,46], we denote the operation of getting responses from the language models as " + }, + { + "bbox": [ + 130, + 376, + 482, + 556 + ], + "type": "inline_equation", + "content": "LLM" + }, + { + "bbox": [ + 130, + 376, + 482, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 376, + 482, + 556 + ], + "type": "inline_equation", + "content": "MLLM" + }, + { + "bbox": [ + 130, + 376, + 482, + 556 + ], + "type": "text", + "content": ", and use [,] to represent the concatenation operation of two pacts of texts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 573, + 259, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 573, + 259, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 573, + 259, + 586 + ], + "type": "text", + "content": "3.2 Caption Extraction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 653 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 653 + ], + "type": "text", + "content": "To begin our approach, we leverage the powerful image captioning capabilities of the MLLM to transform the images into informative and detailed captions. Instead of briefly describing the images, we prompt the MLLM to generate more detailed captions centered around the main subject CLS. The prompt we used in this step is stated as follow:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 147, + 653, + 395, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 653, + 395, + 666 + ], + "spans": [ + { + "bbox": [ + 147, + 653, + 395, + 666 + ], + "type": "inline_equation", + "content": "P_{1} =" + }, + { + "bbox": [ + 147, + 653, + 395, + 666 + ], + "type": "text", + "content": " \"Describe the image of the subject CLS in detail.\"" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Y. Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 137, + 468, + 235 + ], + "blocks": [ + { + "bbox": [ + 148, + 137, + 468, + 235 + ], + "lines": [ + { + "bbox": [ + 148, + 137, + 468, + 235 + ], + "spans": [ + { + "bbox": [ + 148, + 137, + 468, + 235 + ], + "type": "table", + "html": "
Step 1 Caption ExtractionStep 2-1 Dimension Generation
Input: Dataset: Dimg, MLLMInput: Captions: C, LLM
Output: Image Captions: COutput: dimensions: Dims
1: for i in range(NumOfIterations)1: for i in range(NumOfIterations)
2: img = Dimg.sample()2: c = C/sample(NumOfSamples)
3: c = MLLM(img, P1)3: S.append(LLM([P2^1, c]))
4: C.append(c)4: end for
5: end for5: Dims = LLMSummary(S)
", + "image_path": "42f35fcf09f848c5e2546c225603f45735886993dac04ea59ab019fbada94f0a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 253, + 268, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 253, + 268, + 264 + ], + "spans": [ + { + "bbox": [ + 132, + 253, + 268, + 264 + ], + "type": "text", + "content": "3.3 Criteria Initialization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 275, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 407 + ], + "type": "text", + "content": "To discover the hidden subpopulation structures within the dataset, we employ an LLM to delve into the information-rich captions generated in the previous step. Our objective is to identify certain criteria that effectively partition the images into several distinct subgroups. Beyond simply dividing the dataset into subgroups, we articulate the classification dimension for the partition and record all the resulting attributes generated from the classification process. Along with the class information and the resulting subpopulations, this criteria naturally form a four-layer structure, class-, dimension-, attribute-, and subpopulation-. Noticing criteria encompass multiple dimensions and their corresponding attributes, we adopt a generate-and-select paradigm with the LLM to discover the dimensions and the attributes sequentially." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 407, + 482, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 526 + ], + "type": "text", + "content": "To determine the dimensions and attributes, we employ an iterative sampling approach, repeatedly prompting the LLM to propose dimensions and attributes based on batches of image captions. In each iteration, the LLM generates candidate dimensions and attributes, which are subsequently processed through an LLM summarization process. This sample-and-summarize approach effectively addresses the challenges when processing large datasets. Since the number of dimensions that can differentiate images in a dataset is relatively small, and these dimensions have an appearance in numerous images, our sample-based approach effectively identifies relevant dimensions. The prompts we used in this step are stated as follow, omitting Chain-of-thought examples for simplicity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 527, + 480, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 527, + 480, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 527, + 480, + 550 + ], + "type": "inline_equation", + "content": "P_2^1 = \\text{\"Suggest some dimensions that can differentiate the following image captions.}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 551, + 480, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 551, + 480, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 551, + 480, + 574 + ], + "type": "inline_equation", + "content": "P_2^2 = \\text{\"Suggest a complete criterion to differentiate the following image captions by the given dimension.}\"" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 595, + 261, + 606 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 595, + 261, + 606 + ], + "spans": [ + { + "bbox": [ + 132, + 595, + 261, + 606 + ], + "type": "text", + "content": "3.4 Criteria Refinement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "To further refine the criteria and ensure its effectiveness in classifying image captions across the dataset, we implement a recursive refining process. This approach proposes a novel method for identifying image captions requiring further refinement utilizing the self-consistency of LLM responses as an indicator [45]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "type": "text", + "content": "Subpopulation Structure Discovery with Large Language Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 272, + 128 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 272, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 272, + 128 + ], + "type": "text", + "content": "Step 2-2 Attribute Generation" + } + ] + } + ], + "index": 2, + "type": "text" + }, + { + "bbox": [ + 132, + 130, + 418, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 130, + 418, + 141 + ], + "spans": [ + { + "bbox": [ + 132, + 130, + 418, + 141 + ], + "type": "text", + "content": "Input: Dimensions: Dims, Captions: " + }, + { + "bbox": [ + 132, + 130, + 418, + 141 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 132, + 130, + 418, + 141 + ], + "type": "text", + "content": ", Large language model: LLM" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 142, + 289, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 142, + 289, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 142, + 289, + 152 + ], + "type": "text", + "content": "Output: Initialized criteria: Criteria" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 153, + 356, + 251 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 137, + 153, + 236, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 153, + 236, + 163 + ], + "spans": [ + { + "bbox": [ + 137, + 153, + 236, + 163 + ], + "type": "text", + "content": "1: for dim in Dims do" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 163, + 309, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 163, + 309, + 175 + ], + "spans": [ + { + "bbox": [ + 137, + 163, + 309, + 175 + ], + "type": "text", + "content": "2: for i in range(NumOfIterations) do" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 175, + 293, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 175, + 293, + 186 + ], + "spans": [ + { + "bbox": [ + 137, + 175, + 293, + 186 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 137, + 175, + 293, + 186 + ], + "type": "inline_equation", + "content": "c = C" + }, + { + "bbox": [ + 137, + 175, + 293, + 186 + ], + "type": "text", + "content": ".sample(NumOfSamples)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 186, + 289, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 186, + 289, + 198 + ], + "spans": [ + { + "bbox": [ + 137, + 186, + 289, + 198 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 137, + 186, + 289, + 198 + ], + "type": "inline_equation", + "content": "S.\\mathrm{append}(\\mathrm{LLM}([P_2^2, dim, c]))" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 198, + 193, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 198, + 193, + 206 + ], + "spans": [ + { + "bbox": [ + 137, + 198, + 193, + 206 + ], + "type": "text", + "content": "5: end for" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 208, + 356, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 208, + 356, + 219 + ], + "spans": [ + { + "bbox": [ + 137, + 208, + 356, + 219 + ], + "type": "text", + "content": "6: Attributes " + }, + { + "bbox": [ + 137, + 208, + 356, + 219 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 137, + 208, + 356, + 219 + ], + "type": "text", + "content": " LLMSumarry(S) *list of attributes" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 219, + 270, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 219, + 270, + 230 + ], + "spans": [ + { + "bbox": [ + 137, + 219, + 270, + 230 + ], + "type": "text", + "content": "7: Criteria[dim] = Attributes" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 230, + 194, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 230, + 194, + 241 + ], + "spans": [ + { + "bbox": [ + 137, + 230, + 194, + 241 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 137, + 230, + 194, + 241 + ], + "type": "inline_equation", + "content": "S.\\mathrm{reset}()" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 137, + 242, + 183, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 242, + 183, + 251 + ], + "spans": [ + { + "bbox": [ + 137, + 242, + 183, + 251 + ], + "type": "text", + "content": "9: end for" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 279, + 481, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 481, + 350 + ], + "type": "text", + "content": "This choice stems from our empirical observation that if an image can be accurately classified according to a particular dimension, it should consistently be classified into the same attribute multiple times. Inconsistent responses, however, suggest that the current criteria require further refinement, either to merge redundant attributes or to include new attributes. The prompts we used in this step are stated as follow and the pseudocode is included in the appendix:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 147, + 351, + 392, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 351, + 392, + 363 + ], + "spans": [ + { + "bbox": [ + 147, + 351, + 392, + 363 + ], + "type": "inline_equation", + "content": "P_3^1 = \\text{\"Classify the caption by the criteria listed below.\"}" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 364, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 411 + ], + "type": "inline_equation", + "content": "P_{3}^{2} = \\text{\"We are unable to classify the following image caption using the provided criteria due to attributes redundancy or disappearance. If redundancy, please prune the redundant attributes. If missing, please suggest an additional attribute that would match the image caption.\"}" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 131, + 430, + 296, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 430, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 131, + 430, + 296, + 442 + ], + "type": "text", + "content": "3.5 Subpopulation Assignment" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 450, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 481, + 533 + ], + "type": "text", + "content": "Equipped with the comprehensive criteria, we proceed to systematically assign each image to the specific attributes of each dimension. Images assigned to the same attributes across all dimensions form distinct subgroups within the dataset, revealing the intrinsic subpopulation structures hidden within the data. These subpopulation structures can then be leveraged to perform various downstream tasks, completing our overall pipeline for employing an LLM to analyze the image dataset. The prompt we used in this step is stated as follow:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 146, + 534, + 478, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 534, + 478, + 546 + ], + "spans": [ + { + "bbox": [ + 146, + 534, + 478, + 546 + ], + "type": "inline_equation", + "content": "P_{4} =" + }, + { + "bbox": [ + 146, + 534, + 478, + 546 + ], + "type": "text", + "content": " \"Please assign following caption to one attribute of given dimension.\"" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 568, + 287, + 581 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 287, + 581 + ], + "type": "text", + "content": "Step 4 Subpopulation Assignment" + } + ] + } + ], + "index": 21, + "type": "text" + }, + { + "bbox": [ + 132, + 583, + 412, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 583, + 412, + 595 + ], + "spans": [ + { + "bbox": [ + 132, + 583, + 412, + 595 + ], + "type": "text", + "content": "Input: Captions: " + }, + { + "bbox": [ + 132, + 583, + 412, + 595 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 132, + 583, + 412, + 595 + ], + "type": "text", + "content": ", Large language model: LLM, criteria: Criteria" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 595, + 335, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 595, + 335, + 606 + ], + "spans": [ + { + "bbox": [ + 132, + 595, + 335, + 606 + ], + "type": "text", + "content": "Output: Further assignments for each caption " + }, + { + "bbox": [ + 132, + 595, + 335, + 606 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 132, + 595, + 335, + 606 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 137, + 607, + 318, + 660 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 137, + 607, + 207, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 607, + 207, + 616 + ], + "spans": [ + { + "bbox": [ + 137, + 607, + 207, + 616 + ], + "type": "text", + "content": "1: for " + }, + { + "bbox": [ + 137, + 607, + 207, + 616 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 137, + 607, + 207, + 616 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 137, + 607, + 207, + 616 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 137, + 607, + 207, + 616 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 137, + 616, + 309, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 616, + 309, + 628 + ], + "spans": [ + { + "bbox": [ + 137, + 616, + 309, + 628 + ], + "type": "text", + "content": "2: for [dim, Attributes] in Criteria do" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 137, + 628, + 318, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 628, + 318, + 639 + ], + "spans": [ + { + "bbox": [ + 137, + 628, + 318, + 639 + ], + "type": "text", + "content": "3: c.Assign(LLM([P4,c, dim, Attributes]))" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 137, + 640, + 193, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 640, + 193, + 649 + ], + "spans": [ + { + "bbox": [ + 137, + 640, + 193, + 649 + ], + "type": "text", + "content": "4: end for" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 137, + 651, + 184, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 651, + 184, + 660 + ], + "spans": [ + { + "bbox": [ + 137, + 651, + 184, + 660 + ], + "type": "text", + "content": "5: end for" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Y. Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 117, + 469, + 264 + ], + "blocks": [ + { + "bbox": [ + 137, + 117, + 469, + 264 + ], + "lines": [ + { + "bbox": [ + 137, + 117, + 469, + 264 + ], + "spans": [ + { + "bbox": [ + 137, + 117, + 469, + 264 + ], + "type": "image", + "image_path": "f3520e9641c319c381fc8d660d62f50d3451f8644963803942cfc7492cb72cd9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 274, + 453, + 285 + ], + "lines": [ + { + "bbox": [ + 159, + 274, + 453, + 285 + ], + "spans": [ + { + "bbox": [ + 159, + 274, + 453, + 285 + ], + "type": "text", + "content": "Fig. 4: A visualization of organised subpopulations in a dataset of cats." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 309, + 379, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 309, + 379, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 309, + 379, + 321 + ], + "type": "text", + "content": "3.6 Task-specific Tuning for Downstream Tasks" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 413 + ], + "type": "text", + "content": "Once we have identified the criteria and subpopulations within the dataset, we can leverage this information to tackle several downstream tasks effectively. This includes organizing the subpopulations, which can reveal valuable insights into the data, such as potential subpopulation biases and the presence of long-tail attributes. In practice, the subpopulation structures can be used to improve model performance on various tasks if combined with extra operations, including handling subpopulation shifts and slice discovery." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 415, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 533 + ], + "type": "text", + "content": "Dataset Subpopulation Organization Organizing subpopulations within a dataset can provide beneficial information about the numbers of data points in each subpopulation and can reflect dataset biases and help us identify some long-tail subgroups. We conduct experiments on the task of image clustering conditioned on human-specified criteria to evaluate the quality of subpopulation organization. In specific, when organising the subpopulations of a given image dataset, we first select out the relevant dimensions and then attach attributes assigned by SSD-LLM directly to the images accordingly. Our method, SSD-LLM, automates subpopulation organization within datasets, and thus holds the potential to revolutionize the way hierarchical datasets are constructed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "Subpopulation Shift Subpopulation shift [48] stands as a common challenge in machine learning, occurring when the proportion of some subpopulations between training and deployment changes, and is shown to be of significant influence to model performances [39, 48]. SSD-LLM, combined with image generation, offers a solution to better handle the scenarios of subpopulation shifts. In our experiments, after we apply SSD-LLM to the datasets, we collect statistics the number of images contained in each subpopulation and utilize diffusion model to generate images for underrepresented subpopulations. Subsequently, we sample attributes from the subpopulation structure for each underrepresented subpopulation and employ LLM to make complete sentences based on these words as the input prompt of a diffusion model. The diffusion model generates images augmented" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "type": "text", + "content": "Subpopulation Structure Discovery with Large Language Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 192, + 114, + 422, + 215 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 422, + 215 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 422, + 215 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 422, + 215 + ], + "type": "table", + "html": "
DatasetCriterionSCAN*IC|TCOurs
Stanford 40 ActionAction0.3460.7470.817
Location0.3570.6710.705
Mood0.2760.7460.768
Place365Place0.332-0.696
PPMIMusical Instruction0.5980.9340.955
Cifar10Object0.8390.9110.921
STL10Object0.7980.9860.988
", + "image_path": "6224993018e615fe55668b739eeea2e4c53df0945387539b58b6622607faa8f3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 216, + 480, + 248 + ], + "lines": [ + { + "bbox": [ + 130, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 480, + 248 + ], + "type": "text", + "content": "Table 1: Quantitative results of Dataset Subpopulation Organization. Method labeled with * is evaluated by having a human provided ground truth labels, cause the method itself is an unsupervised learning paradigm." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 282, + 479, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 479, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 479, + 329 + ], + "type": "text", + "content": "to the image dataset which further helps to achieve balanced classes and attributes. Moreover, we propose to harness an LLM to suggest extra dimensions and attributes based on the current sets in this task for enriched subpopulation structure, generating more diverse images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 331, + 480, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 480, + 450 + ], + "type": "text", + "content": "Slice Discovery Slice discovery is a task aiming at uncovering subpopulations within a dataset where a machine learning model consistently exhibits poor performances. These subpopulations with underperformances, or slices, provide valuable insights into the model's limitations, potential biases [14], and how to improve the performances. SSD-LLM conducts slice discovery for an image dataset with the help of the assigned attributes. In detail, we first calculate the error rates on all subpopulations discovered with the SSD-LLM. Then we identify out the subpopulations with the highest error rate and use the LLM to summarize out discriptions based on the attributes of the subpopulations in the form of texts representing the slice topics, completing the task of slice discovery." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 468, + 230, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 468, + 230, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 230, + 483 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 495, + 479, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 479, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 479, + 542 + ], + "type": "text", + "content": "We now present experimental results demonstrating the effectiveness of SSD-LLM. In particular, we present the main settings and results in this section and defer extra details, including various visualization result, to appendix. In our experiments, we mainly use LLaVA1.5 [26] for MLLM and GPT-4 [1] for LLM." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 543, + 479, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 479, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 479, + 602 + ], + "type": "text", + "content": "We conduct experiments on Dataset Subpopulation Organization, Subpopulation Shift, and Slice Discovery. Our superior performance underscores the method's efficacy in identifying and analyzing subgroups, further affirming its utility in addressing related challenges. Moreover, it illustrates that the unified paradigm can effectively address a variety of downstream tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 620, + 418, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 418, + 633 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 418, + 633 + ], + "type": "text", + "content": "4.1 Evaluation on Dataset Subpopulation Organization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Setup The data assignment process facilitated by SSD-LLM, can be considered a form of clustering. To evaluate the quality of these identified subpopulations," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Y. Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 114, + 476, + 253 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 476, + 253 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 476, + 253 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 476, + 253 + ], + "type": "table", + "html": "
TypeMethodAverage AccuracyWorst Group Accuracy
WaterbirdsMetashiftNico++AverageWaterbirdsMetashiftNico++Average
VanillaERM84.191.276.383.769.182.117.856.3
Subgroup Robust MethodsGroupDRO86.991.574.084.173.183.112.256.1
JTT88.991.277.585.971.282.615.656.5
LfF86.680.477.581.575.072.315.654.3
LISA89.291.475.085.277.079.018.958.3
Imbalanced LearningResample86.292.277.385.270.081.016.755.9
Reweight86.291.573.883.871.983.112.255.7
Focal89.391.673.184.771.681.016.756.4
CBLoss86.891.476.384.874.483.112.256.6
BSoftmax88.491.374.284.674.182.616.757.8
Traditional Data AugmentationMixup89.291.473.084.577.579.014.457.0
RandAug86.390.972.083.171.480.916.756.3
DiffusionClass Prompt85.991.578.085.171.382.718.557.5
Class-Attribute89.191.478.686.473.583.818.858.7
CiP88.091.178.385.873.582.419.358.4
LLM+DiffusionSSD-LLM (Ours)90.593.080.488.079.184.822.162.0
", + "image_path": "551c3d42a03159f814d7fd0af7def7c604ca79065a3794096dfdf962a01662a0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 254, + 476, + 263 + ], + "lines": [ + { + "bbox": [ + 136, + 254, + 476, + 263 + ], + "spans": [ + { + "bbox": [ + 136, + 254, + 476, + 263 + ], + "type": "text", + "content": "Table 2: Comparison of methods for image classification with subpopulation shifts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 299, + 479, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 479, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 479, + 334 + ], + "type": "text", + "content": "we assess the clustering accuracy by comparing images against secondary labels that reflect subgroup attributes derived via SSD-LLM. Full information for datasets, text criterion, and model selection can be found in appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 335, + 480, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 480, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 480, + 382 + ], + "type": "text", + "content": "Comparison Methods SCAN [41] is a two-stage clustering method that decouples feature learning and clustering. IC|TC [19] is a new paradigm for image clustering that supports human interaction. It utilizes the given Text Criteria to accurately control the quality of the clustering results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 384, + 480, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 480, + 479 + ], + "type": "text", + "content": "Results and Analysis In Table. 1, we report the average accuracy achieved by each method based on the predefined textual criteria. When compared to IC|TC, SSD-LLM demonstrates competitive performance. It is important to note that IC|TC incorporate artificial judgment in its process, which leads to poor scalability when handling large datasets. In contrast, our approach is fully automated. We conduct visualization of our organized subpopulations, as shown in Figure. 4. We can visually observe that the pictures and dimensions are indeed consistent, indicating the effectiveness of our mining and assigning process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 498, + 334, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 334, + 511 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 334, + 511 + ], + "type": "text", + "content": "4.2 Evaluation on Subpopulation Shift" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 628 + ], + "type": "text", + "content": "Setup We evaluate subpopulation shifts on three commonly used image datasets, Metashift (Cats vs Dogs), Waterbirds (Landbirds vs Waterbirds), and NICO++. We choose Average Accuracy and Worst Group Accuracy as evaluation metrics. To ensure a fair comparison, following [48], we conduct a random search of 16 trials over a joint distribution of all hyperparameters. We then use the validation set to select the best hyperparameters for each algorithm, fix them, and rerun the experiments under five different random seeds to report the final average results. To make the evaluation more realistic, we consider the model selection setting Attributes are unknown in both training and validation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "content": "Comparison Methods Following recent benchmarking efforts [48], we compare SSD-LLM with several types of methods: (1) vanilla: ERM [42], (2) Subgroup Robust Methods: GroupDRO [38], LfF [32], JTT [24], LISA [49], (3) Imbalanced" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "type": "text", + "content": "Subpopulation Structure Discovery with Large Language Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 114, + 467, + 205 + ], + "blocks": [ + { + "bbox": [ + 148, + 114, + 467, + 205 + ], + "lines": [ + { + "bbox": [ + 148, + 114, + 467, + 205 + ], + "spans": [ + { + "bbox": [ + 148, + 114, + 467, + 205 + ], + "type": "table", + "html": "
Method|CategoriesBoatBirdCarCatDogTruckTopic Error Rate
ImageNet4.330.8111.3311.140.6911.716.72
General Prompt47.8212.1143.5514.2210.1912.6523.42
GPT-Suggest57.5512.8743.5912.7116.3428.1228.53
Domino(Bert)76.2642.2654.2133.8924.5029.5443.44
B2T77.6230.0458.1736.3619.8033.4742.58
SSD-LLM (Ours)79.3145.6760.3432.9726.4839.5747.39
", + "image_path": "a6244cd2546454c418c04d2c2a6cf3dcbb68363509120b03f209e0b960ee9913.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 205, + 452, + 214 + ], + "lines": [ + { + "bbox": [ + 160, + 205, + 452, + 214 + ], + "spans": [ + { + "bbox": [ + 160, + 205, + 452, + 214 + ], + "type": "text", + "content": "Table 3: Results of slice discovery on Imagenet-1K with various SDMs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 247, + 481, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 481, + 283 + ], + "type": "text", + "content": "Learning: ReSample [13], ReWeight [13], Focal [23], CBLoss [5], Bsoftmax [36], (4) Traditional Data Augmentation: Mixup [50], RandAug [3], (5) Diffusion: Class Prompt [40], Class-Attribute Prompt [40], CiP [20]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 283, + 481, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 481, + 462 + ], + "type": "text", + "content": "Results and Analysis From Table 2, SSD-LLM surpasses previous methods, with a " + }, + { + "bbox": [ + 130, + 283, + 481, + 462 + ], + "type": "inline_equation", + "content": "+1.6\\%" + }, + { + "bbox": [ + 130, + 283, + 481, + 462 + ], + "type": "text", + "content": " improvement in average accuracy and " + }, + { + "bbox": [ + 130, + 283, + 481, + 462 + ], + "type": "inline_equation", + "content": "+3.3\\%" + }, + { + "bbox": [ + 130, + 283, + 481, + 462 + ], + "type": "text", + "content": " in worst group accuracy across three datasets. The analysis is as follows: (1) Despite being based on conventional ERM, data-based approaches show competitive performance compared to model-based algorithms, highlighting their potential. (2) For diffusion-based methods, class-attribute prompts outperform class prompts, underscoring the importance of understanding dataset imbalanced attributes for effective image generation. However, the need for pre-identifying these attributes emphasizes the value of SSD-LLM, which automates attribute discovery and provides detailed annotations, enhancing performance. (3) The superior performance of CiP over Class Prompt underscores the significance of diverse text prompts. (4) Our method's superior performance results from a comprehensive analysis of subpopulation imbalances within the dataset. The strategic text diversity achieved through attribute sampling and LLM sentence-making effectively addresses subpopulation shifts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 479, + 309, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 309, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 309, + 491 + ], + "type": "text", + "content": "4.3 Evaluation on Slice Discovery" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 545 + ], + "type": "text", + "content": "Setup In contrast to typical slice discovery tasks, we redefine evaluation pipeline following [11]. In this study, we evaluate bugs found of ImageNet models. Specifically, a classification is deemed incorrect when an image containing target object is erroneously identified by the model as containing an unrelated object." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "text", + "content": "Comparison Methods Domino [9] represents a state-of-the-art method in slice discovery, which effectively clusters errors identified in the validation set and characterizes them through captions generated automatically. B2T [16] is a recently proposed framework which identifies and interprets visual biases in vision models using keyword extraction from captions of mispredicted images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Results and Analysis We evaluate the effectiveness of our method in slice discovery on 6 representative superclasses in Imagenet. As shown in Table. 3, our SSD-LLM overcomes all other SDMs, including Domino [9], by a significant margin. We find topics given by Domino tend to encounter two unsatisfactory cases: loss of semantics, and missing class information. These cases are also discussed" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Y. Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 144, + 114, + 470, + 203 + ], + "blocks": [ + { + "bbox": [ + 144, + 114, + 470, + 203 + ], + "lines": [ + { + "bbox": [ + 144, + 114, + 470, + 203 + ], + "spans": [ + { + "bbox": [ + 144, + 114, + 470, + 203 + ], + "type": "table", + "html": "
MethodsIndexCriteria InitializationCriteria RefinementAccuracy
DimensionAttributeActionLocationMood
SSD LLMA10 × 2020 × 20×75.667.371.6
B5 × 2020 × 20×65.260.066.2
C20 × 2020 × 20×75.967.171.3
D10 × 2010 × 40×70.368.969.5
E10 × 2040 × 10×80.270.074.4
F10 × 2020 × 20✓81.770.576.8
", + "image_path": "38e3b2cfb85ac22a1de5e9de4972769cc02ab21cec3af3f9ccf6c9386a378871.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 204, + 480, + 225 + ], + "lines": [ + { + "bbox": [ + 132, + 204, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 480, + 225 + ], + "type": "text", + "content": "Table 4: Ablation study on the number of samples (NUM of rounds × NUM per round), and Criteria Refinement (with or without)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 258, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 389 + ], + "type": "text", + "content": "in [11], where they reasoned this phenomenon into the inherent difficulties of automatic SDMs. However, evidence suggests that our SSD-LLM can handle these problems, while keeping a high error rate and maintaining automation. Specifically, SSD-LLM achieves an average error rate of " + }, + { + "bbox": [ + 130, + 258, + 482, + 389 + ], + "type": "inline_equation", + "content": "47.39\\%" + }, + { + "bbox": [ + 130, + 258, + 482, + 389 + ], + "type": "text", + "content": ", surpassing Domino [9] by " + }, + { + "bbox": [ + 130, + 258, + 482, + 389 + ], + "type": "inline_equation", + "content": "3.95\\%" + }, + { + "bbox": [ + 130, + 258, + 482, + 389 + ], + "type": "text", + "content": ". Furthermore, when we trace back to origin dataset, the discovered slice is also very consistent (detailed visualizations included in appendix). Interestingly, we find the data mining process of SSD-LLM is just the same as human data scientists, who take up hypotheses and improve model performance by viewing batches of bad subpopulations. Experiments show that our method overcome the inherent difficulties while maintaining automaton, paving the way for data-centric methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 407, + 238, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 238, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 238, + 419 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 557 + ], + "type": "text", + "content": "Hyperparameters of Criteria Initialization/Refinement The " + }, + { + "bbox": [ + 130, + 426, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\mathrm{N} \\times \\mathrm{M}" + }, + { + "bbox": [ + 130, + 426, + 482, + 557 + ], + "type": "text", + "content": " in Table. 4 represents the rounds and samples of suggestions. Setting A serves as our baseline. For A, B, and C, results show that insufficient suggested dimensions lead to decreased performance, while enough dimension samples lead to stable performance, as the ICTC task requires only the most appropriate match. Too few suggestions may fail to find suitable dimensions, resulting in mismatched attribute generation. For A, D, and E, results highlight the importance of the number of captions in summarizing attributes. Too many captions can cause some to be overlooked, reducing total identified attributes and performance. The improvement from A to F demonstrates that Criteria Refinement enhances attribute comprehensiveness and final performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "Diffusion Generation Strategy From Table. 5, we analyze key components for employing diffusion models to address subpopulation shift. For " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{A}\\rightarrow \\mathrm{B}" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": ", we confirm that managing imbalance attributes within datasets helps solve the task. For " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{B}\\rightarrow \\mathrm{D}" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": ", we illustrate that a comprehensive subpopulation structure benefits the task. For " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{D}\\rightarrow \\mathrm{E}" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": ", balanced subpopulation sampling improves data quality and model training. For " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{E}\\rightarrow \\mathrm{F}" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": ", LLM-generated prompts produce more reasonable images, enhancing results. For " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{F}\\rightarrow \\mathrm{G}" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": ", LLM suggests additional attributes to enrich subpopulation structures, generating more diverse images and improving model generalization. For " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{F}\\rightarrow \\mathrm{H}" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": ", we verify the scaling capability of our SSD-LLM." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "type": "text", + "content": "Subpopulation Structure Discovery with Large Language Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 114, + 480, + 207 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 480, + 207 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 480, + 207 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 480, + 207 + ], + "type": "table", + "html": "
MethodsIndexAttribute ModeSample ModeSD Prompt ModeNumberAverage AccuracyWorst Group Accuracy
ERM————————×184.169.1
SSD-LLMAGT Attribute UnknownrandomDirect×185.971.3
BGT Attribute KnownrandomDirect×189.173.5
CGT Attribute KnownweightedDirect×189.373.8
DSSD-LLM AttributerandomDirect×189.576.2
ESSD-LLM AttributeweightedDirect×189.877.4
FSSD-LLM AttributeweightedLLM Sentence×190.178.3
GSSD-LLM & LLM SuggestweightedLLM Sentence×190.579.1
HSSD-LLM AttributeweightedLLM Sentence×291.179.2
", + "image_path": "e9d152beedeead21b134e13c9db385deb11ab1f49f7c4ba1cbb23019501e2fb2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 165, + 208, + 447, + 217 + ], + "lines": [ + { + "bbox": [ + 165, + 208, + 447, + 217 + ], + "spans": [ + { + "bbox": [ + 165, + 208, + 447, + 217 + ], + "type": "text", + "content": "Table 5: Ablation study on the sample mode and SD prompt mode." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 247, + 217, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 217, + 260 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 217, + 260 + ], + "type": "text", + "content": "5 Discussion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 270, + 479, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 479, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 479, + 330 + ], + "type": "text", + "content": "Although our method, SSD-LLM, has already shown effectiveness in various settings, the current exploration of the algorithm is limited to image datasets. Besides, this approach may bear the potential bias from the MLLMs and LLMs. However, it may be reduced from extra human-in-the-loop guidances if available. For future works, we suggest the following promising directions:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 336, + 479, + 453 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 138, + 336, + 479, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 336, + 479, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 336, + 479, + 358 + ], + "type": "text", + "content": "- Structure Format The four-layer subpopulation structure can be expanded to more suitable structures according to specific task requirements." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 360, + 479, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 360, + 479, + 382 + ], + "spans": [ + { + "bbox": [ + 137, + 360, + 479, + 382 + ], + "type": "text", + "content": "- Downstream Tasks SSD-LLM can have more applications in various computer vision and multimodality tasks, e.g. object detection and VQA." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 383, + 479, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 383, + 479, + 417 + ], + "spans": [ + { + "bbox": [ + 137, + 383, + 479, + 417 + ], + "type": "text", + "content": "- Dataset Construction The subpopulation structure obtained from SSD-LLM holds the potential to guide dataset construction with better fairness [44] or further supporting the construction of unbiased datasets [29]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 418, + 479, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 418, + 479, + 453 + ], + "spans": [ + { + "bbox": [ + 137, + 418, + 479, + 453 + ], + "type": "text", + "content": "- Technical Extensions The core procedures of SSD-LLM, using LLM to conduct group-level summarizations, can be extended to more types of contents including patterns of model hallucinations." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 471, + 220, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 220, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 220, + 483 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 480, + 601 + ], + "type": "text", + "content": "In this work, we present the first systematic exploration of subpopulation structure discovery. We provide a precise definition of subpopulation structure and introduce a fine-grained criteria to determine the structures. We propose SSD-LLM for automatic subpopulation structure discovery incorporating elaborate prompt engineering techniques. SSD-LLM can be combined with subsequent operations to better tackle several downstream tasks, including dataset subpopulation organization, subpopulation shift and slice discovery, with minor Task-specific Tuning. We conduct extensive experiments to verify our proposed methods, demonstrating the remarkable effectiveness and and generality of SSD-LLM." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 618, + 247, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 618, + 247, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 618, + 247, + 632 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "Shanghai Zhang is supported by the National Science and Technology Major Project of China (No. 2022ZD0117801)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Y. Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "text", + "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 170, + 480, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 170, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 138, + 170, + 480, + 190 + ], + "type": "text", + "content": "2. Chen, M., Li, Y., Xu, Q.: Hibug: On human-interpretable model debug. In: NeurIPS (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 191, + 480, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 191, + 480, + 233 + ], + "spans": [ + { + "bbox": [ + 138, + 191, + 480, + 233 + ], + "type": "text", + "content": "3. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 702-703 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 234, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 480, + 266 + ], + "type": "text", + "content": "4. Cui, J., Li, Z., Yan, Y., Chen, B., Yuan, L.: Chatlaw: Open-source legal large language model with integrated external knowledge bases. arXiv preprint arXiv:2306.16092 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 267, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 480, + 299 + ], + "type": "text", + "content": "5. Cui, Y., Jia, M., Lin, T.Y., Song, Y., Belongie, S.: Class-balanced loss based on effective number of samples. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9268-9277 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 300, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 480, + 342 + ], + "type": "text", + "content": "6. Dai, D., Sun, Y., Dong, L., Hao, Y., Ma, S., Sui, Z., Wei, F.: Why can gpt learn in-context? language models implicitly perform gradient descent as meta-optimizers. In: ICLR 2023 Workshop on Mathematical and Empirical Understanding of Foundation Models (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 342, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 363 + ], + "type": "text", + "content": "7. Dong, Q., Li, L., Dai, D., Zheng, C., Wu, Z., Chang, B., Sun, X., Xu, J., Sui, Z.: A survey for in-context learning. arXiv preprint arXiv:2301.00234 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 364, + 480, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 480, + 395 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 480, + 395 + ], + "type": "text", + "content": "8. Dunlap, L., Umino, A., Zhang, H., Yang, J., Gonzalez, J.E., Darrell, T.: Diversify your vision datasets with automatic diffusion-based augmentation. arXiv preprint arXiv:2305.16289 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 396, + 480, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 428 + ], + "type": "text", + "content": "9. Eyuboglu, S., Varma, M., Saab, K., Delbrouck, J.B., Lee-Messer, C., Dunnmon, J., Zou, J., Ré, C.: Domino: Discovering systematic errors with cross-modal embeddings. arXiv preprint arXiv:2203.14960 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 480, + 460 + ], + "type": "text", + "content": "0. Fu, Y., Peng, H., Khot, T., Lapata, M.: Improving language model negotiation with self-play and in-context learning from ai feedback. arXiv preprint arXiv:2305.10142 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 461, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 461, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 138, + 461, + 480, + 492 + ], + "type": "text", + "content": "1. Gao, I., Ilharco, G., Lundberg, S., Ribeiro, M.T.: Adaptive testing of computer vision models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4003-4014 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 493, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 480, + 525 + ], + "type": "text", + "content": "2. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 525, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 546 + ], + "type": "text", + "content": "3. Japkowicz, N.: The class imbalance problem: Significance and strategies. In: Proc. of the Int'l Conf. on artificial intelligence. vol. 56, pp. 111-117 (2000)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 547, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 480, + 600 + ], + "type": "text", + "content": "4. Johnson, N., Cabrera, A., Plumb, G., Talwalkar, A.: Where does my model underperform? a human evaluation of slice discovery algorithms. Proceedings of the AAAI Conference on Human Computation and Crowdsourcing 11(1), 65-76 (Nov 2023). https://doi.org/10.1609/hcomp.v11i1.27548, http://dx.doi.org/10.1609/hcomp.v11i1.27548" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "text", + "content": "5. Kim, Y., Mo, S., Kim, M., Lee, K., Lee, J., Shin, J.: Bias-to-text: Debi-aising unknown visual biases through language interpretation. arXiv preprint arXiv:2301.11104 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "6. Kim, Y., Mo, S., Kim, M., Lee, K., Lee, J., Shin, J.: Bias-to-text: Debiased unknown visual biases through language interpretation. arXiv preprint arXiv:2301.11104 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "type": "text", + "content": "Subpopulation Structure Discovery with Large Language Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "17. Kojima, T., Gu, S.S., Reid, M., Matsuo, Y., Iwasawa, Y.: Large language models are zero-shot reasoners. NeurIPS 35, 22199-22213 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "18. Kwon, S., Park, J., Kim, M., Cho, J., Ryu, E.K., Lee, K.: Image clustering conditioned on text criteria. In: ICLR (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "19. Kwon, S., Park, J., Kim, M., Cho, J., Ryu, E.K., Lee, K.: Image clustering conditioned on text criteria. arXiv preprint arXiv:2310.18297 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "20. Lei, S., Chen, H., Zhang, S., Zhao, B., Tao, D.: Image captions are natural prompts for text-to-image models. arXiv preprint arXiv:2307.08526 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "21. Li, X., Qiu, X.: Mot: Pre-thinking and recalling enable chatgpt to self-improve with memory-of-thoughts. arXiv preprint arXiv:2305.05181 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 228, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 480, + 248 + ], + "type": "text", + "content": "22. Liang, W., Zou, J.: Metashift: A dataset of datasets for evaluating contextual distribution shifts and training conflicts. arXiv preprint arXiv:2202.06523 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "23. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 282, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 324 + ], + "type": "text", + "content": "24. Liu, E.Z., Haghgoo, B., Chen, A.S., Raghunathan, A., Koh, P.W., Sagawa, S., Liang, P., Finn, C.: Just train twice: Improving group robustness without training group information. In: International Conference on Machine Learning. pp. 6781-6792. PMLR (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "type": "text", + "content": "25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "26. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "27. Liu, J., Hu, T., Zhang, Y., Gai, X., Feng, Y., Liu, Z.: A chatgpt aided explainable framework for zero-shot medical image diagnosis. arXiv preprint arXiv:2307.01981 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "28. Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., Neubig, G.: Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys 55(9), 1-35 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 436, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 480, + 456 + ], + "type": "text", + "content": "29. Liu, Z., He, K.: A decade's battle on dataset bias: Are we there yet? (2024), https://arxiv.org/abs/2403.08632" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "text", + "content": "30. Luo, Y., Tang, Y., Shen, C., Zhou, Z., Dong, B.: Prompt engineering through the lens of optimal control. arXiv preprint arXiv:2310.14201 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "text", + "content": "31. Madaan, A., Tandon, N., Gupta, P., Hallinan, S., Gao, L., Wiegreffe, S., Alon, U., Dziri, N., Prabhumoye, S., Yang, Y., et al.: Self-refine: Iterative refinement with self-feedback. arXiv preprint arXiv:2303.17651 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "32. Nam, J., Cha, H., Ahn, S., Lee, J., Shin, J.: Learning from failure: De-biasing classifier from biased classifier. Advances in Neural Information Processing Systems 33, 20673-20684 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "type": "text", + "content": "33. Park, J.S., O'Brien, J.C., Cai, C.J., Morris, M.R., Liang, P., Bernstein, M.S.: Generative agents: Interactive simulacra of human behavior. arXiv preprint arXiv:2304.03442 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 577, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 480, + 599 + ], + "type": "text", + "content": "34. Qian, C., Cong, X., Yang, C., Chen, W., Su, Y., Xu, J., Liu, Z., Sun, M.: Communicative agents for software development. arXiv preprint arXiv:2307.07924 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "35. Qiao, S., Ou, Y., Zhang, N., Chen, X., Yao, Y., Deng, S., Tan, C., Huang, F., Chen, H.: Reasoning with language model prompting: A survey. arXiv preprint arXiv:2212.09597 (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "36. Ren, J., Yu, C., Ma, X., Zhao, H., Yi, S., et al.: Balanced meta-softmax for long-tailed visual recognition. Advances in neural information processing systems 33, 4175-4186 (2020)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Y. Luo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 621 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "37. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115, 211-252 (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "text", + "content": "38. Sagawa, S., Koh, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "type": "text", + "content": "39. Santurkar, S., Tsipras, D., Madry, A.: Breeds: Benchmarks for subpopulation shift (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 247 + ], + "type": "text", + "content": "40. Shipard, J., Wiliem, A., Thanh, K.N., Xiang, W., Fookes, C.: Diversity is definitely needed: Improving model-agnostic zero-shot classification via stable diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 769-778 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 248, + 482, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 282 + ], + "type": "text", + "content": "41. Van Gansbeke, W., Vandenhende, S., Georgoulis, S., Proesmans, M., Van Gool, L.: Scan: Learning to classify images without labels. In: European conference on computer vision. pp. 268-285. Springer (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 282, + 482, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 304 + ], + "type": "text", + "content": "42. Vapnik, V.N.: An overview of statistical learning theory. IEEE transactions on neural networks 10(5), 988-999 (1999)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 304, + 482, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 325 + ], + "type": "text", + "content": "43. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 325, + 482, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 358 + ], + "type": "text", + "content": "44. Wang, A., Liu, A., Zhang, R., Kleiman, A., Kim, L., Zhao, D., Shirai, I., Narayanan, A., Russakovsky, O.: Revise: A tool for measuring and mitigating bias in visual datasets. International Journal of Computer Vision 130(7), 1790-1810 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 358, + 482, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 390 + ], + "type": "text", + "content": "45. Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A., Zhou, D.: Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 390, + 482, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 424 + ], + "type": "text", + "content": "46. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 424, + 482, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 446 + ], + "type": "text", + "content": "47. Xie, S.M., Raghunathan, A., Liang, P., Ma, T.: An explanation of in-context learning as implicit bayesian inference. arXiv preprint arXiv:2111.02080 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 446, + 482, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 468 + ], + "type": "text", + "content": "48. Yang, Y., Zhang, H., Katabi, D., Ghassemi, M.: Change is hard: A closer look at subpopulation shift (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 468, + 482, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 500 + ], + "type": "text", + "content": "49. Yao, H., Wang, Y., Li, S., Zhang, L., Liang, W., Zou, J., Finn, C.: Improving out-of-distribution robustness via selective augmentation. In: International Conference on Machine Learning. pp. 25407-25437. PMLR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 500, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 482, + 522 + ], + "type": "text", + "content": "50. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 522, + 482, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 566 + ], + "type": "text", + "content": "51. Zhang, R., Cai, Z., Yang, H., Liu, Z., Gudovskiy, D., Okuno, T., Nakata, Y., Keutzer, K., Chang, B., Du, Y., et al.: Vecaf: Vlm-empowered collaborative active finetuning with training objective awareness. arXiv preprint arXiv:2401.07853 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 566, + 482, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 566, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 566, + 482, + 588 + ], + "type": "text", + "content": "52. Zhang, X., He, Y., Xu, R., Yu, H., Shen, Z., Cui, P.: Nico++: Towards better benchmarking for domain generalization (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 588, + 482, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 588, + 482, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 482, + 621 + ], + "type": "text", + "content": "53. Zhou, D., Schärli, N., Hou, L., Wei, J., Scales, N., Wang, X., Schuurmans, D., Cui, C., Bousquet, O., Le, Q., et al.: Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 187, + 91, + 447, + 102 + ], + "type": "text", + "content": "Subpopulation Structure Discovery with Large Language Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LLMCO4MR_ LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang/33e2dc3a-2be0-4119-88f6-223087048685_content_list.json b/2024/LLMCO4MR_ LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang/33e2dc3a-2be0-4119-88f6-223087048685_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..56e5e6331ba56d6c1cbe3b83c756394fed151e71 --- /dev/null +++ b/2024/LLMCO4MR_ LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang/33e2dc3a-2be0-4119-88f6-223087048685_content_list.json @@ -0,0 +1,1637 @@ +[ + { + "type": "text", + "text": "LLMCO4MR: LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuqing Zhang $^{1}$ , Hangqi Li $^{1}$ , Shengyu Zhang $^{1,2}$ , Runzhong Wang $^{3}$ , Baoyi He $^{1}$ , Huaiyong Dou $^{1}$ , Junchi Yan $^{3,4}$ , Yongquan Zhang $^{1}$ , and Fei Wu $^{1,2,4}$", + "bbox": [ + 220, + 234, + 781, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Zhejiang University, Hangzhou, China", + "bbox": [ + 367, + 291, + 635, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Shanghai Institute for Advanced Study of Zhejiang University, Shanghai, China", + "bbox": [ + 228, + 306, + 772, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Shanghai Jiao Tong University, Shanghai, China", + "bbox": [ + 333, + 320, + 669, + 334 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Shanghai AI Laboratory, Shanghai, China", + "bbox": [ + 352, + 334, + 648, + 348 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yuqingz7, sy_zhang, wufei}@zju.edu.cn, yanjunchi@sjtu.edu.cn", + "bbox": [ + 256, + 349, + 745, + 362 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Restoring ancient manuscripts fragments, such as those from Dunhuang, is crucial for preserving human historical culture. However, their worldwide dispersal and the shifts in cultural and historical contexts pose significant restoration challenges. Traditional archaeological efforts primarily focus on manually piecing major fragments together, yet the small and more intricate pieces remain largely unexplored, which is technically due to their irregular shapes, sparse textual content, and extensive combinatorial space for reassembly. In this paper, we formalize the task of restoring the ancient manuscript from fragments as a cardinality-constrained combinatorial optimization problem, and propose a framework named LLMCO4MR: (Multimodal) Large Language Model-aided Combinatorial Optimization Neural Networks for Ancient Manuscript Restoration. Specifically, LLMCO4MR encapsulates a neural combinatorial solver equipped with a differentiable optimal transport (OT) layer, to efficiently predict the Top- $K$ likely mutual reassembly candidates. Multimodal Large Language Model (MLLM) is then adopted and prompted to yield pairwise matching confidence and relative directions for final restoration. Experiments on synthetic data and cases studies with real-world famous Dunhuang fragments demonstrate our approach's practical potential in assisting archaeologists. Our method provides a novel perspective for ancient manuscript restoration.", + "bbox": [ + 259, + 406, + 743, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Ancient Manuscript Restoration $\\cdot$ Dunhuang Manuscript $\\cdot$", + "bbox": [ + 259, + 709, + 740, + 723 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Combinatorial Optimization $\\cdot$ Multimodal Large Language Models", + "bbox": [ + 259, + 723, + 705, + 737 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 758, + 375, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The restoration of ancient manuscripts has long been the focus of archaeologists' research and the global community. Traditionally, these codices are frequently", + "bbox": [ + 212, + 787, + 785, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding Authors.", + "bbox": [ + 215, + 825, + 392, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/13fed47de562bd0df09c0cd2641f6ae7246e74da46965579ee1a5e721fbd6b22.jpg", + "image_caption": [ + "(a) Conventional manuscript restoration task" + ], + "image_footnote": [], + "bbox": [ + 228, + 161, + 437, + 232 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8d13fbb16a18f7860c294a586be2a40f57fe501885504e80200f7cc2b089fe56.jpg", + "image_caption": [ + "(b) Our task: Small fragments restoration with outliers", + "Fig. 1: Comparison of conventional manuscript restoration and our task. (a) highlights the conventional restoration task with large fragments. Archaeologists focus on written content as primary guide for restoration, supplemented by texture and edges. (b) outlines our task. In an open-world problem setting, small fragments contain limited text. Recovery relies on contour similarity matching. This is generally difficult for humans to identify fragment pairs from a large collection containing outliers." + ], + "image_footnote": [], + "bbox": [ + 493, + 143, + 754, + 232 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "discovered in a (severely) deteriorated condition due to the passage of time. Recovering these manuscripts involves extensive efforts [37] that are often time-consuming, labor-intensive, and require a high level of expertise, necessitating years of professional training [2].", + "bbox": [ + 212, + 378, + 784, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, vision technologies have demonstrated significant potential in aiding the restoration of ancient manuscripts. For example, [1] developed a sophisticated system based on Convolutional Neural Network (CNN) to identify matching candidates among papyrus fragments by analyzing their fiber patterns. Zhang et al. [51] proposed a self-supervised model using Generative Adversarial Network (GAN) and siamese network for the pairwise stitching of Oracle Bone pieces, utilizing manually annotated contours for similarity comparison. While these vision-assisted pipelines enhance efficiency and accuracy beyond traditional manual efforts, their applications are generally confined to local pairwise matching [31], or regularly-shaped (e.g., rectangle) and large fragments [54]. A major challenge persists in dealing with real-world small fragments, which are often less-textual, shape-irregular, and scattered with missing fragments.", + "bbox": [ + 212, + 441, + 787, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These small fragments have often been overlooked by archaeologists, yet they form a substantial part of the manuscript collections. Recent explorations have revealed that these minor pieces are treasure troves of first-hand materials, offering unique insights into the daily lives of past civilizations [12]. Current deep learning approaches rely on textual [29], contextual [2] or multimodal information [54] to reassembly manuscripts. Such learning-based methods require a large amount of annotated training data for complex multimodal feature extraction, which is impractical for small fragment restoration.", + "bbox": [ + 212, + 626, + 787, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To bridge the gap, we resort to the recent advancements in large language models (LLMs) that have revolutionized traditional tasks in computer vision and natural language processing [7, 26, 38]. Notably, the emergence of Multimodal Large Language Models (MLLMs) has demonstrated exceptional capabilities in extracting and interpreting image features, understanding content, and following complex prompt instructions [4, 20, 48]. Capitalizing on these developments, our", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0a7e97e324120a236faeeaf7f01e2fbf5106846592f1c96c4fbc5dc4f86f6b3f.jpg", + "image_caption": [ + "Fig. 2: Showcases of matching results with case studies on Dunhuang fragments." + ], + "image_footnote": [], + "bbox": [ + 274, + 143, + 727, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "research introduces a novel approach by employing MLLMs to tackle the intricate task of restoring ancient manuscript fragments. MLLMs facilitate efficient pairwise matching by: 1) identifying cross-image connections based on contour similarity and capturing complex features similarity in paper texture and content; 2) replicating the comprehensive integration of multimodal semantics that archaeologists employ during their assembly process; 3) demonstrating zero-shot learning capabilities, which are crucial for investigating manuscripts with limited ground truth availability.", + "bbox": [ + 212, + 305, + 784, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, the complexity of our task surpasses the current capabilities of MLLMs. In practical scenarios of ancient manuscript restoration, it is necessary to first identify potential candidate fragments from a larger collection for subsequent pairwise matching. This collection contains fragments that may not belong to the target manuscript, which we refer to as \"outliers\". Performing pairwise comparisons using MLLMs on the vast number of small fragments, including these outliers, leads to a combinatorial explosion challenge, which poses a significant difficulty for MLLM itself to manage effectively.", + "bbox": [ + 212, + 426, + 785, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these challenges, we formulate the fragment selection subtask as a combinatorial optimization (CO) problem. We introduce a novel framework, named LLMCO4MR, a LLMs-aided Neural Combinatorial Optimization solver for manuscript restoration, transforming the unstructured challenge into a technically feasible task. The highlights of the paper are:", + "bbox": [ + 212, + 547, + 785, + 623 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. More Open-world Problem Setting. Beyond the restricted setting requiring that there are no outlier fragments for recovery, to our best knowledge, we are the first to address the more realistic and challenging setting, i.e., restoring the manuscript from a subset of fragments with outliers.", + "2. New Two-stage Solving Pipeline. We solve the new problem by two stages: fragment selection and pairwise matching. In the first stage, $k$ fragments with the highest mutual assemble scores are selected by certain means while the latter stage involves the traditional closed-world problem setting.", + "3. Neural Combinatorial Solver for Candidate Selection. We introduce a combinatorial network to solve the fragment selection problem by formulating it as a $K$ -cardinality constrained task, with an advantage that the input visual fragments are learned end-to-end directly with the final decision goal. Supervised training is achieved by using the shredded fragments generated by splitting a complete manuscript into grids." + ], + "bbox": [ + 222, + 628, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang", + "bbox": [ + 256, + 114, + 732, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. MLLMs for Pairwise Fragment Matching in Zero-shot. The selected Top- $K$ fragments are then fed into the MLLM by pair, to generate the matching score as well as the relative direction of matching, i.e., horizontal or vertical. Then the final manuscript can be recovered by these local pairwise matchings. To our best knowledge, this is the first time to introduce MLLMs for manuscript recovery or more broadly speaking, jigsaw puzzle-like tasks.", + "bbox": [ + 222, + 145, + 787, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 262, + 397, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Manuscript Restoration. This area has recently witnessed significant advancements [2, 37] aided with vision technology for restoration. These developments aim to overcome the challenges in manual methods, which are time-consuming, labor-intensive, and heavily reliant on expert knowledge [2]. A typical method is to solve the problem from a jigsaw puzzle [10, 34] solving perspective. For puzzle solving and global restoration, several successful methods have emerged. Traditional strategies include curve-based techniques [27, 44] and time series sequence-based approaches [6, 13, 32, 39, 53]. Additionally, deep learning methods have been developed, exemplified by works such as [19] and other studies [17, 25, 28]. Inspired by these works, previous manuscript restoration methods are designed based on the visual features of the fragments. Here, matching often refers to identifying whether two fragments are adjacent and their relative direction (horizontally or vertically). In particular, the feature extraction module has been further advanced by deep learning with various aspects, including fragment contour shape [51, 52, 55], texture details [31], and written characters [29]. In the recent work [54], a multimodal restoration pipeline is devised that considers both written content and contour shape.", + "bbox": [ + 212, + 296, + 787, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite these advances, most current relic restoration methods share a common assumption: the candidate pieces for pairwise matching are given, either by manual pre-selection or other means. However, in many real-world restoration cases, this assumption leaves a gap in addressing the more challenging task of open-world problem setting, where candidate fragments are not predetermined, underscoring the novelty and necessity of our approach.", + "bbox": [ + 212, + 554, + 789, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Combinatorial Solver. The recent growing trend of learnable CO solvers [5] demonstrate the potential of solving CO problems by deep neural networks with higher efficiency [18, 22, 40, 42]. To tackle the extensive combinatorial space of ancient manuscript restoration, we present a cardinality-constrained CO reformulation [8] by selecting the Top- $K$ most likely matching fragments from the candidate set. Following the one-shot CO learning paradigm [41], we adopt [43], which handles cardinality-constrained problems similarly to differentiable optimal transport [46]. [43] has the advantage of controlling the tightness of constraints while preserving the gradient, which is crucial in our task.", + "bbox": [ + 212, + 651, + 787, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Large Language Models. With the recent arisen success of Large Language Model (LLM), such as GPT-4 [26], Qwen [3] and Baichuan [47], LLMs have showcased their efficacy in a myriad of fields including computer vision [23,24], natural", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8927eadc45cf92d434036fa4efd5f9da0435d52e26a7b3d39108f238fccf9283.jpg", + "image_caption": [ + "(a) Previous Ancient Manuscript Restoration Pipelines" + ], + "image_footnote": [], + "bbox": [ + 254, + 145, + 754, + 219 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8037dde0be00c3712ea819cf26f79ccc10207daff02b0758ee916cffd08d0ac9.jpg", + "image_caption": [ + "(b) Our Two-stage LLMCO4MS Pipeline", + "Fig. 3: Comparison between previous restoration approaches and our LLMCO4MR." + ], + "image_footnote": [], + "bbox": [ + 250, + 234, + 754, + 301 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/dc4af29bf9a36812077b0c4f47dbe9fa156192bdc1c052fb19cb97b924dac810.jpg", + "table_caption": [ + "Table 1: Comparison with various ancient fragment restoration SOTA." + ], + "table_footnote": [], + "table_body": "
MethodologyTraining DataAnnotation FreeEliminate OutliersNetwork StructureExtracted Features
JigsawNet [19]Self-constructed✓✘CNNTexture
Papyrus [31]Self-constructed✘✘Siamese CNNTexture
S3-Net [51]GAN-augmented✘✘Siamese CNNContour
LLMCO4MRSynthetic Shredding✓✓CO-MLLMMultimodal
", + "bbox": [ + 215, + 396, + 789, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "language processing [38], medical application [35,36], law judgement [33,50], and optimization [15, 16]. Recent advancements in Multimodal-LLM (MLLM) have catalyzed a proliferation of diverse applications, including image content understanding [56], mathematical reasoning [11], and even meme interpolation [49], among others. Inspired by MLLMs' potential in complex visual and textual content understanding, we proposed LLMCO4MR, the first CO solver and MLLM collaboration pipeline to tackle the complex ancient manuscript restoration task. We leverage MLLMs' potential extracting multimodal features [14,24] to tackle small ancient fragments pairwise matching, while further enhancing CO solver performance figuring similar fragments in real-world cases.", + "bbox": [ + 212, + 555, + 787, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion with Other Restoration Pipelines. We provide a qualitative comparison with other major jigsaw puzzle solving and fragment restoration pipelines, as shown in Tab. 1. The key distinctions between our LLMCO4MR pipeline and earlier approaches are highlighted in Fig. 3. Our pipeline adeptly addresses the combinatorial explosion challenge inherent in fragment selection through CO solver. Furthermore, it innovates by introducing the use of MLLMs for the pairwise matching of ancient manuscripts, an innovative approach for manuscript restoration with open-world setting.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang", + "bbox": [ + 256, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 143, + 380, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we provide a detailed description of our LLMCO4MR pipeline, as composed of two main components: a cardinality-constrained neural solver trained on synthetic data to select candidates for further matching; and the MLLM adoption to achieve effective matching in a zero-shot learning manner.", + "bbox": [ + 212, + 181, + 787, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Problem Formulation", + "text_level": 1, + "bbox": [ + 215, + 270, + 439, + 285 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We consider a set of $m$ fragments, denoted as $\\mathcal{F} = \\{f_{in}^{1},\\dots ,f_{in}^{k},f_{out}^{k + 1},\\dots ,f_{out}^{m}\\}$ . Here the first $k$ fragments $f_{in}^{i}$ refer to those originating from a more complete fragment, while the rest $(m - k)$ fragments denotes those outliers. The goal of LLMCO4MR is to identify such $k$ relevant fragments.", + "bbox": [ + 212, + 299, + 787, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 CO Solver for Fragment Selection", + "text_level": 1, + "bbox": [ + 215, + 388, + 540, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2.1 Preliminaries. Based on consultations with archaeology specialists and their insights gained from experimental endeavors, we have formulated the following hypotheses to guide the restoration:", + "bbox": [ + 212, + 419, + 785, + 465 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Internal Texture and Limited Text. Small real-world fragments typically contain very limited characters, and ancient text recognition is unreliable for matching due to low accuracy. Moreover, even fragments that should be joined together may exhibit different textures due to varying museum collection conditions. Therefore, distinguishing fragments by textual content or texture patterns poses a significant challenge.", + "2. Edge Shapes and Historical Tearing. Small fragments were often deliberately torn by people in ancient times [12], resulting in contours that are not severely eroded. This key feature offers the potential for primary restoration through contour similarity.", + "3. Reassembly Group Size. Fragments are typically reassembled in groups ranging from 3 to 9 pieces. It is uncommon to find a large assembly of pieces that can be combined into a significantly larger fragment." + ], + "bbox": [ + 220, + 479, + 785, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we propose a cardinality-constrained neural CO solver to efficiently and robustly select manuscript fragments, which is further exemplified with MLLM with application on the famous real-world Dunhuang Manuscript.", + "bbox": [ + 212, + 690, + 785, + 737 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2.2 Solver Overview. We formulate fragment selection task as cardinality-constrained optimization problem in Eq. (1a). Specifically, the number of nonnegative elements must be restricted to no greater than $k$ , while minimizing the objective $L_{co}$ . The aim of our formulation is to find the optimal fragments subset with a size of $k$ with highest mutual similarity, as shown in Eq. (1b), where $D_{ij}$", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dd2512891180b48494ff65d0bd37c00f60336b806c5d3938433773ddd9308f27.jpg", + "image_caption": [ + "Fig. 4: LLMCO4MR solving pipeline. 1. Fragment Selection with CO solver. Given a set of $m$ fragments, CO solver first pairs and feeds them into a siamese network to obtain similarity. Next, construct similarity graph $\\mathcal{G}$ and pass it through GNN. Then encode cardinality constraints with differentiate OT layer. Finally output the selected Top- $K$ fragments with highest mutual similarity. 2. Pairwise matching with MLLM in zero-shot. Given the selected Top- $K$ fragments by pair as visual prompt with text prompt, MLLMs generate matching score and relative directions as output." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 782, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "denotes the similarity score between candidates $i$ and $j$ , which are predicted by siamese network as described in the following section.", + "bbox": [ + 212, + 530, + 784, + 560 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {S}} L _ {c o} \\quad \\text {s . t .} \\mathcal {S} \\in \\{0, 1 \\} ^ {m}, \\| \\mathcal {S} \\| _ {0} \\leq k, \\tag {1a}\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 570, + 785, + 593 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} L _ {c o} = - \\sum_ {i \\in \\{i | s _ {i} = 1 \\}} \\left(\\sum_ {j \\in \\{j | s _ {j} = 1, j \\neq i \\}} D _ {i j}\\right). \\tag {1b}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 594, + 785, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This cardinality-constrained CO problem could be tackled by a differentiable OT layer [43] and Sinkhorn algorithm [9]. As shown in Fig. 4, we construct an end-to-end learning pipeline consisting of a siamese feature extraction module as problem encoder network, a graph neural network fused with an OT layer as the solver network. The neural network learns to solve Eq. (1a) under the unsupervised loss described in Eq. (3) as follows.", + "bbox": [ + 212, + 652, + 787, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2.3 Siamese Feature Extraction Network. In this section, we introduce the siamese feature extraction network of CO solver. Given the scarcity of real-world ground truth data, as few archaeologists have successfully restored small manuscript fragments together, we first describe the synthetic shredding process for preparing training data.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang", + "bbox": [ + 256, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a6a77b1041e75cc2b5bec9396bd7f197136b4ace57844f82f871ffa2d79e3d3b.jpg", + "image_caption": [ + "Fig. 5: Illustration of synthesized shredding process." + ], + "image_footnote": [], + "bbox": [ + 246, + 143, + 754, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Synthetic Shredding Process. We draw an analogy between the intricate work of manuscript restoration and grid puzzle assembly. To simulate this, we generate synthetic shredded fragments by breaking larger manuscript pieces into smaller, puzzle-piece-like fragments. Our synthesis approach is conducted in two phases, as illustrated in Fig. 5:", + "bbox": [ + 212, + 380, + 787, + 458 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Contour Extraction. We begin by randomly selecting large, real-world manuscripts from which we extract binary boundaries, specifically focusing on the two horizontal and two vertical edges. We discard contours that are either too long or too short, retaining only those of an appropriate length to form our contour set $\\mathcal{C} = \\{C_1,C_2,\\ldots ,C_m\\}$ . Each contour in this set is represented by a series of two-dimensional coordinates $C_i = \\{(x_1,y_1),(x_2,y_2),\\dots ,(x_n,y_n)\\}$ . Through this method, we ensure the synthetic fragments are realistic in size and shape.", + "2. Synthetic Shredding. We proceed to sample from the set $\\mathcal{C}$ without replacement to create an $M\\times M$ grid-like segmentation mask. Utilizing a real rectangular-shaped manuscript piece as our foundation, we apply this mask to segment it into $M^2$ smaller, discrete fragments. This division simulates realistic fragment sizes and shapes encountered in manuscript restoration." + ], + "bbox": [ + 222, + 470, + 787, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Siamese Network. We employ a siamese feature extraction network designed to assess contour similarity. We start with the original set of fragments with outliers $\\mathcal{F} = \\{f_{in}^{1},\\dots,f_{in}^{k},f_{out}^{k + 1},\\dots,f_{out}^{m}\\}$ . For each fragment, we create permutation combinations of pairwise matches $(f^i,f^j)$ . These paired fragments, specifically their masked contour, are then fed into two identical branches of a CNN-based neural network. The network computes a similarity score $D_{ij} = \\text{Siamese}(f^i,f^j)$ with the range being [0-1]. A score of 1 indicates that two fragments have matching contours that align perfectly, suggesting adjacency. Conversely, a score of 0 implies not adjacent. Eq. (2) describes the loss term, $w$ means learnable parameters, $L_{\\text{siamese}}$ means Binary Cross Entropy (BCE) Loss,", + "bbox": [ + 212, + 688, + 787, + 843 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$p$ means predicted similarity score and $y$ means ground truth label.", + "bbox": [ + 212, + 146, + 702, + 161 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s i a m e s e} = - w [ y \\cdot \\log p + (1 - y) \\cdot \\log (1 - p) ]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 174, + 785, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.2.4 Solver Network. Our CO solver network consists of a graph neural network fused with a differentiable OT layer.", + "text_level": 1, + "bbox": [ + 212, + 213, + 782, + 244 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Graph Neural Network. Following the siamese feature extraction network, we construct an undirected similarity graph $\\mathcal{G} = (\\mathcal{V},\\mathcal{E})$ . Graph vertices set $\\mathcal{V} = \\{v_{1},\\ldots ,v_{m}\\}$ stores the extracted fragment mask feature vector. Given the pairwise similarity scores $D_{ij}$ as edge weights, we add an edge $e_{ij}$ to the edge set $\\mathcal{E}$ if similarity score is greater than threshold. We then constructed a graph neural network to extract features from graph vertices and edges. We use Graph Convolution Network (GCN) as backbone.", + "bbox": [ + 212, + 250, + 784, + 354 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "OT Layer and training. For Top- $K$ fragments selection, we encode the cardinality constraint with differential OT layer following [41]. To encode cardinality constraints in a differentiable manner, we first formulate the constraints as marginal distributions in optimal transport (OT), and then solve OT with Sinkhorn algorithm [9] by iterative row-wise and column-wise normalization. Given $m$ GNN output score $p_i$ indicating the probability of selecting each piece, OT layer moves Top- $K$ probability to $p_{max}$ while moves others to $p_{min}$ . We select $K$ nodes with the highest probability to predict the output and optimize neural network with loss function in Eq. (3). Loss function consists of three components: $L_{siamese}$ penalize incorrect prediction in pairwise similarity; $L_{gnn}$ denotes the BCE loss for node selection; and $L_{co}$ indicates the optimization objective in the previously mentioned cardinality constrained CO.", + "bbox": [ + 212, + 362, + 787, + 544 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\min L = \\min \\left(\\alpha L _ {\\text {s i a m e s e}} + \\beta L _ {\\text {g n n}} + L _ {\\text {c o}}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 559, + 785, + 575 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.2.5 Second-order Matching As detailed in Section 3.1, fragment groups ranging from 3 to 9 pieces can be reconstructed into a final form. We explore different scenarios of grid-like restoration. In the case of a $2 \\times 2$ grid, each ground truth fragment is expected to have two pairwise connections with its neighboring fragments, treating all four fragments as equally important. In contrast, within a $3 \\times 3$ puzzle grid, as illustrated in Fig. 6, fragments positioned at [2, 4, 6, and 8] are observed to share three connections each, highlighting the central fragment [5] as most crucial with up to four connections. Meanwhile, the corner fragments [1, 3, 7, 9] are limited to just two connections. Acknowledging the imperfection of siamese network accuracy and the existence of false positive pairwise matches in real-world scenarios, we implement a second-order matching strategy. We start with identifying the Top-5 central pieces that with a higher number of connections. Subsequently, we refine the graph $\\mathcal{G}$ prune by pruning edges, retaining only those connected to the previously selected Top-5 fragments and eliminating other distracting edges. This strategy is designed to mitigate the impact of inaccuracies and ensure a more reliable restoration process.", + "bbox": [ + 212, + 598, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/93b3975e7be84aeea621c725f1d038c3c2b0ddb8e691959455596412b6178129.jpg", + "image_caption": [ + "(a) Select with Equivalent Importance", + "Fig. 6: Illustration of second-order matching. (a) selects all Top-9 fragments with equivalent importance (dashed lines represent incorrect pairings obtained by siamese network). (b) highlights centered 5 fragments with more connections, then select the rest 4 fragments based on pruned graph $\\mathcal{G}_{prune}$ (most incorrect pairings are eliminated)." + ], + "image_footnote": [], + "bbox": [ + 289, + 143, + 431, + 323 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/0db4f04202d4d6949e8ad06bedcc8db9db77c21e1a2247f9e705914fe658b593.jpg", + "image_caption": [ + "(b) Second-order Matching" + ], + "image_footnote": [], + "bbox": [ + 470, + 142, + 728, + 321 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3.3 LLMCO4MR Pipeline for Manuscript Restoration", + "text_level": 1, + "bbox": [ + 214, + 435, + 674, + 450 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Traditional deep learning methods for manuscript restoration heavily depend on human intervention to filter out numerous incorrect pairings generated by models. This task becomes increasingly challenging with smaller fragments, which are abundant and tiny, making manual review burdensome.", + "bbox": [ + 212, + 462, + 784, + 521 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The advent of MLLMs offers high-quality, efficient solutions to tasks that were previously labor-intensive. Innovatively, we are among the first to harness MLLMs for manuscript restoration. MLLMs like GPT-4V have shown remarkable capabilities akin to those of archaeologists by incorporating multimodal semantics for precise pairwise matching, excelling in identifying local pairwise connections with scoring and directional guidance. However, MLLMs struggle with the vast possibilities in open-world settings, particularly with outliers, limiting their ability to perform restoration independently.", + "bbox": [ + 212, + 523, + 787, + 643 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Addressing these challenges, we present the LLMCO4MR pipeline, a pioneering MLLM-aided Neural Combinatorial Optimization solver tailored for the restoration of small manuscript fragments. LLMCO4MR leverages a CO solver for initial fragment selection, effectively countering the combinatorial explosion by sidelining outliers. Subsequently, the Top- $K$ selected pairs undergo analysis via MLLMs, which utilize custom prompts based on a role-play strategy [45] to refine matching precision. If MLLMs identify any Top- $K$ selected fragments as having no viable pairing, these anomalies are removed from the CO solver's input, suggesting a reevaluation of the Top- $K$ fragment selections. This iterative refinement enhances LLMCO4MR efficacy. Experimental results in Section 4.3, demonstrate how this synergistic approach not only improves the precision in selecting fragments but also significantly enhances the accuracy and efficiency of manuscript restoration in practical scenarios.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Quantitative Evaluation and Case Study", + "text_level": 1, + "bbox": [ + 215, + 143, + 643, + 162 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Our experiments encompasses both synthetic data and real-world manuscript fragments. For fragment selection, we conducted simulation experiments using the CO solver, to demonstrate its capability in selecting appropriate fragment candidates for subsequent pairwise matching. We then deployed our LLMCO4MR pipeline on real-world scenarios. Specifically, we conducted case studies on the famous Dunhuang manuscripts, spanning from the 4th to the 14th century. These manuscripts are highly valued by archaeologists and the global community [12], due to their extensive volume, diverse content, and broad historical span.", + "bbox": [ + 212, + 176, + 787, + 297 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The combination of simulation experiments and real-world case studies provides a controlled and ideal environment for the quantitative evaluation of our method, aligning with the protocols of peer studies [30, 31, 51, 54]. These experiments were performed on a server equipped with Intel i7 CPUs at $3.2\\mathrm{GHz}$ and NVIDIA GeForce RTX 3090 GPUs.", + "bbox": [ + 212, + 297, + 787, + 375 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Experiment Setup", + "text_level": 1, + "bbox": [ + 215, + 397, + 413, + 414 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Data. For fragment selection, simulation experiments were employed to train the CO solver using synthetic shredded data, as elaborated in Section 3.2.3. Our methodology ensured that fragments generated from actual manuscripts, including their contours, accurately reflect the complexities found in real-world restoration conditions. For the case studies involving LLMCO4MR pipeline, high-quality (with annotated ground truth) real-world Dunhuang fragments are indeed scarce. Our annotated fragment cases were provided by the expert team dedicated to Dunhuang culture research, in the form of visual images for evaluation. All images were obtained from the International Dunhuang Project (IDP).", + "bbox": [ + 212, + 429, + 787, + 566 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation. For fragment selection with CO solver, we evaluated on synthetic shredded data with accuracy criterion. For instance, in a scenario where the task is to select Top-4 pieces from 10 candidates, fragment correctly identified as one of the 4 ground truth fragments is deemed a successful selection. For the real-world case study on Dunhuang, we evaluate the performance by measuring the success rate of fragment restoration. A successful restoration is defined as a match that aligns with the ground truth candidate pairings provided by archaeologists. This expert-annotated data serves as our benchmark for assessing the accuracy of our method in real-world scenarios.", + "bbox": [ + 212, + 571, + 787, + 708 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Simulation Experiment on Combinatorial Solver", + "text_level": 1, + "bbox": [ + 215, + 731, + 660, + 747 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Training Details. Aligned with Section 3.1, CO solver is designed to select $2 \\times 2$ and $3 \\times 3$ fragments from a set $\\mathcal{F}$ comprising $m$ fragments, aiming to address the Top- $K$ fragment selection challenge. Thus we initially selected $k$ pieces that originate from the same fragment and then added $(m - k)$ random fragments as outliers to compile a group. We assembled 500 such groups as", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/34b5a26d3902e401402c61cc28eefb4aaf9e199ca3b306b66895d1d40c939759.jpg", + "table_caption": [ + "Table 2: Performance of CO solver siamese feature extraction network." + ], + "table_footnote": [], + "table_body": "
Data QuantityHybrid RatioSiamese AccuracyTop-4 Performance
400001:10.87460.6925
2:30.92120.7025
1:20.91170.7100
800001:10.95340.7250
2:30.94370.7325
1:20.96750.7500
1200001:10.92450.7525
2:30.95860.7575
1:20.97310.7800
", + "bbox": [ + 346, + 170, + 653, + 353 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/221f62da7b5b247341f26dc37bc05b0ace2f0656beb4ec46c7871b5f47c337a3.jpg", + "table_caption": [ + "Table 3: Performance of CO solver on various Top- $K$ fragments selection task." + ], + "table_footnote": [], + "table_body": "
Top-K2×23×3
Pool Size101520202530
Random Select0.40000.26670.20000.45000.36000.3000
Baseline0.74750.31250.25750.67560.57670.5389
OT Layer0.78000.58250.42250.72560.64440.6044
", + "bbox": [ + 263, + 396, + 733, + 489 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "training dataset, reserving an additional 100 groups for evaluation. We trained the siamese feature extraction network with 40K, 80K and 120K positive and negative pairing samples, with the ratio of training and testing being 6:1.", + "bbox": [ + 212, + 521, + 784, + 568 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Siamese Feature Extraction Network. As detailed in Table 2, we evaluate the performance across various training data quantity, positive and negative sample ratios utilizing a ResNet backbone. Our findings indicate that CO solver performance peaks on Top-4 selections with 120K samples and 1:2 hybrid ratio on positive and negative matching samples. This can be attributed to the imbalance between positive and negative samples in pairwise matching manuscript fragments. With more negative samples, the reliability of siamese model is improved and CO solver performance is enhanced.", + "bbox": [ + 212, + 574, + 787, + 696 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "CO Solver. Tab. 3 presents the simulation experiment results under various configurations of Top- $K$ fragment selection from $m$ candidates. \"Random select\" refers to the Top- $K$ random selections mathematical expectation. \"Baseline\" refers to selecting the Top- $K$ probabilities in a greedy manner from GNN output without OT layer. The introduction of the OT layer, which enforces cardinality constraints, is observed to enhance model performance, particularly as the candidate size $m$ increases. This improvement is attributed to the OT layer's rigorous penalization of deviations from the optimization objectives in Eqs. (1a) and (1b).", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3513813278b5b64bc0e1467824790a3b0c62164eb090244422c05c964986a56e.jpg", + "table_caption": [ + "Table 4: Enhancement from the second-order Top- $K$ selection." + ], + "table_footnote": [], + "table_body": "
StrategyTop-9 of 20Top-9 of 25Top-9 of 30
Select at Once0.72560.64440.6044
Second-Order0.73110.68220.6633
", + "bbox": [ + 308, + 170, + 687, + 229 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Enhancement through Second-order Matching Strategy. In the $2 \\times 2$ selection settings, all four ground truth fragments hold equal significance (each fragment only have exactly 2 matching connections with their ground-truth adjacent pieces). Thus we explore the efficacy of the second-order matching strategy introduced in Section 3.2.5 for $3 \\times 3$ scenarios and detail the results in Tab. 4. The second-order matching strategy duplicates the cardinality constraint and applies more stringent penalties for incorrect selections. This method aims to mitigate the effects of erroneous pairwise matches proposed by the siamese network, thereby advancing the accuracy and reliability of fragment selection.", + "bbox": [ + 212, + 260, + 785, + 396 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Simulation experiments demonstrate our proposed neural CO solver's effectiveness in selecting candidates based on highest mutual assemble similarity. This gives promise for the subsequent real-world restoration with LLMCO4MR.", + "bbox": [ + 212, + 397, + 785, + 441 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Real-world Manuscript Case Studies on Dunhuang", + "text_level": 1, + "bbox": [ + 214, + 469, + 679, + 484 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results with LLMCO4MR. We conducted case study experiments utilizing LLMCO4MR and compared its effectiveness against other SOTAs, as detailed in Tab. 5. We input a group of fragments with outliers to CO solver for selection, then use the selected Top- $K$ fragments as image prompts for GPT-4V, which suggests the matching scores and relative directions. For SOTAs [31,51] that lack the capability to filter out outliers, we repeatedly input the fragments pairs and select the matching results in a greedy manner (select the fragment candidate pairs with ranked highest matching score).", + "bbox": [ + 212, + 505, + 785, + 626 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We observed that previous works often operated under restricted conditions, such as assuming no outliers, limiting search spaces, regular shape fragments, or using manually outlined perfect matching contours. These constraints do not reflect the open-world setting addressed by our method. Moreover, contour similarity network does not achieve perfect accuracy in matching judgments and real-world scenarios frequently involve more irregular fragments, underscoring the complexity of the task.", + "bbox": [ + 212, + 628, + 785, + 733 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our pipeline addresses these challenges in two ways. First, the CO solver captures irregularity through data augmentation. We extract contours from real-world Dunhuang manuscripts and apply to shredding to adequately represent real-world challenges and mitigate the risk of overfitting unrealistic characteristics in training. Second, the MLLM component captures cross-image similarity using multimodal features, including contour (even with some imperfect fits and irregular shapes) and other visual features.", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang", + "bbox": [ + 256, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/5d80e979cff52688d877e48ac551b48d9a8b17c4d32235586a85a615c901ffe9.jpg", + "table_caption": [ + "Table 5: Evaluation on real-world Dunhuang fragment restoration case studies." + ], + "table_footnote": [], + "table_body": "
Top-K2×23×3
Pool Size10152025
GPT-4V [26]0.32500.17500.37780.2533
LLaVA [21]0.31500.12500.27780.1556
JigsawNet [19]0.52500.37500.45560.3222
Papyrus [31]0.42500.37500.47780.4111
S3-Net [51]0.31250.25750.38890.2111
CO Solver0.57500.52500.53330.4667
LLMCO4MR (Ours)0.67500.62500.62220.5556
", + "bbox": [ + 285, + 170, + 712, + 333 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ablation Study. We also present ablation study results in Tab. 5. \"CO Solver\" refers to deriving matching pairs directly from the output of CO solver. Few selected MLLMs exhibit capabilities in understanding cross-image fragment relations. These MLLMs' performances were evaluated by processing pairs of fragments without CO solver filtering out outliers. We conclude that similar to retrieval-augmented LLM frameworks, CO solver in our LLMCO4MR is a nontrivial extension, empowers MLLM capability handling combinatorial explosion problems. We detail in the Appendix the current limitations of MLLMs in directly addressing manuscript restoration tasks with combinatorial explosion.", + "bbox": [ + 212, + 362, + 787, + 498 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our multifaceted approach serves two purposes: it complements and enhances traditional contour-based matching, and it helps to stabilize the variable performance of MLLMs in image matching. As a result, our method demonstrates increased effectiveness in handling real-world manuscript restoration tasks, especially in scenarios where either matching strategy alone is insufficient.", + "bbox": [ + 212, + 500, + 787, + 575 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 604, + 359, + 619 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have presented a novel pipeline, LLMCO4MR, designed for the intricate task of restoring ancient manuscripts from fragments. We first formalize the problem as a cardinality-constrained combinatorial problem and then a neural combinatorial solver is devised to produce the Top- $K$ candidates for merging. Then the MLLM is employed, leveraging the candidates as prompt for final restoration. Experiments on both synthetic fragments and real-world Dunhuang data demonstrate the potential of our approach, compared with the traditional approaches which require labors of well-trained archaeologists.", + "bbox": [ + 212, + 641, + 787, + 763 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "**Limitation:** Like existing works, our current approach also cannot handle the case with missing fragments and the fragments are required to be basically in a grid layout. Another unsolved case is the existence of both small and large fragments, which we leave for future work. Also, more case studies on broader scope need to be studied when such data becomes available.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was in part supported by National Science and Technology Major Project (2022ZD0119100), National Natural Science Foundation of China (No. 62441605, 72342023), National Social Science Fund of China (No. 21BYY142, 14AZS001), Key Research and Development Program of Zhejiang Province (No. 2024C03270), the StarryNight Science Fund of Zhejiang University Shanghai Institute for Advanced Study (SN-ZJU-SIAS-0010).", + "bbox": [ + 212, + 175, + 787, + 267 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 287, + 321, + 303 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abitbol, R., Shimshoni, I., Ben-Dov, J.: Machine learning based assembly of fragments of ancient papyrus. Journal on Computing and Cultural Heritage (JOCCH) 14(3), 1-21 (2021)", + "2. Assael, Y., Sommerschield, T., Shillingford, B., Bordbar, M., Pavlopoulos, J., Chatzipanagiotou, M., Androutsopoulos, I., Prag, J., de Freitas, N.: Restoring and attributing ancient texts using deep neural networks. Nature 603(7900), 280-283 (2022)", + "3. Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., Hui, B., Ji, L., Li, M., Lin, J., Lin, R., Liu, D., Liu, G., Lu, C., Lu, K., Ma, J., Men, R., Ren, X., Ren, X., Tan, C., Tan, S., Tu, J., Wang, P., Wang, S., Wang, W., Wu, S., Xu, B., Xu, J., Yang, A., Yang, H., Yang, J., Yang, S., Yao, Y., Yu, B., Yuan, H., Yuan, Z., Zhang, J., Zhang, X., Zhang, Y., Zhang, Z., Zhou, C., Zhou, J., Zhou, X., Zhu, T.: Qwen technical report (2023)", + "4. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond (2023)", + "5. Bengio, Y., Lodi, A., Prouvost, A.: Machine learning for combinatorial optimization: a methodological tour d'horizon. European Journal of Operational Research 290(2), 405-421 (2021)", + "6. Besl, P.J., McKay, N.D.: Method for registration of 3-d shapes. In: Sensor fusion IV: control paradigms and data structures. vol. 1611, pp. 586-606. Spie (1992)", + "7. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)", + "8. Buchbinder, N., Feldman, M., Naor, J., Schwartz, R.: Submodular maximization with cardinality constraints. In: Proceedings of the twenty-fifth annual ACM-SIAM symposium on Discrete algorithms. pp. 1433-1452. SIAM (2014)", + "9. Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems 26 (2013)", + "0. Derech, N., Tal, A., Shimshoni, I.: Solving archaeological puzzles. Pattern Recognition 119, 108065 (2021)", + "11. Driess, D., Xia, F., Sajjadi, M.S.M., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., Huang, W., Chebotar, Y., Sermanet, P., Duckworth, D., Levine, S., Vanhoucke, V., Hausman, K., Toussaint, M., Greff, K., Zeng, A., Mordatch, I., Florence, P.: Palm-e: An embodied multimodal language model (2023)", + "12. Galambos, I.: Dunhuang Manuscript Culture: End of the First Millennium, vol. 22. Walter de Gruyter GmbH & Co KG (2020)" + ], + "bbox": [ + 218, + 316, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. da Gama Leitao, H.C., Stolfi, J.: A multiscale method for the reassembly of two-dimensional fragmented objects. IEEE Transactions on Pattern Analysis and Machine Intelligence 24(9), 1239-1251 (2002)", + "14. Ge, J., Luo, H., Qian, S., Gan, Y., Fu, J., Zhang, S.: Chain of thought prompt tuning in vision language models (2023)", + "15. Guo, P.F., Chen, Y.H., Tsai, Y.D., Lin, S.D.: Towards optimizing with large language models. arXiv preprint arXiv:2310.05204 (2023)", + "16. Guo, Q., Wang, R., Guo, J., Li, B., Song, K., Tan, X., Liu, G., Bian, J., Yang, Y.: Connecting large language models with evolutionary algorithms yields powerful prompt optimizers. arXiv preprint arXiv:2309.08532 (2023)", + "17. Hossieni, S.S., Shabani, M.A., Irandoust, S., Furukawa, Y.: Puzzlefusion: Unleashing the power of diffusion models for spatial puzzle solving. Advances in Neural Information Processing Systems 36 (2024)", + "18. Karalias, N., Loukas, A.: Erdos goes neural: an unsupervised learning framework for combinatorial optimization on graphs. Advances in Neural Information Processing Systems 33, 6659-6672 (2020)", + "19. Le, C., Li, X.: Jigsawnet: Shredded image reassembly using convolutional neural network and loop-based composition. IEEE Transactions on Image Processing 28(8), 4000-4015 (2019)", + "20. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023)", + "21. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)", + "22. Lu, J., Sun, Y., Huang, Q.: Jigsaw: Learning to assemble multiple fractured objects. arXiv preprint arXiv:2305.17975 (2023)", + "23. Lyu, C., Wu, M., Wang, L., Huang, X., Liu, B., Du, Z., Shi, S., Tu, Z.: Macaw-lm: Multi-modal language modeling with image, audio, video, and text integration (2023)", + "24. Maaz, M., Rasheed, H., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models (2023)", + "25. Markaki, S., Panagiotakis, C.: Jigsaw puzzle solving techniques and applications: a survey. The Visual Computer pp. 1-17 (2022)", + "26. OpenAI: Gpt-4 technical report (2023)", + "27. Panagiotakis, C., Markaki, S., Kokinou, E., Papadakis, H.: Coastline matching via a graph-based approach. Computational Geosciences 26(6), 1439-1448 (2022)", + "28. Paumard, M.M., Picard, D., Tabia, H.: Jigsaw puzzle solving using local feature co-occurrences in deep neural networks. In: 2018 25th IEEE international conference on image processing (ICIP). pp. 1018-1022. IEEE (2018)", + "29. Pengcheng, G., Gang, G., Jiangqin, W., Baogang, W.: Chinese calligraphic style representation for recognition. International Journal on Document Analysis and Recognition (IJDAR) 20, 59-68 (2017)", + "30. Pirrone, A., Aimar, M.B., Journet, N.: Papy-s-net: A siamese network to match papyrus fragments. In: Proceedings of the 5th International Workshop on Historical Document Imaging and Processing. pp. 78-83 (2019)", + "31. Pirrone, A., Beurton-Aimar, M., Journet, N.: Self-supervised deep metric learning for ancient papyrus fragments retrieval. International Journal on Document Analysis and Recognition (IJDAR) 24(3), 219-234 (2021)", + "32. Rusinkiewicz, S., Levoy, M.: Efficient variants of the icp algorithm. In: Proceedings third international conference on 3-D digital imaging and modeling. pp. 145-152. IEEE (2001)", + "33. Savelka, J., Ashley, K.D., Gray, M.A., Westermann, H., Xu, H.: Explaining legal concepts with augmented large language models (gpt-4) (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Savino, P., Tonazzini, A.: Digital restoration of ancient color manuscripts from geometrically misaligned recto-verso pairs. Journal of Cultural Heritage 19, 511-521 (2016)", + "35. Shuai, R.W., Ruffolo, J.A., Gray, J.J.: Generative language modeling for antibody design. bioRxiv (2022). https://doi.org/10.1101/2021.12.13.472419, https://www.biorxiv.org/content/early/2022/12/20/2021.12.13.472419", + "36. Singhal, K., Azizi, S., Tu, T., Mahdavi, S.S., Wei, J., Chung, H.W., Scales, N., Tanwani, A., Cole-Lewis, H., Pfohl, S., Payne, P., Seneviratne, M., Gamble, P., Kelly, C., Scharli, N., Chowdhery, A., Mansfield, P., y Arcas, B.A., Webster, D., Corrado, G.S., Matias, Y., Chou, K., Gottweis, J., Tomasev, N., Liu, Y., Rajkomar, A., Barral, J., Semturs, C., Karthikesalingam, A., Natarajan, V.: Large language models encode clinical knowledge (2022)", + "37. Sommerscheid, T., Assael, Y., Pavlopoulos, J., Stefanak, V., Senior, A., Dyer, C., Bodel, J., Prag, J., Androutsopoulos, I., de Freitas, N.: Machine learning for ancient languages: A survey. Computational Linguistics pp. 1-44 (2023)", + "38. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)", + "39. Tsamoura, E., Pitas, I.: Automatic color based reassembly of fragmented images and paintings. IEEE Transactions on Image Processing 19(3), 680-690 (2009)", + "40. Vinyals, O., Fortunato, M., Jaitly, N.: Pointer networks. Advances in neural information processing systems 28 (2015)", + "41. Wang, R., Shen, L., Chen, Y., Yang, X., Tao, D., Yan, J.: Towards one-shot neural combinatorial solvers: Theoretical and empirical notes on the cardinality-constrained case. In: The Eleventh International Conference on Learning Representations (2022)", + "42. Wang, R., Yan, J., Yang, X.: Learning combinatorial embedding networks for deep graph matching. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3056-3065 (2019)", + "43. Wang, R., Zhang, Y., Guo, Z., Chen, T., Yang, X., Yan, J.: Linsatnet: The positive linear satisfiability neural networks. In: International Conference on Machine Learning (ICML) (2023)", + "44. Wolfson, H., Schonberg, E., Kalvin, A., Lamdan, Y.: Solving jigsaw puzzles by computer. Annals of Operations Research 12(1), 51-64 (1988)", + "45. Xi, Z., Chen, W., Guo, X., He, W., Ding, Y., Hong, B., Zhang, M., Wang, J., Jin, S., Zhou, E., et al.: The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864 (2023)", + "46. Xie, Y., Dai, H., Chen, M., Dai, B., Zhao, T., Zha, H., Wei, W., Pfister, T.: Differentiable top-k with optimal transport. Advances in Neural Information Processing Systems 33, 20520-20531 (2020)", + "47. Yang, A., Xiao, B., Wang, B., Zhang, B., Bian, C., Yin, C., Lv, C., Pan, D., Wang, D., Yan, D., Yang, F., Deng, F., Wang, F., Liu, F., Ai, G., Dong, G., Zhao, H., Xu, H., Sun, H., Zhang, H., Liu, H., Ji, J., Xie, J., Dai, J., Fang, K., Su, L., Song, L., Liu, L., Ru, L., Ma, L., Wang, M., Liu, M., Lin, M., Nie, N., Guo, P., Sun, R., Zhang, T., Li, T., Li, T., Cheng, W., Chen, W., Zeng, X., Wang, X., Chen, X., Men, X., Yu, X., Pan, X., Shen, Y., Wang, Y., Li, Y., Jiang, Y., Gao, Y., Zhang, Y., Zhou, Z., Wu, Z.: Baichuan 2: Open large-scale language models (2023)", + "48. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of lmmss: Preliminary explorations with gpt-4v(ison) (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Yang, Z., Li, L., Wang, J., Lin, K., Azarnasab, E., Ahmed, F., Liu, Z., Liu, C., Zeng, M., Wang, L.: Mm-react: Prompting chatgpt for multimodal reasoning and action (2023)", + "50. Yu, F., Quartey, L., Schilder, F.: Legal prompting: Teaching a language model to think like a lawyer (2022)", + "51. Zhang, C., Wang, B., Chen, K., Zong, R., Mo, B.f., Men, Y., Almpanidis, G., Chen, S., Zhang, X.: Data-driven oracle bone rejoining: A dataset and practical self-supervised learning scheme. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 4482-4492 (2022)", + "52. Zhang, C., Zong, R., Cao, S., Men, Y., Mo, B.: Ai-powered oracle bone inscriptions recognition and fragments rejoining. In: Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence. pp. 5309-5311 (2021)", + "53. Zhang, K., Li, X.: A graph-based optimization algorithm for fragmented image reassembly. Graphical Models 76(5), 484-495 (2014)", + "54. Zhang, Y., Fang, Z., Yang, X., Zhang, S., He, B., Dou, H., Yan, J., Zhang, Y., Wu, F.: Reconnecting the broken civilization: Patchwork integration of fragments from ancient manuscripts. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 1157-1166 (2023)", + "55. Zhang, Z., Wang, Y.T., Li, B., Guo, A., Liu, C.L.: Deep rejoining model for oracle bone fragment image. In: Asian Conference on Pattern Recognition. pp. 3-15. Springer (2021)", + "56. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 479 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Zhang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/LLMCO4MR_ LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang/33e2dc3a-2be0-4119-88f6-223087048685_model.json b/2024/LLMCO4MR_ LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang/33e2dc3a-2be0-4119-88f6-223087048685_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f3b6eb19377ec8a6626f134ef8599c1ddbfce5bf --- /dev/null +++ b/2024/LLMCO4MR_ LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang/33e2dc3a-2be0-4119-88f6-223087048685_model.json @@ -0,0 +1,2403 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.219, + 0.141, + 0.784, + 0.21 + ], + "angle": 0, + "content": "LLMCO4MR: LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.235, + 0.782, + 0.281 + ], + "angle": 0, + "content": "Yuqing Zhang\\(^{1}\\), Hangqi Li\\(^{1}\\), Shengyu Zhang\\(^{1,2}\\), Runzhong Wang\\(^{3}\\), Baoyi He\\(^{1}\\), Huaiyong Dou\\(^{1}\\), Junchi Yan\\(^{3,4}\\), Yongquan Zhang\\(^{1}\\), and Fei Wu\\(^{1,2,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.292, + 0.636, + 0.307 + ], + "angle": 0, + "content": "1 Zhejiang University, Hangzhou, China" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.307, + 0.773, + 0.321 + ], + "angle": 0, + "content": "\\(^{2}\\) Shanghai Institute for Advanced Study of Zhejiang University, Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.321, + 0.67, + 0.335 + ], + "angle": 0, + "content": "3 Shanghai Jiao Tong University, Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.354, + 0.335, + 0.649, + 0.349 + ], + "angle": 0, + "content": "4 Shanghai AI Laboratory, Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.35, + 0.746, + 0.363 + ], + "angle": 0, + "content": "{yuqingz7, sy_zhang, wufei}@zju.edu.cn, yanjunchi@sjtu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.407, + 0.744, + 0.698 + ], + "angle": 0, + "content": "Abstract. Restoring ancient manuscripts fragments, such as those from Dunhuang, is crucial for preserving human historical culture. However, their worldwide dispersal and the shifts in cultural and historical contexts pose significant restoration challenges. Traditional archaeological efforts primarily focus on manually piecing major fragments together, yet the small and more intricate pieces remain largely unexplored, which is technically due to their irregular shapes, sparse textual content, and extensive combinatorial space for reassembly. In this paper, we formalize the task of restoring the ancient manuscript from fragments as a cardinality-constrained combinatorial optimization problem, and propose a framework named LLMCO4MR: (Multimodal) Large Language Model-aided Combinatorial Optimization Neural Networks for Ancient Manuscript Restoration. Specifically, LLMCO4MR encapsulates a neural combinatorial solver equipped with a differentiable optimal transport (OT) layer, to efficiently predict the Top-\\(K\\) likely mutual reassembly candidates. Multimodal Large Language Model (MLLM) is then adopted and prompted to yield pairwise matching confidence and relative directions for final restoration. Experiments on synthetic data and cases studies with real-world famous Dunhuang fragments demonstrate our approach's practical potential in assisting archaeologists. Our method provides a novel perspective for ancient manuscript restoration." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.71, + 0.741, + 0.724 + ], + "angle": 0, + "content": "Keywords: Ancient Manuscript Restoration \\(\\cdot\\) Dunhuang Manuscript \\(\\cdot\\)" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.724, + 0.707, + 0.738 + ], + "angle": 0, + "content": "Combinatorial Optimization \\(\\cdot\\) Multimodal Large Language Models" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.76, + 0.377, + 0.777 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.788, + 0.787, + 0.819 + ], + "angle": 0, + "content": "The restoration of ancient manuscripts has long been the focus of archaeologists' research and the global community. Traditionally, these codices are frequently" + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.826, + 0.393, + 0.84 + ], + "angle": 0, + "content": "* Corresponding Authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.162, + 0.438, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.236, + 0.462, + 0.25 + ], + "angle": 0, + "content": "(a) Conventional manuscript restoration task" + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.144, + 0.756, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.482, + 0.236, + 0.785, + 0.25 + ], + "angle": 0, + "content": "(b) Our task: Small fragments restoration with outliers" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.261, + 0.788, + 0.346 + ], + "angle": 0, + "content": "Fig. 1: Comparison of conventional manuscript restoration and our task. (a) highlights the conventional restoration task with large fragments. Archaeologists focus on written content as primary guide for restoration, supplemented by texture and edges. (b) outlines our task. In an open-world problem setting, small fragments contain limited text. Recovery relies on contour similarity matching. This is generally difficult for humans to identify fragment pairs from a large collection containing outliers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.785, + 0.442 + ], + "angle": 0, + "content": "discovered in a (severely) deteriorated condition due to the passage of time. Recovering these manuscripts involves extensive efforts [37] that are often time-consuming, labor-intensive, and require a high level of expertise, necessitating years of professional training [2]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.625 + ], + "angle": 0, + "content": "On the other hand, vision technologies have demonstrated significant potential in aiding the restoration of ancient manuscripts. For example, [1] developed a sophisticated system based on Convolutional Neural Network (CNN) to identify matching candidates among papyrus fragments by analyzing their fiber patterns. Zhang et al. [51] proposed a self-supervised model using Generative Adversarial Network (GAN) and siamese network for the pairwise stitching of Oracle Bone pieces, utilizing manually annotated contours for similarity comparison. While these vision-assisted pipelines enhance efficiency and accuracy beyond traditional manual efforts, their applications are generally confined to local pairwise matching [31], or regularly-shaped (e.g., rectangle) and large fragments [54]. A major challenge persists in dealing with real-world small fragments, which are often less-textual, shape-irregular, and scattered with missing fragments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.788, + 0.748 + ], + "angle": 0, + "content": "These small fragments have often been overlooked by archaeologists, yet they form a substantial part of the manuscript collections. Recent explorations have revealed that these minor pieces are treasure troves of first-hand materials, offering unique insights into the daily lives of past civilizations [12]. Current deep learning approaches rely on textual [29], contextual [2] or multimodal information [54] to reassembly manuscripts. Such learning-based methods require a large amount of annotated training data for complex multimodal feature extraction, which is impractical for small fragment restoration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To bridge the gap, we resort to the recent advancements in large language models (LLMs) that have revolutionized traditional tasks in computer vision and natural language processing [7, 26, 38]. Notably, the emergence of Multimodal Large Language Models (MLLMs) has demonstrated exceptional capabilities in extracting and interpreting image features, understanding content, and following complex prompt instructions [4, 20, 48]. Capitalizing on these developments, our" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.144, + 0.728, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.263, + 0.768, + 0.278 + ], + "angle": 0, + "content": "Fig. 2: Showcases of matching results with case studies on Dunhuang fragments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.306, + 0.785, + 0.426 + ], + "angle": 0, + "content": "research introduces a novel approach by employing MLLMs to tackle the intricate task of restoring ancient manuscript fragments. MLLMs facilitate efficient pairwise matching by: 1) identifying cross-image connections based on contour similarity and capturing complex features similarity in paper texture and content; 2) replicating the comprehensive integration of multimodal semantics that archaeologists employ during their assembly process; 3) demonstrating zero-shot learning capabilities, which are crucial for investigating manuscripts with limited ground truth availability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.787, + 0.547 + ], + "angle": 0, + "content": "However, the complexity of our task surpasses the current capabilities of MLLMs. In practical scenarios of ancient manuscript restoration, it is necessary to first identify potential candidate fragments from a larger collection for subsequent pairwise matching. This collection contains fragments that may not belong to the target manuscript, which we refer to as \"outliers\". Performing pairwise comparisons using MLLMs on the vast number of small fragments, including these outliers, leads to a combinatorial explosion challenge, which poses a significant difficulty for MLLM itself to manage effectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.787, + 0.624 + ], + "angle": 0, + "content": "To address these challenges, we formulate the fragment selection subtask as a combinatorial optimization (CO) problem. We introduce a novel framework, named LLMCO4MR, a LLMs-aided Neural Combinatorial Optimization solver for manuscript restoration, transforming the unstructured challenge into a technically feasible task. The highlights of the paper are:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.63, + 0.785, + 0.689 + ], + "angle": 0, + "content": "1. More Open-world Problem Setting. Beyond the restricted setting requiring that there are no outlier fragments for recovery, to our best knowledge, we are the first to address the more realistic and challenging setting, i.e., restoring the manuscript from a subset of fragments with outliers." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.69, + 0.785, + 0.749 + ], + "angle": 0, + "content": "2. New Two-stage Solving Pipeline. We solve the new problem by two stages: fragment selection and pairwise matching. In the first stage, \\( k \\) fragments with the highest mutual assemble scores are selected by certain means while the latter stage involves the traditional closed-world problem setting." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "3. Neural Combinatorial Solver for Candidate Selection. We introduce a combinatorial network to solve the fragment selection problem by formulating it as a \\(K\\)-cardinality constrained task, with an advantage that the input visual fragments are learned end-to-end directly with the final decision goal. Supervised training is achieved by using the shredded fragments generated by splitting a complete manuscript into grids." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.63, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.146, + 0.788, + 0.239 + ], + "angle": 0, + "content": "4. MLLMs for Pairwise Fragment Matching in Zero-shot. The selected Top-\\(K\\) fragments are then fed into the MLLM by pair, to generate the matching score as well as the relative direction of matching, i.e., horizontal or vertical. Then the final manuscript can be recovered by these local pairwise matchings. To our best knowledge, this is the first time to introduce MLLMs for manuscript recovery or more broadly speaking, jigsaw puzzle-like tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.263, + 0.398, + 0.28 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.789, + 0.554 + ], + "angle": 0, + "content": "Manuscript Restoration. This area has recently witnessed significant advancements [2, 37] aided with vision technology for restoration. These developments aim to overcome the challenges in manual methods, which are time-consuming, labor-intensive, and heavily reliant on expert knowledge [2]. A typical method is to solve the problem from a jigsaw puzzle [10, 34] solving perspective. For puzzle solving and global restoration, several successful methods have emerged. Traditional strategies include curve-based techniques [27, 44] and time series sequence-based approaches [6, 13, 32, 39, 53]. Additionally, deep learning methods have been developed, exemplified by works such as [19] and other studies [17, 25, 28]. Inspired by these works, previous manuscript restoration methods are designed based on the visual features of the fragments. Here, matching often refers to identifying whether two fragments are adjacent and their relative direction (horizontally or vertically). In particular, the feature extraction module has been further advanced by deep learning with various aspects, including fragment contour shape [51, 52, 55], texture details [31], and written characters [29]. In the recent work [54], a multimodal restoration pipeline is devised that considers both written content and contour shape." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.79, + 0.646 + ], + "angle": 0, + "content": "Despite these advances, most current relic restoration methods share a common assumption: the candidate pieces for pairwise matching are given, either by manual pre-selection or other means. However, in many real-world restoration cases, this assumption leaves a gap in addressing the more challenging task of open-world problem setting, where candidate fragments are not predetermined, underscoring the novelty and necessity of our approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.789, + 0.789 + ], + "angle": 0, + "content": "Combinatorial Solver. The recent growing trend of learnable CO solvers [5] demonstrate the potential of solving CO problems by deep neural networks with higher efficiency [18, 22, 40, 42]. To tackle the extensive combinatorial space of ancient manuscript restoration, we present a cardinality-constrained CO reformulation [8] by selecting the Top- \\(K\\) most likely matching fragments from the candidate set. Following the one-shot CO learning paradigm [41], we adopt [43], which handles cardinality-constrained problems similarly to differentiable optimal transport [46]. [43] has the advantage of controlling the tightness of constraints while preserving the gradient, which is crucial in our task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Large Language Models. With the recent arisen success of Large Language Model (LLM), such as GPT-4 [26], Qwen [3] and Baichuan [47], LLMs have showcased their efficacy in a myriad of fields including computer vision [23,24], natural" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.146, + 0.756, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.22, + 0.647, + 0.233 + ], + "angle": 0, + "content": "(a) Previous Ancient Manuscript Restoration Pipelines" + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.235, + 0.756, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.304, + 0.606, + 0.317 + ], + "angle": 0, + "content": "(b) Our Two-stage LLMCO4MS Pipeline" + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.328, + 0.779, + 0.343 + ], + "angle": 0, + "content": "Fig. 3: Comparison between previous restoration approaches and our LLMCO4MR." + }, + { + "type": "table_caption", + "bbox": [ + 0.262, + 0.371, + 0.739, + 0.385 + ], + "angle": 0, + "content": "Table 1: Comparison with various ancient fragment restoration SOTA." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.397, + 0.79, + 0.514 + ], + "angle": 0, + "content": "
MethodologyTraining DataAnnotation FreeEliminate OutliersNetwork StructureExtracted Features
JigsawNet [19]Self-constructed✓✘CNNTexture
Papyrus [31]Self-constructed✘✘Siamese CNNTexture
S3-Net [51]GAN-augmented✘✘Siamese CNNContour
LLMCO4MRSynthetic Shredding✓✓CO-MLLMMultimodal
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.788, + 0.708 + ], + "angle": 0, + "content": "language processing [38], medical application [35,36], law judgement [33,50], and optimization [15, 16]. Recent advancements in Multimodal-LLM (MLLM) have catalyzed a proliferation of diverse applications, including image content understanding [56], mathematical reasoning [11], and even meme interpolation [49], among others. Inspired by MLLMs' potential in complex visual and textual content understanding, we proposed LLMCO4MR, the first CO solver and MLLM collaboration pipeline to tackle the complex ancient manuscript restoration task. We leverage MLLMs' potential extracting multimodal features [14,24] to tackle small ancient fragments pairwise matching, while further enhancing CO solver performance figuring similar fragments in real-world cases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Discussion with Other Restoration Pipelines. We provide a qualitative comparison with other major jigsaw puzzle solving and fragment restoration pipelines, as shown in Tab. 1. The key distinctions between our LLMCO4MR pipeline and earlier approaches are highlighted in Fig. 3. Our pipeline adeptly addresses the combinatorial explosion challenge inherent in fragment selection through CO solver. Furthermore, it innovates by introducing the use of MLLMs for the pairwise matching of ancient manuscripts, an innovative approach for manuscript restoration with open-world setting." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.381, + 0.163 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.183, + 0.788, + 0.243 + ], + "angle": 0, + "content": "In this section, we provide a detailed description of our LLMCO4MR pipeline, as composed of two main components: a cardinality-constrained neural solver trained on synthetic data to select candidates for further matching; and the MLLM adoption to achieve effective matching in a zero-shot learning manner." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.271, + 0.441, + 0.286 + ], + "angle": 0, + "content": "3.1 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.362 + ], + "angle": 0, + "content": "We consider a set of \\(m\\) fragments, denoted as \\(\\mathcal{F} = \\{f_{in}^{1},\\dots ,f_{in}^{k},f_{out}^{k + 1},\\dots ,f_{out}^{m}\\}\\). Here the first \\(k\\) fragments \\(f_{in}^{i}\\) refer to those originating from a more complete fragment, while the rest \\((m - k)\\) fragments denotes those outliers. The goal of LLMCO4MR is to identify such \\(k\\) relevant fragments." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.39, + 0.542, + 0.405 + ], + "angle": 0, + "content": "3.2 CO Solver for Fragment Selection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.787, + 0.466 + ], + "angle": 0, + "content": "3.2.1 Preliminaries. Based on consultations with archaeology specialists and their insights gained from experimental endeavors, we have formulated the following hypotheses to guide the restoration:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.48, + 0.787, + 0.57 + ], + "angle": 0, + "content": "1. Internal Texture and Limited Text. Small real-world fragments typically contain very limited characters, and ancient text recognition is unreliable for matching due to low accuracy. Moreover, even fragments that should be joined together may exhibit different textures due to varying museum collection conditions. Therefore, distinguishing fragments by textual content or texture patterns poses a significant challenge." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.571, + 0.787, + 0.632 + ], + "angle": 0, + "content": "2. Edge Shapes and Historical Tearing. Small fragments were often deliberately torn by people in ancient times [12], resulting in contours that are not severely eroded. This key feature offers the potential for primary restoration through contour similarity." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.633, + 0.787, + 0.679 + ], + "angle": 0, + "content": "3. Reassembly Group Size. Fragments are typically reassembled in groups ranging from 3 to 9 pieces. It is uncommon to find a large assembly of pieces that can be combined into a significantly larger fragment." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.48, + 0.787, + 0.679 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.691, + 0.787, + 0.738 + ], + "angle": 0, + "content": "In this section, we propose a cardinality-constrained neural CO solver to efficiently and robustly select manuscript fragments, which is further exemplified with MLLM with application on the famous real-world Dunhuang Manuscript." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "3.2.2 Solver Overview. We formulate fragment selection task as cardinality-constrained optimization problem in Eq. (1a). Specifically, the number of nonnegative elements must be restricted to no greater than \\( k \\), while minimizing the objective \\( L_{co} \\). The aim of our formulation is to find the optimal fragments subset with a size of \\( k \\) with highest mutual similarity, as shown in Eq. (1b), where \\( D_{ij} \\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.784, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.404, + 0.788, + 0.502 + ], + "angle": 0, + "content": "Fig. 4: LLMCO4MR solving pipeline. 1. Fragment Selection with CO solver. Given a set of \\( m \\) fragments, CO solver first pairs and feeds them into a siamese network to obtain similarity. Next, construct similarity graph \\( \\mathcal{G} \\) and pass it through GNN. Then encode cardinality constraints with differentiate OT layer. Finally output the selected Top- \\( K \\) fragments with highest mutual similarity. 2. Pairwise matching with MLLM in zero-shot. Given the selected Top- \\( K \\) fragments by pair as visual prompt with text prompt, MLLMs generate matching score and relative directions as output." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.785, + 0.561 + ], + "angle": 0, + "content": "denotes the similarity score between candidates \\(i\\) and \\(j\\), which are predicted by siamese network as described in the following section." + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.571, + 0.786, + 0.594 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {S}} L _ {c o} \\quad \\text {s . t .} \\mathcal {S} \\in \\{0, 1 \\} ^ {m}, \\| \\mathcal {S} \\| _ {0} \\leq k, \\tag {1a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.595, + 0.787, + 0.644 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} L _ {c o} = - \\sum_ {i \\in \\{i | s _ {i} = 1 \\}} \\left(\\sum_ {j \\in \\{j | s _ {j} = 1, j \\neq i \\}} D _ {i j}\\right). \\tag {1b}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.653, + 0.788, + 0.745 + ], + "angle": 0, + "content": "This cardinality-constrained CO problem could be tackled by a differentiable OT layer [43] and Sinkhorn algorithm [9]. As shown in Fig. 4, we construct an end-to-end learning pipeline consisting of a siamese feature extraction module as problem encoder network, a graph neural network fused with an OT layer as the solver network. The neural network learns to solve Eq. (1a) under the unsupervised loss described in Eq. (3) as follows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "3.2.3 Siamese Feature Extraction Network. In this section, we introduce the siamese feature extraction network of CO solver. Given the scarcity of real-world ground truth data, as few archaeologists have successfully restored small manuscript fragments together, we first describe the synthetic shredding process for preparing training data." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.144, + 0.756, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.336, + 0.677, + 0.351 + ], + "angle": 0, + "content": "Fig. 5: Illustration of synthesized shredding process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.788, + 0.459 + ], + "angle": 0, + "content": "Synthetic Shredding Process. We draw an analogy between the intricate work of manuscript restoration and grid puzzle assembly. To simulate this, we generate synthetic shredded fragments by breaking larger manuscript pieces into smaller, puzzle-piece-like fragments. Our synthesis approach is conducted in two phases, as illustrated in Fig. 5:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.472, + 0.788, + 0.593 + ], + "angle": 0, + "content": "1. Contour Extraction. We begin by randomly selecting large, real-world manuscripts from which we extract binary boundaries, specifically focusing on the two horizontal and two vertical edges. We discard contours that are either too long or too short, retaining only those of an appropriate length to form our contour set \\(\\mathcal{C} = \\{C_1,C_2,\\ldots ,C_m\\}\\). Each contour in this set is represented by a series of two-dimensional coordinates \\(C_i = \\{(x_1,y_1),(x_2,y_2),\\dots ,(x_n,y_n)\\}\\). Through this method, we ensure the synthetic fragments are realistic in size and shape." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.593, + 0.788, + 0.67 + ], + "angle": 0, + "content": "2. Synthetic Shredding. We proceed to sample from the set \\(\\mathcal{C}\\) without replacement to create an \\(M\\times M\\) grid-like segmentation mask. Utilizing a real rectangular-shaped manuscript piece as our foundation, we apply this mask to segment it into \\(M^2\\) smaller, discrete fragments. This division simulates realistic fragment sizes and shapes encountered in manuscript restoration." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.472, + 0.788, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Siamese Network. We employ a siamese feature extraction network designed to assess contour similarity. We start with the original set of fragments with outliers \\(\\mathcal{F} = \\{f_{in}^{1},\\dots,f_{in}^{k},f_{out}^{k + 1},\\dots,f_{out}^{m}\\}\\). For each fragment, we create permutation combinations of pairwise matches \\((f^i,f^j)\\). These paired fragments, specifically their masked contour, are then fed into two identical branches of a CNN-based neural network. The network computes a similarity score \\(D_{ij} = \\text{Siamese}(f^i,f^j)\\) with the range being [0-1]. A score of 1 indicates that two fragments have matching contours that align perfectly, suggesting adjacency. Conversely, a score of 0 implies not adjacent. Eq. (2) describes the loss term, \\(w\\) means learnable parameters, \\(L_{\\text{siamese}}\\) means Binary Cross Entropy (BCE) Loss," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.703, + 0.162 + ], + "angle": 0, + "content": "\\(p\\) means predicted similarity score and \\(y\\) means ground truth label." + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.175, + 0.786, + 0.192 + ], + "angle": 0, + "content": "\\[\nL _ {s i a m e s e} = - w [ y \\cdot \\log p + (1 - y) \\cdot \\log (1 - p) ]. \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.214, + 0.784, + 0.245 + ], + "angle": 0, + "content": "3.2.4 Solver Network. Our CO solver network consists of a graph neural network fused with a differentiable OT layer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.25, + 0.785, + 0.356 + ], + "angle": 0, + "content": "Graph Neural Network. Following the siamese feature extraction network, we construct an undirected similarity graph \\(\\mathcal{G} = (\\mathcal{V},\\mathcal{E})\\). Graph vertices set \\(\\mathcal{V} = \\{v_{1},\\ldots ,v_{m}\\}\\) stores the extracted fragment mask feature vector. Given the pairwise similarity scores \\(D_{ij}\\) as edge weights, we add an edge \\(e_{ij}\\) to the edge set \\(\\mathcal{E}\\) if similarity score is greater than threshold. We then constructed a graph neural network to extract features from graph vertices and edges. We use Graph Convolution Network (GCN) as backbone." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.788, + 0.545 + ], + "angle": 0, + "content": "OT Layer and training. For Top- \\( K \\) fragments selection, we encode the cardinality constraint with differential OT layer following [41]. To encode cardinality constraints in a differentiable manner, we first formulate the constraints as marginal distributions in optimal transport (OT), and then solve OT with Sinkhorn algorithm [9] by iterative row-wise and column-wise normalization. Given \\( m \\) GNN output score \\( p_i \\) indicating the probability of selecting each piece, OT layer moves Top- \\( K \\) probability to \\( p_{max} \\) while moves others to \\( p_{min} \\). We select \\( K \\) nodes with the highest probability to predict the output and optimize neural network with loss function in Eq. (3). Loss function consists of three components: \\( L_{siamese} \\) penalize incorrect prediction in pairwise similarity; \\( L_{gnn} \\) denotes the BCE loss for node selection; and \\( L_{co} \\) indicates the optimization objective in the previously mentioned cardinality constrained CO." + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.56, + 0.786, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\min L = \\min \\left(\\alpha L _ {\\text {s i a m e s e}} + \\beta L _ {\\text {g n n}} + L _ {\\text {c o}}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.842 + ], + "angle": 0, + "content": "3.2.5 Second-order Matching As detailed in Section 3.1, fragment groups ranging from 3 to 9 pieces can be reconstructed into a final form. We explore different scenarios of grid-like restoration. In the case of a \\(2 \\times 2\\) grid, each ground truth fragment is expected to have two pairwise connections with its neighboring fragments, treating all four fragments as equally important. In contrast, within a \\(3 \\times 3\\) puzzle grid, as illustrated in Fig. 6, fragments positioned at [2, 4, 6, and 8] are observed to share three connections each, highlighting the central fragment [5] as most crucial with up to four connections. Meanwhile, the corner fragments [1, 3, 7, 9] are limited to just two connections. Acknowledging the imperfection of siamese network accuracy and the existence of false positive pairwise matches in real-world scenarios, we implement a second-order matching strategy. We start with identifying the Top-5 central pieces that with a higher number of connections. Subsequently, we refine the graph \\(\\mathcal{G}\\) prune by pruning edges, retaining only those connected to the previously selected Top-5 fragments and eliminating other distracting edges. This strategy is designed to mitigate the impact of inaccuracies and ensure a more reliable restoration process." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.145, + 0.432, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.327, + 0.461, + 0.34 + ], + "angle": 0, + "content": "(a) Select with Equivalent Importance" + }, + { + "type": "image", + "bbox": [ + 0.472, + 0.143, + 0.73, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.326, + 0.671, + 0.339 + ], + "angle": 0, + "content": "(b) Second-order Matching" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.35, + 0.788, + 0.407 + ], + "angle": 0, + "content": "Fig. 6: Illustration of second-order matching. (a) selects all Top-9 fragments with equivalent importance (dashed lines represent incorrect pairings obtained by siamese network). (b) highlights centered 5 fragments with more connections, then select the rest 4 fragments based on pruned graph \\(\\mathcal{G}_{prune}\\) (most incorrect pairings are eliminated)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.436, + 0.676, + 0.452 + ], + "angle": 0, + "content": "3.3 LLMCO4MR Pipeline for Manuscript Restoration" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.785, + 0.522 + ], + "angle": 0, + "content": "Traditional deep learning methods for manuscript restoration heavily depend on human intervention to filter out numerous incorrect pairings generated by models. This task becomes increasingly challenging with smaller fragments, which are abundant and tiny, making manual review burdensome." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.644 + ], + "angle": 0, + "content": "The advent of MLLMs offers high-quality, efficient solutions to tasks that were previously labor-intensive. Innovatively, we are among the first to harness MLLMs for manuscript restoration. MLLMs like GPT-4V have shown remarkable capabilities akin to those of archaeologists by incorporating multimodal semantics for precise pairwise matching, excelling in identifying local pairwise connections with scoring and directional guidance. However, MLLMs struggle with the vast possibilities in open-world settings, particularly with outliers, limiting their ability to perform restoration independently." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Addressing these challenges, we present the LLMCO4MR pipeline, a pioneering MLLM-aided Neural Combinatorial Optimization solver tailored for the restoration of small manuscript fragments. LLMCO4MR leverages a CO solver for initial fragment selection, effectively countering the combinatorial explosion by sidelining outliers. Subsequently, the Top-\\(K\\) selected pairs undergo analysis via MLLMs, which utilize custom prompts based on a role-play strategy [45] to refine matching precision. If MLLMs identify any Top-\\(K\\) selected fragments as having no viable pairing, these anomalies are removed from the CO solver's input, suggesting a reevaluation of the Top-\\(K\\) fragment selections. This iterative refinement enhances LLMCO4MR efficacy. Experimental results in Section 4.3, demonstrate how this synergistic approach not only improves the precision in selecting fragments but also significantly enhances the accuracy and efficiency of manuscript restoration in practical scenarios." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.644, + 0.163 + ], + "angle": 0, + "content": "4 Quantitative Evaluation and Case Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.299 + ], + "angle": 0, + "content": "Our experiments encompasses both synthetic data and real-world manuscript fragments. For fragment selection, we conducted simulation experiments using the CO solver, to demonstrate its capability in selecting appropriate fragment candidates for subsequent pairwise matching. We then deployed our LLMCO4MR pipeline on real-world scenarios. Specifically, we conducted case studies on the famous Dunhuang manuscripts, spanning from the 4th to the 14th century. These manuscripts are highly valued by archaeologists and the global community [12], due to their extensive volume, diverse content, and broad historical span." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.788, + 0.375 + ], + "angle": 0, + "content": "The combination of simulation experiments and real-world case studies provides a controlled and ideal environment for the quantitative evaluation of our method, aligning with the protocols of peer studies [30, 31, 51, 54]. These experiments were performed on a server equipped with Intel i7 CPUs at \\(3.2\\mathrm{GHz}\\) and NVIDIA GeForce RTX 3090 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.398, + 0.414, + 0.415 + ], + "angle": 0, + "content": "4.1 Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.568 + ], + "angle": 0, + "content": "Data. For fragment selection, simulation experiments were employed to train the CO solver using synthetic shredded data, as elaborated in Section 3.2.3. Our methodology ensured that fragments generated from actual manuscripts, including their contours, accurately reflect the complexities found in real-world restoration conditions. For the case studies involving LLMCO4MR pipeline, high-quality (with annotated ground truth) real-world Dunhuang fragments are indeed scarce. Our annotated fragment cases were provided by the expert team dedicated to Dunhuang culture research, in the form of visual images for evaluation. All images were obtained from the International Dunhuang Project (IDP)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.573, + 0.788, + 0.709 + ], + "angle": 0, + "content": "Evaluation. For fragment selection with CO solver, we evaluated on synthetic shredded data with accuracy criterion. For instance, in a scenario where the task is to select Top-4 pieces from 10 candidates, fragment correctly identified as one of the 4 ground truth fragments is deemed a successful selection. For the real-world case study on Dunhuang, we evaluate the performance by measuring the success rate of fragment restoration. A successful restoration is defined as a match that aligns with the ground truth candidate pairings provided by archaeologists. This expert-annotated data serves as our benchmark for assessing the accuracy of our method in real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.732, + 0.661, + 0.748 + ], + "angle": 0, + "content": "4.2 Simulation Experiment on Combinatorial Solver" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Training Details. Aligned with Section 3.1, CO solver is designed to select \\(2 \\times 2\\) and \\(3 \\times 3\\) fragments from a set \\(\\mathcal{F}\\) comprising \\(m\\) fragments, aiming to address the Top-\\(K\\) fragment selection challenge. Thus we initially selected \\(k\\) pieces that originate from the same fragment and then added \\((m - k)\\) random fragments as outliers to compile a group. We assembled 500 such groups as" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.262, + 0.145, + 0.74, + 0.159 + ], + "angle": 0, + "content": "Table 2: Performance of CO solver siamese feature extraction network." + }, + { + "type": "table", + "bbox": [ + 0.347, + 0.171, + 0.654, + 0.354 + ], + "angle": 0, + "content": "
Data QuantityHybrid RatioSiamese AccuracyTop-4 Performance
400001:10.87460.6925
2:30.92120.7025
1:20.91170.7100
800001:10.95340.7250
2:30.94370.7325
1:20.96750.7500
1200001:10.92450.7525
2:30.95860.7575
1:20.97310.7800
" + }, + { + "type": "table_caption", + "bbox": [ + 0.236, + 0.37, + 0.766, + 0.385 + ], + "angle": 0, + "content": "Table 3: Performance of CO solver on various Top- \\( K \\) fragments selection task." + }, + { + "type": "table", + "bbox": [ + 0.264, + 0.397, + 0.735, + 0.49 + ], + "angle": 0, + "content": "
Top-K2×23×3
Pool Size101520202530
Random Select0.40000.26670.20000.45000.36000.3000
Baseline0.74750.31250.25750.67560.57670.5389
OT Layer0.78000.58250.42250.72560.64440.6044
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.785, + 0.569 + ], + "angle": 0, + "content": "training dataset, reserving an additional 100 groups for evaluation. We trained the siamese feature extraction network with 40K, 80K and 120K positive and negative pairing samples, with the ratio of training and testing being 6:1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.788, + 0.697 + ], + "angle": 0, + "content": "Siamese Feature Extraction Network. As detailed in Table 2, we evaluate the performance across various training data quantity, positive and negative sample ratios utilizing a ResNet backbone. Our findings indicate that CO solver performance peaks on Top-4 selections with 120K samples and 1:2 hybrid ratio on positive and negative matching samples. This can be attributed to the imbalance between positive and negative samples in pairwise matching manuscript fragments. With more negative samples, the reliability of siamese model is improved and CO solver performance is enhanced." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "CO Solver. Tab. 3 presents the simulation experiment results under various configurations of Top- \\(K\\) fragment selection from \\(m\\) candidates. \"Random select\" refers to the Top- \\(K\\) random selections mathematical expectation. \"Baseline\" refers to selecting the Top- \\(K\\) probabilities in a greedy manner from GNN output without OT layer. The introduction of the OT layer, which enforces cardinality constraints, is observed to enhance model performance, particularly as the candidate size \\(m\\) increases. This improvement is attributed to the OT layer's rigorous penalization of deviations from the optimization objectives in Eqs. (1a) and (1b)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.289, + 0.145, + 0.713, + 0.159 + ], + "angle": 0, + "content": "Table 4: Enhancement from the second-order Top- \\( K \\) selection." + }, + { + "type": "table", + "bbox": [ + 0.31, + 0.171, + 0.689, + 0.23 + ], + "angle": 0, + "content": "
StrategyTop-9 of 20Top-9 of 25Top-9 of 30
Select at Once0.72560.64440.6044
Second-Order0.73110.68220.6633
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.261, + 0.787, + 0.397 + ], + "angle": 0, + "content": "Enhancement through Second-order Matching Strategy. In the \\(2 \\times 2\\) selection settings, all four ground truth fragments hold equal significance (each fragment only have exactly 2 matching connections with their ground-truth adjacent pieces). Thus we explore the efficacy of the second-order matching strategy introduced in Section 3.2.5 for \\(3 \\times 3\\) scenarios and detail the results in Tab. 4. The second-order matching strategy duplicates the cardinality constraint and applies more stringent penalties for incorrect selections. This method aims to mitigate the effects of erroneous pairwise matches proposed by the siamese network, thereby advancing the accuracy and reliability of fragment selection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.787, + 0.443 + ], + "angle": 0, + "content": "Simulation experiments demonstrate our proposed neural CO solver's effectiveness in selecting candidates based on highest mutual assemble similarity. This gives promise for the subsequent real-world restoration with LLMCO4MR." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.47, + 0.681, + 0.485 + ], + "angle": 0, + "content": "4.3 Real-world Manuscript Case Studies on Dunhuang" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.787, + 0.627 + ], + "angle": 0, + "content": "Results with LLMCO4MR. We conducted case study experiments utilizing LLMCO4MR and compared its effectiveness against other SOTAs, as detailed in Tab. 5. We input a group of fragments with outliers to CO solver for selection, then use the selected Top- \\( K \\) fragments as image prompts for GPT-4V, which suggests the matching scores and relative directions. For SOTAs [31,51] that lack the capability to filter out outliers, we repeatedly input the fragments pairs and select the matching results in a greedy manner (select the fragment candidate pairs with ranked highest matching score)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.787, + 0.734 + ], + "angle": 0, + "content": "We observed that previous works often operated under restricted conditions, such as assuming no outliers, limiting search spaces, regular shape fragments, or using manually outlined perfect matching contours. These constraints do not reflect the open-world setting addressed by our method. Moreover, contour similarity network does not achieve perfect accuracy in matching judgments and real-world scenarios frequently involve more irregular fragments, underscoring the complexity of the task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Our pipeline addresses these challenges in two ways. First, the CO solver captures irregularity through data augmentation. We extract contours from real-world Dunhuang manuscripts and apply to shredding to adequately represent real-world challenges and mitigate the risk of overfitting unrealistic characteristics in training. Second, the MLLM component captures cross-image similarity using multimodal features, including contour (even with some imperfect fits and irregular shapes) and other visual features." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.235, + 0.145, + 0.768, + 0.16 + ], + "angle": 0, + "content": "Table 5: Evaluation on real-world Dunhuang fragment restoration case studies." + }, + { + "type": "table", + "bbox": [ + 0.286, + 0.171, + 0.714, + 0.334 + ], + "angle": 0, + "content": "
Top-K2×23×3
Pool Size10152025
GPT-4V [26]0.32500.17500.37780.2533
LLaVA [21]0.31500.12500.27780.1556
JigsawNet [19]0.52500.37500.45560.3222
Papyrus [31]0.42500.37500.47780.4111
S3-Net [51]0.31250.25750.38890.2111
CO Solver0.57500.52500.53330.4667
LLMCO4MR (Ours)0.67500.62500.62220.5556
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.788, + 0.499 + ], + "angle": 0, + "content": "Ablation Study. We also present ablation study results in Tab. 5. \"CO Solver\" refers to deriving matching pairs directly from the output of CO solver. Few selected MLLMs exhibit capabilities in understanding cross-image fragment relations. These MLLMs' performances were evaluated by processing pairs of fragments without CO solver filtering out outliers. We conclude that similar to retrieval-augmented LLM frameworks, CO solver in our LLMCO4MR is a nontrivial extension, empowers MLLM capability handling combinatorial explosion problems. We detail in the Appendix the current limitations of MLLMs in directly addressing manuscript restoration tasks with combinatorial explosion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.788, + 0.577 + ], + "angle": 0, + "content": "Our multifaceted approach serves two purposes: it complements and enhances traditional contour-based matching, and it helps to stabilize the variable performance of MLLMs in image matching. As a result, our method demonstrates increased effectiveness in handling real-world manuscript restoration tasks, especially in scenarios where either matching strategy alone is insufficient." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.605, + 0.36, + 0.621 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.788, + 0.764 + ], + "angle": 0, + "content": "We have presented a novel pipeline, LLMCO4MR, designed for the intricate task of restoring ancient manuscripts from fragments. We first formalize the problem as a cardinality-constrained combinatorial problem and then a neural combinatorial solver is devised to produce the Top- \\( K \\) candidates for merging. Then the MLLM is employed, leveraging the candidates as prompt for final restoration. Experiments on both synthetic fragments and real-world Dunhuang data demonstrate the potential of our approach, compared with the traditional approaches which require labors of well-trained archaeologists." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "**Limitation:** Like existing works, our current approach also cannot handle the case with missing fragments and the fragments are required to be basically in a grid layout. Another unsolved case is the existence of both small and large fragments, which we leave for future work. Also, more case studies on broader scope need to be studied when such data becomes available." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.268 + ], + "angle": 0, + "content": "This work was in part supported by National Science and Technology Major Project (2022ZD0119100), National Natural Science Foundation of China (No. 62441605, 72342023), National Social Science Fund of China (No. 21BYY142, 14AZS001), Key Research and Development Program of Zhejiang Province (No. 2024C03270), the StarryNight Science Fund of Zhejiang University Shanghai Institute for Advanced Study (SN-ZJU-SIAS-0010)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.288, + 0.323, + 0.304 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.36 + ], + "angle": 0, + "content": "1. Abitbol, R., Shimshoni, I., Ben-Dov, J.: Machine learning based assembly of fragments of ancient papyrus. Journal on Computing and Cultural Heritage (JOCCH) 14(3), 1-21 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.36, + 0.785, + 0.414 + ], + "angle": 0, + "content": "2. Assael, Y., Sommerschield, T., Shillingford, B., Bordbar, M., Pavlopoulos, J., Chatzipanagiotou, M., Androutsopoulos, I., Prag, J., de Freitas, N.: Restoring and attributing ancient texts using deep neural networks. Nature 603(7900), 280-283 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.415, + 0.785, + 0.497 + ], + "angle": 0, + "content": "3. Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., Hui, B., Ji, L., Li, M., Lin, J., Lin, R., Liu, D., Liu, G., Lu, C., Lu, K., Ma, J., Men, R., Ren, X., Ren, X., Tan, C., Tan, S., Tu, J., Wang, P., Wang, S., Wang, W., Wu, S., Xu, B., Xu, J., Yang, A., Yang, H., Yang, J., Yang, S., Yao, Y., Yu, B., Yuan, H., Yuan, Z., Zhang, J., Zhang, X., Zhang, Y., Zhang, Z., Zhou, C., Zhou, J., Zhou, X., Zhu, T.: Qwen technical report (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.498, + 0.785, + 0.538 + ], + "angle": 0, + "content": "4. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.539, + 0.785, + 0.58 + ], + "angle": 0, + "content": "5. Bengio, Y., Lodi, A., Prouvost, A.: Machine learning for combinatorial optimization: a methodological tour d'horizon. European Journal of Operational Research 290(2), 405-421 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.785, + 0.607 + ], + "angle": 0, + "content": "6. Besl, P.J., McKay, N.D.: Method for registration of 3-d shapes. In: Sensor fusion IV: control paradigms and data structures. vol. 1611, pp. 586-606. Spie (1992)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.608, + 0.785, + 0.648 + ], + "angle": 0, + "content": "7. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.649, + 0.785, + 0.689 + ], + "angle": 0, + "content": "8. Buchbinder, N., Feldman, M., Naor, J., Schwartz, R.: Submodular maximization with cardinality constraints. In: Proceedings of the twenty-fifth annual ACM-SIAM symposium on Discrete algorithms. pp. 1433-1452. SIAM (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.69, + 0.785, + 0.717 + ], + "angle": 0, + "content": "9. Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems 26 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.744 + ], + "angle": 0, + "content": "0. Derech, N., Tal, A., Shimshoni, I.: Solving archaeological puzzles. Pattern Recognition 119, 108065 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.745, + 0.785, + 0.812 + ], + "angle": 0, + "content": "11. Driess, D., Xia, F., Sajjadi, M.S.M., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., Huang, W., Chebotar, Y., Sermanet, P., Duckworth, D., Levine, S., Vanhoucke, V., Hausman, K., Toussaint, M., Greff, K., Zeng, A., Mordatch, I., Florence, P.: Palm-e: An embodied multimodal language model (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "12. Galambos, I.: Dunhuang Manuscript Culture: End of the First Millennium, vol. 22. Walter de Gruyter GmbH & Co KG (2020)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.318, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "13. da Gama Leitao, H.C., Stolfi, J.: A multiscale method for the reassembly of two-dimensional fragmented objects. IEEE Transactions on Pattern Analysis and Machine Intelligence 24(9), 1239-1251 (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "14. Ge, J., Luo, H., Qian, S., Gan, Y., Fu, J., Zhang, S.: Chain of thought prompt tuning in vision language models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.785, + 0.243 + ], + "angle": 0, + "content": "15. Guo, P.F., Chen, Y.H., Tsai, Y.D., Lin, S.D.: Towards optimizing with large language models. arXiv preprint arXiv:2310.05204 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.785, + 0.284 + ], + "angle": 0, + "content": "16. Guo, Q., Wang, R., Guo, J., Li, B., Song, K., Tan, X., Liu, G., Bian, J., Yang, Y.: Connecting large language models with evolutionary algorithms yields powerful prompt optimizers. arXiv preprint arXiv:2309.08532 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.284, + 0.785, + 0.325 + ], + "angle": 0, + "content": "17. Hossieni, S.S., Shabani, M.A., Irandoust, S., Furukawa, Y.: Puzzlefusion: Unleashing the power of diffusion models for spatial puzzle solving. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.325, + 0.785, + 0.366 + ], + "angle": 0, + "content": "18. Karalias, N., Loukas, A.: Erdos goes neural: an unsupervised learning framework for combinatorial optimization on graphs. Advances in Neural Information Processing Systems 33, 6659-6672 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.366, + 0.785, + 0.407 + ], + "angle": 0, + "content": "19. Le, C., Li, X.: Jigsawnet: Shredded image reassembly using convolutional neural network and loop-based composition. IEEE Transactions on Image Processing 28(8), 4000-4015 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.688, + 0.42 + ], + "angle": 0, + "content": "20. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.447 + ], + "angle": 0, + "content": "21. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.473 + ], + "angle": 0, + "content": "22. Lu, J., Sun, Y., Huang, Q.: Jigsaw: Learning to assemble multiple fractured objects. arXiv preprint arXiv:2305.17975 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "23. Lyu, C., Wu, M., Wang, L., Huang, X., Liu, B., Du, Z., Shi, S., Tu, Z.: Macaw-lm: Multi-modal language modeling with image, audio, video, and text integration (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.542 + ], + "angle": 0, + "content": "24. Maaz, M., Rasheed, H., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.569 + ], + "angle": 0, + "content": "25. Markaki, S., Panagiotakis, C.: Jigsaw puzzle solving techniques and applications: a survey. The Visual Computer pp. 1-17 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.503, + 0.582 + ], + "angle": 0, + "content": "26. OpenAI: Gpt-4 technical report (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.785, + 0.609 + ], + "angle": 0, + "content": "27. Panagiotakis, C., Markaki, S., Kokinou, E., Papadakis, H.: Coastline matching via a graph-based approach. Computational Geosciences 26(6), 1439-1448 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.785, + 0.65 + ], + "angle": 0, + "content": "28. Paumard, M.M., Picard, D., Tabia, H.: Jigsaw puzzle solving using local feature co-occurrences in deep neural networks. In: 2018 25th IEEE international conference on image processing (ICIP). pp. 1018-1022. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.691 + ], + "angle": 0, + "content": "29. Pengcheng, G., Gang, G., Jiangqin, W., Baogang, W.: Chinese calligraphic style representation for recognition. International Journal on Document Analysis and Recognition (IJDAR) 20, 59-68 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.732 + ], + "angle": 0, + "content": "30. Pirrone, A., Aimar, M.B., Journet, N.: Papy-s-net: A siamese network to match papyrus fragments. In: Proceedings of the 5th International Workshop on Historical Document Imaging and Processing. pp. 78-83 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.772 + ], + "angle": 0, + "content": "31. Pirrone, A., Beurton-Aimar, M., Journet, N.: Self-supervised deep metric learning for ancient papyrus fragments retrieval. International Journal on Document Analysis and Recognition (IJDAR) 24(3), 219-234 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "32. Rusinkiewicz, S., Levoy, M.: Efficient variants of the icp algorithm. In: Proceedings third international conference on 3-D digital imaging and modeling. pp. 145-152. IEEE (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.841 + ], + "angle": 0, + "content": "33. Savelka, J., Ashley, K.D., Gray, M.A., Westermann, H., Xu, H.: Explaining legal concepts with augmented large language models (gpt-4) (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "34. Savino, P., Tonazzini, A.: Digital restoration of ancient color manuscripts from geometrically misaligned recto-verso pairs. Journal of Cultural Heritage 19, 511-521 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "35. Shuai, R.W., Ruffolo, J.A., Gray, J.J.: Generative language modeling for antibody design. bioRxiv (2022). https://doi.org/10.1101/2021.12.13.472419, https://www.biorxiv.org/content/early/2022/12/20/2021.12.13.472419" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.233, + 0.788, + 0.316 + ], + "angle": 0, + "content": "36. Singhal, K., Azizi, S., Tu, T., Mahdavi, S.S., Wei, J., Chung, H.W., Scales, N., Tanwani, A., Cole-Lewis, H., Pfohl, S., Payne, P., Seneviratne, M., Gamble, P., Kelly, C., Scharli, N., Chowdhery, A., Mansfield, P., y Arcas, B.A., Webster, D., Corrado, G.S., Matias, Y., Chou, K., Gottweis, J., Tomasev, N., Liu, Y., Rajkomar, A., Barral, J., Semturs, C., Karthikesalingam, A., Natarajan, V.: Large language models encode clinical knowledge (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.317, + 0.788, + 0.358 + ], + "angle": 0, + "content": "37. Sommerscheid, T., Assael, Y., Pavlopoulos, J., Stefanak, V., Senior, A., Dyer, C., Bodel, J., Prag, J., Androutsopoulos, I., de Freitas, N.: Machine learning for ancient languages: A survey. Computational Linguistics pp. 1-44 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.36, + 0.788, + 0.401 + ], + "angle": 0, + "content": "38. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.402, + 0.788, + 0.431 + ], + "angle": 0, + "content": "39. Tsamoura, E., Pitas, I.: Automatic color based reassembly of fragmented images and paintings. IEEE Transactions on Image Processing 19(3), 680-690 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.459 + ], + "angle": 0, + "content": "40. Vinyals, O., Fortunato, M., Jaitly, N.: Pointer networks. Advances in neural information processing systems 28 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.514 + ], + "angle": 0, + "content": "41. Wang, R., Shen, L., Chen, Y., Yang, X., Tao, D., Yan, J.: Towards one-shot neural combinatorial solvers: Theoretical and empirical notes on the cardinality-constrained case. In: The Eleventh International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.557 + ], + "angle": 0, + "content": "42. Wang, R., Yan, J., Yang, X.: Learning combinatorial embedding networks for deep graph matching. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3056-3065 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.558, + 0.788, + 0.6 + ], + "angle": 0, + "content": "43. Wang, R., Zhang, Y., Guo, Z., Chen, T., Yang, X., Yan, J.: Linsatnet: The positive linear satisfiability neural networks. In: International Conference on Machine Learning (ICML) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.601, + 0.788, + 0.629 + ], + "angle": 0, + "content": "44. Wolfson, H., Schonberg, E., Kalvin, A., Lamdan, Y.: Solving jigsaw puzzles by computer. Annals of Operations Research 12(1), 51-64 (1988)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.671 + ], + "angle": 0, + "content": "45. Xi, Z., Chen, W., Guo, X., He, W., Ding, Y., Hong, B., Zhang, M., Wang, J., Jin, S., Zhou, E., et al.: The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.672, + 0.788, + 0.713 + ], + "angle": 0, + "content": "46. Xie, Y., Dai, H., Chen, M., Dai, B., Zhao, T., Zha, H., Wei, W., Pfister, T.: Differentiable top-k with optimal transport. Advances in Neural Information Processing Systems 33, 20520-20531 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.714, + 0.788, + 0.811 + ], + "angle": 0, + "content": "47. Yang, A., Xiao, B., Wang, B., Zhang, B., Bian, C., Yin, C., Lv, C., Pan, D., Wang, D., Yan, D., Yang, F., Deng, F., Wang, F., Liu, F., Ai, G., Dong, G., Zhao, H., Xu, H., Sun, H., Zhang, H., Liu, H., Ji, J., Xie, J., Dai, J., Fang, K., Su, L., Song, L., Liu, L., Ru, L., Ma, L., Wang, M., Liu, M., Lin, M., Nie, N., Guo, P., Sun, R., Zhang, T., Li, T., Li, T., Cheng, W., Chen, W., Zeng, X., Wang, X., Chen, X., Men, X., Yu, X., Pan, X., Shen, Y., Wang, Y., Li, Y., Jiang, Y., Gao, Y., Zhang, Y., Zhou, Z., Wu, Z.: Baichuan 2: Open large-scale language models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.812, + 0.788, + 0.841 + ], + "angle": 0, + "content": "48. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of lmmss: Preliminary explorations with gpt-4v(ison) (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "Y. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "49. Yang, Z., Li, L., Wang, J., Lin, K., Azarnasab, E., Ahmed, F., Liu, Z., Liu, C., Zeng, M., Wang, L.: Mm-react: Prompting chatgpt for multimodal reasoning and action (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "50. Yu, F., Quartey, L., Schilder, F.: Legal prompting: Teaching a language model to think like a lawyer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.272 + ], + "angle": 0, + "content": "51. Zhang, C., Wang, B., Chen, K., Zong, R., Mo, B.f., Men, Y., Almpanidis, G., Chen, S., Zhang, X.: Data-driven oracle bone rejoining: A dataset and practical self-supervised learning scheme. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 4482-4492 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.327 + ], + "angle": 0, + "content": "52. Zhang, C., Zong, R., Cao, S., Men, Y., Mo, B.: Ai-powered oracle bone inscriptions recognition and fragments rejoining. In: Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence. pp. 5309-5311 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "53. Zhang, K., Li, X.: A graph-based optimization algorithm for fragmented image reassembly. Graphical Models 76(5), 484-495 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.41 + ], + "angle": 0, + "content": "54. Zhang, Y., Fang, Z., Yang, X., Zhang, S., He, B., Dou, H., Yan, J., Zhang, Y., Wu, F.: Reconnecting the broken civilization: Patchwork integration of fragments from ancient manuscripts. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 1157-1166 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.452 + ], + "angle": 0, + "content": "55. Zhang, Z., Wang, Y.T., Li, B., Guo, A., Liu, C.L.: Deep rejoining model for oracle bone fragment image. In: Asian Conference on Pattern Recognition. pp. 3-15. Springer (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.481 + ], + "angle": 0, + "content": "56. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.481 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/LLMCO4MR_ LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang/33e2dc3a-2be0-4119-88f6-223087048685_origin.pdf b/2024/LLMCO4MR_ LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang/33e2dc3a-2be0-4119-88f6-223087048685_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..59eb8a38b1d28300a8cde4aaef9984dc2baadc9f --- /dev/null +++ b/2024/LLMCO4MR_ LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang/33e2dc3a-2be0-4119-88f6-223087048685_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:841a134ea040ca9d34aa1f14bf03c97f619f9ef62c896193cb7bbb4f8718dec6 +size 3574080 diff --git a/2024/LLMCO4MR_ LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang/full.md b/2024/LLMCO4MR_ LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e4fe72b42471f0e0dc191d83f776550e7b29e7c2 --- /dev/null +++ b/2024/LLMCO4MR_ LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang/full.md @@ -0,0 +1,284 @@ +# LLMCO4MR: LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang + +Yuqing Zhang $^{1}$ , Hangqi Li $^{1}$ , Shengyu Zhang $^{1,2}$ , Runzhong Wang $^{3}$ , Baoyi He $^{1}$ , Huaiyong Dou $^{1}$ , Junchi Yan $^{3,4}$ , Yongquan Zhang $^{1}$ , and Fei Wu $^{1,2,4}$ + +1 Zhejiang University, Hangzhou, China + +$^{2}$ Shanghai Institute for Advanced Study of Zhejiang University, Shanghai, China + +3 Shanghai Jiao Tong University, Shanghai, China + +4 Shanghai AI Laboratory, Shanghai, China + +{yuqingz7, sy_zhang, wufei}@zju.edu.cn, yanjunchi@sjtu.edu.cn + +Abstract. Restoring ancient manuscripts fragments, such as those from Dunhuang, is crucial for preserving human historical culture. However, their worldwide dispersal and the shifts in cultural and historical contexts pose significant restoration challenges. Traditional archaeological efforts primarily focus on manually piecing major fragments together, yet the small and more intricate pieces remain largely unexplored, which is technically due to their irregular shapes, sparse textual content, and extensive combinatorial space for reassembly. In this paper, we formalize the task of restoring the ancient manuscript from fragments as a cardinality-constrained combinatorial optimization problem, and propose a framework named LLMCO4MR: (Multimodal) Large Language Model-aided Combinatorial Optimization Neural Networks for Ancient Manuscript Restoration. Specifically, LLMCO4MR encapsulates a neural combinatorial solver equipped with a differentiable optimal transport (OT) layer, to efficiently predict the Top- $K$ likely mutual reassembly candidates. Multimodal Large Language Model (MLLM) is then adopted and prompted to yield pairwise matching confidence and relative directions for final restoration. Experiments on synthetic data and cases studies with real-world famous Dunhuang fragments demonstrate our approach's practical potential in assisting archaeologists. Our method provides a novel perspective for ancient manuscript restoration. + +Keywords: Ancient Manuscript Restoration $\cdot$ Dunhuang Manuscript $\cdot$ + +Combinatorial Optimization $\cdot$ Multimodal Large Language Models + +# 1 Introduction + +The restoration of ancient manuscripts has long been the focus of archaeologists' research and the global community. Traditionally, these codices are frequently + +![](images/13fed47de562bd0df09c0cd2641f6ae7246e74da46965579ee1a5e721fbd6b22.jpg) +(a) Conventional manuscript restoration task + +![](images/8d13fbb16a18f7860c294a586be2a40f57fe501885504e80200f7cc2b089fe56.jpg) +(b) Our task: Small fragments restoration with outliers +Fig. 1: Comparison of conventional manuscript restoration and our task. (a) highlights the conventional restoration task with large fragments. Archaeologists focus on written content as primary guide for restoration, supplemented by texture and edges. (b) outlines our task. In an open-world problem setting, small fragments contain limited text. Recovery relies on contour similarity matching. This is generally difficult for humans to identify fragment pairs from a large collection containing outliers. + +discovered in a (severely) deteriorated condition due to the passage of time. Recovering these manuscripts involves extensive efforts [37] that are often time-consuming, labor-intensive, and require a high level of expertise, necessitating years of professional training [2]. + +On the other hand, vision technologies have demonstrated significant potential in aiding the restoration of ancient manuscripts. For example, [1] developed a sophisticated system based on Convolutional Neural Network (CNN) to identify matching candidates among papyrus fragments by analyzing their fiber patterns. Zhang et al. [51] proposed a self-supervised model using Generative Adversarial Network (GAN) and siamese network for the pairwise stitching of Oracle Bone pieces, utilizing manually annotated contours for similarity comparison. While these vision-assisted pipelines enhance efficiency and accuracy beyond traditional manual efforts, their applications are generally confined to local pairwise matching [31], or regularly-shaped (e.g., rectangle) and large fragments [54]. A major challenge persists in dealing with real-world small fragments, which are often less-textual, shape-irregular, and scattered with missing fragments. + +These small fragments have often been overlooked by archaeologists, yet they form a substantial part of the manuscript collections. Recent explorations have revealed that these minor pieces are treasure troves of first-hand materials, offering unique insights into the daily lives of past civilizations [12]. Current deep learning approaches rely on textual [29], contextual [2] or multimodal information [54] to reassembly manuscripts. Such learning-based methods require a large amount of annotated training data for complex multimodal feature extraction, which is impractical for small fragment restoration. + +To bridge the gap, we resort to the recent advancements in large language models (LLMs) that have revolutionized traditional tasks in computer vision and natural language processing [7, 26, 38]. Notably, the emergence of Multimodal Large Language Models (MLLMs) has demonstrated exceptional capabilities in extracting and interpreting image features, understanding content, and following complex prompt instructions [4, 20, 48]. Capitalizing on these developments, our + +![](images/0a7e97e324120a236faeeaf7f01e2fbf5106846592f1c96c4fbc5dc4f86f6b3f.jpg) +Fig. 2: Showcases of matching results with case studies on Dunhuang fragments. + +research introduces a novel approach by employing MLLMs to tackle the intricate task of restoring ancient manuscript fragments. MLLMs facilitate efficient pairwise matching by: 1) identifying cross-image connections based on contour similarity and capturing complex features similarity in paper texture and content; 2) replicating the comprehensive integration of multimodal semantics that archaeologists employ during their assembly process; 3) demonstrating zero-shot learning capabilities, which are crucial for investigating manuscripts with limited ground truth availability. + +However, the complexity of our task surpasses the current capabilities of MLLMs. In practical scenarios of ancient manuscript restoration, it is necessary to first identify potential candidate fragments from a larger collection for subsequent pairwise matching. This collection contains fragments that may not belong to the target manuscript, which we refer to as "outliers". Performing pairwise comparisons using MLLMs on the vast number of small fragments, including these outliers, leads to a combinatorial explosion challenge, which poses a significant difficulty for MLLM itself to manage effectively. + +To address these challenges, we formulate the fragment selection subtask as a combinatorial optimization (CO) problem. We introduce a novel framework, named LLMCO4MR, a LLMs-aided Neural Combinatorial Optimization solver for manuscript restoration, transforming the unstructured challenge into a technically feasible task. The highlights of the paper are: + +1. More Open-world Problem Setting. Beyond the restricted setting requiring that there are no outlier fragments for recovery, to our best knowledge, we are the first to address the more realistic and challenging setting, i.e., restoring the manuscript from a subset of fragments with outliers. +2. New Two-stage Solving Pipeline. We solve the new problem by two stages: fragment selection and pairwise matching. In the first stage, $k$ fragments with the highest mutual assemble scores are selected by certain means while the latter stage involves the traditional closed-world problem setting. +3. Neural Combinatorial Solver for Candidate Selection. We introduce a combinatorial network to solve the fragment selection problem by formulating it as a $K$ -cardinality constrained task, with an advantage that the input visual fragments are learned end-to-end directly with the final decision goal. Supervised training is achieved by using the shredded fragments generated by splitting a complete manuscript into grids. + +4. MLLMs for Pairwise Fragment Matching in Zero-shot. The selected Top- $K$ fragments are then fed into the MLLM by pair, to generate the matching score as well as the relative direction of matching, i.e., horizontal or vertical. Then the final manuscript can be recovered by these local pairwise matchings. To our best knowledge, this is the first time to introduce MLLMs for manuscript recovery or more broadly speaking, jigsaw puzzle-like tasks. + +# 2 Related Works + +Manuscript Restoration. This area has recently witnessed significant advancements [2, 37] aided with vision technology for restoration. These developments aim to overcome the challenges in manual methods, which are time-consuming, labor-intensive, and heavily reliant on expert knowledge [2]. A typical method is to solve the problem from a jigsaw puzzle [10, 34] solving perspective. For puzzle solving and global restoration, several successful methods have emerged. Traditional strategies include curve-based techniques [27, 44] and time series sequence-based approaches [6, 13, 32, 39, 53]. Additionally, deep learning methods have been developed, exemplified by works such as [19] and other studies [17, 25, 28]. Inspired by these works, previous manuscript restoration methods are designed based on the visual features of the fragments. Here, matching often refers to identifying whether two fragments are adjacent and their relative direction (horizontally or vertically). In particular, the feature extraction module has been further advanced by deep learning with various aspects, including fragment contour shape [51, 52, 55], texture details [31], and written characters [29]. In the recent work [54], a multimodal restoration pipeline is devised that considers both written content and contour shape. + +Despite these advances, most current relic restoration methods share a common assumption: the candidate pieces for pairwise matching are given, either by manual pre-selection or other means. However, in many real-world restoration cases, this assumption leaves a gap in addressing the more challenging task of open-world problem setting, where candidate fragments are not predetermined, underscoring the novelty and necessity of our approach. + +Combinatorial Solver. The recent growing trend of learnable CO solvers [5] demonstrate the potential of solving CO problems by deep neural networks with higher efficiency [18, 22, 40, 42]. To tackle the extensive combinatorial space of ancient manuscript restoration, we present a cardinality-constrained CO reformulation [8] by selecting the Top- $K$ most likely matching fragments from the candidate set. Following the one-shot CO learning paradigm [41], we adopt [43], which handles cardinality-constrained problems similarly to differentiable optimal transport [46]. [43] has the advantage of controlling the tightness of constraints while preserving the gradient, which is crucial in our task. + +Large Language Models. With the recent arisen success of Large Language Model (LLM), such as GPT-4 [26], Qwen [3] and Baichuan [47], LLMs have showcased their efficacy in a myriad of fields including computer vision [23,24], natural + +![](images/8927eadc45cf92d434036fa4efd5f9da0435d52e26a7b3d39108f238fccf9283.jpg) +(a) Previous Ancient Manuscript Restoration Pipelines + +![](images/8037dde0be00c3712ea819cf26f79ccc10207daff02b0758ee916cffd08d0ac9.jpg) +(b) Our Two-stage LLMCO4MS Pipeline +Fig. 3: Comparison between previous restoration approaches and our LLMCO4MR. + +Table 1: Comparison with various ancient fragment restoration SOTA. + +
MethodologyTraining DataAnnotation FreeEliminate OutliersNetwork StructureExtracted Features
JigsawNet [19]Self-constructed✓✘CNNTexture
Papyrus [31]Self-constructed✘✘Siamese CNNTexture
S3-Net [51]GAN-augmented✘✘Siamese CNNContour
LLMCO4MRSynthetic Shredding✓✓CO-MLLMMultimodal
+ +language processing [38], medical application [35,36], law judgement [33,50], and optimization [15, 16]. Recent advancements in Multimodal-LLM (MLLM) have catalyzed a proliferation of diverse applications, including image content understanding [56], mathematical reasoning [11], and even meme interpolation [49], among others. Inspired by MLLMs' potential in complex visual and textual content understanding, we proposed LLMCO4MR, the first CO solver and MLLM collaboration pipeline to tackle the complex ancient manuscript restoration task. We leverage MLLMs' potential extracting multimodal features [14,24] to tackle small ancient fragments pairwise matching, while further enhancing CO solver performance figuring similar fragments in real-world cases. + +Discussion with Other Restoration Pipelines. We provide a qualitative comparison with other major jigsaw puzzle solving and fragment restoration pipelines, as shown in Tab. 1. The key distinctions between our LLMCO4MR pipeline and earlier approaches are highlighted in Fig. 3. Our pipeline adeptly addresses the combinatorial explosion challenge inherent in fragment selection through CO solver. Furthermore, it innovates by introducing the use of MLLMs for the pairwise matching of ancient manuscripts, an innovative approach for manuscript restoration with open-world setting. + +# 3 Methodology + +In this section, we provide a detailed description of our LLMCO4MR pipeline, as composed of two main components: a cardinality-constrained neural solver trained on synthetic data to select candidates for further matching; and the MLLM adoption to achieve effective matching in a zero-shot learning manner. + +# 3.1 Problem Formulation + +We consider a set of $m$ fragments, denoted as $\mathcal{F} = \{f_{in}^{1},\dots ,f_{in}^{k},f_{out}^{k + 1},\dots ,f_{out}^{m}\}$ . Here the first $k$ fragments $f_{in}^{i}$ refer to those originating from a more complete fragment, while the rest $(m - k)$ fragments denotes those outliers. The goal of LLMCO4MR is to identify such $k$ relevant fragments. + +# 3.2 CO Solver for Fragment Selection + +3.2.1 Preliminaries. Based on consultations with archaeology specialists and their insights gained from experimental endeavors, we have formulated the following hypotheses to guide the restoration: + +1. Internal Texture and Limited Text. Small real-world fragments typically contain very limited characters, and ancient text recognition is unreliable for matching due to low accuracy. Moreover, even fragments that should be joined together may exhibit different textures due to varying museum collection conditions. Therefore, distinguishing fragments by textual content or texture patterns poses a significant challenge. +2. Edge Shapes and Historical Tearing. Small fragments were often deliberately torn by people in ancient times [12], resulting in contours that are not severely eroded. This key feature offers the potential for primary restoration through contour similarity. +3. Reassembly Group Size. Fragments are typically reassembled in groups ranging from 3 to 9 pieces. It is uncommon to find a large assembly of pieces that can be combined into a significantly larger fragment. + +In this section, we propose a cardinality-constrained neural CO solver to efficiently and robustly select manuscript fragments, which is further exemplified with MLLM with application on the famous real-world Dunhuang Manuscript. + +3.2.2 Solver Overview. We formulate fragment selection task as cardinality-constrained optimization problem in Eq. (1a). Specifically, the number of nonnegative elements must be restricted to no greater than $k$ , while minimizing the objective $L_{co}$ . The aim of our formulation is to find the optimal fragments subset with a size of $k$ with highest mutual similarity, as shown in Eq. (1b), where $D_{ij}$ + +![](images/dd2512891180b48494ff65d0bd37c00f60336b806c5d3938433773ddd9308f27.jpg) +Fig. 4: LLMCO4MR solving pipeline. 1. Fragment Selection with CO solver. Given a set of $m$ fragments, CO solver first pairs and feeds them into a siamese network to obtain similarity. Next, construct similarity graph $\mathcal{G}$ and pass it through GNN. Then encode cardinality constraints with differentiate OT layer. Finally output the selected Top- $K$ fragments with highest mutual similarity. 2. Pairwise matching with MLLM in zero-shot. Given the selected Top- $K$ fragments by pair as visual prompt with text prompt, MLLMs generate matching score and relative directions as output. + +denotes the similarity score between candidates $i$ and $j$ , which are predicted by siamese network as described in the following section. + +$$ +\min _ {\mathcal {S}} L _ {c o} \quad \text {s . t .} \mathcal {S} \in \{0, 1 \} ^ {m}, \| \mathcal {S} \| _ {0} \leq k, \tag {1a} +$$ + +$$ +\text {w h e r e} L _ {c o} = - \sum_ {i \in \{i | s _ {i} = 1 \}} \left(\sum_ {j \in \{j | s _ {j} = 1, j \neq i \}} D _ {i j}\right). \tag {1b} +$$ + +This cardinality-constrained CO problem could be tackled by a differentiable OT layer [43] and Sinkhorn algorithm [9]. As shown in Fig. 4, we construct an end-to-end learning pipeline consisting of a siamese feature extraction module as problem encoder network, a graph neural network fused with an OT layer as the solver network. The neural network learns to solve Eq. (1a) under the unsupervised loss described in Eq. (3) as follows. + +3.2.3 Siamese Feature Extraction Network. In this section, we introduce the siamese feature extraction network of CO solver. Given the scarcity of real-world ground truth data, as few archaeologists have successfully restored small manuscript fragments together, we first describe the synthetic shredding process for preparing training data. + +![](images/a6a77b1041e75cc2b5bec9396bd7f197136b4ace57844f82f871ffa2d79e3d3b.jpg) +Fig. 5: Illustration of synthesized shredding process. + +Synthetic Shredding Process. We draw an analogy between the intricate work of manuscript restoration and grid puzzle assembly. To simulate this, we generate synthetic shredded fragments by breaking larger manuscript pieces into smaller, puzzle-piece-like fragments. Our synthesis approach is conducted in two phases, as illustrated in Fig. 5: + +1. Contour Extraction. We begin by randomly selecting large, real-world manuscripts from which we extract binary boundaries, specifically focusing on the two horizontal and two vertical edges. We discard contours that are either too long or too short, retaining only those of an appropriate length to form our contour set $\mathcal{C} = \{C_1,C_2,\ldots ,C_m\}$ . Each contour in this set is represented by a series of two-dimensional coordinates $C_i = \{(x_1,y_1),(x_2,y_2),\dots ,(x_n,y_n)\}$ . Through this method, we ensure the synthetic fragments are realistic in size and shape. +2. Synthetic Shredding. We proceed to sample from the set $\mathcal{C}$ without replacement to create an $M\times M$ grid-like segmentation mask. Utilizing a real rectangular-shaped manuscript piece as our foundation, we apply this mask to segment it into $M^2$ smaller, discrete fragments. This division simulates realistic fragment sizes and shapes encountered in manuscript restoration. + +Siamese Network. We employ a siamese feature extraction network designed to assess contour similarity. We start with the original set of fragments with outliers $\mathcal{F} = \{f_{in}^{1},\dots,f_{in}^{k},f_{out}^{k + 1},\dots,f_{out}^{m}\}$ . For each fragment, we create permutation combinations of pairwise matches $(f^i,f^j)$ . These paired fragments, specifically their masked contour, are then fed into two identical branches of a CNN-based neural network. The network computes a similarity score $D_{ij} = \text{Siamese}(f^i,f^j)$ with the range being [0-1]. A score of 1 indicates that two fragments have matching contours that align perfectly, suggesting adjacency. Conversely, a score of 0 implies not adjacent. Eq. (2) describes the loss term, $w$ means learnable parameters, $L_{\text{siamese}}$ means Binary Cross Entropy (BCE) Loss, + +$p$ means predicted similarity score and $y$ means ground truth label. + +$$ +L _ {s i a m e s e} = - w [ y \cdot \log p + (1 - y) \cdot \log (1 - p) ]. \tag {2} +$$ + +# 3.2.4 Solver Network. Our CO solver network consists of a graph neural network fused with a differentiable OT layer. + +Graph Neural Network. Following the siamese feature extraction network, we construct an undirected similarity graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$ . Graph vertices set $\mathcal{V} = \{v_{1},\ldots ,v_{m}\}$ stores the extracted fragment mask feature vector. Given the pairwise similarity scores $D_{ij}$ as edge weights, we add an edge $e_{ij}$ to the edge set $\mathcal{E}$ if similarity score is greater than threshold. We then constructed a graph neural network to extract features from graph vertices and edges. We use Graph Convolution Network (GCN) as backbone. + +OT Layer and training. For Top- $K$ fragments selection, we encode the cardinality constraint with differential OT layer following [41]. To encode cardinality constraints in a differentiable manner, we first formulate the constraints as marginal distributions in optimal transport (OT), and then solve OT with Sinkhorn algorithm [9] by iterative row-wise and column-wise normalization. Given $m$ GNN output score $p_i$ indicating the probability of selecting each piece, OT layer moves Top- $K$ probability to $p_{max}$ while moves others to $p_{min}$ . We select $K$ nodes with the highest probability to predict the output and optimize neural network with loss function in Eq. (3). Loss function consists of three components: $L_{siamese}$ penalize incorrect prediction in pairwise similarity; $L_{gnn}$ denotes the BCE loss for node selection; and $L_{co}$ indicates the optimization objective in the previously mentioned cardinality constrained CO. + +$$ +\min L = \min \left(\alpha L _ {\text {s i a m e s e}} + \beta L _ {\text {g n n}} + L _ {\text {c o}}\right). \tag {3} +$$ + +3.2.5 Second-order Matching As detailed in Section 3.1, fragment groups ranging from 3 to 9 pieces can be reconstructed into a final form. We explore different scenarios of grid-like restoration. In the case of a $2 \times 2$ grid, each ground truth fragment is expected to have two pairwise connections with its neighboring fragments, treating all four fragments as equally important. In contrast, within a $3 \times 3$ puzzle grid, as illustrated in Fig. 6, fragments positioned at [2, 4, 6, and 8] are observed to share three connections each, highlighting the central fragment [5] as most crucial with up to four connections. Meanwhile, the corner fragments [1, 3, 7, 9] are limited to just two connections. Acknowledging the imperfection of siamese network accuracy and the existence of false positive pairwise matches in real-world scenarios, we implement a second-order matching strategy. We start with identifying the Top-5 central pieces that with a higher number of connections. Subsequently, we refine the graph $\mathcal{G}$ prune by pruning edges, retaining only those connected to the previously selected Top-5 fragments and eliminating other distracting edges. This strategy is designed to mitigate the impact of inaccuracies and ensure a more reliable restoration process. + +![](images/93b3975e7be84aeea621c725f1d038c3c2b0ddb8e691959455596412b6178129.jpg) +(a) Select with Equivalent Importance +Fig. 6: Illustration of second-order matching. (a) selects all Top-9 fragments with equivalent importance (dashed lines represent incorrect pairings obtained by siamese network). (b) highlights centered 5 fragments with more connections, then select the rest 4 fragments based on pruned graph $\mathcal{G}_{prune}$ (most incorrect pairings are eliminated). + +![](images/0db4f04202d4d6949e8ad06bedcc8db9db77c21e1a2247f9e705914fe658b593.jpg) +(b) Second-order Matching + +# 3.3 LLMCO4MR Pipeline for Manuscript Restoration + +Traditional deep learning methods for manuscript restoration heavily depend on human intervention to filter out numerous incorrect pairings generated by models. This task becomes increasingly challenging with smaller fragments, which are abundant and tiny, making manual review burdensome. + +The advent of MLLMs offers high-quality, efficient solutions to tasks that were previously labor-intensive. Innovatively, we are among the first to harness MLLMs for manuscript restoration. MLLMs like GPT-4V have shown remarkable capabilities akin to those of archaeologists by incorporating multimodal semantics for precise pairwise matching, excelling in identifying local pairwise connections with scoring and directional guidance. However, MLLMs struggle with the vast possibilities in open-world settings, particularly with outliers, limiting their ability to perform restoration independently. + +Addressing these challenges, we present the LLMCO4MR pipeline, a pioneering MLLM-aided Neural Combinatorial Optimization solver tailored for the restoration of small manuscript fragments. LLMCO4MR leverages a CO solver for initial fragment selection, effectively countering the combinatorial explosion by sidelining outliers. Subsequently, the Top- $K$ selected pairs undergo analysis via MLLMs, which utilize custom prompts based on a role-play strategy [45] to refine matching precision. If MLLMs identify any Top- $K$ selected fragments as having no viable pairing, these anomalies are removed from the CO solver's input, suggesting a reevaluation of the Top- $K$ fragment selections. This iterative refinement enhances LLMCO4MR efficacy. Experimental results in Section 4.3, demonstrate how this synergistic approach not only improves the precision in selecting fragments but also significantly enhances the accuracy and efficiency of manuscript restoration in practical scenarios. + +# 4 Quantitative Evaluation and Case Study + +Our experiments encompasses both synthetic data and real-world manuscript fragments. For fragment selection, we conducted simulation experiments using the CO solver, to demonstrate its capability in selecting appropriate fragment candidates for subsequent pairwise matching. We then deployed our LLMCO4MR pipeline on real-world scenarios. Specifically, we conducted case studies on the famous Dunhuang manuscripts, spanning from the 4th to the 14th century. These manuscripts are highly valued by archaeologists and the global community [12], due to their extensive volume, diverse content, and broad historical span. + +The combination of simulation experiments and real-world case studies provides a controlled and ideal environment for the quantitative evaluation of our method, aligning with the protocols of peer studies [30, 31, 51, 54]. These experiments were performed on a server equipped with Intel i7 CPUs at $3.2\mathrm{GHz}$ and NVIDIA GeForce RTX 3090 GPUs. + +# 4.1 Experiment Setup + +Data. For fragment selection, simulation experiments were employed to train the CO solver using synthetic shredded data, as elaborated in Section 3.2.3. Our methodology ensured that fragments generated from actual manuscripts, including their contours, accurately reflect the complexities found in real-world restoration conditions. For the case studies involving LLMCO4MR pipeline, high-quality (with annotated ground truth) real-world Dunhuang fragments are indeed scarce. Our annotated fragment cases were provided by the expert team dedicated to Dunhuang culture research, in the form of visual images for evaluation. All images were obtained from the International Dunhuang Project (IDP). + +Evaluation. For fragment selection with CO solver, we evaluated on synthetic shredded data with accuracy criterion. For instance, in a scenario where the task is to select Top-4 pieces from 10 candidates, fragment correctly identified as one of the 4 ground truth fragments is deemed a successful selection. For the real-world case study on Dunhuang, we evaluate the performance by measuring the success rate of fragment restoration. A successful restoration is defined as a match that aligns with the ground truth candidate pairings provided by archaeologists. This expert-annotated data serves as our benchmark for assessing the accuracy of our method in real-world scenarios. + +# 4.2 Simulation Experiment on Combinatorial Solver + +Training Details. Aligned with Section 3.1, CO solver is designed to select $2 \times 2$ and $3 \times 3$ fragments from a set $\mathcal{F}$ comprising $m$ fragments, aiming to address the Top- $K$ fragment selection challenge. Thus we initially selected $k$ pieces that originate from the same fragment and then added $(m - k)$ random fragments as outliers to compile a group. We assembled 500 such groups as + +Table 2: Performance of CO solver siamese feature extraction network. + +
Data QuantityHybrid RatioSiamese AccuracyTop-4 Performance
400001:10.87460.6925
2:30.92120.7025
1:20.91170.7100
800001:10.95340.7250
2:30.94370.7325
1:20.96750.7500
1200001:10.92450.7525
2:30.95860.7575
1:20.97310.7800
+ +Table 3: Performance of CO solver on various Top- $K$ fragments selection task. + +
Top-K2×23×3
Pool Size101520202530
Random Select0.40000.26670.20000.45000.36000.3000
Baseline0.74750.31250.25750.67560.57670.5389
OT Layer0.78000.58250.42250.72560.64440.6044
+ +training dataset, reserving an additional 100 groups for evaluation. We trained the siamese feature extraction network with 40K, 80K and 120K positive and negative pairing samples, with the ratio of training and testing being 6:1. + +Siamese Feature Extraction Network. As detailed in Table 2, we evaluate the performance across various training data quantity, positive and negative sample ratios utilizing a ResNet backbone. Our findings indicate that CO solver performance peaks on Top-4 selections with 120K samples and 1:2 hybrid ratio on positive and negative matching samples. This can be attributed to the imbalance between positive and negative samples in pairwise matching manuscript fragments. With more negative samples, the reliability of siamese model is improved and CO solver performance is enhanced. + +CO Solver. Tab. 3 presents the simulation experiment results under various configurations of Top- $K$ fragment selection from $m$ candidates. "Random select" refers to the Top- $K$ random selections mathematical expectation. "Baseline" refers to selecting the Top- $K$ probabilities in a greedy manner from GNN output without OT layer. The introduction of the OT layer, which enforces cardinality constraints, is observed to enhance model performance, particularly as the candidate size $m$ increases. This improvement is attributed to the OT layer's rigorous penalization of deviations from the optimization objectives in Eqs. (1a) and (1b). + +Table 4: Enhancement from the second-order Top- $K$ selection. + +
StrategyTop-9 of 20Top-9 of 25Top-9 of 30
Select at Once0.72560.64440.6044
Second-Order0.73110.68220.6633
+ +Enhancement through Second-order Matching Strategy. In the $2 \times 2$ selection settings, all four ground truth fragments hold equal significance (each fragment only have exactly 2 matching connections with their ground-truth adjacent pieces). Thus we explore the efficacy of the second-order matching strategy introduced in Section 3.2.5 for $3 \times 3$ scenarios and detail the results in Tab. 4. The second-order matching strategy duplicates the cardinality constraint and applies more stringent penalties for incorrect selections. This method aims to mitigate the effects of erroneous pairwise matches proposed by the siamese network, thereby advancing the accuracy and reliability of fragment selection. + +Simulation experiments demonstrate our proposed neural CO solver's effectiveness in selecting candidates based on highest mutual assemble similarity. This gives promise for the subsequent real-world restoration with LLMCO4MR. + +# 4.3 Real-world Manuscript Case Studies on Dunhuang + +Results with LLMCO4MR. We conducted case study experiments utilizing LLMCO4MR and compared its effectiveness against other SOTAs, as detailed in Tab. 5. We input a group of fragments with outliers to CO solver for selection, then use the selected Top- $K$ fragments as image prompts for GPT-4V, which suggests the matching scores and relative directions. For SOTAs [31,51] that lack the capability to filter out outliers, we repeatedly input the fragments pairs and select the matching results in a greedy manner (select the fragment candidate pairs with ranked highest matching score). + +We observed that previous works often operated under restricted conditions, such as assuming no outliers, limiting search spaces, regular shape fragments, or using manually outlined perfect matching contours. These constraints do not reflect the open-world setting addressed by our method. Moreover, contour similarity network does not achieve perfect accuracy in matching judgments and real-world scenarios frequently involve more irregular fragments, underscoring the complexity of the task. + +Our pipeline addresses these challenges in two ways. First, the CO solver captures irregularity through data augmentation. We extract contours from real-world Dunhuang manuscripts and apply to shredding to adequately represent real-world challenges and mitigate the risk of overfitting unrealistic characteristics in training. Second, the MLLM component captures cross-image similarity using multimodal features, including contour (even with some imperfect fits and irregular shapes) and other visual features. + +Table 5: Evaluation on real-world Dunhuang fragment restoration case studies. + +
Top-K2×23×3
Pool Size10152025
GPT-4V [26]0.32500.17500.37780.2533
LLaVA [21]0.31500.12500.27780.1556
JigsawNet [19]0.52500.37500.45560.3222
Papyrus [31]0.42500.37500.47780.4111
S3-Net [51]0.31250.25750.38890.2111
CO Solver0.57500.52500.53330.4667
LLMCO4MR (Ours)0.67500.62500.62220.5556
+ +Ablation Study. We also present ablation study results in Tab. 5. "CO Solver" refers to deriving matching pairs directly from the output of CO solver. Few selected MLLMs exhibit capabilities in understanding cross-image fragment relations. These MLLMs' performances were evaluated by processing pairs of fragments without CO solver filtering out outliers. We conclude that similar to retrieval-augmented LLM frameworks, CO solver in our LLMCO4MR is a nontrivial extension, empowers MLLM capability handling combinatorial explosion problems. We detail in the Appendix the current limitations of MLLMs in directly addressing manuscript restoration tasks with combinatorial explosion. + +Our multifaceted approach serves two purposes: it complements and enhances traditional contour-based matching, and it helps to stabilize the variable performance of MLLMs in image matching. As a result, our method demonstrates increased effectiveness in handling real-world manuscript restoration tasks, especially in scenarios where either matching strategy alone is insufficient. + +# 5 Conclusion + +We have presented a novel pipeline, LLMCO4MR, designed for the intricate task of restoring ancient manuscripts from fragments. We first formalize the problem as a cardinality-constrained combinatorial problem and then a neural combinatorial solver is devised to produce the Top- $K$ candidates for merging. Then the MLLM is employed, leveraging the candidates as prompt for final restoration. Experiments on both synthetic fragments and real-world Dunhuang data demonstrate the potential of our approach, compared with the traditional approaches which require labors of well-trained archaeologists. + +**Limitation:** Like existing works, our current approach also cannot handle the case with missing fragments and the fragments are required to be basically in a grid layout. Another unsolved case is the existence of both small and large fragments, which we leave for future work. Also, more case studies on broader scope need to be studied when such data becomes available. + +# Acknowledgements + +This work was in part supported by National Science and Technology Major Project (2022ZD0119100), National Natural Science Foundation of China (No. 62441605, 72342023), National Social Science Fund of China (No. 21BYY142, 14AZS001), Key Research and Development Program of Zhejiang Province (No. 2024C03270), the StarryNight Science Fund of Zhejiang University Shanghai Institute for Advanced Study (SN-ZJU-SIAS-0010). + +# References + +1. Abitbol, R., Shimshoni, I., Ben-Dov, J.: Machine learning based assembly of fragments of ancient papyrus. Journal on Computing and Cultural Heritage (JOCCH) 14(3), 1-21 (2021) +2. Assael, Y., Sommerschield, T., Shillingford, B., Bordbar, M., Pavlopoulos, J., Chatzipanagiotou, M., Androutsopoulos, I., Prag, J., de Freitas, N.: Restoring and attributing ancient texts using deep neural networks. Nature 603(7900), 280-283 (2022) +3. Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., Hui, B., Ji, L., Li, M., Lin, J., Lin, R., Liu, D., Liu, G., Lu, C., Lu, K., Ma, J., Men, R., Ren, X., Ren, X., Tan, C., Tan, S., Tu, J., Wang, P., Wang, S., Wang, W., Wu, S., Xu, B., Xu, J., Yang, A., Yang, H., Yang, J., Yang, S., Yao, Y., Yu, B., Yuan, H., Yuan, Z., Zhang, J., Zhang, X., Zhang, Y., Zhang, Z., Zhou, C., Zhou, J., Zhou, X., Zhu, T.: Qwen technical report (2023) +4. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond (2023) +5. Bengio, Y., Lodi, A., Prouvost, A.: Machine learning for combinatorial optimization: a methodological tour d'horizon. European Journal of Operational Research 290(2), 405-421 (2021) +6. Besl, P.J., McKay, N.D.: Method for registration of 3-d shapes. In: Sensor fusion IV: control paradigms and data structures. vol. 1611, pp. 586-606. Spie (1992) +7. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) +8. Buchbinder, N., Feldman, M., Naor, J., Schwartz, R.: Submodular maximization with cardinality constraints. In: Proceedings of the twenty-fifth annual ACM-SIAM symposium on Discrete algorithms. pp. 1433-1452. SIAM (2014) +9. Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems 26 (2013) +0. Derech, N., Tal, A., Shimshoni, I.: Solving archaeological puzzles. Pattern Recognition 119, 108065 (2021) +11. Driess, D., Xia, F., Sajjadi, M.S.M., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., Huang, W., Chebotar, Y., Sermanet, P., Duckworth, D., Levine, S., Vanhoucke, V., Hausman, K., Toussaint, M., Greff, K., Zeng, A., Mordatch, I., Florence, P.: Palm-e: An embodied multimodal language model (2023) +12. Galambos, I.: Dunhuang Manuscript Culture: End of the First Millennium, vol. 22. Walter de Gruyter GmbH & Co KG (2020) + +13. da Gama Leitao, H.C., Stolfi, J.: A multiscale method for the reassembly of two-dimensional fragmented objects. IEEE Transactions on Pattern Analysis and Machine Intelligence 24(9), 1239-1251 (2002) +14. Ge, J., Luo, H., Qian, S., Gan, Y., Fu, J., Zhang, S.: Chain of thought prompt tuning in vision language models (2023) +15. Guo, P.F., Chen, Y.H., Tsai, Y.D., Lin, S.D.: Towards optimizing with large language models. arXiv preprint arXiv:2310.05204 (2023) +16. Guo, Q., Wang, R., Guo, J., Li, B., Song, K., Tan, X., Liu, G., Bian, J., Yang, Y.: Connecting large language models with evolutionary algorithms yields powerful prompt optimizers. arXiv preprint arXiv:2309.08532 (2023) +17. Hossieni, S.S., Shabani, M.A., Irandoust, S., Furukawa, Y.: Puzzlefusion: Unleashing the power of diffusion models for spatial puzzle solving. Advances in Neural Information Processing Systems 36 (2024) +18. Karalias, N., Loukas, A.: Erdos goes neural: an unsupervised learning framework for combinatorial optimization on graphs. Advances in Neural Information Processing Systems 33, 6659-6672 (2020) +19. Le, C., Li, X.: Jigsawnet: Shredded image reassembly using convolutional neural network and loop-based composition. IEEE Transactions on Image Processing 28(8), 4000-4015 (2019) +20. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023) +21. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024) +22. Lu, J., Sun, Y., Huang, Q.: Jigsaw: Learning to assemble multiple fractured objects. arXiv preprint arXiv:2305.17975 (2023) +23. Lyu, C., Wu, M., Wang, L., Huang, X., Liu, B., Du, Z., Shi, S., Tu, Z.: Macaw-lm: Multi-modal language modeling with image, audio, video, and text integration (2023) +24. Maaz, M., Rasheed, H., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models (2023) +25. Markaki, S., Panagiotakis, C.: Jigsaw puzzle solving techniques and applications: a survey. The Visual Computer pp. 1-17 (2022) +26. OpenAI: Gpt-4 technical report (2023) +27. Panagiotakis, C., Markaki, S., Kokinou, E., Papadakis, H.: Coastline matching via a graph-based approach. Computational Geosciences 26(6), 1439-1448 (2022) +28. Paumard, M.M., Picard, D., Tabia, H.: Jigsaw puzzle solving using local feature co-occurrences in deep neural networks. In: 2018 25th IEEE international conference on image processing (ICIP). pp. 1018-1022. IEEE (2018) +29. Pengcheng, G., Gang, G., Jiangqin, W., Baogang, W.: Chinese calligraphic style representation for recognition. International Journal on Document Analysis and Recognition (IJDAR) 20, 59-68 (2017) +30. Pirrone, A., Aimar, M.B., Journet, N.: Papy-s-net: A siamese network to match papyrus fragments. In: Proceedings of the 5th International Workshop on Historical Document Imaging and Processing. pp. 78-83 (2019) +31. Pirrone, A., Beurton-Aimar, M., Journet, N.: Self-supervised deep metric learning for ancient papyrus fragments retrieval. International Journal on Document Analysis and Recognition (IJDAR) 24(3), 219-234 (2021) +32. Rusinkiewicz, S., Levoy, M.: Efficient variants of the icp algorithm. In: Proceedings third international conference on 3-D digital imaging and modeling. pp. 145-152. IEEE (2001) +33. Savelka, J., Ashley, K.D., Gray, M.A., Westermann, H., Xu, H.: Explaining legal concepts with augmented large language models (gpt-4) (2023) + +34. Savino, P., Tonazzini, A.: Digital restoration of ancient color manuscripts from geometrically misaligned recto-verso pairs. Journal of Cultural Heritage 19, 511-521 (2016) +35. Shuai, R.W., Ruffolo, J.A., Gray, J.J.: Generative language modeling for antibody design. bioRxiv (2022). https://doi.org/10.1101/2021.12.13.472419, https://www.biorxiv.org/content/early/2022/12/20/2021.12.13.472419 +36. Singhal, K., Azizi, S., Tu, T., Mahdavi, S.S., Wei, J., Chung, H.W., Scales, N., Tanwani, A., Cole-Lewis, H., Pfohl, S., Payne, P., Seneviratne, M., Gamble, P., Kelly, C., Scharli, N., Chowdhery, A., Mansfield, P., y Arcas, B.A., Webster, D., Corrado, G.S., Matias, Y., Chou, K., Gottweis, J., Tomasev, N., Liu, Y., Rajkomar, A., Barral, J., Semturs, C., Karthikesalingam, A., Natarajan, V.: Large language models encode clinical knowledge (2022) +37. Sommerscheid, T., Assael, Y., Pavlopoulos, J., Stefanak, V., Senior, A., Dyer, C., Bodel, J., Prag, J., Androutsopoulos, I., de Freitas, N.: Machine learning for ancient languages: A survey. Computational Linguistics pp. 1-44 (2023) +38. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) +39. Tsamoura, E., Pitas, I.: Automatic color based reassembly of fragmented images and paintings. IEEE Transactions on Image Processing 19(3), 680-690 (2009) +40. Vinyals, O., Fortunato, M., Jaitly, N.: Pointer networks. Advances in neural information processing systems 28 (2015) +41. Wang, R., Shen, L., Chen, Y., Yang, X., Tao, D., Yan, J.: Towards one-shot neural combinatorial solvers: Theoretical and empirical notes on the cardinality-constrained case. In: The Eleventh International Conference on Learning Representations (2022) +42. Wang, R., Yan, J., Yang, X.: Learning combinatorial embedding networks for deep graph matching. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3056-3065 (2019) +43. Wang, R., Zhang, Y., Guo, Z., Chen, T., Yang, X., Yan, J.: Linsatnet: The positive linear satisfiability neural networks. In: International Conference on Machine Learning (ICML) (2023) +44. Wolfson, H., Schonberg, E., Kalvin, A., Lamdan, Y.: Solving jigsaw puzzles by computer. Annals of Operations Research 12(1), 51-64 (1988) +45. Xi, Z., Chen, W., Guo, X., He, W., Ding, Y., Hong, B., Zhang, M., Wang, J., Jin, S., Zhou, E., et al.: The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864 (2023) +46. Xie, Y., Dai, H., Chen, M., Dai, B., Zhao, T., Zha, H., Wei, W., Pfister, T.: Differentiable top-k with optimal transport. Advances in Neural Information Processing Systems 33, 20520-20531 (2020) +47. Yang, A., Xiao, B., Wang, B., Zhang, B., Bian, C., Yin, C., Lv, C., Pan, D., Wang, D., Yan, D., Yang, F., Deng, F., Wang, F., Liu, F., Ai, G., Dong, G., Zhao, H., Xu, H., Sun, H., Zhang, H., Liu, H., Ji, J., Xie, J., Dai, J., Fang, K., Su, L., Song, L., Liu, L., Ru, L., Ma, L., Wang, M., Liu, M., Lin, M., Nie, N., Guo, P., Sun, R., Zhang, T., Li, T., Li, T., Cheng, W., Chen, W., Zeng, X., Wang, X., Chen, X., Men, X., Yu, X., Pan, X., Shen, Y., Wang, Y., Li, Y., Jiang, Y., Gao, Y., Zhang, Y., Zhou, Z., Wu, Z.: Baichuan 2: Open large-scale language models (2023) +48. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of lmmss: Preliminary explorations with gpt-4v(ison) (2023) + +49. Yang, Z., Li, L., Wang, J., Lin, K., Azarnasab, E., Ahmed, F., Liu, Z., Liu, C., Zeng, M., Wang, L.: Mm-react: Prompting chatgpt for multimodal reasoning and action (2023) +50. Yu, F., Quartey, L., Schilder, F.: Legal prompting: Teaching a language model to think like a lawyer (2022) +51. Zhang, C., Wang, B., Chen, K., Zong, R., Mo, B.f., Men, Y., Almpanidis, G., Chen, S., Zhang, X.: Data-driven oracle bone rejoining: A dataset and practical self-supervised learning scheme. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 4482-4492 (2022) +52. Zhang, C., Zong, R., Cao, S., Men, Y., Mo, B.: Ai-powered oracle bone inscriptions recognition and fragments rejoining. In: Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence. pp. 5309-5311 (2021) +53. Zhang, K., Li, X.: A graph-based optimization algorithm for fragmented image reassembly. Graphical Models 76(5), 484-495 (2014) +54. Zhang, Y., Fang, Z., Yang, X., Zhang, S., He, B., Dou, H., Yan, J., Zhang, Y., Wu, F.: Reconnecting the broken civilization: Patchwork integration of fragments from ancient manuscripts. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 1157-1166 (2023) +55. Zhang, Z., Wang, Y.T., Li, B., Guo, A., Liu, C.L.: Deep rejoining model for oracle bone fragment image. In: Asian Conference on Pattern Recognition. pp. 3-15. Springer (2021) +56. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models (2023) \ No newline at end of file diff --git a/2024/LLMCO4MR_ LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang/images.zip b/2024/LLMCO4MR_ LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..10e58d98084bf56ca7f9e250fd8d3a684fe285c2 --- /dev/null +++ b/2024/LLMCO4MR_ LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a203dd5328d56f703ef7b7d05fc239a3117f911892db91b17c896e7ae3ab6d3 +size 422030 diff --git a/2024/LLMCO4MR_ LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang/layout.json b/2024/LLMCO4MR_ LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..83cca2e82f97283d01e06965e1f55d8f0d0a8ff2 --- /dev/null +++ b/2024/LLMCO4MR_ LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang/layout.json @@ -0,0 +1,9033 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 166 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 166 + ], + "type": "text", + "content": "LLMCO4MR: LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 135, + 186, + 478, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 186, + 478, + 222 + ], + "spans": [ + { + "bbox": [ + 135, + 186, + 478, + 222 + ], + "type": "text", + "content": "Yuqing Zhang" + }, + { + "bbox": [ + 135, + 186, + 478, + 222 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 186, + 478, + 222 + ], + "type": "text", + "content": ", Hangqi Li" + }, + { + "bbox": [ + 135, + 186, + 478, + 222 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 186, + 478, + 222 + ], + "type": "text", + "content": ", Shengyu Zhang" + }, + { + "bbox": [ + 135, + 186, + 478, + 222 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 135, + 186, + 478, + 222 + ], + "type": "text", + "content": ", Runzhong Wang" + }, + { + "bbox": [ + 135, + 186, + 478, + 222 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 135, + 186, + 478, + 222 + ], + "type": "text", + "content": ", Baoyi He" + }, + { + "bbox": [ + 135, + 186, + 478, + 222 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 186, + 478, + 222 + ], + "type": "text", + "content": ", Huaiyong Dou" + }, + { + "bbox": [ + 135, + 186, + 478, + 222 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 186, + 478, + 222 + ], + "type": "text", + "content": ", Junchi Yan" + }, + { + "bbox": [ + 135, + 186, + 478, + 222 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 135, + 186, + 478, + 222 + ], + "type": "text", + "content": ", Yongquan Zhang" + }, + { + "bbox": [ + 135, + 186, + 478, + 222 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 186, + 478, + 222 + ], + "type": "text", + "content": ", and Fei Wu" + }, + { + "bbox": [ + 135, + 186, + 478, + 222 + ], + "type": "inline_equation", + "content": "^{1,2,4}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 225, + 231, + 389, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 231, + 389, + 243 + ], + "spans": [ + { + "bbox": [ + 225, + 231, + 389, + 243 + ], + "type": "text", + "content": "1 Zhejiang University, Hangzhou, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 243, + 473, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 243, + 473, + 254 + ], + "spans": [ + { + "bbox": [ + 140, + 243, + 473, + 254 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 243, + 473, + 254 + ], + "type": "text", + "content": " Shanghai Institute for Advanced Study of Zhejiang University, Shanghai, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 204, + 254, + 410, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 254, + 410, + 265 + ], + "spans": [ + { + "bbox": [ + 204, + 254, + 410, + 265 + ], + "type": "text", + "content": "3 Shanghai Jiao Tong University, Shanghai, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 216, + 265, + 397, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 265, + 397, + 276 + ], + "spans": [ + { + "bbox": [ + 216, + 265, + 397, + 276 + ], + "type": "text", + "content": "4 Shanghai AI Laboratory, Shanghai, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 157, + 277, + 456, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 277, + 456, + 287 + ], + "spans": [ + { + "bbox": [ + 157, + 277, + 456, + 287 + ], + "type": "text", + "content": "{yuqingz7, sy_zhang, wufei}@zju.edu.cn, yanjunchi@sjtu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 322, + 455, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 322, + 455, + 552 + ], + "spans": [ + { + "bbox": [ + 159, + 322, + 455, + 552 + ], + "type": "text", + "content": "Abstract. Restoring ancient manuscripts fragments, such as those from Dunhuang, is crucial for preserving human historical culture. However, their worldwide dispersal and the shifts in cultural and historical contexts pose significant restoration challenges. Traditional archaeological efforts primarily focus on manually piecing major fragments together, yet the small and more intricate pieces remain largely unexplored, which is technically due to their irregular shapes, sparse textual content, and extensive combinatorial space for reassembly. In this paper, we formalize the task of restoring the ancient manuscript from fragments as a cardinality-constrained combinatorial optimization problem, and propose a framework named LLMCO4MR: (Multimodal) Large Language Model-aided Combinatorial Optimization Neural Networks for Ancient Manuscript Restoration. Specifically, LLMCO4MR encapsulates a neural combinatorial solver equipped with a differentiable optimal transport (OT) layer, to efficiently predict the Top-" + }, + { + "bbox": [ + 159, + 322, + 455, + 552 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 159, + 322, + 455, + 552 + ], + "type": "text", + "content": " likely mutual reassembly candidates. Multimodal Large Language Model (MLLM) is then adopted and prompted to yield pairwise matching confidence and relative directions for final restoration. Experiments on synthetic data and cases studies with real-world famous Dunhuang fragments demonstrate our approach's practical potential in assisting archaeologists. Our method provides a novel perspective for ancient manuscript restoration." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 562, + 453, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 562, + 453, + 573 + ], + "spans": [ + { + "bbox": [ + 159, + 562, + 453, + 573 + ], + "type": "text", + "content": "Keywords: Ancient Manuscript Restoration " + }, + { + "bbox": [ + 159, + 562, + 453, + 573 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 562, + 453, + 573 + ], + "type": "text", + "content": " Dunhuang Manuscript " + }, + { + "bbox": [ + 159, + 562, + 453, + 573 + ], + "type": "inline_equation", + "content": "\\cdot" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 159, + 573, + 432, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 573, + 432, + 584 + ], + "spans": [ + { + "bbox": [ + 159, + 573, + 432, + 584 + ], + "type": "text", + "content": "Combinatorial Optimization " + }, + { + "bbox": [ + 159, + 573, + 432, + 584 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 573, + 432, + 584 + ], + "type": "text", + "content": " Multimodal Large Language Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 601, + 230, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 230, + 615 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 230, + 615 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 624, + 481, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 481, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 481, + 648 + ], + "type": "text", + "content": "The restoration of ancient manuscripts has long been the focus of archaeologists' research and the global community. Traditionally, these codices are frequently" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 654, + 240, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 654, + 240, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 654, + 240, + 665 + ], + "type": "text", + "content": "* Corresponding Authors." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 128, + 268, + 184 + ], + "blocks": [ + { + "bbox": [ + 140, + 128, + 268, + 184 + ], + "lines": [ + { + "bbox": [ + 140, + 128, + 268, + 184 + ], + "spans": [ + { + "bbox": [ + 140, + 128, + 268, + 184 + ], + "type": "image", + "image_path": "13fed47de562bd0df09c0cd2641f6ae7246e74da46965579ee1a5e721fbd6b22.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 186, + 282, + 198 + ], + "lines": [ + { + "bbox": [ + 133, + 186, + 282, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 186, + 282, + 198 + ], + "type": "text", + "content": "(a) Conventional manuscript restoration task" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 302, + 114, + 462, + 184 + ], + "blocks": [ + { + "bbox": [ + 302, + 114, + 462, + 184 + ], + "lines": [ + { + "bbox": [ + 302, + 114, + 462, + 184 + ], + "spans": [ + { + "bbox": [ + 302, + 114, + 462, + 184 + ], + "type": "image", + "image_path": "8d13fbb16a18f7860c294a586be2a40f57fe501885504e80200f7cc2b089fe56.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 294, + 186, + 480, + 198 + ], + "lines": [ + { + "bbox": [ + 294, + 186, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 294, + 186, + 480, + 198 + ], + "type": "text", + "content": "(b) Our task: Small fragments restoration with outliers" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 206, + 482, + 274 + ], + "lines": [ + { + "bbox": [ + 130, + 206, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 206, + 482, + 274 + ], + "type": "text", + "content": "Fig. 1: Comparison of conventional manuscript restoration and our task. (a) highlights the conventional restoration task with large fragments. Archaeologists focus on written content as primary guide for restoration, supplemented by texture and edges. (b) outlines our task. In an open-world problem setting, small fragments contain limited text. Recovery relies on contour similarity matching. This is generally difficult for humans to identify fragment pairs from a large collection containing outliers." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 480, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 480, + 350 + ], + "type": "text", + "content": "discovered in a (severely) deteriorated condition due to the passage of time. Recovering these manuscripts involves extensive efforts [37] that are often time-consuming, labor-intensive, and require a high level of expertise, necessitating years of professional training [2]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 350, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 495 + ], + "type": "text", + "content": "On the other hand, vision technologies have demonstrated significant potential in aiding the restoration of ancient manuscripts. For example, [1] developed a sophisticated system based on Convolutional Neural Network (CNN) to identify matching candidates among papyrus fragments by analyzing their fiber patterns. Zhang et al. [51] proposed a self-supervised model using Generative Adversarial Network (GAN) and siamese network for the pairwise stitching of Oracle Bone pieces, utilizing manually annotated contours for similarity comparison. While these vision-assisted pipelines enhance efficiency and accuracy beyond traditional manual efforts, their applications are generally confined to local pairwise matching [31], or regularly-shaped (e.g., rectangle) and large fragments [54]. A major challenge persists in dealing with real-world small fragments, which are often less-textual, shape-irregular, and scattered with missing fragments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 496, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 482, + 592 + ], + "type": "text", + "content": "These small fragments have often been overlooked by archaeologists, yet they form a substantial part of the manuscript collections. Recent explorations have revealed that these minor pieces are treasure troves of first-hand materials, offering unique insights into the daily lives of past civilizations [12]. Current deep learning approaches rely on textual [29], contextual [2] or multimodal information [54] to reassembly manuscripts. Such learning-based methods require a large amount of annotated training data for complex multimodal feature extraction, which is impractical for small fragment restoration." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "To bridge the gap, we resort to the recent advancements in large language models (LLMs) that have revolutionized traditional tasks in computer vision and natural language processing [7, 26, 38]. Notably, the emergence of Multimodal Large Language Models (MLLMs) has demonstrated exceptional capabilities in extracting and interpreting image features, understanding content, and following complex prompt instructions [4, 20, 48]. Capitalizing on these developments, our" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 168, + 114, + 445, + 199 + ], + "blocks": [ + { + "bbox": [ + 168, + 114, + 445, + 199 + ], + "lines": [ + { + "bbox": [ + 168, + 114, + 445, + 199 + ], + "spans": [ + { + "bbox": [ + 168, + 114, + 445, + 199 + ], + "type": "image", + "image_path": "0a7e97e324120a236faeeaf7f01e2fbf5106846592f1c96c4fbc5dc4f86f6b3f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 208, + 470, + 220 + ], + "lines": [ + { + "bbox": [ + 141, + 208, + 470, + 220 + ], + "spans": [ + { + "bbox": [ + 141, + 208, + 470, + 220 + ], + "type": "text", + "content": "Fig. 2: Showcases of matching results with case studies on Dunhuang fragments." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 242, + 480, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 242, + 480, + 337 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 480, + 337 + ], + "type": "text", + "content": "research introduces a novel approach by employing MLLMs to tackle the intricate task of restoring ancient manuscript fragments. MLLMs facilitate efficient pairwise matching by: 1) identifying cross-image connections based on contour similarity and capturing complex features similarity in paper texture and content; 2) replicating the comprehensive integration of multimodal semantics that archaeologists employ during their assembly process; 3) demonstrating zero-shot learning capabilities, which are crucial for investigating manuscripts with limited ground truth availability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 338, + 481, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 481, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 481, + 433 + ], + "type": "text", + "content": "However, the complexity of our task surpasses the current capabilities of MLLMs. In practical scenarios of ancient manuscript restoration, it is necessary to first identify potential candidate fragments from a larger collection for subsequent pairwise matching. This collection contains fragments that may not belong to the target manuscript, which we refer to as \"outliers\". Performing pairwise comparisons using MLLMs on the vast number of small fragments, including these outliers, leads to a combinatorial explosion challenge, which poses a significant difficulty for MLLM itself to manage effectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 434, + 481, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 481, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 481, + 494 + ], + "type": "text", + "content": "To address these challenges, we formulate the fragment selection subtask as a combinatorial optimization (CO) problem. We introduce a novel framework, named LLMCO4MR, a LLMs-aided Neural Combinatorial Optimization solver for manuscript restoration, transforming the unstructured challenge into a technically feasible task. The highlights of the paper are:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 498, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 136, + 498, + 480, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 498, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 136, + 498, + 480, + 545 + ], + "type": "text", + "content": "1. More Open-world Problem Setting. Beyond the restricted setting requiring that there are no outlier fragments for recovery, to our best knowledge, we are the first to address the more realistic and challenging setting, i.e., restoring the manuscript from a subset of fragments with outliers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 546, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 546, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 136, + 546, + 480, + 593 + ], + "type": "text", + "content": "2. New Two-stage Solving Pipeline. We solve the new problem by two stages: fragment selection and pairwise matching. In the first stage, " + }, + { + "bbox": [ + 136, + 546, + 480, + 593 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 136, + 546, + 480, + 593 + ], + "type": "text", + "content": " fragments with the highest mutual assemble scores are selected by certain means while the latter stage involves the traditional closed-world problem setting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 594, + 481, + 665 + ], + "type": "text", + "content": "3. Neural Combinatorial Solver for Candidate Selection. We introduce a combinatorial network to solve the fragment selection problem by formulating it as a " + }, + { + "bbox": [ + 136, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 136, + 594, + 481, + 665 + ], + "type": "text", + "content": "-cardinality constrained task, with an advantage that the input visual fragments are learned end-to-end directly with the final decision goal. Supervised training is achieved by using the shredded fragments generated by splitting a complete manuscript into grids." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 448, + 103 + ], + "type": "text", + "content": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 136, + 115, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 115, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 482, + 189 + ], + "type": "text", + "content": "4. MLLMs for Pairwise Fragment Matching in Zero-shot. The selected Top-" + }, + { + "bbox": [ + 136, + 115, + 482, + 189 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 136, + 115, + 482, + 189 + ], + "type": "text", + "content": " fragments are then fed into the MLLM by pair, to generate the matching score as well as the relative direction of matching, i.e., horizontal or vertical. Then the final manuscript can be recovered by these local pairwise matchings. To our best knowledge, this is the first time to introduce MLLMs for manuscript recovery or more broadly speaking, jigsaw puzzle-like tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 208, + 243, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 208, + 243, + 221 + ], + "spans": [ + { + "bbox": [ + 132, + 208, + 243, + 221 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 235, + 482, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 482, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 482, + 438 + ], + "type": "text", + "content": "Manuscript Restoration. This area has recently witnessed significant advancements [2, 37] aided with vision technology for restoration. These developments aim to overcome the challenges in manual methods, which are time-consuming, labor-intensive, and heavily reliant on expert knowledge [2]. A typical method is to solve the problem from a jigsaw puzzle [10, 34] solving perspective. For puzzle solving and global restoration, several successful methods have emerged. Traditional strategies include curve-based techniques [27, 44] and time series sequence-based approaches [6, 13, 32, 39, 53]. Additionally, deep learning methods have been developed, exemplified by works such as [19] and other studies [17, 25, 28]. Inspired by these works, previous manuscript restoration methods are designed based on the visual features of the fragments. Here, matching often refers to identifying whether two fragments are adjacent and their relative direction (horizontally or vertically). In particular, the feature extraction module has been further advanced by deep learning with various aspects, including fragment contour shape [51, 52, 55], texture details [31], and written characters [29]. In the recent work [54], a multimodal restoration pipeline is devised that considers both written content and contour shape." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 439, + 483, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 483, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 483, + 511 + ], + "type": "text", + "content": "Despite these advances, most current relic restoration methods share a common assumption: the candidate pieces for pairwise matching are given, either by manual pre-selection or other means. However, in many real-world restoration cases, this assumption leaves a gap in addressing the more challenging task of open-world problem setting, where candidate fragments are not predetermined, underscoring the novelty and necessity of our approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "content": "Combinatorial Solver. The recent growing trend of learnable CO solvers [5] demonstrate the potential of solving CO problems by deep neural networks with higher efficiency [18, 22, 40, 42]. To tackle the extensive combinatorial space of ancient manuscript restoration, we present a cardinality-constrained CO reformulation [8] by selecting the Top- " + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "content": " most likely matching fragments from the candidate set. Following the one-shot CO learning paradigm [41], we adopt [43], which handles cardinality-constrained problems similarly to differentiable optimal transport [46]. [43] has the advantage of controlling the tightness of constraints while preserving the gradient, which is crucial in our task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Large Language Models. With the recent arisen success of Large Language Model (LLM), such as GPT-4 [26], Qwen [3] and Baichuan [47], LLMs have showcased their efficacy in a myriad of fields including computer vision [23,24], natural" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 156, + 115, + 462, + 174 + ], + "blocks": [ + { + "bbox": [ + 156, + 115, + 462, + 174 + ], + "lines": [ + { + "bbox": [ + 156, + 115, + 462, + 174 + ], + "spans": [ + { + "bbox": [ + 156, + 115, + 462, + 174 + ], + "type": "image", + "image_path": "8927eadc45cf92d434036fa4efd5f9da0435d52e26a7b3d39108f238fccf9283.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 174, + 395, + 184 + ], + "lines": [ + { + "bbox": [ + 214, + 174, + 395, + 184 + ], + "spans": [ + { + "bbox": [ + 214, + 174, + 395, + 184 + ], + "type": "text", + "content": "(a) Previous Ancient Manuscript Restoration Pipelines" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 153, + 186, + 462, + 239 + ], + "blocks": [ + { + "bbox": [ + 153, + 186, + 462, + 239 + ], + "lines": [ + { + "bbox": [ + 153, + 186, + 462, + 239 + ], + "spans": [ + { + "bbox": [ + 153, + 186, + 462, + 239 + ], + "type": "image", + "image_path": "8037dde0be00c3712ea819cf26f79ccc10207daff02b0758ee916cffd08d0ac9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 240, + 370, + 251 + ], + "lines": [ + { + "bbox": [ + 233, + 240, + 370, + 251 + ], + "spans": [ + { + "bbox": [ + 233, + 240, + 370, + 251 + ], + "type": "text", + "content": "(b) Our Two-stage LLMCO4MS Pipeline" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 259, + 476, + 271 + ], + "lines": [ + { + "bbox": [ + 135, + 259, + 476, + 271 + ], + "spans": [ + { + "bbox": [ + 135, + 259, + 476, + 271 + ], + "type": "text", + "content": "Fig. 3: Comparison between previous restoration approaches and our LLMCO4MR." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 132, + 314, + 483, + 407 + ], + "blocks": [ + { + "bbox": [ + 160, + 293, + 452, + 304 + ], + "lines": [ + { + "bbox": [ + 160, + 293, + 452, + 304 + ], + "spans": [ + { + "bbox": [ + 160, + 293, + 452, + 304 + ], + "type": "text", + "content": "Table 1: Comparison with various ancient fragment restoration SOTA." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 314, + 483, + 407 + ], + "lines": [ + { + "bbox": [ + 132, + 314, + 483, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 483, + 407 + ], + "type": "table", + "html": "
MethodologyTraining DataAnnotation FreeEliminate OutliersNetwork StructureExtracted Features
JigsawNet [19]Self-constructed✓✘CNNTexture
Papyrus [31]Self-constructed✘✘Siamese CNNTexture
S3-Net [51]GAN-augmented✘✘Siamese CNNContour
LLMCO4MRSynthetic Shredding✓✓CO-MLLMMultimodal
", + "image_path": "dc4af29bf9a36812077b0c4f47dbe9fa156192bdc1c052fb19cb97b924dac810.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 440, + 482, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 560 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 560 + ], + "type": "text", + "content": "language processing [38], medical application [35,36], law judgement [33,50], and optimization [15, 16]. Recent advancements in Multimodal-LLM (MLLM) have catalyzed a proliferation of diverse applications, including image content understanding [56], mathematical reasoning [11], and even meme interpolation [49], among others. Inspired by MLLMs' potential in complex visual and textual content understanding, we proposed LLMCO4MR, the first CO solver and MLLM collaboration pipeline to tackle the complex ancient manuscript restoration task. We leverage MLLMs' potential extracting multimodal features [14,24] to tackle small ancient fragments pairwise matching, while further enhancing CO solver performance figuring similar fragments in real-world cases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Discussion with Other Restoration Pipelines. We provide a qualitative comparison with other major jigsaw puzzle solving and fragment restoration pipelines, as shown in Tab. 1. The key distinctions between our LLMCO4MR pipeline and earlier approaches are highlighted in Fig. 3. Our pipeline adeptly addresses the combinatorial explosion challenge inherent in fragment selection through CO solver. Furthermore, it innovates by introducing the use of MLLMs for the pairwise matching of ancient manuscripts, an innovative approach for manuscript restoration with open-world setting." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 448, + 102 + ], + "type": "text", + "content": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 233, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 233, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 233, + 129 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 144, + 482, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 144, + 482, + 192 + ], + "spans": [ + { + "bbox": [ + 130, + 144, + 482, + 192 + ], + "type": "text", + "content": "In this section, we provide a detailed description of our LLMCO4MR pipeline, as composed of two main components: a cardinality-constrained neural solver trained on synthetic data to select candidates for further matching; and the MLLM adoption to achieve effective matching in a zero-shot learning manner." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 214, + 269, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 269, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 269, + 226 + ], + "type": "text", + "content": "3.1 Problem Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 237, + 482, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 286 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 286 + ], + "type": "text", + "content": "We consider a set of " + }, + { + "bbox": [ + 130, + 237, + 482, + 286 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 237, + 482, + 286 + ], + "type": "text", + "content": " fragments, denoted as " + }, + { + "bbox": [ + 130, + 237, + 482, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = \\{f_{in}^{1},\\dots ,f_{in}^{k},f_{out}^{k + 1},\\dots ,f_{out}^{m}\\}" + }, + { + "bbox": [ + 130, + 237, + 482, + 286 + ], + "type": "text", + "content": ". Here the first " + }, + { + "bbox": [ + 130, + 237, + 482, + 286 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 237, + 482, + 286 + ], + "type": "text", + "content": " fragments " + }, + { + "bbox": [ + 130, + 237, + 482, + 286 + ], + "type": "inline_equation", + "content": "f_{in}^{i}" + }, + { + "bbox": [ + 130, + 237, + 482, + 286 + ], + "type": "text", + "content": " refer to those originating from a more complete fragment, while the rest " + }, + { + "bbox": [ + 130, + 237, + 482, + 286 + ], + "type": "inline_equation", + "content": "(m - k)" + }, + { + "bbox": [ + 130, + 237, + 482, + 286 + ], + "type": "text", + "content": " fragments denotes those outliers. The goal of LLMCO4MR is to identify such " + }, + { + "bbox": [ + 130, + 237, + 482, + 286 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 237, + 482, + 286 + ], + "type": "text", + "content": " relevant fragments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 308, + 331, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 331, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 331, + 320 + ], + "type": "text", + "content": "3.2 CO Solver for Fragment Selection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 332, + 481, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 481, + 369 + ], + "type": "text", + "content": "3.2.1 Preliminaries. Based on consultations with archaeology specialists and their insights gained from experimental endeavors, we have formulated the following hypotheses to guide the restoration:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 135, + 380, + 481, + 537 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 136, + 380, + 481, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 380, + 481, + 451 + ], + "spans": [ + { + "bbox": [ + 136, + 380, + 481, + 451 + ], + "type": "text", + "content": "1. Internal Texture and Limited Text. Small real-world fragments typically contain very limited characters, and ancient text recognition is unreliable for matching due to low accuracy. Moreover, even fragments that should be joined together may exhibit different textures due to varying museum collection conditions. Therefore, distinguishing fragments by textual content or texture patterns poses a significant challenge." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 452, + 481, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 452, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 135, + 452, + 481, + 500 + ], + "type": "text", + "content": "2. Edge Shapes and Historical Tearing. Small fragments were often deliberately torn by people in ancient times [12], resulting in contours that are not severely eroded. This key feature offers the potential for primary restoration through contour similarity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 135, + 501, + 481, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 501, + 481, + 537 + ], + "spans": [ + { + "bbox": [ + 135, + 501, + 481, + 537 + ], + "type": "text", + "content": "3. Reassembly Group Size. Fragments are typically reassembled in groups ranging from 3 to 9 pieces. It is uncommon to find a large assembly of pieces that can be combined into a significantly larger fragment." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 547, + 481, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 547, + 481, + 584 + ], + "spans": [ + { + "bbox": [ + 130, + 547, + 481, + 584 + ], + "type": "text", + "content": "In this section, we propose a cardinality-constrained neural CO solver to efficiently and robustly select manuscript fragments, which is further exemplified with MLLM with application on the famous real-world Dunhuang Manuscript." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "3.2.2 Solver Overview. We formulate fragment selection task as cardinality-constrained optimization problem in Eq. (1a). Specifically, the number of nonnegative elements must be restricted to no greater than " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ", while minimizing the objective " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{co}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ". The aim of our formulation is to find the optimal fragments subset with a size of " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " with highest mutual similarity, as shown in Eq. (1b), where " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "D_{ij}" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 479, + 309 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 479, + 309 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 479, + 309 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 479, + 309 + ], + "type": "image", + "image_path": "dd2512891180b48494ff65d0bd37c00f60336b806c5d3938433773ddd9308f27.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 319, + 482, + 397 + ], + "lines": [ + { + "bbox": [ + 132, + 319, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 482, + 397 + ], + "type": "text", + "content": "Fig. 4: LLMCO4MR solving pipeline. 1. Fragment Selection with CO solver. Given a set of " + }, + { + "bbox": [ + 132, + 319, + 482, + 397 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 132, + 319, + 482, + 397 + ], + "type": "text", + "content": " fragments, CO solver first pairs and feeds them into a siamese network to obtain similarity. Next, construct similarity graph " + }, + { + "bbox": [ + 132, + 319, + 482, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 132, + 319, + 482, + 397 + ], + "type": "text", + "content": " and pass it through GNN. Then encode cardinality constraints with differentiate OT layer. Finally output the selected Top- " + }, + { + "bbox": [ + 132, + 319, + 482, + 397 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 319, + 482, + 397 + ], + "type": "text", + "content": " fragments with highest mutual similarity. 2. Pairwise matching with MLLM in zero-shot. Given the selected Top- " + }, + { + "bbox": [ + 132, + 319, + 482, + 397 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 319, + 482, + 397 + ], + "type": "text", + "content": " fragments by pair as visual prompt with text prompt, MLLMs generate matching score and relative directions as output." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 420, + 480, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 480, + 444 + ], + "type": "text", + "content": "denotes the similarity score between candidates " + }, + { + "bbox": [ + 130, + 420, + 480, + 444 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 420, + 480, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 420, + 480, + 444 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 420, + 480, + 444 + ], + "type": "text", + "content": ", which are predicted by siamese network as described in the following section." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 232, + 452, + 481, + 470 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 452, + 481, + 470 + ], + "spans": [ + { + "bbox": [ + 232, + 452, + 481, + 470 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {S}} L _ {c o} \\quad \\text {s . t .} \\mathcal {S} \\in \\{0, 1 \\} ^ {m}, \\| \\mathcal {S} \\| _ {0} \\leq k, \\tag {1a}", + "image_path": "3a5ff0ddd1c6e78757da2b954871afa838a966ce676668b410c1477ba224186f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 202, + 471, + 481, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 471, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 202, + 471, + 481, + 510 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} L _ {c o} = - \\sum_ {i \\in \\{i | s _ {i} = 1 \\}} \\left(\\sum_ {j \\in \\{j | s _ {j} = 1, j \\neq i \\}} D _ {i j}\\right). \\tag {1b}", + "image_path": "15a810c7f80e6655907fe460e2ff4fbb41462eb36184cf49f5036f261845d7b2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 517, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 590 + ], + "type": "text", + "content": "This cardinality-constrained CO problem could be tackled by a differentiable OT layer [43] and Sinkhorn algorithm [9]. As shown in Fig. 4, we construct an end-to-end learning pipeline consisting of a siamese feature extraction module as problem encoder network, a graph neural network fused with an OT layer as the solver network. The neural network learns to solve Eq. (1a) under the unsupervised loss described in Eq. (3) as follows." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "3.2.3 Siamese Feature Extraction Network. In this section, we introduce the siamese feature extraction network of CO solver. Given the scarcity of real-world ground truth data, as few archaeologists have successfully restored small manuscript fragments together, we first describe the synthetic shredding process for preparing training data." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 448, + 102 + ], + "type": "text", + "content": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 114, + 462, + 257 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 462, + 257 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 462, + 257 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 462, + 257 + ], + "type": "image", + "image_path": "a6a77b1041e75cc2b5bec9396bd7f197136b4ace57844f82f871ffa2d79e3d3b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 266, + 414, + 277 + ], + "lines": [ + { + "bbox": [ + 198, + 266, + 414, + 277 + ], + "spans": [ + { + "bbox": [ + 198, + 266, + 414, + 277 + ], + "type": "text", + "content": "Fig. 5: Illustration of synthesized shredding process." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 301, + 482, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 363 + ], + "type": "text", + "content": "Synthetic Shredding Process. We draw an analogy between the intricate work of manuscript restoration and grid puzzle assembly. To simulate this, we generate synthetic shredded fragments by breaking larger manuscript pieces into smaller, puzzle-piece-like fragments. Our synthesis approach is conducted in two phases, as illustrated in Fig. 5:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 373, + 482, + 530 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 136, + 373, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 373, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 136, + 373, + 482, + 469 + ], + "type": "text", + "content": "1. Contour Extraction. We begin by randomly selecting large, real-world manuscripts from which we extract binary boundaries, specifically focusing on the two horizontal and two vertical edges. We discard contours that are either too long or too short, retaining only those of an appropriate length to form our contour set " + }, + { + "bbox": [ + 136, + 373, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{C_1,C_2,\\ldots ,C_m\\}" + }, + { + "bbox": [ + 136, + 373, + 482, + 469 + ], + "type": "text", + "content": ". Each contour in this set is represented by a series of two-dimensional coordinates " + }, + { + "bbox": [ + 136, + 373, + 482, + 469 + ], + "type": "inline_equation", + "content": "C_i = \\{(x_1,y_1),(x_2,y_2),\\dots ,(x_n,y_n)\\}" + }, + { + "bbox": [ + 136, + 373, + 482, + 469 + ], + "type": "text", + "content": ". Through this method, we ensure the synthetic fragments are realistic in size and shape." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 469, + 482, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 469, + 482, + 530 + ], + "spans": [ + { + "bbox": [ + 136, + 469, + 482, + 530 + ], + "type": "text", + "content": "2. Synthetic Shredding. We proceed to sample from the set " + }, + { + "bbox": [ + 136, + 469, + 482, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 136, + 469, + 482, + 530 + ], + "type": "text", + "content": " without replacement to create an " + }, + { + "bbox": [ + 136, + 469, + 482, + 530 + ], + "type": "inline_equation", + "content": "M\\times M" + }, + { + "bbox": [ + 136, + 469, + 482, + 530 + ], + "type": "text", + "content": " grid-like segmentation mask. Utilizing a real rectangular-shaped manuscript piece as our foundation, we apply this mask to segment it into " + }, + { + "bbox": [ + 136, + 469, + 482, + 530 + ], + "type": "inline_equation", + "content": "M^2" + }, + { + "bbox": [ + 136, + 469, + 482, + 530 + ], + "type": "text", + "content": " smaller, discrete fragments. This division simulates realistic fragment sizes and shapes encountered in manuscript restoration." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": "Siamese Network. We employ a siamese feature extraction network designed to assess contour similarity. We start with the original set of fragments with outliers " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = \\{f_{in}^{1},\\dots,f_{in}^{k},f_{out}^{k + 1},\\dots,f_{out}^{m}\\}" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": ". For each fragment, we create permutation combinations of pairwise matches " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "(f^i,f^j)" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": ". These paired fragments, specifically their masked contour, are then fed into two identical branches of a CNN-based neural network. The network computes a similarity score " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "D_{ij} = \\text{Siamese}(f^i,f^j)" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " with the range being [0-1]. A score of 1 indicates that two fragments have matching contours that align perfectly, suggesting adjacency. Conversely, a score of 0 implies not adjacent. Eq. (2) describes the loss term, " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " means learnable parameters, " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "L_{\\text{siamese}}" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " means Binary Cross Entropy (BCE) Loss," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 430, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 430, + 128 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 430, + 128 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 116, + 430, + 128 + ], + "type": "text", + "content": " means predicted similarity score and " + }, + { + "bbox": [ + 130, + 116, + 430, + 128 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 116, + 430, + 128 + ], + "type": "text", + "content": " means ground truth label." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 205, + 138, + 481, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 138, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 205, + 138, + 481, + 152 + ], + "type": "interline_equation", + "content": "L _ {s i a m e s e} = - w [ y \\cdot \\log p + (1 - y) \\cdot \\log (1 - p) ]. \\tag {2}", + "image_path": "971657dfb7740eab6e38784ad9b59fd1c0720543e7e36a4b010ec6b5e7148d7e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 169, + 479, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 169, + 479, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 169, + 479, + 194 + ], + "type": "text", + "content": "3.2.4 Solver Network. Our CO solver network consists of a graph neural network fused with a differentiable OT layer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": "Graph Neural Network. Following the siamese feature extraction network, we construct an undirected similarity graph " + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = (\\mathcal{V},\\mathcal{E})" + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": ". Graph vertices set " + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{v_{1},\\ldots ,v_{m}\\}" + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": " stores the extracted fragment mask feature vector. Given the pairwise similarity scores " + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "inline_equation", + "content": "D_{ij}" + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": " as edge weights, we add an edge " + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "inline_equation", + "content": "e_{ij}" + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": " to the edge set " + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 198, + 480, + 281 + ], + "type": "text", + "content": " if similarity score is greater than threshold. We then constructed a graph neural network to extract features from graph vertices and edges. We use Graph Convolution Network (GCN) as backbone." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "text", + "content": "OT Layer and training. For Top- " + }, + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "text", + "content": " fragments selection, we encode the cardinality constraint with differential OT layer following [41]. To encode cardinality constraints in a differentiable manner, we first formulate the constraints as marginal distributions in optimal transport (OT), and then solve OT with Sinkhorn algorithm [9] by iterative row-wise and column-wise normalization. Given " + }, + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "text", + "content": " GNN output score " + }, + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "text", + "content": " indicating the probability of selecting each piece, OT layer moves Top- " + }, + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "text", + "content": " probability to " + }, + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "inline_equation", + "content": "p_{max}" + }, + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "text", + "content": " while moves others to " + }, + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "inline_equation", + "content": "p_{min}" + }, + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "text", + "content": ". We select " + }, + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "text", + "content": " nodes with the highest probability to predict the output and optimize neural network with loss function in Eq. (3). Loss function consists of three components: " + }, + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "inline_equation", + "content": "L_{siamese}" + }, + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "text", + "content": " penalize incorrect prediction in pairwise similarity; " + }, + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "inline_equation", + "content": "L_{gnn}" + }, + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "text", + "content": " denotes the BCE loss for node selection; and " + }, + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "inline_equation", + "content": "L_{co}" + }, + { + "bbox": [ + 130, + 287, + 482, + 431 + ], + "type": "text", + "content": " indicates the optimization objective in the previously mentioned cardinality constrained CO." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 219, + 443, + 481, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 443, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 219, + 443, + 481, + 456 + ], + "type": "interline_equation", + "content": "\\min L = \\min \\left(\\alpha L _ {\\text {s i a m e s e}} + \\beta L _ {\\text {g n n}} + L _ {\\text {c o}}\\right). \\tag {3}", + "image_path": "94f2657b34ec23b4e77dfc0369579a3ea502033b75767e480844232c12218f8d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": "3.2.5 Second-order Matching As detailed in Section 3.1, fragment groups ranging from 3 to 9 pieces can be reconstructed into a final form. We explore different scenarios of grid-like restoration. In the case of a " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " grid, each ground truth fragment is expected to have two pairwise connections with its neighboring fragments, treating all four fragments as equally important. In contrast, within a " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " puzzle grid, as illustrated in Fig. 6, fragments positioned at [2, 4, 6, and 8] are observed to share three connections each, highlighting the central fragment [5] as most crucial with up to four connections. Meanwhile, the corner fragments [1, 3, 7, 9] are limited to just two connections. Acknowledging the imperfection of siamese network accuracy and the existence of false positive pairwise matches in real-world scenarios, we implement a second-order matching strategy. We start with identifying the Top-5 central pieces that with a higher number of connections. Subsequently, we refine the graph " + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": " prune by pruning edges, retaining only those connected to the previously selected Top-5 fragments and eliminating other distracting edges. This strategy is designed to mitigate the impact of inaccuracies and ensure a more reliable restoration process." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 177, + 114, + 264, + 256 + ], + "blocks": [ + { + "bbox": [ + 177, + 114, + 264, + 256 + ], + "lines": [ + { + "bbox": [ + 177, + 114, + 264, + 256 + ], + "spans": [ + { + "bbox": [ + 177, + 114, + 264, + 256 + ], + "type": "image", + "image_path": "93b3975e7be84aeea621c725f1d038c3c2b0ddb8e691959455596412b6178129.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 258, + 282, + 269 + ], + "lines": [ + { + "bbox": [ + 168, + 258, + 282, + 269 + ], + "spans": [ + { + "bbox": [ + 168, + 258, + 282, + 269 + ], + "type": "text", + "content": "(a) Select with Equivalent Importance" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 277, + 482, + 322 + ], + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 322 + ], + "type": "text", + "content": "Fig. 6: Illustration of second-order matching. (a) selects all Top-9 fragments with equivalent importance (dashed lines represent incorrect pairings obtained by siamese network). (b) highlights centered 5 fragments with more connections, then select the rest 4 fragments based on pruned graph " + }, + { + "bbox": [ + 130, + 277, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{prune}" + }, + { + "bbox": [ + 130, + 277, + 482, + 322 + ], + "type": "text", + "content": " (most incorrect pairings are eliminated)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 288, + 113, + 446, + 255 + ], + "blocks": [ + { + "bbox": [ + 288, + 113, + 446, + 255 + ], + "lines": [ + { + "bbox": [ + 288, + 113, + 446, + 255 + ], + "spans": [ + { + "bbox": [ + 288, + 113, + 446, + 255 + ], + "type": "image", + "image_path": "0db4f04202d4d6949e8ad06bedcc8db9db77c21e1a2247f9e705914fe658b593.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 258, + 410, + 268 + ], + "lines": [ + { + "bbox": [ + 328, + 258, + 410, + 268 + ], + "spans": [ + { + "bbox": [ + 328, + 258, + 410, + 268 + ], + "type": "text", + "content": "(b) Second-order Matching" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 345, + 413, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 345, + 413, + 357 + ], + "spans": [ + { + "bbox": [ + 131, + 345, + 413, + 357 + ], + "type": "text", + "content": "3.3 LLMCO4MR Pipeline for Manuscript Restoration" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 366, + 480, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 480, + 413 + ], + "type": "text", + "content": "Traditional deep learning methods for manuscript restoration heavily depend on human intervention to filter out numerous incorrect pairings generated by models. This task becomes increasingly challenging with smaller fragments, which are abundant and tiny, making manual review burdensome." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 415, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 510 + ], + "type": "text", + "content": "The advent of MLLMs offers high-quality, efficient solutions to tasks that were previously labor-intensive. Innovatively, we are among the first to harness MLLMs for manuscript restoration. MLLMs like GPT-4V have shown remarkable capabilities akin to those of archaeologists by incorporating multimodal semantics for precise pairwise matching, excelling in identifying local pairwise connections with scoring and directional guidance. However, MLLMs struggle with the vast possibilities in open-world settings, particularly with outliers, limiting their ability to perform restoration independently." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Addressing these challenges, we present the LLMCO4MR pipeline, a pioneering MLLM-aided Neural Combinatorial Optimization solver tailored for the restoration of small manuscript fragments. LLMCO4MR leverages a CO solver for initial fragment selection, effectively countering the combinatorial explosion by sidelining outliers. Subsequently, the Top-" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " selected pairs undergo analysis via MLLMs, which utilize custom prompts based on a role-play strategy [45] to refine matching precision. If MLLMs identify any Top-" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " selected fragments as having no viable pairing, these anomalies are removed from the CO solver's input, suggesting a reevaluation of the Top-" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " fragment selections. This iterative refinement enhances LLMCO4MR efficacy. Experimental results in Section 4.3, demonstrate how this synergistic approach not only improves the precision in selecting fragments but also significantly enhances the accuracy and efficiency of manuscript restoration in practical scenarios." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 394, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 394, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 394, + 129 + ], + "type": "text", + "content": "4 Quantitative Evaluation and Case Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "content": "Our experiments encompasses both synthetic data and real-world manuscript fragments. For fragment selection, we conducted simulation experiments using the CO solver, to demonstrate its capability in selecting appropriate fragment candidates for subsequent pairwise matching. We then deployed our LLMCO4MR pipeline on real-world scenarios. Specifically, we conducted case studies on the famous Dunhuang manuscripts, spanning from the 4th to the 14th century. These manuscripts are highly valued by archaeologists and the global community [12], due to their extensive volume, diverse content, and broad historical span." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 297 + ], + "type": "text", + "content": "The combination of simulation experiments and real-world case studies provides a controlled and ideal environment for the quantitative evaluation of our method, aligning with the protocols of peer studies [30, 31, 51, 54]. These experiments were performed on a server equipped with Intel i7 CPUs at " + }, + { + "bbox": [ + 130, + 236, + 482, + 297 + ], + "type": "inline_equation", + "content": "3.2\\mathrm{GHz}" + }, + { + "bbox": [ + 130, + 236, + 482, + 297 + ], + "type": "text", + "content": " and NVIDIA GeForce RTX 3090 GPUs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 315, + 253, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 253, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 253, + 328 + ], + "type": "text", + "content": "4.1 Experiment Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 449 + ], + "type": "text", + "content": "Data. For fragment selection, simulation experiments were employed to train the CO solver using synthetic shredded data, as elaborated in Section 3.2.3. Our methodology ensured that fragments generated from actual manuscripts, including their contours, accurately reflect the complexities found in real-world restoration conditions. For the case studies involving LLMCO4MR pipeline, high-quality (with annotated ground truth) real-world Dunhuang fragments are indeed scarce. Our annotated fragment cases were provided by the expert team dedicated to Dunhuang culture research, in the form of visual images for evaluation. All images were obtained from the International Dunhuang Project (IDP)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 453, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 561 + ], + "type": "text", + "content": "Evaluation. For fragment selection with CO solver, we evaluated on synthetic shredded data with accuracy criterion. For instance, in a scenario where the task is to select Top-4 pieces from 10 candidates, fragment correctly identified as one of the 4 ground truth fragments is deemed a successful selection. For the real-world case study on Dunhuang, we evaluate the performance by measuring the success rate of fragment restoration. A successful restoration is defined as a match that aligns with the ground truth candidate pairings provided by archaeologists. This expert-annotated data serves as our benchmark for assessing the accuracy of our method in real-world scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 579, + 404, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 404, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 404, + 592 + ], + "type": "text", + "content": "4.2 Simulation Experiment on Combinatorial Solver" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Training Details. Aligned with Section 3.1, CO solver is designed to select " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " fragments from a set " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " comprising " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " fragments, aiming to address the Top-" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " fragment selection challenge. Thus we initially selected " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " pieces that originate from the same fragment and then added " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "(m - k)" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " random fragments as outliers to compile a group. We assembled 500 such groups as" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 212, + 135, + 400, + 280 + ], + "blocks": [ + { + "bbox": [ + 160, + 114, + 452, + 125 + ], + "lines": [ + { + "bbox": [ + 160, + 114, + 452, + 125 + ], + "spans": [ + { + "bbox": [ + 160, + 114, + 452, + 125 + ], + "type": "text", + "content": "Table 2: Performance of CO solver siamese feature extraction network." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 212, + 135, + 400, + 280 + ], + "lines": [ + { + "bbox": [ + 212, + 135, + 400, + 280 + ], + "spans": [ + { + "bbox": [ + 212, + 135, + 400, + 280 + ], + "type": "table", + "html": "
Data QuantityHybrid RatioSiamese AccuracyTop-4 Performance
400001:10.87460.6925
2:30.92120.7025
1:20.91170.7100
800001:10.95340.7250
2:30.94370.7325
1:20.96750.7500
1200001:10.92450.7525
2:30.95860.7575
1:20.97310.7800
", + "image_path": "34b5a26d3902e401402c61cc28eefb4aaf9e199ca3b306b66895d1d40c939759.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 161, + 314, + 449, + 388 + ], + "blocks": [ + { + "bbox": [ + 144, + 293, + 468, + 304 + ], + "lines": [ + { + "bbox": [ + 144, + 293, + 468, + 304 + ], + "spans": [ + { + "bbox": [ + 144, + 293, + 468, + 304 + ], + "type": "text", + "content": "Table 3: Performance of CO solver on various Top- " + }, + { + "bbox": [ + 144, + 293, + 468, + 304 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 144, + 293, + 468, + 304 + ], + "type": "text", + "content": " fragments selection task." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 161, + 314, + 449, + 388 + ], + "lines": [ + { + "bbox": [ + 161, + 314, + 449, + 388 + ], + "spans": [ + { + "bbox": [ + 161, + 314, + 449, + 388 + ], + "type": "table", + "html": "
Top-K2×23×3
Pool Size101520202530
Random Select0.40000.26670.20000.45000.36000.3000
Baseline0.74750.31250.25750.67560.57670.5389
OT Layer0.78000.58250.42250.72560.64440.6044
", + "image_path": "221f62da7b5b247341f26dc37bc05b0ace2f0656beb4ec46c7871b5f47c337a3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 413, + 480, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 480, + 450 + ], + "type": "text", + "content": "training dataset, reserving an additional 100 groups for evaluation. We trained the siamese feature extraction network with 40K, 80K and 120K positive and negative pairing samples, with the ratio of training and testing being 6:1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 455, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 482, + 552 + ], + "type": "text", + "content": "Siamese Feature Extraction Network. As detailed in Table 2, we evaluate the performance across various training data quantity, positive and negative sample ratios utilizing a ResNet backbone. Our findings indicate that CO solver performance peaks on Top-4 selections with 120K samples and 1:2 hybrid ratio on positive and negative matching samples. This can be attributed to the imbalance between positive and negative samples in pairwise matching manuscript fragments. With more negative samples, the reliability of siamese model is improved and CO solver performance is enhanced." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "CO Solver. Tab. 3 presents the simulation experiment results under various configurations of Top- " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " fragment selection from " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " candidates. \"Random select\" refers to the Top- " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " random selections mathematical expectation. \"Baseline\" refers to selecting the Top- " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " probabilities in a greedy manner from GNN output without OT layer. The introduction of the OT layer, which enforces cardinality constraints, is observed to enhance model performance, particularly as the candidate size " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " increases. This improvement is attributed to the OT layer's rigorous penalization of deviations from the optimization objectives in Eqs. (1a) and (1b)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 189, + 135, + 421, + 182 + ], + "blocks": [ + { + "bbox": [ + 176, + 114, + 436, + 125 + ], + "lines": [ + { + "bbox": [ + 176, + 114, + 436, + 125 + ], + "spans": [ + { + "bbox": [ + 176, + 114, + 436, + 125 + ], + "type": "text", + "content": "Table 4: Enhancement from the second-order Top- " + }, + { + "bbox": [ + 176, + 114, + 436, + 125 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 176, + 114, + 436, + 125 + ], + "type": "text", + "content": " selection." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 189, + 135, + 421, + 182 + ], + "lines": [ + { + "bbox": [ + 189, + 135, + 421, + 182 + ], + "spans": [ + { + "bbox": [ + 189, + 135, + 421, + 182 + ], + "type": "table", + "html": "
StrategyTop-9 of 20Top-9 of 25Top-9 of 30
Select at Once0.72560.64440.6044
Second-Order0.73110.68220.6633
", + "image_path": "3513813278b5b64bc0e1467824790a3b0c62164eb090244422c05c964986a56e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 206, + 481, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 206, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 206, + 481, + 314 + ], + "type": "text", + "content": "Enhancement through Second-order Matching Strategy. In the " + }, + { + "bbox": [ + 130, + 206, + 481, + 314 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 130, + 206, + 481, + 314 + ], + "type": "text", + "content": " selection settings, all four ground truth fragments hold equal significance (each fragment only have exactly 2 matching connections with their ground-truth adjacent pieces). Thus we explore the efficacy of the second-order matching strategy introduced in Section 3.2.5 for " + }, + { + "bbox": [ + 130, + 206, + 481, + 314 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 206, + 481, + 314 + ], + "type": "text", + "content": " scenarios and detail the results in Tab. 4. The second-order matching strategy duplicates the cardinality constraint and applies more stringent penalties for incorrect selections. This method aims to mitigate the effects of erroneous pairwise matches proposed by the siamese network, thereby advancing the accuracy and reliability of fragment selection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 315, + 481, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 481, + 350 + ], + "type": "text", + "content": "Simulation experiments demonstrate our proposed neural CO solver's effectiveness in selecting candidates based on highest mutual assemble similarity. This gives promise for the subsequent real-world restoration with LLMCO4MR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 372, + 416, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 372, + 416, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 372, + 416, + 384 + ], + "type": "text", + "content": "4.3 Real-world Manuscript Case Studies on Dunhuang" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 400, + 481, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 481, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 481, + 496 + ], + "type": "text", + "content": "Results with LLMCO4MR. We conducted case study experiments utilizing LLMCO4MR and compared its effectiveness against other SOTAs, as detailed in Tab. 5. We input a group of fragments with outliers to CO solver for selection, then use the selected Top- " + }, + { + "bbox": [ + 130, + 400, + 481, + 496 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 400, + 481, + 496 + ], + "type": "text", + "content": " fragments as image prompts for GPT-4V, which suggests the matching scores and relative directions. For SOTAs [31,51] that lack the capability to filter out outliers, we repeatedly input the fragments pairs and select the matching results in a greedy manner (select the fragment candidate pairs with ranked highest matching score)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 581 + ], + "type": "text", + "content": "We observed that previous works often operated under restricted conditions, such as assuming no outliers, limiting search spaces, regular shape fragments, or using manually outlined perfect matching contours. These constraints do not reflect the open-world setting addressed by our method. Moreover, contour similarity network does not achieve perfect accuracy in matching judgments and real-world scenarios frequently involve more irregular fragments, underscoring the complexity of the task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "Our pipeline addresses these challenges in two ways. First, the CO solver captures irregularity through data augmentation. We extract contours from real-world Dunhuang manuscripts and apply to shredding to adequately represent real-world challenges and mitigate the risk of overfitting unrealistic characteristics in training. Second, the MLLM component captures cross-image similarity using multimodal features, including contour (even with some imperfect fits and irregular shapes) and other visual features." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 90, + 447, + 102 + ], + "type": "text", + "content": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 175, + 135, + 436, + 264 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 470, + 126 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 470, + 126 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 470, + 126 + ], + "type": "text", + "content": "Table 5: Evaluation on real-world Dunhuang fragment restoration case studies." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 175, + 135, + 436, + 264 + ], + "lines": [ + { + "bbox": [ + 175, + 135, + 436, + 264 + ], + "spans": [ + { + "bbox": [ + 175, + 135, + 436, + 264 + ], + "type": "table", + "html": "
Top-K2×23×3
Pool Size10152025
GPT-4V [26]0.32500.17500.37780.2533
LLaVA [21]0.31500.12500.27780.1556
JigsawNet [19]0.52500.37500.45560.3222
Papyrus [31]0.42500.37500.47780.4111
S3-Net [51]0.31250.25750.38890.2111
CO Solver0.57500.52500.53330.4667
LLMCO4MR (Ours)0.67500.62500.62220.5556
", + "image_path": "5d80e979cff52688d877e48ac551b48d9a8b17c4d32235586a85a615c901ffe9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 287, + 482, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 482, + 395 + ], + "type": "text", + "content": "Ablation Study. We also present ablation study results in Tab. 5. \"CO Solver\" refers to deriving matching pairs directly from the output of CO solver. Few selected MLLMs exhibit capabilities in understanding cross-image fragment relations. These MLLMs' performances were evaluated by processing pairs of fragments without CO solver filtering out outliers. We conclude that similar to retrieval-augmented LLM frameworks, CO solver in our LLMCO4MR is a nontrivial extension, empowers MLLM capability handling combinatorial explosion problems. We detail in the Appendix the current limitations of MLLMs in directly addressing manuscript restoration tasks with combinatorial explosion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 396, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 456 + ], + "type": "text", + "content": "Our multifaceted approach serves two purposes: it complements and enhances traditional contour-based matching, and it helps to stabilize the variable performance of MLLMs in image matching. As a result, our method demonstrates increased effectiveness in handling real-world manuscript restoration tasks, especially in scenarios where either matching strategy alone is insufficient." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 479, + 220, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 220, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 220, + 491 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 508, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 605 + ], + "type": "text", + "content": "We have presented a novel pipeline, LLMCO4MR, designed for the intricate task of restoring ancient manuscripts from fragments. We first formalize the problem as a cardinality-constrained combinatorial problem and then a neural combinatorial solver is devised to produce the Top- " + }, + { + "bbox": [ + 130, + 508, + 482, + 605 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 508, + 482, + 605 + ], + "type": "text", + "content": " candidates for merging. Then the MLLM is employed, leveraging the candidates as prompt for final restoration. Experiments on both synthetic fragments and real-world Dunhuang data demonstrate the potential of our approach, compared with the traditional approaches which require labors of well-trained archaeologists." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "**Limitation:** Like existing works, our current approach also cannot handle the case with missing fragments and the fragments are required to be basically in a grid layout. Another unsolved case is the existence of both small and large fragments, which we leave for future work. Also, more case studies on broader scope need to be studied when such data becomes available." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 212 + ], + "type": "text", + "content": "This work was in part supported by National Science and Technology Major Project (2022ZD0119100), National Natural Science Foundation of China (No. 62441605, 72342023), National Social Science Fund of China (No. 21BYY142, 14AZS001), Key Research and Development Program of Zhejiang Province (No. 2024C03270), the StarryNight Science Fund of Zhejiang University Shanghai Institute for Advanced Study (SN-ZJU-SIAS-0010)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 228, + 197, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 228, + 197, + 240 + ], + "spans": [ + { + "bbox": [ + 133, + 228, + 197, + 240 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 251, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 251, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 285 + ], + "type": "text", + "content": "1. Abitbol, R., Shimshoni, I., Ben-Dov, J.: Machine learning based assembly of fragments of ancient papyrus. Journal on Computing and Cultural Heritage (JOCCH) 14(3), 1-21 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 285, + 480, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 480, + 327 + ], + "type": "text", + "content": "2. Assael, Y., Sommerschield, T., Shillingford, B., Bordbar, M., Pavlopoulos, J., Chatzipanagiotou, M., Androutsopoulos, I., Prag, J., de Freitas, N.: Restoring and attributing ancient texts using deep neural networks. Nature 603(7900), 280-283 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 328, + 480, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 328, + 480, + 393 + ], + "spans": [ + { + "bbox": [ + 138, + 328, + 480, + 393 + ], + "type": "text", + "content": "3. Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., Hui, B., Ji, L., Li, M., Lin, J., Lin, R., Liu, D., Liu, G., Lu, C., Lu, K., Ma, J., Men, R., Ren, X., Ren, X., Tan, C., Tan, S., Tu, J., Wang, P., Wang, S., Wang, W., Wu, S., Xu, B., Xu, J., Yang, A., Yang, H., Yang, J., Yang, S., Yao, Y., Yu, B., Yuan, H., Yuan, Z., Zhang, J., Zhang, X., Zhang, Y., Zhang, Z., Zhou, C., Zhou, J., Zhou, X., Zhu, T.: Qwen technical report (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 394, + 480, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 394, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 138, + 394, + 480, + 426 + ], + "type": "text", + "content": "4. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 426, + 480, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 480, + 459 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 480, + 459 + ], + "type": "text", + "content": "5. Bengio, Y., Lodi, A., Prouvost, A.: Machine learning for combinatorial optimization: a methodological tour d'horizon. European Journal of Operational Research 290(2), 405-421 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 459, + 480, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 480, + 480 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 480, + 480 + ], + "type": "text", + "content": "6. Besl, P.J., McKay, N.D.: Method for registration of 3-d shapes. In: Sensor fusion IV: control paradigms and data structures. vol. 1611, pp. 586-606. Spie (1992)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 481, + 480, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 481, + 480, + 513 + ], + "spans": [ + { + "bbox": [ + 138, + 481, + 480, + 513 + ], + "type": "text", + "content": "7. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 514, + 480, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 545 + ], + "type": "text", + "content": "8. Buchbinder, N., Feldman, M., Naor, J., Schwartz, R.: Submodular maximization with cardinality constraints. In: Proceedings of the twenty-fifth annual ACM-SIAM symposium on Discrete algorithms. pp. 1433-1452. SIAM (2014)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 546, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 480, + 567 + ], + "type": "text", + "content": "9. Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems 26 (2013)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 568, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 589 + ], + "type": "text", + "content": "0. Derech, N., Tal, A., Shimshoni, I.: Solving archaeological puzzles. Pattern Recognition 119, 108065 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 590, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 590, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 590, + 480, + 643 + ], + "type": "text", + "content": "11. Driess, D., Xia, F., Sajjadi, M.S.M., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., Huang, W., Chebotar, Y., Sermanet, P., Duckworth, D., Levine, S., Vanhoucke, V., Hausman, K., Toussaint, M., Greff, K., Zeng, A., Mordatch, I., Florence, P.: Palm-e: An embodied multimodal language model (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "text", + "content": "12. Galambos, I.: Dunhuang Manuscript Culture: End of the First Millennium, vol. 22. Walter de Gruyter GmbH & Co KG (2020)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "13. da Gama Leitao, H.C., Stolfi, J.: A multiscale method for the reassembly of two-dimensional fragmented objects. IEEE Transactions on Pattern Analysis and Machine Intelligence 24(9), 1239-1251 (2002)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "type": "text", + "content": "14. Ge, J., Luo, H., Qian, S., Gan, Y., Fu, J., Zhang, S.: Chain of thought prompt tuning in vision language models (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 192 + ], + "type": "text", + "content": "15. Guo, P.F., Chen, Y.H., Tsai, Y.D., Lin, S.D.: Towards optimizing with large language models. arXiv preprint arXiv:2310.05204 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 480, + 224 + ], + "type": "text", + "content": "16. Guo, Q., Wang, R., Guo, J., Li, B., Song, K., Tan, X., Liu, G., Bian, J., Yang, Y.: Connecting large language models with evolutionary algorithms yields powerful prompt optimizers. arXiv preprint arXiv:2309.08532 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 224, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 224, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 133, + 224, + 480, + 257 + ], + "type": "text", + "content": "17. Hossieni, S.S., Shabani, M.A., Irandoust, S., Furukawa, Y.: Puzzlefusion: Unleashing the power of diffusion models for spatial puzzle solving. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 257, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 257, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 480, + 289 + ], + "type": "text", + "content": "18. Karalias, N., Loukas, A.: Erdos goes neural: an unsupervised learning framework for combinatorial optimization on graphs. Advances in Neural Information Processing Systems 33, 6659-6672 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 289, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 289, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 133, + 289, + 480, + 322 + ], + "type": "text", + "content": "19. Le, C., Li, X.: Jigsawnet: Shredded image reassembly using convolutional neural network and loop-based composition. IEEE Transactions on Image Processing 28(8), 4000-4015 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 322, + 421, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 421, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 421, + 332 + ], + "type": "text", + "content": "20. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 332, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 354 + ], + "type": "text", + "content": "21. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "text", + "content": "22. Lu, J., Sun, Y., Huang, Q.: Jigsaw: Learning to assemble multiple fractured objects. arXiv preprint arXiv:2305.17975 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "text", + "content": "23. Lyu, C., Wu, M., Wang, L., Huang, X., Liu, B., Du, Z., Shi, S., Tu, Z.: Macaw-lm: Multi-modal language modeling with image, audio, video, and text integration (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 429 + ], + "type": "text", + "content": "24. Maaz, M., Rasheed, H., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 429, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 450 + ], + "type": "text", + "content": "25. Markaki, S., Panagiotakis, C.: Jigsaw puzzle solving techniques and applications: a survey. The Visual Computer pp. 1-17 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 450, + 307, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 307, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 307, + 460 + ], + "type": "text", + "content": "26. OpenAI: Gpt-4 technical report (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 460, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 480, + 482 + ], + "type": "text", + "content": "27. Panagiotakis, C., Markaki, S., Kokinou, E., Papadakis, H.: Coastline matching via a graph-based approach. Computational Geosciences 26(6), 1439-1448 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "type": "text", + "content": "28. Paumard, M.M., Picard, D., Tabia, H.: Jigsaw puzzle solving using local feature co-occurrences in deep neural networks. In: 2018 25th IEEE international conference on image processing (ICIP). pp. 1018-1022. IEEE (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 514, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 547 + ], + "type": "text", + "content": "29. Pengcheng, G., Gang, G., Jiangqin, W., Baogang, W.: Chinese calligraphic style representation for recognition. International Journal on Document Analysis and Recognition (IJDAR) 20, 59-68 (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "type": "text", + "content": "30. Pirrone, A., Aimar, M.B., Journet, N.: Papy-s-net: A siamese network to match papyrus fragments. In: Proceedings of the 5th International Workshop on Historical Document Imaging and Processing. pp. 78-83 (2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "type": "text", + "content": "31. Pirrone, A., Beurton-Aimar, M., Journet, N.: Self-supervised deep metric learning for ancient papyrus fragments retrieval. International Journal on Document Analysis and Recognition (IJDAR) 24(3), 219-234 (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "32. Rusinkiewicz, S., Levoy, M.: Efficient variants of the icp algorithm. In: Proceedings third international conference on 3-D digital imaging and modeling. pp. 145-152. IEEE (2001)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "text", + "content": "33. Savelka, J., Ashley, K.D., Gray, M.A., Westermann, H., Xu, H.: Explaining legal concepts with augmented large language models (gpt-4) (2023)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "34. Savino, P., Tonazzini, A.: Digital restoration of ancient color manuscripts from geometrically misaligned recto-verso pairs. Journal of Cultural Heritage 19, 511-521 (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "35. Shuai, R.W., Ruffolo, J.A., Gray, J.J.: Generative language modeling for antibody design. bioRxiv (2022). https://doi.org/10.1101/2021.12.13.472419, https://www.biorxiv.org/content/early/2022/12/20/2021.12.13.472419" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 184, + 482, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 184, + 482, + 250 + ], + "spans": [ + { + "bbox": [ + 130, + 184, + 482, + 250 + ], + "type": "text", + "content": "36. Singhal, K., Azizi, S., Tu, T., Mahdavi, S.S., Wei, J., Chung, H.W., Scales, N., Tanwani, A., Cole-Lewis, H., Pfohl, S., Payne, P., Seneviratne, M., Gamble, P., Kelly, C., Scharli, N., Chowdhery, A., Mansfield, P., y Arcas, B.A., Webster, D., Corrado, G.S., Matias, Y., Chou, K., Gottweis, J., Tomasev, N., Liu, Y., Rajkomar, A., Barral, J., Semturs, C., Karthikesalingam, A., Natarajan, V.: Large language models encode clinical knowledge (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 251, + 482, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 283 + ], + "type": "text", + "content": "37. Sommerscheid, T., Assael, Y., Pavlopoulos, J., Stefanak, V., Senior, A., Dyer, C., Bodel, J., Prag, J., Androutsopoulos, I., de Freitas, N.: Machine learning for ancient languages: A survey. Computational Linguistics pp. 1-44 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 285, + 482, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 317 + ], + "type": "text", + "content": "38. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 318, + 482, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 341 + ], + "type": "text", + "content": "39. Tsamoura, E., Pitas, I.: Automatic color based reassembly of fragmented images and paintings. IEEE Transactions on Image Processing 19(3), 680-690 (2009)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 341, + 482, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 363 + ], + "type": "text", + "content": "40. Vinyals, O., Fortunato, M., Jaitly, N.: Pointer networks. Advances in neural information processing systems 28 (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 363, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 407 + ], + "type": "text", + "content": "41. Wang, R., Shen, L., Chen, Y., Yang, X., Tao, D., Yan, J.: Towards one-shot neural combinatorial solvers: Theoretical and empirical notes on the cardinality-constrained case. In: The Eleventh International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 408, + 482, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 441 + ], + "type": "text", + "content": "42. Wang, R., Yan, J., Yang, X.: Learning combinatorial embedding networks for deep graph matching. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3056-3065 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 441, + 482, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 475 + ], + "type": "text", + "content": "43. Wang, R., Zhang, Y., Guo, Z., Chen, T., Yang, X., Yan, J.: Linsatnet: The positive linear satisfiability neural networks. In: International Conference on Machine Learning (ICML) (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 475, + 482, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 498 + ], + "type": "text", + "content": "44. Wolfson, H., Schonberg, E., Kalvin, A., Lamdan, Y.: Solving jigsaw puzzles by computer. Annals of Operations Research 12(1), 51-64 (1988)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 498, + 482, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 531 + ], + "type": "text", + "content": "45. Xi, Z., Chen, W., Guo, X., He, W., Ding, Y., Hong, B., Zhang, M., Wang, J., Jin, S., Zhou, E., et al.: The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 532, + 482, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 482, + 564 + ], + "type": "text", + "content": "46. Xie, Y., Dai, H., Chen, M., Dai, B., Zhao, T., Zha, H., Wei, W., Pfister, T.: Differentiable top-k with optimal transport. Advances in Neural Information Processing Systems 33, 20520-20531 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 565, + 482, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 565, + 482, + 642 + ], + "spans": [ + { + "bbox": [ + 130, + 565, + 482, + 642 + ], + "type": "text", + "content": "47. Yang, A., Xiao, B., Wang, B., Zhang, B., Bian, C., Yin, C., Lv, C., Pan, D., Wang, D., Yan, D., Yang, F., Deng, F., Wang, F., Liu, F., Ai, G., Dong, G., Zhao, H., Xu, H., Sun, H., Zhang, H., Liu, H., Ji, J., Xie, J., Dai, J., Fang, K., Su, L., Song, L., Liu, L., Ru, L., Ma, L., Wang, M., Liu, M., Lin, M., Nie, N., Guo, P., Sun, R., Zhang, T., Li, T., Li, T., Cheng, W., Chen, W., Zeng, X., Wang, X., Chen, X., Men, X., Yu, X., Pan, X., Shen, Y., Wang, Y., Li, Y., Jiang, Y., Gao, Y., Zhang, Y., Zhou, Z., Wu, Z.: Baichuan 2: Open large-scale language models (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "type": "text", + "content": "48. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of lmmss: Preliminary explorations with gpt-4v(ison) (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "LLMCO4MR: Manuscript Restoration with Case Studies on Dunhuang" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 380 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 148 + ], + "type": "text", + "content": "49. Yang, Z., Li, L., Wang, J., Lin, K., Azarnasab, E., Ahmed, F., Liu, Z., Liu, C., Zeng, M., Wang, L.: Mm-react: Prompting chatgpt for multimodal reasoning and action (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "text", + "content": "50. Yu, F., Quartey, L., Schilder, F.: Legal prompting: Teaching a language model to think like a lawyer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 215 + ], + "type": "text", + "content": "51. Zhang, C., Wang, B., Chen, K., Zong, R., Mo, B.f., Men, Y., Almpanidis, G., Chen, S., Zhang, X.: Data-driven oracle bone rejoining: A dataset and practical self-supervised learning scheme. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 4482-4492 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "type": "text", + "content": "52. Zhang, C., Zong, R., Cao, S., Men, Y., Mo, B.: Ai-powered oracle bone inscriptions recognition and fragments rejoining. In: Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence. pp. 5309-5311 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "53. Zhang, K., Li, X.: A graph-based optimization algorithm for fragmented image reassembly. Graphical Models 76(5), 484-495 (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 324 + ], + "type": "text", + "content": "54. Zhang, Y., Fang, Z., Yang, X., Zhang, S., He, B., Dou, H., Yan, J., Zhang, Y., Wu, F.: Reconnecting the broken civilization: Patchwork integration of fragments from ancient manuscripts. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 1157-1166 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 357 + ], + "type": "text", + "content": "55. Zhang, Z., Wang, Y.T., Li, B., Guo, A., Liu, C.L.: Deep rejoining model for oracle bone fragment image. In: Asian Conference on Pattern Recognition. pp. 3-15. Springer (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 380 + ], + "type": "text", + "content": "56. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models (2023)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "Y. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LLMGA_ Multimodal Large Language Model based Generation Assistant/d91671aa-dc99-4c79-b4c9-9d98440c12f4_content_list.json b/2024/LLMGA_ Multimodal Large Language Model based Generation Assistant/d91671aa-dc99-4c79-b4c9-9d98440c12f4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8122bdab56826cfc71d9eb0267b3884413c2db03 --- /dev/null +++ b/2024/LLMGA_ Multimodal Large Language Model based Generation Assistant/d91671aa-dc99-4c79-b4c9-9d98440c12f4_content_list.json @@ -0,0 +1,1932 @@ +[ + { + "type": "text", + "text": "LLMGA: Multimodal Large Language Model based Generation Assistant", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bin Xia $^{1}$ , Shiyin Wang $^{2}$ , Yingfan Tao $^{2}$ , Yitong Wang $^{2}$ , and Jiaya Jia $^{1}$", + "bbox": [ + 246, + 210, + 750, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The Chinese University of Hong Kong", + "bbox": [ + 364, + 238, + 635, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 ByteDance Inc", + "bbox": [ + 444, + 253, + 558, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://llmga.github.io/", + "bbox": [ + 419, + 268, + 581, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this paper, we introduce a Multimodal Large Language Model-based Generation Assistant (LLMGA), leveraging the vast reservoir of knowledge and proficiency in reasoning, comprehension, and response inherent in Large Language Models (LLMs) to assist users in image generation and editing. Diverging from existing approaches where Multimodal Large Language Models (MLLMs) generate fixed-size embeddings to control Stable Diffusion (SD), our LLMGA provides a detailed language generation prompt for precise control over SD. This not only augments LLM context understanding but also reduces noise in generation prompts, yields images with more intricate and precise content, and elevates the interpretability of the network. To this end, we curate a comprehensive dataset comprising prompt refinement, similar image generation, inpainting & outpainting, and instruction-based editing. Moreover, we propose a two-stage training scheme. In the first stage, we train the MLLM to grasp the properties of image generation and editing, enabling it to generate detailed prompts. In the second stage, we optimize SD to align with the MLLM's generation prompts. Additionally, we propose a reference-based restoration network to alleviate texture, brightness, and contrast disparities between generated and preserved regions during inpainting and outpainting. Extensive results show that LLMGA has promising generation and editing capabilities and can enable more flexible and expansive applications in an interactive manner.", + "bbox": [ + 261, + 315, + 738, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Interactive generation and editing $\\cdot$ LLM $\\cdot$ Diffusion Model", + "bbox": [ + 261, + 633, + 738, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 672, + 375, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Artificial Intelligence Generated Content (AIGC) has witnessed remarkable advancements, particularly propelled by the evolution of large language models (LLMs) [8, 10, 54] for text generation and diffusion models (DMs) [21, 42] for image generation. LLMs, in particular, have received considerable acclaim for their exceptional ability to comprehend, reason, make decisions, possess extensive knowledge, and generate text with unparalleled accuracy and fluency.", + "bbox": [ + 212, + 703, + 785, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent studies have begun delving deeper into Multimodal Large Language Models (MLLMs) [1,69] built upon LLMs, aiming to empower LLMs to comprehend inputs extending beyond text. For example, BLIP-2 [30] and LLaVA [33]", + "bbox": [ + 212, + 795, + 785, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9e911b04cc38e7b9e961d81ffec92e55e0cf1f64fdcd19496e0c4d3167263829.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 143, + 356, + 241 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bad945006d7a95d2b4fb1e61f9fc57c1b0f858a8f474ec7e9d2dd97684d30b66.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 356, + 143, + 491, + 241 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c6aa2e25d53add88442e917afb5fe46ff0d46775bfe210dc6fd635fe442f1036.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 143, + 663, + 241 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/edc64691018c7e8e081f9a9db223816f40e255a2683ff23a295fc36a8dccb3cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 143, + 807, + 247 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/51323782904c03d52c3a1a5ba51ec6489ec41152a06c08e989e8c100c6efe123.jpg", + "image_caption": [ + "Fig. 1: Some examples of LLMGA for assisting in image generation and editing. (1) T2I generation. LLMGA can refine the user's generation prompt to produce more vivid and vibrant images. (2) Similar image generation. LLMGA can understand the component and layout of the input images and generate a similar image. (3) Inpainting & Outpainting. LLMGA can provide detailed generation prompts based on user preferences and input images. (4) Instruction based editing. LLMGA can understand user instructions and realize accurate editing. (5) Interactive image generation and editing exemplify the comprehensive capabilities of LLMGA. Users can design satisfactory images by engaging in interactions with LLMGA, leveraging its vast knowledge." + ], + "image_footnote": [], + "bbox": [ + 218, + 243, + 356, + 348 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e9325f1b8254c5a5da9fa606c8b65118ca571514c3bc01c3b4b329e299a7ace7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 356, + 243, + 491, + 348 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7196089690010dc3ce9acb2ce6d66a849548ef96d897bc1d86b8455f88583739.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 243, + 660, + 348 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ee491b89156c24e2a42b8f0e84bc42a93043ee3b1525a3d9c8020413879d1880.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 660, + 243, + 807, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "employ visual encoders to transform images into input embeddings, enabling them to be used as prompts alongside text input for the LLM, thus achieving LLMs with the visual modality. Furthermore, recent works focused on extending the capabilities of LLMs to generate multimodal outputs. For example, GILL [29] involves instructing LLMs to predict fixed-size visual embeddings aligned with CLIP [40] space to control the Stable Diffusion [42] (SD) for image generation.", + "bbox": [ + 212, + 520, + 787, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, existing works [29, 52] merely focus on enabling LLM to output images but do not aim to assist users in generating or editing images to enhance quality. In this paper, we aim to develop a Multimodal Large Language Model-based Generation Assistant (LLMGA) to better assist image generation models, making them more user-friendly and capable of producing high-quality images. In contrast to certain methods [29, 52] that leverage MLLMs to predict fixed-size visual embeddings for implicit SD control, our approach is straightforward. We guide the generation of SD using detailed language prompts from MLLM based on five reasons. (1) The embeddings predicted by the MLLM are often filled with noise. This can be filtered out by mapping them to a fixed language domain, enabling precise control of SD. (2) Detailed language prompts can make the network more transparent and interactive, allowing users to understand MLLM's thoughts for generating images. (3) MLLM is pre-trained on vast textual datasets. Explicit language prompts rather than implicit embeddings are more advantageous for MLLM to generate prompts and comprehend context. (4)", + "bbox": [ + 212, + 613, + 787, + 843 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia", + "bbox": [ + 274, + 114, + 709, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic-sized language prompt facilitates the addition of generation requests during interactions. (5) Training is more simple and efficient.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, we face several challenges: (1) MLLM may reject the execution of generation instructions due to its nature as a language assistant. (2) MLLM lacks a comprehensive understanding of image generation and editing, and cannot provide an accurate and detailed generation prompt. (3) Determining which part of texts generated by MLLM to guide SD generation. (4) SD's CLIP encode only 75 tokens. Additionally, SD is trained on short captions, whereas our LLMGA typically generates detailed prompts exceeding 150 tokens. This discrepancy poses a challenge for SD in following the detailed prompt of LLMGA.", + "bbox": [ + 212, + 178, + 785, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we have devised a two-stage training scheme. First, we construct a training dataset: prompt refinement, similar image generation, inpainting & outpainting, and visual question answering. We then train LLMGA on these four datasets to cultivate four fundamental capabilities: (1) For concise user prompts, LLMGA can refine the generation of intricate details, encompassing attire, background, and characters. (2) LLMGA can precisely regenerate an image it observes. (3) LLMGA can generate or refine prompts for inpainting & outpainting based on its understanding of the image. (4) LLMGA can generate accurate prompts for instruction-based editing according to users' requirements and given images. Additionally, we make LLMGA use special symbols and to distinguish generation prompts and responses. In the second stage, we freeze the parameters of LLMGA's MLLM and initiate joint training with the SD. This process enables the SD to acclimate to the detailed prompt produced by the MLLM. Notably, when the input token count exceeds 75, we iteratively apply the CLIP [40] encoder to the surplus tokens.", + "bbox": [ + 212, + 301, + 787, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, we have identified noticeable disparities in texture, contrast, and brightness between the newly generated and preserved sections in inpainting & outpainting. Therefore, we propose a Diffusion-based Reference Restoration Network (DiffRIR). Specifically, aside from images generated by SD, we add masked images as reference inputs into DiffRIR. This enables the DiffRIR to refer to the texture, contrast, and brightness of the preserved regions for restoration. Additionally, we introduce perturbations to contrast and brightness during training, enabling DiffRIR to correct contrast and brightness disparities in the images.", + "bbox": [ + 212, + 531, + 787, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 1, LLMGA is a unified and interactive framework for image generation and editing, endowed with a wide array of capabilities: (1) LLMGA leverages its extensive world knowledge and powerful reasoning abilities to assist image generation and editing and significantly improve results. (2) LLMGA can be integrated with external plugins, like ControlNet. (3) Most importantly, users can interact with LLMGA to design satisfying images in a more convenient, flexible, and enjoyable way. In summary, our contributions are as follows:", + "bbox": [ + 212, + 654, + 787, + 761 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "LLMGA: Multimodal Large Language Model based Generation Assistant", + "bbox": [ + 243, + 113, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "- We proposed LLMGA, a simple yet powerful interactive generation and editing framework. Experiments affirm the efficacy of LLMGA in enhancing generation and editing thanks to its vast knowledge and interactive features. Plus, LLMGA can integrate with external plugins for wider applications.", + "bbox": [ + 225, + 779, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We construct a training dataset, including four parts: prompt refinement, similar image generation, inpainting & outpainting, and instruction-based editing. This enhances LLMGA's comprehension of generation and editing tasks while standardizing response formats.", + "- We proposed a restoration network DiffRIR, which introduces reference images and training perturbations to contrast and brightness. DiffRIR can alleviate texture, contrast, and brightness discrepancies between newly generated and preserved regions for edited images.", + "- Open-source. The following assets are released: the generated data, the codebase for model training, the model checkpoint, and a demo." + ], + "bbox": [ + 223, + 146, + 784, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 313, + 387, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion Model. Diffusion Models (DMs) [2,4,7,13,19,21,22,28,35,38,49,51] have achieved remarkable results in image generation. DMs adopt a parameterized Markov chain to optimize the lower variational bound on the likelihood function. In this way, it can generate realistic images from Gaussian noise. After that, several DM methods [3,5,6,12,16,17,26,27,34,39,44,45,55,59,60,62] have been tailored to enhance the text-to-image (T2I) generation and editing. Notably, GLIDE [36] pioneered the incorporation of text features into transformer blocks during the denoising process. Subsequently, DALL-E [41], Imagen [46], and Stable Diffusion [42] have made substantial strides in improving T2I generation. Subsequently, some works [43,66,67] introduced conditioning controls to the DMs to facilitate a more convenient and precise manipulation of the generation process. Overall, enhancing the user-friendliness of DMs is a key focus within the community. In this paper, we introduce LLMGA, leveraging the extensive knowledge and powerful reasoning capabilities of LLM to facilitate users in achieving more easily attainable and satisfactory image designs.", + "bbox": [ + 212, + 340, + 784, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multimodal Large Language Models. Recently, LLMs have undeniably made profound impacts and revolutions within the entire AI community and beyond. For example, exemplary LLMs, such as ChatGPT and GPT4 [37], have showcased remarkable abilities in comprehension, reasoning, responses, and knowledge reservoirs. Subsequently, a range of LLMs [11], including Vicuna [9], LLaMA [54], and Alpaca [53] have been released as open-source models, substantially propelling advancements of the community.", + "bbox": [ + 212, + 568, + 784, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Afterward, the community began focusing on the development of the Multimodal Large Language Model [14, 15, 18, 23, 63, 64, 69]. They aim to enable LLMs to comprehend both images and text and provide textual responses. For instance, Flamingo [1] encodes images and feeds them into the LLM's attention layer. BLIP-2 [30] employs Q-Former to encode input images into queries. Additionally, LLaVA [33] leverages CLIP [40] to encode images into visual embeddings, which are then concatenated with text embeddings.", + "bbox": [ + 212, + 674, + 784, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent concurrent works, such as Next-GPT [58], have extended the capabilities to encompass audio and video modalities. Moreover, Visual-ChatGPT [57] and HuggingGPT [48] make LLMs act as agents capable of invoking various pretrained visual models to achieve MLLM. However, these works focus on making", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia", + "bbox": [ + 274, + 113, + 707, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ac16f1c191015e38321bbbc44c46d40be6e0f8fa0223e8990c44fa1ded187dd5.jpg", + "image_caption": [ + "Fig. 2: The overview of LLMGA. (a) In the first training stage, we train the MLLM to produce generation prompts based on provided instructions. Moreover, we construct a training dataset including four categories: prompt refinement, similar image generation, inpainting & outpainting, and instruction-based editing. (b) In the second training stage, we optimize SD to adapt to the detailed generation prompts from MLLM. (c) In the inference stage, LLMGA can respond to user queries and assist in various tasks, such as image generation, inpainting & outpainting, and editing." + ], + "image_footnote": [], + "bbox": [ + 225, + 151, + 555, + 338 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/89b4e111b668e96c8e2ed4ec0e9287c59b03a9f2210e5350cc4d65c448bde191.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 151, + 779, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LLM determine the combined invocation of modules (such as detection, recognition, and generation) to fulfill user requirements. However, these methods are not tailored for generation and editing and use concise prompts that lack the capability to enhance results. Thus, we propose LLMGA, which is designed to assist with various image generation and editing tasks. It can achieve satisfactory results by strong reasoning capability and flexible interaction with users.", + "bbox": [ + 212, + 479, + 787, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 590, + 380, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overview of LLMGA", + "text_level": 1, + "bbox": [ + 215, + 619, + 434, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we aim to design a MLLM-based Generation Assistant (LLMGA). Our LLMGA produces detailed language-based generation prompts to control SD rather than predicting fixed-sized visual embeddings [29, 52] to govern SD. This has five advantages: (1) Visual embeddings contain noise, and mapping them to the language domain can filter out this noise, enabling precise SD control. (2) Language-based generation prompts facilitate users in understanding the LLMGA's thoughts, enhancing interaction. (3) Dynamic-sized language-based generation prompt enables the addition of generation requests. (4) MLLM is pre-trained on textual datasets. Language prompts rather than implicit visual embeddings are more advantageous for MLLM to generate accurate prompts and comprehend context. (5) Training is simpler and more efficient.", + "bbox": [ + 212, + 643, + 787, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, we need to address several issues: (1) As a language assistant, MLLM may decline the execution of generation instructions. (2) MLLM lacks", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "LLMGA: Multimodal Large Language Model based Generation Assistant", + "bbox": [ + 243, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a nuanced understanding of image generation and editing, and cannot produce precise and detailed generation prompts. (3) MLLM needs to decide which part of the output text serves as generation prompts to guide generation. (4) SD's CLIP encode only 75 tokens. Moreover, SD is primarily trained on short captions, while detailed prompts generated by LLMGA may exceed 150 tokens. This disparity makes it hard for SD to understand the instructions from LLMGA. To address the aforementioned challenges, we construct a training dataset and design two-stage training schemes to train the MLLM (Sec. 3.2) and SD (Sec. 3.3).", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The network structure and pipeline of LLMGA are illustrated in Fig. 2. Specifically, as shown in Fig. 2 (a) and (c), the images $\\mathbf{I}_{input}$ are encoded into image embeddings by CLIP vision encoder and a projection layer. Subsequently, the image embedding is concatenated with the text embedding and fed into the LLM to obtain text output $\\mathbf{T}_{output}$ . This process can be formulated as:", + "bbox": [ + 212, + 267, + 787, + 343 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {\\text {o u t p u t}} = \\mathcal {F} _ {M L L M} \\left(\\mathbf {T} _ {\\text {i n p u t}}, \\mathbf {I} _ {\\text {i n p u t}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 358, + 784, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{T}_{input}$ indicates the input text instructions from users. It is notable that $\\mathcal{F}_{MLLM}$ can process only $\\mathbf{T}_{input}$ as input.", + "bbox": [ + 212, + 378, + 785, + 410 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The text output $\\mathbf{T}_{output}$ can comprise two components: text response $\\mathbf{T}_R$ and generation prompt $\\mathbf{T}_P$ . To distinguish between $\\mathbf{T}_R$ and $\\mathbf{T}_P$ , we adopt new special tokens, i.e., $<\\text{gen\\_img}>$ and $$ , to encompass $\\mathbf{T}_P$ .", + "bbox": [ + 212, + 410, + 785, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present $\\mathbf{T}_R$ as the immediate text response to users. Concurrently, $\\mathbf{T}_P$ is further fed into the subsequent SD to guide T2I generation (Eq. 2), inpainting & outpainting (Eq. 3), and instruction-based image editing (Eq. 4).", + "bbox": [ + 212, + 455, + 785, + 501 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {g} = \\mathcal {F} _ {T 2 I} (\\mathbf {T} _ {P}, \\mathbf {Z}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 512, + 784, + 527 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {p} = \\mathcal {F} _ {\\text {P a i n t}} (\\mathbf {I} _ {\\text {i n p u t}}, \\mathbf {I} _ {\\text {m a s k}}, \\mathbf {T} _ {P}, \\mathbf {Z}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 554, + 784, + 570 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {e} = \\mathcal {F} _ {\\text {E d i t}} \\left(\\mathbf {I} _ {\\text {i n p u t}}, \\mathbf {T} _ {P}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 590, + 784, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where $\\mathbf{Z}$ denotes the random Gaussian noise. Furthermore, to ensure the encoding of all $\\mathbf{T}_P$ for SD, we iteratively run the CLIP text encoder until all prompts are encoded. For inpainting & outpainting, except the input image $\\mathbf{I}_{input}$ , an additional mask $\\mathbf{I}_{mask}$ are essential inputs for the inpainting SD to specify the region requiring generation. For instruction-based image editing, $\\mathcal{F}_{Edit}$ performs inversion [25, 50] and prompt-to-prompt [20] based on T2I SD.", + "bbox": [ + 212, + 612, + 787, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 MLLM Training", + "text_level": 1, + "bbox": [ + 214, + 724, + 398, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As described in Sec. 3.1, original MLLMs are specifically designed and trained as language assistants, but they lack the proficiency to assist in image generation, and editing. Notably, considering the input of images and the performance of open-sourced MLLMs, employing few-shot learning to guide the model to achieve the desired results proves challenging and inefficient. Therefore, it is crucial to train the MLLM to serve as proficient assistants in image generation and editing", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia", + "bbox": [ + 274, + 113, + 709, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tasks, understanding the expected response formats and enhancing their comprehension of image generation and editing properties. To this end, as depicted in Fig. 2 (a), we construct a training dataset consisting of four categories:", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Prompt Refinement. We establish this dataset to cultivate the prompt refinement ability of the MLLM. Specifically, we utilize GPT4-V to furnish detailed descriptions of images in MSCOCO [31]. These detailed descriptions, along with the original MSCOCO brief descriptions, constitute a training text pair. During training, we input the brief MSCOCO captions and randomly select and append a generation instruction. When a generation instruction is included in the prompt, we add and on the later description.", + "(2) Similar Image Generation. We select images from the MSCOCO dataset along with corresponding detailed descriptions generated by GPT4-V to create the Similar Image Generation dataset. During training, we input the images along with a generation instruction. In cases where a generation instruction is provided, we add and on the subsequent description.", + "(3) Inpainting & Outpainting. We use pairs of detailed descriptions and images from the Similar Image Generation dataset. During training, we input masked images with inpainting or outpainting instructions. Besides, we include and on the subsequent description.", + "(4) Instruction-based editing. We fine-tune Mixtral-8x7B [24] to enable it to generate editing data based on detailed descriptions from MSCOCO. Subsequently, we clean the generated data. During training, we can input images or original captions and provide corresponding editing instructions, aiming to train LLMGA to output the target caption. Additionally, we include and tags on the target caption during training." + ], + "bbox": [ + 212, + 191, + 785, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During training, alongside the above data, we integrated the image designing and Alpaca [53] dataset, to enhance LLMGA's question-answering (QA). Notably, we excluded certain visual multimodal incompatible question-answer pairs from it. Furthermore, we incorporated the LLaVA v1.5 mix665k dataset [32] to endow LLMGA with Visual Question Answering (VQA) capabilities. We provide more details in the supplementary material.", + "bbox": [ + 212, + 523, + 785, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2 (a), we freeze the CLIP vision encoder and optimize the projection layer and LLM. The model is trained end-to-end using the autoregressive cross-entropy loss $(\\mathcal{L}_{MLLM})$ for text generation. Given the ground-truth targets $\\mathbf{T}_{GT}$ , this loss can be formulated as:", + "bbox": [ + 212, + 614, + 785, + 675 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M L L M} = \\mathbf {C E} \\left(\\mathbf {T} _ {\\text {o u t p u t}}, \\mathbf {T} _ {G T}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 686, + 785, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Stable Diffusion Training", + "text_level": 1, + "bbox": [ + 214, + 724, + 470, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As described in Sec. 3.1, the original SD's CLIP text encoder only encodes 75 tokens, which cannot handle the entire MLLM's generation prompt. Moreover, the original SD is trained on brief captions, which cannot fully understand the generation prompts. Thus, we repeatedly use the CLIP text encoder to encode all instances of $\\mathbf{T}_P$ for SD. Besides, we train the T2I SD model and the inpainting SD model, respectively. For both generation and inpainting & outpainting tasks,", + "bbox": [ + 212, + 750, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "LLMGA: Multimodal Large Language Model based Generation Assistant", + "bbox": [ + 243, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the generation prompts of MLLM are detailed descriptions of images. Therefore, during training, we can instruct MLLM to provide a detailed description $\\mathbf{T}_P$ for images from the LAION-Aesthetics [47] and MSCOCO datasets. Subsequently, $\\mathbf{T}_P$ is fed into T2I SD or inpainting SD for joint training. Notably, we only optimize the SD unet while freezing the parameters of other networks. To accelerate the training, we record the prompt $\\mathbf{T}_P$ of MLLM to avoid redundant calculations. The model is trained using SD loss (Eq. 6). For instruction-based editing, we directly adopt the pre-trained T2I SD model.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S D} = \\mathbb {E} _ {\\mathbf {Z} _ {t}, \\mathbf {C}, \\epsilon , t} \\left(\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\mathbf {Z} _ {t}, \\mathbf {C}\\right) \\| _ {2} ^ {2}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 279, + 785, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{Z}_t = \\sqrt{\\overline{\\alpha}_t}\\mathbf{Z}_0 + \\sqrt{1 - \\overline{\\alpha}_t}\\epsilon$ represents the noised feature map at timestep $t$ . Ground truth images are encoded into latent space to derive $\\mathbf{Z}_0$ . Here, $\\epsilon \\in \\mathcal{N}(0,\\mathbf{I})$ represents Gaussian noise, and $\\epsilon_{\\theta}$ refers to the SD unet. $\\mathbf{C}$ indicates the conditional information. For T2I generation, $\\mathbf{C}$ is $\\mathbf{T}_P$ . For inpainting and outpainting, $\\mathbf{C}$ contains $\\mathbf{T}_P$ , the mask, and the VAE-encoded masked image.", + "bbox": [ + 212, + 314, + 787, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Restoration Network Training", + "text_level": 1, + "bbox": [ + 214, + 412, + 511, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For SD inpainting & outpainting, we observed noticeable disparities between the preserved and newly generated regions in the edited images. To enhance the consistency between the newly generated and the preserved regions, we introduce a reference-based restoration scheme. Specifically, existing restoration methods [56, 60, 61] take low-quality (LQ) images as input and produce high-quality (HQ) images, but they often do not leverage the preserved information from the given masked image. Different from them, we concatenate the LQ image $I_{LQ}$ and the masked image, i.e., $(1 - \\mathbf{I}_{mask})\\mathbf{I}_{GT}$ , as inputs and feed them into the restoration model $\\mathcal{F}_R$ . This process can be formulated as:", + "bbox": [ + 212, + 436, + 787, + 571 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {H Q} = \\mathcal {F} _ {R} (\\operatorname {c o n c a t} \\left(\\mathbf {I} _ {L Q}, (1 - \\mathbf {I} _ {\\text {m a s k}}) \\mathbf {I} _ {G T}\\right)). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 585, + 785, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further mitigate the brightness and contrast disparities, we introduced additional color degradation (i.e., random brightness and contrast disturbance) into the training process of the restoration model, which is formulated as:", + "bbox": [ + 212, + 612, + 785, + 657 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {2} (\\mathbf {x}) = c _ {1} \\mathbf {x} + c _ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 671, + 785, + 685 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {L Q} = \\mathcal {D} _ {2} \\left(\\mathcal {D} _ {1} \\left(\\mathbf {I} _ {G T}\\right)\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 428, + 712, + 785, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_2$ indicates contrast and brightness disturbance, $\\mathbf{x}$ denotes the input image. $c_{1}$ is the contrast gain, randomly varied within the range of [0.94, 1.06], while $c_{2}$ is the brightness bias, randomly varied within the range of $[-0.05, 0.05]$ . $\\mathcal{D}_1$ represents the real-world degradation process used in restoration model [56, 60, 61]. $\\mathbf{I}_{GT}$ is ground truth image, and $\\mathbf{I}_{LQ}$ is LQ image. Here, we adopt the network structure of the SOTA restoration model DiffIR [60] and apply our schemes to it to obtain our Diffusion-based Reference Restoration Network (DiffRIR).", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia", + "bbox": [ + 274, + 113, + 709, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 176, + 457, + 193 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For the first stage of training, we train the MLLM (including projection and LLM) on $8 \\times \\mathrm{A}100$ . The batch size is set to 128. Besides, the training datasets include VQA (LLaVA v1.5 mix665k), QA, prompt refinement, similar image generation, inpainting & outpainting, and instruction-based editing.", + "bbox": [ + 212, + 202, + 782, + 262 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For the second stage of training, we adopt the Stable Diffusion 1.5 (SD1.5) as the initial image generation or inpainting & outpainting model. We train these models on $8 \\times \\mathrm{A}100$ . The batch size is set to 32.", + "bbox": [ + 212, + 262, + 784, + 306 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For the restoration network, we train DiffRIR on $8 \\times \\mathrm{V}100$ . The batch size is set to 64. Please check more details in the supplementary material.", + "bbox": [ + 212, + 306, + 784, + 338 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 361, + 441, + 376 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation on T2I Generation. The results are shown in Tab. 1. Notably, SD1.5 is the original Stable Diffusion 1.5 while SD1.5-ft indicates the finetuned SD1.5 in LLMGA. We also compare our LLMGA with the recently proposed multimodal generative model GILL [29]. (1) Comparing SD1.5, our LLMGA-7B achieves notable 5.4847 FID and 5.05 IS improvements, underscoring the effectiveness of LLMGA. (2) Moreover,", + "bbox": [ + 215, + 386, + 486, + 568 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "our LLMGA-7B significantly outperforms GILL. (3) In the 3rd and 4th rows of Tab. 1, we use LLMGA-7B to refine the short MSCOCO caption to the detailed generation prompt and send it to the SD1.5 and SD1.5-ft, respectively. Our LLMGA-7B with SD1.5-ft achieves significant 5.0712 FID and 3.92 IS improvements, respectively. This demonstrates our second-stage training can help SD1.5-ft to better follow detailed prompts from MLLM. (4) Comparing the 4th and 6th rows of Tab. 1, LLMGA-13B exhibits better performance than LLMGA-7B due to its superior reasoning ability.", + "bbox": [ + 212, + 568, + 784, + 688 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The qualitative results are shown in Fig. 3. (1) LLMGA excels in refining prompts by incorporating details to generate visually rich and pleasing images. For example, in the first row, LLMGA crafts a battle attire for the husky, depicting engaging scenarios of battling monsters. This makes user usage more convenient, eliminating the need for them to think about generating image details themselves. (2) LLMGA can leverage its extensive knowledge base to generate images, even for concepts users may not be familiar with, like a space elevator.", + "bbox": [ + 212, + 688, + 785, + 792 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation on Instruction-based Editing. For instruction-based editing, we utilize LLMGA to provide a detailed description of the edited image based on the input image and user editing instructions. We then employ Direct Inversion [25]", + "bbox": [ + 212, + 794, + 785, + 842 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/a48c202f5cd6d9b0c40536c66491d58b03991e17bc0e0797298d320e1bad46c7.jpg", + "table_caption": [ + "Table 1: Quantitative comparison on T2I generation on the MSCOCO [31] dataset." + ], + "table_footnote": [], + "table_body": "
MethodFID↓IS↑
SD1.524.008135.99
GILL [29]25.112334.20
LLMGA-7b(SD1.5)23.594637.12
LLMGA-7b(SD1.5-ft, Ours)18.523441.04
LLMGA-13b(SD1.5)23.582837.58
LLMGA-13b(SD1.5-ft, Ours)18.406341.16
", + "bbox": [ + 501, + 416, + 787, + 540 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "LLMGA: Multimodal Large Language Model based Generation Assistant", + "bbox": [ + 243, + 113, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6c5ca517e99e665420423b72be44130bd6d43d781b60bd96ed1459b91812afc6.jpg", + "image_caption": [ + "Fig. 3: T2I visual comparison. LLMGA can produce accurate and high-quality results." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 772, + 362 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/aae02e8db1f0fb0aca33f00dbd56354f0d3d4493609cedd1659aa01f246d472b.jpg", + "table_caption": [ + "Table 2: Quantitative comparison for image editing on MagicBrush test set [65]." + ], + "table_footnote": [], + "table_body": "
MethodsL1↓L2↓CLIP-I↑DINO↑CLIP-T↑
InstructPix2Pix0.11970.04160.84420.72520.2909
MagicBrush0.06470.02240.92930.89130.2979
LLMGA (Ours)0.08140.02180.89360.87680.3137
", + "bbox": [ + 305, + 433, + 697, + 508 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "and prompt-to-prompt [20] methods to obtain the edited image. (1) The quantitative results are presented in Tab. 2. Notably, our LLMGA did not train SD for image editing like InstructPix2Pix [7] and MagicBrush [65]. However, our zero-shot performance on the MagicBrush test set surpassed that of InstructPix2Pix, achieving performance similar to that of MagicBrush. (2) Additionally, our LLMGA offers a superior user experience in interactive editing, allowing image modifications to be carried out conversationally. In contrast, InstructPix2Pix only supports input via instructions and output as images. Visualized results are shown in Fig. 4, by leveraging the powerful reasoning capabilities of LLM, our LLMGA can provide more accurate and reasonable editing results.", + "bbox": [ + 212, + 537, + 787, + 688 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation on Inpainting and Outpainting. The results are shown in Tab. 3. (1) Comparing the 1st and 3rd rows of Tab. 3, our LLMGA-7B achieves significant improvements of 2.5681 FID over the SD1.5 in outpainting under wide masks. (2) In the 2nd and 3rd rows of Tab. 3, we make LLMGA imagine the complete generation prompts for the given masked images, which are then input into the later SD. Notably, our LLMGA-7B (with SD1.5-ft) demonstrates a significant FID improvement over LLMGA-7B (with SD1.5) in both outpainting and inpainting. This demonstrates the second stage of training makes SD better follow the prompts from MLLM. (3) Comparing the 3rd and 5th rows, our LLMGA-13B outperforms LLMGA-7B due to its superior reasoning capabilities.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia", + "bbox": [ + 274, + 113, + 710, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/48c3f14583c3f198ad34333c853599803b868249c6aa9821c67403fb3f3e8411.jpg", + "image_caption": [ + "Fig. 4: Visual comparison on instruction-based editing." + ], + "image_footnote": [], + "bbox": [ + 243, + 148, + 764, + 392 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a9bba7404f34e57eec5f9063c44aac7f517e89e5e69d473791037185d2c836cc.jpg", + "table_caption": [ + "Table 3: Quantitative comparison for outpainting & inpainting on Places [68]." + ], + "table_footnote": [], + "table_body": "
MethodOutpaintingInpainting
NarrowMasksWideMasksNarrowMasksWideMasks
FID↓LPIPS↓FID↓LPIPS↓FID↓LPIPS↓FID↓LPIPS↓
SD1.52.01670.22835.00900.37341.07950.12361.28550.1434
LLMGA-7b (SD1.5)1.55300.22633.26300.36881.06920.12331.09830.1427
LLMGA-7b (SD1.5-ft, Ours)1.29730.22152.44090.36160.80270.11710.98070.1405
LLMGA-13b (SD1.5)1.56310.22633.08450.36791.03260.12281.09780.1426
LLMGA-13b (SD1.5-ft, Ours)1.21600.22102.36630.36090.79920.11660.97800.1400
", + "bbox": [ + 220, + 470, + 787, + 602 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The qualitative results are shown in Fig. 5. We can see that LLMGA can deduce and imagine complete images based on masked input images. For example, in the 3rd row of Fig. 5, LLMGA can infer the presence of wind turbines on the mountain based on the given environment. Overall, LLMGA's powerful reasoning capability and extensive knowledge can assist users in conveniently making accurate and visually pleasing inpainting and outpainting.", + "bbox": [ + 212, + 638, + 787, + 731 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation on ControlNet. Our LLMGA demonstrates exceptional scalability, enabling integration with external plugins like ControlNet [66]. Here, we utilize LLMGA to create detailed prompts derived from input images and user requirements, working alongside ControlNet to guide image generation. As depicted in Fig. 6, our LLMGA significantly enhances the diversity and richness of outcomes in picture reference-guided image generation. Furthermore, more external plugins can also be integrated into the interactive framework of LLMGA,", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "LLMGA: Multimodal Large Language Model based Generation Assistant", + "bbox": [ + 243, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f61f8cd5dc4c567f736bd55e7b552c52df09cc2f898d5f77908e49f71cfc9897.jpg", + "image_caption": [ + "Fig. 5: Visual comparison on inpainting and outpainting." + ], + "image_footnote": [], + "bbox": [ + 232, + 143, + 772, + 366 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/39cbd39d76c37d0e07e7cebf169c819561ff40536a55f56fa04e2d503d2f016f.jpg", + "image_caption": [ + "Fig. 6: Visualization of LLMGA plus ControlNet. Our LLMGA can enhance the details in generated images, producing visually pleasing images." + ], + "image_footnote": [], + "bbox": [ + 232, + 412, + 772, + 577 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "combining with LLMGA's reasoning design capabilities and all previous functions to allow for a broader range of creative and engaging applications.", + "bbox": [ + 212, + 651, + 784, + 681 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Evaluation on Image Restoration. The results are shown in Tab. 4. For comparisons, we validate DiffIR [60] and our DiffRIR on the outpainting image generated by LLMGA. (1) Comparing the 2nd and 3rd rows of Tab. 4, it is evident that introducing a reference scheme can significantly improve restoration performance. (2) When comparing the 3rd and 4th rows of Tab. 4, it can be observed that introducing color degradation helps alleviate the bright and contrast distortion caused by SD. (3) Comparing the 1st and 4th rows of Tab. 4, our DiffRIR yields significant improvement, validating the effectiveness of DiffRIR.", + "bbox": [ + 212, + 684, + 787, + 805 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "As shown in Fig. 7, our DiffRIR (i.e., $\\mathrm{DiffRIR}_2$ in Tab. 4) can alleviate the texture, brightness, and contrast discrepancies, and generate realistic details.", + "bbox": [ + 212, + 806, + 787, + 838 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia", + "bbox": [ + 274, + 113, + 709, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/16b2f5cca6a5cf80ccae3ac84dc00423609833342c7a1e6222cc2cfffdfbe0b9.jpg", + "table_caption": [ + "Table 4: Quantitative comparison on image restoration for Places [68] outpainting." + ], + "table_footnote": [], + "table_body": "
MethodColorFID↓LPIPS↓
Reference Degradation
LLMGA-7BXX2.44090.3616
LLMGA+DiffIR [60]XX2.35870.3612
LLMGA+DiffRIR1✓X2.29930.3609
LLMGA+DiffRIR2 (Ours)✓✓2.26870.3607
", + "bbox": [ + 303, + 171, + 699, + 260 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d61d967b941e41a79902a688b01a89953ded62a49980f7f9c6a0d288fa24ad7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 232, + 280, + 318, + 349 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1eb7e259f0810d9986e8570445abe9d6da67065a241ea358e8be52203e35ebab.jpg", + "image_caption": [ + "LLMGA" + ], + "image_footnote": [], + "bbox": [ + 320, + 281, + 410, + 349 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c9ad9fd42e21199e45b8b081645cd824c84eb78466adb47ada4112f7db65db87.jpg", + "image_caption": [ + "Masked Image" + ], + "image_footnote": [], + "bbox": [ + 411, + 281, + 500, + 349 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ff9e7e78133ad38a7383ed810676b6498be9903d3e35ac82c9217cc432aa7f2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 281, + 589, + 349 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/75a3a4b1c266273737fa5c1332f15c326d5ada07257358fbe6372715fd2da9c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 281, + 679, + 349 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9c53a09691581ae00e8ddc042ab7d668b9c619b726ca9dbe8057f83c9e108b42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 281, + 771, + 349 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/bfd0c6c1db5eb97d5b7db4ef18f14322762dd56b98557a47b15655ecfc1cca97.jpg", + "image_caption": [ + "Masked Image", + "DiffIR", + "Fig. 7: Visual comparison of image restoration methods. DiffRIR can alleviate the texture, contrast, and brightness disparities in inpainting & outpainting results." + ], + "image_footnote": [], + "bbox": [ + 230, + 366, + 320, + 433 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c4e0d036b74494b9d7b8132081ffe26561ca341b31efb0931527214ac7eced22.jpg", + "image_caption": [ + "DiffRIR (Ours)" + ], + "image_footnote": [], + "bbox": [ + 320, + 366, + 410, + 433 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/39ee74a6f76f0311326040847369f1ed36efc9463b8d7d0f9f31d580740f6c61.jpg", + "image_caption": [ + "DiffIR" + ], + "image_footnote": [], + "bbox": [ + 411, + 366, + 500, + 433 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7fe13a363b7c88b0aba6269d13a976fab0287419fb6c00a23b21be4a5e77ff98.jpg", + "image_caption": [ + "LLMGA", + "DiffRIR (Ours)" + ], + "image_footnote": [], + "bbox": [ + 501, + 366, + 589, + 433 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c099d4e7b008a726198e3aee2f5912569ba7db41e4f23094f58749cf7cbbb948.jpg", + "image_caption": [ + "Masked Image", + "DiffIR" + ], + "image_footnote": [], + "bbox": [ + 591, + 366, + 679, + 433 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/368b233adacba37700fedfdae6dd1fbd1200efdb273bebf78a4438ac48c87174.jpg", + "image_caption": [ + "LLMGA", + "DiffRIR (Ours)" + ], + "image_footnote": [], + "bbox": [ + 681, + 366, + 771, + 433 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Control SD using detailed language prompt or embedding?", + "bbox": [ + 212, + 522, + 486, + 553 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The results are shown in Fig. 5. We compare two approaches: GILL [29], which makes LLM estimate a fixed-sized embedding to control SD generation, and LLMGA Embedding, a variant of LLMGA where the language prompt is replaced with embedding, undergoing the same training process as LLMGA. The evaluation is conducted on MSCOCO by instructing these methods to generate images with the same prompts in multiple times in conversation form. (1) The quality of generated images (Fig. 5) in embedding-based methods", + "bbox": [ + 212, + 553, + 486, + 779 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "(i.e., GILL and LLMGA Embedding) deteriorate rapidly as the number of conversation turns increases. In contrast, our LLMGA remains unaffected. This discrepancy arises from the inherent noise present in the embeddings predicted by LLM. As the number of conversation turns rises, these generated embeddings", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/031046ecd895a4ef2f28532213d538f8fe3ee36e92da00499658e49e761ce098.jpg", + "image_caption": [ + "Table 5: T2I performance comparison of SD control schemes based on detailed language prompt and embedding." + ], + "image_footnote": [], + "bbox": [ + 509, + 527, + 777, + 685 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "LLMGA: Multimodal Large Language Model based Generation Assistant", + "bbox": [ + 243, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0d398b852fc49e5cb0ebd0827dcefd780a2de08b4f92255f968a2840ef0e3004.jpg", + "table_caption": [ + "Table 6: Datasets comparison. We conducted comparisons of FID on T2I and outpainting. The $\\checkmark$ signifies the utilization of the complete dataset during training. Conversely, the absence of the $\\checkmark$ indicates a reduction to only $10\\%$ of the original datasets." + ], + "table_footnote": [], + "table_body": "
MethodPrompt RefinementInpainting &OutpaintingSimilar Image GenerationInstruction-based EditingT2IOutpainting
LLMGA1✓✓✓21.44602.6533
LLMGA2✓✓✓19.19143.1054
LLMGA3✓✓✓19.56982.6612
LLMGA4✓✓✓19.06422.8047
LLMGA5 (Ours)✓✓✓✓18.52342.4409
", + "bbox": [ + 218, + 198, + 785, + 292 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "integrate with the preceding conversations, introducing even more noise. This poses challenges for the precise control of SD-generated content. Our LLMGA addresses this issue by mapping the embedding to the fixed language domain, effectively eliminating such noise. (2) Additionally, LLMGA Embedding also outperforms GILL, indicating that the prompt size used to guide SD generation should be adaptive in content, rather than a fixed size.", + "bbox": [ + 215, + 320, + 784, + 409 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Contribution of Training Data. To assess the impact of training data, we downsized one of the four training datasets in $\\mathrm{LLMGA}_5$ to $10\\%$ of its original magnitude, ensuring that LLMGA remains capable of furnishing responses in the prescribed format. The results are shown in Tab. 6. It is evident that prompt refinement, inpainting & outpainting, and instruction-based editing datasets enhance LLMGA's comprehension of image generation and editing properties, resulting in superior images. Moreover, comparing $\\mathrm{LLMGA}_3$ and $\\mathrm{LLMGA}_5$ , we can see that engaging in similar image generation training further improves the performance of LLMGA in both generation and editing.", + "bbox": [ + 215, + 411, + 785, + 546 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 217, + 568, + 356, + 583 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "LLM possesses an extensive reservoir of knowledge and powerful comprehension and reasoning capabilities. In this paper, we introduce a MLLM-based generation assistant (LLMGA), aiming to exploit LLM's capabilities in an interactive manner to facilitate more efficient and convenient image generation and editing. Compared to relying on LLM to predict a fixed-size embedding to control SD, we employ detailed generation prompts. These prompts prove to be more favorable for enhancing LLM's contextual comprehension and generating more accurate and rich content. To this end, we develop a two-stage training scheme and curate a dataset, including four parts: prompt refinement, similar image generation, inpainting & outpainting, and instruction-based editing. For the first stage, we train MLLM to understand the properties of image generation and editing, enabling it to give fitting responses. For the second stage, we optimize the SD unet to adapt to the generation prompt. Moreover, we propose a DM-based reference restoration network (DiffRIR) to mitigate disparities in texture, contrast, and brightness for image editing. Consequently, LLMGA can offer design suggestions and enhance results based on user's requests during interactions.", + "bbox": [ + 215, + 598, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia", + "bbox": [ + 274, + 113, + 707, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by the Research Grants Council under the Areas of Excellence scheme grant AoE/E-601/22-R.", + "bbox": [ + 215, + 178, + 787, + 209 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 233, + 321, + 250 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. NeurIPS (2022)", + "2. Austin, J., Johnson, D.D., Ho, J., Tarlow, D., van den Berg, R.: Structured denoising diffusion models in discrete state-spaces. NeurIPS (2021)", + "3. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)", + "4. Bao, F., Nie, S., Xue, K., Cao, Y., Li, C., Su, H., Zhu, J.: All are worth words: A vit backbone for diffusion models. In: CVPR (2023)", + "5. Bao, F., Nie, S., Xue, K., Li, C., Pu, S., Wang, Y., Yue, G., Cao, Y., Su, H., Zhu, J.: One transformer fits all distributions in multi-modal diffusion at scale. arXiv preprint arXiv:2303.06555 (2023)", + "6. Batzolis, G., Stanczuk, J., Schonlieb, C.B., Etmann, C.: Conditional image generation with score-based diffusion models. arXiv preprint arXiv:2111.13606 (2021)", + "7. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023)", + "8. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. NeurIPS (2020)", + "9. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ chatgpt quality (March 2023), https://lmsys.org/blog/2023-03-30-vicuna/", + "10. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022)", + "11. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022)", + "12. Dai, X., Hou, J., Ma, C.Y., Tsai, S., Wang, J., Wang, R., Zhang, P., Vandenhende, S., Wang, X., Dubey, A., et al.: Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807 (2023)", + "13. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS (2021)", + "14. Dong, R., Han, C., Peng, Y., Qi, Z., Ge, Z., Yang, J., Zhao, L., Sun, J., Zhou, H., Wei, H., et al.: Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499 (2023)", + "15. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Thompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023)" + ], + "bbox": [ + 218, + 266, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "LLMGA: Multimodal Large Language Model based Generation Assistant", + "bbox": [ + 243, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Fan, W.C., Chen, Y.C., Chen, D., Cheng, Y., Yuan, L., Wang, Y.C.F.: Frido: Feature pyramid diffusion for complex scene image synthesis. In: AAAI (2023)", + "17. Feng, Z., Zhang, Z., Yu, X., Fang, Y., Li, L., Chen, X., Lu, Y., Liu, J., Yin, W., Feng, S., et al.: Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. In: CVPR (2023)", + "18. Ge, Y., Ge, Y., Zeng, Z., Wang, X., Shan, Y.: Planting a seed of vision in large language model. arXiv preprint arXiv:2307.08041 (2023)", + "19. Gu, S., Chen, D., Bao, J., Wen, F., Zhang, B., Chen, D., Yuan, L., Guo, B.: Vector quantized diffusion model for text-to-image synthesis. In: CVPR (2022)", + "20. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022)", + "21. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020)", + "22. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. JMLR (2022)", + "23. Huang, S., Dong, L., Wang, W., Hao, Y., Singhal, S., Ma, S., Lv, T., Cui, L., Mohammed, O.K., Liu, Q., et al.: Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045 (2023)", + "24. Jiang, A.Q., Sablayrolles, A., Roux, A., Mensch, A., Savary, B., Bamford, C., Chaplot, D.S., Casas, D.d.l., Hanna, E.B., Bressand, F., et al.: Mixtral of experts. arXiv preprint arXiv:2401.04088 (2024)", + "25. Ju, X., Zeng, A., Bian, Y., Liu, S., Xu, Q.: Direct inversion: Boosting diffusion-based editing with 3 lines of code. ICLR (2024)", + "26. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: CVPR (2023)", + "27. Kim, G., Kwon, T., Ye, J.C.: Diffusionclip: Text-guided diffusion models for robust image manipulation. In: CVPR (2022)", + "28. Kingma, D., Salimans, T., Poole, B., Ho, J.: Variational diffusion models. NeurIPS (2021)", + "29. Koh, J.Y., Fried, D., Salakhutdinov, R.: Generating images with multimodal language models. arXiv preprint arXiv:2305.17216 (2023)", + "30. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)", + "31. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)", + "32. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023)", + "33. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. NeurIPS (2023)", + "34. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: CVPR (2022)", + "35. Meng, C., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073 (2021)", + "36. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)", + "37. OpenAI: Gpt-4 technical report (2023)", + "38. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: ICCV (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia", + "bbox": [ + 274, + 113, + 707, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. NeurIPS (2023)", + "40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)", + "41. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021)", + "42. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)", + "43. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023)", + "44. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH (2022)", + "45. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)", + "46. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)", + "47. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS (2022)", + "48. Shen, Y., Song, K., Tan, X., Li, D., Lu, W., Zhuang, Y.: Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580 (2023)", + "49. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015)", + "50. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)", + "51. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)", + "52. Sun, Q., Yu, Q., Cui, Y., Zhang, F., Zhang, X., Wang, Y., Gao, H., Liu, J., Huang, T., Wang, X.: Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222 (2023)", + "53. Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca (2023)", + "54. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)", + "55. Valevski, D., Kalman, M., Matias, Y., Leviathan, Y.: Unitune: Text-driven image editing by fine tuning an image generation model on a single image. arXiv preprint arXiv:2210.09477 (2022)", + "56. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: ICCVW (2021)", + "57. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "LLMGA: Multimodal Large Language Model based Generation Assistant", + "bbox": [ + 243, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "58. Wu, S., Fei, H., Qu, L., Ji, W., Chua, T.S.: Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519 (2023)", + "59. Xia, B., Zhang, Y., Wang, S., Wang, Y., Wu, X., Tian, Y., Yang, W., Timotfe, R., Van Gool, L.: Diffie2i: Efficient diffusion model for image-to-image translation. arXiv preprint arXiv:2308.13767 (2023)", + "60. Xia, B., Zhang, Y., Wang, S., Wang, Y., Wu, X., Tian, Y., Yang, W., Van Gool, L.: Diffir: Efficient diffusion model for image restoration. ICCV (2023)", + "61. Xia, B., Zhang, Y., Wang, Y., Tian, Y., Yang, W., Timofte, R., Van Gool, L.: Knowledge distillation based degradation estimation for blind super-resolution. ICLR (2023)", + "62. Xue, Z., Song, G., Guo, Q., Liu, B., Zong, Z., Liu, Y., Luo, P.: Raphael: Text-to-image generation via large mixture of diffusion paths. arXiv preprint arXiv:2305.18295 (2023)", + "63. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)", + "64. Yu, L., Shi, B., Pasunuru, R., Muller, B., Golovneva, O., Wang, T., Babu, A., Tang, B., Karrer, B., Sheynin, S., et al.: Scaling autoregressive multi-modal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591 (2023)", + "65. Zhang, K., Mo, L., Chen, W., Sun, H., Su, Y.: Magicbrush: A manually annotated dataset for instruction-guided image editing. NeurIPS (2024)", + "66. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)", + "67. Zhong, S., Huang, Z., Wen, W., Qin, J., Lin, L.: Sur-adapter: Enhancing text-to-image pre-trained diffusion models with large language models. In: ACM MM (2023)", + "68. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. TPAMI (2017)", + "69. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 575 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia", + "bbox": [ + 274, + 114, + 707, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/LLMGA_ Multimodal Large Language Model based Generation Assistant/d91671aa-dc99-4c79-b4c9-9d98440c12f4_model.json b/2024/LLMGA_ Multimodal Large Language Model based Generation Assistant/d91671aa-dc99-4c79-b4c9-9d98440c12f4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dd86098c0c472e1da6c4da8723b7b99608536538 --- /dev/null +++ b/2024/LLMGA_ Multimodal Large Language Model based Generation Assistant/d91671aa-dc99-4c79-b4c9-9d98440c12f4_model.json @@ -0,0 +1,2854 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.245, + 0.141, + 0.761, + 0.185 + ], + "angle": 0, + "content": "LLMGA: Multimodal Large Language Model based Generation Assistant" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.212, + 0.751, + 0.228 + ], + "angle": 0, + "content": "Bin Xia\\(^{1}\\), Shiyin Wang\\(^{2}\\), Yingfan Tao\\(^{2}\\), Yitong Wang\\(^{2}\\), and Jiaya Jia\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.366, + 0.239, + 0.637, + 0.254 + ], + "angle": 0, + "content": "1 The Chinese University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.445, + 0.254, + 0.56, + 0.268 + ], + "angle": 0, + "content": "2 ByteDance Inc" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.269, + 0.582, + 0.282 + ], + "angle": 0, + "content": "https://llmga.github.io/" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.316, + 0.74, + 0.621 + ], + "angle": 0, + "content": "Abstract. In this paper, we introduce a Multimodal Large Language Model-based Generation Assistant (LLMGA), leveraging the vast reservoir of knowledge and proficiency in reasoning, comprehension, and response inherent in Large Language Models (LLMs) to assist users in image generation and editing. Diverging from existing approaches where Multimodal Large Language Models (MLLMs) generate fixed-size embeddings to control Stable Diffusion (SD), our LLMGA provides a detailed language generation prompt for precise control over SD. This not only augments LLM context understanding but also reduces noise in generation prompts, yields images with more intricate and precise content, and elevates the interpretability of the network. To this end, we curate a comprehensive dataset comprising prompt refinement, similar image generation, inpainting & outpainting, and instruction-based editing. Moreover, we propose a two-stage training scheme. In the first stage, we train the MLLM to grasp the properties of image generation and editing, enabling it to generate detailed prompts. In the second stage, we optimize SD to align with the MLLM's generation prompts. Additionally, we propose a reference-based restoration network to alleviate texture, brightness, and contrast disparities between generated and preserved regions during inpainting and outpainting. Extensive results show that LLMGA has promising generation and editing capabilities and can enable more flexible and expansive applications in an interactive manner." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.635, + 0.74, + 0.649 + ], + "angle": 0, + "content": "Keywords: Interactive generation and editing \\(\\cdot\\) LLM \\(\\cdot\\) Diffusion Model" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.674, + 0.377, + 0.691 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.786, + 0.795 + ], + "angle": 0, + "content": "Artificial Intelligence Generated Content (AIGC) has witnessed remarkable advancements, particularly propelled by the evolution of large language models (LLMs) [8, 10, 54] for text generation and diffusion models (DMs) [21, 42] for image generation. LLMs, in particular, have received considerable acclaim for their exceptional ability to comprehend, reason, make decisions, possess extensive knowledge, and generate text with unparalleled accuracy and fluency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Recent studies have begun delving deeper into Multimodal Large Language Models (MLLMs) [1,69] built upon LLMs, aiming to empower LLMs to comprehend inputs extending beyond text. For example, BLIP-2 [30] and LLaVA [33]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.276, + 0.115, + 0.71, + 0.129 + ], + "angle": 0, + "content": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.357, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.145, + 0.493, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.145, + 0.665, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.145, + 0.808, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.244, + 0.357, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.244, + 0.493, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.244, + 0.661, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.244, + 0.808, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.362, + 0.788, + 0.487 + ], + "angle": 0, + "content": "Fig. 1: Some examples of LLMGA for assisting in image generation and editing. (1) T2I generation. LLMGA can refine the user's generation prompt to produce more vivid and vibrant images. (2) Similar image generation. LLMGA can understand the component and layout of the input images and generate a similar image. (3) Inpainting & Outpainting. LLMGA can provide detailed generation prompts based on user preferences and input images. (4) Instruction based editing. LLMGA can understand user instructions and realize accurate editing. (5) Interactive image generation and editing exemplify the comprehensive capabilities of LLMGA. Users can design satisfactory images by engaging in interactions with LLMGA, leveraging its vast knowledge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.612 + ], + "angle": 0, + "content": "employ visual encoders to transform images into input embeddings, enabling them to be used as prompts alongside text input for the LLM, thus achieving LLMs with the visual modality. Furthermore, recent works focused on extending the capabilities of LLMs to generate multimodal outputs. For example, GILL [29] involves instructing LLMs to predict fixed-size visual embeddings aligned with CLIP [40] space to control the Stable Diffusion [42] (SD) for image generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.844 + ], + "angle": 0, + "content": "However, existing works [29, 52] merely focus on enabling LLM to output images but do not aim to assist users in generating or editing images to enhance quality. In this paper, we aim to develop a Multimodal Large Language Model-based Generation Assistant (LLMGA) to better assist image generation models, making them more user-friendly and capable of producing high-quality images. In contrast to certain methods [29, 52] that leverage MLLMs to predict fixed-size visual embeddings for implicit SD control, our approach is straightforward. We guide the generation of SD using detailed language prompts from MLLM based on five reasons. (1) The embeddings predicted by the MLLM are often filled with noise. This can be filtered out by mapping them to a fixed language domain, enabling precise control of SD. (2) Detailed language prompts can make the network more transparent and interactive, allowing users to understand MLLM's thoughts for generating images. (3) MLLM is pre-trained on vast textual datasets. Explicit language prompts rather than implicit embeddings are more advantageous for MLLM to generate prompts and comprehend context. (4)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.244, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LLMGA: Multimodal Large Language Model based Generation Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "Dynamic-sized language prompt facilitates the addition of generation requests during interactions. (5) Training is more simple and efficient." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.787, + 0.302 + ], + "angle": 0, + "content": "However, we face several challenges: (1) MLLM may reject the execution of generation instructions due to its nature as a language assistant. (2) MLLM lacks a comprehensive understanding of image generation and editing, and cannot provide an accurate and detailed generation prompt. (3) Determining which part of texts generated by MLLM to guide SD generation. (4) SD's CLIP encode only 75 tokens. Additionally, SD is trained on short captions, whereas our LLMGA typically generates detailed prompts exceeding 150 tokens. This discrepancy poses a challenge for SD in following the detailed prompt of LLMGA." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.303, + 0.788, + 0.53 + ], + "angle": 0, + "content": "To this end, we have devised a two-stage training scheme. First, we construct a training dataset: prompt refinement, similar image generation, inpainting & outpainting, and visual question answering. We then train LLMGA on these four datasets to cultivate four fundamental capabilities: (1) For concise user prompts, LLMGA can refine the generation of intricate details, encompassing attire, background, and characters. (2) LLMGA can precisely regenerate an image it observes. (3) LLMGA can generate or refine prompts for inpainting & outpainting based on its understanding of the image. (4) LLMGA can generate accurate prompts for instruction-based editing according to users' requirements and given images. Additionally, we make LLMGA use special symbols and to distinguish generation prompts and responses. In the second stage, we freeze the parameters of LLMGA's MLLM and initiate joint training with the SD. This process enables the SD to acclimate to the detailed prompt produced by the MLLM. Notably, when the input token count exceeds 75, we iteratively apply the CLIP [40] encoder to the surplus tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.532, + 0.788, + 0.653 + ], + "angle": 0, + "content": "Moreover, we have identified noticeable disparities in texture, contrast, and brightness between the newly generated and preserved sections in inpainting & outpainting. Therefore, we propose a Diffusion-based Reference Restoration Network (DiffRIR). Specifically, aside from images generated by SD, we add masked images as reference inputs into DiffRIR. This enables the DiffRIR to refer to the texture, contrast, and brightness of the preserved regions for restoration. Additionally, we introduce perturbations to contrast and brightness during training, enabling DiffRIR to correct contrast and brightness disparities in the images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.788, + 0.762 + ], + "angle": 0, + "content": "As shown in Fig. 1, LLMGA is a unified and interactive framework for image generation and editing, endowed with a wide array of capabilities: (1) LLMGA leverages its extensive world knowledge and powerful reasoning abilities to assist image generation and editing and significantly improve results. (2) LLMGA can be integrated with external plugins, like ControlNet. (3) Most importantly, users can interact with LLMGA to design satisfying images in a more convenient, flexible, and enjoyable way. In summary, our contributions are as follows:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.226, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "- We proposed LLMGA, a simple yet powerful interactive generation and editing framework. Experiments affirm the efficacy of LLMGA in enhancing generation and editing thanks to its vast knowledge and interactive features. Plus, LLMGA can integrate with external plugins for wider applications." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.275, + 0.114, + 0.709, + 0.129 + ], + "angle": 0, + "content": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.784, + 0.206 + ], + "angle": 0, + "content": "- We construct a training dataset, including four parts: prompt refinement, similar image generation, inpainting & outpainting, and instruction-based editing. This enhances LLMGA's comprehension of generation and editing tasks while standardizing response formats." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.207, + 0.785, + 0.265 + ], + "angle": 0, + "content": "- We proposed a restoration network DiffRIR, which introduces reference images and training perturbations to contrast and brightness. DiffRIR can alleviate texture, contrast, and brightness discrepancies between newly generated and preserved regions for edited images." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.266, + 0.785, + 0.295 + ], + "angle": 0, + "content": "- Open-source. The following assets are released: the generated data, the codebase for model training, the model checkpoint, and a demo." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.785, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.314, + 0.388, + 0.33 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.785, + 0.567 + ], + "angle": 0, + "content": "Diffusion Model. Diffusion Models (DMs) [2,4,7,13,19,21,22,28,35,38,49,51] have achieved remarkable results in image generation. DMs adopt a parameterized Markov chain to optimize the lower variational bound on the likelihood function. In this way, it can generate realistic images from Gaussian noise. After that, several DM methods [3,5,6,12,16,17,26,27,34,39,44,45,55,59,60,62] have been tailored to enhance the text-to-image (T2I) generation and editing. Notably, GLIDE [36] pioneered the incorporation of text features into transformer blocks during the denoising process. Subsequently, DALL-E [41], Imagen [46], and Stable Diffusion [42] have made substantial strides in improving T2I generation. Subsequently, some works [43,66,67] introduced conditioning controls to the DMs to facilitate a more convenient and precise manipulation of the generation process. Overall, enhancing the user-friendliness of DMs is a key focus within the community. In this paper, we introduce LLMGA, leveraging the extensive knowledge and powerful reasoning capabilities of LLM to facilitate users in achieving more easily attainable and satisfactory image designs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.785, + 0.674 + ], + "angle": 0, + "content": "Multimodal Large Language Models. Recently, LLMs have undeniably made profound impacts and revolutions within the entire AI community and beyond. For example, exemplary LLMs, such as ChatGPT and GPT4 [37], have showcased remarkable abilities in comprehension, reasoning, responses, and knowledge reservoirs. Subsequently, a range of LLMs [11], including Vicuna [9], LLaMA [54], and Alpaca [53] have been released as open-source models, substantially propelling advancements of the community." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.785, + 0.779 + ], + "angle": 0, + "content": "Afterward, the community began focusing on the development of the Multimodal Large Language Model [14, 15, 18, 23, 63, 64, 69]. They aim to enable LLMs to comprehend both images and text and provide textual responses. For instance, Flamingo [1] encodes images and feeds them into the LLM's attention layer. BLIP-2 [30] employs Q-Former to encode input images into queries. Additionally, LLaVA [33] leverages CLIP [40] to encode images into visual embeddings, which are then concatenated with text embeddings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Recent concurrent works, such as Next-GPT [58], have extended the capabilities to encompass audio and video modalities. Moreover, Visual-ChatGPT [57] and HuggingGPT [48] make LLMs act as agents capable of invoking various pretrained visual models to achieve MLLM. However, these works focus on making" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LLMGA: Multimodal Large Language Model based Generation Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.152, + 0.557, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.152, + 0.78, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.452 + ], + "angle": 0, + "content": "Fig. 2: The overview of LLMGA. (a) In the first training stage, we train the MLLM to produce generation prompts based on provided instructions. Moreover, we construct a training dataset including four categories: prompt refinement, similar image generation, inpainting & outpainting, and instruction-based editing. (b) In the second training stage, we optimize SD to adapt to the detailed generation prompts from MLLM. (c) In the inference stage, LLMGA can respond to user queries and assist in various tasks, such as image generation, inpainting & outpainting, and editing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.571 + ], + "angle": 0, + "content": "LLM determine the combined invocation of modules (such as detection, recognition, and generation) to fulfill user requirements. However, these methods are not tailored for generation and editing and use concise prompts that lack the capability to enhance results. Thus, we propose LLMGA, which is designed to assist with various image generation and editing tasks. It can achieve satisfactory results by strong reasoning capability and flexible interaction with users." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.591, + 0.381, + 0.609 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.621, + 0.436, + 0.635 + ], + "angle": 0, + "content": "3.1 Overview of LLMGA" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.81 + ], + "angle": 0, + "content": "In this paper, we aim to design a MLLM-based Generation Assistant (LLMGA). Our LLMGA produces detailed language-based generation prompts to control SD rather than predicting fixed-sized visual embeddings [29, 52] to govern SD. This has five advantages: (1) Visual embeddings contain noise, and mapping them to the language domain can filter out this noise, enabling precise SD control. (2) Language-based generation prompts facilitate users in understanding the LLMGA's thoughts, enhancing interaction. (3) Dynamic-sized language-based generation prompt enables the addition of generation requests. (4) MLLM is pre-trained on textual datasets. Language prompts rather than implicit visual embeddings are more advantageous for MLLM to generate accurate prompts and comprehend context. (5) Training is simpler and more efficient." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "However, we need to address several issues: (1) As a language assistant, MLLM may decline the execution of generation instructions. (2) MLLM lacks" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.276, + 0.114, + 0.71, + 0.129 + ], + "angle": 0, + "content": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "a nuanced understanding of image generation and editing, and cannot produce precise and detailed generation prompts. (3) MLLM needs to decide which part of the output text serves as generation prompts to guide generation. (4) SD's CLIP encode only 75 tokens. Moreover, SD is primarily trained on short captions, while detailed prompts generated by LLMGA may exceed 150 tokens. This disparity makes it hard for SD to understand the instructions from LLMGA. To address the aforementioned challenges, we construct a training dataset and design two-stage training schemes to train the MLLM (Sec. 3.2) and SD (Sec. 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.344 + ], + "angle": 0, + "content": "The network structure and pipeline of LLMGA are illustrated in Fig. 2. Specifically, as shown in Fig. 2 (a) and (c), the images \\(\\mathbf{I}_{input}\\) are encoded into image embeddings by CLIP vision encoder and a projection layer. Subsequently, the image embedding is concatenated with the text embedding and fed into the LLM to obtain text output \\(\\mathbf{T}_{output}\\). This process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.359, + 0.785, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {\\text {o u t p u t}} = \\mathcal {F} _ {M L L M} \\left(\\mathbf {T} _ {\\text {i n p u t}}, \\mathbf {I} _ {\\text {i n p u t}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.786, + 0.411 + ], + "angle": 0, + "content": "where \\(\\mathbf{T}_{input}\\) indicates the input text instructions from users. It is notable that \\(\\mathcal{F}_{MLLM}\\) can process only \\(\\mathbf{T}_{input}\\) as input." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.786, + 0.456 + ], + "angle": 0, + "content": "The text output \\(\\mathbf{T}_{output}\\) can comprise two components: text response \\(\\mathbf{T}_R\\) and generation prompt \\(\\mathbf{T}_P\\). To distinguish between \\(\\mathbf{T}_R\\) and \\(\\mathbf{T}_P\\), we adopt new special tokens, i.e., \\(<\\text{gen\\_img}>\\) and \\(\\), to encompass \\(\\mathbf{T}_P\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.787, + 0.502 + ], + "angle": 0, + "content": "We present \\(\\mathbf{T}_R\\) as the immediate text response to users. Concurrently, \\(\\mathbf{T}_P\\) is further fed into the subsequent SD to guide T2I generation (Eq. 2), inpainting & outpainting (Eq. 3), and instruction-based image editing (Eq. 4)." + }, + { + "type": "equation", + "bbox": [ + 0.436, + 0.513, + 0.785, + 0.529 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {g} = \\mathcal {F} _ {T 2 I} (\\mathbf {T} _ {P}, \\mathbf {Z}), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.555, + 0.785, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {p} = \\mathcal {F} _ {\\text {P a i n t}} (\\mathbf {I} _ {\\text {i n p u t}}, \\mathbf {I} _ {\\text {m a s k}}, \\mathbf {T} _ {P}, \\mathbf {Z}), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.592, + 0.785, + 0.608 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {e} = \\mathcal {F} _ {\\text {E d i t}} \\left(\\mathbf {I} _ {\\text {i n p u t}}, \\mathbf {T} _ {P}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.705 + ], + "angle": 0, + "content": "Where \\(\\mathbf{Z}\\) denotes the random Gaussian noise. Furthermore, to ensure the encoding of all \\(\\mathbf{T}_P\\) for SD, we iteratively run the CLIP text encoder until all prompts are encoded. For inpainting & outpainting, except the input image \\(\\mathbf{I}_{input}\\), an additional mask \\(\\mathbf{I}_{mask}\\) are essential inputs for the inpainting SD to specify the region requiring generation. For instruction-based image editing, \\(\\mathcal{F}_{Edit}\\) performs inversion [25, 50] and prompt-to-prompt [20] based on T2I SD." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.725, + 0.399, + 0.741 + ], + "angle": 0, + "content": "3.2 MLLM Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As described in Sec. 3.1, original MLLMs are specifically designed and trained as language assistants, but they lack the proficiency to assist in image generation, and editing. Notably, considering the input of images and the performance of open-sourced MLLMs, employing few-shot learning to guide the model to achieve the desired results proves challenging and inefficient. Therefore, it is crucial to train the MLLM to serve as proficient assistants in image generation and editing" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.244, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "LLMGA: Multimodal Large Language Model based Generation Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "tasks, understanding the expected response formats and enhancing their comprehension of image generation and editing properties. To this end, as depicted in Fig. 2 (a), we construct a training dataset consisting of four categories:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.787, + 0.298 + ], + "angle": 0, + "content": "(1) Prompt Refinement. We establish this dataset to cultivate the prompt refinement ability of the MLLM. Specifically, we utilize GPT4-V to furnish detailed descriptions of images in MSCOCO [31]. These detailed descriptions, along with the original MSCOCO brief descriptions, constitute a training text pair. During training, we input the brief MSCOCO captions and randomly select and append a generation instruction. When a generation instruction is included in the prompt, we add and on the later description." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.787, + 0.373 + ], + "angle": 0, + "content": "(2) Similar Image Generation. We select images from the MSCOCO dataset along with corresponding detailed descriptions generated by GPT4-V to create the Similar Image Generation dataset. During training, we input the images along with a generation instruction. In cases where a generation instruction is provided, we add and on the subsequent description." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.787, + 0.434 + ], + "angle": 0, + "content": "(3) Inpainting & Outpainting. We use pairs of detailed descriptions and images from the Similar Image Generation dataset. During training, we input masked images with inpainting or outpainting instructions. Besides, we include and on the subsequent description." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.787, + 0.525 + ], + "angle": 0, + "content": "(4) Instruction-based editing. We fine-tune Mixtral-8x7B [24] to enable it to generate editing data based on detailed descriptions from MSCOCO. Subsequently, we clean the generated data. During training, we can input images or original captions and provide corresponding editing instructions, aiming to train LLMGA to output the target caption. Additionally, we include and tags on the target caption during training." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.193, + 0.787, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.787, + 0.615 + ], + "angle": 0, + "content": "During training, alongside the above data, we integrated the image designing and Alpaca [53] dataset, to enhance LLMGA's question-answering (QA). Notably, we excluded certain visual multimodal incompatible question-answer pairs from it. Furthermore, we incorporated the LLaVA v1.5 mix665k dataset [32] to endow LLMGA with Visual Question Answering (VQA) capabilities. We provide more details in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.787, + 0.676 + ], + "angle": 0, + "content": "As illustrated in Fig. 2 (a), we freeze the CLIP vision encoder and optimize the projection layer and LLM. The model is trained end-to-end using the autoregressive cross-entropy loss \\((\\mathcal{L}_{MLLM})\\) for text generation. Given the ground-truth targets \\(\\mathbf{T}_{GT}\\), this loss can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.687, + 0.786, + 0.705 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M L L M} = \\mathbf {C E} \\left(\\mathbf {T} _ {\\text {o u t p u t}}, \\mathbf {T} _ {G T}\\right). \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.725, + 0.472, + 0.74 + ], + "angle": 0, + "content": "3.3 Stable Diffusion Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.841 + ], + "angle": 0, + "content": "As described in Sec. 3.1, the original SD's CLIP text encoder only encodes 75 tokens, which cannot handle the entire MLLM's generation prompt. Moreover, the original SD is trained on brief captions, which cannot fully understand the generation prompts. Thus, we repeatedly use the CLIP text encoder to encode all instances of \\(\\mathbf{T}_P\\) for SD. Besides, we train the T2I SD model and the inpainting SD model, respectively. For both generation and inpainting & outpainting tasks," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.275, + 0.114, + 0.71, + 0.129 + ], + "angle": 0, + "content": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "the generation prompts of MLLM are detailed descriptions of images. Therefore, during training, we can instruct MLLM to provide a detailed description \\(\\mathbf{T}_P\\) for images from the LAION-Aesthetics [47] and MSCOCO datasets. Subsequently, \\(\\mathbf{T}_P\\) is fed into T2I SD or inpainting SD for joint training. Notably, we only optimize the SD unet while freezing the parameters of other networks. To accelerate the training, we record the prompt \\(\\mathbf{T}_P\\) of MLLM to avoid redundant calculations. The model is trained using SD loss (Eq. 6). For instruction-based editing, we directly adopt the pre-trained T2I SD model." + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.28, + 0.786, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S D} = \\mathbb {E} _ {\\mathbf {Z} _ {t}, \\mathbf {C}, \\epsilon , t} \\left(\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\mathbf {Z} _ {t}, \\mathbf {C}\\right) \\| _ {2} ^ {2}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.392 + ], + "angle": 0, + "content": "where \\(\\mathbf{Z}_t = \\sqrt{\\overline{\\alpha}_t}\\mathbf{Z}_0 + \\sqrt{1 - \\overline{\\alpha}_t}\\epsilon\\) represents the noised feature map at timestep \\(t\\). Ground truth images are encoded into latent space to derive \\(\\mathbf{Z}_0\\). Here, \\(\\epsilon \\in \\mathcal{N}(0,\\mathbf{I})\\) represents Gaussian noise, and \\(\\epsilon_{\\theta}\\) refers to the SD unet. \\(\\mathbf{C}\\) indicates the conditional information. For T2I generation, \\(\\mathbf{C}\\) is \\(\\mathbf{T}_P\\). For inpainting and outpainting, \\(\\mathbf{C}\\) contains \\(\\mathbf{T}_P\\), the mask, and the VAE-encoded masked image." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.413, + 0.513, + 0.429 + ], + "angle": 0, + "content": "3.4 Restoration Network Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.437, + 0.788, + 0.573 + ], + "angle": 0, + "content": "For SD inpainting & outpainting, we observed noticeable disparities between the preserved and newly generated regions in the edited images. To enhance the consistency between the newly generated and the preserved regions, we introduce a reference-based restoration scheme. Specifically, existing restoration methods [56, 60, 61] take low-quality (LQ) images as input and produce high-quality (HQ) images, but they often do not leverage the preserved information from the given masked image. Different from them, we concatenate the LQ image \\( I_{LQ} \\) and the masked image, i.e., \\( (1 - \\mathbf{I}_{mask})\\mathbf{I}_{GT} \\), as inputs and feed them into the restoration model \\( \\mathcal{F}_R \\). This process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.356, + 0.587, + 0.786, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {H Q} = \\mathcal {F} _ {R} (\\operatorname {c o n c a t} \\left(\\mathbf {I} _ {L Q}, (1 - \\mathbf {I} _ {\\text {m a s k}}) \\mathbf {I} _ {G T}\\right)). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.787, + 0.658 + ], + "angle": 0, + "content": "To further mitigate the brightness and contrast disparities, we introduced additional color degradation (i.e., random brightness and contrast disturbance) into the training process of the restoration model, which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.439, + 0.672, + 0.786, + 0.686 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {2} (\\mathbf {x}) = c _ {1} \\mathbf {x} + c _ {2}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.429, + 0.713, + 0.786, + 0.729 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {L Q} = \\mathcal {D} _ {2} \\left(\\mathcal {D} _ {1} \\left(\\mathbf {I} _ {G T}\\right)\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_2\\) indicates contrast and brightness disturbance, \\(\\mathbf{x}\\) denotes the input image. \\(c_{1}\\) is the contrast gain, randomly varied within the range of [0.94, 1.06], while \\(c_{2}\\) is the brightness bias, randomly varied within the range of \\([-0.05, 0.05]\\). \\(\\mathcal{D}_1\\) represents the real-world degradation process used in restoration model [56, 60, 61]. \\(\\mathbf{I}_{GT}\\) is ground truth image, and \\(\\mathbf{I}_{LQ}\\) is LQ image. Here, we adopt the network structure of the SOTA restoration model DiffIR [60] and apply our schemes to it to obtain our Diffusion-based Reference Restoration Network (DiffRIR)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.244, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LLMGA: Multimodal Large Language Model based Generation Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.377, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.178, + 0.458, + 0.194 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.203, + 0.784, + 0.263 + ], + "angle": 0, + "content": "For the first stage of training, we train the MLLM (including projection and LLM) on \\(8 \\times \\mathrm{A}100\\). The batch size is set to 128. Besides, the training datasets include VQA (LLaVA v1.5 mix665k), QA, prompt refinement, similar image generation, inpainting & outpainting, and instruction-based editing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.263, + 0.785, + 0.307 + ], + "angle": 0, + "content": "For the second stage of training, we adopt the Stable Diffusion 1.5 (SD1.5) as the initial image generation or inpainting & outpainting model. We train these models on \\(8 \\times \\mathrm{A}100\\). The batch size is set to 32." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.785, + 0.339 + ], + "angle": 0, + "content": "For the restoration network, we train DiffRIR on \\(8 \\times \\mathrm{V}100\\). The batch size is set to 64. Please check more details in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.362, + 0.442, + 0.377 + ], + "angle": 0, + "content": "4.2 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.387, + 0.488, + 0.569 + ], + "angle": 0, + "content": "Evaluation on T2I Generation. The results are shown in Tab. 1. Notably, SD1.5 is the original Stable Diffusion 1.5 while SD1.5-ft indicates the finetuned SD1.5 in LLMGA. We also compare our LLMGA with the recently proposed multimodal generative model GILL [29]. (1) Comparing SD1.5, our LLMGA-7B achieves notable 5.4847 FID and 5.05 IS improvements, underscoring the effectiveness of LLMGA. (2) Moreover," + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.785, + 0.689 + ], + "angle": 0, + "content": "our LLMGA-7B significantly outperforms GILL. (3) In the 3rd and 4th rows of Tab. 1, we use LLMGA-7B to refine the short MSCOCO caption to the detailed generation prompt and send it to the SD1.5 and SD1.5-ft, respectively. Our LLMGA-7B with SD1.5-ft achieves significant 5.0712 FID and 3.92 IS improvements, respectively. This demonstrates our second-stage training can help SD1.5-ft to better follow detailed prompts from MLLM. (4) Comparing the 4th and 6th rows of Tab. 1, LLMGA-13B exhibits better performance than LLMGA-7B due to its superior reasoning ability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.794 + ], + "angle": 0, + "content": "The qualitative results are shown in Fig. 3. (1) LLMGA excels in refining prompts by incorporating details to generate visually rich and pleasing images. For example, in the first row, LLMGA crafts a battle attire for the husky, depicting engaging scenarios of battling monsters. This makes user usage more convenient, eliminating the need for them to think about generating image details themselves. (2) LLMGA can leverage its extensive knowledge base to generate images, even for concepts users may not be familiar with, like a space elevator." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.843 + ], + "angle": 0, + "content": "Evaluation on Instruction-based Editing. For instruction-based editing, we utilize LLMGA to provide a detailed description of the edited image based on the input image and user editing instructions. We then employ Direct Inversion [25]" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.388, + 0.788, + 0.417 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison on T2I generation on the MSCOCO [31] dataset." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.417, + 0.789, + 0.541 + ], + "angle": 0, + "content": "
MethodFID↓IS↑
SD1.524.008135.99
GILL [29]25.112334.20
LLMGA-7b(SD1.5)23.594637.12
LLMGA-7b(SD1.5-ft, Ours)18.523441.04
LLMGA-13b(SD1.5)23.582837.58
LLMGA-13b(SD1.5-ft, Ours)18.406341.16
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.275, + 0.114, + 0.711, + 0.129 + ], + "angle": 0, + "content": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.145, + 0.773, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.376, + 0.785, + 0.391 + ], + "angle": 0, + "content": "Fig. 3: T2I visual comparison. LLMGA can produce accurate and high-quality results." + }, + { + "type": "table_caption", + "bbox": [ + 0.223, + 0.406, + 0.779, + 0.421 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison for image editing on MagicBrush test set [65]." + }, + { + "type": "table", + "bbox": [ + 0.307, + 0.434, + 0.698, + 0.51 + ], + "angle": 0, + "content": "
MethodsL1↓L2↓CLIP-I↑DINO↑CLIP-T↑
InstructPix2Pix0.11970.04160.84420.72520.2909
MagicBrush0.06470.02240.92930.89130.2979
LLMGA (Ours)0.08140.02180.89360.87680.3137
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.689 + ], + "angle": 0, + "content": "and prompt-to-prompt [20] methods to obtain the edited image. (1) The quantitative results are presented in Tab. 2. Notably, our LLMGA did not train SD for image editing like InstructPix2Pix [7] and MagicBrush [65]. However, our zero-shot performance on the MagicBrush test set surpassed that of InstructPix2Pix, achieving performance similar to that of MagicBrush. (2) Additionally, our LLMGA offers a superior user experience in interactive editing, allowing image modifications to be carried out conversationally. In contrast, InstructPix2Pix only supports input via instructions and output as images. Visualized results are shown in Fig. 4, by leveraging the powerful reasoning capabilities of LLM, our LLMGA can provide more accurate and reasonable editing results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Evaluation on Inpainting and Outpainting. The results are shown in Tab. 3. (1) Comparing the 1st and 3rd rows of Tab. 3, our LLMGA-7B achieves significant improvements of 2.5681 FID over the SD1.5 in outpainting under wide masks. (2) In the 2nd and 3rd rows of Tab. 3, we make LLMGA imagine the complete generation prompts for the given masked images, which are then input into the later SD. Notably, our LLMGA-7B (with SD1.5-ft) demonstrates a significant FID improvement over LLMGA-7B (with SD1.5) in both outpainting and inpainting. This demonstrates the second stage of training makes SD better follow the prompts from MLLM. (3) Comparing the 3rd and 5th rows, our LLMGA-13B outperforms LLMGA-7B due to its superior reasoning capabilities." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LLMGA: Multimodal Large Language Model based Generation Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.149, + 0.765, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.303, + 0.408, + 0.698, + 0.423 + ], + "angle": 0, + "content": "Fig. 4: Visual comparison on instruction-based editing." + }, + { + "type": "table_caption", + "bbox": [ + 0.226, + 0.445, + 0.776, + 0.46 + ], + "angle": 0, + "content": "Table 3: Quantitative comparison for outpainting & inpainting on Places [68]." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.472, + 0.789, + 0.603 + ], + "angle": 0, + "content": "
MethodOutpaintingInpainting
NarrowMasksWideMasksNarrowMasksWideMasks
FID↓LPIPS↓FID↓LPIPS↓FID↓LPIPS↓FID↓LPIPS↓
SD1.52.01670.22835.00900.37341.07950.12361.28550.1434
LLMGA-7b (SD1.5)1.55300.22633.26300.36881.06920.12331.09830.1427
LLMGA-7b (SD1.5-ft, Ours)1.29730.22152.44090.36160.80270.11710.98070.1405
LLMGA-13b (SD1.5)1.56310.22633.08450.36791.03260.12281.09780.1426
LLMGA-13b (SD1.5-ft, Ours)1.21600.22102.36630.36090.79920.11660.97800.1400
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.788, + 0.732 + ], + "angle": 0, + "content": "The qualitative results are shown in Fig. 5. We can see that LLMGA can deduce and imagine complete images based on masked input images. For example, in the 3rd row of Fig. 5, LLMGA can infer the presence of wind turbines on the mountain based on the given environment. Overall, LLMGA's powerful reasoning capability and extensive knowledge can assist users in conveniently making accurate and visually pleasing inpainting and outpainting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Evaluation on ControlNet. Our LLMGA demonstrates exceptional scalability, enabling integration with external plugins like ControlNet [66]. Here, we utilize LLMGA to create detailed prompts derived from input images and user requirements, working alongside ControlNet to guide image generation. As depicted in Fig. 6, our LLMGA significantly enhances the diversity and richness of outcomes in picture reference-guided image generation. Furthermore, more external plugins can also be integrated into the interactive framework of LLMGA," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.275, + 0.114, + 0.71, + 0.129 + ], + "angle": 0, + "content": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.145, + 0.773, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.377, + 0.706, + 0.392 + ], + "angle": 0, + "content": "Fig. 5: Visual comparison on inpainting and outpainting." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.413, + 0.773, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.59, + 0.788, + 0.62 + ], + "angle": 0, + "content": "Fig. 6: Visualization of LLMGA plus ControlNet. Our LLMGA can enhance the details in generated images, producing visually pleasing images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.785, + 0.683 + ], + "angle": 0, + "content": "combining with LLMGA's reasoning design capabilities and all previous functions to allow for a broader range of creative and engaging applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.685, + 0.788, + 0.806 + ], + "angle": 0, + "content": "Evaluation on Image Restoration. The results are shown in Tab. 4. For comparisons, we validate DiffIR [60] and our DiffRIR on the outpainting image generated by LLMGA. (1) Comparing the 2nd and 3rd rows of Tab. 4, it is evident that introducing a reference scheme can significantly improve restoration performance. (2) When comparing the 3rd and 4th rows of Tab. 4, it can be observed that introducing color degradation helps alleviate the bright and contrast distortion caused by SD. (3) Comparing the 1st and 4th rows of Tab. 4, our DiffRIR yields significant improvement, validating the effectiveness of DiffRIR." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.808, + 0.788, + 0.839 + ], + "angle": 0, + "content": "As shown in Fig. 7, our DiffRIR (i.e., \\(\\mathrm{DiffRIR}_2\\) in Tab. 4) can alleviate the texture, brightness, and contrast discrepancies, and generate realistic details." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.244, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LLMGA: Multimodal Large Language Model based Generation Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.784, + 0.16 + ], + "angle": 0, + "content": "Table 4: Quantitative comparison on image restoration for Places [68] outpainting." + }, + { + "type": "table", + "bbox": [ + 0.304, + 0.172, + 0.7, + 0.261 + ], + "angle": 0, + "content": "
MethodColorFID↓LPIPS↓
Reference Degradation
LLMGA-7BXX2.44090.3616
LLMGA+DiffIR [60]XX2.35870.3612
LLMGA+DiffRIR1✓X2.29930.3609
LLMGA+DiffRIR2 (Ours)✓✓2.26870.3607
" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.281, + 0.32, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.354, + 0.311, + 0.365 + ], + "angle": 0, + "content": "Masked Image" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.282, + 0.411, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.354, + 0.389, + 0.364 + ], + "angle": 0, + "content": "LLMGA" + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.282, + 0.5, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.353, + 0.49, + 0.364 + ], + "angle": 0, + "content": "Masked Image" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.282, + 0.59, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.354, + 0.569, + 0.364 + ], + "angle": 0, + "content": "LLMGA" + }, + { + "type": "image", + "bbox": [ + 0.592, + 0.282, + 0.681, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.603, + 0.354, + 0.672, + 0.365 + ], + "angle": 0, + "content": "Masked Image" + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.282, + 0.772, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.354, + 0.748, + 0.364 + ], + "angle": 0, + "content": "LLMGA" + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.367, + 0.321, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.437, + 0.296, + 0.446 + ], + "angle": 0, + "content": "DiffIR" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.367, + 0.411, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.437, + 0.404, + 0.447 + ], + "angle": 0, + "content": "DiffRIR (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.367, + 0.5, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.437, + 0.474, + 0.446 + ], + "angle": 0, + "content": "DiffIR" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.367, + 0.59, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.437, + 0.582, + 0.447 + ], + "angle": 0, + "content": "DiffRIR (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.592, + 0.367, + 0.681, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.437, + 0.655, + 0.446 + ], + "angle": 0, + "content": "DiffIR" + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.367, + 0.772, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.437, + 0.763, + 0.447 + ], + "angle": 0, + "content": "DiffRIR (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.459, + 0.787, + 0.487 + ], + "angle": 0, + "content": "Fig. 7: Visual comparison of image restoration methods. DiffRIR can alleviate the texture, contrast, and brightness disparities in inpainting & outpainting results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.487, + 0.554 + ], + "angle": 0, + "content": "Control SD using detailed language prompt or embedding?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.487, + 0.78 + ], + "angle": 0, + "content": "The results are shown in Fig. 5. We compare two approaches: GILL [29], which makes LLM estimate a fixed-sized embedding to control SD generation, and LLMGA Embedding, a variant of LLMGA where the language prompt is replaced with embedding, undergoing the same training process as LLMGA. The evaluation is conducted on MSCOCO by instructing these methods to generate images with the same prompts in multiple times in conversation form. (1) The quality of generated images (Fig. 5) in embedding-based methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "(i.e., GILL and LLMGA Embedding) deteriorate rapidly as the number of conversation turns increases. In contrast, our LLMGA remains unaffected. This discrepancy arises from the inherent noise present in the embeddings predicted by LLM. As the number of conversation turns rises, these generated embeddings" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.529, + 0.779, + 0.686 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.698, + 0.788, + 0.74 + ], + "angle": 0, + "content": "Table 5: T2I performance comparison of SD control schemes based on detailed language prompt and embedding." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.275, + 0.114, + 0.709, + 0.129 + ], + "angle": 0, + "content": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.187 + ], + "angle": 0, + "content": "Table 6: Datasets comparison. We conducted comparisons of FID on T2I and outpainting. The \\(\\checkmark\\) signifies the utilization of the complete dataset during training. Conversely, the absence of the \\(\\checkmark\\) indicates a reduction to only \\(10\\%\\) of the original datasets." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.199, + 0.787, + 0.294 + ], + "angle": 0, + "content": "
MethodPrompt RefinementInpainting &OutpaintingSimilar Image GenerationInstruction-based EditingT2IOutpainting
LLMGA1✓✓✓21.44602.6533
LLMGA2✓✓✓19.19143.1054
LLMGA3✓✓✓19.56982.6612
LLMGA4✓✓✓19.06422.8047
LLMGA5 (Ours)✓✓✓✓18.52342.4409
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.321, + 0.785, + 0.41 + ], + "angle": 0, + "content": "integrate with the preceding conversations, introducing even more noise. This poses challenges for the precise control of SD-generated content. Our LLMGA addresses this issue by mapping the embedding to the fixed language domain, effectively eliminating such noise. (2) Additionally, LLMGA Embedding also outperforms GILL, indicating that the prompt size used to guide SD generation should be adaptive in content, rather than a fixed size." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.412, + 0.786, + 0.547 + ], + "angle": 0, + "content": "Contribution of Training Data. To assess the impact of training data, we downsized one of the four training datasets in \\(\\mathrm{LLMGA}_5\\) to \\(10\\%\\) of its original magnitude, ensuring that LLMGA remains capable of furnishing responses in the prescribed format. The results are shown in Tab. 6. It is evident that prompt refinement, inpainting & outpainting, and instruction-based editing datasets enhance LLMGA's comprehension of image generation and editing properties, resulting in superior images. Moreover, comparing \\(\\mathrm{LLMGA}_3\\) and \\(\\mathrm{LLMGA}_5\\), we can see that engaging in similar image generation training further improves the performance of LLMGA in both generation and editing." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.569, + 0.357, + 0.584 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.599, + 0.786, + 0.84 + ], + "angle": 0, + "content": "LLM possesses an extensive reservoir of knowledge and powerful comprehension and reasoning capabilities. In this paper, we introduce a MLLM-based generation assistant (LLMGA), aiming to exploit LLM's capabilities in an interactive manner to facilitate more efficient and convenient image generation and editing. Compared to relying on LLM to predict a fixed-size embedding to control SD, we employ detailed generation prompts. These prompts prove to be more favorable for enhancing LLM's contextual comprehension and generating more accurate and rich content. To this end, we develop a two-stage training scheme and curate a dataset, including four parts: prompt refinement, similar image generation, inpainting & outpainting, and instruction-based editing. For the first stage, we train MLLM to understand the properties of image generation and editing, enabling it to give fitting responses. For the second stage, we optimize the SD unet to adapt to the generation prompt. Moreover, we propose a DM-based reference restoration network (DiffRIR) to mitigate disparities in texture, contrast, and brightness for image editing. Consequently, LLMGA can offer design suggestions and enhance results based on user's requests during interactions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.244, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LLMGA: Multimodal Large Language Model based Generation Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.788, + 0.21 + ], + "angle": 0, + "content": "This work was supported in part by the Research Grants Council under the Areas of Excellence scheme grant AoE/E-601/22-R." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.234, + 0.323, + 0.25 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.786, + 0.309 + ], + "angle": 0, + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.787, + 0.338 + ], + "angle": 0, + "content": "2. Austin, J., Johnson, D.D., Ho, J., Tarlow, D., van den Berg, R.: Structured denoising diffusion models in discrete state-spaces. NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.787, + 0.38 + ], + "angle": 0, + "content": "3. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.381, + 0.787, + 0.407 + ], + "angle": 0, + "content": "4. Bao, F., Nie, S., Xue, K., Cao, Y., Li, C., Su, H., Zhu, J.: All are worth words: A vit backbone for diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.409, + 0.787, + 0.449 + ], + "angle": 0, + "content": "5. Bao, F., Nie, S., Xue, K., Li, C., Pu, S., Wang, Y., Yue, G., Cao, Y., Su, H., Zhu, J.: One transformer fits all distributions in multi-modal diffusion at scale. arXiv preprint arXiv:2303.06555 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.451, + 0.787, + 0.478 + ], + "angle": 0, + "content": "6. Batzolis, G., Stanczuk, J., Schonlieb, C.B., Etmann, C.: Conditional image generation with score-based diffusion models. arXiv preprint arXiv:2111.13606 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.479, + 0.787, + 0.506 + ], + "angle": 0, + "content": "7. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.507, + 0.787, + 0.547 + ], + "angle": 0, + "content": "8. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.549, + 0.787, + 0.603 + ], + "angle": 0, + "content": "9. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality (March 2023), https://lmsys.org/blog/2023-03-30-vicuna/" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.605, + 0.787, + 0.645 + ], + "angle": 0, + "content": "10. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.647, + 0.787, + 0.686 + ], + "angle": 0, + "content": "11. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.688, + 0.787, + 0.729 + ], + "angle": 0, + "content": "12. Dai, X., Hou, J., Ma, C.Y., Tsai, S., Wang, J., Wang, R., Zhang, P., Vandenhende, S., Wang, X., Dubey, A., et al.: Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.73, + 0.787, + 0.756 + ], + "angle": 0, + "content": "13. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "14. Dong, R., Han, C., Peng, Y., Qi, Z., Ge, Z., Yang, J., Zhao, L., Sun, J., Zhou, H., Wei, H., et al.: Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "15. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Thompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.219, + 0.267, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.275, + 0.114, + 0.709, + 0.129 + ], + "angle": 0, + "content": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Fan, W.C., Chen, Y.C., Chen, D., Cheng, Y., Yuan, L., Wang, Y.C.F.: Frido: Feature pyramid diffusion for complex scene image synthesis. In: AAAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.217 + ], + "angle": 0, + "content": "17. Feng, Z., Zhang, Z., Yu, X., Fang, Y., Li, L., Chen, X., Lu, Y., Liu, J., Yin, W., Feng, S., et al.: Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "18. Ge, Y., Ge, Y., Zeng, Z., Wang, X., Shan, Y.: Planting a seed of vision in large language model. arXiv preprint arXiv:2307.08041 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.245, + 0.785, + 0.272 + ], + "angle": 0, + "content": "19. Gu, S., Chen, D., Bao, J., Wen, F., Zhang, B., Chen, D., Yuan, L., Guo, B.: Vector quantized diffusion model for text-to-image synthesis. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.313 + ], + "angle": 0, + "content": "20. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.341 + ], + "angle": 0, + "content": "21. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.37 + ], + "angle": 0, + "content": "22. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. JMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "23. Huang, S., Dong, L., Wang, W., Hao, Y., Singhal, S., Ma, S., Lv, T., Cui, L., Mohammed, O.K., Liu, Q., et al.: Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "24. Jiang, A.Q., Sablayrolles, A., Roux, A., Mensch, A., Savary, B., Bamford, C., Chaplot, D.S., Casas, D.d.l., Hanna, E.B., Bressand, F., et al.: Mixtral of experts. arXiv preprint arXiv:2401.04088 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "25. Ju, X., Zeng, A., Bian, Y., Liu, S., Xu, Q.: Direct inversion: Boosting diffusion-based editing with 3 lines of code. ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.785, + 0.508 + ], + "angle": 0, + "content": "26. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "27. Kim, G., Kwon, T., Ye, J.C.: Diffusionclip: Text-guided diffusion models for robust image manipulation. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "28. Kingma, D., Salimans, T., Poole, B., Ho, J.: Variational diffusion models. NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "29. Koh, J.Y., Fried, D., Salakhutdinov, R.: Generating images with multimodal language models. arXiv preprint arXiv:2305.17216 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.632 + ], + "angle": 0, + "content": "30. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.66 + ], + "angle": 0, + "content": "31. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.687 + ], + "angle": 0, + "content": "32. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.753, + 0.702 + ], + "angle": 0, + "content": "33. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.785, + 0.73 + ], + "angle": 0, + "content": "34. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.771 + ], + "angle": 0, + "content": "35. Meng, C., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "36. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.503, + 0.826 + ], + "angle": 0, + "content": "37. OpenAI: Gpt-4 technical report (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.827, + 0.785, + 0.84 + ], + "angle": 0, + "content": "38. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: ICCV (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.245, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LLMGA: Multimodal Large Language Model based Generation Assistant" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "39. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.229 + ], + "angle": 0, + "content": "40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.785, + 0.256 + ], + "angle": 0, + "content": "41. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.283 + ], + "angle": 0, + "content": "42. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.324 + ], + "angle": 0, + "content": "43. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.351 + ], + "angle": 0, + "content": "44. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.392 + ], + "angle": 0, + "content": "45. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.433 + ], + "angle": 0, + "content": "46. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.473 + ], + "angle": 0, + "content": "47. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.514 + ], + "angle": 0, + "content": "48. Shen, Y., Song, K., Tan, X., Li, D., Lu, W., Zhuang, Y.: Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.541 + ], + "angle": 0, + "content": "49. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.567 + ], + "angle": 0, + "content": "50. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.568, + 0.785, + 0.608 + ], + "angle": 0, + "content": "51. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.785, + 0.65 + ], + "angle": 0, + "content": "52. Sun, Q., Yu, Q., Cui, Y., Zhang, F., Zhang, X., Wang, Y., Gao, H., Liu, J., Huang, T., Wang, X.: Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.69 + ], + "angle": 0, + "content": "53. Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.731 + ], + "angle": 0, + "content": "54. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.771 + ], + "angle": 0, + "content": "55. Valevski, D., Kalman, M., Matias, Y., Leviathan, Y.: Unitune: Text-driven image editing by fine tuning an image generation model on a single image. arXiv preprint arXiv:2210.09477 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "56. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: ICCVW (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "57. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.276, + 0.115, + 0.709, + 0.129 + ], + "angle": 0, + "content": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "58. Wu, S., Fei, H., Qu, L., Ji, W., Chua, T.S.: Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.217 + ], + "angle": 0, + "content": "59. Xia, B., Zhang, Y., Wang, S., Wang, Y., Wu, X., Tian, Y., Yang, W., Timotfe, R., Van Gool, L.: Diffie2i: Efficient diffusion model for image-to-image translation. arXiv preprint arXiv:2308.13767 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "60. Xia, B., Zhang, Y., Wang, S., Wang, Y., Wu, X., Tian, Y., Yang, W., Van Gool, L.: Diffir: Efficient diffusion model for image restoration. ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.285 + ], + "angle": 0, + "content": "61. Xia, B., Zhang, Y., Wang, Y., Tian, Y., Yang, W., Timofte, R., Van Gool, L.: Knowledge distillation based degradation estimation for blind super-resolution. ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "62. Xue, Z., Song, G., Guo, Q., Liu, B., Zong, Z., Liu, Y., Luo, P.: Raphael: Text-to-image generation via large mixture of diffusion paths. arXiv preprint arXiv:2305.18295 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.369 + ], + "angle": 0, + "content": "63. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.411 + ], + "angle": 0, + "content": "64. Yu, L., Shi, B., Pasunuru, R., Muller, B., Golovneva, O., Wang, T., Babu, A., Tang, B., Karrer, B., Sheynin, S., et al.: Scaling autoregressive multi-modal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.438 + ], + "angle": 0, + "content": "65. Zhang, K., Mo, L., Chen, W., Sun, H., Su, Y.: Magicbrush: A manually annotated dataset for instruction-guided image editing. NeurIPS (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "66. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "67. Zhong, S., Huang, Z., Wen, W., Qin, J., Lin, L.: Sur-adapter: Enhancing text-to-image pre-trained diffusion models with large language models. In: ACM MM (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.785, + 0.536 + ], + "angle": 0, + "content": "68. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. TPAMI (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "69. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.577 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/LLMGA_ Multimodal Large Language Model based Generation Assistant/d91671aa-dc99-4c79-b4c9-9d98440c12f4_origin.pdf b/2024/LLMGA_ Multimodal Large Language Model based Generation Assistant/d91671aa-dc99-4c79-b4c9-9d98440c12f4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..302bbcaf94457bf03a08fe3cbaa59ab7805c2f3d --- /dev/null +++ b/2024/LLMGA_ Multimodal Large Language Model based Generation Assistant/d91671aa-dc99-4c79-b4c9-9d98440c12f4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09dd1c8f3f0c428d14c998e10d45676b3da0acfc060932b972820812f504439a +size 10054918 diff --git a/2024/LLMGA_ Multimodal Large Language Model based Generation Assistant/full.md b/2024/LLMGA_ Multimodal Large Language Model based Generation Assistant/full.md new file mode 100644 index 0000000000000000000000000000000000000000..81bfff4fcf9063a5928588709e3ad22663e0fc21 --- /dev/null +++ b/2024/LLMGA_ Multimodal Large Language Model based Generation Assistant/full.md @@ -0,0 +1,357 @@ +# LLMGA: Multimodal Large Language Model based Generation Assistant + +Bin Xia $^{1}$ , Shiyin Wang $^{2}$ , Yingfan Tao $^{2}$ , Yitong Wang $^{2}$ , and Jiaya Jia $^{1}$ + +1 The Chinese University of Hong Kong + +2 ByteDance Inc + +https://llmga.github.io/ + +Abstract. In this paper, we introduce a Multimodal Large Language Model-based Generation Assistant (LLMGA), leveraging the vast reservoir of knowledge and proficiency in reasoning, comprehension, and response inherent in Large Language Models (LLMs) to assist users in image generation and editing. Diverging from existing approaches where Multimodal Large Language Models (MLLMs) generate fixed-size embeddings to control Stable Diffusion (SD), our LLMGA provides a detailed language generation prompt for precise control over SD. This not only augments LLM context understanding but also reduces noise in generation prompts, yields images with more intricate and precise content, and elevates the interpretability of the network. To this end, we curate a comprehensive dataset comprising prompt refinement, similar image generation, inpainting & outpainting, and instruction-based editing. Moreover, we propose a two-stage training scheme. In the first stage, we train the MLLM to grasp the properties of image generation and editing, enabling it to generate detailed prompts. In the second stage, we optimize SD to align with the MLLM's generation prompts. Additionally, we propose a reference-based restoration network to alleviate texture, brightness, and contrast disparities between generated and preserved regions during inpainting and outpainting. Extensive results show that LLMGA has promising generation and editing capabilities and can enable more flexible and expansive applications in an interactive manner. + +Keywords: Interactive generation and editing $\cdot$ LLM $\cdot$ Diffusion Model + +# 1 Introduction + +Artificial Intelligence Generated Content (AIGC) has witnessed remarkable advancements, particularly propelled by the evolution of large language models (LLMs) [8, 10, 54] for text generation and diffusion models (DMs) [21, 42] for image generation. LLMs, in particular, have received considerable acclaim for their exceptional ability to comprehend, reason, make decisions, possess extensive knowledge, and generate text with unparalleled accuracy and fluency. + +Recent studies have begun delving deeper into Multimodal Large Language Models (MLLMs) [1,69] built upon LLMs, aiming to empower LLMs to comprehend inputs extending beyond text. For example, BLIP-2 [30] and LLaVA [33] + +![](images/9e911b04cc38e7b9e961d81ffec92e55e0cf1f64fdcd19496e0c4d3167263829.jpg) + +![](images/bad945006d7a95d2b4fb1e61f9fc57c1b0f858a8f474ec7e9d2dd97684d30b66.jpg) + +![](images/c6aa2e25d53add88442e917afb5fe46ff0d46775bfe210dc6fd635fe442f1036.jpg) + +![](images/edc64691018c7e8e081f9a9db223816f40e255a2683ff23a295fc36a8dccb3cc.jpg) + +![](images/51323782904c03d52c3a1a5ba51ec6489ec41152a06c08e989e8c100c6efe123.jpg) +Fig. 1: Some examples of LLMGA for assisting in image generation and editing. (1) T2I generation. LLMGA can refine the user's generation prompt to produce more vivid and vibrant images. (2) Similar image generation. LLMGA can understand the component and layout of the input images and generate a similar image. (3) Inpainting & Outpainting. LLMGA can provide detailed generation prompts based on user preferences and input images. (4) Instruction based editing. LLMGA can understand user instructions and realize accurate editing. (5) Interactive image generation and editing exemplify the comprehensive capabilities of LLMGA. Users can design satisfactory images by engaging in interactions with LLMGA, leveraging its vast knowledge. + +![](images/e9325f1b8254c5a5da9fa606c8b65118ca571514c3bc01c3b4b329e299a7ace7.jpg) + +![](images/7196089690010dc3ce9acb2ce6d66a849548ef96d897bc1d86b8455f88583739.jpg) + +![](images/ee491b89156c24e2a42b8f0e84bc42a93043ee3b1525a3d9c8020413879d1880.jpg) + +employ visual encoders to transform images into input embeddings, enabling them to be used as prompts alongside text input for the LLM, thus achieving LLMs with the visual modality. Furthermore, recent works focused on extending the capabilities of LLMs to generate multimodal outputs. For example, GILL [29] involves instructing LLMs to predict fixed-size visual embeddings aligned with CLIP [40] space to control the Stable Diffusion [42] (SD) for image generation. + +However, existing works [29, 52] merely focus on enabling LLM to output images but do not aim to assist users in generating or editing images to enhance quality. In this paper, we aim to develop a Multimodal Large Language Model-based Generation Assistant (LLMGA) to better assist image generation models, making them more user-friendly and capable of producing high-quality images. In contrast to certain methods [29, 52] that leverage MLLMs to predict fixed-size visual embeddings for implicit SD control, our approach is straightforward. We guide the generation of SD using detailed language prompts from MLLM based on five reasons. (1) The embeddings predicted by the MLLM are often filled with noise. This can be filtered out by mapping them to a fixed language domain, enabling precise control of SD. (2) Detailed language prompts can make the network more transparent and interactive, allowing users to understand MLLM's thoughts for generating images. (3) MLLM is pre-trained on vast textual datasets. Explicit language prompts rather than implicit embeddings are more advantageous for MLLM to generate prompts and comprehend context. (4) + +Dynamic-sized language prompt facilitates the addition of generation requests during interactions. (5) Training is more simple and efficient. + +However, we face several challenges: (1) MLLM may reject the execution of generation instructions due to its nature as a language assistant. (2) MLLM lacks a comprehensive understanding of image generation and editing, and cannot provide an accurate and detailed generation prompt. (3) Determining which part of texts generated by MLLM to guide SD generation. (4) SD's CLIP encode only 75 tokens. Additionally, SD is trained on short captions, whereas our LLMGA typically generates detailed prompts exceeding 150 tokens. This discrepancy poses a challenge for SD in following the detailed prompt of LLMGA. + +To this end, we have devised a two-stage training scheme. First, we construct a training dataset: prompt refinement, similar image generation, inpainting & outpainting, and visual question answering. We then train LLMGA on these four datasets to cultivate four fundamental capabilities: (1) For concise user prompts, LLMGA can refine the generation of intricate details, encompassing attire, background, and characters. (2) LLMGA can precisely regenerate an image it observes. (3) LLMGA can generate or refine prompts for inpainting & outpainting based on its understanding of the image. (4) LLMGA can generate accurate prompts for instruction-based editing according to users' requirements and given images. Additionally, we make LLMGA use special symbols and to distinguish generation prompts and responses. In the second stage, we freeze the parameters of LLMGA's MLLM and initiate joint training with the SD. This process enables the SD to acclimate to the detailed prompt produced by the MLLM. Notably, when the input token count exceeds 75, we iteratively apply the CLIP [40] encoder to the surplus tokens. + +Moreover, we have identified noticeable disparities in texture, contrast, and brightness between the newly generated and preserved sections in inpainting & outpainting. Therefore, we propose a Diffusion-based Reference Restoration Network (DiffRIR). Specifically, aside from images generated by SD, we add masked images as reference inputs into DiffRIR. This enables the DiffRIR to refer to the texture, contrast, and brightness of the preserved regions for restoration. Additionally, we introduce perturbations to contrast and brightness during training, enabling DiffRIR to correct contrast and brightness disparities in the images. + +As shown in Fig. 1, LLMGA is a unified and interactive framework for image generation and editing, endowed with a wide array of capabilities: (1) LLMGA leverages its extensive world knowledge and powerful reasoning abilities to assist image generation and editing and significantly improve results. (2) LLMGA can be integrated with external plugins, like ControlNet. (3) Most importantly, users can interact with LLMGA to design satisfying images in a more convenient, flexible, and enjoyable way. In summary, our contributions are as follows: + +- We construct a training dataset, including four parts: prompt refinement, similar image generation, inpainting & outpainting, and instruction-based editing. This enhances LLMGA's comprehension of generation and editing tasks while standardizing response formats. +- We proposed a restoration network DiffRIR, which introduces reference images and training perturbations to contrast and brightness. DiffRIR can alleviate texture, contrast, and brightness discrepancies between newly generated and preserved regions for edited images. +- Open-source. The following assets are released: the generated data, the codebase for model training, the model checkpoint, and a demo. + +# 2 Related Work + +Diffusion Model. Diffusion Models (DMs) [2,4,7,13,19,21,22,28,35,38,49,51] have achieved remarkable results in image generation. DMs adopt a parameterized Markov chain to optimize the lower variational bound on the likelihood function. In this way, it can generate realistic images from Gaussian noise. After that, several DM methods [3,5,6,12,16,17,26,27,34,39,44,45,55,59,60,62] have been tailored to enhance the text-to-image (T2I) generation and editing. Notably, GLIDE [36] pioneered the incorporation of text features into transformer blocks during the denoising process. Subsequently, DALL-E [41], Imagen [46], and Stable Diffusion [42] have made substantial strides in improving T2I generation. Subsequently, some works [43,66,67] introduced conditioning controls to the DMs to facilitate a more convenient and precise manipulation of the generation process. Overall, enhancing the user-friendliness of DMs is a key focus within the community. In this paper, we introduce LLMGA, leveraging the extensive knowledge and powerful reasoning capabilities of LLM to facilitate users in achieving more easily attainable and satisfactory image designs. + +Multimodal Large Language Models. Recently, LLMs have undeniably made profound impacts and revolutions within the entire AI community and beyond. For example, exemplary LLMs, such as ChatGPT and GPT4 [37], have showcased remarkable abilities in comprehension, reasoning, responses, and knowledge reservoirs. Subsequently, a range of LLMs [11], including Vicuna [9], LLaMA [54], and Alpaca [53] have been released as open-source models, substantially propelling advancements of the community. + +Afterward, the community began focusing on the development of the Multimodal Large Language Model [14, 15, 18, 23, 63, 64, 69]. They aim to enable LLMs to comprehend both images and text and provide textual responses. For instance, Flamingo [1] encodes images and feeds them into the LLM's attention layer. BLIP-2 [30] employs Q-Former to encode input images into queries. Additionally, LLaVA [33] leverages CLIP [40] to encode images into visual embeddings, which are then concatenated with text embeddings. + +Recent concurrent works, such as Next-GPT [58], have extended the capabilities to encompass audio and video modalities. Moreover, Visual-ChatGPT [57] and HuggingGPT [48] make LLMs act as agents capable of invoking various pretrained visual models to achieve MLLM. However, these works focus on making + +![](images/ac16f1c191015e38321bbbc44c46d40be6e0f8fa0223e8990c44fa1ded187dd5.jpg) +Fig. 2: The overview of LLMGA. (a) In the first training stage, we train the MLLM to produce generation prompts based on provided instructions. Moreover, we construct a training dataset including four categories: prompt refinement, similar image generation, inpainting & outpainting, and instruction-based editing. (b) In the second training stage, we optimize SD to adapt to the detailed generation prompts from MLLM. (c) In the inference stage, LLMGA can respond to user queries and assist in various tasks, such as image generation, inpainting & outpainting, and editing. + +![](images/89b4e111b668e96c8e2ed4ec0e9287c59b03a9f2210e5350cc4d65c448bde191.jpg) + +LLM determine the combined invocation of modules (such as detection, recognition, and generation) to fulfill user requirements. However, these methods are not tailored for generation and editing and use concise prompts that lack the capability to enhance results. Thus, we propose LLMGA, which is designed to assist with various image generation and editing tasks. It can achieve satisfactory results by strong reasoning capability and flexible interaction with users. + +# 3 Methodology + +# 3.1 Overview of LLMGA + +In this paper, we aim to design a MLLM-based Generation Assistant (LLMGA). Our LLMGA produces detailed language-based generation prompts to control SD rather than predicting fixed-sized visual embeddings [29, 52] to govern SD. This has five advantages: (1) Visual embeddings contain noise, and mapping them to the language domain can filter out this noise, enabling precise SD control. (2) Language-based generation prompts facilitate users in understanding the LLMGA's thoughts, enhancing interaction. (3) Dynamic-sized language-based generation prompt enables the addition of generation requests. (4) MLLM is pre-trained on textual datasets. Language prompts rather than implicit visual embeddings are more advantageous for MLLM to generate accurate prompts and comprehend context. (5) Training is simpler and more efficient. + +However, we need to address several issues: (1) As a language assistant, MLLM may decline the execution of generation instructions. (2) MLLM lacks + +a nuanced understanding of image generation and editing, and cannot produce precise and detailed generation prompts. (3) MLLM needs to decide which part of the output text serves as generation prompts to guide generation. (4) SD's CLIP encode only 75 tokens. Moreover, SD is primarily trained on short captions, while detailed prompts generated by LLMGA may exceed 150 tokens. This disparity makes it hard for SD to understand the instructions from LLMGA. To address the aforementioned challenges, we construct a training dataset and design two-stage training schemes to train the MLLM (Sec. 3.2) and SD (Sec. 3.3). + +The network structure and pipeline of LLMGA are illustrated in Fig. 2. Specifically, as shown in Fig. 2 (a) and (c), the images $\mathbf{I}_{input}$ are encoded into image embeddings by CLIP vision encoder and a projection layer. Subsequently, the image embedding is concatenated with the text embedding and fed into the LLM to obtain text output $\mathbf{T}_{output}$ . This process can be formulated as: + +$$ +\mathbf {T} _ {\text {o u t p u t}} = \mathcal {F} _ {M L L M} \left(\mathbf {T} _ {\text {i n p u t}}, \mathbf {I} _ {\text {i n p u t}}\right), \tag {1} +$$ + +where $\mathbf{T}_{input}$ indicates the input text instructions from users. It is notable that $\mathcal{F}_{MLLM}$ can process only $\mathbf{T}_{input}$ as input. + +The text output $\mathbf{T}_{output}$ can comprise two components: text response $\mathbf{T}_R$ and generation prompt $\mathbf{T}_P$ . To distinguish between $\mathbf{T}_R$ and $\mathbf{T}_P$ , we adopt new special tokens, i.e., $<\text{gen\_img}>$ and $$ , to encompass $\mathbf{T}_P$ . + +We present $\mathbf{T}_R$ as the immediate text response to users. Concurrently, $\mathbf{T}_P$ is further fed into the subsequent SD to guide T2I generation (Eq. 2), inpainting & outpainting (Eq. 3), and instruction-based image editing (Eq. 4). + +$$ +\mathbf {I} _ {g} = \mathcal {F} _ {T 2 I} (\mathbf {T} _ {P}, \mathbf {Z}), \tag {2} +$$ + +$$ +\mathbf {I} _ {p} = \mathcal {F} _ {\text {P a i n t}} (\mathbf {I} _ {\text {i n p u t}}, \mathbf {I} _ {\text {m a s k}}, \mathbf {T} _ {P}, \mathbf {Z}), \tag {3} +$$ + +$$ +\mathbf {I} _ {e} = \mathcal {F} _ {\text {E d i t}} \left(\mathbf {I} _ {\text {i n p u t}}, \mathbf {T} _ {P}\right), \tag {4} +$$ + +Where $\mathbf{Z}$ denotes the random Gaussian noise. Furthermore, to ensure the encoding of all $\mathbf{T}_P$ for SD, we iteratively run the CLIP text encoder until all prompts are encoded. For inpainting & outpainting, except the input image $\mathbf{I}_{input}$ , an additional mask $\mathbf{I}_{mask}$ are essential inputs for the inpainting SD to specify the region requiring generation. For instruction-based image editing, $\mathcal{F}_{Edit}$ performs inversion [25, 50] and prompt-to-prompt [20] based on T2I SD. + +# 3.2 MLLM Training + +As described in Sec. 3.1, original MLLMs are specifically designed and trained as language assistants, but they lack the proficiency to assist in image generation, and editing. Notably, considering the input of images and the performance of open-sourced MLLMs, employing few-shot learning to guide the model to achieve the desired results proves challenging and inefficient. Therefore, it is crucial to train the MLLM to serve as proficient assistants in image generation and editing + +tasks, understanding the expected response formats and enhancing their comprehension of image generation and editing properties. To this end, as depicted in Fig. 2 (a), we construct a training dataset consisting of four categories: + +(1) Prompt Refinement. We establish this dataset to cultivate the prompt refinement ability of the MLLM. Specifically, we utilize GPT4-V to furnish detailed descriptions of images in MSCOCO [31]. These detailed descriptions, along with the original MSCOCO brief descriptions, constitute a training text pair. During training, we input the brief MSCOCO captions and randomly select and append a generation instruction. When a generation instruction is included in the prompt, we add and on the later description. +(2) Similar Image Generation. We select images from the MSCOCO dataset along with corresponding detailed descriptions generated by GPT4-V to create the Similar Image Generation dataset. During training, we input the images along with a generation instruction. In cases where a generation instruction is provided, we add and on the subsequent description. +(3) Inpainting & Outpainting. We use pairs of detailed descriptions and images from the Similar Image Generation dataset. During training, we input masked images with inpainting or outpainting instructions. Besides, we include and on the subsequent description. +(4) Instruction-based editing. We fine-tune Mixtral-8x7B [24] to enable it to generate editing data based on detailed descriptions from MSCOCO. Subsequently, we clean the generated data. During training, we can input images or original captions and provide corresponding editing instructions, aiming to train LLMGA to output the target caption. Additionally, we include and tags on the target caption during training. + +During training, alongside the above data, we integrated the image designing and Alpaca [53] dataset, to enhance LLMGA's question-answering (QA). Notably, we excluded certain visual multimodal incompatible question-answer pairs from it. Furthermore, we incorporated the LLaVA v1.5 mix665k dataset [32] to endow LLMGA with Visual Question Answering (VQA) capabilities. We provide more details in the supplementary material. + +As illustrated in Fig. 2 (a), we freeze the CLIP vision encoder and optimize the projection layer and LLM. The model is trained end-to-end using the autoregressive cross-entropy loss $(\mathcal{L}_{MLLM})$ for text generation. Given the ground-truth targets $\mathbf{T}_{GT}$ , this loss can be formulated as: + +$$ +\mathcal {L} _ {M L L M} = \mathbf {C E} \left(\mathbf {T} _ {\text {o u t p u t}}, \mathbf {T} _ {G T}\right). \tag {5} +$$ + +# 3.3 Stable Diffusion Training + +As described in Sec. 3.1, the original SD's CLIP text encoder only encodes 75 tokens, which cannot handle the entire MLLM's generation prompt. Moreover, the original SD is trained on brief captions, which cannot fully understand the generation prompts. Thus, we repeatedly use the CLIP text encoder to encode all instances of $\mathbf{T}_P$ for SD. Besides, we train the T2I SD model and the inpainting SD model, respectively. For both generation and inpainting & outpainting tasks, + +the generation prompts of MLLM are detailed descriptions of images. Therefore, during training, we can instruct MLLM to provide a detailed description $\mathbf{T}_P$ for images from the LAION-Aesthetics [47] and MSCOCO datasets. Subsequently, $\mathbf{T}_P$ is fed into T2I SD or inpainting SD for joint training. Notably, we only optimize the SD unet while freezing the parameters of other networks. To accelerate the training, we record the prompt $\mathbf{T}_P$ of MLLM to avoid redundant calculations. The model is trained using SD loss (Eq. 6). For instruction-based editing, we directly adopt the pre-trained T2I SD model. + +$$ +\mathcal {L} _ {S D} = \mathbb {E} _ {\mathbf {Z} _ {t}, \mathbf {C}, \epsilon , t} \left(\| \epsilon - \epsilon_ {\theta} \left(\mathbf {Z} _ {t}, \mathbf {C}\right) \| _ {2} ^ {2}\right), \tag {6} +$$ + +where $\mathbf{Z}_t = \sqrt{\overline{\alpha}_t}\mathbf{Z}_0 + \sqrt{1 - \overline{\alpha}_t}\epsilon$ represents the noised feature map at timestep $t$ . Ground truth images are encoded into latent space to derive $\mathbf{Z}_0$ . Here, $\epsilon \in \mathcal{N}(0,\mathbf{I})$ represents Gaussian noise, and $\epsilon_{\theta}$ refers to the SD unet. $\mathbf{C}$ indicates the conditional information. For T2I generation, $\mathbf{C}$ is $\mathbf{T}_P$ . For inpainting and outpainting, $\mathbf{C}$ contains $\mathbf{T}_P$ , the mask, and the VAE-encoded masked image. + +# 3.4 Restoration Network Training + +For SD inpainting & outpainting, we observed noticeable disparities between the preserved and newly generated regions in the edited images. To enhance the consistency between the newly generated and the preserved regions, we introduce a reference-based restoration scheme. Specifically, existing restoration methods [56, 60, 61] take low-quality (LQ) images as input and produce high-quality (HQ) images, but they often do not leverage the preserved information from the given masked image. Different from them, we concatenate the LQ image $I_{LQ}$ and the masked image, i.e., $(1 - \mathbf{I}_{mask})\mathbf{I}_{GT}$ , as inputs and feed them into the restoration model $\mathcal{F}_R$ . This process can be formulated as: + +$$ +\mathbf {I} _ {H Q} = \mathcal {F} _ {R} (\operatorname {c o n c a t} \left(\mathbf {I} _ {L Q}, (1 - \mathbf {I} _ {\text {m a s k}}) \mathbf {I} _ {G T}\right)). \tag {7} +$$ + +To further mitigate the brightness and contrast disparities, we introduced additional color degradation (i.e., random brightness and contrast disturbance) into the training process of the restoration model, which is formulated as: + +$$ +\mathcal {D} _ {2} (\mathbf {x}) = c _ {1} \mathbf {x} + c _ {2}, \tag {8} +$$ + +$$ +\mathbf {I} _ {L Q} = \mathcal {D} _ {2} \left(\mathcal {D} _ {1} \left(\mathbf {I} _ {G T}\right)\right), \tag {9} +$$ + +where $\mathcal{D}_2$ indicates contrast and brightness disturbance, $\mathbf{x}$ denotes the input image. $c_{1}$ is the contrast gain, randomly varied within the range of [0.94, 1.06], while $c_{2}$ is the brightness bias, randomly varied within the range of $[-0.05, 0.05]$ . $\mathcal{D}_1$ represents the real-world degradation process used in restoration model [56, 60, 61]. $\mathbf{I}_{GT}$ is ground truth image, and $\mathbf{I}_{LQ}$ is LQ image. Here, we adopt the network structure of the SOTA restoration model DiffIR [60] and apply our schemes to it to obtain our Diffusion-based Reference Restoration Network (DiffRIR). + +# 4 Experiments + +# 4.1 Implementation Details + +For the first stage of training, we train the MLLM (including projection and LLM) on $8 \times \mathrm{A}100$ . The batch size is set to 128. Besides, the training datasets include VQA (LLaVA v1.5 mix665k), QA, prompt refinement, similar image generation, inpainting & outpainting, and instruction-based editing. + +For the second stage of training, we adopt the Stable Diffusion 1.5 (SD1.5) as the initial image generation or inpainting & outpainting model. We train these models on $8 \times \mathrm{A}100$ . The batch size is set to 32. + +For the restoration network, we train DiffRIR on $8 \times \mathrm{V}100$ . The batch size is set to 64. Please check more details in the supplementary material. + +# 4.2 Experimental Results + +Evaluation on T2I Generation. The results are shown in Tab. 1. Notably, SD1.5 is the original Stable Diffusion 1.5 while SD1.5-ft indicates the finetuned SD1.5 in LLMGA. We also compare our LLMGA with the recently proposed multimodal generative model GILL [29]. (1) Comparing SD1.5, our LLMGA-7B achieves notable 5.4847 FID and 5.05 IS improvements, underscoring the effectiveness of LLMGA. (2) Moreover, + +our LLMGA-7B significantly outperforms GILL. (3) In the 3rd and 4th rows of Tab. 1, we use LLMGA-7B to refine the short MSCOCO caption to the detailed generation prompt and send it to the SD1.5 and SD1.5-ft, respectively. Our LLMGA-7B with SD1.5-ft achieves significant 5.0712 FID and 3.92 IS improvements, respectively. This demonstrates our second-stage training can help SD1.5-ft to better follow detailed prompts from MLLM. (4) Comparing the 4th and 6th rows of Tab. 1, LLMGA-13B exhibits better performance than LLMGA-7B due to its superior reasoning ability. + +The qualitative results are shown in Fig. 3. (1) LLMGA excels in refining prompts by incorporating details to generate visually rich and pleasing images. For example, in the first row, LLMGA crafts a battle attire for the husky, depicting engaging scenarios of battling monsters. This makes user usage more convenient, eliminating the need for them to think about generating image details themselves. (2) LLMGA can leverage its extensive knowledge base to generate images, even for concepts users may not be familiar with, like a space elevator. + +Evaluation on Instruction-based Editing. For instruction-based editing, we utilize LLMGA to provide a detailed description of the edited image based on the input image and user editing instructions. We then employ Direct Inversion [25] + +Table 1: Quantitative comparison on T2I generation on the MSCOCO [31] dataset. + +
MethodFID↓IS↑
SD1.524.008135.99
GILL [29]25.112334.20
LLMGA-7b(SD1.5)23.594637.12
LLMGA-7b(SD1.5-ft, Ours)18.523441.04
LLMGA-13b(SD1.5)23.582837.58
LLMGA-13b(SD1.5-ft, Ours)18.406341.16
+ +![](images/6c5ca517e99e665420423b72be44130bd6d43d781b60bd96ed1459b91812afc6.jpg) +Fig. 3: T2I visual comparison. LLMGA can produce accurate and high-quality results. + +Table 2: Quantitative comparison for image editing on MagicBrush test set [65]. + +
MethodsL1↓L2↓CLIP-I↑DINO↑CLIP-T↑
InstructPix2Pix0.11970.04160.84420.72520.2909
MagicBrush0.06470.02240.92930.89130.2979
LLMGA (Ours)0.08140.02180.89360.87680.3137
+ +and prompt-to-prompt [20] methods to obtain the edited image. (1) The quantitative results are presented in Tab. 2. Notably, our LLMGA did not train SD for image editing like InstructPix2Pix [7] and MagicBrush [65]. However, our zero-shot performance on the MagicBrush test set surpassed that of InstructPix2Pix, achieving performance similar to that of MagicBrush. (2) Additionally, our LLMGA offers a superior user experience in interactive editing, allowing image modifications to be carried out conversationally. In contrast, InstructPix2Pix only supports input via instructions and output as images. Visualized results are shown in Fig. 4, by leveraging the powerful reasoning capabilities of LLM, our LLMGA can provide more accurate and reasonable editing results. + +Evaluation on Inpainting and Outpainting. The results are shown in Tab. 3. (1) Comparing the 1st and 3rd rows of Tab. 3, our LLMGA-7B achieves significant improvements of 2.5681 FID over the SD1.5 in outpainting under wide masks. (2) In the 2nd and 3rd rows of Tab. 3, we make LLMGA imagine the complete generation prompts for the given masked images, which are then input into the later SD. Notably, our LLMGA-7B (with SD1.5-ft) demonstrates a significant FID improvement over LLMGA-7B (with SD1.5) in both outpainting and inpainting. This demonstrates the second stage of training makes SD better follow the prompts from MLLM. (3) Comparing the 3rd and 5th rows, our LLMGA-13B outperforms LLMGA-7B due to its superior reasoning capabilities. + +![](images/48c3f14583c3f198ad34333c853599803b868249c6aa9821c67403fb3f3e8411.jpg) +Fig. 4: Visual comparison on instruction-based editing. + +Table 3: Quantitative comparison for outpainting & inpainting on Places [68]. + +
MethodOutpaintingInpainting
NarrowMasksWideMasksNarrowMasksWideMasks
FID↓LPIPS↓FID↓LPIPS↓FID↓LPIPS↓FID↓LPIPS↓
SD1.52.01670.22835.00900.37341.07950.12361.28550.1434
LLMGA-7b (SD1.5)1.55300.22633.26300.36881.06920.12331.09830.1427
LLMGA-7b (SD1.5-ft, Ours)1.29730.22152.44090.36160.80270.11710.98070.1405
LLMGA-13b (SD1.5)1.56310.22633.08450.36791.03260.12281.09780.1426
LLMGA-13b (SD1.5-ft, Ours)1.21600.22102.36630.36090.79920.11660.97800.1400
+ +The qualitative results are shown in Fig. 5. We can see that LLMGA can deduce and imagine complete images based on masked input images. For example, in the 3rd row of Fig. 5, LLMGA can infer the presence of wind turbines on the mountain based on the given environment. Overall, LLMGA's powerful reasoning capability and extensive knowledge can assist users in conveniently making accurate and visually pleasing inpainting and outpainting. + +Evaluation on ControlNet. Our LLMGA demonstrates exceptional scalability, enabling integration with external plugins like ControlNet [66]. Here, we utilize LLMGA to create detailed prompts derived from input images and user requirements, working alongside ControlNet to guide image generation. As depicted in Fig. 6, our LLMGA significantly enhances the diversity and richness of outcomes in picture reference-guided image generation. Furthermore, more external plugins can also be integrated into the interactive framework of LLMGA, + +![](images/f61f8cd5dc4c567f736bd55e7b552c52df09cc2f898d5f77908e49f71cfc9897.jpg) +Fig. 5: Visual comparison on inpainting and outpainting. + +![](images/39cbd39d76c37d0e07e7cebf169c819561ff40536a55f56fa04e2d503d2f016f.jpg) +Fig. 6: Visualization of LLMGA plus ControlNet. Our LLMGA can enhance the details in generated images, producing visually pleasing images. + +combining with LLMGA's reasoning design capabilities and all previous functions to allow for a broader range of creative and engaging applications. + +Evaluation on Image Restoration. The results are shown in Tab. 4. For comparisons, we validate DiffIR [60] and our DiffRIR on the outpainting image generated by LLMGA. (1) Comparing the 2nd and 3rd rows of Tab. 4, it is evident that introducing a reference scheme can significantly improve restoration performance. (2) When comparing the 3rd and 4th rows of Tab. 4, it can be observed that introducing color degradation helps alleviate the bright and contrast distortion caused by SD. (3) Comparing the 1st and 4th rows of Tab. 4, our DiffRIR yields significant improvement, validating the effectiveness of DiffRIR. + +As shown in Fig. 7, our DiffRIR (i.e., $\mathrm{DiffRIR}_2$ in Tab. 4) can alleviate the texture, brightness, and contrast discrepancies, and generate realistic details. + +Table 4: Quantitative comparison on image restoration for Places [68] outpainting. + +
MethodColorFID↓LPIPS↓
Reference Degradation
LLMGA-7BXX2.44090.3616
LLMGA+DiffIR [60]XX2.35870.3612
LLMGA+DiffRIR1✓X2.29930.3609
LLMGA+DiffRIR2 (Ours)✓✓2.26870.3607
+ +![](images/d61d967b941e41a79902a688b01a89953ded62a49980f7f9c6a0d288fa24ad7a.jpg) + +![](images/1eb7e259f0810d9986e8570445abe9d6da67065a241ea358e8be52203e35ebab.jpg) +LLMGA + +![](images/c9ad9fd42e21199e45b8b081645cd824c84eb78466adb47ada4112f7db65db87.jpg) +Masked Image + +![](images/ff9e7e78133ad38a7383ed810676b6498be9903d3e35ac82c9217cc432aa7f2c.jpg) + +![](images/75a3a4b1c266273737fa5c1332f15c326d5ada07257358fbe6372715fd2da9c0.jpg) + +![](images/9c53a09691581ae00e8ddc042ab7d668b9c619b726ca9dbe8057f83c9e108b42.jpg) + +![](images/bfd0c6c1db5eb97d5b7db4ef18f14322762dd56b98557a47b15655ecfc1cca97.jpg) +Masked Image +DiffIR +Fig. 7: Visual comparison of image restoration methods. DiffRIR can alleviate the texture, contrast, and brightness disparities in inpainting & outpainting results. + +![](images/c4e0d036b74494b9d7b8132081ffe26561ca341b31efb0931527214ac7eced22.jpg) +DiffRIR (Ours) + +![](images/39ee74a6f76f0311326040847369f1ed36efc9463b8d7d0f9f31d580740f6c61.jpg) +DiffIR + +![](images/7fe13a363b7c88b0aba6269d13a976fab0287419fb6c00a23b21be4a5e77ff98.jpg) +LLMGA +DiffRIR (Ours) + +![](images/c099d4e7b008a726198e3aee2f5912569ba7db41e4f23094f58749cf7cbbb948.jpg) +Masked Image +DiffIR + +![](images/368b233adacba37700fedfdae6dd1fbd1200efdb273bebf78a4438ac48c87174.jpg) +LLMGA +DiffRIR (Ours) + +Control SD using detailed language prompt or embedding? + +The results are shown in Fig. 5. We compare two approaches: GILL [29], which makes LLM estimate a fixed-sized embedding to control SD generation, and LLMGA Embedding, a variant of LLMGA where the language prompt is replaced with embedding, undergoing the same training process as LLMGA. The evaluation is conducted on MSCOCO by instructing these methods to generate images with the same prompts in multiple times in conversation form. (1) The quality of generated images (Fig. 5) in embedding-based methods + +(i.e., GILL and LLMGA Embedding) deteriorate rapidly as the number of conversation turns increases. In contrast, our LLMGA remains unaffected. This discrepancy arises from the inherent noise present in the embeddings predicted by LLM. As the number of conversation turns rises, these generated embeddings + +![](images/031046ecd895a4ef2f28532213d538f8fe3ee36e92da00499658e49e761ce098.jpg) +Table 5: T2I performance comparison of SD control schemes based on detailed language prompt and embedding. + +Table 6: Datasets comparison. We conducted comparisons of FID on T2I and outpainting. The $\checkmark$ signifies the utilization of the complete dataset during training. Conversely, the absence of the $\checkmark$ indicates a reduction to only $10\%$ of the original datasets. + +
MethodPrompt RefinementInpainting &OutpaintingSimilar Image GenerationInstruction-based EditingT2IOutpainting
LLMGA1✓✓✓21.44602.6533
LLMGA2✓✓✓19.19143.1054
LLMGA3✓✓✓19.56982.6612
LLMGA4✓✓✓19.06422.8047
LLMGA5 (Ours)✓✓✓✓18.52342.4409
+ +integrate with the preceding conversations, introducing even more noise. This poses challenges for the precise control of SD-generated content. Our LLMGA addresses this issue by mapping the embedding to the fixed language domain, effectively eliminating such noise. (2) Additionally, LLMGA Embedding also outperforms GILL, indicating that the prompt size used to guide SD generation should be adaptive in content, rather than a fixed size. + +Contribution of Training Data. To assess the impact of training data, we downsized one of the four training datasets in $\mathrm{LLMGA}_5$ to $10\%$ of its original magnitude, ensuring that LLMGA remains capable of furnishing responses in the prescribed format. The results are shown in Tab. 6. It is evident that prompt refinement, inpainting & outpainting, and instruction-based editing datasets enhance LLMGA's comprehension of image generation and editing properties, resulting in superior images. Moreover, comparing $\mathrm{LLMGA}_3$ and $\mathrm{LLMGA}_5$ , we can see that engaging in similar image generation training further improves the performance of LLMGA in both generation and editing. + +# 5 Conclusion + +LLM possesses an extensive reservoir of knowledge and powerful comprehension and reasoning capabilities. In this paper, we introduce a MLLM-based generation assistant (LLMGA), aiming to exploit LLM's capabilities in an interactive manner to facilitate more efficient and convenient image generation and editing. Compared to relying on LLM to predict a fixed-size embedding to control SD, we employ detailed generation prompts. These prompts prove to be more favorable for enhancing LLM's contextual comprehension and generating more accurate and rich content. To this end, we develop a two-stage training scheme and curate a dataset, including four parts: prompt refinement, similar image generation, inpainting & outpainting, and instruction-based editing. For the first stage, we train MLLM to understand the properties of image generation and editing, enabling it to give fitting responses. For the second stage, we optimize the SD unet to adapt to the generation prompt. Moreover, we propose a DM-based reference restoration network (DiffRIR) to mitigate disparities in texture, contrast, and brightness for image editing. Consequently, LLMGA can offer design suggestions and enhance results based on user's requests during interactions. + +# Acknowledgements + +This work was supported in part by the Research Grants Council under the Areas of Excellence scheme grant AoE/E-601/22-R. + +# References + +1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. NeurIPS (2022) +2. Austin, J., Johnson, D.D., Ho, J., Tarlow, D., van den Berg, R.: Structured denoising diffusion models in discrete state-spaces. NeurIPS (2021) +3. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022) +4. Bao, F., Nie, S., Xue, K., Cao, Y., Li, C., Su, H., Zhu, J.: All are worth words: A vit backbone for diffusion models. In: CVPR (2023) +5. Bao, F., Nie, S., Xue, K., Li, C., Pu, S., Wang, Y., Yue, G., Cao, Y., Su, H., Zhu, J.: One transformer fits all distributions in multi-modal diffusion at scale. arXiv preprint arXiv:2303.06555 (2023) +6. Batzolis, G., Stanczuk, J., Schonlieb, C.B., Etmann, C.: Conditional image generation with score-based diffusion models. arXiv preprint arXiv:2111.13606 (2021) +7. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023) +8. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. NeurIPS (2020) +9. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ chatgpt quality (March 2023), https://lmsys.org/blog/2023-03-30-vicuna/ +10. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022) +11. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022) +12. Dai, X., Hou, J., Ma, C.Y., Tsai, S., Wang, J., Wang, R., Zhang, P., Vandenhende, S., Wang, X., Dubey, A., et al.: Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807 (2023) +13. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS (2021) +14. Dong, R., Han, C., Peng, Y., Qi, Z., Ge, Z., Yang, J., Zhao, L., Sun, J., Zhou, H., Wei, H., et al.: Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499 (2023) +15. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Thompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023) + +16. Fan, W.C., Chen, Y.C., Chen, D., Cheng, Y., Yuan, L., Wang, Y.C.F.: Frido: Feature pyramid diffusion for complex scene image synthesis. In: AAAI (2023) +17. Feng, Z., Zhang, Z., Yu, X., Fang, Y., Li, L., Chen, X., Lu, Y., Liu, J., Yin, W., Feng, S., et al.: Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. In: CVPR (2023) +18. Ge, Y., Ge, Y., Zeng, Z., Wang, X., Shan, Y.: Planting a seed of vision in large language model. arXiv preprint arXiv:2307.08041 (2023) +19. Gu, S., Chen, D., Bao, J., Wen, F., Zhang, B., Chen, D., Yuan, L., Guo, B.: Vector quantized diffusion model for text-to-image synthesis. In: CVPR (2022) +20. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022) +21. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020) +22. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. JMLR (2022) +23. Huang, S., Dong, L., Wang, W., Hao, Y., Singhal, S., Ma, S., Lv, T., Cui, L., Mohammed, O.K., Liu, Q., et al.: Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045 (2023) +24. Jiang, A.Q., Sablayrolles, A., Roux, A., Mensch, A., Savary, B., Bamford, C., Chaplot, D.S., Casas, D.d.l., Hanna, E.B., Bressand, F., et al.: Mixtral of experts. arXiv preprint arXiv:2401.04088 (2024) +25. Ju, X., Zeng, A., Bian, Y., Liu, S., Xu, Q.: Direct inversion: Boosting diffusion-based editing with 3 lines of code. ICLR (2024) +26. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: CVPR (2023) +27. Kim, G., Kwon, T., Ye, J.C.: Diffusionclip: Text-guided diffusion models for robust image manipulation. In: CVPR (2022) +28. Kingma, D., Salimans, T., Poole, B., Ho, J.: Variational diffusion models. NeurIPS (2021) +29. Koh, J.Y., Fried, D., Salakhutdinov, R.: Generating images with multimodal language models. arXiv preprint arXiv:2305.17216 (2023) +30. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) +31. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014) +32. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023) +33. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. NeurIPS (2023) +34. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: CVPR (2022) +35. Meng, C., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073 (2021) +36. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021) +37. OpenAI: Gpt-4 technical report (2023) +38. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: ICCV (2023) + +39. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. NeurIPS (2023) +40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) +41. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021) +42. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) +43. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023) +44. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH (2022) +45. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022) +46. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022) +47. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS (2022) +48. Shen, Y., Song, K., Tan, X., Li, D., Lu, W., Zhuang, Y.: Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580 (2023) +49. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015) +50. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) +51. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020) +52. Sun, Q., Yu, Q., Cui, Y., Zhang, F., Zhang, X., Wang, Y., Gao, H., Liu, J., Huang, T., Wang, X.: Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222 (2023) +53. Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca (2023) +54. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) +55. Valevski, D., Kalman, M., Matias, Y., Leviathan, Y.: Unitune: Text-driven image editing by fine tuning an image generation model on a single image. arXiv preprint arXiv:2210.09477 (2022) +56. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: ICCVW (2021) +57. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023) + +58. Wu, S., Fei, H., Qu, L., Ji, W., Chua, T.S.: Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519 (2023) +59. Xia, B., Zhang, Y., Wang, S., Wang, Y., Wu, X., Tian, Y., Yang, W., Timotfe, R., Van Gool, L.: Diffie2i: Efficient diffusion model for image-to-image translation. arXiv preprint arXiv:2308.13767 (2023) +60. Xia, B., Zhang, Y., Wang, S., Wang, Y., Wu, X., Tian, Y., Yang, W., Van Gool, L.: Diffir: Efficient diffusion model for image restoration. ICCV (2023) +61. Xia, B., Zhang, Y., Wang, Y., Tian, Y., Yang, W., Timofte, R., Van Gool, L.: Knowledge distillation based degradation estimation for blind super-resolution. ICLR (2023) +62. Xue, Z., Song, G., Guo, Q., Liu, B., Zong, Z., Liu, Y., Luo, P.: Raphael: Text-to-image generation via large mixture of diffusion paths. arXiv preprint arXiv:2305.18295 (2023) +63. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023) +64. Yu, L., Shi, B., Pasunuru, R., Muller, B., Golovneva, O., Wang, T., Babu, A., Tang, B., Karrer, B., Sheynin, S., et al.: Scaling autoregressive multi-modal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591 (2023) +65. Zhang, K., Mo, L., Chen, W., Sun, H., Su, Y.: Magicbrush: A manually annotated dataset for instruction-guided image editing. NeurIPS (2024) +66. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023) +67. Zhong, S., Huang, Z., Wen, W., Qin, J., Lin, L.: Sur-adapter: Enhancing text-to-image pre-trained diffusion models with large language models. In: ACM MM (2023) +68. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. TPAMI (2017) +69. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) \ No newline at end of file diff --git a/2024/LLMGA_ Multimodal Large Language Model based Generation Assistant/images.zip b/2024/LLMGA_ Multimodal Large Language Model based Generation Assistant/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ba89cebc70d00937d3b5fafca21ca2496254d8fb --- /dev/null +++ b/2024/LLMGA_ Multimodal Large Language Model based Generation Assistant/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a54b89ce8d0771e5314d74f746a1e1fd6c6cfab5630266378f3f7b88cb3996f +size 811666 diff --git a/2024/LLMGA_ Multimodal Large Language Model based Generation Assistant/layout.json b/2024/LLMGA_ Multimodal Large Language Model based Generation Assistant/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..05aa18d96f570ba9cc8d012d0c454c40449bb789 --- /dev/null +++ b/2024/LLMGA_ Multimodal Large Language Model based Generation Assistant/layout.json @@ -0,0 +1,10245 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 146 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 146 + ], + "type": "text", + "content": "LLMGA: Multimodal Large Language Model based Generation Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 151, + 167, + 459, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 167, + 459, + 180 + ], + "spans": [ + { + "bbox": [ + 151, + 167, + 459, + 180 + ], + "type": "text", + "content": "Bin Xia" + }, + { + "bbox": [ + 151, + 167, + 459, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 151, + 167, + 459, + 180 + ], + "type": "text", + "content": ", Shiyin Wang" + }, + { + "bbox": [ + 151, + 167, + 459, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 151, + 167, + 459, + 180 + ], + "type": "text", + "content": ", Yingfan Tao" + }, + { + "bbox": [ + 151, + 167, + 459, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 151, + 167, + 459, + 180 + ], + "type": "text", + "content": ", Yitong Wang" + }, + { + "bbox": [ + 151, + 167, + 459, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 151, + 167, + 459, + 180 + ], + "type": "text", + "content": ", and Jiaya Jia" + }, + { + "bbox": [ + 151, + 167, + 459, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 223, + 189, + 389, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 189, + 389, + 201 + ], + "spans": [ + { + "bbox": [ + 223, + 189, + 389, + 201 + ], + "type": "text", + "content": "1 The Chinese University of Hong Kong" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 272, + 201, + 342, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 201, + 342, + 212 + ], + "spans": [ + { + "bbox": [ + 272, + 201, + 342, + 212 + ], + "type": "text", + "content": "2 ByteDance Inc" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 257, + 213, + 356, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 213, + 356, + 223 + ], + "spans": [ + { + "bbox": [ + 257, + 213, + 356, + 223 + ], + "type": "text", + "content": "https://llmga.github.io/" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 250, + 452, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 250, + 452, + 491 + ], + "spans": [ + { + "bbox": [ + 160, + 250, + 452, + 491 + ], + "type": "text", + "content": "Abstract. In this paper, we introduce a Multimodal Large Language Model-based Generation Assistant (LLMGA), leveraging the vast reservoir of knowledge and proficiency in reasoning, comprehension, and response inherent in Large Language Models (LLMs) to assist users in image generation and editing. Diverging from existing approaches where Multimodal Large Language Models (MLLMs) generate fixed-size embeddings to control Stable Diffusion (SD), our LLMGA provides a detailed language generation prompt for precise control over SD. This not only augments LLM context understanding but also reduces noise in generation prompts, yields images with more intricate and precise content, and elevates the interpretability of the network. To this end, we curate a comprehensive dataset comprising prompt refinement, similar image generation, inpainting & outpainting, and instruction-based editing. Moreover, we propose a two-stage training scheme. In the first stage, we train the MLLM to grasp the properties of image generation and editing, enabling it to generate detailed prompts. In the second stage, we optimize SD to align with the MLLM's generation prompts. Additionally, we propose a reference-based restoration network to alleviate texture, brightness, and contrast disparities between generated and preserved regions during inpainting and outpainting. Extensive results show that LLMGA has promising generation and editing capabilities and can enable more flexible and expansive applications in an interactive manner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 502, + 452, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 502, + 452, + 514 + ], + "spans": [ + { + "bbox": [ + 160, + 502, + 452, + 514 + ], + "type": "text", + "content": "Keywords: Interactive generation and editing " + }, + { + "bbox": [ + 160, + 502, + 452, + 514 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 502, + 452, + 514 + ], + "type": "text", + "content": " LLM " + }, + { + "bbox": [ + 160, + 502, + 452, + 514 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 502, + 452, + 514 + ], + "type": "text", + "content": " Diffusion Model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 629 + ], + "type": "text", + "content": "Artificial Intelligence Generated Content (AIGC) has witnessed remarkable advancements, particularly propelled by the evolution of large language models (LLMs) [8, 10, 54] for text generation and diffusion models (DMs) [21, 42] for image generation. LLMs, in particular, have received considerable acclaim for their exceptional ability to comprehend, reason, make decisions, possess extensive knowledge, and generate text with unparalleled accuracy and fluency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "content": "Recent studies have begun delving deeper into Multimodal Large Language Models (MLLMs) [1,69] built upon LLMs, aiming to empower LLMs to comprehend inputs extending beyond text. For example, BLIP-2 [30] and LLaVA [33]" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 218, + 191 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 218, + 191 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 218, + 191 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 218, + 191 + ], + "type": "image", + "image_path": "9e911b04cc38e7b9e961d81ffec92e55e0cf1f64fdcd19496e0c4d3167263829.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 218, + 114, + 301, + 191 + ], + "blocks": [ + { + "bbox": [ + 218, + 114, + 301, + 191 + ], + "lines": [ + { + "bbox": [ + 218, + 114, + 301, + 191 + ], + "spans": [ + { + "bbox": [ + 218, + 114, + 301, + 191 + ], + "type": "image", + "image_path": "bad945006d7a95d2b4fb1e61f9fc57c1b0f858a8f474ec7e9d2dd97684d30b66.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 301, + 114, + 406, + 191 + ], + "blocks": [ + { + "bbox": [ + 301, + 114, + 406, + 191 + ], + "lines": [ + { + "bbox": [ + 301, + 114, + 406, + 191 + ], + "spans": [ + { + "bbox": [ + 301, + 114, + 406, + 191 + ], + "type": "image", + "image_path": "c6aa2e25d53add88442e917afb5fe46ff0d46775bfe210dc6fd635fe442f1036.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 408, + 114, + 494, + 196 + ], + "blocks": [ + { + "bbox": [ + 408, + 114, + 494, + 196 + ], + "lines": [ + { + "bbox": [ + 408, + 114, + 494, + 196 + ], + "spans": [ + { + "bbox": [ + 408, + 114, + 494, + 196 + ], + "type": "image", + "image_path": "edc64691018c7e8e081f9a9db223816f40e255a2683ff23a295fc36a8dccb3cc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 134, + 193, + 218, + 276 + ], + "blocks": [ + { + "bbox": [ + 134, + 193, + 218, + 276 + ], + "lines": [ + { + "bbox": [ + 134, + 193, + 218, + 276 + ], + "spans": [ + { + "bbox": [ + 134, + 193, + 218, + 276 + ], + "type": "image", + "image_path": "51323782904c03d52c3a1a5ba51ec6489ec41152a06c08e989e8c100c6efe123.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 286, + 482, + 385 + ], + "lines": [ + { + "bbox": [ + 130, + 286, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 482, + 385 + ], + "type": "text", + "content": "Fig. 1: Some examples of LLMGA for assisting in image generation and editing. (1) T2I generation. LLMGA can refine the user's generation prompt to produce more vivid and vibrant images. (2) Similar image generation. LLMGA can understand the component and layout of the input images and generate a similar image. (3) Inpainting & Outpainting. LLMGA can provide detailed generation prompts based on user preferences and input images. (4) Instruction based editing. LLMGA can understand user instructions and realize accurate editing. (5) Interactive image generation and editing exemplify the comprehensive capabilities of LLMGA. Users can design satisfactory images by engaging in interactions with LLMGA, leveraging its vast knowledge." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 218, + 193, + 301, + 276 + ], + "blocks": [ + { + "bbox": [ + 218, + 193, + 301, + 276 + ], + "lines": [ + { + "bbox": [ + 218, + 193, + 301, + 276 + ], + "spans": [ + { + "bbox": [ + 218, + 193, + 301, + 276 + ], + "type": "image", + "image_path": "e9325f1b8254c5a5da9fa606c8b65118ca571514c3bc01c3b4b329e299a7ace7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 301, + 193, + 404, + 276 + ], + "blocks": [ + { + "bbox": [ + 301, + 193, + 404, + 276 + ], + "lines": [ + { + "bbox": [ + 301, + 193, + 404, + 276 + ], + "spans": [ + { + "bbox": [ + 301, + 193, + 404, + 276 + ], + "type": "image", + "image_path": "7196089690010dc3ce9acb2ce6d66a849548ef96d897bc1d86b8455f88583739.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 404, + 193, + 494, + 276 + ], + "blocks": [ + { + "bbox": [ + 404, + 193, + 494, + 276 + ], + "lines": [ + { + "bbox": [ + 404, + 193, + 494, + 276 + ], + "spans": [ + { + "bbox": [ + 404, + 193, + 494, + 276 + ], + "type": "image", + "image_path": "ee491b89156c24e2a42b8f0e84bc42a93043ee3b1525a3d9c8020413879d1880.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 412, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 484 + ], + "type": "text", + "content": "employ visual encoders to transform images into input embeddings, enabling them to be used as prompts alongside text input for the LLM, thus achieving LLMs with the visual modality. Furthermore, recent works focused on extending the capabilities of LLMs to generate multimodal outputs. For example, GILL [29] involves instructing LLMs to predict fixed-size visual embeddings aligned with CLIP [40] space to control the Stable Diffusion [42] (SD) for image generation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "text", + "content": "However, existing works [29, 52] merely focus on enabling LLM to output images but do not aim to assist users in generating or editing images to enhance quality. In this paper, we aim to develop a Multimodal Large Language Model-based Generation Assistant (LLMGA) to better assist image generation models, making them more user-friendly and capable of producing high-quality images. In contrast to certain methods [29, 52] that leverage MLLMs to predict fixed-size visual embeddings for implicit SD control, our approach is straightforward. We guide the generation of SD using detailed language prompts from MLLM based on five reasons. (1) The embeddings predicted by the MLLM are often filled with noise. This can be filtered out by mapping them to a fixed language domain, enabling precise control of SD. (2) Detailed language prompts can make the network more transparent and interactive, allowing users to understand MLLM's thoughts for generating images. (3) MLLM is pre-trained on vast textual datasets. Explicit language prompts rather than implicit embeddings are more advantageous for MLLM to generate prompts and comprehend context. (4)" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 168, + 91, + 434, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 91, + 434, + 102 + ], + "spans": [ + { + "bbox": [ + 168, + 91, + 434, + 102 + ], + "type": "text", + "content": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "Dynamic-sized language prompt facilitates the addition of generation requests during interactions. (5) Training is more simple and efficient." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 481, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 481, + 239 + ], + "type": "text", + "content": "However, we face several challenges: (1) MLLM may reject the execution of generation instructions due to its nature as a language assistant. (2) MLLM lacks a comprehensive understanding of image generation and editing, and cannot provide an accurate and detailed generation prompt. (3) Determining which part of texts generated by MLLM to guide SD generation. (4) SD's CLIP encode only 75 tokens. Additionally, SD is trained on short captions, whereas our LLMGA typically generates detailed prompts exceeding 150 tokens. This discrepancy poses a challenge for SD in following the detailed prompt of LLMGA." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 239, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 419 + ], + "type": "text", + "content": "To this end, we have devised a two-stage training scheme. First, we construct a training dataset: prompt refinement, similar image generation, inpainting & outpainting, and visual question answering. We then train LLMGA on these four datasets to cultivate four fundamental capabilities: (1) For concise user prompts, LLMGA can refine the generation of intricate details, encompassing attire, background, and characters. (2) LLMGA can precisely regenerate an image it observes. (3) LLMGA can generate or refine prompts for inpainting & outpainting based on its understanding of the image. (4) LLMGA can generate accurate prompts for instruction-based editing according to users' requirements and given images. Additionally, we make LLMGA use special symbols and to distinguish generation prompts and responses. In the second stage, we freeze the parameters of LLMGA's MLLM and initiate joint training with the SD. This process enables the SD to acclimate to the detailed prompt produced by the MLLM. Notably, when the input token count exceeds 75, we iteratively apply the CLIP [40] encoder to the surplus tokens." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 421, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 482, + 517 + ], + "type": "text", + "content": "Moreover, we have identified noticeable disparities in texture, contrast, and brightness between the newly generated and preserved sections in inpainting & outpainting. Therefore, we propose a Diffusion-based Reference Restoration Network (DiffRIR). Specifically, aside from images generated by SD, we add masked images as reference inputs into DiffRIR. This enables the DiffRIR to refer to the texture, contrast, and brightness of the preserved regions for restoration. Additionally, we introduce perturbations to contrast and brightness during training, enabling DiffRIR to correct contrast and brightness disparities in the images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 518, + 482, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 482, + 603 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 482, + 603 + ], + "type": "text", + "content": "As shown in Fig. 1, LLMGA is a unified and interactive framework for image generation and editing, endowed with a wide array of capabilities: (1) LLMGA leverages its extensive world knowledge and powerful reasoning abilities to assist image generation and editing and significantly improve results. (2) LLMGA can be integrated with external plugins, like ControlNet. (3) Most importantly, users can interact with LLMGA to design satisfying images in a more convenient, flexible, and enjoyable way. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 90, + 447, + 102 + ], + "type": "text", + "content": "LLMGA: Multimodal Large Language Model based Generation Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 617, + 482, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 617, + 482, + 666 + ], + "type": "text", + "content": "- We proposed LLMGA, a simple yet powerful interactive generation and editing framework. Experiments affirm the efficacy of LLMGA in enhancing generation and editing thanks to its vast knowledge and interactive features. Plus, LLMGA can integrate with external plugins for wider applications." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 480, + 233 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 479, + 163 + ], + "type": "text", + "content": "- We construct a training dataset, including four parts: prompt refinement, similar image generation, inpainting & outpainting, and instruction-based editing. This enhances LLMGA's comprehension of generation and editing tasks while standardizing response formats." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 163, + 480, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 163, + 480, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 163, + 480, + 209 + ], + "type": "text", + "content": "- We proposed a restoration network DiffRIR, which introduces reference images and training perturbations to contrast and brightness. DiffRIR can alleviate texture, contrast, and brightness discrepancies between newly generated and preserved regions for edited images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 210, + 480, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 210, + 480, + 233 + ], + "spans": [ + { + "bbox": [ + 138, + 210, + 480, + 233 + ], + "type": "text", + "content": "- Open-source. The following assets are released: the generated data, the codebase for model training, the model checkpoint, and a demo." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 248, + 237, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 237, + 261 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 237, + 261 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 270, + 480, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 480, + 449 + ], + "type": "text", + "content": "Diffusion Model. Diffusion Models (DMs) [2,4,7,13,19,21,22,28,35,38,49,51] have achieved remarkable results in image generation. DMs adopt a parameterized Markov chain to optimize the lower variational bound on the likelihood function. In this way, it can generate realistic images from Gaussian noise. After that, several DM methods [3,5,6,12,16,17,26,27,34,39,44,45,55,59,60,62] have been tailored to enhance the text-to-image (T2I) generation and editing. Notably, GLIDE [36] pioneered the incorporation of text features into transformer blocks during the denoising process. Subsequently, DALL-E [41], Imagen [46], and Stable Diffusion [42] have made substantial strides in improving T2I generation. Subsequently, some works [43,66,67] introduced conditioning controls to the DMs to facilitate a more convenient and precise manipulation of the generation process. Overall, enhancing the user-friendliness of DMs is a key focus within the community. In this paper, we introduce LLMGA, leveraging the extensive knowledge and powerful reasoning capabilities of LLM to facilitate users in achieving more easily attainable and satisfactory image designs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 450, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 480, + 533 + ], + "type": "text", + "content": "Multimodal Large Language Models. Recently, LLMs have undeniably made profound impacts and revolutions within the entire AI community and beyond. For example, exemplary LLMs, such as ChatGPT and GPT4 [37], have showcased remarkable abilities in comprehension, reasoning, responses, and knowledge reservoirs. Subsequently, a range of LLMs [11], including Vicuna [9], LLaMA [54], and Alpaca [53] have been released as open-source models, substantially propelling advancements of the community." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 534, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 480, + 616 + ], + "type": "text", + "content": "Afterward, the community began focusing on the development of the Multimodal Large Language Model [14, 15, 18, 23, 63, 64, 69]. They aim to enable LLMs to comprehend both images and text and provide textual responses. For instance, Flamingo [1] encodes images and feeds them into the LLM's attention layer. BLIP-2 [30] employs Q-Former to encode input images into queries. Additionally, LLaVA [33] leverages CLIP [40] to encode images into visual embeddings, which are then concatenated with text embeddings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "Recent concurrent works, such as Next-GPT [58], have extended the capabilities to encompass audio and video modalities. Moreover, Visual-ChatGPT [57] and HuggingGPT [48] make LLMs act as agents capable of invoking various pretrained visual models to achieve MLLM. However, these works focus on making" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 168, + 90, + 433, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 90, + 433, + 102 + ], + "spans": [ + { + "bbox": [ + 168, + 90, + 433, + 102 + ], + "type": "text", + "content": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 120, + 340, + 268 + ], + "blocks": [ + { + "bbox": [ + 138, + 120, + 340, + 268 + ], + "lines": [ + { + "bbox": [ + 138, + 120, + 340, + 268 + ], + "spans": [ + { + "bbox": [ + 138, + 120, + 340, + 268 + ], + "type": "image", + "image_path": "ac16f1c191015e38321bbbc44c46d40be6e0f8fa0223e8990c44fa1ded187dd5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 280, + 482, + 357 + ], + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 357 + ], + "type": "text", + "content": "Fig. 2: The overview of LLMGA. (a) In the first training stage, we train the MLLM to produce generation prompts based on provided instructions. Moreover, we construct a training dataset including four categories: prompt refinement, similar image generation, inpainting & outpainting, and instruction-based editing. (b) In the second training stage, we optimize SD to adapt to the detailed generation prompts from MLLM. (c) In the inference stage, LLMGA can respond to user queries and assist in various tasks, such as image generation, inpainting & outpainting, and editing." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 342, + 120, + 477, + 269 + ], + "blocks": [ + { + "bbox": [ + 342, + 120, + 477, + 269 + ], + "lines": [ + { + "bbox": [ + 342, + 120, + 477, + 269 + ], + "spans": [ + { + "bbox": [ + 342, + 120, + 477, + 269 + ], + "type": "image", + "image_path": "89b4e111b668e96c8e2ed4ec0e9287c59b03a9f2210e5350cc4d65c448bde191.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 380, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 452 + ], + "type": "text", + "content": "LLM determine the combined invocation of modules (such as detection, recognition, and generation) to fulfill user requirements. However, these methods are not tailored for generation and editing and use concise prompts that lack the capability to enhance results. Thus, we propose LLMGA, which is designed to assist with various image generation and editing tasks. It can achieve satisfactory results by strong reasoning capability and flexible interaction with users." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 468, + 233, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 233, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 233, + 482 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 491, + 266, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 266, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 266, + 502 + ], + "type": "text", + "content": "3.1 Overview of LLMGA" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": "In this paper, we aim to design a MLLM-based Generation Assistant (LLMGA). Our LLMGA produces detailed language-based generation prompts to control SD rather than predicting fixed-sized visual embeddings [29, 52] to govern SD. This has five advantages: (1) Visual embeddings contain noise, and mapping them to the language domain can filter out this noise, enabling precise SD control. (2) Language-based generation prompts facilitate users in understanding the LLMGA's thoughts, enhancing interaction. (3) Dynamic-sized language-based generation prompt enables the addition of generation requests. (4) MLLM is pre-trained on textual datasets. Language prompts rather than implicit visual embeddings are more advantageous for MLLM to generate accurate prompts and comprehend context. (5) Training is simpler and more efficient." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "However, we need to address several issues: (1) As a language assistant, MLLM may decline the execution of generation instructions. (2) MLLM lacks" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "text", + "content": "LLMGA: Multimodal Large Language Model based Generation Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "a nuanced understanding of image generation and editing, and cannot produce precise and detailed generation prompts. (3) MLLM needs to decide which part of the output text serves as generation prompts to guide generation. (4) SD's CLIP encode only 75 tokens. Moreover, SD is primarily trained on short captions, while detailed prompts generated by LLMGA may exceed 150 tokens. This disparity makes it hard for SD to understand the instructions from LLMGA. To address the aforementioned challenges, we construct a training dataset and design two-stage training schemes to train the MLLM (Sec. 3.2) and SD (Sec. 3.3)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 272 + ], + "type": "text", + "content": "The network structure and pipeline of LLMGA are illustrated in Fig. 2. Specifically, as shown in Fig. 2 (a) and (c), the images " + }, + { + "bbox": [ + 130, + 212, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{input}" + }, + { + "bbox": [ + 130, + 212, + 482, + 272 + ], + "type": "text", + "content": " are encoded into image embeddings by CLIP vision encoder and a projection layer. Subsequently, the image embedding is concatenated with the text embedding and fed into the LLM to obtain text output " + }, + { + "bbox": [ + 130, + 212, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{output}" + }, + { + "bbox": [ + 130, + 212, + 482, + 272 + ], + "type": "text", + "content": ". This process can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 233, + 284, + 480, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 284, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 233, + 284, + 480, + 297 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {\\text {o u t p u t}} = \\mathcal {F} _ {M L L M} \\left(\\mathbf {T} _ {\\text {i n p u t}}, \\mathbf {I} _ {\\text {i n p u t}}\\right), \\tag {1}", + "image_path": "94ace391f344bc0bb9a1fc127605856c1bf1ca4401efa28645be5a81fe458fe4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 481, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 481, + 325 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 300, + 481, + 325 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{input}" + }, + { + "bbox": [ + 130, + 300, + 481, + 325 + ], + "type": "text", + "content": " indicates the input text instructions from users. It is notable that " + }, + { + "bbox": [ + 130, + 300, + 481, + 325 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{MLLM}" + }, + { + "bbox": [ + 130, + 300, + 481, + 325 + ], + "type": "text", + "content": " can process only " + }, + { + "bbox": [ + 130, + 300, + 481, + 325 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{input}" + }, + { + "bbox": [ + 130, + 300, + 481, + 325 + ], + "type": "text", + "content": " as input." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "type": "text", + "content": "The text output " + }, + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{output}" + }, + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "type": "text", + "content": " can comprise two components: text response " + }, + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_R" + }, + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "type": "text", + "content": " and generation prompt " + }, + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_P" + }, + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "type": "text", + "content": ". To distinguish between " + }, + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_R" + }, + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_P" + }, + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "type": "text", + "content": ", we adopt new special tokens, i.e., " + }, + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "type": "inline_equation", + "content": "<\\text{gen\\_img}>" + }, + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "type": "text", + "content": ", to encompass " + }, + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_P" + }, + { + "bbox": [ + 130, + 325, + 481, + 361 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 361, + 481, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 481, + 397 + ], + "type": "text", + "content": "We present " + }, + { + "bbox": [ + 130, + 361, + 481, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_R" + }, + { + "bbox": [ + 130, + 361, + 481, + 397 + ], + "type": "text", + "content": " as the immediate text response to users. Concurrently, " + }, + { + "bbox": [ + 130, + 361, + 481, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_P" + }, + { + "bbox": [ + 130, + 361, + 481, + 397 + ], + "type": "text", + "content": " is further fed into the subsequent SD to guide T2I generation (Eq. 2), inpainting & outpainting (Eq. 3), and instruction-based image editing (Eq. 4)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 266, + 406, + 480, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 406, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 266, + 406, + 480, + 418 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {g} = \\mathcal {F} _ {T 2 I} (\\mathbf {T} _ {P}, \\mathbf {Z}), \\tag {2}", + "image_path": "2d75f4003670bb0dc38b1e88e5dbfd4a360e0e6a854991bd3cc8cfc2e2951413.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 234, + 439, + 480, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 439, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 234, + 439, + 480, + 452 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {p} = \\mathcal {F} _ {\\text {P a i n t}} (\\mathbf {I} _ {\\text {i n p u t}}, \\mathbf {I} _ {\\text {m a s k}}, \\mathbf {T} _ {P}, \\mathbf {Z}), \\tag {3}", + "image_path": "fc4934d666a9b86e117333928174d6226009361159658fe45226e0e28418fe6f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 256, + 468, + 480, + 481 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 468, + 480, + 481 + ], + "spans": [ + { + "bbox": [ + 256, + 468, + 480, + 481 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {e} = \\mathcal {F} _ {\\text {E d i t}} \\left(\\mathbf {I} _ {\\text {i n p u t}}, \\mathbf {T} _ {P}\\right), \\tag {4}", + "image_path": "8ec4bb0ee9498a61ac237a05bcf07d19145197979e1c946590cb18f2b2de87dd.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 485, + 482, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 558 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 130, + 485, + 482, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 130, + 485, + 482, + 558 + ], + "type": "text", + "content": " denotes the random Gaussian noise. Furthermore, to ensure the encoding of all " + }, + { + "bbox": [ + 130, + 485, + 482, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_P" + }, + { + "bbox": [ + 130, + 485, + 482, + 558 + ], + "type": "text", + "content": " for SD, we iteratively run the CLIP text encoder until all prompts are encoded. For inpainting & outpainting, except the input image " + }, + { + "bbox": [ + 130, + 485, + 482, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{input}" + }, + { + "bbox": [ + 130, + 485, + 482, + 558 + ], + "type": "text", + "content": ", an additional mask " + }, + { + "bbox": [ + 130, + 485, + 482, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{mask}" + }, + { + "bbox": [ + 130, + 485, + 482, + 558 + ], + "type": "text", + "content": " are essential inputs for the inpainting SD to specify the region requiring generation. For instruction-based image editing, " + }, + { + "bbox": [ + 130, + 485, + 482, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{Edit}" + }, + { + "bbox": [ + 130, + 485, + 482, + 558 + ], + "type": "text", + "content": " performs inversion [25, 50] and prompt-to-prompt [20] based on T2I SD." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 574, + 244, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 244, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 244, + 586 + ], + "type": "text", + "content": "3.2 MLLM Training" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "As described in Sec. 3.1, original MLLMs are specifically designed and trained as language assistants, but they lack the proficiency to assist in image generation, and editing. Notably, considering the input of images and the performance of open-sourced MLLMs, employing few-shot learning to guide the model to achieve the desired results proves challenging and inefficient. Therefore, it is crucial to train the MLLM to serve as proficient assistants in image generation and editing" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 168, + 90, + 434, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 90, + 434, + 102 + ], + "spans": [ + { + "bbox": [ + 168, + 90, + 434, + 102 + ], + "type": "text", + "content": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "tasks, understanding the expected response formats and enhancing their comprehension of image generation and editing properties. To this end, as depicted in Fig. 2 (a), we construct a training dataset consisting of four categories:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 481, + 415 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 152, + 481, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 481, + 236 + ], + "type": "text", + "content": "(1) Prompt Refinement. We establish this dataset to cultivate the prompt refinement ability of the MLLM. Specifically, we utilize GPT4-V to furnish detailed descriptions of images in MSCOCO [31]. These detailed descriptions, along with the original MSCOCO brief descriptions, constitute a training text pair. During training, we input the brief MSCOCO captions and randomly select and append a generation instruction. When a generation instruction is included in the prompt, we add and on the later description." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 481, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 481, + 295 + ], + "type": "text", + "content": "(2) Similar Image Generation. We select images from the MSCOCO dataset along with corresponding detailed descriptions generated by GPT4-V to create the Similar Image Generation dataset. During training, we input the images along with a generation instruction. In cases where a generation instruction is provided, we add and on the subsequent description." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 295, + 481, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 481, + 343 + ], + "type": "text", + "content": "(3) Inpainting & Outpainting. We use pairs of detailed descriptions and images from the Similar Image Generation dataset. During training, we input masked images with inpainting or outpainting instructions. Besides, we include and on the subsequent description." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 342, + 481, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 481, + 415 + ], + "type": "text", + "content": "(4) Instruction-based editing. We fine-tune Mixtral-8x7B [24] to enable it to generate editing data based on detailed descriptions from MSCOCO. Subsequently, we clean the generated data. During training, we can input images or original captions and provide corresponding editing instructions, aiming to train LLMGA to output the target caption. Additionally, we include and tags on the target caption during training." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 415, + 481, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 481, + 487 + ], + "type": "text", + "content": "During training, alongside the above data, we integrated the image designing and Alpaca [53] dataset, to enhance LLMGA's question-answering (QA). Notably, we excluded certain visual multimodal incompatible question-answer pairs from it. Furthermore, we incorporated the LLaVA v1.5 mix665k dataset [32] to endow LLMGA with Visual Question Answering (VQA) capabilities. We provide more details in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 487, + 481, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 481, + 535 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 481, + 535 + ], + "type": "text", + "content": "As illustrated in Fig. 2 (a), we freeze the CLIP vision encoder and optimize the projection layer and LLM. The model is trained end-to-end using the autoregressive cross-entropy loss " + }, + { + "bbox": [ + 130, + 487, + 481, + 535 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{MLLM})" + }, + { + "bbox": [ + 130, + 487, + 481, + 535 + ], + "type": "text", + "content": " for text generation. Given the ground-truth targets " + }, + { + "bbox": [ + 130, + 487, + 481, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{GT}" + }, + { + "bbox": [ + 130, + 487, + 481, + 535 + ], + "type": "text", + "content": ", this loss can be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 237, + 544, + 481, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 544, + 481, + 558 + ], + "spans": [ + { + "bbox": [ + 237, + 544, + 481, + 558 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M L L M} = \\mathbf {C E} \\left(\\mathbf {T} _ {\\text {o u t p u t}}, \\mathbf {T} _ {G T}\\right). \\tag {5}", + "image_path": "99ab9ce34b3bae3502ee71e82822c01f4dcccb93ced9a09b073be8bb290c5ca7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 574, + 288, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 288, + 586 + ], + "type": "text", + "content": "3.3 Stable Diffusion Training" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "content": "As described in Sec. 3.1, the original SD's CLIP text encoder only encodes 75 tokens, which cannot handle the entire MLLM's generation prompt. Moreover, the original SD is trained on brief captions, which cannot fully understand the generation prompts. Thus, we repeatedly use the CLIP text encoder to encode all instances of " + }, + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_P" + }, + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "content": " for SD. Besides, we train the T2I SD model and the inpainting SD model, respectively. For both generation and inpainting & outpainting tasks," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 448, + 102 + ], + "type": "text", + "content": "LLMGA: Multimodal Large Language Model based Generation Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "the generation prompts of MLLM are detailed descriptions of images. Therefore, during training, we can instruct MLLM to provide a detailed description " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_P" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " for images from the LAION-Aesthetics [47] and MSCOCO datasets. Subsequently, " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_P" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " is fed into T2I SD or inpainting SD for joint training. Notably, we only optimize the SD unet while freezing the parameters of other networks. To accelerate the training, we record the prompt " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_P" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " of MLLM to avoid redundant calculations. The model is trained using SD loss (Eq. 6). For instruction-based editing, we directly adopt the pre-trained T2I SD model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 228, + 221, + 481, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 221, + 481, + 240 + ], + "spans": [ + { + "bbox": [ + 228, + 221, + 481, + 240 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S D} = \\mathbb {E} _ {\\mathbf {Z} _ {t}, \\mathbf {C}, \\epsilon , t} \\left(\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\mathbf {Z} _ {t}, \\mathbf {C}\\right) \\| _ {2} ^ {2}\\right), \\tag {6}", + "image_path": "e159a30fec5fb870ac568d2cd2e875090e10341e1f1ccba3b9bb6be0373a17f4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_t = \\sqrt{\\overline{\\alpha}_t}\\mathbf{Z}_0 + \\sqrt{1 - \\overline{\\alpha}_t}\\epsilon" + }, + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "text", + "content": " represents the noised feature map at timestep " + }, + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "text", + "content": ". Ground truth images are encoded into latent space to derive " + }, + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_0" + }, + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\epsilon \\in \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "text", + "content": " represents Gaussian noise, and " + }, + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "text", + "content": " refers to the SD unet. " + }, + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "text", + "content": " indicates the conditional information. For T2I generation, " + }, + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_P" + }, + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "text", + "content": ". For inpainting and outpainting, " + }, + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_P" + }, + { + "bbox": [ + 130, + 249, + 482, + 310 + ], + "type": "text", + "content": ", the mask, and the VAE-encoded masked image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 327, + 313, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 327, + 313, + 339 + ], + "spans": [ + { + "bbox": [ + 131, + 327, + 313, + 339 + ], + "type": "text", + "content": "3.4 Restoration Network Training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 346, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 453 + ], + "type": "text", + "content": "For SD inpainting & outpainting, we observed noticeable disparities between the preserved and newly generated regions in the edited images. To enhance the consistency between the newly generated and the preserved regions, we introduce a reference-based restoration scheme. Specifically, existing restoration methods [56, 60, 61] take low-quality (LQ) images as input and produce high-quality (HQ) images, but they often do not leverage the preserved information from the given masked image. Different from them, we concatenate the LQ image " + }, + { + "bbox": [ + 130, + 346, + 482, + 453 + ], + "type": "inline_equation", + "content": "I_{LQ}" + }, + { + "bbox": [ + 130, + 346, + 482, + 453 + ], + "type": "text", + "content": " and the masked image, i.e., " + }, + { + "bbox": [ + 130, + 346, + 482, + 453 + ], + "type": "inline_equation", + "content": "(1 - \\mathbf{I}_{mask})\\mathbf{I}_{GT}" + }, + { + "bbox": [ + 130, + 346, + 482, + 453 + ], + "type": "text", + "content": ", as inputs and feed them into the restoration model " + }, + { + "bbox": [ + 130, + 346, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_R" + }, + { + "bbox": [ + 130, + 346, + 482, + 453 + ], + "type": "text", + "content": ". This process can be formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 217, + 464, + 481, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 464, + 481, + 476 + ], + "spans": [ + { + "bbox": [ + 217, + 464, + 481, + 476 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {H Q} = \\mathcal {F} _ {R} (\\operatorname {c o n c a t} \\left(\\mathbf {I} _ {L Q}, (1 - \\mathbf {I} _ {\\text {m a s k}}) \\mathbf {I} _ {G T}\\right)). \\tag {7}", + "image_path": "aaebce74e707003c20127f638ae13f4d9c9c61e2abad2d9bc152fd42bd98a7f4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "text", + "content": "To further mitigate the brightness and contrast disparities, we introduced additional color degradation (i.e., random brightness and contrast disturbance) into the training process of the restoration model, which is formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 268, + 532, + 481, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 532, + 481, + 543 + ], + "spans": [ + { + "bbox": [ + 268, + 532, + 481, + 543 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {2} (\\mathbf {x}) = c _ {1} \\mathbf {x} + c _ {2}, \\tag {8}", + "image_path": "cdef7c65caf5a40fa8894efdcd57a91fca48dca7de662f60e667ebb31d99d6a3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 262, + 564, + 481, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 564, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 262, + 564, + 481, + 577 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {L Q} = \\mathcal {D} _ {2} \\left(\\mathcal {D} _ {1} \\left(\\mathbf {I} _ {G T}\\right)\\right), \\tag {9}", + "image_path": "ea23218867d388af1ec70e201b7b5142159c02f3e69a55aba6c028fe69b3e1a2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " indicates contrast and brightness disturbance, " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " denotes the input image. " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "c_{1}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " is the contrast gain, randomly varied within the range of [0.94, 1.06], while " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "c_{2}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " is the brightness bias, randomly varied within the range of " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "[-0.05, 0.05]" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " represents the real-world degradation process used in restoration model [56, 60, 61]. " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{GT}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " is ground truth image, and " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{LQ}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " is LQ image. Here, we adopt the network structure of the SOTA restoration model DiffIR [60] and apply our schemes to it to obtain our Diffusion-based Reference Restoration Network (DiffRIR)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 168, + 90, + 434, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 90, + 434, + 102 + ], + "spans": [ + { + "bbox": [ + 168, + 90, + 434, + 102 + ], + "type": "text", + "content": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 280, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 280, + 153 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 280, + 153 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 479, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 479, + 208 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 479, + 208 + ], + "type": "text", + "content": "For the first stage of training, we train the MLLM (including projection and LLM) on " + }, + { + "bbox": [ + 130, + 160, + 479, + 208 + ], + "type": "inline_equation", + "content": "8 \\times \\mathrm{A}100" + }, + { + "bbox": [ + 130, + 160, + 479, + 208 + ], + "type": "text", + "content": ". The batch size is set to 128. Besides, the training datasets include VQA (LLaVA v1.5 mix665k), QA, prompt refinement, similar image generation, inpainting & outpainting, and instruction-based editing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 208, + 480, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 208, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 208, + 480, + 243 + ], + "type": "text", + "content": "For the second stage of training, we adopt the Stable Diffusion 1.5 (SD1.5) as the initial image generation or inpainting & outpainting model. We train these models on " + }, + { + "bbox": [ + 130, + 208, + 480, + 243 + ], + "type": "inline_equation", + "content": "8 \\times \\mathrm{A}100" + }, + { + "bbox": [ + 130, + 208, + 480, + 243 + ], + "type": "text", + "content": ". The batch size is set to 32." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 243, + 480, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 480, + 268 + ], + "type": "text", + "content": "For the restoration network, we train DiffRIR on " + }, + { + "bbox": [ + 130, + 243, + 480, + 268 + ], + "type": "inline_equation", + "content": "8 \\times \\mathrm{V}100" + }, + { + "bbox": [ + 130, + 243, + 480, + 268 + ], + "type": "text", + "content": ". The batch size is set to 64. Please check more details in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 286, + 270, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 286, + 270, + 298 + ], + "spans": [ + { + "bbox": [ + 132, + 286, + 270, + 298 + ], + "type": "text", + "content": "4.2 Experimental Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 306, + 298, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 306, + 298, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 306, + 298, + 450 + ], + "type": "text", + "content": "Evaluation on T2I Generation. The results are shown in Tab. 1. Notably, SD1.5 is the original Stable Diffusion 1.5 while SD1.5-ft indicates the finetuned SD1.5 in LLMGA. We also compare our LLMGA with the recently proposed multimodal generative model GILL [29]. (1) Comparing SD1.5, our LLMGA-7B achieves notable 5.4847 FID and 5.05 IS improvements, underscoring the effectiveness of LLMGA. (2) Moreover," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 450, + 480, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 480, + 545 + ], + "type": "text", + "content": "our LLMGA-7B significantly outperforms GILL. (3) In the 3rd and 4th rows of Tab. 1, we use LLMGA-7B to refine the short MSCOCO caption to the detailed generation prompt and send it to the SD1.5 and SD1.5-ft, respectively. Our LLMGA-7B with SD1.5-ft achieves significant 5.0712 FID and 3.92 IS improvements, respectively. This demonstrates our second-stage training can help SD1.5-ft to better follow detailed prompts from MLLM. (4) Comparing the 4th and 6th rows of Tab. 1, LLMGA-13B exhibits better performance than LLMGA-7B due to its superior reasoning ability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 628 + ], + "type": "text", + "content": "The qualitative results are shown in Fig. 3. (1) LLMGA excels in refining prompts by incorporating details to generate visually rich and pleasing images. For example, in the first row, LLMGA crafts a battle attire for the husky, depicting engaging scenarios of battling monsters. This makes user usage more convenient, eliminating the need for them to think about generating image details themselves. (2) LLMGA can leverage its extensive knowledge base to generate images, even for concepts users may not be familiar with, like a space elevator." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": "Evaluation on Instruction-based Editing. For instruction-based editing, we utilize LLMGA to provide a detailed description of the edited image based on the input image and user editing instructions. We then employ Direct Inversion [25]" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 307, + 330, + 482, + 428 + ], + "blocks": [ + { + "bbox": [ + 305, + 307, + 482, + 330 + ], + "lines": [ + { + "bbox": [ + 305, + 307, + 482, + 330 + ], + "spans": [ + { + "bbox": [ + 305, + 307, + 482, + 330 + ], + "type": "text", + "content": "Table 1: Quantitative comparison on T2I generation on the MSCOCO [31] dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 330, + 482, + 428 + ], + "lines": [ + { + "bbox": [ + 307, + 330, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 307, + 330, + 482, + 428 + ], + "type": "table", + "html": "
MethodFID↓IS↑
SD1.524.008135.99
GILL [29]25.112334.20
LLMGA-7b(SD1.5)23.594637.12
LLMGA-7b(SD1.5-ft, Ours)18.523441.04
LLMGA-13b(SD1.5)23.582837.58
LLMGA-13b(SD1.5-ft, Ours)18.406341.16
", + "image_path": "a48c202f5cd6d9b0c40536c66491d58b03991e17bc0e0797298d320e1bad46c7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 90, + 447, + 102 + ], + "type": "text", + "content": "LLMGA: Multimodal Large Language Model based Generation Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 473, + 287 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 473, + 287 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 473, + 287 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 473, + 287 + ], + "type": "image", + "image_path": "6c5ca517e99e665420423b72be44130bd6d43d781b60bd96ed1459b91812afc6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 297, + 480, + 309 + ], + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 309 + ], + "type": "text", + "content": "Fig. 3: T2I visual comparison. LLMGA can produce accurate and high-quality results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 187, + 343, + 427, + 403 + ], + "blocks": [ + { + "bbox": [ + 136, + 321, + 476, + 333 + ], + "lines": [ + { + "bbox": [ + 136, + 321, + 476, + 333 + ], + "spans": [ + { + "bbox": [ + 136, + 321, + 476, + 333 + ], + "type": "text", + "content": "Table 2: Quantitative comparison for image editing on MagicBrush test set [65]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 343, + 427, + 403 + ], + "lines": [ + { + "bbox": [ + 187, + 343, + 427, + 403 + ], + "spans": [ + { + "bbox": [ + 187, + 343, + 427, + 403 + ], + "type": "table", + "html": "
MethodsL1↓L2↓CLIP-I↑DINO↑CLIP-T↑
InstructPix2Pix0.11970.04160.84420.72520.2909
MagicBrush0.06470.02240.92930.89130.2979
LLMGA (Ours)0.08140.02180.89360.87680.3137
", + "image_path": "aae02e8db1f0fb0aca33f00dbd56354f0d3d4493609cedd1659aa01f246d472b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": "and prompt-to-prompt [20] methods to obtain the edited image. (1) The quantitative results are presented in Tab. 2. Notably, our LLMGA did not train SD for image editing like InstructPix2Pix [7] and MagicBrush [65]. However, our zero-shot performance on the MagicBrush test set surpassed that of InstructPix2Pix, achieving performance similar to that of MagicBrush. (2) Additionally, our LLMGA offers a superior user experience in interactive editing, allowing image modifications to be carried out conversationally. In contrast, InstructPix2Pix only supports input via instructions and output as images. Visualized results are shown in Fig. 4, by leveraging the powerful reasoning capabilities of LLM, our LLMGA can provide more accurate and reasonable editing results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Evaluation on Inpainting and Outpainting. The results are shown in Tab. 3. (1) Comparing the 1st and 3rd rows of Tab. 3, our LLMGA-7B achieves significant improvements of 2.5681 FID over the SD1.5 in outpainting under wide masks. (2) In the 2nd and 3rd rows of Tab. 3, we make LLMGA imagine the complete generation prompts for the given masked images, which are then input into the later SD. Notably, our LLMGA-7B (with SD1.5-ft) demonstrates a significant FID improvement over LLMGA-7B (with SD1.5) in both outpainting and inpainting. This demonstrates the second stage of training makes SD better follow the prompts from MLLM. (3) Comparing the 3rd and 5th rows, our LLMGA-13B outperforms LLMGA-7B due to its superior reasoning capabilities." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 168, + 90, + 435, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 90, + 435, + 102 + ], + "spans": [ + { + "bbox": [ + 168, + 90, + 435, + 102 + ], + "type": "text", + "content": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 118, + 468, + 311 + ], + "blocks": [ + { + "bbox": [ + 149, + 118, + 468, + 311 + ], + "lines": [ + { + "bbox": [ + 149, + 118, + 468, + 311 + ], + "spans": [ + { + "bbox": [ + 149, + 118, + 468, + 311 + ], + "type": "image", + "image_path": "48c3f14583c3f198ad34333c853599803b868249c6aa9821c67403fb3f3e8411.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 185, + 323, + 427, + 335 + ], + "lines": [ + { + "bbox": [ + 185, + 323, + 427, + 335 + ], + "spans": [ + { + "bbox": [ + 185, + 323, + 427, + 335 + ], + "type": "text", + "content": "Fig. 4: Visual comparison on instruction-based editing." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 135, + 373, + 482, + 477 + ], + "blocks": [ + { + "bbox": [ + 138, + 352, + 474, + 364 + ], + "lines": [ + { + "bbox": [ + 138, + 352, + 474, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 352, + 474, + 364 + ], + "type": "text", + "content": "Table 3: Quantitative comparison for outpainting & inpainting on Places [68]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 373, + 482, + 477 + ], + "lines": [ + { + "bbox": [ + 135, + 373, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 135, + 373, + 482, + 477 + ], + "type": "table", + "html": "
MethodOutpaintingInpainting
NarrowMasksWideMasksNarrowMasksWideMasks
FID↓LPIPS↓FID↓LPIPS↓FID↓LPIPS↓FID↓LPIPS↓
SD1.52.01670.22835.00900.37341.07950.12361.28550.1434
LLMGA-7b (SD1.5)1.55300.22633.26300.36881.06920.12331.09830.1427
LLMGA-7b (SD1.5-ft, Ours)1.29730.22152.44090.36160.80270.11710.98070.1405
LLMGA-13b (SD1.5)1.56310.22633.08450.36791.03260.12281.09780.1426
LLMGA-13b (SD1.5-ft, Ours)1.21600.22102.36630.36090.79920.11660.97800.1400
", + "image_path": "a9bba7404f34e57eec5f9063c44aac7f517e89e5e69d473791037185d2c836cc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 506, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 579 + ], + "type": "text", + "content": "The qualitative results are shown in Fig. 5. We can see that LLMGA can deduce and imagine complete images based on masked input images. For example, in the 3rd row of Fig. 5, LLMGA can infer the presence of wind turbines on the mountain based on the given environment. Overall, LLMGA's powerful reasoning capability and extensive knowledge can assist users in conveniently making accurate and visually pleasing inpainting and outpainting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Evaluation on ControlNet. Our LLMGA demonstrates exceptional scalability, enabling integration with external plugins like ControlNet [66]. Here, we utilize LLMGA to create detailed prompts derived from input images and user requirements, working alongside ControlNet to guide image generation. As depicted in Fig. 6, our LLMGA significantly enhances the diversity and richness of outcomes in picture reference-guided image generation. Furthermore, more external plugins can also be integrated into the interactive framework of LLMGA," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "text", + "content": "LLMGA: Multimodal Large Language Model based Generation Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 114, + 473, + 290 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 473, + 290 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 473, + 290 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 473, + 290 + ], + "type": "image", + "image_path": "f61f8cd5dc4c567f736bd55e7b552c52df09cc2f898d5f77908e49f71cfc9897.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 298, + 432, + 310 + ], + "lines": [ + { + "bbox": [ + 181, + 298, + 432, + 310 + ], + "spans": [ + { + "bbox": [ + 181, + 298, + 432, + 310 + ], + "type": "text", + "content": "Fig. 5: Visual comparison on inpainting and outpainting." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 142, + 327, + 473, + 457 + ], + "blocks": [ + { + "bbox": [ + 142, + 327, + 473, + 457 + ], + "lines": [ + { + "bbox": [ + 142, + 327, + 473, + 457 + ], + "spans": [ + { + "bbox": [ + 142, + 327, + 473, + 457 + ], + "type": "image", + "image_path": "39cbd39d76c37d0e07e7cebf169c819561ff40536a55f56fa04e2d503d2f016f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 467, + 482, + 491 + ], + "lines": [ + { + "bbox": [ + 130, + 467, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 482, + 491 + ], + "type": "text", + "content": "Fig. 6: Visualization of LLMGA plus ControlNet. Our LLMGA can enhance the details in generated images, producing visually pleasing images." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 516, + 480, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 480, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 480, + 540 + ], + "type": "text", + "content": "combining with LLMGA's reasoning design capabilities and all previous functions to allow for a broader range of creative and engaging applications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 542, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 542, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 542, + 482, + 638 + ], + "type": "text", + "content": "Evaluation on Image Restoration. The results are shown in Tab. 4. For comparisons, we validate DiffIR [60] and our DiffRIR on the outpainting image generated by LLMGA. (1) Comparing the 2nd and 3rd rows of Tab. 4, it is evident that introducing a reference scheme can significantly improve restoration performance. (2) When comparing the 3rd and 4th rows of Tab. 4, it can be observed that introducing color degradation helps alleviate the bright and contrast distortion caused by SD. (3) Comparing the 1st and 4th rows of Tab. 4, our DiffRIR yields significant improvement, validating the effectiveness of DiffRIR." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 639, + 482, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 639, + 482, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 639, + 482, + 664 + ], + "type": "text", + "content": "As shown in Fig. 7, our DiffRIR (i.e., " + }, + { + "bbox": [ + 130, + 639, + 482, + 664 + ], + "type": "inline_equation", + "content": "\\mathrm{DiffRIR}_2" + }, + { + "bbox": [ + 130, + 639, + 482, + 664 + ], + "type": "text", + "content": " in Tab. 4) can alleviate the texture, brightness, and contrast discrepancies, and generate realistic details." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 168, + 90, + 434, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 90, + 434, + 102 + ], + "spans": [ + { + "bbox": [ + 168, + 90, + 434, + 102 + ], + "type": "text", + "content": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 186, + 136, + 428, + 206 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "type": "text", + "content": "Table 4: Quantitative comparison on image restoration for Places [68] outpainting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 186, + 136, + 428, + 206 + ], + "lines": [ + { + "bbox": [ + 186, + 136, + 428, + 206 + ], + "spans": [ + { + "bbox": [ + 186, + 136, + 428, + 206 + ], + "type": "table", + "html": "
MethodColorFID↓LPIPS↓
Reference Degradation
LLMGA-7BXX2.44090.3616
LLMGA+DiffIR [60]XX2.35870.3612
LLMGA+DiffRIR1✓X2.29930.3609
LLMGA+DiffRIR2 (Ours)✓✓2.26870.3607
", + "image_path": "16b2f5cca6a5cf80ccae3ac84dc00423609833342c7a1e6222cc2cfffdfbe0b9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 142, + 222, + 195, + 277 + ], + "blocks": [ + { + "bbox": [ + 142, + 222, + 195, + 277 + ], + "lines": [ + { + "bbox": [ + 142, + 222, + 195, + 277 + ], + "spans": [ + { + "bbox": [ + 142, + 222, + 195, + 277 + ], + "type": "image", + "image_path": "d61d967b941e41a79902a688b01a89953ded62a49980f7f9c6a0d288fa24ad7a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 196, + 223, + 251, + 277 + ], + "blocks": [ + { + "bbox": [ + 196, + 223, + 251, + 277 + ], + "lines": [ + { + "bbox": [ + 196, + 223, + 251, + 277 + ], + "spans": [ + { + "bbox": [ + 196, + 223, + 251, + 277 + ], + "type": "image", + "image_path": "1eb7e259f0810d9986e8570445abe9d6da67065a241ea358e8be52203e35ebab.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 280, + 238, + 288 + ], + "lines": [ + { + "bbox": [ + 212, + 280, + 238, + 288 + ], + "spans": [ + { + "bbox": [ + 212, + 280, + 238, + 288 + ], + "type": "text", + "content": "LLMGA" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 252, + 223, + 306, + 277 + ], + "blocks": [ + { + "bbox": [ + 252, + 223, + 306, + 277 + ], + "lines": [ + { + "bbox": [ + 252, + 223, + 306, + 277 + ], + "spans": [ + { + "bbox": [ + 252, + 223, + 306, + 277 + ], + "type": "image", + "image_path": "c9ad9fd42e21199e45b8b081645cd824c84eb78466adb47ada4112f7db65db87.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 256, + 279, + 299, + 288 + ], + "lines": [ + { + "bbox": [ + 256, + 279, + 299, + 288 + ], + "spans": [ + { + "bbox": [ + 256, + 279, + 299, + 288 + ], + "type": "text", + "content": "Masked Image" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 223, + 361, + 277 + ], + "blocks": [ + { + "bbox": [ + 307, + 223, + 361, + 277 + ], + "lines": [ + { + "bbox": [ + 307, + 223, + 361, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 223, + 361, + 277 + ], + "type": "image", + "image_path": "ff9e7e78133ad38a7383ed810676b6498be9903d3e35ac82c9217cc432aa7f2c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 362, + 223, + 416, + 277 + ], + "blocks": [ + { + "bbox": [ + 362, + 223, + 416, + 277 + ], + "lines": [ + { + "bbox": [ + 362, + 223, + 416, + 277 + ], + "spans": [ + { + "bbox": [ + 362, + 223, + 416, + 277 + ], + "type": "image", + "image_path": "75a3a4b1c266273737fa5c1332f15c326d5ada07257358fbe6372715fd2da9c0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 417, + 223, + 472, + 277 + ], + "blocks": [ + { + "bbox": [ + 417, + 223, + 472, + 277 + ], + "lines": [ + { + "bbox": [ + 417, + 223, + 472, + 277 + ], + "spans": [ + { + "bbox": [ + 417, + 223, + 472, + 277 + ], + "type": "image", + "image_path": "9c53a09691581ae00e8ddc042ab7d668b9c619b726ca9dbe8057f83c9e108b42.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 141, + 290, + 196, + 343 + ], + "blocks": [ + { + "bbox": [ + 146, + 280, + 190, + 289 + ], + "lines": [ + { + "bbox": [ + 146, + 280, + 190, + 289 + ], + "spans": [ + { + "bbox": [ + 146, + 280, + 190, + 289 + ], + "type": "text", + "content": "Masked Image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 290, + 196, + 343 + ], + "lines": [ + { + "bbox": [ + 141, + 290, + 196, + 343 + ], + "spans": [ + { + "bbox": [ + 141, + 290, + 196, + 343 + ], + "type": "image", + "image_path": "bfd0c6c1db5eb97d5b7db4ef18f14322762dd56b98557a47b15655ecfc1cca97.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 346, + 181, + 353 + ], + "lines": [ + { + "bbox": [ + 161, + 346, + 181, + 353 + ], + "spans": [ + { + "bbox": [ + 161, + 346, + 181, + 353 + ], + "type": "text", + "content": "DiffIR" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 363, + 481, + 385 + ], + "lines": [ + { + "bbox": [ + 132, + 363, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 481, + 385 + ], + "type": "text", + "content": "Fig. 7: Visual comparison of image restoration methods. DiffRIR can alleviate the texture, contrast, and brightness disparities in inpainting & outpainting results." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 196, + 290, + 251, + 343 + ], + "blocks": [ + { + "bbox": [ + 196, + 290, + 251, + 343 + ], + "lines": [ + { + "bbox": [ + 196, + 290, + 251, + 343 + ], + "spans": [ + { + "bbox": [ + 196, + 290, + 251, + 343 + ], + "type": "image", + "image_path": "c4e0d036b74494b9d7b8132081ffe26561ca341b31efb0931527214ac7eced22.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 202, + 346, + 247, + 354 + ], + "lines": [ + { + "bbox": [ + 202, + 346, + 247, + 354 + ], + "spans": [ + { + "bbox": [ + 202, + 346, + 247, + 354 + ], + "type": "text", + "content": "DiffRIR (Ours)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 252, + 290, + 306, + 343 + ], + "blocks": [ + { + "bbox": [ + 252, + 290, + 306, + 343 + ], + "lines": [ + { + "bbox": [ + 252, + 290, + 306, + 343 + ], + "spans": [ + { + "bbox": [ + 252, + 290, + 306, + 343 + ], + "type": "image", + "image_path": "39ee74a6f76f0311326040847369f1ed36efc9463b8d7d0f9f31d580740f6c61.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 346, + 290, + 353 + ], + "lines": [ + { + "bbox": [ + 270, + 346, + 290, + 353 + ], + "spans": [ + { + "bbox": [ + 270, + 346, + 290, + 353 + ], + "type": "text", + "content": "DiffIR" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 307, + 290, + 361, + 343 + ], + "blocks": [ + { + "bbox": [ + 321, + 280, + 348, + 288 + ], + "lines": [ + { + "bbox": [ + 321, + 280, + 348, + 288 + ], + "spans": [ + { + "bbox": [ + 321, + 280, + 348, + 288 + ], + "type": "text", + "content": "LLMGA" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 290, + 361, + 343 + ], + "lines": [ + { + "bbox": [ + 307, + 290, + 361, + 343 + ], + "spans": [ + { + "bbox": [ + 307, + 290, + 361, + 343 + ], + "type": "image", + "image_path": "7fe13a363b7c88b0aba6269d13a976fab0287419fb6c00a23b21be4a5e77ff98.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 346, + 356, + 354 + ], + "lines": [ + { + "bbox": [ + 312, + 346, + 356, + 354 + ], + "spans": [ + { + "bbox": [ + 312, + 346, + 356, + 354 + ], + "type": "text", + "content": "DiffRIR (Ours)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 362, + 290, + 416, + 343 + ], + "blocks": [ + { + "bbox": [ + 369, + 280, + 411, + 289 + ], + "lines": [ + { + "bbox": [ + 369, + 280, + 411, + 289 + ], + "spans": [ + { + "bbox": [ + 369, + 280, + 411, + 289 + ], + "type": "text", + "content": "Masked Image" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 362, + 290, + 416, + 343 + ], + "lines": [ + { + "bbox": [ + 362, + 290, + 416, + 343 + ], + "spans": [ + { + "bbox": [ + 362, + 290, + 416, + 343 + ], + "type": "image", + "image_path": "c099d4e7b008a726198e3aee2f5912569ba7db41e4f23094f58749cf7cbbb948.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 346, + 400, + 353 + ], + "lines": [ + { + "bbox": [ + 381, + 346, + 400, + 353 + ], + "spans": [ + { + "bbox": [ + 381, + 346, + 400, + 353 + ], + "type": "text", + "content": "DiffIR" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 417, + 290, + 472, + 343 + ], + "blocks": [ + { + "bbox": [ + 432, + 280, + 457, + 288 + ], + "lines": [ + { + "bbox": [ + 432, + 280, + 457, + 288 + ], + "spans": [ + { + "bbox": [ + 432, + 280, + 457, + 288 + ], + "type": "text", + "content": "LLMGA" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 290, + 472, + 343 + ], + "lines": [ + { + "bbox": [ + 417, + 290, + 472, + 343 + ], + "spans": [ + { + "bbox": [ + 417, + 290, + 472, + 343 + ], + "type": "image", + "image_path": "368b233adacba37700fedfdae6dd1fbd1200efdb273bebf78a4438ac48c87174.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 346, + 466, + 354 + ], + "lines": [ + { + "bbox": [ + 422, + 346, + 466, + 354 + ], + "spans": [ + { + "bbox": [ + 422, + 346, + 466, + 354 + ], + "type": "text", + "content": "DiffRIR (Ours)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 130, + 414, + 298, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 298, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 298, + 438 + ], + "type": "text", + "content": "Control SD using detailed language prompt or embedding?" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 130, + 438, + 298, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 298, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 298, + 617 + ], + "type": "text", + "content": "The results are shown in Fig. 5. We compare two approaches: GILL [29], which makes LLM estimate a fixed-sized embedding to control SD generation, and LLMGA Embedding, a variant of LLMGA where the language prompt is replaced with embedding, undergoing the same training process as LLMGA. The evaluation is conducted on MSCOCO by instructing these methods to generate images with the same prompts in multiple times in conversation form. (1) The quality of generated images (Fig. 5) in embedding-based methods" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "(i.e., GILL and LLMGA Embedding) deteriorate rapidly as the number of conversation turns increases. In contrast, our LLMGA remains unaffected. This discrepancy arises from the inherent noise present in the embeddings predicted by LLM. As the number of conversation turns rises, these generated embeddings" + } + ] + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 312, + 418, + 476, + 543 + ], + "blocks": [ + { + "bbox": [ + 312, + 418, + 476, + 543 + ], + "lines": [ + { + "bbox": [ + 312, + 418, + 476, + 543 + ], + "spans": [ + { + "bbox": [ + 312, + 418, + 476, + 543 + ], + "type": "image", + "image_path": "031046ecd895a4ef2f28532213d538f8fe3ee36e92da00499658e49e761ce098.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 552, + 482, + 586 + ], + "lines": [ + { + "bbox": [ + 304, + 552, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 482, + 586 + ], + "type": "text", + "content": "Table 5: T2I performance comparison of SD control schemes based on detailed language prompt and embedding." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "text", + "content": "LLMGA: Multimodal Large Language Model based Generation Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 157, + 481, + 232 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "text", + "content": "Table 6: Datasets comparison. We conducted comparisons of FID on T2I and outpainting. The " + }, + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "text", + "content": " signifies the utilization of the complete dataset during training. Conversely, the absence of the " + }, + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "text", + "content": " indicates a reduction to only " + }, + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "text", + "content": " of the original datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 157, + 481, + 232 + ], + "lines": [ + { + "bbox": [ + 134, + 157, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 134, + 157, + 481, + 232 + ], + "type": "table", + "html": "
MethodPrompt RefinementInpainting &OutpaintingSimilar Image GenerationInstruction-based EditingT2IOutpainting
LLMGA1✓✓✓21.44602.6533
LLMGA2✓✓✓19.19143.1054
LLMGA3✓✓✓19.56982.6612
LLMGA4✓✓✓19.06422.8047
LLMGA5 (Ours)✓✓✓✓18.52342.4409
", + "image_path": "0d398b852fc49e5cb0ebd0827dcefd780a2de08b4f92255f968a2840ef0e3004.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 254, + 480, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 254, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 254, + 480, + 324 + ], + "type": "text", + "content": "integrate with the preceding conversations, introducing even more noise. This poses challenges for the precise control of SD-generated content. Our LLMGA addresses this issue by mapping the embedding to the fixed language domain, effectively eliminating such noise. (2) Additionally, LLMGA Embedding also outperforms GILL, indicating that the prompt size used to guide SD generation should be adaptive in content, rather than a fixed size." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 326, + 481, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 433 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 433 + ], + "type": "text", + "content": "Contribution of Training Data. To assess the impact of training data, we downsized one of the four training datasets in " + }, + { + "bbox": [ + 132, + 326, + 481, + 433 + ], + "type": "inline_equation", + "content": "\\mathrm{LLMGA}_5" + }, + { + "bbox": [ + 132, + 326, + 481, + 433 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 326, + 481, + 433 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 132, + 326, + 481, + 433 + ], + "type": "text", + "content": " of its original magnitude, ensuring that LLMGA remains capable of furnishing responses in the prescribed format. The results are shown in Tab. 6. It is evident that prompt refinement, inpainting & outpainting, and instruction-based editing datasets enhance LLMGA's comprehension of image generation and editing properties, resulting in superior images. Moreover, comparing " + }, + { + "bbox": [ + 132, + 326, + 481, + 433 + ], + "type": "inline_equation", + "content": "\\mathrm{LLMGA}_3" + }, + { + "bbox": [ + 132, + 326, + 481, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 326, + 481, + 433 + ], + "type": "inline_equation", + "content": "\\mathrm{LLMGA}_5" + }, + { + "bbox": [ + 132, + 326, + 481, + 433 + ], + "type": "text", + "content": ", we can see that engaging in similar image generation training further improves the performance of LLMGA in both generation and editing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 450, + 218, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 450, + 218, + 462 + ], + "spans": [ + { + "bbox": [ + 133, + 450, + 218, + 462 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 474, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 474, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 474, + 481, + 665 + ], + "type": "text", + "content": "LLM possesses an extensive reservoir of knowledge and powerful comprehension and reasoning capabilities. In this paper, we introduce a MLLM-based generation assistant (LLMGA), aiming to exploit LLM's capabilities in an interactive manner to facilitate more efficient and convenient image generation and editing. Compared to relying on LLM to predict a fixed-size embedding to control SD, we employ detailed generation prompts. These prompts prove to be more favorable for enhancing LLM's contextual comprehension and generating more accurate and rich content. To this end, we develop a two-stage training scheme and curate a dataset, including four parts: prompt refinement, similar image generation, inpainting & outpainting, and instruction-based editing. For the first stage, we train MLLM to understand the properties of image generation and editing, enabling it to give fitting responses. For the second stage, we optimize the SD unet to adapt to the generation prompt. Moreover, we propose a DM-based reference restoration network (DiffRIR) to mitigate disparities in texture, contrast, and brightness for image editing. Consequently, LLMGA can offer design suggestions and enhance results based on user's requests during interactions." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 168, + 90, + 433, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 90, + 433, + 102 + ], + "spans": [ + { + "bbox": [ + 168, + 90, + 433, + 102 + ], + "type": "text", + "content": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 482, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 482, + 166 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 482, + 166 + ], + "type": "text", + "content": "This work was supported in part by the Research Grants Council under the Areas of Excellence scheme grant AoE/E-601/22-R." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 211, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 211, + 481, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 481, + 244 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 481, + 244 + ], + "type": "text", + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. NeurIPS (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 267 + ], + "type": "text", + "content": "2. Austin, J., Johnson, D.D., Ho, J., Tarlow, D., van den Berg, R.: Structured denoising diffusion models in discrete state-spaces. NeurIPS (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 269, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 481, + 300 + ], + "type": "text", + "content": "3. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 301, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 481, + 322 + ], + "type": "text", + "content": "4. Bao, F., Nie, S., Xue, K., Cao, Y., Li, C., Su, H., Zhu, J.: All are worth words: A vit backbone for diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 323, + 481, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 481, + 355 + ], + "type": "text", + "content": "5. Bao, F., Nie, S., Xue, K., Li, C., Pu, S., Wang, Y., Yue, G., Cao, Y., Su, H., Zhu, J.: One transformer fits all distributions in multi-modal diffusion at scale. arXiv preprint arXiv:2303.06555 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 357, + 481, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 481, + 378 + ], + "type": "text", + "content": "6. Batzolis, G., Stanczuk, J., Schonlieb, C.B., Etmann, C.: Conditional image generation with score-based diffusion models. arXiv preprint arXiv:2111.13606 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 379, + 481, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 481, + 400 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 481, + 400 + ], + "type": "text", + "content": "7. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 401, + 481, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 401, + 481, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 401, + 481, + 433 + ], + "type": "text", + "content": "8. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. NeurIPS (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 434, + 481, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 481, + 477 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 481, + 477 + ], + "type": "text", + "content": "9. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 138, + 434, + 481, + 477 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 138, + 434, + 481, + 477 + ], + "type": "text", + "content": " chatgpt quality (March 2023), https://lmsys.org/blog/2023-03-30-vicuna/" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 479, + 481, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 479, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 134, + 479, + 481, + 510 + ], + "type": "text", + "content": "10. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 512, + 481, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 512, + 481, + 543 + ], + "spans": [ + { + "bbox": [ + 134, + 512, + 481, + 543 + ], + "type": "text", + "content": "11. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 544, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 544, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 134, + 544, + 481, + 577 + ], + "type": "text", + "content": "12. Dai, X., Hou, J., Ma, C.Y., Tsai, S., Wang, J., Wang, R., Zhang, P., Vandenhende, S., Wang, X., Dubey, A., et al.: Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 578, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 578, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 134, + 578, + 481, + 598 + ], + "type": "text", + "content": "13. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "type": "text", + "content": "14. Dong, R., Han, C., Peng, Y., Qi, Z., Ge, Z., Yang, J., Zhao, L., Sun, J., Zhou, H., Wei, H., et al.: Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "text", + "content": "15. Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Thompson, J., Vuong, Q., Yu, T., et al.: Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "text", + "content": "LLMGA: Multimodal Large Language Model based Generation Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "16. Fan, W.C., Chen, Y.C., Chen, D., Cheng, Y., Yuan, L., Wang, Y.C.F.: Frido: Feature pyramid diffusion for complex scene image synthesis. In: AAAI (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "text", + "content": "17. Feng, Z., Zhang, Z., Yu, X., Fang, Y., Li, L., Chen, X., Lu, Y., Liu, J., Yin, W., Feng, S., et al.: Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. In: CVPR (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "type": "text", + "content": "18. Ge, Y., Ge, Y., Zeng, Z., Wang, X., Shan, Y.: Planting a seed of vision in large language model. arXiv preprint arXiv:2307.08041 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 215 + ], + "type": "text", + "content": "19. Gu, S., Chen, D., Bao, J., Wen, F., Zhang, B., Chen, D., Yuan, L., Guo, B.: Vector quantized diffusion model for text-to-image synthesis. In: CVPR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "text", + "content": "20. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "21. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "text", + "content": "22. Ho, J., Sahara, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. JMLR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "23. Huang, S., Dong, L., Wang, W., Hao, Y., Singhal, S., Ma, S., Lv, T., Cui, L., Mohammed, O.K., Liu, Q., et al.: Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "24. Jiang, A.Q., Sablayrolles, A., Roux, A., Mensch, A., Savary, B., Bamford, C., Chaplot, D.S., Casas, D.d.l., Hanna, E.B., Bressand, F., et al.: Mixtral of experts. arXiv preprint arXiv:2401.04088 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "25. Ju, X., Zeng, A., Bian, Y., Liu, S., Xu, Q.: Direct inversion: Boosting diffusion-based editing with 3 lines of code. ICLR (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "26. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "27. Kim, G., Kwon, T., Ye, J.C.: Diffusionclip: Text-guided diffusion models for robust image manipulation. In: CVPR (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "text", + "content": "28. Kingma, D., Salimans, T., Poole, B., Ho, J.: Variational diffusion models. NeurIPS (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "text", + "content": "29. Koh, J.Y., Fried, D., Salakhutdinov, R.: Generating images with multimodal language models. arXiv preprint arXiv:2305.17216 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "30. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "text", + "content": "31. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "text", + "content": "32. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 545, + 460, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 460, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 460, + 555 + ], + "type": "text", + "content": "33. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. NeurIPS (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 556, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 480, + 578 + ], + "type": "text", + "content": "34. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: CVPR (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "text", + "content": "35. Meng, C., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073 (2021)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "36. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 644, + 307, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 307, + 654 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 307, + 654 + ], + "type": "text", + "content": "37. OpenAI: Gpt-4 technical report (2023)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 654, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 654, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 654, + 480, + 665 + ], + "type": "text", + "content": "38. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: ICCV (2023)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 168, + 90, + 433, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 90, + 433, + 102 + ], + "spans": [ + { + "bbox": [ + 168, + 90, + 433, + 102 + ], + "type": "text", + "content": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "text", + "content": "39. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. NeurIPS (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 181 + ], + "type": "text", + "content": "40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "text", + "content": "41. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "42. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 256 + ], + "type": "text", + "content": "43. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 257, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 277 + ], + "type": "text", + "content": "44. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 310 + ], + "type": "text", + "content": "45. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 311, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 342 + ], + "type": "text", + "content": "46. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 343, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 374 + ], + "type": "text", + "content": "47. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "text", + "content": "48. Shen, Y., Song, K., Tan, X., Li, D., Lu, W., Zhuang, Y.: Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "type": "text", + "content": "49. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 429, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 449 + ], + "type": "text", + "content": "50. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 449, + 480, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 480, + 481 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 480, + 481 + ], + "type": "text", + "content": "51. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "type": "text", + "content": "52. Sun, Q., Yu, Q., Cui, Y., Zhang, F., Zhang, X., Wang, Y., Gao, H., Liu, J., Huang, T., Wang, X.: Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "type": "text", + "content": "53. Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 547, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 578 + ], + "type": "text", + "content": "54. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 579, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 610 + ], + "type": "text", + "content": "55. Valevski, D., Kalman, M., Matias, Y., Leviathan, Y.: Unitune: Text-driven image editing by fine tuning an image generation model on a single image. arXiv preprint arXiv:2210.09477 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "56. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: ICCVW (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "57. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 149, + 91, + 447, + 102 + ], + "type": "text", + "content": "LLMGA: Multimodal Large Language Model based Generation Assistant" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 456 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "58. Wu, S., Fei, H., Qu, L., Ji, W., Chua, T.S.: Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "text", + "content": "59. Xia, B., Zhang, Y., Wang, S., Wang, Y., Wu, X., Tian, Y., Yang, W., Timotfe, R., Van Gool, L.: Diffie2i: Efficient diffusion model for image-to-image translation. arXiv preprint arXiv:2308.13767 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "60. Xia, B., Zhang, Y., Wang, S., Wang, Y., Wu, X., Tian, Y., Yang, W., Van Gool, L.: Diffir: Efficient diffusion model for image restoration. ICCV (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "text", + "content": "61. Xia, B., Zhang, Y., Wang, Y., Tian, Y., Yang, W., Timofte, R., Van Gool, L.: Knowledge distillation based degradation estimation for blind super-resolution. ICLR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "62. Xue, Z., Song, G., Guo, Q., Liu, B., Zong, Z., Liu, Y., Luo, P.: Raphael: Text-to-image generation via large mixture of diffusion paths. arXiv preprint arXiv:2305.18295 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "text", + "content": "63. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "64. Yu, L., Shi, B., Pasunuru, R., Muller, B., Golovneva, O., Wang, T., Babu, A., Tang, B., Karrer, B., Sheynin, S., et al.: Scaling autoregressive multi-modal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "type": "text", + "content": "65. Zhang, K., Mo, L., Chen, W., Sun, H., Su, Y.: Magicbrush: A manually annotated dataset for instruction-guided image editing. NeurIPS (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "66. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "67. Zhong, S., Huang, Z., Wen, W., Qin, J., Lin, L.: Sur-adapter: Enhancing text-to-image pre-trained diffusion models with large language models. In: ACM MM (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 424 + ], + "type": "text", + "content": "68. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. TPAMI (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "69. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 168, + 91, + 433, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 91, + 433, + 102 + ], + "spans": [ + { + "bbox": [ + 168, + 91, + 433, + 102 + ], + "type": "text", + "content": "Bin Xia, Shiyin Wang, Yingfan Tao, Yitong Wang, and Jiaya Jia" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LLaMA-VID_ An Image is Worth 2 Tokens in Large Language Models/022c7268-b560-44a0-867d-3ec58c86bb1a_content_list.json b/2024/LLaMA-VID_ An Image is Worth 2 Tokens in Large Language Models/022c7268-b560-44a0-867d-3ec58c86bb1a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fc44bb64da131a82e9feea9ef23630e111ef91bb --- /dev/null +++ b/2024/LLaMA-VID_ An Image is Worth 2 Tokens in Large Language Models/022c7268-b560-44a0-867d-3ec58c86bb1a_content_list.json @@ -0,0 +1,2070 @@ +[ + { + "type": "text", + "text": "LLaMA-VID: An Image is Worth 2 Tokens in Large Language Models", + "text_level": 1, + "bbox": [ + 256, + 140, + 746, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yanwei Li $^{1\\star}$ Chengyao Wang $^{1*}$ Jiaya Jia $^{1,2}$", + "bbox": [ + 328, + 210, + 673, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CUHK $^{1}$ SmartMore $^{2}$", + "bbox": [ + 424, + 238, + 576, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this work, we present a novel method to tackle the token generation challenge in Vision Language Models (VLMs) for video and image understanding, called LLaMA-VID. Current VLMs, while proficient in tasks like image captioning and visual question answering, face computational burdens when processing long videos due to the excessive visual tokens. LLaMA-VID addresses this issue by representing each frame with two distinct tokens, namely context token and content token. The context token encodes the overall image context based on user input, whereas the content token encapsulates visual cues in each frame. This dual-token strategy significantly reduces the overload of long videos while preserving critical information. Generally, LLaMA-VID empowers existing frameworks to support hour-long videos and pushes their upper limit with an extra context token. It is demonstrated to surpass previous methods on most of video- or image-based benchmarks. Code and models are available at https://github.com/dvlab-research/LLaMA-VID.", + "bbox": [ + 261, + 292, + 743, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 526, + 375, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large Language Models (LLMs) [44, 54, 66], through their capacity to generate contextually accurate responses, have significantly advanced the field of AI. Drawing from the strengths of LLMs, Vision Language Models (VLMs) [13,33,45] have been developed to extend these capabilities to visual data, demonstrating their adeptness in tasks like image captioning and visual question answering. However, a substantial challenge emerges in the context of long video, where an excessive number of tokens are required to represent consecutive frames. The computational demands escalate with the video length, thereby constraining the practical application of VLMs for extensive videos.", + "bbox": [ + 212, + 558, + 787, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, several approaches have been proposed to handle videos, moving beyond image-only VLMs. These methods aim to alleviate the token issue by utilizing representative queries [30, 64] or applying temporal compression [38, 39, 52]. Despite these efforts, the challenge of long videos remains unresolved. The primary obstacle stems from the excessive number of tokens required for each video frame. For instance, models like BLIP [13, 29] and LLaVA [33] require 32 and over 256 tokens respectively for a single image. A video containing 10K frames would thus necessitate over 320K tokens, exceeding the capacity of current VLMs.", + "bbox": [ + 212, + 695, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* equal contribution", + "bbox": [ + 217, + 825, + 357, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f29a384f07e4c9034fdc8fc584dea006ebd3d391385baac25d928b725800a86a.jpg", + "image_caption": [ + "(a) Video-based benchmarks" + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 498, + 323 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/31ed8192d04c010148629d519032cf3f54fca4eb3ee276cc14e594824bef8d96.jpg", + "image_caption": [ + "(b) Image-based benchmarks", + "Fig. 1: The proposed LLaMA-VID achieves leading performance on most of benchmarks with 7B LLMs. The video-based and image-based benchmarks are noted in blue and purple color, respectively. Please refer to Tables 1, 2, and 5 for more details." + ], + "image_footnote": [], + "bbox": [ + 527, + 142, + 759, + 324 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, simple temporal compression can significantly damage the representation over long-term intervals. This drawback hampers their performance, thereby underscoring the need for a robust solution.", + "bbox": [ + 212, + 433, + 784, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present LLaMA-VID, a novel approach to effectively manage the token generation issue in long videos. Our core idea is to represent each video frame with two distinct tokens: context token and content token. The context token is designed to encode the overall context of the image based on user input, which efficiently condenses the broader picture into a single token. Simultaneously, the content token captures finer aspects of each frame. According to computational constraints, the length of content token can be extended to include more details, e.g., 1 token/frame for video input and beyond 256 token/frame for single image. In this way, the overload of long videos can be significantly reduced without sacrificing critical information.", + "bbox": [ + 212, + 484, + 787, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In particular, our method employs a dual-token generation strategy that is both efficient and effective. For each frame, we first extract image features using a pre-trained vision transformer [15], akin to other VLMs [13, 33]. The key question is how to generate the context-related token according to user instructions. We provide the solution by leveraging the cross-modality design [14,29] for instruction-guided queries, which carry the interactive intention from users. For context token, these queries interact with previously generated image features in the designed attention module, termed as context attention. To generate content token, the image features are average pooled to formulate tokens that adapt to different settings. For instance, global pooling is adopted to maintain efficiency for video input while details are preserved with more tokens for single image input. The context and content tokens are subsequently projected to the space of LLMs with simple linear layers for final prediction. Furthermore, to better sup", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "L. Yanwei et al.", + "bbox": [ + 271, + 114, + 379, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "port hour-long videos in VLMs, we construct an instruction dataset that contains 9K movie-level conversations for plot reasoning and detail understanding.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generally, LLaMA-VID can be distinguished from two aspects. On one hand, with the dual-token paradigm, each frame can be efficiently encoded with only two tokens, which empowers existing LLMs to support long videos. On the other hand, the context token aggregates the most informative feature of each image, which further extends the upper limit of VLMs with an extra token.", + "bbox": [ + 212, + 176, + 784, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall framework, dubbed LLaMA-VID, can be easily instantiated with various decoders and LLMs, as elaborated in Section 3. Extensive empirical studies are conducted in Section 4 to reveal the effectiveness of each component. Remarkably, our model can complete training within 2 days on a single machine with $8 \\times \\mathrm{A}100$ GPUs, and it outperforms previous leading methods on most of video- and image-based benchmarks, as shown in Figure 1.", + "bbox": [ + 212, + 252, + 785, + 342 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 364, + 387, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first review large language models and delve into recent advances in vision language models.", + "bbox": [ + 214, + 397, + 782, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Large Language Models", + "text_level": 1, + "bbox": [ + 214, + 450, + 459, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The field of Natural Language Processing (NLP) has witnessed tremendous advancements with the evolution of LLMs. Transformer [55] marked a pivotal milestone, with subsequent language models [14, 35, 66] demonstrating remarkable capabilities. GPT [6] revolutionized this field by utilizing generative pre-trained transformers for auto-regressive prediction, which is proved to be a potent language modeling paradigm. Recent groundbreaking works, such as ChatGPT [44], GPT-4 [45], and LLaMA [54], have pushed the boundaries even further. Trained on vast amounts of text data, these models exhibit exceptional capabilities in complex linguistic tasks. To leverage the potential of pre-trained LLMs, instruction tuning [46, 57] is a crucial component for high-quality output. This strategy is widely adopted in open-source models like Alpaca [53] and Vicuna [12], which improve over LLaMA [54] using specially designed instruction pairs. There are also researches [58, 61] that utilize the reasoning ability of LLMs and invoke pre-defined tools for visual applications. Different from them, we collect multimodality instruction data that contains text, images, and videos in this work, which is employed to empower LLMs for long video processing.", + "bbox": [ + 212, + 474, + 785, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Vision Language Models", + "text_level": 1, + "bbox": [ + 214, + 739, + 464, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The advancements in computer vision and NLP have led to the emergence of vision-language models (VLMs) that integrate vision models with language models for cross-modality understanding [10,59] and reasoning [19,27,37]. Pioneering large-scale VLMs like CLIP [47] and ALIGN [24] have extended language models to vision-language tasks. The recent progress has seen an increasing focus", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "LLaMA-VID", + "bbox": [ + 642, + 113, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3386a0651921af16022c75a39b6ad5c4586f118f8c530a9f1590adf29a9a5540.jpg", + "image_caption": [ + "Fig. 2: The framework of LLaMA-VID. With user directive, it operates by taking either a single image or video frames as input, and generates responses from LLM. The process initiates with a visual encoder that transforms input frames into the visual embedding. Then, the text decoder produces text queries based on the user input. In context attention, the text query aggregates text-related visual cues. For efficiency, an option is provided to downsample the visual embedding to various token sizes, or even to a single token. The text-guided context token and the visually-enriched content token are then formulated using a linear projector to represent each frame at time $t$ . Finally, the LLM takes the user directive and all visual tokens as input and gives responses." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 782, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "on leveraging the power of LLMs. Notably, Flamingo [2] and BLIP-2 [29] utilize web-scale image-text pairs for cross-modality alignment, thereby enhancing learning performance. To further exploit the potential of such pre-trained models, InstructBLIP [13] and MiniGPT-4 [67] construct high-quality instruction pairs based on BLIP-2 and achieve superior results. Simultaneously, LLaVA [33] employs a simple linear projector with a few learnable parameters to align the image and text space of LLaMA. Given the tailored instruction data, this straightforward approach demonstrates strong capabilities. To support video understanding in LLMs, several studies [30,52,64] attempt to utilize BLIP-2 for video embedding or text-only caption [63] extraction, while Video-ChatGPT [39] proposes spatial and temporal pooling for video features. However, given the substantial number of tokens required for each frame, LLMs encounter significant challenges when processing extensive video sequences. It prevents previous work from representing long video sequences that exceed a duration of one hour in LLMs. To solve the issue, we propose to efficiently encode each frame with only 2 tokens, which supports long video understanding in existing LLMs.", + "bbox": [ + 212, + 446, + 787, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 LLaMA-VID", + "text_level": 1, + "bbox": [ + 215, + 722, + 377, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The framework of LLaMA-VID is conceptually simple: encoder and decoder are adopted to produce visual embedding and text-guided features, respectively; context token and content token are transformed with the tailored token generation strategy; instruction tuning is designed to unleash the potential of LLMs for image and video.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "L. Yanwei et al.", + "bbox": [ + 271, + 114, + 377, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Encoder and Decoder", + "text_level": 1, + "bbox": [ + 215, + 146, + 441, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed LLaMA-VID can be utilized to interact with single image or long videos. For clarity, we assume the input image is captured from a video sequence, as presented in Figure 2. Given a video frame $\\mathbf{V}_t \\in \\mathbb{R}^{H \\times W \\times 3}$ at time $t$ , a transformer-based visual encoder is first employed to produce the visual embedding $\\mathbf{X}_t \\in \\mathbb{R}^{N \\times C}$ . Here, $N = H / p \\times W / p$ and $C$ indicate the number of image patches and embedding channels, respectively. The patch size $p$ is typically set to 14 for ViT-based backbones [15, 16, 47]. Meanwhile, we take the user instruction as input and generate the text-guided query $\\mathbf{Q}_t \\in \\mathbb{R}^{M \\times C}$ with the produced $\\mathbf{X}_t$ , where $M$ denotes the number of queries. As depicted in Figure 2, this cross-modality interaction predominantly occurs in the text decoder, which can be easily instantiated with BERT [14] or QFormer [13], as compared in Table 8. In this way, the text query $\\mathbf{Q}_t$ contains highlighted visual cues that are most related to the user instruction.", + "bbox": [ + 212, + 176, + 787, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Token Generation", + "text_level": 1, + "bbox": [ + 215, + 401, + 410, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the text query $\\mathbf{Q}_t$ and visual embedding $\\mathbf{X}_t$ , we can easily generate representative tokens for LLMs. Specifically, context attention is designed to aggregate text-related visual features and condense them to a single context token. As shown in Figure 2, it takes $\\mathbf{Q}_t$ and $\\mathbf{X}_t$ as input and formulates the context-related embedding $\\mathbf{E}_t \\in \\mathbb{R}^{1 \\times C}$ as", + "bbox": [ + 212, + 431, + 785, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} _ {t} = \\operatorname {M e a n} \\left(\\operatorname {S o f t m a x} \\left(\\left(\\mathbf {Q} _ {t} \\times \\mathbf {X} _ {t} ^ {\\top}\\right) / \\sqrt {C}\\right) \\times \\mathbf {X} _ {t}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 525, + 785, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the Softmax function and Mean operation are conducted along the $N$ and $M$ dimensions, respectively. Unlike QFormer [13] that adopts 32 visual queries as LLMs tokens, we only utilize the text query $\\mathbf{Q}_t$ to aggregate the visual features with high-response scores to input instructions. As a result, the most crucial visual cues related to user input are efficiently preserved in the condensed embedding $\\mathbf{E}_t$ . The effectiveness of this context-related token generation is demonstrated in Table 6 and Figure 6. Subsequently, a linear projector is utilized to transform the embedding $\\mathbf{E}_t$ into the context token $\\mathbf{E}_t^T \\in \\mathbb{R}^{1 \\times C}$ , which aligns with the language space of LLMs. Meanwhile, we employ an adaptive pooling strategy for the visual embedding according to computational constraints to produce the content token $\\mathbf{E}_t^V \\in \\mathbb{R}^{n \\times C}$ , where $n \\in [1, N]$ . For instance, we maintain the original resolution of visual embedding $\\mathbf{X}_t$ when input single image, while we downsample $\\mathbf{X}_t$ to 1 token for long videos. This approach significantly reduces the overload of LLMs for each frame, thereby supporting hour-long videos effectively. Finally, the generated context token $\\mathbf{E}_t^T$ and the content token $\\mathbf{E}_t^V$ are concatenated to represent the frame at time $t$ . Along with frames at other timestamps, the entire video sequence is translated into the language space in token format, which is then used to generate responses from LLMs. The whole process is summarized in Algorithm 1.", + "bbox": [ + 212, + 551, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "LLaMA-VID", + "bbox": [ + 642, + 113, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1 Pseudo Code for Token Generation.", + "Stage 1: Modality Alignment", + "Stage 3: Long Video Tuning (Optional)" + ], + "code_body": "B: batch size; C: channel size; n: content shape\n# M: query length; N: shape of flatten image pacthes;\n# text_q: text query in shape (B, M, C)\n# vis_embedding: visual embedding in shape (B, N, C)\n# Key part 1: calculate context-related embedding\nctx_embedding = text_q @ vis_embedding.transpose(-1,-2)\nctx_embedding = ctx_embedding / (vis_embedding.shape[-1]**0.5)\nctx_embedding = (ctx_embedding-largest(-1)@vis_embedding).mean(1)\nctx_embedding = self.batchproj(ctx_embedding[:,None])\n# Key part 2: calculate visual embedding\ncur_shape = int(vis_embedding.shape[1]**0.5)\nvis_embedding = vis_embedding.reshape(B, cur_shape, -1, C)\nvis_embedding = F AVG_pool2d(vis_embedding.permute(0,3,1,2), kernel_size=cur_shape//n,\nstride=cur_shape//n)\nvis_embedding = vis_embedding.permute(0,2,3,1).flatten(1,2)\nvis_embedding = self.visproj(vis_embedding)\n# concat token in shape (B, n+1, C), n in [1,N]\nfinal_token = torch.cat([ctx_embedding, vis_embedding], dim=1)", + "guess_lang": "python", + "bbox": [ + 215, + 171, + 759, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Training Strategy", + "text_level": 1, + "bbox": [ + 215, + 454, + 408, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training strategy, particularly instruction tuning, has proven to be crucial in LLMs [12, 53, 54] and VLMs [13, 32, 33]. Considering training efficiency, in this work, we divide the training procedure into three stages, i.e., modality alignment, instruction tuning, and long video tuning.", + "bbox": [ + 214, + 479, + 787, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Modality Alignment. As shown in Figure 2, each video frame is projected into the space of LLMs in each forward pass. Therefore, it is crucial to ensure visual features are well aligned with the language space. To this end, we construct a compact dataset that contains 790K high-quality image- and video-caption pairs.", + "bbox": [ + 214, + 555, + 787, + 617 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/62f3d2892bcd1ea522a3318d308948c4fa02c4266572185b880f28f9a46a99fc.jpg", + "image_caption": [ + "Fig.3: Multimodal data distribution and instruction format for each model training stage. and denote the token for single image and the $i$ -th video frame, respectively." + ], + "image_footnote": [], + "bbox": [ + 222, + 655, + 473, + 779 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/27d3cdddae9ee3ab385bc27e54e68df0c0f44ca9c5723681c3e3b91829be4218.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 655, + 779, + 709 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "L. Yanwei et al.", + "bbox": [ + 269, + 114, + 377, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7f197feedc3a18f177efe7ec95c499f55cbe7cb9742d4a77850dcf4ece39ae61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 148, + 493, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2c5aa9aa91eea4ed1cf89a2e60911465da7c1cc646a24ac7c0814ae5d634e8de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 204, + 491, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/75471a82147bf9acafe1c681dd3906adc4b3c01023236f39b97cb330adccabda.jpg", + "image_caption": [ + "Fig. 4: An example to construct instruction pairs for the movie Titanic. Given the movie synopsis and script, we utilize the developed LLMs like GPT-4 [45] and Claude-2 [3] to generate movie summaries, plot-related QA pairs, and general reasoning pairs." + ], + "image_footnote": [], + "bbox": [ + 222, + 258, + 491, + 313 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/813f71ff703751aa520a4a478429818f3458d252c130cf47ef598f6f9366c557.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 147, + 779, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4d01d043fddc5e8a8368b8dec20a40f587859395a77c4999585584e5f6b82104.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 200, + 779, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4db7cfbd623810400fd8af925e5d2dbca6d2d171ab7188b95a9d53632cc526af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 257, + 779, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As illustrated in Figure 3, it mainly consists of 558K image-caption pairs from the LLaVA-filtered CC3M dataset [49] and 232K video-caption pairs sampled from the WebVid 2.5M dataset [5]. The instruction format for each modality is presented in Figure 3. In this stage, we optimize the projectors in Figure 2, while freezing the pre-trained modules like the visual encoder and text decoder.", + "bbox": [ + 214, + 401, + 787, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Instruction Tuning. To enhance the multi-modality understanding of LLMs, we build the instruction pairs from [32] and [39]. In particular, the constructed dataset mainly involves content from three sources, i.e., 40K text conversations from ShareGPT [1], 625K single- or multi-turn visual QA pairs from [19,22,25, 26,33,41-43,48,50], and 98K video QA pairs from [7]. For the instruction, we adopt different formats for text, image, and video input, as shown in Figure 3. The input prompt and answer vary with datasets. Please refer to [32] and [39] for more details. Meanwhile, the image token is randomly inserted at the beginning or end of the user input during our training. In instruction tuning, all the modules are optimized except the visual encoder.", + "bbox": [ + 214, + 492, + 787, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Long Video Tuning. To further unleash the potential for hour-long videos, we construct 15K long QA pairs, including 9K conversions in movie scenes and 6K data sampled from LongLoRA [11] for token expanding. Specifically, we utilize more than 400 long movies and corresponding scripts in MovieNet [21] to build the training set. The key components for instruction generation are visualized in Figure 4. Generally, the generated dataset includes QA pairs from three aspects: video summary, movie plot, and detail reasoning. For video summaries, we collect movie synopses to produce brief and detailed summaries for each movie using developed LLMs like GPT-4 [45]. It brings about 1K summary-level instruction pairs in total. For plot-level data, we take the entire movie synopsis as input and leverage GPT-4 [45] to generate plot-related and character-related QA pairs. These include plot understanding, description, analysis, character relationship,", + "bbox": [ + 214, + 657, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "LLaMA-VID", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "personality, and behavior. In particular, we generate 5 plot-related pairs and 5 character-related pairs for each movie, resulting in 4K plot-level QA data. As for detail-level data, we feed the long movie script into Claude-2 [3] and generate 5 plot-related reasoning pairs and 5 detail-related descriptions for each movie, which brings 4K pairs in total. With long videos and the generated pairs, we perform instruction tuning by concatenating visual tokens and subtitle tokens for each frame, as depicted in Figure 3. In this way, LLaMA-VID can well support 64K tokens with more than 3-hour video as input.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 292, + 375, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we provide the experimental setup and comparisons with leading methods on several benchmarks.", + "bbox": [ + 212, + 325, + 785, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 380, + 428, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation Details. In this work, we instantiate the model with the pre-trained EVA-G [16] for visual encoder and QFormer [13] for text decoder by default. During training, we keep the visual encoder fixed in all stages and freeze the text decoder, as well as the LLM, in the modality alignment stage, except for the BERT module in Table 8 that is not pre-trained. Following the strategy in [32], we optimize trainable parameters with the designed data and instructions in Figure 3, running for 1 epoch in each stage. For video input, we extract frames at 1 FPS. All models are trained using $8 \\times$ NVIDIA A100 GPUs.", + "bbox": [ + 212, + 407, + 787, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. In this study, we construct the training set mainly from [5,21,32,39], as illustrated in Section 3.3. Moreover, we report results on several video- and image-based benchmarks. In particular, for video input, we evaluate the zero-shot performance on the open-ended QA benchmarks like MSVD [8], MSRVTT [59], ActivityNet [7], EgoSchema [40], SSv2 (Something-Something v2) [18], and the newly-proposed generative performance benchmark [39]. As for image-based evaluation, we conduct experiments on several widely-adopted benchmarks, including GQA [22], MMB (MMBench) [36], MME [17], POPE [31], SEED [28], $\\mathrm{SQA}^{\\mathrm{I}}$ [37], $\\mathrm{VQAT}^{\\mathrm{T}}$ (TextVQA) [51], VizWiz [20], and $\\mathrm{VQAv}^{\\mathrm{v2}}$ (VQA V2) [19].", + "bbox": [ + 212, + 546, + 787, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 215, + 705, + 372, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on Video-based Benchmarks. In Table 1, we provide a comparative evaluation of LLaMA-VID against various state-of-the-art methods across zero-shot video QA benchmarks: MSVD-QA [8], MSRVTT-QA [59], ActivityNet-QA [7], EgoSchema [40], and SSv2 [18]. Notably, the results are reported with only two tokens for each frame. It is evident that LLaMA-VID, employing Vicuna-7B and Vicuna-13B as the LLMs, consistently delivers superior performance across all datasets. On the MSVD-QA and MSRVTT-QA datasets, it", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "L. Yanwei et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bbc70c2198b547a88367ceb7ab6089b92a7bdc2d6bb12c35d6bafa5cd837d5ce.jpg", + "table_caption": [ + "Table 1: Comparison with leading methods on 3 zero-shot video QA datasets. We report results with 2 tokens for each frame. For fair comparisons, our model is trained without long video tuning (stage 3) in Figure 3. Res indicates image resolution." + ], + "table_footnote": [], + "table_body": "
MethodLLMRes.MSVD-QAMSRVTT-QAActivityNet-QA
AccScoreAccScoreAccScore
VideoLLaMA [64]Vicuna-7B22451.62.529.61.812.41.1
LLaMA-Adapter [65]LLaMA-7B22454.93.143.82.734.22.7
VideoChat [30]Vicuna-7B22456.32.845.02.526.52.2
Video-ChatGPT [39]Vicuna-7B22464.93.349.32.835.22.7
BT-Adapter [34]Vicuna-7B-67.53.757.03.245.73.2
LLaMA-VIDVicuna-7B22469.73.757.73.247.43.3
LLaMA-VIDVicuna-13B22470.03.758.93.347.53.3
", + "bbox": [ + 228, + 198, + 774, + 319 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/c9c9a7f00be14f8e5b0431d694a4ef6cb05a05dfd83e96e643573b794ef7206f.jpg", + "table_caption": [ + "Table 2: Comparison with leading methods on the video generative benchmark [39]. We report results with 2 tokens for each frame. For fair comparisons, our model is trained without long video tuning (stage 3) in Figure 3. Res indicates image resolution. Correct, Detail, Context, Temporal, and Consist indicate evaluation metrics in [39]." + ], + "table_footnote": [], + "table_body": "
MethodLLMRes.Correct.DetailContextTemporalConsist.
VideoLLaMA [64]Vicuna-7B2241.962.182.161.821.79
LLaMA-Adapter [65]LLaMA-7B2242.032.322.301.982.15
VideoChat [30]Vicuna-7B2242.232.502.531.942.24
Video-ChatGPT [39]Vicuna-7B2242.402.522.621.982.37
BT-Adapter [34]Vicuna-7B-2.682.693.272.342.46
LLaMA-VIDVicuna-7B2242.963.003.532.462.51
LLaMA-VIDVicuna-13B2243.073.053.602.582.63
", + "bbox": [ + 220, + 400, + 782, + 507 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/63acb632e1784d863f70b8c4485afcec79a535e947bb673de9b84d54801b89d4.jpg", + "table_caption": [ + "Table 3: Comparison with leading methods on the zero-shot benchmark EgoSchema [40]. We report results with 2 tokens for each frame on different settings." + ], + "table_footnote": [], + "table_body": "
FrameFrozenBiLM [60]mPLUG-Owl [62]InternVideo [56]LLaMA-VID
1026.4%29.6%31.4%41.2%
30-20.0%31.8%41.7%
9026.9%-32.1%40.5%
180---41.7%
", + "bbox": [ + 241, + 561, + 754, + 645 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "achieves the accuracy of $69.7\\%$ and $57.7\\%$ with Vicuna-7B, surpassing the previous leading approach [34] with absolute gains of $2.2\\%$ and $0.7\\%$ , respectively. For the ActivityNet-QA dataset, LLaMA-VID attains top performance in accuracy and the highest score of 3.3. In Table 2, we also carry out experiments on the newly proposed video-based generative performance benchmark [39]. As for the long EgoSchema and SSv2 dataset, the proposed method surpasses other leading approaches in different frame settings, as presented in Table 3 and Table 4. Specifically, we evaluate the model on the SSv2 dataset similar with that in MSRVTT-QA. Our LLaMA-VID is validated to achieve the best performance across all the evaluation metrics, surpassing previous approaches by a large margin. In general, LLaMA-VID is proved to bring robust results on all benchmarks.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "LLaMA-VID", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/b05e475af42b7dbbba3a374f8fcc6b17050ed770fc001b60f8f059cc3a09c641.jpg", + "table_caption": [ + "Table 4: Comparisons with zero-shot methods on SSv2 [18] dataset." + ], + "table_footnote": [], + "table_body": "
MethodLLMRes.AccScore
Video-ChatGPT [38]Vicuna-7B22411.592.10
LLaMA-VIDVicuna-7B22420.122.14
", + "bbox": [ + 328, + 170, + 671, + 226 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/b35063c6794095e7b019e6b045e62d610ecf382e4b9c4bfed9c918fa5c8d5573.jpg", + "table_caption": [ + "Table 5: Comparison with leading methods on 8 benchmarks. Here, we use the same training and instruction finetuning data as that in LLaVA-1.5. We report results with 1 context token and $n$ content tokens, where $n$ is kept the same with that in LLaVA-1.5, i.e., $n = (336 / 14)^2 = 576$ . Res indicates input image resolution. * and † denote the train subset is included for training and the data is not publicly available, respectively." + ], + "table_footnote": [], + "table_body": "
MethodLLMRes.GQAMMBMMEPOPESEED\\( \\mathbf{SQA}^{\\dagger} \\)VizWiz\\( VQA^{v2} \\)
InstructBLIP [13]Vicuna-7B22449.236.0--53.460.534.5-
IDEFICS-9B [23]LLaMA-7B22438.448.2----35.550.9
Qwen-VL† [4]Qwen-7B44859.3*38.2--56.367.135.278.8*
Qwen-VL-Chat† [4]Qwen-7B44857.5*60.61487.5-58.268.238.978.2*
LLaVA-1.5 [32]Vicuna-7B33662.0*64.31510.785.958.666.850.078.5*
LLaMA-VIDVicuna-7B33664.3*65.11521.486.059.968.354.279.3*
BLIP-2 [29]Vicuna-13B22441.0-1293.885.346.461.019.641.0
InstructBLIP [13]Vicuna-13B22449.5-1212.878.9-63.133.4-
Shikra [9]Vicuna-13B224-58.8-----77.4*
IDEFICS-80B [23]LLaMA-65B22445.254.5----36.060.0
LLaVA-1.5 [32]Vicuna-13B33663.3*67.71531.385.961.671.653.680.0*
LLaMA-VIDVicuna-13B33665.0*66.61542.386.062.370.054.380.0*
", + "bbox": [ + 220, + 319, + 782, + 483 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results on Image-based Benchmarks. As illustrated in Section 3.2, LLaMA-VID not only efficiently represents long video, but also expands the upper limit of VLMs with an additional context token. In Table 5, we perform comparisons with leading VLMs on 8 widely-adopted benchmarks. It should be noted that we maintain the same training data and image resolution as in LLaVA-1.5 [32] for fair comparisons. It is evident that LLaMA-VID outperforms other leading methods across most of benchmarks with different LLMs. In particular, with Vicuna-7B, LLaMA-VID attains the best results across all the datasets and surpasses LLaVA-1.5 with significant gains in GQA, MME, and VizWiz, where the improvement reaches up to $2.3\\%$ , 10.7, and $4.2\\%$ , respectively. With a larger Vicuna-13B, LLaMA-VID also outperforms other methods in 6 benchmarks and achieves top-2 in the other datasets. This demonstrates the generality of LLaMA-VID, which can be scaled up with a stronger foundation model.", + "bbox": [ + 212, + 508, + 787, + 705 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Qualitative Results. In Figure 5, we apply LLaMA-VID to different types of data, including short videos and long movies. We represent each frame with 2 tokens for videos. LLaMA-VID demonstrates various capabilities with different inputs. Given a short trailer video, LLaMA-VID summarizes the overall plot and infers the movie name, producer, and the type of movie. It shows that the proposed LLaMA-VID can well recognize the typical characters with the produced tokens. As for a 3-hour movie, the proposed model adeptly describes the storyline and demonstrates plot-related reasoning and detailed understanding.", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "L. Yanwei et al.", + "bbox": [ + 271, + 114, + 377, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/72a665b4f944cc1c8215ff56a24bf462941473040496baa5701c5c2718a46530.jpg", + "image_caption": [ + "User: Please describe this video in detail." + ], + "image_footnote": [], + "bbox": [ + 228, + 143, + 480, + 220 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8ee3bb725204f6e157a720e757acc358757d42f42151a99240501c19c0c3461b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 455, + 227, + 477, + 243 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fd7f68283ef81c51cf52c956375506abddf2e4a5c95f9e658b813886db30e4ff.jpg", + "image_caption": [ + "3-hour VideoAvatar" + ], + "image_footnote": [], + "bbox": [ + 557, + 178, + 581, + 193 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/734c9af289904b0876a4ea61cfea6e1ded8258f3424ea95681f14d03d6a2c762.jpg", + "image_caption": [ + "User: Summarize the plot of this movie." + ], + "image_footnote": [], + "bbox": [ + 607, + 146, + 674, + 220 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/bdec7b6c096e13dffb02b5a47ab1d8e1bd53886907c113822141ae1c715de62c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 146, + 767, + 220 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/77b1b3cf4b8ffccce0291d0c13432ce2f4503730d2118dff6cd9adc708585953.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 229, + 766, + 243 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d81eb905ae1ee06d782c52e2f95e2ed44489a292b3915b8bc612ad94bd36283f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 253, + 263, + 279 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Assistant: The video starts with a carto on character standing in front of a castle, which is surrounded by a beautiful lan descape. The character is holding a bucket and appears to be observing the surroundings. The scene then transitions to a t train, where the character is seen sitting inside the train. The train is surrounded by lush greenery and appears to be mov ing through a scenic route....", + "bbox": [ + 269, + 257, + 473, + 354 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e050285158efd78ed3d3d8396ffab588de67a13fa8fe849fcfcb7d0406126027.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 253, + 553, + 280 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Assistant:Avatar follows Jake Sully, a disabled former Marine who is recruited by the RDA to work on Pandora, a moon rich in a valuable mineral called unobtanium. Jake's twin brother, Tommy, died in a helmet-cam accident, and Jake is offered his place in the Avatar Program, which allows humans to link with the native Na' vi, a humanoid species. Jake is assigned to the avatar of his deceased brother....", + "bbox": [ + 558, + 257, + 761, + 354 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "User: What's the name of the movie described in this video?", + "bbox": [ + 238, + 368, + 444, + 388 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8fb7ed9008fb0fc006e34c39774d688179bad2dbfaa53a23949bf75d2f0f1065.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 367, + 475, + 383 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f6d9d45df6c90a54aedd838c34e4ac056cd2e79da034450845f24c0a1c459aca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 745, + 364, + 764, + 380 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9add4a71043559c97faef4fd0b846eab03a20bdbc35205b8ba24eb418d99a528.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 238, + 400, + 264, + 425 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Assistant: The movie described in this video is \"Zootopia.\"", + "bbox": [ + 271, + 402, + 470, + 424 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ecfc7d758e0961100b562c8c9cfe8d12ca884d4463bbdfc8f7928d4e7212910b.jpg", + "image_caption": [ + "User: How human work on Pandora?" + ], + "image_footnote": [], + "bbox": [ + 527, + 388, + 553, + 414 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Assistant: Humans work on Pandora by using avatars, which are grown from human DNA mixed with the DNA of the native Na'vi, to interact with the environment and the Na'vi.", + "bbox": [ + 560, + 392, + 759, + 441 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "User: Who produce the movie?", + "bbox": [ + 238, + 438, + 382, + 449 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8ae39fbd05ec4df51649fca78a50950ff4cfe12dcb215ec2440482673f4cc56a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 435, + 475, + 450 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "User: Why Jake Sully decide to fight againt humanity with the Na'vi?", + "bbox": [ + 529, + 454, + 736, + 476 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ef43d0dd38c7d488cc75d3a24093f39d1d94417afbcdf9f7baa4eef1a2fb0be7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 453, + 766, + 468 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b10e02a5fc1e5b2c2b28de1f37525f77a539f2dabbeaf4b9f5c76c9a89eeee6f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 460, + 263, + 484 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Assistant: The movie is produced by Walt Disney Pictures.", + "bbox": [ + 271, + 464, + 468, + 484 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b40b3eba5c1f3b56ff1d4399214042fe8fc836c591d3176d58ccdcbc41a1197a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 484, + 553, + 510 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Assistant: Jake Sully decides to fight a gainst humanity with the Na'vi after he falls in love with the Na'vi and their way of life, and after he realizes that the human military is destroying their home and culture.", + "bbox": [ + 558, + 488, + 759, + 547 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "User: What's the type of this movie?", + "bbox": [ + 238, + 500, + 424, + 511 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/287e9de54548bfde8d7bae7dff2287c3eb4372defa68ef5b5d0a5ae21d59c3dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 455, + 497, + 475, + 513 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/429f15612f6aee68abd959212c1f33a3cdcbcfde3fe754d04f0f37ca20a2ceed.jpg", + "image_caption": [ + "Fig. 5: Examples of LLaMA-VID with short video and hour-long video." + ], + "image_footnote": [], + "bbox": [ + 236, + 521, + 263, + 547 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Assistant: The movie is a Disney animat ed film.", + "bbox": [ + 271, + 526, + 462, + 546 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Component-wise Analysis", + "text_level": 1, + "bbox": [ + 215, + 609, + 473, + 626 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this subsection, we conduct ablation studies with input resolution 224 and 2 tokens for each image by default.", + "bbox": [ + 214, + 638, + 785, + 670 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Generated Token Types. As illustrated in Figure 2, each image is represented with a context token and a content token in LLMs. To validate the effectiveness of each part, we conduct experiments with different types of tokens in Table 6. Without the context token, the compressed content token, which encodes each image with 1 token, cannot adjust to input instructions, leading to subpar performance. Compared with a single content token, the instruction-guided context token results in significant gains across all datasets with only 1 token. With both tokens for each image, the model achieves the best performance across all image- and video-based benchmarks. It shows that both instruction cues in the context token and the image content itself in the content token are important.", + "bbox": [ + 214, + 688, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "LLaMA-VID", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/70c7b59a3c7cf7d0d0a6ad1fa43c184b287f7c3da4ac101ae08d59e80f4f8741.jpg", + "table_caption": [ + "Table 6: Comparison with different token types. We report results with 1 context token (if exists) and 1 content token on image-based and video-based benchmarks (MSVD)." + ], + "table_footnote": [], + "table_body": "
contextcontentGQAPOPE\\( \\mathbf{SQA^I} \\)\\( VQA^T \\)MSVD
✘✓53.380.966.146.565.9
✓✘54.382.467.748.369.3
✓✓55.583.168.849.069.7
", + "bbox": [ + 287, + 184, + 710, + 250 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/9093f7acdb1f419ecad104afa6078a9d0c5bd9e0d46d3c046aa1a65c8405bdc9.jpg", + "table_caption": [ + "Table 7: Comparison with different token numbers. We report results with various numbers of context token and content token on image-based benchmarks for efficiency." + ], + "table_footnote": [], + "table_body": "
contextcontentGQAPOPE\\( \\mathbf{SQA^I} \\)\\( VQA^T \\)
025661.985.567.553.0
125663.086.667.753.8
16460.885.168.752.3
11658.283.167.450.8
1456.283.568.749.1
1155.583.168.849.0
", + "bbox": [ + 321, + 305, + 676, + 414 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Generated Token Numbers. In Table 7, we conduct experiments with different numbers of tokens for further investigation. With an image size $224 \\times 224$ , we set up experiments with $n$ content tokens, where $n = (224 / 14)^{2} = 256$ for uncompressed settings in the first two rows. The results clearly show that the context token consistently improves performance across different benchmarks with only 1 extra token. When we compress the content token to $1 / 4$ with $n = 64$ , the performance drops about $1\\%$ to $2\\%$ but increases $1\\%$ in $\\mathrm{SQA}^{\\mathrm{I}}$ . Considering the extra efficient setting for hour-long videos, we compress the content token to $1 / 256$ with $n = 1$ by default. Compared to the original setting without context token, we can reduce the computational cost to $1 / 128$ with about $2\\% - 6\\%$ performance drop, which is generally acceptable. The linear increase in performance presents significant potential for token compression. For instance, we can dynamically compress the content token to different numbers according to resource budget and content importance. Interestingly, the model achieves peak performance in $\\mathrm{SQA}^{\\mathrm{I}}$ with only 2 tokens. This could be attributed to the fact that problems in ScienceQA [37] focus more on visual-based reasoning rather than image details. As demonstrated in Tables 1 and 2, with only 2 tokens for each image, LLaMA-VID still outperforms all previous work in different video-based benchmarks. This makes it feasible to enable LLMs for hour-long video processing.", + "bbox": [ + 212, + 443, + 787, + 731 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Text Decoder. As depicted in Figure 2, the text decoder plays an essential role in producing instruction-guided context cues. Here, we further perform comparisons with different text decoders in Table 8. We mainly instantiate the text decoder with two types of modules, namely BERT [14] and QFormer [13]. For BERT, we randomly initialize it as a cross-modality decoder and only retain the first two layers. As for QFormer, we utilize the pre-trained modules and fix them", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "L. Yanwei et al.", + "bbox": [ + 271, + 114, + 377, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/1df1c4843c5d5d130c5f17f2b83721b5fadac6df3e9622453e9642d8c9760f56.jpg", + "table_caption": [ + "Table 8: Comparison with different text decoders. We report results with 1 context token (if exists) and 1 content token on image-based benchmarks for efficiency." + ], + "table_footnote": [], + "table_body": "
textGQAPOPE\\( \\mathbf{SQA^I} \\)\\( \\mathbf{VQA^T} \\)
-53.380.966.146.5
BERT54.180.867.948.1
QFormer55.583.168.849.0
", + "bbox": [ + 346, + 184, + 653, + 252 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/10d0d34e19921c09cd1e7a61c1f27450189cefc50de0c76ef8a35874f02f596d.jpg", + "table_caption": [ + "Table 9: Average module latency of video frame on a A100 GPU. Encoder, Content, Decoder, and Context denote the latency for text-free visual encoder and content token, text-related text decoder and context token generation, respectively." + ], + "table_footnote": [], + "table_body": "
textText-freeText-relatedTotal (ms)
EncoderContentDecoderContext
BERT4.890.010.250.025.17
QFormer4.870.011.080.025.98
", + "bbox": [ + 292, + 320, + 705, + 388 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "for modality alignment. Even with a simple 2-layer BERT, as shown in Table 8, the generated context token achieves significant gains in most of benchmarks. This proves the effectiveness of the paradigm for context token generation. With a pre-trained text decoder like QFormer, the model can be further enhanced and attains peak performance in all datasets with $2.2\\%$ to $2.7\\%$ significant gain.", + "bbox": [ + 212, + 417, + 784, + 494 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Latency Analysis. To further analyse the module latency, we keep the original setting in Table 8 and report average module latency over 100 videos in Table 9. Here, we divide the modules for generation into two types, namely text-free and text-related. In particular, we generate visual tokens one time for each frame with text-free modules. And text-related modules are required to produce different context tokens according to user inputs. It is clear that the text-free tokens account for the most computation, while the text-related modules only cost $5\\%$ to $18.3\\%$ for each user instruction. That means the total cost will not increase greatly given variant user questions because of the efficient text-related modules.", + "bbox": [ + 212, + 515, + 787, + 652 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Response in Context Attention. To more vividly explore the context attention, we visualize the high response areas with the top 20 scores in Figure 6. Specifically, we draw the normalized heatmap for the first four queries in $\\mathbf{Q}_t$ before applying the Softmax function, which is used to formulate context token in Equation 1. As shown in Figure 6, the text-guided query $\\mathbf{Q}_t$ effectively focuses on important areas relevant to the input questions. For example, in the second row, when inquiring whether the image depicts a fishing village, the query $\\mathbf{Q}_t$ focuses more on buildings along the river and a seagull. These are all typical characteristics to distinguish a fishing village in common sense. Other examples also confirm that the designed context attention successfully achieves its goal of formulating the context token under instruction guidance.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "LLaMA-VID", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9b55caa5a4cc5c34b647ac1b059e14c923a33db2c50956c2cf8c7da0eec0fff7.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 218, + 147, + 331, + 205 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a950cf1de1a709c4d6d996327a5f19b7fb4ab06775a37a6b6c8b3d95b75105f9.jpg", + "image_caption": [ + "Query 1" + ], + "image_footnote": [], + "bbox": [ + 339, + 147, + 447, + 205 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/fc726282f4090ce1c536c1dcf5cad305881be27b371fc37c3d3c92898c39da40.jpg", + "image_caption": [ + "Query 2" + ], + "image_footnote": [], + "bbox": [ + 450, + 147, + 558, + 205 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1b25ca521e7ade249fee96d8831230fa7aec15eb136f840c08db05105a628e02.jpg", + "image_caption": [ + "Query 3" + ], + "image_footnote": [], + "bbox": [ + 560, + 147, + 669, + 205 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f8d85417b652785ef5650a64562b7076d43d452be66eb6e5451e615020cff279.jpg", + "image_caption": [ + "Query 4" + ], + "image_footnote": [], + "bbox": [ + 671, + 147, + 782, + 205 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/6499112d0933b04822419699d45b4cccf51b8b29c975d2d4467da848846ef8d4.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 220, + 244, + 331, + 309 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/fb5d34a3fe3c591ab99e4e3e0c568403a37785ecf775d331ad906b9bede29fe1.jpg", + "image_caption": [ + "Question: Are the elephants fighting?", + "Query 1" + ], + "image_footnote": [], + "bbox": [ + 339, + 244, + 447, + 309 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/032f9634bb7e2c3bf8e3e8ad8571c44938c864467212772bb8c6e7f5a0a4d9d4.jpg", + "image_caption": [ + "Query 2" + ], + "image_footnote": [], + "bbox": [ + 450, + 244, + 558, + 309 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2dc0580fd2dffdb9dcbb92165e32c82db59c0bb3e6251519d7d94e3caa5cd1a0.jpg", + "image_caption": [ + "Query 3" + ], + "image_footnote": [], + "bbox": [ + 560, + 244, + 669, + 309 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/41507cc8848fab782616d302aebb8e79163fe5ba25fc3b34e16025f4698a5a4a.jpg", + "image_caption": [ + "Query 4" + ], + "image_footnote": [], + "bbox": [ + 671, + 244, + 782, + 309 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3d0fd57ad55436340e87cfe02370b24f4ec48bde01cfc2728062f6acb966a407.jpg", + "image_caption": [ + "Input", + "Fig. 6: High response areas with top scores to input question in Equation 1. We present the response of first four queries in $\\mathbf{Q}_t$ . Images are sampled from VQA V2 test-dev set." + ], + "image_footnote": [], + "bbox": [ + 220, + 352, + 331, + 411 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9dfe53c6317b6802cf82358d866028009350c3050522da42907dc85ec0290f62.jpg", + "image_caption": [ + "Question: Is this a fishing village?", + "Query 1", + "Question: What tall animals are displayed in the background?" + ], + "image_footnote": [], + "bbox": [ + 339, + 352, + 447, + 411 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/bbc34d68c49ec350d0f17fce2a6552fea86bf83ea4cb4b13746b11af6dcd8c5a.jpg", + "image_caption": [ + "Query 2" + ], + "image_footnote": [], + "bbox": [ + 450, + 352, + 558, + 411 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c8ca05f8bef6da9eec0e0b50a3ef145e2fdc79664f5730d70d1c0c3d5ea895a7.jpg", + "image_caption": [ + "Query 3" + ], + "image_footnote": [], + "bbox": [ + 560, + 352, + 669, + 411 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/924c4718926950fe8014999ce472fbd48fde2a71e3e93bfa0dbe187ec57acbfc.jpg", + "image_caption": [ + "Query 4" + ], + "image_footnote": [], + "bbox": [ + 671, + 352, + 782, + 411 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 214, + 506, + 509, + 522 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have introduced LLaMA-VID, a simple yet effective token generation approach for VLMs. The central concept behind LLaMA-VID is to represent an image with the context token and the content token. In particular, the context token is generated according to input instructions, and the content token is produced based on the image content. Depending on the budget, the content token can be compressed to one token or expressed without compression. It allows us to represent a single image with preserved details and efficiently encode each video frame with only two tokens. Moreover, we have constructed an instruction dataset for hour-long video understanding. Our experiments on several video-and image-based benchmarks prove the superiority of our method. We hope that LLaMA-VID can serve as a strong benchmark for efficient visual representation.", + "bbox": [ + 212, + 537, + 787, + 703 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "There still exists certain limitations in the current method. Although LLaMAVID succeeds to represent hour-long videos in LLMs, it focuses more on temporal consistency and may lose spatial details in each frame. In the future work, we plan to maintain dense spatial information for each frame outside the LLM and conduct the visual token fusion within LLMs to avoid such detail loss.", + "bbox": [ + 212, + 704, + 787, + 779 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported in part by the Research Grants Council under the Areas of Excellence scheme grant AoE/E-601/22-R.", + "bbox": [ + 214, + 794, + 785, + 825 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "L. Yanwei et al.", + "bbox": [ + 271, + 114, + 377, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Sharegpt. https://sharegpt.com/ (2023) 7", + "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: NeurIPS (2022) 4", + "3. Anthropic: Claude 2. https://www.anthropic.com/index/claude-2 (2023) 7, 8", + "4. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv:2308.12966 (2023) 10", + "5. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: ICCV (2021) 7, 8", + "6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS (2020) 3", + "7. Caba Heilbron, F., Escorcia, V., Ghanem, B., Carlos Niebles, J.: Activitynet: A large-scale video benchmark for human activity understanding. In: CVPR (2015) 7, 8", + "8. Chen, D., Dolan, W.B.: Collecting highly parallel data for paraphrase evaluation. In: ACL (2011) 8", + "9. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv:2306.15195 (2023) 10", + "10. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollár, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv:1504.00325 (2015) 3", + "11. Chen, Y., Qian, S., Tang, H., Lai, X., Liu, Z., Han, S., Jia, J.: Longlora: Efficient fine-tuning of long-context large language models. arXiv:2309.12307 (2023) 7", + "12. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ chatgpt quality. https://lmsys.org/blog/2023-03-30-vicuna/ (2023) 3, 6", + "13. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv:2305.06500 (2023) 1, 2, 4, 5, 6, 8, 10, 12", + "14. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv:1810.04805 (2018) 2, 3, 5, 12", + "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) 2, 5", + "16. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. In: CVPR (2023) 5, 8", + "17. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv:2306.13394 (2023) 8", + "18. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The\" something something\" video database for learning and evaluating visual common sense. In: ICCV (2017) 8, 10" + ], + "bbox": [ + 218, + 176, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "LLaMA-VID", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: CVPR (2017) 3, 7, 8", + "20. Gurari, D., Li, Q., Stangl, A.J., Guo, A., Lin, C., Grauman, K., Luo, J., Bigham, J.P.: Vizwiz grand challenge: Answering visual questions from blind people. In: CVPR (2018) 8", + "21. Huang, Q., Xiong, Y., Rao, A., Wang, J., Lin, D.: Movienet: A holistic dataset for movie understanding. In: ECCV (2020) 7, 8", + "22. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: CVPR (2019) 7, 8", + "23. IDEFICS: Introducing idefics: An open reproduction of state-of-the-art visual language model. https://huggingface.co/blog/idefics (2023) 10", + "24. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021) 3", + "25. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: Referitgame: Referring to objects in photographs of natural scenes. In: EMNLP (2014) 7", + "26. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV (2017) 7", + "27. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. arXiv:2308.00692 (2023) 3", + "28. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv:2307.16125 (2023) 8", + "29. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv:2301.12597 (2023) 1, 2, 4, 10", + "30. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chat-centric video understanding. arXiv:2305.06355 (2023) 1, 4, 9", + "31. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv:2305.10355 (2023) 8", + "32. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv:2310.03744 (2023) 6, 7, 8, 10", + "33. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeruIPS (2023) 1, 2, 4, 6, 7", + "34. Liu, R., Li, C., Ge, Y., Shan, Y., Li, T.H., Li, G.: One for all: Video conversation is feasible without video instruction tuning. arXiv:2309.15785 (2023) 9", + "35. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. arXiv:1907.11692 (2019) 3", + "36. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv:2307.06281 (2023) 8", + "37. Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. In: NeurIPS (2022) 3, 8, 12", + "38. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability. arXiv:2306.07207 (2023) 1" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "L. Yanwei et al.", + "bbox": [ + 271, + 114, + 377, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Maaz, M., Rasheed, H., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv:2306.05424 (2023) 1, 4, 7, 8, 9", + "40. Mangalam, K., Akshulakov, R., Malik, J.: Egoschema: A diagnostic benchmark for very long-form video language understanding. NeurIPS (2024) 8, 9", + "41. Mao, J., Huang, J., Toshev, A., Camburu, O., Yuille, A.L., Murphy, K.: Generation and comprehension of unambiguous object descriptions. In: CVPR (2016) 7", + "42. Mao, J., Huang, J., Toshev, A., Camburu, O., Yuille, A.L., Murphy, K.: Generation and comprehension of unambiguous object descriptions. In: CVPR (2016) 7", + "43. Mishra, A., Shekhar, S., Singh, A.K., Chakraborty, A.: Ocr-vqa: Visual question answering by reading text in images. In: ICDAR (2019) 7", + "44. OpenAI: Chatgpt. https://openai.com/blog/chatgpt/ (2023) 1, 3", + "45. OpenAI: Gpt-4 technical report. arXiv:2303.08774 (2023) 1, 3, 7", + "46. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. In: NeurIPS (2022) 3", + "47. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 3, 5", + "48. Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: A-okvqa: A benchmark for visual question answering using world knowledge. In: ECCV (2022) 7", + "49. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: ACL (2018) 7", + "50. Sidorov, O., Hu, R., Rohrbach, M., Singh, A.: Textcaps: a dataset for image captioning with reading comprehension. In: ECCV (2020) 7", + "51. Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards vqa models that can read. In: CVPR (2019) 8", + "52. Song, E., Chai, W., Wang, G., Zhang, Y., Zhou, H., Wu, F., Chi, H., Guo, X., Ye, T., Zhang, Y., et al.: Moviechat: From dense token to sparse memory for long video understanding. In: CVPR (2024) 1, 4", + "53. Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca (2023) 3, 6", + "54. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models. arXiv:2302.13971 (2023) 1, 3, 6", + "55. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 3", + "56. Wang, Y., Li, K., Li, Y., He, Y., Huang, B., Zhao, Z., Zhang, H., Xu, J., Liu, Y., Wang, Z., et al.: Intervideo: General video foundation models via generative and discriminative learning. arXiv:2212.03191 (2022) 9", + "57. Wei, J., Bosma, M., Zhao, V.Y., Guu, K., Yu, A.W., Lester, B., Du, N., Dai, A.M., Le, Q.V.: Finetuned language models are zero-shot learners. arXiv:2109.01652 (2021) 3", + "58. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv:2303.04671 (2023) 3", + "59. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: CVPR (2016) 3, 8" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "LLaMA-VID", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "60. Yang, A., Miech, A., Sivic, J., Laptev, I., Schmid, C.: Zero-shot video question answering via frozen bidirectional language models. In: NeurIPS (2022) 9", + "61. Yang, R., Song, L., Li, Y., Zhao, S., Ge, Y., Li, X., Shan, Y.: Gpt4tools: Teaching large language model to use tools via self-instruction. arXiv:2305.18752 (2023) 3", + "62. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv:2304.14178 (2023) 9", + "63. Zhang, C., Lu, T., Islam, M.M., Wang, Z., Yu, S., Bansal, M., Bertasius, G.: A simple lvm framework for long-range video question-answering. arXiv:2312.17235 (2023) 4", + "64. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv:2306.02858 (2023) 1, 4, 9", + "65. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv:2303.16199 (2023) 9", + "66. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv:2205.01068 (2022) 1, 3", + "67. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv:2304.10592 (2023) 4" + ], + "bbox": [ + 215, + 146, + 784, + 436 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "L. Yanwei et al.", + "bbox": [ + 271, + 114, + 377, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/LLaMA-VID_ An Image is Worth 2 Tokens in Large Language Models/022c7268-b560-44a0-867d-3ec58c86bb1a_model.json b/2024/LLaMA-VID_ An Image is Worth 2 Tokens in Large Language Models/022c7268-b560-44a0-867d-3ec58c86bb1a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..59974bd7a607befa2802732b46469a2e5c5cd716 --- /dev/null +++ b/2024/LLaMA-VID_ An Image is Worth 2 Tokens in Large Language Models/022c7268-b560-44a0-867d-3ec58c86bb1a_model.json @@ -0,0 +1,3019 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.258, + 0.141, + 0.747, + 0.187 + ], + "angle": 0, + "content": "LLaMA-VID: An Image is Worth 2 Tokens in Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.212, + 0.674, + 0.228 + ], + "angle": 0, + "content": "Yanwei Li\\(^{1\\star}\\) Chengyao Wang\\(^{1*}\\) Jiaya Jia\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.425, + 0.239, + 0.578, + 0.254 + ], + "angle": 0, + "content": "CUHK\\(^{1}\\) SmartMore\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.293, + 0.744, + 0.502 + ], + "angle": 0, + "content": "Abstract. In this work, we present a novel method to tackle the token generation challenge in Vision Language Models (VLMs) for video and image understanding, called LLaMA-VID. Current VLMs, while proficient in tasks like image captioning and visual question answering, face computational burdens when processing long videos due to the excessive visual tokens. LLaMA-VID addresses this issue by representing each frame with two distinct tokens, namely context token and content token. The context token encodes the overall image context based on user input, whereas the content token encapsulates visual cues in each frame. This dual-token strategy significantly reduces the overload of long videos while preserving critical information. Generally, LLaMA-VID empowers existing frameworks to support hour-long videos and pushes their upper limit with an extra context token. It is demonstrated to surpass previous methods on most of video- or image-based benchmarks. Code and models are available at https://github.com/dvlab-research/LLaMA-VID." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.527, + 0.377, + 0.543 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.559, + 0.788, + 0.695 + ], + "angle": 0, + "content": "Large Language Models (LLMs) [44, 54, 66], through their capacity to generate contextually accurate responses, have significantly advanced the field of AI. Drawing from the strengths of LLMs, Vision Language Models (VLMs) [13,33,45] have been developed to extend these capabilities to visual data, demonstrating their adeptness in tasks like image captioning and visual question answering. However, a substantial challenge emerges in the context of long video, where an excessive number of tokens are required to represent consecutive frames. The computational demands escalate with the video length, thereby constraining the practical application of VLMs for extensive videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.696, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Recently, several approaches have been proposed to handle videos, moving beyond image-only VLMs. These methods aim to alleviate the token issue by utilizing representative queries [30, 64] or applying temporal compression [38, 39, 52]. Despite these efforts, the challenge of long videos remains unresolved. The primary obstacle stems from the excessive number of tokens required for each video frame. For instance, models like BLIP [13, 29] and LLaVA [33] require 32 and over 256 tokens respectively for a single image. A video containing 10K frames would thus necessitate over 320K tokens, exceeding the capacity of current VLMs." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.826, + 0.359, + 0.84 + ], + "angle": 0, + "content": "* equal contribution" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.38, + 0.128 + ], + "angle": 0, + "content": "L. Yanwei et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.499, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.326, + 0.456, + 0.337 + ], + "angle": 0, + "content": "(a) Video-based benchmarks" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.143, + 0.761, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.325, + 0.729, + 0.337 + ], + "angle": 0, + "content": "(b) Image-based benchmarks" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.349, + 0.788, + 0.391 + ], + "angle": 0, + "content": "Fig. 1: The proposed LLaMA-VID achieves leading performance on most of benchmarks with 7B LLMs. The video-based and image-based benchmarks are noted in blue and purple color, respectively. Please refer to Tables 1, 2, and 5 for more details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.785, + 0.48 + ], + "angle": 0, + "content": "Furthermore, simple temporal compression can significantly damage the representation over long-term intervals. This drawback hampers their performance, thereby underscoring the need for a robust solution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.637 + ], + "angle": 0, + "content": "In this work, we present LLaMA-VID, a novel approach to effectively manage the token generation issue in long videos. Our core idea is to represent each video frame with two distinct tokens: context token and content token. The context token is designed to encode the overall context of the image based on user input, which efficiently condenses the broader picture into a single token. Simultaneously, the content token captures finer aspects of each frame. According to computational constraints, the length of content token can be extended to include more details, e.g., 1 token/frame for video input and beyond 256 token/frame for single image. In this way, the overload of long videos can be significantly reduced without sacrificing critical information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In particular, our method employs a dual-token generation strategy that is both efficient and effective. For each frame, we first extract image features using a pre-trained vision transformer [15], akin to other VLMs [13, 33]. The key question is how to generate the context-related token according to user instructions. We provide the solution by leveraging the cross-modality design [14,29] for instruction-guided queries, which carry the interactive intention from users. For context token, these queries interact with previously generated image features in the designed attention module, termed as context attention. To generate content token, the image features are average pooled to formulate tokens that adapt to different settings. For instance, global pooling is adopted to maintain efficiency for video input while details are preserved with more tokens for single image input. The context and content tokens are subsequently projected to the space of LLMs with simple linear layers for final prediction. Furthermore, to better sup" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "LLaMA-VID" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "port hour-long videos in VLMs, we construct an instruction dataset that contains 9K movie-level conversations for plot reasoning and detail understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.252 + ], + "angle": 0, + "content": "Generally, LLaMA-VID can be distinguished from two aspects. On one hand, with the dual-token paradigm, each frame can be efficiently encoded with only two tokens, which empowers existing LLMs to support long videos. On the other hand, the context token aggregates the most informative feature of each image, which further extends the upper limit of VLMs with an extra token." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.787, + 0.343 + ], + "angle": 0, + "content": "The overall framework, dubbed LLaMA-VID, can be easily instantiated with various decoders and LLMs, as elaborated in Section 3. Extensive empirical studies are conducted in Section 4 to reveal the effectiveness of each component. Remarkably, our model can complete training within 2 days on a single machine with \\(8 \\times \\mathrm{A}100\\) GPUs, and it outperforms previous leading methods on most of video- and image-based benchmarks, as shown in Figure 1." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.366, + 0.388, + 0.383 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.398, + 0.784, + 0.429 + ], + "angle": 0, + "content": "In this section, we first review large language models and delve into recent advances in vision language models." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.451, + 0.46, + 0.466 + ], + "angle": 0, + "content": "2.1 Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.787, + 0.718 + ], + "angle": 0, + "content": "The field of Natural Language Processing (NLP) has witnessed tremendous advancements with the evolution of LLMs. Transformer [55] marked a pivotal milestone, with subsequent language models [14, 35, 66] demonstrating remarkable capabilities. GPT [6] revolutionized this field by utilizing generative pre-trained transformers for auto-regressive prediction, which is proved to be a potent language modeling paradigm. Recent groundbreaking works, such as ChatGPT [44], GPT-4 [45], and LLaMA [54], have pushed the boundaries even further. Trained on vast amounts of text data, these models exhibit exceptional capabilities in complex linguistic tasks. To leverage the potential of pre-trained LLMs, instruction tuning [46, 57] is a crucial component for high-quality output. This strategy is widely adopted in open-source models like Alpaca [53] and Vicuna [12], which improve over LLaMA [54] using specially designed instruction pairs. There are also researches [58, 61] that utilize the reasoning ability of LLMs and invoke pre-defined tools for visual applications. Different from them, we collect multimodality instruction data that contains text, images, and videos in this work, which is employed to empower LLMs for long video processing." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.74, + 0.465, + 0.755 + ], + "angle": 0, + "content": "2.2 Vision Language Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The advancements in computer vision and NLP have led to the emergence of vision-language models (VLMs) that integrate vision models with language models for cross-modality understanding [10,59] and reasoning [19,27,37]. Pioneering large-scale VLMs like CLIP [47] and ALIGN [24] have extended language models to vision-language tasks. The recent progress has seen an increasing focus" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.378, + 0.127 + ], + "angle": 0, + "content": "L. Yanwei et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.146, + 0.784, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.289, + 0.789, + 0.415 + ], + "angle": 0, + "content": "Fig. 2: The framework of LLaMA-VID. With user directive, it operates by taking either a single image or video frames as input, and generates responses from LLM. The process initiates with a visual encoder that transforms input frames into the visual embedding. Then, the text decoder produces text queries based on the user input. In context attention, the text query aggregates text-related visual cues. For efficiency, an option is provided to downsample the visual embedding to various token sizes, or even to a single token. The text-guided context token and the visually-enriched content token are then formulated using a linear projector to represent each frame at time \\( t \\). Finally, the LLM takes the user directive and all visual tokens as input and gives responses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.789, + 0.691 + ], + "angle": 0, + "content": "on leveraging the power of LLMs. Notably, Flamingo [2] and BLIP-2 [29] utilize web-scale image-text pairs for cross-modality alignment, thereby enhancing learning performance. To further exploit the potential of such pre-trained models, InstructBLIP [13] and MiniGPT-4 [67] construct high-quality instruction pairs based on BLIP-2 and achieve superior results. Simultaneously, LLaVA [33] employs a simple linear projector with a few learnable parameters to align the image and text space of LLaMA. Given the tailored instruction data, this straightforward approach demonstrates strong capabilities. To support video understanding in LLMs, several studies [30,52,64] attempt to utilize BLIP-2 for video embedding or text-only caption [63] extraction, while Video-ChatGPT [39] proposes spatial and temporal pooling for video features. However, given the substantial number of tokens required for each frame, LLMs encounter significant challenges when processing extensive video sequences. It prevents previous work from representing long video sequences that exceed a duration of one hour in LLMs. To solve the issue, we propose to efficiently encode each frame with only 2 tokens, which supports long video understanding in existing LLMs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.723, + 0.378, + 0.739 + ], + "angle": 0, + "content": "3 LLaMA-VID" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The framework of LLaMA-VID is conceptually simple: encoder and decoder are adopted to produce visual embedding and text-guided features, respectively; context token and content token are transformed with the tailored token generation strategy; instruction tuning is designed to unleash the potential of LLMs for image and video." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "LLaMA-VID" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.442, + 0.161 + ], + "angle": 0, + "content": "3.1 Encoder and Decoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.374 + ], + "angle": 0, + "content": "The proposed LLaMA-VID can be utilized to interact with single image or long videos. For clarity, we assume the input image is captured from a video sequence, as presented in Figure 2. Given a video frame \\(\\mathbf{V}_t \\in \\mathbb{R}^{H \\times W \\times 3}\\) at time \\(t\\), a transformer-based visual encoder is first employed to produce the visual embedding \\(\\mathbf{X}_t \\in \\mathbb{R}^{N \\times C}\\). Here, \\(N = H / p \\times W / p\\) and \\(C\\) indicate the number of image patches and embedding channels, respectively. The patch size \\(p\\) is typically set to 14 for ViT-based backbones [15, 16, 47]. Meanwhile, we take the user instruction as input and generate the text-guided query \\(\\mathbf{Q}_t \\in \\mathbb{R}^{M \\times C}\\) with the produced \\(\\mathbf{X}_t\\), where \\(M\\) denotes the number of queries. As depicted in Figure 2, this cross-modality interaction predominantly occurs in the text decoder, which can be easily instantiated with BERT [14] or QFormer [13], as compared in Table 8. In this way, the text query \\(\\mathbf{Q}_t\\) contains highlighted visual cues that are most related to the user instruction." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.402, + 0.411, + 0.416 + ], + "angle": 0, + "content": "3.2 Token Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.787, + 0.509 + ], + "angle": 0, + "content": "With the text query \\(\\mathbf{Q}_t\\) and visual embedding \\(\\mathbf{X}_t\\), we can easily generate representative tokens for LLMs. Specifically, context attention is designed to aggregate text-related visual features and condense them to a single context token. As shown in Figure 2, it takes \\(\\mathbf{Q}_t\\) and \\(\\mathbf{X}_t\\) as input and formulates the context-related embedding \\(\\mathbf{E}_t \\in \\mathbb{R}^{1 \\times C}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.343, + 0.526, + 0.786, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} _ {t} = \\operatorname {M e a n} \\left(\\operatorname {S o f t m a x} \\left(\\left(\\mathbf {Q} _ {t} \\times \\mathbf {X} _ {t} ^ {\\top}\\right) / \\sqrt {C}\\right) \\times \\mathbf {X} _ {t}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.842 + ], + "angle": 0, + "content": "where the Softmax function and Mean operation are conducted along the \\(N\\) and \\(M\\) dimensions, respectively. Unlike QFormer [13] that adopts 32 visual queries as LLMs tokens, we only utilize the text query \\(\\mathbf{Q}_t\\) to aggregate the visual features with high-response scores to input instructions. As a result, the most crucial visual cues related to user input are efficiently preserved in the condensed embedding \\(\\mathbf{E}_t\\). The effectiveness of this context-related token generation is demonstrated in Table 6 and Figure 6. Subsequently, a linear projector is utilized to transform the embedding \\(\\mathbf{E}_t\\) into the context token \\(\\mathbf{E}_t^T \\in \\mathbb{R}^{1 \\times C}\\), which aligns with the language space of LLMs. Meanwhile, we employ an adaptive pooling strategy for the visual embedding according to computational constraints to produce the content token \\(\\mathbf{E}_t^V \\in \\mathbb{R}^{n \\times C}\\), where \\(n \\in [1, N]\\). For instance, we maintain the original resolution of visual embedding \\(\\mathbf{X}_t\\) when input single image, while we downsample \\(\\mathbf{X}_t\\) to 1 token for long videos. This approach significantly reduces the overload of LLMs for each frame, thereby supporting hour-long videos effectively. Finally, the generated context token \\(\\mathbf{E}_t^T\\) and the content token \\(\\mathbf{E}_t^V\\) are concatenated to represent the frame at time \\(t\\). Along with frames at other timestamps, the entire video sequence is translated into the language space in token format, which is then used to generate responses from LLMs. The whole process is summarized in Algorithm 1." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.379, + 0.128 + ], + "angle": 0, + "content": "L. Yanwei et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.147, + 0.582, + 0.162 + ], + "angle": 0, + "content": "Algorithm 1 Pseudo Code for Token Generation." + }, + { + "type": "code", + "bbox": [ + 0.216, + 0.172, + 0.761, + 0.4 + ], + "angle": 0, + "content": "B: batch size; C: channel size; n: content shape\n# M: query length; N: shape of flatten image pacthes;\n# text_q: text query in shape (B, M, C)\n# vis_embedding: visual embedding in shape (B, N, C)\n# Key part 1: calculate context-related embedding\nctx_embedding = text_q @ vis_embedding.transpose(-1,-2)\nctx_embedding = ctx_embedding / (vis_embedding.shape[-1]**0.5)\nctx_embedding = (ctx_embedding-largest(-1)@vis_embedding).mean(1)\nctx_embedding = self.batchproj(ctx_embedding[:,None])\n# Key part 2: calculate visual embedding\ncur_shape = int(vis_embedding.shape[1]**0.5)\nvis_embedding = vis_embedding.reshape(B, cur_shape, -1, C)\nvis_embedding = F AVG_pool2d(vis_embedding.permute(0,3,1,2), kernel_size=cur_shape//n,\nstride=cur_shape//n)\nvis_embedding = vis_embedding.permute(0,2,3,1).flatten(1,2)\nvis_embedding = self.visproj(vis_embedding)\n# concat token in shape (B, n+1, C), n in [1,N]\nfinal_token = torch.cat([ctx_embedding, vis_embedding], dim=1)" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.455, + 0.409, + 0.47 + ], + "angle": 0, + "content": "3.3 Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.48, + 0.789, + 0.542 + ], + "angle": 0, + "content": "Training strategy, particularly instruction tuning, has proven to be crucial in LLMs [12, 53, 54] and VLMs [13, 32, 33]. Considering training efficiency, in this work, we divide the training procedure into three stages, i.e., modality alignment, instruction tuning, and long video tuning." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.556, + 0.789, + 0.618 + ], + "angle": 0, + "content": "Modality Alignment. As shown in Figure 2, each video frame is projected into the space of LLMs in each forward pass. Therefore, it is crucial to ensure visual features are well aligned with the language space. To this end, we construct a compact dataset that contains 790K high-quality image- and video-caption pairs." + }, + { + "type": "code_caption", + "bbox": [ + 0.224, + 0.645, + 0.338, + 0.655 + ], + "angle": 0, + "content": "Stage 1: Modality Alignment" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.656, + 0.475, + 0.78 + ], + "angle": 0, + "content": null + }, + { + "type": "code_caption", + "bbox": [ + 0.506, + 0.645, + 0.655, + 0.655 + ], + "angle": 0, + "content": "Stage 3: Long Video Tuning (Optional)" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.656, + 0.78, + 0.71 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.794, + 0.789, + 0.837 + ], + "angle": 0, + "content": "Fig.3: Multimodal data distribution and instruction format for each model training stage. and denote the token for single image and the \\(i\\)-th video frame, respectively." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LLaMA-VID" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.149, + 0.494, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.205, + 0.493, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.259, + 0.493, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.148, + 0.78, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.202, + 0.78, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.258, + 0.78, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.33, + 0.788, + 0.374 + ], + "angle": 0, + "content": "Fig. 4: An example to construct instruction pairs for the movie Titanic. Given the movie synopsis and script, we utilize the developed LLMs like GPT-4 [45] and Claude-2 [3] to generate movie summaries, plot-related QA pairs, and general reasoning pairs." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.402, + 0.789, + 0.478 + ], + "angle": 0, + "content": "As illustrated in Figure 3, it mainly consists of 558K image-caption pairs from the LLaVA-filtered CC3M dataset [49] and 232K video-caption pairs sampled from the WebVid 2.5M dataset [5]. The instruction format for each modality is presented in Figure 3. In this stage, we optimize the projectors in Figure 2, while freezing the pre-trained modules like the visual encoder and text decoder." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.493, + 0.789, + 0.644 + ], + "angle": 0, + "content": "Instruction Tuning. To enhance the multi-modality understanding of LLMs, we build the instruction pairs from [32] and [39]. In particular, the constructed dataset mainly involves content from three sources, i.e., 40K text conversations from ShareGPT [1], 625K single- or multi-turn visual QA pairs from [19,22,25, 26,33,41-43,48,50], and 98K video QA pairs from [7]. For the instruction, we adopt different formats for text, image, and video input, as shown in Figure 3. The input prompt and answer vary with datasets. Please refer to [32] and [39] for more details. Meanwhile, the image token is randomly inserted at the beginning or end of the user input during our training. In instruction tuning, all the modules are optimized except the visual encoder." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Long Video Tuning. To further unleash the potential for hour-long videos, we construct 15K long QA pairs, including 9K conversions in movie scenes and 6K data sampled from LongLoRA [11] for token expanding. Specifically, we utilize more than 400 long movies and corresponding scripts in MovieNet [21] to build the training set. The key components for instruction generation are visualized in Figure 4. Generally, the generated dataset includes QA pairs from three aspects: video summary, movie plot, and detail reasoning. For video summaries, we collect movie synopses to produce brief and detailed summaries for each movie using developed LLMs like GPT-4 [45]. It brings about 1K summary-level instruction pairs in total. For plot-level data, we take the entire movie synopsis as input and leverage GPT-4 [45] to generate plot-related and character-related QA pairs. These include plot understanding, description, analysis, character relationship," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "L. Yanwei et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "personality, and behavior. In particular, we generate 5 plot-related pairs and 5 character-related pairs for each movie, resulting in 4K plot-level QA data. As for detail-level data, we feed the long movie script into Claude-2 [3] and generate 5 plot-related reasoning pairs and 5 detail-related descriptions for each movie, which brings 4K pairs in total. With long videos and the generated pairs, we perform instruction tuning by concatenating visual tokens and subtitle tokens for each frame, as depicted in Figure 3. In this way, LLaMA-VID can well support 64K tokens with more than 3-hour video as input." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.293, + 0.377, + 0.312 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.787, + 0.358 + ], + "angle": 0, + "content": "In this section, we provide the experimental setup and comparisons with leading methods on several benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.381, + 0.429, + 0.398 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.409, + 0.788, + 0.531 + ], + "angle": 0, + "content": "Implementation Details. In this work, we instantiate the model with the pre-trained EVA-G [16] for visual encoder and QFormer [13] for text decoder by default. During training, we keep the visual encoder fixed in all stages and freeze the text decoder, as well as the LLM, in the modality alignment stage, except for the BERT module in Table 8 that is not pre-trained. Following the strategy in [32], we optimize trainable parameters with the designed data and instructions in Figure 3, running for 1 epoch in each stage. For video input, we extract frames at 1 FPS. All models are trained using \\(8 \\times\\) NVIDIA A100 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.788, + 0.685 + ], + "angle": 0, + "content": "Datasets. In this study, we construct the training set mainly from [5,21,32,39], as illustrated in Section 3.3. Moreover, we report results on several video- and image-based benchmarks. In particular, for video input, we evaluate the zero-shot performance on the open-ended QA benchmarks like MSVD [8], MSRVTT [59], ActivityNet [7], EgoSchema [40], SSv2 (Something-Something v2) [18], and the newly-proposed generative performance benchmark [39]. As for image-based evaluation, we conduct experiments on several widely-adopted benchmarks, including GQA [22], MMB (MMBench) [36], MME [17], POPE [31], SEED [28], \\(\\mathrm{SQA}^{\\mathrm{I}}\\) [37], \\(\\mathrm{VQAT}^{\\mathrm{T}}\\) (TextVQA) [51], VizWiz [20], and \\(\\mathrm{VQAv}^{\\mathrm{v2}}\\) (VQA V2) [19]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.707, + 0.373, + 0.722 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Results on Video-based Benchmarks. In Table 1, we provide a comparative evaluation of LLaMA-VID against various state-of-the-art methods across zero-shot video QA benchmarks: MSVD-QA [8], MSRVTT-QA [59], ActivityNet-QA [7], EgoSchema [40], and SSv2 [18]. Notably, the results are reported with only two tokens for each frame. It is evident that LLaMA-VID, employing Vicuna-7B and Vicuna-13B as the LLMs, consistently delivers superior performance across all datasets. On the MSVD-QA and MSRVTT-QA datasets, it" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LLaMA-VID" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 1: Comparison with leading methods on 3 zero-shot video QA datasets. We report results with 2 tokens for each frame. For fair comparisons, our model is trained without long video tuning (stage 3) in Figure 3. Res indicates image resolution." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.199, + 0.776, + 0.32 + ], + "angle": 0, + "content": "
MethodLLMRes.MSVD-QAMSRVTT-QAActivityNet-QA
AccScoreAccScoreAccScore
VideoLLaMA [64]Vicuna-7B22451.62.529.61.812.41.1
LLaMA-Adapter [65]LLaMA-7B22454.93.143.82.734.22.7
VideoChat [30]Vicuna-7B22456.32.845.02.526.52.2
Video-ChatGPT [39]Vicuna-7B22464.93.349.32.835.22.7
BT-Adapter [34]Vicuna-7B-67.53.757.03.245.73.2
LLaMA-VIDVicuna-7B22469.73.757.73.247.43.3
LLaMA-VIDVicuna-13B22470.03.758.93.347.53.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.333, + 0.788, + 0.391 + ], + "angle": 0, + "content": "Table 2: Comparison with leading methods on the video generative benchmark [39]. We report results with 2 tokens for each frame. For fair comparisons, our model is trained without long video tuning (stage 3) in Figure 3. Res indicates image resolution. Correct, Detail, Context, Temporal, and Consist indicate evaluation metrics in [39]." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.401, + 0.784, + 0.508 + ], + "angle": 0, + "content": "
MethodLLMRes.Correct.DetailContextTemporalConsist.
VideoLLaMA [64]Vicuna-7B2241.962.182.161.821.79
LLaMA-Adapter [65]LLaMA-7B2242.032.322.301.982.15
VideoChat [30]Vicuna-7B2242.232.502.531.942.24
Video-ChatGPT [39]Vicuna-7B2242.402.522.621.982.37
BT-Adapter [34]Vicuna-7B-2.682.693.272.342.46
LLaMA-VIDVicuna-7B2242.963.003.532.462.51
LLaMA-VIDVicuna-13B2243.073.053.602.582.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.522, + 0.788, + 0.551 + ], + "angle": 0, + "content": "Table 3: Comparison with leading methods on the zero-shot benchmark EgoSchema [40]. We report results with 2 tokens for each frame on different settings." + }, + { + "type": "table", + "bbox": [ + 0.243, + 0.562, + 0.756, + 0.646 + ], + "angle": 0, + "content": "
FrameFrozenBiLM [60]mPLUG-Owl [62]InternVideo [56]LLaMA-VID
1026.4%29.6%31.4%41.2%
30-20.0%31.8%41.7%
9026.9%-32.1%40.5%
180---41.7%
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.842 + ], + "angle": 0, + "content": "achieves the accuracy of \\(69.7\\%\\) and \\(57.7\\%\\) with Vicuna-7B, surpassing the previous leading approach [34] with absolute gains of \\(2.2\\%\\) and \\(0.7\\%\\), respectively. For the ActivityNet-QA dataset, LLaMA-VID attains top performance in accuracy and the highest score of 3.3. In Table 2, we also carry out experiments on the newly proposed video-based generative performance benchmark [39]. As for the long EgoSchema and SSv2 dataset, the proposed method surpasses other leading approaches in different frame settings, as presented in Table 3 and Table 4. Specifically, we evaluate the model on the SSv2 dataset similar with that in MSRVTT-QA. Our LLaMA-VID is validated to achieve the best performance across all the evaluation metrics, surpassing previous approaches by a large margin. In general, LLaMA-VID is proved to bring robust results on all benchmarks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.127 + ], + "angle": 0, + "content": "L. Yanwei et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.272, + 0.145, + 0.729, + 0.16 + ], + "angle": 0, + "content": "Table 4: Comparisons with zero-shot methods on SSv2 [18] dataset." + }, + { + "type": "table", + "bbox": [ + 0.33, + 0.171, + 0.673, + 0.227 + ], + "angle": 0, + "content": "
MethodLLMRes.AccScore
Video-ChatGPT [38]Vicuna-7B22411.592.10
LLaMA-VIDVicuna-7B22420.122.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.238, + 0.788, + 0.309 + ], + "angle": 0, + "content": "Table 5: Comparison with leading methods on 8 benchmarks. Here, we use the same training and instruction finetuning data as that in LLaVA-1.5. We report results with 1 context token and \\( n \\) content tokens, where \\( n \\) is kept the same with that in LLaVA-1.5, i.e., \\( n = (336 / 14)^2 = 576 \\). Res indicates input image resolution. * and † denote the train subset is included for training and the data is not publicly available, respectively." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.32, + 0.784, + 0.484 + ], + "angle": 0, + "content": "
MethodLLMRes.GQAMMBMMEPOPESEED\\( \\mathbf{SQA}^{\\dagger} \\)VizWiz\\( VQA^{v2} \\)
InstructBLIP [13]Vicuna-7B22449.236.0--53.460.534.5-
IDEFICS-9B [23]LLaMA-7B22438.448.2----35.550.9
Qwen-VL† [4]Qwen-7B44859.3*38.2--56.367.135.278.8*
Qwen-VL-Chat† [4]Qwen-7B44857.5*60.61487.5-58.268.238.978.2*
LLaVA-1.5 [32]Vicuna-7B33662.0*64.31510.785.958.666.850.078.5*
LLaMA-VIDVicuna-7B33664.3*65.11521.486.059.968.354.279.3*
BLIP-2 [29]Vicuna-13B22441.0-1293.885.346.461.019.641.0
InstructBLIP [13]Vicuna-13B22449.5-1212.878.9-63.133.4-
Shikra [9]Vicuna-13B224-58.8-----77.4*
IDEFICS-80B [23]LLaMA-65B22445.254.5----36.060.0
LLaVA-1.5 [32]Vicuna-13B33663.3*67.71531.385.961.671.653.680.0*
LLaMA-VIDVicuna-13B33665.0*66.61542.386.062.370.054.380.0*
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.789, + 0.707 + ], + "angle": 0, + "content": "Results on Image-based Benchmarks. As illustrated in Section 3.2, LLaMA-VID not only efficiently represents long video, but also expands the upper limit of VLMs with an additional context token. In Table 5, we perform comparisons with leading VLMs on 8 widely-adopted benchmarks. It should be noted that we maintain the same training data and image resolution as in LLaVA-1.5 [32] for fair comparisons. It is evident that LLaMA-VID outperforms other leading methods across most of benchmarks with different LLMs. In particular, with Vicuna-7B, LLaMA-VID attains the best results across all the datasets and surpasses LLaVA-1.5 with significant gains in GQA, MME, and VizWiz, where the improvement reaches up to \\(2.3\\%\\), 10.7, and \\(4.2\\%\\), respectively. With a larger Vicuna-13B, LLaMA-VID also outperforms other methods in 6 benchmarks and achieves top-2 in the other datasets. This demonstrates the generality of LLaMA-VID, which can be scaled up with a stronger foundation model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Qualitative Results. In Figure 5, we apply LLaMA-VID to different types of data, including short videos and long movies. We represent each frame with 2 tokens for videos. LLaMA-VID demonstrates various capabilities with different inputs. Given a short trailer video, LLaMA-VID summarizes the overall plot and infers the movie name, producer, and the type of movie. It shows that the proposed LLaMA-VID can well recognize the typical characters with the produced tokens. As for a 3-hour movie, the proposed model adeptly describes the storyline and demonstrates plot-related reasoning and detailed understanding." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LLaMA-VID" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.145, + 0.482, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.231, + 0.445, + 0.243 + ], + "angle": 0, + "content": "User: Please describe this video in detail." + }, + { + "type": "image", + "bbox": [ + 0.456, + 0.228, + 0.478, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.179, + 0.582, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.196, + 0.602, + 0.218 + ], + "angle": 0, + "content": "3-hour VideoAvatar" + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.147, + 0.676, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.147, + 0.769, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.232, + 0.726, + 0.243 + ], + "angle": 0, + "content": "User: Summarize the plot of this movie." + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.23, + 0.767, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.255, + 0.264, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.271, + 0.258, + 0.475, + 0.356 + ], + "angle": 0, + "content": "Assistant: The video starts with a carto on character standing in front of a castle, which is surrounded by a beautiful lan descape. The character is holding a bucket and appears to be observing the surroundings. The scene then transitions to a t train, where the character is seen sitting inside the train. The train is surrounded by lush greenery and appears to be mov ing through a scenic route...." + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.255, + 0.554, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.56, + 0.258, + 0.763, + 0.356 + ], + "angle": 0, + "content": "Assistant:Avatar follows Jake Sully, a disabled former Marine who is recruited by the RDA to work on Pandora, a moon rich in a valuable mineral called unobtanium. Jake's twin brother, Tommy, died in a helmet-cam accident, and Jake is offered his place in the Avatar Program, which allows humans to link with the native Na' vi, a humanoid species. Jake is assigned to the avatar of his deceased brother...." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.369, + 0.445, + 0.39 + ], + "angle": 0, + "content": "User: What's the name of the movie described in this video?" + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.368, + 0.477, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.369, + 0.706, + 0.38 + ], + "angle": 0, + "content": "User: How human work on Pandora?" + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.366, + 0.766, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.239, + 0.401, + 0.265, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.404, + 0.472, + 0.425 + ], + "angle": 0, + "content": "Assistant: The movie described in this video is \"Zootopia.\"" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.389, + 0.554, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.561, + 0.393, + 0.761, + 0.442 + ], + "angle": 0, + "content": "Assistant: Humans work on Pandora by using avatars, which are grown from human DNA mixed with the DNA of the native Na'vi, to interact with the environment and the Na'vi." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.439, + 0.383, + 0.45 + ], + "angle": 0, + "content": "User: Who produce the movie?" + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.436, + 0.477, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.455, + 0.737, + 0.477 + ], + "angle": 0, + "content": "User: Why Jake Sully decide to fight againt humanity with the Na'vi?" + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.454, + 0.767, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.462, + 0.264, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.465, + 0.469, + 0.486 + ], + "angle": 0, + "content": "Assistant: The movie is produced by Walt Disney Pictures." + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.486, + 0.554, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.56, + 0.489, + 0.761, + 0.548 + ], + "angle": 0, + "content": "Assistant: Jake Sully decides to fight a gainst humanity with the Na'vi after he falls in love with the Na'vi and their way of life, and after he realizes that the human military is destroying their home and culture." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.5, + 0.425, + 0.512 + ], + "angle": 0, + "content": "User: What's the type of this movie?" + }, + { + "type": "image", + "bbox": [ + 0.456, + 0.498, + 0.476, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.522, + 0.264, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.527, + 0.463, + 0.547 + ], + "angle": 0, + "content": "Assistant: The movie is a Disney animat ed film." + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.565, + 0.741, + 0.58 + ], + "angle": 0, + "content": "Fig. 5: Examples of LLaMA-VID with short video and hour-long video." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.611, + 0.475, + 0.627 + ], + "angle": 0, + "content": "4.3 Component-wise Analysis" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.64, + 0.787, + 0.671 + ], + "angle": 0, + "content": "In this subsection, we conduct ablation studies with input resolution 224 and 2 tokens for each image by default." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Generated Token Types. As illustrated in Figure 2, each image is represented with a context token and a content token in LLMs. To validate the effectiveness of each part, we conduct experiments with different types of tokens in Table 6. Without the context token, the compressed content token, which encodes each image with 1 token, cannot adjust to input instructions, leading to subpar performance. Compared with a single content token, the instruction-guided context token results in significant gains across all datasets with only 1 token. With both tokens for each image, the model achieves the best performance across all image- and video-based benchmarks. It shows that both instruction cues in the context token and the image content itself in the content token are important." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.127 + ], + "angle": 0, + "content": "L. Yanwei et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 6: Comparison with different token types. We report results with 1 context token (if exists) and 1 content token on image-based and video-based benchmarks (MSVD)." + }, + { + "type": "table", + "bbox": [ + 0.289, + 0.185, + 0.711, + 0.251 + ], + "angle": 0, + "content": "
contextcontentGQAPOPE\\( \\mathbf{SQA^I} \\)\\( VQA^T \\)MSVD
✘✓53.380.966.146.565.9
✓✘54.382.467.748.369.3
✓✓55.583.168.849.069.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.265, + 0.787, + 0.294 + ], + "angle": 0, + "content": "Table 7: Comparison with different token numbers. We report results with various numbers of context token and content token on image-based benchmarks for efficiency." + }, + { + "type": "table", + "bbox": [ + 0.322, + 0.306, + 0.678, + 0.415 + ], + "angle": 0, + "content": "
contextcontentGQAPOPE\\( \\mathbf{SQA^I} \\)\\( VQA^T \\)
025661.985.567.553.0
125663.086.667.753.8
16460.885.168.752.3
11658.283.167.450.8
1456.283.568.749.1
1155.583.168.849.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.788, + 0.732 + ], + "angle": 0, + "content": "Generated Token Numbers. In Table 7, we conduct experiments with different numbers of tokens for further investigation. With an image size \\(224 \\times 224\\), we set up experiments with \\(n\\) content tokens, where \\(n = (224 / 14)^{2} = 256\\) for uncompressed settings in the first two rows. The results clearly show that the context token consistently improves performance across different benchmarks with only 1 extra token. When we compress the content token to \\(1 / 4\\) with \\(n = 64\\), the performance drops about \\(1\\%\\) to \\(2\\%\\) but increases \\(1\\%\\) in \\(\\mathrm{SQA}^{\\mathrm{I}}\\). Considering the extra efficient setting for hour-long videos, we compress the content token to \\(1 / 256\\) with \\(n = 1\\) by default. Compared to the original setting without context token, we can reduce the computational cost to \\(1 / 128\\) with about \\(2\\% - 6\\%\\) performance drop, which is generally acceptable. The linear increase in performance presents significant potential for token compression. For instance, we can dynamically compress the content token to different numbers according to resource budget and content importance. Interestingly, the model achieves peak performance in \\(\\mathrm{SQA}^{\\mathrm{I}}\\) with only 2 tokens. This could be attributed to the fact that problems in ScienceQA [37] focus more on visual-based reasoning rather than image details. As demonstrated in Tables 1 and 2, with only 2 tokens for each image, LLaMA-VID still outperforms all previous work in different video-based benchmarks. This makes it feasible to enable LLMs for hour-long video processing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Text Decoder. As depicted in Figure 2, the text decoder plays an essential role in producing instruction-guided context cues. Here, we further perform comparisons with different text decoders in Table 8. We mainly instantiate the text decoder with two types of modules, namely BERT [14] and QFormer [13]. For BERT, we randomly initialize it as a cross-modality decoder and only retain the first two layers. As for QFormer, we utilize the pre-trained modules and fix them" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LLaMA-VID" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 8: Comparison with different text decoders. We report results with 1 context token (if exists) and 1 content token on image-based benchmarks for efficiency." + }, + { + "type": "table", + "bbox": [ + 0.347, + 0.185, + 0.654, + 0.253 + ], + "angle": 0, + "content": "
textGQAPOPE\\( \\mathbf{SQA^I} \\)\\( \\mathbf{VQA^T} \\)
-53.380.966.146.5
BERT54.180.867.948.1
QFormer55.583.168.849.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.267, + 0.787, + 0.31 + ], + "angle": 0, + "content": "Table 9: Average module latency of video frame on a A100 GPU. Encoder, Content, Decoder, and Context denote the latency for text-free visual encoder and content token, text-related text decoder and context token generation, respectively." + }, + { + "type": "table", + "bbox": [ + 0.294, + 0.321, + 0.706, + 0.389 + ], + "angle": 0, + "content": "
textText-freeText-relatedTotal (ms)
EncoderContentDecoderContext
BERT4.890.010.250.025.17
QFormer4.870.011.080.025.98
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.785, + 0.496 + ], + "angle": 0, + "content": "for modality alignment. Even with a simple 2-layer BERT, as shown in Table 8, the generated context token achieves significant gains in most of benchmarks. This proves the effectiveness of the paradigm for context token generation. With a pre-trained text decoder like QFormer, the model can be further enhanced and attains peak performance in all datasets with \\(2.2\\%\\) to \\(2.7\\%\\) significant gain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.653 + ], + "angle": 0, + "content": "Latency Analysis. To further analyse the module latency, we keep the original setting in Table 8 and report average module latency over 100 videos in Table 9. Here, we divide the modules for generation into two types, namely text-free and text-related. In particular, we generate visual tokens one time for each frame with text-free modules. And text-related modules are required to produce different context tokens according to user inputs. It is clear that the text-free tokens account for the most computation, while the text-related modules only cost \\(5\\%\\) to \\(18.3\\%\\) for each user instruction. That means the total cost will not increase greatly given variant user questions because of the efficient text-related modules." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Response in Context Attention. To more vividly explore the context attention, we visualize the high response areas with the top 20 scores in Figure 6. Specifically, we draw the normalized heatmap for the first four queries in \\(\\mathbf{Q}_t\\) before applying the Softmax function, which is used to formulate context token in Equation 1. As shown in Figure 6, the text-guided query \\(\\mathbf{Q}_t\\) effectively focuses on important areas relevant to the input questions. For example, in the second row, when inquiring whether the image depicts a fishing village, the query \\(\\mathbf{Q}_t\\) focuses more on buildings along the river and a seagull. These are all typical characteristics to distinguish a fishing village in common sense. Other examples also confirm that the designed context attention successfully achieves its goal of formulating the context token under instruction guidance." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.127 + ], + "angle": 0, + "content": "L. Yanwei et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.148, + 0.332, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.207, + 0.291, + 0.216 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.148, + 0.449, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.207, + 0.414, + 0.216 + ], + "angle": 0, + "content": "Query 1" + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.148, + 0.56, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.487, + 0.207, + 0.525, + 0.216 + ], + "angle": 0, + "content": "Query 2" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.148, + 0.671, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.207, + 0.637, + 0.216 + ], + "angle": 0, + "content": "Query 3" + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.148, + 0.783, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.708, + 0.207, + 0.747, + 0.216 + ], + "angle": 0, + "content": "Query 4" + }, + { + "type": "image_caption", + "bbox": [ + 0.399, + 0.22, + 0.605, + 0.234 + ], + "angle": 0, + "content": "Question: Are the elephants fighting?" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.246, + 0.332, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.312, + 0.291, + 0.321 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.246, + 0.449, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.376, + 0.312, + 0.414, + 0.321 + ], + "angle": 0, + "content": "Query 1" + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.246, + 0.56, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.487, + 0.312, + 0.525, + 0.321 + ], + "angle": 0, + "content": "Query 2" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.246, + 0.671, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.312, + 0.637, + 0.321 + ], + "angle": 0, + "content": "Query 3" + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.246, + 0.783, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.708, + 0.312, + 0.747, + 0.321 + ], + "angle": 0, + "content": "Query 4" + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.327, + 0.596, + 0.34 + ], + "angle": 0, + "content": "Question: Is this a fishing village?" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.353, + 0.332, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.413, + 0.291, + 0.422 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.353, + 0.449, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.376, + 0.413, + 0.414, + 0.421 + ], + "angle": 0, + "content": "Query 1" + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.353, + 0.56, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.487, + 0.413, + 0.525, + 0.421 + ], + "angle": 0, + "content": "Query 2" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.353, + 0.671, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.413, + 0.637, + 0.422 + ], + "angle": 0, + "content": "Query 3" + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.353, + 0.783, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.708, + 0.413, + 0.747, + 0.422 + ], + "angle": 0, + "content": "Query 4" + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.426, + 0.666, + 0.439 + ], + "angle": 0, + "content": "Question: What tall animals are displayed in the background?" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.45, + 0.788, + 0.48 + ], + "angle": 0, + "content": "Fig. 6: High response areas with top scores to input question in Equation 1. We present the response of first four queries in \\(\\mathbf{Q}_t\\). Images are sampled from VQA V2 test-dev set." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.507, + 0.51, + 0.523 + ], + "angle": 0, + "content": "5 Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.788, + 0.704 + ], + "angle": 0, + "content": "We have introduced LLaMA-VID, a simple yet effective token generation approach for VLMs. The central concept behind LLaMA-VID is to represent an image with the context token and the content token. In particular, the context token is generated according to input instructions, and the content token is produced based on the image content. Depending on the budget, the content token can be compressed to one token or expressed without compression. It allows us to represent a single image with preserved details and efficiently encode each video frame with only two tokens. Moreover, we have constructed an instruction dataset for hour-long video understanding. Our experiments on several video-and image-based benchmarks prove the superiority of our method. We hope that LLaMA-VID can serve as a strong benchmark for efficient visual representation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.78 + ], + "angle": 0, + "content": "There still exists certain limitations in the current method. Although LLaMAVID succeeds to represent hour-long videos in LLMs, it focuses more on temporal consistency and may lose spatial details in each frame. In the future work, we plan to maintain dense spatial information for each frame outside the LLM and conduct the visual token fusion within LLMs to avoid such detail loss." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.787, + 0.827 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported in part by the Research Grants Council under the Areas of Excellence scheme grant AoE/E-601/22-R." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LLaMA-VID" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.54, + 0.19 + ], + "angle": 0, + "content": "1. Sharegpt. https://sharegpt.com/ (2023) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.191, + 0.787, + 0.232 + ], + "angle": 0, + "content": "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: NeurIPS (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.778, + 0.245 + ], + "angle": 0, + "content": "3. Anthropic: Claude 2. https://www.anthropic.com/index/claude-2 (2023) 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "4. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv:2308.12966 (2023) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.786, + 0.314 + ], + "angle": 0, + "content": "5. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: ICCV (2021) 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.315, + 0.786, + 0.355 + ], + "angle": 0, + "content": "6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.357, + 0.786, + 0.396 + ], + "angle": 0, + "content": "7. Caba Heilbron, F., Escorcia, V., Ghanem, B., Carlos Niebles, J.: Activitynet: A large-scale video benchmark for human activity understanding. In: CVPR (2015) 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.786, + 0.424 + ], + "angle": 0, + "content": "8. Chen, D., Dolan, W.B.: Collecting highly parallel data for paraphrase evaluation. In: ACL (2011) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.786, + 0.452 + ], + "angle": 0, + "content": "9. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv:2306.15195 (2023) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.453, + 0.786, + 0.494 + ], + "angle": 0, + "content": "10. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollár, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv:1504.00325 (2015) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.495, + 0.786, + 0.522 + ], + "angle": 0, + "content": "11. Chen, Y., Qian, S., Tang, H., Lai, X., Liu, Z., Han, S., Jia, J.: Longlora: Efficient fine-tuning of long-context large language models. arXiv:2309.12307 (2023) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.523, + 0.786, + 0.577 + ], + "angle": 0, + "content": "12. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality. https://lmsys.org/blog/2023-03-30-vicuna/ (2023) 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.578, + 0.786, + 0.618 + ], + "angle": 0, + "content": "13. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv:2305.06500 (2023) 1, 2, 4, 5, 6, 8, 10, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.619, + 0.786, + 0.658 + ], + "angle": 0, + "content": "14. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv:1810.04805 (2018) 2, 3, 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.66, + 0.786, + 0.702 + ], + "angle": 0, + "content": "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.702, + 0.786, + 0.743 + ], + "angle": 0, + "content": "16. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. In: CVPR (2023) 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.744, + 0.786, + 0.785 + ], + "angle": 0, + "content": "17. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv:2306.13394 (2023) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.786, + 0.786, + 0.84 + ], + "angle": 0, + "content": "18. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The\" something something\" video database for learning and evaluating visual common sense. In: ICCV (2017) 8, 10" + }, + { + "type": "list", + "bbox": [ + 0.219, + 0.177, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.127 + ], + "angle": 0, + "content": "L. Yanwei et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "19. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: CVPR (2017) 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.786, + 0.232 + ], + "angle": 0, + "content": "20. Gurari, D., Li, Q., Stangl, A.J., Guo, A., Lin, C., Grauman, K., Luo, J., Bigham, J.P.: Vizwiz grand challenge: Answering visual questions from blind people. In: CVPR (2018) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.786, + 0.261 + ], + "angle": 0, + "content": "21. Huang, Q., Xiong, Y., Rao, A., Wang, J., Lin, D.: Movienet: A holistic dataset for movie understanding. In: ECCV (2020) 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.786, + 0.289 + ], + "angle": 0, + "content": "22. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: CVPR (2019) 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.786, + 0.318 + ], + "angle": 0, + "content": "23. IDEFICS: Introducing idefics: An open reproduction of state-of-the-art visual language model. https://huggingface.co/blog/idefics (2023) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.319, + 0.786, + 0.36 + ], + "angle": 0, + "content": "24. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.786, + 0.388 + ], + "angle": 0, + "content": "25. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: Referitgame: Referring to objects in photographs of natural scenes. In: EMNLP (2014) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.786, + 0.431 + ], + "angle": 0, + "content": "26. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV (2017) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.786, + 0.459 + ], + "angle": 0, + "content": "27. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. arXiv:2308.00692 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.786, + 0.487 + ], + "angle": 0, + "content": "28. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv:2307.16125 (2023) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.786, + 0.529 + ], + "angle": 0, + "content": "29. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv:2301.12597 (2023) 1, 2, 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.786, + 0.558 + ], + "angle": 0, + "content": "30. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chat-centric video understanding. arXiv:2305.06355 (2023) 1, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.786, + 0.586 + ], + "angle": 0, + "content": "31. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv:2305.10355 (2023) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.588, + 0.786, + 0.614 + ], + "angle": 0, + "content": "32. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv:2310.03744 (2023) 6, 7, 8, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.786, + 0.642 + ], + "angle": 0, + "content": "33. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeruIPS (2023) 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.786, + 0.672 + ], + "angle": 0, + "content": "34. Liu, R., Li, C., Ge, Y., Shan, Y., Li, T.H., Li, G.: One for all: Video conversation is feasible without video instruction tuning. arXiv:2309.15785 (2023) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.673, + 0.786, + 0.713 + ], + "angle": 0, + "content": "35. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. arXiv:1907.11692 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.715, + 0.786, + 0.756 + ], + "angle": 0, + "content": "36. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv:2307.06281 (2023) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.786, + 0.799 + ], + "angle": 0, + "content": "37. Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. In: NeurIPS (2022) 3, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.786, + 0.841 + ], + "angle": 0, + "content": "38. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability. arXiv:2306.07207 (2023) 1" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LLaMA-VID" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.188 + ], + "angle": 0, + "content": "39. Maaz, M., Rasheed, H., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv:2306.05424 (2023) 1, 4, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.787, + 0.217 + ], + "angle": 0, + "content": "40. Mangalam, K., Akshulakov, R., Malik, J.: Egoschema: A diagnostic benchmark for very long-form video language understanding. NeurIPS (2024) 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.217, + 0.787, + 0.243 + ], + "angle": 0, + "content": "41. Mao, J., Huang, J., Toshev, A., Camburu, O., Yuille, A.L., Murphy, K.: Generation and comprehension of unambiguous object descriptions. In: CVPR (2016) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.243, + 0.787, + 0.27 + ], + "angle": 0, + "content": "42. Mao, J., Huang, J., Toshev, A., Camburu, O., Yuille, A.L., Murphy, K.: Generation and comprehension of unambiguous object descriptions. In: CVPR (2016) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.27, + 0.787, + 0.298 + ], + "angle": 0, + "content": "43. Mishra, A., Shekhar, S., Singh, A.K., Chakraborty, A.: Ocr-vqa: Visual question answering by reading text in images. In: ICDAR (2019) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.298, + 0.698, + 0.31 + ], + "angle": 0, + "content": "44. OpenAI: Chatgpt. https://openai.com/blog/chatgpt/ (2023) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.311, + 0.675, + 0.324 + ], + "angle": 0, + "content": "45. OpenAI: Gpt-4 technical report. arXiv:2303.08774 (2023) 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.324, + 0.787, + 0.364 + ], + "angle": 0, + "content": "46. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. In: NeurIPS (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.364, + 0.787, + 0.405 + ], + "angle": 0, + "content": "47. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.405, + 0.787, + 0.444 + ], + "angle": 0, + "content": "48. Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: A-okvqa: A benchmark for visual question answering using world knowledge. In: ECCV (2022) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.446, + 0.787, + 0.487 + ], + "angle": 0, + "content": "49. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: ACL (2018) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.487, + 0.787, + 0.514 + ], + "angle": 0, + "content": "50. Sidorov, O., Hu, R., Rohrbach, M., Singh, A.: Textcaps: a dataset for image captioning with reading comprehension. In: ECCV (2020) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.514, + 0.787, + 0.542 + ], + "angle": 0, + "content": "51. Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards vqa models that can read. In: CVPR (2019) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.542, + 0.787, + 0.582 + ], + "angle": 0, + "content": "52. Song, E., Chai, W., Wang, G., Zhang, Y., Zhou, H., Wu, F., Chi, H., Guo, X., Ye, T., Zhang, Y., et al.: Moviechat: From dense token to sparse memory for long video understanding. In: CVPR (2024) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.582, + 0.787, + 0.623 + ], + "angle": 0, + "content": "53. Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca (2023) 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.623, + 0.787, + 0.678 + ], + "angle": 0, + "content": "54. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models. arXiv:2302.13971 (2023) 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.678, + 0.787, + 0.704 + ], + "angle": 0, + "content": "55. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.745 + ], + "angle": 0, + "content": "56. Wang, Y., Li, K., Li, Y., He, Y., Huang, B., Zhao, Z., Zhang, H., Xu, J., Liu, Y., Wang, Z., et al.: Intervideo: General video foundation models via generative and discriminative learning. arXiv:2212.03191 (2022) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.745, + 0.787, + 0.786 + ], + "angle": 0, + "content": "57. Wei, J., Bosma, M., Zhao, V.Y., Guu, K., Yu, A.W., Lester, B., Du, N., Dai, A.M., Le, Q.V.: Finetuned language models are zero-shot learners. arXiv:2109.01652 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.786, + 0.787, + 0.814 + ], + "angle": 0, + "content": "58. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv:2303.04671 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.814, + 0.787, + 0.84 + ], + "angle": 0, + "content": "59. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: CVPR (2016) 3, 8" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.127 + ], + "angle": 0, + "content": "L. Yanwei et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "60. Yang, A., Miech, A., Sivic, J., Laptev, I., Schmid, C.: Zero-shot video question answering via frozen bidirectional language models. In: NeurIPS (2022) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.203 + ], + "angle": 0, + "content": "61. Yang, R., Song, L., Li, Y., Zhao, S., Ge, Y., Li, X., Shan, Y.: Gpt4tools: Teaching large language model to use tools via self-instruction. arXiv:2305.18752 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.245 + ], + "angle": 0, + "content": "62. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv:2304.14178 (2023) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.285 + ], + "angle": 0, + "content": "63. Zhang, C., Lu, T., Islam, M.M., Wang, Z., Yu, S., Bansal, M., Bertasius, G.: A simple lvm framework for long-range video question-answering. arXiv:2312.17235 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.314 + ], + "angle": 0, + "content": "64. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv:2306.02858 (2023) 1, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.314, + 0.785, + 0.355 + ], + "angle": 0, + "content": "65. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv:2303.16199 (2023) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.355, + 0.785, + 0.397 + ], + "angle": 0, + "content": "66. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv:2205.01068 (2022) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.397, + 0.785, + 0.438 + ], + "angle": 0, + "content": "67. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv:2304.10592 (2023) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.438 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/LLaMA-VID_ An Image is Worth 2 Tokens in Large Language Models/022c7268-b560-44a0-867d-3ec58c86bb1a_origin.pdf b/2024/LLaMA-VID_ An Image is Worth 2 Tokens in Large Language Models/022c7268-b560-44a0-867d-3ec58c86bb1a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1353324015443d60cba2abfd52195f7b80bbd06a --- /dev/null +++ b/2024/LLaMA-VID_ An Image is Worth 2 Tokens in Large Language Models/022c7268-b560-44a0-867d-3ec58c86bb1a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f3a98c6d70356b9a24d9d4d679881bab1a44dcbb2f61877231e21ec604f3f25 +size 1840065 diff --git a/2024/LLaMA-VID_ An Image is Worth 2 Tokens in Large Language Models/full.md b/2024/LLaMA-VID_ An Image is Worth 2 Tokens in Large Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9bf85a8cb8a9f89e42e5af51ceddf970be4e61e8 --- /dev/null +++ b/2024/LLaMA-VID_ An Image is Worth 2 Tokens in Large Language Models/full.md @@ -0,0 +1,388 @@ +# LLaMA-VID: An Image is Worth 2 Tokens in Large Language Models + +Yanwei Li $^{1\star}$ Chengyao Wang $^{1*}$ Jiaya Jia $^{1,2}$ + +CUHK $^{1}$ SmartMore $^{2}$ + +Abstract. In this work, we present a novel method to tackle the token generation challenge in Vision Language Models (VLMs) for video and image understanding, called LLaMA-VID. Current VLMs, while proficient in tasks like image captioning and visual question answering, face computational burdens when processing long videos due to the excessive visual tokens. LLaMA-VID addresses this issue by representing each frame with two distinct tokens, namely context token and content token. The context token encodes the overall image context based on user input, whereas the content token encapsulates visual cues in each frame. This dual-token strategy significantly reduces the overload of long videos while preserving critical information. Generally, LLaMA-VID empowers existing frameworks to support hour-long videos and pushes their upper limit with an extra context token. It is demonstrated to surpass previous methods on most of video- or image-based benchmarks. Code and models are available at https://github.com/dvlab-research/LLaMA-VID. + +# 1 Introduction + +Large Language Models (LLMs) [44, 54, 66], through their capacity to generate contextually accurate responses, have significantly advanced the field of AI. Drawing from the strengths of LLMs, Vision Language Models (VLMs) [13,33,45] have been developed to extend these capabilities to visual data, demonstrating their adeptness in tasks like image captioning and visual question answering. However, a substantial challenge emerges in the context of long video, where an excessive number of tokens are required to represent consecutive frames. The computational demands escalate with the video length, thereby constraining the practical application of VLMs for extensive videos. + +Recently, several approaches have been proposed to handle videos, moving beyond image-only VLMs. These methods aim to alleviate the token issue by utilizing representative queries [30, 64] or applying temporal compression [38, 39, 52]. Despite these efforts, the challenge of long videos remains unresolved. The primary obstacle stems from the excessive number of tokens required for each video frame. For instance, models like BLIP [13, 29] and LLaVA [33] require 32 and over 256 tokens respectively for a single image. A video containing 10K frames would thus necessitate over 320K tokens, exceeding the capacity of current VLMs. + +![](images/f29a384f07e4c9034fdc8fc584dea006ebd3d391385baac25d928b725800a86a.jpg) +(a) Video-based benchmarks + +![](images/31ed8192d04c010148629d519032cf3f54fca4eb3ee276cc14e594824bef8d96.jpg) +(b) Image-based benchmarks +Fig. 1: The proposed LLaMA-VID achieves leading performance on most of benchmarks with 7B LLMs. The video-based and image-based benchmarks are noted in blue and purple color, respectively. Please refer to Tables 1, 2, and 5 for more details. + +Furthermore, simple temporal compression can significantly damage the representation over long-term intervals. This drawback hampers their performance, thereby underscoring the need for a robust solution. + +In this work, we present LLaMA-VID, a novel approach to effectively manage the token generation issue in long videos. Our core idea is to represent each video frame with two distinct tokens: context token and content token. The context token is designed to encode the overall context of the image based on user input, which efficiently condenses the broader picture into a single token. Simultaneously, the content token captures finer aspects of each frame. According to computational constraints, the length of content token can be extended to include more details, e.g., 1 token/frame for video input and beyond 256 token/frame for single image. In this way, the overload of long videos can be significantly reduced without sacrificing critical information. + +In particular, our method employs a dual-token generation strategy that is both efficient and effective. For each frame, we first extract image features using a pre-trained vision transformer [15], akin to other VLMs [13, 33]. The key question is how to generate the context-related token according to user instructions. We provide the solution by leveraging the cross-modality design [14,29] for instruction-guided queries, which carry the interactive intention from users. For context token, these queries interact with previously generated image features in the designed attention module, termed as context attention. To generate content token, the image features are average pooled to formulate tokens that adapt to different settings. For instance, global pooling is adopted to maintain efficiency for video input while details are preserved with more tokens for single image input. The context and content tokens are subsequently projected to the space of LLMs with simple linear layers for final prediction. Furthermore, to better sup + +port hour-long videos in VLMs, we construct an instruction dataset that contains 9K movie-level conversations for plot reasoning and detail understanding. + +Generally, LLaMA-VID can be distinguished from two aspects. On one hand, with the dual-token paradigm, each frame can be efficiently encoded with only two tokens, which empowers existing LLMs to support long videos. On the other hand, the context token aggregates the most informative feature of each image, which further extends the upper limit of VLMs with an extra token. + +The overall framework, dubbed LLaMA-VID, can be easily instantiated with various decoders and LLMs, as elaborated in Section 3. Extensive empirical studies are conducted in Section 4 to reveal the effectiveness of each component. Remarkably, our model can complete training within 2 days on a single machine with $8 \times \mathrm{A}100$ GPUs, and it outperforms previous leading methods on most of video- and image-based benchmarks, as shown in Figure 1. + +# 2 Related Work + +In this section, we first review large language models and delve into recent advances in vision language models. + +# 2.1 Large Language Models + +The field of Natural Language Processing (NLP) has witnessed tremendous advancements with the evolution of LLMs. Transformer [55] marked a pivotal milestone, with subsequent language models [14, 35, 66] demonstrating remarkable capabilities. GPT [6] revolutionized this field by utilizing generative pre-trained transformers for auto-regressive prediction, which is proved to be a potent language modeling paradigm. Recent groundbreaking works, such as ChatGPT [44], GPT-4 [45], and LLaMA [54], have pushed the boundaries even further. Trained on vast amounts of text data, these models exhibit exceptional capabilities in complex linguistic tasks. To leverage the potential of pre-trained LLMs, instruction tuning [46, 57] is a crucial component for high-quality output. This strategy is widely adopted in open-source models like Alpaca [53] and Vicuna [12], which improve over LLaMA [54] using specially designed instruction pairs. There are also researches [58, 61] that utilize the reasoning ability of LLMs and invoke pre-defined tools for visual applications. Different from them, we collect multimodality instruction data that contains text, images, and videos in this work, which is employed to empower LLMs for long video processing. + +# 2.2 Vision Language Models + +The advancements in computer vision and NLP have led to the emergence of vision-language models (VLMs) that integrate vision models with language models for cross-modality understanding [10,59] and reasoning [19,27,37]. Pioneering large-scale VLMs like CLIP [47] and ALIGN [24] have extended language models to vision-language tasks. The recent progress has seen an increasing focus + +![](images/3386a0651921af16022c75a39b6ad5c4586f118f8c530a9f1590adf29a9a5540.jpg) +Fig. 2: The framework of LLaMA-VID. With user directive, it operates by taking either a single image or video frames as input, and generates responses from LLM. The process initiates with a visual encoder that transforms input frames into the visual embedding. Then, the text decoder produces text queries based on the user input. In context attention, the text query aggregates text-related visual cues. For efficiency, an option is provided to downsample the visual embedding to various token sizes, or even to a single token. The text-guided context token and the visually-enriched content token are then formulated using a linear projector to represent each frame at time $t$ . Finally, the LLM takes the user directive and all visual tokens as input and gives responses. + +on leveraging the power of LLMs. Notably, Flamingo [2] and BLIP-2 [29] utilize web-scale image-text pairs for cross-modality alignment, thereby enhancing learning performance. To further exploit the potential of such pre-trained models, InstructBLIP [13] and MiniGPT-4 [67] construct high-quality instruction pairs based on BLIP-2 and achieve superior results. Simultaneously, LLaVA [33] employs a simple linear projector with a few learnable parameters to align the image and text space of LLaMA. Given the tailored instruction data, this straightforward approach demonstrates strong capabilities. To support video understanding in LLMs, several studies [30,52,64] attempt to utilize BLIP-2 for video embedding or text-only caption [63] extraction, while Video-ChatGPT [39] proposes spatial and temporal pooling for video features. However, given the substantial number of tokens required for each frame, LLMs encounter significant challenges when processing extensive video sequences. It prevents previous work from representing long video sequences that exceed a duration of one hour in LLMs. To solve the issue, we propose to efficiently encode each frame with only 2 tokens, which supports long video understanding in existing LLMs. + +# 3 LLaMA-VID + +The framework of LLaMA-VID is conceptually simple: encoder and decoder are adopted to produce visual embedding and text-guided features, respectively; context token and content token are transformed with the tailored token generation strategy; instruction tuning is designed to unleash the potential of LLMs for image and video. + +# 3.1 Encoder and Decoder + +The proposed LLaMA-VID can be utilized to interact with single image or long videos. For clarity, we assume the input image is captured from a video sequence, as presented in Figure 2. Given a video frame $\mathbf{V}_t \in \mathbb{R}^{H \times W \times 3}$ at time $t$ , a transformer-based visual encoder is first employed to produce the visual embedding $\mathbf{X}_t \in \mathbb{R}^{N \times C}$ . Here, $N = H / p \times W / p$ and $C$ indicate the number of image patches and embedding channels, respectively. The patch size $p$ is typically set to 14 for ViT-based backbones [15, 16, 47]. Meanwhile, we take the user instruction as input and generate the text-guided query $\mathbf{Q}_t \in \mathbb{R}^{M \times C}$ with the produced $\mathbf{X}_t$ , where $M$ denotes the number of queries. As depicted in Figure 2, this cross-modality interaction predominantly occurs in the text decoder, which can be easily instantiated with BERT [14] or QFormer [13], as compared in Table 8. In this way, the text query $\mathbf{Q}_t$ contains highlighted visual cues that are most related to the user instruction. + +# 3.2 Token Generation + +With the text query $\mathbf{Q}_t$ and visual embedding $\mathbf{X}_t$ , we can easily generate representative tokens for LLMs. Specifically, context attention is designed to aggregate text-related visual features and condense them to a single context token. As shown in Figure 2, it takes $\mathbf{Q}_t$ and $\mathbf{X}_t$ as input and formulates the context-related embedding $\mathbf{E}_t \in \mathbb{R}^{1 \times C}$ as + +$$ +\mathbf {E} _ {t} = \operatorname {M e a n} \left(\operatorname {S o f t m a x} \left(\left(\mathbf {Q} _ {t} \times \mathbf {X} _ {t} ^ {\top}\right) / \sqrt {C}\right) \times \mathbf {X} _ {t}\right), \tag {1} +$$ + +where the Softmax function and Mean operation are conducted along the $N$ and $M$ dimensions, respectively. Unlike QFormer [13] that adopts 32 visual queries as LLMs tokens, we only utilize the text query $\mathbf{Q}_t$ to aggregate the visual features with high-response scores to input instructions. As a result, the most crucial visual cues related to user input are efficiently preserved in the condensed embedding $\mathbf{E}_t$ . The effectiveness of this context-related token generation is demonstrated in Table 6 and Figure 6. Subsequently, a linear projector is utilized to transform the embedding $\mathbf{E}_t$ into the context token $\mathbf{E}_t^T \in \mathbb{R}^{1 \times C}$ , which aligns with the language space of LLMs. Meanwhile, we employ an adaptive pooling strategy for the visual embedding according to computational constraints to produce the content token $\mathbf{E}_t^V \in \mathbb{R}^{n \times C}$ , where $n \in [1, N]$ . For instance, we maintain the original resolution of visual embedding $\mathbf{X}_t$ when input single image, while we downsample $\mathbf{X}_t$ to 1 token for long videos. This approach significantly reduces the overload of LLMs for each frame, thereby supporting hour-long videos effectively. Finally, the generated context token $\mathbf{E}_t^T$ and the content token $\mathbf{E}_t^V$ are concatenated to represent the frame at time $t$ . Along with frames at other timestamps, the entire video sequence is translated into the language space in token format, which is then used to generate responses from LLMs. The whole process is summarized in Algorithm 1. + +Algorithm 1 Pseudo Code for Token Generation. +Stage 1: Modality Alignment +Stage 3: Long Video Tuning (Optional) +```python +B: batch size; C: channel size; n: content shape +# M: query length; N: shape of flatten image pacthes; +# text_q: text query in shape (B, M, C) +# vis_embedding: visual embedding in shape (B, N, C) +# Key part 1: calculate context-related embedding +ctx_embedding = text_q @ vis_embedding.transpose(-1,-2) +ctx_embedding = ctx_embedding / (vis_embedding.shape[-1]**0.5) +ctx_embedding = (ctx_embedding-largest(-1)@vis_embedding).mean(1) +ctx_embedding = self.batchproj(ctx_embedding[:,None]) +# Key part 2: calculate visual embedding +cur_shape = int(vis_embedding.shape[1]**0.5) +vis_embedding = vis_embedding.reshape(B, cur_shape, -1, C) +vis_embedding = F AVG_pool2d(vis_embedding.permute(0,3,1,2), kernel_size=cur_shape//n, +stride=cur_shape//n) +vis_embedding = vis_embedding.permute(0,2,3,1).flatten(1,2) +vis_embedding = self.visproj(vis_embedding) +# concat token in shape (B, n+1, C), n in [1,N] +final_token = torch.cat([ctx_embedding, vis_embedding], dim=1) +``` + +# 3.3 Training Strategy + +Training strategy, particularly instruction tuning, has proven to be crucial in LLMs [12, 53, 54] and VLMs [13, 32, 33]. Considering training efficiency, in this work, we divide the training procedure into three stages, i.e., modality alignment, instruction tuning, and long video tuning. + +Modality Alignment. As shown in Figure 2, each video frame is projected into the space of LLMs in each forward pass. Therefore, it is crucial to ensure visual features are well aligned with the language space. To this end, we construct a compact dataset that contains 790K high-quality image- and video-caption pairs. + +![](images/62f3d2892bcd1ea522a3318d308948c4fa02c4266572185b880f28f9a46a99fc.jpg) +Fig.3: Multimodal data distribution and instruction format for each model training stage. and denote the token for single image and the $i$ -th video frame, respectively. + +![](images/27d3cdddae9ee3ab385bc27e54e68df0c0f44ca9c5723681c3e3b91829be4218.jpg) + +![](images/7f197feedc3a18f177efe7ec95c499f55cbe7cb9742d4a77850dcf4ece39ae61.jpg) + +![](images/2c5aa9aa91eea4ed1cf89a2e60911465da7c1cc646a24ac7c0814ae5d634e8de.jpg) + +![](images/75471a82147bf9acafe1c681dd3906adc4b3c01023236f39b97cb330adccabda.jpg) +Fig. 4: An example to construct instruction pairs for the movie Titanic. Given the movie synopsis and script, we utilize the developed LLMs like GPT-4 [45] and Claude-2 [3] to generate movie summaries, plot-related QA pairs, and general reasoning pairs. + +![](images/813f71ff703751aa520a4a478429818f3458d252c130cf47ef598f6f9366c557.jpg) + +![](images/4d01d043fddc5e8a8368b8dec20a40f587859395a77c4999585584e5f6b82104.jpg) + +![](images/4db7cfbd623810400fd8af925e5d2dbca6d2d171ab7188b95a9d53632cc526af.jpg) + +As illustrated in Figure 3, it mainly consists of 558K image-caption pairs from the LLaVA-filtered CC3M dataset [49] and 232K video-caption pairs sampled from the WebVid 2.5M dataset [5]. The instruction format for each modality is presented in Figure 3. In this stage, we optimize the projectors in Figure 2, while freezing the pre-trained modules like the visual encoder and text decoder. + +Instruction Tuning. To enhance the multi-modality understanding of LLMs, we build the instruction pairs from [32] and [39]. In particular, the constructed dataset mainly involves content from three sources, i.e., 40K text conversations from ShareGPT [1], 625K single- or multi-turn visual QA pairs from [19,22,25, 26,33,41-43,48,50], and 98K video QA pairs from [7]. For the instruction, we adopt different formats for text, image, and video input, as shown in Figure 3. The input prompt and answer vary with datasets. Please refer to [32] and [39] for more details. Meanwhile, the image token is randomly inserted at the beginning or end of the user input during our training. In instruction tuning, all the modules are optimized except the visual encoder. + +Long Video Tuning. To further unleash the potential for hour-long videos, we construct 15K long QA pairs, including 9K conversions in movie scenes and 6K data sampled from LongLoRA [11] for token expanding. Specifically, we utilize more than 400 long movies and corresponding scripts in MovieNet [21] to build the training set. The key components for instruction generation are visualized in Figure 4. Generally, the generated dataset includes QA pairs from three aspects: video summary, movie plot, and detail reasoning. For video summaries, we collect movie synopses to produce brief and detailed summaries for each movie using developed LLMs like GPT-4 [45]. It brings about 1K summary-level instruction pairs in total. For plot-level data, we take the entire movie synopsis as input and leverage GPT-4 [45] to generate plot-related and character-related QA pairs. These include plot understanding, description, analysis, character relationship, + +personality, and behavior. In particular, we generate 5 plot-related pairs and 5 character-related pairs for each movie, resulting in 4K plot-level QA data. As for detail-level data, we feed the long movie script into Claude-2 [3] and generate 5 plot-related reasoning pairs and 5 detail-related descriptions for each movie, which brings 4K pairs in total. With long videos and the generated pairs, we perform instruction tuning by concatenating visual tokens and subtitle tokens for each frame, as depicted in Figure 3. In this way, LLaMA-VID can well support 64K tokens with more than 3-hour video as input. + +# 4 Experiments + +In this section, we provide the experimental setup and comparisons with leading methods on several benchmarks. + +# 4.1 Experimental Setup + +Implementation Details. In this work, we instantiate the model with the pre-trained EVA-G [16] for visual encoder and QFormer [13] for text decoder by default. During training, we keep the visual encoder fixed in all stages and freeze the text decoder, as well as the LLM, in the modality alignment stage, except for the BERT module in Table 8 that is not pre-trained. Following the strategy in [32], we optimize trainable parameters with the designed data and instructions in Figure 3, running for 1 epoch in each stage. For video input, we extract frames at 1 FPS. All models are trained using $8 \times$ NVIDIA A100 GPUs. + +Datasets. In this study, we construct the training set mainly from [5,21,32,39], as illustrated in Section 3.3. Moreover, we report results on several video- and image-based benchmarks. In particular, for video input, we evaluate the zero-shot performance on the open-ended QA benchmarks like MSVD [8], MSRVTT [59], ActivityNet [7], EgoSchema [40], SSv2 (Something-Something v2) [18], and the newly-proposed generative performance benchmark [39]. As for image-based evaluation, we conduct experiments on several widely-adopted benchmarks, including GQA [22], MMB (MMBench) [36], MME [17], POPE [31], SEED [28], $\mathrm{SQA}^{\mathrm{I}}$ [37], $\mathrm{VQAT}^{\mathrm{T}}$ (TextVQA) [51], VizWiz [20], and $\mathrm{VQAv}^{\mathrm{v2}}$ (VQA V2) [19]. + +# 4.2 Main Results + +Results on Video-based Benchmarks. In Table 1, we provide a comparative evaluation of LLaMA-VID against various state-of-the-art methods across zero-shot video QA benchmarks: MSVD-QA [8], MSRVTT-QA [59], ActivityNet-QA [7], EgoSchema [40], and SSv2 [18]. Notably, the results are reported with only two tokens for each frame. It is evident that LLaMA-VID, employing Vicuna-7B and Vicuna-13B as the LLMs, consistently delivers superior performance across all datasets. On the MSVD-QA and MSRVTT-QA datasets, it + +Table 1: Comparison with leading methods on 3 zero-shot video QA datasets. We report results with 2 tokens for each frame. For fair comparisons, our model is trained without long video tuning (stage 3) in Figure 3. Res indicates image resolution. + +
MethodLLMRes.MSVD-QAMSRVTT-QAActivityNet-QA
AccScoreAccScoreAccScore
VideoLLaMA [64]Vicuna-7B22451.62.529.61.812.41.1
LLaMA-Adapter [65]LLaMA-7B22454.93.143.82.734.22.7
VideoChat [30]Vicuna-7B22456.32.845.02.526.52.2
Video-ChatGPT [39]Vicuna-7B22464.93.349.32.835.22.7
BT-Adapter [34]Vicuna-7B-67.53.757.03.245.73.2
LLaMA-VIDVicuna-7B22469.73.757.73.247.43.3
LLaMA-VIDVicuna-13B22470.03.758.93.347.53.3
+ +Table 2: Comparison with leading methods on the video generative benchmark [39]. We report results with 2 tokens for each frame. For fair comparisons, our model is trained without long video tuning (stage 3) in Figure 3. Res indicates image resolution. Correct, Detail, Context, Temporal, and Consist indicate evaluation metrics in [39]. + +
MethodLLMRes.Correct.DetailContextTemporalConsist.
VideoLLaMA [64]Vicuna-7B2241.962.182.161.821.79
LLaMA-Adapter [65]LLaMA-7B2242.032.322.301.982.15
VideoChat [30]Vicuna-7B2242.232.502.531.942.24
Video-ChatGPT [39]Vicuna-7B2242.402.522.621.982.37
BT-Adapter [34]Vicuna-7B-2.682.693.272.342.46
LLaMA-VIDVicuna-7B2242.963.003.532.462.51
LLaMA-VIDVicuna-13B2243.073.053.602.582.63
+ +Table 3: Comparison with leading methods on the zero-shot benchmark EgoSchema [40]. We report results with 2 tokens for each frame on different settings. + +
FrameFrozenBiLM [60]mPLUG-Owl [62]InternVideo [56]LLaMA-VID
1026.4%29.6%31.4%41.2%
30-20.0%31.8%41.7%
9026.9%-32.1%40.5%
180---41.7%
+ +achieves the accuracy of $69.7\%$ and $57.7\%$ with Vicuna-7B, surpassing the previous leading approach [34] with absolute gains of $2.2\%$ and $0.7\%$ , respectively. For the ActivityNet-QA dataset, LLaMA-VID attains top performance in accuracy and the highest score of 3.3. In Table 2, we also carry out experiments on the newly proposed video-based generative performance benchmark [39]. As for the long EgoSchema and SSv2 dataset, the proposed method surpasses other leading approaches in different frame settings, as presented in Table 3 and Table 4. Specifically, we evaluate the model on the SSv2 dataset similar with that in MSRVTT-QA. Our LLaMA-VID is validated to achieve the best performance across all the evaluation metrics, surpassing previous approaches by a large margin. In general, LLaMA-VID is proved to bring robust results on all benchmarks. + +Table 4: Comparisons with zero-shot methods on SSv2 [18] dataset. + +
MethodLLMRes.AccScore
Video-ChatGPT [38]Vicuna-7B22411.592.10
LLaMA-VIDVicuna-7B22420.122.14
+ +Table 5: Comparison with leading methods on 8 benchmarks. Here, we use the same training and instruction finetuning data as that in LLaVA-1.5. We report results with 1 context token and $n$ content tokens, where $n$ is kept the same with that in LLaVA-1.5, i.e., $n = (336 / 14)^2 = 576$ . Res indicates input image resolution. * and † denote the train subset is included for training and the data is not publicly available, respectively. + +
MethodLLMRes.GQAMMBMMEPOPESEED\( \mathbf{SQA}^{\dagger} \)VizWiz\( VQA^{v2} \)
InstructBLIP [13]Vicuna-7B22449.236.0--53.460.534.5-
IDEFICS-9B [23]LLaMA-7B22438.448.2----35.550.9
Qwen-VL† [4]Qwen-7B44859.3*38.2--56.367.135.278.8*
Qwen-VL-Chat† [4]Qwen-7B44857.5*60.61487.5-58.268.238.978.2*
LLaVA-1.5 [32]Vicuna-7B33662.0*64.31510.785.958.666.850.078.5*
LLaMA-VIDVicuna-7B33664.3*65.11521.486.059.968.354.279.3*
BLIP-2 [29]Vicuna-13B22441.0-1293.885.346.461.019.641.0
InstructBLIP [13]Vicuna-13B22449.5-1212.878.9-63.133.4-
Shikra [9]Vicuna-13B224-58.8-----77.4*
IDEFICS-80B [23]LLaMA-65B22445.254.5----36.060.0
LLaVA-1.5 [32]Vicuna-13B33663.3*67.71531.385.961.671.653.680.0*
LLaMA-VIDVicuna-13B33665.0*66.61542.386.062.370.054.380.0*
+ +Results on Image-based Benchmarks. As illustrated in Section 3.2, LLaMA-VID not only efficiently represents long video, but also expands the upper limit of VLMs with an additional context token. In Table 5, we perform comparisons with leading VLMs on 8 widely-adopted benchmarks. It should be noted that we maintain the same training data and image resolution as in LLaVA-1.5 [32] for fair comparisons. It is evident that LLaMA-VID outperforms other leading methods across most of benchmarks with different LLMs. In particular, with Vicuna-7B, LLaMA-VID attains the best results across all the datasets and surpasses LLaVA-1.5 with significant gains in GQA, MME, and VizWiz, where the improvement reaches up to $2.3\%$ , 10.7, and $4.2\%$ , respectively. With a larger Vicuna-13B, LLaMA-VID also outperforms other methods in 6 benchmarks and achieves top-2 in the other datasets. This demonstrates the generality of LLaMA-VID, which can be scaled up with a stronger foundation model. + +Qualitative Results. In Figure 5, we apply LLaMA-VID to different types of data, including short videos and long movies. We represent each frame with 2 tokens for videos. LLaMA-VID demonstrates various capabilities with different inputs. Given a short trailer video, LLaMA-VID summarizes the overall plot and infers the movie name, producer, and the type of movie. It shows that the proposed LLaMA-VID can well recognize the typical characters with the produced tokens. As for a 3-hour movie, the proposed model adeptly describes the storyline and demonstrates plot-related reasoning and detailed understanding. + +![](images/72a665b4f944cc1c8215ff56a24bf462941473040496baa5701c5c2718a46530.jpg) +User: Please describe this video in detail. + +![](images/8ee3bb725204f6e157a720e757acc358757d42f42151a99240501c19c0c3461b.jpg) + +![](images/fd7f68283ef81c51cf52c956375506abddf2e4a5c95f9e658b813886db30e4ff.jpg) +3-hour VideoAvatar + +![](images/734c9af289904b0876a4ea61cfea6e1ded8258f3424ea95681f14d03d6a2c762.jpg) +User: Summarize the plot of this movie. + +![](images/bdec7b6c096e13dffb02b5a47ab1d8e1bd53886907c113822141ae1c715de62c.jpg) + +![](images/77b1b3cf4b8ffccce0291d0c13432ce2f4503730d2118dff6cd9adc708585953.jpg) + +![](images/d81eb905ae1ee06d782c52e2f95e2ed44489a292b3915b8bc612ad94bd36283f.jpg) + +Assistant: The video starts with a carto on character standing in front of a castle, which is surrounded by a beautiful lan descape. The character is holding a bucket and appears to be observing the surroundings. The scene then transitions to a t train, where the character is seen sitting inside the train. The train is surrounded by lush greenery and appears to be mov ing through a scenic route.... + +![](images/e050285158efd78ed3d3d8396ffab588de67a13fa8fe849fcfcb7d0406126027.jpg) + +Assistant:Avatar follows Jake Sully, a disabled former Marine who is recruited by the RDA to work on Pandora, a moon rich in a valuable mineral called unobtanium. Jake's twin brother, Tommy, died in a helmet-cam accident, and Jake is offered his place in the Avatar Program, which allows humans to link with the native Na' vi, a humanoid species. Jake is assigned to the avatar of his deceased brother.... + +User: What's the name of the movie described in this video? + +![](images/8fb7ed9008fb0fc006e34c39774d688179bad2dbfaa53a23949bf75d2f0f1065.jpg) + +![](images/f6d9d45df6c90a54aedd838c34e4ac056cd2e79da034450845f24c0a1c459aca.jpg) + +![](images/9add4a71043559c97faef4fd0b846eab03a20bdbc35205b8ba24eb418d99a528.jpg) + +Assistant: The movie described in this video is "Zootopia." + +![](images/ecfc7d758e0961100b562c8c9cfe8d12ca884d4463bbdfc8f7928d4e7212910b.jpg) +User: How human work on Pandora? + +Assistant: Humans work on Pandora by using avatars, which are grown from human DNA mixed with the DNA of the native Na'vi, to interact with the environment and the Na'vi. + +User: Who produce the movie? + +![](images/8ae39fbd05ec4df51649fca78a50950ff4cfe12dcb215ec2440482673f4cc56a.jpg) + +User: Why Jake Sully decide to fight againt humanity with the Na'vi? + +![](images/ef43d0dd38c7d488cc75d3a24093f39d1d94417afbcdf9f7baa4eef1a2fb0be7.jpg) + +![](images/b10e02a5fc1e5b2c2b28de1f37525f77a539f2dabbeaf4b9f5c76c9a89eeee6f.jpg) + +Assistant: The movie is produced by Walt Disney Pictures. + +![](images/b40b3eba5c1f3b56ff1d4399214042fe8fc836c591d3176d58ccdcbc41a1197a.jpg) + +Assistant: Jake Sully decides to fight a gainst humanity with the Na'vi after he falls in love with the Na'vi and their way of life, and after he realizes that the human military is destroying their home and culture. + +User: What's the type of this movie? + +![](images/287e9de54548bfde8d7bae7dff2287c3eb4372defa68ef5b5d0a5ae21d59c3dd.jpg) + +![](images/429f15612f6aee68abd959212c1f33a3cdcbcfde3fe754d04f0f37ca20a2ceed.jpg) +Fig. 5: Examples of LLaMA-VID with short video and hour-long video. + +Assistant: The movie is a Disney animat ed film. + +# 4.3 Component-wise Analysis + +In this subsection, we conduct ablation studies with input resolution 224 and 2 tokens for each image by default. + +Generated Token Types. As illustrated in Figure 2, each image is represented with a context token and a content token in LLMs. To validate the effectiveness of each part, we conduct experiments with different types of tokens in Table 6. Without the context token, the compressed content token, which encodes each image with 1 token, cannot adjust to input instructions, leading to subpar performance. Compared with a single content token, the instruction-guided context token results in significant gains across all datasets with only 1 token. With both tokens for each image, the model achieves the best performance across all image- and video-based benchmarks. It shows that both instruction cues in the context token and the image content itself in the content token are important. + +Table 6: Comparison with different token types. We report results with 1 context token (if exists) and 1 content token on image-based and video-based benchmarks (MSVD). + +
contextcontentGQAPOPE\( \mathbf{SQA^I} \)\( VQA^T \)MSVD
✘✓53.380.966.146.565.9
✓✘54.382.467.748.369.3
✓✓55.583.168.849.069.7
+ +Table 7: Comparison with different token numbers. We report results with various numbers of context token and content token on image-based benchmarks for efficiency. + +
contextcontentGQAPOPE\( \mathbf{SQA^I} \)\( VQA^T \)
025661.985.567.553.0
125663.086.667.753.8
16460.885.168.752.3
11658.283.167.450.8
1456.283.568.749.1
1155.583.168.849.0
+ +Generated Token Numbers. In Table 7, we conduct experiments with different numbers of tokens for further investigation. With an image size $224 \times 224$ , we set up experiments with $n$ content tokens, where $n = (224 / 14)^{2} = 256$ for uncompressed settings in the first two rows. The results clearly show that the context token consistently improves performance across different benchmarks with only 1 extra token. When we compress the content token to $1 / 4$ with $n = 64$ , the performance drops about $1\%$ to $2\%$ but increases $1\%$ in $\mathrm{SQA}^{\mathrm{I}}$ . Considering the extra efficient setting for hour-long videos, we compress the content token to $1 / 256$ with $n = 1$ by default. Compared to the original setting without context token, we can reduce the computational cost to $1 / 128$ with about $2\% - 6\%$ performance drop, which is generally acceptable. The linear increase in performance presents significant potential for token compression. For instance, we can dynamically compress the content token to different numbers according to resource budget and content importance. Interestingly, the model achieves peak performance in $\mathrm{SQA}^{\mathrm{I}}$ with only 2 tokens. This could be attributed to the fact that problems in ScienceQA [37] focus more on visual-based reasoning rather than image details. As demonstrated in Tables 1 and 2, with only 2 tokens for each image, LLaMA-VID still outperforms all previous work in different video-based benchmarks. This makes it feasible to enable LLMs for hour-long video processing. + +Text Decoder. As depicted in Figure 2, the text decoder plays an essential role in producing instruction-guided context cues. Here, we further perform comparisons with different text decoders in Table 8. We mainly instantiate the text decoder with two types of modules, namely BERT [14] and QFormer [13]. For BERT, we randomly initialize it as a cross-modality decoder and only retain the first two layers. As for QFormer, we utilize the pre-trained modules and fix them + +Table 8: Comparison with different text decoders. We report results with 1 context token (if exists) and 1 content token on image-based benchmarks for efficiency. + +
textGQAPOPE\( \mathbf{SQA^I} \)\( \mathbf{VQA^T} \)
-53.380.966.146.5
BERT54.180.867.948.1
QFormer55.583.168.849.0
+ +Table 9: Average module latency of video frame on a A100 GPU. Encoder, Content, Decoder, and Context denote the latency for text-free visual encoder and content token, text-related text decoder and context token generation, respectively. + +
textText-freeText-relatedTotal (ms)
EncoderContentDecoderContext
BERT4.890.010.250.025.17
QFormer4.870.011.080.025.98
+ +for modality alignment. Even with a simple 2-layer BERT, as shown in Table 8, the generated context token achieves significant gains in most of benchmarks. This proves the effectiveness of the paradigm for context token generation. With a pre-trained text decoder like QFormer, the model can be further enhanced and attains peak performance in all datasets with $2.2\%$ to $2.7\%$ significant gain. + +Latency Analysis. To further analyse the module latency, we keep the original setting in Table 8 and report average module latency over 100 videos in Table 9. Here, we divide the modules for generation into two types, namely text-free and text-related. In particular, we generate visual tokens one time for each frame with text-free modules. And text-related modules are required to produce different context tokens according to user inputs. It is clear that the text-free tokens account for the most computation, while the text-related modules only cost $5\%$ to $18.3\%$ for each user instruction. That means the total cost will not increase greatly given variant user questions because of the efficient text-related modules. + +Response in Context Attention. To more vividly explore the context attention, we visualize the high response areas with the top 20 scores in Figure 6. Specifically, we draw the normalized heatmap for the first four queries in $\mathbf{Q}_t$ before applying the Softmax function, which is used to formulate context token in Equation 1. As shown in Figure 6, the text-guided query $\mathbf{Q}_t$ effectively focuses on important areas relevant to the input questions. For example, in the second row, when inquiring whether the image depicts a fishing village, the query $\mathbf{Q}_t$ focuses more on buildings along the river and a seagull. These are all typical characteristics to distinguish a fishing village in common sense. Other examples also confirm that the designed context attention successfully achieves its goal of formulating the context token under instruction guidance. + +![](images/9b55caa5a4cc5c34b647ac1b059e14c923a33db2c50956c2cf8c7da0eec0fff7.jpg) +Input + +![](images/a950cf1de1a709c4d6d996327a5f19b7fb4ab06775a37a6b6c8b3d95b75105f9.jpg) +Query 1 + +![](images/fc726282f4090ce1c536c1dcf5cad305881be27b371fc37c3d3c92898c39da40.jpg) +Query 2 + +![](images/1b25ca521e7ade249fee96d8831230fa7aec15eb136f840c08db05105a628e02.jpg) +Query 3 + +![](images/f8d85417b652785ef5650a64562b7076d43d452be66eb6e5451e615020cff279.jpg) +Query 4 + +![](images/6499112d0933b04822419699d45b4cccf51b8b29c975d2d4467da848846ef8d4.jpg) +Input + +![](images/fb5d34a3fe3c591ab99e4e3e0c568403a37785ecf775d331ad906b9bede29fe1.jpg) +Question: Are the elephants fighting? +Query 1 + +![](images/032f9634bb7e2c3bf8e3e8ad8571c44938c864467212772bb8c6e7f5a0a4d9d4.jpg) +Query 2 + +![](images/2dc0580fd2dffdb9dcbb92165e32c82db59c0bb3e6251519d7d94e3caa5cd1a0.jpg) +Query 3 + +![](images/41507cc8848fab782616d302aebb8e79163fe5ba25fc3b34e16025f4698a5a4a.jpg) +Query 4 + +![](images/3d0fd57ad55436340e87cfe02370b24f4ec48bde01cfc2728062f6acb966a407.jpg) +Input +Fig. 6: High response areas with top scores to input question in Equation 1. We present the response of first four queries in $\mathbf{Q}_t$ . Images are sampled from VQA V2 test-dev set. + +![](images/9dfe53c6317b6802cf82358d866028009350c3050522da42907dc85ec0290f62.jpg) +Question: Is this a fishing village? +Query 1 +Question: What tall animals are displayed in the background? + +![](images/bbc34d68c49ec350d0f17fce2a6552fea86bf83ea4cb4b13746b11af6dcd8c5a.jpg) +Query 2 + +![](images/c8ca05f8bef6da9eec0e0b50a3ef145e2fdc79664f5730d70d1c0c3d5ea895a7.jpg) +Query 3 + +![](images/924c4718926950fe8014999ce472fbd48fde2a71e3e93bfa0dbe187ec57acbfc.jpg) +Query 4 + +# 5 Discussion and Conclusion + +We have introduced LLaMA-VID, a simple yet effective token generation approach for VLMs. The central concept behind LLaMA-VID is to represent an image with the context token and the content token. In particular, the context token is generated according to input instructions, and the content token is produced based on the image content. Depending on the budget, the content token can be compressed to one token or expressed without compression. It allows us to represent a single image with preserved details and efficiently encode each video frame with only two tokens. Moreover, we have constructed an instruction dataset for hour-long video understanding. Our experiments on several video-and image-based benchmarks prove the superiority of our method. We hope that LLaMA-VID can serve as a strong benchmark for efficient visual representation. + +There still exists certain limitations in the current method. Although LLaMAVID succeeds to represent hour-long videos in LLMs, it focuses more on temporal consistency and may lose spatial details in each frame. In the future work, we plan to maintain dense spatial information for each frame outside the LLM and conduct the visual token fusion within LLMs to avoid such detail loss. + +Acknowledgement. This work was supported in part by the Research Grants Council under the Areas of Excellence scheme grant AoE/E-601/22-R. + +# References + +1. Sharegpt. https://sharegpt.com/ (2023) 7 +2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: NeurIPS (2022) 4 +3. Anthropic: Claude 2. https://www.anthropic.com/index/claude-2 (2023) 7, 8 +4. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv:2308.12966 (2023) 10 +5. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: ICCV (2021) 7, 8 +6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS (2020) 3 +7. Caba Heilbron, F., Escorcia, V., Ghanem, B., Carlos Niebles, J.: Activitynet: A large-scale video benchmark for human activity understanding. In: CVPR (2015) 7, 8 +8. Chen, D., Dolan, W.B.: Collecting highly parallel data for paraphrase evaluation. In: ACL (2011) 8 +9. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv:2306.15195 (2023) 10 +10. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollár, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv:1504.00325 (2015) 3 +11. Chen, Y., Qian, S., Tang, H., Lai, X., Liu, Z., Han, S., Jia, J.: Longlora: Efficient fine-tuning of long-context large language models. arXiv:2309.12307 (2023) 7 +12. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ chatgpt quality. https://lmsys.org/blog/2023-03-30-vicuna/ (2023) 3, 6 +13. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv:2305.06500 (2023) 1, 2, 4, 5, 6, 8, 10, 12 +14. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv:1810.04805 (2018) 2, 3, 5, 12 +15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) 2, 5 +16. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. In: CVPR (2023) 5, 8 +17. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv:2306.13394 (2023) 8 +18. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The" something something" video database for learning and evaluating visual common sense. In: ICCV (2017) 8, 10 + +19. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: CVPR (2017) 3, 7, 8 +20. Gurari, D., Li, Q., Stangl, A.J., Guo, A., Lin, C., Grauman, K., Luo, J., Bigham, J.P.: Vizwiz grand challenge: Answering visual questions from blind people. In: CVPR (2018) 8 +21. Huang, Q., Xiong, Y., Rao, A., Wang, J., Lin, D.: Movienet: A holistic dataset for movie understanding. In: ECCV (2020) 7, 8 +22. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: CVPR (2019) 7, 8 +23. IDEFICS: Introducing idefics: An open reproduction of state-of-the-art visual language model. https://huggingface.co/blog/idefics (2023) 10 +24. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021) 3 +25. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: Referitgame: Referring to objects in photographs of natural scenes. In: EMNLP (2014) 7 +26. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV (2017) 7 +27. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. arXiv:2308.00692 (2023) 3 +28. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv:2307.16125 (2023) 8 +29. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv:2301.12597 (2023) 1, 2, 4, 10 +30. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chat-centric video understanding. arXiv:2305.06355 (2023) 1, 4, 9 +31. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv:2305.10355 (2023) 8 +32. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv:2310.03744 (2023) 6, 7, 8, 10 +33. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeruIPS (2023) 1, 2, 4, 6, 7 +34. Liu, R., Li, C., Ge, Y., Shan, Y., Li, T.H., Li, G.: One for all: Video conversation is feasible without video instruction tuning. arXiv:2309.15785 (2023) 9 +35. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. arXiv:1907.11692 (2019) 3 +36. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv:2307.06281 (2023) 8 +37. Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. In: NeurIPS (2022) 3, 8, 12 +38. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability. arXiv:2306.07207 (2023) 1 + +39. Maaz, M., Rasheed, H., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv:2306.05424 (2023) 1, 4, 7, 8, 9 +40. Mangalam, K., Akshulakov, R., Malik, J.: Egoschema: A diagnostic benchmark for very long-form video language understanding. NeurIPS (2024) 8, 9 +41. Mao, J., Huang, J., Toshev, A., Camburu, O., Yuille, A.L., Murphy, K.: Generation and comprehension of unambiguous object descriptions. In: CVPR (2016) 7 +42. Mao, J., Huang, J., Toshev, A., Camburu, O., Yuille, A.L., Murphy, K.: Generation and comprehension of unambiguous object descriptions. In: CVPR (2016) 7 +43. Mishra, A., Shekhar, S., Singh, A.K., Chakraborty, A.: Ocr-vqa: Visual question answering by reading text in images. In: ICDAR (2019) 7 +44. OpenAI: Chatgpt. https://openai.com/blog/chatgpt/ (2023) 1, 3 +45. OpenAI: Gpt-4 technical report. arXiv:2303.08774 (2023) 1, 3, 7 +46. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. In: NeurIPS (2022) 3 +47. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 3, 5 +48. Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: A-okvqa: A benchmark for visual question answering using world knowledge. In: ECCV (2022) 7 +49. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: ACL (2018) 7 +50. Sidorov, O., Hu, R., Rohrbach, M., Singh, A.: Textcaps: a dataset for image captioning with reading comprehension. In: ECCV (2020) 7 +51. Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards vqa models that can read. In: CVPR (2019) 8 +52. Song, E., Chai, W., Wang, G., Zhang, Y., Zhou, H., Wu, F., Chi, H., Guo, X., Ye, T., Zhang, Y., et al.: Moviechat: From dense token to sparse memory for long video understanding. In: CVPR (2024) 1, 4 +53. Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca (2023) 3, 6 +54. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models. arXiv:2302.13971 (2023) 1, 3, 6 +55. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 3 +56. Wang, Y., Li, K., Li, Y., He, Y., Huang, B., Zhao, Z., Zhang, H., Xu, J., Liu, Y., Wang, Z., et al.: Intervideo: General video foundation models via generative and discriminative learning. arXiv:2212.03191 (2022) 9 +57. Wei, J., Bosma, M., Zhao, V.Y., Guu, K., Yu, A.W., Lester, B., Du, N., Dai, A.M., Le, Q.V.: Finetuned language models are zero-shot learners. arXiv:2109.01652 (2021) 3 +58. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv:2303.04671 (2023) 3 +59. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: CVPR (2016) 3, 8 + +60. Yang, A., Miech, A., Sivic, J., Laptev, I., Schmid, C.: Zero-shot video question answering via frozen bidirectional language models. In: NeurIPS (2022) 9 +61. Yang, R., Song, L., Li, Y., Zhao, S., Ge, Y., Li, X., Shan, Y.: Gpt4tools: Teaching large language model to use tools via self-instruction. arXiv:2305.18752 (2023) 3 +62. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv:2304.14178 (2023) 9 +63. Zhang, C., Lu, T., Islam, M.M., Wang, Z., Yu, S., Bansal, M., Bertasius, G.: A simple lvm framework for long-range video question-answering. arXiv:2312.17235 (2023) 4 +64. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv:2306.02858 (2023) 1, 4, 9 +65. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv:2303.16199 (2023) 9 +66. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv:2205.01068 (2022) 1, 3 +67. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv:2304.10592 (2023) 4 \ No newline at end of file diff --git a/2024/LLaMA-VID_ An Image is Worth 2 Tokens in Large Language Models/images.zip b/2024/LLaMA-VID_ An Image is Worth 2 Tokens in Large Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f03a0b327f501f6b85c025943e5b980c6dcfe1fd --- /dev/null +++ b/2024/LLaMA-VID_ An Image is Worth 2 Tokens in Large Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9481d91b45a5a13f19551bc984db0d0c3b44aaf2c344ad9d89624b22c8f3b7a +size 581494 diff --git a/2024/LLaMA-VID_ An Image is Worth 2 Tokens in Large Language Models/layout.json b/2024/LLaMA-VID_ An Image is Worth 2 Tokens in Large Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..af624e15182d96aa35c177eb8a0bb2f449dfb3e0 --- /dev/null +++ b/2024/LLaMA-VID_ An Image is Worth 2 Tokens in Large Language Models/layout.json @@ -0,0 +1,11151 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 157, + 111, + 457, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 111, + 457, + 148 + ], + "spans": [ + { + "bbox": [ + 157, + 111, + 457, + 148 + ], + "type": "text", + "content": "LLaMA-VID: An Image is Worth 2 Tokens in Large Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 201, + 167, + 412, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 167, + 412, + 180 + ], + "spans": [ + { + "bbox": [ + 201, + 167, + 412, + 180 + ], + "type": "text", + "content": "Yanwei Li" + }, + { + "bbox": [ + 201, + 167, + 412, + 180 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 201, + 167, + 412, + 180 + ], + "type": "text", + "content": " Chengyao Wang" + }, + { + "bbox": [ + 201, + 167, + 412, + 180 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 201, + 167, + 412, + 180 + ], + "type": "text", + "content": " Jiaya Jia" + }, + { + "bbox": [ + 201, + 167, + 412, + 180 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 260, + 189, + 353, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 189, + 353, + 201 + ], + "spans": [ + { + "bbox": [ + 260, + 189, + 353, + 201 + ], + "type": "text", + "content": "CUHK" + }, + { + "bbox": [ + 260, + 189, + 353, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 260, + 189, + 353, + 201 + ], + "type": "text", + "content": " SmartMore" + }, + { + "bbox": [ + 260, + 189, + 353, + 201 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 232, + 455, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 232, + 455, + 397 + ], + "spans": [ + { + "bbox": [ + 160, + 232, + 455, + 397 + ], + "type": "text", + "content": "Abstract. In this work, we present a novel method to tackle the token generation challenge in Vision Language Models (VLMs) for video and image understanding, called LLaMA-VID. Current VLMs, while proficient in tasks like image captioning and visual question answering, face computational burdens when processing long videos due to the excessive visual tokens. LLaMA-VID addresses this issue by representing each frame with two distinct tokens, namely context token and content token. The context token encodes the overall image context based on user input, whereas the content token encapsulates visual cues in each frame. This dual-token strategy significantly reduces the overload of long videos while preserving critical information. Generally, LLaMA-VID empowers existing frameworks to support hour-long videos and pushes their upper limit with an extra context token. It is demonstrated to surpass previous methods on most of video- or image-based benchmarks. Code and models are available at https://github.com/dvlab-research/LLaMA-VID." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 417, + 230, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 417, + 230, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 417, + 230, + 430 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 442, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 442, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 442, + 482, + 550 + ], + "type": "text", + "content": "Large Language Models (LLMs) [44, 54, 66], through their capacity to generate contextually accurate responses, have significantly advanced the field of AI. Drawing from the strengths of LLMs, Vision Language Models (VLMs) [13,33,45] have been developed to extend these capabilities to visual data, demonstrating their adeptness in tasks like image captioning and visual question answering. However, a substantial challenge emerges in the context of long video, where an excessive number of tokens are required to represent consecutive frames. The computational demands escalate with the video length, thereby constraining the practical application of VLMs for extensive videos." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 551, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 551, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 551, + 482, + 647 + ], + "type": "text", + "content": "Recently, several approaches have been proposed to handle videos, moving beyond image-only VLMs. These methods aim to alleviate the token issue by utilizing representative queries [30, 64] or applying temporal compression [38, 39, 52]. Despite these efforts, the challenge of long videos remains unresolved. The primary obstacle stems from the excessive number of tokens required for each video frame. For instance, models like BLIP [13, 29] and LLaVA [33] require 32 and over 256 tokens respectively for a single image. A video containing 10K frames would thus necessitate over 320K tokens, exceeding the capacity of current VLMs." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 654, + 219, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 654, + 219, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 654, + 219, + 665 + ], + "type": "text", + "content": "* equal contribution" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 305, + 256 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 305, + 256 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 305, + 256 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 305, + 256 + ], + "type": "image", + "image_path": "f29a384f07e4c9034fdc8fc584dea006ebd3d391385baac25d928b725800a86a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 258, + 279, + 266 + ], + "lines": [ + { + "bbox": [ + 176, + 258, + 279, + 266 + ], + "spans": [ + { + "bbox": [ + 176, + 258, + 279, + 266 + ], + "type": "text", + "content": "(a) Video-based benchmarks" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 323, + 113, + 465, + 257 + ], + "blocks": [ + { + "bbox": [ + 323, + 113, + 465, + 257 + ], + "lines": [ + { + "bbox": [ + 323, + 113, + 465, + 257 + ], + "spans": [ + { + "bbox": [ + 323, + 113, + 465, + 257 + ], + "type": "image", + "image_path": "31ed8192d04c010148629d519032cf3f54fca4eb3ee276cc14e594824bef8d96.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 257, + 446, + 266 + ], + "lines": [ + { + "bbox": [ + 342, + 257, + 446, + 266 + ], + "spans": [ + { + "bbox": [ + 342, + 257, + 446, + 266 + ], + "type": "text", + "content": "(b) Image-based benchmarks" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 276, + 482, + 309 + ], + "lines": [ + { + "bbox": [ + 131, + 276, + 482, + 309 + ], + "spans": [ + { + "bbox": [ + 131, + 276, + 482, + 309 + ], + "type": "text", + "content": "Fig. 1: The proposed LLaMA-VID achieves leading performance on most of benchmarks with 7B LLMs. The video-based and image-based benchmarks are noted in blue and purple color, respectively. Please refer to Tables 1, 2, and 5 for more details." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 343, + 480, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 480, + 380 + ], + "type": "text", + "content": "Furthermore, simple temporal compression can significantly damage the representation over long-term intervals. This drawback hampers their performance, thereby underscoring the need for a robust solution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 384, + 482, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 504 + ], + "type": "text", + "content": "In this work, we present LLaMA-VID, a novel approach to effectively manage the token generation issue in long videos. Our core idea is to represent each video frame with two distinct tokens: context token and content token. The context token is designed to encode the overall context of the image based on user input, which efficiently condenses the broader picture into a single token. Simultaneously, the content token captures finer aspects of each frame. According to computational constraints, the length of content token can be extended to include more details, e.g., 1 token/frame for video input and beyond 256 token/frame for single image. In this way, the overload of long videos can be significantly reduced without sacrificing critical information." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "In particular, our method employs a dual-token generation strategy that is both efficient and effective. For each frame, we first extract image features using a pre-trained vision transformer [15], akin to other VLMs [13, 33]. The key question is how to generate the context-related token according to user instructions. We provide the solution by leveraging the cross-modality design [14,29] for instruction-guided queries, which carry the interactive intention from users. For context token, these queries interact with previously generated image features in the designed attention module, termed as context attention. To generate content token, the image features are average pooled to formulate tokens that adapt to different settings. For instance, global pooling is adopted to maintain efficiency for video input while details are preserved with more tokens for single image input. The context and content tokens are subsequently projected to the space of LLMs with simple linear layers for final prediction. Furthermore, to better sup" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 101 + ], + "type": "text", + "content": "L. Yanwei et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "port hour-long videos in VLMs, we construct an instruction dataset that contains 9K movie-level conversations for plot reasoning and detail understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 199 + ], + "type": "text", + "content": "Generally, LLaMA-VID can be distinguished from two aspects. On one hand, with the dual-token paradigm, each frame can be efficiently encoded with only two tokens, which empowers existing LLMs to support long videos. On the other hand, the context token aggregates the most informative feature of each image, which further extends the upper limit of VLMs with an extra token." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 200, + 481, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 481, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 481, + 271 + ], + "type": "text", + "content": "The overall framework, dubbed LLaMA-VID, can be easily instantiated with various decoders and LLMs, as elaborated in Section 3. Extensive empirical studies are conducted in Section 4 to reveal the effectiveness of each component. Remarkably, our model can complete training within 2 days on a single machine with " + }, + { + "bbox": [ + 130, + 200, + 481, + 271 + ], + "type": "inline_equation", + "content": "8 \\times \\mathrm{A}100" + }, + { + "bbox": [ + 130, + 200, + 481, + 271 + ], + "type": "text", + "content": " GPUs, and it outperforms previous leading methods on most of video- and image-based benchmarks, as shown in Figure 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 289, + 237, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 237, + 303 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 237, + 303 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 315, + 479, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 315, + 479, + 339 + ], + "spans": [ + { + "bbox": [ + 131, + 315, + 479, + 339 + ], + "type": "text", + "content": "In this section, we first review large language models and delve into recent advances in vision language models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 357, + 281, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 357, + 281, + 369 + ], + "spans": [ + { + "bbox": [ + 131, + 357, + 281, + 369 + ], + "type": "text", + "content": "2.1 Large Language Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 376, + 481, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 481, + 568 + ], + "type": "text", + "content": "The field of Natural Language Processing (NLP) has witnessed tremendous advancements with the evolution of LLMs. Transformer [55] marked a pivotal milestone, with subsequent language models [14, 35, 66] demonstrating remarkable capabilities. GPT [6] revolutionized this field by utilizing generative pre-trained transformers for auto-regressive prediction, which is proved to be a potent language modeling paradigm. Recent groundbreaking works, such as ChatGPT [44], GPT-4 [45], and LLaMA [54], have pushed the boundaries even further. Trained on vast amounts of text data, these models exhibit exceptional capabilities in complex linguistic tasks. To leverage the potential of pre-trained LLMs, instruction tuning [46, 57] is a crucial component for high-quality output. This strategy is widely adopted in open-source models like Alpaca [53] and Vicuna [12], which improve over LLaMA [54] using specially designed instruction pairs. There are also researches [58, 61] that utilize the reasoning ability of LLMs and invoke pre-defined tools for visual applications. Different from them, we collect multimodality instruction data that contains text, images, and videos in this work, which is employed to empower LLMs for long video processing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 586, + 284, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 284, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 284, + 597 + ], + "type": "text", + "content": "2.2 Vision Language Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "The advancements in computer vision and NLP have led to the emergence of vision-language models (VLMs) that integrate vision models with language models for cross-modality understanding [10,59] and reasoning [19,27,37]. Pioneering large-scale VLMs like CLIP [47] and ALIGN [24] have extended language models to vision-language tasks. The recent progress has seen an increasing focus" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 393, + 90, + 447, + 101 + ], + "type": "text", + "content": "LLaMA-VID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 479, + 220 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 479, + 220 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 479, + 220 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 479, + 220 + ], + "type": "image", + "image_path": "3386a0651921af16022c75a39b6ad5c4586f118f8c530a9f1590adf29a9a5540.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 228, + 482, + 328 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 328 + ], + "type": "text", + "content": "Fig. 2: The framework of LLaMA-VID. With user directive, it operates by taking either a single image or video frames as input, and generates responses from LLM. The process initiates with a visual encoder that transforms input frames into the visual embedding. Then, the text decoder produces text queries based on the user input. In context attention, the text query aggregates text-related visual cues. For efficiency, an option is provided to downsample the visual embedding to various token sizes, or even to a single token. The text-guided context token and the visually-enriched content token are then formulated using a linear projector to represent each frame at time " + }, + { + "bbox": [ + 130, + 228, + 482, + 328 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 228, + 482, + 328 + ], + "type": "text", + "content": ". Finally, the LLM takes the user directive and all visual tokens as input and gives responses." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 354, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 547 + ], + "type": "text", + "content": "on leveraging the power of LLMs. Notably, Flamingo [2] and BLIP-2 [29] utilize web-scale image-text pairs for cross-modality alignment, thereby enhancing learning performance. To further exploit the potential of such pre-trained models, InstructBLIP [13] and MiniGPT-4 [67] construct high-quality instruction pairs based on BLIP-2 and achieve superior results. Simultaneously, LLaVA [33] employs a simple linear projector with a few learnable parameters to align the image and text space of LLaMA. Given the tailored instruction data, this straightforward approach demonstrates strong capabilities. To support video understanding in LLMs, several studies [30,52,64] attempt to utilize BLIP-2 for video embedding or text-only caption [63] extraction, while Video-ChatGPT [39] proposes spatial and temporal pooling for video features. However, given the substantial number of tokens required for each frame, LLMs encounter significant challenges when processing extensive video sequences. It prevents previous work from representing long video sequences that exceed a duration of one hour in LLMs. To solve the issue, we propose to efficiently encode each frame with only 2 tokens, which supports long video understanding in existing LLMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 572, + 231, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 572, + 231, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 572, + 231, + 585 + ], + "type": "text", + "content": "3 LLaMA-VID" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "The framework of LLaMA-VID is conceptually simple: encoder and decoder are adopted to produce visual embedding and text-guided features, respectively; context token and content token are transformed with the tailored token generation strategy; instruction tuning is designed to unleash the potential of LLMs for image and video." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "type": "text", + "content": "L. Yanwei et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 270, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 270, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 270, + 127 + ], + "type": "text", + "content": "3.1 Encoder and Decoder" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "text", + "content": "The proposed LLaMA-VID can be utilized to interact with single image or long videos. For clarity, we assume the input image is captured from a video sequence, as presented in Figure 2. Given a video frame " + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_t \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "text", + "content": ", a transformer-based visual encoder is first employed to produce the visual embedding " + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "inline_equation", + "content": "N = H / p \\times W / p" + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "text", + "content": " indicate the number of image patches and embedding channels, respectively. The patch size " + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "text", + "content": " is typically set to 14 for ViT-based backbones [15, 16, 47]. Meanwhile, we take the user instruction as input and generate the text-guided query " + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t \\in \\mathbb{R}^{M \\times C}" + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "text", + "content": " with the produced " + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t" + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "text", + "content": " denotes the number of queries. As depicted in Figure 2, this cross-modality interaction predominantly occurs in the text decoder, which can be easily instantiated with BERT [14] or QFormer [13], as compared in Table 8. In this way, the text query " + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t" + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "text", + "content": " contains highlighted visual cues that are most related to the user instruction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 318, + 251, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 251, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 251, + 329 + ], + "type": "text", + "content": "3.2 Token Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 481, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 481, + 403 + ], + "type": "text", + "content": "With the text query " + }, + { + "bbox": [ + 130, + 342, + 481, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t" + }, + { + "bbox": [ + 130, + 342, + 481, + 403 + ], + "type": "text", + "content": " and visual embedding " + }, + { + "bbox": [ + 130, + 342, + 481, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t" + }, + { + "bbox": [ + 130, + 342, + 481, + 403 + ], + "type": "text", + "content": ", we can easily generate representative tokens for LLMs. Specifically, context attention is designed to aggregate text-related visual features and condense them to a single context token. As shown in Figure 2, it takes " + }, + { + "bbox": [ + 130, + 342, + 481, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t" + }, + { + "bbox": [ + 130, + 342, + 481, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 481, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t" + }, + { + "bbox": [ + 130, + 342, + 481, + 403 + ], + "type": "text", + "content": " as input and formulates the context-related embedding " + }, + { + "bbox": [ + 130, + 342, + 481, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_t \\in \\mathbb{R}^{1 \\times C}" + }, + { + "bbox": [ + 130, + 342, + 481, + 403 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 209, + 416, + 481, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 416, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 209, + 416, + 481, + 430 + ], + "type": "interline_equation", + "content": "\\mathbf {E} _ {t} = \\operatorname {M e a n} \\left(\\operatorname {S o f t m a x} \\left(\\left(\\mathbf {Q} _ {t} \\times \\mathbf {X} _ {t} ^ {\\top}\\right) / \\sqrt {C}\\right) \\times \\mathbf {X} _ {t}\\right), \\tag {1}", + "image_path": "1c0ee028d751d58fb23fdea13b928d6d6ed638e487e58b6587b931ae6468b51f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": "where the Softmax function and Mean operation are conducted along the " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": " dimensions, respectively. Unlike QFormer [13] that adopts 32 visual queries as LLMs tokens, we only utilize the text query " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": " to aggregate the visual features with high-response scores to input instructions. As a result, the most crucial visual cues related to user input are efficiently preserved in the condensed embedding " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_t" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": ". The effectiveness of this context-related token generation is demonstrated in Table 6 and Figure 6. Subsequently, a linear projector is utilized to transform the embedding " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_t" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": " into the context token " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_t^T \\in \\mathbb{R}^{1 \\times C}" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": ", which aligns with the language space of LLMs. Meanwhile, we employ an adaptive pooling strategy for the visual embedding according to computational constraints to produce the content token " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_t^V \\in \\mathbb{R}^{n \\times C}" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "n \\in [1, N]" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": ". For instance, we maintain the original resolution of visual embedding " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": " when input single image, while we downsample " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": " to 1 token for long videos. This approach significantly reduces the overload of LLMs for each frame, thereby supporting hour-long videos effectively. Finally, the generated context token " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_t^T" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": " and the content token " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_t^V" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": " are concatenated to represent the frame at time " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": ". Along with frames at other timestamps, the entire video sequence is translated into the language space in token format, which is then used to generate responses from LLMs. The whole process is summarized in Algorithm 1." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 393, + 90, + 447, + 101 + ], + "type": "text", + "content": "LLaMA-VID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 136, + 465, + 316 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 356, + 128 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 356, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 356, + 128 + ], + "type": "text", + "content": "Algorithm 1 Pseudo Code for Token Generation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 136, + 465, + 316 + ], + "lines": [ + { + "bbox": [ + 132, + 136, + 465, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 136, + 465, + 316 + ], + "type": "text", + "content": "B: batch size; C: channel size; n: content shape\n# M: query length; N: shape of flatten image pacthes;\n# text_q: text query in shape (B, M, C)\n# vis_embedding: visual embedding in shape (B, N, C)\n# Key part 1: calculate context-related embedding\nctx_embedding = text_q @ vis_embedding.transpose(-1,-2)\nctx_embedding = ctx_embedding / (vis_embedding.shape[-1]**0.5)\nctx_embedding = (ctx_embedding-largest(-1)@vis_embedding).mean(1)\nctx_embedding = self.batchproj(ctx_embedding[:,None])\n# Key part 2: calculate visual embedding\ncur_shape = int(vis_embedding.shape[1]**0.5)\nvis_embedding = vis_embedding.reshape(B, cur_shape, -1, C)\nvis_embedding = F AVG_pool2d(vis_embedding.permute(0,3,1,2), kernel_size=cur_shape//n,\nstride=cur_shape//n)\nvis_embedding = vis_embedding.permute(0,2,3,1).flatten(1,2)\nvis_embedding = self.visproj(vis_embedding)\n# concat token in shape (B, n+1, C), n in [1,N]\nfinal_token = torch.cat([ctx_embedding, vis_embedding], dim=1)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + }, + { + "bbox": [ + 137, + 510, + 206, + 518 + ], + "lines": [ + { + "bbox": [ + 137, + 510, + 206, + 518 + ], + "spans": [ + { + "bbox": [ + 137, + 510, + 206, + 518 + ], + "type": "text", + "content": "Stage 1: Modality Alignment" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 309, + 510, + 400, + 518 + ], + "lines": [ + { + "bbox": [ + 309, + 510, + 400, + 518 + ], + "spans": [ + { + "bbox": [ + 309, + 510, + 400, + 518 + ], + "type": "text", + "content": "Stage 3: Long Video Tuning (Optional)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_caption" + } + ], + "index": 3, + "sub_type": "code", + "guess_lang": "python" + }, + { + "bbox": [ + 132, + 360, + 250, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 360, + 250, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 360, + 250, + 372 + ], + "type": "text", + "content": "3.3 Training Strategy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 380, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 380, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 482, + 429 + ], + "type": "text", + "content": "Training strategy, particularly instruction tuning, has proven to be crucial in LLMs [12, 53, 54] and VLMs [13, 32, 33]. Considering training efficiency, in this work, we divide the training procedure into three stages, i.e., modality alignment, instruction tuning, and long video tuning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 440, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 440, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 131, + 440, + 482, + 489 + ], + "type": "text", + "content": "Modality Alignment. As shown in Figure 2, each video frame is projected into the space of LLMs in each forward pass. Therefore, it is crucial to ensure visual features are well aligned with the language space. To this end, we construct a compact dataset that contains 790K high-quality image- and video-caption pairs." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 136, + 519, + 290, + 617 + ], + "blocks": [ + { + "bbox": [ + 136, + 519, + 290, + 617 + ], + "lines": [ + { + "bbox": [ + 136, + 519, + 290, + 617 + ], + "spans": [ + { + "bbox": [ + 136, + 519, + 290, + 617 + ], + "type": "image", + "image_path": "62f3d2892bcd1ea522a3318d308948c4fa02c4266572185b880f28f9a46a99fc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 628, + 482, + 662 + ], + "lines": [ + { + "bbox": [ + 130, + 628, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 482, + 662 + ], + "type": "text", + "content": "Fig.3: Multimodal data distribution and instruction format for each model training stage. and denote the token for single image and the " + }, + { + "bbox": [ + 130, + 628, + 482, + 662 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 628, + 482, + 662 + ], + "type": "text", + "content": "-th video frame, respectively." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 519, + 477, + 562 + ], + "blocks": [ + { + "bbox": [ + 309, + 519, + 477, + 562 + ], + "lines": [ + { + "bbox": [ + 309, + 519, + 477, + 562 + ], + "spans": [ + { + "bbox": [ + 309, + 519, + 477, + 562 + ], + "type": "image", + "image_path": "27d3cdddae9ee3ab385bc27e54e68df0c0f44ca9c5723681c3e3b91829be4218.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 231, + 101 + ], + "type": "text", + "content": "L. Yanwei et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 118, + 302, + 152 + ], + "blocks": [ + { + "bbox": [ + 135, + 118, + 302, + 152 + ], + "lines": [ + { + "bbox": [ + 135, + 118, + 302, + 152 + ], + "spans": [ + { + "bbox": [ + 135, + 118, + 302, + 152 + ], + "type": "image", + "image_path": "7f197feedc3a18f177efe7ec95c499f55cbe7cb9742d4a77850dcf4ece39ae61.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 136, + 162, + 301, + 198 + ], + "blocks": [ + { + "bbox": [ + 136, + 162, + 301, + 198 + ], + "lines": [ + { + "bbox": [ + 136, + 162, + 301, + 198 + ], + "spans": [ + { + "bbox": [ + 136, + 162, + 301, + 198 + ], + "type": "image", + "image_path": "2c5aa9aa91eea4ed1cf89a2e60911465da7c1cc646a24ac7c0814ae5d634e8de.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 205, + 301, + 248 + ], + "blocks": [ + { + "bbox": [ + 136, + 205, + 301, + 248 + ], + "lines": [ + { + "bbox": [ + 136, + 205, + 301, + 248 + ], + "spans": [ + { + "bbox": [ + 136, + 205, + 301, + 248 + ], + "type": "image", + "image_path": "75471a82147bf9acafe1c681dd3906adc4b3c01023236f39b97cb330adccabda.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 261, + 482, + 296 + ], + "lines": [ + { + "bbox": [ + 131, + 261, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 131, + 261, + 482, + 296 + ], + "type": "text", + "content": "Fig. 4: An example to construct instruction pairs for the movie Titanic. Given the movie synopsis and script, we utilize the developed LLMs like GPT-4 [45] and Claude-2 [3] to generate movie summaries, plot-related QA pairs, and general reasoning pairs." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 117, + 477, + 152 + ], + "blocks": [ + { + "bbox": [ + 310, + 117, + 477, + 152 + ], + "lines": [ + { + "bbox": [ + 310, + 117, + 477, + 152 + ], + "spans": [ + { + "bbox": [ + 310, + 117, + 477, + 152 + ], + "type": "image", + "image_path": "813f71ff703751aa520a4a478429818f3458d252c130cf47ef598f6f9366c557.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 311, + 159, + 477, + 198 + ], + "blocks": [ + { + "bbox": [ + 311, + 159, + 477, + 198 + ], + "lines": [ + { + "bbox": [ + 311, + 159, + 477, + 198 + ], + "spans": [ + { + "bbox": [ + 311, + 159, + 477, + 198 + ], + "type": "image", + "image_path": "4d01d043fddc5e8a8368b8dec20a40f587859395a77c4999585584e5f6b82104.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 311, + 204, + 477, + 248 + ], + "blocks": [ + { + "bbox": [ + 311, + 204, + 477, + 248 + ], + "lines": [ + { + "bbox": [ + 311, + 204, + 477, + 248 + ], + "spans": [ + { + "bbox": [ + 311, + 204, + 477, + 248 + ], + "type": "image", + "image_path": "4db7cfbd623810400fd8af925e5d2dbca6d2d171ab7188b95a9d53632cc526af.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 318, + 482, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 318, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 131, + 318, + 482, + 378 + ], + "type": "text", + "content": "As illustrated in Figure 3, it mainly consists of 558K image-caption pairs from the LLaVA-filtered CC3M dataset [49] and 232K video-caption pairs sampled from the WebVid 2.5M dataset [5]. The instruction format for each modality is presented in Figure 3. In this stage, we optimize the projectors in Figure 2, while freezing the pre-trained modules like the visual encoder and text decoder." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 390, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 390, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 390, + 482, + 510 + ], + "type": "text", + "content": "Instruction Tuning. To enhance the multi-modality understanding of LLMs, we build the instruction pairs from [32] and [39]. In particular, the constructed dataset mainly involves content from three sources, i.e., 40K text conversations from ShareGPT [1], 625K single- or multi-turn visual QA pairs from [19,22,25, 26,33,41-43,48,50], and 98K video QA pairs from [7]. For the instruction, we adopt different formats for text, image, and video input, as shown in Figure 3. The input prompt and answer vary with datasets. Please refer to [32] and [39] for more details. Meanwhile, the image token is randomly inserted at the beginning or end of the user input during our training. In instruction tuning, all the modules are optimized except the visual encoder." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 482, + 666 + ], + "type": "text", + "content": "Long Video Tuning. To further unleash the potential for hour-long videos, we construct 15K long QA pairs, including 9K conversions in movie scenes and 6K data sampled from LongLoRA [11] for token expanding. Specifically, we utilize more than 400 long movies and corresponding scripts in MovieNet [21] to build the training set. The key components for instruction generation are visualized in Figure 4. Generally, the generated dataset includes QA pairs from three aspects: video summary, movie plot, and detail reasoning. For video summaries, we collect movie synopses to produce brief and detailed summaries for each movie using developed LLMs like GPT-4 [45]. It brings about 1K summary-level instruction pairs in total. For plot-level data, we take the entire movie synopsis as input and leverage GPT-4 [45] to generate plot-related and character-related QA pairs. These include plot understanding, description, analysis, character relationship," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "LLaMA-VID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "personality, and behavior. In particular, we generate 5 plot-related pairs and 5 character-related pairs for each movie, resulting in 4K plot-level QA data. As for detail-level data, we feed the long movie script into Claude-2 [3] and generate 5 plot-related reasoning pairs and 5 detail-related descriptions for each movie, which brings 4K pairs in total. With long videos and the generated pairs, we perform instruction tuning by concatenating visual tokens and subtitle tokens for each frame, as depicted in Figure 3. In this way, LLaMA-VID can well support 64K tokens with more than 3-hour video as input." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 232, + 230, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 230, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 230, + 247 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 481, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 481, + 283 + ], + "type": "text", + "content": "In this section, we provide the experimental setup and comparisons with leading methods on several benchmarks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 301, + 262, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 301, + 262, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 301, + 262, + 315 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 323, + 482, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 420 + ], + "type": "text", + "content": "Implementation Details. In this work, we instantiate the model with the pre-trained EVA-G [16] for visual encoder and QFormer [13] for text decoder by default. During training, we keep the visual encoder fixed in all stages and freeze the text decoder, as well as the LLM, in the modality alignment stage, except for the BERT module in Table 8 that is not pre-trained. Following the strategy in [32], we optimize trainable parameters with the designed data and instructions in Figure 3, running for 1 epoch in each stage. For video input, we extract frames at 1 FPS. All models are trained using " + }, + { + "bbox": [ + 130, + 323, + 482, + 420 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 130, + 323, + 482, + 420 + ], + "type": "text", + "content": " NVIDIA A100 GPUs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 433, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 482, + 542 + ], + "type": "text", + "content": "Datasets. In this study, we construct the training set mainly from [5,21,32,39], as illustrated in Section 3.3. Moreover, we report results on several video- and image-based benchmarks. In particular, for video input, we evaluate the zero-shot performance on the open-ended QA benchmarks like MSVD [8], MSRVTT [59], ActivityNet [7], EgoSchema [40], SSv2 (Something-Something v2) [18], and the newly-proposed generative performance benchmark [39]. As for image-based evaluation, we conduct experiments on several widely-adopted benchmarks, including GQA [22], MMB (MMBench) [36], MME [17], POPE [31], SEED [28], " + }, + { + "bbox": [ + 130, + 433, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\mathrm{SQA}^{\\mathrm{I}}" + }, + { + "bbox": [ + 130, + 433, + 482, + 542 + ], + "type": "text", + "content": " [37], " + }, + { + "bbox": [ + 130, + 433, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\mathrm{VQAT}^{\\mathrm{T}}" + }, + { + "bbox": [ + 130, + 433, + 482, + 542 + ], + "type": "text", + "content": " (TextVQA) [51], VizWiz [20], and " + }, + { + "bbox": [ + 130, + 433, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\mathrm{VQAv}^{\\mathrm{v2}}" + }, + { + "bbox": [ + 130, + 433, + 482, + 542 + ], + "type": "text", + "content": " (VQA V2) [19]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 559, + 228, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 228, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 228, + 571 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Results on Video-based Benchmarks. In Table 1, we provide a comparative evaluation of LLaMA-VID against various state-of-the-art methods across zero-shot video QA benchmarks: MSVD-QA [8], MSRVTT-QA [59], ActivityNet-QA [7], EgoSchema [40], and SSv2 [18]. Notably, the results are reported with only two tokens for each frame. It is evident that LLaMA-VID, employing Vicuna-7B and Vicuna-13B as the LLMs, consistently delivers superior performance across all datasets. On the MSVD-QA and MSRVTT-QA datasets, it" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "L. Yanwei et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 157, + 474, + 253 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 1: Comparison with leading methods on 3 zero-shot video QA datasets. We report results with 2 tokens for each frame. For fair comparisons, our model is trained without long video tuning (stage 3) in Figure 3. Res indicates image resolution." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 157, + 474, + 253 + ], + "lines": [ + { + "bbox": [ + 140, + 157, + 474, + 253 + ], + "spans": [ + { + "bbox": [ + 140, + 157, + 474, + 253 + ], + "type": "table", + "html": "
MethodLLMRes.MSVD-QAMSRVTT-QAActivityNet-QA
AccScoreAccScoreAccScore
VideoLLaMA [64]Vicuna-7B22451.62.529.61.812.41.1
LLaMA-Adapter [65]LLaMA-7B22454.93.143.82.734.22.7
VideoChat [30]Vicuna-7B22456.32.845.02.526.52.2
Video-ChatGPT [39]Vicuna-7B22464.93.349.32.835.22.7
BT-Adapter [34]Vicuna-7B-67.53.757.03.245.73.2
LLaMA-VIDVicuna-7B22469.73.757.73.247.43.3
LLaMA-VIDVicuna-13B22470.03.758.93.347.53.3
", + "image_path": "bbc70c2198b547a88367ceb7ab6089b92a7bdc2d6bb12c35d6bafa5cd837d5ce.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 317, + 479, + 402 + ], + "blocks": [ + { + "bbox": [ + 131, + 263, + 482, + 309 + ], + "lines": [ + { + "bbox": [ + 131, + 263, + 482, + 309 + ], + "spans": [ + { + "bbox": [ + 131, + 263, + 482, + 309 + ], + "type": "text", + "content": "Table 2: Comparison with leading methods on the video generative benchmark [39]. We report results with 2 tokens for each frame. For fair comparisons, our model is trained without long video tuning (stage 3) in Figure 3. Res indicates image resolution. Correct, Detail, Context, Temporal, and Consist indicate evaluation metrics in [39]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 317, + 479, + 402 + ], + "lines": [ + { + "bbox": [ + 135, + 317, + 479, + 402 + ], + "spans": [ + { + "bbox": [ + 135, + 317, + 479, + 402 + ], + "type": "table", + "html": "
MethodLLMRes.Correct.DetailContextTemporalConsist.
VideoLLaMA [64]Vicuna-7B2241.962.182.161.821.79
LLaMA-Adapter [65]LLaMA-7B2242.032.322.301.982.15
VideoChat [30]Vicuna-7B2242.232.502.531.942.24
Video-ChatGPT [39]Vicuna-7B2242.402.522.621.982.37
BT-Adapter [34]Vicuna-7B-2.682.693.272.342.46
LLaMA-VIDVicuna-7B2242.963.003.532.462.51
LLaMA-VIDVicuna-13B2243.073.053.602.582.63
", + "image_path": "c9c9a7f00be14f8e5b0431d694a4ef6cb05a05dfd83e96e643573b794ef7206f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 148, + 445, + 462, + 511 + ], + "blocks": [ + { + "bbox": [ + 131, + 413, + 482, + 436 + ], + "lines": [ + { + "bbox": [ + 131, + 413, + 482, + 436 + ], + "spans": [ + { + "bbox": [ + 131, + 413, + 482, + 436 + ], + "type": "text", + "content": "Table 3: Comparison with leading methods on the zero-shot benchmark EgoSchema [40]. We report results with 2 tokens for each frame on different settings." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 445, + 462, + 511 + ], + "lines": [ + { + "bbox": [ + 148, + 445, + 462, + 511 + ], + "spans": [ + { + "bbox": [ + 148, + 445, + 462, + 511 + ], + "type": "table", + "html": "
FrameFrozenBiLM [60]mPLUG-Owl [62]InternVideo [56]LLaMA-VID
1026.4%29.6%31.4%41.2%
30-20.0%31.8%41.7%
9026.9%-32.1%40.5%
180---41.7%
", + "image_path": "63acb632e1784d863f70b8c4485afcec79a535e947bb673de9b84d54801b89d4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "achieves the accuracy of " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "69.7\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "57.7\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " with Vicuna-7B, surpassing the previous leading approach [34] with absolute gains of " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", respectively. For the ActivityNet-QA dataset, LLaMA-VID attains top performance in accuracy and the highest score of 3.3. In Table 2, we also carry out experiments on the newly proposed video-based generative performance benchmark [39]. As for the long EgoSchema and SSv2 dataset, the proposed method surpasses other leading approaches in different frame settings, as presented in Table 3 and Table 4. Specifically, we evaluate the model on the SSv2 dataset similar with that in MSRVTT-QA. Our LLaMA-VID is validated to achieve the best performance across all the evaluation metrics, surpassing previous approaches by a large margin. In general, LLaMA-VID is proved to bring robust results on all benchmarks." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "LLaMA-VID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 201, + 135, + 411, + 179 + ], + "blocks": [ + { + "bbox": [ + 166, + 114, + 446, + 126 + ], + "lines": [ + { + "bbox": [ + 166, + 114, + 446, + 126 + ], + "spans": [ + { + "bbox": [ + 166, + 114, + 446, + 126 + ], + "type": "text", + "content": "Table 4: Comparisons with zero-shot methods on SSv2 [18] dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 201, + 135, + 411, + 179 + ], + "lines": [ + { + "bbox": [ + 201, + 135, + 411, + 179 + ], + "spans": [ + { + "bbox": [ + 201, + 135, + 411, + 179 + ], + "type": "table", + "html": "
MethodLLMRes.AccScore
Video-ChatGPT [38]Vicuna-7B22411.592.10
LLaMA-VIDVicuna-7B22420.122.14
", + "image_path": "b05e475af42b7dbbba3a374f8fcc6b17050ed770fc001b60f8f059cc3a09c641.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 253, + 479, + 383 + ], + "blocks": [ + { + "bbox": [ + 132, + 188, + 482, + 244 + ], + "lines": [ + { + "bbox": [ + 132, + 188, + 482, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 188, + 482, + 244 + ], + "type": "text", + "content": "Table 5: Comparison with leading methods on 8 benchmarks. Here, we use the same training and instruction finetuning data as that in LLaVA-1.5. We report results with 1 context token and " + }, + { + "bbox": [ + 132, + 188, + 482, + 244 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 132, + 188, + 482, + 244 + ], + "type": "text", + "content": " content tokens, where " + }, + { + "bbox": [ + 132, + 188, + 482, + 244 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 132, + 188, + 482, + 244 + ], + "type": "text", + "content": " is kept the same with that in LLaVA-1.5, i.e., " + }, + { + "bbox": [ + 132, + 188, + 482, + 244 + ], + "type": "inline_equation", + "content": "n = (336 / 14)^2 = 576" + }, + { + "bbox": [ + 132, + 188, + 482, + 244 + ], + "type": "text", + "content": ". Res indicates input image resolution. * and † denote the train subset is included for training and the data is not publicly available, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 253, + 479, + 383 + ], + "lines": [ + { + "bbox": [ + 135, + 253, + 479, + 383 + ], + "spans": [ + { + "bbox": [ + 135, + 253, + 479, + 383 + ], + "type": "table", + "html": "
MethodLLMRes.GQAMMBMMEPOPESEED\\( \\mathbf{SQA}^{\\dagger} \\)VizWiz\\( VQA^{v2} \\)
InstructBLIP [13]Vicuna-7B22449.236.0--53.460.534.5-
IDEFICS-9B [23]LLaMA-7B22438.448.2----35.550.9
Qwen-VL† [4]Qwen-7B44859.3*38.2--56.367.135.278.8*
Qwen-VL-Chat† [4]Qwen-7B44857.5*60.61487.5-58.268.238.978.2*
LLaVA-1.5 [32]Vicuna-7B33662.0*64.31510.785.958.666.850.078.5*
LLaMA-VIDVicuna-7B33664.3*65.11521.486.059.968.354.279.3*
BLIP-2 [29]Vicuna-13B22441.0-1293.885.346.461.019.641.0
InstructBLIP [13]Vicuna-13B22449.5-1212.878.9-63.133.4-
Shikra [9]Vicuna-13B224-58.8-----77.4*
IDEFICS-80B [23]LLaMA-65B22445.254.5----36.060.0
LLaVA-1.5 [32]Vicuna-13B33663.3*67.71531.385.961.671.653.680.0*
LLaMA-VIDVicuna-13B33665.0*66.61542.386.062.370.054.380.0*
", + "image_path": "b35063c6794095e7b019e6b045e62d610ecf382e4b9c4bfed9c918fa5c8d5573.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 403, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 559 + ], + "type": "text", + "content": "Results on Image-based Benchmarks. As illustrated in Section 3.2, LLaMA-VID not only efficiently represents long video, but also expands the upper limit of VLMs with an additional context token. In Table 5, we perform comparisons with leading VLMs on 8 widely-adopted benchmarks. It should be noted that we maintain the same training data and image resolution as in LLaVA-1.5 [32] for fair comparisons. It is evident that LLaMA-VID outperforms other leading methods across most of benchmarks with different LLMs. In particular, with Vicuna-7B, LLaMA-VID attains the best results across all the datasets and surpasses LLaVA-1.5 with significant gains in GQA, MME, and VizWiz, where the improvement reaches up to " + }, + { + "bbox": [ + 130, + 403, + 482, + 559 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 130, + 403, + 482, + 559 + ], + "type": "text", + "content": ", 10.7, and " + }, + { + "bbox": [ + 130, + 403, + 482, + 559 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 130, + 403, + 482, + 559 + ], + "type": "text", + "content": ", respectively. With a larger Vicuna-13B, LLaMA-VID also outperforms other methods in 6 benchmarks and achieves top-2 in the other datasets. This demonstrates the generality of LLaMA-VID, which can be scaled up with a stronger foundation model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Qualitative Results. In Figure 5, we apply LLaMA-VID to different types of data, including short videos and long movies. We represent each frame with 2 tokens for videos. LLaMA-VID demonstrates various capabilities with different inputs. Given a short trailer video, LLaMA-VID summarizes the overall plot and infers the movie name, producer, and the type of movie. It shows that the proposed LLaMA-VID can well recognize the typical characters with the produced tokens. As for a 3-hour movie, the proposed model adeptly describes the storyline and demonstrates plot-related reasoning and detailed understanding." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "type": "text", + "content": "L. Yanwei et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 294, + 175 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 294, + 175 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 294, + 175 + ], + "type": "image", + "image_path": "72a665b4f944cc1c8215ff56a24bf462941473040496baa5701c5c2718a46530.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 182, + 272, + 192 + ], + "lines": [ + { + "bbox": [ + 146, + 182, + 272, + 192 + ], + "spans": [ + { + "bbox": [ + 146, + 182, + 272, + 192 + ], + "type": "text", + "content": "User: Please describe this video in detail." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 279, + 180, + 292, + 193 + ], + "blocks": [ + { + "bbox": [ + 279, + 180, + 292, + 193 + ], + "lines": [ + { + "bbox": [ + 279, + 180, + 292, + 193 + ], + "spans": [ + { + "bbox": [ + 279, + 180, + 292, + 193 + ], + "type": "image", + "image_path": "8ee3bb725204f6e157a720e757acc358757d42f42151a99240501c19c0c3461b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 341, + 141, + 356, + 153 + ], + "blocks": [ + { + "bbox": [ + 341, + 141, + 356, + 153 + ], + "lines": [ + { + "bbox": [ + 341, + 141, + 356, + 153 + ], + "spans": [ + { + "bbox": [ + 341, + 141, + 356, + 153 + ], + "type": "image", + "image_path": "fd7f68283ef81c51cf52c956375506abddf2e4a5c95f9e658b813886db30e4ff.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 155, + 368, + 172 + ], + "lines": [ + { + "bbox": [ + 328, + 155, + 368, + 172 + ], + "spans": [ + { + "bbox": [ + 328, + 155, + 368, + 172 + ], + "type": "text", + "content": "3-hour VideoAvatar" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 372, + 116, + 413, + 175 + ], + "blocks": [ + { + "bbox": [ + 372, + 116, + 413, + 175 + ], + "lines": [ + { + "bbox": [ + 372, + 116, + 413, + 175 + ], + "spans": [ + { + "bbox": [ + 372, + 116, + 413, + 175 + ], + "type": "image", + "image_path": "734c9af289904b0876a4ea61cfea6e1ded8258f3424ea95681f14d03d6a2c762.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 183, + 444, + 192 + ], + "lines": [ + { + "bbox": [ + 324, + 183, + 444, + 192 + ], + "spans": [ + { + "bbox": [ + 324, + 183, + 444, + 192 + ], + "type": "text", + "content": "User: Summarize the plot of this movie." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 413, + 116, + 470, + 175 + ], + "blocks": [ + { + "bbox": [ + 413, + 116, + 470, + 175 + ], + "lines": [ + { + "bbox": [ + 413, + 116, + 470, + 175 + ], + "spans": [ + { + "bbox": [ + 413, + 116, + 470, + 175 + ], + "type": "image", + "image_path": "bdec7b6c096e13dffb02b5a47ab1d8e1bd53886907c113822141ae1c715de62c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 457, + 182, + 469, + 193 + ], + "blocks": [ + { + "bbox": [ + 457, + 182, + 469, + 193 + ], + "lines": [ + { + "bbox": [ + 457, + 182, + 469, + 193 + ], + "spans": [ + { + "bbox": [ + 457, + 182, + 469, + 193 + ], + "type": "image", + "image_path": "77b1b3cf4b8ffccce0291d0c13432ce2f4503730d2118dff6cd9adc708585953.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 145, + 201, + 161, + 221 + ], + "blocks": [ + { + "bbox": [ + 145, + 201, + 161, + 221 + ], + "lines": [ + { + "bbox": [ + 145, + 201, + 161, + 221 + ], + "spans": [ + { + "bbox": [ + 145, + 201, + 161, + 221 + ], + "type": "image", + "image_path": "d81eb905ae1ee06d782c52e2f95e2ed44489a292b3915b8bc612ad94bd36283f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 165, + 204, + 290, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 204, + 290, + 281 + ], + "spans": [ + { + "bbox": [ + 165, + 204, + 290, + 281 + ], + "type": "text", + "content": "Assistant: The video starts with a carto on character standing in front of a castle, which is surrounded by a beautiful lan descape. The character is holding a bucket and appears to be observing the surroundings. The scene then transitions to a t train, where the character is seen sitting inside the train. The train is surrounded by lush greenery and appears to be mov ing through a scenic route...." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 323, + 201, + 339, + 222 + ], + "blocks": [ + { + "bbox": [ + 323, + 201, + 339, + 222 + ], + "lines": [ + { + "bbox": [ + 323, + 201, + 339, + 222 + ], + "spans": [ + { + "bbox": [ + 323, + 201, + 339, + 222 + ], + "type": "image", + "image_path": "e050285158efd78ed3d3d8396ffab588de67a13fa8fe849fcfcb7d0406126027.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 342, + 204, + 466, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 204, + 466, + 281 + ], + "spans": [ + { + "bbox": [ + 342, + 204, + 466, + 281 + ], + "type": "text", + "content": "Assistant:Avatar follows Jake Sully, a disabled former Marine who is recruited by the RDA to work on Pandora, a moon rich in a valuable mineral called unobtanium. Jake's twin brother, Tommy, died in a helmet-cam accident, and Jake is offered his place in the Avatar Program, which allows humans to link with the native Na' vi, a humanoid species. Jake is assigned to the avatar of his deceased brother...." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 146, + 292, + 272, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 292, + 272, + 308 + ], + "spans": [ + { + "bbox": [ + 146, + 292, + 272, + 308 + ], + "type": "text", + "content": "User: What's the name of the movie described in this video?" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 280, + 291, + 291, + 304 + ], + "blocks": [ + { + "bbox": [ + 280, + 291, + 291, + 304 + ], + "lines": [ + { + "bbox": [ + 280, + 291, + 291, + 304 + ], + "spans": [ + { + "bbox": [ + 280, + 291, + 291, + 304 + ], + "type": "image", + "image_path": "8fb7ed9008fb0fc006e34c39774d688179bad2dbfaa53a23949bf75d2f0f1065.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 456, + 289, + 468, + 301 + ], + "blocks": [ + { + "bbox": [ + 456, + 289, + 468, + 301 + ], + "lines": [ + { + "bbox": [ + 456, + 289, + 468, + 301 + ], + "spans": [ + { + "bbox": [ + 456, + 289, + 468, + 301 + ], + "type": "image", + "image_path": "f6d9d45df6c90a54aedd838c34e4ac056cd2e79da034450845f24c0a1c459aca.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 146, + 317, + 162, + 337 + ], + "blocks": [ + { + "bbox": [ + 146, + 317, + 162, + 337 + ], + "lines": [ + { + "bbox": [ + 146, + 317, + 162, + 337 + ], + "spans": [ + { + "bbox": [ + 146, + 317, + 162, + 337 + ], + "type": "image", + "image_path": "9add4a71043559c97faef4fd0b846eab03a20bdbc35205b8ba24eb418d99a528.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 166, + 319, + 288, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 319, + 288, + 336 + ], + "spans": [ + { + "bbox": [ + 166, + 319, + 288, + 336 + ], + "type": "text", + "content": "Assistant: The movie described in this video is \"Zootopia.\"" + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 323, + 308, + 339, + 328 + ], + "blocks": [ + { + "bbox": [ + 325, + 292, + 432, + 300 + ], + "lines": [ + { + "bbox": [ + 325, + 292, + 432, + 300 + ], + "spans": [ + { + "bbox": [ + 325, + 292, + 432, + 300 + ], + "type": "text", + "content": "User: How human work on Pandora?" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 308, + 339, + 328 + ], + "lines": [ + { + "bbox": [ + 323, + 308, + 339, + 328 + ], + "spans": [ + { + "bbox": [ + 323, + 308, + 339, + 328 + ], + "type": "image", + "image_path": "ecfc7d758e0961100b562c8c9cfe8d12ca884d4463bbdfc8f7928d4e7212910b.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 343, + 311, + 465, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 311, + 465, + 350 + ], + "spans": [ + { + "bbox": [ + 343, + 311, + 465, + 350 + ], + "type": "text", + "content": "Assistant: Humans work on Pandora by using avatars, which are grown from human DNA mixed with the DNA of the native Na'vi, to interact with the environment and the Na'vi." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 146, + 347, + 234, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 347, + 234, + 356 + ], + "spans": [ + { + "bbox": [ + 146, + 347, + 234, + 356 + ], + "type": "text", + "content": "User: Who produce the movie?" + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 280, + 345, + 291, + 357 + ], + "blocks": [ + { + "bbox": [ + 280, + 345, + 291, + 357 + ], + "lines": [ + { + "bbox": [ + 280, + 345, + 291, + 357 + ], + "spans": [ + { + "bbox": [ + 280, + 345, + 291, + 357 + ], + "type": "image", + "image_path": "8ae39fbd05ec4df51649fca78a50950ff4cfe12dcb215ec2440482673f4cc56a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 324, + 360, + 451, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 360, + 451, + 377 + ], + "spans": [ + { + "bbox": [ + 324, + 360, + 451, + 377 + ], + "type": "text", + "content": "User: Why Jake Sully decide to fight againt humanity with the Na'vi?" + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 457, + 359, + 469, + 371 + ], + "blocks": [ + { + "bbox": [ + 457, + 359, + 469, + 371 + ], + "lines": [ + { + "bbox": [ + 457, + 359, + 469, + 371 + ], + "spans": [ + { + "bbox": [ + 457, + 359, + 469, + 371 + ], + "type": "image", + "image_path": "ef43d0dd38c7d488cc75d3a24093f39d1d94417afbcdf9f7baa4eef1a2fb0be7.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 145, + 365, + 161, + 384 + ], + "blocks": [ + { + "bbox": [ + 145, + 365, + 161, + 384 + ], + "lines": [ + { + "bbox": [ + 145, + 365, + 161, + 384 + ], + "spans": [ + { + "bbox": [ + 145, + 365, + 161, + 384 + ], + "type": "image", + "image_path": "b10e02a5fc1e5b2c2b28de1f37525f77a539f2dabbeaf4b9f5c76c9a89eeee6f.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 166, + 368, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 368, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 166, + 368, + 287, + 384 + ], + "type": "text", + "content": "Assistant: The movie is produced by Walt Disney Pictures." + } + ] + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 323, + 384, + 339, + 404 + ], + "blocks": [ + { + "bbox": [ + 323, + 384, + 339, + 404 + ], + "lines": [ + { + "bbox": [ + 323, + 384, + 339, + 404 + ], + "spans": [ + { + "bbox": [ + 323, + 384, + 339, + 404 + ], + "type": "image", + "image_path": "b40b3eba5c1f3b56ff1d4399214042fe8fc836c591d3176d58ccdcbc41a1197a.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 342, + 387, + 465, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 387, + 465, + 434 + ], + "spans": [ + { + "bbox": [ + 342, + 387, + 465, + 434 + ], + "type": "text", + "content": "Assistant: Jake Sully decides to fight a gainst humanity with the Na'vi after he falls in love with the Na'vi and their way of life, and after he realizes that the human military is destroying their home and culture." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 146, + 396, + 260, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 396, + 260, + 405 + ], + "spans": [ + { + "bbox": [ + 146, + 396, + 260, + 405 + ], + "type": "text", + "content": "User: What's the type of this movie?" + } + ] + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 279, + 394, + 291, + 407 + ], + "blocks": [ + { + "bbox": [ + 279, + 394, + 291, + 407 + ], + "lines": [ + { + "bbox": [ + 279, + 394, + 291, + 407 + ], + "spans": [ + { + "bbox": [ + 279, + 394, + 291, + 407 + ], + "type": "image", + "image_path": "287e9de54548bfde8d7bae7dff2287c3eb4372defa68ef5b5d0a5ae21d59c3dd.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 145, + 413, + 161, + 434 + ], + "blocks": [ + { + "bbox": [ + 145, + 413, + 161, + 434 + ], + "lines": [ + { + "bbox": [ + 145, + 413, + 161, + 434 + ], + "spans": [ + { + "bbox": [ + 145, + 413, + 161, + 434 + ], + "type": "image", + "image_path": "429f15612f6aee68abd959212c1f33a3cdcbcfde3fe754d04f0f37ca20a2ceed.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 447, + 453, + 459 + ], + "lines": [ + { + "bbox": [ + 159, + 447, + 453, + 459 + ], + "spans": [ + { + "bbox": [ + 159, + 447, + 453, + 459 + ], + "type": "text", + "content": "Fig. 5: Examples of LLaMA-VID with short video and hour-long video." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "bbox": [ + 166, + 417, + 283, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 417, + 283, + 433 + ], + "spans": [ + { + "bbox": [ + 166, + 417, + 283, + 433 + ], + "type": "text", + "content": "Assistant: The movie is a Disney animat ed film." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 132, + 483, + 290, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 290, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 290, + 496 + ], + "type": "text", + "content": "4.3 Component-wise Analysis" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 131, + 506, + 481, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 506, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 131, + 506, + 481, + 531 + ], + "type": "text", + "content": "In this subsection, we conduct ablation studies with input resolution 224 and 2 tokens for each image by default." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 131, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 545, + 482, + 666 + ], + "type": "text", + "content": "Generated Token Types. As illustrated in Figure 2, each image is represented with a context token and a content token in LLMs. To validate the effectiveness of each part, we conduct experiments with different types of tokens in Table 6. Without the context token, the compressed content token, which encodes each image with 1 token, cannot adjust to input instructions, leading to subpar performance. Compared with a single content token, the instruction-guided context token results in significant gains across all datasets with only 1 token. With both tokens for each image, the model achieves the best performance across all image- and video-based benchmarks. It shows that both instruction cues in the context token and the image content itself in the content token are important." + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "LLaMA-VID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 176, + 146, + 435, + 198 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 6: Comparison with different token types. We report results with 1 context token (if exists) and 1 content token on image-based and video-based benchmarks (MSVD)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 176, + 146, + 435, + 198 + ], + "lines": [ + { + "bbox": [ + 176, + 146, + 435, + 198 + ], + "spans": [ + { + "bbox": [ + 176, + 146, + 435, + 198 + ], + "type": "table", + "html": "
contextcontentGQAPOPE\\( \\mathbf{SQA^I} \\)\\( VQA^T \\)MSVD
✘✓53.380.966.146.565.9
✓✘54.382.467.748.369.3
✓✓55.583.168.849.069.7
", + "image_path": "70c7b59a3c7cf7d0d0a6ad1fa43c184b287f7c3da4ac101ae08d59e80f4f8741.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 197, + 242, + 414, + 328 + ], + "blocks": [ + { + "bbox": [ + 132, + 209, + 481, + 232 + ], + "lines": [ + { + "bbox": [ + 132, + 209, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 481, + 232 + ], + "type": "text", + "content": "Table 7: Comparison with different token numbers. We report results with various numbers of context token and content token on image-based benchmarks for efficiency." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 197, + 242, + 414, + 328 + ], + "lines": [ + { + "bbox": [ + 197, + 242, + 414, + 328 + ], + "spans": [ + { + "bbox": [ + 197, + 242, + 414, + 328 + ], + "type": "table", + "html": "
contextcontentGQAPOPE\\( \\mathbf{SQA^I} \\)\\( VQA^T \\)
025661.985.567.553.0
125663.086.667.753.8
16460.885.168.752.3
11658.283.167.450.8
1456.283.568.749.1
1155.583.168.849.0
", + "image_path": "9093f7acdb1f419ecad104afa6078a9d0c5bd9e0d46d3c046aa1a65c8405bdc9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "text", + "content": "Generated Token Numbers. In Table 7, we conduct experiments with different numbers of tokens for further investigation. With an image size " + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "text", + "content": ", we set up experiments with " + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "text", + "content": " content tokens, where " + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "inline_equation", + "content": "n = (224 / 14)^{2} = 256" + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "text", + "content": " for uncompressed settings in the first two rows. The results clearly show that the context token consistently improves performance across different benchmarks with only 1 extra token. When we compress the content token to " + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "inline_equation", + "content": "1 / 4" + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "inline_equation", + "content": "n = 64" + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "text", + "content": ", the performance drops about " + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "text", + "content": " but increases " + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathrm{SQA}^{\\mathrm{I}}" + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "text", + "content": ". Considering the extra efficient setting for hour-long videos, we compress the content token to " + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "inline_equation", + "content": "1 / 256" + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "inline_equation", + "content": "n = 1" + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "text", + "content": " by default. Compared to the original setting without context token, we can reduce the computational cost to " + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "inline_equation", + "content": "1 / 128" + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "text", + "content": " with about " + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "inline_equation", + "content": "2\\% - 6\\%" + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "text", + "content": " performance drop, which is generally acceptable. The linear increase in performance presents significant potential for token compression. For instance, we can dynamically compress the content token to different numbers according to resource budget and content importance. Interestingly, the model achieves peak performance in " + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathrm{SQA}^{\\mathrm{I}}" + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "text", + "content": " with only 2 tokens. This could be attributed to the fact that problems in ScienceQA [37] focus more on visual-based reasoning rather than image details. As demonstrated in Tables 1 and 2, with only 2 tokens for each image, LLaMA-VID still outperforms all previous work in different video-based benchmarks. This makes it feasible to enable LLMs for hour-long video processing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Text Decoder. As depicted in Figure 2, the text decoder plays an essential role in producing instruction-guided context cues. Here, we further perform comparisons with different text decoders in Table 8. We mainly instantiate the text decoder with two types of modules, namely BERT [14] and QFormer [13]. For BERT, we randomly initialize it as a cross-modality decoder and only retain the first two layers. As for QFormer, we utilize the pre-trained modules and fix them" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "type": "text", + "content": "L. Yanwei et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 212, + 146, + 400, + 200 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 8: Comparison with different text decoders. We report results with 1 context token (if exists) and 1 content token on image-based benchmarks for efficiency." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 212, + 146, + 400, + 200 + ], + "lines": [ + { + "bbox": [ + 212, + 146, + 400, + 200 + ], + "spans": [ + { + "bbox": [ + 212, + 146, + 400, + 200 + ], + "type": "table", + "html": "
textGQAPOPE\\( \\mathbf{SQA^I} \\)\\( \\mathbf{VQA^T} \\)
-53.380.966.146.5
BERT54.180.867.948.1
QFormer55.583.168.849.0
", + "image_path": "1df1c4843c5d5d130c5f17f2b83721b5fadac6df3e9622453e9642d8c9760f56.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 179, + 254, + 432, + 308 + ], + "blocks": [ + { + "bbox": [ + 132, + 211, + 481, + 245 + ], + "lines": [ + { + "bbox": [ + 132, + 211, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 211, + 481, + 245 + ], + "type": "text", + "content": "Table 9: Average module latency of video frame on a A100 GPU. Encoder, Content, Decoder, and Context denote the latency for text-free visual encoder and content token, text-related text decoder and context token generation, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 179, + 254, + 432, + 308 + ], + "lines": [ + { + "bbox": [ + 179, + 254, + 432, + 308 + ], + "spans": [ + { + "bbox": [ + 179, + 254, + 432, + 308 + ], + "type": "table", + "html": "
textText-freeText-relatedTotal (ms)
EncoderContentDecoderContext
BERT4.890.010.250.025.17
QFormer4.870.011.080.025.98
", + "image_path": "10d0d34e19921c09cd1e7a61c1f27450189cefc50de0c76ef8a35874f02f596d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 331, + 480, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 480, + 392 + ], + "type": "text", + "content": "for modality alignment. Even with a simple 2-layer BERT, as shown in Table 8, the generated context token achieves significant gains in most of benchmarks. This proves the effectiveness of the paradigm for context token generation. With a pre-trained text decoder like QFormer, the model can be further enhanced and attains peak performance in all datasets with " + }, + { + "bbox": [ + 130, + 331, + 480, + 392 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 130, + 331, + 480, + 392 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 331, + 480, + 392 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 130, + 331, + 480, + 392 + ], + "type": "text", + "content": " significant gain." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 408, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 517 + ], + "type": "text", + "content": "Latency Analysis. To further analyse the module latency, we keep the original setting in Table 8 and report average module latency over 100 videos in Table 9. Here, we divide the modules for generation into two types, namely text-free and text-related. In particular, we generate visual tokens one time for each frame with text-free modules. And text-related modules are required to produce different context tokens according to user inputs. It is clear that the text-free tokens account for the most computation, while the text-related modules only cost " + }, + { + "bbox": [ + 130, + 408, + 482, + 517 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 408, + 482, + 517 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 408, + 482, + 517 + ], + "type": "inline_equation", + "content": "18.3\\%" + }, + { + "bbox": [ + 130, + 408, + 482, + 517 + ], + "type": "text", + "content": " for each user instruction. That means the total cost will not increase greatly given variant user questions because of the efficient text-related modules." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Response in Context Attention. To more vividly explore the context attention, we visualize the high response areas with the top 20 scores in Figure 6. Specifically, we draw the normalized heatmap for the first four queries in " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " before applying the Softmax function, which is used to formulate context token in Equation 1. As shown in Figure 6, the text-guided query " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " effectively focuses on important areas relevant to the input questions. For example, in the second row, when inquiring whether the image depicts a fishing village, the query " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " focuses more on buildings along the river and a seagull. These are all typical characteristics to distinguish a fishing village in common sense. Other examples also confirm that the designed context attention successfully achieves its goal of formulating the context token under instruction guidance." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "LLaMA-VID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 117, + 203, + 163 + ], + "blocks": [ + { + "bbox": [ + 134, + 117, + 203, + 163 + ], + "lines": [ + { + "bbox": [ + 134, + 117, + 203, + 163 + ], + "spans": [ + { + "bbox": [ + 134, + 117, + 203, + 163 + ], + "type": "image", + "image_path": "9b55caa5a4cc5c34b647ac1b059e14c923a33db2c50956c2cf8c7da0eec0fff7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 163, + 178, + 171 + ], + "lines": [ + { + "bbox": [ + 161, + 163, + 178, + 171 + ], + "spans": [ + { + "bbox": [ + 161, + 163, + 178, + 171 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 208, + 117, + 274, + 163 + ], + "blocks": [ + { + "bbox": [ + 208, + 117, + 274, + 163 + ], + "lines": [ + { + "bbox": [ + 208, + 117, + 274, + 163 + ], + "spans": [ + { + "bbox": [ + 208, + 117, + 274, + 163 + ], + "type": "image", + "image_path": "a950cf1de1a709c4d6d996327a5f19b7fb4ab06775a37a6b6c8b3d95b75105f9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 163, + 253, + 171 + ], + "lines": [ + { + "bbox": [ + 229, + 163, + 253, + 171 + ], + "spans": [ + { + "bbox": [ + 229, + 163, + 253, + 171 + ], + "type": "text", + "content": "Query 1" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 276, + 117, + 342, + 163 + ], + "blocks": [ + { + "bbox": [ + 276, + 117, + 342, + 163 + ], + "lines": [ + { + "bbox": [ + 276, + 117, + 342, + 163 + ], + "spans": [ + { + "bbox": [ + 276, + 117, + 342, + 163 + ], + "type": "image", + "image_path": "fc726282f4090ce1c536c1dcf5cad305881be27b371fc37c3d3c92898c39da40.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 163, + 321, + 171 + ], + "lines": [ + { + "bbox": [ + 298, + 163, + 321, + 171 + ], + "spans": [ + { + "bbox": [ + 298, + 163, + 321, + 171 + ], + "type": "text", + "content": "Query 2" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 343, + 117, + 410, + 163 + ], + "blocks": [ + { + "bbox": [ + 343, + 117, + 410, + 163 + ], + "lines": [ + { + "bbox": [ + 343, + 117, + 410, + 163 + ], + "spans": [ + { + "bbox": [ + 343, + 117, + 410, + 163 + ], + "type": "image", + "image_path": "1b25ca521e7ade249fee96d8831230fa7aec15eb136f840c08db05105a628e02.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 163, + 389, + 171 + ], + "lines": [ + { + "bbox": [ + 365, + 163, + 389, + 171 + ], + "spans": [ + { + "bbox": [ + 365, + 163, + 389, + 171 + ], + "type": "text", + "content": "Query 3" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 411, + 117, + 479, + 163 + ], + "blocks": [ + { + "bbox": [ + 411, + 117, + 479, + 163 + ], + "lines": [ + { + "bbox": [ + 411, + 117, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 411, + 117, + 479, + 163 + ], + "type": "image", + "image_path": "f8d85417b652785ef5650a64562b7076d43d452be66eb6e5451e615020cff279.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 163, + 457, + 171 + ], + "lines": [ + { + "bbox": [ + 433, + 163, + 457, + 171 + ], + "spans": [ + { + "bbox": [ + 433, + 163, + 457, + 171 + ], + "type": "text", + "content": "Query 4" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 135, + 194, + 203, + 245 + ], + "blocks": [ + { + "bbox": [ + 135, + 194, + 203, + 245 + ], + "lines": [ + { + "bbox": [ + 135, + 194, + 203, + 245 + ], + "spans": [ + { + "bbox": [ + 135, + 194, + 203, + 245 + ], + "type": "image", + "image_path": "6499112d0933b04822419699d45b4cccf51b8b29c975d2d4467da848846ef8d4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 247, + 178, + 254 + ], + "lines": [ + { + "bbox": [ + 161, + 247, + 178, + 254 + ], + "spans": [ + { + "bbox": [ + 161, + 247, + 178, + 254 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 208, + 194, + 274, + 245 + ], + "blocks": [ + { + "bbox": [ + 244, + 174, + 370, + 185 + ], + "lines": [ + { + "bbox": [ + 244, + 174, + 370, + 185 + ], + "spans": [ + { + "bbox": [ + 244, + 174, + 370, + 185 + ], + "type": "text", + "content": "Question: Are the elephants fighting?" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 208, + 194, + 274, + 245 + ], + "lines": [ + { + "bbox": [ + 208, + 194, + 274, + 245 + ], + "spans": [ + { + "bbox": [ + 208, + 194, + 274, + 245 + ], + "type": "image", + "image_path": "fb5d34a3fe3c591ab99e4e3e0c568403a37785ecf775d331ad906b9bede29fe1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 247, + 253, + 254 + ], + "lines": [ + { + "bbox": [ + 230, + 247, + 253, + 254 + ], + "spans": [ + { + "bbox": [ + 230, + 247, + 253, + 254 + ], + "type": "text", + "content": "Query 1" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 276, + 194, + 342, + 245 + ], + "blocks": [ + { + "bbox": [ + 276, + 194, + 342, + 245 + ], + "lines": [ + { + "bbox": [ + 276, + 194, + 342, + 245 + ], + "spans": [ + { + "bbox": [ + 276, + 194, + 342, + 245 + ], + "type": "image", + "image_path": "032f9634bb7e2c3bf8e3e8ad8571c44938c864467212772bb8c6e7f5a0a4d9d4.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 247, + 321, + 254 + ], + "lines": [ + { + "bbox": [ + 298, + 247, + 321, + 254 + ], + "spans": [ + { + "bbox": [ + 298, + 247, + 321, + 254 + ], + "type": "text", + "content": "Query 2" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 343, + 194, + 410, + 245 + ], + "blocks": [ + { + "bbox": [ + 343, + 194, + 410, + 245 + ], + "lines": [ + { + "bbox": [ + 343, + 194, + 410, + 245 + ], + "spans": [ + { + "bbox": [ + 343, + 194, + 410, + 245 + ], + "type": "image", + "image_path": "2dc0580fd2dffdb9dcbb92165e32c82db59c0bb3e6251519d7d94e3caa5cd1a0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 247, + 389, + 254 + ], + "lines": [ + { + "bbox": [ + 365, + 247, + 389, + 254 + ], + "spans": [ + { + "bbox": [ + 365, + 247, + 389, + 254 + ], + "type": "text", + "content": "Query 3" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 411, + 194, + 479, + 245 + ], + "blocks": [ + { + "bbox": [ + 411, + 194, + 479, + 245 + ], + "lines": [ + { + "bbox": [ + 411, + 194, + 479, + 245 + ], + "spans": [ + { + "bbox": [ + 411, + 194, + 479, + 245 + ], + "type": "image", + "image_path": "41507cc8848fab782616d302aebb8e79163fe5ba25fc3b34e16025f4698a5a4a.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 247, + 457, + 254 + ], + "lines": [ + { + "bbox": [ + 433, + 247, + 457, + 254 + ], + "spans": [ + { + "bbox": [ + 433, + 247, + 457, + 254 + ], + "type": "text", + "content": "Query 4" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 135, + 279, + 203, + 326 + ], + "blocks": [ + { + "bbox": [ + 135, + 279, + 203, + 326 + ], + "lines": [ + { + "bbox": [ + 135, + 279, + 203, + 326 + ], + "spans": [ + { + "bbox": [ + 135, + 279, + 203, + 326 + ], + "type": "image", + "image_path": "3d0fd57ad55436340e87cfe02370b24f4ec48bde01cfc2728062f6acb966a407.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 327, + 178, + 334 + ], + "lines": [ + { + "bbox": [ + 161, + 327, + 178, + 334 + ], + "spans": [ + { + "bbox": [ + 161, + 327, + 178, + 334 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 356, + 482, + 380 + ], + "lines": [ + { + "bbox": [ + 131, + 356, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 131, + 356, + 482, + 380 + ], + "type": "text", + "content": "Fig. 6: High response areas with top scores to input question in Equation 1. We present the response of first four queries in " + }, + { + "bbox": [ + 131, + 356, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t" + }, + { + "bbox": [ + 131, + 356, + 482, + 380 + ], + "type": "text", + "content": ". Images are sampled from VQA V2 test-dev set." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 208, + 279, + 274, + 326 + ], + "blocks": [ + { + "bbox": [ + 250, + 258, + 364, + 269 + ], + "lines": [ + { + "bbox": [ + 250, + 258, + 364, + 269 + ], + "spans": [ + { + "bbox": [ + 250, + 258, + 364, + 269 + ], + "type": "text", + "content": "Question: Is this a fishing village?" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 208, + 279, + 274, + 326 + ], + "lines": [ + { + "bbox": [ + 208, + 279, + 274, + 326 + ], + "spans": [ + { + "bbox": [ + 208, + 279, + 274, + 326 + ], + "type": "image", + "image_path": "9dfe53c6317b6802cf82358d866028009350c3050522da42907dc85ec0290f62.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 327, + 253, + 333 + ], + "lines": [ + { + "bbox": [ + 230, + 327, + 253, + 333 + ], + "spans": [ + { + "bbox": [ + 230, + 327, + 253, + 333 + ], + "type": "text", + "content": "Query 1" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 206, + 337, + 407, + 347 + ], + "lines": [ + { + "bbox": [ + 206, + 337, + 407, + 347 + ], + "spans": [ + { + "bbox": [ + 206, + 337, + 407, + 347 + ], + "type": "text", + "content": "Question: What tall animals are displayed in the background?" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 276, + 279, + 342, + 326 + ], + "blocks": [ + { + "bbox": [ + 276, + 279, + 342, + 326 + ], + "lines": [ + { + "bbox": [ + 276, + 279, + 342, + 326 + ], + "spans": [ + { + "bbox": [ + 276, + 279, + 342, + 326 + ], + "type": "image", + "image_path": "bbc34d68c49ec350d0f17fce2a6552fea86bf83ea4cb4b13746b11af6dcd8c5a.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 327, + 321, + 333 + ], + "lines": [ + { + "bbox": [ + 298, + 327, + 321, + 333 + ], + "spans": [ + { + "bbox": [ + 298, + 327, + 321, + 333 + ], + "type": "text", + "content": "Query 2" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 343, + 279, + 410, + 326 + ], + "blocks": [ + { + "bbox": [ + 343, + 279, + 410, + 326 + ], + "lines": [ + { + "bbox": [ + 343, + 279, + 410, + 326 + ], + "spans": [ + { + "bbox": [ + 343, + 279, + 410, + 326 + ], + "type": "image", + "image_path": "c8ca05f8bef6da9eec0e0b50a3ef145e2fdc79664f5730d70d1c0c3d5ea895a7.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 327, + 389, + 334 + ], + "lines": [ + { + "bbox": [ + 365, + 327, + 389, + 334 + ], + "spans": [ + { + "bbox": [ + 365, + 327, + 389, + 334 + ], + "type": "text", + "content": "Query 3" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 411, + 279, + 479, + 326 + ], + "blocks": [ + { + "bbox": [ + 411, + 279, + 479, + 326 + ], + "lines": [ + { + "bbox": [ + 411, + 279, + 479, + 326 + ], + "spans": [ + { + "bbox": [ + 411, + 279, + 479, + 326 + ], + "type": "image", + "image_path": "924c4718926950fe8014999ce472fbd48fde2a71e3e93bfa0dbe187ec57acbfc.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 327, + 457, + 334 + ], + "lines": [ + { + "bbox": [ + 433, + 327, + 457, + 334 + ], + "spans": [ + { + "bbox": [ + 433, + 327, + 457, + 334 + ], + "type": "text", + "content": "Query 4" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "bbox": [ + 131, + 401, + 312, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 401, + 312, + 414 + ], + "spans": [ + { + "bbox": [ + 131, + 401, + 312, + 414 + ], + "type": "text", + "content": "5 Discussion and Conclusion" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 130, + 426, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 557 + ], + "type": "text", + "content": "We have introduced LLaMA-VID, a simple yet effective token generation approach for VLMs. The central concept behind LLaMA-VID is to represent an image with the context token and the content token. In particular, the context token is generated according to input instructions, and the content token is produced based on the image content. Depending on the budget, the content token can be compressed to one token or expressed without compression. It allows us to represent a single image with preserved details and efficiently encode each video frame with only two tokens. Moreover, we have constructed an instruction dataset for hour-long video understanding. Our experiments on several video-and image-based benchmarks prove the superiority of our method. We hope that LLaMA-VID can serve as a strong benchmark for efficient visual representation." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "type": "text", + "content": "There still exists certain limitations in the current method. Although LLaMAVID succeeds to represent hour-long videos in LLMs, it focuses more on temporal consistency and may lose spatial details in each frame. In the future work, we plan to maintain dense spatial information for each frame outside the LLM and conduct the visual token fusion within LLMs to avoid such detail loss." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 131, + 629, + 481, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 481, + 654 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 481, + 654 + ], + "type": "text", + "content": "Acknowledgement. This work was supported in part by the Research Grants Council under the Areas of Excellence scheme grant AoE/E-601/22-R." + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "type": "text", + "content": "L. Yanwei et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 140, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 140, + 330, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 330, + 150 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 330, + 150 + ], + "type": "text", + "content": "1. Sharegpt. https://sharegpt.com/ (2023) 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 151, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 151, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 151, + 481, + 183 + ], + "type": "text", + "content": "2. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. In: NeurIPS (2022) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 184, + 476, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 476, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 476, + 194 + ], + "type": "text", + "content": "3. Anthropic: Claude 2. https://www.anthropic.com/index/claude-2 (2023) 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 481, + 226 + ], + "type": "text", + "content": "4. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv:2308.12966 (2023) 10" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 228, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 481, + 248 + ], + "type": "text", + "content": "5. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: ICCV (2021) 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 481, + 281 + ], + "type": "text", + "content": "6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. In: NeurIPS (2020) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 282, + 481, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 282, + 481, + 313 + ], + "spans": [ + { + "bbox": [ + 138, + 282, + 481, + 313 + ], + "type": "text", + "content": "7. Caba Heilbron, F., Escorcia, V., Ghanem, B., Carlos Niebles, J.: Activitynet: A large-scale video benchmark for human activity understanding. In: CVPR (2015) 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 315, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 481, + 335 + ], + "type": "text", + "content": "8. Chen, D., Dolan, W.B.: Collecting highly parallel data for paraphrase evaluation. In: ACL (2011) 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 337, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 481, + 357 + ], + "type": "text", + "content": "9. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv:2306.15195 (2023) 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 358, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 358, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 134, + 358, + 481, + 391 + ], + "type": "text", + "content": "10. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollár, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv:1504.00325 (2015) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 134, + 392, + 481, + 413 + ], + "type": "text", + "content": "11. Chen, Y., Qian, S., Tang, H., Lai, X., Liu, Z., Han, S., Jia, J.: Longlora: Efficient fine-tuning of long-context large language models. arXiv:2309.12307 (2023) 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 414, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 414, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 134, + 414, + 481, + 456 + ], + "type": "text", + "content": "12. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 134, + 414, + 481, + 456 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 134, + 414, + 481, + 456 + ], + "type": "text", + "content": " chatgpt quality. https://lmsys.org/blog/2023-03-30-vicuna/ (2023) 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 457, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 457, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 134, + 457, + 481, + 489 + ], + "type": "text", + "content": "13. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv:2305.06500 (2023) 1, 2, 4, 5, 6, 8, 10, 12" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 490, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 490, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 134, + 490, + 481, + 521 + ], + "type": "text", + "content": "14. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv:1810.04805 (2018) 2, 3, 5, 12" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 522, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 522, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 134, + 522, + 481, + 555 + ], + "type": "text", + "content": "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 555, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 481, + 588 + ], + "type": "text", + "content": "16. Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale. In: CVPR (2023) 5, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 589, + 481, + 621 + ], + "type": "text", + "content": "17. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv:2306.13394 (2023) 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 481, + 665 + ], + "type": "text", + "content": "18. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The\" something something\" video database for learning and evaluating visual common sense. In: ICCV (2017) 8, 10" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "LLaMA-VID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "19. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In: CVPR (2017) 3, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "20. Gurari, D., Li, Q., Stangl, A.J., Guo, A., Lin, C., Grauman, K., Luo, J., Bigham, J.P.: Vizwiz grand challenge: Answering visual questions from blind people. In: CVPR (2018) 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 206 + ], + "type": "text", + "content": "21. Huang, Q., Xiong, Y., Rao, A., Wang, J., Lin, D.: Movienet: A holistic dataset for movie understanding. In: ECCV (2020) 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 481, + 228 + ], + "type": "text", + "content": "22. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: CVPR (2019) 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 229, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 481, + 251 + ], + "type": "text", + "content": "23. IDEFICS: Introducing idefics: An open reproduction of state-of-the-art visual language model. https://huggingface.co/blog/idefics (2023) 10" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 252, + 481, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 252, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 481, + 285 + ], + "type": "text", + "content": "24. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 285, + 481, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 307 + ], + "type": "text", + "content": "25. Kazemzadeh, S., Ordonez, V., Matten, M., Berg, T.: Referitgame: Referring to objects in photographs of natural scenes. In: EMNLP (2014) 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 341 + ], + "type": "text", + "content": "26. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV (2017) 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 342, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 363 + ], + "type": "text", + "content": "27. Lai, X., Tian, Z., Chen, Y., Li, Y., Yuan, Y., Liu, S., Jia, J.: Lisa: Reasoning segmentation via large language model. arXiv:2308.00692 (2023) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 364, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 481, + 385 + ], + "type": "text", + "content": "28. Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv:2307.16125 (2023) 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "29. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv:2301.12597 (2023) 1, 2, 4, 10" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "type": "text", + "content": "30. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chat-centric video understanding. arXiv:2305.06355 (2023) 1, 4, 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 442, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 464 + ], + "type": "text", + "content": "31. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv:2305.10355 (2023) 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 465, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 465, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 465, + 481, + 486 + ], + "type": "text", + "content": "32. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv:2310.03744 (2023) 6, 7, 8, 10" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 487, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 508 + ], + "type": "text", + "content": "33. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeruIPS (2023) 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 509, + 481, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 481, + 532 + ], + "type": "text", + "content": "34. Liu, R., Li, C., Ge, Y., Shan, Y., Li, T.H., Li, G.: One for all: Video conversation is feasible without video instruction tuning. arXiv:2309.15785 (2023) 9" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 533, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 481, + 564 + ], + "type": "text", + "content": "35. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. arXiv:1907.11692 (2019) 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 566, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 481, + 598 + ], + "type": "text", + "content": "36. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv:2307.06281 (2023) 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "text", + "content": "37. Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. In: NeurIPS (2022) 3, 8, 12" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "38. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability. arXiv:2306.07207 (2023) 1" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "type": "text", + "content": "L. Yanwei et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 148 + ], + "type": "text", + "content": "39. Maaz, M., Rasheed, H., Khan, S., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv:2306.05424 (2023) 1, 4, 7, 8, 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 481, + 171 + ], + "type": "text", + "content": "40. Mangalam, K., Akshulakov, R., Malik, J.: Egoschema: A diagnostic benchmark for very long-form video language understanding. NeurIPS (2024) 8, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 171, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 171, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 481, + 192 + ], + "type": "text", + "content": "41. Mao, J., Huang, J., Toshev, A., Camburu, O., Yuille, A.L., Murphy, K.: Generation and comprehension of unambiguous object descriptions. In: CVPR (2016) 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 192, + 481, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 192, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 481, + 213 + ], + "type": "text", + "content": "42. Mao, J., Huang, J., Toshev, A., Camburu, O., Yuille, A.L., Murphy, K.: Generation and comprehension of unambiguous object descriptions. In: CVPR (2016) 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 213, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 481, + 236 + ], + "type": "text", + "content": "43. Mishra, A., Shekhar, S., Singh, A.K., Chakraborty, A.: Ocr-vqa: Visual question answering by reading text in images. In: ICDAR (2019) 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 236, + 427, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 427, + 245 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 427, + 245 + ], + "type": "text", + "content": "44. OpenAI: Chatgpt. https://openai.com/blog/chatgpt/ (2023) 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 246, + 413, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 413, + 256 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 413, + 256 + ], + "type": "text", + "content": "45. OpenAI: Gpt-4 technical report. arXiv:2303.08774 (2023) 1, 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 256, + 481, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 256, + 481, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 481, + 288 + ], + "type": "text", + "content": "46. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. In: NeurIPS (2022) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 288, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 481, + 320 + ], + "type": "text", + "content": "47. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 320, + 481, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 481, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 481, + 351 + ], + "type": "text", + "content": "48. Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: A-okvqa: A benchmark for visual question answering using world knowledge. In: ECCV (2022) 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 353, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 481, + 385 + ], + "type": "text", + "content": "49. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: ACL (2018) 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 385, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 481, + 407 + ], + "type": "text", + "content": "50. Sidorov, O., Hu, R., Rohrbach, M., Singh, A.: Textcaps: a dataset for image captioning with reading comprehension. In: ECCV (2020) 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 407, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 481, + 429 + ], + "type": "text", + "content": "51. Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards vqa models that can read. In: CVPR (2019) 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 429, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 481, + 460 + ], + "type": "text", + "content": "52. Song, E., Chai, W., Wang, G., Zhang, Y., Zhou, H., Wu, F., Chi, H., Guo, X., Ye, T., Zhang, Y., et al.: Moviechat: From dense token to sparse memory for long video understanding. In: CVPR (2024) 1, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 460, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 481, + 493 + ], + "type": "text", + "content": "53. Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., Hashimoto, T.B.: Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca (2023) 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 493, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 481, + 536 + ], + "type": "text", + "content": "54. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models. arXiv:2302.13971 (2023) 1, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 536, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 481, + 557 + ], + "type": "text", + "content": "55. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017) 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 557, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 590 + ], + "type": "text", + "content": "56. Wang, Y., Li, K., Li, Y., He, Y., Huang, B., Zhao, Z., Zhang, H., Xu, J., Liu, Y., Wang, Z., et al.: Intervideo: General video foundation models via generative and discriminative learning. arXiv:2212.03191 (2022) 9" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 590, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 481, + 622 + ], + "type": "text", + "content": "57. Wei, J., Bosma, M., Zhao, V.Y., Guu, K., Yu, A.W., Lester, B., Du, N., Dai, A.M., Le, Q.V.: Finetuned language models are zero-shot learners. arXiv:2109.01652 (2021) 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 622, + 481, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 481, + 644 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 481, + 644 + ], + "type": "text", + "content": "58. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv:2303.04671 (2023) 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 644, + 481, + 665 + ], + "type": "text", + "content": "59. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: CVPR (2016) 3, 8" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "LLaMA-VID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 346 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "60. Yang, A., Miech, A., Sivic, J., Laptev, I., Schmid, C.: Zero-shot video question answering via frozen bidirectional language models. In: NeurIPS (2022) 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "type": "text", + "content": "61. Yang, R., Song, L., Li, Y., Zhao, S., Ge, Y., Li, X., Shan, Y.: Gpt4tools: Teaching large language model to use tools via self-instruction. arXiv:2305.18752 (2023) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 194 + ], + "type": "text", + "content": "62. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv:2304.14178 (2023) 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "text", + "content": "63. Zhang, C., Lu, T., Islam, M.M., Wang, Z., Yu, S., Bansal, M., Bertasius, G.: A simple lvm framework for long-range video question-answering. arXiv:2312.17235 (2023) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 225, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 248 + ], + "type": "text", + "content": "64. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv:2306.02858 (2023) 1, 4, 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 248, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 480, + 281 + ], + "type": "text", + "content": "65. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv:2303.16199 (2023) 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "text", + "content": "66. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv:2205.01068 (2022) 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 314, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 480, + 346 + ], + "type": "text", + "content": "67. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv:2304.10592 (2023) 4" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "type": "text", + "content": "L. Yanwei et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file